diff --git a/.gitattributes b/.gitattributes index bb497928986ee643e6d489dfc1155f0107016816..907023af39d0664ea3c447a14961373c535f61e6 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2204,3 +2204,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Self-Reinforcing[[:space:]]Prototype[[:space:]]Evolution[[:space:]]with[[:space:]]Dual-Knowledge[[:space:]]Cooperation[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Lifelong[[:space:]]Person[[:space:]]Re-Identification/0d8ec9cc-91fd-4f6f-adc9-3bba1cad24c3_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Self-Supervised[[:space:]]Monocular[[:space:]]4D[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]for[[:space:]]Egocentric[[:space:]]Videos/0f34d99d-368d-4ba9-b31d-032659745be1_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Self-Supervised[[:space:]]Sparse[[:space:]]Sensor[[:space:]]Fusion[[:space:]]for[[:space:]]Long[[:space:]]Range[[:space:]]Perception/6d4fd112-3fac-483b-b978-3c134ff1cdd1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Self-supervised[[:space:]]Learning[[:space:]]of[[:space:]]Hybrid[[:space:]]Part-aware[[:space:]]3D[[:space:]]Representations[[:space:]]of[[:space:]]2D[[:space:]]Gaussians[[:space:]]and[[:space:]]Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SemGes_[[:space:]]Semantics-aware[[:space:]]Co-Speech[[:space:]]Gesture[[:space:]]Generation[[:space:]]using[[:space:]]Semantic[[:space:]]Coherence[[:space:]]and[[:space:]]Relevance[[:space:]]Learning/d193eee8-ef57-4a70-9609-6d91a9953312_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SemTalk_[[:space:]]Holistic[[:space:]]Co-speech[[:space:]]Motion[[:space:]]Generation[[:space:]]with[[:space:]]Frame-level[[:space:]]Semantic[[:space:]]Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]Alignment[[:space:]]and[[:space:]]Reinforcement[[:space:]]for[[:space:]]Data-Free[[:space:]]Quantization[[:space:]]of[[:space:]]Vision[[:space:]]Transformers/33638821-3dce-4a55-a900-82748a75aeee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]Causality-Aware[[:space:]]Vision-Based[[:space:]]3D[[:space:]]Occupancy[[:space:]]Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]Discrepancy-aware[[:space:]]Detector[[:space:]]for[[:space:]]Image[[:space:]]Forgery[[:space:]]Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]Equitable[[:space:]]Clustering_[[:space:]]A[[:space:]]Simple[[:space:]]and[[:space:]]Effective[[:space:]]Strategy[[:space:]]for[[:space:]]Clustering[[:space:]]Vision[[:space:]]Tokens/755b1314-9e24-437c-8f51-19b61bb62095_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]Watermarking[[:space:]]Reinvented_[[:space:]]Enhancing[[:space:]]Robustness[[:space:]]and[[:space:]]Generation[[:space:]]Quality[[:space:]]with[[:space:]]Fourier[[:space:]]Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic[[:space:]]versus[[:space:]]Identity_[[:space:]]A[[:space:]]Divide-and-Conquer[[:space:]]Approach[[:space:]]towards[[:space:]]Adjustable[[:space:]]Medical[[:space:]]Image[[:space:]]De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semantic-guided[[:space:]]Camera[[:space:]]Ray[[:space:]]Regression[[:space:]]for[[:space:]]Visual[[:space:]]Localization/d007a0e2-f228-4563-89f9-6176e65aa729_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semi-ViM_[[:space:]]Bidirectional[[:space:]]State[[:space:]]Space[[:space:]]Model[[:space:]]for[[:space:]]Mitigating[[:space:]]Label[[:space:]]Imbalance[[:space:]]in[[:space:]]Semi-Supervised[[:space:]]Learning/694bc6b8-9d54-4920-9947-577e19b17917_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semi-supervised[[:space:]]Concept[[:space:]]Bottleneck[[:space:]]Models/790be7e9-6d82-433e-9aab-92f2c72b545a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Semi-supervised[[:space:]]Deep[[:space:]]Transfer[[:space:]]for[[:space:]]Regression[[:space:]]without[[:space:]]Domain[[:space:]]Alignment/60f16256-f509-4463-b38b-322283424ecd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SemiVisBooster_[[:space:]]Boosting[[:space:]]Semi-Supervised[[:space:]]Learning[[:space:]]for[[:space:]]Fine-Grained[[:space:]]Classification[[:space:]]through[[:space:]]Pseudo-Label[[:space:]]Semantic[[:space:]]Guidance/2707eef4-2e69-485f-a5ee-896dce310667_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Separation[[:space:]]for[[:space:]]Better[[:space:]]Integration_[[:space:]]Disentangling[[:space:]]Edge[[:space:]]and[[:space:]]Motion[[:space:]]in[[:space:]]Event-based[[:space:]]Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SeqGrowGraph_[[:space:]]Learning[[:space:]]Lane[[:space:]]Topology[[:space:]]as[[:space:]]a[[:space:]]Chain[[:space:]]of[[:space:]]Graph[[:space:]]Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sequential[[:space:]]Gaussian[[:space:]]Avatars[[:space:]]with[[:space:]]Hierarchical[[:space:]]Motion[[:space:]]Context/ce271790-0752-41a2-ac20-e12f48492a27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sequential[[:space:]]keypoint[[:space:]]density[[:space:]]estimator_[[:space:]]an[[:space:]]overlooked[[:space:]]baseline[[:space:]]of[[:space:]]skeleton-based[[:space:]]video[[:space:]]anomaly[[:space:]]detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Serialization[[:space:]]based[[:space:]]Point[[:space:]]Cloud[[:space:]]Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ShadowHack_[[:space:]]Hacking[[:space:]]Shadows[[:space:]]via[[:space:]]Luminance-Color[[:space:]]Divide[[:space:]]and[[:space:]]Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Shape[[:space:]]of[[:space:]]Motion_[[:space:]]4D[[:space:]]Reconstruction[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ShortFT_[[:space:]]Diffusion[[:space:]]Model[[:space:]]Alignment[[:space:]]via[[:space:]]Shortcut-based[[:space:]]Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ShortV_[[:space:]]Efficient[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]by[[:space:]]Freezing[[:space:]]Visual[[:space:]]Tokens[[:space:]]in[[:space:]]Ineffective[[:space:]]Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Shot-by-Shot_[[:space:]]Film-Grammar-Aware[[:space:]]Training-Free[[:space:]]Audio[[:space:]]Description[[:space:]]Generation/b2b3829b-f130-4214-bc1a-c2339869b439_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SiM3D_[[:space:]]Single-instance[[:space:]]Multiview[[:space:]]Multimodal[[:space:]]and[[:space:]]Multisetup[[:space:]]3D[[:space:]]Anomaly[[:space:]]Detection[[:space:]]Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sibai_[[:space:]]A[[:space:]]Few-Shot[[:space:]]Meta-Classifier[[:space:]]for[[:space:]]Poisoning[[:space:]]Detection[[:space:]]in[[:space:]]Federated[[:space:]]Learning/c815ddc4-d777-4075-b792-39640cd474e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SignRep_[[:space:]]Enhancing[[:space:]]Self-Supervised[[:space:]]Sign[[:space:]]Representations/b0873b30-3769-4f25-84b2-07088f813479_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Signs[[:space:]]as[[:space:]]Tokens_[[:space:]]A[[:space:]]Retrieval-Enhanced[[:space:]]Multilingual[[:space:]]Sign[[:space:]]Language[[:space:]]Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sim-DETR_[[:space:]]Unlock[[:space:]]DETR[[:space:]]for[[:space:]]Temporal[[:space:]]Sentence[[:space:]]Grounding/461e9907-9af3-47f5-b610-f09041cded55_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SimMLM_[[:space:]]A[[:space:]]Simple[[:space:]]Framework[[:space:]]for[[:space:]]Multi-modal[[:space:]]Learning[[:space:]]with[[:space:]]Missing[[:space:]]Modality/398740ea-2adc-472d-af59-846c4c676c5a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Similarity[[:space:]]Memory[[:space:]]Prior[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]for[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SimpleVQA_[[:space:]]Multimodal[[:space:]]Factuality[[:space:]]Evaluation[[:space:]]for[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Simulating[[:space:]]Dual-Pixel[[:space:]]Images[[:space:]]From[[:space:]]Ray[[:space:]]Tracing[[:space:]]For[[:space:]]Depth[[:space:]]Estimation/daa8d996-9820-46cb-8540-26b2ea341407_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Simultaneous[[:space:]]Motion[[:space:]]And[[:space:]]Noise[[:space:]]Estimation[[:space:]]with[[:space:]]Event[[:space:]]Cameras/6f9dcd49-a085-404d-b287-141b9737e694_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Single-Scanline[[:space:]]Relative[[:space:]]Pose[[:space:]]Estimation[[:space:]]for[[:space:]]Rolling[[:space:]]Shutter[[:space:]]Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Skeleton[[:space:]]Motion[[:space:]]Words[[:space:]]for[[:space:]]Unsupervised[[:space:]]Skeleton-Based[[:space:]]Temporal[[:space:]]Action[[:space:]]Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SketchSplat_[[:space:]]3D[[:space:]]Edge[[:space:]]Reconstruction[[:space:]]via[[:space:]]Differentiable[[:space:]]Multi-view[[:space:]]Sketch[[:space:]]Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Skip-Vision_[[:space:]]Efficient[[:space:]]and[[:space:]]Scalable[[:space:]]Acceleration[[:space:]]of[[:space:]]Vision-Language[[:space:]]Models[[:space:]]via[[:space:]]Adaptive[[:space:]]Token[[:space:]]Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SkySense[[:space:]]V2_[[:space:]]A[[:space:]]Unified[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Multi-modal[[:space:]]Remote[[:space:]]Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sliced[[:space:]]Wasserstein[[:space:]]Bridge[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SliderSpace_[[:space:]]Decomposing[[:space:]]the[[:space:]]Visual[[:space:]]Capabilities[[:space:]]of[[:space:]]Diffusion[[:space:]]Models/50d20dec-7e48-4238-88fb-f013284e5706_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SmolDocling_[[:space:]]An[[:space:]]ultra-compact[[:space:]]vision-language[[:space:]]model[[:space:]]for[[:space:]]end-to-end[[:space:]]multi-modal[[:space:]]document[[:space:]]conversion/52fdd715-223a-408c-b30d-b6b134926ef6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Snakes[[:space:]]and[[:space:]]Ladders_[[:space:]]Two[[:space:]]Steps[[:space:]]Up[[:space:]]for[[:space:]]VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Social[[:space:]]Debiasing[[:space:]]for[[:space:]]Fair[[:space:]]Multi-modal[[:space:]]LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Soft[[:space:]]Local[[:space:]]Completeness_[[:space:]]Rethinking[[:space:]]Completeness[[:space:]]in[[:space:]]XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Soft[[:space:]]Separation[[:space:]]and[[:space:]]Distillation_[[:space:]]Toward[[:space:]]Global[[:space:]]Uniformity[[:space:]]in[[:space:]]Federated[[:space:]]Unsupervised[[:space:]]Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparfels_[[:space:]]Fast[[:space:]]Reconstruction[[:space:]]from[[:space:]]Sparse[[:space:]]Unposed[[:space:]]Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparse[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Transformers[[:space:]]for[[:space:]]Generative[[:space:]]Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparse-Dense[[:space:]]Side-Tuner[[:space:]]for[[:space:]]efficient[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding/87762bca-719d-433c-abf0-05e54d888e6d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseFlex_[[:space:]]High-Resolution[[:space:]]and[[:space:]]Arbitrary-Topology[[:space:]]3D[[:space:]]Shape[[:space:]]Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseLaneSTP_[[:space:]]Leveraging[[:space:]]Spatio-Temporal[[:space:]]Priors[[:space:]]with[[:space:]]Sparse[[:space:]]Transformers[[:space:]]for[[:space:]]3D[[:space:]]Lane[[:space:]]Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseMM_[[:space:]]Head[[:space:]]Sparsity[[:space:]]Emerges[[:space:]]from[[:space:]]Visual[[:space:]]Concept[[:space:]]Responses[[:space:]]in[[:space:]]MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseRecon_[[:space:]]Neural[[:space:]]Implicit[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]from[[:space:]]Sparse[[:space:]]Views[[:space:]]with[[:space:]]Feature[[:space:]]and[[:space:]]Depth[[:space:]]Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseVILA_[[:space:]]Decoupling[[:space:]]Visual[[:space:]]Sparsity[[:space:]]for[[:space:]]Efficient[[:space:]]VLM[[:space:]]Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparsity[[:space:]]Outperforms[[:space:]]Low-Rank[[:space:]]Projections[[:space:]]in[[:space:]]Few-Shot[[:space:]]Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial[[:space:]]Alignment[[:space:]]and[[:space:]]Temporal[[:space:]]Matching[[:space:]]Adapter[[:space:]]for[[:space:]]Video-Radar[[:space:]]Remote[[:space:]]Physiological[[:space:]]Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial[[:space:]]Preference[[:space:]]Rewarding[[:space:]]for[[:space:]]MLLMs[[:space:]]Spatial[[:space:]]Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial-Temporal[[:space:]]Aware[[:space:]]Visuomotor[[:space:]]Diffusion[[:space:]]Policy[[:space:]]Learning/635ea516-8718-4cd2-bbf5-79fa32431478_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial-Temporal[[:space:]]Forgery[[:space:]]Trace[[:space:]]based[[:space:]]Forgery[[:space:]]Image[[:space:]]Identification/b7964ec0-870b-444a-822c-785edaff983a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialCrafter_[[:space:]]Unleashing[[:space:]]the[[:space:]]Imagination[[:space:]]of[[:space:]]Video[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]from[[:space:]]Limited[[:space:]]Observations/359a0256-c24c-45c3-9f07-0205db934b25_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialSplat_[[:space:]]Efficient[[:space:]]Semantic[[:space:]]3D[[:space:]]from[[:space:]]Sparse[[:space:]]Unposed[[:space:]]Images/b5246924-8fb7-4786-baa2-29a57614062d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialTrackerV2_[[:space:]]Advancing[[:space:]]3D[[:space:]]Point[[:space:]]Tracking[[:space:]]with[[:space:]]Explicit[[:space:]]Camera[[:space:]]Motion/925f4fd6-2e60-44c7-a108-57075a837941_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatially-Varying[[:space:]]Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatio-Spectral[[:space:]]Pattern[[:space:]]Illumination[[:space:]]for[[:space:]]Direct[[:space:]]and[[:space:]]Indirect[[:space:]]Separation[[:space:]]from[[:space:]]a[[:space:]]Single[[:space:]]Hyperspectral[[:space:]]Image/c416257e-42cc-4603-837a-a5cb6ea0c724_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_content_list.json b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb2eb7b4647419cdc5e648953cc2998b8ee5cfde --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_content_list.json @@ -0,0 +1,1914 @@ +[ + { + "type": "text", + "text": "Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics", + "text_level": 1, + "bbox": [ + 158, + 130, + 839, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhirui Gao Renjiao Yi† Yuhang Huang Wei Chen Chenyang Zhu Kai Xu† National University of Defense Technology zhirui-gao.github.io/PartGS", + "bbox": [ + 158, + 202, + 836, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-level 3D representations, such as point clouds, meshes, NeRFs and 3D Gaussians, are commonly used for modeling 3D objects and scenes. However, cognitive studies indicate that human perception operates at higher levels and interprets 3D environments by decomposing them into meaningful structural parts, rather than low-level elements like points or voxels. Structured geometric decomposition enhances scene interpretability and facilitates downstream tasks requiring component-level manipulation. In this work, we introduce PartGS, a self-supervised part-aware reconstruction framework that integrates 2D Gaussians and superquadrics to parse objects and scenes into an interpretable decomposition, leveraging multi-view image inputs to uncover 3D structural information. Our method jointly optimizes superquadric meshes and Gaussians by coupling their parameters within a hybrid representation. On one hand, superquadrics enable the representation of a wide range of shape primitives, facilitating flexible and meaningful decompositions. On the other hand, 2D Gaussians capture detailed texture and geometric details, ensuring high-fidelity appearance and geometry reconstruction. Operating in a self-supervised manner, our approach demonstrates superior performance compared to state-of-the-art methods across extensive experiments on the DTU, ShapeNet, and real-world datasets.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D reconstruction from multi-view images is a long-standing challenge in 3D vision and graphics [12, 17, 54]. Most reconstructed scenes are in low-level representations such as point clouds, voxels, or meshes. However, humans tend to understand 3D scenes as reasonable parts [35]. For instance, when observing a scene, we naturally construct high-level structural information, such as scene graphs, instead of focusing on low-level details like point clouds or voxels. Motivated by it, we propose a part-aware reconstruction framework", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9b5112f0f0000698fc8e41981e9bcd34a1d0411e232cd65435bab3a447439ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 295, + 602, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7960d6e57f33b96c3164e2a169b996caa297f6960aa0b28d639943a539939304.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 296, + 710, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c70eff9910d387520dac0b67d636cab4ff433cabe32591d4789c5b760d389a60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 296, + 895, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fe9a196fd51c54769e7333f9a65c7064bf3e2dae3d63580702f399f6004a1b43.jpg", + "image_caption": [ + "3D decomposition", + "Prior work (EMS)" + ], + "image_footnote": [], + "bbox": [ + 517, + 383, + 604, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66b22a31d89c1e7f0e5b6e079ec97e74c5ccfba2e479122679f26b944fa75d42.jpg", + "image_caption": [ + "Block-level" + ], + "image_footnote": [], + "bbox": [ + 622, + 383, + 705, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c5961e98e585a5e6c1936f5358c07eb210b0bf051b91053e8f3f57be874c6ec.jpg", + "image_caption": [ + "Point-level", + "Ours", + "Figure 1. Prior works [32, 36] model the scene using primitives, while the proposed method can further model precise geometry details and textures. Here, EMS [32] takes point cloud inputs and reconstructs non-textured primitives." + ], + "image_footnote": [], + "bbox": [ + 714, + 378, + 803, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d654b2ecedca9d5196535791b9dbc553c6b5ddfd3bad5fe8253689a9104a1689.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 388, + 895, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that decomposes objects or scenes into meaningful shapes or parts, facilitating tasks such as physical simulation, editing, content generation and understanding [18, 28, 52, 53, 56].", + "bbox": [ + 511, + 550, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several prior works [29, 32, 33, 37, 55, 68] have explored part-aware reconstruction or 3D decomposition. However, most of them rely heavily on 3D supervision and often struggle to retain fine-grained geometric details, limiting their practicality in real-world scenarios. Recent advances [47, 65] have focused on extending the neural radiance field framework for part-aware reconstruction, building upon its remarkable success in 3D reconstruction from multi-view images. For example, PartNeRF [47] models objects using multiple neural radiance fields. Yet, the intricate composition of implicit fields complicates the learning process, leading to suboptimal rendering quality and inefficient decomposition. Recently, DBW [36] proposes a novel approach that decomposes scenes into block-based representations using superquadric primitives [2], optimizing both their parameters and UV texture maps through rendering loss minimization. While this method demonstrates effective scene decomposition into coherent components, it faces challenges in achieving accurate part-aware reconstruction of both geometry and appearance, as illustrated in Fig. 1.", + "bbox": [ + 511, + 598, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors", + "bbox": [ + 114, + 886, + 246, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9649", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose PartGS, a self-supervised part-aware hybrid representation that integrates 2D Gaussians [21] and superquadrics [2], to achieve both high-quality texture reconstruction and geometrically accurate part decomposition. There are previous approaches [14, 20, 49] combining mesh reconstruction and Gaussian splatting, where they first reconstruct the mesh and then attach Gaussians to it. In contrast to their primary focus on acquiring the representations, the proposed approach is to achieve part-aware decomposition reconstruction. Our method establishes a coupled optimization framework that simultaneously learns superquadric meshes and Gaussians through parameter sharing, with each part constrained to a single superquadric. The differentiable rendering of Gaussians drives this hybrid representation, leveraging the inherent convexity of superquadrics for qualitative 3D decomposition. The self-supervised training of part-aware reconstruction is built upon the assumption that each part of most objects should be a basic shape and can be represented by a superquadric. Meanwhile, Gaussian splatting, renowned for its superior rendering quality and efficient training, is incorporated to capture intricate texture details, speeding up reconstruction and improving rendering quality.", + "bbox": [ + 89, + 90, + 485, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the hybrid representation, 2D Gaussians are distributed on superquadric surfaces to form structured blocks. The pose and shape of Gaussians within each block are determined by their corresponding superquadrics rather than being optimized independently. The parameter space encompasses global controls for superquadric properties (shape, pose, and opacity) and local spherical harmonic coefficients for individual Gaussians. Compared to standard Gaussian splating [21, 25], which populates the occupancy volume with independently parameterized Gaussians, the coupled representation is more compact and efficient.", + "bbox": [ + 89, + 441, + 483, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training, the parameters are optimized through rendering loss without additional supervision. However, we notice that image rendering loss often leads to local minima in superquadric shape optimization. To tackle this, we introduce several regularizers to maintain global consistency between the 3D representation and the input 2D information. This strategic implementation allows us to segment parts fully self-supervised, achieving block-level reconstruction. Finally, to better capture irregular shapes, we implement a point-level refinement step that frees 2D Gaussians to deviate from the surface, thereby enhancing geometric fidelity. Extensive experiments show that PartGS, at block-level and point-level, achieves $33.3\\%$ and $75.9\\%$ improvements in reconstruction accuracy, 3.18 and 16.13 increases in PSNR, and 4X and 3X speedups compared to the state-of-the-art baseline. Our contributions are summarized as follows:", + "bbox": [ + 89, + 609, + 485, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a novel hybrid representation for part-aware 3D reconstruction, combining the strengths of superquadrics and Gaussian splatting to achieve reasonable", + "bbox": [ + 94, + 854, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "decomposition and high-quality rendering.", + "bbox": [ + 529, + 90, + 813, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose several novel regularizers to enforce consistency between 3D decomposition and 2D observations, enabling self-supervised part decomposition.", + "- Compared to prior works, the method takes one step forward, by reconstructing both the block-level and point-level part-aware reconstructions, preserving both part segmentation and reconstruction precision." + ], + "bbox": [ + 517, + 107, + 906, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 228, + 655, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Shape Decomposition and Abstraction", + "text_level": 1, + "bbox": [ + 511, + 253, + 844, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Structured shape representation learning decomposes objects or scenes into coherent geometric primitives, facilitating shape understanding and generation [9, 15, 19, 46, 57]. Early approaches like Blocks World [42] and Generalized Cylinders [3] emphasized compact representations. Modern methods typically process 3D inputs (point clouds, meshes, voxels) by decomposing them into primitive ensembles, including cuboids [41, 48], superquadrics [32, 38, 39, 58], and convex shapes [8, 10]. For instance, MonteBoxFinder [41] integrates clustering [11, 50, 60], cuboid representations, and Monte Carlo Tree Search, while EMS [32] adopts a probabilistic approach for superquadric recovery. However, these methods are limited to coarse shape representations. Recent advances in shape abstraction [13, 22, 31, 40, 45] enable more detailed shape representations through flexible primitive deformation.", + "bbox": [ + 511, + 276, + 906, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some studies [1, 16, 36, 47, 65, 66] attempt to create structure-aware 3D representations directly from images. PartNeRF [47] introduces ellipsoid representations within NeRFs, its reliance on multiple implicit fields results in inefficient 3D decomposition. ISCO [1] and DBW [36] use 3D superquadrics for shape decomposition, which enables more meaningful structural separation. However, their simple shape parameters struggle to capture complex geometries, leading to poor geometry and appearance reconstruction. DPA-Net [65] has advanced 3D shape abstraction from sparse views but generates redundant parts and struggles with realistic texture rendering. A concurrent work, GaussianBlock [24], employs SAM [27] to guide superquadric splitting and fusion for 3D decomposition, yet its computational efficiency remains limited, typically requiring several hours per scene. In contrast, our approach accomplishes self-supervised [51, 59] part-aware scene reconstruction through an efficient hybrid representation that simultaneously maintains geometric fidelity and photorealistic rendering quality.", + "bbox": [ + 511, + 518, + 908, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Mesh-based Gaussian Splatting", + "text_level": 1, + "bbox": [ + 511, + 816, + 792, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian splitting (GS) [25] has been rapidly adopted in multiple fields due to its remarkable rendering capability. Several studies [7, 14, 20, 49] attempt to align Gaussians with mesh surfaces for easier editing and animation. Among", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9650", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1725736d806238fdb08ae4afbf2598d85df1ee5bf8813240447110681c446bb3.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. PartGS takes multi-view images to learn a parametric hybrid representation of superquadrics and 2D Gaussians. It initializes from random superquadrics and is gradually optimized during training to obtain a block-level reconstruction. Then, we free the constraints of Gaussians to model detailed geometry, achieving point-level reconstruction." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "them, SuGaR [20] uses flat 3D Gaussians to enforce the alignment with the scene surface during optimization, minimizing the difference between the signed distance functions (SDF) of the desired Gaussians and actual Gaussians. GaMeS [49] introduces a hybrid representation of Gaussians and mesh, where Gaussians are attached to triangular facets of the mesh. Similarly, Gao et al. [14] proposes a mesh-based GS to achieve large-scale deformation effects on objects. Recently, 2DGS [21] proposes 2D Gaussians for surface modeling and significantly enhances the geometric quality. Inspired by them, we propose a representation that combines 2D Gaussians with superquadrics. A key distinction between our method and previous mesh-based GS approaches is that these methods require first reconstructing the scene's mesh and then binding Gaussians to the mesh surface, which results in a non-continuous mesh. In contrast, our method directly optimizes the mesh through a rendering loss, enabling part-aware mesh reconstruction.", + "bbox": [ + 91, + 412, + 483, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 705, + 181, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of calibrated multi-view images and foreground masks, we aim to learn part-aware 3D representations, a meaningful decomposition of both geometry and appearance. Our approach adopts a two-stage optimization strategy: first, decomposing the object into basic shapes using a mixture of Gaussians and superquadrics at the block-level, followed by refining the decomposition at the point-level to achieve precise geometry. In Sec 3.1, we parameterize the hybrid representation. Sec. 3.2 then elaborates on leveraging this representation, enhanced by novel regularizers and an adaptive control strategy, to achieve self-supervised block-level", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decomposition. Finally, Sec. 3.3 presents the process for obtaining detailed part-aware results.", + "bbox": [ + 511, + 412, + 906, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Parametrizing the Hybrid Representation", + "text_level": 1, + "bbox": [ + 511, + 450, + 870, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, to leverage the strengths of both, we attach Gaussians to the surface of superquadric meshes. This representation retains the superquadric's ability to parse a 3D scene into distinct parts. Meanwhile, spherical harmonics of Gaussians enable complex texture rendering via Gaussian splatting, addressing the texture learning limitations of prior work [32, 36, 41, 64]. Sharing pose parameters between superquadrics and 2D Gaussians further improves the representation's efficiency.", + "bbox": [ + 509, + 470, + 906, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The parametric representation is controlled by both primitive and Gaussian parameters, which are optimized simultaneously. Given a 3D scene $S$ , the proposed method decomposes it into multiple hybrid blocks, each consisting of a superquadric with associated Gaussians. Each scene is denoted by the hybrid representation: $S = \\mathcal{B}_1 \\cup \\ldots \\cup \\mathcal{B}_i \\cup \\mathcal{B}_M$ , where $\\mathcal{B}_i$ is the $i$ -th hybrid block, and $M$ is the total number of blocks. Blocks are defined using manually designed parameters that control pose, opacity, scale, shape, and texture. These parameters are optimized via differentiable rendering to parse the target scene.", + "bbox": [ + 511, + 608, + 908, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shape and Scale Parameters. For each hybrid block $\\mathcal{B}_i$ , its geometry is controlled by superquadric parameters [2]. Specifically, there are two shape parameters $\\epsilon_1, \\epsilon_2$ that define its shape, along with three scale parameters $s_1, s_2, s_3$ , which scale the 3D axes. Analogous to icosphere vertex positioning, superquadric vertex coordinates are computed by:", + "bbox": [ + 511, + 773, + 908, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\left[ s _ {1} \\cos^ {\\epsilon 1} (\\theta) \\cos^ {\\epsilon 2} (\\varphi); s _ {2} \\sin^ {\\epsilon 1} (\\theta); s _ {3} \\cos^ {\\epsilon 1} (\\theta) \\sin^ {\\epsilon 2} (\\varphi) \\right], \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 867, + 906, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ and $\\varphi$ represent the azimuth and elevation defined", + "bbox": [ + 511, + 885, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9651", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the spherical coordinate. The shape and scale parameters govern block deformation to learn part-aware geometry.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Parameters. The pose of the $i$ -th hybrid block is defined by its rotation $\\mathbf{R}_i$ and translation $\\mathbf{t}_i$ . The vertex position from Eq. 1 is transformed from the local coordinate to world space as: $\\hat{\\mathbf{v}}_i^j = \\mathbf{R}_i\\mathbf{v}_i^j + \\mathbf{t}_i$ , where $j$ indexes the vertices. Previous approaches, including SuGaR [20] and GaMeS [49], position Gaussians on reconstructed meshes with independent pose and shape parameters to enhance appearance modeling. In contrast, our method employs Gaussians to construct a differentiable rendering that bridges images and superquadrics. To achieve this, we couple the parameters of Gaussians with superquadrics.", + "bbox": [ + 88, + 121, + 482, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a posed superquadric, Gaussian centers are uniformly sampled on triangular faces, with their poses determined by face vertices. Following GaMeS [49], each Gaussian's rotation matrix $\\mathrm{R}_v = [r_1,r_2,r_3]$ and scaling $\\mathrm{S}_v$ are computed from vertex positions. Given a triangular face $V$ with vertices $v_{1},v_{2},v_{3}\\in \\mathbb{R}^{3}$ , orthonormal vectors are constructed such that $r_1$ aligns with the face normal, $r_2$ points from the centroid $m = \\mathrm{mean}(v_1,v_2,v_3)$ to $v_{1}$ . $r_3$ is obtained by orthonormalizing the vector from the center to the second vertex concerning the existing $r_1$ and $r_2$ :", + "bbox": [ + 89, + 287, + 483, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {3} = \\frac {\\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2}\\right)}{\\left\\| \\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2} \\right. \\right\\|}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 449, + 483, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{ort}$ represents the orthonormalization in Gram-Schmidt [4]. For the scaling of 2D Gaussians, we use:", + "bbox": [ + 89, + 486, + 483, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} _ {v} = \\operatorname {d i a g} \\left(s _ {v} ^ {2}, s _ {v} ^ {3}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 523, + 482, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_v^2 = c||m - v_1||$ , $s_v^3 = c < v_2, r_3 >$ , and $c$ is a size control hyperparameter. We place a fixed number of Gaussians on each triangular face. These designs eliminate the need to learn the geometry of Gaussians, thereby enhancing the efficiency of the representation.", + "bbox": [ + 89, + 551, + 483, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Opacity Parameters. We define the total number of hybrid blocks as $M$ . However, a typical scene does not contain exactly $M$ blocks. Therefore, only the meaningful blocks are retained. To achieve this, we introduce a learnable parameter $\\tau_{i}$ to represent each block's opacity. During the optimization, only blocks with $\\tau_{i}$ greater than a certain threshold are retained. Note that Gaussians within the same block share the same $\\tau$ for opacity in rasterization.", + "bbox": [ + 89, + 628, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture Parameters. The texture is modeled using 2D Gaussians positioned on the surface of the superquadrics. Spherical harmonics of each Gaussian control the texture and are optimized for rendering view-dependent images [25].", + "bbox": [ + 89, + 750, + 483, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Block-level Decomposition", + "text_level": 1, + "bbox": [ + 89, + 816, + 328, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section describes the optimization of the hybrid representation. We observed that minimizing the rendering loss across multi-view images alone led to instability in positioning hybrid blocks. Therefore, several regularization terms", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are introduced to optimize the composition for maximal image formation consistency.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Optimization", + "text_level": 1, + "bbox": [ + 511, + 128, + 651, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the image rendering loss, we employ regularizers consisting of coverage, overlap, parsimony, and opacity entropy. The coverage loss encourages hybrid blocks to cover only meaningful regions; the overlap loss prevents blocks from overlapping; the parsimony loss regularizes the number of existing blocks; and the opacity entropy encourages block opacities close to binary.", + "bbox": [ + 511, + 147, + 906, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rendering Loss. The rendering loss is based on 3DGS [25], where it combines an $L_{1}$ term with a D-SSIM term:", + "bbox": [ + 511, + 253, + 906, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n}} = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 297, + 906, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Coverage Loss.** The coverage loss ensures that the block set $\\{\\mathcal{B}_1 \\ldots \\mathcal{B}_M\\}$ covers the object while preventing it from extending beyond its boundaries. To determine the 3D occupancy field based on the blocks, we first define the approximate signed distance of a 3D point $x$ to the $i$ -th block, $D_i(x) = \\Psi_i(x) - 1$ . Here, $\\Psi_i$ denotes the superquadric inside-outside function [2] associated with the block $i$ . Consequently, $D_i(x) \\leq 0$ if the point $x$ lies inside the $i$ -th block, and $D_i(x) > 0$ if $x$ is outside the block. Inspired by NeRF [34], we sample a ray set $\\mathcal{R}$ using ray-casting based on camera poses. Given the object mask, we associate each ray $r \\in \\mathcal{R}$ with a binary label: $l_r = 1$ if the ray $r$ is inside mask, otherwise $l_r = 0$ . The coverage loss is defined as:", + "bbox": [ + 511, + 320, + 906, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {R}) = \\sum_ {r \\in \\mathcal {R}} l _ {r} L _ {\\text {c r o s s}} (r) + \\left(1 - l _ {r}\\right) L _ {\\text {n o n - c r o s s}} (r). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 526, + 906, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $L_{\\mathrm{cross}}$ encourages rays to intersect with blocks, while $L_{\\mathrm{non - cross}}$ discourages rays from intersecting with blocks:", + "bbox": [ + 511, + 568, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c r o s s}} (r) = \\operatorname {R e L U} \\left(\\min _ {i \\in \\mathcal {M}} \\min _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 609, + 906, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o n - c r o s s}} (r) = \\operatorname {R e L U} \\left(\\max _ {i \\in \\mathcal {M}} \\max _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} - D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 645, + 906, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{j}^{r}$ means the $j$ -th sampled point along the ray $r$ and $\\mathcal{M} = \\{1, \\dots, M\\}$ . Intuitively, this implies that at least one sampled points $\\mathcal{X}_{r}$ along the ray $r$ inside the mask that lies within a certain block, while all points along the ray outside the mask should not belong to any block.", + "bbox": [ + 511, + 681, + 905, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overlap Loss. We introduce a regularization term to prevent overlap between individual blocks. Given the difficulty of directly calculating block overlap, we adopt a Monte Carlo method similar to [36]. Specifically, we sample multiple 3D points in space and penalize those that lie inside more than $k$ blocks. Based on the superquadric inside-outside function, a soft occupancy function is defined as:", + "bbox": [ + 511, + 757, + 906, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {i} ^ {x} = \\tau_ {i} \\operatorname {s i g m o i d} \\left(\\frac {- D _ {i} (x)}{\\gamma}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 872, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9652", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a temperature parameter. Thus, for a set of $N$ sampled 3D points $\\Omega$ , the overlap loss is expressed as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o v e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\operatorname {R e L U} \\left(\\sum_ {i \\in \\mathcal {M}} \\mathcal {O} _ {i} ^ {x} - k\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 131, + 483, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parsimony Loss. To promote the use of the minimal number of blocks and achieve parsimony in decomposition, we introduce a regularization term that penalizes block opacity $(\\tau)$ . This loss is defined as:", + "bbox": [ + 89, + 184, + 483, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p a r}} = \\frac {1}{M} \\sum_ {i \\in \\mathcal {M}} \\sqrt {\\tau_ {i}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 253, + 483, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Opacity Entropy Loss. During optimization, the opacity of the block inside the object region tends to approach 1, while the opacity of the block outside the object region tends to approach 0. To facilitate this, we associate the opacities of blocks with the labeled rays and apply a cross-entropy loss $L_{ce}$ between the block opacity and the mask labels, defined as follows:", + "bbox": [ + 89, + 301, + 483, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o p a}} (\\mathcal {R}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} L _ {c e} \\left(\\max _ {i \\in \\mathcal {M}} \\tau_ {i} \\left(x ^ {r}\\right), l _ {r}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 415, + 483, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, only points $x^r$ inside the blocks are sampled, e.g., $\\{x^r \\in \\mathcal{X}_r \\mid D_i(x^r) \\leq 0\\}$ .", + "bbox": [ + 89, + 463, + 483, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss is the weighted sum of the loss terms described above:", + "bbox": [ + 89, + 493, + 483, + 522 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n}} + \\lambda_ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + \\lambda_ {\\text {o v e r}} \\mathcal {L} _ {\\text {o v e r}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {o p a}} \\mathcal {L} _ {\\text {o p a}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 535, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Adaptive Number of Blocks.", + "text_level": 1, + "bbox": [ + 89, + 563, + 333, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that the number of parts may vary across different scenes, we allow the number of blocks to adjust during optimization adaptively. Specifically, when the opacity of an existing block falls below a threshold $t$ , this block is removed immediately. Furthermore, we explore a block-adding mechanism to dynamically incorporate new components, ensuring comprehensive coverage of target objects. Specifically, DBSCAN [44] is employed to cluster initial point clouds that are not covered by any blocks, where the point clouds are either randomly initialized or derived as by-products from COLMAP [43]. For each point cloud cluster, a new block is introduced at its center, with the remaining parameters of the block initialized randomly.", + "bbox": [ + 89, + 582, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final block-level reconstruction generates a textured 3D scene represented by multiple superquadrics, with surface details rendered through Gaussians, as illustrated in Fig. 2.", + "bbox": [ + 89, + 779, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Point-level Decomposition", + "text_level": 1, + "bbox": [ + 89, + 848, + 326, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The primitive-based hybrid representation effectively decomposes the shape into parts but exhibits suboptimally for", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "complex objects. To address this limitation, we perform a refinement stage that enhances geometric fitting. In this stage, the constraint between Gaussians and superquadrics is decoupled, enabling independent Gaussian optimization. Furthermore, to prevent Gaussian components from one block passing through adjacent blocks and disrupting the continuity and plausibility of the part decomposition, we impose a new constraint to minimize the negative signed distance of each Gaussian entering other blocks:", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e n t e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\sum_ {m \\in \\mathcal {M} \\backslash \\{\\delta \\}} \\operatorname {R e L U} (- D _ {m} (x)), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 232, + 906, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega$ is the sampled Gaussian set, and $\\delta$ represents the identifier of the block to which the Gaussian $x$ belongs. As seen in Fig. 2, the reconstruction becomes more accurate and aligns better with the target shape.", + "bbox": [ + 511, + 276, + 905, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 349, + 645, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We comprehensively evaluate our approach across four key aspects: 3D reconstruction quality, view synthesis performance, shape parsimony, and computational efficiency, comparing it with state-of-the-art methods in the field.", + "bbox": [ + 511, + 373, + 906, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 441, + 614, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extensive experiments are conducted on two widely-used public datasets: DTU [23] and ShapeNet [6]. Specifically, 15 standard scenes from the DTU dataset that are widely adopted for reconstruction quality assessment are used for evaluation. Each scene contains either 49 or 64 images. Additionally, we construct a ShapeNet subset consisting of four categories: Chair, Table, Gun, and Airplane, with 15 objects selected per category. For each object, 50 rendered images are generated for training. We further validate real-world applicability on both the BlendedMVS dataset [61] and self-captured scenes, demonstrating robust performance across synthetic and natural environments.", + "bbox": [ + 511, + 463, + 908, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 652, + 730, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We build upon the 2DGS [21] architecture and add a nonlearnable part attribute to each Gaussian component to enforce the constraint in Eq. 13. In the ShapeNet benchmark, we observe a common degradation in mesh reconstruction as shown in Fig. 7. To address this, we propose Gaussian scale regularization, as validated in Sec. 4.4. We set the initial number of primitives $M$ to 8. The temperature parameter $\\gamma$ in Eq. 8 is 0.005 and the overlapping number $k$ in Eq. 9 is 1.95. We train the hybrid representation for 30k iterations, followed by refinement optimization for another 30k iterations. During block-level optimization, the block-adding operation is carried out at the 5k-th and 10k-th iterations. The loss weights $\\lambda_{\\mathrm{cov}}$ , $\\lambda_{\\mathrm{over}}$ , $\\lambda_{\\mathrm{par}}$ , and $\\lambda_{\\mathrm{opa}}$ are set to 10, 1, 0.002, and 0.01, respectively. For more details about datasets, metrics, and baselines, please refer to Supplementary.", + "bbox": [ + 511, + 674, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a2893f0459bd4436057f44847546ed02f6e3881a4ec7cc693ace7254912d196.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on DTU [23] and ShapeNet [23]. The first two rows are DTU examples, and the last two are ShapeNet examples, respectively. Our method is the only one that provides reasonable 3D part decomposition while capturing detailed geometry." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 397 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e5f3de70384992d0ad68fdd1fc6188968ad3b4bf69304e88bbf5d51915da125.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputRenderableChamfer distance per sceneMeanMean
S24S37S40S55S63S65S69S83S97S105S106S110S114S118S122CD#P
EMS [32]3D GTX6.323.542.994.304.164.013.753.244.974.344.167.627.584.464.034.657.7
MBF [41]3D GTX3.122.663.842.541.592.112.192.012.322.452.172.123.832.022.552.5034.1
EMS [32] + Neus [54]ImageX5.995.564.434.325.426.143.753.964.634.345.885.114.294.833.534.978.87
MBF [41] + Neus [54]ImageX2.693.373.222.693.632.602.593.132.852.512.453.722.242.492.522.8546.7
PartNeRF [47]Image✓9.3810.469.088.636.047.257.229.158.7210.016.729.857.858.689.218.548.0
DBW [36]Image✓5.418.351.573.083.404.157.463.946.634.854.384.656.294.343.044.764.8
Ours (Block-level)Image✓5.684.911.852.613.754.663.757.574.274.383.494.483.614.213.704.195.9
Ours (Point-level)Image✓0.701.170.550.651.061.231.101.361.370.780.921.410.691.050.710.985.9
", + "bbox": [ + 91, + 455, + 903, + 578 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0a458dcd15aa25583c5f3d0ea2026d6871595758658bcfb85d33ef84a9c85d01.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on DTU [23]. The Chamfer distance between the 3D reconstruction and the ground-truth is reported in 15 scenes. The best results are bolded, and the average numbers of primitives found (#P) that are greater than 10 are underlined." + ], + "table_footnote": [ + "Table 2. Quantitative results on DTU [23]. Our method outperforms all part-aware approaches in image synthesis quality, reconstruction accuracy, and efficiency. Neuralangelo's results are from the original paper, with all times measured on an RTX 3090 GPU." + ], + "table_body": "
MethodPart-awareCD ↓PSNR ↑SSIM ↑LIPPS ↓Time ↓
Neuralangelo [30]X0.6133.84--> 10 h
2DGS [21]X0.8134.070.990.019~ 10 m
PartNeRF [47]✓9.5917.970.770.246~ 8 h
DBW [36]✓4.7316.440.750.201~ 2 h
Ours (Block-level)✓4.1919.840.820.189~ 30 m
Ours (Point-level)✓0.9835.040.990.015~ 40 m
", + "bbox": [ + 93, + 632, + 480, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluations", + "text_level": 1, + "bbox": [ + 89, + 794, + 218, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. DTU and Shapenet Benchmark", + "text_level": 1, + "bbox": [ + 89, + 819, + 356, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry Reconstruction. In Tab. 1 and Tab. 3, we compare our geometry reconstruction to SOTA shape decomposition methods on Chamfer distance and training time using DTU and Shapenet dataset. The Chamfer distance", + "bbox": [ + 89, + 839, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a74c8f1cb64e244e76ba97da6521ae10573bb16f939493226b77961fd2ab850d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInputChamfer Distance ↓Primitives (#P)MeanMean
AirplaneTableChairGunAirplaneTableChairGunCD#P
EMS [32]3D GT3.406.9219.02.029.47.8810.38.47.849.0
MBF [41]3D GT2.832.181.592.3210.8513.913.414.32.2113.1
PartNeRF [47]Image2.292.772.302.468.08.08.08.02.468.0
DBW [36]Image3.617.336.192.092.75.23.63.34.813.7
Ours (Block-level)Image2.472.152.321.783.96.67.65.02.185.8
Ours (Point-level)Image1.291.720.941.073.96.67.65.01.255.8
", + "bbox": [ + 514, + 632, + 905, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison on ShapeNet [6]. We report Chamfer distance and the number of parts. The best results are bolded, and the second-best results are underlined.", + "bbox": [ + 511, + 729, + 906, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metrics reported for ShapeNet are scaled by a factor of 100 for readability. Our method consistently outperforms in all scenes compared to prior works. As shown in Fig. 3, our approach consistently produces interpretable 3D decompositions with further refinement achieving more detailed geometry (the last two columns). MBF [41] achieves a low CD error but at the cost of using significantly more prim", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c0ed874782a95993fdfe8b944acb6d17edbc29d67c181e5cee8edef5ddc27757.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPart-awareChamfer Distance ↓PSNR ↑SSIM ↑LIPPS ↓
AirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGun
2DGS [21]✗1.472.370.501.0340.8939.8039.0541.740.9940.9900.9900.9940.0090.0270.0170.009
PartNeRF [47]✓2.292.772.302.4619.6320.6619.0821.970.8980.8550.8750.9160.0860.1610.1360.083
DBW [36]✓3.617.336.192.0926.1123.8420.2528.720.9500.9150.8920.9600.0740.1360.1320.042
Ours (Block-level)✓2.472.152.321.7827.9427.9824.9229.950.9590.9250.9060.9630.0720.1290.1220.047
Ours (Point-level)✓1.291.720.941.0741.1836.8036.0739.510.9920.9730.9770.9890.0140.0700.0380.021
", + "bbox": [ + 91, + 93, + 903, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24ae7e5a67815b404bd07fee031db8fba68f96091208d7f64af3ec24f97d31af.jpg", + "image_caption": [ + "Figure 4. Qualitative results on BlendedMVS [61] and self-captured data. We present RGB renderings and decomposed parts from novel views. The top examples are from the BlendedMVS dataset, and the last example is from our captured scenes." + ], + "image_footnote": [], + "bbox": [ + 91, + 218, + 230, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c926f3ba22a56d86a32c5e06a2a597aa03f4573cc5f5d110f7785cc401a7d7e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 218, + 310, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f27da3159df474a17048759326d7e7f2cf043bbb41b0910ae41fe2687b249422.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 218, + 387, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a9fecf992c3b20d510816d220905d3c02c3f955ff86d2a7b933ec78efbd4807.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with SAM-based methods. Visual results demonstrate that our method produces more structurally coherent decompositions, whereas SAM-based approaches frequently exhibit spatial discontinuities." + ], + "image_footnote": [], + "bbox": [ + 405, + 218, + 470, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0cf1c6b0b7ee14a1103a679235523a6b7e3143fe7142ee49d274421f4a041d4a.jpg", + "image_caption": [ + "Figure 5. Ablation studies on key strategies. The block-level visual comparisons illustrate the impact of adopting our proposed strategy. The first row shows results without the strategy, and the second with the strategy implemented." + ], + "image_footnote": [], + "bbox": [ + 91, + 484, + 478, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ities, leading to over-decomposition and the inclusion of meaningless parts. Moreover, as shown in Tab. 2 and Tab. 4, our approach achieves competitive results with the advanced Gaussian-based 2DGS [21] and surface reconstruction Neuralangelo [30], which are limited to producing unstructured meshes. Notably, our model demonstrates excellent efficiency, with a reconstruction speed approximately $10\\mathrm{X}$ faster than PartNeRF [47] and over $3\\mathrm{X}$ faster than DBW [36].", + "bbox": [ + 89, + 732, + 483, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Appearance Reconstruction. In addition to part decomposition, our method enables high-fidelity image synthesis, attributed to integrating Gaussian splatting within hybrid", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/307d665fea83eea93e22db493cd3d7d1a8350fa763b31b558dd767118b793551.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 218, + 589, + 369 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/70407c2d856a6e1c769162ddf9593c72e1736e682f1fbac84e02effa1cfda752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 218, + 709, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10cdb9e196463b04c8405d36179586a3b92d152ed2795781a4014c35037727cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 218, + 823, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4111f7b72c108581367c01f4837d509695d917e6f151090dc0b0072f30e0d65e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 218, + 903, + 371 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec3d9a1a29ac736dfa20c8af29ad64edcb2050e51c874a82c80940963d1f7e11.jpg", + "table_caption": [ + "Table 4. Quantitative results on ShapeNet [6]. We report the Chamfer distance and novel view synthesis results across four categories." + ], + "table_footnote": [], + "table_body": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
Complete model2.3224.920.9060.1220.9436.070.9770.0386.6
w/o Lover2.1125.170.9140.1181.3436.040.9770.0377.6
w/o Lopa5.7221.120.8530.1741.0635.600.9750.0403.8
w/o Lcov3.2122.960.8960.1341.3436.000.9760.0388.53
w/o Adaptive3.1622.740.8800.1411.0535.740.9740.0406.7
w/o Lpar1.9125.480.9180.1150.9336.180.9770.03710.1
", + "bbox": [ + 516, + 454, + 903, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the ShapeNet [6]. We report Chamfer Distance, rendering metrics, and the number of primitives (#P).", + "bbox": [ + 511, + 544, + 906, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "representations. EMS [32] and MBF [41] operate directly on point clouds and consequently lack image rendering capabilities, while PartNeRF [47] and DBW [36] yield low-quality view synthesis. In contrast, our method achieves high-quality appearance rendering results, as demonstrated in Tab. 2 and Tab. 4.", + "bbox": [ + 511, + 601, + 906, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Real-life Data", + "text_level": 1, + "bbox": [ + 511, + 700, + 656, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the applicability of our method for learning shape decomposition, we test our model on the real-life images from the BlendedMVS dataset [61] and a self-captured dataset. As shown in Fig. 4, our approach can robustly produce both realistic appearances and reasonable 3D decompositions across a variety of data types. More results are provided in the supplementary material.", + "bbox": [ + 511, + 720, + 905, + 827 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3. Compared to SAM-based Methods", + "text_level": 1, + "bbox": [ + 511, + 835, + 802, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct comparisons to SAM-based [27] methods for this task, as shown in Fig. 6. Despite impressive advances in 3D segmentation and editing [5, 26, 62, 63] by", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9655", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c82033cd52c3d624329aede2b26145e5a5fa59c98af17cea26b89603aa1ec4e6.jpg", + "image_caption": [ + "Figure 7. Impact of Gaussian scale regularization (SR). The degraded mesh produced by 2DGS [21] is effectively improved." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 480, + 159 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/143f6800a56837a75210e2b7e2342ddd66ce215d4f93d1ed9af327ee18b64fad.jpg", + "image_caption": [ + "Figure 8. Impact of the initial number of primitives $(M)$ . Increasing $M$ yields a finer-grained decomposition, while decreasing it produces a coarser decomposition." + ], + "image_footnote": [], + "bbox": [ + 91, + 210, + 202, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4fe97c9d7eee1036ca983cb6d1ff5a65d638c7a2d84c52711fd4a9c3ebf044ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 210, + 290, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9e5052eafdcbb60ada0535cf39aab79f0c65599843c8f28937ab275a24121cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 210, + 380, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37e420abf3cb308307587cda0307fd45dcd6cef15b6656db7d4300f0038b3a18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 210, + 473, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2fcb7e24646fc7aec39c0b4580b93c9fd57fc7ce80c30d2386654825a3e18cd5.jpg", + "image_caption": [ + "Figure 9. Applications. Part-Aware Editing (top): After optimization, we can easily edit the scene by adding, scaling, or moving specific parts. 3D Content Generation (bottom): By combining parts of different objects, we can create new 3D content." + ], + "image_footnote": [], + "bbox": [ + 93, + 345, + 483, + 498 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/39608ccb0875c5ec68a808e27c2807976861aa0c9c971369efe22f9aae7c3dba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
λpar=0.16.4616.100.7210.2551.0834.300.9870.0182.7
λpar=0.01 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
λpar=0.0014.0120.180.8350.1801.0335.040.9880.0157.1
M=44.9719.060.7940.2131.0634.930.9870.0165.0
M=8 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
M=163.9920.870.830.1761.0735.050.9880.0158.3
", + "bbox": [ + 93, + 579, + 480, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Effect of parsimony weight $(\\lambda_{\\mathrm{par}})$ and initial primitives count $(M)$ on DTU [23]. We report Chamfer distance, rendering metrics, and primitives count (#P). * denotes the default setting.", + "bbox": [ + 89, + 678, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "distilling 2D information, achieving semantic disentanglement but remains challenging due to indistinct textures and severe cross-view 2D inconsistencies. Lifting 2D segmentation features to 3D segmentation will also lead to feature misalignment, leading to visual artifacts. In contrast, our method performs direct 3D space segmentation, yielding more reasonable and structurally coherent decompositions.", + "bbox": [ + 89, + 729, + 483, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 89, + 847, + 202, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we first analyze the design choices by isolating each strategy to assess its impact. We report the averaged", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "quantitative performance on 15 instances of the chair category in Tab. 5 and provide visual comparisons in Fig. 5. Removing overlapping loss provides better reconstruction accuracy but more overlapping parts. The opacity loss drastically affects the number of primitives and doubles the CD, highlighting its role in controlling primitive presence and reconstruction quality. Coverage loss is crucial in reconstruction accuracy, ensuring primitives align correctly with the target. The adaptive primitive strategy fills in missing parts and improves the reconstruction metric. Without parsimony loss, CD and rendering metrics are optimal but result in over-decomposition. Note that the full model did not yield the highest accuracy because the part-aware reconstruction trades off the accuracy and rationality of the part decomposition. The objective is to achieve high accuracy while preserving reasonable parts.", + "bbox": [ + 511, + 90, + 906, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also validate the effectiveness of Gaussian scale regularization in Fig. 7. 2DGS [21] tends to use large-scale Gaussians in texture-less areas and create holes in the constructed mesh by TSDF integration [67]. The presence of holes stems from the inability of large-scale Gaussians to maintain view-consistent depth. Scale regularization effectively addresses this issue by suppressing large-scale Gaussians.", + "bbox": [ + 511, + 332, + 908, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, in Tab. 6, we analyze the influence of two key hyperparameters on decomposition granularity: the parsimony loss weight $\\lambda_{\\mathrm{par}}$ and the initial primitive count $M$ . Stronger parsimony regularization reduces the number of primitives, while weaker regularization increases it. As $M$ rises, both reconstruction and view synthesis performance slightly improve. This demonstrates that adjusting $M$ and $\\lambda_{\\mathrm{par}}$ enables our method to effectively control the granularity of object or scene decomposition. Fig. 8 visually illustrates this impact.", + "bbox": [ + 511, + 438, + 908, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Applications", + "text_level": 1, + "bbox": [ + 511, + 583, + 648, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 9 illustrates two applications of our method that the original 3DGS and NeRF-based approaches do not support. Firstly, after optimization, we obtain the part decomposition, facilitating easy editing of specific objects or scene components, e.g., adding, moving, removing, or scaling. Secondly, by combining parts of different objects, our method enables the creation of new high-quality 3D content.", + "bbox": [ + 511, + 604, + 908, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 724, + 640, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce PartGS, a hybrid representation of superquadrics and 2D Gaussians, to learn 3D scenes in a part-aware representation. Compared to prior works, the proposed method retains geometry details, supporting high-quality image rendering. It obtains state-of-the-art performance in comprehensive evaluations. One limitation is that it takes background-free scene images as inputs, leveraging segmentation tools such as SAM [27]. In the future, we aim to explore how to model backgrounds and extend to larger and more complex scenes.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9656", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104).", + "bbox": [ + 89, + 113, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18013-18023, 2023. 2", + "[2] Alan H. Barr. Superquadrics and Angle-Preserving Transformations. IEEE Computer Graphics and Applications, 1981. 1, 2, 3, 4", + "[3] Thomas Binford. Visual Perception by Computer. In IEEE Conference on Systems and Control, 1971. 2", + "[4] Åke Björck. Numerics of gram-schmidt orthogonalization. Linear Algebra and Its Applications, 197:297-316, 1994. 4", + "[5] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 7", + "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012 [cs.CV], 2015. 5, 6, 7", + "[7] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2", + "[8] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[9] Zhiqin Chen, Qimin Chen, Hang Zhou, and Hao Zhang. Dae-net: Deforming auto-encoder for fine-grained shape co-segmentation. arXiv preprint arXiv:2311.13125, 2023. 2", + "[10] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[11] Martin A. Fischler and Robert C. Bolles. Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography. Communications of the ACM, 1981. 2", + "[12] Huachen Gao, Shihe Shen, Zhe Zhang, Kaiqiang Xiong, Rui Peng, Zhirui Gao, Qi Wang, Yugui Xie, and Ronggang Wang. Fdc-nerf: learning pose-free neural radiance fields with flow-depth consistency. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3615-3619. IEEE, 2024. 1", + "[13] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network" + ], + "bbox": [ + 93, + 234, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2", + "[14] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based gaussian splatt-ting for real-time large-scale deformation. arXiv preprint arXiv:2402.04796, 2024. 2, 3", + "[15] Zhirui Gao, Renjiao Yi, Zheng Qin, Yunfan Ye, Chenyang Zhu, and Kai Xu. Learning accurate template matching with differentiable coarse-to-fine correspondence refinement. Computational Visual Media, 10(2):309-330, 2024. 2", + "[16] Zhirui Gao, Renjiao Yi, Yaqiao Dai, Xuening Zhu, Wei Chen, Chenyang Zhu, and Kai Xu. Curve-aware gaussian splatting for 3d parametric curve reconstruction, 2025. 2", + "[17] Zhirui Gao, Renjiao Yi, Chenyang Zhu, Ke Zhuang, Wei Chen, and Kai Xu. Generic objects as pose probes for few-shot view synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 1", + "[18] Zeqi Gu, Yin Cui, Zhaoshuo Li, Fangyin Wei, Yunhao Ge, Jinwei Gu, Ming-Yu Liu, Abe Davis, and Yifan Ding. Artiscene: Language-driven artistic 3d scene generation through image intermediary. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2891-2901, 2025. 1", + "[19] Yanran Guan, Han Liu, Kun Liu, Kangxue Yin, Ruizhen Hu, Oliver van Kaick, Yan Zhang, Ersin Yumer, Nathan Carr, Radomir Mech, et al. Fame: 3d shape generation via functionality-aware model evolution. IEEE Transactions on Visualization and Computer Graphics, 28(4):1758-1772, 2020. 2", + "[20] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3, 4", + "[21] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. SIGGRAPH, 2024. 2, 3, 5, 6, 7, 8", + "[22] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18572-18582, 2022. 2", + "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanaes. Large Scale Multi-view Stereopsis Evaluation. In CVPR, 2014. 5, 6, 8", + "[24] Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, and Na Zhao. Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaussians. arXiv preprint arXiv:2410.01535, 2024. 2", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 4", + "[26] Chung Min* Kim, Mingxuan* Wu, Justin* Kerr, Matthew Tancik, Ken Goldberg, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9657", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 7, 8", + "[28] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 1", + "[29] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised Fitting of Geometric Primitives to 3D Point Clouds. In CVPR, 2019. 1", + "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 6, 7", + "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023. 2", + "[32] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: a Probabilistic Approach. In CVPR, 2022. 1, 2, 3, 6, 7", + "[33] Romain Loiseau, Elliot Vincent, Mathieu Aubry, and Loic Landrieu. Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans. arXiv:2304.09704 [cs.CV], 2023. 1", + "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 4", + "[35] Niloy Mitra, Michael Wand, Hao (Richard) Zhang, Daniel Cohen-Or, Vladimir Kim, and Qi-Xing Huang. Structure-aware shape processing. In SIGGRAPH Asia 2013 Courses, 2013. 1", + "[36] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives. In NeurIPS, 2023. 1, 2, 3, 4, 6, 7", + "[37] Chengjie Niu, Jun Li, and Kai Xu. Im2struct: Recovering 3d shape structure from a single rgb image. Cornell University - arXiv, Cornell University - arXiv, 2018. 1", + "[38] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics Revisited: Learning 3D Shape Parsing Beyond Cuboids. In CVPR, 2019. 2", + "[39] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning Unsupervised Hierarchical Part Decomposition of 3D Objects from a Single RGB Image. In CVPR, 2020. 2", + "[40] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In ECCV, 2022. 2, 3, 6, 7", + "[42] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2", + "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5", + "[44] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems (TODS), 42(3):1-21, 2017. 5", + "[45] Qingyao Shuai, Chi Zhang, Kaizhi Yang, and Xuejin Chen. Dpf-net: combining explicit shape priors in deformable primitive field for unsupervised structural reconstruction of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14321-14329, 2023. 2", + "[46] Chunyu Sun, Yuqi Yang, Haoxiang Guo, Pengshuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media, 2022. 2", + "[47] Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yannis Avrithis, and Leonidas Guibas. PartNeRF: Generating Part-Aware Editable 3D Shapes without 3D Supervision. In CVPR, 2023. 1, 2, 6, 7", + "[48] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. In CVPR, 2017. 2", + "[49] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2, 3, 4", + "[50] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 2", + "[51] Xinhang Wan, Jiyuan Liu, Hao Yu, Qian Qu, Ao Li, Xinwang Liu, Ke Liang, Zhibin Dong, and En Zhu. Contrastive continual multiview clustering with filtered structural fusion. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2", + "[52] Fengxiang Wang, Mingshuo Chen, Yueying Li, Di Wang, Haotian Wang, Zonghao Guo, Zefan Wang, Boqi Shan, Long Lan, Yulin Wang, et al. Geollava-8k: Scaling remote-sensing multimodal large language models to 8k resolution. arXiv preprint arXiv:2505.21375, 2025. 1", + "[53] Fengxiang Wang, Hongzhen Wang, Zonghao Guo, Di Wang, Yulin Wang, Mingshuo Chen, Qiang Ma, Long Lan, Wenjing Yang, Jing Zhang, et al. Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14325-14336, 2025. 1", + "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9658", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1, 6", + "[55] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliafaschi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 1", + "[56] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Vastsd: Learning 3d vascular tree-state space diffusion model for angiography synthesis. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 15693-15702, 2025. 1", + "[57] Tong Wu, Lin Gao, Ling-Xiao Zhang, Yu-Kun Lai, and Hao Zhang. Star-tm: Structure aware reconstruction of textured mesh from single image. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2021. 2", + "[58] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S. Chirikjian. Primitive-based Shape Abstraction via Nonparametric Bayesian Inference. In ECCV, 2022. 2", + "[59] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2054-2065, 2022. 2", + "[60] Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. Cluster-guided contrastive graph clustering network. In Proceedings of the AAAI conference on artificial intelligence, pages 10834-10842, 2023. 2", + "[61] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks. In CVPR, 2020. 5, 7", + "[62] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 7", + "[63] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. arXiv preprint arXiv:2311.11666, 2023. 7", + "[64] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9491-9500, 2019. 3", + "[65] Fenggen Yu, Yimin Qian, Xu Zhang, Francisca Gil-Ureta, Brian Jackson, Eric Bennett, and Hao Zhang. Dpa-net: Structured 3d abstraction from sparse views via differentiable primitive assembly. arXiv preprint arXiv:2404.00875, 2024. 1, 2", + "[66] Hongyi Zhou, Xiaogang Wang, Yulan Guo, and Kai Xu. Monomobility: Zero-shot 3d mobility analysis from monocular videos, 2025. 2", + "[67] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.8", + "[68] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "shape co-segmentation with group consistency loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2020. 1", + "bbox": [ + 545, + 90, + 906, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9659", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_model.json b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_model.json new file mode 100644 index 0000000000000000000000000000000000000000..29142beb459799eddff3d87d4fa16c2a8e244c67 --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_model.json @@ -0,0 +1,2697 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.159, + 0.131, + 0.84, + 0.177 + ], + "angle": 0, + "content": "Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.203, + 0.838, + 0.258 + ], + "angle": 0, + "content": "Zhirui Gao Renjiao Yi† Yuhang Huang Wei Chen Chenyang Zhu Kai Xu† National University of Defense Technology zhirui-gao.github.io/PartGS" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Low-level 3D representations, such as point clouds, meshes, NeRFs and 3D Gaussians, are commonly used for modeling 3D objects and scenes. However, cognitive studies indicate that human perception operates at higher levels and interprets 3D environments by decomposing them into meaningful structural parts, rather than low-level elements like points or voxels. Structured geometric decomposition enhances scene interpretability and facilitates downstream tasks requiring component-level manipulation. In this work, we introduce PartGS, a self-supervised part-aware reconstruction framework that integrates 2D Gaussians and superquadrics to parse objects and scenes into an interpretable decomposition, leveraging multi-view image inputs to uncover 3D structural information. Our method jointly optimizes superquadric meshes and Gaussians by coupling their parameters within a hybrid representation. On one hand, superquadrics enable the representation of a wide range of shape primitives, facilitating flexible and meaningful decompositions. On the other hand, 2D Gaussians capture detailed texture and geometric details, ensuring high-fidelity appearance and geometry reconstruction. Operating in a self-supervised manner, our approach demonstrates superior performance compared to state-of-the-art methods across extensive experiments on the DTU, ShapeNet, and real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.485, + 0.877 + ], + "angle": 0, + "content": "3D reconstruction from multi-view images is a long-standing challenge in 3D vision and graphics [12, 17, 54]. Most reconstructed scenes are in low-level representations such as point clouds, voxels, or meshes. However, humans tend to understand 3D scenes as reasonable parts [35]. For instance, when observing a scene, we naturally construct high-level structural information, such as scene graphs, instead of focusing on low-level details like point clouds or voxels. Motivated by it, we propose a part-aware reconstruction framework" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.296, + 0.603, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.374, + 0.602, + 0.384 + ], + "angle": 0, + "content": "3D decomposition" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.297, + 0.712, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.374, + 0.689, + 0.383 + ], + "angle": 0, + "content": "Block-level" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.297, + 0.896, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.374, + 0.785, + 0.382 + ], + "angle": 0, + "content": "Point-level" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.385, + 0.605, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.45, + 0.596, + 0.459 + ], + "angle": 0, + "content": "Prior work (EMS)" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.384, + 0.707, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.38, + 0.805, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.45, + 0.762, + 0.459 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.389, + 0.896, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.473, + 0.908, + 0.529 + ], + "angle": 0, + "content": "Figure 1. Prior works [32, 36] model the scene using primitives, while the proposed method can further model precise geometry details and textures. Here, EMS [32] takes point cloud inputs and reconstructs non-textured primitives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.597 + ], + "angle": 0, + "content": "that decomposes objects or scenes into meaningful shapes or parts, facilitating tasks such as physical simulation, editing, content generation and understanding [18, 28, 52, 53, 56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Several prior works [29, 32, 33, 37, 55, 68] have explored part-aware reconstruction or 3D decomposition. However, most of them rely heavily on 3D supervision and often struggle to retain fine-grained geometric details, limiting their practicality in real-world scenarios. Recent advances [47, 65] have focused on extending the neural radiance field framework for part-aware reconstruction, building upon its remarkable success in 3D reconstruction from multi-view images. For example, PartNeRF [47] models objects using multiple neural radiance fields. Yet, the intricate composition of implicit fields complicates the learning process, leading to suboptimal rendering quality and inefficient decomposition. Recently, DBW [36] proposes a novel approach that decomposes scenes into block-based representations using superquadric primitives [2], optimizing both their parameters and UV texture maps through rendering loss minimization. While this method demonstrates effective scene decomposition into coherent components, it faces challenges in achieving accurate part-aware reconstruction of both geometry and appearance, as illustrated in Fig. 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.248, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.439 + ], + "angle": 0, + "content": "To address these limitations, we propose PartGS, a self-supervised part-aware hybrid representation that integrates 2D Gaussians [21] and superquadrics [2], to achieve both high-quality texture reconstruction and geometrically accurate part decomposition. There are previous approaches [14, 20, 49] combining mesh reconstruction and Gaussian splatting, where they first reconstruct the mesh and then attach Gaussians to it. In contrast to their primary focus on acquiring the representations, the proposed approach is to achieve part-aware decomposition reconstruction. Our method establishes a coupled optimization framework that simultaneously learns superquadric meshes and Gaussians through parameter sharing, with each part constrained to a single superquadric. The differentiable rendering of Gaussians drives this hybrid representation, leveraging the inherent convexity of superquadrics for qualitative 3D decomposition. The self-supervised training of part-aware reconstruction is built upon the assumption that each part of most objects should be a basic shape and can be represented by a superquadric. Meanwhile, Gaussian splatting, renowned for its superior rendering quality and efficient training, is incorporated to capture intricate texture details, speeding up reconstruction and improving rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.485, + 0.608 + ], + "angle": 0, + "content": "In the hybrid representation, 2D Gaussians are distributed on superquadric surfaces to form structured blocks. The pose and shape of Gaussians within each block are determined by their corresponding superquadrics rather than being optimized independently. The parameter space encompasses global controls for superquadric properties (shape, pose, and opacity) and local spherical harmonic coefficients for individual Gaussians. Compared to standard Gaussian splating [21, 25], which populates the occupancy volume with independently parameterized Gaussians, the coupled representation is more compact and efficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.486, + 0.852 + ], + "angle": 0, + "content": "During training, the parameters are optimized through rendering loss without additional supervision. However, we notice that image rendering loss often leads to local minima in superquadric shape optimization. To tackle this, we introduce several regularizers to maintain global consistency between the 3D representation and the input 2D information. This strategic implementation allows us to segment parts fully self-supervised, achieving block-level reconstruction. Finally, to better capture irregular shapes, we implement a point-level refinement step that frees 2D Gaussians to deviate from the surface, thereby enhancing geometric fidelity. Extensive experiments show that PartGS, at block-level and point-level, achieves \\(33.3\\%\\) and \\(75.9\\%\\) improvements in reconstruction accuracy, 3.18 and 16.13 increases in PSNR, and 4X and 3X speedups compared to the state-of-the-art baseline. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "- We introduce a novel hybrid representation for part-aware 3D reconstruction, combining the strengths of superquadrics and Gaussian splatting to achieve reasonable" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.815, + 0.107 + ], + "angle": 0, + "content": "decomposition and high-quality rendering." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.108, + 0.907, + 0.152 + ], + "angle": 0, + "content": "- We propose several novel regularizers to enforce consistency between 3D decomposition and 2D observations, enabling self-supervised part decomposition." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.153, + 0.907, + 0.213 + ], + "angle": 0, + "content": "- Compared to prior works, the method takes one step forward, by reconstructing both the block-level and point-level part-aware reconstructions, preserving both part segmentation and reconstruction precision." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.108, + 0.907, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.229, + 0.656, + 0.244 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.254, + 0.845, + 0.27 + ], + "angle": 0, + "content": "2.1. Shape Decomposition and Abstraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.908, + 0.518 + ], + "angle": 0, + "content": "Structured shape representation learning decomposes objects or scenes into coherent geometric primitives, facilitating shape understanding and generation [9, 15, 19, 46, 57]. Early approaches like Blocks World [42] and Generalized Cylinders [3] emphasized compact representations. Modern methods typically process 3D inputs (point clouds, meshes, voxels) by decomposing them into primitive ensembles, including cuboids [41, 48], superquadrics [32, 38, 39, 58], and convex shapes [8, 10]. For instance, MonteBoxFinder [41] integrates clustering [11, 50, 60], cuboid representations, and Monte Carlo Tree Search, while EMS [32] adopts a probabilistic approach for superquadric recovery. However, these methods are limited to coarse shape representations. Recent advances in shape abstraction [13, 22, 31, 40, 45] enable more detailed shape representations through flexible primitive deformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Some studies [1, 16, 36, 47, 65, 66] attempt to create structure-aware 3D representations directly from images. PartNeRF [47] introduces ellipsoid representations within NeRFs, its reliance on multiple implicit fields results in inefficient 3D decomposition. ISCO [1] and DBW [36] use 3D superquadrics for shape decomposition, which enables more meaningful structural separation. However, their simple shape parameters struggle to capture complex geometries, leading to poor geometry and appearance reconstruction. DPA-Net [65] has advanced 3D shape abstraction from sparse views but generates redundant parts and struggles with realistic texture rendering. A concurrent work, GaussianBlock [24], employs SAM [27] to guide superquadric splitting and fusion for 3D decomposition, yet its computational efficiency remains limited, typically requiring several hours per scene. In contrast, our approach accomplishes self-supervised [51, 59] part-aware scene reconstruction through an efficient hybrid representation that simultaneously maintains geometric fidelity and photorealistic rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.794, + 0.835 + ], + "angle": 0, + "content": "2.2. Mesh-based Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Gaussian splitting (GS) [25] has been rapidly adopted in multiple fields due to its remarkable rendering capability. Several studies [7, 14, 20, 49] attempt to align Gaussians with mesh surfaces for easier editing and animation. Among" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.885, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.351, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. PartGS takes multi-view images to learn a parametric hybrid representation of superquadrics and 2D Gaussians. It initializes from random superquadrics and is gradually optimized during training to obtain a block-level reconstruction. Then, we free the constraints of Gaussians to model detailed geometry, achieving point-level reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.413, + 0.485, + 0.685 + ], + "angle": 0, + "content": "them, SuGaR [20] uses flat 3D Gaussians to enforce the alignment with the scene surface during optimization, minimizing the difference between the signed distance functions (SDF) of the desired Gaussians and actual Gaussians. GaMeS [49] introduces a hybrid representation of Gaussians and mesh, where Gaussians are attached to triangular facets of the mesh. Similarly, Gao et al. [14] proposes a mesh-based GS to achieve large-scale deformation effects on objects. Recently, 2DGS [21] proposes 2D Gaussians for surface modeling and significantly enhances the geometric quality. Inspired by them, we propose a representation that combines 2D Gaussians with superquadrics. A key distinction between our method and previous mesh-based GS approaches is that these methods require first reconstructing the scene's mesh and then binding Gaussians to the mesh surface, which results in a non-continuous mesh. In contrast, our method directly optimizes the mesh through a rendering loss, enabling part-aware mesh reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.182, + 0.722 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Given a set of calibrated multi-view images and foreground masks, we aim to learn part-aware 3D representations, a meaningful decomposition of both geometry and appearance. Our approach adopts a two-stage optimization strategy: first, decomposing the object into basic shapes using a mixture of Gaussians and superquadrics at the block-level, followed by refining the decomposition at the point-level to achieve precise geometry. In Sec 3.1, we parameterize the hybrid representation. Sec. 3.2 then elaborates on leveraging this representation, enhanced by novel regularizers and an adaptive control strategy, to achieve self-supervised block-level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.444 + ], + "angle": 0, + "content": "decomposition. Finally, Sec. 3.3 presents the process for obtaining detailed part-aware results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.871, + 0.467 + ], + "angle": 0, + "content": "3.1. Parametrizing the Hybrid Representation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.472, + 0.908, + 0.608 + ], + "angle": 0, + "content": "As shown in Fig. 2, to leverage the strengths of both, we attach Gaussians to the surface of superquadric meshes. This representation retains the superquadric's ability to parse a 3D scene into distinct parts. Meanwhile, spherical harmonics of Gaussians enable complex texture rendering via Gaussian splatting, addressing the texture learning limitations of prior work [32, 36, 41, 64]. Sharing pose parameters between superquadrics and 2D Gaussians further improves the representation's efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.774 + ], + "angle": 0, + "content": "The parametric representation is controlled by both primitive and Gaussian parameters, which are optimized simultaneously. Given a 3D scene \\(S\\), the proposed method decomposes it into multiple hybrid blocks, each consisting of a superquadric with associated Gaussians. Each scene is denoted by the hybrid representation: \\(S = \\mathcal{B}_1 \\cup \\ldots \\cup \\mathcal{B}_i \\cup \\mathcal{B}_M\\), where \\(\\mathcal{B}_i\\) is the \\(i\\)-th hybrid block, and \\(M\\) is the total number of blocks. Blocks are defined using manually designed parameters that control pose, opacity, scale, shape, and texture. These parameters are optimized via differentiable rendering to parse the target scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.909, + 0.866 + ], + "angle": 0, + "content": "Shape and Scale Parameters. For each hybrid block \\(\\mathcal{B}_i\\), its geometry is controlled by superquadric parameters [2]. Specifically, there are two shape parameters \\(\\epsilon_1, \\epsilon_2\\) that define its shape, along with three scale parameters \\(s_1, s_2, s_3\\), which scale the 3D axes. Analogous to icosphere vertex positioning, superquadric vertex coordinates are computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.868, + 0.907, + 0.883 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\left[ s _ {1} \\cos^ {\\epsilon 1} (\\theta) \\cos^ {\\epsilon 2} (\\varphi); s _ {2} \\sin^ {\\epsilon 1} (\\theta); s _ {3} \\cos^ {\\epsilon 1} (\\theta) \\sin^ {\\epsilon 2} (\\varphi) \\right], \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta\\) and \\(\\varphi\\) represent the azimuth and elevation defined" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9651" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "in the spherical coordinate. The shape and scale parameters govern block deformation to learn part-aware geometry." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Pose Parameters. The pose of the \\(i\\)-th hybrid block is defined by its rotation \\(\\mathbf{R}_i\\) and translation \\(\\mathbf{t}_i\\). The vertex position from Eq. 1 is transformed from the local coordinate to world space as: \\(\\hat{\\mathbf{v}}_i^j = \\mathbf{R}_i\\mathbf{v}_i^j + \\mathbf{t}_i\\), where \\(j\\) indexes the vertices. Previous approaches, including SuGaR [20] and GaMeS [49], position Gaussians on reconstructed meshes with independent pose and shape parameters to enhance appearance modeling. In contrast, our method employs Gaussians to construct a differentiable rendering that bridges images and superquadrics. To achieve this, we couple the parameters of Gaussians with superquadrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.439 + ], + "angle": 0, + "content": "For a posed superquadric, Gaussian centers are uniformly sampled on triangular faces, with their poses determined by face vertices. Following GaMeS [49], each Gaussian's rotation matrix \\(\\mathrm{R}_v = [r_1,r_2,r_3]\\) and scaling \\(\\mathrm{S}_v\\) are computed from vertex positions. Given a triangular face \\(V\\) with vertices \\(v_{1},v_{2},v_{3}\\in \\mathbb{R}^{3}\\), orthonormal vectors are constructed such that \\(r_1\\) aligns with the face normal, \\(r_2\\) points from the centroid \\(m = \\mathrm{mean}(v_1,v_2,v_3)\\) to \\(v_{1}\\). \\(r_3\\) is obtained by orthonormalizing the vector from the center to the second vertex concerning the existing \\(r_1\\) and \\(r_2\\):" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.45, + 0.484, + 0.482 + ], + "angle": 0, + "content": "\\[\nr _ {3} = \\frac {\\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2}\\right)}{\\left\\| \\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2} \\right. \\right\\|}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.517 + ], + "angle": 0, + "content": "where \\( \\text{ort} \\) represents the orthonormalization in Gram-Schmidt [4]. For the scaling of 2D Gaussians, we use:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.525, + 0.483, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} _ {v} = \\operatorname {d i a g} \\left(s _ {v} ^ {2}, s _ {v} ^ {3}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.628 + ], + "angle": 0, + "content": "where \\( s_v^2 = c||m - v_1|| \\), \\( s_v^3 = c < v_2, r_3 > \\), and \\( c \\) is a size control hyperparameter. We place a fixed number of Gaussians on each triangular face. These designs eliminate the need to learn the geometry of Gaussians, thereby enhancing the efficiency of the representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Opacity Parameters. We define the total number of hybrid blocks as \\( M \\). However, a typical scene does not contain exactly \\( M \\) blocks. Therefore, only the meaningful blocks are retained. To achieve this, we introduce a learnable parameter \\( \\tau_{i} \\) to represent each block's opacity. During the optimization, only blocks with \\( \\tau_{i} \\) greater than a certain threshold are retained. Note that Gaussians within the same block share the same \\( \\tau \\) for opacity in rasterization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.811 + ], + "angle": 0, + "content": "Texture Parameters. The texture is modeled using 2D Gaussians positioned on the surface of the superquadrics. Spherical harmonics of each Gaussian control the texture and are optimized for rendering view-dependent images [25]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.33, + 0.835 + ], + "angle": 0, + "content": "3.2. Block-level Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This section describes the optimization of the hybrid representation. We observed that minimizing the rendering loss across multi-view images alone led to instability in positioning hybrid blocks. Therefore, several regularization terms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "are introduced to optimize the composition for maximal image formation consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.129, + 0.652, + 0.144 + ], + "angle": 0, + "content": "3.2.1. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.148, + 0.907, + 0.253 + ], + "angle": 0, + "content": "In addition to the image rendering loss, we employ regularizers consisting of coverage, overlap, parsimony, and opacity entropy. The coverage loss encourages hybrid blocks to cover only meaningful regions; the overlap loss prevents blocks from overlapping; the parsimony loss regularizes the number of existing blocks; and the opacity entropy encourages block opacities close to binary." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Rendering Loss. The rendering loss is based on 3DGS [25], where it combines an \\( L_{1} \\) term with a D-SSIM term:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.298, + 0.907, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n}} = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.321, + 0.907, + 0.518 + ], + "angle": 0, + "content": "**Coverage Loss.** The coverage loss ensures that the block set \\(\\{\\mathcal{B}_1 \\ldots \\mathcal{B}_M\\}\\) covers the object while preventing it from extending beyond its boundaries. To determine the 3D occupancy field based on the blocks, we first define the approximate signed distance of a 3D point \\(x\\) to the \\(i\\)-th block, \\(D_i(x) = \\Psi_i(x) - 1\\). Here, \\(\\Psi_i\\) denotes the superquadric inside-outside function [2] associated with the block \\(i\\). Consequently, \\(D_i(x) \\leq 0\\) if the point \\(x\\) lies inside the \\(i\\)-th block, and \\(D_i(x) > 0\\) if \\(x\\) is outside the block. Inspired by NeRF [34], we sample a ray set \\(\\mathcal{R}\\) using ray-casting based on camera poses. Given the object mask, we associate each ray \\(r \\in \\mathcal{R}\\) with a binary label: \\(l_r = 1\\) if the ray \\(r\\) is inside mask, otherwise \\(l_r = 0\\). The coverage loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.527, + 0.907, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {R}) = \\sum_ {r \\in \\mathcal {R}} l _ {r} L _ {\\text {c r o s s}} (r) + \\left(1 - l _ {r}\\right) L _ {\\text {n o n - c r o s s}} (r). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.906, + 0.6 + ], + "angle": 0, + "content": "Here, \\(L_{\\mathrm{cross}}\\) encourages rays to intersect with blocks, while \\(L_{\\mathrm{non - cross}}\\) discourages rays from intersecting with blocks:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.61, + 0.907, + 0.64 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c r o s s}} (r) = \\operatorname {R e L U} \\left(\\min _ {i \\in \\mathcal {M}} \\min _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.646, + 0.907, + 0.676 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o n - c r o s s}} (r) = \\operatorname {R e L U} \\left(\\max _ {i \\in \\mathcal {M}} \\max _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} - D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.757 + ], + "angle": 0, + "content": "where \\( x_{j}^{r} \\) means the \\( j \\)-th sampled point along the ray \\( r \\) and \\( \\mathcal{M} = \\{1, \\dots, M\\} \\). Intuitively, this implies that at least one sampled points \\( \\mathcal{X}_{r} \\) along the ray \\( r \\) inside the mask that lies within a certain block, while all points along the ray outside the mask should not belong to any block." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.863 + ], + "angle": 0, + "content": "Overlap Loss. We introduce a regularization term to prevent overlap between individual blocks. Given the difficulty of directly calculating block overlap, we adopt a Monte Carlo method similar to [36]. Specifically, we sample multiple 3D points in space and penalize those that lie inside more than \\( k \\) blocks. Based on the superquadric inside-outside function, a soft occupancy function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.873, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {i} ^ {x} = \\tau_ {i} \\operatorname {s i g m o i d} \\left(\\frac {- D _ {i} (x)}{\\gamma}\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a temperature parameter. Thus, for a set of \\(N\\) sampled 3D points \\(\\Omega\\), the overlap loss is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.132, + 0.484, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o v e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\operatorname {R e L U} \\left(\\sum_ {i \\in \\mathcal {M}} \\mathcal {O} _ {i} ^ {x} - k\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.485, + 0.245 + ], + "angle": 0, + "content": "Parsimony Loss. To promote the use of the minimal number of blocks and achieve parsimony in decomposition, we introduce a regularization term that penalizes block opacity \\((\\tau)\\). This loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.255, + 0.484, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p a r}} = \\frac {1}{M} \\sum_ {i \\in \\mathcal {M}} \\sqrt {\\tau_ {i}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.406 + ], + "angle": 0, + "content": "Opacity Entropy Loss. During optimization, the opacity of the block inside the object region tends to approach 1, while the opacity of the block outside the object region tends to approach 0. To facilitate this, we associate the opacities of blocks with the labeled rays and apply a cross-entropy loss \\( L_{ce} \\) between the block opacity and the mask labels, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.416, + 0.484, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o p a}} (\\mathcal {R}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} L _ {c e} \\left(\\max _ {i \\in \\mathcal {M}} \\tau_ {i} \\left(x ^ {r}\\right), l _ {r}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.464, + 0.484, + 0.494 + ], + "angle": 0, + "content": "Here, only points \\( x^r \\) inside the blocks are sampled, e.g., \\( \\{x^r \\in \\mathcal{X}_r \\mid D_i(x^r) \\leq 0\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.494, + 0.484, + 0.523 + ], + "angle": 0, + "content": "The total loss is the weighted sum of the loss terms described above:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.536, + 0.483, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e n}} + \\lambda_ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + \\lambda_ {\\text {o v e r}} \\mathcal {L} _ {\\text {o v e r}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {o p a}} \\mathcal {L} _ {\\text {o p a}}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.334, + 0.579 + ], + "angle": 0, + "content": "3.2.2. Adaptive Number of Blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Given that the number of parts may vary across different scenes, we allow the number of blocks to adjust during optimization adaptively. Specifically, when the opacity of an existing block falls below a threshold \\( t \\), this block is removed immediately. Furthermore, we explore a block-adding mechanism to dynamically incorporate new components, ensuring comprehensive coverage of target objects. Specifically, DBSCAN [44] is employed to cluster initial point clouds that are not covered by any blocks, where the point clouds are either randomly initialized or derived as by-products from COLMAP [43]. For each point cloud cluster, a new block is introduced at its center, with the remaining parameters of the block initialized randomly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.84 + ], + "angle": 0, + "content": "The final block-level reconstruction generates a textured 3D scene represented by multiple superquadrics, with surface details rendered through Gaussians, as illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.327, + 0.866 + ], + "angle": 0, + "content": "3.3. Point-level Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The primitive-based hybrid representation effectively decomposes the shape into parts but exhibits suboptimally for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "complex objects. To address this limitation, we perform a refinement stage that enhances geometric fitting. In this stage, the constraint between Gaussians and superquadrics is decoupled, enabling independent Gaussian optimization. Furthermore, to prevent Gaussian components from one block passing through adjacent blocks and disrupting the continuity and plausibility of the part decomposition, we impose a new constraint to minimize the negative signed distance of each Gaussian entering other blocks:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.233, + 0.907, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e n t e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\sum_ {m \\in \\mathcal {M} \\backslash \\{\\delta \\}} \\operatorname {R e L U} (- D _ {m} (x)), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.906, + 0.338 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is the sampled Gaussian set, and \\(\\delta\\) represents the identifier of the block to which the Gaussian \\(x\\) belongs. As seen in Fig. 2, the reconstruction becomes more accurate and aligns better with the target shape." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.646, + 0.366 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.908, + 0.435 + ], + "angle": 0, + "content": "We comprehensively evaluate our approach across four key aspects: 3D reconstruction quality, view synthesis performance, shape parsimony, and computational efficiency, comparing it with state-of-the-art methods in the field." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.442, + 0.616, + 0.457 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.646 + ], + "angle": 0, + "content": "Extensive experiments are conducted on two widely-used public datasets: DTU [23] and ShapeNet [6]. Specifically, 15 standard scenes from the DTU dataset that are widely adopted for reconstruction quality assessment are used for evaluation. Each scene contains either 49 or 64 images. Additionally, we construct a ShapeNet subset consisting of four categories: Chair, Table, Gun, and Airplane, with 15 objects selected per category. For each object, 50 rendered images are generated for training. We further validate real-world applicability on both the BlendedMVS dataset [61] and self-captured scenes, demonstrating robust performance across synthetic and natural environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.731, + 0.669 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We build upon the 2DGS [21] architecture and add a nonlearnable part attribute to each Gaussian component to enforce the constraint in Eq. 13. In the ShapeNet benchmark, we observe a common degradation in mesh reconstruction as shown in Fig. 7. To address this, we propose Gaussian scale regularization, as validated in Sec. 4.4. We set the initial number of primitives \\(M\\) to 8. The temperature parameter \\(\\gamma\\) in Eq. 8 is 0.005 and the overlapping number \\(k\\) in Eq. 9 is 1.95. We train the hybrid representation for 30k iterations, followed by refinement optimization for another 30k iterations. During block-level optimization, the block-adding operation is carried out at the 5k-th and 10k-th iterations. The loss weights \\(\\lambda_{\\mathrm{cov}}\\), \\(\\lambda_{\\mathrm{over}}\\), \\(\\lambda_{\\mathrm{par}}\\), and \\(\\lambda_{\\mathrm{opa}}\\) are set to 10, 1, 0.002, and 0.01, respectively. For more details about datasets, metrics, and baselines, please refer to Supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9653" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.411, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on DTU [23] and ShapeNet [23]. The first two rows are DTU examples, and the last two are ShapeNet examples, respectively. Our method is the only one that provides reasonable 3D part decomposition while capturing detailed geometry." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.457, + 0.905, + 0.579 + ], + "angle": 0, + "content": "
MethodInputRenderableChamfer distance per sceneMeanMean
S24S37S40S55S63S65S69S83S97S105S106S110S114S118S122CD#P
EMS [32]3D GTX6.323.542.994.304.164.013.753.244.974.344.167.627.584.464.034.657.7
MBF [41]3D GTX3.122.663.842.541.592.112.192.012.322.452.172.123.832.022.552.5034.1
EMS [32] + Neus [54]ImageX5.995.564.434.325.426.143.753.964.634.345.885.114.294.833.534.978.87
MBF [41] + Neus [54]ImageX2.693.373.222.693.632.602.593.132.852.512.453.722.242.492.522.8546.7
PartNeRF [47]Image✓9.3810.469.088.636.047.257.229.158.7210.016.729.857.858.689.218.548.0
DBW [36]Image✓5.418.351.573.083.404.157.463.946.634.854.384.656.294.343.044.764.8
Ours (Block-level)Image✓5.684.911.852.613.754.663.757.574.274.383.494.483.614.213.704.195.9
Ours (Point-level)Image✓0.701.170.550.651.061.231.101.361.370.780.921.410.691.050.710.985.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.589, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on DTU [23]. The Chamfer distance between the 3D reconstruction and the ground-truth is reported in 15 scenes. The best results are bolded, and the average numbers of primitives found (#P) that are greater than 10 are underlined." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.633, + 0.481, + 0.711 + ], + "angle": 0, + "content": "
MethodPart-awareCD ↓PSNR ↑SSIM ↑LIPPS ↓Time ↓
Neuralangelo [30]X0.6133.84--> 10 h
2DGS [21]X0.8134.070.990.019~ 10 m
PartNeRF [47]✓9.5917.970.770.246~ 8 h
DBW [36]✓4.7316.440.750.201~ 2 h
Ours (Block-level)✓4.1919.840.820.189~ 30 m
Ours (Point-level)✓0.9835.040.990.015~ 40 m
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.721, + 0.486, + 0.778 + ], + "angle": 0, + "content": "Table 2. Quantitative results on DTU [23]. Our method outperforms all part-aware approaches in image synthesis quality, reconstruction accuracy, and efficiency. Neuralangelo's results are from the original paper, with all times measured on an RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.795, + 0.22, + 0.81 + ], + "angle": 0, + "content": "4.3. Evaluations" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.357, + 0.835 + ], + "angle": 0, + "content": "4.3.1. DTU and Shapenet Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Geometry Reconstruction. In Tab. 1 and Tab. 3, we compare our geometry reconstruction to SOTA shape decomposition methods on Chamfer distance and training time using DTU and Shapenet dataset. The Chamfer distance" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.633, + 0.906, + 0.721 + ], + "angle": 0, + "content": "
MethodInputChamfer Distance ↓Primitives (#P)MeanMean
AirplaneTableChairGunAirplaneTableChairGunCD#P
EMS [32]3D GT3.406.9219.02.029.47.8810.38.47.849.0
MBF [41]3D GT2.832.181.592.3210.8513.913.414.32.2113.1
PartNeRF [47]Image2.292.772.302.468.08.08.08.02.468.0
DBW [36]Image3.617.336.192.092.75.23.63.34.813.7
Ours (Block-level)Image2.472.152.321.783.96.67.65.02.185.8
Ours (Point-level)Image1.291.720.941.073.96.67.65.01.255.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.731, + 0.908, + 0.772 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison on ShapeNet [6]. We report Chamfer distance and the number of parts. The best results are bolded, and the second-best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "metrics reported for ShapeNet are scaled by a factor of 100 for readability. Our method consistently outperforms in all scenes compared to prior works. As shown in Fig. 3, our approach consistently produces interpretable 3D decompositions with further refinement achieving more detailed geometry (the last two columns). MBF [41] achieves a low CD error but at the cost of using significantly more prim" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9654" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.094, + 0.904, + 0.178 + ], + "angle": 0, + "content": "
MethodPart-awareChamfer Distance ↓PSNR ↑SSIM ↑LIPPS ↓
AirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGun
2DGS [21]✗1.472.370.501.0340.8939.8039.0541.740.9940.9900.9900.9940.0090.0270.0170.009
PartNeRF [47]✓2.292.772.302.4619.6320.6619.0821.970.8980.8550.8750.9160.0860.1610.1360.083
DBW [36]✓3.617.336.192.0926.1123.8420.2528.720.9500.9150.8920.9600.0740.1360.1320.042
Ours (Block-level)✓2.472.152.321.7827.9427.9824.9229.950.9590.9250.9060.9630.0720.1290.1220.047
Ours (Point-level)✓1.291.720.941.0741.1836.8036.0739.510.9920.9730.9770.9890.0140.0700.0380.021
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.187, + 0.9, + 0.202 + ], + "angle": 0, + "content": "Table 4. Quantitative results on ShapeNet [6]. We report the Chamfer distance and novel view synthesis results across four categories." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.219, + 0.232, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.219, + 0.312, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.219, + 0.388, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.219, + 0.471, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.409, + 0.486, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on BlendedMVS [61] and self-captured data. We present RGB renderings and decomposed parts from novel views. The top examples are from the BlendedMVS dataset, and the last example is from our captured scenes." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.486, + 0.48, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.713 + ], + "angle": 0, + "content": "Figure 5. Ablation studies on key strategies. The block-level visual comparisons illustrate the impact of adopting our proposed strategy. The first row shows results without the strategy, and the second with the strategy implemented." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.854 + ], + "angle": 0, + "content": "ities, leading to over-decomposition and the inclusion of meaningless parts. Moreover, as shown in Tab. 2 and Tab. 4, our approach achieves competitive results with the advanced Gaussian-based 2DGS [21] and surface reconstruction Neuralangelo [30], which are limited to producing unstructured meshes. Notably, our model demonstrates excellent efficiency, with a reconstruction speed approximately \\(10\\mathrm{X}\\) faster than PartNeRF [47] and over \\(3\\mathrm{X}\\) faster than DBW [36]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Appearance Reconstruction. In addition to part decomposition, our method enables high-fidelity image synthesis, attributed to integrating Gaussian splatting within hybrid" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.219, + 0.591, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.219, + 0.71, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.219, + 0.824, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.219, + 0.905, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.385, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with SAM-based methods. Visual results demonstrate that our method produces more structurally coherent decompositions, whereas SAM-based approaches frequently exhibit spatial discontinuities." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.535 + ], + "angle": 0, + "content": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
Complete model2.3224.920.9060.1220.9436.070.9770.0386.6
w/o Lover2.1125.170.9140.1181.3436.040.9770.0377.6
w/o Lopa5.7221.120.8530.1741.0635.600.9750.0403.8
w/o Lcov3.2122.960.8960.1341.3436.000.9760.0388.53
w/o Adaptive3.1622.740.8800.1411.0535.740.9740.0406.7
w/o Lpar1.9125.480.9180.1150.9336.180.9770.03710.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.574 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the ShapeNet [6]. We report Chamfer Distance, rendering metrics, and the number of primitives (#P)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.691 + ], + "angle": 0, + "content": "representations. EMS [32] and MBF [41] operate directly on point clouds and consequently lack image rendering capabilities, while PartNeRF [47] and DBW [36] yield low-quality view synthesis. In contrast, our method achieves high-quality appearance rendering results, as demonstrated in Tab. 2 and Tab. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.657, + 0.715 + ], + "angle": 0, + "content": "4.3.2. Real-life Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.828 + ], + "angle": 0, + "content": "To further demonstrate the applicability of our method for learning shape decomposition, we test our model on the real-life images from the BlendedMVS dataset [61] and a self-captured dataset. As shown in Fig. 4, our approach can robustly produce both realistic appearances and reasonable 3D decompositions across a variety of data types. More results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.803, + 0.851 + ], + "angle": 0, + "content": "4.3.3. Compared to SAM-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We also conduct comparisons to SAM-based [27] methods for this task, as shown in Fig. 6. Despite impressive advances in 3D segmentation and editing [5, 26, 62, 63] by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.482, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Figure 7. Impact of Gaussian scale regularization (SR). The degraded mesh produced by 2DGS [21] is effectively improved." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.211, + 0.203, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.211, + 0.291, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.212, + 0.382, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.392, + 0.212, + 0.474, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.485, + 0.34 + ], + "angle": 0, + "content": "Figure 8. Impact of the initial number of primitives \\((M)\\). Increasing \\(M\\) yields a finer-grained decomposition, while decreasing it produces a coarser decomposition." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.347, + 0.485, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Figure 9. Applications. Part-Aware Editing (top): After optimization, we can easily edit the scene by adding, scaling, or moving specific parts. 3D Content Generation (bottom): By combining parts of different objects, we can create new 3D content." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.58, + 0.482, + 0.668 + ], + "angle": 0, + "content": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
λpar=0.16.4616.100.7210.2551.0834.300.9870.0182.7
λpar=0.01 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
λpar=0.0014.0120.180.8350.1801.0335.040.9880.0157.1
M=44.9719.060.7940.2131.0634.930.9870.0165.0
M=8 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
M=163.9920.870.830.1761.0735.050.9880.0158.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.721 + ], + "angle": 0, + "content": "Table 6. Effect of parsimony weight \\((\\lambda_{\\mathrm{par}})\\) and initial primitives count \\((M)\\) on DTU [23]. We report Chamfer distance, rendering metrics, and primitives count (#P). * denotes the default setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.837 + ], + "angle": 0, + "content": "distilling 2D information, achieving semantic disentanglement but remains challenging due to indistinct textures and severe cross-view 2D inconsistencies. Lifting 2D segmentation features to 3D segmentation will also lead to feature misalignment, leading to visual artifacts. In contrast, our method performs direct 3D space segmentation, yielding more reasonable and structurally coherent decompositions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.203, + 0.863 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this section, we first analyze the design choices by isolating each strategy to assess its impact. We report the averaged" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.333 + ], + "angle": 0, + "content": "quantitative performance on 15 instances of the chair category in Tab. 5 and provide visual comparisons in Fig. 5. Removing overlapping loss provides better reconstruction accuracy but more overlapping parts. The opacity loss drastically affects the number of primitives and doubles the CD, highlighting its role in controlling primitive presence and reconstruction quality. Coverage loss is crucial in reconstruction accuracy, ensuring primitives align correctly with the target. The adaptive primitive strategy fills in missing parts and improves the reconstruction metric. Without parsimony loss, CD and rendering metrics are optimal but result in over-decomposition. Note that the full model did not yield the highest accuracy because the part-aware reconstruction trades off the accuracy and rationality of the part decomposition. The objective is to achieve high accuracy while preserving reasonable parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.909, + 0.438 + ], + "angle": 0, + "content": "We also validate the effectiveness of Gaussian scale regularization in Fig. 7. 2DGS [21] tends to use large-scale Gaussians in texture-less areas and create holes in the constructed mesh by TSDF integration [67]. The presence of holes stems from the inability of large-scale Gaussians to maintain view-consistent depth. Scale regularization effectively addresses this issue by suppressing large-scale Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.909, + 0.577 + ], + "angle": 0, + "content": "Lastly, in Tab. 6, we analyze the influence of two key hyperparameters on decomposition granularity: the parsimony loss weight \\(\\lambda_{\\mathrm{par}}\\) and the initial primitive count \\(M\\). Stronger parsimony regularization reduces the number of primitives, while weaker regularization increases it. As \\(M\\) rises, both reconstruction and view synthesis performance slightly improve. This demonstrates that adjusting \\(M\\) and \\(\\lambda_{\\mathrm{par}}\\) enables our method to effectively control the granularity of object or scene decomposition. Fig. 8 visually illustrates this impact." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.584, + 0.649, + 0.601 + ], + "angle": 0, + "content": "4.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Fig. 9 illustrates two applications of our method that the original 3DGS and NeRF-based approaches do not support. Firstly, after optimization, we obtain the part decomposition, facilitating easy editing of specific objects or scene components, e.g., adding, moving, removing, or scaling. Secondly, by combining parts of different objects, our method enables the creation of new high-quality 3D content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.642, + 0.74 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduce PartGS, a hybrid representation of superquadrics and 2D Gaussians, to learn 3D scenes in a part-aware representation. Compared to prior works, the proposed method retains geometry details, supporting high-quality image rendering. It obtains state-of-the-art performance in comprehensive evaluations. One limitation is that it takes background-free scene images as inputs, leveraging segmentation tools such as SAM [27]. In the future, we aim to explore how to model backgrounds and extend to larger and more complex scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9656" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.184 + ], + "angle": 0, + "content": "This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.486, + 0.305 + ], + "angle": 0, + "content": "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18013-18023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.307, + 0.485, + 0.348 + ], + "angle": 0, + "content": "[2] Alan H. Barr. Superquadrics and Angle-Preserving Transformations. IEEE Computer Graphics and Applications, 1981. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.351, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[3] Thomas Binford. Visual Perception by Computer. In IEEE Conference on Systems and Control, 1971. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.483, + 0.406 + ], + "angle": 0, + "content": "[4] Åke Björck. Numerics of gram-schmidt orthogonalization. Linear Algebra and Its Applications, 197:297-316, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[5] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.519 + ], + "angle": 0, + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012 [cs.CV], 2015. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[7] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[8] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.485, + 0.661 + ], + "angle": 0, + "content": "[9] Zhiqin Chen, Qimin Chen, Hang Zhou, and Hao Zhang. Dae-net: Deforming auto-encoder for fine-grained shape co-segmentation. arXiv preprint arXiv:2311.13125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.485, + 0.731 + ], + "angle": 0, + "content": "[10] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[11] Martin A. Fischler and Robert C. Bolles. Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography. Communications of the ACM, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[12] Huachen Gao, Shihe Shen, Zhe Zhang, Kaiqiang Xiong, Rui Peng, Zhirui Gao, Qi Wang, Yugui Xie, and Ronggang Wang. Fdc-nerf: learning pose-free neural radiance fields with flow-depth consistency. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3615-3619. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.236, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.178 + ], + "angle": 0, + "content": "[14] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based gaussian splatt-ting for real-time large-scale deformation. arXiv preprint arXiv:2402.04796, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.907, + 0.236 + ], + "angle": 0, + "content": "[15] Zhirui Gao, Renjiao Yi, Zheng Qin, Yunfan Ye, Chenyang Zhu, and Kai Xu. Learning accurate template matching with differentiable coarse-to-fine correspondence refinement. Computational Visual Media, 10(2):309-330, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.239, + 0.907, + 0.28 + ], + "angle": 0, + "content": "[16] Zhirui Gao, Renjiao Yi, Yaqiao Dai, Xuening Zhu, Wei Chen, Chenyang Zhu, and Kai Xu. Curve-aware gaussian splatting for 3d parametric curve reconstruction, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.907, + 0.338 + ], + "angle": 0, + "content": "[17] Zhirui Gao, Renjiao Yi, Chenyang Zhu, Ke Zhuang, Wei Chen, and Kai Xu. Generic objects as pose probes for few-shot view synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.908, + 0.422 + ], + "angle": 0, + "content": "[18] Zeqi Gu, Yin Cui, Zhaoshuo Li, Fangyin Wei, Yunhao Ge, Jinwei Gu, Ming-Yu Liu, Abe Davis, and Yifan Ding. Artiscene: Language-driven artistic 3d scene generation through image intermediary. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2891-2901, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.907, + 0.508 + ], + "angle": 0, + "content": "[19] Yanran Guan, Han Liu, Kun Liu, Kangxue Yin, Ruizhen Hu, Oliver van Kaick, Yan Zhang, Ersin Yumer, Nathan Carr, Radomir Mech, et al. Fame: 3d shape generation via functionality-aware model evolution. IEEE Transactions on Visualization and Computer Graphics, 28(4):1758-1772, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.908, + 0.567 + ], + "angle": 0, + "content": "[20] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.907, + 0.611 + ], + "angle": 0, + "content": "[21] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. SIGGRAPH, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.907, + 0.682 + ], + "angle": 0, + "content": "[22] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18572-18582, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.907, + 0.727 + ], + "angle": 0, + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanaes. Large Scale Multi-view Stereopsis Evaluation. In CVPR, 2014. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.908, + 0.785 + ], + "angle": 0, + "content": "[24] Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, and Na Zhao. Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaussians. arXiv preprint arXiv:2410.01535, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.907, + 0.842 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Chung Min* Kim, Mingxuan* Wu, Justin* Kerr, Matthew Tancik, Ken Goldberg, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9657" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[28] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[29] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised Fitting of Geometric Primitives to 3D Point Clouds. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.438, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[32] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: a Probabilistic Approach. In CVPR, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.482, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[33] Romain Loiseau, Elliot Vincent, Mathieu Aubry, and Loic Landrieu. Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans. arXiv:2304.09704 [cs.CV], 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.484, + 0.638 + ], + "angle": 0, + "content": "[35] Niloy Mitra, Michael Wand, Hao (Richard) Zhang, Daniel Cohen-Or, Vladimir Kim, and Qi-Xing Huang. Structure-aware shape processing. In SIGGRAPH Asia 2013 Courses, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.484, + 0.696 + ], + "angle": 0, + "content": "[36] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives. In NeurIPS, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[37] Chengjie Niu, Jun Li, and Kai Xu. Im2struct: Recovering 3d shape structure from a single rgb image. Cornell University - arXiv, Cornell University - arXiv, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[38] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics Revisited: Learning 3D Shape Parsing Beyond Cuboids. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[39] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning Unsupervised Hierarchical Part Decomposition of 3D Objects from a Single RGB Image. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[41] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In ECCV, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[42] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[44] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems (TODS), 42(3):1-21, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[45] Qingyao Shuai, Chi Zhang, Kaizhi Yang, and Xuejin Chen. Dpf-net: combining explicit shape priors in deformable primitive field for unsupervised structural reconstruction of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14321-14329, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[46] Chunyu Sun, Yuqi Yang, Haoxiang Guo, Pengshuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[47] Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yannis Avrithis, and Leonidas Guibas. PartNeRF: Generating Part-Aware Editable 3D Shapes without 3D Supervision. In CVPR, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.51 + ], + "angle": 0, + "content": "[48] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[49] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.907, + 0.634 + ], + "angle": 0, + "content": "[50] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[51] Xinhang Wan, Jiyuan Liu, Hao Yu, Qian Qu, Ao Li, Xinwang Liu, Ke Liang, Zhibin Dong, and En Zhu. Contrastive continual multiview clustering with filtered structural fusion. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[52] Fengxiang Wang, Mingshuo Chen, Yueying Li, Di Wang, Haotian Wang, Zonghao Guo, Zefan Wang, Boqi Shan, Long Lan, Yulin Wang, et al. Geollava-8k: Scaling remote-sensing multimodal large language models to 8k resolution. arXiv preprint arXiv:2505.21375, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[53] Fengxiang Wang, Hongzhen Wang, Zonghao Guo, Di Wang, Yulin Wang, Mingshuo Chen, Qiang Ma, Long Lan, Wenjing Yang, Jing Zhang, et al. Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14325-14336, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9658" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[55] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliafaschi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.484, + 0.25 + ], + "angle": 0, + "content": "[56] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Vastsd: Learning 3d vascular tree-state space diffusion model for angiography synthesis. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 15693-15702, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.483, + 0.308 + ], + "angle": 0, + "content": "[57] Tong Wu, Lin Gao, Ling-Xiao Zhang, Yu-Kun Lai, and Hao Zhang. Star-tm: Structure aware reconstruction of textured mesh from single image. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[58] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S. Chirikjian. Primitive-based Shape Abstraction via Nonparametric Bayesian Inference. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[59] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2054-2065, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.484, + 0.481 + ], + "angle": 0, + "content": "[60] Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. Cluster-guided contrastive graph clustering network. In Proceedings of the AAAI conference on artificial intelligence, pages 10834-10842, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[61] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks. In CVPR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.484, + 0.582 + ], + "angle": 0, + "content": "[62] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[63] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. arXiv preprint arXiv:2311.11666, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[64] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9491-9500, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.484, + 0.782 + ], + "angle": 0, + "content": "[65] Fenggen Yu, Yimin Qian, Xu Zhang, Francisca Gil-Ureta, Brian Jackson, Eric Bennett, and Hao Zhang. Dpa-net: Structured 3d abstraction from sparse views via differentiable primitive assembly. arXiv preprint arXiv:2404.00875, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.484, + 0.826 + ], + "angle": 0, + "content": "[66] Hongyi Zhou, Xiaogang Wang, Yulan Guo, and Kai Xu. Monomobility: Zero-shot 3d mobility analysis from monocular videos, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.829, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[67] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[68] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "shape co-segmentation with group consistency loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2020. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9659" + } + ] +] \ No newline at end of file diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_origin.pdf b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf92fbc2a57b6e7a9c8328261fb82cfa79ee755d --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/b5799a7c-6527-47cd-81fd-ab5ff79ff559_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03362f2a5c1ad24ecfc0c58aeed241dee00bee1f7c46e007798c3c8e88e7240c +size 8252259 diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/full.md b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d32c9f69f7cd7bee6c7e423c830a56610e5ad6c9 --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/full.md @@ -0,0 +1,391 @@ +# Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics + +Zhirui Gao Renjiao Yi† Yuhang Huang Wei Chen Chenyang Zhu Kai Xu† National University of Defense Technology zhirui-gao.github.io/PartGS + +# Abstract + +Low-level 3D representations, such as point clouds, meshes, NeRFs and 3D Gaussians, are commonly used for modeling 3D objects and scenes. However, cognitive studies indicate that human perception operates at higher levels and interprets 3D environments by decomposing them into meaningful structural parts, rather than low-level elements like points or voxels. Structured geometric decomposition enhances scene interpretability and facilitates downstream tasks requiring component-level manipulation. In this work, we introduce PartGS, a self-supervised part-aware reconstruction framework that integrates 2D Gaussians and superquadrics to parse objects and scenes into an interpretable decomposition, leveraging multi-view image inputs to uncover 3D structural information. Our method jointly optimizes superquadric meshes and Gaussians by coupling their parameters within a hybrid representation. On one hand, superquadrics enable the representation of a wide range of shape primitives, facilitating flexible and meaningful decompositions. On the other hand, 2D Gaussians capture detailed texture and geometric details, ensuring high-fidelity appearance and geometry reconstruction. Operating in a self-supervised manner, our approach demonstrates superior performance compared to state-of-the-art methods across extensive experiments on the DTU, ShapeNet, and real-world datasets. + +# 1. Introduction + +3D reconstruction from multi-view images is a long-standing challenge in 3D vision and graphics [12, 17, 54]. Most reconstructed scenes are in low-level representations such as point clouds, voxels, or meshes. However, humans tend to understand 3D scenes as reasonable parts [35]. For instance, when observing a scene, we naturally construct high-level structural information, such as scene graphs, instead of focusing on low-level details like point clouds or voxels. Motivated by it, we propose a part-aware reconstruction framework + +![](images/e9b5112f0f0000698fc8e41981e9bcd34a1d0411e232cd65435bab3a447439ce.jpg) + +![](images/7960d6e57f33b96c3164e2a169b996caa297f6960aa0b28d639943a539939304.jpg) + +![](images/c70eff9910d387520dac0b67d636cab4ff433cabe32591d4789c5b760d389a60.jpg) + +![](images/fe9a196fd51c54769e7333f9a65c7064bf3e2dae3d63580702f399f6004a1b43.jpg) +3D decomposition +Prior work (EMS) + +![](images/66b22a31d89c1e7f0e5b6e079ec97e74c5ccfba2e479122679f26b944fa75d42.jpg) +Block-level + +![](images/3c5961e98e585a5e6c1936f5358c07eb210b0bf051b91053e8f3f57be874c6ec.jpg) +Point-level +Ours +Figure 1. Prior works [32, 36] model the scene using primitives, while the proposed method can further model precise geometry details and textures. Here, EMS [32] takes point cloud inputs and reconstructs non-textured primitives. + +![](images/d654b2ecedca9d5196535791b9dbc553c6b5ddfd3bad5fe8253689a9104a1689.jpg) + +that decomposes objects or scenes into meaningful shapes or parts, facilitating tasks such as physical simulation, editing, content generation and understanding [18, 28, 52, 53, 56]. + +Several prior works [29, 32, 33, 37, 55, 68] have explored part-aware reconstruction or 3D decomposition. However, most of them rely heavily on 3D supervision and often struggle to retain fine-grained geometric details, limiting their practicality in real-world scenarios. Recent advances [47, 65] have focused on extending the neural radiance field framework for part-aware reconstruction, building upon its remarkable success in 3D reconstruction from multi-view images. For example, PartNeRF [47] models objects using multiple neural radiance fields. Yet, the intricate composition of implicit fields complicates the learning process, leading to suboptimal rendering quality and inefficient decomposition. Recently, DBW [36] proposes a novel approach that decomposes scenes into block-based representations using superquadric primitives [2], optimizing both their parameters and UV texture maps through rendering loss minimization. While this method demonstrates effective scene decomposition into coherent components, it faces challenges in achieving accurate part-aware reconstruction of both geometry and appearance, as illustrated in Fig. 1. + +To address these limitations, we propose PartGS, a self-supervised part-aware hybrid representation that integrates 2D Gaussians [21] and superquadrics [2], to achieve both high-quality texture reconstruction and geometrically accurate part decomposition. There are previous approaches [14, 20, 49] combining mesh reconstruction and Gaussian splatting, where they first reconstruct the mesh and then attach Gaussians to it. In contrast to their primary focus on acquiring the representations, the proposed approach is to achieve part-aware decomposition reconstruction. Our method establishes a coupled optimization framework that simultaneously learns superquadric meshes and Gaussians through parameter sharing, with each part constrained to a single superquadric. The differentiable rendering of Gaussians drives this hybrid representation, leveraging the inherent convexity of superquadrics for qualitative 3D decomposition. The self-supervised training of part-aware reconstruction is built upon the assumption that each part of most objects should be a basic shape and can be represented by a superquadric. Meanwhile, Gaussian splatting, renowned for its superior rendering quality and efficient training, is incorporated to capture intricate texture details, speeding up reconstruction and improving rendering quality. + +In the hybrid representation, 2D Gaussians are distributed on superquadric surfaces to form structured blocks. The pose and shape of Gaussians within each block are determined by their corresponding superquadrics rather than being optimized independently. The parameter space encompasses global controls for superquadric properties (shape, pose, and opacity) and local spherical harmonic coefficients for individual Gaussians. Compared to standard Gaussian splating [21, 25], which populates the occupancy volume with independently parameterized Gaussians, the coupled representation is more compact and efficient. + +During training, the parameters are optimized through rendering loss without additional supervision. However, we notice that image rendering loss often leads to local minima in superquadric shape optimization. To tackle this, we introduce several regularizers to maintain global consistency between the 3D representation and the input 2D information. This strategic implementation allows us to segment parts fully self-supervised, achieving block-level reconstruction. Finally, to better capture irregular shapes, we implement a point-level refinement step that frees 2D Gaussians to deviate from the surface, thereby enhancing geometric fidelity. Extensive experiments show that PartGS, at block-level and point-level, achieves $33.3\%$ and $75.9\%$ improvements in reconstruction accuracy, 3.18 and 16.13 increases in PSNR, and 4X and 3X speedups compared to the state-of-the-art baseline. Our contributions are summarized as follows: + +- We introduce a novel hybrid representation for part-aware 3D reconstruction, combining the strengths of superquadrics and Gaussian splatting to achieve reasonable + +decomposition and high-quality rendering. + +- We propose several novel regularizers to enforce consistency between 3D decomposition and 2D observations, enabling self-supervised part decomposition. +- Compared to prior works, the method takes one step forward, by reconstructing both the block-level and point-level part-aware reconstructions, preserving both part segmentation and reconstruction precision. + +# 2. Related Work + +# 2.1. Shape Decomposition and Abstraction + +Structured shape representation learning decomposes objects or scenes into coherent geometric primitives, facilitating shape understanding and generation [9, 15, 19, 46, 57]. Early approaches like Blocks World [42] and Generalized Cylinders [3] emphasized compact representations. Modern methods typically process 3D inputs (point clouds, meshes, voxels) by decomposing them into primitive ensembles, including cuboids [41, 48], superquadrics [32, 38, 39, 58], and convex shapes [8, 10]. For instance, MonteBoxFinder [41] integrates clustering [11, 50, 60], cuboid representations, and Monte Carlo Tree Search, while EMS [32] adopts a probabilistic approach for superquadric recovery. However, these methods are limited to coarse shape representations. Recent advances in shape abstraction [13, 22, 31, 40, 45] enable more detailed shape representations through flexible primitive deformation. + +Some studies [1, 16, 36, 47, 65, 66] attempt to create structure-aware 3D representations directly from images. PartNeRF [47] introduces ellipsoid representations within NeRFs, its reliance on multiple implicit fields results in inefficient 3D decomposition. ISCO [1] and DBW [36] use 3D superquadrics for shape decomposition, which enables more meaningful structural separation. However, their simple shape parameters struggle to capture complex geometries, leading to poor geometry and appearance reconstruction. DPA-Net [65] has advanced 3D shape abstraction from sparse views but generates redundant parts and struggles with realistic texture rendering. A concurrent work, GaussianBlock [24], employs SAM [27] to guide superquadric splitting and fusion for 3D decomposition, yet its computational efficiency remains limited, typically requiring several hours per scene. In contrast, our approach accomplishes self-supervised [51, 59] part-aware scene reconstruction through an efficient hybrid representation that simultaneously maintains geometric fidelity and photorealistic rendering quality. + +# 2.2. Mesh-based Gaussian Splatting + +Gaussian splitting (GS) [25] has been rapidly adopted in multiple fields due to its remarkable rendering capability. Several studies [7, 14, 20, 49] attempt to align Gaussians with mesh surfaces for easier editing and animation. Among + +![](images/1725736d806238fdb08ae4afbf2598d85df1ee5bf8813240447110681c446bb3.jpg) +Figure 2. Overview of our pipeline. PartGS takes multi-view images to learn a parametric hybrid representation of superquadrics and 2D Gaussians. It initializes from random superquadrics and is gradually optimized during training to obtain a block-level reconstruction. Then, we free the constraints of Gaussians to model detailed geometry, achieving point-level reconstruction. + +them, SuGaR [20] uses flat 3D Gaussians to enforce the alignment with the scene surface during optimization, minimizing the difference between the signed distance functions (SDF) of the desired Gaussians and actual Gaussians. GaMeS [49] introduces a hybrid representation of Gaussians and mesh, where Gaussians are attached to triangular facets of the mesh. Similarly, Gao et al. [14] proposes a mesh-based GS to achieve large-scale deformation effects on objects. Recently, 2DGS [21] proposes 2D Gaussians for surface modeling and significantly enhances the geometric quality. Inspired by them, we propose a representation that combines 2D Gaussians with superquadrics. A key distinction between our method and previous mesh-based GS approaches is that these methods require first reconstructing the scene's mesh and then binding Gaussians to the mesh surface, which results in a non-continuous mesh. In contrast, our method directly optimizes the mesh through a rendering loss, enabling part-aware mesh reconstruction. + +# 3. Method + +Given a set of calibrated multi-view images and foreground masks, we aim to learn part-aware 3D representations, a meaningful decomposition of both geometry and appearance. Our approach adopts a two-stage optimization strategy: first, decomposing the object into basic shapes using a mixture of Gaussians and superquadrics at the block-level, followed by refining the decomposition at the point-level to achieve precise geometry. In Sec 3.1, we parameterize the hybrid representation. Sec. 3.2 then elaborates on leveraging this representation, enhanced by novel regularizers and an adaptive control strategy, to achieve self-supervised block-level + +decomposition. Finally, Sec. 3.3 presents the process for obtaining detailed part-aware results. + +# 3.1. Parametrizing the Hybrid Representation + +As shown in Fig. 2, to leverage the strengths of both, we attach Gaussians to the surface of superquadric meshes. This representation retains the superquadric's ability to parse a 3D scene into distinct parts. Meanwhile, spherical harmonics of Gaussians enable complex texture rendering via Gaussian splatting, addressing the texture learning limitations of prior work [32, 36, 41, 64]. Sharing pose parameters between superquadrics and 2D Gaussians further improves the representation's efficiency. + +The parametric representation is controlled by both primitive and Gaussian parameters, which are optimized simultaneously. Given a 3D scene $S$ , the proposed method decomposes it into multiple hybrid blocks, each consisting of a superquadric with associated Gaussians. Each scene is denoted by the hybrid representation: $S = \mathcal{B}_1 \cup \ldots \cup \mathcal{B}_i \cup \mathcal{B}_M$ , where $\mathcal{B}_i$ is the $i$ -th hybrid block, and $M$ is the total number of blocks. Blocks are defined using manually designed parameters that control pose, opacity, scale, shape, and texture. These parameters are optimized via differentiable rendering to parse the target scene. + +Shape and Scale Parameters. For each hybrid block $\mathcal{B}_i$ , its geometry is controlled by superquadric parameters [2]. Specifically, there are two shape parameters $\epsilon_1, \epsilon_2$ that define its shape, along with three scale parameters $s_1, s_2, s_3$ , which scale the 3D axes. Analogous to icosphere vertex positioning, superquadric vertex coordinates are computed by: + +$$ +\mathbf {v} = \left[ s _ {1} \cos^ {\epsilon 1} (\theta) \cos^ {\epsilon 2} (\varphi); s _ {2} \sin^ {\epsilon 1} (\theta); s _ {3} \cos^ {\epsilon 1} (\theta) \sin^ {\epsilon 2} (\varphi) \right], \quad (1) +$$ + +where $\theta$ and $\varphi$ represent the azimuth and elevation defined + +in the spherical coordinate. The shape and scale parameters govern block deformation to learn part-aware geometry. + +Pose Parameters. The pose of the $i$ -th hybrid block is defined by its rotation $\mathbf{R}_i$ and translation $\mathbf{t}_i$ . The vertex position from Eq. 1 is transformed from the local coordinate to world space as: $\hat{\mathbf{v}}_i^j = \mathbf{R}_i\mathbf{v}_i^j + \mathbf{t}_i$ , where $j$ indexes the vertices. Previous approaches, including SuGaR [20] and GaMeS [49], position Gaussians on reconstructed meshes with independent pose and shape parameters to enhance appearance modeling. In contrast, our method employs Gaussians to construct a differentiable rendering that bridges images and superquadrics. To achieve this, we couple the parameters of Gaussians with superquadrics. + +For a posed superquadric, Gaussian centers are uniformly sampled on triangular faces, with their poses determined by face vertices. Following GaMeS [49], each Gaussian's rotation matrix $\mathrm{R}_v = [r_1,r_2,r_3]$ and scaling $\mathrm{S}_v$ are computed from vertex positions. Given a triangular face $V$ with vertices $v_{1},v_{2},v_{3}\in \mathbb{R}^{3}$ , orthonormal vectors are constructed such that $r_1$ aligns with the face normal, $r_2$ points from the centroid $m = \mathrm{mean}(v_1,v_2,v_3)$ to $v_{1}$ . $r_3$ is obtained by orthonormalizing the vector from the center to the second vertex concerning the existing $r_1$ and $r_2$ : + +$$ +r _ {3} = \frac {\operatorname {o r t} \left(v _ {2} - m ; r _ {1} , r _ {2}\right)}{\left\| \operatorname {o r t} \left(v _ {2} - m ; r _ {1} , r _ {2} \right. \right\|}, \tag {2} +$$ + +where $\text{ort}$ represents the orthonormalization in Gram-Schmidt [4]. For the scaling of 2D Gaussians, we use: + +$$ +\mathrm {S} _ {v} = \operatorname {d i a g} \left(s _ {v} ^ {2}, s _ {v} ^ {3}\right), \tag {3} +$$ + +where $s_v^2 = c||m - v_1||$ , $s_v^3 = c < v_2, r_3 >$ , and $c$ is a size control hyperparameter. We place a fixed number of Gaussians on each triangular face. These designs eliminate the need to learn the geometry of Gaussians, thereby enhancing the efficiency of the representation. + +Opacity Parameters. We define the total number of hybrid blocks as $M$ . However, a typical scene does not contain exactly $M$ blocks. Therefore, only the meaningful blocks are retained. To achieve this, we introduce a learnable parameter $\tau_{i}$ to represent each block's opacity. During the optimization, only blocks with $\tau_{i}$ greater than a certain threshold are retained. Note that Gaussians within the same block share the same $\tau$ for opacity in rasterization. + +Texture Parameters. The texture is modeled using 2D Gaussians positioned on the surface of the superquadrics. Spherical harmonics of each Gaussian control the texture and are optimized for rendering view-dependent images [25]. + +# 3.2. Block-level Decomposition + +This section describes the optimization of the hybrid representation. We observed that minimizing the rendering loss across multi-view images alone led to instability in positioning hybrid blocks. Therefore, several regularization terms + +are introduced to optimize the composition for maximal image formation consistency. + +# 3.2.1. Optimization + +In addition to the image rendering loss, we employ regularizers consisting of coverage, overlap, parsimony, and opacity entropy. The coverage loss encourages hybrid blocks to cover only meaningful regions; the overlap loss prevents blocks from overlapping; the parsimony loss regularizes the number of existing blocks; and the opacity entropy encourages block opacities close to binary. + +Rendering Loss. The rendering loss is based on 3DGS [25], where it combines an $L_{1}$ term with a D-SSIM term: + +$$ +\mathcal {L} _ {\text {r e n}} = (1 - \lambda) L _ {1} + \lambda L _ {\mathrm {D} - \mathrm {S S I M}}. \tag {4} +$$ + +**Coverage Loss.** The coverage loss ensures that the block set $\{\mathcal{B}_1 \ldots \mathcal{B}_M\}$ covers the object while preventing it from extending beyond its boundaries. To determine the 3D occupancy field based on the blocks, we first define the approximate signed distance of a 3D point $x$ to the $i$ -th block, $D_i(x) = \Psi_i(x) - 1$ . Here, $\Psi_i$ denotes the superquadric inside-outside function [2] associated with the block $i$ . Consequently, $D_i(x) \leq 0$ if the point $x$ lies inside the $i$ -th block, and $D_i(x) > 0$ if $x$ is outside the block. Inspired by NeRF [34], we sample a ray set $\mathcal{R}$ using ray-casting based on camera poses. Given the object mask, we associate each ray $r \in \mathcal{R}$ with a binary label: $l_r = 1$ if the ray $r$ is inside mask, otherwise $l_r = 0$ . The coverage loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {c o v}} (\mathcal {R}) = \sum_ {r \in \mathcal {R}} l _ {r} L _ {\text {c r o s s}} (r) + \left(1 - l _ {r}\right) L _ {\text {n o n - c r o s s}} (r). \tag {5} +$$ + +Here, $L_{\mathrm{cross}}$ encourages rays to intersect with blocks, while $L_{\mathrm{non - cross}}$ discourages rays from intersecting with blocks: + +$$ +L _ {\text {c r o s s}} (r) = \operatorname {R e L U} \left(\min _ {i \in \mathcal {M}} \min _ {x _ {j} ^ {r} \in \mathcal {X} _ {r}} D _ {i} \left(x _ {j} ^ {r}\right)\right), \tag {6} +$$ + +$$ +L _ {\text {n o n - c r o s s}} (r) = \operatorname {R e L U} \left(\max _ {i \in \mathcal {M}} \max _ {x _ {j} ^ {r} \in \mathcal {X} _ {r}} - D _ {i} \left(x _ {j} ^ {r}\right)\right), \tag {7} +$$ + +where $x_{j}^{r}$ means the $j$ -th sampled point along the ray $r$ and $\mathcal{M} = \{1, \dots, M\}$ . Intuitively, this implies that at least one sampled points $\mathcal{X}_{r}$ along the ray $r$ inside the mask that lies within a certain block, while all points along the ray outside the mask should not belong to any block. + +Overlap Loss. We introduce a regularization term to prevent overlap between individual blocks. Given the difficulty of directly calculating block overlap, we adopt a Monte Carlo method similar to [36]. Specifically, we sample multiple 3D points in space and penalize those that lie inside more than $k$ blocks. Based on the superquadric inside-outside function, a soft occupancy function is defined as: + +$$ +\mathcal {O} _ {i} ^ {x} = \tau_ {i} \operatorname {s i g m o i d} \left(\frac {- D _ {i} (x)}{\gamma}\right), \tag {8} +$$ + +where $\gamma$ is a temperature parameter. Thus, for a set of $N$ sampled 3D points $\Omega$ , the overlap loss is expressed as: + +$$ +\mathcal {L} _ {\text {o v e r}} = \frac {1}{N} \sum_ {x \in \Omega} \operatorname {R e L U} \left(\sum_ {i \in \mathcal {M}} \mathcal {O} _ {i} ^ {x} - k\right). \tag {9} +$$ + +Parsimony Loss. To promote the use of the minimal number of blocks and achieve parsimony in decomposition, we introduce a regularization term that penalizes block opacity $(\tau)$ . This loss is defined as: + +$$ +\mathcal {L} _ {\text {p a r}} = \frac {1}{M} \sum_ {i \in \mathcal {M}} \sqrt {\tau_ {i}}. \tag {10} +$$ + +Opacity Entropy Loss. During optimization, the opacity of the block inside the object region tends to approach 1, while the opacity of the block outside the object region tends to approach 0. To facilitate this, we associate the opacities of blocks with the labeled rays and apply a cross-entropy loss $L_{ce}$ between the block opacity and the mask labels, defined as follows: + +$$ +\mathcal {L} _ {\mathrm {o p a}} (\mathcal {R}) = \frac {1}{| \mathcal {R} |} \sum_ {r \in \mathcal {R}} L _ {c e} \left(\max _ {i \in \mathcal {M}} \tau_ {i} \left(x ^ {r}\right), l _ {r}\right). \tag {11} +$$ + +Here, only points $x^r$ inside the blocks are sampled, e.g., $\{x^r \in \mathcal{X}_r \mid D_i(x^r) \leq 0\}$ . + +The total loss is the weighted sum of the loss terms described above: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e n}} + \lambda_ {\text {c o v}} \mathcal {L} _ {\text {c o v}} + \lambda_ {\text {o v e r}} \mathcal {L} _ {\text {o v e r}} + \lambda_ {\text {p a r}} \mathcal {L} _ {\text {p a r}} + \lambda_ {\text {o p a}} \mathcal {L} _ {\text {o p a}}. \tag {12} +$$ + +# 3.2.2. Adaptive Number of Blocks. + +Given that the number of parts may vary across different scenes, we allow the number of blocks to adjust during optimization adaptively. Specifically, when the opacity of an existing block falls below a threshold $t$ , this block is removed immediately. Furthermore, we explore a block-adding mechanism to dynamically incorporate new components, ensuring comprehensive coverage of target objects. Specifically, DBSCAN [44] is employed to cluster initial point clouds that are not covered by any blocks, where the point clouds are either randomly initialized or derived as by-products from COLMAP [43]. For each point cloud cluster, a new block is introduced at its center, with the remaining parameters of the block initialized randomly. + +The final block-level reconstruction generates a textured 3D scene represented by multiple superquadrics, with surface details rendered through Gaussians, as illustrated in Fig. 2. + +# 3.3. Point-level Decomposition + +The primitive-based hybrid representation effectively decomposes the shape into parts but exhibits suboptimally for + +complex objects. To address this limitation, we perform a refinement stage that enhances geometric fitting. In this stage, the constraint between Gaussians and superquadrics is decoupled, enabling independent Gaussian optimization. Furthermore, to prevent Gaussian components from one block passing through adjacent blocks and disrupting the continuity and plausibility of the part decomposition, we impose a new constraint to minimize the negative signed distance of each Gaussian entering other blocks: + +$$ +\mathcal {L} _ {\text {e n t e r}} = \frac {1}{N} \sum_ {x \in \Omega} \sum_ {m \in \mathcal {M} \backslash \{\delta \}} \operatorname {R e L U} (- D _ {m} (x)), \tag {13} +$$ + +where $\Omega$ is the sampled Gaussian set, and $\delta$ represents the identifier of the block to which the Gaussian $x$ belongs. As seen in Fig. 2, the reconstruction becomes more accurate and aligns better with the target shape. + +# 4. Experiments + +We comprehensively evaluate our approach across four key aspects: 3D reconstruction quality, view synthesis performance, shape parsimony, and computational efficiency, comparing it with state-of-the-art methods in the field. + +# 4.1. Datasets + +Extensive experiments are conducted on two widely-used public datasets: DTU [23] and ShapeNet [6]. Specifically, 15 standard scenes from the DTU dataset that are widely adopted for reconstruction quality assessment are used for evaluation. Each scene contains either 49 or 64 images. Additionally, we construct a ShapeNet subset consisting of four categories: Chair, Table, Gun, and Airplane, with 15 objects selected per category. For each object, 50 rendered images are generated for training. We further validate real-world applicability on both the BlendedMVS dataset [61] and self-captured scenes, demonstrating robust performance across synthetic and natural environments. + +# 4.2. Implementation Details + +We build upon the 2DGS [21] architecture and add a nonlearnable part attribute to each Gaussian component to enforce the constraint in Eq. 13. In the ShapeNet benchmark, we observe a common degradation in mesh reconstruction as shown in Fig. 7. To address this, we propose Gaussian scale regularization, as validated in Sec. 4.4. We set the initial number of primitives $M$ to 8. The temperature parameter $\gamma$ in Eq. 8 is 0.005 and the overlapping number $k$ in Eq. 9 is 1.95. We train the hybrid representation for 30k iterations, followed by refinement optimization for another 30k iterations. During block-level optimization, the block-adding operation is carried out at the 5k-th and 10k-th iterations. The loss weights $\lambda_{\mathrm{cov}}$ , $\lambda_{\mathrm{over}}$ , $\lambda_{\mathrm{par}}$ , and $\lambda_{\mathrm{opa}}$ are set to 10, 1, 0.002, and 0.01, respectively. For more details about datasets, metrics, and baselines, please refer to Supplementary. + +![](images/9a2893f0459bd4436057f44847546ed02f6e3881a4ec7cc693ace7254912d196.jpg) +Figure 3. Qualitative comparison on DTU [23] and ShapeNet [23]. The first two rows are DTU examples, and the last two are ShapeNet examples, respectively. Our method is the only one that provides reasonable 3D part decomposition while capturing detailed geometry. + +
MethodInputRenderableChamfer distance per sceneMeanMean
S24S37S40S55S63S65S69S83S97S105S106S110S114S118S122CD#P
EMS [32]3D GTX6.323.542.994.304.164.013.753.244.974.344.167.627.584.464.034.657.7
MBF [41]3D GTX3.122.663.842.541.592.112.192.012.322.452.172.123.832.022.552.5034.1
EMS [32] + Neus [54]ImageX5.995.564.434.325.426.143.753.964.634.345.885.114.294.833.534.978.87
MBF [41] + Neus [54]ImageX2.693.373.222.693.632.602.593.132.852.512.453.722.242.492.522.8546.7
PartNeRF [47]Image✓9.3810.469.088.636.047.257.229.158.7210.016.729.857.858.689.218.548.0
DBW [36]Image✓5.418.351.573.083.404.157.463.946.634.854.384.656.294.343.044.764.8
Ours (Block-level)Image✓5.684.911.852.613.754.663.757.574.274.383.494.483.614.213.704.195.9
Ours (Point-level)Image✓0.701.170.550.651.061.231.101.361.370.780.921.410.691.050.710.985.9
+ +Table 1. Quantitative comparison on DTU [23]. The Chamfer distance between the 3D reconstruction and the ground-truth is reported in 15 scenes. The best results are bolded, and the average numbers of primitives found (#P) that are greater than 10 are underlined. + +
MethodPart-awareCD ↓PSNR ↑SSIM ↑LIPPS ↓Time ↓
Neuralangelo [30]X0.6133.84--> 10 h
2DGS [21]X0.8134.070.990.019~ 10 m
PartNeRF [47]✓9.5917.970.770.246~ 8 h
DBW [36]✓4.7316.440.750.201~ 2 h
Ours (Block-level)✓4.1919.840.820.189~ 30 m
Ours (Point-level)✓0.9835.040.990.015~ 40 m
+ +Table 2. Quantitative results on DTU [23]. Our method outperforms all part-aware approaches in image synthesis quality, reconstruction accuracy, and efficiency. Neuralangelo's results are from the original paper, with all times measured on an RTX 3090 GPU. + +# 4.3. Evaluations + +# 4.3.1. DTU and Shapenet Benchmark + +Geometry Reconstruction. In Tab. 1 and Tab. 3, we compare our geometry reconstruction to SOTA shape decomposition methods on Chamfer distance and training time using DTU and Shapenet dataset. The Chamfer distance + +
MethodInputChamfer Distance ↓Primitives (#P)MeanMean
AirplaneTableChairGunAirplaneTableChairGunCD#P
EMS [32]3D GT3.406.9219.02.029.47.8810.38.47.849.0
MBF [41]3D GT2.832.181.592.3210.8513.913.414.32.2113.1
PartNeRF [47]Image2.292.772.302.468.08.08.08.02.468.0
DBW [36]Image3.617.336.192.092.75.23.63.34.813.7
Ours (Block-level)Image2.472.152.321.783.96.67.65.02.185.8
Ours (Point-level)Image1.291.720.941.073.96.67.65.01.255.8
+ +Table 3. Quantitative comparison on ShapeNet [6]. We report Chamfer distance and the number of parts. The best results are bolded, and the second-best results are underlined. + +metrics reported for ShapeNet are scaled by a factor of 100 for readability. Our method consistently outperforms in all scenes compared to prior works. As shown in Fig. 3, our approach consistently produces interpretable 3D decompositions with further refinement achieving more detailed geometry (the last two columns). MBF [41] achieves a low CD error but at the cost of using significantly more prim + +
MethodPart-awareChamfer Distance ↓PSNR ↑SSIM ↑LIPPS ↓
AirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGun
2DGS [21]✗1.472.370.501.0340.8939.8039.0541.740.9940.9900.9900.9940.0090.0270.0170.009
PartNeRF [47]✓2.292.772.302.4619.6320.6619.0821.970.8980.8550.8750.9160.0860.1610.1360.083
DBW [36]✓3.617.336.192.0926.1123.8420.2528.720.9500.9150.8920.9600.0740.1360.1320.042
Ours (Block-level)✓2.472.152.321.7827.9427.9824.9229.950.9590.9250.9060.9630.0720.1290.1220.047
Ours (Point-level)✓1.291.720.941.0741.1836.8036.0739.510.9920.9730.9770.9890.0140.0700.0380.021
+ +![](images/24ae7e5a67815b404bd07fee031db8fba68f96091208d7f64af3ec24f97d31af.jpg) +Figure 4. Qualitative results on BlendedMVS [61] and self-captured data. We present RGB renderings and decomposed parts from novel views. The top examples are from the BlendedMVS dataset, and the last example is from our captured scenes. + +![](images/c926f3ba22a56d86a32c5e06a2a597aa03f4573cc5f5d110f7785cc401a7d7e3.jpg) + +![](images/f27da3159df474a17048759326d7e7f2cf043bbb41b0910ae41fe2687b249422.jpg) + +![](images/0a9fecf992c3b20d510816d220905d3c02c3f955ff86d2a7b933ec78efbd4807.jpg) +Figure 6. Qualitative comparison with SAM-based methods. Visual results demonstrate that our method produces more structurally coherent decompositions, whereas SAM-based approaches frequently exhibit spatial discontinuities. + +![](images/0cf1c6b0b7ee14a1103a679235523a6b7e3143fe7142ee49d274421f4a041d4a.jpg) +Figure 5. Ablation studies on key strategies. The block-level visual comparisons illustrate the impact of adopting our proposed strategy. The first row shows results without the strategy, and the second with the strategy implemented. + +ities, leading to over-decomposition and the inclusion of meaningless parts. Moreover, as shown in Tab. 2 and Tab. 4, our approach achieves competitive results with the advanced Gaussian-based 2DGS [21] and surface reconstruction Neuralangelo [30], which are limited to producing unstructured meshes. Notably, our model demonstrates excellent efficiency, with a reconstruction speed approximately $10\mathrm{X}$ faster than PartNeRF [47] and over $3\mathrm{X}$ faster than DBW [36]. + +Appearance Reconstruction. In addition to part decomposition, our method enables high-fidelity image synthesis, attributed to integrating Gaussian splatting within hybrid + +![](images/307d665fea83eea93e22db493cd3d7d1a8350fa763b31b558dd767118b793551.jpg) + +![](images/70407c2d856a6e1c769162ddf9593c72e1736e682f1fbac84e02effa1cfda752.jpg) + +![](images/10cdb9e196463b04c8405d36179586a3b92d152ed2795781a4014c35037727cf.jpg) + +![](images/4111f7b72c108581367c01f4837d509695d917e6f151090dc0b0072f30e0d65e.jpg) + +Table 4. Quantitative results on ShapeNet [6]. We report the Chamfer distance and novel view synthesis results across four categories. + +
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
Complete model2.3224.920.9060.1220.9436.070.9770.0386.6
w/o Lover2.1125.170.9140.1181.3436.040.9770.0377.6
w/o Lopa5.7221.120.8530.1741.0635.600.9750.0403.8
w/o Lcov3.2122.960.8960.1341.3436.000.9760.0388.53
w/o Adaptive3.1622.740.8800.1411.0535.740.9740.0406.7
w/o Lpar1.9125.480.9180.1150.9336.180.9770.03710.1
+ +Table 5. Ablation studies on the ShapeNet [6]. We report Chamfer Distance, rendering metrics, and the number of primitives (#P). + +representations. EMS [32] and MBF [41] operate directly on point clouds and consequently lack image rendering capabilities, while PartNeRF [47] and DBW [36] yield low-quality view synthesis. In contrast, our method achieves high-quality appearance rendering results, as demonstrated in Tab. 2 and Tab. 4. + +# 4.3.2. Real-life Data + +To further demonstrate the applicability of our method for learning shape decomposition, we test our model on the real-life images from the BlendedMVS dataset [61] and a self-captured dataset. As shown in Fig. 4, our approach can robustly produce both realistic appearances and reasonable 3D decompositions across a variety of data types. More results are provided in the supplementary material. + +# 4.3.3. Compared to SAM-based Methods + +We also conduct comparisons to SAM-based [27] methods for this task, as shown in Fig. 6. Despite impressive advances in 3D segmentation and editing [5, 26, 62, 63] by + +![](images/c82033cd52c3d624329aede2b26145e5a5fa59c98af17cea26b89603aa1ec4e6.jpg) +Figure 7. Impact of Gaussian scale regularization (SR). The degraded mesh produced by 2DGS [21] is effectively improved. + +![](images/143f6800a56837a75210e2b7e2342ddd66ce215d4f93d1ed9af327ee18b64fad.jpg) +Figure 8. Impact of the initial number of primitives $(M)$ . Increasing $M$ yields a finer-grained decomposition, while decreasing it produces a coarser decomposition. + +![](images/4fe97c9d7eee1036ca983cb6d1ff5a65d638c7a2d84c52711fd4a9c3ebf044ab.jpg) + +![](images/b9e5052eafdcbb60ada0535cf39aab79f0c65599843c8f28937ab275a24121cd.jpg) + +![](images/37e420abf3cb308307587cda0307fd45dcd6cef15b6656db7d4300f0038b3a18.jpg) + +![](images/2fcb7e24646fc7aec39c0b4580b93c9fd57fc7ce80c30d2386654825a3e18cd5.jpg) +Figure 9. Applications. Part-Aware Editing (top): After optimization, we can easily edit the scene by adding, scaling, or moving specific parts. 3D Content Generation (bottom): By combining parts of different objects, we can create new 3D content. + +
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
λpar=0.16.4616.100.7210.2551.0834.300.9870.0182.7
λpar=0.01 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
λpar=0.0014.0120.180.8350.1801.0335.040.9880.0157.1
M=44.9719.060.7940.2131.0634.930.9870.0165.0
M=8 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
M=163.9920.870.830.1761.0735.050.9880.0158.3
+ +Table 6. Effect of parsimony weight $(\lambda_{\mathrm{par}})$ and initial primitives count $(M)$ on DTU [23]. We report Chamfer distance, rendering metrics, and primitives count (#P). * denotes the default setting. + +distilling 2D information, achieving semantic disentanglement but remains challenging due to indistinct textures and severe cross-view 2D inconsistencies. Lifting 2D segmentation features to 3D segmentation will also lead to feature misalignment, leading to visual artifacts. In contrast, our method performs direct 3D space segmentation, yielding more reasonable and structurally coherent decompositions. + +# 4.4. Ablations + +In this section, we first analyze the design choices by isolating each strategy to assess its impact. We report the averaged + +quantitative performance on 15 instances of the chair category in Tab. 5 and provide visual comparisons in Fig. 5. Removing overlapping loss provides better reconstruction accuracy but more overlapping parts. The opacity loss drastically affects the number of primitives and doubles the CD, highlighting its role in controlling primitive presence and reconstruction quality. Coverage loss is crucial in reconstruction accuracy, ensuring primitives align correctly with the target. The adaptive primitive strategy fills in missing parts and improves the reconstruction metric. Without parsimony loss, CD and rendering metrics are optimal but result in over-decomposition. Note that the full model did not yield the highest accuracy because the part-aware reconstruction trades off the accuracy and rationality of the part decomposition. The objective is to achieve high accuracy while preserving reasonable parts. + +We also validate the effectiveness of Gaussian scale regularization in Fig. 7. 2DGS [21] tends to use large-scale Gaussians in texture-less areas and create holes in the constructed mesh by TSDF integration [67]. The presence of holes stems from the inability of large-scale Gaussians to maintain view-consistent depth. Scale regularization effectively addresses this issue by suppressing large-scale Gaussians. + +Lastly, in Tab. 6, we analyze the influence of two key hyperparameters on decomposition granularity: the parsimony loss weight $\lambda_{\mathrm{par}}$ and the initial primitive count $M$ . Stronger parsimony regularization reduces the number of primitives, while weaker regularization increases it. As $M$ rises, both reconstruction and view synthesis performance slightly improve. This demonstrates that adjusting $M$ and $\lambda_{\mathrm{par}}$ enables our method to effectively control the granularity of object or scene decomposition. Fig. 8 visually illustrates this impact. + +# 4.5. Applications + +Fig. 9 illustrates two applications of our method that the original 3DGS and NeRF-based approaches do not support. Firstly, after optimization, we obtain the part decomposition, facilitating easy editing of specific objects or scene components, e.g., adding, moving, removing, or scaling. Secondly, by combining parts of different objects, our method enables the creation of new high-quality 3D content. + +# 5. Conclusions + +We introduce PartGS, a hybrid representation of superquadrics and 2D Gaussians, to learn 3D scenes in a part-aware representation. Compared to prior works, the proposed method retains geometry details, supporting high-quality image rendering. It obtains state-of-the-art performance in comprehensive evaluations. One limitation is that it takes background-free scene images as inputs, leveraging segmentation tools such as SAM [27]. In the future, we aim to explore how to model backgrounds and extend to larger and more complex scenes. + +# Acknowledgements + +This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104). + +# References + +[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18013-18023, 2023. 2 +[2] Alan H. Barr. Superquadrics and Angle-Preserving Transformations. IEEE Computer Graphics and Applications, 1981. 1, 2, 3, 4 +[3] Thomas Binford. Visual Perception by Computer. In IEEE Conference on Systems and Control, 1971. 2 +[4] Åke Björck. Numerics of gram-schmidt orthogonalization. Linear Algebra and Its Applications, 197:297-316, 1994. 4 +[5] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 7 +[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012 [cs.CV], 2015. 5, 6, 7 +[7] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2 +[8] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[9] Zhiqin Chen, Qimin Chen, Hang Zhou, and Hao Zhang. Dae-net: Deforming auto-encoder for fine-grained shape co-segmentation. arXiv preprint arXiv:2311.13125, 2023. 2 +[10] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[11] Martin A. Fischler and Robert C. Bolles. Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography. Communications of the ACM, 1981. 2 +[12] Huachen Gao, Shihe Shen, Zhe Zhang, Kaiqiang Xiong, Rui Peng, Zhirui Gao, Qi Wang, Yugui Xie, and Ronggang Wang. Fdc-nerf: learning pose-free neural radiance fields with flow-depth consistency. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3615-3619. IEEE, 2024. 1 +[13] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network + +for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2 +[14] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based gaussian splatt-ting for real-time large-scale deformation. arXiv preprint arXiv:2402.04796, 2024. 2, 3 +[15] Zhirui Gao, Renjiao Yi, Zheng Qin, Yunfan Ye, Chenyang Zhu, and Kai Xu. Learning accurate template matching with differentiable coarse-to-fine correspondence refinement. Computational Visual Media, 10(2):309-330, 2024. 2 +[16] Zhirui Gao, Renjiao Yi, Yaqiao Dai, Xuening Zhu, Wei Chen, Chenyang Zhu, and Kai Xu. Curve-aware gaussian splatting for 3d parametric curve reconstruction, 2025. 2 +[17] Zhirui Gao, Renjiao Yi, Chenyang Zhu, Ke Zhuang, Wei Chen, and Kai Xu. Generic objects as pose probes for few-shot view synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 1 +[18] Zeqi Gu, Yin Cui, Zhaoshuo Li, Fangyin Wei, Yunhao Ge, Jinwei Gu, Ming-Yu Liu, Abe Davis, and Yifan Ding. Artiscene: Language-driven artistic 3d scene generation through image intermediary. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2891-2901, 2025. 1 +[19] Yanran Guan, Han Liu, Kun Liu, Kangxue Yin, Ruizhen Hu, Oliver van Kaick, Yan Zhang, Ersin Yumer, Nathan Carr, Radomir Mech, et al. Fame: 3d shape generation via functionality-aware model evolution. IEEE Transactions on Visualization and Computer Graphics, 28(4):1758-1772, 2020. 2 +[20] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3, 4 +[21] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. SIGGRAPH, 2024. 2, 3, 5, 6, 7, 8 +[22] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18572-18582, 2022. 2 +[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanaes. Large Scale Multi-view Stereopsis Evaluation. In CVPR, 2014. 5, 6, 8 +[24] Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, and Na Zhao. Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaussians. arXiv preprint arXiv:2410.01535, 2024. 2 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 4 +[26] Chung Min* Kim, Mingxuan* Wu, Justin* Kerr, Matthew Tancik, Ken Goldberg, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7 + +[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 7, 8 +[28] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 1 +[29] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised Fitting of Geometric Primitives to 3D Point Clouds. In CVPR, 2019. 1 +[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 6, 7 +[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023. 2 +[32] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: a Probabilistic Approach. In CVPR, 2022. 1, 2, 3, 6, 7 +[33] Romain Loiseau, Elliot Vincent, Mathieu Aubry, and Loic Landrieu. Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans. arXiv:2304.09704 [cs.CV], 2023. 1 +[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 4 +[35] Niloy Mitra, Michael Wand, Hao (Richard) Zhang, Daniel Cohen-Or, Vladimir Kim, and Qi-Xing Huang. Structure-aware shape processing. In SIGGRAPH Asia 2013 Courses, 2013. 1 +[36] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives. In NeurIPS, 2023. 1, 2, 3, 4, 6, 7 +[37] Chengjie Niu, Jun Li, and Kai Xu. Im2struct: Recovering 3d shape structure from a single rgb image. Cornell University - arXiv, Cornell University - arXiv, 2018. 1 +[38] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics Revisited: Learning 3D Shape Parsing Beyond Cuboids. In CVPR, 2019. 2 +[39] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning Unsupervised Hierarchical Part Decomposition of 3D Objects from a Single RGB Image. In CVPR, 2020. 2 +[40] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2 + +[41] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In ECCV, 2022. 2, 3, 6, 7 +[42] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2 +[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5 +[44] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems (TODS), 42(3):1-21, 2017. 5 +[45] Qingyao Shuai, Chi Zhang, Kaizhi Yang, and Xuejin Chen. Dpf-net: combining explicit shape priors in deformable primitive field for unsupervised structural reconstruction of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14321-14329, 2023. 2 +[46] Chunyu Sun, Yuqi Yang, Haoxiang Guo, Pengshuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media, 2022. 2 +[47] Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yannis Avrithis, and Leonidas Guibas. PartNeRF: Generating Part-Aware Editable 3D Shapes without 3D Supervision. In CVPR, 2023. 1, 2, 6, 7 +[48] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. In CVPR, 2017. 2 +[49] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2, 3, 4 +[50] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 2 +[51] Xinhang Wan, Jiyuan Liu, Hao Yu, Qian Qu, Ao Li, Xinwang Liu, Ke Liang, Zhibin Dong, and En Zhu. Contrastive continual multiview clustering with filtered structural fusion. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2 +[52] Fengxiang Wang, Mingshuo Chen, Yueying Li, Di Wang, Haotian Wang, Zonghao Guo, Zefan Wang, Boqi Shan, Long Lan, Yulin Wang, et al. Geollava-8k: Scaling remote-sensing multimodal large language models to 8k resolution. arXiv preprint arXiv:2505.21375, 2025. 1 +[53] Fengxiang Wang, Hongzhen Wang, Zonghao Guo, Di Wang, Yulin Wang, Mingshuo Chen, Qiang Ma, Long Lan, Wenjing Yang, Jing Zhang, et al. Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14325-14336, 2025. 1 +[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit + +Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1, 6 +[55] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliafaschi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 1 +[56] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Vastsd: Learning 3d vascular tree-state space diffusion model for angiography synthesis. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 15693-15702, 2025. 1 +[57] Tong Wu, Lin Gao, Ling-Xiao Zhang, Yu-Kun Lai, and Hao Zhang. Star-tm: Structure aware reconstruction of textured mesh from single image. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2021. 2 +[58] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S. Chirikjian. Primitive-based Shape Abstraction via Nonparametric Bayesian Inference. In ECCV, 2022. 2 +[59] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2054-2065, 2022. 2 +[60] Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. Cluster-guided contrastive graph clustering network. In Proceedings of the AAAI conference on artificial intelligence, pages 10834-10842, 2023. 2 +[61] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks. In CVPR, 2020. 5, 7 +[62] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 7 +[63] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. arXiv preprint arXiv:2311.11666, 2023. 7 +[64] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9491-9500, 2019. 3 +[65] Fenggen Yu, Yimin Qian, Xu Zhang, Francisca Gil-Ureta, Brian Jackson, Eric Bennett, and Hao Zhang. Dpa-net: Structured 3d abstraction from sparse views via differentiable primitive assembly. arXiv preprint arXiv:2404.00875, 2024. 1, 2 +[66] Hongyi Zhou, Xiaogang Wang, Yulan Guo, and Kai Xu. Monomobility: Zero-shot 3d mobility analysis from monocular videos, 2025. 2 +[67] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.8 +[68] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive + +shape co-segmentation with group consistency loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2020. 1 \ No newline at end of file diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/images.zip b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7101a7490e0d51e04be6bfa2aa9114fae90c306d --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7ffe0eaf2e7929ffb0ad805cbde11c23f913de25a5c5f47f8d7b5b44d009184 +size 756809 diff --git a/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/layout.json b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7cc8d2bfe8279173f05bc86846f555d42ad6ffc --- /dev/null +++ b/2025/Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics/layout.json @@ -0,0 +1,10251 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 514, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 514, + 140 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 514, + 140 + ], + "type": "text", + "content": "Self-supervised Learning of Hybrid Part-aware 3D Representations of 2D Gaussians and Superquadrics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 512, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 512, + 204 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 512, + 204 + ], + "type": "text", + "content": "Zhirui Gao Renjiao Yi† Yuhang Huang Wei Chen Chenyang Zhu Kai Xu† National University of Defense Technology zhirui-gao.github.io/PartGS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "Low-level 3D representations, such as point clouds, meshes, NeRFs and 3D Gaussians, are commonly used for modeling 3D objects and scenes. However, cognitive studies indicate that human perception operates at higher levels and interprets 3D environments by decomposing them into meaningful structural parts, rather than low-level elements like points or voxels. Structured geometric decomposition enhances scene interpretability and facilitates downstream tasks requiring component-level manipulation. In this work, we introduce PartGS, a self-supervised part-aware reconstruction framework that integrates 2D Gaussians and superquadrics to parse objects and scenes into an interpretable decomposition, leveraging multi-view image inputs to uncover 3D structural information. Our method jointly optimizes superquadric meshes and Gaussians by coupling their parameters within a hybrid representation. On one hand, superquadrics enable the representation of a wide range of shape primitives, facilitating flexible and meaningful decompositions. On the other hand, 2D Gaussians capture detailed texture and geometric details, ensuring high-fidelity appearance and geometry reconstruction. Operating in a self-supervised manner, our approach demonstrates superior performance compared to state-of-the-art methods across extensive experiments on the DTU, ShapeNet, and real-world datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "3D reconstruction from multi-view images is a long-standing challenge in 3D vision and graphics [12, 17, 54]. Most reconstructed scenes are in low-level representations such as point clouds, voxels, or meshes. However, humans tend to understand 3D scenes as reasonable parts [35]. For instance, when observing a scene, we naturally construct high-level structural information, such as scene graphs, instead of focusing on low-level details like point clouds or voxels. Motivated by it, we propose a part-aware reconstruction framework" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 234, + 369, + 291 + ], + "blocks": [ + { + "bbox": [ + 318, + 234, + 369, + 291 + ], + "lines": [ + { + "bbox": [ + 318, + 234, + 369, + 291 + ], + "spans": [ + { + "bbox": [ + 318, + 234, + 369, + 291 + ], + "type": "image", + "image_path": "e9b5112f0f0000698fc8e41981e9bcd34a1d0411e232cd65435bab3a447439ce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 374, + 235, + 435, + 289 + ], + "blocks": [ + { + "bbox": [ + 374, + 235, + 435, + 289 + ], + "lines": [ + { + "bbox": [ + 374, + 235, + 435, + 289 + ], + "spans": [ + { + "bbox": [ + 374, + 235, + 435, + 289 + ], + "type": "image", + "image_path": "7960d6e57f33b96c3164e2a169b996caa297f6960aa0b28d639943a539939304.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 443, + 235, + 548, + 291 + ], + "blocks": [ + { + "bbox": [ + 443, + 235, + 548, + 291 + ], + "lines": [ + { + "bbox": [ + 443, + 235, + 548, + 291 + ], + "spans": [ + { + "bbox": [ + 443, + 235, + 548, + 291 + ], + "type": "image", + "image_path": "c70eff9910d387520dac0b67d636cab4ff433cabe32591d4789c5b760d389a60.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 304, + 370, + 349 + ], + "blocks": [ + { + "bbox": [ + 323, + 296, + 368, + 304 + ], + "lines": [ + { + "bbox": [ + 323, + 296, + 368, + 304 + ], + "spans": [ + { + "bbox": [ + 323, + 296, + 368, + 304 + ], + "type": "text", + "content": "3D decomposition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 304, + 370, + 349 + ], + "lines": [ + { + "bbox": [ + 317, + 304, + 370, + 349 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 370, + 349 + ], + "type": "image", + "image_path": "fe9a196fd51c54769e7333f9a65c7064bf3e2dae3d63580702f399f6004a1b43.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 356, + 364, + 363 + ], + "lines": [ + { + "bbox": [ + 322, + 356, + 364, + 363 + ], + "spans": [ + { + "bbox": [ + 322, + 356, + 364, + 363 + ], + "type": "text", + "content": "Prior work (EMS)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 381, + 304, + 432, + 352 + ], + "blocks": [ + { + "bbox": [ + 393, + 296, + 421, + 303 + ], + "lines": [ + { + "bbox": [ + 393, + 296, + 421, + 303 + ], + "spans": [ + { + "bbox": [ + 393, + 296, + 421, + 303 + ], + "type": "text", + "content": "Block-level" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 304, + 432, + 352 + ], + "lines": [ + { + "bbox": [ + 381, + 304, + 432, + 352 + ], + "spans": [ + { + "bbox": [ + 381, + 304, + 432, + 352 + ], + "type": "image", + "image_path": "66b22a31d89c1e7f0e5b6e079ec97e74c5ccfba2e479122679f26b944fa75d42.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 437, + 300, + 492, + 354 + ], + "blocks": [ + { + "bbox": [ + 453, + 296, + 480, + 302 + ], + "lines": [ + { + "bbox": [ + 453, + 296, + 480, + 302 + ], + "spans": [ + { + "bbox": [ + 453, + 296, + 480, + 302 + ], + "type": "text", + "content": "Point-level" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 300, + 492, + 354 + ], + "lines": [ + { + "bbox": [ + 437, + 300, + 492, + 354 + ], + "spans": [ + { + "bbox": [ + 437, + 300, + 492, + 354 + ], + "type": "image", + "image_path": "3c5961e98e585a5e6c1936f5358c07eb210b0bf051b91053e8f3f57be874c6ec.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 356, + 466, + 363 + ], + "lines": [ + { + "bbox": [ + 452, + 356, + 466, + 363 + ], + "spans": [ + { + "bbox": [ + 452, + 356, + 466, + 363 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 374, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Prior works [32, 36] model the scene using primitives, while the proposed method can further model precise geometry details and textures. Here, EMS [32] takes point cloud inputs and reconstructs non-textured primitives." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 492, + 308, + 548, + 354 + ], + "blocks": [ + { + "bbox": [ + 492, + 308, + 548, + 354 + ], + "lines": [ + { + "bbox": [ + 492, + 308, + 548, + 354 + ], + "spans": [ + { + "bbox": [ + 492, + 308, + 548, + 354 + ], + "type": "image", + "image_path": "d654b2ecedca9d5196535791b9dbc553c6b5ddfd3bad5fe8253689a9104a1689.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 436, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 472 + ], + "type": "text", + "content": "that decomposes objects or scenes into meaningful shapes or parts, facilitating tasks such as physical simulation, editing, content generation and understanding [18, 28, 52, 53, 56]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": "Several prior works [29, 32, 33, 37, 55, 68] have explored part-aware reconstruction or 3D decomposition. However, most of them rely heavily on 3D supervision and often struggle to retain fine-grained geometric details, limiting their practicality in real-world scenarios. Recent advances [47, 65] have focused on extending the neural radiance field framework for part-aware reconstruction, building upon its remarkable success in 3D reconstruction from multi-view images. For example, PartNeRF [47] models objects using multiple neural radiance fields. Yet, the intricate composition of implicit fields complicates the learning process, leading to suboptimal rendering quality and inefficient decomposition. Recently, DBW [36] proposes a novel approach that decomposes scenes into block-based representations using superquadric primitives [2], optimizing both their parameters and UV texture maps through rendering loss minimization. While this method demonstrates effective scene decomposition into coherent components, it faces challenges in achieving accurate part-aware reconstruction of both geometry and appearance, as illustrated in Fig. 1." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 151, + 712 + ], + "type": "text", + "content": "† Corresponding authors" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9649" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 347 + ], + "type": "text", + "content": "To address these limitations, we propose PartGS, a self-supervised part-aware hybrid representation that integrates 2D Gaussians [21] and superquadrics [2], to achieve both high-quality texture reconstruction and geometrically accurate part decomposition. There are previous approaches [14, 20, 49] combining mesh reconstruction and Gaussian splatting, where they first reconstruct the mesh and then attach Gaussians to it. In contrast to their primary focus on acquiring the representations, the proposed approach is to achieve part-aware decomposition reconstruction. Our method establishes a coupled optimization framework that simultaneously learns superquadric meshes and Gaussians through parameter sharing, with each part constrained to a single superquadric. The differentiable rendering of Gaussians drives this hybrid representation, leveraging the inherent convexity of superquadrics for qualitative 3D decomposition. The self-supervised training of part-aware reconstruction is built upon the assumption that each part of most objects should be a basic shape and can be represented by a superquadric. Meanwhile, Gaussian splatting, renowned for its superior rendering quality and efficient training, is incorporated to capture intricate texture details, speeding up reconstruction and improving rendering quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 481 + ], + "type": "text", + "content": "In the hybrid representation, 2D Gaussians are distributed on superquadric surfaces to form structured blocks. The pose and shape of Gaussians within each block are determined by their corresponding superquadrics rather than being optimized independently. The parameter space encompasses global controls for superquadric properties (shape, pose, and opacity) and local spherical harmonic coefficients for individual Gaussians. Compared to standard Gaussian splating [21, 25], which populates the occupancy volume with independently parameterized Gaussians, the coupled representation is more compact and efficient." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "text", + "content": "During training, the parameters are optimized through rendering loss without additional supervision. However, we notice that image rendering loss often leads to local minima in superquadric shape optimization. To tackle this, we introduce several regularizers to maintain global consistency between the 3D representation and the input 2D information. This strategic implementation allows us to segment parts fully self-supervised, achieving block-level reconstruction. Finally, to better capture irregular shapes, we implement a point-level refinement step that frees 2D Gaussians to deviate from the surface, thereby enhancing geometric fidelity. Extensive experiments show that PartGS, at block-level and point-level, achieves " + }, + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "inline_equation", + "content": "33.3\\%" + }, + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "inline_equation", + "content": "75.9\\%" + }, + { + "bbox": [ + 55, + 483, + 297, + 674 + ], + "type": "text", + "content": " improvements in reconstruction accuracy, 3.18 and 16.13 increases in PSNR, and 4X and 3X speedups compared to the state-of-the-art baseline. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 297, + 714 + ], + "type": "text", + "content": "- We introduce a novel hybrid representation for part-aware 3D reconstruction, combining the strengths of superquadrics and Gaussian splatting to achieve reasonable" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 324, + 72, + 498, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 498, + 84 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 498, + 84 + ], + "type": "text", + "content": "decomposition and high-quality rendering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 85, + 555, + 168 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 317, + 85, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 85, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 317, + 85, + 555, + 120 + ], + "type": "text", + "content": "- We propose several novel regularizers to enforce consistency between 3D decomposition and 2D observations, enabling self-supervised part decomposition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 121, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 555, + 168 + ], + "type": "text", + "content": "- Compared to prior works, the method takes one step forward, by reconstructing both the block-level and point-level part-aware reconstructions, preserving both part segmentation and reconstruction precision." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 181, + 401, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 181, + 401, + 193 + ], + "spans": [ + { + "bbox": [ + 314, + 181, + 401, + 193 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 517, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 517, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 517, + 213 + ], + "type": "text", + "content": "2.1. Shape Decomposition and Abstraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 219, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 410 + ], + "type": "text", + "content": "Structured shape representation learning decomposes objects or scenes into coherent geometric primitives, facilitating shape understanding and generation [9, 15, 19, 46, 57]. Early approaches like Blocks World [42] and Generalized Cylinders [3] emphasized compact representations. Modern methods typically process 3D inputs (point clouds, meshes, voxels) by decomposing them into primitive ensembles, including cuboids [41, 48], superquadrics [32, 38, 39, 58], and convex shapes [8, 10]. For instance, MonteBoxFinder [41] integrates clustering [11, 50, 60], cuboid representations, and Monte Carlo Tree Search, while EMS [32] adopts a probabilistic approach for superquadric recovery. However, these methods are limited to coarse shape representations. Recent advances in shape abstraction [13, 22, 31, 40, 45] enable more detailed shape representations through flexible primitive deformation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 639 + ], + "type": "text", + "content": "Some studies [1, 16, 36, 47, 65, 66] attempt to create structure-aware 3D representations directly from images. PartNeRF [47] introduces ellipsoid representations within NeRFs, its reliance on multiple implicit fields results in inefficient 3D decomposition. ISCO [1] and DBW [36] use 3D superquadrics for shape decomposition, which enables more meaningful structural separation. However, their simple shape parameters struggle to capture complex geometries, leading to poor geometry and appearance reconstruction. DPA-Net [65] has advanced 3D shape abstraction from sparse views but generates redundant parts and struggles with realistic texture rendering. A concurrent work, GaussianBlock [24], employs SAM [27] to guide superquadric splitting and fusion for 3D decomposition, yet its computational efficiency remains limited, typically requiring several hours per scene. In contrast, our approach accomplishes self-supervised [51, 59] part-aware scene reconstruction through an efficient hybrid representation that simultaneously maintains geometric fidelity and photorealistic rendering quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 485, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 485, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 485, + 661 + ], + "type": "text", + "content": "2.2. Mesh-based Gaussian Splatting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Gaussian splitting (GS) [25] has been rapidly adopted in multiple fields due to its remarkable rendering capability. Several studies [7, 14, 20, 49] attempt to align Gaussians with mesh surfaces for easier editing and animation. Among" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9650" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 270 + ], + "type": "image", + "image_path": "1725736d806238fdb08ae4afbf2598d85df1ee5bf8813240447110681c446bb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 312 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. PartGS takes multi-view images to learn a parametric hybrid representation of superquadrics and 2D Gaussians. It initializes from random superquadrics and is gradually optimized during training to obtain a block-level reconstruction. Then, we free the constraints of Gaussians to model detailed geometry, achieving point-level reconstruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 327, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 542 + ], + "type": "text", + "content": "them, SuGaR [20] uses flat 3D Gaussians to enforce the alignment with the scene surface during optimization, minimizing the difference between the signed distance functions (SDF) of the desired Gaussians and actual Gaussians. GaMeS [49] introduces a hybrid representation of Gaussians and mesh, where Gaussians are attached to triangular facets of the mesh. Similarly, Gao et al. [14] proposes a mesh-based GS to achieve large-scale deformation effects on objects. Recently, 2DGS [21] proposes 2D Gaussians for surface modeling and significantly enhances the geometric quality. Inspired by them, we propose a representation that combines 2D Gaussians with superquadrics. A key distinction between our method and previous mesh-based GS approaches is that these methods require first reconstructing the scene's mesh and then binding Gaussians to the mesh surface, which results in a non-continuous mesh. In contrast, our method directly optimizes the mesh through a rendering loss, enabling part-aware mesh reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 111, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 111, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 111, + 571 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "Given a set of calibrated multi-view images and foreground masks, we aim to learn part-aware 3D representations, a meaningful decomposition of both geometry and appearance. Our approach adopts a two-stage optimization strategy: first, decomposing the object into basic shapes using a mixture of Gaussians and superquadrics at the block-level, followed by refining the decomposition at the point-level to achieve precise geometry. In Sec 3.1, we parameterize the hybrid representation. Sec. 3.2 then elaborates on leveraging this representation, enhanced by novel regularizers and an adaptive control strategy, to achieve self-supervised block-level" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 351 + ], + "type": "text", + "content": "decomposition. Finally, Sec. 3.3 presents the process for obtaining detailed part-aware results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 357, + 533, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 533, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 533, + 369 + ], + "type": "text", + "content": "3.1. Parametrizing the Hybrid Representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 373, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 373, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 312, + 373, + 555, + 481 + ], + "type": "text", + "content": "As shown in Fig. 2, to leverage the strengths of both, we attach Gaussians to the surface of superquadric meshes. This representation retains the superquadric's ability to parse a 3D scene into distinct parts. Meanwhile, spherical harmonics of Gaussians enable complex texture rendering via Gaussian splatting, addressing the texture learning limitations of prior work [32, 36, 41, 64]. Sharing pose parameters between superquadrics and 2D Gaussians further improves the representation's efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": "The parametric representation is controlled by both primitive and Gaussian parameters, which are optimized simultaneously. Given a 3D scene " + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": ", the proposed method decomposes it into multiple hybrid blocks, each consisting of a superquadric with associated Gaussians. Each scene is denoted by the hybrid representation: " + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "inline_equation", + "content": "S = \\mathcal{B}_1 \\cup \\ldots \\cup \\mathcal{B}_i \\cup \\mathcal{B}_M" + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": "-th hybrid block, and " + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 482, + 556, + 613 + ], + "type": "text", + "content": " is the total number of blocks. Blocks are defined using manually designed parameters that control pose, opacity, scale, shape, and texture. These parameters are optimized via differentiable rendering to parse the target scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "text", + "content": "Shape and Scale Parameters. For each hybrid block " + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "text", + "content": ", its geometry is controlled by superquadric parameters [2]. Specifically, there are two shape parameters " + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "inline_equation", + "content": "\\epsilon_1, \\epsilon_2" + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "text", + "content": " that define its shape, along with three scale parameters " + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "inline_equation", + "content": "s_1, s_2, s_3" + }, + { + "bbox": [ + 313, + 613, + 556, + 685 + ], + "type": "text", + "content": ", which scale the 3D axes. Analogous to icosphere vertex positioning, superquadric vertex coordinates are computed by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 687, + 555, + 699 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 687, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 322, + 687, + 555, + 699 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\left[ s _ {1} \\cos^ {\\epsilon 1} (\\theta) \\cos^ {\\epsilon 2} (\\varphi); s _ {2} \\sin^ {\\epsilon 1} (\\theta); s _ {3} \\cos^ {\\epsilon 1} (\\theta) \\sin^ {\\epsilon 2} (\\varphi) \\right], \\quad (1)", + "image_path": "5e533662ba077f0a8c0269fbe13f35685c79a245b980e754f17bda856b78dec3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 701, + 555, + 713 + ], + "type": "text", + "content": " represent the azimuth and elevation defined" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9651" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "in the spherical coordinate. The shape and scale parameters govern block deformation to learn part-aware geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": "Pose Parameters. The pose of the " + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": "-th hybrid block is defined by its rotation " + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i" + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": ". The vertex position from Eq. 1 is transformed from the local coordinate to world space as: " + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_i^j = \\mathbf{R}_i\\mathbf{v}_i^j + \\mathbf{t}_i" + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 96, + 295, + 227 + ], + "type": "text", + "content": " indexes the vertices. Previous approaches, including SuGaR [20] and GaMeS [49], position Gaussians on reconstructed meshes with independent pose and shape parameters to enhance appearance modeling. In contrast, our method employs Gaussians to construct a differentiable rendering that bridges images and superquadrics. To achieve this, we couple the parameters of Gaussians with superquadrics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": "For a posed superquadric, Gaussian centers are uniformly sampled on triangular faces, with their poses determined by face vertices. Following GaMeS [49], each Gaussian's rotation matrix " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\mathrm{R}_v = [r_1,r_2,r_3]" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " and scaling " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\mathrm{S}_v" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " are computed from vertex positions. Given a triangular face " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " with vertices " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "v_{1},v_{2},v_{3}\\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": ", orthonormal vectors are constructed such that " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " aligns with the face normal, " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " points from the centroid " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "m = \\mathrm{mean}(v_1,v_2,v_3)" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "r_3" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " is obtained by orthonormalizing the vector from the center to the second vertex concerning the existing " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 55, + 228, + 296, + 347 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 356, + 296, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 356, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 356, + 296, + 381 + ], + "type": "interline_equation", + "content": "r _ {3} = \\frac {\\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2}\\right)}{\\left\\| \\operatorname {o r t} \\left(v _ {2} - m ; r _ {1} , r _ {2} \\right. \\right\\|}, \\tag {2}", + "image_path": "02ac56c70fba462f230e602bba87e55c0a54c6493ae776696290858feff38e02.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 385, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\text{ort}" + }, + { + "bbox": [ + 55, + 385, + 296, + 409 + ], + "type": "text", + "content": " represents the orthonormalization in Gram-Schmidt [4]. For the scaling of 2D Gaussians, we use:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 415, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 415, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 135, + 415, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathrm {S} _ {v} = \\operatorname {d i a g} \\left(s _ {v} ^ {2}, s _ {v} ^ {3}\\right), \\tag {3}", + "image_path": "c6866e2480006da2c524a4422e031c132ab698a719849127034704733b40b3aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "s_v^2 = c||m - v_1||" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "s_v^3 = c < v_2, r_3 >" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": " is a size control hyperparameter. We place a fixed number of Gaussians on each triangular face. These designs eliminate the need to learn the geometry of Gaussians, thereby enhancing the efficiency of the representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": "Opacity Parameters. We define the total number of hybrid blocks as " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": ". However, a typical scene does not contain exactly " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " blocks. Therefore, only the meaningful blocks are retained. To achieve this, we introduce a learnable parameter " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " to represent each block's opacity. During the optimization, only blocks with " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " greater than a certain threshold are retained. Note that Gaussians within the same block share the same " + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": " for opacity in rasterization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 642 + ], + "type": "text", + "content": "Texture Parameters. The texture is modeled using 2D Gaussians positioned on the surface of the superquadrics. Spherical harmonics of each Gaussian control the texture and are optimized for rendering view-dependent images [25]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 201, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 201, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 201, + 661 + ], + "type": "text", + "content": "3.2. Block-level Decomposition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "This section describes the optimization of the hybrid representation. We observed that minimizing the rendering loss across multi-view images alone led to instability in positioning hybrid blocks. Therefore, several regularization terms" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "are introduced to optimize the composition for maximal image formation consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 102, + 399, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 399, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 399, + 114 + ], + "type": "text", + "content": "3.2.1. Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 117, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 555, + 200 + ], + "type": "text", + "content": "In addition to the image rendering loss, we employ regularizers consisting of coverage, overlap, parsimony, and opacity entropy. The coverage loss encourages hybrid blocks to cover only meaningful regions; the overlap loss prevents blocks from overlapping; the parsimony loss regularizes the number of existing blocks; and the opacity entropy encourages block opacities close to binary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "text", + "content": "Rendering Loss. The rendering loss is based on 3DGS [25], where it combines an " + }, + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 201, + 555, + 224 + ], + "type": "text", + "content": " term with a D-SSIM term:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 236, + 555, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 236, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 369, + 236, + 555, + 249 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n}} = (1 - \\lambda) L _ {1} + \\lambda L _ {\\mathrm {D} - \\mathrm {S S I M}}. \\tag {4}", + "image_path": "797472ae8d55dcf2e7cf92ffc5bdcc3085fdc9270d494f389859a245b291017d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": "**Coverage Loss.** The coverage loss ensures that the block set " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_1 \\ldots \\mathcal{B}_M\\}" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " covers the object while preventing it from extending beyond its boundaries. To determine the 3D occupancy field based on the blocks, we first define the approximate signed distance of a 3D point " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": "-th block, " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "D_i(x) = \\Psi_i(x) - 1" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\Psi_i" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " denotes the superquadric inside-outside function [2] associated with the block " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "D_i(x) \\leq 0" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " if the point " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " lies inside the " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": "-th block, and " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "D_i(x) > 0" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " is outside the block. Inspired by NeRF [34], we sample a ray set " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " using ray-casting based on camera poses. Given the object mask, we associate each ray " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "r \\in \\mathcal{R}" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " with a binary label: " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "l_r = 1" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " if the ray " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": " is inside mask, otherwise " + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "inline_equation", + "content": "l_r = 0" + }, + { + "bbox": [ + 313, + 254, + 555, + 410 + ], + "type": "text", + "content": ". The coverage loss is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 417, + 555, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 417, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 326, + 417, + 555, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {R}) = \\sum_ {r \\in \\mathcal {R}} l _ {r} L _ {\\text {c r o s s}} (r) + \\left(1 - l _ {r}\\right) L _ {\\text {n o n - c r o s s}} (r). \\tag {5}", + "image_path": "ae32758f9ff01e55300bda4ec820bdb90557a365965c0a5f3bea95ad0140c865.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cross}}" + }, + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "text", + "content": " encourages rays to intersect with blocks, while " + }, + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{non - cross}}" + }, + { + "bbox": [ + 313, + 450, + 554, + 475 + ], + "type": "text", + "content": " discourages rays from intersecting with blocks:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 351, + 483, + 555, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 351, + 483, + 555, + 506 + ], + "type": "interline_equation", + "content": "L _ {\\text {c r o s s}} (r) = \\operatorname {R e L U} \\left(\\min _ {i \\in \\mathcal {M}} \\min _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {6}", + "image_path": "3ed4f38c91af8f0ae9baf91b663ccd465f4ca7c1db9d3a65370b907e66949065.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 341, + 511, + 555, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 511, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 341, + 511, + 555, + 535 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o n - c r o s s}} (r) = \\operatorname {R e L U} \\left(\\max _ {i \\in \\mathcal {M}} \\max _ {x _ {j} ^ {r} \\in \\mathcal {X} _ {r}} - D _ {i} \\left(x _ {j} ^ {r}\\right)\\right), \\tag {7}", + "image_path": "ff3074de881777e36c5efdf8a287063d93cbc227c6c2b8714d9132cfcdacd2b4.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "x_{j}^{r}" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": "-th sampled point along the ray " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{1, \\dots, M\\}" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": ". Intuitively, this implies that at least one sampled points " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{r}" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": " along the ray " + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 540, + 554, + 599 + ], + "type": "text", + "content": " inside the mask that lies within a certain block, while all points along the ray outside the mask should not belong to any block." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 600, + 555, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 683 + ], + "type": "text", + "content": "Overlap Loss. We introduce a regularization term to prevent overlap between individual blocks. Given the difficulty of directly calculating block overlap, we adopt a Monte Carlo method similar to [36]. Specifically, we sample multiple 3D points in space and penalize those that lie inside more than " + }, + { + "bbox": [ + 313, + 600, + 555, + 683 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 600, + 555, + 683 + ], + "type": "text", + "content": " blocks. Based on the superquadric inside-outside function, a soft occupancy function is defined as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 375, + 691, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 691, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 375, + 691, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {i} ^ {x} = \\tau_ {i} \\operatorname {s i g m o i d} \\left(\\frac {- D _ {i} (x)}{\\gamma}\\right), \\tag {8}", + "image_path": "72fa171e275f9caab210694c431b6da339b3655680081fe176c9201a61d84090.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9652" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is a temperature parameter. Thus, for a set of " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " sampled 3D points " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", the overlap loss is expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 104, + 296, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 104, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 94, + 104, + 296, + 137 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o v e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\operatorname {R e L U} \\left(\\sum_ {i \\in \\mathcal {M}} \\mathcal {O} _ {i} ^ {x} - k\\right). \\tag {9}", + "image_path": "5eb6fd2c817aef597c8b9d316feb9fedf662da1b92564327ffbd8f1acd1fda16.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 194 + ], + "type": "text", + "content": "Parsimony Loss. To promote the use of the minimal number of blocks and achieve parsimony in decomposition, we introduce a regularization term that penalizes block opacity " + }, + { + "bbox": [ + 55, + 146, + 296, + 194 + ], + "type": "inline_equation", + "content": "(\\tau)" + }, + { + "bbox": [ + 55, + 146, + 296, + 194 + ], + "type": "text", + "content": ". This loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 201, + 296, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 296, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p a r}} = \\frac {1}{M} \\sum_ {i \\in \\mathcal {M}} \\sqrt {\\tau_ {i}}. \\tag {10}", + "image_path": "22bb8208f1565e715779071fe1f69c0873c7640b6fcba48f388314b4fd92b3ea.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 321 + ], + "type": "text", + "content": "Opacity Entropy Loss. During optimization, the opacity of the block inside the object region tends to approach 1, while the opacity of the block outside the object region tends to approach 0. To facilitate this, we associate the opacities of blocks with the labeled rays and apply a cross-entropy loss " + }, + { + "bbox": [ + 55, + 239, + 296, + 321 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 55, + 239, + 296, + 321 + ], + "type": "text", + "content": " between the block opacity and the mask labels, defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 329, + 296, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 329, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 77, + 329, + 296, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o p a}} (\\mathcal {R}) = \\frac {1}{| \\mathcal {R} |} \\sum_ {r \\in \\mathcal {R}} L _ {c e} \\left(\\max _ {i \\in \\mathcal {M}} \\tau_ {i} \\left(x ^ {r}\\right), l _ {r}\\right). \\tag {11}", + "image_path": "20c03c6f5751f9a3d07b5b6c5a83f328cd3d2d63ae16f5a4e17a55066de3bd0d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "text", + "content": "Here, only points " + }, + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "inline_equation", + "content": "x^r" + }, + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "text", + "content": " inside the blocks are sampled, e.g., " + }, + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "inline_equation", + "content": "\\{x^r \\in \\mathcal{X}_r \\mid D_i(x^r) \\leq 0\\}" + }, + { + "bbox": [ + 55, + 367, + 296, + 391 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 391, + 296, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 414 + ], + "type": "text", + "content": "The total loss is the weighted sum of the loss terms described above:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 424, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 424, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 61, + 424, + 295, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e n}} + \\lambda_ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + \\lambda_ {\\text {o v e r}} \\mathcal {L} _ {\\text {o v e r}} + \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} + \\lambda_ {\\text {o p a}} \\mathcal {L} _ {\\text {o p a}}. \\tag {12}", + "image_path": "f0378ed4c621b50c9e63e21b64c916e253d13ce265595e015f3d5416d50e8250.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 446, + 204, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 204, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 204, + 458 + ], + "type": "text", + "content": "3.2.2. Adaptive Number of Blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": "Given that the number of parts may vary across different scenes, we allow the number of blocks to adjust during optimization adaptively. Specifically, when the opacity of an existing block falls below a threshold " + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 461, + 296, + 616 + ], + "type": "text", + "content": ", this block is removed immediately. Furthermore, we explore a block-adding mechanism to dynamically incorporate new components, ensuring comprehensive coverage of target objects. Specifically, DBSCAN [44] is employed to cluster initial point clouds that are not covered by any blocks, where the point clouds are either randomly initialized or derived as by-products from COLMAP [43]. For each point cloud cluster, a new block is introduced at its center, with the remaining parameters of the block initialized randomly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": "The final block-level reconstruction generates a textured 3D scene represented by multiple superquadrics, with surface details rendered through Gaussians, as illustrated in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 200, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 200, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 200, + 685 + ], + "type": "text", + "content": "3.3. Point-level Decomposition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The primitive-based hybrid representation effectively decomposes the shape into parts but exhibits suboptimally for" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "complex objects. To address this limitation, we perform a refinement stage that enhances geometric fitting. In this stage, the constraint between Gaussians and superquadrics is decoupled, enabling independent Gaussian optimization. Furthermore, to prevent Gaussian components from one block passing through adjacent blocks and disrupting the continuity and plausibility of the part decomposition, we impose a new constraint to minimize the negative signed distance of each Gaussian entering other blocks:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 336, + 184, + 555, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 184, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 336, + 184, + 555, + 214 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e n t e r}} = \\frac {1}{N} \\sum_ {x \\in \\Omega} \\sum_ {m \\in \\mathcal {M} \\backslash \\{\\delta \\}} \\operatorname {R e L U} (- D _ {m} (x)), \\tag {13}", + "image_path": "af4d41041906d9c5319f1a6d3f1222a1f3de205664ff09bf0989bd0e7b7a7fac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "content": " is the sampled Gaussian set, and " + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "content": " represents the identifier of the block to which the Gaussian " + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 219, + 554, + 267 + ], + "type": "text", + "content": " belongs. As seen in Fig. 2, the reconstruction becomes more accurate and aligns better with the target shape." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 277, + 395, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 395, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 395, + 289 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 296, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 344 + ], + "type": "text", + "content": "We comprehensively evaluate our approach across four key aspects: 3D reconstruction quality, view synthesis performance, shape parsimony, and computational efficiency, comparing it with state-of-the-art methods in the field." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 350, + 376, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 376, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 376, + 361 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 367, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 511 + ], + "type": "text", + "content": "Extensive experiments are conducted on two widely-used public datasets: DTU [23] and ShapeNet [6]. Specifically, 15 standard scenes from the DTU dataset that are widely adopted for reconstruction quality assessment are used for evaluation. Each scene contains either 49 or 64 images. Additionally, we construct a ShapeNet subset consisting of four categories: Chair, Table, Gun, and Airplane, with 15 objects selected per category. For each object, 50 rendered images are generated for training. We further validate real-world applicability on both the BlendedMVS dataset [61] and self-captured scenes, demonstrating robust performance across synthetic and natural environments." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 517, + 447, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 447, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 447, + 529 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": "We build upon the 2DGS [21] architecture and add a nonlearnable part attribute to each Gaussian component to enforce the constraint in Eq. 13. In the ShapeNet benchmark, we observe a common degradation in mesh reconstruction as shown in Fig. 7. To address this, we propose Gaussian scale regularization, as validated in Sec. 4.4. We set the initial number of primitives " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " to 8. The temperature parameter " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " in Eq. 8 is 0.005 and the overlapping number " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " in Eq. 9 is 1.95. We train the hybrid representation for 30k iterations, followed by refinement optimization for another 30k iterations. During block-level optimization, the block-adding operation is carried out at the 5k-th and 10k-th iterations. The loss weights " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{cov}}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{over}}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{par}}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{opa}}" + }, + { + "bbox": [ + 313, + 534, + 556, + 714 + ], + "type": "text", + "content": " are set to 10, 1, 0.002, and 0.01, respectively. For more details about datasets, metrics, and baselines, please refer to Supplementary." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9653" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 315 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 315 + ], + "type": "image", + "image_path": "9a2893f0459bd4436057f44847546ed02f6e3881a4ec7cc693ace7254912d196.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 325, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 325, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 555, + 348 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on DTU [23] and ShapeNet [23]. The first two rows are DTU examples, and the last two are ShapeNet examples, respectively. Our method is the only one that provides reasonable 3D part decomposition while capturing detailed geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 361, + 553, + 458 + ], + "blocks": [ + { + "bbox": [ + 56, + 361, + 553, + 458 + ], + "lines": [ + { + "bbox": [ + 56, + 361, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 553, + 458 + ], + "type": "table", + "html": "
MethodInputRenderableChamfer distance per sceneMeanMean
S24S37S40S55S63S65S69S83S97S105S106S110S114S118S122CD#P
EMS [32]3D GTX6.323.542.994.304.164.013.753.244.974.344.167.627.584.464.034.657.7
MBF [41]3D GTX3.122.663.842.541.592.112.192.012.322.452.172.123.832.022.552.5034.1
EMS [32] + Neus [54]ImageX5.995.564.434.325.426.143.753.964.634.345.885.114.294.833.534.978.87
MBF [41] + Neus [54]ImageX2.693.373.222.693.632.602.593.132.852.512.453.722.242.492.522.8546.7
PartNeRF [47]Image✓9.3810.469.088.636.047.257.229.158.7210.016.729.857.858.689.218.548.0
DBW [36]Image✓5.418.351.573.083.404.157.463.946.634.854.384.656.294.343.044.764.8
Ours (Block-level)Image✓5.684.911.852.613.754.663.757.574.274.383.494.483.614.213.704.195.9
Ours (Point-level)Image✓0.701.170.550.651.061.231.101.361.370.780.921.410.691.050.710.985.9
", + "image_path": "8e5f3de70384992d0ad68fdd1fc6188968ad3b4bf69304e88bbf5d51915da125.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 501, + 294, + 563 + ], + "blocks": [ + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on DTU [23]. The Chamfer distance between the 3D reconstruction and the ground-truth is reported in 15 scenes. The best results are bolded, and the average numbers of primitives found (#P) that are greater than 10 are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 501, + 294, + 563 + ], + "lines": [ + { + "bbox": [ + 57, + 501, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 294, + 563 + ], + "type": "table", + "html": "
MethodPart-awareCD ↓PSNR ↑SSIM ↑LIPPS ↓Time ↓
Neuralangelo [30]X0.6133.84--> 10 h
2DGS [21]X0.8134.070.990.019~ 10 m
PartNeRF [47]✓9.5917.970.770.246~ 8 h
DBW [36]✓4.7316.440.750.201~ 2 h
Ours (Block-level)✓4.1919.840.820.189~ 30 m
Ours (Point-level)✓0.9835.040.990.015~ 40 m
", + "image_path": "0a458dcd15aa25583c5f3d0ea2026d6871595758658bcfb85d33ef84a9c85d01.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 571, + 297, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 616 + ], + "type": "text", + "content": "Table 2. Quantitative results on DTU [23]. Our method outperforms all part-aware approaches in image synthesis quality, reconstruction accuracy, and efficiency. Neuralangelo's results are from the original paper, with all times measured on an RTX 3090 GPU." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 134, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 134, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 134, + 641 + ], + "type": "text", + "content": "4.3. Evaluations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 649, + 218, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 218, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 218, + 661 + ], + "type": "text", + "content": "4.3.1. DTU and Shapenet Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "content": "Geometry Reconstruction. In Tab. 1 and Tab. 3, we compare our geometry reconstruction to SOTA shape decomposition methods on Chamfer distance and training time using DTU and Shapenet dataset. The Chamfer distance" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 501, + 554, + 571 + ], + "blocks": [ + { + "bbox": [ + 315, + 501, + 554, + 571 + ], + "lines": [ + { + "bbox": [ + 315, + 501, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 315, + 501, + 554, + 571 + ], + "type": "table", + "html": "
MethodInputChamfer Distance ↓Primitives (#P)MeanMean
AirplaneTableChairGunAirplaneTableChairGunCD#P
EMS [32]3D GT3.406.9219.02.029.47.8810.38.47.849.0
MBF [41]3D GT2.832.181.592.3210.8513.913.414.32.2113.1
PartNeRF [47]Image2.292.772.302.468.08.08.08.02.468.0
DBW [36]Image3.617.336.192.092.75.23.63.34.813.7
Ours (Block-level)Image2.472.152.321.783.96.67.65.02.185.8
Ours (Point-level)Image1.291.720.941.073.96.67.65.01.255.8
", + "image_path": "a74c8f1cb64e244e76ba97da6521ae10573bb16f939493226b77961fd2ab850d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 578, + 555, + 611 + ], + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 611 + ], + "type": "text", + "content": "Table 3. Quantitative comparison on ShapeNet [6]. We report Chamfer distance and the number of parts. The best results are bolded, and the second-best results are underlined." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "metrics reported for ShapeNet are scaled by a factor of 100 for readability. Our method consistently outperforms in all scenes compared to prior works. As shown in Fig. 3, our approach consistently produces interpretable 3D decompositions with further refinement achieving more detailed geometry (the last two columns). MBF [41] achieves a low CD error but at the cost of using significantly more prim" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9654" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 74, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 553, + 140 + ], + "type": "table", + "html": "
MethodPart-awareChamfer Distance ↓PSNR ↑SSIM ↑LIPPS ↓
AirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGunAirplaneTableChairGun
2DGS [21]✗1.472.370.501.0340.8939.8039.0541.740.9940.9900.9900.9940.0090.0270.0170.009
PartNeRF [47]✓2.292.772.302.4619.6320.6619.0821.970.8980.8550.8750.9160.0860.1610.1360.083
DBW [36]✓3.617.336.192.0926.1123.8420.2528.720.9500.9150.8920.9600.0740.1360.1320.042
Ours (Block-level)✓2.472.152.321.7827.9427.9824.9229.950.9590.9250.9060.9630.0720.1290.1220.047
Ours (Point-level)✓1.291.720.941.0741.1836.8036.0739.510.9920.9730.9770.9890.0140.0700.0380.021
", + "image_path": "c0ed874782a95993fdfe8b944acb6d17edbc29d67c181e5cee8edef5ddc27757.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 173, + 141, + 316 + ], + "blocks": [ + { + "bbox": [ + 56, + 173, + 141, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 173, + 141, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 141, + 316 + ], + "type": "image", + "image_path": "24ae7e5a67815b404bd07fee031db8fba68f96091208d7f64af3ec24f97d31af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 297, + 369 + ], + "type": "text", + "content": "Figure 4. Qualitative results on BlendedMVS [61] and self-captured data. We present RGB renderings and decomposed parts from novel views. The top examples are from the BlendedMVS dataset, and the last example is from our captured scenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 150, + 173, + 190, + 316 + ], + "blocks": [ + { + "bbox": [ + 150, + 173, + 190, + 316 + ], + "lines": [ + { + "bbox": [ + 150, + 173, + 190, + 316 + ], + "spans": [ + { + "bbox": [ + 150, + 173, + 190, + 316 + ], + "type": "image", + "image_path": "c926f3ba22a56d86a32c5e06a2a597aa03f4573cc5f5d110f7785cc401a7d7e3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 199, + 173, + 237, + 308 + ], + "blocks": [ + { + "bbox": [ + 199, + 173, + 237, + 308 + ], + "lines": [ + { + "bbox": [ + 199, + 173, + 237, + 308 + ], + "spans": [ + { + "bbox": [ + 199, + 173, + 237, + 308 + ], + "type": "image", + "image_path": "f27da3159df474a17048759326d7e7f2cf043bbb41b0910ae41fe2687b249422.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 248, + 173, + 288, + 308 + ], + "blocks": [ + { + "bbox": [ + 248, + 173, + 288, + 308 + ], + "lines": [ + { + "bbox": [ + 248, + 173, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 248, + 173, + 288, + 308 + ], + "type": "image", + "image_path": "0a9fecf992c3b20d510816d220905d3c02c3f955ff86d2a7b933ec78efbd4807.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 304, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 348 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with SAM-based methods. Visual results demonstrate that our method produces more structurally coherent decompositions, whereas SAM-based approaches frequently exhibit spatial discontinuities." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 384, + 293, + 513 + ], + "blocks": [ + { + "bbox": [ + 56, + 384, + 293, + 513 + ], + "lines": [ + { + "bbox": [ + 56, + 384, + 293, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 293, + 513 + ], + "type": "image", + "image_path": "0cf1c6b0b7ee14a1103a679235523a6b7e3143fe7142ee49d274421f4a041d4a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 295, + 564 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 564 + ], + "type": "text", + "content": "Figure 5. Ablation studies on key strategies. The block-level visual comparisons illustrate the impact of adopting our proposed strategy. The first row shows results without the strategy, and the second with the strategy implemented." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "text", + "content": "ities, leading to over-decomposition and the inclusion of meaningless parts. Moreover, as shown in Tab. 2 and Tab. 4, our approach achieves competitive results with the advanced Gaussian-based 2DGS [21] and surface reconstruction Neuralangelo [30], which are limited to producing unstructured meshes. Notably, our model demonstrates excellent efficiency, with a reconstruction speed approximately " + }, + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "inline_equation", + "content": "10\\mathrm{X}" + }, + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "text", + "content": " faster than PartNeRF [47] and over " + }, + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "inline_equation", + "content": "3\\mathrm{X}" + }, + { + "bbox": [ + 55, + 580, + 296, + 676 + ], + "type": "text", + "content": " faster than DBW [36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Appearance Reconstruction. In addition to part decomposition, our method enables high-fidelity image synthesis, attributed to integrating Gaussian splatting within hybrid" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 173, + 361, + 293 + ], + "blocks": [ + { + "bbox": [ + 315, + 173, + 361, + 293 + ], + "lines": [ + { + "bbox": [ + 315, + 173, + 361, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 361, + 293 + ], + "type": "image", + "image_path": "307d665fea83eea93e22db493cd3d7d1a8350fa763b31b558dd767118b793551.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 362, + 173, + 434, + 294 + ], + "blocks": [ + { + "bbox": [ + 362, + 173, + 434, + 294 + ], + "lines": [ + { + "bbox": [ + 362, + 173, + 434, + 294 + ], + "spans": [ + { + "bbox": [ + 362, + 173, + 434, + 294 + ], + "type": "image", + "image_path": "70407c2d856a6e1c769162ddf9593c72e1736e682f1fbac84e02effa1cfda752.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 173, + 504, + 294 + ], + "blocks": [ + { + "bbox": [ + 434, + 173, + 504, + 294 + ], + "lines": [ + { + "bbox": [ + 434, + 173, + 504, + 294 + ], + "spans": [ + { + "bbox": [ + 434, + 173, + 504, + 294 + ], + "type": "image", + "image_path": "10cdb9e196463b04c8405d36179586a3b92d152ed2795781a4014c35037727cf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 504, + 173, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 504, + 173, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 504, + 173, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 504, + 173, + 553, + 294 + ], + "type": "image", + "image_path": "4111f7b72c108581367c01f4837d509695d917e6f151090dc0b0072f30e0d65e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 316, + 360, + 553, + 423 + ], + "blocks": [ + { + "bbox": [ + 58, + 148, + 550, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 148, + 550, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 148, + 550, + 159 + ], + "type": "text", + "content": "Table 4. Quantitative results on ShapeNet [6]. We report the Chamfer distance and novel view synthesis results across four categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 360, + 553, + 423 + ], + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 423 + ], + "type": "table", + "html": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
Complete model2.3224.920.9060.1220.9436.070.9770.0386.6
w/o Lover2.1125.170.9140.1181.3436.040.9770.0377.6
w/o Lopa5.7221.120.8530.1741.0635.600.9750.0403.8
w/o Lcov3.2122.960.8960.1341.3436.000.9760.0388.53
w/o Adaptive3.1622.740.8800.1411.0535.740.9740.0406.7
w/o Lpar1.9125.480.9180.1150.9336.180.9770.03710.1
", + "image_path": "ec3d9a1a29ac736dfa20c8af29ad64edcb2050e51c874a82c80940963d1f7e11.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 454 + ], + "type": "text", + "content": "Table 5. Ablation studies on the ShapeNet [6]. We report Chamfer Distance, rendering metrics, and the number of primitives (#P)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 476, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 547 + ], + "type": "text", + "content": "representations. EMS [32] and MBF [41] operate directly on point clouds and consequently lack image rendering capabilities, while PartNeRF [47] and DBW [36] yield low-quality view synthesis. In contrast, our method achieves high-quality appearance rendering results, as demonstrated in Tab. 2 and Tab. 4." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 555, + 402, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 402, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 402, + 566 + ], + "type": "text", + "content": "4.3.2. Real-life Data" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 655 + ], + "type": "text", + "content": "To further demonstrate the applicability of our method for learning shape decomposition, we test our model on the real-life images from the BlendedMVS dataset [61] and a self-captured dataset. As shown in Fig. 4, our approach can robustly produce both realistic appearances and reasonable 3D decompositions across a variety of data types. More results are provided in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 662, + 491, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 491, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 491, + 673 + ], + "type": "text", + "content": "4.3.3. Compared to SAM-based Methods" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We also conduct comparisons to SAM-based [27] methods for this task, as shown in Fig. 6. Despite impressive advances in 3D segmentation and editing [5, 26, 62, 63] by" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9655" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 294, + 126 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 294, + 126 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 294, + 126 + ], + "type": "image", + "image_path": "c82033cd52c3d624329aede2b26145e5a5fa59c98af17cea26b89603aa1ec4e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "type": "text", + "content": "Figure 7. Impact of Gaussian scale regularization (SR). The degraded mesh produced by 2DGS [21] is effectively improved." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 167, + 124, + 227 + ], + "blocks": [ + { + "bbox": [ + 56, + 167, + 124, + 227 + ], + "lines": [ + { + "bbox": [ + 56, + 167, + 124, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 124, + 227 + ], + "type": "image", + "image_path": "143f6800a56837a75210e2b7e2342ddd66ce215d4f93d1ed9af327ee18b64fad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "type": "text", + "content": "Figure 8. Impact of the initial number of primitives " + }, + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "type": "text", + "content": ". Increasing " + }, + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 236, + 296, + 269 + ], + "type": "text", + "content": " yields a finer-grained decomposition, while decreasing it produces a coarser decomposition." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 126, + 167, + 178, + 223 + ], + "blocks": [ + { + "bbox": [ + 126, + 167, + 178, + 223 + ], + "lines": [ + { + "bbox": [ + 126, + 167, + 178, + 223 + ], + "spans": [ + { + "bbox": [ + 126, + 167, + 178, + 223 + ], + "type": "image", + "image_path": "4fe97c9d7eee1036ca983cb6d1ff5a65d638c7a2d84c52711fd4a9c3ebf044ab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 179, + 167, + 233, + 223 + ], + "blocks": [ + { + "bbox": [ + 179, + 167, + 233, + 223 + ], + "lines": [ + { + "bbox": [ + 179, + 167, + 233, + 223 + ], + "spans": [ + { + "bbox": [ + 179, + 167, + 233, + 223 + ], + "type": "image", + "image_path": "b9e5052eafdcbb60ada0535cf39aab79f0c65599843c8f28937ab275a24121cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 167, + 290, + 225 + ], + "blocks": [ + { + "bbox": [ + 239, + 167, + 290, + 225 + ], + "lines": [ + { + "bbox": [ + 239, + 167, + 290, + 225 + ], + "spans": [ + { + "bbox": [ + 239, + 167, + 290, + 225 + ], + "type": "image", + "image_path": "37e420abf3cb308307587cda0307fd45dcd6cef15b6656db7d4300f0038b3a18.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 274, + 296, + 395 + ], + "blocks": [ + { + "bbox": [ + 57, + 274, + 296, + 395 + ], + "lines": [ + { + "bbox": [ + 57, + 274, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 57, + 274, + 296, + 395 + ], + "type": "image", + "image_path": "2fcb7e24646fc7aec39c0b4580b93c9fd57fc7ce80c30d2386654825a3e18cd5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 404, + 296, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 449 + ], + "type": "text", + "content": "Figure 9. Applications. Part-Aware Editing (top): After optimization, we can easily edit the scene by adding, scaling, or moving specific parts. 3D Content Generation (bottom): By combining parts of different objects, we can create new 3D content." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 57, + 459, + 294, + 529 + ], + "blocks": [ + { + "bbox": [ + 57, + 459, + 294, + 529 + ], + "lines": [ + { + "bbox": [ + 57, + 459, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 294, + 529 + ], + "type": "table", + "html": "
MethodBlock-levelPoint-level#P
CD ↓PSNR ↑SSIM ↑LPIPS ↓CD ↓PSNR ↑SSIM ↑LPIPS ↓
λpar=0.16.4616.100.7210.2551.0834.300.9870.0182.7
λpar=0.01 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
λpar=0.0014.0120.180.8350.1801.0335.040.9880.0157.1
M=44.9719.060.7940.2131.0634.930.9870.0165.0
M=8 (*)4.1919.840.8200.1891.0535.040.9880.0155.9
M=163.9920.870.830.1761.0735.050.9880.0158.3
", + "image_path": "39608ccb0875c5ec68a808e27c2807976861aa0c9c971369efe22f9aae7c3dba.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "type": "text", + "content": "Table 6. Effect of parsimony weight " + }, + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{par}})" + }, + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "type": "text", + "content": " and initial primitives count " + }, + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 55, + 537, + 295, + 571 + ], + "type": "text", + "content": " on DTU [23]. We report Chamfer distance, rendering metrics, and primitives count (#P). * denotes the default setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "text", + "content": "distilling 2D information, achieving semantic disentanglement but remains challenging due to indistinct textures and severe cross-view 2D inconsistencies. Lifting 2D segmentation features to 3D segmentation will also lead to feature misalignment, leading to visual artifacts. In contrast, our method performs direct 3D space segmentation, yielding more reasonable and structurally coherent decompositions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 124, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 124, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 124, + 683 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In this section, we first analyze the design choices by isolating each strategy to assess its impact. We report the averaged" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": "quantitative performance on 15 instances of the chair category in Tab. 5 and provide visual comparisons in Fig. 5. Removing overlapping loss provides better reconstruction accuracy but more overlapping parts. The opacity loss drastically affects the number of primitives and doubles the CD, highlighting its role in controlling primitive presence and reconstruction quality. Coverage loss is crucial in reconstruction accuracy, ensuring primitives align correctly with the target. The adaptive primitive strategy fills in missing parts and improves the reconstruction metric. Without parsimony loss, CD and rendering metrics are optimal but result in over-decomposition. Note that the full model did not yield the highest accuracy because the part-aware reconstruction trades off the accuracy and rationality of the part decomposition. The objective is to achieve high accuracy while preserving reasonable parts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 263, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 556, + 346 + ], + "type": "text", + "content": "We also validate the effectiveness of Gaussian scale regularization in Fig. 7. 2DGS [21] tends to use large-scale Gaussians in texture-less areas and create holes in the constructed mesh by TSDF integration [67]. The presence of holes stems from the inability of large-scale Gaussians to maintain view-consistent depth. Scale regularization effectively addresses this issue by suppressing large-scale Gaussians." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": "Lastly, in Tab. 6, we analyze the influence of two key hyperparameters on decomposition granularity: the parsimony loss weight " + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{par}}" + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": " and the initial primitive count " + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": ". Stronger parsimony regularization reduces the number of primitives, while weaker regularization increases it. As " + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": " rises, both reconstruction and view synthesis performance slightly improve. This demonstrates that adjusting " + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{par}}" + }, + { + "bbox": [ + 313, + 347, + 556, + 456 + ], + "type": "text", + "content": " enables our method to effectively control the granularity of object or scene decomposition. Fig. 8 visually illustrates this impact." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 397, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 397, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 397, + 475 + ], + "type": "text", + "content": "4.5. Applications" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 479, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 564 + ], + "type": "text", + "content": "Fig. 9 illustrates two applications of our method that the original 3DGS and NeRF-based approaches do not support. Firstly, after optimization, we obtain the part decomposition, facilitating easy editing of specific objects or scene components, e.g., adding, moving, removing, or scaling. Secondly, by combining parts of different objects, our method enables the creation of new high-quality 3D content." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 392, + 586 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "We introduce PartGS, a hybrid representation of superquadrics and 2D Gaussians, to learn 3D scenes in a part-aware representation. Compared to prior works, the proposed method retains geometry details, supporting high-quality image rendering. It obtains state-of-the-art performance in comprehensive evaluations. One limitation is that it takes background-free scene images as inputs, leveraging segmentation tools such as SAM [27]. In the future, we aim to explore how to model backgrounds and extend to larger and more complex scenes." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9656" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 145 + ], + "type": "text", + "content": "This work is supported in part by the NSFC (62325211, 62132021, 62372457), the Major Program of Xiangjiang Laboratory (23XJ01009), Young Elite Scientists Sponsorship Program by CAST (2023QNRC001), the Natural Science Foundation of Hunan Province of China (2022RC1104)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 186, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 186, + 297, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 297, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 297, + 241 + ], + "type": "text", + "content": "[1] Stephan Alaniz, Massimiliano Mancini, and Zeynep Akata. Iterative superquadric decomposition of 3d objects from multiple views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18013-18023, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 275 + ], + "type": "text", + "content": "[2] Alan H. Barr. Superquadrics and Angle-Preserving Transformations. IEEE Computer Graphics and Applications, 1981. 1, 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 298 + ], + "type": "text", + "content": "[3] Thomas Binford. Visual Perception by Computer. In IEEE Conference on Systems and Control, 1971. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 300, + 295, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 295, + 321 + ], + "type": "text", + "content": "[4] Åke Björck. Numerics of gram-schmidt orthogonalization. Linear Algebra and Its Applications, 197:297-316, 1994. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 295, + 355 + ], + "type": "text", + "content": "[5] Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Segment any 3d gaussians. arXiv preprint arXiv:2312.00860, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "text", + "content": "[6] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. arXiv:1512.03012 [cs.CV], 2015. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 412, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 295, + 445 + ], + "type": "text", + "content": "[7] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "text", + "content": "[8] Zhiqin Chen, Andrea Tagliasacchi, and Hao Zhang. Bsp-net: Generating compact meshes via binary space partitioning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 523 + ], + "type": "text", + "content": "[9] Zhiqin Chen, Qimin Chen, Hang Zhou, and Hao Zhang. Dae-net: Deforming auto-encoder for fine-grained shape co-segmentation. arXiv preprint arXiv:2311.13125, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 578 + ], + "type": "text", + "content": "[10] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. Cvxnet: Learnable convex decomposition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "text", + "content": "[11] Martin A. Fischler and Robert C. Bolles. Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography. Communications of the ACM, 1981. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 690 + ], + "type": "text", + "content": "[12] Huachen Gao, Shihe Shen, Zhe Zhang, Kaiqiang Xiong, Rui Peng, Zhirui Gao, Qi Wang, Yugui Xie, and Ronggang Wang. Fdc-nerf: learning pose-free neural radiance fields with flow-depth consistency. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3615-3619. IEEE, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Lin Gao, Jie Yang, Tong Wu, Yu-Jie Yuan, Hongbo Fu, Yu-Kun Lai, and Hao Zhang. Sdm-net: Deep generative network" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "for structured deformable mesh. ACM Transactions on Graphics (TOG), 38(6):1-15, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[14] Lin Gao, Jie Yang, Bo-Tao Zhang, Jia-Mu Sun, Yu-Jie Yuan, Hongbo Fu, and Yu-Kun Lai. Mesh-based gaussian splatt-ting for real-time large-scale deformation. arXiv preprint arXiv:2402.04796, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 555, + 186 + ], + "type": "text", + "content": "[15] Zhirui Gao, Renjiao Yi, Zheng Qin, Yunfan Ye, Chenyang Zhu, and Kai Xu. Learning accurate template matching with differentiable coarse-to-fine correspondence refinement. Computational Visual Media, 10(2):309-330, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 189, + 555, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 189, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 316, + 189, + 555, + 221 + ], + "type": "text", + "content": "[16] Zhirui Gao, Renjiao Yi, Yaqiao Dai, Xuening Zhu, Wei Chen, Chenyang Zhu, and Kai Xu. Curve-aware gaussian splatting for 3d parametric curve reconstruction, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 224, + 555, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 555, + 267 + ], + "type": "text", + "content": "[17] Zhirui Gao, Renjiao Yi, Chenyang Zhu, Ke Zhuang, Wei Chen, and Kai Xu. Generic objects as pose probes for few-shot view synthesis. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 270, + 555, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 555, + 334 + ], + "type": "text", + "content": "[18] Zeqi Gu, Yin Cui, Zhaoshuo Li, Fangyin Wei, Yunhao Ge, Jinwei Gu, Ming-Yu Liu, Abe Davis, and Yifan Ding. Artiscene: Language-driven artistic 3d scene generation through image intermediary. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2891-2901, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 337, + 555, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 555, + 402 + ], + "type": "text", + "content": "[19] Yanran Guan, Han Liu, Kun Liu, Kangxue Yin, Ruizhen Hu, Oliver van Kaick, Yan Zhang, Ersin Yumer, Nathan Carr, Radomir Mech, et al. Fame: 3d shape generation via functionality-aware model evolution. IEEE Transactions on Visualization and Computer Graphics, 28(4):1758-1772, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 405, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 449 + ], + "type": "text", + "content": "[20] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. arXiv preprint arXiv:2311.12775, 2023. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 451, + 555, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 555, + 483 + ], + "type": "text", + "content": "[21] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. SIGGRAPH, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 486, + 555, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 555, + 540 + ], + "type": "text", + "content": "[22] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural template: Topology-aware reconstruction and disentangled generation of 3d meshes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18572-18582, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "type": "text", + "content": "[23] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanaes. Large Scale Multi-view Stereopsis Evaluation. In CVPR, 2014. 5, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 621 + ], + "type": "text", + "content": "[24] Shuyi Jiang, Qihao Zhao, Hossein Rahmani, De Wen Soh, Jun Liu, and Na Zhao. Gaussianblock: Building part-aware compositional and editable 3d scene by primitives and gaussians. arXiv preprint arXiv:2410.01535, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 666 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1-14, 2023. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[26] Chung Min* Kim, Mingxuan* Wu, Justin* Kerr, Matthew Tancik, Ken Goldberg, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9657" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "text", + "content": "[28] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics (TOG), 36(4):1-14, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 295, + 208 + ], + "type": "text", + "content": "[29] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised Fitting of Geometric Primitives to 3D Point Clouds. In CVPR, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 296, + 275 + ], + "type": "text", + "content": "[30] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 278, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 343 + ], + "type": "text", + "content": "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 346, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 346, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 346, + 295, + 379 + ], + "type": "text", + "content": "[32] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and Accurate Superquadric Recovery: a Probabilistic Approach. In CVPR, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 381, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 381, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 295, + 414 + ], + "type": "text", + "content": "[33] Romain Loiseau, Elliot Vincent, Mathieu Aubry, and Loic Landrieu. Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans. arXiv:2304.09704 [cs.CV], 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 459 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 296, + 505 + ], + "type": "text", + "content": "[35] Niloy Mitra, Michael Wand, Hao (Richard) Zhang, Daniel Cohen-Or, Vladimir Kim, and Qi-Xing Huang. Structure-aware shape processing. In SIGGRAPH Asia 2013 Courses, 2013. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 508, + 296, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 296, + 551 + ], + "type": "text", + "content": "[36] Tom Monnier, Jake Austin, Angjoo Kanazawa, Alexei A. Efros, and Mathieu Aubry. Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives. In NeurIPS, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 296, + 586 + ], + "type": "text", + "content": "[37] Chengjie Niu, Jun Li, and Kai Xu. Im2struct: Recovering 3d shape structure from a single rgb image. Cornell University - arXiv, Cornell University - arXiv, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 589, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 295, + 620 + ], + "type": "text", + "content": "[38] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics Revisited: Learning 3D Shape Parsing Beyond Cuboids. In CVPR, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "text", + "content": "[39] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning Unsupervised Hierarchical Part Decomposition of 3D Objects from a Single RGB Image. In CVPR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[40] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[41] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud. In ECCV, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "text", + "content": "[42] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 172 + ], + "type": "text", + "content": "[43] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 555, + 216 + ], + "type": "text", + "content": "[44] Erich Schubert, Jörg Sander, Martin Ester, Hans Peter Kriegel, and Xiaowei Xu. Dbscan revisited, revisited: why and how you should (still) use dbscan. ACM Transactions on Database Systems (TODS), 42(3):1-21, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "text", + "content": "[45] Qingyao Shuai, Chi Zhang, Kaizhi Yang, and Xuejin Chen. Dpf-net: combining explicit shape priors in deformable primitive field for unsupervised structural reconstruction of 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14321-14329, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "type": "text", + "content": "[46] Chunyu Sun, Yuqi Yang, Haoxiang Guo, Pengshuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Semi-supervised 3d shape segmentation with multilevel consistency and part substitution. Computational Visual Media, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 317, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 555, + 370 + ], + "type": "text", + "content": "[47] Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yannis Avrithis, and Leonidas Guibas. PartNeRF: Generating Part-Aware Editable 3D Shapes without 3D Supervision. In CVPR, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 403 + ], + "type": "text", + "content": "[48] Shubham Tulsiani, Hao Su, Leonidas J. Guibas, Alexei A. Efros, and Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. In CVPR, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 555, + 448 + ], + "type": "text", + "content": "[49] Joanna Waczyńska, Piotr Borycki, Sławomir Tadeja, Jacek Tabor, and Przemysław Spurek. Games: Mesh-based adapting and modification of gaussian splatting. arXiv preprint arXiv:2402.01459, 2024. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 502 + ], + "type": "text", + "content": "[50] Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. One-step multiview clustering with diverse representation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 555, + 557 + ], + "type": "text", + "content": "[51] Xinhang Wan, Jiyuan Liu, Hao Yu, Qian Qu, Ao Li, Xinwang Liu, Ke Liang, Zhibin Dong, and En Zhu. Contrastive continual multiview clustering with filtered structural fusion. IEEE Transactions on Neural Networks and Learning Systems, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "type": "text", + "content": "[52] Fengxiang Wang, Mingshuo Chen, Yueying Li, Di Wang, Haotian Wang, Zonghao Guo, Zefan Wang, Boqi Shan, Long Lan, Yulin Wang, et al. Geollava-8k: Scaling remote-sensing multimodal large language models to 8k resolution. arXiv preprint arXiv:2505.21375, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 615, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 689 + ], + "type": "text", + "content": "[53] Fengxiang Wang, Hongzhen Wang, Zonghao Guo, Di Wang, Yulin Wang, Mingshuo Chen, Qiang Ma, Long Lan, Wenjing Yang, Jing Zhang, et al. Xlrs-bench: Could your multimodal llms understand extremely large ultra-high-resolution remote sensing imagery? In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14325-14336, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[54] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning Neural Implicit" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9658" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "Surfaces by Volume Rendering for Multi-view Reconstruction. In NeurIPS, 2021. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[55] Xiaogang Wang, Yuelang Xu, Kai Xu, Andrea Tagliafaschi, Bin Zhou, Ali Mahdavi-Amiri, and Hao Zhang. Pie-net: Parametric inference of point cloud edges. Advances in neural information processing systems, 33:20167-20178, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 296, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 296, + 198 + ], + "type": "text", + "content": "[56] Zhifeng Wang, Renjiao Yi, Xin Wen, Chenyang Zhu, and Kai Xu. Vastsd: Learning 3d vascular tree-state space diffusion model for angiography synthesis. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 15693-15702, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 295, + 243 + ], + "type": "text", + "content": "[57] Tong Wu, Lin Gao, Ling-Xiao Zhang, Yu-Kun Lai, and Hao Zhang. Star-tm: Structure aware reconstruction of textured mesh from single image. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-14, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 278 + ], + "type": "text", + "content": "[58] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S. Chirikjian. Primitive-based Shape Abstraction via Nonparametric Bayesian Inference. In ECCV, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 296, + 323 + ], + "type": "text", + "content": "[59] Xihong Yang, Xiaochang Hu, Sihang Zhou, Xinwang Liu, and En Zhu. Interpolation-based contrastive learning for few-label semi-supervised learning. IEEE Transactions on Neural Networks and Learning Systems, 35(2):2054-2065, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 296, + 380 + ], + "type": "text", + "content": "[60] Xihong Yang, Yue Liu, Sihang Zhou, Siwei Wang, Wenxuan Tu, Qun Zheng, Xinwang Liu, Liming Fang, and En Zhu. Cluster-guided contrastive graph clustering network. In Proceedings of the AAAI conference on artificial intelligence, pages 10834-10842, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 296, + 425 + ], + "type": "text", + "content": "[61] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A Large-scale Dataset for Generalized Multi-view Stereo Networks. In CVPR, 2020. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 428, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 296, + 460 + ], + "type": "text", + "content": "[62] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 462, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 296, + 506 + ], + "type": "text", + "content": "[63] Haiyang Ying, Yixuan Yin, Jinzhi Zhang, Fan Wang, Tao Yu, Ruqi Huang, and Lu Fang. Omniseg3d: Omniversal 3d segmentation via hierarchical contrastive learning. arXiv preprint arXiv:2311.11666, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 509, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 296, + 563 + ], + "type": "text", + "content": "[64] Fenggen Yu, Kun Liu, Yan Zhang, Chenyang Zhu, and Kai Xu. Partnet: A recursive part decomposition network for fine-grained and hierarchical shape segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9491-9500, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 296, + 619 + ], + "type": "text", + "content": "[65] Fenggen Yu, Yimin Qian, Xu Zhang, Francisca Gil-Ureta, Brian Jackson, Eric Bennett, and Hao Zhang. Dpa-net: Structured 3d abstraction from sparse views via differentiable primitive assembly. arXiv preprint arXiv:2404.00875, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "type": "text", + "content": "[66] Hongyi Zhou, Xiaogang Wang, Yulan Guo, and Kai Xu. Monomobility: Zero-shot 3d mobility analysis from monocular videos, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 656, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 296, + 689 + ], + "type": "text", + "content": "[67] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018.8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 714 + ], + "type": "text", + "content": "[68] Chenyang Zhu, Kai Xu, Siddhartha Chaudhuri, Li Yi, Leonidas J Guibas, and Hao Zhang. Adacoseg: Adaptive" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 555, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 555, + 106 + ], + "type": "text", + "content": "shape co-segmentation with group consistency loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2020. 1" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9659" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_content_list.json b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a65d414239036e7647fb0f9be3f6405c60382ff0 --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "SemGes: Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning", + "text_level": 1, + "bbox": [ + 125, + 130, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lanmiao Liu $^{1,2,3}$ Esam Ghaleb $^{1,2}$ Asli Özyurek $^{1,2}$ and Zerrin Yumak $^{3}$", + "bbox": [ + 235, + 202, + 795, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Psycholinguistics $^{2}$ Donders Institute for Brain Cognition and Behaviour $^{3}$ Utrecht University {lanmiao.liu, esam.ghaleb, asli.ozyurek}@mpi.nl z.yumak@uu.nl", + "bbox": [ + 122, + 222, + 910, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 258, + 291, + 336, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating a virtual avatar with semantically coherent gestures that are aligned with speech is a challenging task. Existing gesture generation research mainly focused on generating rhythmic beat gestures, neglecting the semantic context of the gestures. In this paper, we propose a novel approach for semantic grounding in co-speech gesture generation that integrates semantic information at both fine-grained and global levels. Our approach starts with learning the motion prior through a vector-quantized variational autoencoder. Built on this model, a second-stage module is applied to automatically generate gestures from speech, text-based semantics and speaker identity that ensures consistency between the semantic relevance of generated gestures and co-occurring speech semantics through semantic coherence and relevance modules. Experimental results demonstrate that our approach enhances the realism and coherence of semantic gestures. Extensive experiments and user studies show that our method outperforms state-of-the-art approaches across two benchmarks in co-speech gesture generation in both objective and subjective metrics. The qualitative results of our model, code, dataset and pre-trained models can be viewed at https://semgesture.github.io/.", + "bbox": [ + 109, + 324, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 718, + 243, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human language is inherently multimodal, with gestures and speech complementing each other to convey pragmatic and semantic information [21, 35]. Cospeech gestures are non-verbal cues that are uniquely related to co-occurring speech, pragmatically, semantically, and temporally. For example, representational iconic gestures that visually express the semantic content of speech and interact with spoken language [13, 14, 19, 21, 41]. A long-standing goal in Computer Vision is to create digital humans that use non-verbal cues in sync with speech. Gesture generation—synthesizing", + "bbox": [ + 109, + 744, + 483, + 912 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd16a7daec2693e07ed1d0fd1d4ecb8a33f5cb99f942f7a9f8240ae716d35d3e.jpg", + "image_caption": [ + "Figure 1. SemGes integrates audio, text-based semantics, and speaker identity to produce both contextually relevant (discourse-level) and fine-grained (local) gestures. A semantic coherence module aligns text and motion embeddings. The multimodal consistency loss synchronizes the quantized multimodal representations to match the quantized learned motion features for final speech-driven semantics-aware gesture generation. The semantic relevance loss selectively emphasizes gestures with semantic annotations." + ], + "image_footnote": [], + "bbox": [ + 535, + 299, + 867, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "movements from co-occurring speech, masked motion, or speaker identity—has advanced to enhance AI agents' expressiveness and realism [29]. However, much of the focus has gone into generating rhythmic beat gestures with limited semantic information, leaving representational gestures that convey semantic messages (e.g., iconic) less explored [29, 40].", + "bbox": [ + 511, + 625, + 883, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating spontaneous and semantically rich gestures from speech comes with multiple challenges. First, it requires capturing global discourse-level information and local fine-grained details (e.g., salient words) to generate speech-driven gestures that reflect the intended meaning and align with speech temporally and semantically. Second, existing methods often generate repetitive and short sequences that do not span the full range of expressive motions required for natural communication. To leverage semantics when generating gestures, researchers have attempted to align motion with speech representations at a global level, e.g., by leveraging pre", + "bbox": [ + 511, + 731, + 883, + 912 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13963", + "bbox": [ + 480, + 950, + 517, + 963 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trained semantic representations such as CLIP [58] or focusing on semantically important keywords [6, 57]. Nonetheless, they often fail to (i) unify global and local semantic modelling within a single framework and (ii) exploit the relevance of the semantic information in guiding gesture generation [27]. At the same time, raw audio features and speaker identity are relevant to the timing and style of gestures. In this paper, we address these limitations by integrating speech, speech semantics, and gesturing style, exploiting semantic information at different levels.", + "bbox": [ + 109, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a two-stage framework, namely, SemGes, that integrates speech, text-based semantics, and speaker identity into a unified gesture-generation model (see Figure 1). In stage 1, we build motion prior of holistic gestures (i.e., body and hands) by training a vector-quantized variational autoencoder (VQ-VAE) to learn an efficient, compositional motion latent space. This stage results in a robust motion encoder & decoder and quantized codebooks that can reconstruct naturalistic gestures while allowing the reuse of learned codebook entries. Stage 2 leverages the learned motion priors to drive gesture synthesis by fusing three modalities using a cross-modal Transformer encoder: (i) text-based semantics, (ii) raw-audio speech features, and (iii) speaker identity for style consistency. We impose a semantic coherence loss that aligns text-based embeddings with the VQ-VAE motion latent space and a semantic relevance loss that emphasises representational gestures (e.g. iconic and metaphoric gestures). A multimodal consistency objective ensures the fused multimodal representations are compatible with the learned motion codebooks, enabling the generation of gestures that are both semantically rich and visually natural. Finally, we introduce a simple but effective long-sequence inference strategy that smoothly combines overlapping motion clips for extended durations. To summarize our contributions,", + "bbox": [ + 109, + 260, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel framework, SemGes, that first learns a robust VQ-VAE motion prior for body and hand gestures, and then generates gestures driven by fused speech audio, text-based semantics, and speaker identity in a cross-modal transformer.", + "- Our method jointly captures discourse-level context via a semantic coherence loss and fine-grained representational gestures (e.g., iconic, metaphoric) via a semantic relevance loss.", + "- We propose an overlap-and-combine inference algorithm that maintains smooth continuity over extended durations.", + "- Extensive experiments on two benchmarks, namely, the BEAT [27] and TED Expressive [33] datasets show that our method outperforms recent baselines in both objective metrics (e.g., Fréchet Gesture Distance" + ], + "bbox": [ + 112, + 670, + 482, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(FGD), diversity, semantic alignment) and user judgment of generated gestures.", + "bbox": [ + 527, + 90, + 880, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 137, + 653, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-driven Co-Speech Gesture Generation. Current gesture generation approaches are based on generative deep neural networks. These approaches use advanced models such as Transformers [28], Generative Adversarial Networks [23], Normalizing Flows [18, 30], Vector Quantized Variational Autoencoder(VQ-VAE) [16] and Denoising Diffusion Probabilistic Models [47]. In addition, researchers have explored the impact of different model inputs on the naturalness and appropriateness of generated gestures. Various modal inputs have been used, such as text [15], audio [50, 59], image [32, 39], and speaking style [3]. For a comprehensive survey, we refer to Nyatsanga et al. [40]. Although there have been significant improvements in this field, current methods fall short in generating semantically grounded gestures at a fine-grained level. In other words, while the generated motions look convincing at first glance, they do not match well with the meaning of the text, or they mostly focus on beat-type gestures.", + "bbox": [ + 511, + 162, + 883, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantics-aware Co-Speech Gesture Generation.", + "text_level": 1, + "bbox": [ + 511, + 473, + 880, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A group of work focused on semantics-aware gesture generation where the semantic information is handled in two ways: global semantics and local semantics. Methods that focus on global semantic information [10, 22, 58] align gestures with text or audio, but they fall short in generating gesture types matching the semantic context, such as iconic, metaphoric and deictic gestures. To capture a wider range of semantic gestures, works like [5, 6, 27] adopt local semantic-aware modelling by integrating the semantic salient words to the neural network. However, these approaches often fail to ensure that the generated gestures align with both the broader audio or textual context and a combination of global and local semantics. Liang et al. [26], Voß and Kopp [49] incorporate both global and local semantics, however, they require extensive annotations. Recently, Zhang et al. [57] employed a generative retrieval framework based on LLMs to address the sparsity problem in datasets with semantic gestures. However, they do not explicitly model the different types of gestures [27] or gesture phases [12] grounded in linguistic research. Moreover, there is still not enough understanding of the impact of different annotations and fine-grained semantics.", + "bbox": [ + 511, + 489, + 883, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Substantial research [6, 28, 57, 58] focused on two-stage latent space generative modelling to overcome the limitations of co-speech gesture generation and to generate more naturalistic and diverse gestures.", + "bbox": [ + 511, + 851, + 883, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13964", + "bbox": [ + 480, + 950, + 517, + 962 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71af008c2730c688b5250cfcba0cceaa9c17e67ca54a1d376357b4a68f77ac40.jpg", + "image_caption": [ + "Figure 2. We pre-train two VQ-VAEs by reconstructing body and hand motions with a dedicated codebook for each." + ], + "image_footnote": [], + "bbox": [ + 130, + 90, + 467, + 141 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These approaches first learn a latent space and then model gestures probabilistically, effectively integrating the strengths of different methods in different stages. Liu et al. [28], Zhang et al. [57] capture complex dependencies in the latent space using VQ-VAE while Zhi et al. [58] employs CLIP [46, 56] to align text and motion embeddings. Ao et al. [6] introduces a diffusion-based model that leverages semantic awareness, while Liu et al. [28] utilizes a transformer-based approach to generate holistic body gestures. In contrast with two-stage generative modeling approaches, end-to-end methods such as [5, 59], are prone to jittering artifacts, especially in hand-motion generation.", + "bbox": [ + 109, + 207, + 483, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model contributes to the research line on semantics-aware co-speech gesture generation by taking into account both global and local semantics. Inspired by the previous work, we employ a two-stage latent space generative modelling for high-quality motion representation. We learn semantic coherence between text and gestures globally with cosine similarity. Moreover, our model takes into account the semantic relevancy of gesture types with minimally required annotations. In contrast with other semantic learning models, we focus on annotations with different gesture types embedded in linguistic research. Our work is closest to CAMN [27] in that sense; however, CAMN does not include semantic coherence learning by aligning text and gestures' latent space globally.", + "bbox": [ + 109, + 404, + 483, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 112, + 648, + 246, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a two-stage approach that generates cospeech gestures by grounding them in raw speech, text-based semantics, and speaker identity. In Section 3.1, we introduce a VQ-VAE encoder-decoder that learns a robust motion prior. Section 3.2 details our gesture synthesis and inference pipeline based on speech, semantics, and identity.", + "bbox": [ + 109, + 674, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation. Our goal is to generate hand gestures $\\pmb{G}^{h} = (g_{1}^{h},\\dots,g_{T}^{h})\\in \\mathbb{R}^{T\\times J}$ and body gestures $\\pmb{G}^{b} = (g_{1}^{b},\\dots,g_{T}^{b})\\in \\mathbb{R}^{T\\times J}$ , where $T$ is the number of time steps and $J$ the number of joints (e.g., 38 for hands, 9 for body). Each motion vector $g_{t}^{h}$ or $g_{t}^{b}$ is encoded in a Rot6D representation, capturing joint rotations at time $t$ .", + "bbox": [ + 109, + 805, + 483, + 910 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To model human motion of body and hands, we first learn a motion generator $\\mathcal{M}_g$ (Stage 1), which synthesizes a plausible motion sequence:", + "bbox": [ + 511, + 90, + 883, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathcal {M} _ {g}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(g _ {1}, \\dots , g _ {T}\\right) \\right\\|. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 142, + 883, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we condition on (i) the raw input audio $\\mathbf{A} = (a_{1},\\ldots ,a_{T})$ , (ii) the speaker identity embedding $I$ , and (iii) the text-based semantic embeddings of the speech $S = (s_{1},\\dots ,s_{T})$ . Our second-stage model $\\mathcal{M}_{a,s,i}$ uses these inputs to generate a latent sequence that the motion generator $\\mathcal{M}_g$ then decodes into naturalistic gestures:", + "bbox": [ + 511, + 178, + 883, + 270 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\mathcal {M} _ {a, s, i}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(\\mathcal {M} _ {a, s, i} (\\boldsymbol {A}, \\boldsymbol {S}, I)\\right) \\right\\|. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 280, + 883, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stage 1: Learning Efficient Codebooks & Compositional Motion Priors", + "text_level": 1, + "bbox": [ + 511, + 309, + 880, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Realistic co-speech gestures require modelling the sequential motion of both body and hand joints. Rather than learning a single representation for the entire body, we adopt a compositional approach, using a discrete codebook of learned representations specific to each part (hands & body). Any gesture motion can then be represented by selecting appropriate codebook entries. Following [28, 48, 53], we employ a VQ-VAE architecture (see Fig. 2) with encoder $\\mathcal{E}_m$ and decoder $\\mathcal{D}_m$ . Given hand motion $G^h \\in \\mathbb{R}^{T \\times J}$ and body motion $G^b \\in \\mathbb{R}^{T \\times J}$ , the encoder produces latent vectors $\\hat{z}^h$ and $\\hat{z}^b$ , which are quantized by selecting the nearest entries in the codebooks. Formally,", + "bbox": [ + 511, + 345, + 883, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} (\\hat {\\boldsymbol {z}}) = \\arg \\min _ {z ^ {i} \\in \\mathcal {Z}} \\| \\hat {z} ^ {j} - z ^ {i} \\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 555, + 883, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z^i$ are the learned codebook entries, and $\\hat{z}^j$ denotes an element of the latent vector for either hand or body. We train the VQ-VAE via a straight-through gradient estimator, minimizing:", + "bbox": [ + 511, + 584, + 883, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\left\\| \\mathbf {g} - \\hat {\\mathbf {g}} \\right\\| ^ {2} + \\left\\| \\dot {\\mathbf {g}} - \\hat {\\dot {\\mathbf {g}}} \\right\\| ^ {2} + \\left\\| \\ddot {\\mathbf {g}} - \\hat {\\ddot {\\mathbf {g}}} \\right\\| ^ {2} \\tag {4} \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {E} (\\mathbf {g}) ] - \\mathbf {q} (\\hat {z}) \\right\\| ^ {2} + \\left\\| \\mathbf {E} (\\mathbf {g}) - \\operatorname {s g} [ \\mathbf {q} (\\hat {z}) ] \\right\\| ^ {2}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 662, + 883, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the first three terms reconstruct joint positions, velocities, and accelerations, and the last two terms implement the VQ-VAE commitment loss [48].", + "bbox": [ + 511, + 700, + 880, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By the end of this stage, we have motion $(m)$ encoder $(\\mathcal{E}_m)$ , decoder $(\\mathcal{D}_m)$ and codebooks $(\\text{Quant}^m(\\cdot))$ for hands and body. In the next section (Section 3.2), we show how this discretized motion of hands and body guides speech, semantics and speaker identity-driven generation to produce realized co-speech gestures.", + "bbox": [ + 511, + 746, + 883, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stage 2: Speech and Identity Driven Semantic Gesture Generator", + "text_level": 1, + "bbox": [ + 511, + 844, + 880, + 873 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This stage focuses on generating gestures conditioned on three inputs: speech embeddings, text-based seman", + "bbox": [ + 511, + 881, + 883, + 912 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13965", + "bbox": [ + 480, + 950, + 517, + 962 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f390537d0780d93a79329de2bb3a674c27d7e5f1c1a52b19a07db673cc813312.jpg", + "image_caption": [ + "Figure 3. SemGes employs three training pathways: (1) Global semantic coherence, which minimizes latent disparities between gesture and text encoders; (2) Multimodal Quantization learning, where integrated multimodal representation codes are aligned with quantized motion to decode them into hand and body movements; and (3) Semantic relevance learning, which emphasizes semantic gestures." + ], + "image_footnote": [], + "bbox": [ + 147, + 98, + 857, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tic embeddings, and speaker identity. As illustrated in Figure 3, the second-stage architecture has three main modules, which we elaborate on in the following subsections.", + "bbox": [ + 111, + 359, + 483, + 419 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79def40ab4ae068b8c0a3781378d3b936df9ae93d335c8b434390dd14871c50a.jpg", + "image_caption": [ + "Figure 4. Semantic Coherence Embedding Learning." + ], + "image_footnote": [], + "bbox": [ + 179, + 446, + 428, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Semantic Coherence Embedding Learning", + "text_level": 1, + "bbox": [ + 112, + 683, + 455, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To align text-based semantics with motion embeddings, we introduce a shared embedding space for both motion priors and speech transcripts. Specifically, we embed word tokens using a pre-trained FastText model [8], then feed these embeddings into a trainable text-based semantic encoder $\\mathcal{E}_s$ . At the same time, we use the pre-trained motion encoder $\\mathcal{E}_m$ from Stage 1 to encode ground-truth gesture sequences. Thus, for a batch of paired (gesture, transcript) samples, we get:", + "bbox": [ + 111, + 702, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Z} ^ {S} = \\mathcal {E} _ {s} (S), \\quad \\mathcal {Z} ^ {h} = \\mathcal {E} _ {m} ^ {h} (G ^ {h}), \\quad \\mathcal {Z} ^ {b} = \\mathcal {E} _ {m} ^ {b} (G ^ {b}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 844, + 482, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S$ is the tokenized speech transcript, and $G^{h}$ and $G^{b}$ correspond to the ground-truth hand gesture sequence and body gesture sequence, respectively. $\\mathcal{Z}^h$ and", + "bbox": [ + 111, + 866, + 483, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{Z}^b$ represent the hand and body ground-truth motion encodings from Stage 1, and $\\mathcal{Z}^s$ denotes the text-based semantic encoder output.", + "bbox": [ + 511, + 359, + 882, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Coherence Loss. We maximize the similarity of correct (gesture, transcript) pairs and minimize it for mismatched pairs, enforcing semantic coherence. This aligns gestures and textual semantics in a common space while keeping $\\mathcal{E}_m$ frozen, as illustrated in Figure 4. We impose the semantic coherence constraint separately on both hand and body movements to align gestures with transcripts in the shared embedding space. Specifically, we introduce two distinct cosine similarity losses: one between the text encoder output and the hand motion latent encoding and another between the text encoder output and the body motion latent encoding. Formally, we minimize:", + "bbox": [ + 511, + 425, + 883, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} = 1 - \\cos \\left(\\mathcal {Z} ^ {h}, \\mathcal {Z} ^ {s}\\right) + 1 - \\cos \\left(\\mathcal {Z} ^ {b}, \\mathcal {Z} ^ {s}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 645, + 883, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the function $\\cos (\\cdot ,\\cdot)$ measures cosine similarity.", + "bbox": [ + 511, + 672, + 875, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Crossmodal Integration", + "text_level": 1, + "bbox": [ + 511, + 696, + 725, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SemGes supports multi-modal inputs in the second training stage by combining audio features and speaker identity with semantic text embeddings using a Transformer encoder with self and cross-attention layers (see Figure 3). We begin by extracting HuBERT features [20] from raw speech (keeping the HuBERT encoder frozen). We concatenate audio features $\\mathcal{Z}^a$ and speaker embeddings $\\mathcal{Z}^i$ , resulting in $\\mathcal{Z}^r$ , which we feed into a self-attention layer.", + "bbox": [ + 511, + 715, + 883, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we use a cross-attention layer that takes $\\mathcal{Z}^r$ as the query and the motion-aligned text-based semantic features $\\mathcal{Z}^s$ as the key-value pair. The final hidden representation $\\mathcal{Z}^f$ serves as the multimodal latent code", + "bbox": [ + 511, + 851, + 883, + 912 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13966", + "bbox": [ + 480, + 950, + 519, + 963 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that drives gesture synthesis when passed to our vector quantization and VQ-VAE-based motion decoder, which is learned in our first stage (see the yellow box in Figure 3).", + "bbox": [ + 111, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Quantization Consistency Loss. SemGes quantizes the multimodal latent code using separate hand and body codebooks. To align this code with the ground-truth motion latent codes, we apply independent quantization consistency losses for each component. Specifically, the quantization loss is defined as:", + "bbox": [ + 111, + 169, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {q u a n t i z a t i o n}} = \\left\\| Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {h}\\right) \\right\\| ^ {2} + \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 280, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {b}\\right) \\right\\| ^ {2} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 253, + 303, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Quant}^h(\\cdot)$ and $\\text{Quant}^b(\\cdot)$ denote the quantization functions for the hand and body codebooks, respectively.", + "bbox": [ + 111, + 330, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multimodal quantization loss aligns the integrated latent code $\\mathcal{Z}^f$ with the learned motion code, a critical step since gesture synthesis is obtained through the quantized multimodal representation. Specifically, $\\mathcal{Z}^f$ is vector-quantized using separate hand and body codebooks before being decoded by their respective VQ decoders. This process ensures that both hand and body movements contribute effectively to the final output. Formally, the generated gestures are given by:", + "bbox": [ + 111, + 377, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {G} = \\hat {G} ^ {h} \\oplus \\hat {G} ^ {b} = \\mathcal {D} _ {m} ^ {h} \\left(\\operatorname {Q u a n t} ^ {h} \\left(\\mathcal {Z} ^ {f}\\right)\\right) \\oplus \\mathcal {D} _ {m} ^ {b} \\left(\\operatorname {Q u a n t} ^ {b} \\left(\\mathcal {Z} ^ {f}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 518, + 482, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes concatenation, jointly synthesizing hand and body motions (i.e. $\\hat{G}$ ).", + "bbox": [ + 111, + 544, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Gesture Semantic Relevance Loss", + "text_level": 1, + "bbox": [ + 112, + 580, + 393, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prioritize the generation of semantically meaningful gestures (e.g., iconic, metaphoric, or deictic), which are less frequent than beat gestures, we introduce a semantic relevance loss. This loss emphasizes semantic annotations while preventing over-penalization of minor deviations. Formally, it is defined as:", + "bbox": [ + 111, + 599, + 483, + 690 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} = \\mathbb {E} \\left[ \\lambda \\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 698, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is the annotation relevance factor, and $\\Psi(\\cdot)$ is a piecewise function that applies a quadratic penalty for small errors and a linear penalty for larger ones:", + "bbox": [ + 111, + 724, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (\\mathbf {G} - \\hat {\\mathbf {G}}) ^ {2}, & \\text {i f} | \\mathbf {G} - \\hat {\\mathbf {G}} | < \\alpha , \\\\ \\alpha \\left(| \\mathbf {G} - \\hat {\\mathbf {G}} | - \\frac {1}{2} \\alpha\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 777, + 482, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\alpha = 0.01$", + "bbox": [ + 112, + 821, + 215, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combined Objective Functions. Finally, the overall objective is:", + "bbox": [ + 111, + 854, + 483, + 883 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {S e m G e s}} = \\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} + \\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} + \\mathcal {L} _ {\\text {q u a n t i z a t i o n}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 893, + 482, + 907 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Long Gesture Sequence Algorithm" + ], + "code_body": "Require: Audio $\\mathcal{A}$ aligned speech transcript $S$ and speaker ID $\\mathcal{I}$ Pre-trained codebooks and motion decoder (Stage 1) \nEnsure: Long-sequence gesture M \n1: Partition $(\\mathcal{A},\\mathcal{S},\\mathcal{I})$ into clips $\\{(\\mathcal{A}_c,\\mathcal{S}_c,\\mathcal{I}_c)\\}_{c = 1}^C$ \n2: Compute latent representation: $\\mathcal{Z}^{f}\\gets$ Encode(A,S,I) \n3: Quantize: $\\mathcal{Z}^e\\gets$ VectorQuantize $(\\mathcal{Z}^{f})$ \n4: Decode initial clip: $\\hat{M}_1\\gets \\mathrm{Dec}(\\mathcal{Z}^e)$ \n5: for each clip $c = 2$ to $C$ do \n6: Set first 4 frames of $\\hat{M}_c$ to the last 4 frames of $\\hat{M}_{c - 1}$ \n7: Generate remaining frames of $\\hat{M}_c$ \n8: end for \n9: return $\\hat{M}$", + "bbox": [ + 514, + 109, + 883, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which jointly optimizes the model to generate gestures that are semantically coherent at both global and fine-grained levels while remaining faithful to the Stage 1 motion prior.", + "bbox": [ + 511, + 364, + 883, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Inference of Long Gesture Sequences", + "text_level": 1, + "bbox": [ + 511, + 435, + 834, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating long sequences of gestures is challenging due to the need to maintain coherence and smooth transitions. Our Long-Sequence Gesture Motion algorithm (Alg. 1) addresses these challenges by partitioning the input speech, transcript, and speaker identity into aligned clips. For each clip, a multimodal latent representation is computed using our cross-modal encoder, vector-quantized via the Stage 1 codebooks, and decoded into gesture motions. Overlapping 4-frame segments between clips provides continuity, resulting in extended, naturalistic gesture sequences.", + "bbox": [ + 511, + 458, + 883, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 638, + 705, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 662, + 614, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed methodology is evaluated on two benchmarks, namely, BEAT [27] and the TED expressive dataset [33]. The BEAT dataset consists of 76 hours of multimodal recordings, which include speech audio recordings, speech transcriptions, and, more importantly, motion data collected from 30 participants, leveraging Motion Capture (MOCAP) technology. The participants expressed emotions in eight distinct scenarios across four languages. The motion data contains joint rotation angles, which were designed for consistency across varying body sizes. The TED Expressive dataset [33] is segmented from TED Talk videos into smaller shots based on scene boundaries. Liu et al. [33] extracted each frame's 2D human pose using OpenPose BEAT [9]. Using these 2D pose priors, ExPose [43] was", + "bbox": [ + 511, + 685, + 883, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13967", + "bbox": [ + 480, + 950, + 517, + 963 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fe64ef64ce5708037abbe668334be41eacceca7f62ef9e0e84f81d5cab389ab9.jpg", + "table_caption": [ + "Table 1. Comparison of SemGesGen with other methods on the BEAT and TED-Expressive datasets. For BEAT, we compare with CaMN [27], DiffGesture [59], LivelySpeaker [58], and DiffSheg [10]. The same methods are evaluated on TED-Expressive. SRGR is not applicable (denoted with $-$ ) for TED-Expressive as it does not contain annotations for semantic relevance of gestures." + ], + "table_footnote": [], + "table_body": "
BEATTED-Expressive
MethodFGD ↓BC ↑Diversity ↑SRGR ↑MethodFGD ↓BC ↑Diversity ↑
CaMN [27]8.5100.797206.7890.231CaMN [27]7.6730.642156.236
DiffGesture [59]9.6320.876210.6780.106DiffGesture [59]9.3260.662119.889
LivelySpeaker [58]13.3780.891214.9460.229LivelySpeaker [58]8.1450.691119.231
DiffSheg [10]6.6230.922257.6740.250DiffSheg [10]8.4570.712108.972
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
", + "bbox": [ + 148, + 142, + 846, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "employed to annotate the 3D upper body keypoints, including 13 upper body joints and 30 finger joints. Both datasets' training and validation samples are divided into 34-frame clips.", + "bbox": [ + 111, + 276, + 483, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Validation. We evaluate our approach on the BEAT dataset, following the protocol in [27], where the data is randomly split into a 19:2:2 ratio for training, validation, and testing. Similarly, for the TED Expressive dataset, we adapt the protocol in [33], using a random split of 8:1:1 for training, validation, and testing.", + "bbox": [ + 111, + 357, + 483, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The details of the model architectures and training are provided in Section 2 of the Supplementary Materials.", + "bbox": [ + 111, + 468, + 483, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. State-of-the-Art Baselines", + "text_level": 1, + "bbox": [ + 112, + 523, + 348, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare SemGes against a set of representative state-of-the-art models that focus on semantic-driven gesture generation. The selected models achieved strong performance on the BEAT and TED-Expressive datasets, making them suitable for a fair comparison with our method. The selected models are as follows:", + "bbox": [ + 111, + 546, + 483, + 637 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Cascaded Motion Network(CaMN) [27] is the current benchmark model for the BEAT dataset. CaMN is based on LSTMs and integrates multiple input modalities, including audio, text, facial expressions, and emotion. Additionally, like SemGes, it leverages semantic relevance annotations to enhance gesture generation.", + "2. DiffSHEG [10] is a state-of-the-art diffusion-based model for real-time speech-driven holistic gesture generation. It is conditioned on noisy motion, audio, and speaker ID. DiffSHEG introduces a Fast Out-painting-based Partial Autoregressive Sampling method to efficiently generate arbitrary-length sequences in real time.", + "3. LivelySpeaker [58] generates semantically and rhythmically aware co-speech gestures by leveraging an MLP-based diffusion model. The model conditions gesture generation on text, noised motion," + ], + "bbox": [ + 112, + 638, + 483, + 912 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "speaker ID, and audio to enable text-driven gesture control while incorporating global semantics.", + "bbox": [ + 532, + 276, + 882, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. DiffGes [59] models the diffusion and denoising processes within the gesture domain, enabling the generation of high-fidelity, audio-driven gestures conditioned on both audio and gesture inputs. Several recent studies[6, 26] have also demonstrated strong performance in this area.", + "bbox": [ + 511, + 306, + 883, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We exclude certain models from our comparison. For instance, SEEG [26] and [57] rely on additional data annotations (e.g., Semantic Prompt Gallery or ChatGPT-generated annotations) that are not uniformly available. In addition, other works, such as Ao et al. [6], Pang et al. [42], Zhang et al. [57], are excluded from our analysis due to the inaccessibility of their codebase. Voß and Kopp [49] is omitted due to its high computational cost and the unavailability of annotations. Liu et al. [31, 34], Mughal et al. [36, 37], Ng et al. [38], Yi et al. [54] are excluded as they primarily focus on holistic gestures with face and mesh data, which fall outside the scope of this work. Similarly, Chhatre et al. [11], Qi et al. [44] are excluded, as their emphasis lies in emotion-driven gesture generation rather than the semantic aspects. Furthermore, Ahuja et al. [1, 2], Alexander et al. [4], Habibie et al. [17], Liu et al. [33], Sun et al. [45], Yang et al. [51], Ye et al. [52] are omitted due to their lack of relevance to semantic-driven gesture generation.", + "bbox": [ + 511, + 400, + 883, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Quantitative Objective Evaluations", + "text_level": 1, + "bbox": [ + 511, + 719, + 815, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We employ four standard objective metrics for evaluating the quality of gesture generation, namely, Fréchet Gesture Distance (FGD) [55], Beat Consistency Score (BC) [25], Diversity [24], and Semantic-Relevant Gesture Recall (SRGR) [27].", + "bbox": [ + 511, + 744, + 882, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FGD measures how the generated gestures resemble real motion distributions by embedding sequences into a latent space via a pre-trained autoencoder. In contrast, BC focuses on synchronization with speech, measuring the alignment between speech onsets (audio beats) and motion beats, which are identified as velocity minima in", + "bbox": [ + 511, + 821, + 883, + 912 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13968", + "bbox": [ + 480, + 950, + 517, + 962 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ce76706a7dfae3d06af8432d0a430ad28cd1f590b81667879d4f45900acd19f.jpg", + "table_caption": [ + "Table 2. Ablation studies evaluating the contributions of key components in SemGes on the BEAT and TED-Expressive Datasets. For BEAT, performance is measured using FGD (lower is better), BC, Diversity, and SRGR, while for TED-Expressive, SRGR is not applicable (denoted as -)." + ], + "table_footnote": [], + "table_body": "
BEATTED-Expressive
Model VariantsFGD ↓BC ↑Diversity ↑SRGR ↑Model VariantsFGD ↓BC ↑Diversity ↑
Baseline (VQVAE)10.3480.564198.5680.176Baseline (VQVAE)10.6820.612114.692
w/o Semantic Coherence Module8.0530.556249.5500.180w/o Semantic Coherence Module7.9240.623109.256
w/o Semantic Relevance Module7.5490.573245.3190.195w/o Semantic Relevance Module---
w/ SpeechCLIP Encoder6.7870.468289.6210.245w/ SpeechCLIP Encoder7.3410.605245.680
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
", + "bbox": [ + 148, + 141, + 844, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "upper-body joints (excluding fingers). Meanwhile, **Diversity** captures the variability of generated motions by computing the average $L1$ distance between pairs of $N$ generated clips. Finally, SRGR assesses semantic relevance by determining how well generated gestures align with the annotated semantic gestures. Further details on the objective metrics are included in the Supplementary Materials (Section 1).", + "bbox": [ + 109, + 252, + 480, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with Other Models. Table 1 compares the performance of our approach against four baseline methods across four evaluation metrics. As highlighted in the table, SemGes outperforms the baselines in FGD, Diversity, and SRGR.", + "bbox": [ + 109, + 396, + 480, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the BEAT dataset, our approach achieves the highest SRGR, which we attribute to the exploitation of semantic relevance information in our training objectives. In addition, our approach shows a significant improvement in FGD and Diversity, indicating a closer alignment with the ground truth gesture distribution and a broader range of generated gestures compared to the second-best baselines. The performance on the Beat Consistency (BC) metric is lower for our method. This is expected given our focus on improving semantic awareness of the generated gestures rather than optimizing for strict temporal alignment between rhythmic beat gestures and speech. In addition, the BC metric can be sensitive to rapid, jittery movements; even minor motion artefacts may be mistakenly counted as additional beats, thereby increasing the BC score artificially- a phenomenon also observed in the diffusion-based baselines, as further illustrated in our supplementary video.", + "bbox": [ + 109, + 472, + 482, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate how our model handles the trade-off between semantic and beat scores by testing the model on beat-dominant gestures (without semantic content). The results show a significantly higher Beat score (0.689) than the full dataset Beat score (0.453). This confirms rhythmic consistency in beat-focused contexts. We provide additional evaluation in the supplementary materials (Section 3) to show how the model handles difficult cases (such as noisy speech or misaligned speech).", + "bbox": [ + 109, + 744, + 482, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that the TED Expressive dataset lacks annotations for gesture semantic relevance, so SRGR is not", + "bbox": [ + 111, + 882, + 483, + 912 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "applicable, and the semantic relevance loss was omitted during training. Nevertheless, SemGes produces diverse, naturalistic gestures on TED Expressive, outperforming baselines in FGD and Diversity metrics.", + "bbox": [ + 511, + 252, + 883, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We evaluate the contributions of key components in SemGes through ablation experiments. First, we assess a baseline VQ-VAE model (Stage 1 only), which uses two stacked encoder-decoder blocks and an MLP. In this experiment, we test its ability to generate gestures, conditioned on audio, masked motion, and speaker identity. As shown in Table 2, this baseline underperforms compared to state-of-the-art methods (Table 1). As a result, we motivate our two-stage design where the VQ-VAE is reserved to learn the motion latent space and Stage 2 leverages speech and identity conditioning to generate gestures.", + "bbox": [ + 511, + 330, + 883, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we examine Stage 2 by removing its components: (i) the Semantic Coherence Loss, (ii) the Semantic Relevance Loss, and (iii) by replacing the HuBERT-based speech encoder with SpeechCLIP. Results in Table 2 show that removing either the Semantic Coherence or Relevance Loss degrades FGD, Diversity, and SRGR scores, highlighting their roles in aligning gesture representations with textual semantics and capturing semantic importance. In addition, replacing the speech encoder results in marginal gains. The semantic encoder is fixed as FastText, which we believe is sufficient to capture the necessary semantic information [7]. Overall, these results confirm the importance of each module in generating semantics-aware gestures.", + "bbox": [ + 511, + 512, + 883, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Qualitative & Subjective Evaluations", + "text_level": 1, + "bbox": [ + 511, + 734, + 848, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization Comparisons. Before presenting the subjective ratings of the generated gestures, Figure 5 provides a visual comparison between the ground truth, results from our approach and two baseline models. We use examples from the BEAT dataset. It is clear from the figure that our approach not only achieves better speech-gesture alignment but also produces gestures that are more naturalistic, diverse, and semantically aware. For example, while CaMN generates smooth movements, its gestures tend to be slower and less varied compared to", + "bbox": [ + 511, + 760, + 883, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13969", + "bbox": [ + 480, + 950, + 519, + 963 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4028ad1da7211a158b3965e028ce1b10b9d5550fdc235e8203825ca3cebacf9.jpg", + "image_caption": [ + "Figure 5. Comparisons with baselines and ground truth gestures. Compared to the baseline method, our approach generates gestures that are aligned with speech content (semantics). For instance, when the speaker says \"remix\", our method produces gestures where the character raises both hands to emphasize the word before gradually lowering them—a movement that other methods fail to achieve. Similarly, when uttering \"first\", our method generates a raised hand gesture, producing an iconic gesture." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 869, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our model. Additionally, the baseline methods show varying degrees of jitter—DiffGesture shows the highest jitter, followed by LivelySpeaker and DiffSheg, with CaMN displaying the least. Although CaMN includes semantic information, our approach strikes a more effective balance, generating gestures that align with actual motion, as shown also with the objective metrics. Based on these qualitative observations, our subsequent rating study focuses on evaluating gestures produced by the ground truth, our model, CaMN, and DiffSHEG.", + "bbox": [ + 109, + 316, + 483, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Ratings of Generated Gestures. We conducted a user study using 40-second video clips from the BEAT test set, featuring subjects narrating six topics. Thirty native English speakers from the United Kingdom and the United States participated, with an average age of $36 \\pm 20$ years and a female-to-male ratio of approximately 2:1. Each participant evaluated 24 videos generated by the ground truth, CaMN, DiffSHEG, and our model over a study duration of that lasted on average $27 \\pm 5$ minutes. For data quality, participants were required to pass attention verification questions, i.e., correctly answering at least two out of four questions regarding the narration topic. Participants rated the videos on three criteria: naturalness, diversity, and alignment with speech content and timing on a scale from 1 to 5. The videos were presented in a randomized order to avoid bias. In Section 3 of the Supplementary Materials, we provide screenshots and more details on the user study and interface.", + "bbox": [ + 109, + 488, + 482, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 shows that ground-truth gestures received an average rating of 4 across all metrics, establishing an upper bound and validating the participant survey. Our model received the highest ratings among the generated gestures, significantly outperforming CaMN and Diff-SHEG in naturalness, synchronization, and diversity (indicated by the in Figure 6). These results prove that our approach produces gestures that are more natural, better aligned with speech, and more diverse than those gener", + "bbox": [ + 109, + 776, + 482, + 912 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ated by SOTA baselines.", + "bbox": [ + 513, + 316, + 679, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3d5c4222b692542165121d51ee5f693894562de6161cdae88f72d56ef8b033e.jpg", + "image_caption": [ + "Figure 6. Average ratings of users for ground truth gestures and gestures generated through our approach, CAMN, and DiffSHEG. The bars illustrate the average user ratings across three metrics: naturalness, diversity, and alignment with speech content and timing. Statistical t-tests show that our approach received significantly higher ratings than CAMN and DiffSHEG, with $p < 0.05$ ." + ], + "image_footnote": [], + "bbox": [ + 535, + 347, + 849, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 643, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed SemGes, a novel two-stage approach to semantic grounding in co-speech gesture generation by integrating semantic information at both fine-grained and global levels. In the first stage, a motion prior generation module is trained using a vector-quantized variational autoencoder to produce realistic and smooth gesture motions. Building upon this model, the second stage generates gestures from speech, text-based semantics, and speaker identity while maintaining consistency between gesture semantics and co-occurring speech through semantic coherence and relevance modules. Subjective and objective evaluations show that our work achieves state-of-the-art performance across two public benchmarks, generating semantics-aware and diverse gestures. Future direction and limitations are discussed in Section 5 of the Supplementary Materials.", + "bbox": [ + 511, + 670, + 883, + 912 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13970", + "bbox": [ + 480, + 950, + 519, + 963 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 112, + 90, + 272, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The project is funded by the Max Planck Society. We thank Sachit Misra for his invaluable assistance with rendering Avatar characters. We extend our gratitude to the members of the Multimodal Language Department at Max Planck Institute for Psycholinguistics for their feedback.", + "bbox": [ + 111, + 114, + 483, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 228, + 209, + 244 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566-20576, 2022. 6", + "[2] Chaitanya Ahuja, Pratik Joshi, Ryo Ishii, and Louis-Philippe Morency. Continual learning for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20893-20903, 2023. 6", + "[3] Simon Alexander, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2", + "[4] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audiodriven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 6", + "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6): 1-19, 2022. 2, 3", + "[6] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2, 3, 6", + "[7] Ben Athiwaratkun, Andrew Gordon Wilson, and Anima Anandkumar. Probabilistic fasttext for multi-sense word embeddings. arXiv preprint arXiv:1806.02901, 2018. 7", + "[8] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 4", + "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291–7299, 2017. 5", + "[10] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In CVPR, 2024. 2, 6", + "[11] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al." + ], + "bbox": [ + 114, + 253, + 483, + 912 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 6", + "[12] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Adversarial gesture generation with realistic gesture phasing. Computers & Graphics, 89:117-130, 2020. 2", + "[13] Esam Ghaleb, Bulat Khaertdinov, Wim Pouw, Marlou Rasenberg, Judith Holler, Asli Ozyurek, and Raquel Fernandez. Learning co-speech gesture representations in dialogue through contrastive learning: An intrinsic evaluation. In Proceedings of the 26th International Conference on Multimodal Interaction, pages 274-283, 2024. 1", + "[14] Esam Ghaleb, Bulat Khaertdinov, Asli Özyürek, and Raquel Fernández. I see what you mean: Co-speech gestures for reference resolution in multimodal dialogue. In Proceedings of the 63rd Conference of the Association for Computational Linguistics (ACL Findings), 2025. To appear. 1", + "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2", + "[16] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2", + "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 6", + "[18] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2", + "[19] Judith Holler and Stephen C Levinson. Multimodal language processing in human communication. Trends in Cognitive Sciences, 23(8):639-652, 2019. 1", + "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451-3460, 2021. 4", + "[21] Adam Kendon. Gesture units, gesture phrases and speech. In *Gesture: Visible Action as Utterance*, chapter 7, page 108–126. Cambridge University Press, 2004. 1", + "[22] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2" + ], + "bbox": [ + 516, + 92, + 883, + 912 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13971", + "bbox": [ + 480, + 950, + 517, + 962 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1279, 2022. 2", + "[24] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 6", + "[25] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6", + "[26] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2, 6", + "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6", + "[28] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 2, 3", + "[29] Li Liu, Lufei Gao, Wentao Lei, Fengji Ma, Xiaotian Lin, and Jinting Wang. A survey on deep multi-modal learning for body language recognition and generation. arXiv preprint arXiv:2308.08849, 2023. 1", + "[30] Lanmiao Liu, Chuang Yu, Siyang Song, Zhidong Su, and Adriana Tapus. Human gesture recognition with a flow-based model for human robot interaction. In *Compan- ion of the 2023 ACM/IEEE International Conference on Human-Robot Interaction*, pages 548–551, 2023. 2", + "[31] Pinxin Liu, Luchuan Song, Junhua Huang, Haiyang Liu, and Chenliang Xu. Gesturelsm: Latent shortcut based co-speech gesture generation with spatial-temporal modeling. arXiv preprint arXiv:2501.18898, 2025. 6", + "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. Advances in Neural Information Processing Systems, 35:21386-21399, 2022. 2", + "[33] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for co-speech gesture generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10462-10472, 2022. 2, 5, 6", + "[34] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated" + ], + "bbox": [ + 114, + 90, + 485, + 912 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 6", + "[35] David McNeill. Hand and mind. Advances in Visual Semiotics, 351, 1992. 1", + "[36] Muhammad Hamza Mughal, Rishabh Dabral, Ikhsanul Habibie, Lucia Donatelli, Marc Habermann, and Christian Theobalt. Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1388-1398, 2024. 6", + "[37] M Hamza Mughal, Rishabh Dabral, Merel CJ Scholman, Vera Demberg, and Christian Theobalt. Retrieving semantics from the deep: an rag solution for gesture synthesis. arXiv preprint arXiv:2412.06786, 2024. 6", + "[38] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 6", + "[39] Mang Ning, Mingxiao Li, Jianlin Su, Haozhe Jia, Lanmiao Liu, Martin Beneš, Wenshuo Chen, Albert Ali Salah, and Itir Onal Ertugrul. Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2412.15032, 2024. 2", + "[40] Simbarashe Nyatsanga, Taras Kucherenko, Chaitanya Ahuja, Gustav Eje Henter, and Michael Neff. A comprehensive review of data-driven co-speech gesture generation. In Computer Graphics Forum, pages 569-596. Wiley Online Library, 2023. 1, 2", + "[41] Asli Özyürek. Hearing and seeing meaning in speech and gesture: Insights from brain and behaviour. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1651):20130296, 2014. 1", + "[42] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5", + "[44] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10424-10434, 2024. 6", + "[45] Mingyang Sun, Mengchen Zhao, Yaqing Hou, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2331-2340, 2023. 6" + ], + "bbox": [ + 516, + 92, + 883, + 912 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13972", + "bbox": [ + 480, + 950, + 519, + 963 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 3", + "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3", + "[49] Hendric Voß and Stefan Kopp. Augmented co-speech gesture generation: Including form and meaning features to guide learning-based gesture synthesis. In Proceedings of the 23rd ACM International Conference on Intelligent Virtual Agents, pages 1-8, 2023. 2, 6", + "[50] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven cospeech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2", + "[51] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2321-2330, 2023. 6", + "[52] Sheng Ye, Yu-Hui Wen, Yanan Sun, Ying He, Ziyang Zhang, Yaoyuan Wang, Weihua He, and Yong-Jin Liu. Audio-driven stylized gesture generation with flow-based model. In European Conference on Computer Vision, pages 712-728. Springer, 2022. 6", + "[53] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 3", + "[54] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yan-dong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 6", + "[55] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 6", + "[56] Pengfei Zhang, Pinxin Liu, Hyeongwoo Kim, Pablo Garrido, and Bindita Chaudhuri. Kinmo: Kinematic-aware human motion understanding and generation. arXiv preprint arXiv:2411.15472, 2024.3", + "[57] Zeyi Zhang, Tenglong Ao, Yuyao Zhang, Qingzhe Gao, Chuan Lin, Baoquan Chen, and Libin Liu. Semantic gesticulator: Semantics-aware co-speech gesture synthesis. ACM Transactions on Graphics (TOG), 43(4):1-17, 2024. 2, 3, 6", + "[58] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker:" + ], + "bbox": [ + 114, + 92, + 482, + 911 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 3, 6", + "[59] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 6" + ], + "bbox": [ + 516, + 93, + 883, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13973", + "bbox": [ + 480, + 950, + 517, + 962 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_model.json b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e46690320f8bc183bce7283de4a36ca08100b5b5 --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.872, + 0.177 + ], + "angle": 0, + "content": "SemGes: Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.203, + 0.796, + 0.221 + ], + "angle": 0, + "content": "Lanmiao Liu\\(^{1,2,3}\\) Esam Ghaleb\\(^{1,2}\\) Asli Özyurek\\(^{1,2}\\) and Zerrin Yumak\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.223, + 0.911, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Psycholinguistics \\(^{2}\\)Donders Institute for Brain Cognition and Behaviour \\(^{3}\\)Utrecht University {lanmiao.liu, esam.ghaleb, asli.ozyurek}@mpi.nl z.yumak@uu.nl" + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.292, + 0.338, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.325, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Creating a virtual avatar with semantically coherent gestures that are aligned with speech is a challenging task. Existing gesture generation research mainly focused on generating rhythmic beat gestures, neglecting the semantic context of the gestures. In this paper, we propose a novel approach for semantic grounding in co-speech gesture generation that integrates semantic information at both fine-grained and global levels. Our approach starts with learning the motion prior through a vector-quantized variational autoencoder. Built on this model, a second-stage module is applied to automatically generate gestures from speech, text-based semantics and speaker identity that ensures consistency between the semantic relevance of generated gestures and co-occurring speech semantics through semantic coherence and relevance modules. Experimental results demonstrate that our approach enhances the realism and coherence of semantic gestures. Extensive experiments and user studies show that our method outperforms state-of-the-art approaches across two benchmarks in co-speech gesture generation in both objective and subjective metrics. The qualitative results of our model, code, dataset and pre-trained models can be viewed at https://semgesture.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.719, + 0.245, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.746, + 0.485, + 0.913 + ], + "angle": 0, + "content": "Human language is inherently multimodal, with gestures and speech complementing each other to convey pragmatic and semantic information [21, 35]. Cospeech gestures are non-verbal cues that are uniquely related to co-occurring speech, pragmatically, semantically, and temporally. For example, representational iconic gestures that visually express the semantic content of speech and interact with spoken language [13, 14, 19, 21, 41]. A long-standing goal in Computer Vision is to create digital humans that use non-verbal cues in sync with speech. Gesture generation—synthesizing" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.3, + 0.868, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.477, + 0.885, + 0.603 + ], + "angle": 0, + "content": "Figure 1. SemGes integrates audio, text-based semantics, and speaker identity to produce both contextually relevant (discourse-level) and fine-grained (local) gestures. A semantic coherence module aligns text and motion embeddings. The multimodal consistency loss synchronizes the quantized multimodal representations to match the quantized learned motion features for final speech-driven semantics-aware gesture generation. The semantic relevance loss selectively emphasizes gestures with semantic annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.885, + 0.731 + ], + "angle": 0, + "content": "movements from co-occurring speech, masked motion, or speaker identity—has advanced to enhance AI agents' expressiveness and realism [29]. However, much of the focus has gone into generating rhythmic beat gestures with limited semantic information, leaving representational gestures that convey semantic messages (e.g., iconic) less explored [29, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Generating spontaneous and semantically rich gestures from speech comes with multiple challenges. First, it requires capturing global discourse-level information and local fine-grained details (e.g., salient words) to generate speech-driven gestures that reflect the intended meaning and align with speech temporally and semantically. Second, existing methods often generate repetitive and short sequences that do not span the full range of expressive motions required for natural communication. To leverage semantics when generating gestures, researchers have attempted to align motion with speech representations at a global level, e.g., by leveraging pre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "13963" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "trained semantic representations such as CLIP [58] or focusing on semantically important keywords [6, 57]. Nonetheless, they often fail to (i) unify global and local semantic modelling within a single framework and (ii) exploit the relevance of the semantic information in guiding gesture generation [27]. At the same time, raw audio features and speaker identity are relevant to the timing and style of gestures. In this paper, we address these limitations by integrating speech, speech semantics, and gesturing style, exploiting semantic information at different levels." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.261, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Specifically, we propose a two-stage framework, namely, SemGes, that integrates speech, text-based semantics, and speaker identity into a unified gesture-generation model (see Figure 1). In stage 1, we build motion prior of holistic gestures (i.e., body and hands) by training a vector-quantized variational autoencoder (VQ-VAE) to learn an efficient, compositional motion latent space. This stage results in a robust motion encoder & decoder and quantized codebooks that can reconstruct naturalistic gestures while allowing the reuse of learned codebook entries. Stage 2 leverages the learned motion priors to drive gesture synthesis by fusing three modalities using a cross-modal Transformer encoder: (i) text-based semantics, (ii) raw-audio speech features, and (iii) speaker identity for style consistency. We impose a semantic coherence loss that aligns text-based embeddings with the VQ-VAE motion latent space and a semantic relevance loss that emphasises representational gestures (e.g. iconic and metaphoric gestures). A multimodal consistency objective ensures the fused multimodal representations are compatible with the learned motion codebooks, enabling the generation of gestures that are both semantically rich and visually natural. Finally, we introduce a simple but effective long-sequence inference strategy that smoothly combines overlapping motion clips for extended durations. To summarize our contributions," + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.671, + 0.483, + 0.746 + ], + "angle": 0, + "content": "- We introduce a novel framework, SemGes, that first learns a robust VQ-VAE motion prior for body and hand gestures, and then generates gestures driven by fused speech audio, text-based semantics, and speaker identity in a cross-modal transformer." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.747, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- Our method jointly captures discourse-level context via a semantic coherence loss and fine-grained representational gestures (e.g., iconic, metaphoric) via a semantic relevance loss." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.808, + 0.483, + 0.85 + ], + "angle": 0, + "content": "- We propose an overlap-and-combine inference algorithm that maintains smooth continuity over extended durations." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.852, + 0.483, + 0.912 + ], + "angle": 0, + "content": "- Extensive experiments on two benchmarks, namely, the BEAT [27] and TED Expressive [33] datasets show that our method outperforms recent baselines in both objective metrics (e.g., Fréchet Gesture Distance" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.671, + 0.483, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.882, + 0.122 + ], + "angle": 0, + "content": "(FGD), diversity, semantic alignment) and user judgment of generated gestures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.138, + 0.655, + 0.154 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.164, + 0.885, + 0.452 + ], + "angle": 0, + "content": "Data-driven Co-Speech Gesture Generation. Current gesture generation approaches are based on generative deep neural networks. These approaches use advanced models such as Transformers [28], Generative Adversarial Networks [23], Normalizing Flows [18, 30], Vector Quantized Variational Autoencoder(VQ-VAE) [16] and Denoising Diffusion Probabilistic Models [47]. In addition, researchers have explored the impact of different model inputs on the naturalness and appropriateness of generated gestures. Various modal inputs have been used, such as text [15], audio [50, 59], image [32, 39], and speaking style [3]. For a comprehensive survey, we refer to Nyatsanga et al. [40]. Although there have been significant improvements in this field, current methods fall short in generating semantically grounded gestures at a fine-grained level. In other words, while the generated motions look convincing at first glance, they do not match well with the meaning of the text, or they mostly focus on beat-type gestures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.474, + 0.882, + 0.489 + ], + "angle": 0, + "content": "Semantics-aware Co-Speech Gesture Generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.49, + 0.884, + 0.849 + ], + "angle": 0, + "content": "A group of work focused on semantics-aware gesture generation where the semantic information is handled in two ways: global semantics and local semantics. Methods that focus on global semantic information [10, 22, 58] align gestures with text or audio, but they fall short in generating gesture types matching the semantic context, such as iconic, metaphoric and deictic gestures. To capture a wider range of semantic gestures, works like [5, 6, 27] adopt local semantic-aware modelling by integrating the semantic salient words to the neural network. However, these approaches often fail to ensure that the generated gestures align with both the broader audio or textual context and a combination of global and local semantics. Liang et al. [26], Voß and Kopp [49] incorporate both global and local semantics, however, they require extensive annotations. Recently, Zhang et al. [57] employed a generative retrieval framework based on LLMs to address the sparsity problem in datasets with semantic gestures. However, they do not explicitly model the different types of gestures [27] or gesture phases [12] grounded in linguistic research. Moreover, there is still not enough understanding of the impact of different annotations and fine-grained semantics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.852, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Substantial research [6, 28, 57, 58] focused on two-stage latent space generative modelling to overcome the limitations of co-speech gesture generation and to generate more naturalistic and diverse gestures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "13964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.092, + 0.468, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.157, + 0.483, + 0.185 + ], + "angle": 0, + "content": "Figure 2. We pre-train two VQ-VAEs by reconstructing body and hand motions with a dedicated codebook for each." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.208, + 0.484, + 0.404 + ], + "angle": 0, + "content": "These approaches first learn a latent space and then model gestures probabilistically, effectively integrating the strengths of different methods in different stages. Liu et al. [28], Zhang et al. [57] capture complex dependencies in the latent space using VQ-VAE while Zhi et al. [58] employs CLIP [46, 56] to align text and motion embeddings. Ao et al. [6] introduces a diffusion-based model that leverages semantic awareness, while Liu et al. [28] utilizes a transformer-based approach to generate holistic body gestures. In contrast with two-stage generative modeling approaches, end-to-end methods such as [5, 59], are prone to jittering artifacts, especially in hand-motion generation." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.405, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Our model contributes to the research line on semantics-aware co-speech gesture generation by taking into account both global and local semantics. Inspired by the previous work, we employ a two-stage latent space generative modelling for high-quality motion representation. We learn semantic coherence between text and gestures globally with cosine similarity. Moreover, our model takes into account the semantic relevancy of gesture types with minimally required annotations. In contrast with other semantic learning models, we focus on annotations with different gesture types embedded in linguistic research. Our work is closest to CAMN [27] in that sense; however, CAMN does not include semantic coherence learning by aligning text and gestures' latent space globally." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.649, + 0.248, + 0.667 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.675, + 0.483, + 0.781 + ], + "angle": 0, + "content": "We propose a two-stage approach that generates cospeech gestures by grounding them in raw speech, text-based semantics, and speaker identity. In Section 3.1, we introduce a VQ-VAE encoder-decoder that learns a robust motion prior. Section 3.2 details our gesture synthesis and inference pipeline based on speech, semantics, and identity." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.806, + 0.484, + 0.911 + ], + "angle": 0, + "content": "Problem formulation. Our goal is to generate hand gestures \\( \\pmb{G}^{h} = (g_{1}^{h},\\dots,g_{T}^{h})\\in \\mathbb{R}^{T\\times J} \\) and body gestures \\( \\pmb{G}^{b} = (g_{1}^{b},\\dots,g_{T}^{b})\\in \\mathbb{R}^{T\\times J} \\), where \\( T \\) is the number of time steps and \\( J \\) the number of joints (e.g., 38 for hands, 9 for body). Each motion vector \\( g_{t}^{h} \\) or \\( g_{t}^{b} \\) is encoded in a Rot6D representation, capturing joint rotations at time \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.138 + ], + "angle": 0, + "content": "To model human motion of body and hands, we first learn a motion generator \\(\\mathcal{M}_g\\) (Stage 1), which synthesizes a plausible motion sequence:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.143, + 0.884, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathcal {M} _ {g}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(g _ {1}, \\dots , g _ {T}\\right) \\right\\|. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.179, + 0.884, + 0.271 + ], + "angle": 0, + "content": "Next, we condition on (i) the raw input audio \\( \\mathbf{A} = (a_{1},\\ldots ,a_{T}) \\), (ii) the speaker identity embedding \\( I \\), and (iii) the text-based semantic embeddings of the speech \\( S = (s_{1},\\dots ,s_{T}) \\). Our second-stage model \\( \\mathcal{M}_{a,s,i} \\) uses these inputs to generate a latent sequence that the motion generator \\( \\mathcal{M}_g \\) then decodes into naturalistic gestures:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.281, + 0.884, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\mathcal {M} _ {a, s, i}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(\\mathcal {M} _ {a, s, i} (\\boldsymbol {A}, \\boldsymbol {S}, I)\\right) \\right\\|. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.31, + 0.882, + 0.34 + ], + "angle": 0, + "content": "3.1. Stage 1: Learning Efficient Codebooks & Compositional Motion Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.884, + 0.543 + ], + "angle": 0, + "content": "Realistic co-speech gestures require modelling the sequential motion of both body and hand joints. Rather than learning a single representation for the entire body, we adopt a compositional approach, using a discrete codebook of learned representations specific to each part (hands & body). Any gesture motion can then be represented by selecting appropriate codebook entries. Following [28, 48, 53], we employ a VQ-VAE architecture (see Fig. 2) with encoder \\(\\mathcal{E}_m\\) and decoder \\(\\mathcal{D}_m\\). Given hand motion \\(G^h \\in \\mathbb{R}^{T \\times J}\\) and body motion \\(G^b \\in \\mathbb{R}^{T \\times J}\\), the encoder produces latent vectors \\(\\hat{z}^h\\) and \\(\\hat{z}^b\\), which are quantized by selecting the nearest entries in the codebooks. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.556, + 0.884, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} (\\hat {\\boldsymbol {z}}) = \\arg \\min _ {z ^ {i} \\in \\mathcal {Z}} \\| \\hat {z} ^ {j} - z ^ {i} \\|, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.884, + 0.645 + ], + "angle": 0, + "content": "where \\(z^i\\) are the learned codebook entries, and \\(\\hat{z}^j\\) denotes an element of the latent vector for either hand or body. We train the VQ-VAE via a straight-through gradient estimator, minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.663, + 0.884, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\left\\| \\mathbf {g} - \\hat {\\mathbf {g}} \\right\\| ^ {2} + \\left\\| \\dot {\\mathbf {g}} - \\hat {\\dot {\\mathbf {g}}} \\right\\| ^ {2} + \\left\\| \\ddot {\\mathbf {g}} - \\hat {\\ddot {\\mathbf {g}}} \\right\\| ^ {2} \\tag {4} \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {E} (\\mathbf {g}) ] - \\mathbf {q} (\\hat {z}) \\right\\| ^ {2} + \\left\\| \\mathbf {E} (\\mathbf {g}) - \\operatorname {s g} [ \\mathbf {q} (\\hat {z}) ] \\right\\| ^ {2}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.882, + 0.746 + ], + "angle": 0, + "content": "where the first three terms reconstruct joint positions, velocities, and accelerations, and the last two terms implement the VQ-VAE commitment loss [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.884, + 0.838 + ], + "angle": 0, + "content": "By the end of this stage, we have motion \\((m)\\) encoder \\((\\mathcal{E}_m)\\), decoder \\((\\mathcal{D}_m)\\) and codebooks \\((\\text{Quant}^m(\\cdot))\\) for hands and body. In the next section (Section 3.2), we show how this discretized motion of hands and body guides speech, semantics and speaker identity-driven generation to produce realized co-speech gestures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.882, + 0.874 + ], + "angle": 0, + "content": "3.2. Stage 2: Speech and Identity Driven Semantic Gesture Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.882, + 0.884, + 0.913 + ], + "angle": 0, + "content": "This stage focuses on generating gestures conditioned on three inputs: speech embeddings, text-based seman" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "13965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.099, + 0.859, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.285, + 0.887, + 0.342 + ], + "angle": 0, + "content": "Figure 3. SemGes employs three training pathways: (1) Global semantic coherence, which minimizes latent disparities between gesture and text encoders; (2) Multimodal Quantization learning, where integrated multimodal representation codes are aligned with quantized motion to decode them into hand and body movements; and (3) Semantic relevance learning, which emphasizes semantic gestures." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.36, + 0.484, + 0.42 + ], + "angle": 0, + "content": "tic embeddings, and speaker identity. As illustrated in Figure 3, the second-stage architecture has three main modules, which we elaborate on in the following subsections." + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.447, + 0.429, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.654, + 0.456, + 0.669 + ], + "angle": 0, + "content": "Figure 4. Semantic Coherence Embedding Learning." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.684, + 0.456, + 0.7 + ], + "angle": 0, + "content": "3.2.1. Semantic Coherence Embedding Learning" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.703, + 0.484, + 0.839 + ], + "angle": 0, + "content": "To align text-based semantics with motion embeddings, we introduce a shared embedding space for both motion priors and speech transcripts. Specifically, we embed word tokens using a pre-trained FastText model [8], then feed these embeddings into a trainable text-based semantic encoder \\(\\mathcal{E}_s\\). At the same time, we use the pre-trained motion encoder \\(\\mathcal{E}_m\\) from Stage 1 to encode ground-truth gesture sequences. Thus, for a batch of paired (gesture, transcript) samples, we get:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.845, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Z} ^ {S} = \\mathcal {E} _ {s} (S), \\quad \\mathcal {Z} ^ {h} = \\mathcal {E} _ {m} ^ {h} (G ^ {h}), \\quad \\mathcal {Z} ^ {b} = \\mathcal {E} _ {m} ^ {b} (G ^ {b}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.867, + 0.484, + 0.913 + ], + "angle": 0, + "content": "where \\(S\\) is the tokenized speech transcript, and \\(G^{h}\\) and \\(G^{b}\\) correspond to the ground-truth hand gesture sequence and body gesture sequence, respectively. \\(\\mathcal{Z}^h\\) and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.36, + 0.883, + 0.406 + ], + "angle": 0, + "content": "\\(\\mathcal{Z}^b\\) represent the hand and body ground-truth motion encodings from Stage 1, and \\(\\mathcal{Z}^s\\) denotes the text-based semantic encoder output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.884, + 0.622 + ], + "angle": 0, + "content": "Semantic Coherence Loss. We maximize the similarity of correct (gesture, transcript) pairs and minimize it for mismatched pairs, enforcing semantic coherence. This aligns gestures and textual semantics in a common space while keeping \\(\\mathcal{E}_m\\) frozen, as illustrated in Figure 4. We impose the semantic coherence constraint separately on both hand and body movements to align gestures with transcripts in the shared embedding space. Specifically, we introduce two distinct cosine similarity losses: one between the text encoder output and the hand motion latent encoding and another between the text encoder output and the body motion latent encoding. Formally, we minimize:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.646, + 0.884, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} = 1 - \\cos \\left(\\mathcal {Z} ^ {h}, \\mathcal {Z} ^ {s}\\right) + 1 - \\cos \\left(\\mathcal {Z} ^ {b}, \\mathcal {Z} ^ {s}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.876, + 0.689 + ], + "angle": 0, + "content": "where the function \\(\\cos (\\cdot ,\\cdot)\\) measures cosine similarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.727, + 0.712 + ], + "angle": 0, + "content": "3.2.2. Crossmodal Integration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.884, + 0.851 + ], + "angle": 0, + "content": "SemGes supports multi-modal inputs in the second training stage by combining audio features and speaker identity with semantic text embeddings using a Transformer encoder with self and cross-attention layers (see Figure 3). We begin by extracting HuBERT features [20] from raw speech (keeping the HuBERT encoder frozen). We concatenate audio features \\(\\mathcal{Z}^a\\) and speaker embeddings \\(\\mathcal{Z}^i\\), resulting in \\(\\mathcal{Z}^r\\), which we feed into a self-attention layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.852, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Next, we use a cross-attention layer that takes \\(\\mathcal{Z}^r\\) as the query and the motion-aligned text-based semantic features \\(\\mathcal{Z}^s\\) as the key-value pair. The final hidden representation \\(\\mathcal{Z}^f\\) serves as the multimodal latent code" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "13966" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "that drives gesture synthesis when passed to our vector quantization and VQ-VAE-based motion decoder, which is learned in our first stage (see the yellow box in Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.17, + 0.484, + 0.274 + ], + "angle": 0, + "content": "Multimodal Quantization Consistency Loss. SemGes quantizes the multimodal latent code using separate hand and body codebooks. To align this code with the ground-truth motion latent codes, we apply independent quantization consistency losses for each component. Specifically, the quantization loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.281, + 0.484, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {q u a n t i z a t i o n}} = \\left\\| Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {h}\\right) \\right\\| ^ {2} + \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.254, + 0.304, + 0.484, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\left\\| Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {b}\\right) \\right\\| ^ {2} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.331, + 0.483, + 0.376 + ], + "angle": 0, + "content": "where \\( \\text{Quant}^h(\\cdot) \\) and \\( \\text{Quant}^b(\\cdot) \\) denote the quantization functions for the hand and body codebooks, respectively." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.378, + 0.484, + 0.513 + ], + "angle": 0, + "content": "The multimodal quantization loss aligns the integrated latent code \\(\\mathcal{Z}^f\\) with the learned motion code, a critical step since gesture synthesis is obtained through the quantized multimodal representation. Specifically, \\(\\mathcal{Z}^f\\) is vector-quantized using separate hand and body codebooks before being decoded by their respective VQ decoders. This process ensures that both hand and body movements contribute effectively to the final output. Formally, the generated gestures are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.52, + 0.483, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\hat {G} = \\hat {G} ^ {h} \\oplus \\hat {G} ^ {b} = \\mathcal {D} _ {m} ^ {h} \\left(\\operatorname {Q u a n t} ^ {h} \\left(\\mathcal {Z} ^ {f}\\right)\\right) \\oplus \\mathcal {D} _ {m} ^ {b} \\left(\\operatorname {Q u a n t} ^ {b} \\left(\\mathcal {Z} ^ {f}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.545, + 0.484, + 0.576 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes concatenation, jointly synthesizing hand and body motions (i.e. \\(\\hat{G}\\))." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.582, + 0.394, + 0.595 + ], + "angle": 0, + "content": "3.2.3. Gesture Semantic Relevance Loss" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.6, + 0.484, + 0.691 + ], + "angle": 0, + "content": "To prioritize the generation of semantically meaningful gestures (e.g., iconic, metaphoric, or deictic), which are less frequent than beat gestures, we introduce a semantic relevance loss. This loss emphasizes semantic annotations while preventing over-penalization of minor deviations. Formally, it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.699, + 0.483, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} = \\mathbb {E} \\left[ \\lambda \\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.726, + 0.484, + 0.772 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the annotation relevance factor, and \\(\\Psi(\\cdot)\\) is a piecewise function that applies a quadratic penalty for small errors and a linear penalty for larger ones:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.779, + 0.483, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (\\mathbf {G} - \\hat {\\mathbf {G}}) ^ {2}, & \\text {i f} | \\mathbf {G} - \\hat {\\mathbf {G}} | < \\alpha , \\\\ \\alpha \\left(| \\mathbf {G} - \\hat {\\mathbf {G}} | - \\frac {1}{2} \\alpha\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.822, + 0.217, + 0.835 + ], + "angle": 0, + "content": "with \\(\\alpha = 0.01\\)" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.855, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Combined Objective Functions. Finally, the overall objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.895, + 0.483, + 0.909 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {S e m G e s}} = \\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} + \\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} + \\mathcal {L} _ {\\text {q u a n t i z a t i o n}}, \\tag {11}\n\\]" + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.836, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Long Gesture Sequence Algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.111, + 0.885, + 0.337 + ], + "angle": 0, + "content": "Require: Audio \\(\\mathcal{A}\\) aligned speech transcript \\(S\\) and speaker ID \\(\\mathcal{I}\\) Pre-trained codebooks and motion decoder (Stage 1) \nEnsure: Long-sequence gesture M \n1: Partition \\((\\mathcal{A},\\mathcal{S},\\mathcal{I})\\) into clips \\(\\{(\\mathcal{A}_c,\\mathcal{S}_c,\\mathcal{I}_c)\\}_{c = 1}^C\\) \n2: Compute latent representation: \\(\\mathcal{Z}^{f}\\gets\\) Encode(A,S,I) \n3: Quantize: \\(\\mathcal{Z}^e\\gets\\) VectorQuantize \\((\\mathcal{Z}^{f})\\) \n4: Decode initial clip: \\(\\hat{M}_1\\gets \\mathrm{Dec}(\\mathcal{Z}^e)\\) \n5: for each clip \\(c = 2\\) to \\(C\\) do \n6: Set first 4 frames of \\(\\hat{M}_c\\) to the last 4 frames of \\(\\hat{M}_{c - 1}\\) \n7: Generate remaining frames of \\(\\hat{M}_c\\) \n8: end for \n9: return \\(\\hat{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.884, + 0.427 + ], + "angle": 0, + "content": "which jointly optimizes the model to generate gestures that are semantically coherent at both global and fine-grained levels while remaining faithful to the Stage 1 motion prior." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.835, + 0.453 + ], + "angle": 0, + "content": "3.3. Inference of Long Gesture Sequences" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.884, + 0.626 + ], + "angle": 0, + "content": "Generating long sequences of gestures is challenging due to the need to maintain coherence and smooth transitions. Our Long-Sequence Gesture Motion algorithm (Alg. 1) addresses these challenges by partitioning the input speech, transcript, and speaker identity into aligned clips. For each clip, a multimodal latent representation is computed using our cross-modal encoder, vector-quantized via the Stage 1 codebooks, and decoded into gesture motions. Overlapping 4-frame segments between clips provides continuity, resulting in extended, naturalistic gesture sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.706, + 0.657 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.616, + 0.679 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Our proposed methodology is evaluated on two benchmarks, namely, BEAT [27] and the TED expressive dataset [33]. The BEAT dataset consists of 76 hours of multimodal recordings, which include speech audio recordings, speech transcriptions, and, more importantly, motion data collected from 30 participants, leveraging Motion Capture (MOCAP) technology. The participants expressed emotions in eight distinct scenarios across four languages. The motion data contains joint rotation angles, which were designed for consistency across varying body sizes. The TED Expressive dataset [33] is segmented from TED Talk videos into smaller shots based on scene boundaries. Liu et al. [33] extracted each frame's 2D human pose using OpenPose BEAT [9]. Using these 2D pose priors, ExPose [43] was" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "13967" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.089, + 0.885, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison of SemGesGen with other methods on the BEAT and TED-Expressive datasets. For BEAT, we compare with CaMN [27], DiffGesture [59], LivelySpeaker [58], and DiffSheg [10]. The same methods are evaluated on TED-Expressive. SRGR is not applicable (denoted with \\(-\\)) for TED-Expressive as it does not contain annotations for semantic relevance of gestures." + }, + { + "type": "table", + "bbox": [ + 0.15, + 0.143, + 0.847, + 0.26 + ], + "angle": 0, + "content": "
BEATTED-Expressive
MethodFGD ↓BC ↑Diversity ↑SRGR ↑MethodFGD ↓BC ↑Diversity ↑
CaMN [27]8.5100.797206.7890.231CaMN [27]7.6730.642156.236
DiffGesture [59]9.6320.876210.6780.106DiffGesture [59]9.3260.662119.889
LivelySpeaker [58]13.3780.891214.9460.229LivelySpeaker [58]8.1450.691119.231
DiffSheg [10]6.6230.922257.6740.250DiffSheg [10]8.4570.712108.972
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.277, + 0.484, + 0.339 + ], + "angle": 0, + "content": "employed to annotate the 3D upper body keypoints, including 13 upper body joints and 30 finger joints. Both datasets' training and validation samples are divided into 34-frame clips." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.358, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Cross-Validation. We evaluate our approach on the BEAT dataset, following the protocol in [27], where the data is randomly split into a 19:2:2 ratio for training, validation, and testing. Similarly, for the TED Expressive dataset, we adapt the protocol in [33], using a random split of 8:1:1 for training, validation, and testing." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.469, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Implementation Details. The details of the model architectures and training are provided in Section 2 of the Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.524, + 0.349, + 0.54 + ], + "angle": 0, + "content": "4.2. State-of-the-Art Baselines" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.547, + 0.484, + 0.638 + ], + "angle": 0, + "content": "We compare SemGes against a set of representative state-of-the-art models that focus on semantic-driven gesture generation. The selected models achieved strong performance on the BEAT and TED-Expressive datasets, making them suitable for a fair comparison with our method. The selected models are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.64, + 0.483, + 0.746 + ], + "angle": 0, + "content": "1. Cascaded Motion Network(CaMN) [27] is the current benchmark model for the BEAT dataset. CaMN is based on LSTMs and integrates multiple input modalities, including audio, text, facial expressions, and emotion. Additionally, like SemGes, it leverages semantic relevance annotations to enhance gesture generation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.747, + 0.483, + 0.852 + ], + "angle": 0, + "content": "2. DiffSHEG [10] is a state-of-the-art diffusion-based model for real-time speech-driven holistic gesture generation. It is conditioned on noisy motion, audio, and speaker ID. DiffSHEG introduces a Fast Out-painting-based Partial Autoregressive Sampling method to efficiently generate arbitrary-length sequences in real time." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.852, + 0.484, + 0.913 + ], + "angle": 0, + "content": "3. LivelySpeaker [58] generates semantically and rhythmically aware co-speech gestures by leveraging an MLP-based diffusion model. The model conditions gesture generation on text, noised motion," + }, + { + "type": "list", + "bbox": [ + 0.113, + 0.64, + 0.484, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.277, + 0.883, + 0.307 + ], + "angle": 0, + "content": "speaker ID, and audio to enable text-driven gesture control while incorporating global semantics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.884, + 0.398 + ], + "angle": 0, + "content": "4. DiffGes [59] models the diffusion and denoising processes within the gesture domain, enabling the generation of high-fidelity, audio-driven gestures conditioned on both audio and gesture inputs. Several recent studies[6, 26] have also demonstrated strong performance in this area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.884, + 0.705 + ], + "angle": 0, + "content": "We exclude certain models from our comparison. For instance, SEEG [26] and [57] rely on additional data annotations (e.g., Semantic Prompt Gallery or ChatGPT-generated annotations) that are not uniformly available. In addition, other works, such as Ao et al. [6], Pang et al. [42], Zhang et al. [57], are excluded from our analysis due to the inaccessibility of their codebase. Voß and Kopp [49] is omitted due to its high computational cost and the unavailability of annotations. Liu et al. [31, 34], Mughal et al. [36, 37], Ng et al. [38], Yi et al. [54] are excluded as they primarily focus on holistic gestures with face and mesh data, which fall outside the scope of this work. Similarly, Chhatre et al. [11], Qi et al. [44] are excluded, as their emphasis lies in emotion-driven gesture generation rather than the semantic aspects. Furthermore, Ahuja et al. [1, 2], Alexander et al. [4], Habibie et al. [17], Liu et al. [33], Sun et al. [45], Yang et al. [51], Ye et al. [52] are omitted due to their lack of relevance to semantic-driven gesture generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.72, + 0.816, + 0.737 + ], + "angle": 0, + "content": "5. Quantitative Objective Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.883, + 0.821 + ], + "angle": 0, + "content": "Evaluation Metrics. We employ four standard objective metrics for evaluating the quality of gesture generation, namely, Fréchet Gesture Distance (FGD) [55], Beat Consistency Score (BC) [25], Diversity [24], and Semantic-Relevant Gesture Recall (SRGR) [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.884, + 0.913 + ], + "angle": 0, + "content": "FGD measures how the generated gestures resemble real motion distributions by embedding sequences into a latent space via a pre-trained autoencoder. In contrast, BC focuses on synchronization with speech, measuring the alignment between speech onsets (audio beats) and motion beats, which are identified as velocity minima in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "13968" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.089, + 0.884, + 0.133 + ], + "angle": 0, + "content": "Table 2. Ablation studies evaluating the contributions of key components in SemGes on the BEAT and TED-Expressive Datasets. For BEAT, performance is measured using FGD (lower is better), BC, Diversity, and SRGR, while for TED-Expressive, SRGR is not applicable (denoted as -)." + }, + { + "type": "table", + "bbox": [ + 0.15, + 0.142, + 0.846, + 0.235 + ], + "angle": 0, + "content": "
BEATTED-Expressive
Model VariantsFGD ↓BC ↑Diversity ↑SRGR ↑Model VariantsFGD ↓BC ↑Diversity ↑
Baseline (VQVAE)10.3480.564198.5680.176Baseline (VQVAE)10.6820.612114.692
w/o Semantic Coherence Module8.0530.556249.5500.180w/o Semantic Coherence Module7.9240.623109.256
w/o Semantic Relevance Module7.5490.573245.3190.195w/o Semantic Relevance Module---
w/ SpeechCLIP Encoder6.7870.468289.6210.245w/ SpeechCLIP Encoder7.3410.605245.680
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.253, + 0.482, + 0.372 + ], + "angle": 0, + "content": "upper-body joints (excluding fingers). Meanwhile, **Diversity** captures the variability of generated motions by computing the average \\( L1 \\) distance between pairs of \\( N \\) generated clips. Finally, SRGR assesses semantic relevance by determining how well generated gestures align with the annotated semantic gestures. Further details on the objective metrics are included in the Supplementary Materials (Section 1)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.397, + 0.482, + 0.472 + ], + "angle": 0, + "content": "Comparisons with Other Models. Table 1 compares the performance of our approach against four baseline methods across four evaluation metrics. As highlighted in the table, SemGes outperforms the baselines in FGD, Diversity, and SRGR." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.473, + 0.483, + 0.744 + ], + "angle": 0, + "content": "For the BEAT dataset, our approach achieves the highest SRGR, which we attribute to the exploitation of semantic relevance information in our training objectives. In addition, our approach shows a significant improvement in FGD and Diversity, indicating a closer alignment with the ground truth gesture distribution and a broader range of generated gestures compared to the second-best baselines. The performance on the Beat Consistency (BC) metric is lower for our method. This is expected given our focus on improving semantic awareness of the generated gestures rather than optimizing for strict temporal alignment between rhythmic beat gestures and speech. In addition, the BC metric can be sensitive to rapid, jittery movements; even minor motion artefacts may be mistakenly counted as additional beats, thereby increasing the BC score artificially- a phenomenon also observed in the diffusion-based baselines, as further illustrated in our supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.746, + 0.483, + 0.882 + ], + "angle": 0, + "content": "We evaluate how our model handles the trade-off between semantic and beat scores by testing the model on beat-dominant gestures (without semantic content). The results show a significantly higher Beat score (0.689) than the full dataset Beat score (0.453). This confirms rhythmic consistency in beat-focused contexts. We provide additional evaluation in the supplementary materials (Section 3) to show how the model handles difficult cases (such as noisy speech or misaligned speech)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.883, + 0.484, + 0.913 + ], + "angle": 0, + "content": "Note that the TED Expressive dataset lacks annotations for gesture semantic relevance, so SRGR is not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.885, + 0.314 + ], + "angle": 0, + "content": "applicable, and the semantic relevance loss was omitted during training. Nevertheless, SemGes produces diverse, naturalistic gestures on TED Expressive, outperforming baselines in FGD and Diversity metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.884, + 0.512 + ], + "angle": 0, + "content": "Ablation Study. We evaluate the contributions of key components in SemGes through ablation experiments. First, we assess a baseline VQ-VAE model (Stage 1 only), which uses two stacked encoder-decoder blocks and an MLP. In this experiment, we test its ability to generate gestures, conditioned on audio, masked motion, and speaker identity. As shown in Table 2, this baseline underperforms compared to state-of-the-art methods (Table 1). As a result, we motivate our two-stage design where the VQ-VAE is reserved to learn the motion latent space and Stage 2 leverages speech and identity conditioning to generate gestures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.884, + 0.724 + ], + "angle": 0, + "content": "Next, we examine Stage 2 by removing its components: (i) the Semantic Coherence Loss, (ii) the Semantic Relevance Loss, and (iii) by replacing the HuBERT-based speech encoder with SpeechCLIP. Results in Table 2 show that removing either the Semantic Coherence or Relevance Loss degrades FGD, Diversity, and SRGR scores, highlighting their roles in aligning gesture representations with textual semantics and capturing semantic importance. In addition, replacing the speech encoder results in marginal gains. The semantic encoder is fixed as FastText, which we believe is sufficient to capture the necessary semantic information [7]. Overall, these results confirm the importance of each module in generating semantics-aware gestures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.849, + 0.755 + ], + "angle": 0, + "content": "6. Qualitative & Subjective Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Visualization Comparisons. Before presenting the subjective ratings of the generated gestures, Figure 5 provides a visual comparison between the ground truth, results from our approach and two baseline models. We use examples from the BEAT dataset. It is clear from the figure that our approach not only achieves better speech-gesture alignment but also produces gestures that are more naturalistic, diverse, and semantically aware. For example, while CaMN generates smooth movements, its gestures tend to be slower and less varied compared to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "13969" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.09, + 0.87, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.241, + 0.885, + 0.3 + ], + "angle": 0, + "content": "Figure 5. Comparisons with baselines and ground truth gestures. Compared to the baseline method, our approach generates gestures that are aligned with speech content (semantics). For instance, when the speaker says \"remix\", our method produces gestures where the character raises both hands to emphasize the word before gradually lowering them—a movement that other methods fail to achieve. Similarly, when uttering \"first\", our method generates a raised hand gesture, producing an iconic gesture." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.317, + 0.484, + 0.468 + ], + "angle": 0, + "content": "our model. Additionally, the baseline methods show varying degrees of jitter—DiffGesture shows the highest jitter, followed by LivelySpeaker and DiffSheg, with CaMN displaying the least. Although CaMN includes semantic information, our approach strikes a more effective balance, generating gestures that align with actual motion, as shown also with the objective metrics. Based on these qualitative observations, our subsequent rating study focuses on evaluating gestures produced by the ground truth, our model, CaMN, and DiffSHEG." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.489, + 0.483, + 0.775 + ], + "angle": 0, + "content": "User Ratings of Generated Gestures. We conducted a user study using 40-second video clips from the BEAT test set, featuring subjects narrating six topics. Thirty native English speakers from the United Kingdom and the United States participated, with an average age of \\(36 \\pm 20\\) years and a female-to-male ratio of approximately 2:1. Each participant evaluated 24 videos generated by the ground truth, CaMN, DiffSHEG, and our model over a study duration of that lasted on average \\(27 \\pm 5\\) minutes. For data quality, participants were required to pass attention verification questions, i.e., correctly answering at least two out of four questions regarding the narration topic. Participants rated the videos on three criteria: naturalness, diversity, and alignment with speech content and timing on a scale from 1 to 5. The videos were presented in a randomized order to avoid bias. In Section 3 of the Supplementary Materials, we provide screenshots and more details on the user study and interface." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.777, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Figure 6 shows that ground-truth gestures received an average rating of 4 across all metrics, establishing an upper bound and validating the participant survey. Our model received the highest ratings among the generated gestures, significantly outperforming CaMN and Diff-SHEG in naturalness, synchronization, and diversity (indicated by the in Figure 6). These results prove that our approach produces gestures that are more natural, better aligned with speech, and more diverse than those gener" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.317, + 0.68, + 0.332 + ], + "angle": 0, + "content": "ated by SOTA baselines." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.348, + 0.85, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.522, + 0.885, + 0.62 + ], + "angle": 0, + "content": "Figure 6. Average ratings of users for ground truth gestures and gestures generated through our approach, CAMN, and DiffSHEG. The bars illustrate the average user ratings across three metrics: naturalness, diversity, and alignment with speech content and timing. Statistical t-tests show that our approach received significantly higher ratings than CAMN and DiffSHEG, with \\( p < 0.05 \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.645, + 0.634, + 0.661 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.885, + 0.913 + ], + "angle": 0, + "content": "We proposed SemGes, a novel two-stage approach to semantic grounding in co-speech gesture generation by integrating semantic information at both fine-grained and global levels. In the first stage, a motion prior generation module is trained using a vector-quantized variational autoencoder to produce realistic and smooth gesture motions. Building upon this model, the second stage generates gestures from speech, text-based semantics, and speaker identity while maintaining consistency between gesture semantics and co-occurring speech through semantic coherence and relevance modules. Subjective and objective evaluations show that our work achieves state-of-the-art performance across two public benchmarks, generating semantics-aware and diverse gestures. Future direction and limitations are discussed in Section 5 of the Supplementary Materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "13970" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.091, + 0.274, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.115, + 0.484, + 0.201 + ], + "angle": 0, + "content": "The project is funded by the Max Planck Society. We thank Sachit Misra for his invaluable assistance with rendering Avatar characters. We extend our gratitude to the members of the Multimodal Language Department at Max Planck Institute for Psycholinguistics for their feedback." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.229, + 0.21, + 0.245 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.254, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566-20576, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.326, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[2] Chaitanya Ahuja, Pratik Joshi, Ryo Ishii, and Louis-Philippe Morency. Continual learning for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20893-20903, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.397, + 0.484, + 0.467 + ], + "angle": 0, + "content": "[3] Simon Alexander, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.468, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[4] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audiodriven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.526, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6): 1-19, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.597, + 0.484, + 0.652 + ], + "angle": 0, + "content": "[6] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.655, + 0.484, + 0.697 + ], + "angle": 0, + "content": "[7] Ben Athiwaratkun, Andrew Gordon Wilson, and Anima Anandkumar. Probabilistic fasttext for multi-sense word embeddings. arXiv preprint arXiv:1806.02901, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.698, + 0.484, + 0.754 + ], + "angle": 0, + "content": "[8] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.755, + 0.484, + 0.825 + ], + "angle": 0, + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291–7299, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.827, + 0.484, + 0.883 + ], + "angle": 0, + "content": "[10] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.885, + 0.484, + 0.913 + ], + "angle": 0, + "content": "[11] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.254, + 0.484, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.885, + 0.148 + ], + "angle": 0, + "content": "Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.884, + 0.19 + ], + "angle": 0, + "content": "[12] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Adversarial gesture generation with realistic gesture phasing. Computers & Graphics, 89:117-130, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.885, + 0.285 + ], + "angle": 0, + "content": "[13] Esam Ghaleb, Bulat Khaertdinov, Wim Pouw, Marlou Rasenberg, Judith Holler, Asli Ozyurek, and Raquel Fernandez. Learning co-speech gesture representations in dialogue through contrastive learning: An intrinsic evaluation. In Proceedings of the 26th International Conference on Multimodal Interaction, pages 274-283, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.287, + 0.885, + 0.37 + ], + "angle": 0, + "content": "[14] Esam Ghaleb, Bulat Khaertdinov, Asli Özyürek, and Raquel Fernández. I see what you mean: Co-speech gestures for reference resolution in multimodal dialogue. In Proceedings of the 63rd Conference of the Association for Computational Linguistics (ACL Findings), 2025. To appear. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.885, + 0.44 + ], + "angle": 0, + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.885, + 0.509 + ], + "angle": 0, + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.51, + 0.885, + 0.593 + ], + "angle": 0, + "content": "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.593, + 0.885, + 0.648 + ], + "angle": 0, + "content": "[18] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.885, + 0.69 + ], + "angle": 0, + "content": "[19] Judith Holler and Stephen C Levinson. Multimodal language processing in human communication. Trends in Cognitive Sciences, 23(8):639-652, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.691, + 0.885, + 0.773 + ], + "angle": 0, + "content": "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451-3460, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.885, + 0.828 + ], + "angle": 0, + "content": "[21] Adam Kendon. Gesture units, gesture phrases and speech. In *Gesture: Visible Action as Utterance*, chapter 7, page 108–126. Cambridge University Press, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.885, + 0.913 + ], + "angle": 0, + "content": "[22] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.885, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "13971" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[23] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1279, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.162, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[24] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.246, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[25] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[26] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.385, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.468, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[28] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[29] Li Liu, Lufei Gao, Wentao Lei, Fengji Ma, Xiaotian Lin, and Jinting Wang. A survey on deep multi-modal learning for body language recognition and generation. arXiv preprint arXiv:2308.08849, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.621, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[30] Lanmiao Liu, Chuang Yu, Siyang Song, Zhidong Su, and Adriana Tapus. Human gesture recognition with a flow-based model for human robot interaction. In *Compan- ion of the 2023 ACM/IEEE International Conference on Human-Robot Interaction*, pages 548–551, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.69, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[31] Pinxin Liu, Luchuan Song, Junhua Huang, Haiyang Liu, and Chenliang Xu. Gesturelsm: Latent shortcut based co-speech gesture generation with spatial-temporal modeling. arXiv preprint arXiv:2501.18898, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. Advances in Neural Information Processing Systems, 35:21386-21399, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.801, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[33] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for co-speech gesture generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10462-10472, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.885, + 0.483, + 0.913 + ], + "angle": 0, + "content": "[34] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.092, + 0.486, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.885, + 0.134 + ], + "angle": 0, + "content": "holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.884, + 0.161 + ], + "angle": 0, + "content": "[35] David McNeill. Hand and mind. Advances in Visual Semiotics, 351, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.884, + 0.245 + ], + "angle": 0, + "content": "[36] Muhammad Hamza Mughal, Rishabh Dabral, Ikhsanul Habibie, Lucia Donatelli, Marc Habermann, and Christian Theobalt. Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1388-1398, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.883, + 0.301 + ], + "angle": 0, + "content": "[37] M Hamza Mughal, Rishabh Dabral, Merel CJ Scholman, Vera Demberg, and Christian Theobalt. Retrieving semantics from the deep: an rag solution for gesture synthesis. arXiv preprint arXiv:2412.06786, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.883, + 0.384 + ], + "angle": 0, + "content": "[38] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.883, + 0.453 + ], + "angle": 0, + "content": "[39] Mang Ning, Mingxiao Li, Jianlin Su, Haozhe Jia, Lanmiao Liu, Martin Beneš, Wenshuo Chen, Albert Ali Salah, and Itir Onal Ertugrul. Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2412.15032, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.883, + 0.523 + ], + "angle": 0, + "content": "[40] Simbarashe Nyatsanga, Taras Kucherenko, Chaitanya Ahuja, Gustav Eje Henter, and Michael Neff. A comprehensive review of data-driven co-speech gesture generation. In Computer Graphics Forum, pages 569-596. Wiley Online Library, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.883, + 0.579 + ], + "angle": 0, + "content": "[41] Asli Özyürek. Hearing and seeing meaning in speech and gesture: Insights from brain and behaviour. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1651):20130296, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.883, + 0.648 + ], + "angle": 0, + "content": "[42] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.884, + 0.732 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.883, + 0.828 + ], + "angle": 0, + "content": "[44] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10424-10434, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.883, + 0.913 + ], + "angle": 0, + "content": "[45] Mingyang Sun, Mengchen Zhao, Yaqing Hou, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2331-2340, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.885, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.52, + 0.964 + ], + "angle": 0, + "content": "13972" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[49] Hendric Voß and Stefan Kopp. Augmented co-speech gesture generation: Including form and meaning features to guide learning-based gesture synthesis. In Proceedings of the 23rd ACM International Conference on Intelligent Virtual Agents, pages 1-8, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.317, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[50] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven cospeech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.385, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[51] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2321-2330, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.482, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[52] Sheng Ye, Yu-Hui Wen, Yanan Sun, Ying He, Ziyang Zhang, Yaoyuan Wang, Weihua He, and Yong-Jin Liu. Audio-driven stylized gesture generation with flow-based model. In European Conference on Computer Vision, pages 712-728. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.552, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[53] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.608, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[54] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yan-dong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.691, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[55] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.761, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[56] Pengfei Zhang, Pinxin Liu, Hyeongwoo Kim, Pablo Garrido, and Bindita Chaudhuri. Kinmo: Kinematic-aware human motion understanding and generation. arXiv preprint arXiv:2411.15472, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.816, + 0.482, + 0.883 + ], + "angle": 0, + "content": "[57] Zeyi Zhang, Tenglong Ao, Yuyao Zhang, Qingzhe Gao, Chuan Lin, Baoquan Chen, and Libin Liu. Semantic gesticulator: Semantics-aware co-speech gesture synthesis. ACM Transactions on Graphics (TOG), 43(4):1-17, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.886, + 0.482, + 0.912 + ], + "angle": 0, + "content": "[58] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker:" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.093, + 0.483, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.882, + 0.147 + ], + "angle": 0, + "content": "Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.884, + 0.219 + ], + "angle": 0, + "content": "[59] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.884, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "13973" + } + ] +] \ No newline at end of file diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_origin.pdf b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c39e42544f9934aa0df4034d981dea714c149f5 --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/d193eee8-ef57-4a70-9609-6d91a9953312_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cafb5bef682ae8d821e531952f6236f5dff33cae1fa0c50c2440e4c6744b70e5 +size 5470833 diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/full.md b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1760ba671db9fc548d3702261c1a93545e312dff --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/full.md @@ -0,0 +1,337 @@ +# SemGes: Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning + +Lanmiao Liu $^{1,2,3}$ Esam Ghaleb $^{1,2}$ Asli Özyurek $^{1,2}$ and Zerrin Yumak $^{3}$ + +$^{1}$ Max Planck Institute for Psycholinguistics $^{2}$ Donders Institute for Brain Cognition and Behaviour $^{3}$ Utrecht University {lanmiao.liu, esam.ghaleb, asli.ozyurek}@mpi.nl z.yumak@uu.nl + +# Abstract + +Creating a virtual avatar with semantically coherent gestures that are aligned with speech is a challenging task. Existing gesture generation research mainly focused on generating rhythmic beat gestures, neglecting the semantic context of the gestures. In this paper, we propose a novel approach for semantic grounding in co-speech gesture generation that integrates semantic information at both fine-grained and global levels. Our approach starts with learning the motion prior through a vector-quantized variational autoencoder. Built on this model, a second-stage module is applied to automatically generate gestures from speech, text-based semantics and speaker identity that ensures consistency between the semantic relevance of generated gestures and co-occurring speech semantics through semantic coherence and relevance modules. Experimental results demonstrate that our approach enhances the realism and coherence of semantic gestures. Extensive experiments and user studies show that our method outperforms state-of-the-art approaches across two benchmarks in co-speech gesture generation in both objective and subjective metrics. The qualitative results of our model, code, dataset and pre-trained models can be viewed at https://semgesture.github.io/. + +# 1. Introduction + +Human language is inherently multimodal, with gestures and speech complementing each other to convey pragmatic and semantic information [21, 35]. Cospeech gestures are non-verbal cues that are uniquely related to co-occurring speech, pragmatically, semantically, and temporally. For example, representational iconic gestures that visually express the semantic content of speech and interact with spoken language [13, 14, 19, 21, 41]. A long-standing goal in Computer Vision is to create digital humans that use non-verbal cues in sync with speech. Gesture generation—synthesizing + +![](images/cd16a7daec2693e07ed1d0fd1d4ecb8a33f5cb99f942f7a9f8240ae716d35d3e.jpg) +Figure 1. SemGes integrates audio, text-based semantics, and speaker identity to produce both contextually relevant (discourse-level) and fine-grained (local) gestures. A semantic coherence module aligns text and motion embeddings. The multimodal consistency loss synchronizes the quantized multimodal representations to match the quantized learned motion features for final speech-driven semantics-aware gesture generation. The semantic relevance loss selectively emphasizes gestures with semantic annotations. + +movements from co-occurring speech, masked motion, or speaker identity—has advanced to enhance AI agents' expressiveness and realism [29]. However, much of the focus has gone into generating rhythmic beat gestures with limited semantic information, leaving representational gestures that convey semantic messages (e.g., iconic) less explored [29, 40]. + +Generating spontaneous and semantically rich gestures from speech comes with multiple challenges. First, it requires capturing global discourse-level information and local fine-grained details (e.g., salient words) to generate speech-driven gestures that reflect the intended meaning and align with speech temporally and semantically. Second, existing methods often generate repetitive and short sequences that do not span the full range of expressive motions required for natural communication. To leverage semantics when generating gestures, researchers have attempted to align motion with speech representations at a global level, e.g., by leveraging pre + +trained semantic representations such as CLIP [58] or focusing on semantically important keywords [6, 57]. Nonetheless, they often fail to (i) unify global and local semantic modelling within a single framework and (ii) exploit the relevance of the semantic information in guiding gesture generation [27]. At the same time, raw audio features and speaker identity are relevant to the timing and style of gestures. In this paper, we address these limitations by integrating speech, speech semantics, and gesturing style, exploiting semantic information at different levels. + +Specifically, we propose a two-stage framework, namely, SemGes, that integrates speech, text-based semantics, and speaker identity into a unified gesture-generation model (see Figure 1). In stage 1, we build motion prior of holistic gestures (i.e., body and hands) by training a vector-quantized variational autoencoder (VQ-VAE) to learn an efficient, compositional motion latent space. This stage results in a robust motion encoder & decoder and quantized codebooks that can reconstruct naturalistic gestures while allowing the reuse of learned codebook entries. Stage 2 leverages the learned motion priors to drive gesture synthesis by fusing three modalities using a cross-modal Transformer encoder: (i) text-based semantics, (ii) raw-audio speech features, and (iii) speaker identity for style consistency. We impose a semantic coherence loss that aligns text-based embeddings with the VQ-VAE motion latent space and a semantic relevance loss that emphasises representational gestures (e.g. iconic and metaphoric gestures). A multimodal consistency objective ensures the fused multimodal representations are compatible with the learned motion codebooks, enabling the generation of gestures that are both semantically rich and visually natural. Finally, we introduce a simple but effective long-sequence inference strategy that smoothly combines overlapping motion clips for extended durations. To summarize our contributions, + +- We introduce a novel framework, SemGes, that first learns a robust VQ-VAE motion prior for body and hand gestures, and then generates gestures driven by fused speech audio, text-based semantics, and speaker identity in a cross-modal transformer. +- Our method jointly captures discourse-level context via a semantic coherence loss and fine-grained representational gestures (e.g., iconic, metaphoric) via a semantic relevance loss. +- We propose an overlap-and-combine inference algorithm that maintains smooth continuity over extended durations. +- Extensive experiments on two benchmarks, namely, the BEAT [27] and TED Expressive [33] datasets show that our method outperforms recent baselines in both objective metrics (e.g., Fréchet Gesture Distance + +(FGD), diversity, semantic alignment) and user judgment of generated gestures. + +# 2. Related Work + +Data-driven Co-Speech Gesture Generation. Current gesture generation approaches are based on generative deep neural networks. These approaches use advanced models such as Transformers [28], Generative Adversarial Networks [23], Normalizing Flows [18, 30], Vector Quantized Variational Autoencoder(VQ-VAE) [16] and Denoising Diffusion Probabilistic Models [47]. In addition, researchers have explored the impact of different model inputs on the naturalness and appropriateness of generated gestures. Various modal inputs have been used, such as text [15], audio [50, 59], image [32, 39], and speaking style [3]. For a comprehensive survey, we refer to Nyatsanga et al. [40]. Although there have been significant improvements in this field, current methods fall short in generating semantically grounded gestures at a fine-grained level. In other words, while the generated motions look convincing at first glance, they do not match well with the meaning of the text, or they mostly focus on beat-type gestures. + +# Semantics-aware Co-Speech Gesture Generation. + +A group of work focused on semantics-aware gesture generation where the semantic information is handled in two ways: global semantics and local semantics. Methods that focus on global semantic information [10, 22, 58] align gestures with text or audio, but they fall short in generating gesture types matching the semantic context, such as iconic, metaphoric and deictic gestures. To capture a wider range of semantic gestures, works like [5, 6, 27] adopt local semantic-aware modelling by integrating the semantic salient words to the neural network. However, these approaches often fail to ensure that the generated gestures align with both the broader audio or textual context and a combination of global and local semantics. Liang et al. [26], Voß and Kopp [49] incorporate both global and local semantics, however, they require extensive annotations. Recently, Zhang et al. [57] employed a generative retrieval framework based on LLMs to address the sparsity problem in datasets with semantic gestures. However, they do not explicitly model the different types of gestures [27] or gesture phases [12] grounded in linguistic research. Moreover, there is still not enough understanding of the impact of different annotations and fine-grained semantics. + +Substantial research [6, 28, 57, 58] focused on two-stage latent space generative modelling to overcome the limitations of co-speech gesture generation and to generate more naturalistic and diverse gestures. + +![](images/71af008c2730c688b5250cfcba0cceaa9c17e67ca54a1d376357b4a68f77ac40.jpg) +Figure 2. We pre-train two VQ-VAEs by reconstructing body and hand motions with a dedicated codebook for each. + +These approaches first learn a latent space and then model gestures probabilistically, effectively integrating the strengths of different methods in different stages. Liu et al. [28], Zhang et al. [57] capture complex dependencies in the latent space using VQ-VAE while Zhi et al. [58] employs CLIP [46, 56] to align text and motion embeddings. Ao et al. [6] introduces a diffusion-based model that leverages semantic awareness, while Liu et al. [28] utilizes a transformer-based approach to generate holistic body gestures. In contrast with two-stage generative modeling approaches, end-to-end methods such as [5, 59], are prone to jittering artifacts, especially in hand-motion generation. + +Our model contributes to the research line on semantics-aware co-speech gesture generation by taking into account both global and local semantics. Inspired by the previous work, we employ a two-stage latent space generative modelling for high-quality motion representation. We learn semantic coherence between text and gestures globally with cosine similarity. Moreover, our model takes into account the semantic relevancy of gesture types with minimally required annotations. In contrast with other semantic learning models, we focus on annotations with different gesture types embedded in linguistic research. Our work is closest to CAMN [27] in that sense; however, CAMN does not include semantic coherence learning by aligning text and gestures' latent space globally. + +# 3. Methodology + +We propose a two-stage approach that generates cospeech gestures by grounding them in raw speech, text-based semantics, and speaker identity. In Section 3.1, we introduce a VQ-VAE encoder-decoder that learns a robust motion prior. Section 3.2 details our gesture synthesis and inference pipeline based on speech, semantics, and identity. + +Problem formulation. Our goal is to generate hand gestures $\pmb{G}^{h} = (g_{1}^{h},\dots,g_{T}^{h})\in \mathbb{R}^{T\times J}$ and body gestures $\pmb{G}^{b} = (g_{1}^{b},\dots,g_{T}^{b})\in \mathbb{R}^{T\times J}$ , where $T$ is the number of time steps and $J$ the number of joints (e.g., 38 for hands, 9 for body). Each motion vector $g_{t}^{h}$ or $g_{t}^{b}$ is encoded in a Rot6D representation, capturing joint rotations at time $t$ . + +To model human motion of body and hands, we first learn a motion generator $\mathcal{M}_g$ (Stage 1), which synthesizes a plausible motion sequence: + +$$ +\arg \min _ {\mathcal {M} _ {g}} \left\| \boldsymbol {G} - \mathcal {M} _ {g} \left(g _ {1}, \dots , g _ {T}\right) \right\|. \tag {1} +$$ + +Next, we condition on (i) the raw input audio $\mathbf{A} = (a_{1},\ldots ,a_{T})$ , (ii) the speaker identity embedding $I$ , and (iii) the text-based semantic embeddings of the speech $S = (s_{1},\dots ,s_{T})$ . Our second-stage model $\mathcal{M}_{a,s,i}$ uses these inputs to generate a latent sequence that the motion generator $\mathcal{M}_g$ then decodes into naturalistic gestures: + +$$ +\arg \min _ {\mathcal {M} _ {a, s, i}} \left\| \boldsymbol {G} - \mathcal {M} _ {g} \left(\mathcal {M} _ {a, s, i} (\boldsymbol {A}, \boldsymbol {S}, I)\right) \right\|. \tag {2} +$$ + +# 3.1. Stage 1: Learning Efficient Codebooks & Compositional Motion Priors + +Realistic co-speech gestures require modelling the sequential motion of both body and hand joints. Rather than learning a single representation for the entire body, we adopt a compositional approach, using a discrete codebook of learned representations specific to each part (hands & body). Any gesture motion can then be represented by selecting appropriate codebook entries. Following [28, 48, 53], we employ a VQ-VAE architecture (see Fig. 2) with encoder $\mathcal{E}_m$ and decoder $\mathcal{D}_m$ . Given hand motion $G^h \in \mathbb{R}^{T \times J}$ and body motion $G^b \in \mathbb{R}^{T \times J}$ , the encoder produces latent vectors $\hat{z}^h$ and $\hat{z}^b$ , which are quantized by selecting the nearest entries in the codebooks. Formally, + +$$ +\mathbf {q} (\hat {\boldsymbol {z}}) = \arg \min _ {z ^ {i} \in \mathcal {Z}} \| \hat {z} ^ {j} - z ^ {i} \|, \tag {3} +$$ + +where $z^i$ are the learned codebook entries, and $\hat{z}^j$ denotes an element of the latent vector for either hand or body. We train the VQ-VAE via a straight-through gradient estimator, minimizing: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {V Q - V A E}} = \left\| \mathbf {g} - \hat {\mathbf {g}} \right\| ^ {2} + \left\| \dot {\mathbf {g}} - \hat {\dot {\mathbf {g}}} \right\| ^ {2} + \left\| \ddot {\mathbf {g}} - \hat {\ddot {\mathbf {g}}} \right\| ^ {2} \tag {4} \\ + \left\| \operatorname {s g} [ \mathbf {E} (\mathbf {g}) ] - \mathbf {q} (\hat {z}) \right\| ^ {2} + \left\| \mathbf {E} (\mathbf {g}) - \operatorname {s g} [ \mathbf {q} (\hat {z}) ] \right\| ^ {2}, \\ \end{array} +$$ + +where the first three terms reconstruct joint positions, velocities, and accelerations, and the last two terms implement the VQ-VAE commitment loss [48]. + +By the end of this stage, we have motion $(m)$ encoder $(\mathcal{E}_m)$ , decoder $(\mathcal{D}_m)$ and codebooks $(\text{Quant}^m(\cdot))$ for hands and body. In the next section (Section 3.2), we show how this discretized motion of hands and body guides speech, semantics and speaker identity-driven generation to produce realized co-speech gestures. + +# 3.2. Stage 2: Speech and Identity Driven Semantic Gesture Generator + +This stage focuses on generating gestures conditioned on three inputs: speech embeddings, text-based seman + +![](images/f390537d0780d93a79329de2bb3a674c27d7e5f1c1a52b19a07db673cc813312.jpg) +Figure 3. SemGes employs three training pathways: (1) Global semantic coherence, which minimizes latent disparities between gesture and text encoders; (2) Multimodal Quantization learning, where integrated multimodal representation codes are aligned with quantized motion to decode them into hand and body movements; and (3) Semantic relevance learning, which emphasizes semantic gestures. + +tic embeddings, and speaker identity. As illustrated in Figure 3, the second-stage architecture has three main modules, which we elaborate on in the following subsections. + +![](images/79def40ab4ae068b8c0a3781378d3b936df9ae93d335c8b434390dd14871c50a.jpg) +Figure 4. Semantic Coherence Embedding Learning. + +# 3.2.1. Semantic Coherence Embedding Learning + +To align text-based semantics with motion embeddings, we introduce a shared embedding space for both motion priors and speech transcripts. Specifically, we embed word tokens using a pre-trained FastText model [8], then feed these embeddings into a trainable text-based semantic encoder $\mathcal{E}_s$ . At the same time, we use the pre-trained motion encoder $\mathcal{E}_m$ from Stage 1 to encode ground-truth gesture sequences. Thus, for a batch of paired (gesture, transcript) samples, we get: + +$$ +\mathcal {Z} ^ {S} = \mathcal {E} _ {s} (S), \quad \mathcal {Z} ^ {h} = \mathcal {E} _ {m} ^ {h} (G ^ {h}), \quad \mathcal {Z} ^ {b} = \mathcal {E} _ {m} ^ {b} (G ^ {b}), \tag {5} +$$ + +where $S$ is the tokenized speech transcript, and $G^{h}$ and $G^{b}$ correspond to the ground-truth hand gesture sequence and body gesture sequence, respectively. $\mathcal{Z}^h$ and + +$\mathcal{Z}^b$ represent the hand and body ground-truth motion encodings from Stage 1, and $\mathcal{Z}^s$ denotes the text-based semantic encoder output. + +Semantic Coherence Loss. We maximize the similarity of correct (gesture, transcript) pairs and minimize it for mismatched pairs, enforcing semantic coherence. This aligns gestures and textual semantics in a common space while keeping $\mathcal{E}_m$ frozen, as illustrated in Figure 4. We impose the semantic coherence constraint separately on both hand and body movements to align gestures with transcripts in the shared embedding space. Specifically, we introduce two distinct cosine similarity losses: one between the text encoder output and the hand motion latent encoding and another between the text encoder output and the body motion latent encoding. Formally, we minimize: + +$$ +\mathcal {L} _ {\text {s e m a n t i c - c o h e r e n c e}} = 1 - \cos \left(\mathcal {Z} ^ {h}, \mathcal {Z} ^ {s}\right) + 1 - \cos \left(\mathcal {Z} ^ {b}, \mathcal {Z} ^ {s}\right), \tag {6} +$$ + +where the function $\cos (\cdot ,\cdot)$ measures cosine similarity. + +# 3.2.2. Crossmodal Integration + +SemGes supports multi-modal inputs in the second training stage by combining audio features and speaker identity with semantic text embeddings using a Transformer encoder with self and cross-attention layers (see Figure 3). We begin by extracting HuBERT features [20] from raw speech (keeping the HuBERT encoder frozen). We concatenate audio features $\mathcal{Z}^a$ and speaker embeddings $\mathcal{Z}^i$ , resulting in $\mathcal{Z}^r$ , which we feed into a self-attention layer. + +Next, we use a cross-attention layer that takes $\mathcal{Z}^r$ as the query and the motion-aligned text-based semantic features $\mathcal{Z}^s$ as the key-value pair. The final hidden representation $\mathcal{Z}^f$ serves as the multimodal latent code + +that drives gesture synthesis when passed to our vector quantization and VQ-VAE-based motion decoder, which is learned in our first stage (see the yellow box in Figure 3). + +Multimodal Quantization Consistency Loss. SemGes quantizes the multimodal latent code using separate hand and body codebooks. To align this code with the ground-truth motion latent codes, we apply independent quantization consistency losses for each component. Specifically, the quantization loss is defined as: + +$$ +\mathcal {L} _ {\text {q u a n t i z a t i o n}} = \left\| Q u a n t ^ {h} \left(\mathcal {Z} ^ {f}\right) - Q u a n t ^ {h} \left(\mathcal {Z} ^ {h}\right) \right\| ^ {2} + \tag {7} +$$ + +$$ +\left\| Q u a n t ^ {b} \left(\mathcal {Z} ^ {f}\right) - Q u a n t ^ {b} \left(\mathcal {Z} ^ {b}\right) \right\| ^ {2} \tag {17} +$$ + +where $\text{Quant}^h(\cdot)$ and $\text{Quant}^b(\cdot)$ denote the quantization functions for the hand and body codebooks, respectively. + +The multimodal quantization loss aligns the integrated latent code $\mathcal{Z}^f$ with the learned motion code, a critical step since gesture synthesis is obtained through the quantized multimodal representation. Specifically, $\mathcal{Z}^f$ is vector-quantized using separate hand and body codebooks before being decoded by their respective VQ decoders. This process ensures that both hand and body movements contribute effectively to the final output. Formally, the generated gestures are given by: + +$$ +\hat {G} = \hat {G} ^ {h} \oplus \hat {G} ^ {b} = \mathcal {D} _ {m} ^ {h} \left(\operatorname {Q u a n t} ^ {h} \left(\mathcal {Z} ^ {f}\right)\right) \oplus \mathcal {D} _ {m} ^ {b} \left(\operatorname {Q u a n t} ^ {b} \left(\mathcal {Z} ^ {f}\right)\right), \tag {8} +$$ + +where $\oplus$ denotes concatenation, jointly synthesizing hand and body motions (i.e. $\hat{G}$ ). + +# 3.2.3. Gesture Semantic Relevance Loss + +To prioritize the generation of semantically meaningful gestures (e.g., iconic, metaphoric, or deictic), which are less frequent than beat gestures, we introduce a semantic relevance loss. This loss emphasizes semantic annotations while preventing over-penalization of minor deviations. Formally, it is defined as: + +$$ +\mathcal {L} _ {\text {s e m a n t i c - r e l e v a n c e}} = \mathbb {E} \left[ \lambda \Psi (\mathbf {G} - \hat {\mathbf {G}}) \right], \tag {9} +$$ + +where $\lambda$ is the annotation relevance factor, and $\Psi(\cdot)$ is a piecewise function that applies a quadratic penalty for small errors and a linear penalty for larger ones: + +$$ +\Psi (\mathbf {G} - \hat {\mathbf {G}}) = \left\{ \begin{array}{l l} \frac {1}{2} (\mathbf {G} - \hat {\mathbf {G}}) ^ {2}, & \text {i f} | \mathbf {G} - \hat {\mathbf {G}} | < \alpha , \\ \alpha \left(| \mathbf {G} - \hat {\mathbf {G}} | - \frac {1}{2} \alpha\right), & \text {o t h e r w i s e}, \end{array} \right. \tag {10} +$$ + +with $\alpha = 0.01$ + +Combined Objective Functions. Finally, the overall objective is: + +$$ +\mathcal {L} _ {\text {S e m G e s}} = \mathcal {L} _ {\text {s e m a n t i c - c o h e r e n c e}} + \mathcal {L} _ {\text {s e m a n t i c - r e l e v a n c e}} + \mathcal {L} _ {\text {q u a n t i z a t i o n}}, \tag {11} +$$ + +Algorithm 1 Long Gesture Sequence Algorithm +Require: Audio $\mathcal{A}$ aligned speech transcript $S$ and speaker ID $\mathcal{I}$ Pre-trained codebooks and motion decoder (Stage 1) +Ensure: Long-sequence gesture M +1: Partition $(\mathcal{A},\mathcal{S},\mathcal{I})$ into clips $\{(\mathcal{A}_c,\mathcal{S}_c,\mathcal{I}_c)\}_{c = 1}^C$ +2: Compute latent representation: $\mathcal{Z}^{f}\gets$ Encode(A,S,I) +3: Quantize: $\mathcal{Z}^e\gets$ VectorQuantize $(\mathcal{Z}^{f})$ +4: Decode initial clip: $\hat{M}_1\gets \mathrm{Dec}(\mathcal{Z}^e)$ +5: for each clip $c = 2$ to $C$ do +6: Set first 4 frames of $\hat{M}_c$ to the last 4 frames of $\hat{M}_{c - 1}$ +7: Generate remaining frames of $\hat{M}_c$ +8: end for +9: return $\hat{M}$ + +which jointly optimizes the model to generate gestures that are semantically coherent at both global and fine-grained levels while remaining faithful to the Stage 1 motion prior. + +# 3.3. Inference of Long Gesture Sequences + +Generating long sequences of gestures is challenging due to the need to maintain coherence and smooth transitions. Our Long-Sequence Gesture Motion algorithm (Alg. 1) addresses these challenges by partitioning the input speech, transcript, and speaker identity into aligned clips. For each clip, a multimodal latent representation is computed using our cross-modal encoder, vector-quantized via the Stage 1 codebooks, and decoded into gesture motions. Overlapping 4-frame segments between clips provides continuity, resulting in extended, naturalistic gesture sequences. + +# 4. Experimental Setup + +# 4.1. Datasets + +Our proposed methodology is evaluated on two benchmarks, namely, BEAT [27] and the TED expressive dataset [33]. The BEAT dataset consists of 76 hours of multimodal recordings, which include speech audio recordings, speech transcriptions, and, more importantly, motion data collected from 30 participants, leveraging Motion Capture (MOCAP) technology. The participants expressed emotions in eight distinct scenarios across four languages. The motion data contains joint rotation angles, which were designed for consistency across varying body sizes. The TED Expressive dataset [33] is segmented from TED Talk videos into smaller shots based on scene boundaries. Liu et al. [33] extracted each frame's 2D human pose using OpenPose BEAT [9]. Using these 2D pose priors, ExPose [43] was + +Table 1. Comparison of SemGesGen with other methods on the BEAT and TED-Expressive datasets. For BEAT, we compare with CaMN [27], DiffGesture [59], LivelySpeaker [58], and DiffSheg [10]. The same methods are evaluated on TED-Expressive. SRGR is not applicable (denoted with $-$ ) for TED-Expressive as it does not contain annotations for semantic relevance of gestures. + +
BEATTED-Expressive
MethodFGD ↓BC ↑Diversity ↑SRGR ↑MethodFGD ↓BC ↑Diversity ↑
CaMN [27]8.5100.797206.7890.231CaMN [27]7.6730.642156.236
DiffGesture [59]9.6320.876210.6780.106DiffGesture [59]9.3260.662119.889
LivelySpeaker [58]13.3780.891214.9460.229LivelySpeaker [58]8.1450.691119.231
DiffSheg [10]6.6230.922257.6740.250DiffSheg [10]8.4570.712108.972
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
+ +employed to annotate the 3D upper body keypoints, including 13 upper body joints and 30 finger joints. Both datasets' training and validation samples are divided into 34-frame clips. + +Cross-Validation. We evaluate our approach on the BEAT dataset, following the protocol in [27], where the data is randomly split into a 19:2:2 ratio for training, validation, and testing. Similarly, for the TED Expressive dataset, we adapt the protocol in [33], using a random split of 8:1:1 for training, validation, and testing. + +Implementation Details. The details of the model architectures and training are provided in Section 2 of the Supplementary Materials. + +# 4.2. State-of-the-Art Baselines + +We compare SemGes against a set of representative state-of-the-art models that focus on semantic-driven gesture generation. The selected models achieved strong performance on the BEAT and TED-Expressive datasets, making them suitable for a fair comparison with our method. The selected models are as follows: + +1. Cascaded Motion Network(CaMN) [27] is the current benchmark model for the BEAT dataset. CaMN is based on LSTMs and integrates multiple input modalities, including audio, text, facial expressions, and emotion. Additionally, like SemGes, it leverages semantic relevance annotations to enhance gesture generation. +2. DiffSHEG [10] is a state-of-the-art diffusion-based model for real-time speech-driven holistic gesture generation. It is conditioned on noisy motion, audio, and speaker ID. DiffSHEG introduces a Fast Out-painting-based Partial Autoregressive Sampling method to efficiently generate arbitrary-length sequences in real time. +3. LivelySpeaker [58] generates semantically and rhythmically aware co-speech gestures by leveraging an MLP-based diffusion model. The model conditions gesture generation on text, noised motion, + +speaker ID, and audio to enable text-driven gesture control while incorporating global semantics. + +4. DiffGes [59] models the diffusion and denoising processes within the gesture domain, enabling the generation of high-fidelity, audio-driven gestures conditioned on both audio and gesture inputs. Several recent studies[6, 26] have also demonstrated strong performance in this area. + +We exclude certain models from our comparison. For instance, SEEG [26] and [57] rely on additional data annotations (e.g., Semantic Prompt Gallery or ChatGPT-generated annotations) that are not uniformly available. In addition, other works, such as Ao et al. [6], Pang et al. [42], Zhang et al. [57], are excluded from our analysis due to the inaccessibility of their codebase. Voß and Kopp [49] is omitted due to its high computational cost and the unavailability of annotations. Liu et al. [31, 34], Mughal et al. [36, 37], Ng et al. [38], Yi et al. [54] are excluded as they primarily focus on holistic gestures with face and mesh data, which fall outside the scope of this work. Similarly, Chhatre et al. [11], Qi et al. [44] are excluded, as their emphasis lies in emotion-driven gesture generation rather than the semantic aspects. Furthermore, Ahuja et al. [1, 2], Alexander et al. [4], Habibie et al. [17], Liu et al. [33], Sun et al. [45], Yang et al. [51], Ye et al. [52] are omitted due to their lack of relevance to semantic-driven gesture generation. + +# 5. Quantitative Objective Evaluations + +Evaluation Metrics. We employ four standard objective metrics for evaluating the quality of gesture generation, namely, Fréchet Gesture Distance (FGD) [55], Beat Consistency Score (BC) [25], Diversity [24], and Semantic-Relevant Gesture Recall (SRGR) [27]. + +FGD measures how the generated gestures resemble real motion distributions by embedding sequences into a latent space via a pre-trained autoencoder. In contrast, BC focuses on synchronization with speech, measuring the alignment between speech onsets (audio beats) and motion beats, which are identified as velocity minima in + +Table 2. Ablation studies evaluating the contributions of key components in SemGes on the BEAT and TED-Expressive Datasets. For BEAT, performance is measured using FGD (lower is better), BC, Diversity, and SRGR, while for TED-Expressive, SRGR is not applicable (denoted as -). + +
BEATTED-Expressive
Model VariantsFGD ↓BC ↑Diversity ↑SRGR ↑Model VariantsFGD ↓BC ↑Diversity ↑
Baseline (VQVAE)10.3480.564198.5680.176Baseline (VQVAE)10.6820.612114.692
w/o Semantic Coherence Module8.0530.556249.5500.180w/o Semantic Coherence Module7.9240.623109.256
w/o Semantic Relevance Module7.5490.573245.3190.195w/o Semantic Relevance Module---
w/ SpeechCLIP Encoder6.7870.468289.6210.245w/ SpeechCLIP Encoder7.3410.605245.680
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
+ +upper-body joints (excluding fingers). Meanwhile, **Diversity** captures the variability of generated motions by computing the average $L1$ distance between pairs of $N$ generated clips. Finally, SRGR assesses semantic relevance by determining how well generated gestures align with the annotated semantic gestures. Further details on the objective metrics are included in the Supplementary Materials (Section 1). + +Comparisons with Other Models. Table 1 compares the performance of our approach against four baseline methods across four evaluation metrics. As highlighted in the table, SemGes outperforms the baselines in FGD, Diversity, and SRGR. + +For the BEAT dataset, our approach achieves the highest SRGR, which we attribute to the exploitation of semantic relevance information in our training objectives. In addition, our approach shows a significant improvement in FGD and Diversity, indicating a closer alignment with the ground truth gesture distribution and a broader range of generated gestures compared to the second-best baselines. The performance on the Beat Consistency (BC) metric is lower for our method. This is expected given our focus on improving semantic awareness of the generated gestures rather than optimizing for strict temporal alignment between rhythmic beat gestures and speech. In addition, the BC metric can be sensitive to rapid, jittery movements; even minor motion artefacts may be mistakenly counted as additional beats, thereby increasing the BC score artificially- a phenomenon also observed in the diffusion-based baselines, as further illustrated in our supplementary video. + +We evaluate how our model handles the trade-off between semantic and beat scores by testing the model on beat-dominant gestures (without semantic content). The results show a significantly higher Beat score (0.689) than the full dataset Beat score (0.453). This confirms rhythmic consistency in beat-focused contexts. We provide additional evaluation in the supplementary materials (Section 3) to show how the model handles difficult cases (such as noisy speech or misaligned speech). + +Note that the TED Expressive dataset lacks annotations for gesture semantic relevance, so SRGR is not + +applicable, and the semantic relevance loss was omitted during training. Nevertheless, SemGes produces diverse, naturalistic gestures on TED Expressive, outperforming baselines in FGD and Diversity metrics. + +Ablation Study. We evaluate the contributions of key components in SemGes through ablation experiments. First, we assess a baseline VQ-VAE model (Stage 1 only), which uses two stacked encoder-decoder blocks and an MLP. In this experiment, we test its ability to generate gestures, conditioned on audio, masked motion, and speaker identity. As shown in Table 2, this baseline underperforms compared to state-of-the-art methods (Table 1). As a result, we motivate our two-stage design where the VQ-VAE is reserved to learn the motion latent space and Stage 2 leverages speech and identity conditioning to generate gestures. + +Next, we examine Stage 2 by removing its components: (i) the Semantic Coherence Loss, (ii) the Semantic Relevance Loss, and (iii) by replacing the HuBERT-based speech encoder with SpeechCLIP. Results in Table 2 show that removing either the Semantic Coherence or Relevance Loss degrades FGD, Diversity, and SRGR scores, highlighting their roles in aligning gesture representations with textual semantics and capturing semantic importance. In addition, replacing the speech encoder results in marginal gains. The semantic encoder is fixed as FastText, which we believe is sufficient to capture the necessary semantic information [7]. Overall, these results confirm the importance of each module in generating semantics-aware gestures. + +# 6. Qualitative & Subjective Evaluations + +Visualization Comparisons. Before presenting the subjective ratings of the generated gestures, Figure 5 provides a visual comparison between the ground truth, results from our approach and two baseline models. We use examples from the BEAT dataset. It is clear from the figure that our approach not only achieves better speech-gesture alignment but also produces gestures that are more naturalistic, diverse, and semantically aware. For example, while CaMN generates smooth movements, its gestures tend to be slower and less varied compared to + +![](images/b4028ad1da7211a158b3965e028ce1b10b9d5550fdc235e8203825ca3cebacf9.jpg) +Figure 5. Comparisons with baselines and ground truth gestures. Compared to the baseline method, our approach generates gestures that are aligned with speech content (semantics). For instance, when the speaker says "remix", our method produces gestures where the character raises both hands to emphasize the word before gradually lowering them—a movement that other methods fail to achieve. Similarly, when uttering "first", our method generates a raised hand gesture, producing an iconic gesture. + +our model. Additionally, the baseline methods show varying degrees of jitter—DiffGesture shows the highest jitter, followed by LivelySpeaker and DiffSheg, with CaMN displaying the least. Although CaMN includes semantic information, our approach strikes a more effective balance, generating gestures that align with actual motion, as shown also with the objective metrics. Based on these qualitative observations, our subsequent rating study focuses on evaluating gestures produced by the ground truth, our model, CaMN, and DiffSHEG. + +User Ratings of Generated Gestures. We conducted a user study using 40-second video clips from the BEAT test set, featuring subjects narrating six topics. Thirty native English speakers from the United Kingdom and the United States participated, with an average age of $36 \pm 20$ years and a female-to-male ratio of approximately 2:1. Each participant evaluated 24 videos generated by the ground truth, CaMN, DiffSHEG, and our model over a study duration of that lasted on average $27 \pm 5$ minutes. For data quality, participants were required to pass attention verification questions, i.e., correctly answering at least two out of four questions regarding the narration topic. Participants rated the videos on three criteria: naturalness, diversity, and alignment with speech content and timing on a scale from 1 to 5. The videos were presented in a randomized order to avoid bias. In Section 3 of the Supplementary Materials, we provide screenshots and more details on the user study and interface. + +Figure 6 shows that ground-truth gestures received an average rating of 4 across all metrics, establishing an upper bound and validating the participant survey. Our model received the highest ratings among the generated gestures, significantly outperforming CaMN and Diff-SHEG in naturalness, synchronization, and diversity (indicated by the in Figure 6). These results prove that our approach produces gestures that are more natural, better aligned with speech, and more diverse than those gener + +ated by SOTA baselines. + +![](images/e3d5c4222b692542165121d51ee5f693894562de6161cdae88f72d56ef8b033e.jpg) +Figure 6. Average ratings of users for ground truth gestures and gestures generated through our approach, CAMN, and DiffSHEG. The bars illustrate the average user ratings across three metrics: naturalness, diversity, and alignment with speech content and timing. Statistical t-tests show that our approach received significantly higher ratings than CAMN and DiffSHEG, with $p < 0.05$ . + +# 7. Conclusion + +We proposed SemGes, a novel two-stage approach to semantic grounding in co-speech gesture generation by integrating semantic information at both fine-grained and global levels. In the first stage, a motion prior generation module is trained using a vector-quantized variational autoencoder to produce realistic and smooth gesture motions. Building upon this model, the second stage generates gestures from speech, text-based semantics, and speaker identity while maintaining consistency between gesture semantics and co-occurring speech through semantic coherence and relevance modules. Subjective and objective evaluations show that our work achieves state-of-the-art performance across two public benchmarks, generating semantics-aware and diverse gestures. Future direction and limitations are discussed in Section 5 of the Supplementary Materials. + +# Acknowledgement + +The project is funded by the Max Planck Society. We thank Sachit Misra for his invaluable assistance with rendering Avatar characters. We extend our gratitude to the members of the Multimodal Language Department at Max Planck Institute for Psycholinguistics for their feedback. + +# References + +[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566-20576, 2022. 6 +[2] Chaitanya Ahuja, Pratik Joshi, Ryo Ishii, and Louis-Philippe Morency. Continual learning for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20893-20903, 2023. 6 +[3] Simon Alexander, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2 +[4] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audiodriven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 6 +[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6): 1-19, 2022. 2, 3 +[6] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2, 3, 6 +[7] Ben Athiwaratkun, Andrew Gordon Wilson, and Anima Anandkumar. Probabilistic fasttext for multi-sense word embeddings. arXiv preprint arXiv:1806.02901, 2018. 7 +[8] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 4 +[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291–7299, 2017. 5 +[10] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In CVPR, 2024. 2, 6 +[11] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al. + +Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 6 +[12] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Adversarial gesture generation with realistic gesture phasing. Computers & Graphics, 89:117-130, 2020. 2 +[13] Esam Ghaleb, Bulat Khaertdinov, Wim Pouw, Marlou Rasenberg, Judith Holler, Asli Ozyurek, and Raquel Fernandez. Learning co-speech gesture representations in dialogue through contrastive learning: An intrinsic evaluation. In Proceedings of the 26th International Conference on Multimodal Interaction, pages 274-283, 2024. 1 +[14] Esam Ghaleb, Bulat Khaertdinov, Asli Özyürek, and Raquel Fernández. I see what you mean: Co-speech gestures for reference resolution in multimodal dialogue. In Proceedings of the 63rd Conference of the Association for Computational Linguistics (ACL Findings), 2025. To appear. 1 +[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2 +[16] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2 +[17] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 6 +[18] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2 +[19] Judith Holler and Stephen C Levinson. Multimodal language processing in human communication. Trends in Cognitive Sciences, 23(8):639-652, 2019. 1 +[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451-3460, 2021. 4 +[21] Adam Kendon. Gesture units, gesture phrases and speech. In *Gesture: Visible Action as Utterance*, chapter 7, page 108–126. Cambridge University Press, 2004. 1 +[22] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2 + +[23] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1279, 2022. 2 +[24] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 6 +[25] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6 +[26] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2, 6 +[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6 +[28] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 2, 3 +[29] Li Liu, Lufei Gao, Wentao Lei, Fengji Ma, Xiaotian Lin, and Jinting Wang. A survey on deep multi-modal learning for body language recognition and generation. arXiv preprint arXiv:2308.08849, 2023. 1 +[30] Lanmiao Liu, Chuang Yu, Siyang Song, Zhidong Su, and Adriana Tapus. Human gesture recognition with a flow-based model for human robot interaction. In *Compan- ion of the 2023 ACM/IEEE International Conference on Human-Robot Interaction*, pages 548–551, 2023. 2 +[31] Pinxin Liu, Luchuan Song, Junhua Huang, Haiyang Liu, and Chenliang Xu. Gesturelsm: Latent shortcut based co-speech gesture generation with spatial-temporal modeling. arXiv preprint arXiv:2501.18898, 2025. 6 +[32] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. Advances in Neural Information Processing Systems, 35:21386-21399, 2022. 2 +[33] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for co-speech gesture generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10462-10472, 2022. 2, 5, 6 +[34] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated + +holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 6 +[35] David McNeill. Hand and mind. Advances in Visual Semiotics, 351, 1992. 1 +[36] Muhammad Hamza Mughal, Rishabh Dabral, Ikhsanul Habibie, Lucia Donatelli, Marc Habermann, and Christian Theobalt. Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1388-1398, 2024. 6 +[37] M Hamza Mughal, Rishabh Dabral, Merel CJ Scholman, Vera Demberg, and Christian Theobalt. Retrieving semantics from the deep: an rag solution for gesture synthesis. arXiv preprint arXiv:2412.06786, 2024. 6 +[38] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 6 +[39] Mang Ning, Mingxiao Li, Jianlin Su, Haozhe Jia, Lanmiao Liu, Martin Beneš, Wenshuo Chen, Albert Ali Salah, and Itir Onal Ertugrul. Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2412.15032, 2024. 2 +[40] Simbarashe Nyatsanga, Taras Kucherenko, Chaitanya Ahuja, Gustav Eje Henter, and Michael Neff. A comprehensive review of data-driven co-speech gesture generation. In Computer Graphics Forum, pages 569-596. Wiley Online Library, 2023. 1, 2 +[41] Asli Özyürek. Hearing and seeing meaning in speech and gesture: Insights from brain and behaviour. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1651):20130296, 2014. 1 +[42] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5 +[44] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10424-10434, 2024. 6 +[45] Mingyang Sun, Mengchen Zhao, Yaqing Hou, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2331-2340, 2023. 6 + +[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 3 +[47] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2022. 2 +[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3 +[49] Hendric Voß and Stefan Kopp. Augmented co-speech gesture generation: Including form and meaning features to guide learning-based gesture synthesis. In Proceedings of the 23rd ACM International Conference on Intelligent Virtual Agents, pages 1-8, 2023. 2, 6 +[50] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven cospeech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2 +[51] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2321-2330, 2023. 6 +[52] Sheng Ye, Yu-Hui Wen, Yanan Sun, Ying He, Ziyang Zhang, Yaoyuan Wang, Weihua He, and Yong-Jin Liu. Audio-driven stylized gesture generation with flow-based model. In European Conference on Computer Vision, pages 712-728. Springer, 2022. 6 +[53] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 3 +[54] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yan-dong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 6 +[55] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 6 +[56] Pengfei Zhang, Pinxin Liu, Hyeongwoo Kim, Pablo Garrido, and Bindita Chaudhuri. Kinmo: Kinematic-aware human motion understanding and generation. arXiv preprint arXiv:2411.15472, 2024.3 +[57] Zeyi Zhang, Tenglong Ao, Yuyao Zhang, Qingzhe Gao, Chuan Lin, Baoquan Chen, and Libin Liu. Semantic gesticulator: Semantics-aware co-speech gesture synthesis. ACM Transactions on Graphics (TOG), 43(4):1-17, 2024. 2, 3, 6 +[58] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: + +Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 3, 6 +[59] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 6 \ No newline at end of file diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/images.zip b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4d1f6499355709699bfeb1701e678893d6b7def --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37caf8fb614e1ea0ff177d6dbd1fc5b984bdfc144501d18b65c22b2da3e83f00 +size 342436 diff --git a/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/layout.json b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2c162e079ecbd3072e9e48dd8bedb7e475c4983 --- /dev/null +++ b/2025/SemGes_ Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning/layout.json @@ -0,0 +1,8439 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 533, + 140 + ], + "type": "text", + "content": "SemGes: Semantics-aware Co-Speech Gesture Generation using Semantic Coherence and Relevance Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "text", + "content": "Lanmiao Liu" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "text", + "content": " Esam Ghaleb" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "text", + "content": " Asli Özyurek" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "text", + "content": " and Zerrin Yumak" + }, + { + "bbox": [ + 144, + 160, + 487, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "spans": [ + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "text", + "content": "Max Planck Institute for Psycholinguistics " + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "text", + "content": "Donders Institute for Brain Cognition and Behaviour " + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 176, + 557, + 203 + ], + "type": "text", + "content": "Utrecht University {lanmiao.liu, esam.ghaleb, asli.ozyurek}@mpi.nl z.yumak@uu.nl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 231, + 206, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 231, + 206, + 243 + ], + "spans": [ + { + "bbox": [ + 158, + 231, + 206, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 257, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 297, + 544 + ], + "type": "text", + "content": "Creating a virtual avatar with semantically coherent gestures that are aligned with speech is a challenging task. Existing gesture generation research mainly focused on generating rhythmic beat gestures, neglecting the semantic context of the gestures. In this paper, we propose a novel approach for semantic grounding in co-speech gesture generation that integrates semantic information at both fine-grained and global levels. Our approach starts with learning the motion prior through a vector-quantized variational autoencoder. Built on this model, a second-stage module is applied to automatically generate gestures from speech, text-based semantics and speaker identity that ensures consistency between the semantic relevance of generated gestures and co-occurring speech semantics through semantic coherence and relevance modules. Experimental results demonstrate that our approach enhances the realism and coherence of semantic gestures. Extensive experiments and user studies show that our method outperforms state-of-the-art approaches across two benchmarks in co-speech gesture generation in both objective and subjective metrics. The qualitative results of our model, code, dataset and pre-trained models can be viewed at https://semgesture.github.io/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 569, + 149, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 590, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 296, + 723 + ], + "type": "text", + "content": "Human language is inherently multimodal, with gestures and speech complementing each other to convey pragmatic and semantic information [21, 35]. Cospeech gestures are non-verbal cues that are uniquely related to co-occurring speech, pragmatically, semantically, and temporally. For example, representational iconic gestures that visually express the semantic content of speech and interact with spoken language [13, 14, 19, 21, 41]. A long-standing goal in Computer Vision is to create digital humans that use non-verbal cues in sync with speech. Gesture generation—synthesizing" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 237, + 531, + 361 + ], + "blocks": [ + { + "bbox": [ + 328, + 237, + 531, + 361 + ], + "lines": [ + { + "bbox": [ + 328, + 237, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 328, + 237, + 531, + 361 + ], + "type": "image", + "image_path": "cd16a7daec2693e07ed1d0fd1d4ecb8a33f5cb99f942f7a9f8240ae716d35d3e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 541, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 541, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 541, + 477 + ], + "type": "text", + "content": "Figure 1. SemGes integrates audio, text-based semantics, and speaker identity to produce both contextually relevant (discourse-level) and fine-grained (local) gestures. A semantic coherence module aligns text and motion embeddings. The multimodal consistency loss synchronizes the quantized multimodal representations to match the quantized learned motion features for final speech-driven semantics-aware gesture generation. The semantic relevance loss selectively emphasizes gestures with semantic annotations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 541, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 541, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 541, + 578 + ], + "type": "text", + "content": "movements from co-occurring speech, masked motion, or speaker identity—has advanced to enhance AI agents' expressiveness and realism [29]. However, much of the focus has gone into generating rhythmic beat gestures with limited semantic information, leaving representational gestures that convey semantic messages (e.g., iconic) less explored [29, 40]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 579, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 541, + 723 + ], + "type": "text", + "content": "Generating spontaneous and semantically rich gestures from speech comes with multiple challenges. First, it requires capturing global discourse-level information and local fine-grained details (e.g., salient words) to generate speech-driven gestures that reflect the intended meaning and align with speech temporally and semantically. Second, existing methods often generate repetitive and short sequences that do not span the full range of expressive motions required for natural communication. To leverage semantics when generating gestures, researchers have attempted to align motion with speech representations at a global level, e.g., by leveraging pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "type": "text", + "content": "13963" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "type": "text", + "content": "trained semantic representations such as CLIP [58] or focusing on semantically important keywords [6, 57]. Nonetheless, they often fail to (i) unify global and local semantic modelling within a single framework and (ii) exploit the relevance of the semantic information in guiding gesture generation [27]. At the same time, raw audio features and speaker identity are relevant to the timing and style of gestures. In this paper, we address these limitations by integrating speech, speech semantics, and gesturing style, exploiting semantic information at different levels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 206, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 206, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 67, + 206, + 295, + 529 + ], + "type": "text", + "content": "Specifically, we propose a two-stage framework, namely, SemGes, that integrates speech, text-based semantics, and speaker identity into a unified gesture-generation model (see Figure 1). In stage 1, we build motion prior of holistic gestures (i.e., body and hands) by training a vector-quantized variational autoencoder (VQ-VAE) to learn an efficient, compositional motion latent space. This stage results in a robust motion encoder & decoder and quantized codebooks that can reconstruct naturalistic gestures while allowing the reuse of learned codebook entries. Stage 2 leverages the learned motion priors to drive gesture synthesis by fusing three modalities using a cross-modal Transformer encoder: (i) text-based semantics, (ii) raw-audio speech features, and (iii) speaker identity for style consistency. We impose a semantic coherence loss that aligns text-based embeddings with the VQ-VAE motion latent space and a semantic relevance loss that emphasises representational gestures (e.g. iconic and metaphoric gestures). A multimodal consistency objective ensures the fused multimodal representations are compatible with the learned motion codebooks, enabling the generation of gestures that are both semantically rich and visually natural. Finally, we introduce a simple but effective long-sequence inference strategy that smoothly combines overlapping motion clips for extended durations. To summarize our contributions," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 531, + 295, + 722 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 69, + 531, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 531, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 69, + 531, + 295, + 590 + ], + "type": "text", + "content": "- We introduce a novel framework, SemGes, that first learns a robust VQ-VAE motion prior for body and hand gestures, and then generates gestures driven by fused speech audio, text-based semantics, and speaker identity in a cross-modal transformer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 591, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 591, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 69, + 591, + 295, + 638 + ], + "type": "text", + "content": "- Our method jointly captures discourse-level context via a semantic coherence loss and fine-grained representational gestures (e.g., iconic, metaphoric) via a semantic relevance loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 639, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 639, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 69, + 639, + 295, + 673 + ], + "type": "text", + "content": "- We propose an overlap-and-combine inference algorithm that maintains smooth continuity over extended durations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 674, + 295, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 295, + 722 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 295, + 722 + ], + "type": "text", + "content": "- Extensive experiments on two benchmarks, namely, the BEAT [27] and TED Expressive [33] datasets show that our method outperforms recent baselines in both objective metrics (e.g., Fréchet Gesture Distance" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 539, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 539, + 96 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 539, + 96 + ], + "type": "text", + "content": "(FGD), diversity, semantic alignment) and user judgment of generated gestures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 109, + 400, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 109, + 400, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 109, + 400, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 129, + 541, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 541, + 357 + ], + "type": "text", + "content": "Data-driven Co-Speech Gesture Generation. Current gesture generation approaches are based on generative deep neural networks. These approaches use advanced models such as Transformers [28], Generative Adversarial Networks [23], Normalizing Flows [18, 30], Vector Quantized Variational Autoencoder(VQ-VAE) [16] and Denoising Diffusion Probabilistic Models [47]. In addition, researchers have explored the impact of different model inputs on the naturalness and appropriateness of generated gestures. Various modal inputs have been used, such as text [15], audio [50, 59], image [32, 39], and speaking style [3]. For a comprehensive survey, we refer to Nyatsanga et al. [40]. Although there have been significant improvements in this field, current methods fall short in generating semantically grounded gestures at a fine-grained level. In other words, while the generated motions look convincing at first glance, they do not match well with the meaning of the text, or they mostly focus on beat-type gestures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 375, + 539, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 539, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 539, + 387 + ], + "type": "text", + "content": "Semantics-aware Co-Speech Gesture Generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 541, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 541, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 541, + 672 + ], + "type": "text", + "content": "A group of work focused on semantics-aware gesture generation where the semantic information is handled in two ways: global semantics and local semantics. Methods that focus on global semantic information [10, 22, 58] align gestures with text or audio, but they fall short in generating gesture types matching the semantic context, such as iconic, metaphoric and deictic gestures. To capture a wider range of semantic gestures, works like [5, 6, 27] adopt local semantic-aware modelling by integrating the semantic salient words to the neural network. However, these approaches often fail to ensure that the generated gestures align with both the broader audio or textual context and a combination of global and local semantics. Liang et al. [26], Voß and Kopp [49] incorporate both global and local semantics, however, they require extensive annotations. Recently, Zhang et al. [57] employed a generative retrieval framework based on LLMs to address the sparsity problem in datasets with semantic gestures. However, they do not explicitly model the different types of gestures [27] or gesture phases [12] grounded in linguistic research. Moreover, there is still not enough understanding of the impact of different annotations and fine-grained semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "content": "Substantial research [6, 28, 57, 58] focused on two-stage latent space generative modelling to overcome the limitations of co-speech gesture generation and to generate more naturalistic and diverse gestures." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "text", + "content": "13964" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 286, + 112 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 286, + 112 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 286, + 112 + ], + "type": "image", + "image_path": "71af008c2730c688b5250cfcba0cceaa9c17e67ca54a1d376357b4a68f77ac40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 124, + 295, + 146 + ], + "lines": [ + { + "bbox": [ + 69, + 124, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 69, + 124, + 295, + 146 + ], + "type": "text", + "content": "Figure 2. We pre-train two VQ-VAEs by reconstructing body and hand motions with a dedicated codebook for each." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 164, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 164, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 67, + 164, + 296, + 319 + ], + "type": "text", + "content": "These approaches first learn a latent space and then model gestures probabilistically, effectively integrating the strengths of different methods in different stages. Liu et al. [28], Zhang et al. [57] capture complex dependencies in the latent space using VQ-VAE while Zhi et al. [58] employs CLIP [46, 56] to align text and motion embeddings. Ao et al. [6] introduces a diffusion-based model that leverages semantic awareness, while Liu et al. [28] utilizes a transformer-based approach to generate holistic body gestures. In contrast with two-stage generative modeling approaches, end-to-end methods such as [5, 59], are prone to jittering artifacts, especially in hand-motion generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 320, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 320, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 67, + 320, + 296, + 499 + ], + "type": "text", + "content": "Our model contributes to the research line on semantics-aware co-speech gesture generation by taking into account both global and local semantics. Inspired by the previous work, we employ a two-stage latent space generative modelling for high-quality motion representation. We learn semantic coherence between text and gestures globally with cosine similarity. Moreover, our model takes into account the semantic relevancy of gesture types with minimally required annotations. In contrast with other semantic learning models, we focus on annotations with different gesture types embedded in linguistic research. Our work is closest to CAMN [27] in that sense; however, CAMN does not include semantic coherence learning by aligning text and gestures' latent space globally." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 514, + 151, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 514, + 151, + 528 + ], + "spans": [ + { + "bbox": [ + 69, + 514, + 151, + 528 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 534, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 534, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 295, + 618 + ], + "type": "text", + "content": "We propose a two-stage approach that generates cospeech gestures by grounding them in raw speech, text-based semantics, and speaker identity. In Section 3.1, we introduce a VQ-VAE encoder-decoder that learns a robust motion prior. Section 3.2 details our gesture synthesis and inference pipeline based on speech, semantics, and identity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": "Problem formulation. Our goal is to generate hand gestures " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{h} = (g_{1}^{h},\\dots,g_{T}^{h})\\in \\mathbb{R}^{T\\times J}" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": " and body gestures " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{b} = (g_{1}^{b},\\dots,g_{T}^{b})\\in \\mathbb{R}^{T\\times J}" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": " is the number of time steps and " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": " the number of joints (e.g., 38 for hands, 9 for body). Each motion vector " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "g_{t}^{h}" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "g_{t}^{b}" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": " is encoded in a Rot6D representation, capturing joint rotations at time " + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 638, + 296, + 721 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 541, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 109 + ], + "type": "text", + "content": "To model human motion of body and hands, we first learn a motion generator " + }, + { + "bbox": [ + 313, + 72, + 541, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_g" + }, + { + "bbox": [ + 313, + 72, + 541, + 109 + ], + "type": "text", + "content": " (Stage 1), which synthesizes a plausible motion sequence:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 113, + 541, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 113, + 541, + 136 + ], + "spans": [ + { + "bbox": [ + 359, + 113, + 541, + 136 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathcal {M} _ {g}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(g _ {1}, \\dots , g _ {T}\\right) \\right\\|. \\tag {1}", + "image_path": "0730adf0b76a5d6b5bd11adac0986b90fa9f3bdba2342a82cab88678922660b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": "Next, we condition on (i) the raw input audio " + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = (a_{1},\\ldots ,a_{T})" + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": ", (ii) the speaker identity embedding " + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": ", and (iii) the text-based semantic embeddings of the speech " + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "inline_equation", + "content": "S = (s_{1},\\dots ,s_{T})" + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": ". Our second-stage model " + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{a,s,i}" + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": " uses these inputs to generate a latent sequence that the motion generator " + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_g" + }, + { + "bbox": [ + 313, + 141, + 541, + 214 + ], + "type": "text", + "content": " then decodes into naturalistic gestures:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 222, + 541, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 222, + 541, + 239 + ], + "spans": [ + { + "bbox": [ + 358, + 222, + 541, + 239 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\mathcal {M} _ {a, s, i}} \\left\\| \\boldsymbol {G} - \\mathcal {M} _ {g} \\left(\\mathcal {M} _ {a, s, i} (\\boldsymbol {A}, \\boldsymbol {S}, I)\\right) \\right\\|. \\tag {2}", + "image_path": "383d4e8310827f0871168cd1ac647ad23f450207f88f40052f57fa6c56a5b7eb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 539, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 539, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 539, + 269 + ], + "type": "text", + "content": "3.1. Stage 1: Learning Efficient Codebooks & Compositional Motion Priors" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": "Realistic co-speech gestures require modelling the sequential motion of both body and hand joints. Rather than learning a single representation for the entire body, we adopt a compositional approach, using a discrete codebook of learned representations specific to each part (hands & body). Any gesture motion can then be represented by selecting appropriate codebook entries. Following [28, 48, 53], we employ a VQ-VAE architecture (see Fig. 2) with encoder " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": ". Given hand motion " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "G^h \\in \\mathbb{R}^{T \\times J}" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": " and body motion " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "G^b \\in \\mathbb{R}^{T \\times J}" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": ", the encoder produces latent vectors " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "\\hat{z}^h" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "inline_equation", + "content": "\\hat{z}^b" + }, + { + "bbox": [ + 313, + 274, + 541, + 430 + ], + "type": "text", + "content": ", which are quantized by selecting the nearest entries in the codebooks. Formally," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 440, + 541, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 440, + 541, + 459 + ], + "spans": [ + { + "bbox": [ + 372, + 440, + 541, + 459 + ], + "type": "interline_equation", + "content": "\\mathbf {q} (\\hat {\\boldsymbol {z}}) = \\arg \\min _ {z ^ {i} \\in \\mathcal {Z}} \\| \\hat {z} ^ {j} - z ^ {i} \\|, \\tag {3}", + "image_path": "1d4ed1dc9382446e70f1c3a7a483c055bf70411fb7cc6ae16c658c1c9df9ee08.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "inline_equation", + "content": "z^i" + }, + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "text", + "content": " are the learned codebook entries, and " + }, + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "inline_equation", + "content": "\\hat{z}^j" + }, + { + "bbox": [ + 313, + 463, + 541, + 510 + ], + "type": "text", + "content": " denotes an element of the latent vector for either hand or body. We train the VQ-VAE via a straight-through gradient estimator, minimizing:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 334, + 525, + 541, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 525, + 541, + 551 + ], + "spans": [ + { + "bbox": [ + 334, + 525, + 541, + 551 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {V Q - V A E}} = \\left\\| \\mathbf {g} - \\hat {\\mathbf {g}} \\right\\| ^ {2} + \\left\\| \\dot {\\mathbf {g}} - \\hat {\\dot {\\mathbf {g}}} \\right\\| ^ {2} + \\left\\| \\ddot {\\mathbf {g}} - \\hat {\\ddot {\\mathbf {g}}} \\right\\| ^ {2} \\tag {4} \\\\ + \\left\\| \\operatorname {s g} [ \\mathbf {E} (\\mathbf {g}) ] - \\mathbf {q} (\\hat {z}) \\right\\| ^ {2} + \\left\\| \\mathbf {E} (\\mathbf {g}) - \\operatorname {s g} [ \\mathbf {q} (\\hat {z}) ] \\right\\| ^ {2}, \\\\ \\end{array}", + "image_path": "62149f5e430eac318e9f7945307abaccc2cac728cf5f154ae40c6e9890d5c234.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 539, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 539, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 539, + 590 + ], + "type": "text", + "content": "where the first three terms reconstruct joint positions, velocities, and accelerations, and the last two terms implement the VQ-VAE commitment loss [48]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "content": "By the end of this stage, we have motion " + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "inline_equation", + "content": "(m)" + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "content": " encoder " + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_m)" + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "content": ", decoder " + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_m)" + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "content": " and codebooks " + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "inline_equation", + "content": "(\\text{Quant}^m(\\cdot))" + }, + { + "bbox": [ + 313, + 591, + 541, + 663 + ], + "type": "text", + "content": " for hands and body. In the next section (Section 3.2), we show how this discretized motion of hands and body guides speech, semantics and speaker identity-driven generation to produce realized co-speech gestures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 669, + 539, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 539, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 539, + 692 + ], + "type": "text", + "content": "3.2. Stage 2: Speech and Identity Driven Semantic Gesture Generator" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 698, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 698, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 698, + 541, + 723 + ], + "type": "text", + "content": "This stage focuses on generating gestures conditioned on three inputs: speech embeddings, text-based seman" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "text", + "content": "13965" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 78, + 525, + 213 + ], + "blocks": [ + { + "bbox": [ + 90, + 78, + 525, + 213 + ], + "lines": [ + { + "bbox": [ + 90, + 78, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 90, + 78, + 525, + 213 + ], + "type": "image", + "image_path": "f390537d0780d93a79329de2bb3a674c27d7e5f1c1a52b19a07db673cc813312.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 225, + 542, + 270 + ], + "lines": [ + { + "bbox": [ + 68, + 225, + 542, + 270 + ], + "spans": [ + { + "bbox": [ + 68, + 225, + 542, + 270 + ], + "type": "text", + "content": "Figure 3. SemGes employs three training pathways: (1) Global semantic coherence, which minimizes latent disparities between gesture and text encoders; (2) Multimodal Quantization learning, where integrated multimodal representation codes are aligned with quantized motion to decode them into hand and body movements; and (3) Semantic relevance learning, which emphasizes semantic gestures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 285, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 285, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 68, + 285, + 296, + 332 + ], + "type": "text", + "content": "tic embeddings, and speaker identity. As illustrated in Figure 3, the second-stage architecture has three main modules, which we elaborate on in the following subsections." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 110, + 354, + 262, + 497 + ], + "blocks": [ + { + "bbox": [ + 110, + 354, + 262, + 497 + ], + "lines": [ + { + "bbox": [ + 110, + 354, + 262, + 497 + ], + "spans": [ + { + "bbox": [ + 110, + 354, + 262, + 497 + ], + "type": "image", + "image_path": "79def40ab4ae068b8c0a3781378d3b936df9ae93d335c8b434390dd14871c50a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 517, + 279, + 529 + ], + "lines": [ + { + "bbox": [ + 85, + 517, + 279, + 529 + ], + "spans": [ + { + "bbox": [ + 85, + 517, + 279, + 529 + ], + "type": "text", + "content": "Figure 4. Semantic Coherence Embedding Learning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 541, + 279, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 541, + 279, + 554 + ], + "spans": [ + { + "bbox": [ + 69, + 541, + 279, + 554 + ], + "type": "text", + "content": "3.2.1. Semantic Coherence Embedding Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "text", + "content": "To align text-based semantics with motion embeddings, we introduce a shared embedding space for both motion priors and speech transcripts. Specifically, we embed word tokens using a pre-trained FastText model [8], then feed these embeddings into a trainable text-based semantic encoder " + }, + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_s" + }, + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "text", + "content": ". At the same time, we use the pre-trained motion encoder " + }, + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 68, + 556, + 296, + 664 + ], + "type": "text", + "content": " from Stage 1 to encode ground-truth gesture sequences. Thus, for a batch of paired (gesture, transcript) samples, we get:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 669, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 669, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 99, + 669, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {Z} ^ {S} = \\mathcal {E} _ {s} (S), \\quad \\mathcal {Z} ^ {h} = \\mathcal {E} _ {m} ^ {h} (G ^ {h}), \\quad \\mathcal {Z} ^ {b} = \\mathcal {E} _ {m} ^ {b} (G ^ {b}), \\tag {5}", + "image_path": "20e39b74405b1666440da50c5ec2aaf7b9cca1d00b261b3851def676f1cd95d2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "content": " is the tokenized speech transcript, and " + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "inline_equation", + "content": "G^{h}" + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "inline_equation", + "content": "G^{b}" + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "content": " correspond to the ground-truth hand gesture sequence and body gesture sequence, respectively. " + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^h" + }, + { + "bbox": [ + 68, + 686, + 296, + 723 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^b" + }, + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "type": "text", + "content": " represent the hand and body ground-truth motion encodings from Stage 1, and " + }, + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^s" + }, + { + "bbox": [ + 313, + 285, + 540, + 321 + ], + "type": "text", + "content": " denotes the text-based semantic encoder output." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 337, + 541, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 541, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 541, + 492 + ], + "type": "text", + "content": "Semantic Coherence Loss. We maximize the similarity of correct (gesture, transcript) pairs and minimize it for mismatched pairs, enforcing semantic coherence. This aligns gestures and textual semantics in a common space while keeping " + }, + { + "bbox": [ + 313, + 337, + 541, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 313, + 337, + 541, + 492 + ], + "type": "text", + "content": " frozen, as illustrated in Figure 4. We impose the semantic coherence constraint separately on both hand and body movements to align gestures with transcripts in the shared embedding space. Specifically, we introduce two distinct cosine similarity losses: one between the text encoder output and the hand motion latent encoding and another between the text encoder output and the body motion latent encoding. Formally, we minimize:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 511, + 541, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 511, + 541, + 525 + ], + "spans": [ + { + "bbox": [ + 324, + 511, + 541, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} = 1 - \\cos \\left(\\mathcal {Z} ^ {h}, \\mathcal {Z} ^ {s}\\right) + 1 - \\cos \\left(\\mathcal {Z} ^ {b}, \\mathcal {Z} ^ {s}\\right), \\tag {6}", + "image_path": "fca88b59437933c4e37d47f9bb90954ae6865a11577f6c1b3115c3e87be0ecdf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 536, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 536, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 536, + 545 + ], + "type": "text", + "content": "where the function " + }, + { + "bbox": [ + 313, + 533, + 536, + 545 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 533, + 536, + 545 + ], + "type": "text", + "content": " measures cosine similarity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 444, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 444, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 444, + 563 + ], + "type": "text", + "content": "3.2.2. Crossmodal Integration" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "text", + "content": "SemGes supports multi-modal inputs in the second training stage by combining audio features and speaker identity with semantic text embeddings using a Transformer encoder with self and cross-attention layers (see Figure 3). We begin by extracting HuBERT features [20] from raw speech (keeping the HuBERT encoder frozen). We concatenate audio features " + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^a" + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "text", + "content": " and speaker embeddings " + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^i" + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^r" + }, + { + "bbox": [ + 313, + 567, + 541, + 673 + ], + "type": "text", + "content": ", which we feed into a self-attention layer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "content": "Next, we use a cross-attention layer that takes " + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^r" + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "content": " as the query and the motion-aligned text-based semantic features " + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^s" + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "content": " as the key-value pair. The final hidden representation " + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^f" + }, + { + "bbox": [ + 313, + 674, + 541, + 723 + ], + "type": "text", + "content": " serves as the multimodal latent code" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "text", + "content": "13966" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 296, + 120 + ], + "type": "text", + "content": "that drives gesture synthesis when passed to our vector quantization and VQ-VAE-based motion decoder, which is learned in our first stage (see the yellow box in Figure 3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 134, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 134, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 68, + 134, + 296, + 217 + ], + "type": "text", + "content": "Multimodal Quantization Consistency Loss. SemGes quantizes the multimodal latent code using separate hand and body codebooks. To align this code with the ground-truth motion latent codes, we apply independent quantization consistency losses for each component. Specifically, the quantization loss is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 101, + 222, + 296, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 222, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 101, + 222, + 296, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {q u a n t i z a t i o n}} = \\left\\| Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {h} \\left(\\mathcal {Z} ^ {h}\\right) \\right\\| ^ {2} + \\tag {7}", + "image_path": "0ecf9d2c6541564ae8a81899f9589716b9580280e93088cef90b21088e04ed3d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 240, + 296, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 240, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 155, + 240, + 296, + 255 + ], + "type": "interline_equation", + "content": "\\left\\| Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {f}\\right) - Q u a n t ^ {b} \\left(\\mathcal {Z} ^ {b}\\right) \\right\\| ^ {2} \\tag {17}", + "image_path": "618b01282e026d20a415d64fd748e13dd23af7476c22f8b56c31daee0ef7e257.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\text{Quant}^h(\\cdot)" + }, + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\text{Quant}^b(\\cdot)" + }, + { + "bbox": [ + 68, + 262, + 295, + 297 + ], + "type": "text", + "content": " denote the quantization functions for the hand and body codebooks, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "text", + "content": "The multimodal quantization loss aligns the integrated latent code " + }, + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^f" + }, + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "text", + "content": " with the learned motion code, a critical step since gesture synthesis is obtained through the quantized multimodal representation. Specifically, " + }, + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^f" + }, + { + "bbox": [ + 68, + 299, + 296, + 406 + ], + "type": "text", + "content": " is vector-quantized using separate hand and body codebooks before being decoded by their respective VQ decoders. This process ensures that both hand and body movements contribute effectively to the final output. Formally, the generated gestures are given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 411, + 295, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 411, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 77, + 411, + 295, + 425 + ], + "type": "interline_equation", + "content": "\\hat {G} = \\hat {G} ^ {h} \\oplus \\hat {G} ^ {b} = \\mathcal {D} _ {m} ^ {h} \\left(\\operatorname {Q u a n t} ^ {h} \\left(\\mathcal {Z} ^ {f}\\right)\\right) \\oplus \\mathcal {D} _ {m} ^ {b} \\left(\\operatorname {Q u a n t} ^ {b} \\left(\\mathcal {Z} ^ {f}\\right)\\right), \\tag {8}", + "image_path": "af7798b1c54ba5ac5c217ac834cef42c0bb254af1bfabf92c2d7570d7ed00de9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "text", + "content": " denotes concatenation, jointly synthesizing hand and body motions (i.e. " + }, + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{G}" + }, + { + "bbox": [ + 68, + 431, + 296, + 456 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 460, + 241, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 460, + 241, + 471 + ], + "spans": [ + { + "bbox": [ + 69, + 460, + 241, + 471 + ], + "type": "text", + "content": "3.2.3. Gesture Semantic Relevance Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 475, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 475, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 68, + 475, + 296, + 547 + ], + "type": "text", + "content": "To prioritize the generation of semantically meaningful gestures (e.g., iconic, metaphoric, or deictic), which are less frequent than beat gestures, we introduce a semantic relevance loss. This loss emphasizes semantic annotations while preventing over-penalization of minor deviations. Formally, it is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 553, + 295, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 553, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 295, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} = \\mathbb {E} \\left[ \\lambda \\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) \\right], \\tag {9}", + "image_path": "abb6dee3fdc042cb70c885e7832033d803479947650a4791af669bca53c41d02.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "text", + "content": " is the annotation relevance factor, and " + }, + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 68, + 574, + 296, + 611 + ], + "type": "text", + "content": " is a piecewise function that applies a quadratic penalty for small errors and a linear penalty for larger ones:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 616, + 295, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 616, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 86, + 616, + 295, + 645 + ], + "type": "interline_equation", + "content": "\\Psi (\\mathbf {G} - \\hat {\\mathbf {G}}) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (\\mathbf {G} - \\hat {\\mathbf {G}}) ^ {2}, & \\text {i f} | \\mathbf {G} - \\hat {\\mathbf {G}} | < \\alpha , \\\\ \\alpha \\left(| \\mathbf {G} - \\hat {\\mathbf {G}} | - \\frac {1}{2} \\alpha\\right), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {10}", + "image_path": "fb696550ae4f95c500ab04d4f48e6a57e5d95cceba004d3d16ce9bac866a4d53.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 651, + 132, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 651, + 132, + 661 + ], + "spans": [ + { + "bbox": [ + 69, + 651, + 132, + 661 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 69, + 651, + 132, + 661 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 677, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 677, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 68, + 677, + 296, + 700 + ], + "type": "text", + "content": "Combined Objective Functions. Finally, the overall objective is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 82, + 708, + 295, + 719 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 708, + 295, + 719 + ], + "spans": [ + { + "bbox": [ + 82, + 708, + 295, + 719 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {S e m G e s}} = \\mathcal {L} _ {\\text {s e m a n t i c - c o h e r e n c e}} + \\mathcal {L} _ {\\text {s e m a n t i c - r e l e v a n c e}} + \\mathcal {L} _ {\\text {q u a n t i z a t i o n}}, \\tag {11}", + "image_path": "c328dc3580e326fae802c51ce18b3726671b63013e2a4f2dbb221a1974e093cd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "type": "code", + "bbox": [ + 315, + 87, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 511, + 84 + ], + "type": "text", + "content": "Algorithm 1 Long Gesture Sequence Algorithm" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": "Require: Audio " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " aligned speech transcript " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " and speaker ID " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " Pre-trained codebooks and motion decoder (Stage 1) \nEnsure: Long-sequence gesture M \n1: Partition " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "(\\mathcal{A},\\mathcal{S},\\mathcal{I})" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " into clips " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\{(\\mathcal{A}_c,\\mathcal{S}_c,\\mathcal{I}_c)\\}_{c = 1}^C" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " \n2: Compute latent representation: " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^{f}\\gets" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " Encode(A,S,I) \n3: Quantize: " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^e\\gets" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " VectorQuantize " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "(\\mathcal{Z}^{f})" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " \n4: Decode initial clip: " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\hat{M}_1\\gets \\mathrm{Dec}(\\mathcal{Z}^e)" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " \n5: for each clip " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "c = 2" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " do \n6: Set first 4 frames of " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\hat{M}_c" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " to the last 4 frames of " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\hat{M}_{c - 1}" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " \n7: Generate remaining frames of " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\hat{M}_c" + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "text", + "content": " \n8: end for \n9: return " + }, + { + "bbox": [ + 315, + 87, + 541, + 266 + ], + "type": "inline_equation", + "content": "\\hat{M}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "code_body" + } + ], + "index": 17, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 289, + 541, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 541, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 541, + 338 + ], + "type": "text", + "content": "which jointly optimizes the model to generate gestures that are semantically coherent at both global and fine-grained levels while remaining faithful to the Stage 1 motion prior." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 345, + 511, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 511, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 511, + 358 + ], + "type": "text", + "content": "3.3. Inference of Long Gesture Sequences" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 363, + 541, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 541, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 541, + 495 + ], + "type": "text", + "content": "Generating long sequences of gestures is challenging due to the need to maintain coherence and smooth transitions. Our Long-Sequence Gesture Motion algorithm (Alg. 1) addresses these challenges by partitioning the input speech, transcript, and speaker identity into aligned clips. For each clip, a multimodal latent representation is computed using our cross-modal encoder, vector-quantized via the Stage 1 codebooks, and decoded into gesture motions. Overlapping 4-frame segments between clips provides continuity, resulting in extended, naturalistic gesture sequences." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 506, + 432, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 432, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 432, + 520 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 525, + 376, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 376, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 376, + 537 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 543, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 541, + 723 + ], + "type": "text", + "content": "Our proposed methodology is evaluated on two benchmarks, namely, BEAT [27] and the TED expressive dataset [33]. The BEAT dataset consists of 76 hours of multimodal recordings, which include speech audio recordings, speech transcriptions, and, more importantly, motion data collected from 30 participants, leveraging Motion Capture (MOCAP) technology. The participants expressed emotions in eight distinct scenarios across four languages. The motion data contains joint rotation angles, which were designed for consistency across varying body sizes. The TED Expressive dataset [33] is segmented from TED Talk videos into smaller shots based on scene boundaries. Liu et al. [33] extracted each frame's 2D human pose using OpenPose BEAT [9]. Using these 2D pose priors, ExPose [43] was" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 763 + ], + "type": "text", + "content": "13967" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 113, + 518, + 205 + ], + "blocks": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "type": "text", + "content": "Table 1. Comparison of SemGesGen with other methods on the BEAT and TED-Expressive datasets. For BEAT, we compare with CaMN [27], DiffGesture [59], LivelySpeaker [58], and DiffSheg [10]. The same methods are evaluated on TED-Expressive. SRGR is not applicable (denoted with " + }, + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "type": "inline_equation", + "content": "-" + }, + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "type": "text", + "content": ") for TED-Expressive as it does not contain annotations for semantic relevance of gestures." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 113, + 518, + 205 + ], + "lines": [ + { + "bbox": [ + 91, + 113, + 518, + 205 + ], + "spans": [ + { + "bbox": [ + 91, + 113, + 518, + 205 + ], + "type": "table", + "html": "
BEATTED-Expressive
MethodFGD ↓BC ↑Diversity ↑SRGR ↑MethodFGD ↓BC ↑Diversity ↑
CaMN [27]8.5100.797206.7890.231CaMN [27]7.6730.642156.236
DiffGesture [59]9.6320.876210.6780.106DiffGesture [59]9.3260.662119.889
LivelySpeaker [58]13.3780.891214.9460.229LivelySpeaker [58]8.1450.691119.231
DiffSheg [10]6.6230.922257.6740.250DiffSheg [10]8.4570.712108.972
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
", + "image_path": "fe64ef64ce5708037abbe668334be41eacceca7f62ef9e0e84f81d5cab389ab9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 219, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 219, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 68, + 219, + 296, + 268 + ], + "type": "text", + "content": "employed to annotate the 3D upper body keypoints, including 13 upper body joints and 30 finger joints. Both datasets' training and validation samples are divided into 34-frame clips." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 283, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 283, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 68, + 283, + 296, + 357 + ], + "type": "text", + "content": "Cross-Validation. We evaluate our approach on the BEAT dataset, following the protocol in [27], where the data is randomly split into a 19:2:2 ratio for training, validation, and testing. Similarly, for the TED Expressive dataset, we adapt the protocol in [33], using a random split of 8:1:1 for training, validation, and testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 371, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 371, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 68, + 371, + 296, + 407 + ], + "type": "text", + "content": "Implementation Details. The details of the model architectures and training are provided in Section 2 of the Supplementary Materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 415, + 213, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 415, + 213, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 415, + 213, + 427 + ], + "type": "text", + "content": "4.2. State-of-the-Art Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 433, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 433, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 68, + 433, + 296, + 505 + ], + "type": "text", + "content": "We compare SemGes against a set of representative state-of-the-art models that focus on semantic-driven gesture generation. The selected models achieved strong performance on the BEAT and TED-Expressive datasets, making them suitable for a fair comparison with our method. The selected models are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 506, + 296, + 723 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 506, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 506, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 69, + 506, + 295, + 590 + ], + "type": "text", + "content": "1. Cascaded Motion Network(CaMN) [27] is the current benchmark model for the BEAT dataset. CaMN is based on LSTMs and integrates multiple input modalities, including audio, text, facial expressions, and emotion. Additionally, like SemGes, it leverages semantic relevance annotations to enhance gesture generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 591, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 591, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 69, + 591, + 295, + 674 + ], + "type": "text", + "content": "2. DiffSHEG [10] is a state-of-the-art diffusion-based model for real-time speech-driven holistic gesture generation. It is conditioned on noisy motion, audio, and speaker ID. DiffSHEG introduces a Fast Out-painting-based Partial Autoregressive Sampling method to efficiently generate arbitrary-length sequences in real time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 674, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 296, + 723 + ], + "type": "text", + "content": "3. LivelySpeaker [58] generates semantically and rhythmically aware co-speech gestures by leveraging an MLP-based diffusion model. The model conditions gesture generation on text, noised motion," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 219, + 540, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 219, + 540, + 243 + ], + "spans": [ + { + "bbox": [ + 326, + 219, + 540, + 243 + ], + "type": "text", + "content": "speaker ID, and audio to enable text-driven gesture control while incorporating global semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 541, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 541, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 541, + 315 + ], + "type": "text", + "content": "4. DiffGes [59] models the diffusion and denoising processes within the gesture domain, enabling the generation of high-fidelity, audio-driven gestures conditioned on both audio and gesture inputs. Several recent studies[6, 26] have also demonstrated strong performance in this area." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 541, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 541, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 541, + 558 + ], + "type": "text", + "content": "We exclude certain models from our comparison. For instance, SEEG [26] and [57] rely on additional data annotations (e.g., Semantic Prompt Gallery or ChatGPT-generated annotations) that are not uniformly available. In addition, other works, such as Ao et al. [6], Pang et al. [42], Zhang et al. [57], are excluded from our analysis due to the inaccessibility of their codebase. Voß and Kopp [49] is omitted due to its high computational cost and the unavailability of annotations. Liu et al. [31, 34], Mughal et al. [36, 37], Ng et al. [38], Yi et al. [54] are excluded as they primarily focus on holistic gestures with face and mesh data, which fall outside the scope of this work. Similarly, Chhatre et al. [11], Qi et al. [44] are excluded, as their emphasis lies in emotion-driven gesture generation rather than the semantic aspects. Furthermore, Ahuja et al. [1, 2], Alexander et al. [4], Habibie et al. [17], Liu et al. [33], Sun et al. [45], Yang et al. [51], Ye et al. [52] are omitted due to their lack of relevance to semantic-driven gesture generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 499, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 499, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 499, + 583 + ], + "type": "text", + "content": "5. Quantitative Objective Evaluations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 590, + 540, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 540, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 540, + 650 + ], + "type": "text", + "content": "Evaluation Metrics. We employ four standard objective metrics for evaluating the quality of gesture generation, namely, Fréchet Gesture Distance (FGD) [55], Beat Consistency Score (BC) [25], Diversity [24], and Semantic-Relevant Gesture Recall (SRGR) [27]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 541, + 723 + ], + "type": "text", + "content": "FGD measures how the generated gestures resemble real motion distributions by embedding sequences into a latent space via a pre-trained autoencoder. In contrast, BC focuses on synchronization with speech, measuring the alignment between speech onsets (audio beats) and motion beats, which are identified as velocity minima in" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "text", + "content": "13968" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 112, + 517, + 186 + ], + "blocks": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 541, + 105 + ], + "type": "text", + "content": "Table 2. Ablation studies evaluating the contributions of key components in SemGes on the BEAT and TED-Expressive Datasets. For BEAT, performance is measured using FGD (lower is better), BC, Diversity, and SRGR, while for TED-Expressive, SRGR is not applicable (denoted as -)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 112, + 517, + 186 + ], + "lines": [ + { + "bbox": [ + 91, + 112, + 517, + 186 + ], + "spans": [ + { + "bbox": [ + 91, + 112, + 517, + 186 + ], + "type": "table", + "html": "
BEATTED-Expressive
Model VariantsFGD ↓BC ↑Diversity ↑SRGR ↑Model VariantsFGD ↓BC ↑Diversity ↑
Baseline (VQVAE)10.3480.564198.5680.176Baseline (VQVAE)10.6820.612114.692
w/o Semantic Coherence Module8.0530.556249.5500.180w/o Semantic Coherence Module7.9240.623109.256
w/o Semantic Relevance Module7.5490.573245.3190.195w/o Semantic Relevance Module---
w/ SpeechCLIP Encoder6.7870.468289.6210.245w/ SpeechCLIP Encoder7.3410.605245.680
SemGes (Ours)4.4670.453305.7060.256SemGes (Ours)7.2630.671302.772
", + "image_path": "6ce76706a7dfae3d06af8432d0a430ad28cd1f590b81667879d4f45900acd19f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "text", + "content": "upper-body joints (excluding fingers). Meanwhile, **Diversity** captures the variability of generated motions by computing the average " + }, + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "text", + "content": " distance between pairs of " + }, + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 200, + 294, + 294 + ], + "type": "text", + "content": " generated clips. Finally, SRGR assesses semantic relevance by determining how well generated gestures align with the annotated semantic gestures. Further details on the objective metrics are included in the Supplementary Materials (Section 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 314, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 314, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 67, + 314, + 294, + 373 + ], + "type": "text", + "content": "Comparisons with Other Models. Table 1 compares the performance of our approach against four baseline methods across four evaluation metrics. As highlighted in the table, SemGes outperforms the baselines in FGD, Diversity, and SRGR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 374, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 374, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 67, + 374, + 295, + 589 + ], + "type": "text", + "content": "For the BEAT dataset, our approach achieves the highest SRGR, which we attribute to the exploitation of semantic relevance information in our training objectives. In addition, our approach shows a significant improvement in FGD and Diversity, indicating a closer alignment with the ground truth gesture distribution and a broader range of generated gestures compared to the second-best baselines. The performance on the Beat Consistency (BC) metric is lower for our method. This is expected given our focus on improving semantic awareness of the generated gestures rather than optimizing for strict temporal alignment between rhythmic beat gestures and speech. In addition, the BC metric can be sensitive to rapid, jittery movements; even minor motion artefacts may be mistakenly counted as additional beats, thereby increasing the BC score artificially- a phenomenon also observed in the diffusion-based baselines, as further illustrated in our supplementary video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 590, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 295, + 698 + ], + "type": "text", + "content": "We evaluate how our model handles the trade-off between semantic and beat scores by testing the model on beat-dominant gestures (without semantic content). The results show a significantly higher Beat score (0.689) than the full dataset Beat score (0.453). This confirms rhythmic consistency in beat-focused contexts. We provide additional evaluation in the supplementary materials (Section 3) to show how the model handles difficult cases (such as noisy speech or misaligned speech)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 699, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 699, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 699, + 296, + 723 + ], + "type": "text", + "content": "Note that the TED Expressive dataset lacks annotations for gesture semantic relevance, so SRGR is not" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 200, + 541, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 541, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 541, + 248 + ], + "type": "text", + "content": "applicable, and the semantic relevance loss was omitted during training. Nevertheless, SemGes produces diverse, naturalistic gestures on TED Expressive, outperforming baselines in FGD and Diversity metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 262, + 541, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 541, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 541, + 405 + ], + "type": "text", + "content": "Ablation Study. We evaluate the contributions of key components in SemGes through ablation experiments. First, we assess a baseline VQ-VAE model (Stage 1 only), which uses two stacked encoder-decoder blocks and an MLP. In this experiment, we test its ability to generate gestures, conditioned on audio, masked motion, and speaker identity. As shown in Table 2, this baseline underperforms compared to state-of-the-art methods (Table 1). As a result, we motivate our two-stage design where the VQ-VAE is reserved to learn the motion latent space and Stage 2 leverages speech and identity conditioning to generate gestures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 406, + 541, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 541, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 541, + 573 + ], + "type": "text", + "content": "Next, we examine Stage 2 by removing its components: (i) the Semantic Coherence Loss, (ii) the Semantic Relevance Loss, and (iii) by replacing the HuBERT-based speech encoder with SpeechCLIP. Results in Table 2 show that removing either the Semantic Coherence or Relevance Loss degrades FGD, Diversity, and SRGR scores, highlighting their roles in aligning gesture representations with textual semantics and capturing semantic importance. In addition, replacing the speech encoder results in marginal gains. The semantic encoder is fixed as FastText, which we believe is sufficient to capture the necessary semantic information [7]. Overall, these results confirm the importance of each module in generating semantics-aware gestures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 519, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 519, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 519, + 597 + ], + "type": "text", + "content": "6. Qualitative & Subjective Evaluations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 602, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 541, + 723 + ], + "type": "text", + "content": "Visualization Comparisons. Before presenting the subjective ratings of the generated gestures, Figure 5 provides a visual comparison between the ground truth, results from our approach and two baseline models. We use examples from the BEAT dataset. It is clear from the figure that our approach not only achieves better speech-gesture alignment but also produces gestures that are more naturalistic, diverse, and semantically aware. For example, while CaMN generates smooth movements, its gestures tend to be slower and less varied compared to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "text", + "content": "13969" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 532, + 178 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 532, + 178 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 532, + 178 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 532, + 178 + ], + "type": "image", + "image_path": "b4028ad1da7211a158b3965e028ce1b10b9d5550fdc235e8203825ca3cebacf9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 190, + 541, + 237 + ], + "lines": [ + { + "bbox": [ + 67, + 190, + 541, + 237 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 541, + 237 + ], + "type": "text", + "content": "Figure 5. Comparisons with baselines and ground truth gestures. Compared to the baseline method, our approach generates gestures that are aligned with speech content (semantics). For instance, when the speaker says \"remix\", our method produces gestures where the character raises both hands to emphasize the word before gradually lowering them—a movement that other methods fail to achieve. Similarly, when uttering \"first\", our method generates a raised hand gesture, producing an iconic gesture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 251, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 251, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 296, + 370 + ], + "type": "text", + "content": "our model. Additionally, the baseline methods show varying degrees of jitter—DiffGesture shows the highest jitter, followed by LivelySpeaker and DiffSheg, with CaMN displaying the least. Although CaMN includes semantic information, our approach strikes a more effective balance, generating gestures that align with actual motion, as shown also with the objective metrics. Based on these qualitative observations, our subsequent rating study focuses on evaluating gestures produced by the ground truth, our model, CaMN, and DiffSHEG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "text", + "content": "User Ratings of Generated Gestures. We conducted a user study using 40-second video clips from the BEAT test set, featuring subjects narrating six topics. Thirty native English speakers from the United Kingdom and the United States participated, with an average age of " + }, + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "inline_equation", + "content": "36 \\pm 20" + }, + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "text", + "content": " years and a female-to-male ratio of approximately 2:1. Each participant evaluated 24 videos generated by the ground truth, CaMN, DiffSHEG, and our model over a study duration of that lasted on average " + }, + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "inline_equation", + "content": "27 \\pm 5" + }, + { + "bbox": [ + 67, + 387, + 295, + 613 + ], + "type": "text", + "content": " minutes. For data quality, participants were required to pass attention verification questions, i.e., correctly answering at least two out of four questions regarding the narration topic. Participants rated the videos on three criteria: naturalness, diversity, and alignment with speech content and timing on a scale from 1 to 5. The videos were presented in a randomized order to avoid bias. In Section 3 of the Supplementary Materials, we provide screenshots and more details on the user study and interface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 615, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 615, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 615, + 295, + 723 + ], + "type": "text", + "content": "Figure 6 shows that ground-truth gestures received an average rating of 4 across all metrics, establishing an upper bound and validating the participant survey. Our model received the highest ratings among the generated gestures, significantly outperforming CaMN and Diff-SHEG in naturalness, synchronization, and diversity (indicated by the in Figure 6). These results prove that our approach produces gestures that are more natural, better aligned with speech, and more diverse than those gener" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 251, + 416, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 416, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 416, + 262 + ], + "type": "text", + "content": "ated by SOTA baselines." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 275, + 520, + 403 + ], + "blocks": [ + { + "bbox": [ + 328, + 275, + 520, + 403 + ], + "lines": [ + { + "bbox": [ + 328, + 275, + 520, + 403 + ], + "spans": [ + { + "bbox": [ + 328, + 275, + 520, + 403 + ], + "type": "image", + "image_path": "e3d5c4222b692542165121d51ee5f693894562de6161cdae88f72d56ef8b033e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 413, + 541, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 541, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 541, + 491 + ], + "type": "text", + "content": "Figure 6. Average ratings of users for ground truth gestures and gestures generated through our approach, CAMN, and DiffSHEG. The bars illustrate the average user ratings across three metrics: naturalness, diversity, and alignment with speech content and timing. Statistical t-tests show that our approach received significantly higher ratings than CAMN and DiffSHEG, with " + }, + { + "bbox": [ + 313, + 413, + 541, + 491 + ], + "type": "inline_equation", + "content": "p < 0.05" + }, + { + "bbox": [ + 313, + 413, + 541, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 388, + 523 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 531, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 541, + 723 + ], + "type": "text", + "content": "We proposed SemGes, a novel two-stage approach to semantic grounding in co-speech gesture generation by integrating semantic information at both fine-grained and global levels. In the first stage, a motion prior generation module is trained using a vector-quantized variational autoencoder to produce realistic and smooth gesture motions. Building upon this model, the second stage generates gestures from speech, text-based semantics, and speaker identity while maintaining consistency between gesture semantics and co-occurring speech through semantic coherence and relevance modules. Subjective and objective evaluations show that our work achieves state-of-the-art performance across two public benchmarks, generating semantics-aware and diverse gestures. Future direction and limitations are discussed in Section 5 of the Supplementary Materials." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "text", + "content": "13970" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 91, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 91, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 68, + 91, + 296, + 159 + ], + "type": "text", + "content": "The project is funded by the Max Planck Society. We thank Sachit Misra for his invaluable assistance with rendering Avatar characters. We extend our gratitude to the members of the Multimodal Language Department at Max Planck Institute for Psycholinguistics for their feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 181, + 128, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 181, + 128, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 181, + 128, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 201, + 296, + 723 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 74, + 201, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 201, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 74, + 201, + 296, + 257 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566-20576, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 258, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 258, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 74, + 258, + 296, + 312 + ], + "type": "text", + "content": "[2] Chaitanya Ahuja, Pratik Joshi, Ryo Ishii, and Louis-Philippe Morency. Continual learning for personalized co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20893-20903, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 314, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 314, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 74, + 314, + 296, + 369 + ], + "type": "text", + "content": "[3] Simon Alexander, Gustav Eje Henter, Taras Kucherenko, and Jonas Beskow. Style-controllable speech-driven gesture synthesis using normalising flows. In Computer Graphics Forum, pages 487-496. Wiley Online Library, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 370, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 370, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 74, + 370, + 296, + 415 + ], + "type": "text", + "content": "[4] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audiodriven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 416, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 416, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 74, + 416, + 296, + 470 + ], + "type": "text", + "content": "[5] Tenglong Ao, Qingzhe Gao, Yuke Lou, Baoquan Chen, and Libin Liu. Rhythmic gesticulator: Rhythm-aware co-speech gesture synthesis with hierarchical neural embeddings. ACM Transactions on Graphics (TOG), 41(6): 1-19, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 472, + 296, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 472, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 74, + 472, + 296, + 516 + ], + "type": "text", + "content": "[6] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 518, + 296, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 518, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 74, + 518, + 296, + 552 + ], + "type": "text", + "content": "[7] Ben Athiwaratkun, Andrew Gordon Wilson, and Anima Anandkumar. Probabilistic fasttext for multi-sense word embeddings. arXiv preprint arXiv:1806.02901, 2018. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 552, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 552, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 74, + 552, + 296, + 597 + ], + "type": "text", + "content": "[8] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the association for computational linguistics, 5:135-146, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 597, + 296, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 597, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 74, + 597, + 296, + 653 + ], + "type": "text", + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291–7299, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 654, + 296, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 654, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 70, + 654, + 296, + 699 + ], + "type": "text", + "content": "[10] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 700, + 296, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 700, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 70, + 700, + 296, + 723 + ], + "type": "text", + "content": "[11] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 541, + 723 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 541, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 541, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 541, + 117 + ], + "type": "text", + "content": "Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 541, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 541, + 150 + ], + "type": "text", + "content": "[12] Ylva Ferstl, Michael Neff, and Rachel McDonnell. Adversarial gesture generation with realistic gesture phasing. Computers & Graphics, 89:117-130, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 541, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 541, + 225 + ], + "type": "text", + "content": "[13] Esam Ghaleb, Bulat Khaertdinov, Wim Pouw, Marlou Rasenberg, Judith Holler, Asli Ozyurek, and Raquel Fernandez. Learning co-speech gesture representations in dialogue through contrastive learning: An intrinsic evaluation. In Proceedings of the 26th International Conference on Multimodal Interaction, pages 274-283, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 227, + 541, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 227, + 541, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 541, + 293 + ], + "type": "text", + "content": "[14] Esam Ghaleb, Bulat Khaertdinov, Asli Özyürek, and Raquel Fernández. I see what you mean: Co-speech gestures for reference resolution in multimodal dialogue. In Proceedings of the 63rd Conference of the Association for Computational Linguistics (ACL Findings), 2025. To appear. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 293, + 541, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 541, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 541, + 348 + ], + "type": "text", + "content": "[15] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 348, + 541, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 541, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 541, + 403 + ], + "type": "text", + "content": "[16] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 403, + 541, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 541, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 541, + 469 + ], + "type": "text", + "content": "[17] Ikhsanul Habibie, Mohamed Elgharib, Kripasindhu Sarkar, Ahsan Abdullah, Simbarashe Nyatsanga, Michael Neff, and Christian Theobalt. A motion matching-based framework for controllable gesture synthesis from speech. In ACM SIGGRAPH 2022 conference proceedings, pages 1-9, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 469, + 541, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 469, + 541, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 469, + 541, + 513 + ], + "type": "text", + "content": "[18] Gustav Eje Henter, Simon Alexanderson, and Jonas Beskow. Moglow: Probabilistic and controllable motion synthesis using normalising flows. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 541, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 541, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 541, + 546 + ], + "type": "text", + "content": "[19] Judith Holler and Stephen C Levinson. Multimodal language processing in human communication. Trends in Cognitive Sciences, 23(8):639-652, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 547, + 541, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 547, + 541, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 547, + 541, + 612 + ], + "type": "text", + "content": "[20] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29:3451-3460, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 541, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 541, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 541, + 655 + ], + "type": "text", + "content": "[21] Adam Kendon. Gesture units, gesture phrases and speech. In *Gesture: Visible Action as Utterance*, chapter 7, page 108–126. Cambridge University Press, 2004. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 656, + 541, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 541, + 723 + ], + "type": "text", + "content": "[22] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "text", + "content": "13971" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 297, + 723 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 70, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 297, + 127 + ], + "type": "text", + "content": "[23] Buyu Li, Yongchi Zhao, Shi Zhelun, and Lu Sheng. Danceformer: Music conditioned 3d dance generation with parametric motion transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1279, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 128, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 128, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 128, + 295, + 194 + ], + "type": "text", + "content": "[24] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 194, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 194, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 70, + 194, + 295, + 249 + ], + "type": "text", + "content": "[25] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 70, + 250, + 295, + 304 + ], + "type": "text", + "content": "[26] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 304, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 304, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 70, + 304, + 295, + 369 + ], + "type": "text", + "content": "[27] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European Conference on Computer Vision, pages 612-630. Springer, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 370, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 370, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 70, + 370, + 295, + 446 + ], + "type": "text", + "content": "[28] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 70, + 447, + 295, + 491 + ], + "type": "text", + "content": "[29] Li Liu, Lufei Gao, Wentao Lei, Fengji Ma, Xiaotian Lin, and Jinting Wang. A survey on deep multi-modal learning for body language recognition and generation. arXiv preprint arXiv:2308.08849, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 491, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 491, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 70, + 491, + 295, + 545 + ], + "type": "text", + "content": "[30] Lanmiao Liu, Chuang Yu, Siyang Song, Zhidong Su, and Adriana Tapus. Human gesture recognition with a flow-based model for human robot interaction. In *Compan- ion of the 2023 ACM/IEEE International Conference on Human-Robot Interaction*, pages 548–551, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 546, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 546, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 70, + 546, + 295, + 590 + ], + "type": "text", + "content": "[31] Pinxin Liu, Luchuan Song, Junhua Huang, Haiyang Liu, and Chenliang Xu. Gesturelsm: Latent shortcut based co-speech gesture generation with spatial-temporal modeling. arXiv preprint arXiv:2501.18898, 2025. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 70, + 590, + 295, + 633 + ], + "type": "text", + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yuanqi Du, Wayne Wu, Dahua Lin, and Ziwei Liu. Audio-driven co-speech gesture video generation. Advances in Neural Information Processing Systems, 35:21386-21399, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 634, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 634, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 70, + 634, + 295, + 700 + ], + "type": "text", + "content": "[33] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei Zhou. Learning hierarchical cross-modal association for co-speech gesture generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10462-10472, 2022. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 700, + 295, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 700, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 70, + 700, + 295, + 723 + ], + "type": "text", + "content": "[34] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 541, + 723 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 541, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 541, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 541, + 106 + ], + "type": "text", + "content": "holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 541, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 541, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 541, + 127 + ], + "type": "text", + "content": "[35] David McNeill. Hand and mind. Advances in Visual Semiotics, 351, 1992. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 541, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 541, + 194 + ], + "type": "text", + "content": "[36] Muhammad Hamza Mughal, Rishabh Dabral, Ikhsanul Habibie, Lucia Donatelli, Marc Habermann, and Christian Theobalt. Convofusion: Multi-modal conversational diffusion for co-speech gesture synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1388-1398, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 194, + 540, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 540, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 540, + 238 + ], + "type": "text", + "content": "[37] M Hamza Mughal, Rishabh Dabral, Merel CJ Scholman, Vera Demberg, and Christian Theobalt. Retrieving semantics from the deep: an rag solution for gesture synthesis. arXiv preprint arXiv:2412.06786, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 540, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 540, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 540, + 304 + ], + "type": "text", + "content": "[38] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 540, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 540, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 540, + 358 + ], + "type": "text", + "content": "[39] Mang Ning, Mingxiao Li, Jianlin Su, Haozhe Jia, Lanmiao Liu, Martin Beneš, Wenshuo Chen, Albert Ali Salah, and Itir Onal Ertugrul. Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2412.15032, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 360, + 540, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 540, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 540, + 414 + ], + "type": "text", + "content": "[40] Simbarashe Nyatsanga, Taras Kucherenko, Chaitanya Ahuja, Gustav Eje Henter, and Michael Neff. A comprehensive review of data-driven co-speech gesture generation. In Computer Graphics Forum, pages 569-596. Wiley Online Library, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 540, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 540, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 540, + 458 + ], + "type": "text", + "content": "[41] Asli Özyürek. Hearing and seeing meaning in speech and gesture: Insights from brain and behaviour. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1651):20130296, 2014. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 459, + 540, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 540, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 540, + 513 + ], + "type": "text", + "content": "[42] Kunkun Pang, Dafei Qin, Yingruo Fan, Julian Habekost, Takaaki Shiratori, Junichi Yamagishi, and Taku Komura. Bodyformer: Semantics-guided 3d body gesture synthesis with transformer. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 541, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 541, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 541, + 579 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 580, + 540, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 540, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 540, + 655 + ], + "type": "text", + "content": "[44] Xingqun Qi, Jiahao Pan, Peng Li, Ruibin Yuan, Xiaowei Chi, Mengfei Li, Wenhan Luo, Wei Xue, Shanghang Zhang, Qifeng Liu, et al. Weakly-supervised emotion transition learning for diverse 3d co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10424-10434, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 656, + 540, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 540, + 723 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 540, + 723 + ], + "type": "text", + "content": "[45] Mingyang Sun, Mengchen Zhao, Yaqing Hou, Minglei Li, Huang Xu, Songcen Xu, and Jianye Hao. Cospeech gesture synthesis by reinforcement learning with contrastive pre-trained rewards. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2331-2340, 2023. 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 318, + 763 + ], + "type": "text", + "content": "13972" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 73, + 295, + 722 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 70, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 70, + 73, + 294, + 116 + ], + "type": "text", + "content": "[46] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 118, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 118, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 71, + 118, + 295, + 160 + ], + "type": "text", + "content": "[47] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 71, + 162, + 294, + 194 + ], + "type": "text", + "content": "[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 70, + 195, + 294, + 248 + ], + "type": "text", + "content": "[49] Hendric Voß and Stefan Kopp. Augmented co-speech gesture generation: Including form and meaning features to guide learning-based gesture synthesis. In Proceedings of the 23rd ACM International Conference on Intelligent Virtual Agents, pages 1-8, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 251, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 251, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 70, + 251, + 295, + 303 + ], + "type": "text", + "content": "[50] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven cospeech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 304, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 304, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 70, + 304, + 294, + 380 + ], + "type": "text", + "content": "[51] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, and Haolin Zhuang. Qpgesture: Quantization-based and phase-guided motion matching for natural speech-driven gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2321-2330, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 381, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 381, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 70, + 381, + 294, + 435 + ], + "type": "text", + "content": "[52] Sheng Ye, Yu-Hui Wen, Yanan Sun, Ying He, Ziyang Zhang, Yaoyuan Wang, Weihua He, and Yong-Jin Liu. Audio-driven stylized gesture generation with flow-based model. In European Conference on Computer Vision, pages 712-728. Springer, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 437, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 437, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 70, + 437, + 294, + 479 + ], + "type": "text", + "content": "[53] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 481, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 481, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 70, + 481, + 294, + 544 + ], + "type": "text", + "content": "[54] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yan-dong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 70, + 547, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 547, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 70, + 547, + 294, + 600 + ], + "type": "text", + "content": "[55] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39(6):1-16, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 70, + 602, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 602, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 70, + 602, + 294, + 644 + ], + "type": "text", + "content": "[56] Pengfei Zhang, Pinxin Liu, Hyeongwoo Kim, Pablo Garrido, and Bindita Chaudhuri. Kinmo: Kinematic-aware human motion understanding and generation. arXiv preprint arXiv:2411.15472, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 646, + 294, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 646, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 70, + 646, + 294, + 699 + ], + "type": "text", + "content": "[57] Zeyi Zhang, Tenglong Ao, Yuyao Zhang, Qingzhe Gao, Chuan Lin, Baoquan Chen, and Libin Liu. Semantic gesticulator: Semantics-aware co-speech gesture synthesis. ACM Transactions on Graphics (TOG), 43(4):1-17, 2024. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 701, + 294, + 722 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 701, + 294, + 722 + ], + "spans": [ + { + "bbox": [ + 70, + 701, + 294, + 722 + ], + "type": "text", + "content": "[58] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 541, + 173 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 74, + 539, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 539, + 116 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 539, + 116 + ], + "type": "text", + "content": "Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 541, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 541, + 173 + ], + "type": "text", + "content": "[59] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audiodriven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2, 3, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 294, + 753, + 317, + 762 + ], + "type": "text", + "content": "13973" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_content_list.json b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2466b964cb0ff2f7a76b99952d5305cb5b7a250e --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_content_list.json @@ -0,0 +1,1446 @@ +[ + { + "type": "text", + "text": "SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis", + "text_level": 1, + "bbox": [ + 163, + 128, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71477c7f42ee897a304a1baaff56e265bfc8ff144b3bf0b898e4065bbbead50f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 200, + 875, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49b2f3acde280b07ad3698249e8c8c9186570f4d3f6fe253805a2336977286cd.jpg", + "image_caption": [ + "Figure 1. On the left, we analyze semantic labels from the BEAT2 dataset [31] and visualize frame-level motion, revealing that semantically relevant motions are rare and sparse, aligning with real-life observations. On the right, this observation drives the design of SemTalk, which establishes a rhythm-aligned base motion and dynamically emphasizes sparse semantic gestures at the frame-level. In this example, SemTalk amplifies expressiveness on words like \"watching\" and \"just,\" enhancing gesture and torso movements. The semantic scores below are automatically generated by SemTalk to modulate semantic emphasis over time." + ], + "image_footnote": [], + "bbox": [ + 114, + 282, + 370, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d782c86a891238469b65982dca38bade915fe2835c8fecb4deb855fcb404705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 282, + 885, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 546, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A good co-speech motion generation cannot be achieved without a careful integration of common rhythmic motion and rare yet essential semantic motion. In this work, we propose SemTalk for holistic co-speech motion generation with frame-level semantic emphasis. Our key insight is to separately learn base motions and sparse motions, and then adaptively fuse them. In particular, coarse2fine cross-attention module and rhythmic consistency learning are explored to establish rhythm-related base motion, ensuring a coherent foundation that synchronizes gestures with the speech rhythm. Subsequently, semantic emphasis learning is designed to generate semantic-aware sparse motion, focusing on frame-level semantic cues. Finally, to integrate sparse motion into the base motion and generate semantic-emphasized co-speech gestures, we further leverage a learned semantic score for adaptive synthesis. Qualitative and quantitative comparisons on two public datasets demonstrate that our method outperforms the state-of-the-art, delivering high-quality co-speech motion", + "bbox": [ + 88, + 578, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with enhanced semantic richness over a stable base motion.", + "bbox": [ + 513, + 547, + 906, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 613, + 645, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonverbal communication, including body language, hand gestures, and facial expressions, is integral to human interactions. It enriches conversations with contextual cues and enhances understanding among participants [6, 14, 20, 24]. This aspect is particularly significant in holistic co-speech motion generation, where the challenge lies in synthesizing gestures that align with speech rhythm while also capturing the infrequent yet critical semantic gestures [25, 38].", + "bbox": [ + 509, + 641, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing methods [17, 30, 48] rely heavily on rhythm-related audio features as conditions for gesture generation. While these rhythm-based features successfully align gestures with the timing of speech, they often overshadow the sparse yet expressive semantic motion (see Fig. 1). As a result, the generated motions may lack the contextual depth necessary and nuanced expressiveness for natural interaction. Some methods try to address this by incorporating semantic information like emotion, style, and", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "content[10, 12, 23, 32]. However, the rhythm features tend to dominate, making the models difficult to capture sparse, semantically relevant gestures at the frame level. These rare but impactful gestures are often diluted or overlooked, highlighting the challenge of balancing rhythmic alignment with semantic expressiveness in co-speech motion generation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In real-world human conversations, we have an observation that while most speech-related gestures are indeed rhythm-related, only a limited number of frames involve semantically emphasized gestures. This insight suggests that co-speech motions can be decomposed into two distinct components: (i) Rhythm-related base motion. These provide a continuous, coherent base motion aligned with the speech rhythm, reflecting the natural timing of speaking. (ii) Semantic-aware sparse motion: These occur infrequently but are essential for conveying specific meanings or emphasizing key points within the conversation.", + "bbox": [ + 89, + 184, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this observation, we propose a new framework SemTalk. SemTalk models the base motion and the sparse motion separately and then fuses them adaptively to generate high-fidelity co-speech motion. Specifically, we first focus on generating rhythm-related base motion by introducing coarse2fine cross-attention module and rhythmic consistency learning. We design a hierarchical coarse2fine cross-attention module, which progressively refines the base motion cues in a coarse-to-fine manner, starting from the face and moving through the hands, upper body, and lower body. This approach ensures consistent rhythmic transmission across all body parts, enhancing coherence base motion. Moreover, we propose a local-global rhythmic consistency learning approach, which enforces alignment at both the frame and sequence levels. Locally, a frame-level consistency loss ensures that each frame is precisely synchronized with its corresponding speech features, guaranteeing accurate temporal alignment. Globally, a sequence-level consistency loss sustains a coherent rhythmic flow across the entire motion sequence, preserving consistency throughout the generated gestures.", + "bbox": [ + 91, + 353, + 482, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we introduce semantic emphasis learning approach, which focuses on generating semantic-aware sparse motion. This approach utilizes frame-level semantic cues from textual information, high-level speech features, and emotion to identify frames that require emphasis through a learned semantic score produced by a gating strategy, i.e., sem-gate. The sem-gate is designed to dynamically activate semantic motions at key frames through two weighting methods applied on the motion condition and the loss, respectively, and semantic label guidance, allowing the model to produce motion that enhances the motion with deeper semantic meaning and contextual relevance.", + "bbox": [ + 89, + 672, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the base motion and sparse motion are integrated through semantic score-based motion fusion, which adaptively amplifies expressiveness by incorporating semantic-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aware key frames into the rhythm-related base motion. Our contributions are summarized below:", + "bbox": [ + 513, + 90, + 872, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SemTalk, a novel framework for holistic co-speech motion generation that separately models rhythm-related base motion and semantic-aware sparse motion, adaptively integrating them via a learned semantic gate.", + "- We propose a hierarchical coarse2fine cross-attention module to refine base motion and a local-global rhythmic consistency learning to integrate latent face and hand features with rhythm-related priors, ensuring coherence and rhythmic consistency. We then propose semantic emphasis learning to generate semantic gestures at certain frames, enhancing semantic-aware sparse motion.", + "- Experimental results show that our model surpasses state-of-the-art methods qualitatively and quantitatively, achieving higher motion quality and richer semantics." + ], + "bbox": [ + 514, + 121, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Co-speech Gesture Generation. Co-speech gesture generation aims to produce gestures aligned with speech. Early rule-based methods [7, 19, 21, 22, 41] lacked variability, while deterministic models [5, 7, 29, 36, 46, 49] mapped speech directly to gestures. Probabilistic models, including GANs [1, 17, 40] and diffusion models [2, 10, 47, 54], introduced variability. Some methods incorporated semantic cues, such as HA2G [32] and SEEG [28], which used hierarchical networks and alignment techniques. SynTalk [8] employs prompt-based control but treats inputs as signal strengths rather than fully interpreting semantics. LivelySpeaker [53] combines rhythmic features and semantic cues using CLIP [39] but struggles to integrate gestures with rhythm and capture semantics consistently, moreover, it only provides global control, limiting fine-grained refinement. DisCo [29] disentangles content and rhythm but lacks explicit modeling of sparse semantic gestures. SemTalk addresses this by separately modeling rhythm-related base motion and semantic-aware sparse motion, integrating them adaptively through a learned semantic score.", + "bbox": [ + 511, + 372, + 903, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Holistic Co-speech Motion Generation. Generating synchronized, expressive full-body motion from speech remains challenging, especially in coordinating the face, hands, and torso [9, 31, 34, 37, 48, 52]. Early methods introduced generative models to improve synchronization, but issues persisted. TalkSHOW [48] improved with VQ-VAE [42] cross-conditioning but handled facial expressions separately, causing fragmented outputs. DiffSHEG [9] and EMAGE [31] used separate encoders for expressions and gestures, but their unidirectional flow limited coherence. ProbTalk [33] leverages PQ-VAE [43] for improved body-facial synchronization but mainly relies on rhythmic cues, risking the loss of nuanced semantic gestures. Inspired by TM2D [15], which decomposes dance motion into music-related components, we separately model co-speech motion", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13762", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into rhythm-related and semantic-aware motion.", + "bbox": [ + 89, + 90, + 408, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 119, + 181, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary on RVQ-VAE", + "text_level": 1, + "bbox": [ + 89, + 145, + 328, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [4, 16, 51], our approach uses a residual vector-quantized autoencoder (RVQ-VAE) to progressively capture complex body movements in a few players. To retain unique motion characteristics across body regions, we segment the body into four parts—face, upper body, hands, and lower body—each with a dedicated RVQ-VAE, following [3, 31]. This segmentation preserves each part's dynamics and prevents feature entanglement.", + "bbox": [ + 89, + 167, + 483, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 89, + 297, + 200, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, our SemTalk pipeline includes two main components: the Base Motion Blocks $f_{r}(\\cdot)$ and the Sparse Motion Blocks $f_{b}(\\cdot)$ . Given rhythmic features $\\gamma_{b}, \\gamma_{h}$ , a seed pose $\\tilde{m}$ , and a speaker ID $id$ , the Base Motion Blocks generate rhythm-aligned codes $q^{b}$ , forming the rhythmic foundation of the base motion:", + "bbox": [ + 89, + 321, + 483, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r}: \\left(\\gamma_ {b}, \\gamma_ {h}, \\tilde {m}, i d; \\theta_ {f _ {r}}\\right)\\rightarrow q ^ {b}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 422, + 480, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{f_r}$ denotes the learnable parameters of the Base Motion Blocks. The Sparse Motion Blocks then take semantic features $\\phi_l$ , $\\phi_g$ , $\\phi_e$ , along with $\\gamma_h$ , $\\tilde{m}$ and $id$ , to produce frame-level semantic codes $q^s$ and semantic score $\\psi$ . $\\psi$ then triggers these codes only for semantically significant frames, producing a sparse motion representation:", + "bbox": [ + 89, + 453, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s}: \\left(\\phi_ {l}, \\phi_ {g}, \\phi_ {e}, \\tilde {m}, i d; \\theta_ {f _ {a}}\\right)\\rightarrow \\left(q ^ {s}, \\psi\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 556, + 480, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{f_s}$ represents the Sparse Motion Block parameters. Finally, the semantic emphasis mechanism $\\mathcal{E}$ combines $q^b$ and $q^s$ , guided by $\\psi$ , to form the final motion codes $q^m$ :", + "bbox": [ + 89, + 585, + 482, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {m} = \\mathcal {E} \\left(q ^ {b}, q ^ {s}; \\psi\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 642, + 480, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The motion decoder then uses $q^{m}$ to generate the output $m'$ .", + "bbox": [ + 89, + 671, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Generating Rhythm-related Base Motion", + "text_level": 1, + "bbox": [ + 89, + 710, + 441, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Base Motion Generation (Fig. 3 a) in SemTalk establishes a rhythmically aligned foundation by leveraging both rhythmic and speaker-specific features, enhancing the naturalness and personalization of generated motion.", + "bbox": [ + 89, + 734, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rhythmic Speech Encoding. To synchronize motion with speech, SemTalk incorporates rhythmic features: beats $\\gamma_{b}$ and HuBERT features $\\gamma_{h}$ . $\\gamma_{b}$ , derived from amplitude, short-time energy [11], and onset detection, mark key rhythmic points for aligning gestures with speech. Meanwhile, $\\gamma_{h}$ , extracted by the HuBERT encoder [18], captures high-level audio traits. In addition to rhythmic features $\\gamma$ ,", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1299c229de3e4d85275790079a0d9224f4c1d716d31f7f6960ad7aecdf6289bd.jpg", + "image_caption": [ + "Figure 2. An overview of the SemTalk pipeline. SemTalk generates holistic co-speech motion by first constructing rhythm-aligned $q^r$ in $f_r$ , guided by rhythmic consistency loss $L_{\\mathrm{Rhy}}$ . Meanwhile, $f_s$ produce frame-level semantic codes $q^s$ , activated selectively by the semantic score $\\psi$ . Finally, $q^m$ is achieved by fusing $q^r$ and $q^s$ based on $\\psi$ , with motion decoder, yielding synchronized and contextually enriched motions." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 887, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SemTalk uses a seed pose $\\tilde{m}$ and speaker identity $id$ to generate a personalized, rhythm-aligned latent pose $p$ . Then MLP-based Face Enhancement and Body Part-Aware modules utilize $\\gamma$ , $p$ and $id$ to obtain latent face $f_{e}$ , hands $f_{h}$ , upper body $f_{u}$ and lower body $f_{l}$ .", + "bbox": [ + 511, + 526, + 905, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse2Fine Cross-Attention Module. To facilitate the learning of base motion, we first proposed a transformer-based hierarchical Coarse2Fine Cross-Attn Module utilize $f_{e}$ , $f_{h}$ , $f_{u}$ and $f_{l}$ to obtain latent base motion $f_{b}$ . The refinement begins with $\\gamma$ for $f_{e}$ , which guides the rhythmic representation for $f_{h}$ , followed by conditioning $f_{u}$ and finally influencing $f_{l}$ . Since mouth movements closely correspond to speech syllables with minimal delay, we use the face to guide hand motions, inspired by DiffSHEG [9]. As the upper and lower body movements are less directly driven by speech and instead reflect the natural swinging of the hands and torso, we adopt cascading guidance: hands influence the upper body, which in turn drives the lower body. This structured approach, moving from the face to the hands, upper body, and lower body, ensures smooth and coherent motion propagation across the entire body.", + "bbox": [ + 511, + 607, + 906, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rhythmic Consistency Learning. Inspired by CoG's use of InfoNCE loss [45] to synchronize facial expressions with audio cues, our approach adopts a similar philosophy of", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/144d417bbc31eefff39cb9d8e50d3900e837a9275475db832198428d94277b60.jpg", + "image_caption": [ + "Figure 3. Architecture of SemTalk. SemTalk generates holistic co-speech motion in three stages. (a) Base Motion Generation uses rhythmic consistency learning to produce rhythm-aligned codes $q^b$ , conditioned on rhythmic features $\\gamma_b, \\gamma_h$ . (b) Sparse Motion Generation employs semantic emphasis learning to generate semantic codes $q^s$ , activated by semantic score $\\psi$ . (c) Adaptively Fusion automatically combines $q^b$ and $q^s$ based on $\\psi$ to produce mixed codes $q^m$ at frame level for rhythmically aligned and contextually rich motions." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 859, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aligning motion and speech rhythm. It can be defined as:", + "bbox": [ + 89, + 382, + 465, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R h y}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {j}\\right) / \\tau\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 409, + 480, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the number of frames(or the batch size), $\\tau$ denotes the temperature hyperparameter, $h(\\cdot)$ is the projection head for latent motion, $f_{i}$ and $\\gamma_h^i$ are the latent motion and rhythmic features at frame (or sample) $i$ , and $\\mathrm{sim}(\\cdot)$ represents cosine similarity.", + "bbox": [ + 89, + 468, + 482, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike CoG, our approach fundamentally differs by incorporating separate local and global rhythmic consistency losses, which are applied to both latent face $f_{e}$ and latent hands $f_{h}$ , ensuring a more cohesive and synchronized representation across the entire motion sequence. This rhythmic consistency loss ensures that the motions are not only synchronized at the frame level but also maintain a consistent rhythmic flow across the entire sequence.", + "bbox": [ + 89, + 545, + 482, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The local frame-level consistency loss $\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}$ aligns the motion features of each frame with the corresponding rhythmic cues $\\gamma_{h}$ . By leveraging HuBERT features $\\gamma_{h}$ instead of basic beat features $\\gamma_{b}$ , which only capture rhythmic pauses, we incorporate rich, high-level audio representations that enhance the model's ability to capture rhythm-related motion patterns and maintain temporal coherence.", + "bbox": [ + 89, + 666, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The global sentence-level consistency loss $\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}$ is designed to ensure rhythmic coherence at a global level. Unlike local loss, $\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}$ reinforces rhythm consistency throughout the sequence, ensuring that the generated motion maintains smooth and rhythm-aligned throughout its duration.", + "bbox": [ + 89, + 773, + 482, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By jointly minimizing $\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}$ and $\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}$ , rhythmic consistency learning enables SemTalk to produce base motions that are rhythmically aligned and temporally cohesive,", + "bbox": [ + 89, + 852, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "forming a solid rhythm-related base motion foundation.", + "bbox": [ + 511, + 382, + 879, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Generating Semantic-aware Sparse Motion", + "text_level": 1, + "bbox": [ + 511, + 409, + 880, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Sparse Motion Generation (Fig. 3 b) in SemTalk adds semantic-aware sparse motion to base motion by incorporating semantic cues drawn from speech content and emotional tone. By separating rhythm and semantics, this stage enhances motion generation by emphasizing contextually meaningful motion at key semantic moments.", + "bbox": [ + 511, + 431, + 903, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Speech Encoding. To capture semantic cues in speech, similar to [10], Semantic Emphasis Learning combines frame-level text embeddings $\\phi_t$ , sentence-level features $\\phi_g$ from the CLIP model [39], and emotion features $\\phi_e$ from the emotion2vec model [35]. These features form a comprehensive semantic representation $f_t$ , together with audio feature $\\gamma_h$ , that reflects both the content and emotional undertones of speech, enabling SemTalk to activate motions that are sensitive to nuanced semantic cues.", + "bbox": [ + 511, + 522, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Emphasis Learning. The process begins by generating $f_{t}$ , combining local and global cues from text, speech, emotion embeddings and HuBERT features $\\gamma_{h}$ . Then, the sem-gate leverages multi-modal inputs to generate a semantic score, identifying frames that require enhanced semantic emphasis. The sem-gate in SemTalk refines keyframe motion by applying two forms of weighting methods $\\mathcal{W}$ : feature weighting $\\mathcal{W}_{f}$ and loss weighting $\\mathcal{W}_{l}$ . Using $f_{t}$ and $\\gamma_{h}$ , SemTalk computes a semantic score $\\psi$ which dynamically scales feature weighting—filtering back semantic features $f_{t}$ to activate frames with significant relevance, ensuring that the model emphasizes frames aligned with specific communicative intentions. Second, the loss weighting is applied by supervising $\\psi$ , with a classification loss $\\mathcal{L}_{cls}^{G}$ based on semantic labels, further enhancing the model's ability to identify key frames. The two weight-", + "bbox": [ + 511, + 659, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c2aefad23af9a6868a75803373bfe37fae4fe3203eb7f4810dda18663f9b900a.jpg", + "image_caption": [ + "Figure 4. Concept comparison with LivelySpeaker [53]. (Top) LivelySpeaker generates semantic gestures with CLIP embeddings in SAG and refines rhythm-related gestures separately using diffusion, causing potential jitter. (Bottom) SemTalk integrates text and speech, uses a semantic gate for fine-grained control, and unifies rhythm and semantics for smoother, more coherent motions." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 460, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing methods allow SemTalk to selectively enhance semantic gestures while suppressing uninformative motion, leading to more expressive co-speech motion.", + "bbox": [ + 89, + 320, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once $\\psi$ is established, it modulates the integration of rhythm-aligned base motion $f_{b}$ and sparse semantic motion $f_{s}$ . Through alpha-blending, frames with high semantic relevance draw more from $f_{s}$ , while others rely on $f_{b}$ . The final motion codes $q^{s}$ are computed as:", + "bbox": [ + 89, + 364, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {s} = M L P \\left(\\psi f _ {s} + (1 - \\psi) f _ {b}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 455, + 482, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure cohesive propagation of semantic emphasis across body regions, we employ the Coarse2Fine Cross-Attention Module, similar to Sec. 3.3. In this stage, we focuses solely on body motion, excluding facial movements, as body gestures play a more critical role in conveying semantic meaning in co-speech interactions.", + "bbox": [ + 89, + 478, + 482, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To foster diverse motion generation, SemTalk includes a code classification loss $\\mathcal{L}_{cls}$ and a reconstruction loss $\\mathcal{L}_{rec}$ . These losses are specifically focused on frames with high semantic scores, guiding the model to prioritize the generation of sparse, meaningful gestures.", + "bbox": [ + 89, + 569, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. Recently, LivelySpeaker [53] designs the Semantic-Aware Generator (SAG) and Rhythm-Aware Generator (RAG) for co-speech gesture generation, combining them through beat empowerment. While effective, key differences exist between LivelySpeaker and SemTalk, see Fig. 4. First, SAG generates gestures from text using CLIP embeddings, but bridging words and expressive gestures is challenging, causing jitter. SemTalk incorporates speech features (pitch, tone, emotion) alongside text and GT supervision for adaptive gestures. Second, LivelySpeaker applies global control, missing local semantic details, while SemTalk uses fine-grained, frame-level semantic control for subtle variations. Third, LivelySpeaker fuses SAG and RAG in separate latent spaces, leading to misalignment and inconsistencies. SemTalk jointly models rhythm and semantics in a unified framework, ensuring smoother transitions and coherence. We further compare SAG with our", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantic gate in experiments.", + "bbox": [ + 513, + 92, + 712, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Semantic Score-based motion fusion", + "text_level": 1, + "bbox": [ + 513, + 117, + 828, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Adaptive Fusion stage (Fig. 3 c) in SemTalk seamlessly integrates semantic-aware sparse motion into the rhythmic-related base motion. By strategically enhancing frames based on their semantic importance, it maintains a smooth and natural motion flow across sequences. For each frame $i$ , the semantic score $\\psi_i$ computed during the Sparse Motion Generation stage is compared to a threshold $\\beta$ . If $\\psi_i > \\beta$ , the base motion's latent code $q_i^r$ is replaced with the sparse semantic code $q_i^s$ , effectively highlighting expressive gestures where they are most relevant; otherwise, $q_i = q_i^r$ .", + "bbox": [ + 511, + 140, + 903, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This selective replacement emphasizes semantically critical gestures while preserving the natural rhythmic base motion. By blending $q^{b}$ and $q^{s}$ based on semantic scores, SemTalk adapts to the expressive needs of the speech context while ensuring coherence. Additionally, the convolution structure of the RVQ-VAE decoder ensures smooth transitions between frames, preserving motion continuity.", + "bbox": [ + 511, + 292, + 903, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 412, + 645, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 513, + 439, + 702, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For training and evaluation, we use two datasets: BEAT2 and SHOW. BEAT2, introduced in EMAGE [31], extends BEAT [30] with 76 hours of data from 30 speakers, standardized into a mesh representation with paired audio, text, and frame-level semantic labels. We follow [31] and use the BEAT2-standard subset with an $85\\% / 7.5\\% / 7.5\\%$ train/val/test split. SHOW [48] includes 26.9 hours of high-quality talk show videos with 3D body meshes at 30fps. Since it lacks frame-level semantic labels, we use the semgate from SemTalk, pre-trained on BEAT2, to generate them. Following [48], we select video clips longer than 10 seconds and split the data $80\\% / 10\\% / 10\\%$ for train/val/test.", + "bbox": [ + 511, + 460, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our model is trained on a single NVIDIA A100 GPU for 200 epochs with a batch size of 64. We use RVQ-VAE [42], downscaling by 4. The residual quantization has 6 layers, a codebook size of 256 and a dropout rate of 0.2. We use five transformer layers to predict the last five layer codes. In Base Motion Learning, $\\tau = 0.1$ ; in Sparse Motion Learning, $\\beta = 0.5$ empirically. The training uses ADAM with a 1e-4 learning rate. Following [31], we start with a 4-frame seed pose, gradually increasing masked frames from 0 to $40\\%$ over 120 epochs.", + "bbox": [ + 511, + 643, + 908, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate generated body gestures using FGD [50] to measure distributional alignment with GT, reflecting realism. DIV [26] quantifies gesture variation via the average L1 distance across clips. BC [27] assesses speech-motion synchrony. For facial expressions, we use MSE [47] to quantify positional differences and LVD [48] to measure discrepancies between GT and generated facial vertices.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ecfa083670d714d79dd4c95afe184dc95c93b4c4df5646584ffc0173e277becc.jpg", + "image_caption": [ + "Figure 5. Comparison on BEAT2 [31] Dataset. SemTalk* refers to the model trained solely on the Base Motion Generation stage, capturing rhythmic alignment but lacking semantic gestures. In contrast, SemTalk successfully emphasized sparse yet vivid motions. For instance, when saying \"my opinion,\" SemTalk generates a hand-raising gesture followed by an index finger extension for emphasis. Similarly, for \"never tell,\" our model produces a clear, repeated gesture matching the rhythm, reinforcing the intended emphasis." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 897, + 510 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f09db62adf066365352e12b4bbf519f8db3b6c167df8738489b0a3ac7649784f.jpg", + "image_caption": [ + "Figure 6. Comparison on SHOW [48] Dataset. Our method performs better in motion diversity and semantic richness." + ], + "image_footnote": [], + "bbox": [ + 133, + 578, + 282, + 672 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d15e04660b009631faa3ac6e9e6bff96583a48bd97717a24d032bc6b4078ce89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 578, + 864, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 705, + 274, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We encourage readers to watch our demo video for a clearer understanding of SemTalk's qualitative performance. Our method achieves superior speech-motion alignment, generating more realistic, diverse, and semantically consistent gestures than the baselines. As shown in Fig. 5, LivelySpeaker, TalkSHOW, EMAGE, and DiffSHEG exhibit jitter—EMAGE mainly in the legs and shoulders, while TalkSHOW affects the entire body. LivelySpeaker and DiffSHEG, which focus primarily on the upper body, produce slow and inconsistent motions, especially at speech clip boundaries. DiffSHEG improves", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gesture diversity over IMAGE and TalkSHOW, though IMAGE maintains greater naturalness. SemTalk surpasses all baselines in both realism and diversity. Compared to SemTalk*, SemTalk generates more expressive gestures, emphasizing key phrases (e.g., raising hands for “dream job” or pointing for “that is why”). While SemTalk* ensures rhythmic consistency, it lacks semantic expressiveness. By integrating frame-level semantic emphasis, SemTalk aligns motion with both rhythm and semantics, demonstrating the effectiveness of rhythmic consistency learning and semantic emphasis learning. In facial comparisons (Fig. 7), IMAGE shows minimal lip movement, while both DiffSHEG and", + "bbox": [ + 511, + 705, + 908, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13766", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a717cf177762c731d3be4305e5fb997193b304e071da862c3175abdaaafa8d47.jpg", + "image_caption": [ + "Figure 7. Facial Comparison on the BEAT2 [31] Dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 486, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a501893e2c1a842880f342ab42f131c7d71c2b26442d82b3b6605150e683eb60.jpg", + "image_caption": [ + "Figure 8. Qualitative study on semantic score. Semantic score aligns with keywords, influencing gesture intensity." + ], + "image_footnote": [], + "bbox": [ + 109, + 246, + 455, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EMAGE reveal inconsistencies between lip motion and the rhythm of speech. In contrast, SemTalk produces smooth, natural transitions across syllables, resulting in realistic and expressive lips, significantly surpassing the baselines.", + "bbox": [ + 89, + 446, + 482, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the SHOW dataset (Fig. 6), SemTalk shows more agile gestures than all baselines, when applied to unseen data. Our method captures natural and contextually rich gestures, particularly in moments of emphasis such as \"I like to do\" and \"relaxing,\" where our model produces lively hand and body movements that align with the speech content.", + "bbox": [ + 89, + 508, + 482, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Score. Fig. 8 shows how semantic emphasis influences gesture intensity, with peaks in the semantic score aligning with keywords like \"comes,\" \"fantastic,\" and \"captured.\" By extracting semantic scores from key frames, we track gesture emphasis trends. Furthermore, as shown in Fig. 9, SemTalk adapts to different emotional tones even when the text remains unchanged. This adaptability prevents overfitting to the text itself, allowing the model to generate gestures that vary according to the emotional delivery of the speech. The learned semantic score provides fine-grained, frame-level control, keeping gestures both rhythmically synchronized and semantically aligned in real time. User Study. We conducted a user study with 10 video samples and 25 participants from diverse backgrounds, evaluating realism, semantic consistency, motion-speech synchrony, and diversity. Participants were required to rank shuffled videos across different methods. As shown in Fig. 10, our approach received dominant preferences across all metrics, especially in semantic consistency and realism.", + "bbox": [ + 89, + 598, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/760d1fd802be9416574db3e923376768c8d359df700277ab247316886991aed9.jpg", + "image_caption": [ + "Figure 9. Same words with different speech from the internet. \"emo\" represents different emotional tones extracted from speech. SemTalk can generate different motions, even when the text script is the same, preventing overfitting to the text itself." + ], + "image_footnote": [], + "bbox": [ + 521, + 88, + 905, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1f5db3de63a6fab1a877bb143eb01bf4d2b4cb70944e878ac8e7d740c826275.jpg", + "image_caption": [ + "Figure 10. Results of the user study." + ], + "image_footnote": [], + "bbox": [ + 539, + 291, + 885, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 436, + 705, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Baselines. As shown in Tab. 1, SemTalk outperforms previous methods on BEAT2, achieving lower FGD, MSE, and LVD, indicating better distribution alignment and reduced motion errors. For fairness, we follow [31] and add a lower-body VQ-VAE to TalkSHOW, Diff-SHEG, and SemTalk. Notably, SemTalk significantly reduces FGD, ensuring strong distribution matching. While TalkSHOW and EMAGE achieve competitive diversity (DIV) scores, SemTalk balances high semantic relevance with natural motion flow.", + "bbox": [ + 511, + 459, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the SHOW dataset, SemTalk excels with the lowest FGD, MSE, and the highest BC, indicating precise beat alignment with the audio and enhanced semantic consistency in generated motions. Although EMAGE exhibits high DIV, our model achieves comparable results while maintaining smooth, realistic motion free from jitter.", + "bbox": [ + 511, + 612, + 905, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sem-gate. Tab. 2 highlights the effectiveness of sem-gate. Without sem-gate, the model fails to emphasize key moments. Randomized semantic scores led to poor performance by preventing meaningful frame distinction. Introducing a learned sem-gate even (w/o $\\mathcal{W}$ ) significantly improves semantic alignment and classification accuracy. Refinement is further enhanced through weighting strategies: feature weighting $\\mathcal{W}_f$ enhances motion emphasis, while loss weighting $\\mathcal{W}_l$ improves FGD and overall accuracy. These results suggest that weighting methods enhance the accuracy of the semantic score and help the model prioritize important frames. The best results come from applying two weighting methods together, where frames with stronger se", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/58edef6b5eff11c192d309fb2a029ed365faf04ed21a1a6efe927bab1cd944e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodFGD↓BC↑DIV↑MSE↓LVD↓
BEAT2FaceFormer [13]---7.7877.593
CodeTalker [44]---8.0267.766
CaMN [30]6.6446.76910.86--
DSG [47]8.8117.24111.49--
LivelySpeaker [17]11.806.65911.28--
Habibie et al. [17]9.0407.7168.2138.6148.043
TalkSHOW [48]6.2096.94713.477.7917.771
EMAGE [31]5.5127.72413.067.6807.556
DiffSHEG [9]8.9867.14211.917.6658.673
SemTalk (Ours)4.2787.77012.916.1536.938
SHOWFaceFormer [13]---138.143.69
CodeTalker [44]---140.745.84
CaMN [30]22.127.71210.37--
DSG [47]24.848.02710.23--
LivelySpeaker [17]32.177.84410.14--
Habibie et al. [17]27.228.2098.541145.647.35
TalkSHOW [48]24.438.24910.98139.645.17
EMAGE [31]22.128.28012.46136.142.44
DiffSHEG [9]24.878.06110.79139.045.77
SemTalk (Ours)20.188.30411.36134.139.15
", + "bbox": [ + 99, + 88, + 475, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mantic signals receive higher emphasis. We also compare sem-gate with LivelySpeaker's SAG [53]. We find that replacing the Sparse Motion stage with SAG and substituting motion using GT semantic labels led to poor performance. SAG relies only on text-motion alignment, ignoring emotional tone, making it more prone to overfitting the text. In contrast, our sem-gate applies GT supervision with two weighting methods, achieving more accurate and stable semantic motion.", + "bbox": [ + 89, + 441, + 483, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on Components. We assess the impact of each component of our model on BEAT2 and present the results in Tab. 3, which reveals several key insights (more ablation results please see supplementary material):", + "bbox": [ + 89, + 580, + 483, + 642 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Rhythmic Consistency Learning (RC) not only boosts performance on key metrics like FGD, LVD, and BC but also reduces the MSE, contributing to smoother and more realistic base motion.", + "- Semantic Emphasis Learning (SE) proves essential for selectively enhancing semantic-rich gestures. The inclusion of SE, as shown in rows with SE enabled, improves both diversity (DIV) and FGD, enabling the model to emphasize semantically relevant motions. SE demonstrates its effectiveness in focusing on frame-level semantic information, which contributes to the generation of lifelike gestures with enriched contextual meaning.", + "- Coarse2Fine Cross-Attention Module (C2F) effectively refines motion details, improving BC, FGD, and DIV. When combined with RVQ and RC, C2F achieves the best MSE and LVD, highlighting its role in enhancing motion realism and diversity hierarchically." + ], + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46044737566a791b5e411fd91acb5ab4722fbdb7212022197ec7a2d652deb453.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with SOTA. SemTalk consistently outperforms baselines across both the BEAT2 and SHOW datasets. Lower values are better for FMD, FGD, MSE, and LVD. Higher values are better for BC and DIV. We report $\\mathrm{FGD} \\times 10^{-1}$ , $\\mathrm{BC} \\times 10^{-1}$ , $\\mathrm{MSE} \\times 10^{-8}$ and $\\mathrm{LVD} \\times 10^{-5}$ for simplify." + ], + "table_footnote": [], + "table_body": "
MethodFGD↓BC↑DIV↑Acc(%)↑
w/o Sem-gate4.8937.70212.42-
SAG (LivelySpeaker [53])4.6187.68212.45-
Sem-gate (Random ψ)4.6347.70012.4450.07
Sem-gate (w/o W)4.4957.63312.2672.32
Sem-gate (w/ Wf)4.4087.67912.2878.52
Sem-gate (w/ Wl)4.3667.77211.9477.83
Sem-gate (ours)4.2787.77012.9182.76
", + "bbox": [ + 526, + 88, + 895, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/44f4ce8af71fc71a6a99c3ea070234be949d6df598113e8ea29ab0ea3b82177e.jpg", + "table_caption": [ + "Table 2. Ablation study on Sem-gate. \"Acc\" denotes semantic classification performance on BEAT2. \"w/o Sem-gate\" means directly input $f_{t}$ and $\\gamma_{h}$ without Sem-gate. \"SAG (LivelySpeaker [53])\" replaces the Sparse Motion Generation stage with LivelySpeaker's SAG method. \"Random $\\psi$ \" assigns frame-level scores randomly. \"w/o $\\mathcal{W}$ \" applies the semantic gate but excludes frame-level weighting. \"w/ $\\mathcal{W}_{f}$ \" applies feature weighting. \"w/ $\\mathcal{W}_{l}$ \" applies loss weighting. (as mentioned in Sec. 3.4). Sem-gate (ours) integrates both the semantic gate and frame-level weighting to enhance emphasis." + ], + "table_footnote": [], + "table_body": "
RCSEC2FRVQFGD↓BC↑DIV↑MSE↓LVD↓
----6.2347.62811.448.2397.831
---√5.4847.64111.8413.88215.42
√--√4.8677.70112.386.2016.928
√√-√4.5267.75112.836.2156.997
--√√4.8977.70212.4213.41615.72
√√--5.8317.75811.976.5877.106
√-√√4.3977.77612.496.1006.898
√√√√4.2787.77012.916.1536.938
", + "bbox": [ + 526, + 345, + 895, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on each key component. \"RC\" denotes rhythmic consistency learning, \"SE\" denotes the semantic emphasis learning, and \"C2F\" denotes Coarse2Fine Cross-Att Module, \"RVQ\" denotes the RVQ-VAE.", + "bbox": [ + 511, + 465, + 903, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "- RVQ-VAE (RVQ) enhances the diversity and realism of generated motion. Though it slightly increases MSE and LVD, it notably improves FGD, leading to more natural motion generation compared to standard VQ-VAE.", + "bbox": [ + 513, + 542, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose SemTalk, a novel approach for holistic cospeech motion generation with frame-level semantic emphasis. Our method addresses the integration of sparse yet expressive motion into foundational rhythm-related motion, which has received less attention in previous works. We develop a framework that separately learns rhythm-related base motion through coarse2fine cross-attention module and rhythmic consistency learning, while capturing semantic-aware motion through Semantic Emphasis Learning. These components are then adaptively fused based on a learned semantic score. Our approach has demonstrated state-of-the-art performance on two public datasets quantitatively and qualitatively. The qualitative results and user study show that our method can generate high-quality cospeech motion sequences that enhance frame-level semantics over robust base motions, reflecting the full spectrum of human expressiveness.", + "bbox": [ + 509, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13768", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by Alibaba Research Intern Program, the Young Scientists Fund of the National Natural Science Foundation of China No. 624B2110, the National Key Research and Development Program of China No. 2024YFC3015600, the Fundamental Research Funds for Central Universities No.2042023KF0180 & No.2042025KF0053. The numerical calculation is supported by supercomputing system in Super-computing Center of Wuhan University and Tongyi Lab, Alibaba Group.", + "bbox": [ + 89, + 90, + 485, + 233 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 257, + 187, + 273 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566–20576, 2022. 2", + "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2", + "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 3", + "[4] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 31:2523-2533, 2023. 3", + "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2", + "[6] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1", + "[7] Justine Cassell, Hannes Högni Vilhjalmsson, and Timothy Bickmore. Beat: the behavior expression animation toolkit. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 477-486, 2001. 2", + "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2", + "[9] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7352-7361, 2024. 2, 3, 8" + ], + "bbox": [ + 99, + 282, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al. Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 2, 4", + "[11] Selina Chu, Shrikanth Narayanan, and C-C Jay Kuo. Environmental sound recognition with time-frequency audio features. IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142-1158, 2009. 3", + "[12] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 2", + "[13] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 8", + "[14] Susan Goldin-Meadow. The role of gesture in communication and thinking. Trends in cognitive sciences, 3(11):419-429, 1999. 1", + "[15] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942–9952, 2023. 2", + "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 3", + "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 1, 2, 8", + "[18] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 3", + "[19] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In Proceedings of the seventh annual ACM/IEEE international conference on Human-Robot Interaction, pages 25-32, 2012. 2", + "[20] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1", + "[21] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2", + "[22] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thorisson, and Hannes Vilhjalmsson. Towards a common" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13769", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2", + "[23] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2", + "[24] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In Proceedings of the 26th International Conference on Intelligent User Interfaces, pages 11-21, 2021. 1", + "[25] Alex Lascarides and Matthew Stone. A formal semantic analysis of gesture. Journal of Semantics, 26(4):393-449, 2009. 1", + "[26] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 5", + "[27] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 5", + "[28] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2", + "[29] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM international conference on multimedia, pages 3764-3773, 2022. 2", + "[30] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European conference on computer vision, pages 612-630. Springer, 2022. 1, 5, 8", + "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 1, 2, 3, 5, 6, 7, 8", + "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2", + "[33] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 2", + "[34] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2", + "[35] Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023. 4", + "[36] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics symposium on computer animation, pages 25-35, 2013. 2", + "[37] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 2", + "[38] Ashi Özyürek, Roel M Willems, Sotaro Kita, and Peter Hagoort. On-line integration of semantic information from speech and gesture: Insights from event-related brain potentials. Journal of cognitive neuroscience, 19(4):605-616, 2007. 1", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[40] Manuel Rebol, Christian Güttl, and Krzysztof Pietroszek. Real-time gesture animation generation from speech for virtual human interaction. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-4, 2021. 2", + "[41] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized talking head synthesis. arXiv preprint arXiv:2301.03786, 2(4):5, 2023. 2", + "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5", + "[43] Hanwei Wu and Markus Flierl. Learning product codebooks using vector-quantized autoencoders for image retrieval. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5. IEEE, 2019. 2", + "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13770", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 8", + "[45] Zunnan Xu, Yachao Zhang, Sicheng Yang, Ronghui Li, and Xiu Li. Chain of generation: Multi-modal gesture synthesis via cascaded conditional control. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6387-6395, 2024. 3", + "[46] Sicheng Yang, Zilin Wang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Qiaochu Huang, Lei Hao, Songcen Xu, Xiaofei Wu, Changpeng Yang, et al. Unified gesture: A unified gesture synthesis model for multiple skeletons. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1033-1044, 2023. 2", + "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 5, 8", + "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 1, 2, 5, 6, 8", + "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2", + "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 5", + "[51] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 3", + "[52] Jinsong Zhang, Minjie Zhu, Yuxiang Zhang, Zerong Zheng, Yebin Liu, and Kun Li. Speechact: Towards generating whole-body motion from speech. IEEE Transactions on Visualization and Computer Graphics, 2025. 2", + "[53] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 5, 8", + "[54] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2" + ], + "bbox": [ + 91, + 90, + 483, + 852 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13771", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_model.json b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c750cd023128447efa183ed926edd00187b3a043 --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.13, + 0.834, + 0.178 + ], + "angle": 0, + "content": "SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.202, + 0.877, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.284, + 0.372, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.284, + 0.886, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.459, + 0.908, + 0.531 + ], + "angle": 0, + "content": "Figure 1. On the left, we analyze semantic labels from the BEAT2 dataset [31] and visualize frame-level motion, revealing that semantically relevant motions are rare and sparse, aligning with real-life observations. On the right, this observation drives the design of SemTalk, which establishes a rhythm-aligned base motion and dynamically emphasizes sparse semantic gestures at the frame-level. In this example, SemTalk amplifies expressiveness on words like \"watching\" and \"just,\" enhancing gesture and torso movements. The semantic scores below are automatically generated by SemTalk to modulate semantic emphasis over time." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.547, + 0.328, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.579, + 0.486, + 0.868 + ], + "angle": 0, + "content": "A good co-speech motion generation cannot be achieved without a careful integration of common rhythmic motion and rare yet essential semantic motion. In this work, we propose SemTalk for holistic co-speech motion generation with frame-level semantic emphasis. Our key insight is to separately learn base motions and sparse motions, and then adaptively fuse them. In particular, coarse2fine cross-attention module and rhythmic consistency learning are explored to establish rhythm-related base motion, ensuring a coherent foundation that synchronizes gestures with the speech rhythm. Subsequently, semantic emphasis learning is designed to generate semantic-aware sparse motion, focusing on frame-level semantic cues. Finally, to integrate sparse motion into the base motion and generate semantic-emphasized co-speech gestures, we further leverage a learned semantic score for adaptive synthesis. Qualitative and quantitative comparisons on two public datasets demonstrate that our method outperforms the state-of-the-art, delivering high-quality co-speech motion" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.549, + 0.907, + 0.563 + ], + "angle": 0, + "content": "with enhanced semantic richness over a stable base motion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.614, + 0.646, + 0.63 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.642, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Nonverbal communication, including body language, hand gestures, and facial expressions, is integral to human interactions. It enriches conversations with contextual cues and enhances understanding among participants [6, 14, 20, 24]. This aspect is particularly significant in holistic co-speech motion generation, where the challenge lies in synthesizing gestures that align with speech rhythm while also capturing the infrequent yet critical semantic gestures [25, 38]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Most existing methods [17, 30, 48] rely heavily on rhythm-related audio features as conditions for gesture generation. While these rhythm-based features successfully align gestures with the timing of speech, they often overshadow the sparse yet expressive semantic motion (see Fig. 1). As a result, the generated motions may lack the contextual depth necessary and nuanced expressiveness for natural interaction. Some methods try to address this by incorporating semantic information like emotion, style, and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.24, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13761" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "content[10, 12, 23, 32]. However, the rhythm features tend to dominate, making the models difficult to capture sparse, semantically relevant gestures at the frame level. These rare but impactful gestures are often diluted or overlooked, highlighting the challenge of balancing rhythmic alignment with semantic expressiveness in co-speech motion generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.35 + ], + "angle": 0, + "content": "In real-world human conversations, we have an observation that while most speech-related gestures are indeed rhythm-related, only a limited number of frames involve semantically emphasized gestures. This insight suggests that co-speech motions can be decomposed into two distinct components: (i) Rhythm-related base motion. These provide a continuous, coherent base motion aligned with the speech rhythm, reflecting the natural timing of speaking. (ii) Semantic-aware sparse motion: These occur infrequently but are essential for conveying specific meanings or emphasizing key points within the conversation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.354, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Inspired by this observation, we propose a new framework SemTalk. SemTalk models the base motion and the sparse motion separately and then fuses them adaptively to generate high-fidelity co-speech motion. Specifically, we first focus on generating rhythm-related base motion by introducing coarse2fine cross-attention module and rhythmic consistency learning. We design a hierarchical coarse2fine cross-attention module, which progressively refines the base motion cues in a coarse-to-fine manner, starting from the face and moving through the hands, upper body, and lower body. This approach ensures consistent rhythmic transmission across all body parts, enhancing coherence base motion. Moreover, we propose a local-global rhythmic consistency learning approach, which enforces alignment at both the frame and sequence levels. Locally, a frame-level consistency loss ensures that each frame is precisely synchronized with its corresponding speech features, guaranteeing accurate temporal alignment. Globally, a sequence-level consistency loss sustains a coherent rhythmic flow across the entire motion sequence, preserving consistency throughout the generated gestures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Furthermore, we introduce semantic emphasis learning approach, which focuses on generating semantic-aware sparse motion. This approach utilizes frame-level semantic cues from textual information, high-level speech features, and emotion to identify frames that require emphasis through a learned semantic score produced by a gating strategy, i.e., sem-gate. The sem-gate is designed to dynamically activate semantic motions at key frames through two weighting methods applied on the motion condition and the loss, respectively, and semantic label guidance, allowing the model to produce motion that enhances the motion with deeper semantic meaning and contextual relevance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, the base motion and sparse motion are integrated through semantic score-based motion fusion, which adaptively amplifies expressiveness by incorporating semantic-" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.874, + 0.121 + ], + "angle": 0, + "content": "aware key frames into the rhythm-related base motion. Our contributions are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We propose SemTalk, a novel framework for holistic co-speech motion generation that separately models rhythm-related base motion and semantic-aware sparse motion, adaptively integrating them via a learned semantic gate." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.183, + 0.905, + 0.287 + ], + "angle": 0, + "content": "- We propose a hierarchical coarse2fine cross-attention module to refine base motion and a local-global rhythmic consistency learning to integrate latent face and hand features with rhythm-related priors, ensuring coherence and rhythmic consistency. We then propose semantic emphasis learning to generate semantic gestures at certain frames, enhancing semantic-aware sparse motion." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.289, + 0.904, + 0.333 + ], + "angle": 0, + "content": "- Experimental results show that our model surpasses state-of-the-art methods qualitatively and quantitatively, achieving higher motion quality and richer semantics." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.905, + 0.674 + ], + "angle": 0, + "content": "Co-speech Gesture Generation. Co-speech gesture generation aims to produce gestures aligned with speech. Early rule-based methods [7, 19, 21, 22, 41] lacked variability, while deterministic models [5, 7, 29, 36, 46, 49] mapped speech directly to gestures. Probabilistic models, including GANs [1, 17, 40] and diffusion models [2, 10, 47, 54], introduced variability. Some methods incorporated semantic cues, such as HA2G [32] and SEEG [28], which used hierarchical networks and alignment techniques. SynTalk [8] employs prompt-based control but treats inputs as signal strengths rather than fully interpreting semantics. LivelySpeaker [53] combines rhythmic features and semantic cues using CLIP [39] but struggles to integrate gestures with rhythm and capture semantics consistently, moreover, it only provides global control, limiting fine-grained refinement. DisCo [29] disentangles content and rhythm but lacks explicit modeling of sparse semantic gestures. SemTalk addresses this by separately modeling rhythm-related base motion and semantic-aware sparse motion, integrating them adaptively through a learned semantic score." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Holistic Co-speech Motion Generation. Generating synchronized, expressive full-body motion from speech remains challenging, especially in coordinating the face, hands, and torso [9, 31, 34, 37, 48, 52]. Early methods introduced generative models to improve synchronization, but issues persisted. TalkSHOW [48] improved with VQ-VAE [42] cross-conditioning but handled facial expressions separately, causing fragmented outputs. DiffSHEG [9] and EMAGE [31] used separate encoders for expressions and gestures, but their unidirectional flow limited coherence. ProbTalk [33] leverages PQ-VAE [43] for improved body-facial synchronization but mainly relies on rhythmic cues, risking the loss of nuanced semantic gestures. Inspired by TM2D [15], which decomposes dance motion into music-related components, we separately model co-speech motion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13762" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.41, + 0.106 + ], + "angle": 0, + "content": "into rhythm-related and semantic-aware motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.121, + 0.182, + 0.136 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.146, + 0.329, + 0.163 + ], + "angle": 0, + "content": "3.1. Preliminary on RVQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.484, + 0.29 + ], + "angle": 0, + "content": "Following [4, 16, 51], our approach uses a residual vector-quantized autoencoder (RVQ-VAE) to progressively capture complex body movements in a few players. To retain unique motion characteristics across body regions, we segment the body into four parts—face, upper body, hands, and lower body—each with a dedicated RVQ-VAE, following [3, 31]. This segmentation preserves each part's dynamics and prevents feature entanglement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.299, + 0.201, + 0.314 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.412 + ], + "angle": 0, + "content": "As shown in Figure 2, our SemTalk pipeline includes two main components: the Base Motion Blocks \\( f_{r}(\\cdot) \\) and the Sparse Motion Blocks \\( f_{b}(\\cdot) \\). Given rhythmic features \\( \\gamma_{b}, \\gamma_{h} \\), a seed pose \\( \\tilde{m} \\), and a speaker ID \\( id \\), the Base Motion Blocks generate rhythm-aligned codes \\( q^{b} \\), forming the rhythmic foundation of the base motion:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.424, + 0.482, + 0.442 + ], + "angle": 0, + "content": "\\[\nf _ {r}: \\left(\\gamma_ {b}, \\gamma_ {h}, \\tilde {m}, i d; \\theta_ {f _ {r}}\\right)\\rightarrow q ^ {b}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.545 + ], + "angle": 0, + "content": "where \\(\\theta_{f_r}\\) denotes the learnable parameters of the Base Motion Blocks. The Sparse Motion Blocks then take semantic features \\(\\phi_l\\), \\(\\phi_g\\), \\(\\phi_e\\), along with \\(\\gamma_h\\), \\(\\tilde{m}\\) and \\(id\\), to produce frame-level semantic codes \\(q^s\\) and semantic score \\(\\psi\\). \\(\\psi\\) then triggers these codes only for semantically significant frames, producing a sparse motion representation:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.557, + 0.482, + 0.574 + ], + "angle": 0, + "content": "\\[\nf _ {s}: \\left(\\phi_ {l}, \\phi_ {g}, \\phi_ {e}, \\tilde {m}, i d; \\theta_ {f _ {a}}\\right)\\rightarrow \\left(q ^ {s}, \\psi\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.632 + ], + "angle": 0, + "content": "where \\(\\theta_{f_s}\\) represents the Sparse Motion Block parameters. Finally, the semantic emphasis mechanism \\(\\mathcal{E}\\) combines \\(q^b\\) and \\(q^s\\), guided by \\(\\psi\\), to form the final motion codes \\(q^m\\):" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.643, + 0.482, + 0.66 + ], + "angle": 0, + "content": "\\[\nq ^ {m} = \\mathcal {E} \\left(q ^ {b}, q ^ {s}; \\psi\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.687 + ], + "angle": 0, + "content": "The motion decoder then uses \\( q^{m} \\) to generate the output \\( m' \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.442, + 0.729 + ], + "angle": 0, + "content": "3.3. Generating Rhythm-related Base Motion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.795 + ], + "angle": 0, + "content": "The Base Motion Generation (Fig. 3 a) in SemTalk establishes a rhythmically aligned foundation by leveraging both rhythmic and speaker-specific features, enhancing the naturalness and personalization of generated motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Rhythmic Speech Encoding. To synchronize motion with speech, SemTalk incorporates rhythmic features: beats \\(\\gamma_{b}\\) and HuBERT features \\(\\gamma_{h}\\). \\(\\gamma_{b}\\), derived from amplitude, short-time energy [11], and onset detection, mark key rhythmic points for aligning gestures with speech. Meanwhile, \\(\\gamma_{h}\\), extracted by the HuBERT encoder [18], captures high-level audio traits. In addition to rhythmic features \\(\\gamma\\)," + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.888, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.403, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Figure 2. An overview of the SemTalk pipeline. SemTalk generates holistic co-speech motion by first constructing rhythm-aligned \\( q^r \\) in \\( f_r \\), guided by rhythmic consistency loss \\( L_{\\mathrm{Rhy}} \\). Meanwhile, \\( f_s \\) produce frame-level semantic codes \\( q^s \\), activated selectively by the semantic score \\( \\psi \\). Finally, \\( q^m \\) is achieved by fusing \\( q^r \\) and \\( q^s \\) based on \\( \\psi \\), with motion decoder, yielding synchronized and contextually enriched motions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.603 + ], + "angle": 0, + "content": "SemTalk uses a seed pose \\(\\tilde{m}\\) and speaker identity \\(id\\) to generate a personalized, rhythm-aligned latent pose \\(p\\). Then MLP-based Face Enhancement and Body Part-Aware modules utilize \\(\\gamma\\), \\(p\\) and \\(id\\) to obtain latent face \\(f_{e}\\), hands \\(f_{h}\\), upper body \\(f_{u}\\) and lower body \\(f_{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Coarse2Fine Cross-Attention Module. To facilitate the learning of base motion, we first proposed a transformer-based hierarchical Coarse2Fine Cross-Attn Module utilize \\( f_{e} \\), \\( f_{h} \\), \\( f_{u} \\) and \\( f_{l} \\) to obtain latent base motion \\( f_{b} \\). The refinement begins with \\( \\gamma \\) for \\( f_{e} \\), which guides the rhythmic representation for \\( f_{h} \\), followed by conditioning \\( f_{u} \\) and finally influencing \\( f_{l} \\). Since mouth movements closely correspond to speech syllables with minimal delay, we use the face to guide hand motions, inspired by DiffSHEG [9]. As the upper and lower body movements are less directly driven by speech and instead reflect the natural swinging of the hands and torso, we adopt cascading guidance: hands influence the upper body, which in turn drives the lower body. This structured approach, moving from the face to the hands, upper body, and lower body, ensures smooth and coherent motion propagation across the entire body." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Rhythmic Consistency Learning. Inspired by CoG's use of InfoNCE loss [45] to synchronize facial expressions with audio cues, our approach adopts a similar philosophy of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.861, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Architecture of SemTalk. SemTalk generates holistic co-speech motion in three stages. (a) Base Motion Generation uses rhythmic consistency learning to produce rhythm-aligned codes \\( q^b \\), conditioned on rhythmic features \\( \\gamma_b, \\gamma_h \\). (b) Sparse Motion Generation employs semantic emphasis learning to generate semantic codes \\( q^s \\), activated by semantic score \\( \\psi \\). (c) Adaptively Fusion automatically combines \\( q^b \\) and \\( q^s \\) based on \\( \\psi \\) to produce mixed codes \\( q^m \\) at frame level for rhythmically aligned and contextually rich motions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.383, + 0.467, + 0.399 + ], + "angle": 0, + "content": "aligning motion and speech rhythm. It can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.41, + 0.482, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R h y}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {j}\\right) / \\tau\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.545 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of frames(or the batch size), \\(\\tau\\) denotes the temperature hyperparameter, \\(h(\\cdot)\\) is the projection head for latent motion, \\(f_{i}\\) and \\(\\gamma_h^i\\) are the latent motion and rhythmic features at frame (or sample) \\(i\\), and \\(\\mathrm{sim}(\\cdot)\\) represents cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.483, + 0.666 + ], + "angle": 0, + "content": "Unlike CoG, our approach fundamentally differs by incorporating separate local and global rhythmic consistency losses, which are applied to both latent face \\( f_{e} \\) and latent hands \\( f_{h} \\), ensuring a more cohesive and synchronized representation across the entire motion sequence. This rhythmic consistency loss ensures that the motions are not only synchronized at the frame level but also maintain a consistent rhythmic flow across the entire sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.773 + ], + "angle": 0, + "content": "The local frame-level consistency loss \\(\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}\\) aligns the motion features of each frame with the corresponding rhythmic cues \\(\\gamma_{h}\\). By leveraging HuBERT features \\(\\gamma_{h}\\) instead of basic beat features \\(\\gamma_{b}\\), which only capture rhythmic pauses, we incorporate rich, high-level audio representations that enhance the model's ability to capture rhythm-related motion patterns and maintain temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.483, + 0.853 + ], + "angle": 0, + "content": "The global sentence-level consistency loss \\(\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}\\) is designed to ensure rhythmic coherence at a global level. Unlike local loss, \\(\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}\\) reinforces rhythm consistency throughout the sequence, ensuring that the generated motion maintains smooth and rhythm-aligned throughout its duration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.853, + 0.484, + 0.902 + ], + "angle": 0, + "content": "By jointly minimizing \\(\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}\\) and \\(\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}\\), rhythmic consistency learning enables SemTalk to produce base motions that are rhythmically aligned and temporally cohesive," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.383, + 0.88, + 0.399 + ], + "angle": 0, + "content": "forming a solid rhythm-related base motion foundation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.881, + 0.426 + ], + "angle": 0, + "content": "3.4. Generating Semantic-aware Sparse Motion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.522 + ], + "angle": 0, + "content": "The Sparse Motion Generation (Fig. 3 b) in SemTalk adds semantic-aware sparse motion to base motion by incorporating semantic cues drawn from speech content and emotional tone. By separating rhythm and semantics, this stage enhances motion generation by emphasizing contextually meaningful motion at key semantic moments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Semantic Speech Encoding. To capture semantic cues in speech, similar to [10], Semantic Emphasis Learning combines frame-level text embeddings \\(\\phi_t\\), sentence-level features \\(\\phi_g\\) from the CLIP model [39], and emotion features \\(\\phi_e\\) from the emotion2vec model [35]. These features form a comprehensive semantic representation \\(f_t\\), together with audio feature \\(\\gamma_h\\), that reflects both the content and emotional undertones of speech, enabling SemTalk to activate motions that are sensitive to nuanced semantic cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Semantic Emphasis Learning. The process begins by generating \\( f_{t} \\), combining local and global cues from text, speech, emotion embeddings and HuBERT features \\( \\gamma_{h} \\). Then, the sem-gate leverages multi-modal inputs to generate a semantic score, identifying frames that require enhanced semantic emphasis. The sem-gate in SemTalk refines keyframe motion by applying two forms of weighting methods \\( \\mathcal{W} \\): feature weighting \\( \\mathcal{W}_{f} \\) and loss weighting \\( \\mathcal{W}_{l} \\). Using \\( f_{t} \\) and \\( \\gamma_{h} \\), SemTalk computes a semantic score \\( \\psi \\) which dynamically scales feature weighting—filtering back semantic features \\( f_{t} \\) to activate frames with significant relevance, ensuring that the model emphasizes frames aligned with specific communicative intentions. Second, the loss weighting is applied by supervising \\( \\psi \\), with a classification loss \\( \\mathcal{L}_{cls}^{G} \\) based on semantic labels, further enhancing the model's ability to identify key frames. The two weight-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.462, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.485, + 0.309 + ], + "angle": 0, + "content": "Figure 4. Concept comparison with LivelySpeaker [53]. (Top) LivelySpeaker generates semantic gestures with CLIP embeddings in SAG and refines rhythm-related gestures separately using diffusion, causing potential jitter. (Bottom) SemTalk integrates text and speech, uses a semantic gate for fine-grained control, and unifies rhythm and semantics for smoother, more coherent motions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.366 + ], + "angle": 0, + "content": "ing methods allow SemTalk to selectively enhance semantic gestures while suppressing uninformative motion, leading to more expressive co-speech motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Once \\(\\psi\\) is established, it modulates the integration of rhythm-aligned base motion \\(f_{b}\\) and sparse semantic motion \\(f_{s}\\). Through alpha-blending, frames with high semantic relevance draw more from \\(f_{s}\\), while others rely on \\(f_{b}\\). The final motion codes \\(q^{s}\\) are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.456, + 0.483, + 0.473 + ], + "angle": 0, + "content": "\\[\nq ^ {s} = M L P \\left(\\psi f _ {s} + (1 - \\psi) f _ {b}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.569 + ], + "angle": 0, + "content": "To ensure cohesive propagation of semantic emphasis across body regions, we employ the Coarse2Fine Cross-Attention Module, similar to Sec. 3.3. In this stage, we focuses solely on body motion, excluding facial movements, as body gestures play a more critical role in conveying semantic meaning in co-speech interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.645 + ], + "angle": 0, + "content": "To foster diverse motion generation, SemTalk includes a code classification loss \\(\\mathcal{L}_{cls}\\) and a reconstruction loss \\(\\mathcal{L}_{rec}\\). These losses are specifically focused on frames with high semantic scores, guiding the model to prioritize the generation of sparse, meaningful gestures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Discussion. Recently, LivelySpeaker [53] designs the Semantic-Aware Generator (SAG) and Rhythm-Aware Generator (RAG) for co-speech gesture generation, combining them through beat empowerment. While effective, key differences exist between LivelySpeaker and SemTalk, see Fig. 4. First, SAG generates gestures from text using CLIP embeddings, but bridging words and expressive gestures is challenging, causing jitter. SemTalk incorporates speech features (pitch, tone, emotion) alongside text and GT supervision for adaptive gestures. Second, LivelySpeaker applies global control, missing local semantic details, while SemTalk uses fine-grained, frame-level semantic control for subtle variations. Third, LivelySpeaker fuses SAG and RAG in separate latent spaces, leading to misalignment and inconsistencies. SemTalk jointly models rhythm and semantics in a unified framework, ensuring smoother transitions and coherence. We further compare SAG with our" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.714, + 0.107 + ], + "angle": 0, + "content": "semantic gate in experiments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.118, + 0.829, + 0.133 + ], + "angle": 0, + "content": "3.5. Semantic Score-based motion fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.905, + 0.292 + ], + "angle": 0, + "content": "The Adaptive Fusion stage (Fig. 3 c) in SemTalk seamlessly integrates semantic-aware sparse motion into the rhythmic-related base motion. By strategically enhancing frames based on their semantic importance, it maintains a smooth and natural motion flow across sequences. For each frame \\( i \\), the semantic score \\( \\psi_i \\) computed during the Sparse Motion Generation stage is compared to a threshold \\( \\beta \\). If \\( \\psi_i > \\beta \\), the base motion's latent code \\( q_i^r \\) is replaced with the sparse semantic code \\( q_i^s \\), effectively highlighting expressive gestures where they are most relevant; otherwise, \\( q_i = q_i^r \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.4 + ], + "angle": 0, + "content": "This selective replacement emphasizes semantically critical gestures while preserving the natural rhythmic base motion. By blending \\( q^{b} \\) and \\( q^{s} \\) based on semantic scores, SemTalk adapts to the expressive needs of the speech context while ensuring coherence. Additionally, the convolution structure of the RVQ-VAE decoder ensures smooth transitions between frames, preserving motion continuity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.414, + 0.646, + 0.431 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.44, + 0.704, + 0.456 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Datasets. For training and evaluation, we use two datasets: BEAT2 and SHOW. BEAT2, introduced in EMAGE [31], extends BEAT [30] with 76 hours of data from 30 speakers, standardized into a mesh representation with paired audio, text, and frame-level semantic labels. We follow [31] and use the BEAT2-standard subset with an \\(85\\% / 7.5\\% / 7.5\\%\\) train/val/test split. SHOW [48] includes 26.9 hours of high-quality talk show videos with 3D body meshes at 30fps. Since it lacks frame-level semantic labels, we use the semgate from SemTalk, pre-trained on BEAT2, to generate them. Following [48], we select video clips longer than 10 seconds and split the data \\(80\\% / 10\\% / 10\\%\\) for train/val/test." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Implementation Details. Our model is trained on a single NVIDIA A100 GPU for 200 epochs with a batch size of 64. We use RVQ-VAE [42], downscaling by 4. The residual quantization has 6 layers, a codebook size of 256 and a dropout rate of 0.2. We use five transformer layers to predict the last five layer codes. In Base Motion Learning, \\(\\tau = 0.1\\); in Sparse Motion Learning, \\(\\beta = 0.5\\) empirically. The training uses ADAM with a 1e-4 learning rate. Following [31], we start with a 4-frame seed pose, gradually increasing masked frames from 0 to \\(40\\%\\) over 120 epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics. We evaluate generated body gestures using FGD [50] to measure distributional alignment with GT, reflecting realism. DIV [26] quantifies gesture variation via the average L1 distance across clips. BC [27] assesses speech-motion synchrony. For facial expressions, we use MSE [47] to quantify positional differences and LVD [48] to measure discrepancies between GT and generated facial vertices." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.092, + 0.898, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.514, + 0.908, + 0.573 + ], + "angle": 0, + "content": "Figure 5. Comparison on BEAT2 [31] Dataset. SemTalk* refers to the model trained solely on the Base Motion Generation stage, capturing rhythmic alignment but lacking semantic gestures. In contrast, SemTalk successfully emphasized sparse yet vivid motions. For instance, when saying \"my opinion,\" SemTalk generates a hand-raising gesture followed by an index finger extension for emphasis. Similarly, for \"never tell,\" our model produces a clear, repeated gesture matching the rhythm, reinforcing the intended emphasis." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.579, + 0.283, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.579, + 0.865, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.684, + 0.854, + 0.699 + ], + "angle": 0, + "content": "Figure 6. Comparison on SHOW [48] Dataset. Our method performs better in motion diversity and semantic richness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.275, + 0.722 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparisons. We encourage readers to watch our demo video for a clearer understanding of SemTalk's qualitative performance. Our method achieves superior speech-motion alignment, generating more realistic, diverse, and semantically consistent gestures than the baselines. As shown in Fig. 5, LivelySpeaker, TalkSHOW, EMAGE, and DiffSHEG exhibit jitter—EMAGE mainly in the legs and shoulders, while TalkSHOW affects the entire body. LivelySpeaker and DiffSHEG, which focus primarily on the upper body, produce slow and inconsistent motions, especially at speech clip boundaries. DiffSHEG improves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.909, + 0.89 + ], + "angle": 0, + "content": "gesture diversity over IMAGE and TalkSHOW, though IMAGE maintains greater naturalness. SemTalk surpasses all baselines in both realism and diversity. Compared to SemTalk*, SemTalk generates more expressive gestures, emphasizing key phrases (e.g., raising hands for “dream job” or pointing for “that is why”). While SemTalk* ensures rhythmic consistency, it lacks semantic expressiveness. By integrating frame-level semantic emphasis, SemTalk aligns motion with both rhythm and semantics, demonstrating the effectiveness of rhythmic consistency learning and semantic emphasis learning. In facial comparisons (Fig. 7), IMAGE shows minimal lip movement, while both DiffSHEG and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.488, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.226, + 0.465, + 0.24 + ], + "angle": 0, + "content": "Figure 7. Facial Comparison on the BEAT2 [31] Dataset." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.247, + 0.456, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.405, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Figure 8. Qualitative study on semantic score. Semantic score aligns with keywords, influencing gesture intensity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.509 + ], + "angle": 0, + "content": "EMAGE reveal inconsistencies between lip motion and the rhythm of speech. In contrast, SemTalk produces smooth, natural transitions across syllables, resulting in realistic and expressive lips, significantly surpassing the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.599 + ], + "angle": 0, + "content": "On the SHOW dataset (Fig. 6), SemTalk shows more agile gestures than all baselines, when applied to unseen data. Our method captures natural and contextually rich gestures, particularly in moments of emphasis such as \"I like to do\" and \"relaxing,\" where our model produces lively hand and body movements that align with the speech content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Semantic Score. Fig. 8 shows how semantic emphasis influences gesture intensity, with peaks in the semantic score aligning with keywords like \"comes,\" \"fantastic,\" and \"captured.\" By extracting semantic scores from key frames, we track gesture emphasis trends. Furthermore, as shown in Fig. 9, SemTalk adapts to different emotional tones even when the text remains unchanged. This adaptability prevents overfitting to the text itself, allowing the model to generate gestures that vary according to the emotional delivery of the speech. The learned semantic score provides fine-grained, frame-level control, keeping gestures both rhythmically synchronized and semantically aligned in real time. User Study. We conducted a user study with 10 video samples and 25 participants from diverse backgrounds, evaluating realism, semantic consistency, motion-speech synchrony, and diversity. Participants were required to rank shuffled videos across different methods. As shown in Fig. 10, our approach received dominant preferences across all metrics, especially in semantic consistency and realism." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.089, + 0.906, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.231, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Figure 9. Same words with different speech from the internet. \"emo\" represents different emotional tones extracted from speech. SemTalk can generate different motions, even when the text script is the same, preventing overfitting to the text itself." + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.292, + 0.886, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.413, + 0.822, + 0.427 + ], + "angle": 0, + "content": "Figure 10. Results of the user study." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.707, + 0.453 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Comparison with Baselines. As shown in Tab. 1, SemTalk outperforms previous methods on BEAT2, achieving lower FGD, MSE, and LVD, indicating better distribution alignment and reduced motion errors. For fairness, we follow [31] and add a lower-body VQ-VAE to TalkSHOW, Diff-SHEG, and SemTalk. Notably, SemTalk significantly reduces FGD, ensuring strong distribution matching. While TalkSHOW and EMAGE achieve competitive diversity (DIV) scores, SemTalk balances high semantic relevance with natural motion flow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.703 + ], + "angle": 0, + "content": "On the SHOW dataset, SemTalk excels with the lowest FGD, MSE, and the highest BC, indicating precise beat alignment with the audio and enhanced semantic consistency in generated motions. Although EMAGE exhibits high DIV, our model achieves comparable results while maintaining smooth, realistic motion free from jitter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Sem-gate. Tab. 2 highlights the effectiveness of sem-gate. Without sem-gate, the model fails to emphasize key moments. Randomized semantic scores led to poor performance by preventing meaningful frame distinction. Introducing a learned sem-gate even (w/o \\(\\mathcal{W}\\)) significantly improves semantic alignment and classification accuracy. Refinement is further enhanced through weighting strategies: feature weighting \\(\\mathcal{W}_f\\) enhances motion emphasis, while loss weighting \\(\\mathcal{W}_l\\) improves FGD and overall accuracy. These results suggest that weighting methods enhance the accuracy of the semantic score and help the model prioritize important frames. The best results come from applying two weighting methods together, where frames with stronger se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13767" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.477, + 0.344 + ], + "angle": 0, + "content": "
DatasetMethodFGD↓BC↑DIV↑MSE↓LVD↓
BEAT2FaceFormer [13]---7.7877.593
CodeTalker [44]---8.0267.766
CaMN [30]6.6446.76910.86--
DSG [47]8.8117.24111.49--
LivelySpeaker [17]11.806.65911.28--
Habibie et al. [17]9.0407.7168.2138.6148.043
TalkSHOW [48]6.2096.94713.477.7917.771
EMAGE [31]5.5127.72413.067.6807.556
DiffSHEG [9]8.9867.14211.917.6658.673
SemTalk (Ours)4.2787.77012.916.1536.938
SHOWFaceFormer [13]---138.143.69
CodeTalker [44]---140.745.84
CaMN [30]22.127.71210.37--
DSG [47]24.848.02710.23--
LivelySpeaker [17]32.177.84410.14--
Habibie et al. [17]27.228.2098.541145.647.35
TalkSHOW [48]24.438.24910.98139.645.17
EMAGE [31]22.128.28012.46136.142.44
DiffSHEG [9]24.878.06110.79139.045.77
SemTalk (Ours)20.188.30411.36134.139.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.347, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with SOTA. SemTalk consistently outperforms baselines across both the BEAT2 and SHOW datasets. Lower values are better for FMD, FGD, MSE, and LVD. Higher values are better for BC and DIV. We report \\(\\mathrm{FGD} \\times 10^{-1}\\), \\(\\mathrm{BC} \\times 10^{-1}\\), \\(\\mathrm{MSE} \\times 10^{-8}\\) and \\(\\mathrm{LVD} \\times 10^{-5}\\) for simplify." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.578 + ], + "angle": 0, + "content": "mantic signals receive higher emphasis. We also compare sem-gate with LivelySpeaker's SAG [53]. We find that replacing the Sparse Motion stage with SAG and substituting motion using GT semantic labels led to poor performance. SAG relies only on text-motion alignment, ignoring emotional tone, making it more prone to overfitting the text. In contrast, our sem-gate applies GT supervision with two weighting methods, achieving more accurate and stable semantic motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Ablation Study on Components. We assess the impact of each component of our model on BEAT2 and present the results in Tab. 3, which reveals several key insights (more ablation results please see supplementary material):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.703 + ], + "angle": 0, + "content": "- Rhythmic Consistency Learning (RC) not only boosts performance on key metrics like FGD, LVD, and BC but also reduces the MSE, contributing to smoother and more realistic base motion." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- Semantic Emphasis Learning (SE) proves essential for selectively enhancing semantic-rich gestures. The inclusion of SE, as shown in rows with SE enabled, improves both diversity (DIV) and FGD, enabling the model to emphasize semantically relevant motions. SE demonstrates its effectiveness in focusing on frame-level semantic information, which contributes to the generation of lifelike gestures with enriched contextual meaning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Coarse2Fine Cross-Attention Module (C2F) effectively refines motion details, improving BC, FGD, and DIV. When combined with RVQ and RC, C2F achieves the best MSE and LVD, highlighting its role in enhancing motion realism and diversity hierarchically." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.897, + 0.201 + ], + "angle": 0, + "content": "
MethodFGD↓BC↑DIV↑Acc(%)↑
w/o Sem-gate4.8937.70212.42-
SAG (LivelySpeaker [53])4.6187.68212.45-
Sem-gate (Random ψ)4.6347.70012.4450.07
Sem-gate (w/o W)4.4957.63312.2672.32
Sem-gate (w/ Wf)4.4087.67912.2878.52
Sem-gate (w/ Wl)4.3667.77211.9477.83
Sem-gate (ours)4.2787.77012.9182.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.206, + 0.905, + 0.345 + ], + "angle": 0, + "content": "Table 2. Ablation study on Sem-gate. \"Acc\" denotes semantic classification performance on BEAT2. \"w/o Sem-gate\" means directly input \\( f_{t} \\) and \\( \\gamma_{h} \\) without Sem-gate. \"SAG (LivelySpeaker [53])\" replaces the Sparse Motion Generation stage with LivelySpeaker's SAG method. \"Random \\( \\psi \\)\" assigns frame-level scores randomly. \"w/o \\( \\mathcal{W} \\)\" applies the semantic gate but excludes frame-level weighting. \"w/ \\( \\mathcal{W}_{f} \\)\" applies feature weighting. \"w/ \\( \\mathcal{W}_{l} \\)\" applies loss weighting. (as mentioned in Sec. 3.4). Sem-gate (ours) integrates both the semantic gate and frame-level weighting to enhance emphasis." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.347, + 0.897, + 0.462 + ], + "angle": 0, + "content": "
RCSEC2FRVQFGD↓BC↑DIV↑MSE↓LVD↓
----6.2347.62811.448.2397.831
---√5.4847.64111.8413.88215.42
√--√4.8677.70112.386.2016.928
√√-√4.5267.75112.836.2156.997
--√√4.8977.70212.4213.41615.72
√√--5.8317.75811.976.5877.106
√-√√4.3977.77612.496.1006.898
√√√√4.2787.77012.916.1536.938
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.466, + 0.905, + 0.522 + ], + "angle": 0, + "content": "Table 3. Ablation study on each key component. \"RC\" denotes rhythmic consistency learning, \"SE\" denotes the semantic emphasis learning, and \"C2F\" denotes Coarse2Fine Cross-Att Module, \"RVQ\" denotes the RVQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.543, + 0.907, + 0.605 + ], + "angle": 0, + "content": "- RVQ-VAE (RVQ) enhances the diversity and realism of generated motion. Though it slightly increases MSE and LVD, it notably improves FGD, leading to more natural motion generation compared to standard VQ-VAE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose SemTalk, a novel approach for holistic cospeech motion generation with frame-level semantic emphasis. Our method addresses the integration of sparse yet expressive motion into foundational rhythm-related motion, which has received less attention in previous works. We develop a framework that separately learns rhythm-related base motion through coarse2fine cross-attention module and rhythmic consistency learning, while capturing semantic-aware motion through Semantic Emphasis Learning. These components are then adaptively fused based on a learned semantic score. Our approach has demonstrated state-of-the-art performance on two public datasets quantitatively and qualitatively. The qualitative results and user study show that our method can generate high-quality cospeech motion sequences that enhance frame-level semantics over robust base motions, reflecting the full spectrum of human expressiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13768" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.234 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by Alibaba Research Intern Program, the Young Scientists Fund of the National Natural Science Foundation of China No. 624B2110, the National Key Research and Development Program of China No. 2024YFC3015600, the Fundamental Research Funds for Central Universities No.2042023KF0180 & No.2042025KF0053. The numerical calculation is supported by supercomputing system in Super-computing Center of Wuhan University and Tongyi Lab, Alibaba Group." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.258, + 0.188, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566–20576, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[4] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 31:2523-2533, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[6] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[7] Justine Cassell, Hannes Högni Vilhjalmsson, and Timothy Bickmore. Beat: the behavior expression animation toolkit. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 477-486, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[9] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7352-7361, 2024. 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.284, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[10] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al. Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[11] Selina Chu, Shrikanth Narayanan, and C-C Jay Kuo. Environmental sound recognition with time-frequency audio features. IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142-1158, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[12] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.306, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[13] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[14] Susan Goldin-Meadow. The role of gesture in communication and thinking. Trends in cognitive sciences, 3(11):419-429, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[15] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942–9952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[18] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[19] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In Proceedings of the seventh annual ACM/IEEE international conference on Human-Robot Interaction, pages 25-32, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[20] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[21] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thorisson, and Hannes Vilhjalmsson. Towards a common" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13769" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.161 + ], + "angle": 0, + "content": "framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[23] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[24] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In Proceedings of the 26th International Conference on Intelligent User Interfaces, pages 11-21, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[25] Alex Lascarides and Matthew Stone. A formal semantic analysis of gesture. Journal of Semantics, 26(4):393-449, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[26] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[27] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[28] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[29] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM international conference on multimedia, pages 3764-3773, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[30] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European conference on computer vision, pages 612-630. Springer, 2022. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.149 + ], + "angle": 0, + "content": "Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[33] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[34] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[35] Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[36] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics symposium on computer animation, pages 25-35, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[37] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[38] Ashi Özyürek, Roel M Willems, Sotaro Kita, and Peter Hagoort. On-line integration of semantic information from speech and gesture: Insights from event-related brain potentials. Journal of cognitive neuroscience, 19(4):605-616, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[40] Manuel Rebol, Christian Güttl, and Krzysztof Pietroszek. Real-time gesture animation generation from speech for virtual human interaction. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-4, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[41] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized talking head synthesis. arXiv preprint arXiv:2301.03786, 2(4):5, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[43] Hanwei Wu and Markus Flierl. Learning product codebooks using vector-quantized autoencoders for image retrieval. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13770" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[45] Zunnan Xu, Yachao Zhang, Sicheng Yang, Ronghui Li, and Xiu Li. Chain of generation: Multi-modal gesture synthesis via cascaded conditional control. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6387-6395, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[46] Sicheng Yang, Zilin Wang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Qiaochu Huang, Lei Hao, Songcen Xu, Xiaofei Wu, Changpeng Yang, et al. Unified gesture: A unified gesture synthesis model for multiple skeletons. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1033-1044, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.483, + 0.513 + ], + "angle": 0, + "content": "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.483, + 0.653 + ], + "angle": 0, + "content": "[51] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.656, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[52] Jinsong Zhang, Minjie Zhu, Yuxiang Zhang, Zerong Zheng, Yebin Liu, and Kun Li. Speechact: Towards generating whole-body motion from speech. IEEE Transactions on Visualization and Computer Graphics, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.483, + 0.782 + ], + "angle": 0, + "content": "[53] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.784, + 0.483, + 0.853 + ], + "angle": 0, + "content": "[54] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13771" + } + ] +] \ No newline at end of file diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_origin.pdf b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8b1926d4cb96255da73016d61c2dbcc6082400c --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/d0c52370-d24a-4c84-9ea3-ee5cd6ca6239_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2224ed3409a318d3f2294388518ba4046521a18cdb8b83410a4db03be8a1b811 +size 14239274 diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/full.md b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..46e12daf04aa526f4cea86ae05dbc3cf5183528f --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/full.md @@ -0,0 +1,283 @@ +# SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis + +![](images/71477c7f42ee897a304a1baaff56e265bfc8ff144b3bf0b898e4065bbbead50f.jpg) + +![](images/49b2f3acde280b07ad3698249e8c8c9186570f4d3f6fe253805a2336977286cd.jpg) +Figure 1. On the left, we analyze semantic labels from the BEAT2 dataset [31] and visualize frame-level motion, revealing that semantically relevant motions are rare and sparse, aligning with real-life observations. On the right, this observation drives the design of SemTalk, which establishes a rhythm-aligned base motion and dynamically emphasizes sparse semantic gestures at the frame-level. In this example, SemTalk amplifies expressiveness on words like "watching" and "just," enhancing gesture and torso movements. The semantic scores below are automatically generated by SemTalk to modulate semantic emphasis over time. + +![](images/0d782c86a891238469b65982dca38bade915fe2835c8fecb4deb855fcb404705.jpg) + +# Abstract + +A good co-speech motion generation cannot be achieved without a careful integration of common rhythmic motion and rare yet essential semantic motion. In this work, we propose SemTalk for holistic co-speech motion generation with frame-level semantic emphasis. Our key insight is to separately learn base motions and sparse motions, and then adaptively fuse them. In particular, coarse2fine cross-attention module and rhythmic consistency learning are explored to establish rhythm-related base motion, ensuring a coherent foundation that synchronizes gestures with the speech rhythm. Subsequently, semantic emphasis learning is designed to generate semantic-aware sparse motion, focusing on frame-level semantic cues. Finally, to integrate sparse motion into the base motion and generate semantic-emphasized co-speech gestures, we further leverage a learned semantic score for adaptive synthesis. Qualitative and quantitative comparisons on two public datasets demonstrate that our method outperforms the state-of-the-art, delivering high-quality co-speech motion + +with enhanced semantic richness over a stable base motion. + +# 1. Introduction + +Nonverbal communication, including body language, hand gestures, and facial expressions, is integral to human interactions. It enriches conversations with contextual cues and enhances understanding among participants [6, 14, 20, 24]. This aspect is particularly significant in holistic co-speech motion generation, where the challenge lies in synthesizing gestures that align with speech rhythm while also capturing the infrequent yet critical semantic gestures [25, 38]. + +Most existing methods [17, 30, 48] rely heavily on rhythm-related audio features as conditions for gesture generation. While these rhythm-based features successfully align gestures with the timing of speech, they often overshadow the sparse yet expressive semantic motion (see Fig. 1). As a result, the generated motions may lack the contextual depth necessary and nuanced expressiveness for natural interaction. Some methods try to address this by incorporating semantic information like emotion, style, and + +content[10, 12, 23, 32]. However, the rhythm features tend to dominate, making the models difficult to capture sparse, semantically relevant gestures at the frame level. These rare but impactful gestures are often diluted or overlooked, highlighting the challenge of balancing rhythmic alignment with semantic expressiveness in co-speech motion generation. + +In real-world human conversations, we have an observation that while most speech-related gestures are indeed rhythm-related, only a limited number of frames involve semantically emphasized gestures. This insight suggests that co-speech motions can be decomposed into two distinct components: (i) Rhythm-related base motion. These provide a continuous, coherent base motion aligned with the speech rhythm, reflecting the natural timing of speaking. (ii) Semantic-aware sparse motion: These occur infrequently but are essential for conveying specific meanings or emphasizing key points within the conversation. + +Inspired by this observation, we propose a new framework SemTalk. SemTalk models the base motion and the sparse motion separately and then fuses them adaptively to generate high-fidelity co-speech motion. Specifically, we first focus on generating rhythm-related base motion by introducing coarse2fine cross-attention module and rhythmic consistency learning. We design a hierarchical coarse2fine cross-attention module, which progressively refines the base motion cues in a coarse-to-fine manner, starting from the face and moving through the hands, upper body, and lower body. This approach ensures consistent rhythmic transmission across all body parts, enhancing coherence base motion. Moreover, we propose a local-global rhythmic consistency learning approach, which enforces alignment at both the frame and sequence levels. Locally, a frame-level consistency loss ensures that each frame is precisely synchronized with its corresponding speech features, guaranteeing accurate temporal alignment. Globally, a sequence-level consistency loss sustains a coherent rhythmic flow across the entire motion sequence, preserving consistency throughout the generated gestures. + +Furthermore, we introduce semantic emphasis learning approach, which focuses on generating semantic-aware sparse motion. This approach utilizes frame-level semantic cues from textual information, high-level speech features, and emotion to identify frames that require emphasis through a learned semantic score produced by a gating strategy, i.e., sem-gate. The sem-gate is designed to dynamically activate semantic motions at key frames through two weighting methods applied on the motion condition and the loss, respectively, and semantic label guidance, allowing the model to produce motion that enhances the motion with deeper semantic meaning and contextual relevance. + +Finally, the base motion and sparse motion are integrated through semantic score-based motion fusion, which adaptively amplifies expressiveness by incorporating semantic- + +aware key frames into the rhythm-related base motion. Our contributions are summarized below: + +- We propose SemTalk, a novel framework for holistic co-speech motion generation that separately models rhythm-related base motion and semantic-aware sparse motion, adaptively integrating them via a learned semantic gate. +- We propose a hierarchical coarse2fine cross-attention module to refine base motion and a local-global rhythmic consistency learning to integrate latent face and hand features with rhythm-related priors, ensuring coherence and rhythmic consistency. We then propose semantic emphasis learning to generate semantic gestures at certain frames, enhancing semantic-aware sparse motion. +- Experimental results show that our model surpasses state-of-the-art methods qualitatively and quantitatively, achieving higher motion quality and richer semantics. + +# 2. Related Work + +Co-speech Gesture Generation. Co-speech gesture generation aims to produce gestures aligned with speech. Early rule-based methods [7, 19, 21, 22, 41] lacked variability, while deterministic models [5, 7, 29, 36, 46, 49] mapped speech directly to gestures. Probabilistic models, including GANs [1, 17, 40] and diffusion models [2, 10, 47, 54], introduced variability. Some methods incorporated semantic cues, such as HA2G [32] and SEEG [28], which used hierarchical networks and alignment techniques. SynTalk [8] employs prompt-based control but treats inputs as signal strengths rather than fully interpreting semantics. LivelySpeaker [53] combines rhythmic features and semantic cues using CLIP [39] but struggles to integrate gestures with rhythm and capture semantics consistently, moreover, it only provides global control, limiting fine-grained refinement. DisCo [29] disentangles content and rhythm but lacks explicit modeling of sparse semantic gestures. SemTalk addresses this by separately modeling rhythm-related base motion and semantic-aware sparse motion, integrating them adaptively through a learned semantic score. + +Holistic Co-speech Motion Generation. Generating synchronized, expressive full-body motion from speech remains challenging, especially in coordinating the face, hands, and torso [9, 31, 34, 37, 48, 52]. Early methods introduced generative models to improve synchronization, but issues persisted. TalkSHOW [48] improved with VQ-VAE [42] cross-conditioning but handled facial expressions separately, causing fragmented outputs. DiffSHEG [9] and EMAGE [31] used separate encoders for expressions and gestures, but their unidirectional flow limited coherence. ProbTalk [33] leverages PQ-VAE [43] for improved body-facial synchronization but mainly relies on rhythmic cues, risking the loss of nuanced semantic gestures. Inspired by TM2D [15], which decomposes dance motion into music-related components, we separately model co-speech motion + +into rhythm-related and semantic-aware motion. + +# 3. Method + +# 3.1. Preliminary on RVQ-VAE + +Following [4, 16, 51], our approach uses a residual vector-quantized autoencoder (RVQ-VAE) to progressively capture complex body movements in a few players. To retain unique motion characteristics across body regions, we segment the body into four parts—face, upper body, hands, and lower body—each with a dedicated RVQ-VAE, following [3, 31]. This segmentation preserves each part's dynamics and prevents feature entanglement. + +# 3.2. Overview + +As shown in Figure 2, our SemTalk pipeline includes two main components: the Base Motion Blocks $f_{r}(\cdot)$ and the Sparse Motion Blocks $f_{b}(\cdot)$ . Given rhythmic features $\gamma_{b}, \gamma_{h}$ , a seed pose $\tilde{m}$ , and a speaker ID $id$ , the Base Motion Blocks generate rhythm-aligned codes $q^{b}$ , forming the rhythmic foundation of the base motion: + +$$ +f _ {r}: \left(\gamma_ {b}, \gamma_ {h}, \tilde {m}, i d; \theta_ {f _ {r}}\right)\rightarrow q ^ {b}, \tag {1} +$$ + +where $\theta_{f_r}$ denotes the learnable parameters of the Base Motion Blocks. The Sparse Motion Blocks then take semantic features $\phi_l$ , $\phi_g$ , $\phi_e$ , along with $\gamma_h$ , $\tilde{m}$ and $id$ , to produce frame-level semantic codes $q^s$ and semantic score $\psi$ . $\psi$ then triggers these codes only for semantically significant frames, producing a sparse motion representation: + +$$ +f _ {s}: \left(\phi_ {l}, \phi_ {g}, \phi_ {e}, \tilde {m}, i d; \theta_ {f _ {a}}\right)\rightarrow \left(q ^ {s}, \psi\right), \tag {2} +$$ + +where $\theta_{f_s}$ represents the Sparse Motion Block parameters. Finally, the semantic emphasis mechanism $\mathcal{E}$ combines $q^b$ and $q^s$ , guided by $\psi$ , to form the final motion codes $q^m$ : + +$$ +q ^ {m} = \mathcal {E} \left(q ^ {b}, q ^ {s}; \psi\right). \tag {3} +$$ + +The motion decoder then uses $q^{m}$ to generate the output $m'$ . + +# 3.3. Generating Rhythm-related Base Motion + +The Base Motion Generation (Fig. 3 a) in SemTalk establishes a rhythmically aligned foundation by leveraging both rhythmic and speaker-specific features, enhancing the naturalness and personalization of generated motion. + +Rhythmic Speech Encoding. To synchronize motion with speech, SemTalk incorporates rhythmic features: beats $\gamma_{b}$ and HuBERT features $\gamma_{h}$ . $\gamma_{b}$ , derived from amplitude, short-time energy [11], and onset detection, mark key rhythmic points for aligning gestures with speech. Meanwhile, $\gamma_{h}$ , extracted by the HuBERT encoder [18], captures high-level audio traits. In addition to rhythmic features $\gamma$ , + +![](images/1299c229de3e4d85275790079a0d9224f4c1d716d31f7f6960ad7aecdf6289bd.jpg) +Figure 2. An overview of the SemTalk pipeline. SemTalk generates holistic co-speech motion by first constructing rhythm-aligned $q^r$ in $f_r$ , guided by rhythmic consistency loss $L_{\mathrm{Rhy}}$ . Meanwhile, $f_s$ produce frame-level semantic codes $q^s$ , activated selectively by the semantic score $\psi$ . Finally, $q^m$ is achieved by fusing $q^r$ and $q^s$ based on $\psi$ , with motion decoder, yielding synchronized and contextually enriched motions. + +SemTalk uses a seed pose $\tilde{m}$ and speaker identity $id$ to generate a personalized, rhythm-aligned latent pose $p$ . Then MLP-based Face Enhancement and Body Part-Aware modules utilize $\gamma$ , $p$ and $id$ to obtain latent face $f_{e}$ , hands $f_{h}$ , upper body $f_{u}$ and lower body $f_{l}$ . + +Coarse2Fine Cross-Attention Module. To facilitate the learning of base motion, we first proposed a transformer-based hierarchical Coarse2Fine Cross-Attn Module utilize $f_{e}$ , $f_{h}$ , $f_{u}$ and $f_{l}$ to obtain latent base motion $f_{b}$ . The refinement begins with $\gamma$ for $f_{e}$ , which guides the rhythmic representation for $f_{h}$ , followed by conditioning $f_{u}$ and finally influencing $f_{l}$ . Since mouth movements closely correspond to speech syllables with minimal delay, we use the face to guide hand motions, inspired by DiffSHEG [9]. As the upper and lower body movements are less directly driven by speech and instead reflect the natural swinging of the hands and torso, we adopt cascading guidance: hands influence the upper body, which in turn drives the lower body. This structured approach, moving from the face to the hands, upper body, and lower body, ensures smooth and coherent motion propagation across the entire body. + +Rhythmic Consistency Learning. Inspired by CoG's use of InfoNCE loss [45] to synchronize facial expressions with audio cues, our approach adopts a similar philosophy of + +![](images/144d417bbc31eefff39cb9d8e50d3900e837a9275475db832198428d94277b60.jpg) +Figure 3. Architecture of SemTalk. SemTalk generates holistic co-speech motion in three stages. (a) Base Motion Generation uses rhythmic consistency learning to produce rhythm-aligned codes $q^b$ , conditioned on rhythmic features $\gamma_b, \gamma_h$ . (b) Sparse Motion Generation employs semantic emphasis learning to generate semantic codes $q^s$ , activated by semantic score $\psi$ . (c) Adaptively Fusion automatically combines $q^b$ and $q^s$ based on $\psi$ to produce mixed codes $q^m$ at frame level for rhythmically aligned and contextually rich motions. + +aligning motion and speech rhythm. It can be defined as: + +$$ +\mathcal {L} _ {\mathrm {R h y}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\operatorname {s i m} \left(h \left(f _ {i}\right) , \gamma_ {h} ^ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {N} \exp \left(\operatorname {s i m} \left(h \left(f _ {i}\right) , \gamma_ {h} ^ {j}\right) / \tau\right)}, \tag {4} +$$ + +where $N$ denotes the number of frames(or the batch size), $\tau$ denotes the temperature hyperparameter, $h(\cdot)$ is the projection head for latent motion, $f_{i}$ and $\gamma_h^i$ are the latent motion and rhythmic features at frame (or sample) $i$ , and $\mathrm{sim}(\cdot)$ represents cosine similarity. + +Unlike CoG, our approach fundamentally differs by incorporating separate local and global rhythmic consistency losses, which are applied to both latent face $f_{e}$ and latent hands $f_{h}$ , ensuring a more cohesive and synchronized representation across the entire motion sequence. This rhythmic consistency loss ensures that the motions are not only synchronized at the frame level but also maintain a consistent rhythmic flow across the entire sequence. + +The local frame-level consistency loss $\mathcal{L}_{\mathrm{Rhy}}^{(L)}$ aligns the motion features of each frame with the corresponding rhythmic cues $\gamma_{h}$ . By leveraging HuBERT features $\gamma_{h}$ instead of basic beat features $\gamma_{b}$ , which only capture rhythmic pauses, we incorporate rich, high-level audio representations that enhance the model's ability to capture rhythm-related motion patterns and maintain temporal coherence. + +The global sentence-level consistency loss $\mathcal{L}_{\mathrm{Rhy}}^{(G)}$ is designed to ensure rhythmic coherence at a global level. Unlike local loss, $\mathcal{L}_{\mathrm{Rhy}}^{(G)}$ reinforces rhythm consistency throughout the sequence, ensuring that the generated motion maintains smooth and rhythm-aligned throughout its duration. + +By jointly minimizing $\mathcal{L}_{\mathrm{Rhy}}^{(L)}$ and $\mathcal{L}_{\mathrm{Rhy}}^{(G)}$ , rhythmic consistency learning enables SemTalk to produce base motions that are rhythmically aligned and temporally cohesive, + +forming a solid rhythm-related base motion foundation. + +# 3.4. Generating Semantic-aware Sparse Motion + +The Sparse Motion Generation (Fig. 3 b) in SemTalk adds semantic-aware sparse motion to base motion by incorporating semantic cues drawn from speech content and emotional tone. By separating rhythm and semantics, this stage enhances motion generation by emphasizing contextually meaningful motion at key semantic moments. + +Semantic Speech Encoding. To capture semantic cues in speech, similar to [10], Semantic Emphasis Learning combines frame-level text embeddings $\phi_t$ , sentence-level features $\phi_g$ from the CLIP model [39], and emotion features $\phi_e$ from the emotion2vec model [35]. These features form a comprehensive semantic representation $f_t$ , together with audio feature $\gamma_h$ , that reflects both the content and emotional undertones of speech, enabling SemTalk to activate motions that are sensitive to nuanced semantic cues. + +Semantic Emphasis Learning. The process begins by generating $f_{t}$ , combining local and global cues from text, speech, emotion embeddings and HuBERT features $\gamma_{h}$ . Then, the sem-gate leverages multi-modal inputs to generate a semantic score, identifying frames that require enhanced semantic emphasis. The sem-gate in SemTalk refines keyframe motion by applying two forms of weighting methods $\mathcal{W}$ : feature weighting $\mathcal{W}_{f}$ and loss weighting $\mathcal{W}_{l}$ . Using $f_{t}$ and $\gamma_{h}$ , SemTalk computes a semantic score $\psi$ which dynamically scales feature weighting—filtering back semantic features $f_{t}$ to activate frames with significant relevance, ensuring that the model emphasizes frames aligned with specific communicative intentions. Second, the loss weighting is applied by supervising $\psi$ , with a classification loss $\mathcal{L}_{cls}^{G}$ based on semantic labels, further enhancing the model's ability to identify key frames. The two weight- + +![](images/c2aefad23af9a6868a75803373bfe37fae4fe3203eb7f4810dda18663f9b900a.jpg) +Figure 4. Concept comparison with LivelySpeaker [53]. (Top) LivelySpeaker generates semantic gestures with CLIP embeddings in SAG and refines rhythm-related gestures separately using diffusion, causing potential jitter. (Bottom) SemTalk integrates text and speech, uses a semantic gate for fine-grained control, and unifies rhythm and semantics for smoother, more coherent motions. + +ing methods allow SemTalk to selectively enhance semantic gestures while suppressing uninformative motion, leading to more expressive co-speech motion. + +Once $\psi$ is established, it modulates the integration of rhythm-aligned base motion $f_{b}$ and sparse semantic motion $f_{s}$ . Through alpha-blending, frames with high semantic relevance draw more from $f_{s}$ , while others rely on $f_{b}$ . The final motion codes $q^{s}$ are computed as: + +$$ +q ^ {s} = M L P \left(\psi f _ {s} + (1 - \psi) f _ {b}\right), \tag {5} +$$ + +To ensure cohesive propagation of semantic emphasis across body regions, we employ the Coarse2Fine Cross-Attention Module, similar to Sec. 3.3. In this stage, we focuses solely on body motion, excluding facial movements, as body gestures play a more critical role in conveying semantic meaning in co-speech interactions. + +To foster diverse motion generation, SemTalk includes a code classification loss $\mathcal{L}_{cls}$ and a reconstruction loss $\mathcal{L}_{rec}$ . These losses are specifically focused on frames with high semantic scores, guiding the model to prioritize the generation of sparse, meaningful gestures. + +Discussion. Recently, LivelySpeaker [53] designs the Semantic-Aware Generator (SAG) and Rhythm-Aware Generator (RAG) for co-speech gesture generation, combining them through beat empowerment. While effective, key differences exist between LivelySpeaker and SemTalk, see Fig. 4. First, SAG generates gestures from text using CLIP embeddings, but bridging words and expressive gestures is challenging, causing jitter. SemTalk incorporates speech features (pitch, tone, emotion) alongside text and GT supervision for adaptive gestures. Second, LivelySpeaker applies global control, missing local semantic details, while SemTalk uses fine-grained, frame-level semantic control for subtle variations. Third, LivelySpeaker fuses SAG and RAG in separate latent spaces, leading to misalignment and inconsistencies. SemTalk jointly models rhythm and semantics in a unified framework, ensuring smoother transitions and coherence. We further compare SAG with our + +semantic gate in experiments. + +# 3.5. Semantic Score-based motion fusion + +The Adaptive Fusion stage (Fig. 3 c) in SemTalk seamlessly integrates semantic-aware sparse motion into the rhythmic-related base motion. By strategically enhancing frames based on their semantic importance, it maintains a smooth and natural motion flow across sequences. For each frame $i$ , the semantic score $\psi_i$ computed during the Sparse Motion Generation stage is compared to a threshold $\beta$ . If $\psi_i > \beta$ , the base motion's latent code $q_i^r$ is replaced with the sparse semantic code $q_i^s$ , effectively highlighting expressive gestures where they are most relevant; otherwise, $q_i = q_i^r$ . + +This selective replacement emphasizes semantically critical gestures while preserving the natural rhythmic base motion. By blending $q^{b}$ and $q^{s}$ based on semantic scores, SemTalk adapts to the expressive needs of the speech context while ensuring coherence. Additionally, the convolution structure of the RVQ-VAE decoder ensures smooth transitions between frames, preserving motion continuity. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. For training and evaluation, we use two datasets: BEAT2 and SHOW. BEAT2, introduced in EMAGE [31], extends BEAT [30] with 76 hours of data from 30 speakers, standardized into a mesh representation with paired audio, text, and frame-level semantic labels. We follow [31] and use the BEAT2-standard subset with an $85\% / 7.5\% / 7.5\%$ train/val/test split. SHOW [48] includes 26.9 hours of high-quality talk show videos with 3D body meshes at 30fps. Since it lacks frame-level semantic labels, we use the semgate from SemTalk, pre-trained on BEAT2, to generate them. Following [48], we select video clips longer than 10 seconds and split the data $80\% / 10\% / 10\%$ for train/val/test. + +Implementation Details. Our model is trained on a single NVIDIA A100 GPU for 200 epochs with a batch size of 64. We use RVQ-VAE [42], downscaling by 4. The residual quantization has 6 layers, a codebook size of 256 and a dropout rate of 0.2. We use five transformer layers to predict the last five layer codes. In Base Motion Learning, $\tau = 0.1$ ; in Sparse Motion Learning, $\beta = 0.5$ empirically. The training uses ADAM with a 1e-4 learning rate. Following [31], we start with a 4-frame seed pose, gradually increasing masked frames from 0 to $40\%$ over 120 epochs. + +Metrics. We evaluate generated body gestures using FGD [50] to measure distributional alignment with GT, reflecting realism. DIV [26] quantifies gesture variation via the average L1 distance across clips. BC [27] assesses speech-motion synchrony. For facial expressions, we use MSE [47] to quantify positional differences and LVD [48] to measure discrepancies between GT and generated facial vertices. + +![](images/ecfa083670d714d79dd4c95afe184dc95c93b4c4df5646584ffc0173e277becc.jpg) +Figure 5. Comparison on BEAT2 [31] Dataset. SemTalk* refers to the model trained solely on the Base Motion Generation stage, capturing rhythmic alignment but lacking semantic gestures. In contrast, SemTalk successfully emphasized sparse yet vivid motions. For instance, when saying "my opinion," SemTalk generates a hand-raising gesture followed by an index finger extension for emphasis. Similarly, for "never tell," our model produces a clear, repeated gesture matching the rhythm, reinforcing the intended emphasis. + +![](images/f09db62adf066365352e12b4bbf519f8db3b6c167df8738489b0a3ac7649784f.jpg) +Figure 6. Comparison on SHOW [48] Dataset. Our method performs better in motion diversity and semantic richness. + +![](images/d15e04660b009631faa3ac6e9e6bff96583a48bd97717a24d032bc6b4078ce89.jpg) + +# 4.2. Qualitative Results + +Qualitative Comparisons. We encourage readers to watch our demo video for a clearer understanding of SemTalk's qualitative performance. Our method achieves superior speech-motion alignment, generating more realistic, diverse, and semantically consistent gestures than the baselines. As shown in Fig. 5, LivelySpeaker, TalkSHOW, EMAGE, and DiffSHEG exhibit jitter—EMAGE mainly in the legs and shoulders, while TalkSHOW affects the entire body. LivelySpeaker and DiffSHEG, which focus primarily on the upper body, produce slow and inconsistent motions, especially at speech clip boundaries. DiffSHEG improves + +gesture diversity over IMAGE and TalkSHOW, though IMAGE maintains greater naturalness. SemTalk surpasses all baselines in both realism and diversity. Compared to SemTalk*, SemTalk generates more expressive gestures, emphasizing key phrases (e.g., raising hands for “dream job” or pointing for “that is why”). While SemTalk* ensures rhythmic consistency, it lacks semantic expressiveness. By integrating frame-level semantic emphasis, SemTalk aligns motion with both rhythm and semantics, demonstrating the effectiveness of rhythmic consistency learning and semantic emphasis learning. In facial comparisons (Fig. 7), IMAGE shows minimal lip movement, while both DiffSHEG and + +![](images/a717cf177762c731d3be4305e5fb997193b304e071da862c3175abdaaafa8d47.jpg) +Figure 7. Facial Comparison on the BEAT2 [31] Dataset. + +![](images/a501893e2c1a842880f342ab42f131c7d71c2b26442d82b3b6605150e683eb60.jpg) +Figure 8. Qualitative study on semantic score. Semantic score aligns with keywords, influencing gesture intensity. + +EMAGE reveal inconsistencies between lip motion and the rhythm of speech. In contrast, SemTalk produces smooth, natural transitions across syllables, resulting in realistic and expressive lips, significantly surpassing the baselines. + +On the SHOW dataset (Fig. 6), SemTalk shows more agile gestures than all baselines, when applied to unseen data. Our method captures natural and contextually rich gestures, particularly in moments of emphasis such as "I like to do" and "relaxing," where our model produces lively hand and body movements that align with the speech content. + +Semantic Score. Fig. 8 shows how semantic emphasis influences gesture intensity, with peaks in the semantic score aligning with keywords like "comes," "fantastic," and "captured." By extracting semantic scores from key frames, we track gesture emphasis trends. Furthermore, as shown in Fig. 9, SemTalk adapts to different emotional tones even when the text remains unchanged. This adaptability prevents overfitting to the text itself, allowing the model to generate gestures that vary according to the emotional delivery of the speech. The learned semantic score provides fine-grained, frame-level control, keeping gestures both rhythmically synchronized and semantically aligned in real time. User Study. We conducted a user study with 10 video samples and 25 participants from diverse backgrounds, evaluating realism, semantic consistency, motion-speech synchrony, and diversity. Participants were required to rank shuffled videos across different methods. As shown in Fig. 10, our approach received dominant preferences across all metrics, especially in semantic consistency and realism. + +![](images/760d1fd802be9416574db3e923376768c8d359df700277ab247316886991aed9.jpg) +Figure 9. Same words with different speech from the internet. "emo" represents different emotional tones extracted from speech. SemTalk can generate different motions, even when the text script is the same, preventing overfitting to the text itself. + +![](images/d1f5db3de63a6fab1a877bb143eb01bf4d2b4cb70944e878ac8e7d740c826275.jpg) +Figure 10. Results of the user study. + +# 4.3.Quantitative Results + +Comparison with Baselines. As shown in Tab. 1, SemTalk outperforms previous methods on BEAT2, achieving lower FGD, MSE, and LVD, indicating better distribution alignment and reduced motion errors. For fairness, we follow [31] and add a lower-body VQ-VAE to TalkSHOW, Diff-SHEG, and SemTalk. Notably, SemTalk significantly reduces FGD, ensuring strong distribution matching. While TalkSHOW and EMAGE achieve competitive diversity (DIV) scores, SemTalk balances high semantic relevance with natural motion flow. + +On the SHOW dataset, SemTalk excels with the lowest FGD, MSE, and the highest BC, indicating precise beat alignment with the audio and enhanced semantic consistency in generated motions. Although EMAGE exhibits high DIV, our model achieves comparable results while maintaining smooth, realistic motion free from jitter. + +Sem-gate. Tab. 2 highlights the effectiveness of sem-gate. Without sem-gate, the model fails to emphasize key moments. Randomized semantic scores led to poor performance by preventing meaningful frame distinction. Introducing a learned sem-gate even (w/o $\mathcal{W}$ ) significantly improves semantic alignment and classification accuracy. Refinement is further enhanced through weighting strategies: feature weighting $\mathcal{W}_f$ enhances motion emphasis, while loss weighting $\mathcal{W}_l$ improves FGD and overall accuracy. These results suggest that weighting methods enhance the accuracy of the semantic score and help the model prioritize important frames. The best results come from applying two weighting methods together, where frames with stronger se + +
DatasetMethodFGD↓BC↑DIV↑MSE↓LVD↓
BEAT2FaceFormer [13]---7.7877.593
CodeTalker [44]---8.0267.766
CaMN [30]6.6446.76910.86--
DSG [47]8.8117.24111.49--
LivelySpeaker [17]11.806.65911.28--
Habibie et al. [17]9.0407.7168.2138.6148.043
TalkSHOW [48]6.2096.94713.477.7917.771
EMAGE [31]5.5127.72413.067.6807.556
DiffSHEG [9]8.9867.14211.917.6658.673
SemTalk (Ours)4.2787.77012.916.1536.938
SHOWFaceFormer [13]---138.143.69
CodeTalker [44]---140.745.84
CaMN [30]22.127.71210.37--
DSG [47]24.848.02710.23--
LivelySpeaker [17]32.177.84410.14--
Habibie et al. [17]27.228.2098.541145.647.35
TalkSHOW [48]24.438.24910.98139.645.17
EMAGE [31]22.128.28012.46136.142.44
DiffSHEG [9]24.878.06110.79139.045.77
SemTalk (Ours)20.188.30411.36134.139.15
+ +mantic signals receive higher emphasis. We also compare sem-gate with LivelySpeaker's SAG [53]. We find that replacing the Sparse Motion stage with SAG and substituting motion using GT semantic labels led to poor performance. SAG relies only on text-motion alignment, ignoring emotional tone, making it more prone to overfitting the text. In contrast, our sem-gate applies GT supervision with two weighting methods, achieving more accurate and stable semantic motion. + +Ablation Study on Components. We assess the impact of each component of our model on BEAT2 and present the results in Tab. 3, which reveals several key insights (more ablation results please see supplementary material): + +- Rhythmic Consistency Learning (RC) not only boosts performance on key metrics like FGD, LVD, and BC but also reduces the MSE, contributing to smoother and more realistic base motion. +- Semantic Emphasis Learning (SE) proves essential for selectively enhancing semantic-rich gestures. The inclusion of SE, as shown in rows with SE enabled, improves both diversity (DIV) and FGD, enabling the model to emphasize semantically relevant motions. SE demonstrates its effectiveness in focusing on frame-level semantic information, which contributes to the generation of lifelike gestures with enriched contextual meaning. +- Coarse2Fine Cross-Attention Module (C2F) effectively refines motion details, improving BC, FGD, and DIV. When combined with RVQ and RC, C2F achieves the best MSE and LVD, highlighting its role in enhancing motion realism and diversity hierarchically. + +Table 1. Quantitative comparison with SOTA. SemTalk consistently outperforms baselines across both the BEAT2 and SHOW datasets. Lower values are better for FMD, FGD, MSE, and LVD. Higher values are better for BC and DIV. We report $\mathrm{FGD} \times 10^{-1}$ , $\mathrm{BC} \times 10^{-1}$ , $\mathrm{MSE} \times 10^{-8}$ and $\mathrm{LVD} \times 10^{-5}$ for simplify. + +
MethodFGD↓BC↑DIV↑Acc(%)↑
w/o Sem-gate4.8937.70212.42-
SAG (LivelySpeaker [53])4.6187.68212.45-
Sem-gate (Random ψ)4.6347.70012.4450.07
Sem-gate (w/o W)4.4957.63312.2672.32
Sem-gate (w/ Wf)4.4087.67912.2878.52
Sem-gate (w/ Wl)4.3667.77211.9477.83
Sem-gate (ours)4.2787.77012.9182.76
+ +Table 2. Ablation study on Sem-gate. "Acc" denotes semantic classification performance on BEAT2. "w/o Sem-gate" means directly input $f_{t}$ and $\gamma_{h}$ without Sem-gate. "SAG (LivelySpeaker [53])" replaces the Sparse Motion Generation stage with LivelySpeaker's SAG method. "Random $\psi$ " assigns frame-level scores randomly. "w/o $\mathcal{W}$ " applies the semantic gate but excludes frame-level weighting. "w/ $\mathcal{W}_{f}$ " applies feature weighting. "w/ $\mathcal{W}_{l}$ " applies loss weighting. (as mentioned in Sec. 3.4). Sem-gate (ours) integrates both the semantic gate and frame-level weighting to enhance emphasis. + +
RCSEC2FRVQFGD↓BC↑DIV↑MSE↓LVD↓
----6.2347.62811.448.2397.831
---√5.4847.64111.8413.88215.42
√--√4.8677.70112.386.2016.928
√√-√4.5267.75112.836.2156.997
--√√4.8977.70212.4213.41615.72
√√--5.8317.75811.976.5877.106
√-√√4.3977.77612.496.1006.898
√√√√4.2787.77012.916.1536.938
+ +Table 3. Ablation study on each key component. "RC" denotes rhythmic consistency learning, "SE" denotes the semantic emphasis learning, and "C2F" denotes Coarse2Fine Cross-Att Module, "RVQ" denotes the RVQ-VAE. + +- RVQ-VAE (RVQ) enhances the diversity and realism of generated motion. Though it slightly increases MSE and LVD, it notably improves FGD, leading to more natural motion generation compared to standard VQ-VAE. + +# 5. Conclusion + +We propose SemTalk, a novel approach for holistic cospeech motion generation with frame-level semantic emphasis. Our method addresses the integration of sparse yet expressive motion into foundational rhythm-related motion, which has received less attention in previous works. We develop a framework that separately learns rhythm-related base motion through coarse2fine cross-attention module and rhythmic consistency learning, while capturing semantic-aware motion through Semantic Emphasis Learning. These components are then adaptively fused based on a learned semantic score. Our approach has demonstrated state-of-the-art performance on two public datasets quantitatively and qualitatively. The qualitative results and user study show that our method can generate high-quality cospeech motion sequences that enhance frame-level semantics over robust base motions, reflecting the full spectrum of human expressiveness. + +Acknowledgments. This work was supported by Alibaba Research Intern Program, the Young Scientists Fund of the National Natural Science Foundation of China No. 624B2110, the National Key Research and Development Program of China No. 2024YFC3015600, the Fundamental Research Funds for Central Universities No.2042023KF0180 & No.2042025KF0053. The numerical calculation is supported by supercomputing system in Super-computing Center of Wuhan University and Tongyi Lab, Alibaba Group. + +# References + +[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566–20576, 2022. 2 +[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2 +[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 3 +[4] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 31:2523-2533, 2023. 3 +[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2 +[6] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1 +[7] Justine Cassell, Hannes Högni Vilhjalmsson, and Timothy Bickmore. Beat: the behavior expression animation toolkit. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 477-486, 2001. 2 +[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2 +[9] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7352-7361, 2024. 2, 3, 8 + +[10] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al. Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 2, 4 +[11] Selina Chu, Shrikanth Narayanan, and C-C Jay Kuo. Environmental sound recognition with time-frequency audio features. IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142-1158, 2009. 3 +[12] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 2 +[13] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 8 +[14] Susan Goldin-Meadow. The role of gesture in communication and thinking. Trends in cognitive sciences, 3(11):419-429, 1999. 1 +[15] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942–9952, 2023. 2 +[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 3 +[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 1, 2, 8 +[18] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 3 +[19] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In Proceedings of the seventh annual ACM/IEEE international conference on Human-Robot Interaction, pages 25-32, 2012. 2 +[20] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1 +[21] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2 +[22] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thorisson, and Hannes Vilhjalmsson. Towards a common + +framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2 +[23] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2 +[24] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In Proceedings of the 26th International Conference on Intelligent User Interfaces, pages 11-21, 2021. 1 +[25] Alex Lascarides and Matthew Stone. A formal semantic analysis of gesture. Journal of Semantics, 26(4):393-449, 2009. 1 +[26] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 5 +[27] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 5 +[28] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2 +[29] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM international conference on multimedia, pages 3764-3773, 2022. 2 +[30] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European conference on computer vision, pages 612-630. Springer, 2022. 1, 5, 8 +[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 1, 2, 3, 5, 6, 7, 8 +[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei + +Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2 +[33] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 2 +[34] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2 +[35] Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023. 4 +[36] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics symposium on computer animation, pages 25-35, 2013. 2 +[37] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 2 +[38] Ashi Özyürek, Roel M Willems, Sotaro Kita, and Peter Hagoort. On-line integration of semantic information from speech and gesture: Insights from event-related brain potentials. Journal of cognitive neuroscience, 19(4):605-616, 2007. 1 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4 +[40] Manuel Rebol, Christian Güttl, and Krzysztof Pietroszek. Real-time gesture animation generation from speech for virtual human interaction. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-4, 2021. 2 +[41] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized talking head synthesis. arXiv preprint arXiv:2301.03786, 2(4):5, 2023. 2 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5 +[43] Hanwei Wu and Markus Flierl. Learning product codebooks using vector-quantized autoencoders for image retrieval. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5. IEEE, 2019. 2 +[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven + +3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 8 +[45] Zunnan Xu, Yachao Zhang, Sicheng Yang, Ronghui Li, and Xiu Li. Chain of generation: Multi-modal gesture synthesis via cascaded conditional control. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6387-6395, 2024. 3 +[46] Sicheng Yang, Zilin Wang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Qiaochu Huang, Lei Hao, Songcen Xu, Xiaofei Wu, Changpeng Yang, et al. Unified gesture: A unified gesture synthesis model for multiple skeletons. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1033-1044, 2023. 2 +[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 5, 8 +[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 1, 2, 5, 6, 8 +[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2 +[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 5 +[51] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 3 +[52] Jinsong Zhang, Minjie Zhu, Yuxiang Zhang, Zerong Zheng, Yebin Liu, and Kun Li. Speechact: Towards generating whole-body motion from speech. IEEE Transactions on Visualization and Computer Graphics, 2025. 2 +[53] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 5, 8 +[54] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2 \ No newline at end of file diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/images.zip b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08a33e8393176428aa720bcffc9685d28b7fa0d1 --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8d1c5d025316f3c328805c7149160f84bbe6b4715bbe4d421ad983f71d7dbcd +size 694650 diff --git a/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/layout.json b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..32fae9626f7ee41c177c173456099ac1171c2d14 --- /dev/null +++ b/2025/SemTalk_ Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis/layout.json @@ -0,0 +1,9230 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 102, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 102, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 510, + 140 + ], + "type": "text", + "content": "SemTalk: Holistic Co-speech Motion Generation with Frame-level Semantic Emphasis" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 159, + 536, + 219 + ], + "blocks": [ + { + "bbox": [ + 71, + 159, + 536, + 219 + ], + "lines": [ + { + "bbox": [ + 71, + 159, + 536, + 219 + ], + "spans": [ + { + "bbox": [ + 71, + 159, + 536, + 219 + ], + "type": "image", + "image_path": "71477c7f42ee897a304a1baaff56e265bfc8ff144b3bf0b898e4065bbbead50f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 224, + 227, + 358 + ], + "blocks": [ + { + "bbox": [ + 70, + 224, + 227, + 358 + ], + "lines": [ + { + "bbox": [ + 70, + 224, + 227, + 358 + ], + "spans": [ + { + "bbox": [ + 70, + 224, + 227, + 358 + ], + "type": "image", + "image_path": "49b2f3acde280b07ad3698249e8c8c9186570f4d3f6fe253805a2336977286cd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 363, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 54, + 363, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 555, + 420 + ], + "type": "text", + "content": "Figure 1. On the left, we analyze semantic labels from the BEAT2 dataset [31] and visualize frame-level motion, revealing that semantically relevant motions are rare and sparse, aligning with real-life observations. On the right, this observation drives the design of SemTalk, which establishes a rhythm-aligned base motion and dynamically emphasizes sparse semantic gestures at the frame-level. In this example, SemTalk amplifies expressiveness on words like \"watching\" and \"just,\" enhancing gesture and torso movements. The semantic scores below are automatically generated by SemTalk to modulate semantic emphasis over time." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 224, + 542, + 358 + ], + "blocks": [ + { + "bbox": [ + 230, + 224, + 542, + 358 + ], + "lines": [ + { + "bbox": [ + 230, + 224, + 542, + 358 + ], + "spans": [ + { + "bbox": [ + 230, + 224, + 542, + 358 + ], + "type": "image", + "image_path": "0d782c86a891238469b65982dca38bade915fe2835c8fecb4deb855fcb404705.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 433, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 433, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 151, + 433, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 458, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 458, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 458, + 297, + 687 + ], + "type": "text", + "content": "A good co-speech motion generation cannot be achieved without a careful integration of common rhythmic motion and rare yet essential semantic motion. In this work, we propose SemTalk for holistic co-speech motion generation with frame-level semantic emphasis. Our key insight is to separately learn base motions and sparse motions, and then adaptively fuse them. In particular, coarse2fine cross-attention module and rhythmic consistency learning are explored to establish rhythm-related base motion, ensuring a coherent foundation that synchronizes gestures with the speech rhythm. Subsequently, semantic emphasis learning is designed to generate semantic-aware sparse motion, focusing on frame-level semantic cues. Finally, to integrate sparse motion into the base motion and generate semantic-emphasized co-speech gestures, we further leverage a learned semantic score for adaptive synthesis. Qualitative and quantitative comparisons on two public datasets demonstrate that our method outperforms the state-of-the-art, delivering high-quality co-speech motion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 434, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 555, + 445 + ], + "type": "text", + "content": "with enhanced semantic richness over a stable base motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 486, + 395, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 486, + 395, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 486, + 395, + 498 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 508, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 508, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 508, + 555, + 604 + ], + "type": "text", + "content": "Nonverbal communication, including body language, hand gestures, and facial expressions, is integral to human interactions. It enriches conversations with contextual cues and enhances understanding among participants [6, 14, 20, 24]. This aspect is particularly significant in holistic co-speech motion generation, where the challenge lies in synthesizing gestures that align with speech rhythm while also capturing the infrequent yet critical semantic gestures [25, 38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "Most existing methods [17, 30, 48] rely heavily on rhythm-related audio features as conditions for gesture generation. While these rhythm-based features successfully align gestures with the timing of speech, they often overshadow the sparse yet expressive semantic motion (see Fig. 1). As a result, the generated motions may lack the contextual depth necessary and nuanced expressiveness for natural interaction. Some methods try to address this by incorporating semantic information like emotion, style, and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13761" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "content[10, 12, 23, 32]. However, the rhythm features tend to dominate, making the models difficult to capture sparse, semantically relevant gestures at the frame level. These rare but impactful gestures are often diluted or overlooked, highlighting the challenge of balancing rhythmic alignment with semantic expressiveness in co-speech motion generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 277 + ], + "type": "text", + "content": "In real-world human conversations, we have an observation that while most speech-related gestures are indeed rhythm-related, only a limited number of frames involve semantically emphasized gestures. This insight suggests that co-speech motions can be decomposed into two distinct components: (i) Rhythm-related base motion. These provide a continuous, coherent base motion aligned with the speech rhythm, reflecting the natural timing of speaking. (ii) Semantic-aware sparse motion: These occur infrequently but are essential for conveying specific meanings or emphasizing key points within the conversation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 280, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 295, + 530 + ], + "type": "text", + "content": "Inspired by this observation, we propose a new framework SemTalk. SemTalk models the base motion and the sparse motion separately and then fuses them adaptively to generate high-fidelity co-speech motion. Specifically, we first focus on generating rhythm-related base motion by introducing coarse2fine cross-attention module and rhythmic consistency learning. We design a hierarchical coarse2fine cross-attention module, which progressively refines the base motion cues in a coarse-to-fine manner, starting from the face and moving through the hands, upper body, and lower body. This approach ensures consistent rhythmic transmission across all body parts, enhancing coherence base motion. Moreover, we propose a local-global rhythmic consistency learning approach, which enforces alignment at both the frame and sequence levels. Locally, a frame-level consistency loss ensures that each frame is precisely synchronized with its corresponding speech features, guaranteeing accurate temporal alignment. Globally, a sequence-level consistency loss sustains a coherent rhythmic flow across the entire motion sequence, preserving consistency throughout the generated gestures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "type": "text", + "content": "Furthermore, we introduce semantic emphasis learning approach, which focuses on generating semantic-aware sparse motion. This approach utilizes frame-level semantic cues from textual information, high-level speech features, and emotion to identify frames that require emphasis through a learned semantic score produced by a gating strategy, i.e., sem-gate. The sem-gate is designed to dynamically activate semantic motions at key frames through two weighting methods applied on the motion condition and the loss, respectively, and semantic label guidance, allowing the model to produce motion that enhances the motion with deeper semantic meaning and contextual relevance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Finally, the base motion and sparse motion are integrated through semantic score-based motion fusion, which adaptively amplifies expressiveness by incorporating semantic-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 534, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 534, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 534, + 95 + ], + "type": "text", + "content": "aware key frames into the rhythm-related base motion. Our contributions are summarized below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 96, + 553, + 263 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 96, + 553, + 144 + ], + "type": "text", + "content": "- We propose SemTalk, a novel framework for holistic co-speech motion generation that separately models rhythm-related base motion and semantic-aware sparse motion, adaptively integrating them via a learned semantic gate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 144, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 144, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 553, + 227 + ], + "type": "text", + "content": "- We propose a hierarchical coarse2fine cross-attention module to refine base motion and a local-global rhythmic consistency learning to integrate latent face and hand features with rhythm-related priors, ensuring coherence and rhythmic consistency. We then propose semantic emphasis learning to generate semantic gestures at certain frames, enhancing semantic-aware sparse motion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 228, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 228, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 553, + 263 + ], + "type": "text", + "content": "- Experimental results show that our model surpasses state-of-the-art methods qualitatively and quantitatively, achieving higher motion quality and richer semantics." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 295, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 533 + ], + "type": "text", + "content": "Co-speech Gesture Generation. Co-speech gesture generation aims to produce gestures aligned with speech. Early rule-based methods [7, 19, 21, 22, 41] lacked variability, while deterministic models [5, 7, 29, 36, 46, 49] mapped speech directly to gestures. Probabilistic models, including GANs [1, 17, 40] and diffusion models [2, 10, 47, 54], introduced variability. Some methods incorporated semantic cues, such as HA2G [32] and SEEG [28], which used hierarchical networks and alignment techniques. SynTalk [8] employs prompt-based control but treats inputs as signal strengths rather than fully interpreting semantics. LivelySpeaker [53] combines rhythmic features and semantic cues using CLIP [39] but struggles to integrate gestures with rhythm and capture semantics consistently, moreover, it only provides global control, limiting fine-grained refinement. DisCo [29] disentangles content and rhythm but lacks explicit modeling of sparse semantic gestures. SemTalk addresses this by separately modeling rhythm-related base motion and semantic-aware sparse motion, integrating them adaptively through a learned semantic score." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Holistic Co-speech Motion Generation. Generating synchronized, expressive full-body motion from speech remains challenging, especially in coordinating the face, hands, and torso [9, 31, 34, 37, 48, 52]. Early methods introduced generative models to improve synchronization, but issues persisted. TalkSHOW [48] improved with VQ-VAE [42] cross-conditioning but handled facial expressions separately, causing fragmented outputs. DiffSHEG [9] and EMAGE [31] used separate encoders for expressions and gestures, but their unidirectional flow limited coherence. ProbTalk [33] leverages PQ-VAE [43] for improved body-facial synchronization but mainly relies on rhythmic cues, risking the loss of nuanced semantic gestures. Inspired by TM2D [15], which decomposes dance motion into music-related components, we separately model co-speech motion" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13762" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 250, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 250, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 250, + 83 + ], + "type": "text", + "content": "into rhythm-related and semantic-aware motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 111, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 111, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 111, + 107 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 115, + 201, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 201, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 201, + 129 + ], + "type": "text", + "content": "3.1. Preliminary on RVQ-VAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 133, + 296, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 229 + ], + "type": "text", + "content": "Following [4, 16, 51], our approach uses a residual vector-quantized autoencoder (RVQ-VAE) to progressively capture complex body movements in a few players. To retain unique motion characteristics across body regions, we segment the body into four parts—face, upper body, hands, and lower body—each with a dedicated RVQ-VAE, following [3, 31]. This segmentation preserves each part's dynamics and prevents feature entanglement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 123, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 123, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 123, + 248 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": "As shown in Figure 2, our SemTalk pipeline includes two main components: the Base Motion Blocks " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "f_{r}(\\cdot)" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": " and the Sparse Motion Blocks " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "f_{b}(\\cdot)" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": ". Given rhythmic features " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\gamma_{b}, \\gamma_{h}" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": ", a seed pose " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": ", and a speaker ID " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "id" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": ", the Base Motion Blocks generate rhythm-aligned codes " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "q^{b}" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": ", forming the rhythmic foundation of the base motion:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 335, + 294, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 335, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 113, + 335, + 294, + 350 + ], + "type": "interline_equation", + "content": "f _ {r}: \\left(\\gamma_ {b}, \\gamma_ {h}, \\tilde {m}, i d; \\theta_ {f _ {r}}\\right)\\rightarrow q ^ {b}, \\tag {1}", + "image_path": "0a91918ccb6a96d2a03dfdc748d1ae6654a067c437eabdc2bea688e88acd3c74.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\theta_{f_r}" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " denotes the learnable parameters of the Base Motion Blocks. The Sparse Motion Blocks then take semantic features " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\phi_g" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\phi_e" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ", along with " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\gamma_h" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "id" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ", to produce frame-level semantic codes " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " and semantic score " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 359, + 296, + 431 + ], + "type": "text", + "content": " then triggers these codes only for semantically significant frames, producing a sparse motion representation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 441, + 294, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 441, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 96, + 441, + 294, + 454 + ], + "type": "interline_equation", + "content": "f _ {s}: \\left(\\phi_ {l}, \\phi_ {g}, \\phi_ {e}, \\tilde {m}, i d; \\theta_ {f _ {a}}\\right)\\rightarrow \\left(q ^ {s}, \\psi\\right), \\tag {2}", + "image_path": "c3c8cb010268c1e7558695f63a2ba6a20951a5324186d5044f4e9a47103cf47c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\theta_{f_s}" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": " represents the Sparse Motion Block parameters. Finally, the semantic emphasis mechanism " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": " combines " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "q^b" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": ", guided by " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": ", to form the final motion codes " + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "inline_equation", + "content": "q^m" + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 509, + 294, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 509, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 135, + 509, + 294, + 522 + ], + "type": "interline_equation", + "content": "q ^ {m} = \\mathcal {E} \\left(q ^ {b}, q ^ {s}; \\psi\\right). \\tag {3}", + "image_path": "ca1b861354d2f5561d99e67c6efb64e0d1e75edf31cdcada10c93f961a35e27a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "text", + "content": "The motion decoder then uses " + }, + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "inline_equation", + "content": "q^{m}" + }, + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "text", + "content": " to generate the output " + }, + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 55, + 532, + 295, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 563, + 270, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 270, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 270, + 577 + ], + "type": "text", + "content": "3.3. Generating Rhythm-related Base Motion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "type": "text", + "content": "The Base Motion Generation (Fig. 3 a) in SemTalk establishes a rhythmically aligned foundation by leveraging both rhythmic and speaker-specific features, enhancing the naturalness and personalization of generated motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Rhythmic Speech Encoding. To synchronize motion with speech, SemTalk incorporates rhythmic features: beats " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and HuBERT features " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", derived from amplitude, short-time energy [11], and onset detection, mark key rhythmic points for aligning gestures with speech. Meanwhile, " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", extracted by the HuBERT encoder [18], captures high-level audio traits. In addition to rhythmic features " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 543, + 312 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 543, + 312 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 543, + 312 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 543, + 312 + ], + "type": "image", + "image_path": "1299c229de3e4d85275790079a0d9224f4c1d716d31f7f6960ad7aecdf6289bd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": "Figure 2. An overview of the SemTalk pipeline. SemTalk generates holistic co-speech motion by first constructing rhythm-aligned " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "q^r" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "f_r" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": ", guided by rhythmic consistency loss " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{Rhy}}" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": " produce frame-level semantic codes " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": ", activated selectively by the semantic score " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "q^m" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": " is achieved by fusing " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "q^r" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 319, + 555, + 396 + ], + "type": "text", + "content": ", with motion decoder, yielding synchronized and contextually enriched motions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": "SemTalk uses a seed pose " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "\\tilde{m}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " and speaker identity " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "id" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " to generate a personalized, rhythm-aligned latent pose " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": ". Then MLP-based Face Enhancement and Body Part-Aware modules utilize " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "id" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " to obtain latent face " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": ", hands " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": ", upper body " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": " and lower body " + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 313, + 417, + 554, + 477 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": "Coarse2Fine Cross-Attention Module. To facilitate the learning of base motion, we first proposed a transformer-based hierarchical Coarse2Fine Cross-Attn Module utilize " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": " to obtain latent base motion " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ". The refinement begins with " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ", which guides the rhythmic representation for " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ", followed by conditioning " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": " and finally influencing " + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 313, + 481, + 555, + 673 + ], + "type": "text", + "content": ". Since mouth movements closely correspond to speech syllables with minimal delay, we use the face to guide hand motions, inspired by DiffSHEG [9]. As the upper and lower body movements are less directly driven by speech and instead reflect the natural swinging of the hands and torso, we adopt cascading guidance: hands influence the upper body, which in turn drives the lower body. This structured approach, moving from the face to the hands, upper body, and lower body, ensures smooth and coherent motion propagation across the entire body." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Rhythmic Consistency Learning. Inspired by CoG's use of InfoNCE loss [45] to synchronize facial expressions with audio cues, our approach adopts a similar philosophy of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13763" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 526, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 526, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 526, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 526, + 242 + ], + "type": "image", + "image_path": "144d417bbc31eefff39cb9d8e50d3900e837a9275475db832198428d94277b60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. Architecture of SemTalk. SemTalk generates holistic co-speech motion in three stages. (a) Base Motion Generation uses rhythmic consistency learning to produce rhythm-aligned codes " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "q^b" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": ", conditioned on rhythmic features " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\gamma_b, \\gamma_h" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": ". (b) Sparse Motion Generation employs semantic emphasis learning to generate semantic codes " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": ", activated by semantic score " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": ". (c) Adaptively Fusion automatically combines " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "q^b" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "q^s" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": " to produce mixed codes " + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "q^m" + }, + { + "bbox": [ + 54, + 247, + 555, + 293 + ], + "type": "text", + "content": " at frame level for rhythmically aligned and contextually rich motions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 285, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 285, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 285, + 316 + ], + "type": "text", + "content": "aligning motion and speech rhythm. It can be defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 324, + 294, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 324, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 324, + 294, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R h y}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(h \\left(f _ {i}\\right) , \\gamma_ {h} ^ {j}\\right) / \\tau\\right)}, \\tag {4}", + "image_path": "0198b267bd742e8ce74b727d728882ab8c3e2e534a79de7af0fea4063ffb2e9a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " denotes the number of frames(or the batch size), " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " denotes the temperature hyperparameter, " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " is the projection head for latent motion, " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\gamma_h^i" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " are the latent motion and rhythmic features at frame (or sample) " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 55, + 371, + 295, + 431 + ], + "type": "text", + "content": " represents cosine similarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "text", + "content": "Unlike CoG, our approach fundamentally differs by incorporating separate local and global rhythmic consistency losses, which are applied to both latent face " + }, + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "text", + "content": " and latent hands " + }, + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 55, + 432, + 295, + 527 + ], + "type": "text", + "content": ", ensuring a more cohesive and synchronized representation across the entire motion sequence. This rhythmic consistency loss ensures that the motions are not only synchronized at the frame level but also maintain a consistent rhythmic flow across the entire sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": "The local frame-level consistency loss " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": " aligns the motion features of each frame with the corresponding rhythmic cues " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": ". By leveraging HuBERT features " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": " instead of basic beat features " + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "inline_equation", + "content": "\\gamma_{b}" + }, + { + "bbox": [ + 55, + 528, + 295, + 612 + ], + "type": "text", + "content": ", which only capture rhythmic pauses, we incorporate rich, high-level audio representations that enhance the model's ability to capture rhythm-related motion patterns and maintain temporal coherence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "text", + "content": "The global sentence-level consistency loss " + }, + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}" + }, + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "text", + "content": " is designed to ensure rhythmic coherence at a global level. Unlike local loss, " + }, + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}" + }, + { + "bbox": [ + 55, + 613, + 295, + 675 + ], + "type": "text", + "content": " reinforces rhythm consistency throughout the sequence, ensuring that the generated motion maintains smooth and rhythm-aligned throughout its duration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": "By jointly minimizing " + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Rhy}}^{(L)}" + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Rhy}}^{(G)}" + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": ", rhythmic consistency learning enables SemTalk to produce base motions that are rhythmically aligned and temporally cohesive," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 538, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 538, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 538, + 316 + ], + "type": "text", + "content": "forming a solid rhythm-related base motion foundation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 539, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 539, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 539, + 337 + ], + "type": "text", + "content": "3.4. Generating Semantic-aware Sparse Motion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 413 + ], + "type": "text", + "content": "The Sparse Motion Generation (Fig. 3 b) in SemTalk adds semantic-aware sparse motion to base motion by incorporating semantic cues drawn from speech content and emotional tone. By separating rhythm and semantics, this stage enhances motion generation by emphasizing contextually meaningful motion at key semantic moments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": "Semantic Speech Encoding. To capture semantic cues in speech, similar to [10], Semantic Emphasis Learning combines frame-level text embeddings " + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\phi_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": ", sentence-level features " + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\phi_g" + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": " from the CLIP model [39], and emotion features " + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\phi_e" + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": " from the emotion2vec model [35]. These features form a comprehensive semantic representation " + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "inline_equation", + "content": "f_t" + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": ", together with audio feature " + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\gamma_h" + }, + { + "bbox": [ + 313, + 414, + 554, + 521 + ], + "type": "text", + "content": ", that reflects both the content and emotional undertones of speech, enabling SemTalk to activate motions that are sensitive to nuanced semantic cues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": "Semantic Emphasis Learning. The process begins by generating " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ", combining local and global cues from text, speech, emotion embeddings and HuBERT features " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ". Then, the sem-gate leverages multi-modal inputs to generate a semantic score, identifying frames that require enhanced semantic emphasis. The sem-gate in SemTalk refines keyframe motion by applying two forms of weighting methods " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ": feature weighting " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{f}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": " and loss weighting " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{l}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ", SemTalk computes a semantic score " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": " which dynamically scales feature weighting—filtering back semantic features " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": " to activate frames with significant relevance, ensuring that the model emphasizes frames aligned with specific communicative intentions. Second, the loss weighting is applied by supervising " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": ", with a classification loss " + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}^{G}" + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": " based on semantic labels, further enhancing the model's ability to identify key frames. The two weight-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13764" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 282, + 173 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 282, + 173 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 282, + 173 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 282, + 173 + ], + "type": "image", + "image_path": "c2aefad23af9a6868a75803373bfe37fae4fe3203eb7f4810dda18663f9b900a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 177, + 296, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 244 + ], + "type": "text", + "content": "Figure 4. Concept comparison with LivelySpeaker [53]. (Top) LivelySpeaker generates semantic gestures with CLIP embeddings in SAG and refines rhythm-related gestures separately using diffusion, causing potential jitter. (Bottom) SemTalk integrates text and speech, uses a semantic gate for fine-grained control, and unifies rhythm and semantics for smoother, more coherent motions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 289 + ], + "type": "text", + "content": "ing methods allow SemTalk to selectively enhance semantic gestures while suppressing uninformative motion, leading to more expressive co-speech motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " is established, it modulates the integration of rhythm-aligned base motion " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " and sparse semantic motion " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": ". Through alpha-blending, frames with high semantic relevance draw more from " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": ", while others rely on " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": ". The final motion codes " + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "inline_equation", + "content": "q^{s}" + }, + { + "bbox": [ + 55, + 289, + 296, + 350 + ], + "type": "text", + "content": " are computed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 361, + 295, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 361, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 110, + 361, + 295, + 374 + ], + "type": "interline_equation", + "content": "q ^ {s} = M L P \\left(\\psi f _ {s} + (1 - \\psi) f _ {b}\\right), \\tag {5}", + "image_path": "548b0d81b3e7ec7eb45b02b3c53a8cae2e6f47e440d6d35f4f00258dcae84105.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 450 + ], + "type": "text", + "content": "To ensure cohesive propagation of semantic emphasis across body regions, we employ the Coarse2Fine Cross-Attention Module, similar to Sec. 3.3. In this stage, we focuses solely on body motion, excluding facial movements, as body gestures play a more critical role in conveying semantic meaning in co-speech interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "text", + "content": "To foster diverse motion generation, SemTalk includes a code classification loss " + }, + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "text", + "content": " and a reconstruction loss " + }, + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 55, + 451, + 295, + 510 + ], + "type": "text", + "content": ". These losses are specifically focused on frames with high semantic scores, guiding the model to prioritize the generation of sparse, meaningful gestures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Discussion. Recently, LivelySpeaker [53] designs the Semantic-Aware Generator (SAG) and Rhythm-Aware Generator (RAG) for co-speech gesture generation, combining them through beat empowerment. While effective, key differences exist between LivelySpeaker and SemTalk, see Fig. 4. First, SAG generates gestures from text using CLIP embeddings, but bridging words and expressive gestures is challenging, causing jitter. SemTalk incorporates speech features (pitch, tone, emotion) alongside text and GT supervision for adaptive gestures. Second, LivelySpeaker applies global control, missing local semantic details, while SemTalk uses fine-grained, frame-level semantic control for subtle variations. Third, LivelySpeaker fuses SAG and RAG in separate latent spaces, leading to misalignment and inconsistencies. SemTalk jointly models rhythm and semantics in a unified framework, ensuring smoother transitions and coherence. We further compare SAG with our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 73, + 436, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 436, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 436, + 84 + ], + "type": "text", + "content": "semantic gate in experiments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 93, + 507, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 93, + 507, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 93, + 507, + 105 + ], + "type": "text", + "content": "3.5. Semantic Score-based motion fusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": "The Adaptive Fusion stage (Fig. 3 c) in SemTalk seamlessly integrates semantic-aware sparse motion into the rhythmic-related base motion. By strategically enhancing frames based on their semantic importance, it maintains a smooth and natural motion flow across sequences. For each frame " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": ", the semantic score " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\psi_i" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": " computed during the Sparse Motion Generation stage is compared to a threshold " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "\\psi_i > \\beta" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": ", the base motion's latent code " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "q_i^r" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": " is replaced with the sparse semantic code " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "q_i^s" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": ", effectively highlighting expressive gestures where they are most relevant; otherwise, " + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "inline_equation", + "content": "q_i = q_i^r" + }, + { + "bbox": [ + 313, + 111, + 553, + 231 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": "This selective replacement emphasizes semantically critical gestures while preserving the natural rhythmic base motion. By blending " + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "inline_equation", + "content": "q^{b}" + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "inline_equation", + "content": "q^{s}" + }, + { + "bbox": [ + 313, + 232, + 553, + 316 + ], + "type": "text", + "content": " based on semantic scores, SemTalk adapts to the expressive needs of the speech context while ensuring coherence. Additionally, the convolution structure of the RVQ-VAE decoder ensures smooth transitions between frames, preserving motion continuity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 327, + 395, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 395, + 341 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 395, + 341 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 348, + 430, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 430, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 430, + 361 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "content": "Datasets. For training and evaluation, we use two datasets: BEAT2 and SHOW. BEAT2, introduced in EMAGE [31], extends BEAT [30] with 76 hours of data from 30 speakers, standardized into a mesh representation with paired audio, text, and frame-level semantic labels. We follow [31] and use the BEAT2-standard subset with an " + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "inline_equation", + "content": "85\\% / 7.5\\% / 7.5\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "content": " train/val/test split. SHOW [48] includes 26.9 hours of high-quality talk show videos with 3D body meshes at 30fps. Since it lacks frame-level semantic labels, we use the semgate from SemTalk, pre-trained on BEAT2, to generate them. Following [48], we select video clips longer than 10 seconds and split the data " + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "inline_equation", + "content": "80\\% / 10\\% / 10\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "content": " for train/val/test." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": "Implementation Details. Our model is trained on a single NVIDIA A100 GPU for 200 epochs with a batch size of 64. We use RVQ-VAE [42], downscaling by 4. The residual quantization has 6 layers, a codebook size of 256 and a dropout rate of 0.2. We use five transformer layers to predict the last five layer codes. In Base Motion Learning, " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\tau = 0.1" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": "; in Sparse Motion Learning, " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " empirically. The training uses ADAM with a 1e-4 learning rate. Following [31], we start with a 4-frame seed pose, gradually increasing masked frames from 0 to " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " over 120 epochs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Metrics. We evaluate generated body gestures using FGD [50] to measure distributional alignment with GT, reflecting realism. DIV [26] quantifies gesture variation via the average L1 distance across clips. BC [27] assesses speech-motion synchrony. For facial expressions, we use MSE [47] to quantify positional differences and LVD [48] to measure discrepancies between GT and generated facial vertices." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13765" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 549, + 404 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 549, + 404 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 549, + 404 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 549, + 404 + ], + "type": "image", + "image_path": "ecfa083670d714d79dd4c95afe184dc95c93b4c4df5646584ffc0173e277becc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 407, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 555, + 453 + ], + "type": "text", + "content": "Figure 5. Comparison on BEAT2 [31] Dataset. SemTalk* refers to the model trained solely on the Base Motion Generation stage, capturing rhythmic alignment but lacking semantic gestures. In contrast, SemTalk successfully emphasized sparse yet vivid motions. For instance, when saying \"my opinion,\" SemTalk generates a hand-raising gesture followed by an index finger extension for emphasis. Similarly, for \"never tell,\" our model produces a clear, repeated gesture matching the rhythm, reinforcing the intended emphasis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 458, + 173, + 533 + ], + "blocks": [ + { + "bbox": [ + 82, + 458, + 173, + 533 + ], + "lines": [ + { + "bbox": [ + 82, + 458, + 173, + 533 + ], + "spans": [ + { + "bbox": [ + 82, + 458, + 173, + 533 + ], + "type": "image", + "image_path": "f09db62adf066365352e12b4bbf519f8db3b6c167df8738489b0a3ac7649784f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 541, + 522, + 553 + ], + "lines": [ + { + "bbox": [ + 86, + 541, + 522, + 553 + ], + "spans": [ + { + "bbox": [ + 86, + 541, + 522, + 553 + ], + "type": "text", + "content": "Figure 6. Comparison on SHOW [48] Dataset. Our method performs better in motion diversity and semantic richness." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 458, + 529, + 534 + ], + "blocks": [ + { + "bbox": [ + 173, + 458, + 529, + 534 + ], + "lines": [ + { + "bbox": [ + 173, + 458, + 529, + 534 + ], + "spans": [ + { + "bbox": [ + 173, + 458, + 529, + 534 + ], + "type": "image", + "image_path": "d15e04660b009631faa3ac6e9e6bff96583a48bd97717a24d032bc6b4078ce89.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 168, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 168, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 168, + 571 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Qualitative Comparisons. We encourage readers to watch our demo video for a clearer understanding of SemTalk's qualitative performance. Our method achieves superior speech-motion alignment, generating more realistic, diverse, and semantically consistent gestures than the baselines. As shown in Fig. 5, LivelySpeaker, TalkSHOW, EMAGE, and DiffSHEG exhibit jitter—EMAGE mainly in the legs and shoulders, while TalkSHOW affects the entire body. LivelySpeaker and DiffSHEG, which focus primarily on the upper body, produce slow and inconsistent motions, especially at speech clip boundaries. DiffSHEG improves" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 556, + 704 + ], + "type": "text", + "content": "gesture diversity over IMAGE and TalkSHOW, though IMAGE maintains greater naturalness. SemTalk surpasses all baselines in both realism and diversity. Compared to SemTalk*, SemTalk generates more expressive gestures, emphasizing key phrases (e.g., raising hands for “dream job” or pointing for “that is why”). While SemTalk* ensures rhythmic consistency, it lacks semantic expressiveness. By integrating frame-level semantic emphasis, SemTalk aligns motion with both rhythm and semantics, demonstrating the effectiveness of rhythmic consistency learning and semantic emphasis learning. In facial comparisons (Fig. 7), IMAGE shows minimal lip movement, while both DiffSHEG and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13766" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 298, + 173 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 298, + 173 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 298, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 298, + 173 + ], + "type": "image", + "image_path": "a717cf177762c731d3be4305e5fb997193b304e071da862c3175abdaaafa8d47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 178, + 284, + 190 + ], + "lines": [ + { + "bbox": [ + 66, + 178, + 284, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 178, + 284, + 190 + ], + "type": "text", + "content": "Figure 7. Facial Comparison on the BEAT2 [31] Dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 195, + 279, + 318 + ], + "blocks": [ + { + "bbox": [ + 67, + 195, + 279, + 318 + ], + "lines": [ + { + "bbox": [ + 67, + 195, + 279, + 318 + ], + "spans": [ + { + "bbox": [ + 67, + 195, + 279, + 318 + ], + "type": "image", + "image_path": "a501893e2c1a842880f342ab42f131c7d71c2b26442d82b3b6605150e683eb60.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "type": "text", + "content": "Figure 8. Qualitative study on semantic score. Semantic score aligns with keywords, influencing gesture intensity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "type": "text", + "content": "EMAGE reveal inconsistencies between lip motion and the rhythm of speech. In contrast, SemTalk produces smooth, natural transitions across syllables, resulting in realistic and expressive lips, significantly surpassing the baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 474 + ], + "type": "text", + "content": "On the SHOW dataset (Fig. 6), SemTalk shows more agile gestures than all baselines, when applied to unseen data. Our method captures natural and contextually rich gestures, particularly in moments of emphasis such as \"I like to do\" and \"relaxing,\" where our model produces lively hand and body movements that align with the speech content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 700 + ], + "type": "text", + "content": "Semantic Score. Fig. 8 shows how semantic emphasis influences gesture intensity, with peaks in the semantic score aligning with keywords like \"comes,\" \"fantastic,\" and \"captured.\" By extracting semantic scores from key frames, we track gesture emphasis trends. Furthermore, as shown in Fig. 9, SemTalk adapts to different emotional tones even when the text remains unchanged. This adaptability prevents overfitting to the text itself, allowing the model to generate gestures that vary according to the emotional delivery of the speech. The learned semantic score provides fine-grained, frame-level control, keeping gestures both rhythmically synchronized and semantically aligned in real time. User Study. We conducted a user study with 10 video samples and 25 participants from diverse backgrounds, evaluating realism, semantic consistency, motion-speech synchrony, and diversity. Participants were required to rank shuffled videos across different methods. As shown in Fig. 10, our approach received dominant preferences across all metrics, especially in semantic consistency and realism." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 70, + 554, + 177 + ], + "blocks": [ + { + "bbox": [ + 319, + 70, + 554, + 177 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 554, + 177 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 554, + 177 + ], + "type": "image", + "image_path": "760d1fd802be9416574db3e923376768c8d359df700277ab247316886991aed9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 227 + ], + "type": "text", + "content": "Figure 9. Same words with different speech from the internet. \"emo\" represents different emotional tones extracted from speech. SemTalk can generate different motions, even when the text script is the same, preventing overfitting to the text itself." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 330, + 231, + 542, + 319 + ], + "blocks": [ + { + "bbox": [ + 330, + 231, + 542, + 319 + ], + "lines": [ + { + "bbox": [ + 330, + 231, + 542, + 319 + ], + "spans": [ + { + "bbox": [ + 330, + 231, + 542, + 319 + ], + "type": "image", + "image_path": "d1f5db3de63a6fab1a877bb143eb01bf4d2b4cb70944e878ac8e7d740c826275.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 327, + 503, + 338 + ], + "lines": [ + { + "bbox": [ + 365, + 327, + 503, + 338 + ], + "spans": [ + { + "bbox": [ + 365, + 327, + 503, + 338 + ], + "type": "text", + "content": "Figure 10. Results of the user study." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 432, + 358 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 483 + ], + "type": "text", + "content": "Comparison with Baselines. As shown in Tab. 1, SemTalk outperforms previous methods on BEAT2, achieving lower FGD, MSE, and LVD, indicating better distribution alignment and reduced motion errors. For fairness, we follow [31] and add a lower-body VQ-VAE to TalkSHOW, Diff-SHEG, and SemTalk. Notably, SemTalk significantly reduces FGD, ensuring strong distribution matching. While TalkSHOW and EMAGE achieve competitive diversity (DIV) scores, SemTalk balances high semantic relevance with natural motion flow." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 556 + ], + "type": "text", + "content": "On the SHOW dataset, SemTalk excels with the lowest FGD, MSE, and the highest BC, indicating precise beat alignment with the audio and enhanced semantic consistency in generated motions. Although EMAGE exhibits high DIV, our model achieves comparable results while maintaining smooth, realistic motion free from jitter." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Sem-gate. Tab. 2 highlights the effectiveness of sem-gate. Without sem-gate, the model fails to emphasize key moments. Randomized semantic scores led to poor performance by preventing meaningful frame distinction. Introducing a learned sem-gate even (w/o " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ") significantly improves semantic alignment and classification accuracy. Refinement is further enhanced through weighting strategies: feature weighting " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_f" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " enhances motion emphasis, while loss weighting " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_l" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " improves FGD and overall accuracy. These results suggest that weighting methods enhance the accuracy of the semantic score and help the model prioritize important frames. The best results come from applying two weighting methods together, where frames with stronger se" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13767" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 272 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 272 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 272 + ], + "type": "table", + "html": "
DatasetMethodFGD↓BC↑DIV↑MSE↓LVD↓
BEAT2FaceFormer [13]---7.7877.593
CodeTalker [44]---8.0267.766
CaMN [30]6.6446.76910.86--
DSG [47]8.8117.24111.49--
LivelySpeaker [17]11.806.65911.28--
Habibie et al. [17]9.0407.7168.2138.6148.043
TalkSHOW [48]6.2096.94713.477.7917.771
EMAGE [31]5.5127.72413.067.6807.556
DiffSHEG [9]8.9867.14211.917.6658.673
SemTalk (Ours)4.2787.77012.916.1536.938
SHOWFaceFormer [13]---138.143.69
CodeTalker [44]---140.745.84
CaMN [30]22.127.71210.37--
DSG [47]24.848.02710.23--
LivelySpeaker [17]32.177.84410.14--
Habibie et al. [17]27.228.2098.541145.647.35
TalkSHOW [48]24.438.24910.98139.645.17
EMAGE [31]22.128.28012.46136.142.44
DiffSHEG [9]24.878.06110.79139.045.77
SemTalk (Ours)20.188.30411.36134.139.15
", + "image_path": "58edef6b5eff11c192d309fb2a029ed365faf04ed21a1a6efe927bab1cd944e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 457 + ], + "type": "text", + "content": "mantic signals receive higher emphasis. We also compare sem-gate with LivelySpeaker's SAG [53]. We find that replacing the Sparse Motion stage with SAG and substituting motion using GT semantic labels led to poor performance. SAG relies only on text-motion alignment, ignoring emotional tone, making it more prone to overfitting the text. In contrast, our sem-gate applies GT supervision with two weighting methods, achieving more accurate and stable semantic motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": "Ablation Study on Components. We assess the impact of each component of our model on BEAT2 and present the results in Tab. 3, which reveals several key insights (more ablation results please see supplementary material):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 556 + ], + "type": "text", + "content": "- Rhythmic Consistency Learning (RC) not only boosts performance on key metrics like FGD, LVD, and BC but also reduces the MSE, contributing to smoother and more realistic base motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "- Semantic Emphasis Learning (SE) proves essential for selectively enhancing semantic-rich gestures. The inclusion of SE, as shown in rows with SE enabled, improves both diversity (DIV) and FGD, enabling the model to emphasize semantically relevant motions. SE demonstrates its effectiveness in focusing on frame-level semantic information, which contributes to the generation of lifelike gestures with enriched contextual meaning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "- Coarse2Fine Cross-Attention Module (C2F) effectively refines motion details, improving BC, FGD, and DIV. When combined with RVQ and RC, C2F achieves the best MSE and LVD, highlighting its role in enhancing motion realism and diversity hierarchically." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 548, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with SOTA. SemTalk consistently outperforms baselines across both the BEAT2 and SHOW datasets. Lower values are better for FMD, FGD, MSE, and LVD. Higher values are better for BC and DIV. We report " + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{FGD} \\times 10^{-1}" + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{BC} \\times 10^{-1}" + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE} \\times 10^{-8}" + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{LVD} \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 274, + 295, + 331 + ], + "type": "text", + "content": " for simplify." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 70, + 548, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 548, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 548, + 159 + ], + "type": "table", + "html": "
MethodFGD↓BC↑DIV↑Acc(%)↑
w/o Sem-gate4.8937.70212.42-
SAG (LivelySpeaker [53])4.6187.68212.45-
Sem-gate (Random ψ)4.6347.70012.4450.07
Sem-gate (w/o W)4.4957.63312.2672.32
Sem-gate (w/ Wf)4.4087.67912.2878.52
Sem-gate (w/ Wl)4.3667.77211.9477.83
Sem-gate (ours)4.2787.77012.9182.76
", + "image_path": "46044737566a791b5e411fd91acb5ab4722fbdb7212022197ec7a2d652deb453.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 274, + 548, + 365 + ], + "blocks": [ + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": "Table 2. Ablation study on Sem-gate. \"Acc\" denotes semantic classification performance on BEAT2. \"w/o Sem-gate\" means directly input " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\gamma_{h}" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": " without Sem-gate. \"SAG (LivelySpeaker [53])\" replaces the Sparse Motion Generation stage with LivelySpeaker's SAG method. \"Random " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": "\" assigns frame-level scores randomly. \"w/o " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": "\" applies the semantic gate but excludes frame-level weighting. \"w/ " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{f}" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": "\" applies feature weighting. \"w/ " + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{l}" + }, + { + "bbox": [ + 313, + 163, + 553, + 273 + ], + "type": "text", + "content": "\" applies loss weighting. (as mentioned in Sec. 3.4). Sem-gate (ours) integrates both the semantic gate and frame-level weighting to enhance emphasis." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 274, + 548, + 365 + ], + "lines": [ + { + "bbox": [ + 322, + 274, + 548, + 365 + ], + "spans": [ + { + "bbox": [ + 322, + 274, + 548, + 365 + ], + "type": "table", + "html": "
RCSEC2FRVQFGD↓BC↑DIV↑MSE↓LVD↓
----6.2347.62811.448.2397.831
---√5.4847.64111.8413.88215.42
√--√4.8677.70112.386.2016.928
√√-√4.5267.75112.836.2156.997
--√√4.8977.70212.4213.41615.72
√√--5.8317.75811.976.5877.106
√-√√4.3977.77612.496.1006.898
√√√√4.2787.77012.916.1536.938
", + "image_path": "44f4ce8af71fc71a6a99c3ea070234be949d6df598113e8ea29ab0ea3b82177e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 413 + ], + "type": "text", + "content": "Table 3. Ablation study on each key component. \"RC\" denotes rhythmic consistency learning, \"SE\" denotes the semantic emphasis learning, and \"C2F\" denotes Coarse2Fine Cross-Att Module, \"RVQ\" denotes the RVQ-VAE." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 430, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 555, + 479 + ], + "type": "text", + "content": "- RVQ-VAE (RVQ) enhances the diversity and realism of generated motion. Though it slightly increases MSE and LVD, it notably improves FGD, leading to more natural motion generation compared to standard VQ-VAE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "content": "We propose SemTalk, a novel approach for holistic cospeech motion generation with frame-level semantic emphasis. Our method addresses the integration of sparse yet expressive motion into foundational rhythm-related motion, which has received less attention in previous works. We develop a framework that separately learns rhythm-related base motion through coarse2fine cross-attention module and rhythmic consistency learning, while capturing semantic-aware motion through Semantic Emphasis Learning. These components are then adaptively fused based on a learned semantic score. Our approach has demonstrated state-of-the-art performance on two public datasets quantitatively and qualitatively. The qualitative results and user study show that our method can generate high-quality cospeech motion sequences that enhance frame-level semantics over robust base motions, reflecting the full spectrum of human expressiveness." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13768" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by Alibaba Research Intern Program, the Young Scientists Fund of the National Natural Science Foundation of China No. 624B2110, the National Key Research and Development Program of China No. 2024YFC3015600, the Fundamental Research Funds for Central Universities No.2042023KF0180 & No.2042025KF0053. The numerical calculation is supported by supercomputing system in Super-computing Center of Wuhan University and Tongyi Lab, Alibaba Group." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 224, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "type": "text", + "content": "[1] Chaitanya Ahuja, Dong Won Lee, and Louis-Philippe Morency. Low-resource adaptation for personalized cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20566–20576, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 281, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 294, + 323 + ], + "type": "text", + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 294, + 357 + ], + "type": "text", + "content": "[3] Tenglong Ao, Zeyi Zhang, and Libin Liu. Gesture diffuclip: Gesture diffusion model with clip latents. ACM Transactions on Graphics (TOG), 42(4):1-18, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 359, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 294, + 423 + ], + "type": "text", + "content": "[4] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 31:2523-2533, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 501 + ], + "type": "text", + "content": "[5] Justine Cassell, Catherine Pelachaud, Norman Badler, Mark Steedman, Brett Achorn, Tripp Becket, Brett Douville, Scott Prevost, and Matthew Stone. Animated conversation: rule-based generation of facial expression, gesture & spoken intonation for multiple conversational agents. In Proceedings of the 21st annual conference on Computer graphics and interactive techniques, pages 413-420, 1994. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 502, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 545 + ], + "type": "text", + "content": "[6] Justine Cassell, David McNeill, and Karl-Erik McCullough. Speech-gesture mismatches: Evidence for one underlying representation of linguistic and nonlinguistic information. *Pragmatics & cognition*, 7(1):1-34, 1999. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 548, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 590 + ], + "type": "text", + "content": "[7] Justine Cassell, Hannes Högni Vilhjalmsson, and Timothy Bickmore. Beat: the behavior expression animation toolkit. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 477-486, 2001. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 646 + ], + "type": "text", + "content": "[8] Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. Enabling synergistic full-body control in prompt-based co-speech motion generation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6774-6783, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[9] Junming Chen, Yunfei Liu, Jianan Wang, Ailing Zeng, Yu Li, and Qifeng Chen. Diffsheg: A diffusion-based approach for real-time speech-driven holistic 3d expression and gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7352-7361, 2024. 2, 3, 8" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[10] Kiran Chhatre, Nikos Athanasiou, Giorgio Becherini, Christopher Peters, Michael J Black, Timo Bolkart, et al. Emotional speech-driven 3d body animation via disentangled latent diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1942-1953, 2024. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 184 + ], + "type": "text", + "content": "[11] Selina Chu, Shrikanth Narayanan, and C-C Jay Kuo. Environmental sound recognition with time-frequency audio features. IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142-1158, 2009. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[12] Radek Daneček, Kiran Chhatre, Shashank Tripathi, Yandong Wen, Michael Black, and Timo Bolkart. Emotional speech-driven animation with content-emotion disentanglement. In SIGGRAPH Asia 2023 Conference Papers, pages 1-13, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 242, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 242, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 553, + 296 + ], + "type": "text", + "content": "[13] Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, and Taku Komura. Faceformer: Speech-driven 3d facial animation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18770-18780, 2022. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 298, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 329 + ], + "type": "text", + "content": "[14] Susan Goldin-Meadow. The role of gesture in communication and thinking. Trends in cognitive sciences, 3(11):419-429, 1999. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 386 + ], + "type": "text", + "content": "[15] Kehong Gong, Dongze Lian, Heng Chang, Chuan Guo, Zihang Jiang, Xinxin Zuo, Michael Bi Mi, and Xinchao Wang. Tm2d: Bimodality driven 3d dance generation via music-text integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9942–9952, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[16] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 445, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 509 + ], + "type": "text", + "content": "[17] Ikhsanul Habibie, Weipeng Xu, Dushyant Mehta, Lingjie Liu, Hans-Peter Seidel, Gerard Pons-Moll, Mohamed Elgharib, and Christian Theobalt. Learning speech-driven 3d conversational gestures from video. In Proceedings of the 21st ACM International Conference on Intelligent Virtual Agents, pages 101-108, 2021. 1, 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 511, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 575 + ], + "type": "text", + "content": "[18] Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units. IEEE/ACM transactions on audio, speech, and language processing, 29: 3451-3460, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 578, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 621 + ], + "type": "text", + "content": "[19] Chien-Ming Huang and Bilge Mutlu. Robot behavior toolkit: generating effective social behaviors for robots. In Proceedings of the seventh annual ACM/IEEE international conference on Human-Robot Interaction, pages 25-32, 2012. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 624, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 644 + ], + "type": "text", + "content": "[20] Adam Kendon. Gesture: Visible action as utterance. Cambridge University Press, 2004. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 647, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 677 + ], + "type": "text", + "content": "[21] Michael Kipp. Gesture generation by imitation: From human behavior to computer character animation. Universal-Publishers, 2005. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 712 + ], + "type": "text", + "content": "[22] Stefan Kopp, Brigitte Krenn, Stacy Marsella, Andrew N Marshall, Catherine Pelachaud, Hannes Pirker, Kristinn R Thorisson, and Hannes Vilhjalmsson. Towards a common" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13769" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "text", + "content": "framework for multimodal generation: The behavior markup language. In Intelligent Virtual Agents: 6th International Conference, IVA 2006, Marina Del Rey, CA, USA, August 21-23, 2006. Proceedings 6, pages 205-217. Springer, 2006. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 195 + ], + "type": "text", + "content": "[23] Taras Kucherenko, Patrik Jonell, Sanne Van Waveren, Gustav Eje Henter, Simon Alexandersson, Iolanda Leite, and Hedvig Kjellström. Gesticulator: A framework for semantically-aware speech-driven gesture generation. In Proceedings of the 2020 international conference on multimodal interaction, pages 242-250, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 262 + ], + "type": "text", + "content": "[24] Taras Kucherenko, Patrik Jonell, Youngwoo Yoon, Pieter Wolfert, and Gustav Eje Henter. A large, crowdsourced evaluation of gesture generation systems on common data: The genea challenge 2020. In Proceedings of the 26th International Conference on Intelligent User Interfaces, pages 11-21, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "type": "text", + "content": "[25] Alex Lascarides and Matthew Stone. A formal semantic analysis of gesture. Journal of Semantics, 26(4):393-449, 2009. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "type": "text", + "content": "[26] Jing Li, Di Kang, Wenjie Pei, Xuefei Zhe, Ying Zhang, Zhenyu He, and Linchao Bao. Audio2gestures: Generating diverse gestures from speech audio with conditional variational autoencoders. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11293-11302, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 420 + ], + "type": "text", + "content": "[27] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13401-13412, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "text", + "content": "[28] Yuanzhi Liang, Qianyu Feng, Linchao Zhu, Li Hu, Pan Pan, and Yi Yang. Seeg: Semantic energized co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10473-10482, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "text", + "content": "[29] Haiyang Liu, Naoya Iwamoto, Zihao Zhu, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Disco: Disentangled implicit content and rhythm learning for diverse co-speech gestures synthesis. In Proceedings of the 30th ACM international conference on multimedia, pages 3764-3773, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 545, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 612 + ], + "type": "text", + "content": "[30] Haiyang Liu, Zihao Zhu, Naoya Iwamoto, Yichen Peng, Zhengqing Li, You Zhou, Elif Bozkurt, and Bo Zheng. Beat: A large-scale semantic and emotional multi-modal dataset for conversational gestures synthesis. In European conference on computer vision, pages 612-630. Springer, 2022. 1, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "type": "text", + "content": "[31] Haiyang Liu, Zihao Zhu, Giorgio Becherini, Yichen Peng, Mingyang Su, You Zhou, Xuefei Zhe, Naoya Iwamoto, Bo Zheng, and Michael J Black. Emage: Towards unified holistic co-speech gesture generation via expressive masked audio gesture modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1144-1154, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[32] Xian Liu, Qianyi Wu, Hang Zhou, Yinghao Xu, Rui Qian, Xinyi Lin, Xiaowei Zhou, Wayne Wu, Bo Dai, and Bolei" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 118 + ], + "type": "text", + "content": "Zhou. Learning hierarchical cross-modal association for cospeech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10462-10472, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[33] Yifei Liu, Qiong Cao, Yandong Wen, Huaiguang Jiang, and Changxing Ding. Towards variable and coordinated holistic co-speech motion generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1566-1576, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[34] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. arXiv preprint arXiv:2310.12978, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[35] Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[36] Stacy Marsella, Yuyu Xu, Margaux Lhommet, Andrew Feng, Stefan Scherer, and Ari Shapiro. Virtual character performance from speech. In Proceedings of the 12th ACM SIGGRAPH/Eurographics symposium on computer animation, pages 25-35, 2013. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 321, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 386 + ], + "type": "text", + "content": "[37] Evonne Ng, Javier Romero, Timur Bagautdinov, Shaojie Bai, Trevor Darrell, Angjoo Kanazawa, and Alexander Richard. From audio to photoreal embodiment: Synthesizing humans in conversations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1001-1010, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[38] Ashi Özyürek, Roel M Willems, Sotaro Kita, and Peter Hagoort. On-line integration of semantic information from speech and gesture: Insights from event-related brain potentials. Journal of cognitive neuroscience, 19(4):605-616, 2007. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 509 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 566 + ], + "type": "text", + "content": "[40] Manuel Rebol, Christian Güttl, and Krzysztof Pietroszek. Real-time gesture animation generation from speech for virtual human interaction. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1-4, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "text", + "content": "[41] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized talking head synthesis. arXiv preprint arXiv:2301.03786, 2(4):5, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "type": "text", + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[43] Hanwei Wu and Markus Flierl. Learning product codebooks using vector-quantized autoencoders for image retrieval. In 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP), pages 1-5. IEEE, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[44] Jinbo Xing, Menghan Xia, Yuechen Zhang, Xiaodong Cun, Jue Wang, and Tien-Tsin Wong. Codetalker: Speech-driven" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13770" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 675 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "3d facial animation with discrete motion prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12780-12790, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[45] Zunnan Xu, Yachao Zhang, Sicheng Yang, Ronghui Li, and Xiu Li. Chain of generation: Multi-modal gesture synthesis via cascaded conditional control. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6387-6395, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "type": "text", + "content": "[46] Sicheng Yang, Zilin Wang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Qiaochu Huang, Lei Hao, Songcen Xu, Xiaofei Wu, Changpeng Yang, et al. Unified gesture: A unified gesture synthesis model for multiple skeletons. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1033-1044, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 284 + ], + "type": "text", + "content": "[47] Sicheng Yang, Zhiyong Wu, Minglei Li, Zhensong Zhang, Lei Hao, Weihong Bao, Ming Cheng, and Long Xiao. Diffusestylegesture: Stylized audio-driven co-speech gesture generation with diffusion models. arXiv preprint arXiv:2305.04919, 2023. 2, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 340 + ], + "type": "text", + "content": "[48] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 469-480, 2023. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 406 + ], + "type": "text", + "content": "[49] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 295, + 462 + ], + "type": "text", + "content": "[50] Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. Speech gesture generation from the trimodal context of text, audio, and speaker identity. ACM Transactions on Graphics (TOG), 39 (6):1-16, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 464, + 295, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 295, + 517 + ], + "type": "text", + "content": "[51] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "type": "text", + "content": "[52] Jinsong Zhang, Minjie Zhu, Yuxiang Zhang, Zerong Zheng, Yebin Liu, and Kun Li. Speechact: Towards generating whole-body motion from speech. IEEE Transactions on Visualization and Computer Graphics, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 565, + 295, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 295, + 619 + ], + "type": "text", + "content": "[53] Yihao Zhi, Xiaodong Cun, Xuelin Chen, Xi Shen, Wen Guo, Shaoli Huang, and Shenghua Gao. Livelyspeaker: Towards semantic-aware co-speech gesture generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20807-20817, 2023. 2, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 295, + 675 + ], + "type": "text", + "content": "[54] Lingting Zhu, Xian Liu, Xuanyu Liu, Rui Qian, Ziwei Liu, and Lequan Yu. Taming diffusion models for audio-driven co-speech gesture generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10544-10553, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13771" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_content_list.json b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e212e2d95ce030ce00f9475b35980340ef967d73 --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_content_list.json @@ -0,0 +1,2022 @@ +[ + { + "type": "text", + "text": "Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers", + "text_level": 1, + "bbox": [ + 107, + 128, + 888, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunshan Zhong $^{1,2}$ , Yuyao Zhou $^{2}$ , Yuxin Zhang $^{2}$ , Wanchen Sui $^{3}$ , Shen Li $^{3}$ , Yong Li $^{3}$ , Fei Chao $^{2}$ , Rongrong Ji $^{1,2*}$", + "bbox": [ + 163, + 202, + 833, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Artificial Intelligence, Xiamen University", + "bbox": [ + 284, + 238, + 712, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ MAC Lab, School of Informatics, Xiamen University $^{3}$ Alibaba Group", + "bbox": [ + 214, + 255, + 782, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "viper.zhong@gmail.com, {yuyaozhou, yuxinzhang}@stu.xmu.edu.cn \n{wanchen.swc, litan.ps, jiufeng.ly}@alibaba-inc.com, {fchao, rrji}@xmu.edu.cn", + "bbox": [ + 156, + 277, + 834, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Data-free quantization (DFQ) enables model quantization without accessing real data, addressing concerns regarding data security and privacy. With the growing adoption of Vision Transformers (ViTs), DFQ for ViTs has garnered significant attention. However, existing DFQ methods exhibit two limitations: (1) semantic distortion, where the semantics of synthetic images deviate substantially from those of real images, and (2) semantic inadequacy, where synthetic images contain extensive regions with limited content and oversimplified textures, leading to suboptimal quantization performance. To address these limitations, we propose SARDFQ, a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs. To address semantic distortion, SARDFQ incorporates Attention Priors Alignment (APA), which optimizes synthetic images to follow randomly generated structure attention priors. To mitigate semantic inadequacy, SARDFQ introduces Multi-Semantic Reinforcement (MSR), leveraging localized patch optimization to enhance semantic richness across synthetic images. Furthermore, SARDFQ employs Soft-Label Learning (SL), wherein multiple semantic targets are adapted to facilitate the learning of multi-semantic images augmented by MSR. Extensive experiments demonstrate the effectiveness of SARDFQ, significantly surpassing existing methods. For example, SARDFQ improves top-1 accuracy on ImageNet by $15.52\\%$ for W4A4 ViT-B1.", + "bbox": [ + 89, + 378, + 483, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 803, + 220, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) [20, 29, 36] have demonstrated remarkable success across various computer vision tasks [1,", + "bbox": [ + 89, + 830, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b50f3a45edef337a0f4073e0ff2e040144da215e4fb9126beb74c732643f2fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 344, + 846, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(a) The t-SNE[69] visualization of the penultimate-layer features (extracted by DeiT-S) of synthetic images. Each marker (circle, star, triangle) represents a distinct category. The red dashed circles highlight the features extracted from our APA and real images. Notably, the features produced by PSAQ-ViT[47] exhibit substantial deviation from those of real images, indicating semantic distortion. In contrast, our APA yields features that more closely align with those of real images, suggesting improved semantics alignment.", + "bbox": [ + 511, + 518, + 906, + 616 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/249fb76fabd4d8d549bcfc3fb1bc6e46377268668c0690f604df88d19982974d.jpg", + "image_caption": [ + "PSAQ" + ], + "image_footnote": [], + "bbox": [ + 573, + 617, + 665, + 744 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39086b1071ec1e2cac7df6a959b641a108c4ee0baa0e9e30a61a0d1ebef8644d.jpg", + "image_caption": [ + "PSAQ V2" + ], + "image_footnote": [], + "bbox": [ + 666, + 617, + 754, + 744 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faf3b7d4f308588c04140baf283cbb66741b4e3f32e492455dcd0bf875558327.jpg", + "image_caption": [ + "SARDFQ (Ours)" + ], + "image_footnote": [], + "bbox": [ + 754, + 617, + 844, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(b) The images of PSAQ-ViT and PSAQ-ViT V2 exhibit numerous dull regions with limited content and simplified textures, reflecting semantic inadequacy. In comparison, our SARDFQ generates images with greater diversity in both content and texture, demonstrating enhanced semantics.", + "bbox": [ + 511, + 757, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Illustration of (a) semantic distortion and (b) semantic inadequacy.", + "bbox": [ + 511, + 819, + 906, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "7, 56, 68]. However, their high computational cost and substantial memory footprint hinder deployment in resource-", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: rrji@xmu.edu.cn", + "bbox": [ + 107, + 875, + 334, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The code is at https://github.com/zysxmu/SARDFQ.", + "bbox": [ + 109, + 887, + 447, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12479", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "constrained environments [11, 30, 32, 46, 55, 67, 82]. To address these limitations, quantization [39] has emerged as a promising solution, which reduces model complexity by converting full-precision weight and activations into low-bit representations.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional quantization methods typically require access to the original training dataset, which raises data privacy and security concerns [6, 27, 77, 84]. As a result, data-free quantization (DFQ) has gained increasing attention, allowing quantization without the need for real data [14, 76, 81]. However, most existing DFQ methods are designed specifically for convolutional neural networks (CNNs) and are not directly applicable to vision transformers (ViTs). These methods generally rely on batch normalization statistics (BNS), which capture the distribution of real data, to synthesize in-distribution synthetic data [6, 76, 81, 84]. Yet, BNS is unavailable for ViTs, which use layer normalization (LN) to dynamically compute distribution statistics during inference [47]. Recently, several DFQ methods have been proposed for ViTs [16, 33, 47, 48, 62]. For example, PSAQ-ViT [47] introduces patch similarity entropy (PSE) loss to optimize Gaussian noise towards usable synthetic images.", + "bbox": [ + 89, + 167, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, we observe that existing methods suffer from semantic distortion and inadequacy issues. As shown in Fig. 1a, features of synthetic images generated by PSAQ-ViT deviate significantly from those of real images. Tab. 1 shows that the cosine similarity between synthetic images from PSAQ-ViT and real images is notably low, also indicating significant distortion. These results highlight the issue of semantic distortion. Moreover, as shown in Fig. $1\\mathrm{b}^2$ synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit many regions with limited content diversity and overly simplified textures. These low-quality dull regions are useless or even detrimental to model learning [33], highlighting the issue of semantic inadequacy. Consequently, quantized models trained on such low-quality images suffer from degraded performance.", + "bbox": [ + 89, + 424, + 482, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above analysis, we propose a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs, termed SARDFQ. The overall framework is depicted in Fig.2. To address the semantic distortion issue, SARDFQ introduces Attention Priors Alignment (APA), where synthetic images are optimized to follow structured attention priors generated using Gaussian Mixture Models (GMMs). APA effectively aligns the semantics of synthetic images with real images, as validated by both visual and quantitative analyses. As shown in Fig. 1a, features of APA exhibit a closer alignment to those of real images, while quantitative results in Tab. 1 also confirm that the semantics of APA are more consistent to the real images, indicating enhanced semantics alignment. To address the semantic inadequacy issue, SARDFQ incorpo", + "bbox": [ + 89, + 651, + 482, + 876 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rates Multi-Semantic Reinforcement (MSR) and Softlabel Learning (SL). MSR utilizes localized patch optimization, which encourages different sub-patches of synthetic images to capture various semantics, reinforcing the rich semantics across images. SL applies multiple semantic targets to accommodate the learning of multi-semantic images augmented by MSR. As shown in Fig. 1b, synthetic images after applying MSR exhibit greater diversity in content and texture, providing reinforced semantics.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results across various ViT models and tasks demonstrate that SARDFQ presents substantial performance improvements. For example, SARDFQ achieves a $15.52\\%$ increase in top-1 accuracy on the ImageNet dataset for the W4A4 ViT-B model.", + "bbox": [ + 511, + 227, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 319, + 661, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision Transformers", + "text_level": 1, + "bbox": [ + 511, + 344, + 707, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The great success of transformers in the natural language processing field has driven widespread attempts in the computer vision community to apply them to vision tasks [12, 28, 72]. ViT [20] is the pioneer that builds a transformer-based model to handle images, boosting the performance on the image classification task. DeiT [68] introduces an efficient teacher-student training strategy where a distillation token is employed to distill knowledge from the teacher model to the student model. Swin Transformers [54] builds an efficient and effective hierarchical model by introducing a shifted window-based self-attention mechanism. Other than the image classification task, the applications of ViTs also have broadened considerably, manifesting groundbreaking performance in object detection [7], image segmentation [9, 83], low-level vision [50], video recognition [1, 59], and medical image processing [64], etc. Nevertheless, the impressive performance of ViTs relies on a high number of parameters and significant computational overhead, preventing deployment in resource-constrained environments. Several recent efforts design lightweight ViTs, such as MobileViT [57], MiniVit [80], and TinyViT [73]. However, the model complexity is still unsatisfactory [47].", + "bbox": [ + 511, + 367, + 903, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Network Quantization", + "text_level": 1, + "bbox": [ + 511, + 710, + 720, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-Driven Quantization. Model quantization reduces the complexity of neural networks by replacing full-precision weight and activation with the low-bit format. Data-driven quantization can be roughly divided into two categories: quantization-aware training (QAT) and posttraining quantization (PTQ). QAT is compute-heavy since it re-trains the quantized model with the full training data to retain performance [22, 25, 41, 43, 46, 52, 65, 74]. PTQ perform quantization with a tiny dataset and a reduced time overhead, harvesting widespread attention [3, 42]. The specific architecture of ViTs, such as LayerNorm and the self", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2More results are presented in the appendix.", + "bbox": [ + 107, + 886, + 344, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12480", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1b7b933fd4ed13c1c2693d8e289055bfe0cf694f3afb23c5f54be139c7e73d04.jpg", + "image_caption": [ + "Figure 2. SARDFQ Framework overview: Attention Priors Alignment (APA) employs randomly generated attention priors to improve semantics alignment. Multi-Semantic Reinforcement (MSR) learns the different regions of synthetic images with various semantics to enhance overall semantic richness. Meanwhile, Softlabel Learning (SL) adopts multiple semantic targets to ensure consistent learning of multi-semantic images augmented by MSR." + ], + "image_footnote": [], + "bbox": [ + 171, + 85, + 823, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attention module, urges distinct PTQ methods compared to CNNs [18, 24, 46, 51, 53, 86]. For example, Liu et al. [55] develop a ranking loss to maintain the relative order of the self-attention activation. Unfortunately, both QAT and PTQ involve the original training data, causing concerns about data privacy and security issues in data-sensitive scenarios.", + "bbox": [ + 88, + 316, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-Free Quantization. DFQ quantizes models without accessing real data [2, 13-15, 27, 35, 40, 44, 45, 61]. Most previous DFQ methods focus on CNN, where the BNS can be adopted as the regularization term [6, 81]. However, BNS is infeasible for ViTs built on the LN. Recently, few efforts have been explored to accommodate ViTs [16, 33, 47, 48, 62]. PSAQ-ViT [47] introduces the first DFQ method for ViTs. They discover that Gaussian noise yields homogeneous patches, while the real image yields heterogeneous patches. Thus, patch similarity entropy (PSE) loss is proposed to optimize the Gaussian noise towards real-like images by making them showcase heterogeneous patches. Based on PSAQ-ViT, PSAQ-ViT V2 [48] further introduces an adversarial learning strategy [26]. [62] incorporates contrastive learning and proposes an iterative generation-quantization PTQ-based DFQ method. [33] proposes a sparse generation method to remove noisy and hallucination backgrounds in synthetic images.", + "bbox": [ + 93, + 410, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 698, + 181, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 724, + 230, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1.Quantizers", + "text_level": 1, + "bbox": [ + 89, + 747, + 215, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ the linear quantizer for all weights and activations, except for the attention scores, which use a log2 quantizer to handle highly non-negative and uneven values [24, 49, 51]. For the linear quantizer, given a full-precision input $\\mathbf{x}$ and bit-width $b$ , the quantized value $\\mathbf{x}_q$ and the de-quantized value $\\bar{\\mathbf{x}}$ are computed as follows:", + "bbox": [ + 89, + 767, + 482, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot \\left(\\mathbf {x} _ {q} - z\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 869, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lfloor \\cdot \\rfloor$ denotes rounding to the nearest integer, and clip limits the value to $[0,2^{b} - 1]$ . Here, $\\Delta$ and $z$ are the scale factor and zero-point, respectively. For the log2 quantizer:", + "bbox": [ + 511, + 316, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor - \\log_ {2} \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil , 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot 2 ^ {- \\mathbf {x} _ {q}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 375, + 906, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Data Synthesis", + "text_level": 1, + "bbox": [ + 511, + 411, + 663, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DFQ methods parameterize synthetic images and optimize them toward real-like images with a pre-trained full-precision model $F$ . Given a image $\\tilde{I}$ initialized from Gaussian noise, the one-hot loss [75] is introduced to learn label-related semantics:", + "bbox": [ + 511, + 430, + 905, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {O H}} (\\tilde {\\boldsymbol {I}}) = C E (F (\\tilde {\\boldsymbol {I}}), c) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 516, + 906, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $CE(\\cdot, \\cdot)$ represents the cross entropy, $c$ is a random class label, and $F(\\cdot)$ returns the predicted probability for image $\\tilde{I}$ .", + "bbox": [ + 511, + 547, + 905, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, total variance (TV) [77] loss is a smoothing regularization term to improve the image quality:", + "bbox": [ + 511, + 593, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}) = \\iint | \\nabla \\tilde {\\boldsymbol {I}} (\\tau_ {1}, \\tau_ {2}) | d \\tau_ {1} d \\tau_ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 633, + 906, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\nabla \\tilde{I} (\\tau_1,\\tau_2)$ denotes the gradient at $\\tilde{I}$ at $(\\tau_{1},\\tau_{2})$", + "bbox": [ + 511, + 678, + 870, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To perform DFQ for ViTs, PSAQ-ViT [47] proposes patch similarity entropy (PSE) loss. It first compute patch similarity $\\Gamma_{l}[i,j] = \\frac{u_{i} \\cdot u_{j}}{||u_{i}||||u_{j}||}$ , where $u_{i}, u_{j}$ are feature vectors of MHSA outputs in $l$ -th block and $||\\cdot||$ denotes the $l_{2}$ norm. Then, it estimates the density function $\\hat{f}_{l}(x) = \\frac{1}{Mh}\\sum_{m=1}^{M}K\\left(\\frac{x - x_{m}}{h}\\right)$ , where $K(\\cdot)$ is a normal kernel, $h$ is the bandwidth, and $x_{m}$ is the kernel center derived from $\\Gamma_{l}$ . Finally, the PSE loss is defined as:", + "bbox": [ + 511, + 695, + 905, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {P S E}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = 1} ^ {L} \\int \\hat {f} _ {l} (x) \\log [ \\hat {f} _ {l} (x) ] d x, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 833, + 906, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L$ is the block number of the model.", + "bbox": [ + 511, + 885, + 797, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12481", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/8e49475232f1a4334a49796ef4e3b995aa80fada2b27fb929cd6cf44f8dd8986.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method ClassesRealPSAQ-ViT [47]APA (Ours)
Class 10.680.440.64
Class 20.320.260.31
Class 30.410.310.36
", + "bbox": [ + 94, + 88, + 480, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Average cosine similarity of three randomly selected classes. For real images, the similarity is measured within the class itself, while for PSAQ-ViT and APA, the similarity is measured between synthetic and real images of the same class. The results show that APA achieves higher similarity than PSE loss, indicating aligned semantics.", + "bbox": [ + 89, + 186, + 483, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Observations", + "text_level": 1, + "bbox": [ + 89, + 297, + 230, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing DFQ methods have made significant progress. However, through carefully analyzing the synthetic images, we reveal that these images suffer from issues of semantic distortion and semantic inadequacy, both of which hinder further advancement in the DFQ of ViTs.", + "bbox": [ + 89, + 321, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The semantic distortion issue refers to the significant divergence between the semantics of synthetic images and real images. To demonstrate this, we visualize the features of synthetic and real images in Fig. 1a. Note that the penultimate feature typically is regarded to represent the semantics of the input [4, 58, 78, 79]. It is clear that the features of PSAQ-ViT diverge significantly from real images, suggesting that these images fail to capture the true semantic distribution of real data. Tab. 1 quantitatively measures the semantics using the average cosine similarity (ranging from $-1$ to $1$ ). We also report the intra-class similarity within real images as an approximate upper bound for comparison. This result further supports the observation of low similarity between synthetic images from PSAQ-ViT and real images. For instance, for Class 1, the intra-class similarity within real images is 0.68, while PSAQ-ViT achieves only 0.44.", + "bbox": [ + 89, + 398, + 483, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The semantic inadequacy issue refers to the presence of dull regions in synthetic images, which contain redundant or non-semantic content [33, 37], hindering the model's learning process. As indicated in [13, 14], a diverse content and textures generally suggests rich information. As shown in Fig. 1b, many regions of synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit a lack of diversity in content, with overly simplified textures. Specifically, the central region of PSAQ-ViT images only contains faint object structures, while PSAQ-ViT V2 images appear excessively smoothed and indistinct.", + "bbox": [ + 89, + 642, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For high-bit quantization, where model capacity is largely retained [42], the performance degradation remains relatively minor even using semantic distorted and inadequate images [47, 62]. However, in low-bit quantization, where model capacity is severely damaged and informative images are essential for recovering performance [13, 84],", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fine-tuning on these poor-quality images leads to inferior generalization to real datasets, resulting in limited performance. For example, as shown in Tab. 2, the W4A4 ViT-B fine-tuned on real images yields $68.16\\%$ , whereas PSAQ-ViT only achieves $36.32\\%$ .", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a76c498bc5d105f8c3d2cd2c8ad6b212c421d9237e70e6e0570cd8cd08d8f7fd.jpg", + "image_caption": [ + "Figure 3. Comparison between attention maps." + ], + "image_footnote": [], + "bbox": [ + 573, + 179, + 848, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantics Alignment and Reinforcement Data-Free Quantization", + "text_level": 1, + "bbox": [ + 511, + 321, + 905, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we introduce the proposed SARDFQ, whose framework is illustrated in Fig. 2.", + "bbox": [ + 511, + 359, + 905, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Attention Priors Alignment", + "text_level": 1, + "bbox": [ + 511, + 397, + 751, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In ViTs, the self-attention mechanism encodes semantic correlations between image regions, where high-response areas in attention maps strongly correlate with semantic-discriminative content [10, 20, 31]. However, existing DFQ methods overlook this intrinsic property in the generation process. As a result, as shown in Fig. 3, synthetic images often exhibit disordered and unnatural attention patterns, with attention maps either overly diffuse or misaligned toward peripheral regions. This undermines their ability to preserve semantic-discriminative content, causing semantic distortion, as demonstrated in Fig. 1a and Tab. 1. In response, we propose Attention Priors Alignment (APA), which improves semantics alignment by optimizing synthetic images to follow randomly generated structure attention priors.", + "bbox": [ + 511, + 416, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given a synthetic image $\\tilde{I}$ , we first obtain its attention maps in the $h$ -th head of the $l$ -th block, denoting as $\\mathbf{A}_{l,h} \\in \\mathbb{R}^{N \\times N}$ , where $N$ represents the total number of tokens. In DeiT, the attention of the classification token toward other tokens serves as the indicator for semantic versus non-semantic parts [5]. Thus, we extract $\\mathbf{A}_{l,h}^c \\in \\mathbb{R}^{1 \\times (N - 1)}$ from $\\mathbf{A}_{l,h}$ , representing the attention of the classification token to all tokens except itself. We then randomly generate attention priors $\\tilde{\\mathbf{A}}_{l,h}$ , whose generation is detailed in the next part, and align $\\mathbf{A}_{l,h}^c$ with $\\tilde{\\mathbf{A}}_{l,h}$ by:", + "bbox": [ + 511, + 628, + 906, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}) = \\operatorname {M S E} \\left(\\mathbf {A} _ {l, h} ^ {c} - \\tilde {\\mathbf {A}} _ {l, h}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 794, + 906, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MSE represents the mean squared error. For Swin models that do not use a classification token, we substitute $\\mathbf{A}_{l,h}^{c}$ in Eq.6 with the average attention map of all tokens [10]. As noted in [17], ViTs initially focus on all regions to capture low-level information in shallow blocks", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12482", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/47a751ddad2314c95508a2a12f85e247727bb9004ac501eed58fe564b277eaf4.jpg", + "image_caption": [ + "Figure 4. Examples of generated attention priors." + ], + "image_footnote": [], + "bbox": [ + 153, + 88, + 217, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c735d92212b727d406c580060dcabab8f5fe18157239ef6f24724c4a4698a16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 222, + 88, + 285, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9511c52c1df3beaa3c5e411286edf72afd6a1e11835c3eaf0e2f6b22a8878de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 88, + 354, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d1652aa491f07440ea2ee43d6f9c93ab483199aba85c303bd36c784fca49c7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 423, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and gradually shift their focus toward semantic regions in deeper blocks to extract high-level semantic information. Leveraging this property, we selectively apply $\\mathcal{L}_{l,h}^{\\mathrm{APA}}$ to deeper blocks, progressively aligning attention towards semantically relevant areas. The total APA loss is computed as a depth-weighted sum of the individual Eq. 6 across these deeper blocks:", + "bbox": [ + 89, + 179, + 483, + 285 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = S} ^ {L} \\sum_ {h = 1} ^ {H} \\frac {l}{L} \\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 294, + 480, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ is a pre-given hyper-parameter denoting the start of deep blocks and are experimentally set $S = \\frac{L}{2}$ .", + "bbox": [ + 89, + 344, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention Priors Generation. To generate attention priors $\\tilde{\\mathbf{A}}_{l,h}$ , Gaussian Mixture Models (GMMs) are employed as it is the most commonly used distribution with high flexibility. ViTs utilize different attention heads to capture diverse patterns and learn varied, informative representations [8]. Thus, we use distinct GMMs for each head. Note that the goal here is not to replicate real attention maps precisely, but to generate simulated structure attention priors to guide the learning of synthetic images.", + "bbox": [ + 89, + 375, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, we first initialize an all zero matrix $\\tilde{\\mathbf{P}}\\in$ $\\mathbb{R}^{H\\times W}$ , where $H = W = \\sqrt{N - 1}$ for DeiT, $H =$ $W = \\sqrt{N}$ for Swin. For example, for DeiT-S, the $H =$ $W = \\sqrt{196} = 14$ . Then, we generate $k$ two-dimensional Gaussian distributions, where $k$ is randomly sampled from $1\\sim K_{APA}$ and $K_{APA}$ is set to 5 in all experiments. Each Gaussian has its mean and covariance3. Consequently, the matrix element at the $i$ -th row and $j$ -th column, $\\tilde{\\mathbf{P}} [i,j]$ , is determined by:", + "bbox": [ + 89, + 510, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {P}} [ i, j ] = \\max _ {m = 1, \\dots , k} \\mathbf {G} ^ {m} [ i, j ]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 654, + 482, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, $\\tilde{\\mathbf{P}}$ is normalized by:", + "bbox": [ + 89, + 691, + 267, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {P}} _ {n} = \\frac {\\tilde {\\mathbf {P}}}{\\sum \\tilde {\\mathbf {P}}} \\cdot (1 - x). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 715, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, for DeiT that incorporates the classification token, $x$ is randomly sampled from a uniform distribution $U(0,1)$ , representing the proportion of the attention score that the classification token allocates to itself. For Swin, which does not use a classification token, $x$ is set to 0. Finally, $\\tilde{\\mathbf{P}}_n$ is flatten to match the dimensionality:", + "bbox": [ + 89, + 760, + 483, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {A}} _ {l, h} = \\operatorname {f l a t t e n} \\left(\\tilde {\\mathbf {P}} _ {n}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 859, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 displays examples of generated attention priors.", + "bbox": [ + 531, + 90, + 888, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. APA prevents attention disorder, ensuring that synthetic images exhibit more coherent and natural attention patterns, as demonstrated in Fig. 3. As a result, APA selectively enhances the responses in certain regions, effectively emphasizing semantic-discriminative regions within synthetic images and thus prompting the discriminative features. Although simple, APA enables synthetic images to align better with the real semantics, as validated by both visual and quantitative evaluations. As shown in Fig. 1a, compared to PSAQ-ViT, features obtained after applying the APA loss are more closely aligned with real images, indicating better semantic alignment. The quantitative results in Tab. 1 further support that APA achieves superior semantic alignment. For example, in Class 1, the intra-class similarity for PSAQ-ViT is 0.44, whereas APA achieves a higher similarity of 0.64.", + "bbox": [ + 511, + 107, + 906, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Multi-Semantic Reinforcement", + "text_level": 1, + "bbox": [ + 511, + 356, + 774, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current DFQ methods for ViTs [47, 48, 62] optimize images through global optimization, treating the entire image as a single semantic unit. However, this is affected by low-rank structural regularity [34], where adjacent pixels exhibit strong similarity, leading to dull regions with redundant or non-semantic content for model learning [37]. This issue is further exacerbated by the tokenization mechanism in ViTs, as processing images in fixed-size patches make dull regions increase at the patch level [19, 70]. Consequently, as shown in Fig. 1b, synthetic images generated by existing methods exhibit large dull regions, resulting in semantic inadequacy [37]. In response, we propose Multi-Semantic Reinforcement (MSR), which applies localized patch optimization to enhance semantic richness by learning local patches with distinct semantics.", + "bbox": [ + 511, + 375, + 906, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for a synthetic image $\\tilde{I}$ , instead of feeding only the entire image, we also feed its patches and optimize them individually. Initially, we select $m$ nonoverlapping patches, where $m$ is chosen randomly from the set $\\{1,2,\\dots,K_{MSR}\\}$ , with $K_{MSR}$ set to 4 in all experiments. These $m$ patches are then cropped and resized to match the model's input dimensions:", + "bbox": [ + 511, + 601, + 905, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\tilde {\\boldsymbol {I}} _ {M S R} ^ {i} \\} _ {i = 1, \\dots , m} = \\operatorname {r e s i z e} \\left(\\operatorname {c r o p} _ {m} (\\tilde {\\boldsymbol {I}})\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 718, + 903, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{crop}_m(\\cdot)$ crops $m$ non-overlapping patches from input, and $\\text{resize}(\\cdot)$ is the resize function. Each patch, denoted as $\\tilde{I}_{MSR}^i$ , is treated as a new image with an assigned semantic target $c^i$ . Note that the gradient is backpropagated only to update the corresponding patch in the original image, leaving the rest of the image unaffected.", + "bbox": [ + 511, + 750, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Softlabel Learning. The one-hot loss (Eq. 3) only learns the semantics of the target class, making it unsuitable for $\\tilde{\\pmb{I}}$ under MSR, as its patches $\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}$ contain distinct semantics. In response, we propose Softlabel Learning", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3The pseudo code is detailed in appendix.", + "bbox": [ + 107, + 886, + 331, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(SL), which applies multiple semantic targets to accommodate the learning of images augmented by MSR. Specifically, we first sample $Z \\in \\mathbb{R}^{C} \\sim U(0,1)$ , then modify its values by:", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} Z [ c ^ {i} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}} _ {M S R} ^ {i}, \\\\ Z [ c ^ {1}, \\ldots , c ^ {m} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}}, \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 161, + 428, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $U(\\epsilon_1, \\epsilon_2)$ denotes the uniform distribution over the interval $[\\epsilon_1, \\epsilon_2]$ , $m$ is the number of patches determined in MSR, and $\\epsilon_1$ and $\\epsilon_2$ control the softness, both empirically set consistently to 5 and 10 in all experiments. The soft target is defined as $T_s = \\mathrm{softmax}(Z)$ , and SL loss is:", + "bbox": [ + 89, + 210, + 483, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\mathrm {S L}} \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}\\right) = S C E \\left(F \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}, T _ {s}\\right), \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 295, + 480, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $SCE(\\cdot, \\cdot)$ is soft cross entropy and $F(\\cdot)$ returns the predicted probability for its input. SL facilitates smooth learning between across semantic targets, ensuring that MSR-enhanced images receive consistent supervision rather than conflicting supervision.", + "bbox": [ + 89, + 321, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. By leveraging localized patch optimization, MSR ensures that each patch contributes unique semantics, forcing synthetic images to capture diverse features rather than being dominated by large homogeneous dull regions. As demonstrated by the richer and more diverse content and textures in Fig. 1b, MSR effectively reduces dull regions and enhances semantic richness in synthetic images. Moreover, MSR transforms synthetic images into composites of multiple semantic objects rather than a single unit, thereby providing more distinct semantic samples, i.e., $\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}$ , for model training. Unlike traditional cropping used in data augmentation, which aims to improve classification robustness by training with cropped patches labeled with the original class, MSR aims for semantic richness within synthetic images, ultimately enabling accurate data-free quantization.", + "bbox": [ + 89, + 398, + 483, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Overall Pipeline", + "text_level": 1, + "bbox": [ + 89, + 648, + 251, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall pipeline consists of two stages: data synthesis and quantized network learning. The first stage uses the proposed SARDFQ to produce synthetic images. The second stage fine-tunes the quantized model using the generated synthetic images.", + "bbox": [ + 89, + 671, + 483, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.1. Data Synthesis", + "text_level": 1, + "bbox": [ + 89, + 753, + 241, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the data synthesis stage, we combine the proposed APA loss of Eq. 7, SL loss of Eq. 12, and TV loss of Eq. 4 to formulate the objective function as follows:", + "bbox": [ + 89, + 772, + 483, + 818 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G} (\\tilde {\\boldsymbol {I}}) = \\alpha_ {1} \\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) + \\mathcal {L} ^ {\\mathrm {S L}} (\\tilde {\\boldsymbol {I}}) + 0. 0 5 \\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 827, + 483, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ is hyperparameters and is determined by gridsearch. Note that the weight of TV loss is fixed to 0.05, following [47], to avoid a cumbersome hyperparameter search.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.2. Quantized Network Learning", + "text_level": 1, + "bbox": [ + 513, + 90, + 766, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently DFQ methods have introduced the PTQ methods in learning quantized models due to their advantages of speed, memory efficiency, and performance [35, 62]. Thus, following the success of [42, 71], we fine-tune the quantized network block-wisely. Specifically, denote $\\mathbf{X}_l$ as the outputs of the $l$ -th block of the full-precision model, and $\\bar{\\mathbf{X}}_l$ represent outputs of the quantized counterpart. The reconstruction loss is defined as:", + "bbox": [ + 511, + 109, + 906, + 231 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l} = \\left\\| \\mathbf {X} _ {l} - \\bar {\\mathbf {X}} _ {l} \\right\\| _ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 239, + 906, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_l$ is only backward to update weights within $l$ -th block. Note that for a fair comparison, all compared methods adopt the same quantized network learning stage.", + "bbox": [ + 511, + 266, + 906, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 324, + 638, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 349, + 730, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models and Tasks. We evaluate the performance of SARDFQ by test quantized ViT-S/B [20], DeiT-T/S/B [68], and Swin-S/B [54] on the classification task using ImageNet [63]. The pre-trained models are downloaded from the timm library. In appendix, we further provide results on detection and segmentation tasks.", + "bbox": [ + 511, + 371, + 905, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We compare our SARDFQ against Gaussian noise, real images, and previous methods including SMI [33] and PSAQ-ViT [47] and its subsequent version, PSAQ-ViT V2 [48]. For a fair comparison, we generate synthetic images using their methods and apply our quantized network learning strategy. We use the official code for SMI and PSAQ-ViT to reproduce their images, while PSAQ-ViT V2 is re-implemented by us, as no official code is available.", + "bbox": [ + 511, + 462, + 905, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental settings. All experiments were conducted using the PyTorch framework [60] on a single NVIDIA 3090 GPU. In the data synthesis stage, synthetic images were initialized with standard Gaussian noise, generating 32 images in total. The Adam optimizer [38] with $\\beta_{1} = 0.5$ , $\\beta_{2} = 0.9$ was used, with learning rates of 0.25 for Swin and 0.2 for others, and a total of 1,000 iterations. For all models, $K_{APA}$ , $K_{MSR}$ , $\\epsilon_{1}$ , and $\\epsilon_{2}$ were set to 5, 4, 5, and 10, respectively, based on a search with W4A4 DeiT-S. The value of $\\alpha_{1}$ was determined by grid search for each model: 1e5 for DeiT-T/S, 1e4 for DeiT-B, 100 for ViT-B and Swin-B, 10 for Swin-S, and 1 for ViT-S. Although further hyperparameter search may improve performance, the current settings already yield superior results. In the quantized network learning stage, following [85], the Adam optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ was used, with weight decay set to 0 and an initial learning rate of 4e-5, adjusted", + "bbox": [ + 511, + 598, + 906, + 854 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4In appendix, we present practical efficiency results and more performance comparisons with more methods such as CLAMP-ViT [62] on W8/A8 and W4/A8 settings.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12484", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/33299b202b35ff497d1ff72b39a86307dea4795633c4b17dd51632f5f95dd691.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelW/ARealGaussian noisePSAQ-ViT [47]PSAQ-ViT V2 [48]SMI [33]SARDFQ (Ours)
ViT-S (81.39)4/466.576.0247.2441.5324.3329.4150.32
5/576.6936.7771.5968.4161.3365.1974.31
6/679.4661.2077.2074.7672.9572.4678.40
ViT-B (84.54)4/468.160.1536.3226.3235.2719.6751.84
5/579.214.1668.4867.9567.5357.1370.70
6/681.8955.1876.6571.8776.3369.8279.16
DeiT-T (72.21)4/456.6017.4347.7530.2030.1413.1852.06
5/567.0943.4964.1055.1656.4439.3566.41
6/669.8156.2368.3762.7764.0344.3969.73
DeiT-S (79.85)4/468.4620.8958.2845.5342.7711.7162.29
5/575.0641.0671.9063.1462.8829.1374.06
6/677.8765.6375.8568.8571.6537.6977.31
DeiT-B (81.85)4/477.0747.2071.7566.4365.3359.0472.17
5/579.8665.4678.4576.7776.7475.3378.72
6/680.9062.7980.0079.2278.8177.6680.15
Swin-S (83.20)4/478.1231.9273.1965.5565.8574.74
5/580.5152.1078.1574.3775.4179.56
6/680.6065.6679.7478.5078.2580.56
Swin-B (85.27)4/478.8030.1471.8467.4265.2376.42
5/582.5135.2878.5077.2075.2580.82
6/682.6467.3782.0081.4180.3083.03
", + "bbox": [ + 102, + 89, + 893, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantization results on ImageNet dataset, with top-1 accuracy $(\\%)$ reported. The performance of the full-precision model is listed below the model name. \"W/A\" denotes the bit-width of weights/activations. \"Real\" refers to using real images. For SMI [33], we provide the performance of using dense (normal-sized numbers) and sparse (smaller-sized numbers) synthetic images, respectively. Note that for Swin models, we do not provide the results for sparse synthetic images as the sparse generation method of SMI is infeasible.", + "bbox": [ + 89, + 484, + 906, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "via cosine decay for 100 iterations. A channel-wise quantizer was used for weights, and a layer-wise quantizer for activations, with all matrix multiplications in ViTs quantized [47-49].", + "bbox": [ + 89, + 556, + 482, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantization Results", + "text_level": 1, + "bbox": [ + 89, + 633, + 289, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantization results are presented in Tab.2. Our SARDFQ demonstrates consistent improvements across various quantization bit-width configurations, particularly with low bit-width settings. Specifically, for ViT-S, SARDFQ improves the performance by $3.08\\%$ in the W4/A4 setting, $2.72\\%$ in the W5/A5 setting, and $1.20\\%$ in the W6/A6 setting. For ViT-B, SARDFQ achieves performance gains of $15.52\\%$ in the W4/A4 setting, $2.22\\%$ in the W5/A5 setting, and $2.51\\%$ in the W6/A6 setting. Results on DeiT also demonstrate the effectiveness of SARDFQ. For example, on DeiT-T, SARDFQ shows a marked improvement by increasing top-1 accuracy by $4.31\\%$ in the W4/A4 setting, $2.31\\%$ in the W5/A5 setting, and $1.36\\%$ in the W6/A6 setting. For DeiT-S, SARDFQ enhances top-1 accuracy by $4.01\\%$ in the W4/A4 setting, $2.16\\%$ in the W5/A5 setting, and $1.46\\%$ in the W6/A6 setting. The quan", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tization results of Swin-S/B also affirm the superiority of our SARDFQ in enhancing model accuracy under different quantization configurations. In particular, for Swin-S, the proposed SARDFQ increases the accuracy by $1.55\\%$ for the W4/A4 setting, $1.41\\%$ for the W5/A5 setting, and $0.82\\%$ for the W6/A6 setting, respectively. When it comes to Swin-B, the proposed SARDFQ increases the accuracy by $4.58\\%$ for the W4/A4 setting, $2.32\\%$ for the W5/A5 setting, and $1.03\\%$ for the W6/A6 setting, respectively.", + "bbox": [ + 511, + 556, + 906, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 708, + 666, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation studies are conducted on the W4A4 DeiT-S.", + "bbox": [ + 511, + 732, + 885, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of APA, MSR, and SL. We analyze the effectiveness of the proposed APA (Sec.,3.3.1), MSR (Sec.,3.3.2), and SL (Eq.,12) in Tab.,3. Adding APA and SL individually to the baseline increases accuracy. Notably, APA boosts performance from $51.73\\%$ to $60.26\\%$ , confirming its effectiveness in aligning semantics (Sec.,3.3.1). Applying MSR alone slightly decreases accuracy from $51.73\\%$ to $50.75\\%$ , indicating that one-hot loss is unsuitable for MSR-augmented synthetic images. However, when both MSR and SL are applied, accuracy rises to $56.08\\%$ , sug-", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12485", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/206a27669ba58a5391cf71ee5a1676d1fde71dbe5309b8c26d17b7bd17575392.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
APAMSRSLAcc. (%)
Baseline51.73
✓60.26
✓50.75
✓52.02
✓✓61.58
✓✓60.51
✓✓56.08
✓✓✓62.29
", + "bbox": [ + 169, + 88, + 406, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gesting SL is more compatible with MSR than one-hot loss. Combining APA with either MSR or SL further improves performance. For example, APA and SL together yield an accuracy of $60.51\\%$ , and when all three strategies are used, the best performance of $62.29\\%$ is achieved.", + "bbox": [ + 89, + 321, + 482, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of priors distribution. Tab. 4a showcases the results of using other distributions to formulate the attention priors. The unevenly distributed GMM and Laplace present comparable performance of $62.29\\%$ and $62.16\\%$ , respectively. Moreover, GMM provides a similar performance to the real's, indicating it performs well in imitating the patterns of real images.", + "bbox": [ + 89, + 397, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9648db1c30d9fa309266ebee1ec85a0f5984e694d58af35381fa07d3dce66ed.jpg", + "image_caption": [ + "(a)", + "Figure 5. Effect of varying (a) $\\alpha_{1}$ and (b) $K_{MSR}$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 515, + 282, + 631 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e1333dc94a67c2dbdb0700a59515b6c89c5e3a10c0ae3c609fbbe5e2a6d3f60.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 282, + 515, + 478, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of $\\alpha_{1}, K_{APA}$ , and $K_{MSR}$ . The $\\alpha_{1}$ from Eq. 13 balance the importance of the proposed APA loss during the update of the synthetic images. Fig. 5a demonstrates that the optimal performance is achieved when $\\alpha_{1} = 1e5$ . Incrementally increasing $\\alpha_{1}$ improves performance up to $62.29\\%$ at $\\alpha_{1} = 1e5$ . However, further increases in $\\alpha_{1}$ subsequently degrade performance. The $K_{APA}$ in APA is the upper limit on the number of the Gaussian distributions used for priors generation. Tab. 4b displays the ablation study for different values of $K_{APA}$ . The best accuracy is achieved when $K_{APA} = 5$ . The $K_{MSR}$ in MSR is the upper limit on the number of patches. Fig. 5b demonstrates that when $K_{MSR} = 4$ , the optimal performance is achieved. Using $K_{MSR}$ larger than 4 will hurt the accuracy. We consider this", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3e3d62c782164115c35624252c98cb9d7fdd5faa0e95401c20fe58b13586962c.jpg", + "table_caption": [ + "Table 3. Influence of the proposed APA, MSR, and SL on accuracy. The baseline adopts the one-hot loss. Applying APA, MSR, and SL yields SARDFQ." + ], + "table_footnote": [], + "table_body": "
Priors DistributionTop-1
GMM62.29
Laplace62.16
Real63.19
", + "bbox": [ + 514, + 103, + 720, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29b607cecea06dacbd324c39babe14929a287cf17e47dc3308ce9afe27f79bbd.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
KAPATop-1
161.13
361.52
562.29
761.53
961.05
", + "bbox": [ + 751, + 88, + 880, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/428c915ab844ae292e8cac279f6eac909d561e07aa4b86888bd7f7747fb61b2f.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [], + "table_body": "
STop-1
061.96
L/262.29
", + "bbox": [ + 563, + 217, + 676, + 280 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8bb3b141338774aeae4a45994c2d336555dc3fc8f80865f42494534157eafc84.jpg", + "table_caption": [ + "(c)" + ], + "table_footnote": [], + "table_body": "
w. l/LTop-1
✓62.29
×61.32
", + "bbox": [ + 740, + 217, + 856, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(d)", + "bbox": [ + 790, + 282, + 808, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Effect of varying (a) priors types; (b) $K_{APA}$ ; (c) $S$ ; (d) $\\frac{l}{L}$ .", + "bbox": [ + 511, + 306, + 903, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to limited patch resolution if using a too large $K_{MSR}$ .", + "bbox": [ + 511, + 352, + 898, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of APA loss. Here, we conduct the ablation study by considering the $S$ and scale $\\frac{l}{L}$ in Eq.7. Tab.4c presents the effect of varying $S$ in Eq.7. If applying APA loss to all blocks $(S = 0)$ , the top-1 accuracy decreases to $61.96\\%$ . From Tab.4d, it can be seen that absorbing the scale $\\frac{l}{L}$ in Eq.7 presents $0.97\\%$ performance gains.", + "bbox": [ + 511, + 367, + 905, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 472, + 635, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further discuss some limitations of the proposed SARDFQ, which will guide future research directions. First, although SARDFQ shows substantial performance improvement, a performance gap between SARDFQ and real data remains challenging, highlighting the need for a stronger semantics alignment and reinforcement method. Second, SARDFQ currently lacks a theoretical foundation. Future work could establish a theoretical framework for SARDFQ, particularly in understanding how APA and MSR influence synthetic images in a formalized manner.", + "bbox": [ + 511, + 497, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we investigate the DFQ method for ViTs. We first identify that synthetic images generated by existing methods suffer from semantic distortion and inadequacy issues, and propose SARDFQ to address these issues. To mitigate semantic distortion, SARDFQ introduces APA, which guides synthetic images to align with randomly generated structural attention patterns. To tackle semantic inadequacy, SARDFQ incorporates MSR. MSR optimizes different regions of synthetic images with unique semantics, thereby enhancing overall semantic richness. Moreover, SARDFQ employs SL, which adopts multiple semantic targets to ensure seamless learning of images augmented by MSR. Extensive experiments on various ViT models and tasks validate the effectiveness of SARDFQ.", + "bbox": [ + 509, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12486", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (Np. 624B2119, No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001).", + "bbox": [ + 89, + 113, + 485, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 6836-6846, 2021. 1, 2", + "[2] Jianhong Bai, Yuchen Yang, Huanpeng Chu, Hualiang Wang, Zuozhu Liu, Ruizhe Chen, Xiaoxuan He, Lianrui Mu, Chengfei Cai, and Haoji Hu. Robustness-guided image synthesis for data-free quantization. arXiv preprint arXiv:2310.03661, 2023. 3", + "[3] Ron Banner, Yury Nahshan, Daniel Soudry, et al. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems (NeurIPS), pages 7950-7958, 2019. 2", + "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(8):1798-1828, 2013. 4", + "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 4", + "[6] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13169-13178, 2020. 2, 3", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 1, 2", + "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 397-406, 2021. 5", + "[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12299-12310, 2021. 2", + "[10] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general" + ], + "bbox": [ + 93, + 306, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "coarse-to-fine method for vision transformer. In AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 4", + "[11] Mengzhao Chen, Wenqi Shao, Peng Xu, Mingbao Lin, Kaipeng Zhang, Fei Chao, Rongrong Ji, Yu Qiao, and Ping Luo. Diffrate: Differentiable compression rate for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17164-17174, 2023. 2", + "[12] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2021. 2", + "[13] Xinrui Chen, Yizhi Wang, Renao Yan, Yiqing Liu, Tian Guan, and Yonghong He. Texq: Zero-shot network quantization with texture feature distribution calibration. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4, 1", + "[14] Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. In Advances in Neural Information Processing Systems (NeurIPS), pages 14835-14847, 2021. 2, 4, 1", + "[15] Kanghyun Choi, Hye Yoon Lee, Deokki Hong, Joonsang Yu, Noseong Park, Youngsok Kim, and Jinho Lee. It's all in the teacher: Zero-shot quantization brought closer to the teacher. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8311-8321, 2022. 3", + "[16] Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, and Jinho Lee. Mimiq: Low-bit data-free quantization of vision transformers. arXiv preprint arXiv:2407.20021, 2024. 2, 3", + "[17] Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-attention and convolutional layers. In International Conference on Learning Representations (ICLR), 2020. 4", + "[18] Yifu Ding, Haotong Qin, Qinghua Yan, Zhenhua Chai, Junjie Liu, Xiaolin Wei, and Xianglong Liu. Towards accurate posttraining quantization for vision transformer. In 30th ACM International Conference on Multimedia (ACMMM), pages 5380-5388, 2022. 3", + "[19] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International conference on machine learning (ICML), pages 2793-2803. PMLR, 2021. 5", + "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2, 4, 6", + "[21] Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero-shot quantization. In International Conference on Machine Learning (ICML), 2024. 2", + "[22] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learned step size quantization. In International Conference on Learning Representations (ICLR), 2020. 2", + "[23] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), 2021. 2", + "[24] Natalia Frumkin, Dibakar Gope, and Diana Marculescu. Jumping through local minima: Quantization in the loss landscape of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16978-16988, 2023. 3", + "[25] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4852–4861, 2019. 2", + "[26] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), pages 2672–2680, 2014. 3", + "[27] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quantization via diagonal hessian approximation. In The Eleventh International Conference on Learning Representations (ICLR), 2022. 2, 3", + "[28] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[29] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):87-110, 2022. 1", + "[30] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2", + "[31] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 773-783, 2023. 4", + "[32] Zejiang Hou and Sun-Yuan Kung. Multi-dimensional vision transformer compression via dependency guided gaussian process search. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3669-3678, 2022. 2", + "[33] Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Lei Li, Chun Yuan, and Dacheng Tao. Sparse model inversion: Efficient inversion of vision transformers for data-free applications. In International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 6, 7, 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Jinggang Huang and David Mumford. Statistics of natural images and models. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 541-547. IEEE, 1999. 5", + "[35] Yongkweon Jeon, Chungman Lee, and Ho-young Kim. Genie: show me the data for quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12073, 2023. 3, 6", + "[36] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM Computing Surveys (CSUR), 2021. 1", + "[37] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning (ICML), pages 11102-11118. PMLR, 2022. 4, 5", + "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 6", + "[39] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 2", + "[40] Huantong Li, Xiangmiao Wu, Fanbing Lv, Daihai Liao, Thomas H Li, Yonggang Zhang, Bo Han, and Mingkui Tan. Hard sample matters a lot in zero-shot quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24417-24426, 2023. 3", + "[41] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In International Conference on Learning Representations (ICLR), 2020. 2", + "[42] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations (ICLR), 2021. 2, 4, 6", + "[43] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 34451-34463, 2022. 2", + "[44] Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Stableq: Enhancing data-scarce quantization with text-to-image data. arXiv preprint arXiv:2312.05272, 2023. 3", + "[45] Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, and Priyadarshini Panda. Genq: Quantization in low data regimes with generative synthetic data. In European Conference on Computer Vision (ECCV), pages 216-235. Springer, 2024. 3", + "[46] Zhikai Li and Qingyi Gu. I-vit: Integer-only quantization for efficient vision transformer inference. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17065-17075, 2023. 2, 3", + "[47] Zhikai Li, Liping Ma, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12488", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision transformers. In European Conference on Computer Vision (ECCV), pages 154-170. Springer, 2022. 1, 2, 3, 4, 5, 6, 7", + "[48] Zhikai Li, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2023. 2, 3, 5, 6, 7, 1", + "[49] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17227-17236, 2023. 3, 7", + "[50] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 1833-1844, 2021. 2", + "[51] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), pages 1173–1179, 2022. 3", + "[52] Shih-Yang Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning (ICML), pages 21813-21824, 2023. 2", + "[53] Yijiang Liu, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20321-20330, 2023. 3", + "[54] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF international conference on computer vision (ICCV), pages 10012-10022, 2021. 2, 6", + "[55] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 28092-28103, 2021. 2, 3", + "[56] Yiwei Ma, Jiayi Ji, Xiaoshuai Sun, Yiyi Zhou, and Rongrong Ji. Towards local visual modeling for image captioning. Pattern Recognition, 138:109420, 2023. 1", + "[57] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations (ICLR), 2022. 2", + "[58] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 23296-23308, 2021. 4", + "[59] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 3163-3172, 2021. 2" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), pages 8026-8037, 2019. 6", + "[61] Biao Qian, Yang Wang, Richang Hong, and Meng Wang. Adaptive data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7960-7968, 2023. 3", + "[62] Akshit Ramachandran, Souvik Kundu, and Tushar Krishna. Clamp-vit: contrastive data-free learning for adaptive posttraining quantization of vits. In European Conference on Computer Vision (ECCV). Springer, 2024. 2, 3, 4, 5, 6, 1", + "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115:211-252, 2015. 6", + "[64] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transformers in medical imaging: A survey. Medical Image Analysis, page 102802, 2023. 2", + "[65] Huixin Sun, Runqi Wang, Yanjing Li, Xianbin Cao, Xiaolong Jiang, Yao Hu, and Baochang Zhang. P4q: Learning to prompt for quantization in visual-language models. arXiv preprint arXiv:2409.17634, 2024. 2", + "[66] Yehui Tang, Yunhe Wang, Yixing Xu, Yiping Deng, Chao Xu, Dacheng Tao, and Chang Xu. Manifold regularized dynamic network pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5028, 2021. 2", + "[67] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12165-12174, 2022. 2", + "[68] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357. PMLR, 2021. 1, 2, 6", + "[69] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9:2579-2605, 2008. 1", + "[70] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 5", + "[71] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. In International Conference on Learning Representations (ICLR), 2022. 6", + "[72] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In IEEE/CVF In" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12489", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Computer Vision (ICCV), pages 10033-10041, 2021. 2", + "[73] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision (ECCV), pages 68-85, 2022. 2", + "[74] Zhiqiang Shen Xijie Huang and Kwang-Ting Cheng. Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331, 2023. 2", + "[75] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17, 2020. 3", + "[76] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17. Springer, 2020. 2", + "[77] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transffer via deep-inversion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8715-8724, 2020. 2, 3, 1", + "[78] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015. 4", + "[79] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 818-833, 2014. 4", + "[80] Jinnian Zhang, Houwen Peng, Kan Wu, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Minivit: Compressing vision transformers with weight multiplexing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12145-12154, 2022. 2", + "[81] Xiangguo Zhang, Haotong Qin, Yifu Ding, Ruihao Gong, Qinghua Yan, Renshuai Tao, Yuhang Li, Fengwei Yu, and Xianglong Liu. Diversifying sample generation for accurate data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15658-15667, 2021. 2, 3", + "[82] Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, and Yunhe Wang. Less is more: Focus attention for efficient detr. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6674-6683, 2023. 2", + "[83] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 6881-6890, 2021. 2", + "[84] Yunshan Zhong, Mingbao Lin, Gongrui Nan, Jianzhuang Liu, Baochang Zhang, Yonghong Tian, and Rongrong Ji. Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12339-12348, 2022. 2, 4, 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&S-vit: An inclusive & stable method for pushing the limit of post-training vits quantization. arXiv preprint arXiv:2311.10126, 2023. 6", + "[86] Yunshan Zhong, Jiawei Hu, You Huang, Yuxin Zhang, and Rongrong Ji. Erq: Error reduction for post-training quantization of vision transformers. In International Conference on Machine Learning (ICML), 2024. 3" + ], + "bbox": [ + 516, + 90, + 905, + 204 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12490", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_model.json b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d8b288e92a03ebcc97f45f87b59d936ea7414f2 --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_model.json @@ -0,0 +1,2974 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.13, + 0.89, + 0.175 + ], + "angle": 0, + "content": "Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.203, + 0.834, + 0.24 + ], + "angle": 0, + "content": "Yunshan Zhong\\(^{1,2}\\), Yuyao Zhou\\(^{2}\\), Yuxin Zhang\\(^{2}\\), Wanchen Sui\\(^{3}\\), Shen Li\\(^{3}\\), Yong Li\\(^{3}\\), Fei Chao\\(^{2}\\), Rongrong Ji\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.239, + 0.714, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Artificial Intelligence, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.256, + 0.783, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)MAC Lab, School of Informatics, Xiamen University \\(^{3}\\)Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.278, + 0.836, + 0.312 + ], + "angle": 0, + "content": "viper.zhong@gmail.com, {yuyaozhou, yuxinzhang}@stu.xmu.edu.cn \n{wanchen.swc, litan.ps, jiufeng.ly}@alibaba-inc.com, {fchao, rrji}@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.485, + 0.773 + ], + "angle": 0, + "content": "Data-free quantization (DFQ) enables model quantization without accessing real data, addressing concerns regarding data security and privacy. With the growing adoption of Vision Transformers (ViTs), DFQ for ViTs has garnered significant attention. However, existing DFQ methods exhibit two limitations: (1) semantic distortion, where the semantics of synthetic images deviate substantially from those of real images, and (2) semantic inadequacy, where synthetic images contain extensive regions with limited content and oversimplified textures, leading to suboptimal quantization performance. To address these limitations, we propose SARDFQ, a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs. To address semantic distortion, SARDFQ incorporates Attention Priors Alignment (APA), which optimizes synthetic images to follow randomly generated structure attention priors. To mitigate semantic inadequacy, SARDFQ introduces Multi-Semantic Reinforcement (MSR), leveraging localized patch optimization to enhance semantic richness across synthetic images. Furthermore, SARDFQ employs Soft-Label Learning (SL), wherein multiple semantic targets are adapted to facilitate the learning of multi-semantic images augmented by MSR. Extensive experiments demonstrate the effectiveness of SARDFQ, significantly surpassing existing methods. For example, SARDFQ improves top-1 accuracy on ImageNet by \\(15.52\\%\\) for W4A4 ViT-B1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.804, + 0.222, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.831, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) [20, 29, 36] have demonstrated remarkable success across various computer vision tasks [1," + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.345, + 0.847, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.617 + ], + "angle": 0, + "content": "(a) The t-SNE[69] visualization of the penultimate-layer features (extracted by DeiT-S) of synthetic images. Each marker (circle, star, triangle) represents a distinct category. The red dashed circles highlight the features extracted from our APA and real images. Notably, the features produced by PSAQ-ViT[47] exhibit substantial deviation from those of real images, indicating semantic distortion. In contrast, our APA yields features that more closely align with those of real images, suggesting improved semantics alignment." + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.618, + 0.666, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.746, + 0.635, + 0.756 + ], + "angle": 0, + "content": "PSAQ" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.618, + 0.755, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.745, + 0.736, + 0.756 + ], + "angle": 0, + "content": "PSAQ V2" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.618, + 0.846, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.745, + 0.844, + 0.757 + ], + "angle": 0, + "content": "SARDFQ (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.758, + 0.907, + 0.809 + ], + "angle": 0, + "content": "(b) The images of PSAQ-ViT and PSAQ-ViT V2 exhibit numerous dull regions with limited content and simplified textures, reflecting semantic inadequacy. In comparison, our SARDFQ generates images with greater diversity in both content and texture, demonstrating enhanced semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.82, + 0.907, + 0.848 + ], + "angle": 0, + "content": "Figure 1. Illustration of (a) semantic distortion and (b) semantic inadequacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "7, 56, 68]. However, their high computational cost and substantial memory footprint hinder deployment in resource-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.335, + 0.888 + ], + "angle": 0, + "content": "*Corresponding Author: rrji@xmu.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.449, + 0.9 + ], + "angle": 0, + "content": "1The code is at https://github.com/zysxmu/SARDFQ." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.449, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "constrained environments [11, 30, 32, 46, 55, 67, 82]. To address these limitations, quantization [39] has emerged as a promising solution, which reduces model complexity by converting full-precision weight and activations into low-bit representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Traditional quantization methods typically require access to the original training dataset, which raises data privacy and security concerns [6, 27, 77, 84]. As a result, data-free quantization (DFQ) has gained increasing attention, allowing quantization without the need for real data [14, 76, 81]. However, most existing DFQ methods are designed specifically for convolutional neural networks (CNNs) and are not directly applicable to vision transformers (ViTs). These methods generally rely on batch normalization statistics (BNS), which capture the distribution of real data, to synthesize in-distribution synthetic data [6, 76, 81, 84]. Yet, BNS is unavailable for ViTs, which use layer normalization (LN) to dynamically compute distribution statistics during inference [47]. Recently, several DFQ methods have been proposed for ViTs [16, 33, 47, 48, 62]. For example, PSAQ-ViT [47] introduces patch similarity entropy (PSE) loss to optimize Gaussian noise towards usable synthetic images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.65 + ], + "angle": 0, + "content": "Nevertheless, we observe that existing methods suffer from semantic distortion and inadequacy issues. As shown in Fig. 1a, features of synthetic images generated by PSAQ-ViT deviate significantly from those of real images. Tab. 1 shows that the cosine similarity between synthetic images from PSAQ-ViT and real images is notably low, also indicating significant distortion. These results highlight the issue of semantic distortion. Moreover, as shown in Fig. \\(1\\mathrm{b}^2\\) synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit many regions with limited content diversity and overly simplified textures. These low-quality dull regions are useless or even detrimental to model learning [33], highlighting the issue of semantic inadequacy. Consequently, quantized models trained on such low-quality images suffer from degraded performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Motivated by the above analysis, we propose a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs, termed SARDFQ. The overall framework is depicted in Fig.2. To address the semantic distortion issue, SARDFQ introduces Attention Priors Alignment (APA), where synthetic images are optimized to follow structured attention priors generated using Gaussian Mixture Models (GMMs). APA effectively aligns the semantics of synthetic images with real images, as validated by both visual and quantitative analyses. As shown in Fig. 1a, features of APA exhibit a closer alignment to those of real images, while quantitative results in Tab. 1 also confirm that the semantics of APA are more consistent to the real images, indicating enhanced semantics alignment. To address the semantic inadequacy issue, SARDFQ incorpo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "rates Multi-Semantic Reinforcement (MSR) and Softlabel Learning (SL). MSR utilizes localized patch optimization, which encourages different sub-patches of synthetic images to capture various semantics, reinforcing the rich semantics across images. SL applies multiple semantic targets to accommodate the learning of multi-semantic images augmented by MSR. As shown in Fig. 1b, synthetic images after applying MSR exhibit greater diversity in content and texture, providing reinforced semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Experimental results across various ViT models and tasks demonstrate that SARDFQ presents substantial performance improvements. For example, SARDFQ achieves a \\(15.52\\%\\) increase in top-1 accuracy on the ImageNet dataset for the W4A4 ViT-B model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.663, + 0.336 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.708, + 0.36 + ], + "angle": 0, + "content": "2.1. Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.701 + ], + "angle": 0, + "content": "The great success of transformers in the natural language processing field has driven widespread attempts in the computer vision community to apply them to vision tasks [12, 28, 72]. ViT [20] is the pioneer that builds a transformer-based model to handle images, boosting the performance on the image classification task. DeiT [68] introduces an efficient teacher-student training strategy where a distillation token is employed to distill knowledge from the teacher model to the student model. Swin Transformers [54] builds an efficient and effective hierarchical model by introducing a shifted window-based self-attention mechanism. Other than the image classification task, the applications of ViTs also have broadened considerably, manifesting groundbreaking performance in object detection [7], image segmentation [9, 83], low-level vision [50], video recognition [1, 59], and medical image processing [64], etc. Nevertheless, the impressive performance of ViTs relies on a high number of parameters and significant computational overhead, preventing deployment in resource-constrained environments. Several recent efforts design lightweight ViTs, such as MobileViT [57], MiniVit [80], and TinyViT [73]. However, the model complexity is still unsatisfactory [47]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.722, + 0.728 + ], + "angle": 0, + "content": "2.2. Network Quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Data-Driven Quantization. Model quantization reduces the complexity of neural networks by replacing full-precision weight and activation with the low-bit format. Data-driven quantization can be roughly divided into two categories: quantization-aware training (QAT) and posttraining quantization (PTQ). QAT is compute-heavy since it re-trains the quantized model with the full training data to retain performance [22, 25, 41, 43, 46, 52, 65, 74]. PTQ perform quantization with a tiny dataset and a reduced time overhead, harvesting widespread attention [3, 42]. The specific architecture of ViTs, such as LayerNorm and the self" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.345, + 0.901 + ], + "angle": 0, + "content": "2More results are presented in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.087, + 0.825, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.247, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 2. SARDFQ Framework overview: Attention Priors Alignment (APA) employs randomly generated attention priors to improve semantics alignment. Multi-Semantic Reinforcement (MSR) learns the different regions of synthetic images with various semantics to enhance overall semantic richness. Meanwhile, Softlabel Learning (SL) adopts multiple semantic targets to ensure consistent learning of multi-semantic images augmented by MSR." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.482, + 0.409 + ], + "angle": 0, + "content": "attention module, urges distinct PTQ methods compared to CNNs [18, 24, 46, 51, 53, 86]. For example, Liu et al. [55] develop a ranking loss to maintain the relative order of the self-attention activation. Unfortunately, both QAT and PTQ involve the original training data, causing concerns about data privacy and security issues in data-sensitive scenarios." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.411, + 0.485, + 0.683 + ], + "angle": 0, + "content": "Data-Free Quantization. DFQ quantizes models without accessing real data [2, 13-15, 27, 35, 40, 44, 45, 61]. Most previous DFQ methods focus on CNN, where the BNS can be adopted as the regularization term [6, 81]. However, BNS is infeasible for ViTs built on the LN. Recently, few efforts have been explored to accommodate ViTs [16, 33, 47, 48, 62]. PSAQ-ViT [47] introduces the first DFQ method for ViTs. They discover that Gaussian noise yields homogeneous patches, while the real image yields heterogeneous patches. Thus, patch similarity entropy (PSE) loss is proposed to optimize the Gaussian noise towards real-like images by making them showcase heterogeneous patches. Based on PSAQ-ViT, PSAQ-ViT V2 [48] further introduces an adversarial learning strategy [26]. [62] incorporates contrastive learning and proposes an iterative generation-quantization PTQ-based DFQ method. [33] proposes a sparse generation method to remove noisy and hallucination backgrounds in synthetic images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.699, + 0.182, + 0.715 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.232, + 0.74 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.216, + 0.763 + ], + "angle": 0, + "content": "3.1.1.Quantizers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.859 + ], + "angle": 0, + "content": "We employ the linear quantizer for all weights and activations, except for the attention scores, which use a log2 quantizer to handle highly non-negative and uneven values [24, 49, 51]. For the linear quantizer, given a full-precision input \\(\\mathbf{x}\\) and bit-width \\(b\\), the quantized value \\(\\mathbf{x}_q\\) and the de-quantized value \\(\\bar{\\mathbf{x}}\\) are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot \\left(\\mathbf {x} _ {q} - z\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.365 + ], + "angle": 0, + "content": "where \\(\\lfloor \\cdot \\rfloor\\) denotes rounding to the nearest integer, and clip limits the value to \\([0,2^{b} - 1]\\). Here, \\(\\Delta\\) and \\(z\\) are the scale factor and zero-point, respectively. For the log2 quantizer:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.375, + 0.907, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor - \\log_ {2} \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil , 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot 2 ^ {- \\mathbf {x} _ {q}}. \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.664, + 0.427 + ], + "angle": 0, + "content": "3.1.2. Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.505 + ], + "angle": 0, + "content": "DFQ methods parameterize synthetic images and optimize them toward real-like images with a pre-trained full-precision model \\( F \\). Given a image \\( \\tilde{I} \\) initialized from Gaussian noise, the one-hot loss [75] is introduced to learn label-related semantics:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.517, + 0.907, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {O H}} (\\tilde {\\boldsymbol {I}}) = C E (F (\\tilde {\\boldsymbol {I}}), c) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.593 + ], + "angle": 0, + "content": "where \\( CE(\\cdot, \\cdot) \\) represents the cross entropy, \\( c \\) is a random class label, and \\( F(\\cdot) \\) returns the predicted probability for image \\( \\tilde{I} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Moreover, total variance (TV) [77] loss is a smoothing regularization term to improve the image quality:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.635, + 0.907, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}) = \\iint | \\nabla \\tilde {\\boldsymbol {I}} (\\tau_ {1}, \\tau_ {2}) | d \\tau_ {1} d \\tau_ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.871, + 0.695 + ], + "angle": 0, + "content": "where \\(\\nabla \\tilde{I} (\\tau_1,\\tau_2)\\) denotes the gradient at \\(\\tilde{I}\\) at \\((\\tau_{1},\\tau_{2})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.821 + ], + "angle": 0, + "content": "To perform DFQ for ViTs, PSAQ-ViT [47] proposes patch similarity entropy (PSE) loss. It first compute patch similarity \\(\\Gamma_{l}[i,j] = \\frac{u_{i} \\cdot u_{j}}{||u_{i}||||u_{j}||}\\), where \\(u_{i}, u_{j}\\) are feature vectors of MHSA outputs in \\(l\\)-th block and \\(||\\cdot||\\) denotes the \\(l_{2}\\) norm. Then, it estimates the density function \\(\\hat{f}_{l}(x) = \\frac{1}{Mh}\\sum_{m=1}^{M}K\\left(\\frac{x - x_{m}}{h}\\right)\\), where \\(K(\\cdot)\\) is a normal kernel, \\(h\\) is the bandwidth, and \\(x_{m}\\) is the kernel center derived from \\(\\Gamma_{l}\\). Finally, the PSE loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.834, + 0.907, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {P S E}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = 1} ^ {L} \\int \\hat {f} _ {l} (x) \\log [ \\hat {f} _ {l} (x) ] d x, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.798, + 0.9 + ], + "angle": 0, + "content": "where \\(L\\) is the block number of the model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12481" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.176 + ], + "angle": 0, + "content": "
Method ClassesRealPSAQ-ViT [47]APA (Ours)
Class 10.680.440.64
Class 20.320.260.31
Class 30.410.310.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.187, + 0.485, + 0.271 + ], + "angle": 0, + "content": "Table 1. Average cosine similarity of three randomly selected classes. For real images, the similarity is measured within the class itself, while for PSAQ-ViT and APA, the similarity is measured between synthetic and real images of the same class. The results show that APA achieves higher similarity than PSE loss, indicating aligned semantics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.299, + 0.231, + 0.315 + ], + "angle": 0, + "content": "3.2. Observations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.398 + ], + "angle": 0, + "content": "Existing DFQ methods have made significant progress. However, through carefully analyzing the synthetic images, we reveal that these images suffer from issues of semantic distortion and semantic inadequacy, both of which hinder further advancement in the DFQ of ViTs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.642 + ], + "angle": 0, + "content": "The semantic distortion issue refers to the significant divergence between the semantics of synthetic images and real images. To demonstrate this, we visualize the features of synthetic and real images in Fig. 1a. Note that the penultimate feature typically is regarded to represent the semantics of the input [4, 58, 78, 79]. It is clear that the features of PSAQ-ViT diverge significantly from real images, suggesting that these images fail to capture the true semantic distribution of real data. Tab. 1 quantitatively measures the semantics using the average cosine similarity (ranging from \\(-1\\) to \\(1\\)). We also report the intra-class similarity within real images as an approximate upper bound for comparison. This result further supports the observation of low similarity between synthetic images from PSAQ-ViT and real images. For instance, for Class 1, the intra-class similarity within real images is 0.68, while PSAQ-ViT achieves only 0.44." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The semantic inadequacy issue refers to the presence of dull regions in synthetic images, which contain redundant or non-semantic content [33, 37], hindering the model's learning process. As indicated in [13, 14], a diverse content and textures generally suggests rich information. As shown in Fig. 1b, many regions of synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit a lack of diversity in content, with overly simplified textures. Specifically, the central region of PSAQ-ViT images only contains faint object structures, while PSAQ-ViT V2 images appear excessively smoothed and indistinct." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For high-bit quantization, where model capacity is largely retained [42], the performance degradation remains relatively minor even using semantic distorted and inadequate images [47, 62]. However, in low-bit quantization, where model capacity is severely damaged and informative images are essential for recovering performance [13, 84]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "fine-tuning on these poor-quality images leads to inferior generalization to real datasets, resulting in limited performance. For example, as shown in Tab. 2, the W4A4 ViT-B fine-tuned on real images yields \\(68.16\\%\\), whereas PSAQ-ViT only achieves \\(36.32\\%\\)." + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.18, + 0.849, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.292, + 0.851, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Comparison between attention maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.322, + 0.906, + 0.354 + ], + "angle": 0, + "content": "3.3. Semantics Alignment and Reinforcement Data-Free Quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.906, + 0.391 + ], + "angle": 0, + "content": "In the following, we introduce the proposed SARDFQ, whose framework is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.753, + 0.414 + ], + "angle": 0, + "content": "3.3.1. Attention Priors Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.629 + ], + "angle": 0, + "content": "In ViTs, the self-attention mechanism encodes semantic correlations between image regions, where high-response areas in attention maps strongly correlate with semantic-discriminative content [10, 20, 31]. However, existing DFQ methods overlook this intrinsic property in the generation process. As a result, as shown in Fig. 3, synthetic images often exhibit disordered and unnatural attention patterns, with attention maps either overly diffuse or misaligned toward peripheral regions. This undermines their ability to preserve semantic-discriminative content, causing semantic distortion, as demonstrated in Fig. 1a and Tab. 1. In response, we propose Attention Priors Alignment (APA), which improves semantics alignment by optimizing synthetic images to follow randomly generated structure attention priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.783 + ], + "angle": 0, + "content": "Specifically, given a synthetic image \\(\\tilde{I}\\), we first obtain its attention maps in the \\(h\\)-th head of the \\(l\\)-th block, denoting as \\(\\mathbf{A}_{l,h} \\in \\mathbb{R}^{N \\times N}\\), where \\(N\\) represents the total number of tokens. In DeiT, the attention of the classification token toward other tokens serves as the indicator for semantic versus non-semantic parts [5]. Thus, we extract \\(\\mathbf{A}_{l,h}^c \\in \\mathbb{R}^{1 \\times (N - 1)}\\) from \\(\\mathbf{A}_{l,h}\\), representing the attention of the classification token to all tokens except itself. We then randomly generate attention priors \\(\\tilde{\\mathbf{A}}_{l,h}\\), whose generation is detailed in the next part, and align \\(\\mathbf{A}_{l,h}^c\\) with \\(\\tilde{\\mathbf{A}}_{l,h}\\) by:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.795, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}) = \\operatorname {M S E} \\left(\\mathbf {A} _ {l, h} ^ {c} - \\tilde {\\mathbf {A}} _ {l, h}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where MSE represents the mean squared error. For Swin models that do not use a classification token, we substitute \\(\\mathbf{A}_{l,h}^{c}\\) in Eq.6 with the average attention map of all tokens [10]. As noted in [17], ViTs initially focus on all regions to capture low-level information in shallow blocks" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.089, + 0.218, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.089, + 0.287, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.355, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.089, + 0.424, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.151, + 0.432, + 0.166 + ], + "angle": 0, + "content": "Figure 4. Examples of generated attention priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.484, + 0.286 + ], + "angle": 0, + "content": "and gradually shift their focus toward semantic regions in deeper blocks to extract high-level semantic information. Leveraging this property, we selectively apply \\(\\mathcal{L}_{l,h}^{\\mathrm{APA}}\\) to deeper blocks, progressively aligning attention towards semantically relevant areas. The total APA loss is computed as a depth-weighted sum of the individual Eq. 6 across these deeper blocks:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.295, + 0.482, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = S} ^ {L} \\sum_ {h = 1} ^ {H} \\frac {l}{L} \\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.376 + ], + "angle": 0, + "content": "where \\( S \\) is a pre-given hyper-parameter denoting the start of deep blocks and are experimentally set \\( S = \\frac{L}{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Attention Priors Generation. To generate attention priors \\(\\tilde{\\mathbf{A}}_{l,h}\\), Gaussian Mixture Models (GMMs) are employed as it is the most commonly used distribution with high flexibility. ViTs utilize different attention heads to capture diverse patterns and learn varied, informative representations [8]. Thus, we use distinct GMMs for each head. Note that the goal here is not to replicate real attention maps precisely, but to generate simulated structure attention priors to guide the learning of synthetic images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.648 + ], + "angle": 0, + "content": "In particular, we first initialize an all zero matrix \\(\\tilde{\\mathbf{P}}\\in\\) \\(\\mathbb{R}^{H\\times W}\\) , where \\(H = W = \\sqrt{N - 1}\\) for DeiT, \\(H =\\) \\(W = \\sqrt{N}\\) for Swin. For example, for DeiT-S, the \\(H =\\) \\(W = \\sqrt{196} = 14\\) . Then, we generate \\(k\\) two-dimensional Gaussian distributions, where \\(k\\) is randomly sampled from \\(1\\sim K_{APA}\\) and \\(K_{APA}\\) is set to 5 in all experiments. Each Gaussian has its mean and covariance3. Consequently, the matrix element at the \\(i\\) -th row and \\(j\\) -th column, \\(\\tilde{\\mathbf{P}} [i,j]\\) , is determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.655, + 0.483, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {P}} [ i, j ] = \\max _ {m = 1, \\dots , k} \\mathbf {G} ^ {m} [ i, j ]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.692, + 0.268, + 0.707 + ], + "angle": 0, + "content": "Then, \\(\\tilde{\\mathbf{P}}\\) is normalized by:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.717, + 0.483, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {P}} _ {n} = \\frac {\\tilde {\\mathbf {P}}}{\\sum \\tilde {\\mathbf {P}}} \\cdot (1 - x). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Here, for DeiT that incorporates the classification token, \\( x \\) is randomly sampled from a uniform distribution \\( U(0,1) \\), representing the proportion of the attention score that the classification token allocates to itself. For Swin, which does not use a classification token, \\( x \\) is set to 0. Finally, \\( \\tilde{\\mathbf{P}}_n \\) is flatten to match the dimensionality:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.86, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {A}} _ {l, h} = \\operatorname {f l a t t e n} \\left(\\tilde {\\mathbf {P}} _ {n}\\right). \\tag {10}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.332, + 0.901 + ], + "angle": 0, + "content": "3The pseudo code is detailed in appendix." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "Fig. 4 displays examples of generated attention priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Discussion. APA prevents attention disorder, ensuring that synthetic images exhibit more coherent and natural attention patterns, as demonstrated in Fig. 3. As a result, APA selectively enhances the responses in certain regions, effectively emphasizing semantic-discriminative regions within synthetic images and thus prompting the discriminative features. Although simple, APA enables synthetic images to align better with the real semantics, as validated by both visual and quantitative evaluations. As shown in Fig. 1a, compared to PSAQ-ViT, features obtained after applying the APA loss are more closely aligned with real images, indicating better semantic alignment. The quantitative results in Tab. 1 further support that APA achieves superior semantic alignment. For example, in Class 1, the intra-class similarity for PSAQ-ViT is 0.44, whereas APA achieves a higher similarity of 0.64." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.357, + 0.776, + 0.371 + ], + "angle": 0, + "content": "3.3.2. Multi-Semantic Reinforcement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Current DFQ methods for ViTs [47, 48, 62] optimize images through global optimization, treating the entire image as a single semantic unit. However, this is affected by low-rank structural regularity [34], where adjacent pixels exhibit strong similarity, leading to dull regions with redundant or non-semantic content for model learning [37]. This issue is further exacerbated by the tokenization mechanism in ViTs, as processing images in fixed-size patches make dull regions increase at the patch level [19, 70]. Consequently, as shown in Fig. 1b, synthetic images generated by existing methods exhibit large dull regions, resulting in semantic inadequacy [37]. In response, we propose Multi-Semantic Reinforcement (MSR), which applies localized patch optimization to enhance semantic richness by learning local patches with distinct semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.906, + 0.709 + ], + "angle": 0, + "content": "Specifically, for a synthetic image \\(\\tilde{I}\\), instead of feeding only the entire image, we also feed its patches and optimize them individually. Initially, we select \\(m\\) nonoverlapping patches, where \\(m\\) is chosen randomly from the set \\(\\{1,2,\\dots,K_{MSR}\\}\\), with \\(K_{MSR}\\) set to 4 in all experiments. These \\(m\\) patches are then cropped and resized to match the model's input dimensions:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.719, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\{\\tilde {\\boldsymbol {I}} _ {M S R} ^ {i} \\} _ {i = 1, \\dots , m} = \\operatorname {r e s i z e} \\left(\\operatorname {c r o p} _ {m} (\\tilde {\\boldsymbol {I}})\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.84 + ], + "angle": 0, + "content": "where \\( \\text{crop}_m(\\cdot) \\) crops \\( m \\) non-overlapping patches from input, and \\( \\text{resize}(\\cdot) \\) is the resize function. Each patch, denoted as \\( \\tilde{I}_{MSR}^i \\), is treated as a new image with an assigned semantic target \\( c^i \\). Note that the gradient is backpropagated only to update the corresponding patch in the original image, leaving the rest of the image unaffected." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Softlabel Learning. The one-hot loss (Eq. 3) only learns the semantics of the target class, making it unsuitable for \\(\\tilde{\\pmb{I}}\\) under MSR, as its patches \\(\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}\\) contain distinct semantics. In response, we propose Softlabel Learning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "(SL), which applies multiple semantic targets to accommodate the learning of images augmented by MSR. Specifically, we first sample \\( Z \\in \\mathbb{R}^{C} \\sim U(0,1) \\), then modify its values by:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.162, + 0.429, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} Z [ c ^ {i} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}} _ {M S R} ^ {i}, \\\\ Z [ c ^ {1}, \\ldots , c ^ {m} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}}, \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.211, + 0.484, + 0.288 + ], + "angle": 0, + "content": "where \\( U(\\epsilon_1, \\epsilon_2) \\) denotes the uniform distribution over the interval \\( [\\epsilon_1, \\epsilon_2] \\), \\( m \\) is the number of patches determined in MSR, and \\( \\epsilon_1 \\) and \\( \\epsilon_2 \\) control the softness, both empirically set consistently to 5 and 10 in all experiments. The soft target is defined as \\( T_s = \\mathrm{softmax}(Z) \\), and SL loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.296, + 0.482, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\mathrm {S L}} \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}\\right) = S C E \\left(F \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}, T _ {s}\\right), \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.484, + 0.4 + ], + "angle": 0, + "content": "where \\( SCE(\\cdot, \\cdot) \\) is soft cross entropy and \\( F(\\cdot) \\) returns the predicted probability for its input. SL facilitates smooth learning between across semantic targets, ensuring that MSR-enhanced images receive consistent supervision rather than conflicting supervision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Discussion. By leveraging localized patch optimization, MSR ensures that each patch contributes unique semantics, forcing synthetic images to capture diverse features rather than being dominated by large homogeneous dull regions. As demonstrated by the richer and more diverse content and textures in Fig. 1b, MSR effectively reduces dull regions and enhances semantic richness in synthetic images. Moreover, MSR transforms synthetic images into composites of multiple semantic objects rather than a single unit, thereby providing more distinct semantic samples, i.e., \\(\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}\\), for model training. Unlike traditional cropping used in data augmentation, which aims to improve classification robustness by training with cropped patches labeled with the original class, MSR aims for semantic richness within synthetic images, ultimately enabling accurate data-free quantization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.253, + 0.667 + ], + "angle": 0, + "content": "3.4. Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.748 + ], + "angle": 0, + "content": "The overall pipeline consists of two stages: data synthesis and quantized network learning. The first stage uses the proposed SARDFQ to produce synthetic images. The second stage fine-tunes the quantized model using the generated synthetic images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.242, + 0.77 + ], + "angle": 0, + "content": "3.4.1. Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.484, + 0.819 + ], + "angle": 0, + "content": "In the data synthesis stage, we combine the proposed APA loss of Eq. 7, SL loss of Eq. 12, and TV loss of Eq. 4 to formulate the objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.828, + 0.484, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G} (\\tilde {\\boldsymbol {I}}) = \\alpha_ {1} \\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) + \\mathcal {L} ^ {\\mathrm {S L}} (\\tilde {\\boldsymbol {I}}) + 0. 0 5 \\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) is hyperparameters and is determined by gridsearch. Note that the weight of TV loss is fixed to 0.05, following [47], to avoid a cumbersome hyperparameter search." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.767, + 0.108 + ], + "angle": 0, + "content": "3.4.2. Quantized Network Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.907, + 0.232 + ], + "angle": 0, + "content": "Recently DFQ methods have introduced the PTQ methods in learning quantized models due to their advantages of speed, memory efficiency, and performance [35, 62]. Thus, following the success of [42, 71], we fine-tune the quantized network block-wisely. Specifically, denote \\(\\mathbf{X}_l\\) as the outputs of the \\(l\\)-th block of the full-precision model, and \\(\\bar{\\mathbf{X}}_l\\) represent outputs of the quantized counterpart. The reconstruction loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.241, + 0.907, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l} = \\left\\| \\mathbf {X} _ {l} - \\bar {\\mathbf {X}} _ {l} \\right\\| _ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_l\\) is only backward to update weights within \\(l\\)-th block. Note that for a fair comparison, all compared methods adopt the same quantized network learning stage." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.325, + 0.64, + 0.343 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.731, + 0.367 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Models and Tasks. We evaluate the performance of SARDFQ by test quantized ViT-S/B [20], DeiT-T/S/B [68], and Swin-S/B [54] on the classification task using ImageNet [63]. The pre-trained models are downloaded from the timm library. In appendix, we further provide results on detection and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.597 + ], + "angle": 0, + "content": "Comparison methods. We compare our SARDFQ against Gaussian noise, real images, and previous methods including SMI [33] and PSAQ-ViT [47] and its subsequent version, PSAQ-ViT V2 [48]. For a fair comparison, we generate synthetic images using their methods and apply our quantized network learning strategy. We use the official code for SMI and PSAQ-ViT to reproduce their images, while PSAQ-ViT V2 is re-implemented by us, as no official code is available." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Experimental settings. All experiments were conducted using the PyTorch framework [60] on a single NVIDIA 3090 GPU. In the data synthesis stage, synthetic images were initialized with standard Gaussian noise, generating 32 images in total. The Adam optimizer [38] with \\(\\beta_{1} = 0.5\\), \\(\\beta_{2} = 0.9\\) was used, with learning rates of 0.25 for Swin and 0.2 for others, and a total of 1,000 iterations. For all models, \\(K_{APA}\\), \\(K_{MSR}\\), \\(\\epsilon_{1}\\), and \\(\\epsilon_{2}\\) were set to 5, 4, 5, and 10, respectively, based on a search with W4A4 DeiT-S. The value of \\(\\alpha_{1}\\) was determined by grid search for each model: 1e5 for DeiT-T/S, 1e4 for DeiT-B, 100 for ViT-B and Swin-B, 10 for Swin-S, and 1 for ViT-S. Although further hyperparameter search may improve performance, the current settings already yield superior results. In the quantized network learning stage, following [85], the Adam optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\) was used, with weight decay set to 0 and an initial learning rate of 4e-5, adjusted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "4In appendix, we present practical efficiency results and more performance comparisons with more methods such as CLAMP-ViT [62] on W8/A8 and W4/A8 settings." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12484" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.09, + 0.894, + 0.475 + ], + "angle": 0, + "content": "
ModelW/ARealGaussian noisePSAQ-ViT [47]PSAQ-ViT V2 [48]SMI [33]SARDFQ (Ours)
ViT-S (81.39)4/466.576.0247.2441.5324.3329.4150.32
5/576.6936.7771.5968.4161.3365.1974.31
6/679.4661.2077.2074.7672.9572.4678.40
ViT-B (84.54)4/468.160.1536.3226.3235.2719.6751.84
5/579.214.1668.4867.9567.5357.1370.70
6/681.8955.1876.6571.8776.3369.8279.16
DeiT-T (72.21)4/456.6017.4347.7530.2030.1413.1852.06
5/567.0943.4964.1055.1656.4439.3566.41
6/669.8156.2368.3762.7764.0344.3969.73
DeiT-S (79.85)4/468.4620.8958.2845.5342.7711.7162.29
5/575.0641.0671.9063.1462.8829.1374.06
6/677.8765.6375.8568.8571.6537.6977.31
DeiT-B (81.85)4/477.0747.2071.7566.4365.3359.0472.17
5/579.8665.4678.4576.7776.7475.3378.72
6/680.9062.7980.0079.2278.8177.6680.15
Swin-S (83.20)4/478.1231.9273.1965.5565.8574.74
5/580.5152.1078.1574.3775.4179.56
6/680.6065.6679.7478.5078.2580.56
Swin-B (85.27)4/478.8030.1471.8467.4265.2376.42
5/582.5135.2878.5077.2075.2580.82
6/682.6467.3782.0081.4180.3083.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.486, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Table 2. Quantization results on ImageNet dataset, with top-1 accuracy \\((\\%)\\) reported. The performance of the full-precision model is listed below the model name. \"W/A\" denotes the bit-width of weights/activations. \"Real\" refers to using real images. For SMI [33], we provide the performance of using dense (normal-sized numbers) and sparse (smaller-sized numbers) synthetic images, respectively. Note that for Swin models, we do not provide the results for sparse synthetic images as the sparse generation method of SMI is infeasible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.618 + ], + "angle": 0, + "content": "via cosine decay for 100 iterations. A channel-wise quantizer was used for weights, and a layer-wise quantizer for activations, with all matrix multiplications in ViTs quantized [47-49]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.29, + 0.65 + ], + "angle": 0, + "content": "4.2. Quantization Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The quantization results are presented in Tab.2. Our SARDFQ demonstrates consistent improvements across various quantization bit-width configurations, particularly with low bit-width settings. Specifically, for ViT-S, SARDFQ improves the performance by \\(3.08\\%\\) in the W4/A4 setting, \\(2.72\\%\\) in the W5/A5 setting, and \\(1.20\\%\\) in the W6/A6 setting. For ViT-B, SARDFQ achieves performance gains of \\(15.52\\%\\) in the W4/A4 setting, \\(2.22\\%\\) in the W5/A5 setting, and \\(2.51\\%\\) in the W6/A6 setting. Results on DeiT also demonstrate the effectiveness of SARDFQ. For example, on DeiT-T, SARDFQ shows a marked improvement by increasing top-1 accuracy by \\(4.31\\%\\) in the W4/A4 setting, \\(2.31\\%\\) in the W5/A5 setting, and \\(1.36\\%\\) in the W6/A6 setting. For DeiT-S, SARDFQ enhances top-1 accuracy by \\(4.01\\%\\) in the W4/A4 setting, \\(2.16\\%\\) in the W5/A5 setting, and \\(1.46\\%\\) in the W6/A6 setting. The quan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.694 + ], + "angle": 0, + "content": "tization results of Swin-S/B also affirm the superiority of our SARDFQ in enhancing model accuracy under different quantization configurations. In particular, for Swin-S, the proposed SARDFQ increases the accuracy by \\(1.55\\%\\) for the W4/A4 setting, \\(1.41\\%\\) for the W5/A5 setting, and \\(0.82\\%\\) for the W6/A6 setting, respectively. When it comes to Swin-B, the proposed SARDFQ increases the accuracy by \\(4.58\\%\\) for the W4/A4 setting, \\(2.32\\%\\) for the W5/A5 setting, and \\(1.03\\%\\) for the W6/A6 setting, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.667, + 0.725 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.887, + 0.748 + ], + "angle": 0, + "content": "All ablation studies are conducted on the W4A4 DeiT-S." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Analysis of APA, MSR, and SL. We analyze the effectiveness of the proposed APA (Sec.,3.3.1), MSR (Sec.,3.3.2), and SL (Eq.,12) in Tab.,3. Adding APA and SL individually to the baseline increases accuracy. Notably, APA boosts performance from \\(51.73\\%\\) to \\(60.26\\%\\), confirming its effectiveness in aligning semantics (Sec.,3.3.1). Applying MSR alone slightly decreases accuracy from \\(51.73\\%\\) to \\(50.75\\%\\), indicating that one-hot loss is unsuitable for MSR-augmented synthetic images. However, when both MSR and SL are applied, accuracy rises to \\(56.08\\%\\), sug-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12485" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.17, + 0.089, + 0.408, + 0.25 + ], + "angle": 0, + "content": "
APAMSRSLAcc. (%)
Baseline51.73
✓60.26
✓50.75
✓52.02
✓✓61.58
✓✓60.51
✓✓56.08
✓✓✓62.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.303 + ], + "angle": 0, + "content": "Table 3. Influence of the proposed APA, MSR, and SL on accuracy. The baseline adopts the one-hot loss. Applying APA, MSR, and SL yields SARDFQ." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.397 + ], + "angle": 0, + "content": "gesting SL is more compatible with MSR than one-hot loss. Combining APA with either MSR or SL further improves performance. For example, APA and SL together yield an accuracy of \\(60.51\\%\\), and when all three strategies are used, the best performance of \\(62.29\\%\\) is achieved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Analysis of priors distribution. Tab. 4a showcases the results of using other distributions to formulate the attention priors. The unevenly distributed GMM and Laplace present comparable performance of \\(62.29\\%\\) and \\(62.16\\%\\), respectively. Moreover, GMM provides a similar performance to the real's, indicating it performs well in imitating the patterns of real images." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.516, + 0.283, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.637, + 0.198, + 0.647 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.516, + 0.48, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.636, + 0.391, + 0.647 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.66, + 0.436, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Effect of varying (a) \\(\\alpha_{1}\\) and (b) \\(K_{MSR}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis of \\(\\alpha_{1}, K_{APA}\\), and \\(K_{MSR}\\). The \\(\\alpha_{1}\\) from Eq. 13 balance the importance of the proposed APA loss during the update of the synthetic images. Fig. 5a demonstrates that the optimal performance is achieved when \\(\\alpha_{1} = 1e5\\). Incrementally increasing \\(\\alpha_{1}\\) improves performance up to \\(62.29\\%\\) at \\(\\alpha_{1} = 1e5\\). However, further increases in \\(\\alpha_{1}\\) subsequently degrade performance. The \\(K_{APA}\\) in APA is the upper limit on the number of the Gaussian distributions used for priors generation. Tab. 4b displays the ablation study for different values of \\(K_{APA}\\). The best accuracy is achieved when \\(K_{APA} = 5\\). The \\(K_{MSR}\\) in MSR is the upper limit on the number of patches. Fig. 5b demonstrates that when \\(K_{MSR} = 4\\), the optimal performance is achieved. Using \\(K_{MSR}\\) larger than 4 will hurt the accuracy. We consider this" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.104, + 0.722, + 0.183 + ], + "angle": 0, + "content": "
Priors DistributionTop-1
GMM62.29
Laplace62.16
Real63.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.185, + 0.613, + 0.196 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.753, + 0.089, + 0.882, + 0.197 + ], + "angle": 0, + "content": "
KAPATop-1
161.13
361.52
562.29
761.53
961.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.807, + 0.199, + 0.826, + 0.211 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.218, + 0.678, + 0.281 + ], + "angle": 0, + "content": "
STop-1
061.96
L/262.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.613, + 0.284, + 0.631, + 0.296 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "table", + "bbox": [ + 0.741, + 0.218, + 0.857, + 0.281 + ], + "angle": 0, + "content": "
w. l/LTop-1
✓62.29
×61.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.791, + 0.284, + 0.809, + 0.296 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.338 + ], + "angle": 0, + "content": "Table 4. Effect of varying (a) priors types; (b) \\( K_{APA} \\); (c) \\( S \\); (d) \\( \\frac{l}{L} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.899, + 0.368 + ], + "angle": 0, + "content": "due to limited patch resolution if using a too large \\(K_{MSR}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.906, + 0.46 + ], + "angle": 0, + "content": "Analysis of APA loss. Here, we conduct the ablation study by considering the \\(S\\) and scale \\(\\frac{l}{L}\\) in Eq.7. Tab.4c presents the effect of varying \\(S\\) in Eq.7. If applying APA loss to all blocks \\((S = 0)\\), the top-1 accuracy decreases to \\(61.96\\%\\). From Tab.4d, it can be seen that absorbing the scale \\(\\frac{l}{L}\\) in Eq.7 presents \\(0.97\\%\\) performance gains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.473, + 0.637, + 0.489 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.651 + ], + "angle": 0, + "content": "We further discuss some limitations of the proposed SARDFQ, which will guide future research directions. First, although SARDFQ shows substantial performance improvement, a performance gap between SARDFQ and real data remains challenging, highlighting the need for a stronger semantics alignment and reinforcement method. Second, SARDFQ currently lacks a theoretical foundation. Future work could establish a theoretical framework for SARDFQ, particularly in understanding how APA and MSR influence synthetic images in a formalized manner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we investigate the DFQ method for ViTs. We first identify that synthetic images generated by existing methods suffer from semantic distortion and inadequacy issues, and propose SARDFQ to address these issues. To mitigate semantic distortion, SARDFQ introduces APA, which guides synthetic images to align with randomly generated structural attention patterns. To tackle semantic inadequacy, SARDFQ incorporates MSR. MSR optimizes different regions of synthetic images with unique semantics, thereby enhancing overall semantic richness. Moreover, SARDFQ employs SL, which adopts multiple semantic targets to ensure seamless learning of images augmented by MSR. Extensive experiments on various ViT models and tasks validate the effectiveness of SARDFQ." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12486" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.256 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (Np. 624B2119, No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.282, + 0.188, + 0.298 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.307, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 6836-6846, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.483, + 0.434 + ], + "angle": 0, + "content": "[2] Jianhong Bai, Yuchen Yang, Huanpeng Chu, Hualiang Wang, Zuozhu Liu, Ruizhe Chen, Xiaoxuan He, Lianrui Mu, Chengfei Cai, and Haoji Hu. Robustness-guided image synthesis for data-free quantization. arXiv preprint arXiv:2310.03661, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[3] Ron Banner, Yury Nahshan, Daniel Soudry, et al. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems (NeurIPS), pages 7950-7958, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.484, + 0.547 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(8):1798-1828, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[6] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13169-13178, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 397-406, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.307, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "coarse-to-fine method for vision transformer. In AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[11] Mengzhao Chen, Wenqi Shao, Peng Xu, Mingbao Lin, Kaipeng Zhang, Fei Chao, Rongrong Ji, Yu Qiao, and Ping Luo. Diffrate: Differentiable compression rate for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17164-17174, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[12] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[13] Xinrui Chen, Yizhi Wang, Renao Yan, Yiqing Liu, Tian Guan, and Yonghong He. Texq: Zero-shot network quantization with texture feature distribution calibration. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[14] Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. In Advances in Neural Information Processing Systems (NeurIPS), pages 14835-14847, 2021. 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[15] Kanghyun Choi, Hye Yoon Lee, Deokki Hong, Joonsang Yu, Noseong Park, Youngsok Kim, and Jinho Lee. It's all in the teacher: Zero-shot quantization brought closer to the teacher. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8311-8321, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[16] Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, and Jinho Lee. Mimiq: Low-bit data-free quantization of vision transformers. arXiv preprint arXiv:2407.20021, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-attention and convolutional layers. In International Conference on Learning Representations (ICLR), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[18] Yifu Ding, Haotong Qin, Qinghua Yan, Zhenhua Chai, Junjie Liu, Xiaolin Wei, and Xianglong Liu. Towards accurate posttraining quantization for vision transformer. In 30th ACM International Conference on Multimedia (ACMMM), pages 5380-5388, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[19] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International conference on machine learning (ICML), pages 2793-2803. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[21] Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero-shot quantization. In International Conference on Machine Learning (ICML), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12487" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Learned step size quantization. In International Conference on Learning Representations (ICLR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[23] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[24] Natalia Frumkin, Dibakar Gope, and Diana Marculescu. Jumping through local minima: Quantization in the loss landscape of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16978-16988, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[25] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4852–4861, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[26] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), pages 2672–2680, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[27] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quantization via diagonal hessian approximation. In The Eleventh International Conference on Learning Representations (ICLR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[28] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[29] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):87-110, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[30] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[31] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 773-783, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[32] Zejiang Hou and Sun-Yuan Kung. Multi-dimensional vision transformer compression via dependency guided gaussian process search. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3669-3678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[33] Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Lei Li, Chun Yuan, and Dacheng Tao. Sparse model inversion: Efficient inversion of vision transformers for data-free applications. In International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 6, 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[34] Jinggang Huang and David Mumford. Statistics of natural images and models. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 541-547. IEEE, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[35] Yongkweon Jeon, Chungman Lee, and Ho-young Kim. Genie: show me the data for quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12073, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[36] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM Computing Surveys (CSUR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[37] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning (ICML), pages 11102-11118. PMLR, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[39] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[40] Huantong Li, Xiangmiao Wu, Fanbing Lv, Daihai Liao, Thomas H Li, Yonggang Zhang, Bo Han, and Mingkui Tan. Hard sample matters a lot in zero-shot quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24417-24426, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[41] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In International Conference on Learning Representations (ICLR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[42] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations (ICLR), 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[43] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 34451-34463, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[44] Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Stableq: Enhancing data-scarce quantization with text-to-image data. arXiv preprint arXiv:2312.05272, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[45] Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, and Priyadarshini Panda. Genq: Quantization in low data regimes with generative synthetic data. In European Conference on Computer Vision (ECCV), pages 216-235. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[46] Zhikai Li and Qingyi Gu. I-vit: Integer-only quantization for efficient vision transformer inference. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17065-17075, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Zhikai Li, Liping Ma, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "vision transformers. In European Conference on Computer Vision (ECCV), pages 154-170. Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[48] Zhikai Li, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2023. 2, 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[49] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17227-17236, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[50] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 1833-1844, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.484, + 0.407 + ], + "angle": 0, + "content": "[51] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), pages 1173–1179, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[52] Shih-Yang Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning (ICML), pages 21813-21824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[53] Yijiang Liu, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20321-20330, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[54] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF international conference on computer vision (ICCV), pages 10012-10022, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.613, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[55] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 28092-28103, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[56] Yiwei Ma, Jiayi Ji, Xiaoshuai Sun, Yiyi Zhou, and Rongrong Ji. Towards local visual modeling for image captioning. Pattern Recognition, 138:109420, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[57] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[58] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 23296-23308, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[59] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 3163-3172, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[60] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), pages 8026-8037, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[61] Biao Qian, Yang Wang, Richang Hong, and Meng Wang. Adaptive data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7960-7968, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[62] Akshit Ramachandran, Souvik Kundu, and Tushar Krishna. Clamp-vit: contrastive data-free learning for adaptive posttraining quantization of vits. In European Conference on Computer Vision (ECCV). Springer, 2024. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[64] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transformers in medical imaging: A survey. Medical Image Analysis, page 102802, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[65] Huixin Sun, Runqi Wang, Yanjing Li, Xianbin Cao, Xiaolong Jiang, Yao Hu, and Baochang Zhang. P4q: Learning to prompt for quantization in visual-language models. arXiv preprint arXiv:2409.17634, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[66] Yehui Tang, Yunhe Wang, Yixing Xu, Yiping Deng, Chao Xu, Dacheng Tao, and Chang Xu. Manifold regularized dynamic network pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5028, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.615 + ], + "angle": 0, + "content": "[67] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12165-12174, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[68] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357. PMLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[69] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9:2579-2605, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[70] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[71] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. In International Conference on Learning Representations (ICLR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[72] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In IEEE/CVF In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12489" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ternational Conference on Computer Vision (ICCV), pages 10033-10041, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[73] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision (ECCV), pages 68-85, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[74] Zhiqiang Shen Xijie Huang and Kwang-Ting Cheng. Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.278 + ], + "angle": 0, + "content": "[75] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[76] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[77] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transffer via deep-inversion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8715-8724, 2020. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[78] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[79] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 818-833, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[80] Jinnian Zhang, Houwen Peng, Kan Wu, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Minivit: Compressing vision transformers with weight multiplexing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12145-12154, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.592, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[81] Xiangguo Zhang, Haotong Qin, Yifu Ding, Ruihao Gong, Qinghua Yan, Renshuai Tao, Yuhang Li, Fengwei Yu, and Xianglong Liu. Diversifying sample generation for accurate data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15658-15667, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[82] Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, and Yunhe Wang. Less is more: Focus attention for efficient detr. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6674-6683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[83] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 6881-6890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[84] Yunshan Zhong, Mingbao Lin, Gongrui Nan, Jianzhuang Liu, Baochang Zhang, Yonghong Tian, and Rongrong Ji. Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12339-12348, 2022. 2, 4, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[85] Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&S-vit: An inclusive & stable method for pushing the limit of post-training vits quantization. arXiv preprint arXiv:2311.10126, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[86] Yunshan Zhong, Jiawei Hu, You Huang, Yuxin Zhang, and Rongrong Ji. Erq: Error reduction for post-training quantization of vision transformers. In International Conference on Machine Learning (ICML), 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12490" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_origin.pdf b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..174c2a1ae992e0395e997024e7b8b63fd36758b0 --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/33638821-3dce-4a55-a900-82748a75aeee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a24929928c29339fcdb7983134465b5ac7b62bd61902eb273eaf994850c6aae +size 688280 diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/full.md b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..560c5cfc89a693414c10b4a9ef7b843a92df2fa1 --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/full.md @@ -0,0 +1,419 @@ +# Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers + +Yunshan Zhong $^{1,2}$ , Yuyao Zhou $^{2}$ , Yuxin Zhang $^{2}$ , Wanchen Sui $^{3}$ , Shen Li $^{3}$ , Yong Li $^{3}$ , Fei Chao $^{2}$ , Rongrong Ji $^{1,2*}$ + +$^{1}$ Institute of Artificial Intelligence, Xiamen University + +$^{2}$ MAC Lab, School of Informatics, Xiamen University $^{3}$ Alibaba Group + +viper.zhong@gmail.com, {yuyaozhou, yuxinzhang}@stu.xmu.edu.cn +{wanchen.swc, litan.ps, jiufeng.ly}@alibaba-inc.com, {fchao, rrji}@xmu.edu.cn + +# Abstract + +Data-free quantization (DFQ) enables model quantization without accessing real data, addressing concerns regarding data security and privacy. With the growing adoption of Vision Transformers (ViTs), DFQ for ViTs has garnered significant attention. However, existing DFQ methods exhibit two limitations: (1) semantic distortion, where the semantics of synthetic images deviate substantially from those of real images, and (2) semantic inadequacy, where synthetic images contain extensive regions with limited content and oversimplified textures, leading to suboptimal quantization performance. To address these limitations, we propose SARDFQ, a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs. To address semantic distortion, SARDFQ incorporates Attention Priors Alignment (APA), which optimizes synthetic images to follow randomly generated structure attention priors. To mitigate semantic inadequacy, SARDFQ introduces Multi-Semantic Reinforcement (MSR), leveraging localized patch optimization to enhance semantic richness across synthetic images. Furthermore, SARDFQ employs Soft-Label Learning (SL), wherein multiple semantic targets are adapted to facilitate the learning of multi-semantic images augmented by MSR. Extensive experiments demonstrate the effectiveness of SARDFQ, significantly surpassing existing methods. For example, SARDFQ improves top-1 accuracy on ImageNet by $15.52\%$ for W4A4 ViT-B1. + +# 1. Introduction + +Vision Transformers (ViTs) [20, 29, 36] have demonstrated remarkable success across various computer vision tasks [1, + +![](images/4b50f3a45edef337a0f4073e0ff2e040144da215e4fb9126beb74c732643f2fe.jpg) + +(a) The t-SNE[69] visualization of the penultimate-layer features (extracted by DeiT-S) of synthetic images. Each marker (circle, star, triangle) represents a distinct category. The red dashed circles highlight the features extracted from our APA and real images. Notably, the features produced by PSAQ-ViT[47] exhibit substantial deviation from those of real images, indicating semantic distortion. In contrast, our APA yields features that more closely align with those of real images, suggesting improved semantics alignment. + +![](images/249fb76fabd4d8d549bcfc3fb1bc6e46377268668c0690f604df88d19982974d.jpg) +PSAQ + +![](images/39086b1071ec1e2cac7df6a959b641a108c4ee0baa0e9e30a61a0d1ebef8644d.jpg) +PSAQ V2 + +![](images/faf3b7d4f308588c04140baf283cbb66741b4e3f32e492455dcd0bf875558327.jpg) +SARDFQ (Ours) + +(b) The images of PSAQ-ViT and PSAQ-ViT V2 exhibit numerous dull regions with limited content and simplified textures, reflecting semantic inadequacy. In comparison, our SARDFQ generates images with greater diversity in both content and texture, demonstrating enhanced semantics. + +Figure 1. Illustration of (a) semantic distortion and (b) semantic inadequacy. + +7, 56, 68]. However, their high computational cost and substantial memory footprint hinder deployment in resource- + +constrained environments [11, 30, 32, 46, 55, 67, 82]. To address these limitations, quantization [39] has emerged as a promising solution, which reduces model complexity by converting full-precision weight and activations into low-bit representations. + +Traditional quantization methods typically require access to the original training dataset, which raises data privacy and security concerns [6, 27, 77, 84]. As a result, data-free quantization (DFQ) has gained increasing attention, allowing quantization without the need for real data [14, 76, 81]. However, most existing DFQ methods are designed specifically for convolutional neural networks (CNNs) and are not directly applicable to vision transformers (ViTs). These methods generally rely on batch normalization statistics (BNS), which capture the distribution of real data, to synthesize in-distribution synthetic data [6, 76, 81, 84]. Yet, BNS is unavailable for ViTs, which use layer normalization (LN) to dynamically compute distribution statistics during inference [47]. Recently, several DFQ methods have been proposed for ViTs [16, 33, 47, 48, 62]. For example, PSAQ-ViT [47] introduces patch similarity entropy (PSE) loss to optimize Gaussian noise towards usable synthetic images. + +Nevertheless, we observe that existing methods suffer from semantic distortion and inadequacy issues. As shown in Fig. 1a, features of synthetic images generated by PSAQ-ViT deviate significantly from those of real images. Tab. 1 shows that the cosine similarity between synthetic images from PSAQ-ViT and real images is notably low, also indicating significant distortion. These results highlight the issue of semantic distortion. Moreover, as shown in Fig. $1\mathrm{b}^2$ synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit many regions with limited content diversity and overly simplified textures. These low-quality dull regions are useless or even detrimental to model learning [33], highlighting the issue of semantic inadequacy. Consequently, quantized models trained on such low-quality images suffer from degraded performance. + +Motivated by the above analysis, we propose a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs, termed SARDFQ. The overall framework is depicted in Fig.2. To address the semantic distortion issue, SARDFQ introduces Attention Priors Alignment (APA), where synthetic images are optimized to follow structured attention priors generated using Gaussian Mixture Models (GMMs). APA effectively aligns the semantics of synthetic images with real images, as validated by both visual and quantitative analyses. As shown in Fig. 1a, features of APA exhibit a closer alignment to those of real images, while quantitative results in Tab. 1 also confirm that the semantics of APA are more consistent to the real images, indicating enhanced semantics alignment. To address the semantic inadequacy issue, SARDFQ incorpo + +rates Multi-Semantic Reinforcement (MSR) and Softlabel Learning (SL). MSR utilizes localized patch optimization, which encourages different sub-patches of synthetic images to capture various semantics, reinforcing the rich semantics across images. SL applies multiple semantic targets to accommodate the learning of multi-semantic images augmented by MSR. As shown in Fig. 1b, synthetic images after applying MSR exhibit greater diversity in content and texture, providing reinforced semantics. + +Experimental results across various ViT models and tasks demonstrate that SARDFQ presents substantial performance improvements. For example, SARDFQ achieves a $15.52\%$ increase in top-1 accuracy on the ImageNet dataset for the W4A4 ViT-B model. + +# 2. Related Works + +# 2.1. Vision Transformers + +The great success of transformers in the natural language processing field has driven widespread attempts in the computer vision community to apply them to vision tasks [12, 28, 72]. ViT [20] is the pioneer that builds a transformer-based model to handle images, boosting the performance on the image classification task. DeiT [68] introduces an efficient teacher-student training strategy where a distillation token is employed to distill knowledge from the teacher model to the student model. Swin Transformers [54] builds an efficient and effective hierarchical model by introducing a shifted window-based self-attention mechanism. Other than the image classification task, the applications of ViTs also have broadened considerably, manifesting groundbreaking performance in object detection [7], image segmentation [9, 83], low-level vision [50], video recognition [1, 59], and medical image processing [64], etc. Nevertheless, the impressive performance of ViTs relies on a high number of parameters and significant computational overhead, preventing deployment in resource-constrained environments. Several recent efforts design lightweight ViTs, such as MobileViT [57], MiniVit [80], and TinyViT [73]. However, the model complexity is still unsatisfactory [47]. + +# 2.2. Network Quantization + +Data-Driven Quantization. Model quantization reduces the complexity of neural networks by replacing full-precision weight and activation with the low-bit format. Data-driven quantization can be roughly divided into two categories: quantization-aware training (QAT) and posttraining quantization (PTQ). QAT is compute-heavy since it re-trains the quantized model with the full training data to retain performance [22, 25, 41, 43, 46, 52, 65, 74]. PTQ perform quantization with a tiny dataset and a reduced time overhead, harvesting widespread attention [3, 42]. The specific architecture of ViTs, such as LayerNorm and the self + +![](images/1b7b933fd4ed13c1c2693d8e289055bfe0cf694f3afb23c5f54be139c7e73d04.jpg) +Figure 2. SARDFQ Framework overview: Attention Priors Alignment (APA) employs randomly generated attention priors to improve semantics alignment. Multi-Semantic Reinforcement (MSR) learns the different regions of synthetic images with various semantics to enhance overall semantic richness. Meanwhile, Softlabel Learning (SL) adopts multiple semantic targets to ensure consistent learning of multi-semantic images augmented by MSR. + +attention module, urges distinct PTQ methods compared to CNNs [18, 24, 46, 51, 53, 86]. For example, Liu et al. [55] develop a ranking loss to maintain the relative order of the self-attention activation. Unfortunately, both QAT and PTQ involve the original training data, causing concerns about data privacy and security issues in data-sensitive scenarios. + +Data-Free Quantization. DFQ quantizes models without accessing real data [2, 13-15, 27, 35, 40, 44, 45, 61]. Most previous DFQ methods focus on CNN, where the BNS can be adopted as the regularization term [6, 81]. However, BNS is infeasible for ViTs built on the LN. Recently, few efforts have been explored to accommodate ViTs [16, 33, 47, 48, 62]. PSAQ-ViT [47] introduces the first DFQ method for ViTs. They discover that Gaussian noise yields homogeneous patches, while the real image yields heterogeneous patches. Thus, patch similarity entropy (PSE) loss is proposed to optimize the Gaussian noise towards real-like images by making them showcase heterogeneous patches. Based on PSAQ-ViT, PSAQ-ViT V2 [48] further introduces an adversarial learning strategy [26]. [62] incorporates contrastive learning and proposes an iterative generation-quantization PTQ-based DFQ method. [33] proposes a sparse generation method to remove noisy and hallucination backgrounds in synthetic images. + +# 3. Method + +# 3.1. Preliminaries + +# 3.1.1.Quantizers + +We employ the linear quantizer for all weights and activations, except for the attention scores, which use a log2 quantizer to handle highly non-negative and uneven values [24, 49, 51]. For the linear quantizer, given a full-precision input $\mathbf{x}$ and bit-width $b$ , the quantized value $\mathbf{x}_q$ and the de-quantized value $\bar{\mathbf{x}}$ are computed as follows: + +$$ +\mathbf {x} _ {q} = \operatorname {c l i p} \left(\left\lfloor \frac {\mathbf {x}}{\Delta} \right\rceil + z, 0, 2 ^ {b} - 1\right), \bar {\mathbf {x}} = \Delta \cdot \left(\mathbf {x} _ {q} - z\right), \tag {1} +$$ + +where $\lfloor \cdot \rfloor$ denotes rounding to the nearest integer, and clip limits the value to $[0,2^{b} - 1]$ . Here, $\Delta$ and $z$ are the scale factor and zero-point, respectively. For the log2 quantizer: + +$$ +\mathbf {x} _ {q} = \operatorname {c l i p} \left(\left\lfloor - \log_ {2} \frac {\mathbf {x}}{\Delta} \right\rceil , 0, 2 ^ {b} - 1\right), \bar {\mathbf {x}} = \Delta \cdot 2 ^ {- \mathbf {x} _ {q}}. \tag {2} +$$ + +# 3.1.2. Data Synthesis + +DFQ methods parameterize synthetic images and optimize them toward real-like images with a pre-trained full-precision model $F$ . Given a image $\tilde{I}$ initialized from Gaussian noise, the one-hot loss [75] is introduced to learn label-related semantics: + +$$ +\mathcal {L} ^ {\mathrm {O H}} (\tilde {\boldsymbol {I}}) = C E (F (\tilde {\boldsymbol {I}}), c) \tag {3} +$$ + +where $CE(\cdot, \cdot)$ represents the cross entropy, $c$ is a random class label, and $F(\cdot)$ returns the predicted probability for image $\tilde{I}$ . + +Moreover, total variance (TV) [77] loss is a smoothing regularization term to improve the image quality: + +$$ +\mathcal {L} ^ {\mathrm {T V}} (\tilde {\boldsymbol {I}}) = \iint | \nabla \tilde {\boldsymbol {I}} (\tau_ {1}, \tau_ {2}) | d \tau_ {1} d \tau_ {2}. \tag {4} +$$ + +where $\nabla \tilde{I} (\tau_1,\tau_2)$ denotes the gradient at $\tilde{I}$ at $(\tau_{1},\tau_{2})$ + +To perform DFQ for ViTs, PSAQ-ViT [47] proposes patch similarity entropy (PSE) loss. It first compute patch similarity $\Gamma_{l}[i,j] = \frac{u_{i} \cdot u_{j}}{||u_{i}||||u_{j}||}$ , where $u_{i}, u_{j}$ are feature vectors of MHSA outputs in $l$ -th block and $||\cdot||$ denotes the $l_{2}$ norm. Then, it estimates the density function $\hat{f}_{l}(x) = \frac{1}{Mh}\sum_{m=1}^{M}K\left(\frac{x - x_{m}}{h}\right)$ , where $K(\cdot)$ is a normal kernel, $h$ is the bandwidth, and $x_{m}$ is the kernel center derived from $\Gamma_{l}$ . Finally, the PSE loss is defined as: + +$$ +\mathcal {L} ^ {\mathrm {P S E}} (\tilde {\boldsymbol {I}}) = \sum_ {l = 1} ^ {L} \int \hat {f} _ {l} (x) \log [ \hat {f} _ {l} (x) ] d x, \tag {5} +$$ + +where $L$ is the block number of the model. + +
Method ClassesRealPSAQ-ViT [47]APA (Ours)
Class 10.680.440.64
Class 20.320.260.31
Class 30.410.310.36
+ +Table 1. Average cosine similarity of three randomly selected classes. For real images, the similarity is measured within the class itself, while for PSAQ-ViT and APA, the similarity is measured between synthetic and real images of the same class. The results show that APA achieves higher similarity than PSE loss, indicating aligned semantics. + +# 3.2. Observations + +Existing DFQ methods have made significant progress. However, through carefully analyzing the synthetic images, we reveal that these images suffer from issues of semantic distortion and semantic inadequacy, both of which hinder further advancement in the DFQ of ViTs. + +The semantic distortion issue refers to the significant divergence between the semantics of synthetic images and real images. To demonstrate this, we visualize the features of synthetic and real images in Fig. 1a. Note that the penultimate feature typically is regarded to represent the semantics of the input [4, 58, 78, 79]. It is clear that the features of PSAQ-ViT diverge significantly from real images, suggesting that these images fail to capture the true semantic distribution of real data. Tab. 1 quantitatively measures the semantics using the average cosine similarity (ranging from $-1$ to $1$ ). We also report the intra-class similarity within real images as an approximate upper bound for comparison. This result further supports the observation of low similarity between synthetic images from PSAQ-ViT and real images. For instance, for Class 1, the intra-class similarity within real images is 0.68, while PSAQ-ViT achieves only 0.44. + +The semantic inadequacy issue refers to the presence of dull regions in synthetic images, which contain redundant or non-semantic content [33, 37], hindering the model's learning process. As indicated in [13, 14], a diverse content and textures generally suggests rich information. As shown in Fig. 1b, many regions of synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit a lack of diversity in content, with overly simplified textures. Specifically, the central region of PSAQ-ViT images only contains faint object structures, while PSAQ-ViT V2 images appear excessively smoothed and indistinct. + +For high-bit quantization, where model capacity is largely retained [42], the performance degradation remains relatively minor even using semantic distorted and inadequate images [47, 62]. However, in low-bit quantization, where model capacity is severely damaged and informative images are essential for recovering performance [13, 84], + +fine-tuning on these poor-quality images leads to inferior generalization to real datasets, resulting in limited performance. For example, as shown in Tab. 2, the W4A4 ViT-B fine-tuned on real images yields $68.16\%$ , whereas PSAQ-ViT only achieves $36.32\%$ . + +![](images/a76c498bc5d105f8c3d2cd2c8ad6b212c421d9237e70e6e0570cd8cd08d8f7fd.jpg) +Figure 3. Comparison between attention maps. + +# 3.3. Semantics Alignment and Reinforcement Data-Free Quantization + +In the following, we introduce the proposed SARDFQ, whose framework is illustrated in Fig. 2. + +# 3.3.1. Attention Priors Alignment + +In ViTs, the self-attention mechanism encodes semantic correlations between image regions, where high-response areas in attention maps strongly correlate with semantic-discriminative content [10, 20, 31]. However, existing DFQ methods overlook this intrinsic property in the generation process. As a result, as shown in Fig. 3, synthetic images often exhibit disordered and unnatural attention patterns, with attention maps either overly diffuse or misaligned toward peripheral regions. This undermines their ability to preserve semantic-discriminative content, causing semantic distortion, as demonstrated in Fig. 1a and Tab. 1. In response, we propose Attention Priors Alignment (APA), which improves semantics alignment by optimizing synthetic images to follow randomly generated structure attention priors. + +Specifically, given a synthetic image $\tilde{I}$ , we first obtain its attention maps in the $h$ -th head of the $l$ -th block, denoting as $\mathbf{A}_{l,h} \in \mathbb{R}^{N \times N}$ , where $N$ represents the total number of tokens. In DeiT, the attention of the classification token toward other tokens serves as the indicator for semantic versus non-semantic parts [5]. Thus, we extract $\mathbf{A}_{l,h}^c \in \mathbb{R}^{1 \times (N - 1)}$ from $\mathbf{A}_{l,h}$ , representing the attention of the classification token to all tokens except itself. We then randomly generate attention priors $\tilde{\mathbf{A}}_{l,h}$ , whose generation is detailed in the next part, and align $\mathbf{A}_{l,h}^c$ with $\tilde{\mathbf{A}}_{l,h}$ by: + +$$ +\mathcal {L} _ {l, h} (\tilde {\boldsymbol {I}}) = \operatorname {M S E} \left(\mathbf {A} _ {l, h} ^ {c} - \tilde {\mathbf {A}} _ {l, h}\right), \tag {6} +$$ + +where MSE represents the mean squared error. For Swin models that do not use a classification token, we substitute $\mathbf{A}_{l,h}^{c}$ in Eq.6 with the average attention map of all tokens [10]. As noted in [17], ViTs initially focus on all regions to capture low-level information in shallow blocks + +![](images/47a751ddad2314c95508a2a12f85e247727bb9004ac501eed58fe564b277eaf4.jpg) +Figure 4. Examples of generated attention priors. + +![](images/3c735d92212b727d406c580060dcabab8f5fe18157239ef6f24724c4a4698a16.jpg) + +![](images/e9511c52c1df3beaa3c5e411286edf72afd6a1e11835c3eaf0e2f6b22a8878de.jpg) + +![](images/0d1652aa491f07440ea2ee43d6f9c93ab483199aba85c303bd36c784fca49c7c.jpg) + +and gradually shift their focus toward semantic regions in deeper blocks to extract high-level semantic information. Leveraging this property, we selectively apply $\mathcal{L}_{l,h}^{\mathrm{APA}}$ to deeper blocks, progressively aligning attention towards semantically relevant areas. The total APA loss is computed as a depth-weighted sum of the individual Eq. 6 across these deeper blocks: + +$$ +\mathcal {L} ^ {\mathrm {A P A}} (\tilde {\boldsymbol {I}}) = \sum_ {l = S} ^ {L} \sum_ {h = 1} ^ {H} \frac {l}{L} \mathcal {L} _ {l, h} (\tilde {\boldsymbol {I}}), \tag {7} +$$ + +where $S$ is a pre-given hyper-parameter denoting the start of deep blocks and are experimentally set $S = \frac{L}{2}$ . + +Attention Priors Generation. To generate attention priors $\tilde{\mathbf{A}}_{l,h}$ , Gaussian Mixture Models (GMMs) are employed as it is the most commonly used distribution with high flexibility. ViTs utilize different attention heads to capture diverse patterns and learn varied, informative representations [8]. Thus, we use distinct GMMs for each head. Note that the goal here is not to replicate real attention maps precisely, but to generate simulated structure attention priors to guide the learning of synthetic images. + +In particular, we first initialize an all zero matrix $\tilde{\mathbf{P}}\in$ $\mathbb{R}^{H\times W}$ , where $H = W = \sqrt{N - 1}$ for DeiT, $H =$ $W = \sqrt{N}$ for Swin. For example, for DeiT-S, the $H =$ $W = \sqrt{196} = 14$ . Then, we generate $k$ two-dimensional Gaussian distributions, where $k$ is randomly sampled from $1\sim K_{APA}$ and $K_{APA}$ is set to 5 in all experiments. Each Gaussian has its mean and covariance3. Consequently, the matrix element at the $i$ -th row and $j$ -th column, $\tilde{\mathbf{P}} [i,j]$ , is determined by: + +$$ +\tilde {\mathbf {P}} [ i, j ] = \max _ {m = 1, \dots , k} \mathbf {G} ^ {m} [ i, j ]. \tag {8} +$$ + +Then, $\tilde{\mathbf{P}}$ is normalized by: + +$$ +\tilde {\mathbf {P}} _ {n} = \frac {\tilde {\mathbf {P}}}{\sum \tilde {\mathbf {P}}} \cdot (1 - x). \tag {9} +$$ + +Here, for DeiT that incorporates the classification token, $x$ is randomly sampled from a uniform distribution $U(0,1)$ , representing the proportion of the attention score that the classification token allocates to itself. For Swin, which does not use a classification token, $x$ is set to 0. Finally, $\tilde{\mathbf{P}}_n$ is flatten to match the dimensionality: + +$$ +\tilde {\mathbf {A}} _ {l, h} = \operatorname {f l a t t e n} \left(\tilde {\mathbf {P}} _ {n}\right). \tag {10} +$$ + +Fig. 4 displays examples of generated attention priors. + +Discussion. APA prevents attention disorder, ensuring that synthetic images exhibit more coherent and natural attention patterns, as demonstrated in Fig. 3. As a result, APA selectively enhances the responses in certain regions, effectively emphasizing semantic-discriminative regions within synthetic images and thus prompting the discriminative features. Although simple, APA enables synthetic images to align better with the real semantics, as validated by both visual and quantitative evaluations. As shown in Fig. 1a, compared to PSAQ-ViT, features obtained after applying the APA loss are more closely aligned with real images, indicating better semantic alignment. The quantitative results in Tab. 1 further support that APA achieves superior semantic alignment. For example, in Class 1, the intra-class similarity for PSAQ-ViT is 0.44, whereas APA achieves a higher similarity of 0.64. + +# 3.3.2. Multi-Semantic Reinforcement + +Current DFQ methods for ViTs [47, 48, 62] optimize images through global optimization, treating the entire image as a single semantic unit. However, this is affected by low-rank structural regularity [34], where adjacent pixels exhibit strong similarity, leading to dull regions with redundant or non-semantic content for model learning [37]. This issue is further exacerbated by the tokenization mechanism in ViTs, as processing images in fixed-size patches make dull regions increase at the patch level [19, 70]. Consequently, as shown in Fig. 1b, synthetic images generated by existing methods exhibit large dull regions, resulting in semantic inadequacy [37]. In response, we propose Multi-Semantic Reinforcement (MSR), which applies localized patch optimization to enhance semantic richness by learning local patches with distinct semantics. + +Specifically, for a synthetic image $\tilde{I}$ , instead of feeding only the entire image, we also feed its patches and optimize them individually. Initially, we select $m$ nonoverlapping patches, where $m$ is chosen randomly from the set $\{1,2,\dots,K_{MSR}\}$ , with $K_{MSR}$ set to 4 in all experiments. These $m$ patches are then cropped and resized to match the model's input dimensions: + +$$ +\{\tilde {\boldsymbol {I}} _ {M S R} ^ {i} \} _ {i = 1, \dots , m} = \operatorname {r e s i z e} \left(\operatorname {c r o p} _ {m} (\tilde {\boldsymbol {I}})\right), \tag {11} +$$ + +where $\text{crop}_m(\cdot)$ crops $m$ non-overlapping patches from input, and $\text{resize}(\cdot)$ is the resize function. Each patch, denoted as $\tilde{I}_{MSR}^i$ , is treated as a new image with an assigned semantic target $c^i$ . Note that the gradient is backpropagated only to update the corresponding patch in the original image, leaving the rest of the image unaffected. + +Softlabel Learning. The one-hot loss (Eq. 3) only learns the semantics of the target class, making it unsuitable for $\tilde{\pmb{I}}$ under MSR, as its patches $\{\tilde{I}_{MSR}^i\}_{i = 1,\dots ,m}$ contain distinct semantics. In response, we propose Softlabel Learning + +(SL), which applies multiple semantic targets to accommodate the learning of images augmented by MSR. Specifically, we first sample $Z \in \mathbb{R}^{C} \sim U(0,1)$ , then modify its values by: + +$$ +\left\{ \begin{array}{l l} Z [ c ^ {i} ] \sim U (\epsilon_ {1}, \epsilon_ {2}), & \text {f o r} \tilde {\boldsymbol {I}} _ {M S R} ^ {i}, \\ Z [ c ^ {1}, \ldots , c ^ {m} ] \sim U (\epsilon_ {1}, \epsilon_ {2}), & \text {f o r} \tilde {\boldsymbol {I}}, \end{array} \right. +$$ + +where $U(\epsilon_1, \epsilon_2)$ denotes the uniform distribution over the interval $[\epsilon_1, \epsilon_2]$ , $m$ is the number of patches determined in MSR, and $\epsilon_1$ and $\epsilon_2$ control the softness, both empirically set consistently to 5 and 10 in all experiments. The soft target is defined as $T_s = \mathrm{softmax}(Z)$ , and SL loss is: + +$$ +\mathcal {L} ^ {\mathrm {S L}} \left(\tilde {I} / \tilde {I} _ {M S R} ^ {i}\right) = S C E \left(F \left(\tilde {I} / \tilde {I} _ {M S R} ^ {i}, T _ {s}\right), \right. \tag {12} +$$ + +where $SCE(\cdot, \cdot)$ is soft cross entropy and $F(\cdot)$ returns the predicted probability for its input. SL facilitates smooth learning between across semantic targets, ensuring that MSR-enhanced images receive consistent supervision rather than conflicting supervision. + +Discussion. By leveraging localized patch optimization, MSR ensures that each patch contributes unique semantics, forcing synthetic images to capture diverse features rather than being dominated by large homogeneous dull regions. As demonstrated by the richer and more diverse content and textures in Fig. 1b, MSR effectively reduces dull regions and enhances semantic richness in synthetic images. Moreover, MSR transforms synthetic images into composites of multiple semantic objects rather than a single unit, thereby providing more distinct semantic samples, i.e., $\{\tilde{I}_{MSR}^i\}_{i = 1,\dots ,m}$ , for model training. Unlike traditional cropping used in data augmentation, which aims to improve classification robustness by training with cropped patches labeled with the original class, MSR aims for semantic richness within synthetic images, ultimately enabling accurate data-free quantization. + +# 3.4. Overall Pipeline + +The overall pipeline consists of two stages: data synthesis and quantized network learning. The first stage uses the proposed SARDFQ to produce synthetic images. The second stage fine-tunes the quantized model using the generated synthetic images. + +# 3.4.1. Data Synthesis + +In the data synthesis stage, we combine the proposed APA loss of Eq. 7, SL loss of Eq. 12, and TV loss of Eq. 4 to formulate the objective function as follows: + +$$ +\mathcal {L} _ {G} (\tilde {\boldsymbol {I}}) = \alpha_ {1} \mathcal {L} ^ {\mathrm {A P A}} (\tilde {\boldsymbol {I}}) + \mathcal {L} ^ {\mathrm {S L}} (\tilde {\boldsymbol {I}}) + 0. 0 5 \mathcal {L} ^ {\mathrm {T V}} (\tilde {\boldsymbol {I}}). \tag {13} +$$ + +where $\alpha_{1}$ is hyperparameters and is determined by gridsearch. Note that the weight of TV loss is fixed to 0.05, following [47], to avoid a cumbersome hyperparameter search. + +# 3.4.2. Quantized Network Learning + +Recently DFQ methods have introduced the PTQ methods in learning quantized models due to their advantages of speed, memory efficiency, and performance [35, 62]. Thus, following the success of [42, 71], we fine-tune the quantized network block-wisely. Specifically, denote $\mathbf{X}_l$ as the outputs of the $l$ -th block of the full-precision model, and $\bar{\mathbf{X}}_l$ represent outputs of the quantized counterpart. The reconstruction loss is defined as: + +$$ +\mathcal {L} _ {l} = \left\| \mathbf {X} _ {l} - \bar {\mathbf {X}} _ {l} \right\| _ {2}. \tag {14} +$$ + +Here, $\mathcal{L}_l$ is only backward to update weights within $l$ -th block. Note that for a fair comparison, all compared methods adopt the same quantized network learning stage. + +# 4. Experiment + +# 4.1. Implementation Details + +Models and Tasks. We evaluate the performance of SARDFQ by test quantized ViT-S/B [20], DeiT-T/S/B [68], and Swin-S/B [54] on the classification task using ImageNet [63]. The pre-trained models are downloaded from the timm library. In appendix, we further provide results on detection and segmentation tasks. + +Comparison methods. We compare our SARDFQ against Gaussian noise, real images, and previous methods including SMI [33] and PSAQ-ViT [47] and its subsequent version, PSAQ-ViT V2 [48]. For a fair comparison, we generate synthetic images using their methods and apply our quantized network learning strategy. We use the official code for SMI and PSAQ-ViT to reproduce their images, while PSAQ-ViT V2 is re-implemented by us, as no official code is available. + +Experimental settings. All experiments were conducted using the PyTorch framework [60] on a single NVIDIA 3090 GPU. In the data synthesis stage, synthetic images were initialized with standard Gaussian noise, generating 32 images in total. The Adam optimizer [38] with $\beta_{1} = 0.5$ , $\beta_{2} = 0.9$ was used, with learning rates of 0.25 for Swin and 0.2 for others, and a total of 1,000 iterations. For all models, $K_{APA}$ , $K_{MSR}$ , $\epsilon_{1}$ , and $\epsilon_{2}$ were set to 5, 4, 5, and 10, respectively, based on a search with W4A4 DeiT-S. The value of $\alpha_{1}$ was determined by grid search for each model: 1e5 for DeiT-T/S, 1e4 for DeiT-B, 100 for ViT-B and Swin-B, 10 for Swin-S, and 1 for ViT-S. Although further hyperparameter search may improve performance, the current settings already yield superior results. In the quantized network learning stage, following [85], the Adam optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ was used, with weight decay set to 0 and an initial learning rate of 4e-5, adjusted + +
ModelW/ARealGaussian noisePSAQ-ViT [47]PSAQ-ViT V2 [48]SMI [33]SARDFQ (Ours)
ViT-S (81.39)4/466.576.0247.2441.5324.3329.4150.32
5/576.6936.7771.5968.4161.3365.1974.31
6/679.4661.2077.2074.7672.9572.4678.40
ViT-B (84.54)4/468.160.1536.3226.3235.2719.6751.84
5/579.214.1668.4867.9567.5357.1370.70
6/681.8955.1876.6571.8776.3369.8279.16
DeiT-T (72.21)4/456.6017.4347.7530.2030.1413.1852.06
5/567.0943.4964.1055.1656.4439.3566.41
6/669.8156.2368.3762.7764.0344.3969.73
DeiT-S (79.85)4/468.4620.8958.2845.5342.7711.7162.29
5/575.0641.0671.9063.1462.8829.1374.06
6/677.8765.6375.8568.8571.6537.6977.31
DeiT-B (81.85)4/477.0747.2071.7566.4365.3359.0472.17
5/579.8665.4678.4576.7776.7475.3378.72
6/680.9062.7980.0079.2278.8177.6680.15
Swin-S (83.20)4/478.1231.9273.1965.5565.8574.74
5/580.5152.1078.1574.3775.4179.56
6/680.6065.6679.7478.5078.2580.56
Swin-B (85.27)4/478.8030.1471.8467.4265.2376.42
5/582.5135.2878.5077.2075.2580.82
6/682.6467.3782.0081.4180.3083.03
+ +Table 2. Quantization results on ImageNet dataset, with top-1 accuracy $(\%)$ reported. The performance of the full-precision model is listed below the model name. "W/A" denotes the bit-width of weights/activations. "Real" refers to using real images. For SMI [33], we provide the performance of using dense (normal-sized numbers) and sparse (smaller-sized numbers) synthetic images, respectively. Note that for Swin models, we do not provide the results for sparse synthetic images as the sparse generation method of SMI is infeasible. + +via cosine decay for 100 iterations. A channel-wise quantizer was used for weights, and a layer-wise quantizer for activations, with all matrix multiplications in ViTs quantized [47-49]. + +# 4.2. Quantization Results + +The quantization results are presented in Tab.2. Our SARDFQ demonstrates consistent improvements across various quantization bit-width configurations, particularly with low bit-width settings. Specifically, for ViT-S, SARDFQ improves the performance by $3.08\%$ in the W4/A4 setting, $2.72\%$ in the W5/A5 setting, and $1.20\%$ in the W6/A6 setting. For ViT-B, SARDFQ achieves performance gains of $15.52\%$ in the W4/A4 setting, $2.22\%$ in the W5/A5 setting, and $2.51\%$ in the W6/A6 setting. Results on DeiT also demonstrate the effectiveness of SARDFQ. For example, on DeiT-T, SARDFQ shows a marked improvement by increasing top-1 accuracy by $4.31\%$ in the W4/A4 setting, $2.31\%$ in the W5/A5 setting, and $1.36\%$ in the W6/A6 setting. For DeiT-S, SARDFQ enhances top-1 accuracy by $4.01\%$ in the W4/A4 setting, $2.16\%$ in the W5/A5 setting, and $1.46\%$ in the W6/A6 setting. The quan + +tization results of Swin-S/B also affirm the superiority of our SARDFQ in enhancing model accuracy under different quantization configurations. In particular, for Swin-S, the proposed SARDFQ increases the accuracy by $1.55\%$ for the W4/A4 setting, $1.41\%$ for the W5/A5 setting, and $0.82\%$ for the W6/A6 setting, respectively. When it comes to Swin-B, the proposed SARDFQ increases the accuracy by $4.58\%$ for the W4/A4 setting, $2.32\%$ for the W5/A5 setting, and $1.03\%$ for the W6/A6 setting, respectively. + +# 4.3. Ablation Study + +All ablation studies are conducted on the W4A4 DeiT-S. + +Analysis of APA, MSR, and SL. We analyze the effectiveness of the proposed APA (Sec.,3.3.1), MSR (Sec.,3.3.2), and SL (Eq.,12) in Tab.,3. Adding APA and SL individually to the baseline increases accuracy. Notably, APA boosts performance from $51.73\%$ to $60.26\%$ , confirming its effectiveness in aligning semantics (Sec.,3.3.1). Applying MSR alone slightly decreases accuracy from $51.73\%$ to $50.75\%$ , indicating that one-hot loss is unsuitable for MSR-augmented synthetic images. However, when both MSR and SL are applied, accuracy rises to $56.08\%$ , sug- + +
APAMSRSLAcc. (%)
Baseline51.73
✓60.26
✓50.75
✓52.02
✓✓61.58
✓✓60.51
✓✓56.08
✓✓✓62.29
+ +gesting SL is more compatible with MSR than one-hot loss. Combining APA with either MSR or SL further improves performance. For example, APA and SL together yield an accuracy of $60.51\%$ , and when all three strategies are used, the best performance of $62.29\%$ is achieved. + +Analysis of priors distribution. Tab. 4a showcases the results of using other distributions to formulate the attention priors. The unevenly distributed GMM and Laplace present comparable performance of $62.29\%$ and $62.16\%$ , respectively. Moreover, GMM provides a similar performance to the real's, indicating it performs well in imitating the patterns of real images. + +![](images/a9648db1c30d9fa309266ebee1ec85a0f5984e694d58af35381fa07d3dce66ed.jpg) +(a) +Figure 5. Effect of varying (a) $\alpha_{1}$ and (b) $K_{MSR}$ . + +![](images/3e1333dc94a67c2dbdb0700a59515b6c89c5e3a10c0ae3c609fbbe5e2a6d3f60.jpg) +(b) + +Analysis of $\alpha_{1}, K_{APA}$ , and $K_{MSR}$ . The $\alpha_{1}$ from Eq. 13 balance the importance of the proposed APA loss during the update of the synthetic images. Fig. 5a demonstrates that the optimal performance is achieved when $\alpha_{1} = 1e5$ . Incrementally increasing $\alpha_{1}$ improves performance up to $62.29\%$ at $\alpha_{1} = 1e5$ . However, further increases in $\alpha_{1}$ subsequently degrade performance. The $K_{APA}$ in APA is the upper limit on the number of the Gaussian distributions used for priors generation. Tab. 4b displays the ablation study for different values of $K_{APA}$ . The best accuracy is achieved when $K_{APA} = 5$ . The $K_{MSR}$ in MSR is the upper limit on the number of patches. Fig. 5b demonstrates that when $K_{MSR} = 4$ , the optimal performance is achieved. Using $K_{MSR}$ larger than 4 will hurt the accuracy. We consider this + +Table 3. Influence of the proposed APA, MSR, and SL on accuracy. The baseline adopts the one-hot loss. Applying APA, MSR, and SL yields SARDFQ. + +
Priors DistributionTop-1
GMM62.29
Laplace62.16
Real63.19
+ +(a) + +
KAPATop-1
161.13
361.52
562.29
761.53
961.05
+ +(b) + +
STop-1
061.96
L/262.29
+ +(c) + +
w. l/LTop-1
✓62.29
×61.32
+ +(d) + +Table 4. Effect of varying (a) priors types; (b) $K_{APA}$ ; (c) $S$ ; (d) $\frac{l}{L}$ . + +due to limited patch resolution if using a too large $K_{MSR}$ . + +Analysis of APA loss. Here, we conduct the ablation study by considering the $S$ and scale $\frac{l}{L}$ in Eq.7. Tab.4c presents the effect of varying $S$ in Eq.7. If applying APA loss to all blocks $(S = 0)$ , the top-1 accuracy decreases to $61.96\%$ . From Tab.4d, it can be seen that absorbing the scale $\frac{l}{L}$ in Eq.7 presents $0.97\%$ performance gains. + +# 5. Limitations + +We further discuss some limitations of the proposed SARDFQ, which will guide future research directions. First, although SARDFQ shows substantial performance improvement, a performance gap between SARDFQ and real data remains challenging, highlighting the need for a stronger semantics alignment and reinforcement method. Second, SARDFQ currently lacks a theoretical foundation. Future work could establish a theoretical framework for SARDFQ, particularly in understanding how APA and MSR influence synthetic images in a formalized manner. + +# 6. Conclusion + +In this paper, we investigate the DFQ method for ViTs. We first identify that synthetic images generated by existing methods suffer from semantic distortion and inadequacy issues, and propose SARDFQ to address these issues. To mitigate semantic distortion, SARDFQ introduces APA, which guides synthetic images to align with randomly generated structural attention patterns. To tackle semantic inadequacy, SARDFQ incorporates MSR. MSR optimizes different regions of synthetic images with unique semantics, thereby enhancing overall semantic richness. Moreover, SARDFQ employs SL, which adopts multiple semantic targets to ensure seamless learning of images augmented by MSR. Extensive experiments on various ViT models and tasks validate the effectiveness of SARDFQ. + +# 7. Acknowledgments + +This work was supported by National Science and Technology Major Project (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (Np. 624B2119, No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001). + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 6836-6846, 2021. 1, 2 +[2] Jianhong Bai, Yuchen Yang, Huanpeng Chu, Hualiang Wang, Zuozhu Liu, Ruizhe Chen, Xiaoxuan He, Lianrui Mu, Chengfei Cai, and Haoji Hu. Robustness-guided image synthesis for data-free quantization. arXiv preprint arXiv:2310.03661, 2023. 3 +[3] Ron Banner, Yury Nahshan, Daniel Soudry, et al. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems (NeurIPS), pages 7950-7958, 2019. 2 +[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(8):1798-1828, 2013. 4 +[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 4 +[6] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13169-13178, 2020. 2, 3 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 1, 2 +[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 397-406, 2021. 5 +[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12299-12310, 2021. 2 +[10] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general + +coarse-to-fine method for vision transformer. In AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 4 +[11] Mengzhao Chen, Wenqi Shao, Peng Xu, Mingbao Lin, Kaipeng Zhang, Fei Chao, Rongrong Ji, Yu Qiao, and Ping Luo. Diffrate: Differentiable compression rate for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17164-17174, 2023. 2 +[12] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2021. 2 +[13] Xinrui Chen, Yizhi Wang, Renao Yan, Yiqing Liu, Tian Guan, and Yonghong He. Texq: Zero-shot network quantization with texture feature distribution calibration. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4, 1 +[14] Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. In Advances in Neural Information Processing Systems (NeurIPS), pages 14835-14847, 2021. 2, 4, 1 +[15] Kanghyun Choi, Hye Yoon Lee, Deokki Hong, Joonsang Yu, Noseong Park, Youngsok Kim, and Jinho Lee. It's all in the teacher: Zero-shot quantization brought closer to the teacher. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8311-8321, 2022. 3 +[16] Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, and Jinho Lee. Mimiq: Low-bit data-free quantization of vision transformers. arXiv preprint arXiv:2407.20021, 2024. 2, 3 +[17] Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-attention and convolutional layers. In International Conference on Learning Representations (ICLR), 2020. 4 +[18] Yifu Ding, Haotong Qin, Qinghua Yan, Zhenhua Chai, Junjie Liu, Xiaolin Wei, and Xianglong Liu. Towards accurate posttraining quantization for vision transformer. In 30th ACM International Conference on Multimedia (ACMMM), pages 5380-5388, 2022. 3 +[19] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International conference on machine learning (ICML), pages 2793-2803. PMLR, 2021. 5 +[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2, 4, 6 +[21] Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero-shot quantization. In International Conference on Machine Learning (ICML), 2024. 2 +[22] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha. + +Learned step size quantization. In International Conference on Learning Representations (ICLR), 2020. 2 +[23] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), 2021. 2 +[24] Natalia Frumkin, Dibakar Gope, and Diana Marculescu. Jumping through local minima: Quantization in the loss landscape of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16978-16988, 2023. 3 +[25] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4852–4861, 2019. 2 +[26] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), pages 2672–2680, 2014. 3 +[27] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quantization via diagonal hessian approximation. In The Eleventh International Conference on Learning Representations (ICLR), 2022. 2, 3 +[28] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[29] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):87-110, 2022. 1 +[30] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2 +[31] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 773-783, 2023. 4 +[32] Zejiang Hou and Sun-Yuan Kung. Multi-dimensional vision transformer compression via dependency guided gaussian process search. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3669-3678, 2022. 2 +[33] Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Lei Li, Chun Yuan, and Dacheng Tao. Sparse model inversion: Efficient inversion of vision transformers for data-free applications. In International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 6, 7, 1 + +[34] Jinggang Huang and David Mumford. Statistics of natural images and models. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 541-547. IEEE, 1999. 5 +[35] Yongkweon Jeon, Chungman Lee, and Ho-young Kim. Genie: show me the data for quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12073, 2023. 3, 6 +[36] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM Computing Surveys (CSUR), 2021. 1 +[37] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning (ICML), pages 11102-11118. PMLR, 2022. 4, 5 +[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 6 +[39] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 2 +[40] Huantong Li, Xiangmiao Wu, Fanbing Lv, Daihai Liao, Thomas H Li, Yonggang Zhang, Bo Han, and Mingkui Tan. Hard sample matters a lot in zero-shot quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24417-24426, 2023. 3 +[41] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In International Conference on Learning Representations (ICLR), 2020. 2 +[42] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations (ICLR), 2021. 2, 4, 6 +[43] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 34451-34463, 2022. 2 +[44] Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Stableq: Enhancing data-scarce quantization with text-to-image data. arXiv preprint arXiv:2312.05272, 2023. 3 +[45] Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, and Priyadarshini Panda. Genq: Quantization in low data regimes with generative synthetic data. In European Conference on Computer Vision (ECCV), pages 216-235. Springer, 2024. 3 +[46] Zhikai Li and Qingyi Gu. I-vit: Integer-only quantization for efficient vision transformer inference. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17065-17075, 2023. 2, 3 +[47] Zhikai Li, Liping Ma, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for + +vision transformers. In European Conference on Computer Vision (ECCV), pages 154-170. Springer, 2022. 1, 2, 3, 4, 5, 6, 7 +[48] Zhikai Li, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2023. 2, 3, 5, 6, 7, 1 +[49] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17227-17236, 2023. 3, 7 +[50] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 1833-1844, 2021. 2 +[51] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), pages 1173–1179, 2022. 3 +[52] Shih-Yang Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning (ICML), pages 21813-21824, 2023. 2 +[53] Yijiang Liu, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20321-20330, 2023. 3 +[54] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF international conference on computer vision (ICCV), pages 10012-10022, 2021. 2, 6 +[55] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 28092-28103, 2021. 2, 3 +[56] Yiwei Ma, Jiayi Ji, Xiaoshuai Sun, Yiyi Zhou, and Rongrong Ji. Towards local visual modeling for image captioning. Pattern Recognition, 138:109420, 2023. 1 +[57] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations (ICLR), 2022. 2 +[58] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 23296-23308, 2021. 4 +[59] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 3163-3172, 2021. 2 + +[60] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), pages 8026-8037, 2019. 6 +[61] Biao Qian, Yang Wang, Richang Hong, and Meng Wang. Adaptive data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7960-7968, 2023. 3 +[62] Akshit Ramachandran, Souvik Kundu, and Tushar Krishna. Clamp-vit: contrastive data-free learning for adaptive posttraining quantization of vits. In European Conference on Computer Vision (ECCV). Springer, 2024. 2, 3, 4, 5, 6, 1 +[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115:211-252, 2015. 6 +[64] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transformers in medical imaging: A survey. Medical Image Analysis, page 102802, 2023. 2 +[65] Huixin Sun, Runqi Wang, Yanjing Li, Xianbin Cao, Xiaolong Jiang, Yao Hu, and Baochang Zhang. P4q: Learning to prompt for quantization in visual-language models. arXiv preprint arXiv:2409.17634, 2024. 2 +[66] Yehui Tang, Yunhe Wang, Yixing Xu, Yiping Deng, Chao Xu, Dacheng Tao, and Chang Xu. Manifold regularized dynamic network pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5028, 2021. 2 +[67] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12165-12174, 2022. 2 +[68] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357. PMLR, 2021. 1, 2, 6 +[69] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9:2579-2605, 2008. 1 +[70] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 5 +[71] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. In International Conference on Learning Representations (ICLR), 2022. 6 +[72] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In IEEE/CVF In + +ternational Conference on Computer Vision (ICCV), pages 10033-10041, 2021. 2 +[73] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision (ECCV), pages 68-85, 2022. 2 +[74] Zhiqiang Shen Xijie Huang and Kwang-Ting Cheng. Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331, 2023. 2 +[75] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17, 2020. 3 +[76] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17. Springer, 2020. 2 +[77] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transffer via deep-inversion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8715-8724, 2020. 2, 3, 1 +[78] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015. 4 +[79] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 818-833, 2014. 4 +[80] Jinnian Zhang, Houwen Peng, Kan Wu, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Minivit: Compressing vision transformers with weight multiplexing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12145-12154, 2022. 2 +[81] Xiangguo Zhang, Haotong Qin, Yifu Ding, Ruihao Gong, Qinghua Yan, Renshuai Tao, Yuhang Li, Fengwei Yu, and Xianglong Liu. Diversifying sample generation for accurate data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15658-15667, 2021. 2, 3 +[82] Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, and Yunhe Wang. Less is more: Focus attention for efficient detr. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6674-6683, 2023. 2 +[83] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 6881-6890, 2021. 2 +[84] Yunshan Zhong, Mingbao Lin, Gongrui Nan, Jianzhuang Liu, Baochang Zhang, Yonghong Tian, and Rongrong Ji. Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12339-12348, 2022. 2, 4, 1 + +[85] Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&S-vit: An inclusive & stable method for pushing the limit of post-training vits quantization. arXiv preprint arXiv:2311.10126, 2023. 6 +[86] Yunshan Zhong, Jiawei Hu, You Huang, Yuxin Zhang, and Rongrong Ji. Erq: Error reduction for post-training quantization of vision transformers. In International Conference on Machine Learning (ICML), 2024. 3 \ No newline at end of file diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/images.zip b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4dee51d2a21e48b33b6a903da6bbc3a3524e87ba --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ebf1979a465a0550ce85f555f5d2c94cbec6158a57443841286ba94f1c4c923 +size 427558 diff --git a/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/layout.json b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47b0fa75ca0b656550c55a4eab4d6f405598e4ca --- /dev/null +++ b/2025/Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers/layout.json @@ -0,0 +1,12437 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 544, + 138 + ], + "type": "text", + "content": "Semantic Alignment and Reinforcement for Data-Free Quantization of Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "spans": [ + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": "Yunshan Zhong" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Yuyao Zhou" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Yuxin Zhang" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Wanchen Sui" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Shen Li" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Yong Li" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Fei Chao" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "text", + "content": ", Rongrong Ji" + }, + { + "bbox": [ + 100, + 160, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 174, + 189, + 436, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 189, + 436, + 203 + ], + "spans": [ + { + "bbox": [ + 174, + 189, + 436, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 174, + 189, + 436, + 203 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Xiamen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "type": "text", + "content": "MAC Lab, School of Informatics, Xiamen University " + }, + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 131, + 202, + 479, + 217 + ], + "type": "text", + "content": "Alibaba Group" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 220, + 511, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 220, + 511, + 247 + ], + "spans": [ + { + "bbox": [ + 96, + 220, + 511, + 247 + ], + "type": "text", + "content": "viper.zhong@gmail.com, {yuyaozhou, yuxinzhang}@stu.xmu.edu.cn \n{wanchen.swc, litan.ps, jiufeng.ly}@alibaba-inc.com, {fchao, rrji}@xmu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 300, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 612 + ], + "type": "text", + "content": "Data-free quantization (DFQ) enables model quantization without accessing real data, addressing concerns regarding data security and privacy. With the growing adoption of Vision Transformers (ViTs), DFQ for ViTs has garnered significant attention. However, existing DFQ methods exhibit two limitations: (1) semantic distortion, where the semantics of synthetic images deviate substantially from those of real images, and (2) semantic inadequacy, where synthetic images contain extensive regions with limited content and oversimplified textures, leading to suboptimal quantization performance. To address these limitations, we propose SARDFQ, a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs. To address semantic distortion, SARDFQ incorporates Attention Priors Alignment (APA), which optimizes synthetic images to follow randomly generated structure attention priors. To mitigate semantic inadequacy, SARDFQ introduces Multi-Semantic Reinforcement (MSR), leveraging localized patch optimization to enhance semantic richness across synthetic images. Furthermore, SARDFQ employs Soft-Label Learning (SL), wherein multiple semantic targets are adapted to facilitate the learning of multi-semantic images augmented by MSR. Extensive experiments demonstrate the effectiveness of SARDFQ, significantly surpassing existing methods. For example, SARDFQ improves top-1 accuracy on ImageNet by " + }, + { + "bbox": [ + 55, + 300, + 296, + 612 + ], + "type": "inline_equation", + "content": "15.52\\%" + }, + { + "bbox": [ + 55, + 300, + 296, + 612 + ], + "type": "text", + "content": " for W4A4 ViT-B1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 658, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 682 + ], + "type": "text", + "content": "Vision Transformers (ViTs) [20, 29, 36] have demonstrated remarkable success across various computer vision tasks [1," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 352, + 273, + 518, + 411 + ], + "blocks": [ + { + "bbox": [ + 352, + 273, + 518, + 411 + ], + "lines": [ + { + "bbox": [ + 352, + 273, + 518, + 411 + ], + "spans": [ + { + "bbox": [ + 352, + 273, + 518, + 411 + ], + "type": "image", + "image_path": "4b50f3a45edef337a0f4073e0ff2e040144da215e4fb9126beb74c732643f2fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 488 + ], + "type": "text", + "content": "(a) The t-SNE[69] visualization of the penultimate-layer features (extracted by DeiT-S) of synthetic images. Each marker (circle, star, triangle) represents a distinct category. The red dashed circles highlight the features extracted from our APA and real images. Notably, the features produced by PSAQ-ViT[47] exhibit substantial deviation from those of real images, indicating semantic distortion. In contrast, our APA yields features that more closely align with those of real images, suggesting improved semantics alignment." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 351, + 489, + 407, + 590 + ], + "blocks": [ + { + "bbox": [ + 351, + 489, + 407, + 590 + ], + "lines": [ + { + "bbox": [ + 351, + 489, + 407, + 590 + ], + "spans": [ + { + "bbox": [ + 351, + 489, + 407, + 590 + ], + "type": "image", + "image_path": "249fb76fabd4d8d549bcfc3fb1bc6e46377268668c0690f604df88d19982974d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 590, + 388, + 598 + ], + "lines": [ + { + "bbox": [ + 367, + 590, + 388, + 598 + ], + "spans": [ + { + "bbox": [ + 367, + 590, + 388, + 598 + ], + "type": "text", + "content": "PSAQ" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 408, + 489, + 462, + 590 + ], + "blocks": [ + { + "bbox": [ + 408, + 489, + 462, + 590 + ], + "lines": [ + { + "bbox": [ + 408, + 489, + 462, + 590 + ], + "spans": [ + { + "bbox": [ + 408, + 489, + 462, + 590 + ], + "type": "image", + "image_path": "39086b1071ec1e2cac7df6a959b641a108c4ee0baa0e9e30a61a0d1ebef8644d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 590, + 450, + 598 + ], + "lines": [ + { + "bbox": [ + 418, + 590, + 450, + 598 + ], + "spans": [ + { + "bbox": [ + 418, + 590, + 450, + 598 + ], + "type": "text", + "content": "PSAQ V2" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 462, + 489, + 517, + 590 + ], + "blocks": [ + { + "bbox": [ + 462, + 489, + 517, + 590 + ], + "lines": [ + { + "bbox": [ + 462, + 489, + 517, + 590 + ], + "spans": [ + { + "bbox": [ + 462, + 489, + 517, + 590 + ], + "type": "image", + "image_path": "faf3b7d4f308588c04140baf283cbb66741b4e3f32e492455dcd0bf875558327.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 590, + 516, + 599 + ], + "lines": [ + { + "bbox": [ + 463, + 590, + 516, + 599 + ], + "spans": [ + { + "bbox": [ + 463, + 590, + 516, + 599 + ], + "type": "text", + "content": "SARDFQ (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 600, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 640 + ], + "type": "text", + "content": "(b) The images of PSAQ-ViT and PSAQ-ViT V2 exhibit numerous dull regions with limited content and simplified textures, reflecting semantic inadequacy. In comparison, our SARDFQ generates images with greater diversity in both content and texture, demonstrating enhanced semantics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 649, + 555, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 555, + 671 + ], + "type": "text", + "content": "Figure 1. Illustration of (a) semantic distortion and (b) semantic inadequacy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "7, 56, 68]. However, their high computational cost and substantial memory footprint hinder deployment in resource-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 205, + 703 + ], + "type": "text", + "content": "*Corresponding Author: rrji@xmu.edu.cn" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 703, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 274, + 712 + ], + "type": "text", + "content": "1The code is at https://github.com/zysxmu/SARDFQ." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12479" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "constrained environments [11, 30, 32, 46, 55, 67, 82]. To address these limitations, quantization [39] has emerged as a promising solution, which reduces model complexity by converting full-precision weight and activations into low-bit representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 335 + ], + "type": "text", + "content": "Traditional quantization methods typically require access to the original training dataset, which raises data privacy and security concerns [6, 27, 77, 84]. As a result, data-free quantization (DFQ) has gained increasing attention, allowing quantization without the need for real data [14, 76, 81]. However, most existing DFQ methods are designed specifically for convolutional neural networks (CNNs) and are not directly applicable to vision transformers (ViTs). These methods generally rely on batch normalization statistics (BNS), which capture the distribution of real data, to synthesize in-distribution synthetic data [6, 76, 81, 84]. Yet, BNS is unavailable for ViTs, which use layer normalization (LN) to dynamically compute distribution statistics during inference [47]. Recently, several DFQ methods have been proposed for ViTs [16, 33, 47, 48, 62]. For example, PSAQ-ViT [47] introduces patch similarity entropy (PSE) loss to optimize Gaussian noise towards usable synthetic images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 514 + ], + "type": "text", + "content": "Nevertheless, we observe that existing methods suffer from semantic distortion and inadequacy issues. As shown in Fig. 1a, features of synthetic images generated by PSAQ-ViT deviate significantly from those of real images. Tab. 1 shows that the cosine similarity between synthetic images from PSAQ-ViT and real images is notably low, also indicating significant distortion. These results highlight the issue of semantic distortion. Moreover, as shown in Fig. " + }, + { + "bbox": [ + 55, + 336, + 295, + 514 + ], + "type": "inline_equation", + "content": "1\\mathrm{b}^2" + }, + { + "bbox": [ + 55, + 336, + 295, + 514 + ], + "type": "text", + "content": " synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit many regions with limited content diversity and overly simplified textures. These low-quality dull regions are useless or even detrimental to model learning [33], highlighting the issue of semantic inadequacy. Consequently, quantized models trained on such low-quality images suffer from degraded performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 694 + ], + "type": "text", + "content": "Motivated by the above analysis, we propose a novel Semantics Alignment and Reinforcement Data-Free Quantization method for ViTs, termed SARDFQ. The overall framework is depicted in Fig.2. To address the semantic distortion issue, SARDFQ introduces Attention Priors Alignment (APA), where synthetic images are optimized to follow structured attention priors generated using Gaussian Mixture Models (GMMs). APA effectively aligns the semantics of synthetic images with real images, as validated by both visual and quantitative analyses. As shown in Fig. 1a, features of APA exhibit a closer alignment to those of real images, while quantitative results in Tab. 1 also confirm that the semantics of APA are more consistent to the real images, indicating enhanced semantics alignment. To address the semantic inadequacy issue, SARDFQ incorpo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "rates Multi-Semantic Reinforcement (MSR) and Softlabel Learning (SL). MSR utilizes localized patch optimization, which encourages different sub-patches of synthetic images to capture various semantics, reinforcing the rich semantics across images. SL applies multiple semantic targets to accommodate the learning of multi-semantic images augmented by MSR. As shown in Fig. 1b, synthetic images after applying MSR exhibit greater diversity in content and texture, providing reinforced semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 239 + ], + "type": "text", + "content": "Experimental results across various ViT models and tasks demonstrate that SARDFQ presents substantial performance improvements. For example, SARDFQ achieves a " + }, + { + "bbox": [ + 313, + 180, + 553, + 239 + ], + "type": "inline_equation", + "content": "15.52\\%" + }, + { + "bbox": [ + 313, + 180, + 553, + 239 + ], + "type": "text", + "content": " increase in top-1 accuracy on the ImageNet dataset for the W4A4 ViT-B model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 253, + 405, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 405, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 405, + 266 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 273, + 433, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 433, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 433, + 285 + ], + "type": "text", + "content": "2.1. Vision Transformers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 555 + ], + "type": "text", + "content": "The great success of transformers in the natural language processing field has driven widespread attempts in the computer vision community to apply them to vision tasks [12, 28, 72]. ViT [20] is the pioneer that builds a transformer-based model to handle images, boosting the performance on the image classification task. DeiT [68] introduces an efficient teacher-student training strategy where a distillation token is employed to distill knowledge from the teacher model to the student model. Swin Transformers [54] builds an efficient and effective hierarchical model by introducing a shifted window-based self-attention mechanism. Other than the image classification task, the applications of ViTs also have broadened considerably, manifesting groundbreaking performance in object detection [7], image segmentation [9, 83], low-level vision [50], video recognition [1, 59], and medical image processing [64], etc. Nevertheless, the impressive performance of ViTs relies on a high number of parameters and significant computational overhead, preventing deployment in resource-constrained environments. Several recent efforts design lightweight ViTs, such as MobileViT [57], MiniVit [80], and TinyViT [73]. However, the model complexity is still unsatisfactory [47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 441, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 441, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 441, + 576 + ], + "type": "text", + "content": "2.2. Network Quantization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Data-Driven Quantization. Model quantization reduces the complexity of neural networks by replacing full-precision weight and activation with the low-bit format. Data-driven quantization can be roughly divided into two categories: quantization-aware training (QAT) and posttraining quantization (PTQ). QAT is compute-heavy since it re-trains the quantized model with the full training data to retain performance [22, 25, 41, 43, 46, 52, 65, 74]. PTQ perform quantization with a tiny dataset and a reduced time overhead, harvesting widespread attention [3, 42]. The specific architecture of ViTs, such as LayerNorm and the self" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 211, + 713 + ], + "type": "text", + "content": "2More results are presented in the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12480" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 68, + 504, + 187 + ], + "blocks": [ + { + "bbox": [ + 105, + 68, + 504, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 68, + 504, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 68, + 504, + 187 + ], + "type": "image", + "image_path": "1b7b933fd4ed13c1c2693d8e289055bfe0cf694f3afb23c5f54be139c7e73d04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "type": "text", + "content": "Figure 2. SARDFQ Framework overview: Attention Priors Alignment (APA) employs randomly generated attention priors to improve semantics alignment. Multi-Semantic Reinforcement (MSR) learns the different regions of synthetic images with various semantics to enhance overall semantic richness. Meanwhile, Softlabel Learning (SL) adopts multiple semantic targets to ensure consistent learning of multi-semantic images augmented by MSR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 251, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 294, + 323 + ], + "type": "text", + "content": "attention module, urges distinct PTQ methods compared to CNNs [18, 24, 46, 51, 53, 86]. For example, Liu et al. [55] develop a ranking loss to maintain the relative order of the self-attention activation. Unfortunately, both QAT and PTQ involve the original training data, causing concerns about data privacy and security issues in data-sensitive scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 325, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 325, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 325, + 296, + 540 + ], + "type": "text", + "content": "Data-Free Quantization. DFQ quantizes models without accessing real data [2, 13-15, 27, 35, 40, 44, 45, 61]. Most previous DFQ methods focus on CNN, where the BNS can be adopted as the regularization term [6, 81]. However, BNS is infeasible for ViTs built on the LN. Recently, few efforts have been explored to accommodate ViTs [16, 33, 47, 48, 62]. PSAQ-ViT [47] introduces the first DFQ method for ViTs. They discover that Gaussian noise yields homogeneous patches, while the real image yields heterogeneous patches. Thus, patch similarity entropy (PSE) loss is proposed to optimize the Gaussian noise towards real-like images by making them showcase heterogeneous patches. Based on PSAQ-ViT, PSAQ-ViT V2 [48] further introduces an adversarial learning strategy [26]. [62] incorporates contrastive learning and proposes an iterative generation-quantization PTQ-based DFQ method. [33] proposes a sparse generation method to remove noisy and hallucination backgrounds in synthetic images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 553, + 111, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 111, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 111, + 566 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 141, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 141, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 141, + 586 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 132, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 132, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 132, + 604 + ], + "type": "text", + "content": "3.1.1.Quantizers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "content": "We employ the linear quantizer for all weights and activations, except for the attention scores, which use a log2 quantizer to handle highly non-negative and uneven values [24, 49, 51]. For the linear quantizer, given a full-precision input " + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "content": " and bit-width " + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "content": ", the quantized value " + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_q" + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "content": " and the de-quantized value " + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 608, + 295, + 680 + ], + "type": "text", + "content": " are computed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 689, + 295, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 689, + 295, + 712 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil + z, 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot \\left(\\mathbf {x} _ {q} - z\\right), \\tag {1}", + "image_path": "589453dbf02f79c533cfeacf49047b443f63cf943834d9113a67bf9c79840da2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "content": " denotes rounding to the nearest integer, and clip limits the value to " + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "inline_equation", + "content": "[0,2^{b} - 1]" + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 251, + 555, + 289 + ], + "type": "text", + "content": " are the scale factor and zero-point, respectively. For the log2 quantizer:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 297, + 555, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 322, + 297, + 555, + 319 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {q} = \\operatorname {c l i p} \\left(\\left\\lfloor - \\log_ {2} \\frac {\\mathbf {x}}{\\Delta} \\right\\rceil , 0, 2 ^ {b} - 1\\right), \\bar {\\mathbf {x}} = \\Delta \\cdot 2 ^ {- \\mathbf {x} _ {q}}. \\tag {2}", + "image_path": "384644b72cd93884760f545a9fd2a54001b4aea5aeca946093a73c840e6b95be.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 326, + 406, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 406, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 406, + 338 + ], + "type": "text", + "content": "3.1.2. Data Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "text", + "content": "DFQ methods parameterize synthetic images and optimize them toward real-like images with a pre-trained full-precision model " + }, + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "text", + "content": ". Given a image " + }, + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 341, + 554, + 399 + ], + "type": "text", + "content": " initialized from Gaussian noise, the one-hot loss [75] is introduced to learn label-related semantics:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 409, + 555, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 409, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 383, + 409, + 555, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {O H}} (\\tilde {\\boldsymbol {I}}) = C E (F (\\tilde {\\boldsymbol {I}}), c) \\tag {3}", + "image_path": "9532079ffa5eeea000b656a8635840d8997c15a3f12f14749b245dfe16959e44.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "CE(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": " represents the cross entropy, " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": " is a random class label, and " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": " returns the predicted probability for image " + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 434, + 554, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 470, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 495 + ], + "type": "text", + "content": "Moreover, total variance (TV) [77] loss is a smoothing regularization term to improve the image quality:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 502, + 555, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 502, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 361, + 502, + 555, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}) = \\iint | \\nabla \\tilde {\\boldsymbol {I}} (\\tau_ {1}, \\tau_ {2}) | d \\tau_ {1} d \\tau_ {2}. \\tag {4}", + "image_path": "31a48dd0768c129219f6eb8dc9f091b7e9fd57216a32cf1980757e3f09481dc8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "inline_equation", + "content": "\\nabla \\tilde{I} (\\tau_1,\\tau_2)" + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "text", + "content": " denotes the gradient at " + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 537, + 533, + 550 + ], + "type": "inline_equation", + "content": "(\\tau_{1},\\tau_{2})" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": "To perform DFQ for ViTs, PSAQ-ViT [47] proposes patch similarity entropy (PSE) loss. It first compute patch similarity " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\Gamma_{l}[i,j] = \\frac{u_{i} \\cdot u_{j}}{||u_{i}||||u_{j}||}" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "u_{i}, u_{j}" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " are feature vectors of MHSA outputs in " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": "-th block and " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "||\\cdot||" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " norm. Then, it estimates the density function " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\hat{f}_{l}(x) = \\frac{1}{Mh}\\sum_{m=1}^{M}K\\left(\\frac{x - x_{m}}{h}\\right)" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "K(\\cdot)" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " is a normal kernel, " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " is the bandwidth, and " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "x_{m}" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": " is the kernel center derived from " + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\Gamma_{l}" + }, + { + "bbox": [ + 313, + 551, + 554, + 650 + ], + "type": "text", + "content": ". Finally, the PSE loss is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 660, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 660, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 353, + 660, + 555, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {P S E}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = 1} ^ {L} \\int \\hat {f} _ {l} (x) \\log [ \\hat {f} _ {l} (x) ] d x, \\tag {5}", + "image_path": "f2f3f7223ef3e118f8e5db0a64c240f96189bc3327085d60959330e5c777964e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 701, + 488, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 488, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 488, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 488, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 701, + 488, + 712 + ], + "type": "text", + "content": " is the block number of the model." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12481" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 139 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "type": "table", + "html": "
Method ClassesRealPSAQ-ViT [47]APA (Ours)
Class 10.680.440.64
Class 20.320.260.31
Class 30.410.310.36
", + "image_path": "8e49475232f1a4334a49796ef4e3b995aa80fada2b27fb929cd6cf44f8dd8986.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 296, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 214 + ], + "type": "text", + "content": "Table 1. Average cosine similarity of three randomly selected classes. For real images, the similarity is measured within the class itself, while for PSAQ-ViT and APA, the similarity is measured between synthetic and real images of the same class. The results show that APA achieves higher similarity than PSE loss, indicating aligned semantics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 236, + 141, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 141, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 141, + 249 + ], + "type": "text", + "content": "3.2. Observations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 315 + ], + "type": "text", + "content": "Existing DFQ methods have made significant progress. However, through carefully analyzing the synthetic images, we reveal that these images suffer from issues of semantic distortion and semantic inadequacy, both of which hinder further advancement in the DFQ of ViTs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": "The semantic distortion issue refers to the significant divergence between the semantics of synthetic images and real images. To demonstrate this, we visualize the features of synthetic and real images in Fig. 1a. Note that the penultimate feature typically is regarded to represent the semantics of the input [4, 58, 78, 79]. It is clear that the features of PSAQ-ViT diverge significantly from real images, suggesting that these images fail to capture the true semantic distribution of real data. Tab. 1 quantitatively measures the semantics using the average cosine similarity (ranging from " + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "inline_equation", + "content": "1" + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": "). We also report the intra-class similarity within real images as an approximate upper bound for comparison. This result further supports the observation of low similarity between synthetic images from PSAQ-ViT and real images. For instance, for Class 1, the intra-class similarity within real images is 0.68, while PSAQ-ViT achieves only 0.44." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 640 + ], + "type": "text", + "content": "The semantic inadequacy issue refers to the presence of dull regions in synthetic images, which contain redundant or non-semantic content [33, 37], hindering the model's learning process. As indicated in [13, 14], a diverse content and textures generally suggests rich information. As shown in Fig. 1b, many regions of synthetic images generated by PSAQ-ViT and PSAQ-ViT V2 exhibit a lack of diversity in content, with overly simplified textures. Specifically, the central region of PSAQ-ViT images only contains faint object structures, while PSAQ-ViT V2 images appear excessively smoothed and indistinct." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "For high-bit quantization, where model capacity is largely retained [42], the performance degradation remains relatively minor even using semantic distorted and inadequate images [47, 62]. However, in low-bit quantization, where model capacity is severely damaged and informative images are essential for recovering performance [13, 84]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "fine-tuning on these poor-quality images leads to inferior generalization to real datasets, resulting in limited performance. For example, as shown in Tab. 2, the W4A4 ViT-B fine-tuned on real images yields " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "68.16\\%" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", whereas PSAQ-ViT only achieves " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "36.32\\%" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 351, + 142, + 519, + 223 + ], + "blocks": [ + { + "bbox": [ + 351, + 142, + 519, + 223 + ], + "lines": [ + { + "bbox": [ + 351, + 142, + 519, + 223 + ], + "spans": [ + { + "bbox": [ + 351, + 142, + 519, + 223 + ], + "type": "image", + "image_path": "a76c498bc5d105f8c3d2cd2c8ad6b212c421d9237e70e6e0570cd8cd08d8f7fd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 231, + 520, + 243 + ], + "lines": [ + { + "bbox": [ + 348, + 231, + 520, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 231, + 520, + 243 + ], + "type": "text", + "content": "Figure 3. Comparison between attention maps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 554, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 280 + ], + "type": "text", + "content": "3.3. Semantics Alignment and Reinforcement Data-Free Quantization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 554, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 309 + ], + "type": "text", + "content": "In the following, we introduce the proposed SARDFQ, whose framework is illustrated in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 315, + 460, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 460, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 460, + 327 + ], + "type": "text", + "content": "3.3.1. Attention Priors Alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 330, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 498 + ], + "type": "text", + "content": "In ViTs, the self-attention mechanism encodes semantic correlations between image regions, where high-response areas in attention maps strongly correlate with semantic-discriminative content [10, 20, 31]. However, existing DFQ methods overlook this intrinsic property in the generation process. As a result, as shown in Fig. 3, synthetic images often exhibit disordered and unnatural attention patterns, with attention maps either overly diffuse or misaligned toward peripheral regions. This undermines their ability to preserve semantic-discriminative content, causing semantic distortion, as demonstrated in Fig. 1a and Tab. 1. In response, we propose Attention Priors Alignment (APA), which improves semantics alignment by optimizing synthetic images to follow randomly generated structure attention priors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": "Specifically, given a synthetic image " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": ", we first obtain its attention maps in the " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": "-th head of the " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": "-th block, denoting as " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,h} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": " represents the total number of tokens. In DeiT, the attention of the classification token toward other tokens serves as the indicator for semantic versus non-semantic parts [5]. Thus, we extract " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,h}^c \\in \\mathbb{R}^{1 \\times (N - 1)}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,h}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": ", representing the attention of the classification token to all tokens except itself. We then randomly generate attention priors " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}_{l,h}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": ", whose generation is detailed in the next part, and align " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,h}^c" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}_{l,h}" + }, + { + "bbox": [ + 313, + 498, + 555, + 620 + ], + "type": "text", + "content": " by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 629, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 629, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 370, + 629, + 555, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}) = \\operatorname {M S E} \\left(\\mathbf {A} _ {l, h} ^ {c} - \\tilde {\\mathbf {A}} _ {l, h}\\right), \\tag {6}", + "image_path": "5a79418d3efcac799991af598759c8ec27a6698c33e5fa17cd2558eae724bd23.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "where MSE represents the mean squared error. For Swin models that do not use a classification token, we substitute " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,h}^{c}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " in Eq.6 with the average attention map of all tokens [10]. As noted in [17], ViTs initially focus on all regions to capture low-level information in shallow blocks" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12482" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 70, + 133, + 110 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 133, + 110 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 133, + 110 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 133, + 110 + ], + "type": "image", + "image_path": "47a751ddad2314c95508a2a12f85e247727bb9004ac501eed58fe564b277eaf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 119, + 264, + 131 + ], + "lines": [ + { + "bbox": [ + 85, + 119, + 264, + 131 + ], + "spans": [ + { + "bbox": [ + 85, + 119, + 264, + 131 + ], + "type": "text", + "content": "Figure 4. Examples of generated attention priors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 70, + 175, + 110 + ], + "blocks": [ + { + "bbox": [ + 136, + 70, + 175, + 110 + ], + "lines": [ + { + "bbox": [ + 136, + 70, + 175, + 110 + ], + "spans": [ + { + "bbox": [ + 136, + 70, + 175, + 110 + ], + "type": "image", + "image_path": "3c735d92212b727d406c580060dcabab8f5fe18157239ef6f24724c4a4698a16.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 217, + 110 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 217, + 110 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 217, + 110 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 217, + 110 + ], + "type": "image", + "image_path": "e9511c52c1df3beaa3c5e411286edf72afd6a1e11835c3eaf0e2f6b22a8878de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 259, + 110 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 259, + 110 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 259, + 110 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 259, + 110 + ], + "type": "image", + "image_path": "0d1652aa491f07440ea2ee43d6f9c93ab483199aba85c303bd36c784fca49c7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 142, + 296, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 226 + ], + "type": "text", + "content": "and gradually shift their focus toward semantic regions in deeper blocks to extract high-level semantic information. Leveraging this property, we selectively apply " + }, + { + "bbox": [ + 55, + 142, + 296, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l,h}^{\\mathrm{APA}}" + }, + { + "bbox": [ + 55, + 142, + 296, + 226 + ], + "type": "text", + "content": " to deeper blocks, progressively aligning attention towards semantically relevant areas. The total APA loss is computed as a depth-weighted sum of the individual Eq. 6 across these deeper blocks:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 233, + 294, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 233, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 112, + 233, + 294, + 266 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) = \\sum_ {l = S} ^ {L} \\sum_ {h = 1} ^ {H} \\frac {l}{L} \\mathcal {L} _ {l, h} (\\tilde {\\boldsymbol {I}}), \\tag {7}", + "image_path": "4f1c0a8a1a252d98cb0dbfd64bd2def1a6f99b21e74a412c94abd9fd4e03c62c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "text", + "content": " is a pre-given hyper-parameter denoting the start of deep blocks and are experimentally set " + }, + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "inline_equation", + "content": "S = \\frac{L}{2}" + }, + { + "bbox": [ + 55, + 273, + 296, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "text", + "content": "Attention Priors Generation. To generate attention priors " + }, + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}_{l,h}" + }, + { + "bbox": [ + 55, + 297, + 296, + 403 + ], + "type": "text", + "content": ", Gaussian Mixture Models (GMMs) are employed as it is the most commonly used distribution with high flexibility. ViTs utilize different attention heads to capture diverse patterns and learn varied, informative representations [8]. Thus, we use distinct GMMs for each head. Note that the goal here is not to replicate real attention maps precisely, but to generate simulated structure attention priors to guide the learning of synthetic images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": "In particular, we first initialize an all zero matrix " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}}\\in" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "H = W = \\sqrt{N - 1}" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " for DeiT, " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "H =" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "W = \\sqrt{N}" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " for Swin. For example, for DeiT-S, the " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "H =" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "W = \\sqrt{196} = 14" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " . Then, we generate " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " two-dimensional Gaussian distributions, where " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " is randomly sampled from " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "1\\sim K_{APA}" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "K_{APA}" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " is set to 5 in all experiments. Each Gaussian has its mean and covariance3. Consequently, the matrix element at the " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " -th row and " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " -th column, " + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}} [i,j]" + }, + { + "bbox": [ + 55, + 404, + 296, + 513 + ], + "type": "text", + "content": " , is determined by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 518, + 295, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 518, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 117, + 518, + 295, + 538 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {P}} [ i, j ] = \\max _ {m = 1, \\dots , k} \\mathbf {G} ^ {m} [ i, j ]. \\tag {8}", + "image_path": "110da8a652cd4b4333a6272e46fde393c55a335f095c881a4574b1c646dee98a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 548, + 164, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 164, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 164, + 559 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 55, + 548, + 164, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 548, + 164, + 559 + ], + "type": "text", + "content": " is normalized by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 567, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 567, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 129, + 567, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {P}} _ {n} = \\frac {\\tilde {\\mathbf {P}}}{\\sum \\tilde {\\mathbf {P}}} \\cdot (1 - x). \\tag {9}", + "image_path": "84a2540c16f1ec978a87ed9d9d8d01ef468ac4459895011c38e7e27ce3c2e63c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": "Here, for DeiT that incorporates the classification token, " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": " is randomly sampled from a uniform distribution " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "U(0,1)" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": ", representing the proportion of the attention score that the classification token allocates to itself. For Swin, which does not use a classification token, " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": " is set to 0. Finally, " + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{P}}_n" + }, + { + "bbox": [ + 55, + 602, + 296, + 674 + ], + "type": "text", + "content": " is flatten to match the dimensionality:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 133, + 681, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 681, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 133, + 681, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {A}} _ {l, h} = \\operatorname {f l a t t e n} \\left(\\tilde {\\mathbf {P}} _ {n}\\right). \\tag {10}", + "image_path": "27923c194c9481f238cd1c7735bcd86512e0d32ad1a325badfa5b28af54eab53.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 544, + 84 + ], + "type": "text", + "content": "Fig. 4 displays examples of generated attention priors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 85, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 276 + ], + "type": "text", + "content": "Discussion. APA prevents attention disorder, ensuring that synthetic images exhibit more coherent and natural attention patterns, as demonstrated in Fig. 3. As a result, APA selectively enhances the responses in certain regions, effectively emphasizing semantic-discriminative regions within synthetic images and thus prompting the discriminative features. Although simple, APA enables synthetic images to align better with the real semantics, as validated by both visual and quantitative evaluations. As shown in Fig. 1a, compared to PSAQ-ViT, features obtained after applying the APA loss are more closely aligned with real images, indicating better semantic alignment. The quantitative results in Tab. 1 further support that APA achieves superior semantic alignment. For example, in Class 1, the intra-class similarity for PSAQ-ViT is 0.44, whereas APA achieves a higher similarity of 0.64." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 282, + 474, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 474, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 474, + 293 + ], + "type": "text", + "content": "3.3.2. Multi-Semantic Reinforcement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 297, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 476 + ], + "type": "text", + "content": "Current DFQ methods for ViTs [47, 48, 62] optimize images through global optimization, treating the entire image as a single semantic unit. However, this is affected by low-rank structural regularity [34], where adjacent pixels exhibit strong similarity, leading to dull regions with redundant or non-semantic content for model learning [37]. This issue is further exacerbated by the tokenization mechanism in ViTs, as processing images in fixed-size patches make dull regions increase at the patch level [19, 70]. Consequently, as shown in Fig. 1b, synthetic images generated by existing methods exhibit large dull regions, resulting in semantic inadequacy [37]. In response, we propose Multi-Semantic Reinforcement (MSR), which applies localized patch optimization to enhance semantic richness by learning local patches with distinct semantics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": "Specifically, for a synthetic image " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{I}" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": ", instead of feeding only the entire image, we also feed its patches and optimize them individually. Initially, we select " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": " nonoverlapping patches, where " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": " is chosen randomly from the set " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots,K_{MSR}\\}" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": " set to 4 in all experiments. These " + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 476, + 554, + 561 + ], + "type": "text", + "content": " patches are then cropped and resized to match the model's input dimensions:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 569, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 569, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 354, + 569, + 553, + 585 + ], + "type": "interline_equation", + "content": "\\{\\tilde {\\boldsymbol {I}} _ {M S R} ^ {i} \\} _ {i = 1, \\dots , m} = \\operatorname {r e s i z e} \\left(\\operatorname {c r o p} _ {m} (\\tilde {\\boldsymbol {I}})\\right), \\tag {11}", + "image_path": "76aa5037d48aabcd9a89419f8db99128ddf2d5049ee58e99ff1d9928b13814bc.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\text{crop}_m(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " crops " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " non-overlapping patches from input, and " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\text{resize}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " is the resize function. Each patch, denoted as " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{I}_{MSR}^i" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": ", is treated as a new image with an assigned semantic target " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": ". Note that the gradient is backpropagated only to update the corresponding patch in the original image, leaving the rest of the image unaffected." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Softlabel Learning. The one-hot loss (Eq. 3) only learns the semantics of the target class, making it unsuitable for " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{I}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " under MSR, as its patches " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " contain distinct semantics. In response, we propose Softlabel Learning" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 203, + 713 + ], + "type": "text", + "content": "3The pseudo code is detailed in appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12483" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "(SL), which applies multiple semantic targets to accommodate the learning of images augmented by MSR. Specifically, we first sample " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{C} \\sim U(0,1)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", then modify its values by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 128, + 262, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 128, + 262, + 159 + ], + "spans": [ + { + "bbox": [ + 89, + 128, + 262, + 159 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} Z [ c ^ {i} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}} _ {M S R} ^ {i}, \\\\ Z [ c ^ {1}, \\ldots , c ^ {m} ] \\sim U (\\epsilon_ {1}, \\epsilon_ {2}), & \\text {f o r} \\tilde {\\boldsymbol {I}}, \\end{array} \\right.", + "image_path": "a82da8b1886a95716371e801f7160284601a320f5d3e333f7117da2f6241ef95.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "U(\\epsilon_1, \\epsilon_2)" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " denotes the uniform distribution over the interval " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "[\\epsilon_1, \\epsilon_2]" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " is the number of patches determined in MSR, and " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_1" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\epsilon_2" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": " control the softness, both empirically set consistently to 5 and 10 in all experiments. The soft target is defined as " + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "inline_equation", + "content": "T_s = \\mathrm{softmax}(Z)" + }, + { + "bbox": [ + 55, + 167, + 296, + 228 + ], + "type": "text", + "content": ", and SL loss is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 234, + 294, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 234, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 79, + 234, + 294, + 250 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\mathrm {S L}} \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}\\right) = S C E \\left(F \\left(\\tilde {I} / \\tilde {I} _ {M S R} ^ {i}, T _ {s}\\right), \\right. \\tag {12}", + "image_path": "e8200eada5a3b35a095b7b6580d28edac669cbf477207b55a355a999f3eca55b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "inline_equation", + "content": "SCE(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "text", + "content": " is soft cross entropy and " + }, + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 55, + 255, + 296, + 316 + ], + "type": "text", + "content": " returns the predicted probability for its input. SL facilitates smooth learning between across semantic targets, ensuring that MSR-enhanced images receive consistent supervision rather than conflicting supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": "Discussion. By leveraging localized patch optimization, MSR ensures that each patch contributes unique semantics, forcing synthetic images to capture diverse features rather than being dominated by large homogeneous dull regions. As demonstrated by the richer and more diverse content and textures in Fig. 1b, MSR effectively reduces dull regions and enhances semantic richness in synthetic images. Moreover, MSR transforms synthetic images into composites of multiple semantic objects rather than a single unit, thereby providing more distinct semantic samples, i.e., " + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\{\\tilde{I}_{MSR}^i\\}_{i = 1,\\dots ,m}" + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": ", for model training. Unlike traditional cropping used in data augmentation, which aims to improve classification robustness by training with cropped patches labeled with the original class, MSR aims for semantic richness within synthetic images, ultimately enabling accurate data-free quantization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 154, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 154, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 154, + 528 + ], + "type": "text", + "content": "3.4. Overall Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 532, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 592 + ], + "type": "text", + "content": "The overall pipeline consists of two stages: data synthesis and quantized network learning. The first stage uses the proposed SARDFQ to produce synthetic images. The second stage fine-tunes the quantized model using the generated synthetic images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 148, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 148, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 148, + 609 + ], + "type": "text", + "content": "3.4.1. Data Synthesis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 648 + ], + "type": "text", + "content": "In the data synthesis stage, we combine the proposed APA loss of Eq. 7, SL loss of Eq. 12, and TV loss of Eq. 4 to formulate the objective function as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 655, + 296, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 655, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 66, + 655, + 296, + 671 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G} (\\tilde {\\boldsymbol {I}}) = \\alpha_ {1} \\mathcal {L} ^ {\\mathrm {A P A}} (\\tilde {\\boldsymbol {I}}) + \\mathcal {L} ^ {\\mathrm {S L}} (\\tilde {\\boldsymbol {I}}) + 0. 0 5 \\mathcal {L} ^ {\\mathrm {T V}} (\\tilde {\\boldsymbol {I}}). \\tag {13}", + "image_path": "2bf814de1fea04afd925e58b90b38bb7757d5f997080cab53d080a8af6106484.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is hyperparameters and is determined by gridsearch. Note that the weight of TV loss is fixed to 0.05, following [47], to avoid a cumbersome hyperparameter search." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 469, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 469, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 469, + 85 + ], + "type": "text", + "content": "3.4.2. Quantized Network Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "text", + "content": "Recently DFQ methods have introduced the PTQ methods in learning quantized models due to their advantages of speed, memory efficiency, and performance [35, 62]. Thus, following the success of [42, 71], we fine-tune the quantized network block-wisely. Specifically, denote " + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "text", + "content": " as the outputs of the " + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "text", + "content": "-th block of the full-precision model, and " + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{X}}_l" + }, + { + "bbox": [ + 313, + 87, + 555, + 183 + ], + "type": "text", + "content": " represent outputs of the quantized counterpart. The reconstruction loss is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 394, + 190, + 555, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l} = \\left\\| \\mathbf {X} _ {l} - \\bar {\\mathbf {X}} _ {l} \\right\\| _ {2}. \\tag {14}", + "image_path": "1aafba27b2b933c1b256cddbeb2158eb066d71acb7f47e700985c1aee83e36a9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_l" + }, + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "text", + "content": " is only backward to update weights within " + }, + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 211, + 555, + 248 + ], + "type": "text", + "content": "-th block. Note that for a fair comparison, all compared methods adopt the same quantized network learning stage." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 257, + 391, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 257, + 391, + 271 + ], + "spans": [ + { + "bbox": [ + 314, + 257, + 391, + 271 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 277, + 447, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 447, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 447, + 290 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 294, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 366 + ], + "type": "text", + "content": "Models and Tasks. We evaluate the performance of SARDFQ by test quantized ViT-S/B [20], DeiT-T/S/B [68], and Swin-S/B [54] on the classification task using ImageNet [63]. The pre-trained models are downloaded from the timm library. In appendix, we further provide results on detection and segmentation tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 366, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 472 + ], + "type": "text", + "content": "Comparison methods. We compare our SARDFQ against Gaussian noise, real images, and previous methods including SMI [33] and PSAQ-ViT [47] and its subsequent version, PSAQ-ViT V2 [48]. For a fair comparison, we generate synthetic images using their methods and apply our quantized network learning strategy. We use the official code for SMI and PSAQ-ViT to reproduce their images, while PSAQ-ViT V2 is re-implemented by us, as no official code is available." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": "Experimental settings. All experiments were conducted using the PyTorch framework [60] on a single NVIDIA 3090 GPU. In the data synthesis stage, synthetic images were initialized with standard Gaussian noise, generating 32 images in total. The Adam optimizer [38] with " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.5" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.9" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": " was used, with learning rates of 0.25 for Swin and 0.2 for others, and a total of 1,000 iterations. For all models, " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "K_{APA}" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon_{1}" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon_{2}" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": " were set to 5, 4, 5, and 10, respectively, based on a search with W4A4 DeiT-S. The value of " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": " was determined by grid search for each model: 1e5 for DeiT-T/S, 1e4 for DeiT-B, 100 for ViT-B and Swin-B, 10 for Swin-S, and 1 for ViT-S. Although further hyperparameter search may improve performance, the current settings already yield superior results. In the quantized network learning stage, following [85], the Adam optimizer with " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 474, + 555, + 677 + ], + "type": "text", + "content": " was used, with weight decay set to 0 and an initial learning rate of 4e-5, adjusted" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "4In appendix, we present practical efficiency results and more performance comparisons with more methods such as CLAMP-ViT [62] on W8/A8 and W4/A8 settings." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12484" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 71, + 547, + 376 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 547, + 376 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 547, + 376 + ], + "type": "table", + "html": "
ModelW/ARealGaussian noisePSAQ-ViT [47]PSAQ-ViT V2 [48]SMI [33]SARDFQ (Ours)
ViT-S (81.39)4/466.576.0247.2441.5324.3329.4150.32
5/576.6936.7771.5968.4161.3365.1974.31
6/679.4661.2077.2074.7672.9572.4678.40
ViT-B (84.54)4/468.160.1536.3226.3235.2719.6751.84
5/579.214.1668.4867.9567.5357.1370.70
6/681.8955.1876.6571.8776.3369.8279.16
DeiT-T (72.21)4/456.6017.4347.7530.2030.1413.1852.06
5/567.0943.4964.1055.1656.4439.3566.41
6/669.8156.2368.3762.7764.0344.3969.73
DeiT-S (79.85)4/468.4620.8958.2845.5342.7711.7162.29
5/575.0641.0671.9063.1462.8829.1374.06
6/677.8765.6375.8568.8571.6537.6977.31
DeiT-B (81.85)4/477.0747.2071.7566.4365.3359.0472.17
5/579.8665.4678.4576.7776.7475.3378.72
6/680.9062.7980.0079.2278.8177.6680.15
Swin-S (83.20)4/478.1231.9273.1965.5565.8574.74
5/580.5152.1078.1574.3775.4179.56
6/680.6065.6679.7478.5078.2580.56
Swin-B (85.27)4/478.8030.1471.8467.4265.2376.42
5/582.5135.2878.5077.2075.2580.82
6/682.6467.3782.0081.4180.3083.03
", + "image_path": "33299b202b35ff497d1ff72b39a86307dea4795633c4b17dd51632f5f95dd691.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "type": "text", + "content": "Table 2. Quantization results on ImageNet dataset, with top-1 accuracy " + }, + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 384, + 555, + 430 + ], + "type": "text", + "content": " reported. The performance of the full-precision model is listed below the model name. \"W/A\" denotes the bit-width of weights/activations. \"Real\" refers to using real images. For SMI [33], we provide the performance of using dense (normal-sized numbers) and sparse (smaller-sized numbers) synthetic images, respectively. Note that for Swin models, we do not provide the results for sparse synthetic images as the sparse generation method of SMI is infeasible." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 441, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 489 + ], + "type": "text", + "content": "via cosine decay for 100 iterations. A channel-wise quantizer was used for weights, and a layer-wise quantizer for activations, with all matrix multiplications in ViTs quantized [47-49]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 502, + 177, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 177, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 177, + 514 + ], + "type": "text", + "content": "4.2. Quantization Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "The quantization results are presented in Tab.2. Our SARDFQ demonstrates consistent improvements across various quantization bit-width configurations, particularly with low bit-width settings. Specifically, for ViT-S, SARDFQ improves the performance by " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "3.08\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W4/A4 setting, " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.72\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W5/A5 setting, and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.20\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W6/A6 setting. For ViT-B, SARDFQ achieves performance gains of " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "15.52\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W4/A4 setting, " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.22\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W5/A5 setting, and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.51\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W6/A6 setting. Results on DeiT also demonstrate the effectiveness of SARDFQ. For example, on DeiT-T, SARDFQ shows a marked improvement by increasing top-1 accuracy by " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "4.31\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W4/A4 setting, " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.31\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W5/A5 setting, and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.36\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W6/A6 setting. For DeiT-S, SARDFQ enhances top-1 accuracy by " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "4.01\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W4/A4 setting, " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.16\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W5/A5 setting, and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.46\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " in the W6/A6 setting. The quan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": "tization results of Swin-S/B also affirm the superiority of our SARDFQ in enhancing model accuracy under different quantization configurations. In particular, for Swin-S, the proposed SARDFQ increases the accuracy by " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "1.55\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W4/A4 setting, " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "1.41\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W5/A5 setting, and " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "0.82\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W6/A6 setting, respectively. When it comes to Swin-B, the proposed SARDFQ increases the accuracy by " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "4.58\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W4/A4 setting, " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "2.32\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W5/A5 setting, and " + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "inline_equation", + "content": "1.03\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 549 + ], + "type": "text", + "content": " for the W6/A6 setting, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 561, + 408, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 408, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 408, + 574 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 580, + 542, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 542, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 542, + 592 + ], + "type": "text", + "content": "All ablation studies are conducted on the W4A4 DeiT-S." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Analysis of APA, MSR, and SL. We analyze the effectiveness of the proposed APA (Sec.,3.3.1), MSR (Sec.,3.3.2), and SL (Eq.,12) in Tab.,3. Adding APA and SL individually to the baseline increases accuracy. Notably, APA boosts performance from " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "51.73\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "60.26\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", confirming its effectiveness in aligning semantics (Sec.,3.3.1). Applying MSR alone slightly decreases accuracy from " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "51.73\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "50.75\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", indicating that one-hot loss is unsuitable for MSR-augmented synthetic images. However, when both MSR and SL are applied, accuracy rises to " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "56.08\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", sug-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12485" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 70, + 249, + 198 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 249, + 198 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 249, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 249, + 198 + ], + "type": "table", + "html": "
APAMSRSLAcc. (%)
Baseline51.73
✓60.26
✓50.75
✓52.02
✓✓61.58
✓✓60.51
✓✓56.08
✓✓✓62.29
", + "image_path": "206a27669ba58a5391cf71ee5a1676d1fde71dbe5309b8c26d17b7bd17575392.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "text", + "content": "gesting SL is more compatible with MSR than one-hot loss. Combining APA with either MSR or SL further improves performance. For example, APA and SL together yield an accuracy of " + }, + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "inline_equation", + "content": "60.51\\%" + }, + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "text", + "content": ", and when all three strategies are used, the best performance of " + }, + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "inline_equation", + "content": "62.29\\%" + }, + { + "bbox": [ + 55, + 255, + 295, + 314 + ], + "type": "text", + "content": " is achieved." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "text", + "content": "Analysis of priors distribution. Tab. 4a showcases the results of using other distributions to formulate the attention priors. The unevenly distributed GMM and Laplace present comparable performance of " + }, + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "inline_equation", + "content": "62.29\\%" + }, + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "inline_equation", + "content": "62.16\\%" + }, + { + "bbox": [ + 55, + 315, + 296, + 399 + ], + "type": "text", + "content": ", respectively. Moreover, GMM provides a similar performance to the real's, indicating it performs well in imitating the patterns of real images." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 408, + 173, + 500 + ], + "blocks": [ + { + "bbox": [ + 58, + 408, + 173, + 500 + ], + "lines": [ + { + "bbox": [ + 58, + 408, + 173, + 500 + ], + "spans": [ + { + "bbox": [ + 58, + 408, + 173, + 500 + ], + "type": "image", + "image_path": "a9648db1c30d9fa309266ebee1ec85a0f5984e694d58af35381fa07d3dce66ed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 504, + 121, + 512 + ], + "lines": [ + { + "bbox": [ + 111, + 504, + 121, + 512 + ], + "spans": [ + { + "bbox": [ + 111, + 504, + 121, + 512 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "lines": [ + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "spans": [ + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "type": "text", + "content": "Figure 5. Effect of varying (a) " + }, + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "type": "text", + "content": " and (b) " + }, + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 84, + 522, + 266, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 408, + 293, + 500 + ], + "blocks": [ + { + "bbox": [ + 173, + 408, + 293, + 500 + ], + "lines": [ + { + "bbox": [ + 173, + 408, + 293, + 500 + ], + "spans": [ + { + "bbox": [ + 173, + 408, + 293, + 500 + ], + "type": "image", + "image_path": "3e1333dc94a67c2dbdb0700a59515b6c89c5e3a10c0ae3c609fbbe5e2a6d3f60.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 503, + 239, + 512 + ], + "lines": [ + { + "bbox": [ + 228, + 503, + 239, + 512 + ], + "spans": [ + { + "bbox": [ + 228, + 503, + 239, + 512 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Analysis of " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1}, K_{APA}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " from Eq. 13 balance the importance of the proposed APA loss during the update of the synthetic images. Fig. 5a demonstrates that the optimal performance is achieved when " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 1e5" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". Incrementally increasing " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " improves performance up to " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "62.29\\%" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 1e5" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". However, further increases in " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " subsequently degrade performance. The " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{APA}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " in APA is the upper limit on the number of the Gaussian distributions used for priors generation. Tab. 4b displays the ablation study for different values of " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{APA}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". The best accuracy is achieved when " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{APA} = 5" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " in MSR is the upper limit on the number of patches. Fig. 5b demonstrates that when " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{MSR} = 4" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", the optimal performance is achieved. Using " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " larger than 4 will hurt the accuracy. We consider this" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 82, + 441, + 144 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 239 + ], + "type": "text", + "content": "Table 3. Influence of the proposed APA, MSR, and SL on accuracy. The baseline adopts the one-hot loss. Applying APA, MSR, and SL yields SARDFQ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 82, + 441, + 144 + ], + "lines": [ + { + "bbox": [ + 315, + 82, + 441, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 441, + 144 + ], + "type": "table", + "html": "
Priors DistributionTop-1
GMM62.29
Laplace62.16
Real63.19
", + "image_path": "3e3d62c782164115c35624252c98cb9d7fdd5faa0e95401c20fe58b13586962c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 460, + 70, + 539, + 156 + ], + "blocks": [ + { + "bbox": [ + 364, + 146, + 375, + 155 + ], + "lines": [ + { + "bbox": [ + 364, + 146, + 375, + 155 + ], + "spans": [ + { + "bbox": [ + 364, + 146, + 375, + 155 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 460, + 70, + 539, + 156 + ], + "lines": [ + { + "bbox": [ + 460, + 70, + 539, + 156 + ], + "spans": [ + { + "bbox": [ + 460, + 70, + 539, + 156 + ], + "type": "table", + "html": "
KAPATop-1
161.13
361.52
562.29
761.53
961.05
", + "image_path": "29b607cecea06dacbd324c39babe14929a287cf17e47dc3308ce9afe27f79bbd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 345, + 172, + 414, + 222 + ], + "blocks": [ + { + "bbox": [ + 493, + 157, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 493, + 157, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 493, + 157, + 505, + 167 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 172, + 414, + 222 + ], + "lines": [ + { + "bbox": [ + 345, + 172, + 414, + 222 + ], + "spans": [ + { + "bbox": [ + 345, + 172, + 414, + 222 + ], + "type": "table", + "html": "
STop-1
061.96
L/262.29
", + "image_path": "428c915ab844ae292e8cac279f6eac909d561e07aa4b86888bd7f7747fb61b2f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 453, + 172, + 524, + 222 + ], + "blocks": [ + { + "bbox": [ + 375, + 224, + 386, + 234 + ], + "lines": [ + { + "bbox": [ + 375, + 224, + 386, + 234 + ], + "spans": [ + { + "bbox": [ + 375, + 224, + 386, + 234 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 453, + 172, + 524, + 222 + ], + "lines": [ + { + "bbox": [ + 453, + 172, + 524, + 222 + ], + "spans": [ + { + "bbox": [ + 453, + 172, + 524, + 222 + ], + "type": "table", + "html": "
w. l/LTop-1
✓62.29
×61.32
", + "image_path": "8bb3b141338774aeae4a45994c2d336555dc3fc8f80865f42494534157eafc84.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 484, + 224, + 495, + 234 + ], + "lines": [ + { + "bbox": [ + 484, + 224, + 495, + 234 + ], + "spans": [ + { + "bbox": [ + 484, + 224, + 495, + 234 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "text", + "content": "Table 4. Effect of varying (a) priors types; (b) " + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "inline_equation", + "content": "K_{APA}" + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "text", + "content": "; (c) " + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "text", + "content": "; (d) " + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "inline_equation", + "content": "\\frac{l}{L}" + }, + { + "bbox": [ + 313, + 243, + 553, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 279, + 550, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 550, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 550, + 291 + ], + "type": "text", + "content": "due to limited patch resolution if using a too large " + }, + { + "bbox": [ + 313, + 279, + 550, + 291 + ], + "type": "inline_equation", + "content": "K_{MSR}" + }, + { + "bbox": [ + 313, + 279, + 550, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": "Analysis of APA loss. Here, we conduct the ablation study by considering the " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "\\frac{l}{L}" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": " in Eq.7. Tab.4c presents the effect of varying " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": " in Eq.7. If applying APA loss to all blocks " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "(S = 0)" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": ", the top-1 accuracy decreases to " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "61.96\\%" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": ". From Tab.4d, it can be seen that absorbing the scale " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "\\frac{l}{L}" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": " in Eq.7 presents " + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "inline_equation", + "content": "0.97\\%" + }, + { + "bbox": [ + 313, + 291, + 554, + 364 + ], + "type": "text", + "content": " performance gains." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 374, + 389, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 389, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 389, + 387 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 394, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 515 + ], + "type": "text", + "content": "We further discuss some limitations of the proposed SARDFQ, which will guide future research directions. First, although SARDFQ shows substantial performance improvement, a performance gap between SARDFQ and real data remains challenging, highlighting the need for a stronger semantics alignment and reinforcement method. Second, SARDFQ currently lacks a theoretical foundation. Future work could establish a theoretical framework for SARDFQ, particularly in understanding how APA and MSR influence synthetic images in a formalized manner." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we investigate the DFQ method for ViTs. We first identify that synthetic images generated by existing methods suffer from semantic distortion and inadequacy issues, and propose SARDFQ to address these issues. To mitigate semantic distortion, SARDFQ introduces APA, which guides synthetic images to align with randomly generated structural attention patterns. To tackle semantic inadequacy, SARDFQ incorporates MSR. MSR optimizes different regions of synthetic images with unique semantics, thereby enhancing overall semantic richness. Moreover, SARDFQ employs SL, which adopts multiple semantic targets to ensure seamless learning of images augmented by MSR. Extensive experiments on various ViT models and tasks validate the effectiveness of SARDFQ." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12486" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project (No. 2022ZD0118202), the National Science Fund for Distinguished Young Scholars (No.62025603), the National Natural Science Foundation of China (Np. 624B2119, No. U21B2037, No. U22B2051, No. 62176222, No. 62176223, No. 62176226, No. 62072386, No. 62072387, No. 62072389, No. 62002305 and No. 62272401), and the Natural Science Foundation of Fujian Province of China (No.2021J01002, No.2022J06001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 243, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 243, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 294, + 287 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 6836-6846, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 295, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 343 + ], + "type": "text", + "content": "[2] Jianhong Bai, Yuchen Yang, Huanpeng Chu, Hualiang Wang, Zuozhu Liu, Ruizhe Chen, Xiaoxuan He, Lianrui Mu, Chengfei Cai, and Haoji Hu. Robustness-guided image synthesis for data-free quantization. arXiv preprint arXiv:2310.03661, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 344, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 295, + 388 + ], + "type": "text", + "content": "[3] Ron Banner, Yury Nahshan, Daniel Soudry, et al. Post training 4-bit quantization of convolutional networks for rapid-deployment. In Advances in Neural Information Processing Systems (NeurIPS), pages 7950-7958, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 433 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(8):1798-1828, 2013. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 478 + ], + "type": "text", + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "type": "text", + "content": "[6] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13169-13178, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 589 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "type": "text", + "content": "[8] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 397-406, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 690 + ], + "type": "text", + "content": "[9] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Mengzhao Chen, Mingbao Lin, Ke Li, Yunhang Shen, Yongjian Wu, Fei Chao, and Rongrong Ji. Cf-vit: A general" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "coarse-to-fine method for vision transformer. In AAAI Conference on Artificial Intelligence, pages 7042-7052, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[11] Mengzhao Chen, Wenqi Shao, Peng Xu, Mingbao Lin, Kaipeng Zhang, Fei Chao, Rongrong Ji, Yu Qiao, and Ping Luo. Diffrate: Differentiable compression rate for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17164-17174, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "text", + "content": "[12] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8126-8135, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[13] Xinrui Chen, Yizhi Wang, Renao Yan, Yiqing Liu, Tian Guan, and Yonghong He. Texq: Zero-shot network quantization with texture feature distribution calibration. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[14] Kanghyun Choi, Deokki Hong, Noseong Park, Youngsok Kim, and Jinho Lee. Qimera: Data-free quantization with synthetic boundary supporting samples. In Advances in Neural Information Processing Systems (NeurIPS), pages 14835-14847, 2021. 2, 4, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 371 + ], + "type": "text", + "content": "[15] Kanghyun Choi, Hye Yoon Lee, Deokki Hong, Joonsang Yu, Noseong Park, Youngsok Kim, and Jinho Lee. It's all in the teacher: Zero-shot quantization brought closer to the teacher. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8311-8321, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[16] Kanghyun Choi, Hye Yoon Lee, Dain Kwon, SunJong Park, Kyuyeun Kim, Noseong Park, and Jinho Lee. Mimiq: Low-bit data-free quantization of vision transformers. arXiv preprint arXiv:2407.20021, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[17] Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the relationship between self-attention and convolutional layers. In International Conference on Learning Representations (ICLR), 2020. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[18] Yifu Ding, Haotong Qin, Qinghua Yan, Zhenhua Chai, Junjie Liu, Xiaolin Wei, and Xianglong Liu. Towards accurate posttraining quantization for vision transformer. In 30th ACM International Conference on Multimedia (ACMMM), pages 5380-5388, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 568 + ], + "type": "text", + "content": "[19] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International conference on machine learning (ICML), pages 2793-2803. PMLR, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 570, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 645 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "text", + "content": "[21] Hoang Anh Dung, Cuong Pham, Trung Le, Jianfei Cai, and Thanh-Toan Do. Sharpness-aware data generation for zero-shot quantization. In International Conference on Machine Learning (ICML), 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[22] Steven K. Esser, Jeffrey L. McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S. Modha." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12487" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Learned step size quantization. In International Conference on Learning Representations (ICLR), 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[23] Gongfan Fang, Jie Song, Xinchao Wang, Chengchao Shen, Xingen Wang, and Mingli Song. Contrastive model inversion for data-free knowledge distillation. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[24] Natalia Frumkin, Dibakar Gope, and Diana Marculescu. Jumping through local minima: Quantization in the loss landscape of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16978-16988, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[25] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tianxiang Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differentiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4852–4861, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[26] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), pages 2672–2680, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[27] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. Squant: On-the-fly data-free quantization via diagonal hessian approximation. In The Eleventh International Conference on Learning Representations (ICLR), 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[28] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[29] Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(1):87-110, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[30] Zhiwei Hao, Jianyuan Guo, Ding Jia, Kai Han, Yehui Tang, Chao Zhang, Han Hu, and Yunhe Wang. Learning efficient vision transformers via fine-grained manifold distillation. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 601 + ], + "type": "text", + "content": "[31] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 773-783, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[32] Zejiang Hou and Sun-Yuan Kung. Multi-dimensional vision transformer compression via dependency guided gaussian process search. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3669-3678, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[33] Zixuan Hu, Yongxian Wei, Li Shen, Zhenyi Wang, Lei Li, Chun Yuan, and Dacheng Tao. Sparse model inversion: Efficient inversion of vision transformers for data-free applications. In International Conference on Machine Learning (ICML), 2024. 2, 3, 4, 6, 7, 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[34] Jinggang Huang and David Mumford. Statistics of natural images and models. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 541-547. IEEE, 1999. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[35] Yongkweon Jeon, Chungman Lee, and Ho-young Kim. Genie: show me the data for quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12064-12073, 2023. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[36] Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey. ACM Computing Surveys (CSUR), 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 270 + ], + "type": "text", + "content": "[37] Jang-Hyun Kim, Jinuk Kim, Seong Joon Oh, Sangdoo Yun, Hwanjun Song, Joonhyun Jeong, Jung-Woo Ha, and Hyun Oh Song. Dataset condensation via efficient synthetic-data parameterization. In International Conference on Machine Learning (ICML), pages 11102-11118. PMLR, 2022. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[38] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 337 + ], + "type": "text", + "content": "[39] Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[40] Huantong Li, Xiangmiao Wu, Fanbing Lv, Daihai Liao, Thomas H Li, Yonggang Zhang, Bo Han, and Mingkui Tan. Hard sample matters a lot in zero-shot quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24417-24426, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[41] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for neural networks. In International Conference on Learning Representations (ICLR), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[42] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations (ICLR), 2021. 2, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "text", + "content": "[43] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 34451-34463, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[44] Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Stableq: Enhancing data-scarce quantization with text-to-image data. arXiv preprint arXiv:2312.05272, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 645 + ], + "type": "text", + "content": "[45] Yuhang Li, Youngeun Kim, Donghyun Lee, Souvik Kundu, and Priyadarshini Panda. Genq: Quantization in low data regimes with generative synthetic data. In European Conference on Computer Vision (ECCV), pages 216-235. Springer, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[46] Zhikai Li and Qingyi Gu. I-vit: Integer-only quantization for efficient vision transformer inference. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17065-17075, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[47] Zhikai Li, Liping Ma, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Patch similarity aware data-free quantization for" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12488" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "vision transformers. In European Conference on Computer Vision (ECCV), pages 154-170. Springer, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "text", + "content": "[48] Zhikai Li, Mengjuan Chen, Junrui Xiao, and Qingyi Gu. Psaq-vit v2: Toward accurate and general data-free quantization for vision transformers. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 2023. 2, 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 219 + ], + "type": "text", + "content": "[49] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 17227-17236, 2023. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "type": "text", + "content": "[50] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In IEEE/CVF international conference on computer vision (ICCV), pages 1833-1844, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 296, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 296, + 322 + ], + "type": "text", + "content": "[51] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In Thirty-First International Joint Conference on Artificial Intelligence, (IJCAI), pages 1173–1179, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 296, + 369 + ], + "type": "text", + "content": "[52] Shih-Yang Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning (ICML), pages 21813-21824, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "text", + "content": "[53] Yijiang Liu, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Noisyquant: Noisy bias-enhanced post-training activation quantization for vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20321-20330, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 296, + 483 + ], + "type": "text", + "content": "[54] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF international conference on computer vision (ICCV), pages 10012-10022, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 485, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 485, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 296, + 529 + ], + "type": "text", + "content": "[55] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In Advances in Neural Information Processing Systems (NeurIPS), pages 28092-28103, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 296, + 564 + ], + "type": "text", + "content": "[56] Yiwei Ma, Jiayi Ji, Xiaoshuai Sun, Yiyi Zhou, and Rongrong Ji. Towards local visual modeling for image captioning. Pattern Recognition, 138:109420, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 610 + ], + "type": "text", + "content": "[57] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 667 + ], + "type": "text", + "content": "[58] Muhammad Muzammal Naseer, Kanchana Ranasinghe, Salman H Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Intriguing properties of vision transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 23296-23308, 2021. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[59] Daniel Neimark, Omri Bar, Maya Zohar, and Dotan Asselmann. Video transformer network. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 3163-3172, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "text", + "content": "[60] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), pages 8026-8037, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[61] Biao Qian, Yang Wang, Richang Hong, and Meng Wang. Adaptive data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7960-7968, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "text", + "content": "[62] Akshit Ramachandran, Souvik Kundu, and Tushar Krishna. Clamp-vit: contrastive data-free learning for adaptive posttraining quantization of vits. In European Conference on Computer Vision (ECCV). Springer, 2024. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 554, + 285 + ], + "type": "text", + "content": "[63] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision (IJCV), 115:211-252, 2015. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 331 + ], + "type": "text", + "content": "[64] Fahad Shamshad, Salman Khan, Syed Waqas Zamir, Muhammad Haris Khan, Munawar Hayat, Fahad Shahbaz Khan, and Huazhu Fu. Transformers in medical imaging: A survey. Medical Image Analysis, page 102802, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "text", + "content": "[65] Huixin Sun, Runqi Wang, Yanjing Li, Xianbin Cao, Xiaolong Jiang, Yao Hu, and Baochang Zhang. P4q: Learning to prompt for quantization in visual-language models. arXiv preprint arXiv:2409.17634, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "text", + "content": "[66] Yehui Tang, Yunhe Wang, Yixing Xu, Yiping Deng, Chao Xu, Dacheng Tao, and Chang Xu. Manifold regularized dynamic network pruning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5018-5028, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "text", + "content": "[67] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12165-12174, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 488, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 554, + 544 + ], + "type": "text", + "content": "[68] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), pages 10347-10357. PMLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 578 + ], + "type": "text", + "content": "[69] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research (JMLR), 9:2579-2605, 2008. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 623 + ], + "type": "text", + "content": "[70] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 677 + ], + "type": "text", + "content": "[71] Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization. In International Conference on Learning Representations (ICLR), 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[72] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In IEEE/CVF In" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12489" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "ternational Conference on Computer Vision (ICCV), pages 10033-10041, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[73] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European Conference on Computer Vision (ECCV), pages 68-85, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[74] Zhiqiang Shen Xijie Huang and Kwang-Ting Cheng. Variation-aware vision transformer quantization. arXiv preprint arXiv:2307.00331, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 220 + ], + "type": "text", + "content": "[75] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 265 + ], + "type": "text", + "content": "[76] Shoukai Xu, Haokun Li, Bohan Zhuang, Jing Liu, Jiezhang Cao, Chuangrun Liang, and Mingkui Tan. Generative low-bitwidth data free quantization. In European Conference on Computer Vision (ECCV), pages 1-17. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "type": "text", + "content": "[77] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transffer via deep-inversion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8715-8724, 2020. 2, 3, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "text", + "content": "[78] Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization. arXiv preprint arXiv:1506.06579, 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "type": "text", + "content": "[79] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 818-833, 2014. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 412, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 294, + 467 + ], + "type": "text", + "content": "[80] Jinnian Zhang, Houwen Peng, Kan Wu, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Minivit: Compressing vision transformers with weight multiplexing. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12145-12154, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 533 + ], + "type": "text", + "content": "[81] Xiangguo Zhang, Haotong Qin, Yifu Ding, Ruihao Gong, Qinghua Yan, Renshuai Tao, Yuhang Li, Fengwei Yu, and Xianglong Liu. Diversifying sample generation for accurate data-free quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15658-15667, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "text", + "content": "[82] Dehua Zheng, Wenhui Dong, Hailin Hu, Xinghao Chen, and Yunhe Wang. Less is more: Focus attention for efficient detr. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 6674-6683, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 647 + ], + "type": "text", + "content": "[83] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 6881-6890, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[84] Yunshan Zhong, Mingbao Lin, Gongrui Nan, Jianzhuang Liu, Baochang Zhang, Yonghong Tian, and Rongrong Ji. Intraq: Learning synthetic images with intra-class heterogeneity for zero-shot network quantization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12339-12348, 2022. 2, 4, 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 162 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 117 + ], + "type": "text", + "content": "[85] Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&S-vit: An inclusive & stable method for pushing the limit of post-training vits quantization. arXiv preprint arXiv:2311.10126, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[86] Yunshan Zhong, Jiawei Hu, You Huang, Yuxin Zhang, and Rongrong Ji. Erq: Error reduction for post-training quantization of vision transformers. In International Conference on Machine Learning (ICML), 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12490" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_content_list.json b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4bd300623f5be72045a0f4a640b160168d75257f --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_content_list.json @@ -0,0 +1,1949 @@ +[ + { + "type": "text", + "text": "Semantic Causality-Aware Vision-Based 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dubing Chen $^{1}$ , Huan Zheng $^{1}$ , Yucheng Zhou $^{1}$ , Xianfei Li $^{2}$ , Wenlong Liao $^{2}$ , Tao He $^{2}$ , Pai Peng $^{2}$ , Jianbing Shen $^{1\\boxtimes}$ $^{1}$ SKL-IOTSC, CIS, University of Macau $^{2}$ COWAROBOT Co. Ltd. https://github.com/cdb342/CausalOcc", + "bbox": [ + 223, + 180, + 774, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-based 3D semantic occupancy prediction is a critical task in 3D vision that integrates volumetric 3D reconstruction with semantic understanding. Existing methods, however, often rely on modular pipelines. These modules are typically optimized independently or use pre-configured inputs, leading to cascading errors. In this paper, we address this limitation by designing a novel causal loss that enables holistic, end-to-end supervision of the modular 2D-to-3D transformation pipeline. Grounded in the principle of 2D-to-3D semantic causality, this loss regulates the gradient flow from 3D voxel representations back to the 2D features. Consequently, it renders the entire pipeline differentiable, unifying the learning process and making previously non-trainable components fully learnable. Building on this principle, we propose the Semantic Causality-Aware 2D-to-3D Transformation, which comprises three components guided by our causal loss: Channel-Groupled Lifting for adaptive semantic mapping, Learnable Camera Offsets for enhanced robustness against camera perturbations, and Normalized Convolution for effective feature propagation. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the Occ3D benchmark, demonstrating significant robustness to camera perturbations and improved 2D-to-3D semantic consistency.", + "bbox": [ + 86, + 335, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 728, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting dense 3D semantic occupancy is a fundamental task in 3D vision, providing a fine-grained voxel representation of scene geometry and semantics [39, 40, 43, 47]. The challenge of performing this prediction from vision alone,", + "bbox": [ + 89, + 753, + 483, + 816 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/688f0050e8c1b14bbd9adc04a36708ea013cc74e6c54793c41befb5f3f624927.jpg", + "image_caption": [ + "(a) An Example of Inaccurate 2D-to-3D Transformation" + ], + "image_footnote": [], + "bbox": [ + 519, + 285, + 898, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/716ae459e91b2d010789cd4c87e05ea9d2b792c95a6dc3f8e68044f365e20358.jpg", + "image_caption": [ + "(b) Modular 2D-to-3D Transformation" + ], + "image_footnote": [], + "bbox": [ + 521, + 411, + 898, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bc752e1a1159bbd1d1122c17f2aedd2d4b26b23bcdbc93afbee95d150546f4f.jpg", + "image_caption": [ + "(c) Our End-to-End Supervised 2D-to-3D Transformation", + "Figure 1. (a) Illustrates a Visual Analysis of Semantic Ambiguity in VisionOcc. Inaccurate 2D-to-3D transformations may lead to positional shifts, misaligning supervision signals and resulting in semantic ambiguity. (b) Depicts the Conventional Modular 2D-to-3D Transformation Paradigm [7, 13, 27, 35], which employs depth supervision for geometry estimation, pre-calibrated camera parameters, and fixed mapping for lifting. (c) Presents Our Holistic, End-to-End Supervised 2D-to-3D Transformation Paradigm, which eliminates the need for separate modular supervision or pre-calibration, enabling unified error propagation to supervise all components." + ], + "image_footnote": [], + "bbox": [ + 522, + 481, + 898, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an approach known as vision-based 3D semantic occupancy prediction (VisionOcc), has recently become a focal point of research. By leveraging only commodity cameras, VisionOcc is pivotal for a wide range of 3D applications, serving as a comprehensive digital replica of the environment for tasks like analysis, simulation, and interactive visualization [7, 9, 14, 27, 39, 40, 44].", + "bbox": [ + 511, + 731, + 906, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VisionOcc unifies the challenges of feed-forward 3D reconstruction and dense semantic understanding. Existing pipelines typically decompose this task into two metaphases [7, 14, 27, 30, 35]. The initial 2D-to-3D transfor", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\boxtimes$ Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL and CG2025-IOTSC, the University of Macau SRG2022-00023-IOTSC grant, and the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC).", + "bbox": [ + 89, + 827, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24878", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mation uses large receptive field operators like view lifting [7, 37] or cross attention [14, 27] to construct an initial 3D feature volume. Subsequently, a 3D representation learning phase employs operators with a small receptive field (e.g., 3D convolutions or local self-attention) to refine 3D features and produce the final prediction. Our work targets the 2D-to-3D transformation phase, which is more critical and error-prone. Fig. 1a showcases a primary failure mode where features of one class (e.g., a 2D 'car') are erroneously transformed to the 3D location of another (e.g., a 'tree'). This creates a flawed learning objective, forcing the model to learn a spurious association between 'car' features and a 'tree' label. Such semantic ambiguity is a principal obstacle to achieving high performance.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The prevailing VisionOcc methods, typically based on Lift-Splat-Shoot (LSS), employ a modular approach for the 2D-to-3D transformation (Fig. 1b) [7, 35, 37]. This involves supervising geometry with a proxy depth loss while relying on fixed, pre-calibrated camera parameters and a static lifting map. However, this modularity raises critical questions about robustness and optimality. First, it is susceptible to compounding errors; for example, the reliance on fixed camera parameters makes the system vulnerable to real-world perturbations like camera jitter during motion. More fundamentally, the optimality of such proxy supervision is questionable. An intermediate representation ideal for depth estimation may not be optimal for the final semantic occupancy task, inherently limiting the transformation's expressive power due to this objective misalignment. This motivates our central research question: Can we devise an end-to-end supervision framework1 that holistically optimizes the entire 2D-to-3D transformation, enabling unified semantic-aware error backpropagation and allowing traditionally fixed modules to become fully learnable?", + "bbox": [ + 91, + 303, + 482, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We approach this problem from a causal perspective. In VisionOcc, the 2D image semantics are the \"cause\" of the final 3D semantic \"effect\". Semantic misalignment arises from disrupted information flow from cause to effect (Fig. 1a). Therefore, instead of correcting the erroneous output, we propose to directly regularize the information flow itself. We posit that a 3D prediction for a given class should be influenced predominantly by 2D image regions of that same class. To enforce this information flow, we leverage gradients as a proxy, inspired by prior work [18, 42, 53]. For each semantic class, the gradient of its aggregated 3D features is computed w.r.t. the 2D feature map, producing a saliency-like map of 2D influence. This map is then directly supervised with the ground truth 2D segmentation mask. As shown in Fig. 1c, this establishes a principled, end-to-end supervision signal for the 2D-to-3D transformation, enabling holistic optimization of all its components.", + "bbox": [ + 89, + 606, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully leverage our end-to-end supervision, we introduce a more expressive and learnable 2D-to-3D view transformation, termed the Semantic Causality-Aware Transformation (SCAT). A key challenge is that directly supervising gradients is inherently unstable. Therefore, the entire SCAT module is designed to constrain its gradient flow to a stable [0, 1] range. Specifically, SCAT introduces three targeted designs: i) Channel-Grouped Lifting: To better disentangle semantics, we move beyond LSS's uniform weighting and apply distinct learnable weights to different groups of feature channels. ii) Learnable Camera Offsets: To mitigate motion-induced pose errors, we introduce learnable offsets to the camera parameters, which are implicitly supervised by the 2D-3D semantic consistency enforced by our causal loss. iii) Normalized Convolution: Finally, we employ a normalized convolution to densify the sparse 3D features from LSS [28], ensuring this final step also adheres to our global gradient stability requirement.", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows: i) We systematically analyze the 2D-to-3D transformation in VisionOCC, identifying a critical failure mode we term semantic ambiguity. We provide a theoretical analysis proving how the modularity of prior methods leads to error propagation, offering clear guidance for future work. ii) To address these problems, we propose the Causal Loss that directly regularizes the information flow of the 2D-to-3D transformation. This enables true end-to-end optimization of all constituent modules, mitigating error accumulation and making previously fixed components, such as camera parameters, fully learnable. iii) We instantiate our principles in the Semantic Causality-Aware Transformation, a novel 2D-to-3D transformation architecture. SCAT incorporates Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution to explicitly tackle the challenges of semantic confusion, camera perturbations, and limited learnability. iv) Extensive experiments show our method significantly boosts existing models, achieving a $3.2\\%$ absolute mIoU gain on BEVDet. Furthermore, it demonstrates superior robustness to camera perturbations, reducing the relative performance drop on BEVDet from a severe $-32.2\\%$ to a mere $-7.3\\%$ .", + "bbox": [ + 511, + 364, + 903, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 729, + 653, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semantic Scene Completion", + "text_level": 1, + "bbox": [ + 511, + 756, + 761, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Scene Completion (SSC) [10, 20, 24, 31, 51] refers to the task of simultaneously predicting both the occupancy and semantic labels of a scene. Existing methods can be classified into indoor and outdoor approaches based on the scene type, with the former focusing on occupancy and semantic label prediction in controlled environments [10, 20, 31], while the latter shifts towards more complex outdoor settings, particularly in the context of au", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Here, \"end to end\" refers to a unified supervision scheme for a process, distinct from the concept of a monolithic network architecture.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24879", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tonomous driving [1, 6]. The core principle of SSC lies in its ability to infer the unseen, effectively bridging gaps in incomplete observations with accurate semantic understanding. MonoScene [6] introduces a 3D SSC framework that infers dense geometry and semantics from a single monocular RGB image. VoxFormer [24] presents a Transformer-based semantic scene completion framework that generates complete 3D volumetric semantics from 2D images by first predicting sparse visible voxel queries and then densifying them through self-attention with a masked autoencoder design. OccFormer [51] introduces a dual-path transformer network for 3D semantic occupancy prediction, efficiently processing camera-generated 3D voxel features through local and global pathways, and enhancing occupancy decoding with preserve-pooling and class-guided sampling to address sparsity and class imbalance.", + "bbox": [ + 89, + 90, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision-based 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 89, + 340, + 426, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-based 3D Occupancy Prediction [9, 14-16, 34, 38-40, 43?] aims to predict the spatial and semantic features of 3D voxel grids surrounding an autonomous vehicle from image data. This task is closely related to SSC, emphasizing the importance of multi-perspective joint perception for effective autonomous navigation. TPVFormer [14] is prior work that lifts image features into the 3D TPV space by leveraging an attention mechanism [26, 41]. Different from TPVFormer, which relies on sparse point clouds for supervision, subsequent studies, including OccNet [40], SurroundOcc [44], Occ3D [39], and OpenOccupancy [43], have developed denser occupancy annotations by incorporating temporal information or instance-level labels. Methods such as BEVDet [13], FBOcc [27], COTR [35], and ALOcc [7] leverage depth-based LSS [22, 23, 25, 37] for explicit geometric transformation, demonstrating strong performance. Some methods [2, 4, 17, 36, 46, 49, 52] have explored rendering-based methods that utilize 2D signal supervision, thereby bypassing the need for 3D annotations. Furthermore, recent research like [3, 7, 8, 21, 29, 32, 40] introduced 3D occupancy flow prediction, which addresses the movement of foreground objects in dynamic scenes by embedding 3D flow information to capture per-voxel dynamics. Unlike the above methods, we analyze the 2D-to-3D transformation process from the perspectives of error propagation and semantic causal consistency, proposing a novel approach that enhances causal consistency.", + "bbox": [ + 91, + 363, + 483, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 784, + 181, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. VisionOcc predicts a dense 3D semantic occupancy grid from surround-view images, modeled as a causal dependency chain as shown in Fig. 2. This process involves key variables: input image $\\mathbf{I}$ , estimated geometry $\\mathbf{G}$ , camera parameters $P$ (intrinsic & extrinsic), potential camera parameter errors $e_{P}$ , intermediate 3D fea", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6dc0ed532564f3ef101e617c132dcadb62b9d4e31e01db2e726f7aeeb832c15.jpg", + "image_caption": [ + "Figure 2. The Causal Structure of VisionOcc. It illustrates the dependency chain from the image input $\\mathbf{I}$ to the semantic occupancy output $\\mathbf{O}$ . $\\mathbf{G}$ : geometry for 2D-to-3D transformation. $P$ : camera intrinsic and extrinsic. $\\mathbf{L}$ : intermediate 3D feature. $e_P$ : errors in camera parameters." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 898, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture $\\mathbf{L}$ , and final occupancy output $\\mathbf{O}$ . In LSS-based methods [7, 13, 35], the pipeline starts with the image backbone extracting features $\\mathbf{f}_i = F_i(\\mathbf{I}), \\mathbf{f}_i \\in \\mathbb{R}^{U,V,C}$ ; geometry $\\mathbf{G} \\in \\mathbb{R}^{U,V,D}$ is predicted as a probability distribution over discretized depth bins $D$ , i.e., $\\mathbf{G} = F_g(\\mathbf{f}_i)$ ; 2D features are then transformed to 3D via an outer product $\\mathbf{f}_L' = \\mathbf{G} \\otimes \\mathbf{f}_i, \\mathbf{f}_L' \\in \\mathbb{R}^{U,V,D,C}$ ; camera parameters $P$ map these to voxel coordinates $R_P(u,v,d) \\to (h,w,z) \\in [0,H-1] \\times [0,W-1] \\times [0,Z-1]$ , yielding $\\mathbf{f}_L \\in \\mathbb{R}^{H \\times W \\times Z}$ , where $H \\times W \\times Z$ defines the occupancy grid resolution; finally, $\\mathbf{L}$ is decoded to produce the semantic occupancy output: $\\mathbf{O} = F_o(\\mathbf{f}_L), \\mathbf{O} \\in \\mathbb{R}^{H \\times W \\times Z \\times S}$ , where $F_o$ is the decoding function and $S$ is the number of semantic classes. The prediction $\\mathbf{O}$ is supervised by the ground-truth $\\tilde{\\mathbf{O}}$ .", + "bbox": [ + 511, + 268, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Error Propagation in Depth-Based LSS", + "text_level": 1, + "bbox": [ + 511, + 489, + 852, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. In Depth-Based LSS methods with a fixed $2D$ -to-3D mapping $M_{\\text{fixed}}$ , inherent mapping error $\\delta M$ leads to gradient deviations, preventing convergence to an $\\epsilon$ -optimal solution. This is formalized as:", + "bbox": [ + 511, + 511, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {f i x e d} = M _ {i d e a l} + \\delta M \\Longrightarrow \\nabla_ {\\theta} L _ {L S S} \\neq \\nabla_ {\\theta} L _ {i d e a l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 583, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M_{ideal}$ is the ideal mapping and $L_{ideal}$ is the loss function using $M_{ideal}$ .", + "bbox": [ + 511, + 611, + 905, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. Mapping Error Quantification: Let $\\mathbf{x} \\in \\mathbb{R}^2$ be 2D pixel coordinates and $d(\\mathbf{x})$ be the ground truth depth. Estimated depth is $\\hat{d}(\\mathbf{x}) = d(\\mathbf{x}) + \\epsilon_d(\\mathbf{x})$ , with $\\epsilon_d(\\mathbf{x})$ as depth error. Let $\\mathbf{K}_{ideal} \\in \\mathbb{R}^{3 \\times 3}$ be the ideal camera intrinsics, and $\\mathbf{K} = \\mathbf{K}_{ideal} + \\epsilon_K$ be the estimated camera intrinsics with error $\\epsilon_K$ . The fixed mapping is $M_{fixed} = M_{ideal} + \\delta M$ , where $\\delta M$ encompasses errors from various sources, including depth estimation error $\\epsilon_d(\\mathbf{x})$ and camera extrinsic error $\\epsilon_K$ . We assume the total mapping error is bounded $\\| \\delta M \\|_F \\leq \\Delta_M < \\infty$ . The 3D coordinates are:", + "bbox": [ + 511, + 650, + 906, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = M _ {\\text {f i x e d}} (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 816, + 903, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {i d e a l}} = M _ {\\text {i d e a l}} (\\mathbf {x}, d (\\mathbf {x}), \\mathbf {K} _ {\\text {i d e a l}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 835, + 903, + 852 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {X} = \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} = \\delta M (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 856, + 903, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\Delta \\mathbf{X}\\| _2\\leq C\\cdot \\Delta_M$ for bounded inputs.", + "bbox": [ + 511, + 885, + 820, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24880", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/653d4f3314c43e2cb0d40b3486023ef5483b3b62d55a030e67bddeed13ad7dbf.jpg", + "image_caption": [ + "Figure 3. The Overall Framework of the Proposed Semantic Causality-Aware VisionOcc. The proposed framework consists of three primary components: a backbone network for extracting 2D features, an SCAT module for transforming these features into 3D space, and an Encoder-Decoder network for learning 3D semantics. The SCAT module is supervised by our causal loss." + ], + "image_footnote": [], + "bbox": [ + 181, + 85, + 816, + 210 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/96e2c267bf887b5ae625422ad3037bb568c80c636a8db887b2fc84ab9c78d7e5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoUmIoUDIoU
Depth-Based LSS44.540.478.9
SCL-Aware LSS50.5↑6.046.9↑6.585.7↑6.8
", + "bbox": [ + 120, + 276, + 450, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Performance of Depth-Based LSS vs. SCL-Aware LSS on Occ3D (in Ideal Conditions). BEVDetOcc is the baseline.", + "bbox": [ + 89, + 338, + 482, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Space Deviation and Loss: Let $F_{2D}(\\mathbf{x})$ be 2D features and $F_{3D}(\\cdot) = \\text{Lift}(F_{2D}(\\mathbf{x}), \\cdot)$ . Assuming Lift is $L_{\\text{Lift}}$ -Lipschitz continuous, the 3D feature deviation is:", + "bbox": [ + 89, + 375, + 483, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| F _ {3 D} (\\mathbf {X}) - F _ {3 D} \\left(\\mathbf {X} _ {\\text {i d e a l}}\\right) \\right\\| _ {F} \\leq L _ {\\text {L i f t}} \\| \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} \\| _ {2} \\tag {5} \\\\ \\leq L _ {L i f t} \\| \\Delta \\mathbf {X} \\| _ {2} \\leq \\Delta_ {F _ {3 D}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 431, + 482, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The loss function is $L_{LSS} = \\mathcal{L}(P_{3D}(\\mathbf{X}),GT_{3D})$ , where $P_{3D}(\\mathbf{X}) = Seg_{3D}(F_{3D}(\\mathbf{X}))$", + "bbox": [ + 89, + 478, + 482, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gradient Deviation and Optimization Limit: The gradient of $L_{LSS}$ w.r.t. parameters $\\theta$ is given by chain rule. However, since $M_{fixed}$ is fixed, $\\frac{\\partial\\mathbf{X}}{\\partial\\theta} = 0$ . Thus, the gradient becomes:", + "bbox": [ + 89, + 508, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} L _ {L S S} = \\frac {\\partial L _ {L S S}}{\\partial P _ {3 D}} \\frac {\\partial P _ {3 D}}{\\partial F _ {3 D}} \\left(\\frac {\\partial F _ {3 D}}{\\partial F _ {2 D}} \\frac {\\partial F _ {2 D}}{\\partial \\theta}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 577, + 482, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to $\\Delta_{F_{3D}} > 0$ , the computed gradient $\\nabla_{\\theta}L_{LSS}$ is based on the deviated feature space, i.e.,", + "bbox": [ + 89, + 619, + 482, + 651 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} L _ {L S S} (\\theta) = \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} (\\mathbf {X}), G T _ {3 D}\\right) \\tag {7} \\\\ \\neq \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} \\left(\\mathbf {X} _ {i d e a l}\\right), G T _ {3 D}\\right) = \\nabla_ {\\theta} L _ {i d e a l} (\\theta). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 661, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gradient deviation prevents mapping's direct optimization and limits convergence to an $\\epsilon$ -optimal solution.", + "bbox": [ + 89, + 708, + 483, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The theoretical analysis reveals that the inherent error in the fixed 2D-to-3D mapping of Depth-Based LSS methods fundamentally hinders gradient-based optimization from achieving optimal performance.", + "bbox": [ + 89, + 750, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Semantic Causal Locality in VisionOcc", + "text_level": 1, + "bbox": [ + 89, + 816, + 424, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As revealed in our theoretical analysis (Sec. 3.1), Depth-Based LSS methods suffer from inherent error propagation due to their fixed 2D-to-3D mapping, which limits optimization efficacy and potential performance. To overcome", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these limitations, we particularly focus on strengthening the semantic causality of the 2D-to-3D transformation. We argue that VisionOcc's 2D-to-3D semantic occupancy prediction should exhibit semantic causal locality (SCL) for robust perception in autonomous driving. Ideally, 2D causes should drive 3D semantic effects. For instance, a predicted \"car\" at 3D location $(h,w,z)$ should originate from a matching 2D image region of \"car\". Per the causal chain, camera parameters $P$ and estimated geometry $\\mathbf{G}$ enable this dependency, with $\\mathbf{G}$ being crucial to maintain SCL.", + "bbox": [ + 511, + 280, + 906, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we formulate the ideal SCL condition. For a 2D pixel $(u,v)$ with semantic label $s$ , the projection probability $p_d$ (the value of $\\mathbf{G}$ at this coordinate and depth $d \\in D$ ) should be high if its corresponding 3D ground-truth semantic is $s$ , and low otherwise:", + "bbox": [ + 511, + 430, + 906, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {d} \\propto \\mathbb {1} (\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) = s),\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 516, + 830, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}$ is the indicator function, and $e_P$ represents the potential coordinate transformation error caused by factors such as camera pose error. In practice, $p_d$ acts as a weight multiplied by 2D features (Eq. (8)), enabling a probabilistic mapping that supports differentiable backpropagation.", + "bbox": [ + 511, + 544, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations of Depth-Based LSS in SCL. Depth-Based LSS does not fully account for semantic causal consistency. It only preserves semantic causal locality intuitively under ideal conditions where $e_P = 0$ and depth estimation is perfectly accurate. However, with coordinate transformation errors $e_P$ , even a high $p_d$ for the ideal depth may project 2D semantics to incorrect 3D locations, i.e.,", + "bbox": [ + 511, + 619, + 906, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + 0) = s, \\text {b u t} \\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) \\neq s.\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 734, + 898, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This misalignment causes 2D semantics $s$ (e.g., \"car\") to link with wrong 3D semantics (e.g., \"tree\"), causing semantic ambiguity and hindering training. Moreover, depth-based LSS often propagates semantics to surface points, weakening semantic propagation to occluded regions [7].", + "bbox": [ + 511, + 763, + 905, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirical Validation. We conduct an empirical study to validate our analysis of SCL, comparing two ideal geometric transformations. Using BEVDetOcc [13] as the baseline, we replace its estimated depth-based geometry for LSS", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with: $i$ ) ground-truth LiDAR depths; $ii$ ) SCL-Aware geometry, which computes $p_d$ from 2D and 3D semantic ground truths, where for $(u,v,d,s)$ , $p_d = 1$ if $\\tilde{\\mathbf{O}}(R_P(u,v,d)) = s$ , else $p_d = 0$ . We evaluate their performance in semantic occupancy prediction. Tab. 1 demonstrates that 2D-to-3D Transformation achieves significant performance improvements over depth-based LSS in ideal conditions, validating the benefits of semantic causal locality.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary. Building on the above analysis, we propose our solution to enforce semantic causality constraints during training. The overall framework is shown in Fig. 3.", + "bbox": [ + 89, + 213, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Semantic Causality-Aware Causal Loss", + "text_level": 1, + "bbox": [ + 89, + 268, + 426, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For lifting methods like LSS, we could directly supervise the transformation geometry $\\mathbf{G}$ using the causal semantic geometry derived from ground-truth labels (as described in the previous section). However, we aim to enhance the lifting method in the next section, rendering direct supervision impractical. Thus, we design a gradient-based approach to enforce semantic causality.", + "bbox": [ + 89, + 292, + 483, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin within the LSS framework. For a 2D pixel feature at location $(u, v)$ , LSS multiplies it by the depth-related transformation probability $p_d$ and projects it to the 3D coordinate corresponding to depth $d$ :", + "bbox": [ + 89, + 398, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {L} \\left(R _ {P} (u, v, d)\\right) = p _ {d} (u, v, d) \\cdot \\mathbf {f} _ {i} (u, v, d). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 472, + 482, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{f}_L(R_P(u,v,d))\\in \\mathbb{R}^C$ represents the 3D voxel feature at the projected location, with $e_{P}$ omitted for notational simplicity. $\\mathbf{f}_i(u,v,d)\\in \\mathbb{R}^C$ denotes the 2D image feature at location $(u,v)$ . We backpropagate gradients from $\\mathbf{f}_L$ to $\\mathbf{f}_i$ , i.e.,", + "bbox": [ + 89, + 501, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(R _ {P} (u , v , d)\\right)}{\\partial \\mathbf {f} _ {i} (u , v , d)} = p _ {d} \\cdot \\mathbf {I}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 577, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}$ is the all-ones vector.", + "bbox": [ + 89, + 617, + 290, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each semantic class $s$ , we aggregate the features $\\mathbf{f}_L$ at all 3D positions where the ground truth class equals $s$ . This aggregation is backpropagated to the 2D features $\\mathbf{f}_i$ , yielding a gradient map $\\nabla_s \\in \\mathbb{R}^{U \\times V \\times C}$ for class $s$ :", + "bbox": [ + 89, + 633, + 483, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {s} (u, v, c) = \\sum_ {\\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\Omega_ {s}} \\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c\\right)}{\\partial \\mathbf {f} _ {i} (u , v , c)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 705, + 482, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega_s = \\{(h', w', z') \\mid O(h', w', z') = s\\}$ is the set of 3D positions with semantic label $s$ , and $c$ indexes the feature channels. Averaging over channel $C$ produces an attention map $A_s \\in \\mathbb{R}^{U \\times V}$ :", + "bbox": [ + 89, + 757, + 483, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {s} (u, v) = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\nabla_ {s} (u, v, c). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 832, + 482, + 872 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we enforce per-pixel constraints using 2D", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ground-truth labels with a binary cross-entropy loss:", + "bbox": [ + 511, + 90, + 859, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {b c e} ^ {s} = - \\frac {1}{U \\cdot V} \\sum_ {u, v} \\left[ Y _ {s} (u, v) \\log A _ {s} (u, v) \\right. \\tag {12} \\\\ \\left. + \\left(1 - Y _ {s} (u, v)\\right) \\log (1 - A _ {s} (u, v)) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 114, + 903, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{s}(u,v) \\in \\{0,1\\}$ is the 2D ground-truth label for semantic class $s$ at pixel $(u,v)$ .", + "bbox": [ + 511, + 181, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gradient computation can be performed using the automatic differentiation calculator like torch.autograd, requiring $S$ backward passes to iterate over the $S$ semantic classes. This incurs a computational overhead scaling linearly with the class amount $S$ . To mitigate this overhead, we reformulate the loss computations in terms of the expectation of an unbiased estimator [11]. We define the expected BCE loss across all semantic classes $s \\in \\{1, \\dots, S\\}$ as:", + "bbox": [ + 511, + 212, + 905, + 333 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ L _ {b c e} ^ {s} \\right] = \\frac {1}{S} \\sum_ {s = 1} ^ {S} L _ {b c e} ^ {s}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 343, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this relationship, we uniformly sample a single semantic class $s$ during training:", + "bbox": [ + 511, + 392, + 903, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c a u s a l} = L _ {b c e} ^ {s}, \\quad s \\sim \\operatorname {U n i f o r m} (1, S). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 433, + 903, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This sampling preserves the unbiased nature of the gradient and loss estimates and reduces the computational cost to $\\frac{1}{S}$ . $L_{\\text{causal}}$ focuses on enhancing the geometric transformation, serving as an auxiliary term to complement the occupancy loss (case-by-case, e.g., cross-entropy in BEVDet [13]).", + "bbox": [ + 511, + 459, + 905, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Semantic Causality-Aware Transformation", + "text_level": 1, + "bbox": [ + 511, + 544, + 877, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose semantic causality-aware 2D-to-3D transformation to enhance 2D-to-3D lifting, as shown in Fig. 4. Eq. (14) constrains the geometry $\\mathbf{G}$ using 2D and 3D semantics. This overcomes the rigid, per-location hard alignment of geometric probabilities (e.g., using LiDAR depth supervision) in prior methods [7, 13, 22, 23, 27]. It enables advanced lifting designs, addressing errors from camera pose and other distortions.", + "bbox": [ + 511, + 566, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Channel-Grouped Lifting", + "text_level": 1, + "bbox": [ + 511, + 694, + 740, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vanilla LSS applies uniform weights to all 2D feature channels. We argue this is trivial as 2D and 3D features have distinct locality biases. For instance, a 2D \"car\" edge may capture \"tree\" semantics via convolution, but in 3D, these objects are distant. Uniform weighting both semantics causes ambiguity. Since different channels typically encode distinct semantics, we group the feature channels and learn unique weights for each group:", + "bbox": [ + 511, + 712, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {L, g} \\left(R _ {P} (u, v, d)\\right) = \\omega_ {g, d} \\cdot \\mathbf {f} _ {i, g} (u, v, d), g \\in \\{1, \\dots , N _ {g} \\}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 844, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{L,g}\\in \\mathbb{R}^{C / N_g}$ and $\\mathbf{f}_{i,g}\\in \\mathbb{R}^{C / N_g}$ are the 3D and 2D features for group $g$ . $\\omega_{g,d}$ is the learned weight for group", + "bbox": [ + 511, + 868, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24882", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b35fc75d231592dfad58e5042f4dec235ce01c814d7024733f9396f3c81da568.jpg", + "image_caption": [ + "(a) Channel-Grouped Lifting" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/984cc9ef1d47c4edd313521fd39c3ec47f8a062f101a03dee85ded2f278e13f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 203, + 478, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ea1a77761edd9899688dd59f5533e25f88bab654f20db6ecd83d107b909228b.jpg", + "image_caption": [ + "(b) Coordinate Mapping with Learnable Camera Offsets", + "(c) Post-Processing with Normalized Convolution", + "Figure 4. The Detailed Structure of the Proposed Modules in Semantic Causality-Aware 2D-to-3D Transformation. There are three key components, including (a): Channel-Groupled Lifting, (b): Learnable Camera Offsets, and (c): Normalized Convolution, enabling accurate and robust 2D-to-3D transformation." + ], + "image_footnote": [], + "bbox": [ + 109, + 290, + 473, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$g$ , replacing $p_d$ which uniformly lifts all channels. $N_g$ is the number of groups. This preserves semantic distinction, ensuring channel-specific causal alignment.", + "bbox": [ + 89, + 454, + 483, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.2. Learnable Camera Offsets", + "text_level": 1, + "bbox": [ + 89, + 508, + 321, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address camera parameter errors, especially pose inaccuracies, we introduce learnable offsets into camera parameters. First, we ensure the lifting process is coordinatedifferentiable. This is crucial for the offsets to receive gradients and adapt during training. The transformation of 2D image coordinates $(u,v)$ and depth $d$ to 3D voxel coordinates can be represented as matrix multiplication:", + "bbox": [ + 89, + 527, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n[ h, w, z ] ^ {T} = P \\cdot [ u \\cdot d, v \\cdot d, d, 1 ] ^ {T}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 643, + 482, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{R}^{3 \\times 4}$ is the camera projection matrix combining intrinsics and extrinsics. LSS typically rounds floating-point coordinates to integers, rendering them nondifferentiable. Following ALOcc [7], we use the soft filling to enable differentiability w.r.t. position. This method calculates distances between floating-point 3D coordinates and their eight surrounding integer coordinates. These distances serve as weights to distribute a 2D feature at $(u, v, d)$ across multiple 3D locations. Lifting can be rewritten to", + "bbox": [ + 89, + 674, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {f} _ {L, g} \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) = \\omega_ {g, d} \\cdot \\omega_ {h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}} \\cdot \\mathbf {f} _ {i, g} (u, v, d), \\tag {17} \\\\ \\forall \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\text {n e i g h b o r s}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 821, + 482, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega_{h',w',z'}$ are trilinear interpolation weights.", + "bbox": [ + 89, + 869, + 419, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we propose learning two offsets. First, we directly", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "predict an offset applied to the camera parameters:", + "bbox": [ + 511, + 90, + 849, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP := P + \\Delta P, \\quad \\Delta P = F _ {\\text {o f f s e t 1}} \\left(\\mathbf {f} _ {i}, P\\right), \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 114, + 903, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Delta P$ is the predicted parameter offset, and $F_{offset1}$ denotes the network. Second, we estimate per-position offsets for each $(u,v,d)$ in the image coordinate system:", + "bbox": [ + 511, + 138, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(u, v, d) := (u + \\Delta u, v + \\Delta v, d + \\Delta d), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 194, + 903, + 217 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\Delta u, \\Delta v, \\Delta d\\right) = F _ {\\text {o f f s e t} 2} \\left(\\mathbf {f} _ {i} (u, v, d)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 212, + 843, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F_{offset2}$ is another network. These offsets enable the model to adaptively compensate for camera parameter errors $e_P$ , improving geometric accuracy while preserving semantic causal locality under such errors.", + "bbox": [ + 511, + 237, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.3. Normalized Convolution", + "text_level": 1, + "bbox": [ + 511, + 304, + 730, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prior work [28] notes that the direct mapping in LSS yields sparse 3D features. To address this, an intuitive solution is to use local feature propagation operators (e.g., convolutions) with causal loss supervision for preserving semantic causality. However, vanilla convolutions lack gradient constraints during causal loss computation, producing poor results. We address this by normalizing convolution weights to keep gradient map values in [0, 1]. Given the difficulty of normalizing standard convolution weights, we follow MobileNet and ConvNext to adopt depthwise (spatial) and pointwise (channel) decomposition. For the depthwise kernel $W_{\\mathrm{spatial}} \\in \\mathbb{R}^{3 \\times 3 \\times 3 \\times C}$ , we apply softmax across spatial dimensions $(h, w, z)$ per channel $c$ :", + "bbox": [ + 511, + 321, + 906, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {s p a t i a l}} ^ {\\prime} [ h, w, z, c ] = \\frac {\\exp \\left(W _ {\\text {s p a t i a l}} [ h , w , z , c ]\\right)}{\\sum_ {h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime}} \\exp \\left(W _ {\\text {s p a t i a l}} \\left[ h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c \\right]\\right)}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 526, + 903, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the pointwise kernel $W_{\\mathrm{channel}} \\in \\mathbb{R}^{C \\times C}$ , we apply softmax across input channels per output channel:", + "bbox": [ + 511, + 563, + 903, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {c h a n n e l}} ^ {\\prime} \\left[ c _ {\\text {i n}}, c _ {\\text {o u t}} \\right] = \\frac {\\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} \\right]\\right)}{\\sum_ {c _ {\\text {o u t}} ^ {\\prime}} \\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} ^ {\\prime} \\right]\\right)}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 599, + 903, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we use transposed convolution for depthwise convolutions. It diffuses features from non-zero to zero positions, but the zero position does not affect others. We prove in the supplement that the derived gradient mask remains within [0, 1] even with our semantic causality-aware 2D-to-3D transformation.", + "bbox": [ + 511, + 643, + 905, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Validation of Gradient Error Mitigation", + "text_level": 1, + "bbox": [ + 511, + 742, + 856, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 shows the occupancy loss curves during training for BEVDet, comparing performance with and without our method. The results show that BEVDet integrated with our approach (blue line) achieves a significantly faster and steeper loss reduction compared to the original BEVDet (red line). This empirical evidence validates our theoretical analysis of gradient error mitigation. As Theorem Theorem 1 formalizes, Depth-Based LSS methods are inherently limited by gradient deviations due to fixed 2D-to-3D", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24883", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c7e40b5c0db6937e75f1633b2673a8b1d62dfa14ed19263daf7d4c34633847c.jpg", + "image_caption": [ + "Figure 5. Occupancy Loss Curves with/without Our Method. Our method reduces training loss through semantic causal 2D-to-3D geometry transformation, as shown in the comparison before (red) and after (blue) its application." + ], + "image_footnote": [], + "bbox": [ + 145, + 88, + 428, + 265 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/221af839a449acb9fb887c1b836d27fe4f4cf7e5bd09b236ac2b8d06ba334b47.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoU ↑DropmIoUD↑DropIoU ↑Drop
BEVDetOcc [12]37.1-32.3%30.2-49.0%70.4-8.5%
BEVDetOcc w/ Noise25.115.464.4
BEVDetOcc+Ours38.3-7.3%31.5-10.8%71.2-1.4%
BEVDetOcc+Ours w/ Noise35.528.170.2
ALoCC [7]40.1-21.9%34.3-28.6%70.2-8.4%
ALoCC w/ Noise31.324.564.3
ALoCC+Ours40.9-3.3%35.5-4.8%70.7-1.0%
ALoCC+Ours w/ Noise39.633.870.0
", + "bbox": [ + 91, + 334, + 485, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance Comparison on the Occ3D Dataset with Gaussian Noise Added to Camera Parameters. The \"Drop (%) columns show degradation, with our methods (BEVDetOcc+Ours, ALOcc+Ours) achieving much smaller drops (e.g., -7.3% mIoU vs. -32.4% for BEVDetOcc).", + "bbox": [ + 89, + 473, + 483, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mapping errors. In contrast, our method aims to alleviate this by enabling a learnable mapping and incorporating a causal loss. It indicates that by mitigating gradient error through semantic causality-aware 2D-to-3D transformation, our approach facilitates more efficient gradient-based learning, leading to faster convergence and a lower loss curve.", + "bbox": [ + 89, + 553, + 483, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 657, + 217, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 681, + 281, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset. In this study, we leverage the Occ3D-nuScenes dataset [5, 39], a comprehensive dataset with diverse scenes for autonomous driving research. It encompasses 700 scenes for training, 150 for validation, and 150 for testing. Each scene integrates a 32-beam LiDAR point cloud alongside 6 RGB images, captured from multiple perspectives encircling the ego vehicle. Occ3D [39] introduces voxel-based annotations, covering a spatial extent of $-40\\mathrm{m}$ to $40\\mathrm{m}$ along the X and Y axes, and $-1\\mathrm{m}$ to $5.4\\mathrm{m}$ along the Z axis, with a consistent voxel size of $0.4\\mathrm{m}$ across all dimensions. Occ3D delineates 18 semantic categories, comprising 17 distinct object classes plus an empty class to signify unoccupied regions. Following [7, 27, 35, 39], we evalu", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a25e7a9ca91d4b12c60c8b544975b4dae0d485f86571abe220867e095cae644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 160 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36e3fcbb843c288828c0a7573947a8a3e5ddecb152ab327f76a7f277a3d5dfac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 162, + 903, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0136db64c9a04fd9b4b8a19aa6469c5da59aba4c623a0ecb0d5af6e88d9e22bb.jpg", + "image_caption": [ + "(a) BEVDet" + ], + "image_footnote": [], + "bbox": [ + 519, + 234, + 705, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc7f33cc3030f9bbcf9ee9c392bfbdbb23a50a0f4c906b31ae2a168ac094e284.jpg", + "image_caption": [ + "(b) BEVDet + Ours", + "Figure 6. Visualization of 2D-to-3D Semantic Causal Consistency Using LayerCAM [18]. We enhance BEVDet with our method for comparison. Attention maps are computed for critical traffic classes: \"traffic cone\", \"pedestrian\", and \"car\". Areas of greatest difference are marked with boxes. Each class-specific localization highlights our method's precise focus over vanilla BEVDet, showing improved semantic alignment." + ], + "image_footnote": [], + "bbox": [ + 709, + 234, + 903, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ate the occupancy prediction performance with mIoU across 17 semantic object categories, $\\mathrm{mIoU}_D$ for 8 dynamic categories, and occupied/unoccupied IoU for scene geometry.", + "bbox": [ + 511, + 431, + 905, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We integrate our approach into BEVDet [12] and ALOcc [7] for performance evaluation. The model parameters, image, and BEV augmentation strategies are retrained as the original. For ALOcc, we remove its ground-truth depth denoising module to adapt to our method. We use multiple convolutional layers to predict the two camera parameter offsets. The loss weight of $L_{\\text{causal}}$ is set to 0.02 in the main experiments. We optimize using AdamW [33] with a learning rate of $2 \\times 10^{-4}$ , a global batch size of 16, and for 24 epochs. Experiments use single-frame surrounding images, emphasizing improvements from enhanced 2D-to-3D transformations.", + "bbox": [ + 511, + 477, + 905, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation of Camera Perturbation Robustness", + "text_level": 1, + "bbox": [ + 511, + 666, + 903, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess our method's robustness under extreme noise, we add Gaussian noise with (0.1 variances) to camera parameters in training and testing, as shown in Tab. 2. Compared to vanilla BEVDetOcc and ALOcc, our methods show smaller performance drops. BEVDetOcc+Ours has a mIoU drop of $-7.3\\%$ vs. $-32.3\\%$ for BEVDetOcc, and ALOcc+Ours drops $-3.3\\%$ vs. $-21.9\\%$ for ALOcc, demonstrating enhanced resilience to noisy parameters. This effectively counters motion-induced errors, benefiting self-driving tasks.", + "bbox": [ + 511, + 688, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Semantic Causality Visualization", + "text_level": 1, + "bbox": [ + 511, + 833, + 803, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the Fig. 6, we visualize 3D-to-2D semantic causal consistency using LayerCAM [18]. We backpropagate the final 3D semantic occupancy predictions of dis", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24884", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/20d18994ba23a7982a510eb6e699ce1cc1597209f8aaf830df71988bd7982675.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneInput SizemIoUmIoUDIoU
MonoScene [6]ResNet-101928 × 16006.15.4-
CTF-Occ [39]ResNet-101928 × 160028.527.4-
TPVFormer [14]ResNet-101928 × 160027.827.2-
COTR [35]ResNet-50256 × 70439.133.869.6
ProtoOcc [19]ResNet-50256 × 70439.634.3-
LightOcc-S [50]ResNet-50256 × 70437.932.4-
DHD-S [45]ResNet-50256 × 70436.530.7-
FlashOCC [48]ResNet-50256 × 70432.024.765.3
FB-Occ [27]ResNet-50256 × 70435.730.966.5
BEVDetOcc [12]ResNet-50256 × 70437.130.270.4
BEVDetOcc+OursResNet-50256 × 70438.3 ↑1.231.5 ↑1.371.2 ↑1.2
ALOcc [7]ResNet-50256 × 70440.134.370.2
ALOcc+OursResNet-50256 × 70440.9 ↑0.835.5 ↑1.170.7 ↑0.5
", + "bbox": [ + 94, + 88, + 478, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tinct classes to the 2D feature maps fed into SCAT. Notably, we are the first to apply LayerCAM for cross-dimensional analysis. Fig. 6 (b) shows our method precisely focuses on class-specific locations. This confirms LayerCAM's cross-dimensional effectiveness. Our approach surpasses vanilla BEVDet (Fig. 6 (a)) in targeting class-associated objects. This proves improved semantic localization.", + "bbox": [ + 89, + 364, + 483, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Benchmarking with Previous Methods", + "text_level": 1, + "bbox": [ + 89, + 481, + 421, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 3, we compare our method with leading 3D semantic occupancy prediction approaches on Occ3D [39]. Specifically, compared to baseline models BEVDet and ALOcc, our method achieves significant improvements in mIoU, mIoU $_{\\mathrm{D}}$ , and IoU. For instance, the BEVDetOcc+Ours variant achieves an mIoU of 38.3, surpassing BEVDetOcc [12] by 1.2, while improving mIoU $_{\\mathrm{D}}$ by 1.3 and IoU by 1.2. Similarly, ALOcc+Ours shows gains of 0.8 in mIoU, 1.1 in mIoU $_{\\mathrm{D}}$ , and 0.5 in IoU over ALOcc [7]. These results validate the superiority of our semantic causality-aware 2D-to-3D transformation.", + "bbox": [ + 89, + 503, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 681, + 243, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Causal Loss. We first investigate the effectiveness of the proposed Causal Loss in enhancing the occupancy prediction performance. As demonstrated in Tab. 4, the impact of the proposed Causal Loss is evaluated through a series of experiments. Specifically, using BEVDetOcc as the baseline (Exp. 0), we conducted two ablation studies: one removing the depth supervision loss (Exp. 1), and another incorporating the proposed Causal Loss (Exps. 2, 3). The results reveal that the removal of the depth supervision loss leads to marginal decrease in performance, whereas the addition of the proposed Causal Loss yields a significant improvement. This suggests that Causal Loss facilitates superior 2D-to-3D transformation and ultimately enhances the", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aab475daf81f77cbb53c8fbf8bcdcf502b02c53e114491b7d448639be8e953b5.jpg", + "table_caption": [ + "Table 3. Comparison of 3D Semantic Occupancy Prediction Using Single Frame on the Occ3D Dataset, Evaluated mIoU, mIoUD, and IoU Metrics. Performance gains are indicated by red arrows $\\uparrow$ . Our proposed approach (+Ours) consistently demonstrates superior enhancement over existing methods." + ], + "table_footnote": [], + "table_body": "
Exp.MethodmIoUDiff.mIoUDIoULatency
0Baseline (BEVDetOcc) [12]37.1-30.270.4416/125
1w/o Depth Sup36.8-0.329.670.3414/125
2+ Causal Loss37.6+0.831.070.1450/125
3+ Unbiased Estimator37.5-0.130.770.5417/125
4w/o Post Conv37.3-0.230.770.2379/122
5+ Channel-Grouped Lifting37.6+0.330.770.7419/128
6+ Soft Filling37.6-30.670.7434/149
7+ Learnable Camera Offset37.9+0.331.171.0446/150
8+ Normalized Convolution38.3+0.431.571.2466/159
", + "bbox": [ + 517, + 90, + 903, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Study of 3D Semantic Occupancy Prediction on Occ3D. We comprehensively evaluated the impact of the individual strategy (bolded rows) proposed in our paper, with BEVDetOcc as the baseline. The final column reports single-frame training/inference latency (ms) on an RTX 4090 GPU.", + "bbox": [ + 511, + 258, + 906, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "precision of 3D semantic occupancy prediction. Comparing Exp. 3 to Exp. 2, the Unbiased Estimator simplifies the Causal Loss computation, reducing training overhead.", + "bbox": [ + 511, + 339, + 903, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Each Module. In Tab. 4 (Exps. 5-8), we systematically validate the effectiveness of the three proposed modules. We first remove the two post-lifting convolutional layers from the original BEVDet (Exp. 4), as they serve a similar role to our proposed modules in refining volume features. Subsequently, we incrementally integrate the three proposed modules (Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution) into the baseline model. To ensure gradient flow for Learnable Camera Offsets, we introduce the Soft Filling strategy from ALOcc [7] in Exp. 6, enabling effective training of the camera parameter offsets in Exp. 7. The results show progressive performance improvements with each added component (Exps. 5, 7, 8), confirming the efficacy of the proposed SCAT method. Additionally, the proposed modules incur acceptable computational overhead.", + "bbox": [ + 511, + 383, + 905, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Please refer to the supplementary material for more comprehensive experimental results.", + "bbox": [ + 511, + 626, + 905, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 666, + 633, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel approach leveraging causal principles to address the limitation that ignored the reliability and interpretability by existing methods. By exploring the causal foundations of 3D semantic occupancy prediction, we propose a causal loss that enhances semantic causal consistency. In addition, we develop the SCAT module with three main components: Channel-Groupled Lifting, Learnable Camera Offsets and Normalized Convolution. This approach effectively mitigates transformation inaccuracies arising from uniform mapping weights, camera perturbations, and sparse mappings. Experiments demonstrate that our approach achieves significant improvements in accuracy, robustness to camera perturbations, and semantic causal consistency in 2D-to-3D transformations.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24885", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 3", + "[2] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Langocc: Self-supervised open vocabulary occupancy estimation via volume rendering. arXiv preprint arXiv:2407.17310, 2024. 3", + "[3] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Occfownet: Towards self-supervised occupancy estimation via differentiable rendering and occupancy flow. arXiv preprint arXiv:2402.12792, 2024. 3", + "[4] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Gaussianflowocc: Sparse and weakly supervised occupancy estimation using gaussian splatting and temporal flow. arXiv preprint arXiv:2502.17288, 2025. 3", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 7", + "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 8", + "[7] Dubing Chen, Jin Fang, Wencheng Han, Xinjing Cheng, Junbo Yin, Chengzhong Xu, Fahad Shahbaz Khan, and Jianbing Shen. Alocc: adaptive lifting-based 3d semantic occupancy and cost volume-based flow prediction. arXiv preprint arXiv:2411.07725, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[8] Dubing Chen, Wencheng Han, Jin Fang, and Jianbing Shen. Adaocc: Adaptive forward view transformation and flow modeling for 3d occupancy and flow prediction. arXiv preprint arXiv:2407.01436, 2024. 3", + "[9] Dubing Chen, Huan Zheng, Jin Fang, Xingping Dong, Xianfei Li, Wenlong Liao, Tao He, Pai Peng, and Jianbing Shen. Rethinking temporal fusion with a unified gradient descent view for 3d semantic occupancy prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1505-1515, 2025. 1, 3", + "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4193-4202, 2020. 2", + "[11] Judy Hoffman, Daniel A Roberts, and Sho Yaida. Robust learning with jacobian regularization. arXiv preprint arXiv:1908.02729, 2019. 5", + "[12] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7, 8", + "[13] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 4, 5", + "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023. 1, 2, 3, 8", + "[15] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024.", + "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376-393. Springer, 2024. 3", + "[17] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding. arXiv preprint arXiv:2412.13193, 2024. 3", + "[18] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 7", + "[19] Jungho Kim, Changwon Kang, Dongyoung Lee, Sehwan Choi, and Jun Won Choi. Protoocc: Accurate, efficient 3d occupancy prediction using dual branch encoder-prototype query decoder. arXiv preprint arXiv:2412.08774, 2024. 8", + "[20] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2019. 2", + "[21] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3", + "[22] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 5", + "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 3, 5", + "[24] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 2, 3", + "[25] Yangguang Li, Bin Huang, Zeren Chen, Yufeng Cui, Feng Liang, Mingzhu Shen, Fenggang Liu, Enze Xie, Lu Sheng, Wanli Ouyang, et al. Fast-bev: A fast and strong bird's" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24886", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eye view perception baseline. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3", + "[27] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 2, 3, 5, 7, 8", + "[28] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 6", + "[29] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR 2024 Autonomous Grand Challenge Track On Occupancy and Flow, 2024. 3", + "[30] Lizhe Liu, Bohua Wang, Hongwei Xie, Daqi Liu, Li Liu, Zhiqiang Tian, Kuiyuan Yang, and Bing Wang. Surroundsdf: Implicit 3d scene understanding based on signed distance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[31] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[32] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 3", + "[33] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. 7", + "[34] Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023. 3", + "[35] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. arXiv preprint arXiv:2312.01919, 2023. 1, 2, 3, 7, 8", + "[36] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12404-12411. IEEE, 2024. 3", + "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2, 3" + ], + "bbox": [ + 91, + 92, + 482, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as set of points. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3", + "[39] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7, 8", + "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 1, 3", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[42] Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160, 2023. 2", + "[43] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 3", + "[44] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551, 2023. 1, 3", + "[45] Yuan Wu, Zhiqiang Yan, Zhengxue Wang, Xiang Li, Le Hui, and Jian Yang. Deep height decoupling for precise vision-based 3d occupancy prediction. arXiv preprint arXiv:2409.07972, 2024.8", + "[46] Ziyang Yan, Wenzhen Dong, Yihua Shao, Yuhang Lu, Liu Haiyang, Jingwen Liu, Haozhe Wang, Zhe Wang, Yan Wang, Fabio Remondino, et al. Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356, 2024.3", + "[47] Shichao Yang, Yulan Huang, and Sebastian Scherer. Semantic 3d occupancy mapping through efficient high order crfs. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 590-597. IEEE, 2017. 1", + "[48] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 8", + "[49] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 3", + "[50] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Lightweight spatial embedding for vision-based 3d occupancy prediction. arXiv preprint arXiv:2412.05976, 2024.8" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24887", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 2, 3", + "[52] Jilai Zheng, Pin Tang, Zhongdao Wang, Guoqing Wang, Xiangxuan Ren, Bailan Feng, and Chao Ma. Veon: Vocabulary-enhanced occupancy prediction. In European Conference on Computer Vision, pages 92-108. Springer, 2024. 3", + "[53] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 15890-15902, 2024. 2" + ], + "bbox": [ + 89, + 90, + 482, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24888", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_model.json b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e66e0207bd655a7f966fbdf4a072b3e42918d28 --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Semantic Causality-Aware Vision-Based 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.181, + 0.775, + 0.253 + ], + "angle": 0, + "content": "Dubing Chen\\(^{1}\\), Huan Zheng\\(^{1}\\), Yucheng Zhou\\(^{1}\\), Xianfei Li\\(^{2}\\), Wenlong Liao\\(^{2}\\), Tao He\\(^{2}\\), Pai Peng\\(^{2}\\), Jianbing Shen\\(^{1\\boxtimes}\\) \n\\(^{1}\\)SKL-IOTSC, CIS, University of Macau \\(^{2}\\)COWAROBOT Co. Ltd. https://github.com/cdb342/CausalOcc" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.336, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Vision-based 3D semantic occupancy prediction is a critical task in 3D vision that integrates volumetric 3D reconstruction with semantic understanding. Existing methods, however, often rely on modular pipelines. These modules are typically optimized independently or use pre-configured inputs, leading to cascading errors. In this paper, we address this limitation by designing a novel causal loss that enables holistic, end-to-end supervision of the modular 2D-to-3D transformation pipeline. Grounded in the principle of 2D-to-3D semantic causality, this loss regulates the gradient flow from 3D voxel representations back to the 2D features. Consequently, it renders the entire pipeline differentiable, unifying the learning process and making previously non-trainable components fully learnable. Building on this principle, we propose the Semantic Causality-Aware 2D-to-3D Transformation, which comprises three components guided by our causal loss: Channel-Groupled Lifting for adaptive semantic mapping, Learnable Camera Offsets for enhanced robustness against camera perturbations, and Normalized Convolution for effective feature propagation. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the Occ3D benchmark, demonstrating significant robustness to camera perturbations and improved 2D-to-3D semantic consistency." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.729, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Predicting dense 3D semantic occupancy is a fundamental task in 3D vision, providing a fine-grained voxel representation of scene geometry and semantics [39, 40, 43, 47]. The challenge of performing this prediction from vision alone," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.9 + ], + "angle": 0, + "content": "\\(\\boxtimes\\) Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL and CG2025-IOTSC, the University of Macau SRG2022-00023-IOTSC grant, and the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC)." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.286, + 0.9, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.396, + 0.86, + 0.409 + ], + "angle": 0, + "content": "(a) An Example of Inaccurate 2D-to-3D Transformation" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.412, + 0.9, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.47, + 0.813, + 0.481 + ], + "angle": 0, + "content": "(b) Modular 2D-to-3D Transformation" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.482, + 0.9, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.537, + 0.864, + 0.55 + ], + "angle": 0, + "content": "(c) Our End-to-End Supervised 2D-to-3D Transformation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Figure 1. (a) Illustrates a Visual Analysis of Semantic Ambiguity in VisionOcc. Inaccurate 2D-to-3D transformations may lead to positional shifts, misaligning supervision signals and resulting in semantic ambiguity. (b) Depicts the Conventional Modular 2D-to-3D Transformation Paradigm [7, 13, 27, 35], which employs depth supervision for geometry estimation, pre-calibrated camera parameters, and fixed mapping for lifting. (c) Presents Our Holistic, End-to-End Supervised 2D-to-3D Transformation Paradigm, which eliminates the need for separate modular supervision or pre-calibration, enabling unified error propagation to supervise all components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.838 + ], + "angle": 0, + "content": "an approach known as vision-based 3D semantic occupancy prediction (VisionOcc), has recently become a focal point of research. By leveraging only commodity cameras, VisionOcc is pivotal for a wide range of 3D applications, serving as a comprehensive digital replica of the environment for tasks like analysis, simulation, and interactive visualization [7, 9, 14, 27, 39, 40, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "VisionOcc unifies the challenges of feed-forward 3D reconstruction and dense semantic understanding. Existing pipelines typically decompose this task into two metaphases [7, 14, 27, 30, 35]. The initial 2D-to-3D transfor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24878" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "mation uses large receptive field operators like view lifting [7, 37] or cross attention [14, 27] to construct an initial 3D feature volume. Subsequently, a 3D representation learning phase employs operators with a small receptive field (e.g., 3D convolutions or local self-attention) to refine 3D features and produce the final prediction. Our work targets the 2D-to-3D transformation phase, which is more critical and error-prone. Fig. 1a showcases a primary failure mode where features of one class (e.g., a 2D 'car') are erroneously transformed to the 3D location of another (e.g., a 'tree'). This creates a flawed learning objective, forcing the model to learn a spurious association between 'car' features and a 'tree' label. Such semantic ambiguity is a principal obstacle to achieving high performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.605 + ], + "angle": 0, + "content": "The prevailing VisionOcc methods, typically based on Lift-Splat-Shoot (LSS), employ a modular approach for the 2D-to-3D transformation (Fig. 1b) [7, 35, 37]. This involves supervising geometry with a proxy depth loss while relying on fixed, pre-calibrated camera parameters and a static lifting map. However, this modularity raises critical questions about robustness and optimality. First, it is susceptible to compounding errors; for example, the reliance on fixed camera parameters makes the system vulnerable to real-world perturbations like camera jitter during motion. More fundamentally, the optimality of such proxy supervision is questionable. An intermediate representation ideal for depth estimation may not be optimal for the final semantic occupancy task, inherently limiting the transformation's expressive power due to this objective misalignment. This motivates our central research question: Can we devise an end-to-end supervision framework1 that holistically optimizes the entire 2D-to-3D transformation, enabling unified semantic-aware error backpropagation and allowing traditionally fixed modules to become fully learnable?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.864 + ], + "angle": 0, + "content": "We approach this problem from a causal perspective. In VisionOcc, the 2D image semantics are the \"cause\" of the final 3D semantic \"effect\". Semantic misalignment arises from disrupted information flow from cause to effect (Fig. 1a). Therefore, instead of correcting the erroneous output, we propose to directly regularize the information flow itself. We posit that a 3D prediction for a given class should be influenced predominantly by 2D image regions of that same class. To enforce this information flow, we leverage gradients as a proxy, inspired by prior work [18, 42, 53]. For each semantic class, the gradient of its aggregated 3D features is computed w.r.t. the 2D feature map, producing a saliency-like map of 2D influence. This map is then directly supervised with the ground truth 2D segmentation mask. As shown in Fig. 1c, this establishes a principled, end-to-end supervision signal for the 2D-to-3D transformation, enabling holistic optimization of all its components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "To fully leverage our end-to-end supervision, we introduce a more expressive and learnable 2D-to-3D view transformation, termed the Semantic Causality-Aware Transformation (SCAT). A key challenge is that directly supervising gradients is inherently unstable. Therefore, the entire SCAT module is designed to constrain its gradient flow to a stable [0, 1] range. Specifically, SCAT introduces three targeted designs: i) Channel-Grouped Lifting: To better disentangle semantics, we move beyond LSS's uniform weighting and apply distinct learnable weights to different groups of feature channels. ii) Learnable Camera Offsets: To mitigate motion-induced pose errors, we introduce learnable offsets to the camera parameters, which are implicitly supervised by the 2D-3D semantic consistency enforced by our causal loss. iii) Normalized Convolution: Finally, we employ a normalized convolution to densify the sparse 3D features from LSS [28], ensuring this final step also adheres to our global gradient stability requirement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.905, + 0.711 + ], + "angle": 0, + "content": "Our contributions are as follows: i) We systematically analyze the 2D-to-3D transformation in VisionOCC, identifying a critical failure mode we term semantic ambiguity. We provide a theoretical analysis proving how the modularity of prior methods leads to error propagation, offering clear guidance for future work. ii) To address these problems, we propose the Causal Loss that directly regularizes the information flow of the 2D-to-3D transformation. This enables true end-to-end optimization of all constituent modules, mitigating error accumulation and making previously fixed components, such as camera parameters, fully learnable. iii) We instantiate our principles in the Semantic Causality-Aware Transformation, a novel 2D-to-3D transformation architecture. SCAT incorporates Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution to explicitly tackle the challenges of semantic confusion, camera perturbations, and limited learnability. iv) Extensive experiments show our method significantly boosts existing models, achieving a \\(3.2\\%\\) absolute mIoU gain on BEVDet. Furthermore, it demonstrates superior robustness to camera perturbations, reducing the relative performance drop on BEVDet from a severe \\(-32.2\\%\\) to a mere \\(-7.3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.655, + 0.745 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.763, + 0.772 + ], + "angle": 0, + "content": "2.1. Semantic Scene Completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Semantic Scene Completion (SSC) [10, 20, 24, 31, 51] refers to the task of simultaneously predicting both the occupancy and semantic labels of a scene. Existing methods can be classified into indoor and outdoor approaches based on the scene type, with the former focusing on occupancy and semantic label prediction in controlled environments [10, 20, 31], while the latter shifts towards more complex outdoor settings, particularly in the context of au" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Here, \"end to end\" refers to a unified supervision scheme for a process, distinct from the concept of a monolithic network architecture." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24879" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.333 + ], + "angle": 0, + "content": "tonomous driving [1, 6]. The core principle of SSC lies in its ability to infer the unseen, effectively bridging gaps in incomplete observations with accurate semantic understanding. MonoScene [6] introduces a 3D SSC framework that infers dense geometry and semantics from a single monocular RGB image. VoxFormer [24] presents a Transformer-based semantic scene completion framework that generates complete 3D volumetric semantics from 2D images by first predicting sparse visible voxel queries and then densifying them through self-attention with a masked autoencoder design. OccFormer [51] introduces a dual-path transformer network for 3D semantic occupancy prediction, efficiently processing camera-generated 3D voxel features through local and global pathways, and enhancing occupancy decoding with preserve-pooling and class-guided sampling to address sparsity and class imbalance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.342, + 0.427, + 0.359 + ], + "angle": 0, + "content": "2.2. Vision-based 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.364, + 0.485, + 0.773 + ], + "angle": 0, + "content": "Vision-based 3D Occupancy Prediction [9, 14-16, 34, 38-40, 43?] aims to predict the spatial and semantic features of 3D voxel grids surrounding an autonomous vehicle from image data. This task is closely related to SSC, emphasizing the importance of multi-perspective joint perception for effective autonomous navigation. TPVFormer [14] is prior work that lifts image features into the 3D TPV space by leveraging an attention mechanism [26, 41]. Different from TPVFormer, which relies on sparse point clouds for supervision, subsequent studies, including OccNet [40], SurroundOcc [44], Occ3D [39], and OpenOccupancy [43], have developed denser occupancy annotations by incorporating temporal information or instance-level labels. Methods such as BEVDet [13], FBOcc [27], COTR [35], and ALOcc [7] leverage depth-based LSS [22, 23, 25, 37] for explicit geometric transformation, demonstrating strong performance. Some methods [2, 4, 17, 36, 46, 49, 52] have explored rendering-based methods that utilize 2D signal supervision, thereby bypassing the need for 3D annotations. Furthermore, recent research like [3, 7, 8, 21, 29, 32, 40] introduced 3D occupancy flow prediction, which addresses the movement of foreground objects in dynamic scenes by embedding 3D flow information to capture per-voxel dynamics. Unlike the above methods, we analyze the 2D-to-3D transformation process from the perspectives of error propagation and semantic causal consistency, proposing a novel approach that enhances causal consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.182, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Preliminary. VisionOcc predicts a dense 3D semantic occupancy grid from surround-view images, modeled as a causal dependency chain as shown in Fig. 2. This process involves key variables: input image \\(\\mathbf{I}\\), estimated geometry \\(\\mathbf{G}\\), camera parameters \\(P\\) (intrinsic & extrinsic), potential camera parameter errors \\(e_{P}\\), intermediate 3D fea" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.9, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Figure 2. The Causal Structure of VisionOcc. It illustrates the dependency chain from the image input \\(\\mathbf{I}\\) to the semantic occupancy output \\(\\mathbf{O}\\). \\(\\mathbf{G}\\): geometry for 2D-to-3D transformation. \\(P\\): camera intrinsic and extrinsic. \\(\\mathbf{L}\\): intermediate 3D feature. \\(e_P\\): errors in camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.481 + ], + "angle": 0, + "content": "ture \\(\\mathbf{L}\\), and final occupancy output \\(\\mathbf{O}\\). In LSS-based methods [7, 13, 35], the pipeline starts with the image backbone extracting features \\(\\mathbf{f}_i = F_i(\\mathbf{I}), \\mathbf{f}_i \\in \\mathbb{R}^{U,V,C}\\); geometry \\(\\mathbf{G} \\in \\mathbb{R}^{U,V,D}\\) is predicted as a probability distribution over discretized depth bins \\(D\\), i.e., \\(\\mathbf{G} = F_g(\\mathbf{f}_i)\\); 2D features are then transformed to 3D via an outer product \\(\\mathbf{f}_L' = \\mathbf{G} \\otimes \\mathbf{f}_i, \\mathbf{f}_L' \\in \\mathbb{R}^{U,V,D,C}\\); camera parameters \\(P\\) map these to voxel coordinates \\(R_P(u,v,d) \\to (h,w,z) \\in [0,H-1] \\times [0,W-1] \\times [0,Z-1]\\), yielding \\(\\mathbf{f}_L \\in \\mathbb{R}^{H \\times W \\times Z}\\), where \\(H \\times W \\times Z\\) defines the occupancy grid resolution; finally, \\(\\mathbf{L}\\) is decoded to produce the semantic occupancy output: \\(\\mathbf{O} = F_o(\\mathbf{f}_L), \\mathbf{O} \\in \\mathbb{R}^{H \\times W \\times Z \\times S}\\), where \\(F_o\\) is the decoding function and \\(S\\) is the number of semantic classes. The prediction \\(\\mathbf{O}\\) is supervised by the ground-truth \\(\\tilde{\\mathbf{O}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.49, + 0.854, + 0.507 + ], + "angle": 0, + "content": "3.1. Error Propagation in Depth-Based LSS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Theorem 1. In Depth-Based LSS methods with a fixed \\(2D\\)-to-3D mapping \\(M_{\\text{fixed}}\\), inherent mapping error \\(\\delta M\\) leads to gradient deviations, preventing convergence to an \\(\\epsilon\\)-optimal solution. This is formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.584, + 0.907, + 0.601 + ], + "angle": 0, + "content": "\\[\nM _ {f i x e d} = M _ {i d e a l} + \\delta M \\Longrightarrow \\nabla_ {\\theta} L _ {L S S} \\neq \\nabla_ {\\theta} L _ {i d e a l}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.643 + ], + "angle": 0, + "content": "where \\( M_{ideal} \\) is the ideal mapping and \\( L_{ideal} \\) is the loss function using \\( M_{ideal} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.805 + ], + "angle": 0, + "content": "Proof. Mapping Error Quantification: Let \\(\\mathbf{x} \\in \\mathbb{R}^2\\) be 2D pixel coordinates and \\(d(\\mathbf{x})\\) be the ground truth depth. Estimated depth is \\(\\hat{d}(\\mathbf{x}) = d(\\mathbf{x}) + \\epsilon_d(\\mathbf{x})\\), with \\(\\epsilon_d(\\mathbf{x})\\) as depth error. Let \\(\\mathbf{K}_{ideal} \\in \\mathbb{R}^{3 \\times 3}\\) be the ideal camera intrinsics, and \\(\\mathbf{K} = \\mathbf{K}_{ideal} + \\epsilon_K\\) be the estimated camera intrinsics with error \\(\\epsilon_K\\). The fixed mapping is \\(M_{fixed} = M_{ideal} + \\delta M\\), where \\(\\delta M\\) encompasses errors from various sources, including depth estimation error \\(\\epsilon_d(\\mathbf{x})\\) and camera extrinsic error \\(\\epsilon_K\\). We assume the total mapping error is bounded \\(\\| \\delta M \\|_F \\leq \\Delta_M < \\infty\\). The 3D coordinates are:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.817, + 0.905, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = M _ {\\text {f i x e d}} (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.837, + 0.905, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {i d e a l}} = M _ {\\text {i d e a l}} (\\mathbf {x}, d (\\mathbf {x}), \\mathbf {K} _ {\\text {i d e a l}}) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.857, + 0.905, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {X} = \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} = \\delta M (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.821, + 0.902 + ], + "angle": 0, + "content": "where \\(\\| \\Delta \\mathbf{X}\\| _2\\leq C\\cdot \\Delta_M\\) for bounded inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24880" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.087, + 0.818, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.223, + 0.908, + 0.266 + ], + "angle": 0, + "content": "Figure 3. The Overall Framework of the Proposed Semantic Causality-Aware VisionOcc. The proposed framework consists of three primary components: a backbone network for extracting 2D features, an SCAT module for transforming these features into 3D space, and an Encoder-Decoder network for learning 3D semantics. The SCAT module is supervised by our causal loss." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.277, + 0.451, + 0.336 + ], + "angle": 0, + "content": "
MethodmIoUmIoUDIoU
Depth-Based LSS44.540.478.9
SCL-Aware LSS50.5↑6.046.9↑6.585.7↑6.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.339, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Table 1. Performance of Depth-Based LSS vs. SCL-Aware LSS on Occ3D (in Ideal Conditions). BEVDetOcc is the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Feature Space Deviation and Loss: Let \\( F_{2D}(\\mathbf{x}) \\) be 2D features and \\( F_{3D}(\\cdot) = \\text{Lift}(F_{2D}(\\mathbf{x}), \\cdot) \\). Assuming Lift is \\( L_{\\text{Lift}} \\)-Lipschitz continuous, the 3D feature deviation is:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.432, + 0.483, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| F _ {3 D} (\\mathbf {X}) - F _ {3 D} \\left(\\mathbf {X} _ {\\text {i d e a l}}\\right) \\right\\| _ {F} \\leq L _ {\\text {L i f t}} \\| \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} \\| _ {2} \\tag {5} \\\\ \\leq L _ {L i f t} \\| \\Delta \\mathbf {X} \\| _ {2} \\leq \\Delta_ {F _ {3 D}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.509 + ], + "angle": 0, + "content": "The loss function is \\(L_{LSS} = \\mathcal{L}(P_{3D}(\\mathbf{X}),GT_{3D})\\), where \\(P_{3D}(\\mathbf{X}) = Seg_{3D}(F_{3D}(\\mathbf{X}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Gradient Deviation and Optimization Limit: The gradient of \\(L_{LSS}\\) w.r.t. parameters \\(\\theta\\) is given by chain rule. However, since \\(M_{fixed}\\) is fixed, \\(\\frac{\\partial\\mathbf{X}}{\\partial\\theta} = 0\\). Thus, the gradient becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.578, + 0.483, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} L _ {L S S} = \\frac {\\partial L _ {L S S}}{\\partial P _ {3 D}} \\frac {\\partial P _ {3 D}}{\\partial F _ {3 D}} \\left(\\frac {\\partial F _ {3 D}}{\\partial F _ {2 D}} \\frac {\\partial F _ {2 D}}{\\partial \\theta}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.652 + ], + "angle": 0, + "content": "Due to \\(\\Delta_{F_{3D}} > 0\\), the computed gradient \\(\\nabla_{\\theta}L_{LSS}\\) is based on the deviated feature space, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.662, + 0.484, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} L _ {L S S} (\\theta) = \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} (\\mathbf {X}), G T _ {3 D}\\right) \\tag {7} \\\\ \\neq \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} \\left(\\mathbf {X} _ {i d e a l}\\right), G T _ {3 D}\\right) = \\nabla_ {\\theta} L _ {i d e a l} (\\theta). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.74 + ], + "angle": 0, + "content": "The gradient deviation prevents mapping's direct optimization and limits convergence to an \\(\\epsilon\\)-optimal solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The theoretical analysis reveals that the inherent error in the fixed 2D-to-3D mapping of Depth-Based LSS methods fundamentally hinders gradient-based optimization from achieving optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.425, + 0.835 + ], + "angle": 0, + "content": "3.2. Semantic Causal Locality in VisionOcc" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As revealed in our theoretical analysis (Sec. 3.1), Depth-Based LSS methods suffer from inherent error propagation due to their fixed 2D-to-3D mapping, which limits optimization efficacy and potential performance. To overcome" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.431 + ], + "angle": 0, + "content": "these limitations, we particularly focus on strengthening the semantic causality of the 2D-to-3D transformation. We argue that VisionOcc's 2D-to-3D semantic occupancy prediction should exhibit semantic causal locality (SCL) for robust perception in autonomous driving. Ideally, 2D causes should drive 3D semantic effects. For instance, a predicted \"car\" at 3D location \\((h,w,z)\\) should originate from a matching 2D image region of \"car\". Per the causal chain, camera parameters \\(P\\) and estimated geometry \\(\\mathbf{G}\\) enable this dependency, with \\(\\mathbf{G}\\) being crucial to maintain SCL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Next, we formulate the ideal SCL condition. For a 2D pixel \\((u,v)\\) with semantic label \\(s\\), the projection probability \\(p_d\\) (the value of \\(\\mathbf{G}\\) at this coordinate and depth \\(d \\in D\\)) should be high if its corresponding 3D ground-truth semantic is \\(s\\), and low otherwise:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.517, + 0.831, + 0.535 + ], + "angle": 0, + "content": "\\[\np _ {d} \\propto \\mathbb {1} (\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) = s),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.621 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\) is the indicator function, and \\(e_P\\) represents the potential coordinate transformation error caused by factors such as camera pose error. In practice, \\(p_d\\) acts as a weight multiplied by 2D features (Eq. (8)), enabling a probabilistic mapping that supports differentiable backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Limitations of Depth-Based LSS in SCL. Depth-Based LSS does not fully account for semantic causal consistency. It only preserves semantic causal locality intuitively under ideal conditions where \\( e_P = 0 \\) and depth estimation is perfectly accurate. However, with coordinate transformation errors \\( e_P \\), even a high \\( p_d \\) for the ideal depth may project 2D semantics to incorrect 3D locations, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.736, + 0.899, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + 0) = s, \\text {b u t} \\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) \\neq s.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.841 + ], + "angle": 0, + "content": "This misalignment causes 2D semantics \\( s \\) (e.g., \"car\") to link with wrong 3D semantics (e.g., \"tree\"), causing semantic ambiguity and hindering training. Moreover, depth-based LSS often propagates semantics to surface points, weakening semantic propagation to occluded regions [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Empirical Validation. We conduct an empirical study to validate our analysis of SCL, comparing two ideal geometric transformations. Using BEVDetOcc [13] as the baseline, we replace its estimated depth-based geometry for LSS" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.212 + ], + "angle": 0, + "content": "with: \\(i\\) ) ground-truth LiDAR depths; \\(ii\\) ) SCL-Aware geometry, which computes \\(p_d\\) from 2D and 3D semantic ground truths, where for \\((u,v,d,s)\\), \\(p_d = 1\\) if \\(\\tilde{\\mathbf{O}}(R_P(u,v,d)) = s\\), else \\(p_d = 0\\). We evaluate their performance in semantic occupancy prediction. Tab. 1 demonstrates that 2D-to-3D Transformation achieves significant performance improvements over depth-based LSS in ideal conditions, validating the benefits of semantic causal locality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.259 + ], + "angle": 0, + "content": "Summary. Building on the above analysis, we propose our solution to enforce semantic causality constraints during training. The overall framework is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.27, + 0.428, + 0.287 + ], + "angle": 0, + "content": "3.3. Semantic Causality-Aware Causal Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.399 + ], + "angle": 0, + "content": "For lifting methods like LSS, we could directly supervise the transformation geometry \\(\\mathbf{G}\\) using the causal semantic geometry derived from ground-truth labels (as described in the previous section). However, we aim to enhance the lifting method in the next section, rendering direct supervision impractical. Thus, we design a gradient-based approach to enforce semantic causality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.461 + ], + "angle": 0, + "content": "We begin within the LSS framework. For a 2D pixel feature at location \\((u, v)\\), LSS multiplies it by the depth-related transformation probability \\(p_d\\) and projects it to the 3D coordinate corresponding to depth \\(d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.473, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {L} \\left(R _ {P} (u, v, d)\\right) = p _ {d} (u, v, d) \\cdot \\mathbf {f} _ {i} (u, v, d). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{f}_L(R_P(u,v,d))\\in \\mathbb{R}^C\\) represents the 3D voxel feature at the projected location, with \\(e_{P}\\) omitted for notational simplicity. \\(\\mathbf{f}_i(u,v,d)\\in \\mathbb{R}^C\\) denotes the 2D image feature at location \\((u,v)\\). We backpropagate gradients from \\(\\mathbf{f}_L\\) to \\(\\mathbf{f}_i\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.578, + 0.483, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(R _ {P} (u , v , d)\\right)}{\\partial \\mathbf {f} _ {i} (u , v , d)} = p _ {d} \\cdot \\mathbf {I}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.618, + 0.291, + 0.632 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}\\) is the all-ones vector." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.484, + 0.695 + ], + "angle": 0, + "content": "For each semantic class \\(s\\), we aggregate the features \\(\\mathbf{f}_L\\) at all 3D positions where the ground truth class equals \\(s\\). This aggregation is backpropagated to the 2D features \\(\\mathbf{f}_i\\), yielding a gradient map \\(\\nabla_s \\in \\mathbb{R}^{U \\times V \\times C}\\) for class \\(s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.706, + 0.483, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {s} (u, v, c) = \\sum_ {\\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\Omega_ {s}} \\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c\\right)}{\\partial \\mathbf {f} _ {i} (u , v , c)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.82 + ], + "angle": 0, + "content": "where \\(\\Omega_s = \\{(h', w', z') \\mid O(h', w', z') = s\\}\\) is the set of 3D positions with semantic label \\(s\\), and \\(c\\) indexes the feature channels. Averaging over channel \\(C\\) produces an attention map \\(A_s \\in \\mathbb{R}^{U \\times V}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.833, + 0.483, + 0.873 + ], + "angle": 0, + "content": "\\[\nA _ {s} (u, v) = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\nabla_ {s} (u, v, c). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Finally, we enforce per-pixel constraints using 2D" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.86, + 0.108 + ], + "angle": 0, + "content": "ground-truth labels with a binary cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.116, + 0.905, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {b c e} ^ {s} = - \\frac {1}{U \\cdot V} \\sum_ {u, v} \\left[ Y _ {s} (u, v) \\log A _ {s} (u, v) \\right. \\tag {12} \\\\ \\left. + \\left(1 - Y _ {s} (u, v)\\right) \\log (1 - A _ {s} (u, v)) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.905, + 0.213 + ], + "angle": 0, + "content": "where \\( Y_{s}(u,v) \\in \\{0,1\\} \\) is the 2D ground-truth label for semantic class \\( s \\) at pixel \\( (u,v) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.906, + 0.334 + ], + "angle": 0, + "content": "The gradient computation can be performed using the automatic differentiation calculator like torch.autograd, requiring \\( S \\) backward passes to iterate over the \\( S \\) semantic classes. This incurs a computational overhead scaling linearly with the class amount \\( S \\). To mitigate this overhead, we reformulate the loss computations in terms of the expectation of an unbiased estimator [11]. We define the expected BCE loss across all semantic classes \\( s \\in \\{1, \\dots, S\\} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ L _ {b c e} ^ {s} \\right] = \\frac {1}{S} \\sum_ {s = 1} ^ {S} L _ {b c e} ^ {s}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Based on this relationship, we uniformly sample a single semantic class \\(s\\) during training:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.434, + 0.905, + 0.451 + ], + "angle": 0, + "content": "\\[\nL _ {c a u s a l} = L _ {b c e} ^ {s}, \\quad s \\sim \\operatorname {U n i f o r m} (1, S). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.46, + 0.906, + 0.537 + ], + "angle": 0, + "content": "This sampling preserves the unbiased nature of the gradient and loss estimates and reduces the computational cost to \\( \\frac{1}{S} \\). \\( L_{\\text{causal}} \\) focuses on enhancing the geometric transformation, serving as an auxiliary term to complement the occupancy loss (case-by-case, e.g., cross-entropy in BEVDet [13])." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.879, + 0.56 + ], + "angle": 0, + "content": "3.4. Semantic Causality-Aware Transformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.688 + ], + "angle": 0, + "content": "We propose semantic causality-aware 2D-to-3D transformation to enhance 2D-to-3D lifting, as shown in Fig. 4. Eq. (14) constrains the geometry \\(\\mathbf{G}\\) using 2D and 3D semantics. This overcomes the rigid, per-location hard alignment of geometric probabilities (e.g., using LiDAR depth supervision) in prior methods [7, 13, 22, 23, 27]. It enables advanced lifting designs, addressing errors from camera pose and other distortions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.741, + 0.71 + ], + "angle": 0, + "content": "3.4.1. Channel-Grouped Lifting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.835 + ], + "angle": 0, + "content": "Vanilla LSS applies uniform weights to all 2D feature channels. We argue this is trivial as 2D and 3D features have distinct locality biases. For instance, a 2D \"car\" edge may capture \"tree\" semantics via convolution, but in 3D, these objects are distant. Uniform weighting both semantics causes ambiguity. Since different channels typically encode distinct semantics, we group the feature channels and learn unique weights for each group:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.845, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {L, g} \\left(R _ {P} (u, v, d)\\right) = \\omega_ {g, d} \\cdot \\mathbf {f} _ {i, g} (u, v, d), g \\in \\{1, \\dots , N _ {g} \\}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.869, + 0.905, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{L,g}\\in \\mathbb{R}^{C / N_g}\\) and \\(\\mathbf{f}_{i,g}\\in \\mathbb{R}^{C / N_g}\\) are the 3D and 2D features for group \\(g\\). \\(\\omega_{g,d}\\) is the learned weight for group" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.482, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.188, + 0.362, + 0.202 + ], + "angle": 0, + "content": "(a) Channel-Grouped Lifting" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.204, + 0.48, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.274, + 0.432, + 0.287 + ], + "angle": 0, + "content": "(b) Coordinate Mapping with Learnable Camera Offsets" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.291, + 0.475, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.349, + 0.416, + 0.361 + ], + "angle": 0, + "content": "(c) Post-Processing with Normalized Convolution" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Figure 4. The Detailed Structure of the Proposed Modules in Semantic Causality-Aware 2D-to-3D Transformation. There are three key components, including (a): Channel-Groupled Lifting, (b): Learnable Camera Offsets, and (c): Normalized Convolution, enabling accurate and robust 2D-to-3D transformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.502 + ], + "angle": 0, + "content": "\\(g\\), replacing \\(p_d\\) which uniformly lifts all channels. \\(N_g\\) is the number of groups. This preserves semantic distinction, ensuring channel-specific causal alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.51, + 0.323, + 0.523 + ], + "angle": 0, + "content": "3.4.2. Learnable Camera Offsets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.635 + ], + "angle": 0, + "content": "To address camera parameter errors, especially pose inaccuracies, we introduce learnable offsets into camera parameters. First, we ensure the lifting process is coordinatedifferentiable. This is crucial for the offsets to receive gradients and adapt during training. The transformation of 2D image coordinates \\((u,v)\\) and depth \\(d\\) to 3D voxel coordinates can be represented as matrix multiplication:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.645, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\n[ h, w, z ] ^ {T} = P \\cdot [ u \\cdot d, v \\cdot d, d, 1 ] ^ {T}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.812 + ], + "angle": 0, + "content": "where \\( P \\in \\mathbb{R}^{3 \\times 4} \\) is the camera projection matrix combining intrinsics and extrinsics. LSS typically rounds floating-point coordinates to integers, rendering them nondifferentiable. Following ALOcc [7], we use the soft filling to enable differentiability w.r.t. position. This method calculates distances between floating-point 3D coordinates and their eight surrounding integer coordinates. These distances serve as weights to distribute a 2D feature at \\( (u, v, d) \\) across multiple 3D locations. Lifting can be rewritten to" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.823, + 0.483, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {f} _ {L, g} \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) = \\omega_ {g, d} \\cdot \\omega_ {h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}} \\cdot \\mathbf {f} _ {i, g} (u, v, d), \\tag {17} \\\\ \\forall \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\text {n e i g h b o r s}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.421, + 0.885 + ], + "angle": 0, + "content": "where \\(\\omega_{h',w',z'}\\) are trilinear interpolation weights." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Next, we propose learning two offsets. First, we directly" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.85, + 0.107 + ], + "angle": 0, + "content": "predict an offset applied to the camera parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.115, + 0.905, + 0.133 + ], + "angle": 0, + "content": "\\[\nP := P + \\Delta P, \\quad \\Delta P = F _ {\\text {o f f s e t 1}} \\left(\\mathbf {f} _ {i}, P\\right), \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.14, + 0.905, + 0.187 + ], + "angle": 0, + "content": "where \\(\\Delta P\\) is the predicted parameter offset, and \\(F_{offset1}\\) denotes the network. Second, we estimate per-position offsets for each \\((u,v,d)\\) in the image coordinate system:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.195, + 0.905, + 0.218 + ], + "angle": 0, + "content": "\\[\n(u, v, d) := (u + \\Delta u, v + \\Delta v, d + \\Delta d), \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.213, + 0.844, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\left(\\Delta u, \\Delta v, \\Delta d\\right) = F _ {\\text {o f f s e t} 2} \\left(\\mathbf {f} _ {i} (u, v, d)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.907, + 0.299 + ], + "angle": 0, + "content": "where \\( F_{offset2} \\) is another network. These offsets enable the model to adaptively compensate for camera parameter errors \\( e_P \\), improving geometric accuracy while preserving semantic causal locality under such errors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.731, + 0.318 + ], + "angle": 0, + "content": "3.4.3. Normalized Convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Prior work [28] notes that the direct mapping in LSS yields sparse 3D features. To address this, an intuitive solution is to use local feature propagation operators (e.g., convolutions) with causal loss supervision for preserving semantic causality. However, vanilla convolutions lack gradient constraints during causal loss computation, producing poor results. We address this by normalizing convolution weights to keep gradient map values in [0, 1]. Given the difficulty of normalizing standard convolution weights, we follow MobileNet and ConvNext to adopt depthwise (spatial) and pointwise (channel) decomposition. For the depthwise kernel \\( W_{\\mathrm{spatial}} \\in \\mathbb{R}^{3 \\times 3 \\times 3 \\times C} \\), we apply softmax across spatial dimensions \\( (h, w, z) \\) per channel \\( c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.527, + 0.905, + 0.556 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {s p a t i a l}} ^ {\\prime} [ h, w, z, c ] = \\frac {\\exp \\left(W _ {\\text {s p a t i a l}} [ h , w , z , c ]\\right)}{\\sum_ {h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime}} \\exp \\left(W _ {\\text {s p a t i a l}} \\left[ h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c \\right]\\right)}. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.564, + 0.905, + 0.595 + ], + "angle": 0, + "content": "For the pointwise kernel \\( W_{\\mathrm{channel}} \\in \\mathbb{R}^{C \\times C} \\), we apply softmax across input channels per output channel:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.601, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {c h a n n e l}} ^ {\\prime} \\left[ c _ {\\text {i n}}, c _ {\\text {o u t}} \\right] = \\frac {\\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} \\right]\\right)}{\\sum_ {c _ {\\text {o u t}} ^ {\\prime}} \\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} ^ {\\prime} \\right]\\right)}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Specifically, we use transposed convolution for depthwise convolutions. It diffuses features from non-zero to zero positions, but the zero position does not affect others. We prove in the supplement that the derived gradient mask remains within [0, 1] even with our semantic causality-aware 2D-to-3D transformation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.857, + 0.76 + ], + "angle": 0, + "content": "3.5. Validation of Gradient Error Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 shows the occupancy loss curves during training for BEVDet, comparing performance with and without our method. The results show that BEVDet integrated with our approach (blue line) achieves a significantly faster and steeper loss reduction compared to the original BEVDet (red line). This empirical evidence validates our theoretical analysis of gradient error mitigation. As Theorem Theorem 1 formalizes, Depth-Based LSS methods are inherently limited by gradient deviations due to fixed 2D-to-3D" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24883" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.089, + 0.429, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.328 + ], + "angle": 0, + "content": "Figure 5. Occupancy Loss Curves with/without Our Method. Our method reduces training loss through semantic causal 2D-to-3D geometry transformation, as shown in the comparison before (red) and after (blue) its application." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.335, + 0.486, + 0.468 + ], + "angle": 0, + "content": "
MethodmIoU ↑DropmIoUD↑DropIoU ↑Drop
BEVDetOcc [12]37.1-32.3%30.2-49.0%70.4-8.5%
BEVDetOcc w/ Noise25.115.464.4
BEVDetOcc+Ours38.3-7.3%31.5-10.8%71.2-1.4%
BEVDetOcc+Ours w/ Noise35.528.170.2
ALoCC [7]40.1-21.9%34.3-28.6%70.2-8.4%
ALoCC w/ Noise31.324.564.3
ALoCC+Ours40.9-3.3%35.5-4.8%70.7-1.0%
ALoCC+Ours w/ Noise39.633.870.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.545 + ], + "angle": 0, + "content": "Table 2. Performance Comparison on the Occ3D Dataset with Gaussian Noise Added to Camera Parameters. The \"Drop (%) columns show degradation, with our methods (BEVDetOcc+Ours, ALOcc+Ours) achieving much smaller drops (e.g., -7.3% mIoU vs. -32.4% for BEVDetOcc)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.645 + ], + "angle": 0, + "content": "mapping errors. In contrast, our method aims to alleviate this by enabling a learnable mapping and incorporating a causal loss. It indicates that by mitigating gradient error through semantic causality-aware 2D-to-3D transformation, our approach facilitates more efficient gradient-based learning, leading to faster convergence and a lower loss curve." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.218, + 0.675 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.282, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset. In this study, we leverage the Occ3D-nuScenes dataset [5, 39], a comprehensive dataset with diverse scenes for autonomous driving research. It encompasses 700 scenes for training, 150 for validation, and 150 for testing. Each scene integrates a 32-beam LiDAR point cloud alongside 6 RGB images, captured from multiple perspectives encircling the ego vehicle. Occ3D [39] introduces voxel-based annotations, covering a spatial extent of \\(-40\\mathrm{m}\\) to \\(40\\mathrm{m}\\) along the X and Y axes, and \\(-1\\mathrm{m}\\) to \\(5.4\\mathrm{m}\\) along the Z axis, with a consistent voxel size of \\(0.4\\mathrm{m}\\) across all dimensions. Occ3D delineates 18 semantic categories, comprising 17 distinct object classes plus an empty class to signify unoccupied regions. Following [7, 27, 35, 39], we evalu" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.163, + 0.904, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.235, + 0.707, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.308, + 0.648, + 0.32 + ], + "angle": 0, + "content": "(a) BEVDet" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.236, + 0.904, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.308, + 0.864, + 0.32 + ], + "angle": 0, + "content": "(b) BEVDet + Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Figure 6. Visualization of 2D-to-3D Semantic Causal Consistency Using LayerCAM [18]. We enhance BEVDet with our method for comparison. Attention maps are computed for critical traffic classes: \"traffic cone\", \"pedestrian\", and \"car\". Areas of greatest difference are marked with boxes. Each class-specific localization highlights our method's precise focus over vanilla BEVDet, showing improved semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.478 + ], + "angle": 0, + "content": "ate the occupancy prediction performance with mIoU across 17 semantic object categories, \\(\\mathrm{mIoU}_D\\) for 8 dynamic categories, and occupied/unoccupied IoU for scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Implementation Details. We integrate our approach into BEVDet [12] and ALOcc [7] for performance evaluation. The model parameters, image, and BEV augmentation strategies are retrained as the original. For ALOcc, we remove its ground-truth depth denoising module to adapt to our method. We use multiple convolutional layers to predict the two camera parameter offsets. The loss weight of \\( L_{\\text{causal}} \\) is set to 0.02 in the main experiments. We optimize using AdamW [33] with a learning rate of \\( 2 \\times 10^{-4} \\), a global batch size of 16, and for 24 epochs. Experiments use single-frame surrounding images, emphasizing improvements from enhanced 2D-to-3D transformations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.905, + 0.683 + ], + "angle": 0, + "content": "4.2. Evaluation of Camera Perturbation Robustness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.825 + ], + "angle": 0, + "content": "To assess our method's robustness under extreme noise, we add Gaussian noise with (0.1 variances) to camera parameters in training and testing, as shown in Tab. 2. Compared to vanilla BEVDetOcc and ALOcc, our methods show smaller performance drops. BEVDetOcc+Ours has a mIoU drop of \\(-7.3\\%\\) vs. \\(-32.3\\%\\) for BEVDetOcc, and ALOcc+Ours drops \\(-3.3\\%\\) vs. \\(-21.9\\%\\) for ALOcc, demonstrating enhanced resilience to noisy parameters. This effectively counters motion-induced errors, benefiting self-driving tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.805, + 0.85 + ], + "angle": 0, + "content": "4.3. Semantic Causality Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in the Fig. 6, we visualize 3D-to-2D semantic causal consistency using LayerCAM [18]. We backpropagate the final 3D semantic occupancy predictions of dis" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24884" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.272 + ], + "angle": 0, + "content": "
MethodBackboneInput SizemIoUmIoUDIoU
MonoScene [6]ResNet-101928 × 16006.15.4-
CTF-Occ [39]ResNet-101928 × 160028.527.4-
TPVFormer [14]ResNet-101928 × 160027.827.2-
COTR [35]ResNet-50256 × 70439.133.869.6
ProtoOcc [19]ResNet-50256 × 70439.634.3-
LightOcc-S [50]ResNet-50256 × 70437.932.4-
DHD-S [45]ResNet-50256 × 70436.530.7-
FlashOCC [48]ResNet-50256 × 70432.024.765.3
FB-Occ [27]ResNet-50256 × 70435.730.966.5
BEVDetOcc [12]ResNet-50256 × 70437.130.270.4
BEVDetOcc+OursResNet-50256 × 70438.3 ↑1.231.5 ↑1.371.2 ↑1.2
ALOcc [7]ResNet-50256 × 70440.134.370.2
ALOcc+OursResNet-50256 × 70440.9 ↑0.835.5 ↑1.170.7 ↑0.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.354 + ], + "angle": 0, + "content": "Table 3. Comparison of 3D Semantic Occupancy Prediction Using Single Frame on the Occ3D Dataset, Evaluated mIoU, mIoUD, and IoU Metrics. Performance gains are indicated by red arrows \\(\\uparrow\\). Our proposed approach (+Ours) consistently demonstrates superior enhancement over existing methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.472 + ], + "angle": 0, + "content": "tinct classes to the 2D feature maps fed into SCAT. Notably, we are the first to apply LayerCAM for cross-dimensional analysis. Fig. 6 (b) shows our method precisely focuses on class-specific locations. This confirms LayerCAM's cross-dimensional effectiveness. Our approach surpasses vanilla BEVDet (Fig. 6 (a)) in targeting class-associated objects. This proves improved semantic localization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.422, + 0.499 + ], + "angle": 0, + "content": "4.4. Benchmarking with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.673 + ], + "angle": 0, + "content": "As shown in Tab. 3, we compare our method with leading 3D semantic occupancy prediction approaches on Occ3D [39]. Specifically, compared to baseline models BEVDet and ALOcc, our method achieves significant improvements in mIoU, mIoU\\(_{\\mathrm{D}}\\), and IoU. For instance, the BEVDetOcc+Ours variant achieves an mIoU of 38.3, surpassing BEVDetOcc [12] by 1.2, while improving mIoU\\(_{\\mathrm{D}}\\) by 1.3 and IoU by 1.2. Similarly, ALOcc+Ours shows gains of 0.8 in mIoU, 1.1 in mIoU\\(_{\\mathrm{D}}\\), and 0.5 in IoU over ALOcc [7]. These results validate the superiority of our semantic causality-aware 2D-to-3D transformation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.245, + 0.699 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of Causal Loss. We first investigate the effectiveness of the proposed Causal Loss in enhancing the occupancy prediction performance. As demonstrated in Tab. 4, the impact of the proposed Causal Loss is evaluated through a series of experiments. Specifically, using BEVDetOcc as the baseline (Exp. 0), we conducted two ablation studies: one removing the depth supervision loss (Exp. 1), and another incorporating the proposed Causal Loss (Exps. 2, 3). The results reveal that the removal of the depth supervision loss leads to marginal decrease in performance, whereas the addition of the proposed Causal Loss yields a significant improvement. This suggests that Causal Loss facilitates superior 2D-to-3D transformation and ultimately enhances the" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.252 + ], + "angle": 0, + "content": "
Exp.MethodmIoUDiff.mIoUDIoULatency
0Baseline (BEVDetOcc) [12]37.1-30.270.4416/125
1w/o Depth Sup36.8-0.329.670.3414/125
2+ Causal Loss37.6+0.831.070.1450/125
3+ Unbiased Estimator37.5-0.130.770.5417/125
4w/o Post Conv37.3-0.230.770.2379/122
5+ Channel-Grouped Lifting37.6+0.330.770.7419/128
6+ Soft Filling37.6-30.670.7434/149
7+ Learnable Camera Offset37.9+0.331.171.0446/150
8+ Normalized Convolution38.3+0.431.571.2466/159
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.26, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 4. Ablation Study of 3D Semantic Occupancy Prediction on Occ3D. We comprehensively evaluated the impact of the individual strategy (bolded rows) proposed in our paper, with BEVDetOcc as the baseline. The final column reports single-frame training/inference latency (ms) on an RTX 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.385 + ], + "angle": 0, + "content": "precision of 3D semantic occupancy prediction. Comparing Exp. 3 to Exp. 2, the Unbiased Estimator simplifies the Causal Loss computation, reducing training overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Effect of Each Module. In Tab. 4 (Exps. 5-8), we systematically validate the effectiveness of the three proposed modules. We first remove the two post-lifting convolutional layers from the original BEVDet (Exp. 4), as they serve a similar role to our proposed modules in refining volume features. Subsequently, we incrementally integrate the three proposed modules (Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution) into the baseline model. To ensure gradient flow for Learnable Camera Offsets, we introduce the Soft Filling strategy from ALOcc [7] in Exp. 6, enabling effective training of the camera parameter offsets in Exp. 7. The results show progressive performance improvements with each added component (Exps. 5, 7, 8), confirming the efficacy of the proposed SCAT method. Additionally, the proposed modules incur acceptable computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Please refer to the supplementary material for more comprehensive experimental results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.634, + 0.683 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced a novel approach leveraging causal principles to address the limitation that ignored the reliability and interpretability by existing methods. By exploring the causal foundations of 3D semantic occupancy prediction, we propose a causal loss that enhances semantic causal consistency. In addition, we develop the SCAT module with three main components: Channel-Groupled Lifting, Learnable Camera Offsets and Normalized Convolution. This approach effectively mitigates transformation inaccuracies arising from uniform mapping weights, camera perturbations, and sparse mappings. Experiments demonstrate that our approach achieves significant improvements in accuracy, robustness to camera perturbations, and semantic causal consistency in 2D-to-3D transformations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24885" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[2] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Langocc: Self-supervised open vocabulary occupancy estimation via volume rendering. arXiv preprint arXiv:2407.17310, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Occfownet: Towards self-supervised occupancy estimation via differentiable rendering and occupancy flow. arXiv preprint arXiv:2402.12792, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Gaussianflowocc: Sparse and weakly supervised occupancy estimation using gaussian splatting and temporal flow. arXiv preprint arXiv:2502.17288, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Dubing Chen, Jin Fang, Wencheng Han, Xinjing Cheng, Junbo Yin, Chengzhong Xu, Fahad Shahbaz Khan, and Jianbing Shen. Alocc: adaptive lifting-based 3d semantic occupancy and cost volume-based flow prediction. arXiv preprint arXiv:2411.07725, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Dubing Chen, Wencheng Han, Jin Fang, and Jianbing Shen. Adaocc: Adaptive forward view transformation and flow modeling for 3d occupancy and flow prediction. arXiv preprint arXiv:2407.01436, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Dubing Chen, Huan Zheng, Jin Fang, Xingping Dong, Xianfei Li, Wenlong Liao, Tao He, Pai Peng, and Jianbing Shen. Rethinking temporal fusion with a unified gradient descent view for 3d semantic occupancy prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1505-1515, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4193-4202, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[11] Judy Hoffman, Daniel A Roberts, and Sho Yaida. Robust learning with jacobian regularization. arXiv preprint arXiv:1908.02729, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376-393. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[17] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding. arXiv preprint arXiv:2412.13193, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[18] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[19] Jungho Kim, Changwon Kang, Dongyoung Lee, Sehwan Choi, and Jun Won Choi. Protoocc: Accurate, efficient 3d occupancy prediction using dual branch encoder-prototype query decoder. arXiv preprint arXiv:2412.08774, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[22] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[24] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yangguang Li, Bin Huang, Zeren Chen, Yufeng Cui, Feng Liang, Mingzhu Shen, Fenggang Liu, Enze Xie, Lu Sheng, Wanli Ouyang, et al. Fast-bev: A fast and strong bird's" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24886" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "eye view perception baseline. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[27] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[28] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[29] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR 2024 Autonomous Grand Challenge Track On Occupancy and Flow, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.482, + 0.452 + ], + "angle": 0, + "content": "[30] Lizhe Liu, Bohua Wang, Hongwei Xie, Daqi Liu, Li Liu, Zhiqiang Tian, Kuiyuan Yang, and Bing Wang. Surroundsdf: Implicit 3d scene understanding based on signed distance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[31] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[32] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[34] Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[35] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. arXiv preprint arXiv:2312.01919, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[36] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12404-12411. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as set of points. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[39] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[42] Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[43] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.53 + ], + "angle": 0, + "content": "[44] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[45] Yuan Wu, Zhiqiang Yan, Zhengxue Wang, Xiang Li, Le Hui, and Jian Yang. Deep height decoupling for precise vision-based 3d occupancy prediction. arXiv preprint arXiv:2409.07972, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[46] Ziyang Yan, Wenzhen Dong, Yihua Shao, Yuhang Lu, Liu Haiyang, Jingwen Liu, Haozhe Wang, Zhe Wang, Yan Wang, Fabio Remondino, et al. Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[47] Shichao Yang, Yulan Huang, and Sebastian Scherer. Semantic 3d occupancy mapping through efficient high order crfs. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 590-597. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[48] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[49] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[50] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Lightweight spatial embedding for vision-based 3d occupancy prediction. arXiv preprint arXiv:2412.05976, 2024.8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24887" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[51] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[52] Jilai Zheng, Pin Tang, Zhongdao Wang, Guoqing Wang, Xiangxuan Ren, Bailan Feng, and Chao Ma. Veon: Vocabulary-enhanced occupancy prediction. In European Conference on Computer Vision, pages 92-108. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[53] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 15890-15902, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "24888" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_origin.pdf b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b5f321464b714f41e1b9afa1bc8af50a20885d4c --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/218a3708-3415-4c53-9a50-ef5f14ed0f9a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be183a2ca17b9b55b3acd3cfa8db65e6feda3fa168bcaa740ddc691c34dc1070 +size 1213523 diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/full.md b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7930f83141d4bdbdafe01ff01a6e137889a6ecec --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/full.md @@ -0,0 +1,405 @@ +# Semantic Causality-Aware Vision-Based 3D Occupancy Prediction + +Dubing Chen $^{1}$ , Huan Zheng $^{1}$ , Yucheng Zhou $^{1}$ , Xianfei Li $^{2}$ , Wenlong Liao $^{2}$ , Tao He $^{2}$ , Pai Peng $^{2}$ , Jianbing Shen $^{1\boxtimes}$ $^{1}$ SKL-IOTSC, CIS, University of Macau $^{2}$ COWAROBOT Co. Ltd. https://github.com/cdb342/CausalOcc + +# Abstract + +Vision-based 3D semantic occupancy prediction is a critical task in 3D vision that integrates volumetric 3D reconstruction with semantic understanding. Existing methods, however, often rely on modular pipelines. These modules are typically optimized independently or use pre-configured inputs, leading to cascading errors. In this paper, we address this limitation by designing a novel causal loss that enables holistic, end-to-end supervision of the modular 2D-to-3D transformation pipeline. Grounded in the principle of 2D-to-3D semantic causality, this loss regulates the gradient flow from 3D voxel representations back to the 2D features. Consequently, it renders the entire pipeline differentiable, unifying the learning process and making previously non-trainable components fully learnable. Building on this principle, we propose the Semantic Causality-Aware 2D-to-3D Transformation, which comprises three components guided by our causal loss: Channel-Groupled Lifting for adaptive semantic mapping, Learnable Camera Offsets for enhanced robustness against camera perturbations, and Normalized Convolution for effective feature propagation. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the Occ3D benchmark, demonstrating significant robustness to camera perturbations and improved 2D-to-3D semantic consistency. + +# 1. Introduction + +Predicting dense 3D semantic occupancy is a fundamental task in 3D vision, providing a fine-grained voxel representation of scene geometry and semantics [39, 40, 43, 47]. The challenge of performing this prediction from vision alone, + +![](images/688f0050e8c1b14bbd9adc04a36708ea013cc74e6c54793c41befb5f3f624927.jpg) +(a) An Example of Inaccurate 2D-to-3D Transformation + +![](images/716ae459e91b2d010789cd4c87e05ea9d2b792c95a6dc3f8e68044f365e20358.jpg) +(b) Modular 2D-to-3D Transformation + +![](images/6bc752e1a1159bbd1d1122c17f2aedd2d4b26b23bcdbc93afbee95d150546f4f.jpg) +(c) Our End-to-End Supervised 2D-to-3D Transformation +Figure 1. (a) Illustrates a Visual Analysis of Semantic Ambiguity in VisionOcc. Inaccurate 2D-to-3D transformations may lead to positional shifts, misaligning supervision signals and resulting in semantic ambiguity. (b) Depicts the Conventional Modular 2D-to-3D Transformation Paradigm [7, 13, 27, 35], which employs depth supervision for geometry estimation, pre-calibrated camera parameters, and fixed mapping for lifting. (c) Presents Our Holistic, End-to-End Supervised 2D-to-3D Transformation Paradigm, which eliminates the need for separate modular supervision or pre-calibration, enabling unified error propagation to supervise all components. + +an approach known as vision-based 3D semantic occupancy prediction (VisionOcc), has recently become a focal point of research. By leveraging only commodity cameras, VisionOcc is pivotal for a wide range of 3D applications, serving as a comprehensive digital replica of the environment for tasks like analysis, simulation, and interactive visualization [7, 9, 14, 27, 39, 40, 44]. + +VisionOcc unifies the challenges of feed-forward 3D reconstruction and dense semantic understanding. Existing pipelines typically decompose this task into two metaphases [7, 14, 27, 30, 35]. The initial 2D-to-3D transfor + +mation uses large receptive field operators like view lifting [7, 37] or cross attention [14, 27] to construct an initial 3D feature volume. Subsequently, a 3D representation learning phase employs operators with a small receptive field (e.g., 3D convolutions or local self-attention) to refine 3D features and produce the final prediction. Our work targets the 2D-to-3D transformation phase, which is more critical and error-prone. Fig. 1a showcases a primary failure mode where features of one class (e.g., a 2D 'car') are erroneously transformed to the 3D location of another (e.g., a 'tree'). This creates a flawed learning objective, forcing the model to learn a spurious association between 'car' features and a 'tree' label. Such semantic ambiguity is a principal obstacle to achieving high performance. + +The prevailing VisionOcc methods, typically based on Lift-Splat-Shoot (LSS), employ a modular approach for the 2D-to-3D transformation (Fig. 1b) [7, 35, 37]. This involves supervising geometry with a proxy depth loss while relying on fixed, pre-calibrated camera parameters and a static lifting map. However, this modularity raises critical questions about robustness and optimality. First, it is susceptible to compounding errors; for example, the reliance on fixed camera parameters makes the system vulnerable to real-world perturbations like camera jitter during motion. More fundamentally, the optimality of such proxy supervision is questionable. An intermediate representation ideal for depth estimation may not be optimal for the final semantic occupancy task, inherently limiting the transformation's expressive power due to this objective misalignment. This motivates our central research question: Can we devise an end-to-end supervision framework1 that holistically optimizes the entire 2D-to-3D transformation, enabling unified semantic-aware error backpropagation and allowing traditionally fixed modules to become fully learnable? + +We approach this problem from a causal perspective. In VisionOcc, the 2D image semantics are the "cause" of the final 3D semantic "effect". Semantic misalignment arises from disrupted information flow from cause to effect (Fig. 1a). Therefore, instead of correcting the erroneous output, we propose to directly regularize the information flow itself. We posit that a 3D prediction for a given class should be influenced predominantly by 2D image regions of that same class. To enforce this information flow, we leverage gradients as a proxy, inspired by prior work [18, 42, 53]. For each semantic class, the gradient of its aggregated 3D features is computed w.r.t. the 2D feature map, producing a saliency-like map of 2D influence. This map is then directly supervised with the ground truth 2D segmentation mask. As shown in Fig. 1c, this establishes a principled, end-to-end supervision signal for the 2D-to-3D transformation, enabling holistic optimization of all its components. + +To fully leverage our end-to-end supervision, we introduce a more expressive and learnable 2D-to-3D view transformation, termed the Semantic Causality-Aware Transformation (SCAT). A key challenge is that directly supervising gradients is inherently unstable. Therefore, the entire SCAT module is designed to constrain its gradient flow to a stable [0, 1] range. Specifically, SCAT introduces three targeted designs: i) Channel-Grouped Lifting: To better disentangle semantics, we move beyond LSS's uniform weighting and apply distinct learnable weights to different groups of feature channels. ii) Learnable Camera Offsets: To mitigate motion-induced pose errors, we introduce learnable offsets to the camera parameters, which are implicitly supervised by the 2D-3D semantic consistency enforced by our causal loss. iii) Normalized Convolution: Finally, we employ a normalized convolution to densify the sparse 3D features from LSS [28], ensuring this final step also adheres to our global gradient stability requirement. + +Our contributions are as follows: i) We systematically analyze the 2D-to-3D transformation in VisionOCC, identifying a critical failure mode we term semantic ambiguity. We provide a theoretical analysis proving how the modularity of prior methods leads to error propagation, offering clear guidance for future work. ii) To address these problems, we propose the Causal Loss that directly regularizes the information flow of the 2D-to-3D transformation. This enables true end-to-end optimization of all constituent modules, mitigating error accumulation and making previously fixed components, such as camera parameters, fully learnable. iii) We instantiate our principles in the Semantic Causality-Aware Transformation, a novel 2D-to-3D transformation architecture. SCAT incorporates Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution to explicitly tackle the challenges of semantic confusion, camera perturbations, and limited learnability. iv) Extensive experiments show our method significantly boosts existing models, achieving a $3.2\%$ absolute mIoU gain on BEVDet. Furthermore, it demonstrates superior robustness to camera perturbations, reducing the relative performance drop on BEVDet from a severe $-32.2\%$ to a mere $-7.3\%$ . + +# 2. Related Work + +# 2.1. Semantic Scene Completion + +Semantic Scene Completion (SSC) [10, 20, 24, 31, 51] refers to the task of simultaneously predicting both the occupancy and semantic labels of a scene. Existing methods can be classified into indoor and outdoor approaches based on the scene type, with the former focusing on occupancy and semantic label prediction in controlled environments [10, 20, 31], while the latter shifts towards more complex outdoor settings, particularly in the context of au + +tonomous driving [1, 6]. The core principle of SSC lies in its ability to infer the unseen, effectively bridging gaps in incomplete observations with accurate semantic understanding. MonoScene [6] introduces a 3D SSC framework that infers dense geometry and semantics from a single monocular RGB image. VoxFormer [24] presents a Transformer-based semantic scene completion framework that generates complete 3D volumetric semantics from 2D images by first predicting sparse visible voxel queries and then densifying them through self-attention with a masked autoencoder design. OccFormer [51] introduces a dual-path transformer network for 3D semantic occupancy prediction, efficiently processing camera-generated 3D voxel features through local and global pathways, and enhancing occupancy decoding with preserve-pooling and class-guided sampling to address sparsity and class imbalance. + +# 2.2. Vision-based 3D Occupancy Prediction + +Vision-based 3D Occupancy Prediction [9, 14-16, 34, 38-40, 43?] aims to predict the spatial and semantic features of 3D voxel grids surrounding an autonomous vehicle from image data. This task is closely related to SSC, emphasizing the importance of multi-perspective joint perception for effective autonomous navigation. TPVFormer [14] is prior work that lifts image features into the 3D TPV space by leveraging an attention mechanism [26, 41]. Different from TPVFormer, which relies on sparse point clouds for supervision, subsequent studies, including OccNet [40], SurroundOcc [44], Occ3D [39], and OpenOccupancy [43], have developed denser occupancy annotations by incorporating temporal information or instance-level labels. Methods such as BEVDet [13], FBOcc [27], COTR [35], and ALOcc [7] leverage depth-based LSS [22, 23, 25, 37] for explicit geometric transformation, demonstrating strong performance. Some methods [2, 4, 17, 36, 46, 49, 52] have explored rendering-based methods that utilize 2D signal supervision, thereby bypassing the need for 3D annotations. Furthermore, recent research like [3, 7, 8, 21, 29, 32, 40] introduced 3D occupancy flow prediction, which addresses the movement of foreground objects in dynamic scenes by embedding 3D flow information to capture per-voxel dynamics. Unlike the above methods, we analyze the 2D-to-3D transformation process from the perspectives of error propagation and semantic causal consistency, proposing a novel approach that enhances causal consistency. + +# 3. Method + +Preliminary. VisionOcc predicts a dense 3D semantic occupancy grid from surround-view images, modeled as a causal dependency chain as shown in Fig. 2. This process involves key variables: input image $\mathbf{I}$ , estimated geometry $\mathbf{G}$ , camera parameters $P$ (intrinsic & extrinsic), potential camera parameter errors $e_{P}$ , intermediate 3D fea + +![](images/e6dc0ed532564f3ef101e617c132dcadb62b9d4e31e01db2e726f7aeeb832c15.jpg) +Figure 2. The Causal Structure of VisionOcc. It illustrates the dependency chain from the image input $\mathbf{I}$ to the semantic occupancy output $\mathbf{O}$ . $\mathbf{G}$ : geometry for 2D-to-3D transformation. $P$ : camera intrinsic and extrinsic. $\mathbf{L}$ : intermediate 3D feature. $e_P$ : errors in camera parameters. + +ture $\mathbf{L}$ , and final occupancy output $\mathbf{O}$ . In LSS-based methods [7, 13, 35], the pipeline starts with the image backbone extracting features $\mathbf{f}_i = F_i(\mathbf{I}), \mathbf{f}_i \in \mathbb{R}^{U,V,C}$ ; geometry $\mathbf{G} \in \mathbb{R}^{U,V,D}$ is predicted as a probability distribution over discretized depth bins $D$ , i.e., $\mathbf{G} = F_g(\mathbf{f}_i)$ ; 2D features are then transformed to 3D via an outer product $\mathbf{f}_L' = \mathbf{G} \otimes \mathbf{f}_i, \mathbf{f}_L' \in \mathbb{R}^{U,V,D,C}$ ; camera parameters $P$ map these to voxel coordinates $R_P(u,v,d) \to (h,w,z) \in [0,H-1] \times [0,W-1] \times [0,Z-1]$ , yielding $\mathbf{f}_L \in \mathbb{R}^{H \times W \times Z}$ , where $H \times W \times Z$ defines the occupancy grid resolution; finally, $\mathbf{L}$ is decoded to produce the semantic occupancy output: $\mathbf{O} = F_o(\mathbf{f}_L), \mathbf{O} \in \mathbb{R}^{H \times W \times Z \times S}$ , where $F_o$ is the decoding function and $S$ is the number of semantic classes. The prediction $\mathbf{O}$ is supervised by the ground-truth $\tilde{\mathbf{O}}$ . + +# 3.1. Error Propagation in Depth-Based LSS + +Theorem 1. In Depth-Based LSS methods with a fixed $2D$ -to-3D mapping $M_{\text{fixed}}$ , inherent mapping error $\delta M$ leads to gradient deviations, preventing convergence to an $\epsilon$ -optimal solution. This is formalized as: + +$$ +M _ {f i x e d} = M _ {i d e a l} + \delta M \Longrightarrow \nabla_ {\theta} L _ {L S S} \neq \nabla_ {\theta} L _ {i d e a l}, \tag {1} +$$ + +where $M_{ideal}$ is the ideal mapping and $L_{ideal}$ is the loss function using $M_{ideal}$ . + +Proof. Mapping Error Quantification: Let $\mathbf{x} \in \mathbb{R}^2$ be 2D pixel coordinates and $d(\mathbf{x})$ be the ground truth depth. Estimated depth is $\hat{d}(\mathbf{x}) = d(\mathbf{x}) + \epsilon_d(\mathbf{x})$ , with $\epsilon_d(\mathbf{x})$ as depth error. Let $\mathbf{K}_{ideal} \in \mathbb{R}^{3 \times 3}$ be the ideal camera intrinsics, and $\mathbf{K} = \mathbf{K}_{ideal} + \epsilon_K$ be the estimated camera intrinsics with error $\epsilon_K$ . The fixed mapping is $M_{fixed} = M_{ideal} + \delta M$ , where $\delta M$ encompasses errors from various sources, including depth estimation error $\epsilon_d(\mathbf{x})$ and camera extrinsic error $\epsilon_K$ . We assume the total mapping error is bounded $\| \delta M \|_F \leq \Delta_M < \infty$ . The 3D coordinates are: + +$$ +\mathbf {X} = M _ {\text {f i x e d}} (\mathbf {x}, \hat {d} (\mathbf {x}), \mathbf {K}) \tag {2} +$$ + +$$ +\mathbf {X} _ {\text {i d e a l}} = M _ {\text {i d e a l}} (\mathbf {x}, d (\mathbf {x}), \mathbf {K} _ {\text {i d e a l}}) \tag {3} +$$ + +$$ +\Delta \mathbf {X} = \mathbf {X} - \mathbf {X} _ {\text {i d e a l}} = \delta M (\mathbf {x}, \hat {d} (\mathbf {x}), \mathbf {K}), \tag {4} +$$ + +where $\| \Delta \mathbf{X}\| _2\leq C\cdot \Delta_M$ for bounded inputs. + +![](images/653d4f3314c43e2cb0d40b3486023ef5483b3b62d55a030e67bddeed13ad7dbf.jpg) +Figure 3. The Overall Framework of the Proposed Semantic Causality-Aware VisionOcc. The proposed framework consists of three primary components: a backbone network for extracting 2D features, an SCAT module for transforming these features into 3D space, and an Encoder-Decoder network for learning 3D semantics. The SCAT module is supervised by our causal loss. + +
MethodmIoUmIoUDIoU
Depth-Based LSS44.540.478.9
SCL-Aware LSS50.5↑6.046.9↑6.585.7↑6.8
+ +Table 1. Performance of Depth-Based LSS vs. SCL-Aware LSS on Occ3D (in Ideal Conditions). BEVDetOcc is the baseline. + +Feature Space Deviation and Loss: Let $F_{2D}(\mathbf{x})$ be 2D features and $F_{3D}(\cdot) = \text{Lift}(F_{2D}(\mathbf{x}), \cdot)$ . Assuming Lift is $L_{\text{Lift}}$ -Lipschitz continuous, the 3D feature deviation is: + +$$ +\begin{array}{l} \left\| F _ {3 D} (\mathbf {X}) - F _ {3 D} \left(\mathbf {X} _ {\text {i d e a l}}\right) \right\| _ {F} \leq L _ {\text {L i f t}} \| \mathbf {X} - \mathbf {X} _ {\text {i d e a l}} \| _ {2} \tag {5} \\ \leq L _ {L i f t} \| \Delta \mathbf {X} \| _ {2} \leq \Delta_ {F _ {3 D}}. \\ \end{array} +$$ + +The loss function is $L_{LSS} = \mathcal{L}(P_{3D}(\mathbf{X}),GT_{3D})$ , where $P_{3D}(\mathbf{X}) = Seg_{3D}(F_{3D}(\mathbf{X}))$ + +Gradient Deviation and Optimization Limit: The gradient of $L_{LSS}$ w.r.t. parameters $\theta$ is given by chain rule. However, since $M_{fixed}$ is fixed, $\frac{\partial\mathbf{X}}{\partial\theta} = 0$ . Thus, the gradient becomes: + +$$ +\nabla_ {\theta} L _ {L S S} = \frac {\partial L _ {L S S}}{\partial P _ {3 D}} \frac {\partial P _ {3 D}}{\partial F _ {3 D}} \left(\frac {\partial F _ {3 D}}{\partial F _ {2 D}} \frac {\partial F _ {2 D}}{\partial \theta}\right). \tag {6} +$$ + +Due to $\Delta_{F_{3D}} > 0$ , the computed gradient $\nabla_{\theta}L_{LSS}$ is based on the deviated feature space, i.e., + +$$ +\begin{array}{l} \nabla_ {\theta} L _ {L S S} (\theta) = \nabla_ {\theta} \mathcal {L} \left(P _ {3 D} (\mathbf {X}), G T _ {3 D}\right) \tag {7} \\ \neq \nabla_ {\theta} \mathcal {L} \left(P _ {3 D} \left(\mathbf {X} _ {i d e a l}\right), G T _ {3 D}\right) = \nabla_ {\theta} L _ {i d e a l} (\theta). \\ \end{array} +$$ + +The gradient deviation prevents mapping's direct optimization and limits convergence to an $\epsilon$ -optimal solution. + +The theoretical analysis reveals that the inherent error in the fixed 2D-to-3D mapping of Depth-Based LSS methods fundamentally hinders gradient-based optimization from achieving optimal performance. + +# 3.2. Semantic Causal Locality in VisionOcc + +As revealed in our theoretical analysis (Sec. 3.1), Depth-Based LSS methods suffer from inherent error propagation due to their fixed 2D-to-3D mapping, which limits optimization efficacy and potential performance. To overcome + +these limitations, we particularly focus on strengthening the semantic causality of the 2D-to-3D transformation. We argue that VisionOcc's 2D-to-3D semantic occupancy prediction should exhibit semantic causal locality (SCL) for robust perception in autonomous driving. Ideally, 2D causes should drive 3D semantic effects. For instance, a predicted "car" at 3D location $(h,w,z)$ should originate from a matching 2D image region of "car". Per the causal chain, camera parameters $P$ and estimated geometry $\mathbf{G}$ enable this dependency, with $\mathbf{G}$ being crucial to maintain SCL. + +Next, we formulate the ideal SCL condition. For a 2D pixel $(u,v)$ with semantic label $s$ , the projection probability $p_d$ (the value of $\mathbf{G}$ at this coordinate and depth $d \in D$ ) should be high if its corresponding 3D ground-truth semantic is $s$ , and low otherwise: + +$$ +p _ {d} \propto \mathbb {1} (\tilde {\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) = s), +$$ + +where $\mathbb{1}$ is the indicator function, and $e_P$ represents the potential coordinate transformation error caused by factors such as camera pose error. In practice, $p_d$ acts as a weight multiplied by 2D features (Eq. (8)), enabling a probabilistic mapping that supports differentiable backpropagation. + +Limitations of Depth-Based LSS in SCL. Depth-Based LSS does not fully account for semantic causal consistency. It only preserves semantic causal locality intuitively under ideal conditions where $e_P = 0$ and depth estimation is perfectly accurate. However, with coordinate transformation errors $e_P$ , even a high $p_d$ for the ideal depth may project 2D semantics to incorrect 3D locations, i.e., + +$$ +\tilde {\mathbf {O}} (R _ {P} (u, v, d) + 0) = s, \text {b u t} \tilde {\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) \neq s. +$$ + +This misalignment causes 2D semantics $s$ (e.g., "car") to link with wrong 3D semantics (e.g., "tree"), causing semantic ambiguity and hindering training. Moreover, depth-based LSS often propagates semantics to surface points, weakening semantic propagation to occluded regions [7]. + +Empirical Validation. We conduct an empirical study to validate our analysis of SCL, comparing two ideal geometric transformations. Using BEVDetOcc [13] as the baseline, we replace its estimated depth-based geometry for LSS + +with: $i$ ) ground-truth LiDAR depths; $ii$ ) SCL-Aware geometry, which computes $p_d$ from 2D and 3D semantic ground truths, where for $(u,v,d,s)$ , $p_d = 1$ if $\tilde{\mathbf{O}}(R_P(u,v,d)) = s$ , else $p_d = 0$ . We evaluate their performance in semantic occupancy prediction. Tab. 1 demonstrates that 2D-to-3D Transformation achieves significant performance improvements over depth-based LSS in ideal conditions, validating the benefits of semantic causal locality. + +Summary. Building on the above analysis, we propose our solution to enforce semantic causality constraints during training. The overall framework is shown in Fig. 3. + +# 3.3. Semantic Causality-Aware Causal Loss + +For lifting methods like LSS, we could directly supervise the transformation geometry $\mathbf{G}$ using the causal semantic geometry derived from ground-truth labels (as described in the previous section). However, we aim to enhance the lifting method in the next section, rendering direct supervision impractical. Thus, we design a gradient-based approach to enforce semantic causality. + +We begin within the LSS framework. For a 2D pixel feature at location $(u, v)$ , LSS multiplies it by the depth-related transformation probability $p_d$ and projects it to the 3D coordinate corresponding to depth $d$ : + +$$ +\mathbf {f} _ {L} \left(R _ {P} (u, v, d)\right) = p _ {d} (u, v, d) \cdot \mathbf {f} _ {i} (u, v, d). \tag {8} +$$ + +Here, $\mathbf{f}_L(R_P(u,v,d))\in \mathbb{R}^C$ represents the 3D voxel feature at the projected location, with $e_{P}$ omitted for notational simplicity. $\mathbf{f}_i(u,v,d)\in \mathbb{R}^C$ denotes the 2D image feature at location $(u,v)$ . We backpropagate gradients from $\mathbf{f}_L$ to $\mathbf{f}_i$ , i.e., + +$$ +\frac {\partial \sum \mathbf {f} _ {L} \left(R _ {P} (u , v , d)\right)}{\partial \mathbf {f} _ {i} (u , v , d)} = p _ {d} \cdot \mathbf {I}, \tag {9} +$$ + +where $\mathbf{I}$ is the all-ones vector. + +For each semantic class $s$ , we aggregate the features $\mathbf{f}_L$ at all 3D positions where the ground truth class equals $s$ . This aggregation is backpropagated to the 2D features $\mathbf{f}_i$ , yielding a gradient map $\nabla_s \in \mathbb{R}^{U \times V \times C}$ for class $s$ : + +$$ +\nabla_ {s} (u, v, c) = \sum_ {\left(h ^ {\prime}, w ^ {\prime}, z ^ {\prime}\right) \in \Omega_ {s}} \frac {\partial \sum \mathbf {f} _ {L} \left(h ^ {\prime} , w ^ {\prime} , z ^ {\prime} , c\right)}{\partial \mathbf {f} _ {i} (u , v , c)}, \tag {10} +$$ + +where $\Omega_s = \{(h', w', z') \mid O(h', w', z') = s\}$ is the set of 3D positions with semantic label $s$ , and $c$ indexes the feature channels. Averaging over channel $C$ produces an attention map $A_s \in \mathbb{R}^{U \times V}$ : + +$$ +A _ {s} (u, v) = \frac {1}{C} \sum_ {c = 1} ^ {C} \nabla_ {s} (u, v, c). \tag {11} +$$ + +Finally, we enforce per-pixel constraints using 2D + +ground-truth labels with a binary cross-entropy loss: + +$$ +\begin{array}{l} L _ {b c e} ^ {s} = - \frac {1}{U \cdot V} \sum_ {u, v} \left[ Y _ {s} (u, v) \log A _ {s} (u, v) \right. \tag {12} \\ \left. + \left(1 - Y _ {s} (u, v)\right) \log (1 - A _ {s} (u, v)) \right], \\ \end{array} +$$ + +where $Y_{s}(u,v) \in \{0,1\}$ is the 2D ground-truth label for semantic class $s$ at pixel $(u,v)$ . + +The gradient computation can be performed using the automatic differentiation calculator like torch.autograd, requiring $S$ backward passes to iterate over the $S$ semantic classes. This incurs a computational overhead scaling linearly with the class amount $S$ . To mitigate this overhead, we reformulate the loss computations in terms of the expectation of an unbiased estimator [11]. We define the expected BCE loss across all semantic classes $s \in \{1, \dots, S\}$ as: + +$$ +\mathbb {E} \left[ L _ {b c e} ^ {s} \right] = \frac {1}{S} \sum_ {s = 1} ^ {S} L _ {b c e} ^ {s}. \tag {13} +$$ + +Based on this relationship, we uniformly sample a single semantic class $s$ during training: + +$$ +L _ {c a u s a l} = L _ {b c e} ^ {s}, \quad s \sim \operatorname {U n i f o r m} (1, S). \tag {14} +$$ + +This sampling preserves the unbiased nature of the gradient and loss estimates and reduces the computational cost to $\frac{1}{S}$ . $L_{\text{causal}}$ focuses on enhancing the geometric transformation, serving as an auxiliary term to complement the occupancy loss (case-by-case, e.g., cross-entropy in BEVDet [13]). + +# 3.4. Semantic Causality-Aware Transformation + +We propose semantic causality-aware 2D-to-3D transformation to enhance 2D-to-3D lifting, as shown in Fig. 4. Eq. (14) constrains the geometry $\mathbf{G}$ using 2D and 3D semantics. This overcomes the rigid, per-location hard alignment of geometric probabilities (e.g., using LiDAR depth supervision) in prior methods [7, 13, 22, 23, 27]. It enables advanced lifting designs, addressing errors from camera pose and other distortions. + +# 3.4.1. Channel-Grouped Lifting + +Vanilla LSS applies uniform weights to all 2D feature channels. We argue this is trivial as 2D and 3D features have distinct locality biases. For instance, a 2D "car" edge may capture "tree" semantics via convolution, but in 3D, these objects are distant. Uniform weighting both semantics causes ambiguity. Since different channels typically encode distinct semantics, we group the feature channels and learn unique weights for each group: + +$$ +\mathbf {f} _ {L, g} \left(R _ {P} (u, v, d)\right) = \omega_ {g, d} \cdot \mathbf {f} _ {i, g} (u, v, d), g \in \{1, \dots , N _ {g} \}, \tag {15} +$$ + +where $\mathbf{f}_{L,g}\in \mathbb{R}^{C / N_g}$ and $\mathbf{f}_{i,g}\in \mathbb{R}^{C / N_g}$ are the 3D and 2D features for group $g$ . $\omega_{g,d}$ is the learned weight for group + +![](images/b35fc75d231592dfad58e5042f4dec235ce01c814d7024733f9396f3c81da568.jpg) +(a) Channel-Grouped Lifting + +![](images/984cc9ef1d47c4edd313521fd39c3ec47f8a062f101a03dee85ded2f278e13f4.jpg) + +![](images/0ea1a77761edd9899688dd59f5533e25f88bab654f20db6ecd83d107b909228b.jpg) +(b) Coordinate Mapping with Learnable Camera Offsets +(c) Post-Processing with Normalized Convolution +Figure 4. The Detailed Structure of the Proposed Modules in Semantic Causality-Aware 2D-to-3D Transformation. There are three key components, including (a): Channel-Groupled Lifting, (b): Learnable Camera Offsets, and (c): Normalized Convolution, enabling accurate and robust 2D-to-3D transformation. + +$g$ , replacing $p_d$ which uniformly lifts all channels. $N_g$ is the number of groups. This preserves semantic distinction, ensuring channel-specific causal alignment. + +# 3.4.2. Learnable Camera Offsets + +To address camera parameter errors, especially pose inaccuracies, we introduce learnable offsets into camera parameters. First, we ensure the lifting process is coordinatedifferentiable. This is crucial for the offsets to receive gradients and adapt during training. The transformation of 2D image coordinates $(u,v)$ and depth $d$ to 3D voxel coordinates can be represented as matrix multiplication: + +$$ +[ h, w, z ] ^ {T} = P \cdot [ u \cdot d, v \cdot d, d, 1 ] ^ {T}, \tag {16} +$$ + +where $P \in \mathbb{R}^{3 \times 4}$ is the camera projection matrix combining intrinsics and extrinsics. LSS typically rounds floating-point coordinates to integers, rendering them nondifferentiable. Following ALOcc [7], we use the soft filling to enable differentiability w.r.t. position. This method calculates distances between floating-point 3D coordinates and their eight surrounding integer coordinates. These distances serve as weights to distribute a 2D feature at $(u, v, d)$ across multiple 3D locations. Lifting can be rewritten to + +$$ +\begin{array}{l} \mathbf {f} _ {L, g} \left(h ^ {\prime}, w ^ {\prime}, z ^ {\prime}\right) = \omega_ {g, d} \cdot \omega_ {h ^ {\prime}, w ^ {\prime}, z ^ {\prime}} \cdot \mathbf {f} _ {i, g} (u, v, d), \tag {17} \\ \forall \left(h ^ {\prime}, w ^ {\prime}, z ^ {\prime}\right) \in \text {n e i g h b o r s}, \\ \end{array} +$$ + +where $\omega_{h',w',z'}$ are trilinear interpolation weights. + +Next, we propose learning two offsets. First, we directly + +predict an offset applied to the camera parameters: + +$$ +P := P + \Delta P, \quad \Delta P = F _ {\text {o f f s e t 1}} \left(\mathbf {f} _ {i}, P\right), \tag {18} +$$ + +where $\Delta P$ is the predicted parameter offset, and $F_{offset1}$ denotes the network. Second, we estimate per-position offsets for each $(u,v,d)$ in the image coordinate system: + +$$ +(u, v, d) := (u + \Delta u, v + \Delta v, d + \Delta d), \tag {19} +$$ + +$$ +\left(\Delta u, \Delta v, \Delta d\right) = F _ {\text {o f f s e t} 2} \left(\mathbf {f} _ {i} (u, v, d)\right), +$$ + +where $F_{offset2}$ is another network. These offsets enable the model to adaptively compensate for camera parameter errors $e_P$ , improving geometric accuracy while preserving semantic causal locality under such errors. + +# 3.4.3. Normalized Convolution + +Prior work [28] notes that the direct mapping in LSS yields sparse 3D features. To address this, an intuitive solution is to use local feature propagation operators (e.g., convolutions) with causal loss supervision for preserving semantic causality. However, vanilla convolutions lack gradient constraints during causal loss computation, producing poor results. We address this by normalizing convolution weights to keep gradient map values in [0, 1]. Given the difficulty of normalizing standard convolution weights, we follow MobileNet and ConvNext to adopt depthwise (spatial) and pointwise (channel) decomposition. For the depthwise kernel $W_{\mathrm{spatial}} \in \mathbb{R}^{3 \times 3 \times 3 \times C}$ , we apply softmax across spatial dimensions $(h, w, z)$ per channel $c$ : + +$$ +W _ {\text {s p a t i a l}} ^ {\prime} [ h, w, z, c ] = \frac {\exp \left(W _ {\text {s p a t i a l}} [ h , w , z , c ]\right)}{\sum_ {h ^ {\prime} , w ^ {\prime} , z ^ {\prime}} \exp \left(W _ {\text {s p a t i a l}} \left[ h ^ {\prime} , w ^ {\prime} , z ^ {\prime} , c \right]\right)}. \tag {20} +$$ + +For the pointwise kernel $W_{\mathrm{channel}} \in \mathbb{R}^{C \times C}$ , we apply softmax across input channels per output channel: + +$$ +W _ {\text {c h a n n e l}} ^ {\prime} \left[ c _ {\text {i n}}, c _ {\text {o u t}} \right] = \frac {\exp \left(W _ {\text {c h a n n e l}} \left[ c _ {\text {i n}} , c _ {\text {o u t}} \right]\right)}{\sum_ {c _ {\text {o u t}} ^ {\prime}} \exp \left(W _ {\text {c h a n n e l}} \left[ c _ {\text {i n}} , c _ {\text {o u t}} ^ {\prime} \right]\right)}. \tag {21} +$$ + +Specifically, we use transposed convolution for depthwise convolutions. It diffuses features from non-zero to zero positions, but the zero position does not affect others. We prove in the supplement that the derived gradient mask remains within [0, 1] even with our semantic causality-aware 2D-to-3D transformation. + +# 3.5. Validation of Gradient Error Mitigation + +Fig. 5 shows the occupancy loss curves during training for BEVDet, comparing performance with and without our method. The results show that BEVDet integrated with our approach (blue line) achieves a significantly faster and steeper loss reduction compared to the original BEVDet (red line). This empirical evidence validates our theoretical analysis of gradient error mitigation. As Theorem Theorem 1 formalizes, Depth-Based LSS methods are inherently limited by gradient deviations due to fixed 2D-to-3D + +![](images/8c7e40b5c0db6937e75f1633b2673a8b1d62dfa14ed19263daf7d4c34633847c.jpg) +Figure 5. Occupancy Loss Curves with/without Our Method. Our method reduces training loss through semantic causal 2D-to-3D geometry transformation, as shown in the comparison before (red) and after (blue) its application. + +
MethodmIoU ↑DropmIoUD↑DropIoU ↑Drop
BEVDetOcc [12]37.1-32.3%30.2-49.0%70.4-8.5%
BEVDetOcc w/ Noise25.115.464.4
BEVDetOcc+Ours38.3-7.3%31.5-10.8%71.2-1.4%
BEVDetOcc+Ours w/ Noise35.528.170.2
ALoCC [7]40.1-21.9%34.3-28.6%70.2-8.4%
ALoCC w/ Noise31.324.564.3
ALoCC+Ours40.9-3.3%35.5-4.8%70.7-1.0%
ALoCC+Ours w/ Noise39.633.870.0
+ +Table 2. Performance Comparison on the Occ3D Dataset with Gaussian Noise Added to Camera Parameters. The "Drop (%) columns show degradation, with our methods (BEVDetOcc+Ours, ALOcc+Ours) achieving much smaller drops (e.g., -7.3% mIoU vs. -32.4% for BEVDetOcc). + +mapping errors. In contrast, our method aims to alleviate this by enabling a learnable mapping and incorporating a causal loss. It indicates that by mitigating gradient error through semantic causality-aware 2D-to-3D transformation, our approach facilitates more efficient gradient-based learning, leading to faster convergence and a lower loss curve. + +# 4. Experiment + +# 4.1. Experimental Setup + +Dataset. In this study, we leverage the Occ3D-nuScenes dataset [5, 39], a comprehensive dataset with diverse scenes for autonomous driving research. It encompasses 700 scenes for training, 150 for validation, and 150 for testing. Each scene integrates a 32-beam LiDAR point cloud alongside 6 RGB images, captured from multiple perspectives encircling the ego vehicle. Occ3D [39] introduces voxel-based annotations, covering a spatial extent of $-40\mathrm{m}$ to $40\mathrm{m}$ along the X and Y axes, and $-1\mathrm{m}$ to $5.4\mathrm{m}$ along the Z axis, with a consistent voxel size of $0.4\mathrm{m}$ across all dimensions. Occ3D delineates 18 semantic categories, comprising 17 distinct object classes plus an empty class to signify unoccupied regions. Following [7, 27, 35, 39], we evalu + +![](images/6a25e7a9ca91d4b12c60c8b544975b4dae0d485f86571abe220867e095cae644.jpg) + +![](images/36e3fcbb843c288828c0a7573947a8a3e5ddecb152ab327f76a7f277a3d5dfac.jpg) + +![](images/0136db64c9a04fd9b4b8a19aa6469c5da59aba4c623a0ecb0d5af6e88d9e22bb.jpg) +(a) BEVDet + +![](images/fc7f33cc3030f9bbcf9ee9c392bfbdbb23a50a0f4c906b31ae2a168ac094e284.jpg) +(b) BEVDet + Ours +Figure 6. Visualization of 2D-to-3D Semantic Causal Consistency Using LayerCAM [18]. We enhance BEVDet with our method for comparison. Attention maps are computed for critical traffic classes: "traffic cone", "pedestrian", and "car". Areas of greatest difference are marked with boxes. Each class-specific localization highlights our method's precise focus over vanilla BEVDet, showing improved semantic alignment. + +ate the occupancy prediction performance with mIoU across 17 semantic object categories, $\mathrm{mIoU}_D$ for 8 dynamic categories, and occupied/unoccupied IoU for scene geometry. + +Implementation Details. We integrate our approach into BEVDet [12] and ALOcc [7] for performance evaluation. The model parameters, image, and BEV augmentation strategies are retrained as the original. For ALOcc, we remove its ground-truth depth denoising module to adapt to our method. We use multiple convolutional layers to predict the two camera parameter offsets. The loss weight of $L_{\text{causal}}$ is set to 0.02 in the main experiments. We optimize using AdamW [33] with a learning rate of $2 \times 10^{-4}$ , a global batch size of 16, and for 24 epochs. Experiments use single-frame surrounding images, emphasizing improvements from enhanced 2D-to-3D transformations. + +# 4.2. Evaluation of Camera Perturbation Robustness + +To assess our method's robustness under extreme noise, we add Gaussian noise with (0.1 variances) to camera parameters in training and testing, as shown in Tab. 2. Compared to vanilla BEVDetOcc and ALOcc, our methods show smaller performance drops. BEVDetOcc+Ours has a mIoU drop of $-7.3\%$ vs. $-32.3\%$ for BEVDetOcc, and ALOcc+Ours drops $-3.3\%$ vs. $-21.9\%$ for ALOcc, demonstrating enhanced resilience to noisy parameters. This effectively counters motion-induced errors, benefiting self-driving tasks. + +# 4.3. Semantic Causality Visualization + +As shown in the Fig. 6, we visualize 3D-to-2D semantic causal consistency using LayerCAM [18]. We backpropagate the final 3D semantic occupancy predictions of dis + +
MethodBackboneInput SizemIoUmIoUDIoU
MonoScene [6]ResNet-101928 × 16006.15.4-
CTF-Occ [39]ResNet-101928 × 160028.527.4-
TPVFormer [14]ResNet-101928 × 160027.827.2-
COTR [35]ResNet-50256 × 70439.133.869.6
ProtoOcc [19]ResNet-50256 × 70439.634.3-
LightOcc-S [50]ResNet-50256 × 70437.932.4-
DHD-S [45]ResNet-50256 × 70436.530.7-
FlashOCC [48]ResNet-50256 × 70432.024.765.3
FB-Occ [27]ResNet-50256 × 70435.730.966.5
BEVDetOcc [12]ResNet-50256 × 70437.130.270.4
BEVDetOcc+OursResNet-50256 × 70438.3 ↑1.231.5 ↑1.371.2 ↑1.2
ALOcc [7]ResNet-50256 × 70440.134.370.2
ALOcc+OursResNet-50256 × 70440.9 ↑0.835.5 ↑1.170.7 ↑0.5
+ +tinct classes to the 2D feature maps fed into SCAT. Notably, we are the first to apply LayerCAM for cross-dimensional analysis. Fig. 6 (b) shows our method precisely focuses on class-specific locations. This confirms LayerCAM's cross-dimensional effectiveness. Our approach surpasses vanilla BEVDet (Fig. 6 (a)) in targeting class-associated objects. This proves improved semantic localization. + +# 4.4. Benchmarking with Previous Methods + +As shown in Tab. 3, we compare our method with leading 3D semantic occupancy prediction approaches on Occ3D [39]. Specifically, compared to baseline models BEVDet and ALOcc, our method achieves significant improvements in mIoU, mIoU $_{\mathrm{D}}$ , and IoU. For instance, the BEVDetOcc+Ours variant achieves an mIoU of 38.3, surpassing BEVDetOcc [12] by 1.2, while improving mIoU $_{\mathrm{D}}$ by 1.3 and IoU by 1.2. Similarly, ALOcc+Ours shows gains of 0.8 in mIoU, 1.1 in mIoU $_{\mathrm{D}}$ , and 0.5 in IoU over ALOcc [7]. These results validate the superiority of our semantic causality-aware 2D-to-3D transformation. + +# 4.5. Ablation Study + +Effect of Causal Loss. We first investigate the effectiveness of the proposed Causal Loss in enhancing the occupancy prediction performance. As demonstrated in Tab. 4, the impact of the proposed Causal Loss is evaluated through a series of experiments. Specifically, using BEVDetOcc as the baseline (Exp. 0), we conducted two ablation studies: one removing the depth supervision loss (Exp. 1), and another incorporating the proposed Causal Loss (Exps. 2, 3). The results reveal that the removal of the depth supervision loss leads to marginal decrease in performance, whereas the addition of the proposed Causal Loss yields a significant improvement. This suggests that Causal Loss facilitates superior 2D-to-3D transformation and ultimately enhances the + +Table 3. Comparison of 3D Semantic Occupancy Prediction Using Single Frame on the Occ3D Dataset, Evaluated mIoU, mIoUD, and IoU Metrics. Performance gains are indicated by red arrows $\uparrow$ . Our proposed approach (+Ours) consistently demonstrates superior enhancement over existing methods. + +
Exp.MethodmIoUDiff.mIoUDIoULatency
0Baseline (BEVDetOcc) [12]37.1-30.270.4416/125
1w/o Depth Sup36.8-0.329.670.3414/125
2+ Causal Loss37.6+0.831.070.1450/125
3+ Unbiased Estimator37.5-0.130.770.5417/125
4w/o Post Conv37.3-0.230.770.2379/122
5+ Channel-Grouped Lifting37.6+0.330.770.7419/128
6+ Soft Filling37.6-30.670.7434/149
7+ Learnable Camera Offset37.9+0.331.171.0446/150
8+ Normalized Convolution38.3+0.431.571.2466/159
+ +Table 4. Ablation Study of 3D Semantic Occupancy Prediction on Occ3D. We comprehensively evaluated the impact of the individual strategy (bolded rows) proposed in our paper, with BEVDetOcc as the baseline. The final column reports single-frame training/inference latency (ms) on an RTX 4090 GPU. + +precision of 3D semantic occupancy prediction. Comparing Exp. 3 to Exp. 2, the Unbiased Estimator simplifies the Causal Loss computation, reducing training overhead. + +Effect of Each Module. In Tab. 4 (Exps. 5-8), we systematically validate the effectiveness of the three proposed modules. We first remove the two post-lifting convolutional layers from the original BEVDet (Exp. 4), as they serve a similar role to our proposed modules in refining volume features. Subsequently, we incrementally integrate the three proposed modules (Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution) into the baseline model. To ensure gradient flow for Learnable Camera Offsets, we introduce the Soft Filling strategy from ALOcc [7] in Exp. 6, enabling effective training of the camera parameter offsets in Exp. 7. The results show progressive performance improvements with each added component (Exps. 5, 7, 8), confirming the efficacy of the proposed SCAT method. Additionally, the proposed modules incur acceptable computational overhead. + +Please refer to the supplementary material for more comprehensive experimental results. + +# 5. Conclusion + +In this paper, we introduced a novel approach leveraging causal principles to address the limitation that ignored the reliability and interpretability by existing methods. By exploring the causal foundations of 3D semantic occupancy prediction, we propose a causal loss that enhances semantic causal consistency. In addition, we develop the SCAT module with three main components: Channel-Groupled Lifting, Learnable Camera Offsets and Normalized Convolution. This approach effectively mitigates transformation inaccuracies arising from uniform mapping weights, camera perturbations, and sparse mappings. Experiments demonstrate that our approach achieves significant improvements in accuracy, robustness to camera perturbations, and semantic causal consistency in 2D-to-3D transformations. + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 3 +[2] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Langocc: Self-supervised open vocabulary occupancy estimation via volume rendering. arXiv preprint arXiv:2407.17310, 2024. 3 +[3] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Occfownet: Towards self-supervised occupancy estimation via differentiable rendering and occupancy flow. arXiv preprint arXiv:2402.12792, 2024. 3 +[4] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Gaussianflowocc: Sparse and weakly supervised occupancy estimation using gaussian splatting and temporal flow. arXiv preprint arXiv:2502.17288, 2025. 3 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 7 +[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 8 +[7] Dubing Chen, Jin Fang, Wencheng Han, Xinjing Cheng, Junbo Yin, Chengzhong Xu, Fahad Shahbaz Khan, and Jianbing Shen. Alocc: adaptive lifting-based 3d semantic occupancy and cost volume-based flow prediction. arXiv preprint arXiv:2411.07725, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[8] Dubing Chen, Wencheng Han, Jin Fang, and Jianbing Shen. Adaocc: Adaptive forward view transformation and flow modeling for 3d occupancy and flow prediction. arXiv preprint arXiv:2407.01436, 2024. 3 +[9] Dubing Chen, Huan Zheng, Jin Fang, Xingping Dong, Xianfei Li, Wenlong Liao, Tao He, Pai Peng, and Jianbing Shen. Rethinking temporal fusion with a unified gradient descent view for 3d semantic occupancy prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1505-1515, 2025. 1, 3 +[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4193-4202, 2020. 2 +[11] Judy Hoffman, Daniel A Roberts, and Sho Yaida. Robust learning with jacobian regularization. arXiv preprint arXiv:1908.02729, 2019. 5 +[12] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7, 8 +[13] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de + +tection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 4, 5 +[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023. 1, 2, 3, 8 +[15] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024. +[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376-393. Springer, 2024. 3 +[17] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding. arXiv preprint arXiv:2412.13193, 2024. 3 +[18] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 7 +[19] Jungho Kim, Changwon Kang, Dongyoung Lee, Sehwan Choi, and Jun Won Choi. Protoocc: Accurate, efficient 3d occupancy prediction using dual branch encoder-prototype query decoder. arXiv preprint arXiv:2412.08774, 2024. 8 +[20] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2019. 2 +[21] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3 +[22] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 5 +[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 3, 5 +[24] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 2, 3 +[25] Yangguang Li, Bin Huang, Zeren Chen, Yufeng Cui, Feng Liang, Mingzhu Shen, Fenggang Liu, Enze Xie, Lu Sheng, Wanli Ouyang, et al. Fast-bev: A fast and strong bird's + +eye view perception baseline. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3 +[27] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 2, 3, 5, 7, 8 +[28] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 6 +[29] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR 2024 Autonomous Grand Challenge Track On Occupancy and Flow, 2024. 3 +[30] Lizhe Liu, Bohua Wang, Hongwei Xie, Daqi Liu, Li Liu, Zhiqiang Tian, Kuiyuan Yang, and Bing Wang. Surroundsdf: Implicit 3d scene understanding based on signed distance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1 +[31] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. Advances in Neural Information Processing Systems, 31, 2018. 2 +[32] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 3 +[33] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. 7 +[34] Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023. 3 +[35] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. arXiv preprint arXiv:2312.01919, 2023. 1, 2, 3, 7, 8 +[36] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12404-12411. IEEE, 2024. 3 +[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2, 3 + +[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as set of points. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3 +[39] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7, 8 +[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 1, 3 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[42] Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160, 2023. 2 +[43] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 3 +[44] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551, 2023. 1, 3 +[45] Yuan Wu, Zhiqiang Yan, Zhengxue Wang, Xiang Li, Le Hui, and Jian Yang. Deep height decoupling for precise vision-based 3d occupancy prediction. arXiv preprint arXiv:2409.07972, 2024.8 +[46] Ziyang Yan, Wenzhen Dong, Yihua Shao, Yuhang Lu, Liu Haiyang, Jingwen Liu, Haozhe Wang, Zhe Wang, Yan Wang, Fabio Remondino, et al. Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356, 2024.3 +[47] Shichao Yang, Yulan Huang, and Sebastian Scherer. Semantic 3d occupancy mapping through efficient high order crfs. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 590-597. IEEE, 2017. 1 +[48] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 8 +[49] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 3 +[50] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Lightweight spatial embedding for vision-based 3d occupancy prediction. arXiv preprint arXiv:2412.05976, 2024.8 + +[51] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 2, 3 +[52] Jilai Zheng, Pin Tang, Zhongdao Wang, Guoqing Wang, Xiangxuan Ren, Bailan Feng, and Chao Ma. Veon: Vocabulary-enhanced occupancy prediction. In European Conference on Computer Vision, pages 92-108. Springer, 2024. 3 +[53] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 15890-15902, 2024. 2 \ No newline at end of file diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/images.zip b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b633b0d23b69a39326e40adb2d6bb3e0a3d4b4b0 --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf343dda4be9569d47b4af2a77daaa7351fcaf16a2d0ca92d2db69dd67973a79 +size 514360 diff --git a/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/layout.json b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..426a61ef8dd24ed55c9600f1779b18ffa4b63f23 --- /dev/null +++ b/2025/Semantic Causality-Aware Vision-Based 3D Occupancy Prediction/layout.json @@ -0,0 +1,11187 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "Semantic Causality-Aware Vision-Based 3D Occupancy Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "spans": [ + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": "Dubing Chen" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Huan Zheng" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Yucheng Zhou" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Xianfei Li" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Wenlong Liao" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Tao He" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Pai Peng" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": ", Jianbing Shen" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": "SKL-IOTSC, CIS, University of Macau " + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 143, + 474, + 200 + ], + "type": "text", + "content": "COWAROBOT Co. Ltd. https://github.com/cdb342/CausalOcc" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 296, + 553 + ], + "type": "text", + "content": "Vision-based 3D semantic occupancy prediction is a critical task in 3D vision that integrates volumetric 3D reconstruction with semantic understanding. Existing methods, however, often rely on modular pipelines. These modules are typically optimized independently or use pre-configured inputs, leading to cascading errors. In this paper, we address this limitation by designing a novel causal loss that enables holistic, end-to-end supervision of the modular 2D-to-3D transformation pipeline. Grounded in the principle of 2D-to-3D semantic causality, this loss regulates the gradient flow from 3D voxel representations back to the 2D features. Consequently, it renders the entire pipeline differentiable, unifying the learning process and making previously non-trainable components fully learnable. Building on this principle, we propose the Semantic Causality-Aware 2D-to-3D Transformation, which comprises three components guided by our causal loss: Channel-Groupled Lifting for adaptive semantic mapping, Learnable Camera Offsets for enhanced robustness against camera perturbations, and Normalized Convolution for effective feature propagation. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the Occ3D benchmark, demonstrating significant robustness to camera perturbations and improved 2D-to-3D semantic consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 597, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 647 + ], + "type": "text", + "content": "Predicting dense 3D semantic occupancy is a fundamental task in 3D vision, providing a fine-grained voxel representation of scene geometry and semantics [39, 40, 43, 47]. The challenge of performing this prediction from vision alone," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 226, + 550, + 311 + ], + "blocks": [ + { + "bbox": [ + 318, + 226, + 550, + 311 + ], + "lines": [ + { + "bbox": [ + 318, + 226, + 550, + 311 + ], + "spans": [ + { + "bbox": [ + 318, + 226, + 550, + 311 + ], + "type": "image", + "image_path": "688f0050e8c1b14bbd9adc04a36708ea013cc74e6c54793c41befb5f3f624927.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 313, + 526, + 323 + ], + "lines": [ + { + "bbox": [ + 345, + 313, + 526, + 323 + ], + "spans": [ + { + "bbox": [ + 345, + 313, + 526, + 323 + ], + "type": "text", + "content": "(a) An Example of Inaccurate 2D-to-3D Transformation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 326, + 550, + 370 + ], + "blocks": [ + { + "bbox": [ + 319, + 326, + 550, + 370 + ], + "lines": [ + { + "bbox": [ + 319, + 326, + 550, + 370 + ], + "spans": [ + { + "bbox": [ + 319, + 326, + 550, + 370 + ], + "type": "image", + "image_path": "716ae459e91b2d010789cd4c87e05ea9d2b792c95a6dc3f8e68044f365e20358.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 372, + 497, + 380 + ], + "lines": [ + { + "bbox": [ + 372, + 372, + 497, + 380 + ], + "spans": [ + { + "bbox": [ + 372, + 372, + 497, + 380 + ], + "type": "text", + "content": "(b) Modular 2D-to-3D Transformation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 381, + 550, + 423 + ], + "blocks": [ + { + "bbox": [ + 320, + 381, + 550, + 423 + ], + "lines": [ + { + "bbox": [ + 320, + 381, + 550, + 423 + ], + "spans": [ + { + "bbox": [ + 320, + 381, + 550, + 423 + ], + "type": "image", + "image_path": "6bc752e1a1159bbd1d1122c17f2aedd2d4b26b23bcdbc93afbee95d150546f4f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 425, + 528, + 435 + ], + "lines": [ + { + "bbox": [ + 342, + 425, + 528, + 435 + ], + "spans": [ + { + "bbox": [ + 342, + 425, + 528, + 435 + ], + "type": "text", + "content": "(c) Our End-to-End Supervised 2D-to-3D Transformation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 445, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 567 + ], + "type": "text", + "content": "Figure 1. (a) Illustrates a Visual Analysis of Semantic Ambiguity in VisionOcc. Inaccurate 2D-to-3D transformations may lead to positional shifts, misaligning supervision signals and resulting in semantic ambiguity. (b) Depicts the Conventional Modular 2D-to-3D Transformation Paradigm [7, 13, 27, 35], which employs depth supervision for geometry estimation, pre-calibrated camera parameters, and fixed mapping for lifting. (c) Presents Our Holistic, End-to-End Supervised 2D-to-3D Transformation Paradigm, which eliminates the need for separate modular supervision or pre-calibration, enabling unified error propagation to supervise all components." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "content": "an approach known as vision-based 3D semantic occupancy prediction (VisionOcc), has recently become a focal point of research. By leveraging only commodity cameras, VisionOcc is pivotal for a wide range of 3D applications, serving as a comprehensive digital replica of the environment for tasks like analysis, simulation, and interactive visualization [7, 9, 14, 27, 39, 40, 44]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "VisionOcc unifies the challenges of feed-forward 3D reconstruction and dense semantic understanding. Existing pipelines typically decompose this task into two metaphases [7, 14, 27, 30, 35]. The initial 2D-to-3D transfor" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 55, + 655, + 296, + 712 + ], + "type": "text", + "content": " Corresponding author: Jianbing Shen. This work was supported in part by the Science and Technology Development Fund of Macau SAR (FDCT) under grants 0102/2023/RIA2 and 0154/2022/A3 and 001/2024/SKL and CG2025-IOTSC, the University of Macau SRG2022-00023-IOTSC grant, and the Jiangyin Hi-tech Industrial Development Zone under the Taihu Innovation Scheme (EF2025-00003-SKL-IOTSC)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24878" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "mation uses large receptive field operators like view lifting [7, 37] or cross attention [14, 27] to construct an initial 3D feature volume. Subsequently, a 3D representation learning phase employs operators with a small receptive field (e.g., 3D convolutions or local self-attention) to refine 3D features and produce the final prediction. Our work targets the 2D-to-3D transformation phase, which is more critical and error-prone. Fig. 1a showcases a primary failure mode where features of one class (e.g., a 2D 'car') are erroneously transformed to the 3D location of another (e.g., a 'tree'). This creates a flawed learning objective, forcing the model to learn a spurious association between 'car' features and a 'tree' label. Such semantic ambiguity is a principal obstacle to achieving high performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 479 + ], + "type": "text", + "content": "The prevailing VisionOcc methods, typically based on Lift-Splat-Shoot (LSS), employ a modular approach for the 2D-to-3D transformation (Fig. 1b) [7, 35, 37]. This involves supervising geometry with a proxy depth loss while relying on fixed, pre-calibrated camera parameters and a static lifting map. However, this modularity raises critical questions about robustness and optimality. First, it is susceptible to compounding errors; for example, the reliance on fixed camera parameters makes the system vulnerable to real-world perturbations like camera jitter during motion. More fundamentally, the optimality of such proxy supervision is questionable. An intermediate representation ideal for depth estimation may not be optimal for the final semantic occupancy task, inherently limiting the transformation's expressive power due to this objective misalignment. This motivates our central research question: Can we devise an end-to-end supervision framework1 that holistically optimizes the entire 2D-to-3D transformation, enabling unified semantic-aware error backpropagation and allowing traditionally fixed modules to become fully learnable?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "type": "text", + "content": "We approach this problem from a causal perspective. In VisionOcc, the 2D image semantics are the \"cause\" of the final 3D semantic \"effect\". Semantic misalignment arises from disrupted information flow from cause to effect (Fig. 1a). Therefore, instead of correcting the erroneous output, we propose to directly regularize the information flow itself. We posit that a 3D prediction for a given class should be influenced predominantly by 2D image regions of that same class. To enforce this information flow, we leverage gradients as a proxy, inspired by prior work [18, 42, 53]. For each semantic class, the gradient of its aggregated 3D features is computed w.r.t. the 2D feature map, producing a saliency-like map of 2D influence. This map is then directly supervised with the ground truth 2D segmentation mask. As shown in Fig. 1c, this establishes a principled, end-to-end supervision signal for the 2D-to-3D transformation, enabling holistic optimization of all its components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "To fully leverage our end-to-end supervision, we introduce a more expressive and learnable 2D-to-3D view transformation, termed the Semantic Causality-Aware Transformation (SCAT). A key challenge is that directly supervising gradients is inherently unstable. Therefore, the entire SCAT module is designed to constrain its gradient flow to a stable [0, 1] range. Specifically, SCAT introduces three targeted designs: i) Channel-Grouped Lifting: To better disentangle semantics, we move beyond LSS's uniform weighting and apply distinct learnable weights to different groups of feature channels. ii) Learnable Camera Offsets: To mitigate motion-induced pose errors, we introduce learnable offsets to the camera parameters, which are implicitly supervised by the 2D-3D semantic consistency enforced by our causal loss. iii) Normalized Convolution: Finally, we employ a normalized convolution to densify the sparse 3D features from LSS [28], ensuring this final step also adheres to our global gradient stability requirement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "text", + "content": "Our contributions are as follows: i) We systematically analyze the 2D-to-3D transformation in VisionOCC, identifying a critical failure mode we term semantic ambiguity. We provide a theoretical analysis proving how the modularity of prior methods leads to error propagation, offering clear guidance for future work. ii) To address these problems, we propose the Causal Loss that directly regularizes the information flow of the 2D-to-3D transformation. This enables true end-to-end optimization of all constituent modules, mitigating error accumulation and making previously fixed components, such as camera parameters, fully learnable. iii) We instantiate our principles in the Semantic Causality-Aware Transformation, a novel 2D-to-3D transformation architecture. SCAT incorporates Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution to explicitly tackle the challenges of semantic confusion, camera perturbations, and limited learnability. iv) Extensive experiments show our method significantly boosts existing models, achieving a " + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "text", + "content": " absolute mIoU gain on BEVDet. Furthermore, it demonstrates superior robustness to camera perturbations, reducing the relative performance drop on BEVDet from a severe " + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "inline_equation", + "content": "-32.2\\%" + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "text", + "content": " to a mere " + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "inline_equation", + "content": "-7.3\\%" + }, + { + "bbox": [ + 313, + 289, + 553, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 578, + 400, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 400, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 400, + 590 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 599, + 466, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 466, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 466, + 611 + ], + "type": "text", + "content": "2.1. Semantic Scene Completion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Semantic Scene Completion (SSC) [10, 20, 24, 31, 51] refers to the task of simultaneously predicting both the occupancy and semantic labels of a scene. Existing methods can be classified into indoor and outdoor approaches based on the scene type, with the former focusing on occupancy and semantic label prediction in controlled environments [10, 20, 31], while the latter shifts towards more complex outdoor settings, particularly in the context of au" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Here, \"end to end\" refers to a unified supervision scheme for a process, distinct from the concept of a monolithic network architecture." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24879" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "content": "tonomous driving [1, 6]. The core principle of SSC lies in its ability to infer the unseen, effectively bridging gaps in incomplete observations with accurate semantic understanding. MonoScene [6] introduces a 3D SSC framework that infers dense geometry and semantics from a single monocular RGB image. VoxFormer [24] presents a Transformer-based semantic scene completion framework that generates complete 3D volumetric semantics from 2D images by first predicting sparse visible voxel queries and then densifying them through self-attention with a masked autoencoder design. OccFormer [51] introduces a dual-path transformer network for 3D semantic occupancy prediction, efficiently processing camera-generated 3D voxel features through local and global pathways, and enhancing occupancy decoding with preserve-pooling and class-guided sampling to address sparsity and class imbalance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 261, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 261, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 261, + 284 + ], + "type": "text", + "content": "2.2. Vision-based 3D Occupancy Prediction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 288, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 296, + 612 + ], + "type": "text", + "content": "Vision-based 3D Occupancy Prediction [9, 14-16, 34, 38-40, 43?] aims to predict the spatial and semantic features of 3D voxel grids surrounding an autonomous vehicle from image data. This task is closely related to SSC, emphasizing the importance of multi-perspective joint perception for effective autonomous navigation. TPVFormer [14] is prior work that lifts image features into the 3D TPV space by leveraging an attention mechanism [26, 41]. Different from TPVFormer, which relies on sparse point clouds for supervision, subsequent studies, including OccNet [40], SurroundOcc [44], Occ3D [39], and OpenOccupancy [43], have developed denser occupancy annotations by incorporating temporal information or instance-level labels. Methods such as BEVDet [13], FBOcc [27], COTR [35], and ALOcc [7] leverage depth-based LSS [22, 23, 25, 37] for explicit geometric transformation, demonstrating strong performance. Some methods [2, 4, 17, 36, 46, 49, 52] have explored rendering-based methods that utilize 2D signal supervision, thereby bypassing the need for 3D annotations. Furthermore, recent research like [3, 7, 8, 21, 29, 32, 40] introduced 3D occupancy flow prediction, which addresses the movement of foreground objects in dynamic scenes by embedding 3D flow information to capture per-voxel dynamics. Unlike the above methods, we analyze the 2D-to-3D transformation process from the perspectives of error propagation and semantic causal consistency, proposing a novel approach that enhances causal consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Preliminary. VisionOcc predicts a dense 3D semantic occupancy grid from surround-view images, modeled as a causal dependency chain as shown in Fig. 2. This process involves key variables: input image " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ", estimated geometry " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ", camera parameters " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " (intrinsic & extrinsic), potential camera parameter errors " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "e_{P}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ", intermediate 3D fea" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 550, + 148 + ], + "type": "image", + "image_path": "e6dc0ed532564f3ef101e617c132dcadb62b9d4e31e01db2e726f7aeeb832c15.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": "Figure 2. The Causal Structure of VisionOcc. It illustrates the dependency chain from the image input " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": " to the semantic occupancy output " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": ": geometry for 2D-to-3D transformation. " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": ": camera intrinsic and extrinsic. " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": ": intermediate 3D feature. " + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "inline_equation", + "content": "e_P" + }, + { + "bbox": [ + 313, + 149, + 555, + 205 + ], + "type": "text", + "content": ": errors in camera parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "ture " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", and final occupancy output " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ". In LSS-based methods [7, 13, 35], the pipeline starts with the image backbone extracting features " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i = F_i(\\mathbf{I}), \\mathbf{f}_i \\in \\mathbb{R}^{U,V,C}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "; geometry " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{U,V,D}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " is predicted as a probability distribution over discretized depth bins " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = F_g(\\mathbf{f}_i)" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "; 2D features are then transformed to 3D via an outer product " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_L' = \\mathbf{G} \\otimes \\mathbf{f}_i, \\mathbf{f}_L' \\in \\mathbb{R}^{U,V,D,C}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "; camera parameters " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " map these to voxel coordinates " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "R_P(u,v,d) \\to (h,w,z) \\in [0,H-1] \\times [0,W-1] \\times [0,Z-1]" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_L \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "H \\times W \\times Z" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " defines the occupancy grid resolution; finally, " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " is decoded to produce the semantic occupancy output: " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = F_o(\\mathbf{f}_L), \\mathbf{O} \\in \\mathbb{R}^{H \\times W \\times Z \\times S}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "F_o" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " is the decoding function and " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " is the number of semantic classes. The prediction " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " is supervised by the ground-truth " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{O}}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 388, + 522, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 522, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 522, + 401 + ], + "type": "text", + "content": "3.1. Error Propagation in Depth-Based LSS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": "Theorem 1. In Depth-Based LSS methods with a fixed " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": "-to-3D mapping " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "M_{\\text{fixed}}" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": ", inherent mapping error " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\delta M" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": " leads to gradient deviations, preventing convergence to an " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": "-optimal solution. This is formalized as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 462, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 462, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 319, + 462, + 555, + 475 + ], + "type": "interline_equation", + "content": "M _ {f i x e d} = M _ {i d e a l} + \\delta M \\Longrightarrow \\nabla_ {\\theta} L _ {L S S} \\neq \\nabla_ {\\theta} L _ {i d e a l}, \\tag {1}", + "image_path": "5a973482ec7146e019a78e70df01265aa1b8cddbbbbfdbef857c6f8d51881f2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "inline_equation", + "content": "M_{ideal}" + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "text", + "content": " is the ideal mapping and " + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "inline_equation", + "content": "L_{ideal}" + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "text", + "content": " is the loss function using " + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "inline_equation", + "content": "M_{ideal}" + }, + { + "bbox": [ + 313, + 484, + 554, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": "Proof. Mapping Error Quantification: Let " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " be 2D pixel coordinates and " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " be the ground truth depth. Estimated depth is " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\hat{d}(\\mathbf{x}) = d(\\mathbf{x}) + \\epsilon_d(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_d(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " as depth error. Let " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{ideal} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " be the ideal camera intrinsics, and " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{K}_{ideal} + \\epsilon_K" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " be the estimated camera intrinsics with error " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_K" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": ". The fixed mapping is " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "M_{fixed} = M_{ideal} + \\delta M" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\delta M" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " encompasses errors from various sources, including depth estimation error " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_d(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": " and camera extrinsic error " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_K" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": ". We assume the total mapping error is bounded " + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "inline_equation", + "content": "\\| \\delta M \\|_F \\leq \\Delta_M < \\infty" + }, + { + "bbox": [ + 313, + 515, + 555, + 637 + ], + "type": "text", + "content": ". The 3D coordinates are:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 647, + 553, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 647, + 553, + 660 + ], + "spans": [ + { + "bbox": [ + 365, + 647, + 553, + 660 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = M _ {\\text {f i x e d}} (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}) \\tag {2}", + "image_path": "dae650e19a081514432da8e7c717e25f60015cbf6002786123bb4e01b6dca51b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 662, + 553, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 662, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 348, + 662, + 553, + 675 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {i d e a l}} = M _ {\\text {i d e a l}} (\\mathbf {x}, d (\\mathbf {x}), \\mathbf {K} _ {\\text {i d e a l}}) \\tag {3}", + "image_path": "b3f752f1a081d919e52010145af8cc02af25d554fa414be0ba91af5f815df3f5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 678, + 553, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 678, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 358, + 678, + 553, + 693 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {X} = \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} = \\delta M (\\mathbf {x}, \\hat {d} (\\mathbf {x}), \\mathbf {K}), \\tag {4}", + "image_path": "033261805109a46b7548fa89afba143e0b31c3d009667f7144760224149b465c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 701, + 502, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 502, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 502, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 502, + 714 + ], + "type": "inline_equation", + "content": "\\| \\Delta \\mathbf{X}\\| _2\\leq C\\cdot \\Delta_M" + }, + { + "bbox": [ + 313, + 701, + 502, + 714 + ], + "type": "text", + "content": " for bounded inputs." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24880" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 68, + 500, + 167 + ], + "blocks": [ + { + "bbox": [ + 111, + 68, + 500, + 167 + ], + "lines": [ + { + "bbox": [ + 111, + 68, + 500, + 167 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 500, + 167 + ], + "type": "image", + "image_path": "653d4f3314c43e2cb0d40b3486023ef5483b3b62d55a030e67bddeed13ad7dbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 176, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 54, + 176, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 54, + 176, + 555, + 210 + ], + "type": "text", + "content": "Figure 3. The Overall Framework of the Proposed Semantic Causality-Aware VisionOcc. The proposed framework consists of three primary components: a backbone network for extracting 2D features, an SCAT module for transforming these features into 3D space, and an Encoder-Decoder network for learning 3D semantics. The SCAT module is supervised by our causal loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 219, + 276, + 266 + ], + "blocks": [ + { + "bbox": [ + 74, + 219, + 276, + 266 + ], + "lines": [ + { + "bbox": [ + 74, + 219, + 276, + 266 + ], + "spans": [ + { + "bbox": [ + 74, + 219, + 276, + 266 + ], + "type": "table", + "html": "
MethodmIoUmIoUDIoU
Depth-Based LSS44.540.478.9
SCL-Aware LSS50.5↑6.046.9↑6.585.7↑6.8
", + "image_path": "96e2c267bf887b5ae625422ad3037bb568c80c636a8db887b2fc84ab9c78d7e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 291 + ], + "type": "text", + "content": "Table 1. Performance of Depth-Based LSS vs. SCL-Aware LSS on Occ3D (in Ideal Conditions). BEVDetOcc is the baseline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "content": "Feature Space Deviation and Loss: Let " + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "inline_equation", + "content": "F_{2D}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "content": " be 2D features and " + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "inline_equation", + "content": "F_{3D}(\\cdot) = \\text{Lift}(F_{2D}(\\mathbf{x}), \\cdot)" + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "content": ". Assuming Lift is " + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "inline_equation", + "content": "L_{\\text{Lift}}" + }, + { + "bbox": [ + 55, + 297, + 296, + 334 + ], + "type": "text", + "content": "-Lipschitz continuous, the 3D feature deviation is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 342, + 295, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 295, + 371 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| F _ {3 D} (\\mathbf {X}) - F _ {3 D} \\left(\\mathbf {X} _ {\\text {i d e a l}}\\right) \\right\\| _ {F} \\leq L _ {\\text {L i f t}} \\| \\mathbf {X} - \\mathbf {X} _ {\\text {i d e a l}} \\| _ {2} \\tag {5} \\\\ \\leq L _ {L i f t} \\| \\Delta \\mathbf {X} \\| _ {2} \\leq \\Delta_ {F _ {3 D}}. \\\\ \\end{array}", + "image_path": "bff5483229b9d48aa1bf41b1183f5edc5146f390f2b13d8547fcc5c261f22bfe.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "content": "The loss function is " + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "inline_equation", + "content": "L_{LSS} = \\mathcal{L}(P_{3D}(\\mathbf{X}),GT_{3D})" + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 379, + 295, + 403 + ], + "type": "inline_equation", + "content": "P_{3D}(\\mathbf{X}) = Seg_{3D}(F_{3D}(\\mathbf{X}))" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": "Gradient Deviation and Optimization Limit: The gradient of " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "L_{LSS}" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": " w.r.t. parameters " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": " is given by chain rule. However, since " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "M_{fixed}" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": " is fixed, " + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathbf{X}}{\\partial\\theta} = 0" + }, + { + "bbox": [ + 55, + 403, + 296, + 450 + ], + "type": "text", + "content": ". Thus, the gradient becomes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 457, + 295, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 457, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 83, + 457, + 295, + 484 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} L _ {L S S} = \\frac {\\partial L _ {L S S}}{\\partial P _ {3 D}} \\frac {\\partial P _ {3 D}}{\\partial F _ {3 D}} \\left(\\frac {\\partial F _ {3 D}}{\\partial F _ {2 D}} \\frac {\\partial F _ {2 D}}{\\partial \\theta}\\right). \\tag {6}", + "image_path": "2728cdea5a3e045fe8179ee5f59036d757208f1b6dbf6e862879710ebe767c9c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "text", + "content": "Due to " + }, + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\Delta_{F_{3D}} > 0" + }, + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "text", + "content": ", the computed gradient " + }, + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}L_{LSS}" + }, + { + "bbox": [ + 55, + 491, + 295, + 516 + ], + "type": "text", + "content": " is based on the deviated feature space, i.e.," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 524, + 296, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 553 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} L _ {L S S} (\\theta) = \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} (\\mathbf {X}), G T _ {3 D}\\right) \\tag {7} \\\\ \\neq \\nabla_ {\\theta} \\mathcal {L} \\left(P _ {3 D} \\left(\\mathbf {X} _ {i d e a l}\\right), G T _ {3 D}\\right) = \\nabla_ {\\theta} L _ {i d e a l} (\\theta). \\\\ \\end{array}", + "image_path": "c984e67ef31c96b9ef5ad7a92c94752177e547fbac30ae6818d310c3297c354b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 561, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 586 + ], + "type": "text", + "content": "The gradient deviation prevents mapping's direct optimization and limits convergence to an " + }, + { + "bbox": [ + 55, + 561, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 561, + 296, + 586 + ], + "type": "text", + "content": "-optimal solution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 640 + ], + "type": "text", + "content": "The theoretical analysis reveals that the inherent error in the fixed 2D-to-3D mapping of Depth-Based LSS methods fundamentally hinders gradient-based optimization from achieving optimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 647, + 260, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 260, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 260, + 661 + ], + "type": "text", + "content": "3.2. Semantic Causal Locality in VisionOcc" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "As revealed in our theoretical analysis (Sec. 3.1), Depth-Based LSS methods suffer from inherent error propagation due to their fixed 2D-to-3D mapping, which limits optimization efficacy and potential performance. To overcome" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "content": "these limitations, we particularly focus on strengthening the semantic causality of the 2D-to-3D transformation. We argue that VisionOcc's 2D-to-3D semantic occupancy prediction should exhibit semantic causal locality (SCL) for robust perception in autonomous driving. Ideally, 2D causes should drive 3D semantic effects. For instance, a predicted \"car\" at 3D location " + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "inline_equation", + "content": "(h,w,z)" + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "content": " should originate from a matching 2D image region of \"car\". Per the causal chain, camera parameters " + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "content": " and estimated geometry " + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "content": " enable this dependency, with " + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 222, + 555, + 341 + ], + "type": "text", + "content": " being crucial to maintain SCL." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": "Next, we formulate the ideal SCL condition. For a 2D pixel " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": " with semantic label " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": ", the projection probability " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": " (the value of " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": " at this coordinate and depth " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "d \\in D" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": ") should be high if its corresponding 3D ground-truth semantic is " + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 341, + 555, + 400 + ], + "type": "text", + "content": ", and low otherwise:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 409, + 508, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 409, + 508, + 423 + ], + "spans": [ + { + "bbox": [ + 359, + 409, + 508, + 423 + ], + "type": "interline_equation", + "content": "p _ {d} \\propto \\mathbb {1} (\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) = s),", + "image_path": "91a6951579e57b0e3e059dc1a794b770ca0845cab857801cfb0a108e2adf3b82.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": " is the indicator function, and " + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "inline_equation", + "content": "e_P" + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": " represents the potential coordinate transformation error caused by factors such as camera pose error. In practice, " + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 431, + 554, + 491 + ], + "type": "text", + "content": " acts as a weight multiplied by 2D features (Eq. (8)), enabling a probabilistic mapping that supports differentiable backpropagation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": "Limitations of Depth-Based LSS in SCL. Depth-Based LSS does not fully account for semantic causal consistency. It only preserves semantic causal locality intuitively under ideal conditions where " + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "inline_equation", + "content": "e_P = 0" + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": " and depth estimation is perfectly accurate. However, with coordinate transformation errors " + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "inline_equation", + "content": "e_P" + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": ", even a high " + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 313, + 491, + 555, + 575 + ], + "type": "text", + "content": " for the ideal depth may project 2D semantics to incorrect 3D locations, i.e.," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 582, + 550, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 582, + 550, + 597 + ], + "spans": [ + { + "bbox": [ + 318, + 582, + 550, + 597 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + 0) = s, \\text {b u t} \\tilde {\\mathbf {O}} (R _ {P} (u, v, d) + e _ {P}) \\neq s.", + "image_path": "d1231ae33bbdbbd07d40db15b79741c10178bd1eea69abc49395eba811bea8b9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "content": "This misalignment causes 2D semantics " + }, + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "content": " (e.g., \"car\") to link with wrong 3D semantics (e.g., \"tree\"), causing semantic ambiguity and hindering training. Moreover, depth-based LSS often propagates semantics to surface points, weakening semantic propagation to occluded regions [7]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Empirical Validation. We conduct an empirical study to validate our analysis of SCL, comparing two ideal geometric transformations. Using BEVDetOcc [13] as the baseline, we replace its estimated depth-based geometry for LSS" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24881" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "with: " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " ) ground-truth LiDAR depths; " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " ) SCL-Aware geometry, which computes " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " from 2D and 3D semantic ground truths, where for " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "(u,v,d,s)" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "p_d = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{O}}(R_P(u,v,d)) = s" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ", else " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "p_d = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": ". We evaluate their performance in semantic occupancy prediction. Tab. 1 demonstrates that 2D-to-3D Transformation achieves significant performance improvements over depth-based LSS in ideal conditions, validating the benefits of semantic causal locality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 205 + ], + "type": "text", + "content": "Summary. Building on the above analysis, we propose our solution to enforce semantic causality constraints during training. The overall framework is shown in Fig. 3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 261, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 261, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 261, + 227 + ], + "type": "text", + "content": "3.3. Semantic Causality-Aware Causal Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": "For lifting methods like LSS, we could directly supervise the transformation geometry " + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 232, + 296, + 316 + ], + "type": "text", + "content": " using the causal semantic geometry derived from ground-truth labels (as described in the previous section). However, we aim to enhance the lifting method in the next section, rendering direct supervision impractical. Thus, we design a gradient-based approach to enforce semantic causality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": "We begin within the LSS framework. For a 2D pixel feature at location " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": ", LSS multiplies it by the depth-related transformation probability " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": " and projects it to the 3D coordinate corresponding to depth " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 374, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 374, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {L} \\left(R _ {P} (u, v, d)\\right) = p _ {d} (u, v, d) \\cdot \\mathbf {f} _ {i} (u, v, d). \\tag {8}", + "image_path": "5f5b5ab4b6af83501bf7d433a8c314f2b029119af4fb6e5a1a66fe750871c867.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_L(R_P(u,v,d))\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": " represents the 3D voxel feature at the projected location, with " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "e_{P}" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": " omitted for notational simplicity. " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i(u,v,d)\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": " denotes the 2D image feature at location " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": ". We backpropagate gradients from " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_L" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 55, + 397, + 296, + 456 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 457, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 457, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 113, + 457, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(R _ {P} (u , v , d)\\right)}{\\partial \\mathbf {f} _ {i} (u , v , d)} = p _ {d} \\cdot \\mathbf {I}, \\tag {9}", + "image_path": "f66087b1ecbc7a3f57ac61237c892ac6c2c35e8e191a5e7ec4d2159f8d65e0e2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 489, + 178, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 178, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 178, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 489, + 178, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 489, + 178, + 500 + ], + "type": "text", + "content": " is the all-ones vector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": "For each semantic class " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ", we aggregate the features " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_L" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " at all 3D positions where the ground truth class equals " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ". This aggregation is backpropagated to the 2D features " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ", yielding a gradient map " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\nabla_s \\in \\mathbb{R}^{U \\times V \\times C}" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 502, + 296, + 550 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 559, + 295, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 295, + 591 + ], + "type": "interline_equation", + "content": "\\nabla_ {s} (u, v, c) = \\sum_ {\\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\Omega_ {s}} \\frac {\\partial \\sum \\mathbf {f} _ {L} \\left(h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c\\right)}{\\partial \\mathbf {f} _ {i} (u , v , c)}, \\tag {10}", + "image_path": "158be0e24244d16fad43cc55328c339c5d740a74d5276a9731ece3e60473919b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\Omega_s = \\{(h', w', z') \\mid O(h', w', z') = s\\}" + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": " is the set of 3D positions with semantic label " + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": " indexes the feature channels. Averaging over channel " + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": " produces an attention map " + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "inline_equation", + "content": "A_s \\in \\mathbb{R}^{U \\times V}" + }, + { + "bbox": [ + 55, + 600, + 296, + 649 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 112, + 659, + 295, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 659, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 112, + 659, + 295, + 691 + ], + "type": "interline_equation", + "content": "A _ {s} (u, v) = \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\nabla_ {s} (u, v, c). \\tag {11}", + "image_path": "a75b6f9dce2b872534bb06f06895046ddf4e68ead8eda488cd4fc71be62c1cbb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Finally, we enforce per-pixel constraints using 2D" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 526, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 85 + ], + "type": "text", + "content": "ground-truth labels with a binary cross-entropy loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 91, + 553, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 91, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 337, + 91, + 553, + 137 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {b c e} ^ {s} = - \\frac {1}{U \\cdot V} \\sum_ {u, v} \\left[ Y _ {s} (u, v) \\log A _ {s} (u, v) \\right. \\tag {12} \\\\ \\left. + \\left(1 - Y _ {s} (u, v)\\right) \\log (1 - A _ {s} (u, v)) \\right], \\\\ \\end{array}", + "image_path": "209b4227d9eab98c53c15d35aefa63d108b72ae3362b060ab91b9cc391e7efec.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "Y_{s}(u,v) \\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": " is the 2D ground-truth label for semantic class " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "content": "The gradient computation can be performed using the automatic differentiation calculator like torch.autograd, requiring " + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "content": " backward passes to iterate over the " + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "content": " semantic classes. This incurs a computational overhead scaling linearly with the class amount " + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "content": ". To mitigate this overhead, we reformulate the loss computations in terms of the expectation of an unbiased estimator [11]. We define the expected BCE loss across all semantic classes " + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "inline_equation", + "content": "s \\in \\{1, \\dots, S\\}" + }, + { + "bbox": [ + 313, + 168, + 554, + 264 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 272, + 553, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 386, + 272, + 553, + 304 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ L _ {b c e} ^ {s} \\right] = \\frac {1}{S} \\sum_ {s = 1} ^ {S} L _ {b c e} ^ {s}. \\tag {13}", + "image_path": "04390459d0beb3b358e9f463d5f050f071f37c2d1df71cc0841ced7f1b6a5ca6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 311, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 335 + ], + "type": "text", + "content": "Based on this relationship, we uniformly sample a single semantic class " + }, + { + "bbox": [ + 313, + 311, + 553, + 335 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 311, + 553, + 335 + ], + "type": "text", + "content": " during training:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 343, + 553, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 343, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 354, + 343, + 553, + 357 + ], + "type": "interline_equation", + "content": "L _ {c a u s a l} = L _ {b c e} ^ {s}, \\quad s \\sim \\operatorname {U n i f o r m} (1, S). \\tag {14}", + "image_path": "22197f6c52435fd05c520e3105138abecfef8b048d780a8e761b5b94ebb2a1b1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "content": "This sampling preserves the unbiased nature of the gradient and loss estimates and reduces the computational cost to " + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "inline_equation", + "content": "\\frac{1}{S}" + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "inline_equation", + "content": "L_{\\text{causal}}" + }, + { + "bbox": [ + 313, + 364, + 554, + 425 + ], + "type": "text", + "content": " focuses on enhancing the geometric transformation, serving as an auxiliary term to complement the occupancy loss (case-by-case, e.g., cross-entropy in BEVDet [13])." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 431, + 537, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 537, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 537, + 443 + ], + "type": "text", + "content": "3.4. Semantic Causality-Aware Transformation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": "We propose semantic causality-aware 2D-to-3D transformation to enhance 2D-to-3D lifting, as shown in Fig. 4. Eq. (14) constrains the geometry " + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": " using 2D and 3D semantics. This overcomes the rigid, per-location hard alignment of geometric probabilities (e.g., using LiDAR depth supervision) in prior methods [7, 13, 22, 23, 27]. It enables advanced lifting designs, addressing errors from camera pose and other distortions." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 550, + 453, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 453, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 453, + 562 + ], + "type": "text", + "content": "3.4.1. Channel-Grouped Lifting" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 564, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 661 + ], + "type": "text", + "content": "Vanilla LSS applies uniform weights to all 2D feature channels. We argue this is trivial as 2D and 3D features have distinct locality biases. For instance, a 2D \"car\" edge may capture \"tree\" semantics via convolution, but in 3D, these objects are distant. Uniform weighting both semantics causes ambiguity. Since different channels typically encode distinct semantics, we group the feature channels and learn unique weights for each group:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 669, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 669, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 319, + 669, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {L, g} \\left(R _ {P} (u, v, d)\\right) = \\omega_ {g, d} \\cdot \\mathbf {f} _ {i, g} (u, v, d), g \\in \\{1, \\dots , N _ {g} \\}, \\tag {15}", + "image_path": "95e02f611cdd74049fd3f8e2c51de6482ef1fdfd19f7827e05ff7cc695100f49.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{L,g}\\in \\mathbb{R}^{C / N_g}" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{i,g}\\in \\mathbb{R}^{C / N_g}" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": " are the 3D and 2D features for group " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\omega_{g,d}" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": " is the learned weight for group" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24882" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 148 + ], + "type": "image", + "image_path": "b35fc75d231592dfad58e5042f4dec235ce01c814d7024733f9396f3c81da568.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 148, + 221, + 159 + ], + "lines": [ + { + "bbox": [ + 127, + 148, + 221, + 159 + ], + "spans": [ + { + "bbox": [ + 127, + 148, + 221, + 159 + ], + "type": "text", + "content": "(a) Channel-Grouped Lifting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 161, + 293, + 213 + ], + "blocks": [ + { + "bbox": [ + 59, + 161, + 293, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 161, + 293, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 293, + 213 + ], + "type": "image", + "image_path": "984cc9ef1d47c4edd313521fd39c3ec47f8a062f101a03dee85ded2f278e13f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 230, + 290, + 274 + ], + "blocks": [ + { + "bbox": [ + 84, + 217, + 264, + 227 + ], + "lines": [ + { + "bbox": [ + 84, + 217, + 264, + 227 + ], + "spans": [ + { + "bbox": [ + 84, + 217, + 264, + 227 + ], + "type": "text", + "content": "(b) Coordinate Mapping with Learnable Camera Offsets" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 230, + 290, + 274 + ], + "lines": [ + { + "bbox": [ + 67, + 230, + 290, + 274 + ], + "spans": [ + { + "bbox": [ + 67, + 230, + 290, + 274 + ], + "type": "image", + "image_path": "0ea1a77761edd9899688dd59f5533e25f88bab654f20db6ecd83d107b909228b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 276, + 254, + 285 + ], + "lines": [ + { + "bbox": [ + 94, + 276, + 254, + 285 + ], + "spans": [ + { + "bbox": [ + 94, + 276, + 254, + 285 + ], + "type": "text", + "content": "(c) Post-Processing with Normalized Convolution" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 294, + 295, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 350 + ], + "type": "text", + "content": "Figure 4. The Detailed Structure of the Proposed Modules in Semantic Causality-Aware 2D-to-3D Transformation. There are three key components, including (a): Channel-Groupled Lifting, (b): Learnable Camera Offsets, and (c): Normalized Convolution, enabling accurate and robust 2D-to-3D transformation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": ", replacing " + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": " which uniformly lifts all channels. " + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "inline_equation", + "content": "N_g" + }, + { + "bbox": [ + 55, + 360, + 296, + 397 + ], + "type": "text", + "content": " is the number of groups. This preserves semantic distinction, ensuring channel-specific causal alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 403, + 197, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 197, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 197, + 414 + ], + "type": "text", + "content": "3.4.2. Learnable Camera Offsets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "text", + "content": "To address camera parameter errors, especially pose inaccuracies, we introduce learnable offsets into camera parameters. First, we ensure the lifting process is coordinatedifferentiable. This is crucial for the offsets to receive gradients and adapt during training. The transformation of 2D image coordinates " + }, + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 418, + 296, + 502 + ], + "type": "text", + "content": " to 3D voxel coordinates can be represented as matrix multiplication:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 510, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 510, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 104, + 510, + 295, + 526 + ], + "type": "interline_equation", + "content": "[ h, w, z ] ^ {T} = P \\cdot [ u \\cdot d, v \\cdot d, d, 1 ] ^ {T}, \\tag {16}", + "image_path": "11b8ff647687a16c8a3fd8302c83065d7c7d0be09486f2199163b3bb85a8986d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": " is the camera projection matrix combining intrinsics and extrinsics. LSS typically rounds floating-point coordinates to integers, rendering them nondifferentiable. Following ALOcc [7], we use the soft filling to enable differentiability w.r.t. position. This method calculates distances between floating-point 3D coordinates and their eight surrounding integer coordinates. These distances serve as weights to distribute a 2D feature at " + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "inline_equation", + "content": "(u, v, d)" + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": " across multiple 3D locations. Lifting can be rewritten to" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 651, + 295, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 651, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 71, + 651, + 295, + 679 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {f} _ {L, g} \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) = \\omega_ {g, d} \\cdot \\omega_ {h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}} \\cdot \\mathbf {f} _ {i, g} (u, v, d), \\tag {17} \\\\ \\forall \\left(h ^ {\\prime}, w ^ {\\prime}, z ^ {\\prime}\\right) \\in \\text {n e i g h b o r s}, \\\\ \\end{array}", + "image_path": "33b367f963bbed12e11a64b1bdf7b10f3d88fde251f054f3ec31c679ca0f36e0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 257, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 257, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 257, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 257, + 700 + ], + "type": "inline_equation", + "content": "\\omega_{h',w',z'}" + }, + { + "bbox": [ + 55, + 689, + 257, + 700 + ], + "type": "text", + "content": " are trilinear interpolation weights." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "Next, we propose learning two offsets. First, we directly" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 520, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 520, + 84 + ], + "type": "text", + "content": "predict an offset applied to the camera parameters:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 91, + 553, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 91, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 349, + 91, + 553, + 105 + ], + "type": "interline_equation", + "content": "P := P + \\Delta P, \\quad \\Delta P = F _ {\\text {o f f s e t 1}} \\left(\\mathbf {f} _ {i}, P\\right), \\tag {18}", + "image_path": "344c2530aceaead6977caf9e7af2dcf9060914da0c944097d0dde9fed0f9a213.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "text", + "content": " is the predicted parameter offset, and " + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "inline_equation", + "content": "F_{offset1}" + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "text", + "content": " denotes the network. Second, we estimate per-position offsets for each " + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "inline_equation", + "content": "(u,v,d)" + }, + { + "bbox": [ + 313, + 110, + 553, + 148 + ], + "type": "text", + "content": " in the image coordinate system:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 351, + 154, + 553, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 154, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 351, + 154, + 553, + 172 + ], + "type": "interline_equation", + "content": "(u, v, d) := (u + \\Delta u, v + \\Delta v, d + \\Delta d), \\tag {19}", + "image_path": "887882a4b639fd8bcbc3526c9e37890edd2d252b0d0f182acde2062c90e0a10c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 168, + 516, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 168, + 516, + 182 + ], + "spans": [ + { + "bbox": [ + 361, + 168, + 516, + 182 + ], + "type": "interline_equation", + "content": "\\left(\\Delta u, \\Delta v, \\Delta d\\right) = F _ {\\text {o f f s e t} 2} \\left(\\mathbf {f} _ {i} (u, v, d)\\right),", + "image_path": "29392d3c8e878d24b83758090b672d434e10fc767fe549a73b3d63a685ff2696.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "inline_equation", + "content": "F_{offset2}" + }, + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "text", + "content": " is another network. These offsets enable the model to adaptively compensate for camera parameter errors " + }, + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "inline_equation", + "content": "e_P" + }, + { + "bbox": [ + 313, + 188, + 555, + 236 + ], + "type": "text", + "content": ", improving geometric accuracy while preserving semantic causal locality under such errors." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 241, + 447, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 447, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 447, + 251 + ], + "type": "text", + "content": "3.4.3. Normalized Convolution" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "text", + "content": "Prior work [28] notes that the direct mapping in LSS yields sparse 3D features. To address this, an intuitive solution is to use local feature propagation operators (e.g., convolutions) with causal loss supervision for preserving semantic causality. However, vanilla convolutions lack gradient constraints during causal loss computation, producing poor results. We address this by normalizing convolution weights to keep gradient map values in [0, 1]. Given the difficulty of normalizing standard convolution weights, we follow MobileNet and ConvNext to adopt depthwise (spatial) and pointwise (channel) decomposition. For the depthwise kernel " + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{spatial}} \\in \\mathbb{R}^{3 \\times 3 \\times 3 \\times C}" + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "text", + "content": ", we apply softmax across spatial dimensions " + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "inline_equation", + "content": "(h, w, z)" + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "text", + "content": " per channel " + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 255, + 555, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 417, + 553, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 417, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 318, + 417, + 553, + 440 + ], + "type": "interline_equation", + "content": "W _ {\\text {s p a t i a l}} ^ {\\prime} [ h, w, z, c ] = \\frac {\\exp \\left(W _ {\\text {s p a t i a l}} [ h , w , z , c ]\\right)}{\\sum_ {h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime}} \\exp \\left(W _ {\\text {s p a t i a l}} \\left[ h ^ {\\prime} , w ^ {\\prime} , z ^ {\\prime} , c \\right]\\right)}. \\tag {20}", + "image_path": "108036d881cbdd351367fbd7498c85bdd90d40d9230f9200ec15b3aad91ce108.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 446, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 553, + 471 + ], + "type": "text", + "content": "For the pointwise kernel " + }, + { + "bbox": [ + 313, + 446, + 553, + 471 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{channel}} \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 313, + 446, + 553, + 471 + ], + "type": "text", + "content": ", we apply softmax across input channels per output channel:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 326, + 475, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 475, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 326, + 475, + 553, + 506 + ], + "type": "interline_equation", + "content": "W _ {\\text {c h a n n e l}} ^ {\\prime} \\left[ c _ {\\text {i n}}, c _ {\\text {o u t}} \\right] = \\frac {\\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} \\right]\\right)}{\\sum_ {c _ {\\text {o u t}} ^ {\\prime}} \\exp \\left(W _ {\\text {c h a n n e l}} \\left[ c _ {\\text {i n}} , c _ {\\text {o u t}} ^ {\\prime} \\right]\\right)}. \\tag {21}", + "image_path": "8ba34291a8765349670d01b2c403dbc4e71f7d32695f9f1bc6e20d6beaf1189e.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": "Specifically, we use transposed convolution for depthwise convolutions. It diffuses features from non-zero to zero positions, but the zero position does not affect others. We prove in the supplement that the derived gradient mask remains within [0, 1] even with our semantic causality-aware 2D-to-3D transformation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 588, + 524, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 524, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 524, + 601 + ], + "type": "text", + "content": "3.5. Validation of Gradient Error Mitigation" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Fig. 5 shows the occupancy loss curves during training for BEVDet, comparing performance with and without our method. The results show that BEVDet integrated with our approach (blue line) achieves a significantly faster and steeper loss reduction compared to the original BEVDet (red line). This empirical evidence validates our theoretical analysis of gradient error mitigation. As Theorem Theorem 1 formalizes, Depth-Based LSS methods are inherently limited by gradient deviations due to fixed 2D-to-3D" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24883" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 70, + 262, + 210 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 262, + 210 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 262, + 210 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 262, + 210 + ], + "type": "image", + "image_path": "8c7e40b5c0db6937e75f1633b2673a8b1d62dfa14ed19263daf7d4c34633847c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 295, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 259 + ], + "type": "text", + "content": "Figure 5. Occupancy Loss Curves with/without Our Method. Our method reduces training loss through semantic causal 2D-to-3D geometry transformation, as shown in the comparison before (red) and after (blue) its application." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 265, + 297, + 370 + ], + "blocks": [ + { + "bbox": [ + 56, + 265, + 297, + 370 + ], + "lines": [ + { + "bbox": [ + 56, + 265, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 297, + 370 + ], + "type": "table", + "html": "
MethodmIoU ↑DropmIoUD↑DropIoU ↑Drop
BEVDetOcc [12]37.1-32.3%30.2-49.0%70.4-8.5%
BEVDetOcc w/ Noise25.115.464.4
BEVDetOcc+Ours38.3-7.3%31.5-10.8%71.2-1.4%
BEVDetOcc+Ours w/ Noise35.528.170.2
ALoCC [7]40.1-21.9%34.3-28.6%70.2-8.4%
ALoCC w/ Noise31.324.564.3
ALoCC+Ours40.9-3.3%35.5-4.8%70.7-1.0%
ALoCC+Ours w/ Noise39.633.870.0
", + "image_path": "221af839a449acb9fb887c1b836d27fe4f4cf7e5bd09b236ac2b8d06ba334b47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 296, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 431 + ], + "type": "text", + "content": "Table 2. Performance Comparison on the Occ3D Dataset with Gaussian Noise Added to Camera Parameters. The \"Drop (%) columns show degradation, with our methods (BEVDetOcc+Ours, ALOcc+Ours) achieving much smaller drops (e.g., -7.3% mIoU vs. -32.4% for BEVDetOcc)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 510 + ], + "type": "text", + "content": "mapping errors. In contrast, our method aims to alleviate this by enabling a learnable mapping and incorporating a causal loss. It indicates that by mitigating gradient error through semantic causality-aware 2D-to-3D transformation, our approach facilitates more efficient gradient-based learning, leading to faster convergence and a lower loss curve." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 133, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 133, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 133, + 534 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 172, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 172, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 172, + 553 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Dataset. In this study, we leverage the Occ3D-nuScenes dataset [5, 39], a comprehensive dataset with diverse scenes for autonomous driving research. It encompasses 700 scenes for training, 150 for validation, and 150 for testing. Each scene integrates a 32-beam LiDAR point cloud alongside 6 RGB images, captured from multiple perspectives encircling the ego vehicle. Occ3D [39] introduces voxel-based annotations, covering a spatial extent of " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "-40\\mathrm{m}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "40\\mathrm{m}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " along the X and Y axes, and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "-1\\mathrm{m}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "5.4\\mathrm{m}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " along the Z axis, with a consistent voxel size of " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " across all dimensions. Occ3D delineates 18 semantic categories, comprising 17 distinct object classes plus an empty class to signify unoccupied regions. Following [7, 27, 35, 39], we evalu" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 127 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 127 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 127 + ], + "type": "image", + "image_path": "6a25e7a9ca91d4b12c60c8b544975b4dae0d485f86571abe220867e095cae644.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 129, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "image", + "image_path": "36e3fcbb843c288828c0a7573947a8a3e5ddecb152ab327f76a7f277a3d5dfac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 186, + 432, + 241 + ], + "blocks": [ + { + "bbox": [ + 318, + 186, + 432, + 241 + ], + "lines": [ + { + "bbox": [ + 318, + 186, + 432, + 241 + ], + "spans": [ + { + "bbox": [ + 318, + 186, + 432, + 241 + ], + "type": "image", + "image_path": "0136db64c9a04fd9b4b8a19aa6469c5da59aba4c623a0ecb0d5af6e88d9e22bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 243, + 396, + 253 + ], + "lines": [ + { + "bbox": [ + 352, + 243, + 396, + 253 + ], + "spans": [ + { + "bbox": [ + 352, + 243, + 396, + 253 + ], + "type": "text", + "content": "(a) BEVDet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 434, + 186, + 553, + 241 + ], + "blocks": [ + { + "bbox": [ + 434, + 186, + 553, + 241 + ], + "lines": [ + { + "bbox": [ + 434, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 434, + 186, + 553, + 241 + ], + "type": "image", + "image_path": "fc7f33cc3030f9bbcf9ee9c392bfbdbb23a50a0f4c906b31ae2a168ac094e284.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 243, + 528, + 253 + ], + "lines": [ + { + "bbox": [ + 457, + 243, + 528, + 253 + ], + "spans": [ + { + "bbox": [ + 457, + 243, + 528, + 253 + ], + "type": "text", + "content": "(b) BEVDet + Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 256, + 555, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 334 + ], + "type": "text", + "content": "Figure 6. Visualization of 2D-to-3D Semantic Causal Consistency Using LayerCAM [18]. We enhance BEVDet with our method for comparison. Attention maps are computed for critical traffic classes: \"traffic cone\", \"pedestrian\", and \"car\". Areas of greatest difference are marked with boxes. Each class-specific localization highlights our method's precise focus over vanilla BEVDet, showing improved semantic alignment." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "content": "ate the occupancy prediction performance with mIoU across 17 semantic object categories, " + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "inline_equation", + "content": "\\mathrm{mIoU}_D" + }, + { + "bbox": [ + 313, + 342, + 554, + 378 + ], + "type": "text", + "content": " for 8 dynamic categories, and occupied/unoccupied IoU for scene geometry." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": "Implementation Details. We integrate our approach into BEVDet [12] and ALOcc [7] for performance evaluation. The model parameters, image, and BEV augmentation strategies are retrained as the original. For ALOcc, we remove its ground-truth depth denoising module to adapt to our method. We use multiple convolutional layers to predict the two camera parameter offsets. The loss weight of " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "L_{\\text{causal}}" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": " is set to 0.02 in the main experiments. We optimize using AdamW [33] with a learning rate of " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": ", a global batch size of 16, and for 24 epochs. Experiments use single-frame surrounding images, emphasizing improvements from enhanced 2D-to-3D transformations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 528, + 553, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 553, + 540 + ], + "type": "text", + "content": "4.2. Evaluation of Camera Perturbation Robustness" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": "To assess our method's robustness under extreme noise, we add Gaussian noise with (0.1 variances) to camera parameters in training and testing, as shown in Tab. 2. Compared to vanilla BEVDetOcc and ALOcc, our methods show smaller performance drops. BEVDetOcc+Ours has a mIoU drop of " + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "inline_equation", + "content": "-7.3\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "inline_equation", + "content": "-32.3\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": " for BEVDetOcc, and ALOcc+Ours drops " + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "inline_equation", + "content": "-3.3\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "inline_equation", + "content": "-21.9\\%" + }, + { + "bbox": [ + 313, + 545, + 554, + 653 + ], + "type": "text", + "content": " for ALOcc, demonstrating enhanced resilience to noisy parameters. This effectively counters motion-induced errors, benefiting self-driving tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 660, + 492, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 492, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 492, + 673 + ], + "type": "text", + "content": "4.3. Semantic Causality Visualization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As shown in the Fig. 6, we visualize 3D-to-2D semantic causal consistency using LayerCAM [18]. We backpropagate the final 3D semantic occupancy predictions of dis" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24884" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 215 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 215 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 215 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 215 + ], + "type": "table", + "html": "
MethodBackboneInput SizemIoUmIoUDIoU
MonoScene [6]ResNet-101928 × 16006.15.4-
CTF-Occ [39]ResNet-101928 × 160028.527.4-
TPVFormer [14]ResNet-101928 × 160027.827.2-
COTR [35]ResNet-50256 × 70439.133.869.6
ProtoOcc [19]ResNet-50256 × 70439.634.3-
LightOcc-S [50]ResNet-50256 × 70437.932.4-
DHD-S [45]ResNet-50256 × 70436.530.7-
FlashOCC [48]ResNet-50256 × 70432.024.765.3
FB-Occ [27]ResNet-50256 × 70435.730.966.5
BEVDetOcc [12]ResNet-50256 × 70437.130.270.4
BEVDetOcc+OursResNet-50256 × 70438.3 ↑1.231.5 ↑1.371.2 ↑1.2
ALOcc [7]ResNet-50256 × 70440.134.370.2
ALOcc+OursResNet-50256 × 70440.9 ↑0.835.5 ↑1.170.7 ↑0.5
", + "image_path": "20d18994ba23a7982a510eb6e699ce1cc1597209f8aaf830df71988bd7982675.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "content": "tinct classes to the 2D feature maps fed into SCAT. Notably, we are the first to apply LayerCAM for cross-dimensional analysis. Fig. 6 (b) shows our method precisely focuses on class-specific locations. This confirms LayerCAM's cross-dimensional effectiveness. Our approach surpasses vanilla BEVDet (Fig. 6 (a)) in targeting class-associated objects. This proves improved semantic localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 258, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 258, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 258, + 395 + ], + "type": "text", + "content": "4.4. Benchmarking with Previous Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "text", + "content": "As shown in Tab. 3, we compare our method with leading 3D semantic occupancy prediction approaches on Occ3D [39]. Specifically, compared to baseline models BEVDet and ALOcc, our method achieves significant improvements in mIoU, mIoU" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "inline_equation", + "content": "_{\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "text", + "content": ", and IoU. For instance, the BEVDetOcc+Ours variant achieves an mIoU of 38.3, surpassing BEVDetOcc [12] by 1.2, while improving mIoU" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "inline_equation", + "content": "_{\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "text", + "content": " by 1.3 and IoU by 1.2. Similarly, ALOcc+Ours shows gains of 0.8 in mIoU, 1.1 in mIoU" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "inline_equation", + "content": "_{\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 533 + ], + "type": "text", + "content": ", and 0.5 in IoU over ALOcc [7]. These results validate the superiority of our semantic causality-aware 2D-to-3D transformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 149, + 553 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Effect of Causal Loss. We first investigate the effectiveness of the proposed Causal Loss in enhancing the occupancy prediction performance. As demonstrated in Tab. 4, the impact of the proposed Causal Loss is evaluated through a series of experiments. Specifically, using BEVDetOcc as the baseline (Exp. 0), we conducted two ablation studies: one removing the depth supervision loss (Exp. 1), and another incorporating the proposed Causal Loss (Exps. 2, 3). The results reveal that the removal of the depth supervision loss leads to marginal decrease in performance, whereas the addition of the proposed Causal Loss yields a significant improvement. This suggests that Causal Loss facilitates superior 2D-to-3D transformation and ultimately enhances the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 553, + 199 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 295, + 280 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 280 + ], + "type": "text", + "content": "Table 3. Comparison of 3D Semantic Occupancy Prediction Using Single Frame on the Occ3D Dataset, Evaluated mIoU, mIoUD, and IoU Metrics. Performance gains are indicated by red arrows " + }, + { + "bbox": [ + 55, + 224, + 295, + 280 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 224, + 295, + 280 + ], + "type": "text", + "content": ". Our proposed approach (+Ours) consistently demonstrates superior enhancement over existing methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 199 + ], + "type": "table", + "html": "
Exp.MethodmIoUDiff.mIoUDIoULatency
0Baseline (BEVDetOcc) [12]37.1-30.270.4416/125
1w/o Depth Sup36.8-0.329.670.3414/125
2+ Causal Loss37.6+0.831.070.1450/125
3+ Unbiased Estimator37.5-0.130.770.5417/125
4w/o Post Conv37.3-0.230.770.2379/122
5+ Channel-Grouped Lifting37.6+0.330.770.7419/128
6+ Soft Filling37.6-30.670.7434/149
7+ Learnable Camera Offset37.9+0.331.171.0446/150
8+ Normalized Convolution38.3+0.431.571.2466/159
", + "image_path": "aab475daf81f77cbb53c8fbf8bcdcf502b02c53e114491b7d448639be8e953b5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 205, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 262 + ], + "type": "text", + "content": "Table 4. Ablation Study of 3D Semantic Occupancy Prediction on Occ3D. We comprehensively evaluated the impact of the individual strategy (bolded rows) proposed in our paper, with BEVDetOcc as the baseline. The final column reports single-frame training/inference latency (ms) on an RTX 4090 GPU." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 304 + ], + "type": "text", + "content": "precision of 3D semantic occupancy prediction. Comparing Exp. 3 to Exp. 2, the Unbiased Estimator simplifies the Causal Loss computation, reducing training overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 496 + ], + "type": "text", + "content": "Effect of Each Module. In Tab. 4 (Exps. 5-8), we systematically validate the effectiveness of the three proposed modules. We first remove the two post-lifting convolutional layers from the original BEVDet (Exp. 4), as they serve a similar role to our proposed modules in refining volume features. Subsequently, we incrementally integrate the three proposed modules (Channel-Groupled Lifting, Learnable Camera Offsets, and Normalized Convolution) into the baseline model. To ensure gradient flow for Learnable Camera Offsets, we introduce the Soft Filling strategy from ALOcc [7] in Exp. 6, enabling effective training of the camera parameter offsets in Exp. 7. The results show progressive performance improvements with each added component (Exps. 5, 7, 8), confirming the efficacy of the proposed SCAT method. Additionally, the proposed modules incur acceptable computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 520 + ], + "type": "text", + "content": "Please refer to the supplementary material for more comprehensive experimental results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 388, + 540 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced a novel approach leveraging causal principles to address the limitation that ignored the reliability and interpretability by existing methods. By exploring the causal foundations of 3D semantic occupancy prediction, we propose a causal loss that enhances semantic causal consistency. In addition, we develop the SCAT module with three main components: Channel-Groupled Lifting, Learnable Camera Offsets and Normalized Convolution. This approach effectively mitigates transformation inaccuracies arising from uniform mapping weights, camera perturbations, and sparse mappings. Experiments demonstrate that our approach achieves significant improvements in accuracy, robustness to camera perturbations, and semantic causal consistency in 2D-to-3D transformations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24885" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9297-9307, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 200 + ], + "type": "text", + "content": "[2] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Langocc: Self-supervised open vocabulary occupancy estimation via volume rendering. arXiv preprint arXiv:2407.17310, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "text", + "content": "[3] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Occfownet: Towards self-supervised occupancy estimation via differentiable rendering and occupancy flow. arXiv preprint arXiv:2402.12792, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[4] Simon Boeder, Fabian Gigengack, and Benjamin Risse. Gaussianflowocc: Sparse and weakly supervised occupancy estimation using gaussian splatting and temporal flow. arXiv preprint arXiv:2502.17288, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621-11631, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "text", + "content": "[6] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "text", + "content": "[7] Dubing Chen, Jin Fang, Wencheng Han, Xinjing Cheng, Junbo Yin, Chengzhong Xu, Fahad Shahbaz Khan, and Jianbing Shen. Alocc: adaptive lifting-based 3d semantic occupancy and cost volume-based flow prediction. arXiv preprint arXiv:2411.07725, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 501 + ], + "type": "text", + "content": "[8] Dubing Chen, Wencheng Han, Jin Fang, and Jianbing Shen. Adaocc: Adaptive forward view transformation and flow modeling for 3d occupancy and flow prediction. arXiv preprint arXiv:2407.01436, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "type": "text", + "content": "[9] Dubing Chen, Huan Zheng, Jin Fang, Xingping Dong, Xianfei Li, Wenlong Liao, Tao He, Pai Peng, and Jianbing Shen. Rethinking temporal fusion with a unified gradient descent view for 3d semantic occupancy prediction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1505-1515, 2025. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 623 + ], + "type": "text", + "content": "[10] Xiaokang Chen, Kwan-Yee Lin, Chen Qian, Gang Zeng, and Hongsheng Li. 3d sketch-aware semantic scene completion via semi-supervised structure prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4193-4202, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[11] Judy Hoffman, Daniel A Roberts, and Sho Yaida. Robust learning with jacobian regularization. arXiv preprint arXiv:1908.02729, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "text", + "content": "[12] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. arXiv preprint arXiv:2203.17054, 2022. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Junjie Huang, Guan Huang, Zheng Zhu, Yun Ye, and Dalong Du. Bevdet: High-performance multi-camera 3d object de" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "tection in bird-eye-view. arXiv preprint arXiv:2112.11790, 2021. 1, 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2302.07817, 2023. 1, 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 195 + ], + "type": "text", + "content": "[15] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[16] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In European Conference on Computer Vision, pages 376-393. Springer, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "type": "text", + "content": "[17] Haoyi Jiang, Liu Liu, Tianheng Cheng, Xinjie Wang, Tianwei Lin, Zhizhong Su, Wenyu Liu, and Xinggang Wang. Gausstr: Foundation model-aligned gaussian transformer for self-supervised 3d spatial understanding. arXiv preprint arXiv:2412.13193, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 353 + ], + "type": "text", + "content": "[18] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "text", + "content": "[19] Jungho Kim, Changwon Kang, Dongyoung Lee, Sehwan Choi, and Jun Won Choi. Protoocc: Accurate, efficient 3d occupancy prediction using dual branch encoder-prototype query decoder. arXiv preprint arXiv:2412.08774, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 400, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 553, + 454 + ], + "type": "text", + "content": "[20] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[21] Jinke Li, Xiao He, Chonghua Zhou, Xiaogiang Cheng, Yang Wen, and Dan Zhang. Viewformer: Exploring spatiotemporal modeling for multi-view 3d occupancy perception via view-guided transformers. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[22] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. arXiv preprint arXiv:2206.10092, 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "text", + "content": "[23] Yinhao Li, Han Bao, Zheng Ge, Jinrong Yang, Jianjian Sun, and Zeming Li. Bevstereo: Enhancing depth estimation in multi-view 3d object detection with temporal stereo. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1486-1494, 2023. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "type": "text", + "content": "[24] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9087-9098, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Yangguang Li, Bin Huang, Zeren Chen, Yufeng Cui, Feng Liang, Mingzhu Shen, Fenggang Liu, Enze Xie, Lu Sheng, Wanli Ouyang, et al. Fast-bev: A fast and strong bird's" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24886" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "eye view perception baseline. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "type": "text", + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. arXiv preprint arXiv:2203.17270, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 198 + ], + "type": "text", + "content": "[27] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "type": "text", + "content": "[28] Zhiqi Li, Zhiding Yu, Wenhai Wang, Anima Anandkumar, Tong Lu, and Jose M Alvarez. Fb-bev: Bev representation from forward-backward view transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6919-6928, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 300 + ], + "type": "text", + "content": "[29] Zhimin Liao and Ping Wei. Cascadeflow: 3d occupancy and flow prediction with cascaded sparsity sampling refinement framework. CVPR 2024 Autonomous Grand Challenge Track On Occupancy and Flow, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 303, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 294, + 357 + ], + "type": "text", + "content": "[30] Lizhe Liu, Bohua Wang, Hongwei Xie, Daqi Liu, Li Liu, Zhiqiang Tian, Kuiyuan Yang, and Bing Wang. Surroundsdf: Implicit 3d scene understanding based on signed distance field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[31] Shice Liu, Yu Hu, Yiming Zeng, Qiankun Tang, Beibei Jin, Yinhe Han, and Xiaowei Li. See and think: Disentangling semantic scene completion. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "text", + "content": "[32] Yili Liu, Linzhan Mou, Xuan Yu, Chenrui Han, Sitong Mao, Rong Xiong, and Yue Wang. Let occ flow: Self-supervised 3d occupancy flow prediction. arXiv preprint arXiv:2407.07587, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 452, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 294, + 483 + ], + "type": "text", + "content": "[33] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[34] Yuhang Lu, Xinge Zhu, Tai Wang, and Yuexin Ma. Octreeocc: Efficient and multi-granularity occupancy prediction using octree queries. arXiv preprint arXiv:2312.03774, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "text", + "content": "[35] Qihang Ma, Xin Tan, Yanyun Qu, Lizhuang Ma, Zhizhong Zhang, and Yuan Xie. Cotr: Compact occupancy transformer for vision-based 3d occupancy prediction. arXiv preprint arXiv:2312.01919, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "text", + "content": "[36] Mingjie Pan, Jiaming Liu, Renrui Zhang, Peixiang Huang, Xiaoqi Li, Hongwei Xie, Bing Wang, Li Liu, and Shang-hang Zhang. Renderocc: Vision-centric 3d occupancy prediction with 2d rendering supervision. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12404-12411. IEEE, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 711 + ], + "type": "text", + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[38] Yang Shi, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Xinggang Wang. Occupancy as set of points. In Computer Vision-ECCV 2024: 18th European Conference, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "type": "text", + "content": "[39] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "text", + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 317 + ], + "type": "text", + "content": "[42] Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "type": "text", + "content": "[43] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xinggang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 554, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 419 + ], + "type": "text", + "content": "[44] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. arXiv preprint arXiv:2303.09551, 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "text", + "content": "[45] Yuan Wu, Zhiqiang Yan, Zhengxue Wang, Xiang Li, Le Hui, and Jian Yang. Deep height decoupling for precise vision-based 3d occupancy prediction. arXiv preprint arXiv:2409.07972, 2024.8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 520 + ], + "type": "text", + "content": "[46] Ziyang Yan, Wenzhen Dong, Yihua Shao, Yuhang Lu, Liu Haiyang, Jingwen Liu, Haozhe Wang, Zhe Wang, Yan Wang, Fabio Remondino, et al. Renderworld: World model with self-supervised 3d label. arXiv preprint arXiv:2409.11356, 2024.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "text", + "content": "[47] Shichao Yang, Yulan Huang, and Sebastian Scherer. Semantic 3d occupancy mapping through efficient high order crfs. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 590-597. IEEE, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 569, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 554, + 621 + ], + "type": "text", + "content": "[48] Zichen Yu, Changyong Shu, Jiajun Deng, Kangjie Lu, Zong-dai Liu, Jiangyong Yu, Dawei Yang, Hui Li, and Yan Chen. Flashocc: Fast and memory-efficient occupancy prediction via channel-to-height plugin. arXiv preprint arXiv:2311.12058, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[49] Chubin Zhang, Juncheng Yan, Yi Wei, Jiaxin Li, Li Liu, Yansong Tang, Yueqi Duan, and Jiwen Lu. Occnerf: Self-supervised multi-camera occupancy prediction with neural radiance fields. arXiv preprint arXiv:2312.09243, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[50] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Lightweight spatial embedding for vision-based 3d occupancy prediction. arXiv preprint arXiv:2412.05976, 2024.8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24887" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[51] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. arXiv preprint arXiv:2304.05316, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "text", + "content": "[52] Jilai Zheng, Pin Tang, Zhongdao Wang, Guoqing Wang, Xiangxuan Ren, Bailan Feng, and Chao Ma. Veon: Vocabulary-enhanced occupancy prediction. In European Conference on Computer Vision, pages 92-108. Springer, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "text", + "content": "[53] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. In Findings of the Association for Computational Linguistics ACL 2024, pages 15890-15902, 2024. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24888" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_content_list.json b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d21db95ba1ff7f2924938f73d48ce4d041e729bc --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "Semantic Discrepancy-aware Detector for Image Forgery Identification", + "text_level": 1, + "bbox": [ + 138, + 130, + 856, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziye Wang Minghang Yu Chunyan Xu* Zhen Cui Nanjing University of Science and Technology, Nanjing, China", + "bbox": [ + 250, + 180, + 748, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wzynjust,mhyu,cyx}@njust.edu.cn，zhen.cui@bnu.edu.cn", + "bbox": [ + 256, + 219, + 733, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of image generation techniques, robust forgery detection has become increasingly imperative to ensure the trustworthiness of digital media. Recent research indicates that the learned semantic concepts of pre-trained models are critical for identifying fake images. However, the misalignment between the forgery and semantic concept spaces hinders the model's forgery detection performance. To address this problem, we propose a novel Semantic Discrepancy-aware Detector (SDD) that leverages reconstruction learning to align the two spaces at a fine-grained visual level. By exploiting the conceptual knowledge embedded in the pre-trained vision-language model, we specifically design a semantic token sampling module to mitigate the space shifts caused by features irrelevant to both forgery traces and semantic concepts. A concept-level forgery discrepancy learning module, based on reconstruction, enhances the interaction between semantic concepts and forgery traces, effectively capturing discrepancies under the concepts' guidance. Finally, the low-level forgery feature enhancement integrates the learned concept-level forgery discrepancies to minimize redundant forgery information. Experiments conducted on two standard image forgery datasets demonstrate the efficacy of the proposed SDD, which achieves superior results compared to existing methods. The code is available at https://github.com/wzy111111/SSD.", + "bbox": [ + 89, + 301, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 724, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the thriving of generative AI technologies, like Generative Adversarial Networks (GANs) [14] and diffusion models [2], the images generated by these models can easily create confusion by passing off the spurious as genuine. Therefore, it is crucial to develop a universal method for detecting fake images to mitigate the widespread dissemination of disinformation.", + "bbox": [ + 89, + 751, + 482, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pioneering research [26, 32] has shown that projecting", + "bbox": [ + 109, + 858, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebad5eb90f3ca9bb87c11fd2b66f803687de28196141364d2508083b6c397c5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 266, + 888, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81954e406961bea5aa5a7ed67c1b28aa41cf3ad135b9f7908fb60b5698e8f433.jpg", + "image_caption": [ + "Figure 1. The phenomenon of misalignment between semantic concept space and forgery space. Since $\\cos \\theta$ can reflect the similarity of image descriptions, we model the feature space in polar coordinates. As the semantic concept space in [32] is frozen, fake samples sharing similar concepts with real ones can be easily misclassified. With forgery-adaptive space like [26], the model can correctly distinguish between them based on re-learned forgery features. Nevertheless, due to the semantic concept bias introduced by coarse text prompts, the target samples may be projected into an inaccurate semantic concept dimension, causing them to drift away from the real source samples along the fake dimension." + ], + "image_footnote": [], + "bbox": [ + 534, + 337, + 883, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images in a joint embedding space of texts and images can effectively capture discrepancies between fake and real images. In contrast, methods [6, 13, 44, 50] overlooking the interplay between forgery traces and semantic concepts perform poorly when confronted with unseen generative models.", + "bbox": [ + 511, + 718, + 903, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To investigate the visual semantic concepts of pretrained models, we conduct a statistical analysis of the output features from CNNSpot [50] and CLIP-ViT [32] (See Appendix A for more details). Under different categories, CNNSpot exhibits a synchronized difference between real and fake features in its training space. However, when tran", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3211c3b1e370bfa02685bd229d7de6b14fbe69599dd492db2cb33e55aba0b7cb.jpg", + "image_caption": [ + "Figure 2. Different paradigms of image forgery identification with pre-trained vision-language model. (a) Fine-tune the frozen model only by fully connected (FC) layers [32]. (b) Prompt-based designs are tuned on text prompts and contrastive objectives [26]. (c) Our paradigm incorporating visual clues can capture fine-grained forge traces by reconstruction learning." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 465, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sitioning to the CLIP's space, these differences become inconsistent. From this, we infer a nuanced relationship between semantic concepts and forgery traces: Different semantic concepts may guide the model to uncover distinct forgery traces.", + "bbox": [ + 89, + 338, + 483, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intuitively, relying on a frozen pre-trained vision-language model like UnivFD [32] is essential, but this tends to overlook fine-grained forgery details. Although Fat-Former [26] achieves a substantial enhancement in generalization by employing the forgery-aware adaptive transformer, we observe that soft prompts based on simple [CLASS] embeddings have an intrinsic limitation in their semantic description granularity (See Appendix B for more details). The constrained breadth of the conveyed concepts may lead the detection toward incorrect predictions. This limitation highlights a misalignment between the visual semantic concept space and the target forgery space, as illustrated in Fig. 1.", + "bbox": [ + 89, + 415, + 483, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, one empirical approach is to design more detailed text descriptions, but this method struggles to describe all visual forgery details due to the limited length of texts and brings more computational overhead. Drawing from the aforementioned findings and analysis, we make a first attempt to align the CLIP's visual semantic concept space with the forgery space by reconstructing semantic features.", + "bbox": [ + 89, + 612, + 483, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop a vision-based paradigm, as outlined in Fig. 2. First, employing a pre-trained model only with nearest neighbor or linear probing (e.g. UnivFD [32], Fig. 2 (a)) is suboptimal for image forgery detection. Second, modifying the pre-trained model with task-specific prompts (e.g. FatFormer [26], Fig. 2 (b)) may favor models biased towards any particular semantic concept. These studies pave the way for exploring pre-trained space with rich semantic concepts. Inspired by image reconstruction [43, 53], our paradigm amplifies the concept-level forgery discrepancies of forgery images, which empowers the model to detect sus", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "picious forgery traces with the assistance of semantic concepts.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present a novel Semantic Discrepancy-aware Detector (SDD) to accurately align the semantic concept space and the forgery space. Firstly, to mitigate interference from features unrelated to learned semantic concepts and forgery traces, we divide the real images into non-overlapping blocks and feed them to the frozen CLIP [36] to obtain diverse semantic patch tokens. These tokens acting as visual clues smoothly align the semantic concepts' space and forgery space. It is noteworthy that these tokens sampled by JS divergence are universally representative of the real semantic distribution. Then, the visual clues are fused into a concept-level forgery discrepancy module. Unlike FatFormer, LoRA layers are incorporated into the image encoder. The goal is to preserve the completeness and diversity of the learned semantic concepts of CLIP, while the forgery features sharing similar semantic concepts should be highlighted. During reconstruction, we only narrow the reconstruction gap for real samples to reinforce the reconstructed discrepancies of the synthetic images. Finally, we present low-level forgery feature enhancement to let the reconstruction difference map enhance the extraction of the highly generalizable forgery features while introducing minimal additional parameters. The main challenge is how to capture forgery features with strong semantic concept correlation and features with high forgery relevance but weak semantic concept ties to ensure the model converges to powerful features. Motivated by this, we apply convolutional modules and adaptive weight parameters to avoid over-relying on semantic concepts.", + "bbox": [ + 511, + 121, + 906, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We thoroughly evaluate the generalization performance of our model on a UnivFD benchmark [32] and a SynRIS benchmark [5]. Surprisingly, our method achieves superior performance by a $ap_{m}$ of 98.51% and a $acc_{m}$ of 93.61% on the UnivFD benchmark [32] and an average AUROC of 95.1% on the SynRIS benchmark [5]. In summary, our contributions are as follows:", + "bbox": [ + 511, + 559, + 908, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a robust model (SDD) for forgery detection, specifically designed to align the visual concept space and forgery space in terms of visual information.", + "- We sample semantic tokens to mitigate the space shifts and align the two spaces through reconstruction learning. Additionally, we strengthen low-level forgery features to enhance the model's robustness.", + "- Our method achieves superior performance on two benchmarks, demonstrating its superior capability in comparison to existing approaches." + ], + "bbox": [ + 511, + 665, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 829, + 653, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AI-generated Images Detection. Extensive efforts have been devoted to enhancing the performance of AI-generated image detection. Early works like [25, 44, 45] tend to mine", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18389", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94dd20070c1e79972c6f8254b876b80beacef33b3db35c97c2297de23b51f24e.jpg", + "image_caption": [ + "Figure 3. The architecture of SDD. First, Next, we sample semantic tokens from real images to learn features related to both concepts and forgery. the input images are mapped into a joint space of visual semantic concepts and forgery, which are transformed into learnable features $V_{H}$ . Then, we use transformer-based encoder and decoder to get reconstructed features $\\mathcal{R}_f$ . A reconstruction difference map $\\mathcal{D}_S$ is obtained and goes through the multi-scale convolutional network to refine forgery features. Finally, we concatenate the CLIP's CLS token with this output along the same dimension for classification. The whole system is trained by jointly minimizing the binary cross-entropy loss $L_{bce}$ , the reconstruction loss $L_{r}$ , and the triplet loss $L_{tri}$" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the common forgery traces between all real and fake images, such as noise patterns, texture statistics, and frequency signals. As an illustration, Liu et al. [24] designed a network that learns the consistent noise patterns in images for fake detection. Liu et al. [28] proposed to leverage the gram matrix to discover the global anomalous texture of fake images. An effective approach [13] demonstrated that frequency representation is an important factor in improving fake detection performance. However, these differences are rigorously specific to the monotonous features, which contribute to the issue of overfitting. Cutting-edge research [26, 32] shifted attention toward the semantic properties of images. Ojha et al. [32] showed that projecting images into the feature space of pre-trained vision-language model enables strong generalization ability. To build generalized forgery representations, Liu et al. [26] constructed forgery adaptive space by a forgery-aware adapter. The above research [5, 26, 32] has suggested that concept attributes are vital in the image forgery detection task. Assuming that diffusion-based models leave distinct forgery traces that are characteristic of specific concept distributions, we aim to extract robust forgery features guided by semantic concepts, rather than suppressing them. Therefore, even \"useless\" information can be useful by providing significant certainty about the content of the image.", + "bbox": [ + 91, + 494, + 483, + 872 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reconstruction Learning. Reconstruction learning has", + "bbox": [ + 89, + 873, + 482, + 888 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "great potential in unsupervised representation learning [16, 27]. Some works [5, 39] utilized reconstruction learning to reveal the nuances between real and fake images. For example, Wang et al. [51] found that reconstructing images by DDIM exposes an error between real images and their reconstructed replica. The new synthetic image detection method[5] used text-conditioned inversion maps to learn internal representations, which is conducive to predicting whether an image is fake. Ricker et al. [39] offered a simple detection approach by applying AE to measure the reconstruction error. Notably, these works are committed to reconstructing the distributions of both real and fake samples by leveraging generative models. Unlike previous works, we focus solely on reconstructing real images in the finetuned CLIP space in light of the authenticity and richness of semantic concepts. The distribution of real samples, learned from pre-trained vision-language model, helps to define an optimal boundary, thus alleviating overfitting.", + "bbox": [ + 511, + 494, + 908, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 782, + 648, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to align forgery and visual semantic concept spaces using reconstruction techniques for robust and generalizable synthetic image detection. To achieve this, we introduce a fine-grained model named Semantic Discrepancy-aware Detector (SDD). Building on prior works, we harness the generalization capability of vision-language mod", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18390", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "els. Semantic concept space: The ideal joint embedding space of images and texts with four properties: semantic alignment, modality invariance, locality consistency, and structure preservation. Forgery space: The ideal space covers forgery traces. Notably, we derive semantic concept space via a vision language model pretrained solely on real images; thus we treat the two spaces as independent.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the Semantic Tokens Sampling (STS) module utilizes Jensen-Shannon (JS) divergence to sample semantic patch tokens, which serve as a transitional bridge, facilitating model to establish the association between real and forgery images accurately. Next, the Concept-level forgery Discrepancy Learning (CFDL) module employs reconstruction learning to explore the forgery discrepancies within the visual semantic concept space, which focuses on identifying subtle variations between reconstructed forgery features. Finally, the reconstruction difference map is trained with Low-level forgery Feature Enhancement module, which aims to refine forgery features with more visual details.", + "bbox": [ + 89, + 198, + 482, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Semantic Tokens Sampling", + "text_level": 1, + "bbox": [ + 89, + 387, + 334, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, we considered directly aligning the visual semantic concept space and forgery space by leveraging fine-grained reconstruction learning to model real and fake semantic distributions. However, this strategy would treat the differences in features unrelated to semantics and forgery as crucial factors for identifying image's authenticity. To eliminate these redundant features, we sample real semantic image patch tokens as visual clues to bridge real and forged semantic domains. This module enables the model to focus on concept-related forgery traces and highlight the distinctions between real and fake images. In a tangible way, the image encoder of CLIP: ViT-L/14 is adapted to transform a real image $x_{r}$ into a set of features $f_{r}$ , without the image CLS token. We define the transformation as:", + "bbox": [ + 89, + 409, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {r} = \\phi (x _ {r}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 632, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi (\\cdot)$ is the CLIP:ViT-L/14's visual encoder, $x_{r}\\in$ $\\mathbb{I}_r^{H\\times W\\times 3}$ represents a real image characterized by a height of $H$ and a width of $W$ . Besides, $f_{r}\\in \\mathbb{R}^{N\\times D}$ , where $N$ is the number of tokens and $D$ denotes the dimension of each patch token.", + "bbox": [ + 89, + 657, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since integrating all real patch tokens into the image reconstruction module is computationally intensive and memory-consuming, it is urgent to select a suitable subset of these tokens. From a distribution perspective, the Jensen-Shannon (JS) divergence, derived from the Kullback-Leibler divergence [48], is a symmetric and finite metric that can effectively measure the similarity between tokens by quantifying differences in their distributions.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To calculate the JS divergence between two tokens, both are converted into computable probability distribution space using the softmax function. Let $f_{s} \\in \\mathbb{F}_{r}^{M \\times D}$ be the selected", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "semantic patch tokens with the num of tokens $M = 1 / \\delta$ and dimension $D$ in terms of sampling rate $\\delta$ ( $0 \\leq \\delta \\leq 1$ , $\\delta$ is user-defined). Once the initial token $\\tilde{r}$ and $\\delta$ are determined, the JS divergence between $\\tilde{r}$ and other tokens $r$ falls within the range [0, 1]. Subsequently, the sampling interval is split into $M$ equal segments with one token selected from each segment. As a consequence, the semantic tokens sampling module can be formulated as:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {s} = \\mathcal {S} \\left(\\mathbb {R} ^ {N \\times D}, \\delta\\right) \\\\ = A _ {c} ^ {N _ {a} \\times M} \\times \\mathbb {R} ^ {N \\times D}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 214, + 707, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "s.t. JS(softmax $(\\tilde{r})$ ,softmax $(r)) = \\frac{i}{M}$ , if $a_{ij} = 1$ (2)", + "bbox": [ + 529, + 256, + 903, + 290 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i = 0} ^ {N _ {a}} \\sum_ {j = 0} ^ {M} a _ {i j} = M; \\sum_ {j = 0} ^ {M} a _ {i j} = 1 \\text {o r} 0,\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 287, + 792, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni = 1, \\dots , N _ {a}; j = 1, \\dots , M,\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 333, + 759, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S(\\cdot, \\cdot)$ represents the sampling process. $A_{c}^{N_{a} \\times M}$ is a constraint matrix of size $N_{a} \\times M$ whose element $a_{ij}$ is constrained to the binary pattern of $\\{0, 1\\}$ . Here JS $(\\cdot)$ refers to the Jensen-Shannon divergence, $N_{a}$ denotes the total number of real image patch tokens sampled from the training dataset of UnivFD and $M$ represents the required subset size. The softmax $(\\cdot)$ is the softmax function. The sampling tokens help the reconstruction module avoid becoming biased towards any particular forgery-unrelated distribution. Meanwhile, it avoids the semantic bias often introduced by text prompts, since the tokens are evenly distributed in a unified CLIP space.", + "bbox": [ + 511, + 357, + 906, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Concept-level Forged Discrepancy Learning", + "text_level": 1, + "bbox": [ + 511, + 546, + 885, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A few words alone can hardly paint a picture. We argue that the fine-grained visual details can uncover more forgery traces concealed in the images. As such, we mix sampling tokens with extracted features and capitalize on reconstruction learning to compensate for the omission of forgery traces brought by coarse prompts. As previous work [26] has demonstrated that the pre-trained vision-language model necessitates fine-tuning to adapt to the forgery detection task. Therefore, we integrate LoRA [17] with the CLIP-ViT model to capture discriminative forgery features by making use of the bread semantic concepts. This method, denoted as LoRA-CLIP [54], is more streamlined and flexible. Given an input image $\\mathcal{I} \\in \\mathbb{I}^{H \\times W \\times 3}$ , we can get high-level visual features $V_{H}$ , as follows:", + "bbox": [ + 511, + 568, + 906, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {H} = \\mathcal {F} _ {L o R A} (\\mathcal {I}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 786, + 903, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathcal{F}_{LoRA}$ refers to the CLIP image encoder fine-tuned by LoRA. The reconstruction module of CFDL encompasses two submodules, i.e., transformer-based encoder and decoder. Thanks to the transformer's capability of long-range relationship modeling, we capitalize on the multi-head attention (MHA) mechanism, the core mechanism of", + "bbox": [ + 511, + 809, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the transformer, to obtain more discriminative distribution by utilizing contextual information, which is set as:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{r l} \\operatorname {h e a d} _ {i} & = \\operatorname {A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\\\ & = \\operatorname {S o f t m a x} \\left(\\frac {Q W _ {i} ^ {Q} \\left(K W _ {i} ^ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right) V W _ {i} ^ {V}, \\\\ \\operatorname {M H A} (Q, K, V) & = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 131, + 482, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q$ (Query), $K$ (Key), $V$ (Value) refer to the input, $W_{i}^{Q}, W_{i}^{K}, W_{i}^{V}$ separately denote the corresponding weights of linear projection, $\\mathrm{Attn}(\\cdot)$ denotes the function of the scaled dot product, Softmax is the softmax function, $d$ refers to the dimension of input, Concat ( $\\cdot$ ) represents the concatenation used to stitch the discrete attention outputs of head $1 \\sim h$ together.", + "bbox": [ + 89, + 227, + 483, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To amplify the discrepancy between a fake image and its reconstructed counterpart, the sampled visual clues are employed for the initial processing by the encoder. The encoder's process can be formulated as follows:", + "bbox": [ + 89, + 333, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(f _ {s}, V _ {H}, V _ {H}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 406, + 480, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(R _ {1}, V _ {H}, V _ {H}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 425, + 480, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{LN}(\\cdot)$ denotes the Layer Normalization. Then, the encoder's outputs used as queries are injected into the decoder to get the final reconstructed features, which are similar to the encoder process and perform the following operation:", + "bbox": [ + 89, + 453, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {3} = \\mathrm {L N} (\\mathrm {M H A} (R _ {2}, R _ {2}, R _ {2})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 541, + 480, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {e} = \\mathrm {L N} (\\mathrm {M H A} \\left(R _ {1}, R _ {3}, R _ {3}\\right)). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 560, + 480, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the reconstruction process, we just calculate the reconstruction loss $\\mathcal{L}_r$ between the real input features and their reconstructed counterparts $\\mathcal{R}_e$ within a mini-batch as follows:", + "bbox": [ + 89, + 588, + 483, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\operatorname {M S E} \\left(R _ {e}, V _ {H}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 645, + 482, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{MSE}(\\cdot ,\\cdot)$ is mean squared error. Facilitating $\\mathcal{L}_r$ encourages preserving the completeness and richness of the visual semantic concept space and highlighting the concept-related forgery features. Given the reconstructed features $R_{f}$ and the original feature $f_{r}$ , the reconstruction difference map can be formally expressed as:", + "bbox": [ + 89, + 691, + 483, + 784 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {s} = \\left| R _ {f} - f _ {r} \\right|, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 795, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\cdot |$ denotes the absolute value function.", + "bbox": [ + 89, + 823, + 390, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Low-level Forgery Feature Enhancement", + "text_level": 1, + "bbox": [ + 89, + 848, + 441, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing methods based on pre-trained vision-language models [26, 32] overlook the importance of concept", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de568ffeb00d073298b5f7dc1c0e06b896ff4b86f32ec8d1e4c74be043a2f038.jpg", + "image_caption": [ + "Figure 4. The curve of exponential inverse. In the \"fast\" interval, the value drops sharply. In the \"low\" interval, the curve flattens out, showing a decay towards 0." + ], + "image_footnote": [], + "bbox": [ + 584, + 90, + 836, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weakly-related features. We believe that a thorough alignment between the visual semantic concept space and the forgery space should include the exploration of concept weakly-related forgery features. To eliminate redundant forgery features, we come up with a novel feature enhancement that refines low-level forgery features. Empowered by the reconstruction difference map, our detector orchestrates the extraction of multi-scale features with exceptional robustness and markedly enhanced effectiveness. As shown in Fig. 3, the enhancer follows the typical architecture of a convolutional network. It involves the repeated application of convolutions, each followed by a batch normalization (BN) and a rectified linear unit (ReLU). For a given stage $n$ , $F(n)$ ( $N = 1,2,3$ ) corresponds to its output features. Then, We deconvolve the semantic difference map $D_{s}$ to the shape same as $F(N)$ and perform pixel-wise multiplication with $F(n)$ to get $F'(n)$ as:", + "bbox": [ + 511, + 306, + 906, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\prime} (n) = \\operatorname {d e c o n v} (F (n)) \\otimes D _ {s}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 575, + 903, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ is the element-wise multiplication, deconv $(\\cdot)$ represents deconvolution operation and $\\mathcal{F}'(n)$ is the low-level feature aggregated with semantic information. To further enhance the reliability of the extracted features, we compute an adaptive weight coefficient $\\frac{1}{e_n}$ to indicate the importance of $D_{s}$ to $F(n)$ :", + "bbox": [ + 511, + 606, + 905, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{e _ {n}} = \\frac {1}{e ^ {\\left| F ^ {\\prime} (n) - F (n) \\right|}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 707, + 903, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we explain the role of the exponential inverse through Fig. 4. As $x$ grows large, the curve of $e^x$ becomes flatter. Therefore, in the \"fast\" interval, forgery features with a significant divergence from the semantic difference map will be assigned smaller weights, which mobilizes the network to capture concept strongly-related features. However, in the \"low\" interval, features strongly associated with forgery can avoid being misguided by semantic concepts, which indicates that the order of importance is reversed. Next, we have the attended output features $F_{low}$", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18392", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87d0f709d3c5beebf16fa5d4102d64bc12877c31ade452d4291d5c56b5d7b629.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDallemAP
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200Steps200w/cfg100Steps100w/ CFG10027502710
CNN-SpotCVPR2020100.093.4784.5099.5489.4998.1589.0273.7559.4798.2498.4073.7270.6271.0070.5480.6584.9182.0770.5983.58
PatchForECCV202080.8872.8471.6685.7565.9969.2576.5576.1976.3474.5268.5275.0387.1086.7286.4085.3783.7378.3875.6777.73
Co-occurrenceElect.Imag.99.7480.9550.6198.6353.1167.9959.1468.9860.4273.0687.2170.2091.2189.0292.3989.3288.3582.7980.9678.11
Freq-specWIFS201955.39100.075.0855.1166.08100.045.1847.4657.1253.6150.9857.7277.7277.2576.4768.5864.5861.9267.7766.21
DireICCV2023100.083.5981.5096.5081.7099.8895.7362.5169.9897.3198.6279.5375.5273.4276.4586.2889.0088.3451.3583.54
UnivFDCVPR2023100.099.4699.5997.2499.9899.6082.4561.3279.0296.7299.0087.7799.1492.1599.1794.7495.3494.5797.1593.38
NPRCVPR2024100.099.5096.5099.8096.80100.092.2073.1078.7087.2064.8065.8099.8099.8099.8099.7099.8099.8098.6092.19
FatFormerCVPR2024100.0100.099.9899.75100.0100.097.9997.9481.2399.8499.9391.9299.8399.2299.8999.2799.5099.3399.8498.18
Ours100.099.7799.9399.4899.9899.9797.2397.9193.1099.7999.9692.0699.8898.9599.9298.0698.2997.7399.8198.51
", + "bbox": [ + 94, + 88, + 906, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ce8a5a892e90254564779c552ba0bbf2b349de34e09a0b5faf6cd58b3ff19c3.jpg", + "table_caption": [ + "Table 1. Average precision comparisons with different methods on the UnivFD dataset. We replicate the results of CNNSpot, Patchfor, Co-occurrence, Freq-spec, and UnivFD from the paper[32]. In addition, we obtained the results for Dire, NPR and FatFormer using either the official pre-trained models or our re-implemented versions. Red and underline indicates the best and the second best result, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDalleAvg-acc
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200 Steps200 wcfg100 Steps100 2750 27100 10
CNN-SpotCVPR202099.9985.2070.2085.7078.9591.7053.4766.6748.6986.3186.2660.0754.0354.9654.1460.7863.8065.6655.5869.58
PatchForECCV202075.0368.9768.4779.1664.2363.9475.5475.1475.2872.3355.3067.4176.5076.1075.7774.8173.2868.5267.9171.24
Co-occurrenceElect.Imag.97.7063.1553.7592.5051.1054.7057.1063.0655.8565.6565.8060.5070.7070.5571.0070.2569.6069.9067.5566.86
Freq-specWIFS201949.9099.9050.5049.9050.3099.7050.1050.0048.0050.6050.1050.9050.4050.4050.3051.7051.4050.4050.0055.45
DireICCV202399.8673.4760.6872.3965.1593.6088.8652.7856.3990.0794.0561.0559.3559.9560.6569.3072,7071.0052.7571.19
UnivFDCVPR2023100.098.5094.5082.0099.5097.0066.6063.0057.5059.5072.0070.0394.1973.7694.3679.0779.8578.1486.7881.38
NPRCVPR202499.8092.0089.5096.3087.6099.7079.4061.4070.6074.5057.1055.2397.4098.7097.9097.0097.9097.0088.8086.20
FatFormerCVPR202499.8999.3699.5097.1299.4399.7593.2581.3968.0469.4769.4776.0098.5594.8598.6094.3094.6094.1598.7090.86
Ours99.8895.7696.7098.0898.4699.1791.8283.6177.4595.4096.4779.5598.0594.6098.2592.2093.3591.8098.0093.61
", + "bbox": [ + 93, + 324, + 906, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Accuracy comparisons with different methods on the UnivFD dataset.", + "bbox": [ + 263, + 506, + 730, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by the residual connection:", + "bbox": [ + 89, + 530, + 272, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l o w} (n) = F ^ {\\prime} (n) + \\frac {F ^ {\\prime} (n)}{e _ {n}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 551, + 480, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For optimizing the anchor features $\\tilde{f}_a$ of the enhancer, the following triplet loss [35] is employed to bring positive samples $f_{p}$ closer while pushing negative samples $f_{n}$ apart:", + "bbox": [ + 89, + 584, + 482, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t r i} = \\max \\left(0, d \\left(f _ {p}, f _ {a}\\right) - d \\left(f _ {n}, f _ {a}\\right) + \\alpha\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 637, + 480, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $d(\\cdot)$ represents the Euclidean distance between samples and $\\alpha$ is the margin.", + "bbox": [ + 89, + 662, + 482, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On top of that, we concatenate the LoRA-CLIP's CLS token $T_{CLS}$ with $F_{low}$ along the same dimension to yield the refined representation $F_{out}$ . This ensures that forged features exhibit distinctiveness across different semantic identities while preserving their uniformity within similar semantic identities. The process is formulated as:", + "bbox": [ + 89, + 691, + 483, + 782 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {o u t} = F _ {l o w} \\| T _ {C L S}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 791, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $F^{out}$ is strategically integrated with a linear classifier to enable the execution of binary classification. Eventually, the total loss function $\\mathcal{L}$ of the proposed framework can be defined as:", + "bbox": [ + 89, + 815, + 483, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {b c e} + \\lambda_ {1} \\mathcal {L} _ {t r i} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{bce}$ presents the binary cross-entropy loss and $\\mathcal{L}_{tri}$ is triplet loss. $\\mathcal{L}_{tri}$ and $\\mathcal{L}_r$ are scaled by the hyperparameters $\\lambda_{1}$ and $\\lambda_{2}$ , respectively.", + "bbox": [ + 511, + 529, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 593, + 645, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setups", + "text_level": 1, + "bbox": [ + 511, + 619, + 694, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We follow the protocol described in [32], using ProGAN's real and fake images as training data. Additionally, we adopt the protocol from [5], where the training data is composed of fake Stable Diffusion v1 images [52] and random real LAION images [41]. The UnivFD dataset [32] cover a broad range of generative models, primarily including GANs and diffusion models, such as ProGAN [18], StyleGAN [19], BigGAN [4], CycleGAN [59], StarGAN [10], GauGAN [46], CRN [9], IMLE [22], SAN [11], SITD [7], DeepFakes [20], Guided [12], Glide [31], LDM [40] and DALL-E [37]. The SynRIS dataset [5] is designed to avoid bias toward any specific topic, theme, or style and contains high-fidelity images generated by text-to-image models, such as Kandinsky2 [38], Kandinsky3 [1], PixArt- $\\alpha$ [8], SDXL-DPO [49], SDXL [34], SegMoE [23], SSD-1B [42], Stable-Cascade [33], Segmind-Vega [15], and Würstchen2 [33], Midjourney [29], DALL.E 3 [3]", + "bbox": [ + 511, + 643, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/21d2a7a74f868328d782afdbbed51e059299eb60369d9ada291a02909bf435ec.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. AUROC comparisons with different methods on the SynRIS dataset. We retrieve the results of CNNSpot, UnivFD, and FakeInversion from [5] and obtain the results for Dire, NPR, Fat-Former, and PatchFor using re-implemented models. Red and underline indicate the best and the second best result, respectively." + ], + "table_body": "
MethodsCNN-SpotFreq-specDireUniv FDNPRFatF ormerFakeIn versionPatch ForOurs
Kandinsky260.057.071.656.297.575.669.953.597.13
Kandinsky365.945.774.961.493.780.174.351.494.8
PixArt-α62.756.481.564.789.575.373.049.690.2
SDXL-DPO84.369.869.970.297.686.088.154.595.5
Segmind-Vega74.265.381.062.397.182.481.153.197.0
SDXL81.461.286.266.396.085.180.763.997.6
Seg-MoE66.354.671.962.093.670.871.349.897.6
SSD-1B72.667.879.862.899.670.179.461.299.8
Stable-Cascade70.562.174.168.297.681.674.957.499.2
Würstchen261.063.374.269.790.972.970.547.298.2
Midjourney63.050.972.459.258.573.666.453.790.0
Playground58.252.367.958.793.181.462.554.192.8
DALL-E371.659.980.848.069.179.275.950.185.9
Average68.658.975.962.390.378.074.553.895.1
", + "bbox": [ + 106, + 88, + 472, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and playground [21].", + "bbox": [ + 89, + 476, + 236, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics: As standard evaluation metrics, the average precision (AP), the accuracy (ACC) and AUCROC are considered to measure the effectiveness of different methods.", + "bbox": [ + 89, + 491, + 482, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: In our experiments, we perform thorough comparisons with state-of-the-art methods, as follows: 1) CNNSpot [50]: The method relies only on one CNN generator. 2) PatchFor [6]: The method performs detection on a patch level. 3) Co-occurrence [30]: The method converts input images into co-occurrence matrices for classification. 4) Freq-spec [57]: The method employs the frequency spectrum of images. 5) Dire [51]: The method exploits the error between an input image and its reconstruction counterpart. 6) UnivFD [32]: The method uses the pre-trained language-vision model to determine the authenticity of images. 7) NPR [45]: The method captures the generalized artifacts according to the local interdependence among image pixels. 8) FatFormer [26]: The method is aimed at extracting forgery-adaptive features based on UnivFD. 9) Fakeinversion [5]: The method employs text-conditioned inversion maps extracted from Stable Diffusion.", + "bbox": [ + 89, + 537, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implement details: Our training and testing settings are adapted from the approach outlined in the previous study [26] with several key modifications. Specifically, early stopping was employed during model training, with an initial learning rate of $1 \\times 10^{-4}$ and a batch size of 32. Additionally, the Lora layers are configured with hyperparameters $lora_{r} = 6$ , $lora_{\\alpha} = 6$ , and a dropout rate of 0.8,", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8e851f0a6c0a3b01c9ead4418bfc9f62fb6a7e0eaf71b764adcc5459768e91e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Ablation study of the proposed modules on the UnivFD Dataset. We show the mean accuracy $(acc_m)$ and average precision $(ap_m)$ . Red and underline indicate the best and the second-best result, respectively." + ], + "table_body": "
#STS moduleCFDL modulefeature enhancementUnivFD apmDataset accm
1✓97.3781.64
2✓✓97.4190.17
3✓✓97.3989.98
4✓✓✓98.5293.61
", + "bbox": [ + 529, + 89, + 890, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while $\\alpha$ is set to 8.0. The proposed method is implemented using Pytorch on 2 Nvidia GeForce RTX A6000 GPUs.", + "bbox": [ + 511, + 250, + 903, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison Results", + "text_level": 1, + "bbox": [ + 511, + 287, + 710, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The UnivFD dataset includes a diverse range of models, allowing for a comprehensive evaluation of our method across both GAN and diffusion generative models. In addition, the SynRIS dataset provides images generated by cutting-edge generative models. The overall experimental results are presented in Tab. 1, Tab. 2, and Tab. 3.", + "bbox": [ + 511, + 309, + 903, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on UnivFD dataset. Results show that our proposed method achieves superior performance compared to the UnivFD and FatFormer. Notably, without the biased interpretation introduced by coarse-grained text prompts, SDD surpasses the latest state-of-the-art method FatFormer by the mean AP $(ap_{m})$ of $0.34\\%$ and the mean acc $(acc_{m})$ of $2.75\\%$ . Moreover, compared with methods based on relatively monotonous forgery features in Tab. 1 and Tab. 2, our approach can outperform all of them with a large improvement. The above evidence indicates effective combination of visual concepts and forgery features can contribute model to extract sufficient forgery patterns and eliminate the superfluous features.", + "bbox": [ + 511, + 402, + 905, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on SynRIS dataset. As shown in Tab. 3, when confronted with high-fidelity images generated by text-to-image models, methods leveraging pre-trained vision-language models, such as UnivFD and FatFormer, lose their competitiveness. In contrast, NPR, which focuses on neighboring pixel relationships, retains its edge. We assume that current generative methods grasp the relationships between visual information and semantic concepts in images but cannot refine local forgery details at the pixel level. Considering that excessive reliance on concepts misses abnormal pixel arrangements and focusing on monotonous forgery patterns can cause overfitting, our detector, which emphasizes low-level features with visual concepts, is trained on lower-fidelity fake images generated by Stable Diffusion [52] to capture concept-specific lacunae. We follow the evaluation protocol from SynRIS [5]. In comparison, our detector achieves an impressive mean AUROC of $95.1\\%$ , surpassing the state-of-the-art method by $4.8\\%$ . This demonstrates its superior ability to tackle the challenges posed by evolving generative models.", + "bbox": [ + 511, + 599, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18394", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fc9b0eeef54d2a107d0b1c0f6d6c4e979f1491fe980fcbbc07675e0143222a1.jpg", + "image_caption": [ + "Figure 5. Performance of function on Figure 6. T-SNE visualization of real and fake images [47]. The feature space is based on our adaptive weights. classifier. Each randomly samples 500 real and 500 fake images." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 310, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ba89e1496d6b16bdccd57c575440594fea729573cfd64ec69d877d04d24eeca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 87, + 475, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53ea1b5f353a96c9738269ec99c6192074f98e4c9590ea3b568427e6f48b7455.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 88, + 617, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12508e5e81b25cf63bb1033d94fba7fce7241302bce0d94e1bdf758f953e09b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 759, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88bdb2dfb8a557c7e1f64a77e634075a41c2959d745cddd5ef7d47a86cb6ada7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 88, + 906, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f926a12a11783a8f3703d8b1e19a10696a9a49ef46cb72f9c96a4c3486c118f.jpg", + "image_caption": [ + "Figure 7. The showcase of attention on images which are input into our model." + ], + "image_footnote": [], + "bbox": [ + 91, + 261, + 480, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 462, + 240, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform comprehensive ablation studies on the UnivFD dataset under the original experimental configurations, reporting the mean accuracy $(acc_{m})$ and mean average precision $(ap_{m})$ as the primary evaluation metrics.", + "bbox": [ + 89, + 486, + 483, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Each Component: We study the effects of removing STS module, CFDL module, and feature enhancement in our method. The results, presented in Tab. 4, demonstrate that these components are essential for improving performance in generalization on unseen models. This empirical finding suggests that the CFDL effectively captures forgery discrepancies associated with semantic concepts, while the enhancer module plays a crucial role in identifying robust forgery artifacts. The collaboration of all modules enhances the model's ability to distinguish between real and fake images.", + "bbox": [ + 89, + 549, + 483, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of function on adaptive weights: To check how well the proposed function works in the SDD, we select two conventional functions for comparison: $f(x) = |x|$ and $f(x) = x^2$ . The corresponding results are presented in Fig. 5. We find that our proposed function yields improvements in both $ap_m$ and $acc_m$ compared to the selected functions. These results demonstrate that our proposed function is conducive to capturing robust and distinctive forgery features.", + "bbox": [ + 89, + 717, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of learned latent space: As shown in Fig. 6, the input images can be distinctly categorized into two clusters: real and fake. Nevertheless, why does the di", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vided boundary of ProGAN appear ambiguous in contrast to other models? Additionally, why do the real clusters of CycleGAN and StyleGAN separate from each other? We attribute these to the influence of visual semantic concepts. Perceptively, with the supervision of visual semantic concepts, the learned boundary of ProGAN is more complex and nuanced, rather than just simple straight lines or curves. Similarly, the images generated by StyleGAN and CycleGAN are projected into the corresponding semantic concept distribution and then separated from the real images based on the visual semantic concepts.", + "bbox": [ + 511, + 265, + 906, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of attention on images: We apply Class Activation Mapping (CAM) [58] to visualize the learned representations. From our perspective, Fig. 7 illustrates that with the aid of semantic information, our model can focus on different regions of fake images, including the background, local object regions, and marginal details. This suggests that our fine-grained model is capable of capturing intricate discrepancies generalized to unseen models. Notably, the real images nearly always show no forgery discrepancy regions, which demonstrates the effectiveness of the reconstruction loss in the forgery detection task.", + "bbox": [ + 511, + 431, + 908, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method, SDD, for generalizable forgery image detection. The findings show that our method establishes a new state-of-the-art in detecting images generated by generative models from different periods, which underscores its robustness and superior generalization capability. To the best of our knowledge, in pre-trained vision-language paradigms, our approach is the first to rely solely on visual information, without text prompts. Based on experimental results, we conclude that leveraging sampled tokens and reconstruction techniques effectively aligns the visual semantic concept space with the forgery space. Additionally, refining low-level forgery features under the supervision of visual semantic concepts enhances the performance of forgery detection. Although SDD performs well across various generative methods, there is still room for improvement as generative technologies continue to advance. Future research will explore this further.", + "bbox": [ + 511, + 636, + 908, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (Grants Nos. 62372238,62476133)", + "bbox": [ + 91, + 114, + 483, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] V.Ya. Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. ArXiv, abs/2312.03511, 2023. 6", + "[2] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. ArXiv, abs/2107.03006, 2021. 1", + "[3] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 6", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096, 2018. 6", + "[5] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fake inversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10759-10769, 2024. 2, 3, 6, 7", + "[6] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European Conference on Computer Vision, 2020. 1, 7", + "[7] Cheng Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6", + "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- $\\delta$ : Fast and controllable image generation with latent consistency models. ArXiv, abs/2401.05252, 2024. 6", + "[9] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1520-1529, 2017. 6", + "[10] Yunjey Choi, Min-Je Choi, Mun Su Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2017. 6", + "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shutao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 6", + "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 6" + ], + "bbox": [ + 93, + 186, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Joel Cameron Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. ArXiv, abs/2003.08685, 2020. 1, 3", + "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 1", + "[15] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick von Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss. ArXiv, abs/2401.02677, 2024. 6", + "[16] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloud-vae: Unsupervised feature learning for 3d point clouds from multiple angles by joint self-reconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441-10450. IEEE, 2019. 3", + "[17] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 4", + "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. ArXiv, abs/1710.10196, 2017. 6", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2018. 6", + "[20] Prabhat Kumar, Mayank Vatsa, and Richa Singh. Detecting face2face facial reenactment in videos. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2578-2586, 2020. 6", + "[21] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. ArXiv, abs/2402.17245, 2024. 7", + "[22] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4219-4228, 2018. 6", + "[23] Zhenghong Li, Hao Chen, Jiangjiang Wu, Jun Li, and Ning Jing. Segmind: Semisupervised remote sensing image semantic segmentation with masked image modeling and contrastive learning method. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 6", + "[24] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, 2022. 3", + "[25] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 772-781, 2021. 2", + "[26] Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Yao Zhao, and Jingdong Wang. Forgery-aware adaptive transformer for generalizable synthetic image detection." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10770-10780, 2023. 1, 2, 3, 4, 5, 7", + "[27] Xinhai Liu, Xinchen Liu, Zhizhong Han, and Yu-Shen Liu. Spu-net: Self-supervised point cloud upsampling by coarse-to-fine reconstruction with self-projection optimization. IEEE Transactions on Image Processing, 31:4213-4226, 2020. 3", + "[28] Zhengzhe Liu, Xiaojuan Qi, Jiaya Jia, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8066, 2020. 3", + "[29] Midjourney. Midjourney, n.d. Accessed: 2025-03-04. 6", + "[30] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, B. S. Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H. Bappy, and Amit K. Roy-Chowdhury. Detecting gan generated fake images using co-occurrence matrices. ArXiv, abs/1903.06836, 2019. 7", + "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 6", + "[32] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 5, 6, 7", + "[33] Pablo Pernias, Dominic Rampas, Mats L. Richter, and Marc Aubreville. Wuerstchen: Efficient pretraining of text-to-image models. ArXiv, abs/2306.00637, 2023. 6", + "[34] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 6", + "[35] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 763-772, 2020. 6", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2", + "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 6", + "[38] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, V.Ya. Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. In Conference on Empirical Methods in Natural Language Processing, 2023. 6", + "[39] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "autoencoder reconstruction error. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9130-9140, 2024. 3, 2", + "[40] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 6", + "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6", + "[42] Segmind. Announcing ssd-1b: A leap in efficient t2i generation., 2023. 6", + "[43] Minghe Shen, Hongping Gan, Chao Ning, Yi Hua, and Tao Zhang. Transc: A transformer-based hybrid architecture for image compressed sensing. IEEE Transactions on Image Processing, 31:6991-7005, 2022. 2", + "[44] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12105-12114, 2023. 1, 2", + "[45] Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28130-28139, 2024. 2, 7", + "[46] Devavrat Tomar, Manana Lortkipanidze, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. Self-attentive spatial adaptive normalization for cross-modality domain adaptation. IEEE Transactions on Medical Imaging, 40:2926-2938, 2021. 6", + "[47] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8", + "[48] Tim van Erven and Peter Harremoës. Rényi divergence and kullback-leibler divergence. IEEE Transactions on Information Theory, 60:3797-3820, 2012. 4", + "[49] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 6", + "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8692-8701, 2019. 1, 7, 2, 3", + "[51] Zhendong Wang, Jianmin Bao, Wen gang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion-generated image detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22388-22398, 2023. 3, 7", + "[52] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. In Annual Meeting of the Association for Computational Linguistics, 2022. 6, 7", + "[53] Syed Waqas Zamir, Aditya Arora, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for fast image restoration and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1934-1948, 2022. 2", + "[54] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1593-1603, 2024. 4", + "[55] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. ArXiv, abs/2306.10012, 2023. 4, 5", + "[56] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7545-7556, 2023. 5", + "[57] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. 2019 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 7", + "[58] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2015. 8", + "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242-2251, 2017. 6" + ], + "bbox": [ + 91, + 90, + 482, + 654 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18398", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_model.json b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..986423bb006c1b0a21a03fa676be72e9c1d934ce --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.857, + 0.152 + ], + "angle": 0, + "content": "Semantic Discrepancy-aware Detector for Image Forgery Identification" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.181, + 0.749, + 0.218 + ], + "angle": 0, + "content": "Ziye Wang Minghang Yu Chunyan Xu* Zhen Cui Nanjing University of Science and Technology, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.22, + 0.735, + 0.234 + ], + "angle": 0, + "content": "{wzynjust,mhyu,cyx}@njust.edu.cn，zhen.cui@bnu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.695 + ], + "angle": 0, + "content": "With the rapid advancement of image generation techniques, robust forgery detection has become increasingly imperative to ensure the trustworthiness of digital media. Recent research indicates that the learned semantic concepts of pre-trained models are critical for identifying fake images. However, the misalignment between the forgery and semantic concept spaces hinders the model's forgery detection performance. To address this problem, we propose a novel Semantic Discrepancy-aware Detector (SDD) that leverages reconstruction learning to align the two spaces at a fine-grained visual level. By exploiting the conceptual knowledge embedded in the pre-trained vision-language model, we specifically design a semantic token sampling module to mitigate the space shifts caused by features irrelevant to both forgery traces and semantic concepts. A concept-level forgery discrepancy learning module, based on reconstruction, enhances the interaction between semantic concepts and forgery traces, effectively capturing discrepancies under the concepts' guidance. Finally, the low-level forgery feature enhancement integrates the learned concept-level forgery discrepancies to minimize redundant forgery information. Experiments conducted on two standard image forgery datasets demonstrate the efficacy of the proposed SDD, which achieves superior results compared to existing methods. The code is available at https://github.com/wzy111111/SSD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.726, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.483, + 0.856 + ], + "angle": 0, + "content": "With the thriving of generative AI technologies, like Generative Adversarial Networks (GANs) [14] and diffusion models [2], the images generated by these models can easily create confusion by passing off the spurious as genuine. Therefore, it is crucial to develop a universal method for detecting fake images to mitigate the widespread dissemination of disinformation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.859, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Pioneering research [26, 32] has shown that projecting" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.267, + 0.889, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.338, + 0.885, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.709 + ], + "angle": 0, + "content": "Figure 1. The phenomenon of misalignment between semantic concept space and forgery space. Since \\(\\cos \\theta\\) can reflect the similarity of image descriptions, we model the feature space in polar coordinates. As the semantic concept space in [32] is frozen, fake samples sharing similar concepts with real ones can be easily misclassified. With forgery-adaptive space like [26], the model can correctly distinguish between them based on re-learned forgery features. Nevertheless, due to the semantic concept bias introduced by coarse text prompts, the target samples may be projected into an inaccurate semantic concept dimension, causing them to drift away from the real source samples along the fake dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.809 + ], + "angle": 0, + "content": "images in a joint embedding space of texts and images can effectively capture discrepancies between fake and real images. In contrast, methods [6, 13, 44, 50] overlooking the interplay between forgery traces and semantic concepts perform poorly when confronted with unseen generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To investigate the visual semantic concepts of pretrained models, we conduct a statistical analysis of the output features from CNNSpot [50] and CLIP-ViT [32] (See Appendix A for more details). Under different categories, CNNSpot exhibits a synchronized difference between real and fake features in its training space. However, when tran" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18388" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.466, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.33 + ], + "angle": 0, + "content": "Figure 2. Different paradigms of image forgery identification with pre-trained vision-language model. (a) Fine-tune the frozen model only by fully connected (FC) layers [32]. (b) Prompt-based designs are tuned on text prompts and contrastive objectives [26]. (c) Our paradigm incorporating visual clues can capture fine-grained forge traces by reconstruction learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.415 + ], + "angle": 0, + "content": "sitioning to the CLIP's space, these differences become inconsistent. From this, we infer a nuanced relationship between semantic concepts and forgery traces: Different semantic concepts may guide the model to uncover distinct forgery traces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Intuitively, relying on a frozen pre-trained vision-language model like UnivFD [32] is essential, but this tends to overlook fine-grained forgery details. Although Fat-Former [26] achieves a substantial enhancement in generalization by employing the forgery-aware adaptive transformer, we observe that soft prompts based on simple [CLASS] embeddings have an intrinsic limitation in their semantic description granularity (See Appendix B for more details). The constrained breadth of the conveyed concepts may lead the detection toward incorrect predictions. This limitation highlights a misalignment between the visual semantic concept space and the target forgery space, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.732 + ], + "angle": 0, + "content": "To address this, one empirical approach is to design more detailed text descriptions, but this method struggles to describe all visual forgery details due to the limited length of texts and brings more computational overhead. Drawing from the aforementioned findings and analysis, we make a first attempt to align the CLIP's visual semantic concept space with the forgery space by reconstructing semantic features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We develop a vision-based paradigm, as outlined in Fig. 2. First, employing a pre-trained model only with nearest neighbor or linear probing (e.g. UnivFD [32], Fig. 2 (a)) is suboptimal for image forgery detection. Second, modifying the pre-trained model with task-specific prompts (e.g. FatFormer [26], Fig. 2 (b)) may favor models biased towards any particular semantic concept. These studies pave the way for exploring pre-trained space with rich semantic concepts. Inspired by image reconstruction [43, 53], our paradigm amplifies the concept-level forgery discrepancies of forgery images, which empowers the model to detect sus" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "picious forgery traces with the assistance of semantic concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.56 + ], + "angle": 0, + "content": "In this work, we present a novel Semantic Discrepancy-aware Detector (SDD) to accurately align the semantic concept space and the forgery space. Firstly, to mitigate interference from features unrelated to learned semantic concepts and forgery traces, we divide the real images into non-overlapping blocks and feed them to the frozen CLIP [36] to obtain diverse semantic patch tokens. These tokens acting as visual clues smoothly align the semantic concepts' space and forgery space. It is noteworthy that these tokens sampled by JS divergence are universally representative of the real semantic distribution. Then, the visual clues are fused into a concept-level forgery discrepancy module. Unlike FatFormer, LoRA layers are incorporated into the image encoder. The goal is to preserve the completeness and diversity of the learned semantic concepts of CLIP, while the forgery features sharing similar semantic concepts should be highlighted. During reconstruction, we only narrow the reconstruction gap for real samples to reinforce the reconstructed discrepancies of the synthetic images. Finally, we present low-level forgery feature enhancement to let the reconstruction difference map enhance the extraction of the highly generalizable forgery features while introducing minimal additional parameters. The main challenge is how to capture forgery features with strong semantic concept correlation and features with high forgery relevance but weak semantic concept ties to ensure the model converges to powerful features. Motivated by this, we apply convolutional modules and adaptive weight parameters to avoid over-relying on semantic concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.664 + ], + "angle": 0, + "content": "We thoroughly evaluate the generalization performance of our model on a UnivFD benchmark [32] and a SynRIS benchmark [5]. Surprisingly, our method achieves superior performance by a \\( ap_{m} \\) of 98.51% and a \\( acc_{m} \\) of 93.61% on the UnivFD benchmark [32] and an average AUROC of 95.1% on the SynRIS benchmark [5]. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.71 + ], + "angle": 0, + "content": "- We propose a robust model (SDD) for forgery detection, specifically designed to align the visual concept space and forgery space in terms of visual information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.711, + 0.905, + 0.77 + ], + "angle": 0, + "content": "- We sample semantic tokens to mitigate the space shifts and align the two spaces through reconstruction learning. Additionally, we strengthen low-level forgery features to enhance the model's robustness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.772, + 0.905, + 0.817 + ], + "angle": 0, + "content": "- Our method achieves superior performance on two benchmarks, demonstrating its superior capability in comparison to existing approaches." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.655, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "AI-generated Images Detection. Extensive efforts have been devoted to enhancing the performance of AI-generated image detection. Early works like [25, 44, 45] tend to mine" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18389" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.907, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 3. The architecture of SDD. First, Next, we sample semantic tokens from real images to learn features related to both concepts and forgery. the input images are mapped into a joint space of visual semantic concepts and forgery, which are transformed into learnable features \\(V_{H}\\) . Then, we use transformer-based encoder and decoder to get reconstructed features \\(\\mathcal{R}_f\\) . A reconstruction difference map \\(\\mathcal{D}_S\\) is obtained and goes through the multi-scale convolutional network to refine forgery features. Finally, we concatenate the CLIP's CLS token with this output along the same dimension for classification. The whole system is trained by jointly minimizing the binary cross-entropy loss \\(L_{bce}\\) , the reconstruction loss \\(L_{r}\\) , and the triplet loss \\(L_{tri}\\)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.495, + 0.485, + 0.873 + ], + "angle": 0, + "content": "the common forgery traces between all real and fake images, such as noise patterns, texture statistics, and frequency signals. As an illustration, Liu et al. [24] designed a network that learns the consistent noise patterns in images for fake detection. Liu et al. [28] proposed to leverage the gram matrix to discover the global anomalous texture of fake images. An effective approach [13] demonstrated that frequency representation is an important factor in improving fake detection performance. However, these differences are rigorously specific to the monotonous features, which contribute to the issue of overfitting. Cutting-edge research [26, 32] shifted attention toward the semantic properties of images. Ojha et al. [32] showed that projecting images into the feature space of pre-trained vision-language model enables strong generalization ability. To build generalized forgery representations, Liu et al. [26] constructed forgery adaptive space by a forgery-aware adapter. The above research [5, 26, 32] has suggested that concept attributes are vital in the image forgery detection task. Assuming that diffusion-based models leave distinct forgery traces that are characteristic of specific concept distributions, we aim to extract robust forgery features guided by semantic concepts, rather than suppressing them. Therefore, even \"useless\" information can be useful by providing significant certainty about the content of the image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.874, + 0.483, + 0.889 + ], + "angle": 0, + "content": "Reconstruction Learning. Reconstruction learning has" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.909, + 0.77 + ], + "angle": 0, + "content": "great potential in unsupervised representation learning [16, 27]. Some works [5, 39] utilized reconstruction learning to reveal the nuances between real and fake images. For example, Wang et al. [51] found that reconstructing images by DDIM exposes an error between real images and their reconstructed replica. The new synthetic image detection method[5] used text-conditioned inversion maps to learn internal representations, which is conducive to predicting whether an image is fake. Ricker et al. [39] offered a simple detection approach by applying AE to measure the reconstruction error. Notably, these works are committed to reconstructing the distributions of both real and fake samples by leveraging generative models. Unlike previous works, we focus solely on reconstructing real images in the finetuned CLIP space in light of the authenticity and richness of semantic concepts. The distribution of real samples, learned from pre-trained vision-language model, helps to define an optimal boundary, thus alleviating overfitting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.65, + 0.801 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our goal is to align forgery and visual semantic concept spaces using reconstruction techniques for robust and generalizable synthetic image detection. To achieve this, we introduce a fine-grained model named Semantic Discrepancy-aware Detector (SDD). Building on prior works, we harness the generalization capability of vision-language mod" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18390" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "els. Semantic concept space: The ideal joint embedding space of images and texts with four properties: semantic alignment, modality invariance, locality consistency, and structure preservation. Forgery space: The ideal space covers forgery traces. Notably, we derive semantic concept space via a vision language model pretrained solely on real images; thus we treat the two spaces as independent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.379 + ], + "angle": 0, + "content": "First, the Semantic Tokens Sampling (STS) module utilizes Jensen-Shannon (JS) divergence to sample semantic patch tokens, which serve as a transitional bridge, facilitating model to establish the association between real and forgery images accurately. Next, the Concept-level forgery Discrepancy Learning (CFDL) module employs reconstruction learning to explore the forgery discrepancies within the visual semantic concept space, which focuses on identifying subtle variations between reconstructed forgery features. Finally, the reconstruction difference map is trained with Low-level forgery Feature Enhancement module, which aims to refine forgery features with more visual details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.388, + 0.336, + 0.404 + ], + "angle": 0, + "content": "3.1. Semantic Tokens Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Initially, we considered directly aligning the visual semantic concept space and forgery space by leveraging fine-grained reconstruction learning to model real and fake semantic distributions. However, this strategy would treat the differences in features unrelated to semantics and forgery as crucial factors for identifying image's authenticity. To eliminate these redundant features, we sample real semantic image patch tokens as visual clues to bridge real and forged semantic domains. This module enables the model to focus on concept-related forgery traces and highlight the distinctions between real and fake images. In a tangible way, the image encoder of CLIP: ViT-L/14 is adapted to transform a real image \\( x_{r} \\) into a set of features \\( f_{r} \\), without the image CLS token. We define the transformation as:" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.633, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\nf _ {r} = \\phi (x _ {r}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.735 + ], + "angle": 0, + "content": "where \\(\\phi (\\cdot)\\) is the CLIP:ViT-L/14's visual encoder, \\(x_{r}\\in\\) \\(\\mathbb{I}_r^{H\\times W\\times 3}\\) represents a real image characterized by a height of \\(H\\) and a width of \\(W\\). Besides, \\(f_{r}\\in \\mathbb{R}^{N\\times D}\\), where \\(N\\) is the number of tokens and \\(D\\) denotes the dimension of each patch token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Since integrating all real patch tokens into the image reconstruction module is computationally intensive and memory-consuming, it is urgent to select a suitable subset of these tokens. From a distribution perspective, the Jensen-Shannon (JS) divergence, derived from the Kullback-Leibler divergence [48], is a symmetric and finite metric that can effectively measure the similarity between tokens by quantifying differences in their distributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To calculate the JS divergence between two tokens, both are converted into computable probability distribution space using the softmax function. Let \\( f_{s} \\in \\mathbb{F}_{r}^{M \\times D} \\) be the selected" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "semantic patch tokens with the num of tokens \\( M = 1 / \\delta \\) and dimension \\( D \\) in terms of sampling rate \\( \\delta \\) (\\( 0 \\leq \\delta \\leq 1 \\), \\( \\delta \\) is user-defined). Once the initial token \\( \\tilde{r} \\) and \\( \\delta \\) are determined, the JS divergence between \\( \\tilde{r} \\) and other tokens \\( r \\) falls within the range [0, 1]. Subsequently, the sampling interval is split into \\( M \\) equal segments with one token selected from each segment. As a consequence, the semantic tokens sampling module can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.215, + 0.708, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {s} = \\mathcal {S} \\left(\\mathbb {R} ^ {N \\times D}, \\delta\\right) \\\\ = A _ {c} ^ {N _ {a} \\times M} \\times \\mathbb {R} ^ {N \\times D}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.257, + 0.905, + 0.291 + ], + "angle": 0, + "content": "s.t. JS(softmax \\((\\tilde{r})\\) ,softmax \\((r)) = \\frac{i}{M}\\) , if \\(a_{ij} = 1\\) (2)" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.289, + 0.794, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i = 0} ^ {N _ {a}} \\sum_ {j = 0} ^ {M} a _ {i j} = M; \\sum_ {j = 0} ^ {M} a _ {i j} = 1 \\text {o r} 0,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.334, + 0.761, + 0.351 + ], + "angle": 0, + "content": "\\[\ni = 1, \\dots , N _ {a}; j = 1, \\dots , M,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.541 + ], + "angle": 0, + "content": "where \\( S(\\cdot, \\cdot) \\) represents the sampling process. \\( A_{c}^{N_{a} \\times M} \\) is a constraint matrix of size \\( N_{a} \\times M \\) whose element \\( a_{ij} \\) is constrained to the binary pattern of \\( \\{0, 1\\} \\). Here JS \\( (\\cdot) \\) refers to the Jensen-Shannon divergence, \\( N_{a} \\) denotes the total number of real image patch tokens sampled from the training dataset of UnivFD and \\( M \\) represents the required subset size. The softmax \\( (\\cdot) \\) is the softmax function. The sampling tokens help the reconstruction module avoid becoming biased towards any particular forgery-unrelated distribution. Meanwhile, it avoids the semantic bias often introduced by text prompts, since the tokens are evenly distributed in a unified CLIP space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.887, + 0.564 + ], + "angle": 0, + "content": "3.2. Concept-level Forged Discrepancy Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.781 + ], + "angle": 0, + "content": "A few words alone can hardly paint a picture. We argue that the fine-grained visual details can uncover more forgery traces concealed in the images. As such, we mix sampling tokens with extracted features and capitalize on reconstruction learning to compensate for the omission of forgery traces brought by coarse prompts. As previous work [26] has demonstrated that the pre-trained vision-language model necessitates fine-tuning to adapt to the forgery detection task. Therefore, we integrate LoRA [17] with the CLIP-ViT model to capture discriminative forgery features by making use of the bread semantic concepts. This method, denoted as LoRA-CLIP [54], is more streamlined and flexible. Given an input image \\(\\mathcal{I} \\in \\mathbb{I}^{H \\times W \\times 3}\\), we can get high-level visual features \\(V_{H}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.787, + 0.905, + 0.804 + ], + "angle": 0, + "content": "\\[\nV _ {H} = \\mathcal {F} _ {L o R A} (\\mathcal {I}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Here \\(\\mathcal{F}_{LoRA}\\) refers to the CLIP image encoder fine-tuned by LoRA. The reconstruction module of CFDL encompasses two submodules, i.e., transformer-based encoder and decoder. Thanks to the transformer's capability of long-range relationship modeling, we capitalize on the multi-head attention (MHA) mechanism, the core mechanism of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18391" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "the transformer, to obtain more discriminative distribution by utilizing contextual information, which is set as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.132, + 0.483, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{r l} \\operatorname {h e a d} _ {i} & = \\operatorname {A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\\\ & = \\operatorname {S o f t m a x} \\left(\\frac {Q W _ {i} ^ {Q} \\left(K W _ {i} ^ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right) V W _ {i} ^ {V}, \\\\ \\operatorname {M H A} (Q, K, V) & = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.333 + ], + "angle": 0, + "content": "where \\(Q\\) (Query), \\(K\\) (Key), \\(V\\) (Value) refer to the input, \\(W_{i}^{Q}, W_{i}^{K}, W_{i}^{V}\\) separately denote the corresponding weights of linear projection, \\(\\mathrm{Attn}(\\cdot)\\) denotes the function of the scaled dot product, Softmax is the softmax function, \\(d\\) refers to the dimension of input, Concat (\\(\\cdot\\)) represents the concatenation used to stitch the discrete attention outputs of head \\(1 \\sim h\\) together." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.395 + ], + "angle": 0, + "content": "To amplify the discrepancy between a fake image and its reconstructed counterpart, the sampled visual clues are employed for the initial processing by the encoder. The encoder's process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.407, + 0.482, + 0.423 + ], + "angle": 0, + "content": "\\[\nR _ {1} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(f _ {s}, V _ {H}, V _ {H}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.426, + 0.482, + 0.442 + ], + "angle": 0, + "content": "\\[\nR _ {2} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(R _ {1}, V _ {H}, V _ {H}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.528 + ], + "angle": 0, + "content": "where \\(\\mathrm{LN}(\\cdot)\\) denotes the Layer Normalization. Then, the encoder's outputs used as queries are injected into the decoder to get the final reconstructed features, which are similar to the encoder process and perform the following operation:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.542, + 0.482, + 0.558 + ], + "angle": 0, + "content": "\\[\nR _ {3} = \\mathrm {L N} (\\mathrm {M H A} (R _ {2}, R _ {2}, R _ {2})), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.561, + 0.482, + 0.577 + ], + "angle": 0, + "content": "\\[\nR _ {e} = \\mathrm {L N} (\\mathrm {M H A} \\left(R _ {1}, R _ {3}, R _ {3}\\right)). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.647 + ], + "angle": 0, + "content": "During the reconstruction process, we just calculate the reconstruction loss \\(\\mathcal{L}_r\\) between the real input features and their reconstructed counterparts \\(\\mathcal{R}_e\\) within a mini-batch as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.646, + 0.483, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\operatorname {M S E} \\left(R _ {e}, V _ {H}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.785 + ], + "angle": 0, + "content": "where \\(\\mathrm{MSE}(\\cdot ,\\cdot)\\) is mean squared error. Facilitating \\(\\mathcal{L}_r\\) encourages preserving the completeness and richness of the visual semantic concept space and highlighting the concept-related forgery features. Given the reconstructed features \\(R_{f}\\) and the original feature \\(f_{r}\\), the reconstruction difference map can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.796, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {s} = \\left| R _ {f} - f _ {r} \\right|, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.824, + 0.392, + 0.839 + ], + "angle": 0, + "content": "where \\(|\\cdot |\\) denotes the absolute value function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.442, + 0.866 + ], + "angle": 0, + "content": "3.3. Low-level Forgery Feature Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Existing methods based on pre-trained vision-language models [26, 32] overlook the importance of concept" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.091, + 0.837, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.256, + 0.905, + 0.299 + ], + "angle": 0, + "content": "Figure 4. The curve of exponential inverse. In the \"fast\" interval, the value drops sharply. In the \"low\" interval, the curve flattens out, showing a decay towards 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.565 + ], + "angle": 0, + "content": "weakly-related features. We believe that a thorough alignment between the visual semantic concept space and the forgery space should include the exploration of concept weakly-related forgery features. To eliminate redundant forgery features, we come up with a novel feature enhancement that refines low-level forgery features. Empowered by the reconstruction difference map, our detector orchestrates the extraction of multi-scale features with exceptional robustness and markedly enhanced effectiveness. As shown in Fig. 3, the enhancer follows the typical architecture of a convolutional network. It involves the repeated application of convolutions, each followed by a batch normalization (BN) and a rectified linear unit (ReLU). For a given stage \\( n \\), \\( F(n) \\) (\\( N = 1,2,3 \\)) corresponds to its output features. Then, We deconvolve the semantic difference map \\( D_{s} \\) to the shape same as \\( F(N) \\) and perform pixel-wise multiplication with \\( F(n) \\) to get \\( F'(n) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.576, + 0.905, + 0.595 + ], + "angle": 0, + "content": "\\[\nF ^ {\\prime} (n) = \\operatorname {d e c o n v} (F (n)) \\otimes D _ {s}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.698 + ], + "angle": 0, + "content": "where \\(\\otimes\\) is the element-wise multiplication, deconv \\((\\cdot)\\) represents deconvolution operation and \\(\\mathcal{F}'(n)\\) is the low-level feature aggregated with semantic information. To further enhance the reliability of the extracted features, we compute an adaptive weight coefficient \\(\\frac{1}{e_n}\\) to indicate the importance of \\(D_{s}\\) to \\(F(n)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.708, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{e _ {n}} = \\frac {1}{e ^ {\\left| F ^ {\\prime} (n) - F (n) \\right|}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Here we explain the role of the exponential inverse through Fig. 4. As \\( x \\) grows large, the curve of \\( e^x \\) becomes flatter. Therefore, in the \"fast\" interval, forgery features with a significant divergence from the semantic difference map will be assigned smaller weights, which mobilizes the network to capture concept strongly-related features. However, in the \"low\" interval, features strongly associated with forgery can avoid being misguided by semantic concepts, which indicates that the order of importance is reversed. Next, we have the attended output features \\( F_{low} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18392" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.26 + ], + "angle": 0, + "content": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDallemAP
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200Steps200w/cfg100Steps100w/ CFG10027502710
CNN-SpotCVPR2020100.093.4784.5099.5489.4998.1589.0273.7559.4798.2498.4073.7270.6271.0070.5480.6584.9182.0770.5983.58
PatchForECCV202080.8872.8471.6685.7565.9969.2576.5576.1976.3474.5268.5275.0387.1086.7286.4085.3783.7378.3875.6777.73
Co-occurrenceElect.Imag.99.7480.9550.6198.6353.1167.9959.1468.9860.4273.0687.2170.2091.2189.0292.3989.3288.3582.7980.9678.11
Freq-specWIFS201955.39100.075.0855.1166.08100.045.1847.4657.1253.6150.9857.7277.7277.2576.4768.5864.5861.9267.7766.21
DireICCV2023100.083.5981.5096.5081.7099.8895.7362.5169.9897.3198.6279.5375.5273.4276.4586.2889.0088.3451.3583.54
UnivFDCVPR2023100.099.4699.5997.2499.9899.6082.4561.3279.0296.7299.0087.7799.1492.1599.1794.7495.3494.5797.1593.38
NPRCVPR2024100.099.5096.5099.8096.80100.092.2073.1078.7087.2064.8065.8099.8099.8099.8099.7099.8099.8098.6092.19
FatFormerCVPR2024100.0100.099.9899.75100.0100.097.9997.9481.2399.8499.9391.9299.8399.2299.8999.2799.5099.3399.8498.18
Ours100.099.7799.9399.4899.9899.9797.2397.9193.1099.7999.9692.0699.8898.9599.9298.0698.2997.7399.8198.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.27, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Table 1. Average precision comparisons with different methods on the UnivFD dataset. We replicate the results of CNNSpot, Patchfor, Co-occurrence, Freq-spec, and UnivFD from the paper[32]. In addition, we obtained the results for Dire, NPR and FatFormer using either the official pre-trained models or our re-implemented versions. Red and underline indicates the best and the second best result, respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.325, + 0.907, + 0.497 + ], + "angle": 0, + "content": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDalleAvg-acc
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200 Steps200 wcfg100 Steps100 2750 27100 10
CNN-SpotCVPR202099.9985.2070.2085.7078.9591.7053.4766.6748.6986.3186.2660.0754.0354.9654.1460.7863.8065.6655.5869.58
PatchForECCV202075.0368.9768.4779.1664.2363.9475.5475.1475.2872.3355.3067.4176.5076.1075.7774.8173.2868.5267.9171.24
Co-occurrenceElect.Imag.97.7063.1553.7592.5051.1054.7057.1063.0655.8565.6565.8060.5070.7070.5571.0070.2569.6069.9067.5566.86
Freq-specWIFS201949.9099.9050.5049.9050.3099.7050.1050.0048.0050.6050.1050.9050.4050.4050.3051.7051.4050.4050.0055.45
DireICCV202399.8673.4760.6872.3965.1593.6088.8652.7856.3990.0794.0561.0559.3559.9560.6569.3072,7071.0052.7571.19
UnivFDCVPR2023100.098.5094.5082.0099.5097.0066.6063.0057.5059.5072.0070.0394.1973.7694.3679.0779.8578.1486.7881.38
NPRCVPR202499.8092.0089.5096.3087.6099.7079.4061.4070.6074.5057.1055.2397.4098.7097.9097.0097.9097.0088.8086.20
FatFormerCVPR202499.8999.3699.5097.1299.4399.7593.2581.3968.0469.4769.4776.0098.5594.8598.6094.3094.6094.1598.7090.86
Ours99.8895.7696.7098.0898.4699.1791.8283.6177.4595.4096.4779.5598.0594.6098.2592.2093.3591.8098.0093.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.507, + 0.731, + 0.521 + ], + "angle": 0, + "content": "Table 2. Accuracy comparisons with different methods on the UnivFD dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.531, + 0.273, + 0.545 + ], + "angle": 0, + "content": "by the residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.553, + 0.482, + 0.586 + ], + "angle": 0, + "content": "\\[\nF _ {l o w} (n) = F ^ {\\prime} (n) + \\frac {F ^ {\\prime} (n)}{e _ {n}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.632 + ], + "angle": 0, + "content": "For optimizing the anchor features \\(\\tilde{f}_a\\) of the enhancer, the following triplet loss [35] is employed to bring positive samples \\(f_{p}\\) closer while pushing negative samples \\(f_{n}\\) apart:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.638, + 0.482, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t r i} = \\max \\left(0, d \\left(f _ {p}, f _ {a}\\right) - d \\left(f _ {n}, f _ {a}\\right) + \\alpha\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.693 + ], + "angle": 0, + "content": "where \\(d(\\cdot)\\) represents the Euclidean distance between samples and \\(\\alpha\\) is the margin." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.784 + ], + "angle": 0, + "content": "On top of that, we concatenate the LoRA-CLIP's CLS token \\( T_{CLS} \\) with \\( F_{low} \\) along the same dimension to yield the refined representation \\( F_{out} \\). This ensures that forged features exhibit distinctiveness across different semantic identities while preserving their uniformity within similar semantic identities. The process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.792, + 0.483, + 0.809 + ], + "angle": 0, + "content": "\\[\nF _ {o u t} = F _ {l o w} \\| T _ {C L S}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.484, + 0.876 + ], + "angle": 0, + "content": "where \\( F^{out} \\) is strategically integrated with a linear classifier to enable the execution of binary classification. Eventually, the total loss function \\( \\mathcal{L} \\) of the proposed framework can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {b c e} + \\lambda_ {1} \\mathcal {L} _ {t r i} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.577 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{bce}\\) presents the binary cross-entropy loss and \\(\\mathcal{L}_{tri}\\) is triplet loss. \\(\\mathcal{L}_{tri}\\) and \\(\\mathcal{L}_r\\) are scaled by the hyperparameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.594, + 0.646, + 0.611 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.696, + 0.637 + ], + "angle": 0, + "content": "4.1. Experiment Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Datasets: We follow the protocol described in [32], using ProGAN's real and fake images as training data. Additionally, we adopt the protocol from [5], where the training data is composed of fake Stable Diffusion v1 images [52] and random real LAION images [41]. The UnivFD dataset [32] cover a broad range of generative models, primarily including GANs and diffusion models, such as ProGAN [18], StyleGAN [19], BigGAN [4], CycleGAN [59], StarGAN [10], GauGAN [46], CRN [9], IMLE [22], SAN [11], SITD [7], DeepFakes [20], Guided [12], Glide [31], LDM [40] and DALL-E [37]. The SynRIS dataset [5] is designed to avoid bias toward any specific topic, theme, or style and contains high-fidelity images generated by text-to-image models, such as Kandinsky2 [38], Kandinsky3 [1], PixArt- \\(\\alpha\\) [8], SDXL-DPO [49], SDXL [34], SegMoE [23], SSD-1B [42], Stable-Cascade [33], Segmind-Vega [15], and Würstchen2 [33], Midjourney [29], DALL.E 3 [3]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18393" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.473, + 0.382 + ], + "angle": 0, + "content": "
MethodsCNN-SpotFreq-specDireUniv FDNPRFatF ormerFakeIn versionPatch ForOurs
Kandinsky260.057.071.656.297.575.669.953.597.13
Kandinsky365.945.774.961.493.780.174.351.494.8
PixArt-α62.756.481.564.789.575.373.049.690.2
SDXL-DPO84.369.869.970.297.686.088.154.595.5
Segmind-Vega74.265.381.062.397.182.481.153.197.0
SDXL81.461.286.266.396.085.180.763.997.6
Seg-MoE66.354.671.962.093.670.871.349.897.6
SSD-1B72.667.879.862.899.670.179.461.299.8
Stable-Cascade70.562.174.168.297.681.674.957.499.2
Würstchen261.063.374.269.790.972.970.547.298.2
Midjourney63.050.972.459.258.573.666.453.790.0
Playground58.252.367.958.793.181.462.554.192.8
DALL-E371.659.980.848.069.179.275.950.185.9
Average68.658.975.962.390.378.074.553.895.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.466 + ], + "angle": 0, + "content": "Table 3. AUROC comparisons with different methods on the SynRIS dataset. We retrieve the results of CNNSpot, UnivFD, and FakeInversion from [5] and obtain the results for Dire, NPR, Fat-Former, and PatchFor using re-implemented models. Red and underline indicate the best and the second best result, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.238, + 0.492 + ], + "angle": 0, + "content": "and playground [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.536 + ], + "angle": 0, + "content": "Metrics: As standard evaluation metrics, the average precision (AP), the accuracy (ACC) and AUCROC are considered to measure the effectiveness of different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Baselines: In our experiments, we perform thorough comparisons with state-of-the-art methods, as follows: 1) CNNSpot [50]: The method relies only on one CNN generator. 2) PatchFor [6]: The method performs detection on a patch level. 3) Co-occurrence [30]: The method converts input images into co-occurrence matrices for classification. 4) Freq-spec [57]: The method employs the frequency spectrum of images. 5) Dire [51]: The method exploits the error between an input image and its reconstruction counterpart. 6) UnivFD [32]: The method uses the pre-trained language-vision model to determine the authenticity of images. 7) NPR [45]: The method captures the generalized artifacts according to the local interdependence among image pixels. 8) FatFormer [26]: The method is aimed at extracting forgery-adaptive features based on UnivFD. 9) Fakeinversion [5]: The method employs text-conditioned inversion maps extracted from Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implement details: Our training and testing settings are adapted from the approach outlined in the previous study [26] with several key modifications. Specifically, early stopping was employed during model training, with an initial learning rate of \\(1 \\times 10^{-4}\\) and a batch size of 32. Additionally, the Lora layers are configured with hyperparameters \\(lora_{r} = 6\\), \\(lora_{\\alpha} = 6\\), and a dropout rate of 0.8," + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.09, + 0.892, + 0.173 + ], + "angle": 0, + "content": "
#STS moduleCFDL modulefeature enhancementUnivFD apmDataset accm
1✓97.3781.64
2✓✓97.4190.17
3✓✓97.3989.98
4✓✓✓98.5293.61
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.187, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Table 4. Ablation study of the proposed modules on the UnivFD Dataset. We show the mean accuracy \\((acc_m)\\) and average precision \\((ap_m)\\). Red and underline indicate the best and the second-best result, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.25, + 0.905, + 0.28 + ], + "angle": 0, + "content": "while \\(\\alpha\\) is set to 8.0. The proposed method is implemented using Pytorch on 2 Nvidia GeForce RTX A6000 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.712, + 0.305 + ], + "angle": 0, + "content": "4.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.401 + ], + "angle": 0, + "content": "The UnivFD dataset includes a diverse range of models, allowing for a comprehensive evaluation of our method across both GAN and diffusion generative models. In addition, the SynRIS dataset provides images generated by cutting-edge generative models. The overall experimental results are presented in Tab. 1, Tab. 2, and Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Results on UnivFD dataset. Results show that our proposed method achieves superior performance compared to the UnivFD and FatFormer. Notably, without the biased interpretation introduced by coarse-grained text prompts, SDD surpasses the latest state-of-the-art method FatFormer by the mean AP \\((ap_{m})\\) of \\(0.34\\%\\) and the mean acc \\((acc_{m})\\) of \\(2.75\\%\\). Moreover, compared with methods based on relatively monotonous forgery features in Tab. 1 and Tab. 2, our approach can outperform all of them with a large improvement. The above evidence indicates effective combination of visual concepts and forgery features can contribute model to extract sufficient forgery patterns and eliminate the superfluous features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Results on SynRIS dataset. As shown in Tab. 3, when confronted with high-fidelity images generated by text-to-image models, methods leveraging pre-trained vision-language models, such as UnivFD and FatFormer, lose their competitiveness. In contrast, NPR, which focuses on neighboring pixel relationships, retains its edge. We assume that current generative methods grasp the relationships between visual information and semantic concepts in images but cannot refine local forgery details at the pixel level. Considering that excessive reliance on concepts misses abnormal pixel arrangements and focusing on monotonous forgery patterns can cause overfitting, our detector, which emphasizes low-level features with visual concepts, is trained on lower-fidelity fake images generated by Stable Diffusion [52] to capture concept-specific lacunae. We follow the evaluation protocol from SynRIS [5]. In comparison, our detector achieves an impressive mean AUROC of \\(95.1\\%\\), surpassing the state-of-the-art method by \\(4.8\\%\\). This demonstrates its superior ability to tackle the challenges posed by evolving generative models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18394" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.126, + 0.091, + 0.312, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.088, + 0.477, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.089, + 0.619, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.761, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.089, + 0.907, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.227, + 0.915, + 0.256 + ], + "angle": 0, + "content": "Figure 5. Performance of function on Figure 6. T-SNE visualization of real and fake images [47]. The feature space is based on our adaptive weights. classifier. Each randomly samples 500 real and 500 fake images." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.262, + 0.482, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.454 + ], + "angle": 0, + "content": "Figure 7. The showcase of attention on images which are input into our model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.241, + 0.479 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.547 + ], + "angle": 0, + "content": "We perform comprehensive ablation studies on the UnivFD dataset under the original experimental configurations, reporting the mean accuracy \\((acc_{m})\\) and mean average precision \\((ap_{m})\\) as the primary evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Effect of Each Component: We study the effects of removing STS module, CFDL module, and feature enhancement in our method. The results, presented in Tab. 4, demonstrate that these components are essential for improving performance in generalization on unseen models. This empirical finding suggests that the CFDL effectively captures forgery discrepancies associated with semantic concepts, while the enhancer module plays a crucial role in identifying robust forgery artifacts. The collaboration of all modules enhances the model's ability to distinguish between real and fake images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Effect of function on adaptive weights: To check how well the proposed function works in the SDD, we select two conventional functions for comparison: \\( f(x) = |x| \\) and \\( f(x) = x^2 \\). The corresponding results are presented in Fig. 5. We find that our proposed function yields improvements in both \\( ap_m \\) and \\( acc_m \\) compared to the selected functions. These results demonstrate that our proposed function is conducive to capturing robust and distinctive forgery features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Visualization of learned latent space: As shown in Fig. 6, the input images can be distinctly categorized into two clusters: real and fake. Nevertheless, why does the di" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.431 + ], + "angle": 0, + "content": "vided boundary of ProGAN appear ambiguous in contrast to other models? Additionally, why do the real clusters of CycleGAN and StyleGAN separate from each other? We attribute these to the influence of visual semantic concepts. Perceptively, with the supervision of visual semantic concepts, the learned boundary of ProGAN is more complex and nuanced, rather than just simple straight lines or curves. Similarly, the images generated by StyleGAN and CycleGAN are projected into the corresponding semantic concept distribution and then separated from the real images based on the visual semantic concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.599 + ], + "angle": 0, + "content": "Visualization of attention on images: We apply Class Activation Mapping (CAM) [58] to visualize the learned representations. From our perspective, Fig. 7 illustrates that with the aid of semantic information, our model can focus on different regions of fake images, including the background, local object regions, and marginal details. This suggests that our fine-grained model is capable of capturing intricate discrepancies generalized to unseen models. Notably, the real images nearly always show no forgery discrepancy regions, which demonstrates the effectiveness of the reconstruction loss in the forgery detection task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.634, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.893 + ], + "angle": 0, + "content": "In this paper, we propose a novel method, SDD, for generalizable forgery image detection. The findings show that our method establishes a new state-of-the-art in detecting images generated by generative models from different periods, which underscores its robustness and superior generalization capability. To the best of our knowledge, in pre-trained vision-language paradigms, our approach is the first to rely solely on visual information, without text prompts. Based on experimental results, we conclude that leveraging sampled tokens and reconstruction techniques effectively aligns the visual semantic concept space with the forgery space. Additionally, refining low-level forgery features under the supervision of visual semantic concepts enhances the performance of forgery detection. Although SDD performs well across various generative methods, there is still room for improvement as generative technologies continue to advance. Future research will explore this further." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18395" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.116, + 0.484, + 0.147 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (Grants Nos. 62372238,62476133)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[1] V.Ya. Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. ArXiv, abs/2312.03511, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[2] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. ArXiv, abs/2107.03006, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[5] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fake inversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10759-10769, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[6] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European Conference on Computer Vision, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[7] Cheng Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- \\(\\delta\\): Fast and controllable image generation with latent consistency models. ArXiv, abs/2401.05252, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[9] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1520-1529, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Yunjey Choi, Min-Je Choi, Mun Su Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shutao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Joel Cameron Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. ArXiv, abs/2003.08685, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[15] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick von Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss. ArXiv, abs/2401.02677, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[16] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloud-vae: Unsupervised feature learning for 3d point clouds from multiple angles by joint self-reconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441-10450. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[17] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. ArXiv, abs/1710.10196, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Prabhat Kumar, Mayank Vatsa, and Richa Singh. Detecting face2face facial reenactment in videos. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2578-2586, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[21] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. ArXiv, abs/2402.17245, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[22] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4219-4228, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[23] Zhenghong Li, Hao Chen, Jiangjiang Wu, Jun Li, and Ning Jing. Segmind: Semisupervised remote sensing image semantic segmentation with masked image modeling and contrastive learning method. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[24] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 772-781, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Yao Zhao, and Jingdong Wang. Forgery-aware adaptive transformer for generalizable synthetic image detection." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18396" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10770-10780, 2023. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[27] Xinhai Liu, Xinchen Liu, Zhizhong Han, and Yu-Shen Liu. Spu-net: Self-supervised point cloud upsampling by coarse-to-fine reconstruction with self-projection optimization. IEEE Transactions on Image Processing, 31:4213-4226, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[28] Zhengzhe Liu, Xiaojuan Qi, Jiaya Jia, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8066, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.458, + 0.277 + ], + "angle": 0, + "content": "[29] Midjourney. Midjourney, n.d. Accessed: 2025-03-04. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[30] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, B. S. Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H. Bappy, and Amit K. Roy-Chowdhury. Detecting gan generated fake images using co-occurrence matrices. ArXiv, abs/1903.06836, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[33] Pablo Pernias, Dominic Rampas, Mats L. Richter, and Marc Aubreville. Wuerstchen: Efficient pretraining of text-to-image models. ArXiv, abs/2306.00637, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[35] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 763-772, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, V.Ya. Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. In Conference on Empirical Methods in Natural Language Processing, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[39] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "autoencoder reconstruction error. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9130-9140, 2024. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[42] Segmind. Announcing ssd-1b: A leap in efficient t2i generation., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[43] Minghe Shen, Hongping Gan, Chao Ning, Yi Hua, and Tao Zhang. Transc: A transformer-based hybrid architecture for image compressed sensing. IEEE Transactions on Image Processing, 31:6991-7005, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[44] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12105-12114, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[45] Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28130-28139, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[46] Devavrat Tomar, Manana Lortkipanidze, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. Self-attentive spatial adaptive normalization for cross-modality domain adaptation. IEEE Transactions on Medical Imaging, 40:2926-2938, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[47] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[48] Tim van Erven and Peter Harremoës. Rényi divergence and kullback-leibler divergence. IEEE Transactions on Information Theory, 60:3797-3820, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[49] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8692-8701, 2019. 1, 7, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[51] Zhendong Wang, Jianmin Bao, Wen gang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18397" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "diffusion-generated image detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22388-22398, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[52] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. In Annual Meeting of the Association for Computational Linguistics, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for fast image restoration and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1934-1948, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[54] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1593-1603, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[55] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. ArXiv, abs/2306.10012, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[56] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7545-7556, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[57] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. 2019 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[58] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242-2251, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18398" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_origin.pdf b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78b6443bc46ac81a250161478b26982eb814b957 --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/d257d4bc-770d-463b-8ab1-d4cb1749eea8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8c6ea47e387de77af490cbdebb7e4566f3efb6bc76f15398e633ef5e18f096f +size 4618878 diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/full.md b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0ceb0af72a26e592eb393d3b12d73572e962080 --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/full.md @@ -0,0 +1,344 @@ +# Semantic Discrepancy-aware Detector for Image Forgery Identification + +Ziye Wang Minghang Yu Chunyan Xu* Zhen Cui Nanjing University of Science and Technology, Nanjing, China + +{wzynjust,mhyu,cyx}@njust.edu.cn，zhen.cui@bnu.edu.cn + +# Abstract + +With the rapid advancement of image generation techniques, robust forgery detection has become increasingly imperative to ensure the trustworthiness of digital media. Recent research indicates that the learned semantic concepts of pre-trained models are critical for identifying fake images. However, the misalignment between the forgery and semantic concept spaces hinders the model's forgery detection performance. To address this problem, we propose a novel Semantic Discrepancy-aware Detector (SDD) that leverages reconstruction learning to align the two spaces at a fine-grained visual level. By exploiting the conceptual knowledge embedded in the pre-trained vision-language model, we specifically design a semantic token sampling module to mitigate the space shifts caused by features irrelevant to both forgery traces and semantic concepts. A concept-level forgery discrepancy learning module, based on reconstruction, enhances the interaction between semantic concepts and forgery traces, effectively capturing discrepancies under the concepts' guidance. Finally, the low-level forgery feature enhancement integrates the learned concept-level forgery discrepancies to minimize redundant forgery information. Experiments conducted on two standard image forgery datasets demonstrate the efficacy of the proposed SDD, which achieves superior results compared to existing methods. The code is available at https://github.com/wzy111111/SSD. + +# 1. Introduction + +With the thriving of generative AI technologies, like Generative Adversarial Networks (GANs) [14] and diffusion models [2], the images generated by these models can easily create confusion by passing off the spurious as genuine. Therefore, it is crucial to develop a universal method for detecting fake images to mitigate the widespread dissemination of disinformation. + +Pioneering research [26, 32] has shown that projecting + +![](images/ebad5eb90f3ca9bb87c11fd2b66f803687de28196141364d2508083b6c397c5a.jpg) + +![](images/81954e406961bea5aa5a7ed67c1b28aa41cf3ad135b9f7908fb60b5698e8f433.jpg) +Figure 1. The phenomenon of misalignment between semantic concept space and forgery space. Since $\cos \theta$ can reflect the similarity of image descriptions, we model the feature space in polar coordinates. As the semantic concept space in [32] is frozen, fake samples sharing similar concepts with real ones can be easily misclassified. With forgery-adaptive space like [26], the model can correctly distinguish between them based on re-learned forgery features. Nevertheless, due to the semantic concept bias introduced by coarse text prompts, the target samples may be projected into an inaccurate semantic concept dimension, causing them to drift away from the real source samples along the fake dimension. + +images in a joint embedding space of texts and images can effectively capture discrepancies between fake and real images. In contrast, methods [6, 13, 44, 50] overlooking the interplay between forgery traces and semantic concepts perform poorly when confronted with unseen generative models. + +To investigate the visual semantic concepts of pretrained models, we conduct a statistical analysis of the output features from CNNSpot [50] and CLIP-ViT [32] (See Appendix A for more details). Under different categories, CNNSpot exhibits a synchronized difference between real and fake features in its training space. However, when tran + +![](images/3211c3b1e370bfa02685bd229d7de6b14fbe69599dd492db2cb33e55aba0b7cb.jpg) +Figure 2. Different paradigms of image forgery identification with pre-trained vision-language model. (a) Fine-tune the frozen model only by fully connected (FC) layers [32]. (b) Prompt-based designs are tuned on text prompts and contrastive objectives [26]. (c) Our paradigm incorporating visual clues can capture fine-grained forge traces by reconstruction learning. + +sitioning to the CLIP's space, these differences become inconsistent. From this, we infer a nuanced relationship between semantic concepts and forgery traces: Different semantic concepts may guide the model to uncover distinct forgery traces. + +Intuitively, relying on a frozen pre-trained vision-language model like UnivFD [32] is essential, but this tends to overlook fine-grained forgery details. Although Fat-Former [26] achieves a substantial enhancement in generalization by employing the forgery-aware adaptive transformer, we observe that soft prompts based on simple [CLASS] embeddings have an intrinsic limitation in their semantic description granularity (See Appendix B for more details). The constrained breadth of the conveyed concepts may lead the detection toward incorrect predictions. This limitation highlights a misalignment between the visual semantic concept space and the target forgery space, as illustrated in Fig. 1. + +To address this, one empirical approach is to design more detailed text descriptions, but this method struggles to describe all visual forgery details due to the limited length of texts and brings more computational overhead. Drawing from the aforementioned findings and analysis, we make a first attempt to align the CLIP's visual semantic concept space with the forgery space by reconstructing semantic features. + +We develop a vision-based paradigm, as outlined in Fig. 2. First, employing a pre-trained model only with nearest neighbor or linear probing (e.g. UnivFD [32], Fig. 2 (a)) is suboptimal for image forgery detection. Second, modifying the pre-trained model with task-specific prompts (e.g. FatFormer [26], Fig. 2 (b)) may favor models biased towards any particular semantic concept. These studies pave the way for exploring pre-trained space with rich semantic concepts. Inspired by image reconstruction [43, 53], our paradigm amplifies the concept-level forgery discrepancies of forgery images, which empowers the model to detect sus + +picious forgery traces with the assistance of semantic concepts. + +In this work, we present a novel Semantic Discrepancy-aware Detector (SDD) to accurately align the semantic concept space and the forgery space. Firstly, to mitigate interference from features unrelated to learned semantic concepts and forgery traces, we divide the real images into non-overlapping blocks and feed them to the frozen CLIP [36] to obtain diverse semantic patch tokens. These tokens acting as visual clues smoothly align the semantic concepts' space and forgery space. It is noteworthy that these tokens sampled by JS divergence are universally representative of the real semantic distribution. Then, the visual clues are fused into a concept-level forgery discrepancy module. Unlike FatFormer, LoRA layers are incorporated into the image encoder. The goal is to preserve the completeness and diversity of the learned semantic concepts of CLIP, while the forgery features sharing similar semantic concepts should be highlighted. During reconstruction, we only narrow the reconstruction gap for real samples to reinforce the reconstructed discrepancies of the synthetic images. Finally, we present low-level forgery feature enhancement to let the reconstruction difference map enhance the extraction of the highly generalizable forgery features while introducing minimal additional parameters. The main challenge is how to capture forgery features with strong semantic concept correlation and features with high forgery relevance but weak semantic concept ties to ensure the model converges to powerful features. Motivated by this, we apply convolutional modules and adaptive weight parameters to avoid over-relying on semantic concepts. + +We thoroughly evaluate the generalization performance of our model on a UnivFD benchmark [32] and a SynRIS benchmark [5]. Surprisingly, our method achieves superior performance by a $ap_{m}$ of 98.51% and a $acc_{m}$ of 93.61% on the UnivFD benchmark [32] and an average AUROC of 95.1% on the SynRIS benchmark [5]. In summary, our contributions are as follows: + +- We propose a robust model (SDD) for forgery detection, specifically designed to align the visual concept space and forgery space in terms of visual information. +- We sample semantic tokens to mitigate the space shifts and align the two spaces through reconstruction learning. Additionally, we strengthen low-level forgery features to enhance the model's robustness. +- Our method achieves superior performance on two benchmarks, demonstrating its superior capability in comparison to existing approaches. + +# 2. Related Work + +AI-generated Images Detection. Extensive efforts have been devoted to enhancing the performance of AI-generated image detection. Early works like [25, 44, 45] tend to mine + +![](images/94dd20070c1e79972c6f8254b876b80beacef33b3db35c97c2297de23b51f24e.jpg) +Figure 3. The architecture of SDD. First, Next, we sample semantic tokens from real images to learn features related to both concepts and forgery. the input images are mapped into a joint space of visual semantic concepts and forgery, which are transformed into learnable features $V_{H}$ . Then, we use transformer-based encoder and decoder to get reconstructed features $\mathcal{R}_f$ . A reconstruction difference map $\mathcal{D}_S$ is obtained and goes through the multi-scale convolutional network to refine forgery features. Finally, we concatenate the CLIP's CLS token with this output along the same dimension for classification. The whole system is trained by jointly minimizing the binary cross-entropy loss $L_{bce}$ , the reconstruction loss $L_{r}$ , and the triplet loss $L_{tri}$ + +the common forgery traces between all real and fake images, such as noise patterns, texture statistics, and frequency signals. As an illustration, Liu et al. [24] designed a network that learns the consistent noise patterns in images for fake detection. Liu et al. [28] proposed to leverage the gram matrix to discover the global anomalous texture of fake images. An effective approach [13] demonstrated that frequency representation is an important factor in improving fake detection performance. However, these differences are rigorously specific to the monotonous features, which contribute to the issue of overfitting. Cutting-edge research [26, 32] shifted attention toward the semantic properties of images. Ojha et al. [32] showed that projecting images into the feature space of pre-trained vision-language model enables strong generalization ability. To build generalized forgery representations, Liu et al. [26] constructed forgery adaptive space by a forgery-aware adapter. The above research [5, 26, 32] has suggested that concept attributes are vital in the image forgery detection task. Assuming that diffusion-based models leave distinct forgery traces that are characteristic of specific concept distributions, we aim to extract robust forgery features guided by semantic concepts, rather than suppressing them. Therefore, even "useless" information can be useful by providing significant certainty about the content of the image. + +Reconstruction Learning. Reconstruction learning has + +great potential in unsupervised representation learning [16, 27]. Some works [5, 39] utilized reconstruction learning to reveal the nuances between real and fake images. For example, Wang et al. [51] found that reconstructing images by DDIM exposes an error between real images and their reconstructed replica. The new synthetic image detection method[5] used text-conditioned inversion maps to learn internal representations, which is conducive to predicting whether an image is fake. Ricker et al. [39] offered a simple detection approach by applying AE to measure the reconstruction error. Notably, these works are committed to reconstructing the distributions of both real and fake samples by leveraging generative models. Unlike previous works, we focus solely on reconstructing real images in the finetuned CLIP space in light of the authenticity and richness of semantic concepts. The distribution of real samples, learned from pre-trained vision-language model, helps to define an optimal boundary, thus alleviating overfitting. + +# 3. Methodology + +Our goal is to align forgery and visual semantic concept spaces using reconstruction techniques for robust and generalizable synthetic image detection. To achieve this, we introduce a fine-grained model named Semantic Discrepancy-aware Detector (SDD). Building on prior works, we harness the generalization capability of vision-language mod + +els. Semantic concept space: The ideal joint embedding space of images and texts with four properties: semantic alignment, modality invariance, locality consistency, and structure preservation. Forgery space: The ideal space covers forgery traces. Notably, we derive semantic concept space via a vision language model pretrained solely on real images; thus we treat the two spaces as independent. + +First, the Semantic Tokens Sampling (STS) module utilizes Jensen-Shannon (JS) divergence to sample semantic patch tokens, which serve as a transitional bridge, facilitating model to establish the association between real and forgery images accurately. Next, the Concept-level forgery Discrepancy Learning (CFDL) module employs reconstruction learning to explore the forgery discrepancies within the visual semantic concept space, which focuses on identifying subtle variations between reconstructed forgery features. Finally, the reconstruction difference map is trained with Low-level forgery Feature Enhancement module, which aims to refine forgery features with more visual details. + +# 3.1. Semantic Tokens Sampling + +Initially, we considered directly aligning the visual semantic concept space and forgery space by leveraging fine-grained reconstruction learning to model real and fake semantic distributions. However, this strategy would treat the differences in features unrelated to semantics and forgery as crucial factors for identifying image's authenticity. To eliminate these redundant features, we sample real semantic image patch tokens as visual clues to bridge real and forged semantic domains. This module enables the model to focus on concept-related forgery traces and highlight the distinctions between real and fake images. In a tangible way, the image encoder of CLIP: ViT-L/14 is adapted to transform a real image $x_{r}$ into a set of features $f_{r}$ , without the image CLS token. We define the transformation as: + +$$ +f _ {r} = \phi (x _ {r}), \tag {1} +$$ + +where $\phi (\cdot)$ is the CLIP:ViT-L/14's visual encoder, $x_{r}\in$ $\mathbb{I}_r^{H\times W\times 3}$ represents a real image characterized by a height of $H$ and a width of $W$ . Besides, $f_{r}\in \mathbb{R}^{N\times D}$ , where $N$ is the number of tokens and $D$ denotes the dimension of each patch token. + +Since integrating all real patch tokens into the image reconstruction module is computationally intensive and memory-consuming, it is urgent to select a suitable subset of these tokens. From a distribution perspective, the Jensen-Shannon (JS) divergence, derived from the Kullback-Leibler divergence [48], is a symmetric and finite metric that can effectively measure the similarity between tokens by quantifying differences in their distributions. + +To calculate the JS divergence between two tokens, both are converted into computable probability distribution space using the softmax function. Let $f_{s} \in \mathbb{F}_{r}^{M \times D}$ be the selected + +semantic patch tokens with the num of tokens $M = 1 / \delta$ and dimension $D$ in terms of sampling rate $\delta$ ( $0 \leq \delta \leq 1$ , $\delta$ is user-defined). Once the initial token $\tilde{r}$ and $\delta$ are determined, the JS divergence between $\tilde{r}$ and other tokens $r$ falls within the range [0, 1]. Subsequently, the sampling interval is split into $M$ equal segments with one token selected from each segment. As a consequence, the semantic tokens sampling module can be formulated as: + +$$ +\begin{array}{l} f _ {s} = \mathcal {S} \left(\mathbb {R} ^ {N \times D}, \delta\right) \\ = A _ {c} ^ {N _ {a} \times M} \times \mathbb {R} ^ {N \times D}, \\ \end{array} +$$ + +s.t. JS(softmax $(\tilde{r})$ ,softmax $(r)) = \frac{i}{M}$ , if $a_{ij} = 1$ (2) + +$$ +\sum_ {i = 0} ^ {N _ {a}} \sum_ {j = 0} ^ {M} a _ {i j} = M; \sum_ {j = 0} ^ {M} a _ {i j} = 1 \text {o r} 0, +$$ + +$$ +i = 1, \dots , N _ {a}; j = 1, \dots , M, +$$ + +where $S(\cdot, \cdot)$ represents the sampling process. $A_{c}^{N_{a} \times M}$ is a constraint matrix of size $N_{a} \times M$ whose element $a_{ij}$ is constrained to the binary pattern of $\{0, 1\}$ . Here JS $(\cdot)$ refers to the Jensen-Shannon divergence, $N_{a}$ denotes the total number of real image patch tokens sampled from the training dataset of UnivFD and $M$ represents the required subset size. The softmax $(\cdot)$ is the softmax function. The sampling tokens help the reconstruction module avoid becoming biased towards any particular forgery-unrelated distribution. Meanwhile, it avoids the semantic bias often introduced by text prompts, since the tokens are evenly distributed in a unified CLIP space. + +# 3.2. Concept-level Forged Discrepancy Learning + +A few words alone can hardly paint a picture. We argue that the fine-grained visual details can uncover more forgery traces concealed in the images. As such, we mix sampling tokens with extracted features and capitalize on reconstruction learning to compensate for the omission of forgery traces brought by coarse prompts. As previous work [26] has demonstrated that the pre-trained vision-language model necessitates fine-tuning to adapt to the forgery detection task. Therefore, we integrate LoRA [17] with the CLIP-ViT model to capture discriminative forgery features by making use of the bread semantic concepts. This method, denoted as LoRA-CLIP [54], is more streamlined and flexible. Given an input image $\mathcal{I} \in \mathbb{I}^{H \times W \times 3}$ , we can get high-level visual features $V_{H}$ , as follows: + +$$ +V _ {H} = \mathcal {F} _ {L o R A} (\mathcal {I}). \tag {3} +$$ + +Here $\mathcal{F}_{LoRA}$ refers to the CLIP image encoder fine-tuned by LoRA. The reconstruction module of CFDL encompasses two submodules, i.e., transformer-based encoder and decoder. Thanks to the transformer's capability of long-range relationship modeling, we capitalize on the multi-head attention (MHA) mechanism, the core mechanism of + +the transformer, to obtain more discriminative distribution by utilizing contextual information, which is set as: + +$$ +\left\{ \begin{array}{r l} \operatorname {h e a d} _ {i} & = \operatorname {A t t n} \left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\right) \\ & = \operatorname {S o f t m a x} \left(\frac {Q W _ {i} ^ {Q} \left(K W _ {i} ^ {K}\right) ^ {\top}}{\sqrt {d}}\right) V W _ {i} ^ {V}, \\ \operatorname {M H A} (Q, K, V) & = \operatorname {C o n c a t} \left(\operatorname {h e a d} _ {1}, \dots , \operatorname {h e a d} _ {h}\right) W ^ {O}, \end{array} \right. \tag {4} +$$ + +where $Q$ (Query), $K$ (Key), $V$ (Value) refer to the input, $W_{i}^{Q}, W_{i}^{K}, W_{i}^{V}$ separately denote the corresponding weights of linear projection, $\mathrm{Attn}(\cdot)$ denotes the function of the scaled dot product, Softmax is the softmax function, $d$ refers to the dimension of input, Concat ( $\cdot$ ) represents the concatenation used to stitch the discrete attention outputs of head $1 \sim h$ together. + +To amplify the discrepancy between a fake image and its reconstructed counterpart, the sampled visual clues are employed for the initial processing by the encoder. The encoder's process can be formulated as follows: + +$$ +R _ {1} = \operatorname {L N} \left(\mathrm {M H A} \left(f _ {s}, V _ {H}, V _ {H}\right)\right), \tag {5} +$$ + +$$ +R _ {2} = \operatorname {L N} \left(\mathrm {M H A} \left(R _ {1}, V _ {H}, V _ {H}\right)\right), \tag {6} +$$ + +where $\mathrm{LN}(\cdot)$ denotes the Layer Normalization. Then, the encoder's outputs used as queries are injected into the decoder to get the final reconstructed features, which are similar to the encoder process and perform the following operation: + +$$ +R _ {3} = \mathrm {L N} (\mathrm {M H A} (R _ {2}, R _ {2}, R _ {2})), \tag {7} +$$ + +$$ +R _ {e} = \mathrm {L N} (\mathrm {M H A} \left(R _ {1}, R _ {3}, R _ {3}\right)). \tag {8} +$$ + +During the reconstruction process, we just calculate the reconstruction loss $\mathcal{L}_r$ between the real input features and their reconstructed counterparts $\mathcal{R}_e$ within a mini-batch as follows: + +$$ +\mathcal {L} _ {r} = \frac {1}{B} \sum_ {i = 0} ^ {B} \operatorname {M S E} \left(R _ {e}, V _ {H}\right), \tag {9} +$$ + +where $\mathrm{MSE}(\cdot ,\cdot)$ is mean squared error. Facilitating $\mathcal{L}_r$ encourages preserving the completeness and richness of the visual semantic concept space and highlighting the concept-related forgery features. Given the reconstructed features $R_{f}$ and the original feature $f_{r}$ , the reconstruction difference map can be formally expressed as: + +$$ +\mathcal {D} _ {s} = \left| R _ {f} - f _ {r} \right|, \tag {10} +$$ + +where $|\cdot |$ denotes the absolute value function. + +# 3.3. Low-level Forgery Feature Enhancement + +Existing methods based on pre-trained vision-language models [26, 32] overlook the importance of concept + +![](images/de568ffeb00d073298b5f7dc1c0e06b896ff4b86f32ec8d1e4c74be043a2f038.jpg) +Figure 4. The curve of exponential inverse. In the "fast" interval, the value drops sharply. In the "low" interval, the curve flattens out, showing a decay towards 0. + +weakly-related features. We believe that a thorough alignment between the visual semantic concept space and the forgery space should include the exploration of concept weakly-related forgery features. To eliminate redundant forgery features, we come up with a novel feature enhancement that refines low-level forgery features. Empowered by the reconstruction difference map, our detector orchestrates the extraction of multi-scale features with exceptional robustness and markedly enhanced effectiveness. As shown in Fig. 3, the enhancer follows the typical architecture of a convolutional network. It involves the repeated application of convolutions, each followed by a batch normalization (BN) and a rectified linear unit (ReLU). For a given stage $n$ , $F(n)$ ( $N = 1,2,3$ ) corresponds to its output features. Then, We deconvolve the semantic difference map $D_{s}$ to the shape same as $F(N)$ and perform pixel-wise multiplication with $F(n)$ to get $F'(n)$ as: + +$$ +F ^ {\prime} (n) = \operatorname {d e c o n v} (F (n)) \otimes D _ {s}, \tag {11} +$$ + +where $\otimes$ is the element-wise multiplication, deconv $(\cdot)$ represents deconvolution operation and $\mathcal{F}'(n)$ is the low-level feature aggregated with semantic information. To further enhance the reliability of the extracted features, we compute an adaptive weight coefficient $\frac{1}{e_n}$ to indicate the importance of $D_{s}$ to $F(n)$ : + +$$ +\frac {1}{e _ {n}} = \frac {1}{e ^ {\left| F ^ {\prime} (n) - F (n) \right|}}. \tag {12} +$$ + +Here we explain the role of the exponential inverse through Fig. 4. As $x$ grows large, the curve of $e^x$ becomes flatter. Therefore, in the "fast" interval, forgery features with a significant divergence from the semantic difference map will be assigned smaller weights, which mobilizes the network to capture concept strongly-related features. However, in the "low" interval, features strongly associated with forgery can avoid being misguided by semantic concepts, which indicates that the order of importance is reversed. Next, we have the attended output features $F_{low}$ + +
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDallemAP
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200Steps200w/cfg100Steps100w/ CFG10027502710
CNN-SpotCVPR2020100.093.4784.5099.5489.4998.1589.0273.7559.4798.2498.4073.7270.6271.0070.5480.6584.9182.0770.5983.58
PatchForECCV202080.8872.8471.6685.7565.9969.2576.5576.1976.3474.5268.5275.0387.1086.7286.4085.3783.7378.3875.6777.73
Co-occurrenceElect.Imag.99.7480.9550.6198.6353.1167.9959.1468.9860.4273.0687.2170.2091.2189.0292.3989.3288.3582.7980.9678.11
Freq-specWIFS201955.39100.075.0855.1166.08100.045.1847.4657.1253.6150.9857.7277.7277.2576.4768.5864.5861.9267.7766.21
DireICCV2023100.083.5981.5096.5081.7099.8895.7362.5169.9897.3198.6279.5375.5273.4276.4586.2889.0088.3451.3583.54
UnivFDCVPR2023100.099.4699.5997.2499.9899.6082.4561.3279.0296.7299.0087.7799.1492.1599.1794.7495.3494.5797.1593.38
NPRCVPR2024100.099.5096.5099.8096.80100.092.2073.1078.7087.2064.8065.8099.8099.8099.8099.7099.8099.8098.6092.19
FatFormerCVPR2024100.0100.099.9899.75100.0100.097.9997.9481.2399.8499.9391.9299.8399.2299.8999.2799.5099.3399.8498.18
Ours100.099.7799.9399.4899.9899.9797.2397.9193.1099.7999.9692.0699.8898.9599.9298.0698.2997.7399.8198.51
+ +Table 1. Average precision comparisons with different methods on the UnivFD dataset. We replicate the results of CNNSpot, Patchfor, Co-occurrence, Freq-spec, and UnivFD from the paper[32]. In addition, we obtained the results for Dire, NPR and FatFormer using either the official pre-trained models or our re-implemented versions. Red and underline indicates the best and the second best result, respectively. + +
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDalleAvg-acc
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200 Steps200 wcfg100 Steps100 2750 27100 10
CNN-SpotCVPR202099.9985.2070.2085.7078.9591.7053.4766.6748.6986.3186.2660.0754.0354.9654.1460.7863.8065.6655.5869.58
PatchForECCV202075.0368.9768.4779.1664.2363.9475.5475.1475.2872.3355.3067.4176.5076.1075.7774.8173.2868.5267.9171.24
Co-occurrenceElect.Imag.97.7063.1553.7592.5051.1054.7057.1063.0655.8565.6565.8060.5070.7070.5571.0070.2569.6069.9067.5566.86
Freq-specWIFS201949.9099.9050.5049.9050.3099.7050.1050.0048.0050.6050.1050.9050.4050.4050.3051.7051.4050.4050.0055.45
DireICCV202399.8673.4760.6872.3965.1593.6088.8652.7856.3990.0794.0561.0559.3559.9560.6569.3072,7071.0052.7571.19
UnivFDCVPR2023100.098.5094.5082.0099.5097.0066.6063.0057.5059.5072.0070.0394.1973.7694.3679.0779.8578.1486.7881.38
NPRCVPR202499.8092.0089.5096.3087.6099.7079.4061.4070.6074.5057.1055.2397.4098.7097.9097.0097.9097.0088.8086.20
FatFormerCVPR202499.8999.3699.5097.1299.4399.7593.2581.3968.0469.4769.4776.0098.5594.8598.6094.3094.6094.1598.7090.86
Ours99.8895.7696.7098.0898.4699.1791.8283.6177.4595.4096.4779.5598.0594.6098.2592.2093.3591.8098.0093.61
+ +Table 2. Accuracy comparisons with different methods on the UnivFD dataset. + +by the residual connection: + +$$ +F _ {l o w} (n) = F ^ {\prime} (n) + \frac {F ^ {\prime} (n)}{e _ {n}}. \tag {13} +$$ + +For optimizing the anchor features $\tilde{f}_a$ of the enhancer, the following triplet loss [35] is employed to bring positive samples $f_{p}$ closer while pushing negative samples $f_{n}$ apart: + +$$ +\mathcal {L} _ {t r i} = \max \left(0, d \left(f _ {p}, f _ {a}\right) - d \left(f _ {n}, f _ {a}\right) + \alpha\right), \tag {14} +$$ + +where $d(\cdot)$ represents the Euclidean distance between samples and $\alpha$ is the margin. + +On top of that, we concatenate the LoRA-CLIP's CLS token $T_{CLS}$ with $F_{low}$ along the same dimension to yield the refined representation $F_{out}$ . This ensures that forged features exhibit distinctiveness across different semantic identities while preserving their uniformity within similar semantic identities. The process is formulated as: + +$$ +F _ {o u t} = F _ {l o w} \| T _ {C L S}, \tag {15} +$$ + +where $F^{out}$ is strategically integrated with a linear classifier to enable the execution of binary classification. Eventually, the total loss function $\mathcal{L}$ of the proposed framework can be defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {b c e} + \lambda_ {1} \mathcal {L} _ {t r i} + \lambda_ {2} \mathcal {L} _ {r}, \tag {16} +$$ + +where $\mathcal{L}_{bce}$ presents the binary cross-entropy loss and $\mathcal{L}_{tri}$ is triplet loss. $\mathcal{L}_{tri}$ and $\mathcal{L}_r$ are scaled by the hyperparameters $\lambda_{1}$ and $\lambda_{2}$ , respectively. + +# 4. Experiments + +# 4.1. Experiment Setups + +Datasets: We follow the protocol described in [32], using ProGAN's real and fake images as training data. Additionally, we adopt the protocol from [5], where the training data is composed of fake Stable Diffusion v1 images [52] and random real LAION images [41]. The UnivFD dataset [32] cover a broad range of generative models, primarily including GANs and diffusion models, such as ProGAN [18], StyleGAN [19], BigGAN [4], CycleGAN [59], StarGAN [10], GauGAN [46], CRN [9], IMLE [22], SAN [11], SITD [7], DeepFakes [20], Guided [12], Glide [31], LDM [40] and DALL-E [37]. The SynRIS dataset [5] is designed to avoid bias toward any specific topic, theme, or style and contains high-fidelity images generated by text-to-image models, such as Kandinsky2 [38], Kandinsky3 [1], PixArt- $\alpha$ [8], SDXL-DPO [49], SDXL [34], SegMoE [23], SSD-1B [42], Stable-Cascade [33], Segmind-Vega [15], and Würstchen2 [33], Midjourney [29], DALL.E 3 [3] + +
MethodsCNN-SpotFreq-specDireUniv FDNPRFatF ormerFakeIn versionPatch ForOurs
Kandinsky260.057.071.656.297.575.669.953.597.13
Kandinsky365.945.774.961.493.780.174.351.494.8
PixArt-α62.756.481.564.789.575.373.049.690.2
SDXL-DPO84.369.869.970.297.686.088.154.595.5
Segmind-Vega74.265.381.062.397.182.481.153.197.0
SDXL81.461.286.266.396.085.180.763.997.6
Seg-MoE66.354.671.962.093.670.871.349.897.6
SSD-1B72.667.879.862.899.670.179.461.299.8
Stable-Cascade70.562.174.168.297.681.674.957.499.2
Würstchen261.063.374.269.790.972.970.547.298.2
Midjourney63.050.972.459.258.573.666.453.790.0
Playground58.252.367.958.793.181.462.554.192.8
DALL-E371.659.980.848.069.179.275.950.185.9
Average68.658.975.962.390.378.074.553.895.1
+ +Table 3. AUROC comparisons with different methods on the SynRIS dataset. We retrieve the results of CNNSpot, UnivFD, and FakeInversion from [5] and obtain the results for Dire, NPR, Fat-Former, and PatchFor using re-implemented models. Red and underline indicate the best and the second best result, respectively. + +and playground [21]. + +Metrics: As standard evaluation metrics, the average precision (AP), the accuracy (ACC) and AUCROC are considered to measure the effectiveness of different methods. + +Baselines: In our experiments, we perform thorough comparisons with state-of-the-art methods, as follows: 1) CNNSpot [50]: The method relies only on one CNN generator. 2) PatchFor [6]: The method performs detection on a patch level. 3) Co-occurrence [30]: The method converts input images into co-occurrence matrices for classification. 4) Freq-spec [57]: The method employs the frequency spectrum of images. 5) Dire [51]: The method exploits the error between an input image and its reconstruction counterpart. 6) UnivFD [32]: The method uses the pre-trained language-vision model to determine the authenticity of images. 7) NPR [45]: The method captures the generalized artifacts according to the local interdependence among image pixels. 8) FatFormer [26]: The method is aimed at extracting forgery-adaptive features based on UnivFD. 9) Fakeinversion [5]: The method employs text-conditioned inversion maps extracted from Stable Diffusion. + +Implement details: Our training and testing settings are adapted from the approach outlined in the previous study [26] with several key modifications. Specifically, early stopping was employed during model training, with an initial learning rate of $1 \times 10^{-4}$ and a batch size of 32. Additionally, the Lora layers are configured with hyperparameters $lora_{r} = 6$ , $lora_{\alpha} = 6$ , and a dropout rate of 0.8, + +
#STS moduleCFDL modulefeature enhancementUnivFD apmDataset accm
1✓97.3781.64
2✓✓97.4190.17
3✓✓97.3989.98
4✓✓✓98.5293.61
+ +Table 4. Ablation study of the proposed modules on the UnivFD Dataset. We show the mean accuracy $(acc_m)$ and average precision $(ap_m)$ . Red and underline indicate the best and the second-best result, respectively. + +while $\alpha$ is set to 8.0. The proposed method is implemented using Pytorch on 2 Nvidia GeForce RTX A6000 GPUs. + +# 4.2. Comparison Results + +The UnivFD dataset includes a diverse range of models, allowing for a comprehensive evaluation of our method across both GAN and diffusion generative models. In addition, the SynRIS dataset provides images generated by cutting-edge generative models. The overall experimental results are presented in Tab. 1, Tab. 2, and Tab. 3. + +Results on UnivFD dataset. Results show that our proposed method achieves superior performance compared to the UnivFD and FatFormer. Notably, without the biased interpretation introduced by coarse-grained text prompts, SDD surpasses the latest state-of-the-art method FatFormer by the mean AP $(ap_{m})$ of $0.34\%$ and the mean acc $(acc_{m})$ of $2.75\%$ . Moreover, compared with methods based on relatively monotonous forgery features in Tab. 1 and Tab. 2, our approach can outperform all of them with a large improvement. The above evidence indicates effective combination of visual concepts and forgery features can contribute model to extract sufficient forgery patterns and eliminate the superfluous features. + +Results on SynRIS dataset. As shown in Tab. 3, when confronted with high-fidelity images generated by text-to-image models, methods leveraging pre-trained vision-language models, such as UnivFD and FatFormer, lose their competitiveness. In contrast, NPR, which focuses on neighboring pixel relationships, retains its edge. We assume that current generative methods grasp the relationships between visual information and semantic concepts in images but cannot refine local forgery details at the pixel level. Considering that excessive reliance on concepts misses abnormal pixel arrangements and focusing on monotonous forgery patterns can cause overfitting, our detector, which emphasizes low-level features with visual concepts, is trained on lower-fidelity fake images generated by Stable Diffusion [52] to capture concept-specific lacunae. We follow the evaluation protocol from SynRIS [5]. In comparison, our detector achieves an impressive mean AUROC of $95.1\%$ , surpassing the state-of-the-art method by $4.8\%$ . This demonstrates its superior ability to tackle the challenges posed by evolving generative models. + +![](images/0fc9b0eeef54d2a107d0b1c0f6d6c4e979f1491fe980fcbbc07675e0143222a1.jpg) +Figure 5. Performance of function on Figure 6. T-SNE visualization of real and fake images [47]. The feature space is based on our adaptive weights. classifier. Each randomly samples 500 real and 500 fake images. + +![](images/9ba89e1496d6b16bdccd57c575440594fea729573cfd64ec69d877d04d24eeca.jpg) + +![](images/53ea1b5f353a96c9738269ec99c6192074f98e4c9590ea3b568427e6f48b7455.jpg) + +![](images/12508e5e81b25cf63bb1033d94fba7fce7241302bce0d94e1bdf758f953e09b0.jpg) + +![](images/88bdb2dfb8a557c7e1f64a77e634075a41c2959d745cddd5ef7d47a86cb6ada7.jpg) + +![](images/3f926a12a11783a8f3703d8b1e19a10696a9a49ef46cb72f9c96a4c3486c118f.jpg) +Figure 7. The showcase of attention on images which are input into our model. + +# 4.3. Ablation study + +We perform comprehensive ablation studies on the UnivFD dataset under the original experimental configurations, reporting the mean accuracy $(acc_{m})$ and mean average precision $(ap_{m})$ as the primary evaluation metrics. + +Effect of Each Component: We study the effects of removing STS module, CFDL module, and feature enhancement in our method. The results, presented in Tab. 4, demonstrate that these components are essential for improving performance in generalization on unseen models. This empirical finding suggests that the CFDL effectively captures forgery discrepancies associated with semantic concepts, while the enhancer module plays a crucial role in identifying robust forgery artifacts. The collaboration of all modules enhances the model's ability to distinguish between real and fake images. + +Effect of function on adaptive weights: To check how well the proposed function works in the SDD, we select two conventional functions for comparison: $f(x) = |x|$ and $f(x) = x^2$ . The corresponding results are presented in Fig. 5. We find that our proposed function yields improvements in both $ap_m$ and $acc_m$ compared to the selected functions. These results demonstrate that our proposed function is conducive to capturing robust and distinctive forgery features. + +Visualization of learned latent space: As shown in Fig. 6, the input images can be distinctly categorized into two clusters: real and fake. Nevertheless, why does the di + +vided boundary of ProGAN appear ambiguous in contrast to other models? Additionally, why do the real clusters of CycleGAN and StyleGAN separate from each other? We attribute these to the influence of visual semantic concepts. Perceptively, with the supervision of visual semantic concepts, the learned boundary of ProGAN is more complex and nuanced, rather than just simple straight lines or curves. Similarly, the images generated by StyleGAN and CycleGAN are projected into the corresponding semantic concept distribution and then separated from the real images based on the visual semantic concepts. + +Visualization of attention on images: We apply Class Activation Mapping (CAM) [58] to visualize the learned representations. From our perspective, Fig. 7 illustrates that with the aid of semantic information, our model can focus on different regions of fake images, including the background, local object regions, and marginal details. This suggests that our fine-grained model is capable of capturing intricate discrepancies generalized to unseen models. Notably, the real images nearly always show no forgery discrepancy regions, which demonstrates the effectiveness of the reconstruction loss in the forgery detection task. + +# 5. Conclusion + +In this paper, we propose a novel method, SDD, for generalizable forgery image detection. The findings show that our method establishes a new state-of-the-art in detecting images generated by generative models from different periods, which underscores its robustness and superior generalization capability. To the best of our knowledge, in pre-trained vision-language paradigms, our approach is the first to rely solely on visual information, without text prompts. Based on experimental results, we conclude that leveraging sampled tokens and reconstruction techniques effectively aligns the visual semantic concept space with the forgery space. Additionally, refining low-level forgery features under the supervision of visual semantic concepts enhances the performance of forgery detection. Although SDD performs well across various generative methods, there is still room for improvement as generative technologies continue to advance. Future research will explore this further. + +# Acknowledgement + +This work is supported by the National Natural Science Foundation of China (Grants Nos. 62372238,62476133) + +# References + +[1] V.Ya. Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. ArXiv, abs/2312.03511, 2023. 6 +[2] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. ArXiv, abs/2107.03006, 2021. 1 +[3] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 6 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096, 2018. 6 +[5] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fake inversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10759-10769, 2024. 2, 3, 6, 7 +[6] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European Conference on Computer Vision, 2020. 1, 7 +[7] Cheng Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6 +[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- $\delta$ : Fast and controllable image generation with latent consistency models. ArXiv, abs/2401.05252, 2024. 6 +[9] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1520-1529, 2017. 6 +[10] Yunjey Choi, Min-Je Choi, Mun Su Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2017. 6 +[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shutao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 6 +[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 6 + +[13] Joel Cameron Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. ArXiv, abs/2003.08685, 2020. 1, 3 +[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 1 +[15] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick von Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss. ArXiv, abs/2401.02677, 2024. 6 +[16] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloud-vae: Unsupervised feature learning for 3d point clouds from multiple angles by joint self-reconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441-10450. IEEE, 2019. 3 +[17] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 4 +[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. ArXiv, abs/1710.10196, 2017. 6 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2018. 6 +[20] Prabhat Kumar, Mayank Vatsa, and Richa Singh. Detecting face2face facial reenactment in videos. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2578-2586, 2020. 6 +[21] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. ArXiv, abs/2402.17245, 2024. 7 +[22] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4219-4228, 2018. 6 +[23] Zhenghong Li, Hao Chen, Jiangjiang Wu, Jun Li, and Ning Jing. Segmind: Semisupervised remote sensing image semantic segmentation with masked image modeling and contrastive learning method. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 6 +[24] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, 2022. 3 +[25] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 772-781, 2021. 2 +[26] Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Yao Zhao, and Jingdong Wang. Forgery-aware adaptive transformer for generalizable synthetic image detection. + +2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10770-10780, 2023. 1, 2, 3, 4, 5, 7 +[27] Xinhai Liu, Xinchen Liu, Zhizhong Han, and Yu-Shen Liu. Spu-net: Self-supervised point cloud upsampling by coarse-to-fine reconstruction with self-projection optimization. IEEE Transactions on Image Processing, 31:4213-4226, 2020. 3 +[28] Zhengzhe Liu, Xiaojuan Qi, Jiaya Jia, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8066, 2020. 3 +[29] Midjourney. Midjourney, n.d. Accessed: 2025-03-04. 6 +[30] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, B. S. Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H. Bappy, and Amit K. Roy-Chowdhury. Detecting gan generated fake images using co-occurrence matrices. ArXiv, abs/1903.06836, 2019. 7 +[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 6 +[32] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 5, 6, 7 +[33] Pablo Pernias, Dominic Rampas, Mats L. Richter, and Marc Aubreville. Wuerstchen: Efficient pretraining of text-to-image models. ArXiv, abs/2306.00637, 2023. 6 +[34] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 6 +[35] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 763-772, 2020. 6 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 6 +[38] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, V.Ya. Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. In Conference on Empirical Methods in Natural Language Processing, 2023. 6 +[39] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using + +autoencoder reconstruction error. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9130-9140, 2024. 3, 2 +[40] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 6 +[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6 +[42] Segmind. Announcing ssd-1b: A leap in efficient t2i generation., 2023. 6 +[43] Minghe Shen, Hongping Gan, Chao Ning, Yi Hua, and Tao Zhang. Transc: A transformer-based hybrid architecture for image compressed sensing. IEEE Transactions on Image Processing, 31:6991-7005, 2022. 2 +[44] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12105-12114, 2023. 1, 2 +[45] Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28130-28139, 2024. 2, 7 +[46] Devavrat Tomar, Manana Lortkipanidze, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. Self-attentive spatial adaptive normalization for cross-modality domain adaptation. IEEE Transactions on Medical Imaging, 40:2926-2938, 2021. 6 +[47] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8 +[48] Tim van Erven and Peter Harremoës. Rényi divergence and kullback-leibler divergence. IEEE Transactions on Information Theory, 60:3797-3820, 2012. 4 +[49] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 6 +[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8692-8701, 2019. 1, 7, 2, 3 +[51] Zhendong Wang, Jianmin Bao, Wen gang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for + +diffusion-generated image detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22388-22398, 2023. 3, 7 +[52] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. In Annual Meeting of the Association for Computational Linguistics, 2022. 6, 7 +[53] Syed Waqas Zamir, Aditya Arora, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for fast image restoration and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1934-1948, 2022. 2 +[54] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1593-1603, 2024. 4 +[55] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. ArXiv, abs/2306.10012, 2023. 4, 5 +[56] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7545-7556, 2023. 5 +[57] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. 2019 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 7 +[58] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2015. 8 +[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242-2251, 2017. 6 \ No newline at end of file diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/images.zip b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9dd24b6acdf732c3a64f5055cb040a53b5c06d70 --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aeff7e917a54dd90c917afc13b29d7ab399174ecf1ec80c99187c549202ee48 +size 748640 diff --git a/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/layout.json b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f06e4102cf65974304061ec1f86081a80e2e9da --- /dev/null +++ b/2025/Semantic Discrepancy-aware Detector for Image Forgery Identification/layout.json @@ -0,0 +1,9276 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 524, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 524, + 120 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 524, + 120 + ], + "type": "text", + "content": "Semantic Discrepancy-aware Detector for Image Forgery Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 143, + 458, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 143, + 458, + 172 + ], + "spans": [ + { + "bbox": [ + 153, + 143, + 458, + 172 + ], + "type": "text", + "content": "Ziye Wang Minghang Yu Chunyan Xu* Zhen Cui Nanjing University of Science and Technology, Nanjing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 174, + 449, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 174, + 449, + 185 + ], + "spans": [ + { + "bbox": [ + 157, + 174, + 449, + 185 + ], + "type": "text", + "content": "{wzynjust,mhyu,cyx}@njust.edu.cn，zhen.cui@bnu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": "With the rapid advancement of image generation techniques, robust forgery detection has become increasingly imperative to ensure the trustworthiness of digital media. Recent research indicates that the learned semantic concepts of pre-trained models are critical for identifying fake images. However, the misalignment between the forgery and semantic concept spaces hinders the model's forgery detection performance. To address this problem, we propose a novel Semantic Discrepancy-aware Detector (SDD) that leverages reconstruction learning to align the two spaces at a fine-grained visual level. By exploiting the conceptual knowledge embedded in the pre-trained vision-language model, we specifically design a semantic token sampling module to mitigate the space shifts caused by features irrelevant to both forgery traces and semantic concepts. A concept-level forgery discrepancy learning module, based on reconstruction, enhances the interaction between semantic concepts and forgery traces, effectively capturing discrepancies under the concepts' guidance. Finally, the low-level forgery feature enhancement integrates the learned concept-level forgery discrepancies to minimize redundant forgery information. Experiments conducted on two standard image forgery datasets demonstrate the efficacy of the proposed SDD, which achieves superior results compared to existing methods. The code is available at https://github.com/wzy111111/SSD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 574, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 595, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 677 + ], + "type": "text", + "content": "With the thriving of generative AI technologies, like Generative Adversarial Networks (GANs) [14] and diffusion models [2], the images generated by these models can easily create confusion by passing off the spurious as genuine. Therefore, it is crucial to develop a universal method for detecting fake images to mitigate the widespread dissemination of disinformation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 680, + 295, + 693 + ], + "type": "text", + "content": "Pioneering research [26, 32] has shown that projecting" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 211, + 544, + 266 + ], + "blocks": [ + { + "bbox": [ + 326, + 211, + 544, + 266 + ], + "lines": [ + { + "bbox": [ + 326, + 211, + 544, + 266 + ], + "spans": [ + { + "bbox": [ + 326, + 211, + 544, + 266 + ], + "type": "image", + "image_path": "ebad5eb90f3ca9bb87c11fd2b66f803687de28196141364d2508083b6c397c5a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 327, + 267, + 541, + 430 + ], + "blocks": [ + { + "bbox": [ + 327, + 267, + 541, + 430 + ], + "lines": [ + { + "bbox": [ + 327, + 267, + 541, + 430 + ], + "spans": [ + { + "bbox": [ + 327, + 267, + 541, + 430 + ], + "type": "image", + "image_path": "81954e406961bea5aa5a7ed67c1b28aa41cf3ad135b9f7908fb60b5698e8f433.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 555, + 561 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 561 + ], + "type": "text", + "content": "Figure 1. The phenomenon of misalignment between semantic concept space and forgery space. Since " + }, + { + "bbox": [ + 313, + 440, + 555, + 561 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 313, + 440, + 555, + 561 + ], + "type": "text", + "content": " can reflect the similarity of image descriptions, we model the feature space in polar coordinates. As the semantic concept space in [32] is frozen, fake samples sharing similar concepts with real ones can be easily misclassified. With forgery-adaptive space like [26], the model can correctly distinguish between them based on re-learned forgery features. Nevertheless, due to the semantic concept bias introduced by coarse text prompts, the target samples may be projected into an inaccurate semantic concept dimension, causing them to drift away from the real source samples along the fake dimension." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": "images in a joint embedding space of texts and images can effectively capture discrepancies between fake and real images. In contrast, methods [6, 13, 44, 50] overlooking the interplay between forgery traces and semantic concepts perform poorly when confronted with unseen generative models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "To investigate the visual semantic concepts of pretrained models, we conduct a statistical analysis of the output features from CNNSpot [50] and CLIP-ViT [32] (See Appendix A for more details). Under different categories, CNNSpot exhibits a synchronized difference between real and fake features in its training space. However, when tran" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18388" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 285, + 186 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 285, + 186 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 285, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 285, + 186 + ], + "type": "image", + "image_path": "3211c3b1e370bfa02685bd229d7de6b14fbe69599dd492db2cb33e55aba0b7cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 296, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 261 + ], + "type": "text", + "content": "Figure 2. Different paradigms of image forgery identification with pre-trained vision-language model. (a) Fine-tune the frozen model only by fully connected (FC) layers [32]. (b) Prompt-based designs are tuned on text prompts and contrastive objectives [26]. (c) Our paradigm incorporating visual clues can capture fine-grained forge traces by reconstruction learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 328 + ], + "type": "text", + "content": "sitioning to the CLIP's space, these differences become inconsistent. From this, we infer a nuanced relationship between semantic concepts and forgery traces: Different semantic concepts may guide the model to uncover distinct forgery traces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 484 + ], + "type": "text", + "content": "Intuitively, relying on a frozen pre-trained vision-language model like UnivFD [32] is essential, but this tends to overlook fine-grained forgery details. Although Fat-Former [26] achieves a substantial enhancement in generalization by employing the forgery-aware adaptive transformer, we observe that soft prompts based on simple [CLASS] embeddings have an intrinsic limitation in their semantic description granularity (See Appendix B for more details). The constrained breadth of the conveyed concepts may lead the detection toward incorrect predictions. This limitation highlights a misalignment between the visual semantic concept space and the target forgery space, as illustrated in Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 579 + ], + "type": "text", + "content": "To address this, one empirical approach is to design more detailed text descriptions, but this method struggles to describe all visual forgery details due to the limited length of texts and brings more computational overhead. Drawing from the aforementioned findings and analysis, we make a first attempt to align the CLIP's visual semantic concept space with the forgery space by reconstructing semantic features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "We develop a vision-based paradigm, as outlined in Fig. 2. First, employing a pre-trained model only with nearest neighbor or linear probing (e.g. UnivFD [32], Fig. 2 (a)) is suboptimal for image forgery detection. Second, modifying the pre-trained model with task-specific prompts (e.g. FatFormer [26], Fig. 2 (b)) may favor models biased towards any particular semantic concept. These studies pave the way for exploring pre-trained space with rich semantic concepts. Inspired by image reconstruction [43, 53], our paradigm amplifies the concept-level forgery discrepancies of forgery images, which empowers the model to detect sus" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "picious forgery traces with the assistance of semantic concepts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 96, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 443 + ], + "type": "text", + "content": "In this work, we present a novel Semantic Discrepancy-aware Detector (SDD) to accurately align the semantic concept space and the forgery space. Firstly, to mitigate interference from features unrelated to learned semantic concepts and forgery traces, we divide the real images into non-overlapping blocks and feed them to the frozen CLIP [36] to obtain diverse semantic patch tokens. These tokens acting as visual clues smoothly align the semantic concepts' space and forgery space. It is noteworthy that these tokens sampled by JS divergence are universally representative of the real semantic distribution. Then, the visual clues are fused into a concept-level forgery discrepancy module. Unlike FatFormer, LoRA layers are incorporated into the image encoder. The goal is to preserve the completeness and diversity of the learned semantic concepts of CLIP, while the forgery features sharing similar semantic concepts should be highlighted. During reconstruction, we only narrow the reconstruction gap for real samples to reinforce the reconstructed discrepancies of the synthetic images. Finally, we present low-level forgery feature enhancement to let the reconstruction difference map enhance the extraction of the highly generalizable forgery features while introducing minimal additional parameters. The main challenge is how to capture forgery features with strong semantic concept correlation and features with high forgery relevance but weak semantic concept ties to ensure the model converges to powerful features. Motivated by this, we apply convolutional modules and adaptive weight parameters to avoid over-relying on semantic concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "text", + "content": "We thoroughly evaluate the generalization performance of our model on a UnivFD benchmark [32] and a SynRIS benchmark [5]. Surprisingly, our method achieves superior performance by a " + }, + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "inline_equation", + "content": "ap_{m}" + }, + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "text", + "content": " of 98.51% and a " + }, + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "inline_equation", + "content": "acc_{m}" + }, + { + "bbox": [ + 313, + 443, + 556, + 525 + ], + "type": "text", + "content": " of 93.61% on the UnivFD benchmark [32] and an average AUROC of 95.1% on the SynRIS benchmark [5]. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 527, + 553, + 647 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 527, + 553, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 553, + 562 + ], + "type": "text", + "content": "- We propose a robust model (SDD) for forgery detection, specifically designed to align the visual concept space and forgery space in terms of visual information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 553, + 609 + ], + "type": "text", + "content": "- We sample semantic tokens to mitigate the space shifts and align the two spaces through reconstruction learning. Additionally, we strengthen low-level forgery features to enhance the model's robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 611, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 553, + 647 + ], + "type": "text", + "content": "- Our method achieves superior performance on two benchmarks, demonstrating its superior capability in comparison to existing approaches." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "AI-generated Images Detection. Extensive efforts have been devoted to enhancing the performance of AI-generated image detection. Early works like [25, 44, 45] tend to mine" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18389" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 305 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 305 + ], + "type": "image", + "image_path": "94dd20070c1e79972c6f8254b876b80beacef33b3db35c97c2297de23b51f24e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": "Figure 3. The architecture of SDD. First, Next, we sample semantic tokens from real images to learn features related to both concepts and forgery. the input images are mapped into a joint space of visual semantic concepts and forgery, which are transformed into learnable features " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "V_{H}" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": " . Then, we use transformer-based encoder and decoder to get reconstructed features " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_f" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": " . A reconstruction difference map " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_S" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": " is obtained and goes through the multi-scale convolutional network to refine forgery features. Finally, we concatenate the CLIP's CLS token with this output along the same dimension for classification. The whole system is trained by jointly minimizing the binary cross-entropy loss " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "L_{bce}" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": " , the reconstruction loss " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "L_{r}" + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "text", + "content": " , and the triplet loss " + }, + { + "bbox": [ + 55, + 313, + 555, + 381 + ], + "type": "inline_equation", + "content": "L_{tri}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 392, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 296, + 691 + ], + "type": "text", + "content": "the common forgery traces between all real and fake images, such as noise patterns, texture statistics, and frequency signals. As an illustration, Liu et al. [24] designed a network that learns the consistent noise patterns in images for fake detection. Liu et al. [28] proposed to leverage the gram matrix to discover the global anomalous texture of fake images. An effective approach [13] demonstrated that frequency representation is an important factor in improving fake detection performance. However, these differences are rigorously specific to the monotonous features, which contribute to the issue of overfitting. Cutting-edge research [26, 32] shifted attention toward the semantic properties of images. Ojha et al. [32] showed that projecting images into the feature space of pre-trained vision-language model enables strong generalization ability. To build generalized forgery representations, Liu et al. [26] constructed forgery adaptive space by a forgery-aware adapter. The above research [5, 26, 32] has suggested that concept attributes are vital in the image forgery detection task. Assuming that diffusion-based models leave distinct forgery traces that are characteristic of specific concept distributions, we aim to extract robust forgery features guided by semantic concepts, rather than suppressing them. Therefore, even \"useless\" information can be useful by providing significant certainty about the content of the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 692, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 692, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 692, + 295, + 704 + ], + "type": "text", + "content": "Reconstruction Learning. Reconstruction learning has" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 392, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 609 + ], + "type": "text", + "content": "great potential in unsupervised representation learning [16, 27]. Some works [5, 39] utilized reconstruction learning to reveal the nuances between real and fake images. For example, Wang et al. [51] found that reconstructing images by DDIM exposes an error between real images and their reconstructed replica. The new synthetic image detection method[5] used text-conditioned inversion maps to learn internal representations, which is conducive to predicting whether an image is fake. Ricker et al. [39] offered a simple detection approach by applying AE to measure the reconstruction error. Notably, these works are committed to reconstructing the distributions of both real and fake samples by leveraging generative models. Unlike previous works, we focus solely on reconstructing real images in the finetuned CLIP space in light of the authenticity and richness of semantic concepts. The distribution of real samples, learned from pre-trained vision-language model, helps to define an optimal boundary, thus alleviating overfitting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 620, + 397, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 397, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 397, + 634 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Our goal is to align forgery and visual semantic concept spaces using reconstruction techniques for robust and generalizable synthetic image detection. To achieve this, we introduce a fine-grained model named Semantic Discrepancy-aware Detector (SDD). Building on prior works, we harness the generalization capability of vision-language mod" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18390" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "els. Semantic concept space: The ideal joint embedding space of images and texts with four properties: semantic alignment, modality invariance, locality consistency, and structure preservation. Forgery space: The ideal space covers forgery traces. Notably, we derive semantic concept space via a vision language model pretrained solely on real images; thus we treat the two spaces as independent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 300 + ], + "type": "text", + "content": "First, the Semantic Tokens Sampling (STS) module utilizes Jensen-Shannon (JS) divergence to sample semantic patch tokens, which serve as a transitional bridge, facilitating model to establish the association between real and forgery images accurately. Next, the Concept-level forgery Discrepancy Learning (CFDL) module employs reconstruction learning to explore the forgery discrepancies within the visual semantic concept space, which focuses on identifying subtle variations between reconstructed forgery features. Finally, the reconstruction difference map is trained with Low-level forgery Feature Enhancement module, which aims to refine forgery features with more visual details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 307, + 205, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 205, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 205, + 319 + ], + "type": "text", + "content": "3.1. Semantic Tokens Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "text", + "content": "Initially, we considered directly aligning the visual semantic concept space and forgery space by leveraging fine-grained reconstruction learning to model real and fake semantic distributions. However, this strategy would treat the differences in features unrelated to semantics and forgery as crucial factors for identifying image's authenticity. To eliminate these redundant features, we sample real semantic image patch tokens as visual clues to bridge real and forged semantic domains. This module enables the model to focus on concept-related forgery traces and highlight the distinctions between real and fake images. In a tangible way, the image encoder of CLIP: ViT-L/14 is adapted to transform a real image " + }, + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "inline_equation", + "content": "x_{r}" + }, + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "text", + "content": " into a set of features " + }, + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 55, + 324, + 296, + 491 + ], + "type": "text", + "content": ", without the image CLS token. We define the transformation as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 501, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 501, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 149, + 501, + 295, + 514 + ], + "type": "interline_equation", + "content": "f _ {r} = \\phi (x _ {r}), \\tag {1}", + "image_path": "f927597f238a779d31e5a487eb53bd3da19e4faf6d892f20d6521945abd9926c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": " is the CLIP:ViT-L/14's visual encoder, " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "x_{r}\\in" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_r^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": " represents a real image characterized by a height of " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": " and a width of " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": ". Besides, " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "f_{r}\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": " is the number of tokens and " + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 521, + 295, + 582 + ], + "type": "text", + "content": " denotes the dimension of each patch token." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "Since integrating all real patch tokens into the image reconstruction module is computationally intensive and memory-consuming, it is urgent to select a suitable subset of these tokens. From a distribution perspective, the Jensen-Shannon (JS) divergence, derived from the Kullback-Leibler divergence [48], is a symmetric and finite metric that can effectively measure the similarity between tokens by quantifying differences in their distributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To calculate the JS divergence between two tokens, both are converted into computable probability distribution space using the softmax function. Let " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{s} \\in \\mathbb{F}_{r}^{M \\times D}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " be the selected" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "semantic patch tokens with the num of tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "M = 1 / \\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and dimension " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " in terms of sampling rate " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "0 \\leq \\delta \\leq 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " is user-defined). Once the initial token " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\tilde{r}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " are determined, the JS divergence between " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\tilde{r}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and other tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " falls within the range [0, 1]. Subsequently, the sampling interval is split into " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " equal segments with one token selected from each segment. As a consequence, the semantic tokens sampling module can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 332, + 170, + 433, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 170, + 433, + 201 + ], + "spans": [ + { + "bbox": [ + 332, + 170, + 433, + 201 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {s} = \\mathcal {S} \\left(\\mathbb {R} ^ {N \\times D}, \\delta\\right) \\\\ = A _ {c} ^ {N _ {a} \\times M} \\times \\mathbb {R} ^ {N \\times D}, \\\\ \\end{array}", + "image_path": "b1673ef70479f64400aa62e1a90168013dfe36d23dbc75d23f8eb56c258378aa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "text", + "content": "s.t. JS(softmax " + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "inline_equation", + "content": "(\\tilde{r})" + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "text", + "content": " ,softmax " + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "inline_equation", + "content": "(r)) = \\frac{i}{M}" + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "text", + "content": " , if " + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "inline_equation", + "content": "a_{ij} = 1" + }, + { + "bbox": [ + 324, + 203, + 553, + 230 + ], + "type": "text", + "content": " (2)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 228, + 485, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 228, + 485, + 262 + ], + "spans": [ + { + "bbox": [ + 343, + 228, + 485, + 262 + ], + "type": "interline_equation", + "content": "\\sum_ {i = 0} ^ {N _ {a}} \\sum_ {j = 0} ^ {M} a _ {i j} = M; \\sum_ {j = 0} ^ {M} a _ {i j} = 1 \\text {o r} 0,", + "image_path": "22131de446311ff01b584ce196be7d780869ef135d597925e22a2e48d3b6a0c3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 264, + 465, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 264, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 341, + 264, + 465, + 277 + ], + "type": "interline_equation", + "content": "i = 1, \\dots , N _ {a}; j = 1, \\dots , M,", + "image_path": "45ffcb7cbdb7365aeb1d69b12ccd78f601b1f42cd109261d8ec31750b7a6add6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " represents the sampling process. " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "A_{c}^{N_{a} \\times M}" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " is a constraint matrix of size " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "N_{a} \\times M" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " whose element " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "a_{ij}" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " is constrained to the binary pattern of " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "\\{0, 1\\}" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": ". Here JS " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " refers to the Jensen-Shannon divergence, " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "N_{a}" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " denotes the total number of real image patch tokens sampled from the training dataset of UnivFD and " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " represents the required subset size. The softmax " + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 283, + 555, + 428 + ], + "type": "text", + "content": " is the softmax function. The sampling tokens help the reconstruction module avoid becoming biased towards any particular forgery-unrelated distribution. Meanwhile, it avoids the semantic bias often introduced by text prompts, since the tokens are evenly distributed in a unified CLIP space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 433, + 542, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 542, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 542, + 446 + ], + "type": "text", + "content": "3.2. Concept-level Forged Discrepancy Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "content": "A few words alone can hardly paint a picture. We argue that the fine-grained visual details can uncover more forgery traces concealed in the images. As such, we mix sampling tokens with extracted features and capitalize on reconstruction learning to compensate for the omission of forgery traces brought by coarse prompts. As previous work [26] has demonstrated that the pre-trained vision-language model necessitates fine-tuning to adapt to the forgery detection task. Therefore, we integrate LoRA [17] with the CLIP-ViT model to capture discriminative forgery features by making use of the bread semantic concepts. This method, denoted as LoRA-CLIP [54], is more streamlined and flexible. Given an input image " + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{I}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "content": ", we can get high-level visual features " + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "inline_equation", + "content": "V_{H}" + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 396, + 623, + 553, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 623, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 396, + 623, + 553, + 636 + ], + "type": "interline_equation", + "content": "V _ {H} = \\mathcal {F} _ {L o R A} (\\mathcal {I}). \\tag {3}", + "image_path": "314740cadca648129e1efc3dc3692b99cf4e100e085964874ef8ba845ddc990a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 712 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 641, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{LoRA}" + }, + { + "bbox": [ + 313, + 641, + 555, + 712 + ], + "type": "text", + "content": " refers to the CLIP image encoder fine-tuned by LoRA. The reconstruction module of CFDL encompasses two submodules, i.e., transformer-based encoder and decoder. Thanks to the transformer's capability of long-range relationship modeling, we capitalize on the multi-head attention (MHA) mechanism, the core mechanism of" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18391" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "the transformer, to obtain more discriminative distribution by utilizing contextual information, which is set as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 104, + 295, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 104, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 66, + 104, + 295, + 179 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{r l} \\operatorname {h e a d} _ {i} & = \\operatorname {A t t n} \\left(Q W _ {i} ^ {Q}, K W _ {i} ^ {K}, V W _ {i} ^ {V}\\right) \\\\ & = \\operatorname {S o f t m a x} \\left(\\frac {Q W _ {i} ^ {Q} \\left(K W _ {i} ^ {K}\\right) ^ {\\top}}{\\sqrt {d}}\\right) V W _ {i} ^ {V}, \\\\ \\operatorname {M H A} (Q, K, V) & = \\operatorname {C o n c a t} \\left(\\operatorname {h e a d} _ {1}, \\dots , \\operatorname {h e a d} _ {h}\\right) W ^ {O}, \\end{array} \\right. \\tag {4}", + "image_path": "c175dbe07601fc84121dccf61a36fc2ea1d94595dba67b1068d5f85ac84ce077.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " (Query), " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " (Key), " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " (Value) refer to the input, " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "W_{i}^{Q}, W_{i}^{K}, W_{i}^{V}" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " separately denote the corresponding weights of linear projection, " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}(\\cdot)" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " denotes the function of the scaled dot product, Softmax is the softmax function, " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " refers to the dimension of input, Concat (" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": ") represents the concatenation used to stitch the discrete attention outputs of head " + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "inline_equation", + "content": "1 \\sim h" + }, + { + "bbox": [ + 55, + 180, + 296, + 263 + ], + "type": "text", + "content": " together." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 312 + ], + "type": "text", + "content": "To amplify the discrepancy between a fake image and its reconstructed counterpart, the sampled visual clues are employed for the initial processing by the encoder. The encoder's process can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 322, + 294, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 322, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 110, + 322, + 294, + 335 + ], + "type": "interline_equation", + "content": "R _ {1} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(f _ {s}, V _ {H}, V _ {H}\\right)\\right), \\tag {5}", + "image_path": "df6b3ec424ab3677b5fc6a432a0f5df2401f7f07dfd5de79679080faf3fd1550.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 337, + 294, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 337, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 111, + 337, + 294, + 350 + ], + "type": "interline_equation", + "content": "R _ {2} = \\operatorname {L N} \\left(\\mathrm {M H A} \\left(R _ {1}, V _ {H}, V _ {H}\\right)\\right), \\tag {6}", + "image_path": "e8a28d88ab4d834415bcb0a67ba5a44365295d5ea0021622d7905beab46893db.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 359, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 55, + 359, + 296, + 418 + ], + "type": "text", + "content": " denotes the Layer Normalization. Then, the encoder's outputs used as queries are injected into the decoder to get the final reconstructed features, which are similar to the encoder process and perform the following operation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 429, + 294, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 429, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 111, + 429, + 294, + 441 + ], + "type": "interline_equation", + "content": "R _ {3} = \\mathrm {L N} (\\mathrm {M H A} (R _ {2}, R _ {2}, R _ {2})), \\tag {7}", + "image_path": "aa92fde74c974d56acdb710d5503650cf873df09f827e5f18f01f71635b0971c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 444, + 294, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 444, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 113, + 444, + 294, + 456 + ], + "type": "interline_equation", + "content": "R _ {e} = \\mathrm {L N} (\\mathrm {M H A} \\left(R _ {1}, R _ {3}, R _ {3}\\right)). \\tag {8}", + "image_path": "41c58f4ee86a40e85c45405934de624b27c459aacdb4aa89d0900ca8f56d104b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "text", + "content": "During the reconstruction process, we just calculate the reconstruction loss " + }, + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "text", + "content": " between the real input features and their reconstructed counterparts " + }, + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_e" + }, + { + "bbox": [ + 55, + 466, + 296, + 512 + ], + "type": "text", + "content": " within a mini-batch as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 511, + 295, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 511, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 116, + 511, + 295, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\frac {1}{B} \\sum_ {i = 0} ^ {B} \\operatorname {M S E} \\left(R _ {e}, V _ {H}\\right), \\tag {9}", + "image_path": "be3a4806567dafc287eefb6cd45fc30d8fbe79ce80ec399304f6c863e40b472b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\mathrm{MSE}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "content": " is mean squared error. Facilitating " + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "content": " encourages preserving the completeness and richness of the visual semantic concept space and highlighting the concept-related forgery features. Given the reconstructed features " + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "inline_equation", + "content": "R_{f}" + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "content": " and the original feature " + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 55, + 548, + 296, + 621 + ], + "type": "text", + "content": ", the reconstruction difference map can be formally expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 139, + 630, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 630, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 139, + 630, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {s} = \\left| R _ {f} - f _ {r} \\right|, \\tag {10}", + "image_path": "fdbca3dba04bcb0edfb29513e8d25a6eab435c19726612566d32df084f949629.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 652, + 239, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 239, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 239, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 652, + 239, + 664 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 55, + 652, + 239, + 664 + ], + "type": "text", + "content": " denotes the absolute value function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 672, + 270, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 270, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 270, + 685 + ], + "type": "text", + "content": "3.3. Low-level Forgery Feature Enhancement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Existing methods based on pre-trained vision-language models [26, 32] overlook the importance of concept" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 358, + 72, + 512, + 196 + ], + "blocks": [ + { + "bbox": [ + 358, + 72, + 512, + 196 + ], + "lines": [ + { + "bbox": [ + 358, + 72, + 512, + 196 + ], + "spans": [ + { + "bbox": [ + 358, + 72, + 512, + 196 + ], + "type": "image", + "image_path": "de568ffeb00d073298b5f7dc1c0e06b896ff4b86f32ec8d1e4c74be043a2f038.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "type": "text", + "content": "Figure 4. The curve of exponential inverse. In the \"fast\" interval, the value drops sharply. In the \"low\" interval, the curve flattens out, showing a decay towards 0." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": "weakly-related features. We believe that a thorough alignment between the visual semantic concept space and the forgery space should include the exploration of concept weakly-related forgery features. To eliminate redundant forgery features, we come up with a novel feature enhancement that refines low-level forgery features. Empowered by the reconstruction difference map, our detector orchestrates the extraction of multi-scale features with exceptional robustness and markedly enhanced effectiveness. As shown in Fig. 3, the enhancer follows the typical architecture of a convolutional network. It involves the repeated application of convolutions, each followed by a batch normalization (BN) and a rectified linear unit (ReLU). For a given stage " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "F(n)" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "N = 1,2,3" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": ") corresponds to its output features. Then, We deconvolve the semantic difference map " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": " to the shape same as " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "F(N)" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": " and perform pixel-wise multiplication with " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "F(n)" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "inline_equation", + "content": "F'(n)" + }, + { + "bbox": [ + 313, + 243, + 555, + 447 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 371, + 456, + 553, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 456, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 371, + 456, + 553, + 471 + ], + "type": "interline_equation", + "content": "F ^ {\\prime} (n) = \\operatorname {d e c o n v} (F (n)) \\otimes D _ {s}, \\tag {11}", + "image_path": "08999378248378ab40d4ad095d40cc5d59f086f722d8b22410f8d387efc09933.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": " is the element-wise multiplication, deconv " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": " represents deconvolution operation and " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{F}'(n)" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": " is the low-level feature aggregated with semantic information. To further enhance the reliability of the extracted features, we compute an adaptive weight coefficient " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\frac{1}{e_n}" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": " to indicate the importance of " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "inline_equation", + "content": "F(n)" + }, + { + "bbox": [ + 313, + 480, + 554, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 391, + 560, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 560, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 391, + 560, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\frac {1}{e _ {n}} = \\frac {1}{e ^ {\\left| F ^ {\\prime} (n) - F (n) \\right|}}. \\tag {12}", + "image_path": "2a8f6ac00182802ba9ac88bd0885ef01cecad3c4c78b9b298c6bbe26df378045.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Here we explain the role of the exponential inverse through Fig. 4. As " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " grows large, the curve of " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "e^x" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " becomes flatter. Therefore, in the \"fast\" interval, forgery features with a significant divergence from the semantic difference map will be assigned smaller weights, which mobilizes the network to capture concept strongly-related features. However, in the \"low\" interval, features strongly associated with forgery can avoid being misguided by semantic concepts, which indicates that the order of importance is reversed. Next, we have the attended output features " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{low}" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18392" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 205 + ], + "type": "table", + "html": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDallemAP
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200Steps200w/cfg100Steps100w/ CFG10027502710
CNN-SpotCVPR2020100.093.4784.5099.5489.4998.1589.0273.7559.4798.2498.4073.7270.6271.0070.5480.6584.9182.0770.5983.58
PatchForECCV202080.8872.8471.6685.7565.9969.2576.5576.1976.3474.5268.5275.0387.1086.7286.4085.3783.7378.3875.6777.73
Co-occurrenceElect.Imag.99.7480.9550.6198.6353.1167.9959.1468.9860.4273.0687.2170.2091.2189.0292.3989.3288.3582.7980.9678.11
Freq-specWIFS201955.39100.075.0855.1166.08100.045.1847.4657.1253.6150.9857.7277.7277.2576.4768.5864.5861.9267.7766.21
DireICCV2023100.083.5981.5096.5081.7099.8895.7362.5169.9897.3198.6279.5375.5273.4276.4586.2889.0088.3451.3583.54
UnivFDCVPR2023100.099.4699.5997.2499.9899.6082.4561.3279.0296.7299.0087.7799.1492.1599.1794.7495.3494.5797.1593.38
NPRCVPR2024100.099.5096.5099.8096.80100.092.2073.1078.7087.2064.8065.8099.8099.8099.8099.7099.8099.8098.6092.19
FatFormerCVPR2024100.0100.099.9899.75100.0100.097.9997.9481.2399.8499.9391.9299.8399.2299.8999.2799.5099.3399.8498.18
Ours100.099.7799.9399.4899.9899.9797.2397.9193.1099.7999.9692.0699.8898.9599.9298.0698.2997.7399.8198.51
", + "image_path": "87d0f709d3c5beebf16fa5d4102d64bc12877c31ade452d4291d5c56b5d7b629.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 257, + 555, + 393 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 247 + ], + "type": "text", + "content": "Table 1. Average precision comparisons with different methods on the UnivFD dataset. We replicate the results of CNNSpot, Patchfor, Co-occurrence, Freq-spec, and UnivFD from the paper[32]. In addition, we obtained the results for Dire, NPR and FatFormer using either the official pre-trained models or our re-implemented versions. Red and underline indicates the best and the second best result, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 257, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 57, + 257, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 57, + 257, + 555, + 393 + ], + "type": "table", + "html": "
MethodsRefGANDeep fakesLow levelPerceptual lossGuidedLDMGlideDalleAvg-acc
Pro-GANCycle-GANBig-GANStyle-GANGau-GANStar-GANSITDSANCRNIMLE200 Steps200 wcfg100 Steps100 2750 27100 10
CNN-SpotCVPR202099.9985.2070.2085.7078.9591.7053.4766.6748.6986.3186.2660.0754.0354.9654.1460.7863.8065.6655.5869.58
PatchForECCV202075.0368.9768.4779.1664.2363.9475.5475.1475.2872.3355.3067.4176.5076.1075.7774.8173.2868.5267.9171.24
Co-occurrenceElect.Imag.97.7063.1553.7592.5051.1054.7057.1063.0655.8565.6565.8060.5070.7070.5571.0070.2569.6069.9067.5566.86
Freq-specWIFS201949.9099.9050.5049.9050.3099.7050.1050.0048.0050.6050.1050.9050.4050.4050.3051.7051.4050.4050.0055.45
DireICCV202399.8673.4760.6872.3965.1593.6088.8652.7856.3990.0794.0561.0559.3559.9560.6569.3072,7071.0052.7571.19
UnivFDCVPR2023100.098.5094.5082.0099.5097.0066.6063.0057.5059.5072.0070.0394.1973.7694.3679.0779.8578.1486.7881.38
NPRCVPR202499.8092.0089.5096.3087.6099.7079.4061.4070.6074.5057.1055.2397.4098.7097.9097.0097.9097.0088.8086.20
FatFormerCVPR202499.8999.3699.5097.1299.4399.7593.2581.3968.0469.4769.4776.0098.5594.8598.6094.3094.6094.1598.7090.86
Ours99.8895.7696.7098.0898.4699.1791.8283.6177.4595.4096.4779.5598.0594.6098.2592.2093.3591.8098.0093.61
", + "image_path": "2ce8a5a892e90254564779c552ba0bbf2b349de34e09a0b5faf6cd58b3ff19c3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 401, + 447, + 412 + ], + "lines": [ + { + "bbox": [ + 161, + 401, + 447, + 412 + ], + "spans": [ + { + "bbox": [ + 161, + 401, + 447, + 412 + ], + "type": "text", + "content": "Table 2. Accuracy comparisons with different methods on the UnivFD dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "type": "text", + "content": "by the residual connection:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 437, + 294, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 437, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 437, + 294, + 464 + ], + "type": "interline_equation", + "content": "F _ {l o w} (n) = F ^ {\\prime} (n) + \\frac {F ^ {\\prime} (n)}{e _ {n}}. \\tag {13}", + "image_path": "991a2b93fcf68991da2ea441485030fb0ba0c01a67a96ef081cd220d9243405b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "text", + "content": "For optimizing the anchor features " + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\tilde{f}_a" + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "text", + "content": " of the enhancer, the following triplet loss [35] is employed to bring positive samples " + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "text", + "content": " closer while pushing negative samples " + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 55, + 463, + 295, + 500 + ], + "type": "text", + "content": " apart:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 505, + 294, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 505, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 78, + 505, + 294, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t r i} = \\max \\left(0, d \\left(f _ {p}, f _ {a}\\right) - d \\left(f _ {n}, f _ {a}\\right) + \\alpha\\right), \\tag {14}", + "image_path": "53d151e6cf2f196bee8193296abb356b2ebb4787ee0a1cf02479e20ffd8cb315.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": " represents the Euclidean distance between samples and " + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": " is the margin." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": "On top of that, we concatenate the LoRA-CLIP's CLS token " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "T_{CLS}" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "F_{low}" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": " along the same dimension to yield the refined representation " + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "inline_equation", + "content": "F_{out}" + }, + { + "bbox": [ + 55, + 548, + 296, + 620 + ], + "type": "text", + "content": ". This ensures that forged features exhibit distinctiveness across different semantic identities while preserving their uniformity within similar semantic identities. The process is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 627, + 295, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 627, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 627, + 295, + 640 + ], + "type": "interline_equation", + "content": "F _ {o u t} = F _ {l o w} \\| T _ {C L S}, \\tag {15}", + "image_path": "7197ac7b52a1f65c6eedd5a96ff6c2e5c5e892f668e5331b72954d4d14567d12.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "inline_equation", + "content": "F^{out}" + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": " is strategically integrated with a linear classifier to enable the execution of binary classification. Eventually, the total loss function " + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": " of the proposed framework can be defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 116, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {b c e} + \\lambda_ {1} \\mathcal {L} _ {t r i} + \\lambda_ {2} \\mathcal {L} _ {r}, \\tag {16}", + "image_path": "eeda4d37e8bdf767295153547708777738fe9e4ef89b079c76865f10857cc5f0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bce}" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": " presents the binary cross-entropy loss and " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tri}" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": " is triplet loss. " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tri}" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": " are scaled by the hyperparameters " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 419, + 555, + 456 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 470, + 395, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 395, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 395, + 483 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 491, + 425, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 425, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 425, + 504 + ], + "type": "text", + "content": "4.1. Experiment Setups" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": "Datasets: We follow the protocol described in [32], using ProGAN's real and fake images as training data. Additionally, we adopt the protocol from [5], where the training data is composed of fake Stable Diffusion v1 images [52] and random real LAION images [41]. The UnivFD dataset [32] cover a broad range of generative models, primarily including GANs and diffusion models, such as ProGAN [18], StyleGAN [19], BigGAN [4], CycleGAN [59], StarGAN [10], GauGAN [46], CRN [9], IMLE [22], SAN [11], SITD [7], DeepFakes [20], Guided [12], Glide [31], LDM [40] and DALL-E [37]. The SynRIS dataset [5] is designed to avoid bias toward any specific topic, theme, or style and contains high-fidelity images generated by text-to-image models, such as Kandinsky2 [38], Kandinsky3 [1], PixArt- " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " [8], SDXL-DPO [49], SDXL [34], SegMoE [23], SSD-1B [42], Stable-Cascade [33], Segmind-Vega [15], and Würstchen2 [33], Midjourney [29], DALL.E 3 [3]" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18393" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 289, + 302 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 289, + 302 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 289, + 302 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 289, + 302 + ], + "type": "table", + "html": "
MethodsCNN-SpotFreq-specDireUniv FDNPRFatF ormerFakeIn versionPatch ForOurs
Kandinsky260.057.071.656.297.575.669.953.597.13
Kandinsky365.945.774.961.493.780.174.351.494.8
PixArt-α62.756.481.564.789.575.373.049.690.2
SDXL-DPO84.369.869.970.297.686.088.154.595.5
Segmind-Vega74.265.381.062.397.182.481.153.197.0
SDXL81.461.286.266.396.085.180.763.997.6
Seg-MoE66.354.671.962.093.670.871.349.897.6
SSD-1B72.667.879.862.899.670.179.461.299.8
Stable-Cascade70.562.174.168.297.681.674.957.499.2
Würstchen261.063.374.269.790.972.970.547.298.2
Midjourney63.050.972.459.258.573.666.453.790.0
Playground58.252.367.958.793.181.462.554.192.8
DALL-E371.659.980.848.069.179.275.950.185.9
Average68.658.975.962.390.378.074.553.895.1
", + "image_path": "21d2a7a74f868328d782afdbbed51e059299eb60369d9ada291a02909bf435ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 312, + 295, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 369 + ], + "type": "text", + "content": "Table 3. AUROC comparisons with different methods on the SynRIS dataset. We retrieve the results of CNNSpot, UnivFD, and FakeInversion from [5] and obtain the results for Dire, NPR, Fat-Former, and PatchFor using re-implemented models. Red and underline indicate the best and the second best result, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 377, + 145, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 145, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 145, + 389 + ], + "type": "text", + "content": "and playground [21]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 424 + ], + "type": "text", + "content": "Metrics: As standard evaluation metrics, the average precision (AP), the accuracy (ACC) and AUCROC are considered to measure the effectiveness of different methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 628 + ], + "type": "text", + "content": "Baselines: In our experiments, we perform thorough comparisons with state-of-the-art methods, as follows: 1) CNNSpot [50]: The method relies only on one CNN generator. 2) PatchFor [6]: The method performs detection on a patch level. 3) Co-occurrence [30]: The method converts input images into co-occurrence matrices for classification. 4) Freq-spec [57]: The method employs the frequency spectrum of images. 5) Dire [51]: The method exploits the error between an input image and its reconstruction counterpart. 6) UnivFD [32]: The method uses the pre-trained language-vision model to determine the authenticity of images. 7) NPR [45]: The method captures the generalized artifacts according to the local interdependence among image pixels. 8) FatFormer [26]: The method is aimed at extracting forgery-adaptive features based on UnivFD. 9) Fakeinversion [5]: The method employs text-conditioned inversion maps extracted from Stable Diffusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Implement details: Our training and testing settings are adapted from the approach outlined in the previous study [26] with several key modifications. Specifically, early stopping was employed during model training, with an initial learning rate of " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": " and a batch size of 32. Additionally, the Lora layers are configured with hyperparameters " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "lora_{r} = 6" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "inline_equation", + "content": "lora_{\\alpha} = 6" + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": ", and a dropout rate of 0.8," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 71, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 324, + 71, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 324, + 71, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 324, + 71, + 545, + 137 + ], + "type": "table", + "html": "
#STS moduleCFDL modulefeature enhancementUnivFD apmDataset accm
1✓97.3781.64
2✓✓97.4190.17
3✓✓97.3989.98
4✓✓✓98.5293.61
", + "image_path": "b8e851f0a6c0a3b01c9ead4418bfc9f62fb6a7e0eaf71b764adcc5459768e91e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "type": "text", + "content": "Table 4. Ablation study of the proposed modules on the UnivFD Dataset. We show the mean accuracy " + }, + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "type": "inline_equation", + "content": "(acc_m)" + }, + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "type": "text", + "content": " and average precision " + }, + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "type": "inline_equation", + "content": "(ap_m)" + }, + { + "bbox": [ + 313, + 148, + 553, + 192 + ], + "type": "text", + "content": ". Red and underline indicate the best and the second-best result, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 198, + 553, + 221 + ], + "type": "text", + "content": " is set to 8.0. The proposed method is implemented using Pytorch on 2 Nvidia GeForce RTX A6000 GPUs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 435, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 435, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 435, + 241 + ], + "type": "text", + "content": "4.2. Comparison Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 317 + ], + "type": "text", + "content": "The UnivFD dataset includes a diverse range of models, allowing for a comprehensive evaluation of our method across both GAN and diffusion generative models. In addition, the SynRIS dataset provides images generated by cutting-edge generative models. The overall experimental results are presented in Tab. 1, Tab. 2, and Tab. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "content": "Results on UnivFD dataset. Results show that our proposed method achieves superior performance compared to the UnivFD and FatFormer. Notably, without the biased interpretation introduced by coarse-grained text prompts, SDD surpasses the latest state-of-the-art method FatFormer by the mean AP " + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "inline_equation", + "content": "(ap_{m})" + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "inline_equation", + "content": "0.34\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "content": " and the mean acc " + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "inline_equation", + "content": "(acc_{m})" + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "inline_equation", + "content": "2.75\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 473 + ], + "type": "text", + "content": ". Moreover, compared with methods based on relatively monotonous forgery features in Tab. 1 and Tab. 2, our approach can outperform all of them with a large improvement. The above evidence indicates effective combination of visual concepts and forgery features can contribute model to extract sufficient forgery patterns and eliminate the superfluous features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "text", + "content": "Results on SynRIS dataset. As shown in Tab. 3, when confronted with high-fidelity images generated by text-to-image models, methods leveraging pre-trained vision-language models, such as UnivFD and FatFormer, lose their competitiveness. In contrast, NPR, which focuses on neighboring pixel relationships, retains its edge. We assume that current generative methods grasp the relationships between visual information and semantic concepts in images but cannot refine local forgery details at the pixel level. Considering that excessive reliance on concepts misses abnormal pixel arrangements and focusing on monotonous forgery patterns can cause overfitting, our detector, which emphasizes low-level features with visual concepts, is trained on lower-fidelity fake images generated by Stable Diffusion [52] to capture concept-specific lacunae. We follow the evaluation protocol from SynRIS [5]. In comparison, our detector achieves an impressive mean AUROC of " + }, + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "inline_equation", + "content": "95.1\\%" + }, + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "text", + "content": ", surpassing the state-of-the-art method by " + }, + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "inline_equation", + "content": "4.8\\%" + }, + { + "bbox": [ + 313, + 475, + 554, + 713 + ], + "type": "text", + "content": ". This demonstrates its superior ability to tackle the challenges posed by evolving generative models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18394" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 190, + 169 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 190, + 169 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 190, + 169 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 190, + 169 + ], + "type": "image", + "image_path": "0fc9b0eeef54d2a107d0b1c0f6d6c4e979f1491fe980fcbbc07675e0143222a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 179, + 559, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 559, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 559, + 202 + ], + "type": "text", + "content": "Figure 5. Performance of function on Figure 6. T-SNE visualization of real and fake images [47]. The feature space is based on our adaptive weights. classifier. Each randomly samples 500 real and 500 fake images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 211, + 69, + 291, + 171 + ], + "blocks": [ + { + "bbox": [ + 211, + 69, + 291, + 171 + ], + "lines": [ + { + "bbox": [ + 211, + 69, + 291, + 171 + ], + "spans": [ + { + "bbox": [ + 211, + 69, + 291, + 171 + ], + "type": "image", + "image_path": "9ba89e1496d6b16bdccd57c575440594fea729573cfd64ec69d877d04d24eeca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 378, + 171 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 378, + 171 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 378, + 171 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 378, + 171 + ], + "type": "image", + "image_path": "53ea1b5f353a96c9738269ec99c6192074f98e4c9590ea3b568427e6f48b7455.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 465, + 171 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 465, + 171 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 465, + 171 + ], + "type": "image", + "image_path": "12508e5e81b25cf63bb1033d94fba7fce7241302bce0d94e1bdf758f953e09b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 478, + 70, + 555, + 171 + ], + "blocks": [ + { + "bbox": [ + 478, + 70, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 478, + 70, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 478, + 70, + 555, + 171 + ], + "type": "image", + "image_path": "88bdb2dfb8a557c7e1f64a77e634075a41c2959d745cddd5ef7d47a86cb6ada7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 207, + 294, + 330 + ], + "blocks": [ + { + "bbox": [ + 56, + 207, + 294, + 330 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 330 + ], + "type": "image", + "image_path": "3f926a12a11783a8f3703d8b1e19a10696a9a49ef46cb72f9c96a4c3486c118f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 336, + 295, + 359 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 359 + ], + "type": "text", + "content": "Figure 7. The showcase of attention on images which are input into our model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 147, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 147, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 147, + 379 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "content": "We perform comprehensive ablation studies on the UnivFD dataset under the original experimental configurations, reporting the mean accuracy " + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "inline_equation", + "content": "(acc_{m})" + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "content": " and mean average precision " + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "inline_equation", + "content": "(ap_{m})" + }, + { + "bbox": [ + 55, + 385, + 296, + 433 + ], + "type": "text", + "content": " as the primary evaluation metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 567 + ], + "type": "text", + "content": "Effect of Each Component: We study the effects of removing STS module, CFDL module, and feature enhancement in our method. The results, presented in Tab. 4, demonstrate that these components are essential for improving performance in generalization on unseen models. This empirical finding suggests that the CFDL effectively captures forgery discrepancies associated with semantic concepts, while the enhancer module plays a crucial role in identifying robust forgery artifacts. The collaboration of all modules enhances the model's ability to distinguish between real and fake images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "content": "Effect of function on adaptive weights: To check how well the proposed function works in the SDD, we select two conventional functions for comparison: " + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "inline_equation", + "content": "f(x) = |x|" + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "inline_equation", + "content": "f(x) = x^2" + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "content": ". The corresponding results are presented in Fig. 5. We find that our proposed function yields improvements in both " + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "inline_equation", + "content": "ap_m" + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "inline_equation", + "content": "acc_m" + }, + { + "bbox": [ + 55, + 568, + 295, + 674 + ], + "type": "text", + "content": " compared to the selected functions. These results demonstrate that our proposed function is conducive to capturing robust and distinctive forgery features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Visualization of learned latent space: As shown in Fig. 6, the input images can be distinctly categorized into two clusters: real and fake. Nevertheless, why does the di" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 210, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 341 + ], + "type": "text", + "content": "vided boundary of ProGAN appear ambiguous in contrast to other models? Additionally, why do the real clusters of CycleGAN and StyleGAN separate from each other? We attribute these to the influence of visual semantic concepts. Perceptively, with the supervision of visual semantic concepts, the learned boundary of ProGAN is more complex and nuanced, rather than just simple straight lines or curves. Similarly, the images generated by StyleGAN and CycleGAN are projected into the corresponding semantic concept distribution and then separated from the real images based on the visual semantic concepts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": "Visualization of attention on images: We apply Class Activation Mapping (CAM) [58] to visualize the learned representations. From our perspective, Fig. 7 illustrates that with the aid of semantic information, our model can focus on different regions of fake images, including the background, local object regions, and marginal details. This suggests that our fine-grained model is capable of capturing intricate discrepancies generalized to unseen models. Notably, the real images nearly always show no forgery discrepancy regions, which demonstrates the effectiveness of the reconstruction loss in the forgery detection task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 707 + ], + "type": "text", + "content": "In this paper, we propose a novel method, SDD, for generalizable forgery image detection. The findings show that our method establishes a new state-of-the-art in detecting images generated by generative models from different periods, which underscores its robustness and superior generalization capability. To the best of our knowledge, in pre-trained vision-language paradigms, our approach is the first to rely solely on visual information, without text prompts. Based on experimental results, we conclude that leveraging sampled tokens and reconstruction techniques effectively aligns the visual semantic concept space with the forgery space. Additionally, refining low-level forgery features under the supervision of visual semantic concepts enhances the performance of forgery detection. Although SDD performs well across various generative methods, there is still room for improvement as generative technologies continue to advance. Future research will explore this further." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18395" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 296, + 116 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (Grants Nos. 62372238,62476133)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "text", + "content": "[1] V.Ya. Arkhipkin, Andrei Filatov, Viacheslav Vasilev, Anastasia Maltseva, Said Azizov, Igor Pavlov, Julia Agafonova, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 3.0 technical report. ArXiv, abs/2312.03511, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 236 + ], + "type": "text", + "content": "[2] Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Structured denoising diffusion models in discrete state-spaces. ArXiv, abs/2107.03006, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 294 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 295, + 327 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 296, + 395 + ], + "type": "text", + "content": "[5] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fake inversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10759-10769, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "text", + "content": "[6] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In European Conference on Computer Vision, 2020. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 485 + ], + "type": "text", + "content": "[7] Cheng Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3291-3300, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": ": Fast and controllable image generation with latent consistency models. ArXiv, abs/2401.05252, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[9] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 1520-1529, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 633 + ], + "type": "text", + "content": "[10] Yunjey Choi, Min-Je Choi, Mun Su Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8789-8797, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shutao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11057-11066, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. ArXiv, abs/2105.05233, 2021. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Joel Cameron Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. ArXiv, abs/2003.08685, 2020. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[15] Yatharth Gupta, Vishnu V. Jaddipal, Harish Prabhala, Sayak Paul, and Patrick von Platen. Progressive knowledge distillation of stable diffusion xl using layer level loss. ArXiv, abs/2401.02677, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "type": "text", + "content": "[16] Zhizhong Han, Xiyang Wang, Yu-Shen Liu, and Matthias Zwicker. Multi-angle point cloud-vae: Unsupervised feature learning for 3d point clouds from multiple angles by joint self-reconstruction and half-to-half prediction. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10441-10450. IEEE, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "text", + "content": "[17] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 348 + ], + "type": "text", + "content": "[18] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. ArXiv, abs/1710.10196, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4396-4405, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "text", + "content": "[20] Prabhat Kumar, Mayank Vatsa, and Richa Singh. Detecting face2face facial reenactment in videos. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 2578-2586, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[21] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground v2.5: Three insights towards enhancing aesthetic quality in text-to-image generation. ArXiv, abs/2402.17245, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[22] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4219-4228, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[23] Zhenghong Li, Hao Chen, Jiangjiang Wu, Jun Li, and Ning Jing. Segmind: Semisupervised remote sensing image semantic segmentation with masked image modeling and contrastive learning method. IEEE Transactions on Geoscience and Remote Sensing, 61:1-17, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "text", + "content": "[24] Bo Liu, Fan Yang, Xiuli Bi, Bin Xiao, Weisheng Li, and Xinbo Gao. Detecting generated images by real images. In European Conference on Computer Vision, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 615, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 679 + ], + "type": "text", + "content": "[25] Honggu Liu, Xiaodan Li, Wenbo Zhou, Yuefeng Chen, Yuan He, Hui Xue, Weiming Zhang, and Nenghai Yu. Spatial-phase shallow learning: Rethinking face forgery detection in frequency domain. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 772-781, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Huan Liu, Zichang Tan, Chuangchuang Tan, Yunchao Wei, Yao Zhao, and Jingdong Wang. Forgery-aware adaptive transformer for generalizable synthetic image detection." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18396" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10770-10780, 2023. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[27] Xinhai Liu, Xinchen Liu, Zhizhong Han, and Yu-Shen Liu. Spu-net: Self-supervised point cloud upsampling by coarse-to-fine reconstruction with self-projection optimization. IEEE Transactions on Image Processing, 31:4213-4226, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[28] Zhengzhe Liu, Xiaojuan Qi, Jiaya Jia, and Philip H. S. Torr. Global texture enhancement for fake face detection in the wild. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8057-8066, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 280, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 280, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 280, + 219 + ], + "type": "text", + "content": "[29] Midjourney. Midjourney, n.d. Accessed: 2025-03-04. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[30] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, B. S. Manjunath, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H. Bappy, and Amit K. Roy-Chowdhury. Detecting gan generated fake images using co-occurrence matrices. ArXiv, abs/1903.06836, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "text", + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 387 + ], + "type": "text", + "content": "[32] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[33] Pablo Pernias, Dominic Rampas, Mats L. Richter, and Marc Aubreville. Wuerstchen: Efficient pretraining of text-to-image models. ArXiv, abs/2306.00637, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "type": "text", + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 510 + ], + "type": "text", + "content": "[35] Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. Fcanet: Frequency channel attention networks. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 763-772, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 578 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. ArXiv, abs/2102.12092, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 690 + ], + "type": "text", + "content": "[38] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, V.Ya. Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky: an improved text-to-image synthesis with image prior and latent diffusion. In Conference on Empirical Methods in Natural Language Processing, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "type": "text", + "content": "[39] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "autoencoder reconstruction error. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9130-9140, 2024. 3, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "text", + "content": "[40] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10674-10685, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 250 + ], + "type": "text", + "content": "[41] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 274 + ], + "type": "text", + "content": "[42] Segmind. Announcing ssd-1b: A leap in efficient t2i generation., 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[43] Minghe Shen, Hongping Gan, Chao Ning, Yi Hua, and Tao Zhang. Transc: A transformer-based hybrid architecture for image compressed sensing. IEEE Transactions on Image Processing, 31:6991-7005, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "text", + "content": "[44] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12105-12114, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "text", + "content": "[45] Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28130-28139, 2024. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[46] Devavrat Tomar, Manana Lortkipanidze, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. Self-attentive spatial adaptive normalization for cross-modality domain adaptation. IEEE Transactions on Medical Imaging, 40:2926-2938, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[47] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "text", + "content": "[48] Tim van Erven and Peter Harremoës. Rényi divergence and kullback-leibler divergence. IEEE Transactions on Information Theory, 60:3797-3820, 2012. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 569, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 634 + ], + "type": "text", + "content": "[49] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8692-8701, 2019. 1, 7, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 715 + ], + "type": "text", + "content": "[51] Zhendong Wang, Jianmin Bao, Wen gang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18397" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 518 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "diffusion-generated image detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 22388-22398, 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[52] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. In Annual Meeting of the Association for Computational Linguistics, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[53] Syed Waqas Zamir, Aditya Arora, Salman Hameed Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Learning enriched features for fast image restoration and enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:1934-1948, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[54] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1593-1603, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[55] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. ArXiv, abs/2306.10012, 2023. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "text", + "content": "[56] Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, and Jianbo Shi. Perceptual artifacts localization for image synthesis tasks. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7545-7556, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[57] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. 2019 IEEE International Workshop on Information Forensics and Security (WIFS), pages 1-6, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 462 + ], + "type": "text", + "content": "[58] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921-2929, 2015. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "text", + "content": "[59] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242-2251, 2017. 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18398" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_content_list.json b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e19f55a906c3f549fe0aedf45a4aa1177f4d28fb --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_content_list.json @@ -0,0 +1,1520 @@ +[ + { + "type": "text", + "text": "Semantic Equitable Clustering: A Simple and Effective Strategy for Clustering Vision Tokens", + "text_level": 1, + "bbox": [ + 86, + 130, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihang Fan $^{1,2}$ , Huaibo Huang $^{1*}$ , Mingrui Chen $^{1,2}$ , Ran He $^{1,2}$", + "bbox": [ + 243, + 202, + 723, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China \n $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China \nfanqihang.159@gmail.com, huaibo.huang@cripac.ia.ac.cn, pharmier@hust.edu.cn, rhe@nlpr.ia.ac.cn", + "bbox": [ + 112, + 220, + 854, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Vision Transformer (ViT) has gained prominence for its superior relational modeling prowess. However, its global attention mechanism's quadratic complexity poses substantial computational burdens. A common remedy spatially groups tokens for self-attention, reducing computational requirements. Nonetheless, this strategy neglects semantic information in tokens, possibly scattering semantically-linked tokens across distinct groups, thus compromising the efficacy of self-attention intended for modeling inter-token dependencies. Motivated by these insights, we introduce a fast and balanced clustering method, named Semantic Equitable Clustering (SEC). SEC clusters tokens based on their global semantic relevance in an efficient, straightforward manner. In contrast to traditional clustering methods requiring multiple iterations, our method achieves token clustering in a single pass. Additionally, SEC regulates the number of tokens per cluster, ensuring a balanced distribution for effective parallel processing on current computational platforms without necessitating further optimization. Capitalizing on SEC, we propose a versatile vision backbone, SECViT. Comprehensive experiments in image classification, object detection, instance segmentation, and semantic segmentation validate the effectiveness of SECViT. Moreover, SEC can be conveniently and swiftly applied to multimodal large language models (MLLM), such as LLaVA, to serve as a vision language connector, effectively accelerating the model's efficiency while maintaining unchanged or better performance.", + "bbox": [ + 75, + 358, + 473, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 808, + 209, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since its inception, the Vision Transformer (ViT)[11] has drawn considerable interest from the research community due to its robust modeling prowess. However, the quadratic", + "bbox": [ + 75, + 833, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfedae319fa83a0de59aba8a9fed4488ae3931ea412df57b82603b7d4dda41be.jpg", + "image_caption": [ + "(a) Window Partition", + "X Semantic Information $\\checkmark$ Cluster Efficiency $\\checkmark$ Equi-partition", + "Figure 1. Comparison among Window Partition, Dynamic Group by k-means, and Semantic Equitable Clustering. Our Semantic Equitable Clustering incorporates image semantics while maintaining efficient clustering, eliminating the need for iterative processes such as in k-means. Furthermore, it enables equipartitioning of tokens, promoting efficient GPU processing without necessitating additional CUDA optimization." + ], + "image_footnote": [], + "bbox": [ + 503, + 335, + 620, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d5a36a6d9ba18eca47eadf175213d7a8aec7ab9e2fefba6be9152e6d3af78d6.jpg", + "image_caption": [ + "(b) k-means", + "$\\checkmark$ Semantic Information \n $\\times$ Cluster Efficiency \n $\\times$ Equi-partition", + "(c)Semantic Equitable Clustering" + ], + "image_footnote": [], + "bbox": [ + 627, + 335, + 745, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ca7c9ba783fe071e548b9dfb0324f7fa73d27c460c85350ab380b43ec772af5.jpg", + "image_caption": [ + "$\\checkmark$ Semantic Information $\\checkmark$ Cluster Efficiency $\\checkmark$ Equi-partition" + ], + "image_footnote": [], + "bbox": [ + 754, + 335, + 874, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "complexity of Self-Attention leads to significant computational overhead, thus constraining the practicality of ViT. A variety of strategies have been devised to alleviate this computational load, the most prevalent of which involves token grouping, thereby constraining the attention span of each token[10, 38, 44, 47].", + "bbox": [ + 496, + 580, + 890, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Specifically, the Swin-Transformer [38] partitions tokens into multiple small windows, restricting token attention within each window. The CSWin-Transformer [10] adopts a cross-shaped grouping, endowing each token with a global receptive field. MaxViT [44] amalgamates window and grid attention, facilitating intra-window tokens to attend to their counterparts in other windows. However, these methods, solely reliant on spatial positioning, neglect token semantics, potentially restricting the self-attention's capacity to model semantic dependencies. To mitigate this, DGT [37] employs k-means clustering for query grouping, considering the semantic information of tokens for enhanced feature learning. Nonetheless, the iterative nature of k-means clustering and the potential for uneven token counts per cluster can impact the efficiency of parallel attention operations.", + "bbox": [ + 496, + 674, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Huaibo Huang is the corresponding author.", + "bbox": [ + 94, + 886, + 330, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4019", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/82c9bce75d7bf1a1970ecdfcfbe91b46fd54fe2042e2030049cec92004ce354c.jpg", + "image_caption": [ + "Figure 2. Left: Top-1 accuracy v.s. FLOPs on ImageNet-1K of recent SOTA models. Right: Comparison among different vision language connectors on LLaVA-1.5" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 467, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given these considerations, an optimal token partitioning scheme should efficiently segregate tokens, incorporate semantic information, and efficiently utilize computational resources (e.g., GPU). In response, we introduce a simple, fast and equitable clustering approach named Semantic Equitable Clustering (SEC). SEC segments tokens based on their relevance to global semantic information. Specifically, we employ global pooling to generate a global token encapsulating global semantic information. The similarity between this global token and all other tokens is then computed, reflecting global semantic relevance. Upon obtaining the similarity matrix, tokens (excluding the global token) are sorted by similarity scores, and the tokens with similar scores are grouped into clusters, ensuring uniform token distribution across clusters. As depicted in Fig. 1, SEC comprehensively considers token semantics and completes the clustering process in a single iteration, unlike the multi-iteration k-means. The resulting clusters, containing an equal number of tokens, can be processed in parallel by the GPU efficiently.", + "bbox": [ + 75, + 364, + 468, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon Semantic Equitable Clustering (SEC), we introduce the Semantic Equitable Clustering Vision Transformer (SECViT), a versatile vision backbone that is adaptable to a wide spectrum of downstream tasks. As demonstrated in Fig. 2, SECViT exhibits significant performance improvements compared to previous state-of-the-art (SOTA) models. Impressively, SECViT attains an accuracy of $84.3\\%$ utilizing merely 4.6GFLOPS, without the need for additional training data or supervision. This superior performance is maintained across different model scales. Furthermore, SECViT proves its proficiency in downstream tasks, including but not limited to, object detection, instance segmentation, and semantic segmentation.", + "bbox": [ + 75, + 670, + 468, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond vision tasks, we also apply SEC to multimodal large language models (MLLM) such as LLaVA-1.5 [36] to", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serve as an efficient vision language connector. Specifically, we use SEC to cluster the vision tokens, and then merge all the tokens at corresponding positions within each cluster into a single token. Experiments demonstrate that this approach significantly enhances the efficiency of LLaVA-1.5 while improving the model's performance.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 199, + 650, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformer. The Vision Transformer (ViT) [11] is considered a powerful visual architecture. Many works have improved the Vision Transformer, including enhancing its training efficiency and reducing its computational cost [10, 24, 38, 43, 61]. DeiT [43] uses distillation loss and incorporates extensive data augmentation methods into the ViT training process. Hierarchical structures represented by PVT [16, 41, 45, 46, 51] reduce the number of tokens in global attention by downsampling the keys and values (KV), thereby low the computational cost. In addition to them, some methods directly prune tokens based on their importance, retaining important tokens [32, 40]. This reduces the number of tokens and subsequently lowers the computational cost of the model. Another highly representative approach is to group all tokens such that each token can only attend to tokens within its own group [9, 10, 37, 38, 61]. This method also significantly reduces the computational cost of self-attention.", + "bbox": [ + 496, + 226, + 893, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grouping-Based Vision Transformer. Most grouping-based attention mechanisms perform grouping based on spatial structure [9, 10, 37, 38, 44]. Specifically, the SwinTransformer [38] divides all tokens into equally sized windows based on their spatial positions, where each token can only attend to tokens within its own window. This significantly reduces the model's computational cost. In addition to dividing tokens into small windows along the spatial dimension, DaViT [9] also splits channels into multiple groups along the channel dimension. Unlike the above methods that only consider positional information for grouping, DGT [37] takes semantic information into account by using k-means clustering to group the queries.", + "bbox": [ + 496, + 513, + 893, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Language Connector. The vision language connector is a critical component in MLLMs [3, 23, 36]. It aligns vision tokens with language tokens. Typical vision language connectors include MLP [36], Resampler [1], C-Abstractor [3], and others. Although MLP performs well, it introduces a significant number of vision tokens, which hampers the model's efficiency. On the other hand, connectors like Resampler improve the model's efficiency, but at the cost of reduced performance. Unlike these methods, our proposed SEC consider the semantic information of each token and significantly enhances the model's efficiency while maintaining its performance.", + "bbox": [ + 496, + 727, + 893, + 910 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4020", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5ac3a4564f17d87dc7600153d4694641f9a1af6bd66688dad7304a0f57e71375.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 869, + 246 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a61e8739d6a4850d2400576335e0fa1a8e01d661b4868da9f8fff3f2bc62a9de.jpg", + "image_caption": [ + "Figure 3. (a) Illustration of SECViT (b) Applying SEC to vision language connector. (c) Illustration of Semantic Equitable Clustering for ViT and vision language connector." + ], + "image_footnote": [], + "bbox": [ + 99, + 248, + 872, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 465, + 171, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 76, + 491, + 274, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of SECViT is shown in Fig. 3(a). SECViT consists of four stages with downsampling factors of $\\frac{1}{4}$ , $\\frac{1}{8}$ , $\\frac{1}{16}$ , and $\\frac{1}{32}$ , respectively. This structural design facilitates downstream tasks, such as object detection, in constructing feature pyramids. A SECViT block is composed of three modules. For each block, the input tensor $X_{in} \\in \\mathbb{R}^{C \\times H \\times W}$ is fed into the CPE to introduce the positional information. Then, The Self-Attention based on the Semantic Equitable Clustering (SEC) is employed to serve as the token mixer. The final FFN is utilized to integrate channel-wise information of tokens.", + "bbox": [ + 75, + 513, + 470, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beyond the design of the backbone, we also utilize SEC in the design of the vision language connector in MLLM [36]. For the vision tokens output by ViT, we use SEC to cluster the tokens. For each position corresponding to a cluster, we use attentive pooling to merge them into a single token, thereby reducing the number of vision tokens. The process is shown in Fig. 3(b).", + "bbox": [ + 75, + 681, + 468, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic Equitable Clustering", + "text_level": 1, + "bbox": [ + 76, + 800, + 349, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As previously mentioned, the design objectives of Semantic Equitable Clustering are threefold: 1) Fully consider the semantic information contained in different tokens during clustering. 2) Unlike k-means and other clustering methods that require multiple iterations, Semantic Equitable Cluster-", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing can complete clustering in a single step. 3) Ensure an equal number of tokens in each cluster to facilitate parallel processing on GPUs. In the following paragraphs, we will describe in detail how our Semantic Equitable Clustering achieves these three objectives. And the whole process is illustrated in the Fig. 3(c).", + "bbox": [ + 496, + 467, + 892, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single Clustering Center Related to Semantics. k-means is relatively complex for two reasons. First, it has multiple cluster centers, and each token needs to calculate its distance to each cluster center to determine its cluster membership. Second, the determination of each cluster center in k-means is not precise and requires multiple iterations to accurately establish the cluster centers.", + "bbox": [ + 496, + 574, + 892, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these two issues, we first discard the use of multiple cluster centers and instead calculate the distance between each token and a single center. Based on each token's distance to this center, we divide the tokens into different intervals. Then, to ensure that our chosen center contains the most comprehensive semantic information, we directly use the result of average pooling of all tokens as the center token. This is because, in most vision foundation models, the output of the average pool is assumed to contain the richest semantic information and is thus used for classification [6, 10, 12, 38]. Specifically, the process for determining the cluster center is shown in Eq. 1:", + "bbox": [ + 496, + 680, + 893, + 862 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {Q} X, K = W _ {K} X, V = W _ {V} X, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 867, + 890, + 891 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk _ {c} = \\operatorname {P o o l} (K).\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 887, + 754, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4021", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $W_{K}$ is a learnable matrix. $k_{c}$ is the determined cluster center. $X$ is the set of input tokens.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distance Metric Suitable for ViT. Unlike the Euclidean distance calculation used in the k-means algorithm for computing the distance between tokens, during the actual computation of Self-Attention, similarity between query and key is computed through dot product. To better adapt to the characteristics of Self-Attention, we also measure the distance between tokens using a method similar to dot product. Specifically, we calculate the cosine similarity between the cluster center and each token, and then sort the tokens according to the magnitude of the computed results. The specific process is shown in Eq. 2:", + "bbox": [ + 76, + 143, + 468, + 310 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns i m = \\frac {K \\cdot k _ {c}}{| | K | | \\cdot | | k _ {c} | |},\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 321, + 354, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni d x = \\operatorname {a r g s o r t} (s i m), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 352, + 468, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {*} = Q [ i d x ], K ^ {*} = K [ i d x ], V ^ {*} = V [ i d x ].\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 376, + 418, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $sim$ is the similarity matrix between $K$ and $k_{c}$ , the $\\operatorname{argsort}(sim)$ returns the indices of $sim$ sorted in descending order. $Q^{*}, K^{*}, V^{*}$ are $Q, K, V$ rearranged according to $\\operatorname{argsort}(sim)$ .", + "bbox": [ + 76, + 405, + 468, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equally Partition Tokens based on Distance. The obtained $Q^{*}, K^{*}$ , and $V^{*}$ from the previous step have been sorted based on their distances to the cluster center. For the design of vision backbone, we directly group them, so tokens with similar distances to the cluster center are classified into the same cluster. This allows us to directly control an equal number of tokens in each cluster. This process can be clearly illustrated in Fig. 3(c) and denoted as follows:", + "bbox": [ + 76, + 488, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {m} = Q ^ {*} [ m \\times N: (m + 1) N ],\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 623, + 383, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {m} = K ^ {*} [ m \\times N: (m + 1) N ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 641, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {m} = V ^ {*} [ m \\times N: (m + 1) N ].\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 660, + 383, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the basic token number of each cluster for equal partition and $m$ is the index of the cluster", + "bbox": [ + 76, + 689, + 468, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above steps, we have completed the clustering process that captures semantic information in the image with minimal sorting cost. Moreover, compared to k-means, we have achieved equi-partitioning of each cluster. After clustering is completed, we apply standard Self-Attention to the tokens within each cluster, thereby completing the interaction of information between tokens:", + "bbox": [ + 76, + 720, + 468, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {m} = \\operatorname {A t t n} \\left(Q _ {m}, K _ {m}, V _ {m}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 840, + 468, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the design of vision language connector, we group the tokens according to their similarity, and the tokens", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "within each group are interleaved, as shown in Eq. 5:", + "bbox": [ + 500, + 90, + 849, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {n} = Q ^ {*} [ n: N: L ],\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 119, + 767, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {n} = K ^ {*} [ n: N: L ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 138, + 890, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {n} = V ^ {*} [ n: N: L ].\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 157, + 767, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $L$ is the token's sequence length, $n$ is the index of group tokens. $N$ is the basic token number of each cluster. After obtaining the token groups, we perform pooling on $Q$ to effectively reduce the number of tokens input to the LLM, with each group's output becoming a single token, as shown in Eq 6.", + "bbox": [ + 498, + 184, + 890, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {n} = \\operatorname {A t t n} \\left(\\operatorname {P o o l} \\left(Q _ {n}\\right), K _ {n}, V _ {n}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 287, + 890, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Difference between SEC and EViT.", + "text_level": 1, + "bbox": [ + 500, + 316, + 803, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the most representative example, EViT [32], to illustrate the differences between SEC and other methods based on the similarity between the global token and other tokens.", + "bbox": [ + 498, + 340, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pruning v.s. Clustering. Most similarity-based methods, such as EViT, are pruning methods, where tokens with low similarity to the [cls] token are merged during the forward process, thereby reducing the number of tokens and decreasing computational cost. In contrast, our proposed SECViT employs a clustering-based approach, performing attention operations within each cluster.", + "bbox": [ + 498, + 387, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The role of the [cls] token. In methods like EViT, the [cls] token serves as a measure of importance of a token. Each token computes its similarity to the [cls] token, with higher similarity tokens deemed more important. The less important tokens are abandoned. In contrast, in SEC, the [cls] token (obtained by average pooling over all tokens) measures similarity between tokens. Each token computes its similarity score to the [cls] token; tokens with similar scores are considered to be more similar and grouped into one cluster. Attention is calculated only within the same cluster.", + "bbox": [ + 498, + 493, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 674, + 632, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first make strict comparison with hierarchical/plain baselines. Then we conduct experiments on a wide range of vision tasks for SECViT, including image classification, object detection, instance segmentation, and semantic segmentation. We also verify the role of SEC in MLLM based on LLaVA-1.5 [36]. More details, experiments, and comparison of models' efficiency can be found in the Appendix.", + "bbox": [ + 498, + 700, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. SEC for vision models", + "text_level": 1, + "bbox": [ + 500, + 816, + 705, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Strict Comparison with Baselines. We select two baselines: hierarchical backbone Swin-Transformer [38] and plain backbone DeiT [43] to make a comparison with our SEC based model. In the comparison models (SEC-Swin", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4022", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3b595e4907f31355ede69baec4904abb44339840974f98f930ec98ce67040d79.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
DeiT-S [43]224.6320479.844.540.143.0
EViT-DeiT-S (keeprate=0.9)224.0342879.8not suitnot suitnot suit
SEC-DeiT-S (num_cluster=4)224.1341280.5 (+0.7)47.7 (+3.2)42.7 (+2.6)47.5 (+4.5)
DeiT-B8617.6150281.8---
SEC-DeiT-B8614.8168282.4 (+0.6)---
Swin-T294.5172381.343.739.844.5
SEC-Swin-T294.8148283.8 (+2.5)48.3 (+4.6)43.4 (+3.6)49.3 (+4.8)
Swin-S508.8100683.045.741.147.6
SEC-Swin-S509.280485.0 (+2.0)50.2 (+4.5)44.7 (+3.6)51.3 (+3.7)
", + "bbox": [ + 86, + 78, + 460, + 291 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f5037ae6cb005af4cbb8916d1bbe35e765478d36308ea9fd60728bbabd12f5ce.jpg", + "table_caption": [ + "Table 1. Comparison with Hierarchy/Plain baselines. Inference speed are measured on the A100 GPU." + ], + "table_footnote": [], + "table_body": "
ModelParams (M)FLOPs (G)MethodPretrain epochAcc(%)
Swin-B [38]8815.4Supervised-83.5
ConvNeXt V2-B [50]8815.4Supervised-84.3
SEC-Swin-B8816.2Supervised-85.3
Swin-B [38]8815.4SimMIM [53]80084.0(+0.5)
ConvNeXt V2-B [50]8815.4FCMAE [50]80084.6(+0.3)
SEC-Swin-B8816.2SimMIM [53]80085.9(+0.6)
", + "bbox": [ + 83, + 325, + 465, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and SEC-DeiT), we merely substitute the attention mechanism in the original model with our SEC based Self-Attention and without introducing any other modules. As shown in Tab. 1, we conduct experiments on image classification, object detection, insatance segmentation and semantic segmentation, the simple replacement of the attention mechanism yields significant advantages in both performance and efficiency.", + "bbox": [ + 75, + 469, + 467, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the supervised scenario, we also train the model with SimMIM [53] in the self-supervised scenario. As shown in Tab. 2, SEC also performs exceptionally well in the self-supervised scenario.", + "bbox": [ + 75, + 590, + 467, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Classification. We compare our SECViT with numerous state-of-the-art models, the results are shown in Tab.3. We adopt the training strategy proposed in DeiT [43], with the only supervision is cross entropy loss. All of our models are trained from scratch for 300 epochs with the input resolution of $224 \\times 224$ . SECViT consistently outperforms preceding models across all scales. Notably, SECViT-S attains a Top1-accuracy of $84.3\\%$ with a mere 27M parameters and 4.6G FLOPs. The comparison of the models' efficiency can be found in Appendix.", + "bbox": [ + 75, + 669, + 467, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object Detection and Instance Segmentation. We utilize MMDetection [4] to implement Mask-RCNN [21], Cascade Mask R-CNN [2], and RetinaNet [33] to evaluate the performance of the SECViT. Tab. 5 and Tab. 4 show the", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/645e9c7be184e6fea1bb24e8fe9ec185ec719a723beef8558815a8fdfc664342.jpg", + "table_caption": [ + "Table 2. Comparison with baselines on self-supervised setting." + ], + "table_footnote": [], + "table_body": "
CostModelParmas (M)FLOPs (G)Top1-acc (%)
tiny model ~ 2.5GPVTv2-b1 [46]132.178.7
TCFormer-light [59]143.879.4
QuadTree-B-b1 [42]142.380.0
MPViT-XS [28]112.980.9
BiFormer-T [61]132.281.4
CrossFormer-T [47]282.981.5
FAT-B2 [12]142.081.9
GC-ViT-XT [20]202.682.0
SMT-T [34]122.482.2
RMT-T [13]142.582.4
SECViT-T152.582.7
small model ~ 4.5GPS-ViT-B14 [58]215.481.7
DVT-T2T-ViT-19 [49]396.281.9
ConvNeXt-T [39]294.582.1
TCFormer [59]265.882.3
SG-Former-S [15]234.883.2
StructViT-S-8-1 [25]245.483.3
InternImage-T [48]305.083.5
MLLA-T [18]254.283.5
MaxViT-T [44]315.683.6
FAT-B3 [12]294.483.6
SMT-S [34]204.883.7
BiFormer-S [61]264.583.8
SECViT-S274.684.3
base model ~ 9.0GConvNeXt-S [39]508.783.1
NAT-S [19]517.883.7
Quadtree-B-b4 [42]6411.584.0
MOAT-1 [54]429.184.2
InternImage-S [48]508.084.2
GC-ViT-S [20]518.584.3
BiFormer-B [61]579.884.3
iFormer-B [41]489.484.6
FAT-B4 [12]529.384.8
SECViT-B579.885.2
large model ~ 18.0GCrossFormer-L [47]9216.184.0
SMT-L [34]8117.784.6
DaViT-B [9]8815.584.6
SG-Former-B [15]7815.684.7
iFormer-L [41]8714.084.8
InterImage-B [48]9716.084.9
GC-ViT-B [20]9014.885.0
RMT-L [13]9518.285.5
SECViT-L10118.285.7
", + "bbox": [ + 519, + 88, + 880, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Comparison with the state-of-the-art on ImageNet-1K classification.", + "bbox": [ + 500, + 625, + 890, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "results of SECViT with different detection frameworks. The results show that SECViT performs better than its counterparts in all comparisons.", + "bbox": [ + 498, + 667, + 890, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Segmentation. We utilize Semantic FPN [26] and UperNet [52] to validate our SECViT's performance, implementing these frameworks via MMSegmentation [7]. The results of semantic segmentation can be found in the Tab. 6. All the FLOPs are measured with the input resolution of $512 \\times 2048$ . SECViT achieves the best performance in all settings.", + "bbox": [ + 496, + 731, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. SEC for MLLM", + "text_level": 1, + "bbox": [ + 500, + 845, + 660, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SEC can greatly facilitate the design of vision language connectors in MLLMs. First, we conduct a rigorous compari", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4023", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/30cb13a7c8ecab4dd19dc323b6db6be27d56a6c8b1818e27935212006b10083b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneParams (M)FLOPs (G)Mask R-CNN 1×Params (M)FLOPs (G)RetinaNet 1×
APbAPb50APb75APmAPm50APm75APbAPb50APb75APsAPbAPL
PVTv2-B1 [46]3324341.854.345.938.861.241.62322541.261.943.925.444.554.3
FAT-B2 [12]3321545.267.949.041.364.644.02319644.065.247.227.547.758.8
RMT-T [13]3321847.168.851.742.665.845.92319945.166.248.128.848.961.1
SECViT-T3422147.869.552.543.066.746.32420245.866.849.229.149.860.9
MPViT-S [28]4326846.468.651.242.465.645.73224845.757.348.828.749.759.2
MLLA-T [18]4425546.869.551.542.166.445.0--------
STViT-S [22]4425247.670.052.343.166.846.5--------
RMT-S [13]4626249.070.853.943.967.847.43624447.869.151.832.151.863.5
SECViT-S4526249.970.954.744.668.347.73524048.469.452.031.353.363.8
ScalableViT-B [56]9534946.868.751.542.565.845.98533045.867.349.229.949.561.0
InternImage-S [48]6934047.869.852.843.367.146.7--------
MLLA-S [18]6331949.271.553.944.268.547.2--------
STViT-B [22]7035949.771.754.744.868.948.7--------
SECViT-B7637151.572.956.745.469.948.76334949.370.352.932.053.864.8
Focal-B [55]11053347.870.252.543.267.346.510151446.368.049.831.750.460.8
CSwin-B [10]9752648.770.453.943.967.847.3--------
InternImage-B [48]11550148.870.954.044.067.847.4--------
MLLA-B [18]11550250.572.055.445.069.348.6--------
SECViT-L11955052.073.557.346.370.649.810552750.271.453.933.254.566.3
", + "bbox": [ + 117, + 77, + 849, + 330 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8fbbe1628c6000ad91d39777b38ef9a2b5f60ca0170a1966b2b3e1f8fcedee81.jpg", + "table_caption": [ + "Table 4. Comparison to other backbones using \"1×\" schedule on COCO." + ], + "table_footnote": [], + "table_body": "
BackboneParams (M)FLOPs (G)\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Mask R-CNN 3×+MS
GC-ViT-T [20]4829147.970.152.843.267.046.7
MLLA-T [18]4425548.871.053.643.868.046.8
SMT-S [34]4026549.070.153.443.467.346.7
InternImage-T [48]4927049.170.454.143.767.347.3
RMT-S [13]4626250.771.955.644.969.148.4
SECViT-S4526251.672.555.945.669.948.8
NAT-S [19]7033048.469.853.243.266.946.4
InternImage-S [48]6934049.771.154.544.568.547.8
SMT-B [34]5232849.871.054.444.068.047.3
MLLA-S [18]6331950.571.855.244.969.148.2
RMT-B [13]7337352.272.957.046.170.449.9
SECViT-B7537152.873.657.746.470.849.9
Cascade Mask R-CNN 3×+MS
GC-ViT-T [20]8577051.670.456.144.667.848.3
SMT-S [34]7874451.970.556.344.767.848.6
UniFormer-S [30]7974752.171.156.645.268.348.9
RMT-S [13]8374153.272.057.846.169.849.8
SECViT-S8374154.172.858.647.070.351.0
NAT-S [19]10880951.970.456.244.968.248.6
GC-ViT-S [20]10886652.471.057.145.468.549.3
CSWin-S [10]9282053.772.258.446.469.650.6
UniFormer-B [30]10787853.872.858.546.469.950.4
RMT-B [13]11185254.572.859.047.270.551.4
SECViT-B11484955.474.159.947.871.751.7
", + "bbox": [ + 96, + 349, + 457, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "son between SEC and various baseline vision language connectors based on LLaVA-1.5. Then, we compare LLaVA-1.5+SEC with several popular contemporary MLLMs.", + "bbox": [ + 76, + 715, + 467, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Strict Comparison with Baselines. In Tab. 7, we strictly compare various commonly used vision language connectors, including MLP, Resampler [1], Pooling, and EViT [32], which has achieved success in the design of ViT. Among these, MLP is the original design in LLaVA-1.5 [36], capable of achieving good results. However, it incurs significant computational cost due to the excessive vision tokens. To address this issue, some connectors at", + "bbox": [ + 76, + 779, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f3df275ca490173923467b7266a9a53400b06adca108bc4a933aeaed53d1e182.jpg", + "table_caption": [ + "Table 5. Comparison with other backbones using $3 \\times +\\mathrm{MS}$ schedule on COCO." + ], + "table_footnote": [], + "table_body": "
ModelSemantic FPN 80K Params FLOPsUpernet 160K Params FLOPs
(M)(G)(%)(M)(G)(%)
PVTv2-B1 [46]1813642.5---
VAN-B1 [17]1814042.9---
RMT-T [13]1713646.4---
SECViT-T1813647.24489448.8
StructViT-S [25]2627146.9---
MogaNet-S [31]2918947.75594649.2
SMT-S [34]---5093549.2
SGFormer-S [15]2520549.052.598949.9
RMT-S [13]3018049.45693749.8
SECViT-S3018049.65693650.6
MogaNet-B [31]---74105050.1
InterImage-S [48]---80101750.2
StructViT-B [25]5452948.5---
RMT-B [13]5729450.483105152.0
SECViT-B6029150.786104852.2
MogaNet-L [31]---113117650.9
MLLA-B [18]---128118351.9
SGFormer-B [15]8147550.6109130452.0
RMT-L [13]9848251.4125124152.8
SECViT-L10347552.2131125653.8
", + "bbox": [ + 524, + 359, + 867, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6. Comparison with the state-of-the-art on ADE20K.", + "bbox": [ + 519, + 667, + 869, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tempt to use fewer vision tokens to accelerate LLaVA-1.5. Nonetheless, these adjustments inevitably lead to performance degradation. The results in Tab. 7 show that using SEC can effectively accelerate the inference of LLaVA-1.5 without causing performance degradation, and can even improve the performance of LLaVA-1.5 to a certain extent.", + "bbox": [ + 496, + 698, + 890, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Popular MLLMs. In Tab. 8 and Tab. 9, we compare LLaVA-1.5 equipped with SEC as a vision-language connector with other MLLMs. It is evident that SEC not only enhances the performance of MLLMs across various benchmarks but also significantly improves the efficiency of the models. This fully demonstrates the", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4024", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff193ad143bc4fe3d22a757cc80a23d5395898256a23378743805a5e730243f8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelConnectorV-T NumTimeSpeedTextVQAGQAVQAv2POPEMME
LLaVA-1.5MLP576+1194s1.0×58.262.078.586.11510.7
LLaVA-1.5+ResamplerResampler288+1126s1.5×52.156.876.083.11393.2
LLaVA-1.5+EViTMLP+EViT288+1126s1.5×54.660.077.984.31483.2
LLaVA-1.5+SECMLP+SEC288+1126s1.5×60.163.578.987.71510.7
LLaVA-1.5+ResamplerResampler256+1116s1.7×51.656.075.282.71387.2
LLaVA-1.5+PoolMLP+Pool256+1116s1.7×52.457.676.483.31415.5
LLaVA-1.5+EViTMLP+EViT256+1116s1.7×52.859.677.183.71443.7
LLaVA-1.5+SECMLP+SEC256+1116s1.7×59.663.278.687.11505.2
LLaVA-1.5+ResamplerResampler192+1102s1.9×50.155.274.382.71337.6
LLaVA-1.5+EViTMLP+EViT192+1102s1.9×51.658.676.383.11427.6
LLaVA-1.5+SECMLP+SEC192+1102s1.9×57.762.778.486.71500.1
LLaVA-1.5+ResamplerResampler144+194s2.1×47.654.672.081.91293.7
LLaVA-1.5+PoolMLP+Pool144+194s2.1×50.056.273.681.91310.7
LLaVA-1.5+EViTMLP+EViT144+194s2.1×51.258.076.083.11393.6
LLaVA-1.5+SECMLP+SEC144+194s2.1×56.862.078.086.11487.1
", + "bbox": [ + 125, + 77, + 843, + 275 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab4eade270e932acb2b264e8e6112de399977f63626d29c8b15ff839c83aad8f.jpg", + "table_caption": [ + "Table 7. Comparison of different vision language connectors on LLaVA-1.5. \"V-T Num\" denotes the quantity of visual tokens. The computation expense is impacted by V-T Num, with larger values resulting in higher costs. \"Speed\" refers to the comparative inference velocity relative to LLaVA-1.5. \"Time\" is the average inference time. Inference speed are measured on the A100." + ], + "table_footnote": [], + "table_body": "
ModelLLMConnectorV-T NumResTextVQAGQAVQAv2VisWiz\\( SQA_{img} \\)Speed (↑)
7B LLM
Shikra [5]Vicuna-7BMLP257224--77.4---
IDEFICS-9B [27]LLaMA-7BCross Attn257224-38.450.935.5--
Qwen-VL [1]Qwen-7BResampler256448-59.378.835.267.1-
Qwen-VL-Chat [1]Qwen-7BResampler256448-57.578.238.968.2-
LLaVA-1.5 [35]Vicuna-7BMLP57733658.262.078.550.066.81.0×
LLaVA-1.5+SEC (ours)Vicuna-7BMLP+SEC25733659.663.278.952.869.61.7×
13B LLM
InstructBLIP [8]Vicuna-13BQ-Former32224-49.5-33.463.1-
BLIP-2 [29]Vicuna-13BQ-Former32224-41.041.019.561.0-
LLaVA-1.5 [35]Vicuna-13BMLP57733661.263.380.053.671.61.0×
LLaVA1.5+SEC (ours)Vicuna-13BMLP+SEC25733662.364.380.054.772.01.8×
", + "bbox": [ + 117, + 321, + 854, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Results on General VQA tasks.", + "bbox": [ + 364, + 494, + 602, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "effectiveness of SEC in extracting visual information.", + "bbox": [ + 76, + 523, + 431, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 550, + 230, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present some of the ablation study results for SEC, and more results can be found in the Appendix.", + "bbox": [ + 76, + 574, + 468, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Vision Tokens in Each Clusters. The number of vision tokens has a significant impact on the performance and speed of the model. We thoroughly investigate the effect of the number of vision tokens on SECViT. As shown in Tab. 10, the number of vision tokens in each cluster greatly influences the model's performance. Specifically, in downstream dense prediction tasks, having too few tokens in each cluster leads to substantial performance degradation. When the number of tokens in each cluster is too large, the model's performance does not see a significant improvement, but its speed decreases.", + "bbox": [ + 75, + 625, + 468, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Visualization of SEC", + "text_level": 1, + "bbox": [ + 76, + 800, + 272, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further understand the working mechanism of SEC, we visualize some clustering results for SECViT. As shown in Fig. 4, the left side presents the clustering results of vision tokens at different stages of the model. From the clustering results, we analyze that in the shallow layers, the model", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "distinguishes fine-grained features well, while in the deeper layers, it captures global semantic features effectively. The right side shows the Grad-CAM diagrams at different stages of the model, from which we can draw similar conclusions to the clustering results. More visualization results can be found in Appendix.", + "bbox": [ + 496, + 523, + 890, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Vision Tokens Outputs by SEC. MLLM is quite sensitive to the number of vision tokens. We conduct a detailed exploration based on LLaVA-1.5 regarding the number of vision tokens output by SEC, as shown in Tab. 11. The first row represents the speed and performance of the original LLaVA-1.5 without using SEC. Compared to LLaVA-1.5, employing SEC effectively reduces the number of vision tokens and improves training efficiency. As the number of vision tokens decreases, the model's performance shows a slight decline, but its efficiency is further enhanced.", + "bbox": [ + 496, + 633, + 892, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 814, + 619, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We propose a simple and straightforward clustering method for vision tokens—Semantic Equitable Clustering (SEC). This method assigns each token to a cluster by calculating the similarity between each token and a global token, and", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4025", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c1bba0b8bb52a32dc26933269128d730ddd2a9865c45bfa9fe9a360ff72e4cc1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLLMConnectorV-T NumResPOPEMMBMM-VetSpeed (↑)
7B LLM
MiniGPT-4 [60]Vicuna-7BResampler3222472.224.322.1-
mPLUG-Ow12 [57]LLaMA2-7BResampler32224-49.4--
LLaMA-AdapterV2 [14]LLaMA2-7BLLaMA-Adapter257224-41.031.4-
Shikra [5]Vicuna-7BMLP257224-58.8--
Qwen-VL [1]Qwen-7BResampler256448-38.2--
Qwen-VL-Chat [1]Qwen-7BResampler256448-60.6--
LLaVA-1.5 [35]Vicuna-7BMLP57733686.164.331.11.0×
LLaVA1.5+SEC (ours)Vicuna-7BMLP+SEC14533686.168.431.72.1×
13B LLM
MiniGPT-4 [60]Vicuna-13BResampler32224--24.4-
BLIP-2 [29]Vicuna-13BQ-Former3222485.3-22.4-
LLaVA-1.5 [35]Vicuna-13BMLP57733686.267.736.11.0×
LLaVA-1.5+SEC (ours)Vicuna-13BMLP+SEC14533686.469.237.32.2×
", + "bbox": [ + 102, + 88, + 867, + 284 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c760fd9f8f0d078c2bb8be621b61cf32610c29ee8b18ee63cb67cebf90532cad.jpg", + "table_caption": [ + "Table 9. Results on benchmark designed for MLLMs." + ], + "table_footnote": [], + "table_body": "
V-T numParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
98152.5200482.747.843.047.2
196153.1172283.0 (+0.3)48.2 (+0.4)43.4 (+0.4)47.5 (+0.3)
64152.5194682.7 (+0.0)47.8 (+0.0)42.8 (-0.2)46.9 (-0.3)
49152.4210282.6 (-0.1)47.5 (-0.3)42.7 (-0.3)47.7 (-0.5)
24152.3218682.0 (-0.7)45.9 (-1.9)40.6 (-2.4)44.6 (-2.6)
", + "bbox": [ + 98, + 306, + 447, + 455 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/739005363da79434f53381e97d751e222d3b9da27106c33b781ac1e713904cbb.jpg", + "table_caption": [ + "Table 10. Effect of the number of vision tokens in each cluster. \"V-T num\" means the number of vision tokens in each cluster. The experiments are conducted based on SECViT-T." + ], + "table_footnote": [], + "table_body": "
V-T numTime SpeedTextVQAGQAVQAv2POPEMM-Vet
576+121h1.0×58.262.078.586.131.1
288+114h1.5×60.1(+1.9)63.5(+1.5)78.9(+0.4)87.7(+1.6)33.2(+2.1)
256+113h1.6×59.6(+1.4)63.2(+0.3)78.6(+0.1)87.1(+1.0)32.7(+1.6)
192+111h1.9×57.7(-0.5)62.7(+0.7)78.4(-0.1)86.7(+0.6)32.1(+1.0)
144+110h2.1×56.8(-1.4)62.0(+0.0)78.0(-0.5)86.1(+0.0)31.7(+0.6)
", + "bbox": [ + 81, + 503, + 468, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11. Effect of the number of vision tokens outputs by SEC. \"V-T num\" means the number of vision tokens output by SEC. The experiments are conducted based on LLaVA-1.5 [36].", + "bbox": [ + 76, + 590, + 468, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "completes the whole clustering process in only one step. Our clustering method takes into account the semantic information contained in the tokens, and ensures an equal number of tokens in each cluster, facilitating efficient parallel processing on modern GPUs. Based on Semantic Equitable Clustering, we designed SECViT, a versatile vision backbone that achieves impressive results across various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Besides, SEC can also be conveniently applied to multimodal large language models (MLLM) to serve as a vision language connector and benefits the model's efficiency.", + "bbox": [ + 75, + 648, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 845, + 264, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially funded by Beijing Natural Science Foundation (4252054), Youth Innovation Promotion", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dad77df489835b054acdb275ca0d0404b2aec373fd532a5ac9ebaf2e3f7b9f41.jpg", + "image_caption": [ + "Figure 4. Visualization for SEC." + ], + "image_footnote": [], + "bbox": [ + 500, + 305, + 888, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Association CAS(Grant No.2022132), Beijing Nova Program(20230484276), and CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024005).", + "bbox": [ + 498, + 825, + 890, + 872 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4026", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2, 6, 7, 8", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 5", + "[3] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In CVPR, 2024. 2", + "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5", + "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 7, 8", + "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In ICLR, 2023. 3", + "[7] MMSegmentation Contributors. Mmsegmentation, an open source semantic segmentation toolbox, 2020. 5", + "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7", + "[9] Mingyu Ding, Bin Xiao, Noel Codella, et al. Davit: Dual attention vision transformers. In ECCV, 2022. 2, 5", + "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, et al. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 3, 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2", + "[12] Qihang Fan, Huaibo Huang, Xiaogiang Zhou, and Ran He. Lightweight vision transformer with bidirectional interaction. In NeurIPS, 2023. 3, 5, 6", + "[13] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 5, 6", + "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 8", + "[15] SG-Former: Self guided Transformer with Evolving Token Reallocation. Sucheng ren, xingyi yang, songhua liu, xinchao wang. In ICCV, 2023. 5, 6", + "[16] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 2", + "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6", + "[18] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 5, 6", + "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 5, 6", + "[20] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In ICML, 2023. 5, 6", + "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask r-cnn. In ICCV, 2017. 5", + "[22] Huaibo Huang, Xiaogiang Zhou, Jie Cao, Ran He, and Tieniu Tan. Vision transformer with super token sampling. In CVPR, 2023. 6", + "[23] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver: General perception with iterative attention, 2021. 2", + "[24] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 2", + "[25] Manjin Kim, Paul Hongsuck Seo, Cordelia Schmid, and Minsu Cho. Learning correlation structures for vision transformers. In CVPR, 2024. 5, 6", + "[26] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 5", + "[27] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 7", + "[28] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, 2022. 5, 6", + "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 7, 8", + "[30] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning, 2022. 6", + "[31] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z. Li. Moganet: Multi-order gated aggregation network. In ICLR, 2024. 6", + "[32] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 2, 4, 6", + "[33] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, and Kaiming He and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5", + "[34] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4027", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 7, 8", + "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 3, 4, 6, 8", + "[37] Kai Liu, Tianyi Wu, Cong Liu, and Guodong Guo. Dynamic group transformer: A general vision transformer backbone with dynamic group attention. In IJCAI, 2022. 1, 2", + "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 3, 4, 5", + "[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, et al. A convnet for the 2020s. In CVPR, 2022. 5", + "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, 2021. 2", + "[41] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng YAN. Inception transformer. In NeurIPS, 2022. 2, 5", + "[42] Shitao Tang, Jiahui Zhang, Siyu Zhu, et al. Quadtree attention for vision transformers. In ICLR, 2022. 5", + "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, et al. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2, 4, 5", + "[44] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 1, 2, 5", + "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 2", + "[46] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):1-10, 2022. 2, 5, 6", + "[47] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. In ICLR, 2022. 1, 5", + "[48] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 5, 6", + "[49] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic vision transformers with adaptive sequence length. In NeurIPS, 2021. 5", + "[50] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. 5", + "[51] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 5", + "[53] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 5", + "[54] Chenglin Yang, Siyuan Qiao, Qihang Yu, et al. Moat: Alternating mobile convolution and attention brings strong vision models. In ICLR, 2023. 5", + "[55] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. In NeurIPS, 2021. 6", + "[56] Rui Yang, Hailong Ma, Jie Wu, Yansong Tang, Xuefeng Xiao, Min Zheng, and Xiu Li. Scalablevit: Rethinking the context-oriented generalization of vision transformer. In ECCV, 2022. 6", + "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024. 8", + "[58] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In ICCV, 2021. 5", + "[59] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In CVPR, 2022. 5", + "[60] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 8", + "[61] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2, 5" + ], + "bbox": [ + 501, + 90, + 890, + 614 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4028", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_model.json b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82d0d308780804881e7158ea36a97f0632822dfe --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Semantic Equitable Clustering: A Simple and Effective Strategy for Clustering Vision Tokens" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.203, + 0.724, + 0.221 + ], + "angle": 0, + "content": "Qihang Fan \\(^{1,2}\\), Huaibo Huang\\(^{1*}\\), Mingrui Chen\\(^{1,2}\\), Ran He\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.221, + 0.856, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China \n\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China \nfanqihang.159@gmail.com, huaibo.huang@cripac.ia.ac.cn, pharmier@hust.edu.cn, rhe@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.784 + ], + "angle": 0, + "content": "The Vision Transformer (ViT) has gained prominence for its superior relational modeling prowess. However, its global attention mechanism's quadratic complexity poses substantial computational burdens. A common remedy spatially groups tokens for self-attention, reducing computational requirements. Nonetheless, this strategy neglects semantic information in tokens, possibly scattering semantically-linked tokens across distinct groups, thus compromising the efficacy of self-attention intended for modeling inter-token dependencies. Motivated by these insights, we introduce a fast and balanced clustering method, named Semantic Equitable Clustering (SEC). SEC clusters tokens based on their global semantic relevance in an efficient, straightforward manner. In contrast to traditional clustering methods requiring multiple iterations, our method achieves token clustering in a single pass. Additionally, SEC regulates the number of tokens per cluster, ensuring a balanced distribution for effective parallel processing on current computational platforms without necessitating further optimization. Capitalizing on SEC, we propose a versatile vision backbone, SECViT. Comprehensive experiments in image classification, object detection, instance segmentation, and semantic segmentation validate the effectiveness of SECViT. Moreover, SEC can be conveniently and swiftly applied to multimodal large language models (MLLM), such as LLaVA, to serve as a vision language connector, effectively accelerating the model's efficiency while maintaining unchanged or better performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.809, + 0.21, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.834, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Since its inception, the Vision Transformer (ViT)[11] has drawn considerable interest from the research community due to its robust modeling prowess. However, the quadratic" + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.326, + 0.612, + 0.336 + ], + "angle": 0, + "content": "(a) Window Partition" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.336, + 0.621, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.431, + 0.614, + 0.461 + ], + "angle": 0, + "content": "X Semantic Information \\(\\checkmark\\) Cluster Efficiency \\(\\checkmark\\) Equi-partition" + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.326, + 0.71, + 0.336 + ], + "angle": 0, + "content": "(b) k-means" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.336, + 0.746, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.431, + 0.739, + 0.461 + ], + "angle": 0, + "content": "\\(\\checkmark\\) Semantic Information \n\\(\\times\\) Cluster Efficiency \n\\(\\times\\) Equi-partition" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.326, + 0.887, + 0.336 + ], + "angle": 0, + "content": "(c)Semantic Equitable Clustering" + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.336, + 0.875, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.431, + 0.867, + 0.461 + ], + "angle": 0, + "content": "\\(\\checkmark\\) Semantic Information \\(\\checkmark\\) Cluster Efficiency \\(\\checkmark\\) Equi-partition" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Comparison among Window Partition, Dynamic Group by k-means, and Semantic Equitable Clustering. Our Semantic Equitable Clustering incorporates image semantics while maintaining efficient clustering, eliminating the need for iterative processes such as in k-means. Furthermore, it enables equipartitioning of tokens, promoting efficient GPU processing without necessitating additional CUDA optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.672 + ], + "angle": 0, + "content": "complexity of Self-Attention leads to significant computational overhead, thus constraining the practicality of ViT. A variety of strategies have been devised to alleviate this computational load, the most prevalent of which involves token grouping, thereby constraining the attention span of each token[10, 38, 44, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Specifically, the Swin-Transformer [38] partitions tokens into multiple small windows, restricting token attention within each window. The CSWin-Transformer [10] adopts a cross-shaped grouping, endowing each token with a global receptive field. MaxViT [44] amalgamates window and grid attention, facilitating intra-window tokens to attend to their counterparts in other windows. However, these methods, solely reliant on spatial positioning, neglect token semantics, potentially restricting the self-attention's capacity to model semantic dependencies. To mitigate this, DGT [37] employs k-means clustering for query grouping, considering the semantic information of tokens for enhanced feature learning. Nonetheless, the iterative nature of k-means clustering and the potential for uneven token counts per cluster can impact the efficiency of parallel attention operations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.331, + 0.9 + ], + "angle": 0, + "content": "*Huaibo Huang is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4019" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.468, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.472, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Left: Top-1 accuracy v.s. FLOPs on ImageNet-1K of recent SOTA models. Right: Comparison among different vision language connectors on LLaVA-1.5" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Given these considerations, an optimal token partitioning scheme should efficiently segregate tokens, incorporate semantic information, and efficiently utilize computational resources (e.g., GPU). In response, we introduce a simple, fast and equitable clustering approach named Semantic Equitable Clustering (SEC). SEC segments tokens based on their relevance to global semantic information. Specifically, we employ global pooling to generate a global token encapsulating global semantic information. The similarity between this global token and all other tokens is then computed, reflecting global semantic relevance. Upon obtaining the similarity matrix, tokens (excluding the global token) are sorted by similarity scores, and the tokens with similar scores are grouped into clusters, ensuring uniform token distribution across clusters. As depicted in Fig. 1, SEC comprehensively considers token semantics and completes the clustering process in a single iteration, unlike the multi-iteration k-means. The resulting clusters, containing an equal number of tokens, can be processed in parallel by the GPU efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Building upon Semantic Equitable Clustering (SEC), we introduce the Semantic Equitable Clustering Vision Transformer (SECViT), a versatile vision backbone that is adaptable to a wide spectrum of downstream tasks. As demonstrated in Fig. 2, SECViT exhibits significant performance improvements compared to previous state-of-the-art (SOTA) models. Impressively, SECViT attains an accuracy of \\(84.3\\%\\) utilizing merely 4.6GFLOPS, without the need for additional training data or supervision. This superior performance is maintained across different model scales. Furthermore, SECViT proves its proficiency in downstream tasks, including but not limited to, object detection, instance segmentation, and semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Beyond vision tasks, we also apply SEC to multimodal large language models (MLLM) such as LLaVA-1.5 [36] to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "serve as an efficient vision language connector. Specifically, we use SEC to cluster the vision tokens, and then merge all the tokens at corresponding positions within each cluster into a single token. Experiments demonstrate that this approach significantly enhances the efficiency of LLaVA-1.5 while improving the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.2, + 0.651, + 0.216 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.895, + 0.501 + ], + "angle": 0, + "content": "Vision Transformer. The Vision Transformer (ViT) [11] is considered a powerful visual architecture. Many works have improved the Vision Transformer, including enhancing its training efficiency and reducing its computational cost [10, 24, 38, 43, 61]. DeiT [43] uses distillation loss and incorporates extensive data augmentation methods into the ViT training process. Hierarchical structures represented by PVT [16, 41, 45, 46, 51] reduce the number of tokens in global attention by downsampling the keys and values (KV), thereby low the computational cost. In addition to them, some methods directly prune tokens based on their importance, retaining important tokens [32, 40]. This reduces the number of tokens and subsequently lowers the computational cost of the model. Another highly representative approach is to group all tokens such that each token can only attend to tokens within its own group [9, 10, 37, 38, 61]. This method also significantly reduces the computational cost of self-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.895, + 0.713 + ], + "angle": 0, + "content": "Grouping-Based Vision Transformer. Most grouping-based attention mechanisms perform grouping based on spatial structure [9, 10, 37, 38, 44]. Specifically, the SwinTransformer [38] divides all tokens into equally sized windows based on their spatial positions, where each token can only attend to tokens within its own window. This significantly reduces the model's computational cost. In addition to dividing tokens into small windows along the spatial dimension, DaViT [9] also splits channels into multiple groups along the channel dimension. Unlike the above methods that only consider positional information for grouping, DGT [37] takes semantic information into account by using k-means clustering to group the queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.895, + 0.911 + ], + "angle": 0, + "content": "Vision Language Connector. The vision language connector is a critical component in MLLMs [3, 23, 36]. It aligns vision tokens with language tokens. Typical vision language connectors include MLP [36], Resampler [1], C-Abstractor [3], and others. Although MLP performs well, it introduces a significant number of vision tokens, which hampers the model's efficiency. On the other hand, connectors like Resampler improve the model's efficiency, but at the cost of reduced performance. Unlike these methods, our proposed SEC consider the semantic information of each token and significantly enhances the model's efficiency while maintaining its performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4020" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.87, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.249, + 0.873, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.427, + 0.894, + 0.457 + ], + "angle": 0, + "content": "Figure 3. (a) Illustration of SECViT (b) Applying SEC to vision language connector. (c) Illustration of Semantic Equitable Clustering for ViT and vision language connector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.172, + 0.483 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.492, + 0.275, + 0.506 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.471, + 0.681 + ], + "angle": 0, + "content": "The overall architecture of SECViT is shown in Fig. 3(a). SECViT consists of four stages with downsampling factors of \\(\\frac{1}{4}\\), \\(\\frac{1}{8}\\), \\(\\frac{1}{16}\\), and \\(\\frac{1}{32}\\), respectively. This structural design facilitates downstream tasks, such as object detection, in constructing feature pyramids. A SECViT block is composed of three modules. For each block, the input tensor \\(X_{in} \\in \\mathbb{R}^{C \\times H \\times W}\\) is fed into the CPE to introduce the positional information. Then, The Self-Attention based on the Semantic Equitable Clustering (SEC) is employed to serve as the token mixer. The final FFN is utilized to integrate channel-wise information of tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Beyond the design of the backbone, we also utilize SEC in the design of the vision language connector in MLLM [36]. For the vision tokens output by ViT, we use SEC to cluster the tokens. For each position corresponding to a cluster, we use attentive pooling to merge them into a single token, thereby reducing the number of vision tokens. The process is shown in Fig. 3(b)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.35, + 0.818 + ], + "angle": 0, + "content": "3.2. Semantic Equitable Clustering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As previously mentioned, the design objectives of Semantic Equitable Clustering are threefold: 1) Fully consider the semantic information contained in different tokens during clustering. 2) Unlike k-means and other clustering methods that require multiple iterations, Semantic Equitable Cluster-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.893, + 0.559 + ], + "angle": 0, + "content": "ing can complete clustering in a single step. 3) Ensure an equal number of tokens in each cluster to facilitate parallel processing on GPUs. In the following paragraphs, we will describe in detail how our Semantic Equitable Clustering achieves these three objectives. And the whole process is illustrated in the Fig. 3(c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Single Clustering Center Related to Semantics. k-means is relatively complex for two reasons. First, it has multiple cluster centers, and each token needs to calculate its distance to each cluster center to determine its cluster membership. Second, the determination of each cluster center in k-means is not precise and requires multiple iterations to accurately establish the cluster centers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.895, + 0.863 + ], + "angle": 0, + "content": "To address these two issues, we first discard the use of multiple cluster centers and instead calculate the distance between each token and a single center. Based on each token's distance to this center, we divide the tokens into different intervals. Then, to ensure that our chosen center contains the most comprehensive semantic information, we directly use the result of average pooling of all tokens as the center token. This is because, in most vision foundation models, the output of the average pool is assumed to contain the richest semantic information and is thus used for classification [6, 10, 12, 38]. Specifically, the process for determining the cluster center is shown in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.868, + 0.892, + 0.892 + ], + "angle": 0, + "content": "\\[\nQ = W _ {Q} X, K = W _ {K} X, V = W _ {V} X, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.888, + 0.756, + 0.904 + ], + "angle": 0, + "content": "\\[\nk _ {c} = \\operatorname {P o o l} (K).\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4021" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "Where \\( W_{K} \\) is a learnable matrix. \\( k_{c} \\) is the determined cluster center. \\( X \\) is the set of input tokens." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.144, + 0.47, + 0.311 + ], + "angle": 0, + "content": "Distance Metric Suitable for ViT. Unlike the Euclidean distance calculation used in the k-means algorithm for computing the distance between tokens, during the actual computation of Self-Attention, similarity between query and key is computed through dot product. To better adapt to the characteristics of Self-Attention, we also measure the distance between tokens using a method similar to dot product. Specifically, we calculate the cosine similarity between the cluster center and each token, and then sort the tokens according to the magnitude of the computed results. The specific process is shown in Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.323, + 0.356, + 0.355 + ], + "angle": 0, + "content": "\\[\ns i m = \\frac {K \\cdot k _ {c}}{| | K | | \\cdot | | k _ {c} | |},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.353, + 0.469, + 0.374 + ], + "angle": 0, + "content": "\\[\ni d x = \\operatorname {a r g s o r t} (s i m), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.377, + 0.419, + 0.394 + ], + "angle": 0, + "content": "\\[\nQ ^ {*} = Q [ i d x ], K ^ {*} = K [ i d x ], V ^ {*} = V [ i d x ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.406, + 0.469, + 0.467 + ], + "angle": 0, + "content": "Where \\(sim\\) is the similarity matrix between \\(K\\) and \\(k_{c}\\), the \\(\\operatorname{argsort}(sim)\\) returns the indices of \\(sim\\) sorted in descending order. \\(Q^{*}, K^{*}, V^{*}\\) are \\(Q, K, V\\) rearranged according to \\(\\operatorname{argsort}(sim)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.469, + 0.61 + ], + "angle": 0, + "content": "Equally Partition Tokens based on Distance. The obtained \\( Q^{*}, K^{*} \\), and \\( V^{*} \\) from the previous step have been sorted based on their distances to the cluster center. For the design of vision backbone, we directly group them, so tokens with similar distances to the cluster center are classified into the same cluster. This allows us to directly control an equal number of tokens in each cluster. This process can be clearly illustrated in Fig. 3(c) and denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.624, + 0.384, + 0.639 + ], + "angle": 0, + "content": "\\[\nQ _ {m} = Q ^ {*} [ m \\times N: (m + 1) N ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.642, + 0.469, + 0.658 + ], + "angle": 0, + "content": "\\[\nK _ {m} = K ^ {*} [ m \\times N: (m + 1) N ], \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.661, + 0.384, + 0.677 + ], + "angle": 0, + "content": "\\[\nV _ {m} = V ^ {*} [ m \\times N: (m + 1) N ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.69, + 0.469, + 0.719 + ], + "angle": 0, + "content": "where \\(N\\) is the basic token number of each cluster for equal partition and \\(m\\) is the index of the cluster" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.721, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Based on the above steps, we have completed the clustering process that captures semantic information in the image with minimal sorting cost. Moreover, compared to k-means, we have achieved equi-partitioning of each cluster. After clustering is completed, we apply standard Self-Attention to the tokens within each cluster, thereby completing the interaction of information between tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.841, + 0.469, + 0.857 + ], + "angle": 0, + "content": "\\[\nY _ {m} = \\operatorname {A t t n} \\left(Q _ {m}, K _ {m}, V _ {m}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For the design of vision language connector, we group the tokens according to their similarity, and the tokens" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.85, + 0.107 + ], + "angle": 0, + "content": "within each group are interleaved, as shown in Eq. 5:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.12, + 0.769, + 0.136 + ], + "angle": 0, + "content": "\\[\nQ _ {n} = Q ^ {*} [ n: N: L ],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.139, + 0.891, + 0.154 + ], + "angle": 0, + "content": "\\[\nK _ {n} = K ^ {*} [ n: N: L ], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.158, + 0.769, + 0.173 + ], + "angle": 0, + "content": "\\[\nV _ {n} = V ^ {*} [ n: N: L ].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.276 + ], + "angle": 0, + "content": "in which \\( L \\) is the token's sequence length, \\( n \\) is the index of group tokens. \\( N \\) is the basic token number of each cluster. After obtaining the token groups, we perform pooling on \\( Q \\) to effectively reduce the number of tokens input to the LLM, with each group's output becoming a single token, as shown in Eq 6." + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.289, + 0.891, + 0.306 + ], + "angle": 0, + "content": "\\[\nY _ {n} = \\operatorname {A t t n} \\left(\\operatorname {P o o l} \\left(Q _ {n}\\right), K _ {n}, V _ {n}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.805, + 0.333 + ], + "angle": 0, + "content": "3.3. Difference between SEC and EViT." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.386 + ], + "angle": 0, + "content": "We use the most representative example, EViT [32], to illustrate the differences between SEC and other methods based on the similarity between the global token and other tokens." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Pruning v.s. Clustering. Most similarity-based methods, such as EViT, are pruning methods, where tokens with low similarity to the [cls] token are merged during the forward process, thereby reducing the number of tokens and decreasing computational cost. In contrast, our proposed SECViT employs a clustering-based approach, performing attention operations within each cluster." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.494, + 0.892, + 0.658 + ], + "angle": 0, + "content": "The role of the [cls] token. In methods like EViT, the [cls] token serves as a measure of importance of a token. Each token computes its similarity to the [cls] token, with higher similarity tokens deemed more important. The less important tokens are abandoned. In contrast, in SEC, the [cls] token (obtained by average pooling over all tokens) measures similarity between tokens. Each token computes its similarity score to the [cls] token; tokens with similar scores are considered to be more similar and grouped into one cluster. Attention is calculated only within the same cluster." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.675, + 0.633, + 0.691 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.701, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We first make strict comparison with hierarchical/plain baselines. Then we conduct experiments on a wide range of vision tasks for SECViT, including image classification, object detection, instance segmentation, and semantic segmentation. We also verify the role of SEC in MLLM based on LLaVA-1.5 [36]. More details, experiments, and comparison of models' efficiency can be found in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.706, + 0.832 + ], + "angle": 0, + "content": "4.1. SEC for vision models" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Strict Comparison with Baselines. We select two baselines: hierarchical backbone Swin-Transformer [38] and plain backbone DeiT [43] to make a comparison with our SEC based model. In the comparison models (SEC-Swin" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4022" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.079, + 0.462, + 0.292 + ], + "angle": 0, + "content": "
ModelParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
DeiT-S [43]224.6320479.844.540.143.0
EViT-DeiT-S (keeprate=0.9)224.0342879.8not suitnot suitnot suit
SEC-DeiT-S (num_cluster=4)224.1341280.5 (+0.7)47.7 (+3.2)42.7 (+2.6)47.5 (+4.5)
DeiT-B8617.6150281.8---
SEC-DeiT-B8614.8168282.4 (+0.6)---
Swin-T294.5172381.343.739.844.5
SEC-Swin-T294.8148283.8 (+2.5)48.3 (+4.6)43.4 (+3.6)49.3 (+4.8)
Swin-S508.8100683.045.741.147.6
SEC-Swin-S509.280485.0 (+2.0)50.2 (+4.5)44.7 (+3.6)51.3 (+3.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.295, + 0.468, + 0.322 + ], + "angle": 0, + "content": "Table 1. Comparison with Hierarchy/Plain baselines. Inference speed are measured on the A100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.327, + 0.467, + 0.437 + ], + "angle": 0, + "content": "
ModelParams (M)FLOPs (G)MethodPretrain epochAcc(%)
Swin-B [38]8815.4Supervised-83.5
ConvNeXt V2-B [50]8815.4Supervised-84.3
SEC-Swin-B8816.2Supervised-85.3
Swin-B [38]8815.4SimMIM [53]80084.0(+0.5)
ConvNeXt V2-B [50]8815.4FCMAE [50]80084.6(+0.3)
SEC-Swin-B8816.2SimMIM [53]80085.9(+0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.44, + 0.458, + 0.453 + ], + "angle": 0, + "content": "Table 2. Comparison with baselines on self-supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.468, + 0.59 + ], + "angle": 0, + "content": "and SEC-DeiT), we merely substitute the attention mechanism in the original model with our SEC based Self-Attention and without introducing any other modules. As shown in Tab. 1, we conduct experiments on image classification, object detection, insatance segmentation and semantic segmentation, the simple replacement of the attention mechanism yields significant advantages in both performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.468, + 0.651 + ], + "angle": 0, + "content": "In addition to the supervised scenario, we also train the model with SimMIM [53] in the self-supervised scenario. As shown in Tab. 2, SEC also performs exceptionally well in the self-supervised scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.468, + 0.822 + ], + "angle": 0, + "content": "Image Classification. We compare our SECViT with numerous state-of-the-art models, the results are shown in Tab.3. We adopt the training strategy proposed in DeiT [43], with the only supervision is cross entropy loss. All of our models are trained from scratch for 300 epochs with the input resolution of \\(224 \\times 224\\). SECViT consistently outperforms preceding models across all scales. Notably, SECViT-S attains a Top1-accuracy of \\(84.3\\%\\) with a mere 27M parameters and 4.6G FLOPs. The comparison of the models' efficiency can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Object Detection and Instance Segmentation. We utilize MMDetection [4] to implement Mask-RCNN [21], Cascade Mask R-CNN [2], and RetinaNet [33] to evaluate the performance of the SECViT. Tab. 5 and Tab. 4 show the" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.882, + 0.624 + ], + "angle": 0, + "content": "
CostModelParmas (M)FLOPs (G)Top1-acc (%)
tiny model ~ 2.5GPVTv2-b1 [46]132.178.7
TCFormer-light [59]143.879.4
QuadTree-B-b1 [42]142.380.0
MPViT-XS [28]112.980.9
BiFormer-T [61]132.281.4
CrossFormer-T [47]282.981.5
FAT-B2 [12]142.081.9
GC-ViT-XT [20]202.682.0
SMT-T [34]122.482.2
RMT-T [13]142.582.4
SECViT-T152.582.7
small model ~ 4.5GPS-ViT-B14 [58]215.481.7
DVT-T2T-ViT-19 [49]396.281.9
ConvNeXt-T [39]294.582.1
TCFormer [59]265.882.3
SG-Former-S [15]234.883.2
StructViT-S-8-1 [25]245.483.3
InternImage-T [48]305.083.5
MLLA-T [18]254.283.5
MaxViT-T [44]315.683.6
FAT-B3 [12]294.483.6
SMT-S [34]204.883.7
BiFormer-S [61]264.583.8
SECViT-S274.684.3
base model ~ 9.0GConvNeXt-S [39]508.783.1
NAT-S [19]517.883.7
Quadtree-B-b4 [42]6411.584.0
MOAT-1 [54]429.184.2
InternImage-S [48]508.084.2
GC-ViT-S [20]518.584.3
BiFormer-B [61]579.884.3
iFormer-B [41]489.484.6
FAT-B4 [12]529.384.8
SECViT-B579.885.2
large model ~ 18.0GCrossFormer-L [47]9216.184.0
SMT-L [34]8117.784.6
DaViT-B [9]8815.584.6
SG-Former-B [15]7815.684.7
iFormer-L [41]8714.084.8
InterImage-B [48]9716.084.9
GC-ViT-B [20]9014.885.0
RMT-L [13]9518.285.5
SECViT-L10118.285.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.626, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-art on ImageNet-1K classification." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.669, + 0.892, + 0.715 + ], + "angle": 0, + "content": "results of SECViT with different detection frameworks. The results show that SECViT performs better than its counterparts in all comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Semantic Segmentation. We utilize Semantic FPN [26] and UperNet [52] to validate our SECViT's performance, implementing these frameworks via MMSegmentation [7]. The results of semantic segmentation can be found in the Tab. 6. All the FLOPs are measured with the input resolution of \\(512 \\times 2048\\). SECViT achieves the best performance in all settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.661, + 0.862 + ], + "angle": 0, + "content": "4.2. SEC for MLLM" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "SEC can greatly facilitate the design of vision language connectors in MLLMs. First, we conduct a rigorous compari" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4023" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.078, + 0.85, + 0.332 + ], + "angle": 0, + "content": "
BackboneParams (M)FLOPs (G)Mask R-CNN 1×Params (M)FLOPs (G)RetinaNet 1×
APbAPb50APb75APmAPm50APm75APbAPb50APb75APsAPbAPL
PVTv2-B1 [46]3324341.854.345.938.861.241.62322541.261.943.925.444.554.3
FAT-B2 [12]3321545.267.949.041.364.644.02319644.065.247.227.547.758.8
RMT-T [13]3321847.168.851.742.665.845.92319945.166.248.128.848.961.1
SECViT-T3422147.869.552.543.066.746.32420245.866.849.229.149.860.9
MPViT-S [28]4326846.468.651.242.465.645.73224845.757.348.828.749.759.2
MLLA-T [18]4425546.869.551.542.166.445.0--------
STViT-S [22]4425247.670.052.343.166.846.5--------
RMT-S [13]4626249.070.853.943.967.847.43624447.869.151.832.151.863.5
SECViT-S4526249.970.954.744.668.347.73524048.469.452.031.353.363.8
ScalableViT-B [56]9534946.868.751.542.565.845.98533045.867.349.229.949.561.0
InternImage-S [48]6934047.869.852.843.367.146.7--------
MLLA-S [18]6331949.271.553.944.268.547.2--------
STViT-B [22]7035949.771.754.744.868.948.7--------
SECViT-B7637151.572.956.745.469.948.76334949.370.352.932.053.864.8
Focal-B [55]11053347.870.252.543.267.346.510151446.368.049.831.750.460.8
CSwin-B [10]9752648.770.453.943.967.847.3--------
InternImage-B [48]11550148.870.954.044.067.847.4--------
MLLA-B [18]11550250.572.055.445.069.348.6--------
SECViT-L11955052.073.557.346.370.649.810552750.271.453.933.254.566.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.267, + 0.334, + 0.7, + 0.347 + ], + "angle": 0, + "content": "Table 4. Comparison to other backbones using \"1×\" schedule on COCO." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.351, + 0.459, + 0.677 + ], + "angle": 0, + "content": "
BackboneParams (M)FLOPs (G)\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Mask R-CNN 3×+MS
GC-ViT-T [20]4829147.970.152.843.267.046.7
MLLA-T [18]4425548.871.053.643.868.046.8
SMT-S [34]4026549.070.153.443.467.346.7
InternImage-T [48]4927049.170.454.143.767.347.3
RMT-S [13]4626250.771.955.644.969.148.4
SECViT-S4526251.672.555.945.669.948.8
NAT-S [19]7033048.469.853.243.266.946.4
InternImage-S [48]6934049.771.154.544.568.547.8
SMT-B [34]5232849.871.054.444.068.047.3
MLLA-S [18]6331950.571.855.244.969.148.2
RMT-B [13]7337352.272.957.046.170.449.9
SECViT-B7537152.873.657.746.470.849.9
Cascade Mask R-CNN 3×+MS
GC-ViT-T [20]8577051.670.456.144.667.848.3
SMT-S [34]7874451.970.556.344.767.848.6
UniFormer-S [30]7974752.171.156.645.268.348.9
RMT-S [13]8374153.272.057.846.169.849.8
SECViT-S8374154.172.858.647.070.351.0
NAT-S [19]10880951.970.456.244.968.248.6
GC-ViT-S [20]10886652.471.057.145.468.549.3
CSWin-S [10]9282053.772.258.446.469.650.6
UniFormer-B [30]10787853.872.858.546.469.950.4
RMT-B [13]11185254.572.859.047.270.551.4
SECViT-B11484955.474.159.947.871.751.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.679, + 0.468, + 0.706 + ], + "angle": 0, + "content": "Table 5. Comparison with other backbones using \\(3 \\times +\\mathrm{MS}\\) schedule on COCO." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.716, + 0.468, + 0.762 + ], + "angle": 0, + "content": "son between SEC and various baseline vision language connectors based on LLaVA-1.5. Then, we compare LLaVA-1.5+SEC with several popular contemporary MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Strict Comparison with Baselines. In Tab. 7, we strictly compare various commonly used vision language connectors, including MLP, Resampler [1], Pooling, and EViT [32], which has achieved success in the design of ViT. Among these, MLP is the original design in LLaVA-1.5 [36], capable of achieving good results. However, it incurs significant computational cost due to the excessive vision tokens. To address this issue, some connectors at" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.361, + 0.869, + 0.666 + ], + "angle": 0, + "content": "
ModelSemantic FPN 80K Params FLOPsUpernet 160K Params FLOPs
(M)(G)(%)(M)(G)(%)
PVTv2-B1 [46]1813642.5---
VAN-B1 [17]1814042.9---
RMT-T [13]1713646.4---
SECViT-T1813647.24489448.8
StructViT-S [25]2627146.9---
MogaNet-S [31]2918947.75594649.2
SMT-S [34]---5093549.2
SGFormer-S [15]2520549.052.598949.9
RMT-S [13]3018049.45693749.8
SECViT-S3018049.65693650.6
MogaNet-B [31]---74105050.1
InterImage-S [48]---80101750.2
StructViT-B [25]5452948.5---
RMT-B [13]5729450.483105152.0
SECViT-B6029150.786104852.2
MogaNet-L [31]---113117650.9
MLLA-B [18]---128118351.9
SGFormer-B [15]8147550.6109130452.0
RMT-L [13]9848251.4125124152.8
SECViT-L10347552.2131125653.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.668, + 0.87, + 0.681 + ], + "angle": 0, + "content": "Table 6. Comparison with the state-of-the-art on ADE20K." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.79 + ], + "angle": 0, + "content": "tempt to use fewer vision tokens to accelerate LLaVA-1.5. Nonetheless, these adjustments inevitably lead to performance degradation. The results in Tab. 7 show that using SEC can effectively accelerate the inference of LLaVA-1.5 without causing performance degradation, and can even improve the performance of LLaVA-1.5 to a certain extent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with Popular MLLMs. In Tab. 8 and Tab. 9, we compare LLaVA-1.5 equipped with SEC as a vision-language connector with other MLLMs. It is evident that SEC not only enhances the performance of MLLMs across various benchmarks but also significantly improves the efficiency of the models. This fully demonstrates the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4024" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.078, + 0.844, + 0.276 + ], + "angle": 0, + "content": "
ModelConnectorV-T NumTimeSpeedTextVQAGQAVQAv2POPEMME
LLaVA-1.5MLP576+1194s1.0×58.262.078.586.11510.7
LLaVA-1.5+ResamplerResampler288+1126s1.5×52.156.876.083.11393.2
LLaVA-1.5+EViTMLP+EViT288+1126s1.5×54.660.077.984.31483.2
LLaVA-1.5+SECMLP+SEC288+1126s1.5×60.163.578.987.71510.7
LLaVA-1.5+ResamplerResampler256+1116s1.7×51.656.075.282.71387.2
LLaVA-1.5+PoolMLP+Pool256+1116s1.7×52.457.676.483.31415.5
LLaVA-1.5+EViTMLP+EViT256+1116s1.7×52.859.677.183.71443.7
LLaVA-1.5+SECMLP+SEC256+1116s1.7×59.663.278.687.11505.2
LLaVA-1.5+ResamplerResampler192+1102s1.9×50.155.274.382.71337.6
LLaVA-1.5+EViTMLP+EViT192+1102s1.9×51.658.676.383.11427.6
LLaVA-1.5+SECMLP+SEC192+1102s1.9×57.762.778.486.71500.1
LLaVA-1.5+ResamplerResampler144+194s2.1×47.654.672.081.91293.7
LLaVA-1.5+PoolMLP+Pool144+194s2.1×50.056.273.681.91310.7
LLaVA-1.5+EViTMLP+EViT144+194s2.1×51.258.076.083.11393.6
LLaVA-1.5+SECMLP+SEC144+194s2.1×56.862.078.086.11487.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Table 7. Comparison of different vision language connectors on LLaVA-1.5. \"V-T Num\" denotes the quantity of visual tokens. The computation expense is impacted by V-T Num, with larger values resulting in higher costs. \"Speed\" refers to the comparative inference velocity relative to LLaVA-1.5. \"Time\" is the average inference time. Inference speed are measured on the A100." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.323, + 0.855, + 0.495 + ], + "angle": 0, + "content": "
ModelLLMConnectorV-T NumResTextVQAGQAVQAv2VisWiz\\( SQA_{img} \\)Speed (↑)
7B LLM
Shikra [5]Vicuna-7BMLP257224--77.4---
IDEFICS-9B [27]LLaMA-7BCross Attn257224-38.450.935.5--
Qwen-VL [1]Qwen-7BResampler256448-59.378.835.267.1-
Qwen-VL-Chat [1]Qwen-7BResampler256448-57.578.238.968.2-
LLaVA-1.5 [35]Vicuna-7BMLP57733658.262.078.550.066.81.0×
LLaVA-1.5+SEC (ours)Vicuna-7BMLP+SEC25733659.663.278.952.869.61.7×
13B LLM
InstructBLIP [8]Vicuna-13BQ-Former32224-49.5-33.463.1-
BLIP-2 [29]Vicuna-13BQ-Former32224-41.041.019.561.0-
LLaVA-1.5 [35]Vicuna-13BMLP57733661.263.380.053.671.61.0×
LLaVA1.5+SEC (ours)Vicuna-13BMLP+SEC25733662.364.380.054.772.01.8×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.365, + 0.495, + 0.604, + 0.508 + ], + "angle": 0, + "content": "Table 8. Results on General VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.525, + 0.432, + 0.54 + ], + "angle": 0, + "content": "effectiveness of SEC in extracting visual information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.231, + 0.567 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.575, + 0.47, + 0.606 + ], + "angle": 0, + "content": "In this section, we present some of the ablation study results for SEC, and more results can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.791 + ], + "angle": 0, + "content": "Number of Vision Tokens in Each Clusters. The number of vision tokens has a significant impact on the performance and speed of the model. We thoroughly investigate the effect of the number of vision tokens on SECViT. As shown in Tab. 10, the number of vision tokens in each cluster greatly influences the model's performance. Specifically, in downstream dense prediction tasks, having too few tokens in each cluster leads to substantial performance degradation. When the number of tokens in each cluster is too large, the model's performance does not see a significant improvement, but its speed decreases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.273, + 0.817 + ], + "angle": 0, + "content": "4.4. Visualization of SEC" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To further understand the working mechanism of SEC, we visualize some clustering results for SECViT. As shown in Fig. 4, the left side presents the clustering results of vision tokens at different stages of the model. From the clustering results, we analyze that in the shallow layers, the model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.617 + ], + "angle": 0, + "content": "distinguishes fine-grained features well, while in the deeper layers, it captures global semantic features effectively. The right side shows the Grad-CAM diagrams at different stages of the model, from which we can draw similar conclusions to the clustering results. More visualization results can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.893, + 0.801 + ], + "angle": 0, + "content": "Number of Vision Tokens Outputs by SEC. MLLM is quite sensitive to the number of vision tokens. We conduct a detailed exploration based on LLaVA-1.5 regarding the number of vision tokens output by SEC, as shown in Tab. 11. The first row represents the speed and performance of the original LLaVA-1.5 without using SEC. Compared to LLaVA-1.5, employing SEC effectively reduces the number of vision tokens and improves training efficiency. As the number of vision tokens decreases, the model's performance shows a slight decline, but its efficiency is further enhanced." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.62, + 0.831 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We propose a simple and straightforward clustering method for vision tokens—Semantic Equitable Clustering (SEC). This method assigns each token to a cluster by calculating the similarity between each token and a global token, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4025" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.868, + 0.285 + ], + "angle": 0, + "content": "
ModelLLMConnectorV-T NumResPOPEMMBMM-VetSpeed (↑)
7B LLM
MiniGPT-4 [60]Vicuna-7BResampler3222472.224.322.1-
mPLUG-Ow12 [57]LLaMA2-7BResampler32224-49.4--
LLaMA-AdapterV2 [14]LLaMA2-7BLLaMA-Adapter257224-41.031.4-
Shikra [5]Vicuna-7BMLP257224-58.8--
Qwen-VL [1]Qwen-7BResampler256448-38.2--
Qwen-VL-Chat [1]Qwen-7BResampler256448-60.6--
LLaVA-1.5 [35]Vicuna-7BMLP57733686.164.331.11.0×
LLaVA1.5+SEC (ours)Vicuna-7BMLP+SEC14533686.168.431.72.1×
13B LLM
MiniGPT-4 [60]Vicuna-13BResampler32224--24.4-
BLIP-2 [29]Vicuna-13BQ-Former3222485.3-22.4-
LLaVA-1.5 [35]Vicuna-13BMLP57733686.267.736.11.0×
LLaVA-1.5+SEC (ours)Vicuna-13BMLP+SEC14533686.469.237.32.2×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.325, + 0.287, + 0.644, + 0.3 + ], + "angle": 0, + "content": "Table 9. Results on benchmark designed for MLLMs." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.307, + 0.449, + 0.457 + ], + "angle": 0, + "content": "
V-T numParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
98152.5200482.747.843.047.2
196153.1172283.0 (+0.3)48.2 (+0.4)43.4 (+0.4)47.5 (+0.3)
64152.5194682.7 (+0.0)47.8 (+0.0)42.8 (-0.2)46.9 (-0.3)
49152.4210282.6 (-0.1)47.5 (-0.3)42.7 (-0.3)47.7 (-0.5)
24152.3218682.0 (-0.7)45.9 (-1.9)40.6 (-2.4)44.6 (-2.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.458, + 0.47, + 0.5 + ], + "angle": 0, + "content": "Table 10. Effect of the number of vision tokens in each cluster. \"V-T num\" means the number of vision tokens in each cluster. The experiments are conducted based on SECViT-T." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.504, + 0.469, + 0.59 + ], + "angle": 0, + "content": "
V-T numTime SpeedTextVQAGQAVQAv2POPEMM-Vet
576+121h1.0×58.262.078.586.131.1
288+114h1.5×60.1(+1.9)63.5(+1.5)78.9(+0.4)87.7(+1.6)33.2(+2.1)
256+113h1.6×59.6(+1.4)63.2(+0.3)78.6(+0.1)87.1(+1.0)32.7(+1.6)
192+111h1.9×57.7(-0.5)62.7(+0.7)78.4(-0.1)86.7(+0.6)32.1(+1.0)
144+110h2.1×56.8(-1.4)62.0(+0.0)78.0(-0.5)86.1(+0.0)31.7(+0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.592, + 0.47, + 0.634 + ], + "angle": 0, + "content": "Table 11. Effect of the number of vision tokens outputs by SEC. \"V-T num\" means the number of vision tokens output by SEC. The experiments are conducted based on LLaVA-1.5 [36]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.47, + 0.832 + ], + "angle": 0, + "content": "completes the whole clustering process in only one step. Our clustering method takes into account the semantic information contained in the tokens, and ensures an equal number of tokens in each cluster, facilitating efficient parallel processing on modern GPUs. Based on Semantic Equitable Clustering, we designed SECViT, a versatile vision backbone that achieves impressive results across various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Besides, SEC can also be conveniently applied to multimodal large language models (MLLM) to serve as a vision language connector and benefits the model's efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.265, + 0.863 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This work is partially funded by Beijing Natural Science Foundation (4252054), Youth Innovation Promotion" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.306, + 0.89, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.797, + 0.794, + 0.811 + ], + "angle": 0, + "content": "Figure 4. Visualization for SEC." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.827, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Association CAS(Grant No.2022132), Beijing Nova Program(20230484276), and CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024005)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4026" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[3] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.468, + 0.343 + ], + "angle": 0, + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.468, + 0.414 + ], + "angle": 0, + "content": "[7] MMSegmentation Contributors. Mmsegmentation, an open source semantic segmentation toolbox, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.468, + 0.485 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.468, + 0.514 + ], + "angle": 0, + "content": "[9] Mingyu Ding, Bin Xiao, Noel Codella, et al. Davit: Dual attention vision transformers. In ECCV, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.468, + 0.557 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, et al. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.468, + 0.642 + ], + "angle": 0, + "content": "[12] Qihang Fan, Huaibo Huang, Xiaogiang Zhou, and Ran He. Lightweight vision transformer with bidirectional interaction. In NeurIPS, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[13] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[15] SG-Former: Self guided Transformer with Evolving Token Reallocation. Sucheng ren, xingyi yang, songhua liu, xinchao wang. In ICCV, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[16] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[18] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[20] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In ICML, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.21, + 0.892, + 0.236 + ], + "angle": 0, + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask r-cnn. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[22] Huaibo Huang, Xiaogiang Zhou, Jie Cao, Ran He, and Tieniu Tan. Vision transformer with super token sampling. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[23] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver: General perception with iterative attention, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.327, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[24] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[25] Manjin Kim, Paul Hongsuck Seo, Cordelia Schmid, and Minsu Cho. Learning correlation structures for vision transformers. In CVPR, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.471, + 0.892, + 0.54 + ], + "angle": 0, + "content": "[27] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[28] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.643, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[30] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[31] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z. Li. Moganet: Multi-order gated aggregation network. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[32] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, and Kaiming He and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[34] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4027" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[37] Kai Liu, Tianyi Wu, Cong Liu, and Guodong Guo. Dynamic group transformer: A general vision transformer backbone with dynamic group attention. In IJCAI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.272 + ], + "angle": 0, + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, et al. A convnet for the 2020s. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.469, + 0.356 + ], + "angle": 0, + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[41] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng YAN. Inception transformer. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.426 + ], + "angle": 0, + "content": "[42] Shitao Tang, Jiahui Zhang, Siyu Zhu, et al. Quadtree attention for vision transformers. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, et al. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[44] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[46] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):1-10, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[47] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. In ICLR, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[48] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[49] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic vision transformers with adaptive sequence length. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[50] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[51] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[52] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[53] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[54] Chenglin Yang, Siyuan Qiao, Qihang Yu, et al. Moat: Alternating mobile convolution and attention brings strong vision models. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[55] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[56] Rui Yang, Hailong Ma, Jie Wu, Yansong Tang, Xuefeng Xiao, Min Zheng, and Xiu Li. Scalablevit: Rethinking the context-oriented generalization of vision transformer. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[58] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[59] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[60] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[61] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4028" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_origin.pdf b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9b4d6be90215f4cfc9b2f74e0970d0583b3038c7 --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/755b1314-9e24-437c-8f51-19b61bb62095_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6192dd5e0b2121b2c6c78312e1d16fe963b8198ea17948a55133c7eada8e7c8f +size 2543314 diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/full.md b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..be235510a0d63360abb7fef1b6af9e032dded400 --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/full.md @@ -0,0 +1,329 @@ +# Semantic Equitable Clustering: A Simple and Effective Strategy for Clustering Vision Tokens + +Qihang Fan $^{1,2}$ , Huaibo Huang $^{1*}$ , Mingrui Chen $^{1,2}$ , Ran He $^{1,2}$ + +$^{1}$ MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China + $^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China +fanqihang.159@gmail.com, huaibo.huang@cripac.ia.ac.cn, pharmier@hust.edu.cn, rhe@nlpr.ia.ac.cn + +# Abstract + +The Vision Transformer (ViT) has gained prominence for its superior relational modeling prowess. However, its global attention mechanism's quadratic complexity poses substantial computational burdens. A common remedy spatially groups tokens for self-attention, reducing computational requirements. Nonetheless, this strategy neglects semantic information in tokens, possibly scattering semantically-linked tokens across distinct groups, thus compromising the efficacy of self-attention intended for modeling inter-token dependencies. Motivated by these insights, we introduce a fast and balanced clustering method, named Semantic Equitable Clustering (SEC). SEC clusters tokens based on their global semantic relevance in an efficient, straightforward manner. In contrast to traditional clustering methods requiring multiple iterations, our method achieves token clustering in a single pass. Additionally, SEC regulates the number of tokens per cluster, ensuring a balanced distribution for effective parallel processing on current computational platforms without necessitating further optimization. Capitalizing on SEC, we propose a versatile vision backbone, SECViT. Comprehensive experiments in image classification, object detection, instance segmentation, and semantic segmentation validate the effectiveness of SECViT. Moreover, SEC can be conveniently and swiftly applied to multimodal large language models (MLLM), such as LLaVA, to serve as a vision language connector, effectively accelerating the model's efficiency while maintaining unchanged or better performance. + +# 1. Introduction + +Since its inception, the Vision Transformer (ViT)[11] has drawn considerable interest from the research community due to its robust modeling prowess. However, the quadratic + +![](images/bfedae319fa83a0de59aba8a9fed4488ae3931ea412df57b82603b7d4dda41be.jpg) +(a) Window Partition +X Semantic Information $\checkmark$ Cluster Efficiency $\checkmark$ Equi-partition +Figure 1. Comparison among Window Partition, Dynamic Group by k-means, and Semantic Equitable Clustering. Our Semantic Equitable Clustering incorporates image semantics while maintaining efficient clustering, eliminating the need for iterative processes such as in k-means. Furthermore, it enables equipartitioning of tokens, promoting efficient GPU processing without necessitating additional CUDA optimization. + +![](images/0d5a36a6d9ba18eca47eadf175213d7a8aec7ab9e2fefba6be9152e6d3af78d6.jpg) +(b) k-means +$\checkmark$ Semantic Information + $\times$ Cluster Efficiency + $\times$ Equi-partition +(c)Semantic Equitable Clustering + +![](images/6ca7c9ba783fe071e548b9dfb0324f7fa73d27c460c85350ab380b43ec772af5.jpg) +$\checkmark$ Semantic Information $\checkmark$ Cluster Efficiency $\checkmark$ Equi-partition + +complexity of Self-Attention leads to significant computational overhead, thus constraining the practicality of ViT. A variety of strategies have been devised to alleviate this computational load, the most prevalent of which involves token grouping, thereby constraining the attention span of each token[10, 38, 44, 47]. + +Specifically, the Swin-Transformer [38] partitions tokens into multiple small windows, restricting token attention within each window. The CSWin-Transformer [10] adopts a cross-shaped grouping, endowing each token with a global receptive field. MaxViT [44] amalgamates window and grid attention, facilitating intra-window tokens to attend to their counterparts in other windows. However, these methods, solely reliant on spatial positioning, neglect token semantics, potentially restricting the self-attention's capacity to model semantic dependencies. To mitigate this, DGT [37] employs k-means clustering for query grouping, considering the semantic information of tokens for enhanced feature learning. Nonetheless, the iterative nature of k-means clustering and the potential for uneven token counts per cluster can impact the efficiency of parallel attention operations. + +![](images/82c9bce75d7bf1a1970ecdfcfbe91b46fd54fe2042e2030049cec92004ce354c.jpg) +Figure 2. Left: Top-1 accuracy v.s. FLOPs on ImageNet-1K of recent SOTA models. Right: Comparison among different vision language connectors on LLaVA-1.5 + +Given these considerations, an optimal token partitioning scheme should efficiently segregate tokens, incorporate semantic information, and efficiently utilize computational resources (e.g., GPU). In response, we introduce a simple, fast and equitable clustering approach named Semantic Equitable Clustering (SEC). SEC segments tokens based on their relevance to global semantic information. Specifically, we employ global pooling to generate a global token encapsulating global semantic information. The similarity between this global token and all other tokens is then computed, reflecting global semantic relevance. Upon obtaining the similarity matrix, tokens (excluding the global token) are sorted by similarity scores, and the tokens with similar scores are grouped into clusters, ensuring uniform token distribution across clusters. As depicted in Fig. 1, SEC comprehensively considers token semantics and completes the clustering process in a single iteration, unlike the multi-iteration k-means. The resulting clusters, containing an equal number of tokens, can be processed in parallel by the GPU efficiently. + +Building upon Semantic Equitable Clustering (SEC), we introduce the Semantic Equitable Clustering Vision Transformer (SECViT), a versatile vision backbone that is adaptable to a wide spectrum of downstream tasks. As demonstrated in Fig. 2, SECViT exhibits significant performance improvements compared to previous state-of-the-art (SOTA) models. Impressively, SECViT attains an accuracy of $84.3\%$ utilizing merely 4.6GFLOPS, without the need for additional training data or supervision. This superior performance is maintained across different model scales. Furthermore, SECViT proves its proficiency in downstream tasks, including but not limited to, object detection, instance segmentation, and semantic segmentation. + +Beyond vision tasks, we also apply SEC to multimodal large language models (MLLM) such as LLaVA-1.5 [36] to + +serve as an efficient vision language connector. Specifically, we use SEC to cluster the vision tokens, and then merge all the tokens at corresponding positions within each cluster into a single token. Experiments demonstrate that this approach significantly enhances the efficiency of LLaVA-1.5 while improving the model's performance. + +# 2. Related Works + +Vision Transformer. The Vision Transformer (ViT) [11] is considered a powerful visual architecture. Many works have improved the Vision Transformer, including enhancing its training efficiency and reducing its computational cost [10, 24, 38, 43, 61]. DeiT [43] uses distillation loss and incorporates extensive data augmentation methods into the ViT training process. Hierarchical structures represented by PVT [16, 41, 45, 46, 51] reduce the number of tokens in global attention by downsampling the keys and values (KV), thereby low the computational cost. In addition to them, some methods directly prune tokens based on their importance, retaining important tokens [32, 40]. This reduces the number of tokens and subsequently lowers the computational cost of the model. Another highly representative approach is to group all tokens such that each token can only attend to tokens within its own group [9, 10, 37, 38, 61]. This method also significantly reduces the computational cost of self-attention. + +Grouping-Based Vision Transformer. Most grouping-based attention mechanisms perform grouping based on spatial structure [9, 10, 37, 38, 44]. Specifically, the SwinTransformer [38] divides all tokens into equally sized windows based on their spatial positions, where each token can only attend to tokens within its own window. This significantly reduces the model's computational cost. In addition to dividing tokens into small windows along the spatial dimension, DaViT [9] also splits channels into multiple groups along the channel dimension. Unlike the above methods that only consider positional information for grouping, DGT [37] takes semantic information into account by using k-means clustering to group the queries. + +Vision Language Connector. The vision language connector is a critical component in MLLMs [3, 23, 36]. It aligns vision tokens with language tokens. Typical vision language connectors include MLP [36], Resampler [1], C-Abstractor [3], and others. Although MLP performs well, it introduces a significant number of vision tokens, which hampers the model's efficiency. On the other hand, connectors like Resampler improve the model's efficiency, but at the cost of reduced performance. Unlike these methods, our proposed SEC consider the semantic information of each token and significantly enhances the model's efficiency while maintaining its performance. + +![](images/5ac3a4564f17d87dc7600153d4694641f9a1af6bd66688dad7304a0f57e71375.jpg) + +![](images/a61e8739d6a4850d2400576335e0fa1a8e01d661b4868da9f8fff3f2bc62a9de.jpg) +Figure 3. (a) Illustration of SECViT (b) Applying SEC to vision language connector. (c) Illustration of Semantic Equitable Clustering for ViT and vision language connector. + +# 3. Method + +# 3.1. Overall Architecture + +The overall architecture of SECViT is shown in Fig. 3(a). SECViT consists of four stages with downsampling factors of $\frac{1}{4}$ , $\frac{1}{8}$ , $\frac{1}{16}$ , and $\frac{1}{32}$ , respectively. This structural design facilitates downstream tasks, such as object detection, in constructing feature pyramids. A SECViT block is composed of three modules. For each block, the input tensor $X_{in} \in \mathbb{R}^{C \times H \times W}$ is fed into the CPE to introduce the positional information. Then, The Self-Attention based on the Semantic Equitable Clustering (SEC) is employed to serve as the token mixer. The final FFN is utilized to integrate channel-wise information of tokens. + +Beyond the design of the backbone, we also utilize SEC in the design of the vision language connector in MLLM [36]. For the vision tokens output by ViT, we use SEC to cluster the tokens. For each position corresponding to a cluster, we use attentive pooling to merge them into a single token, thereby reducing the number of vision tokens. The process is shown in Fig. 3(b). + +# 3.2. Semantic Equitable Clustering + +As previously mentioned, the design objectives of Semantic Equitable Clustering are threefold: 1) Fully consider the semantic information contained in different tokens during clustering. 2) Unlike k-means and other clustering methods that require multiple iterations, Semantic Equitable Cluster- + +ing can complete clustering in a single step. 3) Ensure an equal number of tokens in each cluster to facilitate parallel processing on GPUs. In the following paragraphs, we will describe in detail how our Semantic Equitable Clustering achieves these three objectives. And the whole process is illustrated in the Fig. 3(c). + +Single Clustering Center Related to Semantics. k-means is relatively complex for two reasons. First, it has multiple cluster centers, and each token needs to calculate its distance to each cluster center to determine its cluster membership. Second, the determination of each cluster center in k-means is not precise and requires multiple iterations to accurately establish the cluster centers. + +To address these two issues, we first discard the use of multiple cluster centers and instead calculate the distance between each token and a single center. Based on each token's distance to this center, we divide the tokens into different intervals. Then, to ensure that our chosen center contains the most comprehensive semantic information, we directly use the result of average pooling of all tokens as the center token. This is because, in most vision foundation models, the output of the average pool is assumed to contain the richest semantic information and is thus used for classification [6, 10, 12, 38]. Specifically, the process for determining the cluster center is shown in Eq. 1: + +$$ +Q = W _ {Q} X, K = W _ {K} X, V = W _ {V} X, \tag {1} +$$ + +$$ +k _ {c} = \operatorname {P o o l} (K). +$$ + +Where $W_{K}$ is a learnable matrix. $k_{c}$ is the determined cluster center. $X$ is the set of input tokens. + +Distance Metric Suitable for ViT. Unlike the Euclidean distance calculation used in the k-means algorithm for computing the distance between tokens, during the actual computation of Self-Attention, similarity between query and key is computed through dot product. To better adapt to the characteristics of Self-Attention, we also measure the distance between tokens using a method similar to dot product. Specifically, we calculate the cosine similarity between the cluster center and each token, and then sort the tokens according to the magnitude of the computed results. The specific process is shown in Eq. 2: + +$$ +s i m = \frac {K \cdot k _ {c}}{| | K | | \cdot | | k _ {c} | |}, +$$ + +$$ +i d x = \operatorname {a r g s o r t} (s i m), \tag {2} +$$ + +$$ +Q ^ {*} = Q [ i d x ], K ^ {*} = K [ i d x ], V ^ {*} = V [ i d x ]. +$$ + +Where $sim$ is the similarity matrix between $K$ and $k_{c}$ , the $\operatorname{argsort}(sim)$ returns the indices of $sim$ sorted in descending order. $Q^{*}, K^{*}, V^{*}$ are $Q, K, V$ rearranged according to $\operatorname{argsort}(sim)$ . + +Equally Partition Tokens based on Distance. The obtained $Q^{*}, K^{*}$ , and $V^{*}$ from the previous step have been sorted based on their distances to the cluster center. For the design of vision backbone, we directly group them, so tokens with similar distances to the cluster center are classified into the same cluster. This allows us to directly control an equal number of tokens in each cluster. This process can be clearly illustrated in Fig. 3(c) and denoted as follows: + +$$ +Q _ {m} = Q ^ {*} [ m \times N: (m + 1) N ], +$$ + +$$ +K _ {m} = K ^ {*} [ m \times N: (m + 1) N ], \tag {3} +$$ + +$$ +V _ {m} = V ^ {*} [ m \times N: (m + 1) N ]. +$$ + +where $N$ is the basic token number of each cluster for equal partition and $m$ is the index of the cluster + +Based on the above steps, we have completed the clustering process that captures semantic information in the image with minimal sorting cost. Moreover, compared to k-means, we have achieved equi-partitioning of each cluster. After clustering is completed, we apply standard Self-Attention to the tokens within each cluster, thereby completing the interaction of information between tokens: + +$$ +Y _ {m} = \operatorname {A t t n} \left(Q _ {m}, K _ {m}, V _ {m}\right). \tag {4} +$$ + +For the design of vision language connector, we group the tokens according to their similarity, and the tokens + +within each group are interleaved, as shown in Eq. 5: + +$$ +Q _ {n} = Q ^ {*} [ n: N: L ], +$$ + +$$ +K _ {n} = K ^ {*} [ n: N: L ], \tag {5} +$$ + +$$ +V _ {n} = V ^ {*} [ n: N: L ]. +$$ + +in which $L$ is the token's sequence length, $n$ is the index of group tokens. $N$ is the basic token number of each cluster. After obtaining the token groups, we perform pooling on $Q$ to effectively reduce the number of tokens input to the LLM, with each group's output becoming a single token, as shown in Eq 6. + +$$ +Y _ {n} = \operatorname {A t t n} \left(\operatorname {P o o l} \left(Q _ {n}\right), K _ {n}, V _ {n}\right). \tag {6} +$$ + +# 3.3. Difference between SEC and EViT. + +We use the most representative example, EViT [32], to illustrate the differences between SEC and other methods based on the similarity between the global token and other tokens. + +Pruning v.s. Clustering. Most similarity-based methods, such as EViT, are pruning methods, where tokens with low similarity to the [cls] token are merged during the forward process, thereby reducing the number of tokens and decreasing computational cost. In contrast, our proposed SECViT employs a clustering-based approach, performing attention operations within each cluster. + +The role of the [cls] token. In methods like EViT, the [cls] token serves as a measure of importance of a token. Each token computes its similarity to the [cls] token, with higher similarity tokens deemed more important. The less important tokens are abandoned. In contrast, in SEC, the [cls] token (obtained by average pooling over all tokens) measures similarity between tokens. Each token computes its similarity score to the [cls] token; tokens with similar scores are considered to be more similar and grouped into one cluster. Attention is calculated only within the same cluster. + +# 4. Experiments + +We first make strict comparison with hierarchical/plain baselines. Then we conduct experiments on a wide range of vision tasks for SECViT, including image classification, object detection, instance segmentation, and semantic segmentation. We also verify the role of SEC in MLLM based on LLaVA-1.5 [36]. More details, experiments, and comparison of models' efficiency can be found in the Appendix. + +# 4.1. SEC for vision models + +Strict Comparison with Baselines. We select two baselines: hierarchical backbone Swin-Transformer [38] and plain backbone DeiT [43] to make a comparison with our SEC based model. In the comparison models (SEC-Swin + +
ModelParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
DeiT-S [43]224.6320479.844.540.143.0
EViT-DeiT-S (keeprate=0.9)224.0342879.8not suitnot suitnot suit
SEC-DeiT-S (num_cluster=4)224.1341280.5 (+0.7)47.7 (+3.2)42.7 (+2.6)47.5 (+4.5)
DeiT-B8617.6150281.8---
SEC-DeiT-B8614.8168282.4 (+0.6)---
Swin-T294.5172381.343.739.844.5
SEC-Swin-T294.8148283.8 (+2.5)48.3 (+4.6)43.4 (+3.6)49.3 (+4.8)
Swin-S508.8100683.045.741.147.6
SEC-Swin-S509.280485.0 (+2.0)50.2 (+4.5)44.7 (+3.6)51.3 (+3.7)
+ +Table 1. Comparison with Hierarchy/Plain baselines. Inference speed are measured on the A100 GPU. + +
ModelParams (M)FLOPs (G)MethodPretrain epochAcc(%)
Swin-B [38]8815.4Supervised-83.5
ConvNeXt V2-B [50]8815.4Supervised-84.3
SEC-Swin-B8816.2Supervised-85.3
Swin-B [38]8815.4SimMIM [53]80084.0(+0.5)
ConvNeXt V2-B [50]8815.4FCMAE [50]80084.6(+0.3)
SEC-Swin-B8816.2SimMIM [53]80085.9(+0.6)
+ +and SEC-DeiT), we merely substitute the attention mechanism in the original model with our SEC based Self-Attention and without introducing any other modules. As shown in Tab. 1, we conduct experiments on image classification, object detection, insatance segmentation and semantic segmentation, the simple replacement of the attention mechanism yields significant advantages in both performance and efficiency. + +In addition to the supervised scenario, we also train the model with SimMIM [53] in the self-supervised scenario. As shown in Tab. 2, SEC also performs exceptionally well in the self-supervised scenario. + +Image Classification. We compare our SECViT with numerous state-of-the-art models, the results are shown in Tab.3. We adopt the training strategy proposed in DeiT [43], with the only supervision is cross entropy loss. All of our models are trained from scratch for 300 epochs with the input resolution of $224 \times 224$ . SECViT consistently outperforms preceding models across all scales. Notably, SECViT-S attains a Top1-accuracy of $84.3\%$ with a mere 27M parameters and 4.6G FLOPs. The comparison of the models' efficiency can be found in Appendix. + +Object Detection and Instance Segmentation. We utilize MMDetection [4] to implement Mask-RCNN [21], Cascade Mask R-CNN [2], and RetinaNet [33] to evaluate the performance of the SECViT. Tab. 5 and Tab. 4 show the + +Table 2. Comparison with baselines on self-supervised setting. + +
CostModelParmas (M)FLOPs (G)Top1-acc (%)
tiny model ~ 2.5GPVTv2-b1 [46]132.178.7
TCFormer-light [59]143.879.4
QuadTree-B-b1 [42]142.380.0
MPViT-XS [28]112.980.9
BiFormer-T [61]132.281.4
CrossFormer-T [47]282.981.5
FAT-B2 [12]142.081.9
GC-ViT-XT [20]202.682.0
SMT-T [34]122.482.2
RMT-T [13]142.582.4
SECViT-T152.582.7
small model ~ 4.5GPS-ViT-B14 [58]215.481.7
DVT-T2T-ViT-19 [49]396.281.9
ConvNeXt-T [39]294.582.1
TCFormer [59]265.882.3
SG-Former-S [15]234.883.2
StructViT-S-8-1 [25]245.483.3
InternImage-T [48]305.083.5
MLLA-T [18]254.283.5
MaxViT-T [44]315.683.6
FAT-B3 [12]294.483.6
SMT-S [34]204.883.7
BiFormer-S [61]264.583.8
SECViT-S274.684.3
base model ~ 9.0GConvNeXt-S [39]508.783.1
NAT-S [19]517.883.7
Quadtree-B-b4 [42]6411.584.0
MOAT-1 [54]429.184.2
InternImage-S [48]508.084.2
GC-ViT-S [20]518.584.3
BiFormer-B [61]579.884.3
iFormer-B [41]489.484.6
FAT-B4 [12]529.384.8
SECViT-B579.885.2
large model ~ 18.0GCrossFormer-L [47]9216.184.0
SMT-L [34]8117.784.6
DaViT-B [9]8815.584.6
SG-Former-B [15]7815.684.7
iFormer-L [41]8714.084.8
InterImage-B [48]9716.084.9
GC-ViT-B [20]9014.885.0
RMT-L [13]9518.285.5
SECViT-L10118.285.7
+ +Table 3. Comparison with the state-of-the-art on ImageNet-1K classification. + +results of SECViT with different detection frameworks. The results show that SECViT performs better than its counterparts in all comparisons. + +Semantic Segmentation. We utilize Semantic FPN [26] and UperNet [52] to validate our SECViT's performance, implementing these frameworks via MMSegmentation [7]. The results of semantic segmentation can be found in the Tab. 6. All the FLOPs are measured with the input resolution of $512 \times 2048$ . SECViT achieves the best performance in all settings. + +# 4.2. SEC for MLLM + +SEC can greatly facilitate the design of vision language connectors in MLLMs. First, we conduct a rigorous compari + +
BackboneParams (M)FLOPs (G)Mask R-CNN 1×Params (M)FLOPs (G)RetinaNet 1×
APbAPb50APb75APmAPm50APm75APbAPb50APb75APsAPbAPL
PVTv2-B1 [46]3324341.854.345.938.861.241.62322541.261.943.925.444.554.3
FAT-B2 [12]3321545.267.949.041.364.644.02319644.065.247.227.547.758.8
RMT-T [13]3321847.168.851.742.665.845.92319945.166.248.128.848.961.1
SECViT-T3422147.869.552.543.066.746.32420245.866.849.229.149.860.9
MPViT-S [28]4326846.468.651.242.465.645.73224845.757.348.828.749.759.2
MLLA-T [18]4425546.869.551.542.166.445.0--------
STViT-S [22]4425247.670.052.343.166.846.5--------
RMT-S [13]4626249.070.853.943.967.847.43624447.869.151.832.151.863.5
SECViT-S4526249.970.954.744.668.347.73524048.469.452.031.353.363.8
ScalableViT-B [56]9534946.868.751.542.565.845.98533045.867.349.229.949.561.0
InternImage-S [48]6934047.869.852.843.367.146.7--------
MLLA-S [18]6331949.271.553.944.268.547.2--------
STViT-B [22]7035949.771.754.744.868.948.7--------
SECViT-B7637151.572.956.745.469.948.76334949.370.352.932.053.864.8
Focal-B [55]11053347.870.252.543.267.346.510151446.368.049.831.750.460.8
CSwin-B [10]9752648.770.453.943.967.847.3--------
InternImage-B [48]11550148.870.954.044.067.847.4--------
MLLA-B [18]11550250.572.055.445.069.348.6--------
SECViT-L11955052.073.557.346.370.649.810552750.271.453.933.254.566.3
+ +Table 4. Comparison to other backbones using "1×" schedule on COCO. + +
BackboneParams (M)FLOPs (G)\(AP^b\)\(AP_{50}^b\)\(AP_{75}^b\)\(AP^m\)\(AP_{50}^m\)\(AP_{75}^m\)
Mask R-CNN 3×+MS
GC-ViT-T [20]4829147.970.152.843.267.046.7
MLLA-T [18]4425548.871.053.643.868.046.8
SMT-S [34]4026549.070.153.443.467.346.7
InternImage-T [48]4927049.170.454.143.767.347.3
RMT-S [13]4626250.771.955.644.969.148.4
SECViT-S4526251.672.555.945.669.948.8
NAT-S [19]7033048.469.853.243.266.946.4
InternImage-S [48]6934049.771.154.544.568.547.8
SMT-B [34]5232849.871.054.444.068.047.3
MLLA-S [18]6331950.571.855.244.969.148.2
RMT-B [13]7337352.272.957.046.170.449.9
SECViT-B7537152.873.657.746.470.849.9
Cascade Mask R-CNN 3×+MS
GC-ViT-T [20]8577051.670.456.144.667.848.3
SMT-S [34]7874451.970.556.344.767.848.6
UniFormer-S [30]7974752.171.156.645.268.348.9
RMT-S [13]8374153.272.057.846.169.849.8
SECViT-S8374154.172.858.647.070.351.0
NAT-S [19]10880951.970.456.244.968.248.6
GC-ViT-S [20]10886652.471.057.145.468.549.3
CSWin-S [10]9282053.772.258.446.469.650.6
UniFormer-B [30]10787853.872.858.546.469.950.4
RMT-B [13]11185254.572.859.047.270.551.4
SECViT-B11484955.474.159.947.871.751.7
+ +son between SEC and various baseline vision language connectors based on LLaVA-1.5. Then, we compare LLaVA-1.5+SEC with several popular contemporary MLLMs. + +Strict Comparison with Baselines. In Tab. 7, we strictly compare various commonly used vision language connectors, including MLP, Resampler [1], Pooling, and EViT [32], which has achieved success in the design of ViT. Among these, MLP is the original design in LLaVA-1.5 [36], capable of achieving good results. However, it incurs significant computational cost due to the excessive vision tokens. To address this issue, some connectors at + +Table 5. Comparison with other backbones using $3 \times +\mathrm{MS}$ schedule on COCO. + +
ModelSemantic FPN 80K Params FLOPsUpernet 160K Params FLOPs
(M)(G)(%)(M)(G)(%)
PVTv2-B1 [46]1813642.5---
VAN-B1 [17]1814042.9---
RMT-T [13]1713646.4---
SECViT-T1813647.24489448.8
StructViT-S [25]2627146.9---
MogaNet-S [31]2918947.75594649.2
SMT-S [34]---5093549.2
SGFormer-S [15]2520549.052.598949.9
RMT-S [13]3018049.45693749.8
SECViT-S3018049.65693650.6
MogaNet-B [31]---74105050.1
InterImage-S [48]---80101750.2
StructViT-B [25]5452948.5---
RMT-B [13]5729450.483105152.0
SECViT-B6029150.786104852.2
MogaNet-L [31]---113117650.9
MLLA-B [18]---128118351.9
SGFormer-B [15]8147550.6109130452.0
RMT-L [13]9848251.4125124152.8
SECViT-L10347552.2131125653.8
+ +Table 6. Comparison with the state-of-the-art on ADE20K. + +tempt to use fewer vision tokens to accelerate LLaVA-1.5. Nonetheless, these adjustments inevitably lead to performance degradation. The results in Tab. 7 show that using SEC can effectively accelerate the inference of LLaVA-1.5 without causing performance degradation, and can even improve the performance of LLaVA-1.5 to a certain extent. + +Comparison with Popular MLLMs. In Tab. 8 and Tab. 9, we compare LLaVA-1.5 equipped with SEC as a vision-language connector with other MLLMs. It is evident that SEC not only enhances the performance of MLLMs across various benchmarks but also significantly improves the efficiency of the models. This fully demonstrates the + +
ModelConnectorV-T NumTimeSpeedTextVQAGQAVQAv2POPEMME
LLaVA-1.5MLP576+1194s1.0×58.262.078.586.11510.7
LLaVA-1.5+ResamplerResampler288+1126s1.5×52.156.876.083.11393.2
LLaVA-1.5+EViTMLP+EViT288+1126s1.5×54.660.077.984.31483.2
LLaVA-1.5+SECMLP+SEC288+1126s1.5×60.163.578.987.71510.7
LLaVA-1.5+ResamplerResampler256+1116s1.7×51.656.075.282.71387.2
LLaVA-1.5+PoolMLP+Pool256+1116s1.7×52.457.676.483.31415.5
LLaVA-1.5+EViTMLP+EViT256+1116s1.7×52.859.677.183.71443.7
LLaVA-1.5+SECMLP+SEC256+1116s1.7×59.663.278.687.11505.2
LLaVA-1.5+ResamplerResampler192+1102s1.9×50.155.274.382.71337.6
LLaVA-1.5+EViTMLP+EViT192+1102s1.9×51.658.676.383.11427.6
LLaVA-1.5+SECMLP+SEC192+1102s1.9×57.762.778.486.71500.1
LLaVA-1.5+ResamplerResampler144+194s2.1×47.654.672.081.91293.7
LLaVA-1.5+PoolMLP+Pool144+194s2.1×50.056.273.681.91310.7
LLaVA-1.5+EViTMLP+EViT144+194s2.1×51.258.076.083.11393.6
LLaVA-1.5+SECMLP+SEC144+194s2.1×56.862.078.086.11487.1
+ +Table 7. Comparison of different vision language connectors on LLaVA-1.5. "V-T Num" denotes the quantity of visual tokens. The computation expense is impacted by V-T Num, with larger values resulting in higher costs. "Speed" refers to the comparative inference velocity relative to LLaVA-1.5. "Time" is the average inference time. Inference speed are measured on the A100. + +
ModelLLMConnectorV-T NumResTextVQAGQAVQAv2VisWiz\( SQA_{img} \)Speed (↑)
7B LLM
Shikra [5]Vicuna-7BMLP257224--77.4---
IDEFICS-9B [27]LLaMA-7BCross Attn257224-38.450.935.5--
Qwen-VL [1]Qwen-7BResampler256448-59.378.835.267.1-
Qwen-VL-Chat [1]Qwen-7BResampler256448-57.578.238.968.2-
LLaVA-1.5 [35]Vicuna-7BMLP57733658.262.078.550.066.81.0×
LLaVA-1.5+SEC (ours)Vicuna-7BMLP+SEC25733659.663.278.952.869.61.7×
13B LLM
InstructBLIP [8]Vicuna-13BQ-Former32224-49.5-33.463.1-
BLIP-2 [29]Vicuna-13BQ-Former32224-41.041.019.561.0-
LLaVA-1.5 [35]Vicuna-13BMLP57733661.263.380.053.671.61.0×
LLaVA1.5+SEC (ours)Vicuna-13BMLP+SEC25733662.364.380.054.772.01.8×
+ +Table 8. Results on General VQA tasks. + +effectiveness of SEC in extracting visual information. + +# 4.3. Ablation Study + +In this section, we present some of the ablation study results for SEC, and more results can be found in the Appendix. + +Number of Vision Tokens in Each Clusters. The number of vision tokens has a significant impact on the performance and speed of the model. We thoroughly investigate the effect of the number of vision tokens on SECViT. As shown in Tab. 10, the number of vision tokens in each cluster greatly influences the model's performance. Specifically, in downstream dense prediction tasks, having too few tokens in each cluster leads to substantial performance degradation. When the number of tokens in each cluster is too large, the model's performance does not see a significant improvement, but its speed decreases. + +# 4.4. Visualization of SEC + +To further understand the working mechanism of SEC, we visualize some clustering results for SECViT. As shown in Fig. 4, the left side presents the clustering results of vision tokens at different stages of the model. From the clustering results, we analyze that in the shallow layers, the model + +distinguishes fine-grained features well, while in the deeper layers, it captures global semantic features effectively. The right side shows the Grad-CAM diagrams at different stages of the model, from which we can draw similar conclusions to the clustering results. More visualization results can be found in Appendix. + +Number of Vision Tokens Outputs by SEC. MLLM is quite sensitive to the number of vision tokens. We conduct a detailed exploration based on LLaVA-1.5 regarding the number of vision tokens output by SEC, as shown in Tab. 11. The first row represents the speed and performance of the original LLaVA-1.5 without using SEC. Compared to LLaVA-1.5, employing SEC effectively reduces the number of vision tokens and improves training efficiency. As the number of vision tokens decreases, the model's performance shows a slight decline, but its efficiency is further enhanced. + +# 5. Conclusion + +We propose a simple and straightforward clustering method for vision tokens—Semantic Equitable Clustering (SEC). This method assigns each token to a cluster by calculating the similarity between each token and a global token, and + +
ModelLLMConnectorV-T NumResPOPEMMBMM-VetSpeed (↑)
7B LLM
MiniGPT-4 [60]Vicuna-7BResampler3222472.224.322.1-
mPLUG-Ow12 [57]LLaMA2-7BResampler32224-49.4--
LLaMA-AdapterV2 [14]LLaMA2-7BLLaMA-Adapter257224-41.031.4-
Shikra [5]Vicuna-7BMLP257224-58.8--
Qwen-VL [1]Qwen-7BResampler256448-38.2--
Qwen-VL-Chat [1]Qwen-7BResampler256448-60.6--
LLaVA-1.5 [35]Vicuna-7BMLP57733686.164.331.11.0×
LLaVA1.5+SEC (ours)Vicuna-7BMLP+SEC14533686.168.431.72.1×
13B LLM
MiniGPT-4 [60]Vicuna-13BResampler32224--24.4-
BLIP-2 [29]Vicuna-13BQ-Former3222485.3-22.4-
LLaVA-1.5 [35]Vicuna-13BMLP57733686.267.736.11.0×
LLaVA-1.5+SEC (ours)Vicuna-13BMLP+SEC14533686.469.237.32.2×
+ +Table 9. Results on benchmark designed for MLLMs. + +
V-T numParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
98152.5200482.747.843.047.2
196153.1172283.0 (+0.3)48.2 (+0.4)43.4 (+0.4)47.5 (+0.3)
64152.5194682.7 (+0.0)47.8 (+0.0)42.8 (-0.2)46.9 (-0.3)
49152.4210282.6 (-0.1)47.5 (-0.3)42.7 (-0.3)47.7 (-0.5)
24152.3218682.0 (-0.7)45.9 (-1.9)40.6 (-2.4)44.6 (-2.6)
+ +Table 10. Effect of the number of vision tokens in each cluster. "V-T num" means the number of vision tokens in each cluster. The experiments are conducted based on SECViT-T. + +
V-T numTime SpeedTextVQAGQAVQAv2POPEMM-Vet
576+121h1.0×58.262.078.586.131.1
288+114h1.5×60.1(+1.9)63.5(+1.5)78.9(+0.4)87.7(+1.6)33.2(+2.1)
256+113h1.6×59.6(+1.4)63.2(+0.3)78.6(+0.1)87.1(+1.0)32.7(+1.6)
192+111h1.9×57.7(-0.5)62.7(+0.7)78.4(-0.1)86.7(+0.6)32.1(+1.0)
144+110h2.1×56.8(-1.4)62.0(+0.0)78.0(-0.5)86.1(+0.0)31.7(+0.6)
+ +Table 11. Effect of the number of vision tokens outputs by SEC. "V-T num" means the number of vision tokens output by SEC. The experiments are conducted based on LLaVA-1.5 [36]. + +completes the whole clustering process in only one step. Our clustering method takes into account the semantic information contained in the tokens, and ensures an equal number of tokens in each cluster, facilitating efficient parallel processing on modern GPUs. Based on Semantic Equitable Clustering, we designed SECViT, a versatile vision backbone that achieves impressive results across various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Besides, SEC can also be conveniently applied to multimodal large language models (MLLM) to serve as a vision language connector and benefits the model's efficiency. + +# 6. Acknowledgements + +This work is partially funded by Beijing Natural Science Foundation (4252054), Youth Innovation Promotion + +![](images/dad77df489835b054acdb275ca0d0404b2aec373fd532a5ac9ebaf2e3f7b9f41.jpg) +Figure 4. Visualization for SEC. + +Association CAS(Grant No.2022132), Beijing Nova Program(20230484276), and CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024005). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2, 6, 7, 8 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 5 +[3] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In CVPR, 2024. 2 +[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5 +[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 7, 8 +[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In ICLR, 2023. 3 +[7] MMSegmentation Contributors. Mmsegmentation, an open source semantic segmentation toolbox, 2020. 5 +[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7 +[9] Mingyu Ding, Bin Xiao, Noel Codella, et al. Davit: Dual attention vision transformers. In ECCV, 2022. 2, 5 +[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, et al. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 3, 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2 +[12] Qihang Fan, Huaibo Huang, Xiaogiang Zhou, and Ran He. Lightweight vision transformer with bidirectional interaction. In NeurIPS, 2023. 3, 5, 6 +[13] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 5, 6 +[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 8 +[15] SG-Former: Self guided Transformer with Evolving Token Reallocation. Sucheng ren, xingyi yang, songhua liu, xinchao wang. In ICCV, 2023. 5, 6 +[16] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 2 +[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6 +[18] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and + +Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 5, 6 +[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 5, 6 +[20] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In ICML, 2023. 5, 6 +[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask r-cnn. In ICCV, 2017. 5 +[22] Huaibo Huang, Xiaogiang Zhou, Jie Cao, Ran He, and Tieniu Tan. Vision transformer with super token sampling. In CVPR, 2023. 6 +[23] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver: General perception with iterative attention, 2021. 2 +[24] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 2 +[25] Manjin Kim, Paul Hongsuck Seo, Cordelia Schmid, and Minsu Cho. Learning correlation structures for vision transformers. In CVPR, 2024. 5, 6 +[26] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 5 +[27] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 7 +[28] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, 2022. 5, 6 +[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 7, 8 +[30] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning, 2022. 6 +[31] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z. Li. Moganet: Multi-order gated aggregation network. In ICLR, 2024. 6 +[32] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 2, 4, 6 +[33] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, and Kaiming He and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5 +[34] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6 + +[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 7, 8 +[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 3, 4, 6, 8 +[37] Kai Liu, Tianyi Wu, Cong Liu, and Guodong Guo. Dynamic group transformer: A general vision transformer backbone with dynamic group attention. In IJCAI, 2022. 1, 2 +[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 3, 4, 5 +[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, et al. A convnet for the 2020s. In CVPR, 2022. 5 +[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, 2021. 2 +[41] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng YAN. Inception transformer. In NeurIPS, 2022. 2, 5 +[42] Shitao Tang, Jiahui Zhang, Siyu Zhu, et al. Quadtree attention for vision transformers. In ICLR, 2022. 5 +[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, et al. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2, 4, 5 +[44] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 1, 2, 5 +[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 2 +[46] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):1-10, 2022. 2, 5, 6 +[47] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. In ICLR, 2022. 1, 5 +[48] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 5, 6 +[49] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic vision transformers with adaptive sequence length. In NeurIPS, 2021. 5 +[50] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. 5 +[51] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, 2022. 2 + +[52] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 5 +[53] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 5 +[54] Chenglin Yang, Siyuan Qiao, Qihang Yu, et al. Moat: Alternating mobile convolution and attention brings strong vision models. In ICLR, 2023. 5 +[55] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. In NeurIPS, 2021. 6 +[56] Rui Yang, Hailong Ma, Jie Wu, Yansong Tang, Xuefeng Xiao, Min Zheng, and Xiu Li. Scalablevit: Rethinking the context-oriented generalization of vision transformer. In ECCV, 2022. 6 +[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024. 8 +[58] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In ICCV, 2021. 5 +[59] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In CVPR, 2022. 5 +[60] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 8 +[61] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2, 5 \ No newline at end of file diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/images.zip b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7601c3b9fbc1f7f0880b7d4f6256e6fdec8c928c --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c981bb4f80f0f9865f693c960e23d5a6435cfc570e3a3d800a9c32341eeacadc +size 1123941 diff --git a/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/layout.json b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b62f146e78cb4729be7eadbbca834ca5ebab301a --- /dev/null +++ b/2025/Semantic Equitable Clustering_ A Simple and Effective Strategy for Clustering Vision Tokens/layout.json @@ -0,0 +1,7733 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 541, + 138 + ], + "type": "text", + "content": "Semantic Equitable Clustering: A Simple and Effective Strategy for Clustering Vision Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "spans": [ + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "text", + "content": "Qihang Fan " + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "text", + "content": ", Huaibo Huang" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "text", + "content": ", Mingrui Chen" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "text", + "content": ", Ran He" + }, + { + "bbox": [ + 149, + 160, + 443, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "type": "text", + "content": "MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China \n" + }, + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 175, + 523, + 232 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China \nfanqihang.159@gmail.com, huaibo.huang@cripac.ia.ac.cn, pharmier@hust.edu.cn, rhe@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 284, + 290, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 620 + ], + "type": "text", + "content": "The Vision Transformer (ViT) has gained prominence for its superior relational modeling prowess. However, its global attention mechanism's quadratic complexity poses substantial computational burdens. A common remedy spatially groups tokens for self-attention, reducing computational requirements. Nonetheless, this strategy neglects semantic information in tokens, possibly scattering semantically-linked tokens across distinct groups, thus compromising the efficacy of self-attention intended for modeling inter-token dependencies. Motivated by these insights, we introduce a fast and balanced clustering method, named Semantic Equitable Clustering (SEC). SEC clusters tokens based on their global semantic relevance in an efficient, straightforward manner. In contrast to traditional clustering methods requiring multiple iterations, our method achieves token clustering in a single pass. Additionally, SEC regulates the number of tokens per cluster, ensuring a balanced distribution for effective parallel processing on current computational platforms without necessitating further optimization. Capitalizing on SEC, we propose a versatile vision backbone, SECViT. Comprehensive experiments in image classification, object detection, instance segmentation, and semantic segmentation validate the effectiveness of SECViT. Moreover, SEC can be conveniently and swiftly applied to multimodal large language models (MLLM), such as LLaVA, to serve as a vision language connector, effectively accelerating the model's efficiency while maintaining unchanged or better performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 640, + 128, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 128, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 128, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "content": "Since its inception, the Vision Transformer (ViT)[11] has drawn considerable interest from the research community due to its robust modeling prowess. However, the quadratic" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 266, + 380, + 338 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 374, + 266 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 374, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 374, + 266 + ], + "type": "text", + "content": "(a) Window Partition" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 266, + 380, + 338 + ], + "lines": [ + { + "bbox": [ + 308, + 266, + 380, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 380, + 338 + ], + "type": "image", + "image_path": "bfedae319fa83a0de59aba8a9fed4488ae3931ea412df57b82603b7d4dda41be.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "type": "text", + "content": "X Semantic Information " + }, + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "type": "text", + "content": " Cluster Efficiency " + }, + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 310, + 341, + 375, + 365 + ], + "type": "text", + "content": " Equi-partition" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 366, + 545, + 443 + ], + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 443 + ], + "type": "text", + "content": "Figure 1. Comparison among Window Partition, Dynamic Group by k-means, and Semantic Equitable Clustering. Our Semantic Equitable Clustering incorporates image semantics while maintaining efficient clustering, eliminating the need for iterative processes such as in k-means. Furthermore, it enables equipartitioning of tokens, promoting efficient GPU processing without necessitating additional CUDA optimization." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 384, + 266, + 456, + 338 + ], + "blocks": [ + { + "bbox": [ + 400, + 258, + 434, + 266 + ], + "lines": [ + { + "bbox": [ + 400, + 258, + 434, + 266 + ], + "spans": [ + { + "bbox": [ + 400, + 258, + 434, + 266 + ], + "type": "text", + "content": "(b) k-means" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 266, + 456, + 338 + ], + "lines": [ + { + "bbox": [ + 384, + 266, + 456, + 338 + ], + "spans": [ + { + "bbox": [ + 384, + 266, + 456, + 338 + ], + "type": "image", + "image_path": "0d5a36a6d9ba18eca47eadf175213d7a8aec7ab9e2fefba6be9152e6d3af78d6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "lines": [ + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "spans": [ + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "text", + "content": " Semantic Information \n" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "text", + "content": " Cluster Efficiency \n" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 385, + 341, + 452, + 365 + ], + "type": "text", + "content": " Equi-partition" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 453, + 258, + 542, + 266 + ], + "lines": [ + { + "bbox": [ + 453, + 258, + 542, + 266 + ], + "spans": [ + { + "bbox": [ + 453, + 258, + 542, + 266 + ], + "type": "text", + "content": "(c)Semantic Equitable Clustering" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 462, + 266, + 535, + 338 + ], + "blocks": [ + { + "bbox": [ + 462, + 266, + 535, + 338 + ], + "lines": [ + { + "bbox": [ + 462, + 266, + 535, + 338 + ], + "spans": [ + { + "bbox": [ + 462, + 266, + 535, + 338 + ], + "type": "image", + "image_path": "6ca7c9ba783fe071e548b9dfb0324f7fa73d27c460c85350ab380b43ec772af5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "lines": [ + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "spans": [ + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "text", + "content": " Semantic Information " + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "text", + "content": " Cluster Efficiency " + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 464, + 341, + 530, + 365 + ], + "type": "text", + "content": " Equi-partition" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 532 + ], + "type": "text", + "content": "complexity of Self-Attention leads to significant computational overhead, thus constraining the practicality of ViT. A variety of strategies have been devised to alleviate this computational load, the most prevalent of which involves token grouping, thereby constraining the attention span of each token[10, 38, 44, 47]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 715 + ], + "type": "text", + "content": "Specifically, the Swin-Transformer [38] partitions tokens into multiple small windows, restricting token attention within each window. The CSWin-Transformer [10] adopts a cross-shaped grouping, endowing each token with a global receptive field. MaxViT [44] amalgamates window and grid attention, facilitating intra-window tokens to attend to their counterparts in other windows. However, these methods, solely reliant on spatial positioning, neglect token semantics, potentially restricting the self-attention's capacity to model semantic dependencies. To mitigate this, DGT [37] employs k-means clustering for query grouping, considering the semantic information of tokens for enhanced feature learning. Nonetheless, the iterative nature of k-means clustering and the potential for uneven token counts per cluster can impact the efficiency of parallel attention operations." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 202, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 202, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 202, + 712 + ], + "type": "text", + "content": "*Huaibo Huang is the corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4019" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 286, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 286, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 286, + 240 + ], + "type": "image", + "image_path": "82c9bce75d7bf1a1970ecdfcfbe91b46fd54fe2042e2030049cec92004ce354c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 288, + 275 + ], + "type": "text", + "content": "Figure 2. Left: Top-1 accuracy v.s. FLOPs on ImageNet-1K of recent SOTA models. Right: Comparison among different vision language connectors on LLaVA-1.5" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 529 + ], + "type": "text", + "content": "Given these considerations, an optimal token partitioning scheme should efficiently segregate tokens, incorporate semantic information, and efficiently utilize computational resources (e.g., GPU). In response, we introduce a simple, fast and equitable clustering approach named Semantic Equitable Clustering (SEC). SEC segments tokens based on their relevance to global semantic information. Specifically, we employ global pooling to generate a global token encapsulating global semantic information. The similarity between this global token and all other tokens is then computed, reflecting global semantic relevance. Upon obtaining the similarity matrix, tokens (excluding the global token) are sorted by similarity scores, and the tokens with similar scores are grouped into clusters, ensuring uniform token distribution across clusters. As depicted in Fig. 1, SEC comprehensively considers token semantics and completes the clustering process in a single iteration, unlike the multi-iteration k-means. The resulting clusters, containing an equal number of tokens, can be processed in parallel by the GPU efficiently." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 531, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 687 + ], + "type": "text", + "content": "Building upon Semantic Equitable Clustering (SEC), we introduce the Semantic Equitable Clustering Vision Transformer (SECViT), a versatile vision backbone that is adaptable to a wide spectrum of downstream tasks. As demonstrated in Fig. 2, SECViT exhibits significant performance improvements compared to previous state-of-the-art (SOTA) models. Impressively, SECViT attains an accuracy of " + }, + { + "bbox": [ + 46, + 531, + 287, + 687 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 46, + 531, + 287, + 687 + ], + "type": "text", + "content": " utilizing merely 4.6GFLOPS, without the need for additional training data or supervision. This superior performance is maintained across different model scales. Furthermore, SECViT proves its proficiency in downstream tasks, including but not limited to, object detection, instance segmentation, and semantic segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Beyond vision tasks, we also apply SEC to multimodal large language models (MLLM) such as LLaVA-1.5 [36] to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "serve as an efficient vision language connector. Specifically, we use SEC to cluster the vision tokens, and then merge all the tokens at corresponding positions within each cluster into a single token. Experiments demonstrate that this approach significantly enhances the efficiency of LLaVA-1.5 while improving the model's performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 158, + 398, + 171 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 179, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 547, + 396 + ], + "type": "text", + "content": "Vision Transformer. The Vision Transformer (ViT) [11] is considered a powerful visual architecture. Many works have improved the Vision Transformer, including enhancing its training efficiency and reducing its computational cost [10, 24, 38, 43, 61]. DeiT [43] uses distillation loss and incorporates extensive data augmentation methods into the ViT training process. Hierarchical structures represented by PVT [16, 41, 45, 46, 51] reduce the number of tokens in global attention by downsampling the keys and values (KV), thereby low the computational cost. In addition to them, some methods directly prune tokens based on their importance, retaining important tokens [32, 40]. This reduces the number of tokens and subsequently lowers the computational cost of the model. Another highly representative approach is to group all tokens such that each token can only attend to tokens within its own group [9, 10, 37, 38, 61]. This method also significantly reduces the computational cost of self-attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 564 + ], + "type": "text", + "content": "Grouping-Based Vision Transformer. Most grouping-based attention mechanisms perform grouping based on spatial structure [9, 10, 37, 38, 44]. Specifically, the SwinTransformer [38] divides all tokens into equally sized windows based on their spatial positions, where each token can only attend to tokens within its own window. This significantly reduces the model's computational cost. In addition to dividing tokens into small windows along the spatial dimension, DaViT [9] also splits channels into multiple groups along the channel dimension. Unlike the above methods that only consider positional information for grouping, DGT [37] takes semantic information into account by using k-means clustering to group the queries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 576, + 547, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 721 + ], + "type": "text", + "content": "Vision Language Connector. The vision language connector is a critical component in MLLMs [3, 23, 36]. It aligns vision tokens with language tokens. Typical vision language connectors include MLP [36], Resampler [1], C-Abstractor [3], and others. Although MLP performs well, it introduces a significant number of vision tokens, which hampers the model's efficiency. On the other hand, connectors like Resampler improve the model's efficiency, but at the cost of reduced performance. Unlike these methods, our proposed SEC consider the semantic information of each token and significantly enhances the model's efficiency while maintaining its performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4020" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 532, + 195 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 532, + 195 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 532, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 532, + 195 + ], + "type": "image", + "image_path": "5ac3a4564f17d87dc7600153d4694641f9a1af6bd66688dad7304a0f57e71375.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 197, + 534, + 337 + ], + "blocks": [ + { + "bbox": [ + 61, + 197, + 534, + 337 + ], + "lines": [ + { + "bbox": [ + 61, + 197, + 534, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 534, + 337 + ], + "type": "image", + "image_path": "a61e8739d6a4850d2400576335e0fa1a8e01d661b4868da9f8fff3f2bc62a9de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 547, + 361 + ], + "type": "text", + "content": "Figure 3. (a) Illustration of SECViT (b) Applying SEC to vision language connector. (c) Illustration of Semantic Equitable Clustering for ViT and vision language connector." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 105, + 382 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 389, + 168, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 168, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 168, + 400 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": "The overall architecture of SECViT is shown in Fig. 3(a). SECViT consists of four stages with downsampling factors of " + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\frac{1}{16}" + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": ", respectively. This structural design facilitates downstream tasks, such as object detection, in constructing feature pyramids. A SECViT block is composed of three modules. For each block, the input tensor " + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "inline_equation", + "content": "X_{in} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 46, + 407, + 288, + 539 + ], + "type": "text", + "content": " is fed into the CPE to introduce the positional information. Then, The Self-Attention based on the Semantic Equitable Clustering (SEC) is employed to serve as the token mixer. The final FFN is utilized to integrate channel-wise information of tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 540, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 625 + ], + "type": "text", + "content": "Beyond the design of the backbone, we also utilize SEC in the design of the vision language connector in MLLM [36]. For the vision tokens output by ViT, we use SEC to cluster the tokens. For each position corresponding to a cluster, we use attentive pooling to merge them into a single token, thereby reducing the number of vision tokens. The process is shown in Fig. 3(b)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 634, + 214, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 214, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 214, + 647 + ], + "type": "text", + "content": "3.2. Semantic Equitable Clustering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "As previously mentioned, the design objectives of Semantic Equitable Clustering are threefold: 1) Fully consider the semantic information contained in different tokens during clustering. 2) Unlike k-means and other clustering methods that require multiple iterations, Semantic Equitable Cluster-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 370, + 546, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 546, + 442 + ], + "type": "text", + "content": "ing can complete clustering in a single step. 3) Ensure an equal number of tokens in each cluster to facilitate parallel processing on GPUs. In the following paragraphs, we will describe in detail how our Semantic Equitable Clustering achieves these three objectives. And the whole process is illustrated in the Fig. 3(c)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 455, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 546, + 538 + ], + "type": "text", + "content": "Single Clustering Center Related to Semantics. k-means is relatively complex for two reasons. First, it has multiple cluster centers, and each token needs to calculate its distance to each cluster center to determine its cluster membership. Second, the determination of each cluster center in k-means is not precise and requires multiple iterations to accurately establish the cluster centers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 539, + 547, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 547, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 547, + 683 + ], + "type": "text", + "content": "To address these two issues, we first discard the use of multiple cluster centers and instead calculate the distance between each token and a single center. Based on each token's distance to this center, we divide the tokens into different intervals. Then, to ensure that our chosen center contains the most comprehensive semantic information, we directly use the result of average pooling of all tokens as the center token. This is because, in most vision foundation models, the output of the average pool is assumed to contain the richest semantic information and is thus used for classification [6, 10, 12, 38]. Specifically, the process for determining the cluster center is shown in Eq. 1:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 687, + 545, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 687, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 348, + 687, + 545, + 706 + ], + "type": "interline_equation", + "content": "Q = W _ {Q} X, K = W _ {K} X, V = W _ {V} X, \\tag {1}", + "image_path": "7c66d24e922297f1f185225c6fb80a822332e2b6af439fa52e149b7293bb6c2c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 399, + 703, + 462, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 703, + 462, + 715 + ], + "spans": [ + { + "bbox": [ + 399, + 703, + 462, + 715 + ], + "type": "interline_equation", + "content": "k _ {c} = \\operatorname {P o o l} (K).", + "image_path": "9822a00ccb724ed1ef84af182527e3aac3be9214b1fe52b0a52ee30f7db80a69.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4021" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "W_{K}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is a learnable matrix. " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "k_{c}" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the determined cluster center. " + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": " is the set of input tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 114, + 287, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 287, + 246 + ], + "type": "text", + "content": "Distance Metric Suitable for ViT. Unlike the Euclidean distance calculation used in the k-means algorithm for computing the distance between tokens, during the actual computation of Self-Attention, similarity between query and key is computed through dot product. To better adapt to the characteristics of Self-Attention, we also measure the distance between tokens using a method similar to dot product. Specifically, we calculate the cosine similarity between the cluster center and each token, and then sort the tokens according to the magnitude of the computed results. The specific process is shown in Eq. 2:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 255, + 217, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 255, + 217, + 281 + ], + "spans": [ + { + "bbox": [ + 132, + 255, + 217, + 281 + ], + "type": "interline_equation", + "content": "s i m = \\frac {K \\cdot k _ {c}}{| | K | | \\cdot | | k _ {c} | |},", + "image_path": "accbea372ce74b7a5a51a19cb754c6b5bcece85378b248f239dfa68bc251fae4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 279, + 287, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 279, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 132, + 279, + 287, + 296 + ], + "type": "interline_equation", + "content": "i d x = \\operatorname {a r g s o r t} (s i m), \\tag {2}", + "image_path": "295bc07cc75ce88ecc80ecb52081b2f76b26d5c4b0f3c55b0a5b6cf32bab78ff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 298, + 256, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 298, + 256, + 312 + ], + "spans": [ + { + "bbox": [ + 79, + 298, + 256, + 312 + ], + "type": "interline_equation", + "content": "Q ^ {*} = Q [ i d x ], K ^ {*} = K [ i d x ], V ^ {*} = V [ i d x ].", + "image_path": "fe76e8fb4feeda2259f496d0ade2da236f48e5fc119e1fa0fcd371535f057106.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "sim" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " is the similarity matrix between " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "k_{c}" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\operatorname{argsort}(sim)" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " returns the indices of " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "sim" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " sorted in descending order. " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "Q^{*}, K^{*}, V^{*}" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": " rearranged according to " + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\operatorname{argsort}(sim)" + }, + { + "bbox": [ + 47, + 321, + 287, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "text", + "content": "Equally Partition Tokens based on Distance. The obtained " + }, + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "inline_equation", + "content": "Q^{*}, K^{*}" + }, + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 47, + 387, + 287, + 483 + ], + "type": "text", + "content": " from the previous step have been sorted based on their distances to the cluster center. For the design of vision backbone, we directly group them, so tokens with similar distances to the cluster center are classified into the same cluster. This allows us to directly control an equal number of tokens in each cluster. This process can be clearly illustrated in Fig. 3(c) and denoted as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 494, + 235, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 494, + 235, + 506 + ], + "spans": [ + { + "bbox": [ + 101, + 494, + 235, + 506 + ], + "type": "interline_equation", + "content": "Q _ {m} = Q ^ {*} [ m \\times N: (m + 1) N ],", + "image_path": "241b62440ade778946baefa313224e4f1cd5915e9861bd47fd229bd8e6514faf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 101, + 508, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 508, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 101, + 508, + 287, + 521 + ], + "type": "interline_equation", + "content": "K _ {m} = K ^ {*} [ m \\times N: (m + 1) N ], \\tag {3}", + "image_path": "e8eca9e0920092a16b73e6775002018e16f909401317fe14d1d6bf48544b81b3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 523, + 235, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 523, + 235, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 235, + 536 + ], + "type": "interline_equation", + "content": "V _ {m} = V ^ {*} [ m \\times N: (m + 1) N ].", + "image_path": "ad535f098a88da19d5d9b591c0e8774eef3f09e57b0f0e6080ac87b850262810.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "text", + "content": " is the basic token number of each cluster for equal partition and " + }, + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 47, + 546, + 287, + 569 + ], + "type": "text", + "content": " is the index of the cluster" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 571, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 287, + 653 + ], + "type": "text", + "content": "Based on the above steps, we have completed the clustering process that captures semantic information in the image with minimal sorting cost. Moreover, compared to k-means, we have achieved equi-partitioning of each cluster. After clustering is completed, we apply standard Self-Attention to the tokens within each cluster, thereby completing the interaction of information between tokens:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 666, + 287, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 666, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 110, + 666, + 287, + 678 + ], + "type": "interline_equation", + "content": "Y _ {m} = \\operatorname {A t t n} \\left(Q _ {m}, K _ {m}, V _ {m}\\right). \\tag {4}", + "image_path": "80ac186e59cc728354a6250846199c00306357e170a4fbcba6818f021dac450b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "For the design of vision language connector, we group the tokens according to their similarity, and the tokens" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 520, + 84 + ], + "type": "text", + "content": "within each group are interleaved, as shown in Eq. 5:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 95, + 470, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 95, + 470, + 107 + ], + "spans": [ + { + "bbox": [ + 383, + 95, + 470, + 107 + ], + "type": "interline_equation", + "content": "Q _ {n} = Q ^ {*} [ n: N: L ],", + "image_path": "e5e358dda324f0b08cd03e90d629b185ae93062c22a85a0a22dfd6b3f3d61673.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 110, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 110, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 382, + 110, + 545, + 121 + ], + "type": "interline_equation", + "content": "K _ {n} = K ^ {*} [ n: N: L ], \\tag {5}", + "image_path": "57698ec634d99dc3e151c37372f1414956c17f8c1de7bc7a22aeb44541ff6212.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 125, + 470, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 125, + 470, + 137 + ], + "spans": [ + { + "bbox": [ + 386, + 125, + 470, + 137 + ], + "type": "interline_equation", + "content": "V _ {n} = V ^ {*} [ n: N: L ].", + "image_path": "ca21017d07e5eb9226d97217c045fc7c13f1929a930bc4c9a595fd3ea8e739a2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "content": " is the token's sequence length, " + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "content": " is the index of group tokens. " + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "content": " is the basic token number of each cluster. After obtaining the token groups, we perform pooling on " + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 305, + 146, + 545, + 218 + ], + "type": "text", + "content": " to effectively reduce the number of tokens input to the LLM, with each group's output becoming a single token, as shown in Eq 6." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 228, + 545, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 228, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 359, + 228, + 545, + 242 + ], + "type": "interline_equation", + "content": "Y _ {n} = \\operatorname {A t t n} \\left(\\operatorname {P o o l} \\left(Q _ {n}\\right), K _ {n}, V _ {n}\\right). \\tag {6}", + "image_path": "3cd5095f2dea5fa80a8c164ebe1d04a03953b3ca6b9e0b46783a8e9fae70b6ac.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 251, + 492, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 492, + 263 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 492, + 263 + ], + "type": "text", + "content": "3.3. Difference between SEC and EViT." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 305 + ], + "type": "text", + "content": "We use the most representative example, EViT [32], to illustrate the differences between SEC and other methods based on the similarity between the global token and other tokens." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 307, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 389 + ], + "type": "text", + "content": "Pruning v.s. Clustering. Most similarity-based methods, such as EViT, are pruning methods, where tokens with low similarity to the [cls] token are merged during the forward process, thereby reducing the number of tokens and decreasing computational cost. In contrast, our proposed SECViT employs a clustering-based approach, performing attention operations within each cluster." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 391, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 391, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 391, + 545, + 521 + ], + "type": "text", + "content": "The role of the [cls] token. In methods like EViT, the [cls] token serves as a measure of importance of a token. Each token computes its similarity to the [cls] token, with higher similarity tokens deemed more important. The less important tokens are abandoned. In contrast, in SEC, the [cls] token (obtained by average pooling over all tokens) measures similarity between tokens. Each token computes its similarity score to the [cls] token; tokens with similar scores are considered to be more similar and grouped into one cluster. Attention is calculated only within the same cluster." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 534, + 387, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 387, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 387, + 547 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 639 + ], + "type": "text", + "content": "We first make strict comparison with hierarchical/plain baselines. Then we conduct experiments on a wide range of vision tasks for SECViT, including image classification, object detection, instance segmentation, and semantic segmentation. We also verify the role of SEC in MLLM based on LLaVA-1.5 [36]. More details, experiments, and comparison of models' efficiency can be found in the Appendix." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 432, + 658 + ], + "type": "text", + "content": "4.1. SEC for vision models" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "Strict Comparison with Baselines. We select two baselines: hierarchical backbone Swin-Transformer [38] and plain backbone DeiT [43] to make a comparison with our SEC based model. In the comparison models (SEC-Swin" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4022" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 62, + 282, + 231 + ], + "blocks": [ + { + "bbox": [ + 53, + 62, + 282, + 231 + ], + "lines": [ + { + "bbox": [ + 53, + 62, + 282, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 62, + 282, + 231 + ], + "type": "table", + "html": "
ModelParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
DeiT-S [43]224.6320479.844.540.143.0
EViT-DeiT-S (keeprate=0.9)224.0342879.8not suitnot suitnot suit
SEC-DeiT-S (num_cluster=4)224.1341280.5 (+0.7)47.7 (+3.2)42.7 (+2.6)47.5 (+4.5)
DeiT-B8617.6150281.8---
SEC-DeiT-B8614.8168282.4 (+0.6)---
Swin-T294.5172381.343.739.844.5
SEC-Swin-T294.8148283.8 (+2.5)48.3 (+4.6)43.4 (+3.6)49.3 (+4.8)
Swin-S508.8100683.045.741.147.6
SEC-Swin-S509.280485.0 (+2.0)50.2 (+4.5)44.7 (+3.6)51.3 (+3.7)
", + "image_path": "3b595e4907f31355ede69baec4904abb44339840974f98f930ec98ce67040d79.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 51, + 258, + 285, + 346 + ], + "blocks": [ + { + "bbox": [ + 47, + 233, + 286, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 233, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 286, + 255 + ], + "type": "text", + "content": "Table 1. Comparison with Hierarchy/Plain baselines. Inference speed are measured on the A100 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 258, + 285, + 346 + ], + "lines": [ + { + "bbox": [ + 51, + 258, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 51, + 258, + 285, + 346 + ], + "type": "table", + "html": "
ModelParams (M)FLOPs (G)MethodPretrain epochAcc(%)
Swin-B [38]8815.4Supervised-83.5
ConvNeXt V2-B [50]8815.4Supervised-84.3
SEC-Swin-B8816.2Supervised-85.3
Swin-B [38]8815.4SimMIM [53]80084.0(+0.5)
ConvNeXt V2-B [50]8815.4FCMAE [50]80084.6(+0.3)
SEC-Swin-B8816.2SimMIM [53]80085.9(+0.6)
", + "image_path": "f5037ae6cb005af4cbb8916d1bbe35e765478d36308ea9fd60728bbabd12f5ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 286, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 286, + 467 + ], + "type": "text", + "content": "and SEC-DeiT), we merely substitute the attention mechanism in the original model with our SEC based Self-Attention and without introducing any other modules. As shown in Tab. 1, we conduct experiments on image classification, object detection, insatance segmentation and semantic segmentation, the simple replacement of the attention mechanism yields significant advantages in both performance and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 468, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 515 + ], + "type": "text", + "content": "In addition to the supervised scenario, we also train the model with SimMIM [53] in the self-supervised scenario. As shown in Tab. 2, SEC also performs exceptionally well in the self-supervised scenario." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "text", + "content": "Image Classification. We compare our SECViT with numerous state-of-the-art models, the results are shown in Tab.3. We adopt the training strategy proposed in DeiT [43], with the only supervision is cross entropy loss. All of our models are trained from scratch for 300 epochs with the input resolution of " + }, + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "text", + "content": ". SECViT consistently outperforms preceding models across all scales. Notably, SECViT-S attains a Top1-accuracy of " + }, + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 46, + 530, + 286, + 651 + ], + "type": "text", + "content": " with a mere 27M parameters and 4.6G FLOPs. The comparison of the models' efficiency can be found in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Object Detection and Instance Segmentation. We utilize MMDetection [4] to implement Mask-RCNN [21], Cascade Mask R-CNN [2], and RetinaNet [33] to evaluate the performance of the SECViT. Tab. 5 and Tab. 4 show the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 539, + 494 + ], + "blocks": [ + { + "bbox": [ + 53, + 348, + 280, + 358 + ], + "lines": [ + { + "bbox": [ + 53, + 348, + 280, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 348, + 280, + 358 + ], + "type": "text", + "content": "Table 2. Comparison with baselines on self-supervised setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 539, + 494 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 539, + 494 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 539, + 494 + ], + "type": "table", + "html": "
CostModelParmas (M)FLOPs (G)Top1-acc (%)
tiny model ~ 2.5GPVTv2-b1 [46]132.178.7
TCFormer-light [59]143.879.4
QuadTree-B-b1 [42]142.380.0
MPViT-XS [28]112.980.9
BiFormer-T [61]132.281.4
CrossFormer-T [47]282.981.5
FAT-B2 [12]142.081.9
GC-ViT-XT [20]202.682.0
SMT-T [34]122.482.2
RMT-T [13]142.582.4
SECViT-T152.582.7
small model ~ 4.5GPS-ViT-B14 [58]215.481.7
DVT-T2T-ViT-19 [49]396.281.9
ConvNeXt-T [39]294.582.1
TCFormer [59]265.882.3
SG-Former-S [15]234.883.2
StructViT-S-8-1 [25]245.483.3
InternImage-T [48]305.083.5
MLLA-T [18]254.283.5
MaxViT-T [44]315.683.6
FAT-B3 [12]294.483.6
SMT-S [34]204.883.7
BiFormer-S [61]264.583.8
SECViT-S274.684.3
base model ~ 9.0GConvNeXt-S [39]508.783.1
NAT-S [19]517.883.7
Quadtree-B-b4 [42]6411.584.0
MOAT-1 [54]429.184.2
InternImage-S [48]508.084.2
GC-ViT-S [20]518.584.3
BiFormer-B [61]579.884.3
iFormer-B [41]489.484.6
FAT-B4 [12]529.384.8
SECViT-B579.885.2
large model ~ 18.0GCrossFormer-L [47]9216.184.0
SMT-L [34]8117.784.6
DaViT-B [9]8815.584.6
SG-Former-B [15]7815.684.7
iFormer-L [41]8714.084.8
InterImage-B [48]9716.084.9
GC-ViT-B [20]9014.885.0
RMT-L [13]9518.285.5
SECViT-L10118.285.7
", + "image_path": "645e9c7be184e6fea1bb24e8fe9ec185ec719a723beef8558815a8fdfc664342.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 545, + 517 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-art on ImageNet-1K classification." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 529, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 545, + 566 + ], + "type": "text", + "content": "results of SECViT with different detection frameworks. The results show that SECViT performs better than its counterparts in all comparisons." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "content": "Semantic Segmentation. We utilize Semantic FPN [26] and UperNet [52] to validate our SECViT's performance, implementing these frameworks via MMSegmentation [7]. The results of semantic segmentation can be found in the Tab. 6. All the FLOPs are measured with the input resolution of " + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "inline_equation", + "content": "512 \\times 2048" + }, + { + "bbox": [ + 304, + 579, + 545, + 664 + ], + "type": "text", + "content": ". SECViT achieves the best performance in all settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 404, + 682 + ], + "type": "text", + "content": "4.2. SEC for MLLM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "SEC can greatly facilitate the design of vision language connectors in MLLMs. First, we conduct a rigorous compari" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4023" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 61, + 520, + 262 + ], + "blocks": [ + { + "bbox": [ + 72, + 61, + 520, + 262 + ], + "lines": [ + { + "bbox": [ + 72, + 61, + 520, + 262 + ], + "spans": [ + { + "bbox": [ + 72, + 61, + 520, + 262 + ], + "type": "table", + "html": "
BackboneParams (M)FLOPs (G)Mask R-CNN 1×Params (M)FLOPs (G)RetinaNet 1×
APbAPb50APb75APmAPm50APm75APbAPb50APb75APsAPbAPL
PVTv2-B1 [46]3324341.854.345.938.861.241.62322541.261.943.925.444.554.3
FAT-B2 [12]3321545.267.949.041.364.644.02319644.065.247.227.547.758.8
RMT-T [13]3321847.168.851.742.665.845.92319945.166.248.128.848.961.1
SECViT-T3422147.869.552.543.066.746.32420245.866.849.229.149.860.9
MPViT-S [28]4326846.468.651.242.465.645.73224845.757.348.828.749.759.2
MLLA-T [18]4425546.869.551.542.166.445.0--------
STViT-S [22]4425247.670.052.343.166.846.5--------
RMT-S [13]4626249.070.853.943.967.847.43624447.869.151.832.151.863.5
SECViT-S4526249.970.954.744.668.347.73524048.469.452.031.353.363.8
ScalableViT-B [56]9534946.868.751.542.565.845.98533045.867.349.229.949.561.0
InternImage-S [48]6934047.869.852.843.367.146.7--------
MLLA-S [18]6331949.271.553.944.268.547.2--------
STViT-B [22]7035949.771.754.744.868.948.7--------
SECViT-B7637151.572.956.745.469.948.76334949.370.352.932.053.864.8
Focal-B [55]11053347.870.252.543.267.346.510151446.368.049.831.750.460.8
CSwin-B [10]9752648.770.453.943.967.847.3--------
InternImage-B [48]11550148.870.954.044.067.847.4--------
MLLA-B [18]11550250.572.055.445.069.348.6--------
SECViT-L11955052.073.557.346.370.649.810552750.271.453.933.254.566.3
", + "image_path": "30cb13a7c8ecab4dd19dc323b6db6be27d56a6c8b1818e27935212006b10083b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 277, + 280, + 536 + ], + "blocks": [ + { + "bbox": [ + 163, + 264, + 428, + 274 + ], + "lines": [ + { + "bbox": [ + 163, + 264, + 428, + 274 + ], + "spans": [ + { + "bbox": [ + 163, + 264, + 428, + 274 + ], + "type": "text", + "content": "Table 4. Comparison to other backbones using \"1×\" schedule on COCO." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 277, + 280, + 536 + ], + "lines": [ + { + "bbox": [ + 59, + 277, + 280, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 277, + 280, + 536 + ], + "type": "table", + "html": "
BackboneParams (M)FLOPs (G)\\(AP^b\\)\\(AP_{50}^b\\)\\(AP_{75}^b\\)\\(AP^m\\)\\(AP_{50}^m\\)\\(AP_{75}^m\\)
Mask R-CNN 3×+MS
GC-ViT-T [20]4829147.970.152.843.267.046.7
MLLA-T [18]4425548.871.053.643.868.046.8
SMT-S [34]4026549.070.153.443.467.346.7
InternImage-T [48]4927049.170.454.143.767.347.3
RMT-S [13]4626250.771.955.644.969.148.4
SECViT-S4526251.672.555.945.669.948.8
NAT-S [19]7033048.469.853.243.266.946.4
InternImage-S [48]6934049.771.154.544.568.547.8
SMT-B [34]5232849.871.054.444.068.047.3
MLLA-S [18]6331950.571.855.244.969.148.2
RMT-B [13]7337352.272.957.046.170.449.9
SECViT-B7537152.873.657.746.470.849.9
Cascade Mask R-CNN 3×+MS
GC-ViT-T [20]8577051.670.456.144.667.848.3
SMT-S [34]7874451.970.556.344.767.848.6
UniFormer-S [30]7974752.171.156.645.268.348.9
RMT-S [13]8374153.272.057.846.169.849.8
SECViT-S8374154.172.858.647.070.351.0
NAT-S [19]10880951.970.456.244.968.248.6
GC-ViT-S [20]10886652.471.057.145.468.549.3
CSWin-S [10]9282053.772.258.446.469.650.6
UniFormer-B [30]10787853.872.858.546.469.950.4
RMT-B [13]11185254.572.859.047.270.551.4
SECViT-B11484955.474.159.947.871.751.7
", + "image_path": "8fbbe1628c6000ad91d39777b38ef9a2b5f60ca0170a1966b2b3e1f8fcedee81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 567, + 286, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 286, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 286, + 603 + ], + "type": "text", + "content": "son between SEC and various baseline vision language connectors based on LLaVA-1.5. Then, we compare LLaVA-1.5+SEC with several popular contemporary MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 617, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 286, + 712 + ], + "type": "text", + "content": "Strict Comparison with Baselines. In Tab. 7, we strictly compare various commonly used vision language connectors, including MLP, Resampler [1], Pooling, and EViT [32], which has achieved success in the design of ViT. Among these, MLP is the original design in LLaVA-1.5 [36], capable of achieving good results. However, it incurs significant computational cost due to the excessive vision tokens. To address this issue, some connectors at" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 285, + 531, + 527 + ], + "blocks": [ + { + "bbox": [ + 47, + 537, + 286, + 559 + ], + "lines": [ + { + "bbox": [ + 47, + 537, + 286, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 286, + 559 + ], + "type": "text", + "content": "Table 5. Comparison with other backbones using " + }, + { + "bbox": [ + 47, + 537, + 286, + 559 + ], + "type": "inline_equation", + "content": "3 \\times +\\mathrm{MS}" + }, + { + "bbox": [ + 47, + 537, + 286, + 559 + ], + "type": "text", + "content": " schedule on COCO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 285, + 531, + 527 + ], + "lines": [ + { + "bbox": [ + 321, + 285, + 531, + 527 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 531, + 527 + ], + "type": "table", + "html": "
ModelSemantic FPN 80K Params FLOPsUpernet 160K Params FLOPs
(M)(G)(%)(M)(G)(%)
PVTv2-B1 [46]1813642.5---
VAN-B1 [17]1814042.9---
RMT-T [13]1713646.4---
SECViT-T1813647.24489448.8
StructViT-S [25]2627146.9---
MogaNet-S [31]2918947.75594649.2
SMT-S [34]---5093549.2
SGFormer-S [15]2520549.052.598949.9
RMT-S [13]3018049.45693749.8
SECViT-S3018049.65693650.6
MogaNet-B [31]---74105050.1
InterImage-S [48]---80101750.2
StructViT-B [25]5452948.5---
RMT-B [13]5729450.483105152.0
SECViT-B6029150.786104852.2
MogaNet-L [31]---113117650.9
MLLA-B [18]---128118351.9
SGFormer-B [15]8147550.6109130452.0
RMT-L [13]9848251.4125124152.8
SECViT-L10347552.2131125653.8
", + "image_path": "f3df275ca490173923467b7266a9a53400b06adca108bc4a933aeaed53d1e182.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 529, + 532, + 539 + ], + "lines": [ + { + "bbox": [ + 318, + 529, + 532, + 539 + ], + "spans": [ + { + "bbox": [ + 318, + 529, + 532, + 539 + ], + "type": "text", + "content": "Table 6. Comparison with the state-of-the-art on ADE20K." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 625 + ], + "type": "text", + "content": "tempt to use fewer vision tokens to accelerate LLaVA-1.5. Nonetheless, these adjustments inevitably lead to performance degradation. The results in Tab. 7 show that using SEC can effectively accelerate the inference of LLaVA-1.5 without causing performance degradation, and can even improve the performance of LLaVA-1.5 to a certain extent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Comparison with Popular MLLMs. In Tab. 8 and Tab. 9, we compare LLaVA-1.5 equipped with SEC as a vision-language connector with other MLLMs. It is evident that SEC not only enhances the performance of MLLMs across various benchmarks but also significantly improves the efficiency of the models. This fully demonstrates the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4024" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 61, + 516, + 218 + ], + "blocks": [ + { + "bbox": [ + 77, + 61, + 516, + 218 + ], + "lines": [ + { + "bbox": [ + 77, + 61, + 516, + 218 + ], + "spans": [ + { + "bbox": [ + 77, + 61, + 516, + 218 + ], + "type": "table", + "html": "
ModelConnectorV-T NumTimeSpeedTextVQAGQAVQAv2POPEMME
LLaVA-1.5MLP576+1194s1.0×58.262.078.586.11510.7
LLaVA-1.5+ResamplerResampler288+1126s1.5×52.156.876.083.11393.2
LLaVA-1.5+EViTMLP+EViT288+1126s1.5×54.660.077.984.31483.2
LLaVA-1.5+SECMLP+SEC288+1126s1.5×60.163.578.987.71510.7
LLaVA-1.5+ResamplerResampler256+1116s1.7×51.656.075.282.71387.2
LLaVA-1.5+PoolMLP+Pool256+1116s1.7×52.457.676.483.31415.5
LLaVA-1.5+EViTMLP+EViT256+1116s1.7×52.859.677.183.71443.7
LLaVA-1.5+SECMLP+SEC256+1116s1.7×59.663.278.687.11505.2
LLaVA-1.5+ResamplerResampler192+1102s1.9×50.155.274.382.71337.6
LLaVA-1.5+EViTMLP+EViT192+1102s1.9×51.658.676.383.11427.6
LLaVA-1.5+SECMLP+SEC192+1102s1.9×57.762.778.486.71500.1
LLaVA-1.5+ResamplerResampler144+194s2.1×47.654.672.081.91293.7
LLaVA-1.5+PoolMLP+Pool144+194s2.1×50.056.273.681.91310.7
LLaVA-1.5+EViTMLP+EViT144+194s2.1×51.258.076.083.11393.6
LLaVA-1.5+SECMLP+SEC144+194s2.1×56.862.078.086.11487.1
", + "image_path": "ff193ad143bc4fe3d22a757cc80a23d5395898256a23378743805a5e730243f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 255, + 523, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "type": "text", + "content": "Table 7. Comparison of different vision language connectors on LLaVA-1.5. \"V-T Num\" denotes the quantity of visual tokens. The computation expense is impacted by V-T Num, with larger values resulting in higher costs. \"Speed\" refers to the comparative inference velocity relative to LLaVA-1.5. \"Time\" is the average inference time. Inference speed are measured on the A100." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 255, + 523, + 392 + ], + "lines": [ + { + "bbox": [ + 72, + 255, + 523, + 392 + ], + "spans": [ + { + "bbox": [ + 72, + 255, + 523, + 392 + ], + "type": "table", + "html": "
ModelLLMConnectorV-T NumResTextVQAGQAVQAv2VisWiz\\( SQA_{img} \\)Speed (↑)
7B LLM
Shikra [5]Vicuna-7BMLP257224--77.4---
IDEFICS-9B [27]LLaMA-7BCross Attn257224-38.450.935.5--
Qwen-VL [1]Qwen-7BResampler256448-59.378.835.267.1-
Qwen-VL-Chat [1]Qwen-7BResampler256448-57.578.238.968.2-
LLaVA-1.5 [35]Vicuna-7BMLP57733658.262.078.550.066.81.0×
LLaVA-1.5+SEC (ours)Vicuna-7BMLP+SEC25733659.663.278.952.869.61.7×
13B LLM
InstructBLIP [8]Vicuna-13BQ-Former32224-49.5-33.463.1-
BLIP-2 [29]Vicuna-13BQ-Former32224-41.041.019.561.0-
LLaVA-1.5 [35]Vicuna-13BMLP57733661.263.380.053.671.61.0×
LLaVA1.5+SEC (ours)Vicuna-13BMLP+SEC25733662.364.380.054.772.01.8×
", + "image_path": "ab4eade270e932acb2b264e8e6112de399977f63626d29c8b15ff839c83aad8f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 223, + 392, + 369, + 402 + ], + "lines": [ + { + "bbox": [ + 223, + 392, + 369, + 402 + ], + "spans": [ + { + "bbox": [ + 223, + 392, + 369, + 402 + ], + "type": "text", + "content": "Table 8. Results on General VQA tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 415, + 264, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 264, + 427 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 264, + 427 + ], + "type": "text", + "content": "effectiveness of SEC in extracting visual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 436, + 141, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 141, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 141, + 449 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "type": "text", + "content": "In this section, we present some of the ablation study results for SEC, and more results can be found in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 626 + ], + "type": "text", + "content": "Number of Vision Tokens in Each Clusters. The number of vision tokens has a significant impact on the performance and speed of the model. We thoroughly investigate the effect of the number of vision tokens on SECViT. As shown in Tab. 10, the number of vision tokens in each cluster greatly influences the model's performance. Specifically, in downstream dense prediction tasks, having too few tokens in each cluster leads to substantial performance degradation. When the number of tokens in each cluster is too large, the model's performance does not see a significant improvement, but its speed decreases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 167, + 647 + ], + "type": "text", + "content": "4.4. Visualization of SEC" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "To further understand the working mechanism of SEC, we visualize some clustering results for SECViT. As shown in Fig. 4, the left side presents the clustering results of vision tokens at different stages of the model. From the clustering results, we analyze that in the shallow layers, the model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 488 + ], + "type": "text", + "content": "distinguishes fine-grained features well, while in the deeper layers, it captures global semantic features effectively. The right side shows the Grad-CAM diagrams at different stages of the model, from which we can draw similar conclusions to the clustering results. More visualization results can be found in Appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 502, + 546, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 546, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 546, + 634 + ], + "type": "text", + "content": "Number of Vision Tokens Outputs by SEC. MLLM is quite sensitive to the number of vision tokens. We conduct a detailed exploration based on LLaVA-1.5 regarding the number of vision tokens output by SEC, as shown in Tab. 11. The first row represents the speed and performance of the original LLaVA-1.5 without using SEC. Compared to LLaVA-1.5, employing SEC effectively reduces the number of vision tokens and improves training efficiency. As the number of vision tokens decreases, the model's performance shows a slight decline, but its efficiency is further enhanced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 645, + 379, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 379, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 379, + 658 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "We propose a simple and straightforward clustering method for vision tokens—Semantic Equitable Clustering (SEC). This method assigns each token to a cluster by calculating the similarity between each token and a global token, and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4025" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 531, + 225 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 531, + 225 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 531, + 225 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 531, + 225 + ], + "type": "table", + "html": "
ModelLLMConnectorV-T NumResPOPEMMBMM-VetSpeed (↑)
7B LLM
MiniGPT-4 [60]Vicuna-7BResampler3222472.224.322.1-
mPLUG-Ow12 [57]LLaMA2-7BResampler32224-49.4--
LLaMA-AdapterV2 [14]LLaMA2-7BLLaMA-Adapter257224-41.031.4-
Shikra [5]Vicuna-7BMLP257224-58.8--
Qwen-VL [1]Qwen-7BResampler256448-38.2--
Qwen-VL-Chat [1]Qwen-7BResampler256448-60.6--
LLaVA-1.5 [35]Vicuna-7BMLP57733686.164.331.11.0×
LLaVA1.5+SEC (ours)Vicuna-7BMLP+SEC14533686.168.431.72.1×
13B LLM
MiniGPT-4 [60]Vicuna-13BResampler32224--24.4-
BLIP-2 [29]Vicuna-13BQ-Former3222485.3-22.4-
LLaVA-1.5 [35]Vicuna-13BMLP57733686.267.736.11.0×
LLaVA-1.5+SEC (ours)Vicuna-13BMLP+SEC14533686.469.237.32.2×
", + "image_path": "c1bba0b8bb52a32dc26933269128d730ddd2a9865c45bfa9fe9a360ff72e4cc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 243, + 274, + 361 + ], + "blocks": [ + { + "bbox": [ + 198, + 227, + 394, + 237 + ], + "lines": [ + { + "bbox": [ + 198, + 227, + 394, + 237 + ], + "spans": [ + { + "bbox": [ + 198, + 227, + 394, + 237 + ], + "type": "text", + "content": "Table 9. Results on benchmark designed for MLLMs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 243, + 274, + 361 + ], + "lines": [ + { + "bbox": [ + 60, + 243, + 274, + 361 + ], + "spans": [ + { + "bbox": [ + 60, + 243, + 274, + 361 + ], + "type": "table", + "html": "
V-T numParams (M)FLOPs (G)Throughput (imgs/s)AccAPbAPmmIoU
98152.5200482.747.843.047.2
196153.1172283.0 (+0.3)48.2 (+0.4)43.4 (+0.4)47.5 (+0.3)
64152.5194682.7 (+0.0)47.8 (+0.0)42.8 (-0.2)46.9 (-0.3)
49152.4210282.6 (-0.1)47.5 (-0.3)42.7 (-0.3)47.7 (-0.5)
24152.3218682.0 (-0.7)45.9 (-1.9)40.6 (-2.4)44.6 (-2.6)
", + "image_path": "c760fd9f8f0d078c2bb8be621b61cf32610c29ee8b18ee63cb67cebf90532cad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 399, + 287, + 467 + ], + "blocks": [ + { + "bbox": [ + 47, + 362, + 287, + 396 + ], + "lines": [ + { + "bbox": [ + 47, + 362, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 287, + 396 + ], + "type": "text", + "content": "Table 10. Effect of the number of vision tokens in each cluster. \"V-T num\" means the number of vision tokens in each cluster. The experiments are conducted based on SECViT-T." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 399, + 287, + 467 + ], + "lines": [ + { + "bbox": [ + 50, + 399, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 50, + 399, + 287, + 467 + ], + "type": "table", + "html": "
V-T numTime SpeedTextVQAGQAVQAv2POPEMM-Vet
576+121h1.0×58.262.078.586.131.1
288+114h1.5×60.1(+1.9)63.5(+1.5)78.9(+0.4)87.7(+1.6)33.2(+2.1)
256+113h1.6×59.6(+1.4)63.2(+0.3)78.6(+0.1)87.1(+1.0)32.7(+1.6)
192+111h1.9×57.7(-0.5)62.7(+0.7)78.4(-0.1)86.7(+0.6)32.1(+1.0)
144+110h2.1×56.8(-1.4)62.0(+0.0)78.0(-0.5)86.1(+0.0)31.7(+0.6)
", + "image_path": "739005363da79434f53381e97d751e222d3b9da27106c33b781ac1e713904cbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 287, + 502 + ], + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 502 + ], + "type": "text", + "content": "Table 11. Effect of the number of vision tokens outputs by SEC. \"V-T num\" means the number of vision tokens output by SEC. The experiments are conducted based on LLaVA-1.5 [36]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 658 + ], + "type": "text", + "content": "completes the whole clustering process in only one step. Our clustering method takes into account the semantic information contained in the tokens, and ensures an equal number of tokens in each cluster, facilitating efficient parallel processing on modern GPUs. Based on Semantic Equitable Clustering, we designed SECViT, a versatile vision backbone that achieves impressive results across various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Besides, SEC can also be conveniently applied to multimodal large language models (MLLM) to serve as a vision language connector and benefits the model's efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 162, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 162, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 162, + 683 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "This work is partially funded by Beijing Natural Science Foundation (4252054), Youth Innovation Promotion" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 242, + 544, + 630 + ], + "blocks": [ + { + "bbox": [ + 306, + 242, + 544, + 630 + ], + "lines": [ + { + "bbox": [ + 306, + 242, + 544, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 544, + 630 + ], + "type": "image", + "image_path": "dad77df489835b054acdb275ca0d0404b2aec373fd532a5ac9ebaf2e3f7b9f41.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 631, + 485, + 642 + ], + "lines": [ + { + "bbox": [ + 366, + 631, + 485, + 642 + ], + "spans": [ + { + "bbox": [ + 366, + 631, + 485, + 642 + ], + "type": "text", + "content": "Figure 4. Visualization for SEC." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 654, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 691 + ], + "type": "text", + "content": "Association CAS(Grant No.2022132), Beijing Nova Program(20230484276), and CCF-Kuaishou Large Model Explorer Fund (NO. CCF-KuaiShou 2024005)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4026" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 286, + 169 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 203 + ], + "type": "text", + "content": "[3] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 237 + ], + "type": "text", + "content": "[4] Kai Chen, Jiaqi Wang, Jiangmiao Pang, et al. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 271 + ], + "type": "text", + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic, 2023. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 272, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 286, + 304 + ], + "type": "text", + "content": "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. In ICLR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 306, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 286, + 327 + ], + "type": "text", + "content": "[7] MMSegmentation Contributors. Mmsegmentation, an open source semantic segmentation toolbox, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 329, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 286, + 384 + ], + "type": "text", + "content": "[8] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 286, + 407 + ], + "type": "text", + "content": "[9] Mingyu Ding, Bin Xiao, Noel Codella, et al. Davit: Dual attention vision transformers. In ECCV, 2022. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 408, + 286, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 286, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 286, + 441 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, et al. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 442, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 286, + 475 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 286, + 508 + ], + "type": "text", + "content": "[12] Qihang Fan, Huaibo Huang, Xiaogiang Zhou, and Ran He. Lightweight vision transformer with bidirectional interaction. In NeurIPS, 2023. 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 510, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 286, + 542 + ], + "type": "text", + "content": "[13] Qihang Fan, Huaibo Huang, Mingrui Chen, Hongmin Liu, and Ran He. Rmt: Retentive networks meet vision transformers. In CVPR, 2024. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 544, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 588 + ], + "type": "text", + "content": "[14] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, Hongsheng Li, and Yu Qiao. Llama-adapter v2: Parameter-efficient visual instruction model, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 622 + ], + "type": "text", + "content": "[15] SG-Former: Self guided Transformer with Evolving Token Reallocation. Sucheng ren, xingyi yang, songhua liu, xinchao wang. In ICCV, 2023. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "type": "text", + "content": "[16] Jianyuan Guo, Kai Han, Han Wu, Chang Xu, Yehui Tang, Chunjing Xu, and Yunhe Wang. Cmt: Convolutional neural networks meet vision transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "text", + "content": "[17] Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Visual attention network. arXiv preprint arXiv:2202.09741, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[18] Dongchen Han, Ziyi Wang, Zhuofan Xia, Yizeng Han, Yifan Pu, Chunjiang Ge, Jun Song, Shiji Song, Bo Zheng, and" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Gao Huang. Demystify mamba in vision: A linear attention perspective. In NeurIPS, 2024. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 545, + 129 + ], + "type": "text", + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 163 + ], + "type": "text", + "content": "[20] Ali Hatamizadeh, Hongxu Yin, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. In ICML, 2023. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 166, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 166, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 166, + 545, + 186 + ], + "type": "text", + "content": "[21] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask r-cnn. In ICCV, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "type": "text", + "content": "[22] Huaibo Huang, Xiaogiang Zhou, Jie Cao, Ran He, and Tieniu Tan. Vision transformer with super token sampling. In CVPR, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 224, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 256 + ], + "type": "text", + "content": "[23] Andrew Jaegle, Felix Gimeno, Andrew Brock, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver: General perception with iterative attention, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 258, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 300 + ], + "type": "text", + "content": "[24] Zi-Hang Jiang, Qibin Hou, Li Yuan, Daquan Zhou, Yujun Shi, Xiaojie Jin, Anran Wang, and Jiashi Feng. All tokens matter: Token labeling for training better vision transformers. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "text", + "content": "[25] Manjin Kim, Paul Hongsuck Seo, Cordelia Schmid, and Minsu Cho. Learning correlation structures for vision transformers. In CVPR, 2024. 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollár. Panoptic feature pyramid networks. In CVPR, 2019. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 373, + 545, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 427 + ], + "type": "text", + "content": "[27] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. In NeurIPS, 2024. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 461 + ], + "type": "text", + "content": "[28] Youngwan Lee, Jonghee Kim, Jeffrey Willette, and Sung Ju Hwang. Mpvit: Multi-path vision transformer for dense prediction. In CVPR, 2022. 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 464, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 506 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 509, + 545, + 542 + ], + "type": "text", + "content": "[30] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning, 2022. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 586 + ], + "type": "text", + "content": "[31] Siyuan Li, Zedong Wang, Zicheng Liu, Cheng Tan, Haitao Lin, Di Wu, Zhiyuan Chen, Jiangbin Zheng, and Stan Z. Li. Moganet: Multi-order gated aggregation network. In ICLR, 2024. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 643 + ], + "type": "text", + "content": "[32] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In International Conference on Learning Representations, 2022. 2, 4, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, and Kaiming He and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[34] Weifeng Lin, Ziheng Wu, Jiayu Chen, Jun Huang, and Lianwen Jin. Scale-aware modulation meet transformer. In ICCV, 2023. 5, 6" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4027" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 172 + ], + "type": "text", + "content": "[37] Kai Liu, Tianyi Wu, Cong Liu, and Guodong Guo. Dynamic group transformer: A general vision transformer backbone with dynamic group attention. In IJCAI, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 215 + ], + "type": "text", + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 216, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 287, + 237 + ], + "type": "text", + "content": "[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, et al. A convnet for the 2020s. In CVPR, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 238, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 287, + 281 + ], + "type": "text", + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 316 + ], + "type": "text", + "content": "[41] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng YAN. Inception transformer. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 316, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 337 + ], + "type": "text", + "content": "[42] Shitao Tang, Jiahui Zhang, Siyu Zhu, et al. Quadtree attention for vision transformers. In ICLR, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, et al. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 2, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 403 + ], + "type": "text", + "content": "[44] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[45] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "text", + "content": "[46] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvtv2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):1-10, 2022. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "text", + "content": "[47] Wenxiao Wang, Lu Yao, Long Chen, Binbin Lin, Deng Cai, Xiaofei He, and Wei Liu. Crossformer: A versatile vision transformer hinging on cross-scale attention. In ICLR, 2022. 1, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "text", + "content": "[48] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In CVPR, 2023. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[49] Yulin Wang, Rui Huang, Shiji Song, Zeyi Huang, and Gao Huang. Not all images are worth 16x16 words: Dynamic vision transformers with adaptive sequence length. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[50] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. arXiv preprint arXiv:2301.00808, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[51] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 487 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[52] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In ECCV, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[53] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 173 + ], + "type": "text", + "content": "[54] Chenglin Yang, Siyuan Qiao, Qihang Yu, et al. Moat: Alternating mobile convolution and attention brings strong vision models. In ICLR, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[55] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[56] Rui Yang, Hailong Ma, Jie Wu, Yansong Tang, Xuefeng Xiao, Min Zheng, and Xiu Li. Scalablevit: Rethinking the context-oriented generalization of vision transformer. In ECCV, 2022. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 330 + ], + "type": "text", + "content": "[57] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl: Modularization empowers large language models with multimodality, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[58] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In ICCV, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 408 + ], + "type": "text", + "content": "[59] Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. In CVPR, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[60] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 487 + ], + "type": "text", + "content": "[61] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson Lau. Bformer: Vision transformer with bi-level routing attention. In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4028" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_content_list.json b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa2a6bd6291ce739ffafa96800eeb6d75faa1413 --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_content_list.json @@ -0,0 +1,1447 @@ +[ + { + "type": "text", + "text": "Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity", + "text_level": 1, + "bbox": [ + 109, + 128, + 885, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sung Ju Lee Nam Ik Cho", + "bbox": [ + 366, + 203, + 630, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dept. of ECE & INMC, Seoul National University, Korea", + "bbox": [ + 269, + 238, + 727, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thomas11809@snu.ac.kr, nicho@snu.ac.kr", + "bbox": [ + 326, + 258, + 665, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at github.com/thomas11809/SFWMark.", + "bbox": [ + 89, + 342, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advancement of diffusion generative models [17, 25, 33, 35, 39, 41], their generated images are being increasingly used across various fields, including creative works, entertainment, and advertisement. In particular, the open-source release of large-scale language-image (LLI) models like Stable Diffusion [37] has led to an exponential growth", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bbcfca8fcbdc3cc4228a5a7344a6d4682b897c62e1a5eec1f2b6ad07fd5a977.jpg", + "image_caption": [ + "Figure 1. Summary of watermarking performance across different semantic watermarking methods, as detailed in Sec. 5.2. All methods follow the merged-in-generation scheme with no additional processing time. Verification is evaluated using TPR@1%FPR (True Positive Rate at $1\\%$ False Positive Rate), while identification is assessed by Identification Accuracy (Perfect Match Rate). The proposed approaches achieve the best balance between detection robustness and image fidelity." + ], + "image_footnote": [], + "bbox": [ + 516, + 308, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in script-based image generation and editing technologies, resulting in a surge of generative content. This has raised new concerns, such as the copyright of AI-generated content and the tracking of images created with malicious intent. As a solution to these problems, the technique of embedding invisible watermarks into content has been proposed.", + "bbox": [ + 511, + 638, + 906, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Digital content watermarking has been extensively studied through both classical signal processing techniques [5, 6, 12, 13, 15, 28, 38, 43, 44, 49] and deep learning-based approaches [1, 18, 19, 21, 26, 30, 42, 52, 57, 58]. Recently, watermarking techniques that directly intervene in the generation process of diffusion models [3, 7, 10, 20, 22, 27, 31, 36, 51, 54] have been explored; however, these approaches require additional setup or configuration, adding more complexity.", + "bbox": [ + 511, + 747, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several studies have explored embedding watermarks di", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18759", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rectly into the latent representation, eliminating the need for external models. These include a steganographic approach that constructs latent noise with stream cipher-randomized bits [50] and semantic watermarking methods that embed geometric patterns in the Fourier frequency domain of the latent representation [11, 48, 53]. Meanwhile, research on pixel-level perturbation-based watermarking has highlighted its vulnerability to regeneration attacks [55, 56], demonstrating that semantic watermarking serves as a more robust alternative against such attacks. However, the aforementioned semantic watermarking methods in the latent Fourier domain, which are the focus of this paper, suffer from degraded detection accuracy and generative quality due to their lack of frequency integrity preservation.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this context, we propose a novel semantic watermarking framework that enhances both detection robustness and image quality by preserving frequency integrity in the latent Fourier domain. Our method is named Hermitian Symmetric Fourier Watermarking (SFW), which ensures that watermark embeddings maintain statistical consistency with the latent noise distribution, leading to improved retrievability and stability in generative models. Additionally, we incorporate a center-aware embedding strategy that enhances robustness against cropping attacks by embedding watermarks in a spatially resilient region of the latent representation. We comprehensively evaluate our method across various attack scenarios, including signal processing distortions, regeneration attacks, and cropping attacks. Experimental results, as illustrated in Fig. 1, demonstrate that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while simultaneously maintaining superior generative quality, as evidenced by FID and CLIP score evaluations. Our contributions can be summarized as follows:", + "bbox": [ + 91, + 303, + 482, + 602 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Hermitian SFW to ensure frequency integrity, leading to improved watermark detection and generative quality.", + "- We introduce center-aware embedding, which significantly enhances robustness against cropping attacks.", + "- We present extensive evaluations demonstrating that our approach outperforms existing baselines in detection accuracy and generative quality." + ], + "bbox": [ + 91, + 604, + 482, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 739, + 238, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Digital Watermarking. Digital watermarking aims to achieve a balance between high embedding capacity and visual quality when inserting invisible watermarks while also ensuring robustness against various attacks [45]. This field originally began with the adoption of classical signal processing techniques. Embedding the watermark in the spatial domain is a straightforward and intuitive watermark insertion [43, 44], but it tends to be less robust against attacks such as filtering or compression. On the other hand,", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79831f3006d8a799d55b67df55ba2fe6115882fb6449f42436d3e6ef4d026d16.jpg", + "image_caption": [ + "Figure 2. Overview of the semantic watermarking process in the latent Fourier domain using the merged-in-generation scheme." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "studies utilizing the frequency domain have demonstrated resilience against JPEG compression by embedding watermarks in the low and middle frequency bands [2, 5, 6, 12, 13, 15, 28, 32, 38, 47, 49].", + "bbox": [ + 511, + 325, + 903, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Watermarks for Latent Diffusion Models. Latent diffusion models (LDMs), such as Stable Diffusion [37], conduct the diffusion process in a lower-dimensional latent space rather than directly on high-resolution images. Recent research has focused on techniques for integrating watermarks into the latent diffusion process to ensure traceability and robustness. However, many of these approaches rely on model fine-tuning [3, 10, 20, 22, 27] or require separately trained encoders and decoders [7, 10, 14, 20, 22, 23, 27, 31, 36, 51, 54] to facilitate watermark embedding and detection. These dependencies constrain the flexibility of watermarking, limiting their applicability across diverse models.", + "bbox": [ + 511, + 387, + 905, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies have explored embedding semantic watermarks in the Fourier domain of latent vectors using a merged-in-generation scheme. Wen et al. [48] introduced a tree-ring-shaped watermark constructed with Gaussian-distributed radii, while Ci et al. [11] proposed a modified pattern with high-energy signed constant rings for watermark identification, along with an additional random noise key embedded in a separate channel. On the other hand, Zodiac [53] followed a post-hoc approach, embedding a tree-ring pattern into generated images through multiple iterations of latent vector optimization and diffusion-based synthesis. However, in addition to its high computational cost, this method suffers from low practicality, as it relies on extensive linear interpolation between the original and optimized images to artificially improve visual quality.", + "bbox": [ + 511, + 584, + 905, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In these methods, when performing inverse Fourier transform, the imaginary component in the spatial domain is discarded, leading to a distorted frequency representation. As a result, the real component of the key region retains only partial information, while the imaginary component", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18760", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc28e9aa714e81edfecb78c3e654df8afd03e20873e931db02496f0b9bbf1ed4.jpg", + "image_caption": [ + "Figure 3. Examples of various semantic watermarking patterns." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is almost entirely lost, creating an empty key region in the frequency domain. Since detection relies on analyzing this incomplete key region, the process is inherently limited, resulting in suboptimal retrieval performance.", + "bbox": [ + 89, + 316, + 482, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the overall pipeline of semantic watermarking with the merged-in-generation scheme. The embedding process begins with the Fourier transform applied to the latent noise, generating the latent Fourier representation. A watermark key is then embedded into a designated key region, followed by the inverse Fourier transform, producing the watermarked latent vector. Finally, text-guided image generation synthesizes the watermarked image. For detection, the process starts with a clean or attacked image, from which the latent query is obtained via DDIM inversion [41]. The presence of a watermark is determined by analyzing the query key region in the latent representation. Further details on detection tasks and evaluation metrics are provided in Sec. 3.1.", + "bbox": [ + 89, + 378, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 607, + 228, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Formulation", + "text_level": 1, + "bbox": [ + 89, + 633, + 264, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The performance of watermark detection is evaluated following RingID's formulation for watermark verification and identification tasks [11]. The metric used to measure the distance $d$ is the $L_{1}$ distance, calculated specifically for the key region where the watermark is embedded.", + "bbox": [ + 89, + 656, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Verification. The objective of verification is to determine whether a watermark is present in an image by analyzing the distance between the reference key and the key region of the query in the latent Fourier domain. Let $\\hat{w}$ denote the watermarked latent Fourier key, and $\\hat{u}$ denote the unwatermarked latent Fourier null key. Verification is based on comparing the distances between the reference key $w$ and the watermarked/unwatermarked keys, i.e., $d(\\hat{w}, w) \\neq d(\\hat{u}, w)$ . Performance is assessed using statistical metrics derived from the ROC curve, considering different distance thresholds.", + "bbox": [ + 89, + 733, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identification. In identification, given that a watermark", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is already embedded, the task is to accurately determine the embedded information. This is achieved by computing the distance between the watermarked key $\\hat{w}$ and multiple reference keys $w_{i}$ . Performance is evaluated based on the accuracy of the estimated message index, defined as $\\hat{i} = \\arg \\min_{i}d(\\hat{w},w_{i})$ .", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Fourier Considerations for Real Latent Noise", + "text_level": 1, + "bbox": [ + 511, + 190, + 897, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces the mathematical properties of the Fourier domain and highlights important considerations when embedding semantic watermarks.", + "bbox": [ + 511, + 212, + 903, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hermitian Symmetry for Real Signals. To obtain a real-valued signal after inverse Fourier transform, modifications in the frequency domain must maintain Hermitian symmetry about the DC center:", + "bbox": [ + 511, + 257, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF [ M - k, N - l ] = \\overline {{F [ k , l ]}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 325, + 903, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F$ represents the discrete Fourier transform, $M$ and $N$ denote the dimensions of the signal in the spatial domain along the row and column axes, respectively. Failure to preserve this symmetry introduces undesired imaginary components in the spatial domain, which are incompatible with the real-valued latent noise required for diffusion models.", + "bbox": [ + 511, + 352, + 905, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Impact of Ignoring Hermitian Symmetry. In existing baselines [11, 48, 53], this issue is simply addressed by discarding the imaginary component, resulting in a loss of frequency integrity. As illustrated in Fig. 3a, this disruption affects both real and imaginary components of the original watermark pattern. While the real component deviates from its intended structure due to frequency distortion, the imaginary component suffers even greater degradation, often becoming entirely empty. This significantly limits detection performance by preventing the effective use of frequency information in the imaginary domain. The impact of the frequency loss is further examined in Sec. 5.3.1.", + "bbox": [ + 511, + 443, + 905, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preservation of Gaussianity. A real Gaussian noise signal in the spatial domain transforms into a complex Gaussian noise signal in the frequency domain while maintaining its statistical properties:", + "bbox": [ + 511, + 625, + 905, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf [ m, n ] \\sim \\mathcal {N} (0, \\sigma^ {2}) \\Rightarrow F [ k, l ] \\sim \\mathcal {C N} (0, M N \\sigma^ {2}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 691, + 903, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ represents the real-valued Gaussian noise signal in the spatial domain. When embedding watermarks in the frequency domain, some perturbation is inevitably introduced, altering the original distribution of complex Gaussian noise. However, if Hermitian symmetry is not preserved, the inverse Fourier transform produces incomplete noise characteristics in the spatial domain, as imaginary components must be discarded. This disrupts the statistical consistency of the latent noise, leading to a deviation from the expected real Gaussian distribution and potential degradation in generative quality. In contrast, embedding with Hermitian symmetry better retains the statistical structure of the latent", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ab9bf4e0e2b5982d3489bb7c83734271773944aa4c33f486db60e72de68a140a.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed framework and qualitative results. (Left) The key components of our approach: Symmetric Fourier Watermarking (SFW) (blue region) and Center-Aware Embedding Strategy (green region). (Right) Qualitative results of Tree-Ring, RingID, HSTR, and HSQR. Notably, RingID exhibits visible ring-like artifacts, highlighting that its high-energy pattern disrupts generative quality, unlike other merged-in-generation semantic watermarking methods that achieve a better balance between robustness and image fidelity." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 509, + 258 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b50f6570acc8fbdaa56c28b3765c7d40fc0eb61030aa1612eec54a8483f7aa4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 88, + 905, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noise, ensuring that the transformed spatial-domain signal remains closer to a real Gaussian distribution. By preserving these statistical properties, the initialization process in diffusion models becomes more stable, ultimately enhancing generative performance, as observed in Sec. 5.2.", + "bbox": [ + 89, + 349, + 485, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 441, + 191, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Hermitian Symmetric Fourier Watermark", + "text_level": 1, + "bbox": [ + 89, + 465, + 452, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed Hermitian SFW refers to a watermark designed to satisfy the Hermitian symmetry introduced in the previous section. As shown in the left side of Fig. 4, SFW allows free usage of the half-region in the frequency domain, while the remaining half is restricted by the symmetry condition. Certain constraints must be satisfied when embedding a pattern in the free half-region under a semantic watermarking scheme: the imaginary part at the DC center must be zero, and if the signal dimensions are even, the imaginary components of the points corresponding to the Nyquist frequency must also be zero. In cases where both frequency axes have even dimensions as in this study, the points where the imaginary part must be zero are $(0,0)$ , $\\left(\\frac{M}{2},0\\right)$ , $(0,\\frac{N}{2})$ , and $\\left(\\frac{M}{2},\\frac{N}{2}\\right)$ . Instead of directly modifying the watermark pattern to comply with locations where the imaginary part must be zero, a more effective strategy is to embed it while avoiding the DC axis as much as possible, as will be introduced in Sec. 4.3.2", + "bbox": [ + 89, + 488, + 483, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Center-Aware Embedding Strategy", + "text_level": 1, + "bbox": [ + 89, + 771, + 397, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing semantic watermarking baselines [11, 48, 53] embed patterns by applying the Fourier transform to the full spatial matrix of the latent vector. However, such methods are vulnerable to spatial attacks, particularly cropping, which can lead to the loss of watermark information and reduced detection performance. To address this issue, we propose a center-aware embedding strategy, which applies", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the Fourier transform only to the central area of the spatial domain before embedding. Specifically, for a latent vector with a signal dimension of 64, we utilize the central $44 \\times 44$ space. As demonstrated in Sec. 5.3.2, this design significantly improves robustness against cropping attacks on various scales.", + "bbox": [ + 511, + 349, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Integrating SFW and Center-Aware Design", + "text_level": 1, + "bbox": [ + 511, + 450, + 880, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3.1. Application to the Baseline", + "text_level": 1, + "bbox": [ + 511, + 474, + 746, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3a, the Tree-Ring [48] baseline suffers from frequency loss. To mitigate this, we impose the Hermitian symmetry condition directly on the watermark patterns. By integrating SFW with the center-aware embedding strategy, we refine the baseline into Hermitian Symmetric Tree-Ring (HSTR). HSTR not only enhances detection performance by fully utilizing the imaginary components of the latent Fourier domain but also improves image generation quality by restoring frequency integrity.", + "bbox": [ + 511, + 493, + 906, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3.2. Hermitian Symmetric QR code", + "text_level": 1, + "bbox": [ + 511, + 638, + 777, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce a novel approach that extends SFW beyond baselines to QR codes, which are widely recognized for their high versatility and robust error correction capabilities [16].", + "bbox": [ + 511, + 657, + 905, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Embedding.** As illustrated in Fig. 3b, the Hermitian Symmetric QR Code (HSQR) watermark is constructed by splitting the QR code in half and embedding each part separately into the real and imaginary components of the free half-region in the Fourier domain. The binary pattern of the QR code is embedded using the following formulation:", + "bbox": [ + 511, + 718, + 905, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {H S Q R} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 1 \\\\ - | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 0, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 834, + 906, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where:", + "bbox": [ + 514, + 886, + 563, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d5f00951289a17b786af1fca0e97da387fb639b201d62c5587a010845934e893.jpg", + "table_caption": [ + "Table 1. Verification performance of different watermarking methods under various attacks. Bit Accuracy is used for bitstream-based methods (DwtDct, DwtDctSvd, RivaGAN, S.Sign), while TPR@1%FPR is used for semantic methods. Best performances are highlighted. Our methods show superior detection accuracy and robustness against signal processing distortions, regeneration, and cropping attacks." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.8630.5720.5220.5160.6770.8590.5320.5230.5210.5190.7290.8100.637
DwtDctSvd1.0000.5550.4730.6021.0001.0000.7840.6480.5960.6440.7440.8610.742
RivaGAN0.9990.8620.9860.8210.9980.9690.9340.5700.5520.6080.9910.9950.857
S.Sign.0.9950.8940.9780.8060.9110.7210.8380.7170.7150.4780.9870.9910.836
Tree-Ring0.9570.4630.9000.5480.9340.4120.8150.5090.5360.5430.5090.7340.655
Zodiac0.9980.8430.9980.9730.9980.8800.9970.9440.9580.9720.9890.9950.962
HSTR (ours)1.0000.8991.0000.9941.0000.8060.9990.9730.9820.9971.0001.0000.971
RingID1.0000.9881.0001.0001.0000.9871.0000.9921.0001.0001.0001.0000.997
HSQR (ours)1.0000.9911.0001.0001.0000.9831.0000.9921.0001.0001.0001.0000.997
SD-PromptsDwtDct0.8190.5570.5160.5060.6850.8220.5300.5130.5120.5090.7230.7940.624
DwtDctSvd1.0000.5370.4590.6100.9990.9980.8590.6590.6200.6230.7430.8600.747
RivaGAN0.9910.8230.9630.8100.9880.9610.9150.5720.5350.5670.9800.9830.841
S.Sign.0.9940.8990.9670.7690.8880.7420.8090.6770.6710.4930.9830.9900.824
Tree-Ring0.9440.4710.8940.4660.9120.4230.8020.5090.5140.5430.4690.7490.641
Zodiac0.9980.7480.9990.9790.9990.9031.0000.9400.9750.9580.9940.9960.957
HSTR (ours)1.0000.7421.0000.9901.0000.8501.0000.9830.9870.9991.0001.0000.963
RingID1.0000.9721.0001.0001.0000.9881.0000.9961.0001.0001.0001.0000.996
HSQR (ours)1.0000.9551.0000.9991.0000.9921.0000.9960.9991.0001.0001.0000.995
DiffusionDBDwtDct0.8420.5630.5150.5090.6720.8290.5260.5130.5140.5120.7230.8010.627
DwtDctSvd0.9980.5580.4630.5930.9970.9950.8300.6580.6080.6210.7420.8600.744
RivaGAN0.9870.8390.9600.7900.9850.9370.8930.5530.5180.5560.9740.9790.831
S.Sign.0.9900.8900.9670.7870.8890.7260.8190.6900.6870.4960.9810.9860.826
Tree-Ring0.9400.4870.8890.4340.9040.3920.7990.4540.5030.4540.4990.7150.622
Zodiac0.9920.7520.9880.9330.9880.8340.9840.9110.9260.9030.9710.9850.931
HSTR (ours)0.9990.7920.9960.9810.9960.7920.9910.9680.9690.9891.0001.0000.956
RingID1.0000.9891.0001.0001.0000.9631.0000.9950.9991.0001.0001.0000.995
HSQR (ours)1.0000.9771.0000.9991.0000.9740.9990.9970.9991.0001.0001.0000.995
", + "bbox": [ + 98, + 143, + 897, + 506 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $x$ denotes the coordinates in the QR code domain.", + "- $\\tilde{x}$ represents the corresponding coordinates in the embedding region of the Fourier domain. Since the QR code is split into two halves, each half is mapped to either the real or imaginary component.", + "- $c \\in \\{\\mathrm{Re}, \\mathrm{Im}\\}$ indicates whether the embedding occurs in the real or imaginary part.", + "- $F(\\tilde{x}, c)$ is the real-valued amplitude of the complex Gaussian noise in the Fourier domain, where a complex number is expressed as $F_{\\mathrm{Re}} + iF_{\\mathrm{Im}}$ , making both real and imaginary amplitudes individually real-valued." + ], + "bbox": [ + 89, + 531, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To maintain symmetry while avoiding numerical instability, the embedding region is positioned one pixel to the right of the vertical DC axis. Additionally, to increase the amount of embedded information and statistically reduce errors, each QR code cell is represented by multiple pixels arranged in a square pattern, e.g., $2 \\times 2$ pixels.", + "bbox": [ + 89, + 698, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection. The detection process for semantic watermarking is performed by computing the L1 distance. In HSQR, the ground truth QR binary pattern is converted into a signed Boolean representation as follows:", + "bbox": [ + 89, + 790, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q R} ^ {*} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 1 \\\\ - \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 0, \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 863, + 482, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Lambda$ is a fixed amplitude for binary encoding.", + "bbox": [ + 511, + 531, + 841, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 564, + 643, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 590, + 718, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Watermarks. The Tree-Ring [48] watermark is embedded into channel 3 with a radius of 14 and HSTR adopts the same watermark parameters as its baseline. For RingID [11], we follow the original settings, embedding a tree-ring pattern with a radius range of 3-14 in channel 3, along with a Gaussian noise key in channel 0. Both SFW methods follow RingID in utilizing a Gaussian noise key in channel 0. For HSQR, we use a version 1 QR code $(21 \\times 21$ cells) with a cell size of 2 pixels and error correction level H, capable of encoding up to 72 bits. The QR code is embedded in channel 3, and the encoding amplitude $\\Lambda$ is set to 45, corresponding to the standard deviation of the real and imaginary components in the Fourier domain of $64 \\times 64$ normal Gaussian latent vector $(\\approx \\sqrt{64^2 / 2})$ . Note that these experiments are conducted on $512 \\times 512$ images. At higher resolutions, a larger embedding space allows for increased capacity through higher QR code versions that encode more bits, as well as the use of stronger watermarking parameters, enabling greater robustness and adaptability. Additionally,", + "bbox": [ + 511, + 613, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f11de02277708da5e7df5b4593c814afaf2b14a2eb23cbe4fbec9d918a1d5525.jpg", + "table_caption": [ + "Table 2. Identification accuracy for different watermarking methods, evaluated across multiple attack conditions. Perfect Match Rate is used for bitstream-based methods. The best performance for each item is highlighted with shading. HSQR achieves state-of-the-art performance, while HSTR consistently outperforms other Gaussian radius-based semantic baselines, particularly under cropping attacks." + ], + "table_footnote": [], + "table_body": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.4660.0440.0000.0000.0380.4420.0000.0000.0000.0000.0000.0000.083
DwtDctSvd1.0000.0440.0190.0000.9990.9980.0370.0000.0000.0030.0000.0000.258
RivaGAN0.9740.2600.7720.0230.9610.6860.3480.0000.0000.0000.8520.9090.482
S.Sign.0.8730.1770.5630.0000.0360.0100.0070.0000.0000.0000.7090.8020.265
Tree-Ring0.3030.0870.2070.0720.2560.0300.1620.0830.0720.0540.0090.0330.114
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.0000.7140.9990.8860.9980.4600.9720.8330.8310.9711.0001.0000.889
RingID1.0000.8751.0000.9751.0000.9190.9960.9780.9700.9980.8740.9780.964
HSQR (ours)1.0000.9581.0000.9941.0000.9010.9990.9800.9870.9991.0001.0000.985
SD-PromptsDwtDct0.2850.0240.0000.0000.0170.2760.0000.0000.0000.0000.0000.0000.050
DwtDctSvd0.9930.0280.0110.0000.9820.9790.0850.0000.0000.0070.0000.0000.257
RivaGAN0.8780.2130.6130.0090.8570.6570.3040.0000.0000.0000.7360.7680.420
S.Sign.0.8130.2630.4200.0000.0210.0150.0060.0000.0000.0000.5760.7160.236
Tree-Ring0.2880.0940.1890.0510.2350.0340.1590.0790.0760.0560.0120.0410.110
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.000
RingID1.0000.8851.0000.9760.9980.8860.9930.9800.9730.9950.8760.9810.962
HSQR (ours)1.0000.9301.0000.9941.0000.9420.9990.9910.9971.0001.0001.0000.988
DiffusionDBDwtDct0.3570.0370.0000.0000.0340.3200.0000.0000.0000.0000.0000.0000.062
DwtDctSvd0.9900.0360.0190.0000.9750.9590.0810.0000.0000.0010.0000.0000.255
RivaGAN0.8580.2130.6250.0200.8480.6150.2210.0000.0000.0000.7560.7800.411
S.Sign.0.7980.2070.4720.0000.0270.0050.0050.0000.0000.0000.6430.7380.241
Tree-Ring0.2800.0950.1900.0590.2330.0370.1450.0810.0720.0500.0130.0390.108
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)0.9960.7210.9920.8540.9890.5630.9580.8300.8210.9520.9960.9960.889
RingID1.0000.8951.0000.9470.9960.8710.9920.9680.9580.9900.8750.9840.956
HSQR (ours)1.0000.9541.0000.9881.0000.9060.9980.9820.9910.9941.0001.0000.984
", + "bbox": [ + 98, + 143, + 897, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the post-hoc semantic watermarking method, Zodiac [53], is included in our comparative experiments.", + "bbox": [ + 89, + 531, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Setup and Datasets. The diffusion model used for the experiments is Stable Diffusion v2-1-base [37], configured to generate images at a resolution of $512 \\times 512$ with a CFG scale of 7.5. Both generation and inversion steps of the DDIM scheduler [41] are set to 50. The datasets used for image generation include 5,000 captions from the MS-COCO-2017 training set [29], 1,001 sampled prompts from the 2M subset of DiffusionDB [46], and 8,192 prompts from the Stable-Diffusion-Prompts test set [40]. For watermark detection, the number of watermarked images used for each dataset is set to 1,000. Specifically, verification is conducted on 1,000 pairs of watermarked and unwatermarked images, while identification is performed on 1,000 watermarked images.", + "bbox": [ + 88, + 564, + 482, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Attack Methods. To evaluate the robustness of our method against various distortions, we apply 11 different attacks, categorized into three major types.", + "bbox": [ + 89, + 777, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Signal Processing Attacks: These include brightness adjustment (ranging from 0 to 7), contrast adjustment (factor of 0.5), JPEG compression (quality factor of 25), Gaussian blur (radius of 5), Gaussian noise ( $\\sigma = 0.05$ ), and BM3D denoising ( $\\sigma = 0.1$ ).", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Regeneration Attacks: We apply two VAE-based image compression models, Bmshj18 [4] (VAE-B) and Cheng20 [8] (VAE-C), both at quality level 3, as well as a diffusion-based regeneration attack by Zhao et al. [56] (referred to as Diff.) using 60 denoising steps.", + "- Cropping Attacks: We apply center crop (C.C.) with a crop scale of 0.5 and random crop (R.C.) with a crop scale of 0.7, where the crop scale represents the ratio of the cropped image area to the original image." + ], + "bbox": [ + 511, + 531, + 906, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The $L_{1}$ distance is computed only within the key region of the complex Fourier domain of latent noise, and unless otherwise specified, 2,048 keys are used for identification accuracy. For verification, semantic watermarking methods use True Positive Rate at $1\\%$ False Positive Rate (TPR@1%FPR) as the evaluation metric. To ensure a fair comparison with bitstream-based approaches such as DwtDct, DwtDctSvd [12], RivaGAN [52], and Stable Signature (S.Sign.) [22], we measure Bit Accuracy for these methods. For identification, where the goal is to match the exact message, Perfect Match Rate is used as the metric for bitstream-based approaches. To evaluate the generation quality, FID [24] is measured between watermarked images and 5,000 MS-COCO ground truth images. We also calculate the CLIP score [34] using the OpenCLIP-ViT/G", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18764", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4074b3b3a96b3c74136a639019085752fdaf96fe7d52348bdf469659caa65a0b.jpg", + "table_caption": [ + "Table 3. Generative quality evaluation of watermarking methods based on FID (MS-COCO ground truth) and CLIP score. The best performance for each item is highlighted with shading, while bold text specifically marks the low CLIP score in RingID. Our proposed methods preserve frequency integrity, achieving the best balance between watermark robustness and generative performance, whereas RingID introduces visible artifacts, compromising perceptual quality. Vrf. and Idf. denote the average detection performance in verification and identification tasks, respectively." + ], + "table_footnote": [], + "table_body": "
Semantic MethodsFID ↓CLIP ↑Vrf.Idf.
Merged in GenerationTree-Ring26.4180.3260.6550.114
RingID27.0520.3240.9970.964
HSTR (ours)25.0620.3290.9710.889
HSQR (ours)24.8950.3300.9970.985
", + "bbox": [ + 109, + 224, + 460, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model [9] to determine how closely the generated image aligns with the input prompt.", + "bbox": [ + 89, + 324, + 482, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 89, + 364, + 334, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we compare the detection performance in both verification and identification tasks, as well as the generation quality of different watermarking methods.", + "bbox": [ + 89, + 386, + 482, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Enhanced Detection Robustness. According to Tab. 1, our proposed methods consistently achieve top-tier detection performance across various attacks. A key observation is that non-semantic watermarking methods [12, 22, 52] exhibit significant degradation under regeneration attacks, highlighting their vulnerability. Among Gaussian radius-based pattern methods (Tree-Ring, Zodiac, and HSTR), HSTR outperforms corresponding baselines in most cases, demonstrating its effectiveness in enhancing detection robustness while maintaining the fundamental tree-ring structure. In contrast, RingID leverages high-energy signed constant ring patterns, securing strong detection performance. However, as acknowledged by the authors, this method introduces noticeable ring-like artifacts in generated images, disrupting the balance between watermark robustness and generative quality. A detailed quantitative evaluation of this trade-off is presented in the following section. Meanwhile, Zodiac suffers from time-consuming processing, taking several minutes to apply the watermark—requiring 7.36 minutes per image on MS-COCO dataset—which further limits its practicality in real-world applications. In contrast, our methods following the merged-in-generation scheme introduce no additional processing time, ensuring seamless watermark embedding.", + "bbox": [ + 89, + 431, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, we compare identification results, which reveal an even greater disparity in detection performance across different watermarking methods. Gaussian radius-based pattern methods, except for HSTR, perform poorly in identification tasks, with Zodiac—designed solely for verification—failing entirely across all scenarios (achieving zero identification accuracy). Further analysis on scalability", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "under different message capacities is provided in Sec. 5.3.3. Notably, our methods exhibit strong resilience against cropping attacks in both tasks, reinforcing the robustness of center-aware embedding. HSQR achieves state-of-the-art identification accuracy across all datasets, further establishing its dominance in watermark retrieval.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Balance with Generative Quality. Tab. 3 presents the generative quality of different semantic watermarking methods following the merged-in-generation scheme, evaluated using FID and CLIP score. RingID, which deviates from a Gaussian distribution by embedding high-energy perturbations, exhibits the worst generative performance, particularly reflected in its low CLIP score. This decline is attributed to the noticeable ring-like artifacts, as discussed earlier. In contrast, our proposed methods, which preserve frequency integrity via SFW, achieve the top two FID scores, demonstrating a better balance between watermark robustness and generative quality.", + "bbox": [ + 511, + 181, + 903, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expanding Frequency Utilization in Latent Watermarking. Traditional image-domain watermarking methods prioritize low-mid frequency embedding to resist compression and filtering attacks. This approach has also been adopted in latent diffusion-based semantic watermarking methods [11, 48, 53] without explicitly considering the differences between pixel-space and latent-space transformations. However, our findings suggest that such frequency constraints may not be necessary in the latent space, where watermark retrieval involves latent encoding and DDIM inversion, altering the impact of frequency perturbations. This distinction is evident in HSQR, which achieves state-of-the-art detection performance while utilizing nearly the entire frequency spectrum. Instead of relying on specific frequency bands, HSQR leverages structured statistical encoding, distributing watermark information across multiple pixels ( $2 \\times 2$ per QR cell). This redundancy enhances robustness against distortions, enabling effective retrieval even with broader frequency usage. These results indicate that latent-space watermarking is not bounded by traditional low-mid frequency constraints. Future semantic watermarking strategies can benefit from statistically robust encoding methods, allowing for effective watermark retrieval across a wider frequency range without compromising detection accuracy.", + "bbox": [ + 511, + 364, + 903, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 753, + 666, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. Impact of SFW on Detection Performance", + "text_level": 1, + "bbox": [ + 511, + 775, + 851, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the impact of Hermitian SFW by comparing detection performance across four cases (Tab. 4). Without SFW, using both real and imaginary components for detection (Case A) results in the lowest performance due to frequency degradation, while restricting detection to the real component (Case B) improves accuracy. In contrast, applying SFW (Cases C and D) preserves frequency integrity,", + "bbox": [ + 511, + 794, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18765", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/108a8b14af9ed889422ea4a061157f4d42af96d64a4cc11791fc617421af6b56.jpg", + "table_caption": [ + "Table 4. Ablation study on detection performance based on frequency integrity ( $\\sqrt{}$ or $\\times$ ) and the number of detection region usage (1: real only, 2: real & imaginary). Vrf. and Idf. represent average detection performance in verification (TPR@1% FPR) and identification (accuracy), respectively. $\\Delta L_1^*$ denotes the normalized $\\Delta L_1$ metric, indicating detection effectiveness." + ], + "table_footnote": [], + "table_body": "
CaseMethodsFreq. Int.# Det.Vrf.Idf.ΔL1*↑
ATree-Ring×20.6530.1140.232
BTree-Ring×10.8050.4160.368
CHSTR (ours)✓10.9360.7750.471
DHSTR (ours)✓20.9710.8890.476
", + "bbox": [ + 93, + 183, + 478, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56d75d67a6ba238f9798ac25f3ba8dde7f8e62510bcad250ae2eb3872e8feb0a.jpg", + "image_caption": [ + "Figure 5. Identification accuracy under center crop and random crop attacks at different crop scales. HSTR and HSQR maintain higher accuracy compared to RingID, demonstrating improved robustness against cropping." + ], + "image_footnote": [], + "bbox": [ + 91, + 268, + 287, + 415 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d4e78eb6f8b84fdb847e3363e8f009b0c427e3a6510d7addd8a6827ca628aae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 268, + 480, + 414 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf4440f4305343d37f581f8f612d9cbd603c4ac20fff19d24a148a5be3abc9c2.jpg", + "image_caption": [ + "Figure 6. Identification accuracy across watermark message capacities. HSQR remains nearly perfect, while RingID and HSTR degrade at higher capacities. Tree-Ring and Zodiac fail to scale effectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 500, + 480, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "leading to a significant boost in detection performance. Notably, HSTR with SFW and full complex detection (Case D) achieves the highest accuracy, demonstrating that leveraging both frequency components maximizes retrieval effectiveness. These results confirm that SFW enables robust semantic watermarking by maintaining frequency integrity and fully utilizing the Fourier domain for detection.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Robustness to Cropping Attacks", + "text_level": 1, + "bbox": [ + 511, + 90, + 785, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the robustness of center-aware embedding against center crop and random crop attacks, measuring identification accuracy across different crop scales (Fig. 5). As crop scale decreases, accuracy declines for all methods. However, RingID exhibits a steeper drop, indicating higher vulnerability to cropping. In contrast, HSTR and HSQR degrade more gradually, demonstrating improved resilience. While extreme cropping (scale 0.2) significantly affects all methods, our center-aware design consistently outperforms RingID, confirming its effectiveness in preserving watermark information under cropping distortions.", + "bbox": [ + 511, + 109, + 906, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3. Impact of Capacity on Identification", + "text_level": 1, + "bbox": [ + 511, + 282, + 812, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate identification accuracy across different watermark message capacities (64 to 8,192) for five methods (Fig. 6). Zodiac's performance collapses, while Tree-Ring exhibits rapid performance degradation, becoming nearly unusable at higher capacities. RingID, HSTR, and HSQR remain robust, maintaining $80\\%+$ accuracy, though HSTR declines faster, and RingID starts degrading from 2,048 onward. HSQR demonstrates the highest scalability, retaining near-perfect accuracy even at the largest capacity, confirming its superior robustness in high-capacity scenarios.", + "bbox": [ + 511, + 301, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 467, + 633, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Hermitian SFW, a novel approach to semantic watermarking in the latent diffusion model framework. Unlike existing methods that fail to preserve frequency integrity, our approach ensures that watermark embeddings maintain a consistent statistical structure in the latent noise distribution. This is achieved through Hermitian symmetry enforcement, which preserves frequency components and enhances detection and generation quality. Additionally, we have proposed center-aware embedding, which significantly improves robustness against cropping attacks by strategically placing watermarks in a spatially resilient region of the latent representation. Through comprehensive experiments, we demonstrated that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while also maintaining superior image generation quality, as shown by FID and CLIP scores. Our study highlights the importance of frequency integrity in Fourier-based watermarking and challenges the assumption that semantic watermarking must be confined to low-mid frequency bands. Experimental results confirm that a properly structured frequency-domain watermark can be effectively embedded and retrieved across the entire frequency spectrum without compromising generative quality. Future work includes exploring adaptive embedding strategies to further enhance robustness against adversarial attacks and extreme distortions, as well as extending our method to more diverse generative architectures beyond LDMs.", + "bbox": [ + 511, + 492, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18766", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the 2024 AI Semiconductor Application/Demonstration Support Program of the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) with Markany Co., Ltd. as the lead organization, and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2025.", + "bbox": [ + 89, + 113, + 485, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 252 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 1", + "[2] Dhruv Arya. A survey of frequency and wavelet domain digital watermarking techniques. 2", + "[3] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. Promark: Proactive diffusion watermarking for causal attribution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10802-10811, 2024. 1, 2", + "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 6", + "[5] Mauro Barni, Franco Bartolini, Vito Cappellini, and Alessandro Piva. A dct-domain system for robust image watermarking. Signal processing, 66(3):357-372, 1998. 1, 2", + "[6] Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet-based watermarking through pixel-wise masking. IEEE transactions on image processing, 10(5): 783-791, 2001. 1, 2", + "[7] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 1, 2", + "[8] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 6", + "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 7", + "[10] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 2", + "[11] Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. arXiv preprint arXiv:2404.14055, 2024. 2, 3, 4, 5, 7" + ], + "bbox": [ + 93, + 261, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 1, 2, 6, 7", + "[13] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE TIP, 6(12):1673-1687, 1997. 1, 2", + "[14] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for data copyright protection against generative diffusion models. ACM SIGKDD Explorations Newsletter, 26(2): 60-75, 2025. 2", + "[15] DEJEY and RS Rajesh. An improved wavelet domain digital watermarking for image protection. International journal of wavelets, multiresolution and information processing, 8(01): 19-31, 2010. 1, 2", + "[16] Denso Wave Incorporated. QR Code Essentials. Denso Wave, 2011. Available at: https://www.qrcode.com/en/about/standards.html.4", + "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[18] Han Fang, Dongdong Chen, Qidong Huang, Jie Zhang, Zehua Ma, Weiming Zhang, and Nenghai Yu. Deep template-based watermarking. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1436-1451, 2020. 1", + "[19] Han Fang, Zhaoyang Jia, Yupeng Qiu, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. De-end: decoder-driven watermarking network. IEEE Transactions on Multimedia, 25: 7571-7581, 2022. 1", + "[20] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2", + "[21] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 1", + "[22] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 2, 6, 7", + "[23] Yiyang Guo, Ruizhe Li, Mude Hui, Hanzhong Guo, Chen Zhang, Chuangjian Cai, Le Wan, and Shangfei Wang. Freqmark: Invisible image watermarking via frequency based optimization in latent space. arXiv preprint arXiv:2410.20824, 2024. 2", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18767", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 1", + "[27] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024. 1, 2", + "[28] Chunlei Li, Zhaoxiang Zhang, Yunhong Wang, Bin Ma, and Di Huang. Dither modulation of significant amplitude difference for wavelet based robust watermarking. Neurocomputing, 166:404-415, 2015. 1, 2", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[30] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 1", + "[31] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 1, 2", + "[32] A Miyazaki and A Okamoto. Analysis of watermarking systems in the frequency domain and its application to design of robust watermarking systems. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205), pages 506-509. IEEE, 2001. 2", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[36] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1, 2", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 6", + "[38] Joseph JK Ō Ruanaidh and Thierry Pun. Rotation, scale and translation invariant spread spectrum digital image watermarking. Signal processing, 66(3):303-317, 1998. 1, 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[40] Gustavo Santana. Gustavosta: Stable-diffusion-prompts. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2022.6", + "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3, 6", + "[42] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1", + "[43] Jun Tian. Reversible data embedding using a difference expansion. IEEE transactions on circuits and systems for video technology, 13(8):890-896, 2003. 1, 2", + "[44] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 1, 2", + "[45] Wenbo Wan, Jun Wang, Yunming Zhang, Jing Li, Hui Yu, and Jiande Sun. A comprehensive survey on robust image watermarking. Neurocomputing, 488:226-247, 2022. 2", + "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 6", + "[47] Xiu-mei Wen, Wei Zhao, and Fan-xing Meng. Research of a digital image watermarking algorithm resisting geometrical attacks in fourier domain. In 2009 International Conference on Computational Intelligence and Security, pages 265-268. IEEE, 2009. 2", + "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 7", + "[49] Raymond B Wolfgang, Christine I Podilchuk, and Edward J Delp. Perceptual watermarks for digital images and video. Proceedings of the IEEE, 87(7):1108-1126, 1999. 1, 2", + "[50] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 2", + "[51] Guokai Zhang, Lanjun Wang, Yuting Su, and An-An Liu. A training-free plug-and-play watermark framework for stable diffusion. arXiv preprint arXiv:2404.05607, 2024. 1, 2", + "[52] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 1, 6, 7" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18768", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Lijun Zhang, Xiao Liu, Antoni Martin, Cindy Bearfield, Yuriy Brun, and Hui Guan. Attack-resilient image watermarking using stable diffusion. Advances in Neural Information Processing Systems, 37:38480-38507, 2025. 2, 3, 4, 6, 7", + "[54] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 1, 2", + "[55] Xuandong Zhao, Kexun Zhang, Yu-Xiang Wang, and Lei Li. Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. 2023. 2", + "[56] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai. Advances in Neural Information Processing Systems, 37:8643-8672, 2025. 2, 6", + "[57] Yimeng Zhao, Chengyou Wang, Xiao Zhou, and Zhiliang Qin. Dari-mark: Deep learning and attention network for robust image watermarking. Mathematics, 11(1):209, 2022. 1", + "[58] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 1" + ], + "bbox": [ + 91, + 90, + 480, + 457 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18769", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_model.json b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ad78ab6af638ba8cfb95f13cabe6f7d9abc0a45f --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_model.json @@ -0,0 +1,2158 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.887, + 0.177 + ], + "angle": 0, + "content": "Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity" + }, + { + "type": "text", + "bbox": [ + 0.367, + 0.204, + 0.631, + 0.222 + ], + "angle": 0, + "content": "Sung Ju Lee Nam Ik Cho" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.239, + 0.728, + 0.258 + ], + "angle": 0, + "content": "Dept. of ECE & INMC, Seoul National University, Korea" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.26, + 0.666, + 0.274 + ], + "angle": 0, + "content": "thomas11809@snu.ac.kr, nicho@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at github.com/thomas11809/SFWMark." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With the advancement of diffusion generative models [17, 25, 33, 35, 39, 41], their generated images are being increasingly used across various fields, including creative works, entertainment, and advertisement. In particular, the open-source release of large-scale language-image (LLI) models like Stable Diffusion [37] has led to an exponential growth" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.309, + 0.907, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Figure 1. Summary of watermarking performance across different semantic watermarking methods, as detailed in Sec. 5.2. All methods follow the merged-in-generation scheme with no additional processing time. Verification is evaluated using TPR@1%FPR (True Positive Rate at \\(1\\%\\) False Positive Rate), while identification is assessed by Identification Accuracy (Perfect Match Rate). The proposed approaches achieve the best balance between detection robustness and image fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.746 + ], + "angle": 0, + "content": "in script-based image generation and editing technologies, resulting in a surge of generative content. This has raised new concerns, such as the copyright of AI-generated content and the tracking of images created with malicious intent. As a solution to these problems, the technique of embedding invisible watermarks into content has been proposed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Digital content watermarking has been extensively studied through both classical signal processing techniques [5, 6, 12, 13, 15, 28, 38, 43, 44, 49] and deep learning-based approaches [1, 18, 19, 21, 26, 30, 42, 52, 57, 58]. Recently, watermarking techniques that directly intervene in the generation process of diffusion models [3, 7, 10, 20, 22, 27, 31, 36, 51, 54] have been explored; however, these approaches require additional setup or configuration, adding more complexity." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Several studies have explored embedding watermarks di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18759" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "rectly into the latent representation, eliminating the need for external models. These include a steganographic approach that constructs latent noise with stream cipher-randomized bits [50] and semantic watermarking methods that embed geometric patterns in the Fourier frequency domain of the latent representation [11, 48, 53]. Meanwhile, research on pixel-level perturbation-based watermarking has highlighted its vulnerability to regeneration attacks [55, 56], demonstrating that semantic watermarking serves as a more robust alternative against such attacks. However, the aforementioned semantic watermarking methods in the latent Fourier domain, which are the focus of this paper, suffer from degraded detection accuracy and generative quality due to their lack of frequency integrity preservation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.603 + ], + "angle": 0, + "content": "In this context, we propose a novel semantic watermarking framework that enhances both detection robustness and image quality by preserving frequency integrity in the latent Fourier domain. Our method is named Hermitian Symmetric Fourier Watermarking (SFW), which ensures that watermark embeddings maintain statistical consistency with the latent noise distribution, leading to improved retrievability and stability in generative models. Additionally, we incorporate a center-aware embedding strategy that enhances robustness against cropping attacks by embedding watermarks in a spatially resilient region of the latent representation. We comprehensively evaluate our method across various attack scenarios, including signal processing distortions, regeneration attacks, and cropping attacks. Experimental results, as illustrated in Fig. 1, demonstrate that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while simultaneously maintaining superior generative quality, as evidenced by FID and CLIP score evaluations. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.605, + 0.483, + 0.65 + ], + "angle": 0, + "content": "- We propose Hermitian SFW to ensure frequency integrity, leading to improved watermark detection and generative quality." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.651, + 0.483, + 0.681 + ], + "angle": 0, + "content": "- We introduce center-aware embedding, which significantly enhances robustness against cropping attacks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.483, + 0.726 + ], + "angle": 0, + "content": "- We present extensive evaluations demonstrating that our approach outperforms existing baselines in detection accuracy and generative quality." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.605, + 0.483, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.74, + 0.24, + 0.755 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Digital Watermarking. Digital watermarking aims to achieve a balance between high embedding capacity and visual quality when inserting invisible watermarks while also ensuring robustness against various attacks [45]. This field originally began with the adoption of classical signal processing techniques. Embedding the watermark in the spatial domain is a straightforward and intuitive watermark insertion [43, 44], but it tends to be less robust against attacks such as filtering or compression. On the other hand," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.272, + 0.905, + 0.301 + ], + "angle": 0, + "content": "Figure 2. Overview of the semantic watermarking process in the latent Fourier domain using the merged-in-generation scheme." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.387 + ], + "angle": 0, + "content": "studies utilizing the frequency domain have demonstrated resilience against JPEG compression by embedding watermarks in the low and middle frequency bands [2, 5, 6, 12, 13, 15, 28, 32, 38, 47, 49]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Watermarks for Latent Diffusion Models. Latent diffusion models (LDMs), such as Stable Diffusion [37], conduct the diffusion process in a lower-dimensional latent space rather than directly on high-resolution images. Recent research has focused on techniques for integrating watermarks into the latent diffusion process to ensure traceability and robustness. However, many of these approaches rely on model fine-tuning [3, 10, 20, 22, 27] or require separately trained encoders and decoders [7, 10, 14, 20, 22, 23, 27, 31, 36, 51, 54] to facilitate watermark embedding and detection. These dependencies constrain the flexibility of watermarking, limiting their applicability across diverse models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Several studies have explored embedding semantic watermarks in the Fourier domain of latent vectors using a merged-in-generation scheme. Wen et al. [48] introduced a tree-ring-shaped watermark constructed with Gaussian-distributed radii, while Ci et al. [11] proposed a modified pattern with high-energy signed constant rings for watermark identification, along with an additional random noise key embedded in a separate channel. On the other hand, Zodiac [53] followed a post-hoc approach, embedding a tree-ring pattern into generated images through multiple iterations of latent vector optimization and diffusion-based synthesis. However, in addition to its high computational cost, this method suffers from low practicality, as it relies on extensive linear interpolation between the original and optimized images to artificially improve visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In these methods, when performing inverse Fourier transform, the imaginary component in the spatial domain is discarded, leading to a distorted frequency representation. As a result, the real component of the key region retains only partial information, while the imaginary component" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.276, + 0.476, + 0.29 + ], + "angle": 0, + "content": "Figure 3. Examples of various semantic watermarking patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.379 + ], + "angle": 0, + "content": "is almost entirely lost, creating an empty key region in the frequency domain. Since detection relies on analyzing this incomplete key region, the process is inherently limited, resulting in suboptimal retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Fig. 2 illustrates the overall pipeline of semantic watermarking with the merged-in-generation scheme. The embedding process begins with the Fourier transform applied to the latent noise, generating the latent Fourier representation. A watermark key is then embedded into a designated key region, followed by the inverse Fourier transform, producing the watermarked latent vector. Finally, text-guided image generation synthesizes the watermarked image. For detection, the process starts with a clean or attacked image, from which the latent query is obtained via DDIM inversion [41]. The presence of a watermark is determined by analyzing the query key region in the latent representation. Further details on detection tasks and evaluation metrics are provided in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.229, + 0.624 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.266, + 0.649 + ], + "angle": 0, + "content": "3.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.732 + ], + "angle": 0, + "content": "The performance of watermark detection is evaluated following RingID's formulation for watermark verification and identification tasks [11]. The metric used to measure the distance \\( d \\) is the \\( L_{1} \\) distance, calculated specifically for the key region where the watermark is embedded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Verification. The objective of verification is to determine whether a watermark is present in an image by analyzing the distance between the reference key and the key region of the query in the latent Fourier domain. Let \\(\\hat{w}\\) denote the watermarked latent Fourier key, and \\(\\hat{u}\\) denote the unwatermarked latent Fourier null key. Verification is based on comparing the distances between the reference key \\(w\\) and the watermarked/unwatermarked keys, i.e., \\(d(\\hat{w}, w) \\neq d(\\hat{u}, w)\\). Performance is assessed using statistical metrics derived from the ROC curve, considering different distance thresholds." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Identification. In identification, given that a watermark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "is already embedded, the task is to accurately determine the embedded information. This is achieved by computing the distance between the watermarked key \\(\\hat{w}\\) and multiple reference keys \\(w_{i}\\). Performance is evaluated based on the accuracy of the estimated message index, defined as \\(\\hat{i} = \\arg \\min_{i}d(\\hat{w},w_{i})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.191, + 0.898, + 0.206 + ], + "angle": 0, + "content": "3.2. Fourier Considerations for Real Latent Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.257 + ], + "angle": 0, + "content": "This section introduces the mathematical properties of the Fourier domain and highlights important considerations when embedding semantic watermarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Hermitian Symmetry for Real Signals. To obtain a real-valued signal after inverse Fourier transform, modifications in the frequency domain must maintain Hermitian symmetry about the DC center:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.326, + 0.905, + 0.345 + ], + "angle": 0, + "content": "\\[\nF [ M - k, N - l ] = \\overline {{F [ k , l ]}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.443 + ], + "angle": 0, + "content": "where \\(F\\) represents the discrete Fourier transform, \\(M\\) and \\(N\\) denote the dimensions of the signal in the spatial domain along the row and column axes, respectively. Failure to preserve this symmetry introduces undesired imaginary components in the spatial domain, which are incompatible with the real-valued latent noise required for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Impact of Ignoring Hermitian Symmetry. In existing baselines [11, 48, 53], this issue is simply addressed by discarding the imaginary component, resulting in a loss of frequency integrity. As illustrated in Fig. 3a, this disruption affects both real and imaginary components of the original watermark pattern. While the real component deviates from its intended structure due to frequency distortion, the imaginary component suffers even greater degradation, often becoming entirely empty. This significantly limits detection performance by preventing the effective use of frequency information in the imaginary domain. The impact of the frequency loss is further examined in Sec. 5.3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.906, + 0.686 + ], + "angle": 0, + "content": "Preservation of Gaussianity. A real Gaussian noise signal in the spatial domain transforms into a complex Gaussian noise signal in the frequency domain while maintaining its statistical properties:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.693, + 0.905, + 0.712 + ], + "angle": 0, + "content": "\\[\nf [ m, n ] \\sim \\mathcal {N} (0, \\sigma^ {2}) \\Rightarrow F [ k, l ] \\sim \\mathcal {C N} (0, M N \\sigma^ {2}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( f \\) represents the real-valued Gaussian noise signal in the spatial domain. When embedding watermarks in the frequency domain, some perturbation is inevitably introduced, altering the original distribution of complex Gaussian noise. However, if Hermitian symmetry is not preserved, the inverse Fourier transform produces incomplete noise characteristics in the spatial domain, as imaginary components must be discarded. This disrupts the statistical consistency of the latent noise, leading to a deviation from the expected real Gaussian distribution and potential degradation in generative quality. In contrast, embedding with Hermitian symmetry better retains the statistical structure of the latent" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18761" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.51, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.268, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed framework and qualitative results. (Left) The key components of our approach: Symmetric Fourier Watermarking (SFW) (blue region) and Center-Aware Embedding Strategy (green region). (Right) Qualitative results of Tree-Ring, RingID, HSTR, and HSQR. Notably, RingID exhibits visible ring-like artifacts, highlighting that its high-energy pattern disrupts generative quality, unlike other merged-in-generation semantic watermarking methods that achieve a better balance between robustness and image fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.486, + 0.427 + ], + "angle": 0, + "content": "noise, ensuring that the transformed spatial-domain signal remains closer to a real Gaussian distribution. By preserving these statistical properties, the initialization process in diffusion models becomes more stable, ultimately enhancing generative performance, as observed in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.442, + 0.192, + 0.456 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.467, + 0.453, + 0.483 + ], + "angle": 0, + "content": "4.1. Hermitian Symmetric Fourier Watermark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.485, + 0.761 + ], + "angle": 0, + "content": "The proposed Hermitian SFW refers to a watermark designed to satisfy the Hermitian symmetry introduced in the previous section. As shown in the left side of Fig. 4, SFW allows free usage of the half-region in the frequency domain, while the remaining half is restricted by the symmetry condition. Certain constraints must be satisfied when embedding a pattern in the free half-region under a semantic watermarking scheme: the imaginary part at the DC center must be zero, and if the signal dimensions are even, the imaginary components of the points corresponding to the Nyquist frequency must also be zero. In cases where both frequency axes have even dimensions as in this study, the points where the imaginary part must be zero are \\((0,0)\\), \\(\\left(\\frac{M}{2},0\\right)\\), \\((0,\\frac{N}{2})\\), and \\(\\left(\\frac{M}{2},\\frac{N}{2}\\right)\\). Instead of directly modifying the watermark pattern to comply with locations where the imaginary part must be zero, a more effective strategy is to embed it while avoiding the DC axis as much as possible, as will be introduced in Sec. 4.3.2" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.398, + 0.789 + ], + "angle": 0, + "content": "4.2. Center-Aware Embedding Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Existing semantic watermarking baselines [11, 48, 53] embed patterns by applying the Fourier transform to the full spatial matrix of the latent vector. However, such methods are vulnerable to spatial attacks, particularly cropping, which can lead to the loss of watermark information and reduced detection performance. To address this issue, we propose a center-aware embedding strategy, which applies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.442 + ], + "angle": 0, + "content": "the Fourier transform only to the central area of the spatial domain before embedding. Specifically, for a latent vector with a signal dimension of 64, we utilize the central \\(44 \\times 44\\) space. As demonstrated in Sec. 5.3.2, this design significantly improves robustness against cropping attacks on various scales." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.882, + 0.469 + ], + "angle": 0, + "content": "4.3. Integrating SFW and Center-Aware Design" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.475, + 0.747, + 0.489 + ], + "angle": 0, + "content": "4.3.1. Application to the Baseline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.631 + ], + "angle": 0, + "content": "As shown in Fig. 3a, the Tree-Ring [48] baseline suffers from frequency loss. To mitigate this, we impose the Hermitian symmetry condition directly on the watermark patterns. By integrating SFW with the center-aware embedding strategy, we refine the baseline into Hermitian Symmetric Tree-Ring (HSTR). HSTR not only enhances detection performance by fully utilizing the imaginary components of the latent Fourier domain but also improves image generation quality by restoring frequency integrity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.779, + 0.654 + ], + "angle": 0, + "content": "4.3.2. Hermitian Symmetric QR code" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.718 + ], + "angle": 0, + "content": "In this section, we introduce a novel approach that extends SFW beyond baselines to QR codes, which are widely recognized for their high versatility and robust error correction capabilities [16]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.719, + 0.906, + 0.81 + ], + "angle": 0, + "content": "**Embedding.** As illustrated in Fig. 3b, the Hermitian Symmetric QR Code (HSQR) watermark is constructed by splitting the QR code in half and embedding each part separately into the real and imaginary components of the free half-region in the Fourier domain. The binary pattern of the QR code is embedded using the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.835, + 0.907, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\operatorname {H S Q R} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 1 \\\\ - | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 0, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.887, + 0.565, + 0.9 + ], + "angle": 0, + "content": "where:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18762" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Verification performance of different watermarking methods under various attacks. Bit Accuracy is used for bitstream-based methods (DwtDct, DwtDctSvd, RivaGAN, S.Sign), while TPR@1%FPR is used for semantic methods. Best performances are highlighted. Our methods show superior detection accuracy and robustness against signal processing distortions, regeneration, and cropping attacks." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.144, + 0.898, + 0.507 + ], + "angle": 0, + "content": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.8630.5720.5220.5160.6770.8590.5320.5230.5210.5190.7290.8100.637
DwtDctSvd1.0000.5550.4730.6021.0001.0000.7840.6480.5960.6440.7440.8610.742
RivaGAN0.9990.8620.9860.8210.9980.9690.9340.5700.5520.6080.9910.9950.857
S.Sign.0.9950.8940.9780.8060.9110.7210.8380.7170.7150.4780.9870.9910.836
Tree-Ring0.9570.4630.9000.5480.9340.4120.8150.5090.5360.5430.5090.7340.655
Zodiac0.9980.8430.9980.9730.9980.8800.9970.9440.9580.9720.9890.9950.962
HSTR (ours)1.0000.8991.0000.9941.0000.8060.9990.9730.9820.9971.0001.0000.971
RingID1.0000.9881.0001.0001.0000.9871.0000.9921.0001.0001.0001.0000.997
HSQR (ours)1.0000.9911.0001.0001.0000.9831.0000.9921.0001.0001.0001.0000.997
SD-PromptsDwtDct0.8190.5570.5160.5060.6850.8220.5300.5130.5120.5090.7230.7940.624
DwtDctSvd1.0000.5370.4590.6100.9990.9980.8590.6590.6200.6230.7430.8600.747
RivaGAN0.9910.8230.9630.8100.9880.9610.9150.5720.5350.5670.9800.9830.841
S.Sign.0.9940.8990.9670.7690.8880.7420.8090.6770.6710.4930.9830.9900.824
Tree-Ring0.9440.4710.8940.4660.9120.4230.8020.5090.5140.5430.4690.7490.641
Zodiac0.9980.7480.9990.9790.9990.9031.0000.9400.9750.9580.9940.9960.957
HSTR (ours)1.0000.7421.0000.9901.0000.8501.0000.9830.9870.9991.0001.0000.963
RingID1.0000.9721.0001.0001.0000.9881.0000.9961.0001.0001.0001.0000.996
HSQR (ours)1.0000.9551.0000.9991.0000.9921.0000.9960.9991.0001.0001.0000.995
DiffusionDBDwtDct0.8420.5630.5150.5090.6720.8290.5260.5130.5140.5120.7230.8010.627
DwtDctSvd0.9980.5580.4630.5930.9970.9950.8300.6580.6080.6210.7420.8600.744
RivaGAN0.9870.8390.9600.7900.9850.9370.8930.5530.5180.5560.9740.9790.831
S.Sign.0.9900.8900.9670.7870.8890.7260.8190.6900.6870.4960.9810.9860.826
Tree-Ring0.9400.4870.8890.4340.9040.3920.7990.4540.5030.4540.4990.7150.622
Zodiac0.9920.7520.9880.9330.9880.8340.9840.9110.9260.9030.9710.9850.931
HSTR (ours)0.9990.7920.9960.9810.9960.7920.9910.9680.9690.9891.0001.0000.956
RingID1.0000.9891.0001.0001.0000.9631.0000.9950.9991.0001.0001.0000.995
HSQR (ours)1.0000.9771.0000.9991.0000.9740.9990.9970.9991.0001.0001.0000.995
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.532, + 0.435, + 0.546 + ], + "angle": 0, + "content": "- \\( x \\) denotes the coordinates in the QR code domain." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.607 + ], + "angle": 0, + "content": "- \\(\\tilde{x}\\) represents the corresponding coordinates in the embedding region of the Fourier domain. Since the QR code is split into two halves, each half is mapped to either the real or imaginary component." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.608, + 0.483, + 0.638 + ], + "angle": 0, + "content": "- \\( c \\in \\{\\mathrm{Re}, \\mathrm{Im}\\} \\) indicates whether the embedding occurs in the real or imaginary part." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.638, + 0.483, + 0.698 + ], + "angle": 0, + "content": "- \\( F(\\tilde{x}, c) \\) is the real-valued amplitude of the complex Gaussian noise in the Fourier domain, where a complex number is expressed as \\( F_{\\mathrm{Re}} + iF_{\\mathrm{Im}} \\), making both real and imaginary amplitudes individually real-valued." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.532, + 0.483, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.79 + ], + "angle": 0, + "content": "To maintain symmetry while avoiding numerical instability, the embedding region is positioned one pixel to the right of the vertical DC axis. Additionally, to increase the amount of embedded information and statistically reduce errors, each QR code cell is represented by multiple pixels arranged in a square pattern, e.g., \\(2 \\times 2\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Detection. The detection process for semantic watermarking is performed by computing the L1 distance. In HSQR, the ground truth QR binary pattern is converted into a signed Boolean representation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.864, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q R} ^ {*} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 1 \\\\ - \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 0, \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.842, + 0.547 + ], + "angle": 0, + "content": "where \\(\\Lambda\\) is a fixed amplitude for binary encoding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.645, + 0.582 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.72, + 0.607 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Semantic Watermarks. The Tree-Ring [48] watermark is embedded into channel 3 with a radius of 14 and HSTR adopts the same watermark parameters as its baseline. For RingID [11], we follow the original settings, embedding a tree-ring pattern with a radius range of 3-14 in channel 3, along with a Gaussian noise key in channel 0. Both SFW methods follow RingID in utilizing a Gaussian noise key in channel 0. For HSQR, we use a version 1 QR code \\((21 \\times 21\\) cells) with a cell size of 2 pixels and error correction level H, capable of encoding up to 72 bits. The QR code is embedded in channel 3, and the encoding amplitude \\(\\Lambda\\) is set to 45, corresponding to the standard deviation of the real and imaginary components in the Fourier domain of \\(64 \\times 64\\) normal Gaussian latent vector \\((\\approx \\sqrt{64^2 / 2})\\). Note that these experiments are conducted on \\(512 \\times 512\\) images. At higher resolutions, a larger embedding space allows for increased capacity through higher QR code versions that encode more bits, as well as the use of stronger watermarking parameters, enabling greater robustness and adaptability. Additionally," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18763" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Identification accuracy for different watermarking methods, evaluated across multiple attack conditions. Perfect Match Rate is used for bitstream-based methods. The best performance for each item is highlighted with shading. HSQR achieves state-of-the-art performance, while HSTR consistently outperforms other Gaussian radius-based semantic baselines, particularly under cropping attacks." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.144, + 0.898, + 0.507 + ], + "angle": 0, + "content": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.4660.0440.0000.0000.0380.4420.0000.0000.0000.0000.0000.0000.083
DwtDctSvd1.0000.0440.0190.0000.9990.9980.0370.0000.0000.0030.0000.0000.258
RivaGAN0.9740.2600.7720.0230.9610.6860.3480.0000.0000.0000.8520.9090.482
S.Sign.0.8730.1770.5630.0000.0360.0100.0070.0000.0000.0000.7090.8020.265
Tree-Ring0.3030.0870.2070.0720.2560.0300.1620.0830.0720.0540.0090.0330.114
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.0000.7140.9990.8860.9980.4600.9720.8330.8310.9711.0001.0000.889
RingID1.0000.8751.0000.9751.0000.9190.9960.9780.9700.9980.8740.9780.964
HSQR (ours)1.0000.9581.0000.9941.0000.9010.9990.9800.9870.9991.0001.0000.985
SD-PromptsDwtDct0.2850.0240.0000.0000.0170.2760.0000.0000.0000.0000.0000.0000.050
DwtDctSvd0.9930.0280.0110.0000.9820.9790.0850.0000.0000.0070.0000.0000.257
RivaGAN0.8780.2130.6130.0090.8570.6570.3040.0000.0000.0000.7360.7680.420
S.Sign.0.8130.2630.4200.0000.0210.0150.0060.0000.0000.0000.5760.7160.236
Tree-Ring0.2880.0940.1890.0510.2350.0340.1590.0790.0760.0560.0120.0410.110
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.000
RingID1.0000.8851.0000.9760.9980.8860.9930.9800.9730.9950.8760.9810.962
HSQR (ours)1.0000.9301.0000.9941.0000.9420.9990.9910.9971.0001.0001.0000.988
DiffusionDBDwtDct0.3570.0370.0000.0000.0340.3200.0000.0000.0000.0000.0000.0000.062
DwtDctSvd0.9900.0360.0190.0000.9750.9590.0810.0000.0000.0010.0000.0000.255
RivaGAN0.8580.2130.6250.0200.8480.6150.2210.0000.0000.0000.7560.7800.411
S.Sign.0.7980.2070.4720.0000.0270.0050.0050.0000.0000.0000.6430.7380.241
Tree-Ring0.2800.0950.1900.0590.2330.0370.1450.0810.0720.0500.0130.0390.108
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)0.9960.7210.9920.8540.9890.5630.9580.8300.8210.9520.9960.9960.889
RingID1.0000.8951.0000.9470.9960.8710.9920.9680.9580.9900.8750.9840.956
HSQR (ours)1.0000.9541.0000.9881.0000.9060.9980.9820.9910.9941.0001.0000.984
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.563 + ], + "angle": 0, + "content": "the post-hoc semantic watermarking method, Zodiac [53], is included in our comparative experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.565, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Model Setup and Datasets. The diffusion model used for the experiments is Stable Diffusion v2-1-base [37], configured to generate images at a resolution of \\(512 \\times 512\\) with a CFG scale of 7.5. Both generation and inversion steps of the DDIM scheduler [41] are set to 50. The datasets used for image generation include 5,000 captions from the MS-COCO-2017 training set [29], 1,001 sampled prompts from the 2M subset of DiffusionDB [46], and 8,192 prompts from the Stable-Diffusion-Prompts test set [40]. For watermark detection, the number of watermarked images used for each dataset is set to 1,000. Specifically, verification is conducted on 1,000 pairs of watermarked and unwatermarked images, while identification is performed on 1,000 watermarked images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Attack Methods. To evaluate the robustness of our method against various distortions, we apply 11 different attacks, categorized into three major types." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Signal Processing Attacks: These include brightness adjustment (ranging from 0 to 7), contrast adjustment (factor of 0.5), JPEG compression (quality factor of 25), Gaussian blur (radius of 5), Gaussian noise (\\(\\sigma = 0.05\\)), and BM3D denoising (\\(\\sigma = 0.1\\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.607 + ], + "angle": 0, + "content": "- Regeneration Attacks: We apply two VAE-based image compression models, Bmshj18 [4] (VAE-B) and Cheng20 [8] (VAE-C), both at quality level 3, as well as a diffusion-based regeneration attack by Zhao et al. [56] (referred to as Diff.) using 60 denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.608, + 0.906, + 0.669 + ], + "angle": 0, + "content": "- Cropping Attacks: We apply center crop (C.C.) with a crop scale of 0.5 and random crop (R.C.) with a crop scale of 0.7, where the crop scale represents the ratio of the cropped image area to the original image." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. The \\(L_{1}\\) distance is computed only within the key region of the complex Fourier domain of latent noise, and unless otherwise specified, 2,048 keys are used for identification accuracy. For verification, semantic watermarking methods use True Positive Rate at \\(1\\%\\) False Positive Rate (TPR@1%FPR) as the evaluation metric. To ensure a fair comparison with bitstream-based approaches such as DwtDct, DwtDctSvd [12], RivaGAN [52], and Stable Signature (S.Sign.) [22], we measure Bit Accuracy for these methods. For identification, where the goal is to match the exact message, Perfect Match Rate is used as the metric for bitstream-based approaches. To evaluate the generation quality, FID [24] is measured between watermarked images and 5,000 MS-COCO ground truth images. We also calculate the CLIP score [34] using the OpenCLIP-ViT/G" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18764" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.215 + ], + "angle": 0, + "content": "Table 3. Generative quality evaluation of watermarking methods based on FID (MS-COCO ground truth) and CLIP score. The best performance for each item is highlighted with shading, while bold text specifically marks the low CLIP score in RingID. Our proposed methods preserve frequency integrity, achieving the best balance between watermark robustness and generative performance, whereas RingID introduces visible artifacts, compromising perceptual quality. Vrf. and Idf. denote the average detection performance in verification and identification tasks, respectively." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.226, + 0.462, + 0.3 + ], + "angle": 0, + "content": "
Semantic MethodsFID ↓CLIP ↑Vrf.Idf.
Merged in GenerationTree-Ring26.4180.3260.6550.114
RingID27.0520.3240.9970.964
HSTR (ours)25.0620.3290.9710.889
HSQR (ours)24.8950.3300.9970.985
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.483, + 0.356 + ], + "angle": 0, + "content": "model [9] to determine how closely the generated image aligns with the input prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.336, + 0.382 + ], + "angle": 0, + "content": "5.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.483, + 0.432 + ], + "angle": 0, + "content": "In this section, we compare the detection performance in both verification and identification tasks, as well as the generation quality of different watermarking methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Enhanced Detection Robustness. According to Tab. 1, our proposed methods consistently achieve top-tier detection performance across various attacks. A key observation is that non-semantic watermarking methods [12, 22, 52] exhibit significant degradation under regeneration attacks, highlighting their vulnerability. Among Gaussian radius-based pattern methods (Tree-Ring, Zodiac, and HSTR), HSTR outperforms corresponding baselines in most cases, demonstrating its effectiveness in enhancing detection robustness while maintaining the fundamental tree-ring structure. In contrast, RingID leverages high-energy signed constant ring patterns, securing strong detection performance. However, as acknowledged by the authors, this method introduces noticeable ring-like artifacts in generated images, disrupting the balance between watermark robustness and generative quality. A detailed quantitative evaluation of this trade-off is presented in the following section. Meanwhile, Zodiac suffers from time-consuming processing, taking several minutes to apply the watermark—requiring 7.36 minutes per image on MS-COCO dataset—which further limits its practicality in real-world applications. In contrast, our methods following the merged-in-generation scheme introduce no additional processing time, ensuring seamless watermark embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In Tab. 2, we compare identification results, which reveal an even greater disparity in detection performance across different watermarking methods. Gaussian radius-based pattern methods, except for HSTR, perform poorly in identification tasks, with Zodiac—designed solely for verification—failing entirely across all scenarios (achieving zero identification accuracy). Further analysis on scalability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "under different message capacities is provided in Sec. 5.3.3. Notably, our methods exhibit strong resilience against cropping attacks in both tasks, reinforcing the robustness of center-aware embedding. HSQR achieves state-of-the-art identification accuracy across all datasets, further establishing its dominance in watermark retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Balance with Generative Quality. Tab. 3 presents the generative quality of different semantic watermarking methods following the merged-in-generation scheme, evaluated using FID and CLIP score. RingID, which deviates from a Gaussian distribution by embedding high-energy perturbations, exhibits the worst generative performance, particularly reflected in its low CLIP score. This decline is attributed to the noticeable ring-like artifacts, as discussed earlier. In contrast, our proposed methods, which preserve frequency integrity via SFW, achieve the top two FID scores, demonstrating a better balance between watermark robustness and generative quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.742 + ], + "angle": 0, + "content": "Expanding Frequency Utilization in Latent Watermarking. Traditional image-domain watermarking methods prioritize low-mid frequency embedding to resist compression and filtering attacks. This approach has also been adopted in latent diffusion-based semantic watermarking methods [11, 48, 53] without explicitly considering the differences between pixel-space and latent-space transformations. However, our findings suggest that such frequency constraints may not be necessary in the latent space, where watermark retrieval involves latent encoding and DDIM inversion, altering the impact of frequency perturbations. This distinction is evident in HSQR, which achieves state-of-the-art detection performance while utilizing nearly the entire frequency spectrum. Instead of relying on specific frequency bands, HSQR leverages structured statistical encoding, distributing watermark information across multiple pixels (\\(2 \\times 2\\) per QR cell). This redundancy enhances robustness against distortions, enabling effective retrieval even with broader frequency usage. These results indicate that latent-space watermarking is not bounded by traditional low-mid frequency constraints. Future semantic watermarking strategies can benefit from statistically robust encoding methods, allowing for effective watermark retrieval across a wider frequency range without compromising detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.667, + 0.77 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.852, + 0.79 + ], + "angle": 0, + "content": "5.3.1. Impact of SFW on Detection Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.903 + ], + "angle": 0, + "content": "We evaluate the impact of Hermitian SFW by comparing detection performance across four cases (Tab. 4). Without SFW, using both real and imaginary components for detection (Case A) results in the lowest performance due to frequency degradation, while restricting detection to the real component (Case B) improves accuracy. In contrast, applying SFW (Cases C and D) preserves frequency integrity," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18765" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.175 + ], + "angle": 0, + "content": "Table 4. Ablation study on detection performance based on frequency integrity (\\(\\sqrt{}\\) or \\(\\times\\)) and the number of detection region usage (1: real only, 2: real & imaginary). Vrf. and Idf. represent average detection performance in verification (TPR@1% FPR) and identification (accuracy), respectively. \\(\\Delta L_1^*\\) denotes the normalized \\(\\Delta L_1\\) metric, indicating detection effectiveness." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.184, + 0.48, + 0.253 + ], + "angle": 0, + "content": "
CaseMethodsFreq. Int.# Det.Vrf.Idf.ΔL1*↑
ATree-Ring×20.6530.1140.232
BTree-Ring×10.8050.4160.368
CHSTR (ours)✓10.9360.7750.471
DHSTR (ours)✓20.9710.8890.476
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.269, + 0.289, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.269, + 0.482, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.427, + 0.485, + 0.485 + ], + "angle": 0, + "content": "Figure 5. Identification accuracy under center crop and random crop attacks at different crop scales. HSTR and HSQR maintain higher accuracy compared to RingID, demonstrating improved robustness against cropping." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.501, + 0.481, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.709, + 0.485, + 0.766 + ], + "angle": 0, + "content": "Figure 6. Identification accuracy across watermark message capacities. HSQR remains nearly perfect, while RingID and HSTR degrade at higher capacities. Tree-Ring and Zodiac fail to scale effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "leading to a significant boost in detection performance. Notably, HSTR with SFW and full complex detection (Case D) achieves the highest accuracy, demonstrating that leveraging both frequency components maximizes retrieval effectiveness. These results confirm that SFW enables robust semantic watermarking by maintaining frequency integrity and fully utilizing the Fourier domain for detection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.787, + 0.108 + ], + "angle": 0, + "content": "5.3.2. Robustness to Cropping Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.907, + 0.277 + ], + "angle": 0, + "content": "We assess the robustness of center-aware embedding against center crop and random crop attacks, measuring identification accuracy across different crop scales (Fig. 5). As crop scale decreases, accuracy declines for all methods. However, RingID exhibits a steeper drop, indicating higher vulnerability to cropping. In contrast, HSTR and HSQR degrade more gradually, demonstrating improved resilience. While extreme cropping (scale 0.2) significantly affects all methods, our center-aware design consistently outperforms RingID, confirming its effectiveness in preserving watermark information under cropping distortions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.813, + 0.3 + ], + "angle": 0, + "content": "5.3.3. Impact of Capacity on Identification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.456 + ], + "angle": 0, + "content": "We evaluate identification accuracy across different watermark message capacities (64 to 8,192) for five methods (Fig. 6). Zodiac's performance collapses, while Tree-Ring exhibits rapid performance degradation, becoming nearly unusable at higher capacities. RingID, HSTR, and HSQR remain robust, maintaining \\(80\\%+\\) accuracy, though HSTR declines faster, and RingID starts degrading from 2,048 onward. HSQR demonstrates the highest scalability, retaining near-perfect accuracy even at the largest capacity, confirming its superior robustness in high-capacity scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.634, + 0.484 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We have introduced Hermitian SFW, a novel approach to semantic watermarking in the latent diffusion model framework. Unlike existing methods that fail to preserve frequency integrity, our approach ensures that watermark embeddings maintain a consistent statistical structure in the latent noise distribution. This is achieved through Hermitian symmetry enforcement, which preserves frequency components and enhances detection and generation quality. Additionally, we have proposed center-aware embedding, which significantly improves robustness against cropping attacks by strategically placing watermarks in a spatially resilient region of the latent representation. Through comprehensive experiments, we demonstrated that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while also maintaining superior image generation quality, as shown by FID and CLIP scores. Our study highlights the importance of frequency integrity in Fourier-based watermarking and challenges the assumption that semantic watermarking must be confined to low-mid frequency bands. Experimental results confirm that a properly structured frequency-domain watermark can be effectively embedded and retrieved across the entire frequency spectrum without compromising generative quality. Future work includes exploring adaptive embedding strategies to further enhance robustness against adversarial attacks and extreme distortions, as well as extending our method to more diverse generative architectures beyond LDMs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18766" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.213 + ], + "angle": 0, + "content": "This research was supported by the 2024 AI Semiconductor Application/Demonstration Support Program of the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) with Markany Co., Ltd. as the lead organization, and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2025." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.238, + 0.188, + 0.253 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[2] Dhruv Arya. A survey of frequency and wavelet domain digital watermarking techniques. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[3] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. Promark: Proactive diffusion watermarking for causal attribution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10802-10811, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.471, + 0.483, + 0.513 + ], + "angle": 0, + "content": "[5] Mauro Barni, Franco Bartolini, Vito Cappellini, and Alessandro Piva. A dct-domain system for robust image watermarking. Signal processing, 66(3):357-372, 1998. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.483, + 0.568 + ], + "angle": 0, + "content": "[6] Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet-based watermarking through pixel-wise masking. IEEE transactions on image processing, 10(5): 783-791, 2001. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.569, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[7] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.483, + 0.707 + ], + "angle": 0, + "content": "[8] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.708, + 0.483, + 0.79 + ], + "angle": 0, + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.791, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. arXiv preprint arXiv:2404.14055, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[13] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE TIP, 6(12):1673-1687, 1997. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[14] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for data copyright protection against generative diffusion models. ACM SIGKDD Explorations Newsletter, 26(2): 60-75, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.304 + ], + "angle": 0, + "content": "[15] DEJEY and RS Rajesh. An improved wavelet domain digital watermarking for image protection. International journal of wavelets, multiresolution and information processing, 8(01): 19-31, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[16] Denso Wave Incorporated. QR Code Essentials. Denso Wave, 2011. Available at: https://www.qrcode.com/en/about/standards.html.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.391 + ], + "angle": 0, + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.906, + 0.448 + ], + "angle": 0, + "content": "[18] Han Fang, Dongdong Chen, Qidong Huang, Jie Zhang, Zehua Ma, Weiming Zhang, and Nenghai Yu. Deep template-based watermarking. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1436-1451, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[19] Han Fang, Zhaoyang Jia, Yupeng Qiu, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. De-end: decoder-driven watermarking network. IEEE Transactions on Multimedia, 25: 7571-7581, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[20] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[21] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[22] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[23] Yiyang Guo, Ruizhe Li, Mude Hui, Hanzhong Guo, Chen Zhang, Chuangjian Cai, Le Wan, and Shangfei Wang. Freqmark: Invisible image watermarking via frequency based optimization in latent space. arXiv preprint arXiv:2410.20824, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18767" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[26] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[27] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[28] Chunlei Li, Zhaoxiang Zhang, Yunhong Wang, Bin Ma, and Di Huang. Dither modulation of significant amplitude difference for wavelet based robust watermarking. Neurocomputing, 166:404-415, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[30] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[31] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[32] A Miyazaki and A Okamoto. Analysis of watermarking systems in the frequency domain and its application to design of robust watermarking systems. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205), pages 506-509. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[36] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Joseph JK Ō Ruanaidh and Thierry Pun. Rotation, scale and translation invariant spread spectrum digital image watermarking. Signal processing, 66(3):303-317, 1998. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[40] Gustavo Santana. Gustavosta: Stable-diffusion-prompts. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[42] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[43] Jun Tian. Reversible data embedding using a difference expansion. IEEE transactions on circuits and systems for video technology, 13(8):890-896, 2003. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[44] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[45] Wenbo Wan, Jun Wang, Yunming Zhang, Jing Li, Hui Yu, and Jiande Sun. A comprehensive survey on robust image watermarking. Neurocomputing, 488:226-247, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[47] Xiu-mei Wen, Wei Zhao, and Fan-xing Meng. Research of a digital image watermarking algorithm resisting geometrical attacks in fourier domain. In 2009 International Conference on Computational Intelligence and Security, pages 265-268. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[49] Raymond B Wolfgang, Christine I Podilchuk, and Edward J Delp. Perceptual watermarks for digital images and video. Proceedings of the IEEE, 87(7):1108-1126, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.799 + ], + "angle": 0, + "content": "[50] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[51] Guokai Zhang, Lanjun Wang, Yuting Su, and An-An Liu. A training-free plug-and-play watermark framework for stable diffusion. arXiv preprint arXiv:2404.05607, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[52] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[53] Lijun Zhang, Xiao Liu, Antoni Martin, Cindy Bearfield, Yuriy Brun, and Hui Guan. Attack-resilient image watermarking using stable diffusion. Advances in Neural Information Processing Systems, 37:38480-38507, 2025. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[54] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[55] Xuandong Zhao, Kexun Zhang, Yu-Xiang Wang, and Lei Li. Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[56] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai. Advances in Neural Information Processing Systems, 37:8643-8672, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[57] Yimeng Zhao, Chengyou Wang, Xiao Zhou, and Zhiliang Qin. Dari-mark: Deep learning and attention network for robust image watermarking. Mathematics, 11(1):209, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[58] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18769" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_origin.pdf b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc6e235a9e5fbbbbe9d0e7b86a7abc14919c766f --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/4150a989-5307-4ff1-8e86-ca9b050e7e76_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:295e3144d32e528fcfacab627682f7aa04278d5cc748edd8c2dc3b19c5abaf6f +size 3168638 diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/full.md b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77ff7c65ce84408e2fac92dcd341e399745eeac1 --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/full.md @@ -0,0 +1,288 @@ +# Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity + +Sung Ju Lee Nam Ik Cho + +Dept. of ECE & INMC, Seoul National University, Korea + +thomas11809@snu.ac.kr, nicho@snu.ac.kr + +# Abstract + +Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at github.com/thomas11809/SFWMark. + +# 1. Introduction + +With the advancement of diffusion generative models [17, 25, 33, 35, 39, 41], their generated images are being increasingly used across various fields, including creative works, entertainment, and advertisement. In particular, the open-source release of large-scale language-image (LLI) models like Stable Diffusion [37] has led to an exponential growth + +![](images/5bbcfca8fcbdc3cc4228a5a7344a6d4682b897c62e1a5eec1f2b6ad07fd5a977.jpg) +Figure 1. Summary of watermarking performance across different semantic watermarking methods, as detailed in Sec. 5.2. All methods follow the merged-in-generation scheme with no additional processing time. Verification is evaluated using TPR@1%FPR (True Positive Rate at $1\%$ False Positive Rate), while identification is assessed by Identification Accuracy (Perfect Match Rate). The proposed approaches achieve the best balance between detection robustness and image fidelity. + +in script-based image generation and editing technologies, resulting in a surge of generative content. This has raised new concerns, such as the copyright of AI-generated content and the tracking of images created with malicious intent. As a solution to these problems, the technique of embedding invisible watermarks into content has been proposed. + +Digital content watermarking has been extensively studied through both classical signal processing techniques [5, 6, 12, 13, 15, 28, 38, 43, 44, 49] and deep learning-based approaches [1, 18, 19, 21, 26, 30, 42, 52, 57, 58]. Recently, watermarking techniques that directly intervene in the generation process of diffusion models [3, 7, 10, 20, 22, 27, 31, 36, 51, 54] have been explored; however, these approaches require additional setup or configuration, adding more complexity. + +Several studies have explored embedding watermarks di + +rectly into the latent representation, eliminating the need for external models. These include a steganographic approach that constructs latent noise with stream cipher-randomized bits [50] and semantic watermarking methods that embed geometric patterns in the Fourier frequency domain of the latent representation [11, 48, 53]. Meanwhile, research on pixel-level perturbation-based watermarking has highlighted its vulnerability to regeneration attacks [55, 56], demonstrating that semantic watermarking serves as a more robust alternative against such attacks. However, the aforementioned semantic watermarking methods in the latent Fourier domain, which are the focus of this paper, suffer from degraded detection accuracy and generative quality due to their lack of frequency integrity preservation. + +In this context, we propose a novel semantic watermarking framework that enhances both detection robustness and image quality by preserving frequency integrity in the latent Fourier domain. Our method is named Hermitian Symmetric Fourier Watermarking (SFW), which ensures that watermark embeddings maintain statistical consistency with the latent noise distribution, leading to improved retrievability and stability in generative models. Additionally, we incorporate a center-aware embedding strategy that enhances robustness against cropping attacks by embedding watermarks in a spatially resilient region of the latent representation. We comprehensively evaluate our method across various attack scenarios, including signal processing distortions, regeneration attacks, and cropping attacks. Experimental results, as illustrated in Fig. 1, demonstrate that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while simultaneously maintaining superior generative quality, as evidenced by FID and CLIP score evaluations. Our contributions can be summarized as follows: + +- We propose Hermitian SFW to ensure frequency integrity, leading to improved watermark detection and generative quality. +- We introduce center-aware embedding, which significantly enhances robustness against cropping attacks. +- We present extensive evaluations demonstrating that our approach outperforms existing baselines in detection accuracy and generative quality. + +# 2. Related Works + +Digital Watermarking. Digital watermarking aims to achieve a balance between high embedding capacity and visual quality when inserting invisible watermarks while also ensuring robustness against various attacks [45]. This field originally began with the adoption of classical signal processing techniques. Embedding the watermark in the spatial domain is a straightforward and intuitive watermark insertion [43, 44], but it tends to be less robust against attacks such as filtering or compression. On the other hand, + +![](images/79831f3006d8a799d55b67df55ba2fe6115882fb6449f42436d3e6ef4d026d16.jpg) +Figure 2. Overview of the semantic watermarking process in the latent Fourier domain using the merged-in-generation scheme. + +studies utilizing the frequency domain have demonstrated resilience against JPEG compression by embedding watermarks in the low and middle frequency bands [2, 5, 6, 12, 13, 15, 28, 32, 38, 47, 49]. + +Watermarks for Latent Diffusion Models. Latent diffusion models (LDMs), such as Stable Diffusion [37], conduct the diffusion process in a lower-dimensional latent space rather than directly on high-resolution images. Recent research has focused on techniques for integrating watermarks into the latent diffusion process to ensure traceability and robustness. However, many of these approaches rely on model fine-tuning [3, 10, 20, 22, 27] or require separately trained encoders and decoders [7, 10, 14, 20, 22, 23, 27, 31, 36, 51, 54] to facilitate watermark embedding and detection. These dependencies constrain the flexibility of watermarking, limiting their applicability across diverse models. + +Several studies have explored embedding semantic watermarks in the Fourier domain of latent vectors using a merged-in-generation scheme. Wen et al. [48] introduced a tree-ring-shaped watermark constructed with Gaussian-distributed radii, while Ci et al. [11] proposed a modified pattern with high-energy signed constant rings for watermark identification, along with an additional random noise key embedded in a separate channel. On the other hand, Zodiac [53] followed a post-hoc approach, embedding a tree-ring pattern into generated images through multiple iterations of latent vector optimization and diffusion-based synthesis. However, in addition to its high computational cost, this method suffers from low practicality, as it relies on extensive linear interpolation between the original and optimized images to artificially improve visual quality. + +In these methods, when performing inverse Fourier transform, the imaginary component in the spatial domain is discarded, leading to a distorted frequency representation. As a result, the real component of the key region retains only partial information, while the imaginary component + +![](images/dc28e9aa714e81edfecb78c3e654df8afd03e20873e931db02496f0b9bbf1ed4.jpg) +Figure 3. Examples of various semantic watermarking patterns. + +is almost entirely lost, creating an empty key region in the frequency domain. Since detection relies on analyzing this incomplete key region, the process is inherently limited, resulting in suboptimal retrieval performance. + +Fig. 2 illustrates the overall pipeline of semantic watermarking with the merged-in-generation scheme. The embedding process begins with the Fourier transform applied to the latent noise, generating the latent Fourier representation. A watermark key is then embedded into a designated key region, followed by the inverse Fourier transform, producing the watermarked latent vector. Finally, text-guided image generation synthesizes the watermarked image. For detection, the process starts with a clean or attacked image, from which the latent query is obtained via DDIM inversion [41]. The presence of a watermark is determined by analyzing the query key region in the latent representation. Further details on detection tasks and evaluation metrics are provided in Sec. 3.1. + +# 3. Preliminaries + +# 3.1. Task Formulation + +The performance of watermark detection is evaluated following RingID's formulation for watermark verification and identification tasks [11]. The metric used to measure the distance $d$ is the $L_{1}$ distance, calculated specifically for the key region where the watermark is embedded. + +Verification. The objective of verification is to determine whether a watermark is present in an image by analyzing the distance between the reference key and the key region of the query in the latent Fourier domain. Let $\hat{w}$ denote the watermarked latent Fourier key, and $\hat{u}$ denote the unwatermarked latent Fourier null key. Verification is based on comparing the distances between the reference key $w$ and the watermarked/unwatermarked keys, i.e., $d(\hat{w}, w) \neq d(\hat{u}, w)$ . Performance is assessed using statistical metrics derived from the ROC curve, considering different distance thresholds. + +Identification. In identification, given that a watermark + +is already embedded, the task is to accurately determine the embedded information. This is achieved by computing the distance between the watermarked key $\hat{w}$ and multiple reference keys $w_{i}$ . Performance is evaluated based on the accuracy of the estimated message index, defined as $\hat{i} = \arg \min_{i}d(\hat{w},w_{i})$ . + +# 3.2. Fourier Considerations for Real Latent Noise + +This section introduces the mathematical properties of the Fourier domain and highlights important considerations when embedding semantic watermarks. + +Hermitian Symmetry for Real Signals. To obtain a real-valued signal after inverse Fourier transform, modifications in the frequency domain must maintain Hermitian symmetry about the DC center: + +$$ +F [ M - k, N - l ] = \overline {{F [ k , l ]}}, \tag {1} +$$ + +where $F$ represents the discrete Fourier transform, $M$ and $N$ denote the dimensions of the signal in the spatial domain along the row and column axes, respectively. Failure to preserve this symmetry introduces undesired imaginary components in the spatial domain, which are incompatible with the real-valued latent noise required for diffusion models. + +Impact of Ignoring Hermitian Symmetry. In existing baselines [11, 48, 53], this issue is simply addressed by discarding the imaginary component, resulting in a loss of frequency integrity. As illustrated in Fig. 3a, this disruption affects both real and imaginary components of the original watermark pattern. While the real component deviates from its intended structure due to frequency distortion, the imaginary component suffers even greater degradation, often becoming entirely empty. This significantly limits detection performance by preventing the effective use of frequency information in the imaginary domain. The impact of the frequency loss is further examined in Sec. 5.3.1. + +Preservation of Gaussianity. A real Gaussian noise signal in the spatial domain transforms into a complex Gaussian noise signal in the frequency domain while maintaining its statistical properties: + +$$ +f [ m, n ] \sim \mathcal {N} (0, \sigma^ {2}) \Rightarrow F [ k, l ] \sim \mathcal {C N} (0, M N \sigma^ {2}), \tag {2} +$$ + +where $f$ represents the real-valued Gaussian noise signal in the spatial domain. When embedding watermarks in the frequency domain, some perturbation is inevitably introduced, altering the original distribution of complex Gaussian noise. However, if Hermitian symmetry is not preserved, the inverse Fourier transform produces incomplete noise characteristics in the spatial domain, as imaginary components must be discarded. This disrupts the statistical consistency of the latent noise, leading to a deviation from the expected real Gaussian distribution and potential degradation in generative quality. In contrast, embedding with Hermitian symmetry better retains the statistical structure of the latent + +![](images/ab9bf4e0e2b5982d3489bb7c83734271773944aa4c33f486db60e72de68a140a.jpg) +Figure 4. Overview of the proposed framework and qualitative results. (Left) The key components of our approach: Symmetric Fourier Watermarking (SFW) (blue region) and Center-Aware Embedding Strategy (green region). (Right) Qualitative results of Tree-Ring, RingID, HSTR, and HSQR. Notably, RingID exhibits visible ring-like artifacts, highlighting that its high-energy pattern disrupts generative quality, unlike other merged-in-generation semantic watermarking methods that achieve a better balance between robustness and image fidelity. + +![](images/b50f6570acc8fbdaa56c28b3765c7d40fc0eb61030aa1612eec54a8483f7aa4e.jpg) + +noise, ensuring that the transformed spatial-domain signal remains closer to a real Gaussian distribution. By preserving these statistical properties, the initialization process in diffusion models becomes more stable, ultimately enhancing generative performance, as observed in Sec. 5.2. + +# 4. Methods + +# 4.1. Hermitian Symmetric Fourier Watermark + +The proposed Hermitian SFW refers to a watermark designed to satisfy the Hermitian symmetry introduced in the previous section. As shown in the left side of Fig. 4, SFW allows free usage of the half-region in the frequency domain, while the remaining half is restricted by the symmetry condition. Certain constraints must be satisfied when embedding a pattern in the free half-region under a semantic watermarking scheme: the imaginary part at the DC center must be zero, and if the signal dimensions are even, the imaginary components of the points corresponding to the Nyquist frequency must also be zero. In cases where both frequency axes have even dimensions as in this study, the points where the imaginary part must be zero are $(0,0)$ , $\left(\frac{M}{2},0\right)$ , $(0,\frac{N}{2})$ , and $\left(\frac{M}{2},\frac{N}{2}\right)$ . Instead of directly modifying the watermark pattern to comply with locations where the imaginary part must be zero, a more effective strategy is to embed it while avoiding the DC axis as much as possible, as will be introduced in Sec. 4.3.2 + +# 4.2. Center-Aware Embedding Strategy + +Existing semantic watermarking baselines [11, 48, 53] embed patterns by applying the Fourier transform to the full spatial matrix of the latent vector. However, such methods are vulnerable to spatial attacks, particularly cropping, which can lead to the loss of watermark information and reduced detection performance. To address this issue, we propose a center-aware embedding strategy, which applies + +the Fourier transform only to the central area of the spatial domain before embedding. Specifically, for a latent vector with a signal dimension of 64, we utilize the central $44 \times 44$ space. As demonstrated in Sec. 5.3.2, this design significantly improves robustness against cropping attacks on various scales. + +# 4.3. Integrating SFW and Center-Aware Design + +# 4.3.1. Application to the Baseline + +As shown in Fig. 3a, the Tree-Ring [48] baseline suffers from frequency loss. To mitigate this, we impose the Hermitian symmetry condition directly on the watermark patterns. By integrating SFW with the center-aware embedding strategy, we refine the baseline into Hermitian Symmetric Tree-Ring (HSTR). HSTR not only enhances detection performance by fully utilizing the imaginary components of the latent Fourier domain but also improves image generation quality by restoring frequency integrity. + +# 4.3.2. Hermitian Symmetric QR code + +In this section, we introduce a novel approach that extends SFW beyond baselines to QR codes, which are widely recognized for their high versatility and robust error correction capabilities [16]. + +**Embedding.** As illustrated in Fig. 3b, the Hermitian Symmetric QR Code (HSQR) watermark is constructed by splitting the QR code in half and embedding each part separately into the real and imaginary components of the free half-region in the Fourier domain. The binary pattern of the QR code is embedded using the following formulation: + +$$ +\operatorname {H S Q R} (\tilde {x}, c) = \left\{ \begin{array}{l l} + | F (\tilde {x}, c) |, & \text {i f} \operatorname {Q R} (x) = 1 \\ - | F (\tilde {x}, c) |, & \text {i f} \operatorname {Q R} (x) = 0, \end{array} \right. \tag {3} +$$ + +where: + +Table 1. Verification performance of different watermarking methods under various attacks. Bit Accuracy is used for bitstream-based methods (DwtDct, DwtDctSvd, RivaGAN, S.Sign), while TPR@1%FPR is used for semantic methods. Best performances are highlighted. Our methods show superior detection accuracy and robustness against signal processing distortions, regeneration, and cropping attacks. + +
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.8630.5720.5220.5160.6770.8590.5320.5230.5210.5190.7290.8100.637
DwtDctSvd1.0000.5550.4730.6021.0001.0000.7840.6480.5960.6440.7440.8610.742
RivaGAN0.9990.8620.9860.8210.9980.9690.9340.5700.5520.6080.9910.9950.857
S.Sign.0.9950.8940.9780.8060.9110.7210.8380.7170.7150.4780.9870.9910.836
Tree-Ring0.9570.4630.9000.5480.9340.4120.8150.5090.5360.5430.5090.7340.655
Zodiac0.9980.8430.9980.9730.9980.8800.9970.9440.9580.9720.9890.9950.962
HSTR (ours)1.0000.8991.0000.9941.0000.8060.9990.9730.9820.9971.0001.0000.971
RingID1.0000.9881.0001.0001.0000.9871.0000.9921.0001.0001.0001.0000.997
HSQR (ours)1.0000.9911.0001.0001.0000.9831.0000.9921.0001.0001.0001.0000.997
SD-PromptsDwtDct0.8190.5570.5160.5060.6850.8220.5300.5130.5120.5090.7230.7940.624
DwtDctSvd1.0000.5370.4590.6100.9990.9980.8590.6590.6200.6230.7430.8600.747
RivaGAN0.9910.8230.9630.8100.9880.9610.9150.5720.5350.5670.9800.9830.841
S.Sign.0.9940.8990.9670.7690.8880.7420.8090.6770.6710.4930.9830.9900.824
Tree-Ring0.9440.4710.8940.4660.9120.4230.8020.5090.5140.5430.4690.7490.641
Zodiac0.9980.7480.9990.9790.9990.9031.0000.9400.9750.9580.9940.9960.957
HSTR (ours)1.0000.7421.0000.9901.0000.8501.0000.9830.9870.9991.0001.0000.963
RingID1.0000.9721.0001.0001.0000.9881.0000.9961.0001.0001.0001.0000.996
HSQR (ours)1.0000.9551.0000.9991.0000.9921.0000.9960.9991.0001.0001.0000.995
DiffusionDBDwtDct0.8420.5630.5150.5090.6720.8290.5260.5130.5140.5120.7230.8010.627
DwtDctSvd0.9980.5580.4630.5930.9970.9950.8300.6580.6080.6210.7420.8600.744
RivaGAN0.9870.8390.9600.7900.9850.9370.8930.5530.5180.5560.9740.9790.831
S.Sign.0.9900.8900.9670.7870.8890.7260.8190.6900.6870.4960.9810.9860.826
Tree-Ring0.9400.4870.8890.4340.9040.3920.7990.4540.5030.4540.4990.7150.622
Zodiac0.9920.7520.9880.9330.9880.8340.9840.9110.9260.9030.9710.9850.931
HSTR (ours)0.9990.7920.9960.9810.9960.7920.9910.9680.9690.9891.0001.0000.956
RingID1.0000.9891.0001.0001.0000.9631.0000.9950.9991.0001.0001.0000.995
HSQR (ours)1.0000.9771.0000.9991.0000.9740.9990.9970.9991.0001.0001.0000.995
+ +- $x$ denotes the coordinates in the QR code domain. +- $\tilde{x}$ represents the corresponding coordinates in the embedding region of the Fourier domain. Since the QR code is split into two halves, each half is mapped to either the real or imaginary component. +- $c \in \{\mathrm{Re}, \mathrm{Im}\}$ indicates whether the embedding occurs in the real or imaginary part. +- $F(\tilde{x}, c)$ is the real-valued amplitude of the complex Gaussian noise in the Fourier domain, where a complex number is expressed as $F_{\mathrm{Re}} + iF_{\mathrm{Im}}$ , making both real and imaginary amplitudes individually real-valued. + +To maintain symmetry while avoiding numerical instability, the embedding region is positioned one pixel to the right of the vertical DC axis. Additionally, to increase the amount of embedded information and statistically reduce errors, each QR code cell is represented by multiple pixels arranged in a square pattern, e.g., $2 \times 2$ pixels. + +Detection. The detection process for semantic watermarking is performed by computing the L1 distance. In HSQR, the ground truth QR binary pattern is converted into a signed Boolean representation as follows: + +$$ +\mathrm {Q R} ^ {*} (\tilde {x}, c) = \left\{ \begin{array}{l l} + \Lambda , & \text {i f} \mathrm {Q R} (x) = 1 \\ - \Lambda , & \text {i f} \mathrm {Q R} (x) = 0, \end{array} \right. \tag {4} +$$ + +where $\Lambda$ is a fixed amplitude for binary encoding. + +# 5. Experiments + +# 5.1. Experimental Settings + +Semantic Watermarks. The Tree-Ring [48] watermark is embedded into channel 3 with a radius of 14 and HSTR adopts the same watermark parameters as its baseline. For RingID [11], we follow the original settings, embedding a tree-ring pattern with a radius range of 3-14 in channel 3, along with a Gaussian noise key in channel 0. Both SFW methods follow RingID in utilizing a Gaussian noise key in channel 0. For HSQR, we use a version 1 QR code $(21 \times 21$ cells) with a cell size of 2 pixels and error correction level H, capable of encoding up to 72 bits. The QR code is embedded in channel 3, and the encoding amplitude $\Lambda$ is set to 45, corresponding to the standard deviation of the real and imaginary components in the Fourier domain of $64 \times 64$ normal Gaussian latent vector $(\approx \sqrt{64^2 / 2})$ . Note that these experiments are conducted on $512 \times 512$ images. At higher resolutions, a larger embedding space allows for increased capacity through higher QR code versions that encode more bits, as well as the use of stronger watermarking parameters, enabling greater robustness and adaptability. Additionally, + +Table 2. Identification accuracy for different watermarking methods, evaluated across multiple attack conditions. Perfect Match Rate is used for bitstream-based methods. The best performance for each item is highlighted with shading. HSQR achieves state-of-the-art performance, while HSTR consistently outperforms other Gaussian radius-based semantic baselines, particularly under cropping attacks. + +
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.4660.0440.0000.0000.0380.4420.0000.0000.0000.0000.0000.0000.083
DwtDctSvd1.0000.0440.0190.0000.9990.9980.0370.0000.0000.0030.0000.0000.258
RivaGAN0.9740.2600.7720.0230.9610.6860.3480.0000.0000.0000.8520.9090.482
S.Sign.0.8730.1770.5630.0000.0360.0100.0070.0000.0000.0000.7090.8020.265
Tree-Ring0.3030.0870.2070.0720.2560.0300.1620.0830.0720.0540.0090.0330.114
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.0000.7140.9990.8860.9980.4600.9720.8330.8310.9711.0001.0000.889
RingID1.0000.8751.0000.9751.0000.9190.9960.9780.9700.9980.8740.9780.964
HSQR (ours)1.0000.9581.0000.9941.0000.9010.9990.9800.9870.9991.0001.0000.985
SD-PromptsDwtDct0.2850.0240.0000.0000.0170.2760.0000.0000.0000.0000.0000.0000.050
DwtDctSvd0.9930.0280.0110.0000.9820.9790.0850.0000.0000.0070.0000.0000.257
RivaGAN0.8780.2130.6130.0090.8570.6570.3040.0000.0000.0000.7360.7680.420
S.Sign.0.8130.2630.4200.0000.0210.0150.0060.0000.0000.0000.5760.7160.236
Tree-Ring0.2880.0940.1890.0510.2350.0340.1590.0790.0760.0560.0120.0410.110
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.000
RingID1.0000.8851.0000.9760.9980.8860.9930.9800.9730.9950.8760.9810.962
HSQR (ours)1.0000.9301.0000.9941.0000.9420.9990.9910.9971.0001.0001.0000.988
DiffusionDBDwtDct0.3570.0370.0000.0000.0340.3200.0000.0000.0000.0000.0000.0000.062
DwtDctSvd0.9900.0360.0190.0000.9750.9590.0810.0000.0000.0010.0000.0000.255
RivaGAN0.8580.2130.6250.0200.8480.6150.2210.0000.0000.0000.7560.7800.411
S.Sign.0.7980.2070.4720.0000.0270.0050.0050.0000.0000.0000.6430.7380.241
Tree-Ring0.2800.0950.1900.0590.2330.0370.1450.0810.0720.0500.0130.0390.108
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)0.9960.7210.9920.8540.9890.5630.9580.8300.8210.9520.9960.9960.889
RingID1.0000.8951.0000.9470.9960.8710.9920.9680.9580.9900.8750.9840.956
HSQR (ours)1.0000.9541.0000.9881.0000.9060.9980.9820.9910.9941.0001.0000.984
+ +the post-hoc semantic watermarking method, Zodiac [53], is included in our comparative experiments. + +Model Setup and Datasets. The diffusion model used for the experiments is Stable Diffusion v2-1-base [37], configured to generate images at a resolution of $512 \times 512$ with a CFG scale of 7.5. Both generation and inversion steps of the DDIM scheduler [41] are set to 50. The datasets used for image generation include 5,000 captions from the MS-COCO-2017 training set [29], 1,001 sampled prompts from the 2M subset of DiffusionDB [46], and 8,192 prompts from the Stable-Diffusion-Prompts test set [40]. For watermark detection, the number of watermarked images used for each dataset is set to 1,000. Specifically, verification is conducted on 1,000 pairs of watermarked and unwatermarked images, while identification is performed on 1,000 watermarked images. + +Attack Methods. To evaluate the robustness of our method against various distortions, we apply 11 different attacks, categorized into three major types. + +- Signal Processing Attacks: These include brightness adjustment (ranging from 0 to 7), contrast adjustment (factor of 0.5), JPEG compression (quality factor of 25), Gaussian blur (radius of 5), Gaussian noise ( $\sigma = 0.05$ ), and BM3D denoising ( $\sigma = 0.1$ ). + +- Regeneration Attacks: We apply two VAE-based image compression models, Bmshj18 [4] (VAE-B) and Cheng20 [8] (VAE-C), both at quality level 3, as well as a diffusion-based regeneration attack by Zhao et al. [56] (referred to as Diff.) using 60 denoising steps. +- Cropping Attacks: We apply center crop (C.C.) with a crop scale of 0.5 and random crop (R.C.) with a crop scale of 0.7, where the crop scale represents the ratio of the cropped image area to the original image. + +Evaluation Metrics. The $L_{1}$ distance is computed only within the key region of the complex Fourier domain of latent noise, and unless otherwise specified, 2,048 keys are used for identification accuracy. For verification, semantic watermarking methods use True Positive Rate at $1\%$ False Positive Rate (TPR@1%FPR) as the evaluation metric. To ensure a fair comparison with bitstream-based approaches such as DwtDct, DwtDctSvd [12], RivaGAN [52], and Stable Signature (S.Sign.) [22], we measure Bit Accuracy for these methods. For identification, where the goal is to match the exact message, Perfect Match Rate is used as the metric for bitstream-based approaches. To evaluate the generation quality, FID [24] is measured between watermarked images and 5,000 MS-COCO ground truth images. We also calculate the CLIP score [34] using the OpenCLIP-ViT/G + +Table 3. Generative quality evaluation of watermarking methods based on FID (MS-COCO ground truth) and CLIP score. The best performance for each item is highlighted with shading, while bold text specifically marks the low CLIP score in RingID. Our proposed methods preserve frequency integrity, achieving the best balance between watermark robustness and generative performance, whereas RingID introduces visible artifacts, compromising perceptual quality. Vrf. and Idf. denote the average detection performance in verification and identification tasks, respectively. + +
Semantic MethodsFID ↓CLIP ↑Vrf.Idf.
Merged in GenerationTree-Ring26.4180.3260.6550.114
RingID27.0520.3240.9970.964
HSTR (ours)25.0620.3290.9710.889
HSQR (ours)24.8950.3300.9970.985
+ +model [9] to determine how closely the generated image aligns with the input prompt. + +# 5.2. Comparison with Baselines + +In this section, we compare the detection performance in both verification and identification tasks, as well as the generation quality of different watermarking methods. + +Enhanced Detection Robustness. According to Tab. 1, our proposed methods consistently achieve top-tier detection performance across various attacks. A key observation is that non-semantic watermarking methods [12, 22, 52] exhibit significant degradation under regeneration attacks, highlighting their vulnerability. Among Gaussian radius-based pattern methods (Tree-Ring, Zodiac, and HSTR), HSTR outperforms corresponding baselines in most cases, demonstrating its effectiveness in enhancing detection robustness while maintaining the fundamental tree-ring structure. In contrast, RingID leverages high-energy signed constant ring patterns, securing strong detection performance. However, as acknowledged by the authors, this method introduces noticeable ring-like artifacts in generated images, disrupting the balance between watermark robustness and generative quality. A detailed quantitative evaluation of this trade-off is presented in the following section. Meanwhile, Zodiac suffers from time-consuming processing, taking several minutes to apply the watermark—requiring 7.36 minutes per image on MS-COCO dataset—which further limits its practicality in real-world applications. In contrast, our methods following the merged-in-generation scheme introduce no additional processing time, ensuring seamless watermark embedding. + +In Tab. 2, we compare identification results, which reveal an even greater disparity in detection performance across different watermarking methods. Gaussian radius-based pattern methods, except for HSTR, perform poorly in identification tasks, with Zodiac—designed solely for verification—failing entirely across all scenarios (achieving zero identification accuracy). Further analysis on scalability + +under different message capacities is provided in Sec. 5.3.3. Notably, our methods exhibit strong resilience against cropping attacks in both tasks, reinforcing the robustness of center-aware embedding. HSQR achieves state-of-the-art identification accuracy across all datasets, further establishing its dominance in watermark retrieval. + +Balance with Generative Quality. Tab. 3 presents the generative quality of different semantic watermarking methods following the merged-in-generation scheme, evaluated using FID and CLIP score. RingID, which deviates from a Gaussian distribution by embedding high-energy perturbations, exhibits the worst generative performance, particularly reflected in its low CLIP score. This decline is attributed to the noticeable ring-like artifacts, as discussed earlier. In contrast, our proposed methods, which preserve frequency integrity via SFW, achieve the top two FID scores, demonstrating a better balance between watermark robustness and generative quality. + +Expanding Frequency Utilization in Latent Watermarking. Traditional image-domain watermarking methods prioritize low-mid frequency embedding to resist compression and filtering attacks. This approach has also been adopted in latent diffusion-based semantic watermarking methods [11, 48, 53] without explicitly considering the differences between pixel-space and latent-space transformations. However, our findings suggest that such frequency constraints may not be necessary in the latent space, where watermark retrieval involves latent encoding and DDIM inversion, altering the impact of frequency perturbations. This distinction is evident in HSQR, which achieves state-of-the-art detection performance while utilizing nearly the entire frequency spectrum. Instead of relying on specific frequency bands, HSQR leverages structured statistical encoding, distributing watermark information across multiple pixels ( $2 \times 2$ per QR cell). This redundancy enhances robustness against distortions, enabling effective retrieval even with broader frequency usage. These results indicate that latent-space watermarking is not bounded by traditional low-mid frequency constraints. Future semantic watermarking strategies can benefit from statistically robust encoding methods, allowing for effective watermark retrieval across a wider frequency range without compromising detection accuracy. + +# 5.3. Ablation Study + +# 5.3.1. Impact of SFW on Detection Performance + +We evaluate the impact of Hermitian SFW by comparing detection performance across four cases (Tab. 4). Without SFW, using both real and imaginary components for detection (Case A) results in the lowest performance due to frequency degradation, while restricting detection to the real component (Case B) improves accuracy. In contrast, applying SFW (Cases C and D) preserves frequency integrity, + +Table 4. Ablation study on detection performance based on frequency integrity ( $\sqrt{}$ or $\times$ ) and the number of detection region usage (1: real only, 2: real & imaginary). Vrf. and Idf. represent average detection performance in verification (TPR@1% FPR) and identification (accuracy), respectively. $\Delta L_1^*$ denotes the normalized $\Delta L_1$ metric, indicating detection effectiveness. + +
CaseMethodsFreq. Int.# Det.Vrf.Idf.ΔL1*↑
ATree-Ring×20.6530.1140.232
BTree-Ring×10.8050.4160.368
CHSTR (ours)✓10.9360.7750.471
DHSTR (ours)✓20.9710.8890.476
+ +![](images/56d75d67a6ba238f9798ac25f3ba8dde7f8e62510bcad250ae2eb3872e8feb0a.jpg) +Figure 5. Identification accuracy under center crop and random crop attacks at different crop scales. HSTR and HSQR maintain higher accuracy compared to RingID, demonstrating improved robustness against cropping. + +![](images/3d4e78eb6f8b84fdb847e3363e8f009b0c427e3a6510d7addd8a6827ca628aae.jpg) + +![](images/bf4440f4305343d37f581f8f612d9cbd603c4ac20fff19d24a148a5be3abc9c2.jpg) +Figure 6. Identification accuracy across watermark message capacities. HSQR remains nearly perfect, while RingID and HSTR degrade at higher capacities. Tree-Ring and Zodiac fail to scale effectively. + +leading to a significant boost in detection performance. Notably, HSTR with SFW and full complex detection (Case D) achieves the highest accuracy, demonstrating that leveraging both frequency components maximizes retrieval effectiveness. These results confirm that SFW enables robust semantic watermarking by maintaining frequency integrity and fully utilizing the Fourier domain for detection. + +# 5.3.2. Robustness to Cropping Attacks + +We assess the robustness of center-aware embedding against center crop and random crop attacks, measuring identification accuracy across different crop scales (Fig. 5). As crop scale decreases, accuracy declines for all methods. However, RingID exhibits a steeper drop, indicating higher vulnerability to cropping. In contrast, HSTR and HSQR degrade more gradually, demonstrating improved resilience. While extreme cropping (scale 0.2) significantly affects all methods, our center-aware design consistently outperforms RingID, confirming its effectiveness in preserving watermark information under cropping distortions. + +# 5.3.3. Impact of Capacity on Identification + +We evaluate identification accuracy across different watermark message capacities (64 to 8,192) for five methods (Fig. 6). Zodiac's performance collapses, while Tree-Ring exhibits rapid performance degradation, becoming nearly unusable at higher capacities. RingID, HSTR, and HSQR remain robust, maintaining $80\%+$ accuracy, though HSTR declines faster, and RingID starts degrading from 2,048 onward. HSQR demonstrates the highest scalability, retaining near-perfect accuracy even at the largest capacity, confirming its superior robustness in high-capacity scenarios. + +# 6. Conclusion + +We have introduced Hermitian SFW, a novel approach to semantic watermarking in the latent diffusion model framework. Unlike existing methods that fail to preserve frequency integrity, our approach ensures that watermark embeddings maintain a consistent statistical structure in the latent noise distribution. This is achieved through Hermitian symmetry enforcement, which preserves frequency components and enhances detection and generation quality. Additionally, we have proposed center-aware embedding, which significantly improves robustness against cropping attacks by strategically placing watermarks in a spatially resilient region of the latent representation. Through comprehensive experiments, we demonstrated that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while also maintaining superior image generation quality, as shown by FID and CLIP scores. Our study highlights the importance of frequency integrity in Fourier-based watermarking and challenges the assumption that semantic watermarking must be confined to low-mid frequency bands. Experimental results confirm that a properly structured frequency-domain watermark can be effectively embedded and retrieved across the entire frequency spectrum without compromising generative quality. Future work includes exploring adaptive embedding strategies to further enhance robustness against adversarial attacks and extreme distortions, as well as extending our method to more diverse generative architectures beyond LDMs. + +# Acknowledgements + +This research was supported by the 2024 AI Semiconductor Application/Demonstration Support Program of the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) with Markany Co., Ltd. as the lead organization, and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2025. + +# References + +[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 1 +[2] Dhruv Arya. A survey of frequency and wavelet domain digital watermarking techniques. 2 +[3] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. Promark: Proactive diffusion watermarking for causal attribution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10802-10811, 2024. 1, 2 +[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 6 +[5] Mauro Barni, Franco Bartolini, Vito Cappellini, and Alessandro Piva. A dct-domain system for robust image watermarking. Signal processing, 66(3):357-372, 1998. 1, 2 +[6] Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet-based watermarking through pixel-wise masking. IEEE transactions on image processing, 10(5): 783-791, 2001. 1, 2 +[7] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 1, 2 +[8] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 6 +[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 7 +[10] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 2 +[11] Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. arXiv preprint arXiv:2404.14055, 2024. 2, 3, 4, 5, 7 + +[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 1, 2, 6, 7 +[13] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE TIP, 6(12):1673-1687, 1997. 1, 2 +[14] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for data copyright protection against generative diffusion models. ACM SIGKDD Explorations Newsletter, 26(2): 60-75, 2025. 2 +[15] DEJEY and RS Rajesh. An improved wavelet domain digital watermarking for image protection. International journal of wavelets, multiresolution and information processing, 8(01): 19-31, 2010. 1, 2 +[16] Denso Wave Incorporated. QR Code Essentials. Denso Wave, 2011. Available at: https://www.qrcode.com/en/about/standards.html.4 +[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[18] Han Fang, Dongdong Chen, Qidong Huang, Jie Zhang, Zehua Ma, Weiming Zhang, and Nenghai Yu. Deep template-based watermarking. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1436-1451, 2020. 1 +[19] Han Fang, Zhaoyang Jia, Yupeng Qiu, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. De-end: decoder-driven watermarking network. IEEE Transactions on Multimedia, 25: 7571-7581, 2022. 1 +[20] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2 +[21] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 1 +[22] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 2, 6, 7 +[23] Yiyang Guo, Ruizhe Li, Mude Hui, Hanzhong Guo, Chen Zhang, Chuangjian Cai, Le Wan, and Shangfei Wang. Freqmark: Invisible image watermarking via frequency based optimization in latent space. arXiv preprint arXiv:2410.20824, 2024. 2 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 + +[26] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 1 +[27] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024. 1, 2 +[28] Chunlei Li, Zhaoxiang Zhang, Yunhong Wang, Bin Ma, and Di Huang. Dither modulation of significant amplitude difference for wavelet based robust watermarking. Neurocomputing, 166:404-415, 2015. 1, 2 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[30] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 1 +[31] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 1, 2 +[32] A Miyazaki and A Okamoto. Analysis of watermarking systems in the frequency domain and its application to design of robust watermarking systems. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205), pages 506-509. IEEE, 2001. 2 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[36] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1, 2 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 6 +[38] Joseph JK Ō Ruanaidh and Thierry Pun. Rotation, scale and translation invariant spread spectrum digital image watermarking. Signal processing, 66(3):303-317, 1998. 1, 2 + +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[40] Gustavo Santana. Gustavosta: Stable-diffusion-prompts. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2022.6 +[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3, 6 +[42] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1 +[43] Jun Tian. Reversible data embedding using a difference expansion. IEEE transactions on circuits and systems for video technology, 13(8):890-896, 2003. 1, 2 +[44] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 1, 2 +[45] Wenbo Wan, Jun Wang, Yunming Zhang, Jing Li, Hui Yu, and Jiande Sun. A comprehensive survey on robust image watermarking. Neurocomputing, 488:226-247, 2022. 2 +[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 6 +[47] Xiu-mei Wen, Wei Zhao, and Fan-xing Meng. Research of a digital image watermarking algorithm resisting geometrical attacks in fourier domain. In 2009 International Conference on Computational Intelligence and Security, pages 265-268. IEEE, 2009. 2 +[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 7 +[49] Raymond B Wolfgang, Christine I Podilchuk, and Edward J Delp. Perceptual watermarks for digital images and video. Proceedings of the IEEE, 87(7):1108-1126, 1999. 1, 2 +[50] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 2 +[51] Guokai Zhang, Lanjun Wang, Yuting Su, and An-An Liu. A training-free plug-and-play watermark framework for stable diffusion. arXiv preprint arXiv:2404.05607, 2024. 1, 2 +[52] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 1, 6, 7 + +[53] Lijun Zhang, Xiao Liu, Antoni Martin, Cindy Bearfield, Yuriy Brun, and Hui Guan. Attack-resilient image watermarking using stable diffusion. Advances in Neural Information Processing Systems, 37:38480-38507, 2025. 2, 3, 4, 6, 7 +[54] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 1, 2 +[55] Xuandong Zhao, Kexun Zhang, Yu-Xiang Wang, and Lei Li. Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. 2023. 2 +[56] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai. Advances in Neural Information Processing Systems, 37:8643-8672, 2025. 2, 6 +[57] Yimeng Zhao, Chengyou Wang, Xiao Zhou, and Zhiliang Qin. Dari-mark: Deep learning and attention network for robust image watermarking. Mathematics, 11(1):209, 2022. 1 +[58] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 1 \ No newline at end of file diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/images.zip b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f8cd0b1151a2d03935ce958fa4b29a5f3d7fc43 --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8abe10af38ca26403bf3f2e4a3771f7f9ed3c26eedc8137a8d623c4320592d12 +size 813323 diff --git a/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/layout.json b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..219ba066a9deb7b0cfcc77d0b50e79963f8e3fec --- /dev/null +++ b/2025/Semantic Watermarking Reinvented_ Enhancing Robustness and Generation Quality with Fourier Integrity/layout.json @@ -0,0 +1,7369 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "text", + "content": "Semantic Watermarking Reinvented: Enhancing Robustness and Generation Quality with Fourier Integrity" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 224, + 161, + 386, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 161, + 386, + 175 + ], + "spans": [ + { + "bbox": [ + 224, + 161, + 386, + 175 + ], + "type": "text", + "content": "Sung Ju Lee Nam Ik Cho" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 189, + 445, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 189, + 445, + 204 + ], + "spans": [ + { + "bbox": [ + 165, + 189, + 445, + 204 + ], + "type": "text", + "content": "Dept. of ECE & INMC, Seoul National University, Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 205, + 407, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 205, + 407, + 217 + ], + "spans": [ + { + "bbox": [ + 200, + 205, + 407, + 217 + ], + "type": "text", + "content": "thomas11809@snu.ac.kr, nicho@snu.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 271, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 594 + ], + "type": "text", + "content": "Semantic watermarking techniques for latent diffusion models (LDMs) are robust against regeneration attacks, but often suffer from detection performance degradation due to the loss of frequency integrity. To tackle this problem, we propose a novel embedding method called Hermitian Symmetric Fourier Watermarking (SFW), which maintains frequency integrity by enforcing Hermitian symmetry. Additionally, we introduce a center-aware embedding strategy that reduces the vulnerability of semantic watermarking due to cropping attacks by ensuring robust information retention. To validate our approach, we apply these techniques to existing semantic watermarking schemes, enhancing their frequency-domain structures for better robustness and retrieval accuracy. Extensive experiments demonstrate that our methods achieve state-of-the-art verification and identification performance, surpassing previous approaches across various attack scenarios. Ablation studies confirm the impact of SFW on detection capabilities, the effectiveness of the center-aware embedding against cropping, and how message capacity influences identification accuracy. Notably, our method achieves the highest detection accuracy while maintaining superior image fidelity, as evidenced by FID and CLIP scores. Conclusively, our proposed SFW is shown to be an effective framework for balancing robustness and image fidelity, addressing the inherent trade-offs in semantic watermarking. Code available at github.com/thomas11809/SFWMark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "With the advancement of diffusion generative models [17, 25, 33, 35, 39, 41], their generated images are being increasingly used across various fields, including creative works, entertainment, and advertisement. In particular, the open-source release of large-scale language-image (LLI) models like Stable Diffusion [37] has led to an exponential growth" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 244, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 316, + 244, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 244, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 555, + 384 + ], + "type": "image", + "image_path": "5bbcfca8fcbdc3cc4228a5a7344a6d4682b897c62e1a5eec1f2b6ad07fd5a977.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 394, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 483 + ], + "type": "text", + "content": "Figure 1. Summary of watermarking performance across different semantic watermarking methods, as detailed in Sec. 5.2. All methods follow the merged-in-generation scheme with no additional processing time. Verification is evaluated using TPR@1%FPR (True Positive Rate at " + }, + { + "bbox": [ + 313, + 394, + 555, + 483 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 394, + 555, + 483 + ], + "type": "text", + "content": " False Positive Rate), while identification is assessed by Identification Accuracy (Perfect Match Rate). The proposed approaches achieve the best balance between detection robustness and image fidelity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "content": "in script-based image generation and editing technologies, resulting in a surge of generative content. This has raised new concerns, such as the copyright of AI-generated content and the tracking of images created with malicious intent. As a solution to these problems, the technique of embedding invisible watermarks into content has been proposed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "type": "text", + "content": "Digital content watermarking has been extensively studied through both classical signal processing techniques [5, 6, 12, 13, 15, 28, 38, 43, 44, 49] and deep learning-based approaches [1, 18, 19, 21, 26, 30, 42, 52, 57, 58]. Recently, watermarking techniques that directly intervene in the generation process of diffusion models [3, 7, 10, 20, 22, 27, 31, 36, 51, 54] have been explored; however, these approaches require additional setup or configuration, adding more complexity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Several studies have explored embedding watermarks di" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18759" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "rectly into the latent representation, eliminating the need for external models. These include a steganographic approach that constructs latent noise with stream cipher-randomized bits [50] and semantic watermarking methods that embed geometric patterns in the Fourier frequency domain of the latent representation [11, 48, 53]. Meanwhile, research on pixel-level perturbation-based watermarking has highlighted its vulnerability to regeneration attacks [55, 56], demonstrating that semantic watermarking serves as a more robust alternative against such attacks. However, the aforementioned semantic watermarking methods in the latent Fourier domain, which are the focus of this paper, suffer from degraded detection accuracy and generative quality due to their lack of frequency integrity preservation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 240, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 477 + ], + "type": "text", + "content": "In this context, we propose a novel semantic watermarking framework that enhances both detection robustness and image quality by preserving frequency integrity in the latent Fourier domain. Our method is named Hermitian Symmetric Fourier Watermarking (SFW), which ensures that watermark embeddings maintain statistical consistency with the latent noise distribution, leading to improved retrievability and stability in generative models. Additionally, we incorporate a center-aware embedding strategy that enhances robustness against cropping attacks by embedding watermarks in a spatially resilient region of the latent representation. We comprehensively evaluate our method across various attack scenarios, including signal processing distortions, regeneration attacks, and cropping attacks. Experimental results, as illustrated in Fig. 1, demonstrate that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while simultaneously maintaining superior generative quality, as evidenced by FID and CLIP score evaluations. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 479, + 295, + 574 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 479, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 514 + ], + "type": "text", + "content": "- We propose Hermitian SFW to ensure frequency integrity, leading to improved watermark detection and generative quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 515, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 539 + ], + "type": "text", + "content": "- We introduce center-aware embedding, which significantly enhances robustness against cropping attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 295, + 574 + ], + "type": "text", + "content": "- We present extensive evaluations demonstrating that our approach outperforms existing baselines in detection accuracy and generative quality." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 586, + 146, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 146, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 146, + 597 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Digital Watermarking. Digital watermarking aims to achieve a balance between high embedding capacity and visual quality when inserting invisible watermarks while also ensuring robustness against various attacks [45]. This field originally began with the adoption of classical signal processing techniques. Embedding the watermark in the spatial domain is a straightforward and intuitive watermark insertion [43, 44], but it tends to be less robust against attacks such as filtering or compression. On the other hand," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 208 + ], + "type": "image", + "image_path": "79831f3006d8a799d55b67df55ba2fe6115882fb6449f42436d3e6ef4d026d16.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 215, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 553, + 238 + ], + "type": "text", + "content": "Figure 2. Overview of the semantic watermarking process in the latent Fourier domain using the merged-in-generation scheme." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 306 + ], + "type": "text", + "content": "studies utilizing the frequency domain have demonstrated resilience against JPEG compression by embedding watermarks in the low and middle frequency bands [2, 5, 6, 12, 13, 15, 28, 32, 38, 47, 49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 307, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 461 + ], + "type": "text", + "content": "Watermarks for Latent Diffusion Models. Latent diffusion models (LDMs), such as Stable Diffusion [37], conduct the diffusion process in a lower-dimensional latent space rather than directly on high-resolution images. Recent research has focused on techniques for integrating watermarks into the latent diffusion process to ensure traceability and robustness. However, many of these approaches rely on model fine-tuning [3, 10, 20, 22, 27] or require separately trained encoders and decoders [7, 10, 14, 20, 22, 23, 27, 31, 36, 51, 54] to facilitate watermark embedding and detection. These dependencies constrain the flexibility of watermarking, limiting their applicability across diverse models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 653 + ], + "type": "text", + "content": "Several studies have explored embedding semantic watermarks in the Fourier domain of latent vectors using a merged-in-generation scheme. Wen et al. [48] introduced a tree-ring-shaped watermark constructed with Gaussian-distributed radii, while Ci et al. [11] proposed a modified pattern with high-energy signed constant rings for watermark identification, along with an additional random noise key embedded in a separate channel. On the other hand, Zodiac [53] followed a post-hoc approach, embedding a tree-ring pattern into generated images through multiple iterations of latent vector optimization and diffusion-based synthesis. However, in addition to its high computational cost, this method suffers from low practicality, as it relies on extensive linear interpolation between the original and optimized images to artificially improve visual quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "In these methods, when performing inverse Fourier transform, the imaginary component in the spatial domain is discarded, leading to a distorted frequency representation. As a result, the real component of the key region retains only partial information, while the imaginary component" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18760" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 209 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 209 + ], + "type": "image", + "image_path": "dc28e9aa714e81edfecb78c3e654df8afd03e20873e931db02496f0b9bbf1ed4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 218, + 291, + 229 + ], + "lines": [ + { + "bbox": [ + 59, + 218, + 291, + 229 + ], + "spans": [ + { + "bbox": [ + 59, + 218, + 291, + 229 + ], + "type": "text", + "content": "Figure 3. Examples of various semantic watermarking patterns." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 300 + ], + "type": "text", + "content": "is almost entirely lost, creating an empty key region in the frequency domain. Since detection relies on analyzing this incomplete key region, the process is inherently limited, resulting in suboptimal retrieval performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 468 + ], + "type": "text", + "content": "Fig. 2 illustrates the overall pipeline of semantic watermarking with the merged-in-generation scheme. The embedding process begins with the Fourier transform applied to the latent noise, generating the latent Fourier representation. A watermark key is then embedded into a designated key region, followed by the inverse Fourier transform, producing the watermarked latent vector. Finally, text-guided image generation synthesizes the watermarked image. For detection, the process starts with a clean or attacked image, from which the latent query is obtained via DDIM inversion [41]. The presence of a watermark is determined by analyzing the query key region in the latent representation. Further details on detection tasks and evaluation metrics are provided in Sec. 3.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 481, + 140, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 140, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 140, + 494 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 162, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 162, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 162, + 514 + ], + "type": "text", + "content": "3.1. Task Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "text", + "content": "The performance of watermark detection is evaluated following RingID's formulation for watermark verification and identification tasks [11]. The metric used to measure the distance " + }, + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 520, + 295, + 579 + ], + "type": "text", + "content": " distance, calculated specifically for the key region where the watermark is embedded." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": "Verification. The objective of verification is to determine whether a watermark is present in an image by analyzing the distance between the reference key and the key region of the query in the latent Fourier domain. Let " + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": " denote the watermarked latent Fourier key, and " + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\hat{u}" + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": " denote the unwatermarked latent Fourier null key. Verification is based on comparing the distances between the reference key " + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": " and the watermarked/unwatermarked keys, i.e., " + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "inline_equation", + "content": "d(\\hat{w}, w) \\neq d(\\hat{u}, w)" + }, + { + "bbox": [ + 55, + 581, + 295, + 700 + ], + "type": "text", + "content": ". Performance is assessed using statistical metrics derived from the ROC curve, considering different distance thresholds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Identification. In identification, given that a watermark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "is already embedded, the task is to accurately determine the embedded information. This is achieved by computing the distance between the watermarked key " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and multiple reference keys " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". Performance is evaluated based on the accuracy of the estimated message index, defined as " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\hat{i} = \\arg \\min_{i}d(\\hat{w},w_{i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 151, + 549, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 549, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 549, + 163 + ], + "type": "text", + "content": "3.2. Fourier Considerations for Real Latent Noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 168, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 203 + ], + "type": "text", + "content": "This section introduces the mathematical properties of the Fourier domain and highlights important considerations when embedding semantic watermarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 251 + ], + "type": "text", + "content": "Hermitian Symmetry for Real Signals. To obtain a real-valued signal after inverse Fourier transform, modifications in the frequency domain must maintain Hermitian symmetry about the DC center:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 376, + 258, + 553, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 258, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 376, + 258, + 553, + 273 + ], + "type": "interline_equation", + "content": "F [ M - k, N - l ] = \\overline {{F [ k , l ]}}, \\tag {1}", + "image_path": "96926b2e3f4b5edce57a52cc1bbf1660dd804f873fc4af9da9a676a3b181189d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": " represents the discrete Fourier transform, " + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": " denote the dimensions of the signal in the spatial domain along the row and column axes, respectively. Failure to preserve this symmetry introduces undesired imaginary components in the spatial domain, which are incompatible with the real-valued latent noise required for diffusion models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 351, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 494 + ], + "type": "text", + "content": "Impact of Ignoring Hermitian Symmetry. In existing baselines [11, 48, 53], this issue is simply addressed by discarding the imaginary component, resulting in a loss of frequency integrity. As illustrated in Fig. 3a, this disruption affects both real and imaginary components of the original watermark pattern. While the real component deviates from its intended structure due to frequency distortion, the imaginary component suffers even greater degradation, often becoming entirely empty. This significantly limits detection performance by preventing the effective use of frequency information in the imaginary domain. The impact of the frequency loss is further examined in Sec. 5.3.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 495, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 543 + ], + "type": "text", + "content": "Preservation of Gaussianity. A real Gaussian noise signal in the spatial domain transforms into a complex Gaussian noise signal in the frequency domain while maintaining its statistical properties:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 548, + 553, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 548, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 318, + 548, + 553, + 563 + ], + "type": "interline_equation", + "content": "f [ m, n ] \\sim \\mathcal {N} (0, \\sigma^ {2}) \\Rightarrow F [ k, l ] \\sim \\mathcal {C N} (0, M N \\sigma^ {2}), \\tag {2}", + "image_path": "eb7deb05265a1e55e5f324232bf18336e202bbf1ba175d9902d4a8c30e5d83d8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " represents the real-valued Gaussian noise signal in the spatial domain. When embedding watermarks in the frequency domain, some perturbation is inevitably introduced, altering the original distribution of complex Gaussian noise. However, if Hermitian symmetry is not preserved, the inverse Fourier transform produces incomplete noise characteristics in the spatial domain, as imaginary components must be discarded. This disrupts the statistical consistency of the latent noise, leading to a deviation from the expected real Gaussian distribution and potential degradation in generative quality. In contrast, embedding with Hermitian symmetry better retains the statistical structure of the latent" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18761" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 312, + 205 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 312, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 312, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 312, + 205 + ], + "type": "image", + "image_path": "ab9bf4e0e2b5982d3489bb7c83734271773944aa4c33f486db60e72de68a140a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed framework and qualitative results. (Left) The key components of our approach: Symmetric Fourier Watermarking (SFW) (blue region) and Center-Aware Embedding Strategy (green region). (Right) Qualitative results of Tree-Ring, RingID, HSTR, and HSQR. Notably, RingID exhibits visible ring-like artifacts, highlighting that its high-energy pattern disrupts generative quality, unlike other merged-in-generation semantic watermarking methods that achieve a better balance between robustness and image fidelity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 554, + 203 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 203 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 203 + ], + "type": "image", + "image_path": "b50f6570acc8fbdaa56c28b3765c7d40fc0eb61030aa1612eec54a8483f7aa4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 297, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 297, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 297, + 338 + ], + "type": "text", + "content": "noise, ensuring that the transformed spatial-domain signal remains closer to a real Gaussian distribution. By preserving these statistical properties, the initialization process in diffusion models becomes more stable, ultimately enhancing generative performance, as observed in Sec. 5.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 117, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 117, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 117, + 361 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 277, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 277, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 277, + 382 + ], + "type": "text", + "content": "4.1. Hermitian Symmetric Fourier Watermark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "content": "The proposed Hermitian SFW refers to a watermark designed to satisfy the Hermitian symmetry introduced in the previous section. As shown in the left side of Fig. 4, SFW allows free usage of the half-region in the frequency domain, while the remaining half is restricted by the symmetry condition. Certain constraints must be satisfied when embedding a pattern in the free half-region under a semantic watermarking scheme: the imaginary part at the DC center must be zero, and if the signal dimensions are even, the imaginary components of the points corresponding to the Nyquist frequency must also be zero. In cases where both frequency axes have even dimensions as in this study, the points where the imaginary part must be zero are " + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\left(\\frac{M}{2},0\\right)" + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "inline_equation", + "content": "(0,\\frac{N}{2})" + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\left(\\frac{M}{2},\\frac{N}{2}\\right)" + }, + { + "bbox": [ + 55, + 387, + 296, + 602 + ], + "type": "text", + "content": ". Instead of directly modifying the watermark pattern to comply with locations where the imaginary part must be zero, a more effective strategy is to embed it while avoiding the DC axis as much as possible, as will be introduced in Sec. 4.3.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 243, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 243, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 243, + 624 + ], + "type": "text", + "content": "4.2. Center-Aware Embedding Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Existing semantic watermarking baselines [11, 48, 53] embed patterns by applying the Fourier transform to the full spatial matrix of the latent vector. However, such methods are vulnerable to spatial attacks, particularly cropping, which can lead to the loss of watermark information and reduced detection performance. To address this issue, we propose a center-aware embedding strategy, which applies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "the Fourier transform only to the central area of the spatial domain before embedding. Specifically, for a latent vector with a signal dimension of 64, we utilize the central " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "44 \\times 44" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": " space. As demonstrated in Sec. 5.3.2, this design significantly improves robustness against cropping attacks on various scales." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 539, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 539, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 539, + 371 + ], + "type": "text", + "content": "4.3. Integrating SFW and Center-Aware Design" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 457, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 457, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 457, + 387 + ], + "type": "text", + "content": "4.3.1. Application to the Baseline" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 499 + ], + "type": "text", + "content": "As shown in Fig. 3a, the Tree-Ring [48] baseline suffers from frequency loss. To mitigate this, we impose the Hermitian symmetry condition directly on the watermark patterns. By integrating SFW with the center-aware embedding strategy, we refine the baseline into Hermitian Symmetric Tree-Ring (HSTR). HSTR not only enhances detection performance by fully utilizing the imaginary components of the latent Fourier domain but also improves image generation quality by restoring frequency integrity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 476, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 476, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 476, + 517 + ], + "type": "text", + "content": "4.3.2. Hermitian Symmetric QR code" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "type": "text", + "content": "In this section, we introduce a novel approach that extends SFW beyond baselines to QR codes, which are widely recognized for their high versatility and robust error correction capabilities [16]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 569, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 641 + ], + "type": "text", + "content": "**Embedding.** As illustrated in Fig. 3b, the Hermitian Symmetric QR Code (HSQR) watermark is constructed by splitting the QR code in half and embedding each part separately into the real and imaginary components of the free half-region in the Fourier domain. The binary pattern of the QR code is embedded using the following formulation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 661, + 555, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 661, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 337, + 661, + 555, + 693 + ], + "type": "interline_equation", + "content": "\\operatorname {H S Q R} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 1 \\\\ - | F (\\tilde {x}, c) |, & \\text {i f} \\operatorname {Q R} (x) = 0, \\end{array} \\right. \\tag {3}", + "image_path": "cb252898df7493235b9ebe1984dd11c072f1b54a6829b73729796585ae70c134.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 702, + 345, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 345, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 345, + 712 + ], + "type": "text", + "content": "where:" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18762" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 114, + 549, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Verification performance of different watermarking methods under various attacks. Bit Accuracy is used for bitstream-based methods (DwtDct, DwtDctSvd, RivaGAN, S.Sign), while TPR@1%FPR is used for semantic methods. Best performances are highlighted. Our methods show superior detection accuracy and robustness against signal processing distortions, regeneration, and cropping attacks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "lines": [ + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "spans": [ + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "type": "table", + "html": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.8630.5720.5220.5160.6770.8590.5320.5230.5210.5190.7290.8100.637
DwtDctSvd1.0000.5550.4730.6021.0001.0000.7840.6480.5960.6440.7440.8610.742
RivaGAN0.9990.8620.9860.8210.9980.9690.9340.5700.5520.6080.9910.9950.857
S.Sign.0.9950.8940.9780.8060.9110.7210.8380.7170.7150.4780.9870.9910.836
Tree-Ring0.9570.4630.9000.5480.9340.4120.8150.5090.5360.5430.5090.7340.655
Zodiac0.9980.8430.9980.9730.9980.8800.9970.9440.9580.9720.9890.9950.962
HSTR (ours)1.0000.8991.0000.9941.0000.8060.9990.9730.9820.9971.0001.0000.971
RingID1.0000.9881.0001.0001.0000.9871.0000.9921.0001.0001.0001.0000.997
HSQR (ours)1.0000.9911.0001.0001.0000.9831.0000.9921.0001.0001.0001.0000.997
SD-PromptsDwtDct0.8190.5570.5160.5060.6850.8220.5300.5130.5120.5090.7230.7940.624
DwtDctSvd1.0000.5370.4590.6100.9990.9980.8590.6590.6200.6230.7430.8600.747
RivaGAN0.9910.8230.9630.8100.9880.9610.9150.5720.5350.5670.9800.9830.841
S.Sign.0.9940.8990.9670.7690.8880.7420.8090.6770.6710.4930.9830.9900.824
Tree-Ring0.9440.4710.8940.4660.9120.4230.8020.5090.5140.5430.4690.7490.641
Zodiac0.9980.7480.9990.9790.9990.9031.0000.9400.9750.9580.9940.9960.957
HSTR (ours)1.0000.7421.0000.9901.0000.8501.0000.9830.9870.9991.0001.0000.963
RingID1.0000.9721.0001.0001.0000.9881.0000.9961.0001.0001.0001.0000.996
HSQR (ours)1.0000.9551.0000.9991.0000.9921.0000.9960.9991.0001.0001.0000.995
DiffusionDBDwtDct0.8420.5630.5150.5090.6720.8290.5260.5130.5140.5120.7230.8010.627
DwtDctSvd0.9980.5580.4630.5930.9970.9950.8300.6580.6080.6210.7420.8600.744
RivaGAN0.9870.8390.9600.7900.9850.9370.8930.5530.5180.5560.9740.9790.831
S.Sign.0.9900.8900.9670.7870.8890.7260.8190.6900.6870.4960.9810.9860.826
Tree-Ring0.9400.4870.8890.4340.9040.3920.7990.4540.5030.4540.4990.7150.622
Zodiac0.9920.7520.9880.9330.9880.8340.9840.9110.9260.9030.9710.9850.931
HSTR (ours)0.9990.7920.9960.9810.9960.7920.9910.9680.9690.9891.0001.0000.956
RingID1.0000.9891.0001.0001.0000.9631.0000.9950.9991.0001.0001.0000.995
HSQR (ours)1.0000.9771.0000.9991.0000.9740.9990.9970.9991.0001.0001.0000.995
", + "image_path": "d5f00951289a17b786af1fca0e97da387fb639b201d62c5587a010845934e893.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 421, + 295, + 552 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 421, + 266, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 266, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 266, + 432 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 421, + 266, + 432 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 421, + 266, + 432 + ], + "type": "text", + "content": " denotes the coordinates in the QR code domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 55, + 433, + 295, + 480 + ], + "type": "text", + "content": " represents the corresponding coordinates in the embedding region of the Fourier domain. Since the QR code is split into two halves, each half is mapped to either the real or imaginary component." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 481, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 505 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 481, + 295, + 505 + ], + "type": "inline_equation", + "content": "c \\in \\{\\mathrm{Re}, \\mathrm{Im}\\}" + }, + { + "bbox": [ + 55, + 481, + 295, + 505 + ], + "type": "text", + "content": " indicates whether the embedding occurs in the real or imaginary part." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "inline_equation", + "content": "F(\\tilde{x}, c)" + }, + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "text", + "content": " is the real-valued amplitude of the complex Gaussian noise in the Fourier domain, where a complex number is expressed as " + }, + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{Re}} + iF_{\\mathrm{Im}}" + }, + { + "bbox": [ + 56, + 505, + 295, + 552 + ], + "type": "text", + "content": ", making both real and imaginary amplitudes individually real-valued." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 553, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 625 + ], + "type": "text", + "content": "To maintain symmetry while avoiding numerical instability, the embedding region is positioned one pixel to the right of the vertical DC axis. Additionally, to increase the amount of embedded information and statistically reduce errors, each QR code cell is represented by multiple pixels arranged in a square pattern, e.g., " + }, + { + "bbox": [ + 55, + 553, + 295, + 625 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 553, + 295, + 625 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 626, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 295, + 673 + ], + "type": "text", + "content": "Detection. The detection process for semantic watermarking is performed by computing the L1 distance. In HSQR, the ground truth QR binary pattern is converted into a signed Boolean representation as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 684, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 684, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 97, + 684, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\mathrm {Q R} ^ {*} (\\tilde {x}, c) = \\left\\{ \\begin{array}{l l} + \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 1 \\\\ - \\Lambda , & \\text {i f} \\mathrm {Q R} (x) = 0, \\end{array} \\right. \\tag {4}", + "image_path": "3033499f83389eea8e533a1e968169057a237d50edb2a3782dbbbd04d19c7752.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 313, + 421, + 515, + 433 + ], + "type": "text", + "content": " is a fixed amplitude for binary encoding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 394, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 394, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 394, + 460 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 440, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 440, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 440, + 480 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": "Semantic Watermarks. The Tree-Ring [48] watermark is embedded into channel 3 with a radius of 14 and HSTR adopts the same watermark parameters as its baseline. For RingID [11], we follow the original settings, embedding a tree-ring pattern with a radius range of 3-14 in channel 3, along with a Gaussian noise key in channel 0. Both SFW methods follow RingID in utilizing a Gaussian noise key in channel 0. For HSQR, we use a version 1 QR code " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "(21 \\times 21" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " cells) with a cell size of 2 pixels and error correction level H, capable of encoding up to 72 bits. The QR code is embedded in channel 3, and the encoding amplitude " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " is set to 45, corresponding to the standard deviation of the real and imaginary components in the Fourier domain of " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " normal Gaussian latent vector " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "(\\approx \\sqrt{64^2 / 2})" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": ". Note that these experiments are conducted on " + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 486, + 555, + 715 + ], + "type": "text", + "content": " images. At higher resolutions, a larger embedding space allows for increased capacity through higher QR code versions that encode more bits, as well as the use of stronger watermarking parameters, enabling greater robustness and adaptability. Additionally," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18763" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 114, + 549, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Identification accuracy for different watermarking methods, evaluated across multiple attack conditions. Perfect Match Rate is used for bitstream-based methods. The best performance for each item is highlighted with shading. HSQR achieves state-of-the-art performance, while HSTR consistently outperforms other Gaussian radius-based semantic baselines, particularly under cropping attacks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "lines": [ + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "spans": [ + { + "bbox": [ + 60, + 114, + 549, + 401 + ], + "type": "table", + "html": "
DatasetsMethodsNo AttackSignal Processing AttackRegeneration AttackCropping AttackAvg
CleanBright.Cont.JPEGBlurNoiseBM3DVAE-BVAE-CDiff.C.C.R.C.
MS-COCODwtDct0.4660.0440.0000.0000.0380.4420.0000.0000.0000.0000.0000.0000.083
DwtDctSvd1.0000.0440.0190.0000.9990.9980.0370.0000.0000.0030.0000.0000.258
RivaGAN0.9740.2600.7720.0230.9610.6860.3480.0000.0000.0000.8520.9090.482
S.Sign.0.8730.1770.5630.0000.0360.0100.0070.0000.0000.0000.7090.8020.265
Tree-Ring0.3030.0870.2070.0720.2560.0300.1620.0830.0720.0540.0090.0330.114
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.0000.7140.9990.8860.9980.4600.9720.8330.8310.9711.0001.0000.889
RingID1.0000.8751.0000.9751.0000.9190.9960.9780.9700.9980.8740.9780.964
HSQR (ours)1.0000.9581.0000.9941.0000.9010.9990.9800.9870.9991.0001.0000.985
SD-PromptsDwtDct0.2850.0240.0000.0000.0170.2760.0000.0000.0000.0000.0000.0000.050
DwtDctSvd0.9930.0280.0110.0000.9820.9790.0850.0000.0000.0070.0000.0000.257
RivaGAN0.8780.2130.6130.0090.8570.6570.3040.0000.0000.0000.7360.7680.420
S.Sign.0.8130.2630.4200.0000.0210.0150.0060.0000.0000.0000.5760.7160.236
Tree-Ring0.2880.0940.1890.0510.2350.0340.1590.0790.0760.0560.0120.0410.110
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)1.000
RingID1.0000.8851.0000.9760.9980.8860.9930.9800.9730.9950.8760.9810.962
HSQR (ours)1.0000.9301.0000.9941.0000.9420.9990.9910.9971.0001.0001.0000.988
DiffusionDBDwtDct0.3570.0370.0000.0000.0340.3200.0000.0000.0000.0000.0000.0000.062
DwtDctSvd0.9900.0360.0190.0000.9750.9590.0810.0000.0000.0010.0000.0000.255
RivaGAN0.8580.2130.6250.0200.8480.6150.2210.0000.0000.0000.7560.7800.411
S.Sign.0.7980.2070.4720.0000.0270.0050.0050.0000.0000.0000.6430.7380.241
Tree-Ring0.2800.0950.1900.0590.2330.0370.1450.0810.0720.0500.0130.0390.108
Zodiac0.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.0000.000
HSTR (ours)0.9960.7210.9920.8540.9890.5630.9580.8300.8210.9520.9960.9960.889
RingID1.0000.8951.0000.9470.9960.8710.9920.9680.9580.9900.8750.9840.956
HSQR (ours)1.0000.9541.0000.9881.0000.9060.9980.9820.9910.9941.0001.0000.984
", + "image_path": "f11de02277708da5e7df5b4593c814afaf2b14a2eb23cbe4fbec9d918a1d5525.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 445 + ], + "type": "text", + "content": "the post-hoc semantic watermarking method, Zodiac [53], is included in our comparative experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 447, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 295, + 614 + ], + "type": "text", + "content": "Model Setup and Datasets. The diffusion model used for the experiments is Stable Diffusion v2-1-base [37], configured to generate images at a resolution of " + }, + { + "bbox": [ + 54, + 447, + 295, + 614 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 447, + 295, + 614 + ], + "type": "text", + "content": " with a CFG scale of 7.5. Both generation and inversion steps of the DDIM scheduler [41] are set to 50. The datasets used for image generation include 5,000 captions from the MS-COCO-2017 training set [29], 1,001 sampled prompts from the 2M subset of DiffusionDB [46], and 8,192 prompts from the Stable-Diffusion-Prompts test set [40]. For watermark detection, the number of watermarked images used for each dataset is set to 1,000. Specifically, verification is conducted on 1,000 pairs of watermarked and unwatermarked images, while identification is performed on 1,000 watermarked images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "Attack Methods. To evaluate the robustness of our method against various distortions, we apply 11 different attacks, categorized into three major types." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "- Signal Processing Attacks: These include brightness adjustment (ranging from 0 to 7), contrast adjustment (factor of 0.5), JPEG compression (quality factor of 25), Gaussian blur (radius of 5), Gaussian noise (" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sigma = 0.05" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "), and BM3D denoising (" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sigma = 0.1" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 421, + 555, + 529 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 421, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 480 + ], + "type": "text", + "content": "- Regeneration Attacks: We apply two VAE-based image compression models, Bmshj18 [4] (VAE-B) and Cheng20 [8] (VAE-C), both at quality level 3, as well as a diffusion-based regeneration attack by Zhao et al. [56] (referred to as Diff.) using 60 denoising steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "type": "text", + "content": "- Cropping Attacks: We apply center crop (C.C.) with a crop scale of 0.5 and random crop (R.C.) with a crop scale of 0.7, where the crop scale represents the ratio of the cropped image area to the original image." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. The " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " distance is computed only within the key region of the complex Fourier domain of latent noise, and unless otherwise specified, 2,048 keys are used for identification accuracy. For verification, semantic watermarking methods use True Positive Rate at " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " False Positive Rate (TPR@1%FPR) as the evaluation metric. To ensure a fair comparison with bitstream-based approaches such as DwtDct, DwtDctSvd [12], RivaGAN [52], and Stable Signature (S.Sign.) [22], we measure Bit Accuracy for these methods. For identification, where the goal is to match the exact message, Perfect Match Rate is used as the metric for bitstream-based approaches. To evaluate the generation quality, FID [24] is measured between watermarked images and 5,000 MS-COCO ground truth images. We also calculate the CLIP score [34] using the OpenCLIP-ViT/G" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18764" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 178, + 282, + 237 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 170 + ], + "type": "text", + "content": "Table 3. Generative quality evaluation of watermarking methods based on FID (MS-COCO ground truth) and CLIP score. The best performance for each item is highlighted with shading, while bold text specifically marks the low CLIP score in RingID. Our proposed methods preserve frequency integrity, achieving the best balance between watermark robustness and generative performance, whereas RingID introduces visible artifacts, compromising perceptual quality. Vrf. and Idf. denote the average detection performance in verification and identification tasks, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 178, + 282, + 237 + ], + "lines": [ + { + "bbox": [ + 67, + 178, + 282, + 237 + ], + "spans": [ + { + "bbox": [ + 67, + 178, + 282, + 237 + ], + "type": "table", + "html": "
Semantic MethodsFID ↓CLIP ↑Vrf.Idf.
Merged in GenerationTree-Ring26.4180.3260.6550.114
RingID27.0520.3240.9970.964
HSTR (ours)25.0620.3290.9710.889
HSQR (ours)24.8950.3300.9970.985
", + "image_path": "4074b3b3a96b3c74136a639019085752fdaf96fe7d52348bdf469659caa65a0b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 295, + 281 + ], + "type": "text", + "content": "model [9] to determine how closely the generated image aligns with the input prompt." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 205, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 205, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 205, + 302 + ], + "type": "text", + "content": "5.2. Comparison with Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 295, + 342 + ], + "type": "text", + "content": "In this section, we compare the detection performance in both verification and identification tasks, as well as the generation quality of different watermarking methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 628 + ], + "type": "text", + "content": "Enhanced Detection Robustness. According to Tab. 1, our proposed methods consistently achieve top-tier detection performance across various attacks. A key observation is that non-semantic watermarking methods [12, 22, 52] exhibit significant degradation under regeneration attacks, highlighting their vulnerability. Among Gaussian radius-based pattern methods (Tree-Ring, Zodiac, and HSTR), HSTR outperforms corresponding baselines in most cases, demonstrating its effectiveness in enhancing detection robustness while maintaining the fundamental tree-ring structure. In contrast, RingID leverages high-energy signed constant ring patterns, securing strong detection performance. However, as acknowledged by the authors, this method introduces noticeable ring-like artifacts in generated images, disrupting the balance between watermark robustness and generative quality. A detailed quantitative evaluation of this trade-off is presented in the following section. Meanwhile, Zodiac suffers from time-consuming processing, taking several minutes to apply the watermark—requiring 7.36 minutes per image on MS-COCO dataset—which further limits its practicality in real-world applications. In contrast, our methods following the merged-in-generation scheme introduce no additional processing time, ensuring seamless watermark embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "In Tab. 2, we compare identification results, which reveal an even greater disparity in detection performance across different watermarking methods. Gaussian radius-based pattern methods, except for HSTR, perform poorly in identification tasks, with Zodiac—designed solely for verification—failing entirely across all scenarios (achieving zero identification accuracy). Further analysis on scalability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "under different message capacities is provided in Sec. 5.3.3. Notably, our methods exhibit strong resilience against cropping attacks in both tasks, reinforcing the robustness of center-aware embedding. HSQR achieves state-of-the-art identification accuracy across all datasets, further establishing its dominance in watermark retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 288 + ], + "type": "text", + "content": "Balance with Generative Quality. Tab. 3 presents the generative quality of different semantic watermarking methods following the merged-in-generation scheme, evaluated using FID and CLIP score. RingID, which deviates from a Gaussian distribution by embedding high-energy perturbations, exhibits the worst generative performance, particularly reflected in its low CLIP score. This decline is attributed to the noticeable ring-like artifacts, as discussed earlier. In contrast, our proposed methods, which preserve frequency integrity via SFW, achieve the top two FID scores, demonstrating a better balance between watermark robustness and generative quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 553, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 587 + ], + "type": "text", + "content": "Expanding Frequency Utilization in Latent Watermarking. Traditional image-domain watermarking methods prioritize low-mid frequency embedding to resist compression and filtering attacks. This approach has also been adopted in latent diffusion-based semantic watermarking methods [11, 48, 53] without explicitly considering the differences between pixel-space and latent-space transformations. However, our findings suggest that such frequency constraints may not be necessary in the latent space, where watermark retrieval involves latent encoding and DDIM inversion, altering the impact of frequency perturbations. This distinction is evident in HSQR, which achieves state-of-the-art detection performance while utilizing nearly the entire frequency spectrum. Instead of relying on specific frequency bands, HSQR leverages structured statistical encoding, distributing watermark information across multiple pixels (" + }, + { + "bbox": [ + 313, + 289, + 553, + 587 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 289, + 553, + 587 + ], + "type": "text", + "content": " per QR cell). This redundancy enhances robustness against distortions, enabling effective retrieval even with broader frequency usage. These results indicate that latent-space watermarking is not bounded by traditional low-mid frequency constraints. Future semantic watermarking strategies can benefit from statistically robust encoding methods, allowing for effective watermark retrieval across a wider frequency range without compromising detection accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 614, + 521, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 521, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 521, + 625 + ], + "type": "text", + "content": "5.3.1. Impact of SFW on Detection Performance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "We evaluate the impact of Hermitian SFW by comparing detection performance across four cases (Tab. 4). Without SFW, using both real and imaginary components for detection (Case A) results in the lowest performance due to frequency degradation, while restricting detection to the real component (Case B) improves accuracy. In contrast, applying SFW (Cases C and D) preserves frequency integrity," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18765" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 145, + 293, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": "Table 4. Ablation study on detection performance based on frequency integrity (" + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "inline_equation", + "content": "\\sqrt{}" + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": ") and the number of detection region usage (1: real only, 2: real & imaginary). Vrf. and Idf. represent average detection performance in verification (TPR@1% FPR) and identification (accuracy), respectively. " + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "inline_equation", + "content": "\\Delta L_1^*" + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": " denotes the normalized " + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "inline_equation", + "content": "\\Delta L_1" + }, + { + "bbox": [ + 55, + 70, + 297, + 138 + ], + "type": "text", + "content": " metric, indicating detection effectiveness." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 145, + 293, + 200 + ], + "lines": [ + { + "bbox": [ + 57, + 145, + 293, + 200 + ], + "spans": [ + { + "bbox": [ + 57, + 145, + 293, + 200 + ], + "type": "table", + "html": "
CaseMethodsFreq. Int.# Det.Vrf.Idf.ΔL1*↑
ATree-Ring×20.6530.1140.232
BTree-Ring×10.8050.4160.368
CHSTR (ours)✓10.9360.7750.471
DHSTR (ours)✓20.9710.8890.476
", + "image_path": "108a8b14af9ed889422ea4a061157f4d42af96d64a4cc11791fc617421af6b56.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 213, + 176, + 329 + ], + "blocks": [ + { + "bbox": [ + 56, + 213, + 176, + 329 + ], + "lines": [ + { + "bbox": [ + 56, + 213, + 176, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 176, + 329 + ], + "type": "image", + "image_path": "56d75d67a6ba238f9798ac25f3ba8dde7f8e62510bcad250ae2eb3872e8feb0a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 296, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 384 + ], + "type": "text", + "content": "Figure 5. Identification accuracy under center crop and random crop attacks at different crop scales. HSTR and HSQR maintain higher accuracy compared to RingID, demonstrating improved robustness against cropping." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 213, + 294, + 328 + ], + "blocks": [ + { + "bbox": [ + 176, + 213, + 294, + 328 + ], + "lines": [ + { + "bbox": [ + 176, + 213, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 176, + 213, + 294, + 328 + ], + "type": "image", + "image_path": "3d4e78eb6f8b84fdb847e3363e8f009b0c427e3a6510d7addd8a6827ca628aae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 396, + 294, + 552 + ], + "blocks": [ + { + "bbox": [ + 56, + 396, + 294, + 552 + ], + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 552 + ], + "type": "image", + "image_path": "bf4440f4305343d37f581f8f612d9cbd603c4ac20fff19d24a148a5be3abc9c2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 561, + 296, + 606 + ], + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 606 + ], + "type": "text", + "content": "Figure 6. Identification accuracy across watermark message capacities. HSQR remains nearly perfect, while RingID and HSTR degrade at higher capacities. Tree-Ring and Zodiac fail to scale effectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "leading to a significant boost in detection performance. Notably, HSTR with SFW and full complex detection (Case D) achieves the highest accuracy, demonstrating that leveraging both frequency components maximizes retrieval effectiveness. These results confirm that SFW enables robust semantic watermarking by maintaining frequency integrity and fully utilizing the Fourier domain for detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 481, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 481, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 481, + 85 + ], + "type": "text", + "content": "5.3.2. Robustness to Cropping Attacks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 219 + ], + "type": "text", + "content": "We assess the robustness of center-aware embedding against center crop and random crop attacks, measuring identification accuracy across different crop scales (Fig. 5). As crop scale decreases, accuracy declines for all methods. However, RingID exhibits a steeper drop, indicating higher vulnerability to cropping. In contrast, HSTR and HSQR degrade more gradually, demonstrating improved resilience. While extreme cropping (scale 0.2) significantly affects all methods, our center-aware design consistently outperforms RingID, confirming its effectiveness in preserving watermark information under cropping distortions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 497, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 497, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 497, + 237 + ], + "type": "text", + "content": "5.3.3. Impact of Capacity on Identification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 361 + ], + "type": "text", + "content": "We evaluate identification accuracy across different watermark message capacities (64 to 8,192) for five methods (Fig. 6). Zodiac's performance collapses, while Tree-Ring exhibits rapid performance degradation, becoming nearly unusable at higher capacities. RingID, HSTR, and HSQR remain robust, maintaining " + }, + { + "bbox": [ + 313, + 239, + 555, + 361 + ], + "type": "inline_equation", + "content": "80\\%+" + }, + { + "bbox": [ + 313, + 239, + 555, + 361 + ], + "type": "text", + "content": " accuracy, though HSTR declines faster, and RingID starts degrading from 2,048 onward. HSQR demonstrates the highest scalability, retaining near-perfect accuracy even at the largest capacity, confirming its superior robustness in high-capacity scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 388, + 383 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 390, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 714 + ], + "type": "text", + "content": "We have introduced Hermitian SFW, a novel approach to semantic watermarking in the latent diffusion model framework. Unlike existing methods that fail to preserve frequency integrity, our approach ensures that watermark embeddings maintain a consistent statistical structure in the latent noise distribution. This is achieved through Hermitian symmetry enforcement, which preserves frequency components and enhances detection and generation quality. Additionally, we have proposed center-aware embedding, which significantly improves robustness against cropping attacks by strategically placing watermarks in a spatially resilient region of the latent representation. Through comprehensive experiments, we demonstrated that our method achieves state-of-the-art detection accuracy in both verification and identification tasks while also maintaining superior image generation quality, as shown by FID and CLIP scores. Our study highlights the importance of frequency integrity in Fourier-based watermarking and challenges the assumption that semantic watermarking must be confined to low-mid frequency bands. Experimental results confirm that a properly structured frequency-domain watermark can be effectively embedded and retrieved across the entire frequency spectrum without compromising generative quality. Future work includes exploring adaptive embedding strategies to further enhance robustness against adversarial attacks and extreme distortions, as well as extending our method to more diverse generative architectures beyond LDMs." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18766" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 168 + ], + "type": "text", + "content": "This research was supported by the 2024 AI Semiconductor Application/Demonstration Support Program of the Ministry of Science and ICT and the National IT Industry Promotion Agency (NIPA) with Markany Co., Ltd. as the lead organization, and in part by the BK21 FOUR program of the Education and Research Program for Future ICT Pioneers, Seoul National University in 2025." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "type": "text", + "content": "[1] Mahdi Ahmadi, Alireza Norouzi, Nader Karimi, Shadrokh Samavi, and Ali Emami. Redmark: Framework for residual diffusion watermarking based on deep networks. Expert Systems with Applications, 146:113157, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 252, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 295, + 274 + ], + "type": "text", + "content": "[2] Dhruv Arya. A survey of frequency and wavelet domain digital watermarking techniques. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 328 + ], + "type": "text", + "content": "[3] Vishal Asnani, John Collomosse, Tu Bui, Xiaoming Liu, and Shruti Agarwal. Promark: Proactive diffusion watermarking for causal attribution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10802-10811, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 371 + ], + "type": "text", + "content": "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 295, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 295, + 406 + ], + "type": "text", + "content": "[5] Mauro Barni, Franco Bartolini, Vito Cappellini, and Alessandro Piva. A dct-domain system for robust image watermarking. Signal processing, 66(3):357-372, 1998. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 295, + 449 + ], + "type": "text", + "content": "[6] Mauro Barni, Franco Bartolini, and Alessandro Piva. Improved wavelet-based watermarking through pixel-wise masking. IEEE transactions on image processing, 10(5): 783-791, 2001. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 450, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 450, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 450, + 295, + 503 + ], + "type": "text", + "content": "[7] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 505, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 295, + 559 + ], + "type": "text", + "content": "[8] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 560, + 295, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 560, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 62, + 560, + 295, + 625 + ], + "type": "text", + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 626, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 626, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 626, + 295, + 667 + ], + "type": "text", + "content": "[10] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "text", + "content": "[11] Hai Ci, Pei Yang, Yiren Song, and Mike Zheng Shou. Ringid: Rethinking tree-ring watermarking for enhanced multi-key identification. arXiv preprint arXiv:2404.14055, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[12] Ingemar Cox, Matthew Miller, Jeffrey Bloom, Jessica Fridrich, and Ton Kalker. Digital watermarking and steganography. Morgan Kaufmann, 2007. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 140 + ], + "type": "text", + "content": "[13] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE TIP, 6(12):1673-1687, 1997. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 195 + ], + "type": "text", + "content": "[14] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for data copyright protection against generative diffusion models. ACM SIGKDD Explorations Newsletter, 26(2): 60-75, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "text", + "content": "[15] DEJEY and RS Rajesh. An improved wavelet domain digital watermarking for image protection. International journal of wavelets, multiresolution and information processing, 8(01): 19-31, 2010. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "text", + "content": "[16] Denso Wave Incorporated. QR Code Essentials. Denso Wave, 2011. Available at: https://www.qrcode.com/en/about/standards.html.4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "type": "text", + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "text", + "content": "[18] Han Fang, Dongdong Chen, Qidong Huang, Jie Zhang, Zehua Ma, Weiming Zhang, and Nenghai Yu. Deep template-based watermarking. IEEE Transactions on Circuits and Systems for Video Technology, 31(4):1436-1451, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "type": "text", + "content": "[19] Han Fang, Zhaoyang Jia, Yupeng Qiu, Jiyi Zhang, Weiming Zhang, and Ee-Chien Chang. De-end: decoder-driven watermarking network. IEEE Transactions on Multimedia, 25: 7571-7581, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 400, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 554, + 454 + ], + "type": "text", + "content": "[20] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 511 + ], + "type": "text", + "content": "[21] Pierre Fernandez, Alexandre Sablayrolles, Teddy Furon, Hervé Jégou, and Matthijs Douze. Watermarking images in self-supervised latent spaces. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3054-3058. IEEE, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 567 + ], + "type": "text", + "content": "[22] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 622 + ], + "type": "text", + "content": "[23] Yiyang Guo, Ruizhe Li, Mude Hui, Hanzhong Guo, Chen Zhang, Chuangjian Cai, Le Wan, and Shangfei Wang. Freqmark: Invisible image watermarking via frequency based optimization in latent space. arXiv preprint arXiv:2410.20824, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 679, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 554, + 714 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18767" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[26] Jae-Eun Lee, Young-Ho Seo, and Dong-Wook Kim. Convolutional neural network-based digital image watermarking adaptive to the resolution of image and watermark. Applied Sciences, 10(19):6854, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[27] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 207 + ], + "type": "text", + "content": "[28] Chunlei Li, Zhaoxiang Zhang, Yunhong Wang, Bin Ma, and Di Huang. Dither modulation of significant amplitude difference for wavelet based robust watermarking. Neurocomputing, 166:404-415, 2015. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 274 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[30] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13548-13557, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "text", + "content": "[31] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "text", + "content": "[32] A Miyazaki and A Okamoto. Analysis of watermarking systems in the frequency domain and its application to design of robust watermarking systems. In Proceedings 2001 International Conference on Image Processing (Cat. No. 01CH37205), pages 506-509. IEEE, 2001. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 464 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 532 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "text", + "content": "[36] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[38] Joseph JK Ō Ruanaidh and Thierry Pun. Rotation, scale and translation invariant spread spectrum digital image watermarking. Signal processing, 66(3):303-317, 1998. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "type": "text", + "content": "[40] Gustavo Santana. Gustavosta: Stable-diffusion-prompts. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2022.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "text", + "content": "[41] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "text", + "content": "[42] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 297 + ], + "type": "text", + "content": "[43] Jun Tian. Reversible data embedding using a difference expansion. IEEE transactions on circuits and systems for video technology, 13(8):890-896, 2003. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 341 + ], + "type": "text", + "content": "[44] Ron G Van Schyndel, Andrew Z Tirkel, and Charles F Osborne. A digital watermark. In Proceedings of 1st international conference on image processing, pages 86-90. IEEE, 1994. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "text", + "content": "[45] Wenbo Wan, Jun Wang, Yunming Zhang, Jing Li, Hui Yu, and Jiande Sun. A comprehensive survey on robust image watermarking. Neurocomputing, 488:226-247, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 430 + ], + "type": "text", + "content": "[46] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusional: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "text", + "content": "[47] Xiu-mei Wen, Wei Zhao, and Fan-xing Meng. Research of a digital image watermarking algorithm resisting geometrical attacks in fourier domain. In 2009 International Conference on Computational Intelligence and Security, pages 265-268. IEEE, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 490, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 554, + 533 + ], + "type": "text", + "content": "[48] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "text", + "content": "[49] Raymond B Wolfgang, Christine I Podilchuk, and Edward J Delp. Perceptual watermarks for digital images and video. Proceedings of the IEEE, 87(7):1108-1126, 1999. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 554, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 632 + ], + "type": "text", + "content": "[50] Zijin Yang, Kai Zeng, Kejiang Chen, Han Fang, Weiming Zhang, and Nenghai Yu. Gaussian shading: Provable performance-lossless image watermarking for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12162-12171, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 668 + ], + "type": "text", + "content": "[51] Guokai Zhang, Lanjun Wang, Yuting Su, and An-An Liu. A training-free plug-and-play watermark framework for stable diffusion. arXiv preprint arXiv:2404.05607, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[52] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 1, 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18768" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 362 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[53] Lijun Zhang, Xiao Liu, Antoni Martin, Cindy Bearfield, Yuriy Brun, and Hui Guan. Attack-resilient image watermarking using stable diffusion. Advances in Neural Information Processing Systems, 37:38480-38507, 2025. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[54] Xuanyu Zhang, Runyi Li, Jiwen Yu, Youmin Xu, Weiqi Li, and Jian Zhang. Editguard: Versatile image watermarking for tamper localization and copyright protection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11964-11974, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[55] Xuandong Zhao, Kexun Zhang, Yu-Xiang Wang, and Lei Li. Generative autoencoders as watermark attackers: Analyses of vulnerabilities and threats. 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 274 + ], + "type": "text", + "content": "[56] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai. Advances in Neural Information Processing Systems, 37:8643-8672, 2025. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 317 + ], + "type": "text", + "content": "[57] Yimeng Zhao, Chengyou Wang, Xiao Zhou, and Zhiliang Qin. Dari-mark: Deep learning and attention network for robust image watermarking. Mathematics, 11(1):209, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 362 + ], + "type": "text", + "content": "[58] Jiren Zhu, Russell Kaplan, Justin Johnson, and Li Fei-Fei. Hidden: Hiding data with deep networks. In Proceedings of the European conference on computer vision (ECCV), pages 657–672, 2018. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18769" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_content_list.json b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..de3f4bf7faece63afebe82d14b0b4c53248b96d9 --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_content_list.json @@ -0,0 +1,1793 @@ +[ + { + "type": "text", + "text": "Semantics versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification", + "text_level": 1, + "bbox": [ + 91, + 130, + 903, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Tian $^{1}$ Shuo Wang $^{2}$ Rongzhao Zhang $^{1}$ Zijian Chen $^{2,1}$ Yankai Jiang $^{1}$ Chunyi Li $^{2,1}$ Xiangyang Zhu $^{1}$ Fang Yan $^{1}$ Qiang Hu $^{3}$ XiaoSong Wang $^{1}$ Guangtao Zhai $^{2,1\\boxplus}$ Shanghai AI Laboratory", + "bbox": [ + 200, + 204, + 795, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University", + "bbox": [ + 220, + 257, + 774, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University", + "bbox": [ + 281, + 275, + 712, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical imaging has significantly advanced computer-aided diagnosis, yet its re-identification (ReID) risks raise critical privacy concerns, calling for de-identification (DeID) techniques. Unfortunately, existing DeID methods neither particularly preserve medical semantics, nor are flexibly adjustable towards different privacy levels. To address these issues, we propose a divide-and-conquer framework comprising two steps: (1) Identity-Blocking, which blocks varying proportions of identity-related regions, to achieve different privacy levels; and (2) Medical-Semantics-Compensation, which leverages pretrained Medical Foundation Models (MFMs) to extract medical semantic features to compensate the blocked regions. Moreover, recognizing that features from MFMs may still contain residual identity information, we introduce a Minimum Description Length principle-based feature decoupling strategy, to effectively decouple and discard such identity components. Extensive evaluations against existing approaches across seven datasets and three downstream tasks, demonstrates our state-of-the-art performance.", + "bbox": [ + 88, + 357, + 483, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 671, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the era of digital medicine, large-scale medical images, such as X-rays and fundus photographs [6], are routinely processed by AI-based diagnostic models [28, 87, 106, 119, 121] to aid clinical decision-making. However, the increasing availability of these images raises significant concerns regarding patient privacy [17, 56, 81, 89], calling for the research on medical image de-identification.", + "bbox": [ + 89, + 696, + 482, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although explicit personal details such as patient name can be easily removed from medical image headers [1, 73, 85] or burned-in texts [102, 123], re-identification (ReID) remains feasible for the intrinsic bio-identifiers, such as anatomical markers visible in chest X-rays [35, 76, 99].", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a90263befa2c62fee1b9677538860a46d6829df0ee37ae5e085e3d9e7645b20b.jpg", + "image_caption": [ + "Figure 1. (a) Given the query medical image, the ReID model can retrieve sensitive patient information from a leaked database. (b) Our DeID framework, removing identity and then compensating medical semantics, ensures adjustable identity protection, while preserving downstream task utility. Besides, a Minimum Description Length (MDL) principle-based code space is introduced, to decouple and discard the identity information in medical features." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 901, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This enables sensitive information breaches [9, 52, 90], compromising patient privacy (see Figure 1(a)).", + "bbox": [ + 511, + 657, + 903, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several studies have attempted to defend against ReID attacks. For instance, some approaches [12, 30, 32, 44] focus on removing facial features to obfuscate identity. However, such methods cannot be applied to other body parts like the chest, where identity information is deeply interwoven with diagnostic semantics. Standard image filtering techniques, such as blurring [103], pixelation [41], and masking [108], indiscriminately degrade critical diagnostic details, thereby impairing downstream medical task performance. Moreover, under high privacy settings, the severe degradation of image quality further deteriorates task performance. Differential privacy methods [19, 27, 57, 112] mitigate identity information via noise injection, but this operation also perturbs diagnostic features. Identity adver", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author", + "bbox": [ + 112, + 886, + 250, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20613", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sarial learning methods [55, 77] train generators by jointly maximizing identity discrepancy between the generated and original images, while minimizing distortion of medical information. Nevertheless, given the inherent entanglement between identity and diagnostic features, these methods fail to preserve diagnostic semantics at high privacy levels adequately. Recently, diagnostic annotation-conditioned generative models [15, 26, 43, 91, 109] have yielded promising results, yet they remain limited to task-specific semantics and cannot offer adjustable privacy levels. In summary, no existing method preserves task-generalizable semantics, while supporting a wide range of adjustable privacy levels.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we introduce a novel divide-and-conquer framework DCM-DeID, which decouples identity removal from semantic preservation, to achieve semantic-rich yet adjustable de-identification. Our approach includes three steps, i.e., $ID$ -Blocking, which masks identity-related regions to achieve adjustable privacy levels; Medical Semantics Extraction, which leverages pre-trained medical foundation models (MFMs) [74, 117] to extract semantic-rich medical features; Image Re-Synthesis, which employs a diffusion model [42, 86] to synthesize de-identified images, given the above ID-masked image and the medical features. Moreover, considering that the features from MFMs may also contain some identity information, we introduce a novel minimum description length [33]-based feature decoupling strategy, which excludes identity-associated information from the vanilla MFM features in a minimum-codelength latent space. This effectively prevents the reintroduction of identity information during the image resynthesis step. Through the above designs, our approach achieves a better trade-off between DeID and diagnostic utilization than prior approaches. Our contributions are:", + "bbox": [ + 91, + 273, + 483, + 589 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal that existing medical DeID methods fall short in preserving task-generalizable semantics, and do not adjust seamlessly across privacy levels. We build the first benchmark for this problem, by reproducing previous approaches fairly on seven datasets.", + "- We propose the DCM-DeID framework, which performs identity removal and medical semantics preservation in separate steps, enabling both adjustable privacy protection and medical task utility.", + "- We introduce a Minimum Description Length-based decoupling strategy, which decouples identity cues from medical features in a compact code space, further improving the privacy protection capability.", + "- Our framework demonstrates state-of-the-art performance. Extensive Analysis is also performed." + ], + "bbox": [ + 109, + 589, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 829, + 241, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Privacy Protection. Early methods applied low-level filters to obscure image details, including downsampling [20], blurring [103], and pixelation [41]. Later, en", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cryption in alternate domains such as JPEG bitstreams [82, 101] and DCT coefficients [114, 115] was explored, though these often introduced severe distortions that hindered downstream tasks. Homomorphic encryption [110, 124] addresses inference on encrypted images, but suffers from high computational cost [78] and limited compatibility with advanced models like Vision Transformers [24, 93, 95]. Additionally, approaches for face images [12, 34, 69] leverage facial priors from StyleGAN [53] or face recognition networks [22, 118], which may not readily generalize to other domains, such as the medical-domain images in our work.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Medical Image De-Identification. Early methods (e.g., FreeSurfer [30], PyDeface [32], SynthStrip [44]) focus on removing facial features in brain MRI. For common medical images, early approaches use pixel-domain filters (like blurring [103] and pixelation [41]) or frequency-domain techniques [31], but these hand-crafted solutions also severely degrade the image details, leading to substantially degraded performance and visual quality [59, 60, 62]. Differential Privacy methods [19, 27, 57, 112] inject noise into the training data, which compromises inference-time utility. More recent generative models [15, 26, 43, 91, 109] synthesize images conditioned on disease labels or lesion masks. However, they tend to lack task generalizability and struggle to balance privacy-utility trade-offs, which are addressed by our approach.", + "bbox": [ + 511, + 257, + 903, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Decoupling. Variational auto-encoder (VAE)-based works [40, 88] decouple representations, by constraining the variables in latent space independent. Generative adversarial network (GAN)-based methods [14, 63] are unsupervised, leaving factors unaligned with explicit semantic or identity information. For face images, there are methods [23, 53, 54, 100] targeting identity separation. However, these methods rely on strong facial priors that may not generalize to medical images. In contrast, motivated by the explicit entropy regularization terms adopted in data compression field [2, 45-47, 61, 92, 94, 96-98], our approach effectively decouples identity in medical images, within a minimum-codelength space.", + "bbox": [ + 511, + 484, + 903, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 686, + 648, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first describe the medical re-identification (ReID) models used for privacy attacks. Next, we introduce our de-identification model, which divides the task into two stages. First, identity information is removed via region blocking with an adjustable threshold. Second, lost medical semantics are compensated. This approach flexibly adjusts privacy while preserving rich, generalizable medical features for downstream tasks.", + "bbox": [ + 511, + 704, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Medical ReID Models", + "text_level": 1, + "bbox": [ + 511, + 833, + 718, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a query medical image, ReID models aim to retrieve all images belonging to the same individual, from a medical record database. Concretely, the model first extracts identity", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20614", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac8e1bfc66603ba13b1bf15a2825c68727523f679f897b3770ce1205b4272735.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed divide-and-conquer framework, DCM-DeID. (a) $ID$ -Blocking: A pre-trained ReID network produces the identity-similarity map, which is binarized by different thresholds to adjust privacy level. (b) $ID$ -Free Medical Semantics Extraction: Medical foundation models (MFMs) extract features that are encoded into a code space under the minimum-codelength regularization. A learned mask partitions the codes into identity- and medical semantics-related ones, where only the latter one is preserved. (c) Image Re-Synthesis: A diffusion model re-synthesizes images that are privacy-preserving and semantics-rich, generalizing to various downstream tasks. We illustrate with X-ray images, but the framework is also applicable to other modalities such as fundus images. $\\diamond$ denotes frozen models, while gray dashed lines indicate components used solely for learning identity-semantic decoupling. The channel number of codes $Q$ is arbitrary; two channels are shown for conciseness. $\\otimes$ denotes the element-wise multiplication." + ], + "image_footnote": [], + "bbox": [ + 130, + 74, + 859, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(ID) embedding from the query image, and then compare it with that of each image within the database. Then, the image with the closest Euclidean distance is adopted as the re-identified image. We build two medical ReID models, i.e., ViT [24] and VisionMamba [122]-based ones, which are separately adopted in the training and the evaluation stages. These ReID models are optimized with a combination of classification loss and triplet loss [39], following the previous object ReID work [38].", + "bbox": [ + 88, + 505, + 480, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. A Divide-and-Conquer Approach", + "text_level": 1, + "bbox": [ + 89, + 651, + 382, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To defend against attacks on medical ReID models, we propose DCM-DeID, a divide-and-conquer approach for medical image de-identification. DCM-DeID operates in three stages: ID Blocking, which removes identity-related image regions; ID-Free Medical Semantics Extraction, which extracts rich medical information without reintroducing identity information; and Image Re-Synthesis, which generates the final de-identified medical image.", + "bbox": [ + 88, + 672, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ID-Blocking. Given an input image $X \\in \\mathbb{R}^{3 \\times H \\times W}$ , where $H$ and $W$ denote the image spatial scales, we use a ViT-based ReID model to extract local features $f \\in \\mathbb{R}^{768 \\times h \\times w}$ , where $h = H / 16$ and $w = W / 16$ . Spatial average pooling is applied to $f$ to obtain an identity embedding $id \\in \\mathbb{R}^{768}$ . For each spatial position in $f$ , the cosine similarity with $id$ is computed, resulting in a similarity map $S[i] =$", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\cos(f[i],id)$ , where $i$ denotes the spatial location. Then, $S$ is binarized by a threshold $T$ . Finally, the ID-blocked image is computed as: $X^{noID} = X \\odot \\mathrm{Upsample}(S > T)$ , where Upsample denotes nearest-neighbor interpolation to match the resolution of $S$ to $X$ .", + "bbox": [ + 511, + 505, + 906, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ID-Free Medical Semantics Extraction. Although $X^{noID}$ effectively removes identity information, it inevitably distorts medical cues such as lung shadows. To amend this, we employ pre-trained medical foundation models (MFMs), e.g., MGCA [105] for X-ray images, to extract rich medical feature $f_{MFM}$ from $X$ . Since $f_{MFM}$ contains both semantic cues and local details that may encode identity, we introduce a feature decoupling strategy (Section 3.3) to decouple and remove the identity information, yielding the identity-free semantic feature $\\hat{f}_{sem}$ .", + "bbox": [ + 511, + 580, + 905, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Re-Synthesis. Given $X^{noID}$ and $\\hat{f}_{sem}$ , a dual-conditioned diffusion model synthesizes the de-identified image that inherits the rich semantics within MFMs, while also protecting privacy. Since the synthesized image is highly realistic, it can be directly deployed to downstream medical AI applications, without further adaptation. The model details are elaborated in Section 3.4.", + "bbox": [ + 511, + 732, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Medical Semantics Decoupling", + "text_level": 1, + "bbox": [ + 511, + 847, + 785, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Medical features extracted by the MFM encode both diagnostic semantics (e.g., lesion morphology) and identity", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20615", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "related cues (e.g., rib patterns in chest X-rays). For effective privacy-preserving, it is imperative to decouple these two types of information, and discard the identity cues. We achieve this by learning a minimum-length code space, and separating the two parts in this space.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical Motivation. From an information-theoretic perspective, the Minimum Description Length (MDL) principle [4, 33] states that the best representation for a given set of data is the one that minimizes the total codelength needed to describe the data, where each group of features tends to capture the independent or low-correlation information parts. In our context, let $Q$ be the latent representation of the MFM feature $f_{\\mathrm{MFM}}$ and let $H(Q)$ denote its expected codelength. The MDL principle objective can be seen as balancing a reconstruction loss and a compression term, i.e., the so-called rate-distortion loss (RD loss) [5]:", + "bbox": [ + 89, + 167, + 482, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o d e - a l l}} = \\min _ {\\mathcal {E}, \\mathcal {D}} \\underbrace {\\| f _ {M F M} - \\hat {f} _ {M F M} \\| _ {2}} _ {\\text {F e a t u r e R e c o n s t r u c t i o n}} + \\underbrace {\\beta H (\\boldsymbol {Q})} _ {\\text {C o d e l e n g t h}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 337, + 482, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q = \\mathcal{E}(f_{MFM})$ , $\\hat{f}_{MFM} = \\mathcal{D}(Q)$ , and $\\beta$ denotes balancing weight. $\\mathcal{E}$ and $\\mathcal{D}$ represent a pair of feature encoder and decoder networks.", + "bbox": [ + 89, + 380, + 482, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discrete Code-based Codelength Estimation. Directly calculating the $H(\\mathbf{Q})$ for the continuous variable $\\mathbf{Q}$ is nontrivial [68]. Fortunately, the neural data compression community [2, 3, 70, 72] have verified that the codelength of integer latent variables can be quite precisely estimated with a learnable entropy model. Therefore, we append the quantization operation at the tail of the encoder $\\mathcal{E}$ , to make elements within $\\mathbf{Q}$ discrete values, and estimate its codelength.", + "bbox": [ + 89, + 426, + 482, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, $\\mathcal{E}$ comprises three residual blocks [36] with 256 channels, followed by a convolutional layer to reduce dimensionality and a rounding operation that outputs a 32-channel integer code $Q$ . The decoder network $\\mathcal{D}$ is symmetric to $\\mathcal{E}$ , except it omits the rounding operation. During training, the straight-through estimator [71] is employed to backpropagate gradients through the rounding step.", + "bbox": [ + 89, + 547, + 482, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [2], the expected codelength of encoding $Q$ is calculated as the log-likelihood, i.e., $H(Q) = -\\log_2p(Q)$ , where the probability $p(Q)$ is modeled using a Gaussian Mixture Model (GMM) [84] with $K$ components:", + "bbox": [ + 89, + 654, + 482, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {Q}) = \\sum_ {k = 1} ^ {K} w ^ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {Q}; \\mu^ {k}, e ^ {\\sigma^ {k}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 715, + 482, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\mathbf{w},\\boldsymbol {\\mu},\\boldsymbol {\\sigma}\\}$ are the learnable mixture weights, means, and log variance scalers of the GMM components, respectively, which are shared across spatial positions, not unshared along the channel axis [2]. Following [16], $K$ is set to three. For each integer element $q\\in$ $Q$ , the probability is computed over the quantization bin [18, 72], $p(q) = \\mathcal{F}(q + 0.5) - \\mathcal{F}(q - 0.5)$ , where $\\mathcal{F}(x) = \\sum_{k = 1}^{K}w^{k}\\Phi \\Big(x;\\mu^{k},e^{\\sigma^{k}}\\Big)$ is the cumulative distribution function (CDF) of the Gaussian Mixture Model", + "bbox": [ + 89, + 758, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(GMM), $\\Phi \\left(x;\\mu ,e^{\\sigma}\\right) = \\frac{1}{2}\\left[1 + \\mathrm{erf}\\left(\\frac{x - \\mu}{\\sqrt{2}e^{\\sigma}}\\right)\\right]$ . We not that the CDF can be efficiently calculated by the modern deep learning framework such as PyTorch [80].", + "bbox": [ + 511, + 87, + 903, + 141 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning of Identity-Associated Code Mask. A single convolution layer predicts a binary mask $\\mathbf{M}$ from $Q$ , with the same dimensions as $Q$ . The Gumbel-Softmax algorithm [50] is applied to enable gradient propagation through the binary mask. The identity-associated codes are then obtained by element-wise masking, $Q_{\\mathrm{id}} = Q\\odot \\mathbf{M}$ . A lightweight convolutional network, composed of three residual blocks followed by average pooling, predicts the identity embedding $\\hat{\\boldsymbol{id}}$ from $Q_{\\mathrm{id}}$ . Then, the RD loss for reconstructing identity can be given by:", + "bbox": [ + 511, + 142, + 903, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o d e - i d}} = \\| \\hat {i d} - i d \\| _ {2} + \\beta H \\left(\\boldsymbol {Q} _ {\\text {i d}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 301, + 903, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H(\\tilde{Q}_{\\mathrm{id}})$ is calculated similarly to $H(Q)$ , sharing the same GMM parameters and balancing weight $\\beta$ as in Equation 1, since they operate in the same latent space.", + "bbox": [ + 511, + 330, + 903, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction of Medical Semantics. By suppressing identity-related codes via the inverse mask $(1 - \\mathbf{M})$ , we obtain the semantics-part codes $Q_{sem} = (1 - \\mathbf{M})\\otimes Q$ . Finally, the final ID-free medical semantic feature is reconstructed as: $\\hat{f}_{sem} = \\mathcal{D}(Q_{sem})$ , which preserves critical diagnostic semantics, excluding the identity information.", + "bbox": [ + 511, + 378, + 905, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Image Re-Synthesis Model", + "text_level": 1, + "bbox": [ + 513, + 481, + 754, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the ID-masked image $X^{noID}$ and the ID-free medical semantic feature $\\hat{f}_{sem}$ , we employ a diffusion model to synthesize de-identified medical images. First, we utilize a Feat-Net to project the high-resolution $X^{noID}$ into the low-resolution feature $f^{noID} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}}$ . The Feat-Net consists of the VAE encoder from Stable Diffusion [86], followed by two convolution layers of kernel size 5 and stride size 2. Next, we adopt a bi-directional crossattention mechanism [13] to fuse $f^{noID}$ and $\\hat{f}_{sem}$ , producing a fused feature $f_{fuse} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}}$ , which is further processed through a series of convolutional layers. This produces a set of features with dimensions matching those of the UNet's intermediate feature maps within the diffusion model. These features are added to the UNet layers, guiding the diffusion process toward two objectives: maintaining the privacy level of $X^{noID}$ , while preserving the medical semantics in $\\hat{f}_{sem}$ .", + "bbox": [ + 511, + 503, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Learning Strategy", + "text_level": 1, + "bbox": [ + 513, + 775, + 689, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The whole framework is end-to-end optimized, with the following objective,", + "bbox": [ + 511, + 797, + 903, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o d e - a l l}} + \\mathcal {L} _ {\\text {c o d e - i d}} + \\mathcal {L} _ {\\text {d i f f u s e}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 837, + 903, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{diffuse}}$ denotes the diffusion loss [42]. We do not introduce the balancing weight, since we found directly adding the loss terms already achieves satisfactory results.", + "bbox": [ + 511, + 859, + 905, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c99259f96619449aa904ee38d86f8de71df13aaea8e70f3bc5f3cc3cdc2adebf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 74, + 297, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/808fc601ee7308eefabd34039ba22fd34d7aab8b1065800b3e7b11889bd437ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 74, + 500, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d21caf0f38b5ab4c364b97222113ffbe277f29a2365623c9bbd21927028cd87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 74, + 702, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d665c432f1e60176e8a8d202611a287d670558f16f9793fe46c229356addef9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 71, + 906, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e557411b02a3bb757f64fcc717e16de5c1cc89d18c05a51fa944b5acff47adc.jpg", + "image_caption": [ + "(a)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 99, + 234, + 297, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2faff10e5a4edea8300716b970cf5868f398bbfba667288db02c7e9e2240961b.jpg", + "image_caption": [ + "(b)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 303, + 234, + 500, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/10ac4a2770487f28e742ddaedb9dea3b441d98ed1049d509b8218643496ae457.jpg", + "image_caption": [ + "(c)", + "(g)" + ], + "image_footnote": [], + "bbox": [ + 506, + 234, + 702, + 383 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1f20bd9d447e113e2051a015d7447551d6395951025e02778cdc5afd2b6fd45.jpg", + "image_caption": [ + "(d)", + "(h)" + ], + "image_footnote": [], + "bbox": [ + 710, + 234, + 906, + 383 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5d8b344a8601089caa65eafa69109a29a5731b4c224a3ba1bb6fa4c569a36ed6.jpg", + "table_caption": [ + "Figure 3. Identity-performance trade-off curves of various medical privacy protection methods." + ], + "table_footnote": [ + "Table 1. Performance comparison of medical image privacy protection methods, under different attack Success-Rates (SR). For measuring SR, we adopt CMC-R1 metric for MIMIC-X, Chest-Xray, CheXpert, EyePACS, and ODIR5K, using ID-R metric for ChestX-Det and Refuge2. Original denotes the performance on original images, which is the performance upper-bound of privacy-removal images." + ], + "table_body": "
Attack SRMethodX-ray Classify AUROC (%)X-ray Caption BLEUX-ray Seg Dice (%)Fundus Classify AUROC (%)Fundus Seg Dice (%)
MIMIC-XChest-XRayCheXpertMIMIC-XChestX-DetEyePACSODIR5KRefuge2
10%Pixel-Blur [103]64.1560.4874.450.12597.5344.8356.1722.35
Feat-Noise [116]65.8466.2566.490.145310.7447.7655.5324.85
ID-Adv [77]59.3661.7573.110.113122.2356.2858.5041.23
Privacy-Net [55]57.0959.5171.470.132915.7552.6357.2939.23
MAE [37]59.1165.2677.050.11869.9450.9156.5027.33
Ours72.8673.6686.120.175027.9865.2359.9654.66
20%Pixel-Blur [103]67.2364.2780.670.130711.2549.5657.4234.11
Feat-Noise [116]68.2769.0180.780.159514.2851.2957.1737.20
ID-Adv [77]64.0862.7177.530.119322.8363.2459.2849.47
Privacy-Net [55]62.1162.6380.250.143420.5561.0158.2547.47
MAE [37]66.4371.3883.520.138313.9954.9758.0140.13
Ours74.3575.0186.320.185929.4969.5960.4162.04
Original82.1384.8287.240.321852.8981.4661.5390.08
", + "bbox": [ + 93, + 412, + 901, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 686, + 223, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 712, + 307, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Med-ReID models, we adopt the AdamW optimizer [66] during training, with a learning rate of 1e-5 scheduled by cosine decaying strategy and a weight decay of 1e-2. The training process consists of 300,000 steps. The batch size is 256. We apply random cropping and blurring as image augmentation strategies, and the input image resolution to networks is $256 \\times 256$ . The ViT-based models are initialized with CLIP-pretrained weights [83], while the VisionMamba-based models are initialized with ImageNet-pretrained weights [21]. Training a single ReID model takes about 24 hours with four NVIDIA RTX 4090 GPUs.", + "bbox": [ + 89, + 734, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For DCM-DeID model, the UNet within the diffusion model follows the same architecture as the Stable Diffusion [86], also performing the diffusion procedure in the latent space. The feature channels within UNet are reduced to [128, 256, 512, 1024], for the four stages of both the down-pathway and up-pathway, to reduce computational cost. The identity-similarity map threshold $T$ is defined as the $r$ -th quantile of the similarity map $S$ . $r$ is selected from [0.95, 0.7, 0.4, 0.2] to cover wide privacy levels. We adopt MGCA-ResNet [105] and RetFound-ViT [121] MFMs for X-ray and fundus images, respectively. During training, we apply random flipping and random cropping $256 \\times 256$ patches for data augmentation. The codelength loss term weight $\\beta$ is set to 0.5. At test time, we resize the shorter", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20617", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "side of the images to 256 and then center-crop the middle $256 \\times 256$ region. The learning rate is set to 1e-4 and is gradually decayed with the cosine annealing strategy [67]. The total number of training steps is 800,000. The mini-batch size is 64. We utilize the AdamW optimizer [66] implemented in PyTorch [80] with CUDA support. The momentum parameters are set as $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.99$ , and the gradient norm is clipped to a maximum value of 1. The entire training process takes about three days on a machine equipped with eight NVIDIA RTX 4090 GPUs.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 89, + 253, + 192, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our approach on two medical image modalities: chest X-rays and eye fundus photographs, with seven public datasets. For the chest X-ray modality, we split the MIMIC-X dataset [51] into training, validation, and test sets using an 8:1:1 ratio. For the Chest-Xray and CheXpert datasets, we randomly select $10\\%$ patients as the test set. We also adopt the ChestX-Det dataset [64] to evaluate the X-ray segmentation task. For the eye fundus modality, we divide the EyePACS dataset [25] into training, validation, and test sets with an 8:1:1 ratio, and we use the Refuge2 dataset [29] to evaluate the fundus segmentation task. ODIR5K [8] is also adopted for evaluating the fundus classification task of systemic diseases such as hypertension. Note that only MIMIC-X and EyePACS are used during training; all other datasets, which differ in environment, demographics, and imaging devices, are never seen during training, to evaluate the domain generalizability of our approach. MIMIC-X, CheXpert, Chest-Xray, EyePACS, ChestXDet, and ODIR5K contain 377K, 224K, 112K, 88K, 3.6K, and 5K images.", + "bbox": [ + 89, + 275, + 482, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Reproduced Privacy Protection Methods", + "text_level": 1, + "bbox": [ + 89, + 589, + 439, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement several privacy protection methods, comparing them with our approach in a fair setting.", + "bbox": [ + 89, + 611, + 482, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pixel-Blur [103]. This method applies a Gaussian blur to the input image. We experiment with standard deviations of $\\{1,5,10,20\\}$ to vary the level of de-identification.", + "bbox": [ + 89, + 642, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feat-Noise [116]. We train an autoencoder [116] and inject Gaussian noise into its latent features. The noise level is selected from $\\{0.1, 0.8, 0.85, 0.9, 0.95\\}$ .", + "bbox": [ + 89, + 686, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ID-Adv [77]. A UNet is trained to generate a de-identified image $Y$ from the original image $X$ , optimizing the loss $\\mathcal{L} = \\lambda \\cos(id_X, id_Y) + \\|med_X - med_Y\\|_2 + \\mathcal{L}_{reg}$ , where $id_X$ and $id_Y$ are identity features extracted by a ViT-based ReID model, and $med_X$ and $med_Y$ are medical features obtained from MFMs same as our approach. $\\mathcal{L}_{reg}$ is a GAN regularization loss ensuring visual plausibility, $\\cos(\\cdot, \\cdot)$ denotes cosine similarity, and $\\| \\cdot \\|_2$ the $\\ell_2$ norm. The trade-off weight $\\lambda$ is chosen from $\\{0.1, 0.5, 1, 2\\}$ .", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Privacy-Net [55]. This method updates the identity model and the de-identification network adversarially, enhanc", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing de-identification performance. The original PrivacyNet focuses solely on segmentation tasks, supervised by segmentation masks. To enable task-agnostic de-identification, we train it using the same objective as ID-Adv. Since the identity model is adversarially updated and are stronger, we use smaller $\\lambda$ values compared to ID-Adv, i.e., $\\{0.05, 0.25, 0.5, 1\\}$ .", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MAE [37]. Following [108], we transfer the concept of masked auto-encoders (MAE) [37] to the adjustable privacy protection problem, by masking a random proportion of patches to obscure identity information. It adopts the same diffusion model as our approach to generate the masked regions. This model can also serve as a degenerated version of our model, where both semantic compensation and identity-region similarity designs are removed.", + "bbox": [ + 511, + 196, + 906, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Downstream Task Models", + "text_level": 1, + "bbox": [ + 511, + 329, + 750, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the identity recognition, we adopt the VisionMamba-based ReID model, which differs from the ViT-based model employed during the training of privacy protection methods, ensuring the method generalization capability across different ReID models. For the $X$ -ray classification, we use the ViT model pre-trained with Med-UniC [104]. For $X$ -ray captioning, we employ the visual-language model CXRLLaVA-v2 [58], which is specifically designed for X-ray images. For $X$ -ray segmentation, we adopt CGRSeg [75]. For fundus classification, we use the ViT model pre-trained with KeepFit [111]. Finally, for fundus segmentation, given the limited dataset size, we employ nnUNet [49].", + "bbox": [ + 511, + 351, + 906, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 542, + 696, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For privacy evaluation, we adopt the cumulative matching characteristics (CMC) [10] at Rank-1, i.e., CMC-R1, on datasets with patient ID information available (i.e., MIMIC-X, Chest-Xray, CheXpert, EyePacs, and ODIR5K). For datasets without patient ID information (i.e., CheX-det and REFuge2), we adopt the recognition rate, i.e., ID-R, which determines whether the distance between the ID feature of the original and de-identified image exceeds a predefined threshold. The thresholds are set to 1.1 and 1.35 for the X-ray and fundus modalities, respectively, based on statistics from the validation sets of MIMIC-X and EyePACS. For the disease diagnosis task, we employ the area under the receiver operating characteristic curve (AUROC) metric [11]; for the image captioning task, we use the bilingual evaluation understudy (BLEU) metric [79]; and for the image segmentation task, we adopt the Dice score metric [7].", + "bbox": [ + 511, + 565, + 908, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Results", + "text_level": 1, + "bbox": [ + 511, + 816, + 607, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "X-ray Classification. As shown in Table 1, our method substantially outperforms other approaches, achieving AU-ROCs of $72.86\\%$ , $73.66\\%$ , and $86.12\\%$ on MIMIC-X, Chest-XRay, and Chexpert, respectively, under CMC", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20618", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathrm{R}1 = 10\\%$ Notably, although our model is trained on MIMIC-X, it generalizes well to the other two datasets.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among the other compared approaches, Feat-Noise obtains the second-best performance, i.e., AUROC of $65.84\\%$ at CMC-R1=10% on MIMIX-X, by condensing image pixels into a compact latent feature space. In contrast, methods that jointly optimize a trade-off between de-identification and medical preservation, i.e., ID-Adv and PrivacyNet, yield unsatisfactory performances. As shown in Figure 3 (a), under the ID-R=5% setting, ID-Adv and PrivacyNet attain AUROCs of only $56.32\\%$ and $54.21\\%$ on MIMIC-X, respectively, which are much lower than the simple pixel blurring baseline $(61.62\\%)$ . This indicates that directly optimizing the two conflicting objectives is suboptimal. In contrast, our approach decouples the objectives into two separate steps, identity removal and medical semantic compensation, achieving consistently superior performance.", + "bbox": [ + 88, + 123, + 482, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As for MAE, which employs the same diffusion model as ours, it achieves competitive results at a high attacking rate, with an AUROC of $76.12\\%$ @CMC-R1=40% on MIMIC-X, outperforming all other approaches except ours. However, at a low attacking rate CMC-R1=10%, it falls behind our method by over $13\\%$ AUROC. This highlights that our superior performance is not solely due to the generative power of the diffusion model, but rather stems from the effectiveness of our core idea of semantic compensation.", + "bbox": [ + 89, + 351, + 483, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "X-ray Caption. As shown in Table 1, our method attains a BLEU score of 0.1750, remarkably surpassing Pixel-Blur (0.1259), Feat-Noise (0.1453), and MAE (0.1186), at CMC-R1=10%. This proves that our approach can comprehensively preserve the clinic-required information, beyond only the classification label.", + "bbox": [ + 89, + 489, + 483, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "X-ray Segmentation. Furthermore, we evaluate the methods on a fine-grained task: segmentation. As shown in Table 1, at ID-R1=10%, our method achieves a Dice score of 27.98%, outperforming Pixel-Blur (7.53%), Feat-Noise (10.74%), Privacy-Net (15.75%), ID-Adv (22.23%), and MAE(9.94%). This proves that our semantic compensation scheme not only preserves the global semantics for classification, but also effectively retains the local semantics for segmentation. Pixel-Blur and Feat-Noise perform poorly, since they severely corrupts the image details. In contrast, ID-Adv and Privacy-Net, which incorporate a medical feature-matching loss, achieve slightly decent performance, but still lag far behind our approach. For instance, under ID-R=80%, our method outperforms Privacy-Net by approximately 8%, as shown in Figure 3(e).", + "bbox": [ + 89, + 582, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fundus Classification. Beyond X-ray images, our method also proves effective on fundus data. For instance, on EyePACS and ORID5K, our approach outperforms the second-best competitor ID-Adv by about $9\\%$ and $1\\%$ , respectively. These results confirm that our approach generalizes well across different imaging modalities.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84ebb715e6aff1f92f6bcac6c42d18e0684e09850eea652fb7261aecfc4e5704.jpg", + "image_caption": [ + "Figure 4. (Left) Ablation on the framework design. (Right) Ablation study on the feature decoupling strategy." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 705, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c067b500d0ecda17b4ae16ba8843054cac61225c4f617c31305272b1439c21d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 897, + 232 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/484413447dbe3f3b1e24e657e8d0c093c97a2de92a3ab18c7a0221e23eca47ab.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 614, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc0f1f5271607d965c29560134977575fbda663fdd274c3aec9228ad50406ce8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 614, + 268, + 710, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d44fe9c5aaac6d1b055fc4c601b4c980c5eb05dd55be3825a532e23efd55cb31.jpg", + "image_caption": [ + "w/oSem", + "Figure 5. Qualitative comparison of different variant models. The models are described in Figure 4 caption. Red arrow denotes the modified identity-related features. Best to view by zooming-in." + ], + "image_footnote": [], + "bbox": [ + 710, + 268, + 800, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/240574cd23d8e87fbad067532886bd56f2b2361796c0e7b8a2dd01987d35b347.jpg", + "image_caption": [ + "w/oDecouple" + ], + "image_footnote": [], + "bbox": [ + 800, + 268, + 900, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fundus Segmentation. Our method achieves a Dice score of $54.66\\%$ on REFUGE2 at ID-R1=10%, largely surpassing MAE (27.33%), Privacy-Net (39.23%), and ID-Adv (41.23%). This further validates that our method also effectively preserves fine-grained semantic cues of eye fundus.", + "bbox": [ + 511, + 411, + 903, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Model Analysis", + "text_level": 1, + "bbox": [ + 511, + 498, + 668, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Framework-Level Ablation Study. As shown in Figure 4 (Left), by removing the semantic branch, the AUROC of the resulted model 'w/oSem' dramatically drops by over $8\\%$ , at CMC-R1=5%. On the other hand, without the identity-semantics decoupling mechanism, the resulting model 'w/oDecouple' leads to about $15\\%$ CMC-R1 increase, for achieving the similar AUROC performance, since substantial identity cues are leaked from the vanilla medical features of MFMs. We further illustrate the protected images from different models. As shown in Figure 5, our results effectively modify identity-related features, such as the shape and location of the clavicle and chest contour. The 'w/oSem' model also removes these regions but significantly alters medical manifestations. In contrast, the 'w/oDecouple' model preserves medical features but fails to sufficiently suppress identity-related features, such as clavicle shape, due to residual identity information in the features from MFMs. These results confirm that both medical semantics and identity-semantics decoupling are essential for our advanced medical DeID approach.", + "bbox": [ + 511, + 522, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides, we attempt to remove ID-Blocking module, tuning privacy via activating more decoupled ID codes. This leads to degraded AUROC (72.01%, 73.47%, 74.68%), compared to using ID-Blocking (72.68%, 74.35%, 79.62%) on MIMIC-X (10%, 20%, 40% at tack SR settings), due to", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20619", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df8bbd1cbc0022cd31f7ab2795244b47bfe48cc3b7d54d867ef2c40f606bbc97.jpg", + "image_caption": [ + "Figure 6. Comparison of semantic and identity information in terms of Bits-per-Pixel (bpp) [65], calculated as the feature code-length divided by the original image size." + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 480, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the complete loss of pixel-wise details.", + "bbox": [ + 89, + 233, + 349, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on the Decoupling Strategy. As shown in Figure 4 (Right), omitting the codelength loss terms ('w/oCodeLengthLoss') fails to effectively remove identity information from MFM features, since the loose space cannot effectively decouple the identity and the semantics information. Moreover, removing the discrete code bottleneck ('w/oCodeSpace') further exacerbates identity leakage, leading to further inferior performance.", + "bbox": [ + 89, + 248, + 482, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we quantitatively compare the overall and identity-related information in MFM features, as shown in Figure 6. First, we notice that a significant portion is identity-related, i.e., around $44\\%$ and $55\\%$ for X-ray and fundus images. Second, the average information amount of the X-ray dataset Chest-Xray is 0.23bpp, much higher than 0.11bpp achieved by the fundus dataset EyePACS. This is aligned with the medical knowledge prior, that X-rays capture multiple organs and tissues, containing much complex information, than the fundus image that only focuses on eyes. This proves that the learned codelength effectively describes the medical data characteristics.", + "bbox": [ + 88, + 369, + 482, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we analyze the impact of the codelength loss weight $\\beta$ and the latent code channel number. As shown in Figure 7 (Left), reducing $\\beta$ from 0.5 to 0.1 significantly increases CMC-R1 from $5.85\\%$ to $12.34\\%$ , as a loosely constrained code space fails to effectively decouple identity information. Conversely, increasing $\\beta$ from 0.5 to 2 has little effect on CMC-R1 but reduces AUROC performance by approximately $6\\%$ , as an overly strong constraint impairs semantic feature reconstruction. The number of code channels also influences performance, by tuning the information capacity of the latent code, as shown in Figure 7 (Right). However, since $\\beta$ directly regulates the code-length term, the impact of the channel number is limited.", + "bbox": [ + 89, + 550, + 482, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion with Label-Conditioned Diffusion Models. These methods [48, 107, 120] employ task-specific labels (e.g., disease labels or text reports) to synthesize images, which are limited to label-associated tasks. In contrast, our approach is task-agnostic and applicable to diverse tasks. Moreover, after fine-tuning our approach towards a single task, i.e., replacing the MFM with a supervised classification network, our method achieves $81.92\\%$ AUROC at CMC-R1 = 0.30%, surpassing the label-conditioned model,", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98358b9df95a800e9e382297e8f44ecd2788cd1858fac5dbc638c897a2b07695.jpg", + "image_caption": [ + "Figure 7. (Left) Impact of the rate-distortion weight $\\beta$ . (Right:) Impact of the code dimension. All experiments are evaluated by masking $95\\%$ identity-related regions, for a fair comparison." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 707, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96fc7bb0e8cda64f277c0f295fa73ce25d3e77aa0311592b67d5d53a02547bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 893, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "i.e., $80.79\\%$ AUROC at CMC-R1 = 0.29%. This confirms that our minimum-codelength representation also benefits the single-task setting, compared to the methods directly using the task labels guiding the diffusion procedure.", + "bbox": [ + 511, + 276, + 903, + 337 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Complexity. All methods and our model comprise about 380M parameters, for a fair comparison. Our inference time is 540 ms on an NVIDIA 4090 GPU, which is similar to MAE (526ms), but slower than Privacy-Net (120 ms), ID-Adv (122 ms), and Feat-Noise (124 ms), due to the multiple inference steps of diffusion procedure. Nonetheless, given the significant performance gains and that the medical imaging procedure itself is time-consuming, the running time is acceptable and does not hinder clinical workflows. In the future, we will integrate the single-step diffusion technique [113] to accelerate the process.", + "bbox": [ + 511, + 340, + 906, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Future Works, and Other", + "text_level": 1, + "bbox": [ + 511, + 534, + 854, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We have presented DCM-DeID, a divide-and-conquer framework for medical image de-identification. By leveraging pre-trained Medical Foundation Models and a minimum codelength-based feature decoupling strategy, our method effectively remove identity cues, while preserving medical task utility. Extensive evaluations demonstrate the superiority of our approach. Future Works. We will extend our approach to multi-slice images, such as those produced by Magnetic Resonance Imaging (MRI). Broader Impacts. Our DeID technique is designed for medical AI applications, aiding the human. We emphasize that all rigorous clinical decisions must be made by human physicians using the original medical images. Furthermore, it is critical to enforce ethical guidelines, working in synergy with technological approaches to achieve medical privacy protection.", + "bbox": [ + 511, + 564, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No.62225112), the Fundamental Research Funds for the Central Universities, National Key R&D Program of China 2021YFE0206700, Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and STCSM 22DZ2229005.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20620", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kadek YE, Aryanto, Andre Broekema, Matthijs Oudkerk, and Peter MA van Ooijen. Implementation of an anonymisation tool for clinical trials using a clinical trial processor integrated with an existing trial patient data information system. European radiology, 22:144-151, 2012. 1", + "[2] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 4", + "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4", + "[4] Andrew Barron, Jorma Rissanen, and Bin Yu. The minimum description length principle in coding and modeling. IEEE transactions on information theory, 44(6):2743-2760, 1998. 4", + "[5] Toby Berger. Rate-distortion theory. Wiley Encyclopedia of Telecommunications, 2003. 4", + "[6] Rui Bernardes, Pedro Serranho, and Conceicao Lobo. Digital ocular fundus imaging: a review. Ophthalmologica, 226 (4):161-181, 2011. 1", + "[7] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 92-100. Springer, 2019. 6", + "[8] Amit Bhati, Neha Gour, Pritee Khanna, and Aparajita Ojha. Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image dataset. Computers in Biology and Medicine, 153:106519, 2023. 6", + "[9] Vincent Bindschaedler, Paul Grubbs, David Cash, Thomas Ristenpart, and Vitaly Shmatikov. The tao of inference in privacy-protected databases. Cryptology ePrint Archive, 2017. 1", + "[10] Ruud M Bolle, Jonathan H Connell, Sharath Pankanti, Nalini K Ratha, and Andrew W Senior. The relation between the roc curve and the cmc. In Fourth IEEE workshop on automatic identification advanced technologies (AutoID'05), pages 15-20. IEEE, 2005. 6", + "[11] Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145-1159, 1997. 6", + "[12] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3334–3342, 2021. 1, 2", + "[13] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4" + ], + "bbox": [ + 99, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2", + "[15] Yan Chen and Pouyan Esmaeilzadeh. Generative ai in medical practice: in-depth exploration of privacy and security challenges. Journal of Medical Internet Research, 26:e53008, 2024. 2", + "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 4", + "[17] I Glenn Cohen and Michelle M Mello. Big data, big tech, and protecting patient privacy. *Jama*, 322(12):1141-1142, 2019. 1", + "[18] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 4", + "[19] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Obfuscation of images via differential privacy: From facial images to general images. Peer-to-Peer Networking and Applications, 14:1705-1733, 2021. 1, 2", + "[20] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2", + "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2", + "[23] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 2", + "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3", + "[25] Emma Dugas, Jared, Jorge, and Will Cukierski. Diabetic retinopathy detection. https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. Kaggle. 6", + "[26] August DuMont Schütte, Jürgen Hetzel, Sergios Gatidis, Tobias Hepp, Benedikt Dietz, Stefan Bauer, and Patrick Schwab. Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation. NPJ digital medicine, 4(1):141, 2021. 2" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20621", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2", + "[28] Weijie Fan, Yi Yang, Jing Qi, Qichuan Zhang, Cuiwei Liao, Li Wen, Shuang Wang, Guangxian Wang, Yu Xia, Qihua Wu, et al. A deep-learning-based framework for identifying and localizing multiple abnormalities and assessing cardiomegaly in chest x-ray. Nature Communications, 15(1): 1347, 2024. 1", + "[29] Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, et al. Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. arXiv preprint arXiv:2202.08994, 2022. 6", + "[30] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774-781, 2012. 1, 2", + "[31] Alex Gaudio, Asim Smailagic, Christos Faloutsos, Shreshta Mohan, Elvin Johnson, Yuhao Liu, Pedro Costa, and Aurélio Campilho. Deepfixcx: Explainable privacy-preserving image compression for medical image analysis. Wiley interdisciplinary reviews: Data mining and knowledge discovery, 13(4):e1495, 2023. 2", + "[32] Satrajit Ghosh, Chris Gorgolewski, et al. pydeface: A tool to remove facial features from mri images, 2012. 1, 2", + "[33] Peter D Grünwald. The minimum description length principle. MIT press, 2007. 2, 4", + "[34] Xiuye Gu, Weixin Luo, Michael S Ryoo, and Yong Jae Lee. Password-conditioned anonymization and deanonymization with face identity transformers. In European conference on computer vision, pages 727-743. Springer, 2020. 2", + "[35] A Gueziec and P Kazanzides. Anatomy-based registration of ct-scan and intraoperative x-ray images for guiding a surgical robot. IEEE Transactions on Medical Imaging, 17(5): 715-728, 1998. 1", + "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5, 6", + "[38] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 3", + "[39] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 3", + "[40] Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017. 2" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Steven Hill, Zhimin Zhou, Lawrence Saul, and Hovav Shacham. On the (in) effectiveness of mosaicing and blurring as tools for document redaction. Proceedings on Privacy Enhancing Technologies, 2016. 1, 2", + "[42] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4", + "[43] Sungmin Hong, Razvan Marinescu, Adrian V Dalca, Anna K Bonkhoff, Martin Bretzner, Natalia S Rost, and Polina Golland. 3d-stylegan: A style-based generative adversarial network for generative modeling of three-dimensional medical images. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 24-34. Springer, 2021. 2", + "[44] Andrew Hoopes, Jocelyn S Mora, Adrian V Dalca, Bruce Fischl, and Malte Hoffmann. Synthstrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. 1, 2", + "[45] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. arXiv preprint arXiv:2501.13630, 2025. 2", + "[46] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 875-885, 2025.", + "[47] Qiang Hu, Houqiang Zhong, Zihan Zheng, Xiaoyun Zhang, Zhengxue Cheng, Li Song, Guangtao Zhai, and Yanfeng Wang. Vrvvc: Variable-rate nerf-based volumetric video compression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3563-3571, 2025. 2", + "[48] Peng Huang, Xue Gao, Lihong Huang, Jing Jiao, Xiaokang Li, Yuanyuan Wang, and Yi Guo. Chest-diffusion: a lightweight text-to-image model for report-to-cxr generation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 8", + "[49] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6", + "[50] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4", + "[51] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 6", + "[52] Georgios A Kaissis, Marcus R Makowski, Daniel Rückert, and Rickmer F Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6):305-311, 2020. 1" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20622", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[54] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[55] Bach Ngoc Kim, Jose Dolz, Pierre-Marc Jodoin, and Christian Desrosiers. Privacy-net: an adversarial approach for identity-obfuscated segmentation of medical images. IEEE Transactions on Medical Imaging, 40(7):1737-1749, 2021. 2, 5, 6", + "[56] Cody A Koch and Wayne F Larrabee. Patient privacy, photographs, and publication. JAMA facial plastic surgery, 15 (5):335-336, 2013. 1", + "[57] Abhinav Kumar, Sanjay Kumar Singh, K Lakshmanan, Sonal Saxena, and Sameer Shrivastava. A novel cloud-assisted secure deep feature classification framework for cancer histopathology images. ACM Transactions on Internet Technology (TOIT), 21(2):1-22, 2021. 1, 2", + "[58] Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim, and Soon Ho Yoon. Cxr-llava: a multimodal large language model for interpreting chest x-ray images. European Radiology, pages 1-13, 2025. 6", + "[59] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8):6833-6846, 2023. 2", + "[60] Chunyi Li, Xiaozhe Li, Zicheng Zhang, Yuan Tian, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Jia Wang, Haodong Duan, Kai Chen, et al. Information density principle for mllm benchmarks. arXiv preprint arXiv:2503.10079, 2025. 2", + "[61] Chunyi Li, Yuan Tian, Xiaoyue Ling, Zicheng Zhang, Haodong Duan, Haoning Wu, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Guo Lu, et al. Image quality assessment: From human to machine preference. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7570-7581, 2025. 2", + "[62] Chunyi Li, Jiaohao Xiao, Jianbo Zhang, Farong Wen, Zicheng Zhang, Yuan Tian, Xiangyang Zhu, Xiaohong Liu, Zhengxue Cheng, Weisi Lin, et al. Perceptual quality assessment for embodied ai. arXiv preprint arXiv:2505.16815, 2025. 2", + "[63] Yuheng Li, Krishna Kumar Singh, Utkarsh Ojha, and Yong Jae Lee. Mixmatch: Multifactor disentanglement and encoding for conditional image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8039-8048, 2020. 2", + "[64] Jie Lian, Jingyu Liu, Shu Zhang, Kai Gao, Xiaqing Liu, Dingwen Zhang, and Yizhou Yu. A structure-aware relation network for thoracic diseases detection and segmentation. IEEE Transactions on Medical Imaging, 40(8):2042-2052, 2021. 6" + ], + "bbox": [ + 99, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Weisi Lin and Li Dong. Adaptive downsampling to improve image compression at low bit rates. IEEE Transactions on Image Processing, 15(9):2513-2521, 2006. 8", + "[66] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6", + "[67] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[68] Charles Marsh. Introduction to continuous entropy. Department of Computer Science, Princeton University, 1034, 2013. 4", + "[69] Maxim Maximov, Ismail Elezi, and Laura Leal-Taixe. Ciagan: Conditional identity anonymization generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5447-5456, 2020. 2", + "[70] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4394-4402, 2018. 4", + "[71] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4", + "[72] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 4", + "[73] Eriksson Monteiro, Carlos Costa, and José Luís Oliveira. A de-identification pipeline for ultrasound medical images in dicom format. Journal of Medical Systems, 41(5):89, 2017. 1", + "[74] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 2", + "[75] Zhenliang Ni, Xinghao Chen, Yingjie Zhai, Yehui Tang, and Yunhe Wang. Context-guided spatial feature reconstruction for efficient semantic segmentation. In European Conference on Computer Vision, pages 239-255. Springer, 2024. 6", + "[76] Kai Packhäuser, Sebastian Gündel, Nicolas Münster, Christopher Syben, Vincent Christlein, and Andreas Maier. Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest x-ray data. Scientific Reports, 12(1):14851, 2022. 1", + "[77] Kai Packhäuser, Sebastian Gündel, Florian Thamm, Felix Denzinger, and Andreas Maier. Deep learning-based anonymization of chest radiographs: a utility-preserving measure for patient privacy. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 262-272. Springer, 2023. 2, 5, 6", + "[78] Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In International conference on the theory and applications of cryptographic techniques, pages 223-238. Springer, 1999. 2" + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6", + "[80] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 6", + "[81] W. Nicholson Price and I. Glenn Cohen. Privacy in the age of medical big data. Nature Medicine, 25(1):37-43, 2019. 1", + "[82] Moo-Ryong Ra, Ramesh Govindan, and Antonio Ortega. P3: Toward {Privacy-Preserving} photo sharing. In 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13), pages 515-528, 2013. 2", + "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[84] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663):3, 2009. 4", + "[85] David Rodríguez González, Trevor Carpenter, Jano I van Hemert, and Joanna Wardlaw. An open source toolkit for medical imaging de-identification. European radiology, 20: 1896-1904, 2010. 1", + "[86] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 5", + "[87] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1", + "[88] Huajie Shao, Yifei Yang, Haohong Lin, Longzhong Lin, Yizhuo Chen, Qinmin Yang, and Han Zhao. Rethinking controllable variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19250-19259, 2022. 2", + "[89] Julie K. Taitsman, Christi Macrina Grimm, and Shantanu Agrawal. Protecting patient privacy and data security. New England Journal of Medicine, 368(11):977-979, 2013. 1", + "[90] Adam Tanner. Our bodies, our data: how companies make billions selling our medical records. Beacon Press, 2017. 1", + "[91] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 2", + "[92] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Yuan Tian, Yichao Yan, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Ean: event adaptive network for enhanced action recognition. International Journal of Computer Vision, 130(10):2453-2471, 2022. 2", + "[94] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Non-semantics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 2", + "[95] Yuan Tian, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. Clsa: a contrastive learning framework with selective aggregation for video rescaling. IEEE Transactions on Image Processing, 32:1300-1314, 2023. 2", + "[96] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[97] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-vsc: Free semantics from visual foundation models for unsupervised video semantic compression. In European Conference on Computer Vision, pages 163-183. Springer, 2024.", + "[98] Yuan Tian, Guo Lu, and Guangtao Zhai. Smc++: Masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2", + "[99] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 1", + "[100] Yuan Tian, Shuo Wang, and Guangtao Zhai. Medical manifestation-aware de-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 26363–26372, 2025. 2", + "[101] Matt Tierney, Ian Spiro, Christoph Bregler, and Lakshminarayanan Subramanian. Cryptogram: Photo privacy for online social media. In Proceedings of the first ACM conference on Online social networks, pages 75-88, 2013. 2", + "[102] Gary Kin-wai Tsui and Tao Chan. Automatic selective removal of embedded patient information from image content of dicom files. American Journal of Roentgenology, 198(4): 769-772, 2012. 1", + "[103] Nishant Vishwamitra, Bart Knijnenburg, Hongxin Hu, Yifang P Kelly Caine, et al. Blur vs. block: Investigating the effectiveness of privacy-enhancing obfuscation for images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 39-47, 2017. 1, 2, 5, 6", + "[104] Zhongwei Wan, Che Liu, Mi Zhang, Jie Fu, Benyou Wang, Sibo Cheng, Lei Ma, César Quilodrán-Casas, and Rossella Arcucci. Med-unic: Unifying cross-lingual medical vision-language pre-training by diminishing bias. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[105] Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu. Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems, 35:33536-33549, 2022. 3, 5" + ], + "bbox": [ + 514, + 92, + 905, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20624", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[106] Guangyu Wang, Xiaohong Liu, Jun Shen, Chengdi Wang, Zhihuan Li, Linsen Ye, Xingwang Wu, Ting Chen, Kai Wang, Xuan Zhang, et al. A deep-learning pipeline for the diagnosis and discrimination of viral, non-viral and COVID-19 pneumonia from chest x-ray images. Nature biomedical engineering, 5(6):509–521, 2021. 1", + "[107] Jinzhuo Wang, Kai Wang, Yunfang Yu, Yuxing Lu, Wenchao Xiao, Zhuo Sun, Fei Liu, Zixing Zou, Yuanxu Gao, Lei Yang, et al. Self-improving generative foundation model for synthetic medical image generation and clinical applications. Nature Medicine, pages 1-9, 2024. 8", + "[108] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Jiankang Deng, Xinchao Wang, Hakan Bilen, and Yang You. Facemae: Privacy-preserving face recognition via masked autoencoders. arXiv preprint arXiv:2205.11090, 2022. 1, 6", + "[109] Shudong Wang, Zhiyuan Zhao, Yawu Zhao, Luqi Wang, Yuanyuan Zhang, Jiehuan Wang, Sibo Qiao, and Zhihan Lyu. A semantic conditional diffusion model for enhanced personal privacy preservation in medical images. IEEE Journal of Biomedical and Health Informatics, 2024. 2", + "[110] Weiru Wang, Chi-Man Vong, Yilong Yang, and Pak-Kin Wong. Encrypted image classification based on multilayer extreme learning machine. Multidimensional Systems and Signal Processing, 28(3):851-865, 2017. 2", + "[111] Ruiqi Wu, Chenran Zhang, Jianle Zhang, Yi Zhou, Tao Zhou, and Huazhu Fu. Mm-retinal: Knowledge-enhanced foundational pretraining with fundus image-text expertise. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 722-732. Springer, 2024. 6", + "[112] Hanyu Xue, Bo Liu, Ming Ding, Tianqing Zhu, Dayong Ye, Li Song, and Wanlei Zhou. Dp-image: Differential privacy for image data in feature space. arXiv preprint arXiv:2103.07073, 2021. 1, 2", + "[113] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 8", + "[114] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. Privacy-preserving photo sharing based on a secure JPEG. In 2015 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pages 185-190. IEEE, 2015. 2", + "[115] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. SecureJPEG scrambling enabling privacy in photo sharing. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 2", + "[116] Junhai Zhai, Sufang Zhang, Junfen Chen, and Qiang He. Autoencoder and its various variants. In 2018 IEEE international conference on systems, man, and cybernetics (SMC), pages 415-419. IEEE, 2018. 5, 6", + "[117] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2", + "[118] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for effectively learning deep face representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10823-10832, 2019. 2", + "[119] Hong-Yu Zhou, Xiaoyu Chen, Yinghao Zhang, Ruibang Luo, Liansheng Wang, and Yizhou Yu. Generalized radiograph representation learning via cross-supervision between images and free-text radiology reports. Nature Machine Intelligence, 4(1):32-40, 2022. 1", + "[120] Xinrui Zhou, Yuhao Huang, Haoran Dou, Shijing Chen, Ao Chang, Jia Liu, Weiran Long, Jian Zheng, Erjiao Xu, Jie Ren, et al. Ctrl-genaug: Controllable generative augmentation for medical sequence classification. arXiv preprint arXiv:2409.17091, 2024. 8", + "[121] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter WoodwardCourt, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 5", + "[122] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In Proceedings of the 41st International Conference on Machine Learning, pages 62429-62442. PMLR, 2024. 3", + "[123] Yingxuan Zhu, PD Singh, Khan Siddiqui, and Michael Gillam. An automatic system to detect and extract texts in medical images for de-identification. In Medical Imaging 2010: Advanced PACS-based Imaging Informatics and Therapeutic Applications, page 762803. SPIE, 2010. 1", + "[124] M Tarek Ibn Ziad, Amr Alanwar, Moustafa Alzantot, and Mani Srivastava. Cryptoimg: Privacy preserving processing over encrypted images. In 2016 IEEE Conference on Communications and Network Security (CNS), pages 570-575. IEEE, 2016. 2" + ], + "bbox": [ + 516, + 92, + 903, + 570 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "20625", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_model.json b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_model.json new file mode 100644 index 0000000000000000000000000000000000000000..06071afbba61d011e88e3505582e3429c17437a3 --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_model.json @@ -0,0 +1,3207 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.905, + 0.177 + ], + "angle": 0, + "content": "Semantics versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.205, + 0.796, + 0.257 + ], + "angle": 0, + "content": "Yuan Tian\\(^{1}\\) Shuo Wang\\(^{2}\\) Rongzhao Zhang\\(^{1}\\) Zijian Chen\\(^{2,1}\\) Yankai Jiang\\(^{1}\\) Chunyi Li\\(^{2,1}\\) Xiangyang Zhu\\(^{1}\\) Fang Yan\\(^{1}\\) Qiang Hu\\(^{3}\\) XiaoSong Wang\\(^{1}\\) Guangtao Zhai\\(^{2,1\\boxplus}\\) Shanghai AI Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.258, + 0.775, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.276, + 0.714, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)Cooperative Medianet Innovation Center, Shanghai Jiao Tong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.485, + 0.661 + ], + "angle": 0, + "content": "Medical imaging has significantly advanced computer-aided diagnosis, yet its re-identification (ReID) risks raise critical privacy concerns, calling for de-identification (DeID) techniques. Unfortunately, existing DeID methods neither particularly preserve medical semantics, nor are flexibly adjustable towards different privacy levels. To address these issues, we propose a divide-and-conquer framework comprising two steps: (1) Identity-Blocking, which blocks varying proportions of identity-related regions, to achieve different privacy levels; and (2) Medical-Semantics-Compensation, which leverages pretrained Medical Foundation Models (MFMs) to extract medical semantic features to compensate the blocked regions. Moreover, recognizing that features from MFMs may still contain residual identity information, we introduce a Minimum Description Length principle-based feature decoupling strategy, to effectively decouple and discard such identity components. Extensive evaluations against existing approaches across seven datasets and three downstream tasks, demonstrates our state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.222, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.803 + ], + "angle": 0, + "content": "In the era of digital medicine, large-scale medical images, such as X-rays and fundus photographs [6], are routinely processed by AI-based diagnostic models [28, 87, 106, 119, 121] to aid clinical decision-making. However, the increasing availability of these images raises significant concerns regarding patient privacy [17, 56, 81, 89], calling for the research on medical image de-identification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Although explicit personal details such as patient name can be easily removed from medical image headers [1, 73, 85] or burned-in texts [102, 123], re-identification (ReID) remains feasible for the intrinsic bio-identifiers, such as anatomical markers visible in chest X-rays [35, 76, 99]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.902, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.654 + ], + "angle": 0, + "content": "Figure 1. (a) Given the query medical image, the ReID model can retrieve sensitive patient information from a leaked database. (b) Our DeID framework, removing identity and then compensating medical semantics, ensures adjustable identity protection, while preserving downstream task utility. Besides, a Minimum Description Length (MDL) principle-based code space is introduced, to decouple and discard the identity information in medical features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.689 + ], + "angle": 0, + "content": "This enables sensitive information breaches [9, 52, 90], compromising patient privacy (see Figure 1(a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Several studies have attempted to defend against ReID attacks. For instance, some approaches [12, 30, 32, 44] focus on removing facial features to obfuscate identity. However, such methods cannot be applied to other body parts like the chest, where identity information is deeply interwoven with diagnostic semantics. Standard image filtering techniques, such as blurring [103], pixelation [41], and masking [108], indiscriminately degrade critical diagnostic details, thereby impairing downstream medical task performance. Moreover, under high privacy settings, the severe degradation of image quality further deteriorates task performance. Differential privacy methods [19, 27, 57, 112] mitigate identity information via noise injection, but this operation also perturbs diagnostic features. Identity adver" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.251, + 0.9 + ], + "angle": 0, + "content": "Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "sarial learning methods [55, 77] train generators by jointly maximizing identity discrepancy between the generated and original images, while minimizing distortion of medical information. Nevertheless, given the inherent entanglement between identity and diagnostic features, these methods fail to preserve diagnostic semantics at high privacy levels adequately. Recently, diagnostic annotation-conditioned generative models [15, 26, 43, 91, 109] have yielded promising results, yet they remain limited to task-specific semantics and cannot offer adjustable privacy levels. In summary, no existing method preserves task-generalizable semantics, while supporting a wide range of adjustable privacy levels." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.274, + 0.485, + 0.59 + ], + "angle": 0, + "content": "To address these issues, we introduce a novel divide-and-conquer framework DCM-DeID, which decouples identity removal from semantic preservation, to achieve semantic-rich yet adjustable de-identification. Our approach includes three steps, i.e., \\( ID \\)-Blocking, which masks identity-related regions to achieve adjustable privacy levels; Medical Semantics Extraction, which leverages pre-trained medical foundation models (MFMs) [74, 117] to extract semantic-rich medical features; Image Re-Synthesis, which employs a diffusion model [42, 86] to synthesize de-identified images, given the above ID-masked image and the medical features. Moreover, considering that the features from MFMs may also contain some identity information, we introduce a novel minimum description length [33]-based feature decoupling strategy, which excludes identity-associated information from the vanilla MFM features in a minimum-codelength latent space. This effectively prevents the reintroduction of identity information during the image resynthesis step. Through the above designs, our approach achieves a better trade-off between DeID and diagnostic utilization than prior approaches. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.59, + 0.483, + 0.666 + ], + "angle": 0, + "content": "- We reveal that existing medical DeID methods fall short in preserving task-generalizable semantics, and do not adjust seamlessly across privacy levels. We build the first benchmark for this problem, by reproducing previous approaches fairly on seven datasets." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.666, + 0.483, + 0.725 + ], + "angle": 0, + "content": "- We propose the DCM-DeID framework, which performs identity removal and medical semantics preservation in separate steps, enabling both adjustable privacy protection and medical task utility." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.726, + 0.483, + 0.786 + ], + "angle": 0, + "content": "- We introduce a Minimum Description Length-based decoupling strategy, which decouples identity cues from medical features in a compact code space, further improving the privacy protection capability." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.787, + 0.483, + 0.817 + ], + "angle": 0, + "content": "- Our framework demonstrates state-of-the-art performance. Extensive Analysis is also performed." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.59, + 0.483, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.242, + 0.847 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Image Privacy Protection. Early methods applied low-level filters to obscure image details, including downsampling [20], blurring [103], and pixelation [41]. Later, en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "cryption in alternate domains such as JPEG bitstreams [82, 101] and DCT coefficients [114, 115] was explored, though these often introduced severe distortions that hindered downstream tasks. Homomorphic encryption [110, 124] addresses inference on encrypted images, but suffers from high computational cost [78] and limited compatibility with advanced models like Vision Transformers [24, 93, 95]. Additionally, approaches for face images [12, 34, 69] leverage facial priors from StyleGAN [53] or face recognition networks [22, 118], which may not readily generalize to other domains, such as the medical-domain images in our work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.483 + ], + "angle": 0, + "content": "Medical Image De-Identification. Early methods (e.g., FreeSurfer [30], PyDeface [32], SynthStrip [44]) focus on removing facial features in brain MRI. For common medical images, early approaches use pixel-domain filters (like blurring [103] and pixelation [41]) or frequency-domain techniques [31], but these hand-crafted solutions also severely degrade the image details, leading to substantially degraded performance and visual quality [59, 60, 62]. Differential Privacy methods [19, 27, 57, 112] inject noise into the training data, which compromises inference-time utility. More recent generative models [15, 26, 43, 91, 109] synthesize images conditioned on disease labels or lesion masks. However, they tend to lack task generalizability and struggle to balance privacy-utility trade-offs, which are addressed by our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.905, + 0.68 + ], + "angle": 0, + "content": "Feature Decoupling. Variational auto-encoder (VAE)-based works [40, 88] decouple representations, by constraining the variables in latent space independent. Generative adversarial network (GAN)-based methods [14, 63] are unsupervised, leaving factors unaligned with explicit semantic or identity information. For face images, there are methods [23, 53, 54, 100] targeting identity separation. However, these methods rely on strong facial priors that may not generalize to medical images. In contrast, motivated by the explicit entropy regularization terms adopted in data compression field [2, 45-47, 61, 92, 94, 96-98], our approach effectively decouples identity in medical images, within a minimum-codelength space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.649, + 0.703 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.825 + ], + "angle": 0, + "content": "In this section, we first describe the medical re-identification (ReID) models used for privacy attacks. Next, we introduce our de-identification model, which divides the task into two stages. First, identity information is removed via region blocking with an adjustable threshold. Second, lost medical semantics are compensated. This approach flexibly adjusts privacy while preserving rich, generalizable medical features for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.719, + 0.849 + ], + "angle": 0, + "content": "3.1. Medical ReID Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Given a query medical image, ReID models aim to retrieve all images belonging to the same individual, from a medical record database. Concretely, the model first extracts identity" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20614" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.075, + 0.861, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.386, + 0.907, + 0.499 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed divide-and-conquer framework, DCM-DeID. (a) \\( ID \\)-Blocking: A pre-trained ReID network produces the identity-similarity map, which is binarized by different thresholds to adjust privacy level. (b) \\( ID \\)-Free Medical Semantics Extraction: Medical foundation models (MFMs) extract features that are encoded into a code space under the minimum-codelength regularization. A learned mask partitions the codes into identity- and medical semantics-related ones, where only the latter one is preserved. (c) Image Re-Synthesis: A diffusion model re-synthesizes images that are privacy-preserving and semantics-rich, generalizing to various downstream tasks. We illustrate with X-ray images, but the framework is also applicable to other modalities such as fundus images. \\(\\diamond\\) denotes frozen models, while gray dashed lines indicate components used solely for learning identity-semantic decoupling. The channel number of codes \\( Q \\) is arbitrary; two channels are shown for conciseness. \\(\\otimes\\) denotes the element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.482, + 0.643 + ], + "angle": 0, + "content": "(ID) embedding from the query image, and then compare it with that of each image within the database. Then, the image with the closest Euclidean distance is adopted as the re-identified image. We build two medical ReID models, i.e., ViT [24] and VisionMamba [122]-based ones, which are separately adopted in the training and the evaluation stages. These ReID models are optimized with a combination of classification loss and triplet loss [39], following the previous object ReID work [38]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.383, + 0.669 + ], + "angle": 0, + "content": "3.2. A Divide-and-Conquer Approach" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.795 + ], + "angle": 0, + "content": "To defend against attacks on medical ReID models, we propose DCM-DeID, a divide-and-conquer approach for medical image de-identification. DCM-DeID operates in three stages: ID Blocking, which removes identity-related image regions; ID-Free Medical Semantics Extraction, which extracts rich medical information without reintroducing identity information; and Image Re-Synthesis, which generates the final de-identified medical image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.904 + ], + "angle": 0, + "content": "ID-Blocking. Given an input image \\(X \\in \\mathbb{R}^{3 \\times H \\times W}\\), where \\(H\\) and \\(W\\) denote the image spatial scales, we use a ViT-based ReID model to extract local features \\(f \\in \\mathbb{R}^{768 \\times h \\times w}\\), where \\(h = H / 16\\) and \\(w = W / 16\\). Spatial average pooling is applied to \\(f\\) to obtain an identity embedding \\(id \\in \\mathbb{R}^{768}\\). For each spatial position in \\(f\\), the cosine similarity with \\(id\\) is computed, resulting in a similarity map \\(S[i] =\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\(\\cos(f[i],id)\\), where \\(i\\) denotes the spatial location. Then, \\(S\\) is binarized by a threshold \\(T\\). Finally, the ID-blocked image is computed as: \\(X^{noID} = X \\odot \\mathrm{Upsample}(S > T)\\), where Upsample denotes nearest-neighbor interpolation to match the resolution of \\(S\\) to \\(X\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.734 + ], + "angle": 0, + "content": "ID-Free Medical Semantics Extraction. Although \\( X^{noID} \\) effectively removes identity information, it inevitably distorts medical cues such as lung shadows. To amend this, we employ pre-trained medical foundation models (MFMs), e.g., MGCA [105] for X-ray images, to extract rich medical feature \\( f_{MFM} \\) from \\( X \\). Since \\( f_{MFM} \\) contains both semantic cues and local details that may encode identity, we introduce a feature decoupling strategy (Section 3.3) to decouple and remove the identity information, yielding the identity-free semantic feature \\( \\hat{f}_{sem} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Image Re-Synthesis. Given \\( X^{noID} \\) and \\( \\hat{f}_{sem} \\), a dual-conditioned diffusion model synthesizes the de-identified image that inherits the rich semantics within MFMs, while also protecting privacy. Since the synthesized image is highly realistic, it can be directly deployed to downstream medical AI applications, without further adaptation. The model details are elaborated in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.787, + 0.866 + ], + "angle": 0, + "content": "3.3. Medical Semantics Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Medical features extracted by the MFM encode both diagnostic semantics (e.g., lesion morphology) and identity" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "related cues (e.g., rib patterns in chest X-rays). For effective privacy-preserving, it is imperative to decouple these two types of information, and discard the identity cues. We achieve this by learning a minimum-length code space, and separating the two parts in this space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Theoretical Motivation. From an information-theoretic perspective, the Minimum Description Length (MDL) principle [4, 33] states that the best representation for a given set of data is the one that minimizes the total codelength needed to describe the data, where each group of features tends to capture the independent or low-correlation information parts. In our context, let \\( Q \\) be the latent representation of the MFM feature \\( f_{\\mathrm{MFM}} \\) and let \\( H(Q) \\) denote its expected codelength. The MDL principle objective can be seen as balancing a reconstruction loss and a compression term, i.e., the so-called rate-distortion loss (RD loss) [5]:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.338, + 0.483, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o d e - a l l}} = \\min _ {\\mathcal {E}, \\mathcal {D}} \\underbrace {\\| f _ {M F M} - \\hat {f} _ {M F M} \\| _ {2}} _ {\\text {F e a t u r e R e c o n s t r u c t i o n}} + \\underbrace {\\beta H (\\boldsymbol {Q})} _ {\\text {C o d e l e n g t h}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.426 + ], + "angle": 0, + "content": "where \\(Q = \\mathcal{E}(f_{MFM})\\), \\(\\hat{f}_{MFM} = \\mathcal{D}(Q)\\), and \\(\\beta\\) denotes balancing weight. \\(\\mathcal{E}\\) and \\(\\mathcal{D}\\) represent a pair of feature encoder and decoder networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.483, + 0.548 + ], + "angle": 0, + "content": "Discrete Code-based Codelength Estimation. Directly calculating the \\( H(\\mathbf{Q}) \\) for the continuous variable \\( \\mathbf{Q} \\) is nontrivial [68]. Fortunately, the neural data compression community [2, 3, 70, 72] have verified that the codelength of integer latent variables can be quite precisely estimated with a learnable entropy model. Therefore, we append the quantization operation at the tail of the encoder \\( \\mathcal{E} \\), to make elements within \\( \\mathbf{Q} \\) discrete values, and estimate its codelength." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Concretely, \\(\\mathcal{E}\\) comprises three residual blocks [36] with 256 channels, followed by a convolutional layer to reduce dimensionality and a rounding operation that outputs a 32-channel integer code \\(Q\\). The decoder network \\(\\mathcal{D}\\) is symmetric to \\(\\mathcal{E}\\), except it omits the rounding operation. During training, the straight-through estimator [71] is employed to backpropagate gradients through the rounding step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Following [2], the expected codelength of encoding \\(Q\\) is calculated as the log-likelihood, i.e., \\(H(Q) = -\\log_2p(Q)\\), where the probability \\(p(Q)\\) is modeled using a Gaussian Mixture Model (GMM) [84] with \\(K\\) components:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.717, + 0.483, + 0.757 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {Q}) = \\sum_ {k = 1} ^ {K} w ^ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {Q}; \\mu^ {k}, e ^ {\\sigma^ {k}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\{\\mathbf{w},\\boldsymbol {\\mu},\\boldsymbol {\\sigma}\\}\\) are the learnable mixture weights, means, and log variance scalers of the GMM components, respectively, which are shared across spatial positions, not unshared along the channel axis [2]. Following [16], \\(K\\) is set to three. For each integer element \\(q\\in\\) \\(Q\\) , the probability is computed over the quantization bin [18, 72], \\(p(q) = \\mathcal{F}(q + 0.5) - \\mathcal{F}(q - 0.5)\\) , where \\(\\mathcal{F}(x) = \\sum_{k = 1}^{K}w^{k}\\Phi \\Big(x;\\mu^{k},e^{\\sigma^{k}}\\Big)\\) is the cumulative distribution function (CDF) of the Gaussian Mixture Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.088, + 0.905, + 0.142 + ], + "angle": 0, + "content": "(GMM), \\(\\Phi \\left(x;\\mu ,e^{\\sigma}\\right) = \\frac{1}{2}\\left[1 + \\mathrm{erf}\\left(\\frac{x - \\mu}{\\sqrt{2}e^{\\sigma}}\\right)\\right]\\). We not that the CDF can be efficiently calculated by the modern deep learning framework such as PyTorch [80]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.143, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Learning of Identity-Associated Code Mask. A single convolution layer predicts a binary mask \\(\\mathbf{M}\\) from \\(Q\\), with the same dimensions as \\(Q\\). The Gumbel-Softmax algorithm [50] is applied to enable gradient propagation through the binary mask. The identity-associated codes are then obtained by element-wise masking, \\(Q_{\\mathrm{id}} = Q\\odot \\mathbf{M}\\). A lightweight convolutional network, composed of three residual blocks followed by average pooling, predicts the identity embedding \\(\\hat{\\boldsymbol{id}}\\) from \\(Q_{\\mathrm{id}}\\). Then, the RD loss for reconstructing identity can be given by:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.303, + 0.905, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o d e - i d}} = \\| \\hat {i d} - i d \\| _ {2} + \\beta H \\left(\\boldsymbol {Q} _ {\\text {i d}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.331, + 0.905, + 0.378 + ], + "angle": 0, + "content": "where \\( H(\\tilde{Q}_{\\mathrm{id}}) \\) is calculated similarly to \\( H(Q) \\), sharing the same GMM parameters and balancing weight \\( \\beta \\) as in Equation 1, since they operate in the same latent space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.906, + 0.47 + ], + "angle": 0, + "content": "Reconstruction of Medical Semantics. By suppressing identity-related codes via the inverse mask \\((1 - \\mathbf{M})\\), we obtain the semantics-part codes \\(Q_{sem} = (1 - \\mathbf{M})\\otimes Q\\). Finally, the final ID-free medical semantic feature is reconstructed as: \\(\\hat{f}_{sem} = \\mathcal{D}(Q_{sem})\\), which preserves critical diagnostic semantics, excluding the identity information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.755, + 0.499 + ], + "angle": 0, + "content": "3.4. Image Re-Synthesis Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Given the ID-masked image \\( X^{noID} \\) and the ID-free medical semantic feature \\( \\hat{f}_{sem} \\), we employ a diffusion model to synthesize de-identified medical images. First, we utilize a Feat-Net to project the high-resolution \\( X^{noID} \\) into the low-resolution feature \\( f^{noID} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}} \\). The Feat-Net consists of the VAE encoder from Stable Diffusion [86], followed by two convolution layers of kernel size 5 and stride size 2. Next, we adopt a bi-directional crossattention mechanism [13] to fuse \\( f^{noID} \\) and \\( \\hat{f}_{sem} \\), producing a fused feature \\( f_{fuse} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}} \\), which is further processed through a series of convolutional layers. This produces a set of features with dimensions matching those of the UNet's intermediate feature maps within the diffusion model. These features are added to the UNet layers, guiding the diffusion process toward two objectives: maintaining the privacy level of \\( X^{noID} \\), while preserving the medical semantics in \\( \\hat{f}_{sem} \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.776, + 0.691, + 0.792 + ], + "angle": 0, + "content": "3.5. Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.799, + 0.905, + 0.83 + ], + "angle": 0, + "content": "The whole framework is end-to-end optimized, with the following objective," + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.838, + 0.905, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o d e - a l l}} + \\mathcal {L} _ {\\text {c o d e - i d}} + \\mathcal {L} _ {\\text {d i f f u s e}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.86, + 0.906, + 0.906 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{diffuse}}\\) denotes the diffusion loss [42]. We do not introduce the balancing weight, since we found directly adding the loss terms already achieves satisfactory results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.075, + 0.298, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.226, + 0.207, + 0.236 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.075, + 0.5, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.226, + 0.411, + 0.236 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.075, + 0.703, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.226, + 0.614, + 0.236 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.073, + 0.907, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.226, + 0.819, + 0.236 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.236, + 0.298, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.387, + 0.208, + 0.396 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.236, + 0.501, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.387, + 0.41, + 0.396 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.236, + 0.704, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.387, + 0.614, + 0.397 + ], + "angle": 0, + "content": "(g)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.236, + 0.907, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.387, + 0.818, + 0.396 + ], + "angle": 0, + "content": "(h)" + }, + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.399, + 0.779, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Identity-performance trade-off curves of various medical privacy protection methods." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.413, + 0.902, + 0.632 + ], + "angle": 0, + "content": "
Attack SRMethodX-ray Classify AUROC (%)X-ray Caption BLEUX-ray Seg Dice (%)Fundus Classify AUROC (%)Fundus Seg Dice (%)
MIMIC-XChest-XRayCheXpertMIMIC-XChestX-DetEyePACSODIR5KRefuge2
10%Pixel-Blur [103]64.1560.4874.450.12597.5344.8356.1722.35
Feat-Noise [116]65.8466.2566.490.145310.7447.7655.5324.85
ID-Adv [77]59.3661.7573.110.113122.2356.2858.5041.23
Privacy-Net [55]57.0959.5171.470.132915.7552.6357.2939.23
MAE [37]59.1165.2677.050.11869.9450.9156.5027.33
Ours72.8673.6686.120.175027.9865.2359.9654.66
20%Pixel-Blur [103]67.2364.2780.670.130711.2549.5657.4234.11
Feat-Noise [116]68.2769.0180.780.159514.2851.2957.1737.20
ID-Adv [77]64.0862.7177.530.119322.8363.2459.2849.47
Privacy-Net [55]62.1162.6380.250.143420.5561.0158.2547.47
MAE [37]66.4371.3883.520.138313.9954.9758.0140.13
Ours74.3575.0186.320.185929.4969.5960.4162.04
Original82.1384.8287.240.321852.8981.4661.5390.08
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.635, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Table 1. Performance comparison of medical image privacy protection methods, under different attack Success-Rates (SR). For measuring SR, we adopt CMC-R1 metric for MIMIC-X, Chest-Xray, CheXpert, EyePACS, and ODIR5K, using ID-R metric for ChestX-Det and Refuge2. Original denotes the performance on original images, which is the performance upper-bound of privacy-removal images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.224, + 0.705 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.308, + 0.729 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.9 + ], + "angle": 0, + "content": "For Med-ReID models, we adopt the AdamW optimizer [66] during training, with a learning rate of 1e-5 scheduled by cosine decaying strategy and a weight decay of 1e-2. The training process consists of 300,000 steps. The batch size is 256. We apply random cropping and blurring as image augmentation strategies, and the input image resolution to networks is \\(256 \\times 256\\). The ViT-based models are initialized with CLIP-pretrained weights [83], while the VisionMamba-based models are initialized with ImageNet-pretrained weights [21]. Training a single ReID model takes about 24 hours with four NVIDIA RTX 4090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For DCM-DeID model, the UNet within the diffusion model follows the same architecture as the Stable Diffusion [86], also performing the diffusion procedure in the latent space. The feature channels within UNet are reduced to [128, 256, 512, 1024], for the four stages of both the down-pathway and up-pathway, to reduce computational cost. The identity-similarity map threshold \\( T \\) is defined as the \\( r \\)-th quantile of the similarity map \\( S \\). \\( r \\) is selected from [0.95, 0.7, 0.4, 0.2] to cover wide privacy levels. We adopt MGCA-ResNet [105] and RetFound-ViT [121] MFMs for X-ray and fundus images, respectively. During training, we apply random flipping and random cropping \\( 256 \\times 256 \\) patches for data augmentation. The codelength loss term weight \\( \\beta \\) is set to 0.5. At test time, we resize the shorter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "side of the images to 256 and then center-crop the middle \\(256 \\times 256\\) region. The learning rate is set to 1e-4 and is gradually decayed with the cosine annealing strategy [67]. The total number of training steps is 800,000. The mini-batch size is 64. We utilize the AdamW optimizer [66] implemented in PyTorch [80] with CUDA support. The momentum parameters are set as \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.99\\), and the gradient norm is clipped to a maximum value of 1. The entire training process takes about three days on a machine equipped with eight NVIDIA RTX 4090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.194, + 0.269 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.58 + ], + "angle": 0, + "content": "We evaluate our approach on two medical image modalities: chest X-rays and eye fundus photographs, with seven public datasets. For the chest X-ray modality, we split the MIMIC-X dataset [51] into training, validation, and test sets using an 8:1:1 ratio. For the Chest-Xray and CheXpert datasets, we randomly select \\(10\\%\\) patients as the test set. We also adopt the ChestX-Det dataset [64] to evaluate the X-ray segmentation task. For the eye fundus modality, we divide the EyePACS dataset [25] into training, validation, and test sets with an 8:1:1 ratio, and we use the Refuge2 dataset [29] to evaluate the fundus segmentation task. ODIR5K [8] is also adopted for evaluating the fundus classification task of systemic diseases such as hypertension. Note that only MIMIC-X and EyePACS are used during training; all other datasets, which differ in environment, demographics, and imaging devices, are never seen during training, to evaluate the domain generalizability of our approach. MIMIC-X, CheXpert, Chest-Xray, EyePACS, ChestXDet, and ODIR5K contain 377K, 224K, 112K, 88K, 3.6K, and 5K images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.44, + 0.607 + ], + "angle": 0, + "content": "4.3. Reproduced Privacy Protection Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.642 + ], + "angle": 0, + "content": "We implement several privacy protection methods, comparing them with our approach in a fair setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Pixel-Blur [103]. This method applies a Gaussian blur to the input image. We experiment with standard deviations of \\(\\{1,5,10,20\\}\\) to vary the level of de-identification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Feat-Noise [116]. We train an autoencoder [116] and inject Gaussian noise into its latent features. The noise level is selected from \\(\\{0.1, 0.8, 0.85, 0.9, 0.95\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "ID-Adv [77]. A UNet is trained to generate a de-identified image \\( Y \\) from the original image \\( X \\), optimizing the loss \\( \\mathcal{L} = \\lambda \\cos(id_X, id_Y) + \\|med_X - med_Y\\|_2 + \\mathcal{L}_{reg} \\), where \\( id_X \\) and \\( id_Y \\) are identity features extracted by a ViT-based ReID model, and \\( med_X \\) and \\( med_Y \\) are medical features obtained from MFMs same as our approach. \\( \\mathcal{L}_{reg} \\) is a GAN regularization loss ensuring visual plausibility, \\( \\cos(\\cdot, \\cdot) \\) denotes cosine similarity, and \\( \\| \\cdot \\|_2 \\) the \\( \\ell_2 \\) norm. The trade-off weight \\( \\lambda \\) is chosen from \\( \\{0.1, 0.5, 1, 2\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Privacy-Net [55]. This method updates the identity model and the de-identification network adversarially, enhanc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ing de-identification performance. The original PrivacyNet focuses solely on segmentation tasks, supervised by segmentation masks. To enable task-agnostic de-identification, we train it using the same objective as ID-Adv. Since the identity model is adversarially updated and are stronger, we use smaller \\(\\lambda\\) values compared to ID-Adv, i.e., \\(\\{0.05, 0.25, 0.5, 1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.32 + ], + "angle": 0, + "content": "MAE [37]. Following [108], we transfer the concept of masked auto-encoders (MAE) [37] to the adjustable privacy protection problem, by masking a random proportion of patches to obscure identity information. It adopts the same diffusion model as our approach to generate the masked regions. This model can also serve as a degenerated version of our model, where both semantic compensation and identity-region similarity designs are removed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.33, + 0.75, + 0.344 + ], + "angle": 0, + "content": "4.4. Downstream Task Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.534 + ], + "angle": 0, + "content": "For the identity recognition, we adopt the VisionMamba-based ReID model, which differs from the ViT-based model employed during the training of privacy protection methods, ensuring the method generalization capability across different ReID models. For the \\(X\\)-ray classification, we use the ViT model pre-trained with Med-UniC [104]. For \\(X\\)-ray captioning, we employ the visual-language model CXRLLaVA-v2 [58], which is specifically designed for X-ray images. For \\(X\\)-ray segmentation, we adopt CGRSeg [75]. For fundus classification, we use the ViT model pre-trained with KeepFit [111]. Finally, for fundus segmentation, given the limited dataset size, we employ nnUNet [49]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.697, + 0.558 + ], + "angle": 0, + "content": "4.5. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.809 + ], + "angle": 0, + "content": "For privacy evaluation, we adopt the cumulative matching characteristics (CMC) [10] at Rank-1, i.e., CMC-R1, on datasets with patient ID information available (i.e., MIMIC-X, Chest-Xray, CheXpert, EyePacs, and ODIR5K). For datasets without patient ID information (i.e., CheX-det and REFuge2), we adopt the recognition rate, i.e., ID-R, which determines whether the distance between the ID feature of the original and de-identified image exceeds a predefined threshold. The thresholds are set to 1.1 and 1.35 for the X-ray and fundus modalities, respectively, based on statistics from the validation sets of MIMIC-X and EyePACS. For the disease diagnosis task, we employ the area under the receiver operating characteristic curve (AUROC) metric [11]; for the image captioning task, we use the bilingual evaluation understudy (BLEU) metric [79]; and for the image segmentation task, we adopt the Dice score metric [7]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.608, + 0.833 + ], + "angle": 0, + "content": "4.6. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "X-ray Classification. As shown in Table 1, our method substantially outperforms other approaches, achieving AU-ROCs of \\(72.86\\%\\), \\(73.66\\%\\), and \\(86.12\\%\\) on MIMIC-X, Chest-XRay, and Chexpert, respectively, under CMC" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20618" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "\\(\\mathrm{R}1 = 10\\%\\) Notably, although our model is trained on MIMIC-X, it generalizes well to the other two datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.35 + ], + "angle": 0, + "content": "Among the other compared approaches, Feat-Noise obtains the second-best performance, i.e., AUROC of \\(65.84\\%\\) at CMC-R1=10% on MIMIX-X, by condensing image pixels into a compact latent feature space. In contrast, methods that jointly optimize a trade-off between de-identification and medical preservation, i.e., ID-Adv and PrivacyNet, yield unsatisfactory performances. As shown in Figure 3 (a), under the ID-R=5% setting, ID-Adv and PrivacyNet attain AUROCs of only \\(56.32\\%\\) and \\(54.21\\%\\) on MIMIC-X, respectively, which are much lower than the simple pixel blurring baseline \\((61.62\\%)\\). This indicates that directly optimizing the two conflicting objectives is suboptimal. In contrast, our approach decouples the objectives into two separate steps, identity removal and medical semantic compensation, achieving consistently superior performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.488 + ], + "angle": 0, + "content": "As for MAE, which employs the same diffusion model as ours, it achieves competitive results at a high attacking rate, with an AUROC of \\(76.12\\%\\) @CMC-R1=40% on MIMIC-X, outperforming all other approaches except ours. However, at a low attacking rate CMC-R1=10%, it falls behind our method by over \\(13\\%\\) AUROC. This highlights that our superior performance is not solely due to the generative power of the diffusion model, but rather stems from the effectiveness of our core idea of semantic compensation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.58 + ], + "angle": 0, + "content": "X-ray Caption. As shown in Table 1, our method attains a BLEU score of 0.1750, remarkably surpassing Pixel-Blur (0.1259), Feat-Noise (0.1453), and MAE (0.1186), at CMC-R1=10%. This proves that our approach can comprehensively preserve the clinic-required information, beyond only the classification label." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.809 + ], + "angle": 0, + "content": "X-ray Segmentation. Furthermore, we evaluate the methods on a fine-grained task: segmentation. As shown in Table 1, at ID-R1=10%, our method achieves a Dice score of 27.98%, outperforming Pixel-Blur (7.53%), Feat-Noise (10.74%), Privacy-Net (15.75%), ID-Adv (22.23%), and MAE(9.94%). This proves that our semantic compensation scheme not only preserves the global semantics for classification, but also effectively retains the local semantics for segmentation. Pixel-Blur and Feat-Noise perform poorly, since they severely corrupts the image details. In contrast, ID-Adv and Privacy-Net, which incorporate a medical feature-matching loss, achieve slightly decent performance, but still lag far behind our approach. For instance, under ID-R=80%, our method outperforms Privacy-Net by approximately 8%, as shown in Figure 3(e)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Fundus Classification. Beyond X-ray images, our method also proves effective on fundus data. For instance, on EyePACS and ORID5K, our approach outperforms the second-best competitor ID-Adv by about \\(9\\%\\) and \\(1\\%\\), respectively. These results confirm that our approach generalizes well across different imaging modalities." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.707, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.091, + 0.898, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Figure 4. (Left) Ablation on the framework design. (Right) Ablation study on the feature decoupling strategy." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.615, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.346, + 0.585, + 0.356 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.269, + 0.712, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.346, + 0.676, + 0.356 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.269, + 0.801, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.346, + 0.777, + 0.356 + ], + "angle": 0, + "content": "w/oSem" + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.269, + 0.901, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.346, + 0.88, + 0.357 + ], + "angle": 0, + "content": "w/oDecouple" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.363, + 0.905, + 0.404 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of different variant models. The models are described in Figure 4 caption. Red arrow denotes the modified identity-related features. Best to view by zooming-in." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Fundus Segmentation. Our method achieves a Dice score of \\(54.66\\%\\) on REFUGE2 at ID-R1=10%, largely surpassing MAE (27.33%), Privacy-Net (39.23%), and ID-Adv (41.23%). This further validates that our method also effectively preserves fine-grained semantic cues of eye fundus." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.669, + 0.515 + ], + "angle": 0, + "content": "4.7. Model Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Framework-Level Ablation Study. As shown in Figure 4 (Left), by removing the semantic branch, the AUROC of the resulted model 'w/oSem' dramatically drops by over \\(8\\%\\), at CMC-R1=5%. On the other hand, without the identity-semantics decoupling mechanism, the resulting model 'w/oDecouple' leads to about \\(15\\%\\) CMC-R1 increase, for achieving the similar AUROC performance, since substantial identity cues are leaked from the vanilla medical features of MFMs. We further illustrate the protected images from different models. As shown in Figure 5, our results effectively modify identity-related features, such as the shape and location of the clavicle and chest contour. The 'w/oSem' model also removes these regions but significantly alters medical manifestations. In contrast, the 'w/oDecouple' model preserves medical features but fails to sufficiently suppress identity-related features, such as clavicle shape, due to residual identity information in the features from MFMs. These results confirm that both medical semantics and identity-semantics decoupling are essential for our advanced medical DeID approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Besides, we attempt to remove ID-Blocking module, tuning privacy via activating more decoupled ID codes. This leads to degraded AUROC (72.01%, 73.47%, 74.68%), compared to using ID-Blocking (72.68%, 74.35%, 79.62%) on MIMIC-X (10%, 20%, 40% at tack SR settings), due to" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.09, + 0.482, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.226 + ], + "angle": 0, + "content": "Figure 6. Comparison of semantic and identity information in terms of Bits-per-Pixel (bpp) [65], calculated as the feature code-length divided by the original image size." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.234, + 0.35, + 0.248 + ], + "angle": 0, + "content": "the complete loss of pixel-wise details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.369 + ], + "angle": 0, + "content": "Ablation Study on the Decoupling Strategy. As shown in Figure 4 (Right), omitting the codelength loss terms ('w/oCodeLengthLoss') fails to effectively remove identity information from MFM features, since the loose space cannot effectively decouple the identity and the semantics information. Moreover, removing the discrete code bottleneck ('w/oCodeSpace') further exacerbates identity leakage, leading to further inferior performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Furthermore, we quantitatively compare the overall and identity-related information in MFM features, as shown in Figure 6. First, we notice that a significant portion is identity-related, i.e., around \\(44\\%\\) and \\(55\\%\\) for X-ray and fundus images. Second, the average information amount of the X-ray dataset Chest-Xray is 0.23bpp, much higher than 0.11bpp achieved by the fundus dataset EyePACS. This is aligned with the medical knowledge prior, that X-rays capture multiple organs and tissues, containing much complex information, than the fundus image that only focuses on eyes. This proves that the learned codelength effectively describes the medical data characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Finally, we analyze the impact of the codelength loss weight \\(\\beta\\) and the latent code channel number. As shown in Figure 7 (Left), reducing \\(\\beta\\) from 0.5 to 0.1 significantly increases CMC-R1 from \\(5.85\\%\\) to \\(12.34\\%\\), as a loosely constrained code space fails to effectively decouple identity information. Conversely, increasing \\(\\beta\\) from 0.5 to 2 has little effect on CMC-R1 but reduces AUROC performance by approximately \\(6\\%\\), as an overly strong constraint impairs semantic feature reconstruction. The number of code channels also influences performance, by tuning the information capacity of the latent code, as shown in Figure 7 (Right). However, since \\(\\beta\\) directly regulates the code-length term, the impact of the channel number is limited." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Discussion with Label-Conditioned Diffusion Models. These methods [48, 107, 120] employ task-specific labels (e.g., disease labels or text reports) to synthesize images, which are limited to label-associated tasks. In contrast, our approach is task-agnostic and applicable to diverse tasks. Moreover, after fine-tuning our approach towards a single task, i.e., replacing the MFM with a supervised classification network, our method achieves \\(81.92\\%\\) AUROC at CMC-R1 = 0.30%, surpassing the label-conditioned model," + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.092, + 0.709, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.091, + 0.895, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.214, + 0.906, + 0.257 + ], + "angle": 0, + "content": "Figure 7. (Left) Impact of the rate-distortion weight \\(\\beta\\). (Right:) Impact of the code dimension. All experiments are evaluated by masking \\(95\\%\\) identity-related regions, for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.338 + ], + "angle": 0, + "content": "i.e., \\(80.79\\%\\) AUROC at CMC-R1 = 0.29%. This confirms that our minimum-codelength representation also benefits the single-task setting, compared to the methods directly using the task labels guiding the diffusion procedure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Model Complexity. All methods and our model comprise about 380M parameters, for a fair comparison. Our inference time is 540 ms on an NVIDIA 4090 GPU, which is similar to MAE (526ms), but slower than Privacy-Net (120 ms), ID-Adv (122 ms), and Feat-Noise (124 ms), due to the multiple inference steps of diffusion procedure. Nonetheless, given the significant performance gains and that the medical imaging procedure itself is time-consuming, the running time is acceptable and does not hinder clinical workflows. In the future, we will integrate the single-step diffusion technique [113] to accelerate the process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.855, + 0.551 + ], + "angle": 0, + "content": "5. Conclusion, Future Works, and Other" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Conclusion. We have presented DCM-DeID, a divide-and-conquer framework for medical image de-identification. By leveraging pre-trained Medical Foundation Models and a minimum codelength-based feature decoupling strategy, our method effectively remove identity cues, while preserving medical task utility. Extensive evaluations demonstrate the superiority of our approach. Future Works. We will extend our approach to multi-slice images, such as those produced by Magnetic Resonance Imaging (MRI). Broader Impacts. Our DeID technique is designed for medical AI applications, aiding the human. We emphasize that all rigorous clinical decisions must be made by human physicians using the original medical images. Furthermore, it is critical to enforce ethical guidelines, working in synergy with technological approaches to achieve medical privacy protection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No.62225112), the Fundamental Research Funds for the Central Universities, National Key R&D Program of China 2021YFE0206700, Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and STCSM 22DZ2229005." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20620" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Kadek YE, Aryanto, Andre Broekema, Matthijs Oudkerk, and Peter MA van Ooijen. Implementation of an anonymisation tool for clinical trials using a clinical trial processor integrated with an existing trial patient data information system. European radiology, 22:144-151, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.186, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.228, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.284, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[4] Andrew Barron, Jorma Rissanen, and Bin Yu. The minimum description length principle in coding and modeling. IEEE transactions on information theory, 44(6):2743-2760, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.34, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[5] Toby Berger. Rate-distortion theory. Wiley Encyclopedia of Telecommunications, 2003. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.369, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[6] Rui Bernardes, Pedro Serranho, and Conceicao Lobo. Digital ocular fundus imaging: a review. Ophthalmologica, 226 (4):161-181, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.411, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 92-100. Springer, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.523, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[8] Amit Bhati, Neha Gour, Pritee Khanna, and Aparajita Ojha. Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image dataset. Computers in Biology and Medicine, 153:106519, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.593, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[9] Vincent Bindschaedler, Paul Grubbs, David Cash, Thomas Ristenpart, and Vitaly Shmatikov. The tao of inference in privacy-protected databases. Cryptology ePrint Archive, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.649, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Ruud M Bolle, Jonathan H Connell, Sharath Pankanti, Nalini K Ratha, and Andrew W Senior. The relation between the roc curve and the cmc. In Fourth IEEE workshop on automatic identification advanced technologies (AutoID'05), pages 15-20. IEEE, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.72, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145-1159, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[12] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3334–3342, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Yan Chen and Pouyan Esmaeilzadeh. Generative ai in medical practice: in-depth exploration of privacy and security challenges. Journal of Medical Internet Research, 26:e53008, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[17] I Glenn Cohen and Michelle M Mello. Big data, big tech, and protecting patient privacy. *Jama*, 322(12):1141-1142, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.334, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[18] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.364, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Obfuscation of images via differential privacy: From facial images to general images. Peer-to-Peer Networking and Applications, 14:1705-1733, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.421, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[20] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.478, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.535, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.606, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[23] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.677, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.775, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[25] Emma Dugas, Jared, Jorge, and Will Cukierski. Diabetic retinopathy detection. https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. Kaggle. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] August DuMont Schütte, Jürgen Hetzel, Sergios Gatidis, Tobias Hepp, Benedikt Dietz, Stefan Bauer, and Patrick Schwab. Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation. NPJ digital medicine, 4(1):141, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20621" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[28] Weijie Fan, Yi Yang, Jing Qi, Qichuan Zhang, Cuiwei Liao, Li Wen, Shuang Wang, Guangxian Wang, Yu Xia, Qihua Wu, et al. A deep-learning-based framework for identifying and localizing multiple abnormalities and assessing cardiomegaly in chest x-ray. Nature Communications, 15(1): 1347, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[29] Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, et al. Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. arXiv preprint arXiv:2202.08994, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[30] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774-781, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[31] Alex Gaudio, Asim Smailagic, Christos Faloutsos, Shreshta Mohan, Elvin Johnson, Yuhao Liu, Pedro Costa, and Aurélio Campilho. Deepfixcx: Explainable privacy-preserving image compression for medical image analysis. Wiley interdisciplinary reviews: Data mining and knowledge discovery, 13(4):e1495, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[32] Satrajit Ghosh, Chris Gorgolewski, et al. pydeface: A tool to remove facial features from mri images, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[33] Peter D Grünwald. The minimum description length principle. MIT press, 2007. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[34] Xiuye Gu, Weixin Luo, Michael S Ryoo, and Yong Jae Lee. Password-conditioned anonymization and deanonymization with face identity transformers. In European conference on computer vision, pages 727-743. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[35] A Gueziec and P Kazanzides. Anatomy-based registration of ct-scan and intraoperative x-ray images for guiding a surgical robot. IEEE Transactions on Medical Imaging, 17(5): 715-728, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[38] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[39] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Steven Hill, Zhimin Zhou, Lawrence Saul, and Hovav Shacham. On the (in) effectiveness of mosaicing and blurring as tools for document redaction. Proceedings on Privacy Enhancing Technologies, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[42] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.19, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[43] Sungmin Hong, Razvan Marinescu, Adrian V Dalca, Anna K Bonkhoff, Martin Bretzner, Natalia S Rost, and Polina Golland. 3d-stylegan: A style-based generative adversarial network for generative modeling of three-dimensional medical images. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 24-34. Springer, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.33, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[44] Andrew Hoopes, Jocelyn S Mora, Adrian V Dalca, Bruce Fischl, and Malte Hoffmann. Synthstrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[45] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. arXiv preprint arXiv:2501.13630, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.441, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[46] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 875-885, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[47] Qiang Hu, Houqiang Zhong, Zihan Zheng, Xiaoyun Zhang, Zhengxue Cheng, Li Song, Guangtao Zhai, and Yanfeng Wang. Vrvvc: Variable-rate nerf-based volumetric video compression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3563-3571, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[48] Peng Huang, Xue Gao, Lihong Huang, Jing Jiao, Xiaokang Li, Yuanyuan Wang, and Yi Guo. Chest-diffusion: a lightweight text-to-image model for report-to-cxr generation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.664, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[49] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.734, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[50] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[51] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Georgios A Kaissis, Marcus R Makowski, Daniel Rückert, and Rickmer F Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6):305-311, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20622" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[53] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.486, + 0.219 + ], + "angle": 0, + "content": "[54] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.485, + 0.289 + ], + "angle": 0, + "content": "[55] Bach Ngoc Kim, Jose Dolz, Pierre-Marc Jodoin, and Christian Desrosiers. Privacy-net: an adversarial approach for identity-obfuscated segmentation of medical images. IEEE Transactions on Medical Imaging, 40(7):1737-1749, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[56] Cody A Koch and Wayne F Larrabee. Patient privacy, photographs, and publication. JAMA facial plastic surgery, 15 (5):335-336, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.335, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[57] Abhinav Kumar, Sanjay Kumar Singh, K Lakshmanan, Sonal Saxena, and Sameer Shrivastava. A novel cloud-assisted secure deep feature classification framework for cancer histopathology images. ACM Transactions on Internet Technology (TOIT), 21(2):1-22, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[58] Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim, and Soon Ho Yoon. Cxr-llava: a multimodal large language model for interpreting chest x-ray images. European Radiology, pages 1-13, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[59] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8):6833-6846, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[60] Chunyi Li, Xiaozhe Li, Zicheng Zhang, Yuan Tian, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Jia Wang, Haodong Duan, Kai Chen, et al. Information density principle for mllm benchmarks. arXiv preprint arXiv:2503.10079, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.485, + 0.688 + ], + "angle": 0, + "content": "[61] Chunyi Li, Yuan Tian, Xiaoyue Ling, Zicheng Zhang, Haodong Duan, Haoning Wu, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Guo Lu, et al. Image quality assessment: From human to machine preference. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7570-7581, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[62] Chunyi Li, Jiaohao Xiao, Jianbo Zhang, Farong Wen, Zicheng Zhang, Yuan Tian, Xiangyang Zhu, Xiaohong Liu, Zhengxue Cheng, Weisi Lin, et al. Perceptual quality assessment for embodied ai. arXiv preprint arXiv:2505.16815, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[63] Yuheng Li, Krishna Kumar Singh, Utkarsh Ojha, and Yong Jae Lee. Mixmatch: Multifactor disentanglement and encoding for conditional image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8039-8048, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[64] Jie Lian, Jingyu Liu, Shu Zhang, Kai Gao, Xiaqing Liu, Dingwen Zhang, and Yizhou Yu. A structure-aware relation network for thoracic diseases detection and segmentation. IEEE Transactions on Medical Imaging, 40(8):2042-2052, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[65] Weisi Lin and Li Dong. Adaptive downsampling to improve image compression at low bit rates. IEEE Transactions on Image Processing, 15(9):2513-2521, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[66] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[67] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[68] Charles Marsh. Introduction to continuous entropy. Department of Computer Science, Princeton University, 1034, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.251, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[69] Maxim Maximov, Ismail Elezi, and Laura Leal-Taixe. Ciagan: Conditional identity anonymization generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5447-5456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.321, + 0.908, + 0.39 + ], + "angle": 0, + "content": "[70] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4394-4402, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.392, + 0.905, + 0.434 + ], + "angle": 0, + "content": "[71] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.435, + 0.905, + 0.491 + ], + "angle": 0, + "content": "[72] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.492, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[73] Eriksson Monteiro, Carlos Costa, and José Luís Oliveira. A de-identification pipeline for ultrasound medical images in dicom format. Journal of Medical Systems, 41(5):89, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.549, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[74] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.619, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[75] Zhenliang Ni, Xinghao Chen, Yingjie Zhai, Yehui Tang, and Yunhe Wang. Context-guided spatial feature reconstruction for efficient semantic segmentation. In European Conference on Computer Vision, pages 239-255. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.689, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[76] Kai Packhäuser, Sebastian Gündel, Nicolas Münster, Christopher Syben, Vincent Christlein, and Andreas Maier. Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest x-ray data. Scientific Reports, 12(1):14851, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.761, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[77] Kai Packhäuser, Sebastian Gündel, Florian Thamm, Felix Denzinger, and Andreas Maier. Deep learning-based anonymization of chest radiographs: a utility-preserving measure for patient privacy. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 262-272. Springer, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.845, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[78] Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In International conference on the theory and applications of cryptographic techniques, pages 223-238. Springer, 1999. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20623" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[79] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.485, + 0.245 + ], + "angle": 0, + "content": "[80] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[81] W. Nicholson Price and I. Glenn Cohen. Privacy in the age of medical big data. Nature Medicine, 25(1):37-43, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[82] Moo-Ryong Ra, Ramesh Govindan, and Antonio Ortega. P3: Toward {Privacy-Preserving} photo sharing. In 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13), pages 515-528, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.427, + 0.484, + 0.454 + ], + "angle": 0, + "content": "[84] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663):3, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.455, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[85] David Rodríguez González, Trevor Carpenter, Jano I van Hemert, and Joanna Wardlaw. An open source toolkit for medical imaging de-identification. European radiology, 20: 1896-1904, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[86] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.581, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[87] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.65, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[88] Huajie Shao, Yifei Yang, Haohong Lin, Longzhong Lin, Yizhuo Chen, Qinmin Yang, and Han Zhao. Rethinking controllable variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19250-19259, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.72, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[89] Julie K. Taitsman, Christi Macrina Grimm, and Shantanu Agrawal. Protecting patient privacy and data security. New England Journal of Medicine, 368(11):977-979, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.762, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[90] Adam Tanner. Our bodies, our data: how companies make billions selling our medical records. Beacon Press, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[91] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[92] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[93] Yuan Tian, Yichao Yan, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Ean: event adaptive network for enhanced action recognition. International Journal of Computer Vision, 130(10):2453-2471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[94] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Non-semantics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[95] Yuan Tian, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. Clsa: a contrastive learning framework with selective aggregation for video rescaling. IEEE Transactions on Image Processing, 32:1300-1314, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[96] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.334, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[97] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-vsc: Free semantics from visual foundation models for unsupervised video semantic compression. In European Conference on Computer Vision, pages 163-183. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.392, + 0.906, + 0.433 + ], + "angle": 0, + "content": "[98] Yuan Tian, Guo Lu, and Guangtao Zhai. Smc++: Masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.435, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[99] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.505, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[100] Yuan Tian, Shuo Wang, and Guangtao Zhai. Medical manifestation-aware de-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 26363–26372, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.562, + 0.906, + 0.618 + ], + "angle": 0, + "content": "[101] Matt Tierney, Ian Spiro, Christoph Bregler, and Lakshminarayanan Subramanian. Cryptogram: Photo privacy for online social media. In Proceedings of the first ACM conference on Online social networks, pages 75-88, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.62, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[102] Gary Kin-wai Tsui and Tao Chan. Automatic selective removal of embedded patient information from image content of dicom files. American Journal of Roentgenology, 198(4): 769-772, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.676, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[103] Nishant Vishwamitra, Bart Knijnenburg, Hongxin Hu, Yifang P Kelly Caine, et al. Blur vs. block: Investigating the effectiveness of privacy-enhancing obfuscation for images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 39-47, 2017. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.761, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[104] Zhongwei Wan, Che Liu, Mi Zhang, Jie Fu, Benyou Wang, Sibo Cheng, Lei Ma, César Quilodrán-Casas, and Rossella Arcucci. Med-unic: Unifying cross-lingual medical vision-language pre-training by diminishing bias. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[105] Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu. Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems, 35:33536-33549, 2022. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[106] Guangyu Wang, Xiaohong Liu, Jun Shen, Chengdi Wang, Zhihuan Li, Linsen Ye, Xingwang Wu, Ting Chen, Kai Wang, Xuan Zhang, et al. A deep-learning pipeline for the diagnosis and discrimination of viral, non-viral and COVID-19 pneumonia from chest x-ray images. Nature biomedical engineering, 5(6):509–521, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[107] Jinzhuo Wang, Kai Wang, Yunfang Yu, Yuxing Lu, Wenchao Xiao, Zhuo Sun, Fei Liu, Zixing Zou, Yuanxu Gao, Lei Yang, et al. Self-improving generative foundation model for synthetic medical image generation and clinical applications. Nature Medicine, pages 1-9, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[108] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Jiankang Deng, Xinchao Wang, Hakan Bilen, and Yang You. Facemae: Privacy-preserving face recognition via masked autoencoders. arXiv preprint arXiv:2205.11090, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[109] Shudong Wang, Zhiyuan Zhao, Yawu Zhao, Luqi Wang, Yuanyuan Zhang, Jiehuan Wang, Sibo Qiao, and Zhihan Lyu. A semantic conditional diffusion model for enhanced personal privacy preservation in medical images. IEEE Journal of Biomedical and Health Informatics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[110] Weiru Wang, Chi-Man Vong, Yilong Yang, and Pak-Kin Wong. Encrypted image classification based on multilayer extreme learning machine. Multidimensional Systems and Signal Processing, 28(3):851-865, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[111] Ruiqi Wu, Chenran Zhang, Jianle Zhang, Yi Zhou, Tao Zhou, and Huazhu Fu. Mm-retinal: Knowledge-enhanced foundational pretraining with fundus image-text expertise. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 722-732. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[112] Hanyu Xue, Bo Liu, Ming Ding, Tianqing Zhu, Dayong Ye, Li Song, and Wanlei Zhou. Dp-image: Differential privacy for image data in feature space. arXiv preprint arXiv:2103.07073, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[113] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[114] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. Privacy-preserving photo sharing based on a secure JPEG. In 2015 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pages 185-190. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[115] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. SecureJPEG scrambling enabling privacy in photo sharing. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[116] Junhai Zhai, Sufang Zhang, Junfen Chen, and Qiang He. Autoencoder and its various variants. In 2018 IEEE international conference on systems, man, and cybernetics (SMC), pages 415-419. IEEE, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[117] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[118] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "for effectively learning deep face representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10823-10832, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[119] Hong-Yu Zhou, Xiaoyu Chen, Yinghao Zhang, Ruibang Luo, Liansheng Wang, and Yizhou Yu. Generalized radiograph representation learning via cross-supervision between images and free-text radiology reports. Nature Machine Intelligence, 4(1):32-40, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[120] Xinrui Zhou, Yuhao Huang, Haoran Dou, Shijing Chen, Ao Chang, Jia Liu, Weiran Long, Jian Zheng, Erjiao Xu, Jie Ren, et al. Ctrl-genaug: Controllable generative augmentation for medical sequence classification. arXiv preprint arXiv:2409.17091, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[121] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter WoodwardCourt, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[122] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In Proceedings of the 41st International Conference on Machine Learning, pages 62429-62442. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[123] Yingxuan Zhu, PD Singh, Khan Siddiqui, and Michael Gillam. An automatic system to detect and extract texts in medical images for de-identification. In Medical Imaging 2010: Advanced PACS-based Imaging Informatics and Therapeutic Applications, page 762803. SPIE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[124] M Tarek Ibn Ziad, Amr Alanwar, Moustafa Alzantot, and Mani Srivastava. Cryptoimg: Privacy preserving processing over encrypted images. In 2016 IEEE Conference on Communications and Network Security (CNS), pages 570-575. IEEE, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20625" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_origin.pdf b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46c1df2a0937009cd9a9d1e647a79faab96e67bc --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/2b39ad1e-d18f-4579-85b8-6d56df6a82db_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1110b5804eb049038e0d3b9fe4e0a4bac31456d67fdbd2cb7e72b4ff6f7fba46 +size 2287192 diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/full.md b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de731e1c41a06daf070b93a3e7e3dc58cd497ab3 --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/full.md @@ -0,0 +1,395 @@ +# Semantics versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification + +Yuan Tian $^{1}$ Shuo Wang $^{2}$ Rongzhao Zhang $^{1}$ Zijian Chen $^{2,1}$ Yankai Jiang $^{1}$ Chunyi Li $^{2,1}$ Xiangyang Zhu $^{1}$ Fang Yan $^{1}$ Qiang Hu $^{3}$ XiaoSong Wang $^{1}$ Guangtao Zhai $^{2,1\boxplus}$ Shanghai AI Laboratory + +$^{2}$ Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University + +$^{3}$ Cooperative Medianet Innovation Center, Shanghai Jiao Tong University + +# Abstract + +Medical imaging has significantly advanced computer-aided diagnosis, yet its re-identification (ReID) risks raise critical privacy concerns, calling for de-identification (DeID) techniques. Unfortunately, existing DeID methods neither particularly preserve medical semantics, nor are flexibly adjustable towards different privacy levels. To address these issues, we propose a divide-and-conquer framework comprising two steps: (1) Identity-Blocking, which blocks varying proportions of identity-related regions, to achieve different privacy levels; and (2) Medical-Semantics-Compensation, which leverages pretrained Medical Foundation Models (MFMs) to extract medical semantic features to compensate the blocked regions. Moreover, recognizing that features from MFMs may still contain residual identity information, we introduce a Minimum Description Length principle-based feature decoupling strategy, to effectively decouple and discard such identity components. Extensive evaluations against existing approaches across seven datasets and three downstream tasks, demonstrates our state-of-the-art performance. + +# 1. Introduction + +In the era of digital medicine, large-scale medical images, such as X-rays and fundus photographs [6], are routinely processed by AI-based diagnostic models [28, 87, 106, 119, 121] to aid clinical decision-making. However, the increasing availability of these images raises significant concerns regarding patient privacy [17, 56, 81, 89], calling for the research on medical image de-identification. + +Although explicit personal details such as patient name can be easily removed from medical image headers [1, 73, 85] or burned-in texts [102, 123], re-identification (ReID) remains feasible for the intrinsic bio-identifiers, such as anatomical markers visible in chest X-rays [35, 76, 99]. + +![](images/a90263befa2c62fee1b9677538860a46d6829df0ee37ae5e085e3d9e7645b20b.jpg) +Figure 1. (a) Given the query medical image, the ReID model can retrieve sensitive patient information from a leaked database. (b) Our DeID framework, removing identity and then compensating medical semantics, ensures adjustable identity protection, while preserving downstream task utility. Besides, a Minimum Description Length (MDL) principle-based code space is introduced, to decouple and discard the identity information in medical features. + +This enables sensitive information breaches [9, 52, 90], compromising patient privacy (see Figure 1(a)). + +Several studies have attempted to defend against ReID attacks. For instance, some approaches [12, 30, 32, 44] focus on removing facial features to obfuscate identity. However, such methods cannot be applied to other body parts like the chest, where identity information is deeply interwoven with diagnostic semantics. Standard image filtering techniques, such as blurring [103], pixelation [41], and masking [108], indiscriminately degrade critical diagnostic details, thereby impairing downstream medical task performance. Moreover, under high privacy settings, the severe degradation of image quality further deteriorates task performance. Differential privacy methods [19, 27, 57, 112] mitigate identity information via noise injection, but this operation also perturbs diagnostic features. Identity adver + +sarial learning methods [55, 77] train generators by jointly maximizing identity discrepancy between the generated and original images, while minimizing distortion of medical information. Nevertheless, given the inherent entanglement between identity and diagnostic features, these methods fail to preserve diagnostic semantics at high privacy levels adequately. Recently, diagnostic annotation-conditioned generative models [15, 26, 43, 91, 109] have yielded promising results, yet they remain limited to task-specific semantics and cannot offer adjustable privacy levels. In summary, no existing method preserves task-generalizable semantics, while supporting a wide range of adjustable privacy levels. + +To address these issues, we introduce a novel divide-and-conquer framework DCM-DeID, which decouples identity removal from semantic preservation, to achieve semantic-rich yet adjustable de-identification. Our approach includes three steps, i.e., $ID$ -Blocking, which masks identity-related regions to achieve adjustable privacy levels; Medical Semantics Extraction, which leverages pre-trained medical foundation models (MFMs) [74, 117] to extract semantic-rich medical features; Image Re-Synthesis, which employs a diffusion model [42, 86] to synthesize de-identified images, given the above ID-masked image and the medical features. Moreover, considering that the features from MFMs may also contain some identity information, we introduce a novel minimum description length [33]-based feature decoupling strategy, which excludes identity-associated information from the vanilla MFM features in a minimum-codelength latent space. This effectively prevents the reintroduction of identity information during the image resynthesis step. Through the above designs, our approach achieves a better trade-off between DeID and diagnostic utilization than prior approaches. Our contributions are: + +- We reveal that existing medical DeID methods fall short in preserving task-generalizable semantics, and do not adjust seamlessly across privacy levels. We build the first benchmark for this problem, by reproducing previous approaches fairly on seven datasets. +- We propose the DCM-DeID framework, which performs identity removal and medical semantics preservation in separate steps, enabling both adjustable privacy protection and medical task utility. +- We introduce a Minimum Description Length-based decoupling strategy, which decouples identity cues from medical features in a compact code space, further improving the privacy protection capability. +- Our framework demonstrates state-of-the-art performance. Extensive Analysis is also performed. + +# 2. Related Works + +Image Privacy Protection. Early methods applied low-level filters to obscure image details, including downsampling [20], blurring [103], and pixelation [41]. Later, en + +cryption in alternate domains such as JPEG bitstreams [82, 101] and DCT coefficients [114, 115] was explored, though these often introduced severe distortions that hindered downstream tasks. Homomorphic encryption [110, 124] addresses inference on encrypted images, but suffers from high computational cost [78] and limited compatibility with advanced models like Vision Transformers [24, 93, 95]. Additionally, approaches for face images [12, 34, 69] leverage facial priors from StyleGAN [53] or face recognition networks [22, 118], which may not readily generalize to other domains, such as the medical-domain images in our work. + +Medical Image De-Identification. Early methods (e.g., FreeSurfer [30], PyDeface [32], SynthStrip [44]) focus on removing facial features in brain MRI. For common medical images, early approaches use pixel-domain filters (like blurring [103] and pixelation [41]) or frequency-domain techniques [31], but these hand-crafted solutions also severely degrade the image details, leading to substantially degraded performance and visual quality [59, 60, 62]. Differential Privacy methods [19, 27, 57, 112] inject noise into the training data, which compromises inference-time utility. More recent generative models [15, 26, 43, 91, 109] synthesize images conditioned on disease labels or lesion masks. However, they tend to lack task generalizability and struggle to balance privacy-utility trade-offs, which are addressed by our approach. + +Feature Decoupling. Variational auto-encoder (VAE)-based works [40, 88] decouple representations, by constraining the variables in latent space independent. Generative adversarial network (GAN)-based methods [14, 63] are unsupervised, leaving factors unaligned with explicit semantic or identity information. For face images, there are methods [23, 53, 54, 100] targeting identity separation. However, these methods rely on strong facial priors that may not generalize to medical images. In contrast, motivated by the explicit entropy regularization terms adopted in data compression field [2, 45-47, 61, 92, 94, 96-98], our approach effectively decouples identity in medical images, within a minimum-codelength space. + +# 3. Methodology + +In this section, we first describe the medical re-identification (ReID) models used for privacy attacks. Next, we introduce our de-identification model, which divides the task into two stages. First, identity information is removed via region blocking with an adjustable threshold. Second, lost medical semantics are compensated. This approach flexibly adjusts privacy while preserving rich, generalizable medical features for downstream tasks. + +# 3.1. Medical ReID Models + +Given a query medical image, ReID models aim to retrieve all images belonging to the same individual, from a medical record database. Concretely, the model first extracts identity + +![](images/ac8e1bfc66603ba13b1bf15a2825c68727523f679f897b3770ce1205b4272735.jpg) +Figure 2. Overview of the proposed divide-and-conquer framework, DCM-DeID. (a) $ID$ -Blocking: A pre-trained ReID network produces the identity-similarity map, which is binarized by different thresholds to adjust privacy level. (b) $ID$ -Free Medical Semantics Extraction: Medical foundation models (MFMs) extract features that are encoded into a code space under the minimum-codelength regularization. A learned mask partitions the codes into identity- and medical semantics-related ones, where only the latter one is preserved. (c) Image Re-Synthesis: A diffusion model re-synthesizes images that are privacy-preserving and semantics-rich, generalizing to various downstream tasks. We illustrate with X-ray images, but the framework is also applicable to other modalities such as fundus images. $\diamond$ denotes frozen models, while gray dashed lines indicate components used solely for learning identity-semantic decoupling. The channel number of codes $Q$ is arbitrary; two channels are shown for conciseness. $\otimes$ denotes the element-wise multiplication. + +(ID) embedding from the query image, and then compare it with that of each image within the database. Then, the image with the closest Euclidean distance is adopted as the re-identified image. We build two medical ReID models, i.e., ViT [24] and VisionMamba [122]-based ones, which are separately adopted in the training and the evaluation stages. These ReID models are optimized with a combination of classification loss and triplet loss [39], following the previous object ReID work [38]. + +# 3.2. A Divide-and-Conquer Approach + +To defend against attacks on medical ReID models, we propose DCM-DeID, a divide-and-conquer approach for medical image de-identification. DCM-DeID operates in three stages: ID Blocking, which removes identity-related image regions; ID-Free Medical Semantics Extraction, which extracts rich medical information without reintroducing identity information; and Image Re-Synthesis, which generates the final de-identified medical image. + +ID-Blocking. Given an input image $X \in \mathbb{R}^{3 \times H \times W}$ , where $H$ and $W$ denote the image spatial scales, we use a ViT-based ReID model to extract local features $f \in \mathbb{R}^{768 \times h \times w}$ , where $h = H / 16$ and $w = W / 16$ . Spatial average pooling is applied to $f$ to obtain an identity embedding $id \in \mathbb{R}^{768}$ . For each spatial position in $f$ , the cosine similarity with $id$ is computed, resulting in a similarity map $S[i] =$ + +$\cos(f[i],id)$ , where $i$ denotes the spatial location. Then, $S$ is binarized by a threshold $T$ . Finally, the ID-blocked image is computed as: $X^{noID} = X \odot \mathrm{Upsample}(S > T)$ , where Upsample denotes nearest-neighbor interpolation to match the resolution of $S$ to $X$ . + +ID-Free Medical Semantics Extraction. Although $X^{noID}$ effectively removes identity information, it inevitably distorts medical cues such as lung shadows. To amend this, we employ pre-trained medical foundation models (MFMs), e.g., MGCA [105] for X-ray images, to extract rich medical feature $f_{MFM}$ from $X$ . Since $f_{MFM}$ contains both semantic cues and local details that may encode identity, we introduce a feature decoupling strategy (Section 3.3) to decouple and remove the identity information, yielding the identity-free semantic feature $\hat{f}_{sem}$ . + +Image Re-Synthesis. Given $X^{noID}$ and $\hat{f}_{sem}$ , a dual-conditioned diffusion model synthesizes the de-identified image that inherits the rich semantics within MFMs, while also protecting privacy. Since the synthesized image is highly realistic, it can be directly deployed to downstream medical AI applications, without further adaptation. The model details are elaborated in Section 3.4. + +# 3.3. Medical Semantics Decoupling + +Medical features extracted by the MFM encode both diagnostic semantics (e.g., lesion morphology) and identity + +related cues (e.g., rib patterns in chest X-rays). For effective privacy-preserving, it is imperative to decouple these two types of information, and discard the identity cues. We achieve this by learning a minimum-length code space, and separating the two parts in this space. + +Theoretical Motivation. From an information-theoretic perspective, the Minimum Description Length (MDL) principle [4, 33] states that the best representation for a given set of data is the one that minimizes the total codelength needed to describe the data, where each group of features tends to capture the independent or low-correlation information parts. In our context, let $Q$ be the latent representation of the MFM feature $f_{\mathrm{MFM}}$ and let $H(Q)$ denote its expected codelength. The MDL principle objective can be seen as balancing a reconstruction loss and a compression term, i.e., the so-called rate-distortion loss (RD loss) [5]: + +$$ +\mathcal {L} _ {\text {c o d e - a l l}} = \min _ {\mathcal {E}, \mathcal {D}} \underbrace {\| f _ {M F M} - \hat {f} _ {M F M} \| _ {2}} _ {\text {F e a t u r e R e c o n s t r u c t i o n}} + \underbrace {\beta H (\boldsymbol {Q})} _ {\text {C o d e l e n g t h}}, \tag {1} +$$ + +where $Q = \mathcal{E}(f_{MFM})$ , $\hat{f}_{MFM} = \mathcal{D}(Q)$ , and $\beta$ denotes balancing weight. $\mathcal{E}$ and $\mathcal{D}$ represent a pair of feature encoder and decoder networks. + +Discrete Code-based Codelength Estimation. Directly calculating the $H(\mathbf{Q})$ for the continuous variable $\mathbf{Q}$ is nontrivial [68]. Fortunately, the neural data compression community [2, 3, 70, 72] have verified that the codelength of integer latent variables can be quite precisely estimated with a learnable entropy model. Therefore, we append the quantization operation at the tail of the encoder $\mathcal{E}$ , to make elements within $\mathbf{Q}$ discrete values, and estimate its codelength. + +Concretely, $\mathcal{E}$ comprises three residual blocks [36] with 256 channels, followed by a convolutional layer to reduce dimensionality and a rounding operation that outputs a 32-channel integer code $Q$ . The decoder network $\mathcal{D}$ is symmetric to $\mathcal{E}$ , except it omits the rounding operation. During training, the straight-through estimator [71] is employed to backpropagate gradients through the rounding step. + +Following [2], the expected codelength of encoding $Q$ is calculated as the log-likelihood, i.e., $H(Q) = -\log_2p(Q)$ , where the probability $p(Q)$ is modeled using a Gaussian Mixture Model (GMM) [84] with $K$ components: + +$$ +p (\boldsymbol {Q}) = \sum_ {k = 1} ^ {K} w ^ {k} \cdot \mathcal {N} \left(\boldsymbol {Q}; \mu^ {k}, e ^ {\sigma^ {k}}\right), \tag {2} +$$ + +where $\{\mathbf{w},\boldsymbol {\mu},\boldsymbol {\sigma}\}$ are the learnable mixture weights, means, and log variance scalers of the GMM components, respectively, which are shared across spatial positions, not unshared along the channel axis [2]. Following [16], $K$ is set to three. For each integer element $q\in$ $Q$ , the probability is computed over the quantization bin [18, 72], $p(q) = \mathcal{F}(q + 0.5) - \mathcal{F}(q - 0.5)$ , where $\mathcal{F}(x) = \sum_{k = 1}^{K}w^{k}\Phi \Big(x;\mu^{k},e^{\sigma^{k}}\Big)$ is the cumulative distribution function (CDF) of the Gaussian Mixture Model + +(GMM), $\Phi \left(x;\mu ,e^{\sigma}\right) = \frac{1}{2}\left[1 + \mathrm{erf}\left(\frac{x - \mu}{\sqrt{2}e^{\sigma}}\right)\right]$ . We not that the CDF can be efficiently calculated by the modern deep learning framework such as PyTorch [80]. + +Learning of Identity-Associated Code Mask. A single convolution layer predicts a binary mask $\mathbf{M}$ from $Q$ , with the same dimensions as $Q$ . The Gumbel-Softmax algorithm [50] is applied to enable gradient propagation through the binary mask. The identity-associated codes are then obtained by element-wise masking, $Q_{\mathrm{id}} = Q\odot \mathbf{M}$ . A lightweight convolutional network, composed of three residual blocks followed by average pooling, predicts the identity embedding $\hat{\boldsymbol{id}}$ from $Q_{\mathrm{id}}$ . Then, the RD loss for reconstructing identity can be given by: + +$$ +\mathcal {L} _ {\text {c o d e - i d}} = \| \hat {i d} - i d \| _ {2} + \beta H \left(\boldsymbol {Q} _ {\text {i d}}\right), \tag {3} +$$ + +where $H(\tilde{Q}_{\mathrm{id}})$ is calculated similarly to $H(Q)$ , sharing the same GMM parameters and balancing weight $\beta$ as in Equation 1, since they operate in the same latent space. + +Reconstruction of Medical Semantics. By suppressing identity-related codes via the inverse mask $(1 - \mathbf{M})$ , we obtain the semantics-part codes $Q_{sem} = (1 - \mathbf{M})\otimes Q$ . Finally, the final ID-free medical semantic feature is reconstructed as: $\hat{f}_{sem} = \mathcal{D}(Q_{sem})$ , which preserves critical diagnostic semantics, excluding the identity information. + +# 3.4. Image Re-Synthesis Model + +Given the ID-masked image $X^{noID}$ and the ID-free medical semantic feature $\hat{f}_{sem}$ , we employ a diffusion model to synthesize de-identified medical images. First, we utilize a Feat-Net to project the high-resolution $X^{noID}$ into the low-resolution feature $f^{noID} \in \mathbb{R}^{512 \times \frac{H}{32} \times \frac{W}{32}}$ . The Feat-Net consists of the VAE encoder from Stable Diffusion [86], followed by two convolution layers of kernel size 5 and stride size 2. Next, we adopt a bi-directional crossattention mechanism [13] to fuse $f^{noID}$ and $\hat{f}_{sem}$ , producing a fused feature $f_{fuse} \in \mathbb{R}^{512 \times \frac{H}{32} \times \frac{W}{32}}$ , which is further processed through a series of convolutional layers. This produces a set of features with dimensions matching those of the UNet's intermediate feature maps within the diffusion model. These features are added to the UNet layers, guiding the diffusion process toward two objectives: maintaining the privacy level of $X^{noID}$ , while preserving the medical semantics in $\hat{f}_{sem}$ . + +# 3.5. Learning Strategy + +The whole framework is end-to-end optimized, with the following objective, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o d e - a l l}} + \mathcal {L} _ {\text {c o d e - i d}} + \mathcal {L} _ {\text {d i f f u s e}}, \tag {4} +$$ + +where $\mathcal{L}_{\mathrm{diffuse}}$ denotes the diffusion loss [42]. We do not introduce the balancing weight, since we found directly adding the loss terms already achieves satisfactory results. + +![](images/c99259f96619449aa904ee38d86f8de71df13aaea8e70f3bc5f3cc3cdc2adebf.jpg) + +![](images/808fc601ee7308eefabd34039ba22fd34d7aab8b1065800b3e7b11889bd437ba.jpg) + +![](images/3d21caf0f38b5ab4c364b97222113ffbe277f29a2365623c9bbd21927028cd87.jpg) + +![](images/d665c432f1e60176e8a8d202611a287d670558f16f9793fe46c229356addef9c.jpg) + +![](images/1e557411b02a3bb757f64fcc717e16de5c1cc89d18c05a51fa944b5acff47adc.jpg) +(a) +(e) + +![](images/2faff10e5a4edea8300716b970cf5868f398bbfba667288db02c7e9e2240961b.jpg) +(b) +(f) + +![](images/10ac4a2770487f28e742ddaedb9dea3b441d98ed1049d509b8218643496ae457.jpg) +(c) +(g) + +![](images/e1f20bd9d447e113e2051a015d7447551d6395951025e02778cdc5afd2b6fd45.jpg) +(d) +(h) + +Figure 3. Identity-performance trade-off curves of various medical privacy protection methods. + +
Attack SRMethodX-ray Classify AUROC (%)X-ray Caption BLEUX-ray Seg Dice (%)Fundus Classify AUROC (%)Fundus Seg Dice (%)
MIMIC-XChest-XRayCheXpertMIMIC-XChestX-DetEyePACSODIR5KRefuge2
10%Pixel-Blur [103]64.1560.4874.450.12597.5344.8356.1722.35
Feat-Noise [116]65.8466.2566.490.145310.7447.7655.5324.85
ID-Adv [77]59.3661.7573.110.113122.2356.2858.5041.23
Privacy-Net [55]57.0959.5171.470.132915.7552.6357.2939.23
MAE [37]59.1165.2677.050.11869.9450.9156.5027.33
Ours72.8673.6686.120.175027.9865.2359.9654.66
20%Pixel-Blur [103]67.2364.2780.670.130711.2549.5657.4234.11
Feat-Noise [116]68.2769.0180.780.159514.2851.2957.1737.20
ID-Adv [77]64.0862.7177.530.119322.8363.2459.2849.47
Privacy-Net [55]62.1162.6380.250.143420.5561.0158.2547.47
MAE [37]66.4371.3883.520.138313.9954.9758.0140.13
Ours74.3575.0186.320.185929.4969.5960.4162.04
Original82.1384.8287.240.321852.8981.4661.5390.08
+ +Table 1. Performance comparison of medical image privacy protection methods, under different attack Success-Rates (SR). For measuring SR, we adopt CMC-R1 metric for MIMIC-X, Chest-Xray, CheXpert, EyePACS, and ODIR5K, using ID-R metric for ChestX-Det and Refuge2. Original denotes the performance on original images, which is the performance upper-bound of privacy-removal images. + +# 4. Experiments + +# 4.1. Implementation Details + +For Med-ReID models, we adopt the AdamW optimizer [66] during training, with a learning rate of 1e-5 scheduled by cosine decaying strategy and a weight decay of 1e-2. The training process consists of 300,000 steps. The batch size is 256. We apply random cropping and blurring as image augmentation strategies, and the input image resolution to networks is $256 \times 256$ . The ViT-based models are initialized with CLIP-pretrained weights [83], while the VisionMamba-based models are initialized with ImageNet-pretrained weights [21]. Training a single ReID model takes about 24 hours with four NVIDIA RTX 4090 GPUs. + +For DCM-DeID model, the UNet within the diffusion model follows the same architecture as the Stable Diffusion [86], also performing the diffusion procedure in the latent space. The feature channels within UNet are reduced to [128, 256, 512, 1024], for the four stages of both the down-pathway and up-pathway, to reduce computational cost. The identity-similarity map threshold $T$ is defined as the $r$ -th quantile of the similarity map $S$ . $r$ is selected from [0.95, 0.7, 0.4, 0.2] to cover wide privacy levels. We adopt MGCA-ResNet [105] and RetFound-ViT [121] MFMs for X-ray and fundus images, respectively. During training, we apply random flipping and random cropping $256 \times 256$ patches for data augmentation. The codelength loss term weight $\beta$ is set to 0.5. At test time, we resize the shorter + +side of the images to 256 and then center-crop the middle $256 \times 256$ region. The learning rate is set to 1e-4 and is gradually decayed with the cosine annealing strategy [67]. The total number of training steps is 800,000. The mini-batch size is 64. We utilize the AdamW optimizer [66] implemented in PyTorch [80] with CUDA support. The momentum parameters are set as $\beta_{1} = 0.9$ and $\beta_{2} = 0.99$ , and the gradient norm is clipped to a maximum value of 1. The entire training process takes about three days on a machine equipped with eight NVIDIA RTX 4090 GPUs. + +# 4.2. Datasets + +We evaluate our approach on two medical image modalities: chest X-rays and eye fundus photographs, with seven public datasets. For the chest X-ray modality, we split the MIMIC-X dataset [51] into training, validation, and test sets using an 8:1:1 ratio. For the Chest-Xray and CheXpert datasets, we randomly select $10\%$ patients as the test set. We also adopt the ChestX-Det dataset [64] to evaluate the X-ray segmentation task. For the eye fundus modality, we divide the EyePACS dataset [25] into training, validation, and test sets with an 8:1:1 ratio, and we use the Refuge2 dataset [29] to evaluate the fundus segmentation task. ODIR5K [8] is also adopted for evaluating the fundus classification task of systemic diseases such as hypertension. Note that only MIMIC-X and EyePACS are used during training; all other datasets, which differ in environment, demographics, and imaging devices, are never seen during training, to evaluate the domain generalizability of our approach. MIMIC-X, CheXpert, Chest-Xray, EyePACS, ChestXDet, and ODIR5K contain 377K, 224K, 112K, 88K, 3.6K, and 5K images. + +# 4.3. Reproduced Privacy Protection Methods + +We implement several privacy protection methods, comparing them with our approach in a fair setting. + +Pixel-Blur [103]. This method applies a Gaussian blur to the input image. We experiment with standard deviations of $\{1,5,10,20\}$ to vary the level of de-identification. + +Feat-Noise [116]. We train an autoencoder [116] and inject Gaussian noise into its latent features. The noise level is selected from $\{0.1, 0.8, 0.85, 0.9, 0.95\}$ . + +ID-Adv [77]. A UNet is trained to generate a de-identified image $Y$ from the original image $X$ , optimizing the loss $\mathcal{L} = \lambda \cos(id_X, id_Y) + \|med_X - med_Y\|_2 + \mathcal{L}_{reg}$ , where $id_X$ and $id_Y$ are identity features extracted by a ViT-based ReID model, and $med_X$ and $med_Y$ are medical features obtained from MFMs same as our approach. $\mathcal{L}_{reg}$ is a GAN regularization loss ensuring visual plausibility, $\cos(\cdot, \cdot)$ denotes cosine similarity, and $\| \cdot \|_2$ the $\ell_2$ norm. The trade-off weight $\lambda$ is chosen from $\{0.1, 0.5, 1, 2\}$ . + +Privacy-Net [55]. This method updates the identity model and the de-identification network adversarially, enhanc + +ing de-identification performance. The original PrivacyNet focuses solely on segmentation tasks, supervised by segmentation masks. To enable task-agnostic de-identification, we train it using the same objective as ID-Adv. Since the identity model is adversarially updated and are stronger, we use smaller $\lambda$ values compared to ID-Adv, i.e., $\{0.05, 0.25, 0.5, 1\}$ . + +MAE [37]. Following [108], we transfer the concept of masked auto-encoders (MAE) [37] to the adjustable privacy protection problem, by masking a random proportion of patches to obscure identity information. It adopts the same diffusion model as our approach to generate the masked regions. This model can also serve as a degenerated version of our model, where both semantic compensation and identity-region similarity designs are removed. + +# 4.4. Downstream Task Models + +For the identity recognition, we adopt the VisionMamba-based ReID model, which differs from the ViT-based model employed during the training of privacy protection methods, ensuring the method generalization capability across different ReID models. For the $X$ -ray classification, we use the ViT model pre-trained with Med-UniC [104]. For $X$ -ray captioning, we employ the visual-language model CXRLLaVA-v2 [58], which is specifically designed for X-ray images. For $X$ -ray segmentation, we adopt CGRSeg [75]. For fundus classification, we use the ViT model pre-trained with KeepFit [111]. Finally, for fundus segmentation, given the limited dataset size, we employ nnUNet [49]. + +# 4.5. Evaluation Metrics + +For privacy evaluation, we adopt the cumulative matching characteristics (CMC) [10] at Rank-1, i.e., CMC-R1, on datasets with patient ID information available (i.e., MIMIC-X, Chest-Xray, CheXpert, EyePacs, and ODIR5K). For datasets without patient ID information (i.e., CheX-det and REFuge2), we adopt the recognition rate, i.e., ID-R, which determines whether the distance between the ID feature of the original and de-identified image exceeds a predefined threshold. The thresholds are set to 1.1 and 1.35 for the X-ray and fundus modalities, respectively, based on statistics from the validation sets of MIMIC-X and EyePACS. For the disease diagnosis task, we employ the area under the receiver operating characteristic curve (AUROC) metric [11]; for the image captioning task, we use the bilingual evaluation understudy (BLEU) metric [79]; and for the image segmentation task, we adopt the Dice score metric [7]. + +# 4.6. Results + +X-ray Classification. As shown in Table 1, our method substantially outperforms other approaches, achieving AU-ROCs of $72.86\%$ , $73.66\%$ , and $86.12\%$ on MIMIC-X, Chest-XRay, and Chexpert, respectively, under CMC + +$\mathrm{R}1 = 10\%$ Notably, although our model is trained on MIMIC-X, it generalizes well to the other two datasets. + +Among the other compared approaches, Feat-Noise obtains the second-best performance, i.e., AUROC of $65.84\%$ at CMC-R1=10% on MIMIX-X, by condensing image pixels into a compact latent feature space. In contrast, methods that jointly optimize a trade-off between de-identification and medical preservation, i.e., ID-Adv and PrivacyNet, yield unsatisfactory performances. As shown in Figure 3 (a), under the ID-R=5% setting, ID-Adv and PrivacyNet attain AUROCs of only $56.32\%$ and $54.21\%$ on MIMIC-X, respectively, which are much lower than the simple pixel blurring baseline $(61.62\%)$ . This indicates that directly optimizing the two conflicting objectives is suboptimal. In contrast, our approach decouples the objectives into two separate steps, identity removal and medical semantic compensation, achieving consistently superior performance. + +As for MAE, which employs the same diffusion model as ours, it achieves competitive results at a high attacking rate, with an AUROC of $76.12\%$ @CMC-R1=40% on MIMIC-X, outperforming all other approaches except ours. However, at a low attacking rate CMC-R1=10%, it falls behind our method by over $13\%$ AUROC. This highlights that our superior performance is not solely due to the generative power of the diffusion model, but rather stems from the effectiveness of our core idea of semantic compensation. + +X-ray Caption. As shown in Table 1, our method attains a BLEU score of 0.1750, remarkably surpassing Pixel-Blur (0.1259), Feat-Noise (0.1453), and MAE (0.1186), at CMC-R1=10%. This proves that our approach can comprehensively preserve the clinic-required information, beyond only the classification label. + +X-ray Segmentation. Furthermore, we evaluate the methods on a fine-grained task: segmentation. As shown in Table 1, at ID-R1=10%, our method achieves a Dice score of 27.98%, outperforming Pixel-Blur (7.53%), Feat-Noise (10.74%), Privacy-Net (15.75%), ID-Adv (22.23%), and MAE(9.94%). This proves that our semantic compensation scheme not only preserves the global semantics for classification, but also effectively retains the local semantics for segmentation. Pixel-Blur and Feat-Noise perform poorly, since they severely corrupts the image details. In contrast, ID-Adv and Privacy-Net, which incorporate a medical feature-matching loss, achieve slightly decent performance, but still lag far behind our approach. For instance, under ID-R=80%, our method outperforms Privacy-Net by approximately 8%, as shown in Figure 3(e). + +Fundus Classification. Beyond X-ray images, our method also proves effective on fundus data. For instance, on EyePACS and ORID5K, our approach outperforms the second-best competitor ID-Adv by about $9\%$ and $1\%$ , respectively. These results confirm that our approach generalizes well across different imaging modalities. + +![](images/84ebb715e6aff1f92f6bcac6c42d18e0684e09850eea652fb7261aecfc4e5704.jpg) +Figure 4. (Left) Ablation on the framework design. (Right) Ablation study on the feature decoupling strategy. + +![](images/c067b500d0ecda17b4ae16ba8843054cac61225c4f617c31305272b1439c21d3.jpg) + +![](images/484413447dbe3f3b1e24e657e8d0c093c97a2de92a3ab18c7a0221e23eca47ab.jpg) +Original + +![](images/fc0f1f5271607d965c29560134977575fbda663fdd274c3aec9228ad50406ce8.jpg) +Ours + +![](images/d44fe9c5aaac6d1b055fc4c601b4c980c5eb05dd55be3825a532e23efd55cb31.jpg) +w/oSem +Figure 5. Qualitative comparison of different variant models. The models are described in Figure 4 caption. Red arrow denotes the modified identity-related features. Best to view by zooming-in. + +![](images/240574cd23d8e87fbad067532886bd56f2b2361796c0e7b8a2dd01987d35b347.jpg) +w/oDecouple + +Fundus Segmentation. Our method achieves a Dice score of $54.66\%$ on REFUGE2 at ID-R1=10%, largely surpassing MAE (27.33%), Privacy-Net (39.23%), and ID-Adv (41.23%). This further validates that our method also effectively preserves fine-grained semantic cues of eye fundus. + +# 4.7. Model Analysis + +Framework-Level Ablation Study. As shown in Figure 4 (Left), by removing the semantic branch, the AUROC of the resulted model 'w/oSem' dramatically drops by over $8\%$ , at CMC-R1=5%. On the other hand, without the identity-semantics decoupling mechanism, the resulting model 'w/oDecouple' leads to about $15\%$ CMC-R1 increase, for achieving the similar AUROC performance, since substantial identity cues are leaked from the vanilla medical features of MFMs. We further illustrate the protected images from different models. As shown in Figure 5, our results effectively modify identity-related features, such as the shape and location of the clavicle and chest contour. The 'w/oSem' model also removes these regions but significantly alters medical manifestations. In contrast, the 'w/oDecouple' model preserves medical features but fails to sufficiently suppress identity-related features, such as clavicle shape, due to residual identity information in the features from MFMs. These results confirm that both medical semantics and identity-semantics decoupling are essential for our advanced medical DeID approach. + +Besides, we attempt to remove ID-Blocking module, tuning privacy via activating more decoupled ID codes. This leads to degraded AUROC (72.01%, 73.47%, 74.68%), compared to using ID-Blocking (72.68%, 74.35%, 79.62%) on MIMIC-X (10%, 20%, 40% at tack SR settings), due to + +![](images/df8bbd1cbc0022cd31f7ab2795244b47bfe48cc3b7d54d867ef2c40f606bbc97.jpg) +Figure 6. Comparison of semantic and identity information in terms of Bits-per-Pixel (bpp) [65], calculated as the feature code-length divided by the original image size. + +the complete loss of pixel-wise details. + +Ablation Study on the Decoupling Strategy. As shown in Figure 4 (Right), omitting the codelength loss terms ('w/oCodeLengthLoss') fails to effectively remove identity information from MFM features, since the loose space cannot effectively decouple the identity and the semantics information. Moreover, removing the discrete code bottleneck ('w/oCodeSpace') further exacerbates identity leakage, leading to further inferior performance. + +Furthermore, we quantitatively compare the overall and identity-related information in MFM features, as shown in Figure 6. First, we notice that a significant portion is identity-related, i.e., around $44\%$ and $55\%$ for X-ray and fundus images. Second, the average information amount of the X-ray dataset Chest-Xray is 0.23bpp, much higher than 0.11bpp achieved by the fundus dataset EyePACS. This is aligned with the medical knowledge prior, that X-rays capture multiple organs and tissues, containing much complex information, than the fundus image that only focuses on eyes. This proves that the learned codelength effectively describes the medical data characteristics. + +Finally, we analyze the impact of the codelength loss weight $\beta$ and the latent code channel number. As shown in Figure 7 (Left), reducing $\beta$ from 0.5 to 0.1 significantly increases CMC-R1 from $5.85\%$ to $12.34\%$ , as a loosely constrained code space fails to effectively decouple identity information. Conversely, increasing $\beta$ from 0.5 to 2 has little effect on CMC-R1 but reduces AUROC performance by approximately $6\%$ , as an overly strong constraint impairs semantic feature reconstruction. The number of code channels also influences performance, by tuning the information capacity of the latent code, as shown in Figure 7 (Right). However, since $\beta$ directly regulates the code-length term, the impact of the channel number is limited. + +Discussion with Label-Conditioned Diffusion Models. These methods [48, 107, 120] employ task-specific labels (e.g., disease labels or text reports) to synthesize images, which are limited to label-associated tasks. In contrast, our approach is task-agnostic and applicable to diverse tasks. Moreover, after fine-tuning our approach towards a single task, i.e., replacing the MFM with a supervised classification network, our method achieves $81.92\%$ AUROC at CMC-R1 = 0.30%, surpassing the label-conditioned model, + +![](images/98358b9df95a800e9e382297e8f44ecd2788cd1858fac5dbc638c897a2b07695.jpg) +Figure 7. (Left) Impact of the rate-distortion weight $\beta$ . (Right:) Impact of the code dimension. All experiments are evaluated by masking $95\%$ identity-related regions, for a fair comparison. + +![](images/96fc7bb0e8cda64f277c0f295fa73ce25d3e77aa0311592b67d5d53a02547bcb.jpg) + +i.e., $80.79\%$ AUROC at CMC-R1 = 0.29%. This confirms that our minimum-codelength representation also benefits the single-task setting, compared to the methods directly using the task labels guiding the diffusion procedure. + +Model Complexity. All methods and our model comprise about 380M parameters, for a fair comparison. Our inference time is 540 ms on an NVIDIA 4090 GPU, which is similar to MAE (526ms), but slower than Privacy-Net (120 ms), ID-Adv (122 ms), and Feat-Noise (124 ms), due to the multiple inference steps of diffusion procedure. Nonetheless, given the significant performance gains and that the medical imaging procedure itself is time-consuming, the running time is acceptable and does not hinder clinical workflows. In the future, we will integrate the single-step diffusion technique [113] to accelerate the process. + +# 5. Conclusion, Future Works, and Other + +Conclusion. We have presented DCM-DeID, a divide-and-conquer framework for medical image de-identification. By leveraging pre-trained Medical Foundation Models and a minimum codelength-based feature decoupling strategy, our method effectively remove identity cues, while preserving medical task utility. Extensive evaluations demonstrate the superiority of our approach. Future Works. We will extend our approach to multi-slice images, such as those produced by Magnetic Resonance Imaging (MRI). Broader Impacts. Our DeID technique is designed for medical AI applications, aiding the human. We emphasize that all rigorous clinical decisions must be made by human physicians using the original medical images. Furthermore, it is critical to enforce ethical guidelines, working in synergy with technological approaches to achieve medical privacy protection. + +Acknowledgment. This work was supported by Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No.62225112), the Fundamental Research Funds for the Central Universities, National Key R&D Program of China 2021YFE0206700, Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and STCSM 22DZ2229005. + +# References + +[1] Kadek YE, Aryanto, Andre Broekema, Matthijs Oudkerk, and Peter MA van Ooijen. Implementation of an anonymisation tool for clinical trials using a clinical trial processor integrated with an existing trial patient data information system. European radiology, 22:144-151, 2012. 1 +[2] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 4 +[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4 +[4] Andrew Barron, Jorma Rissanen, and Bin Yu. The minimum description length principle in coding and modeling. IEEE transactions on information theory, 44(6):2743-2760, 1998. 4 +[5] Toby Berger. Rate-distortion theory. Wiley Encyclopedia of Telecommunications, 2003. 4 +[6] Rui Bernardes, Pedro Serranho, and Conceicao Lobo. Digital ocular fundus imaging: a review. Ophthalmologica, 226 (4):161-181, 2011. 1 +[7] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 92-100. Springer, 2019. 6 +[8] Amit Bhati, Neha Gour, Pritee Khanna, and Aparajita Ojha. Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image dataset. Computers in Biology and Medicine, 153:106519, 2023. 6 +[9] Vincent Bindschaedler, Paul Grubbs, David Cash, Thomas Ristenpart, and Vitaly Shmatikov. The tao of inference in privacy-protected databases. Cryptology ePrint Archive, 2017. 1 +[10] Ruud M Bolle, Jonathan H Connell, Sharath Pankanti, Nalini K Ratha, and Andrew W Senior. The relation between the roc curve and the cmc. In Fourth IEEE workshop on automatic identification advanced technologies (AutoID'05), pages 15-20. IEEE, 2005. 6 +[11] Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145-1159, 1997. 6 +[12] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3334–3342, 2021. 1, 2 +[13] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4 + +[14] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2 +[15] Yan Chen and Pouyan Esmaeilzadeh. Generative ai in medical practice: in-depth exploration of privacy and security challenges. Journal of Medical Internet Research, 26:e53008, 2024. 2 +[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 4 +[17] I Glenn Cohen and Michelle M Mello. Big data, big tech, and protecting patient privacy. *Jama*, 322(12):1141-1142, 2019. 1 +[18] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 4 +[19] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Obfuscation of images via differential privacy: From facial images to general images. Peer-to-Peer Networking and Applications, 14:1705-1733, 2021. 1, 2 +[20] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2 +[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2 +[23] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 2 +[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3 +[25] Emma Dugas, Jared, Jorge, and Will Cukierski. Diabetic retinopathy detection. https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. Kaggle. 6 +[26] August DuMont Schütte, Jürgen Hetzel, Sergios Gatidis, Tobias Hepp, Benedikt Dietz, Stefan Bauer, and Patrick Schwab. Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation. NPJ digital medicine, 4(1):141, 2021. 2 + +[27] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2 +[28] Weijie Fan, Yi Yang, Jing Qi, Qichuan Zhang, Cuiwei Liao, Li Wen, Shuang Wang, Guangxian Wang, Yu Xia, Qihua Wu, et al. A deep-learning-based framework for identifying and localizing multiple abnormalities and assessing cardiomegaly in chest x-ray. Nature Communications, 15(1): 1347, 2024. 1 +[29] Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, et al. Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. arXiv preprint arXiv:2202.08994, 2022. 6 +[30] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774-781, 2012. 1, 2 +[31] Alex Gaudio, Asim Smailagic, Christos Faloutsos, Shreshta Mohan, Elvin Johnson, Yuhao Liu, Pedro Costa, and Aurélio Campilho. Deepfixcx: Explainable privacy-preserving image compression for medical image analysis. Wiley interdisciplinary reviews: Data mining and knowledge discovery, 13(4):e1495, 2023. 2 +[32] Satrajit Ghosh, Chris Gorgolewski, et al. pydeface: A tool to remove facial features from mri images, 2012. 1, 2 +[33] Peter D Grünwald. The minimum description length principle. MIT press, 2007. 2, 4 +[34] Xiuye Gu, Weixin Luo, Michael S Ryoo, and Yong Jae Lee. Password-conditioned anonymization and deanonymization with face identity transformers. In European conference on computer vision, pages 727-743. Springer, 2020. 2 +[35] A Gueziec and P Kazanzides. Anatomy-based registration of ct-scan and intraoperative x-ray images for guiding a surgical robot. IEEE Transactions on Medical Imaging, 17(5): 715-728, 1998. 1 +[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5, 6 +[38] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 3 +[39] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 3 +[40] Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017. 2 + +[41] Steven Hill, Zhimin Zhou, Lawrence Saul, and Hovav Shacham. On the (in) effectiveness of mosaicing and blurring as tools for document redaction. Proceedings on Privacy Enhancing Technologies, 2016. 1, 2 +[42] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4 +[43] Sungmin Hong, Razvan Marinescu, Adrian V Dalca, Anna K Bonkhoff, Martin Bretzner, Natalia S Rost, and Polina Golland. 3d-stylegan: A style-based generative adversarial network for generative modeling of three-dimensional medical images. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 24-34. Springer, 2021. 2 +[44] Andrew Hoopes, Jocelyn S Mora, Adrian V Dalca, Bruce Fischl, and Malte Hoffmann. Synthstrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. 1, 2 +[45] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. arXiv preprint arXiv:2501.13630, 2025. 2 +[46] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 875-885, 2025. +[47] Qiang Hu, Houqiang Zhong, Zihan Zheng, Xiaoyun Zhang, Zhengxue Cheng, Li Song, Guangtao Zhai, and Yanfeng Wang. Vrvvc: Variable-rate nerf-based volumetric video compression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3563-3571, 2025. 2 +[48] Peng Huang, Xue Gao, Lihong Huang, Jing Jiao, Xiaokang Li, Yuanyuan Wang, and Yi Guo. Chest-diffusion: a lightweight text-to-image model for report-to-cxr generation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 8 +[49] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6 +[50] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4 +[51] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 6 +[52] Georgios A Kaissis, Marcus R Makowski, Daniel Rückert, and Rickmer F Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6):305-311, 2020. 1 + +[53] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[54] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[55] Bach Ngoc Kim, Jose Dolz, Pierre-Marc Jodoin, and Christian Desrosiers. Privacy-net: an adversarial approach for identity-obfuscated segmentation of medical images. IEEE Transactions on Medical Imaging, 40(7):1737-1749, 2021. 2, 5, 6 +[56] Cody A Koch and Wayne F Larrabee. Patient privacy, photographs, and publication. JAMA facial plastic surgery, 15 (5):335-336, 2013. 1 +[57] Abhinav Kumar, Sanjay Kumar Singh, K Lakshmanan, Sonal Saxena, and Sameer Shrivastava. A novel cloud-assisted secure deep feature classification framework for cancer histopathology images. ACM Transactions on Internet Technology (TOIT), 21(2):1-22, 2021. 1, 2 +[58] Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim, and Soon Ho Yoon. Cxr-llava: a multimodal large language model for interpreting chest x-ray images. European Radiology, pages 1-13, 2025. 6 +[59] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8):6833-6846, 2023. 2 +[60] Chunyi Li, Xiaozhe Li, Zicheng Zhang, Yuan Tian, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Jia Wang, Haodong Duan, Kai Chen, et al. Information density principle for mllm benchmarks. arXiv preprint arXiv:2503.10079, 2025. 2 +[61] Chunyi Li, Yuan Tian, Xiaoyue Ling, Zicheng Zhang, Haodong Duan, Haoning Wu, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Guo Lu, et al. Image quality assessment: From human to machine preference. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7570-7581, 2025. 2 +[62] Chunyi Li, Jiaohao Xiao, Jianbo Zhang, Farong Wen, Zicheng Zhang, Yuan Tian, Xiangyang Zhu, Xiaohong Liu, Zhengxue Cheng, Weisi Lin, et al. Perceptual quality assessment for embodied ai. arXiv preprint arXiv:2505.16815, 2025. 2 +[63] Yuheng Li, Krishna Kumar Singh, Utkarsh Ojha, and Yong Jae Lee. Mixmatch: Multifactor disentanglement and encoding for conditional image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8039-8048, 2020. 2 +[64] Jie Lian, Jingyu Liu, Shu Zhang, Kai Gao, Xiaqing Liu, Dingwen Zhang, and Yizhou Yu. A structure-aware relation network for thoracic diseases detection and segmentation. IEEE Transactions on Medical Imaging, 40(8):2042-2052, 2021. 6 + +[65] Weisi Lin and Li Dong. Adaptive downsampling to improve image compression at low bit rates. IEEE Transactions on Image Processing, 15(9):2513-2521, 2006. 8 +[66] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6 +[67] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[68] Charles Marsh. Introduction to continuous entropy. Department of Computer Science, Princeton University, 1034, 2013. 4 +[69] Maxim Maximov, Ismail Elezi, and Laura Leal-Taixe. Ciagan: Conditional identity anonymization generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5447-5456, 2020. 2 +[70] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4394-4402, 2018. 4 +[71] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4 +[72] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 4 +[73] Eriksson Monteiro, Carlos Costa, and José Luís Oliveira. A de-identification pipeline for ultrasound medical images in dicom format. Journal of Medical Systems, 41(5):89, 2017. 1 +[74] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 2 +[75] Zhenliang Ni, Xinghao Chen, Yingjie Zhai, Yehui Tang, and Yunhe Wang. Context-guided spatial feature reconstruction for efficient semantic segmentation. In European Conference on Computer Vision, pages 239-255. Springer, 2024. 6 +[76] Kai Packhäuser, Sebastian Gündel, Nicolas Münster, Christopher Syben, Vincent Christlein, and Andreas Maier. Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest x-ray data. Scientific Reports, 12(1):14851, 2022. 1 +[77] Kai Packhäuser, Sebastian Gündel, Florian Thamm, Felix Denzinger, and Andreas Maier. Deep learning-based anonymization of chest radiographs: a utility-preserving measure for patient privacy. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 262-272. Springer, 2023. 2, 5, 6 +[78] Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In International conference on the theory and applications of cryptographic techniques, pages 223-238. Springer, 1999. 2 + +[79] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 +[80] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 6 +[81] W. Nicholson Price and I. Glenn Cohen. Privacy in the age of medical big data. Nature Medicine, 25(1):37-43, 2019. 1 +[82] Moo-Ryong Ra, Ramesh Govindan, and Antonio Ortega. P3: Toward {Privacy-Preserving} photo sharing. In 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13), pages 515-528, 2013. 2 +[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[84] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663):3, 2009. 4 +[85] David Rodríguez González, Trevor Carpenter, Jano I van Hemert, and Joanna Wardlaw. An open source toolkit for medical imaging de-identification. European radiology, 20: 1896-1904, 2010. 1 +[86] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 5 +[87] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1 +[88] Huajie Shao, Yifei Yang, Haohong Lin, Longzhong Lin, Yizhuo Chen, Qinmin Yang, and Han Zhao. Rethinking controllable variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19250-19259, 2022. 2 +[89] Julie K. Taitsman, Christi Macrina Grimm, and Shantanu Agrawal. Protecting patient privacy and data security. New England Journal of Medicine, 368(11):977-979, 2013. 1 +[90] Adam Tanner. Our bodies, our data: how companies make billions selling our medical records. Beacon Press, 2017. 1 +[91] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 2 +[92] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2 + +[93] Yuan Tian, Yichao Yan, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Ean: event adaptive network for enhanced action recognition. International Journal of Computer Vision, 130(10):2453-2471, 2022. 2 +[94] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Non-semantics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 2 +[95] Yuan Tian, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. Clsa: a contrastive learning framework with selective aggregation for video rescaling. IEEE Transactions on Image Processing, 32:1300-1314, 2023. 2 +[96] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[97] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-vsc: Free semantics from visual foundation models for unsupervised video semantic compression. In European Conference on Computer Vision, pages 163-183. Springer, 2024. +[98] Yuan Tian, Guo Lu, and Guangtao Zhai. Smc++: Masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2 +[99] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 1 +[100] Yuan Tian, Shuo Wang, and Guangtao Zhai. Medical manifestation-aware de-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 26363–26372, 2025. 2 +[101] Matt Tierney, Ian Spiro, Christoph Bregler, and Lakshminarayanan Subramanian. Cryptogram: Photo privacy for online social media. In Proceedings of the first ACM conference on Online social networks, pages 75-88, 2013. 2 +[102] Gary Kin-wai Tsui and Tao Chan. Automatic selective removal of embedded patient information from image content of dicom files. American Journal of Roentgenology, 198(4): 769-772, 2012. 1 +[103] Nishant Vishwamitra, Bart Knijnenburg, Hongxin Hu, Yifang P Kelly Caine, et al. Blur vs. block: Investigating the effectiveness of privacy-enhancing obfuscation for images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 39-47, 2017. 1, 2, 5, 6 +[104] Zhongwei Wan, Che Liu, Mi Zhang, Jie Fu, Benyou Wang, Sibo Cheng, Lei Ma, César Quilodrán-Casas, and Rossella Arcucci. Med-unic: Unifying cross-lingual medical vision-language pre-training by diminishing bias. Advances in Neural Information Processing Systems, 36, 2024. 6 +[105] Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu. Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems, 35:33536-33549, 2022. 3, 5 + +[106] Guangyu Wang, Xiaohong Liu, Jun Shen, Chengdi Wang, Zhihuan Li, Linsen Ye, Xingwang Wu, Ting Chen, Kai Wang, Xuan Zhang, et al. A deep-learning pipeline for the diagnosis and discrimination of viral, non-viral and COVID-19 pneumonia from chest x-ray images. Nature biomedical engineering, 5(6):509–521, 2021. 1 +[107] Jinzhuo Wang, Kai Wang, Yunfang Yu, Yuxing Lu, Wenchao Xiao, Zhuo Sun, Fei Liu, Zixing Zou, Yuanxu Gao, Lei Yang, et al. Self-improving generative foundation model for synthetic medical image generation and clinical applications. Nature Medicine, pages 1-9, 2024. 8 +[108] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Jiankang Deng, Xinchao Wang, Hakan Bilen, and Yang You. Facemae: Privacy-preserving face recognition via masked autoencoders. arXiv preprint arXiv:2205.11090, 2022. 1, 6 +[109] Shudong Wang, Zhiyuan Zhao, Yawu Zhao, Luqi Wang, Yuanyuan Zhang, Jiehuan Wang, Sibo Qiao, and Zhihan Lyu. A semantic conditional diffusion model for enhanced personal privacy preservation in medical images. IEEE Journal of Biomedical and Health Informatics, 2024. 2 +[110] Weiru Wang, Chi-Man Vong, Yilong Yang, and Pak-Kin Wong. Encrypted image classification based on multilayer extreme learning machine. Multidimensional Systems and Signal Processing, 28(3):851-865, 2017. 2 +[111] Ruiqi Wu, Chenran Zhang, Jianle Zhang, Yi Zhou, Tao Zhou, and Huazhu Fu. Mm-retinal: Knowledge-enhanced foundational pretraining with fundus image-text expertise. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 722-732. Springer, 2024. 6 +[112] Hanyu Xue, Bo Liu, Ming Ding, Tianqing Zhu, Dayong Ye, Li Song, and Wanlei Zhou. Dp-image: Differential privacy for image data in feature space. arXiv preprint arXiv:2103.07073, 2021. 1, 2 +[113] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 8 +[114] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. Privacy-preserving photo sharing based on a secure JPEG. In 2015 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pages 185-190. IEEE, 2015. 2 +[115] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. SecureJPEG scrambling enabling privacy in photo sharing. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 2 +[116] Junhai Zhai, Sufang Zhang, Junfen Chen, and Qiang He. Autoencoder and its various variants. In 2018 IEEE international conference on systems, man, and cybernetics (SMC), pages 415-419. IEEE, 2018. 5, 6 +[117] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2 +[118] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits + +for effectively learning deep face representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10823-10832, 2019. 2 +[119] Hong-Yu Zhou, Xiaoyu Chen, Yinghao Zhang, Ruibang Luo, Liansheng Wang, and Yizhou Yu. Generalized radiograph representation learning via cross-supervision between images and free-text radiology reports. Nature Machine Intelligence, 4(1):32-40, 2022. 1 +[120] Xinrui Zhou, Yuhao Huang, Haoran Dou, Shijing Chen, Ao Chang, Jia Liu, Weiran Long, Jian Zheng, Erjiao Xu, Jie Ren, et al. Ctrl-genaug: Controllable generative augmentation for medical sequence classification. arXiv preprint arXiv:2409.17091, 2024. 8 +[121] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter WoodwardCourt, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 5 +[122] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In Proceedings of the 41st International Conference on Machine Learning, pages 62429-62442. PMLR, 2024. 3 +[123] Yingxuan Zhu, PD Singh, Khan Siddiqui, and Michael Gillam. An automatic system to detect and extract texts in medical images for de-identification. In Medical Imaging 2010: Advanced PACS-based Imaging Informatics and Therapeutic Applications, page 762803. SPIE, 2010. 1 +[124] M Tarek Ibn Ziad, Amr Alanwar, Moustafa Alzantot, and Mani Srivastava. Cryptoimg: Privacy preserving processing over encrypted images. In 2016 IEEE Conference on Communications and Network Security (CNS), pages 570-575. IEEE, 2016. 2 \ No newline at end of file diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/images.zip b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0a7a881b1176c2fb822fbf78d6ba85cd1e8cb82c --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7988761381ae0849d521129cb549640ddb16eed294d6c5c048cf52709a27486a +size 547244 diff --git a/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/layout.json b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..44109c6cf5bda55904cb8c19c455d7a5da8a8095 --- /dev/null +++ b/2025/Semantic versus Identity_ A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification/layout.json @@ -0,0 +1,12702 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 553, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 140 + ], + "type": "text", + "content": "Semantics versus Identity: A Divide-and-Conquer Approach towards Adjustable Medical Image De-Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "spans": [ + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": "Yuan Tian" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Shuo Wang" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Rongzhao Zhang" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Zijian Chen" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Yankai Jiang" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Chunyi Li" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Xiangyang Zhu" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Fang Yan" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Qiang Hu" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " XiaoSong Wang" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Guangtao Zhai" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "inline_equation", + "content": "^{2,1\\boxplus}" + }, + { + "bbox": [ + 123, + 162, + 487, + 203 + ], + "type": "text", + "content": " Shanghai AI Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 204, + 474, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 204, + 474, + 217 + ], + "spans": [ + { + "bbox": [ + 135, + 204, + 474, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 204, + 474, + 217 + ], + "type": "text", + "content": "Institute of Image Communication and Network Engineering, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 218, + 436, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 218, + 436, + 232 + ], + "spans": [ + { + "bbox": [ + 172, + 218, + 436, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 172, + 218, + 436, + 232 + ], + "type": "text", + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 283, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 296, + 523 + ], + "type": "text", + "content": "Medical imaging has significantly advanced computer-aided diagnosis, yet its re-identification (ReID) risks raise critical privacy concerns, calling for de-identification (DeID) techniques. Unfortunately, existing DeID methods neither particularly preserve medical semantics, nor are flexibly adjustable towards different privacy levels. To address these issues, we propose a divide-and-conquer framework comprising two steps: (1) Identity-Blocking, which blocks varying proportions of identity-related regions, to achieve different privacy levels; and (2) Medical-Semantics-Compensation, which leverages pretrained Medical Foundation Models (MFMs) to extract medical semantic features to compensate the blocked regions. Moreover, recognizing that features from MFMs may still contain residual identity information, we introduce a Minimum Description Length principle-based feature decoupling strategy, to effectively decouple and discard such identity components. Extensive evaluations against existing approaches across seven datasets and three downstream tasks, demonstrates our state-of-the-art performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 635 + ], + "type": "text", + "content": "In the era of digital medicine, large-scale medical images, such as X-rays and fundus photographs [6], are routinely processed by AI-based diagnostic models [28, 87, 106, 119, 121] to aid clinical decision-making. However, the increasing availability of these images raises significant concerns regarding patient privacy [17, 56, 81, 89], calling for the research on medical image de-identification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Although explicit personal details such as patient name can be easily removed from medical image headers [1, 73, 85] or burned-in texts [102, 123], re-identification (ReID) remains feasible for the intrinsic bio-identifiers, such as anatomical markers visible in chest X-rays [35, 76, 99]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 552, + 434 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 552, + 434 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 552, + 434 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 552, + 434 + ], + "type": "image", + "image_path": "a90263befa2c62fee1b9677538860a46d6829df0ee37ae5e085e3d9e7645b20b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 517 + ], + "type": "text", + "content": "Figure 1. (a) Given the query medical image, the ReID model can retrieve sensitive patient information from a leaked database. (b) Our DeID framework, removing identity and then compensating medical semantics, ensures adjustable identity protection, while preserving downstream task utility. Besides, a Minimum Description Length (MDL) principle-based code space is introduced, to decouple and discard the identity information in medical features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "This enables sensitive information breaches [9, 52, 90], compromising patient privacy (see Figure 1(a))." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Several studies have attempted to defend against ReID attacks. For instance, some approaches [12, 30, 32, 44] focus on removing facial features to obfuscate identity. However, such methods cannot be applied to other body parts like the chest, where identity information is deeply interwoven with diagnostic semantics. Standard image filtering techniques, such as blurring [103], pixelation [41], and masking [108], indiscriminately degrade critical diagnostic details, thereby impairing downstream medical task performance. Moreover, under high privacy settings, the severe degradation of image quality further deteriorates task performance. Differential privacy methods [19, 27, 57, 112] mitigate identity information via noise injection, but this operation also perturbs diagnostic features. Identity adver" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 153, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 153, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 153, + 712 + ], + "type": "text", + "content": "Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20613" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "sarial learning methods [55, 77] train generators by jointly maximizing identity discrepancy between the generated and original images, while minimizing distortion of medical information. Nevertheless, given the inherent entanglement between identity and diagnostic features, these methods fail to preserve diagnostic semantics at high privacy levels adequately. Recently, diagnostic annotation-conditioned generative models [15, 26, 43, 91, 109] have yielded promising results, yet they remain limited to task-specific semantics and cannot offer adjustable privacy levels. In summary, no existing method preserves task-generalizable semantics, while supporting a wide range of adjustable privacy levels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 217, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 467 + ], + "type": "text", + "content": "To address these issues, we introduce a novel divide-and-conquer framework DCM-DeID, which decouples identity removal from semantic preservation, to achieve semantic-rich yet adjustable de-identification. Our approach includes three steps, i.e., " + }, + { + "bbox": [ + 56, + 217, + 296, + 467 + ], + "type": "inline_equation", + "content": "ID" + }, + { + "bbox": [ + 56, + 217, + 296, + 467 + ], + "type": "text", + "content": "-Blocking, which masks identity-related regions to achieve adjustable privacy levels; Medical Semantics Extraction, which leverages pre-trained medical foundation models (MFMs) [74, 117] to extract semantic-rich medical features; Image Re-Synthesis, which employs a diffusion model [42, 86] to synthesize de-identified images, given the above ID-masked image and the medical features. Moreover, considering that the features from MFMs may also contain some identity information, we introduce a novel minimum description length [33]-based feature decoupling strategy, which excludes identity-associated information from the vanilla MFM features in a minimum-codelength latent space. This effectively prevents the reintroduction of identity information during the image resynthesis step. Through the above designs, our approach achieves a better trade-off between DeID and diagnostic utilization than prior approaches. Our contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 467, + 295, + 647 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 467, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 467, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 67, + 467, + 295, + 527 + ], + "type": "text", + "content": "- We reveal that existing medical DeID methods fall short in preserving task-generalizable semantics, and do not adjust seamlessly across privacy levels. We build the first benchmark for this problem, by reproducing previous approaches fairly on seven datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 527, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 527, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 67, + 527, + 295, + 574 + ], + "type": "text", + "content": "- We propose the DCM-DeID framework, which performs identity removal and medical semantics preservation in separate steps, enabling both adjustable privacy protection and medical task utility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 574, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 574, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 67, + 574, + 295, + 622 + ], + "type": "text", + "content": "- We introduce a Minimum Description Length-based decoupling strategy, which decouples identity cues from medical features in a compact code space, further improving the privacy protection capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 623, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 623, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 67, + 623, + 295, + 647 + ], + "type": "text", + "content": "- Our framework demonstrates state-of-the-art performance. Extensive Analysis is also performed." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 148, + 670 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Image Privacy Protection. Early methods applied low-level filters to obscure image details, including downsampling [20], blurring [103], and pixelation [41]. Later, en" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "cryption in alternate domains such as JPEG bitstreams [82, 101] and DCT coefficients [114, 115] was explored, though these often introduced severe distortions that hindered downstream tasks. Homomorphic encryption [110, 124] addresses inference on encrypted images, but suffers from high computational cost [78] and limited compatibility with advanced models like Vision Transformers [24, 93, 95]. Additionally, approaches for face images [12, 34, 69] leverage facial priors from StyleGAN [53] or face recognition networks [22, 118], which may not readily generalize to other domains, such as the medical-domain images in our work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 204, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 382 + ], + "type": "text", + "content": "Medical Image De-Identification. Early methods (e.g., FreeSurfer [30], PyDeface [32], SynthStrip [44]) focus on removing facial features in brain MRI. For common medical images, early approaches use pixel-domain filters (like blurring [103] and pixelation [41]) or frequency-domain techniques [31], but these hand-crafted solutions also severely degrade the image details, leading to substantially degraded performance and visual quality [59, 60, 62]. Differential Privacy methods [19, 27, 57, 112] inject noise into the training data, which compromises inference-time utility. More recent generative models [15, 26, 43, 91, 109] synthesize images conditioned on disease labels or lesion masks. However, they tend to lack task generalizability and struggle to balance privacy-utility trade-offs, which are addressed by our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 538 + ], + "type": "text", + "content": "Feature Decoupling. Variational auto-encoder (VAE)-based works [40, 88] decouple representations, by constraining the variables in latent space independent. Generative adversarial network (GAN)-based methods [14, 63] are unsupervised, leaving factors unaligned with explicit semantic or identity information. For face images, there are methods [23, 53, 54, 100] targeting identity separation. However, these methods rely on strong facial priors that may not generalize to medical images. In contrast, motivated by the explicit entropy regularization terms adopted in data compression field [2, 45-47, 61, 92, 94, 96-98], our approach effectively decouples identity in medical images, within a minimum-codelength space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 397, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 397, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 397, + 556 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 653 + ], + "type": "text", + "content": "In this section, we first describe the medical re-identification (ReID) models used for privacy attacks. Next, we introduce our de-identification model, which divides the task into two stages. First, identity information is removed via region blocking with an adjustable threshold. Second, lost medical semantics are compensated. This approach flexibly adjusts privacy while preserving rich, generalizable medical features for downstream tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 660, + 440, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 440, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 440, + 672 + ], + "type": "text", + "content": "3.1. Medical ReID Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Given a query medical image, ReID models aim to retrieve all images belonging to the same individual, from a medical record database. Concretely, the model first extracts identity" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20614" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 59, + 526, + 300 + ], + "blocks": [ + { + "bbox": [ + 80, + 59, + 526, + 300 + ], + "lines": [ + { + "bbox": [ + 80, + 59, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 80, + 59, + 526, + 300 + ], + "type": "image", + "image_path": "ac8e1bfc66603ba13b1bf15a2825c68727523f679f897b3770ce1205b4272735.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed divide-and-conquer framework, DCM-DeID. (a) " + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "inline_equation", + "content": "ID" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": "-Blocking: A pre-trained ReID network produces the identity-similarity map, which is binarized by different thresholds to adjust privacy level. (b) " + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "inline_equation", + "content": "ID" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": "-Free Medical Semantics Extraction: Medical foundation models (MFMs) extract features that are encoded into a code space under the minimum-codelength regularization. A learned mask partitions the codes into identity- and medical semantics-related ones, where only the latter one is preserved. (c) Image Re-Synthesis: A diffusion model re-synthesizes images that are privacy-preserving and semantics-rich, generalizing to various downstream tasks. We illustrate with X-ray images, but the framework is also applicable to other modalities such as fundus images. " + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": " denotes frozen models, while gray dashed lines indicate components used solely for learning identity-semantic decoupling. The channel number of codes " + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": " is arbitrary; two channels are shown for conciseness. " + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 54, + 305, + 555, + 395 + ], + "type": "text", + "content": " denotes the element-wise multiplication." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 400, + 294, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 294, + 509 + ], + "type": "text", + "content": "(ID) embedding from the query image, and then compare it with that of each image within the database. Then, the image with the closest Euclidean distance is adopted as the re-identified image. We build two medical ReID models, i.e., ViT [24] and VisionMamba [122]-based ones, which are separately adopted in the training and the evaluation stages. These ReID models are optimized with a combination of classification loss and triplet loss [39], following the previous object ReID work [38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 234, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 234, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 234, + 529 + ], + "type": "text", + "content": "3.2. A Divide-and-Conquer Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 533, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 629 + ], + "type": "text", + "content": "To defend against attacks on medical ReID models, we propose DCM-DeID, a divide-and-conquer approach for medical image de-identification. DCM-DeID operates in three stages: ID Blocking, which removes identity-related image regions; ID-Free Medical Semantics Extraction, which extracts rich medical information without reintroducing identity information; and Image Re-Synthesis, which generates the final de-identified medical image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "ID-Blocking. Given an input image " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " denote the image spatial scales, we use a ViT-based ReID model to extract local features " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{768 \\times h \\times w}" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "h = H / 16" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "w = W / 16" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ". Spatial average pooling is applied to " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " to obtain an identity embedding " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "id \\in \\mathbb{R}^{768}" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ". For each spatial position in " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": ", the cosine similarity with " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "id" + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": " is computed, resulting in a similarity map " + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "inline_equation", + "content": "S[i] =" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\cos(f[i],id)" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": " denotes the spatial location. Then, " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": " is binarized by a threshold " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": ". Finally, the ID-blocked image is computed as: " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "X^{noID} = X \\odot \\mathrm{Upsample}(S > T)" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": ", where Upsample denotes nearest-neighbor interpolation to match the resolution of " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 400, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": "ID-Free Medical Semantics Extraction. Although " + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "inline_equation", + "content": "X^{noID}" + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": " effectively removes identity information, it inevitably distorts medical cues such as lung shadows. To amend this, we employ pre-trained medical foundation models (MFMs), e.g., MGCA [105] for X-ray images, to extract rich medical feature " + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "inline_equation", + "content": "f_{MFM}" + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "inline_equation", + "content": "f_{MFM}" + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": " contains both semantic cues and local details that may encode identity, we introduce a feature decoupling strategy (Section 3.3) to decouple and remove the identity information, yielding the identity-free semantic feature " + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem}" + }, + { + "bbox": [ + 313, + 460, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "text", + "content": "Image Re-Synthesis. Given " + }, + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "inline_equation", + "content": "X^{noID}" + }, + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem}" + }, + { + "bbox": [ + 313, + 580, + 554, + 664 + ], + "type": "text", + "content": ", a dual-conditioned diffusion model synthesizes the de-identified image that inherits the rich semantics within MFMs, while also protecting privacy. Since the synthesized image is highly realistic, it can be directly deployed to downstream medical AI applications, without further adaptation. The model details are elaborated in Section 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 481, + 685 + ], + "type": "text", + "content": "3.3. Medical Semantics Decoupling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Medical features extracted by the MFM encode both diagnostic semantics (e.g., lesion morphology) and identity" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20615" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "related cues (e.g., rib patterns in chest X-rays). For effective privacy-preserving, it is imperative to decouple these two types of information, and discard the identity cues. We achieve this by learning a minimum-length code space, and separating the two parts in this space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "text", + "content": "Theoretical Motivation. From an information-theoretic perspective, the Minimum Description Length (MDL) principle [4, 33] states that the best representation for a given set of data is the one that minimizes the total codelength needed to describe the data, where each group of features tends to capture the independent or low-correlation information parts. In our context, let " + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "text", + "content": " be the latent representation of the MFM feature " + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{MFM}}" + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "inline_equation", + "content": "H(Q)" + }, + { + "bbox": [ + 55, + 133, + 295, + 264 + ], + "type": "text", + "content": " denote its expected codelength. The MDL principle objective can be seen as balancing a reconstruction loss and a compression term, i.e., the so-called rate-distortion loss (RD loss) [5]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 267, + 295, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 267, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 295, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o d e - a l l}} = \\min _ {\\mathcal {E}, \\mathcal {D}} \\underbrace {\\| f _ {M F M} - \\hat {f} _ {M F M} \\| _ {2}} _ {\\text {F e a t u r e R e c o n s t r u c t i o n}} + \\underbrace {\\beta H (\\boldsymbol {Q})} _ {\\text {C o d e l e n g t h}}, \\tag {1}", + "image_path": "c83c273dd1a707e5274bf71a3cc2df2fb73af316e4a97fe9c5b47a5630e114ec.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "inline_equation", + "content": "Q = \\mathcal{E}(f_{MFM})" + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\hat{f}_{MFM} = \\mathcal{D}(Q)" + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": " denotes balancing weight. " + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 301, + 295, + 337 + ], + "type": "text", + "content": " represent a pair of feature encoder and decoder networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": "Discrete Code-based Codelength Estimation. Directly calculating the " + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "inline_equation", + "content": "H(\\mathbf{Q})" + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": " for the continuous variable " + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": " is nontrivial [68]. Fortunately, the neural data compression community [2, 3, 70, 72] have verified that the codelength of integer latent variables can be quite precisely estimated with a learnable entropy model. Therefore, we append the quantization operation at the tail of the encoder " + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": ", to make elements within " + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 338, + 295, + 434 + ], + "type": "text", + "content": " discrete values, and estimate its codelength." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": "Concretely, " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": " comprises three residual blocks [36] with 256 channels, followed by a convolutional layer to reduce dimensionality and a rounding operation that outputs a 32-channel integer code " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": ". The decoder network " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": " is symmetric to " + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": ", except it omits the rounding operation. During training, the straight-through estimator [71] is employed to backpropagate gradients through the rounding step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": "Following [2], the expected codelength of encoding " + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": " is calculated as the log-likelihood, i.e., " + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "inline_equation", + "content": "H(Q) = -\\log_2p(Q)" + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": ", where the probability " + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "inline_equation", + "content": "p(Q)" + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": " is modeled using a Gaussian Mixture Model (GMM) [84] with " + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 518, + 295, + 566 + ], + "type": "text", + "content": " components:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 567, + 295, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 567, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 102, + 567, + 295, + 599 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {Q}) = \\sum_ {k = 1} ^ {K} w ^ {k} \\cdot \\mathcal {N} \\left(\\boldsymbol {Q}; \\mu^ {k}, e ^ {\\sigma^ {k}}\\right), \\tag {2}", + "image_path": "d0a75b8bdb0eea417fb649ebad0b8a167ccc141d717f716412695df9a31036b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w},\\boldsymbol {\\mu},\\boldsymbol {\\sigma}\\}" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": " are the learnable mixture weights, means, and log variance scalers of the GMM components, respectively, which are shared across spatial positions, not unshared along the channel axis [2]. Following [16], " + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": " is set to three. For each integer element " + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "q\\in" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": " , the probability is computed over the quantization bin [18, 72], " + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "p(q) = \\mathcal{F}(q + 0.5) - \\mathcal{F}(q - 0.5)" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x) = \\sum_{k = 1}^{K}w^{k}\\Phi \\Big(x;\\mu^{k},e^{\\sigma^{k}}\\Big)" + }, + { + "bbox": [ + 55, + 601, + 295, + 713 + ], + "type": "text", + "content": " is the cumulative distribution function (CDF) of the Gaussian Mixture Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 69, + 553, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 69, + 553, + 112 + ], + "spans": [ + { + "bbox": [ + 313, + 69, + 553, + 112 + ], + "type": "text", + "content": "(GMM), " + }, + { + "bbox": [ + 313, + 69, + 553, + 112 + ], + "type": "inline_equation", + "content": "\\Phi \\left(x;\\mu ,e^{\\sigma}\\right) = \\frac{1}{2}\\left[1 + \\mathrm{erf}\\left(\\frac{x - \\mu}{\\sqrt{2}e^{\\sigma}}\\right)\\right]" + }, + { + "bbox": [ + 313, + 69, + 553, + 112 + ], + "type": "text", + "content": ". We not that the CDF can be efficiently calculated by the modern deep learning framework such as PyTorch [80]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": "Learning of Identity-Associated Code Mask. A single convolution layer predicts a binary mask " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": ", with the same dimensions as " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": ". The Gumbel-Softmax algorithm [50] is applied to enable gradient propagation through the binary mask. The identity-associated codes are then obtained by element-wise masking, " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{id}} = Q\\odot \\mathbf{M}" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": ". A lightweight convolutional network, composed of three residual blocks followed by average pooling, predicts the identity embedding " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{id}}" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 113, + 553, + 233 + ], + "type": "text", + "content": ". Then, the RD loss for reconstructing identity can be given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 239, + 553, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 239, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 364, + 239, + 553, + 255 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o d e - i d}} = \\| \\hat {i d} - i d \\| _ {2} + \\beta H \\left(\\boldsymbol {Q} _ {\\text {i d}}\\right), \\tag {3}", + "image_path": "cffd70baa575e8a985397612e65540fc1aded0f4a4c4eaeedc944e6e34f349ad.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "inline_equation", + "content": "H(\\tilde{Q}_{\\mathrm{id}})" + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "text", + "content": " is calculated similarly to " + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "inline_equation", + "content": "H(Q)" + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "text", + "content": ", sharing the same GMM parameters and balancing weight " + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 262, + 553, + 299 + ], + "type": "text", + "content": " as in Equation 1, since they operate in the same latent space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "text", + "content": "Reconstruction of Medical Semantics. By suppressing identity-related codes via the inverse mask " + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "inline_equation", + "content": "(1 - \\mathbf{M})" + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "text", + "content": ", we obtain the semantics-part codes " + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "inline_equation", + "content": "Q_{sem} = (1 - \\mathbf{M})\\otimes Q" + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "text", + "content": ". Finally, the final ID-free medical semantic feature is reconstructed as: " + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem} = \\mathcal{D}(Q_{sem})" + }, + { + "bbox": [ + 313, + 300, + 554, + 372 + ], + "type": "text", + "content": ", which preserves critical diagnostic semantics, excluding the identity information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 381, + 462, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 462, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 462, + 395 + ], + "type": "text", + "content": "3.4. Image Re-Synthesis Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": "Given the ID-masked image " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "X^{noID}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": " and the ID-free medical semantic feature " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": ", we employ a diffusion model to synthesize de-identified medical images. First, we utilize a Feat-Net to project the high-resolution " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "X^{noID}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": " into the low-resolution feature " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "f^{noID} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": ". The Feat-Net consists of the VAE encoder from Stable Diffusion [86], followed by two convolution layers of kernel size 5 and stride size 2. Next, we adopt a bi-directional crossattention mechanism [13] to fuse " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "f^{noID}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": ", producing a fused feature " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "f_{fuse} \\in \\mathbb{R}^{512 \\times \\frac{H}{32} \\times \\frac{W}{32}}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": ", which is further processed through a series of convolutional layers. This produces a set of features with dimensions matching those of the UNet's intermediate feature maps within the diffusion model. These features are added to the UNet layers, guiding the diffusion process toward two objectives: maintaining the privacy level of " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "X^{noID}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": ", while preserving the medical semantics in " + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\hat{f}_{sem}" + }, + { + "bbox": [ + 313, + 399, + 554, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 614, + 422, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 614, + 422, + 627 + ], + "spans": [ + { + "bbox": [ + 314, + 614, + 422, + 627 + ], + "type": "text", + "content": "3.5. Learning Strategy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 553, + 657 + ], + "type": "text", + "content": "The whole framework is end-to-end optimized, with the following objective," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 663, + 553, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 663, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 359, + 663, + 553, + 676 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o d e - a l l}} + \\mathcal {L} _ {\\text {c o d e - i d}} + \\mathcal {L} _ {\\text {d i f f u s e}}, \\tag {4}", + "image_path": "4173ac77344c1a36728d9e63f6733ba4ba007b1e7314af92db19d383e799af64.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 681, + 554, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 681, + 554, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 681, + 554, + 717 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 681, + 554, + 717 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diffuse}}" + }, + { + "bbox": [ + 313, + 681, + 554, + 717 + ], + "type": "text", + "content": " denotes the diffusion loss [42]. We do not introduce the balancing weight, since we found directly adding the loss terms already achieves satisfactory results." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20616" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 59, + 182, + 177 + ], + "blocks": [ + { + "bbox": [ + 61, + 59, + 182, + 177 + ], + "lines": [ + { + "bbox": [ + 61, + 59, + 182, + 177 + ], + "spans": [ + { + "bbox": [ + 61, + 59, + 182, + 177 + ], + "type": "image", + "image_path": "c99259f96619449aa904ee38d86f8de71df13aaea8e70f3bc5f3cc3cdc2adebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 59, + 306, + 177 + ], + "blocks": [ + { + "bbox": [ + 186, + 59, + 306, + 177 + ], + "lines": [ + { + "bbox": [ + 186, + 59, + 306, + 177 + ], + "spans": [ + { + "bbox": [ + 186, + 59, + 306, + 177 + ], + "type": "image", + "image_path": "808fc601ee7308eefabd34039ba22fd34d7aab8b1065800b3e7b11889bd437ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 59, + 430, + 177 + ], + "blocks": [ + { + "bbox": [ + 310, + 59, + 430, + 177 + ], + "lines": [ + { + "bbox": [ + 310, + 59, + 430, + 177 + ], + "spans": [ + { + "bbox": [ + 310, + 59, + 430, + 177 + ], + "type": "image", + "image_path": "3d21caf0f38b5ab4c364b97222113ffbe277f29a2365623c9bbd21927028cd87.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 435, + 57, + 555, + 177 + ], + "blocks": [ + { + "bbox": [ + 435, + 57, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 435, + 57, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 435, + 57, + 555, + 177 + ], + "type": "image", + "image_path": "d665c432f1e60176e8a8d202611a287d670558f16f9793fe46c229356addef9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 61, + 186, + 182, + 304 + ], + "blocks": [ + { + "bbox": [ + 116, + 178, + 126, + 186 + ], + "lines": [ + { + "bbox": [ + 116, + 178, + 126, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 178, + 126, + 186 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 186, + 182, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 186, + 182, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 182, + 304 + ], + "type": "image", + "image_path": "1e557411b02a3bb757f64fcc717e16de5c1cc89d18c05a51fa944b5acff47adc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 306, + 127, + 313 + ], + "lines": [ + { + "bbox": [ + 117, + 306, + 127, + 313 + ], + "spans": [ + { + "bbox": [ + 117, + 306, + 127, + 313 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 186, + 186, + 306, + 304 + ], + "blocks": [ + { + "bbox": [ + 241, + 178, + 251, + 186 + ], + "lines": [ + { + "bbox": [ + 241, + 178, + 251, + 186 + ], + "spans": [ + { + "bbox": [ + 241, + 178, + 251, + 186 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 186, + 306, + 304 + ], + "lines": [ + { + "bbox": [ + 186, + 186, + 306, + 304 + ], + "spans": [ + { + "bbox": [ + 186, + 186, + 306, + 304 + ], + "type": "image", + "image_path": "2faff10e5a4edea8300716b970cf5868f398bbfba667288db02c7e9e2240961b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 306, + 250, + 313 + ], + "lines": [ + { + "bbox": [ + 242, + 306, + 250, + 313 + ], + "spans": [ + { + "bbox": [ + 242, + 306, + 250, + 313 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 186, + 430, + 304 + ], + "blocks": [ + { + "bbox": [ + 365, + 178, + 375, + 186 + ], + "lines": [ + { + "bbox": [ + 365, + 178, + 375, + 186 + ], + "spans": [ + { + "bbox": [ + 365, + 178, + 375, + 186 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 186, + 430, + 304 + ], + "lines": [ + { + "bbox": [ + 310, + 186, + 430, + 304 + ], + "spans": [ + { + "bbox": [ + 310, + 186, + 430, + 304 + ], + "type": "image", + "image_path": "10ac4a2770487f28e742ddaedb9dea3b441d98ed1049d509b8218643496ae457.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 306, + 375, + 314 + ], + "lines": [ + { + "bbox": [ + 365, + 306, + 375, + 314 + ], + "spans": [ + { + "bbox": [ + 365, + 306, + 375, + 314 + ], + "type": "text", + "content": "(g)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 186, + 555, + 304 + ], + "blocks": [ + { + "bbox": [ + 490, + 178, + 501, + 186 + ], + "lines": [ + { + "bbox": [ + 490, + 178, + 501, + 186 + ], + "spans": [ + { + "bbox": [ + 490, + 178, + 501, + 186 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 186, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 435, + 186, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 435, + 186, + 555, + 304 + ], + "type": "image", + "image_path": "e1f20bd9d447e113e2051a015d7447551d6395951025e02778cdc5afd2b6fd45.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 306, + 500, + 313 + ], + "lines": [ + { + "bbox": [ + 490, + 306, + 500, + 313 + ], + "spans": [ + { + "bbox": [ + 490, + 306, + 500, + 313 + ], + "type": "text", + "content": "(h)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 57, + 327, + 552, + 500 + ], + "blocks": [ + { + "bbox": [ + 132, + 316, + 476, + 326 + ], + "lines": [ + { + "bbox": [ + 132, + 316, + 476, + 326 + ], + "spans": [ + { + "bbox": [ + 132, + 316, + 476, + 326 + ], + "type": "text", + "content": "Figure 3. Identity-performance trade-off curves of various medical privacy protection methods." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 327, + 552, + 500 + ], + "lines": [ + { + "bbox": [ + 57, + 327, + 552, + 500 + ], + "spans": [ + { + "bbox": [ + 57, + 327, + 552, + 500 + ], + "type": "table", + "html": "
Attack SRMethodX-ray Classify AUROC (%)X-ray Caption BLEUX-ray Seg Dice (%)Fundus Classify AUROC (%)Fundus Seg Dice (%)
MIMIC-XChest-XRayCheXpertMIMIC-XChestX-DetEyePACSODIR5KRefuge2
10%Pixel-Blur [103]64.1560.4874.450.12597.5344.8356.1722.35
Feat-Noise [116]65.8466.2566.490.145310.7447.7655.5324.85
ID-Adv [77]59.3661.7573.110.113122.2356.2858.5041.23
Privacy-Net [55]57.0959.5171.470.132915.7552.6357.2939.23
MAE [37]59.1165.2677.050.11869.9450.9156.5027.33
Ours72.8673.6686.120.175027.9865.2359.9654.66
20%Pixel-Blur [103]67.2364.2780.670.130711.2549.5657.4234.11
Feat-Noise [116]68.2769.0180.780.159514.2851.2957.1737.20
ID-Adv [77]64.0862.7177.530.119322.8363.2459.2849.47
Privacy-Net [55]62.1162.6380.250.143420.5561.0158.2547.47
MAE [37]66.4371.3883.520.138313.9954.9758.0140.13
Ours74.3575.0186.320.185929.4969.5960.4162.04
Original82.1384.8287.240.321852.8981.4661.5390.08
", + "image_path": "5d8b344a8601089caa65eafa69109a29a5731b4c224a3ba1bb6fa4c569a36ed6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "type": "text", + "content": "Table 1. Performance comparison of medical image privacy protection methods, under different attack Success-Rates (SR). For measuring SR, we adopt CMC-R1 metric for MIMIC-X, Chest-Xray, CheXpert, EyePACS, and ODIR5K, using ID-R metric for ChestX-Det and Refuge2. Original denotes the performance on original images, which is the performance upper-bound of privacy-removal images." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 188, + 577 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": "For Med-ReID models, we adopt the AdamW optimizer [66] during training, with a learning rate of 1e-5 scheduled by cosine decaying strategy and a weight decay of 1e-2. The training process consists of 300,000 steps. The batch size is 256. We apply random cropping and blurring as image augmentation strategies, and the input image resolution to networks is " + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": ". The ViT-based models are initialized with CLIP-pretrained weights [83], while the VisionMamba-based models are initialized with ImageNet-pretrained weights [21]. Training a single ReID model takes about 24 hours with four NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "For DCM-DeID model, the UNet within the diffusion model follows the same architecture as the Stable Diffusion [86], also performing the diffusion procedure in the latent space. The feature channels within UNet are reduced to [128, 256, 512, 1024], for the four stages of both the down-pathway and up-pathway, to reduce computational cost. The identity-similarity map threshold " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " is defined as the " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "-th quantile of the similarity map " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " is selected from [0.95, 0.7, 0.4, 0.2] to cover wide privacy levels. We adopt MGCA-ResNet [105] and RetFound-ViT [121] MFMs for X-ray and fundus images, respectively. During training, we apply random flipping and random cropping " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " patches for data augmentation. The codelength loss term weight " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " is set to 0.5. At test time, we resize the shorter" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20617" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "side of the images to 256 and then center-crop the middle " + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": " region. The learning rate is set to 1e-4 and is gradually decayed with the cosine annealing strategy [67]. The total number of training steps is 800,000. The mini-batch size is 64. We utilize the AdamW optimizer [66] implemented in PyTorch [80] with CUDA support. The momentum parameters are set as " + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": ", and the gradient norm is clipped to a maximum value of 1. The entire training process takes about three days on a machine equipped with eight NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 118, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 118, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 118, + 213 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 459 + ], + "type": "text", + "content": "We evaluate our approach on two medical image modalities: chest X-rays and eye fundus photographs, with seven public datasets. For the chest X-ray modality, we split the MIMIC-X dataset [51] into training, validation, and test sets using an 8:1:1 ratio. For the Chest-Xray and CheXpert datasets, we randomly select " + }, + { + "bbox": [ + 55, + 218, + 295, + 459 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 218, + 295, + 459 + ], + "type": "text", + "content": " patients as the test set. We also adopt the ChestX-Det dataset [64] to evaluate the X-ray segmentation task. For the eye fundus modality, we divide the EyePACS dataset [25] into training, validation, and test sets with an 8:1:1 ratio, and we use the Refuge2 dataset [29] to evaluate the fundus segmentation task. ODIR5K [8] is also adopted for evaluating the fundus classification task of systemic diseases such as hypertension. Note that only MIMIC-X and EyePACS are used during training; all other datasets, which differ in environment, demographics, and imaging devices, are never seen during training, to evaluate the domain generalizability of our approach. MIMIC-X, CheXpert, Chest-Xray, EyePACS, ChestXDet, and ODIR5K contain 377K, 224K, 112K, 88K, 3.6K, and 5K images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 467, + 269, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 269, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 269, + 480 + ], + "type": "text", + "content": "4.3. Reproduced Privacy Protection Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "type": "text", + "content": "We implement several privacy protection methods, comparing them with our approach in a fair setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "content": "Pixel-Blur [103]. This method applies a Gaussian blur to the input image. We experiment with standard deviations of " + }, + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\{1,5,10,20\\}" + }, + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "content": " to vary the level of de-identification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "text", + "content": "Feat-Noise [116]. We train an autoencoder [116] and inject Gaussian noise into its latent features. The noise level is selected from " + }, + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.8, 0.85, 0.9, 0.95\\}" + }, + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "ID-Adv [77]. A UNet is trained to generate a de-identified image " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " from the original image " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": ", optimizing the loss " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\lambda \\cos(id_X, id_Y) + \\|med_X - med_Y\\|_2 + \\mathcal{L}_{reg}" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "id_X" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "id_Y" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " are identity features extracted by a ViT-based ReID model, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "med_X" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "med_Y" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " are medical features obtained from MFMs same as our approach. " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " is a GAN regularization loss ensuring visual plausibility, " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\cos(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " denotes cosine similarity, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " the " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " norm. The trade-off weight " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": " is chosen from " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.5, 1, 2\\}" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Privacy-Net [55]. This method updates the identity model and the de-identification network adversarially, enhanc" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ing de-identification performance. The original PrivacyNet focuses solely on segmentation tasks, supervised by segmentation masks. To enable task-agnostic de-identification, we train it using the same objective as ID-Adv. Since the identity model is adversarially updated and are stronger, we use smaller " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " values compared to ID-Adv, i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\{0.05, 0.25, 0.5, 1\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 253 + ], + "type": "text", + "content": "MAE [37]. Following [108], we transfer the concept of masked auto-encoders (MAE) [37] to the adjustable privacy protection problem, by masking a random proportion of patches to obscure identity information. It adopts the same diffusion model as our approach to generate the masked regions. This model can also serve as a degenerated version of our model, where both semantic compensation and identity-region similarity designs are removed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 459, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 459, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 459, + 272 + ], + "type": "text", + "content": "4.4. Downstream Task Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "content": "For the identity recognition, we adopt the VisionMamba-based ReID model, which differs from the ViT-based model employed during the training of privacy protection methods, ensuring the method generalization capability across different ReID models. For the " + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "content": "-ray classification, we use the ViT model pre-trained with Med-UniC [104]. For " + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "content": "-ray captioning, we employ the visual-language model CXRLLaVA-v2 [58], which is specifically designed for X-ray images. For " + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 278, + 555, + 422 + ], + "type": "text", + "content": "-ray segmentation, we adopt CGRSeg [75]. For fundus classification, we use the ViT model pre-trained with KeepFit [111]. Finally, for fundus segmentation, given the limited dataset size, we employ nnUNet [49]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 426, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 426, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 426, + 441 + ], + "type": "text", + "content": "4.5. Evaluation Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 448, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 640 + ], + "type": "text", + "content": "For privacy evaluation, we adopt the cumulative matching characteristics (CMC) [10] at Rank-1, i.e., CMC-R1, on datasets with patient ID information available (i.e., MIMIC-X, Chest-Xray, CheXpert, EyePacs, and ODIR5K). For datasets without patient ID information (i.e., CheX-det and REFuge2), we adopt the recognition rate, i.e., ID-R, which determines whether the distance between the ID feature of the original and de-identified image exceeds a predefined threshold. The thresholds are set to 1.1 and 1.35 for the X-ray and fundus modalities, respectively, based on statistics from the validation sets of MIMIC-X and EyePACS. For the disease diagnosis task, we employ the area under the receiver operating characteristic curve (AUROC) metric [11]; for the image captioning task, we use the bilingual evaluation understudy (BLEU) metric [79]; and for the image segmentation task, we adopt the Dice score metric [7]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 372, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 372, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 372, + 659 + ], + "type": "text", + "content": "4.6. Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "X-ray Classification. As shown in Table 1, our method substantially outperforms other approaches, achieving AU-ROCs of " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "72.86\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "73.66\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "86.12\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " on MIMIC-X, Chest-XRay, and Chexpert, respectively, under CMC" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20618" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "\\mathrm{R}1 = 10\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " Notably, although our model is trained on MIMIC-X, it generalizes well to the other two datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "content": "Among the other compared approaches, Feat-Noise obtains the second-best performance, i.e., AUROC of " + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "inline_equation", + "content": "65.84\\%" + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "content": " at CMC-R1=10% on MIMIX-X, by condensing image pixels into a compact latent feature space. In contrast, methods that jointly optimize a trade-off between de-identification and medical preservation, i.e., ID-Adv and PrivacyNet, yield unsatisfactory performances. As shown in Figure 3 (a), under the ID-R=5% setting, ID-Adv and PrivacyNet attain AUROCs of only " + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "inline_equation", + "content": "56.32\\%" + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "inline_equation", + "content": "54.21\\%" + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "content": " on MIMIC-X, respectively, which are much lower than the simple pixel blurring baseline " + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "inline_equation", + "content": "(61.62\\%)" + }, + { + "bbox": [ + 54, + 98, + 295, + 277 + ], + "type": "text", + "content": ". This indicates that directly optimizing the two conflicting objectives is suboptimal. In contrast, our approach decouples the objectives into two separate steps, identity removal and medical semantic compensation, achieving consistently superior performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "content": "As for MAE, which employs the same diffusion model as ours, it achieves competitive results at a high attacking rate, with an AUROC of " + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "inline_equation", + "content": "76.12\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "content": " @CMC-R1=40% on MIMIC-X, outperforming all other approaches except ours. However, at a low attacking rate CMC-R1=10%, it falls behind our method by over " + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "content": " AUROC. This highlights that our superior performance is not solely due to the generative power of the diffusion model, but rather stems from the effectiveness of our core idea of semantic compensation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "type": "text", + "content": "X-ray Caption. As shown in Table 1, our method attains a BLEU score of 0.1750, remarkably surpassing Pixel-Blur (0.1259), Feat-Noise (0.1453), and MAE (0.1186), at CMC-R1=10%. This proves that our approach can comprehensively preserve the clinic-required information, beyond only the classification label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "type": "text", + "content": "X-ray Segmentation. Furthermore, we evaluate the methods on a fine-grained task: segmentation. As shown in Table 1, at ID-R1=10%, our method achieves a Dice score of 27.98%, outperforming Pixel-Blur (7.53%), Feat-Noise (10.74%), Privacy-Net (15.75%), ID-Adv (22.23%), and MAE(9.94%). This proves that our semantic compensation scheme not only preserves the global semantics for classification, but also effectively retains the local semantics for segmentation. Pixel-Blur and Feat-Noise perform poorly, since they severely corrupts the image details. In contrast, ID-Adv and Privacy-Net, which incorporate a medical feature-matching loss, achieve slightly decent performance, but still lag far behind our approach. For instance, under ID-R=80%, our method outperforms Privacy-Net by approximately 8%, as shown in Figure 3(e)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Fundus Classification. Beyond X-ray images, our method also proves effective on fundus data. For instance, on EyePACS and ORID5K, our approach outperforms the second-best competitor ID-Adv by about " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", respectively. These results confirm that our approach generalizes well across different imaging modalities." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 432, + 184 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 432, + 184 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 432, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 432, + 184 + ], + "type": "image", + "image_path": "84ebb715e6aff1f92f6bcac6c42d18e0684e09850eea652fb7261aecfc4e5704.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 210 + ], + "type": "text", + "content": "Figure 4. (Left) Ablation on the framework design. (Right) Ablation study on the feature decoupling strategy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 549, + 184 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 549, + 184 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 549, + 184 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 549, + 184 + ], + "type": "image", + "image_path": "c067b500d0ecda17b4ae16ba8843054cac61225c4f617c31305272b1439c21d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 376, + 272 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 376, + 272 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 376, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 376, + 272 + ], + "type": "image", + "image_path": "484413447dbe3f3b1e24e657e8d0c093c97a2de92a3ab18c7a0221e23eca47ab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 274, + 358, + 281 + ], + "lines": [ + { + "bbox": [ + 333, + 274, + 358, + 281 + ], + "spans": [ + { + "bbox": [ + 333, + 274, + 358, + 281 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 213, + 435, + 272 + ], + "blocks": [ + { + "bbox": [ + 376, + 213, + 435, + 272 + ], + "lines": [ + { + "bbox": [ + 376, + 213, + 435, + 272 + ], + "spans": [ + { + "bbox": [ + 376, + 213, + 435, + 272 + ], + "type": "image", + "image_path": "fc0f1f5271607d965c29560134977575fbda663fdd274c3aec9228ad50406ce8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 274, + 413, + 281 + ], + "lines": [ + { + "bbox": [ + 397, + 274, + 413, + 281 + ], + "spans": [ + { + "bbox": [ + 397, + 274, + 413, + 281 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 213, + 490, + 272 + ], + "blocks": [ + { + "bbox": [ + 435, + 213, + 490, + 272 + ], + "lines": [ + { + "bbox": [ + 435, + 213, + 490, + 272 + ], + "spans": [ + { + "bbox": [ + 435, + 213, + 490, + 272 + ], + "type": "image", + "image_path": "d44fe9c5aaac6d1b055fc4c601b4c980c5eb05dd55be3825a532e23efd55cb31.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 274, + 475, + 281 + ], + "lines": [ + { + "bbox": [ + 452, + 274, + 475, + 281 + ], + "spans": [ + { + "bbox": [ + 452, + 274, + 475, + 281 + ], + "type": "text", + "content": "w/oSem" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 287, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 319 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of different variant models. The models are described in Figure 4 caption. Red arrow denotes the modified identity-related features. Best to view by zooming-in." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 490, + 213, + 551, + 272 + ], + "blocks": [ + { + "bbox": [ + 490, + 213, + 551, + 272 + ], + "lines": [ + { + "bbox": [ + 490, + 213, + 551, + 272 + ], + "spans": [ + { + "bbox": [ + 490, + 213, + 551, + 272 + ], + "type": "image", + "image_path": "240574cd23d8e87fbad067532886bd56f2b2361796c0e7b8a2dd01987d35b347.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 274, + 538, + 282 + ], + "lines": [ + { + "bbox": [ + 501, + 274, + 538, + 282 + ], + "spans": [ + { + "bbox": [ + 501, + 274, + 538, + 282 + ], + "type": "text", + "content": "w/oDecouple" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "text", + "content": "Fundus Segmentation. Our method achieves a Dice score of " + }, + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "inline_equation", + "content": "54.66\\%" + }, + { + "bbox": [ + 313, + 326, + 553, + 386 + ], + "type": "text", + "content": " on REFUGE2 at ID-R1=10%, largely surpassing MAE (27.33%), Privacy-Net (39.23%), and ID-Adv (41.23%). This further validates that our method also effectively preserves fine-grained semantic cues of eye fundus." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 395, + 409, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 409, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 409, + 407 + ], + "type": "text", + "content": "4.7. Model Analysis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "text", + "content": "Framework-Level Ablation Study. As shown in Figure 4 (Left), by removing the semantic branch, the AUROC of the resulted model 'w/oSem' dramatically drops by over " + }, + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "text", + "content": ", at CMC-R1=5%. On the other hand, without the identity-semantics decoupling mechanism, the resulting model 'w/oDecouple' leads to about " + }, + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 414, + 555, + 653 + ], + "type": "text", + "content": " CMC-R1 increase, for achieving the similar AUROC performance, since substantial identity cues are leaked from the vanilla medical features of MFMs. We further illustrate the protected images from different models. As shown in Figure 5, our results effectively modify identity-related features, such as the shape and location of the clavicle and chest contour. The 'w/oSem' model also removes these regions but significantly alters medical manifestations. In contrast, the 'w/oDecouple' model preserves medical features but fails to sufficiently suppress identity-related features, such as clavicle shape, due to residual identity information in the features from MFMs. These results confirm that both medical semantics and identity-semantics decoupling are essential for our advanced medical DeID approach." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Besides, we attempt to remove ID-Blocking module, tuning privacy via activating more decoupled ID codes. This leads to degraded AUROC (72.01%, 73.47%, 74.68%), compared to using ID-Blocking (72.68%, 74.35%, 79.62%) on MIMIC-X (10%, 20%, 40% at tack SR settings), due to" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20619" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 294, + 142 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 294, + 142 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 294, + 142 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 294, + 142 + ], + "type": "image", + "image_path": "df8bbd1cbc0022cd31f7ab2795244b47bfe48cc3b7d54d867ef2c40f606bbc97.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 145, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 178 + ], + "type": "text", + "content": "Figure 6. Comparison of semantic and identity information in terms of Bits-per-Pixel (bpp) [65], calculated as the feature code-length divided by the original image size." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 185, + 214, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 214, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 214, + 196 + ], + "type": "text", + "content": "the complete loss of pixel-wise details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 292 + ], + "type": "text", + "content": "Ablation Study on the Decoupling Strategy. As shown in Figure 4 (Right), omitting the codelength loss terms ('w/oCodeLengthLoss') fails to effectively remove identity information from MFM features, since the loose space cannot effectively decouple the identity and the semantics information. Moreover, removing the discrete code bottleneck ('w/oCodeSpace') further exacerbates identity leakage, leading to further inferior performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "text", + "content": "Furthermore, we quantitatively compare the overall and identity-related information in MFM features, as shown in Figure 6. First, we notice that a significant portion is identity-related, i.e., around " + }, + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 54, + 293, + 295, + 434 + ], + "type": "text", + "content": " for X-ray and fundus images. Second, the average information amount of the X-ray dataset Chest-Xray is 0.23bpp, much higher than 0.11bpp achieved by the fundus dataset EyePACS. This is aligned with the medical knowledge prior, that X-rays capture multiple organs and tissues, containing much complex information, than the fundus image that only focuses on eyes. This proves that the learned codelength effectively describes the medical data characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": "Finally, we analyze the impact of the codelength loss weight " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": " and the latent code channel number. As shown in Figure 7 (Left), reducing " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": " from 0.5 to 0.1 significantly increases CMC-R1 from " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "5.85\\%" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "12.34\\%" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": ", as a loosely constrained code space fails to effectively decouple identity information. Conversely, increasing " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": " from 0.5 to 2 has little effect on CMC-R1 but reduces AUROC performance by approximately " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": ", as an overly strong constraint impairs semantic feature reconstruction. The number of code channels also influences performance, by tuning the information capacity of the latent code, as shown in Figure 7 (Right). However, since " + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": " directly regulates the code-length term, the impact of the channel number is limited." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Discussion with Label-Conditioned Diffusion Models. These methods [48, 107, 120] employ task-specific labels (e.g., disease labels or text reports) to synthesize images, which are limited to label-associated tasks. In contrast, our approach is task-agnostic and applicable to diverse tasks. Moreover, after fine-tuning our approach towards a single task, i.e., replacing the MFM with a supervised classification network, our method achieves " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "81.92\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " AUROC at CMC-R1 = 0.30%, surpassing the label-conditioned model," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 433, + 160 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 433, + 160 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 433, + 160 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 433, + 160 + ], + "type": "image", + "image_path": "98358b9df95a800e9e382297e8f44ecd2788cd1858fac5dbc638c897a2b07695.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "lines": [ + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "type": "text", + "content": "Figure 7. (Left) Impact of the rate-distortion weight " + }, + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "type": "text", + "content": ". (Right:) Impact of the code dimension. All experiments are evaluated by masking " + }, + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 169, + 554, + 203 + ], + "type": "text", + "content": " identity-related regions, for a fair comparison." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 547, + 160 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 547, + 160 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 547, + 160 + ], + "type": "image", + "image_path": "96fc7bb0e8cda64f277c0f295fa73ce25d3e77aa0311592b67d5d53a02547bcb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 219, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 267 + ], + "type": "text", + "content": "i.e., " + }, + { + "bbox": [ + 313, + 219, + 553, + 267 + ], + "type": "inline_equation", + "content": "80.79\\%" + }, + { + "bbox": [ + 313, + 219, + 553, + 267 + ], + "type": "text", + "content": " AUROC at CMC-R1 = 0.29%. This confirms that our minimum-codelength representation also benefits the single-task setting, compared to the methods directly using the task labels guiding the diffusion procedure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 402 + ], + "type": "text", + "content": "Model Complexity. All methods and our model comprise about 380M parameters, for a fair comparison. Our inference time is 540 ms on an NVIDIA 4090 GPU, which is similar to MAE (526ms), but slower than Privacy-Net (120 ms), ID-Adv (122 ms), and Feat-Noise (124 ms), due to the multiple inference steps of diffusion procedure. Nonetheless, given the significant performance gains and that the medical imaging procedure itself is time-consuming, the running time is acceptable and does not hinder clinical workflows. In the future, we will integrate the single-step diffusion technique [113] to accelerate the process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 523, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 523, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 523, + 436 + ], + "type": "text", + "content": "5. Conclusion, Future Works, and Other" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 628 + ], + "type": "text", + "content": "Conclusion. We have presented DCM-DeID, a divide-and-conquer framework for medical image de-identification. By leveraging pre-trained Medical Foundation Models and a minimum codelength-based feature decoupling strategy, our method effectively remove identity cues, while preserving medical task utility. Extensive evaluations demonstrate the superiority of our approach. Future Works. We will extend our approach to multi-slice images, such as those produced by Magnetic Resonance Imaging (MRI). Broader Impacts. Our DeID technique is designed for medical AI applications, aiding the human. We emphasize that all rigorous clinical decisions must be made by human physicians using the original medical images. Furthermore, it is critical to enforce ethical guidelines, working in synergy with technological approaches to achieve medical privacy protection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No.62225112), the Fundamental Research Funds for the Central Universities, National Key R&D Program of China 2021YFE0206700, Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and STCSM 22DZ2229005." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20620" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Kadek YE, Aryanto, Andre Broekema, Matthijs Oudkerk, and Peter MA van Ooijen. Implementation of an anonymisation tool for clinical trials using a clinical trial processor integrated with an existing trial patient data information system. European radiology, 22:144-151, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 147, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 147, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 147, + 296, + 178 + ], + "type": "text", + "content": "[2] Johannes Balle, Valero Laparra, and Eero P Simoncelli. End-to-end optimized image compression. arXiv preprint arXiv:1611.01704, 2016. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 295, + 223 + ], + "type": "text", + "content": "[3] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. arXiv preprint arXiv:1802.01436, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 224, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 224, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 295, + 267 + ], + "type": "text", + "content": "[4] Andrew Barron, Jorma Rissanen, and Bin Yu. The minimum description length principle in coding and modeling. IEEE transactions on information theory, 44(6):2743-2760, 1998. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 269, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 269, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 66, + 269, + 296, + 291 + ], + "type": "text", + "content": "[5] Toby Berger. Rate-distortion theory. Wiley Encyclopedia of Telecommunications, 2003. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 292, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 292, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 66, + 292, + 295, + 324 + ], + "type": "text", + "content": "[6] Rui Bernardes, Pedro Serranho, and Conceicao Lobo. Digital ocular fundus imaging: a review. Ophthalmologica, 226 (4):161-181, 2011. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 325, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 325, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 66, + 325, + 295, + 413 + ], + "type": "text", + "content": "[7] Jeroen Bertels, Tom Eelbode, Maxim Berman, Dirk Vandermeulen, Frederik Maes, Raf Bisschops, and Matthew B Blaschko. Optimizing the dice score and jaccard index for medical image segmentation: Theory and practice. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part II 22, pages 92-100. Springer, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 414, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 414, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 66, + 414, + 295, + 467 + ], + "type": "text", + "content": "[8] Amit Bhati, Neha Gour, Pritee Khanna, and Aparajita Ojha. Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image dataset. Computers in Biology and Medicine, 153:106519, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 469, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 469, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 66, + 469, + 295, + 512 + ], + "type": "text", + "content": "[9] Vincent Bindschaedler, Paul Grubbs, David Cash, Thomas Ristenpart, and Vitaly Shmatikov. The tao of inference in privacy-protected databases. Cryptology ePrint Archive, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 514, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 514, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 61, + 514, + 295, + 568 + ], + "type": "text", + "content": "[10] Ruud M Bolle, Jonathan H Connell, Sharath Pankanti, Nalini K Ratha, and Andrew W Senior. The relation between the roc curve and the cmc. In Fourth IEEE workshop on automatic identification advanced technologies (AutoID'05), pages 15-20. IEEE, 2005. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 570, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 570, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 61, + 570, + 295, + 602 + ], + "type": "text", + "content": "[11] Andrew P Bradley. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145-1159, 1997. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 61, + 603, + 295, + 658 + ], + "type": "text", + "content": "[12] Jingyi Cao, Bo Liu, Yunqian Wen, Rong Xie, and Li Song. Personalized and invertible face de-identification by disentangled identity information manipulation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3334–3342, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 295, + 712 + ], + "type": "text", + "content": "[13] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "text", + "content": "[15] Yan Chen and Pouyan Esmaeilzadeh. Generative ai in medical practice: in-depth exploration of privacy and security challenges. Journal of Medical Internet Research, 26:e53008, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "type": "text", + "content": "[16] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7939-7948, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 231, + 553, + 262 + ], + "type": "text", + "content": "[17] I Glenn Cohen and Michelle M Mello. Big data, big tech, and protecting patient privacy. *Jama*, 322(12):1141-1142, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 553, + 286 + ], + "type": "text", + "content": "[18] Thomas M Cover. Elements of information theory. John Wiley & Sons, 1999. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 288, + 553, + 331 + ], + "type": "text", + "content": "[19] William L Croft, Jörg-Rüdiger Sack, and Wei Shi. Obfuscation of images via differential privacy: From facial images to general images. Peer-to-Peer Networking and Applications, 14:1705-1733, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 333, + 553, + 376 + ], + "type": "text", + "content": "[20] Ji Dai, Behrouz Saghafi, Jonathan Wu, Janusz Konrad, and Prakash Ishwar. Towards privacy-preserving recognition of human activities. In 2015 IEEE international conference on image processing (ICIP), pages 4238-4242. IEEE, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 378, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 378, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 320, + 378, + 553, + 422 + ], + "type": "text", + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 423, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 423, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 320, + 423, + 553, + 478 + ], + "type": "text", + "content": "[22] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 479, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 553, + 534 + ], + "type": "text", + "content": "[23] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5154–5163, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 536, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 536, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 320, + 536, + 553, + 612 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 613, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 553, + 658 + ], + "type": "text", + "content": "[25] Emma Dugas, Jared, Jorge, and Will Cukierski. Diabetic retinopathy detection. https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. Kaggle. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[26] August DuMont Schütte, Jürgen Hetzel, Sergios Gatidis, Tobias Hepp, Benedikt Dietz, Stefan Bauer, and Patrick Schwab. Overcoming barriers to data sharing with medical image generation: a comprehensive evaluation. NPJ digital medicine, 4(1):141, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20621" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[27] Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming, pages 1-12. Springer, 2006. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 294, + 171 + ], + "type": "text", + "content": "[28] Weijie Fan, Yi Yang, Jing Qi, Qichuan Zhang, Cuiwei Liao, Li Wen, Shuang Wang, Guangxian Wang, Yu Xia, Qihua Wu, et al. A deep-learning-based framework for identifying and localizing multiple abnormalities and assessing cardiomegaly in chest x-ray. Nature Communications, 15(1): 1347, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 172, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 294, + 227 + ], + "type": "text", + "content": "[29] Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, et al. Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. arXiv preprint arXiv:2202.08994, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 248 + ], + "type": "text", + "content": "[30] Bruce Fischl. Freesurfer. Neuroimage, 62(2):774-781, 2012. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 294, + 316 + ], + "type": "text", + "content": "[31] Alex Gaudio, Asim Smailagic, Christos Faloutsos, Shreshta Mohan, Elvin Johnson, Yuhao Liu, Pedro Costa, and Aurélio Campilho. Deepfixcx: Explainable privacy-preserving image compression for medical image analysis. Wiley interdisciplinary reviews: Data mining and knowledge discovery, 13(4):e1495, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 338 + ], + "type": "text", + "content": "[32] Satrajit Ghosh, Chris Gorgolewski, et al. pydeface: A tool to remove facial features from mri images, 2012. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 360 + ], + "type": "text", + "content": "[33] Peter D Grünwald. The minimum description length principle. MIT press, 2007. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 361, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 413 + ], + "type": "text", + "content": "[34] Xiuye Gu, Weixin Luo, Michael S Ryoo, and Yong Jae Lee. Password-conditioned anonymization and deanonymization with face identity transformers. In European conference on computer vision, pages 727-743. Springer, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "text", + "content": "[35] A Gueziec and P Kazanzides. Anatomy-based registration of ct-scan and intraoperative x-ray images for guiding a surgical robot. IEEE Transactions on Medical Imaging, 17(5): 715-728, 1998. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 503 + ], + "type": "text", + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 294, + 557 + ], + "type": "text", + "content": "[37] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "text", + "content": "[38] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 647 + ], + "type": "text", + "content": "[39] Alexander Hermans, Lucas Beyer, and Bastian Leibe. In defense of the triplet loss for person re-identification. arXiv preprint arXiv:1703.07737, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[40] Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Steven Hill, Zhimin Zhou, Lawrence Saul, and Hovav Shacham. On the (in) effectiveness of mosaicing and blurring as tools for document redaction. Proceedings on Privacy Enhancing Technologies, 2016. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 149 + ], + "type": "text", + "content": "[42] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 150, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 150, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 150, + 553, + 259 + ], + "type": "text", + "content": "[43] Sungmin Hong, Razvan Marinescu, Adrian V Dalca, Anna K Bonkhoff, Martin Bretzner, Natalia S Rost, and Polina Golland. 3d-stylegan: A style-based generative adversarial network for generative modeling of three-dimensional medical images. In Deep Generative Models, and Data Augmentation, Labelling, and Imperfections: First Workshop, DGM4MICCAI 2021, and First Workshop, DALI 2021, Held in Conjunction with MICCAI 2021, Strasbourg, France, October 1, 2021, Proceedings 1, pages 24-34. Springer, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 553, + 293 + ], + "type": "text", + "content": "[44] Andrew Hoopes, Jocelyn S Mora, Adrian V Dalca, Bruce Fischl, and Malte Hoffmann. Synthstrip: skull-stripping for any brain image. NeuroImage, 260:119474, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 553, + 347 + ], + "type": "text", + "content": "[45] Qiang Hu, Qihan He, Houqiang Zhong, Guo Lu, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. Varfvv: View-adaptive real-time interactive free-view video streaming with edge computing. arXiv preprint arXiv:2501.13630, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 349, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 349, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 349, + 555, + 414 + ], + "type": "text", + "content": "[46] Qiang Hu, Zihan Zheng, Houqiang Zhong, Sihua Fu, Li Song, Xiaoyun Zhang, Guangtao Zhai, and Yanfeng Wang. 4dgc: Rate-aware 4d gaussian compression for efficient streamable free-viewpoint video. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 875-885, 2025." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 470 + ], + "type": "text", + "content": "[47] Qiang Hu, Houqiang Zhong, Zihan Zheng, Xiaoyun Zhang, Zhengxue Cheng, Li Song, Guangtao Zhai, and Yanfeng Wang. Vrvvc: Variable-rate nerf-based volumetric video compression. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3563-3571, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 553, + 525 + ], + "type": "text", + "content": "[48] Peng Huang, Xue Gao, Lihong Huang, Jing Jiao, Xiaokang Li, Yuanyuan Wang, and Yi Guo. Chest-diffusion: a lightweight text-to-image model for report-to-cxr generation. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 525, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 553, + 578 + ], + "type": "text", + "content": "[49] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203–211, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 581, + 553, + 613 + ], + "type": "text", + "content": "[50] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 553, + 669 + ], + "type": "text", + "content": "[51] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "text", + "content": "[52] Georgios A Kaissis, Marcus R Makowski, Daniel Rückert, and Rickmer F Braren. Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6):305-311, 2020. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20622" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[53] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 297, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 297, + 173 + ], + "type": "text", + "content": "[54] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 296, + 228 + ], + "type": "text", + "content": "[55] Bach Ngoc Kim, Jose Dolz, Pierre-Marc Jodoin, and Christian Desrosiers. Privacy-net: an adversarial approach for identity-obfuscated segmentation of medical images. IEEE Transactions on Medical Imaging, 40(7):1737-1749, 2021. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 295, + 263 + ], + "type": "text", + "content": "[56] Cody A Koch and Wayne F Larrabee. Patient privacy, photographs, and publication. JAMA facial plastic surgery, 15 (5):335-336, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 265, + 295, + 319 + ], + "type": "text", + "content": "[57] Abhinav Kumar, Sanjay Kumar Singh, K Lakshmanan, Sonal Saxena, and Sameer Shrivastava. A novel cloud-assisted secure deep feature classification framework for cancer histopathology images. ACM Transactions on Internet Technology (TOIT), 21(2):1-22, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "type": "text", + "content": "[58] Seowoo Lee, Jiwon Youn, Hyungjin Kim, Mansu Kim, and Soon Ho Yoon. Cxr-llava: a multimodal large language model for interpreting chest x-ray images. European Radiology, pages 1-13, 2025. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 366, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 295, + 422 + ], + "type": "text", + "content": "[59] Chunyi Li, Zicheng Zhang, Haoning Wu, Wei Sun, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Agiaq-3k: An open database for ai-generated image quality assessment. IEEE Transactions on Circuits and Systems for Video Technology, 34(8):6833-6846, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "text", + "content": "[60] Chunyi Li, Xiaozhe Li, Zicheng Zhang, Yuan Tian, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Jia Wang, Haodong Duan, Kai Chen, et al. Information density principle for mllm benchmarks. arXiv preprint arXiv:2503.10079, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "type": "text", + "content": "[61] Chunyi Li, Yuan Tian, Xiaoyue Ling, Zicheng Zhang, Haodong Duan, Haoning Wu, Ziheng Jia, Xiaohong Liu, Xiongkuo Min, Guo Lu, et al. Image quality assessment: From human to machine preference. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 7570-7581, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "type": "text", + "content": "[62] Chunyi Li, Jiaohao Xiao, Jianbo Zhang, Farong Wen, Zicheng Zhang, Yuan Tian, Xiangyang Zhu, Xiaohong Liu, Zhengxue Cheng, Weisi Lin, et al. Perceptual quality assessment for embodied ai. arXiv preprint arXiv:2505.16815, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 657 + ], + "type": "text", + "content": "[63] Yuheng Li, Krishna Kumar Singh, Utkarsh Ojha, and Yong Jae Lee. Mixmatch: Multifactor disentanglement and encoding for conditional image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8039-8048, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[64] Jie Lian, Jingyu Liu, Shu Zhang, Kai Gao, Xiaqing Liu, Dingwen Zhang, and Yizhou Yu. A structure-aware relation network for thoracic diseases detection and segmentation. IEEE Transactions on Medical Imaging, 40(8):2042-2052, 2021. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[65] Weisi Lin and Li Dong. Adaptive downsampling to improve image compression at low bit rates. IEEE Transactions on Image Processing, 15(9):2513-2521, 2006. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 129 + ], + "type": "text", + "content": "[66] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 320, + 130, + 553, + 163 + ], + "type": "text", + "content": "[67] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "type": "text", + "content": "[68] Charles Marsh. Introduction to continuous entropy. Department of Computer Science, Princeton University, 1034, 2013. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 252 + ], + "type": "text", + "content": "[69] Maxim Maximov, Ismail Elezi, and Laura Leal-Taixe. Ciagan: Conditional identity anonymization generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5447-5456, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "type": "text", + "content": "[70] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Conditional probability models for deep image compression. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4394-4402, 2018. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 310, + 553, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 310, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 320, + 310, + 553, + 343 + ], + "type": "text", + "content": "[71] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 320, + 344, + 553, + 388 + ], + "type": "text", + "content": "[72] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 320, + 389, + 553, + 432 + ], + "type": "text", + "content": "[73] Eriksson Monteiro, Carlos Costa, and José Luís Oliveira. A de-identification pipeline for ultrasound medical images in dicom format. Journal of Medical Systems, 41(5):89, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 434, + 553, + 488 + ], + "type": "text", + "content": "[74] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 544 + ], + "type": "text", + "content": "[75] Zhenliang Ni, Xinghao Chen, Yingjie Zhai, Yehui Tang, and Yunhe Wang. Context-guided spatial feature reconstruction for efficient semantic segmentation. In European Conference on Computer Vision, pages 239-255. Springer, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 545, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 545, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 545, + 553, + 601 + ], + "type": "text", + "content": "[76] Kai Packhäuser, Sebastian Gündel, Nicolas Münster, Christopher Syben, Vincent Christlein, and Andreas Maier. Deep learning-based patient re-identification is able to exploit the biometric nature of medical chest x-ray data. Scientific Reports, 12(1):14851, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 553, + 668 + ], + "type": "text", + "content": "[77] Kai Packhäuser, Sebastian Gündel, Florian Thamm, Felix Denzinger, and Andreas Maier. Deep learning-based anonymization of chest radiographs: a utility-preserving measure for patient privacy. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 262-272. Springer, 2023. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 669, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 669, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 669, + 553, + 714 + ], + "type": "text", + "content": "[78] Pascal Paillier. Public-key cryptosystems based on composite degree residuosity classes. In International conference on the theory and applications of cryptographic techniques, pages 223-238. Springer, 1999. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20623" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 127 + ], + "type": "text", + "content": "[79] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 194 + ], + "type": "text", + "content": "[80] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "type": "text", + "content": "[81] W. Nicholson Price and I. Glenn Cohen. Privacy in the age of medical big data. Nature Medicine, 25(1):37-43, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "text", + "content": "[82] Moo-Ryong Ra, Ramesh Govindan, and Antonio Ortega. P3: Toward {Privacy-Preserving} photo sharing. In 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI 13), pages 515-528, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 337 + ], + "type": "text", + "content": "[83] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 338, + 296, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 296, + 359 + ], + "type": "text", + "content": "[84] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663):3, 2009. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 360, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 61, + 360, + 296, + 403 + ], + "type": "text", + "content": "[85] David Rodríguez González, Trevor Carpenter, Jano I van Hemert, and Joanna Wardlaw. An open source toolkit for medical imaging de-identification. European radiology, 20: 1896-1904, 2010. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 61, + 404, + 296, + 459 + ], + "type": "text", + "content": "[86] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 61, + 460, + 296, + 514 + ], + "type": "text", + "content": "[87] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 514, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 514, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 61, + 514, + 296, + 569 + ], + "type": "text", + "content": "[88] Huajie Shao, Yifei Yang, Haohong Lin, Longzhong Lin, Yizhuo Chen, Qinmin Yang, and Han Zhao. Rethinking controllable variational autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19250-19259, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 570, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 570, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 61, + 570, + 296, + 602 + ], + "type": "text", + "content": "[89] Julie K. Taitsman, Christi Macrina Grimm, and Shantanu Agrawal. Protecting patient privacy and data security. New England Journal of Medicine, 368(11):977-979, 2013. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 603, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 603, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 61, + 603, + 296, + 624 + ], + "type": "text", + "content": "[90] Adam Tanner. Our bodies, our data: how companies make billions selling our medical records. Beacon Press, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 296, + 658 + ], + "type": "text", + "content": "[91] Huan Tian, Tianqing Zhu, and Wanlei Zhou. Fairness and privacy preservation for facial images: Gan-based methods. Computers & Security, 122:102902, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 713 + ], + "type": "text", + "content": "[92] Yuan Tian, Guo Lu, Xiongkuo Min, Zhaohui Che, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Self-conditioned probabilistic learning of video rescaling. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4490-4499, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "text", + "content": "[93] Yuan Tian, Yichao Yan, Guangtao Zhai, Guodong Guo, and Zhiyong Gao. Ean: event adaptive network for enhanced action recognition. International Journal of Computer Vision, 130(10):2453-2471, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 173 + ], + "type": "text", + "content": "[94] Yuan Tian, Guo Lu, Guangtao Zhai, and Zhiyong Gao. Non-semantics suppressed mask learning for unsupervised video semantic compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13610-13622, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 554, + 217 + ], + "type": "text", + "content": "[95] Yuan Tian, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. Clsa: a contrastive learning framework with selective aggregation for video rescaling. IEEE Transactions on Image Processing, 32:1300-1314, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "text", + "content": "[96] Yuan Tian, Guo Lu, Yichao Yan, Guangtao Zhai, Li Chen, and Zhiyong Gao. A coding framework and benchmark towards low-bitrate video understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 554, + 308 + ], + "type": "text", + "content": "[97] Yuan Tian, Guo Lu, and Guangtao Zhai. Free-vsc: Free semantics from visual foundation models for unsupervised video semantic compression. In European Conference on Computer Vision, pages 163-183. Springer, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 310, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 310, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 310, + 554, + 342 + ], + "type": "text", + "content": "[98] Yuan Tian, Guo Lu, and Guangtao Zhai. Smc++: Masked learning of unsupervised video semantic compression. arXiv preprint arXiv:2406.04765, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 344, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 344, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 344, + 554, + 398 + ], + "type": "text", + "content": "[99] Yuan Tian, Kaiyuan Ji, Rongzhao Zhang, Yankai Jiang, Chunyi Li, Xiaosong Wang, and Guangtao Zhai. Towards all-in-one medical image re-identification. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30774–30786, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 554, + 443 + ], + "type": "text", + "content": "[100] Yuan Tian, Shuo Wang, and Guangtao Zhai. Medical manifestation-aware de-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 26363–26372, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 445, + 554, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 554, + 489 + ], + "type": "text", + "content": "[101] Matt Tierney, Ian Spiro, Christoph Bregler, and Lakshminarayanan Subramanian. Cryptogram: Photo privacy for online social media. In Proceedings of the first ACM conference on Online social networks, pages 75-88, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 491, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 554, + 533 + ], + "type": "text", + "content": "[102] Gary Kin-wai Tsui and Tao Chan. Automatic selective removal of embedded patient information from image content of dicom files. American Journal of Roentgenology, 198(4): 769-772, 2012. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 535, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 535, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 315, + 535, + 554, + 600 + ], + "type": "text", + "content": "[103] Nishant Vishwamitra, Bart Knijnenburg, Hongxin Hu, Yifang P Kelly Caine, et al. Blur vs. block: Investigating the effectiveness of privacy-enhancing obfuscation for images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 39-47, 2017. 1, 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 602, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 602, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 315, + 602, + 554, + 657 + ], + "type": "text", + "content": "[104] Zhongwei Wan, Che Liu, Mi Zhang, Jie Fu, Benyou Wang, Sibo Cheng, Lei Ma, César Quilodrán-Casas, and Rossella Arcucci. Med-unic: Unifying cross-lingual medical vision-language pre-training by diminishing bias. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 554, + 713 + ], + "type": "text", + "content": "[105] Fuying Wang, Yuyin Zhou, Shujun Wang, Varut Vardhanabhuti, and Lequan Yu. Multi-granularity cross-modal alignment for generalized medical visual representation learning. Advances in Neural Information Processing Systems, 35:33536-33549, 2022. 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20624" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[106] Guangyu Wang, Xiaohong Liu, Jun Shen, Chengdi Wang, Zhihuan Li, Linsen Ye, Xingwang Wu, Ting Chen, Kai Wang, Xuan Zhang, et al. A deep-learning pipeline for the diagnosis and discrimination of viral, non-viral and COVID-19 pneumonia from chest x-ray images. Nature biomedical engineering, 5(6):509–521, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[107] Jinzhuo Wang, Kai Wang, Yunfang Yu, Yuxing Lu, Wenchao Xiao, Zhuo Sun, Fei Liu, Zixing Zou, Yuanxu Gao, Lei Yang, et al. Self-improving generative foundation model for synthetic medical image generation and clinical applications. Nature Medicine, pages 1-9, 2024. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[108] Kai Wang, Bo Zhao, Xiangyu Peng, Zheng Zhu, Jiankang Deng, Xinchao Wang, Hakan Bilen, and Yang You. Facemae: Privacy-preserving face recognition via masked autoencoders. arXiv preprint arXiv:2205.11090, 2022. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[109] Shudong Wang, Zhiyuan Zhao, Yawu Zhao, Luqi Wang, Yuanyuan Zhang, Jiehuan Wang, Sibo Qiao, and Zhihan Lyu. A semantic conditional diffusion model for enhanced personal privacy preservation in medical images. IEEE Journal of Biomedical and Health Informatics, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[110] Weiru Wang, Chi-Man Vong, Yilong Yang, and Pak-Kin Wong. Encrypted image classification based on multilayer extreme learning machine. Multidimensional Systems and Signal Processing, 28(3):851-865, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "text", + "content": "[111] Ruiqi Wu, Chenran Zhang, Jianle Zhang, Yi Zhou, Tao Zhou, and Huazhu Fu. Mm-retinal: Knowledge-enhanced foundational pretraining with fundus image-text expertise. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 722-732. Springer, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[112] Hanyu Xue, Bo Liu, Ming Ding, Tianqing Zhu, Dayong Ye, Li Song, and Wanlei Zhou. Dp-image: Differential privacy for image data in feature space. arXiv preprint arXiv:2103.07073, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 504 + ], + "type": "text", + "content": "[113] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6613-6623, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[114] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. Privacy-preserving photo sharing based on a secure JPEG. In 2015 IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pages 185-190. IEEE, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[115] Lin Yuan, Pavel Korshunov, and Touradj Ebrahimi. SecureJPEG scrambling enabling privacy in photo sharing. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), pages 1-6. IEEE, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[116] Junhai Zhai, Sufang Zhang, Junfen Chen, and Qiang He. Autoencoder and its various variants. In 2018 IEEE international conference on systems, man, and cybernetics (SMC), pages 415-419. IEEE, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[117] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[118] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 452 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "text", + "content": "for effectively learning deep face representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10823-10832, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[119] Hong-Yu Zhou, Xiaoyu Chen, Yinghao Zhang, Ruibang Luo, Liansheng Wang, and Yizhou Yu. Generalized radiograph representation learning via cross-supervision between images and free-text radiology reports. Nature Machine Intelligence, 4(1):32-40, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[120] Xinrui Zhou, Yuhao Huang, Haoran Dou, Shijing Chen, Ao Chang, Jia Liu, Weiran Long, Jian Zheng, Erjiao Xu, Jie Ren, et al. Ctrl-genaug: Controllable generative augmentation for medical sequence classification. arXiv preprint arXiv:2409.17091, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 284 + ], + "type": "text", + "content": "[121] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter WoodwardCourt, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "text", + "content": "[122] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In Proceedings of the 41st International Conference on Machine Learning, pages 62429-62442. PMLR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[123] Yingxuan Zhu, PD Singh, Khan Siddiqui, and Michael Gillam. An automatic system to detect and extract texts in medical images for de-identification. In Medical Imaging 2010: Advanced PACS-based Imaging Informatics and Therapeutic Applications, page 762803. SPIE, 2010. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[124] M Tarek Ibn Ziad, Amr Alanwar, Moustafa Alzantot, and Mani Srivastava. Cryptoimg: Privacy preserving processing over encrypted images. In 2016 IEEE Conference on Communications and Network Security (CNS), pages 570-575. IEEE, 2016. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20625" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_content_list.json b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7b942ecd351019741a83a2512c6a2f4ce53c8bcd --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_content_list.json @@ -0,0 +1,1621 @@ +[ + { + "type": "text", + "text": "Semantic-guided Camera Ray Regression for Visual Localization", + "text_level": 1, + "bbox": [ + 169, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yesheng Zhang, Xu Zhao* \nSchool of Automation and Intelligent Sensing, Shanghai Jiao Tong University \n{preacher, zhaoxu}@sjtu.edu.cn", + "bbox": [ + 191, + 181, + 805, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work presents a novel framework for Visual Localization (VL), that is, regressing camera rays from query images to derive camera poses. As an overparameterized representation of the camera pose, camera rays possess superior robustness in optimization. Of particular importance, Camera Ray Regression (CRR) is privacy-preserving, rendering it a viable VL approach for real-world applications. Thus, we introduce DINO-based Multi-Mappers, coined DIMM, to achieve VL by CRR. DIMM utilizes DINO as a scene-agnostic encoder to obtain powerful features from images. To mitigate ambiguity, the features integrate both local and global perception, as well as potential geometric constraint. Then, a scene-specific mapper head regresses camera rays from these features. It incorporates a semantic attention module for soft fusion of multiple mappers, utilizing the rich semantic information in DINO features. In extensive experiments on both indoor and outdoor datasets, our methods showcase impressive performance, revealing a promising direction for advancements in VL.", + "bbox": [ + 88, + 301, + 485, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 619, + 220, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Localization (VL) involves estimating 6-degree-of-freedom camera pose of an image taken from a known scene, which is also known as camera relocalization. VL plays a vital role in many vision, robotic and graphic applications, such as augment reality, virtual reality and autonomous driving. Despite extensive research, achieving precise, robust, and privacy-preserving VL [28, 38] continues to present a significant challenge.", + "bbox": [ + 89, + 645, + 483, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current VL methods can be categorized into explicit and implicit methods based on their mapping approaches. Traditional explicit methods [2, 30, 31] involve an explicit 3D map, e.g., point clouds. They establish 2D-3D (image-to-map) correspondences by Feature Matching, and then estimate camera pose via PnP and RANSAC [10, 12]. While advanced methods [11, 39] leads to high accuracy, the stor", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1094e448dc67043bf342f4052a9dad3b42f9463fdd4b0c457783c941a957d08.jpg", + "image_caption": [ + "Figure 1. The ambiguity challenge in Camera Ray Regression (CRR) for VL. Effective global and local disambiguations are critical for CRR. Global ambiguity arises from rays of different patches (cf. $r_2^1$ and $r_2^2$ in the figure), while local ambiguity occurs within the same patch but from different viewing angles (cf. $r_1^1$ and $r_2^2$ ). Semantic information, e.g., the table in the figure to differentiate between chairs, along with geometry constraints (geometry properties of rays $\\{r_1^i\\}_i$ ) are essential for the disambiguations." + ], + "image_footnote": [], + "bbox": [ + 598, + 268, + 823, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age demands of explicit maps are troublesome especially in large-scale scenes. Additionally, explicit maps raise privacy concerns, restricting their applications in real-world settings.", + "bbox": [ + 511, + 532, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Implicit methods, on the other hand, leverage learning models as implicit and light-weight maps, which mainly include two frameworks: Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR). APR directly utilizes networks to regress camera poses (rotations and translations) from images [6, 15], offering enhanced privacy protection. However, the compact nature of absolute pose makes it vulnerable to noise, thereby leading to accuracy issue [6].", + "bbox": [ + 511, + 580, + 908, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To improve precision, employing an overparameterized representation for the camera pose is crucial [35]. Thus, SCR framework [22, 41] is proposed, relying on 2D-3D matches as an indirect yet robust camera pose representation, akin to its explicit counterparts. ACE [4] is a milestone within this framework, utilizing simple networks to regress pixel-aligned scene coordinates. It introduces a scene-centric patch-level training strategy, i.e., Gradient Decorrelation Training (GDT), enabling both efficient and precise mapping. Nevertheless, the unbounded 3D point error is unstable in training, necessitating manual hyperparameter tuning or additional outlier rejection modules [5, 20]. Besides, SCR methods still run the risk of 3D privacy breaches, as they", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 91, + 886, + 223, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25639", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94306d1c810ff5d230ee5804b658da387401ef5a043ed499d24347ee427df6fa.jpg", + "image_caption": [ + "Figure 2. The proposed DIMM approach performing CRR for VL. This approach utilizes DINO as a scene-agnostic feature encoder to output carefully designed patch-level features. Then, a scene-specific mapper head integrates multi-mapper results under semantic guidance, through a semantic attention module. After the scene-centric training technology, DIMM is capable of predicting camera ray parameters for accurate and privacy-preserving VL." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 473, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "disclose detailed 3D scene information [29].", + "bbox": [ + 89, + 339, + 383, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a promising camera pose representation, camera ray [27, 35], has garnered increased attention in the vision field. Its overparameterization is beneficial for patch-level learning [23, 44], aligning well with the effective GDT in VL. This inspires us to introduce camera rays in the VL tasks. In particular, we propose adopting a learning model to regress patch-level camera rays from the image, termed as Camera Ray Regression (CRR). Then, the camera pose can be achieved by solving two linear problems [44]. This implementation of VL could strike a balance between efficacy and privacy preservation. Firstly, it belongs to the implicit mapping category, incurring low storage costs. As a overparameterized pose representation, it can yield improved precision [35] than APR. Also, the ray error is more stable than the unbounded 3D point error [27] in optimization of model training. Hence, CRR obviates the need for intricate training setting like SCR methods [5, 20, 41]. More importantly, it is hard to recover 3D scene information from camera rays, leading to a privacy-preserving VL implementation.", + "bbox": [ + 89, + 356, + 483, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, challenges exist in introducing camera rays in VL, primarily stemming from the issue of ambiguity (cf. Fig. 1). This ambiguity can generally be divided into two levels. The first is global ambiguity, referring to distinguish rays of different image patches, particularly in repetitive or textureless scenes. The similar ambiguity is encountered in SCR [4, 41], where leveraging powerful local features, such as semantic features [20], has proven effective. The second is local ambiguity, unique to CRR, which involves differentiating rays from the same image patch with different view angles. The ambiguity becomes troublesome for patches with continuous depth, where changes in view direction result in minor differences in patch content but obvious variations in the corresponding rays. To eliminate this ambiguity, we emphasize two observations on the CRR task. 1) The entire image often contains patches with depth discontinuities. Their appearance changes under different", + "bbox": [ + 89, + 645, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "viewing angles are significant enough to resolve the local ambiguity. Thus, the image-level perception is important for CRR disambiguation, which is also proven to be beneficial for handling global repetitiveness [41]. 2) The rays within the same image are governed by the same camera pose. This geometric constraint allows confident rays to correct uncertain ones. Meanwhile, the smoothness of the patchray mapping [44] makes this geometric constraint learnable. Therefore, the integration of image-level perception, geometric constraint and robust local features is critical for the global and local disambiguation of CRR.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above analysis, this work presents DINO-based Multi-Mappers (DIMM, Fig. 2), as the first implementation of CRR in VL. Following the scene-centric design of ACE, DIMM consists of a scene-agnostic encoder and a scene-specific head, but possessing several key modules tailed for CRR. Specifically, we utilize DINO [8] as the feature encoder, utilizing its powerful semantic perception for disambiguation. It provides robust patch tokens as local features while supporting CLS tokens for image-level perception. We further incorporate Fourier Encoding [24] of the image position for each patch into the features (XY tokens), as the potential geometric constraint of camera rays are associated with patch locations. Furthermore, we propose a scene-specific mapper head which adopts a semantic attention module to softly fuse ray regression results from multiple Multi-Layer Perceptron (MLP)-based mappers. This head effectively utilizes the inherit semantic prior of the DINO features, to enhance ray regression performance especially in large scenes. Finally, we propose a ray-level RANSAC algorithm for accurate pose estimation, decomposing the rotation and translation calculation [44].", + "bbox": [ + 511, + 258, + 908, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summaries as follows.", + "bbox": [ + 532, + 577, + 823, + 590 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. To the best of our knowledge, we are the first to introduce the camera ray into VL, with a tailored network design to achieve query image poses by camera ray regression.", + "2. We propose carefully designed scene-agnostic features to handle local and global disambiguation in CRR, along with a semantic-guided mapper head that enables a soft multi-mapper ensemble.", + "3. Comprehensive experiments on indoor and outdoor datasets demonstrate the effectiveness of our method, achieving results that rival or surpass the existing state-of-the-art. Additionally, extensive ablation studies validate the contributions of each module." + ], + "bbox": [ + 514, + 597, + 906, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 795, + 653, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Matching-based VL. Traditional VL methods [30, 32, 33] rely on Feature Matching (FM) to determine the position and direction of cameras. These techniques aim to match 3D scene points with 2D image points, enabling the camera pose calculation through $\\mathrm{PnP} + \\mathrm{RANSAC}$ -based", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25640", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e18d4064dc6c3b2e7eba9d92fce0f63485b0302b383fdf56a3b0a6343f91569.jpg", + "image_caption": [ + "Figure 3. The Multi-Mapper network with semantic attention as our scene-specific head. The patch-level features are first projected to geometry and semantic embeddings. The geometry embedding is fed into multiple MLP-based mappers to derive ray embeddings. These embeddings are then fused under the guidance of the semantic embedding to achieve the camera rays, through cross attention modules." + ], + "image_footnote": [], + "bbox": [ + 214, + 88, + 782, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "algorithms [10, 12]. As a result, these approaches necessitate the storage of explicit maps containing 3D points and their feature descriptors. By leveraging robust matching algorithms [11, 19, 39, 42, 45], FM-based localization can deliver precise outcomes. Nevertheless, explicit maps raises concerns about privacy leaks [28]. Furthermore, as the scene size expands, managing the increasingly large explicit maps becomes challenging, leading to storage issues. In contrast, our method employs implicit maps that have lower storage requirements and privacy protection.", + "bbox": [ + 88, + 297, + 485, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Coordinate Regression. Similar to FM-based VL, Scene Coordinate Regression (SCR) also accomplishes camera poses through 3D-2D matching [3]. However, SCR directly employs network to regress pixel-aligned 3D scene coordinates, which addresses the map storage issue through implicit mapping. ACE [4], a prominent approach within SCR, introduces a gradient decorrelation training technique enabling mapping within 5 minutes using a 4MB network. Subsequent methods [5, 14, 20, 22, 41] built upon ACE further enhance VL accuracy mainly by solid feature selection. Nevertheless, the limited network size of ACE pose challenges in fitting large scenes, necessitating spatial segmentation of the large scene for sub-map construction. Despite its efficacy, this approach may result in ambiguous localization at the boundaries of segmented sub-scenes. At the same time, it is noteworthy that the SCR methods output the 3D information of scenes, i.e., 3D coordinates, thereby leading to a potential risk of privacy leak.", + "bbox": [ + 91, + 455, + 483, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Absolute Pose Regression. Absolute Pose Regression (APR) also adopts implicit mapping, but directly utilizes networks to predict the rotation and translation of cameras. However, this compact parameter representation of camera pose is highly sensitive to noises [12, 44]. Regressing pose parameters from images, thus, is challenging [15-17, 36], and the accuracy is insufficient to compare with methods using over-parameterized representations, e.g., 2D-3D correspondences. Recent APR approach [6] opts to predict the camera pose based on the SCR output, resulting in favorable accuracy. However, this success is heavily reliant", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on the precision of SCR methods. In contrast, our method employs camera rays to depict the camera pose. This overparameterized representation is more conducive to learning optimization [25, 27, 44], enabling superior accuracy.", + "bbox": [ + 511, + 297, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera Ray in Vision. Recently, camera rays have emerged as a promising camera parametrization, representing a generic camera configuration [35]. In recent work [23, 43, 44], camera rays have been proven to be well-suited for patch-level learning, with applications in multiple vision tasks. Also, in 3D reconstruction works [25, 27], camera rays yield more robust geometric optimization than classical correspondence-based representations, e.g., 3D-2D matches. This advantage stems from their overparameterized nature and constrained ray error. Likewise, our approach is the first to introduce camera rays in VL, performing Camera Ray Regression (CRR) to achieve camera poses from images. Its patch-level property aligns well with the Gradient Deceleration Training [4], and the bounded ray error is beneficial for optimization in VL Learning as well. Furthermore, the inherent privacy-preserving attribute of CRR positions it as a promising direction for VL implementations.", + "bbox": [ + 511, + 361, + 908, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 631, + 604, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the formulation of CRR for the VL task in Sec. 3.1. Then the proposed DIMM method is described in detail. DIMM adopts a scene-centric setup [4, 14, 41] by dividing the network into a scene-agnostic encoder (Sec. 3.2) and a scene-specific head (Sec. 3.3). Additionally, the mapping process incorporates GDT [4] and utilizes a specially designed loss function (Sec. 3.4). To further enhance accuracy, we propose a RANSAC-based algorithm for estimating camera pose from rays in Sec. 3.5.", + "bbox": [ + 511, + 657, + 908, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 818, + 718, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Employing CRR for VL involves two main steps: 1) regressing the patch-level camera ray parameters from the input image and 2) calculating the camera pose from the ray parameters. Given an input image $I \\in \\mathbb{R}^{H \\times W \\times C}$ , it is first", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25641", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "divided uniformly into a set of patches $(p_i)$ of size $s\\times s$ :", + "bbox": [ + 89, + 90, + 465, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{p _ {i} \\in \\mathbb {R} ^ {s \\times s \\times C} \\right\\} _ {i = 1} ^ {N}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 116, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N = HW / s^2$ . Subsequently, the model $\\mathcal{F}$ regresses the ray parameters corresponding to the center points of image patches, expressed in the world coordinate system, $\\hat{\\mathcal{R}} = \\{r_i\\in \\mathbb{R}^6\\}_{i = 1}^N$", + "bbox": [ + 89, + 146, + 483, + 209 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {R}} = \\mathcal {F} (I), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 222, + 482, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the ray parameters are expressed in Plücker coordinates [12]: $r_i = [d_i, m_i]$ , where $d_i \\in \\mathbb{R}^3$ denotes the ray direction and $m_i \\in \\mathbb{R}^3$ represents the ray moment. Upon obtaining sufficient rays ( $\\geq 2$ ), the camera pose can be acquired by solving two distinct linear problems corresponding to rotation and translation, respectively [44].", + "bbox": [ + 89, + 248, + 483, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, in terms of rotation, given the camera intrinsic, the Plücker coordinates of all camera rays in the camera coordinate system can be achieved, denoted as $\\{r_i^c = [d_i^c, 0]\\}_i^N$ . The rotation matrix $R$ , thus, corresponds to the transformation from the ray directions in the camera coordinate system to those in the world coordinate system:", + "bbox": [ + 89, + 339, + 483, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} = \\arg \\min _ {| | R | | = 1} \\sum_ {i} \\left| \\left| R d _ {i} ^ {c} - d _ {i} \\right| \\right| _ {2} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 441, + 482, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the translation, also known as the world position of the camera optical center denoted as $c$ , is determined. Since $c$ is the intersection point of all camera rays, the optimization problem can be formulated as:", + "bbox": [ + 89, + 479, + 483, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} = \\arg \\min _ {c} \\sum_ {i} \\| c \\times d _ {i} - m _ {i} \\| _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 553, + 482, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through these two linear problems, the camera pose $([\\hat{R} |\\hat{c} ])$ can be deduced from the camera rays.", + "bbox": [ + 89, + 594, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Feature Encoder", + "text_level": 1, + "bbox": [ + 89, + 635, + 256, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the patch-wise independent GDT, it entails that the scene-specific head must learn the mapping from patch-level features to ray parameters. To ensure the learnability of this, ambiguity within patch features, particularly the local ambiguities, needs to be addressed. Thus, based on our prior analyze, patch-level features must possess both robust local representation and image-level perception. Hence, we opt for DINO [26] as our encoder backbone. From an input image $I$ , it can generate patch-level feature tokens, $\\{p_i\\}_i^N$ , and provide the CLS token for image-level perception. According to prior work [13], fine-tuning DINO's last two blocks can effectively enhance the performance of VL tasks. Consequently, we also fine-tune DINO using the ACE Encoder training method [4], resulting in $\\mathcal{D}$ .", + "bbox": [ + 89, + 657, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {T} _ {i} \\right\\} _ {i} ^ {N}, \\mathcal {T} _ {[ \\mathrm {C L S} ]} = \\mathcal {D} (I), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 RC-RANSAC Algorithm", + "text_level": 1, + "bbox": [ + 514, + 90, + 774, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: $\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i = 1}^N$ # camera rays from DIMM", + "bbox": [ + 514, + 109, + 875, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: $\\hat{R} \\in \\mathbb{R}^{3 \\times 3}$ , $\\hat{c} \\in \\mathbb{R}^{3 \\times 1}$ ; # camera pose", + "bbox": [ + 514, + 126, + 823, + 142 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Get Rotation by RANSAC: $\\hat{R} \\gets R$ -RANSAC( $\\{\\hat{d}_i\\}_{i=1}^N$ ); cf. Algorithm 2 of the Suppl.", + "2: Ray Correction: $\\hat{\\mathcal{R}}\\gets \\{[\\hat{R}\\hat{d}_i^c,\\hat{m}_i]\\}_{i = 1}^N;$", + "3: Get Camera Center by RANSAC: $\\hat{c} \\gets C$ -RANSAC( $\\hat{\\mathcal{R}}$ ); cf. Algorithm 3 of the Suppl." + ], + "bbox": [ + 521, + 143, + 906, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{T}_i\\in \\mathbb{R}^D$ represents the patch token corresponding to $p_i$ locally, $\\mathcal{T}_{[\\mathrm{CLS}]}\\in \\mathbb{R}^{D}$ denotes the CLS token with global perception, $N$ signifies the number of patches, and $D$ is the token dimension.", + "bbox": [ + 511, + 251, + 903, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, the geometric constraints of the rays from the same camera are crucial for eliminating local ambiguities. Thus, the network must also explicitly specify the spatial image location of patches during training. To further prevent the positional feature from being overshadowed by other dimensions, we draw inspiration from Nerf [24] and enhance this feature using Fourier Encoding $(FE)$ [40].", + "bbox": [ + 511, + 311, + 905, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\mathrm {x y} _ {i}} = F E \\left(\\left\\{x _ {i}, y _ {i} \\right\\} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times D _ {F}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 426, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\{x_{i},y_{i}\\}_{i}$ represents the positions of image patches $\\{p_i\\}_i$ in the original image, and $D_F$ is the dimension of the Fourier encoding, set to 16. Finally, the scene-independent patch-level feature is a combination of the aforementioned three features. For each patch $p_i$ , we have:", + "bbox": [ + 511, + 455, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {i} ^ {*} := \\mathcal {T} _ {i} \\otimes \\mathcal {T} _ {[ C L S ]} \\otimes \\mathcal {T} _ {\\mathrm {x y} _ {i}} \\in \\mathbb {R} ^ {2 D + D _ {F}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 542, + 906, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\otimes$ denotes concatenation along the last dimension. This carefully designed feature to address ambiguity serves as the foundation of our CRR method.", + "bbox": [ + 511, + 566, + 906, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mapper Head", + "text_level": 1, + "bbox": [ + 513, + 622, + 658, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The scene-specific mapper head is responsible for transforming input features to ray parameters in CRR. A powerful feature encoder allows similar transformations to be achieved with simple MLPs [4, 14, 22, 41]. However, these mappers struggles to handle large scenes due to its limited capacity. To address the issue, spatial clustering is applied to divide the large scene [4, 41], with multiple MLPs fitting different sub-scenes independently. While effective, this hard scene partitioning can lead to ambiguous results near the sub-scene boundaries. In contrast, we adopt a soft map partitioning. As its core, we fuse the outputs of multiple sub-mappers with soft semantic guidance. Through end-to-end training, the sub-mappers are able to adaptively focus on different semantic regions. Particularly, we employ two MLP projection layers $(P_{s}$ and $P_{g})$ to first project the input feature into a geometric embedding $\\mathcal{T}_i^g$ and a semantic embedding $\\mathcal{T}_i^s$ , with $D_h$ hidden dimension respectively.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25642", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/76eb49d63a0a98a85ebe3774d2a05a1be9cf5b03e4e02a25ced85bc57b3cfacc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodMapping w/ 3D info.Mapping TimeMapping Size
FM-basedhloc(SP+SG)No1.5 hours~2GB
SCRDSAC*Yes15 hours28MB
ACENo5 minutes4MB
ACE (×4)No20 minutes14MB
EGFSNo12 minutes4.5MB
EGFS (dual)No21 minutes9MB
D2SYes~9.4 hours22 MB
APRmarepoNo5 minutes98.9MB
mareporeplasNo~15 minutes98.9MB
CRRDIMM (Ours)No~1 hour16.2 MB
", + "bbox": [ + 94, + 88, + 482, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. The mapping size and time comparison between VL methods from different categories. As a new flavor of VL, our DIMM achieves mapping time and memory size comparable to other methods while ensuring privacy protection.", + "bbox": [ + 89, + 231, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {g} = P _ {g} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 306, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} ^ {s} = P _ {s} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 349, + 483, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we omit the indices of the patches for simplification. Next, the sub-mappers $\\mathcal{M} = \\{\\mathcal{M}_j\\}_{j=1}^M$ , composed of $M$ MLPs with depth $D_M$ , take the geometric embedding $\\mathcal{T}^g$ to obtain $M$ ray embeddings $\\{\\mathcal{T}_j^r\\}_{j=1}^M$ .", + "bbox": [ + 89, + 376, + 483, + 440 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {j} ^ {r} = \\mathcal {M} _ {j} \\left(\\mathcal {T} ^ {g}\\right) \\in \\mathbb {R} ^ {D _ {h}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 453, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, in the sequence of L cross attention layers $(\\mathcal{A} = \\{\\mathcal{A}_l\\}_{l=1}^L$ , termed as Semantic Attention), we replicate the semantic embedding $(\\mathcal{T}_i^s)$ $M$ times along the first dimension to serve as the query $q$ .", + "bbox": [ + 89, + 479, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\otimes \\left\\{\\mathcal {T} ^ {s} \\right\\} ^ {M} \\mapsto q \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 553, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All ray embeddings are concatenated along the first dimension to form both the key and value of the first layer.", + "bbox": [ + 89, + 585, + 485, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\otimes \\left\\{\\mathcal {T} _ {j} ^ {r} \\right\\} _ {j} ^ {M} \\mapsto k _ {0}, v _ {0} \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 631, + 483, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By iteratively updating the key and value through the attention layers, we fuse the results of all sub-mappers by weighting them using the semantic guidance from DINO.", + "bbox": [ + 89, + 657, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {l} \\left(q, k _ {l}, v _ {l}\\right) \\mapsto v _ {l + 1}, k _ {l + 1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 718, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final value is first reshaped: $v_{L} \\in \\mathbb{R}^{M \\times D_{h}} \\to \\mathbb{R}^{MD_{h}}$ , followed by projection to obtain the ultimate ray embedding: $\\mathcal{T}_f^r = P_f(v_L) \\in \\mathbb{R}^{D_h}$ . Finally, the ray parameters $\\hat{r} = [\\hat{d},\\hat{m}]$ are obtained through a straightforward MLP projection layer.", + "bbox": [ + 89, + 746, + 483, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {r} = P _ {r} \\left(\\mathcal {T} _ {f} ^ {r}\\right) \\in \\mathbb {R} ^ {6}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 839, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this head, we can achieve the corresponding camera ray from a patch-level embedding from our feature encoder.", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 513, + 90, + 656, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adopting the patch-level Gradient Decorrelation Training, our loss function operates independently between patches. First, we utilize the L2 distance between the ground truth $(r_{GT})$ and predicted rays as the optimization target.", + "bbox": [ + 511, + 112, + 906, + 174 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l 2} = \\left\\| r _ {G T} - \\hat {r} \\right\\| _ {2} ^ {2}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 181, + 906, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, to enhance the model awareness of geometric constraints, we incorporate a center loss $\\mathcal{L}_c$ based on the geometric properties of camera rays. The principle is any camera ray passes through a corresponding camera center [12].", + "bbox": [ + 511, + 210, + 906, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\left\\| c \\times \\hat {d} - \\hat {m} \\right\\| _ {2} ^ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 279, + 906, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, our patch-level loss function is $\\mathcal{L} = \\mathcal{L}_{l2} + \\alpha \\mathcal{L}_c$ , where $\\alpha$ is a balance term, which is empirically set as 0.2. Our training loss is simple yet effective compared to SCR methods [4, 14, 22], benefiting from the robustness of rays in optimization.", + "bbox": [ + 511, + 306, + 906, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. RC-RANSAC", + "text_level": 1, + "bbox": [ + 513, + 391, + 656, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to SCR, it is also not guaranteed that all camera rays can be accurately predicted in CRR. Therefore, employing the RANSAC mechanism to eliminate outliers is a beneficial approach. Unlike the PnP solution, when solving for camera pose from rays, the rotation and translation constitute two linear problems. Hence, the RANSAC process can be separately applied. Given that the rotation solution is independent of translation, we introduce the RC-RANSAC (Algorithm 1) as a modification of the general RANSAC algorithm. Specifically, after obtaining the predicted camera rays $\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i}^{N}$ , we initially use $R$ -RANSAC (cf. Algorithm 2 of the Suppl.) to determine the rotation $\\hat{R}$ from the ray directions only. Subsequently, all camera rays are corrected using the obtained rotation matrix $\\hat{R}$ to get better rays: $\\hat{\\mathcal{R}}^* = \\{[\\hat{R} d_i^c,\\hat{m}_i]\\}_{i}$ . The camera translation $\\hat{c}$ is then derived from the corrected rays by another $C$ -RANSAC process (cf. Algorithm 3 of the Suppl.). Ultimately, combining the rotation and translation yields the final camera pose: $[\\hat{R},\\hat{c} ]$ for VL.", + "bbox": [ + 511, + 412, + 908, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 715, + 645, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 739, + 730, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network Configuration. For the feature encoder $\\mathcal{D}$ , we use the dinov2_vitb14_reg as our backbone, with an embedding dimension of 768 ( $D = 768$ ). The frequency of Fourier Encoding is set as 16, making $D_F = 64$ . On the other hand, the mapper head consists of 4 sub-mappers and a semantic attention model with 4 attention layers. Each sub-mapper has a 4-depth MLP with skip connections. The first 2 layers form the common mapping, and the other 2 layers output direction and moment embedding respectively", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25643", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/10c381c461516ed3021300597feaa254733a89ace05959c76eaf9789d706343a.jpg", + "table_caption": [ + "Indoor-6" + ], + "table_footnote": [], + "table_body": "
CategoryMethodscene1 (cm/°)scene2a (%)scene3 (cm/°)scene4a (%)scene5 (cm/°)scene6 (cm/°)Average (cm/°)
FM-basedHloc [30]3.2/0.564.8-/-51.42.1/0.481.02.9/0.5
Hloc+SLD [9]2.9/0.468.73.4/0.662.71.9/0.381.042.7
2.1/0.4
APRPoseNet [15]159.0/7.50.0-/--141.0/9.30.0118.2/9.3
SCRDSAC* [3]12.3/2.118.77.9/0.928.013.1/2.319.740.7/6.7
ACE [4]13.6/2.124.96.8/0.731.98.1/1.333.014.7/2.3
NBE+SLD(E) [9]7.5/1.228.47.3/0.730.46.2/1.343.56.3/1.0
NBE+SLD [9]6.5/0.938.47.2/0.732.74.4/0.953.06.0/0.9
EGFS [20]-/-46.4-/-60.6-/-56.4-/-
EGFS-q [20]-/-58.5-/-59.1-/-67.0-/-
D2S [5]4.8/0.851.84.0/0.461.13.6/0.760.05.8/0.9
2.4/0.5
CRRDIMM5.5/1.544.44.2/0.866.74.4/1.161.67.6/1.5
DIMM-R5.5/1.545.23.3/0.677.44.2/1.162.56.3/1.5
3.2/0.8
", + "bbox": [ + 91, + 104, + 903, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Localization results on Indoor-6 [9]. We report the median errors in $cm$ for the position, degree $(^{\\circ})$ for the orientation, and recall at $5cm / 5^{\\circ}$ of the Indoor-6 dataset. The best results, the second best and our results are highlighted, except for the FM-based method.", + "bbox": [ + 91, + 297, + 903, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1e2f8c023fb91cac51d7cd5b299efbfc91ce6f72b472aa92a8b19b794d461d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 327, + 359, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98f2f06b7a021c1bffbc2ab4c799d93f723f58fb0808baf763710d83d418992d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 327, + 513, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb74bb96a190dcec59e4d02dccefc5aac2a2f6d1118c2ea2b8a36b1a8e178d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 327, + 665, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/907551be19592bf81aabe16194e29e0d4cdc033081ce81d93199b375ccf0373b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 327, + 816, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be3d98060c8717ce5516bf203bb186714dcc5f8736a714c988303e5042231639.jpg", + "image_caption": [ + "Figure 4. The qualitative comparison between ray regression w/ or w/o global perception in the scene1 of Indoor-6. These results contain predicted and ground truth rays from the same cameras, along with their camera centers. The samples indicate that ray predictions w/o global perception (without the CLS and XY tokens, bottom) lack geometric consistency, although some individual rays are accurate. In contrast, incorporating global perception facilitates learning the geometric constraint of rays, resulting in precise ray regression (top)." + ], + "image_footnote": [], + "bbox": [ + 176, + 444, + 359, + 563 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed54ac662cbd8a7ef5e89ae8140dbacde14ec2b767e69b37dc4ab22c136d669a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 448, + 513, + 563 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d93c7a76b63a7177a208b780b57549b9e4158c7eae2a556eca16418ec7cf8b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 448, + 665, + 563 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf7af604d1a6e14d37c2fa0a6969baed6b39d7a78621881208a1c4dba886d8be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 448, + 816, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(cf. Fig. 3). The hidden dimension $D_{h}$ is set as 256. More ablation experiments can be found in Sec. 4.3.", + "bbox": [ + 89, + 637, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Hardware Details. Firstly, according to related work [13], we fine-tune the last two blocks of DINO in terms of CRR task for better performance. Particularly, we follow the encoder training protocol of [4] and [13], using the first 100 training scenes of ScanNet [7] to update the weights of unfrozen DINO blocks. The training is performed on 4 NVIDIA A800 GPUs using gradient accumulation [4] in parallel. Next, the mapper head is trained on 1 NVIDIA A800 GPU for each scene, with a training buffer of 26M encoder features. The large buffer size is important for CRR training, as most of patches have different rays. We observed a buffer size scaling law in experiments (cf. Suppl. B). We do 20 passes over the training buffer, utilizing a batch size of 5120. The optimization uses AdamW [21] with a learning rate between $5e^{-5}$ and $1e^{-4}$ and a 1 cycle schedule [37].", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 637, + 735, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate VL performance of our method on both indoor and outdoor datasets. For indoor scenes, we adopt the Indoor-6 dataset [9], which is a relatively large indoor dataset [9]. It comprises calibrated images from 6 multi-rooms indoor scenes captured over several days, leading to challenging VL. For outdoor scenes, the Map-Free [1] dataset is employed. We use its first 10 scenes, $s00000 \\sim s00009$ , following [4]. Each scene contains images from two scans of an outdoor location. We split them into mapping and query sets. The ground truth camera poses are computed via SfM method [34].", + "bbox": [ + 511, + 666, + 906, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Indoor Visual Localization. Firstly, the indoor VL experiments are conducted using the Indoor-6 dataset. In the experiments, we compare our DIMM method and its variation enhanced by RANSAC (DIMM-R) with both APR [15]", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25644", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43f4b3597abcf4d8f36ee1f6abd42b8c1385f4cdff5d963f122cda8cb870fa72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scene (°/cm/%)SCRAPRCRR
DSAC*[3]ACE[4]marepo[6]marepos[6]DIMMDIMM-R
Throughput (fps)17.917.955.655.657.720.1
s000001.2/5.1/55.80.7/4.2/62.60.6/3.6/69.20.6/3.2/70.70.7/2.0/93.10.6/2.0/94.0
s000010.8/2.4/94.50.4/1.5/1000.9/1.8/96.00.8/1.6/99.21.1/1.6/96.81.1/1.6/96.8
s000020.8/2.4/69.50.5/2.1/77.70.6/2.3/77.10.5/2.1/78.20.9/3.7/71.20.9/3.0/77.1
s000030.6/2.1/75.80.4/2.4/73.60.9/2.5/80.30.8/2.1/84.50.6/2.4/92.80.6/2.2/93.3
s000041.1/6.3/40.40.7/5.8/45.30.8/5.3/45.90.7/5.1/47.60.7/2.4/82.60.7/2.3/96.2
s000050.7/3.6/64.40.5/2.4/73.30.7/2.7/72.20.5/2.6/77.30.8/2.5/86.20.8/2.1/87.9
s000061.0/4.4/53.40.8/4.5/54.80.8/4.2/58.50.6/3.6/63.80.8/5.6/44.30.8/3.9/64.8
s000070.7/7.8/16.90.6/7.6/19.01.0/6.3/25.10.8/6.1/27.50.6/3.6/69.50.5/3.5/71.1
s000081.2/3.5/68.90.7/3.1/71.60.7/2.3/77.40.7/1.9/82.80.7/2.7/85.10.7/2.4/87.9
s000090.8/3.1/81.20.4/1.2/99.70.8/1.8/94.90.9/1.6/95.10.7/1.3/97.70.7/1.2/97.4
Average0.8/4.1/62.10.6/3.5/67.70.8/3.3/69.70.7/3.0/72.70.8/2.8/81.90.7/2.4/86.6
", + "bbox": [ + 192, + 88, + 803, + 290 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c88a6839f0c5812c4256b76c42d57858b378a3ec3f9f759b6020405a2ef9e0f7.jpg", + "table_caption": [ + "Table 3. Pose accuracy comparison on the outdoor MapFree [1] dataset. We report the median errors in degree (°) for the orientation, cm for the position, and recall at $5\\mathrm{cm} / 5^{\\circ}$ . The best results, the second best and our results are highlighted." + ], + "table_footnote": [], + "table_body": "
Recall-5°/5cm (%)s00000s00001s00002s00003s00004s00005Average
ACE [4]62.610077.773.645.373.372.1
DIMM-CRR (Ours)93.196.871.292.882.686.287.1
DIMM-SCR64.898.674.577.348.272.172.6
DIMM-APR3.212.45.73.60.32.64.6
", + "bbox": [ + 91, + 337, + 480, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and SCR [3-5, 9, 20] approaches. The results are reported in Tab. 2. Additionally, we analyzed the corresponding mapping time and memory costs of various VL methods in Tab. 1, for a comprehensive comparison. DIMM demonstrates better or comparable performance than the previous state-of-the-art, e.g., achieving a recall@5°/5cm of 78.3 in scene6. Moreover, our method achieves generally better results than EGFS-q with \"hard\" multi-mapper combination, proving the efficacy of our \"soft\" mapper ensemble utilizing semantic guidance. The effectiveness of RC-RANSAC is also highlighted by a substantial precision improvement in scene2a from 66.7 to 77.4. Overall, the proposed DIMM-R achieves the best average recall in this dataset. Furthermore, our method enables mapping with reduced computational costs compared to D2S and is trained without 3D supervision, as illustrated in Tab. 1. These findings underscore the potential of our CRR-based method as a promising VL implementation. Some visualization results of DIMM can be found in Fig. 4 and Fig. 7 in the Suppl., where the 3D scene structure is preserved, demonstrating the privacy protection offered by our method.", + "bbox": [ + 88, + 458, + 483, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Outdoor Visual Localization. Our outdoor VL experiments are conducted on the Map-Free Dataset. In addition to the SCR-based ACE and DSAC*, we also compared the recent APR-based method, marepo [6], and its fine-tuned variation, marepos. The results, including median errors for rotation $(^{\\circ})$ and translation $(cm)$ , as well as localization recalls at $5\\mathrm{cm} / 5^{\\circ}$ , are presented in Tab. 3. These results also come from methods with mapping time in Tab. 1. From the table,", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36c54f60241000a11f97dd9378bb29fa63b7c54bcb286f65fa1fcf3a1fe47ba1.jpg", + "table_caption": [ + "Table 4. Ablation Results about Camera Ray Regression (CRR). We evaluate three variations of our model to showcase the effectiveness of CRR, by modifying output to different VL formulations." + ], + "table_footnote": [], + "table_body": "
Feature EncoderMed. R Err. (°)Med. t Err. (cm)Recall-5°/5cm
T_i ∩ T[CLS] ∩ Txy_i0.85.644.3
T_i * T*[CLS] ∩ Txy_i0.86.039.7
T_i ∩ TG[41] ∩ Txy_i1.18.319.7
T_i ∩ T[CLS] ∩ XY_i0.97.627.2
T_i ∩ T[CLS]0.97.426.8
T_i1.111.517.2
", + "bbox": [ + 555, + 337, + 867, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation Study about our Feature Encoder. The experiments are performed on the MapFree s00006 dataset.", + "bbox": [ + 511, + 422, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "it is evident that our methods achieve the best overall results, outperforming both SCR and APR methods. The rotation accuracy of ACE, yet, surpasses our methods, indicating that local ambiguity poses a significant challenge for CRR methods and obtaining precise ray directions from image patches is challenging. However, it is observed that CRR methods facilitate better camera center estimation compared to APR and SCR methods. This improvement can be attributed to the fact that the camera center location integrates contributions from all camera rays, and the bounded ray error [27] enhances the robustness of the center estimation against noise. The efficacy of RC-RANSAC algorithm can also be verified in the table, as it leads to substantial improvements of VL performance, e.g., $44.3 \\rightarrow 64.8$ of recall in s00006. Furthermore, we investigate the computational efficiency of our methods. The throughputs (fps) of all methods are reported in the table, revealing that the lightweight network and linear pose solver of DIMM achieve the best inference efficiency.", + "bbox": [ + 511, + 454, + 908, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Experiments", + "text_level": 1, + "bbox": [ + 511, + 739, + 718, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera Ray Regression. One of our main contributions in this work is introducing Camera Ray Regression (CRR) into the VL task. Thus, we provide ablation study about this formulation in Tab. 4. In particular, we modify the last layer of DIMM to output scene coordinates (DIMM-SCR) or direct camera pose (DIMM-APR), and train these variations separately. The DIMM-SCR is trained with GDT, same as ACE[4]. The DIMM-APR, on the other hand, is trained in image level, as the direct pose vectors cannot", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25645", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3713c77838c9d31b05a8d92c2320b1795b2ee1aa602394120bf515105e0f23f7.jpg", + "image_caption": [ + "Figure 5. Visualization results of sub-mapper attention score. In the experiments, we observed that specific semantic entities elicit responses from different mappers. For instance, the fireplace at the top of the figure frequently results in high attention scores from mapper #3, while mapper #2 exhibits a low attention score. Conversely, mapper #3 seems do not \"care\" the sofa at the bottom, which leads to high scores of mapper #2 instead." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 460, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "be assigned to image patches. The experiments are performed on the MapFree dataset, with the localization recall reported. In Tab. 4, we can see that our CRR formulation significantly outperforms the SCR and APR under the same network architecture, indicating the contribution of our ray-parameterization. DIMM-SCR overcomes ACE slightly in average, but our model is tailored for CRR. DIMM-APR utilizes per-scene pose regression and achieves poor results, but this is consistent with the results in marepo [6]. This suggests the APR model may require much larger datasets to learn pose patterns. Overall, the results in Tab. 4 evaluate our main contribution: CRR formulation in VL.", + "bbox": [ + 89, + 380, + 483, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Encoder. Based on our analysis of CRR, we require patch features to exhibit strong local description while also incorporating global perception and image position encoding to achieve disambiguation. To validate this, we conduct experiments on the MapFree s00006 dataset (cf. Tab. 5). Specifically, we modify the content of the patch feature, including features w/o global perception and image position encoding $(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}$ and $\\mathcal{T}_i)$ , to compare with the proposed version $(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathcal{T}_{\\mathrm{xy}_i})$ . Also, we investigate the efficacy of DINO fine-tuning, and the performance of original DINO $(\\mathcal{T}_i^* \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}^* \\otimes \\mathcal{T}_{\\mathrm{xy}_i})$ is also reported. It is evident that global perception with image position encoding are crucial for CRR, leading to a significant accuracy improvement. The fine-tuning of DINO also improves the performance as well (39.7 vs. 44.3). Visualization comparison about the global perception is presented in Fig. 4. We can see the proposed global perception enables the learning of ray geometry, causing the rays to converge toward the camera center, thereby enhancing precision. We also conduct experiments for a variation without Fourier Encoding $(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathsf{XY}_i)$ . The results in the first and third rows of Tab. 5 show that Fourier Encoding effectively prevents the submergence of positional", + "bbox": [ + 89, + 568, + 486, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa4a5dae98b98a091af77d93805cb15475bd18c246465da52b0e68182e17790a.jpg", + "table_caption": [], + "table_footnote": [ + "1 Take average of the sub-mapper embeddings, eliminating attention layers." + ], + "table_body": "
Atten. Num.Med. R Err. (°)Med. t Err. (cm)Recall@5°/Scmsize (MB)
011.07.231.012.1
20.95.842.614.2
40.85.644.316.2
60.86.337.718.2
81.17.730.420.2
", + "bbox": [ + 537, + 89, + 885, + 155 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation Study about the Number of Semantic Attention Layers. The experiments are performed on the MapFree s00006 dataset. The efficacy of our semantic attention is evaluated.", + "bbox": [ + 511, + 174, + 906, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features. Finally, we explore another possible global perception way provided by [41] $(\\mathcal{T}_i \\otimes \\mathcal{T}_G \\otimes \\mathcal{T}_{xy_i})$ , the feature for image retrieval. The results show that the replaced global feature fails to match the original efficacy, possibly because DINO's global feature is more compatible with its own local patch features, facilitating subsequent mapping learning.", + "bbox": [ + 511, + 228, + 906, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mapper Head. In the same scene, we also conduct ablation experiments on the network structure of the Mapper Head. We first validate the efficacy of the semantic attention module and experiment with its number of layers. When the layer number is set to 0, we directly take the average of output from multiple sub-mappers. The results are presented in Tab. 6, containing median rotation/translation error, localization recall and model size. It is evident that utilizing semantic attention is more effective compared to simple average fusion, and the increase in network size is deemed acceptable. Additional ablation study about the MLP mappers can be found in Suppl. B. To showcase the effectiveness of the soft mapper assemble, we visualize some query images combined with their sub-mapper attention scores in Fig. 5. It can be observed that different sub-mappers get high attention scores to different objects in images. This indicates that our semantic attention mechanism allows the sub-mappers to \"remember\" specific semantic objects.", + "bbox": [ + 511, + 325, + 908, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 626, + 633, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose regressing camera rays to perform visual localization. This overparameterized representation of camera model leads to high precision, and, more importantly, enhances privacy protection. In particular, we introduce DIMM as a learning model to regress camera rays from images. It utilizes DINO as a scene-agnostic encoder to output features that incorporate both local and global perception. A scene-specific mapper then regresses ray parameters from the features, involving a semantic attention module to merge results from multiple mappers. Finally, a ray-level RANSAC algorithm is proposed to improve the ray-to-pose accuracy. In experiments on both indoor and outdoor datasets, our methods demonstrate comparable or superior performance to current approaches, while ensuring privacy preservation. Acknowledgement. This work has been funded in part by the NSFC grants 62176156.", + "bbox": [ + 511, + 656, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25646", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Áron Monszpart, Victor Adrian Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In ECCV, 2022. 6, 7", + "[2] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7525-7534, 2019. 1", + "[3] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE transactions on pattern analysis and machine intelligence, 44 (9):5847-5865, 2021. 3, 6, 7", + "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7", + "[5] Bach-Thuan Bui, Huy-Hoang Bui, Dinh-Tuan Tran, and JooHo Lee. D2s: Representing sparse descriptors and 3d coordinates for camera relocalization. IEEE Robotics and Automation Letters, 2024. 1, 2, 3, 6, 7", + "[6] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1, 3, 7, 8", + "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6", + "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2", + "[9] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 6, 7", + "[10] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 3", + "[11] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2025. 1, 3", + "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 3, 4, 5", + "[13] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17658-17668, 2023. 4, 6" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xudong Jiang, Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. R-score: Revisiting scene coordinate regression for robust large-scale visual localization. arXiv preprint arXiv:2501.01421, 2025. 3, 4, 5", + "[15] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In 2016 IEEE international conference on Robotics and Automation (ICRA), pages 4762-4769. IEEE, 2016. 1, 3, 6, 13", + "[16] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5974-5983, 2017.", + "[17] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 3", + "[18] Sihang Li, Siqi Tan, Bowen Chang, Jing Zhang, Chen Feng, and Yiming Li. Unleashing the power of data synthesis in visual localization. arXiv preprint arXiv:2412.00138, 2024. 13", + "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 3", + "[20] Ting-Ru Liu, Hsuan-Kung Yang, Jou-Min Liu, Chun-Wei Huang, Tsung-Chih Chiang, Quan Kong, Norimasa Kobori, and Chun-Yi Lee. Reprojection errors as prompts for efficient scene coordinate regression. In European Conference on Computer Vision, pages 286-302. Springer, 2025. 1, 2, 3, 6, 7", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6", + "[22] Deben Lu, Wendong Xiao, Teng Ran, Liang Yuan, Kai Lv, and Jianbo Zhang. Attention-based accelerated coordinate encoding network for visual relocalization. In 2024 IEEE 7th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), pages 1675-1680, 2024. 1, 3, 4, 5", + "[23] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 2, 3", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4", + "[25] Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. arXiv preprint arXiv:2412.12392, 2024. 3", + "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DiNov2: Learning robust visual features without supervision, 2023. 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25647", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schonberger. Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV), 2024. 2, 3, 7", + "[28] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 1, 3", + "[29] Francesco Pittaluga, Sanjeev J. Koppal, Sing Bing Kang, and Sudipta N. Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[30] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1, 2, 6", + "[31] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1", + "[32] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 752-765. Springer, 2012. 2", + "[33] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE transactions on pattern analysis and machine intelligence, 39(9):1744-1756, 2016. 2", + "[34] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6", + "[35] Thomas Schops, Viktor Larsson, Marc Pollefeys, and Torsten Sattler. Why having 10,000 parameters in your camera model is better than twelve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2544, 2020. 1, 2, 3", + "[36] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 3", + "[37] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of residual networks using large learning rates. Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, 2019. 6", + "[38] Pablo Speciale, Johannes L. Schonberger, Sing Bing Kang, Sudipta N. Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5488-5498, 2019. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 1, 3", + "[40] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 4", + "[41] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 3, 4, 7, 8", + "[42] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3", + "[43] Yinshuang Xu, Dian Chen, Katherine Liu, Sergey Zakharov, Rares Andrei Ambrus, Kostas Daniilidis, and Vitor Campagnolo Guizilini. $se(3)$ equivariant ray embeddings for implicit multi-view depth estimation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 3", + "[44] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4", + "[45] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20217-20226, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 570 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25648", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_model.json b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94c672dfa7c7d59ccffcdd02fbaca1280026e03e --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.828, + 0.153 + ], + "angle": 0, + "content": "Semantic-guided Camera Ray Regression for Visual Localization" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.182, + 0.806, + 0.234 + ], + "angle": 0, + "content": "Yesheng Zhang, Xu Zhao* \nSchool of Automation and Intelligent Sensing, Shanghai Jiao Tong University \n{preacher, zhaoxu}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.59 + ], + "angle": 0, + "content": "This work presents a novel framework for Visual Localization (VL), that is, regressing camera rays from query images to derive camera poses. As an overparameterized representation of the camera pose, camera rays possess superior robustness in optimization. Of particular importance, Camera Ray Regression (CRR) is privacy-preserving, rendering it a viable VL approach for real-world applications. Thus, we introduce DINO-based Multi-Mappers, coined DIMM, to achieve VL by CRR. DIMM utilizes DINO as a scene-agnostic encoder to obtain powerful features from images. To mitigate ambiguity, the features integrate both local and global perception, as well as potential geometric constraint. Then, a scene-specific mapper head regresses camera rays from these features. It incorporates a semantic attention module for soft fusion of multiple mappers, utilizing the rich semantic information in DINO features. In extensive experiments on both indoor and outdoor datasets, our methods showcase impressive performance, revealing a promising direction for advancements in VL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.222, + 0.636 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.485, + 0.768 + ], + "angle": 0, + "content": "Visual Localization (VL) involves estimating 6-degree-of-freedom camera pose of an image taken from a known scene, which is also known as camera relocalization. VL plays a vital role in many vision, robotic and graphic applications, such as augment reality, virtual reality and autonomous driving. Despite extensive research, achieving precise, robust, and privacy-preserving VL [28, 38] continues to present a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Current VL methods can be categorized into explicit and implicit methods based on their mapping approaches. Traditional explicit methods [2, 30, 31] involve an explicit 3D map, e.g., point clouds. They establish 2D-3D (image-to-map) correspondences by Feature Matching, and then estimate camera pose via PnP and RANSAC [10, 12]. While advanced methods [11, 39] leads to high accuracy, the stor" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.269, + 0.825, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.411, + 0.908, + 0.523 + ], + "angle": 0, + "content": "Figure 1. The ambiguity challenge in Camera Ray Regression (CRR) for VL. Effective global and local disambiguations are critical for CRR. Global ambiguity arises from rays of different patches (cf. \\( r_2^1 \\) and \\( r_2^2 \\) in the figure), while local ambiguity occurs within the same patch but from different viewing angles (cf. \\( r_1^1 \\) and \\( r_2^2 \\)). Semantic information, e.g., the table in the figure to differentiate between chairs, along with geometry constraints (geometry properties of rays \\( \\{r_1^i\\}_i \\)) are essential for the disambiguations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.907, + 0.58 + ], + "angle": 0, + "content": "age demands of explicit maps are troublesome especially in large-scale scenes. Additionally, explicit maps raise privacy concerns, restricting their applications in real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.703 + ], + "angle": 0, + "content": "Implicit methods, on the other hand, leverage learning models as implicit and light-weight maps, which mainly include two frameworks: Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR). APR directly utilizes networks to regress camera poses (rotations and translations) from images [6, 15], offering enhanced privacy protection. However, the compact nature of absolute pose makes it vulnerable to noise, thereby leading to accuracy issue [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To improve precision, employing an overparameterized representation for the camera pose is crucial [35]. Thus, SCR framework [22, 41] is proposed, relying on 2D-3D matches as an indirect yet robust camera pose representation, akin to its explicit counterparts. ACE [4] is a milestone within this framework, utilizing simple networks to regress pixel-aligned scene coordinates. It introduces a scene-centric patch-level training strategy, i.e., Gradient Decorrelation Training (GDT), enabling both efficient and precise mapping. Nevertheless, the unbounded 3D point error is unstable in training, necessitating manual hyperparameter tuning or additional outlier rejection modules [5, 20]. Besides, SCR methods still run the risk of 3D privacy breaches, as they" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.887, + 0.225, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.475, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.332 + ], + "angle": 0, + "content": "Figure 2. The proposed DIMM approach performing CRR for VL. This approach utilizes DINO as a scene-agnostic feature encoder to output carefully designed patch-level features. Then, a scene-specific mapper head integrates multi-mapper results under semantic guidance, through a semantic attention module. After the scene-centric training technology, DIMM is capable of predicting camera ray parameters for accurate and privacy-preserving VL." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.34, + 0.384, + 0.355 + ], + "angle": 0, + "content": "disclose detailed 3D scene information [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.644 + ], + "angle": 0, + "content": "Recently, a promising camera pose representation, camera ray [27, 35], has garnered increased attention in the vision field. Its overparameterization is beneficial for patch-level learning [23, 44], aligning well with the effective GDT in VL. This inspires us to introduce camera rays in the VL tasks. In particular, we propose adopting a learning model to regress patch-level camera rays from the image, termed as Camera Ray Regression (CRR). Then, the camera pose can be achieved by solving two linear problems [44]. This implementation of VL could strike a balance between efficacy and privacy preservation. Firstly, it belongs to the implicit mapping category, incurring low storage costs. As a overparameterized pose representation, it can yield improved precision [35] than APR. Also, the ray error is more stable than the unbounded 3D point error [27] in optimization of model training. Hence, CRR obviates the need for intricate training setting like SCR methods [5, 20, 41]. More importantly, it is hard to recover 3D scene information from camera rays, leading to a privacy-preserving VL implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, challenges exist in introducing camera rays in VL, primarily stemming from the issue of ambiguity (cf. Fig. 1). This ambiguity can generally be divided into two levels. The first is global ambiguity, referring to distinguish rays of different image patches, particularly in repetitive or textureless scenes. The similar ambiguity is encountered in SCR [4, 41], where leveraging powerful local features, such as semantic features [20], has proven effective. The second is local ambiguity, unique to CRR, which involves differentiating rays from the same image patch with different view angles. The ambiguity becomes troublesome for patches with continuous depth, where changes in view direction result in minor differences in patch content but obvious variations in the corresponding rays. To eliminate this ambiguity, we emphasize two observations on the CRR task. 1) The entire image often contains patches with depth discontinuities. Their appearance changes under different" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "viewing angles are significant enough to resolve the local ambiguity. Thus, the image-level perception is important for CRR disambiguation, which is also proven to be beneficial for handling global repetitiveness [41]. 2) The rays within the same image are governed by the same camera pose. This geometric constraint allows confident rays to correct uncertain ones. Meanwhile, the smoothness of the patchray mapping [44] makes this geometric constraint learnable. Therefore, the integration of image-level perception, geometric constraint and robust local features is critical for the global and local disambiguation of CRR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.909, + 0.576 + ], + "angle": 0, + "content": "Based on the above analysis, this work presents DINO-based Multi-Mappers (DIMM, Fig. 2), as the first implementation of CRR in VL. Following the scene-centric design of ACE, DIMM consists of a scene-agnostic encoder and a scene-specific head, but possessing several key modules tailed for CRR. Specifically, we utilize DINO [8] as the feature encoder, utilizing its powerful semantic perception for disambiguation. It provides robust patch tokens as local features while supporting CLS tokens for image-level perception. We further incorporate Fourier Encoding [24] of the image position for each patch into the features (XY tokens), as the potential geometric constraint of camera rays are associated with patch locations. Furthermore, we propose a scene-specific mapper head which adopts a semantic attention module to softly fuse ray regression results from multiple Multi-Layer Perceptron (MLP)-based mappers. This head effectively utilizes the inherit semantic prior of the DINO features, to enhance ray regression performance especially in large scenes. Finally, we propose a ray-level RANSAC algorithm for accurate pose estimation, decomposing the rotation and translation calculation [44]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.578, + 0.825, + 0.591 + ], + "angle": 0, + "content": "Our contributions are summaries as follows." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.598, + 0.905, + 0.643 + ], + "angle": 0, + "content": "1. To the best of our knowledge, we are the first to introduce the camera ray into VL, with a tailored network design to achieve query image poses by camera ray regression." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.643, + 0.905, + 0.702 + ], + "angle": 0, + "content": "2. We propose carefully designed scene-agnostic features to handle local and global disambiguation in CRR, along with a semantic-guided mapper head that enables a soft multi-mapper ensemble." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.703, + 0.907, + 0.779 + ], + "angle": 0, + "content": "3. Comprehensive experiments on indoor and outdoor datasets demonstrate the effectiveness of our method, achieving results that rival or surpass the existing state-of-the-art. Additionally, extensive ablation studies validate the contributions of each module." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.598, + 0.907, + 0.779 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.796, + 0.655, + 0.811 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Feature Matching-based VL. Traditional VL methods [30, 32, 33] rely on Feature Matching (FM) to determine the position and direction of cameras. These techniques aim to match 3D scene points with 2D image points, enabling the camera pose calculation through \\(\\mathrm{PnP} + \\mathrm{RANSAC}\\)-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.215, + 0.089, + 0.784, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.25, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 3. The Multi-Mapper network with semantic attention as our scene-specific head. The patch-level features are first projected to geometry and semantic embeddings. The geometry embedding is fed into multiple MLP-based mappers to derive ray embeddings. These embeddings are then fused under the guidance of the semantic embedding to achieve the camera rays, through cross attention modules." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.299, + 0.486, + 0.45 + ], + "angle": 0, + "content": "algorithms [10, 12]. As a result, these approaches necessitate the storage of explicit maps containing 3D points and their feature descriptors. By leveraging robust matching algorithms [11, 19, 39, 42, 45], FM-based localization can deliver precise outcomes. Nevertheless, explicit maps raises concerns about privacy leaks [28]. Furthermore, as the scene size expands, managing the increasingly large explicit maps becomes challenging, leading to storage issues. In contrast, our method employs implicit maps that have lower storage requirements and privacy protection." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.456, + 0.485, + 0.729 + ], + "angle": 0, + "content": "Scene Coordinate Regression. Similar to FM-based VL, Scene Coordinate Regression (SCR) also accomplishes camera poses through 3D-2D matching [3]. However, SCR directly employs network to regress pixel-aligned 3D scene coordinates, which addresses the map storage issue through implicit mapping. ACE [4], a prominent approach within SCR, introduces a gradient decorrelation training technique enabling mapping within 5 minutes using a 4MB network. Subsequent methods [5, 14, 20, 22, 41] built upon ACE further enhance VL accuracy mainly by solid feature selection. Nevertheless, the limited network size of ACE pose challenges in fitting large scenes, necessitating spatial segmentation of the large scene for sub-map construction. Despite its efficacy, this approach may result in ambiguous localization at the boundaries of segmented sub-scenes. At the same time, it is noteworthy that the SCR methods output the 3D information of scenes, i.e., 3D coordinates, thereby leading to a potential risk of privacy leak." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Absolute Pose Regression. Absolute Pose Regression (APR) also adopts implicit mapping, but directly utilizes networks to predict the rotation and translation of cameras. However, this compact parameter representation of camera pose is highly sensitive to noises [12, 44]. Regressing pose parameters from images, thus, is challenging [15-17, 36], and the accuracy is insufficient to compare with methods using over-parameterized representations, e.g., 2D-3D correspondences. Recent APR approach [6] opts to predict the camera pose based on the SCR output, resulting in favorable accuracy. However, this success is heavily reliant" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.908, + 0.36 + ], + "angle": 0, + "content": "on the precision of SCR methods. In contrast, our method employs camera rays to depict the camera pose. This overparameterized representation is more conducive to learning optimization [25, 27, 44], enabling superior accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.909, + 0.619 + ], + "angle": 0, + "content": "Camera Ray in Vision. Recently, camera rays have emerged as a promising camera parametrization, representing a generic camera configuration [35]. In recent work [23, 43, 44], camera rays have been proven to be well-suited for patch-level learning, with applications in multiple vision tasks. Also, in 3D reconstruction works [25, 27], camera rays yield more robust geometric optimization than classical correspondence-based representations, e.g., 3D-2D matches. This advantage stems from their overparameterized nature and constrained ray error. Likewise, our approach is the first to introduce camera rays in VL, performing Camera Ray Regression (CRR) to achieve camera poses from images. Its patch-level property aligns well with the Gradient Deceleration Training [4], and the bounded ray error is beneficial for optimization in VL Learning as well. Furthermore, the inherent privacy-preserving attribute of CRR positions it as a promising direction for VL implementations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.605, + 0.646 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.809 + ], + "angle": 0, + "content": "In this section, we first present the formulation of CRR for the VL task in Sec. 3.1. Then the proposed DIMM method is described in detail. DIMM adopts a scene-centric setup [4, 14, 41] by dividing the network into a scene-agnostic encoder (Sec. 3.2) and a scene-specific head (Sec. 3.3). Additionally, the mapping process incorporates GDT [4] and utilizes a specially designed loss function (Sec. 3.4). To further enhance accuracy, we propose a RANSAC-based algorithm for estimating camera pose from rays in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.719, + 0.833 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Employing CRR for VL involves two main steps: 1) regressing the patch-level camera ray parameters from the input image and 2) calculating the camera pose from the ray parameters. Given an input image \\(I \\in \\mathbb{R}^{H \\times W \\times C}\\), it is first" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25641" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.467, + 0.107 + ], + "angle": 0, + "content": "divided uniformly into a set of patches \\((p_i)\\) of size \\(s\\times s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.117, + 0.483, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\left\\{p _ {i} \\in \\mathbb {R} ^ {s \\times s \\times C} \\right\\} _ {i = 1} ^ {N}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.147, + 0.485, + 0.21 + ], + "angle": 0, + "content": "where \\(N = HW / s^2\\). Subsequently, the model \\(\\mathcal{F}\\) regresses the ray parameters corresponding to the center points of image patches, expressed in the world coordinate system, \\(\\hat{\\mathcal{R}} = \\{r_i\\in \\mathbb{R}^6\\}_{i = 1}^N\\)" + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.223, + 0.483, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {R}} = \\mathcal {F} (I), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.485, + 0.339 + ], + "angle": 0, + "content": "where the ray parameters are expressed in Plücker coordinates [12]: \\( r_i = [d_i, m_i] \\), where \\( d_i \\in \\mathbb{R}^3 \\) denotes the ray direction and \\( m_i \\in \\mathbb{R}^3 \\) represents the ray moment. Upon obtaining sufficient rays (\\( \\geq 2 \\)), the camera pose can be acquired by solving two distinct linear problems corresponding to rotation and translation, respectively [44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.43 + ], + "angle": 0, + "content": "Firstly, in terms of rotation, given the camera intrinsic, the Plücker coordinates of all camera rays in the camera coordinate system can be achieved, denoted as \\(\\{r_i^c = [d_i^c, 0]\\}_i^N\\). The rotation matrix \\(R\\), thus, corresponds to the transformation from the ray directions in the camera coordinate system to those in the world coordinate system:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.443, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\hat {R} = \\arg \\min _ {| | R | | = 1} \\sum_ {i} \\left| \\left| R d _ {i} ^ {c} - d _ {i} \\right| \\right| _ {2} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.485, + 0.541 + ], + "angle": 0, + "content": "Subsequently, the translation, also known as the world position of the camera optical center denoted as \\( c \\), is determined. Since \\( c \\) is the intersection point of all camera rays, the optimization problem can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.554, + 0.483, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\hat {c} = \\arg \\min _ {c} \\sum_ {i} \\| c \\times d _ {i} - m _ {i} \\| _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Through these two linear problems, the camera pose \\(([\\hat{R} |\\hat{c} ])\\) can be deduced from the camera rays." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.257, + 0.65 + ], + "angle": 0, + "content": "3.2. Feature Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Given the patch-wise independent GDT, it entails that the scene-specific head must learn the mapping from patch-level features to ray parameters. To ensure the learnability of this, ambiguity within patch features, particularly the local ambiguities, needs to be addressed. Thus, based on our prior analyze, patch-level features must possess both robust local representation and image-level perception. Hence, we opt for DINO [26] as our encoder backbone. From an input image \\(I\\), it can generate patch-level feature tokens, \\(\\{p_i\\}_i^N\\), and provide the CLS token for image-level perception. According to prior work [13], fine-tuning DINO's last two blocks can effectively enhance the performance of VL tasks. Consequently, we also fine-tune DINO using the ACE Encoder training method [4], resulting in \\(\\mathcal{D}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {T} _ {i} \\right\\} _ {i} ^ {N}, \\mathcal {T} _ {[ \\mathrm {C L S} ]} = \\mathcal {D} (I), \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.776, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 RC-RANSAC Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.11, + 0.877, + 0.127 + ], + "angle": 0, + "content": "Input: \\(\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i = 1}^N\\) # camera rays from DIMM" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.127, + 0.825, + 0.143 + ], + "angle": 0, + "content": "Output: \\(\\hat{R} \\in \\mathbb{R}^{3 \\times 3}\\), \\(\\hat{c} \\in \\mathbb{R}^{3 \\times 1}\\); # camera pose" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.144, + 0.907, + 0.175 + ], + "angle": 0, + "content": "1: Get Rotation by RANSAC: \\(\\hat{R} \\gets R\\)-RANSAC(\\(\\{\\hat{d}_i\\}_{i=1}^N\\)); cf. Algorithm 2 of the Suppl." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.175, + 0.803, + 0.192 + ], + "angle": 0, + "content": "2: Ray Correction: \\(\\hat{\\mathcal{R}}\\gets \\{[\\hat{R}\\hat{d}_i^c,\\hat{m}_i]\\}_{i = 1}^N;\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.192, + 0.907, + 0.223 + ], + "angle": 0, + "content": "3: Get Camera Center by RANSAC: \\(\\hat{c} \\gets C\\)-RANSAC(\\(\\hat{\\mathcal{R}}\\)); cf. Algorithm 3 of the Suppl." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.144, + 0.907, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.311 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{T}_i\\in \\mathbb{R}^D\\) represents the patch token corresponding to \\(p_i\\) locally, \\(\\mathcal{T}_{[\\mathrm{CLS}]}\\in \\mathbb{R}^{D}\\) denotes the CLS token with global perception, \\(N\\) signifies the number of patches, and \\(D\\) is the token dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.906, + 0.419 + ], + "angle": 0, + "content": "Meanwhile, the geometric constraints of the rays from the same camera are crucial for eliminating local ambiguities. Thus, the network must also explicitly specify the spatial image location of patches during training. To further prevent the positional feature from being overshadowed by other dimensions, we draw inspiration from Nerf [24] and enhance this feature using Fourier Encoding \\((FE)\\) [40]." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.427, + 0.907, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\mathrm {x y} _ {i}} = F E \\left(\\left\\{x _ {i}, y _ {i} \\right\\} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times D _ {F}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Here, \\(\\{x_{i},y_{i}\\}_{i}\\) represents the positions of image patches \\(\\{p_i\\}_i\\) in the original image, and \\(D_F\\) is the dimension of the Fourier encoding, set to 16. Finally, the scene-independent patch-level feature is a combination of the aforementioned three features. For each patch \\(p_i\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.544, + 0.907, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {i} ^ {*} := \\mathcal {T} _ {i} \\otimes \\mathcal {T} _ {[ C L S ]} \\otimes \\mathcal {T} _ {\\mathrm {x y} _ {i}} \\in \\mathbb {R} ^ {2 D + D _ {F}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Here, \\(\\otimes\\) denotes concatenation along the last dimension. This carefully designed feature to address ambiguity serves as the foundation of our CRR method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.623, + 0.659, + 0.639 + ], + "angle": 0, + "content": "3.3. Mapper Head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The scene-specific mapper head is responsible for transforming input features to ray parameters in CRR. A powerful feature encoder allows similar transformations to be achieved with simple MLPs [4, 14, 22, 41]. However, these mappers struggles to handle large scenes due to its limited capacity. To address the issue, spatial clustering is applied to divide the large scene [4, 41], with multiple MLPs fitting different sub-scenes independently. While effective, this hard scene partitioning can lead to ambiguous results near the sub-scene boundaries. In contrast, we adopt a soft map partitioning. As its core, we fuse the outputs of multiple sub-mappers with soft semantic guidance. Through end-to-end training, the sub-mappers are able to adaptively focus on different semantic regions. Particularly, we employ two MLP projection layers \\((P_{s}\\) and \\(P_{g})\\) to first project the input feature into a geometric embedding \\(\\mathcal{T}_i^g\\) and a semantic embedding \\(\\mathcal{T}_i^s\\), with \\(D_h\\) hidden dimension respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25642" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.221 + ], + "angle": 0, + "content": "
CategoryMethodMapping w/ 3D info.Mapping TimeMapping Size
FM-basedhloc(SP+SG)No1.5 hours~2GB
SCRDSAC*Yes15 hours28MB
ACENo5 minutes4MB
ACE (×4)No20 minutes14MB
EGFSNo12 minutes4.5MB
EGFS (dual)No21 minutes9MB
D2SYes~9.4 hours22 MB
APRmarepoNo5 minutes98.9MB
mareporeplasNo~15 minutes98.9MB
CRRDIMM (Ours)No~1 hour16.2 MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.232, + 0.485, + 0.288 + ], + "angle": 0, + "content": "Table 1. The mapping size and time comparison between VL methods from different categories. As a new flavor of VL, our DIMM achieves mapping time and memory size comparable to other methods while ensuring privacy protection." + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.307, + 0.484, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {g} = P _ {g} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.35, + 0.484, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} ^ {s} = P _ {s} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.441 + ], + "angle": 0, + "content": "where we omit the indices of the patches for simplification. Next, the sub-mappers \\(\\mathcal{M} = \\{\\mathcal{M}_j\\}_{j=1}^M\\), composed of \\(M\\) MLPs with depth \\(D_M\\), take the geometric embedding \\(\\mathcal{T}^g\\) to obtain \\(M\\) ray embeddings \\(\\{\\mathcal{T}_j^r\\}_{j=1}^M\\)." + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.454, + 0.484, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {j} ^ {r} = \\mathcal {M} _ {j} \\left(\\mathcal {T} ^ {g}\\right) \\in \\mathbb {R} ^ {D _ {h}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.481, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Afterwards, in the sequence of L cross attention layers \\((\\mathcal{A} = \\{\\mathcal{A}_l\\}_{l=1}^L\\), termed as Semantic Attention), we replicate the semantic embedding \\((\\mathcal{T}_i^s)\\) \\(M\\) times along the first dimension to serve as the query \\(q\\)." + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.554, + 0.483, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\otimes \\left\\{\\mathcal {T} ^ {s} \\right\\} ^ {M} \\mapsto q \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.486, + 0.618 + ], + "angle": 0, + "content": "All ray embeddings are concatenated along the first dimension to form both the key and value of the first layer." + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.632, + 0.484, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\otimes \\left\\{\\mathcal {T} _ {j} ^ {r} \\right\\} _ {j} ^ {M} \\mapsto k _ {0}, v _ {0} \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.706 + ], + "angle": 0, + "content": "By iteratively updating the key and value through the attention layers, we fuse the results of all sub-mappers by weighting them using the semantic guidance from DINO." + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.719, + 0.483, + 0.735 + ], + "angle": 0, + "content": "\\[\nA _ {l} \\left(q, k _ {l}, v _ {l}\\right) \\mapsto v _ {l + 1}, k _ {l + 1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.829 + ], + "angle": 0, + "content": "The final value is first reshaped: \\( v_{L} \\in \\mathbb{R}^{M \\times D_{h}} \\to \\mathbb{R}^{MD_{h}} \\), followed by projection to obtain the ultimate ray embedding: \\( \\mathcal{T}_f^r = P_f(v_L) \\in \\mathbb{R}^{D_h} \\). Finally, the ray parameters \\( \\hat{r} = [\\hat{d},\\hat{m}] \\) are obtained through a straightforward MLP projection layer." + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.84, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\hat {r} = P _ {r} \\left(\\mathcal {T} _ {f} ^ {r}\\right) \\in \\mathbb {R} ^ {6}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Based on this head, we can achieve the corresponding camera ray from a patch-level embedding from our feature encoder." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.658, + 0.108 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.175 + ], + "angle": 0, + "content": "Adopting the patch-level Gradient Decorrelation Training, our loss function operates independently between patches. First, we utilize the L2 distance between the ground truth \\((r_{GT})\\) and predicted rays as the optimization target." + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.183, + 0.907, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l 2} = \\left\\| r _ {G T} - \\hat {r} \\right\\| _ {2} ^ {2}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.211, + 0.908, + 0.272 + ], + "angle": 0, + "content": "Besides, to enhance the model awareness of geometric constraints, we incorporate a center loss \\(\\mathcal{L}_c\\) based on the geometric properties of camera rays. The principle is any camera ray passes through a corresponding camera center [12]." + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.28, + 0.907, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\left\\| c \\times \\hat {d} - \\hat {m} \\right\\| _ {2} ^ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.907, + 0.384 + ], + "angle": 0, + "content": "Finally, our patch-level loss function is \\(\\mathcal{L} = \\mathcal{L}_{l2} + \\alpha \\mathcal{L}_c\\), where \\(\\alpha\\) is a balance term, which is empirically set as 0.2. Our training loss is simple yet effective compared to SCR methods [4, 14, 22], benefiting from the robustness of rays in optimization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.658, + 0.407 + ], + "angle": 0, + "content": "3.5. RC-RANSAC" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.909, + 0.705 + ], + "angle": 0, + "content": "Similar to SCR, it is also not guaranteed that all camera rays can be accurately predicted in CRR. Therefore, employing the RANSAC mechanism to eliminate outliers is a beneficial approach. Unlike the PnP solution, when solving for camera pose from rays, the rotation and translation constitute two linear problems. Hence, the RANSAC process can be separately applied. Given that the rotation solution is independent of translation, we introduce the RC-RANSAC (Algorithm 1) as a modification of the general RANSAC algorithm. Specifically, after obtaining the predicted camera rays \\(\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i}^{N}\\), we initially use \\(R\\)-RANSAC (cf. Algorithm 2 of the Suppl.) to determine the rotation \\(\\hat{R}\\) from the ray directions only. Subsequently, all camera rays are corrected using the obtained rotation matrix \\(\\hat{R}\\) to get better rays: \\(\\hat{\\mathcal{R}}^* = \\{[\\hat{R} d_i^c,\\hat{m}_i]\\}_{i}\\). The camera translation \\(\\hat{c}\\) is then derived from the corrected rays by another \\(C\\)-RANSAC process (cf. Algorithm 3 of the Suppl.). Ultimately, combining the rotation and translation yields the final camera pose: \\([\\hat{R},\\hat{c} ]\\) for VL." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.716, + 0.646, + 0.733 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.731, + 0.757 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Network Configuration. For the feature encoder \\(\\mathcal{D}\\), we use the dinov2_vitb14_reg as our backbone, with an embedding dimension of 768 (\\(D = 768\\)). The frequency of Fourier Encoding is set as 16, making \\(D_F = 64\\). On the other hand, the mapper head consists of 4 sub-mappers and a semantic attention model with 4 attention layers. Each sub-mapper has a 4-depth MLP with skip connections. The first 2 layers form the common mapping, and the other 2 layers output direction and moment embedding respectively" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25643" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.475, + 0.094, + 0.52, + 0.103 + ], + "angle": 0, + "content": "Indoor-6" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.106, + 0.905, + 0.289 + ], + "angle": 0, + "content": "
CategoryMethodscene1 (cm/°)scene2a (%)scene3 (cm/°)scene4a (%)scene5 (cm/°)scene6 (cm/°)Average (cm/°)
FM-basedHloc [30]3.2/0.564.8-/-51.42.1/0.481.02.9/0.5
Hloc+SLD [9]2.9/0.468.73.4/0.662.71.9/0.381.042.7
2.1/0.4
APRPoseNet [15]159.0/7.50.0-/--141.0/9.30.0118.2/9.3
SCRDSAC* [3]12.3/2.118.77.9/0.928.013.1/2.319.740.7/6.7
ACE [4]13.6/2.124.96.8/0.731.98.1/1.333.014.7/2.3
NBE+SLD(E) [9]7.5/1.228.47.3/0.730.46.2/1.343.56.3/1.0
NBE+SLD [9]6.5/0.938.47.2/0.732.74.4/0.953.06.0/0.9
EGFS [20]-/-46.4-/-60.6-/-56.4-/-
EGFS-q [20]-/-58.5-/-59.1-/-67.0-/-
D2S [5]4.8/0.851.84.0/0.461.13.6/0.760.05.8/0.9
2.4/0.5
CRRDIMM5.5/1.544.44.2/0.866.74.4/1.161.67.6/1.5
DIMM-R5.5/1.545.23.3/0.677.44.2/1.162.56.3/1.5
3.2/0.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.299, + 0.905, + 0.327 + ], + "angle": 0, + "content": "Table 2. Localization results on Indoor-6 [9]. We report the median errors in \\(cm\\) for the position, degree \\((^{\\circ})\\) for the orientation, and recall at \\(5cm / 5^{\\circ}\\) of the Indoor-6 dataset. The best results, the second best and our results are highlighted, except for the FM-based method." + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.328, + 0.361, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.328, + 0.514, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.328, + 0.666, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.328, + 0.818, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.445, + 0.361, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.449, + 0.514, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.449, + 0.666, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.449, + 0.818, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.576, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Figure 4. The qualitative comparison between ray regression w/ or w/o global perception in the scene1 of Indoor-6. These results contain predicted and ground truth rays from the same cameras, along with their camera centers. The samples indicate that ray predictions w/o global perception (without the CLS and XY tokens, bottom) lack geometric consistency, although some individual rays are accurate. In contrast, incorporating global perception facilitates learning the geometric constraint of rays, resulting in precise ray regression (top)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.669 + ], + "angle": 0, + "content": "(cf. Fig. 3). The hidden dimension \\( D_{h} \\) is set as 256. More ablation experiments can be found in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training and Hardware Details. Firstly, according to related work [13], we fine-tune the last two blocks of DINO in terms of CRR task for better performance. Particularly, we follow the encoder training protocol of [4] and [13], using the first 100 training scenes of ScanNet [7] to update the weights of unfrozen DINO blocks. The training is performed on 4 NVIDIA A800 GPUs using gradient accumulation [4] in parallel. Next, the mapper head is trained on 1 NVIDIA A800 GPU for each scene, with a training buffer of 26M encoder features. The large buffer size is important for CRR training, as most of patches have different rays. We observed a buffer size scaling law in experiments (cf. Suppl. B). We do 20 passes over the training buffer, utilizing a batch size of 5120. The optimization uses AdamW [21] with a learning rate between \\(5e^{-5}\\) and \\(1e^{-4}\\) and a 1 cycle schedule [37]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.736, + 0.654 + ], + "angle": 0, + "content": "4.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.834 + ], + "angle": 0, + "content": "Datasets. We evaluate VL performance of our method on both indoor and outdoor datasets. For indoor scenes, we adopt the Indoor-6 dataset [9], which is a relatively large indoor dataset [9]. It comprises calibrated images from 6 multi-rooms indoor scenes captured over several days, leading to challenging VL. For outdoor scenes, the Map-Free [1] dataset is employed. We use its first 10 scenes, \\( s00000 \\sim s00009 \\), following [4]. Each scene contains images from two scans of an outdoor location. We split them into mapping and query sets. The ground truth camera poses are computed via SfM method [34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Indoor Visual Localization. Firstly, the indoor VL experiments are conducted using the Indoor-6 dataset. In the experiments, we compare our DIMM method and its variation enhanced by RANSAC (DIMM-R) with both APR [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25644" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.194, + 0.089, + 0.805, + 0.291 + ], + "angle": 0, + "content": "
Scene (°/cm/%)SCRAPRCRR
DSAC*[3]ACE[4]marepo[6]marepos[6]DIMMDIMM-R
Throughput (fps)17.917.955.655.657.720.1
s000001.2/5.1/55.80.7/4.2/62.60.6/3.6/69.20.6/3.2/70.70.7/2.0/93.10.6/2.0/94.0
s000010.8/2.4/94.50.4/1.5/1000.9/1.8/96.00.8/1.6/99.21.1/1.6/96.81.1/1.6/96.8
s000020.8/2.4/69.50.5/2.1/77.70.6/2.3/77.10.5/2.1/78.20.9/3.7/71.20.9/3.0/77.1
s000030.6/2.1/75.80.4/2.4/73.60.9/2.5/80.30.8/2.1/84.50.6/2.4/92.80.6/2.2/93.3
s000041.1/6.3/40.40.7/5.8/45.30.8/5.3/45.90.7/5.1/47.60.7/2.4/82.60.7/2.3/96.2
s000050.7/3.6/64.40.5/2.4/73.30.7/2.7/72.20.5/2.6/77.30.8/2.5/86.20.8/2.1/87.9
s000061.0/4.4/53.40.8/4.5/54.80.8/4.2/58.50.6/3.6/63.80.8/5.6/44.30.8/3.9/64.8
s000070.7/7.8/16.90.6/7.6/19.01.0/6.3/25.10.8/6.1/27.50.6/3.6/69.50.5/3.5/71.1
s000081.2/3.5/68.90.7/3.1/71.60.7/2.3/77.40.7/1.9/82.80.7/2.7/85.10.7/2.4/87.9
s000090.8/3.1/81.20.4/1.2/99.70.8/1.8/94.90.9/1.6/95.10.7/1.3/97.70.7/1.2/97.4
Average0.8/4.1/62.10.6/3.5/67.70.8/3.3/69.70.7/3.0/72.70.8/2.8/81.90.7/2.4/86.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.302, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 3. Pose accuracy comparison on the outdoor MapFree [1] dataset. We report the median errors in degree (°) for the orientation, cm for the position, and recall at \\(5\\mathrm{cm} / 5^{\\circ}\\). The best results, the second best and our results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.338, + 0.482, + 0.398 + ], + "angle": 0, + "content": "
Recall-5°/5cm (%)s00000s00001s00002s00003s00004s00005Average
ACE [4]62.610077.773.645.373.372.1
DIMM-CRR (Ours)93.196.871.292.882.686.287.1
DIMM-SCR64.898.674.577.348.272.172.6
DIMM-APR3.212.45.73.60.32.64.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.408, + 0.485, + 0.451 + ], + "angle": 0, + "content": "Table 4. Ablation Results about Camera Ray Regression (CRR). We evaluate three variations of our model to showcase the effectiveness of CRR, by modifying output to different VL formulations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.485, + 0.776 + ], + "angle": 0, + "content": "and SCR [3-5, 9, 20] approaches. The results are reported in Tab. 2. Additionally, we analyzed the corresponding mapping time and memory costs of various VL methods in Tab. 1, for a comprehensive comparison. DIMM demonstrates better or comparable performance than the previous state-of-the-art, e.g., achieving a recall@5°/5cm of 78.3 in scene6. Moreover, our method achieves generally better results than EGFS-q with \"hard\" multi-mapper combination, proving the efficacy of our \"soft\" mapper ensemble utilizing semantic guidance. The effectiveness of RC-RANSAC is also highlighted by a substantial precision improvement in scene2a from 66.7 to 77.4. Overall, the proposed DIMM-R achieves the best average recall in this dataset. Furthermore, our method enables mapping with reduced computational costs compared to D2S and is trained without 3D supervision, as illustrated in Tab. 1. These findings underscore the potential of our CRR-based method as a promising VL implementation. Some visualization results of DIMM can be found in Fig. 4 and Fig. 7 in the Suppl., where the 3D scene structure is preserved, demonstrating the privacy protection offered by our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Outdoor Visual Localization. Our outdoor VL experiments are conducted on the Map-Free Dataset. In addition to the SCR-based ACE and DSAC*, we also compared the recent APR-based method, marepo [6], and its fine-tuned variation, marepos. The results, including median errors for rotation \\((^{\\circ})\\) and translation \\((cm)\\), as well as localization recalls at \\(5\\mathrm{cm} / 5^{\\circ}\\), are presented in Tab. 3. These results also come from methods with mapping time in Tab. 1. From the table," + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.338, + 0.868, + 0.413 + ], + "angle": 0, + "content": "
Feature EncoderMed. R Err. (°)Med. t Err. (cm)Recall-5°/5cm
T_i ∩ T[CLS] ∩ Txy_i0.85.644.3
T_i * T*[CLS] ∩ Txy_i0.86.039.7
T_i ∩ TG[41] ∩ Txy_i1.18.319.7
T_i ∩ T[CLS] ∩ XY_i0.97.627.2
T_i ∩ T[CLS]0.97.426.8
T_i1.111.517.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Table 5. Ablation Study about our Feature Encoder. The experiments are performed on the MapFree s00006 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.909, + 0.728 + ], + "angle": 0, + "content": "it is evident that our methods achieve the best overall results, outperforming both SCR and APR methods. The rotation accuracy of ACE, yet, surpasses our methods, indicating that local ambiguity poses a significant challenge for CRR methods and obtaining precise ray directions from image patches is challenging. However, it is observed that CRR methods facilitate better camera center estimation compared to APR and SCR methods. This improvement can be attributed to the fact that the camera center location integrates contributions from all camera rays, and the bounded ray error [27] enhances the robustness of the center estimation against noise. The efficacy of RC-RANSAC algorithm can also be verified in the table, as it leads to substantial improvements of VL performance, e.g., \\(44.3 \\rightarrow 64.8\\) of recall in s00006. Furthermore, we investigate the computational efficiency of our methods. The throughputs (fps) of all methods are reported in the table, revealing that the lightweight network and linear pose solver of DIMM achieve the best inference efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.72, + 0.756 + ], + "angle": 0, + "content": "4.3. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Camera Ray Regression. One of our main contributions in this work is introducing Camera Ray Regression (CRR) into the VL task. Thus, we provide ablation study about this formulation in Tab. 4. In particular, we modify the last layer of DIMM to output scene coordinates (DIMM-SCR) or direct camera pose (DIMM-APR), and train these variations separately. The DIMM-SCR is trained with GDT, same as ACE[4]. The DIMM-APR, on the other hand, is trained in image level, as the direct pose vectors cannot" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25645" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.462, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.26, + 0.485, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Visualization results of sub-mapper attention score. In the experiments, we observed that specific semantic entities elicit responses from different mappers. For instance, the fireplace at the top of the figure frequently results in high attention scores from mapper #3, while mapper #2 exhibits a low attention score. Conversely, mapper #3 seems do not \"care\" the sofa at the bottom, which leads to high scores of mapper #2 instead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.562 + ], + "angle": 0, + "content": "be assigned to image patches. The experiments are performed on the MapFree dataset, with the localization recall reported. In Tab. 4, we can see that our CRR formulation significantly outperforms the SCR and APR under the same network architecture, indicating the contribution of our ray-parameterization. DIMM-SCR overcomes ACE slightly in average, but our model is tailored for CRR. DIMM-APR utilizes per-scene pose regression and achieves poor results, but this is consistent with the results in marepo [6]. This suggests the APR model may require much larger datasets to learn pose patterns. Overall, the results in Tab. 4 evaluate our main contribution: CRR formulation in VL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Feature Encoder. Based on our analysis of CRR, we require patch features to exhibit strong local description while also incorporating global perception and image position encoding to achieve disambiguation. To validate this, we conduct experiments on the MapFree s00006 dataset (cf. Tab. 5). Specifically, we modify the content of the patch feature, including features w/o global perception and image position encoding \\((\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}\\) and \\(\\mathcal{T}_i)\\), to compare with the proposed version \\((\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathcal{T}_{\\mathrm{xy}_i})\\). Also, we investigate the efficacy of DINO fine-tuning, and the performance of original DINO \\((\\mathcal{T}_i^* \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}^* \\otimes \\mathcal{T}_{\\mathrm{xy}_i})\\) is also reported. It is evident that global perception with image position encoding are crucial for CRR, leading to a significant accuracy improvement. The fine-tuning of DINO also improves the performance as well (39.7 vs. 44.3). Visualization comparison about the global perception is presented in Fig. 4. We can see the proposed global perception enables the learning of ray geometry, causing the rays to converge toward the camera center, thereby enhancing precision. We also conduct experiments for a variation without Fourier Encoding \\((\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathsf{XY}_i)\\). The results in the first and third rows of Tab. 5 show that Fourier Encoding effectively prevents the submergence of positional" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.09, + 0.887, + 0.156 + ], + "angle": 0, + "content": "
Atten. Num.Med. R Err. (°)Med. t Err. (cm)Recall@5°/Scmsize (MB)
011.07.231.012.1
20.95.842.614.2
40.85.644.316.2
60.86.337.718.2
81.17.730.420.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.545, + 0.156, + 0.794, + 0.165 + ], + "angle": 0, + "content": "1 Take average of the sub-mapper embeddings, eliminating attention layers." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.175, + 0.908, + 0.218 + ], + "angle": 0, + "content": "Table 6. Ablation Study about the Number of Semantic Attention Layers. The experiments are performed on the MapFree s00006 dataset. The efficacy of our semantic attention is evaluated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.32 + ], + "angle": 0, + "content": "features. Finally, we explore another possible global perception way provided by [41] \\((\\mathcal{T}_i \\otimes \\mathcal{T}_G \\otimes \\mathcal{T}_{xy_i})\\), the feature for image retrieval. The results show that the replaced global feature fails to match the original efficacy, possibly because DINO's global feature is more compatible with its own local patch features, facilitating subsequent mapping learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Mapper Head. In the same scene, we also conduct ablation experiments on the network structure of the Mapper Head. We first validate the efficacy of the semantic attention module and experiment with its number of layers. When the layer number is set to 0, we directly take the average of output from multiple sub-mappers. The results are presented in Tab. 6, containing median rotation/translation error, localization recall and model size. It is evident that utilizing semantic attention is more effective compared to simple average fusion, and the increase in network size is deemed acceptable. Additional ablation study about the MLP mappers can be found in Suppl. B. To showcase the effectiveness of the soft mapper assemble, we visualize some query images combined with their sub-mapper attention scores in Fig. 5. It can be observed that different sub-mappers get high attention scores to different objects in images. This indicates that our semantic attention mechanism allows the sub-mappers to \"remember\" specific semantic objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.627, + 0.634, + 0.643 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose regressing camera rays to perform visual localization. This overparameterized representation of camera model leads to high precision, and, more importantly, enhances privacy protection. In particular, we introduce DIMM as a learning model to regress camera rays from images. It utilizes DINO as a scene-agnostic encoder to output features that incorporate both local and global perception. A scene-specific mapper then regresses ray parameters from the features, involving a semantic attention module to merge results from multiple mappers. Finally, a ray-level RANSAC algorithm is proposed to improve the ray-to-pose accuracy. In experiments on both indoor and outdoor datasets, our methods demonstrate comparable or superior performance to current approaches, while ensuring privacy preservation. Acknowledgement. This work has been funded in part by the NSFC grants 62176156." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25646" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Áron Monszpart, Victor Adrian Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In ECCV, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.485, + 0.243 + ], + "angle": 0, + "content": "[2] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7525-7534, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[3] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE transactions on pattern analysis and machine intelligence, 44 (9):5847-5865, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[5] Bach-Thuan Bui, Huy-Hoang Bui, Dinh-Tuan Tran, and JooHo Lee. D2s: Representing sparse descriptors and 3d coordinates for camera relocalization. IEEE Robotics and Automation Letters, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[6] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[9] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[10] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[11] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17658-17668, 2023. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[14] Xudong Jiang, Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. R-score: Revisiting scene coordinate regression for robust large-scale visual localization. arXiv preprint arXiv:2501.01421, 2025. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[15] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In 2016 IEEE international conference on Robotics and Automation (ICRA), pages 4762-4769. IEEE, 2016. 1, 3, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[16] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5974-5983, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[17] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[18] Sihang Li, Siqi Tan, Bowen Chang, Jing Zhang, Chen Feng, and Yiming Li. Unleashing the power of data synthesis in visual localization. arXiv preprint arXiv:2412.00138, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[20] Ting-Ru Liu, Hsuan-Kung Yang, Jou-Min Liu, Chun-Wei Huang, Tsung-Chih Chiang, Quan Kong, Norimasa Kobori, and Chun-Yi Lee. Reprojection errors as prompts for efficient scene coordinate regression. In European Conference on Computer Vision, pages 286-302. Springer, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.906, + 0.519 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[22] Deben Lu, Wendong Xiao, Teng Ran, Liang Yuan, Kai Lv, and Jianbo Zhang. Attention-based accelerated coordinate encoding network for visual relocalization. In 2024 IEEE 7th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), pages 1675-1680, 2024. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.907, + 0.661 + ], + "angle": 0, + "content": "[23] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.663, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.907, + 0.774 + ], + "angle": 0, + "content": "[25] Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. arXiv preprint arXiv:2412.12392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DiNov2: Learning robust visual features without supervision, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.146 + ], + "angle": 0, + "content": "[27] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schonberger. Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV), 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.231 + ], + "angle": 0, + "content": "[28] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.301 + ], + "angle": 0, + "content": "[29] Francesco Pittaluga, Sanjeev J. Koppal, Sing Bing Kang, and Sudipta N. Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[30] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.467 + ], + "angle": 0, + "content": "[31] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[32] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 752-765. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[33] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE transactions on pattern analysis and machine intelligence, 39(9):1744-1756, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[34] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[35] Thomas Schops, Viktor Larsson, Marc Pollefeys, and Torsten Sattler. Why having 10,000 parameters in your camera model is better than twelve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2544, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[36] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[37] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of residual networks using large learning rates. Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[38] Pablo Speciale, Johannes L. Schonberger, Sing Bing Kang, Sudipta N. Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5488-5498, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.16 + ], + "angle": 0, + "content": "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[40] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.907, + 0.316 + ], + "angle": 0, + "content": "[41] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[42] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.907, + 0.457 + ], + "angle": 0, + "content": "[43] Yinshuang Xu, Dian Chen, Katherine Liu, Sergey Zakharov, Rares Andrei Ambrus, Kostas Daniilidis, and Vitor Campagnolo Guizilini. \\(se(3)\\) equivariant ray embeddings for implicit multi-view depth estimation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.907, + 0.514 + ], + "angle": 0, + "content": "[44] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.907, + 0.571 + ], + "angle": 0, + "content": "[45] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20217-20226, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25648" + } + ] +] \ No newline at end of file diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_origin.pdf b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a3050b6c1bdbe109f70ee3df63357b984e3d5e7 --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/d007a0e2-f228-4563-89f9-6176e65aa729_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94ea3c964bf625df52ff97e103b27ab83e08ecea1ccd02e8fa1b7d2c60bc7c0f +size 2203953 diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/full.md b/2025/Semantic-guided Camera Ray Regression for Visual Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b0b4323bc3e9becdede523b347016adb2b23d280 --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/full.md @@ -0,0 +1,331 @@ +# Semantic-guided Camera Ray Regression for Visual Localization + +Yesheng Zhang, Xu Zhao* +School of Automation and Intelligent Sensing, Shanghai Jiao Tong University +{preacher, zhaoxu}@sjtu.edu.cn + +# Abstract + +This work presents a novel framework for Visual Localization (VL), that is, regressing camera rays from query images to derive camera poses. As an overparameterized representation of the camera pose, camera rays possess superior robustness in optimization. Of particular importance, Camera Ray Regression (CRR) is privacy-preserving, rendering it a viable VL approach for real-world applications. Thus, we introduce DINO-based Multi-Mappers, coined DIMM, to achieve VL by CRR. DIMM utilizes DINO as a scene-agnostic encoder to obtain powerful features from images. To mitigate ambiguity, the features integrate both local and global perception, as well as potential geometric constraint. Then, a scene-specific mapper head regresses camera rays from these features. It incorporates a semantic attention module for soft fusion of multiple mappers, utilizing the rich semantic information in DINO features. In extensive experiments on both indoor and outdoor datasets, our methods showcase impressive performance, revealing a promising direction for advancements in VL. + +# 1. Introduction + +Visual Localization (VL) involves estimating 6-degree-of-freedom camera pose of an image taken from a known scene, which is also known as camera relocalization. VL plays a vital role in many vision, robotic and graphic applications, such as augment reality, virtual reality and autonomous driving. Despite extensive research, achieving precise, robust, and privacy-preserving VL [28, 38] continues to present a significant challenge. + +Current VL methods can be categorized into explicit and implicit methods based on their mapping approaches. Traditional explicit methods [2, 30, 31] involve an explicit 3D map, e.g., point clouds. They establish 2D-3D (image-to-map) correspondences by Feature Matching, and then estimate camera pose via PnP and RANSAC [10, 12]. While advanced methods [11, 39] leads to high accuracy, the stor + +![](images/d1094e448dc67043bf342f4052a9dad3b42f9463fdd4b0c457783c941a957d08.jpg) +Figure 1. The ambiguity challenge in Camera Ray Regression (CRR) for VL. Effective global and local disambiguations are critical for CRR. Global ambiguity arises from rays of different patches (cf. $r_2^1$ and $r_2^2$ in the figure), while local ambiguity occurs within the same patch but from different viewing angles (cf. $r_1^1$ and $r_2^2$ ). Semantic information, e.g., the table in the figure to differentiate between chairs, along with geometry constraints (geometry properties of rays $\{r_1^i\}_i$ ) are essential for the disambiguations. + +age demands of explicit maps are troublesome especially in large-scale scenes. Additionally, explicit maps raise privacy concerns, restricting their applications in real-world settings. + +Implicit methods, on the other hand, leverage learning models as implicit and light-weight maps, which mainly include two frameworks: Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR). APR directly utilizes networks to regress camera poses (rotations and translations) from images [6, 15], offering enhanced privacy protection. However, the compact nature of absolute pose makes it vulnerable to noise, thereby leading to accuracy issue [6]. + +To improve precision, employing an overparameterized representation for the camera pose is crucial [35]. Thus, SCR framework [22, 41] is proposed, relying on 2D-3D matches as an indirect yet robust camera pose representation, akin to its explicit counterparts. ACE [4] is a milestone within this framework, utilizing simple networks to regress pixel-aligned scene coordinates. It introduces a scene-centric patch-level training strategy, i.e., Gradient Decorrelation Training (GDT), enabling both efficient and precise mapping. Nevertheless, the unbounded 3D point error is unstable in training, necessitating manual hyperparameter tuning or additional outlier rejection modules [5, 20]. Besides, SCR methods still run the risk of 3D privacy breaches, as they + +![](images/94306d1c810ff5d230ee5804b658da387401ef5a043ed499d24347ee427df6fa.jpg) +Figure 2. The proposed DIMM approach performing CRR for VL. This approach utilizes DINO as a scene-agnostic feature encoder to output carefully designed patch-level features. Then, a scene-specific mapper head integrates multi-mapper results under semantic guidance, through a semantic attention module. After the scene-centric training technology, DIMM is capable of predicting camera ray parameters for accurate and privacy-preserving VL. + +disclose detailed 3D scene information [29]. + +Recently, a promising camera pose representation, camera ray [27, 35], has garnered increased attention in the vision field. Its overparameterization is beneficial for patch-level learning [23, 44], aligning well with the effective GDT in VL. This inspires us to introduce camera rays in the VL tasks. In particular, we propose adopting a learning model to regress patch-level camera rays from the image, termed as Camera Ray Regression (CRR). Then, the camera pose can be achieved by solving two linear problems [44]. This implementation of VL could strike a balance between efficacy and privacy preservation. Firstly, it belongs to the implicit mapping category, incurring low storage costs. As a overparameterized pose representation, it can yield improved precision [35] than APR. Also, the ray error is more stable than the unbounded 3D point error [27] in optimization of model training. Hence, CRR obviates the need for intricate training setting like SCR methods [5, 20, 41]. More importantly, it is hard to recover 3D scene information from camera rays, leading to a privacy-preserving VL implementation. + +However, challenges exist in introducing camera rays in VL, primarily stemming from the issue of ambiguity (cf. Fig. 1). This ambiguity can generally be divided into two levels. The first is global ambiguity, referring to distinguish rays of different image patches, particularly in repetitive or textureless scenes. The similar ambiguity is encountered in SCR [4, 41], where leveraging powerful local features, such as semantic features [20], has proven effective. The second is local ambiguity, unique to CRR, which involves differentiating rays from the same image patch with different view angles. The ambiguity becomes troublesome for patches with continuous depth, where changes in view direction result in minor differences in patch content but obvious variations in the corresponding rays. To eliminate this ambiguity, we emphasize two observations on the CRR task. 1) The entire image often contains patches with depth discontinuities. Their appearance changes under different + +viewing angles are significant enough to resolve the local ambiguity. Thus, the image-level perception is important for CRR disambiguation, which is also proven to be beneficial for handling global repetitiveness [41]. 2) The rays within the same image are governed by the same camera pose. This geometric constraint allows confident rays to correct uncertain ones. Meanwhile, the smoothness of the patchray mapping [44] makes this geometric constraint learnable. Therefore, the integration of image-level perception, geometric constraint and robust local features is critical for the global and local disambiguation of CRR. + +Based on the above analysis, this work presents DINO-based Multi-Mappers (DIMM, Fig. 2), as the first implementation of CRR in VL. Following the scene-centric design of ACE, DIMM consists of a scene-agnostic encoder and a scene-specific head, but possessing several key modules tailed for CRR. Specifically, we utilize DINO [8] as the feature encoder, utilizing its powerful semantic perception for disambiguation. It provides robust patch tokens as local features while supporting CLS tokens for image-level perception. We further incorporate Fourier Encoding [24] of the image position for each patch into the features (XY tokens), as the potential geometric constraint of camera rays are associated with patch locations. Furthermore, we propose a scene-specific mapper head which adopts a semantic attention module to softly fuse ray regression results from multiple Multi-Layer Perceptron (MLP)-based mappers. This head effectively utilizes the inherit semantic prior of the DINO features, to enhance ray regression performance especially in large scenes. Finally, we propose a ray-level RANSAC algorithm for accurate pose estimation, decomposing the rotation and translation calculation [44]. + +Our contributions are summaries as follows. + +1. To the best of our knowledge, we are the first to introduce the camera ray into VL, with a tailored network design to achieve query image poses by camera ray regression. +2. We propose carefully designed scene-agnostic features to handle local and global disambiguation in CRR, along with a semantic-guided mapper head that enables a soft multi-mapper ensemble. +3. Comprehensive experiments on indoor and outdoor datasets demonstrate the effectiveness of our method, achieving results that rival or surpass the existing state-of-the-art. Additionally, extensive ablation studies validate the contributions of each module. + +# 2. Related Work + +Feature Matching-based VL. Traditional VL methods [30, 32, 33] rely on Feature Matching (FM) to determine the position and direction of cameras. These techniques aim to match 3D scene points with 2D image points, enabling the camera pose calculation through $\mathrm{PnP} + \mathrm{RANSAC}$ -based + +![](images/3e18d4064dc6c3b2e7eba9d92fce0f63485b0302b383fdf56a3b0a6343f91569.jpg) +Figure 3. The Multi-Mapper network with semantic attention as our scene-specific head. The patch-level features are first projected to geometry and semantic embeddings. The geometry embedding is fed into multiple MLP-based mappers to derive ray embeddings. These embeddings are then fused under the guidance of the semantic embedding to achieve the camera rays, through cross attention modules. + +algorithms [10, 12]. As a result, these approaches necessitate the storage of explicit maps containing 3D points and their feature descriptors. By leveraging robust matching algorithms [11, 19, 39, 42, 45], FM-based localization can deliver precise outcomes. Nevertheless, explicit maps raises concerns about privacy leaks [28]. Furthermore, as the scene size expands, managing the increasingly large explicit maps becomes challenging, leading to storage issues. In contrast, our method employs implicit maps that have lower storage requirements and privacy protection. + +Scene Coordinate Regression. Similar to FM-based VL, Scene Coordinate Regression (SCR) also accomplishes camera poses through 3D-2D matching [3]. However, SCR directly employs network to regress pixel-aligned 3D scene coordinates, which addresses the map storage issue through implicit mapping. ACE [4], a prominent approach within SCR, introduces a gradient decorrelation training technique enabling mapping within 5 minutes using a 4MB network. Subsequent methods [5, 14, 20, 22, 41] built upon ACE further enhance VL accuracy mainly by solid feature selection. Nevertheless, the limited network size of ACE pose challenges in fitting large scenes, necessitating spatial segmentation of the large scene for sub-map construction. Despite its efficacy, this approach may result in ambiguous localization at the boundaries of segmented sub-scenes. At the same time, it is noteworthy that the SCR methods output the 3D information of scenes, i.e., 3D coordinates, thereby leading to a potential risk of privacy leak. + +Absolute Pose Regression. Absolute Pose Regression (APR) also adopts implicit mapping, but directly utilizes networks to predict the rotation and translation of cameras. However, this compact parameter representation of camera pose is highly sensitive to noises [12, 44]. Regressing pose parameters from images, thus, is challenging [15-17, 36], and the accuracy is insufficient to compare with methods using over-parameterized representations, e.g., 2D-3D correspondences. Recent APR approach [6] opts to predict the camera pose based on the SCR output, resulting in favorable accuracy. However, this success is heavily reliant + +on the precision of SCR methods. In contrast, our method employs camera rays to depict the camera pose. This overparameterized representation is more conducive to learning optimization [25, 27, 44], enabling superior accuracy. + +Camera Ray in Vision. Recently, camera rays have emerged as a promising camera parametrization, representing a generic camera configuration [35]. In recent work [23, 43, 44], camera rays have been proven to be well-suited for patch-level learning, with applications in multiple vision tasks. Also, in 3D reconstruction works [25, 27], camera rays yield more robust geometric optimization than classical correspondence-based representations, e.g., 3D-2D matches. This advantage stems from their overparameterized nature and constrained ray error. Likewise, our approach is the first to introduce camera rays in VL, performing Camera Ray Regression (CRR) to achieve camera poses from images. Its patch-level property aligns well with the Gradient Deceleration Training [4], and the bounded ray error is beneficial for optimization in VL Learning as well. Furthermore, the inherent privacy-preserving attribute of CRR positions it as a promising direction for VL implementations. + +# 3. Method + +In this section, we first present the formulation of CRR for the VL task in Sec. 3.1. Then the proposed DIMM method is described in detail. DIMM adopts a scene-centric setup [4, 14, 41] by dividing the network into a scene-agnostic encoder (Sec. 3.2) and a scene-specific head (Sec. 3.3). Additionally, the mapping process incorporates GDT [4] and utilizes a specially designed loss function (Sec. 3.4). To further enhance accuracy, we propose a RANSAC-based algorithm for estimating camera pose from rays in Sec. 3.5. + +# 3.1. Problem Formulation + +Employing CRR for VL involves two main steps: 1) regressing the patch-level camera ray parameters from the input image and 2) calculating the camera pose from the ray parameters. Given an input image $I \in \mathbb{R}^{H \times W \times C}$ , it is first + +divided uniformly into a set of patches $(p_i)$ of size $s\times s$ : + +$$ +\left\{p _ {i} \in \mathbb {R} ^ {s \times s \times C} \right\} _ {i = 1} ^ {N}, \tag {1} +$$ + +where $N = HW / s^2$ . Subsequently, the model $\mathcal{F}$ regresses the ray parameters corresponding to the center points of image patches, expressed in the world coordinate system, $\hat{\mathcal{R}} = \{r_i\in \mathbb{R}^6\}_{i = 1}^N$ + +$$ +\hat {\mathcal {R}} = \mathcal {F} (I), \tag {2} +$$ + +where the ray parameters are expressed in Plücker coordinates [12]: $r_i = [d_i, m_i]$ , where $d_i \in \mathbb{R}^3$ denotes the ray direction and $m_i \in \mathbb{R}^3$ represents the ray moment. Upon obtaining sufficient rays ( $\geq 2$ ), the camera pose can be acquired by solving two distinct linear problems corresponding to rotation and translation, respectively [44]. + +Firstly, in terms of rotation, given the camera intrinsic, the Plücker coordinates of all camera rays in the camera coordinate system can be achieved, denoted as $\{r_i^c = [d_i^c, 0]\}_i^N$ . The rotation matrix $R$ , thus, corresponds to the transformation from the ray directions in the camera coordinate system to those in the world coordinate system: + +$$ +\hat {R} = \arg \min _ {| | R | | = 1} \sum_ {i} \left| \left| R d _ {i} ^ {c} - d _ {i} \right| \right| _ {2} ^ {2}. \tag {3} +$$ + +Subsequently, the translation, also known as the world position of the camera optical center denoted as $c$ , is determined. Since $c$ is the intersection point of all camera rays, the optimization problem can be formulated as: + +$$ +\hat {c} = \arg \min _ {c} \sum_ {i} \| c \times d _ {i} - m _ {i} \| _ {2} ^ {2}. \tag {4} +$$ + +Through these two linear problems, the camera pose $([\hat{R} |\hat{c} ])$ can be deduced from the camera rays. + +# 3.2. Feature Encoder + +Given the patch-wise independent GDT, it entails that the scene-specific head must learn the mapping from patch-level features to ray parameters. To ensure the learnability of this, ambiguity within patch features, particularly the local ambiguities, needs to be addressed. Thus, based on our prior analyze, patch-level features must possess both robust local representation and image-level perception. Hence, we opt for DINO [26] as our encoder backbone. From an input image $I$ , it can generate patch-level feature tokens, $\{p_i\}_i^N$ , and provide the CLS token for image-level perception. According to prior work [13], fine-tuning DINO's last two blocks can effectively enhance the performance of VL tasks. Consequently, we also fine-tune DINO using the ACE Encoder training method [4], resulting in $\mathcal{D}$ . + +$$ +\left\{\mathcal {T} _ {i} \right\} _ {i} ^ {N}, \mathcal {T} _ {[ \mathrm {C L S} ]} = \mathcal {D} (I), \tag {5} +$$ + +# Algorithm 1 RC-RANSAC Algorithm + +Input: $\hat{\mathcal{R}} = \{[\hat{d}_i,\hat{m}_i]\}_{i = 1}^N$ # camera rays from DIMM + +Output: $\hat{R} \in \mathbb{R}^{3 \times 3}$ , $\hat{c} \in \mathbb{R}^{3 \times 1}$ ; # camera pose + +1: Get Rotation by RANSAC: $\hat{R} \gets R$ -RANSAC( $\{\hat{d}_i\}_{i=1}^N$ ); cf. Algorithm 2 of the Suppl. +2: Ray Correction: $\hat{\mathcal{R}}\gets \{[\hat{R}\hat{d}_i^c,\hat{m}_i]\}_{i = 1}^N;$ +3: Get Camera Center by RANSAC: $\hat{c} \gets C$ -RANSAC( $\hat{\mathcal{R}}$ ); cf. Algorithm 3 of the Suppl. + +Here, $\mathcal{T}_i\in \mathbb{R}^D$ represents the patch token corresponding to $p_i$ locally, $\mathcal{T}_{[\mathrm{CLS}]}\in \mathbb{R}^{D}$ denotes the CLS token with global perception, $N$ signifies the number of patches, and $D$ is the token dimension. + +Meanwhile, the geometric constraints of the rays from the same camera are crucial for eliminating local ambiguities. Thus, the network must also explicitly specify the spatial image location of patches during training. To further prevent the positional feature from being overshadowed by other dimensions, we draw inspiration from Nerf [24] and enhance this feature using Fourier Encoding $(FE)$ [40]. + +$$ +\mathcal {T} _ {\mathrm {x y} _ {i}} = F E \left(\left\{x _ {i}, y _ {i} \right\} _ {i}\right) \in \mathbb {R} ^ {N \times D _ {F}} \tag {6} +$$ + +Here, $\{x_{i},y_{i}\}_{i}$ represents the positions of image patches $\{p_i\}_i$ in the original image, and $D_F$ is the dimension of the Fourier encoding, set to 16. Finally, the scene-independent patch-level feature is a combination of the aforementioned three features. For each patch $p_i$ , we have: + +$$ +\mathcal {T} _ {i} ^ {*} := \mathcal {T} _ {i} \otimes \mathcal {T} _ {[ C L S ]} \otimes \mathcal {T} _ {\mathrm {x y} _ {i}} \in \mathbb {R} ^ {2 D + D _ {F}}, \tag {7} +$$ + +Here, $\otimes$ denotes concatenation along the last dimension. This carefully designed feature to address ambiguity serves as the foundation of our CRR method. + +# 3.3. Mapper Head + +The scene-specific mapper head is responsible for transforming input features to ray parameters in CRR. A powerful feature encoder allows similar transformations to be achieved with simple MLPs [4, 14, 22, 41]. However, these mappers struggles to handle large scenes due to its limited capacity. To address the issue, spatial clustering is applied to divide the large scene [4, 41], with multiple MLPs fitting different sub-scenes independently. While effective, this hard scene partitioning can lead to ambiguous results near the sub-scene boundaries. In contrast, we adopt a soft map partitioning. As its core, we fuse the outputs of multiple sub-mappers with soft semantic guidance. Through end-to-end training, the sub-mappers are able to adaptively focus on different semantic regions. Particularly, we employ two MLP projection layers $(P_{s}$ and $P_{g})$ to first project the input feature into a geometric embedding $\mathcal{T}_i^g$ and a semantic embedding $\mathcal{T}_i^s$ , with $D_h$ hidden dimension respectively. + +
CategoryMethodMapping w/ 3D info.Mapping TimeMapping Size
FM-basedhloc(SP+SG)No1.5 hours~2GB
SCRDSAC*Yes15 hours28MB
ACENo5 minutes4MB
ACE (×4)No20 minutes14MB
EGFSNo12 minutes4.5MB
EGFS (dual)No21 minutes9MB
D2SYes~9.4 hours22 MB
APRmarepoNo5 minutes98.9MB
mareporeplasNo~15 minutes98.9MB
CRRDIMM (Ours)No~1 hour16.2 MB
+ +Table 1. The mapping size and time comparison between VL methods from different categories. As a new flavor of VL, our DIMM achieves mapping time and memory size comparable to other methods while ensuring privacy protection. + +$$ +\mathcal {T} ^ {g} = P _ {g} \left(\mathcal {T} _ {i} ^ {*}\right) \in \mathbb {R} ^ {D _ {h}}, \tag {8} +$$ + +$$ +\mathcal {T} ^ {s} = P _ {s} \left(\mathcal {T} _ {i} ^ {*}\right) \in \mathbb {R} ^ {D _ {h}}, \tag {9} +$$ + +where we omit the indices of the patches for simplification. Next, the sub-mappers $\mathcal{M} = \{\mathcal{M}_j\}_{j=1}^M$ , composed of $M$ MLPs with depth $D_M$ , take the geometric embedding $\mathcal{T}^g$ to obtain $M$ ray embeddings $\{\mathcal{T}_j^r\}_{j=1}^M$ . + +$$ +\mathcal {T} _ {j} ^ {r} = \mathcal {M} _ {j} \left(\mathcal {T} ^ {g}\right) \in \mathbb {R} ^ {D _ {h}}. \tag {10} +$$ + +Afterwards, in the sequence of L cross attention layers $(\mathcal{A} = \{\mathcal{A}_l\}_{l=1}^L$ , termed as Semantic Attention), we replicate the semantic embedding $(\mathcal{T}_i^s)$ $M$ times along the first dimension to serve as the query $q$ . + +$$ +\otimes \left\{\mathcal {T} ^ {s} \right\} ^ {M} \mapsto q \in \mathbb {R} ^ {M \times D _ {h}}. \tag {11} +$$ + +All ray embeddings are concatenated along the first dimension to form both the key and value of the first layer. + +$$ +\otimes \left\{\mathcal {T} _ {j} ^ {r} \right\} _ {j} ^ {M} \mapsto k _ {0}, v _ {0} \in \mathbb {R} ^ {M \times D _ {h}}. \tag {12} +$$ + +By iteratively updating the key and value through the attention layers, we fuse the results of all sub-mappers by weighting them using the semantic guidance from DINO. + +$$ +A _ {l} \left(q, k _ {l}, v _ {l}\right) \mapsto v _ {l + 1}, k _ {l + 1}. \tag {13} +$$ + +The final value is first reshaped: $v_{L} \in \mathbb{R}^{M \times D_{h}} \to \mathbb{R}^{MD_{h}}$ , followed by projection to obtain the ultimate ray embedding: $\mathcal{T}_f^r = P_f(v_L) \in \mathbb{R}^{D_h}$ . Finally, the ray parameters $\hat{r} = [\hat{d},\hat{m}]$ are obtained through a straightforward MLP projection layer. + +$$ +\hat {r} = P _ {r} \left(\mathcal {T} _ {f} ^ {r}\right) \in \mathbb {R} ^ {6}. \tag {14} +$$ + +Based on this head, we can achieve the corresponding camera ray from a patch-level embedding from our feature encoder. + +# 3.4. Training Loss + +Adopting the patch-level Gradient Decorrelation Training, our loss function operates independently between patches. First, we utilize the L2 distance between the ground truth $(r_{GT})$ and predicted rays as the optimization target. + +$$ +\mathcal {L} _ {l 2} = \left\| r _ {G T} - \hat {r} \right\| _ {2} ^ {2}. \tag {15} +$$ + +Besides, to enhance the model awareness of geometric constraints, we incorporate a center loss $\mathcal{L}_c$ based on the geometric properties of camera rays. The principle is any camera ray passes through a corresponding camera center [12]. + +$$ +\mathcal {L} _ {c} = \left\| c \times \hat {d} - \hat {m} \right\| _ {2} ^ {2}. \tag {16} +$$ + +Finally, our patch-level loss function is $\mathcal{L} = \mathcal{L}_{l2} + \alpha \mathcal{L}_c$ , where $\alpha$ is a balance term, which is empirically set as 0.2. Our training loss is simple yet effective compared to SCR methods [4, 14, 22], benefiting from the robustness of rays in optimization. + +# 3.5. RC-RANSAC + +Similar to SCR, it is also not guaranteed that all camera rays can be accurately predicted in CRR. Therefore, employing the RANSAC mechanism to eliminate outliers is a beneficial approach. Unlike the PnP solution, when solving for camera pose from rays, the rotation and translation constitute two linear problems. Hence, the RANSAC process can be separately applied. Given that the rotation solution is independent of translation, we introduce the RC-RANSAC (Algorithm 1) as a modification of the general RANSAC algorithm. Specifically, after obtaining the predicted camera rays $\hat{\mathcal{R}} = \{[\hat{d}_i,\hat{m}_i]\}_{i}^{N}$ , we initially use $R$ -RANSAC (cf. Algorithm 2 of the Suppl.) to determine the rotation $\hat{R}$ from the ray directions only. Subsequently, all camera rays are corrected using the obtained rotation matrix $\hat{R}$ to get better rays: $\hat{\mathcal{R}}^* = \{[\hat{R} d_i^c,\hat{m}_i]\}_{i}$ . The camera translation $\hat{c}$ is then derived from the corrected rays by another $C$ -RANSAC process (cf. Algorithm 3 of the Suppl.). Ultimately, combining the rotation and translation yields the final camera pose: $[\hat{R},\hat{c} ]$ for VL. + +# 4. Experiments + +# 4.1. Implementation Details + +Network Configuration. For the feature encoder $\mathcal{D}$ , we use the dinov2_vitb14_reg as our backbone, with an embedding dimension of 768 ( $D = 768$ ). The frequency of Fourier Encoding is set as 16, making $D_F = 64$ . On the other hand, the mapper head consists of 4 sub-mappers and a semantic attention model with 4 attention layers. Each sub-mapper has a 4-depth MLP with skip connections. The first 2 layers form the common mapping, and the other 2 layers output direction and moment embedding respectively + +Indoor-6 + +
CategoryMethodscene1 (cm/°)scene2a (%)scene3 (cm/°)scene4a (%)scene5 (cm/°)scene6 (cm/°)Average (cm/°)
FM-basedHloc [30]3.2/0.564.8-/-51.42.1/0.481.02.9/0.5
Hloc+SLD [9]2.9/0.468.73.4/0.662.71.9/0.381.042.7
2.1/0.4
APRPoseNet [15]159.0/7.50.0-/--141.0/9.30.0118.2/9.3
SCRDSAC* [3]12.3/2.118.77.9/0.928.013.1/2.319.740.7/6.7
ACE [4]13.6/2.124.96.8/0.731.98.1/1.333.014.7/2.3
NBE+SLD(E) [9]7.5/1.228.47.3/0.730.46.2/1.343.56.3/1.0
NBE+SLD [9]6.5/0.938.47.2/0.732.74.4/0.953.06.0/0.9
EGFS [20]-/-46.4-/-60.6-/-56.4-/-
EGFS-q [20]-/-58.5-/-59.1-/-67.0-/-
D2S [5]4.8/0.851.84.0/0.461.13.6/0.760.05.8/0.9
2.4/0.5
CRRDIMM5.5/1.544.44.2/0.866.74.4/1.161.67.6/1.5
DIMM-R5.5/1.545.23.3/0.677.44.2/1.162.56.3/1.5
3.2/0.8
+ +Table 2. Localization results on Indoor-6 [9]. We report the median errors in $cm$ for the position, degree $(^{\circ})$ for the orientation, and recall at $5cm / 5^{\circ}$ of the Indoor-6 dataset. The best results, the second best and our results are highlighted, except for the FM-based method. + +![](images/b1e2f8c023fb91cac51d7cd5b299efbfc91ce6f72b472aa92a8b19b794d461d5.jpg) + +![](images/98f2f06b7a021c1bffbc2ab4c799d93f723f58fb0808baf763710d83d418992d.jpg) + +![](images/cb74bb96a190dcec59e4d02dccefc5aac2a2f6d1118c2ea2b8a36b1a8e178d96.jpg) + +![](images/907551be19592bf81aabe16194e29e0d4cdc033081ce81d93199b375ccf0373b.jpg) + +![](images/be3d98060c8717ce5516bf203bb186714dcc5f8736a714c988303e5042231639.jpg) +Figure 4. The qualitative comparison between ray regression w/ or w/o global perception in the scene1 of Indoor-6. These results contain predicted and ground truth rays from the same cameras, along with their camera centers. The samples indicate that ray predictions w/o global perception (without the CLS and XY tokens, bottom) lack geometric consistency, although some individual rays are accurate. In contrast, incorporating global perception facilitates learning the geometric constraint of rays, resulting in precise ray regression (top). + +![](images/ed54ac662cbd8a7ef5e89ae8140dbacde14ec2b767e69b37dc4ab22c136d669a.jpg) + +![](images/5d93c7a76b63a7177a208b780b57549b9e4158c7eae2a556eca16418ec7cf8b3.jpg) + +![](images/cf7af604d1a6e14d37c2fa0a6969baed6b39d7a78621881208a1c4dba886d8be.jpg) + +(cf. Fig. 3). The hidden dimension $D_{h}$ is set as 256. More ablation experiments can be found in Sec. 4.3. + +Training and Hardware Details. Firstly, according to related work [13], we fine-tune the last two blocks of DINO in terms of CRR task for better performance. Particularly, we follow the encoder training protocol of [4] and [13], using the first 100 training scenes of ScanNet [7] to update the weights of unfrozen DINO blocks. The training is performed on 4 NVIDIA A800 GPUs using gradient accumulation [4] in parallel. Next, the mapper head is trained on 1 NVIDIA A800 GPU for each scene, with a training buffer of 26M encoder features. The large buffer size is important for CRR training, as most of patches have different rays. We observed a buffer size scaling law in experiments (cf. Suppl. B). We do 20 passes over the training buffer, utilizing a batch size of 5120. The optimization uses AdamW [21] with a learning rate between $5e^{-5}$ and $1e^{-4}$ and a 1 cycle schedule [37]. + +# 4.2.Quantitative Evaluation + +Datasets. We evaluate VL performance of our method on both indoor and outdoor datasets. For indoor scenes, we adopt the Indoor-6 dataset [9], which is a relatively large indoor dataset [9]. It comprises calibrated images from 6 multi-rooms indoor scenes captured over several days, leading to challenging VL. For outdoor scenes, the Map-Free [1] dataset is employed. We use its first 10 scenes, $s00000 \sim s00009$ , following [4]. Each scene contains images from two scans of an outdoor location. We split them into mapping and query sets. The ground truth camera poses are computed via SfM method [34]. + +Indoor Visual Localization. Firstly, the indoor VL experiments are conducted using the Indoor-6 dataset. In the experiments, we compare our DIMM method and its variation enhanced by RANSAC (DIMM-R) with both APR [15] + +
Scene (°/cm/%)SCRAPRCRR
DSAC*[3]ACE[4]marepo[6]marepos[6]DIMMDIMM-R
Throughput (fps)17.917.955.655.657.720.1
s000001.2/5.1/55.80.7/4.2/62.60.6/3.6/69.20.6/3.2/70.70.7/2.0/93.10.6/2.0/94.0
s000010.8/2.4/94.50.4/1.5/1000.9/1.8/96.00.8/1.6/99.21.1/1.6/96.81.1/1.6/96.8
s000020.8/2.4/69.50.5/2.1/77.70.6/2.3/77.10.5/2.1/78.20.9/3.7/71.20.9/3.0/77.1
s000030.6/2.1/75.80.4/2.4/73.60.9/2.5/80.30.8/2.1/84.50.6/2.4/92.80.6/2.2/93.3
s000041.1/6.3/40.40.7/5.8/45.30.8/5.3/45.90.7/5.1/47.60.7/2.4/82.60.7/2.3/96.2
s000050.7/3.6/64.40.5/2.4/73.30.7/2.7/72.20.5/2.6/77.30.8/2.5/86.20.8/2.1/87.9
s000061.0/4.4/53.40.8/4.5/54.80.8/4.2/58.50.6/3.6/63.80.8/5.6/44.30.8/3.9/64.8
s000070.7/7.8/16.90.6/7.6/19.01.0/6.3/25.10.8/6.1/27.50.6/3.6/69.50.5/3.5/71.1
s000081.2/3.5/68.90.7/3.1/71.60.7/2.3/77.40.7/1.9/82.80.7/2.7/85.10.7/2.4/87.9
s000090.8/3.1/81.20.4/1.2/99.70.8/1.8/94.90.9/1.6/95.10.7/1.3/97.70.7/1.2/97.4
Average0.8/4.1/62.10.6/3.5/67.70.8/3.3/69.70.7/3.0/72.70.8/2.8/81.90.7/2.4/86.6
+ +Table 3. Pose accuracy comparison on the outdoor MapFree [1] dataset. We report the median errors in degree (°) for the orientation, cm for the position, and recall at $5\mathrm{cm} / 5^{\circ}$ . The best results, the second best and our results are highlighted. + +
Recall-5°/5cm (%)s00000s00001s00002s00003s00004s00005Average
ACE [4]62.610077.773.645.373.372.1
DIMM-CRR (Ours)93.196.871.292.882.686.287.1
DIMM-SCR64.898.674.577.348.272.172.6
DIMM-APR3.212.45.73.60.32.64.6
+ +and SCR [3-5, 9, 20] approaches. The results are reported in Tab. 2. Additionally, we analyzed the corresponding mapping time and memory costs of various VL methods in Tab. 1, for a comprehensive comparison. DIMM demonstrates better or comparable performance than the previous state-of-the-art, e.g., achieving a recall@5°/5cm of 78.3 in scene6. Moreover, our method achieves generally better results than EGFS-q with "hard" multi-mapper combination, proving the efficacy of our "soft" mapper ensemble utilizing semantic guidance. The effectiveness of RC-RANSAC is also highlighted by a substantial precision improvement in scene2a from 66.7 to 77.4. Overall, the proposed DIMM-R achieves the best average recall in this dataset. Furthermore, our method enables mapping with reduced computational costs compared to D2S and is trained without 3D supervision, as illustrated in Tab. 1. These findings underscore the potential of our CRR-based method as a promising VL implementation. Some visualization results of DIMM can be found in Fig. 4 and Fig. 7 in the Suppl., where the 3D scene structure is preserved, demonstrating the privacy protection offered by our method. + +Outdoor Visual Localization. Our outdoor VL experiments are conducted on the Map-Free Dataset. In addition to the SCR-based ACE and DSAC*, we also compared the recent APR-based method, marepo [6], and its fine-tuned variation, marepos. The results, including median errors for rotation $(^{\circ})$ and translation $(cm)$ , as well as localization recalls at $5\mathrm{cm} / 5^{\circ}$ , are presented in Tab. 3. These results also come from methods with mapping time in Tab. 1. From the table, + +Table 4. Ablation Results about Camera Ray Regression (CRR). We evaluate three variations of our model to showcase the effectiveness of CRR, by modifying output to different VL formulations. + +
Feature EncoderMed. R Err. (°)Med. t Err. (cm)Recall-5°/5cm
T_i ∩ T[CLS] ∩ Txy_i0.85.644.3
T_i * T*[CLS] ∩ Txy_i0.86.039.7
T_i ∩ TG[41] ∩ Txy_i1.18.319.7
T_i ∩ T[CLS] ∩ XY_i0.97.627.2
T_i ∩ T[CLS]0.97.426.8
T_i1.111.517.2
+ +Table 5. Ablation Study about our Feature Encoder. The experiments are performed on the MapFree s00006 dataset. + +it is evident that our methods achieve the best overall results, outperforming both SCR and APR methods. The rotation accuracy of ACE, yet, surpasses our methods, indicating that local ambiguity poses a significant challenge for CRR methods and obtaining precise ray directions from image patches is challenging. However, it is observed that CRR methods facilitate better camera center estimation compared to APR and SCR methods. This improvement can be attributed to the fact that the camera center location integrates contributions from all camera rays, and the bounded ray error [27] enhances the robustness of the center estimation against noise. The efficacy of RC-RANSAC algorithm can also be verified in the table, as it leads to substantial improvements of VL performance, e.g., $44.3 \rightarrow 64.8$ of recall in s00006. Furthermore, we investigate the computational efficiency of our methods. The throughputs (fps) of all methods are reported in the table, revealing that the lightweight network and linear pose solver of DIMM achieve the best inference efficiency. + +# 4.3. Ablation Experiments + +Camera Ray Regression. One of our main contributions in this work is introducing Camera Ray Regression (CRR) into the VL task. Thus, we provide ablation study about this formulation in Tab. 4. In particular, we modify the last layer of DIMM to output scene coordinates (DIMM-SCR) or direct camera pose (DIMM-APR), and train these variations separately. The DIMM-SCR is trained with GDT, same as ACE[4]. The DIMM-APR, on the other hand, is trained in image level, as the direct pose vectors cannot + +![](images/3713c77838c9d31b05a8d92c2320b1795b2ee1aa602394120bf515105e0f23f7.jpg) +Figure 5. Visualization results of sub-mapper attention score. In the experiments, we observed that specific semantic entities elicit responses from different mappers. For instance, the fireplace at the top of the figure frequently results in high attention scores from mapper #3, while mapper #2 exhibits a low attention score. Conversely, mapper #3 seems do not "care" the sofa at the bottom, which leads to high scores of mapper #2 instead. + +be assigned to image patches. The experiments are performed on the MapFree dataset, with the localization recall reported. In Tab. 4, we can see that our CRR formulation significantly outperforms the SCR and APR under the same network architecture, indicating the contribution of our ray-parameterization. DIMM-SCR overcomes ACE slightly in average, but our model is tailored for CRR. DIMM-APR utilizes per-scene pose regression and achieves poor results, but this is consistent with the results in marepo [6]. This suggests the APR model may require much larger datasets to learn pose patterns. Overall, the results in Tab. 4 evaluate our main contribution: CRR formulation in VL. + +Feature Encoder. Based on our analysis of CRR, we require patch features to exhibit strong local description while also incorporating global perception and image position encoding to achieve disambiguation. To validate this, we conduct experiments on the MapFree s00006 dataset (cf. Tab. 5). Specifically, we modify the content of the patch feature, including features w/o global perception and image position encoding $(\mathcal{T}_i \otimes \mathcal{T}_{[\mathrm{CLS}]}$ and $\mathcal{T}_i)$ , to compare with the proposed version $(\mathcal{T}_i \otimes \mathcal{T}_{[\mathrm{CLS}]} \otimes \mathcal{T}_{\mathrm{xy}_i})$ . Also, we investigate the efficacy of DINO fine-tuning, and the performance of original DINO $(\mathcal{T}_i^* \otimes \mathcal{T}_{[\mathrm{CLS}]}^* \otimes \mathcal{T}_{\mathrm{xy}_i})$ is also reported. It is evident that global perception with image position encoding are crucial for CRR, leading to a significant accuracy improvement. The fine-tuning of DINO also improves the performance as well (39.7 vs. 44.3). Visualization comparison about the global perception is presented in Fig. 4. We can see the proposed global perception enables the learning of ray geometry, causing the rays to converge toward the camera center, thereby enhancing precision. We also conduct experiments for a variation without Fourier Encoding $(\mathcal{T}_i \otimes \mathcal{T}_{[\mathrm{CLS}]} \otimes \mathsf{XY}_i)$ . The results in the first and third rows of Tab. 5 show that Fourier Encoding effectively prevents the submergence of positional + +
Atten. Num.Med. R Err. (°)Med. t Err. (cm)Recall@5°/Scmsize (MB)
011.07.231.012.1
20.95.842.614.2
40.85.644.316.2
60.86.337.718.2
81.17.730.420.2
+ +1 Take average of the sub-mapper embeddings, eliminating attention layers. + +Table 6. Ablation Study about the Number of Semantic Attention Layers. The experiments are performed on the MapFree s00006 dataset. The efficacy of our semantic attention is evaluated. + +features. Finally, we explore another possible global perception way provided by [41] $(\mathcal{T}_i \otimes \mathcal{T}_G \otimes \mathcal{T}_{xy_i})$ , the feature for image retrieval. The results show that the replaced global feature fails to match the original efficacy, possibly because DINO's global feature is more compatible with its own local patch features, facilitating subsequent mapping learning. + +Mapper Head. In the same scene, we also conduct ablation experiments on the network structure of the Mapper Head. We first validate the efficacy of the semantic attention module and experiment with its number of layers. When the layer number is set to 0, we directly take the average of output from multiple sub-mappers. The results are presented in Tab. 6, containing median rotation/translation error, localization recall and model size. It is evident that utilizing semantic attention is more effective compared to simple average fusion, and the increase in network size is deemed acceptable. Additional ablation study about the MLP mappers can be found in Suppl. B. To showcase the effectiveness of the soft mapper assemble, we visualize some query images combined with their sub-mapper attention scores in Fig. 5. It can be observed that different sub-mappers get high attention scores to different objects in images. This indicates that our semantic attention mechanism allows the sub-mappers to "remember" specific semantic objects. + +# 5. Conclusion + +In this work, we propose regressing camera rays to perform visual localization. This overparameterized representation of camera model leads to high precision, and, more importantly, enhances privacy protection. In particular, we introduce DIMM as a learning model to regress camera rays from images. It utilizes DINO as a scene-agnostic encoder to output features that incorporate both local and global perception. A scene-specific mapper then regresses ray parameters from the features, involving a semantic attention module to merge results from multiple mappers. Finally, a ray-level RANSAC algorithm is proposed to improve the ray-to-pose accuracy. In experiments on both indoor and outdoor datasets, our methods demonstrate comparable or superior performance to current approaches, while ensuring privacy preservation. Acknowledgement. This work has been funded in part by the NSFC grants 62176156. + +# References + +[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Áron Monszpart, Victor Adrian Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In ECCV, 2022. 6, 7 +[2] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7525-7534, 2019. 1 +[3] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE transactions on pattern analysis and machine intelligence, 44 (9):5847-5865, 2021. 3, 6, 7 +[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7 +[5] Bach-Thuan Bui, Huy-Hoang Bui, Dinh-Tuan Tran, and JooHo Lee. D2s: Representing sparse descriptors and 3d coordinates for camera relocalization. IEEE Robotics and Automation Letters, 2024. 1, 2, 3, 6, 7 +[6] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1, 3, 7, 8 +[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6 +[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2 +[9] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 6, 7 +[10] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 3 +[11] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2025. 1, 3 +[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 3, 4, 5 +[13] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17658-17668, 2023. 4, 6 + +[14] Xudong Jiang, Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. R-score: Revisiting scene coordinate regression for robust large-scale visual localization. arXiv preprint arXiv:2501.01421, 2025. 3, 4, 5 +[15] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In 2016 IEEE international conference on Robotics and Automation (ICRA), pages 4762-4769. IEEE, 2016. 1, 3, 6, 13 +[16] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5974-5983, 2017. +[17] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 3 +[18] Sihang Li, Siqi Tan, Bowen Chang, Jing Zhang, Chen Feng, and Yiming Li. Unleashing the power of data synthesis in visual localization. arXiv preprint arXiv:2412.00138, 2024. 13 +[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 3 +[20] Ting-Ru Liu, Hsuan-Kung Yang, Jou-Min Liu, Chun-Wei Huang, Tsung-Chih Chiang, Quan Kong, Norimasa Kobori, and Chun-Yi Lee. Reprojection errors as prompts for efficient scene coordinate regression. In European Conference on Computer Vision, pages 286-302. Springer, 2025. 1, 2, 3, 6, 7 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6 +[22] Deben Lu, Wendong Xiao, Teng Ran, Liang Yuan, Kai Lv, and Jianbo Zhang. Attention-based accelerated coordinate encoding network for visual relocalization. In 2024 IEEE 7th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), pages 1675-1680, 2024. 1, 3, 4, 5 +[23] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 2, 3 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4 +[25] Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. arXiv preprint arXiv:2412.12392, 2024. 3 +[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DiNov2: Learning robust visual features without supervision, 2023. 4 + +[27] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schonberger. Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV), 2024. 2, 3, 7 +[28] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 1, 3 +[29] Francesco Pittaluga, Sanjeev J. Koppal, Sing Bing Kang, and Sudipta N. Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[30] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1, 2, 6 +[31] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1 +[32] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 752-765. Springer, 2012. 2 +[33] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE transactions on pattern analysis and machine intelligence, 39(9):1744-1756, 2016. 2 +[34] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6 +[35] Thomas Schops, Viktor Larsson, Marc Pollefeys, and Torsten Sattler. Why having 10,000 parameters in your camera model is better than twelve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2544, 2020. 1, 2, 3 +[36] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 3 +[37] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of residual networks using large learning rates. Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, 2019. 6 +[38] Pablo Speciale, Johannes L. Schonberger, Sing Bing Kang, Sudipta N. Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5488-5498, 2019. 1 + +[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 1, 3 +[40] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 4 +[41] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 3, 4, 7, 8 +[42] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3 +[43] Yinshuang Xu, Dian Chen, Katherine Liu, Sergey Zakharov, Rares Andrei Ambrus, Kostas Daniilidis, and Vitor Campagnolo Guizilini. $se(3)$ equivariant ray embeddings for implicit multi-view depth estimation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 3 +[44] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4 +[45] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20217-20226, 2024. 3 \ No newline at end of file diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/images.zip b/2025/Semantic-guided Camera Ray Regression for Visual Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ada13a3f6590b04c8c802b30b515cfff6be05776 --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b16ca0683bc694e92c854b189144906e479ab46f0fff09ed2a6736284cffe02c +size 571521 diff --git a/2025/Semantic-guided Camera Ray Regression for Visual Localization/layout.json b/2025/Semantic-guided Camera Ray Regression for Visual Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b64fe9258d62e3b8ed6ba89851b2e04374caef99 --- /dev/null +++ b/2025/Semantic-guided Camera Ray Regression for Visual Localization/layout.json @@ -0,0 +1,8479 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 506, + 121 + ], + "type": "text", + "content": "Semantic-guided Camera Ray Regression for Visual Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 144, + 493, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 144, + 493, + 185 + ], + "spans": [ + { + "bbox": [ + 117, + 144, + 493, + 185 + ], + "type": "text", + "content": "Yesheng Zhang, Xu Zhao* \nSchool of Automation and Intelligent Sensing, Shanghai Jiao Tong University \n{preacher, zhaoxu}@sjtu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 239, + 297, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 467 + ], + "type": "text", + "content": "This work presents a novel framework for Visual Localization (VL), that is, regressing camera rays from query images to derive camera poses. As an overparameterized representation of the camera pose, camera rays possess superior robustness in optimization. Of particular importance, Camera Ray Regression (CRR) is privacy-preserving, rendering it a viable VL approach for real-world applications. Thus, we introduce DINO-based Multi-Mappers, coined DIMM, to achieve VL by CRR. DIMM utilizes DINO as a scene-agnostic encoder to obtain powerful features from images. To mitigate ambiguity, the features integrate both local and global perception, as well as potential geometric constraint. Then, a scene-specific mapper head regresses camera rays from these features. It incorporates a semantic attention module for soft fusion of multiple mappers, utilizing the rich semantic information in DINO features. In extensive experiments on both indoor and outdoor datasets, our methods showcase impressive performance, revealing a promising direction for advancements in VL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 608 + ], + "type": "text", + "content": "Visual Localization (VL) involves estimating 6-degree-of-freedom camera pose of an image taken from a known scene, which is also known as camera relocalization. VL plays a vital role in many vision, robotic and graphic applications, such as augment reality, virtual reality and autonomous driving. Despite extensive research, achieving precise, robust, and privacy-preserving VL [28, 38] continues to present a significant challenge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "Current VL methods can be categorized into explicit and implicit methods based on their mapping approaches. Traditional explicit methods [2, 30, 31] involve an explicit 3D map, e.g., point clouds. They establish 2D-3D (image-to-map) correspondences by Feature Matching, and then estimate camera pose via PnP and RANSAC [10, 12]. While advanced methods [11, 39] leads to high accuracy, the stor" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 366, + 213, + 504, + 315 + ], + "blocks": [ + { + "bbox": [ + 366, + 213, + 504, + 315 + ], + "lines": [ + { + "bbox": [ + 366, + 213, + 504, + 315 + ], + "spans": [ + { + "bbox": [ + 366, + 213, + 504, + 315 + ], + "type": "image", + "image_path": "d1094e448dc67043bf342f4052a9dad3b42f9463fdd4b0c457783c941a957d08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": "Figure 1. The ambiguity challenge in Camera Ray Regression (CRR) for VL. Effective global and local disambiguations are critical for CRR. Global ambiguity arises from rays of different patches (cf. " + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "inline_equation", + "content": "r_2^1" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "inline_equation", + "content": "r_2^2" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": " in the figure), while local ambiguity occurs within the same patch but from different viewing angles (cf. " + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "inline_equation", + "content": "r_1^1" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "inline_equation", + "content": "r_2^2" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": "). Semantic information, e.g., the table in the figure to differentiate between chairs, along with geometry constraints (geometry properties of rays " + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "inline_equation", + "content": "\\{r_1^i\\}_i" + }, + { + "bbox": [ + 313, + 325, + 555, + 414 + ], + "type": "text", + "content": ") are essential for the disambiguations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 459 + ], + "type": "text", + "content": "age demands of explicit maps are troublesome especially in large-scale scenes. Additionally, explicit maps raise privacy concerns, restricting their applications in real-world settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 556 + ], + "type": "text", + "content": "Implicit methods, on the other hand, leverage learning models as implicit and light-weight maps, which mainly include two frameworks: Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR). APR directly utilizes networks to regress camera poses (rotations and translations) from images [6, 15], offering enhanced privacy protection. However, the compact nature of absolute pose makes it vulnerable to noise, thereby leading to accuracy issue [6]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "To improve precision, employing an overparameterized representation for the camera pose is crucial [35]. Thus, SCR framework [22, 41] is proposed, relying on 2D-3D matches as an indirect yet robust camera pose representation, akin to its explicit counterparts. ACE [4] is a milestone within this framework, utilizing simple networks to regress pixel-aligned scene coordinates. It introduces a scene-centric patch-level training strategy, i.e., Gradient Decorrelation Training (GDT), enabling both efficient and precise mapping. Nevertheless, the unbounded 3D point error is unstable in training, necessitating manual hyperparameter tuning or additional outlier rejection modules [5, 20]. Besides, SCR methods still run the risk of 3D privacy breaches, as they" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 702, + 137, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 137, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 137, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25639" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 290, + 175 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 290, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 290, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 290, + 175 + ], + "type": "image", + "image_path": "94306d1c810ff5d230ee5804b658da387401ef5a043ed499d24347ee427df6fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 295, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 262 + ], + "type": "text", + "content": "Figure 2. The proposed DIMM approach performing CRR for VL. This approach utilizes DINO as a scene-agnostic feature encoder to output carefully designed patch-level features. Then, a scene-specific mapper head integrates multi-mapper results under semantic guidance, through a semantic attention module. After the scene-centric training technology, DIMM is capable of predicting camera ray parameters for accurate and privacy-preserving VL." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 269, + 235, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 235, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 235, + 281 + ], + "type": "text", + "content": "disclose detailed 3D scene information [29]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 510 + ], + "type": "text", + "content": "Recently, a promising camera pose representation, camera ray [27, 35], has garnered increased attention in the vision field. Its overparameterization is beneficial for patch-level learning [23, 44], aligning well with the effective GDT in VL. This inspires us to introduce camera rays in the VL tasks. In particular, we propose adopting a learning model to regress patch-level camera rays from the image, termed as Camera Ray Regression (CRR). Then, the camera pose can be achieved by solving two linear problems [44]. This implementation of VL could strike a balance between efficacy and privacy preservation. Firstly, it belongs to the implicit mapping category, incurring low storage costs. As a overparameterized pose representation, it can yield improved precision [35] than APR. Also, the ray error is more stable than the unbounded 3D point error [27] in optimization of model training. Hence, CRR obviates the need for intricate training setting like SCR methods [5, 20, 41]. More importantly, it is hard to recover 3D scene information from camera rays, leading to a privacy-preserving VL implementation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "content": "However, challenges exist in introducing camera rays in VL, primarily stemming from the issue of ambiguity (cf. Fig. 1). This ambiguity can generally be divided into two levels. The first is global ambiguity, referring to distinguish rays of different image patches, particularly in repetitive or textureless scenes. The similar ambiguity is encountered in SCR [4, 41], where leveraging powerful local features, such as semantic features [20], has proven effective. The second is local ambiguity, unique to CRR, which involves differentiating rays from the same image patch with different view angles. The ambiguity becomes troublesome for patches with continuous depth, where changes in view direction result in minor differences in patch content but obvious variations in the corresponding rays. To eliminate this ambiguity, we emphasize two observations on the CRR task. 1) The entire image often contains patches with depth discontinuities. Their appearance changes under different" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "viewing angles are significant enough to resolve the local ambiguity. Thus, the image-level perception is important for CRR disambiguation, which is also proven to be beneficial for handling global repetitiveness [41]. 2) The rays within the same image are governed by the same camera pose. This geometric constraint allows confident rays to correct uncertain ones. Meanwhile, the smoothness of the patchray mapping [44] makes this geometric constraint learnable. Therefore, the integration of image-level perception, geometric constraint and robust local features is critical for the global and local disambiguation of CRR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 556, + 456 + ], + "type": "text", + "content": "Based on the above analysis, this work presents DINO-based Multi-Mappers (DIMM, Fig. 2), as the first implementation of CRR in VL. Following the scene-centric design of ACE, DIMM consists of a scene-agnostic encoder and a scene-specific head, but possessing several key modules tailed for CRR. Specifically, we utilize DINO [8] as the feature encoder, utilizing its powerful semantic perception for disambiguation. It provides robust patch tokens as local features while supporting CLS tokens for image-level perception. We further incorporate Fourier Encoding [24] of the image position for each patch into the features (XY tokens), as the potential geometric constraint of camera rays are associated with patch locations. Furthermore, we propose a scene-specific mapper head which adopts a semantic attention module to softly fuse ray regression results from multiple Multi-Layer Perceptron (MLP)-based mappers. This head effectively utilizes the inherit semantic prior of the DINO features, to enhance ray regression performance especially in large scenes. Finally, we propose a ray-level RANSAC algorithm for accurate pose estimation, decomposing the rotation and translation calculation [44]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 457, + 504, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 457, + 504, + 468 + ], + "spans": [ + { + "bbox": [ + 326, + 457, + 504, + 468 + ], + "type": "text", + "content": "Our contributions are summaries as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 473, + 555, + 616 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 473, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 473, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 315, + 473, + 553, + 509 + ], + "type": "text", + "content": "1. To the best of our knowledge, we are the first to introduce the camera ray into VL, with a tailored network design to achieve query image poses by camera ray regression." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 509, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 509, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 509, + 553, + 555 + ], + "type": "text", + "content": "2. We propose carefully designed scene-agnostic features to handle local and global disambiguation in CRR, along with a semantic-guided mapper head that enables a soft multi-mapper ensemble." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 556, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 555, + 616 + ], + "type": "text", + "content": "3. Comprehensive experiments on indoor and outdoor datasets demonstrate the effectiveness of our method, achieving results that rival or surpass the existing state-of-the-art. Additionally, extensive ablation studies validate the contributions of each module." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 630, + 400, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 400, + 642 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 400, + 642 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Feature Matching-based VL. Traditional VL methods [30, 32, 33] rely on Feature Matching (FM) to determine the position and direction of cameras. These techniques aim to match 3D scene points with 2D image points, enabling the camera pose calculation through " + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PnP} + \\mathrm{RANSAC}" + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "-based" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25640" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 131, + 70, + 479, + 189 + ], + "blocks": [ + { + "bbox": [ + 131, + 70, + 479, + 189 + ], + "lines": [ + { + "bbox": [ + 131, + 70, + 479, + 189 + ], + "spans": [ + { + "bbox": [ + 131, + 70, + 479, + 189 + ], + "type": "image", + "image_path": "3e18d4064dc6c3b2e7eba9d92fce0f63485b0302b383fdf56a3b0a6343f91569.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 3. The Multi-Mapper network with semantic attention as our scene-specific head. The patch-level features are first projected to geometry and semantic embeddings. The geometry embedding is fed into multiple MLP-based mappers to derive ray embeddings. These embeddings are then fused under the guidance of the semantic embedding to achieve the camera rays, through cross attention modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 236, + 297, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 297, + 356 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 297, + 356 + ], + "type": "text", + "content": "algorithms [10, 12]. As a result, these approaches necessitate the storage of explicit maps containing 3D points and their feature descriptors. By leveraging robust matching algorithms [11, 19, 39, 42, 45], FM-based localization can deliver precise outcomes. Nevertheless, explicit maps raises concerns about privacy leaks [28]. Furthermore, as the scene size expands, managing the increasingly large explicit maps becomes challenging, leading to storage issues. In contrast, our method employs implicit maps that have lower storage requirements and privacy protection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 361, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 577 + ], + "type": "text", + "content": "Scene Coordinate Regression. Similar to FM-based VL, Scene Coordinate Regression (SCR) also accomplishes camera poses through 3D-2D matching [3]. However, SCR directly employs network to regress pixel-aligned 3D scene coordinates, which addresses the map storage issue through implicit mapping. ACE [4], a prominent approach within SCR, introduces a gradient decorrelation training technique enabling mapping within 5 minutes using a 4MB network. Subsequent methods [5, 14, 20, 22, 41] built upon ACE further enhance VL accuracy mainly by solid feature selection. Nevertheless, the limited network size of ACE pose challenges in fitting large scenes, necessitating spatial segmentation of the large scene for sub-map construction. Despite its efficacy, this approach may result in ambiguous localization at the boundaries of segmented sub-scenes. At the same time, it is noteworthy that the SCR methods output the 3D information of scenes, i.e., 3D coordinates, thereby leading to a potential risk of privacy leak." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "Absolute Pose Regression. Absolute Pose Regression (APR) also adopts implicit mapping, but directly utilizes networks to predict the rotation and translation of cameras. However, this compact parameter representation of camera pose is highly sensitive to noises [12, 44]. Regressing pose parameters from images, thus, is challenging [15-17, 36], and the accuracy is insufficient to compare with methods using over-parameterized representations, e.g., 2D-3D correspondences. Recent APR approach [6] opts to predict the camera pose based on the SCR output, resulting in favorable accuracy. However, this success is heavily reliant" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 236, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 285 + ], + "type": "text", + "content": "on the precision of SCR methods. In contrast, our method employs camera rays to depict the camera pose. This overparameterized representation is more conducive to learning optimization [25, 27, 44], enabling superior accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 286, + 556, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 556, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 556, + 490 + ], + "type": "text", + "content": "Camera Ray in Vision. Recently, camera rays have emerged as a promising camera parametrization, representing a generic camera configuration [35]. In recent work [23, 43, 44], camera rays have been proven to be well-suited for patch-level learning, with applications in multiple vision tasks. Also, in 3D reconstruction works [25, 27], camera rays yield more robust geometric optimization than classical correspondence-based representations, e.g., 3D-2D matches. This advantage stems from their overparameterized nature and constrained ray error. Likewise, our approach is the first to introduce camera rays in VL, performing Camera Ray Regression (CRR) to achieve camera poses from images. Its patch-level property aligns well with the Gradient Deceleration Training [4], and the bounded ray error is beneficial for optimization in VL Learning as well. Furthermore, the inherent privacy-preserving attribute of CRR positions it as a promising direction for VL implementations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 370, + 511 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 640 + ], + "type": "text", + "content": "In this section, we first present the formulation of CRR for the VL task in Sec. 3.1. Then the proposed DIMM method is described in detail. DIMM adopts a scene-centric setup [4, 14, 41] by dividing the network into a scene-agnostic encoder (Sec. 3.2) and a scene-specific head (Sec. 3.3). Additionally, the mapping process incorporates GDT [4] and utilizes a specially designed loss function (Sec. 3.4). To further enhance accuracy, we propose a RANSAC-based algorithm for estimating camera pose from rays in Sec. 3.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 648, + 440, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 440, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 440, + 659 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Employing CRR for VL involves two main steps: 1) regressing the patch-level camera ray parameters from the input image and 2) calculating the camera pose from the ray parameters. Given an input image " + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": ", it is first" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25641" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "text", + "content": "divided uniformly into a set of patches " + }, + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "inline_equation", + "content": "(p_i)" + }, + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "inline_equation", + "content": "s\\times s" + }, + { + "bbox": [ + 55, + 72, + 285, + 84 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 92, + 295, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 92, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 134, + 92, + 295, + 108 + ], + "type": "interline_equation", + "content": "\\left\\{p _ {i} \\in \\mathbb {R} ^ {s \\times s \\times C} \\right\\} _ {i = 1} ^ {N}, \\tag {1}", + "image_path": "b5021f82d26418f408d9d16acb62cc71717802b622e109d3e8ae60f28cc062d5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "inline_equation", + "content": "N = HW / s^2" + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "text", + "content": ". Subsequently, the model " + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "text", + "content": " regresses the ray parameters corresponding to the center points of image patches, expressed in the world coordinate system, " + }, + { + "bbox": [ + 55, + 116, + 296, + 166 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} = \\{r_i\\in \\mathbb{R}^6\\}_{i = 1}^N" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 176, + 295, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 176, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 176, + 295, + 190 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {R}} = \\mathcal {F} (I), \\tag {2}", + "image_path": "0c2faf4b8b961f84f8cb89eb024ba2c534fd2b483e4edd32947e25b92cc04d7c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": "where the ray parameters are expressed in Plücker coordinates [12]: " + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "inline_equation", + "content": "r_i = [d_i, m_i]" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "inline_equation", + "content": "d_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": " denotes the ray direction and " + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "inline_equation", + "content": "m_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": " represents the ray moment. Upon obtaining sufficient rays (" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\geq 2" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": "), the camera pose can be acquired by solving two distinct linear problems corresponding to rotation and translation, respectively [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "text", + "content": "Firstly, in terms of rotation, given the camera intrinsic, the Plücker coordinates of all camera rays in the camera coordinate system can be achieved, denoted as " + }, + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "inline_equation", + "content": "\\{r_i^c = [d_i^c, 0]\\}_i^N" + }, + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "text", + "content": ". The rotation matrix " + }, + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 269, + 296, + 340 + ], + "type": "text", + "content": ", thus, corresponds to the transformation from the ray directions in the camera coordinate system to those in the world coordinate system:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 350, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 350, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 295, + 376 + ], + "type": "interline_equation", + "content": "\\hat {R} = \\arg \\min _ {| | R | | = 1} \\sum_ {i} \\left| \\left| R d _ {i} ^ {c} - d _ {i} \\right| \\right| _ {2} ^ {2}. \\tag {3}", + "image_path": "3d1095154207a1ac5087e23915a115249a764da5633f87b416f75af074d8cfaa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "text", + "content": "Subsequently, the translation, also known as the world position of the camera optical center denoted as " + }, + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "text", + "content": ", is determined. Since " + }, + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 380, + 296, + 428 + ], + "type": "text", + "content": " is the intersection point of all camera rays, the optimization problem can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 438, + 295, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 438, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 295, + 463 + ], + "type": "interline_equation", + "content": "\\hat {c} = \\arg \\min _ {c} \\sum_ {i} \\| c \\times d _ {i} - m _ {i} \\| _ {2} ^ {2}. \\tag {4}", + "image_path": "0c57d07bea54096b9c5bf5b80f1763f531845ea6411c3bf835501eeb461648c1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 471, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 495 + ], + "type": "text", + "content": "Through these two linear problems, the camera pose " + }, + { + "bbox": [ + 55, + 471, + 295, + 495 + ], + "type": "inline_equation", + "content": "([\\hat{R} |\\hat{c} ])" + }, + { + "bbox": [ + 55, + 471, + 295, + 495 + ], + "type": "text", + "content": " can be deduced from the camera rays." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 503, + 157, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 157, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 157, + 514 + ], + "type": "text", + "content": "3.2. Feature Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": "Given the patch-wise independent GDT, it entails that the scene-specific head must learn the mapping from patch-level features to ray parameters. To ensure the learnability of this, ambiguity within patch features, particularly the local ambiguities, needs to be addressed. Thus, based on our prior analyze, patch-level features must possess both robust local representation and image-level perception. Hence, we opt for DINO [26] as our encoder backbone. From an input image " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": ", it can generate patch-level feature tokens, " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\{p_i\\}_i^N" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": ", and provide the CLS token for image-level perception. According to prior work [13], fine-tuning DINO's last two blocks can effectively enhance the performance of VL tasks. Consequently, we also fine-tune DINO using the ACE Encoder training method [4], resulting in " + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 521, + 296, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 129, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {T} _ {i} \\right\\} _ {i} ^ {N}, \\mathcal {T} _ {[ \\mathrm {C L S} ]} = \\mathcal {D} (I), \\tag {5}", + "image_path": "0ac98f303d6426f8e529c71ab57896e3ed756c27a9e169512bc7f2f029f89222.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 72, + 474, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 474, + 84 + ], + "type": "text", + "content": "Algorithm 1 RC-RANSAC Algorithm" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 87, + 536, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 536, + 100 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 536, + 100 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 315, + 87, + 536, + 100 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i = 1}^N" + }, + { + "bbox": [ + 315, + 87, + 536, + 100 + ], + "type": "text", + "content": " # camera rays from DIMM" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "spans": [ + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "inline_equation", + "content": "\\hat{R} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "inline_equation", + "content": "\\hat{c} \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 315, + 100, + 504, + 113 + ], + "type": "text", + "content": "; # camera pose" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 114, + 555, + 176 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "text", + "content": "1: Get Rotation by RANSAC: " + }, + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "inline_equation", + "content": "\\hat{R} \\gets R" + }, + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "text", + "content": "-RANSAC(" + }, + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "inline_equation", + "content": "\\{\\hat{d}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 319, + 114, + 555, + 138 + ], + "type": "text", + "content": "); cf. Algorithm 2 of the Suppl." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 138, + 491, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 138, + 491, + 152 + ], + "spans": [ + { + "bbox": [ + 319, + 138, + 491, + 152 + ], + "type": "text", + "content": "2: Ray Correction: " + }, + { + "bbox": [ + 319, + 138, + 491, + 152 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}\\gets \\{[\\hat{R}\\hat{d}_i^c,\\hat{m}_i]\\}_{i = 1}^N;" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "text", + "content": "3: Get Camera Center by RANSAC: " + }, + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "inline_equation", + "content": "\\hat{c} \\gets C" + }, + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "text", + "content": "-RANSAC(" + }, + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}" + }, + { + "bbox": [ + 319, + 152, + 555, + 176 + ], + "type": "text", + "content": "); cf. Algorithm 3 of the Suppl." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": " represents the patch token corresponding to " + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": " locally, " + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{[\\mathrm{CLS}]}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": " denotes the CLS token with global perception, " + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": " signifies the number of patches, and " + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 199, + 553, + 246 + ], + "type": "text", + "content": " is the token dimension." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "content": "Meanwhile, the geometric constraints of the rays from the same camera are crucial for eliminating local ambiguities. Thus, the network must also explicitly specify the spatial image location of patches during training. To further prevent the positional feature from being overshadowed by other dimensions, we draw inspiration from Nerf [24] and enhance this feature using Fourier Encoding " + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "inline_equation", + "content": "(FE)" + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "content": " [40]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 365, + 338, + 555, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 338, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 365, + 338, + 555, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\mathrm {x y} _ {i}} = F E \\left(\\left\\{x _ {i}, y _ {i} \\right\\} _ {i}\\right) \\in \\mathbb {R} ^ {N \\times D _ {F}} \\tag {6}", + "image_path": "5d718194e62af277e2f91c9429af945cbcdd664d82e9070ccf6fb52fdbe55fbb.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\{x_{i},y_{i}\\}_{i}" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": " represents the positions of image patches " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\{p_i\\}_i" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": " in the original image, and " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "D_F" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": " is the dimension of the Fourier encoding, set to 16. Finally, the scene-independent patch-level feature is a combination of the aforementioned three features. For each patch " + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 361, + 554, + 421 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 356, + 430, + 555, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 430, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 356, + 430, + 555, + 447 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {i} ^ {*} := \\mathcal {T} _ {i} \\otimes \\mathcal {T} _ {[ C L S ]} \\otimes \\mathcal {T} _ {\\mathrm {x y} _ {i}} \\in \\mathbb {R} ^ {2 D + D _ {F}}, \\tag {7}", + "image_path": "55d63e2404ecef0bbd7e23b190bfaef1b735087e380df3f032231f10f88b8909.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 449, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 485 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 449, + 555, + 485 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 449, + 555, + 485 + ], + "type": "text", + "content": " denotes concatenation along the last dimension. This carefully designed feature to address ambiguity serves as the foundation of our CRR method." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 493, + 403, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 403, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 403, + 506 + ], + "type": "text", + "content": "3.3. Mapper Head" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "The scene-specific mapper head is responsible for transforming input features to ray parameters in CRR. A powerful feature encoder allows similar transformations to be achieved with simple MLPs [4, 14, 22, 41]. However, these mappers struggles to handle large scenes due to its limited capacity. To address the issue, spatial clustering is applied to divide the large scene [4, 41], with multiple MLPs fitting different sub-scenes independently. While effective, this hard scene partitioning can lead to ambiguous results near the sub-scene boundaries. In contrast, we adopt a soft map partitioning. As its core, we fuse the outputs of multiple sub-mappers with soft semantic guidance. Through end-to-end training, the sub-mappers are able to adaptively focus on different semantic regions. Particularly, we employ two MLP projection layers " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "(P_{s}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "P_{g})" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " to first project the input feature into a geometric embedding " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i^g" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and a semantic embedding " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i^s" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_h" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " hidden dimension respectively." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25642" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 295, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 175 + ], + "type": "table", + "html": "
CategoryMethodMapping w/ 3D info.Mapping TimeMapping Size
FM-basedhloc(SP+SG)No1.5 hours~2GB
SCRDSAC*Yes15 hours28MB
ACENo5 minutes4MB
ACE (×4)No20 minutes14MB
EGFSNo12 minutes4.5MB
EGFS (dual)No21 minutes9MB
D2SYes~9.4 hours22 MB
APRmarepoNo5 minutes98.9MB
mareporeplasNo~15 minutes98.9MB
CRRDIMM (Ours)No~1 hour16.2 MB
", + "image_path": "76eb49d63a0a98a85ebe3774d2a05a1be9cf5b03e4e02a25ced85bc57b3cfacc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 183, + 296, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 296, + 228 + ], + "type": "text", + "content": "Table 1. The mapping size and time comparison between VL methods from different categories. As a new flavor of VL, our DIMM achieves mapping time and memory size comparable to other methods while ensuring privacy protection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 128, + 243, + 296, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 243, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 128, + 243, + 296, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {g} = P _ {g} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {8}", + "image_path": "36ef520798c9ab93b3c72a9bea166961947c58cad57b8a06b6e96b314ac47d7b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 277, + 296, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 277, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 129, + 277, + 296, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {T} ^ {s} = P _ {s} \\left(\\mathcal {T} _ {i} ^ {*}\\right) \\in \\mathbb {R} ^ {D _ {h}}, \\tag {9}", + "image_path": "b93ceabf2ed51349bb30b232245b4687f87adf940618d61a5128718e48c36f18.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": "where we omit the indices of the patches for simplification. Next, the sub-mappers " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathcal{M}_j\\}_{j=1}^M" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": ", composed of " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": " MLPs with depth " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "D_M" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": ", take the geometric embedding " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^g" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": " ray embeddings " + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_j^r\\}_{j=1}^M" + }, + { + "bbox": [ + 55, + 298, + 296, + 349 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 359, + 296, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 359, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 126, + 359, + 296, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {j} ^ {r} = \\mathcal {M} _ {j} \\left(\\mathcal {T} ^ {g}\\right) \\in \\mathbb {R} ^ {D _ {h}}. \\tag {10}", + "image_path": "785c3bae4da9a41a202c37eab0ee7f09ced6372e6d0a5dba3a533022df293e4a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "text", + "content": "Afterwards, in the sequence of L cross attention layers " + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "inline_equation", + "content": "(\\mathcal{A} = \\{\\mathcal{A}_l\\}_{l=1}^L" + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "text", + "content": ", termed as Semantic Attention), we replicate the semantic embedding " + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i^s)" + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "text", + "content": " times along the first dimension to serve as the query " + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 380, + 296, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 438, + 295, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 438, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 119, + 438, + 295, + 453 + ], + "type": "interline_equation", + "content": "\\otimes \\left\\{\\mathcal {T} ^ {s} \\right\\} ^ {M} \\mapsto q \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {11}", + "image_path": "3d98d83325989076db90bca7c7f43cfa11b4bb24b396605fac5477bfee0a36e4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 464, + 297, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 297, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 297, + 489 + ], + "type": "text", + "content": "All ray embeddings are concatenated along the first dimension to form both the key and value of the first layer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 500, + 296, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 500, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 110, + 500, + 296, + 516 + ], + "type": "interline_equation", + "content": "\\otimes \\left\\{\\mathcal {T} _ {j} ^ {r} \\right\\} _ {j} ^ {M} \\mapsto k _ {0}, v _ {0} \\in \\mathbb {R} ^ {M \\times D _ {h}}. \\tag {12}", + "image_path": "ce01ba4afd19f92d36a016b19a65a913ff70cb328e57bfbce32b309246654970.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 559 + ], + "type": "text", + "content": "By iteratively updating the key and value through the attention layers, we fuse the results of all sub-mappers by weighting them using the semantic guidance from DINO." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 120, + 569, + 295, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 569, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 120, + 569, + 295, + 582 + ], + "type": "interline_equation", + "content": "A _ {l} \\left(q, k _ {l}, v _ {l}\\right) \\mapsto v _ {l + 1}, k _ {l + 1}. \\tag {13}", + "image_path": "9461742bb99b7204925115b1ac8f1d816f23f4b137c021111661f0f53311e2e6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "text", + "content": "The final value is first reshaped: " + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "inline_equation", + "content": "v_{L} \\in \\mathbb{R}^{M \\times D_{h}} \\to \\mathbb{R}^{MD_{h}}" + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "text", + "content": ", followed by projection to obtain the ultimate ray embedding: " + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_f^r = P_f(v_L) \\in \\mathbb{R}^{D_h}" + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "text", + "content": ". Finally, the ray parameters " + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "inline_equation", + "content": "\\hat{r} = [\\hat{d},\\hat{m}]" + }, + { + "bbox": [ + 55, + 591, + 296, + 656 + ], + "type": "text", + "content": " are obtained through a straightforward MLP projection layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 135, + 665, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 665, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 135, + 665, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\hat {r} = P _ {r} \\left(\\mathcal {T} _ {f} ^ {r}\\right) \\in \\mathbb {R} ^ {6}. \\tag {14}", + "image_path": "c7a9796c66ad43a34894a32195c7692958604f33d03c323a0638e077ad458235.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "Based on this head, we can achieve the corresponding camera ray from a patch-level embedding from our feature encoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 72, + 402, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 402, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 402, + 85 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "text", + "content": "Adopting the patch-level Gradient Decorrelation Training, our loss function operates independently between patches. First, we utilize the L2 distance between the ground truth " + }, + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "inline_equation", + "content": "(r_{GT})" + }, + { + "bbox": [ + 313, + 89, + 555, + 138 + ], + "type": "text", + "content": " and predicted rays as the optimization target." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 144, + 555, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 144, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 392, + 144, + 555, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l 2} = \\left\\| r _ {G T} - \\hat {r} \\right\\| _ {2} ^ {2}. \\tag {15}", + "image_path": "f695354a1d3283eb50fcdec7b3a782ae2757f9a7cd1404735d39a85528825246.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 167, + 555, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 215 + ], + "type": "text", + "content": "Besides, to enhance the model awareness of geometric constraints, we incorporate a center loss " + }, + { + "bbox": [ + 313, + 167, + 555, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 167, + 555, + 215 + ], + "type": "text", + "content": " based on the geometric properties of camera rays. The principle is any camera ray passes through a corresponding camera center [12]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 390, + 221, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 221, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 390, + 221, + 555, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\left\\| c \\times \\hat {d} - \\hat {m} \\right\\| _ {2} ^ {2}. \\tag {16}", + "image_path": "dd05236cca2a3febfe4bc177a049740917fc532b3b948bc6b4f925292545a75d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "content": "Finally, our patch-level loss function is " + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{l2} + \\alpha \\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 243, + 555, + 304 + ], + "type": "text", + "content": " is a balance term, which is empirically set as 0.2. Our training loss is simple yet effective compared to SCR methods [4, 14, 22], benefiting from the robustness of rays in optimization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 310, + 402, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 402, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 402, + 322 + ], + "type": "text", + "content": "3.5. RC-RANSAC" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": "Similar to SCR, it is also not guaranteed that all camera rays can be accurately predicted in CRR. Therefore, employing the RANSAC mechanism to eliminate outliers is a beneficial approach. Unlike the PnP solution, when solving for camera pose from rays, the rotation and translation constitute two linear problems. Hence, the RANSAC process can be separately applied. Given that the rotation solution is independent of translation, we introduce the RC-RANSAC (Algorithm 1) as a modification of the general RANSAC algorithm. Specifically, after obtaining the predicted camera rays " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}} = \\{[\\hat{d}_i,\\hat{m}_i]\\}_{i}^{N}" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": ", we initially use " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": "-RANSAC (cf. Algorithm 2 of the Suppl.) to determine the rotation " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": " from the ray directions only. Subsequently, all camera rays are corrected using the obtained rotation matrix " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "\\hat{R}" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": " to get better rays: " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{R}}^* = \\{[\\hat{R} d_i^c,\\hat{m}_i]\\}_{i}" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": ". The camera translation " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": " is then derived from the corrected rays by another " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": "-RANSAC process (cf. Algorithm 3 of the Suppl.). Ultimately, combining the rotation and translation yields the final camera pose: " + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "inline_equation", + "content": "[\\hat{R},\\hat{c} ]" + }, + { + "bbox": [ + 313, + 327, + 556, + 558 + ], + "type": "text", + "content": " for VL." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 567, + 395, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 395, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 395, + 580 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 586, + 447, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 447, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 447, + 599 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Network Configuration. For the feature encoder " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", we use the dinov2_vitb14_reg as our backbone, with an embedding dimension of 768 (" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "D = 768" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "). The frequency of Fourier Encoding is set as 16, making " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "D_F = 64" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". On the other hand, the mapper head consists of 4 sub-mappers and a semantic attention model with 4 attention layers. Each sub-mapper has a 4-depth MLP with skip connections. The first 2 layers form the common mapping, and the other 2 layers output direction and moment embedding respectively" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25643" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 83, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 290, + 74, + 318, + 81 + ], + "lines": [ + { + "bbox": [ + 290, + 74, + 318, + 81 + ], + "spans": [ + { + "bbox": [ + 290, + 74, + 318, + 81 + ], + "type": "text", + "content": "Indoor-6" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 83, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 83, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 83, + 553, + 228 + ], + "type": "table", + "html": "
CategoryMethodscene1 (cm/°)scene2a (%)scene3 (cm/°)scene4a (%)scene5 (cm/°)scene6 (cm/°)Average (cm/°)
FM-basedHloc [30]3.2/0.564.8-/-51.42.1/0.481.02.9/0.5
Hloc+SLD [9]2.9/0.468.73.4/0.662.71.9/0.381.042.7
2.1/0.4
APRPoseNet [15]159.0/7.50.0-/--141.0/9.30.0118.2/9.3
SCRDSAC* [3]12.3/2.118.77.9/0.928.013.1/2.319.740.7/6.7
ACE [4]13.6/2.124.96.8/0.731.98.1/1.333.014.7/2.3
NBE+SLD(E) [9]7.5/1.228.47.3/0.730.46.2/1.343.56.3/1.0
NBE+SLD [9]6.5/0.938.47.2/0.732.74.4/0.953.06.0/0.9
EGFS [20]-/-46.4-/-60.6-/-56.4-/-
EGFS-q [20]-/-58.5-/-59.1-/-67.0-/-
D2S [5]4.8/0.851.84.0/0.461.13.6/0.760.05.8/0.9
2.4/0.5
CRRDIMM5.5/1.544.44.2/0.866.74.4/1.161.67.6/1.5
DIMM-R5.5/1.545.23.3/0.677.44.2/1.162.56.3/1.5
3.2/0.8
", + "image_path": "10c381c461516ed3021300597feaa254733a89ace05959c76eaf9789d706343a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "text", + "content": "Table 2. Localization results on Indoor-6 [9]. We report the median errors in " + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "inline_equation", + "content": "cm" + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "text", + "content": " for the position, degree " + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "text", + "content": " for the orientation, and recall at " + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "inline_equation", + "content": "5cm / 5^{\\circ}" + }, + { + "bbox": [ + 56, + 236, + 553, + 258 + ], + "type": "text", + "content": " of the Indoor-6 dataset. The best results, the second best and our results are highlighted, except for the FM-based method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 108, + 259, + 220, + 350 + ], + "blocks": [ + { + "bbox": [ + 108, + 259, + 220, + 350 + ], + "lines": [ + { + "bbox": [ + 108, + 259, + 220, + 350 + ], + "spans": [ + { + "bbox": [ + 108, + 259, + 220, + 350 + ], + "type": "image", + "image_path": "b1e2f8c023fb91cac51d7cd5b299efbfc91ce6f72b472aa92a8b19b794d461d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 224, + 259, + 314, + 350 + ], + "blocks": [ + { + "bbox": [ + 224, + 259, + 314, + 350 + ], + "lines": [ + { + "bbox": [ + 224, + 259, + 314, + 350 + ], + "spans": [ + { + "bbox": [ + 224, + 259, + 314, + 350 + ], + "type": "image", + "image_path": "98f2f06b7a021c1bffbc2ab4c799d93f723f58fb0808baf763710d83d418992d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 259, + 407, + 350 + ], + "blocks": [ + { + "bbox": [ + 318, + 259, + 407, + 350 + ], + "lines": [ + { + "bbox": [ + 318, + 259, + 407, + 350 + ], + "spans": [ + { + "bbox": [ + 318, + 259, + 407, + 350 + ], + "type": "image", + "image_path": "cb74bb96a190dcec59e4d02dccefc5aac2a2f6d1118c2ea2b8a36b1a8e178d96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 410, + 259, + 500, + 350 + ], + "blocks": [ + { + "bbox": [ + 410, + 259, + 500, + 350 + ], + "lines": [ + { + "bbox": [ + 410, + 259, + 500, + 350 + ], + "spans": [ + { + "bbox": [ + 410, + 259, + 500, + 350 + ], + "type": "image", + "image_path": "907551be19592bf81aabe16194e29e0d4cdc033081ce81d93199b375ccf0373b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 108, + 352, + 220, + 446 + ], + "blocks": [ + { + "bbox": [ + 108, + 352, + 220, + 446 + ], + "lines": [ + { + "bbox": [ + 108, + 352, + 220, + 446 + ], + "spans": [ + { + "bbox": [ + 108, + 352, + 220, + 446 + ], + "type": "image", + "image_path": "be3d98060c8717ce5516bf203bb186714dcc5f8736a714c988303e5042231639.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 456, + 555, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 555, + 500 + ], + "type": "text", + "content": "Figure 4. The qualitative comparison between ray regression w/ or w/o global perception in the scene1 of Indoor-6. These results contain predicted and ground truth rays from the same cameras, along with their camera centers. The samples indicate that ray predictions w/o global perception (without the CLS and XY tokens, bottom) lack geometric consistency, although some individual rays are accurate. In contrast, incorporating global perception facilitates learning the geometric constraint of rays, resulting in precise ray regression (top)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 224, + 355, + 314, + 446 + ], + "blocks": [ + { + "bbox": [ + 224, + 355, + 314, + 446 + ], + "lines": [ + { + "bbox": [ + 224, + 355, + 314, + 446 + ], + "spans": [ + { + "bbox": [ + 224, + 355, + 314, + 446 + ], + "type": "image", + "image_path": "ed54ac662cbd8a7ef5e89ae8140dbacde14ec2b767e69b37dc4ab22c136d669a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 355, + 407, + 446 + ], + "blocks": [ + { + "bbox": [ + 318, + 355, + 407, + 446 + ], + "lines": [ + { + "bbox": [ + 318, + 355, + 407, + 446 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 407, + 446 + ], + "type": "image", + "image_path": "5d93c7a76b63a7177a208b780b57549b9e4158c7eae2a556eca16418ec7cf8b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 410, + 355, + 500, + 446 + ], + "blocks": [ + { + "bbox": [ + 410, + 355, + 500, + 446 + ], + "lines": [ + { + "bbox": [ + 410, + 355, + 500, + 446 + ], + "spans": [ + { + "bbox": [ + 410, + 355, + 500, + 446 + ], + "type": "image", + "image_path": "cf7af604d1a6e14d37c2fa0a6969baed6b39d7a78621881208a1c4dba886d8be.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "content": "(cf. Fig. 3). The hidden dimension " + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "inline_equation", + "content": "D_{h}" + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "content": " is set as 256. More ablation experiments can be found in Sec. 4.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Training and Hardware Details. Firstly, according to related work [13], we fine-tune the last two blocks of DINO in terms of CRR task for better performance. Particularly, we follow the encoder training protocol of [4] and [13], using the first 100 training scenes of ScanNet [7] to update the weights of unfrozen DINO blocks. The training is performed on 4 NVIDIA A800 GPUs using gradient accumulation [4] in parallel. Next, the mapper head is trained on 1 NVIDIA A800 GPU for each scene, with a training buffer of 26M encoder features. The large buffer size is important for CRR training, as most of patches have different rays. We observed a buffer size scaling law in experiments (cf. Suppl. B). We do 20 passes over the training buffer, utilizing a batch size of 5120. The optimization uses AdamW [21] with a learning rate between " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "5e^{-5}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " and a 1 cycle schedule [37]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 505, + 450, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 450, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 450, + 517 + ], + "type": "text", + "content": "4.2.Quantitative Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 528, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 660 + ], + "type": "text", + "content": "Datasets. We evaluate VL performance of our method on both indoor and outdoor datasets. For indoor scenes, we adopt the Indoor-6 dataset [9], which is a relatively large indoor dataset [9]. It comprises calibrated images from 6 multi-rooms indoor scenes captured over several days, leading to challenging VL. For outdoor scenes, the Map-Free [1] dataset is employed. We use its first 10 scenes, " + }, + { + "bbox": [ + 313, + 528, + 555, + 660 + ], + "type": "inline_equation", + "content": "s00000 \\sim s00009" + }, + { + "bbox": [ + 313, + 528, + 555, + 660 + ], + "type": "text", + "content": ", following [4]. Each scene contains images from two scans of an outdoor location. We split them into mapping and query sets. The ground truth camera poses are computed via SfM method [34]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Indoor Visual Localization. Firstly, the indoor VL experiments are conducted using the Indoor-6 dataset. In the experiments, we compare our DIMM method and its variation enhanced by RANSAC (DIMM-R) with both APR [15]" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25644" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 70, + 492, + 230 + ], + "blocks": [ + { + "bbox": [ + 118, + 70, + 492, + 230 + ], + "lines": [ + { + "bbox": [ + 118, + 70, + 492, + 230 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 492, + 230 + ], + "type": "table", + "html": "
Scene (°/cm/%)SCRAPRCRR
DSAC*[3]ACE[4]marepo[6]marepos[6]DIMMDIMM-R
Throughput (fps)17.917.955.655.657.720.1
s000001.2/5.1/55.80.7/4.2/62.60.6/3.6/69.20.6/3.2/70.70.7/2.0/93.10.6/2.0/94.0
s000010.8/2.4/94.50.4/1.5/1000.9/1.8/96.00.8/1.6/99.21.1/1.6/96.81.1/1.6/96.8
s000020.8/2.4/69.50.5/2.1/77.70.6/2.3/77.10.5/2.1/78.20.9/3.7/71.20.9/3.0/77.1
s000030.6/2.1/75.80.4/2.4/73.60.9/2.5/80.30.8/2.1/84.50.6/2.4/92.80.6/2.2/93.3
s000041.1/6.3/40.40.7/5.8/45.30.8/5.3/45.90.7/5.1/47.60.7/2.4/82.60.7/2.3/96.2
s000050.7/3.6/64.40.5/2.4/73.30.7/2.7/72.20.5/2.6/77.30.8/2.5/86.20.8/2.1/87.9
s000061.0/4.4/53.40.8/4.5/54.80.8/4.2/58.50.6/3.6/63.80.8/5.6/44.30.8/3.9/64.8
s000070.7/7.8/16.90.6/7.6/19.01.0/6.3/25.10.8/6.1/27.50.6/3.6/69.50.5/3.5/71.1
s000081.2/3.5/68.90.7/3.1/71.60.7/2.3/77.40.7/1.9/82.80.7/2.7/85.10.7/2.4/87.9
s000090.8/3.1/81.20.4/1.2/99.70.8/1.8/94.90.9/1.6/95.10.7/1.3/97.70.7/1.2/97.4
Average0.8/4.1/62.10.6/3.5/67.70.8/3.3/69.70.7/3.0/72.70.8/2.8/81.90.7/2.4/86.6
", + "image_path": "43f4b3597abcf4d8f36ee1f6abd42b8c1385f4cdff5d963f122cda8cb870fa72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 267, + 294, + 315 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "type": "text", + "content": "Table 3. Pose accuracy comparison on the outdoor MapFree [1] dataset. We report the median errors in degree (°) for the orientation, cm for the position, and recall at " + }, + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm} / 5^{\\circ}" + }, + { + "bbox": [ + 55, + 239, + 555, + 262 + ], + "type": "text", + "content": ". The best results, the second best and our results are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 267, + 294, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 267, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 294, + 315 + ], + "type": "table", + "html": "
Recall-5°/5cm (%)s00000s00001s00002s00003s00004s00005Average
ACE [4]62.610077.773.645.373.372.1
DIMM-CRR (Ours)93.196.871.292.882.686.287.1
DIMM-SCR64.898.674.577.348.272.172.6
DIMM-APR3.212.45.73.60.32.64.6
", + "image_path": "c88a6839f0c5812c4256b76c42d57858b378a3ec3f9f759b6020405a2ef9e0f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 363, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 296, + 614 + ], + "type": "text", + "content": "and SCR [3-5, 9, 20] approaches. The results are reported in Tab. 2. Additionally, we analyzed the corresponding mapping time and memory costs of various VL methods in Tab. 1, for a comprehensive comparison. DIMM demonstrates better or comparable performance than the previous state-of-the-art, e.g., achieving a recall@5°/5cm of 78.3 in scene6. Moreover, our method achieves generally better results than EGFS-q with \"hard\" multi-mapper combination, proving the efficacy of our \"soft\" mapper ensemble utilizing semantic guidance. The effectiveness of RC-RANSAC is also highlighted by a substantial precision improvement in scene2a from 66.7 to 77.4. Overall, the proposed DIMM-R achieves the best average recall in this dataset. Furthermore, our method enables mapping with reduced computational costs compared to D2S and is trained without 3D supervision, as illustrated in Tab. 1. These findings underscore the potential of our CRR-based method as a promising VL implementation. Some visualization results of DIMM can be found in Fig. 4 and Fig. 7 in the Suppl., where the 3D scene structure is preserved, demonstrating the privacy protection offered by our method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "Outdoor Visual Localization. Our outdoor VL experiments are conducted on the Map-Free Dataset. In addition to the SCR-based ACE and DSAC*, we also compared the recent APR-based method, marepo [6], and its fine-tuned variation, marepos. The results, including median errors for rotation " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "(cm)" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", as well as localization recalls at " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm} / 5^{\\circ}" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": ", are presented in Tab. 3. These results also come from methods with mapping time in Tab. 1. From the table," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 267, + 531, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "type": "text", + "content": "Table 4. Ablation Results about Camera Ray Regression (CRR). We evaluate three variations of our model to showcase the effectiveness of CRR, by modifying output to different VL formulations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 267, + 531, + 327 + ], + "lines": [ + { + "bbox": [ + 340, + 267, + 531, + 327 + ], + "spans": [ + { + "bbox": [ + 340, + 267, + 531, + 327 + ], + "type": "table", + "html": "
Feature EncoderMed. R Err. (°)Med. t Err. (cm)Recall-5°/5cm
T_i ∩ T[CLS] ∩ Txy_i0.85.644.3
T_i * T*[CLS] ∩ Txy_i0.86.039.7
T_i ∩ TG[41] ∩ Txy_i1.18.319.7
T_i ∩ T[CLS] ∩ XY_i0.97.627.2
T_i ∩ T[CLS]0.97.426.8
T_i1.111.517.2
", + "image_path": "36c54f60241000a11f97dd9378bb29fa63b7c54bcb286f65fa1fcf3a1fe47ba1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 357 + ], + "type": "text", + "content": "Table 5. Ablation Study about our Feature Encoder. The experiments are performed on the MapFree s00006 dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 360, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 576 + ], + "type": "text", + "content": "it is evident that our methods achieve the best overall results, outperforming both SCR and APR methods. The rotation accuracy of ACE, yet, surpasses our methods, indicating that local ambiguity poses a significant challenge for CRR methods and obtaining precise ray directions from image patches is challenging. However, it is observed that CRR methods facilitate better camera center estimation compared to APR and SCR methods. This improvement can be attributed to the fact that the camera center location integrates contributions from all camera rays, and the bounded ray error [27] enhances the robustness of the center estimation against noise. The efficacy of RC-RANSAC algorithm can also be verified in the table, as it leads to substantial improvements of VL performance, e.g., " + }, + { + "bbox": [ + 313, + 360, + 556, + 576 + ], + "type": "inline_equation", + "content": "44.3 \\rightarrow 64.8" + }, + { + "bbox": [ + 313, + 360, + 556, + 576 + ], + "type": "text", + "content": " of recall in s00006. Furthermore, we investigate the computational efficiency of our methods. The throughputs (fps) of all methods are reported in the table, revealing that the lightweight network and linear pose solver of DIMM achieve the best inference efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 440, + 598 + ], + "type": "text", + "content": "4.3. Ablation Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Camera Ray Regression. One of our main contributions in this work is introducing Camera Ray Regression (CRR) into the VL task. Thus, we provide ablation study about this formulation in Tab. 4. In particular, we modify the last layer of DIMM to output scene coordinates (DIMM-SCR) or direct camera pose (DIMM-APR), and train these variations separately. The DIMM-SCR is trained with GDT, same as ACE[4]. The DIMM-APR, on the other hand, is trained in image level, as the direct pose vectors cannot" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25645" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 282, + 198 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 198 + ], + "type": "image", + "image_path": "3713c77838c9d31b05a8d92c2320b1795b2ee1aa602394120bf515105e0f23f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 296, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 283 + ], + "type": "text", + "content": "Figure 5. Visualization results of sub-mapper attention score. In the experiments, we observed that specific semantic entities elicit responses from different mappers. For instance, the fireplace at the top of the figure frequently results in high attention scores from mapper #3, while mapper #2 exhibits a low attention score. Conversely, mapper #3 seems do not \"care\" the sofa at the bottom, which leads to high scores of mapper #2 instead." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 445 + ], + "type": "text", + "content": "be assigned to image patches. The experiments are performed on the MapFree dataset, with the localization recall reported. In Tab. 4, we can see that our CRR formulation significantly outperforms the SCR and APR under the same network architecture, indicating the contribution of our ray-parameterization. DIMM-SCR overcomes ACE slightly in average, but our model is tailored for CRR. DIMM-APR utilizes per-scene pose regression and achieves poor results, but this is consistent with the results in marepo [6]. This suggests the APR model may require much larger datasets to learn pose patterns. Overall, the results in Tab. 4 evaluate our main contribution: CRR formulation in VL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": "Feature Encoder. Based on our analysis of CRR, we require patch features to exhibit strong local description while also incorporating global perception and image position encoding to achieve disambiguation. To validate this, we conduct experiments on the MapFree s00006 dataset (cf. Tab. 5). Specifically, we modify the content of the patch feature, including features w/o global perception and image position encoding " + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}" + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i)" + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": ", to compare with the proposed version " + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathcal{T}_{\\mathrm{xy}_i})" + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": ". Also, we investigate the efficacy of DINO fine-tuning, and the performance of original DINO " + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i^* \\otimes \\mathcal{T}_{[\\mathrm{CLS}]}^* \\otimes \\mathcal{T}_{\\mathrm{xy}_i})" + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": " is also reported. It is evident that global perception with image position encoding are crucial for CRR, leading to a significant accuracy improvement. The fine-tuning of DINO also improves the performance as well (39.7 vs. 44.3). Visualization comparison about the global perception is presented in Fig. 4. We can see the proposed global perception enables the learning of ray geometry, causing the rays to converge toward the camera center, thereby enhancing precision. We also conduct experiments for a variation without Fourier Encoding " + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i \\otimes \\mathcal{T}_{[\\mathrm{CLS}]} \\otimes \\mathsf{XY}_i)" + }, + { + "bbox": [ + 55, + 450, + 298, + 715 + ], + "type": "text", + "content": ". The results in the first and third rows of Tab. 5 show that Fourier Encoding effectively prevents the submergence of positional" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 329, + 71, + 542, + 123 + ], + "blocks": [ + { + "bbox": [ + 329, + 71, + 542, + 123 + ], + "lines": [ + { + "bbox": [ + 329, + 71, + 542, + 123 + ], + "spans": [ + { + "bbox": [ + 329, + 71, + 542, + 123 + ], + "type": "table", + "html": "
Atten. Num.Med. R Err. (°)Med. t Err. (cm)Recall@5°/Scmsize (MB)
011.07.231.012.1
20.95.842.614.2
40.85.644.316.2
60.86.337.718.2
81.17.730.420.2
", + "image_path": "fa4a5dae98b98a091af77d93805cb15475bd18c246465da52b0e68182e17790a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 333, + 123, + 485, + 130 + ], + "lines": [ + { + "bbox": [ + 333, + 123, + 485, + 130 + ], + "spans": [ + { + "bbox": [ + 333, + 123, + 485, + 130 + ], + "type": "text", + "content": "1 Take average of the sub-mapper embeddings, eliminating attention layers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 138, + 555, + 172 + ], + "lines": [ + { + "bbox": [ + 313, + 138, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 555, + 172 + ], + "type": "text", + "content": "Table 6. Ablation Study about the Number of Semantic Attention Layers. The experiments are performed on the MapFree s00006 dataset. The efficacy of our semantic attention is evaluated." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "text", + "content": "features. Finally, we explore another possible global perception way provided by [41] " + }, + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "inline_equation", + "content": "(\\mathcal{T}_i \\otimes \\mathcal{T}_G \\otimes \\mathcal{T}_{xy_i})" + }, + { + "bbox": [ + 313, + 181, + 555, + 253 + ], + "type": "text", + "content": ", the feature for image retrieval. The results show that the replaced global feature fails to match the original efficacy, possibly because DINO's global feature is more compatible with its own local patch features, facilitating subsequent mapping learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 258, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 475 + ], + "type": "text", + "content": "Mapper Head. In the same scene, we also conduct ablation experiments on the network structure of the Mapper Head. We first validate the efficacy of the semantic attention module and experiment with its number of layers. When the layer number is set to 0, we directly take the average of output from multiple sub-mappers. The results are presented in Tab. 6, containing median rotation/translation error, localization recall and model size. It is evident that utilizing semantic attention is more effective compared to simple average fusion, and the increase in network size is deemed acceptable. Additional ablation study about the MLP mappers can be found in Suppl. B. To showcase the effectiveness of the soft mapper assemble, we visualize some query images combined with their sub-mapper attention scores in Fig. 5. It can be observed that different sub-mappers get high attention scores to different objects in images. This indicates that our semantic attention mechanism allows the sub-mappers to \"remember\" specific semantic objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 496, + 388, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 496, + 388, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 496, + 388, + 509 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 713 + ], + "type": "text", + "content": "In this work, we propose regressing camera rays to perform visual localization. This overparameterized representation of camera model leads to high precision, and, more importantly, enhances privacy protection. In particular, we introduce DIMM as a learning model to regress camera rays from images. It utilizes DINO as a scene-agnostic encoder to output features that incorporate both local and global perception. A scene-specific mapper then regresses ray parameters from the features, involving a semantic attention module to merge results from multiple mappers. Finally, a ray-level RANSAC algorithm is proposed to improve the ray-to-pose accuracy. In experiments on both indoor and outdoor datasets, our methods demonstrate comparable or superior performance to current approaches, while ensuring privacy preservation. Acknowledgement. This work has been funded in part by the NSFC grants 62176156." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25646" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Eduardo Arnold, Jamie Wynn, Sara Vicente, Guillermo Garcia-Hernando, Áron Monszpart, Victor Adrian Prisacariu, Daniyar Turmukhambetov, and Eric Brachmann. Map-free visual relocalization: Metric pose relative to a single image. In ECCV, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "text", + "content": "[2] Eric Brachmann and Carsten Rother. Expert sample consensus applied to camera re-localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7525-7534, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "type": "text", + "content": "[3] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE transactions on pattern analysis and machine intelligence, 44 (9):5847-5865, 2021. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "text", + "content": "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "text", + "content": "[5] Bach-Thuan Bui, Huy-Hoang Bui, Dinh-Tuan Tran, and JooHo Lee. D2s: Representing sparse descriptors and 3d coordinates for camera relocalization. IEEE Robotics and Automation Letters, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "type": "text", + "content": "[6] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 1, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 475 + ], + "type": "text", + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 531 + ], + "type": "text", + "content": "[9] Tien Do, Ondrej Miksik, Joseph DeGol, Hyun Soo Park, and Sudipta N Sinha. Learning to detect scene landmarks for camera localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11132-11142, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 577 + ], + "type": "text", + "content": "[10] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 296, + 632 + ], + "type": "text", + "content": "[11] Kartik Garg, Sai Shubodh Puligilla, Shishir Kolathaya, Madhava Krishna, and Sourav Garg. Revisit anything: Visual place recognition via image segment retrieval. In European Conference on Computer Vision, pages 326-343. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[12] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 3, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[13] Sergio Izquierdo and Javier Civera. Optimal transport aggregation for visual place recognition. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17658-17668, 2023. 4, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[14] Xudong Jiang, Fangjinhua Wang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. R-score: Revisiting scene coordinate regression for robust large-scale visual localization. arXiv preprint arXiv:2501.01421, 2025. 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "type": "text", + "content": "[15] Alex Kendall and Roberto Cipolla. Modelling uncertainty in deep learning for camera relocalization. In 2016 IEEE international conference on Robotics and Automation (ICRA), pages 4762-4769. IEEE, 2016. 1, 3, 6, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "text", + "content": "[16] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5974-5983, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 555, + 252 + ], + "type": "text", + "content": "[17] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Proceedings of the IEEE international conference on computer vision, pages 2938-2946, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 296 + ], + "type": "text", + "content": "[18] Sihang Li, Siqi Tan, Bowen Chang, Jing Zhang, Chen Feng, and Yiming Li. Unleashing the power of data synthesis in visual localization. arXiv preprint arXiv:2412.00138, 2024. 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "type": "text", + "content": "[19] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In ICCV, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 333, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 554, + 388 + ], + "type": "text", + "content": "[20] Ting-Ru Liu, Hsuan-Kung Yang, Jou-Min Liu, Chun-Wei Huang, Tsung-Chih Chiang, Quan Kong, Norimasa Kobori, and Chun-Yi Lee. Reprojection errors as prompts for efficient scene coordinate regression. In European Conference on Computer Vision, pages 286-302. Springer, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 389, + 554, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 554, + 411 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. *ICLR*, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 412, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 555, + 477 + ], + "type": "text", + "content": "[22] Deben Lu, Wendong Xiao, Teng Ran, Liang Yuan, Kai Lv, and Jianbo Zhang. Attention-based accelerated coordinate encoding network for visual relocalization. In 2024 IEEE 7th Information Technology, Networking, Electronic and Automation Control Conference (ITNEC), pages 1675-1680, 2024. 1, 3, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 479, + 555, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 555, + 523 + ], + "type": "text", + "content": "[23] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 525, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 555, + 579 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "type": "text", + "content": "[25] Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. arXiv preprint arXiv:2412.12392, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 615, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 712 + ], + "type": "text", + "content": "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DiNov2: Learning robust visual features without supervision, 2023. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25647" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "text", + "content": "[27] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes Lutz Schonberger. Global Structure-from-Motion Revisited. In European Conference on Computer Vision (ECCV), 2024. 2, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 182 + ], + "type": "text", + "content": "[28] Maxime Pietrantoni, Martin Humenberger, Torsten Sattler, and Gabriela Csurka. Segloc: Learning segmentation-based representations for privacy-preserving visual localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15380-15391, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 296, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 238 + ], + "type": "text", + "content": "[29] Francesco Pittaluga, Sanjeev J. Koppal, Sing Bing Kang, and Sudipta N. Sinha. Revealing scenes by inverting structure from motion reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 293 + ], + "type": "text", + "content": "[30] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 369 + ], + "type": "text", + "content": "[31] Paul-Edouard Sarlin, Ajaykumar Unagar, Mans Larsson, Hugo Germain, Carl Toft, Viktor Larsson, Marc Pollefeys, Vincent Lepetit, Lars Hammarstrand, Fredrik Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3247–3257, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "text", + "content": "[32] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence search. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I 12, pages 752-765. Springer, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[33] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE transactions on pattern analysis and machine intelligence, 39(9):1744-1756, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[34] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 296, + 568 + ], + "type": "text", + "content": "[35] Thomas Schops, Viktor Larsson, Marc Pollefeys, and Torsten Sattler. Why having 10,000 parameters in your camera model is better than twelve. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2535-2544, 2020. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 613 + ], + "type": "text", + "content": "[36] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 296, + 658 + ], + "type": "text", + "content": "[37] Leslie N. Smith and Nicholay Topin. Super-convergence: Very fast training of residual networks using large learning rates. Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[38] Pablo Speciale, Johannes L. Schonberger, Sing Bing Kang, Sudipta N. Sinha, and Marc Pollefeys. Privacy preserving image-based localization. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5488-5498, 2019. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 452 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "text", + "content": "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "text", + "content": "[40] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 250 + ], + "type": "text", + "content": "[41] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 306 + ], + "type": "text", + "content": "[42] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient loftr: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "text", + "content": "[43] Yinshuang Xu, Dian Chen, Katherine Liu, Sergey Zakharov, Rares Andrei Ambrus, Kostas Daniilidis, and Vitor Campagnolo Guizilini. " + }, + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "inline_equation", + "content": "se(3)" + }, + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "text", + "content": " equivariant ray embeddings for implicit multi-view depth estimation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "type": "text", + "content": "[44] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 408, + 555, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 555, + 452 + ], + "type": "text", + "content": "[45] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20217-20226, 2024. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25648" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_content_list.json b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b659df69d01d559e50995a9786a01a0f2b9cb859 --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Semi-ViM: Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning", + "text_level": 1, + "bbox": [ + 104, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyang He* \nUniversity of Warwick", + "bbox": [ + 183, + 203, + 367, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyang Xie \nUniversity of Warwick", + "bbox": [ + 413, + 203, + 596, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haochen You Columbia University", + "bbox": [ + 643, + 203, + 813, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Victor Sanchez \nUniversity of Warwick", + "bbox": [ + 405, + 246, + 588, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 316, + 326, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) is often hindered by learning biases when imbalanced datasets are used for training, which limits its effectiveness in real-world applications. In this paper, we propose Semi-ViM, a novel SSL framework based on Vision Mamba, a bidirectional state space model (SSM) that serves as a superior alternative to Transformer-based architectures for visual representation learning. Semi-ViM effectively deals with imbalanced datasets and improves model stability through two key innovations: LyapEMA, a stability-aware parameter update mechanism inspired by Lyapunov theory, and SSMixup, a novel mixup strategy applied at the hidden state level of bidirectional SSMs. Experimental results on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms state-of-the-art SSL models, achieving $85.40\\%$ accuracy with only $10\\%$ of the labeled data, surpassing Transformer-based methods such as Semi-ViT.", + "bbox": [ + 88, + 348, + 485, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 636, + 220, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) aims to mitigate the scarcity of large-scale labeled training data while reducing annotation costs [16]. In the context of classification, however, a challenge remains: Most state-of-the-art (SOTA) SSL models are trained under the assumption that access to a dataset with a balanced number of labeled samples per class is possible, which is often overly idealized and deviates significantly from real-world conditions [20]. In practice, label imbalance induces learning biases in SSL, ultimately compromising model performance [41].", + "bbox": [ + 89, + 662, + 482, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to its outstanding performance, many studies have integrated the vision Transformer (ViT) into SSL methods [4, 43, 46]; e.g., Semi-ViT, adopts a pre-training and fine-tuning approach to achieve SOTA results [4]. How-", + "bbox": [ + 89, + 814, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31f478e572973ac2bff382619233040efc743ffdbc670b2ce3ef03e085494769.jpg", + "image_caption": [ + "Figure 1. Semi-ViT-Small trained with only $1\\%$ of the labeled data of ImageNet-1K. Left: Training fails when FixMatch is used. The loss decreases but test accuracy remains near zero. Right: Training is unstable and fails when EMA-Teacher is used: The loss and accuracy oscillate heavily." + ], + "image_footnote": [], + "bbox": [ + 517, + 316, + 903, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ever, Semi-ViT has a weaker inductive bias, requiring large-scale data for effective training [39]. Moroever, as shown by our experiments, Semi-ViT does not perform well on long-tailed imbalanced datasets, like ImageNet-LT, confirming its limitations in handling label imbalance. Although the performance of Semi-ViT can be improved by increasing the model size, e.g., Semi-ViT-Huge, this comes at the expense of a larger number of trainable parameters, around 632M [4]. Vision Mamba (ViM), which incorporates a bidirectional state space model (SSM), has modeling capabilities comparable to those of ViT, while achieving subquadratic-time computational and linear memory complexities [47].", + "bbox": [ + 511, + 551, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on the previous discussions, this paper integrates ViM into an SSL framework for classification and proposes a ViM-based SSL model, hereinafter called Semi-ViM. Semi-ViM incorporates two key innovations aimed at improving model stability during training and performance: LyapEMA and SSMixup. LyapEMA is a novel parameter update mechanism that leverages the Lyapunov theory to dynamically adjust parameter updates, ensuring a more stable training process. LyapEMA addresses the training failures encountered when using only a small amount of labeled", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "765", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data to train Semi-ViT-Small, as shown in Fig 1. LyapEMA also helps prevent training collapse, commonly encountered when using the exponential moving average (EMA) and FixMatch parameter update mechanisms [4]. SSMixup, on the other hand, is a novel mixup strategy that operates at the hidden state space of SSMs and goes beyond the traditional input-level mixup strategy. Specifically, SSMixup incorporates mixed hidden states as residual units in the training process, improving the model's capability to generalize under sparse label conditions. Our main contributions can be summarized as follows:", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Semi-ViM, an SSL framework that outperforms other SSL methods based on CNNs and Transformers. The number of parameters of Semi-ViM is only around one-fourth of that of Semi-ViT.", + "- We propose LyapEMA, a novel parameter update mechanism that successfully addresses the training instability issues commonly found in Semi-ViT, thus improving model convergence and generalization capabilities.", + "- We propose SSMixup, a hidden state space-level mixup strategy that effectively enhances the learning capability of bidirectional SSMs.", + "- Based on evaluations on the long-tailed ImageNet-LT dataset, we demonstrate Semi-ViM's strong capabilities in mitigating learning biases caused by label imbalance." + ], + "bbox": [ + 89, + 257, + 482, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 482, + 232, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imbalanced datasets are those in which the different classes are distributed unequally, often resulting in long-tailed distributions [7]. SSL aims to leverage a small amount of labeled data and a large amount of unlabeled data to improve performance, which is useful to deal with imbalanced datasets. SSL has achieved significant advancements in several computer vision tasks [10, 11, 21, 38, 45].", + "bbox": [ + 89, + 507, + 482, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional SSL methods, e.g., entropy minimization and consistency regularization [13, 15], exploit unlabeled data to reduce dependence on manual annotation. For example, consistency regularization leverages unlabeled data under the assumption that a model's predictions remain consistent under different input perturbations. Mainstream methods include II-Model [22], Mean Teacher [40], Virtual Adversarial Training (VAT) [30], and FixMatch [38], which is an important benchmark in SSL [38]. Pseudo-Labeling [23] is a common strategy used in SSL. It consists of using the model's predictions on unlabeled data as pseudo-labels for training. Noisy Student, FixMatch [38], and FlexMatch [45] adopt this strategy and use confidence thresholds to filter out pseudo-labels.", + "bbox": [ + 89, + 614, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several SOTA SSL methods are based on FixMatch or any of its variants, e.g., FlexMatch and CoMatch [24], which integrate contrastive learning or self-supervised methods to enhance performance. However, these methods typically assume that the class distribution is balanced,", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which is rarely the case in real-world scenarios [12]. When dealing with imbalanced datasets [8], these SOTA SSL methods face two primary challenges. First, pseudo-label biases, i.e., models tend to predict more frequently the majority classes when generating pseudo-labels, exacerbating the imbalance issue and limiting learning from the minority classes [19]. Second, low-quality pseudo-labels, i.e., models tend to assign low confidence scores to those pseudolabels generated for the minority class samples, thus reducing their impact on training and potentially leading to training collapse [4]. For example, FixMatch relies on a fixed confidence threshold for selecting pseudo-labels, but in an imbalanced dataset setting, pseudo-labels of the minority classes often fail to reach this threshold, leading to insufficient learning from such classes [4, 43].", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 330, + 651, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models and ViM: SSMs have a long-standing history in control systems and signal processing, offering a structured approach to sequential data modeling [47]. Deep learning adopts SSMs to more effectively handle long-range dependencies compared to traditional recurrence-based models. Unlike Transformers, which rely on global self-attention, SSMs leverage structured state transitions for sequential data propagation, enabling subquadratic computational complexity and linear memory scaling [47]. In computer vision, SSMs have been integrated into deep architectures to enhance representation learning by modeling both spatial and temporal dependencies [27, 47]. The SSM is formulated as a discrete-time state-space system [47]:", + "bbox": [ + 511, + 354, + 906, + 551 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = A h _ {t - 1} + B x _ {t}, \\quad y _ {t} = C h _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 561, + 903, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $A \\in \\mathbb{R}^{N \\times N}$ is the state transition matrix; $B \\in \\mathbb{R}^{N \\times 1}$ and $C \\in \\mathbb{R}^{1 \\times N}$ are input and output projection matrices, respectively; $x_{t}$ represents the input sequence at time step $t$ ; and $y_{t}$ is the corresponding output representation. The parameters of an SSM are usually optimized in a data-dependent manner to ensure computational efficiency and parallelization.", + "bbox": [ + 511, + 587, + 903, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The bidirectional extension of SSMs further improves feature learning by capturing both forward and backward information, making it a competitive alternative to self-attention mechanisms used by Transformers. [47]. The bidirectional SSM operates in the forward $(f)$ and backward $(b)$ direction:", + "bbox": [ + 511, + 694, + 903, + 782 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} ^ {f} = A h _ {t - 1} ^ {f} + B x _ {t}, \\quad y _ {t} ^ {f} = C h _ {t} ^ {f}; \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 792, + 903, + 811 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} ^ {b} = A h _ {t + 1} ^ {b} + B x _ {t}, \\quad y _ {t} ^ {b} = C h _ {t} ^ {b}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 820, + 903, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The final output at time step $t$ is obtained by merging the forward and backward outputs through a gating mechanism:", + "bbox": [ + 511, + 844, + 903, + 875 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "766", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3653d4902f23b10c9df92783cd4756241dcf665036e59c57e7732dff742acb1e.jpg", + "image_caption": [ + "(a) FixMatch" + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 328, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f810dfa34687df9fa1847eeb2d61ff8e17d4b2205639902f1751d62b55cef712.jpg", + "image_caption": [ + "(b) EMA-Teacher", + "Figure 2. Differences between (a) FixMatch, (b) EMA-Teacher and (c) Semi-ViM (ours). In the figure, $X_{s}$ and $X_{w}$ are, respectively, strongly and weakly augmented inputs; $p^{s}$ and $p_{student}^{s}$ are the predictions of the student model on $X_{s}$ ; $p^{w}$ and $p_{teacher}^{w}$ are the predictions of the teacher model on $X_{w}$ ; $\\tau$ is the threshold used to select pseudo-labels; and $\\theta, \\theta^{student}, \\theta^{teacher}$ are, respectively, the learnable parameters for the entire model, the student model, and the teacher model." + ], + "image_footnote": [], + "bbox": [ + 370, + 90, + 619, + 271 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3714e3218e742eb564341217c5bc94abe38467f6b3ea0b918b6481af82df601.jpg", + "image_caption": [ + "(c) Semi-ViM(Ours)" + ], + "image_footnote": [], + "bbox": [ + 666, + 90, + 908, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{SiLU}(\\cdot)$ is the Sigmoid-weighted Linear Unit activation function, and $z$ is a gating vector that adaptively balances bidirectional information.", + "bbox": [ + 89, + 385, + 482, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViM adopts a bidirectional SSM for visual representation learning and integrates positional embeddings for spatial awareness. In ViM, the input image is first partitioned into patches and projected onto a high-dimensional feature space before being processed. With a linear computational complexity, ViM provides data-dependent global visual modeling, significantly improving efficiency in high-resolution tasks compared to Transformer-based architectures [26, 47].", + "bbox": [ + 88, + 440, + 482, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixup strategies: These strategies interpolate input samples and labels, and hence can be used as data augmentation techniques to improve generalization capabilities [2]. In SSL, they can enhance pseudo-labeling and provide consistency regularization [34]. Recent variants extend mixup strategies to the feature space and pseudo-label frameworks, e.g., the probabilistic pseudo mixup strategy refines pseudolabel quality by weighting interpolated samples based on their confidence scores, which are derived from the model's softmax outputs or uncertainty estimates. These variants can improve SSL robustness, particularly for ViTs, which benefit from stronger regularization [4].", + "bbox": [ + 88, + 571, + 482, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weak and strong data augmentation: Weak data augmentation involves simple transformation, e.g., random cropping, flipping, or slight color jitter, which introduce small variations to the original data to help a model learn from basic invariances. Strong data augmentation, on the other hand, applies more aggressive transformations, e.g., large crops, strong color distortions, or more significant perturbations, which force a model to learn more robust representations by simulating more challenging scenarios [31].", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Semi-ViM framework", + "text_level": 1, + "bbox": [ + 511, + 383, + 805, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-ViM adopts a teacher-student framework and uses a labeled dataset, $\\mathcal{X}_l = \\{(x_i^l,y_i^l)\\}_{i = 1}^{N_l}$ , and an unlabeled dataset $\\mathcal{X}_u = \\{x_i^u\\}_{i = 1}^{N_u}$ for training. Specifically, the teacher model, $f_{\\mathrm{teacher}}(\\theta^{\\mathrm{teacher}})$ , generates pseudo-labels, $\\{\\hat{y}_i\\}$ , for unlabeled samples, $\\{x_i^u\\}$ , and outputs confidence scores, $\\{o_i\\}$ , for the pseudo-labels. The student model, $f_{\\mathrm{student}}(\\theta^{\\mathrm{student}})$ , is then trained on samples mixed according to their confidence scores. The difference between the training paradigm of Semi-ViM and those of FixMatch and EMA-Teacher is shown in Fig 2.", + "bbox": [ + 511, + 410, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce LyapEMA, which is inspired by Lyapunov stability theory [36] and stabilizes parameter updates, thus preventing model collapse during the training of Semi-ViM. Next, we introduce SSMixup, a new mixup strategy designed for SSM-based models. Unlike traditional mixup strategies that operate solely on the input level, SSMixup mixes the hidden states of bidirectional SSMs. Finally, we detail the training and loss function of Semi-ViM.", + "bbox": [ + 511, + 563, + 908, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. LyapEMA", + "text_level": 1, + "bbox": [ + 511, + 710, + 632, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LyapEMA is a novel optimization framework that integrates Lyapunov stability theory with EMA. LyapEMA is designed for teacher-student frameworks under an SSL setting. It has two main objectives. The first is to ensure stable parameter updates, because training SSL frameworks, like Semi-ViT, with a limited amount of labeled data can lead to training collapse. The second objective is reducing the learning bias associated with pseudo-labels while preserving the advantages of EMA in refining the teacher model.", + "bbox": [ + 511, + 733, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\theta^{\\mathrm{student}}$ and $\\theta^{\\mathrm{teacher}}$ denote the student and teacher model parameters, respectively. EMA updates the teacher", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "767", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model using gradient descent as follows:", + "bbox": [ + 89, + 90, + 361, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {\\text {t e a c h e r}} = \\alpha_ {t} \\theta_ {t - 1} ^ {\\text {t e a c h e r}} + (1 - \\alpha_ {t}) \\theta_ {t} ^ {\\text {s t u d e n t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 119, + 480, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\in (0,1)$ is the EMA decay factor. However, EMA cannot guarantee stability, potentially leading to divergence due to accumulated bias associated with learning from pseudo-labels. To address this issue, LyapEMA introduces Lyapunov stability constraints to guide the updates. Let us define a Lyapunov candidate function, $V(\\theta^{\\mathrm{student}},\\theta^{\\mathrm{teacher}})$ , to measure the discrepancy between the student and teacher models:", + "bbox": [ + 89, + 142, + 483, + 261 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV \\left(\\theta^ {\\text {s t u d e n t}}, \\theta^ {\\text {t e a c h e r}}\\right) = \\frac {1}{2} \\| \\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 282, + 482, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the function in Eq. 6 acts as a stability indicator, ensuring that the student model does not drift too far from the teacher model. The Lyapunov stability criterion requires that the derivative of $V(\\cdot, \\cdot)$ w.r.t. time, i.e., $\\dot{V}$ , satisfies:", + "bbox": [ + 89, + 316, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {V} = \\frac {d V}{d t} = \\left(\\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}}\\right) ^ {\\mathrm {T}} \\left(\\dot {\\theta} ^ {\\text {s t u d e n t}} - \\dot {\\theta} ^ {\\text {t e a c h e r}}\\right) < 0, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 412, + 482, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\dot{\\theta}^{\\mathrm{student}}$ and $\\dot{\\theta}^{\\mathrm{teacher}}$ also derivatives w.r.t. time. The criterion in Eq. 7 guarantees that the discrepancy between the student and teacher models decreases over time, ensuring convergence.", + "bbox": [ + 89, + 450, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enforce stability while leveraging EMA for smooth teacher model updates, LyapEMA updates the student model parameters by using aLyapunov regularization term:", + "bbox": [ + 89, + 513, + 483, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t} ^ {\\text {s t u d e n t}} = \\theta_ {t - 1} ^ {\\text {s t u d e n t}} - \\eta \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) + \\lambda \\left(\\theta_ {t} ^ {\\text {t e a c h e r}} - \\theta_ {t} ^ {\\text {s t u d e n t}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 580, + 482, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\theta^{\\mathrm{student}})$ is a semi-supervised loss function, $\\eta$ is the learning rate, and $\\lambda$ is a stability coefficient that ensures the student model remains close to the teacher model.", + "bbox": [ + 89, + 607, + 482, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LyapEMA updates the teacher model parameters by using EMA as denoted by Eq. 5. However, the EMA decay factor $\\alpha_{t}$ is dynamically adjusted based on the Lyapunov function:", + "bbox": [ + 89, + 652, + 483, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t} = \\operatorname {s i g m o i d} \\left(\\gamma \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t - 1} ^ {\\text {t e a c h e r}} \\| _ {2}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 726, + 480, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ is a sensitivity hyperparameter controlling how $\\alpha_{t}$ adapts based on the student-teacher discrepancy, and the sigmoid function ensures $\\alpha_{t}$ remains within (0, 1), preventing instability in the updates.", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LyapEMA ensures that updates lead to stable training by using the Lyapunov decrease condition:", + "bbox": [ + 89, + 810, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t + 1} - V _ {t} = \\frac {1}{2} \\| \\theta_ {t + 1} ^ {\\text {s t u d e n t}} - \\theta_ {t + 1} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2} - \\frac {1}{2} \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 859, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 LyapEMA", + "text_level": 1, + "bbox": [ + 514, + 90, + 676, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Initial parameters $\\theta^{\\mathrm{student}}$ , $\\theta^{\\mathrm{teacher}}$ , learning rate $\\eta$ , stability coefficient $\\lambda$ , sensitivity $\\gamma$ \nOutput: Stabilized parameters $\\theta^{\\mathrm{student}}$ , $\\theta^{\\mathrm{teacher}}$", + "bbox": [ + 511, + 108, + 903, + 155 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: while training not converged do", + "2: Compute loss $\\mathcal{L}(\\theta^{\\mathrm{student}})$", + "3: Update student model: $\\theta_{t}^{\\mathrm{student}} = \\theta_{t - 1}^{\\mathrm{student}} - \\eta \\nabla_{\\theta}\\mathcal{L}(\\theta^{\\mathrm{student}}) + \\lambda (\\theta_{t}^{\\mathrm{teacher}} - \\theta_{t}^{\\mathrm{student}})$", + "4: Compute adaptive EMA decay in LyapEMA: $\\alpha_{t} = \\mathrm{sigmoid}(\\gamma \\| \\theta_{t}^{\\mathrm{student}} - \\theta_{t - 1}^{\\mathrm{teacher}}\\|_{2})$", + "5: Update teacher model: $\\theta_{t}^{\\mathrm{teacher}} = \\alpha_{t}\\theta_{t - 1}^{\\mathrm{teacher}} + (1 - \\alpha_{t})\\theta_{t}^{\\mathrm{student}}$", + "6: Ensure stability: Verify $V(\\theta)$ decreases", + "7: if $V_{t} \\geq V_{t-1}$ then", + "8: Reduce learning rate: $\\eta \\gets \\eta /2$", + "9: Increase stability coefficient: $\\lambda \\gets \\lambda +\\Delta \\lambda$", + "10: Recompute student and teacher updates", + "11: end if", + "12: end while", + "13: return $\\theta^{\\mathrm{teacher}}$ , $\\theta^{\\mathrm{student}}$" + ], + "bbox": [ + 519, + 157, + 903, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By using the LyapEMA update rules in Eq. 10 and expanding, we obtain:", + "bbox": [ + 511, + 446, + 903, + 477 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {t + 1} - V _ {t} \\approx - \\eta \\| \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) \\| _ {2} ^ {2} - \\lambda \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 505, + 903, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since both terms in Eq. 11 are negative, $V$ decreases monotonically, ensuring stability. The complete LyapEMA framework is summarized in Algorithm 1. In Section 5.1, we conduct experiments not only on Semi-ViM but also on other SSL methods to validate LyapEMA's effectiveness.", + "bbox": [ + 511, + 537, + 906, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. SSMixup", + "text_level": 1, + "bbox": [ + 511, + 626, + 622, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SSMixup introduces mixup strategies in the input and the hidden state space of a bidirectional SSM to effectively leverage unlabeled data and enhance generalization in SSL. As mentioned before, Semi-ViM adopts a teacher-student framework where the student model, $f_{\\mathrm{student}}(\\theta^{\\mathrm{student}})$ , is trained on the mixed samples.", + "bbox": [ + 511, + 648, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In SSMixup, the pseudo-label confidence scores, $o_i$ and $o_j$ , of two unlabeled samples, $x_i^u$ and $x_j^u$ , respectively, are used first to compute a mixup coefficient:", + "bbox": [ + 511, + 739, + 905, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi = \\frac {o _ {i}}{o _ {i} + o _ {j}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 797, + 903, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The mixup coefficient $\\psi$ is used to form the mixed input and labels:", + "bbox": [ + 511, + 840, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {\\text {m i x}} = \\psi x _ {i} ^ {u} + (1 - \\psi) x _ {j} ^ {u}, \\quad y _ {\\text {m i x}} = \\psi \\hat {y} _ {i} + (1 - \\psi) \\hat {y} _ {j}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 885, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "768", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03c75fa5720095ae52eb1d7609617dd683315634110e7abbd6ace3ba5eebf439.jpg", + "image_caption": [ + "Figure 3. The SSMixup strategy. The red part applies mixup at the hidden state level. $L$ denotes to the number of SSMixup encoder layers in Semi-ViM. It is a hyperparameter typically set to 24." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 480, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The mixed input $x_{\\mathrm{mix}}$ is then fed into the forward and backward components of the bidirectional SSM, yielding hidden states:", + "bbox": [ + 89, + 308, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\mathrm {f w d}} = \\operatorname {S S M} _ {\\mathrm {f w d}} \\left(x _ {\\text {m i x}}\\right), \\quad h ^ {\\mathrm {b w d}} = \\operatorname {S S M} _ {\\mathrm {b w d}} \\left(x _ {\\text {m i x}}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 364, + 482, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSMixup mixes the hidden states by using the mixup coefficient, $\\psi$ :", + "bbox": [ + 89, + 393, + 483, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\text {m i x}} = \\psi h ^ {\\mathrm {f w d}} + (1 - \\psi) h ^ {\\mathrm {b w d}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 435, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_{\\mathrm{mix}}$ is computed as a residual unit and $C$ is defined in Eq. 1. The forward and backward outputs of the bidirectional SSM are merged to compute the final output (see Eq. 4):", + "bbox": [ + 89, + 465, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b} + \\nu \\left(C h _ {\\text {m i x}}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 551, + 482, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nu$ is the control coefficient. The computed $y_{t}$ is used in a cross-entropy loss with $y_{\\mathrm{mix}}$ to update $\\theta^{\\mathrm{student}}$ . Parameters $\\theta^{\\mathrm{teacher}}$ are smoothly updated via LyapEMA, maintaining stable pseudo-label generation.", + "bbox": [ + 89, + 580, + 483, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 illustrates the functionality of SSMixup. By introducing mixup strategies at both the input (see Eq. 13) and hidden state levels (see Eq. 15), combined with confidence-driven weighting, SSMixup effectively exploits unlabeled data and significantly improves generalization capabilities in SSL settings.", + "bbox": [ + 89, + 642, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training Semi-ViM", + "text_level": 1, + "bbox": [ + 89, + 742, + 277, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Semi-ViM involves three main steps. The first step involves self-supervised pre-training based on Autoregressive Mamba (ARM) [33], where labeled data is treated as unlabeled data and used as part of the training samples. The second step involves supervised fine-tuning using only labeled data. The third step involves SSL. These three steps are detailed next.", + "bbox": [ + 89, + 763, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised pre-training based on ARM: To improve feature extraction, we propose to leverage ARM", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for self-supervised pre-training of Semi-ViM. ARM is better suited for ViM pre-training than Masked Autoencoders (MAEs) and DINO-V2 (contrastive learning) [32] due to its alignment with ViM's sequential state-space modeling [33]. Unlike MAEs, which rely on masked image reconstruction and bidirectional Transformers, Semi-ViM's bidirectional state-space modeling allows to effectively leverage ARM's autoregressive token prediction, making ARM a more natural fit [17]. Compared to DINO-V2, which depends on global feature alignment and large-batch contrastive learning, ARM directly optimizes sequential token dependencies, avoiding instability from batch size sensitivity and augmentation complexity [9]. By leveraging ViM's structured recurrence, ARM enables faster training, better scalability, and improved efficiency over both methods. These statements are validated by our experiments in Section 5.", + "bbox": [ + 511, + 90, + 906, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fully-supervised fine-tuning: After completing ARM-based pre-training, Semi-ViM undergoes fine-tuning using only labeled data. This process aims to leverage the representations learned during pre-training to adapt the model to specific downstream tasks. During fine-tuning, most pretrained parameters remain unchanged, with only certain layers or global parameters being updated to a limited extent in order to improve performance in SSL. Fine-tuning specifically targets layers responsible for learning bidirectional dependencies, i.e., the forward and backward SSMs are updated to better capture dependencies specific to a classification task. The class token used in image classification tasks is also fine-tuned during this phase.", + "bbox": [ + 511, + 347, + 908, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSL: In the third training step, we use LyapEMA, in conjunction with SSMixup to fully exploit unlabeled data while ensuring training stability and efficiency. This last training step consists of five key components: weak and strong data augmentation, pseudo-label generation, LyapEMA, SSMixup, and minimization of the loss function.", + "bbox": [ + 511, + 545, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each unlabeled sample $x^{u}$ , we generate a weakly augmented sample $x_{w}^{u}$ and a strongly augmented sample $x_{s}^{u}$ . The weakly augmented sample $x_{w}^{u}$ is processed by the teacher model $f_{\\text{teacher}}$ to produce a set of predicted probabilities, one for each class:", + "bbox": [ + 511, + 636, + 905, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\text {t e a c h e r} _ {x _ {w} ^ {u}}} = f _ {\\text {t e a c h e r}} \\left(x _ {w} ^ {u}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 728, + 906, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pseudo-label for $x_w^u$ is then determined as the class associated with the highest probability in set $p_{\\text{teacher}_x u}$ :", + "bbox": [ + 511, + 753, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {u} = \\arg \\max \\left(p _ {\\text {t e a c h e r} _ {x _ {w}} ^ {u}}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 800, + 903, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that only unlabeled samples with a class probability exceeding the threshold $\\tau$ are retained for training; i.e., only those with high-score pseudo-labels are retained. SSMixup is then applied to these unlabeled samples with high-score pseudo labels in order to be used to train the student model.", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "769", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fa5bf86679e1effb6541121f76e6b04eb9398562b3d48ea47b95b2144f768e48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Variant nameBackboneParam.Input size
Semi-ViM-TinyViM-Tiny [47]7M2242
Semi-ViM-SmallViM-Small [47]26M2242
Semi-ViM-BaseViM-Base [47]98M2242
", + "bbox": [ + 94, + 88, + 480, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The different variants of Semi-ViM.", + "bbox": [ + 151, + 170, + 419, + 184 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall loss function consists of the supervised loss $L_{\\text{labeled}}$ and the unsupervised loss $L_{\\text{unlabeled}}$ :", + "bbox": [ + 91, + 208, + 482, + 239 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {l a b e l e d}} + \\mu L _ {\\text {u n l a b e l e d}}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 253, + 480, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The supervised loss, $L_{\\mathrm{labeled}}$ , is based on the standard cross-entropy (CE) loss using labeled samples:", + "bbox": [ + 89, + 273, + 482, + 304 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {l a b e l e d}} = \\frac {1}{N _ {l}} \\sum_ {i = 1} ^ {N _ {l}} \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 315, + 482, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The unsupervised loss $L_{\\mathrm{unlabeled}}$ uses strongly augmented samples, $x_{s}^{u}$ as follows:", + "bbox": [ + 89, + 358, + 482, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {u n l a b e l e d}} = \\frac {1}{N _ {u}} \\sum_ {i = 1} ^ {N _ {u}} \\mathbb {1} \\left(o _ {i} > \\tau\\right) \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {s, i} ^ {u}\\right), \\hat {y} _ {i} ^ {u}\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 407, + 482, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the indicator function $\\mathbb{1}(o > \\tau)$ selects only those unlabeled samples with high-score pseudo-labels, where $o$ denotes the probability of the class associated with the pseduo-label. Finally, the teacher model is updated using LyapEMA to stabilize parameter evolution and improve the quality of pseudo-labels.", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 558, + 223, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 lists the different variants of Semi-ViM we use in our experiments: Tiny, Small, and Base. These variants are trained on ImageNet-1K using 8 NVIDIA A100 (80GB) GPUs with DistributedDataParallel (DDP) PyTorch. We use AdamW as the optimizer with a weight decay of 0.05, and a learning rate that follows a cosine annealing schedule starting from $5 \\times 10^{-4}$ with 10 epochs of warm-up. We use 300 epochs for training with a global batch size of 1024 (128 per GPU) and automatic mixed precision (AMP) enabled. Strong data augmentation includes AutoAugment, Mixup ( $\\alpha = 0.2$ ), CutMix ( $\\alpha = 1.0$ ), and label smoothing ( $\\varepsilon = 0.1$ ). The total training time is approximately 70 hours.", + "bbox": [ + 89, + 583, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct classification experiments on ImageNet-1K [9] and ImageNet-LT [28]. ImageNet-1K contains 1.28M training and 50K validation images across 1,000 classes. ImageNet-LT has a long-tailed class distribution. We use a confidence threshold of $\\tau = 0.7$ for pseudo-label selection, with weak and strong data augmentations applied to unlabeled data. Inference is performed using a $224\\times 224$ center crop on the images.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Results and analysis", + "text_level": 1, + "bbox": [ + 513, + 90, + 705, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semi-ViM demonstrates superior performance across multiple experimental settings as tabulated in Table 2. On ImageNet-1K, Semi-ViM-Base achieves $85.40\\%$ accuracy with only $10\\%$ of the labeled data, outperforming SemiViT-Huge. Compared to SemiFormer, Semi-ViM consistently excels, confirming the suitability of ViM for SSL.", + "bbox": [ + 511, + 112, + 903, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the more challenging ImageNet-LT dataset, Semi-ViM-Base achieves $77.40\\%$ accuracy with only $10\\%$ of the labeled data, significantly surpassing Semi-ViT-Huge and SemiFormer. This demonstrates Semi-ViM's adaptability to long-tailed data distributions and its capability to work with rare category representations. With only $1\\%$ of the labeled data, Semi-ViM-Base achieves $66.30\\%$ accuracy, surpassing Semi-ViT-Huge and demonstrating its strength in SSL.", + "bbox": [ + 511, + 203, + 903, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of the Top-5 accuracy on ImageNet-LT with only $1\\%$ of the labeled data (last column of Table 2), Semi-ViM-Base outperforms Semi-ViT-Huge by $6.17\\%$ and CNN-based methods, like FixMatch and PAWS, by over $15\\%$ , highlighting its superior feature representation, class diversity learning, and robustness against data imbalance.", + "bbox": [ + 511, + 324, + 903, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 tabulates the Top-1 Average Precision (AP) on ImageNet-LT of several fully-supervised models trained exclusively with labeled data. Semi-ViM-Base with just $10\\%$ of the labeled data achieves competitive accuracy. This suggests that our framework efficiently extracts meaningful representations even with significantly fewer labeled samples, making it highly suitable for real-world scenarios where labeled data may be scarce.", + "bbox": [ + 511, + 414, + 903, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 graphically summarizes the performance of Semi-ViM and that of other SOTA SSL models and fully-supervised models.", + "bbox": [ + 511, + 536, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation study", + "text_level": 1, + "bbox": [ + 513, + 590, + 661, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several pre-training strategies, parameter updating methods, and mixup strategies are compared in Table 4, using Semi-ViM-Base. ARM pre-training consistently achieves the best performance, demonstrating its effectiveness in leveraging sequential dependencies inherent to the ViM. ARM reaches $81.90\\%$ (1% of the labeled data) and $85.40\\%$ (10% of the labeled data) when used in conjunction with LyapEMA+SSMixup, significantly outperforming MAEs and DINO-V2 across all settings. This confirms that ARM, designed specifically for sequential state-space models, is a more appropriate fit for ViM than contrastive learning (DINO-V2) or MAEs.", + "bbox": [ + 511, + 613, + 903, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of parameter updating methods, LyapEMA outperforms EMA in all configurations, particularly when used in conjunction with ARM. Unlike EMA, LyapEMA stabilizes updates by leveraging Lyapunov theory, preventing pseudo-label collapse and ensuring more reliable parameter evolution. This is particularly beneficial in low-labeled data scenarios, where training instability is a key challenge.", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "770", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b9948231be431d4eb495d93228581730a8a8dedd222aba536ba81ce3e312f3b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CNN-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
FixMatch [38]ResNet-5025M-67.8371.7425M-52.3257.68-
PAWS [1]ResNet-5027M66.5469.2176.5527M46.3056.1361.8155.19
FreeMatch [42]ResNet-5026M55.4270.1174.5026M-55.5460.94-
SoftMatch [5]ResNet-5028M-70.4076.2028M-57.4161.02-
SimCLRv2(self-distilled) [6]ResNet-15228M76.6578.2380.9028M48.5559.3062.1067.28
SimCLRv2(distilled) [6]ResNet-5028M75.9077.4080.2028M47.7057.2360.8168.50
Co-training methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
Co-Training [35]CT(MLP)56M80.0381.6581.7056M58.6264.7971.2078.30
MCT [35]MCT (MLP)176M80.7582.0885.20176M63.1872.6075.3080.11
Transformer-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top-5 (1%)
DINO [14]ViT-Small22M65.2171.4572.5322M47.8055.3059.9064.25
SemiFormer [43]ViT-S+Conv42M-72.8073.2842M50.1058.0063.5067.87
DPT [44]ViT-Huge604M80.2082.5084.06604M55.9164.7668.1575.40
Semi-ViT-Base [4]ViT-Base86M71.3077.5279.9086M53.2062.9068.4071.50
Semi-ViT-Large [4]ViT-Large307M76.8180.5082.26307M56.8066.1071.2074.70
Semi-ViT-Huge [4]ViT-Huge632M80.4783.2084.27632M58.9068.4073.5076.90
REACT [25]ViT-Huge651M79.1383.2384.74632M59.3668.7173.0877.45
Semi-ViM-TinyViM-Tiny12M76.1078.3081.0012M52.1362.9068.3071.11
Semi-ViM-SmallViM-Small27M80.3081.4083.1027M63.2472.1574.6880.02
Semi-ViM-BaseViM-Base146 M81.9083.2585.40146M66.3075.5277.4083.07
", + "bbox": [ + 94, + 88, + 903, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Accuracy (%) of the different variants of Semi-ViM (in pink) and other SOTA SSL models over 5 folds when 1%, 5% and 10% of the labeled data is used for training. The best performance of other SOTA models is highlighted in peach.", + "bbox": [ + 89, + 492, + 906, + 527 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74f9f7d7480c003c75db1f08bda0e673bc73c1086c6c34d5682cc6bad7ae9c7a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 556, + 352, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13d152dcee76c156b3a779e93ab1e54c4b61496da5c35430cc8c567eee0211d9.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 372, + 556, + 624, + 710 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/040ac5484445f07bd10ca732b9919d3e188c386cb24f38030b3c4a28a93b3445.jpg", + "image_caption": [ + "(c)", + "Figure 4. (a) Training accuracy $(\\%)$ on ImageNet-1K of Semi-ViM-Base and other SOTA SSL methods using only $10\\%$ of the labeled data. (b) Top-1 accuracy $(\\%)$ on ImageNet-LT $(1\\%$ of the labeled data) of Semi-ViM and other SOTA SSL algorithms across different model scales. (c) Top-1 Average Precision $(\\%)$ on ImageNet-LT of Semi-ViM-Base and several fully supervised models." + ], + "image_footnote": [], + "bbox": [ + 645, + 568, + 901, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding mixup strategies, SSMixup provides the largest performance boost, outperforming ProbMixup consistently across all pre-training methods. With ARM, LyapEAM+SSMixup boosts accuracy by $2 - 3\\%$ compared to ProbMixup. This confirms that SSMixup is highly effective", + "bbox": [ + 89, + 821, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for SSL, mitigating pseudo-label noise and improving generalization in low-data settings. The combination of ARM pre-training, and LyapEMA+SSMixup achieves the highest accuracy, underscoring the importance of pre-training alignment, stable parameters updates, and structured aug-", + "bbox": [ + 511, + 821, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "771", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6e402a4a8b9ef0cb506acfd93778c776668c696281363002e8e77095b0f96ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelLabeled dataTop-1 AP
LIFT(ViT-L/14) [37]100%82.90
MAM(ViT-B/16) [18]100%82.30
LIFT(ViT-B/16) [37]100%78.30
BALLAD (ResNet-50) [29]100%76.50
Semi-ViT-Huge [4]10%73.50
Semi-ViM-Small10%74.68
Semi-ViM-Base5%75.52
Semi-ViM-Base10%77.40
", + "bbox": [ + 91, + 88, + 493, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ea01db5993150fd38a4d2a356a2189423d077c8d1b820aa3aa8daec505428f3.jpg", + "table_caption": [ + "Table 3. Top-1 AP (%) of fully-supervised models and the variants of Semi-ViM on ImageNet-LT. Semi-ViM-Base (in purple), using only $10\\%$ of the labeled data, outperforms BALLAD (ResNet $50\\times 16$ )." + ], + "table_footnote": [], + "table_body": "
PretrainedMethod1%10%
Nonefinetune--
EMA42.1147.41
LyapEMA44.7353.26
+ProbMixup44.7849.57
+SSMixup45.2954.73
MAEfinetune63.0573.14
EMA71.3074.20
LyapEMA65.3075.65
+ProbMixup66.0975.87
+SSMixup66.0276.15
DINO-V2finetune62.0272.63
EMA63.4573.57
LyapEMA64.8475.21
+ProbMixup65.5374.04
+SSMixup65.3276.07
ARMfinetune66.5374.21
EMA72.5379.90
LyapEMA76.0278.03
+ProbMixup79.8380.52
+SSMixup81.9085.40
", + "bbox": [ + 130, + 342, + 444, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mentation in semi-supervised models based on ViM.", + "bbox": [ + 89, + 806, + 439, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5 compares different parameter updating methodologies, demonstrating the superior stability and generalization capability of LyapEMA in SSL. The EMA-Teacher suffers from training instability, particularly in Semi-ViT-Small and Semi-ViM-Small with only $1\\%$ of the la", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac97d9fb9485cc723e41253ccdf9961345742cafc511bb0114d5144e997a5a24.jpg", + "table_caption": [ + "Table 4. Top-1 accuracy (%) on ImageNet-1K of Semi-ViM-Base under different pre-training strategies and parameter updating methods using $1\\%$ and $10\\%$ of the labeled data for training. \" $^+$ \" means the mixup strategy used in conjunction with LyapEMA. Best results are highlighted in bold font." + ], + "table_footnote": [], + "table_body": "
ModelMethod1%10%
FreeMatch [42]FixMatch [38]--
EMA-Teacher [3]55.4274.50
LyapEMA56.7075.07
Semi-ViT-S [4]FixMatch [38]-×
EMA-Teacher [3]×75.80
LyapEMA61.3076.60
Semi-ViT-B [4]FixMatch [38]-×
EMA-Teacher [3]66.4178.15
LyapEMA68.8480.43
Semi-ViM-SFixMatch [38]××
EMA-Teacher [3]×74.01
LyapEMA80.3085.40
Semi-ViM-BFixMatch [38]×71.28
EMA-Teacher [3]70.5376.11
LyapEMA81.9085.40
", + "bbox": [ + 522, + 88, + 897, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Top-1 accuracy (%) on ImageNet-1K of FixMatch, EMA-Teacher and LyapEMA with different models using $1\\%$ and $10\\%$ of the labeled data for training. “×” means the training failed (accuracy near 0). “-” means no result is available. Results in light blue indicate risk of training instability and collapse.", + "bbox": [ + 511, + 382, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "beled data, where training fails to converge. In contrast, LyapEMA effectively prevents collapse, achieving $61.30\\%$ and $80.30\\%$ accuracy, respectively, showing its robustness in low-labeled data scenarios. By dynamically adapting the EMA decay factor based on the student-teacher discrepancy, LyapEMA mitigates pseudo-label bias and enhances convergence. Additional ablation studies and the model's performance under our designed extreme imbalance scenarios are provided in the supplementary material.", + "bbox": [ + 511, + 483, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Semi-ViM, a novel semi-supervised learning framework that leverages bidirectional SSMs to mitigate learning bias caused by label imbalance in classification tasks. By integrating LyapEMA, a novel stability-aware parameter update mechanism, and SS-Mixup, a hidden-state-level mixing strategy, Semi-ViM enhances both training stability and feature learning. Our extensive experiments on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms CNN-based and Transformer-based SSL models, achieving superior accuracy with fewer labeled samples. Moreover, Semi-ViM requires far fewer parameters than ViT-based SSL models. These findings highlight the potential of state space models as a competitive alternative to Transformers in semi-supervised learning, particularly in addressing real-world imbalanced datasets.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "772", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8443-8452, 2021. 7", + "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3", + "[3] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 8", + "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. Advances in Neural Information Processing Systems, 35:25697-25710, 2022. 1, 2, 3, 7, 8", + "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921, 2023. 7", + "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7", + "[7] Nieves Crasto. Class imbalance in object detection: An experimental diagnosis and study of mitigation strategies. arXiv preprint arXiv:2403.07113, 2024. 2", + "[8] Charika De Alvis and Suranga Seneviratne. A survey of deep long-tail classification advancements. arXiv preprint arXiv:2404.15593, 2024. 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6", + "[10] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Towards semi-supervised learning with nonrandom missing labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16121-16131, 2023. 2", + "[11] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine-grained learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11829-11837, 2024. 2", + "[12] Marawan Elbatel, Hualiang Wang, Jixiang Chen, Hao Wang, and Xiaomeng Li. Learning unlabeled clients divergence via anchor model aggregation for federated semi-supervised learning. arXiv e-prints, pages arXiv-2407, 2024. 2", + "[13] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 2", + "[14] Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3187-3197, 2023. 7", + "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2", + "[16] Lan-Zhe Guo, Lin-Han Jia, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning in open environments. Frontiers of Computer Science, 19(8):198345, 2025. 1", + "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5", + "[18] Ahmet Iscen, Alireza Fathi, and Cordelia Schmid. Improving image recognition by retrieving from web-scale image-text data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19295-19304, 2023. 8", + "[19] Zhen Jiang, Lingyun Zhao, Yu Lu, Yongzhao Zhan, and Qirong Mao. A semi-supervised resampling method for class-imbalanced learning. Expert Systems with Applications, 221:119733, 2023. 2", + "[20] Suraj Kothawade, Pavan Kumar Reddy, Ganesh Ramakrishnan, and Rishabh Iyer. Basil: Balanced active semi-supervised learning for class imbalanced datasets. arXiv preprint arXiv:2203.05651, 2022. 1", + "[21] Alex Kurakin, Colin Raffel, David Berthelot, Ekin Dogus Cubuk, Han Zhang, Kihyuk Sohn, and Nicholas Carlini. Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. 2020. 2", + "[22] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2", + "[23] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2", + "[24] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 2", + "[25] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 7", + "[26] Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy. arXiv preprint arXiv:2405.04404, 2024. 3", + "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "773", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2", + "[28] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6", + "[29] Teli Ma, Shijie Geng, Mengmeng Wang, Jing Shao, Jiasen Lu, Hongsheng Li, Peng Gao, and Yu Qiao. A simple long-tailed recognition baseline via vision-language model. arXiv preprint arXiv:2111.14745, 2021. 8", + "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2", + "[31] Khanh-Binh Nguyen. Sequencematch: revisiting the design of weak-strong augmentations for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 96-106, 2024. 3", + "[32] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5", + "[33] Sucheng Ren, Xianhang Li, Haoqin Tu, Feng Wang, Fangxun Shu, Lei Zhang, Jieru Mei, Linjie Yang, Peng Wang, Heng Wang, et al. Autoregressive pretraining with mamba in vision. arXiv preprint arXiv:2406.07537, 2024. 5", + "[34] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 3", + "[35] Jay C. Rothenberger and Dimitrios I. Diochnos. Meta cotraining: Two views are better than one. In 2025 Joint Mathematics Meetings (JMM 2025). AMS. 7", + "[36] Shankar Sastry and Shankar Sastry. Lyapunov stability theory. Nonlinear systems: analysis, stability, and control, pages 182-234, 1999. 3", + "[37] Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. Long-tail learning with foundation model: Heavy fine-tuning hurts. In *Forty-first International Conference on Machine Learning*, 2024. 8", + "[38] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 7, 8", + "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 1", + "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Renzhen Wang, Xixi Jia, Quanziang Wang, Yichen Wu, and Deyu Meng. Imbalanced semi-supervised learning with bias adaptive classifier. arXiv preprint arXiv:2207.13856, 2022.1", + "[42] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 7, 8", + "[43] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In European conference on computer vision, pages 605-620. Springer, 2022. 1, 2, 7", + "[44] Zebin You, Yong Zhong, Fan Bao, Jiacheng Sun, Chongxuan Li, and Jun Zhu. Diffusion models and semi-supervised learners benefit mutually with few labels. Advances in Neural Information Processing Systems, 36:43479-43495, 2023. 7", + "[45] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2", + "[46] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8373-8383, 2023. 1", + "[47] L Zhu, B Liao, Q Zhang, X Wang, W Liu, and X Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arxiv 2024. arXiv preprint arXiv:2401.09417. 1, 2, 3, 6" + ], + "bbox": [ + 514, + 90, + 903, + 555 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "774", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_model.json b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e50bacb6a09f859b4b3c7b7d138c5fe828996da --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Semi-ViM: Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.204, + 0.368, + 0.24 + ], + "angle": 0, + "content": "Hongyang He* \nUniversity of Warwick" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.204, + 0.598, + 0.24 + ], + "angle": 0, + "content": "Hongyang Xie \nUniversity of Warwick" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.204, + 0.814, + 0.24 + ], + "angle": 0, + "content": "Haochen You Columbia University" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.247, + 0.589, + 0.283 + ], + "angle": 0, + "content": "Victor Sanchez \nUniversity of Warwick" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.317, + 0.327, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.349, + 0.486, + 0.606 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) is often hindered by learning biases when imbalanced datasets are used for training, which limits its effectiveness in real-world applications. In this paper, we propose Semi-ViM, a novel SSL framework based on Vision Mamba, a bidirectional state space model (SSM) that serves as a superior alternative to Transformer-based architectures for visual representation learning. Semi-ViM effectively deals with imbalanced datasets and improves model stability through two key innovations: LyapEMA, a stability-aware parameter update mechanism inspired by Lyapunov theory, and SSMixup, a novel mixup strategy applied at the hidden state level of bidirectional SSMs. Experimental results on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms state-of-the-art SSL models, achieving \\(85.40\\%\\) accuracy with only \\(10\\%\\) of the labeled data, surpassing Transformer-based methods such as Semi-ViT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.222, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) aims to mitigate the scarcity of large-scale labeled training data while reducing annotation costs [16]. In the context of classification, however, a challenge remains: Most state-of-the-art (SOTA) SSL models are trained under the assumption that access to a dataset with a balanced number of labeled samples per class is possible, which is often overly idealized and deviates significantly from real-world conditions [20]. In practice, label imbalance induces learning biases in SSL, ultimately compromising model performance [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Due to its outstanding performance, many studies have integrated the vision Transformer (ViT) into SSL methods [4, 43, 46]; e.g., Semi-ViT, adopts a pre-training and fine-tuning approach to achieve SOTA results [4]. How-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.453, + 0.908, + 0.524 + ], + "angle": 0, + "content": "Figure 1. Semi-ViT-Small trained with only \\(1\\%\\) of the labeled data of ImageNet-1K. Left: Training fails when FixMatch is used. The loss decreases but test accuracy remains near zero. Right: Training is unstable and fails when EMA-Teacher is used: The loss and accuracy oscillate heavily." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.749 + ], + "angle": 0, + "content": "ever, Semi-ViT has a weaker inductive bias, requiring large-scale data for effective training [39]. Moroever, as shown by our experiments, Semi-ViT does not perform well on long-tailed imbalanced datasets, like ImageNet-LT, confirming its limitations in handling label imbalance. Although the performance of Semi-ViT can be improved by increasing the model size, e.g., Semi-ViT-Huge, this comes at the expense of a larger number of trainable parameters, around 632M [4]. Vision Mamba (ViM), which incorporates a bidirectional state space model (SSM), has modeling capabilities comparable to those of ViT, while achieving subquadratic-time computational and linear memory complexities [47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Based on the previous discussions, this paper integrates ViM into an SSL framework for classification and proposes a ViM-based SSL model, hereinafter called Semi-ViM. Semi-ViM incorporates two key innovations aimed at improving model stability during training and performance: LyapEMA and SSMixup. LyapEMA is a novel parameter update mechanism that leverages the Lyapunov theory to dynamically adjust parameter updates, ensuring a more stable training process. LyapEMA addresses the training failures encountered when using only a small amount of labeled" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "765" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "data to train Semi-ViT-Small, as shown in Fig 1. LyapEMA also helps prevent training collapse, commonly encountered when using the exponential moving average (EMA) and FixMatch parameter update mechanisms [4]. SSMixup, on the other hand, is a novel mixup strategy that operates at the hidden state space of SSMs and goes beyond the traditional input-level mixup strategy. Specifically, SSMixup incorporates mixed hidden states as residual units in the training process, improving the model's capability to generalize under sparse label conditions. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.317 + ], + "angle": 0, + "content": "- We propose Semi-ViM, an SSL framework that outperforms other SSL methods based on CNNs and Transformers. The number of parameters of Semi-ViM is only around one-fourth of that of Semi-ViT." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.318, + 0.483, + 0.379 + ], + "angle": 0, + "content": "- We propose LyapEMA, a novel parameter update mechanism that successfully addresses the training instability issues commonly found in Semi-ViT, thus improving model convergence and generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.379, + 0.482, + 0.422 + ], + "angle": 0, + "content": "- We propose SSMixup, a hidden state space-level mixup strategy that effectively enhances the learning capability of bidirectional SSMs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.424, + 0.483, + 0.469 + ], + "angle": 0, + "content": "- Based on evaluations on the long-tailed ImageNet-LT dataset, we demonstrate Semi-ViM's strong capabilities in mitigating learning biases caused by label imbalance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.258, + 0.483, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.233, + 0.499 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Imbalanced datasets are those in which the different classes are distributed unequally, often resulting in long-tailed distributions [7]. SSL aims to leverage a small amount of labeled data and a large amount of unlabeled data to improve performance, which is useful to deal with imbalanced datasets. SSL has achieved significant advancements in several computer vision tasks [10, 11, 21, 38, 45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Traditional SSL methods, e.g., entropy minimization and consistency regularization [13, 15], exploit unlabeled data to reduce dependence on manual annotation. For example, consistency regularization leverages unlabeled data under the assumption that a model's predictions remain consistent under different input perturbations. Mainstream methods include II-Model [22], Mean Teacher [40], Virtual Adversarial Training (VAT) [30], and FixMatch [38], which is an important benchmark in SSL [38]. Pseudo-Labeling [23] is a common strategy used in SSL. It consists of using the model's predictions on unlabeled data as pseudo-labels for training. Noisy Student, FixMatch [38], and FlexMatch [45] adopt this strategy and use confidence thresholds to filter out pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Several SOTA SSL methods are based on FixMatch or any of its variants, e.g., FlexMatch and CoMatch [24], which integrate contrastive learning or self-supervised methods to enhance performance. However, these methods typically assume that the class distribution is balanced," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "which is rarely the case in real-world scenarios [12]. When dealing with imbalanced datasets [8], these SOTA SSL methods face two primary challenges. First, pseudo-label biases, i.e., models tend to predict more frequently the majority classes when generating pseudo-labels, exacerbating the imbalance issue and limiting learning from the minority classes [19]. Second, low-quality pseudo-labels, i.e., models tend to assign low confidence scores to those pseudolabels generated for the minority class samples, thus reducing their impact on training and potentially leading to training collapse [4]. For example, FixMatch relies on a fixed confidence threshold for selecting pseudo-labels, but in an imbalanced dataset setting, pseudo-labels of the minority classes often fail to reach this threshold, leading to insufficient learning from such classes [4, 43]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.331, + 0.652, + 0.347 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.553 + ], + "angle": 0, + "content": "State Space Models and ViM: SSMs have a long-standing history in control systems and signal processing, offering a structured approach to sequential data modeling [47]. Deep learning adopts SSMs to more effectively handle long-range dependencies compared to traditional recurrence-based models. Unlike Transformers, which rely on global self-attention, SSMs leverage structured state transitions for sequential data propagation, enabling subquadratic computational complexity and linear memory scaling [47]. In computer vision, SSMs have been integrated into deep architectures to enhance representation learning by modeling both spatial and temporal dependencies [27, 47]. The SSM is formulated as a discrete-time state-space system [47]:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.563, + 0.905, + 0.58 + ], + "angle": 0, + "content": "\\[\nh _ {t} = A h _ {t - 1} + B x _ {t}, \\quad y _ {t} = C h _ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.905, + 0.693 + ], + "angle": 0, + "content": "where \\(A \\in \\mathbb{R}^{N \\times N}\\) is the state transition matrix; \\(B \\in \\mathbb{R}^{N \\times 1}\\) and \\(C \\in \\mathbb{R}^{1 \\times N}\\) are input and output projection matrices, respectively; \\(x_{t}\\) represents the input sequence at time step \\(t\\); and \\(y_{t}\\) is the corresponding output representation. The parameters of an SSM are usually optimized in a data-dependent manner to ensure computational efficiency and parallelization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.905, + 0.784 + ], + "angle": 0, + "content": "The bidirectional extension of SSMs further improves feature learning by capturing both forward and backward information, making it a competitive alternative to self-attention mechanisms used by Transformers. [47]. The bidirectional SSM operates in the forward \\((f)\\) and backward \\((b)\\) direction:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.793, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\nh _ {t} ^ {f} = A h _ {t - 1} ^ {f} + B x _ {t}, \\quad y _ {t} ^ {f} = C h _ {t} ^ {f}; \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.821, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\nh _ {t} ^ {b} = A h _ {t + 1} ^ {b} + B x _ {t}, \\quad y _ {t} ^ {b} = C h _ {t} ^ {b}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.905, + 0.875 + ], + "angle": 0, + "content": "The final output at time step \\( t \\) is obtained by merging the forward and backward outputs through a gating mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b}, \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "766" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.329, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.277, + 0.235, + 0.288 + ], + "angle": 0, + "content": "(a) FixMatch" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.091, + 0.62, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.277, + 0.531, + 0.289 + ], + "angle": 0, + "content": "(b) EMA-Teacher" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.092, + 0.91, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.277, + 0.832, + 0.289 + ], + "angle": 0, + "content": "(c) Semi-ViM(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.303, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Differences between (a) FixMatch, (b) EMA-Teacher and (c) Semi-ViM (ours). In the figure, \\( X_{s} \\) and \\( X_{w} \\) are, respectively, strongly and weakly augmented inputs; \\( p^{s} \\) and \\( p_{student}^{s} \\) are the predictions of the student model on \\( X_{s} \\); \\( p^{w} \\) and \\( p_{teacher}^{w} \\) are the predictions of the teacher model on \\( X_{w} \\); \\( \\tau \\) is the threshold used to select pseudo-labels; and \\( \\theta, \\theta^{student}, \\theta^{teacher} \\) are, respectively, the learnable parameters for the entire model, the student model, and the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.431 + ], + "angle": 0, + "content": "where \\(\\mathrm{SiLU}(\\cdot)\\) is the Sigmoid-weighted Linear Unit activation function, and \\(z\\) is a gating vector that adaptively balances bidirectional information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.563 + ], + "angle": 0, + "content": "ViM adopts a bidirectional SSM for visual representation learning and integrates positional embeddings for spatial awareness. In ViM, the input image is first partitioned into patches and projected onto a high-dimensional feature space before being processed. With a linear computational complexity, ViM provides data-dependent global visual modeling, significantly improving efficiency in high-resolution tasks compared to Transformer-based architectures [26, 47]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.573, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Mixup strategies: These strategies interpolate input samples and labels, and hence can be used as data augmentation techniques to improve generalization capabilities [2]. In SSL, they can enhance pseudo-labeling and provide consistency regularization [34]. Recent variants extend mixup strategies to the feature space and pseudo-label frameworks, e.g., the probabilistic pseudo mixup strategy refines pseudolabel quality by weighting interpolated samples based on their confidence scores, which are derived from the model's softmax outputs or uncertainty estimates. These variants can improve SSL robustness, particularly for ViTs, which benefit from stronger regularization [4]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Weak and strong data augmentation: Weak data augmentation involves simple transformation, e.g., random cropping, flipping, or slight color jitter, which introduce small variations to the original data to help a model learn from basic invariances. Strong data augmentation, on the other hand, applies more aggressive transformations, e.g., large crops, strong color distortions, or more significant perturbations, which force a model to learn more robust representations by simulating more challenging scenarios [31]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.385, + 0.806, + 0.403 + ], + "angle": 0, + "content": "4. Proposed Semi-ViM framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.563 + ], + "angle": 0, + "content": "Semi-ViM adopts a teacher-student framework and uses a labeled dataset, \\(\\mathcal{X}_l = \\{(x_i^l,y_i^l)\\}_{i = 1}^{N_l}\\), and an unlabeled dataset \\(\\mathcal{X}_u = \\{x_i^u\\}_{i = 1}^{N_u}\\) for training. Specifically, the teacher model, \\(f_{\\mathrm{teacher}}(\\theta^{\\mathrm{teacher}})\\), generates pseudo-labels, \\(\\{\\hat{y}_i\\}\\), for unlabeled samples, \\(\\{x_i^u\\}\\), and outputs confidence scores, \\(\\{o_i\\}\\), for the pseudo-labels. The student model, \\(f_{\\mathrm{student}}(\\theta^{\\mathrm{student}})\\), is then trained on samples mixed according to their confidence scores. The difference between the training paradigm of Semi-ViM and those of FixMatch and EMA-Teacher is shown in Fig 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.701 + ], + "angle": 0, + "content": "In this section, we first introduce LyapEMA, which is inspired by Lyapunov stability theory [36] and stabilizes parameter updates, thus preventing model collapse during the training of Semi-ViM. Next, we introduce SSMixup, a new mixup strategy designed for SSM-based models. Unlike traditional mixup strategies that operate solely on the input level, SSMixup mixes the hidden states of bidirectional SSMs. Finally, we detail the training and loss function of Semi-ViM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.633, + 0.728 + ], + "angle": 0, + "content": "4.1. LyapEMA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.87 + ], + "angle": 0, + "content": "LyapEMA is a novel optimization framework that integrates Lyapunov stability theory with EMA. LyapEMA is designed for teacher-student frameworks under an SSL setting. It has two main objectives. The first is to ensure stable parameter updates, because training SSL frameworks, like Semi-ViT, with a limited amount of labeled data can lead to training collapse. The second objective is reducing the learning bias associated with pseudo-labels while preserving the advantages of EMA in refining the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Let \\(\\theta^{\\mathrm{student}}\\) and \\(\\theta^{\\mathrm{teacher}}\\) denote the student and teacher model parameters, respectively. EMA updates the teacher" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "767" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.362, + 0.106 + ], + "angle": 0, + "content": "model using gradient descent as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.12, + 0.482, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {\\text {t e a c h e r}} = \\alpha_ {t} \\theta_ {t - 1} ^ {\\text {t e a c h e r}} + (1 - \\alpha_ {t}) \\theta_ {t} ^ {\\text {s t u d e n t}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.143, + 0.484, + 0.262 + ], + "angle": 0, + "content": "where \\(\\alpha \\in (0,1)\\) is the EMA decay factor. However, EMA cannot guarantee stability, potentially leading to divergence due to accumulated bias associated with learning from pseudo-labels. To address this issue, LyapEMA introduces Lyapunov stability constraints to guide the updates. Let us define a Lyapunov candidate function, \\(V(\\theta^{\\mathrm{student}},\\theta^{\\mathrm{teacher}})\\), to measure the discrepancy between the student and teacher models:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.284, + 0.483, + 0.313 + ], + "angle": 0, + "content": "\\[\nV \\left(\\theta^ {\\text {s t u d e n t}}, \\theta^ {\\text {t e a c h e r}}\\right) = \\frac {1}{2} \\| \\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Note that the function in Eq. 6 acts as a stability indicator, ensuring that the student model does not drift too far from the teacher model. The Lyapunov stability criterion requires that the derivative of \\( V(\\cdot, \\cdot) \\) w.r.t. time, i.e., \\( \\dot{V} \\), satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.414, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\dot {V} = \\frac {d V}{d t} = \\left(\\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}}\\right) ^ {\\mathrm {T}} \\left(\\dot {\\theta} ^ {\\text {s t u d e n t}} - \\dot {\\theta} ^ {\\text {t e a c h e r}}\\right) < 0, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.514 + ], + "angle": 0, + "content": "where \\(\\dot{\\theta}^{\\mathrm{student}}\\) and \\(\\dot{\\theta}^{\\mathrm{teacher}}\\) also derivatives w.r.t. time. The criterion in Eq. 7 guarantees that the discrepancy between the student and teacher models decreases over time, ensuring convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.56 + ], + "angle": 0, + "content": "To enforce stability while leveraging EMA for smooth teacher model updates, LyapEMA updates the student model parameters by using aLyapunov regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.581, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t} ^ {\\text {s t u d e n t}} = \\theta_ {t - 1} ^ {\\text {s t u d e n t}} - \\eta \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) + \\lambda \\left(\\theta_ {t} ^ {\\text {t e a c h e r}} - \\theta_ {t} ^ {\\text {s t u d e n t}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.652 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\theta^{\\mathrm{student}})\\) is a semi-supervised loss function, \\(\\eta\\) is the learning rate, and \\(\\lambda\\) is a stability coefficient that ensures the student model remains close to the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.713 + ], + "angle": 0, + "content": "LyapEMA updates the teacher model parameters by using EMA as denoted by Eq. 5. However, the EMA decay factor \\(\\alpha_{t}\\) is dynamically adjusted based on the Lyapunov function:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.727, + 0.482, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t} = \\operatorname {s i g m o i d} \\left(\\gamma \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t - 1} ^ {\\text {t e a c h e r}} \\| _ {2}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is a sensitivity hyperparameter controlling how \\(\\alpha_{t}\\) adapts based on the student-teacher discrepancy, and the sigmoid function ensures \\(\\alpha_{t}\\) remains within (0, 1), preventing instability in the updates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.84 + ], + "angle": 0, + "content": "LyapEMA ensures that updates lead to stable training by using the Lyapunov decrease condition:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.86, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\nV _ {t + 1} - V _ {t} = \\frac {1}{2} \\| \\theta_ {t + 1} ^ {\\text {s t u d e n t}} - \\theta_ {t + 1} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2} - \\frac {1}{2} \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.092, + 0.677, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 LyapEMA" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.905, + 0.156 + ], + "angle": 0, + "content": "Input: Initial parameters \\(\\theta^{\\mathrm{student}}\\), \\(\\theta^{\\mathrm{teacher}}\\), learning rate \\(\\eta\\), stability coefficient \\(\\lambda\\), sensitivity \\(\\gamma\\) \nOutput: Stabilized parameters \\(\\theta^{\\mathrm{student}}\\), \\(\\theta^{\\mathrm{teacher}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.158, + 0.757, + 0.172 + ], + "angle": 0, + "content": "1: while training not converged do" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.173, + 0.722, + 0.188 + ], + "angle": 0, + "content": "2: Compute loss \\(\\mathcal{L}(\\theta^{\\mathrm{student}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.189, + 0.905, + 0.234 + ], + "angle": 0, + "content": "3: Update student model: \\(\\theta_{t}^{\\mathrm{student}} = \\theta_{t - 1}^{\\mathrm{student}} - \\eta \\nabla_{\\theta}\\mathcal{L}(\\theta^{\\mathrm{student}}) + \\lambda (\\theta_{t}^{\\mathrm{teacher}} - \\theta_{t}^{\\mathrm{student}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.234, + 0.857, + 0.265 + ], + "angle": 0, + "content": "4: Compute adaptive EMA decay in LyapEMA: \\(\\alpha_{t} = \\mathrm{sigmoid}(\\gamma \\| \\theta_{t}^{\\mathrm{student}} - \\theta_{t - 1}^{\\mathrm{teacher}}\\|_{2})\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.265, + 0.802, + 0.294 + ], + "angle": 0, + "content": "5: Update teacher model: \\(\\theta_{t}^{\\mathrm{teacher}} = \\alpha_{t}\\theta_{t - 1}^{\\mathrm{teacher}} + (1 - \\alpha_{t})\\theta_{t}^{\\mathrm{student}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.295, + 0.82, + 0.308 + ], + "angle": 0, + "content": "6: Ensure stability: Verify \\( V(\\theta) \\) decreases" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.309, + 0.69, + 0.323 + ], + "angle": 0, + "content": "7: if \\( V_{t} \\geq V_{t-1} \\) then" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.325, + 0.783, + 0.339 + ], + "angle": 0, + "content": "8: Reduce learning rate: \\(\\eta \\gets \\eta /2\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.34, + 0.859, + 0.354 + ], + "angle": 0, + "content": "9: Increase stability coefficient: \\(\\lambda \\gets \\lambda +\\Delta \\lambda\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.355, + 0.837, + 0.369 + ], + "angle": 0, + "content": "10: Recompute student and teacher updates" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.37, + 0.604, + 0.383 + ], + "angle": 0, + "content": "11: end if" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.385, + 0.614, + 0.398 + ], + "angle": 0, + "content": "12: end while" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.399, + 0.686, + 0.413 + ], + "angle": 0, + "content": "13: return \\(\\theta^{\\mathrm{teacher}}\\), \\(\\theta^{\\mathrm{student}}\\)" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.158, + 0.905, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.905, + 0.478 + ], + "angle": 0, + "content": "By using the LyapEMA update rules in Eq. 10 and expanding, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.506, + 0.905, + 0.538 + ], + "angle": 0, + "content": "\\[\nV _ {t + 1} - V _ {t} \\approx - \\eta \\| \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) \\| _ {2} ^ {2} - \\lambda \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.615 + ], + "angle": 0, + "content": "Since both terms in Eq. 11 are negative, \\(V\\) decreases monotonically, ensuring stability. The complete LyapEMA framework is summarized in Algorithm 1. In Section 5.1, we conduct experiments not only on Semi-ViM but also on other SSL methods to validate LyapEMA's effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.623, + 0.643 + ], + "angle": 0, + "content": "4.2. SSMixup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.74 + ], + "angle": 0, + "content": "SSMixup introduces mixup strategies in the input and the hidden state space of a bidirectional SSM to effectively leverage unlabeled data and enhance generalization in SSL. As mentioned before, Semi-ViM adopts a teacher-student framework where the student model, \\( f_{\\mathrm{student}}(\\theta^{\\mathrm{student}}) \\), is trained on the mixed samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.906, + 0.788 + ], + "angle": 0, + "content": "In SSMixup, the pseudo-label confidence scores, \\( o_i \\) and \\( o_j \\), of two unlabeled samples, \\( x_i^u \\) and \\( x_j^u \\), respectively, are used first to compute a mixup coefficient:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.798, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\psi = \\frac {o _ {i}}{o _ {i} + o _ {j}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.842, + 0.905, + 0.87 + ], + "angle": 0, + "content": "The mixup coefficient \\(\\psi\\) is used to form the mixed input and labels:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.886, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nx _ {\\text {m i x}} = \\psi x _ {i} ^ {u} + (1 - \\psi) x _ {j} ^ {u}, \\quad y _ {\\text {m i x}} = \\psi \\hat {y} _ {i} + (1 - \\psi) \\hat {y} _ {j}. \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "768" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.285 + ], + "angle": 0, + "content": "Figure 3. The SSMixup strategy. The red part applies mixup at the hidden state level. \\( L \\) denotes to the number of SSMixup encoder layers in Semi-ViM. It is a hyperparameter typically set to 24." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.355 + ], + "angle": 0, + "content": "The mixed input \\( x_{\\mathrm{mix}} \\) is then fed into the forward and backward components of the bidirectional SSM, yielding hidden states:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.365, + 0.483, + 0.385 + ], + "angle": 0, + "content": "\\[\nh ^ {\\mathrm {f w d}} = \\operatorname {S S M} _ {\\mathrm {f w d}} \\left(x _ {\\text {m i x}}\\right), \\quad h ^ {\\mathrm {b w d}} = \\operatorname {S S M} _ {\\mathrm {b w d}} \\left(x _ {\\text {m i x}}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.425 + ], + "angle": 0, + "content": "SSMixup mixes the hidden states by using the mixup coefficient, \\(\\psi\\):" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.436, + 0.483, + 0.457 + ], + "angle": 0, + "content": "\\[\nh _ {\\text {m i x}} = \\psi h ^ {\\mathrm {f w d}} + (1 - \\psi) h ^ {\\mathrm {b w d}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.528 + ], + "angle": 0, + "content": "where \\( h_{\\mathrm{mix}} \\) is computed as a residual unit and \\( C \\) is defined in Eq. 1. The forward and backward outputs of the bidirectional SSM are merged to compute the final output (see Eq. 4):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.552, + 0.483, + 0.572 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b} + \\nu \\left(C h _ {\\text {m i x}}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.643 + ], + "angle": 0, + "content": "where \\(\\nu\\) is the control coefficient. The computed \\(y_{t}\\) is used in a cross-entropy loss with \\(y_{\\mathrm{mix}}\\) to update \\(\\theta^{\\mathrm{student}}\\). Parameters \\(\\theta^{\\mathrm{teacher}}\\) are smoothly updated via LyapEMA, maintaining stable pseudo-label generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Figure 3 illustrates the functionality of SSMixup. By introducing mixup strategies at both the input (see Eq. 13) and hidden state levels (see Eq. 15), combined with confidence-driven weighting, SSMixup effectively exploits unlabeled data and significantly improves generalization capabilities in SSL settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.279, + 0.76 + ], + "angle": 0, + "content": "4.3. Training Semi-ViM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.869 + ], + "angle": 0, + "content": "Training Semi-ViM involves three main steps. The first step involves self-supervised pre-training based on Autoregressive Mamba (ARM) [33], where labeled data is treated as unlabeled data and used as part of the training samples. The second step involves supervised fine-tuning using only labeled data. The third step involves SSL. These three steps are detailed next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Self-supervised pre-training based on ARM: To improve feature extraction, we propose to leverage ARM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.346 + ], + "angle": 0, + "content": "for self-supervised pre-training of Semi-ViM. ARM is better suited for ViM pre-training than Masked Autoencoders (MAEs) and DINO-V2 (contrastive learning) [32] due to its alignment with ViM's sequential state-space modeling [33]. Unlike MAEs, which rely on masked image reconstruction and bidirectional Transformers, Semi-ViM's bidirectional state-space modeling allows to effectively leverage ARM's autoregressive token prediction, making ARM a more natural fit [17]. Compared to DINO-V2, which depends on global feature alignment and large-batch contrastive learning, ARM directly optimizes sequential token dependencies, avoiding instability from batch size sensitivity and augmentation complexity [9]. By leveraging ViM's structured recurrence, ARM enables faster training, better scalability, and improved efficiency over both methods. These statements are validated by our experiments in Section 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.909, + 0.546 + ], + "angle": 0, + "content": "Fully-supervised fine-tuning: After completing ARM-based pre-training, Semi-ViM undergoes fine-tuning using only labeled data. This process aims to leverage the representations learned during pre-training to adapt the model to specific downstream tasks. During fine-tuning, most pretrained parameters remain unchanged, with only certain layers or global parameters being updated to a limited extent in order to improve performance in SSL. Fine-tuning specifically targets layers responsible for learning bidirectional dependencies, i.e., the forward and backward SSMs are updated to better capture dependencies specific to a classification task. The class token used in image classification tasks is also fine-tuned during this phase." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.636 + ], + "angle": 0, + "content": "SSL: In the third training step, we use LyapEMA, in conjunction with SSMixup to fully exploit unlabeled data while ensuring training stability and efficiency. This last training step consists of five key components: weak and strong data augmentation, pseudo-label generation, LyapEMA, SSMixup, and minimization of the loss function." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.637, + 0.906, + 0.711 + ], + "angle": 0, + "content": "For each unlabeled sample \\( x^{u} \\), we generate a weakly augmented sample \\( x_{w}^{u} \\) and a strongly augmented sample \\( x_{s}^{u} \\). The weakly augmented sample \\( x_{w}^{u} \\) is processed by the teacher model \\( f_{\\text{teacher}} \\) to produce a set of predicted probabilities, one for each class:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.729, + 0.907, + 0.748 + ], + "angle": 0, + "content": "\\[\np _ {\\text {t e a c h e r} _ {x _ {w} ^ {u}}} = f _ {\\text {t e a c h e r}} \\left(x _ {w} ^ {u}\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.754, + 0.906, + 0.789 + ], + "angle": 0, + "content": "The pseudo-label for \\( x_w^u \\) is then determined as the class associated with the highest probability in set \\( p_{\\text{teacher}_x u} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.801, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {u} = \\arg \\max \\left(p _ {\\text {t e a c h e r} _ {x _ {w}} ^ {u}}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Note that only unlabeled samples with a class probability exceeding the threshold \\(\\tau\\) are retained for training; i.e., only those with high-score pseudo-labels are retained. SSMixup is then applied to these unlabeled samples with high-score pseudo labels in order to be used to train the student model." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "769" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.161 + ], + "angle": 0, + "content": "
Variant nameBackboneParam.Input size
Semi-ViM-TinyViM-Tiny [47]7M2242
Semi-ViM-SmallViM-Small [47]26M2242
Semi-ViM-BaseViM-Base [47]98M2242
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.171, + 0.42, + 0.185 + ], + "angle": 0, + "content": "Table 1. The different variants of Semi-ViM." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.209, + 0.483, + 0.24 + ], + "angle": 0, + "content": "The overall loss function consists of the supervised loss \\( L_{\\text{labeled}} \\) and the unsupervised loss \\( L_{\\text{unlabeled}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.254, + 0.482, + 0.27 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {l a b e l e d}} + \\mu L _ {\\text {u n l a b e l e d}}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.274, + 0.483, + 0.305 + ], + "angle": 0, + "content": "The supervised loss, \\( L_{\\mathrm{labeled}} \\), is based on the standard cross-entropy (CE) loss using labeled samples:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.316, + 0.483, + 0.357 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {l a b e l e d}} = \\frac {1}{N _ {l}} \\sum_ {i = 1} ^ {N _ {l}} \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.359, + 0.483, + 0.391 + ], + "angle": 0, + "content": "The unsupervised loss \\( L_{\\mathrm{unlabeled}} \\) uses strongly augmented samples, \\( x_{s}^{u} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.409, + 0.483, + 0.45 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {u n l a b e l e d}} = \\frac {1}{N _ {u}} \\sum_ {i = 1} ^ {N _ {u}} \\mathbb {1} \\left(o _ {i} > \\tau\\right) \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {s, i} ^ {u}\\right), \\hat {y} _ {i} ^ {u}\\right), \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.547 + ], + "angle": 0, + "content": "where the indicator function \\(\\mathbb{1}(o > \\tau)\\) selects only those unlabeled samples with high-score pseudo-labels, where \\(o\\) denotes the probability of the class associated with the pseduo-label. Finally, the teacher model is updated using LyapEMA to stabilize parameter evolution and improve the quality of pseudo-labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.224, + 0.575 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Table 1 lists the different variants of Semi-ViM we use in our experiments: Tiny, Small, and Base. These variants are trained on ImageNet-1K using 8 NVIDIA A100 (80GB) GPUs with DistributedDataParallel (DDP) PyTorch. We use AdamW as the optimizer with a weight decay of 0.05, and a learning rate that follows a cosine annealing schedule starting from \\(5 \\times 10^{-4}\\) with 10 epochs of warm-up. We use 300 epochs for training with a global batch size of 1024 (128 per GPU) and automatic mixed precision (AMP) enabled. Strong data augmentation includes AutoAugment, Mixup (\\(\\alpha = 0.2\\)), CutMix (\\(\\alpha = 1.0\\)), and label smoothing (\\(\\varepsilon = 0.1\\)). The total training time is approximately 70 hours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We conduct classification experiments on ImageNet-1K [9] and ImageNet-LT [28]. ImageNet-1K contains 1.28M training and 50K validation images across 1,000 classes. ImageNet-LT has a long-tailed class distribution. We use a confidence threshold of \\(\\tau = 0.7\\) for pseudo-label selection, with weak and strong data augmentations applied to unlabeled data. Inference is performed using a \\(224\\times 224\\) center crop on the images." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.706, + 0.107 + ], + "angle": 0, + "content": "5.1. Results and analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.204 + ], + "angle": 0, + "content": "Semi-ViM demonstrates superior performance across multiple experimental settings as tabulated in Table 2. On ImageNet-1K, Semi-ViM-Base achieves \\(85.40\\%\\) accuracy with only \\(10\\%\\) of the labeled data, outperforming SemiViT-Huge. Compared to SemiFormer, Semi-ViM consistently excels, confirming the suitability of ViM for SSL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.905, + 0.325 + ], + "angle": 0, + "content": "For the more challenging ImageNet-LT dataset, Semi-ViM-Base achieves \\(77.40\\%\\) accuracy with only \\(10\\%\\) of the labeled data, significantly surpassing Semi-ViT-Huge and SemiFormer. This demonstrates Semi-ViM's adaptability to long-tailed data distributions and its capability to work with rare category representations. With only \\(1\\%\\) of the labeled data, Semi-ViM-Base achieves \\(66.30\\%\\) accuracy, surpassing Semi-ViT-Huge and demonstrating its strength in SSL." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.325, + 0.905, + 0.415 + ], + "angle": 0, + "content": "In terms of the Top-5 accuracy on ImageNet-LT with only \\(1\\%\\) of the labeled data (last column of Table 2), Semi-ViM-Base outperforms Semi-ViT-Huge by \\(6.17\\%\\) and CNN-based methods, like FixMatch and PAWS, by over \\(15\\%\\), highlighting its superior feature representation, class diversity learning, and robustness against data imbalance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.905, + 0.536 + ], + "angle": 0, + "content": "Table 3 tabulates the Top-1 Average Precision (AP) on ImageNet-LT of several fully-supervised models trained exclusively with labeled data. Semi-ViM-Base with just \\(10\\%\\) of the labeled data achieves competitive accuracy. This suggests that our framework efficiently extracts meaningful representations even with significantly fewer labeled samples, making it highly suitable for real-world scenarios where labeled data may be scarce." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.537, + 0.905, + 0.582 + ], + "angle": 0, + "content": "Figure 4 graphically summarizes the performance of Semi-ViM and that of other SOTA SSL models and fully-supervised models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.592, + 0.663, + 0.607 + ], + "angle": 0, + "content": "5.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Several pre-training strategies, parameter updating methods, and mixup strategies are compared in Table 4, using Semi-ViM-Base. ARM pre-training consistently achieves the best performance, demonstrating its effectiveness in leveraging sequential dependencies inherent to the ViM. ARM reaches \\(81.90\\%\\) (1% of the labeled data) and \\(85.40\\%\\) (10% of the labeled data) when used in conjunction with LyapEMA+SSMixup, significantly outperforming MAEs and DINO-V2 across all settings. This confirms that ARM, designed specifically for sequential state-space models, is a more appropriate fit for ViM than contrastive learning (DINO-V2) or MAEs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In terms of parameter updating methods, LyapEMA outperforms EMA in all configurations, particularly when used in conjunction with ARM. Unlike EMA, LyapEMA stabilizes updates by leveraging Lyapunov theory, preventing pseudo-label collapse and ensuring more reliable parameter evolution. This is particularly beneficial in low-labeled data scenarios, where training instability is a key challenge." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "770" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.482 + ], + "angle": 0, + "content": "
CNN-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
FixMatch [38]ResNet-5025M-67.8371.7425M-52.3257.68-
PAWS [1]ResNet-5027M66.5469.2176.5527M46.3056.1361.8155.19
FreeMatch [42]ResNet-5026M55.4270.1174.5026M-55.5460.94-
SoftMatch [5]ResNet-5028M-70.4076.2028M-57.4161.02-
SimCLRv2(self-distilled) [6]ResNet-15228M76.6578.2380.9028M48.5559.3062.1067.28
SimCLRv2(distilled) [6]ResNet-5028M75.9077.4080.2028M47.7057.2360.8168.50
Co-training methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
Co-Training [35]CT(MLP)56M80.0381.6581.7056M58.6264.7971.2078.30
MCT [35]MCT (MLP)176M80.7582.0885.20176M63.1872.6075.3080.11
Transformer-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top-5 (1%)
DINO [14]ViT-Small22M65.2171.4572.5322M47.8055.3059.9064.25
SemiFormer [43]ViT-S+Conv42M-72.8073.2842M50.1058.0063.5067.87
DPT [44]ViT-Huge604M80.2082.5084.06604M55.9164.7668.1575.40
Semi-ViT-Base [4]ViT-Base86M71.3077.5279.9086M53.2062.9068.4071.50
Semi-ViT-Large [4]ViT-Large307M76.8180.5082.26307M56.8066.1071.2074.70
Semi-ViT-Huge [4]ViT-Huge632M80.4783.2084.27632M58.9068.4073.5076.90
REACT [25]ViT-Huge651M79.1383.2384.74632M59.3668.7173.0877.45
Semi-ViM-TinyViM-Tiny12M76.1078.3081.0012M52.1362.9068.3071.11
Semi-ViM-SmallViM-Small27M80.3081.4083.1027M63.2472.1574.6880.02
Semi-ViM-BaseViM-Base146 M81.9083.2585.40146M66.3075.5277.4083.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.493, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Table 2. Accuracy (%) of the different variants of Semi-ViM (in pink) and other SOTA SSL models over 5 folds when 1%, 5% and 10% of the labeled data is used for training. The best performance of other SOTA models is highlighted in peach." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.558, + 0.353, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.729, + 0.234, + 0.74 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.558, + 0.625, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.728, + 0.509, + 0.74 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.569, + 0.902, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.729, + 0.784, + 0.74 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.753, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Figure 4. (a) Training accuracy \\((\\%)\\) on ImageNet-1K of Semi-ViM-Base and other SOTA SSL methods using only \\(10\\%\\) of the labeled data. (b) Top-1 accuracy \\((\\%)\\) on ImageNet-LT \\((1\\%\\) of the labeled data) of Semi-ViM and other SOTA SSL algorithms across different model scales. (c) Top-1 Average Precision \\((\\%)\\) on ImageNet-LT of Semi-ViM-Base and several fully supervised models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.486, + 0.899 + ], + "angle": 0, + "content": "Regarding mixup strategies, SSMixup provides the largest performance boost, outperforming ProbMixup consistently across all pre-training methods. With ARM, LyapEAM+SSMixup boosts accuracy by \\(2 - 3\\%\\) compared to ProbMixup. This confirms that SSMixup is highly effective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.907, + 0.899 + ], + "angle": 0, + "content": "for SSL, mitigating pseudo-label noise and improving generalization in low-data settings. The combination of ARM pre-training, and LyapEMA+SSMixup achieves the highest accuracy, underscoring the importance of pre-training alignment, stable parameters updates, and structured aug-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "771" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.495, + 0.255 + ], + "angle": 0, + "content": "
ModelLabeled dataTop-1 AP
LIFT(ViT-L/14) [37]100%82.90
MAM(ViT-B/16) [18]100%82.30
LIFT(ViT-B/16) [37]100%78.30
BALLAD (ResNet-50) [29]100%76.50
Semi-ViT-Huge [4]10%73.50
Semi-ViM-Small10%74.68
Semi-ViM-Base5%75.52
Semi-ViM-Base10%77.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.325 + ], + "angle": 0, + "content": "Table 3. Top-1 AP (%) of fully-supervised models and the variants of Semi-ViM on ImageNet-LT. Semi-ViM-Base (in purple), using only \\(10\\%\\) of the labeled data, outperforms BALLAD (ResNet\\(50\\times 16\\))." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.343, + 0.446, + 0.696 + ], + "angle": 0, + "content": "
PretrainedMethod1%10%
Nonefinetune--
EMA42.1147.41
LyapEMA44.7353.26
+ProbMixup44.7849.57
+SSMixup45.2954.73
MAEfinetune63.0573.14
EMA71.3074.20
LyapEMA65.3075.65
+ProbMixup66.0975.87
+SSMixup66.0276.15
DINO-V2finetune62.0272.63
EMA63.4573.57
LyapEMA64.8475.21
+ProbMixup65.5374.04
+SSMixup65.3276.07
ARMfinetune66.5374.21
EMA72.5379.90
LyapEMA76.0278.03
+ProbMixup79.8380.52
+SSMixup81.9085.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Table 4. Top-1 accuracy (%) on ImageNet-1K of Semi-ViM-Base under different pre-training strategies and parameter updating methods using \\(1\\%\\) and \\(10\\%\\) of the labeled data for training. \" \\(^+\\) \" means the mixup strategy used in conjunction with LyapEMA. Best results are highlighted in bold font." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.441, + 0.822 + ], + "angle": 0, + "content": "mentation in semi-supervised models based on ViM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Table 5 compares different parameter updating methodologies, demonstrating the superior stability and generalization capability of LyapEMA in SSL. The EMA-Teacher suffers from training instability, particularly in Semi-ViT-Small and Semi-ViM-Small with only \\(1\\%\\) of the la" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.089, + 0.898, + 0.372 + ], + "angle": 0, + "content": "
ModelMethod1%10%
FreeMatch [42]FixMatch [38]--
EMA-Teacher [3]55.4274.50
LyapEMA56.7075.07
Semi-ViT-S [4]FixMatch [38]-×
EMA-Teacher [3]×75.80
LyapEMA61.3076.60
Semi-ViT-B [4]FixMatch [38]-×
EMA-Teacher [3]66.4178.15
LyapEMA68.8480.43
Semi-ViM-SFixMatch [38]××
EMA-Teacher [3]×74.01
LyapEMA80.3085.40
Semi-ViM-BFixMatch [38]×71.28
EMA-Teacher [3]70.5376.11
LyapEMA81.9085.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.383, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Table 5. Top-1 accuracy (%) on ImageNet-1K of FixMatch, EMA-Teacher and LyapEMA with different models using \\(1\\%\\) and \\(10\\%\\) of the labeled data for training. “×” means the training failed (accuracy near 0). “-” means no result is available. Results in light blue indicate risk of training instability and collapse." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.62 + ], + "angle": 0, + "content": "beled data, where training fails to converge. In contrast, LyapEMA effectively prevents collapse, achieving \\(61.30\\%\\) and \\(80.30\\%\\) accuracy, respectively, showing its robustness in low-labeled data scenarios. By dynamically adapting the EMA decay factor based on the student-teacher discrepancy, LyapEMA mitigates pseudo-label bias and enhances convergence. Additional ablation studies and the model's performance under our designed extreme imbalance scenarios are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced Semi-ViM, a novel semi-supervised learning framework that leverages bidirectional SSMs to mitigate learning bias caused by label imbalance in classification tasks. By integrating LyapEMA, a novel stability-aware parameter update mechanism, and SS-Mixup, a hidden-state-level mixing strategy, Semi-ViM enhances both training stability and feature learning. Our extensive experiments on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms CNN-based and Transformer-based SSL models, achieving superior accuracy with fewer labeled samples. Moreover, Semi-ViM requires far fewer parameters than ViT-based SSL models. These findings highlight the potential of state space models as a competitive alternative to Transformers in semi-supervised learning, particularly in addressing real-world imbalanced datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "772" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8443-8452, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. Advances in Neural Information Processing Systems, 35:25697-25710, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Nieves Crasto. Class imbalance in object detection: An experimental diagnosis and study of mitigation strategies. arXiv preprint arXiv:2403.07113, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[8] Charika De Alvis and Suranga Seneviratne. A survey of deep long-tail classification advancements. arXiv preprint arXiv:2404.15593, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Towards semi-supervised learning with nonrandom missing labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16121-16131, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine-grained learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11829-11837, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[12] Marawan Elbatel, Hualiang Wang, Jixiang Chen, Hao Wang, and Xiaomeng Li. Learning unlabeled clients divergence via anchor model aggregation for federated semi-supervised learning. arXiv e-prints, pages arXiv-2407, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3187-3197, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Lan-Zhe Guo, Lin-Han Jia, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning in open environments. Frontiers of Computer Science, 19(8):198345, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[18] Ahmet Iscen, Alireza Fathi, and Cordelia Schmid. Improving image recognition by retrieving from web-scale image-text data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19295-19304, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[19] Zhen Jiang, Lingyun Zhao, Yu Lu, Yongzhao Zhan, and Qirong Mao. A semi-supervised resampling method for class-imbalanced learning. Expert Systems with Applications, 221:119733, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[20] Suraj Kothawade, Pavan Kumar Reddy, Ganesh Ramakrishnan, and Rishabh Iyer. Basil: Balanced active semi-supervised learning for class imbalanced datasets. arXiv preprint arXiv:2203.05651, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Alex Kurakin, Colin Raffel, David Berthelot, Ekin Dogus Cubuk, Han Zhang, Kihyuk Sohn, and Nicholas Carlini. Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[22] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[24] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy. arXiv preprint arXiv:2405.04404, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "773" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[28] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[29] Teli Ma, Shijie Geng, Mengmeng Wang, Jing Shao, Jiasen Lu, Hongsheng Li, Peng Gao, and Yu Qiao. A simple long-tailed recognition baseline via vision-language model. arXiv preprint arXiv:2111.14745, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[31] Khanh-Binh Nguyen. Sequencematch: revisiting the design of weak-strong augmentations for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 96-106, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[32] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[33] Sucheng Ren, Xianhang Li, Haoqin Tu, Feng Wang, Fangxun Shu, Lei Zhang, Jieru Mei, Linjie Yang, Peng Wang, Heng Wang, et al. Autoregressive pretraining with mamba in vision. arXiv preprint arXiv:2406.07537, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[34] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[35] Jay C. Rothenberger and Dimitrios I. Diochnos. Meta cotraining: Two views are better than one. In 2025 Joint Mathematics Meetings (JMM 2025). AMS. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Shankar Sastry and Shankar Sastry. Lyapunov stability theory. Nonlinear systems: analysis, stability, and control, pages 182-234, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[37] Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. Long-tail learning with foundation model: Heavy fine-tuning hurts. In *Forty-first International Conference on Machine Learning*, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[38] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[41] Renzhen Wang, Xixi Jia, Quanziang Wang, Yichen Wu, and Deyu Meng. Imbalanced semi-supervised learning with bias adaptive classifier. arXiv preprint arXiv:2207.13856, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[42] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[43] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In European conference on computer vision, pages 605-620. Springer, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[44] Zebin You, Yong Zhong, Fan Bao, Jiacheng Sun, Chongxuan Li, and Jun Zhu. Diffusion models and semi-supervised learners benefit mutually with few labels. Advances in Neural Information Processing Systems, 36:43479-43495, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[45] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.499 + ], + "angle": 0, + "content": "[46] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8373-8383, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.501, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[47] L Zhu, B Liao, Q Zhang, X Wang, W Liu, and X Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arxiv 2024. arXiv preprint arXiv:2401.09417. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "774" + } + ] +] \ No newline at end of file diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_origin.pdf b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c2cd235cb2fd327783c954f4367aef8fd81b795 --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/694bc6b8-9d54-4920-9947-577e19b17917_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff4419ad37d0a53c5420f00513e742c3801d29171041b5f47420b844ea600a1d +size 709695 diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/full.md b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..974539a763287944b49e11f23b363cfc7a9a0eab --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/full.md @@ -0,0 +1,381 @@ +# Semi-ViM: Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning + +Hongyang He* +University of Warwick + +Hongyang Xie +University of Warwick + +Haochen You Columbia University + +Victor Sanchez +University of Warwick + +# Abstract + +Semi-supervised learning (SSL) is often hindered by learning biases when imbalanced datasets are used for training, which limits its effectiveness in real-world applications. In this paper, we propose Semi-ViM, a novel SSL framework based on Vision Mamba, a bidirectional state space model (SSM) that serves as a superior alternative to Transformer-based architectures for visual representation learning. Semi-ViM effectively deals with imbalanced datasets and improves model stability through two key innovations: LyapEMA, a stability-aware parameter update mechanism inspired by Lyapunov theory, and SSMixup, a novel mixup strategy applied at the hidden state level of bidirectional SSMs. Experimental results on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms state-of-the-art SSL models, achieving $85.40\%$ accuracy with only $10\%$ of the labeled data, surpassing Transformer-based methods such as Semi-ViT. + +# 1. Introduction + +Semi-supervised learning (SSL) aims to mitigate the scarcity of large-scale labeled training data while reducing annotation costs [16]. In the context of classification, however, a challenge remains: Most state-of-the-art (SOTA) SSL models are trained under the assumption that access to a dataset with a balanced number of labeled samples per class is possible, which is often overly idealized and deviates significantly from real-world conditions [20]. In practice, label imbalance induces learning biases in SSL, ultimately compromising model performance [41]. + +Due to its outstanding performance, many studies have integrated the vision Transformer (ViT) into SSL methods [4, 43, 46]; e.g., Semi-ViT, adopts a pre-training and fine-tuning approach to achieve SOTA results [4]. How- + +![](images/31f478e572973ac2bff382619233040efc743ffdbc670b2ce3ef03e085494769.jpg) +Figure 1. Semi-ViT-Small trained with only $1\%$ of the labeled data of ImageNet-1K. Left: Training fails when FixMatch is used. The loss decreases but test accuracy remains near zero. Right: Training is unstable and fails when EMA-Teacher is used: The loss and accuracy oscillate heavily. + +ever, Semi-ViT has a weaker inductive bias, requiring large-scale data for effective training [39]. Moroever, as shown by our experiments, Semi-ViT does not perform well on long-tailed imbalanced datasets, like ImageNet-LT, confirming its limitations in handling label imbalance. Although the performance of Semi-ViT can be improved by increasing the model size, e.g., Semi-ViT-Huge, this comes at the expense of a larger number of trainable parameters, around 632M [4]. Vision Mamba (ViM), which incorporates a bidirectional state space model (SSM), has modeling capabilities comparable to those of ViT, while achieving subquadratic-time computational and linear memory complexities [47]. + +Based on the previous discussions, this paper integrates ViM into an SSL framework for classification and proposes a ViM-based SSL model, hereinafter called Semi-ViM. Semi-ViM incorporates two key innovations aimed at improving model stability during training and performance: LyapEMA and SSMixup. LyapEMA is a novel parameter update mechanism that leverages the Lyapunov theory to dynamically adjust parameter updates, ensuring a more stable training process. LyapEMA addresses the training failures encountered when using only a small amount of labeled + +data to train Semi-ViT-Small, as shown in Fig 1. LyapEMA also helps prevent training collapse, commonly encountered when using the exponential moving average (EMA) and FixMatch parameter update mechanisms [4]. SSMixup, on the other hand, is a novel mixup strategy that operates at the hidden state space of SSMs and goes beyond the traditional input-level mixup strategy. Specifically, SSMixup incorporates mixed hidden states as residual units in the training process, improving the model's capability to generalize under sparse label conditions. Our main contributions can be summarized as follows: + +- We propose Semi-ViM, an SSL framework that outperforms other SSL methods based on CNNs and Transformers. The number of parameters of Semi-ViM is only around one-fourth of that of Semi-ViT. +- We propose LyapEMA, a novel parameter update mechanism that successfully addresses the training instability issues commonly found in Semi-ViT, thus improving model convergence and generalization capabilities. +- We propose SSMixup, a hidden state space-level mixup strategy that effectively enhances the learning capability of bidirectional SSMs. +- Based on evaluations on the long-tailed ImageNet-LT dataset, we demonstrate Semi-ViM's strong capabilities in mitigating learning biases caused by label imbalance. + +# 2. Related Work + +Imbalanced datasets are those in which the different classes are distributed unequally, often resulting in long-tailed distributions [7]. SSL aims to leverage a small amount of labeled data and a large amount of unlabeled data to improve performance, which is useful to deal with imbalanced datasets. SSL has achieved significant advancements in several computer vision tasks [10, 11, 21, 38, 45]. + +Traditional SSL methods, e.g., entropy minimization and consistency regularization [13, 15], exploit unlabeled data to reduce dependence on manual annotation. For example, consistency regularization leverages unlabeled data under the assumption that a model's predictions remain consistent under different input perturbations. Mainstream methods include II-Model [22], Mean Teacher [40], Virtual Adversarial Training (VAT) [30], and FixMatch [38], which is an important benchmark in SSL [38]. Pseudo-Labeling [23] is a common strategy used in SSL. It consists of using the model's predictions on unlabeled data as pseudo-labels for training. Noisy Student, FixMatch [38], and FlexMatch [45] adopt this strategy and use confidence thresholds to filter out pseudo-labels. + +Several SOTA SSL methods are based on FixMatch or any of its variants, e.g., FlexMatch and CoMatch [24], which integrate contrastive learning or self-supervised methods to enhance performance. However, these methods typically assume that the class distribution is balanced, + +which is rarely the case in real-world scenarios [12]. When dealing with imbalanced datasets [8], these SOTA SSL methods face two primary challenges. First, pseudo-label biases, i.e., models tend to predict more frequently the majority classes when generating pseudo-labels, exacerbating the imbalance issue and limiting learning from the minority classes [19]. Second, low-quality pseudo-labels, i.e., models tend to assign low confidence scores to those pseudolabels generated for the minority class samples, thus reducing their impact on training and potentially leading to training collapse [4]. For example, FixMatch relies on a fixed confidence threshold for selecting pseudo-labels, but in an imbalanced dataset setting, pseudo-labels of the minority classes often fail to reach this threshold, leading to insufficient learning from such classes [4, 43]. + +# 3. Preliminaries + +State Space Models and ViM: SSMs have a long-standing history in control systems and signal processing, offering a structured approach to sequential data modeling [47]. Deep learning adopts SSMs to more effectively handle long-range dependencies compared to traditional recurrence-based models. Unlike Transformers, which rely on global self-attention, SSMs leverage structured state transitions for sequential data propagation, enabling subquadratic computational complexity and linear memory scaling [47]. In computer vision, SSMs have been integrated into deep architectures to enhance representation learning by modeling both spatial and temporal dependencies [27, 47]. The SSM is formulated as a discrete-time state-space system [47]: + +$$ +h _ {t} = A h _ {t - 1} + B x _ {t}, \quad y _ {t} = C h _ {t}, \tag {1} +$$ + +where $A \in \mathbb{R}^{N \times N}$ is the state transition matrix; $B \in \mathbb{R}^{N \times 1}$ and $C \in \mathbb{R}^{1 \times N}$ are input and output projection matrices, respectively; $x_{t}$ represents the input sequence at time step $t$ ; and $y_{t}$ is the corresponding output representation. The parameters of an SSM are usually optimized in a data-dependent manner to ensure computational efficiency and parallelization. + +The bidirectional extension of SSMs further improves feature learning by capturing both forward and backward information, making it a competitive alternative to self-attention mechanisms used by Transformers. [47]. The bidirectional SSM operates in the forward $(f)$ and backward $(b)$ direction: + +$$ +h _ {t} ^ {f} = A h _ {t - 1} ^ {f} + B x _ {t}, \quad y _ {t} ^ {f} = C h _ {t} ^ {f}; \tag {2} +$$ + +$$ +h _ {t} ^ {b} = A h _ {t + 1} ^ {b} + B x _ {t}, \quad y _ {t} ^ {b} = C h _ {t} ^ {b}. \tag {3} +$$ + +The final output at time step $t$ is obtained by merging the forward and backward outputs through a gating mechanism: + +$$ +y _ {t} = \operatorname {S i L U} (z) \odot y _ {t} ^ {f} + \operatorname {S i L U} (z) \odot y _ {t} ^ {b}, \tag {4} +$$ + +![](images/3653d4902f23b10c9df92783cd4756241dcf665036e59c57e7732dff742acb1e.jpg) +(a) FixMatch + +![](images/f810dfa34687df9fa1847eeb2d61ff8e17d4b2205639902f1751d62b55cef712.jpg) +(b) EMA-Teacher +Figure 2. Differences between (a) FixMatch, (b) EMA-Teacher and (c) Semi-ViM (ours). In the figure, $X_{s}$ and $X_{w}$ are, respectively, strongly and weakly augmented inputs; $p^{s}$ and $p_{student}^{s}$ are the predictions of the student model on $X_{s}$ ; $p^{w}$ and $p_{teacher}^{w}$ are the predictions of the teacher model on $X_{w}$ ; $\tau$ is the threshold used to select pseudo-labels; and $\theta, \theta^{student}, \theta^{teacher}$ are, respectively, the learnable parameters for the entire model, the student model, and the teacher model. + +![](images/a3714e3218e742eb564341217c5bc94abe38467f6b3ea0b918b6481af82df601.jpg) +(c) Semi-ViM(Ours) + +where $\mathrm{SiLU}(\cdot)$ is the Sigmoid-weighted Linear Unit activation function, and $z$ is a gating vector that adaptively balances bidirectional information. + +ViM adopts a bidirectional SSM for visual representation learning and integrates positional embeddings for spatial awareness. In ViM, the input image is first partitioned into patches and projected onto a high-dimensional feature space before being processed. With a linear computational complexity, ViM provides data-dependent global visual modeling, significantly improving efficiency in high-resolution tasks compared to Transformer-based architectures [26, 47]. + +Mixup strategies: These strategies interpolate input samples and labels, and hence can be used as data augmentation techniques to improve generalization capabilities [2]. In SSL, they can enhance pseudo-labeling and provide consistency regularization [34]. Recent variants extend mixup strategies to the feature space and pseudo-label frameworks, e.g., the probabilistic pseudo mixup strategy refines pseudolabel quality by weighting interpolated samples based on their confidence scores, which are derived from the model's softmax outputs or uncertainty estimates. These variants can improve SSL robustness, particularly for ViTs, which benefit from stronger regularization [4]. + +Weak and strong data augmentation: Weak data augmentation involves simple transformation, e.g., random cropping, flipping, or slight color jitter, which introduce small variations to the original data to help a model learn from basic invariances. Strong data augmentation, on the other hand, applies more aggressive transformations, e.g., large crops, strong color distortions, or more significant perturbations, which force a model to learn more robust representations by simulating more challenging scenarios [31]. + +# 4. Proposed Semi-ViM framework + +Semi-ViM adopts a teacher-student framework and uses a labeled dataset, $\mathcal{X}_l = \{(x_i^l,y_i^l)\}_{i = 1}^{N_l}$ , and an unlabeled dataset $\mathcal{X}_u = \{x_i^u\}_{i = 1}^{N_u}$ for training. Specifically, the teacher model, $f_{\mathrm{teacher}}(\theta^{\mathrm{teacher}})$ , generates pseudo-labels, $\{\hat{y}_i\}$ , for unlabeled samples, $\{x_i^u\}$ , and outputs confidence scores, $\{o_i\}$ , for the pseudo-labels. The student model, $f_{\mathrm{student}}(\theta^{\mathrm{student}})$ , is then trained on samples mixed according to their confidence scores. The difference between the training paradigm of Semi-ViM and those of FixMatch and EMA-Teacher is shown in Fig 2. + +In this section, we first introduce LyapEMA, which is inspired by Lyapunov stability theory [36] and stabilizes parameter updates, thus preventing model collapse during the training of Semi-ViM. Next, we introduce SSMixup, a new mixup strategy designed for SSM-based models. Unlike traditional mixup strategies that operate solely on the input level, SSMixup mixes the hidden states of bidirectional SSMs. Finally, we detail the training and loss function of Semi-ViM. + +# 4.1. LyapEMA + +LyapEMA is a novel optimization framework that integrates Lyapunov stability theory with EMA. LyapEMA is designed for teacher-student frameworks under an SSL setting. It has two main objectives. The first is to ensure stable parameter updates, because training SSL frameworks, like Semi-ViT, with a limited amount of labeled data can lead to training collapse. The second objective is reducing the learning bias associated with pseudo-labels while preserving the advantages of EMA in refining the teacher model. + +Let $\theta^{\mathrm{student}}$ and $\theta^{\mathrm{teacher}}$ denote the student and teacher model parameters, respectively. EMA updates the teacher + +model using gradient descent as follows: + +$$ +\theta_ {t} ^ {\text {t e a c h e r}} = \alpha_ {t} \theta_ {t - 1} ^ {\text {t e a c h e r}} + (1 - \alpha_ {t}) \theta_ {t} ^ {\text {s t u d e n t}}, \tag {5} +$$ + +where $\alpha \in (0,1)$ is the EMA decay factor. However, EMA cannot guarantee stability, potentially leading to divergence due to accumulated bias associated with learning from pseudo-labels. To address this issue, LyapEMA introduces Lyapunov stability constraints to guide the updates. Let us define a Lyapunov candidate function, $V(\theta^{\mathrm{student}},\theta^{\mathrm{teacher}})$ , to measure the discrepancy between the student and teacher models: + +$$ +V \left(\theta^ {\text {s t u d e n t}}, \theta^ {\text {t e a c h e r}}\right) = \frac {1}{2} \| \theta^ {\text {s t u d e n t}} - \theta^ {\text {t e a c h e r}} \| _ {2} ^ {2}. \tag {6} +$$ + +Note that the function in Eq. 6 acts as a stability indicator, ensuring that the student model does not drift too far from the teacher model. The Lyapunov stability criterion requires that the derivative of $V(\cdot, \cdot)$ w.r.t. time, i.e., $\dot{V}$ , satisfies: + +$$ +\dot {V} = \frac {d V}{d t} = \left(\theta^ {\text {s t u d e n t}} - \theta^ {\text {t e a c h e r}}\right) ^ {\mathrm {T}} \left(\dot {\theta} ^ {\text {s t u d e n t}} - \dot {\theta} ^ {\text {t e a c h e r}}\right) < 0, \tag {7} +$$ + +where $\dot{\theta}^{\mathrm{student}}$ and $\dot{\theta}^{\mathrm{teacher}}$ also derivatives w.r.t. time. The criterion in Eq. 7 guarantees that the discrepancy between the student and teacher models decreases over time, ensuring convergence. + +To enforce stability while leveraging EMA for smooth teacher model updates, LyapEMA updates the student model parameters by using aLyapunov regularization term: + +$$ +\theta_ {t} ^ {\text {s t u d e n t}} = \theta_ {t - 1} ^ {\text {s t u d e n t}} - \eta \nabla_ {\theta} \mathcal {L} \left(\theta^ {\text {s t u d e n t}}\right) + \lambda \left(\theta_ {t} ^ {\text {t e a c h e r}} - \theta_ {t} ^ {\text {s t u d e n t}}\right), \tag {8} +$$ + +where $\mathcal{L}(\theta^{\mathrm{student}})$ is a semi-supervised loss function, $\eta$ is the learning rate, and $\lambda$ is a stability coefficient that ensures the student model remains close to the teacher model. + +LyapEMA updates the teacher model parameters by using EMA as denoted by Eq. 5. However, the EMA decay factor $\alpha_{t}$ is dynamically adjusted based on the Lyapunov function: + +$$ +\alpha_ {t} = \operatorname {s i g m o i d} \left(\gamma \| \theta_ {t} ^ {\text {s t u d e n t}} - \theta_ {t - 1} ^ {\text {t e a c h e r}} \| _ {2}\right), \tag {9} +$$ + +where $\gamma$ is a sensitivity hyperparameter controlling how $\alpha_{t}$ adapts based on the student-teacher discrepancy, and the sigmoid function ensures $\alpha_{t}$ remains within (0, 1), preventing instability in the updates. + +LyapEMA ensures that updates lead to stable training by using the Lyapunov decrease condition: + +$$ +V _ {t + 1} - V _ {t} = \frac {1}{2} \| \theta_ {t + 1} ^ {\text {s t u d e n t}} - \theta_ {t + 1} ^ {\text {t e a c h e r}} \| _ {2} ^ {2} - \frac {1}{2} \| \theta_ {t} ^ {\text {s t u d e n t}} - \theta_ {t} ^ {\text {t e a c h e r}} \| _ {2} ^ {2}. \tag {10} +$$ + +# Algorithm 1 LyapEMA + +Input: Initial parameters $\theta^{\mathrm{student}}$ , $\theta^{\mathrm{teacher}}$ , learning rate $\eta$ , stability coefficient $\lambda$ , sensitivity $\gamma$ +Output: Stabilized parameters $\theta^{\mathrm{student}}$ , $\theta^{\mathrm{teacher}}$ + +1: while training not converged do +2: Compute loss $\mathcal{L}(\theta^{\mathrm{student}})$ +3: Update student model: $\theta_{t}^{\mathrm{student}} = \theta_{t - 1}^{\mathrm{student}} - \eta \nabla_{\theta}\mathcal{L}(\theta^{\mathrm{student}}) + \lambda (\theta_{t}^{\mathrm{teacher}} - \theta_{t}^{\mathrm{student}})$ +4: Compute adaptive EMA decay in LyapEMA: $\alpha_{t} = \mathrm{sigmoid}(\gamma \| \theta_{t}^{\mathrm{student}} - \theta_{t - 1}^{\mathrm{teacher}}\|_{2})$ +5: Update teacher model: $\theta_{t}^{\mathrm{teacher}} = \alpha_{t}\theta_{t - 1}^{\mathrm{teacher}} + (1 - \alpha_{t})\theta_{t}^{\mathrm{student}}$ +6: Ensure stability: Verify $V(\theta)$ decreases +7: if $V_{t} \geq V_{t-1}$ then +8: Reduce learning rate: $\eta \gets \eta /2$ +9: Increase stability coefficient: $\lambda \gets \lambda +\Delta \lambda$ +10: Recompute student and teacher updates +11: end if +12: end while +13: return $\theta^{\mathrm{teacher}}$ , $\theta^{\mathrm{student}}$ + +By using the LyapEMA update rules in Eq. 10 and expanding, we obtain: + +$$ +V _ {t + 1} - V _ {t} \approx - \eta \| \nabla_ {\theta} \mathcal {L} \left(\theta^ {\text {s t u d e n t}}\right) \| _ {2} ^ {2} - \lambda \| \theta_ {t} ^ {\text {s t u d e n t}} - \theta_ {t} ^ {\text {t e a c h e r}} \| _ {2} ^ {2}. \tag {11} +$$ + +Since both terms in Eq. 11 are negative, $V$ decreases monotonically, ensuring stability. The complete LyapEMA framework is summarized in Algorithm 1. In Section 5.1, we conduct experiments not only on Semi-ViM but also on other SSL methods to validate LyapEMA's effectiveness. + +# 4.2. SSMixup + +SSMixup introduces mixup strategies in the input and the hidden state space of a bidirectional SSM to effectively leverage unlabeled data and enhance generalization in SSL. As mentioned before, Semi-ViM adopts a teacher-student framework where the student model, $f_{\mathrm{student}}(\theta^{\mathrm{student}})$ , is trained on the mixed samples. + +In SSMixup, the pseudo-label confidence scores, $o_i$ and $o_j$ , of two unlabeled samples, $x_i^u$ and $x_j^u$ , respectively, are used first to compute a mixup coefficient: + +$$ +\psi = \frac {o _ {i}}{o _ {i} + o _ {j}}. \tag {12} +$$ + +The mixup coefficient $\psi$ is used to form the mixed input and labels: + +$$ +x _ {\text {m i x}} = \psi x _ {i} ^ {u} + (1 - \psi) x _ {j} ^ {u}, \quad y _ {\text {m i x}} = \psi \hat {y} _ {i} + (1 - \psi) \hat {y} _ {j}. \tag {13} +$$ + +![](images/03c75fa5720095ae52eb1d7609617dd683315634110e7abbd6ace3ba5eebf439.jpg) +Figure 3. The SSMixup strategy. The red part applies mixup at the hidden state level. $L$ denotes to the number of SSMixup encoder layers in Semi-ViM. It is a hyperparameter typically set to 24. + +The mixed input $x_{\mathrm{mix}}$ is then fed into the forward and backward components of the bidirectional SSM, yielding hidden states: + +$$ +h ^ {\mathrm {f w d}} = \operatorname {S S M} _ {\mathrm {f w d}} \left(x _ {\text {m i x}}\right), \quad h ^ {\mathrm {b w d}} = \operatorname {S S M} _ {\mathrm {b w d}} \left(x _ {\text {m i x}}\right). \tag {14} +$$ + +SSMixup mixes the hidden states by using the mixup coefficient, $\psi$ : + +$$ +h _ {\text {m i x}} = \psi h ^ {\mathrm {f w d}} + (1 - \psi) h ^ {\mathrm {b w d}}, \tag {15} +$$ + +where $h_{\mathrm{mix}}$ is computed as a residual unit and $C$ is defined in Eq. 1. The forward and backward outputs of the bidirectional SSM are merged to compute the final output (see Eq. 4): + +$$ +y _ {t} = \operatorname {S i L U} (z) \odot y _ {t} ^ {f} + \operatorname {S i L U} (z) \odot y _ {t} ^ {b} + \nu \left(C h _ {\text {m i x}}\right). \tag {16} +$$ + +where $\nu$ is the control coefficient. The computed $y_{t}$ is used in a cross-entropy loss with $y_{\mathrm{mix}}$ to update $\theta^{\mathrm{student}}$ . Parameters $\theta^{\mathrm{teacher}}$ are smoothly updated via LyapEMA, maintaining stable pseudo-label generation. + +Figure 3 illustrates the functionality of SSMixup. By introducing mixup strategies at both the input (see Eq. 13) and hidden state levels (see Eq. 15), combined with confidence-driven weighting, SSMixup effectively exploits unlabeled data and significantly improves generalization capabilities in SSL settings. + +# 4.3. Training Semi-ViM + +Training Semi-ViM involves three main steps. The first step involves self-supervised pre-training based on Autoregressive Mamba (ARM) [33], where labeled data is treated as unlabeled data and used as part of the training samples. The second step involves supervised fine-tuning using only labeled data. The third step involves SSL. These three steps are detailed next. + +Self-supervised pre-training based on ARM: To improve feature extraction, we propose to leverage ARM + +for self-supervised pre-training of Semi-ViM. ARM is better suited for ViM pre-training than Masked Autoencoders (MAEs) and DINO-V2 (contrastive learning) [32] due to its alignment with ViM's sequential state-space modeling [33]. Unlike MAEs, which rely on masked image reconstruction and bidirectional Transformers, Semi-ViM's bidirectional state-space modeling allows to effectively leverage ARM's autoregressive token prediction, making ARM a more natural fit [17]. Compared to DINO-V2, which depends on global feature alignment and large-batch contrastive learning, ARM directly optimizes sequential token dependencies, avoiding instability from batch size sensitivity and augmentation complexity [9]. By leveraging ViM's structured recurrence, ARM enables faster training, better scalability, and improved efficiency over both methods. These statements are validated by our experiments in Section 5. + +Fully-supervised fine-tuning: After completing ARM-based pre-training, Semi-ViM undergoes fine-tuning using only labeled data. This process aims to leverage the representations learned during pre-training to adapt the model to specific downstream tasks. During fine-tuning, most pretrained parameters remain unchanged, with only certain layers or global parameters being updated to a limited extent in order to improve performance in SSL. Fine-tuning specifically targets layers responsible for learning bidirectional dependencies, i.e., the forward and backward SSMs are updated to better capture dependencies specific to a classification task. The class token used in image classification tasks is also fine-tuned during this phase. + +SSL: In the third training step, we use LyapEMA, in conjunction with SSMixup to fully exploit unlabeled data while ensuring training stability and efficiency. This last training step consists of five key components: weak and strong data augmentation, pseudo-label generation, LyapEMA, SSMixup, and minimization of the loss function. + +For each unlabeled sample $x^{u}$ , we generate a weakly augmented sample $x_{w}^{u}$ and a strongly augmented sample $x_{s}^{u}$ . The weakly augmented sample $x_{w}^{u}$ is processed by the teacher model $f_{\text{teacher}}$ to produce a set of predicted probabilities, one for each class: + +$$ +p _ {\text {t e a c h e r} _ {x _ {w} ^ {u}}} = f _ {\text {t e a c h e r}} \left(x _ {w} ^ {u}\right). \tag {17} +$$ + +The pseudo-label for $x_w^u$ is then determined as the class associated with the highest probability in set $p_{\text{teacher}_x u}$ : + +$$ +\hat {y} ^ {u} = \arg \max \left(p _ {\text {t e a c h e r} _ {x _ {w}} ^ {u}}\right). \tag {18} +$$ + +Note that only unlabeled samples with a class probability exceeding the threshold $\tau$ are retained for training; i.e., only those with high-score pseudo-labels are retained. SSMixup is then applied to these unlabeled samples with high-score pseudo labels in order to be used to train the student model. + +
Variant nameBackboneParam.Input size
Semi-ViM-TinyViM-Tiny [47]7M2242
Semi-ViM-SmallViM-Small [47]26M2242
Semi-ViM-BaseViM-Base [47]98M2242
+ +Table 1. The different variants of Semi-ViM. + +The overall loss function consists of the supervised loss $L_{\text{labeled}}$ and the unsupervised loss $L_{\text{unlabeled}}$ : + +$$ +L = L _ {\text {l a b e l e d}} + \mu L _ {\text {u n l a b e l e d}}. \tag {19} +$$ + +The supervised loss, $L_{\mathrm{labeled}}$ , is based on the standard cross-entropy (CE) loss using labeled samples: + +$$ +L _ {\text {l a b e l e d}} = \frac {1}{N _ {l}} \sum_ {i = 1} ^ {N _ {l}} \mathrm {C E} \left(f _ {\text {s t u d e n t}} \left(x _ {i} ^ {l}\right), y _ {i} ^ {l}\right). \tag {20} +$$ + +The unsupervised loss $L_{\mathrm{unlabeled}}$ uses strongly augmented samples, $x_{s}^{u}$ as follows: + +$$ +L _ {\text {u n l a b e l e d}} = \frac {1}{N _ {u}} \sum_ {i = 1} ^ {N _ {u}} \mathbb {1} \left(o _ {i} > \tau\right) \mathrm {C E} \left(f _ {\text {s t u d e n t}} \left(x _ {s, i} ^ {u}\right), \hat {y} _ {i} ^ {u}\right), \tag {21} +$$ + +where the indicator function $\mathbb{1}(o > \tau)$ selects only those unlabeled samples with high-score pseudo-labels, where $o$ denotes the probability of the class associated with the pseduo-label. Finally, the teacher model is updated using LyapEMA to stabilize parameter evolution and improve the quality of pseudo-labels. + +# 5. Experiments + +Table 1 lists the different variants of Semi-ViM we use in our experiments: Tiny, Small, and Base. These variants are trained on ImageNet-1K using 8 NVIDIA A100 (80GB) GPUs with DistributedDataParallel (DDP) PyTorch. We use AdamW as the optimizer with a weight decay of 0.05, and a learning rate that follows a cosine annealing schedule starting from $5 \times 10^{-4}$ with 10 epochs of warm-up. We use 300 epochs for training with a global batch size of 1024 (128 per GPU) and automatic mixed precision (AMP) enabled. Strong data augmentation includes AutoAugment, Mixup ( $\alpha = 0.2$ ), CutMix ( $\alpha = 1.0$ ), and label smoothing ( $\varepsilon = 0.1$ ). The total training time is approximately 70 hours. + +We conduct classification experiments on ImageNet-1K [9] and ImageNet-LT [28]. ImageNet-1K contains 1.28M training and 50K validation images across 1,000 classes. ImageNet-LT has a long-tailed class distribution. We use a confidence threshold of $\tau = 0.7$ for pseudo-label selection, with weak and strong data augmentations applied to unlabeled data. Inference is performed using a $224\times 224$ center crop on the images. + +# 5.1. Results and analysis + +Semi-ViM demonstrates superior performance across multiple experimental settings as tabulated in Table 2. On ImageNet-1K, Semi-ViM-Base achieves $85.40\%$ accuracy with only $10\%$ of the labeled data, outperforming SemiViT-Huge. Compared to SemiFormer, Semi-ViM consistently excels, confirming the suitability of ViM for SSL. + +For the more challenging ImageNet-LT dataset, Semi-ViM-Base achieves $77.40\%$ accuracy with only $10\%$ of the labeled data, significantly surpassing Semi-ViT-Huge and SemiFormer. This demonstrates Semi-ViM's adaptability to long-tailed data distributions and its capability to work with rare category representations. With only $1\%$ of the labeled data, Semi-ViM-Base achieves $66.30\%$ accuracy, surpassing Semi-ViT-Huge and demonstrating its strength in SSL. + +In terms of the Top-5 accuracy on ImageNet-LT with only $1\%$ of the labeled data (last column of Table 2), Semi-ViM-Base outperforms Semi-ViT-Huge by $6.17\%$ and CNN-based methods, like FixMatch and PAWS, by over $15\%$ , highlighting its superior feature representation, class diversity learning, and robustness against data imbalance. + +Table 3 tabulates the Top-1 Average Precision (AP) on ImageNet-LT of several fully-supervised models trained exclusively with labeled data. Semi-ViM-Base with just $10\%$ of the labeled data achieves competitive accuracy. This suggests that our framework efficiently extracts meaningful representations even with significantly fewer labeled samples, making it highly suitable for real-world scenarios where labeled data may be scarce. + +Figure 4 graphically summarizes the performance of Semi-ViM and that of other SOTA SSL models and fully-supervised models. + +# 5.2. Ablation study + +Several pre-training strategies, parameter updating methods, and mixup strategies are compared in Table 4, using Semi-ViM-Base. ARM pre-training consistently achieves the best performance, demonstrating its effectiveness in leveraging sequential dependencies inherent to the ViM. ARM reaches $81.90\%$ (1% of the labeled data) and $85.40\%$ (10% of the labeled data) when used in conjunction with LyapEMA+SSMixup, significantly outperforming MAEs and DINO-V2 across all settings. This confirms that ARM, designed specifically for sequential state-space models, is a more appropriate fit for ViM than contrastive learning (DINO-V2) or MAEs. + +In terms of parameter updating methods, LyapEMA outperforms EMA in all configurations, particularly when used in conjunction with ARM. Unlike EMA, LyapEMA stabilizes updates by leveraging Lyapunov theory, preventing pseudo-label collapse and ensuring more reliable parameter evolution. This is particularly beneficial in low-labeled data scenarios, where training instability is a key challenge. + +
CNN-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
FixMatch [38]ResNet-5025M-67.8371.7425M-52.3257.68-
PAWS [1]ResNet-5027M66.5469.2176.5527M46.3056.1361.8155.19
FreeMatch [42]ResNet-5026M55.4270.1174.5026M-55.5460.94-
SoftMatch [5]ResNet-5028M-70.4076.2028M-57.4161.02-
SimCLRv2(self-distilled) [6]ResNet-15228M76.6578.2380.9028M48.5559.3062.1067.28
SimCLRv2(distilled) [6]ResNet-5028M75.9077.4080.2028M47.7057.2360.8168.50
Co-training methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
Co-Training [35]CT(MLP)56M80.0381.6581.7056M58.6264.7971.2078.30
MCT [35]MCT (MLP)176M80.7582.0885.20176M63.1872.6075.3080.11
Transformer-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top-5 (1%)
DINO [14]ViT-Small22M65.2171.4572.5322M47.8055.3059.9064.25
SemiFormer [43]ViT-S+Conv42M-72.8073.2842M50.1058.0063.5067.87
DPT [44]ViT-Huge604M80.2082.5084.06604M55.9164.7668.1575.40
Semi-ViT-Base [4]ViT-Base86M71.3077.5279.9086M53.2062.9068.4071.50
Semi-ViT-Large [4]ViT-Large307M76.8180.5082.26307M56.8066.1071.2074.70
Semi-ViT-Huge [4]ViT-Huge632M80.4783.2084.27632M58.9068.4073.5076.90
REACT [25]ViT-Huge651M79.1383.2384.74632M59.3668.7173.0877.45
Semi-ViM-TinyViM-Tiny12M76.1078.3081.0012M52.1362.9068.3071.11
Semi-ViM-SmallViM-Small27M80.3081.4083.1027M63.2472.1574.6880.02
Semi-ViM-BaseViM-Base146 M81.9083.2585.40146M66.3075.5277.4083.07
+ +Table 2. Accuracy (%) of the different variants of Semi-ViM (in pink) and other SOTA SSL models over 5 folds when 1%, 5% and 10% of the labeled data is used for training. The best performance of other SOTA models is highlighted in peach. + +![](images/74f9f7d7480c003c75db1f08bda0e673bc73c1086c6c34d5682cc6bad7ae9c7a.jpg) +(a) + +![](images/13d152dcee76c156b3a779e93ab1e54c4b61496da5c35430cc8c567eee0211d9.jpg) +(b) + +![](images/040ac5484445f07bd10ca732b9919d3e188c386cb24f38030b3c4a28a93b3445.jpg) +(c) +Figure 4. (a) Training accuracy $(\%)$ on ImageNet-1K of Semi-ViM-Base and other SOTA SSL methods using only $10\%$ of the labeled data. (b) Top-1 accuracy $(\%)$ on ImageNet-LT $(1\%$ of the labeled data) of Semi-ViM and other SOTA SSL algorithms across different model scales. (c) Top-1 Average Precision $(\%)$ on ImageNet-LT of Semi-ViM-Base and several fully supervised models. + +Regarding mixup strategies, SSMixup provides the largest performance boost, outperforming ProbMixup consistently across all pre-training methods. With ARM, LyapEAM+SSMixup boosts accuracy by $2 - 3\%$ compared to ProbMixup. This confirms that SSMixup is highly effective + +for SSL, mitigating pseudo-label noise and improving generalization in low-data settings. The combination of ARM pre-training, and LyapEMA+SSMixup achieves the highest accuracy, underscoring the importance of pre-training alignment, stable parameters updates, and structured aug- + +
ModelLabeled dataTop-1 AP
LIFT(ViT-L/14) [37]100%82.90
MAM(ViT-B/16) [18]100%82.30
LIFT(ViT-B/16) [37]100%78.30
BALLAD (ResNet-50) [29]100%76.50
Semi-ViT-Huge [4]10%73.50
Semi-ViM-Small10%74.68
Semi-ViM-Base5%75.52
Semi-ViM-Base10%77.40
+ +Table 3. Top-1 AP (%) of fully-supervised models and the variants of Semi-ViM on ImageNet-LT. Semi-ViM-Base (in purple), using only $10\%$ of the labeled data, outperforms BALLAD (ResNet $50\times 16$ ). + +
PretrainedMethod1%10%
Nonefinetune--
EMA42.1147.41
LyapEMA44.7353.26
+ProbMixup44.7849.57
+SSMixup45.2954.73
MAEfinetune63.0573.14
EMA71.3074.20
LyapEMA65.3075.65
+ProbMixup66.0975.87
+SSMixup66.0276.15
DINO-V2finetune62.0272.63
EMA63.4573.57
LyapEMA64.8475.21
+ProbMixup65.5374.04
+SSMixup65.3276.07
ARMfinetune66.5374.21
EMA72.5379.90
LyapEMA76.0278.03
+ProbMixup79.8380.52
+SSMixup81.9085.40
+ +mentation in semi-supervised models based on ViM. + +Table 5 compares different parameter updating methodologies, demonstrating the superior stability and generalization capability of LyapEMA in SSL. The EMA-Teacher suffers from training instability, particularly in Semi-ViT-Small and Semi-ViM-Small with only $1\%$ of the la + +Table 4. Top-1 accuracy (%) on ImageNet-1K of Semi-ViM-Base under different pre-training strategies and parameter updating methods using $1\%$ and $10\%$ of the labeled data for training. " $^+$ " means the mixup strategy used in conjunction with LyapEMA. Best results are highlighted in bold font. + +
ModelMethod1%10%
FreeMatch [42]FixMatch [38]--
EMA-Teacher [3]55.4274.50
LyapEMA56.7075.07
Semi-ViT-S [4]FixMatch [38]-×
EMA-Teacher [3]×75.80
LyapEMA61.3076.60
Semi-ViT-B [4]FixMatch [38]-×
EMA-Teacher [3]66.4178.15
LyapEMA68.8480.43
Semi-ViM-SFixMatch [38]××
EMA-Teacher [3]×74.01
LyapEMA80.3085.40
Semi-ViM-BFixMatch [38]×71.28
EMA-Teacher [3]70.5376.11
LyapEMA81.9085.40
+ +Table 5. Top-1 accuracy (%) on ImageNet-1K of FixMatch, EMA-Teacher and LyapEMA with different models using $1\%$ and $10\%$ of the labeled data for training. “×” means the training failed (accuracy near 0). “-” means no result is available. Results in light blue indicate risk of training instability and collapse. + +beled data, where training fails to converge. In contrast, LyapEMA effectively prevents collapse, achieving $61.30\%$ and $80.30\%$ accuracy, respectively, showing its robustness in low-labeled data scenarios. By dynamically adapting the EMA decay factor based on the student-teacher discrepancy, LyapEMA mitigates pseudo-label bias and enhances convergence. Additional ablation studies and the model's performance under our designed extreme imbalance scenarios are provided in the supplementary material. + +# 6. Conclusion + +In this paper, we introduced Semi-ViM, a novel semi-supervised learning framework that leverages bidirectional SSMs to mitigate learning bias caused by label imbalance in classification tasks. By integrating LyapEMA, a novel stability-aware parameter update mechanism, and SS-Mixup, a hidden-state-level mixing strategy, Semi-ViM enhances both training stability and feature learning. Our extensive experiments on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms CNN-based and Transformer-based SSL models, achieving superior accuracy with fewer labeled samples. Moreover, Semi-ViM requires far fewer parameters than ViT-based SSL models. These findings highlight the potential of state space models as a competitive alternative to Transformers in semi-supervised learning, particularly in addressing real-world imbalanced datasets. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8443-8452, 2021. 7 +[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3 +[3] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 8 +[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. Advances in Neural Information Processing Systems, 35:25697-25710, 2022. 1, 2, 3, 7, 8 +[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921, 2023. 7 +[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7 +[7] Nieves Crasto. Class imbalance in object detection: An experimental diagnosis and study of mitigation strategies. arXiv preprint arXiv:2403.07113, 2024. 2 +[8] Charika De Alvis and Suranga Seneviratne. A survey of deep long-tail classification advancements. arXiv preprint arXiv:2404.15593, 2024. 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6 +[10] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Towards semi-supervised learning with nonrandom missing labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16121-16131, 2023. 2 +[11] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine-grained learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11829-11837, 2024. 2 +[12] Marawan Elbatel, Hualiang Wang, Jixiang Chen, Hao Wang, and Xiaomeng Li. Learning unlabeled clients divergence via anchor model aggregation for federated semi-supervised learning. arXiv e-prints, pages arXiv-2407, 2024. 2 +[13] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised + +learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 2 +[14] Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3187-3197, 2023. 7 +[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2 +[16] Lan-Zhe Guo, Lin-Han Jia, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning in open environments. Frontiers of Computer Science, 19(8):198345, 2025. 1 +[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5 +[18] Ahmet Iscen, Alireza Fathi, and Cordelia Schmid. Improving image recognition by retrieving from web-scale image-text data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19295-19304, 2023. 8 +[19] Zhen Jiang, Lingyun Zhao, Yu Lu, Yongzhao Zhan, and Qirong Mao. A semi-supervised resampling method for class-imbalanced learning. Expert Systems with Applications, 221:119733, 2023. 2 +[20] Suraj Kothawade, Pavan Kumar Reddy, Ganesh Ramakrishnan, and Rishabh Iyer. Basil: Balanced active semi-supervised learning for class imbalanced datasets. arXiv preprint arXiv:2203.05651, 2022. 1 +[21] Alex Kurakin, Colin Raffel, David Berthelot, Ekin Dogus Cubuk, Han Zhang, Kihyuk Sohn, and Nicholas Carlini. Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. 2020. 2 +[22] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2 +[23] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2 +[24] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 2 +[25] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 7 +[26] Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy. arXiv preprint arXiv:2405.04404, 2024. 3 +[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan + +Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2 +[28] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6 +[29] Teli Ma, Shijie Geng, Mengmeng Wang, Jing Shao, Jiasen Lu, Hongsheng Li, Peng Gao, and Yu Qiao. A simple long-tailed recognition baseline via vision-language model. arXiv preprint arXiv:2111.14745, 2021. 8 +[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2 +[31] Khanh-Binh Nguyen. Sequencematch: revisiting the design of weak-strong augmentations for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 96-106, 2024. 3 +[32] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5 +[33] Sucheng Ren, Xianhang Li, Haoqin Tu, Feng Wang, Fangxun Shu, Lei Zhang, Jieru Mei, Linjie Yang, Peng Wang, Heng Wang, et al. Autoregressive pretraining with mamba in vision. arXiv preprint arXiv:2406.07537, 2024. 5 +[34] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 3 +[35] Jay C. Rothenberger and Dimitrios I. Diochnos. Meta cotraining: Two views are better than one. In 2025 Joint Mathematics Meetings (JMM 2025). AMS. 7 +[36] Shankar Sastry and Shankar Sastry. Lyapunov stability theory. Nonlinear systems: analysis, stability, and control, pages 182-234, 1999. 3 +[37] Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. Long-tail learning with foundation model: Heavy fine-tuning hurts. In *Forty-first International Conference on Machine Learning*, 2024. 8 +[38] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 7, 8 +[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 1 +[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2 + +[41] Renzhen Wang, Xixi Jia, Quanziang Wang, Yichen Wu, and Deyu Meng. Imbalanced semi-supervised learning with bias adaptive classifier. arXiv preprint arXiv:2207.13856, 2022.1 +[42] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 7, 8 +[43] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In European conference on computer vision, pages 605-620. Springer, 2022. 1, 2, 7 +[44] Zebin You, Yong Zhong, Fan Bao, Jiacheng Sun, Chongxuan Li, and Jun Zhu. Diffusion models and semi-supervised learners benefit mutually with few labels. Advances in Neural Information Processing Systems, 36:43479-43495, 2023. 7 +[45] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2 +[46] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8373-8383, 2023. 1 +[47] L Zhu, B Liao, Q Zhang, X Wang, W Liu, and X Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arxiv 2024. arXiv preprint arXiv:2401.09417. 1, 2, 3, 6 \ No newline at end of file diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/images.zip b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1112cc49745a1302cb86d8dad078b1e889d9cd8c --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:253db359987e361cec776e40444a5ad760494a2c1f000aa941f76cc92d446401 +size 635388 diff --git a/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/layout.json b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..662885fad606b5fb112638c0c9bafb996e7c2f1e --- /dev/null +++ b/2025/Semi-ViM_ Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning/layout.json @@ -0,0 +1,10010 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "text", + "content": "Semi-ViM: Bidirectional State Space Model for Mitigating Label Imbalance in Semi-Supervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 161, + 225, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 161, + 225, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 161, + 225, + 190 + ], + "type": "text", + "content": "Hongyang He* \nUniversity of Warwick" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 253, + 161, + 365, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 161, + 365, + 190 + ], + "spans": [ + { + "bbox": [ + 253, + 161, + 365, + 190 + ], + "type": "text", + "content": "Hongyang Xie \nUniversity of Warwick" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 394, + 161, + 498, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 161, + 498, + 190 + ], + "spans": [ + { + "bbox": [ + 394, + 161, + 498, + 190 + ], + "type": "text", + "content": "Haochen You Columbia University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 248, + 195, + 360, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 195, + 360, + 224 + ], + "spans": [ + { + "bbox": [ + 248, + 195, + 360, + 224 + ], + "type": "text", + "content": "Victor Sanchez \nUniversity of Warwick" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 251, + 200, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 251, + 200, + 264 + ], + "spans": [ + { + "bbox": [ + 152, + 251, + 200, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) is often hindered by learning biases when imbalanced datasets are used for training, which limits its effectiveness in real-world applications. In this paper, we propose Semi-ViM, a novel SSL framework based on Vision Mamba, a bidirectional state space model (SSM) that serves as a superior alternative to Transformer-based architectures for visual representation learning. Semi-ViM effectively deals with imbalanced datasets and improves model stability through two key innovations: LyapEMA, a stability-aware parameter update mechanism inspired by Lyapunov theory, and SSMixup, a novel mixup strategy applied at the hidden state level of bidirectional SSMs. Experimental results on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms state-of-the-art SSL models, achieving " + }, + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "inline_equation", + "content": "85.40\\%" + }, + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "text", + "content": " accuracy with only " + }, + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 276, + 297, + 479 + ], + "type": "text", + "content": " of the labeled data, surpassing Transformer-based methods such as Semi-ViT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 135, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) aims to mitigate the scarcity of large-scale labeled training data while reducing annotation costs [16]. In the context of classification, however, a challenge remains: Most state-of-the-art (SOTA) SSL models are trained under the assumption that access to a dataset with a balanced number of labeled samples per class is possible, which is often overly idealized and deviates significantly from real-world conditions [20]. In practice, label imbalance induces learning biases in SSL, ultimately compromising model performance [41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "content": "Due to its outstanding performance, many studies have integrated the vision Transformer (ViT) into SSL methods [4, 43, 46]; e.g., Semi-ViT, adopts a pre-training and fine-tuning approach to achieve SOTA results [4]. How-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 251, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 350 + ], + "type": "image", + "image_path": "31f478e572973ac2bff382619233040efc743ffdbc670b2ce3ef03e085494769.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. Semi-ViT-Small trained with only " + }, + { + "bbox": [ + 313, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 358, + 555, + 415 + ], + "type": "text", + "content": " of the labeled data of ImageNet-1K. Left: Training fails when FixMatch is used. The loss decreases but test accuracy remains near zero. Right: Training is unstable and fails when EMA-Teacher is used: The loss and accuracy oscillate heavily." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 593 + ], + "type": "text", + "content": "ever, Semi-ViT has a weaker inductive bias, requiring large-scale data for effective training [39]. Moroever, as shown by our experiments, Semi-ViT does not perform well on long-tailed imbalanced datasets, like ImageNet-LT, confirming its limitations in handling label imbalance. Although the performance of Semi-ViT can be improved by increasing the model size, e.g., Semi-ViT-Huge, this comes at the expense of a larger number of trainable parameters, around 632M [4]. Vision Mamba (ViM), which incorporates a bidirectional state space model (SSM), has modeling capabilities comparable to those of ViT, while achieving subquadratic-time computational and linear memory complexities [47]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Based on the previous discussions, this paper integrates ViM into an SSL framework for classification and proposes a ViM-based SSL model, hereinafter called Semi-ViM. Semi-ViM incorporates two key innovations aimed at improving model stability during training and performance: LyapEMA and SSMixup. LyapEMA is a novel parameter update mechanism that leverages the Lyapunov theory to dynamically adjust parameter updates, ensuring a more stable training process. LyapEMA addresses the training failures encountered when using only a small amount of labeled" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "765" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "data to train Semi-ViT-Small, as shown in Fig 1. LyapEMA also helps prevent training collapse, commonly encountered when using the exponential moving average (EMA) and FixMatch parameter update mechanisms [4]. SSMixup, on the other hand, is a novel mixup strategy that operates at the hidden state space of SSMs and goes beyond the traditional input-level mixup strategy. Specifically, SSMixup incorporates mixed hidden states as residual units in the training process, improving the model's capability to generalize under sparse label conditions. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 295, + 371 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 251 + ], + "type": "text", + "content": "- We propose Semi-ViM, an SSL framework that outperforms other SSL methods based on CNNs and Transformers. The number of parameters of Semi-ViM is only around one-fourth of that of Semi-ViT." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 251, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 300 + ], + "type": "text", + "content": "- We propose LyapEMA, a novel parameter update mechanism that successfully addresses the training instability issues commonly found in Semi-ViT, thus improving model convergence and generalization capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 334 + ], + "type": "text", + "content": "- We propose SSMixup, a hidden state space-level mixup strategy that effectively enhances the learning capability of bidirectional SSMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 335, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 295, + 371 + ], + "type": "text", + "content": "- Based on evaluations on the long-tailed ImageNet-LT dataset, we demonstrate Semi-ViM's strong capabilities in mitigating learning biases caused by label imbalance." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 382, + 142, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 142, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 142, + 395 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 402, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 486 + ], + "type": "text", + "content": "Imbalanced datasets are those in which the different classes are distributed unequally, often resulting in long-tailed distributions [7]. SSL aims to leverage a small amount of labeled data and a large amount of unlabeled data to improve performance, which is useful to deal with imbalanced datasets. SSL has achieved significant advancements in several computer vision tasks [10, 11, 21, 38, 45]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 653 + ], + "type": "text", + "content": "Traditional SSL methods, e.g., entropy minimization and consistency regularization [13, 15], exploit unlabeled data to reduce dependence on manual annotation. For example, consistency regularization leverages unlabeled data under the assumption that a model's predictions remain consistent under different input perturbations. Mainstream methods include II-Model [22], Mean Teacher [40], Virtual Adversarial Training (VAT) [30], and FixMatch [38], which is an important benchmark in SSL [38]. Pseudo-Labeling [23] is a common strategy used in SSL. It consists of using the model's predictions on unlabeled data as pseudo-labels for training. Noisy Student, FixMatch [38], and FlexMatch [45] adopt this strategy and use confidence thresholds to filter out pseudo-labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Several SOTA SSL methods are based on FixMatch or any of its variants, e.g., FlexMatch and CoMatch [24], which integrate contrastive learning or self-supervised methods to enhance performance. However, these methods typically assume that the class distribution is balanced," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "which is rarely the case in real-world scenarios [12]. When dealing with imbalanced datasets [8], these SOTA SSL methods face two primary challenges. First, pseudo-label biases, i.e., models tend to predict more frequently the majority classes when generating pseudo-labels, exacerbating the imbalance issue and limiting learning from the minority classes [19]. Second, low-quality pseudo-labels, i.e., models tend to assign low confidence scores to those pseudolabels generated for the minority class samples, thus reducing their impact on training and potentially leading to training collapse [4]. For example, FixMatch relies on a fixed confidence threshold for selecting pseudo-labels, but in an imbalanced dataset setting, pseudo-labels of the minority classes often fail to reach this threshold, leading to insufficient learning from such classes [4, 43]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 399, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 399, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 399, + 274 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": "State Space Models and ViM: SSMs have a long-standing history in control systems and signal processing, offering a structured approach to sequential data modeling [47]. Deep learning adopts SSMs to more effectively handle long-range dependencies compared to traditional recurrence-based models. Unlike Transformers, which rely on global self-attention, SSMs leverage structured state transitions for sequential data propagation, enabling subquadratic computational complexity and linear memory scaling [47]. In computer vision, SSMs have been integrated into deep architectures to enhance representation learning by modeling both spatial and temporal dependencies [27, 47]. The SSM is formulated as a discrete-time state-space system [47]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 445, + 553, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 445, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 365, + 445, + 553, + 459 + ], + "type": "interline_equation", + "content": "h _ {t} = A h _ {t - 1} + B x _ {t}, \\quad y _ {t} = C h _ {t}, \\tag {1}", + "image_path": "dee01423ef24ea3dcde828ec0ebe13047586627bac305674789bb20772ef275b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": " is the state transition matrix; " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": " are input and output projection matrices, respectively; " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": " represents the input sequence at time step " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": "; and " + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 465, + 553, + 548 + ], + "type": "text", + "content": " is the corresponding output representation. The parameters of an SSM are usually optimized in a data-dependent manner to ensure computational efficiency and parallelization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "text", + "content": "The bidirectional extension of SSMs further improves feature learning by capturing both forward and backward information, making it a competitive alternative to self-attention mechanisms used by Transformers. [47]. The bidirectional SSM operates in the forward " + }, + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "text", + "content": " and backward " + }, + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 313, + 550, + 553, + 620 + ], + "type": "text", + "content": " direction:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 628, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 628, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 361, + 628, + 553, + 643 + ], + "type": "interline_equation", + "content": "h _ {t} ^ {f} = A h _ {t - 1} ^ {f} + B x _ {t}, \\quad y _ {t} ^ {f} = C h _ {t} ^ {f}; \\tag {2}", + "image_path": "035c25cddda8f5b1b3219d25e4e224444b570f5e41c06c45b3ea29413b406884.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 650, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 650, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 364, + 650, + 553, + 666 + ], + "type": "interline_equation", + "content": "h _ {t} ^ {b} = A h _ {t + 1} ^ {b} + B x _ {t}, \\quad y _ {t} ^ {b} = C h _ {t} ^ {b}. \\tag {3}", + "image_path": "3adf70d30d1922ddadaf5d83997926c721a07a28a36c1ba4c92f7ec5f548808a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 669, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 553, + 693 + ], + "type": "text", + "content": "The final output at time step " + }, + { + "bbox": [ + 313, + 669, + 553, + 693 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 669, + 553, + 693 + ], + "type": "text", + "content": " is obtained by merging the forward and backward outputs through a gating mechanism:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "y _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b}, \\tag {4}", + "image_path": "1512ebb6ff68b6270ca9d55cd81da58f17491cad63223e1750391460c1ee8a90.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "766" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 201, + 213 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 201, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 201, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 201, + 213 + ], + "type": "image", + "image_path": "3653d4902f23b10c9df92783cd4756241dcf665036e59c57e7732dff742acb1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 219, + 143, + 228 + ], + "lines": [ + { + "bbox": [ + 108, + 219, + 143, + 228 + ], + "spans": [ + { + "bbox": [ + 108, + 219, + 143, + 228 + ], + "type": "text", + "content": "(a) FixMatch" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 379, + 215 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 379, + 215 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 379, + 215 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 379, + 215 + ], + "type": "image", + "image_path": "f810dfa34687df9fa1847eeb2d61ff8e17d4b2205639902f1751d62b55cef712.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 219, + 324, + 228 + ], + "lines": [ + { + "bbox": [ + 277, + 219, + 324, + 228 + ], + "spans": [ + { + "bbox": [ + 277, + 219, + 324, + 228 + ], + "type": "text", + "content": "(b) EMA-Teacher" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Differences between (a) FixMatch, (b) EMA-Teacher and (c) Semi-ViM (ours). In the figure, " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "X_{w}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " are, respectively, strongly and weakly augmented inputs; " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "p^{s}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "p_{student}^{s}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " are the predictions of the student model on " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "X_{s}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "p^{w}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "p_{teacher}^{w}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " are the predictions of the teacher model on " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "X_{w}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " is the threshold used to select pseudo-labels; and " + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\theta, \\theta^{student}, \\theta^{teacher}" + }, + { + "bbox": [ + 54, + 239, + 555, + 285 + ], + "type": "text", + "content": " are, respectively, the learnable parameters for the entire model, the student model, and the teacher model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 556, + 215 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 556, + 215 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 556, + 215 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 556, + 215 + ], + "type": "image", + "image_path": "a3714e3218e742eb564341217c5bc94abe38467f6b3ea0b918b6481af82df601.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 219, + 509, + 228 + ], + "lines": [ + { + "bbox": [ + 456, + 219, + 509, + 228 + ], + "spans": [ + { + "bbox": [ + 456, + 219, + 509, + 228 + ], + "type": "text", + "content": "(c) Semi-ViM(Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "inline_equation", + "content": "\\mathrm{SiLU}(\\cdot)" + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "content": " is the Sigmoid-weighted Linear Unit activation function, and " + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "content": " is a gating vector that adaptively balances bidirectional information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "type": "text", + "content": "ViM adopts a bidirectional SSM for visual representation learning and integrates positional embeddings for spatial awareness. In ViM, the input image is first partitioned into patches and projected onto a high-dimensional feature space before being processed. With a linear computational complexity, ViM provides data-dependent global visual modeling, significantly improving efficiency in high-resolution tasks compared to Transformer-based architectures [26, 47]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 453, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 295, + 598 + ], + "type": "text", + "content": "Mixup strategies: These strategies interpolate input samples and labels, and hence can be used as data augmentation techniques to improve generalization capabilities [2]. In SSL, they can enhance pseudo-labeling and provide consistency regularization [34]. Recent variants extend mixup strategies to the feature space and pseudo-label frameworks, e.g., the probabilistic pseudo mixup strategy refines pseudolabel quality by weighting interpolated samples based on their confidence scores, which are derived from the model's softmax outputs or uncertainty estimates. These variants can improve SSL robustness, particularly for ViTs, which benefit from stronger regularization [4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Weak and strong data augmentation: Weak data augmentation involves simple transformation, e.g., random cropping, flipping, or slight color jitter, which introduce small variations to the original data to help a model learn from basic invariances. Strong data augmentation, on the other hand, applies more aggressive transformations, e.g., large crops, strong color distortions, or more significant perturbations, which force a model to learn more robust representations by simulating more challenging scenarios [31]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 493, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 493, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 493, + 319 + ], + "type": "text", + "content": "4. Proposed Semi-ViM framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": "Semi-ViM adopts a teacher-student framework and uses a labeled dataset, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_l = \\{(x_i^l,y_i^l)\\}_{i = 1}^{N_l}" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", and an unlabeled dataset " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_u = \\{x_i^u\\}_{i = 1}^{N_u}" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": " for training. Specifically, the teacher model, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{teacher}}(\\theta^{\\mathrm{teacher}})" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", generates pseudo-labels, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_i\\}" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", for unlabeled samples, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\{x_i^u\\}" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", and outputs confidence scores, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\{o_i\\}" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", for the pseudo-labels. The student model, " + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{student}}(\\theta^{\\mathrm{student}})" + }, + { + "bbox": [ + 313, + 325, + 555, + 445 + ], + "type": "text", + "content": ", is then trained on samples mixed according to their confidence scores. The difference between the training paradigm of Semi-ViM and those of FixMatch and EMA-Teacher is shown in Fig 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 446, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 555 + ], + "type": "text", + "content": "In this section, we first introduce LyapEMA, which is inspired by Lyapunov stability theory [36] and stabilizes parameter updates, thus preventing model collapse during the training of Semi-ViM. Next, we introduce SSMixup, a new mixup strategy designed for SSM-based models. Unlike traditional mixup strategies that operate solely on the input level, SSMixup mixes the hidden states of bidirectional SSMs. Finally, we detail the training and loss function of Semi-ViM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 563, + 387, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 387, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 387, + 576 + ], + "type": "text", + "content": "4.1. LyapEMA" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 689 + ], + "type": "text", + "content": "LyapEMA is a novel optimization framework that integrates Lyapunov stability theory with EMA. LyapEMA is designed for teacher-student frameworks under an SSL setting. It has two main objectives. The first is to ensure stable parameter updates, because training SSL frameworks, like Semi-ViT, with a limited amount of labeled data can lead to training collapse. The second objective is reducing the learning bias associated with pseudo-labels while preserving the advantages of EMA in refining the teacher model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{student}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{teacher}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denote the student and teacher model parameters, respectively. EMA updates the teacher" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "767" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 221, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 221, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 221, + 83 + ], + "type": "text", + "content": "model using gradient descent as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 98, + 95, + 294, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 95, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 98, + 95, + 294, + 109 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {\\text {t e a c h e r}} = \\alpha_ {t} \\theta_ {t - 1} ^ {\\text {t e a c h e r}} + (1 - \\alpha_ {t}) \\theta_ {t} ^ {\\text {s t u d e n t}}, \\tag {5}", + "image_path": "f06b0da54263248223ab6f6cb188e084a771cf7ae26f124e710f9852f590ec76.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "text", + "content": " is the EMA decay factor. However, EMA cannot guarantee stability, potentially leading to divergence due to accumulated bias associated with learning from pseudo-labels. To address this issue, LyapEMA introduces Lyapunov stability constraints to guide the updates. Let us define a Lyapunov candidate function, " + }, + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "inline_equation", + "content": "V(\\theta^{\\mathrm{student}},\\theta^{\\mathrm{teacher}})" + }, + { + "bbox": [ + 55, + 113, + 296, + 207 + ], + "type": "text", + "content": ", to measure the discrepancy between the student and teacher models:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 224, + 295, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 224, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 87, + 224, + 295, + 247 + ], + "type": "interline_equation", + "content": "V \\left(\\theta^ {\\text {s t u d e n t}}, \\theta^ {\\text {t e a c h e r}}\\right) = \\frac {1}{2} \\| \\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {6}", + "image_path": "e4e0424e1309490cf97d6c356e46a7b89fc2e9dd1297bb3640e3194df8000d54.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "content": "Note that the function in Eq. 6 acts as a stability indicator, ensuring that the student model does not drift too far from the teacher model. The Lyapunov stability criterion requires that the derivative of " + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "inline_equation", + "content": "V(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "content": " w.r.t. time, i.e., " + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\dot{V}" + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "content": ", satisfies:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 327, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 327, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 327, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\dot {V} = \\frac {d V}{d t} = \\left(\\theta^ {\\text {s t u d e n t}} - \\theta^ {\\text {t e a c h e r}}\\right) ^ {\\mathrm {T}} \\left(\\dot {\\theta} ^ {\\text {s t u d e n t}} - \\dot {\\theta} ^ {\\text {t e a c h e r}}\\right) < 0, \\tag {7}", + "image_path": "2753ccb9abc3585224234365471ed1d27ef54f3503fd94be90ed3e0188db0190.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\dot{\\theta}^{\\mathrm{student}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\dot{\\theta}^{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 407 + ], + "type": "text", + "content": " also derivatives w.r.t. time. The criterion in Eq. 7 guarantees that the discrepancy between the student and teacher models decreases over time, ensuring convergence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 443 + ], + "type": "text", + "content": "To enforce stability while leveraging EMA for smooth teacher model updates, LyapEMA updates the student model parameters by using aLyapunov regularization term:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 460, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 460, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 60, + 460, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\theta_ {t} ^ {\\text {s t u d e n t}} = \\theta_ {t - 1} ^ {\\text {s t u d e n t}} - \\eta \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) + \\lambda \\left(\\theta_ {t} ^ {\\text {t e a c h e r}} - \\theta_ {t} ^ {\\text {s t u d e n t}}\\right), \\tag {8}", + "image_path": "7e8005c2f50ead3c60c16cdfff299bd2e74ecb87d48526e5a2ac8536b4be5ab2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta^{\\mathrm{student}})" + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "content": " is a semi-supervised loss function, " + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "content": " is the learning rate, and " + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 481, + 295, + 516 + ], + "type": "text", + "content": " is a stability coefficient that ensures the student model remains close to the teacher model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 564 + ], + "type": "text", + "content": "LyapEMA updates the teacher model parameters by using EMA as denoted by Eq. 5. However, the EMA decay factor " + }, + { + "bbox": [ + 55, + 517, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 517, + 296, + 564 + ], + "type": "text", + "content": " is dynamically adjusted based on the Lyapunov function:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 575, + 294, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 575, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 99, + 575, + 294, + 590 + ], + "type": "interline_equation", + "content": "\\alpha_ {t} = \\operatorname {s i g m o i d} \\left(\\gamma \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t - 1} ^ {\\text {t e a c h e r}} \\| _ {2}\\right), \\tag {9}", + "image_path": "c9ce24f935a12836dd7753139a64268bcbf76b4597a482755146635dbf3dca88.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": " is a sensitivity hyperparameter controlling how " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": " adapts based on the student-teacher discrepancy, and the sigmoid function ensures " + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": " remains within (0, 1), preventing instability in the updates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "type": "text", + "content": "LyapEMA ensures that updates lead to stable training by using the Lyapunov decrease condition:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 681, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 681, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 681, + 296, + 714 + ], + "type": "interline_equation", + "content": "V _ {t + 1} - V _ {t} = \\frac {1}{2} \\| \\theta_ {t + 1} ^ {\\text {s t u d e n t}} - \\theta_ {t + 1} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2} - \\frac {1}{2} \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {10}", + "image_path": "51c539af136c97c660cd1cbe465eb9d80a2979845a67aa630edbf986f31a3b7b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 72, + 414, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 414, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 414, + 84 + ], + "type": "text", + "content": "Algorithm 1 LyapEMA" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": "Input: Initial parameters " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{student}}" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{teacher}}" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": ", stability coefficient " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": ", sensitivity " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": " \nOutput: Stabilized parameters " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{student}}" + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 86, + 553, + 123 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{teacher}}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 125, + 553, + 327 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 321, + 125, + 463, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 125, + 463, + 136 + ], + "spans": [ + { + "bbox": [ + 321, + 125, + 463, + 136 + ], + "type": "text", + "content": "1: while training not converged do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 137, + 441, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 137, + 441, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 137, + 441, + 148 + ], + "type": "text", + "content": "2: Compute loss " + }, + { + "bbox": [ + 321, + 137, + 441, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta^{\\mathrm{student}})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 149, + 553, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 149, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 321, + 149, + 553, + 185 + ], + "type": "text", + "content": "3: Update student model: " + }, + { + "bbox": [ + 321, + 149, + 553, + 185 + ], + "type": "inline_equation", + "content": "\\theta_{t}^{\\mathrm{student}} = \\theta_{t - 1}^{\\mathrm{student}} - \\eta \\nabla_{\\theta}\\mathcal{L}(\\theta^{\\mathrm{student}}) + \\lambda (\\theta_{t}^{\\mathrm{teacher}} - \\theta_{t}^{\\mathrm{student}})" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 185, + 524, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 185, + 524, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 185, + 524, + 209 + ], + "type": "text", + "content": "4: Compute adaptive EMA decay in LyapEMA: " + }, + { + "bbox": [ + 321, + 185, + 524, + 209 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\mathrm{sigmoid}(\\gamma \\| \\theta_{t}^{\\mathrm{student}} - \\theta_{t - 1}^{\\mathrm{teacher}}\\|_{2})" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 209, + 490, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 209, + 490, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 209, + 490, + 232 + ], + "type": "text", + "content": "5: Update teacher model: " + }, + { + "bbox": [ + 321, + 209, + 490, + 232 + ], + "type": "inline_equation", + "content": "\\theta_{t}^{\\mathrm{teacher}} = \\alpha_{t}\\theta_{t - 1}^{\\mathrm{teacher}} + (1 - \\alpha_{t})\\theta_{t}^{\\mathrm{student}}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 233, + 501, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 233, + 501, + 243 + ], + "spans": [ + { + "bbox": [ + 321, + 233, + 501, + 243 + ], + "type": "text", + "content": "6: Ensure stability: Verify " + }, + { + "bbox": [ + 321, + 233, + 501, + 243 + ], + "type": "inline_equation", + "content": "V(\\theta)" + }, + { + "bbox": [ + 321, + 233, + 501, + 243 + ], + "type": "text", + "content": " decreases" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 244, + 422, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 244, + 422, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 244, + 422, + 255 + ], + "type": "text", + "content": "7: if " + }, + { + "bbox": [ + 321, + 244, + 422, + 255 + ], + "type": "inline_equation", + "content": "V_{t} \\geq V_{t-1}" + }, + { + "bbox": [ + 321, + 244, + 422, + 255 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 257, + 479, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 257, + 479, + 268 + ], + "spans": [ + { + "bbox": [ + 321, + 257, + 479, + 268 + ], + "type": "text", + "content": "8: Reduce learning rate: " + }, + { + "bbox": [ + 321, + 257, + 479, + 268 + ], + "type": "inline_equation", + "content": "\\eta \\gets \\eta /2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 269, + 525, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 269, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 321, + 269, + 525, + 280 + ], + "type": "text", + "content": "9: Increase stability coefficient: " + }, + { + "bbox": [ + 321, + 269, + 525, + 280 + ], + "type": "inline_equation", + "content": "\\lambda \\gets \\lambda +\\Delta \\lambda" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 281, + 512, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 281, + 512, + 292 + ], + "spans": [ + { + "bbox": [ + 318, + 281, + 512, + 292 + ], + "type": "text", + "content": "10: Recompute student and teacher updates" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 293, + 369, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 293, + 369, + 303 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 369, + 303 + ], + "type": "text", + "content": "11: end if" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 304, + 375, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 304, + 375, + 315 + ], + "spans": [ + { + "bbox": [ + 318, + 304, + 375, + 315 + ], + "type": "text", + "content": "12: end while" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "spans": [ + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "type": "text", + "content": "13: return " + }, + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{teacher}}" + }, + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 318, + 316, + 419, + 327 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{student}}" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 354, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 378 + ], + "type": "text", + "content": "By using the LyapEMA update rules in Eq. 10 and expanding, we obtain:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 400, + 553, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 400, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 400, + 553, + 426 + ], + "type": "interline_equation", + "content": "V _ {t + 1} - V _ {t} \\approx - \\eta \\| \\nabla_ {\\theta} \\mathcal {L} \\left(\\theta^ {\\text {s t u d e n t}}\\right) \\| _ {2} ^ {2} - \\lambda \\| \\theta_ {t} ^ {\\text {s t u d e n t}} - \\theta_ {t} ^ {\\text {t e a c h e r}} \\| _ {2} ^ {2}. \\tag {11}", + "image_path": "c96e534ce3ef2f748145e5961469c6b09c7bbacdcf5804473ad3c17825eb371f.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 426, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 487 + ], + "type": "text", + "content": "Since both terms in Eq. 11 are negative, " + }, + { + "bbox": [ + 313, + 426, + 555, + 487 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 426, + 555, + 487 + ], + "type": "text", + "content": " decreases monotonically, ensuring stability. The complete LyapEMA framework is summarized in Algorithm 1. In Section 5.1, we conduct experiments not only on Semi-ViM but also on other SSL methods to validate LyapEMA's effectiveness." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 496, + 381, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 381, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 381, + 509 + ], + "type": "text", + "content": "4.2. SSMixup" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 514, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 586 + ], + "type": "text", + "content": "SSMixup introduces mixup strategies in the input and the hidden state space of a bidirectional SSM to effectively leverage unlabeled data and enhance generalization in SSL. As mentioned before, Semi-ViM adopts a teacher-student framework where the student model, " + }, + { + "bbox": [ + 313, + 514, + 555, + 586 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{student}}(\\theta^{\\mathrm{student}})" + }, + { + "bbox": [ + 313, + 514, + 555, + 586 + ], + "type": "text", + "content": ", is trained on the mixed samples." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "content": "In SSMixup, the pseudo-label confidence scores, " + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "content": ", of two unlabeled samples, " + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "inline_equation", + "content": "x_i^u" + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "inline_equation", + "content": "x_j^u" + }, + { + "bbox": [ + 313, + 586, + 554, + 624 + ], + "type": "text", + "content": ", respectively, are used first to compute a mixup coefficient:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 403, + 632, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 632, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 403, + 632, + 553, + 656 + ], + "type": "interline_equation", + "content": "\\psi = \\frac {o _ {i}}{o _ {i} + o _ {j}}. \\tag {12}", + "image_path": "a63fa3276c247f1d11b2be4b665226baca0505ca9c0bee51fc551d0a94319838.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": "The mixup coefficient " + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": " is used to form the mixed input and labels:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 319, + 701, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 701, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 319, + 701, + 553, + 715 + ], + "type": "interline_equation", + "content": "x _ {\\text {m i x}} = \\psi x _ {i} ^ {u} + (1 - \\psi) x _ {j} ^ {u}, \\quad y _ {\\text {m i x}} = \\psi \\hat {y} _ {i} + (1 - \\psi) \\hat {y} _ {j}. \\tag {13}", + "image_path": "1b9bc2da4a636fd7d46931b23a1f88da42282034e58d27683942550db9c097c2.jpg" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "768" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 294, + 183 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 183 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 183 + ], + "type": "image", + "image_path": "03c75fa5720095ae52eb1d7609617dd683315634110e7abbd6ace3ba5eebf439.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 296, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 225 + ], + "type": "text", + "content": "Figure 3. The SSMixup strategy. The red part applies mixup at the hidden state level. " + }, + { + "bbox": [ + 55, + 190, + 296, + 225 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 190, + 296, + 225 + ], + "type": "text", + "content": " denotes to the number of SSMixup encoder layers in Semi-ViM. It is a hyperparameter typically set to 24." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "content": "The mixed input " + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{mix}}" + }, + { + "bbox": [ + 55, + 244, + 296, + 281 + ], + "type": "text", + "content": " is then fed into the forward and backward components of the bidirectional SSM, yielding hidden states:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 289, + 295, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 289, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 68, + 289, + 295, + 304 + ], + "type": "interline_equation", + "content": "h ^ {\\mathrm {f w d}} = \\operatorname {S S M} _ {\\mathrm {f w d}} \\left(x _ {\\text {m i x}}\\right), \\quad h ^ {\\mathrm {b w d}} = \\operatorname {S S M} _ {\\mathrm {b w d}} \\left(x _ {\\text {m i x}}\\right). \\tag {14}", + "image_path": "c407218e4020ea1fe8919cb4180e20de77e6e10f2ff8d9b00b0e6c7f3496307a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 336 + ], + "type": "text", + "content": "SSMixup mixes the hidden states by using the mixup coefficient, " + }, + { + "bbox": [ + 55, + 312, + 296, + 336 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 312, + 296, + 336 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 345, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 345, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 112, + 345, + 295, + 361 + ], + "type": "interline_equation", + "content": "h _ {\\text {m i x}} = \\psi h ^ {\\mathrm {f w d}} + (1 - \\psi) h ^ {\\mathrm {b w d}}, \\tag {15}", + "image_path": "d8860045683cd9dd55058a673c0a2d74340f2bdbdb7da9e757bde4073ec32991.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{mix}}" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": " is computed as a residual unit and " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": " is defined in Eq. 1. The forward and backward outputs of the bidirectional SSM are merged to compute the final output (see Eq. 4):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 437, + 295, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 437, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 65, + 437, + 295, + 453 + ], + "type": "interline_equation", + "content": "y _ {t} = \\operatorname {S i L U} (z) \\odot y _ {t} ^ {f} + \\operatorname {S i L U} (z) \\odot y _ {t} ^ {b} + \\nu \\left(C h _ {\\text {m i x}}\\right). \\tag {16}", + "image_path": "12f0da59960976fb9b1d94ef80c059bc7bd773455eeef2473f79be9ca8d62591.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": " is the control coefficient. The computed " + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": " is used in a cross-entropy loss with " + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{mix}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{student}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": ". Parameters " + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{teacher}}" + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": " are smoothly updated via LyapEMA, maintaining stable pseudo-label generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 582 + ], + "type": "text", + "content": "Figure 3 illustrates the functionality of SSMixup. By introducing mixup strategies at both the input (see Eq. 13) and hidden state levels (see Eq. 15), combined with confidence-driven weighting, SSMixup effectively exploits unlabeled data and significantly improves generalization capabilities in SSL settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "type": "text", + "content": "4.3. Training Semi-ViM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": "Training Semi-ViM involves three main steps. The first step involves self-supervised pre-training based on Autoregressive Mamba (ARM) [33], where labeled data is treated as unlabeled data and used as part of the training samples. The second step involves supervised fine-tuning using only labeled data. The third step involves SSL. These three steps are detailed next." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Self-supervised pre-training based on ARM: To improve feature extraction, we propose to leverage ARM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 274 + ], + "type": "text", + "content": "for self-supervised pre-training of Semi-ViM. ARM is better suited for ViM pre-training than Masked Autoencoders (MAEs) and DINO-V2 (contrastive learning) [32] due to its alignment with ViM's sequential state-space modeling [33]. Unlike MAEs, which rely on masked image reconstruction and bidirectional Transformers, Semi-ViM's bidirectional state-space modeling allows to effectively leverage ARM's autoregressive token prediction, making ARM a more natural fit [17]. Compared to DINO-V2, which depends on global feature alignment and large-batch contrastive learning, ARM directly optimizes sequential token dependencies, avoiding instability from batch size sensitivity and augmentation complexity [9]. By leveraging ViM's structured recurrence, ARM enables faster training, better scalability, and improved efficiency over both methods. These statements are validated by our experiments in Section 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 275, + 556, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 556, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 556, + 432 + ], + "type": "text", + "content": "Fully-supervised fine-tuning: After completing ARM-based pre-training, Semi-ViM undergoes fine-tuning using only labeled data. This process aims to leverage the representations learned during pre-training to adapt the model to specific downstream tasks. During fine-tuning, most pretrained parameters remain unchanged, with only certain layers or global parameters being updated to a limited extent in order to improve performance in SSL. Fine-tuning specifically targets layers responsible for learning bidirectional dependencies, i.e., the forward and backward SSMs are updated to better capture dependencies specific to a classification task. The class token used in image classification tasks is also fine-tuned during this phase." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 432, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 503 + ], + "type": "text", + "content": "SSL: In the third training step, we use LyapEMA, in conjunction with SSMixup to fully exploit unlabeled data while ensuring training stability and efficiency. This last training step consists of five key components: weak and strong data augmentation, pseudo-label generation, LyapEMA, SSMixup, and minimization of the loss function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": "For each unlabeled sample " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "x^{u}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": ", we generate a weakly augmented sample " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "x_{w}^{u}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": " and a strongly augmented sample " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "x_{s}^{u}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": ". The weakly augmented sample " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "x_{w}^{u}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": " is processed by the teacher model " + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "inline_equation", + "content": "f_{\\text{teacher}}" + }, + { + "bbox": [ + 313, + 504, + 554, + 563 + ], + "type": "text", + "content": " to produce a set of predicted probabilities, one for each class:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 577, + 555, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 577, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 383, + 577, + 555, + 592 + ], + "type": "interline_equation", + "content": "p _ {\\text {t e a c h e r} _ {x _ {w} ^ {u}}} = f _ {\\text {t e a c h e r}} \\left(x _ {w} ^ {u}\\right). \\tag {17}", + "image_path": "fa7b0b8aebd66775a9014abb6216c4c4c7b92dd91af989d4b70c2835c990a37d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "text", + "content": "The pseudo-label for " + }, + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "inline_equation", + "content": "x_w^u" + }, + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "text", + "content": " is then determined as the class associated with the highest probability in set " + }, + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "inline_equation", + "content": "p_{\\text{teacher}_x u}" + }, + { + "bbox": [ + 313, + 597, + 554, + 624 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 634, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 634, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 380, + 634, + 553, + 649 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {u} = \\arg \\max \\left(p _ {\\text {t e a c h e r} _ {x _ {w}} ^ {u}}\\right). \\tag {18}", + "image_path": "2c6eadaa77bcf62b1c2446a5ff23e8a90025c41557ecf455531fb87377f038d6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Note that only unlabeled samples with a class probability exceeding the threshold " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " are retained for training; i.e., only those with high-score pseudo-labels are retained. SSMixup is then applied to these unlabeled samples with high-score pseudo labels in order to be used to train the student model." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "769" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 127 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 127 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 127 + ], + "type": "table", + "html": "
Variant nameBackboneParam.Input size
Semi-ViM-TinyViM-Tiny [47]7M2242
Semi-ViM-SmallViM-Small [47]26M2242
Semi-ViM-BaseViM-Base [47]98M2242
", + "image_path": "fa5bf86679e1effb6541121f76e6b04eb9398562b3d48ea47b95b2144f768e48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 135, + 257, + 146 + ], + "lines": [ + { + "bbox": [ + 93, + 135, + 257, + 146 + ], + "spans": [ + { + "bbox": [ + 93, + 135, + 257, + 146 + ], + "type": "text", + "content": "Table 1. The different variants of Semi-ViM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "text", + "content": "The overall loss function consists of the supervised loss " + }, + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "inline_equation", + "content": "L_{\\text{labeled}}" + }, + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "text", + "content": " and the unsupervised loss " + }, + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "inline_equation", + "content": "L_{\\text{unlabeled}}" + }, + { + "bbox": [ + 56, + 165, + 295, + 190 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 201, + 294, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 201, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 123, + 201, + 294, + 213 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {l a b e l e d}} + \\mu L _ {\\text {u n l a b e l e d}}. \\tag {19}", + "image_path": "2fb9266616676ca5aa31c5a6ea64deb9681c47e0da39bbdd9342abf01e50e254.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 217, + 295, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 241 + ], + "type": "text", + "content": "The supervised loss, " + }, + { + "bbox": [ + 55, + 217, + 295, + 241 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{labeled}}" + }, + { + "bbox": [ + 55, + 217, + 295, + 241 + ], + "type": "text", + "content": ", is based on the standard cross-entropy (CE) loss using labeled samples:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 250, + 295, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 250, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 99, + 250, + 295, + 282 + ], + "type": "interline_equation", + "content": "L _ {\\text {l a b e l e d}} = \\frac {1}{N _ {l}} \\sum_ {i = 1} ^ {N _ {l}} \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {i} ^ {l}\\right), y _ {i} ^ {l}\\right). \\tag {20}", + "image_path": "4389765fb75aa8665d102a617790f44cac2ece12bbb99e60ab86e8bc5e89264f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "text", + "content": "The unsupervised loss " + }, + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{unlabeled}}" + }, + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "text", + "content": " uses strongly augmented samples, " + }, + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "inline_equation", + "content": "x_{s}^{u}" + }, + { + "bbox": [ + 55, + 284, + 295, + 309 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 323, + 295, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 295, + 356 + ], + "type": "interline_equation", + "content": "L _ {\\text {u n l a b e l e d}} = \\frac {1}{N _ {u}} \\sum_ {i = 1} ^ {N _ {u}} \\mathbb {1} \\left(o _ {i} > \\tau\\right) \\mathrm {C E} \\left(f _ {\\text {s t u d e n t}} \\left(x _ {s, i} ^ {u}\\right), \\hat {y} _ {i} ^ {u}\\right), \\tag {21}", + "image_path": "0c08bcc708b751e5911f0fb3b405d480af71af418966282a94ee69951df4ae03.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "where the indicator function " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(o > \\tau)" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " selects only those unlabeled samples with high-score pseudo-labels, where " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " denotes the probability of the class associated with the pseduo-label. Finally, the teacher model is updated using LyapEMA to stabilize parameter evolution and improve the quality of pseudo-labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 442, + 137, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 137, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 137, + 455 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "content": "Table 1 lists the different variants of Semi-ViM we use in our experiments: Tiny, Small, and Base. These variants are trained on ImageNet-1K using 8 NVIDIA A100 (80GB) GPUs with DistributedDataParallel (DDP) PyTorch. We use AdamW as the optimizer with a weight decay of 0.05, and a learning rate that follows a cosine annealing schedule starting from " + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "content": " with 10 epochs of warm-up. We use 300 epochs for training with a global batch size of 1024 (128 per GPU) and automatic mixed precision (AMP) enabled. Strong data augmentation includes AutoAugment, Mixup (" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\alpha = 0.2" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "content": "), CutMix (" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\alpha = 1.0" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "content": "), and label smoothing (" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\varepsilon = 0.1" + }, + { + "bbox": [ + 55, + 462, + 295, + 616 + ], + "type": "text", + "content": "). The total training time is approximately 70 hours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "We conduct classification experiments on ImageNet-1K [9] and ImageNet-LT [28]. ImageNet-1K contains 1.28M training and 50K validation images across 1,000 classes. ImageNet-LT has a long-tailed class distribution. We use a confidence threshold of " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 0.7" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " for pseudo-label selection, with weak and strong data augmentations applied to unlabeled data. Inference is performed using a " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "224\\times 224" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " center crop on the images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 432, + 84 + ], + "type": "text", + "content": "5.1. Results and analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "content": "Semi-ViM demonstrates superior performance across multiple experimental settings as tabulated in Table 2. On ImageNet-1K, Semi-ViM-Base achieves " + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "inline_equation", + "content": "85.40\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "content": " accuracy with only " + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 89, + 553, + 161 + ], + "type": "text", + "content": " of the labeled data, outperforming SemiViT-Huge. Compared to SemiFormer, Semi-ViM consistently excels, confirming the suitability of ViM for SSL." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "content": "For the more challenging ImageNet-LT dataset, Semi-ViM-Base achieves " + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "inline_equation", + "content": "77.40\\%" + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "content": " accuracy with only " + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "content": " of the labeled data, significantly surpassing Semi-ViT-Huge and SemiFormer. This demonstrates Semi-ViM's adaptability to long-tailed data distributions and its capability to work with rare category representations. With only " + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "content": " of the labeled data, Semi-ViM-Base achieves " + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "inline_equation", + "content": "66.30\\%" + }, + { + "bbox": [ + 313, + 161, + 553, + 257 + ], + "type": "text", + "content": " accuracy, surpassing Semi-ViT-Huge and demonstrating its strength in SSL." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "content": "In terms of the Top-5 accuracy on ImageNet-LT with only " + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "content": " of the labeled data (last column of Table 2), Semi-ViM-Base outperforms Semi-ViT-Huge by " + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "inline_equation", + "content": "6.17\\%" + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "content": " and CNN-based methods, like FixMatch and PAWS, by over " + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 257, + 553, + 328 + ], + "type": "text", + "content": ", highlighting its superior feature representation, class diversity learning, and robustness against data imbalance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 328, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 553, + 424 + ], + "type": "text", + "content": "Table 3 tabulates the Top-1 Average Precision (AP) on ImageNet-LT of several fully-supervised models trained exclusively with labeled data. Semi-ViM-Base with just " + }, + { + "bbox": [ + 313, + 328, + 553, + 424 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 328, + 553, + 424 + ], + "type": "text", + "content": " of the labeled data achieves competitive accuracy. This suggests that our framework efficiently extracts meaningful representations even with significantly fewer labeled samples, making it highly suitable for real-world scenarios where labeled data may be scarce." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 460 + ], + "type": "text", + "content": "Figure 4 graphically summarizes the performance of Semi-ViM and that of other SOTA SSL models and fully-supervised models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 468, + 405, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 405, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 405, + 480 + ], + "type": "text", + "content": "5.2. Ablation study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "Several pre-training strategies, parameter updating methods, and mixup strategies are compared in Table 4, using Semi-ViM-Base. ARM pre-training consistently achieves the best performance, demonstrating its effectiveness in leveraging sequential dependencies inherent to the ViM. ARM reaches " + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "81.90\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": " (1% of the labeled data) and " + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "85.40\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": " (10% of the labeled data) when used in conjunction with LyapEMA+SSMixup, significantly outperforming MAEs and DINO-V2 across all settings. This confirms that ARM, designed specifically for sequential state-space models, is a more appropriate fit for ViM than contrastive learning (DINO-V2) or MAEs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "In terms of parameter updating methods, LyapEMA outperforms EMA in all configurations, particularly when used in conjunction with ARM. Unlike EMA, LyapEMA stabilizes updates by leveraging Lyapunov theory, preventing pseudo-label collapse and ensuring more reliable parameter evolution. This is particularly beneficial in low-labeled data scenarios, where training instability is a key challenge." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "770" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 381 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 381 + ], + "type": "table", + "html": "
CNN-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
FixMatch [38]ResNet-5025M-67.8371.7425M-52.3257.68-
PAWS [1]ResNet-5027M66.5469.2176.5527M46.3056.1361.8155.19
FreeMatch [42]ResNet-5026M55.4270.1174.5026M-55.5460.94-
SoftMatch [5]ResNet-5028M-70.4076.2028M-57.4161.02-
SimCLRv2(self-distilled) [6]ResNet-15228M76.6578.2380.9028M48.5559.3062.1067.28
SimCLRv2(distilled) [6]ResNet-5028M75.9077.4080.2028M47.7057.2360.8168.50
Co-training methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top5(1%)
Co-Training [35]CT(MLP)56M80.0381.6581.7056M58.6264.7971.2078.30
MCT [35]MCT (MLP)176M80.7582.0885.20176M63.1872.6075.3080.11
Transformer-based methodImageNet-1KImageNet-LT
BackboneParam.1%5%10%Para.m1%5%10%Top-5 (1%)
DINO [14]ViT-Small22M65.2171.4572.5322M47.8055.3059.9064.25
SemiFormer [43]ViT-S+Conv42M-72.8073.2842M50.1058.0063.5067.87
DPT [44]ViT-Huge604M80.2082.5084.06604M55.9164.7668.1575.40
Semi-ViT-Base [4]ViT-Base86M71.3077.5279.9086M53.2062.9068.4071.50
Semi-ViT-Large [4]ViT-Large307M76.8180.5082.26307M56.8066.1071.2074.70
Semi-ViT-Huge [4]ViT-Huge632M80.4783.2084.27632M58.9068.4073.5076.90
REACT [25]ViT-Huge651M79.1383.2384.74632M59.3668.7173.0877.45
Semi-ViM-TinyViM-Tiny12M76.1078.3081.0012M52.1362.9068.3071.11
Semi-ViM-SmallViM-Small27M80.3081.4083.1027M63.2472.1574.6880.02
Semi-ViM-BaseViM-Base146 M81.9083.2585.40146M66.3075.5277.4083.07
", + "image_path": "b9948231be431d4eb495d93228581730a8a8dedd222aba536ba81ce3e312f3b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 390, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 555, + 418 + ], + "type": "text", + "content": "Table 2. Accuracy (%) of the different variants of Semi-ViM (in pink) and other SOTA SSL models over 5 folds when 1%, 5% and 10% of the labeled data is used for training. The best performance of other SOTA models is highlighted in peach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 441, + 216, + 563 + ], + "blocks": [ + { + "bbox": [ + 59, + 441, + 216, + 563 + ], + "lines": [ + { + "bbox": [ + 59, + 441, + 216, + 563 + ], + "spans": [ + { + "bbox": [ + 59, + 441, + 216, + 563 + ], + "type": "image", + "image_path": "74f9f7d7480c003c75db1f08bda0e673bc73c1086c6c34d5682cc6bad7ae9c7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 577, + 143, + 586 + ], + "lines": [ + { + "bbox": [ + 131, + 577, + 143, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 577, + 143, + 586 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 441, + 382, + 563 + ], + "blocks": [ + { + "bbox": [ + 228, + 441, + 382, + 563 + ], + "lines": [ + { + "bbox": [ + 228, + 441, + 382, + 563 + ], + "spans": [ + { + "bbox": [ + 228, + 441, + 382, + 563 + ], + "type": "image", + "image_path": "13d152dcee76c156b3a779e93ab1e54c4b61496da5c35430cc8c567eee0211d9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 576, + 311, + 586 + ], + "lines": [ + { + "bbox": [ + 299, + 576, + 311, + 586 + ], + "spans": [ + { + "bbox": [ + 299, + 576, + 311, + 586 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 450, + 552, + 571 + ], + "blocks": [ + { + "bbox": [ + 395, + 450, + 552, + 571 + ], + "lines": [ + { + "bbox": [ + 395, + 450, + 552, + 571 + ], + "spans": [ + { + "bbox": [ + 395, + 450, + 552, + 571 + ], + "type": "image", + "image_path": "040ac5484445f07bd10ca732b9919d3e188c386cb24f38030b3c4a28a93b3445.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 577, + 479, + 586 + ], + "lines": [ + { + "bbox": [ + 468, + 577, + 479, + 586 + ], + "spans": [ + { + "bbox": [ + 468, + 577, + 479, + 586 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "lines": [ + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": "Figure 4. (a) Training accuracy " + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": " on ImageNet-1K of Semi-ViM-Base and other SOTA SSL methods using only " + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": " of the labeled data. (b) Top-1 accuracy " + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": " on ImageNet-LT " + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "inline_equation", + "content": "(1\\%" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": " of the labeled data) of Semi-ViM and other SOTA SSL algorithms across different model scales. (c) Top-1 Average Precision " + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 596, + 555, + 631 + ], + "type": "text", + "content": " on ImageNet-LT of Semi-ViM-Base and several fully supervised models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "text", + "content": "Regarding mixup strategies, SSMixup provides the largest performance boost, outperforming ProbMixup consistently across all pre-training methods. With ARM, LyapEAM+SSMixup boosts accuracy by " + }, + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "inline_equation", + "content": "2 - 3\\%" + }, + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "text", + "content": " compared to ProbMixup. This confirms that SSMixup is highly effective" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 651, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 712 + ], + "type": "text", + "content": "for SSL, mitigating pseudo-label noise and improving generalization in low-data settings. The combination of ARM pre-training, and LyapEMA+SSMixup achieves the highest accuracy, underscoring the importance of pre-training alignment, stable parameters updates, and structured aug-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "771" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 302, + 201 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 302, + 201 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 302, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 302, + 201 + ], + "type": "table", + "html": "
ModelLabeled dataTop-1 AP
LIFT(ViT-L/14) [37]100%82.90
MAM(ViT-B/16) [18]100%82.30
LIFT(ViT-B/16) [37]100%78.30
BALLAD (ResNet-50) [29]100%76.50
Semi-ViT-Huge [4]10%73.50
Semi-ViM-Small10%74.68
Semi-ViM-Base5%75.52
Semi-ViM-Base10%77.40
", + "image_path": "f6e402a4a8b9ef0cb506acfd93778c776668c696281363002e8e77095b0f96ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 271, + 272, + 551 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "type": "text", + "content": "Table 3. Top-1 AP (%) of fully-supervised models and the variants of Semi-ViM on ImageNet-LT. Semi-ViM-Base (in purple), using only " + }, + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "type": "text", + "content": " of the labeled data, outperforms BALLAD (ResNet" + }, + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "type": "inline_equation", + "content": "50\\times 16" + }, + { + "bbox": [ + 55, + 209, + 296, + 257 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 271, + 272, + 551 + ], + "lines": [ + { + "bbox": [ + 80, + 271, + 272, + 551 + ], + "spans": [ + { + "bbox": [ + 80, + 271, + 272, + 551 + ], + "type": "table", + "html": "
PretrainedMethod1%10%
Nonefinetune--
EMA42.1147.41
LyapEMA44.7353.26
+ProbMixup44.7849.57
+SSMixup45.2954.73
MAEfinetune63.0573.14
EMA71.3074.20
LyapEMA65.3075.65
+ProbMixup66.0975.87
+SSMixup66.0276.15
DINO-V2finetune62.0272.63
EMA63.4573.57
LyapEMA64.8475.21
+ProbMixup65.5374.04
+SSMixup65.3276.07
ARMfinetune66.5374.21
EMA72.5379.90
LyapEMA76.0278.03
+ProbMixup79.8380.52
+SSMixup81.9085.40
", + "image_path": "4ea01db5993150fd38a4d2a356a2189423d077c8d1b820aa3aa8daec505428f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 639, + 269, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 269, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 269, + 651 + ], + "type": "text", + "content": "mentation in semi-supervised models based on ViM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Table 5 compares different parameter updating methodologies, demonstrating the superior stability and generalization capability of LyapEMA in SSL. The EMA-Teacher suffers from training instability, particularly in Semi-ViT-Small and Semi-ViM-Small with only " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " of the la" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 549, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "text", + "content": "Table 4. Top-1 accuracy (%) on ImageNet-1K of Semi-ViM-Base under different pre-training strategies and parameter updating methods using " + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "text", + "content": " of the labeled data for training. \" " + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 558, + 295, + 615 + ], + "type": "text", + "content": " \" means the mixup strategy used in conjunction with LyapEMA. Best results are highlighted in bold font." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 70, + 549, + 294 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 549, + 294 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 549, + 294 + ], + "type": "table", + "html": "
ModelMethod1%10%
FreeMatch [42]FixMatch [38]--
EMA-Teacher [3]55.4274.50
LyapEMA56.7075.07
Semi-ViT-S [4]FixMatch [38]-×
EMA-Teacher [3]×75.80
LyapEMA61.3076.60
Semi-ViT-B [4]FixMatch [38]-×
EMA-Teacher [3]66.4178.15
LyapEMA68.8480.43
Semi-ViM-SFixMatch [38]××
EMA-Teacher [3]×74.01
LyapEMA80.3085.40
Semi-ViM-BFixMatch [38]×71.28
EMA-Teacher [3]70.5376.11
LyapEMA81.9085.40
", + "image_path": "ac97d9fb9485cc723e41253ccdf9961345742cafc511bb0114d5144e997a5a24.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "type": "text", + "content": "Table 5. Top-1 accuracy (%) on ImageNet-1K of FixMatch, EMA-Teacher and LyapEMA with different models using " + }, + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 303, + 555, + 361 + ], + "type": "text", + "content": " of the labeled data for training. “×” means the training failed (accuracy near 0). “-” means no result is available. Results in light blue indicate risk of training instability and collapse." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "text", + "content": "beled data, where training fails to converge. In contrast, LyapEMA effectively prevents collapse, achieving " + }, + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "inline_equation", + "content": "61.30\\%" + }, + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "inline_equation", + "content": "80.30\\%" + }, + { + "bbox": [ + 313, + 383, + 555, + 491 + ], + "type": "text", + "content": " accuracy, respectively, showing its robustness in low-labeled data scenarios. By dynamically adapting the EMA decay factor based on the student-teacher discrepancy, LyapEMA mitigates pseudo-label bias and enhances convergence. Additional ablation studies and the model's performance under our designed extreme imbalance scenarios are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we introduced Semi-ViM, a novel semi-supervised learning framework that leverages bidirectional SSMs to mitigate learning bias caused by label imbalance in classification tasks. By integrating LyapEMA, a novel stability-aware parameter update mechanism, and SS-Mixup, a hidden-state-level mixing strategy, Semi-ViM enhances both training stability and feature learning. Our extensive experiments on ImageNet-1K and ImageNet-LT demonstrate that Semi-ViM significantly outperforms CNN-based and Transformer-based SSL models, achieving superior accuracy with fewer labeled samples. Moreover, Semi-ViM requires far fewer parameters than ViT-based SSL models. These findings highlight the potential of state space models as a competitive alternative to Transformers in semi-supervised learning, particularly in addressing real-world imbalanced datasets." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "772" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8443-8452, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "text", + "content": "[3] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 323 + ], + "type": "text", + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Paolo Favaro, Manchen Wang, Davide Modolo, Rahul Bhotika, Zhuowen Tu, and Stefano Soatto. Semi-supervised vision transformers at scale. Advances in Neural Information Processing Systems, 35:25697-25710, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality trade-off in semi-supervised learning. arXiv preprint arXiv:2301.10921, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255, 2020. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 456 + ], + "type": "text", + "content": "[7] Nieves Crasto. Class imbalance in object detection: An experimental diagnosis and study of mitigation strategies. arXiv preprint arXiv:2403.07113, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 490 + ], + "type": "text", + "content": "[8] Charika De Alvis and Suranga Seneviratne. A survey of deep long-tail classification advancements. arXiv preprint arXiv:2404.15593, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Towards semi-supervised learning with nonrandom missing labels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16121-16131, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 646 + ], + "type": "text", + "content": "[11] Yue Duan, Zhen Zhao, Lei Qi, Luping Zhou, Lei Wang, and Yinghuan Shi. Roll with the punches: Expansion and shrinkage of soft label selection for semi-supervised fine-grained learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11829-11837, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "text", + "content": "[12] Marawan Elbatel, Hualiang Wang, Jixiang Chen, Hao Wang, and Xiaomeng Li. Learning unlabeled clients divergence via anchor model aggregation for federated semi-supervised learning. arXiv e-prints, pages arXiv-2407, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[14] Enrico Fini, Pietro Astolfi, Karteek Alahari, Xavier Alameda-Pineda, Julien Mairal, Moin Nabi, and Elisa Ricci. Semi-supervised learning made simple with self-supervised clustering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3187-3197, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[15] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "text", + "content": "[16] Lan-Zhe Guo, Lin-Han Jia, Jie-Jing Shao, and Yu-Feng Li. Robust semi-supervised learning in open environments. Frontiers of Computer Science, 19(8):198345, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 285 + ], + "type": "text", + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[18] Ahmet Iscen, Alireza Fathi, and Cordelia Schmid. Improving image recognition by retrieving from web-scale image-text data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19295-19304, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[19] Zhen Jiang, Lingyun Zhao, Yu Lu, Yongzhao Zhan, and Qirong Mao. A semi-supervised resampling method for class-imbalanced learning. Expert Systems with Applications, 221:119733, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[20] Suraj Kothawade, Pavan Kumar Reddy, Ganesh Ramakrishnan, and Rishabh Iyer. Basil: Balanced active semi-supervised learning for class imbalanced datasets. arXiv preprint arXiv:2203.05651, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 476 + ], + "type": "text", + "content": "[21] Alex Kurakin, Colin Raffel, David Berthelot, Ekin Dogus Cubuk, Han Zhang, Kihyuk Sohn, and Nicholas Carlini. Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring. 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 509 + ], + "type": "text", + "content": "[22] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "text", + "content": "[23] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[24] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "text", + "content": "[25] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[26] Xiao Liu, Chenxu Zhang, and Lei Zhang. Vision mamba: A comprehensive survey and taxonomy. arXiv preprint arXiv:2405.04404, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "773" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031-103063, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[28] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[29] Teli Ma, Shijie Geng, Mengmeng Wang, Jing Shao, Jiasen Lu, Hongsheng Li, Peng Gao, and Yu Qiao. A simple long-tailed recognition baseline via vision-language model. arXiv preprint arXiv:2111.14745, 2021. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "text", + "content": "[30] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[31] Khanh-Binh Nguyen. Sequencematch: revisiting the design of weak-strong augmentations for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 96-106, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[32] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[33] Sucheng Ren, Xianhang Li, Haoqin Tu, Feng Wang, Fangxun Shu, Lei Zhang, Jieru Mei, Linjie Yang, Peng Wang, Heng Wang, et al. Autoregressive pretraining with mamba in vision. arXiv preprint arXiv:2406.07537, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 446 + ], + "type": "text", + "content": "[34] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 480 + ], + "type": "text", + "content": "[35] Jay C. Rothenberger and Dimitrios I. Diochnos. Meta cotraining: Two views are better than one. In 2025 Joint Mathematics Meetings (JMM 2025). AMS. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[36] Shankar Sastry and Shankar Sastry. Lyapunov stability theory. Nonlinear systems: analysis, stability, and control, pages 182-234, 1999. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "text", + "content": "[37] Jiang-Xin Shi, Tong Wei, Zhi Zhou, Jie-Jing Shao, Xin-Yan Han, and Yu-Feng Li. Long-tail learning with foundation model: Heavy fine-tuning hurts. In *Forty-first International Conference on Machine Learning*, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[38] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[39] Andreas Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 440 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "text", + "content": "[41] Renzhen Wang, Xixi Jia, Quanziang Wang, Yichen Wu, and Deyu Meng. Imbalanced semi-supervised learning with bias adaptive classifier. arXiv preprint arXiv:2207.13856, 2022.1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[42] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[43] Zejia Weng, Xitong Yang, Ang Li, Zuxuan Wu, and Yu-Gang Jiang. Semi-supervised vision transformers. In European conference on computer vision, pages 605-620. Springer, 2022. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 271 + ], + "type": "text", + "content": "[44] Zebin You, Yong Zhong, Fan Bao, Jiacheng Sun, Chongxuan Li, and Jun Zhu. Diffusion models and semi-supervised learners benefit mutually with few labels. Advances in Neural Information Processing Systems, 36:43479-43495, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[45] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems, 34:18408-18419, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 395 + ], + "type": "text", + "content": "[46] Dingyuan Zhang, Dingkang Liang, Zhikang Zou, Jingyu Li, Xiaoqing Ye, Zhe Liu, Xiao Tan, and Xiang Bai. A simple vision transformer for weakly semi-supervised 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8373-8383, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 396, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 440 + ], + "type": "text", + "content": "[47] L Zhu, B Liao, Q Zhang, X Wang, W Liu, and X Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arxiv 2024. arXiv preprint arXiv:2401.09417. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "774" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_content_list.json b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..32cc7966b7ea6983568d47a221eba353cb24685c --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "Semi-supervised Concept Bottleneck Models", + "text_level": 1, + "bbox": [ + 272, + 130, + 723, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lijie Hu\\*,1,2, Tianhao Huang\\*,1,2,3, Huanyi Xie\\*,2, Xin Gong\\*,1,2,4, Chenyang Ren\\*,2, Zhengyu Hu\\*,6, Lu Yu\\*, Ping Ma\\*, Di Wang\\*,1,2", + "bbox": [ + 225, + 178, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Provable Responsible AI and Data Analytics (PRADA) Lab,", + "bbox": [ + 256, + 217, + 740, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ King Abdullah University of Science and Technology,", + "bbox": [ + 279, + 233, + 714, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3University of Virginia, 4University of Georgia, 5Ant Group, 6HKUST(GZ)", + "bbox": [ + 202, + 251, + 795, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lijie.hu, tianhao.huang, huanyi.xie, di.wang}@kaust.edu.sa", + "bbox": [ + 153, + 268, + 843, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept Bottleneck Models (CBMs) have garnered increasing attention due to their ability to provide concept-based explanations for black-box deep learning models while achieving high final prediction accuracy using human-like concepts. However, the training of current CBMs is heavily dependent on the precision and richness of the annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. To address these limitations, we propose a new framework called SsCBM (Semi-supervised Concept Bottleneck Model). Our SSCBM is suitable for practical situations where annotated data is scarce. By leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level, we effectively solve these issues. We propose a strategy to generate pseudo labels and an alignment loss. Experiments demonstrate that our SSCBM is both effective and efficient. With only $10\\%$ labeled data, our model's concept and task accuracy on average across four datasets is only $2.44\\%$ and $3.93\\%$ lower, respectively, compared to the best baseline in the fully supervised learning setting.", + "bbox": [ + 89, + 353, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models like ResNet [13] typically feature complex non-linear architectures, making it difficult for end-users to understand and trust their decisions. This lack of interpretability is a significant obstacle to their adoption, especially in critical fields such as healthcare [41] and finance [30], where transparency is crucial. Explain-", + "bbox": [ + 89, + 771, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "able artificial intelligence (XAI) models have been developed to meet the demand for transparency, providing insights into their behavior and internal mechanisms [16-18, 46]. Concept Bottleneck Models (CBMs) [25] are particularly notable among these XAI models for their ability to clarify the prediction process of end-to-end AI models. CBMs introduce a bottleneck layer that incorporates human-understandable concepts. During prediction, CBMs first predict concept labels from original input, then use these predicted concepts in bottleneck layer to determine the final label. This approach results in a self-explanatory decision-making process that users can comprehend.", + "bbox": [ + 511, + 321, + 906, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A major issue in original CBMs is the need for expert labeling, which is costly in practice. Some researchers address this problem through unsupervised learning. For example, [34] proposes a Label-free CBM that transforms any neural network into an interpretable CBM without requiring labeled concept data while maintaining high accuracy. Similarly, Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. However, these methods have three issues. First, those unsupervised methods heavily rely on large language models like GPT-3, which have reliability issues [26]. Second, the concepts extracted by these models lack evaluation metrics, undermining their interpretability. Third, the assumption that no concept labels are available is too stringent in practice. In reality, obtaining a small portion of concept labels is feasible and cost-effective. Therefore, we can maximize the use of this small labeled concept dataset. This is the motivation for introducing our framework, which focuses on the semi-supervised setting in CBM.", + "bbox": [ + 511, + 506, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a framework called the SS-CBM (Semi-supervised Concept Bottleneck Model). Compared to a supervised setting, semi-supervised CBMs have two main challenges. First, obtaining concept embeddings requires concept labels, so we need to generate pseudo labels to obtain these concept embeddings. To achieve this,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2110", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b72e38c5fb038035c898ebcbac7c7710816c65463a2dfd60d4193e6c0347496.jpg", + "image_caption": [ + "(a) Label Complete and Well-aligned" + ], + "image_footnote": [], + "bbox": [ + 147, + 95, + 290, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/325aa5b1fdfc33c39845046e8bf28a03b67bdedec9aa962d5cfba278dbfe15df.jpg", + "image_caption": [ + "Figure 1. (a) A sample of sparrow class with complete concept labels. (b) A sample of sparrow class with incomplete concept labels. (c) A sample of misalignment between input features and concepts resulting from existing CBM methods. Our framework simultaneously utilizes both (a) and (b) types of data and addresses the issue of (c) through an alignment loss." + ], + "image_footnote": [], + "bbox": [ + 388, + 90, + 545, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71d036ecdeafabac10ff570578e7bb05c5b079de45ffef4b7a1ec66a10413296.jpg", + "image_caption": [ + "(b) Label Incomplete" + ], + "image_footnote": [], + "bbox": [ + 547, + 95, + 612, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/475c52110745db85f13d30e043432371fc73219f227b4f083ed782ef3f6af163.jpg", + "image_caption": [ + "(c) Misaligned" + ], + "image_footnote": [], + "bbox": [ + 630, + 90, + 787, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54b430661fa18cc079ecd653e1b1ee341bed310aee4c5febc1f047f0a4e55ce5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 97, + 852, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSCBM uses a KNN-based algorithm to assign pseudo-concept labels for unlabeled data. Second, while such a simple pseudo-labeling method is effective and has acceptable classification accuracy, we also find that the concept saliency map often misaligns with the input saliency map, meaning concept predictions frequently correspond to irrelevant input features. This misalignment often arises from inaccurate concept annotations or unclear relationships between input features and concepts, which is closely related to the broader issue of annotation alignment. In fact, in the supervised setting, there is a similar misalignment issue [11]. Existing research seeks to improve alignment by connecting textual and image information [15]. However, these methods only focus on supervised settings and cannot be directly applied to our settings, as our pseudo-labels contain noise. Our framework achieves excellent performance in both concept accuracy and concept saliency alignment by leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level. To achieve this, we leverage the relevance between the input image and the concept and get other pseudo-concept labels based on these similarity scores. Finally, we align these two types of pseudo-concept labels to give the concept encoder the ability to extract useful information from features while also inheriting the ability to align concept embeddings with the input. See Figure 1 for an illustration. Our contributions can be summarized as follows:", + "bbox": [ + 91, + 330, + 483, + 737 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SSCBM, a framework designed to address both the semi-supervised annotation problem and the concept semantics alignment problem, which hold significant practical importance in real-world applications. SSCBM tackles them in an elegant and efficient manner. To the best of our knowledge, we are the first to tackle these two problems within a single framework, elucidating the behavior of CBMs through semi-supervised alignment.", + "- Comprehensive experiments demonstrate the superiority of our SSCBM in annotation and concept-saliency align" + ], + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment, indicating its efficiency and effectiveness. Using only $10\\%$ labeled data, our model achieves concept and task accuracy on average across the four datasets that are merely $2.44\\%$ and $3.93\\%$ lower, respectively, than the best baseline under fully supervised learning settings.", + "bbox": [ + 524, + 330, + 906, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 421, + 653, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Bottleneck Models. Concept Bottleneck Model (CBM) [25] is an innovative deep-learning approach for image classification and visual reasoning. By introducing a concept bottleneck layer into deep neural networks, CBMs enhance model generalization and interpretability by learning specific concepts. However, CBMs face two primary challenges: their performance often lags behind that of models without the bottleneck layer due to incomplete information extraction, and they rely heavily on laborious dataset annotation. Researchers have explored various solutions to these challenges. [5] extends CBMs into interactive prediction settings by introducing an interaction policy to determine which concepts to label, thereby improving final predictions. [34] addresses CBM limitations by proposing a Label-free CBM, which transforms any neural network into an interpretable CBM without requiring labeled concept data, maintaining high accuracy. Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. Related work in the image domain includes studies [12, 19, 23, 24, 29, 37, 38]. In the graph concept field, [31] provides a global interpretation of graph neural networks (GNNs) by mapping graphs into a concept space through clustering and offering a human-in-the-loop evaluation. [3, 44] extend this approach by incorporating both global and local explanations. For local explanations, they define a concept set, with each neuron represented as a vector with Boolean values indicating concept activation. However, existing works rarely consider semi-supervised settings, which are practical in real-world", + "bbox": [ + 511, + 446, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2111", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "applications. Our framework addresses these issues effectively.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised Learning. Semi-supervised learning (SSL) combines the two main tasks of machine learning: supervised learning and unsupervised learning [43]. It is typically applied in scenarios where labeled data is scarce. Examples include computer-aided diagnosis [6, 49], medical image analysis [8, 10, 20], and drug discovery [7]. In these cases, collecting detailed annotated data by experts requires considerable time and effort. However, under the assumption of data distribution, unlabeled data can also assist in building better classifiers [43]. SSL, also known as self-labeling or self-teaching in its earliest forms, involves the model iteratively labeling a portion of unlabeled data and adding it to the training set for the next round of training [35]. The expectation-maximization (EM) algorithm proposed by [32] uses both labeled and unlabeled data to produce maximum likelihood estimates of parameters. [27] and [40] focus on consistency regularization. II-model [27] combines both supervised cross-entropy loss and unsupervised consistency loss while perturbing the model and data based on the consistency constraint assumption. A temporal ensembling model integrates predictions from models at various time points. Mean teacher [40] addresses the slow updating issue of the temporal ensembling model on large datasets by averaging model weights instead of predicting labels. MixMatch [4] unifies and refines the previous approaches of consistency regularization, entropy minimization, and traditional regularization into a single loss function, achieving excellent results. Pseudo labeling, as an effective tool for reducing the entropy of unlabeled data [28], has been increasingly attracting the attention of researchers in the field of semi-supervised learning. [2] proposes that directly using the model's predictions as pseudo-labels can achieve good results. FixMatch [39] predicts and retains the model, generating high-confidence pseudo-labels. [36] continuously adjusts the teacher based on feedback from the student, thereby generating better pseudo-labels. While there has been a plethora of work in the semi-supervised learning field, the focus on semi-supervised concept bottleneck models remains largely unexplored. Our work focuses on this new area.", + "bbox": [ + 91, + 119, + 483, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 739, + 230, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Bottleneck Models [25]. We consider a classification task with a concept set denoted as $\\mathcal{C} = \\{p_1,\\dots ,p_k\\}$ with each $p_i$ being a concept given by experts or LLMs, and a training dataset represented $\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}$ . Here, for $i\\in [N]$ , $x^{(i)}\\in \\mathcal{X}\\subseteq \\mathbb{R}^d$ represents the feature vector (e.g., an image's pixels), $y^{(i)}\\in \\mathcal{Y}\\subseteq \\mathbb{R}^l$ denotes the label ( $l$ is the number of classes), $c^{(i)} = (c_i^1,\\dots ,c_i^k)\\in \\mathbb{R}^k$ represents the concept vector (a binary vector of length $k$ , where each value indicates whether the input $x^{(i)}$ contains", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that concept). In CBMs, the goal is to learn two representations: one called a concept encoder that transforms the input space into the concept space, denoted as $g: \\mathbb{R}^d \\to \\mathbb{R}^k$ , and another called label predictor that maps the concept space to the downstream prediction space, denoted as $f: \\mathbb{R}^k \\to \\mathbb{R}^l$ . Usually, the map $f$ is linear. For any input $x$ , we aim to ensure that its predicted concept vector $\\hat{c} = g(x)$ and prediction $\\hat{y} = f(g(x))$ are close to their underlying counterparts, thus capturing the essence of the original CBMs.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Embedding Models [9]. As the original CBM is based solely on concept features to determine the predictions of the model, compared to canonical deep neural networks, it will degrade prediction performance. To further improve the performance of CBMs, CEM is developed by [9]. It achieves this by learning interpretable high-dimensional concept representations (i.e., concept embeddings), thus maintaining high task accuracy while obtaining concept representations that contain meaningful semantic information. For CEMs, we use the same setting as that of [9, 21]. For each input $x$ , the concept encoder learns $k$ concept representations $\\hat{c}_1, \\hat{c}_2, \\dots, \\hat{c}_k$ , each corresponding to one of the $k$ ground truth concepts in the training dataset. In CEMs, each concept $c_i$ is represented using two embeddings $\\hat{c}_i^+, \\hat{c}_i^- \\in \\mathbb{R}^m$ , each with specific semantics, i.e., the concept is True (activate state) and concept is False (negative state), where hyper-parameter $m$ is the embedding dimension. We use a DNN $\\psi(x)$ to learn a latent representation $h \\in \\mathbb{R}^{n_h}$ , which is used as input to the CEM embedding generator, where $n_h$ is the dimension of the latent representation. The CEM embedding generator $\\phi$ feeds $h$ into two fully connected concept-specific layers to learn two concept embeddings in $\\mathbb{R}^m$ .", + "bbox": [ + 511, + 227, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} _ {i} = \\phi_ {i} (\\boldsymbol {h}) = a (W _ {i} \\boldsymbol {h} + \\boldsymbol {b} _ {i}).\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 587, + 810, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then we use a differential scoring function $s: \\mathbb{R}^{2m} \\to [0,1]$ , to achieve the alignment of concept embeddings $\\hat{\\pmb{c}}_i^+$ , $\\hat{\\pmb{c}}_i^-$ and ground-truth concepts $c_i$ . It is trained to predict the probability $\\hat{p}_i := s\\left(\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right) = \\sigma\\left(W_s\\left[\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right] + \\pmb{b}_s\\right)$ that the concept $c_i$ is active in the embedding space. We get the final concept embedding $\\hat{\\pmb{c}}_i$ , as follows:", + "bbox": [ + 511, + 614, + 906, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} _ {i} := \\hat {p} _ {i} \\hat {\\boldsymbol {c}} _ {i} ^ {+} + (1 - \\hat {p} _ {i}) \\hat {\\boldsymbol {c}} _ {i} ^ {-}.\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 750, + 799, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At this point, we understand that we can obtain high-quality concept embeddings rich in semantics through CEMs. In the subsequent section 4, we will effectively utilize these representations of concepts and further optimize their interpretability through our proposed framework SSCBM.", + "bbox": [ + 511, + 779, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised Setting. Now, we consider the setting of semi-supervised learning for concept bottleneck models. As mentioned earlier, a typical training dataset for CBMs", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2112", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can be represented as $\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}$ , where $x^{(i)}\\in \\mathcal{X}$ represents the input feature. However, in semisupervised learning tasks, the set of feature vectors typically consists of two parts, $\\mathcal{X} = \\{\\mathcal{X}_L,\\mathcal{X}_U\\}$ , where $\\mathcal{X}_L$ represents a small subset of labeled data and $\\mathcal{X}_U$ represents the remaining unlabeled data. Here, the unlabeled data refers to data where the class label $y$ is known, while the concept label $c$ is unknown. Generally we assume $|\\mathcal{X}_L|\\ll |\\mathcal{X}_U|$ . We assume that $x^{(j)}\\in \\mathcal{X}_L$ is labeled with a concept vector $c^{(j)}$ and a label $y^{(j)}$ , and for any $x^{(i)}\\in \\mathcal{X}$ , there only exists a corresponding label $y^{(i)}\\in \\mathcal{Y}$ .", + "bbox": [ + 89, + 89, + 483, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under these settings, given a training dataset $\\mathcal{D} = \\mathcal{D}_L \\cup \\mathcal{D}_U$ that includes both labeled and unlabeled data, the goal is to train a CBM using both the labeled data $\\mathcal{D}_L$ and unlabeled data $\\mathcal{D}_U$ . This aims to get better mappings $g: \\mathbb{R}^d \\to \\mathbb{R}^k$ and $f: \\mathbb{R}^k \\to \\mathbb{R}^l$ than those trained by using only labeled data, ultimately achieving higher task accuracy and its corresponding concept-based explanation.", + "bbox": [ + 89, + 257, + 483, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Semi-supervised Concept Bottleneck Models", + "text_level": 1, + "bbox": [ + 89, + 376, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we will elaborate on the details of the proposed SSCBM framework, which is shown in Figure 2. SSCBM follows the main idea of CEM. Specifically, to learn a good concept encoder, we use different processing methods for labeled and unlabeled data. The labeled data first passes through a feature extractor $\\psi$ to be transformed into a latent representation $h$ , which then enters the concept embedding extractor to obtain the concept embeddings and the predicted concept vector $\\hat{c}$ for the labeled data, which is compared to the ground truth concept to compute the concept loss. Moreover, the label predictor predicts $\\hat{y}$ based on $\\hat{c}$ , and calculates the task loss.", + "bbox": [ + 89, + 401, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For unlabeled data, we first extract image features $V$ from the input using an image encoder. Then, we use the KNN algorithm to assign a pseudo-label $\\hat{c}_{img}$ to each unlabeled data, which has been experimentally proven to be simple and effective. In the second step, we compute a heatmap between concept embeddings and the input. After applying a threshold, we obtain the predicted alignment label $\\hat{c}_{align}$ . Finally, we compute the alignment loss between $\\hat{c}_{img}$ and $\\hat{c}_{align}$ . During each training epoch, we simultaneously compute these losses and update the model parameters based on the gradients.", + "bbox": [ + 89, + 582, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Label Anchor: Concept Embedding Encoder", + "text_level": 1, + "bbox": [ + 89, + 757, + 473, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept Embeddings. As described in Section 3, we obtain high-dimensional concept representations with meaningful semantics based on CEMs. Thus, our concept encoder should extract useful information from both labeled and unlabeled data.", + "bbox": [ + 89, + 780, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For labeled training data $\\mathcal{D}_L = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{\\left|\\mathcal{D}_L\\right|}$ we follow the original CEM [9], i.e., using a backbone network (e.g., ResNet50) to extract features $h =$", + "bbox": [ + 89, + 853, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\{\\psi(x^{(i)})\\}_{i=1}^{\\left|\\mathcal{D}_L\\right|}$ . Then, for each feature, it passes through an embedding generator to obtain concept embeddings $\\hat{c}_i \\in \\mathbb{R}^{m \\times k}$ for $i \\in [k]$ . After passing through fully connected layers and activation layers, we obtain the predicted binary concept vector $\\hat{c} \\in \\mathbb{R}^k$ for the labeled data. The specific process can be represented by the following expression:", + "bbox": [ + 511, + 89, + 905, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} ^ {(j)} = \\psi (x ^ {(j)}), \\quad j = 1, \\ldots , | \\mathcal {D} _ {L} |,\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 189, + 830, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {c}} _ {i} ^ {(j)} = \\sigma (\\phi (\\boldsymbol {h} ^ {(j)})), \\quad i = 1, \\dots , k, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 217, + 890, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi, \\phi$ and $\\sigma$ represent the backbone, embedding generator, and activation function, respectively.", + "bbox": [ + 511, + 242, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the interpretability of concept embeddings, we calculate the concept loss utilizing binary cross-entropy to optimize the accuracy of concept predictions by computing $\\mathcal{L}_c$ based on predicted binary concept vector $\\hat{e}$ and ground truth concept labels $c$ :", + "bbox": [ + 511, + 272, + 905, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = B C E (\\hat {\\boldsymbol {c}}, \\boldsymbol {c}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 358, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $BCE$ is the binary cross entropy loss.", + "bbox": [ + 511, + 383, + 812, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task Loss. Since our ultimate task is classification, we also need to incorporate the task loss for the final prediction performance. After obtaining the predicted concept $\\hat{c}$ , we use a label predictor to predict the final class $\\hat{\\pmb{y}}$ . Then we can define the task loss function using the categorical cross-entropy loss to train our classification model as follows:", + "bbox": [ + 511, + 398, + 905, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a s k}} = C E (\\hat {\\boldsymbol {y}}, \\boldsymbol {y}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 503, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that for the unlabeled data, we can also calculate the task loss since their class labels are known, to make full use of the data.", + "bbox": [ + 511, + 526, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Unlabel Alignment: Image-Textual Semantics Alignment", + "text_level": 1, + "bbox": [ + 511, + 580, + 905, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo Labeling. Unlike supervised data, CEM cannot directly extract useful information from unlabeled data, as the concept encoder is a supervised training architecture. Thus, in practical situations lacking labeled data, one direct approach is to get high-quality pseudo concept labels to train the model. In the following, we will introduce the method we use to obtain pseudo-concept labels.", + "bbox": [ + 511, + 617, + 905, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we can naturally think of measuring the similarity between images by calculating the distance in the cosine space. Based on this idea, we can assign pseudo labels to unlabeled data by finding labeled data with similar image features to them. Specifically, for each unlabeled training data $x \\in \\mathcal{D}_U = \\{(x^{(i)},y^{(i)})\\}_{i=1}^{|D_U|}$ , we calculate its cosine distance with all labeled data points $x^{(j)} \\in \\mathcal{D}_L = \\{(x^{(j)},y^{(j)},c^{(j)})\\}_{j=1}^{|D_L|}$ and select the $k$ samples with the smallest distances:", + "bbox": [ + 511, + 723, + 905, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} (x, x ^ {(j)}) = 1 - \\frac {\\Omega (x) \\cdot \\Omega (x ^ {(j)})}{\\| \\Omega (x) \\| _ {2} \\cdot \\| \\Omega (x ^ {(j)}) \\| _ {2}}, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 869, + 923, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2113", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6055d52a8973357c3adf755aef92bc10c105af9ccd0b6cc0e429a6d48343a92.jpg", + "image_caption": [ + "Figure 2. Overall framework of our proposed SSCBM." + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 864, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega$ is a visual encoder. Then, we normalize the reciprocal of the cosine distance between the nearest $k$ data points and $x$ as weights. We use these weights to compute a weighted average of the concept labels of these $k$ data points, obtaining the pseudo concept label for $x$ . In this way, we can obtain pseudo concept labels $\\hat{c}_{img}$ for all $x \\in \\mathcal{D}_U$ .", + "bbox": [ + 89, + 469, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we find that directly feeding pseudo-concept labels generated by KNN to CEM has satisfactory performance. However, this simple labeling method can lead to alignment issues in the concept embeddings learned by the CEM's concept encoder. Specifically, the predicted concepts $\\hat{c}$ could not have a relation to the corresponding features $V$ in the image, which hinders the effectiveness of CEM as a reliable interpretability tool. Moreover, due to misalignment, this will also degrade the prediction performance of the concept encoder. In the following, we aim to address such a misalignment issue.", + "bbox": [ + 89, + 579, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generating Concept Heatmaps. Our above pseudo concept labels via KNN have already contained useful information for prediction. Thus, our goal is to provide these labels with more information about their relation to the corresponding features. To achieve this, we first get another pseudo-concept label via these relations, which are calculated by the similarity between the concept embedding and the input image, namely concept heatmaps. Specifically, given an image $x$ , we first have its feature map $V \\in \\mathbb{R}^{H \\times W \\times m}$ for each concept extracted by $V = \\Omega(x)$ ,", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the height and width of the feature map.", + "bbox": [ + 511, + 469, + 903, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $V$ and the $i$ -th concept embedding $\\pmb{c}_i$ , we can obtain a heatmap $\\mathcal{H}_i$ , i.e., a similarity matrix that measures the similarity between the concept and the image can be obtained by computing their cosine distance:", + "bbox": [ + 511, + 484, + 903, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {p, q, i} = \\frac {\\hat {\\boldsymbol {c}} _ {i} ^ {m} \\cdot V _ {p , q}}{\\| \\hat {\\boldsymbol {c}} _ {i} ^ {m} \\| \\cdot \\| V _ {p , q} \\|}, \\quad p = 1, \\dots , H, \\quad q = 1, \\dots , W\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 554, + 908, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p, q$ are the $p$ -th and $q$ -th positions in the heatmaps, and $\\mathcal{H}_{p,q,i}$ represents a local similarity score between $V$ and $\\hat{c}_i^m$ . Intuitively, $\\mathcal{H}_i$ represents the relation of each part of the image with the $i$ -the concept. Then, we derive heatmaps for all concepts, denoted as $\\{\\mathcal{H}_1, \\mathcal{H}_2, \\dots, \\mathcal{H}_k\\}$ .", + "bbox": [ + 511, + 599, + 905, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calculating Concept Scores and Concept Labels. As average pooling performs better in downstream classification tasks [45], we apply average pooling to the heatmaps to deduce the connection between the image and concepts: $s_i = \\frac{1}{P \\cdot Q} \\sum_{p=1}^{P} \\sum_{q=1}^{Q} \\mathcal{H}_{p,q,i}$ . Intuitively, $s_i$ is the refined similarity score between the image and concept $\\hat{c}_i^m$ . Thus, a concept vector $s$ can be obtained, representing the similarity between an image input $x$ and the set of concepts: $s = (s_1, \\ldots, s_k)^\\top$ . $s$ can be considered as a soft concept label which is got by similarity. Next, we have to transform it into a hard concept label $\\hat{c}_{\\text{align}}$ . We determine the presence of a concept attribute in an image based on a threshold value derived from an experiment. If the value $s_i$ exceeds this threshold, we consider the image to possess that specific concept attribute and set the concept label to be True. We", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2114", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can obtain predicted concept labels for all unlabeled data.", + "bbox": [ + 89, + 90, + 470, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Alignment of Image. Based on our earlier discussions, on the one hand, the concept encoder should learn information from $\\hat{c}_{img}$ . On the other hand, it should also get concept embeddings which can get good similarity-based concept labels $\\hat{c}_{\\text {align }}$ for alignment with the input image. Thus, we need a further alignment loss to achieve these two goals. Specifically, we compute the alignment loss as follows:", + "bbox": [ + 89, + 106, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = B C E \\left(\\hat {c} _ {\\text {i m g}}, \\hat {c} _ {\\text {a l i g n}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 224, + 482, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In fact, during this alignment process, the CBM backbone is taught a reasonable latent space structure by the visual encoder, and thus we can regard it as an implicit Student-Teacher setup framework.", + "bbox": [ + 89, + 253, + 483, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Final Objective", + "text_level": 1, + "bbox": [ + 89, + 324, + 245, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will discuss how we derive the process of network optimization. First, we have the concept loss $\\mathcal{L}_c$ in (1) to improve the interpretability of the concept embeddings. Also, since the concept embeddings are entered into the label predictor to output the final prediction, we also have a task loss between the predictions given by the concept bottleneck and ground truth, which is shown in (2). In the context of binary classification tasks, we employ binary cross-entropy (BCE) as our loss function. For multi-class classification tasks, we use cross-entropy as the measure. Finally, to align the images with the concept labels, we computed the loss of alignment in (3). Formally, the overall loss function of our approach can be formulated as:", + "bbox": [ + 88, + 347, + 483, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {a l i g n}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 556, + 482, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2$ are hyperparameters for the trade-off between interpretability and accuracy.", + "bbox": [ + 89, + 585, + 482, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 630, + 223, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will conduct experimental studies on the performance of our framework. Specifically, we evaluate the utility of our method according to the concept and prediction of the class label. We also design interpretability evaluation and test-time intervention to verify the interpretability and alignment performance of our method. Finally, we give the ablation study to verify why we need the alignment loss. Details and additional results are in Appendix A.", + "bbox": [ + 89, + 655, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 801, + 297, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our methods on four real-world image tasks: CUB[14], AwA2[33], WBCatt[42] and 7-point[22]. See Appendix A.1 for a detailed introduction.", + "bbox": [ + 89, + 824, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our SSCBM with Concept Bottle-neck Model (CBM) and Concept Embedding Model (CEM)", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/18326956d5db39d543f396bdfde12169d7d27915ec56353f965ed9394e907760.jpg", + "table_caption": [ + "Table 1. Concept and task accuracy results of SSCBM at a labeled ratio of 0.1, compared to baselines in the fully supervised setting." + ], + "table_footnote": [], + "table_body": "
MethodCUBAwA2WBCatt7-point
ConceptTaskConceptTaskConceptTaskConceptTask
CBM93.99%67.33%96.48%88.71%94.18%99.71%74.34%75.44%
CEM96.39%79.82%95.91%87.00%95.33%99.71%77.15%75.85%
SSCBM90.88%67.67%96.48%89.77%93.98%99.68%73.67%70.09%
", + "bbox": [ + 516, + 127, + 903, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mentioned in Section 3. Since those baselines do not inherently include settings for semi-supervised learning, to ensure fairness in the evaluation, we use the same pseudolabeling approach as SSCBM. We utilize the KNN algorithm to annotate unlabeled data based on labeled data, which is mentioned in Section 4.2. For CBM [25], we adopt the same setting and architecture as in the original CBM. And for CEM, we follow the same setting as in [9]. We also compare label-free CBM [34], which is an unsupervised CBM method.", + "bbox": [ + 511, + 224, + 906, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the utility, we consider the accuracy for the prediction of class and concept labels. Specifically, concept accuracy measures the model's prediction accuracy for concepts: $\\mathcal{C}_{acc} = \\frac{1}{N}\\cdot \\frac{1}{k}\\sum_{i = 1}^{N}\\sum_{j = 1}^{k}\\mathbb{I}(\\hat{c}_j^{(i)} = c_j^{(i)})$ . Task accuracy measures the model's performance in predicting downstream task classes: $\\mathcal{A}_{acc} = \\frac{1}{N}\\sum_{i = 1}^{N}\\mathbb{I}(\\hat{\\pmb{y}}^{(i)} = \\pmb {y}^{(i)})$", + "bbox": [ + 511, + 377, + 905, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the interpretability, besides the concept accuracy, similar to previous work [9, 25], we show visualization results. Moreover, we evaluate the performance of the test-time intervention.", + "bbox": [ + 511, + 487, + 905, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. All experiments are conducted on a Tesla V100S PCIe 32 GB GPU and an Intel Xeon Processor CPU. See Appendix A.2 for more details.", + "bbox": [ + 511, + 547, + 905, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Results on Utility", + "text_level": 1, + "bbox": [ + 511, + 604, + 769, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively evaluate the model's performance in the semi-supervised learning setting, we simulate real-world application scenarios. We consider labeling K samples for each class in the dataset, with an extreme case where $\\mathrm{K} = 1$ . Clearly, $\\mathrm{K} =$ total data size $\\times$ labeled ratio $\\div$ number of classes. We also conduct experiments with labeled ratios of 0.05, 0.1, 0.15, and 0.2. We also compare the performance of SSCBM with the baseline under the fully supervised setting. The experimental results are shown in Table 1 and 2. Additionally, we also explored the impact of different backbones on model performance, which can be found in the Appendix A.3.", + "bbox": [ + 511, + 628, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the performance comparison of SSCBM at a labeled ratio of 0.1 with CEM and CBM under the fully supervised setting. It can be observed that, although CEM performs the best on CUB, WBCatt and 7-point, SSCBM is only slightly inferior to it. On the AwA2 dataset, its performance even surpasses that of CEM under the fully super", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2115", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5293f30d894b5200af0e407aa5b25326e75fae900e0d4f154395a5a1d6fda5f7.jpg", + "table_caption": [ + "Table 2. Results of concept and task accuracy for different datasets with different portions of labeled data. Since the label-free CBM [34] and LaBo [47] do not use the concept set included in the dataset but instead rely on GPT-3 to generate concepts based on class names, we do not report the concept accuracy here." + ], + "table_footnote": [], + "table_body": "
DatasetLabeled RatioCBM+SSLCEM+SSLLabel-free CBMLaBoSSCBM
ConceptTaskConceptTaskConceptTaskConceptTaskConceptTask
CUBK=183.11%5.51%82.36%59.35%-74.31%-66.38%88.99%66.72%
0.05 (K=2)84.51%8.35%83.72%62.20%-74.31%-66.38%90.04%67.43%
0.1 (K=3)84.96%9.84%84.03%63.12%-74.31%-66.38%90.88%67.67%
0.15 (K=4)85.47%9.96%84.30%64.14%-74.31%-66.38%91.47%68.36%
0.2 (K=5)86.67%16.43%86.83%67.64%-74.31%-66.38%92.09%70.07%
AWA2K=171.32%90.87%65.55%84.44%-73.67%-93.73%85.05%89.02%
0.05 (K=30)77.23%91.11%68.31%85.01%-73.67%-93.73%95.42%89.37%
0.1 (K=60)80.45%91.03%69.96%85.48%-73.67%-93.73%96.48%89.77%
0.15 (K=90)83.46%91.00%72.14%86.61%-73.67%-93.73%96.81%90.54%
0.2 (K=120)85.76%91.12%72.14%86.61%-73.67%-93.73%96.81%90.44%
WBCattK=179.06%99.39%70.27%98.64%-41.92%-92.26%91.48%99.13%
0.05 (K=62)81.08%99.48%73.82%99.52%-41.92%-92.26%93.53%99.61%
0.1 (K=124)85.48%99.32%72.25%99.29%-41.92%-92.26%93.98%99.68%
0.15 (K=186)85.39%99.68%72.68%99.58%-41.92%-92.26%94.42%99.71%
0.2 (K=247)87.07%99.74%74.14%99.52%-41.92%-92.26%94.42%99.71%
7-pointK=159.91%55.95%62.78%66.09%-55.44%-64.97%66.58%66.84%
0.05 (K=5)65.36%57.47%67.85%67.09%-55.44%-64.97%70.98%68.77%
0.1 (K=9)68.82%55.70%72.23%66.33%-55.44%-64.97%73.67%70.09%
0.15 (K=13)66.14%59.75%66.54%67.09%-55.44%-64.97%73.94%72.56%
0.2 (K=17)70.29%60.44%73.04%66.84%-55.44%-64.97%76.52%74.56%
", + "bbox": [ + 174, + 141, + 823, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vised setting, which sufficiently demonstrates the superiority of our method.", + "bbox": [ + 89, + 428, + 482, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, it can be observed that as the proportion of labeled data gradually increases, the concept and class prediction accuracy of all models improve. SSCBM outperforms all baselines. On the CUB dataset, when only one labeled sample per class is available, SSCBM achieves $88.99\\%$ concept accuracy and $66.72\\%$ class accuracy, which are $6.63\\%$ and $7.37\\%$ higher, respectively, than the performance achieved by CEM in a semi-supervised learning setting. This demonstrates the effectiveness of the Image-Textual Semantics Alignment module. The 7-point dataset is the most challenging, and on this dataset, SSCBM still achieves the best performance across all settings. As the amount of labeled data increases, the concept prediction accuracy improves from $66.58\\%$ to $76.52\\%$ . We also note that on WBCatt and 7-point, the performance of Label-free CBM is poor. This is mainly because both datasets contain only five classes, which prevents Label-free CBM from fully leveraging the world knowledge of LLMs to generate an effective concept set, resulting in poor performance. This further demonstrates that semi-supervised learning scenarios are more realistic and practical for real-world datasets.", + "bbox": [ + 91, + 460, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Interpretability Evaluation", + "text_level": 1, + "bbox": [ + 89, + 787, + 334, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the concept accuracy, note that in SSCBM, we also have a pseudo label revived by the alignment between the concept embedding and the input saliency map. We use the alignment loss to inherit this alignment. Thus, we will evaluate the alignment performance here to show the faithfulness of the interpretability given by SSCBM. We measure", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our alignment performance by comparing the correctness of the concept saliency map with the concept semantics in Figure 3. See Appendix A.2 for detailed experimental procedures for generating saliency maps. Results show that our concept saliency map matches the concept semantics, indicating the effectiveness of our alignment loss. In Appendix C, we also provide our additional interpretability evaluation in Figure 6-18.", + "bbox": [ + 511, + 428, + 906, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Test-time Intervention", + "text_level": 1, + "bbox": [ + 511, + 560, + 720, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Test-time intervention enables human users to interact with the model at inference time. We conduct this experiment by correcting the ratio $10\\%$ to $100\\%$ of the concept labels in the concept predictor. We expect that the model performance will steadily increase along with the ratio of concept intervention, indicating that the model learned such correct label information and automatically corrected other labels.", + "bbox": [ + 511, + 583, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Figure 4 demonstrate the robustness of our model and an increasing trend to learn the information about the concept, indicating our interpretability and model prediction performance. During the test phase, the SSCBM model generates probability scores for each concept (e.g., has_wing_color :: brown). Prior to intervention, we employ the COOP algorithm (following CEM) to identify the most critical concepts exhibiting both high predictive uncertainty and task importance. For these selected concepts, we replace the model's predicted values with their corresponding ground truth annotations. This rectified concept vector is subsequently fed into the classification head to produce the final task prediction. Finally, we successfully cause the model to predict the Great Crested Flycatcher instead of the", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2116", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7da7de8127c7047a23aa8346f9d1cd821ee187972096c1a707cf2055e561d899.jpg", + "image_caption": [ + "(a) Original Image" + ], + "image_footnote": [], + "bbox": [ + 166, + 88, + 325, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/517c7b3f9b502cc51979ed2a51369b5e65804b7e6cd043eca99295e0cb1c70c5.jpg", + "image_caption": [ + "(b) Tail Pattern: Multi-colored", + "Figure 3. The concept saliency map for the CUB dataset (savannah sparrow) demonstrates that our proposed SSCBM achieves meaningful alignment between the ground truth concepts and the input image features. The first image on the left is the original input image. The three images on the right show the aligned regions for different concepts using SSCBM." + ], + "image_footnote": [], + "bbox": [ + 336, + 88, + 496, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2040c5184d8c90b24989582fa87a346ddb8e7a58011c416fdc7e64d5ca0dddc.jpg", + "image_caption": [ + "(c) Underparts Color: Brown" + ], + "image_footnote": [], + "bbox": [ + 508, + 88, + 668, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4af6004c8beb3e17a59de06ead3d8a56a8919aadb615af41df688166d676b69c.jpg", + "image_caption": [ + "(d) Wing Color: Brown" + ], + "image_footnote": [], + "bbox": [ + 676, + 88, + 836, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aa34b9d7a97a501ccc0a0a8a04eb3af6137337b6aabf3b19604f4fa466d22b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 306, + 380, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44fcb677be76cff0bb1c49531ccfc493519c5b768fa862338613e9ba8d7d30ab.jpg", + "image_caption": [ + "Figure 4. Left: Performance with different ratios of intervened concepts on CUB dataset. Right: An example of successful intervention." + ], + "image_footnote": [], + "bbox": [ + 398, + 306, + 537, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4eec13b2b797ac0efabcf68556e47c5d7e5bd320be4078f670030c3445ee637d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 318, + 759, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Swainson Warbler. More results are in Appendix B.", + "bbox": [ + 89, + 477, + 431, + 492 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b22c82ffd1aff0be97dc782e68df4a4f75dc8eacf69b1d125410d35402f7f680.jpg", + "table_caption": [ + "Table 3. Results of ablation study." + ], + "table_footnote": [], + "table_body": "
Ratiow/o imgw/o alignfull model
ConceptTaskConceptTaskConceptTask
CUB0.0001 (K=1)87.77%59.89%80.87%65.15%88.99%66.72%
0.05 (K=2)90.01%65.98%80.60%65.67%90.04%67.43%
0.1 (K=3)91.02%67.28%80.60%64.69%90.88%67.67%
0.15 (K=4)90.60%67.97%80.65%65.43%91.47%68.36%
0.2 (K=5)91.94%68.99%80.73%64.45%92.09%70.07%
WBCatt0.0001 (K=1)88.44%99.68%57.80%99.51%91.48%99.13%
0.05 (K=62)92.17%99.29%57.93%99.51%93.53%99.61%
0.1 (K=124)92.91%99.52%60.49%99.64%93.98%99.68%
0.15 (K=186)93.55%99.58%59.13%99.65%94.42%99.71%
0.2 (K=247)94.19%99.61%60.65%99.70%94.42%99.71%
7-point0.0001 (K=1)62.31%64.30%55.07%66.33%66.58%66.84%
0.05 (K=5)69.42%65.06%56.75%68.10%70.98%68.77%
0.1 (K=9)70.06%67.34%56.07%65.82%73.67%70.09%
0.15 (K=13)73.39%68.10%56.23%65.82%73.94%72.56%
0.2 (K=17)74.44%68.61%56.11%65.57%76.52%74.56%
", + "bbox": [ + 94, + 515, + 483, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 742, + 243, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then give a finer study on the two kinds of pseudo-labels to demonstrate that each one is indispensable in bolstering the efficacy of SSCBM. Specifically, based on our method in Section 4.2, we have three types of pseudo labels available: pseudo concept label $\\hat{c}_{img}$ , $\\hat{c}_{align}$ , and label $\\hat{c}$ predicted by the concept embedding. We first remove $\\hat{c}_{img}$ and calculate our alignment loss using $\\hat{c}_{align}$ and $\\hat{c}$ ; conversely, we remove $\\hat{c}_{align}$ and calculate the alignment loss using $\\hat{c}_{img}$ and $\\hat{c}$ .", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the results presented in Table 3, it is evident that removing the $\\hat{c}_{img}$ component significantly degrades the performance of SSCBM at both the concept level and class level. This indicates that the pseudo-concept labels via KNN contain necessary information about the ground truth. The concept encoder needs to extract information from such labels to get better performance. This situation is similar when removing the $\\hat{c}_{align}$ component, indicating that aligning the concept embedding and the input saliency map can further extract useful information from the input image and, thus, is beneficial to improve the performance. Our observations underscore the high degree of joint effectiveness of two kinds of pseudo-concept labels within our objective function, collectively contributing to the enhancement of model prediction and concept label prediction.", + "bbox": [ + 511, + 477, + 906, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The training of current CBMs heavily relies on the accuracy and richness of annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. In this problem, we propose SSCBM, a strategy to generate pseudo labels and an alignment loss to solve these two problems. Results show our effectiveness.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2117", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the funding BAS/1/1689-01-01, URF/1/4663-01-01, REI/1/5232-01-01, REI/1/5332-01-01, and URF/1/5508-01-01 from KAUST, and funding from KAUST - Center of Excellence for Generative AI, under award number 5940.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrea Acevedo, Santiago Alférez, Anna Merino, Laura Puigvi, and José Rodellar. Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine, 180:105020, 2019. 11", + "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 3", + "[3] Pietro Barbiero, Gabriele Ciravegna, Francesco Giannini, Mateo Espinosa Zarlenga, Lucie Charlotte Magister, Alberto Tonda, Pietro Lio, Frédéric Precioso, Mateja Jamnik, and Giuseppe Marra. Interpretable neural-symbolic concept reasoning. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1801-1825. PMLR, 2023. 2", + "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[5] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2", + "[6] Asma Chebli, Akila Djebar, and Hayet Farida Marouani. Semi-supervised learning for medical application: A survey. In 2018 International Conference on Applied Smart Systems (ICASS), pages 1-9, 2018. 3", + "[7] Xing Chen, Biao Ren, Ming Chen, Quanxin Wang, Lixin Zhang, and Guiying Yan. Nllss: predicting synergistic drug combinations based on semi-supervised learning. PLoS computational biology, 12(7):e1004975, 2016. 3", + "[8] Veronika Cheplygina, Marleen de Bruijnne, and Josien P.W. Pluim. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis, 54:280-296, 2019. 3", + "[9] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 3, 4, 6, 11", + "[10] Roman Filipovych and Christos Davatzikos. Semi-supervised pattern classification of medical images: Appli" + ], + "bbox": [ + 93, + 232, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cation to mild cognitive impairment (mci). NeuroImage, 55 (3):1109-1119, 2011. 3", + "[11] Jack Furby, Daniel Cunningham, Dave Braines, and Alun Preece. Can we constrain concept bottleneck models to learn semantically meaningful input features? arXiv preprint arXiv:2402.00912, 2024. 2", + "[12] Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. Advances in Neural Information Processing Systems, 35:23386-23397, 2022. 2", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 11", + "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 6, 11", + "[15] Junlin Hou, Jilan Xu, and Hao Chen. Concept-attention whitening for interpretable skin lesion diagnosis. arXiv preprint arXiv:2404.05997, 2024. 2", + "[16] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Improving faithfulness for vision transformers. arXiv preprint arXiv:2311.17983, 2023. 1", + "[17] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Seat: stable and explainable attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12907-12915, 2023.", + "[18] Lijie Hu, Liang Liu, Shu Yang, Xin Chen, Hongru Xiao, Mengdi Li, Pan Zhou, Muhammad Asif Ali, and Di Wang. A hopfieldian view-based interpretation for chain-of-thought reasoning. arXiv preprint arXiv:2406.12255, 2024. 1", + "[19] Lijie Hu, Chenyang Ren, Zhengyu Hu, Cheng-Long Wang, and Di Wang. Editable concept bottleneck models. arXiv preprint arXiv:2405.15476, 2024. 2", + "[20] Tri Huynh, Aiden Nibali, and Zhen He. Semi-supervised learning for medical image classification using imbalanced training data. Computer Methods and Programs in Biomedicine, 216:106628, 2022. 3", + "[21] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[22] Jeremy Kawahara, Sara Daneshvar, Giuseppe Argenziano, and Ghassan Hamarneh. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE Journal of Biomedical and Health Informatics, 23(2): 538-546, 2019. 6, 11", + "[23] Mert Keser, Gesina Schwalbe, Azarm Nowzad, and Alois Knoll. Interpretable model-agnostic plausibility verification for 2d object detectors using domain-invariant concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890–3899, 2023. 2", + "[24] Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sunghroh Yoon. Probabilistic concept bottleneck models. arXiv preprint arXiv:2306.01574, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2118", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6", + "[26] Songning Lai, Lijie Hu, Junxiao Wang, Laure Berti-Equille, and Di Wang. Faithful vision-language interpretation via concept bottleneck models. In The Twelfth International Conference on Learning Representations, 2023. 1", + "[27] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.3", + "[28] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 3", + "[29] Jia Li, Lijie Hu, Zhixian He, Jingfeng Zhang, Tianhang Zheng, and Di Wang. Text guided image editing with automatic concept locating and forgetting. arXiv preprint arXiv:2405.19708, 2024. 2", + "[30] Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. Large language models in finance: A survey. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 374-382, 2023. 1", + "[31] Lucie Charlotte Magister, Dmitry Kazhdan, Vikash Singh, and Pietro Lio. Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks. arXiv preprint arXiv:2107.11889, 2021. 2", + "[32] Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47-60, 1996. 3", + "[33] Li Niu, Qingtao Tang, Ashok Veeraraghavan, and Ashutosh Sabharwal. Learning from noisy web data with category-level supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7689-7698, 2018. 6, 11", + "[34] Tuomas Oikarinen, Subhro Das, Lam Nguyen, and Lily Weng. Label-free concept bottleneck models. In International Conference on Learning Representations, 2023. 1, 2, 6, 7", + "[35] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv preprint arXiv:2006.05278, 2020. 3", + "[36] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11557-11568, 2021. 3", + "[37] Yoshihide Sawada and Keigo Nakamura. Concept bottleneck model with additional unsupervised concepts. IEEE Access, 10:41758-41765, 2022. 2", + "[38] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in neural information processing systems, 33:596-608, 2020. 3", + "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3", + "[41] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 1", + "[42] Satoshi Tsutsui, Winnie Pang, and Bihan Wen. Wbcatt: A white blood cell dataset annotated with detailed morphological attributes. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6, 11", + "[43] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 3", + "[44] Han Xuanyuan, Pietro Barbiero, Dobrik Georgiev, Lucie Charlotte Magister, and Pietro Lio. Global concept-based interpretability for graph neural networks via neuron analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10675-10683, 2023. 2", + "[45] An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amilcare Gentili, Chun-Nan Hsu, Jingbo Shang, et al. Robust and interpretable medical image classifiers via concept bottleneck models. arXiv preprint arXiv:2310.03182, 2023. 5", + "[46] Shu Yang, Lijie Hu, Lu Yu, Muhammad Asif Ali, and Di Wang. Human-ai interactions in the communication era: Autophagy makes large models achieving local optima. arXiv preprint arXiv:2402.11271, 2024. 1", + "[47] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification, 2023. 7", + "[48] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2", + "[49] Gang Zhang, Shan-Xing Ou, Yong-Hui Huang, and Chun-Ru Wang. Semi-supervised learning methods for large scale healthcare data analysis. International Journal of Computers in Healthcare, 2(2):98-110, 2015. 3" + ], + "bbox": [ + 516, + 92, + 903, + 685 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2119", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_model.json b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a2625676c9d74b607c9a5045064f1b06bc95ac2b --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.273, + 0.131, + 0.725, + 0.153 + ], + "angle": 0, + "content": "Semi-supervised Concept Bottleneck Models" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.179, + 0.77, + 0.217 + ], + "angle": 0, + "content": "Lijie Hu\\*,1,2, Tianhao Huang\\*,1,2,3, Huanyi Xie\\*,2, Xin Gong\\*,1,2,4, Chenyang Ren\\*,2, Zhengyu Hu\\*,6, Lu Yu\\*, Ping Ma\\*, Di Wang\\*,1,2" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.218, + 0.741, + 0.234 + ], + "angle": 0, + "content": "1Provable Responsible AI and Data Analytics (PRADA) Lab," + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.234, + 0.715, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)King Abdullah University of Science and Technology," + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.252, + 0.796, + 0.27 + ], + "angle": 0, + "content": "3University of Virginia, 4University of Georgia, 5Ant Group, 6HKUST(GZ)" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.27, + 0.844, + 0.288 + ], + "angle": 0, + "content": "{lijie.hu, tianhao.huang, huanyi.xie, di.wang}@kaust.edu.sa" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Concept Bottleneck Models (CBMs) have garnered increasing attention due to their ability to provide concept-based explanations for black-box deep learning models while achieving high final prediction accuracy using human-like concepts. However, the training of current CBMs is heavily dependent on the precision and richness of the annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. To address these limitations, we propose a new framework called SsCBM (Semi-supervised Concept Bottleneck Model). Our SSCBM is suitable for practical situations where annotated data is scarce. By leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level, we effectively solve these issues. We propose a strategy to generate pseudo labels and an alignment loss. Experiments demonstrate that our SSCBM is both effective and efficient. With only \\(10\\%\\) labeled data, our model's concept and task accuracy on average across four datasets is only \\(2.44\\%\\) and \\(3.93\\%\\) lower, respectively, compared to the best baseline in the fully supervised learning setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.485, + 0.864 + ], + "angle": 0, + "content": "Deep learning models like ResNet [13] typically feature complex non-linear architectures, making it difficult for end-users to understand and trust their decisions. This lack of interpretability is a significant obstacle to their adoption, especially in critical fields such as healthcare [41] and finance [30], where transparency is crucial. Explain-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.908, + 0.505 + ], + "angle": 0, + "content": "able artificial intelligence (XAI) models have been developed to meet the demand for transparency, providing insights into their behavior and internal mechanisms [16-18, 46]. Concept Bottleneck Models (CBMs) [25] are particularly notable among these XAI models for their ability to clarify the prediction process of end-to-end AI models. CBMs introduce a bottleneck layer that incorporates human-understandable concepts. During prediction, CBMs first predict concept labels from original input, then use these predicted concepts in bottleneck layer to determine the final label. This approach results in a self-explanatory decision-making process that users can comprehend." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.909, + 0.809 + ], + "angle": 0, + "content": "A major issue in original CBMs is the need for expert labeling, which is costly in practice. Some researchers address this problem through unsupervised learning. For example, [34] proposes a Label-free CBM that transforms any neural network into an interpretable CBM without requiring labeled concept data while maintaining high accuracy. Similarly, Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. However, these methods have three issues. First, those unsupervised methods heavily rely on large language models like GPT-3, which have reliability issues [26]. Second, the concepts extracted by these models lack evaluation metrics, undermining their interpretability. Third, the assumption that no concept labels are available is too stringent in practice. In reality, obtaining a small portion of concept labels is feasible and cost-effective. Therefore, we can maximize the use of this small labeled concept dataset. This is the motivation for introducing our framework, which focuses on the semi-supervised setting in CBM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce a framework called the SS-CBM (Semi-supervised Concept Bottleneck Model). Compared to a supervised setting, semi-supervised CBMs have two main challenges. First, obtaining concept embeddings requires concept labels, so we need to generate pseudo labels to obtain these concept embeddings. To achieve this," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.096, + 0.292, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.237, + 0.364, + 0.251 + ], + "angle": 0, + "content": "(a) Label Complete and Well-aligned" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.091, + 0.547, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.237, + 0.56, + 0.251 + ], + "angle": 0, + "content": "(b) Label Incomplete" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.097, + 0.613, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.091, + 0.788, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.098, + 0.854, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.237, + 0.784, + 0.251 + ], + "angle": 0, + "content": "(c) Misaligned" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 1. (a) A sample of sparrow class with complete concept labels. (b) A sample of sparrow class with incomplete concept labels. (c) A sample of misalignment between input features and concepts resulting from existing CBM methods. Our framework simultaneously utilizes both (a) and (b) types of data and addresses the issue of (c) through an alignment loss." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.332, + 0.485, + 0.738 + ], + "angle": 0, + "content": "SSCBM uses a KNN-based algorithm to assign pseudo-concept labels for unlabeled data. Second, while such a simple pseudo-labeling method is effective and has acceptable classification accuracy, we also find that the concept saliency map often misaligns with the input saliency map, meaning concept predictions frequently correspond to irrelevant input features. This misalignment often arises from inaccurate concept annotations or unclear relationships between input features and concepts, which is closely related to the broader issue of annotation alignment. In fact, in the supervised setting, there is a similar misalignment issue [11]. Existing research seeks to improve alignment by connecting textual and image information [15]. However, these methods only focus on supervised settings and cannot be directly applied to our settings, as our pseudo-labels contain noise. Our framework achieves excellent performance in both concept accuracy and concept saliency alignment by leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level. To achieve this, we leverage the relevance between the input image and the concept and get other pseudo-concept labels based on these similarity scores. Finally, we align these two types of pseudo-concept labels to give the concept encoder the ability to extract useful information from features while also inheriting the ability to align concept embeddings with the input. See Figure 1 for an illustration. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- We propose SSCBM, a framework designed to address both the semi-supervised annotation problem and the concept semantics alignment problem, which hold significant practical importance in real-world applications. SSCBM tackles them in an elegant and efficient manner. To the best of our knowledge, we are the first to tackle these two problems within a single framework, elucidating the behavior of CBMs through semi-supervised alignment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate the superiority of our SSCBM in annotation and concept-saliency align" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.332, + 0.907, + 0.408 + ], + "angle": 0, + "content": "ment, indicating its efficiency and effectiveness. Using only \\(10\\%\\) labeled data, our model achieves concept and task accuracy on average across the four datasets that are merely \\(2.44\\%\\) and \\(3.93\\%\\) lower, respectively, than the best baseline under fully supervised learning settings." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.422, + 0.655, + 0.437 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Concept Bottleneck Models. Concept Bottleneck Model (CBM) [25] is an innovative deep-learning approach for image classification and visual reasoning. By introducing a concept bottleneck layer into deep neural networks, CBMs enhance model generalization and interpretability by learning specific concepts. However, CBMs face two primary challenges: their performance often lags behind that of models without the bottleneck layer due to incomplete information extraction, and they rely heavily on laborious dataset annotation. Researchers have explored various solutions to these challenges. [5] extends CBMs into interactive prediction settings by introducing an interaction policy to determine which concepts to label, thereby improving final predictions. [34] addresses CBM limitations by proposing a Label-free CBM, which transforms any neural network into an interpretable CBM without requiring labeled concept data, maintaining high accuracy. Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. Related work in the image domain includes studies [12, 19, 23, 24, 29, 37, 38]. In the graph concept field, [31] provides a global interpretation of graph neural networks (GNNs) by mapping graphs into a concept space through clustering and offering a human-in-the-loop evaluation. [3, 44] extend this approach by incorporating both global and local explanations. For local explanations, they define a concept set, with each neuron represented as a vector with Boolean values indicating concept activation. However, existing works rarely consider semi-supervised settings, which are practical in real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2111" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "applications. Our framework addresses these issues effectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.121, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Semi-supervised Learning. Semi-supervised learning (SSL) combines the two main tasks of machine learning: supervised learning and unsupervised learning [43]. It is typically applied in scenarios where labeled data is scarce. Examples include computer-aided diagnosis [6, 49], medical image analysis [8, 10, 20], and drug discovery [7]. In these cases, collecting detailed annotated data by experts requires considerable time and effort. However, under the assumption of data distribution, unlabeled data can also assist in building better classifiers [43]. SSL, also known as self-labeling or self-teaching in its earliest forms, involves the model iteratively labeling a portion of unlabeled data and adding it to the training set for the next round of training [35]. The expectation-maximization (EM) algorithm proposed by [32] uses both labeled and unlabeled data to produce maximum likelihood estimates of parameters. [27] and [40] focus on consistency regularization. II-model [27] combines both supervised cross-entropy loss and unsupervised consistency loss while perturbing the model and data based on the consistency constraint assumption. A temporal ensembling model integrates predictions from models at various time points. Mean teacher [40] addresses the slow updating issue of the temporal ensembling model on large datasets by averaging model weights instead of predicting labels. MixMatch [4] unifies and refines the previous approaches of consistency regularization, entropy minimization, and traditional regularization into a single loss function, achieving excellent results. Pseudo labeling, as an effective tool for reducing the entropy of unlabeled data [28], has been increasingly attracting the attention of researchers in the field of semi-supervised learning. [2] proposes that directly using the model's predictions as pseudo-labels can achieve good results. FixMatch [39] predicts and retains the model, generating high-confidence pseudo-labels. [36] continuously adjusts the teacher based on feedback from the student, thereby generating better pseudo-labels. While there has been a plethora of work in the semi-supervised learning field, the focus on semi-supervised concept bottleneck models remains largely unexplored. Our work focuses on this new area." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.231, + 0.755 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Concept Bottleneck Models [25]. We consider a classification task with a concept set denoted as \\(\\mathcal{C} = \\{p_1,\\dots ,p_k\\}\\) with each \\(p_i\\) being a concept given by experts or LLMs, and a training dataset represented \\(\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}\\). Here, for \\(i\\in [N]\\), \\(x^{(i)}\\in \\mathcal{X}\\subseteq \\mathbb{R}^d\\) represents the feature vector (e.g., an image's pixels), \\(y^{(i)}\\in \\mathcal{Y}\\subseteq \\mathbb{R}^l\\) denotes the label (\\(l\\) is the number of classes), \\(c^{(i)} = (c_i^1,\\dots ,c_i^k)\\in \\mathbb{R}^k\\) represents the concept vector (a binary vector of length \\(k\\), where each value indicates whether the input \\(x^{(i)}\\) contains" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "that concept). In CBMs, the goal is to learn two representations: one called a concept encoder that transforms the input space into the concept space, denoted as \\( g: \\mathbb{R}^d \\to \\mathbb{R}^k \\), and another called label predictor that maps the concept space to the downstream prediction space, denoted as \\( f: \\mathbb{R}^k \\to \\mathbb{R}^l \\). Usually, the map \\( f \\) is linear. For any input \\( x \\), we aim to ensure that its predicted concept vector \\( \\hat{c} = g(x) \\) and prediction \\( \\hat{y} = f(g(x)) \\) are close to their underlying counterparts, thus capturing the essence of the original CBMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Concept Embedding Models [9]. As the original CBM is based solely on concept features to determine the predictions of the model, compared to canonical deep neural networks, it will degrade prediction performance. To further improve the performance of CBMs, CEM is developed by [9]. It achieves this by learning interpretable high-dimensional concept representations (i.e., concept embeddings), thus maintaining high task accuracy while obtaining concept representations that contain meaningful semantic information. For CEMs, we use the same setting as that of [9, 21]. For each input \\( x \\), the concept encoder learns \\( k \\) concept representations \\( \\hat{c}_1, \\hat{c}_2, \\dots, \\hat{c}_k \\), each corresponding to one of the \\( k \\) ground truth concepts in the training dataset. In CEMs, each concept \\( c_i \\) is represented using two embeddings \\( \\hat{c}_i^+, \\hat{c}_i^- \\in \\mathbb{R}^m \\), each with specific semantics, i.e., the concept is True (activate state) and concept is False (negative state), where hyper-parameter \\( m \\) is the embedding dimension. We use a DNN \\( \\psi(x) \\) to learn a latent representation \\( h \\in \\mathbb{R}^{n_h} \\), which is used as input to the CEM embedding generator, where \\( n_h \\) is the dimension of the latent representation. The CEM embedding generator \\( \\phi \\) feeds \\( h \\) into two fully connected concept-specific layers to learn two concept embeddings in \\( \\mathbb{R}^m \\)." + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.588, + 0.811, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} _ {i} = \\phi_ {i} (\\boldsymbol {h}) = a (W _ {i} \\boldsymbol {h} + \\boldsymbol {b} _ {i}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.616, + 0.907, + 0.738 + ], + "angle": 0, + "content": "Then we use a differential scoring function \\(s: \\mathbb{R}^{2m} \\to [0,1]\\), to achieve the alignment of concept embeddings \\(\\hat{\\pmb{c}}_i^+\\), \\(\\hat{\\pmb{c}}_i^-\\) and ground-truth concepts \\(c_i\\). It is trained to predict the probability \\(\\hat{p}_i := s\\left(\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right) = \\sigma\\left(W_s\\left[\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right] + \\pmb{b}_s\\right)\\) that the concept \\(c_i\\) is active in the embedding space. We get the final concept embedding \\(\\hat{\\pmb{c}}_i\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.75, + 0.8, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} _ {i} := \\hat {p} _ {i} \\hat {\\boldsymbol {c}} _ {i} ^ {+} + (1 - \\hat {p} _ {i}) \\hat {\\boldsymbol {c}} _ {i} ^ {-}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.856 + ], + "angle": 0, + "content": "At this point, we understand that we can obtain high-quality concept embeddings rich in semantics through CEMs. In the subsequent section 4, we will effectively utilize these representations of concepts and further optimize their interpretability through our proposed framework SSCBM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised Setting. Now, we consider the setting of semi-supervised learning for concept bottleneck models. As mentioned earlier, a typical training dataset for CBMs" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.484, + 0.257 + ], + "angle": 0, + "content": "can be represented as \\(\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}\\), where \\(x^{(i)}\\in \\mathcal{X}\\) represents the input feature. However, in semisupervised learning tasks, the set of feature vectors typically consists of two parts, \\(\\mathcal{X} = \\{\\mathcal{X}_L,\\mathcal{X}_U\\}\\), where \\(\\mathcal{X}_L\\) represents a small subset of labeled data and \\(\\mathcal{X}_U\\) represents the remaining unlabeled data. Here, the unlabeled data refers to data where the class label \\(y\\) is known, while the concept label \\(c\\) is unknown. Generally we assume \\(|\\mathcal{X}_L|\\ll |\\mathcal{X}_U|\\). We assume that \\(x^{(j)}\\in \\mathcal{X}_L\\) is labeled with a concept vector \\(c^{(j)}\\) and a label \\(y^{(j)}\\), and for any \\(x^{(i)}\\in \\mathcal{X}\\), there only exists a corresponding label \\(y^{(i)}\\in \\mathcal{Y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Under these settings, given a training dataset \\(\\mathcal{D} = \\mathcal{D}_L \\cup \\mathcal{D}_U\\) that includes both labeled and unlabeled data, the goal is to train a CBM using both the labeled data \\(\\mathcal{D}_L\\) and unlabeled data \\(\\mathcal{D}_U\\). This aims to get better mappings \\(g: \\mathbb{R}^d \\to \\mathbb{R}^k\\) and \\(f: \\mathbb{R}^k \\to \\mathbb{R}^l\\) than those trained by using only labeled data, ultimately achieving higher task accuracy and its corresponding concept-based explanation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.394 + ], + "angle": 0, + "content": "4. Semi-supervised Concept Bottleneck Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.581 + ], + "angle": 0, + "content": "In this section, we will elaborate on the details of the proposed SSCBM framework, which is shown in Figure 2. SSCBM follows the main idea of CEM. Specifically, to learn a good concept encoder, we use different processing methods for labeled and unlabeled data. The labeled data first passes through a feature extractor \\(\\psi\\) to be transformed into a latent representation \\(h\\), which then enters the concept embedding extractor to obtain the concept embeddings and the predicted concept vector \\(\\hat{c}\\) for the labeled data, which is compared to the ground truth concept to compute the concept loss. Moreover, the label predictor predicts \\(\\hat{y}\\) based on \\(\\hat{c}\\), and calculates the task loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.75 + ], + "angle": 0, + "content": "For unlabeled data, we first extract image features \\(V\\) from the input using an image encoder. Then, we use the KNN algorithm to assign a pseudo-label \\(\\hat{c}_{img}\\) to each unlabeled data, which has been experimentally proven to be simple and effective. In the second step, we compute a heatmap between concept embeddings and the input. After applying a threshold, we obtain the predicted alignment label \\(\\hat{c}_{align}\\). Finally, we compute the alignment loss between \\(\\hat{c}_{img}\\) and \\(\\hat{c}_{align}\\). During each training epoch, we simultaneously compute these losses and update the model parameters based on the gradients." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.474, + 0.775 + ], + "angle": 0, + "content": "4.1. Label Anchor: Concept Embedding Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Concept Embeddings. As described in Section 3, we obtain high-dimensional concept representations with meaningful semantics based on CEMs. Thus, our concept encoder should extract useful information from both labeled and unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.854, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For labeled training data \\(\\mathcal{D}_L = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{\\left|\\mathcal{D}_L\\right|}\\) we follow the original CEM [9], i.e., using a backbone network (e.g., ResNet50) to extract features \\(h =\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.906, + 0.184 + ], + "angle": 0, + "content": "\\(\\{\\psi(x^{(i)})\\}_{i=1}^{\\left|\\mathcal{D}_L\\right|}\\). Then, for each feature, it passes through an embedding generator to obtain concept embeddings \\(\\hat{c}_i \\in \\mathbb{R}^{m \\times k}\\) for \\(i \\in [k]\\). After passing through fully connected layers and activation layers, we obtain the predicted binary concept vector \\(\\hat{c} \\in \\mathbb{R}^k\\) for the labeled data. The specific process can be represented by the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.19, + 0.831, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} ^ {(j)} = \\psi (x ^ {(j)}), \\quad j = 1, \\ldots , | \\mathcal {D} _ {L} |,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.218, + 0.892, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {c}} _ {i} ^ {(j)} = \\sigma (\\phi (\\boldsymbol {h} ^ {(j)})), \\quad i = 1, \\dots , k, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.273 + ], + "angle": 0, + "content": "where \\(\\psi, \\phi\\) and \\(\\sigma\\) represent the backbone, embedding generator, and activation function, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.906, + 0.349 + ], + "angle": 0, + "content": "To enhance the interpretability of concept embeddings, we calculate the concept loss utilizing binary cross-entropy to optimize the accuracy of concept predictions by computing \\(\\mathcal{L}_c\\) based on predicted binary concept vector \\(\\hat{e}\\) and ground truth concept labels \\(c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.359, + 0.907, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = B C E (\\hat {\\boldsymbol {c}}, \\boldsymbol {c}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.813, + 0.4 + ], + "angle": 0, + "content": "where \\( BCE \\) is the binary cross entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Task Loss. Since our ultimate task is classification, we also need to incorporate the task loss for the final prediction performance. After obtaining the predicted concept \\(\\hat{c}\\), we use a label predictor to predict the final class \\(\\hat{\\pmb{y}}\\). Then we can define the task loss function using the categorical cross-entropy loss to train our classification model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.505, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a s k}} = C E (\\hat {\\boldsymbol {y}}, \\boldsymbol {y}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.572 + ], + "angle": 0, + "content": "Note that for the unlabeled data, we can also calculate the task loss since their class labels are known, to make full use of the data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.906, + 0.613 + ], + "angle": 0, + "content": "4.2. Unlabel Alignment: Image-Textual Semantics Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.906, + 0.723 + ], + "angle": 0, + "content": "Pseudo Labeling. Unlike supervised data, CEM cannot directly extract useful information from unlabeled data, as the concept encoder is a supervised training architecture. Thus, in practical situations lacking labeled data, one direct approach is to get high-quality pseudo concept labels to train the model. In the following, we will introduce the method we use to obtain pseudo-concept labels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.724, + 0.906, + 0.863 + ], + "angle": 0, + "content": "Firstly, we can naturally think of measuring the similarity between images by calculating the distance in the cosine space. Based on this idea, we can assign pseudo labels to unlabeled data by finding labeled data with similar image features to them. Specifically, for each unlabeled training data \\( x \\in \\mathcal{D}_U = \\{(x^{(i)},y^{(i)})\\}_{i=1}^{|D_U|} \\), we calculate its cosine distance with all labeled data points \\( x^{(j)} \\in \\mathcal{D}_L = \\{(x^{(j)},y^{(j)},c^{(j)})\\}_{j=1}^{|D_L|} \\) and select the \\( k \\) samples with the smallest distances:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.871, + 0.924, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} (x, x ^ {(j)}) = 1 - \\frac {\\Omega (x) \\cdot \\Omega (x ^ {(j)})}{\\| \\Omega (x) \\| _ {2} \\cdot \\| \\Omega (x ^ {(j)}) \\| _ {2}}, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2113" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.088, + 0.865, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.429, + 0.66, + 0.443 + ], + "angle": 0, + "content": "Figure 2. Overall framework of our proposed SSCBM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.575 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is a visual encoder. Then, we normalize the reciprocal of the cosine distance between the nearest \\(k\\) data points and \\(x\\) as weights. We use these weights to compute a weighted average of the concept labels of these \\(k\\) data points, obtaining the pseudo concept label for \\(x\\). In this way, we can obtain pseudo concept labels \\(\\hat{c}_{img}\\) for all \\(x \\in \\mathcal{D}_U\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.746 + ], + "angle": 0, + "content": "In our experiments, we find that directly feeding pseudo-concept labels generated by KNN to CEM has satisfactory performance. However, this simple labeling method can lead to alignment issues in the concept embeddings learned by the CEM's concept encoder. Specifically, the predicted concepts \\(\\hat{c}\\) could not have a relation to the corresponding features \\(V\\) in the image, which hinders the effectiveness of CEM as a reliable interpretability tool. Moreover, due to misalignment, this will also degrade the prediction performance of the concept encoder. In the following, we aim to address such a misalignment issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Generating Concept Heatmaps. Our above pseudo concept labels via KNN have already contained useful information for prediction. Thus, our goal is to provide these labels with more information about their relation to the corresponding features. To achieve this, we first get another pseudo-concept label via these relations, which are calculated by the similarity between the concept embedding and the input image, namely concept heatmaps. Specifically, given an image \\( x \\), we first have its feature map \\( V \\in \\mathbb{R}^{H \\times W \\times m} \\) for each concept extracted by \\( V = \\Omega(x) \\)," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.905, + 0.485 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) are the height and width of the feature map." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.905, + 0.547 + ], + "angle": 0, + "content": "Given \\(V\\) and the \\(i\\)-th concept embedding \\(\\pmb{c}_i\\), we can obtain a heatmap \\(\\mathcal{H}_i\\), i.e., a similarity matrix that measures the similarity between the concept and the image can be obtained by computing their cosine distance:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.555, + 0.909, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {p, q, i} = \\frac {\\hat {\\boldsymbol {c}} _ {i} ^ {m} \\cdot V _ {p , q}}{\\| \\hat {\\boldsymbol {c}} _ {i} ^ {m} \\| \\cdot \\| V _ {p , q} \\|}, \\quad p = 1, \\dots , H, \\quad q = 1, \\dots , W\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.675 + ], + "angle": 0, + "content": "where \\(p, q\\) are the \\(p\\)-th and \\(q\\)-th positions in the heatmaps, and \\(\\mathcal{H}_{p,q,i}\\) represents a local similarity score between \\(V\\) and \\(\\hat{c}_i^m\\). Intuitively, \\(\\mathcal{H}_i\\) represents the relation of each part of the image with the \\(i\\)-the concept. Then, we derive heatmaps for all concepts, denoted as \\(\\{\\mathcal{H}_1, \\mathcal{H}_2, \\dots, \\mathcal{H}_k\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Calculating Concept Scores and Concept Labels. As average pooling performs better in downstream classification tasks [45], we apply average pooling to the heatmaps to deduce the connection between the image and concepts: \\( s_i = \\frac{1}{P \\cdot Q} \\sum_{p=1}^{P} \\sum_{q=1}^{Q} \\mathcal{H}_{p,q,i} \\). Intuitively, \\( s_i \\) is the refined similarity score between the image and concept \\( \\hat{c}_i^m \\). Thus, a concept vector \\( s \\) can be obtained, representing the similarity between an image input \\( x \\) and the set of concepts: \\( s = (s_1, \\ldots, s_k)^\\top \\). \\( s \\) can be considered as a soft concept label which is got by similarity. Next, we have to transform it into a hard concept label \\( \\hat{c}_{\\text{align}} \\). We determine the presence of a concept attribute in an image based on a threshold value derived from an experiment. If the value \\( s_i \\) exceeds this threshold, we consider the image to possess that specific concept attribute and set the concept label to be True. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2114" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.472, + 0.106 + ], + "angle": 0, + "content": "can obtain predicted concept labels for all unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.213 + ], + "angle": 0, + "content": "Alignment of Image. Based on our earlier discussions, on the one hand, the concept encoder should learn information from \\(\\hat{c}_{img}\\). On the other hand, it should also get concept embeddings which can get good similarity-based concept labels \\(\\hat{c}_{\\text {align }}\\) for alignment with the input image. Thus, we need a further alignment loss to achieve these two goals. Specifically, we compute the alignment loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.226, + 0.483, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = B C E \\left(\\hat {c} _ {\\text {i m g}}, \\hat {c} _ {\\text {a l i g n}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.484, + 0.316 + ], + "angle": 0, + "content": "In fact, during this alignment process, the CBM backbone is taught a reasonable latent space structure by the visual encoder, and thus we can regard it as an implicit Student-Teacher setup framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.246, + 0.342 + ], + "angle": 0, + "content": "4.3. Final Objective" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.484, + 0.545 + ], + "angle": 0, + "content": "In this section, we will discuss how we derive the process of network optimization. First, we have the concept loss \\(\\mathcal{L}_c\\) in (1) to improve the interpretability of the concept embeddings. Also, since the concept embeddings are entered into the label predictor to output the final prediction, we also have a task loss between the predictions given by the concept bottleneck and ground truth, which is shown in (2). In the context of binary classification tasks, we employ binary cross-entropy (BCE) as our loss function. For multi-class classification tasks, we use cross-entropy as the measure. Finally, to align the images with the concept labels, we computed the loss of alignment in (3). Formally, the overall loss function of our approach can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.558, + 0.483, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {a l i g n}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.618 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2\\) are hyperparameters for the trade-off between interpretability and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.224, + 0.648 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.793 + ], + "angle": 0, + "content": "In this section, we will conduct experimental studies on the performance of our framework. Specifically, we evaluate the utility of our method according to the concept and prediction of the class label. We also design interpretability evaluation and test-time intervention to verify the interpretability and alignment performance of our method. Finally, we give the ablation study to verify why we need the alignment loss. Details and additional results are in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.298, + 0.82 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Datasets. We evaluate our methods on four real-world image tasks: CUB[14], AwA2[33], WBCatt[42] and 7-point[22]. See Appendix A.1 for a detailed introduction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare our SSCBM with Concept Bottle-neck Model (CBM) and Concept Embedding Model (CEM)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Concept and task accuracy results of SSCBM at a labeled ratio of 0.1, compared to baselines in the fully supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.128, + 0.905, + 0.2 + ], + "angle": 0, + "content": "
MethodCUBAwA2WBCatt7-point
ConceptTaskConceptTaskConceptTaskConceptTask
CBM93.99%67.33%96.48%88.71%94.18%99.71%74.34%75.44%
CEM96.39%79.82%95.91%87.00%95.33%99.71%77.15%75.85%
SSCBM90.88%67.67%96.48%89.77%93.98%99.68%73.67%70.09%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.226, + 0.907, + 0.376 + ], + "angle": 0, + "content": "mentioned in Section 3. Since those baselines do not inherently include settings for semi-supervised learning, to ensure fairness in the evaluation, we use the same pseudolabeling approach as SSCBM. We utilize the KNN algorithm to annotate unlabeled data based on labeled data, which is mentioned in Section 4.2. For CBM [25], we adopt the same setting and architecture as in the original CBM. And for CEM, we follow the same setting as in [9]. We also compare label-free CBM [34], which is an unsupervised CBM method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.378, + 0.906, + 0.488 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the utility, we consider the accuracy for the prediction of class and concept labels. Specifically, concept accuracy measures the model's prediction accuracy for concepts: \\(\\mathcal{C}_{acc} = \\frac{1}{N}\\cdot \\frac{1}{k}\\sum_{i = 1}^{N}\\sum_{j = 1}^{k}\\mathbb{I}(\\hat{c}_j^{(i)} = c_j^{(i)})\\). Task accuracy measures the model's performance in predicting downstream task classes: \\(\\mathcal{A}_{acc} = \\frac{1}{N}\\sum_{i = 1}^{N}\\mathbb{I}(\\hat{\\pmb{y}}^{(i)} = \\pmb {y}^{(i)})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.906, + 0.547 + ], + "angle": 0, + "content": "To evaluate the interpretability, besides the concept accuracy, similar to previous work [9, 25], we show visualization results. Moreover, we evaluate the performance of the test-time intervention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Implementation Details. All experiments are conducted on a Tesla V100S PCIe 32 GB GPU and an Intel Xeon Processor CPU. See Appendix A.2 for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.771, + 0.623 + ], + "angle": 0, + "content": "5.2. Evaluation Results on Utility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.809 + ], + "angle": 0, + "content": "To comprehensively evaluate the model's performance in the semi-supervised learning setting, we simulate real-world application scenarios. We consider labeling K samples for each class in the dataset, with an extreme case where \\( \\mathrm{K} = 1 \\). Clearly, \\( \\mathrm{K} = \\) total data size \\( \\times \\) labeled ratio \\( \\div \\) number of classes. We also conduct experiments with labeled ratios of 0.05, 0.1, 0.15, and 0.2. We also compare the performance of SSCBM with the baseline under the fully supervised setting. The experimental results are shown in Table 1 and 2. Additionally, we also explored the impact of different backbones on model performance, which can be found in the Appendix A.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the performance comparison of SSCBM at a labeled ratio of 0.1 with CEM and CBM under the fully supervised setting. It can be observed that, although CEM performs the best on CUB, WBCatt and 7-point, SSCBM is only slightly inferior to it. On the AwA2 dataset, its performance even surpasses that of CEM under the fully super" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2115" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Results of concept and task accuracy for different datasets with different portions of labeled data. Since the label-free CBM [34] and LaBo [47] do not use the concept set included in the dataset but instead rely on GPT-3 to generate concepts based on class names, we do not report the concept accuracy here." + }, + { + "type": "table", + "bbox": [ + 0.176, + 0.142, + 0.825, + 0.403 + ], + "angle": 0, + "content": "
DatasetLabeled RatioCBM+SSLCEM+SSLLabel-free CBMLaBoSSCBM
ConceptTaskConceptTaskConceptTaskConceptTaskConceptTask
CUBK=183.11%5.51%82.36%59.35%-74.31%-66.38%88.99%66.72%
0.05 (K=2)84.51%8.35%83.72%62.20%-74.31%-66.38%90.04%67.43%
0.1 (K=3)84.96%9.84%84.03%63.12%-74.31%-66.38%90.88%67.67%
0.15 (K=4)85.47%9.96%84.30%64.14%-74.31%-66.38%91.47%68.36%
0.2 (K=5)86.67%16.43%86.83%67.64%-74.31%-66.38%92.09%70.07%
AWA2K=171.32%90.87%65.55%84.44%-73.67%-93.73%85.05%89.02%
0.05 (K=30)77.23%91.11%68.31%85.01%-73.67%-93.73%95.42%89.37%
0.1 (K=60)80.45%91.03%69.96%85.48%-73.67%-93.73%96.48%89.77%
0.15 (K=90)83.46%91.00%72.14%86.61%-73.67%-93.73%96.81%90.54%
0.2 (K=120)85.76%91.12%72.14%86.61%-73.67%-93.73%96.81%90.44%
WBCattK=179.06%99.39%70.27%98.64%-41.92%-92.26%91.48%99.13%
0.05 (K=62)81.08%99.48%73.82%99.52%-41.92%-92.26%93.53%99.61%
0.1 (K=124)85.48%99.32%72.25%99.29%-41.92%-92.26%93.98%99.68%
0.15 (K=186)85.39%99.68%72.68%99.58%-41.92%-92.26%94.42%99.71%
0.2 (K=247)87.07%99.74%74.14%99.52%-41.92%-92.26%94.42%99.71%
7-pointK=159.91%55.95%62.78%66.09%-55.44%-64.97%66.58%66.84%
0.05 (K=5)65.36%57.47%67.85%67.09%-55.44%-64.97%70.98%68.77%
0.1 (K=9)68.82%55.70%72.23%66.33%-55.44%-64.97%73.67%70.09%
0.15 (K=13)66.14%59.75%66.54%67.09%-55.44%-64.97%73.94%72.56%
0.2 (K=17)70.29%60.44%73.04%66.84%-55.44%-64.97%76.52%74.56%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.459 + ], + "angle": 0, + "content": "vised setting, which sufficiently demonstrates the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.779 + ], + "angle": 0, + "content": "In Table 2, it can be observed that as the proportion of labeled data gradually increases, the concept and class prediction accuracy of all models improve. SSCBM outperforms all baselines. On the CUB dataset, when only one labeled sample per class is available, SSCBM achieves \\(88.99\\%\\) concept accuracy and \\(66.72\\%\\) class accuracy, which are \\(6.63\\%\\) and \\(7.37\\%\\) higher, respectively, than the performance achieved by CEM in a semi-supervised learning setting. This demonstrates the effectiveness of the Image-Textual Semantics Alignment module. The 7-point dataset is the most challenging, and on this dataset, SSCBM still achieves the best performance across all settings. As the amount of labeled data increases, the concept prediction accuracy improves from \\(66.58\\%\\) to \\(76.52\\%\\). We also note that on WBCatt and 7-point, the performance of Label-free CBM is poor. This is mainly because both datasets contain only five classes, which prevents Label-free CBM from fully leveraging the world knowledge of LLMs to generate an effective concept set, resulting in poor performance. This further demonstrates that semi-supervised learning scenarios are more realistic and practical for real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.336, + 0.805 + ], + "angle": 0, + "content": "5.3. Interpretability Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Besides the concept accuracy, note that in SSCBM, we also have a pseudo label revived by the alignment between the concept embedding and the input saliency map. We use the alignment loss to inherit this alignment. Thus, we will evaluate the alignment performance here to show the faithfulness of the interpretability given by SSCBM. We measure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.55 + ], + "angle": 0, + "content": "our alignment performance by comparing the correctness of the concept saliency map with the concept semantics in Figure 3. See Appendix A.2 for detailed experimental procedures for generating saliency maps. Results show that our concept saliency map matches the concept semantics, indicating the effectiveness of our alignment loss. In Appendix C, we also provide our additional interpretability evaluation in Figure 6-18." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.722, + 0.577 + ], + "angle": 0, + "content": "5.4. Test-time Intervention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Test-time intervention enables human users to interact with the model at inference time. We conduct this experiment by correcting the ratio \\(10\\%\\) to \\(100\\%\\) of the concept labels in the concept predictor. We expect that the model performance will steadily increase along with the ratio of concept intervention, indicating that the model learned such correct label information and automatically corrected other labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The results in Figure 4 demonstrate the robustness of our model and an increasing trend to learn the information about the concept, indicating our interpretability and model prediction performance. During the test phase, the SSCBM model generates probability scores for each concept (e.g., has_wing_color :: brown). Prior to intervention, we employ the COOP algorithm (following CEM) to identify the most critical concepts exhibiting both high predictive uncertainty and task importance. For these selected concepts, we replace the model's predicted values with their corresponding ground truth annotations. This rectified concept vector is subsequently fed into the classification head to produce the final task prediction. Finally, we successfully cause the model to predict the Great Crested Flycatcher instead of the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2116" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.089, + 0.326, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.214, + 0.288, + 0.224 + ], + "angle": 0, + "content": "(a) Original Image" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.498, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.214, + 0.486, + 0.224 + ], + "angle": 0, + "content": "(b) Tail Pattern: Multi-colored" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.089, + 0.669, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.214, + 0.653, + 0.224 + ], + "angle": 0, + "content": "(c) Underparts Color: Brown" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.089, + 0.838, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.214, + 0.811, + 0.224 + ], + "angle": 0, + "content": "(d) Wing Color: Brown" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 3. The concept saliency map for the CUB dataset (savannah sparrow) demonstrates that our proposed SSCBM achieves meaningful alignment between the ground truth concepts and the input image features. The first image on the left is the original input image. The three images on the right show the aligned regions for different concepts using SSCBM." + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.307, + 0.382, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.307, + 0.538, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.319, + 0.761, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.9, + 0.452 + ], + "angle": 0, + "content": "Figure 4. Left: Performance with different ratios of intervened concepts on CUB dataset. Right: An example of successful intervention." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.478, + 0.433, + 0.493 + ], + "angle": 0, + "content": "Swainson Warbler. More results are in Appendix B." + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.499, + 0.387, + 0.512 + ], + "angle": 0, + "content": "Table 3. Results of ablation study." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.516, + 0.485, + 0.734 + ], + "angle": 0, + "content": "
Ratiow/o imgw/o alignfull model
ConceptTaskConceptTaskConceptTask
CUB0.0001 (K=1)87.77%59.89%80.87%65.15%88.99%66.72%
0.05 (K=2)90.01%65.98%80.60%65.67%90.04%67.43%
0.1 (K=3)91.02%67.28%80.60%64.69%90.88%67.67%
0.15 (K=4)90.60%67.97%80.65%65.43%91.47%68.36%
0.2 (K=5)91.94%68.99%80.73%64.45%92.09%70.07%
WBCatt0.0001 (K=1)88.44%99.68%57.80%99.51%91.48%99.13%
0.05 (K=62)92.17%99.29%57.93%99.51%93.53%99.61%
0.1 (K=124)92.91%99.52%60.49%99.64%93.98%99.68%
0.15 (K=186)93.55%99.58%59.13%99.65%94.42%99.71%
0.2 (K=247)94.19%99.61%60.65%99.70%94.42%99.71%
7-point0.0001 (K=1)62.31%64.30%55.07%66.33%66.58%66.84%
0.05 (K=5)69.42%65.06%56.75%68.10%70.98%68.77%
0.1 (K=9)70.06%67.34%56.07%65.82%73.67%70.09%
0.15 (K=13)73.39%68.10%56.23%65.82%73.94%72.56%
0.2 (K=17)74.44%68.61%56.11%65.57%76.52%74.56%
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.245, + 0.759 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We then give a finer study on the two kinds of pseudo-labels to demonstrate that each one is indispensable in bolstering the efficacy of SSCBM. Specifically, based on our method in Section 4.2, we have three types of pseudo labels available: pseudo concept label \\(\\hat{c}_{img}\\), \\(\\hat{c}_{align}\\), and label \\(\\hat{c}\\) predicted by the concept embedding. We first remove \\(\\hat{c}_{img}\\) and calculate our alignment loss using \\(\\hat{c}_{align}\\) and \\(\\hat{c}\\); conversely, we remove \\(\\hat{c}_{align}\\) and calculate the alignment loss using \\(\\hat{c}_{img}\\) and \\(\\hat{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.705 + ], + "angle": 0, + "content": "From the results presented in Table 3, it is evident that removing the \\(\\hat{c}_{img}\\) component significantly degrades the performance of SSCBM at both the concept level and class level. This indicates that the pseudo-concept labels via KNN contain necessary information about the ground truth. The concept encoder needs to extract information from such labels to get better performance. This situation is similar when removing the \\(\\hat{c}_{align}\\) component, indicating that aligning the concept embedding and the input saliency map can further extract useful information from the input image and, thus, is beneficial to improve the performance. Our observations underscore the high degree of joint effectiveness of two kinds of pseudo-concept labels within our objective function, collectively contributing to the enhancement of model prediction and concept label prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The training of current CBMs heavily relies on the accuracy and richness of annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. In this problem, we propose SSCBM, a strategy to generate pseudo labels and an alignment loss to solve these two problems. Results show our effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2117" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work is supported in part by the funding BAS/1/1689-01-01, URF/1/4663-01-01, REI/1/5232-01-01, REI/1/5332-01-01, and URF/1/5508-01-01 from KAUST, and funding from KAUST - Center of Excellence for Generative AI, under award number 5940." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[1] Andrea Acevedo, Santiago Alférez, Anna Merino, Laura Puigvi, and José Rodellar. Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine, 180:105020, 2019. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[3] Pietro Barbiero, Gabriele Ciravegna, Francesco Giannini, Mateo Espinosa Zarlenga, Lucie Charlotte Magister, Alberto Tonda, Pietro Lio, Frédéric Precioso, Mateja Jamnik, and Giuseppe Marra. Interpretable neural-symbolic concept reasoning. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1801-1825. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[5] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[6] Asma Chebli, Akila Djebar, and Hayet Farida Marouani. Semi-supervised learning for medical application: A survey. In 2018 International Conference on Applied Smart Systems (ICASS), pages 1-9, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[7] Xing Chen, Biao Ren, Ming Chen, Quanxin Wang, Lixin Zhang, and Guiying Yan. Nllss: predicting synergistic drug combinations based on semi-supervised learning. PLoS computational biology, 12(7):e1004975, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[8] Veronika Cheplygina, Marleen de Bruijnne, and Josien P.W. Pluim. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis, 54:280-296, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[9] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 3, 4, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[10] Roman Filipovych and Christos Davatzikos. Semi-supervised pattern classification of medical images: Appli" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "cation to mild cognitive impairment (mci). NeuroImage, 55 (3):1109-1119, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[11] Jack Furby, Daniel Cunningham, Dave Braines, and Alun Preece. Can we constrain concept bottleneck models to learn semantically meaningful input features? arXiv preprint arXiv:2402.00912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[12] Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. Advances in Neural Information Processing Systems, 35:23386-23397, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[15] Junlin Hou, Jilan Xu, and Hao Chen. Concept-attention whitening for interpretable skin lesion diagnosis. arXiv preprint arXiv:2404.05997, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.904, + 0.425 + ], + "angle": 0, + "content": "[16] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Improving faithfulness for vision transformers. arXiv preprint arXiv:2311.17983, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.904, + 0.483 + ], + "angle": 0, + "content": "[17] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Seat: stable and explainable attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12907-12915, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.486, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[18] Lijie Hu, Liang Liu, Shu Yang, Xin Chen, Hongru Xiao, Mengdi Li, Pan Zhou, Muhammad Asif Ali, and Di Wang. A hopfieldian view-based interpretation for chain-of-thought reasoning. arXiv preprint arXiv:2406.12255, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[19] Lijie Hu, Chenyang Ren, Zhengyu Hu, Cheng-Long Wang, and Di Wang. Editable concept bottleneck models. arXiv preprint arXiv:2405.15476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.587, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[20] Tri Huynh, Aiden Nibali, and Zhen He. Semi-supervised learning for medical image classification using imbalanced training data. Computer Methods and Programs in Biomedicine, 216:106628, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[21] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[22] Jeremy Kawahara, Sara Daneshvar, Giuseppe Argenziano, and Ghassan Hamarneh. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE Journal of Biomedical and Health Informatics, 23(2): 538-546, 2019. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[23] Mert Keser, Gesina Schwalbe, Azarm Nowzad, and Alois Knoll. Interpretable model-agnostic plausibility verification for 2d object detectors using domain-invariant concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890–3899, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sunghroh Yoon. Probabilistic concept bottleneck models. arXiv preprint arXiv:2306.01574, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2118" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Songning Lai, Lijie Hu, Junxiao Wang, Laure Berti-Equille, and Di Wang. Faithful vision-language interpretation via concept bottleneck models. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[27] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Jia Li, Lijie Hu, Zhixian He, Jingfeng Zhang, Tianhang Zheng, and Di Wang. Text guided image editing with automatic concept locating and forgetting. arXiv preprint arXiv:2405.19708, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[30] Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. Large language models in finance: A survey. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 374-382, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Lucie Charlotte Magister, Dmitry Kazhdan, Vikash Singh, and Pietro Lio. Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks. arXiv preprint arXiv:2107.11889, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[32] Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47-60, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[33] Li Niu, Qingtao Tang, Ashok Veeraraghavan, and Ashutosh Sabharwal. Learning from noisy web data with category-level supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7689-7698, 2018. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[34] Tuomas Oikarinen, Subhro Das, Lam Nguyen, and Lily Weng. Label-free concept bottleneck models. In International Conference on Learning Representations, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[35] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv preprint arXiv:2006.05278, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[36] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11557-11568, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[37] Yoshihide Sawada and Keigo Nakamura. Concept bottleneck model with additional unsupervised concepts. IEEE Access, 10:41758-41765, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[38] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Advances in neural information processing systems, 33:596-608, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[41] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[42] Satoshi Tsutsui, Winnie Pang, and Bihan Wen. Wbcatt: A white blood cell dataset annotated with detailed morphological attributes. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[43] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[44] Han Xuanyuan, Pietro Barbiero, Dobrik Georgiev, Lucie Charlotte Magister, and Pietro Lio. Global concept-based interpretability for graph neural networks via neuron analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10675-10683, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[45] An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amilcare Gentili, Chun-Nan Hsu, Jingbo Shang, et al. Robust and interpretable medical image classifiers via concept bottleneck models. arXiv preprint arXiv:2310.03182, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[46] Shu Yang, Lijie Hu, Lu Yu, Muhammad Asif Ali, and Di Wang. Human-ai interactions in the communication era: Autophagy makes large models achieving local optima. arXiv preprint arXiv:2402.11271, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[47] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[48] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[49] Gang Zhang, Shan-Xing Ou, Yong-Hui Huang, and Chun-Ru Wang. Semi-supervised learning methods for large scale healthcare data analysis. International Journal of Computers in Healthcare, 2(2):98-110, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2119" + } + ] +] \ No newline at end of file diff --git a/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_origin.pdf b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64f4a30f411ddd3774b93918f0f892fe78bf82f6 --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/790be7e9-6d82-433e-9aab-92f2c72b545a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a286b9ff940f74733a4c4d4a83ee582940bdac7c3011d4b6b3c38b12b1e0b38 +size 2029993 diff --git a/2025/Semi-supervised Concept Bottleneck Models/full.md b/2025/Semi-supervised Concept Bottleneck Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cd9cd0b131d9a628cf21114a92e8432b103ccb31 --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/full.md @@ -0,0 +1,316 @@ +# Semi-supervised Concept Bottleneck Models + +Lijie Hu\*,1,2, Tianhao Huang\*,1,2,3, Huanyi Xie\*,2, Xin Gong\*,1,2,4, Chenyang Ren\*,2, Zhengyu Hu\*,6, Lu Yu\*, Ping Ma\*, Di Wang\*,1,2 + +1Provable Responsible AI and Data Analytics (PRADA) Lab, + +$^{2}$ King Abdullah University of Science and Technology, + +3University of Virginia, 4University of Georgia, 5Ant Group, 6HKUST(GZ) + +{lijie.hu, tianhao.huang, huanyi.xie, di.wang}@kaust.edu.sa + +# Abstract + +Concept Bottleneck Models (CBMs) have garnered increasing attention due to their ability to provide concept-based explanations for black-box deep learning models while achieving high final prediction accuracy using human-like concepts. However, the training of current CBMs is heavily dependent on the precision and richness of the annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. To address these limitations, we propose a new framework called SsCBM (Semi-supervised Concept Bottleneck Model). Our SSCBM is suitable for practical situations where annotated data is scarce. By leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level, we effectively solve these issues. We propose a strategy to generate pseudo labels and an alignment loss. Experiments demonstrate that our SSCBM is both effective and efficient. With only $10\%$ labeled data, our model's concept and task accuracy on average across four datasets is only $2.44\%$ and $3.93\%$ lower, respectively, compared to the best baseline in the fully supervised learning setting. + +# 1. Introduction + +Deep learning models like ResNet [13] typically feature complex non-linear architectures, making it difficult for end-users to understand and trust their decisions. This lack of interpretability is a significant obstacle to their adoption, especially in critical fields such as healthcare [41] and finance [30], where transparency is crucial. Explain- + +able artificial intelligence (XAI) models have been developed to meet the demand for transparency, providing insights into their behavior and internal mechanisms [16-18, 46]. Concept Bottleneck Models (CBMs) [25] are particularly notable among these XAI models for their ability to clarify the prediction process of end-to-end AI models. CBMs introduce a bottleneck layer that incorporates human-understandable concepts. During prediction, CBMs first predict concept labels from original input, then use these predicted concepts in bottleneck layer to determine the final label. This approach results in a self-explanatory decision-making process that users can comprehend. + +A major issue in original CBMs is the need for expert labeling, which is costly in practice. Some researchers address this problem through unsupervised learning. For example, [34] proposes a Label-free CBM that transforms any neural network into an interpretable CBM without requiring labeled concept data while maintaining high accuracy. Similarly, Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. However, these methods have three issues. First, those unsupervised methods heavily rely on large language models like GPT-3, which have reliability issues [26]. Second, the concepts extracted by these models lack evaluation metrics, undermining their interpretability. Third, the assumption that no concept labels are available is too stringent in practice. In reality, obtaining a small portion of concept labels is feasible and cost-effective. Therefore, we can maximize the use of this small labeled concept dataset. This is the motivation for introducing our framework, which focuses on the semi-supervised setting in CBM. + +In this paper, we introduce a framework called the SS-CBM (Semi-supervised Concept Bottleneck Model). Compared to a supervised setting, semi-supervised CBMs have two main challenges. First, obtaining concept embeddings requires concept labels, so we need to generate pseudo labels to obtain these concept embeddings. To achieve this, + +![](images/0b72e38c5fb038035c898ebcbac7c7710816c65463a2dfd60d4193e6c0347496.jpg) +(a) Label Complete and Well-aligned + +![](images/325aa5b1fdfc33c39845046e8bf28a03b67bdedec9aa962d5cfba278dbfe15df.jpg) +Figure 1. (a) A sample of sparrow class with complete concept labels. (b) A sample of sparrow class with incomplete concept labels. (c) A sample of misalignment between input features and concepts resulting from existing CBM methods. Our framework simultaneously utilizes both (a) and (b) types of data and addresses the issue of (c) through an alignment loss. + +![](images/71d036ecdeafabac10ff570578e7bb05c5b079de45ffef4b7a1ec66a10413296.jpg) +(b) Label Incomplete + +![](images/475c52110745db85f13d30e043432371fc73219f227b4f083ed782ef3f6af163.jpg) +(c) Misaligned + +![](images/54b430661fa18cc079ecd653e1b1ee341bed310aee4c5febc1f047f0a4e55ce5.jpg) + +SSCBM uses a KNN-based algorithm to assign pseudo-concept labels for unlabeled data. Second, while such a simple pseudo-labeling method is effective and has acceptable classification accuracy, we also find that the concept saliency map often misaligns with the input saliency map, meaning concept predictions frequently correspond to irrelevant input features. This misalignment often arises from inaccurate concept annotations or unclear relationships between input features and concepts, which is closely related to the broader issue of annotation alignment. In fact, in the supervised setting, there is a similar misalignment issue [11]. Existing research seeks to improve alignment by connecting textual and image information [15]. However, these methods only focus on supervised settings and cannot be directly applied to our settings, as our pseudo-labels contain noise. Our framework achieves excellent performance in both concept accuracy and concept saliency alignment by leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level. To achieve this, we leverage the relevance between the input image and the concept and get other pseudo-concept labels based on these similarity scores. Finally, we align these two types of pseudo-concept labels to give the concept encoder the ability to extract useful information from features while also inheriting the ability to align concept embeddings with the input. See Figure 1 for an illustration. Our contributions can be summarized as follows: + +- We propose SSCBM, a framework designed to address both the semi-supervised annotation problem and the concept semantics alignment problem, which hold significant practical importance in real-world applications. SSCBM tackles them in an elegant and efficient manner. To the best of our knowledge, we are the first to tackle these two problems within a single framework, elucidating the behavior of CBMs through semi-supervised alignment. +- Comprehensive experiments demonstrate the superiority of our SSCBM in annotation and concept-saliency align + +ment, indicating its efficiency and effectiveness. Using only $10\%$ labeled data, our model achieves concept and task accuracy on average across the four datasets that are merely $2.44\%$ and $3.93\%$ lower, respectively, than the best baseline under fully supervised learning settings. + +# 2. Related Work + +Concept Bottleneck Models. Concept Bottleneck Model (CBM) [25] is an innovative deep-learning approach for image classification and visual reasoning. By introducing a concept bottleneck layer into deep neural networks, CBMs enhance model generalization and interpretability by learning specific concepts. However, CBMs face two primary challenges: their performance often lags behind that of models without the bottleneck layer due to incomplete information extraction, and they rely heavily on laborious dataset annotation. Researchers have explored various solutions to these challenges. [5] extends CBMs into interactive prediction settings by introducing an interaction policy to determine which concepts to label, thereby improving final predictions. [34] addresses CBM limitations by proposing a Label-free CBM, which transforms any neural network into an interpretable CBM without requiring labeled concept data, maintaining high accuracy. Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. Related work in the image domain includes studies [12, 19, 23, 24, 29, 37, 38]. In the graph concept field, [31] provides a global interpretation of graph neural networks (GNNs) by mapping graphs into a concept space through clustering and offering a human-in-the-loop evaluation. [3, 44] extend this approach by incorporating both global and local explanations. For local explanations, they define a concept set, with each neuron represented as a vector with Boolean values indicating concept activation. However, existing works rarely consider semi-supervised settings, which are practical in real-world + +applications. Our framework addresses these issues effectively. + +Semi-supervised Learning. Semi-supervised learning (SSL) combines the two main tasks of machine learning: supervised learning and unsupervised learning [43]. It is typically applied in scenarios where labeled data is scarce. Examples include computer-aided diagnosis [6, 49], medical image analysis [8, 10, 20], and drug discovery [7]. In these cases, collecting detailed annotated data by experts requires considerable time and effort. However, under the assumption of data distribution, unlabeled data can also assist in building better classifiers [43]. SSL, also known as self-labeling or self-teaching in its earliest forms, involves the model iteratively labeling a portion of unlabeled data and adding it to the training set for the next round of training [35]. The expectation-maximization (EM) algorithm proposed by [32] uses both labeled and unlabeled data to produce maximum likelihood estimates of parameters. [27] and [40] focus on consistency regularization. II-model [27] combines both supervised cross-entropy loss and unsupervised consistency loss while perturbing the model and data based on the consistency constraint assumption. A temporal ensembling model integrates predictions from models at various time points. Mean teacher [40] addresses the slow updating issue of the temporal ensembling model on large datasets by averaging model weights instead of predicting labels. MixMatch [4] unifies and refines the previous approaches of consistency regularization, entropy minimization, and traditional regularization into a single loss function, achieving excellent results. Pseudo labeling, as an effective tool for reducing the entropy of unlabeled data [28], has been increasingly attracting the attention of researchers in the field of semi-supervised learning. [2] proposes that directly using the model's predictions as pseudo-labels can achieve good results. FixMatch [39] predicts and retains the model, generating high-confidence pseudo-labels. [36] continuously adjusts the teacher based on feedback from the student, thereby generating better pseudo-labels. While there has been a plethora of work in the semi-supervised learning field, the focus on semi-supervised concept bottleneck models remains largely unexplored. Our work focuses on this new area. + +# 3. Preliminaries + +Concept Bottleneck Models [25]. We consider a classification task with a concept set denoted as $\mathcal{C} = \{p_1,\dots ,p_k\}$ with each $p_i$ being a concept given by experts or LLMs, and a training dataset represented $\mathcal{D} = \{(x^{(i)},y^{(i)},c^{(i)})\}_{i = 1}^{N}$ . Here, for $i\in [N]$ , $x^{(i)}\in \mathcal{X}\subseteq \mathbb{R}^d$ represents the feature vector (e.g., an image's pixels), $y^{(i)}\in \mathcal{Y}\subseteq \mathbb{R}^l$ denotes the label ( $l$ is the number of classes), $c^{(i)} = (c_i^1,\dots ,c_i^k)\in \mathbb{R}^k$ represents the concept vector (a binary vector of length $k$ , where each value indicates whether the input $x^{(i)}$ contains + +that concept). In CBMs, the goal is to learn two representations: one called a concept encoder that transforms the input space into the concept space, denoted as $g: \mathbb{R}^d \to \mathbb{R}^k$ , and another called label predictor that maps the concept space to the downstream prediction space, denoted as $f: \mathbb{R}^k \to \mathbb{R}^l$ . Usually, the map $f$ is linear. For any input $x$ , we aim to ensure that its predicted concept vector $\hat{c} = g(x)$ and prediction $\hat{y} = f(g(x))$ are close to their underlying counterparts, thus capturing the essence of the original CBMs. + +Concept Embedding Models [9]. As the original CBM is based solely on concept features to determine the predictions of the model, compared to canonical deep neural networks, it will degrade prediction performance. To further improve the performance of CBMs, CEM is developed by [9]. It achieves this by learning interpretable high-dimensional concept representations (i.e., concept embeddings), thus maintaining high task accuracy while obtaining concept representations that contain meaningful semantic information. For CEMs, we use the same setting as that of [9, 21]. For each input $x$ , the concept encoder learns $k$ concept representations $\hat{c}_1, \hat{c}_2, \dots, \hat{c}_k$ , each corresponding to one of the $k$ ground truth concepts in the training dataset. In CEMs, each concept $c_i$ is represented using two embeddings $\hat{c}_i^+, \hat{c}_i^- \in \mathbb{R}^m$ , each with specific semantics, i.e., the concept is True (activate state) and concept is False (negative state), where hyper-parameter $m$ is the embedding dimension. We use a DNN $\psi(x)$ to learn a latent representation $h \in \mathbb{R}^{n_h}$ , which is used as input to the CEM embedding generator, where $n_h$ is the dimension of the latent representation. The CEM embedding generator $\phi$ feeds $h$ into two fully connected concept-specific layers to learn two concept embeddings in $\mathbb{R}^m$ . + +$$ +\hat {\boldsymbol {c}} _ {i} = \phi_ {i} (\boldsymbol {h}) = a (W _ {i} \boldsymbol {h} + \boldsymbol {b} _ {i}). +$$ + +Then we use a differential scoring function $s: \mathbb{R}^{2m} \to [0,1]$ , to achieve the alignment of concept embeddings $\hat{\pmb{c}}_i^+$ , $\hat{\pmb{c}}_i^-$ and ground-truth concepts $c_i$ . It is trained to predict the probability $\hat{p}_i := s\left(\left[\hat{\pmb{c}}_i^+, \hat{\pmb{c}}_i^-\right]^\top\right) = \sigma\left(W_s\left[\left[\hat{\pmb{c}}_i^+, \hat{\pmb{c}}_i^-\right]^\top\right] + \pmb{b}_s\right)$ that the concept $c_i$ is active in the embedding space. We get the final concept embedding $\hat{\pmb{c}}_i$ , as follows: + +$$ +\hat {\boldsymbol {c}} _ {i} := \hat {p} _ {i} \hat {\boldsymbol {c}} _ {i} ^ {+} + (1 - \hat {p} _ {i}) \hat {\boldsymbol {c}} _ {i} ^ {-}. +$$ + +At this point, we understand that we can obtain high-quality concept embeddings rich in semantics through CEMs. In the subsequent section 4, we will effectively utilize these representations of concepts and further optimize their interpretability through our proposed framework SSCBM. + +Semi-supervised Setting. Now, we consider the setting of semi-supervised learning for concept bottleneck models. As mentioned earlier, a typical training dataset for CBMs + +can be represented as $\mathcal{D} = \{(x^{(i)},y^{(i)},c^{(i)})\}_{i = 1}^{N}$ , where $x^{(i)}\in \mathcal{X}$ represents the input feature. However, in semisupervised learning tasks, the set of feature vectors typically consists of two parts, $\mathcal{X} = \{\mathcal{X}_L,\mathcal{X}_U\}$ , where $\mathcal{X}_L$ represents a small subset of labeled data and $\mathcal{X}_U$ represents the remaining unlabeled data. Here, the unlabeled data refers to data where the class label $y$ is known, while the concept label $c$ is unknown. Generally we assume $|\mathcal{X}_L|\ll |\mathcal{X}_U|$ . We assume that $x^{(j)}\in \mathcal{X}_L$ is labeled with a concept vector $c^{(j)}$ and a label $y^{(j)}$ , and for any $x^{(i)}\in \mathcal{X}$ , there only exists a corresponding label $y^{(i)}\in \mathcal{Y}$ . + +Under these settings, given a training dataset $\mathcal{D} = \mathcal{D}_L \cup \mathcal{D}_U$ that includes both labeled and unlabeled data, the goal is to train a CBM using both the labeled data $\mathcal{D}_L$ and unlabeled data $\mathcal{D}_U$ . This aims to get better mappings $g: \mathbb{R}^d \to \mathbb{R}^k$ and $f: \mathbb{R}^k \to \mathbb{R}^l$ than those trained by using only labeled data, ultimately achieving higher task accuracy and its corresponding concept-based explanation. + +# 4. Semi-supervised Concept Bottleneck Models + +In this section, we will elaborate on the details of the proposed SSCBM framework, which is shown in Figure 2. SSCBM follows the main idea of CEM. Specifically, to learn a good concept encoder, we use different processing methods for labeled and unlabeled data. The labeled data first passes through a feature extractor $\psi$ to be transformed into a latent representation $h$ , which then enters the concept embedding extractor to obtain the concept embeddings and the predicted concept vector $\hat{c}$ for the labeled data, which is compared to the ground truth concept to compute the concept loss. Moreover, the label predictor predicts $\hat{y}$ based on $\hat{c}$ , and calculates the task loss. + +For unlabeled data, we first extract image features $V$ from the input using an image encoder. Then, we use the KNN algorithm to assign a pseudo-label $\hat{c}_{img}$ to each unlabeled data, which has been experimentally proven to be simple and effective. In the second step, we compute a heatmap between concept embeddings and the input. After applying a threshold, we obtain the predicted alignment label $\hat{c}_{align}$ . Finally, we compute the alignment loss between $\hat{c}_{img}$ and $\hat{c}_{align}$ . During each training epoch, we simultaneously compute these losses and update the model parameters based on the gradients. + +# 4.1. Label Anchor: Concept Embedding Encoder + +Concept Embeddings. As described in Section 3, we obtain high-dimensional concept representations with meaningful semantics based on CEMs. Thus, our concept encoder should extract useful information from both labeled and unlabeled data. + +For labeled training data $\mathcal{D}_L = \{(x^{(i)},y^{(i)},c^{(i)})\}_{i = 1}^{\left|\mathcal{D}_L\right|}$ we follow the original CEM [9], i.e., using a backbone network (e.g., ResNet50) to extract features $h =$ + +$\{\psi(x^{(i)})\}_{i=1}^{\left|\mathcal{D}_L\right|}$ . Then, for each feature, it passes through an embedding generator to obtain concept embeddings $\hat{c}_i \in \mathbb{R}^{m \times k}$ for $i \in [k]$ . After passing through fully connected layers and activation layers, we obtain the predicted binary concept vector $\hat{c} \in \mathbb{R}^k$ for the labeled data. The specific process can be represented by the following expression: + +$$ +\boldsymbol {h} ^ {(j)} = \psi (x ^ {(j)}), \quad j = 1, \ldots , | \mathcal {D} _ {L} |, +$$ + +$$ +\hat {\boldsymbol {c}} _ {i} ^ {(j)} = \sigma (\phi (\boldsymbol {h} ^ {(j)})), \quad i = 1, \dots , k, \quad j = 1, \dots , | \mathcal {D} _ {L} |, +$$ + +where $\psi, \phi$ and $\sigma$ represent the backbone, embedding generator, and activation function, respectively. + +To enhance the interpretability of concept embeddings, we calculate the concept loss utilizing binary cross-entropy to optimize the accuracy of concept predictions by computing $\mathcal{L}_c$ based on predicted binary concept vector $\hat{e}$ and ground truth concept labels $c$ : + +$$ +\mathcal {L} _ {c} = B C E (\hat {\boldsymbol {c}}, \boldsymbol {c}), \tag {1} +$$ + +where $BCE$ is the binary cross entropy loss. + +Task Loss. Since our ultimate task is classification, we also need to incorporate the task loss for the final prediction performance. After obtaining the predicted concept $\hat{c}$ , we use a label predictor to predict the final class $\hat{\pmb{y}}$ . Then we can define the task loss function using the categorical cross-entropy loss to train our classification model as follows: + +$$ +\mathcal {L} _ {\text {t a s k}} = C E (\hat {\boldsymbol {y}}, \boldsymbol {y}), \tag {2} +$$ + +Note that for the unlabeled data, we can also calculate the task loss since their class labels are known, to make full use of the data. + +# 4.2. Unlabel Alignment: Image-Textual Semantics Alignment + +Pseudo Labeling. Unlike supervised data, CEM cannot directly extract useful information from unlabeled data, as the concept encoder is a supervised training architecture. Thus, in practical situations lacking labeled data, one direct approach is to get high-quality pseudo concept labels to train the model. In the following, we will introduce the method we use to obtain pseudo-concept labels. + +Firstly, we can naturally think of measuring the similarity between images by calculating the distance in the cosine space. Based on this idea, we can assign pseudo labels to unlabeled data by finding labeled data with similar image features to them. Specifically, for each unlabeled training data $x \in \mathcal{D}_U = \{(x^{(i)},y^{(i)})\}_{i=1}^{|D_U|}$ , we calculate its cosine distance with all labeled data points $x^{(j)} \in \mathcal{D}_L = \{(x^{(j)},y^{(j)},c^{(j)})\}_{j=1}^{|D_L|}$ and select the $k$ samples with the smallest distances: + +$$ +\operatorname {d i s t} (x, x ^ {(j)}) = 1 - \frac {\Omega (x) \cdot \Omega (x ^ {(j)})}{\| \Omega (x) \| _ {2} \cdot \| \Omega (x ^ {(j)}) \| _ {2}}, \quad j = 1, \dots , | \mathcal {D} _ {L} |, +$$ + +![](images/f6055d52a8973357c3adf755aef92bc10c105af9ccd0b6cc0e429a6d48343a92.jpg) +Figure 2. Overall framework of our proposed SSCBM. + +where $\Omega$ is a visual encoder. Then, we normalize the reciprocal of the cosine distance between the nearest $k$ data points and $x$ as weights. We use these weights to compute a weighted average of the concept labels of these $k$ data points, obtaining the pseudo concept label for $x$ . In this way, we can obtain pseudo concept labels $\hat{c}_{img}$ for all $x \in \mathcal{D}_U$ . + +In our experiments, we find that directly feeding pseudo-concept labels generated by KNN to CEM has satisfactory performance. However, this simple labeling method can lead to alignment issues in the concept embeddings learned by the CEM's concept encoder. Specifically, the predicted concepts $\hat{c}$ could not have a relation to the corresponding features $V$ in the image, which hinders the effectiveness of CEM as a reliable interpretability tool. Moreover, due to misalignment, this will also degrade the prediction performance of the concept encoder. In the following, we aim to address such a misalignment issue. + +Generating Concept Heatmaps. Our above pseudo concept labels via KNN have already contained useful information for prediction. Thus, our goal is to provide these labels with more information about their relation to the corresponding features. To achieve this, we first get another pseudo-concept label via these relations, which are calculated by the similarity between the concept embedding and the input image, namely concept heatmaps. Specifically, given an image $x$ , we first have its feature map $V \in \mathbb{R}^{H \times W \times m}$ for each concept extracted by $V = \Omega(x)$ , + +where $H$ and $W$ are the height and width of the feature map. + +Given $V$ and the $i$ -th concept embedding $\pmb{c}_i$ , we can obtain a heatmap $\mathcal{H}_i$ , i.e., a similarity matrix that measures the similarity between the concept and the image can be obtained by computing their cosine distance: + +$$ +\mathcal {H} _ {p, q, i} = \frac {\hat {\boldsymbol {c}} _ {i} ^ {m} \cdot V _ {p , q}}{\| \hat {\boldsymbol {c}} _ {i} ^ {m} \| \cdot \| V _ {p , q} \|}, \quad p = 1, \dots , H, \quad q = 1, \dots , W +$$ + +where $p, q$ are the $p$ -th and $q$ -th positions in the heatmaps, and $\mathcal{H}_{p,q,i}$ represents a local similarity score between $V$ and $\hat{c}_i^m$ . Intuitively, $\mathcal{H}_i$ represents the relation of each part of the image with the $i$ -the concept. Then, we derive heatmaps for all concepts, denoted as $\{\mathcal{H}_1, \mathcal{H}_2, \dots, \mathcal{H}_k\}$ . + +Calculating Concept Scores and Concept Labels. As average pooling performs better in downstream classification tasks [45], we apply average pooling to the heatmaps to deduce the connection between the image and concepts: $s_i = \frac{1}{P \cdot Q} \sum_{p=1}^{P} \sum_{q=1}^{Q} \mathcal{H}_{p,q,i}$ . Intuitively, $s_i$ is the refined similarity score between the image and concept $\hat{c}_i^m$ . Thus, a concept vector $s$ can be obtained, representing the similarity between an image input $x$ and the set of concepts: $s = (s_1, \ldots, s_k)^\top$ . $s$ can be considered as a soft concept label which is got by similarity. Next, we have to transform it into a hard concept label $\hat{c}_{\text{align}}$ . We determine the presence of a concept attribute in an image based on a threshold value derived from an experiment. If the value $s_i$ exceeds this threshold, we consider the image to possess that specific concept attribute and set the concept label to be True. We + +can obtain predicted concept labels for all unlabeled data. + +Alignment of Image. Based on our earlier discussions, on the one hand, the concept encoder should learn information from $\hat{c}_{img}$ . On the other hand, it should also get concept embeddings which can get good similarity-based concept labels $\hat{c}_{\text {align }}$ for alignment with the input image. Thus, we need a further alignment loss to achieve these two goals. Specifically, we compute the alignment loss as follows: + +$$ +\mathcal {L} _ {\text {a l i g n}} = B C E \left(\hat {c} _ {\text {i m g}}, \hat {c} _ {\text {a l i g n}}\right). \tag {3} +$$ + +In fact, during this alignment process, the CBM backbone is taught a reasonable latent space structure by the visual encoder, and thus we can regard it as an implicit Student-Teacher setup framework. + +# 4.3. Final Objective + +In this section, we will discuss how we derive the process of network optimization. First, we have the concept loss $\mathcal{L}_c$ in (1) to improve the interpretability of the concept embeddings. Also, since the concept embeddings are entered into the label predictor to output the final prediction, we also have a task loss between the predictions given by the concept bottleneck and ground truth, which is shown in (2). In the context of binary classification tasks, we employ binary cross-entropy (BCE) as our loss function. For multi-class classification tasks, we use cross-entropy as the measure. Finally, to align the images with the concept labels, we computed the loss of alignment in (3). Formally, the overall loss function of our approach can be formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t a s k}} + \lambda_ {1} \cdot \mathcal {L} _ {c} + \lambda_ {2} \cdot \mathcal {L} _ {\text {a l i g n}}, \tag {4} +$$ + +where $\lambda_1, \lambda_2$ are hyperparameters for the trade-off between interpretability and accuracy. + +# 5. Experiments + +In this section, we will conduct experimental studies on the performance of our framework. Specifically, we evaluate the utility of our method according to the concept and prediction of the class label. We also design interpretability evaluation and test-time intervention to verify the interpretability and alignment performance of our method. Finally, we give the ablation study to verify why we need the alignment loss. Details and additional results are in Appendix A. + +# 5.1. Experimental Settings + +Datasets. We evaluate our methods on four real-world image tasks: CUB[14], AwA2[33], WBCatt[42] and 7-point[22]. See Appendix A.1 for a detailed introduction. + +Baselines. We compare our SSCBM with Concept Bottle-neck Model (CBM) and Concept Embedding Model (CEM) + +Table 1. Concept and task accuracy results of SSCBM at a labeled ratio of 0.1, compared to baselines in the fully supervised setting. + +
MethodCUBAwA2WBCatt7-point
ConceptTaskConceptTaskConceptTaskConceptTask
CBM93.99%67.33%96.48%88.71%94.18%99.71%74.34%75.44%
CEM96.39%79.82%95.91%87.00%95.33%99.71%77.15%75.85%
SSCBM90.88%67.67%96.48%89.77%93.98%99.68%73.67%70.09%
+ +mentioned in Section 3. Since those baselines do not inherently include settings for semi-supervised learning, to ensure fairness in the evaluation, we use the same pseudolabeling approach as SSCBM. We utilize the KNN algorithm to annotate unlabeled data based on labeled data, which is mentioned in Section 4.2. For CBM [25], we adopt the same setting and architecture as in the original CBM. And for CEM, we follow the same setting as in [9]. We also compare label-free CBM [34], which is an unsupervised CBM method. + +Evaluation Metrics. To evaluate the utility, we consider the accuracy for the prediction of class and concept labels. Specifically, concept accuracy measures the model's prediction accuracy for concepts: $\mathcal{C}_{acc} = \frac{1}{N}\cdot \frac{1}{k}\sum_{i = 1}^{N}\sum_{j = 1}^{k}\mathbb{I}(\hat{c}_j^{(i)} = c_j^{(i)})$ . Task accuracy measures the model's performance in predicting downstream task classes: $\mathcal{A}_{acc} = \frac{1}{N}\sum_{i = 1}^{N}\mathbb{I}(\hat{\pmb{y}}^{(i)} = \pmb {y}^{(i)})$ + +To evaluate the interpretability, besides the concept accuracy, similar to previous work [9, 25], we show visualization results. Moreover, we evaluate the performance of the test-time intervention. + +Implementation Details. All experiments are conducted on a Tesla V100S PCIe 32 GB GPU and an Intel Xeon Processor CPU. See Appendix A.2 for more details. + +# 5.2. Evaluation Results on Utility + +To comprehensively evaluate the model's performance in the semi-supervised learning setting, we simulate real-world application scenarios. We consider labeling K samples for each class in the dataset, with an extreme case where $\mathrm{K} = 1$ . Clearly, $\mathrm{K} =$ total data size $\times$ labeled ratio $\div$ number of classes. We also conduct experiments with labeled ratios of 0.05, 0.1, 0.15, and 0.2. We also compare the performance of SSCBM with the baseline under the fully supervised setting. The experimental results are shown in Table 1 and 2. Additionally, we also explored the impact of different backbones on model performance, which can be found in the Appendix A.3. + +Table 1 shows the performance comparison of SSCBM at a labeled ratio of 0.1 with CEM and CBM under the fully supervised setting. It can be observed that, although CEM performs the best on CUB, WBCatt and 7-point, SSCBM is only slightly inferior to it. On the AwA2 dataset, its performance even surpasses that of CEM under the fully super + +Table 2. Results of concept and task accuracy for different datasets with different portions of labeled data. Since the label-free CBM [34] and LaBo [47] do not use the concept set included in the dataset but instead rely on GPT-3 to generate concepts based on class names, we do not report the concept accuracy here. + +
DatasetLabeled RatioCBM+SSLCEM+SSLLabel-free CBMLaBoSSCBM
ConceptTaskConceptTaskConceptTaskConceptTaskConceptTask
CUBK=183.11%5.51%82.36%59.35%-74.31%-66.38%88.99%66.72%
0.05 (K=2)84.51%8.35%83.72%62.20%-74.31%-66.38%90.04%67.43%
0.1 (K=3)84.96%9.84%84.03%63.12%-74.31%-66.38%90.88%67.67%
0.15 (K=4)85.47%9.96%84.30%64.14%-74.31%-66.38%91.47%68.36%
0.2 (K=5)86.67%16.43%86.83%67.64%-74.31%-66.38%92.09%70.07%
AWA2K=171.32%90.87%65.55%84.44%-73.67%-93.73%85.05%89.02%
0.05 (K=30)77.23%91.11%68.31%85.01%-73.67%-93.73%95.42%89.37%
0.1 (K=60)80.45%91.03%69.96%85.48%-73.67%-93.73%96.48%89.77%
0.15 (K=90)83.46%91.00%72.14%86.61%-73.67%-93.73%96.81%90.54%
0.2 (K=120)85.76%91.12%72.14%86.61%-73.67%-93.73%96.81%90.44%
WBCattK=179.06%99.39%70.27%98.64%-41.92%-92.26%91.48%99.13%
0.05 (K=62)81.08%99.48%73.82%99.52%-41.92%-92.26%93.53%99.61%
0.1 (K=124)85.48%99.32%72.25%99.29%-41.92%-92.26%93.98%99.68%
0.15 (K=186)85.39%99.68%72.68%99.58%-41.92%-92.26%94.42%99.71%
0.2 (K=247)87.07%99.74%74.14%99.52%-41.92%-92.26%94.42%99.71%
7-pointK=159.91%55.95%62.78%66.09%-55.44%-64.97%66.58%66.84%
0.05 (K=5)65.36%57.47%67.85%67.09%-55.44%-64.97%70.98%68.77%
0.1 (K=9)68.82%55.70%72.23%66.33%-55.44%-64.97%73.67%70.09%
0.15 (K=13)66.14%59.75%66.54%67.09%-55.44%-64.97%73.94%72.56%
0.2 (K=17)70.29%60.44%73.04%66.84%-55.44%-64.97%76.52%74.56%
+ +vised setting, which sufficiently demonstrates the superiority of our method. + +In Table 2, it can be observed that as the proportion of labeled data gradually increases, the concept and class prediction accuracy of all models improve. SSCBM outperforms all baselines. On the CUB dataset, when only one labeled sample per class is available, SSCBM achieves $88.99\%$ concept accuracy and $66.72\%$ class accuracy, which are $6.63\%$ and $7.37\%$ higher, respectively, than the performance achieved by CEM in a semi-supervised learning setting. This demonstrates the effectiveness of the Image-Textual Semantics Alignment module. The 7-point dataset is the most challenging, and on this dataset, SSCBM still achieves the best performance across all settings. As the amount of labeled data increases, the concept prediction accuracy improves from $66.58\%$ to $76.52\%$ . We also note that on WBCatt and 7-point, the performance of Label-free CBM is poor. This is mainly because both datasets contain only five classes, which prevents Label-free CBM from fully leveraging the world knowledge of LLMs to generate an effective concept set, resulting in poor performance. This further demonstrates that semi-supervised learning scenarios are more realistic and practical for real-world datasets. + +# 5.3. Interpretability Evaluation + +Besides the concept accuracy, note that in SSCBM, we also have a pseudo label revived by the alignment between the concept embedding and the input saliency map. We use the alignment loss to inherit this alignment. Thus, we will evaluate the alignment performance here to show the faithfulness of the interpretability given by SSCBM. We measure + +our alignment performance by comparing the correctness of the concept saliency map with the concept semantics in Figure 3. See Appendix A.2 for detailed experimental procedures for generating saliency maps. Results show that our concept saliency map matches the concept semantics, indicating the effectiveness of our alignment loss. In Appendix C, we also provide our additional interpretability evaluation in Figure 6-18. + +# 5.4. Test-time Intervention + +Test-time intervention enables human users to interact with the model at inference time. We conduct this experiment by correcting the ratio $10\%$ to $100\%$ of the concept labels in the concept predictor. We expect that the model performance will steadily increase along with the ratio of concept intervention, indicating that the model learned such correct label information and automatically corrected other labels. + +The results in Figure 4 demonstrate the robustness of our model and an increasing trend to learn the information about the concept, indicating our interpretability and model prediction performance. During the test phase, the SSCBM model generates probability scores for each concept (e.g., has_wing_color :: brown). Prior to intervention, we employ the COOP algorithm (following CEM) to identify the most critical concepts exhibiting both high predictive uncertainty and task importance. For these selected concepts, we replace the model's predicted values with their corresponding ground truth annotations. This rectified concept vector is subsequently fed into the classification head to produce the final task prediction. Finally, we successfully cause the model to predict the Great Crested Flycatcher instead of the + +![](images/7da7de8127c7047a23aa8346f9d1cd821ee187972096c1a707cf2055e561d899.jpg) +(a) Original Image + +![](images/517c7b3f9b502cc51979ed2a51369b5e65804b7e6cd043eca99295e0cb1c70c5.jpg) +(b) Tail Pattern: Multi-colored +Figure 3. The concept saliency map for the CUB dataset (savannah sparrow) demonstrates that our proposed SSCBM achieves meaningful alignment between the ground truth concepts and the input image features. The first image on the left is the original input image. The three images on the right show the aligned regions for different concepts using SSCBM. + +![](images/a2040c5184d8c90b24989582fa87a346ddb8e7a58011c416fdc7e64d5ca0dddc.jpg) +(c) Underparts Color: Brown + +![](images/4af6004c8beb3e17a59de06ead3d8a56a8919aadb615af41df688166d676b69c.jpg) +(d) Wing Color: Brown + +![](images/2aa34b9d7a97a501ccc0a0a8a04eb3af6137337b6aabf3b19604f4fa466d22b3.jpg) + +![](images/44fcb677be76cff0bb1c49531ccfc493519c5b768fa862338613e9ba8d7d30ab.jpg) +Figure 4. Left: Performance with different ratios of intervened concepts on CUB dataset. Right: An example of successful intervention. + +![](images/4eec13b2b797ac0efabcf68556e47c5d7e5bd320be4078f670030c3445ee637d.jpg) + +Swainson Warbler. More results are in Appendix B. + +Table 3. Results of ablation study. + +
Ratiow/o imgw/o alignfull model
ConceptTaskConceptTaskConceptTask
CUB0.0001 (K=1)87.77%59.89%80.87%65.15%88.99%66.72%
0.05 (K=2)90.01%65.98%80.60%65.67%90.04%67.43%
0.1 (K=3)91.02%67.28%80.60%64.69%90.88%67.67%
0.15 (K=4)90.60%67.97%80.65%65.43%91.47%68.36%
0.2 (K=5)91.94%68.99%80.73%64.45%92.09%70.07%
WBCatt0.0001 (K=1)88.44%99.68%57.80%99.51%91.48%99.13%
0.05 (K=62)92.17%99.29%57.93%99.51%93.53%99.61%
0.1 (K=124)92.91%99.52%60.49%99.64%93.98%99.68%
0.15 (K=186)93.55%99.58%59.13%99.65%94.42%99.71%
0.2 (K=247)94.19%99.61%60.65%99.70%94.42%99.71%
7-point0.0001 (K=1)62.31%64.30%55.07%66.33%66.58%66.84%
0.05 (K=5)69.42%65.06%56.75%68.10%70.98%68.77%
0.1 (K=9)70.06%67.34%56.07%65.82%73.67%70.09%
0.15 (K=13)73.39%68.10%56.23%65.82%73.94%72.56%
0.2 (K=17)74.44%68.61%56.11%65.57%76.52%74.56%
+ +# 5.5. Ablation Study + +We then give a finer study on the two kinds of pseudo-labels to demonstrate that each one is indispensable in bolstering the efficacy of SSCBM. Specifically, based on our method in Section 4.2, we have three types of pseudo labels available: pseudo concept label $\hat{c}_{img}$ , $\hat{c}_{align}$ , and label $\hat{c}$ predicted by the concept embedding. We first remove $\hat{c}_{img}$ and calculate our alignment loss using $\hat{c}_{align}$ and $\hat{c}$ ; conversely, we remove $\hat{c}_{align}$ and calculate the alignment loss using $\hat{c}_{img}$ and $\hat{c}$ . + +From the results presented in Table 3, it is evident that removing the $\hat{c}_{img}$ component significantly degrades the performance of SSCBM at both the concept level and class level. This indicates that the pseudo-concept labels via KNN contain necessary information about the ground truth. The concept encoder needs to extract information from such labels to get better performance. This situation is similar when removing the $\hat{c}_{align}$ component, indicating that aligning the concept embedding and the input saliency map can further extract useful information from the input image and, thus, is beneficial to improve the performance. Our observations underscore the high degree of joint effectiveness of two kinds of pseudo-concept labels within our objective function, collectively contributing to the enhancement of model prediction and concept label prediction. + +# 6. Conclusion + +The training of current CBMs heavily relies on the accuracy and richness of annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. In this problem, we propose SSCBM, a strategy to generate pseudo labels and an alignment loss to solve these two problems. Results show our effectiveness. + +# Acknowledgements + +This work is supported in part by the funding BAS/1/1689-01-01, URF/1/4663-01-01, REI/1/5232-01-01, REI/1/5332-01-01, and URF/1/5508-01-01 from KAUST, and funding from KAUST - Center of Excellence for Generative AI, under award number 5940. + +# References + +[1] Andrea Acevedo, Santiago Alférez, Anna Merino, Laura Puigvi, and José Rodellar. Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine, 180:105020, 2019. 11 +[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 3 +[3] Pietro Barbiero, Gabriele Ciravegna, Francesco Giannini, Mateo Espinosa Zarlenga, Lucie Charlotte Magister, Alberto Tonda, Pietro Lio, Frédéric Precioso, Mateja Jamnik, and Giuseppe Marra. Interpretable neural-symbolic concept reasoning. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1801-1825. PMLR, 2023. 2 +[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[5] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2 +[6] Asma Chebli, Akila Djebar, and Hayet Farida Marouani. Semi-supervised learning for medical application: A survey. In 2018 International Conference on Applied Smart Systems (ICASS), pages 1-9, 2018. 3 +[7] Xing Chen, Biao Ren, Ming Chen, Quanxin Wang, Lixin Zhang, and Guiying Yan. Nllss: predicting synergistic drug combinations based on semi-supervised learning. PLoS computational biology, 12(7):e1004975, 2016. 3 +[8] Veronika Cheplygina, Marleen de Bruijnne, and Josien P.W. Pluim. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis, 54:280-296, 2019. 3 +[9] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 3, 4, 6, 11 +[10] Roman Filipovych and Christos Davatzikos. Semi-supervised pattern classification of medical images: Appli + +cation to mild cognitive impairment (mci). NeuroImage, 55 (3):1109-1119, 2011. 3 +[11] Jack Furby, Daniel Cunningham, Dave Braines, and Alun Preece. Can we constrain concept bottleneck models to learn semantically meaningful input features? arXiv preprint arXiv:2402.00912, 2024. 2 +[12] Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. Advances in Neural Information Processing Systems, 35:23386-23397, 2022. 2 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 11 +[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 6, 11 +[15] Junlin Hou, Jilan Xu, and Hao Chen. Concept-attention whitening for interpretable skin lesion diagnosis. arXiv preprint arXiv:2404.05997, 2024. 2 +[16] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Improving faithfulness for vision transformers. arXiv preprint arXiv:2311.17983, 2023. 1 +[17] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Seat: stable and explainable attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12907-12915, 2023. +[18] Lijie Hu, Liang Liu, Shu Yang, Xin Chen, Hongru Xiao, Mengdi Li, Pan Zhou, Muhammad Asif Ali, and Di Wang. A hopfieldian view-based interpretation for chain-of-thought reasoning. arXiv preprint arXiv:2406.12255, 2024. 1 +[19] Lijie Hu, Chenyang Ren, Zhengyu Hu, Cheng-Long Wang, and Di Wang. Editable concept bottleneck models. arXiv preprint arXiv:2405.15476, 2024. 2 +[20] Tri Huynh, Aiden Nibali, and Zhen He. Semi-supervised learning for medical image classification using imbalanced training data. Computer Methods and Programs in Biomedicine, 216:106628, 2022. 3 +[21] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In The Twelfth International Conference on Learning Representations, 2023. 3 +[22] Jeremy Kawahara, Sara Daneshvar, Giuseppe Argenziano, and Ghassan Hamarneh. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE Journal of Biomedical and Health Informatics, 23(2): 538-546, 2019. 6, 11 +[23] Mert Keser, Gesina Schwalbe, Azarm Nowzad, and Alois Knoll. Interpretable model-agnostic plausibility verification for 2d object detectors using domain-invariant concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890–3899, 2023. 2 +[24] Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sunghroh Yoon. Probabilistic concept bottleneck models. arXiv preprint arXiv:2306.01574, 2023. 2 + +[25] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6 +[26] Songning Lai, Lijie Hu, Junxiao Wang, Laure Berti-Equille, and Di Wang. Faithful vision-language interpretation via concept bottleneck models. In The Twelfth International Conference on Learning Representations, 2023. 1 +[27] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.3 +[28] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 3 +[29] Jia Li, Lijie Hu, Zhixian He, Jingfeng Zhang, Tianhang Zheng, and Di Wang. Text guided image editing with automatic concept locating and forgetting. arXiv preprint arXiv:2405.19708, 2024. 2 +[30] Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. Large language models in finance: A survey. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 374-382, 2023. 1 +[31] Lucie Charlotte Magister, Dmitry Kazhdan, Vikash Singh, and Pietro Lio. Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks. arXiv preprint arXiv:2107.11889, 2021. 2 +[32] Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47-60, 1996. 3 +[33] Li Niu, Qingtao Tang, Ashok Veeraraghavan, and Ashutosh Sabharwal. Learning from noisy web data with category-level supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7689-7698, 2018. 6, 11 +[34] Tuomas Oikarinen, Subhro Das, Lam Nguyen, and Lily Weng. Label-free concept bottleneck models. In International Conference on Learning Representations, 2023. 1, 2, 6, 7 +[35] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv preprint arXiv:2006.05278, 2020. 3 +[36] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11557-11568, 2021. 3 +[37] Yoshihide Sawada and Keigo Nakamura. Concept bottleneck model with additional unsupervised concepts. IEEE Access, 10:41758-41765, 2022. 2 +[38] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. + +Advances in neural information processing systems, 33:596-608, 2020. 3 +[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3 +[41] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 1 +[42] Satoshi Tsutsui, Winnie Pang, and Bihan Wen. Wbcatt: A white blood cell dataset annotated with detailed morphological attributes. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6, 11 +[43] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 3 +[44] Han Xuanyuan, Pietro Barbiero, Dobrik Georgiev, Lucie Charlotte Magister, and Pietro Lio. Global concept-based interpretability for graph neural networks via neuron analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10675-10683, 2023. 2 +[45] An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amilcare Gentili, Chun-Nan Hsu, Jingbo Shang, et al. Robust and interpretable medical image classifiers via concept bottleneck models. arXiv preprint arXiv:2310.03182, 2023. 5 +[46] Shu Yang, Lijie Hu, Lu Yu, Muhammad Asif Ali, and Di Wang. Human-ai interactions in the communication era: Autophagy makes large models achieving local optima. arXiv preprint arXiv:2402.11271, 2024. 1 +[47] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification, 2023. 7 +[48] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2 +[49] Gang Zhang, Shan-Xing Ou, Yong-Hui Huang, and Chun-Ru Wang. Semi-supervised learning methods for large scale healthcare data analysis. International Journal of Computers in Healthcare, 2(2):98-110, 2015. 3 \ No newline at end of file diff --git a/2025/Semi-supervised Concept Bottleneck Models/images.zip b/2025/Semi-supervised Concept Bottleneck Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c1cbeff72a080c7497544595f7b054b5313c81b --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bba941496d1ad5af1a59522861e5bb9be2eb8ac3cf5ffbe0df1544b1f503efc +size 522815 diff --git a/2025/Semi-supervised Concept Bottleneck Models/layout.json b/2025/Semi-supervised Concept Bottleneck Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc3e553b12ae1a05c1792cbb2a29794e417123bc --- /dev/null +++ b/2025/Semi-supervised Concept Bottleneck Models/layout.json @@ -0,0 +1,9747 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "spans": [ + { + "bbox": [ + 167, + 103, + 443, + 121 + ], + "type": "text", + "content": "Semi-supervised Concept Bottleneck Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 471, + 171 + ], + "type": "text", + "content": "Lijie Hu\\*,1,2, Tianhao Huang\\*,1,2,3, Huanyi Xie\\*,2, Xin Gong\\*,1,2,4, Chenyang Ren\\*,2, Zhengyu Hu\\*,6, Lu Yu\\*, Ping Ma\\*, Di Wang\\*,1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 172, + 453, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 172, + 453, + 185 + ], + "spans": [ + { + "bbox": [ + 157, + 172, + 453, + 185 + ], + "type": "text", + "content": "1Provable Responsible AI and Data Analytics (PRADA) Lab," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 185, + 437, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 185, + 437, + 199 + ], + "spans": [ + { + "bbox": [ + 171, + 185, + 437, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 171, + 185, + 437, + 199 + ], + "type": "text", + "content": "King Abdullah University of Science and Technology," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 199, + 487, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 199, + 487, + 213 + ], + "spans": [ + { + "bbox": [ + 124, + 199, + 487, + 213 + ], + "type": "text", + "content": "3University of Virginia, 4University of Georgia, 5Ant Group, 6HKUST(GZ)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 213, + 516, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 213, + 516, + 228 + ], + "spans": [ + { + "bbox": [ + 94, + 213, + 516, + 228 + ], + "type": "text", + "content": "{lijie.hu, tianhao.huang, huanyi.xie, di.wang}@kaust.edu.sa" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "text", + "content": "Concept Bottleneck Models (CBMs) have garnered increasing attention due to their ability to provide concept-based explanations for black-box deep learning models while achieving high final prediction accuracy using human-like concepts. However, the training of current CBMs is heavily dependent on the precision and richness of the annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. To address these limitations, we propose a new framework called SsCBM (Semi-supervised Concept Bottleneck Model). Our SSCBM is suitable for practical situations where annotated data is scarce. By leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level, we effectively solve these issues. We propose a strategy to generate pseudo labels and an alignment loss. Experiments demonstrate that our SSCBM is both effective and efficient. With only " + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "text", + "content": " labeled data, our model's concept and task accuracy on average across four datasets is only " + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "inline_equation", + "content": "2.44\\%" + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "inline_equation", + "content": "3.93\\%" + }, + { + "bbox": [ + 55, + 280, + 297, + 568 + ], + "type": "text", + "content": " lower, respectively, compared to the best baseline in the fully supervised learning setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 684 + ], + "type": "text", + "content": "Deep learning models like ResNet [13] typically feature complex non-linear architectures, making it difficult for end-users to understand and trust their decisions. This lack of interpretability is a significant obstacle to their adoption, especially in critical fields such as healthcare [41] and finance [30], where transparency is crucial. Explain-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 399 + ], + "type": "text", + "content": "able artificial intelligence (XAI) models have been developed to meet the demand for transparency, providing insights into their behavior and internal mechanisms [16-18, 46]. Concept Bottleneck Models (CBMs) [25] are particularly notable among these XAI models for their ability to clarify the prediction process of end-to-end AI models. CBMs introduce a bottleneck layer that incorporates human-understandable concepts. During prediction, CBMs first predict concept labels from original input, then use these predicted concepts in bottleneck layer to determine the final label. This approach results in a self-explanatory decision-making process that users can comprehend." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 401, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 556, + 640 + ], + "type": "text", + "content": "A major issue in original CBMs is the need for expert labeling, which is costly in practice. Some researchers address this problem through unsupervised learning. For example, [34] proposes a Label-free CBM that transforms any neural network into an interpretable CBM without requiring labeled concept data while maintaining high accuracy. Similarly, Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. However, these methods have three issues. First, those unsupervised methods heavily rely on large language models like GPT-3, which have reliability issues [26]. Second, the concepts extracted by these models lack evaluation metrics, undermining their interpretability. Third, the assumption that no concept labels are available is too stringent in practice. In reality, obtaining a small portion of concept labels is feasible and cost-effective. Therefore, we can maximize the use of this small labeled concept dataset. This is the motivation for introducing our framework, which focuses on the semi-supervised setting in CBM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In this paper, we introduce a framework called the SS-CBM (Semi-supervised Concept Bottleneck Model). Compared to a supervised setting, semi-supervised CBMs have two main challenges. First, obtaining concept embeddings requires concept labels, so we need to generate pseudo labels to obtain these concept embeddings. To achieve this," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2110" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 76, + 178, + 184 + ], + "blocks": [ + { + "bbox": [ + 90, + 76, + 178, + 184 + ], + "lines": [ + { + "bbox": [ + 90, + 76, + 178, + 184 + ], + "spans": [ + { + "bbox": [ + 90, + 76, + 178, + 184 + ], + "type": "image", + "image_path": "0b72e38c5fb038035c898ebcbac7c7710816c65463a2dfd60d4193e6c0347496.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 187, + 222, + 198 + ], + "lines": [ + { + "bbox": [ + 95, + 187, + 222, + 198 + ], + "spans": [ + { + "bbox": [ + 95, + 187, + 222, + 198 + ], + "type": "text", + "content": "(a) Label Complete and Well-aligned" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 72, + 334, + 185 + ], + "blocks": [ + { + "bbox": [ + 238, + 72, + 334, + 185 + ], + "lines": [ + { + "bbox": [ + 238, + 72, + 334, + 185 + ], + "spans": [ + { + "bbox": [ + 238, + 72, + 334, + 185 + ], + "type": "image", + "image_path": "325aa5b1fdfc33c39845046e8bf28a03b67bdedec9aa962d5cfba278dbfe15df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 555, + 242 + ], + "type": "text", + "content": "Figure 1. (a) A sample of sparrow class with complete concept labels. (b) A sample of sparrow class with incomplete concept labels. (c) A sample of misalignment between input features and concepts resulting from existing CBM methods. Our framework simultaneously utilizes both (a) and (b) types of data and addresses the issue of (c) through an alignment loss." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 335, + 76, + 375, + 184 + ], + "blocks": [ + { + "bbox": [ + 270, + 187, + 342, + 198 + ], + "lines": [ + { + "bbox": [ + 270, + 187, + 342, + 198 + ], + "spans": [ + { + "bbox": [ + 270, + 187, + 342, + 198 + ], + "type": "text", + "content": "(b) Label Incomplete" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 76, + 375, + 184 + ], + "lines": [ + { + "bbox": [ + 335, + 76, + 375, + 184 + ], + "spans": [ + { + "bbox": [ + 335, + 76, + 375, + 184 + ], + "type": "image", + "image_path": "71d036ecdeafabac10ff570578e7bb05c5b079de45ffef4b7a1ec66a10413296.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 482, + 185 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 482, + 185 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 482, + 185 + ], + "type": "image", + "image_path": "475c52110745db85f13d30e043432371fc73219f227b4f083ed782ef3f6af163.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 187, + 479, + 198 + ], + "lines": [ + { + "bbox": [ + 429, + 187, + 479, + 198 + ], + "spans": [ + { + "bbox": [ + 429, + 187, + 479, + 198 + ], + "type": "text", + "content": "(c) Misaligned" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 485, + 77, + 522, + 184 + ], + "blocks": [ + { + "bbox": [ + 485, + 77, + 522, + 184 + ], + "lines": [ + { + "bbox": [ + 485, + 77, + 522, + 184 + ], + "spans": [ + { + "bbox": [ + 485, + 77, + 522, + 184 + ], + "type": "image", + "image_path": "54b430661fa18cc079ecd653e1b1ee341bed310aee4c5febc1f047f0a4e55ce5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 262, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 296, + 584 + ], + "type": "text", + "content": "SSCBM uses a KNN-based algorithm to assign pseudo-concept labels for unlabeled data. Second, while such a simple pseudo-labeling method is effective and has acceptable classification accuracy, we also find that the concept saliency map often misaligns with the input saliency map, meaning concept predictions frequently correspond to irrelevant input features. This misalignment often arises from inaccurate concept annotations or unclear relationships between input features and concepts, which is closely related to the broader issue of annotation alignment. In fact, in the supervised setting, there is a similar misalignment issue [11]. Existing research seeks to improve alignment by connecting textual and image information [15]. However, these methods only focus on supervised settings and cannot be directly applied to our settings, as our pseudo-labels contain noise. Our framework achieves excellent performance in both concept accuracy and concept saliency alignment by leveraging joint training on both labeled and unlabeled data and aligning the unlabeled data at the concept level. To achieve this, we leverage the relevance between the input image and the concept and get other pseudo-concept labels based on these similarity scores. Finally, we align these two types of pseudo-concept labels to give the concept encoder the ability to extract useful information from features while also inheriting the ability to align concept embeddings with the input. See Figure 1 for an illustration. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "- We propose SSCBM, a framework designed to address both the semi-supervised annotation problem and the concept semantics alignment problem, which hold significant practical importance in real-world applications. SSCBM tackles them in an elegant and efficient manner. To the best of our knowledge, we are the first to tackle these two problems within a single framework, elucidating the behavior of CBMs through semi-supervised alignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate the superiority of our SSCBM in annotation and concept-saliency align" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "text", + "content": "ment, indicating its efficiency and effectiveness. Using only " + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "text", + "content": " labeled data, our model achieves concept and task accuracy on average across the four datasets that are merely " + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "inline_equation", + "content": "2.44\\%" + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "inline_equation", + "content": "3.93\\%" + }, + { + "bbox": [ + 321, + 262, + 555, + 323 + ], + "type": "text", + "content": " lower, respectively, than the best baseline under fully supervised learning settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 400, + 346 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": "Concept Bottleneck Models. Concept Bottleneck Model (CBM) [25] is an innovative deep-learning approach for image classification and visual reasoning. By introducing a concept bottleneck layer into deep neural networks, CBMs enhance model generalization and interpretability by learning specific concepts. However, CBMs face two primary challenges: their performance often lags behind that of models without the bottleneck layer due to incomplete information extraction, and they rely heavily on laborious dataset annotation. Researchers have explored various solutions to these challenges. [5] extends CBMs into interactive prediction settings by introducing an interaction policy to determine which concepts to label, thereby improving final predictions. [34] addresses CBM limitations by proposing a Label-free CBM, which transforms any neural network into an interpretable CBM without requiring labeled concept data, maintaining high accuracy. Post-hoc Concept Bottleneck models [48] can be applied to various neural networks without compromising performance, preserving interpretability advantages. Related work in the image domain includes studies [12, 19, 23, 24, 29, 37, 38]. In the graph concept field, [31] provides a global interpretation of graph neural networks (GNNs) by mapping graphs into a concept space through clustering and offering a human-in-the-loop evaluation. [3, 44] extend this approach by incorporating both global and local explanations. For local explanations, they define a concept set, with each neuron represented as a vector with Boolean values indicating concept activation. However, existing works rarely consider semi-supervised settings, which are practical in real-world" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2111" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "applications. Our framework addresses these issues effectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 574 + ], + "type": "text", + "content": "Semi-supervised Learning. Semi-supervised learning (SSL) combines the two main tasks of machine learning: supervised learning and unsupervised learning [43]. It is typically applied in scenarios where labeled data is scarce. Examples include computer-aided diagnosis [6, 49], medical image analysis [8, 10, 20], and drug discovery [7]. In these cases, collecting detailed annotated data by experts requires considerable time and effort. However, under the assumption of data distribution, unlabeled data can also assist in building better classifiers [43]. SSL, also known as self-labeling or self-teaching in its earliest forms, involves the model iteratively labeling a portion of unlabeled data and adding it to the training set for the next round of training [35]. The expectation-maximization (EM) algorithm proposed by [32] uses both labeled and unlabeled data to produce maximum likelihood estimates of parameters. [27] and [40] focus on consistency regularization. II-model [27] combines both supervised cross-entropy loss and unsupervised consistency loss while perturbing the model and data based on the consistency constraint assumption. A temporal ensembling model integrates predictions from models at various time points. Mean teacher [40] addresses the slow updating issue of the temporal ensembling model on large datasets by averaging model weights instead of predicting labels. MixMatch [4] unifies and refines the previous approaches of consistency regularization, entropy minimization, and traditional regularization into a single loss function, achieving excellent results. Pseudo labeling, as an effective tool for reducing the entropy of unlabeled data [28], has been increasingly attracting the attention of researchers in the field of semi-supervised learning. [2] proposes that directly using the model's predictions as pseudo-labels can achieve good results. FixMatch [39] predicts and retains the model, generating high-confidence pseudo-labels. [36] continuously adjusts the teacher based on feedback from the student, thereby generating better pseudo-labels. While there has been a plethora of work in the semi-supervised learning field, the focus on semi-supervised concept bottleneck models remains largely unexplored. Our work focuses on this new area." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 586, + 141, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 141, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 141, + 597 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Concept Bottleneck Models [25]. We consider a classification task with a concept set denoted as " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{p_1,\\dots ,p_k\\}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " with each " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " being a concept given by experts or LLMs, and a training dataset represented " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ". Here, for " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "i\\in [N]" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "x^{(i)}\\in \\mathcal{X}\\subseteq \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " represents the feature vector (e.g., an image's pixels), " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "y^{(i)}\\in \\mathcal{Y}\\subseteq \\mathbb{R}^l" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " denotes the label (" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " is the number of classes), " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "c^{(i)} = (c_i^1,\\dots ,c_i^k)\\in \\mathbb{R}^k" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " represents the concept vector (a binary vector of length " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ", where each value indicates whether the input " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "x^{(i)}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " contains" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "that concept). In CBMs, the goal is to learn two representations: one called a concept encoder that transforms the input space into the concept space, denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "g: \\mathbb{R}^d \\to \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", and another called label predictor that maps the concept space to the downstream prediction space, denoted as " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^k \\to \\mathbb{R}^l" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ". Usually, the map " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is linear. For any input " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", we aim to ensure that its predicted concept vector " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\hat{c} = g(x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " and prediction " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\hat{y} = f(g(x))" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " are close to their underlying counterparts, thus capturing the essence of the original CBMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": "Concept Embedding Models [9]. As the original CBM is based solely on concept features to determine the predictions of the model, compared to canonical deep neural networks, it will degrade prediction performance. To further improve the performance of CBMs, CEM is developed by [9]. It achieves this by learning interpretable high-dimensional concept representations (i.e., concept embeddings), thus maintaining high task accuracy while obtaining concept representations that contain meaningful semantic information. For CEMs, we use the same setting as that of [9, 21]. For each input " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": ", the concept encoder learns " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " concept representations " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\hat{c}_1, \\hat{c}_2, \\dots, \\hat{c}_k" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": ", each corresponding to one of the " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " ground truth concepts in the training dataset. In CEMs, each concept " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " is represented using two embeddings " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^+, \\hat{c}_i^- \\in \\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": ", each with specific semantics, i.e., the concept is True (activate state) and concept is False (negative state), where hyper-parameter " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " is the embedding dimension. We use a DNN " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\psi(x)" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " to learn a latent representation " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "h \\in \\mathbb{R}^{n_h}" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": ", which is used as input to the CEM embedding generator, where " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "n_h" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " is the dimension of the latent representation. The CEM embedding generator " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " feeds " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": " into two fully connected concept-specific layers to learn two concept embeddings in " + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 180, + 555, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 372, + 465, + 496, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 465, + 496, + 479 + ], + "spans": [ + { + "bbox": [ + 372, + 465, + 496, + 479 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} _ {i} = \\phi_ {i} (\\boldsymbol {h}) = a (W _ {i} \\boldsymbol {h} + \\boldsymbol {b} _ {i}).", + "image_path": "8ec36f2b8129cad66c5bb09f1d3d200d2df03e16a00ab6859f2e2043963dff5b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": "Then we use a differential scoring function " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "s: \\mathbb{R}^{2m} \\to [0,1]" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": ", to achieve the alignment of concept embeddings " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{c}}_i^+" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{c}}_i^-" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": " and ground-truth concepts " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": ". It is trained to predict the probability " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "\\hat{p}_i := s\\left(\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right) = \\sigma\\left(W_s\\left[\\left[\\hat{\\pmb{c}}_i^+, \\hat{\\pmb{c}}_i^-\\right]^\\top\\right] + \\pmb{b}_s\\right)" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": " that the concept " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": " is active in the embedding space. We get the final concept embedding " + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{c}}_i" + }, + { + "bbox": [ + 313, + 487, + 555, + 584 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 378, + 594, + 489, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 594, + 489, + 608 + ], + "spans": [ + { + "bbox": [ + 378, + 594, + 489, + 608 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} _ {i} := \\hat {p} _ {i} \\hat {\\boldsymbol {c}} _ {i} ^ {+} + (1 - \\hat {p} _ {i}) \\hat {\\boldsymbol {c}} _ {i} ^ {-}.", + "image_path": "ffcdfb34a04a9ad9d0ed97cb2b619a124115d68f51608b5e8ab74c41d89d40fc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 677 + ], + "type": "text", + "content": "At this point, we understand that we can obtain high-quality concept embeddings rich in semantics through CEMs. In the subsequent section 4, we will effectively utilize these representations of concepts and further optimize their interpretability through our proposed framework SSCBM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Semi-supervised Setting. Now, we consider the setting of semi-supervised learning for concept bottleneck models. As mentioned earlier, a typical training dataset for CBMs" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2112" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": "can be represented as " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "x^{(i)}\\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " represents the input feature. However, in semisupervised learning tasks, the set of feature vectors typically consists of two parts, " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathcal{X}_L,\\mathcal{X}_U\\}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_L" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " represents a small subset of labeled data and " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_U" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " represents the remaining unlabeled data. Here, the unlabeled data refers to data where the class label " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " is known, while the concept label " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " is unknown. Generally we assume " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "|\\mathcal{X}_L|\\ll |\\mathcal{X}_U|" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": ". We assume that " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "x^{(j)}\\in \\mathcal{X}_L" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " is labeled with a concept vector " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "c^{(j)}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": " and a label " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "y^{(j)}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": ", and for any " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "x^{(i)}\\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": ", there only exists a corresponding label " + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "inline_equation", + "content": "y^{(i)}\\in \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 71, + 296, + 203 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": "Under these settings, given a training dataset " + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\mathcal{D}_L \\cup \\mathcal{D}_U" + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": " that includes both labeled and unlabeled data, the goal is to train a CBM using both the labeled data " + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L" + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": " and unlabeled data " + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_U" + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": ". This aims to get better mappings " + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "inline_equation", + "content": "g: \\mathbb{R}^d \\to \\mathbb{R}^k" + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^k \\to \\mathbb{R}^l" + }, + { + "bbox": [ + 55, + 204, + 296, + 288 + ], + "type": "text", + "content": " than those trained by using only labeled data, ultimately achieving higher task accuracy and its corresponding concept-based explanation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 296, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 312 + ], + "type": "text", + "content": "4. Semi-supervised Concept Bottleneck Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": "In this section, we will elaborate on the details of the proposed SSCBM framework, which is shown in Figure 2. SSCBM follows the main idea of CEM. Specifically, to learn a good concept encoder, we use different processing methods for labeled and unlabeled data. The labeled data first passes through a feature extractor " + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": " to be transformed into a latent representation " + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": ", which then enters the concept embedding extractor to obtain the concept embeddings and the predicted concept vector " + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": " for the labeled data, which is compared to the ground truth concept to compute the concept loss. Moreover, the label predictor predicts " + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 318, + 296, + 460 + ], + "type": "text", + "content": ", and calculates the task loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": "For unlabeled data, we first extract image features " + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": " from the input using an image encoder. Then, we use the KNN algorithm to assign a pseudo-label " + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": " to each unlabeled data, which has been experimentally proven to be simple and effective. In the second step, we compute a heatmap between concept embeddings and the input. After applying a threshold, we obtain the predicted alignment label " + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": ". Finally, we compute the alignment loss between " + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 55, + 461, + 296, + 594 + ], + "type": "text", + "content": ". During each training epoch, we simultaneously compute these losses and update the model parameters based on the gradients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 290, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 290, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 290, + 613 + ], + "type": "text", + "content": "4.1. Label Anchor: Concept Embedding Encoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 676 + ], + "type": "text", + "content": "Concept Embeddings. As described in Section 3, we obtain high-dimensional concept representations with meaningful semantics based on CEMs. Thus, our concept encoder should extract useful information from both labeled and unlabeled data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "type": "text", + "content": "For labeled training data " + }, + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_L = \\{(x^{(i)},y^{(i)},c^{(i)})\\}_{i = 1}^{\\left|\\mathcal{D}_L\\right|}" + }, + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "type": "text", + "content": " we follow the original CEM [9], i.e., using a backbone network (e.g., ResNet50) to extract features " + }, + { + "bbox": [ + 55, + 676, + 296, + 713 + ], + "type": "inline_equation", + "content": "h =" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\{\\psi(x^{(i)})\\}_{i=1}^{\\left|\\mathcal{D}_L\\right|}" + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "text", + "content": ". Then, for each feature, it passes through an embedding generator to obtain concept embeddings " + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\hat{c}_i \\in \\mathbb{R}^{m \\times k}" + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "inline_equation", + "content": "i \\in [k]" + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "text", + "content": ". After passing through fully connected layers and activation layers, we obtain the predicted binary concept vector " + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\hat{c} \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 71, + 554, + 145 + ], + "type": "text", + "content": " for the labeled data. The specific process can be represented by the following expression:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 150, + 508, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 150, + 508, + 167 + ], + "spans": [ + { + "bbox": [ + 359, + 150, + 508, + 167 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} ^ {(j)} = \\psi (x ^ {(j)}), \\quad j = 1, \\ldots , | \\mathcal {D} _ {L} |,", + "image_path": "976d7af610dd2b1981e07f23732837c35ab0789cee675146bb7cc0288696b9b3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 172, + 545, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 172, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 322, + 172, + 545, + 188 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {c}} _ {i} ^ {(j)} = \\sigma (\\phi (\\boldsymbol {h} ^ {(j)})), \\quad i = 1, \\dots , k, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,", + "image_path": "4226d2b0ab3c19d6813f72051bef130239c07e4960ea122795c5a468d8cfe788.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\psi, \\phi" + }, + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 192, + 553, + 216 + ], + "type": "text", + "content": " represent the backbone, embedding generator, and activation function, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "text", + "content": "To enhance the interpretability of concept embeddings, we calculate the concept loss utilizing binary cross-entropy to optimize the accuracy of concept predictions by computing " + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "text", + "content": " based on predicted binary concept vector " + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "inline_equation", + "content": "\\hat{e}" + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "text", + "content": " and ground truth concept labels " + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 216, + 554, + 276 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 395, + 284, + 555, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 284, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 395, + 284, + 555, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = B C E (\\hat {\\boldsymbol {c}}, \\boldsymbol {c}), \\tag {1}", + "image_path": "568ffdef8274cf6f5808e0bb0c13c6954a8bc83fd0724a9b14c873334c869517.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 497, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 497, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 497, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 304, + 497, + 316 + ], + "type": "inline_equation", + "content": "BCE" + }, + { + "bbox": [ + 313, + 304, + 497, + 316 + ], + "type": "text", + "content": " is the binary cross entropy loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "text", + "content": "Task Loss. Since our ultimate task is classification, we also need to incorporate the task loss for the final prediction performance. After obtaining the predicted concept " + }, + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "text", + "content": ", we use a label predictor to predict the final class " + }, + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 313, + 316, + 554, + 389 + ], + "type": "text", + "content": ". Then we can define the task loss function using the categorical cross-entropy loss to train our classification model as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 392, + 399, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 399, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 392, + 399, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a s k}} = C E (\\hat {\\boldsymbol {y}}, \\boldsymbol {y}), \\tag {2}", + "image_path": "834111cac2dbda9a53377e089a3d2cde060f1ea593e80721d58da391902f3720.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 453 + ], + "type": "text", + "content": "Note that for the unlabeled data, we can also calculate the task loss since their class labels are known, to make full use of the data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "content": "4.2. Unlabel Alignment: Image-Textual Semantics Alignment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 572 + ], + "type": "text", + "content": "Pseudo Labeling. Unlike supervised data, CEM cannot directly extract useful information from unlabeled data, as the concept encoder is a supervised training architecture. Thus, in practical situations lacking labeled data, one direct approach is to get high-quality pseudo concept labels to train the model. In the following, we will introduce the method we use to obtain pseudo-concept labels." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "text", + "content": "Firstly, we can naturally think of measuring the similarity between images by calculating the distance in the cosine space. Based on this idea, we can assign pseudo labels to unlabeled data by finding labeled data with similar image features to them. Specifically, for each unlabeled training data " + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}_U = \\{(x^{(i)},y^{(i)})\\}_{i=1}^{|D_U|}" + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "text", + "content": ", we calculate its cosine distance with all labeled data points " + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "inline_equation", + "content": "x^{(j)} \\in \\mathcal{D}_L = \\{(x^{(j)},y^{(j)},c^{(j)})\\}_{j=1}^{|D_L|}" + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "text", + "content": " and select the " + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 573, + 554, + 683 + ], + "type": "text", + "content": " samples with the smallest distances:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 565, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 565, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 565, + 717 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} (x, x ^ {(j)}) = 1 - \\frac {\\Omega (x) \\cdot \\Omega (x ^ {(j)})}{\\| \\Omega (x) \\| _ {2} \\cdot \\| \\Omega (x ^ {(j)}) \\| _ {2}}, \\quad j = 1, \\dots , | \\mathcal {D} _ {L} |,", + "image_path": "d8dc14bd1fbaf982afab32618b1656e0b7eb8f5907e0d44777fed18bcf6ebae8.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2113" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 529, + 331 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 529, + 331 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 529, + 331 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 529, + 331 + ], + "type": "image", + "image_path": "f6055d52a8973357c3adf755aef92bc10c105af9ccd0b6cc0e429a6d48343a92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 339, + 403, + 350 + ], + "lines": [ + { + "bbox": [ + 203, + 339, + 403, + 350 + ], + "spans": [ + { + "bbox": [ + 203, + 339, + 403, + 350 + ], + "type": "text", + "content": "Figure 2. Overall framework of our proposed SSCBM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": " is a visual encoder. Then, we normalize the reciprocal of the cosine distance between the nearest " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": " data points and " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": " as weights. We use these weights to compute a weighted average of the concept labels of these " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": " data points, obtaining the pseudo concept label for " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": ". In this way, we can obtain pseudo concept labels " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}_U" + }, + { + "bbox": [ + 55, + 372, + 295, + 455 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "text", + "content": "In our experiments, we find that directly feeding pseudo-concept labels generated by KNN to CEM has satisfactory performance. However, this simple labeling method can lead to alignment issues in the concept embeddings learned by the CEM's concept encoder. Specifically, the predicted concepts " + }, + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "text", + "content": " could not have a relation to the corresponding features " + }, + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 459, + 295, + 590 + ], + "type": "text", + "content": " in the image, which hinders the effectiveness of CEM as a reliable interpretability tool. Moreover, due to misalignment, this will also degrade the prediction performance of the concept encoder. In the following, we aim to address such a misalignment issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Generating Concept Heatmaps. Our above pseudo concept labels via KNN have already contained useful information for prediction. Thus, our goal is to provide these labels with more information about their relation to the corresponding features. To achieve this, we first get another pseudo-concept label via these relations, which are calculated by the similarity between the concept embedding and the input image, namely concept heatmaps. Specifically, given an image " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", we first have its feature map " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{H \\times W \\times m}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " for each concept extracted by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "V = \\Omega(x)" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 372, + 553, + 384 + ], + "type": "text", + "content": " are the height and width of the feature map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "content": "-th concept embedding " + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\pmb{c}_i" + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "content": ", we can obtain a heatmap " + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 313, + 384, + 553, + 433 + ], + "type": "text", + "content": ", i.e., a similarity matrix that measures the similarity between the concept and the image can be obtained by computing their cosine distance:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 439, + 556, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 556, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 556, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {p, q, i} = \\frac {\\hat {\\boldsymbol {c}} _ {i} ^ {m} \\cdot V _ {p , q}}{\\| \\hat {\\boldsymbol {c}} _ {i} ^ {m} \\| \\cdot \\| V _ {p , q} \\|}, \\quad p = 1, \\dots , H, \\quad q = 1, \\dots , W", + "image_path": "4719330d606d71bbaf5d7e6c749cf89487bfffa4d436eb768c9be3300a732d9a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "p, q" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": "-th positions in the heatmaps, and " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{p,q,i}" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": " represents a local similarity score between " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^m" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_i" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": " represents the relation of each part of the image with the " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": "-the concept. Then, we derive heatmaps for all concepts, denoted as " + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{H}_1, \\mathcal{H}_2, \\dots, \\mathcal{H}_k\\}" + }, + { + "bbox": [ + 313, + 475, + 554, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Calculating Concept Scores and Concept Labels. As average pooling performs better in downstream classification tasks [45], we apply average pooling to the heatmaps to deduce the connection between the image and concepts: " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s_i = \\frac{1}{P \\cdot Q} \\sum_{p=1}^{P} \\sum_{q=1}^{Q} \\mathcal{H}_{p,q,i}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". Intuitively, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is the refined similarity score between the image and concept " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{c}_i^m" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". Thus, a concept vector " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " can be obtained, representing the similarity between an image input " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " and the set of concepts: " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s = (s_1, \\ldots, s_k)^\\top" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " can be considered as a soft concept label which is got by similarity. Next, we have to transform it into a hard concept label " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{c}_{\\text{align}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". We determine the presence of a concept attribute in an image based on a threshold value derived from an experiment. If the value " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " exceeds this threshold, we consider the image to possess that specific concept attribute and set the concept label to be True. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2114" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 288, + 83 + ], + "type": "text", + "content": "can obtain predicted concept labels for all unlabeled data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "text", + "content": "Alignment of Image. Based on our earlier discussions, on the one hand, the concept encoder should learn information from " + }, + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "text", + "content": ". On the other hand, it should also get concept embeddings which can get good similarity-based concept labels " + }, + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\hat{c}_{\\text {align }}" + }, + { + "bbox": [ + 55, + 84, + 296, + 168 + ], + "type": "text", + "content": " for alignment with the input image. Thus, we need a further alignment loss to achieve these two goals. Specifically, we compute the alignment loss as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 178, + 295, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 178, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 110, + 178, + 295, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = B C E \\left(\\hat {c} _ {\\text {i m g}}, \\hat {c} _ {\\text {a l i g n}}\\right). \\tag {3}", + "image_path": "40ca7f8cfd0692bce5742bf9f1c522c74d453c7124ac5a762f4b94eafb43f790.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": "In fact, during this alignment process, the CBM backbone is taught a reasonable latent space structure by the visual encoder, and thus we can regard it as an implicit Student-Teacher setup framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 150, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 150, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 150, + 270 + ], + "type": "text", + "content": "4.3. Final Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 275, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 296, + 431 + ], + "type": "text", + "content": "In this section, we will discuss how we derive the process of network optimization. First, we have the concept loss " + }, + { + "bbox": [ + 54, + 275, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 54, + 275, + 296, + 431 + ], + "type": "text", + "content": " in (1) to improve the interpretability of the concept embeddings. Also, since the concept embeddings are entered into the label predictor to output the final prediction, we also have a task loss between the predictions given by the concept bottleneck and ground truth, which is shown in (2). In the context of binary classification tasks, we employ binary cross-entropy (BCE) as our loss function. For multi-class classification tasks, we use cross-entropy as the measure. Finally, to align the images with the concept labels, we computed the loss of alignment in (3). Formally, the overall loss function of our approach can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 441, + 295, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 441, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 102, + 441, + 295, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\text {a l i g n}}, \\tag {4}", + "image_path": "57eca3c3a3ae031200db7bf5a2ae412e989f98f17ac537b9494843f5305de696.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 464, + 295, + 489 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 55, + 464, + 295, + 489 + ], + "type": "text", + "content": " are hyperparameters for the trade-off between interpretability and accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 499, + 137, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 137, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 137, + 513 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 519, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 628 + ], + "type": "text", + "content": "In this section, we will conduct experimental studies on the performance of our framework. Specifically, we evaluate the utility of our method according to the concept and prediction of the class label. We also design interpretability evaluation and test-time intervention to verify the interpretability and alignment performance of our method. Finally, we give the ablation study to verify why we need the alignment loss. Details and additional results are in Appendix A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 182, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 182, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 182, + 649 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": "Datasets. We evaluate our methods on four real-world image tasks: CUB[14], AwA2[33], WBCatt[42] and 7-point[22]. See Appendix A.1 for a detailed introduction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Baselines. We compare our SSCBM with Concept Bottle-neck Model (CBM) and Concept Embedding Model (CEM)" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 316, + 101, + 553, + 158 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Concept and task accuracy results of SSCBM at a labeled ratio of 0.1, compared to baselines in the fully supervised setting." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 101, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 316, + 101, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 101, + 553, + 158 + ], + "type": "table", + "html": "
MethodCUBAwA2WBCatt7-point
ConceptTaskConceptTaskConceptTaskConceptTask
CBM93.99%67.33%96.48%88.71%94.18%99.71%74.34%75.44%
CEM96.39%79.82%95.91%87.00%95.33%99.71%77.15%75.85%
SSCBM90.88%67.67%96.48%89.77%93.98%99.68%73.67%70.09%
", + "image_path": "18326956d5db39d543f396bdfde12169d7d27915ec56353f965ed9394e907760.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 178, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 297 + ], + "type": "text", + "content": "mentioned in Section 3. Since those baselines do not inherently include settings for semi-supervised learning, to ensure fairness in the evaluation, we use the same pseudolabeling approach as SSCBM. We utilize the KNN algorithm to annotate unlabeled data based on labeled data, which is mentioned in Section 4.2. For CBM [25], we adopt the same setting and architecture as in the original CBM. And for CEM, we follow the same setting as in [9]. We also compare label-free CBM [34], which is an unsupervised CBM method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the utility, we consider the accuracy for the prediction of class and concept labels. Specifically, concept accuracy measures the model's prediction accuracy for concepts: " + }, + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{acc} = \\frac{1}{N}\\cdot \\frac{1}{k}\\sum_{i = 1}^{N}\\sum_{j = 1}^{k}\\mathbb{I}(\\hat{c}_j^{(i)} = c_j^{(i)})" + }, + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "type": "text", + "content": ". Task accuracy measures the model's performance in predicting downstream task classes: " + }, + { + "bbox": [ + 313, + 299, + 554, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{acc} = \\frac{1}{N}\\sum_{i = 1}^{N}\\mathbb{I}(\\hat{\\pmb{y}}^{(i)} = \\pmb {y}^{(i)})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 433 + ], + "type": "text", + "content": "To evaluate the interpretability, besides the concept accuracy, similar to previous work [9, 25], we show visualization results. Moreover, we evaluate the performance of the test-time intervention." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 471 + ], + "type": "text", + "content": "Implementation Details. All experiments are conducted on a Tesla V100S PCIe 32 GB GPU and an Intel Xeon Processor CPU. See Appendix A.2 for more details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 479, + 471, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 471, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 471, + 493 + ], + "type": "text", + "content": "5.2. Evaluation Results on Utility" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "content": "To comprehensively evaluate the model's performance in the semi-supervised learning setting, we simulate real-world application scenarios. We consider labeling K samples for each class in the dataset, with an extreme case where " + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 1" + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "content": ". Clearly, " + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{K} =" + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "content": " total data size " + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "content": " labeled ratio " + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\div" + }, + { + "bbox": [ + 313, + 498, + 555, + 640 + ], + "type": "text", + "content": " number of classes. We also conduct experiments with labeled ratios of 0.05, 0.1, 0.15, and 0.2. We also compare the performance of SSCBM with the baseline under the fully supervised setting. The experimental results are shown in Table 1 and 2. Additionally, we also explored the impact of different backbones on model performance, which can be found in the Appendix A.3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Table 1 shows the performance comparison of SSCBM at a labeled ratio of 0.1 with CEM and CBM under the fully supervised setting. It can be observed that, although CEM performs the best on CUB, WBCatt and 7-point, SSCBM is only slightly inferior to it. On the AwA2 dataset, its performance even surpasses that of CEM under the fully super" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2115" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 112, + 504, + 319 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Results of concept and task accuracy for different datasets with different portions of labeled data. Since the label-free CBM [34] and LaBo [47] do not use the concept set included in the dataset but instead rely on GPT-3 to generate concepts based on class names, we do not report the concept accuracy here." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 112, + 504, + 319 + ], + "lines": [ + { + "bbox": [ + 107, + 112, + 504, + 319 + ], + "spans": [ + { + "bbox": [ + 107, + 112, + 504, + 319 + ], + "type": "table", + "html": "
DatasetLabeled RatioCBM+SSLCEM+SSLLabel-free CBMLaBoSSCBM
ConceptTaskConceptTaskConceptTaskConceptTaskConceptTask
CUBK=183.11%5.51%82.36%59.35%-74.31%-66.38%88.99%66.72%
0.05 (K=2)84.51%8.35%83.72%62.20%-74.31%-66.38%90.04%67.43%
0.1 (K=3)84.96%9.84%84.03%63.12%-74.31%-66.38%90.88%67.67%
0.15 (K=4)85.47%9.96%84.30%64.14%-74.31%-66.38%91.47%68.36%
0.2 (K=5)86.67%16.43%86.83%67.64%-74.31%-66.38%92.09%70.07%
AWA2K=171.32%90.87%65.55%84.44%-73.67%-93.73%85.05%89.02%
0.05 (K=30)77.23%91.11%68.31%85.01%-73.67%-93.73%95.42%89.37%
0.1 (K=60)80.45%91.03%69.96%85.48%-73.67%-93.73%96.48%89.77%
0.15 (K=90)83.46%91.00%72.14%86.61%-73.67%-93.73%96.81%90.54%
0.2 (K=120)85.76%91.12%72.14%86.61%-73.67%-93.73%96.81%90.44%
WBCattK=179.06%99.39%70.27%98.64%-41.92%-92.26%91.48%99.13%
0.05 (K=62)81.08%99.48%73.82%99.52%-41.92%-92.26%93.53%99.61%
0.1 (K=124)85.48%99.32%72.25%99.29%-41.92%-92.26%93.98%99.68%
0.15 (K=186)85.39%99.68%72.68%99.58%-41.92%-92.26%94.42%99.71%
0.2 (K=247)87.07%99.74%74.14%99.52%-41.92%-92.26%94.42%99.71%
7-pointK=159.91%55.95%62.78%66.09%-55.44%-64.97%66.58%66.84%
0.05 (K=5)65.36%57.47%67.85%67.09%-55.44%-64.97%70.98%68.77%
0.1 (K=9)68.82%55.70%72.23%66.33%-55.44%-64.97%73.67%70.09%
0.15 (K=13)66.14%59.75%66.54%67.09%-55.44%-64.97%73.94%72.56%
0.2 (K=17)70.29%60.44%73.04%66.84%-55.44%-64.97%76.52%74.56%
", + "image_path": "5293f30d894b5200af0e407aa5b25326e75fae900e0d4f154395a5a1d6fda5f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 363 + ], + "type": "text", + "content": "vised setting, which sufficiently demonstrates the superiority of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": "In Table 2, it can be observed that as the proportion of labeled data gradually increases, the concept and class prediction accuracy of all models improve. SSCBM outperforms all baselines. On the CUB dataset, when only one labeled sample per class is available, SSCBM achieves " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "88.99\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": " concept accuracy and " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "66.72\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": " class accuracy, which are " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "6.63\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "7.37\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": " higher, respectively, than the performance achieved by CEM in a semi-supervised learning setting. This demonstrates the effectiveness of the Image-Textual Semantics Alignment module. The 7-point dataset is the most challenging, and on this dataset, SSCBM still achieves the best performance across all settings. As the amount of labeled data increases, the concept prediction accuracy improves from " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "66.58\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "inline_equation", + "content": "76.52\\%" + }, + { + "bbox": [ + 56, + 365, + 295, + 616 + ], + "type": "text", + "content": ". We also note that on WBCatt and 7-point, the performance of Label-free CBM is poor. This is mainly because both datasets contain only five classes, which prevents Label-free CBM from fully leveraging the world knowledge of LLMs to generate an effective concept set, resulting in poor performance. This further demonstrates that semi-supervised learning scenarios are more realistic and practical for real-world datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 205, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 205, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 205, + 637 + ], + "type": "text", + "content": "5.3. Interpretability Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Besides the concept accuracy, note that in SSCBM, we also have a pseudo label revived by the alignment between the concept embedding and the input saliency map. We use the alignment loss to inherit this alignment. Thus, we will evaluate the alignment performance here to show the faithfulness of the interpretability given by SSCBM. We measure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "type": "text", + "content": "our alignment performance by comparing the correctness of the concept saliency map with the concept semantics in Figure 3. See Appendix A.2 for detailed experimental procedures for generating saliency maps. Results show that our concept saliency map matches the concept semantics, indicating the effectiveness of our alignment loss. In Appendix C, we also provide our additional interpretability evaluation in Figure 6-18." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 441, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 441, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 441, + 456 + ], + "type": "text", + "content": "5.4. Test-time Intervention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "text", + "content": "Test-time intervention enables human users to interact with the model at inference time. We conduct this experiment by correcting the ratio " + }, + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 462, + 555, + 545 + ], + "type": "text", + "content": " of the concept labels in the concept predictor. We expect that the model performance will steadily increase along with the ratio of concept intervention, indicating that the model learned such correct label information and automatically corrected other labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "The results in Figure 4 demonstrate the robustness of our model and an increasing trend to learn the information about the concept, indicating our interpretability and model prediction performance. During the test phase, the SSCBM model generates probability scores for each concept (e.g., has_wing_color :: brown). Prior to intervention, we employ the COOP algorithm (following CEM) to identify the most critical concepts exhibiting both high predictive uncertainty and task importance. For these selected concepts, we replace the model's predicted values with their corresponding ground truth annotations. This rectified concept vector is subsequently fed into the classification head to produce the final task prediction. Finally, we successfully cause the model to predict the Great Crested Flycatcher instead of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2116" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 70, + 199, + 167 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 199, + 167 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 199, + 167 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 199, + 167 + ], + "type": "image", + "image_path": "7da7de8127c7047a23aa8346f9d1cd821ee187972096c1a707cf2055e561d899.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 169, + 176, + 177 + ], + "lines": [ + { + "bbox": [ + 124, + 169, + 176, + 177 + ], + "spans": [ + { + "bbox": [ + 124, + 169, + 176, + 177 + ], + "type": "text", + "content": "(a) Original Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 304, + 167 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 304, + 167 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 304, + 167 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 304, + 167 + ], + "type": "image", + "image_path": "517c7b3f9b502cc51979ed2a51369b5e65804b7e6cd043eca99295e0cb1c70c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 169, + 297, + 177 + ], + "lines": [ + { + "bbox": [ + 213, + 169, + 297, + 177 + ], + "spans": [ + { + "bbox": [ + 213, + 169, + 297, + 177 + ], + "type": "text", + "content": "(b) Tail Pattern: Multi-colored" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 188, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 555, + 222 + ], + "type": "text", + "content": "Figure 3. The concept saliency map for the CUB dataset (savannah sparrow) demonstrates that our proposed SSCBM achieves meaningful alignment between the ground truth concepts and the input image features. The first image on the left is the original input image. The three images on the right show the aligned regions for different concepts using SSCBM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 311, + 70, + 409, + 167 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 409, + 167 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 409, + 167 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 409, + 167 + ], + "type": "image", + "image_path": "a2040c5184d8c90b24989582fa87a346ddb8e7a58011c416fdc7e64d5ca0dddc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 169, + 399, + 177 + ], + "lines": [ + { + "bbox": [ + 320, + 169, + 399, + 177 + ], + "spans": [ + { + "bbox": [ + 320, + 169, + 399, + 177 + ], + "type": "text", + "content": "(c) Underparts Color: Brown" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 414, + 70, + 512, + 167 + ], + "blocks": [ + { + "bbox": [ + 414, + 70, + 512, + 167 + ], + "lines": [ + { + "bbox": [ + 414, + 70, + 512, + 167 + ], + "spans": [ + { + "bbox": [ + 414, + 70, + 512, + 167 + ], + "type": "image", + "image_path": "4af6004c8beb3e17a59de06ead3d8a56a8919aadb615af41df688166d676b69c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 169, + 496, + 177 + ], + "lines": [ + { + "bbox": [ + 430, + 169, + 496, + 177 + ], + "spans": [ + { + "bbox": [ + 430, + 169, + 496, + 177 + ], + "type": "text", + "content": "(d) Wing Color: Brown" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 144, + 243, + 233, + 331 + ], + "blocks": [ + { + "bbox": [ + 144, + 243, + 233, + 331 + ], + "lines": [ + { + "bbox": [ + 144, + 243, + 233, + 331 + ], + "spans": [ + { + "bbox": [ + 144, + 243, + 233, + 331 + ], + "type": "image", + "image_path": "2aa34b9d7a97a501ccc0a0a8a04eb3af6137337b6aabf3b19604f4fa466d22b3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 244, + 243, + 329, + 333 + ], + "blocks": [ + { + "bbox": [ + 244, + 243, + 329, + 333 + ], + "lines": [ + { + "bbox": [ + 244, + 243, + 329, + 333 + ], + "spans": [ + { + "bbox": [ + 244, + 243, + 329, + 333 + ], + "type": "image", + "image_path": "44fcb677be76cff0bb1c49531ccfc493519c5b768fa862338613e9ba8d7d30ab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 550, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 550, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 550, + 357 + ], + "type": "text", + "content": "Figure 4. Left: Performance with different ratios of intervened concepts on CUB dataset. Right: An example of successful intervention." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 330, + 252, + 465, + 325 + ], + "blocks": [ + { + "bbox": [ + 330, + 252, + 465, + 325 + ], + "lines": [ + { + "bbox": [ + 330, + 252, + 465, + 325 + ], + "spans": [ + { + "bbox": [ + 330, + 252, + 465, + 325 + ], + "type": "image", + "image_path": "4eec13b2b797ac0efabcf68556e47c5d7e5bd320be4078f670030c3445ee637d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 378, + 264, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 264, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 264, + 390 + ], + "type": "text", + "content": "Swainson Warbler. More results are in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 58, + 408, + 296, + 581 + ], + "blocks": [ + { + "bbox": [ + 111, + 395, + 236, + 405 + ], + "lines": [ + { + "bbox": [ + 111, + 395, + 236, + 405 + ], + "spans": [ + { + "bbox": [ + 111, + 395, + 236, + 405 + ], + "type": "text", + "content": "Table 3. Results of ablation study." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 408, + 296, + 581 + ], + "lines": [ + { + "bbox": [ + 58, + 408, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 58, + 408, + 296, + 581 + ], + "type": "table", + "html": "
Ratiow/o imgw/o alignfull model
ConceptTaskConceptTaskConceptTask
CUB0.0001 (K=1)87.77%59.89%80.87%65.15%88.99%66.72%
0.05 (K=2)90.01%65.98%80.60%65.67%90.04%67.43%
0.1 (K=3)91.02%67.28%80.60%64.69%90.88%67.67%
0.15 (K=4)90.60%67.97%80.65%65.43%91.47%68.36%
0.2 (K=5)91.94%68.99%80.73%64.45%92.09%70.07%
WBCatt0.0001 (K=1)88.44%99.68%57.80%99.51%91.48%99.13%
0.05 (K=62)92.17%99.29%57.93%99.51%93.53%99.61%
0.1 (K=124)92.91%99.52%60.49%99.64%93.98%99.68%
0.15 (K=186)93.55%99.58%59.13%99.65%94.42%99.71%
0.2 (K=247)94.19%99.61%60.65%99.70%94.42%99.71%
7-point0.0001 (K=1)62.31%64.30%55.07%66.33%66.58%66.84%
0.05 (K=5)69.42%65.06%56.75%68.10%70.98%68.77%
0.1 (K=9)70.06%67.34%56.07%65.82%73.67%70.09%
0.15 (K=13)73.39%68.10%56.23%65.82%73.94%72.56%
0.2 (K=17)74.44%68.61%56.11%65.57%76.52%74.56%
", + "image_path": "b22c82ffd1aff0be97dc782e68df4a4f75dc8eacf69b1d125410d35402f7f680.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 149, + 601 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "We then give a finer study on the two kinds of pseudo-labels to demonstrate that each one is indispensable in bolstering the efficacy of SSCBM. Specifically, based on our method in Section 4.2, we have three types of pseudo labels available: pseudo concept label " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", and label " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " predicted by the concept embedding. We first remove " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and calculate our alignment loss using " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "; conversely, we remove " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and calculate the alignment loss using " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "text", + "content": "From the results presented in Table 3, it is evident that removing the " + }, + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\hat{c}_{img}" + }, + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "text", + "content": " component significantly degrades the performance of SSCBM at both the concept level and class level. This indicates that the pseudo-concept labels via KNN contain necessary information about the ground truth. The concept encoder needs to extract information from such labels to get better performance. This situation is similar when removing the " + }, + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "inline_equation", + "content": "\\hat{c}_{align}" + }, + { + "bbox": [ + 313, + 378, + 555, + 558 + ], + "type": "text", + "content": " component, indicating that aligning the concept embedding and the input saliency map can further extract useful information from the input image and, thus, is beneficial to improve the performance. Our observations underscore the high degree of joint effectiveness of two kinds of pseudo-concept labels within our objective function, collectively contributing to the enhancement of model prediction and concept label prediction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "The training of current CBMs heavily relies on the accuracy and richness of annotated concepts in the dataset. These concept labels are typically provided by experts, which can be costly and require significant resources and effort. Additionally, concept saliency maps frequently misalign with input saliency maps, causing concept predictions to correspond to irrelevant input features - an issue related to annotation alignment. In this problem, we propose SSCBM, a strategy to generate pseudo labels and an alignment loss to solve these two problems. Results show our effectiveness." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2117" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work is supported in part by the funding BAS/1/1689-01-01, URF/1/4663-01-01, REI/1/5232-01-01, REI/1/5332-01-01, and URF/1/5508-01-01 from KAUST, and funding from KAUST - Center of Excellence for Generative AI, under award number 5940." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "text", + "content": "[1] Andrea Acevedo, Santiago Alférez, Anna Merino, Laura Puigvi, and José Rodellar. Recognition of peripheral blood cell images using convolutional neural networks. Computer Methods and Programs in Biomedicine, 180:105020, 2019. 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In 2020 International joint conference on neural networks (IJCNN), pages 1-8. IEEE, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 294, + 372 + ], + "type": "text", + "content": "[3] Pietro Barbiero, Gabriele Ciravegna, Francesco Giannini, Mateo Espinosa Zarlenga, Lucie Charlotte Magister, Alberto Tonda, Pietro Lio, Frédéric Precioso, Mateja Jamnik, and Giuseppe Marra. Interpretable neural-symbolic concept reasoning. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1801-1825. PMLR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 375, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 294, + 430 + ], + "type": "text", + "content": "[4] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 475 + ], + "type": "text", + "content": "[5] Kushal Chauhan, Rishabh Tiwari, Jan Freyberg, Pradeep Shenoy, and Krishnamurthy Dvijotham. Interactive concept bottleneck models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5948-5955, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 521 + ], + "type": "text", + "content": "[6] Asma Chebli, Akila Djebar, and Hayet Farida Marouani. Semi-supervised learning for medical application: A survey. In 2018 International Conference on Applied Smart Systems (ICASS), pages 1-9, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 566 + ], + "type": "text", + "content": "[7] Xing Chen, Biao Ren, Ming Chen, Quanxin Wang, Lixin Zhang, and Guiying Yan. Nllss: predicting synergistic drug combinations based on semi-supervised learning. PLoS computational biology, 12(7):e1004975, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 612 + ], + "type": "text", + "content": "[8] Veronika Cheplygina, Marleen de Bruijnne, and Josien P.W. Pluim. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis. Medical Image Analysis, 54:280-296, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 689 + ], + "type": "text", + "content": "[9] Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci, Adrian Weller, et al. Concept embedding models: Beyond the accuracy-explainability trade-off. Advances in Neural Information Processing Systems, 35:21400-21413, 2022. 3, 4, 6, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[10] Roman Filipovych and Christos Davatzikos. Semi-supervised pattern classification of medical images: Appli" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "cation to mild cognitive impairment (mci). NeuroImage, 55 (3):1109-1119, 2011. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 140 + ], + "type": "text", + "content": "[11] Jack Furby, Daniel Cunningham, Dave Braines, and Alun Preece. Can we constrain concept bottleneck models to learn semantically meaningful input features? arXiv preprint arXiv:2402.00912, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 143, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 185 + ], + "type": "text", + "content": "[12] Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. Advances in Neural Information Processing Systems, 35:23386-23397, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "type": "text", + "content": "[14] Xiangteng He and Yuxin Peng. Fine-grained visual-textual representation learning. IEEE Transactions on Circuits and Systems for Video Technology, 30(2):520-531, 2019. 6, 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "type": "text", + "content": "[15] Junlin Hou, Jilan Xu, and Hao Chen. Concept-attention whitening for interpretable skin lesion diagnosis. arXiv preprint arXiv:2404.05997, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 336 + ], + "type": "text", + "content": "[16] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Improving faithfulness for vision transformers. arXiv preprint arXiv:2311.17983, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 382 + ], + "type": "text", + "content": "[17] Lijie Hu, Yixin Liu, Ninghao Liu, Mengdi Huai, Lichao Sun, and Di Wang. Seat: stable and explainable attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12907-12915, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 384, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 428 + ], + "type": "text", + "content": "[18] Lijie Hu, Liang Liu, Shu Yang, Xin Chen, Hongru Xiao, Mengdi Li, Pan Zhou, Muhammad Asif Ali, and Di Wang. A hopfieldian view-based interpretation for chain-of-thought reasoning. arXiv preprint arXiv:2406.12255, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 463 + ], + "type": "text", + "content": "[19] Lijie Hu, Chenyang Ren, Zhengyu Hu, Cheng-Long Wang, and Di Wang. Editable concept bottleneck models. arXiv preprint arXiv:2405.15476, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 464, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 464, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 317, + 464, + 553, + 508 + ], + "type": "text", + "content": "[20] Tri Huynh, Aiden Nibali, and Zhen He. Semi-supervised learning for medical image classification using imbalanced training data. Computer Methods and Programs in Biomedicine, 216:106628, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 554 + ], + "type": "text", + "content": "[21] Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, and Kyunghyun Cho. Concept bottleneck generative models. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 556, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 609 + ], + "type": "text", + "content": "[22] Jeremy Kawahara, Sara Daneshvar, Giuseppe Argenziano, and Ghassan Hamarneh. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE Journal of Biomedical and Health Informatics, 23(2): 538-546, 2019. 6, 11" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "text", + "content": "[23] Mert Keser, Gesina Schwalbe, Azarm Nowzad, and Alois Knoll. Interpretable model-agnostic plausibility verification for 2d object detectors using domain-invariant concept bottleneck models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3890–3899, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[24] Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, and Sunghroh Yoon. Probabilistic concept bottleneck models. arXiv preprint arXiv:2306.01574, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2118" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338-5348. PMLR, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[26] Songning Lai, Lijie Hu, Junxiao Wang, Laure Berti-Equille, and Di Wang. Faithful vision-language interpretation via concept bottleneck models. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "type": "text", + "content": "[27] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[28] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, page 896. Atlanta, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Jia Li, Lijie Hu, Zhixian He, Jingfeng Zhang, Tianhang Zheng, and Di Wang. Text guided image editing with automatic concept locating and forgetting. arXiv preprint arXiv:2405.19708, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[30] Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. Large language models in finance: A survey. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 374-382, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[31] Lucie Charlotte Magister, Dmitry Kazhdan, Vikash Singh, and Pietro Lio. Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks. arXiv preprint arXiv:2107.11889, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 410 + ], + "type": "text", + "content": "[32] Todd K Moon. The expectation-maximization algorithm. IEEE Signal processing magazine, 13(6):47-60, 1996. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[33] Li Niu, Qingtao Tang, Ashok Veeraraghavan, and Ashutosh Sabharwal. Learning from noisy web data with category-level supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7689-7698, 2018. 6, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[34] Tuomas Oikarinen, Subhro Das, Lam Nguyen, and Lily Weng. Label-free concept bottleneck models. In International Conference on Learning Representations, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[35] Yassine Ouali, Céline Hudelot, and Myriam Tami. An overview of deep semi-supervised learning. arXiv preprint arXiv:2006.05278, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[36] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V Le. Meta pseudo labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11557-11568, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[37] Yoshihide Sawada and Keigo Nakamura. Concept bottleneck model with additional unsupervised concepts. IEEE Access, 10:41758-41765, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[38] Ivaxi Sheth and Samira Ebrahimi Kahou. Auxiliary losses for learning generalizable concept-based models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[39] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 543 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Advances in neural information processing systems, 33:596-608, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[40] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[41] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "text", + "content": "[42] Satoshi Tsutsui, Winnie Pang, and Bihan Wen. Wbcatt: A white blood cell dataset annotated with detailed morphological attributes. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6, 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[43] Jesper E Van Engelen and Holger H Hoos. A survey on semi-supervised learning. Machine learning, 109(2):373-440, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[44] Han Xuanyuan, Pietro Barbiero, Dobrik Georgiev, Lucie Charlotte Magister, and Pietro Lio. Global concept-based interpretability for graph neural networks via neuron analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10675-10683, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[45] An Yan, Yu Wang, Yiwu Zhong, Zexue He, Petros Karypis, Zihan Wang, Chengyu Dong, Amilcare Gentili, Chun-Nan Hsu, Jingbo Shang, et al. Robust and interpretable medical image classifiers via concept bottleneck models. arXiv preprint arXiv:2310.03182, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[46] Shu Yang, Lijie Hu, Lu Yu, Muhammad Asif Ali, and Di Wang. Human-ai interactions in the communication era: Autophagy makes large models achieving local optima. arXiv preprint arXiv:2402.11271, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[47] Yue Yang, Artemis Panagopoulou, Shenghao Zhou, Daniel Jin, Chris Callison-Burch, and Mark Yatskar. Language in a bottle: Language model guided concept bottlenecks for interpretable image classification, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "text", + "content": "[48] Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 543 + ], + "type": "text", + "content": "[49] Gang Zhang, Shan-Xing Ou, Yong-Hui Huang, and Chun-Ru Wang. Semi-supervised learning methods for large scale healthcare data analysis. International Journal of Computers in Healthcare, 2(2):98-110, 2015. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2119" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_content_list.json b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8c8e819b68da5315bfe2635aa510b36a049c576e --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "Semi-supervised Deep Transfer for Regression without Domain Alignment", + "text_level": 1, + "bbox": [ + 124, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mainak Biswas1,2, Ambedkar Dukkipati3, Devarajan Sridharan1,2,3", + "bbox": [ + 243, + 179, + 772, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Brain, Computation and Data Sciences, $^{2}$ Centre for Neuroscience, $^{3}$ Computer Science and Automation Indian Institute of Science, Bangalore - 560012", + "bbox": [ + 101, + 198, + 916, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mainakbiswas@iisc.ac.in, ambedkar@iisc.ac.in, sridhar@iisc.ac.in", + "bbox": [ + 222, + 237, + 790, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data or reliably labeled target data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT - a Contradistinguisher-based Regularization Approach for Flexible Training - for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and \"brain age\" prediction with structural MRI data. For both datasets, CRAFT yielded up to $9\\%$ improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than $3\\%$ . Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine.", + "bbox": [ + 88, + 321, + 485, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: source-free DA, continuous label prediction, gaze prediction, brain age prediction, healthcare", + "bbox": [ + 89, + 869, + 480, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 286, + 643, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For the successful application of deep learning models in the real world, they must be robust to \"domain shift\" [50]. For example, in biology and medicine, supervised deep learning models are often trained on large, high-quality source datasets collected in resource-rich settings [11, 39]. However, naive transfer or even simple finetuning fails to produce high accuracies when tested with smaller target datasets of mixed quality collected [2, 45]. As a result, state-of-the-art domain adaptation (DA) approaches, which enable the source models to generalize successfully to target datasets, must be employed. Many of these approaches involve explicit alignment of source and target, while learning a shared predictive model for both domains [17, 19, 40].", + "bbox": [ + 511, + 311, + 906, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A particular challenge occurs with DA when the source dataset is unavailable. This can happen either because of privacy or proprietary concerns, such as with medical images or patents [43]. Alternatively, source datasets may be too large, incurring prohibitive storage and computation costs in low-resource settings [39]. Source-free (SF) domain adaptation methods that address these challenges are getting increasingly popular [18, 32]. Such methods generalize the source model to unlabeled or partially target data (unsupervised or semi-supervised DA) with one of two approaches: data-based or model-based. While data-based methods create surrogate source data, model-based methods adapt the source model to the target domain (see Section 2).", + "bbox": [ + 511, + 508, + 908, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address model-based SF-DA in a semi-supervised setting. Specifically, we focus on neuroscience applications where source data are inaccessible, and target labels are reliably available only for a few samples. Regression tasks – like predicting brain age (Fig. 1C) [39], saccades [29], neural activity [20], and stimulus orientation [8] – are ubiquitous in neuroscience; we tackle a subset of these. We build upon Contradistinguisher [3, 4], a recently-proposed tool for effective unsupervised DA (CUDA). Unlike popular representation alignment DA methods [19, 40], CUDA directly learns a common model for both domains. Yet, CUDA requires full access to source data and addresses classification tasks. Therefore, it is not directly applicable to our problem.", + "bbox": [ + 511, + 704, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "827", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb9853653c6b9022712a0579a3ff2529cc99fff34dab54d3756ff8484db4e5c0.jpg", + "image_caption": [ + "A", + "X" + ], + "image_footnote": [], + "bbox": [ + 133, + 98, + 227, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50abad7ebb0e934b69ecdb035b0c1da341db9610f3320d5e761bc874f72f3d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 101, + 359, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7ff5d8c03c3f4ad2d8c212ee190fd4cb9ac2db57efa1226e6b4c1f6914156dc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 97, + 478, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90f5303b5b158277667aad3571bc0e9dbed2b57cdc3db5b3e10ecf16eeeab682.jpg", + "image_caption": [ + "B", + "Figure 1. Schematic of Semisupervised Deep Transfer without Domain Alignment. (A) The approach seeks to transfer a pretrained model to a target domain with limited labeled data when the source dataset is unavailable. (B) Because the source dataset is unavailable, we tune the decision boundaries without explicitly aligning the representation of the domains, combining a supervised loss and the CUDA loss (regularizer). (C) The algorithm targets deep transfer in regression tasks, like brain age prediction." + ], + "image_footnote": [], + "bbox": [ + 106, + 213, + 269, + 291 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c6cd4215a790df2af29f683479d512f61696c30867f90231da3da7ce48aa7b8.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 295, + 195, + 475, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we leverage CUDA to develop a new algorithm - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT) which performs deep transfer, without domain alignment, even when the source dataset is unavailable (Fig. 1A-B). We also extend CUDA to a regression setting. Previous studies approach regression tasks naively by binning continuous labels into discrete bins [40], sometimes employing a rank-ordinal objective [7]. Regardless, the efficacy of such approaches may depend critically on bin sizes. Our CRAFT model uses a principled approach to extend CUDA to predictions involving continuous-valued outputs. We employ binning as a practical strategy to generate and optimize pseudo-labels, without constraining the model to produce discrete outputs. Our study makes the following main contributions.", + "bbox": [ + 89, + 458, + 483, + 684 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- First, we extend CUDA to source-free DA for regression tasks involving continuous-valued outputs.", + "- Second, we formulate a theoretically-motivated semi-supervised objective by combining supervised and unsupervised losses with a tunable hyperparameter $(\\alpha)$ (Section 3). We also derive the unsupervised loss as a maximum entropy prior to model parameters (Appendix A.2).", + "- Third, we show the effectiveness of CRAFT on two neuroscience challenges - gaze prediction from electroencephalographic (EEG) [29] data and brain age prediction from magnetic resonance imaging (MRI) data [45].", + "- Fourth, we also apply CRAFT to two real-world regression benchmarks involving people counting [44, 48] and tumor size prediction [6, 16]." + ], + "bbox": [ + 89, + 689, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fifth, we compare CRAFT against state-of-the-art SF-DA methods [15, 26, 35, 37] and show that the best improvements occur as unlabeled target samples increase, demonstrating its efficacy for semi-supervised SF-DA.", + "- Finally, we analyze the computational complexity and demonstrate numbers competitive with sota models.", + "- Our code is publicly available (see Appendix E.2)." + ], + "bbox": [ + 513, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 210, + 653, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain alignment methods. Popular Unsupervised Domain Adaptation (UDA) approaches rely on learning common intermediate representations across source and target domains. Domain Adversarial training of Neural Networks (DANN) [19] uses a discriminator to learn common representations across both domains adversarially. In Importance Weighting [40] (IW), higher importance is afforded to losses corresponding to source samples more likely to belong to the target domain. Central Moment Discrepancy (CMD) [17] minimizes the central moments of the latent representations between the domains, while others optimize their Maximum Mean Discrepancy (MMD) [22, 34]. Despite being effective, they cannot be used when the source is unavailable and will not be discussed further.", + "bbox": [ + 511, + 236, + 906, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Source-free UDA. To address DA scenarios where source data and its target labels are unavailable, several source-free unsupervised DA approaches have been recently proposed. Broadly, they are categorized as: Data-based and Model-based [32]. Data-based methods generate a source-like distribution [14, 30]. However, they often depend on the availability of generative models trained on the source, which could be unavailable. For example, SF-UDA through domain-specific perturbation (AUGFree) [51] assumes the existence of domain-invariant features for prediction. Since the source is unavailable, AUGFree perturbs the target data and then learns a transformation to the invariant features by adversarially aligning the target and the perturbed target.", + "bbox": [ + 511, + 446, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, model-based methods finetune the source model on the target using strategies like contrastive learning [27], pseudo-labeling [26], or entropy minimization [36]. They are often prone to overfitting to pseudolabels. For example, Class-Balanced Multicentric Dynamic Prototype Strategy for SF-UDA (BMD) [42] follows a pseudo-labeling strategy using class prototypes based on the source model predictions, and in turn uses these labels for training. Source Data-Absent UDA Through Hypothesis Transfer (SHOT++) [33] combines three objectives - a) maximizing mutual information between the predictions and inputs, b) self-supervised loss based on prototype-based clusters, and c) a Mixup-based (See SF-SSDA) loss [37]. Attracting and Dispersing (AaD) [52] defines a neighboring cluster (K nearest neighbors) and a randomly selected background cluster for each sample, and maximizes the log-likelihood of the ratio of their distributions.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "828", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Target-agnostic SF UDA for Regression Tasks (TAS-FAR) [26] works on the principle that prediction uncertainty is correlated with prediction error. It uses Monte-Carlo dropout to estimate label uncertainty and divides the target dataset into confident and uncertain samples. A discrete density map is estimated using the confident data, which is then used as weights to recalibrate the pseudo-labels of uncertain data. SF-UDA to measurement Shift via Bottom-Up Feature Restoration (BUFR/DataFree) [15] is designed on the principle of intermediate feature alignment of the source and the target distributions. During adaptation, the model minimizes the symmetric KL divergence between feature distributions of source (fixed) and target data.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we compare our model's performance against two of the above methods - TASFAR and DataFree. TASFAR was selected as it is a pseudo-labeling-based method suitable for regression. On the other hand, DataFree is a pseudo-label-free feature alignment method, not constrained to classification tasks. By adding a supervised loss (Section 4.2), we evaluate these as semi-supervised models.", + "bbox": [ + 89, + 290, + 482, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Source-free SSDA. Source-free semi-supervised domain adaptation (SF-SSDA) models best address the scenario of SF-DA when few target labels are available. For example, Generalized SF-SSDA [1] minimizes source loss alongside an unsupervised target loss which maximizes the cosine similarity between predictions of k-nearest neighbors. The model is then transferred by adversarial training based on source features generated using a Conditional VAE. Similarly, [23] proposes generating surrogate source data from target samples using gradient-based optimization. Learning from Different Samples for SF-SSDA [28] minimizes the contrastive loss on the predicted probabilities of the target domain. Furthermore, it minimizes supervised and contrastive loss on a set comprising labeled and unlabeled data with low uncertainty. Likewise, Mutual Enhancement training for SS Hypothesis transfer (MESH) [36] minimizes the supervised loss on the labeled samples and entropy on pseudo-labeled samples. Finally, the target manifold is smoothened by making predictions robust to perturbations.", + "bbox": [ + 89, + 398, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Progressive Mixup [37] is trained on augmentations by convex combinations of data on which the source model is confident and ones on which it is uncertain (hybrid mixup). Similarly, another set is obtained by convex combinations of samples in the (pseudo) labeled dataset (self-mixup). In contrast, Bilateral-Branch Consistency Network (BBCN) [35] computes prototypes as the weighted mean of feature representations of the target dataset by the source model. Pseudo-labels are selected based on the proximity of unlabeled data to these prototypes. A separate target branch is trained using the pseudo-labels, and a consistency loss is added between the representations of both branches. The need for class-level prototype generation makes it challenging to adapt this model to regression tasks.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare the performance of CRAFT with two SFSSDA methods - SF-SSDA via progressive Mixup [37], and BBCN [35], as they have little dependence on the source distribution and its derivatives. Moreover, these methods are based on complementary ideas, and the latter [35] has been applied specifically in medical imaging, for tuberculosis recognition.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The CRAFT Model", + "text_level": 1, + "bbox": [ + 511, + 213, + 702, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model formulation", + "text_level": 1, + "bbox": [ + 511, + 239, + 694, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background. Our objective is to develop an approach for source-free semi-supervised domain adaptation, for regression tasks. For this, we build upon the Contradistinguisher-based Unsupervised Domain Adaptation (CUDA) model [4]. CUDA obviates the need for learning a common intermediate representation across the source and the target datasets. It optimizes a joint distribution over the target features and category labels to adapt the source model. Because CUDA is intended for unsupervised domain adaptation (UDA), target labels are unavailable. Hence, the algorithm uses a two-step joint optimization over the target labels and model parameters. First, it generates pseudo-labels on the target dataset from the source model [21]. Then, it updates the model parameters by fixing the pseudo-labels.", + "bbox": [ + 511, + 263, + 903, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this study, we extend CUDA to a source-free semi-supervised setting (SF-SSDA). Briefly, to the supervised label loss, we add a 'regularized', unsupervised loss as a prior on the model parameters. As we will show in the next section, the model can be flexibly applied, either in SF-UDA case or in the SF-SSDA case, depending on the weightage given to the supervised loss. We, moreover, extend the model to continuous label prediction. We call our new formulation CRAFT - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT).", + "bbox": [ + 511, + 474, + 903, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective. We consider a task in which a source model $\\theta^s$ trained on a source dataset - involving continuous label prediction (regression) - is already available to us. The target dataset comprises i.i.d samples from an unknown distribution $p^t (\\mathbf{x},y)$ : $\\mathcal{D}^t = \\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}$ , $\\{\\mathbf{x}_i^t\\}_{i = 1}^{N_{ul}}$ where $\\mathbf{x}_i^t\\in \\mathbb{R}^d$ is a feature-vector, and $y_{i}^{t}\\in \\mathbb{R}$ is its corresponding label. Target training data are partially and sparingly labeled; we define the proportion of unlabeled samples $(N_{ul})$ to the total number of samples $(N)$ , as $N_{ul}:(N_l + N_{ul})$ .", + "bbox": [ + 511, + 627, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate, first, the supervised objective. For continuous label prediction problems, the conditional density of the label given the features is assumed to be normally distributed. The mean is modeled using a neural network $f(:,\\theta)$ , parameterized by $\\theta$ , as: $p(y^{t}|\\mathbf{x}^{t},\\theta) = \\mathcal{N}(y^{t};f(\\mathbf{x}^{t};\\theta),c)$ , and the variance, $c$ is assumed to be constant as it simplifies the optimization. In the final formulation, the variance occurs as the weight of the unsupervised loss, $\\alpha$ (equation 3); therefore, only one of the two", + "bbox": [ + 511, + 765, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "829", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hyperparameters ( $c$ or $\\alpha$ ) needs to be tuned. Hence, the log-likelihood can be written as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log p \\left(\\mathcal {D} ^ {t} | \\theta\\right) \\propto - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 132, + 482, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unsupervised objective is essentially identical to the CUDA objective and is defined as a joint log density over the labels and model parameters as follows:", + "bbox": [ + 89, + 178, + 483, + 224 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log q \\left(\\mathbf {x} ^ {t}, y ^ {t} \\mid \\theta\\right) = \\log \\left[ \\frac {p \\left(y ^ {t} \\mid \\mathbf {x} ^ {t} , \\theta\\right)}{\\sum_ {i = 1} ^ {N} p \\left(y ^ {t} \\mid \\mathbf {x} _ {i} ^ {t} , \\theta\\right)} p \\left(y ^ {t}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 232, + 482, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This joint distribution seeks to reduce the model's predictive bias when faced with a distribution shift in the target domain (see Fig. 1B). The denominator, which reflects the total probability of predicting a particular class, normalizes for biases originating from the source data. In other words, model predictions are normalized as $p(y^{t}|\\mathbf{x}^{t},\\theta) / \\sum_{i = 1}^{N}p(y^{t}|\\mathbf{x}_{i}^{t},\\theta)$ , such that no class has a biased probability of being predicted. Finally, the model predictions align with the target label distribution, using the supplied marginal $p(y^t)$ .", + "bbox": [ + 89, + 282, + 483, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note, that this objective does not depend on whether the samples are labeled or not. Optimizing the joint distribution: 1) normalizes inherent bias towards one label over the other by calibrating the prediction for a label with its total weight in the dataset, 2) and explicitly trains $q$ to match the marginals - i.e., $\\sum_{i=1}^{N} q(\\mathbf{x}_i^t, y^t|\\theta) = p(y^t)$ , thereby matching high-level statistics like per-class proportions in the training distribution. Importantly, with an informative prior $p(y^t)$ that specifies the putative distribution of the target labels, pseudo-label prediction can be biased toward labels that are more likely (see Section 3.2).", + "bbox": [ + 89, + 433, + 483, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining the supervised and the unsupervised objective, the semi-supervised CRAFT objective is defined as:", + "bbox": [ + 89, + 599, + 482, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\mathcal {D} ^ {t}, \\theta\\right) = \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, y _ {i} ^ {t} \\mid \\theta\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 638, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the above equation, $\\alpha$ is the weight associated with the unsupervised objective and is tuned as a hyperparameter on a held-out cross-validation set during training. For deep-transfer, model parameters are initialized to $\\theta^s$ derived from the source model, and then gradient descent is performed to obtain $\\theta^t$ .", + "bbox": [ + 89, + 689, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical motivation. Artificial Neural Networks (ANNs) with hypotheses space $\\Theta$ are trained to learn a parametric approximation $p(\\mathbf{x},y|\\theta)$ , $\\theta \\in \\Theta$ , such that the log-likelihood of the training data given the parameters is maximized. However, training ANNs often does not account for prior distributions over model parameters. In Appendix A.1, we demonstrate that the CRAFT semi-supervised objective can be theoretically derived as MAP", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimation of the parameters $(\\log p(\\theta|\\mathcal{D}))$ , which we write as a sum of the supervised objective, $\\log p(\\mathcal{D}|\\theta)$ , and a prior on the model parameters, $p(\\theta)$ . Specifically, we demonstrate, using the principle of entropy maximization [21], that the unsupervised objective is related to the prior on the model parameters.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Model optimization", + "text_level": 1, + "bbox": [ + 513, + 189, + 702, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We optimize the supervised loss of the CRAFT objective with conventional gradient descent, jointly optimizing both components in the loss (equation 3). The unsupervised component of the loss is dependent on the labels $y^{t}$ , which are unavailable for the vast majority of target data. To tackle this, we employ a 2-step optimization process.", + "bbox": [ + 511, + 210, + 905, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo-label selection. As the first step, we fix the model parameters $\\theta$ and select the label that maximizes the joint distribution for each datapoint:", + "bbox": [ + 511, + 301, + 905, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} q \\left(\\mathbf {x} _ {i} ^ {t}, y ^ {t} \\mid \\theta\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 353, + 906, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{y}_i$ the pseudo-label for feature vector $\\mathbf{x}_i$ . For classification tasks, the optimization in equation 4 can be achieved by selecting the best label from a finite set of class labels that maximizes $q$ [4]. But, for regression problems such as ours, class labels map to real numbers spanning the range of the model outputs $f(:,\\theta):\\mathbb{R}^d\\mapsto [a,b]$ . To address this, we rewrite the above equation under the assumption that the continuous-valued labels are distributed normally, with mean $f(\\mathbf{x}^t;\\theta)$ , and variance $c$ . Hence, the above equation can be written as:", + "bbox": [ + 511, + 387, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} \\frac {\\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {i} ^ {t} ; \\theta\\right) , c\\right) p \\left(y ^ {t}\\right)}{\\sum_ {l = 1} ^ {N} \\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right) , c\\right)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 541, + 906, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Equation 5 can be optimized with gradient ascent with respect to $y^{t}$ for the objective log $q(\\mathbf{x}_i^t, y^t|\\theta)$ . But, the pseudo-labels must be estimated before gradient descent can be performed on the model parameters (see next step, equation 6). This requires nested gradient descent, which increases computational complexity. Moreover, to employ an informative prior over target labels, we fit a class of mixture models (described in Appendix A.3) to estimate $p(y)$ from the data; this makes backpropagation more cumbersome. To overcome these challenges, we divide the entire range of labels into small, discrete bins whose midpoints are the candidate pseudo-labels. Hence, this step in the optimization can be achieved faster pseudo-label selection from a discrete set. Note that this approach does not transform the model outputs to be discrete; it merely represents a convenient optimization strategy to update the parameters. In the experimental section, we provide a recipe for choosing bin sizes and study the impact of varying bin sizes.", + "bbox": [ + 511, + 583, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Maximization. In the second step, we optimize the model parameters by using the labeled data $\\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}$ and fixing the pseudo-labels $\\{\\mathbf{x}_i^t,\\tilde{y}_i^t\\}_{i = 1}^N$ , as:", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "830", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/888e5b032c4d9b15c42bc7c8b1148856bc99f3a289a87f7305873188cd7787c9.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 142, + 104, + 483, + 233 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d0298fa1f99250ae7821fa1d04b391c8c05d5d0613e83037e998a8ea29c239b.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 519, + 119, + 704, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6ceebd152ad3338c057f9ad722eaaa208b690c7917774b10832407957cb3094.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 109, + 879, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00308f5991f4dab43c77e41957efdf5bc73c95df9dd6dd21cd63bedfec5584bd.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 140, + 255, + 488, + 401 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f00e1ab5aa2c1dbc9ec8469fa4c9cc977e74e7b2aaabc0ce703f425165f6f02b.jpg", + "image_caption": [ + "D", + "Figure 2. (A) Schematic of gaze prediction using EEG signals. (Left) The task schematic of the Visual Search and Large Grid datasets from the EEGEyeNet benchmark. (Right) EEG and eye tracking data were recorded [29]. (B) Architecture of the EEGNet-LSTM model, used to decode from EEG data. The parsimonious architecture can be used to decode saccades from EEG data. (C) Schematic of the \"brain age\" prediction task using structural MRI scans. (Left) A sample preprocessed structural MRI scan by SFCN. (right) The age and gender (inset) distribution of the TLSA dataset. (D) Model architecture of the SFCN-based decoder. The SFCN backbone was used to extract features from the MRI scan, which were used by a decoder block to predict brain age." + ], + "image_footnote": [], + "bbox": [ + 524, + 255, + 880, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, \\tilde {y} _ {i} ^ {t} \\mid \\theta\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 547, + 473, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After incorporating distributional assumptions in the above equation and subsuming constants with respect to model parameters (e.g., priors on the labels $p(y^t)$ ):", + "bbox": [ + 89, + 597, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} - \\alpha \\left(\\sum_ {i = 1} ^ {N} \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\right. \\\\ - \\sum_ {i = 1} ^ {N} \\log \\sum_ {l = 1} ^ {N} \\exp \\left(- \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right)\\right) ^ {2}\\right)) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 656, + 503, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, each update is performed with a batch of labeled and unlabeled data. First, pseudo-labels are computed for the unlabeled data. Then, keeping the pseudo-labels fixed, the second and third terms of this objective are optimized, while the labeled data are used to optimize the first term, concurrently. Note that the optimization is performed with the model parameters initialized to $\\theta^s$ , derived from the source model.", + "bbox": [ + 89, + 750, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the supervised loss in Equation 6 encourages model predictions to align with their true labels; because", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model parameters were initialized with the source model, this acts as an implicit regularizer discouraging the adapted model from deviating too far from the source model. The second term forces dissimilar predictions for data points with different pseudo-labels. In other words, providing samples with dissimilar label values during training produces disparate representations for data points with distant labels, which in turn helps to learn a better regression line. We use the log-sum-exponent trick to avoid numerical instability during optimizing of the second term.", + "bbox": [ + 511, + 527, + 906, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Datasets and Competing Methods", + "text_level": 1, + "bbox": [ + 511, + 690, + 821, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ CRAFT to address challenges relating to data scarcity, missing labels, and continuous label prediction in neuroscience.", + "bbox": [ + 511, + 715, + 905, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and models", + "text_level": 1, + "bbox": [ + 511, + 768, + 707, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Gaze prediction from EEG signals", + "text_level": 1, + "bbox": [ + 511, + 791, + 797, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Saccades are rapid ballistic eye movements, putatively linked to higher-order cognitive functions like attention and working memory [5, 13]. Decoding eye movements from neural data could help neuroscientists develop better models of these complex processes. With this objective, the EEGeyenet benchmark dataset [29] was collected from 356", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "831", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "participants. It contains $\\sim 47.5$ hours of 128-channel electroencephalogram (EEG) recordings with eye-tracking data collected concurrently.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we predict saccade amplitudes from these EEG data. We consider two datasets - the \"Large Grid\" and the \"Visual Search\" datasets - from the EEGEyeNet benchmark [29] (Fig. 2A). In the Large Grid paradigm, participants made saccades to a target location cued by a dot on the screen. In the Visual Search paradigm, participants searched for a specific symbol in the display and, thus, made saccades during the search.", + "bbox": [ + 89, + 136, + 480, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Preprocessing. Each dataset was preprocessed using the pipeline described in the benchmark (details in Appendix B.1). We decode saccade magnitudes $(y \\in \\mathbb{R}^{+})$ using the 128-channel EEG timeseries, sampled at $500\\mathrm{Hz}$ $(\\mathbf{x} \\in \\mathbb{R}^{128\\times 500})$ , for corresponding trials. Table 5 (see Appendix B.1) lists the details of the train, cross-validation and test distributions for these datasets. The Visual Search paradigm has more training data and was used to train the source model; this was then transferred to predict saccade amplitudes in the Large Grid dataset (target).", + "bbox": [ + 89, + 257, + 480, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. The EEGEyeNet benchmark showed that Pyramidal CNNs [47] were most effective at predicting saccade amplitude. As an alternative to 2D convolutional networks, we also evaluate the performance of a novel end-to-end Long-Short-Term-Memory (LSTM) model. This model is trained on features extracted by separable convolutions along time and electrode dimensions, following EEGNet [31] (Fig. 2B, more details in the Appendix B.1).", + "bbox": [ + 89, + 407, + 480, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. \"Brain age\" prediction from MRI scans", + "text_level": 1, + "bbox": [ + 89, + 534, + 413, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"Brain age\" is the biological age of the brain and is an important biomarker for cognitive health. Studies have shown that accelerated brain aging is associated with serious cognitive impairments like Alzheimer's disease [9]. Hence, estimating brain age is a critical problem in neuro-medicine.", + "bbox": [ + 89, + 551, + 480, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a first step, attempts have been made to decode the chronological age of healthy participants using T1-weighted structural MRI scans [41]. A 3D-CNN model (see next, Models) was trained using the large UK-Biobank dataset $(N = 14,503,44 - 80$ years, mean age $= 52.7$ years) [39]. We seek to transfer this pretrained model to decode age in a small Indian cohort – the TATA Longitudinal Study of Aging (TLSA) $(N = 188,46 - 88$ years, mean age $= 66$ years) [45] (Fig. 2C), without access to the UKBiobank data.", + "bbox": [ + 89, + 628, + 480, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Preprocessing. We followed a preprocessing pipeline identical to [41]. The T1w-MRI scan was brain-extracted, bias-corrected [46], and registered to standard MNI-152 space [38], and cropped to get preprocessed scans $\\mathbf{x} \\in \\mathbb{R}^{160 \\times 192 \\times 160}$ , which was used to predict brain age.", + "bbox": [ + 89, + 763, + 480, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. Features from the pretrained Simple Fully Convolutional Network (SFCN, Fig. 2D) were used to make the predictions with a regression model; more details are provided in Appendix B.2) [41]. The entire model was then", + "bbox": [ + 89, + 839, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adapted to the target dataset (TLSA).", + "bbox": [ + 513, + 90, + 759, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also apply CRAFT to people counting and tumor prediction tasks (see Appendix B.3, B.4 for dataset details).", + "bbox": [ + 511, + 106, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Competing DA Approaches", + "text_level": 1, + "bbox": [ + 511, + 147, + 761, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline model. We define a naive baseline model that always predicts the mean label of the training set, ignoring the input features of the test samples.", + "bbox": [ + 511, + 170, + 903, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transfer Learning (TL). We naively adapt the source model by minimizing the supervised loss on the labeled samples in the target dataset.", + "bbox": [ + 511, + 215, + 903, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TASFAR. This is an SF-UDA model (See Section 2) for regression problems. It depends on two statistics from the source dataset – the threshold for confident samples and the mapping between uncertainty to prediction errors. Because the source is unavailable, we compute these using the TL model on the target. For a fair comparison, we add the supervised loss on the labeled data to the unsupervised loss, weighted by the same $\\alpha$ as in our model.", + "bbox": [ + 511, + 261, + 903, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DataFree. This is another SF-UDA model (See Section 2). The original model depends on the distribution of intermediate latents on the source distribution. Again, because the source data is unavailable, we computed latent distribution using the TL model and the labeled target data. As before, the objective was modified to accommodate a supervised loss (same $\\alpha$ as in our model). Bin sizes for both TASFAR and DataFree were selected to be the same as in CRAFT.", + "bbox": [ + 511, + 383, + 903, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Progressive Mixup. This is a semi-supervised, source-free domain adaptation tool (See Section 2). We adapt the model to the regression task by replacing cross-entropy losses with mean-squared error loss. Predictive entropy was replaced by variance to estimate uncertainty.", + "bbox": [ + 511, + 503, + 903, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BBCN. Bilateral Cycle Consistency is a prototype-based clustering SF-SSDA paradigm (See Section 2). It depends on class-level prototypes for each model. To extend it to regression, we discretize the output space into bins of the same width as the other methods.", + "bbox": [ + 511, + 580, + 903, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We quantify the prediction performance of continuous valued outputs using two metrics - Root Mean Squared Error (RMSE) and Percentage-Bend Correlation [49] coefficient (R). A lower RMSE reflects a closer correspondence between the predicted and actual labels, a higher R-value quantifies better predictions along the best-fit line between the actual and predicted labels.", + "bbox": [ + 511, + 656, + 903, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In subsequent sections (5.1 and 5.2), we predict saccade amplitude from EEG and brain age from structural MRI scans, respectively.", + "bbox": [ + 511, + 762, + 903, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments: Results", + "text_level": 1, + "bbox": [ + 511, + 821, + 718, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Saccade Amplitude Prediction", + "text_level": 1, + "bbox": [ + 511, + 847, + 782, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, we trained an EEGNet-LSTM model on the Visual Search source dataset (RMSE=67.03, R=0.91). This novel", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "832", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "architecture outperformed the benchmark model (Pyramidal CNNs) by $\\sim 16\\%$ (benchmark: RMSE=79.59 pix, R=0.87). Next, this was transferred to the Large Grid dataset. The upper 3 rows of Table 1 shows the performance when all the data was used for funetuning. In this case also EEGNet-LSTM outperformed the benchmark (Pyramidal CNN [29]) by a large margin ( $\\sim 19\\%$ ). Hence, subsequent SSDA experiments were performed using EEGNet-LSTMs.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All models were trained with mini-batches of 128 samples, and only the best-performing checkpoint on the held-out validation set was used for evaluating the test set. Adam optimizer with an initial learning rate of $10^{-4}$ was used. $\\alpha$ in equation 3 was selected with a grid saerch over $\\{0.01, 0.1, 1.0\\}$ ; empirically, $\\alpha = 0.1$ was always favored.", + "bbox": [ + 89, + 212, + 480, + 303 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0af225e9a794ad1aed3fadb278fc0503c3b1a131b5d671546b8869d0c8be8617.jpg", + "table_caption": [ + "Table 1. Source Free Semi-supervised Deep Transfer of the model trained on the Visual Search task to the Large-Grid Task. The upper part of the table (rows 1-3) shows the performance ceiling when all labeled data in the target is used for training. The lower part (rows 4-9) shows SF-SSDA results when only $1\\%$ of the target samples were labeled (3 seeds)." + ], + "table_footnote": [], + "table_body": "
MethodR ↑RMSE (in pix) ↓
Naive Baseline-149.12 ± 0.02
Benchmark (TL, 100%)0.91 ± 0.0163.84 ± 0.75
EEGNet-LSTM (TL, 100%)0.93 ± 0.0151.47 ± 0.63
SF-SSDA (1% labels)
EEGNet-LSTM + TL0.77 ± 0.0192.26 ± 1.66
EEGNet-LSTM + Mixup0.48 ± 0.02135.70 ± 1.25
EEGNet-LSTM + BBCN0.76 ± 0.0199.8 ± 3.35
EEGNet-LSTM + TASFAR0.76 ± 0.0186.41 ± 1.05
EEGNet-LSTM + DataFree0.80 ± 0.0187.64 ± 3.08
EEGNet-LSTM + CRAFT0.81 ± 0.0284.17 ± 3.95
", + "bbox": [ + 91, + 407, + 483, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the efficacy of CRAFT for SF-SSDA, target labels from the training data of the Large Grid dataset were randomly dropped in a stratified manner. Figure 3A shows how the performance of these models is affected by reducing the fraction of labeled data ( $n_{ul} / N$ is the fraction of unlabeled samples). A lower proportion of labeled data worsens the performance of all the models (RMSE increases), but CRAFT produces the lowest RMSE for all unlabeled data proportions. Figure 3B shows the prediction of saccade magnitudes ( $|\\Delta \\overline{r}|$ ) for the best CRAFT-based EEGNet-LSTM model. Table 1 (lower 6 rows) summarizes the performance of CRAFT when $99\\%$ of the data were unlabeled; CRAFT outperforms supervised training (TL) by $9\\%$ and other SOTA SF-SSDA models by $>4\\%$ in terms of RMSE. The closest competitive methods were TASFAR and DataFree. The correlation coefficient (R) also showed similar trends (Table 1; Fig. 6A).", + "bbox": [ + 91, + 584, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Brain Age Prediction", + "text_level": 1, + "bbox": [ + 89, + 848, + 290, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we transferred the SFCN model trained on the UK Biobank MRI scans to the TLSA dataset. The TLSA", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset has only 188 MRI scans, so we avoid hyperparameter searches. All models were trained for 25 epochs (batch size = 4) using the Adam optimizer (initial learning rate = 10 $^{-4}$ ), and only the final model was used for inference. Here, we used $\\alpha = 0.1$ ; we explore the effect of varying $\\alpha$ for ours and competing models in Appendix D. Each metric was calculated for the prediction on the entire dataset, leaving out one fold at a time (4-folds). The top of Table 2 (row 2) quantifies the performance ceiling when all the training data are used (3 seeds).", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d195487b1e970714d34153890ec0306bc472ae82f7dd5e64bea2a2729d3acf3c.jpg", + "table_caption": [ + "Table 2. Semi-supervised transfer of the SFCN model (source: UKB) to the TLSA dataset. Same as in Table 1, except that the lower part of the table (3-8) shows SF-SSDA results when only $20\\%$ of the target samples were labeled. (3 seeds)" + ], + "table_footnote": [], + "table_body": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-7.91 ± 0.05
SFCN + TL (100%)0.66 ± 0.016.14 ± 0.03
SF-SSDA (20% labels)
SFCN + TL0.41 ± 0.077.41 ± 0.21
SFCN + Mixup0.34 ± 0.047.71 ± 0.14
SFCN + BBCN0.28 ± 0.048.00 ± 0.15
SFCN + TASFAR0.42 ± 0.077.47 ± 0.15
SFCN + DataFree0.50 ± 0.037.36 ± 0.14
SFCN + CRAFT0.51 ± 0.037.14 ± 0.11
", + "bbox": [ + 514, + 320, + 903, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For SF-SSDA, we performed experiments closely similar to the saccade task (Section 5.1). Figure 3C shows how these models' performance is affected when the proportion of unlabeled samples increases. Reducing the fraction of labeled data worsens model performance, but CRAFT continues to remain the most competitive of all models. Figure 3D shows the prediction scatter of the best CRAFT-based model. Table 2 summarizes the performance when $80\\%$ of the data was unlabeled. CRAFT outperforms supervised training (TL) by $\\sim 4\\%$ and other SOTA SF-SSDA models by $>3\\%$ in terms of RMSE. Similar results were obtained with the R values also (Fig. 6B). Again, TASFAR and DataFree emerge as the most competitive methods.", + "bbox": [ + 511, + 493, + 905, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, we evaluated the effectiveness of CRAFT at two other real-world regression benchmarks: People Counting [44, 48] and Tumor Size prediction [6, 16]. CRAFT outperformed other SF-SSDA methods by $>5\\%$ and $>2\\%$ respectively for each dataset; the details are presented in Appendices C.1 and C.2.", + "bbox": [ + 511, + 689, + 905, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mitigating Sampling Biases. Next, we show how the unsupervised CRAFT objective could mitigate sampling bias effects in the target training set. In all of the previous simulations, the label distributions of the target train and test data were matched. Here, we deliberately distorted the label distribution of the target training data: $80\\%$ of the data above the mean age in the training set were removed, creating a heavy bias toward lower ages. Following this, labels were", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "833", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13cc5ff798edf1da9859ff5eaa5b0d5d8a8f76823e76d7401ba5b67b7dfc7c2a.jpg", + "image_caption": [ + "Figure 3. (A, C) RMSE $(\\downarrow)$ for source-free semi-supervised domain adaptation of saccade (A) and brain age (C) prediction tasks, trained with varying proportions of unlabeled target data $n_{ul} / N$ . Dashed black line: Performance ceiling. Circles: individual trials. (B, D) Prediction for the best CRAFT model for saccade amplitude (B) and brain age (D), with $90\\%$ and $40\\%$ unlabeled data, respectively. Triangles: individual participants. Results with a much higher proportion of unlabeled data are shown for the saccade dataset because of the far greater size of this ( $N \\sim 12k$ ) compared to the brain age ( $N \\sim 180$ ) dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 477, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "retained in only $40\\%$ of the training data $(n_{ul} / N = 0.6)$ . No such biases were introduced into the test data. The CRAFT model was then trained by imposing the unbiased (true) marginal label distribution on the unlabeled training set (equation 5). In this scenario, again, we find that CRAFT mitigates the effect of sampling biases in the training data (Table 3). It outperforms transfer learning by $\\sim 5\\%$ , and other sota SF-SSDA models by $>2.5\\%$ (RMSE).", + "bbox": [ + 89, + 522, + 483, + 643 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/176349e5e85828d9434ac21ddb3b5fc2bdbae599b628802a2f56a885a8a4ed7a.jpg", + "table_caption": [ + "Table 3. Analyses with datasets with sampling error. $80\\%$ of the data above the mean age is removed. Furthermore, only $40\\%$ of the labels in this mislabeled data are used for semi-supervised transfer of the SFCN model (source: UKB). (3 seeds)" + ], + "table_footnote": [], + "table_body": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-8.38 ± 0.22
SFCN + TL0.41 ± 0.058.40 ± 0.27
SFCN + Mixup0.32 ± 0.028.19 ± 0.15
SFCN + BBCN0.23 ± 0.028.53 ± 0.16
SFCN + TASFAR0.39 ± 0.048.57 ± 0.22
SFCN + DataFree0.41 ± 0.038.58 ± 0.23
SFCN + CRAFT0.45 ± 0.037.98 ± 0.25
", + "bbox": [ + 91, + 736, + 483, + 877 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Analysis of computational complexity", + "text_level": 1, + "bbox": [ + 511, + 90, + 838, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 shows the computational complexity of CRAFT and the competing models. We compute both the algorithms' theoretical complexity and running time on the saccade and brain age prediction tasks. The theoretical training complexity is defined as the number of gradient updates per epoch. For inference, the complexity is the same for all models as the same architecture is trained using different adaptation paradigms. Training time and complexity are lowest for TASFAR, amongst all other SF-DA tools. The complexity of CRAFT was comparable to DataFree, the next best method.", + "bbox": [ + 511, + 112, + 906, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7daf5f58a8e81844ba2fb3594d433f31abd53c81c9557d05107f68e3b9934c93.jpg", + "table_caption": [ + "Table 4. Computational complexity. Time per training epoch and for all test data (in min). Parentheses: P- no. of model parameters; N- samples; D- latent dimensionality; $\\beta$ - batch size, B- no. of bins, m- no. of inferences for MC Dropout uncertainty estimation." + ], + "table_footnote": [], + "table_body": "
(B) MethodEEGNet-LSTM (P~ 190K)SFCN (P~ 2.95M)
Train /Comp. ↓ (N ~ 12k)Test /Comp. ↓ (N ~ 2.5k)Train/Comp. ↓ (N ~ 150)Test/Comp. ↓ (N ~ 50)
TL0.15 / O(N)0.02 / O(N)0.12 / O(N)0.04 / O(N)
Mixup0.74 / O(Nm)0.02 / O(N)0.74 / O(Nm)0.04 / O(N)
BBCN2.71 / O(NBD)0.02 / O(N)1.29 / O(NBD)0.04 / O(N)
TASFAR0.30 / O(N)0.02 / O(N)0.19 / O(N)0.04 / O(N)
DataFree0.55 / O(ND2)0.02 / O(N)0.32 / O(ND2)0.04 / O(N)
CRAFT0.45 / O(N(β + B))0.02 / O(N)0.36 / O(N(β + B))0.04 / O(N)
", + "bbox": [ + 514, + 354, + 903, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Concluding Remarks", + "text_level": 1, + "bbox": [ + 511, + 482, + 718, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data scarcity poses a serious challenge for applying deep-learning models in real-world settings. Training overparameterized models on a few hundred data points leads to overfitting and poor generalization. Conventional domain adaptation algorithms, to transfer pretrained networks, typically need access to source domain data, but these may not always be available. In this study, we propose a source-free semi-supervised domain adaptation framework that recalibrates the decision boundary without aligning the intermediate representation of the domains. In fields like medicine, resource constraints associated with labeling data are significant; therefore, leveraging the large amounts of unlabeled data becomes critical. We propose CRAFT as an SF-SSDA approach for such source-free, label-sparse settings that are common in real-world applications.", + "bbox": [ + 511, + 507, + 908, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 747, + 700, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the following funding sources: a Prime Minister's Research Fellowship (to M.B.), a DST SwarnaJayanti fellowship, a Pratiksha Trust Intramural grant, a Tata Trusts grant and a Google Research grant (to D.S.). We also thank the Tata Longitudinal Study on Aging at the Indian Institute of Science for access to MRI scans from their cohort.", + "bbox": [ + 511, + 772, + 908, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "834", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiayu An, Changming Zhao, and Dongrui Wu. Semi-supervised generalized source-free domain adaptation (ssgsfda). In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2023. 3", + "[2] APTOS. Aptos 2019 blindness detection dataset, 2019. 1", + "[3] Sourabh Balgi and Ambedkar Dukkipati. CUDA: Contradistinguisher for unsupervised domain adaptation. In Proceedings of the IEEE International Conference on Data Mining, 2019. 1", + "[4] Sourabh Balgi and Ambedkar Dukkipati. Contradistinginguisher: A Vapnik's imperative to unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4730-4747, 2022. 1, 3, 4", + "[5] Scott L. Brincat, Jacob A. Donoghue, Meredith K. Kahnke, Simon Kornblith, Mikael Lundqvist, and Earl K. Miller. Interhemispheric transfer of working memories. Neuron, 109: 1055-1066.e4, 2021. 5", + "[6] Peter Bandy, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, Quanzheng Li, Farhad Ghazvinian Zanjani, Svitlana Zinger, Keisuke Fukuta, Daisuke Komura, Vlado Ovtcharov, Shenghua Cheng, Shaoqun Zeng, Jeppe Thagaard, Anders B. Dahl, Huangjing Lin, Hao Chen, Ludwig Jacobsson, Martin Hedlund, Melih Çetin, Eren Halici, Hunter Jackson, Richard Chen, Fabian Both, Jörg Franke, Heidi Küsters-Vandevelde, Willem Vreuls, Peter Bult, Bram van Ginneken, Jeroen van der Laak, and Geert Litjens. From detection of individual metastases to classification of lymph node status at the patient level: The camelyon17 challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2019. 2, 7, 3", + "[7] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140: 325-331, 2020. 2", + "[8] Thomas A. Carlson. Orientation decoding in human visual cortex: New insights from an unbiased perspective. Journal of Neuroscience, 34(24):8373-8383, 2014. 1", + "[9] J H Cole, S J Ritchie, M E Bastin, M C Valdés Hernández, S Muñoz Maniega, N Royle, J Corley, A Pattie, S E Harris, Q Zhang, N R Wray, P Redmond, R E Marioni, J M Starr, S R Cox, J M Wardlaw, D J Sharp, and I J Deary. Brain age predicts mortality. Molecular Psychiatry, 23:1385-1392, 2018. 6", + "[10] Sébastien M. Crouzet. Fast saccades toward faces: Face detection in just 100 ms. Journal of Vision, 10:1-17, 2010. 2", + "[11] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. PMID: 20144289. 1", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Heiner Deubel and Werner X. Schneider. Saccade target selection and object recognition: Evidence for a common attentional mechanism. Vision Research, 36:1827-1837, 1996. 5", + "[14] Yuhe Ding, Lijun Sheng, Jian Liang, Aihua Zheng, and Ran He. Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks, 167:92-103, 2023. 2", + "[15] Cian Eastwood, Ian Mason, Christopher K I Williams, and Bernhard Scholkopf. In Proceedings of the Tenth International Conference on Learning Representations, 2022. Tenth International Conference on Learning Representations 2022, ICLR 2022; Conference date: 25-04-2022 Through 29-04-2022. 2, 3", + "[16] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen A. W. M. van der Laak, , and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 318(22):2199-2210, 2017. 2, 7, 3", + "[17] Saminger-Platz et al. Central moment discrepancy (cmd) for domain-invariant representation learning. *ICLR*, 2017. 1, 2", + "[18] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 174:106230, 2024. 1", + "[19] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17:1-35, 2016. 1, 2", + "[20] Alessandro T. Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M. Cichy. The algonauts project 2025 challenge: How the human brain makes sense of multimodal movies, 2025. 1", + "[21] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems. MIT Press, 2004. 3, 4", + "[22] Arthur Gretton, Kenji Fukumizu, Zäïd Harchaoui, and Bharath K. Sriperumbudur. A fast, consistent kernel two-sample test. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2", + "[23] Yan Hao, Yuhong Guo, and Chunsheng Yang. Source-free unsupervised domain adaptation with surrogate data generation. In British Machine Vision Conference, 2021. 3", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Computer Vision - ECCV 2016, pages 630-645, Cham, 2016. Springer International Publishing. 3", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3", + "[26] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and Shueng-Han Gary Chan. Target-agnostic source-free domain adaptation for regression tasks. 2024 IEEE 40th In" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "835", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Data Engineering (ICDE), pages 1464-1477, 2023. 2, 3", + "[27] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 2", + "[28] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 3", + "[29] Ard Kastrati, Martyna Plomecka, Damian Pascual Ortiz, Lukas Wolf, Victor Gillioz, Roger Wattenhofer, and Nicolas Langer. Eegeyenet: a simultaneous electroencephalography and eye-tracking dataset and benchmark for eye movement prediction. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 1, 2, 5, 6, 7", + "[30] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method, 2021. 2", + "[31] Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. Eeg-net: a compact convolutional neural network for eeg-based brain-computer interfaces. Journal of Neural Engineering, 15:056013, 2018. 6, 2", + "[32] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A Comprehensive Survey on Source-Free Domain Adaptation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 46(08):5743-5762, 2024. 1, 2", + "[33] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2", + "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the 32nd International Conference on Machine Learning, pages 97-105, Lille, France, 2015. PMLR. 2", + "[35] Jie Ma, Zhijun Xu, Xinyu Xiong, MaYiDiLi NiJiaTi, Dongyu Zhang, Weijun Sun, Feng Gao, and Guanbin Li. Source-free semi-supervised domain adaptation for tuberculosis recognition. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5, 2024. 2, 3", + "[36] Ning Ma, Jiajun Bu, Lixian Lu, Jun Wen, Zhen Zhang, Sheng Zhou, and Xifeng Yan. Semi-supervised hypothesis transfer for source-free domain adaptation, 2021. 2, 3", + "[37] Ning Ma, Haishuai Wang, Zhen Zhang, Sheng Zhou, Hongyang Chen, and Jiajun Bu. Source-free semi-supervised domain adaptation via progressive mixup. Knowledge-Based Systems, 262:110208, 2023. 2, 3", + "[38] Pravat K. Mandal, Rashima Mahajan, and Ivo D. Dinov. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. Journal of Alzheimer's Disease, 31:S169-S188, 2012. 6", + "[39] Karla L Miller, Fidel Alfaro-Almagro, Neal K Bangerter, David L Thomas, Essa Yacoub, Junqian Xu, Andreas J Bartsch, Saad Jbabdi, Stamatios N Sotiropoulos, Jesper L R Andersson, Ludovica Griffanti, Gwenaëlle Douaud," + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Thomas W Okell, Peter Weale, Iulius Dragonu, Steve Garratt, Sarah Hudson, Rory Collins, Mark Jenkinson, Paul M Matthews, and Stephen M Smith. Multimodal population brain imaging in the uk biobank prospective epidemiological study. Nature Neuroscience, 19:1523-1536, 2016. 1, 6", + "[40] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation, 2020. 1, 2", + "[41] Han Peng, Weikang Gong, Christian F Beckmann, Andrea Vedaldi, and Stephen M Smith. Accurate brain age prediction with lightweight deep neural networks. Medical Image Analysis, 68:101871, 2021. 6, 2", + "[42] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In Computer Vision - ECCV 2022, pages 165-182, Cham, 2022. Springer Nature Switzerland. 2", + "[43] Xiaofeng Qu, Li Liu, Lei Zhu, Liqiang Nie, and Huaxiang Zhang. Source-free style-diversity adversarial domain adaptation with privacy-preservation for person re-identification. Knowledge-Based Systems, 283:111150, 2024. 1", + "[44] Vishwanath Sindagi, Rajeev Yasarla, and Vishal Patel. Pushing the Frontiers of Unconstrained Crowd Counting: New Dataset and Benchmark Method. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1221-1231, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 2, 7", + "[45] Jonas S Sundarakumar, Albert Stezin, Abhishek L Menes-gere, and Vijayalakshmi Ravindranath. Rural-urban and gender differences in metabolic syndrome in the aging population from southern india: Two parallel, prospective cohort studies. eClinicalMedicine, 47:101395, 2022. 1, 2, 6", + "[46] J-Donald Tournier, Robert Smith, David Raffelt, Rami Tabbara, Thijs Dhollander, Maximilian Pietsch, Daan Christi-aens, Ben Jeurissen, Chun-Hung Yeh, and Alan Connelly. Mrtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202:116137, 2019. 6", + "[47] Ihsan Ullah and Alfredo Petrosino. About pyramid structure in convolutional neural networks, 2016. 6, 2", + "[48] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2, 7", + "[49] Rand R. Wilcox. The percentage bend correlation coefficient. Psychometrika, 59:601-616, 1994. 6", + "[50] Garrett Wilson and Diane J. Cook. A survey of unsupervised deep domain adaptation. ACM Trans. Intell. Syst. Technol., 11(5), 2020. 1", + "[51] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36(8):3746-3766, 2021. 2", + "[52] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "836", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_model.json b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c73e2d6e5fed6d3140197ab9eeb2c1a5379c51e2 --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Semi-supervised Deep Transfer for Regression without Domain Alignment" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.774, + 0.198 + ], + "angle": 0, + "content": "Mainak Biswas1,2, Ambedkar Dukkipati3, Devarajan Sridharan1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.199, + 0.918, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Brain, Computation and Data Sciences, \\(^{2}\\)Centre for Neuroscience, \\(^{3}\\)Computer Science and Automation Indian Institute of Science, Bangalore - 560012" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.238, + 0.792, + 0.251 + ], + "angle": 0, + "content": "mainakbiswas@iisc.ac.in, ambedkar@iisc.ac.in, sridhar@iisc.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.486, + 0.852 + ], + "angle": 0, + "content": "Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data or reliably labeled target data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT - a Contradistinguisher-based Regularization Approach for Flexible Training - for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and \"brain age\" prediction with structural MRI data. For both datasets, CRAFT yielded up to \\(9\\%\\) improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than \\(3\\%\\). Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.903 + ], + "angle": 0, + "content": "Keywords: source-free DA, continuous label prediction, gaze prediction, brain age prediction, healthcare" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.287, + 0.645, + 0.303 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.907, + 0.508 + ], + "angle": 0, + "content": "For the successful application of deep learning models in the real world, they must be robust to \"domain shift\" [50]. For example, in biology and medicine, supervised deep learning models are often trained on large, high-quality source datasets collected in resource-rich settings [11, 39]. However, naive transfer or even simple finetuning fails to produce high accuracies when tested with smaller target datasets of mixed quality collected [2, 45]. As a result, state-of-the-art domain adaptation (DA) approaches, which enable the source models to generalize successfully to target datasets, must be employed. Many of these approaches involve explicit alignment of source and target, while learning a shared predictive model for both domains [17, 19, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.704 + ], + "angle": 0, + "content": "A particular challenge occurs with DA when the source dataset is unavailable. This can happen either because of privacy or proprietary concerns, such as with medical images or patents [43]. Alternatively, source datasets may be too large, incurring prohibitive storage and computation costs in low-resource settings [39]. Source-free (SF) domain adaptation methods that address these challenges are getting increasingly popular [18, 32]. Such methods generalize the source model to unlabeled or partially target data (unsupervised or semi-supervised DA) with one of two approaches: data-based or model-based. While data-based methods create surrogate source data, model-based methods adapt the source model to the target domain (see Section 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.902 + ], + "angle": 0, + "content": "We address model-based SF-DA in a semi-supervised setting. Specifically, we focus on neuroscience applications where source data are inaccessible, and target labels are reliably available only for a few samples. Regression tasks – like predicting brain age (Fig. 1C) [39], saccades [29], neural activity [20], and stimulus orientation [8] – are ubiquitous in neuroscience; we tackle a subset of these. We build upon Contradistinguisher [3, 4], a recently-proposed tool for effective unsupervised DA (CUDA). Unlike popular representation alignment DA methods [19, 40], CUDA directly learns a common model for both domains. Yet, CUDA requires full access to source data and addresses classification tasks. Therefore, it is not directly applicable to our problem." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "827" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.105, + 0.119, + 0.115 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.099, + 0.228, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.102, + 0.246, + 0.114 + ], + "angle": 0, + "content": "X" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.102, + 0.361, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.098, + 0.48, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.203, + 0.117, + 0.213 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.214, + 0.27, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.203, + 0.293, + 0.213 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.196, + 0.477, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Figure 1. Schematic of Semisupervised Deep Transfer without Domain Alignment. (A) The approach seeks to transfer a pretrained model to a target domain with limited labeled data when the source dataset is unavailable. (B) Because the source dataset is unavailable, we tune the decision boundaries without explicitly aligning the representation of the domains, combining a supervised loss and the CUDA loss (regularizer). (C) The algorithm targets deep transfer in regression tasks, like brain age prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Here, we leverage CUDA to develop a new algorithm - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT) which performs deep transfer, without domain alignment, even when the source dataset is unavailable (Fig. 1A-B). We also extend CUDA to a regression setting. Previous studies approach regression tasks naively by binning continuous labels into discrete bins [40], sometimes employing a rank-ordinal objective [7]. Regardless, the efficacy of such approaches may depend critically on bin sizes. Our CRAFT model uses a principled approach to extend CUDA to predictions involving continuous-valued outputs. We employ binning as a practical strategy to generate and optimize pseudo-labels, without constraining the model to produce discrete outputs. Our study makes the following main contributions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.69, + 0.483, + 0.719 + ], + "angle": 0, + "content": "- First, we extend CUDA to source-free DA for regression tasks involving continuous-valued outputs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- Second, we formulate a theoretically-motivated semi-supervised objective by combining supervised and unsupervised losses with a tunable hyperparameter \\((\\alpha)\\) (Section 3). We also derive the unsupervised loss as a maximum entropy prior to model parameters (Appendix A.2)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.797, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- Third, we show the effectiveness of CRAFT on two neuroscience challenges - gaze prediction from electroencephalographic (EEG) [29] data and brain age prediction from magnetic resonance imaging (MRI) data [45]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.857, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Fourth, we also apply CRAFT to two real-world regression benchmarks involving people counting [44, 48] and tumor size prediction [6, 16]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "- Fifth, we compare CRAFT against state-of-the-art SF-DA methods [15, 26, 35, 37] and show that the best improvements occur as unlabeled target samples increase, demonstrating its efficacy for semi-supervised SF-DA." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Finally, we analyze the computational complexity and demonstrate numbers competitive with sota models." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.861, + 0.198 + ], + "angle": 0, + "content": "- Our code is publicly available (see Appendix E.2)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.211, + 0.655, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.907, + 0.447 + ], + "angle": 0, + "content": "Domain alignment methods. Popular Unsupervised Domain Adaptation (UDA) approaches rely on learning common intermediate representations across source and target domains. Domain Adversarial training of Neural Networks (DANN) [19] uses a discriminator to learn common representations across both domains adversarially. In Importance Weighting [40] (IW), higher importance is afforded to losses corresponding to source samples more likely to belong to the target domain. Central Moment Discrepancy (CMD) [17] minimizes the central moments of the latent representations between the domains, while others optimize their Maximum Mean Discrepancy (MMD) [22, 34]. Despite being effective, they cannot be used when the source is unavailable and will not be discussed further." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Source-free UDA. To address DA scenarios where source data and its target labels are unavailable, several source-free unsupervised DA approaches have been recently proposed. Broadly, they are categorized as: Data-based and Model-based [32]. Data-based methods generate a source-like distribution [14, 30]. However, they often depend on the availability of generative models trained on the source, which could be unavailable. For example, SF-UDA through domain-specific perturbation (AUGFree) [51] assumes the existence of domain-invariant features for prediction. Since the source is unavailable, AUGFree perturbs the target data and then learns a transformation to the invariant features by adversarially aligning the target and the perturbed target." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "On the other hand, model-based methods finetune the source model on the target using strategies like contrastive learning [27], pseudo-labeling [26], or entropy minimization [36]. They are often prone to overfitting to pseudolabels. For example, Class-Balanced Multicentric Dynamic Prototype Strategy for SF-UDA (BMD) [42] follows a pseudo-labeling strategy using class prototypes based on the source model predictions, and in turn uses these labels for training. Source Data-Absent UDA Through Hypothesis Transfer (SHOT++) [33] combines three objectives - a) maximizing mutual information between the predictions and inputs, b) self-supervised loss based on prototype-based clusters, and c) a Mixup-based (See SF-SSDA) loss [37]. Attracting and Dispersing (AaD) [52] defines a neighboring cluster (K nearest neighbors) and a randomly selected background cluster for each sample, and maximizes the log-likelihood of the ratio of their distributions." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Target-agnostic SF UDA for Regression Tasks (TAS-FAR) [26] works on the principle that prediction uncertainty is correlated with prediction error. It uses Monte-Carlo dropout to estimate label uncertainty and divides the target dataset into confident and uncertain samples. A discrete density map is estimated using the confident data, which is then used as weights to recalibrate the pseudo-labels of uncertain data. SF-UDA to measurement Shift via Bottom-Up Feature Restoration (BUFR/DataFree) [15] is designed on the principle of intermediate feature alignment of the source and the target distributions. During adaptation, the model minimizes the symmetric KL divergence between feature distributions of source (fixed) and target data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Here, we compare our model's performance against two of the above methods - TASFAR and DataFree. TASFAR was selected as it is a pseudo-labeling-based method suitable for regression. On the other hand, DataFree is a pseudo-label-free feature alignment method, not constrained to classification tasks. By adding a supervised loss (Section 4.2), we evaluate these as semi-supervised models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.687 + ], + "angle": 0, + "content": "Source-free SSDA. Source-free semi-supervised domain adaptation (SF-SSDA) models best address the scenario of SF-DA when few target labels are available. For example, Generalized SF-SSDA [1] minimizes source loss alongside an unsupervised target loss which maximizes the cosine similarity between predictions of k-nearest neighbors. The model is then transferred by adversarial training based on source features generated using a Conditional VAE. Similarly, [23] proposes generating surrogate source data from target samples using gradient-based optimization. Learning from Different Samples for SF-SSDA [28] minimizes the contrastive loss on the predicted probabilities of the target domain. Furthermore, it minimizes supervised and contrastive loss on a set comprising labeled and unlabeled data with low uncertainty. Likewise, Mutual Enhancement training for SS Hypothesis transfer (MESH) [36] minimizes the supervised loss on the labeled samples and entropy on pseudo-labeled samples. Finally, the target manifold is smoothened by making predictions robust to perturbations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Progressive Mixup [37] is trained on augmentations by convex combinations of data on which the source model is confident and ones on which it is uncertain (hybrid mixup). Similarly, another set is obtained by convex combinations of samples in the (pseudo) labeled dataset (self-mixup). In contrast, Bilateral-Branch Consistency Network (BBCN) [35] computes prototypes as the weighted mean of feature representations of the target dataset by the source model. Pseudo-labels are selected based on the proximity of unlabeled data to these prototypes. A separate target branch is trained using the pseudo-labels, and a consistency loss is added between the representations of both branches. The need for class-level prototype generation makes it challenging to adapt this model to regression tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "We compare the performance of CRAFT with two SFSSDA methods - SF-SSDA via progressive Mixup [37], and BBCN [35], as they have little dependence on the source distribution and its derivatives. Moreover, these methods are based on complementary ideas, and the latter [35] has been applied specifically in medical imaging, for tuberculosis recognition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.214, + 0.704, + 0.23 + ], + "angle": 0, + "content": "3. The CRAFT Model" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.24, + 0.696, + 0.255 + ], + "angle": 0, + "content": "3.1. Model formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.905, + 0.475 + ], + "angle": 0, + "content": "Background. Our objective is to develop an approach for source-free semi-supervised domain adaptation, for regression tasks. For this, we build upon the Contradistinguisher-based Unsupervised Domain Adaptation (CUDA) model [4]. CUDA obviates the need for learning a common intermediate representation across the source and the target datasets. It optimizes a joint distribution over the target features and category labels to adapt the source model. Because CUDA is intended for unsupervised domain adaptation (UDA), target labels are unavailable. Hence, the algorithm uses a two-step joint optimization over the target labels and model parameters. First, it generates pseudo-labels on the target dataset from the source model [21]. Then, it updates the model parameters by fixing the pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.627 + ], + "angle": 0, + "content": "In this study, we extend CUDA to a source-free semi-supervised setting (SF-SSDA). Briefly, to the supervised label loss, we add a 'regularized', unsupervised loss as a prior on the model parameters. As we will show in the next section, the model can be flexibly applied, either in SF-UDA case or in the SF-SSDA case, depending on the weightage given to the supervised loss. We, moreover, extend the model to continuous label prediction. We call our new formulation CRAFT - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Objective. We consider a task in which a source model \\(\\theta^s\\) trained on a source dataset - involving continuous label prediction (regression) - is already available to us. The target dataset comprises i.i.d samples from an unknown distribution \\(p^t (\\mathbf{x},y)\\): \\(\\mathcal{D}^t = \\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}\\), \\(\\{\\mathbf{x}_i^t\\}_{i = 1}^{N_{ul}}\\) where \\(\\mathbf{x}_i^t\\in \\mathbb{R}^d\\) is a feature-vector, and \\(y_{i}^{t}\\in \\mathbb{R}\\) is its corresponding label. Target training data are partially and sparingly labeled; we define the proportion of unlabeled samples \\((N_{ul})\\) to the total number of samples \\((N)\\), as \\(N_{ul}:(N_l + N_{ul})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We formulate, first, the supervised objective. For continuous label prediction problems, the conditional density of the label given the features is assumed to be normally distributed. The mean is modeled using a neural network \\( f(:,\\theta) \\), parameterized by \\( \\theta \\), as: \\( p(y^{t}|\\mathbf{x}^{t},\\theta) = \\mathcal{N}(y^{t};f(\\mathbf{x}^{t};\\theta),c) \\), and the variance, \\( c \\) is assumed to be constant as it simplifies the optimization. In the final formulation, the variance occurs as the weight of the unsupervised loss, \\( \\alpha \\) (equation 3); therefore, only one of the two" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "829" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "hyperparameters (\\(c\\) or \\(\\alpha\\)) needs to be tuned. Hence, the log-likelihood can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.133, + 0.483, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\log p \\left(\\mathcal {D} ^ {t} | \\theta\\right) \\propto - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.485, + 0.226 + ], + "angle": 0, + "content": "The unsupervised objective is essentially identical to the CUDA objective and is defined as a joint log density over the labels and model parameters as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.233, + 0.483, + 0.275 + ], + "angle": 0, + "content": "\\[\n\\log q \\left(\\mathbf {x} ^ {t}, y ^ {t} \\mid \\theta\\right) = \\log \\left[ \\frac {p \\left(y ^ {t} \\mid \\mathbf {x} ^ {t} , \\theta\\right)}{\\sum_ {i = 1} ^ {N} p \\left(y ^ {t} \\mid \\mathbf {x} _ {i} ^ {t} , \\theta\\right)} p \\left(y ^ {t}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.434 + ], + "angle": 0, + "content": "This joint distribution seeks to reduce the model's predictive bias when faced with a distribution shift in the target domain (see Fig. 1B). The denominator, which reflects the total probability of predicting a particular class, normalizes for biases originating from the source data. In other words, model predictions are normalized as \\( p(y^{t}|\\mathbf{x}^{t},\\theta) / \\sum_{i = 1}^{N}p(y^{t}|\\mathbf{x}_{i}^{t},\\theta) \\), such that no class has a biased probability of being predicted. Finally, the model predictions align with the target label distribution, using the supplied marginal \\( p(y^t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Note, that this objective does not depend on whether the samples are labeled or not. Optimizing the joint distribution: 1) normalizes inherent bias towards one label over the other by calibrating the prediction for a label with its total weight in the dataset, 2) and explicitly trains \\( q \\) to match the marginals - i.e., \\( \\sum_{i=1}^{N} q(\\mathbf{x}_i^t, y^t|\\theta) = p(y^t) \\), thereby matching high-level statistics like per-class proportions in the training distribution. Importantly, with an informative prior \\( p(y^t) \\) that specifies the putative distribution of the target labels, pseudo-label prediction can be biased toward labels that are more likely (see Section 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Combining the supervised and the unsupervised objective, the semi-supervised CRAFT objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.64, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\mathcal {D} ^ {t}, \\theta\\right) = \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, y _ {i} ^ {t} \\mid \\theta\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.779 + ], + "angle": 0, + "content": "In the above equation, \\(\\alpha\\) is the weight associated with the unsupervised objective and is tuned as a hyperparameter on a held-out cross-validation set during training. For deep-transfer, model parameters are initialized to \\(\\theta^s\\) derived from the source model, and then gradient descent is performed to obtain \\(\\theta^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Theoretical motivation. Artificial Neural Networks (ANNs) with hypotheses space \\(\\Theta\\) are trained to learn a parametric approximation \\(p(\\mathbf{x},y|\\theta)\\), \\(\\theta \\in \\Theta\\), such that the log-likelihood of the training data given the parameters is maximized. However, training ANNs often does not account for prior distributions over model parameters. In Appendix A.1, we demonstrate that the CRAFT semi-supervised objective can be theoretically derived as MAP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "estimation of the parameters \\((\\log p(\\theta|\\mathcal{D}))\\), which we write as a sum of the supervised objective, \\(\\log p(\\mathcal{D}|\\theta)\\), and a prior on the model parameters, \\(p(\\theta)\\). Specifically, we demonstrate, using the principle of entropy maximization [21], that the unsupervised objective is related to the prior on the model parameters." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.19, + 0.704, + 0.207 + ], + "angle": 0, + "content": "3.2. Model optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.906, + 0.302 + ], + "angle": 0, + "content": "We optimize the supervised loss of the CRAFT objective with conventional gradient descent, jointly optimizing both components in the loss (equation 3). The unsupervised component of the loss is dependent on the labels \\( y^{t} \\), which are unavailable for the vast majority of target data. To tackle this, we employ a 2-step optimization process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.906, + 0.348 + ], + "angle": 0, + "content": "Pseudo-label selection. As the first step, we fix the model parameters \\(\\theta\\) and select the label that maximizes the joint distribution for each datapoint:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.354, + 0.907, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} q \\left(\\mathbf {x} _ {i} ^ {t}, y ^ {t} \\mid \\theta\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.538 + ], + "angle": 0, + "content": "where \\(\\tilde{y}_i\\) the pseudo-label for feature vector \\(\\mathbf{x}_i\\). For classification tasks, the optimization in equation 4 can be achieved by selecting the best label from a finite set of class labels that maximizes \\(q\\) [4]. But, for regression problems such as ours, class labels map to real numbers spanning the range of the model outputs \\(f(:,\\theta):\\mathbb{R}^d\\mapsto [a,b]\\). To address this, we rewrite the above equation under the assumption that the continuous-valued labels are distributed normally, with mean \\(f(\\mathbf{x}^t;\\theta)\\), and variance \\(c\\). Hence, the above equation can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.542, + 0.907, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} \\frac {\\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {i} ^ {t} ; \\theta\\right) , c\\right) p \\left(y ^ {t}\\right)}{\\sum_ {l = 1} ^ {N} \\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right) , c\\right)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Equation 5 can be optimized with gradient ascent with respect to \\( y^{t} \\) for the objective log \\( q(\\mathbf{x}_i^t, y^t|\\theta) \\). But, the pseudo-labels must be estimated before gradient descent can be performed on the model parameters (see next step, equation 6). This requires nested gradient descent, which increases computational complexity. Moreover, to employ an informative prior over target labels, we fit a class of mixture models (described in Appendix A.3) to estimate \\( p(y) \\) from the data; this makes backpropagation more cumbersome. To overcome these challenges, we divide the entire range of labels into small, discrete bins whose midpoints are the candidate pseudo-labels. Hence, this step in the optimization can be achieved faster pseudo-label selection from a discrete set. Note that this approach does not transform the model outputs to be discrete; it merely represents a convenient optimization strategy to update the parameters. In the experimental section, we provide a recipe for choosing bin sizes and study the impact of varying bin sizes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Maximization. In the second step, we optimize the model parameters by using the labeled data \\(\\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}\\) and fixing the pseudo-labels \\(\\{\\mathbf{x}_i^t,\\tilde{y}_i^t\\}_{i = 1}^N\\) , as:" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "830" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.099, + 0.142, + 0.109 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.105, + 0.484, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.099, + 0.526, + 0.111 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.121, + 0.705, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.111, + 0.88, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.248, + 0.139, + 0.26 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.256, + 0.489, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.251, + 0.525, + 0.262 + ], + "angle": 0, + "content": "D" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.256, + 0.882, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.432, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 2. (A) Schematic of gaze prediction using EEG signals. (Left) The task schematic of the Visual Search and Large Grid datasets from the EEGEyeNet benchmark. (Right) EEG and eye tracking data were recorded [29]. (B) Architecture of the EEGNet-LSTM model, used to decode from EEG data. The parsimonious architecture can be used to decode saccades from EEG data. (C) Schematic of the \"brain age\" prediction task using structural MRI scans. (Left) A sample preprocessed structural MRI scan by SFCN. (right) The age and gender (inset) distribution of the TLSA dataset. (D) Model architecture of the SFCN-based decoder. The SFCN backbone was used to extract features from the MRI scan, which were used by a decoder block to predict brain age." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.549, + 0.474, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, \\tilde {y} _ {i} ^ {t} \\mid \\theta\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.598, + 0.484, + 0.645 + ], + "angle": 0, + "content": "After incorporating distributional assumptions in the above equation and subsuming constants with respect to model parameters (e.g., priors on the labels \\( p(y^t) \\)):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.657, + 0.504, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} - \\alpha \\left(\\sum_ {i = 1} ^ {N} \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\right. \\\\ - \\sum_ {i = 1} ^ {N} \\log \\sum_ {l = 1} ^ {N} \\exp \\left(- \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right)\\right) ^ {2}\\right)) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In practice, each update is performed with a batch of labeled and unlabeled data. First, pseudo-labels are computed for the unlabeled data. Then, keeping the pseudo-labels fixed, the second and third terms of this objective are optimized, while the labeled data are used to optimize the first term, concurrently. Note that the optimization is performed with the model parameters initialized to \\(\\theta^s\\), derived from the source model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Intuitively, the supervised loss in Equation 6 encourages model predictions to align with their true labels; because" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.68 + ], + "angle": 0, + "content": "model parameters were initialized with the source model, this acts as an implicit regularizer discouraging the adapted model from deviating too far from the source model. The second term forces dissimilar predictions for data points with different pseudo-labels. In other words, providing samples with dissimilar label values during training produces disparate representations for data points with distant labels, which in turn helps to learn a better regression line. We use the log-sum-exponent trick to avoid numerical instability during optimizing of the second term." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.691, + 0.822, + 0.709 + ], + "angle": 0, + "content": "4. Datasets and Competing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.906, + 0.761 + ], + "angle": 0, + "content": "We employ CRAFT to address challenges relating to data scarcity, missing labels, and continuous label prediction in neuroscience." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.708, + 0.784 + ], + "angle": 0, + "content": "4.1. Datasets and models" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.799, + 0.807 + ], + "angle": 0, + "content": "4.1.1. Gaze prediction from EEG signals" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Saccades are rapid ballistic eye movements, putatively linked to higher-order cognitive functions like attention and working memory [5, 13]. Decoding eye movements from neural data could help neuroscientists develop better models of these complex processes. With this objective, the EEGeyenet benchmark dataset [29] was collected from 356" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "participants. It contains \\(\\sim 47.5\\) hours of 128-channel electroencephalogram (EEG) recordings with eye-tracking data collected concurrently." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Here, we predict saccade amplitudes from these EEG data. We consider two datasets - the \"Large Grid\" and the \"Visual Search\" datasets - from the EEGEyeNet benchmark [29] (Fig. 2A). In the Large Grid paradigm, participants made saccades to a target location cued by a dot on the screen. In the Visual Search paradigm, participants searched for a specific symbol in the display and, thus, made saccades during the search." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.408 + ], + "angle": 0, + "content": "Preprocessing. Each dataset was preprocessed using the pipeline described in the benchmark (details in Appendix B.1). We decode saccade magnitudes \\((y \\in \\mathbb{R}^{+})\\) using the 128-channel EEG timeseries, sampled at \\(500\\mathrm{Hz}\\) \\((\\mathbf{x} \\in \\mathbb{R}^{128\\times 500})\\), for corresponding trials. Table 5 (see Appendix B.1) lists the details of the train, cross-validation and test distributions for these datasets. The Visual Search paradigm has more training data and was used to train the source model; this was then transferred to predict saccade amplitudes in the Large Grid dataset (target)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.482, + 0.53 + ], + "angle": 0, + "content": "Models. The EEGEyeNet benchmark showed that Pyramidal CNNs [47] were most effective at predicting saccade amplitude. As an alternative to 2D convolutional networks, we also evaluate the performance of a novel end-to-end Long-Short-Term-Memory (LSTM) model. This model is trained on features extracted by separable convolutions along time and electrode dimensions, following EEGNet [31] (Fig. 2B, more details in the Appendix B.1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.414, + 0.55 + ], + "angle": 0, + "content": "4.1.2. \"Brain age\" prediction from MRI scans" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.482, + 0.629 + ], + "angle": 0, + "content": "\"Brain age\" is the biological age of the brain and is an important biomarker for cognitive health. Studies have shown that accelerated brain aging is associated with serious cognitive impairments like Alzheimer's disease [9]. Hence, estimating brain age is a critical problem in neuro-medicine." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.482, + 0.764 + ], + "angle": 0, + "content": "As a first step, attempts have been made to decode the chronological age of healthy participants using T1-weighted structural MRI scans [41]. A 3D-CNN model (see next, Models) was trained using the large UK-Biobank dataset \\((N = 14,503,44 - 80\\) years, mean age \\(= 52.7\\) years) [39]. We seek to transfer this pretrained model to decode age in a small Indian cohort – the TATA Longitudinal Study of Aging (TLSA) \\((N = 188,46 - 88\\) years, mean age \\(= 66\\) years) [45] (Fig. 2C), without access to the UKBiobank data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.84 + ], + "angle": 0, + "content": "Preprocessing. We followed a preprocessing pipeline identical to [41]. The T1w-MRI scan was brain-extracted, bias-corrected [46], and registered to standard MNI-152 space [38], and cropped to get preprocessed scans \\(\\mathbf{x} \\in \\mathbb{R}^{160 \\times 192 \\times 160}\\), which was used to predict brain age." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Models. Features from the pretrained Simple Fully Convolutional Network (SFCN, Fig. 2D) were used to make the predictions with a regression model; more details are provided in Appendix B.2) [41]. The entire model was then" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.76, + 0.106 + ], + "angle": 0, + "content": "adapted to the target dataset (TLSA)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.137 + ], + "angle": 0, + "content": "We also apply CRAFT to people counting and tumor prediction tasks (see Appendix B.3, B.4 for dataset details)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.762, + 0.164 + ], + "angle": 0, + "content": "4.2. Competing DA Approaches" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.171, + 0.905, + 0.215 + ], + "angle": 0, + "content": "Baseline model. We define a naive baseline model that always predicts the mean label of the training set, ignoring the input features of the test samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.905, + 0.261 + ], + "angle": 0, + "content": "Transfer Learning (TL). We naively adapt the source model by minimizing the supervised loss on the labeled samples in the target dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.262, + 0.905, + 0.382 + ], + "angle": 0, + "content": "TASFAR. This is an SF-UDA model (See Section 2) for regression problems. It depends on two statistics from the source dataset – the threshold for confident samples and the mapping between uncertainty to prediction errors. Because the source is unavailable, we compute these using the TL model on the target. For a fair comparison, we add the supervised loss on the labeled data to the unsupervised loss, weighted by the same \\(\\alpha\\) as in our model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.905, + 0.503 + ], + "angle": 0, + "content": "DataFree. This is another SF-UDA model (See Section 2). The original model depends on the distribution of intermediate latents on the source distribution. Again, because the source data is unavailable, we computed latent distribution using the TL model and the labeled target data. As before, the objective was modified to accommodate a supervised loss (same \\(\\alpha\\) as in our model). Bin sizes for both TASFAR and DataFree were selected to be the same as in CRAFT." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.905, + 0.58 + ], + "angle": 0, + "content": "Progressive Mixup. This is a semi-supervised, source-free domain adaptation tool (See Section 2). We adapt the model to the regression task by replacing cross-entropy losses with mean-squared error loss. Predictive entropy was replaced by variance to estimate uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.905, + 0.655 + ], + "angle": 0, + "content": "BBCN. Bilateral Cycle Consistency is a prototype-based clustering SF-SSDA paradigm (See Section 2). It depends on class-level prototypes for each model. To extend it to regression, we discretize the output space into bins of the same width as the other methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.905, + 0.762 + ], + "angle": 0, + "content": "Metrics. We quantify the prediction performance of continuous valued outputs using two metrics - Root Mean Squared Error (RMSE) and Percentage-Bend Correlation [49] coefficient (R). A lower RMSE reflects a closer correspondence between the predicted and actual labels, a higher R-value quantifies better predictions along the best-fit line between the actual and predicted labels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In subsequent sections (5.1 and 5.2), we predict saccade amplitude from EEG and brain age from structural MRI scans, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.823, + 0.719, + 0.839 + ], + "angle": 0, + "content": "5. Experiments: Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.783, + 0.864 + ], + "angle": 0, + "content": "5.1. Saccade Amplitude Prediction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Firstly, we trained an EEGNet-LSTM model on the Visual Search source dataset (RMSE=67.03, R=0.91). This novel" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "832" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "architecture outperformed the benchmark model (Pyramidal CNNs) by \\(\\sim 16\\%\\) (benchmark: RMSE=79.59 pix, R=0.87). Next, this was transferred to the Large Grid dataset. The upper 3 rows of Table 1 shows the performance when all the data was used for funetuning. In this case also EEGNet-LSTM outperformed the benchmark (Pyramidal CNN [29]) by a large margin (\\(\\sim 19\\%\\)). Hence, subsequent SSDA experiments were performed using EEGNet-LSTMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.304 + ], + "angle": 0, + "content": "All models were trained with mini-batches of 128 samples, and only the best-performing checkpoint on the held-out validation set was used for evaluating the test set. Adam optimizer with an initial learning rate of \\(10^{-4}\\) was used. \\(\\alpha\\) in equation 3 was selected with a grid saerch over \\(\\{0.01, 0.1, 1.0\\}\\); empirically, \\(\\alpha = 0.1\\) was always favored." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.315, + 0.482, + 0.398 + ], + "angle": 0, + "content": "Table 1. Source Free Semi-supervised Deep Transfer of the model trained on the Visual Search task to the Large-Grid Task. The upper part of the table (rows 1-3) shows the performance ceiling when all labeled data in the target is used for training. The lower part (rows 4-9) shows SF-SSDA results when only \\(1\\%\\) of the target samples were labeled (3 seeds)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.409, + 0.485, + 0.572 + ], + "angle": 0, + "content": "
MethodR ↑RMSE (in pix) ↓
Naive Baseline-149.12 ± 0.02
Benchmark (TL, 100%)0.91 ± 0.0163.84 ± 0.75
EEGNet-LSTM (TL, 100%)0.93 ± 0.0151.47 ± 0.63
SF-SSDA (1% labels)
EEGNet-LSTM + TL0.77 ± 0.0192.26 ± 1.66
EEGNet-LSTM + Mixup0.48 ± 0.02135.70 ± 1.25
EEGNet-LSTM + BBCN0.76 ± 0.0199.8 ± 3.35
EEGNet-LSTM + TASFAR0.76 ± 0.0186.41 ± 1.05
EEGNet-LSTM + DataFree0.80 ± 0.0187.64 ± 3.08
EEGNet-LSTM + CRAFT0.81 ± 0.0284.17 ± 3.95
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.585, + 0.483, + 0.842 + ], + "angle": 0, + "content": "To demonstrate the efficacy of CRAFT for SF-SSDA, target labels from the training data of the Large Grid dataset were randomly dropped in a stratified manner. Figure 3A shows how the performance of these models is affected by reducing the fraction of labeled data (\\(n_{ul} / N\\) is the fraction of unlabeled samples). A lower proportion of labeled data worsens the performance of all the models (RMSE increases), but CRAFT produces the lowest RMSE for all unlabeled data proportions. Figure 3B shows the prediction of saccade magnitudes (\\(|\\Delta \\overline{r}|\\)) for the best CRAFT-based EEGNet-LSTM model. Table 1 (lower 6 rows) summarizes the performance of CRAFT when \\(99\\%\\) of the data were unlabeled; CRAFT outperforms supervised training (TL) by \\(9\\%\\) and other SOTA SF-SSDA models by \\(>4\\%\\) in terms of RMSE. The closest competitive methods were TASFAR and DataFree. The correlation coefficient (R) also showed similar trends (Table 1; Fig. 6A)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.291, + 0.866 + ], + "angle": 0, + "content": "5.2. Brain Age Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Next, we transferred the SFCN model trained on the UK Biobank MRI scans to the TLSA dataset. The TLSA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "dataset has only 188 MRI scans, so we avoid hyperparameter searches. All models were trained for 25 epochs (batch size = 4) using the Adam optimizer (initial learning rate = 10\\(^{-4}\\)), and only the final model was used for inference. Here, we used \\(\\alpha = 0.1\\); we explore the effect of varying \\(\\alpha\\) for ours and competing models in Appendix D. Each metric was calculated for the prediction on the entire dataset, leaving out one fold at a time (4-folds). The top of Table 2 (row 2) quantifies the performance ceiling when all the training data are used (3 seeds)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.254, + 0.905, + 0.311 + ], + "angle": 0, + "content": "Table 2. Semi-supervised transfer of the SFCN model (source: UKB) to the TLSA dataset. Same as in Table 1, except that the lower part of the table (3-8) shows SF-SSDA results when only \\(20\\%\\) of the target samples were labeled. (3 seeds)" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.321, + 0.905, + 0.479 + ], + "angle": 0, + "content": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-7.91 ± 0.05
SFCN + TL (100%)0.66 ± 0.016.14 ± 0.03
SF-SSDA (20% labels)
SFCN + TL0.41 ± 0.077.41 ± 0.21
SFCN + Mixup0.34 ± 0.047.71 ± 0.14
SFCN + BBCN0.28 ± 0.048.00 ± 0.15
SFCN + TASFAR0.42 ± 0.077.47 ± 0.15
SFCN + DataFree0.50 ± 0.037.36 ± 0.14
SFCN + CRAFT0.51 ± 0.037.14 ± 0.11
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.689 + ], + "angle": 0, + "content": "For SF-SSDA, we performed experiments closely similar to the saccade task (Section 5.1). Figure 3C shows how these models' performance is affected when the proportion of unlabeled samples increases. Reducing the fraction of labeled data worsens model performance, but CRAFT continues to remain the most competitive of all models. Figure 3D shows the prediction scatter of the best CRAFT-based model. Table 2 summarizes the performance when \\(80\\%\\) of the data was unlabeled. CRAFT outperforms supervised training (TL) by \\(\\sim 4\\%\\) and other SOTA SF-SSDA models by \\(>3\\%\\) in terms of RMSE. Similar results were obtained with the R values also (Fig. 6B). Again, TASFAR and DataFree emerge as the most competitive methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Finally, we evaluated the effectiveness of CRAFT at two other real-world regression benchmarks: People Counting [44, 48] and Tumor Size prediction [6, 16]. CRAFT outperformed other SF-SSDA methods by \\(>5\\%\\) and \\(>2\\%\\) respectively for each dataset; the details are presented in Appendices C.1 and C.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Mitigating Sampling Biases. Next, we show how the unsupervised CRAFT objective could mitigate sampling bias effects in the target training set. In all of the previous simulations, the label distributions of the target train and test data were matched. Here, we deliberately distorted the label distribution of the target training data: \\(80\\%\\) of the data above the mean age in the training set were removed, creating a heavy bias toward lower ages. Following this, labels were" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.478, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.487, + 0.484 + ], + "angle": 0, + "content": "Figure 3. (A, C) RMSE \\((\\downarrow)\\) for source-free semi-supervised domain adaptation of saccade (A) and brain age (C) prediction tasks, trained with varying proportions of unlabeled target data \\(n_{ul} / N\\). Dashed black line: Performance ceiling. Circles: individual trials. (B, D) Prediction for the best CRAFT model for saccade amplitude (B) and brain age (D), with \\(90\\%\\) and \\(40\\%\\) unlabeled data, respectively. Triangles: individual participants. Results with a much higher proportion of unlabeled data are shown for the saccade dataset because of the far greater size of this (\\(N \\sim 12k\\)) compared to the brain age (\\(N \\sim 180\\)) dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.645 + ], + "angle": 0, + "content": "retained in only \\(40\\%\\) of the training data \\((n_{ul} / N = 0.6)\\). No such biases were introduced into the test data. The CRAFT model was then trained by imposing the unbiased (true) marginal label distribution on the unlabeled training set (equation 5). In this scenario, again, we find that CRAFT mitigates the effect of sampling biases in the training data (Table 3). It outperforms transfer learning by \\(\\sim 5\\%\\), and other sota SF-SSDA models by \\(>2.5\\%\\) (RMSE)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Table 3. Analyses with datasets with sampling error. \\(80\\%\\) of the data above the mean age is removed. Furthermore, only \\(40\\%\\) of the labels in this mislabeled data are used for semi-supervised transfer of the SFCN model (source: UKB). (3 seeds)" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.737, + 0.485, + 0.878 + ], + "angle": 0, + "content": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-8.38 ± 0.22
SFCN + TL0.41 ± 0.058.40 ± 0.27
SFCN + Mixup0.32 ± 0.028.19 ± 0.15
SFCN + BBCN0.23 ± 0.028.53 ± 0.16
SFCN + TASFAR0.39 ± 0.048.57 ± 0.22
SFCN + DataFree0.41 ± 0.038.58 ± 0.23
SFCN + CRAFT0.45 ± 0.037.98 ± 0.25
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.839, + 0.108 + ], + "angle": 0, + "content": "5.3. Analysis of computational complexity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Table 4 shows the computational complexity of CRAFT and the competing models. We compute both the algorithms' theoretical complexity and running time on the saccade and brain age prediction tasks. The theoretical training complexity is defined as the number of gradient updates per epoch. For inference, the complexity is the same for all models as the same architecture is trained using different adaptation paradigms. Training time and complexity are lowest for TASFAR, amongst all other SF-DA tools. The complexity of CRAFT was comparable to DataFree, the next best method." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Table 4. Computational complexity. Time per training epoch and for all test data (in min). Parentheses: P- no. of model parameters; N- samples; D- latent dimensionality; \\(\\beta\\)- batch size, B- no. of bins, m- no. of inferences for MC Dropout uncertainty estimation." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.356, + 0.905, + 0.453 + ], + "angle": 0, + "content": "
(B) MethodEEGNet-LSTM (P~ 190K)SFCN (P~ 2.95M)
Train /Comp. ↓ (N ~ 12k)Test /Comp. ↓ (N ~ 2.5k)Train/Comp. ↓ (N ~ 150)Test/Comp. ↓ (N ~ 50)
TL0.15 / O(N)0.02 / O(N)0.12 / O(N)0.04 / O(N)
Mixup0.74 / O(Nm)0.02 / O(N)0.74 / O(Nm)0.04 / O(N)
BBCN2.71 / O(NBD)0.02 / O(N)1.29 / O(NBD)0.04 / O(N)
TASFAR0.30 / O(N)0.02 / O(N)0.19 / O(N)0.04 / O(N)
DataFree0.55 / O(ND2)0.02 / O(N)0.32 / O(ND2)0.04 / O(N)
CRAFT0.45 / O(N(β + B))0.02 / O(N)0.36 / O(N(β + B))0.04 / O(N)
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.719, + 0.5 + ], + "angle": 0, + "content": "6. Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Data scarcity poses a serious challenge for applying deep-learning models in real-world settings. Training overparameterized models on a few hundred data points leads to overfitting and poor generalization. Conventional domain adaptation algorithms, to transfer pretrained networks, typically need access to source domain data, but these may not always be available. In this study, we propose a source-free semi-supervised domain adaptation framework that recalibrates the decision boundary without aligning the intermediate representation of the domains. In fields like medicine, resource constraints associated with labeling data are significant; therefore, leveraging the large amounts of unlabeled data becomes critical. We propose CRAFT as an SF-SSDA approach for such source-free, label-sparse settings that are common in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.748, + 0.702, + 0.765 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.909, + 0.879 + ], + "angle": 0, + "content": "This work was supported by the following funding sources: a Prime Minister's Research Fellowship (to M.B.), a DST SwarnaJayanti fellowship, a Pratiksha Trust Intramural grant, a Tata Trusts grant and a Google Research grant (to D.S.). We also thank the Tata Longitudinal Study on Aging at the Indian Institute of Science for access to MRI scans from their cohort." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jiayu An, Changming Zhao, and Dongrui Wu. Semi-supervised generalized source-free domain adaptation (ssgsfda). In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.173, + 0.465, + 0.187 + ], + "angle": 0, + "content": "[2] APTOS. Aptos 2019 blindness detection dataset, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.189, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[3] Sourabh Balgi and Ambedkar Dukkipati. CUDA: Contradistinguisher for unsupervised domain adaptation. In Proceedings of the IEEE International Conference on Data Mining, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[4] Sourabh Balgi and Ambedkar Dukkipati. Contradistinginguisher: A Vapnik's imperative to unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4730-4747, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.483, + 0.358 + ], + "angle": 0, + "content": "[5] Scott L. Brincat, Jacob A. Donoghue, Meredith K. Kahnke, Simon Kornblith, Mikael Lundqvist, and Earl K. Miller. Interhemispheric transfer of working memories. Neuron, 109: 1055-1066.e4, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.569 + ], + "angle": 0, + "content": "[6] Peter Bandy, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, Quanzheng Li, Farhad Ghazvinian Zanjani, Svitlana Zinger, Keisuke Fukuta, Daisuke Komura, Vlado Ovtcharov, Shenghua Cheng, Shaoqun Zeng, Jeppe Thagaard, Anders B. Dahl, Huangjing Lin, Hao Chen, Ludwig Jacobsson, Martin Hedlund, Melih Çetin, Eren Halici, Hunter Jackson, Richard Chen, Fabian Both, Jörg Franke, Heidi Küsters-Vandevelde, Willem Vreuls, Peter Bult, Bram van Ginneken, Jeroen van der Laak, and Geert Litjens. From detection of individual metastases to classification of lymph node status at the patient level: The camelyon17 challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2019. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[7] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140: 325-331, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.67 + ], + "angle": 0, + "content": "[8] Thomas A. Carlson. Orientation decoding in human visual cortex: New insights from an unbiased perspective. Journal of Neuroscience, 34(24):8373-8383, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[9] J H Cole, S J Ritchie, M E Bastin, M C Valdés Hernández, S Muñoz Maniega, N Royle, J Corley, A Pattie, S E Harris, Q Zhang, N R Wray, P Redmond, R E Marioni, J M Starr, S R Cox, J M Wardlaw, D J Sharp, and I J Deary. Brain age predicts mortality. Molecular Psychiatry, 23:1385-1392, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Sébastien M. Crouzet. Fast saccades toward faces: Face detection in just 100 ms. Journal of Vision, 10:1-17, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[11] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. PMID: 20144289. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[13] Heiner Deubel and Werner X. Schneider. Saccade target selection and object recognition: Evidence for a common attentional mechanism. Vision Research, 36:1827-1837, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Yuhe Ding, Lijun Sheng, Jian Liang, Aihua Zheng, and Ran He. Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks, 167:92-103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[15] Cian Eastwood, Ian Mason, Christopher K I Williams, and Bernhard Scholkopf. In Proceedings of the Tenth International Conference on Learning Representations, 2022. Tenth International Conference on Learning Representations 2022, ICLR 2022; Conference date: 25-04-2022 Through 29-04-2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[16] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen A. W. M. van der Laak, , and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 318(22):2199-2210, 2017. 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Saminger-Platz et al. Central moment discrepancy (cmd) for domain-invariant representation learning. *ICLR*, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[18] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 174:106230, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[19] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17:1-35, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[20] Alessandro T. Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M. Cichy. The algonauts project 2025 challenge: How the human brain makes sense of multimodal movies, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[21] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems. MIT Press, 2004. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[22] Arthur Gretton, Kenji Fukumizu, Zäïd Harchaoui, and Bharath K. Sriperumbudur. A fast, consistent kernel two-sample test. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Yan Hao, Yuhong Guo, and Chunsheng Yang. Source-free unsupervised domain adaptation with surrogate data generation. In British Machine Vision Conference, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Computer Vision - ECCV 2016, pages 630-645, Cham, 2016. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and Shueng-Han Gary Chan. Target-agnostic source-free domain adaptation for regression tasks. 2024 IEEE 40th In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ternational Conference on Data Engineering (ICDE), pages 1464-1477, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[27] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[28] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[29] Ard Kastrati, Martyna Plomecka, Damian Pascual Ortiz, Lukas Wolf, Victor Gillioz, Roger Wattenhofer, and Nicolas Langer. Eegeyenet: a simultaneous electroencephalography and eye-tracking dataset and benchmark for eye movement prediction. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[30] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[31] Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. Eeg-net: a compact convolutional neural network for eeg-based brain-computer interfaces. Journal of Neural Engineering, 15:056013, 2018. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[32] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A Comprehensive Survey on Source-Free Domain Adaptation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 46(08):5743-5762, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[33] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the 32nd International Conference on Machine Learning, pages 97-105, Lille, France, 2015. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[35] Jie Ma, Zhijun Xu, Xinyu Xiong, MaYiDiLi NiJiaTi, Dongyu Zhang, Weijun Sun, Feng Gao, and Guanbin Li. Source-free semi-supervised domain adaptation for tuberculosis recognition. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[36] Ning Ma, Jiajun Bu, Lixian Lu, Jun Wen, Zhen Zhang, Sheng Zhou, and Xifeng Yan. Semi-supervised hypothesis transfer for source-free domain adaptation, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[37] Ning Ma, Haishuai Wang, Zhen Zhang, Sheng Zhou, Hongyang Chen, and Jiajun Bu. Source-free semi-supervised domain adaptation via progressive mixup. Knowledge-Based Systems, 262:110208, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[38] Pravat K. Mandal, Rashima Mahajan, and Ivo D. Dinov. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. Journal of Alzheimer's Disease, 31:S169-S188, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Karla L Miller, Fidel Alfaro-Almagro, Neal K Bangerter, David L Thomas, Essa Yacoub, Junqian Xu, Andreas J Bartsch, Saad Jbabdi, Stamatios N Sotiropoulos, Jesper L R Andersson, Ludovica Griffanti, Gwenaëlle Douaud," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "Thomas W Okell, Peter Weale, Iulius Dragonu, Steve Garratt, Sarah Hudson, Rory Collins, Mark Jenkinson, Paul M Matthews, and Stephen M Smith. Multimodal population brain imaging in the uk biobank prospective epidemiological study. Nature Neuroscience, 19:1523-1536, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Han Peng, Weikang Gong, Christian F Beckmann, Andrea Vedaldi, and Stephen M Smith. Accurate brain age prediction with lightweight deep neural networks. Medical Image Analysis, 68:101871, 2021. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[42] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In Computer Vision - ECCV 2022, pages 165-182, Cham, 2022. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[43] Xiaofeng Qu, Li Liu, Lei Zhu, Liqiang Nie, and Huaxiang Zhang. Source-free style-diversity adversarial domain adaptation with privacy-preservation for person re-identification. Knowledge-Based Systems, 283:111150, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[44] Vishwanath Sindagi, Rajeev Yasarla, and Vishal Patel. Pushing the Frontiers of Unconstrained Crowd Counting: New Dataset and Benchmark Method. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1221-1231, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[45] Jonas S Sundarakumar, Albert Stezin, Abhishek L Menes-gere, and Vijayalakshmi Ravindranath. Rural-urban and gender differences in metabolic syndrome in the aging population from southern india: Two parallel, prospective cohort studies. eClinicalMedicine, 47:101395, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[46] J-Donald Tournier, Robert Smith, David Raffelt, Rami Tabbara, Thijs Dhollander, Maximilian Pietsch, Daan Christi-aens, Ben Jeurissen, Chun-Hung Yeh, and Alan Connelly. Mrtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202:116137, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[47] Ihsan Ullah and Alfredo Petrosino. About pyramid structure in convolutional neural networks, 2016. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[48] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[49] Rand R. Wilcox. The percentage bend correlation coefficient. Psychometrika, 59:601-616, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[50] Garrett Wilson and Diane J. Cook. A survey of unsupervised deep domain adaptation. ACM Trans. Intell. Syst. Technol., 11(5), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[51] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36(8):3746-3766, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.843, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[52] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "836" + } + ] +] \ No newline at end of file diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_origin.pdf b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50c6994b9569a5b4ace7ae0f1911330387e31d0e --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/60f16256-f509-4463-b38b-322283424ecd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f01c34d3f7c8f1e218e256b12d04d5a7f44f204d7642914a89cfebb40ed9cb3f +size 3292255 diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/full.md b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08be1915980f1f19c6485c7b147c45cf53f7dfef --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/full.md @@ -0,0 +1,323 @@ +# Semi-supervised Deep Transfer for Regression without Domain Alignment + +Mainak Biswas1,2, Ambedkar Dukkipati3, Devarajan Sridharan1,2,3 + +$^{1}$ Brain, Computation and Data Sciences, $^{2}$ Centre for Neuroscience, $^{3}$ Computer Science and Automation Indian Institute of Science, Bangalore - 560012 + +mainakbiswas@iisc.ac.in, ambedkar@iisc.ac.in, sridhar@iisc.ac.in + +# Abstract + +Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data or reliably labeled target data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT - a Contradistinguisher-based Regularization Approach for Flexible Training - for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and "brain age" prediction with structural MRI data. For both datasets, CRAFT yielded up to $9\%$ improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than $3\%$ . Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine. + +Keywords: source-free DA, continuous label prediction, gaze prediction, brain age prediction, healthcare + +# 1. Introduction + +For the successful application of deep learning models in the real world, they must be robust to "domain shift" [50]. For example, in biology and medicine, supervised deep learning models are often trained on large, high-quality source datasets collected in resource-rich settings [11, 39]. However, naive transfer or even simple finetuning fails to produce high accuracies when tested with smaller target datasets of mixed quality collected [2, 45]. As a result, state-of-the-art domain adaptation (DA) approaches, which enable the source models to generalize successfully to target datasets, must be employed. Many of these approaches involve explicit alignment of source and target, while learning a shared predictive model for both domains [17, 19, 40]. + +A particular challenge occurs with DA when the source dataset is unavailable. This can happen either because of privacy or proprietary concerns, such as with medical images or patents [43]. Alternatively, source datasets may be too large, incurring prohibitive storage and computation costs in low-resource settings [39]. Source-free (SF) domain adaptation methods that address these challenges are getting increasingly popular [18, 32]. Such methods generalize the source model to unlabeled or partially target data (unsupervised or semi-supervised DA) with one of two approaches: data-based or model-based. While data-based methods create surrogate source data, model-based methods adapt the source model to the target domain (see Section 2). + +We address model-based SF-DA in a semi-supervised setting. Specifically, we focus on neuroscience applications where source data are inaccessible, and target labels are reliably available only for a few samples. Regression tasks – like predicting brain age (Fig. 1C) [39], saccades [29], neural activity [20], and stimulus orientation [8] – are ubiquitous in neuroscience; we tackle a subset of these. We build upon Contradistinguisher [3, 4], a recently-proposed tool for effective unsupervised DA (CUDA). Unlike popular representation alignment DA methods [19, 40], CUDA directly learns a common model for both domains. Yet, CUDA requires full access to source data and addresses classification tasks. Therefore, it is not directly applicable to our problem. + +![](images/eb9853653c6b9022712a0579a3ff2529cc99fff34dab54d3756ff8484db4e5c0.jpg) +A +X + +![](images/50abad7ebb0e934b69ecdb035b0c1da341db9610f3320d5e761bc874f72f3d96.jpg) + +![](images/7ff5d8c03c3f4ad2d8c212ee190fd4cb9ac2db57efa1226e6b4c1f6914156dc1.jpg) + +![](images/90f5303b5b158277667aad3571bc0e9dbed2b57cdc3db5b3e10ecf16eeeab682.jpg) +B +Figure 1. Schematic of Semisupervised Deep Transfer without Domain Alignment. (A) The approach seeks to transfer a pretrained model to a target domain with limited labeled data when the source dataset is unavailable. (B) Because the source dataset is unavailable, we tune the decision boundaries without explicitly aligning the representation of the domains, combining a supervised loss and the CUDA loss (regularizer). (C) The algorithm targets deep transfer in regression tasks, like brain age prediction. + +![](images/3c6cd4215a790df2af29f683479d512f61696c30867f90231da3da7ce48aa7b8.jpg) +C + +Here, we leverage CUDA to develop a new algorithm - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT) which performs deep transfer, without domain alignment, even when the source dataset is unavailable (Fig. 1A-B). We also extend CUDA to a regression setting. Previous studies approach regression tasks naively by binning continuous labels into discrete bins [40], sometimes employing a rank-ordinal objective [7]. Regardless, the efficacy of such approaches may depend critically on bin sizes. Our CRAFT model uses a principled approach to extend CUDA to predictions involving continuous-valued outputs. We employ binning as a practical strategy to generate and optimize pseudo-labels, without constraining the model to produce discrete outputs. Our study makes the following main contributions. + +- First, we extend CUDA to source-free DA for regression tasks involving continuous-valued outputs. +- Second, we formulate a theoretically-motivated semi-supervised objective by combining supervised and unsupervised losses with a tunable hyperparameter $(\alpha)$ (Section 3). We also derive the unsupervised loss as a maximum entropy prior to model parameters (Appendix A.2). +- Third, we show the effectiveness of CRAFT on two neuroscience challenges - gaze prediction from electroencephalographic (EEG) [29] data and brain age prediction from magnetic resonance imaging (MRI) data [45]. +- Fourth, we also apply CRAFT to two real-world regression benchmarks involving people counting [44, 48] and tumor size prediction [6, 16]. + +- Fifth, we compare CRAFT against state-of-the-art SF-DA methods [15, 26, 35, 37] and show that the best improvements occur as unlabeled target samples increase, demonstrating its efficacy for semi-supervised SF-DA. +- Finally, we analyze the computational complexity and demonstrate numbers competitive with sota models. +- Our code is publicly available (see Appendix E.2). + +# 2. Related Work + +Domain alignment methods. Popular Unsupervised Domain Adaptation (UDA) approaches rely on learning common intermediate representations across source and target domains. Domain Adversarial training of Neural Networks (DANN) [19] uses a discriminator to learn common representations across both domains adversarially. In Importance Weighting [40] (IW), higher importance is afforded to losses corresponding to source samples more likely to belong to the target domain. Central Moment Discrepancy (CMD) [17] minimizes the central moments of the latent representations between the domains, while others optimize their Maximum Mean Discrepancy (MMD) [22, 34]. Despite being effective, they cannot be used when the source is unavailable and will not be discussed further. + +Source-free UDA. To address DA scenarios where source data and its target labels are unavailable, several source-free unsupervised DA approaches have been recently proposed. Broadly, they are categorized as: Data-based and Model-based [32]. Data-based methods generate a source-like distribution [14, 30]. However, they often depend on the availability of generative models trained on the source, which could be unavailable. For example, SF-UDA through domain-specific perturbation (AUGFree) [51] assumes the existence of domain-invariant features for prediction. Since the source is unavailable, AUGFree perturbs the target data and then learns a transformation to the invariant features by adversarially aligning the target and the perturbed target. + +On the other hand, model-based methods finetune the source model on the target using strategies like contrastive learning [27], pseudo-labeling [26], or entropy minimization [36]. They are often prone to overfitting to pseudolabels. For example, Class-Balanced Multicentric Dynamic Prototype Strategy for SF-UDA (BMD) [42] follows a pseudo-labeling strategy using class prototypes based on the source model predictions, and in turn uses these labels for training. Source Data-Absent UDA Through Hypothesis Transfer (SHOT++) [33] combines three objectives - a) maximizing mutual information between the predictions and inputs, b) self-supervised loss based on prototype-based clusters, and c) a Mixup-based (See SF-SSDA) loss [37]. Attracting and Dispersing (AaD) [52] defines a neighboring cluster (K nearest neighbors) and a randomly selected background cluster for each sample, and maximizes the log-likelihood of the ratio of their distributions. + +Target-agnostic SF UDA for Regression Tasks (TAS-FAR) [26] works on the principle that prediction uncertainty is correlated with prediction error. It uses Monte-Carlo dropout to estimate label uncertainty and divides the target dataset into confident and uncertain samples. A discrete density map is estimated using the confident data, which is then used as weights to recalibrate the pseudo-labels of uncertain data. SF-UDA to measurement Shift via Bottom-Up Feature Restoration (BUFR/DataFree) [15] is designed on the principle of intermediate feature alignment of the source and the target distributions. During adaptation, the model minimizes the symmetric KL divergence between feature distributions of source (fixed) and target data. + +Here, we compare our model's performance against two of the above methods - TASFAR and DataFree. TASFAR was selected as it is a pseudo-labeling-based method suitable for regression. On the other hand, DataFree is a pseudo-label-free feature alignment method, not constrained to classification tasks. By adding a supervised loss (Section 4.2), we evaluate these as semi-supervised models. + +Source-free SSDA. Source-free semi-supervised domain adaptation (SF-SSDA) models best address the scenario of SF-DA when few target labels are available. For example, Generalized SF-SSDA [1] minimizes source loss alongside an unsupervised target loss which maximizes the cosine similarity between predictions of k-nearest neighbors. The model is then transferred by adversarial training based on source features generated using a Conditional VAE. Similarly, [23] proposes generating surrogate source data from target samples using gradient-based optimization. Learning from Different Samples for SF-SSDA [28] minimizes the contrastive loss on the predicted probabilities of the target domain. Furthermore, it minimizes supervised and contrastive loss on a set comprising labeled and unlabeled data with low uncertainty. Likewise, Mutual Enhancement training for SS Hypothesis transfer (MESH) [36] minimizes the supervised loss on the labeled samples and entropy on pseudo-labeled samples. Finally, the target manifold is smoothened by making predictions robust to perturbations. + +Progressive Mixup [37] is trained on augmentations by convex combinations of data on which the source model is confident and ones on which it is uncertain (hybrid mixup). Similarly, another set is obtained by convex combinations of samples in the (pseudo) labeled dataset (self-mixup). In contrast, Bilateral-Branch Consistency Network (BBCN) [35] computes prototypes as the weighted mean of feature representations of the target dataset by the source model. Pseudo-labels are selected based on the proximity of unlabeled data to these prototypes. A separate target branch is trained using the pseudo-labels, and a consistency loss is added between the representations of both branches. The need for class-level prototype generation makes it challenging to adapt this model to regression tasks. + +We compare the performance of CRAFT with two SFSSDA methods - SF-SSDA via progressive Mixup [37], and BBCN [35], as they have little dependence on the source distribution and its derivatives. Moreover, these methods are based on complementary ideas, and the latter [35] has been applied specifically in medical imaging, for tuberculosis recognition. + +# 3. The CRAFT Model + +# 3.1. Model formulation + +Background. Our objective is to develop an approach for source-free semi-supervised domain adaptation, for regression tasks. For this, we build upon the Contradistinguisher-based Unsupervised Domain Adaptation (CUDA) model [4]. CUDA obviates the need for learning a common intermediate representation across the source and the target datasets. It optimizes a joint distribution over the target features and category labels to adapt the source model. Because CUDA is intended for unsupervised domain adaptation (UDA), target labels are unavailable. Hence, the algorithm uses a two-step joint optimization over the target labels and model parameters. First, it generates pseudo-labels on the target dataset from the source model [21]. Then, it updates the model parameters by fixing the pseudo-labels. + +In this study, we extend CUDA to a source-free semi-supervised setting (SF-SSDA). Briefly, to the supervised label loss, we add a 'regularized', unsupervised loss as a prior on the model parameters. As we will show in the next section, the model can be flexibly applied, either in SF-UDA case or in the SF-SSDA case, depending on the weightage given to the supervised loss. We, moreover, extend the model to continuous label prediction. We call our new formulation CRAFT - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT). + +Objective. We consider a task in which a source model $\theta^s$ trained on a source dataset - involving continuous label prediction (regression) - is already available to us. The target dataset comprises i.i.d samples from an unknown distribution $p^t (\mathbf{x},y)$ : $\mathcal{D}^t = \{\mathbf{x}_i^t,y_i^t\}_{i = 1}^{N_l}$ , $\{\mathbf{x}_i^t\}_{i = 1}^{N_{ul}}$ where $\mathbf{x}_i^t\in \mathbb{R}^d$ is a feature-vector, and $y_{i}^{t}\in \mathbb{R}$ is its corresponding label. Target training data are partially and sparingly labeled; we define the proportion of unlabeled samples $(N_{ul})$ to the total number of samples $(N)$ , as $N_{ul}:(N_l + N_{ul})$ . + +We formulate, first, the supervised objective. For continuous label prediction problems, the conditional density of the label given the features is assumed to be normally distributed. The mean is modeled using a neural network $f(:,\theta)$ , parameterized by $\theta$ , as: $p(y^{t}|\mathbf{x}^{t},\theta) = \mathcal{N}(y^{t};f(\mathbf{x}^{t};\theta),c)$ , and the variance, $c$ is assumed to be constant as it simplifies the optimization. In the final formulation, the variance occurs as the weight of the unsupervised loss, $\alpha$ (equation 3); therefore, only one of the two + +hyperparameters ( $c$ or $\alpha$ ) needs to be tuned. Hence, the log-likelihood can be written as: + +$$ +\log p \left(\mathcal {D} ^ {t} | \theta\right) \propto - \sum_ {i = 1} ^ {N _ {l}} \left(y _ {i} ^ {t} - f \left(\mathbf {x} _ {i} ^ {t}; \theta\right)\right) ^ {2} \tag {1} +$$ + +The unsupervised objective is essentially identical to the CUDA objective and is defined as a joint log density over the labels and model parameters as follows: + +$$ +\log q \left(\mathbf {x} ^ {t}, y ^ {t} \mid \theta\right) = \log \left[ \frac {p \left(y ^ {t} \mid \mathbf {x} ^ {t} , \theta\right)}{\sum_ {i = 1} ^ {N} p \left(y ^ {t} \mid \mathbf {x} _ {i} ^ {t} , \theta\right)} p \left(y ^ {t}\right) \right] \tag {2} +$$ + +This joint distribution seeks to reduce the model's predictive bias when faced with a distribution shift in the target domain (see Fig. 1B). The denominator, which reflects the total probability of predicting a particular class, normalizes for biases originating from the source data. In other words, model predictions are normalized as $p(y^{t}|\mathbf{x}^{t},\theta) / \sum_{i = 1}^{N}p(y^{t}|\mathbf{x}_{i}^{t},\theta)$ , such that no class has a biased probability of being predicted. Finally, the model predictions align with the target label distribution, using the supplied marginal $p(y^t)$ . + +Note, that this objective does not depend on whether the samples are labeled or not. Optimizing the joint distribution: 1) normalizes inherent bias towards one label over the other by calibrating the prediction for a label with its total weight in the dataset, 2) and explicitly trains $q$ to match the marginals - i.e., $\sum_{i=1}^{N} q(\mathbf{x}_i^t, y^t|\theta) = p(y^t)$ , thereby matching high-level statistics like per-class proportions in the training distribution. Importantly, with an informative prior $p(y^t)$ that specifies the putative distribution of the target labels, pseudo-label prediction can be biased toward labels that are more likely (see Section 3.2). + +Combining the supervised and the unsupervised objective, the semi-supervised CRAFT objective is defined as: + +$$ +\mathcal {L} \left(\mathcal {D} ^ {t}, \theta\right) = \sum_ {i = 1} ^ {N _ {l}} \log p \left(y _ {i} ^ {t} \mid \mathbf {x} _ {i} ^ {t}, \theta\right) + \alpha \sum_ {i = 1} ^ {N} \log q \left(\mathbf {x} _ {i} ^ {t}, y _ {i} ^ {t} \mid \theta\right) \tag {3} +$$ + +In the above equation, $\alpha$ is the weight associated with the unsupervised objective and is tuned as a hyperparameter on a held-out cross-validation set during training. For deep-transfer, model parameters are initialized to $\theta^s$ derived from the source model, and then gradient descent is performed to obtain $\theta^t$ . + +Theoretical motivation. Artificial Neural Networks (ANNs) with hypotheses space $\Theta$ are trained to learn a parametric approximation $p(\mathbf{x},y|\theta)$ , $\theta \in \Theta$ , such that the log-likelihood of the training data given the parameters is maximized. However, training ANNs often does not account for prior distributions over model parameters. In Appendix A.1, we demonstrate that the CRAFT semi-supervised objective can be theoretically derived as MAP + +estimation of the parameters $(\log p(\theta|\mathcal{D}))$ , which we write as a sum of the supervised objective, $\log p(\mathcal{D}|\theta)$ , and a prior on the model parameters, $p(\theta)$ . Specifically, we demonstrate, using the principle of entropy maximization [21], that the unsupervised objective is related to the prior on the model parameters. + +# 3.2. Model optimization + +We optimize the supervised loss of the CRAFT objective with conventional gradient descent, jointly optimizing both components in the loss (equation 3). The unsupervised component of the loss is dependent on the labels $y^{t}$ , which are unavailable for the vast majority of target data. To tackle this, we employ a 2-step optimization process. + +Pseudo-label selection. As the first step, we fix the model parameters $\theta$ and select the label that maximizes the joint distribution for each datapoint: + +$$ +\tilde {y} _ {i} ^ {t} = \underset {y ^ {t} \in \mathcal {Y}} {\operatorname {a r g m a x}} q \left(\mathbf {x} _ {i} ^ {t}, y ^ {t} \mid \theta\right) \tag {4} +$$ + +where $\tilde{y}_i$ the pseudo-label for feature vector $\mathbf{x}_i$ . For classification tasks, the optimization in equation 4 can be achieved by selecting the best label from a finite set of class labels that maximizes $q$ [4]. But, for regression problems such as ours, class labels map to real numbers spanning the range of the model outputs $f(:,\theta):\mathbb{R}^d\mapsto [a,b]$ . To address this, we rewrite the above equation under the assumption that the continuous-valued labels are distributed normally, with mean $f(\mathbf{x}^t;\theta)$ , and variance $c$ . Hence, the above equation can be written as: + +$$ +\tilde {y} _ {i} ^ {t} = \underset {y ^ {t} \in \mathcal {Y}} {\operatorname {a r g m a x}} \frac {\mathcal {N} \left(y ^ {t} ; f \left(\mathbf {x} _ {i} ^ {t} ; \theta\right) , c\right) p \left(y ^ {t}\right)}{\sum_ {l = 1} ^ {N} \mathcal {N} \left(y ^ {t} ; f \left(\mathbf {x} _ {l} ^ {t}; \theta\right) , c\right)} \tag {5} +$$ + +Equation 5 can be optimized with gradient ascent with respect to $y^{t}$ for the objective log $q(\mathbf{x}_i^t, y^t|\theta)$ . But, the pseudo-labels must be estimated before gradient descent can be performed on the model parameters (see next step, equation 6). This requires nested gradient descent, which increases computational complexity. Moreover, to employ an informative prior over target labels, we fit a class of mixture models (described in Appendix A.3) to estimate $p(y)$ from the data; this makes backpropagation more cumbersome. To overcome these challenges, we divide the entire range of labels into small, discrete bins whose midpoints are the candidate pseudo-labels. Hence, this step in the optimization can be achieved faster pseudo-label selection from a discrete set. Note that this approach does not transform the model outputs to be discrete; it merely represents a convenient optimization strategy to update the parameters. In the experimental section, we provide a recipe for choosing bin sizes and study the impact of varying bin sizes. + +Maximization. In the second step, we optimize the model parameters by using the labeled data $\{\mathbf{x}_i^t,y_i^t\}_{i = 1}^{N_l}$ and fixing the pseudo-labels $\{\mathbf{x}_i^t,\tilde{y}_i^t\}_{i = 1}^N$ , as: + +![](images/888e5b032c4d9b15c42bc7c8b1148856bc99f3a289a87f7305873188cd7787c9.jpg) +A + +![](images/0d0298fa1f99250ae7821fa1d04b391c8c05d5d0613e83037e998a8ea29c239b.jpg) +C + +![](images/c6ceebd152ad3338c057f9ad722eaaa208b690c7917774b10832407957cb3094.jpg) + +![](images/00308f5991f4dab43c77e41957efdf5bc73c95df9dd6dd21cd63bedfec5584bd.jpg) +B + +![](images/f00e1ab5aa2c1dbc9ec8469fa4c9cc977e74e7b2aaabc0ce703f425165f6f02b.jpg) +D +Figure 2. (A) Schematic of gaze prediction using EEG signals. (Left) The task schematic of the Visual Search and Large Grid datasets from the EEGEyeNet benchmark. (Right) EEG and eye tracking data were recorded [29]. (B) Architecture of the EEGNet-LSTM model, used to decode from EEG data. The parsimonious architecture can be used to decode saccades from EEG data. (C) Schematic of the "brain age" prediction task using structural MRI scans. (Left) A sample preprocessed structural MRI scan by SFCN. (right) The age and gender (inset) distribution of the TLSA dataset. (D) Model architecture of the SFCN-based decoder. The SFCN backbone was used to extract features from the MRI scan, which were used by a decoder block to predict brain age. + +$$ +\theta^ {*} = \underset {\theta \in \Theta} {\operatorname {a r g m a x}} \sum_ {i = 1} ^ {N _ {l}} \log p \left(y _ {i} ^ {t} \mid \mathbf {x} _ {i} ^ {t}, \theta\right) + \alpha \sum_ {i = 1} ^ {N} \log q \left(\mathbf {x} _ {i} ^ {t}, \tilde {y} _ {i} ^ {t} \mid \theta\right) +$$ + +After incorporating distributional assumptions in the above equation and subsuming constants with respect to model parameters (e.g., priors on the labels $p(y^t)$ ): + +$$ +\begin{array}{l} \theta^ {*} = \underset {\theta \in \Theta} {\operatorname {a r g m a x}} - \sum_ {i = 1} ^ {N _ {l}} \left(y _ {i} ^ {t} - f \left(\mathbf {x} _ {i} ^ {t}; \theta\right)\right) ^ {2} - \alpha \left(\sum_ {i = 1} ^ {N} \left(\tilde {y} _ {i} ^ {t} - f \left(\mathbf {x} _ {i} ^ {t}; \theta\right)\right) ^ {2} \right. \\ - \sum_ {i = 1} ^ {N} \log \sum_ {l = 1} ^ {N} \exp \left(- \left(\tilde {y} _ {i} ^ {t} - f \left(\mathbf {x} _ {l} ^ {t}; \theta\right)\right) ^ {2}\right)) \tag {6} \\ \end{array} +$$ + +In practice, each update is performed with a batch of labeled and unlabeled data. First, pseudo-labels are computed for the unlabeled data. Then, keeping the pseudo-labels fixed, the second and third terms of this objective are optimized, while the labeled data are used to optimize the first term, concurrently. Note that the optimization is performed with the model parameters initialized to $\theta^s$ , derived from the source model. + +Intuitively, the supervised loss in Equation 6 encourages model predictions to align with their true labels; because + +model parameters were initialized with the source model, this acts as an implicit regularizer discouraging the adapted model from deviating too far from the source model. The second term forces dissimilar predictions for data points with different pseudo-labels. In other words, providing samples with dissimilar label values during training produces disparate representations for data points with distant labels, which in turn helps to learn a better regression line. We use the log-sum-exponent trick to avoid numerical instability during optimizing of the second term. + +# 4. Datasets and Competing Methods + +We employ CRAFT to address challenges relating to data scarcity, missing labels, and continuous label prediction in neuroscience. + +# 4.1. Datasets and models + +# 4.1.1. Gaze prediction from EEG signals + +Saccades are rapid ballistic eye movements, putatively linked to higher-order cognitive functions like attention and working memory [5, 13]. Decoding eye movements from neural data could help neuroscientists develop better models of these complex processes. With this objective, the EEGeyenet benchmark dataset [29] was collected from 356 + +participants. It contains $\sim 47.5$ hours of 128-channel electroencephalogram (EEG) recordings with eye-tracking data collected concurrently. + +Here, we predict saccade amplitudes from these EEG data. We consider two datasets - the "Large Grid" and the "Visual Search" datasets - from the EEGEyeNet benchmark [29] (Fig. 2A). In the Large Grid paradigm, participants made saccades to a target location cued by a dot on the screen. In the Visual Search paradigm, participants searched for a specific symbol in the display and, thus, made saccades during the search. + +Preprocessing. Each dataset was preprocessed using the pipeline described in the benchmark (details in Appendix B.1). We decode saccade magnitudes $(y \in \mathbb{R}^{+})$ using the 128-channel EEG timeseries, sampled at $500\mathrm{Hz}$ $(\mathbf{x} \in \mathbb{R}^{128\times 500})$ , for corresponding trials. Table 5 (see Appendix B.1) lists the details of the train, cross-validation and test distributions for these datasets. The Visual Search paradigm has more training data and was used to train the source model; this was then transferred to predict saccade amplitudes in the Large Grid dataset (target). + +Models. The EEGEyeNet benchmark showed that Pyramidal CNNs [47] were most effective at predicting saccade amplitude. As an alternative to 2D convolutional networks, we also evaluate the performance of a novel end-to-end Long-Short-Term-Memory (LSTM) model. This model is trained on features extracted by separable convolutions along time and electrode dimensions, following EEGNet [31] (Fig. 2B, more details in the Appendix B.1). + +# 4.1.2. "Brain age" prediction from MRI scans + +"Brain age" is the biological age of the brain and is an important biomarker for cognitive health. Studies have shown that accelerated brain aging is associated with serious cognitive impairments like Alzheimer's disease [9]. Hence, estimating brain age is a critical problem in neuro-medicine. + +As a first step, attempts have been made to decode the chronological age of healthy participants using T1-weighted structural MRI scans [41]. A 3D-CNN model (see next, Models) was trained using the large UK-Biobank dataset $(N = 14,503,44 - 80$ years, mean age $= 52.7$ years) [39]. We seek to transfer this pretrained model to decode age in a small Indian cohort – the TATA Longitudinal Study of Aging (TLSA) $(N = 188,46 - 88$ years, mean age $= 66$ years) [45] (Fig. 2C), without access to the UKBiobank data. + +Preprocessing. We followed a preprocessing pipeline identical to [41]. The T1w-MRI scan was brain-extracted, bias-corrected [46], and registered to standard MNI-152 space [38], and cropped to get preprocessed scans $\mathbf{x} \in \mathbb{R}^{160 \times 192 \times 160}$ , which was used to predict brain age. + +Models. Features from the pretrained Simple Fully Convolutional Network (SFCN, Fig. 2D) were used to make the predictions with a regression model; more details are provided in Appendix B.2) [41]. The entire model was then + +adapted to the target dataset (TLSA). + +We also apply CRAFT to people counting and tumor prediction tasks (see Appendix B.3, B.4 for dataset details). + +# 4.2. Competing DA Approaches + +Baseline model. We define a naive baseline model that always predicts the mean label of the training set, ignoring the input features of the test samples. + +Transfer Learning (TL). We naively adapt the source model by minimizing the supervised loss on the labeled samples in the target dataset. + +TASFAR. This is an SF-UDA model (See Section 2) for regression problems. It depends on two statistics from the source dataset – the threshold for confident samples and the mapping between uncertainty to prediction errors. Because the source is unavailable, we compute these using the TL model on the target. For a fair comparison, we add the supervised loss on the labeled data to the unsupervised loss, weighted by the same $\alpha$ as in our model. + +DataFree. This is another SF-UDA model (See Section 2). The original model depends on the distribution of intermediate latents on the source distribution. Again, because the source data is unavailable, we computed latent distribution using the TL model and the labeled target data. As before, the objective was modified to accommodate a supervised loss (same $\alpha$ as in our model). Bin sizes for both TASFAR and DataFree were selected to be the same as in CRAFT. + +Progressive Mixup. This is a semi-supervised, source-free domain adaptation tool (See Section 2). We adapt the model to the regression task by replacing cross-entropy losses with mean-squared error loss. Predictive entropy was replaced by variance to estimate uncertainty. + +BBCN. Bilateral Cycle Consistency is a prototype-based clustering SF-SSDA paradigm (See Section 2). It depends on class-level prototypes for each model. To extend it to regression, we discretize the output space into bins of the same width as the other methods. + +Metrics. We quantify the prediction performance of continuous valued outputs using two metrics - Root Mean Squared Error (RMSE) and Percentage-Bend Correlation [49] coefficient (R). A lower RMSE reflects a closer correspondence between the predicted and actual labels, a higher R-value quantifies better predictions along the best-fit line between the actual and predicted labels. + +In subsequent sections (5.1 and 5.2), we predict saccade amplitude from EEG and brain age from structural MRI scans, respectively. + +# 5. Experiments: Results + +# 5.1. Saccade Amplitude Prediction + +Firstly, we trained an EEGNet-LSTM model on the Visual Search source dataset (RMSE=67.03, R=0.91). This novel + +architecture outperformed the benchmark model (Pyramidal CNNs) by $\sim 16\%$ (benchmark: RMSE=79.59 pix, R=0.87). Next, this was transferred to the Large Grid dataset. The upper 3 rows of Table 1 shows the performance when all the data was used for funetuning. In this case also EEGNet-LSTM outperformed the benchmark (Pyramidal CNN [29]) by a large margin ( $\sim 19\%$ ). Hence, subsequent SSDA experiments were performed using EEGNet-LSTMs. + +All models were trained with mini-batches of 128 samples, and only the best-performing checkpoint on the held-out validation set was used for evaluating the test set. Adam optimizer with an initial learning rate of $10^{-4}$ was used. $\alpha$ in equation 3 was selected with a grid saerch over $\{0.01, 0.1, 1.0\}$ ; empirically, $\alpha = 0.1$ was always favored. + +Table 1. Source Free Semi-supervised Deep Transfer of the model trained on the Visual Search task to the Large-Grid Task. The upper part of the table (rows 1-3) shows the performance ceiling when all labeled data in the target is used for training. The lower part (rows 4-9) shows SF-SSDA results when only $1\%$ of the target samples were labeled (3 seeds). + +
MethodR ↑RMSE (in pix) ↓
Naive Baseline-149.12 ± 0.02
Benchmark (TL, 100%)0.91 ± 0.0163.84 ± 0.75
EEGNet-LSTM (TL, 100%)0.93 ± 0.0151.47 ± 0.63
SF-SSDA (1% labels)
EEGNet-LSTM + TL0.77 ± 0.0192.26 ± 1.66
EEGNet-LSTM + Mixup0.48 ± 0.02135.70 ± 1.25
EEGNet-LSTM + BBCN0.76 ± 0.0199.8 ± 3.35
EEGNet-LSTM + TASFAR0.76 ± 0.0186.41 ± 1.05
EEGNet-LSTM + DataFree0.80 ± 0.0187.64 ± 3.08
EEGNet-LSTM + CRAFT0.81 ± 0.0284.17 ± 3.95
+ +To demonstrate the efficacy of CRAFT for SF-SSDA, target labels from the training data of the Large Grid dataset were randomly dropped in a stratified manner. Figure 3A shows how the performance of these models is affected by reducing the fraction of labeled data ( $n_{ul} / N$ is the fraction of unlabeled samples). A lower proportion of labeled data worsens the performance of all the models (RMSE increases), but CRAFT produces the lowest RMSE for all unlabeled data proportions. Figure 3B shows the prediction of saccade magnitudes ( $|\Delta \overline{r}|$ ) for the best CRAFT-based EEGNet-LSTM model. Table 1 (lower 6 rows) summarizes the performance of CRAFT when $99\%$ of the data were unlabeled; CRAFT outperforms supervised training (TL) by $9\%$ and other SOTA SF-SSDA models by $>4\%$ in terms of RMSE. The closest competitive methods were TASFAR and DataFree. The correlation coefficient (R) also showed similar trends (Table 1; Fig. 6A). + +# 5.2. Brain Age Prediction + +Next, we transferred the SFCN model trained on the UK Biobank MRI scans to the TLSA dataset. The TLSA + +dataset has only 188 MRI scans, so we avoid hyperparameter searches. All models were trained for 25 epochs (batch size = 4) using the Adam optimizer (initial learning rate = 10 $^{-4}$ ), and only the final model was used for inference. Here, we used $\alpha = 0.1$ ; we explore the effect of varying $\alpha$ for ours and competing models in Appendix D. Each metric was calculated for the prediction on the entire dataset, leaving out one fold at a time (4-folds). The top of Table 2 (row 2) quantifies the performance ceiling when all the training data are used (3 seeds). + +Table 2. Semi-supervised transfer of the SFCN model (source: UKB) to the TLSA dataset. Same as in Table 1, except that the lower part of the table (3-8) shows SF-SSDA results when only $20\%$ of the target samples were labeled. (3 seeds) + +
MethodR ↑RMSE (in years) ↓
Naive Baseline-7.91 ± 0.05
SFCN + TL (100%)0.66 ± 0.016.14 ± 0.03
SF-SSDA (20% labels)
SFCN + TL0.41 ± 0.077.41 ± 0.21
SFCN + Mixup0.34 ± 0.047.71 ± 0.14
SFCN + BBCN0.28 ± 0.048.00 ± 0.15
SFCN + TASFAR0.42 ± 0.077.47 ± 0.15
SFCN + DataFree0.50 ± 0.037.36 ± 0.14
SFCN + CRAFT0.51 ± 0.037.14 ± 0.11
+ +For SF-SSDA, we performed experiments closely similar to the saccade task (Section 5.1). Figure 3C shows how these models' performance is affected when the proportion of unlabeled samples increases. Reducing the fraction of labeled data worsens model performance, but CRAFT continues to remain the most competitive of all models. Figure 3D shows the prediction scatter of the best CRAFT-based model. Table 2 summarizes the performance when $80\%$ of the data was unlabeled. CRAFT outperforms supervised training (TL) by $\sim 4\%$ and other SOTA SF-SSDA models by $>3\%$ in terms of RMSE. Similar results were obtained with the R values also (Fig. 6B). Again, TASFAR and DataFree emerge as the most competitive methods. + +Finally, we evaluated the effectiveness of CRAFT at two other real-world regression benchmarks: People Counting [44, 48] and Tumor Size prediction [6, 16]. CRAFT outperformed other SF-SSDA methods by $>5\%$ and $>2\%$ respectively for each dataset; the details are presented in Appendices C.1 and C.2. + +Mitigating Sampling Biases. Next, we show how the unsupervised CRAFT objective could mitigate sampling bias effects in the target training set. In all of the previous simulations, the label distributions of the target train and test data were matched. Here, we deliberately distorted the label distribution of the target training data: $80\%$ of the data above the mean age in the training set were removed, creating a heavy bias toward lower ages. Following this, labels were + +![](images/13cc5ff798edf1da9859ff5eaa5b0d5d8a8f76823e76d7401ba5b67b7dfc7c2a.jpg) +Figure 3. (A, C) RMSE $(\downarrow)$ for source-free semi-supervised domain adaptation of saccade (A) and brain age (C) prediction tasks, trained with varying proportions of unlabeled target data $n_{ul} / N$ . Dashed black line: Performance ceiling. Circles: individual trials. (B, D) Prediction for the best CRAFT model for saccade amplitude (B) and brain age (D), with $90\%$ and $40\%$ unlabeled data, respectively. Triangles: individual participants. Results with a much higher proportion of unlabeled data are shown for the saccade dataset because of the far greater size of this ( $N \sim 12k$ ) compared to the brain age ( $N \sim 180$ ) dataset. + +retained in only $40\%$ of the training data $(n_{ul} / N = 0.6)$ . No such biases were introduced into the test data. The CRAFT model was then trained by imposing the unbiased (true) marginal label distribution on the unlabeled training set (equation 5). In this scenario, again, we find that CRAFT mitigates the effect of sampling biases in the training data (Table 3). It outperforms transfer learning by $\sim 5\%$ , and other sota SF-SSDA models by $>2.5\%$ (RMSE). + +Table 3. Analyses with datasets with sampling error. $80\%$ of the data above the mean age is removed. Furthermore, only $40\%$ of the labels in this mislabeled data are used for semi-supervised transfer of the SFCN model (source: UKB). (3 seeds) + +
MethodR ↑RMSE (in years) ↓
Naive Baseline-8.38 ± 0.22
SFCN + TL0.41 ± 0.058.40 ± 0.27
SFCN + Mixup0.32 ± 0.028.19 ± 0.15
SFCN + BBCN0.23 ± 0.028.53 ± 0.16
SFCN + TASFAR0.39 ± 0.048.57 ± 0.22
SFCN + DataFree0.41 ± 0.038.58 ± 0.23
SFCN + CRAFT0.45 ± 0.037.98 ± 0.25
+ +# 5.3. Analysis of computational complexity + +Table 4 shows the computational complexity of CRAFT and the competing models. We compute both the algorithms' theoretical complexity and running time on the saccade and brain age prediction tasks. The theoretical training complexity is defined as the number of gradient updates per epoch. For inference, the complexity is the same for all models as the same architecture is trained using different adaptation paradigms. Training time and complexity are lowest for TASFAR, amongst all other SF-DA tools. The complexity of CRAFT was comparable to DataFree, the next best method. + +Table 4. Computational complexity. Time per training epoch and for all test data (in min). Parentheses: P- no. of model parameters; N- samples; D- latent dimensionality; $\beta$ - batch size, B- no. of bins, m- no. of inferences for MC Dropout uncertainty estimation. + +
(B) MethodEEGNet-LSTM (P~ 190K)SFCN (P~ 2.95M)
Train /Comp. ↓ (N ~ 12k)Test /Comp. ↓ (N ~ 2.5k)Train/Comp. ↓ (N ~ 150)Test/Comp. ↓ (N ~ 50)
TL0.15 / O(N)0.02 / O(N)0.12 / O(N)0.04 / O(N)
Mixup0.74 / O(Nm)0.02 / O(N)0.74 / O(Nm)0.04 / O(N)
BBCN2.71 / O(NBD)0.02 / O(N)1.29 / O(NBD)0.04 / O(N)
TASFAR0.30 / O(N)0.02 / O(N)0.19 / O(N)0.04 / O(N)
DataFree0.55 / O(ND2)0.02 / O(N)0.32 / O(ND2)0.04 / O(N)
CRAFT0.45 / O(N(β + B))0.02 / O(N)0.36 / O(N(β + B))0.04 / O(N)
+ +# 6. Concluding Remarks + +Data scarcity poses a serious challenge for applying deep-learning models in real-world settings. Training overparameterized models on a few hundred data points leads to overfitting and poor generalization. Conventional domain adaptation algorithms, to transfer pretrained networks, typically need access to source domain data, but these may not always be available. In this study, we propose a source-free semi-supervised domain adaptation framework that recalibrates the decision boundary without aligning the intermediate representation of the domains. In fields like medicine, resource constraints associated with labeling data are significant; therefore, leveraging the large amounts of unlabeled data becomes critical. We propose CRAFT as an SF-SSDA approach for such source-free, label-sparse settings that are common in real-world applications. + +# 7. Acknowledgements + +This work was supported by the following funding sources: a Prime Minister's Research Fellowship (to M.B.), a DST SwarnaJayanti fellowship, a Pratiksha Trust Intramural grant, a Tata Trusts grant and a Google Research grant (to D.S.). We also thank the Tata Longitudinal Study on Aging at the Indian Institute of Science for access to MRI scans from their cohort. + +# References + +[1] Jiayu An, Changming Zhao, and Dongrui Wu. Semi-supervised generalized source-free domain adaptation (ssgsfda). In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2023. 3 +[2] APTOS. Aptos 2019 blindness detection dataset, 2019. 1 +[3] Sourabh Balgi and Ambedkar Dukkipati. CUDA: Contradistinguisher for unsupervised domain adaptation. In Proceedings of the IEEE International Conference on Data Mining, 2019. 1 +[4] Sourabh Balgi and Ambedkar Dukkipati. Contradistinginguisher: A Vapnik's imperative to unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4730-4747, 2022. 1, 3, 4 +[5] Scott L. Brincat, Jacob A. Donoghue, Meredith K. Kahnke, Simon Kornblith, Mikael Lundqvist, and Earl K. Miller. Interhemispheric transfer of working memories. Neuron, 109: 1055-1066.e4, 2021. 5 +[6] Peter Bandy, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, Quanzheng Li, Farhad Ghazvinian Zanjani, Svitlana Zinger, Keisuke Fukuta, Daisuke Komura, Vlado Ovtcharov, Shenghua Cheng, Shaoqun Zeng, Jeppe Thagaard, Anders B. Dahl, Huangjing Lin, Hao Chen, Ludwig Jacobsson, Martin Hedlund, Melih Çetin, Eren Halici, Hunter Jackson, Richard Chen, Fabian Both, Jörg Franke, Heidi Küsters-Vandevelde, Willem Vreuls, Peter Bult, Bram van Ginneken, Jeroen van der Laak, and Geert Litjens. From detection of individual metastases to classification of lymph node status at the patient level: The camelyon17 challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2019. 2, 7, 3 +[7] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140: 325-331, 2020. 2 +[8] Thomas A. Carlson. Orientation decoding in human visual cortex: New insights from an unbiased perspective. Journal of Neuroscience, 34(24):8373-8383, 2014. 1 +[9] J H Cole, S J Ritchie, M E Bastin, M C Valdés Hernández, S Muñoz Maniega, N Royle, J Corley, A Pattie, S E Harris, Q Zhang, N R Wray, P Redmond, R E Marioni, J M Starr, S R Cox, J M Wardlaw, D J Sharp, and I J Deary. Brain age predicts mortality. Molecular Psychiatry, 23:1385-1392, 2018. 6 +[10] Sébastien M. Crouzet. Fast saccades toward faces: Face detection in just 100 ms. Journal of Vision, 10:1-17, 2010. 2 +[11] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. PMID: 20144289. 1 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3 + +[13] Heiner Deubel and Werner X. Schneider. Saccade target selection and object recognition: Evidence for a common attentional mechanism. Vision Research, 36:1827-1837, 1996. 5 +[14] Yuhe Ding, Lijun Sheng, Jian Liang, Aihua Zheng, and Ran He. Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks, 167:92-103, 2023. 2 +[15] Cian Eastwood, Ian Mason, Christopher K I Williams, and Bernhard Scholkopf. In Proceedings of the Tenth International Conference on Learning Representations, 2022. Tenth International Conference on Learning Representations 2022, ICLR 2022; Conference date: 25-04-2022 Through 29-04-2022. 2, 3 +[16] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen A. W. M. van der Laak, , and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 318(22):2199-2210, 2017. 2, 7, 3 +[17] Saminger-Platz et al. Central moment discrepancy (cmd) for domain-invariant representation learning. *ICLR*, 2017. 1, 2 +[18] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 174:106230, 2024. 1 +[19] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17:1-35, 2016. 1, 2 +[20] Alessandro T. Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M. Cichy. The algonauts project 2025 challenge: How the human brain makes sense of multimodal movies, 2025. 1 +[21] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems. MIT Press, 2004. 3, 4 +[22] Arthur Gretton, Kenji Fukumizu, Zäïd Harchaoui, and Bharath K. Sriperumbudur. A fast, consistent kernel two-sample test. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2 +[23] Yan Hao, Yuhong Guo, and Chunsheng Yang. Source-free unsupervised domain adaptation with surrogate data generation. In British Machine Vision Conference, 2021. 3 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Computer Vision - ECCV 2016, pages 630-645, Cham, 2016. Springer International Publishing. 3 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3 +[26] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and Shueng-Han Gary Chan. Target-agnostic source-free domain adaptation for regression tasks. 2024 IEEE 40th In + +ternational Conference on Data Engineering (ICDE), pages 1464-1477, 2023. 2, 3 +[27] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 2 +[28] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 3 +[29] Ard Kastrati, Martyna Plomecka, Damian Pascual Ortiz, Lukas Wolf, Victor Gillioz, Roger Wattenhofer, and Nicolas Langer. Eegeyenet: a simultaneous electroencephalography and eye-tracking dataset and benchmark for eye movement prediction. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 1, 2, 5, 6, 7 +[30] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method, 2021. 2 +[31] Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. Eeg-net: a compact convolutional neural network for eeg-based brain-computer interfaces. Journal of Neural Engineering, 15:056013, 2018. 6, 2 +[32] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A Comprehensive Survey on Source-Free Domain Adaptation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 46(08):5743-5762, 2024. 1, 2 +[33] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2 +[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the 32nd International Conference on Machine Learning, pages 97-105, Lille, France, 2015. PMLR. 2 +[35] Jie Ma, Zhijun Xu, Xinyu Xiong, MaYiDiLi NiJiaTi, Dongyu Zhang, Weijun Sun, Feng Gao, and Guanbin Li. Source-free semi-supervised domain adaptation for tuberculosis recognition. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5, 2024. 2, 3 +[36] Ning Ma, Jiajun Bu, Lixian Lu, Jun Wen, Zhen Zhang, Sheng Zhou, and Xifeng Yan. Semi-supervised hypothesis transfer for source-free domain adaptation, 2021. 2, 3 +[37] Ning Ma, Haishuai Wang, Zhen Zhang, Sheng Zhou, Hongyang Chen, and Jiajun Bu. Source-free semi-supervised domain adaptation via progressive mixup. Knowledge-Based Systems, 262:110208, 2023. 2, 3 +[38] Pravat K. Mandal, Rashima Mahajan, and Ivo D. Dinov. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. Journal of Alzheimer's Disease, 31:S169-S188, 2012. 6 +[39] Karla L Miller, Fidel Alfaro-Almagro, Neal K Bangerter, David L Thomas, Essa Yacoub, Junqian Xu, Andreas J Bartsch, Saad Jbabdi, Stamatios N Sotiropoulos, Jesper L R Andersson, Ludovica Griffanti, Gwenaëlle Douaud, + +Thomas W Okell, Peter Weale, Iulius Dragonu, Steve Garratt, Sarah Hudson, Rory Collins, Mark Jenkinson, Paul M Matthews, and Stephen M Smith. Multimodal population brain imaging in the uk biobank prospective epidemiological study. Nature Neuroscience, 19:1523-1536, 2016. 1, 6 +[40] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation, 2020. 1, 2 +[41] Han Peng, Weikang Gong, Christian F Beckmann, Andrea Vedaldi, and Stephen M Smith. Accurate brain age prediction with lightweight deep neural networks. Medical Image Analysis, 68:101871, 2021. 6, 2 +[42] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In Computer Vision - ECCV 2022, pages 165-182, Cham, 2022. Springer Nature Switzerland. 2 +[43] Xiaofeng Qu, Li Liu, Lei Zhu, Liqiang Nie, and Huaxiang Zhang. Source-free style-diversity adversarial domain adaptation with privacy-preservation for person re-identification. Knowledge-Based Systems, 283:111150, 2024. 1 +[44] Vishwanath Sindagi, Rajeev Yasarla, and Vishal Patel. Pushing the Frontiers of Unconstrained Crowd Counting: New Dataset and Benchmark Method. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1221-1231, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 2, 7 +[45] Jonas S Sundarakumar, Albert Stezin, Abhishek L Menes-gere, and Vijayalakshmi Ravindranath. Rural-urban and gender differences in metabolic syndrome in the aging population from southern india: Two parallel, prospective cohort studies. eClinicalMedicine, 47:101395, 2022. 1, 2, 6 +[46] J-Donald Tournier, Robert Smith, David Raffelt, Rami Tabbara, Thijs Dhollander, Maximilian Pietsch, Daan Christi-aens, Ben Jeurissen, Chun-Hung Yeh, and Alan Connelly. Mrtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202:116137, 2019. 6 +[47] Ihsan Ullah and Alfredo Petrosino. About pyramid structure in convolutional neural networks, 2016. 6, 2 +[48] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2, 7 +[49] Rand R. Wilcox. The percentage bend correlation coefficient. Psychometrika, 59:601-616, 1994. 6 +[50] Garrett Wilson and Diane J. Cook. A survey of unsupervised deep domain adaptation. ACM Trans. Intell. Syst. Technol., 11(5), 2020. 1 +[51] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36(8):3746-3766, 2021. 2 +[52] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation, 2022. 2 \ No newline at end of file diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/images.zip b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b8d4f873f79fc1a01f63f2083071fa92be694647 --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:083bfa295f5dfff65dc7cfc6fc5a442f9152bc4a18f3bc58a525398ac20e73a3 +size 418593 diff --git a/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/layout.json b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54b9024b441c012b0849005ded79b08eaab43b09 --- /dev/null +++ b/2025/Semi-supervised Deep Transfer for Regression without Domain Alignment/layout.json @@ -0,0 +1,8911 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 534, + 121 + ], + "type": "text", + "content": "Semi-supervised Deep Transfer for Regression without Domain Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 473, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 473, + 156 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 473, + 156 + ], + "type": "text", + "content": "Mainak Biswas1,2, Ambedkar Dukkipati3, Devarajan Sridharan1,2,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "text", + "content": "Brain, Computation and Data Sciences, " + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "text", + "content": "Centre for Neuroscience, " + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 157, + 561, + 185 + ], + "type": "text", + "content": "Computer Science and Automation Indian Institute of Science, Bangalore - 560012" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 188, + 484, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 188, + 484, + 198 + ], + "spans": [ + { + "bbox": [ + 136, + 188, + 484, + 198 + ], + "type": "text", + "content": "mainakbiswas@iisc.ac.in, ambedkar@iisc.ac.in, sridhar@iisc.ac.in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "text", + "content": "Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data or reliably labeled target data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT - a Contradistinguisher-based Regularization Approach for Flexible Training - for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and \"brain age\" prediction with structural MRI data. For both datasets, CRAFT yielded up to " + }, + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "text", + "content": " improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than " + }, + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 674 + ], + "type": "text", + "content": ". Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 715 + ], + "type": "text", + "content": "Keywords: source-free DA, continuous label prediction, gaze prediction, brain age prediction, healthcare" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 227, + 394, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 227, + 394, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 227, + 394, + 239 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 402 + ], + "type": "text", + "content": "For the successful application of deep learning models in the real world, they must be robust to \"domain shift\" [50]. For example, in biology and medicine, supervised deep learning models are often trained on large, high-quality source datasets collected in resource-rich settings [11, 39]. However, naive transfer or even simple finetuning fails to produce high accuracies when tested with smaller target datasets of mixed quality collected [2, 45]. As a result, state-of-the-art domain adaptation (DA) approaches, which enable the source models to generalize successfully to target datasets, must be employed. Many of these approaches involve explicit alignment of source and target, while learning a shared predictive model for both domains [17, 19, 40]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 557 + ], + "type": "text", + "content": "A particular challenge occurs with DA when the source dataset is unavailable. This can happen either because of privacy or proprietary concerns, such as with medical images or patents [43]. Alternatively, source datasets may be too large, incurring prohibitive storage and computation costs in low-resource settings [39]. Source-free (SF) domain adaptation methods that address these challenges are getting increasingly popular [18, 32]. Such methods generalize the source model to unlabeled or partially target data (unsupervised or semi-supervised DA) with one of two approaches: data-based or model-based. While data-based methods create surrogate source data, model-based methods adapt the source model to the target domain (see Section 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": "We address model-based SF-DA in a semi-supervised setting. Specifically, we focus on neuroscience applications where source data are inaccessible, and target labels are reliably available only for a few samples. Regression tasks – like predicting brain age (Fig. 1C) [39], saccades [29], neural activity [20], and stimulus orientation [8] – are ubiquitous in neuroscience; we tackle a subset of these. We build upon Contradistinguisher [3, 4], a recently-proposed tool for effective unsupervised DA (CUDA). Unlike popular representation alignment DA methods [19, 40], CUDA directly learns a common model for both domains. Yet, CUDA requires full access to source data and addresses classification tasks. Therefore, it is not directly applicable to our problem." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "827" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 78, + 139, + 152 + ], + "blocks": [ + { + "bbox": [ + 64, + 83, + 72, + 91 + ], + "lines": [ + { + "bbox": [ + 64, + 83, + 72, + 91 + ], + "spans": [ + { + "bbox": [ + 64, + 83, + 72, + 91 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 78, + 139, + 152 + ], + "lines": [ + { + "bbox": [ + 82, + 78, + 139, + 152 + ], + "spans": [ + { + "bbox": [ + 82, + 78, + 139, + 152 + ], + "type": "image", + "image_path": "eb9853653c6b9022712a0579a3ff2529cc99fff34dab54d3756ff8484db4e5c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 80, + 150, + 90 + ], + "lines": [ + { + "bbox": [ + 140, + 80, + 150, + 90 + ], + "spans": [ + { + "bbox": [ + 140, + 80, + 150, + 90 + ], + "type": "text", + "content": "X" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 156, + 80, + 220, + 152 + ], + "blocks": [ + { + "bbox": [ + 156, + 80, + 220, + 152 + ], + "lines": [ + { + "bbox": [ + 156, + 80, + 220, + 152 + ], + "spans": [ + { + "bbox": [ + 156, + 80, + 220, + 152 + ], + "type": "image", + "image_path": "50abad7ebb0e934b69ecdb035b0c1da341db9610f3320d5e761bc874f72f3d96.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 77, + 293, + 152 + ], + "blocks": [ + { + "bbox": [ + 223, + 77, + 293, + 152 + ], + "lines": [ + { + "bbox": [ + 223, + 77, + 293, + 152 + ], + "spans": [ + { + "bbox": [ + 223, + 77, + 293, + 152 + ], + "type": "image", + "image_path": "7ff5d8c03c3f4ad2d8c212ee190fd4cb9ac2db57efa1226e6b4c1f6914156dc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 169, + 165, + 231 + ], + "blocks": [ + { + "bbox": [ + 63, + 160, + 71, + 168 + ], + "lines": [ + { + "bbox": [ + 63, + 160, + 71, + 168 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 71, + 168 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 169, + 165, + 231 + ], + "lines": [ + { + "bbox": [ + 65, + 169, + 165, + 231 + ], + "spans": [ + { + "bbox": [ + 65, + 169, + 165, + 231 + ], + "type": "image", + "image_path": "90f5303b5b158277667aad3571bc0e9dbed2b57cdc3db5b3e10ecf16eeeab682.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 252, + 296, + 341 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 341 + ], + "type": "text", + "content": "Figure 1. Schematic of Semisupervised Deep Transfer without Domain Alignment. (A) The approach seeks to transfer a pretrained model to a target domain with limited labeled data when the source dataset is unavailable. (B) Because the source dataset is unavailable, we tune the decision boundaries without explicitly aligning the representation of the domains, combining a supervised loss and the CUDA loss (regularizer). (C) The algorithm targets deep transfer in regression tasks, like brain age prediction." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 155, + 291, + 230 + ], + "blocks": [ + { + "bbox": [ + 171, + 160, + 179, + 168 + ], + "lines": [ + { + "bbox": [ + 171, + 160, + 179, + 168 + ], + "spans": [ + { + "bbox": [ + 171, + 160, + 179, + 168 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 155, + 291, + 230 + ], + "lines": [ + { + "bbox": [ + 181, + 155, + 291, + 230 + ], + "spans": [ + { + "bbox": [ + 181, + 155, + 291, + 230 + ], + "type": "image", + "image_path": "3c6cd4215a790df2af29f683479d512f61696c30867f90231da3da7ce48aa7b8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 363, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 542 + ], + "type": "text", + "content": "Here, we leverage CUDA to develop a new algorithm - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT) which performs deep transfer, without domain alignment, even when the source dataset is unavailable (Fig. 1A-B). We also extend CUDA to a regression setting. Previous studies approach regression tasks naively by binning continuous labels into discrete bins [40], sometimes employing a rank-ordinal objective [7]. Regardless, the efficacy of such approaches may depend critically on bin sizes. Our CRAFT model uses a principled approach to extend CUDA to predictions involving continuous-valued outputs. We employ binning as a practical strategy to generate and optimize pseudo-labels, without constraining the model to produce discrete outputs. Our study makes the following main contributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": "- First, we extend CUDA to source-free DA for regression tasks involving continuous-valued outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": "- Second, we formulate a theoretically-motivated semi-supervised objective by combining supervised and unsupervised losses with a tunable hyperparameter " + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": " (Section 3). We also derive the unsupervised loss as a maximum entropy prior to model parameters (Appendix A.2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 295, + 677 + ], + "type": "text", + "content": "- Third, we show the effectiveness of CRAFT on two neuroscience challenges - gaze prediction from electroencephalographic (EEG) [29] data and brain age prediction from magnetic resonance imaging (MRI) data [45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 714 + ], + "type": "text", + "content": "- Fourth, we also apply CRAFT to two real-world regression benchmarks involving people counting [44, 48] and tumor size prediction [6, 16]." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "type": "text", + "content": "- Fifth, we compare CRAFT against state-of-the-art SF-DA methods [15, 26, 35, 37] and show that the best improvements occur as unlabeled target samples increase, demonstrating its efficacy for semi-supervised SF-DA." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 121, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 144 + ], + "type": "text", + "content": "- Finally, we analyze the computational complexity and demonstrate numbers competitive with sota models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 144, + 526, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 526, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 526, + 156 + ], + "type": "text", + "content": "- Our code is publicly available (see Appendix E.2)." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 187, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 354 + ], + "type": "text", + "content": "Domain alignment methods. Popular Unsupervised Domain Adaptation (UDA) approaches rely on learning common intermediate representations across source and target domains. Domain Adversarial training of Neural Networks (DANN) [19] uses a discriminator to learn common representations across both domains adversarially. In Importance Weighting [40] (IW), higher importance is afforded to losses corresponding to source samples more likely to belong to the target domain. Central Moment Discrepancy (CMD) [17] minimizes the central moments of the latent representations between the domains, while others optimize their Maximum Mean Discrepancy (MMD) [22, 34]. Despite being effective, they cannot be used when the source is unavailable and will not be discussed further." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 510 + ], + "type": "text", + "content": "Source-free UDA. To address DA scenarios where source data and its target labels are unavailable, several source-free unsupervised DA approaches have been recently proposed. Broadly, they are categorized as: Data-based and Model-based [32]. Data-based methods generate a source-like distribution [14, 30]. However, they often depend on the availability of generative models trained on the source, which could be unavailable. For example, SF-UDA through domain-specific perturbation (AUGFree) [51] assumes the existence of domain-invariant features for prediction. Since the source is unavailable, AUGFree perturbs the target data and then learns a transformation to the invariant features by adversarially aligning the target and the perturbed target." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "On the other hand, model-based methods finetune the source model on the target using strategies like contrastive learning [27], pseudo-labeling [26], or entropy minimization [36]. They are often prone to overfitting to pseudolabels. For example, Class-Balanced Multicentric Dynamic Prototype Strategy for SF-UDA (BMD) [42] follows a pseudo-labeling strategy using class prototypes based on the source model predictions, and in turn uses these labels for training. Source Data-Absent UDA Through Hypothesis Transfer (SHOT++) [33] combines three objectives - a) maximizing mutual information between the predictions and inputs, b) self-supervised loss based on prototype-based clusters, and c) a Mixup-based (See SF-SSDA) loss [37]. Attracting and Dispersing (AaD) [52] defines a neighboring cluster (K nearest neighbors) and a randomly selected background cluster for each sample, and maximizes the log-likelihood of the ratio of their distributions." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "828" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "Target-agnostic SF UDA for Regression Tasks (TAS-FAR) [26] works on the principle that prediction uncertainty is correlated with prediction error. It uses Monte-Carlo dropout to estimate label uncertainty and divides the target dataset into confident and uncertain samples. A discrete density map is estimated using the confident data, which is then used as weights to recalibrate the pseudo-labels of uncertain data. SF-UDA to measurement Shift via Bottom-Up Feature Restoration (BUFR/DataFree) [15] is designed on the principle of intermediate feature alignment of the source and the target distributions. During adaptation, the model minimizes the symmetric KL divergence between feature distributions of source (fixed) and target data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 314 + ], + "type": "text", + "content": "Here, we compare our model's performance against two of the above methods - TASFAR and DataFree. TASFAR was selected as it is a pseudo-labeling-based method suitable for regression. On the other hand, DataFree is a pseudo-label-free feature alignment method, not constrained to classification tasks. By adding a supervised loss (Section 4.2), we evaluate these as semi-supervised models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "type": "text", + "content": "Source-free SSDA. Source-free semi-supervised domain adaptation (SF-SSDA) models best address the scenario of SF-DA when few target labels are available. For example, Generalized SF-SSDA [1] minimizes source loss alongside an unsupervised target loss which maximizes the cosine similarity between predictions of k-nearest neighbors. The model is then transferred by adversarial training based on source features generated using a Conditional VAE. Similarly, [23] proposes generating surrogate source data from target samples using gradient-based optimization. Learning from Different Samples for SF-SSDA [28] minimizes the contrastive loss on the predicted probabilities of the target domain. Furthermore, it minimizes supervised and contrastive loss on a set comprising labeled and unlabeled data with low uncertainty. Likewise, Mutual Enhancement training for SS Hypothesis transfer (MESH) [36] minimizes the supervised loss on the labeled samples and entropy on pseudo-labeled samples. Finally, the target manifold is smoothened by making predictions robust to perturbations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Progressive Mixup [37] is trained on augmentations by convex combinations of data on which the source model is confident and ones on which it is uncertain (hybrid mixup). Similarly, another set is obtained by convex combinations of samples in the (pseudo) labeled dataset (self-mixup). In contrast, Bilateral-Branch Consistency Network (BBCN) [35] computes prototypes as the weighted mean of feature representations of the target dataset by the source model. Pseudo-labels are selected based on the proximity of unlabeled data to these prototypes. A separate target branch is trained using the pseudo-labels, and a consistency loss is added between the representations of both branches. The need for class-level prototype generation makes it challenging to adapt this model to regression tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "We compare the performance of CRAFT with two SFSSDA methods - SF-SSDA via progressive Mixup [37], and BBCN [35], as they have little dependence on the source distribution and its derivatives. Moreover, these methods are based on complementary ideas, and the latter [35] has been applied specifically in medical imaging, for tuberculosis recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 169, + 430, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 430, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 430, + 182 + ], + "type": "text", + "content": "3. The CRAFT Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 190, + 425, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 425, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 425, + 201 + ], + "type": "text", + "content": "3.1. Model formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 209, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 376 + ], + "type": "text", + "content": "Background. Our objective is to develop an approach for source-free semi-supervised domain adaptation, for regression tasks. For this, we build upon the Contradistinguisher-based Unsupervised Domain Adaptation (CUDA) model [4]. CUDA obviates the need for learning a common intermediate representation across the source and the target datasets. It optimizes a joint distribution over the target features and category labels to adapt the source model. Because CUDA is intended for unsupervised domain adaptation (UDA), target labels are unavailable. Hence, the algorithm uses a two-step joint optimization over the target labels and model parameters. First, it generates pseudo-labels on the target dataset from the source model [21]. Then, it updates the model parameters by fixing the pseudo-labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 496 + ], + "type": "text", + "content": "In this study, we extend CUDA to a source-free semi-supervised setting (SF-SSDA). Briefly, to the supervised label loss, we add a 'regularized', unsupervised loss as a prior on the model parameters. As we will show in the next section, the model can be flexibly applied, either in SF-UDA case or in the SF-SSDA case, depending on the weightage given to the supervised loss. We, moreover, extend the model to continuous label prediction. We call our new formulation CRAFT - Contradistinguisher-based Regularization Approach for Flexible Training (CRAFT)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": "Objective. We consider a task in which a source model " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\theta^s" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": " trained on a source dataset - involving continuous label prediction (regression) - is already available to us. The target dataset comprises i.i.d samples from an unknown distribution " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "p^t (\\mathbf{x},y)" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t = \\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t\\}_{i = 1}^{N_{ul}}" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": " is a feature-vector, and " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "y_{i}^{t}\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": " is its corresponding label. Target training data are partially and sparingly labeled; we define the proportion of unlabeled samples " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "(N_{ul})" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": " to the total number of samples " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "inline_equation", + "content": "N_{ul}:(N_l + N_{ul})" + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": "We formulate, first, the supervised objective. For continuous label prediction problems, the conditional density of the label given the features is assumed to be normally distributed. The mean is modeled using a neural network " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "f(:,\\theta)" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", as: " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "p(y^{t}|\\mathbf{x}^{t},\\theta) = \\mathcal{N}(y^{t};f(\\mathbf{x}^{t};\\theta),c)" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": ", and the variance, " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " is assumed to be constant as it simplifies the optimization. In the final formulation, the variance occurs as the weight of the unsupervised loss, " + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 606, + 553, + 713 + ], + "type": "text", + "content": " (equation 3); therefore, only one of the two" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "829" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "hyperparameters (" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ") needs to be tuned. Hence, the log-likelihood can be written as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 105, + 295, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 105, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 97, + 105, + 295, + 138 + ], + "type": "interline_equation", + "content": "\\log p \\left(\\mathcal {D} ^ {t} | \\theta\\right) \\propto - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\tag {1}", + "image_path": "ca4535397ea8cb90e9a23341c3b6fad7db49874e05efbd294539d697003320fd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 296, + 178 + ], + "type": "text", + "content": "The unsupervised objective is essentially identical to the CUDA objective and is defined as a joint log density over the labels and model parameters as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 184, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 184, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 184, + 295, + 217 + ], + "type": "interline_equation", + "content": "\\log q \\left(\\mathbf {x} ^ {t}, y ^ {t} \\mid \\theta\\right) = \\log \\left[ \\frac {p \\left(y ^ {t} \\mid \\mathbf {x} ^ {t} , \\theta\\right)}{\\sum_ {i = 1} ^ {N} p \\left(y ^ {t} \\mid \\mathbf {x} _ {i} ^ {t} , \\theta\\right)} p \\left(y ^ {t}\\right) \\right] \\tag {2}", + "image_path": "f401e17f7628d2a1ea2ffe6709da53f32085a58e65781278259cf0c2d7c8b36d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "text", + "content": "This joint distribution seeks to reduce the model's predictive bias when faced with a distribution shift in the target domain (see Fig. 1B). The denominator, which reflects the total probability of predicting a particular class, normalizes for biases originating from the source data. In other words, model predictions are normalized as " + }, + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "inline_equation", + "content": "p(y^{t}|\\mathbf{x}^{t},\\theta) / \\sum_{i = 1}^{N}p(y^{t}|\\mathbf{x}_{i}^{t},\\theta)" + }, + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "text", + "content": ", such that no class has a biased probability of being predicted. Finally, the model predictions align with the target label distribution, using the supplied marginal " + }, + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "inline_equation", + "content": "p(y^t)" + }, + { + "bbox": [ + 55, + 224, + 296, + 343 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "text", + "content": "Note, that this objective does not depend on whether the samples are labeled or not. Optimizing the joint distribution: 1) normalizes inherent bias towards one label over the other by calibrating the prediction for a label with its total weight in the dataset, 2) and explicitly trains " + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "text", + "content": " to match the marginals - i.e., " + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} q(\\mathbf{x}_i^t, y^t|\\theta) = p(y^t)" + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "text", + "content": ", thereby matching high-level statistics like per-class proportions in the training distribution. Importantly, with an informative prior " + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "inline_equation", + "content": "p(y^t)" + }, + { + "bbox": [ + 55, + 343, + 296, + 474 + ], + "type": "text", + "content": " that specifies the putative distribution of the target labels, pseudo-label prediction can be biased toward labels that are more likely (see Section 3.2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": "Combining the supervised and the unsupervised objective, the semi-supervised CRAFT objective is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 506, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 506, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 61, + 506, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\mathcal {D} ^ {t}, \\theta\\right) = \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, y _ {i} ^ {t} \\mid \\theta\\right) \\tag {3}", + "image_path": "e1bad7170d6f4a2cd1947191a4efbbe07157bbadd0cc26f8c9c5123872515e68.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "text", + "content": "In the above equation, " + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "text", + "content": " is the weight associated with the unsupervised objective and is tuned as a hyperparameter on a held-out cross-validation set during training. For deep-transfer, model parameters are initialized to " + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\theta^s" + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "text", + "content": " derived from the source model, and then gradient descent is performed to obtain " + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\theta^t" + }, + { + "bbox": [ + 55, + 546, + 296, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Theoretical motivation. Artificial Neural Networks (ANNs) with hypotheses space " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " are trained to learn a parametric approximation " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},y|\\theta)" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta \\in \\Theta" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", such that the log-likelihood of the training data given the parameters is maximized. However, training ANNs often does not account for prior distributions over model parameters. In Appendix A.1, we demonstrate that the CRAFT semi-supervised objective can be theoretically derived as MAP" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "estimation of the parameters " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "(\\log p(\\theta|\\mathcal{D}))" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", which we write as a sum of the supervised objective, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\log p(\\mathcal{D}|\\theta)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", and a prior on the model parameters, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "p(\\theta)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ". Specifically, we demonstrate, using the principle of entropy maximization [21], that the unsupervised objective is related to the prior on the model parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 150, + 430, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 150, + 430, + 163 + ], + "spans": [ + { + "bbox": [ + 314, + 150, + 430, + 163 + ], + "type": "text", + "content": "3.2. Model optimization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 167, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 239 + ], + "type": "text", + "content": "We optimize the supervised loss of the CRAFT objective with conventional gradient descent, jointly optimizing both components in the loss (equation 3). The unsupervised component of the loss is dependent on the labels " + }, + { + "bbox": [ + 313, + 167, + 554, + 239 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 313, + 167, + 554, + 239 + ], + "type": "text", + "content": ", which are unavailable for the vast majority of target data. To tackle this, we employ a 2-step optimization process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 554, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 275 + ], + "type": "text", + "content": "Pseudo-label selection. As the first step, we fix the model parameters " + }, + { + "bbox": [ + 313, + 239, + 554, + 275 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 239, + 554, + 275 + ], + "type": "text", + "content": " and select the label that maximizes the joint distribution for each datapoint:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 280, + 555, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 280, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 383, + 280, + 555, + 301 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} q \\left(\\mathbf {x} _ {i} ^ {t}, y ^ {t} \\mid \\theta\\right) \\tag {4}", + "image_path": "7f749b0d02ea78914c80ef43697c5d301024a18b42927957e7cb4ff64ec82a2e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": " the pseudo-label for feature vector " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": ". For classification tasks, the optimization in equation 4 can be achieved by selecting the best label from a finite set of class labels that maximizes " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": " [4]. But, for regression problems such as ours, class labels map to real numbers spanning the range of the model outputs " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "f(:,\\theta):\\mathbb{R}^d\\mapsto [a,b]" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": ". To address this, we rewrite the above equation under the assumption that the continuous-valued labels are distributed normally, with mean " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}^t;\\theta)" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": ", and variance " + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": ". Hence, the above equation can be written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 429, + 555, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 429, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 354, + 429, + 555, + 459 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {i} ^ {t} = \\underset {y ^ {t} \\in \\mathcal {Y}} {\\operatorname {a r g m a x}} \\frac {\\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {i} ^ {t} ; \\theta\\right) , c\\right) p \\left(y ^ {t}\\right)}{\\sum_ {l = 1} ^ {N} \\mathcal {N} \\left(y ^ {t} ; f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right) , c\\right)} \\tag {5}", + "image_path": "9b14fa763656d6553eae27f992eb58ffbe957e999443ba7e9639cff6aee34aa6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "text", + "content": "Equation 5 can be optimized with gradient ascent with respect to " + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "inline_equation", + "content": "y^{t}" + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "text", + "content": " for the objective log " + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_i^t, y^t|\\theta)" + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "text", + "content": ". But, the pseudo-labels must be estimated before gradient descent can be performed on the model parameters (see next step, equation 6). This requires nested gradient descent, which increases computational complexity. Moreover, to employ an informative prior over target labels, we fit a class of mixture models (described in Appendix A.3) to estimate " + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "inline_equation", + "content": "p(y)" + }, + { + "bbox": [ + 313, + 462, + 554, + 677 + ], + "type": "text", + "content": " from the data; this makes backpropagation more cumbersome. To overcome these challenges, we divide the entire range of labels into small, discrete bins whose midpoints are the candidate pseudo-labels. Hence, this step in the optimization can be achieved faster pseudo-label selection from a discrete set. Note that this approach does not transform the model outputs to be discrete; it merely represents a convenient optimization strategy to update the parameters. In the experimental section, we provide a recipe for choosing bin sizes and study the impact of varying bin sizes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Maximization. In the second step, we optimize the model parameters by using the labeled data " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t,y_i^t\\}_{i = 1}^{N_l}" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " and fixing the pseudo-labels " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i^t,\\tilde{y}_i^t\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": " , as:" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "830" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 83, + 296, + 185 + ], + "blocks": [ + { + "bbox": [ + 78, + 78, + 86, + 86 + ], + "lines": [ + { + "bbox": [ + 78, + 78, + 86, + 86 + ], + "spans": [ + { + "bbox": [ + 78, + 78, + 86, + 86 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 87, + 83, + 296, + 185 + ], + "lines": [ + { + "bbox": [ + 87, + 83, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 83, + 296, + 185 + ], + "type": "image", + "image_path": "888e5b032c4d9b15c42bc7c8b1148856bc99f3a289a87f7305873188cd7787c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 318, + 95, + 431, + 188 + ], + "blocks": [ + { + "bbox": [ + 313, + 78, + 321, + 87 + ], + "lines": [ + { + "bbox": [ + 313, + 78, + 321, + 87 + ], + "spans": [ + { + "bbox": [ + 313, + 78, + 321, + 87 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 95, + 431, + 188 + ], + "lines": [ + { + "bbox": [ + 318, + 95, + 431, + 188 + ], + "spans": [ + { + "bbox": [ + 318, + 95, + 431, + 188 + ], + "type": "image", + "image_path": "0d0298fa1f99250ae7821fa1d04b391c8c05d5d0613e83037e998a8ea29c239b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 432, + 87, + 538, + 187 + ], + "blocks": [ + { + "bbox": [ + 432, + 87, + 538, + 187 + ], + "lines": [ + { + "bbox": [ + 432, + 87, + 538, + 187 + ], + "spans": [ + { + "bbox": [ + 432, + 87, + 538, + 187 + ], + "type": "image", + "image_path": "c6ceebd152ad3338c057f9ad722eaaa208b690c7917774b10832407957cb3094.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 202, + 299, + 318 + ], + "blocks": [ + { + "bbox": [ + 77, + 196, + 85, + 205 + ], + "lines": [ + { + "bbox": [ + 77, + 196, + 85, + 205 + ], + "spans": [ + { + "bbox": [ + 77, + 196, + 85, + 205 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 202, + 299, + 318 + ], + "lines": [ + { + "bbox": [ + 86, + 202, + 299, + 318 + ], + "spans": [ + { + "bbox": [ + 86, + 202, + 299, + 318 + ], + "type": "image", + "image_path": "00308f5991f4dab43c77e41957efdf5bc73c95df9dd6dd21cd63bedfec5584bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 202, + 539, + 320 + ], + "blocks": [ + { + "bbox": [ + 313, + 198, + 321, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 321, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 321, + 207 + ], + "type": "text", + "content": "D" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 202, + 539, + 320 + ], + "lines": [ + { + "bbox": [ + 321, + 202, + 539, + 320 + ], + "spans": [ + { + "bbox": [ + 321, + 202, + 539, + 320 + ], + "type": "image", + "image_path": "f00e1ab5aa2c1dbc9ec8469fa4c9cc977e74e7b2aaabc0ce703f425165f6f02b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. (A) Schematic of gaze prediction using EEG signals. (Left) The task schematic of the Visual Search and Large Grid datasets from the EEGEyeNet benchmark. (Right) EEG and eye tracking data were recorded [29]. (B) Architecture of the EEGNet-LSTM model, used to decode from EEG data. The parsimonious architecture can be used to decode saccades from EEG data. (C) Schematic of the \"brain age\" prediction task using structural MRI scans. (Left) A sample preprocessed structural MRI scan by SFCN. (right) The age and gender (inset) distribution of the TLSA dataset. (D) Model architecture of the SFCN-based decoder. The SFCN backbone was used to extract features from the MRI scan, which were used by a decoder block to predict brain age." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 434, + 290, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 434, + 290, + 467 + ], + "spans": [ + { + "bbox": [ + 61, + 434, + 290, + 467 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} \\sum_ {i = 1} ^ {N _ {l}} \\log p \\left(y _ {i} ^ {t} \\mid \\mathbf {x} _ {i} ^ {t}, \\theta\\right) + \\alpha \\sum_ {i = 1} ^ {N} \\log q \\left(\\mathbf {x} _ {i} ^ {t}, \\tilde {y} _ {i} ^ {t} \\mid \\theta\\right)", + "image_path": "e701f14dcf81749347209d1c54ac4c03719ff6afb6446f7980ee8d1513c56dc6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": "After incorporating distributional assumptions in the above equation and subsuming constants with respect to model parameters (e.g., priors on the labels " + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "inline_equation", + "content": "p(y^t)" + }, + { + "bbox": [ + 55, + 473, + 296, + 510 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 65, + 520, + 308, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 520, + 308, + 587 + ], + "spans": [ + { + "bbox": [ + 65, + 520, + 308, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {*} = \\underset {\\theta \\in \\Theta} {\\operatorname {a r g m a x}} - \\sum_ {i = 1} ^ {N _ {l}} \\left(y _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} - \\alpha \\left(\\sum_ {i = 1} ^ {N} \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {i} ^ {t}; \\theta\\right)\\right) ^ {2} \\right. \\\\ - \\sum_ {i = 1} ^ {N} \\log \\sum_ {l = 1} ^ {N} \\exp \\left(- \\left(\\tilde {y} _ {i} ^ {t} - f \\left(\\mathbf {x} _ {l} ^ {t}; \\theta\\right)\\right) ^ {2}\\right)) \\tag {6} \\\\ \\end{array}", + "image_path": "3cb1db4710fe1e89213b3b09f7e843355c52204d838f93f3a376db691ac52f80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "text", + "content": "In practice, each update is performed with a batch of labeled and unlabeled data. First, pseudo-labels are computed for the unlabeled data. Then, keeping the pseudo-labels fixed, the second and third terms of this objective are optimized, while the labeled data are used to optimize the first term, concurrently. Note that the optimization is performed with the model parameters initialized to " + }, + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\theta^s" + }, + { + "bbox": [ + 55, + 594, + 295, + 688 + ], + "type": "text", + "content": ", derived from the source model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Intuitively, the supervised loss in Equation 6 encourages model predictions to align with their true labels; because" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 538 + ], + "type": "text", + "content": "model parameters were initialized with the source model, this acts as an implicit regularizer discouraging the adapted model from deviating too far from the source model. The second term forces dissimilar predictions for data points with different pseudo-labels. In other words, providing samples with dissimilar label values during training produces disparate representations for data points with distant labels, which in turn helps to learn a better regression line. We use the log-sum-exponent trick to avoid numerical instability during optimizing of the second term." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 547, + 503, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 503, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 503, + 561 + ], + "type": "text", + "content": "4. Datasets and Competing Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 602 + ], + "type": "text", + "content": "We employ CRAFT to address challenges relating to data scarcity, missing labels, and continuous label prediction in neuroscience." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 609, + 433, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 433, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 433, + 620 + ], + "type": "text", + "content": "4.1. Datasets and models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 627, + 488, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 488, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 488, + 639 + ], + "type": "text", + "content": "4.1.1. Gaze prediction from EEG signals" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Saccades are rapid ballistic eye movements, putatively linked to higher-order cognitive functions like attention and working memory [5, 13]. Decoding eye movements from neural data could help neuroscientists develop better models of these complex processes. With this objective, the EEGeyenet benchmark dataset [29] was collected from 356" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "831" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "participants. It contains " + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "inline_equation", + "content": "\\sim 47.5" + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": " hours of 128-channel electroencephalogram (EEG) recordings with eye-tracking data collected concurrently." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 203 + ], + "type": "text", + "content": "Here, we predict saccade amplitudes from these EEG data. We consider two datasets - the \"Large Grid\" and the \"Visual Search\" datasets - from the EEGEyeNet benchmark [29] (Fig. 2A). In the Large Grid paradigm, participants made saccades to a target location cued by a dot on the screen. In the Visual Search paradigm, participants searched for a specific symbol in the display and, thus, made saccades during the search." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "text", + "content": "Preprocessing. Each dataset was preprocessed using the pipeline described in the benchmark (details in Appendix B.1). We decode saccade magnitudes " + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "inline_equation", + "content": "(y \\in \\mathbb{R}^{+})" + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "text", + "content": " using the 128-channel EEG timeseries, sampled at " + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "inline_equation", + "content": "500\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "inline_equation", + "content": "(\\mathbf{x} \\in \\mathbb{R}^{128\\times 500})" + }, + { + "bbox": [ + 55, + 204, + 294, + 323 + ], + "type": "text", + "content": ", for corresponding trials. Table 5 (see Appendix B.1) lists the details of the train, cross-validation and test distributions for these datasets. The Visual Search paradigm has more training data and was used to train the source model; this was then transferred to predict saccade amplitudes in the Large Grid dataset (target)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 294, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 294, + 419 + ], + "type": "text", + "content": "Models. The EEGEyeNet benchmark showed that Pyramidal CNNs [47] were most effective at predicting saccade amplitude. As an alternative to 2D convolutional networks, we also evaluate the performance of a novel end-to-end Long-Short-Term-Memory (LSTM) model. This model is trained on features extracted by separable convolutions along time and electrode dimensions, following EEGNet [31] (Fig. 2B, more details in the Appendix B.1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 253, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 253, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 253, + 435 + ], + "type": "text", + "content": "4.1.2. \"Brain age\" prediction from MRI scans" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 294, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 294, + 498 + ], + "type": "text", + "content": "\"Brain age\" is the biological age of the brain and is an important biomarker for cognitive health. Studies have shown that accelerated brain aging is associated with serious cognitive impairments like Alzheimer's disease [9]. Hence, estimating brain age is a critical problem in neuro-medicine." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": "As a first step, attempts have been made to decode the chronological age of healthy participants using T1-weighted structural MRI scans [41]. A 3D-CNN model (see next, Models) was trained using the large UK-Biobank dataset " + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "inline_equation", + "content": "(N = 14,503,44 - 80" + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": " years, mean age " + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "inline_equation", + "content": "= 52.7" + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": " years) [39]. We seek to transfer this pretrained model to decode age in a small Indian cohort – the TATA Longitudinal Study of Aging (TLSA) " + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "inline_equation", + "content": "(N = 188,46 - 88" + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": " years, mean age " + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "inline_equation", + "content": "= 66" + }, + { + "bbox": [ + 55, + 498, + 294, + 605 + ], + "type": "text", + "content": " years) [45] (Fig. 2C), without access to the UKBiobank data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 665 + ], + "type": "text", + "content": "Preprocessing. We followed a preprocessing pipeline identical to [41]. The T1w-MRI scan was brain-extracted, bias-corrected [46], and registered to standard MNI-152 space [38], and cropped to get preprocessed scans " + }, + { + "bbox": [ + 55, + 605, + 294, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{160 \\times 192 \\times 160}" + }, + { + "bbox": [ + 55, + 605, + 294, + 665 + ], + "type": "text", + "content": ", which was used to predict brain age." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 294, + 713 + ], + "type": "text", + "content": "Models. Features from the pretrained Simple Fully Convolutional Network (SFCN, Fig. 2D) were used to make the predictions with a regression model; more details are provided in Appendix B.2) [41]. The entire model was then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "type": "text", + "content": "adapted to the target dataset (TLSA)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "type": "text", + "content": "We also apply CRAFT to people counting and tumor prediction tasks (see Appendix B.3, B.4 for dataset details)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 117, + 466, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 466, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 466, + 129 + ], + "type": "text", + "content": "4.2. Competing DA Approaches" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 135, + 553, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 170 + ], + "type": "text", + "content": "Baseline model. We define a naive baseline model that always predicts the mean label of the training set, ignoring the input features of the test samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 171, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 206 + ], + "type": "text", + "content": "Transfer Learning (TL). We naively adapt the source model by minimizing the supervised loss on the labeled samples in the target dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 207, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 302 + ], + "type": "text", + "content": "TASFAR. This is an SF-UDA model (See Section 2) for regression problems. It depends on two statistics from the source dataset – the threshold for confident samples and the mapping between uncertainty to prediction errors. Because the source is unavailable, we compute these using the TL model on the target. For a fair comparison, we add the supervised loss on the labeled data to the unsupervised loss, weighted by the same " + }, + { + "bbox": [ + 313, + 207, + 553, + 302 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 207, + 553, + 302 + ], + "type": "text", + "content": " as in our model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 304, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 398 + ], + "type": "text", + "content": "DataFree. This is another SF-UDA model (See Section 2). The original model depends on the distribution of intermediate latents on the source distribution. Again, because the source data is unavailable, we computed latent distribution using the TL model and the labeled target data. As before, the objective was modified to accommodate a supervised loss (same " + }, + { + "bbox": [ + 313, + 304, + 553, + 398 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 304, + 553, + 398 + ], + "type": "text", + "content": " as in our model). Bin sizes for both TASFAR and DataFree were selected to be the same as in CRAFT." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 399, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 459 + ], + "type": "text", + "content": "Progressive Mixup. This is a semi-supervised, source-free domain adaptation tool (See Section 2). We adapt the model to the regression task by replacing cross-entropy losses with mean-squared error loss. Predictive entropy was replaced by variance to estimate uncertainty." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 518 + ], + "type": "text", + "content": "BBCN. Bilateral Cycle Consistency is a prototype-based clustering SF-SSDA paradigm (See Section 2). It depends on class-level prototypes for each model. To extend it to regression, we discretize the output space into bins of the same width as the other methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 520, + 553, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 553, + 603 + ], + "type": "text", + "content": "Metrics. We quantify the prediction performance of continuous valued outputs using two metrics - Root Mean Squared Error (RMSE) and Percentage-Bend Correlation [49] coefficient (R). A lower RMSE reflects a closer correspondence between the predicted and actual labels, a higher R-value quantifies better predictions along the best-fit line between the actual and predicted labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 604, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 553, + 640 + ], + "type": "text", + "content": "In subsequent sections (5.1 and 5.2), we predict saccade amplitude from EEG and brain age from structural MRI scans, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 651, + 440, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 440, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 440, + 664 + ], + "type": "text", + "content": "5. Experiments: Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 479, + 684 + ], + "type": "text", + "content": "5.1. Saccade Amplitude Prediction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Firstly, we trained an EEGNet-LSTM model on the Visual Search source dataset (RMSE=67.03, R=0.91). This novel" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "832" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "architecture outperformed the benchmark model (Pyramidal CNNs) by " + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\sim 16\\%" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": " (benchmark: RMSE=79.59 pix, R=0.87). Next, this was transferred to the Large Grid dataset. The upper 3 rows of Table 1 shows the performance when all the data was used for funetuning. In this case also EEGNet-LSTM outperformed the benchmark (Pyramidal CNN [29]) by a large margin (" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\sim 19\\%" + }, + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "). Hence, subsequent SSDA experiments were performed using EEGNet-LSTMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": "All models were trained with mini-batches of 128 samples, and only the best-performing checkpoint on the held-out validation set was used for evaluating the test set. Adam optimizer with an initial learning rate of " + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": " was used. " + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": " in equation 3 was selected with a grid saerch over " + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.1, 1.0\\}" + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": "; empirically, " + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": " was always favored." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 323, + 296, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 249, + 294, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 294, + 315 + ], + "type": "text", + "content": "Table 1. Source Free Semi-supervised Deep Transfer of the model trained on the Visual Search task to the Large-Grid Task. The upper part of the table (rows 1-3) shows the performance ceiling when all labeled data in the target is used for training. The lower part (rows 4-9) shows SF-SSDA results when only " + }, + { + "bbox": [ + 55, + 249, + 294, + 315 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 249, + 294, + 315 + ], + "type": "text", + "content": " of the target samples were labeled (3 seeds)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 323, + 296, + 453 + ], + "lines": [ + { + "bbox": [ + 56, + 323, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 296, + 453 + ], + "type": "table", + "html": "
MethodR ↑RMSE (in pix) ↓
Naive Baseline-149.12 ± 0.02
Benchmark (TL, 100%)0.91 ± 0.0163.84 ± 0.75
EEGNet-LSTM (TL, 100%)0.93 ± 0.0151.47 ± 0.63
SF-SSDA (1% labels)
EEGNet-LSTM + TL0.77 ± 0.0192.26 ± 1.66
EEGNet-LSTM + Mixup0.48 ± 0.02135.70 ± 1.25
EEGNet-LSTM + BBCN0.76 ± 0.0199.8 ± 3.35
EEGNet-LSTM + TASFAR0.76 ± 0.0186.41 ± 1.05
EEGNet-LSTM + DataFree0.80 ± 0.0187.64 ± 3.08
EEGNet-LSTM + CRAFT0.81 ± 0.0284.17 ± 3.95
", + "image_path": "0af225e9a794ad1aed3fadb278fc0503c3b1a131b5d671546b8869d0c8be8617.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": "To demonstrate the efficacy of CRAFT for SF-SSDA, target labels from the training data of the Large Grid dataset were randomly dropped in a stratified manner. Figure 3A shows how the performance of these models is affected by reducing the fraction of labeled data (" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "inline_equation", + "content": "n_{ul} / N" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": " is the fraction of unlabeled samples). A lower proportion of labeled data worsens the performance of all the models (RMSE increases), but CRAFT produces the lowest RMSE for all unlabeled data proportions. Figure 3B shows the prediction of saccade magnitudes (" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "inline_equation", + "content": "|\\Delta \\overline{r}|" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": ") for the best CRAFT-based EEGNet-LSTM model. Table 1 (lower 6 rows) summarizes the performance of CRAFT when " + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": " of the data were unlabeled; CRAFT outperforms supervised training (TL) by " + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": " and other SOTA SF-SSDA models by " + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "inline_equation", + "content": ">4\\%" + }, + { + "bbox": [ + 56, + 463, + 295, + 666 + ], + "type": "text", + "content": " in terms of RMSE. The closest competitive methods were TASFAR and DataFree. The correlation coefficient (R) also showed similar trends (Table 1; Fig. 6A)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "type": "text", + "content": "5.2. Brain Age Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "Next, we transferred the SFCN model trained on the UK Biobank MRI scans to the TLSA dataset. The TLSA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "dataset has only 188 MRI scans, so we avoid hyperparameter searches. All models were trained for 25 epochs (batch size = 4) using the Adam optimizer (initial learning rate = 10" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "^{-4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "), and only the final model was used for inference. Here, we used " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "; we explore the effect of varying " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " for ours and competing models in Appendix D. Each metric was calculated for the prediction on the entire dataset, leaving out one fold at a time (4-folds). The top of Table 2 (row 2) quantifies the performance ceiling when all the training data are used (3 seeds)." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 254, + 553, + 379 + ], + "blocks": [ + { + "bbox": [ + 313, + 201, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 246 + ], + "type": "text", + "content": "Table 2. Semi-supervised transfer of the SFCN model (source: UKB) to the TLSA dataset. Same as in Table 1, except that the lower part of the table (3-8) shows SF-SSDA results when only " + }, + { + "bbox": [ + 313, + 201, + 553, + 246 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 201, + 553, + 246 + ], + "type": "text", + "content": " of the target samples were labeled. (3 seeds)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 254, + 553, + 379 + ], + "lines": [ + { + "bbox": [ + 315, + 254, + 553, + 379 + ], + "spans": [ + { + "bbox": [ + 315, + 254, + 553, + 379 + ], + "type": "table", + "html": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-7.91 ± 0.05
SFCN + TL (100%)0.66 ± 0.016.14 ± 0.03
SF-SSDA (20% labels)
SFCN + TL0.41 ± 0.077.41 ± 0.21
SFCN + Mixup0.34 ± 0.047.71 ± 0.14
SFCN + BBCN0.28 ± 0.048.00 ± 0.15
SFCN + TASFAR0.42 ± 0.077.47 ± 0.15
SFCN + DataFree0.50 ± 0.037.36 ± 0.14
SFCN + CRAFT0.51 ± 0.037.14 ± 0.11
", + "image_path": "d195487b1e970714d34153890ec0306bc472ae82f7dd5e64bea2a2729d3acf3c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": "For SF-SSDA, we performed experiments closely similar to the saccade task (Section 5.1). Figure 3C shows how these models' performance is affected when the proportion of unlabeled samples increases. Reducing the fraction of labeled data worsens model performance, but CRAFT continues to remain the most competitive of all models. Figure 3D shows the prediction scatter of the best CRAFT-based model. Table 2 summarizes the performance when " + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": " of the data was unlabeled. CRAFT outperforms supervised training (TL) by " + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\sim 4\\%" + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": " and other SOTA SF-SSDA models by " + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "inline_equation", + "content": ">3\\%" + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": " in terms of RMSE. Similar results were obtained with the R values also (Fig. 6B). Again, TASFAR and DataFree emerge as the most competitive methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": "Finally, we evaluated the effectiveness of CRAFT at two other real-world regression benchmarks: People Counting [44, 48] and Tumor Size prediction [6, 16]. CRAFT outperformed other SF-SSDA methods by " + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": ">5\\%" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": ">2\\%" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": " respectively for each dataset; the details are presented in Appendices C.1 and C.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Mitigating Sampling Biases. Next, we show how the unsupervised CRAFT objective could mitigate sampling bias effects in the target training set. In all of the previous simulations, the label distributions of the target train and test data were matched. Here, we deliberately distorted the label distribution of the target training data: " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " of the data above the mean age in the training set were removed, creating a heavy bias toward lower ages. Following this, labels were" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "833" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 292, + 251 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 292, + 251 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 292, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 292, + 251 + ], + "type": "image", + "image_path": "13cc5ff798edf1da9859ff5eaa5b0d5d8a8f76823e76d7401ba5b67b7dfc7c2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": "Figure 3. (A, C) RMSE " + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": " for source-free semi-supervised domain adaptation of saccade (A) and brain age (C) prediction tasks, trained with varying proportions of unlabeled target data " + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "n_{ul} / N" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": ". Dashed black line: Performance ceiling. Circles: individual trials. (B, D) Prediction for the best CRAFT model for saccade amplitude (B) and brain age (D), with " + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": " unlabeled data, respectively. Triangles: individual participants. Results with a much higher proportion of unlabeled data are shown for the saccade dataset because of the far greater size of this (" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "N \\sim 12k" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": ") compared to the brain age (" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "inline_equation", + "content": "N \\sim 180" + }, + { + "bbox": [ + 55, + 271, + 298, + 383 + ], + "type": "text", + "content": ") dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "content": "retained in only " + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "content": " of the training data " + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "inline_equation", + "content": "(n_{ul} / N = 0.6)" + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "content": ". No such biases were introduced into the test data. The CRAFT model was then trained by imposing the unbiased (true) marginal label distribution on the unlabeled training set (equation 5). In this scenario, again, we find that CRAFT mitigates the effect of sampling biases in the training data (Table 3). It outperforms transfer learning by " + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\sim 5\\%" + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "content": ", and other sota SF-SSDA models by " + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "inline_equation", + "content": ">2.5\\%" + }, + { + "bbox": [ + 55, + 414, + 296, + 510 + ], + "type": "text", + "content": " (RMSE)." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 583, + 296, + 695 + ], + "blocks": [ + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "type": "text", + "content": "Table 3. Analyses with datasets with sampling error. " + }, + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "type": "text", + "content": " of the data above the mean age is removed. Furthermore, only " + }, + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 531, + 296, + 575 + ], + "type": "text", + "content": " of the labels in this mislabeled data are used for semi-supervised transfer of the SFCN model (source: UKB). (3 seeds)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 583, + 296, + 695 + ], + "lines": [ + { + "bbox": [ + 56, + 583, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 56, + 583, + 296, + 695 + ], + "type": "table", + "html": "
MethodR ↑RMSE (in years) ↓
Naive Baseline-8.38 ± 0.22
SFCN + TL0.41 ± 0.058.40 ± 0.27
SFCN + Mixup0.32 ± 0.028.19 ± 0.15
SFCN + BBCN0.23 ± 0.028.53 ± 0.16
SFCN + TASFAR0.39 ± 0.048.57 ± 0.22
SFCN + DataFree0.41 ± 0.038.58 ± 0.23
SFCN + CRAFT0.45 ± 0.037.98 ± 0.25
", + "image_path": "176349e5e85828d9434ac21ddb3b5fc2bdbae599b628802a2f56a885a8a4ed7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 513, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 513, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 513, + 85 + ], + "type": "text", + "content": "5.3. Analysis of computational complexity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "content": "Table 4 shows the computational complexity of CRAFT and the competing models. We compute both the algorithms' theoretical complexity and running time on the saccade and brain age prediction tasks. The theoretical training complexity is defined as the number of gradient updates per epoch. For inference, the complexity is the same for all models as the same architecture is trained using different adaptation paradigms. Training time and complexity are lowest for TASFAR, amongst all other SF-DA tools. The complexity of CRAFT was comparable to DataFree, the next best method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 281, + 553, + 358 + ], + "blocks": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "type": "text", + "content": "Table 4. Computational complexity. Time per training epoch and for all test data (in min). Parentheses: P- no. of model parameters; N- samples; D- latent dimensionality; " + }, + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "type": "text", + "content": "- batch size, B- no. of bins, m- no. of inferences for MC Dropout uncertainty estimation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 281, + 553, + 358 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 553, + 358 + ], + "type": "table", + "html": "
(B) MethodEEGNet-LSTM (P~ 190K)SFCN (P~ 2.95M)
Train /Comp. ↓ (N ~ 12k)Test /Comp. ↓ (N ~ 2.5k)Train/Comp. ↓ (N ~ 150)Test/Comp. ↓ (N ~ 50)
TL0.15 / O(N)0.02 / O(N)0.12 / O(N)0.04 / O(N)
Mixup0.74 / O(Nm)0.02 / O(N)0.74 / O(Nm)0.04 / O(N)
BBCN2.71 / O(NBD)0.02 / O(N)1.29 / O(NBD)0.04 / O(N)
TASFAR0.30 / O(N)0.02 / O(N)0.19 / O(N)0.04 / O(N)
DataFree0.55 / O(ND2)0.02 / O(N)0.32 / O(ND2)0.04 / O(N)
CRAFT0.45 / O(N(β + B))0.02 / O(N)0.36 / O(N(β + B))0.04 / O(N)
", + "image_path": "7daf5f58a8e81844ba2fb3594d433f31abd53c81c9557d05107f68e3b9934c93.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 382, + 440, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 440, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 440, + 396 + ], + "type": "text", + "content": "6. Concluding Remarks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 582 + ], + "type": "text", + "content": "Data scarcity poses a serious challenge for applying deep-learning models in real-world settings. Training overparameterized models on a few hundred data points leads to overfitting and poor generalization. Conventional domain adaptation algorithms, to transfer pretrained networks, typically need access to source domain data, but these may not always be available. In this study, we propose a source-free semi-supervised domain adaptation framework that recalibrates the decision boundary without aligning the intermediate representation of the domains. In fields like medicine, resource constraints associated with labeling data are significant; therefore, leveraging the large amounts of unlabeled data becomes critical. We propose CRAFT as an SF-SSDA approach for such source-free, label-sparse settings that are common in real-world applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 592, + 429, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 429, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 429, + 605 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 556, + 696 + ], + "type": "text", + "content": "This work was supported by the following funding sources: a Prime Minister's Research Fellowship (to M.B.), a DST SwarnaJayanti fellowship, a Pratiksha Trust Intramural grant, a Tata Trusts grant and a Google Research grant (to D.S.). We also thank the Tata Longitudinal Study on Aging at the Indian Institute of Science for access to MRI scans from their cohort." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "834" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jiayu An, Changming Zhao, and Dongrui Wu. Semi-supervised generalized source-free domain adaptation (ssgsfda). In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 137, + 284, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 284, + 148 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 284, + 148 + ], + "type": "text", + "content": "[2] APTOS. Aptos 2019 blindness detection dataset, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 149, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 149, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 294, + 192 + ], + "type": "text", + "content": "[3] Sourabh Balgi and Ambedkar Dukkipati. CUDA: Contradistinguisher for unsupervised domain adaptation. In Proceedings of the IEEE International Conference on Data Mining, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 239 + ], + "type": "text", + "content": "[4] Sourabh Balgi and Ambedkar Dukkipati. Contradistinginguisher: A Vapnik's imperative to unsupervised domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4730-4747, 2022. 1, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 295, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 295, + 283 + ], + "type": "text", + "content": "[5] Scott L. Brincat, Jacob A. Donoghue, Meredith K. Kahnke, Simon Kornblith, Mikael Lundqvist, and Earl K. Miller. Interhemispheric transfer of working memories. Neuron, 109: 1055-1066.e4, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 286, + 295, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 450 + ], + "type": "text", + "content": "[6] Peter Bandy, Oscar Geessink, Quirine Manson, Marcory Van Dijk, Maschenka Balkenhol, Meyke Hermsen, Babak Ehteshami Bejnordi, Byungjae Lee, Kyunghyun Paeng, Aoxiao Zhong, Quanzheng Li, Farhad Ghazvinian Zanjani, Svitlana Zinger, Keisuke Fukuta, Daisuke Komura, Vlado Ovtcharov, Shenghua Cheng, Shaoqun Zeng, Jeppe Thagaard, Anders B. Dahl, Huangjing Lin, Hao Chen, Ludwig Jacobsson, Martin Hedlund, Melih Çetin, Eren Halici, Hunter Jackson, Richard Chen, Fabian Both, Jörg Franke, Heidi Küsters-Vandevelde, Willem Vreuls, Peter Bult, Bram van Ginneken, Jeroen van der Laak, and Geert Litjens. From detection of individual metastases to classification of lymph node status at the patient level: The camelyon17 challenge. IEEE Transactions on Medical Imaging, 38(2):550-560, 2019. 2, 7, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 453, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 495 + ], + "type": "text", + "content": "[7] Wenzhi Cao, Vahid Mirjalili, and Sebastian Raschka. Rank consistent ordinal regression for neural networks with application to age estimation. Pattern Recognition Letters, 140: 325-331, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "text", + "content": "[8] Thomas A. Carlson. Orientation decoding in human visual cortex: New insights from an unbiased perspective. Journal of Neuroscience, 34(24):8373-8383, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "type": "text", + "content": "[9] J H Cole, S J Ritchie, M E Bastin, M C Valdés Hernández, S Muñoz Maniega, N Royle, J Corley, A Pattie, S E Harris, Q Zhang, N R Wray, P Redmond, R E Marioni, J M Starr, S R Cox, J M Wardlaw, D J Sharp, and I J Deary. Brain age predicts mortality. Molecular Psychiatry, 23:1385-1392, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "text", + "content": "[10] Sébastien M. Crouzet. Fast saccades toward faces: Face detection in just 100 ms. Journal of Vision, 10:1-17, 2010. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "text", + "content": "[11] Jorge Cuadros and George Bresnick. Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening. Journal of Diabetes Science and Technology, 3(3):509-516, 2009. PMID: 20144289. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[13] Heiner Deubel and Werner X. Schneider. Saccade target selection and object recognition: Evidence for a common attentional mechanism. Vision Research, 36:1827-1837, 1996. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[14] Yuhe Ding, Lijun Sheng, Jian Liang, Aihua Zheng, and Ran He. Proxymix: Proxy-based mixup training with label refinery for source-free domain adaptation. Neural Networks, 167:92-103, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "text", + "content": "[15] Cian Eastwood, Ian Mason, Christopher K I Williams, and Bernhard Scholkopf. In Proceedings of the Tenth International Conference on Learning Representations, 2022. Tenth International Conference on Learning Representations 2022, ICLR 2022; Conference date: 25-04-2022 Through 29-04-2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 230, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 306 + ], + "type": "text", + "content": "[16] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes van Diest, Bram van Ginneken, Nico Karssemeijer, Geert Litjens, Jeroen A. W. M. van der Laak, , and the CAMELYON16 Consortium. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. JAMA, 318(22):2199-2210, 2017. 2, 7, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "type": "text", + "content": "[17] Saminger-Platz et al. Central moment discrepancy (cmd) for domain-invariant representation learning. *ICLR*, 2017. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[18] Yuqi Fang, Pew-Thian Yap, Weili Lin, Hongtu Zhu, and Mingxia Liu. Source-free unsupervised domain adaptation: A survey. Neural Networks, 174:106230, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 419 + ], + "type": "text", + "content": "[19] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of Machine Learning Research, 17:1-35, 2016. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 475 + ], + "type": "text", + "content": "[20] Alessandro T. Gifford, Domenic Bersch, Marie St-Laurent, Basile Pinsard, Julie Boyle, Lune Bellec, Aude Oliva, Gemma Roig, and Radoslaw M. Cichy. The algonauts project 2025 challenge: How the human brain makes sense of multimodal movies, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "text", + "content": "[21] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Advances in Neural Information Processing Systems. MIT Press, 2004. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[22] Arthur Gretton, Kenji Fukumizu, Zäïd Harchaoui, and Bharath K. Sriperumbudur. A fast, consistent kernel two-sample test. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 589 + ], + "type": "text", + "content": "[23] Yan Hao, Yuhong Guo, and Chunsheng Yang. Source-free unsupervised domain adaptation with surrogate data generation. In British Machine Vision Conference, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 634 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In Computer Vision - ECCV 2016, pages 630-645, Cham, 2016. Springer International Publishing. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 679 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Tianlang He, Zhiqiu Xia, Jierun Chen, Haoliang Li, and Shueng-Han Gary Chan. Target-agnostic source-free domain adaptation for regression tasks. 2024 IEEE 40th In" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "835" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ternational Conference on Data Engineering (ICDE), pages 1464-1477, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "text", + "content": "[27] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "text", + "content": "[28] Xinyang Huang, Chuang Zhu, Bowen Zhang, and Shang-hang Zhang. Learning from different samples: A source-free framework for semi-supervised domain adaptation, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 240 + ], + "type": "text", + "content": "[29] Ard Kastrati, Martyna Plomecka, Damian Pascual Ortiz, Lukas Wolf, Victor Gillioz, Roger Wattenhofer, and Nicolas Langer. Eegeyenet: a simultaneous electroencephalography and eye-tracking dataset and benchmark for eye movement prediction. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "type": "text", + "content": "[30] Vinod K Kurmi, Venkatesh K Subramanian, and Vinay P Namboodiri. Domain impression: A source data free domain adaptation method, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "text", + "content": "[31] Vernon J Lawhern, Amelia J Solon, Nicholas R Waytowich, Stephen M Gordon, Chou P Hung, and Brent J Lance. Eeg-net: a compact convolutional neural network for eeg-based brain-computer interfaces. Journal of Neural Engineering, 15:056013, 2018. 6, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "text", + "content": "[32] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A Comprehensive Survey on Source-Free Domain Adaptation. IEEE Transactions on Pattern Analysis & Machine Intelligence, 46(08):5743-5762, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "text", + "content": "[33] Jian Liang, Dapeng Hu, Yunbo Wang, Ran He, and Jiashi Feng. Source data-absent unsupervised domain adaptation through hypothesis transfer and labeling transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8602-8617, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "text", + "content": "[34] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In Proceedings of the 32nd International Conference on Machine Learning, pages 97-105, Lille, France, 2015. PMLR. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "text", + "content": "[35] Jie Ma, Zhijun Xu, Xinyu Xiong, MaYiDiLi NiJiaTi, Dongyu Zhang, Weijun Sun, Feng Gao, and Guanbin Li. Source-free semi-supervised domain adaptation for tuberculosis recognition. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1-5, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[36] Ning Ma, Jiajun Bu, Lixian Lu, Jun Wen, Zhen Zhang, Sheng Zhou, and Xifeng Yan. Semi-supervised hypothesis transfer for source-free domain adaptation, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "text", + "content": "[37] Ning Ma, Haishuai Wang, Zhen Zhang, Sheng Zhou, Hongyang Chen, and Jiajun Bu. Source-free semi-supervised domain adaptation via progressive mixup. Knowledge-Based Systems, 262:110208, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[38] Pravat K. Mandal, Rashima Mahajan, and Ivo D. Dinov. Structural brain atlases: Design, rationale, and applications in normal and pathological cohorts. Journal of Alzheimer's Disease, 31:S169-S188, 2012. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[39] Karla L Miller, Fidel Alfaro-Almagro, Neal K Bangerter, David L Thomas, Essa Yacoub, Junqian Xu, Andreas J Bartsch, Saad Jbabdi, Stamatios N Sotiropoulos, Jesper L R Andersson, Ludovica Griffanti, Gwenaëlle Douaud," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 700 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 128 + ], + "type": "text", + "content": "Thomas W Okell, Peter Weale, Iulius Dragonu, Steve Garratt, Sarah Hudson, Rory Collins, Mark Jenkinson, Paul M Matthews, and Stephen M Smith. Multimodal population brain imaging in the uk biobank prospective epidemiological study. Nature Neuroscience, 19:1523-1536, 2016. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[40] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[41] Han Peng, Weikang Gong, Christian F Beckmann, Andrea Vedaldi, and Stephen M Smith. Accurate brain age prediction with lightweight deep neural networks. Medical Image Analysis, 68:101871, 2021. 6, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[42] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In Computer Vision - ECCV 2022, pages 165-182, Cham, 2022. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[43] Xiaofeng Qu, Li Liu, Lei Zhu, Liqiang Nie, and Huaxiang Zhang. Source-free style-diversity adversarial domain adaptation with privacy-preservation for person re-identification. Knowledge-Based Systems, 283:111150, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "text", + "content": "[44] Vishwanath Sindagi, Rajeev Yasarla, and Vishal Patel. Pushing the Frontiers of Unconstrained Crowd Counting: New Dataset and Benchmark Method. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1221-1231, Los Alamitos, CA, USA, 2019. IEEE Computer Society. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 430 + ], + "type": "text", + "content": "[45] Jonas S Sundarakumar, Albert Stezin, Abhishek L Menes-gere, and Vijayalakshmi Ravindranath. Rural-urban and gender differences in metabolic syndrome in the aging population from southern india: Two parallel, prospective cohort studies. eClinicalMedicine, 47:101395, 2022. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 495 + ], + "type": "text", + "content": "[46] J-Donald Tournier, Robert Smith, David Raffelt, Rami Tabbara, Thijs Dhollander, Maximilian Pietsch, Daan Christi-aens, Ben Jeurissen, Chun-Hung Yeh, and Alan Connelly. Mrtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202:116137, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 518 + ], + "type": "text", + "content": "[47] Ihsan Ullah and Alfredo Petrosino. About pyramid structure in convolutional neural networks, 2016. 6, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "text", + "content": "[48] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. Nwpu-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 566, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 587 + ], + "type": "text", + "content": "[49] Rand R. Wilcox. The percentage bend correlation coefficient. Psychometrika, 59:601-616, 1994. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 620 + ], + "type": "text", + "content": "[50] Garrett Wilson and Diane J. Cook. A survey of unsupervised deep domain adaptation. ACM Trans. Intell. Syst. Technol., 11(5), 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "type": "text", + "content": "[51] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36(8):3746-3766, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 667, + 553, + 700 + ], + "type": "text", + "content": "[52] Shiqi Yang, Yaxing Wang, Kai Wang, Shangling Jui, and Joost van de Weijer. Attracting and dispersing: A simple approach for source-free domain adaptation, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "836" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_content_list.json b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a8f5c055bf9c2c155e47ebd03c9e50295288615d --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_content_list.json @@ -0,0 +1,1377 @@ +[ + { + "type": "text", + "text": "SemiVisBooster: Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjin Zhang $^{1,2}$ Xinyu Li $^{2}$ Chenyang Gao $^{1,2}$ Ivan Marsic $^{1}$ $^{1}$ Rutgers University ${}^{2}$ Amazon", + "bbox": [ + 236, + 202, + 756, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wz315, cg694}@scarletmail.rutgers.edu, lixinyu.arthur@outlook.com, marsic@rutgers.edu", + "bbox": [ + 109, + 241, + 880, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models rely on large-scale labeled datasets, but collecting such data is expensive and time-consuming. Semi-supervised learning (SSL) mitigates this issue by learning from a small set of labeled samples along with a large pool of unlabeled data. However, existing SSL methods struggle with fine-grained classification when dealing with visually similar classes, as they rely solely on visual features and ignore the semantic information within label names. This paper introduces SemiVisBooster, an SSL enhancement approach that utilizes semantic information from label names to guide visual feature learning, addressing the challenges of fine-grained classification. By aligning text embeddings from label names with visual features, our method helps the model capture subtle visual distinctions that purely visual representations may overlook. To enhance robustness, we propose two key components: (1) text embedding de-similarity (TEDS) to reduce confusion caused by similar text embeddings across different class names, and (2) class-aware visual-text alignment loss to accurately define positive and negative pairs during visual-text alignment. Our method achieves state-of-the-art performance on the latest SSL benchmarks. Additionally, on the challenging Food-101 dataset, which contains many visually similar classes and uses only 404 labeled images, our approach improves performance by approximately $13.6\\%$ over the second-best method. Code is available at SemiVis-Booster", + "bbox": [ + 88, + 325, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 767, + 220, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The superior performance of deep learning heavily relies on supervised training with sufficient labeled data [14, 39]. However, acquiring such massive labeled datasets for real-world applications remains a labor-intensive and expensive process. To mitigate this challenge, semi-supervised learning (SSL) has emerged as a promising method, learning from a small set of labeled data while leveraging a", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ce7554e4fc95fdf41ba04d72aa7921b42de1de71ac5fc46cb85fb88b168328b.jpg", + "image_caption": [ + "(a) FreeMatch" + ], + "image_footnote": [], + "bbox": [ + 553, + 290, + 705, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc6f4f95037ebd0f4ffb1d655c414e1141d935ddcbf5946fed93c9e50d553ff0.jpg", + "image_caption": [ + "(b) SemiVisBooster", + "Figure 1. Comparison of visual feature space distribution for (a) FreeMatch and (b) our method on the Food-101 dataset visualized using t-SNE. We only plot the first 20 classes." + ], + "image_footnote": [], + "bbox": [ + 707, + 290, + 864, + 390 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "large amount of unlabeled data [5, 21, 36, 41, 46]. By significantly reducing the dependence on manual annotations, SSL has demonstrated great potential in practical applications. Despite achieving strong performance in general image classification, particularly on datasets with well-separated class boundaries such as CIFAR-10 [19] and STL-10 [6], SSL methods struggle with fine-grained classification, where visually similar categories exist. Analyzing the confusion matrix of the popular SSL method, FreeMatch [41], on the CIFAR-100 dataset with only 200 labeled images [42], we observed that its accuracy plateaued at $78.9\\%$ due to frequent misclassifications among visually similar classes such as \"woman,\" \"man,\" \"boy,\" and \"girl\" [40]. This observation indicates a key limitation of existing SSL approaches: their reliance on visual feature extraction, which often fails to distinguish subtle inter-class differences. To further validate the observed limitation on SSL, we evaluated FreeMatch on the fine-grained Food101 dataset [3] using 404 labeled images because this dataset contains many visually similar classes. FreeMatch achieved an accuracy of only $60.4\\%$ , while a fully supervised training method using all available labeled data reached $98.8\\%$ [1], highlighting a significant performance gap between SSL and fully supervised learning. This result underscores the need for improvements in SSL methods. The t-SNE visualization (Figure 1) illustrates that while FreeMatch exhibits some level of clustering, it struggles to separate visually similar classes, leading to significant feature overlap. This challenge arises from two key weaknesses in existing SSL methods. (1).", + "bbox": [ + 511, + 446, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1195", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SSL models rely primarily on visual patterns for classification. However, in fine-grained settings, where objects share highly similar appearances, these methods struggle to extract meaningful distinctions, resulting in frequent misclassification. (2). SSL methods depend heavily on pseudolabeling, where high-confidence predictions are used to assign labels to unlabeled samples. When dealing with visually ambiguous cases, incorrect pseudo-labels are generated and subsequently reinforce errors, leading to model bias and further performance degradation.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In many real-world applications, class names are available as meaningful words or phrases, offering an opportunity to use the semantic information. However, existing SSL methods completely ignore the rich semantic information embedded in these labels. Instead, prior SSLs rely exclusively on one-hot-encoded categorical labels, which discard potentially useful semantic knowledge. We hypothesize that introducing label semantics into SSL can provide additional guidance for visual representation learning, thereby improving the model's ability to distinguish visually similar samples. For example, while \"boy\" and \"man\" may have highly similar visual representations in CIFAR-100 dataset, their semantic embeddings are distinct and different. By aligning visual features with distinct label semantics, SSL models can be improved in two ways. (1). Aligning text embeddings from ground-truth label names with the visual representations of labeled images can help the model better capture subtle visual differences through text cues. This makes the initial model more stable and robust. (2) Even if pseudo-labels are incorrect, aligning the text embedding of the pseudo-label with the visual representation of the unlabeled image can still be beneficial. If the incorrect pseudo-label's semantic meaning is close to the true label, this alignment may help guide the visual representation closer to the correct cluster. However, aligning text embeddings from label name with visual embeddings for SSL introduces two major challenges. The first challenge arises from the inherent similarity between text embeddings of different class names. When similar text embeddings are used to guide the visual representation learning, it can introduce unwanted confusion in the visual space (Section 3.4). The second challenge relates to many-to-one alignment in our settings. Unlike CLIP [31] using NT-Xent loss [35], where each image-text pair is uniquely defined, our setting often encounters multiple images within a batch that share the same text embedding because multiple images may belong to the same class. Directly applying existing visual-text alignment loss like NT-Xent [31], which assume a pairwise positive match, would mistakenly treat these samples as negatives, leading to incorrect optimization and degraded performance (Section 3.5).", + "bbox": [ + 91, + 246, + 483, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduced SemiVisBooster, an SSL booster that improves SSL performance by incorpo", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rating label semantics into the learning process. To enrich feature learning beyond basic one-hot-encoded ground truth labels or pseudo-labels, we proposed a label semantic enrichment module that converts categorical labels into meaningful label names. To enable robust alignment, we proposed text embedding de-similarity (TEDS) loss to avoid the impact of inherent similarity between text embeddings and developed class-aware visual-text alignment (CAViTA) strategy, which correctly defines positive and negative pairs in cases where multiple images share the same text embedding. Our approach achieves state-of-the-art results on the latest SSL benchmarks [40]. Specifically, SemiVisBooster consistently surpasses the previous method by more than $2.9\\%$ across multiple benchmarks, including CIFAR-100 (200 labeled images) and STL-10 (40 labeled images). For fine-grained classification tasks, our method achieved $73.6\\%$ accuracy on Food 101 (404 labeled images), significantly surpassing the previous state-of-the-art method, which achieved only $64.8\\%$ accuracy. We summarized our contributions below:", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to directly leverage label name semantics to enhance SSL, achieving state-of-the-art performance on the latest SSL benchmarks and fine-grained classification tasks.", + "- We introduced TEDS to mitigate the inherent similarity between text embeddings of different class names, enhancing the robustness of our method across various text encoders, including a simple word embedding model.", + "- We designed CAViTA to accurately distinguish positive and negative visual-text embedding pairs, ensuring robust visual-text alignment." + ], + "bbox": [ + 513, + 393, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 578, + 653, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semi-supervised Learning", + "text_level": 1, + "bbox": [ + 511, + 604, + 751, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) has been an area of extensive research over recent decades. Recent advances in SSL methodologies can be broadly categorized into two primary approaches: consistency regularization [20, 26, 38, 43] and pseudo-labeling [21, 30]. Consistency-based methods apply perturbations either to the input data or the model and enforce constraints to ensure consistency in the features or probability outputs. For example, the II-model [20] introduces dropout to model weights, generating two outputs for consistency regularization. Pseudo-labeling methods, in contrast, use model predictions as pseudo-labels to guide learning on unlabeled data. This approach has gained recent prominence [4, 21, 30, 44]. Combining these approaches, FixMatch [36] introduced a straightforward SSL framework that applies both pseudo-labeling and consistency regularization, but it uses a fixed confidence threshold. Recent work has focused on refining pseudo-labeling quality by introducing dynamic thresholds [45], class-specific thresh-", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1196", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce24737cb60d26b409a7f0bf37af63fb61b4a07c0bb35b9dbc4a32d29a6de109.jpg", + "image_caption": [ + "Figure 2. Overview of SemiVisBooer framework for enhancing SSL through label semantic information. (Only the flow of unlabeled data is shown. Labeled data are used for both supervised learning and aligning ground truth labels with their corresponding images.) Base SSL branch: One-hot pseudo-labels are generated from weakly augmented images and used to guide predictions on strongly augmented versions of the same images. Boosting Branch: A label semantic enrichment module maps each categorical label to its label name and uses a pre-trained text encoder to extract text embeddings without any prompt operation. A learnable projector transforms all these embeddings into a text embedding bank. The text embedding de-similarity (TEDS) loss enhances robustness by minimizing the similarity between text embeddings of different classes. The class-aware visual-text alignment is used to align selected text embeddings with corresponding visual embeddings to guide visual representation learning." + ], + "image_footnote": [], + "bbox": [ + 127, + 80, + 867, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "olds [46], and dynamic class-specific thresholds [5, 41]. However, even with these adaptive thresholding techniques, existing SSL methods struggle with fine-grained classification as we discussed in Section 1. To boost SSLs, we introduced additional supervision derived from the semantic information in label names. We used FreeMatch[41] as our base SSL.", + "bbox": [ + 88, + 532, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision and Language Alignment", + "text_level": 1, + "bbox": [ + 89, + 651, + 375, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerous studies have demonstrated the benefits of joint image-text representation learning for various vision tasks [9, 13, 16, 23, 25]. Contrastive learning has proven particularly effective for vision-language models by mapping images and text into a shared feature space. Models like CLIP [31], CLOOB [12], ALIGN [17], and DeCLIP [24] achieve this by training text and image encoders on large datasets of image-caption pairs, using contrastive loss to minimize distances between matching pairs while maximizing them for non-matching ones. To handle noise in datasets, ALIGN [16] and DeCLIP [24] use custom metrics, while CLIP [31] employs cosine distance. While these models primarily aim at pre-training large vision-language foundations for zero- and few-shot tasks, our approach enhances semi-supervised learning by aligning visual features", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with pseudo-labels, improving the robustness of SSLs for fine-grained classification.", + "bbox": [ + 511, + 532, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 577, + 604, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 602, + 653, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [36], we formulate the SSL framework for a $C$ -class classification problem. Let $\\mathcal{D}^l = \\{(x_i^l, y_i^l) : i \\in [1, B_L]\\}$ denote a batch of labeled data, and $\\mathcal{D}^u = \\{x_j^u : j \\in [1, B_U]\\}$ represent a batch of unlabeled data, with batch sizes $B_L$ and $B_U$ , respectively. Here, $x_i^l$ denotes a labeled image with label $y_i^l$ , and $x_j^u$ represents an unlabeled image. The model is trained using a combined objective: $\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_s + \\lambda_u \\mathcal{L}_u$ , where $\\mathcal{L}_s$ is the supervised loss over labeled data, $\\mathcal{L}_u$ is the unsupervised loss over unlabeled data, and $\\lambda_u$ is the unsupervised loss weight. The supervised loss $\\mathcal{L}_s$ is defined as:", + "bbox": [ + 511, + 625, + 906, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\frac {1}{B _ {L}} \\sum_ {i = 1} ^ {B _ {L}} \\mathcal {H} \\left(y _ {i} ^ {l}, \\mathcal {F} _ {m} \\left(y \\mid w \\left(x _ {i} ^ {l}\\right)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 801, + 906, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{H}(\\cdot, \\cdot)$ is cross-entropy loss, $w(\\cdot)$ is weak augmentation (i.e., random crop and flip) and $\\mathcal{F}_m(\\cdot)$ is the output probability from the model.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1197", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For unlabeled data, SSL typically dynamically assigns the pseudo-labels if the model's confidence scores surpass the thresholds. The unsupervised loss $\\mathcal{L}_u$ is defined as:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {1}{B _ {U}} \\sum_ {j = 1} ^ {B _ {U}} \\mathbb {1} \\left(\\max \\left(q _ {j}\\right) > \\tau\\right) \\mathcal {H} \\left(\\hat {q} _ {j}, Q _ {j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 159, + 482, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q_{j}$ and $Q_{j}$ denote abbreviation of $\\mathcal{F}_m(y|w(x_j^u))$ and $\\mathcal{F}_m(y|\\mathcal{A}(x_j^u))$ , respectively. $\\hat{q}_j = \\arg \\max (q_j)$ is the one-hot pseudo-label derived from the model's prediction on weakly augmented image. $\\mathcal{A}(\\cdot)$ denotes strong image augmentation [7], and $\\mathbb{1}(\\cdot)$ is an indicator function that selects pseudo-labels based on a threshold $\\tau$ . Our method enhances SSL by enriching pseudo-labels with semantic information to guide visual representation learning.", + "bbox": [ + 89, + 210, + 483, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Model Overview", + "text_level": 1, + "bbox": [ + 89, + 340, + 253, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method enhances the base SSL by incorporating additional guidance from label semantics (Figure 2). In the base SSL branch, weakly augmented images pass through the model first, generating one-hot pseudo-labels based on confidence thresholds. These pseudo-labels then guide the model to predict the same labels for strongly augmented versions of the corresponding unlabeled images [5, 36, 41]. In the boosting branch, label semantic enrichment module performs a mapping operation from categorical label index to its corresponding label name. A pre-trained text encoder extracts text embeddings from these label names, and a learnable projector transforms them into projected text embeddings, named text embedding bank, that covers all classes. To enhance robustness, we apply text embedding de-similarity (TEDS) loss, $\\mathcal{L}_{TEDS}$ , which encourages dissimilarity between different text embeddings. Finally, the one-hot pseudo-labels from the base SSL branch are used to select corresponding text embeddings, which are then aligned with the corresponding visual embeddings through class-aware visual-text alignment loss, $\\mathcal{L}_c$ . The overall training objective is:", + "bbox": [ + 89, + 363, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {T E D S} * \\mathcal {L} _ {T E D S}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 691, + 480, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{c}$ and $\\lambda_{TEDS}$ are the weighting factors for the text embedding de-similarity loss and the class-aware visual-text alignment loss, respectively.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Label Semantic Enrichment", + "text_level": 1, + "bbox": [ + 89, + 772, + 341, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This component converts categorical labels into meaningful words or phrases, allowing the subsequent text encoder module to capture semantic text embeddings. Unlike CLIP [31], which achieves text-image alignment using millions of (image, text) pairs, this is impractical in SSL due to limited annotations. In SSL, we often lack even basic categorical labels, let alone detailed image descriptions.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae1ddcecf2c1f9ca9c2bdcab7ed798661789768a2da9440c27d9fcd51a41f977.jpg", + "image_caption": [ + "T5" + ], + "image_footnote": [], + "bbox": [ + 544, + 85, + 712, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a51bb61f6922a4d879797190692f113bfdc768a6adc3bcfafc063e2787a67e4c.jpg", + "image_caption": [ + "BERT" + ], + "image_footnote": [], + "bbox": [ + 718, + 87, + 870, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efc17289a0f474f48a1e37dc9b2226386b15db3b2e7a5e49d869005fb2d34a09.jpg", + "image_caption": [ + "T5", + "Figure 3. t-SNE visualization of text embeddings for Food101 label names. The first row shows raw text embeddings generated by T5 and BERT. The second row is the embeddings after applying TEDS, resulting in more dispersed and dissimilar embeddings." + ], + "image_footnote": [], + "bbox": [ + 544, + 200, + 712, + 299 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2dcc0713ac24088bb3b08002f9244169e01864666237dd9eaf904cc94e44b9f1.jpg", + "image_caption": [ + "BERT" + ], + "image_footnote": [], + "bbox": [ + 718, + 200, + 870, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, since we have prior knowledge of all label names, our label semantic enrichment module performs a straightforward and fixed mapping from categorical pseudo-labels to their respective semantic names. This fixed mapping also contributes to the efficiency (discussed in Section 3.6). We directly feed the label name to the text encoder without any prompt operations. Using only label names ensures simplicity and reproducibility.", + "bbox": [ + 511, + 383, + 906, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Text Embedding De-Similarity", + "text_level": 1, + "bbox": [ + 511, + 513, + 784, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text embeddings generated by a text encoder are not inherently suitable for guiding visual representation learning because they lack sufficient separation between different classes (see the first row of Figure 3). To quantify this, we calculated the cosine similarity between text embeddings of different class names using the BERT model. The results showed high similarity scores (above 0.92) between classes such as \"Carrot Cake\" and \"Chocolate Cake\", as well as \"Cheesecake\" and \"Chocolate Cake\". This inherent similarity in text embeddings unintentionally pushes corresponding visual representations closer together during visual-text alignment, leading to increased confusion between semantic similar classes.", + "bbox": [ + 511, + 536, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate this issue, we introduce text embedding desimilarity (TEDS), a loss function that encourages greater separation between text embeddings by minimizing the absolute values of pairwise dot products between projected text embeddings. The TEDS loss is formulated as follows:", + "bbox": [ + 511, + 732, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T E D S}} = \\frac {1}{C (C - 1)} \\sum_ {i \\neq j} \\max \\left(\\left| \\frac {\\mathbf {T} _ {i} ^ {\\top} \\mathbf {T} _ {j}}{\\| \\mathbf {T} _ {i} \\| \\| \\mathbf {T} _ {j} \\|} \\right| - \\Delta , 0\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 816, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C$ is the number of classes, $\\mathbf{T}_i$ and $\\mathbf{T}_j$ are individual projected text embeddings, $\\| \\mathbf{T}_i\\|$ and $\\| \\mathbf{T}_j\\|$ represent their", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1198", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf06f05d0cb88275e6b343ede7905b487e71f847713a5c1632fc4eba113a32f4.jpg", + "image_caption": [ + "(a). NT-Xent", + "Figure 4. Comparison between NT-Xent loss and class-aware contrastive loss." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 254, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43175bce05b96c45c9698f444323dcd3c70b6128397d1c7f57a43c24718375b7.jpg", + "image_caption": [ + "(b). Class-aware Contrastive" + ], + "image_footnote": [], + "bbox": [ + 287, + 92, + 460, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "magnitudes and $\\Delta$ is a tolerance margin (e.g., $1 \\times 10^{-5}$ ), defining the lowest acceptable similarity level. Integrating TEDS into the training objective encourages the learnable projector to extract more distinct and discriminative features from text embeddings. Since the text encoder remains frozen, TEDS primarily trains the learnable projector to enhance the separation of projected text embeddings. After applying TEDS, t-SNE visualization (see the second row of Figure 3) confirms that the projected text embeddings become more widely separated, which in turn helps the model learn more distinguishable visual representations.", + "bbox": [ + 89, + 301, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Class-aware Visual-text Alignment", + "text_level": 1, + "bbox": [ + 89, + 476, + 393, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In [31], the NT-Xent-based visual-text alignment forms positive pairs based on specific fine-grained text-visual matches, where each image is paired with a unique text description, and all other text descriptions in the batch are treated as negatives. However, this approach is suboptimal in our SSL context, where a batch may contain multiple images from the same class (i.e., sharing the same text embedding). For example, if one batch has two visual $(V_{0}, V_{1})$ and their corresponding labels $(T_{0}, T_{1})$ where $T_{0} = T_{1}$ , the existing text-visual alignment loss would incorrectly treat $(V_{0}, T_{1})$ and $(V_{1}, T_{0})$ as negative pairs—False Negative Pairs (Figure 4). To address this, we introduce class-aware visual-text alignment which leverages class labels to define positive and negative pairs. Specifically, positive pairs are formed between text-visual pairs from the same class, while negative pairs come from different classes. This is implemented by computing a similarity matrix between visual and text embeddings and applying a mask to identify positive pairs based on class labels. For each visual-text pair within the same class, we maximize similarity, while minimizing similarity for pairs from different classes. The class-aware visual-text alignment loss is defined as:", + "bbox": [ + 89, + 498, + 483, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right) \\cdot \\mathbf {M} _ {i j}}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right)}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 851, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the number of embeddings, $\\tilde{\\mathbf{V}}_i$ is the normalized visual embedding of sample $i$ , $\\tilde{\\mathbf{T}}_j$ is the normalized text embedding of sample $j$ , $\\kappa$ is the temperature scaling factor and $\\mathbf{M}_{ij}$ is a class-aware mask where $\\mathbf{M}_{ij} = 1$ if samples $i$ and $j$ belong to the same class, and $\\mathbf{M}_{ij} = 0$ otherwise.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Implementation Detail", + "text_level": 1, + "bbox": [ + 511, + 195, + 723, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text Encoder. The text encoder plays a crucial role in extracting semantic information from label names to guide visual representation learning. The key requirement is the ability to capture meaningful semantics. Any text encoder that effectively extracts distinct and informative text embeddings can serve this purpose. To ensure flexibility and broad applicability, our method is designed to work with both language models and simple word embedding models that can generate semantically meaningful representations. We evaluate our approach using three language models—T5 [32], BERT [8], and CLIP's text encoder [31]—as well as a lightweight word embedding model, fastText [2]. Among these, the pre-trained CLIP text encoder is particularly well-suited for this task, as it has been specifically trained for visual-text alignment, making it highly effective in guiding visual feature learning. Ablation studies confirm that our method generalizes well across different text encoders, demonstrating its robustness and adaptability across various models.", + "bbox": [ + 511, + 218, + 906, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projection Strategies. In the SSL context, both the visual network architecture and the text encoder can be chosen freely, leading to a dimensional mismatch between text embeddings and visual embeddings. This mismatch poses a challenge for visual-text alignment, as the two modalities exist in different feature spaces. To address this issue, we explore two projection strategies that project text or visual embeddings to the same feature dimension. The first strategy, ProjT2V, is the approach used in Figure 2. In this method, a learnable projection layer is applied to the text embeddings, transforming them into the same dimensional space as the visual embeddings [16]. This transformation enables direct alignment between the projected text embeddings and the visual embeddings using contrastive loss. The second strategy, Proj-V2T, takes the opposite approach by projecting visual embeddings into the dimensional space of the text embeddings [11]. While both strategies successfully create a shared embedding space, our experiment shows that ProjT2V outperforms Proj-V2T, providing a stronger improvement in SSL performance and more effective visual-text alignment.", + "bbox": [ + 511, + 506, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projector Architecture. We implemented the projector architecture by considering two main factors. First, the learnable projector should present a challenging alignment task for the in-training visual feature extraction model. To achieve this, the projector cannot be overly complex, as", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1199", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that would exert too much influence on the visual or text embeddings, potentially distorting their features. Second, we aimed to keep the architecture lightweight and efficient, motivating us to choose a simpler design. Based on these requirements and existing works usually use linear projections [11, 16], we explored two one-layer architectures: a single-layer MLP and self-attention.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficient Implementation. Our method can achieve high efficiency with minimal vRAM usage and without adding significant computation when boosting SSLs for three main reasons. First, by freezing the text encoder, we only perform text encoder inference without the computationally expensive gradient calculations. Second, due to the fixed mapping from categorical labels to label names, the input to the text encoder model is limited to the number of label names. This allows text embeddings to be precomputed once during initialization and stored in a cache, eliminating the need for repeated computations during training. As a result, the text encoder can be offloaded from vRAM, and the necessary text embeddings can be efficiently retrieved from the pre-computed cache. Finally, we reuse the visual features already extracted by the base SSL for alignment, further optimizing efficiency. These optimizations ensure that our approach enhances SSL performance without introducing significant computational cost or memory overhead.", + "bbox": [ + 91, + 199, + 486, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 487, + 215, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 513, + 171, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluated our method on the latest SSL benchmarks [40]. To further evaluate the robustness on fine-grained classification task, we also tested our method on Food 101 dataset [3] as prior SSLs fail on this dataset. For each dataset, we examined two different quantities of labeled images. Detailed settings for each dataset are provided in the Appendix. For instance, we used 404 and 1010 labeled images for the Food-101 dataset. Labeled samples were randomly selected from the training set, except for STL-10 and Semi-Aves. For these two datasets, the labeled and unlabeled splits are predefined.", + "bbox": [ + 89, + 536, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Setup. Our method was based on FreeMatch [41] and implemented with a widely-used SSL codebase, USB [40]. For a fair comparison with baseline methods, we maintained consistent training settings across methods, adjusting only SSL-specific hyperparameters. This approach ensured that all SSLs use the same visual network architectures and pre-training strategies. For the base SSL settings, we initialized the self-adaptive threshold for FreeMatch at 0, whereas FreeMatch typically uses $\\frac{1}{C}$ . In our boosting configuration, we set the weight for the alignment loss term, $\\lambda_{c}$ , to 1 with a temperature factor of 0.07. We also assigned the weight of the text embedding de-similarity term, $\\lambda_{TEDS}$ to 0.01. Our model train", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing used the AdamW optimizer [18] with a cosine annealing learning rate schedule and an initial warmup phase. The training spanned 204,800 steps in total, with the first 10,240 steps dedicated to warmup. Batch sizes were adjusted based on the dataset, with labeled and unlabeled samples sharing the same batch size within each dataset. Specifically, batch sizes were set to 32 for Food-101, CIFAR-100, and STL-10; 64 for Semi-Aves; and 8 for EuroSAT. The base learning rates are also dataset-specific: $1e - 4$ for Food-101 and STL-10, $5e - 4$ for CIFAR-100 and Semi-Aves, and $5e - 5$ for EuroSAT.", + "bbox": [ + 511, + 90, + 906, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Overall Result", + "text_level": 1, + "bbox": [ + 511, + 271, + 661, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To the best of our knowledge, we compared our method with SSL algorithms in the past 4 years, as shown in Table 1. Our approach consistently outperformed prior SSL methods, particularly in fine-grained classification tasks. For example, SemiVisBooster achieved $73.6\\%$ accuracy on Food-101 using only 404 labeled images and $81.8\\%$ accuracy on CIFAR-100 with only 200 labeled images, outperforming the second-best method by margins of $13\\%$ and $2.9\\%$ , respectively. The improvement on Food-101 is more substantial than on CIFAR-100 because CIFAR-100 contains only some fine-grained classes, and existing SSL methods may effectively distinguish most classes with clear boundaries in CIFAR-100. On the Semi-Aves dataset, our method achieved a $2.8\\%$ accuracy improvement over the next best method. In coarse-grained classification tasks, our method also demonstrated strong results, achieving $95.4\\%$ accuracy on the STL-10 dataset with just 20 labeled images and $96.8\\%$ accuracy on the EuroSAT dataset with the same number of labeled images. These results show that our method is not only effective for fine-grained classification but also enhances performance in general coarse-grained classification problems.", + "bbox": [ + 511, + 294, + 906, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with CLIP. To assess the necessity of training a visual model using SSL, we compared our method against CLIP's zero-shot performance [31] and few-shot fine-tuning results. Despite using a smaller visual backbone (ViT Small [40]), our method outperformed both the CLIP-ViT Base and CLIP-ViT Large model on the CIFAR-100 and EuroSAT datasets, where CLIP lacks domain-specific adaptation (Table 1). CLIP performed impressively on Food-101 and STL-10, likely because its training on 400 million image-text pairs exposes it to common food items and general object categories present in these datasets [31]. We further fine-tune CLIP vision encoder with a classifier head on EuroSAT dataset with 20 labeled images. The fine-tuned CLIP-ViT Base and CLIP-ViT Large model achieved $84.63\\%$ and $79.64\\%$ accuracy, respectively. In contrast, our compact model—using only $24\\%$ of CLIP-Base's vision parameters and training from scratch—achieves $96.8\\%$ under USB [40] settings. This comparison highlights the chal", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1200", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c94dade931e8d73aeb31826ed6f32f1b5926569e7118bed7fec8b3066e97767c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Top 1 Accuracy: Bold indicates the highest accuracy and underline indicates the second highest accuracy. Star* marks the result from hugging face model inference [29]." + ], + "table_body": "
DatasetFood-101 [3]Semi-Aves [37]CIFAR-100 [19]STL-10 [6]EuroSAT [15]
# of labels40410103959200400401002040
SSLUDA [43]53.274.368.371.881.885.692.592.395.1
FixMatch [36]54.764.568.271.381.085.792.690.194.1
CoMatch [22]30.648.961.565.675.285.890.894.795.2
Dash [45]57.776.767.870.480.785.792.989.793.9
CRMatch [10]44.469.868.076.182.289.888.588.593.4
AdaMatch [33]60.176.668.479.484.087.993.393.895.4
FlexMatch [46]48.471.867.774.482.188.792.695.495.2
SimMatch [47]55.473.166.277.383.788.493.392.995.6
DeSSL [34]56.675.068.270.380.684.392.491.896.3
FreeMatch [41]60.678.867.578.984.688.592.094.394.7
SoftMatch [5]60.476.968.278.783.887.692.994.995.5
EPASS [27]64.877.665.977.882.691.593.882.390.1
SequenceMatch [28]43.767.968.170.177.486.489.692.992.6
SemiVisBooster73.683.870.381.885.495.495.996.896.7
Zero-shotCLIP-ViT-B[31]84.410.1*65.197.249.4
CLIP-ViT-L[31]92.99.9*77.999.359.9
", + "bbox": [ + 91, + 59, + 903, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lenges general models face in specialized domains and underscores the importance of continuously improving SSL methods.", + "bbox": [ + 89, + 388, + 482, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 440, + 202, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted the ablation studies using the following setup, unless specified: the projection strategy was ProjT2V, the projector architecture was an MLP, the text encoder was the CLIP text encoder, we used both text embedding de-similarity loss and class-aware visual-text alignment strategy for training, and the input to text encoder is the label name.", + "bbox": [ + 89, + 462, + 482, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Projection Strategies and Projector Architecture. We study the impact of two commonly used projection methods: ProjT2V, and Proj-V2T. Following the approach in [16], we implement the ProjT2V module using a single-layer MLP with the text embedding dimension as the input and the visual embedding dimension as the output. Similarly, for the Proj-V2T module, we use a single-layer MLP, but with the input and output dimensions reversed. For self-attention, we use a one-layer self-attention module with a hidden dimension of 512, followed by a linear MLP with an input dimension of 512 and an output dimension matching the visual embedding dimension.", + "bbox": [ + 89, + 568, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 2, we noticed that ProjT2V works best on both setups, while Proj-V2T lags behind. This is expected as ProjT2V directly enhances visual representation learning, whereas Proj-V2T lacks this capability. With ProjV2T, the alignment loss only ensures that the projected visual embeddings resemble the text embeddings, but this approach has two limitations. First, ProjV2T does not address inherent similarities across text embeddings from different classes, which can cause class overlap. Second, even if text embeddings are distinct, ProjV2T does not ensure", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that the pre-projection visual embeddings (i.e. the raw output from the vision model) will naturally align with the text embeddings. This is because the projector itself influences the alignment target, rather than allowing the vision model to learn distinct representations independently. In contrast, ProjT2V promotes direct visual representation learning by projecting the text embeddings and aligning them with the raw visual embeddings directly. Additionally, we ensure that text embeddings are distinct using the TEDS.", + "bbox": [ + 511, + 388, + 906, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth mentioning that we also ablate the projector used for T2V and MLP works slightly better, probably due to the simplicity of the MLP layer. It creates a more challenging task for the visual-text alignment, encouraging more robust visual feature extraction. These results support our decision to use Proj-T2V and MLP as projector.", + "bbox": [ + 511, + 527, + 906, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefit of TEDS and Impact of Text Encoders. We evaluate the impact of TEDS by testing our method both with and without TEDS across different text encoders. To implement our method without TEDS, we set $\\lambda_{t} = 0$ . The pre-trained text encoder models are from HuggingFace repositories.", + "bbox": [ + 511, + 621, + 906, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method consistently outperforms the baseline across different text encoders, including the simple word embedding model, FastText (Table 3). The key to boost SSL is incorporating semantic information to guide visual representations. Thus, our method is not restricted to a specific text model but is adaptable to various text encoders, ranging from lightweight word embedding models to language models, as long as they can extract meaningful semantics from label names. This flexibility enhances the practical applicability of our method in real-world scenarios.", + "bbox": [ + 511, + 700, + 908, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By comparing different text encoders, we found that the CLIP text encoder achieves the highest performance on Food-101 (404 labeled images), followed by fastText and", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1201", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e17432359c4f0bb8cf20f7d5081a8e9b8d5cbd80cddebce82c6fa165f451ac9b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Projection StrategiesProjectorFood-101
4041010
Proj-V2TMLP60.5080.38
Proj-T2VSelf-attention72.5583.83
Proj-T2VMLP73.5583.41
", + "bbox": [ + 96, + 71, + 277, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bc28b6bde5a715c5c45b0a7c220d593607b7006cbef2787be19d3298cd8482a1.jpg", + "table_caption": [ + "Table 2. Choice of Projector Strategies: Proj-T2V with MLP project achieves the best performance." + ], + "table_footnote": [], + "table_body": "
Text EncoderTEDSFood-101
404 1010
CLIP-Text Encoder✓73.55 83.41
BERT✓69.73 83.19
T5✓68.86 83.18
FastText✓72.80 83.13
CLIP-Text Encoder×70.40 82.63
BERT×67.67 82.44
T5×63.92 82.79
FastText×69.57 81.25
", + "bbox": [ + 302, + 70, + 488, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BERT model, while the T5 model performs the worst, particularly in the absence of TEDS. This performance discrepancy can be attributed to the different training objectives of these language models. The CLIP model is explicitly trained for text-image alignment using contrastive learning, creating a shared embedding space that directly supports text-vision alignment. This shared space makes the CLIP text encoder highly effective for guiding visual model training. BERT, while not designed specifically for text-vision alignment, captures rich textual representations that still contribute to the visual model learning process. In contrast, T5 is primarily designed for text-to-text tasks, and its training objectives are less suited for aligning visual and textual embeddings, resulting in its comparatively lower performance. Interestingly, FastText, a simple word embedding model, achieved strong performance by effectively extracting distinct text features from label names. Compared to general language models like BERT and T5, FastText focuses on subword structure rather than contextual meaning [2], allowing it to better preserve distinctions between label names and reduce confusion.", + "bbox": [ + 89, + 226, + 483, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results also indicated that TEDS significantly enhances performance across different text encoders (Table 3). This improvement is expected, as raw text embeddings for different classes often exhibit high similarity (Figure 3), which can mislead the visual model during training, especially the label data is scare. Our TEDS module mitigates this issue by penalizing the similarity between text embeddings from different class names, reducing the impact of inherent similarities. Conversely, our method without TEDS lacks this mechanism, leading to poorer performance due to similarity-induced confusion. This underscores the importance of TEDS in making our method adaptable to various text encoder models.", + "bbox": [ + 89, + 547, + 483, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefit of CAViTA. We evaluate the effectiveness of CAViTA loss by comparing it to the widely used NT-Xent contrastive loss [35], which is commonly used for visual-text alignment in CLIP. Table 4 shows that CAViTA significantly outperforms NT-Xent loss. This result is expected because multiple images in a batch can share the same text embedding. NT-Xent mistakenly classified true positive pairs as negatives, penalizing them and leading to a performance drop. Table 6 reports the False negative rate under NT-Xent loss. It mistakenly penalizes $18.71\\%$ of posi", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/518ae90c316ceaf7aa0f5cf0ed36eb85c3d0b05cc951f981a06711e113827162.jpg", + "table_caption": [ + "Table 3. Benefit of TEDS" + ], + "table_footnote": [], + "table_body": "
Text-visual Alignment LossFood-101
4041010
NT-Xent52.6175.22
CAViTA73.5583.41
", + "bbox": [ + 509, + 70, + 681, + 133 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e593fac45364dbe31dfb585d47d609be990362cd9f30a32ed62ad4b0a0a0c3f.jpg", + "table_caption": [ + "Table 4. Benefit of CAViTA: CAViTA outperforms NT-Xent loss because it correctly define positive pairs." + ], + "table_footnote": [], + "table_body": "
λcλTEDSFood-101
10.0173.6
10.173.8
1172.8
0.10.0161.7
0.010.0154.1
", + "bbox": [ + 710, + 70, + 898, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91e50ade626765def7f86bad6aadefbe940aa82bb5eb31b745b68819f3b81f62.jpg", + "table_caption": [ + "Table 5. Hyperparameter study: our method is sensitivity to ${\\lambda }_{c}$" + ], + "table_footnote": [], + "table_body": "
Dataset (labeled #)FN count(mean±std)FN ratio(mean±std)
Food 101 (404)7.65 ± 3.3418.71% ± 6.96%
EuroSAT (20)3.71 ± 2.0329.72% ± 11.79%
", + "bbox": [ + 540, + 224, + 875, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. False negative (FN) rate under NT-Xent loss", + "bbox": [ + 550, + 276, + 867, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive pairs, leading to a $28.5\\%$ accuracy drop on the Food101 dataset. Furthermore, applying NT-Xent loss within our method does not improve the base SSL method (i.e., FreeMatch) and instead decreases performance by $13.2\\%$ . This finding highlights the importance of correctly defining positive pairs, as misclassifications can significantly degrade model performance.", + "bbox": [ + 511, + 313, + 903, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benefit of Using Label Name. We compared our method with longer text embeddings by extending label names using GPT-4o with the prompt: \"Describe the visual features of {label name} in 2-3 sentences\". These extended descriptions were then used as input to the text encoder to guide visual representation learning. However, using longer text embeddings generated by GPT-4o led to a drop in performance, with accuracy decreasing to $70.0\\%$ on Food101 (404 labeled images) and $80.9\\%$ on CIFAR-100 (200 labeled images), compared to using only the label name. This decline in performance is because long text embedding may include unnecessary tokens that dilute the core semantics and spread out embeddings unnecessarily, resulting in unnecessary confusion between different classes. In contrast, a short label name provides compact and discriminative semantics, which is more suitable for our classification task.", + "bbox": [ + 511, + 425, + 905, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameter Study. Since $\\lambda_{u}$ has been studied in prior SSLs [36, 41], we focus on analyzing $\\lambda_{c}$ and $\\lambda_{TEDS}$ by varying them from 0.01 to 1 in Table 5. Our results indicate that the performance of our method is highly sensitive to $\\lambda_{c}$ . Increasing $\\lambda_{c}$ improves performance, as a stronger visual-text alignment loss helps enforce sufficient regularization on visual representations.", + "bbox": [ + 511, + 672, + 908, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 809, + 633, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SemiVisBooster to improve prior SSL by using semantics from label names to guide visual feature learning and achieved best performance on latest SSL benchmarks[40] and find-grained classification task.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1202", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ardhendu Behera, Zachary Wharton, Pradeep Hewage, and Asish Bera. Context-aware attentional pooling (cap) for fine-grained visual classification. In The Thirty-Fifth AAAI Conference on Artificial Intelligence. AAAI, 2021. 1", + "[2] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5, 2016. 5, 8", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461. Springer, 2014. 1, 6, 7", + "[4] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 194-203, 2021. 2", + "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 4, 7", + "[6] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR. 1, 7", + "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 4", + "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics, 2019. 5", + "[9] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 3", + "[10] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 7" + ], + "bbox": [ + 93, + 205, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Institute of Biomedical Imaging and Bioengineering under NIH award numbers R01LM011834 and R56EB032819.", + "[11] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 5, 6", + "[12] Andreas Fürst, Elisabeth Rumetshofer, Johannes Lehner, Viet T Tran, Fei Tang, Hubert Ramsauer, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto, et al. Cloob: Modern hopfield networks with infoloob outperform clip. Advances in neural information processing systems, 35: 20450-20468, 2022. 3", + "[13] Tanmay Gupta, Kevin Shih, Saurabh Singh, and Derek Hoiem. Aligned image-word representations improve inductive transfer across vision-language tasks. In 2017 IEEE International Conference on Computer Vision (ICCV), page 4223-4232. IEEE, 2017. 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 3, 5, 6, 7", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 7", + "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2", + "[21] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 1, 2", + "[22] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 7", + "[23] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks, page 121-137. Springer International Publishing, 2020. 3" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1203", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021.3", + "[25] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2", + "[27] Khanh-Binh Nguyen. Debiasing, calibrating, and improving Semi-supervised Learning performance via simple Ensemble Projector. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2430-2439, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 7", + "[28] Khanh-Binh Nguyen. Sequencematch revisiting the design of weak-strong augmentations for semi-supervised learning. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 96-105, 2024. 7", + "[29] OpenAI. CLIP: Contrastive Language-Image Pretraining. https://huggingface.co/openai/clip-vit-base-patch32, 2021. Accessed: 2024-11-10. 7", + "[30] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V. Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11557-11568, 2021. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2, 3, 4, 5, 6, 7", + "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5", + "[33] Becca Roelofs, David Berthelot, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: a unified approach to semi-supervised learning and domain adaptation. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2022. 7", + "[34] Hugo Schmutz, Olivier Humbert, and Pierre-Alexandre Mattei. Don't fear the unlabelled: safe semi-supervised learning via simple debiasing. arXiv preprint arXiv:2203.07512, 2022. 7", + "[35] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Proceedings of the 30th International Conference on Neural Information Processing Systems, page 1857-1865, Red Hook, NY, USA, 2016. Curran Associates Inc. 2, 8", + "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3, 4, 7, 8", + "[37] Jong-Chyi Su and Subhransu Maji. The semi-supervised naturalist-aves challenge at fgvc7 workshop. arXiv preprint arXiv:2103.06937, 2021. 7", + "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 1195–1204, Red Hook, NY, USA, 2017. Curran Associates Inc. 2", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1", + "[40] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 1, 2, 6, 8", + "[41] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 1, 3, 4, 6, 7, 8", + "[42] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Semi-supervised-learning. https://github.com/microsoft/Semi-supervised-learning, 2023.1", + "[43] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2, 7", + "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2", + "[45] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In International conference on machine learning, pages 11525-11536. PMLR, 2021. 2, 7", + "[46] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in neural information processing systems, 34:18408-18419, 2021. 1, 3, 7", + "[47] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14471-14481, 2022. 7" + ], + "bbox": [ + 516, + 90, + 903, + 895 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1204", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_model.json b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa7151416eed6463a04d519597d6e3b51f090fed --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.175 + ], + "angle": 0, + "content": "SemiVisBooster: Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.203, + 0.758, + 0.24 + ], + "angle": 0, + "content": "Wenjin Zhang\\(^{1,2}\\) Xinyu Li\\(^{2}\\) Chenyang Gao\\(^{1,2}\\) Ivan Marsic\\(^{1}\\) \n\\(^{1}\\)Rutgers University \\({}^{2}\\)Amazon" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.242, + 0.882, + 0.258 + ], + "angle": 0, + "content": "{wz315, cg694}@scarletmail.rutgers.edu, lixinyu.arthur@outlook.com, marsic@rutgers.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Deep learning models rely on large-scale labeled datasets, but collecting such data is expensive and time-consuming. Semi-supervised learning (SSL) mitigates this issue by learning from a small set of labeled samples along with a large pool of unlabeled data. However, existing SSL methods struggle with fine-grained classification when dealing with visually similar classes, as they rely solely on visual features and ignore the semantic information within label names. This paper introduces SemiVisBooster, an SSL enhancement approach that utilizes semantic information from label names to guide visual feature learning, addressing the challenges of fine-grained classification. By aligning text embeddings from label names with visual features, our method helps the model capture subtle visual distinctions that purely visual representations may overlook. To enhance robustness, we propose two key components: (1) text embedding de-similarity (TEDS) to reduce confusion caused by similar text embeddings across different class names, and (2) class-aware visual-text alignment loss to accurately define positive and negative pairs during visual-text alignment. Our method achieves state-of-the-art performance on the latest SSL benchmarks. Additionally, on the challenging Food-101 dataset, which contains many visually similar classes and uses only 404 labeled images, our approach improves performance by approximately \\(13.6\\%\\) over the second-best method. Code is available at SemiVis-Booster" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.768, + 0.222, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The superior performance of deep learning heavily relies on supervised training with sufficient labeled data [14, 39]. However, acquiring such massive labeled datasets for real-world applications remains a labor-intensive and expensive process. To mitigate this challenge, semi-supervised learning (SSL) has emerged as a promising method, learning from a small set of labeled data while leveraging a" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.291, + 0.707, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.395, + 0.653, + 0.405 + ], + "angle": 0, + "content": "(a) FreeMatch" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.291, + 0.865, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.744, + 0.395, + 0.83, + 0.405 + ], + "angle": 0, + "content": "(b) SemiVisBooster" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 1. Comparison of visual feature space distribution for (a) FreeMatch and (b) our method on the Food-101 dataset visualized using t-SNE. We only plot the first 20 classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.903 + ], + "angle": 0, + "content": "large amount of unlabeled data [5, 21, 36, 41, 46]. By significantly reducing the dependence on manual annotations, SSL has demonstrated great potential in practical applications. Despite achieving strong performance in general image classification, particularly on datasets with well-separated class boundaries such as CIFAR-10 [19] and STL-10 [6], SSL methods struggle with fine-grained classification, where visually similar categories exist. Analyzing the confusion matrix of the popular SSL method, FreeMatch [41], on the CIFAR-100 dataset with only 200 labeled images [42], we observed that its accuracy plateaued at \\(78.9\\%\\) due to frequent misclassifications among visually similar classes such as \"woman,\" \"man,\" \"boy,\" and \"girl\" [40]. This observation indicates a key limitation of existing SSL approaches: their reliance on visual feature extraction, which often fails to distinguish subtle inter-class differences. To further validate the observed limitation on SSL, we evaluated FreeMatch on the fine-grained Food101 dataset [3] using 404 labeled images because this dataset contains many visually similar classes. FreeMatch achieved an accuracy of only \\(60.4\\%\\), while a fully supervised training method using all available labeled data reached \\(98.8\\%\\) [1], highlighting a significant performance gap between SSL and fully supervised learning. This result underscores the need for improvements in SSL methods. The t-SNE visualization (Figure 1) illustrates that while FreeMatch exhibits some level of clustering, it struggles to separate visually similar classes, leading to significant feature overlap. This challenge arises from two key weaknesses in existing SSL methods. (1)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "SSL models rely primarily on visual patterns for classification. However, in fine-grained settings, where objects share highly similar appearances, these methods struggle to extract meaningful distinctions, resulting in frequent misclassification. (2). SSL methods depend heavily on pseudolabeling, where high-confidence predictions are used to assign labels to unlabeled samples. When dealing with visually ambiguous cases, incorrect pseudo-labels are generated and subsequently reinforce errors, leading to model bias and further performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.247, + 0.485, + 0.868 + ], + "angle": 0, + "content": "In many real-world applications, class names are available as meaningful words or phrases, offering an opportunity to use the semantic information. However, existing SSL methods completely ignore the rich semantic information embedded in these labels. Instead, prior SSLs rely exclusively on one-hot-encoded categorical labels, which discard potentially useful semantic knowledge. We hypothesize that introducing label semantics into SSL can provide additional guidance for visual representation learning, thereby improving the model's ability to distinguish visually similar samples. For example, while \"boy\" and \"man\" may have highly similar visual representations in CIFAR-100 dataset, their semantic embeddings are distinct and different. By aligning visual features with distinct label semantics, SSL models can be improved in two ways. (1). Aligning text embeddings from ground-truth label names with the visual representations of labeled images can help the model better capture subtle visual differences through text cues. This makes the initial model more stable and robust. (2) Even if pseudo-labels are incorrect, aligning the text embedding of the pseudo-label with the visual representation of the unlabeled image can still be beneficial. If the incorrect pseudo-label's semantic meaning is close to the true label, this alignment may help guide the visual representation closer to the correct cluster. However, aligning text embeddings from label name with visual embeddings for SSL introduces two major challenges. The first challenge arises from the inherent similarity between text embeddings of different class names. When similar text embeddings are used to guide the visual representation learning, it can introduce unwanted confusion in the visual space (Section 3.4). The second challenge relates to many-to-one alignment in our settings. Unlike CLIP [31] using NT-Xent loss [35], where each image-text pair is uniquely defined, our setting often encounters multiple images within a batch that share the same text embedding because multiple images may belong to the same class. Directly applying existing visual-text alignment loss like NT-Xent [31], which assume a pairwise positive match, would mistakenly treat these samples as negatives, leading to incorrect optimization and degraded performance (Section 3.5)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced SemiVisBooster, an SSL booster that improves SSL performance by incorpo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "rating label semantics into the learning process. To enrich feature learning beyond basic one-hot-encoded ground truth labels or pseudo-labels, we proposed a label semantic enrichment module that converts categorical labels into meaningful label names. To enable robust alignment, we proposed text embedding de-similarity (TEDS) loss to avoid the impact of inherent similarity between text embeddings and developed class-aware visual-text alignment (CAViTA) strategy, which correctly defines positive and negative pairs in cases where multiple images share the same text embedding. Our approach achieves state-of-the-art results on the latest SSL benchmarks [40]. Specifically, SemiVisBooster consistently surpasses the previous method by more than \\(2.9\\%\\) across multiple benchmarks, including CIFAR-100 (200 labeled images) and STL-10 (40 labeled images). For fine-grained classification tasks, our method achieved \\(73.6\\%\\) accuracy on Food 101 (404 labeled images), significantly surpassing the previous state-of-the-art method, which achieved only \\(64.8\\%\\) accuracy. We summarized our contributions below:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.904, + 0.455 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to directly leverage label name semantics to enhance SSL, achieving state-of-the-art performance on the latest SSL benchmarks and fine-grained classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.456, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We introduced TEDS to mitigate the inherent similarity between text embeddings of different class names, enhancing the robustness of our method across various text encoders, including a simple word embedding model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.516, + 0.905, + 0.56 + ], + "angle": 0, + "content": "- We designed CAViTA to accurately distinguish positive and negative visual-text embedding pairs, ensuring robust visual-text alignment." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.579, + 0.655, + 0.595 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.753, + 0.622 + ], + "angle": 0, + "content": "2.1. Semi-supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Semi-supervised learning (SSL) has been an area of extensive research over recent decades. Recent advances in SSL methodologies can be broadly categorized into two primary approaches: consistency regularization [20, 26, 38, 43] and pseudo-labeling [21, 30]. Consistency-based methods apply perturbations either to the input data or the model and enforce constraints to ensure consistency in the features or probability outputs. For example, the II-model [20] introduces dropout to model weights, generating two outputs for consistency regularization. Pseudo-labeling methods, in contrast, use model predictions as pseudo-labels to guide learning on unlabeled data. This approach has gained recent prominence [4, 21, 30, 44]. Combining these approaches, FixMatch [36] introduced a straightforward SSL framework that applies both pseudo-labeling and consistency regularization, but it uses a fixed confidence threshold. Recent work has focused on refining pseudo-labeling quality by introducing dynamic thresholds [45], class-specific thresh-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1196" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.082, + 0.868, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.413, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 2. Overview of SemiVisBooer framework for enhancing SSL through label semantic information. (Only the flow of unlabeled data is shown. Labeled data are used for both supervised learning and aligning ground truth labels with their corresponding images.) Base SSL branch: One-hot pseudo-labels are generated from weakly augmented images and used to guide predictions on strongly augmented versions of the same images. Boosting Branch: A label semantic enrichment module maps each categorical label to its label name and uses a pre-trained text encoder to extract text embeddings without any prompt operation. A learnable projector transforms all these embeddings into a text embedding bank. The text embedding de-similarity (TEDS) loss enhances robustness by minimizing the similarity between text embeddings of different classes. The class-aware visual-text alignment is used to align selected text embeddings with corresponding visual embeddings to guide visual representation learning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.533, + 0.484, + 0.639 + ], + "angle": 0, + "content": "olds [46], and dynamic class-specific thresholds [5, 41]. However, even with these adaptive thresholding techniques, existing SSL methods struggle with fine-grained classification as we discussed in Section 1. To boost SSLs, we introduced additional supervision derived from the semantic information in label names. We used FreeMatch[41] as our base SSL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.376, + 0.669 + ], + "angle": 0, + "content": "2.2. Vision and Language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Numerous studies have demonstrated the benefits of joint image-text representation learning for various vision tasks [9, 13, 16, 23, 25]. Contrastive learning has proven particularly effective for vision-language models by mapping images and text into a shared feature space. Models like CLIP [31], CLOOB [12], ALIGN [17], and DeCLIP [24] achieve this by training text and image encoders on large datasets of image-caption pairs, using contrastive loss to minimize distances between matching pairs while maximizing them for non-matching ones. To handle noise in datasets, ALIGN [16] and DeCLIP [24] use custom metrics, while CLIP [31] employs cosine distance. While these models primarily aim at pre-training large vision-language foundations for zero- and few-shot tasks, our approach enhances semi-supervised learning by aligning visual features" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.907, + 0.563 + ], + "angle": 0, + "content": "with pseudo-labels, improving the robustness of SSLs for fine-grained classification." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.578, + 0.606, + 0.593 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.603, + 0.655, + 0.618 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.625, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Following [36], we formulate the SSL framework for a \\(C\\)-class classification problem. Let \\(\\mathcal{D}^l = \\{(x_i^l, y_i^l) : i \\in [1, B_L]\\}\\) denote a batch of labeled data, and \\(\\mathcal{D}^u = \\{x_j^u : j \\in [1, B_U]\\}\\) represent a batch of unlabeled data, with batch sizes \\(B_L\\) and \\(B_U\\), respectively. Here, \\(x_i^l\\) denotes a labeled image with label \\(y_i^l\\), and \\(x_j^u\\) represents an unlabeled image. The model is trained using a combined objective: \\(\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_s + \\lambda_u \\mathcal{L}_u\\), where \\(\\mathcal{L}_s\\) is the supervised loss over labeled data, \\(\\mathcal{L}_u\\) is the unsupervised loss over unlabeled data, and \\(\\lambda_u\\) is the unsupervised loss weight. The supervised loss \\(\\mathcal{L}_s\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.803, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\frac {1}{B _ {L}} \\sum_ {i = 1} ^ {B _ {L}} \\mathcal {H} \\left(y _ {i} ^ {l}, \\mathcal {F} _ {m} \\left(y \\mid w \\left(x _ {i} ^ {l}\\right)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}(\\cdot, \\cdot)\\) is cross-entropy loss, \\(w(\\cdot)\\) is weak augmentation (i.e., random crop and flip) and \\(\\mathcal{F}_m(\\cdot)\\) is the output probability from the model." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "For unlabeled data, SSL typically dynamically assigns the pseudo-labels if the model's confidence scores surpass the thresholds. The unsupervised loss \\(\\mathcal{L}_u\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.16, + 0.483, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {1}{B _ {U}} \\sum_ {j = 1} ^ {B _ {U}} \\mathbb {1} \\left(\\max \\left(q _ {j}\\right) > \\tau\\right) \\mathcal {H} \\left(\\hat {q} _ {j}, Q _ {j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.334 + ], + "angle": 0, + "content": "where \\( q_{j} \\) and \\( Q_{j} \\) denote abbreviation of \\( \\mathcal{F}_m(y|w(x_j^u)) \\) and \\( \\mathcal{F}_m(y|\\mathcal{A}(x_j^u)) \\), respectively. \\( \\hat{q}_j = \\arg \\max (q_j) \\) is the one-hot pseudo-label derived from the model's prediction on weakly augmented image. \\( \\mathcal{A}(\\cdot) \\) denotes strong image augmentation [7], and \\( \\mathbb{1}(\\cdot) \\) is an indicator function that selects pseudo-labels based on a threshold \\( \\tau \\). Our method enhances SSL by enriching pseudo-labels with semantic information to guide visual representation learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.342, + 0.254, + 0.356 + ], + "angle": 0, + "content": "3.2. Model Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Our method enhances the base SSL by incorporating additional guidance from label semantics (Figure 2). In the base SSL branch, weakly augmented images pass through the model first, generating one-hot pseudo-labels based on confidence thresholds. These pseudo-labels then guide the model to predict the same labels for strongly augmented versions of the corresponding unlabeled images [5, 36, 41]. In the boosting branch, label semantic enrichment module performs a mapping operation from categorical label index to its corresponding label name. A pre-trained text encoder extracts text embeddings from these label names, and a learnable projector transforms them into projected text embeddings, named text embedding bank, that covers all classes. To enhance robustness, we apply text embedding de-similarity (TEDS) loss, \\(\\mathcal{L}_{TEDS}\\), which encourages dissimilarity between different text embeddings. Finally, the one-hot pseudo-labels from the base SSL branch are used to select corresponding text embeddings, which are then aligned with the corresponding visual embeddings through class-aware visual-text alignment loss, \\(\\mathcal{L}_c\\). The overall training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.693, + 0.482, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {T E D S} * \\mathcal {L} _ {T E D S}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "where \\(\\lambda_{c}\\) and \\(\\lambda_{TEDS}\\) are the weighting factors for the text embedding de-similarity loss and the class-aware visual-text alignment loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.343, + 0.788 + ], + "angle": 0, + "content": "3.3. Label Semantic Enrichment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This component converts categorical labels into meaningful words or phrases, allowing the subsequent text encoder module to capture semantic text embeddings. Unlike CLIP [31], which achieves text-image alignment using millions of (image, text) pairs, this is impractical in SSL due to limited annotations. In SSL, we often lack even basic categorical labels, let alone detailed image descriptions." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.086, + 0.713, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.188, + 0.646, + 0.198 + ], + "angle": 0, + "content": "T5" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.088, + 0.871, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.189, + 0.805, + 0.198 + ], + "angle": 0, + "content": "BERT" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.202, + 0.714, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.632, + 0.303, + 0.646, + 0.312 + ], + "angle": 0, + "content": "T5" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.202, + 0.871, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.303, + 0.808, + 0.312 + ], + "angle": 0, + "content": "BERT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.906, + 0.375 + ], + "angle": 0, + "content": "Figure 3. t-SNE visualization of text embeddings for Food101 label names. The first row shows raw text embeddings generated by T5 and BERT. The second row is the embeddings after applying TEDS, resulting in more dispersed and dissimilar embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.507 + ], + "angle": 0, + "content": "However, since we have prior knowledge of all label names, our label semantic enrichment module performs a straightforward and fixed mapping from categorical pseudo-labels to their respective semantic names. This fixed mapping also contributes to the efficiency (discussed in Section 3.6). We directly feed the label name to the text encoder without any prompt operations. Using only label names ensures simplicity and reproducibility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.785, + 0.531 + ], + "angle": 0, + "content": "3.4. Text Embedding De-Similarity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Text embeddings generated by a text encoder are not inherently suitable for guiding visual representation learning because they lack sufficient separation between different classes (see the first row of Figure 3). To quantify this, we calculated the cosine similarity between text embeddings of different class names using the BERT model. The results showed high similarity scores (above 0.92) between classes such as \"Carrot Cake\" and \"Chocolate Cake\", as well as \"Cheesecake\" and \"Chocolate Cake\". This inherent similarity in text embeddings unintentionally pushes corresponding visual representations closer together during visual-text alignment, leading to increased confusion between semantic similar classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To mitigate this issue, we introduce text embedding desimilarity (TEDS), a loss function that encourages greater separation between text embeddings by minimizing the absolute values of pairwise dot products between projected text embeddings. The TEDS loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T E D S}} = \\frac {1}{C (C - 1)} \\sum_ {i \\neq j} \\max \\left(\\left| \\frac {\\mathbf {T} _ {i} ^ {\\top} \\mathbf {T} _ {j}}{\\| \\mathbf {T} _ {i} \\| \\| \\mathbf {T} _ {j} \\|} \\right| - \\Delta , 0\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "where \\(C\\) is the number of classes, \\(\\mathbf{T}_i\\) and \\(\\mathbf{T}_j\\) are individual projected text embeddings, \\(\\| \\mathbf{T}_i\\|\\) and \\(\\| \\mathbf{T}_j\\|\\) represent their" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.092, + 0.256, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.259, + 0.221, + 0.27 + ], + "angle": 0, + "content": "(a). NT-Xent" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.093, + 0.462, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.258, + 0.448, + 0.27 + ], + "angle": 0, + "content": "(b). Class-aware Contrastive" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Comparison between NT-Xent loss and class-aware contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.469 + ], + "angle": 0, + "content": "magnitudes and \\(\\Delta\\) is a tolerance margin (e.g., \\(1 \\times 10^{-5}\\)), defining the lowest acceptable similarity level. Integrating TEDS into the training objective encourages the learnable projector to extract more distinct and discriminative features from text embeddings. Since the text encoder remains frozen, TEDS primarily trains the learnable projector to enhance the separation of projected text embeddings. After applying TEDS, t-SNE visualization (see the second row of Figure 3) confirms that the projected text embeddings become more widely separated, which in turn helps the model learn more distinguishable visual representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.395, + 0.493 + ], + "angle": 0, + "content": "3.5. Class-aware Visual-text Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.833 + ], + "angle": 0, + "content": "In [31], the NT-Xent-based visual-text alignment forms positive pairs based on specific fine-grained text-visual matches, where each image is paired with a unique text description, and all other text descriptions in the batch are treated as negatives. However, this approach is suboptimal in our SSL context, where a batch may contain multiple images from the same class (i.e., sharing the same text embedding). For example, if one batch has two visual \\((V_{0}, V_{1})\\) and their corresponding labels \\((T_{0}, T_{1})\\) where \\(T_{0} = T_{1}\\), the existing text-visual alignment loss would incorrectly treat \\((V_{0}, T_{1})\\) and \\((V_{1}, T_{0})\\) as negative pairs—False Negative Pairs (Figure 4). To address this, we introduce class-aware visual-text alignment which leverages class labels to define positive and negative pairs. Specifically, positive pairs are formed between text-visual pairs from the same class, while negative pairs come from different classes. This is implemented by computing a similarity matrix between visual and text embeddings and applying a mask to identify positive pairs based on class labels. For each visual-text pair within the same class, we maximize similarity, while minimizing similarity for pairs from different classes. The class-aware visual-text alignment loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.852, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right) \\cdot \\mathbf {M} _ {i j}}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right)}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.182 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of embeddings, \\(\\tilde{\\mathbf{V}}_i\\) is the normalized visual embedding of sample \\(i\\), \\(\\tilde{\\mathbf{T}}_j\\) is the normalized text embedding of sample \\(j\\), \\(\\kappa\\) is the temperature scaling factor and \\(\\mathbf{M}_{ij}\\) is a class-aware mask where \\(\\mathbf{M}_{ij} = 1\\) if samples \\(i\\) and \\(j\\) belong to the same class, and \\(\\mathbf{M}_{ij} = 0\\) otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.196, + 0.724, + 0.212 + ], + "angle": 0, + "content": "3.6. Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Text Encoder. The text encoder plays a crucial role in extracting semantic information from label names to guide visual representation learning. The key requirement is the ability to capture meaningful semantics. Any text encoder that effectively extracts distinct and informative text embeddings can serve this purpose. To ensure flexibility and broad applicability, our method is designed to work with both language models and simple word embedding models that can generate semantically meaningful representations. We evaluate our approach using three language models—T5 [32], BERT [8], and CLIP's text encoder [31]—as well as a lightweight word embedding model, fastText [2]. Among these, the pre-trained CLIP text encoder is particularly well-suited for this task, as it has been specifically trained for visual-text alignment, making it highly effective in guiding visual feature learning. Ablation studies confirm that our method generalizes well across different text encoders, demonstrating its robustness and adaptability across various models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Projection Strategies. In the SSL context, both the visual network architecture and the text encoder can be chosen freely, leading to a dimensional mismatch between text embeddings and visual embeddings. This mismatch poses a challenge for visual-text alignment, as the two modalities exist in different feature spaces. To address this issue, we explore two projection strategies that project text or visual embeddings to the same feature dimension. The first strategy, ProjT2V, is the approach used in Figure 2. In this method, a learnable projection layer is applied to the text embeddings, transforming them into the same dimensional space as the visual embeddings [16]. This transformation enables direct alignment between the projected text embeddings and the visual embeddings using contrastive loss. The second strategy, Proj-V2T, takes the opposite approach by projecting visual embeddings into the dimensional space of the text embeddings [11]. While both strategies successfully create a shared embedding space, our experiment shows that ProjT2V outperforms Proj-V2T, providing a stronger improvement in SSL performance and more effective visual-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Projector Architecture. We implemented the projector architecture by considering two main factors. First, the learnable projector should present a challenging alignment task for the in-training visual feature extraction model. To achieve this, the projector cannot be overly complex, as" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1199" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "that would exert too much influence on the visual or text embeddings, potentially distorting their features. Second, we aimed to keep the architecture lightweight and efficient, motivating us to choose a simpler design. Based on these requirements and existing works usually use linear projections [11, 16], we explored two one-layer architectures: a single-layer MLP and self-attention." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.487, + 0.471 + ], + "angle": 0, + "content": "Efficient Implementation. Our method can achieve high efficiency with minimal vRAM usage and without adding significant computation when boosting SSLs for three main reasons. First, by freezing the text encoder, we only perform text encoder inference without the computationally expensive gradient calculations. Second, due to the fixed mapping from categorical labels to label names, the input to the text encoder model is limited to the number of label names. This allows text embeddings to be precomputed once during initialization and stored in a cache, eliminating the need for repeated computations during training. As a result, the text encoder can be offloaded from vRAM, and the necessary text embeddings can be efficiently retrieved from the pre-computed cache. Finally, we reuse the visual features already extracted by the base SSL for alignment, further optimizing efficiency. These optimizations ensure that our approach enhances SSL performance without introducing significant computational cost or memory overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.488, + 0.216, + 0.505 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.514, + 0.173, + 0.53 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Datasets. We evaluated our method on the latest SSL benchmarks [40]. To further evaluate the robustness on fine-grained classification task, we also tested our method on Food 101 dataset [3] as prior SSLs fail on this dataset. For each dataset, we examined two different quantities of labeled images. Detailed settings for each dataset are provided in the Appendix. For instance, we used 404 and 1010 labeled images for the Food-101 dataset. Labeled samples were randomly selected from the training set, except for STL-10 and Semi-Aves. For these two datasets, the labeled and unlabeled splits are predefined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experimental Setup. Our method was based on FreeMatch [41] and implemented with a widely-used SSL codebase, USB [40]. For a fair comparison with baseline methods, we maintained consistent training settings across methods, adjusting only SSL-specific hyperparameters. This approach ensured that all SSLs use the same visual network architectures and pre-training strategies. For the base SSL settings, we initialized the self-adaptive threshold for FreeMatch at 0, whereas FreeMatch typically uses \\(\\frac{1}{C}\\). In our boosting configuration, we set the weight for the alignment loss term, \\(\\lambda_{c}\\), to 1 with a temperature factor of 0.07. We also assigned the weight of the text embedding de-similarity term, \\(\\lambda_{TEDS}\\) to 0.01. Our model train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.258 + ], + "angle": 0, + "content": "ing used the AdamW optimizer [18] with a cosine annealing learning rate schedule and an initial warmup phase. The training spanned 204,800 steps in total, with the first 10,240 steps dedicated to warmup. Batch sizes were adjusted based on the dataset, with labeled and unlabeled samples sharing the same batch size within each dataset. Specifically, batch sizes were set to 32 for Food-101, CIFAR-100, and STL-10; 64 for Semi-Aves; and 8 for EuroSAT. The base learning rates are also dataset-specific: \\( 1e - 4 \\) for Food-101 and STL-10, \\( 5e - 4 \\) for CIFAR-100 and Semi-Aves, and \\( 5e - 5 \\) for EuroSAT." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.663, + 0.287 + ], + "angle": 0, + "content": "4.2. Overall Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.908, + 0.628 + ], + "angle": 0, + "content": "To the best of our knowledge, we compared our method with SSL algorithms in the past 4 years, as shown in Table 1. Our approach consistently outperformed prior SSL methods, particularly in fine-grained classification tasks. For example, SemiVisBooster achieved \\(73.6\\%\\) accuracy on Food-101 using only 404 labeled images and \\(81.8\\%\\) accuracy on CIFAR-100 with only 200 labeled images, outperforming the second-best method by margins of \\(13\\%\\) and \\(2.9\\%\\), respectively. The improvement on Food-101 is more substantial than on CIFAR-100 because CIFAR-100 contains only some fine-grained classes, and existing SSL methods may effectively distinguish most classes with clear boundaries in CIFAR-100. On the Semi-Aves dataset, our method achieved a \\(2.8\\%\\) accuracy improvement over the next best method. In coarse-grained classification tasks, our method also demonstrated strong results, achieving \\(95.4\\%\\) accuracy on the STL-10 dataset with just 20 labeled images and \\(96.8\\%\\) accuracy on the EuroSAT dataset with the same number of labeled images. These results show that our method is not only effective for fine-grained classification but also enhances performance in general coarse-grained classification problems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparison with CLIP. To assess the necessity of training a visual model using SSL, we compared our method against CLIP's zero-shot performance [31] and few-shot fine-tuning results. Despite using a smaller visual backbone (ViT Small [40]), our method outperformed both the CLIP-ViT Base and CLIP-ViT Large model on the CIFAR-100 and EuroSAT datasets, where CLIP lacks domain-specific adaptation (Table 1). CLIP performed impressively on Food-101 and STL-10, likely because its training on 400 million image-text pairs exposes it to common food items and general object categories present in these datasets [31]. We further fine-tune CLIP vision encoder with a classifier head on EuroSAT dataset with 20 labeled images. The fine-tuned CLIP-ViT Base and CLIP-ViT Large model achieved \\(84.63\\%\\) and \\(79.64\\%\\) accuracy, respectively. In contrast, our compact model—using only \\(24\\%\\) of CLIP-Base's vision parameters and training from scratch—achieves \\(96.8\\%\\) under USB [40] settings. This comparison highlights the chal" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1200" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.06, + 0.905, + 0.355 + ], + "angle": 0, + "content": "
DatasetFood-101 [3]Semi-Aves [37]CIFAR-100 [19]STL-10 [6]EuroSAT [15]
# of labels40410103959200400401002040
SSLUDA [43]53.274.368.371.881.885.692.592.395.1
FixMatch [36]54.764.568.271.381.085.792.690.194.1
CoMatch [22]30.648.961.565.675.285.890.894.795.2
Dash [45]57.776.767.870.480.785.792.989.793.9
CRMatch [10]44.469.868.076.182.289.888.588.593.4
AdaMatch [33]60.176.668.479.484.087.993.393.895.4
FlexMatch [46]48.471.867.774.482.188.792.695.495.2
SimMatch [47]55.473.166.277.383.788.493.392.995.6
DeSSL [34]56.675.068.270.380.684.392.491.896.3
FreeMatch [41]60.678.867.578.984.688.592.094.394.7
SoftMatch [5]60.476.968.278.783.887.692.994.995.5
EPASS [27]64.877.665.977.882.691.593.882.390.1
SequenceMatch [28]43.767.968.170.177.486.489.692.992.6
SemiVisBooster73.683.870.381.885.495.495.996.896.7
Zero-shotCLIP-ViT-B[31]84.410.1*65.197.249.4
CLIP-ViT-L[31]92.99.9*77.999.359.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.356, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Table 1. Top 1 Accuracy: Bold indicates the highest accuracy and underline indicates the second highest accuracy. Star* marks the result from hugging face model inference [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.435 + ], + "angle": 0, + "content": "lenges general models face in specialized domains and underscores the importance of continuously improving SSL methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.203, + 0.456 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.568 + ], + "angle": 0, + "content": "We conducted the ablation studies using the following setup, unless specified: the projection strategy was ProjT2V, the projector architecture was an MLP, the text encoder was the CLIP text encoder, we used both text embedding de-similarity loss and class-aware visual-text alignment strategy for training, and the input to text encoder is the label name." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Projection Strategies and Projector Architecture. We study the impact of two commonly used projection methods: ProjT2V, and Proj-V2T. Following the approach in [16], we implement the ProjT2V module using a single-layer MLP with the text embedding dimension as the input and the visual embedding dimension as the output. Similarly, for the Proj-V2T module, we use a single-layer MLP, but with the input and output dimensions reversed. For self-attention, we use a one-layer self-attention module with a hidden dimension of 512, followed by a linear MLP with an input dimension of 512 and an output dimension matching the visual embedding dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "From Table 2, we noticed that ProjT2V works best on both setups, while Proj-V2T lags behind. This is expected as ProjT2V directly enhances visual representation learning, whereas Proj-V2T lacks this capability. With ProjV2T, the alignment loss only ensures that the projected visual embeddings resemble the text embeddings, but this approach has two limitations. First, ProjV2T does not address inherent similarities across text embeddings from different classes, which can cause class overlap. Second, even if text embeddings are distinct, ProjV2T does not ensure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.526 + ], + "angle": 0, + "content": "that the pre-projection visual embeddings (i.e. the raw output from the vision model) will naturally align with the text embeddings. This is because the projector itself influences the alignment target, rather than allowing the vision model to learn distinct representations independently. In contrast, ProjT2V promotes direct visual representation learning by projecting the text embeddings and aligning them with the raw visual embeddings directly. Additionally, we ensure that text embeddings are distinct using the TEDS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.62 + ], + "angle": 0, + "content": "It is worth mentioning that we also ablate the projector used for T2V and MLP works slightly better, probably due to the simplicity of the MLP layer. It creates a more challenging task for the visual-text alignment, encouraging more robust visual feature extraction. These results support our decision to use Proj-T2V and MLP as projector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Benefit of TEDS and Impact of Text Encoders. We evaluate the impact of TEDS by testing our method both with and without TEDS across different text encoders. To implement our method without TEDS, we set \\(\\lambda_{t} = 0\\). The pre-trained text encoder models are from HuggingFace repositories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Our method consistently outperforms the baseline across different text encoders, including the simple word embedding model, FastText (Table 3). The key to boost SSL is incorporating semantic information to guide visual representations. Thus, our method is not restricted to a specific text model but is adaptable to various text encoders, ranging from lightweight word embedding models to language models, as long as they can extract meaningful semantics from label names. This flexibility enhances the practical applicability of our method in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "By comparing different text encoders, we found that the CLIP text encoder achieves the highest performance on Food-101 (404 labeled images), followed by fastText and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.515, + 0.958 + ], + "angle": 0, + "content": "1201" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.072, + 0.278, + 0.145 + ], + "angle": 0, + "content": "
Projection StrategiesProjectorFood-101
4041010
Proj-V2TMLP60.5080.38
Proj-T2VSelf-attention72.5583.83
Proj-T2VMLP73.5583.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.155, + 0.283, + 0.209 + ], + "angle": 0, + "content": "Table 2. Choice of Projector Strategies: Proj-T2V with MLP project achieves the best performance." + }, + { + "type": "table", + "bbox": [ + 0.303, + 0.071, + 0.49, + 0.201 + ], + "angle": 0, + "content": "
Text EncoderTEDSFood-101
404 1010
CLIP-Text Encoder✓73.55 83.41
BERT✓69.73 83.19
T5✓68.86 83.18
FastText✓72.80 83.13
CLIP-Text Encoder×70.40 82.63
BERT×67.67 82.44
T5×63.92 82.79
FastText×69.57 81.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.318, + 0.204, + 0.472, + 0.217 + ], + "angle": 0, + "content": "Table 3. Benefit of TEDS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.485, + 0.544 + ], + "angle": 0, + "content": "BERT model, while the T5 model performs the worst, particularly in the absence of TEDS. This performance discrepancy can be attributed to the different training objectives of these language models. The CLIP model is explicitly trained for text-image alignment using contrastive learning, creating a shared embedding space that directly supports text-vision alignment. This shared space makes the CLIP text encoder highly effective for guiding visual model training. BERT, while not designed specifically for text-vision alignment, captures rich textual representations that still contribute to the visual model learning process. In contrast, T5 is primarily designed for text-to-text tasks, and its training objectives are less suited for aligning visual and textual embeddings, resulting in its comparatively lower performance. Interestingly, FastText, a simple word embedding model, achieved strong performance by effectively extracting distinct text features from label names. Compared to general language models like BERT and T5, FastText focuses on subword structure rather than contextual meaning [2], allowing it to better preserve distinctions between label names and reduce confusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.744 + ], + "angle": 0, + "content": "The results also indicated that TEDS significantly enhances performance across different text encoders (Table 3). This improvement is expected, as raw text embeddings for different classes often exhibit high similarity (Figure 3), which can mislead the visual model during training, especially the label data is scare. Our TEDS module mitigates this issue by penalizing the similarity between text embeddings from different class names, reducing the impact of inherent similarities. Conversely, our method without TEDS lacks this mechanism, leading to poorer performance due to similarity-induced confusion. This underscores the importance of TEDS in making our method adaptable to various text encoder models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Benefit of CAViTA. We evaluate the effectiveness of CAViTA loss by comparing it to the widely used NT-Xent contrastive loss [35], which is commonly used for visual-text alignment in CLIP. Table 4 shows that CAViTA significantly outperforms NT-Xent loss. This result is expected because multiple images in a batch can share the same text embedding. NT-Xent mistakenly classified true positive pairs as negatives, penalizing them and leading to a performance drop. Table 6 reports the False negative rate under NT-Xent loss. It mistakenly penalizes \\(18.71\\%\\) of posi" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.071, + 0.683, + 0.134 + ], + "angle": 0, + "content": "
Text-visual Alignment LossFood-101
4041010
NT-Xent52.6175.22
CAViTA73.5583.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.145, + 0.692, + 0.201 + ], + "angle": 0, + "content": "Table 4. Benefit of CAViTA: CAViTA outperforms NT-Xent loss because it correctly define positive pairs." + }, + { + "type": "table", + "bbox": [ + 0.712, + 0.071, + 0.899, + 0.148 + ], + "angle": 0, + "content": "
λcλTEDSFood-101
10.0173.6
10.173.8
1172.8
0.10.0161.7
0.010.0154.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.71, + 0.158, + 0.9, + 0.187 + ], + "angle": 0, + "content": "Table 5. Hyperparameter study: our method is sensitivity to \\( {\\lambda }_{c} \\)" + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.225, + 0.876, + 0.272 + ], + "angle": 0, + "content": "
Dataset (labeled #)FN count(mean±std)FN ratio(mean±std)
Food 101 (404)7.65 ± 3.3418.71% ± 6.96%
EuroSAT (20)3.71 ± 2.0329.72% ± 11.79%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.277, + 0.868, + 0.291 + ], + "angle": 0, + "content": "Table 6. False negative (FN) rate under NT-Xent loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.905, + 0.42 + ], + "angle": 0, + "content": "tive pairs, leading to a \\(28.5\\%\\) accuracy drop on the Food101 dataset. Furthermore, applying NT-Xent loss within our method does not improve the base SSL method (i.e., FreeMatch) and instead decreases performance by \\(13.2\\%\\). This finding highlights the importance of correctly defining positive pairs, as misclassifications can significantly degrade model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Benefit of Using Label Name. We compared our method with longer text embeddings by extending label names using GPT-4o with the prompt: \"Describe the visual features of {label name} in 2-3 sentences\". These extended descriptions were then used as input to the text encoder to guide visual representation learning. However, using longer text embeddings generated by GPT-4o led to a drop in performance, with accuracy decreasing to \\(70.0\\%\\) on Food101 (404 labeled images) and \\(80.9\\%\\) on CIFAR-100 (200 labeled images), compared to using only the label name. This decline in performance is because long text embedding may include unnecessary tokens that dilute the core semantics and spread out embeddings unnecessarily, resulting in unnecessary confusion between different classes. In contrast, a short label name provides compact and discriminative semantics, which is more suitable for our classification task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Hyper-parameter Study. Since \\(\\lambda_{u}\\) has been studied in prior SSLs [36, 41], we focus on analyzing \\(\\lambda_{c}\\) and \\(\\lambda_{TEDS}\\) by varying them from 0.01 to 1 in Table 5. Our results indicate that the performance of our method is highly sensitive to \\(\\lambda_{c}\\). Increasing \\(\\lambda_{c}\\) improves performance, as a stronger visual-text alignment loss helps enforce sufficient regularization on visual representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.81, + 0.634, + 0.825 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We introduced SemiVisBooster to improve prior SSL by using semantics from label names to guide visual feature learning and achieved best performance on latest SSL benchmarks[40] and find-grained classification task." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1202" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.158 + ], + "angle": 0, + "content": "This work was supported by the National Institute of Biomedical Imaging and Bioengineering under NIH award numbers R01LM011834 and R56EB032819." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[1] Ardhendu Behera, Zachary Wharton, Pradeep Hewage, and Asish Bera. Context-aware attentional pooling (cap) for fine-grained visual classification. In The Thirty-Fifth AAAI Conference on Artificial Intelligence. AAAI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[2] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5, 2016. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461. Springer, 2014. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.456 + ], + "angle": 0, + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 194-203, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.526 + ], + "angle": 0, + "content": "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.528, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[6] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.666, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[9] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.231 + ], + "angle": 0, + "content": "[12] Andreas Fürst, Elisabeth Rumetshofer, Johannes Lehner, Viet T Tran, Fei Tang, Hubert Ramsauer, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto, et al. Cloob: Modern hopfield networks with infoloob outperform clip. Advances in neural information processing systems, 35: 20450-20468, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[13] Tanmay Gupta, Kevin Shih, Saurabh Singh, and Derek Hoiem. Aligned image-word representations improve inductive transfer across vision-language tasks. In 2017 IEEE International Conference on Computer Vision (ICCV), page 4223-4232. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.648 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[21] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[22] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks, page 121-137. Springer International Publishing, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1203" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[27] Khanh-Binh Nguyen. Debiasing, calibrating, and improving Semi-supervised Learning performance via simple Ensemble Projector. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2430-2439, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[28] Khanh-Binh Nguyen. Sequencematch revisiting the design of weak-strong augmentations for semi-supervised learning. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 96-105, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] OpenAI. CLIP: Contrastive Language-Image Pretraining. https://huggingface.co/openai/clip-vit-base-patch32, 2021. Accessed: 2024-11-10. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V. Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11557-11568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[33] Becca Roelofs, David Berthelot, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: a unified approach to semi-supervised learning and domain adaptation. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Hugo Schmutz, Olivier Humbert, and Pierre-Alexandre Mattei. Don't fear the unlabelled: safe semi-supervised learning via simple debiasing. arXiv preprint arXiv:2203.07512, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Proceedings of the 30th International Conference on Neural Information Processing Systems, page 1857-1865, Red Hook, NY, USA, 2016. Curran Associates Inc. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[37] Jong-Chyi Su and Subhransu Maji. The semi-supervised naturalist-aves challenge at fgvc7 workshop. arXiv preprint arXiv:2103.06937, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 1195–1204, Red Hook, NY, USA, 2017. Curran Associates Inc. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[40] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[41] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[42] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Semi-supervised-learning. https://github.com/microsoft/Semi-supervised-learning, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[43] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[45] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In International conference on machine learning, pages 11525-11536. PMLR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.825 + ], + "angle": 0, + "content": "[46] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in neural information processing systems, 34:18408-18419, 2021. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.896 + ], + "angle": 0, + "content": "[47] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14471-14481, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1204" + } + ] +] \ No newline at end of file diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_origin.pdf b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81a56be226b9ce8b216c8c2d677218e948960c27 --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/2707eef4-2e69-485f-a5ee-896dce310667_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb3eebf18143d8184d12fa00e6b3c7c7326d8ff66f5661087c45e419b8c366a6 +size 1822302 diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/full.md b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f7cef5707d314c0ae894a2306d2d116a3c34e50 --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/full.md @@ -0,0 +1,276 @@ +# SemiVisBooster: Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance + +Wenjin Zhang $^{1,2}$ Xinyu Li $^{2}$ Chenyang Gao $^{1,2}$ Ivan Marsic $^{1}$ $^{1}$ Rutgers University ${}^{2}$ Amazon + +{wz315, cg694}@scarletmail.rutgers.edu, lixinyu.arthur@outlook.com, marsic@rutgers.edu + +# Abstract + +Deep learning models rely on large-scale labeled datasets, but collecting such data is expensive and time-consuming. Semi-supervised learning (SSL) mitigates this issue by learning from a small set of labeled samples along with a large pool of unlabeled data. However, existing SSL methods struggle with fine-grained classification when dealing with visually similar classes, as they rely solely on visual features and ignore the semantic information within label names. This paper introduces SemiVisBooster, an SSL enhancement approach that utilizes semantic information from label names to guide visual feature learning, addressing the challenges of fine-grained classification. By aligning text embeddings from label names with visual features, our method helps the model capture subtle visual distinctions that purely visual representations may overlook. To enhance robustness, we propose two key components: (1) text embedding de-similarity (TEDS) to reduce confusion caused by similar text embeddings across different class names, and (2) class-aware visual-text alignment loss to accurately define positive and negative pairs during visual-text alignment. Our method achieves state-of-the-art performance on the latest SSL benchmarks. Additionally, on the challenging Food-101 dataset, which contains many visually similar classes and uses only 404 labeled images, our approach improves performance by approximately $13.6\%$ over the second-best method. Code is available at SemiVis-Booster + +# 1. Introduction + +The superior performance of deep learning heavily relies on supervised training with sufficient labeled data [14, 39]. However, acquiring such massive labeled datasets for real-world applications remains a labor-intensive and expensive process. To mitigate this challenge, semi-supervised learning (SSL) has emerged as a promising method, learning from a small set of labeled data while leveraging a + +![](images/1ce7554e4fc95fdf41ba04d72aa7921b42de1de71ac5fc46cb85fb88b168328b.jpg) +(a) FreeMatch + +![](images/cc6f4f95037ebd0f4ffb1d655c414e1141d935ddcbf5946fed93c9e50d553ff0.jpg) +(b) SemiVisBooster +Figure 1. Comparison of visual feature space distribution for (a) FreeMatch and (b) our method on the Food-101 dataset visualized using t-SNE. We only plot the first 20 classes. + +large amount of unlabeled data [5, 21, 36, 41, 46]. By significantly reducing the dependence on manual annotations, SSL has demonstrated great potential in practical applications. Despite achieving strong performance in general image classification, particularly on datasets with well-separated class boundaries such as CIFAR-10 [19] and STL-10 [6], SSL methods struggle with fine-grained classification, where visually similar categories exist. Analyzing the confusion matrix of the popular SSL method, FreeMatch [41], on the CIFAR-100 dataset with only 200 labeled images [42], we observed that its accuracy plateaued at $78.9\%$ due to frequent misclassifications among visually similar classes such as "woman," "man," "boy," and "girl" [40]. This observation indicates a key limitation of existing SSL approaches: their reliance on visual feature extraction, which often fails to distinguish subtle inter-class differences. To further validate the observed limitation on SSL, we evaluated FreeMatch on the fine-grained Food101 dataset [3] using 404 labeled images because this dataset contains many visually similar classes. FreeMatch achieved an accuracy of only $60.4\%$ , while a fully supervised training method using all available labeled data reached $98.8\%$ [1], highlighting a significant performance gap between SSL and fully supervised learning. This result underscores the need for improvements in SSL methods. The t-SNE visualization (Figure 1) illustrates that while FreeMatch exhibits some level of clustering, it struggles to separate visually similar classes, leading to significant feature overlap. This challenge arises from two key weaknesses in existing SSL methods. (1). + +SSL models rely primarily on visual patterns for classification. However, in fine-grained settings, where objects share highly similar appearances, these methods struggle to extract meaningful distinctions, resulting in frequent misclassification. (2). SSL methods depend heavily on pseudolabeling, where high-confidence predictions are used to assign labels to unlabeled samples. When dealing with visually ambiguous cases, incorrect pseudo-labels are generated and subsequently reinforce errors, leading to model bias and further performance degradation. + +In many real-world applications, class names are available as meaningful words or phrases, offering an opportunity to use the semantic information. However, existing SSL methods completely ignore the rich semantic information embedded in these labels. Instead, prior SSLs rely exclusively on one-hot-encoded categorical labels, which discard potentially useful semantic knowledge. We hypothesize that introducing label semantics into SSL can provide additional guidance for visual representation learning, thereby improving the model's ability to distinguish visually similar samples. For example, while "boy" and "man" may have highly similar visual representations in CIFAR-100 dataset, their semantic embeddings are distinct and different. By aligning visual features with distinct label semantics, SSL models can be improved in two ways. (1). Aligning text embeddings from ground-truth label names with the visual representations of labeled images can help the model better capture subtle visual differences through text cues. This makes the initial model more stable and robust. (2) Even if pseudo-labels are incorrect, aligning the text embedding of the pseudo-label with the visual representation of the unlabeled image can still be beneficial. If the incorrect pseudo-label's semantic meaning is close to the true label, this alignment may help guide the visual representation closer to the correct cluster. However, aligning text embeddings from label name with visual embeddings for SSL introduces two major challenges. The first challenge arises from the inherent similarity between text embeddings of different class names. When similar text embeddings are used to guide the visual representation learning, it can introduce unwanted confusion in the visual space (Section 3.4). The second challenge relates to many-to-one alignment in our settings. Unlike CLIP [31] using NT-Xent loss [35], where each image-text pair is uniquely defined, our setting often encounters multiple images within a batch that share the same text embedding because multiple images may belong to the same class. Directly applying existing visual-text alignment loss like NT-Xent [31], which assume a pairwise positive match, would mistakenly treat these samples as negatives, leading to incorrect optimization and degraded performance (Section 3.5). + +In this paper, we introduced SemiVisBooster, an SSL booster that improves SSL performance by incorpo + +rating label semantics into the learning process. To enrich feature learning beyond basic one-hot-encoded ground truth labels or pseudo-labels, we proposed a label semantic enrichment module that converts categorical labels into meaningful label names. To enable robust alignment, we proposed text embedding de-similarity (TEDS) loss to avoid the impact of inherent similarity between text embeddings and developed class-aware visual-text alignment (CAViTA) strategy, which correctly defines positive and negative pairs in cases where multiple images share the same text embedding. Our approach achieves state-of-the-art results on the latest SSL benchmarks [40]. Specifically, SemiVisBooster consistently surpasses the previous method by more than $2.9\%$ across multiple benchmarks, including CIFAR-100 (200 labeled images) and STL-10 (40 labeled images). For fine-grained classification tasks, our method achieved $73.6\%$ accuracy on Food 101 (404 labeled images), significantly surpassing the previous state-of-the-art method, which achieved only $64.8\%$ accuracy. We summarized our contributions below: + +- To the best of our knowledge, we are the first to directly leverage label name semantics to enhance SSL, achieving state-of-the-art performance on the latest SSL benchmarks and fine-grained classification tasks. +- We introduced TEDS to mitigate the inherent similarity between text embeddings of different class names, enhancing the robustness of our method across various text encoders, including a simple word embedding model. +- We designed CAViTA to accurately distinguish positive and negative visual-text embedding pairs, ensuring robust visual-text alignment. + +# 2. Related Work + +# 2.1. Semi-supervised Learning + +Semi-supervised learning (SSL) has been an area of extensive research over recent decades. Recent advances in SSL methodologies can be broadly categorized into two primary approaches: consistency regularization [20, 26, 38, 43] and pseudo-labeling [21, 30]. Consistency-based methods apply perturbations either to the input data or the model and enforce constraints to ensure consistency in the features or probability outputs. For example, the II-model [20] introduces dropout to model weights, generating two outputs for consistency regularization. Pseudo-labeling methods, in contrast, use model predictions as pseudo-labels to guide learning on unlabeled data. This approach has gained recent prominence [4, 21, 30, 44]. Combining these approaches, FixMatch [36] introduced a straightforward SSL framework that applies both pseudo-labeling and consistency regularization, but it uses a fixed confidence threshold. Recent work has focused on refining pseudo-labeling quality by introducing dynamic thresholds [45], class-specific thresh- + +![](images/ce24737cb60d26b409a7f0bf37af63fb61b4a07c0bb35b9dbc4a32d29a6de109.jpg) +Figure 2. Overview of SemiVisBooer framework for enhancing SSL through label semantic information. (Only the flow of unlabeled data is shown. Labeled data are used for both supervised learning and aligning ground truth labels with their corresponding images.) Base SSL branch: One-hot pseudo-labels are generated from weakly augmented images and used to guide predictions on strongly augmented versions of the same images. Boosting Branch: A label semantic enrichment module maps each categorical label to its label name and uses a pre-trained text encoder to extract text embeddings without any prompt operation. A learnable projector transforms all these embeddings into a text embedding bank. The text embedding de-similarity (TEDS) loss enhances robustness by minimizing the similarity between text embeddings of different classes. The class-aware visual-text alignment is used to align selected text embeddings with corresponding visual embeddings to guide visual representation learning. + +olds [46], and dynamic class-specific thresholds [5, 41]. However, even with these adaptive thresholding techniques, existing SSL methods struggle with fine-grained classification as we discussed in Section 1. To boost SSLs, we introduced additional supervision derived from the semantic information in label names. We used FreeMatch[41] as our base SSL. + +# 2.2. Vision and Language Alignment + +Numerous studies have demonstrated the benefits of joint image-text representation learning for various vision tasks [9, 13, 16, 23, 25]. Contrastive learning has proven particularly effective for vision-language models by mapping images and text into a shared feature space. Models like CLIP [31], CLOOB [12], ALIGN [17], and DeCLIP [24] achieve this by training text and image encoders on large datasets of image-caption pairs, using contrastive loss to minimize distances between matching pairs while maximizing them for non-matching ones. To handle noise in datasets, ALIGN [16] and DeCLIP [24] use custom metrics, while CLIP [31] employs cosine distance. While these models primarily aim at pre-training large vision-language foundations for zero- and few-shot tasks, our approach enhances semi-supervised learning by aligning visual features + +with pseudo-labels, improving the robustness of SSLs for fine-grained classification. + +# 3. Method + +# 3.1. Preliminaries + +Following [36], we formulate the SSL framework for a $C$ -class classification problem. Let $\mathcal{D}^l = \{(x_i^l, y_i^l) : i \in [1, B_L]\}$ denote a batch of labeled data, and $\mathcal{D}^u = \{x_j^u : j \in [1, B_U]\}$ represent a batch of unlabeled data, with batch sizes $B_L$ and $B_U$ , respectively. Here, $x_i^l$ denotes a labeled image with label $y_i^l$ , and $x_j^u$ represents an unlabeled image. The model is trained using a combined objective: $\mathcal{L}_{\mathrm{total}} = \mathcal{L}_s + \lambda_u \mathcal{L}_u$ , where $\mathcal{L}_s$ is the supervised loss over labeled data, $\mathcal{L}_u$ is the unsupervised loss over unlabeled data, and $\lambda_u$ is the unsupervised loss weight. The supervised loss $\mathcal{L}_s$ is defined as: + +$$ +\mathcal {L} _ {s} = \frac {1}{B _ {L}} \sum_ {i = 1} ^ {B _ {L}} \mathcal {H} \left(y _ {i} ^ {l}, \mathcal {F} _ {m} \left(y \mid w \left(x _ {i} ^ {l}\right)\right)\right) \tag {1} +$$ + +where $\mathcal{H}(\cdot, \cdot)$ is cross-entropy loss, $w(\cdot)$ is weak augmentation (i.e., random crop and flip) and $\mathcal{F}_m(\cdot)$ is the output probability from the model. + +For unlabeled data, SSL typically dynamically assigns the pseudo-labels if the model's confidence scores surpass the thresholds. The unsupervised loss $\mathcal{L}_u$ is defined as: + +$$ +\mathcal {L} _ {u} = \frac {1}{B _ {U}} \sum_ {j = 1} ^ {B _ {U}} \mathbb {1} \left(\max \left(q _ {j}\right) > \tau\right) \mathcal {H} \left(\hat {q} _ {j}, Q _ {j}\right) \tag {2} +$$ + +where $q_{j}$ and $Q_{j}$ denote abbreviation of $\mathcal{F}_m(y|w(x_j^u))$ and $\mathcal{F}_m(y|\mathcal{A}(x_j^u))$ , respectively. $\hat{q}_j = \arg \max (q_j)$ is the one-hot pseudo-label derived from the model's prediction on weakly augmented image. $\mathcal{A}(\cdot)$ denotes strong image augmentation [7], and $\mathbb{1}(\cdot)$ is an indicator function that selects pseudo-labels based on a threshold $\tau$ . Our method enhances SSL by enriching pseudo-labels with semantic information to guide visual representation learning. + +# 3.2. Model Overview + +Our method enhances the base SSL by incorporating additional guidance from label semantics (Figure 2). In the base SSL branch, weakly augmented images pass through the model first, generating one-hot pseudo-labels based on confidence thresholds. These pseudo-labels then guide the model to predict the same labels for strongly augmented versions of the corresponding unlabeled images [5, 36, 41]. In the boosting branch, label semantic enrichment module performs a mapping operation from categorical label index to its corresponding label name. A pre-trained text encoder extracts text embeddings from these label names, and a learnable projector transforms them into projected text embeddings, named text embedding bank, that covers all classes. To enhance robustness, we apply text embedding de-similarity (TEDS) loss, $\mathcal{L}_{TEDS}$ , which encourages dissimilarity between different text embeddings. Finally, the one-hot pseudo-labels from the base SSL branch are used to select corresponding text embeddings, which are then aligned with the corresponding visual embeddings through class-aware visual-text alignment loss, $\mathcal{L}_c$ . The overall training objective is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {s} + \lambda_ {u} \mathcal {L} _ {u} + \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {T E D S} * \mathcal {L} _ {T E D S}, \tag {3} +$$ + +where $\lambda_{c}$ and $\lambda_{TEDS}$ are the weighting factors for the text embedding de-similarity loss and the class-aware visual-text alignment loss, respectively. + +# 3.3. Label Semantic Enrichment + +This component converts categorical labels into meaningful words or phrases, allowing the subsequent text encoder module to capture semantic text embeddings. Unlike CLIP [31], which achieves text-image alignment using millions of (image, text) pairs, this is impractical in SSL due to limited annotations. In SSL, we often lack even basic categorical labels, let alone detailed image descriptions. + +![](images/ae1ddcecf2c1f9ca9c2bdcab7ed798661789768a2da9440c27d9fcd51a41f977.jpg) +T5 + +![](images/a51bb61f6922a4d879797190692f113bfdc768a6adc3bcfafc063e2787a67e4c.jpg) +BERT + +![](images/efc17289a0f474f48a1e37dc9b2226386b15db3b2e7a5e49d869005fb2d34a09.jpg) +T5 +Figure 3. t-SNE visualization of text embeddings for Food101 label names. The first row shows raw text embeddings generated by T5 and BERT. The second row is the embeddings after applying TEDS, resulting in more dispersed and dissimilar embeddings. + +![](images/2dcc0713ac24088bb3b08002f9244169e01864666237dd9eaf904cc94e44b9f1.jpg) +BERT + +However, since we have prior knowledge of all label names, our label semantic enrichment module performs a straightforward and fixed mapping from categorical pseudo-labels to their respective semantic names. This fixed mapping also contributes to the efficiency (discussed in Section 3.6). We directly feed the label name to the text encoder without any prompt operations. Using only label names ensures simplicity and reproducibility. + +# 3.4. Text Embedding De-Similarity + +Text embeddings generated by a text encoder are not inherently suitable for guiding visual representation learning because they lack sufficient separation between different classes (see the first row of Figure 3). To quantify this, we calculated the cosine similarity between text embeddings of different class names using the BERT model. The results showed high similarity scores (above 0.92) between classes such as "Carrot Cake" and "Chocolate Cake", as well as "Cheesecake" and "Chocolate Cake". This inherent similarity in text embeddings unintentionally pushes corresponding visual representations closer together during visual-text alignment, leading to increased confusion between semantic similar classes. + +To mitigate this issue, we introduce text embedding desimilarity (TEDS), a loss function that encourages greater separation between text embeddings by minimizing the absolute values of pairwise dot products between projected text embeddings. The TEDS loss is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {T E D S}} = \frac {1}{C (C - 1)} \sum_ {i \neq j} \max \left(\left| \frac {\mathbf {T} _ {i} ^ {\top} \mathbf {T} _ {j}}{\| \mathbf {T} _ {i} \| \| \mathbf {T} _ {j} \|} \right| - \Delta , 0\right), \tag {4} +$$ + +where $C$ is the number of classes, $\mathbf{T}_i$ and $\mathbf{T}_j$ are individual projected text embeddings, $\| \mathbf{T}_i\|$ and $\| \mathbf{T}_j\|$ represent their + +![](images/bf06f05d0cb88275e6b343ede7905b487e71f847713a5c1632fc4eba113a32f4.jpg) +(a). NT-Xent +Figure 4. Comparison between NT-Xent loss and class-aware contrastive loss. + +![](images/43175bce05b96c45c9698f444323dcd3c70b6128397d1c7f57a43c24718375b7.jpg) +(b). Class-aware Contrastive + +magnitudes and $\Delta$ is a tolerance margin (e.g., $1 \times 10^{-5}$ ), defining the lowest acceptable similarity level. Integrating TEDS into the training objective encourages the learnable projector to extract more distinct and discriminative features from text embeddings. Since the text encoder remains frozen, TEDS primarily trains the learnable projector to enhance the separation of projected text embeddings. After applying TEDS, t-SNE visualization (see the second row of Figure 3) confirms that the projected text embeddings become more widely separated, which in turn helps the model learn more distinguishable visual representations. + +# 3.5. Class-aware Visual-text Alignment + +In [31], the NT-Xent-based visual-text alignment forms positive pairs based on specific fine-grained text-visual matches, where each image is paired with a unique text description, and all other text descriptions in the batch are treated as negatives. However, this approach is suboptimal in our SSL context, where a batch may contain multiple images from the same class (i.e., sharing the same text embedding). For example, if one batch has two visual $(V_{0}, V_{1})$ and their corresponding labels $(T_{0}, T_{1})$ where $T_{0} = T_{1}$ , the existing text-visual alignment loss would incorrectly treat $(V_{0}, T_{1})$ and $(V_{1}, T_{0})$ as negative pairs—False Negative Pairs (Figure 4). To address this, we introduce class-aware visual-text alignment which leverages class labels to define positive and negative pairs. Specifically, positive pairs are formed between text-visual pairs from the same class, while negative pairs come from different classes. This is implemented by computing a similarity matrix between visual and text embeddings and applying a mask to identify positive pairs based on class labels. For each visual-text pair within the same class, we maximize similarity, while minimizing similarity for pairs from different classes. The class-aware visual-text alignment loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {c}} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \left(\frac {\sum_ {j = 1} ^ {N} \exp \left(\frac {\tilde {\mathbf {V}} _ {i} \cdot \tilde {\mathbf {T}} _ {j}}{\kappa}\right) \cdot \mathbf {M} _ {i j}}{\sum_ {j = 1} ^ {N} \exp \left(\frac {\tilde {\mathbf {V}} _ {i} \cdot \tilde {\mathbf {T}} _ {j}}{\kappa}\right)}\right) \tag {5} +$$ + +where $N$ represents the number of embeddings, $\tilde{\mathbf{V}}_i$ is the normalized visual embedding of sample $i$ , $\tilde{\mathbf{T}}_j$ is the normalized text embedding of sample $j$ , $\kappa$ is the temperature scaling factor and $\mathbf{M}_{ij}$ is a class-aware mask where $\mathbf{M}_{ij} = 1$ if samples $i$ and $j$ belong to the same class, and $\mathbf{M}_{ij} = 0$ otherwise. + +# 3.6. Implementation Detail + +Text Encoder. The text encoder plays a crucial role in extracting semantic information from label names to guide visual representation learning. The key requirement is the ability to capture meaningful semantics. Any text encoder that effectively extracts distinct and informative text embeddings can serve this purpose. To ensure flexibility and broad applicability, our method is designed to work with both language models and simple word embedding models that can generate semantically meaningful representations. We evaluate our approach using three language models—T5 [32], BERT [8], and CLIP's text encoder [31]—as well as a lightweight word embedding model, fastText [2]. Among these, the pre-trained CLIP text encoder is particularly well-suited for this task, as it has been specifically trained for visual-text alignment, making it highly effective in guiding visual feature learning. Ablation studies confirm that our method generalizes well across different text encoders, demonstrating its robustness and adaptability across various models. + +Projection Strategies. In the SSL context, both the visual network architecture and the text encoder can be chosen freely, leading to a dimensional mismatch between text embeddings and visual embeddings. This mismatch poses a challenge for visual-text alignment, as the two modalities exist in different feature spaces. To address this issue, we explore two projection strategies that project text or visual embeddings to the same feature dimension. The first strategy, ProjT2V, is the approach used in Figure 2. In this method, a learnable projection layer is applied to the text embeddings, transforming them into the same dimensional space as the visual embeddings [16]. This transformation enables direct alignment between the projected text embeddings and the visual embeddings using contrastive loss. The second strategy, Proj-V2T, takes the opposite approach by projecting visual embeddings into the dimensional space of the text embeddings [11]. While both strategies successfully create a shared embedding space, our experiment shows that ProjT2V outperforms Proj-V2T, providing a stronger improvement in SSL performance and more effective visual-text alignment. + +Projector Architecture. We implemented the projector architecture by considering two main factors. First, the learnable projector should present a challenging alignment task for the in-training visual feature extraction model. To achieve this, the projector cannot be overly complex, as + +that would exert too much influence on the visual or text embeddings, potentially distorting their features. Second, we aimed to keep the architecture lightweight and efficient, motivating us to choose a simpler design. Based on these requirements and existing works usually use linear projections [11, 16], we explored two one-layer architectures: a single-layer MLP and self-attention. + +Efficient Implementation. Our method can achieve high efficiency with minimal vRAM usage and without adding significant computation when boosting SSLs for three main reasons. First, by freezing the text encoder, we only perform text encoder inference without the computationally expensive gradient calculations. Second, due to the fixed mapping from categorical labels to label names, the input to the text encoder model is limited to the number of label names. This allows text embeddings to be precomputed once during initialization and stored in a cache, eliminating the need for repeated computations during training. As a result, the text encoder can be offloaded from vRAM, and the necessary text embeddings can be efficiently retrieved from the pre-computed cache. Finally, we reuse the visual features already extracted by the base SSL for alignment, further optimizing efficiency. These optimizations ensure that our approach enhances SSL performance without introducing significant computational cost or memory overhead. + +# 4. Experiment + +# 4.1. Setup + +Datasets. We evaluated our method on the latest SSL benchmarks [40]. To further evaluate the robustness on fine-grained classification task, we also tested our method on Food 101 dataset [3] as prior SSLs fail on this dataset. For each dataset, we examined two different quantities of labeled images. Detailed settings for each dataset are provided in the Appendix. For instance, we used 404 and 1010 labeled images for the Food-101 dataset. Labeled samples were randomly selected from the training set, except for STL-10 and Semi-Aves. For these two datasets, the labeled and unlabeled splits are predefined. + +Experimental Setup. Our method was based on FreeMatch [41] and implemented with a widely-used SSL codebase, USB [40]. For a fair comparison with baseline methods, we maintained consistent training settings across methods, adjusting only SSL-specific hyperparameters. This approach ensured that all SSLs use the same visual network architectures and pre-training strategies. For the base SSL settings, we initialized the self-adaptive threshold for FreeMatch at 0, whereas FreeMatch typically uses $\frac{1}{C}$ . In our boosting configuration, we set the weight for the alignment loss term, $\lambda_{c}$ , to 1 with a temperature factor of 0.07. We also assigned the weight of the text embedding de-similarity term, $\lambda_{TEDS}$ to 0.01. Our model train + +ing used the AdamW optimizer [18] with a cosine annealing learning rate schedule and an initial warmup phase. The training spanned 204,800 steps in total, with the first 10,240 steps dedicated to warmup. Batch sizes were adjusted based on the dataset, with labeled and unlabeled samples sharing the same batch size within each dataset. Specifically, batch sizes were set to 32 for Food-101, CIFAR-100, and STL-10; 64 for Semi-Aves; and 8 for EuroSAT. The base learning rates are also dataset-specific: $1e - 4$ for Food-101 and STL-10, $5e - 4$ for CIFAR-100 and Semi-Aves, and $5e - 5$ for EuroSAT. + +# 4.2. Overall Result + +To the best of our knowledge, we compared our method with SSL algorithms in the past 4 years, as shown in Table 1. Our approach consistently outperformed prior SSL methods, particularly in fine-grained classification tasks. For example, SemiVisBooster achieved $73.6\%$ accuracy on Food-101 using only 404 labeled images and $81.8\%$ accuracy on CIFAR-100 with only 200 labeled images, outperforming the second-best method by margins of $13\%$ and $2.9\%$ , respectively. The improvement on Food-101 is more substantial than on CIFAR-100 because CIFAR-100 contains only some fine-grained classes, and existing SSL methods may effectively distinguish most classes with clear boundaries in CIFAR-100. On the Semi-Aves dataset, our method achieved a $2.8\%$ accuracy improvement over the next best method. In coarse-grained classification tasks, our method also demonstrated strong results, achieving $95.4\%$ accuracy on the STL-10 dataset with just 20 labeled images and $96.8\%$ accuracy on the EuroSAT dataset with the same number of labeled images. These results show that our method is not only effective for fine-grained classification but also enhances performance in general coarse-grained classification problems. + +Comparison with CLIP. To assess the necessity of training a visual model using SSL, we compared our method against CLIP's zero-shot performance [31] and few-shot fine-tuning results. Despite using a smaller visual backbone (ViT Small [40]), our method outperformed both the CLIP-ViT Base and CLIP-ViT Large model on the CIFAR-100 and EuroSAT datasets, where CLIP lacks domain-specific adaptation (Table 1). CLIP performed impressively on Food-101 and STL-10, likely because its training on 400 million image-text pairs exposes it to common food items and general object categories present in these datasets [31]. We further fine-tune CLIP vision encoder with a classifier head on EuroSAT dataset with 20 labeled images. The fine-tuned CLIP-ViT Base and CLIP-ViT Large model achieved $84.63\%$ and $79.64\%$ accuracy, respectively. In contrast, our compact model—using only $24\%$ of CLIP-Base's vision parameters and training from scratch—achieves $96.8\%$ under USB [40] settings. This comparison highlights the chal + +
DatasetFood-101 [3]Semi-Aves [37]CIFAR-100 [19]STL-10 [6]EuroSAT [15]
# of labels40410103959200400401002040
SSLUDA [43]53.274.368.371.881.885.692.592.395.1
FixMatch [36]54.764.568.271.381.085.792.690.194.1
CoMatch [22]30.648.961.565.675.285.890.894.795.2
Dash [45]57.776.767.870.480.785.792.989.793.9
CRMatch [10]44.469.868.076.182.289.888.588.593.4
AdaMatch [33]60.176.668.479.484.087.993.393.895.4
FlexMatch [46]48.471.867.774.482.188.792.695.495.2
SimMatch [47]55.473.166.277.383.788.493.392.995.6
DeSSL [34]56.675.068.270.380.684.392.491.896.3
FreeMatch [41]60.678.867.578.984.688.592.094.394.7
SoftMatch [5]60.476.968.278.783.887.692.994.995.5
EPASS [27]64.877.665.977.882.691.593.882.390.1
SequenceMatch [28]43.767.968.170.177.486.489.692.992.6
SemiVisBooster73.683.870.381.885.495.495.996.896.7
Zero-shotCLIP-ViT-B[31]84.410.1*65.197.249.4
CLIP-ViT-L[31]92.99.9*77.999.359.9
+ +Table 1. Top 1 Accuracy: Bold indicates the highest accuracy and underline indicates the second highest accuracy. Star* marks the result from hugging face model inference [29]. + +lenges general models face in specialized domains and underscores the importance of continuously improving SSL methods. + +# 4.3. Ablations + +We conducted the ablation studies using the following setup, unless specified: the projection strategy was ProjT2V, the projector architecture was an MLP, the text encoder was the CLIP text encoder, we used both text embedding de-similarity loss and class-aware visual-text alignment strategy for training, and the input to text encoder is the label name. + +Projection Strategies and Projector Architecture. We study the impact of two commonly used projection methods: ProjT2V, and Proj-V2T. Following the approach in [16], we implement the ProjT2V module using a single-layer MLP with the text embedding dimension as the input and the visual embedding dimension as the output. Similarly, for the Proj-V2T module, we use a single-layer MLP, but with the input and output dimensions reversed. For self-attention, we use a one-layer self-attention module with a hidden dimension of 512, followed by a linear MLP with an input dimension of 512 and an output dimension matching the visual embedding dimension. + +From Table 2, we noticed that ProjT2V works best on both setups, while Proj-V2T lags behind. This is expected as ProjT2V directly enhances visual representation learning, whereas Proj-V2T lacks this capability. With ProjV2T, the alignment loss only ensures that the projected visual embeddings resemble the text embeddings, but this approach has two limitations. First, ProjV2T does not address inherent similarities across text embeddings from different classes, which can cause class overlap. Second, even if text embeddings are distinct, ProjV2T does not ensure + +that the pre-projection visual embeddings (i.e. the raw output from the vision model) will naturally align with the text embeddings. This is because the projector itself influences the alignment target, rather than allowing the vision model to learn distinct representations independently. In contrast, ProjT2V promotes direct visual representation learning by projecting the text embeddings and aligning them with the raw visual embeddings directly. Additionally, we ensure that text embeddings are distinct using the TEDS. + +It is worth mentioning that we also ablate the projector used for T2V and MLP works slightly better, probably due to the simplicity of the MLP layer. It creates a more challenging task for the visual-text alignment, encouraging more robust visual feature extraction. These results support our decision to use Proj-T2V and MLP as projector. + +Benefit of TEDS and Impact of Text Encoders. We evaluate the impact of TEDS by testing our method both with and without TEDS across different text encoders. To implement our method without TEDS, we set $\lambda_{t} = 0$ . The pre-trained text encoder models are from HuggingFace repositories. + +Our method consistently outperforms the baseline across different text encoders, including the simple word embedding model, FastText (Table 3). The key to boost SSL is incorporating semantic information to guide visual representations. Thus, our method is not restricted to a specific text model but is adaptable to various text encoders, ranging from lightweight word embedding models to language models, as long as they can extract meaningful semantics from label names. This flexibility enhances the practical applicability of our method in real-world scenarios. + +By comparing different text encoders, we found that the CLIP text encoder achieves the highest performance on Food-101 (404 labeled images), followed by fastText and + +
Projection StrategiesProjectorFood-101
4041010
Proj-V2TMLP60.5080.38
Proj-T2VSelf-attention72.5583.83
Proj-T2VMLP73.5583.41
+ +Table 2. Choice of Projector Strategies: Proj-T2V with MLP project achieves the best performance. + +
Text EncoderTEDSFood-101
404 1010
CLIP-Text Encoder✓73.55 83.41
BERT✓69.73 83.19
T5✓68.86 83.18
FastText✓72.80 83.13
CLIP-Text Encoder×70.40 82.63
BERT×67.67 82.44
T5×63.92 82.79
FastText×69.57 81.25
+ +BERT model, while the T5 model performs the worst, particularly in the absence of TEDS. This performance discrepancy can be attributed to the different training objectives of these language models. The CLIP model is explicitly trained for text-image alignment using contrastive learning, creating a shared embedding space that directly supports text-vision alignment. This shared space makes the CLIP text encoder highly effective for guiding visual model training. BERT, while not designed specifically for text-vision alignment, captures rich textual representations that still contribute to the visual model learning process. In contrast, T5 is primarily designed for text-to-text tasks, and its training objectives are less suited for aligning visual and textual embeddings, resulting in its comparatively lower performance. Interestingly, FastText, a simple word embedding model, achieved strong performance by effectively extracting distinct text features from label names. Compared to general language models like BERT and T5, FastText focuses on subword structure rather than contextual meaning [2], allowing it to better preserve distinctions between label names and reduce confusion. + +The results also indicated that TEDS significantly enhances performance across different text encoders (Table 3). This improvement is expected, as raw text embeddings for different classes often exhibit high similarity (Figure 3), which can mislead the visual model during training, especially the label data is scare. Our TEDS module mitigates this issue by penalizing the similarity between text embeddings from different class names, reducing the impact of inherent similarities. Conversely, our method without TEDS lacks this mechanism, leading to poorer performance due to similarity-induced confusion. This underscores the importance of TEDS in making our method adaptable to various text encoder models. + +Benefit of CAViTA. We evaluate the effectiveness of CAViTA loss by comparing it to the widely used NT-Xent contrastive loss [35], which is commonly used for visual-text alignment in CLIP. Table 4 shows that CAViTA significantly outperforms NT-Xent loss. This result is expected because multiple images in a batch can share the same text embedding. NT-Xent mistakenly classified true positive pairs as negatives, penalizing them and leading to a performance drop. Table 6 reports the False negative rate under NT-Xent loss. It mistakenly penalizes $18.71\%$ of posi + +Table 3. Benefit of TEDS + +
Text-visual Alignment LossFood-101
4041010
NT-Xent52.6175.22
CAViTA73.5583.41
+ +Table 4. Benefit of CAViTA: CAViTA outperforms NT-Xent loss because it correctly define positive pairs. + +
λcλTEDSFood-101
10.0173.6
10.173.8
1172.8
0.10.0161.7
0.010.0154.1
+ +Table 5. Hyperparameter study: our method is sensitivity to ${\lambda }_{c}$ + +
Dataset (labeled #)FN count(mean±std)FN ratio(mean±std)
Food 101 (404)7.65 ± 3.3418.71% ± 6.96%
EuroSAT (20)3.71 ± 2.0329.72% ± 11.79%
+ +Table 6. False negative (FN) rate under NT-Xent loss + +tive pairs, leading to a $28.5\%$ accuracy drop on the Food101 dataset. Furthermore, applying NT-Xent loss within our method does not improve the base SSL method (i.e., FreeMatch) and instead decreases performance by $13.2\%$ . This finding highlights the importance of correctly defining positive pairs, as misclassifications can significantly degrade model performance. + +Benefit of Using Label Name. We compared our method with longer text embeddings by extending label names using GPT-4o with the prompt: "Describe the visual features of {label name} in 2-3 sentences". These extended descriptions were then used as input to the text encoder to guide visual representation learning. However, using longer text embeddings generated by GPT-4o led to a drop in performance, with accuracy decreasing to $70.0\%$ on Food101 (404 labeled images) and $80.9\%$ on CIFAR-100 (200 labeled images), compared to using only the label name. This decline in performance is because long text embedding may include unnecessary tokens that dilute the core semantics and spread out embeddings unnecessarily, resulting in unnecessary confusion between different classes. In contrast, a short label name provides compact and discriminative semantics, which is more suitable for our classification task. + +Hyper-parameter Study. Since $\lambda_{u}$ has been studied in prior SSLs [36, 41], we focus on analyzing $\lambda_{c}$ and $\lambda_{TEDS}$ by varying them from 0.01 to 1 in Table 5. Our results indicate that the performance of our method is highly sensitive to $\lambda_{c}$ . Increasing $\lambda_{c}$ improves performance, as a stronger visual-text alignment loss helps enforce sufficient regularization on visual representations. + +# 5. Conclusion + +We introduced SemiVisBooster to improve prior SSL by using semantics from label names to guide visual feature learning and achieved best performance on latest SSL benchmarks[40] and find-grained classification task. + +# Acknowledgments + +# References + +[1] Ardhendu Behera, Zachary Wharton, Pradeep Hewage, and Asish Bera. Context-aware attentional pooling (cap) for fine-grained visual classification. In The Thirty-Fifth AAAI Conference on Artificial Intelligence. AAAI, 2021. 1 +[2] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5, 2016. 5, 8 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461. Springer, 2014. 1, 6, 7 +[4] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 194-203, 2021. 2 +[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 4, 7 +[6] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR. 1, 7 +[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 4 +[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics, 2019. 5 +[9] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 3 +[10] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 7 + +This work was supported by the National Institute of Biomedical Imaging and Bioengineering under NIH award numbers R01LM011834 and R56EB032819. +[11] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 5, 6 +[12] Andreas Fürst, Elisabeth Rumetshofer, Johannes Lehner, Viet T Tran, Fei Tang, Hubert Ramsauer, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto, et al. Cloob: Modern hopfield networks with infoloob outperform clip. Advances in neural information processing systems, 35: 20450-20468, 2022. 3 +[13] Tanmay Gupta, Kevin Shih, Saurabh Singh, and Derek Hoiem. Aligned image-word representations improve inductive transfer across vision-language tasks. In 2017 IEEE International Conference on Computer Vision (ICCV), page 4223-4232. IEEE, 2017. 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 3, 5, 6, 7 +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 7 +[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2 +[21] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 1, 2 +[22] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 7 +[23] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks, page 121-137. Springer International Publishing, 2020. 3 + +[24] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021.3 +[25] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2 +[27] Khanh-Binh Nguyen. Debiasing, calibrating, and improving Semi-supervised Learning performance via simple Ensemble Projector. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2430-2439, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 7 +[28] Khanh-Binh Nguyen. Sequencematch revisiting the design of weak-strong augmentations for semi-supervised learning. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 96-105, 2024. 7 +[29] OpenAI. CLIP: Contrastive Language-Image Pretraining. https://huggingface.co/openai/clip-vit-base-patch32, 2021. Accessed: 2024-11-10. 7 +[30] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V. Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11557-11568, 2021. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2, 3, 4, 5, 6, 7 +[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5 +[33] Becca Roelofs, David Berthelot, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: a unified approach to semi-supervised learning and domain adaptation. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2022. 7 +[34] Hugo Schmutz, Olivier Humbert, and Pierre-Alexandre Mattei. Don't fear the unlabelled: safe semi-supervised learning via simple debiasing. arXiv preprint arXiv:2203.07512, 2022. 7 +[35] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Proceedings of the 30th International Conference on Neural Information Processing Systems, page 1857-1865, Red Hook, NY, USA, 2016. Curran Associates Inc. 2, 8 +[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, + +Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3, 4, 7, 8 +[37] Jong-Chyi Su and Subhransu Maji. The semi-supervised naturalist-aves challenge at fgvc7 workshop. arXiv preprint arXiv:2103.06937, 2021. 7 +[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 1195–1204, Red Hook, NY, USA, 2017. Curran Associates Inc. 2 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1 +[40] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 1, 2, 6, 8 +[41] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 1, 3, 4, 6, 7, 8 +[42] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Semi-supervised-learning. https://github.com/microsoft/Semi-supervised-learning, 2023.1 +[43] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2, 7 +[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2 +[45] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In International conference on machine learning, pages 11525-11536. PMLR, 2021. 2, 7 +[46] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in neural information processing systems, 34:18408-18419, 2021. 1, 3, 7 +[47] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14471-14481, 2022. 7 \ No newline at end of file diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/images.zip b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51a9d06442f69972ef717c12526f69bac502a0c5 --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1e10dc7bd9136a81bf201f4d38818f3afb4402ae8ac443f9eb2edb1df2ec72b +size 447307 diff --git a/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/layout.json b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e0b34112a6556d480a589bca2257cb5d67c4b789 --- /dev/null +++ b/2025/SemiVisBooster_ Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance/layout.json @@ -0,0 +1,8012 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "type": "text", + "content": "SemiVisBooster: Boosting Semi-Supervised Learning for Fine-Grained Classification through Pseudo-Label Semantic Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": "Wenjin Zhang" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": " Xinyu Li" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": " Chenyang Gao" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": " Ivan Marsic" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": "Rutgers University " + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 145, + 160, + 463, + 190 + ], + "type": "text", + "content": "Amazon" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 191, + 539, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 191, + 539, + 204 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 539, + 204 + ], + "type": "text", + "content": "{wz315, cg694}@scarletmail.rutgers.edu, lixinyu.arthur@outlook.com, marsic@rutgers.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 258, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 579 + ], + "type": "text", + "content": "Deep learning models rely on large-scale labeled datasets, but collecting such data is expensive and time-consuming. Semi-supervised learning (SSL) mitigates this issue by learning from a small set of labeled samples along with a large pool of unlabeled data. However, existing SSL methods struggle with fine-grained classification when dealing with visually similar classes, as they rely solely on visual features and ignore the semantic information within label names. This paper introduces SemiVisBooster, an SSL enhancement approach that utilizes semantic information from label names to guide visual feature learning, addressing the challenges of fine-grained classification. By aligning text embeddings from label names with visual features, our method helps the model capture subtle visual distinctions that purely visual representations may overlook. To enhance robustness, we propose two key components: (1) text embedding de-similarity (TEDS) to reduce confusion caused by similar text embeddings across different class names, and (2) class-aware visual-text alignment loss to accurately define positive and negative pairs during visual-text alignment. Our method achieves state-of-the-art performance on the latest SSL benchmarks. Additionally, on the challenging Food-101 dataset, which contains many visually similar classes and uses only 404 labeled images, our approach improves performance by approximately " + }, + { + "bbox": [ + 54, + 258, + 297, + 579 + ], + "type": "inline_equation", + "content": "13.6\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 579 + ], + "type": "text", + "content": " over the second-best method. Code is available at SemiVis-Booster" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 608, + 135, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 135, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 135, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "The superior performance of deep learning heavily relies on supervised training with sufficient labeled data [14, 39]. However, acquiring such massive labeled datasets for real-world applications remains a labor-intensive and expensive process. To mitigate this challenge, semi-supervised learning (SSL) has emerged as a promising method, learning from a small set of labeled data while leveraging a" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 230, + 432, + 309 + ], + "blocks": [ + { + "bbox": [ + 339, + 230, + 432, + 309 + ], + "lines": [ + { + "bbox": [ + 339, + 230, + 432, + 309 + ], + "spans": [ + { + "bbox": [ + 339, + 230, + 432, + 309 + ], + "type": "image", + "image_path": "1ce7554e4fc95fdf41ba04d72aa7921b42de1de71ac5fc46cb85fb88b168328b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 312, + 399, + 320 + ], + "lines": [ + { + "bbox": [ + 361, + 312, + 399, + 320 + ], + "spans": [ + { + "bbox": [ + 361, + 312, + 399, + 320 + ], + "type": "text", + "content": "(a) FreeMatch" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 230, + 529, + 309 + ], + "blocks": [ + { + "bbox": [ + 433, + 230, + 529, + 309 + ], + "lines": [ + { + "bbox": [ + 433, + 230, + 529, + 309 + ], + "spans": [ + { + "bbox": [ + 433, + 230, + 529, + 309 + ], + "type": "image", + "image_path": "cc6f4f95037ebd0f4ffb1d655c414e1141d935ddcbf5946fed93c9e50d553ff0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 312, + 507, + 320 + ], + "lines": [ + { + "bbox": [ + 455, + 312, + 507, + 320 + ], + "spans": [ + { + "bbox": [ + 455, + 312, + 507, + 320 + ], + "type": "text", + "content": "(b) SemiVisBooster" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 320, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 352 + ], + "type": "text", + "content": "Figure 1. Comparison of visual feature space distribution for (a) FreeMatch and (b) our method on the Food-101 dataset visualized using t-SNE. We only plot the first 20 classes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": "large amount of unlabeled data [5, 21, 36, 41, 46]. By significantly reducing the dependence on manual annotations, SSL has demonstrated great potential in practical applications. Despite achieving strong performance in general image classification, particularly on datasets with well-separated class boundaries such as CIFAR-10 [19] and STL-10 [6], SSL methods struggle with fine-grained classification, where visually similar categories exist. Analyzing the confusion matrix of the popular SSL method, FreeMatch [41], on the CIFAR-100 dataset with only 200 labeled images [42], we observed that its accuracy plateaued at " + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "inline_equation", + "content": "78.9\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": " due to frequent misclassifications among visually similar classes such as \"woman,\" \"man,\" \"boy,\" and \"girl\" [40]. This observation indicates a key limitation of existing SSL approaches: their reliance on visual feature extraction, which often fails to distinguish subtle inter-class differences. To further validate the observed limitation on SSL, we evaluated FreeMatch on the fine-grained Food101 dataset [3] using 404 labeled images because this dataset contains many visually similar classes. FreeMatch achieved an accuracy of only " + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "inline_equation", + "content": "60.4\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": ", while a fully supervised training method using all available labeled data reached " + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "inline_equation", + "content": "98.8\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 715 + ], + "type": "text", + "content": " [1], highlighting a significant performance gap between SSL and fully supervised learning. This result underscores the need for improvements in SSL methods. The t-SNE visualization (Figure 1) illustrates that while FreeMatch exhibits some level of clustering, it struggles to separate visually similar classes, leading to significant feature overlap. This challenge arises from two key weaknesses in existing SSL methods. (1)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1195" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "SSL models rely primarily on visual patterns for classification. However, in fine-grained settings, where objects share highly similar appearances, these methods struggle to extract meaningful distinctions, resulting in frequent misclassification. (2). SSL methods depend heavily on pseudolabeling, where high-confidence predictions are used to assign labels to unlabeled samples. When dealing with visually ambiguous cases, incorrect pseudo-labels are generated and subsequently reinforce errors, leading to model bias and further performance degradation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 195, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 296, + 687 + ], + "type": "text", + "content": "In many real-world applications, class names are available as meaningful words or phrases, offering an opportunity to use the semantic information. However, existing SSL methods completely ignore the rich semantic information embedded in these labels. Instead, prior SSLs rely exclusively on one-hot-encoded categorical labels, which discard potentially useful semantic knowledge. We hypothesize that introducing label semantics into SSL can provide additional guidance for visual representation learning, thereby improving the model's ability to distinguish visually similar samples. For example, while \"boy\" and \"man\" may have highly similar visual representations in CIFAR-100 dataset, their semantic embeddings are distinct and different. By aligning visual features with distinct label semantics, SSL models can be improved in two ways. (1). Aligning text embeddings from ground-truth label names with the visual representations of labeled images can help the model better capture subtle visual differences through text cues. This makes the initial model more stable and robust. (2) Even if pseudo-labels are incorrect, aligning the text embedding of the pseudo-label with the visual representation of the unlabeled image can still be beneficial. If the incorrect pseudo-label's semantic meaning is close to the true label, this alignment may help guide the visual representation closer to the correct cluster. However, aligning text embeddings from label name with visual embeddings for SSL introduces two major challenges. The first challenge arises from the inherent similarity between text embeddings of different class names. When similar text embeddings are used to guide the visual representation learning, it can introduce unwanted confusion in the visual space (Section 3.4). The second challenge relates to many-to-one alignment in our settings. Unlike CLIP [31] using NT-Xent loss [35], where each image-text pair is uniquely defined, our setting often encounters multiple images within a batch that share the same text embedding because multiple images may belong to the same class. Directly applying existing visual-text alignment loss like NT-Xent [31], which assume a pairwise positive match, would mistakenly treat these samples as negatives, leading to incorrect optimization and degraded performance (Section 3.5)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In this paper, we introduced SemiVisBooster, an SSL booster that improves SSL performance by incorpo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "rating label semantics into the learning process. To enrich feature learning beyond basic one-hot-encoded ground truth labels or pseudo-labels, we proposed a label semantic enrichment module that converts categorical labels into meaningful label names. To enable robust alignment, we proposed text embedding de-similarity (TEDS) loss to avoid the impact of inherent similarity between text embeddings and developed class-aware visual-text alignment (CAViTA) strategy, which correctly defines positive and negative pairs in cases where multiple images share the same text embedding. Our approach achieves state-of-the-art results on the latest SSL benchmarks [40]. Specifically, SemiVisBooster consistently surpasses the previous method by more than " + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": " across multiple benchmarks, including CIFAR-100 (200 labeled images) and STL-10 (40 labeled images). For fine-grained classification tasks, our method achieved " + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "inline_equation", + "content": "73.6\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": " accuracy on Food 101 (404 labeled images), significantly surpassing the previous state-of-the-art method, which achieved only " + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "inline_equation", + "content": "64.8\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": " accuracy. We summarized our contributions below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 312, + 553, + 443 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 312, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 360 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to directly leverage label name semantics to enhance SSL, achieving state-of-the-art performance on the latest SSL benchmarks and fine-grained classification tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 408 + ], + "type": "text", + "content": "- We introduced TEDS to mitigate the inherent similarity between text embeddings of different class names, enhancing the robustness of our method across various text encoders, including a simple word embedding model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 408, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 553, + 443 + ], + "type": "text", + "content": "- We designed CAViTA to accurately distinguish positive and negative visual-text embedding pairs, ensuring robust visual-text alignment." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 458, + 400, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 458, + 400, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 458, + 400, + 471 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 460, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 460, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 460, + 492 + ], + "type": "text", + "content": "2.1. Semi-supervised Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Semi-supervised learning (SSL) has been an area of extensive research over recent decades. Recent advances in SSL methodologies can be broadly categorized into two primary approaches: consistency regularization [20, 26, 38, 43] and pseudo-labeling [21, 30]. Consistency-based methods apply perturbations either to the input data or the model and enforce constraints to ensure consistency in the features or probability outputs. For example, the II-model [20] introduces dropout to model weights, generating two outputs for consistency regularization. Pseudo-labeling methods, in contrast, use model predictions as pseudo-labels to guide learning on unlabeled data. This approach has gained recent prominence [4, 21, 30, 44]. Combining these approaches, FixMatch [36] introduced a straightforward SSL framework that applies both pseudo-labeling and consistency regularization, but it uses a fixed confidence threshold. Recent work has focused on refining pseudo-labeling quality by introducing dynamic thresholds [45], class-specific thresh-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "1196" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 64, + 531, + 322 + ], + "blocks": [ + { + "bbox": [ + 78, + 64, + 531, + 322 + ], + "lines": [ + { + "bbox": [ + 78, + 64, + 531, + 322 + ], + "spans": [ + { + "bbox": [ + 78, + 64, + 531, + 322 + ], + "type": "image", + "image_path": "ce24737cb60d26b409a7f0bf37af63fb61b4a07c0bb35b9dbc4a32d29a6de109.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 327, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 54, + 327, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 555, + 416 + ], + "type": "text", + "content": "Figure 2. Overview of SemiVisBooer framework for enhancing SSL through label semantic information. (Only the flow of unlabeled data is shown. Labeled data are used for both supervised learning and aligning ground truth labels with their corresponding images.) Base SSL branch: One-hot pseudo-labels are generated from weakly augmented images and used to guide predictions on strongly augmented versions of the same images. Boosting Branch: A label semantic enrichment module maps each categorical label to its label name and uses a pre-trained text encoder to extract text embeddings without any prompt operation. A learnable projector transforms all these embeddings into a text embedding bank. The text embedding de-similarity (TEDS) loss enhances robustness by minimizing the similarity between text embeddings of different classes. The class-aware visual-text alignment is used to align selected text embeddings with corresponding visual embeddings to guide visual representation learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 422, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 296, + 506 + ], + "type": "text", + "content": "olds [46], and dynamic class-specific thresholds [5, 41]. However, even with these adaptive thresholding techniques, existing SSL methods struggle with fine-grained classification as we discussed in Section 1. To boost SSLs, we introduced additional supervision derived from the semantic information in label names. We used FreeMatch[41] as our base SSL." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 230, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 230, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 230, + 529 + ], + "type": "text", + "content": "2.2. Vision and Language Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Numerous studies have demonstrated the benefits of joint image-text representation learning for various vision tasks [9, 13, 16, 23, 25]. Contrastive learning has proven particularly effective for vision-language models by mapping images and text into a shared feature space. Models like CLIP [31], CLOOB [12], ALIGN [17], and DeCLIP [24] achieve this by training text and image encoders on large datasets of image-caption pairs, using contrastive loss to minimize distances between matching pairs while maximizing them for non-matching ones. To handle noise in datasets, ALIGN [16] and DeCLIP [24] use custom metrics, while CLIP [31] employs cosine distance. While these models primarily aim at pre-training large vision-language foundations for zero- and few-shot tasks, our approach enhances semi-supervised learning by aligning visual features" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 445 + ], + "type": "text", + "content": "with pseudo-labels, improving the robustness of SSLs for fine-grained classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 457, + 370, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 370, + 469 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 370, + 469 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 477, + 400, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 477, + 400, + 489 + ], + "spans": [ + { + "bbox": [ + 314, + 477, + 400, + 489 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": "Following [36], we formulate the SSL framework for a " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": "-class classification problem. Let " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(x_i^l, y_i^l) : i \\in [1, B_L]\\}" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " denote a batch of labeled data, and " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{x_j^u : j \\in [1, B_U]\\}" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " represent a batch of unlabeled data, with batch sizes " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "B_L" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "B_U" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": ", respectively. Here, " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "x_i^l" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " denotes a labeled image with label " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "y_i^l" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "x_j^u" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " represents an unlabeled image. The model is trained using a combined objective: " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}} = \\mathcal{L}_s + \\lambda_u \\mathcal{L}_u" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " is the supervised loss over labeled data, " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " is the unsupervised loss over unlabeled data, and " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\lambda_u" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " is the unsupervised loss weight. The supervised loss " + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 495, + 555, + 626 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 635, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 635, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 361, + 635, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\frac {1}{B _ {L}} \\sum_ {i = 1} ^ {B _ {L}} \\mathcal {H} \\left(y _ {i} ^ {l}, \\mathcal {F} _ {m} \\left(y \\mid w \\left(x _ {i} ^ {l}\\right)\\right)\\right) \\tag {1}", + "image_path": "2a9534acbbf3afbe97bcda59c0d6ccac9061cf5ef48a72e4ae459fed28e6348e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is cross-entropy loss, " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "w(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is weak augmentation (i.e., random crop and flip) and " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " is the output probability from the model." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1197" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "For unlabeled data, SSL typically dynamically assigns the pseudo-labels if the model's confidence scores surpass the thresholds. The unsupervised loss " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_u" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 126, + 295, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 126, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 85, + 126, + 295, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {1}{B _ {U}} \\sum_ {j = 1} ^ {B _ {U}} \\mathbb {1} \\left(\\max \\left(q _ {j}\\right) > \\tau\\right) \\mathcal {H} \\left(\\hat {q} _ {j}, Q _ {j}\\right) \\tag {2}", + "image_path": "dfb7b417cbefa18052c6112198c0de1a0e129f2d9942a1d4632aae6455cba2dd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "Q_{j}" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " denote abbreviation of " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m(y|w(x_j^u))" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_m(y|\\mathcal{A}(x_j^u))" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\hat{q}_j = \\arg \\max (q_j)" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " is the one-hot pseudo-label derived from the model's prediction on weakly augmented image. " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " denotes strong image augmentation [7], and " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": " is an indicator function that selects pseudo-labels based on a threshold " + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 167, + 296, + 264 + ], + "type": "text", + "content": ". Our method enhances SSL by enriching pseudo-labels with semantic information to guide visual representation learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 155, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 155, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 155, + 281 + ], + "type": "text", + "content": "3.2. Model Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "text", + "content": "Our method enhances the base SSL by incorporating additional guidance from label semantics (Figure 2). In the base SSL branch, weakly augmented images pass through the model first, generating one-hot pseudo-labels based on confidence thresholds. These pseudo-labels then guide the model to predict the same labels for strongly augmented versions of the corresponding unlabeled images [5, 36, 41]. In the boosting branch, label semantic enrichment module performs a mapping operation from categorical label index to its corresponding label name. A pre-trained text encoder extracts text embeddings from these label names, and a learnable projector transforms them into projected text embeddings, named text embedding bank, that covers all classes. To enhance robustness, we apply text embedding de-similarity (TEDS) loss, " + }, + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{TEDS}" + }, + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "text", + "content": ", which encourages dissimilarity between different text embeddings. Finally, the one-hot pseudo-labels from the base SSL branch are used to select corresponding text embeddings, which are then aligned with the corresponding visual embeddings through class-aware visual-text alignment loss, " + }, + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 288, + 296, + 539 + ], + "type": "text", + "content": ". The overall training objective is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 548, + 294, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 548, + 294, + 561 + ], + "spans": [ + { + "bbox": [ + 69, + 548, + 294, + 561 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {s} + \\lambda_ {u} \\mathcal {L} _ {u} + \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {T E D S} * \\mathcal {L} _ {T E D S}, \\tag {3}", + "image_path": "47f0ab69cf7e501b3674d125585bea948df9c65033d8602f0a066bed890b71a6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\lambda_{TEDS}" + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": " are the weighting factors for the text embedding de-similarity loss and the class-aware visual-text alignment loss, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 209, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 209, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 209, + 624 + ], + "type": "text", + "content": "3.3. Label Semantic Enrichment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "This component converts categorical labels into meaningful words or phrases, allowing the subsequent text encoder module to capture semantic text embeddings. Unlike CLIP [31], which achieves text-image alignment using millions of (image, text) pairs, this is impractical in SSL due to limited annotations. In SSL, we often lack even basic categorical labels, let alone detailed image descriptions." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 68, + 436, + 148 + ], + "blocks": [ + { + "bbox": [ + 333, + 68, + 436, + 148 + ], + "lines": [ + { + "bbox": [ + 333, + 68, + 436, + 148 + ], + "spans": [ + { + "bbox": [ + 333, + 68, + 436, + 148 + ], + "type": "image", + "image_path": "ae1ddcecf2c1f9ca9c2bdcab7ed798661789768a2da9440c27d9fcd51a41f977.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 148, + 395, + 156 + ], + "lines": [ + { + "bbox": [ + 386, + 148, + 395, + 156 + ], + "spans": [ + { + "bbox": [ + 386, + 148, + 395, + 156 + ], + "type": "text", + "content": "T5" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 440, + 69, + 533, + 148 + ], + "blocks": [ + { + "bbox": [ + 440, + 69, + 533, + 148 + ], + "lines": [ + { + "bbox": [ + 440, + 69, + 533, + 148 + ], + "spans": [ + { + "bbox": [ + 440, + 69, + 533, + 148 + ], + "type": "image", + "image_path": "a51bb61f6922a4d879797190692f113bfdc768a6adc3bcfafc063e2787a67e4c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 149, + 492, + 156 + ], + "lines": [ + { + "bbox": [ + 473, + 149, + 492, + 156 + ], + "spans": [ + { + "bbox": [ + 473, + 149, + 492, + 156 + ], + "type": "text", + "content": "BERT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 333, + 159, + 436, + 237 + ], + "blocks": [ + { + "bbox": [ + 333, + 159, + 436, + 237 + ], + "lines": [ + { + "bbox": [ + 333, + 159, + 436, + 237 + ], + "spans": [ + { + "bbox": [ + 333, + 159, + 436, + 237 + ], + "type": "image", + "image_path": "efc17289a0f474f48a1e37dc9b2226386b15db3b2e7a5e49d869005fb2d34a09.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 239, + 395, + 247 + ], + "lines": [ + { + "bbox": [ + 386, + 239, + 395, + 247 + ], + "spans": [ + { + "bbox": [ + 386, + 239, + 395, + 247 + ], + "type": "text", + "content": "T5" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 251, + 554, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 297 + ], + "type": "text", + "content": "Figure 3. t-SNE visualization of text embeddings for Food101 label names. The first row shows raw text embeddings generated by T5 and BERT. The second row is the embeddings after applying TEDS, resulting in more dispersed and dissimilar embeddings." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 159, + 533, + 237 + ], + "blocks": [ + { + "bbox": [ + 440, + 159, + 533, + 237 + ], + "lines": [ + { + "bbox": [ + 440, + 159, + 533, + 237 + ], + "spans": [ + { + "bbox": [ + 440, + 159, + 533, + 237 + ], + "type": "image", + "image_path": "2dcc0713ac24088bb3b08002f9244169e01864666237dd9eaf904cc94e44b9f1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 239, + 494, + 247 + ], + "lines": [ + { + "bbox": [ + 474, + 239, + 494, + 247 + ], + "spans": [ + { + "bbox": [ + 474, + 239, + 494, + 247 + ], + "type": "text", + "content": "BERT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 304, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 401 + ], + "type": "text", + "content": "However, since we have prior knowledge of all label names, our label semantic enrichment module performs a straightforward and fixed mapping from categorical pseudo-labels to their respective semantic names. This fixed mapping also contributes to the efficiency (discussed in Section 3.6). We directly feed the label name to the text encoder without any prompt operations. Using only label names ensures simplicity and reproducibility." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 407, + 480, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 480, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 480, + 420 + ], + "type": "text", + "content": "3.4. Text Embedding De-Similarity" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 425, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 579 + ], + "type": "text", + "content": "Text embeddings generated by a text encoder are not inherently suitable for guiding visual representation learning because they lack sufficient separation between different classes (see the first row of Figure 3). To quantify this, we calculated the cosine similarity between text embeddings of different class names using the BERT model. The results showed high similarity scores (above 0.92) between classes such as \"Carrot Cake\" and \"Chocolate Cake\", as well as \"Cheesecake\" and \"Chocolate Cake\". This inherent similarity in text embeddings unintentionally pushes corresponding visual representations closer together during visual-text alignment, leading to increased confusion between semantic similar classes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 640 + ], + "type": "text", + "content": "To mitigate this issue, we introduce text embedding desimilarity (TEDS), a loss function that encourages greater separation between text embeddings by minimizing the absolute values of pairwise dot products between projected text embeddings. The TEDS loss is formulated as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T E D S}} = \\frac {1}{C (C - 1)} \\sum_ {i \\neq j} \\max \\left(\\left| \\frac {\\mathbf {T} _ {i} ^ {\\top} \\mathbf {T} _ {j}}{\\| \\mathbf {T} _ {i} \\| \\| \\mathbf {T} _ {j} \\|} \\right| - \\Delta , 0\\right), \\tag {4}", + "image_path": "ecade17794b488e28539f8dda41aa5beef0022db921321ace3dba79e3e3a8163.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_j" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " are individual projected text embeddings, " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{T}_i\\|" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{T}_j\\|" + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": " represent their" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1198" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 156, + 202 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 156, + 202 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 156, + 202 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 156, + 202 + ], + "type": "image", + "image_path": "bf06f05d0cb88275e6b343ede7905b487e71f847713a5c1632fc4eba113a32f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 205, + 135, + 213 + ], + "lines": [ + { + "bbox": [ + 91, + 205, + 135, + 213 + ], + "spans": [ + { + "bbox": [ + 91, + 205, + 135, + 213 + ], + "type": "text", + "content": "(a). NT-Xent" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 236 + ], + "type": "text", + "content": "Figure 4. Comparison between NT-Xent loss and class-aware contrastive loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 73, + 282, + 202 + ], + "blocks": [ + { + "bbox": [ + 176, + 73, + 282, + 202 + ], + "lines": [ + { + "bbox": [ + 176, + 73, + 282, + 202 + ], + "spans": [ + { + "bbox": [ + 176, + 73, + 282, + 202 + ], + "type": "image", + "image_path": "43175bce05b96c45c9698f444323dcd3c70b6128397d1c7f57a43c24718375b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 204, + 274, + 213 + ], + "lines": [ + { + "bbox": [ + 181, + 204, + 274, + 213 + ], + "spans": [ + { + "bbox": [ + 181, + 204, + 274, + 213 + ], + "type": "text", + "content": "(b). Class-aware Contrastive" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "content": "magnitudes and " + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "content": " is a tolerance margin (e.g., " + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 239, + 296, + 371 + ], + "type": "text", + "content": "), defining the lowest acceptable similarity level. Integrating TEDS into the training objective encourages the learnable projector to extract more distinct and discriminative features from text embeddings. Since the text encoder remains frozen, TEDS primarily trains the learnable projector to enhance the separation of projected text embeddings. After applying TEDS, t-SNE visualization (see the second row of Figure 3) confirms that the projected text embeddings become more widely separated, which in turn helps the model learn more distinguishable visual representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 377, + 241, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 241, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 241, + 390 + ], + "type": "text", + "content": "3.5. Class-aware Visual-text Alignment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": "In [31], the NT-Xent-based visual-text alignment forms positive pairs based on specific fine-grained text-visual matches, where each image is paired with a unique text description, and all other text descriptions in the batch are treated as negatives. However, this approach is suboptimal in our SSL context, where a batch may contain multiple images from the same class (i.e., sharing the same text embedding). For example, if one batch has two visual " + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "inline_equation", + "content": "(V_{0}, V_{1})" + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": " and their corresponding labels " + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "inline_equation", + "content": "(T_{0}, T_{1})" + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "inline_equation", + "content": "T_{0} = T_{1}" + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": ", the existing text-visual alignment loss would incorrectly treat " + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "inline_equation", + "content": "(V_{0}, T_{1})" + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "inline_equation", + "content": "(V_{1}, T_{0})" + }, + { + "bbox": [ + 55, + 395, + 296, + 659 + ], + "type": "text", + "content": " as negative pairs—False Negative Pairs (Figure 4). To address this, we introduce class-aware visual-text alignment which leverages class labels to define positive and negative pairs. Specifically, positive pairs are formed between text-visual pairs from the same class, while negative pairs come from different classes. This is implemented by computing a similarity matrix between visual and text embeddings and applying a mask to identify positive pairs based on class labels. For each visual-text pair within the same class, we maximize similarity, while minimizing similarity for pairs from different classes. The class-aware visual-text alignment loss is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 674, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 674, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 67, + 674, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\left(\\frac {\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right) \\cdot \\mathbf {M} _ {i j}}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\tilde {\\mathbf {V}} _ {i} \\cdot \\tilde {\\mathbf {T}} _ {j}}{\\kappa}\\right)}\\right) \\tag {5}", + "image_path": "4f988c457f2ca94f6dd4b4dc42e81e99237867379b0ae3d358a936480e4445da.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " represents the number of embeddings, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_i" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is the normalized visual embedding of sample " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}_j" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is the normalized text embedding of sample " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is the temperature scaling factor and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is a class-aware mask where " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij} = 1" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " if samples " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " belong to the same class, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ij} = 0" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 155, + 443, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 443, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 443, + 167 + ], + "type": "text", + "content": "3.6. Implementation Detail" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 173, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 399 + ], + "type": "text", + "content": "Text Encoder. The text encoder plays a crucial role in extracting semantic information from label names to guide visual representation learning. The key requirement is the ability to capture meaningful semantics. Any text encoder that effectively extracts distinct and informative text embeddings can serve this purpose. To ensure flexibility and broad applicability, our method is designed to work with both language models and simple word embedding models that can generate semantically meaningful representations. We evaluate our approach using three language models—T5 [32], BERT [8], and CLIP's text encoder [31]—as well as a lightweight word embedding model, fastText [2]. Among these, the pre-trained CLIP text encoder is particularly well-suited for this task, as it has been specifically trained for visual-text alignment, making it highly effective in guiding visual feature learning. Ablation studies confirm that our method generalizes well across different text encoders, demonstrating its robustness and adaptability across various models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 653 + ], + "type": "text", + "content": "Projection Strategies. In the SSL context, both the visual network architecture and the text encoder can be chosen freely, leading to a dimensional mismatch between text embeddings and visual embeddings. This mismatch poses a challenge for visual-text alignment, as the two modalities exist in different feature spaces. To address this issue, we explore two projection strategies that project text or visual embeddings to the same feature dimension. The first strategy, ProjT2V, is the approach used in Figure 2. In this method, a learnable projection layer is applied to the text embeddings, transforming them into the same dimensional space as the visual embeddings [16]. This transformation enables direct alignment between the projected text embeddings and the visual embeddings using contrastive loss. The second strategy, Proj-V2T, takes the opposite approach by projecting visual embeddings into the dimensional space of the text embeddings [11]. While both strategies successfully create a shared embedding space, our experiment shows that ProjT2V outperforms Proj-V2T, providing a stronger improvement in SSL performance and more effective visual-text alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Projector Architecture. We implemented the projector architecture by considering two main factors. First, the learnable projector should present a challenging alignment task for the in-training visual feature extraction model. To achieve this, the projector cannot be overly complex, as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1199" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "that would exert too much influence on the visual or text embeddings, potentially distorting their features. Second, we aimed to keep the architecture lightweight and efficient, motivating us to choose a simpler design. Based on these requirements and existing works usually use linear projections [11, 16], we explored two one-layer architectures: a single-layer MLP and self-attention." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 298, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 298, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 298, + 373 + ], + "type": "text", + "content": "Efficient Implementation. Our method can achieve high efficiency with minimal vRAM usage and without adding significant computation when boosting SSLs for three main reasons. First, by freezing the text encoder, we only perform text encoder inference without the computationally expensive gradient calculations. Second, due to the fixed mapping from categorical labels to label names, the input to the text encoder model is limited to the number of label names. This allows text embeddings to be precomputed once during initialization and stored in a cache, eliminating the need for repeated computations during training. As a result, the text encoder can be offloaded from vRAM, and the necessary text embeddings can be efficiently retrieved from the pre-computed cache. Finally, we reuse the visual features already extracted by the base SSL for alignment, further optimizing efficiency. These optimizations ensure that our approach enhances SSL performance without introducing significant computational cost or memory overhead." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 386, + 132, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 132, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 132, + 399 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 105, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 105, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 105, + 419 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 556 + ], + "type": "text", + "content": "Datasets. We evaluated our method on the latest SSL benchmarks [40]. To further evaluate the robustness on fine-grained classification task, we also tested our method on Food 101 dataset [3] as prior SSLs fail on this dataset. For each dataset, we examined two different quantities of labeled images. Detailed settings for each dataset are provided in the Appendix. For instance, we used 404 and 1010 labeled images for the Food-101 dataset. Labeled samples were randomly selected from the training set, except for STL-10 and Semi-Aves. For these two datasets, the labeled and unlabeled splits are predefined." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Experimental Setup. Our method was based on FreeMatch [41] and implemented with a widely-used SSL codebase, USB [40]. For a fair comparison with baseline methods, we maintained consistent training settings across methods, adjusting only SSL-specific hyperparameters. This approach ensured that all SSLs use the same visual network architectures and pre-training strategies. For the base SSL settings, we initialized the self-adaptive threshold for FreeMatch at 0, whereas FreeMatch typically uses " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{C}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ". In our boosting configuration, we set the weight for the alignment loss term, " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ", to 1 with a temperature factor of 0.07. We also assigned the weight of the text embedding de-similarity term, " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{TEDS}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " to 0.01. Our model train" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": "ing used the AdamW optimizer [18] with a cosine annealing learning rate schedule and an initial warmup phase. The training spanned 204,800 steps in total, with the first 10,240 steps dedicated to warmup. Batch sizes were adjusted based on the dataset, with labeled and unlabeled samples sharing the same batch size within each dataset. Specifically, batch sizes were set to 32 for Food-101, CIFAR-100, and STL-10; 64 for Semi-Aves; and 8 for EuroSAT. The base learning rates are also dataset-specific: " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " for Food-101 and STL-10, " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " for CIFAR-100 and Semi-Aves, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 313, + 72, + 555, + 204 + ], + "type": "text", + "content": " for EuroSAT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 215, + 405, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 405, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 405, + 227 + ], + "type": "text", + "content": "4.2. Overall Result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": "To the best of our knowledge, we compared our method with SSL algorithms in the past 4 years, as shown in Table 1. Our approach consistently outperformed prior SSL methods, particularly in fine-grained classification tasks. For example, SemiVisBooster achieved " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "73.6\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " accuracy on Food-101 using only 404 labeled images and " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "81.8\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " accuracy on CIFAR-100 with only 200 labeled images, outperforming the second-best method by margins of " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": ", respectively. The improvement on Food-101 is more substantial than on CIFAR-100 because CIFAR-100 contains only some fine-grained classes, and existing SSL methods may effectively distinguish most classes with clear boundaries in CIFAR-100. On the Semi-Aves dataset, our method achieved a " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " accuracy improvement over the next best method. In coarse-grained classification tasks, our method also demonstrated strong results, achieving " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "95.4\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " accuracy on the STL-10 dataset with just 20 labeled images and " + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "inline_equation", + "content": "96.8\\%" + }, + { + "bbox": [ + 313, + 233, + 555, + 497 + ], + "type": "text", + "content": " accuracy on the EuroSAT dataset with the same number of labeled images. These results show that our method is not only effective for fine-grained classification but also enhances performance in general coarse-grained classification problems." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Comparison with CLIP. To assess the necessity of training a visual model using SSL, we compared our method against CLIP's zero-shot performance [31] and few-shot fine-tuning results. Despite using a smaller visual backbone (ViT Small [40]), our method outperformed both the CLIP-ViT Base and CLIP-ViT Large model on the CIFAR-100 and EuroSAT datasets, where CLIP lacks domain-specific adaptation (Table 1). CLIP performed impressively on Food-101 and STL-10, likely because its training on 400 million image-text pairs exposes it to common food items and general object categories present in these datasets [31]. We further fine-tune CLIP vision encoder with a classifier head on EuroSAT dataset with 20 labeled images. The fine-tuned CLIP-ViT Base and CLIP-ViT Large model achieved " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "84.63\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "79.64\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " accuracy, respectively. In contrast, our compact model—using only " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " of CLIP-Base's vision parameters and training from scratch—achieves " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "96.8\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " under USB [40] settings. This comparison highlights the chal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1200" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 47, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 56, + 47, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 47, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 47, + 553, + 281 + ], + "type": "table", + "html": "
DatasetFood-101 [3]Semi-Aves [37]CIFAR-100 [19]STL-10 [6]EuroSAT [15]
# of labels40410103959200400401002040
SSLUDA [43]53.274.368.371.881.885.692.592.395.1
FixMatch [36]54.764.568.271.381.085.792.690.194.1
CoMatch [22]30.648.961.565.675.285.890.894.795.2
Dash [45]57.776.767.870.480.785.792.989.793.9
CRMatch [10]44.469.868.076.182.289.888.588.593.4
AdaMatch [33]60.176.668.479.484.087.993.393.895.4
FlexMatch [46]48.471.867.774.482.188.792.695.495.2
SimMatch [47]55.473.166.277.383.788.493.392.995.6
DeSSL [34]56.675.068.270.380.684.392.491.896.3
FreeMatch [41]60.678.867.578.984.688.592.094.394.7
SoftMatch [5]60.476.968.278.783.887.692.994.995.5
EPASS [27]64.877.665.977.882.691.593.882.390.1
SequenceMatch [28]43.767.968.170.177.486.489.692.992.6
SemiVisBooster73.683.870.381.885.495.495.996.896.7
Zero-shotCLIP-ViT-B[31]84.410.1*65.197.249.4
CLIP-ViT-L[31]92.99.9*77.999.359.9
", + "image_path": "c94dade931e8d73aeb31826ed6f32f1b5926569e7118bed7fec8b3066e97767c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Table 1. Top 1 Accuracy: Bold indicates the highest accuracy and underline indicates the second highest accuracy. Star* marks the result from hugging face model inference [29]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 308, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 344 + ], + "type": "text", + "content": "lenges general models face in specialized domains and underscores the importance of continuously improving SSL methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 124, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 124, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 124, + 361 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 449 + ], + "type": "text", + "content": "We conducted the ablation studies using the following setup, unless specified: the projection strategy was ProjT2V, the projector architecture was an MLP, the text encoder was the CLIP text encoder, we used both text embedding de-similarity loss and class-aware visual-text alignment strategy for training, and the input to text encoder is the label name." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "content": "Projection Strategies and Projector Architecture. We study the impact of two commonly used projection methods: ProjT2V, and Proj-V2T. Following the approach in [16], we implement the ProjT2V module using a single-layer MLP with the text embedding dimension as the input and the visual embedding dimension as the output. Similarly, for the Proj-V2T module, we use a single-layer MLP, but with the input and output dimensions reversed. For self-attention, we use a one-layer self-attention module with a hidden dimension of 512, followed by a linear MLP with an input dimension of 512 and an output dimension matching the visual embedding dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "From Table 2, we noticed that ProjT2V works best on both setups, while Proj-V2T lags behind. This is expected as ProjT2V directly enhances visual representation learning, whereas Proj-V2T lacks this capability. With ProjV2T, the alignment loss only ensures that the projected visual embeddings resemble the text embeddings, but this approach has two limitations. First, ProjV2T does not address inherent similarities across text embeddings from different classes, which can cause class overlap. Second, even if text embeddings are distinct, ProjV2T does not ensure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 416 + ], + "type": "text", + "content": "that the pre-projection visual embeddings (i.e. the raw output from the vision model) will naturally align with the text embeddings. This is because the projector itself influences the alignment target, rather than allowing the vision model to learn distinct representations independently. In contrast, ProjT2V promotes direct visual representation learning by projecting the text embeddings and aligning them with the raw visual embeddings directly. Additionally, we ensure that text embeddings are distinct using the TEDS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 491 + ], + "type": "text", + "content": "It is worth mentioning that we also ablate the projector used for T2V and MLP works slightly better, probably due to the simplicity of the MLP layer. It creates a more challenging task for the visual-text alignment, encouraging more robust visual feature extraction. These results support our decision to use Proj-T2V and MLP as projector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 492, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 553 + ], + "type": "text", + "content": "Benefit of TEDS and Impact of Text Encoders. We evaluate the impact of TEDS by testing our method both with and without TEDS across different text encoders. To implement our method without TEDS, we set " + }, + { + "bbox": [ + 313, + 492, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = 0" + }, + { + "bbox": [ + 313, + 492, + 555, + 553 + ], + "type": "text", + "content": ". The pre-trained text encoder models are from HuggingFace repositories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "content": "Our method consistently outperforms the baseline across different text encoders, including the simple word embedding model, FastText (Table 3). The key to boost SSL is incorporating semantic information to guide visual representations. Thus, our method is not restricted to a specific text model but is adaptable to various text encoders, ranging from lightweight word embedding models to language models, as long as they can extract meaningful semantics from label names. This flexibility enhances the practical applicability of our method in real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "By comparing different text encoders, we found that the CLIP text encoder achieves the highest performance on Food-101 (404 labeled images), followed by fastText and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1201" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 57, + 170, + 114 + ], + "blocks": [ + { + "bbox": [ + 59, + 57, + 170, + 114 + ], + "lines": [ + { + "bbox": [ + 59, + 57, + 170, + 114 + ], + "spans": [ + { + "bbox": [ + 59, + 57, + 170, + 114 + ], + "type": "table", + "html": "
Projection StrategiesProjectorFood-101
4041010
Proj-V2TMLP60.5080.38
Proj-T2VSelf-attention72.5583.83
Proj-T2VMLP73.5583.41
", + "image_path": "e17432359c4f0bb8cf20f7d5081a8e9b8d5cbd80cddebce82c6fa165f451ac9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 185, + 56, + 299, + 159 + ], + "blocks": [ + { + "bbox": [ + 55, + 122, + 173, + 165 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 173, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 173, + 165 + ], + "type": "text", + "content": "Table 2. Choice of Projector Strategies: Proj-T2V with MLP project achieves the best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 185, + 56, + 299, + 159 + ], + "lines": [ + { + "bbox": [ + 185, + 56, + 299, + 159 + ], + "spans": [ + { + "bbox": [ + 185, + 56, + 299, + 159 + ], + "type": "table", + "html": "
Text EncoderTEDSFood-101
404 1010
CLIP-Text Encoder✓73.55 83.41
BERT✓69.73 83.19
T5✓68.86 83.18
FastText✓72.80 83.13
CLIP-Text Encoder×70.40 82.63
BERT×67.67 82.44
T5×63.92 82.79
FastText×69.57 81.25
", + "image_path": "bc28b6bde5a715c5c45b0a7c220d593607b7006cbef2787be19d3298cd8482a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 179, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 296, + 430 + ], + "type": "text", + "content": "BERT model, while the T5 model performs the worst, particularly in the absence of TEDS. This performance discrepancy can be attributed to the different training objectives of these language models. The CLIP model is explicitly trained for text-image alignment using contrastive learning, creating a shared embedding space that directly supports text-vision alignment. This shared space makes the CLIP text encoder highly effective for guiding visual model training. BERT, while not designed specifically for text-vision alignment, captures rich textual representations that still contribute to the visual model learning process. In contrast, T5 is primarily designed for text-to-text tasks, and its training objectives are less suited for aligning visual and textual embeddings, resulting in its comparatively lower performance. Interestingly, FastText, a simple word embedding model, achieved strong performance by effectively extracting distinct text features from label names. Compared to general language models like BERT and T5, FastText focuses on subword structure rather than contextual meaning [2], allowing it to better preserve distinctions between label names and reduce confusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "text", + "content": "The results also indicated that TEDS significantly enhances performance across different text encoders (Table 3). This improvement is expected, as raw text embeddings for different classes often exhibit high similarity (Figure 3), which can mislead the visual model during training, especially the label data is scare. Our TEDS module mitigates this issue by penalizing the similarity between text embeddings from different class names, reducing the impact of inherent similarities. Conversely, our method without TEDS lacks this mechanism, leading to poorer performance due to similarity-induced confusion. This underscores the importance of TEDS in making our method adaptable to various text encoder models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Benefit of CAViTA. We evaluate the effectiveness of CAViTA loss by comparing it to the widely used NT-Xent contrastive loss [35], which is commonly used for visual-text alignment in CLIP. Table 4 shows that CAViTA significantly outperforms NT-Xent loss. This result is expected because multiple images in a batch can share the same text embedding. NT-Xent mistakenly classified true positive pairs as negatives, penalizing them and leading to a performance drop. Table 6 reports the False negative rate under NT-Xent loss. It mistakenly penalizes " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "18.71\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " of posi" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 56, + 417, + 106 + ], + "blocks": [ + { + "bbox": [ + 194, + 161, + 288, + 171 + ], + "lines": [ + { + "bbox": [ + 194, + 161, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 161, + 288, + 171 + ], + "type": "text", + "content": "Table 3. Benefit of TEDS" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 56, + 417, + 106 + ], + "lines": [ + { + "bbox": [ + 312, + 56, + 417, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 56, + 417, + 106 + ], + "type": "table", + "html": "
Text-visual Alignment LossFood-101
4041010
NT-Xent52.6175.22
CAViTA73.5583.41
", + "image_path": "518ae90c316ceaf7aa0f5cf0ed36eb85c3d0b05cc951f981a06711e113827162.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 435, + 56, + 550, + 117 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 423, + 159 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 423, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 423, + 159 + ], + "type": "text", + "content": "Table 4. Benefit of CAViTA: CAViTA outperforms NT-Xent loss because it correctly define positive pairs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 435, + 56, + 550, + 117 + ], + "lines": [ + { + "bbox": [ + 435, + 56, + 550, + 117 + ], + "spans": [ + { + "bbox": [ + 435, + 56, + 550, + 117 + ], + "type": "table", + "html": "
λcλTEDSFood-101
10.0173.6
10.173.8
1172.8
0.10.0161.7
0.010.0154.1
", + "image_path": "4e593fac45364dbe31dfb585d47d609be990362cd9f30a32ed62ad4b0a0a0c3f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 331, + 178, + 536, + 215 + ], + "blocks": [ + { + "bbox": [ + 434, + 125, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 434, + 125, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 434, + 125, + 550, + 148 + ], + "type": "text", + "content": "Table 5. Hyperparameter study: our method is sensitivity to " + }, + { + "bbox": [ + 434, + 125, + 550, + 148 + ], + "type": "inline_equation", + "content": "{\\lambda }_{c}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 178, + 536, + 215 + ], + "lines": [ + { + "bbox": [ + 331, + 178, + 536, + 215 + ], + "spans": [ + { + "bbox": [ + 331, + 178, + 536, + 215 + ], + "type": "table", + "html": "
Dataset (labeled #)FN count(mean±std)FN ratio(mean±std)
Food 101 (404)7.65 ± 3.3418.71% ± 6.96%
EuroSAT (20)3.71 ± 2.0329.72% ± 11.79%
", + "image_path": "91e50ade626765def7f86bad6aadefbe940aa82bb5eb31b745b68819f3b81f62.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 337, + 219, + 531, + 230 + ], + "lines": [ + { + "bbox": [ + 337, + 219, + 531, + 230 + ], + "spans": [ + { + "bbox": [ + 337, + 219, + 531, + 230 + ], + "type": "text", + "content": "Table 6. False negative (FN) rate under NT-Xent loss" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "content": "tive pairs, leading to a " + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "content": " accuracy drop on the Food101 dataset. Furthermore, applying NT-Xent loss within our method does not improve the base SSL method (i.e., FreeMatch) and instead decreases performance by " + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 313, + 248, + 553, + 332 + ], + "type": "text", + "content": ". This finding highlights the importance of correctly defining positive pairs, as misclassifications can significantly degrade model performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "text", + "content": "Benefit of Using Label Name. We compared our method with longer text embeddings by extending label names using GPT-4o with the prompt: \"Describe the visual features of {label name} in 2-3 sentences\". These extended descriptions were then used as input to the text encoder to guide visual representation learning. However, using longer text embeddings generated by GPT-4o led to a drop in performance, with accuracy decreasing to " + }, + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "inline_equation", + "content": "70.0\\%" + }, + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "text", + "content": " on Food101 (404 labeled images) and " + }, + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 313, + 337, + 554, + 529 + ], + "type": "text", + "content": " on CIFAR-100 (200 labeled images), compared to using only the label name. This decline in performance is because long text embedding may include unnecessary tokens that dilute the core semantics and spread out embeddings unnecessarily, resulting in unnecessary confusion between different classes. In contrast, a short label name provides compact and discriminative semantics, which is more suitable for our classification task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": "Hyper-parameter Study. Since " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " has been studied in prior SSLs [36, 41], we focus on analyzing " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{TEDS}" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " by varying them from 0.01 to 1 in Table 5. Our results indicate that the performance of our method is highly sensitive to " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": ". Increasing " + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": " improves performance, as a stronger visual-text alignment loss helps enforce sufficient regularization on visual representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 388, + 653 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We introduced SemiVisBooster to improve prior SSL by using semantics from label names to guide visual feature learning and achieved best performance on latest SSL benchmarks[40] and find-grained classification task." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1202" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 163, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "text", + "content": "[1] Ardhendu Behera, Zachary Wharton, Pradeep Hewage, and Asish Bera. Context-aware attentional pooling (cap) for fine-grained visual classification. In The Thirty-Fifth AAAI Conference on Artificial Intelligence. AAAI, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "text", + "content": "[2] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5, 2016. 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 252, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 296, + 296 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446-461. Springer, 2014. 1, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "text", + "content": "[4] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 194-203, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 416 + ], + "type": "text", + "content": "[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In The Eleventh International Conference on Learning Representations, 2023. 1, 3, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 296, + 471 + ], + "type": "text", + "content": "[6] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, pages 215-223, Fort Lauderdale, FL, USA, 2011. PMLR. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 525 + ], + "type": "text", + "content": "[7] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702-703, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 527, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 527, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 527, + 296, + 613 + ], + "type": "text", + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "text", + "content": "[9] Zi-Yi Dou, Aishwarya Kamath, Zhe Gan, Pengchuan Zhang, Jianfeng Wang, Linjie Li, Zicheng Liu, Ce Liu, Yann LeCun, Nanyun Peng, et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. Advances in neural information processing systems, 35:32942-32956, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[10] Yue Fan, Anna Kukleva, Dengxin Dai, and Bernt Schiele. Revisiting consistency regularization for semi-supervised learning. International Journal of Computer Vision, 131(3): 626-643, 2023. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "type": "text", + "content": "This work was supported by the National Institute of Biomedical Imaging and Bioengineering under NIH award numbers R01LM011834 and R56EB032819." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc'Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. Advances in neural information processing systems, 26, 2013. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 182 + ], + "type": "text", + "content": "[12] Andreas Fürst, Elisabeth Rumetshofer, Johannes Lehner, Viet T Tran, Fei Tang, Hubert Ramsauer, David Kreil, Michael Kopp, Günter Klambauer, Angela Bitto, et al. Cloob: Modern hopfield networks with infoloob outperform clip. Advances in neural information processing systems, 35: 20450-20468, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "type": "text", + "content": "[13] Tanmay Gupta, Kevin Shih, Saurabh Singh, and Derek Hoiem. Aligned image-word representations improve inductive transfer across vision-language tasks. In 2017 IEEE International Conference on Computer Vision (ICCV), page 4223-4232. IEEE, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "text", + "content": "[15] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 414 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 4904-4916. PMLR, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 554, + 480 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 482, + 554, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 513 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 536 + ], + "type": "text", + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "text", + "content": "[20] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[21] Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. ICML 2013 Workshop : Challenges in Representation Learning (WREPL), 2013. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "text", + "content": "[22] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regularization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9475-9484, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 713 + ], + "type": "text", + "content": "[23] Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, Yejin Choi, and Jianfeng Gao. Oscar: Object-Semantics Aligned Pre-training for Vision-Language Tasks, page 121-137. Springer International Publishing, 2020. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1203" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[25] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[26] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979-1993, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 293 + ], + "type": "text", + "content": "[27] Khanh-Binh Nguyen. Debiasing, calibrating, and improving Semi-supervised Learning performance via simple Ensemble Projector. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2430-2439, Los Alamitos, CA, USA, 2024. IEEE Computer Society. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "text", + "content": "[28] Khanh-Binh Nguyen. Sequencematch revisiting the design of weak-strong augmentations for semi-supervised learning. In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 96-105, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[29] OpenAI. CLIP: Contrastive Language-Image Pretraining. https://huggingface.co/openai/clip-vit-base-patch32, 2021. Accessed: 2024-11-10. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[30] Hieu Pham, Zihang Dai, Qizhe Xie, and Quoc V. Le. Meta pseudo labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11557-11568, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "text", + "content": "[32] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[33] Becca Roelofs, David Berthelot, Kihyuk Sohn, Nicholas Carlini, and Alex Kurakin. Adamatch: a unified approach to semi-supervised learning and domain adaptation. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[34] Hugo Schmutz, Olivier Humbert, and Pierre-Alexandre Mattei. Don't fear the unlabelled: safe semi-supervised learning via simple debiasing. arXiv preprint arXiv:2203.07512, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[35] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Proceedings of the 30th International Conference on Neural Information Processing Systems, page 1857-1865, Red Hook, NY, USA, 2016. Curran Associates Inc. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[36] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 709 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 116 + ], + "type": "text", + "content": "Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[37] Jong-Chyi Su and Subhransu Maji. The semi-supervised naturalist-aves challenge at fgvc7 workshop. arXiv preprint arXiv:2103.06937, 2021. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[38] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 1195–1204, Red Hook, NY, USA, 2017. Curran Associates Inc. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[40] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. Advances in Neural Information Processing Systems, 35:3938-3961, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[41] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. arXiv preprint arXiv:2205.07246, 2022. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 375, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 452 + ], + "type": "text", + "content": "[42] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, Heli Qi, Zhen Wu, Yu-Feng Li, Satoshi Nakamura, Wei Ye, Marios Savvides, Bhiksha Raj, Takahiro Shinozaki, Bernt Schiele, Jindong Wang, Xing Xie, and Yue Zhang. Semi-supervised-learning. https://github.com/microsoft/Semi-supervised-learning, 2023.1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[43] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[44] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 553, + 597 + ], + "type": "text", + "content": "[45] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In International conference on machine learning, pages 11525-11536. PMLR, 2021. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 653 + ], + "type": "text", + "content": "[46] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in neural information processing systems, 34:18408-18419, 2021. 1, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 654, + 553, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 709 + ], + "type": "text", + "content": "[47] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14471-14481, 2022. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1204" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_content_list.json b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9011c753f0e5e81aad8c8ceccd403ac4206092ed --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_content_list.json @@ -0,0 +1,1461 @@ +[ + { + "type": "text", + "text": "Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring", + "text_level": 1, + "bbox": [ + 151, + 130, + 846, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufei Zhu $^{1,2}$ , Hao Chen $^{*1,2}$ , Yongjian Deng $^{3}$ , and Wei You $^{4}$", + "bbox": [ + 264, + 214, + 732, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Engineering, Southeast University, Nanjing, China \n $^{2}$ Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China.", + "bbox": [ + 122, + 244, + 872, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ College of Computer Science, Beijing University of Technology, Beijing, China", + "bbox": [ + 179, + 299, + 818, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Advanced Computing and Storage Lab, Huawei Technologies, Beijing, China", + "bbox": [ + 155, + 321, + 776, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{220232213，haochen303}@seu.edu.cn，yjdeng@bjut.edu.cn，youwei22@huawei.com", + "bbox": [ + 166, + 339, + 828, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 398, + 326, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional motion deblurring methods struggle to effectively model motion information within the exposure time. Recently, event cameras have attracted significant research interest for its ability to model motion cues over the exposure duration. However, these methods directly fuse event features with image, overlooking the intrinsic heterogeneity of events. In this paper, we identify that the event modality contains two conflicting types of information: edge features and motion cues. Events accumulated over a short exposure period capture sharp edge details but lose motion information, while those accumulated over a long exposure period blur edge details due to motion. To address this issue, we propose a simple yet effective approach to disentangle these two cues from event features and employ an edge-aware sharpening module along with motion-driven scale-adaptive deblurring module to fully leverage both. Specifically, the first module aids in restoring sharp edges by leveraging the clear edge features provided by events, while the second module leverages motion cues to learn diverse blur kernels, adaptively adjusting the receptive field for optimal deblurring. Extensive experiments on synthetic and real-world datasets validate the effectiveness of our approach and yield a substantial improvement over state-of-the-art single-frame methods and surpasses most multi-frame-based methods. Code will be publicly available.", + "bbox": [ + 86, + 430, + 483, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 825, + 223, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motion blur is often caused by camera shake or the movement of objects. The task of motion deblurring [1, 8, 19-21]", + "bbox": [ + 89, + 845, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/537ef25ec2eeb135a1513e1918b7d7a517e1184f6acb2db4e786ec498a4dd313.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 575, + 398, + 669, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8d6ce0b261fb260abaeed3c13e82b96ed1d1e634fc74184112f47dc1d19701bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 398, + 777, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01edaff7caadb0444c9efbe55b4360bbd9c88c5a9c849a6042c84ea33bf1a23d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 398, + 882, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51b98486babf9b2e163501a268bcd4e4761a412c5ce47b26c0a32c20b391b277.jpg", + "image_caption": [ + "(b)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 540, + 507, + 647, + 625 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0176bf9fbee85f14b9d91bb0407d5dedf163889b5e642216958739b6c5ca652.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 507, + 772, + 589 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9f549a2886fcb970dd8f9d58d1efebf670624d83f0d4f1e94184f3a3fcee2d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 516, + 892, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94d73839f2a813d7498e818a336bd249e0f01371f9acfbe41df402ae1b8b4e26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 617, + 647, + 726 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da57136c57da64e49f60922a322a191a5afb3791bab5ba6904feaf328901fc0f.jpg", + "image_caption": [ + "Figure 1. (a) Visualization of SCER(Symmetric Cumulative Event Representation) [35]. (b) Previous fusion methods between events and image. (c) Ours. Separate events into edge features and motion cues." + ], + "image_footnote": [], + "bbox": [ + 656, + 619, + 772, + 729 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c0fa5a681be0fb37001a6bd936300fc6f08cfd83b8ebac0084ecce4bd7970aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 648, + 887, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aims to recover sharp images from blurred ones by restoring clear edges and capturing fine details, ultimately enhancing the image quality. However, since image features lack the ability to model motion cues over the exposure time, they cannot accurately estimate the variant blur ker", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14732", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nels present in real-world scenarios. This limitation often results in subpar performance of current methods on real-world datasets. Recently, event cameras [2, 9], bio-inspired sensors that asynchronously capture changes in pixel intensities, have been leveraged to aid motion deblurring thanks to their microsecond-level latency and low power consumption.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies [6, 11, 35, 47, 48] aim to integrate image and event features to enhance deblurring performance. For example, Sun et al. [35] introduced EFNet, which considers the inherent differences between event and image modalities and utilizes a cross-attention mechanism, improving upon previous methods that relied on simple concatenation or addition for feature fusion. Kim et al. [12] discovered that event streams contain both clear edge and motion cues and proposed a novel attention-based fusion method, allowing them to leverage these event characteristics to extract discriminative channel and spatial information from the blurred image features. Their approaches directly combine the heterogeneous cues contained in events (edge and motion) with RGB features indiscriminately. However, we find that sharp edge features and detailed motion cues are inherently conflicting in their representation and previous one-step fusion approaches [12, 35, 36] tend to bias the event modality toward capturing either edge or motion information, leading to suboptimal results. Specifically, as shown in Figure 1 (a), it is obvious that events provide sharp edge features within short exposure times(the $2^{nd}$ column), while accumulating the motion trajectory of objects over the entire exposure period(the $3^{rd}$ column). Therefore, sharp edges typically lack motion information, whereas representations rich in motion cues often compromise edge clarity. As a result, previous one-step fusion methods make it challenging for the model to accurately learn both edge and motion features from events, leading to coarse edges and blurred textures(Figure 1 (b)). Our insight is that edge and motion should be modeled separately and then integrated with RGB to respect both perspectives and avoid conflicts. To restore sharp edges and clear textures, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to independently harness the sharp edge features and rich motion cues provided by the event stream (Figure 1 (c)). Contributing to much better leverage of edge and motion cues, our model shows significant improvements in deblurring on both edges and textures compared to previous methods.", + "bbox": [ + 91, + 200, + 483, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A straightforward approach to restore sharp edges is to utilize event features accumulated during a short exposure (the second column in Figure 1 (a)) as edge information and fuse them with image features, but sacrificing the complete event features captured over the full exposure duration. Besides, relying on a fixed time interval to capture sharp edge features is unreliable, as varying motion speeds", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "result in different event trigger frequencies: excessive event accumulation can obscure edge details, while insufficient events may fail to capture edges fully. Therefore, our Edge-aware Sharpening module combines both short-exposure event features with complete event cues, using the former as an edge prior to extract complete edge cues by performing cross-attention with the latter.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for motion cues, previous work [11, 14-16, 44] has typically utilized events as a substitute for optical flow to align frames, achieving significant advancements in video motion deblurring. Some works [11, 33] have used event to predict optical flow for frame-to-frame alignment. FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\" [13, 39, 40], directly using RGB images and events to achieve frame-to-frame alignment without relying on intermediate steps like optical flow prediction. However, these approaches overlook the crucial role that motion cues play in motion deblurring: different motion patterns yield diverse blur kernels. By leveraging motion cues, we can explicitly estimate blur kernels for various motion patterns, enhancing the model's ability to adapt to a wide range of blurring scenarios. To effectively learn blur kernels of varying scales, previous studies [34, 37, 46] have attempted to expand the receptive field by adding more convolutional layers or using larger convolutional kernel sizes. However, traditional convolutional kernels have a fixed size, which limits their robustness when handling diverse blur kernels, reducing the model's adaptability in real-world scenarios. In our work, inspired by [22, 38, 42], we propose a Motion-driven Scale-adaptive Deblurring module that leverages the rich motion cues provided by the event stream to explicitly estimate different blur kernels, enabling adaptive adjustments to the receptive field to better accommodate a variety of blur patterns. By leveraging motion cues effectively, our single-frame motion deblurring model achieves performance that surpasses most video-based approaches.", + "bbox": [ + 511, + 200, + 903, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 532, + 645, + 834, + 659 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We find that the edge and motion cues provided by the event stream exhibit a conflict in their representation which is unexplored in the event community. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to effectively utilize these two distinct features.", + "(2) We find that motion cues are instrumental in learning variant blur kernels. Building on this insight, we utilize motion cues to achieve adaptive adjustments to the receptive field to accommodate variant blur kernels.", + "(3) Conducting extensive experiments on both synthetic and real-world datasets, our method demonstrates superior performance, significantly surpassing existing single-frame-based approaches and even most multi-frame-based methods." + ], + "bbox": [ + 511, + 664, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Frame-based Motion Deblurring on Image", + "text_level": 1, + "bbox": [ + 89, + 114, + 454, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the advancement of deep learning, convolutional neural networks (CNNs) have been increasingly applied to the motion deblurring. To better handle non-uniform and varying blur kernels, researchers have primarily explored three technical approaches: multi-scale architecture designs [7, 10, 18, 34, 37, 46], more sophisticated decoders [26, 29], and novel network layer designs [22, 30]. For example, SRN(Scale-Recurrent Network) [37] achieves deblurring of high-resolution images by guiding the process with low-resolution images through a multi-scale input-output approach. However, traditional convolutional kernels have spatial uniformity and a fixed receptive field, which makes them suboptimal for modeling dynamic scene deblurring. Although Li et al. [22] introduced deformable convolution to fit blur kernels of varying sizes, their approach still faces challenges in fitting non-uniform and varying blur kernels due to limited motion cues. In our work, we incorporate the event modality, leveraging its high temporal resolution and utilize deformable convolution to address this issue effectively.", + "bbox": [ + 89, + 138, + 483, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Event-based Motion Deblurring", + "text_level": 1, + "bbox": [ + 89, + 452, + 372, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early work, based on the imaging principles of event cameras and the process of how blurred images are generated, proposed the Event Double Integral (EDI) model [28, 47, 48]. However, due to the uncertain event-triggering thresholds and abundant noise, these methods have not yielded satisfactory results.", + "bbox": [ + 89, + 474, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other approaches [6, 12, 35, 36] focus on fusing event and image features, training a deblurring network in an end-to-end fashion for improved performance. For example, EFNet [35] utilizes cross-attention mechanism to fuse features from both modalities. REFID [36] employs bidirectional event recurrent (EVR) blocks to extract features from the event streams and introduces an Event-Guided Adaptive Channel Attention module to uncover discriminative channels within the event features. However, these approaches integrate event and image features in a single step, failing to fully account for the edge and motion cues inherent in the event features. In our research, we highlight that event features contain two inherently conflicting aspects: distinct edge information and abundant motion cues. Specifically, distinct edge features help restore the object's sharp boundaries, while abundant motion cues assist the model in learning diverse blur kernels, leading to better recovery of the object's texture details. To leverage each optimally, we introduce an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module.", + "bbox": [ + 89, + 566, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To utilize the rich motion cues offered by events, several studies [11, 14, 15, 33] focus on employing the event", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modality to handle video motion deblurring tasks. For example, FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\", utilizing bidirectional recurrent networks and cascade networks respectively, to achieve frame-to-frame alignment. However, these works use events for frame-to-frame alignment, overlooking the role of motion cues in learning variant blur kernels. In this paper, our Motion-Driven Scale-Adaptive Deblurring module effectively leverages motion cues to guide the adjustment of the model's receptive field, enabling it to adapt to diverse blur kernels. This approach ensures robust handling of various motion patterns and achieves competitive performance, even compared to state-of-the-art video-based methods.", + "bbox": [ + 511, + 90, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 316, + 604, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Firstly, we introduce the main architecture of our model in Sec. 3.1. In Sec. 3.2, an Edge-aware Sharpening module is proposed to utilize sharp edges captured during short exposure times from event to restore the clear boundaries of deblurring objects. In Sec. 3.3, to adaptively learn variant blur kernels, in the Motion-driven Scale-adaptive Deblurring module, we propose using motion cues to explicitly estimate the blur kernels associated with different motion patterns, enabling the model to adaptively adjust its receptive field for improved deblurring performance.", + "bbox": [ + 511, + 343, + 906, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 506, + 700, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based image motion deblurring aims to recover sharp images from a blurred image $\\mathbf{B}$ and an event stream $\\mathcal{E}$ . We use SCER [35] to represent the event $(\\mathcal{E} = \\{E_i\\}_{i=0}^{2N-1})$ . As shown in Figure 2, based on EFNet [35], our model consists of two U-Net [32] networks. During the first stage, an event encoding branch extracts multi-scale features from the event stream $\\mathcal{E}$ .", + "bbox": [ + 511, + 527, + 905, + 632 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{F _ {i} ^ {E} \\} _ {i = 0} ^ {3} = R e s (\\mathcal {E}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 633, + 906, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event features encompass both object edge information and motion cues. Previous approaches [12, 14, 35] often fused these event features directly, which can cause edge information and motion cues to conflict, leading to a model that neither fully utilizes the events to sharpen object edges nor effectively exploits the motion cues for deblurring. Specifically, event stream contains clear edge features from short exposure durations and motion cues accumulated over longer exposure times. This one-step fusion approach can easily cause the model to neglect one aspect of the event information. Therefore, an Edge-aware Sharpening module and a Motion-driven Scale-adaptive Deblurring module are proposed to fully leverage the features contained in the event stream respectively.", + "bbox": [ + 511, + 660, + 906, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{F _ {i} ^ {I} \\} _ {i = 1} ^ {3} = f _ {2} (f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}), F _ {i} ^ {m}) \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14734", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4262f2a4b8b4f113522ba28a561d03b020def9f764ef3f2a43ed9094a7dfa96.jpg", + "image_caption": [ + "Figure 2. The main architecture of our model. Based on EFNet [35], we propose a two-stage U-Net architecture to achieve coarse-to-fine motion deblurring. During the first stage, we separate event into edge features and motion cues to guide the model: edge features enhance the recovery of sharp image edges with an Edge-aware Sharpening module, while motion cues help the model learn diverse blur kernels through Motion-driven Scale-adaptive Deblurring module. In the second stage, similar to EFNet, the output from the first stage is concatenated with the original blurred image. An EMGC (Event Mask Gated Connection) [35] module is then applied to enhance features specifically in the regions events triggered." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{1}$ means Edge-aware Sharpening module and $f_{2}$ means Motion-driven Scale-adaptive Deblurring module. $F_{i}^{e}$ and $F_{i}^{m}$ represent edge features and motion cues from the event stream $\\mathcal{E}$ respectively. Next, a standard U-Net decoder layer with skip connections is employed to predict the deblurred image $I^{initial}$ .", + "bbox": [ + 89, + 604, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{D _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(D _ {i - 1} ^ {I}, F _ {i} ^ {I}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 707, + 482, + 728 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {i n i t i a l} = \\operatorname {c o n v} _ {3 \\times 3} \\left(D _ {3} ^ {I}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 739, + 482, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = P S N R L o s s (I ^ {\\text {i n i t i a l}}, I) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 768, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that event-triggered areas often suffer from more severe blur, in the second stage, the deblurred image $I^{initial}$ is concatenated with the blurred image B and then processed through another U-Net network to focus on restoring the event-triggered areas. EMGC(Event Mask Gated Connection) [35] module generates masks from events to create skip connections from the first stage to the", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "encoder of the second stage, guiding the model to concentrate on event-triggered areas. The encoder layer can be represented as:", + "bbox": [ + 511, + 606, + 906, + 651 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) = c o n v \\left(F _ {i} ^ {I}\\right) + c o n v \\left(F _ {i} ^ {I} \\times (1 - \\mathcal {M})\\right) \\\\ + \\operatorname {c o n v} \\left(D _ {i} ^ {I}\\right) + \\operatorname {c o n v} \\left(D _ {i} ^ {I} \\times \\mathcal {M}\\right) \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 654, + 903, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\hat {F} _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(\\hat {F} _ {i - 1} ^ {I}\\right) + E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 699, + 903, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ represents the mask generated from event stream $\\mathcal{E}$ . Pixels where events triggered are set to 1, while others are set to 0. As a result, feature maps from encoder overlook the event-triggered areas, which may contain severe blur, while feature maps from the decoder enhance these regions, reflecting the restored blurred content. Finally, the final deblurred result $I^{final}$ is obtained after processing through a decoder identical to the one used in the first stage. Loss function is expressed as follows:", + "bbox": [ + 511, + 720, + 906, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {2} = P S N R L o s s \\left(I ^ {f i n a l}, I\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 861, + 903, + 880 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\omega \\times L _ {1} + (1 - \\omega) \\times L _ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14735", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Edge-aware Sharpening Module", + "text_level": 1, + "bbox": [ + 89, + 90, + 375, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the second column in Figure 1 (a), the accumulation of events over a short period often provides clear edge features of objects, which help restore sharp boundaries for blurred objects. Consequently, FEVD [15] employs fast Fourier transform to simultaneously capture the high-frequency signals, which typically contain detailed edge cues, from both events and images. However, event features often contain abundant motion cues, which can blur the sharp edge features and make it difficult to extract clear edge information from events. On the other hand, even objects undergoing rapid motion can accumulate clear edge features within short exposure times. This characteristic is particularly valuable for preserving fine details, as edges play a crucial role in maintaining the integrity of object boundaries and structures in the image. However, when combined with motion cues, which are more prominent in longer exposure times, these edge features can become diluted or lost. Therefore, it is essential to selectively enhance the edge information from the event stream to ensure that sharp boundaries are preserved, while still utilizing the motion information to handle blur effectively.", + "bbox": [ + 89, + 112, + 483, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we propose an Edge-aware Sharpening module that firstly combines the event features accumulated over a short period with the image to enhance the edges of moving objects. These enhanced features serve as queries, allowing the attention mechanism to effectively integrate the edge features from the event stream. Specifically, SCER [35] uses the midpoint as a reference and accumulates $2 \\times N$ frames of events symmetrically from this point. We select the channels closest to the midpoint as edge priors to enhance the image features. Thus, our Edge-aware Sharpening module can be represented as:", + "bbox": [ + 89, + 430, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {e} = \\operatorname {c o n v} \\left(E _ {N - 1} \\| E _ {N}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 606, + 480, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}) =\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 625, + 336, + 643 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA t t \\left(\\operatorname {c o n v} \\left(R e s \\left(F _ {i - 1} ^ {I}\\right) \\| F _ {i} ^ {e}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 646, + 480, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\parallel$ means connection between features and $Att(x,y,z)$ denotes an attention mechanism where $x$ serves as the query, $y$ as the key, and $z$ as the value.", + "bbox": [ + 89, + 672, + 483, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Motion-driven Scale-adaptive Deblurring Module", + "text_level": 1, + "bbox": [ + 89, + 727, + 483, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous approaches have often concentrated on utilizing event streams for frame-to-frame alignment to achieve video motion deblurring. However, these methods frequently overlook the importance of motion features in the context of static image deblurring. The motion cues of objects during exposure time help us explicitly learn variant blur kernels.", + "bbox": [ + 89, + 763, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the third column in Figure 1 (a), the girl exhibits a larger blur kernel, causing the edges accumulated", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in the event stream to appear broader and more diffused. This effect is in stark contrast to the background crowd, where the blur kernel is smaller due to less motion, resulting in sharper and more defined edges in the event accumulation. These varying blur kernels indicate differences in the degree of motion across the scene, which are essential for accurately restoring fine textures in the deblurring process. To effectively handle such diverse blur kernels, we introduce a Motion-driven Scale-adaptive Deblurring module that adaptively adjusts the convolutional receptive field based on the degree of motion captured in the event stream. This is achieved through the use of deformable convolutions, which enable the network to learn dynamic and spatially flexible convolutional kernels. This approach allows the model to handle both large, motion-induced blur (such as that of the girl) and smaller, subtle blur (such as that of the background crowd) simultaneously, improving the overall deblurring performance.", + "bbox": [ + 511, + 90, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned earlier, the event stream contains both edge and motion features. To highlight the motion cues of objects, it is essential to separate the motion cues from the event features. Specifically, we use events from both sides of SCER as motion priors $(E_0$ and $E_{2N - 1})$ and apply cross-attention to capture detailed motion cues from the event features.", + "bbox": [ + 511, + 364, + 908, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {m} = \\operatorname {A t t} \\left(\\operatorname {c o n v} \\left(E _ {0} \\| E _ {2 N - 1}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 479, + 906, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we calculate offsets with the motion cues to explicitly estimate variant blur kernels with different scales, and employ deformable convolutions to capture contextual features, enhancing the ability to express blur effectively.", + "bbox": [ + 511, + 511, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta = \\operatorname {c o n v} \\left(F _ {i} ^ {m}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 585, + 903, + 601 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {2} \\left(f, F _ {i} ^ {m}\\right) = D C N (f, \\Theta) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 614, + 903, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Theta$ means the offsets of deformable convolution.", + "bbox": [ + 511, + 640, + 872, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 670, + 638, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 696, + 805, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic Dataset. To make a fair comparison with previous methods based on events and images, we selected the GoPro dataset [27], which is widely used for deblurring tasks. We utilized ESIM [31] to generate the event stream, applying the same parameters as other studies, specifically using a Gaussian normal distribution with parameters $\\mathcal{N}(\\mu = 0.2,\\sigma = 0.03)$ .", + "bbox": [ + 511, + 718, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world Datasets. To validate the deblurring capabilities of event data in real-world scenarios, REBlur [35] collects events from Dynamic and Active Pixel Vision Sensor (DAVIS) [2] in real environments and uses an electronically controlled stepping motor to relocate the slide-rail", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14736", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to the coordinate corresponding to the timestamp t. However, DAVIS-346 camera has low image quality and resolution. Therefore, high resolution and quality event-based real-world deblurring datasets REVD [15] and EVRB [14] are proposed with Triple-axis Hybrid Camera System.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) [43] are proposed to assess image quality. PSNR calculates the ratio of signal power to noise, with higher values indicating closer similarity to the reference image, though it mainly reflects pixel-level fidelity. SSIM, on the other hand, measures structural, luminance, and contrast similarities, providing a more perceptually relevant quality assessment, with values closer to 1 indicating better similarity.", + "bbox": [ + 89, + 167, + 480, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our model supports end-to-end training with a batch size of 16. The N of SCER [35] is set to 8. For data augmentation, we randomly crop images to $256 \\times 256$ pixels and apply random horizontal and vertical flips. We chose Adam [17] as the optimizer. Training runs for 400,000 iterations, starting with a learning rate of 2e-4, which is progressively reduced to a minimum of 1e-7 using CosineAnnealingLR [25].", + "bbox": [ + 89, + 301, + 480, + 422 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f238a818858af0e03e75d285c3b83dd7f3603cb5a81bddf7d364324bfdbd98a9.jpg", + "table_caption": [ + "Table 1. Quantitative results on the GoPro [27] dataset. $\\uparrow$ indicates that higher values correspond to better performance. The numbers in “()” indicate the publication year of the referenced papers. In the second column, “Single” in black means frame-based method, and red represents methods based on single-frame and event while green means multi-frame event-based methods. The best results are indicated in bold." + ], + "table_footnote": [], + "table_body": "
MethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
Restormer [45] (22')Single32.920.961
NAFNet [4] (22')Single33.690.967
FFTformer [18] (23')Single34.210.969
GRL-B [23] (23')Single33.930.968
MLWNet [10] (24')Single33.830.968
Jianget al. [11] (20')Single31.790.949
REDNet [44] (21')Single35.070.971
EFNet [35] (22')Single35.460.973
SpkNet [5] (23')Single36.120.971
D2Nets [33] (21')3 frames31.760.943
UEVD [12] (22')2 frames35.480.971
REFID [36] (23')2 frames35.910.973
FEVD [15] (24')Multi-frames36.700.978
CMTA-5 [14] (24')5 frames36.550.977
CMTA-7 [14] (24')7 frames36.780.977
OursSingle36.700.977
", + "bbox": [ + 93, + 545, + 478, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compared to State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 848, + 421, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic Dataset. We evaluate our proposed network against state-of-the-art motion deblurring methods on the", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2207bc5ecf85d9bff28b66ddd1868d6608c353a6d0450ed30d03f902c6456383.jpg", + "table_caption": [ + "Table 2. Quantitative results on three real-world datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
REBlurSNR+ [37]Single36.870.970
HINet+ [3]Single37.680.973
EFNet [35]Single38.120.975
OursSingle38.530.977
EVRBUEVD [12]2 frames30.550.915
EFNet [35]Single30.980.927
REFID [36]2 frames30.330.918
CMTA [14]Multi-frames31.380.927
OursSingle32.050.936
REVDeSL-Net [41]Single26.990.787
EFNet [35]Single31.750.920
REDNet [44]Single31.900.920
UEVD [12]2 frames31.970.921
FEVD [15]Multi-frames32.990.933
OursSingle32.140.922
", + "bbox": [ + 517, + 113, + 897, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GoPro [27] dataset. The comparison covers three categories: frame-based image deblurring, event(spike)-based image deblurring and event-based video deblurring. As shown in Table 1, in comparison to SpkNet [5], the existing state-of-the-art spike-based image motion deblurring method, our model achieves an impressive performance boost of $+0.58$ dB. Moreover, compared to CMTA [14], the leading video motion deblurring method, our single-frame model outperforms CMTA-5 (which uses five frames for inference) by $+0.15$ dB and is only $-0.08$ dB behind CMTA-7. Thus, our model demonstrates the remarkable effectiveness of event features in the task of image motion deblurring. We show qualitative results on GoPro in Figure 3, RVRT [24], a video-based motion deblurring method, models long-term temporal dependencies to provide contextual references for deblurring the current frame. MLWNet [10] stands as the leading single-frame motion deblurring approach, while EFNet [35] is an event-based image deblurring method. It's evident that our approach achieves superior detail restoration compared to these previous methods.", + "bbox": [ + 511, + 421, + 903, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world Datasets. To assess our model's effectiveness in real-world settings, we utilized three datasets: REBlur [35], REVD [15], and EVRB [14]. For REBlur, given its limited dataset size, we first train the model on the GoPro [27] dataset and then finetune it on REBlur training set. For FEVD and EVRB, we retrained the model according to the training and testing set splits provided in the official guidelines. As shown in Table 2, our method surpasses event-based image motion deblurring approaches across all three datasets and even outperforms the current top video-based methods on EVRB [14].", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14737", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9798852e165c9a5c59a105ed6da9051949ce718f0e7f77f36d7842b7f645dc59.jpg", + "image_caption": [ + "Figure 3. Visual results with our model and other SOTA methods on GoPro dataset." + ], + "image_footnote": [], + "bbox": [ + 127, + 89, + 844, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 738, + 243, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our baseline model $(B)$ is a standard two-stage U-Net that does not utilize any event features. As shown in Table 3, it is evident that the integration of event modalities results in significant performance enhancements, as demonstrated by both attention-based event fusion and event-driven deformable convolutions, which yield impressive outcomes. As shown in Figure 4, the baseline performs poorly in deblurring, failing to effectively restore either sharp edges or clear textures. And with the introduction of the event", + "bbox": [ + 89, + 763, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "modality, the visual results improve significantly.", + "bbox": [ + 511, + 739, + 839, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the first part in Table 3, simply applying cross-attention to fuse event and image modalities $(B + Event^{Att})$ , yields promising results. However, this straightforward fusion approach overlooks the inherent conflict between motion and edge features in event streams, which prevents the model from accurately capturing sharp edge details. As shown in the first example of Figure 4, when image features are fused with events, the text in the yellow box becomes recognizable but still those in blue", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14738", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25730c6f6a6328d98c9c5ac4f2c1a08f32b20dac26e4efb9c46d28ac8dc3b1d1.jpg", + "image_caption": [ + "Figure 4. Visual results of ablation study on EVRB [14] dataset. It is recommended to zoom in for a clearer view of the details." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 880, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b64544ab5e4973a85ee202be68d18bd295688d2dcbf7c1f679c2b4ee799db00.jpg", + "table_caption": [ + "Table 3. Ablation study on EVRB [14] dataset. $B$ represents our baseline, which does not utilize event streams to enhance the image. $Event^{Att}$ denotes applying cross-attention to integrate image with event features. $Edge^{Att}$ means our Edge-aware Sharpening module. $Img^{DCN}$ and $Event^{DCN}$ represent utilizing image features and event features to calculate offsets for the deformable convolution respectively. $Motion^{DCN}$ is our Motion-driven Scale-adaptive Deblurring module." + ], + "table_footnote": [], + "table_body": "
MethodMetrics
PSNRSSIM
B28.650.900
B+EventAtt30.710.923
B+EdgeAtt31.130.929
B+ImgDCN28.800.894
B+EventDCN30.510.921
B+MotionDCN31.630.930
Ours(B+EdgeAtt + MotionDCN)32.050.936
", + "bbox": [ + 99, + 542, + 470, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "box retain significant blurriness. Instead, our Edge-aware Sharpening module $(B + Edge^{Att})$ strengthens the representation of edge features within the events, producing a more refined result.", + "bbox": [ + 89, + 729, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second part of the Table 3 demonstrates the value of motion cues in learning diverse blur kernels. Initially, we utilize image features to determine offsets for deformable convolutions $(B + Img^{DCN})$ ; however, because image lacks the motion cues, they are unable to accurately capture the diverse blur kernels that correspond to different motion characteristics. As shown in the second example in Fig-", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ure 4, using image features to estimate the blur kernel still results in artifacts around the \"No Parking Sign\". Conversely, training deformable convolutions with event features $(B + Event^{DCN})$ leads to a better performance, but it still results in problematic artifacts in some local regions (in the yellow box). To further improve the capture of motion information, our proposed Motion-driven Scale-adaptive Deblurring module $(B + Motion^{DCN})$ accurately captures object motion cues and estimates non-uniform and varying blur kernels, significantly minimizing artifacts both globally and locally.", + "bbox": [ + 511, + 422, + 906, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, by effectively fusing both motion and edge cues, our model demonstrates substantial improvements in the restoration of both sharp edges and fine texture details.", + "bbox": [ + 511, + 589, + 906, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we observe that event features inherently contain conflicting edge and motion cues: edge features accumulate over short exposure times lack sufficient motion information, while motion cues accumulate over long exposure times compromise edge clarity. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to separately leverage these two aspects of cues. By incorporating edge and motion priors, we enhance the corresponding cues in event features: edge cues restore sharp object edges, while motion cues guide deformable convolutions to adaptively estimate blur kernels, adjusting the receptive field for diverse motion patterns. Extensive experiments on synthetic and real-world datasets demonstrate the superior performance of our approach.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14739", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work is jointly supported by the National Natural Science Foundation of China (NSFC) under Grant 62261160576 and Grant 62203024, the Beijing Natural Science Foundation (4252026), the Research and Development Program of Beijing Municipal Education Commission (KM202310005027), and the Fundamental Research Funds for the Central Universities of China. This research work is also supported by the Big Data Computing Center of Southeast University.", + "bbox": [ + 89, + 114, + 483, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 266, + 187, + 281 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Bahat, Netelee Efrat, and Michal Irani. Non-uniform blind deblurring by reblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3286-3294, 2017. 1", + "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 5", + "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6", + "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 6", + "[5] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems, 36:70279-70292, 2023. 6", + "[6] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2, 3", + "[7] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 3", + "[8] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In ACM SIG International Conference on Computer Graphics and Interactive Techniques, pages 787-794. 2006. 1", + "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 2", + "[10] Xin Gao, Tianheng Qiu, Xinyu Zhang, Hanlin Bai, Kang Liu, Xuan Huang, Hu Wei, Guoying Zhang, and Huaping Liu." + ], + "bbox": [ + 93, + 290, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2733-2742, 2024. 3, 6", + "[11] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2, 3, 6", + "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 3, 6", + "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2", + "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cross-modal temporal alignment for event-guided video deblurring. European Conference on Computer Vision, 2024. 2, 3, 6, 8", + "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2, 3, 5, 6", + "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 2", + "[17] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 6", + "[19] Jan Kotera, Filip Šroubek, and Peyman Milanfar. Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In Computer Analysis of Images and Patterns: 15th International Conference, CAIP 2013, York, UK, August 27-29, 2013, Proceedings, Part II 15, pages 59-66. Springer, 2013. 1", + "[20] Dilip Krishnan, Terence Tay, and Rob Fergus. Blind deconvolution using a normalized sparsity measure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 233-240. IEEE, 2011.", + "[21] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1971. IEEE, 2009. 1", + "[22] Jichun Li, Weimin Tan, and Bo Yan. Perceptual variousness motion deblurring with light global context refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4116-4125, 2021. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14740", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 6", + "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6", + "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[26] Xintian Mao, Qingli Li, and Yan Wang. Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25681-25690, 2024. 3", + "[27] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5, 6", + "[28] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3", + "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 3", + "[30] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurringli2021 perceptual. In Association for the Advancement of Artificial Intelligence, pages 11882-11889, 2020. 3", + "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. pages 969-982. PMLR, 2018. 5", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241. Springer, 2015. 3", + "[33] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2021. 2, 3, 6", + "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606–3615, 2020. 2, 3", + "[35] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 1, 2, 3, 4, 5, 6", + "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 2, 3, 6", + "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 3, 6", + "[38] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 2", + "[39] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 2", + "[40] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2", + "[41] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In European Conference on Computer Vision, pages 155-171. Springer, 2020. 6", + "[42] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 0–0, 2019. 2", + "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6", + "[44] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2, 6", + "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6", + "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 3", + "[47] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2, 3", + "[48] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10734-10744, 2023. 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14742", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_model.json b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..870ed69aaa361acc30c06d71889aa883d5c9eccf --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_model.json @@ -0,0 +1,2037 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.847, + 0.177 + ], + "angle": 0, + "content": "Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.215, + 0.733, + 0.235 + ], + "angle": 0, + "content": "Yufei Zhu\\(^{1,2}\\), Hao Chen\\(^{*1,2}\\), Yongjian Deng\\(^{3}\\), and Wei You\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.246, + 0.874, + 0.3 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Engineering, Southeast University, Nanjing, China \n\\(^{2}\\)Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China." + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.3, + 0.819, + 0.318 + ], + "angle": 0, + "content": "\\(^{3}\\)College of Computer Science, Beijing University of Technology, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.322, + 0.777, + 0.341 + ], + "angle": 0, + "content": "\\(^{4}\\)Advanced Computing and Storage Lab, Huawei Technologies, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.34, + 0.83, + 0.355 + ], + "angle": 0, + "content": "{220232213，haochen303}@seu.edu.cn，yjdeng@bjut.edu.cn，youwei22@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.399, + 0.327, + 0.415 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.431, + 0.485, + 0.81 + ], + "angle": 0, + "content": "Traditional motion deblurring methods struggle to effectively model motion information within the exposure time. Recently, event cameras have attracted significant research interest for its ability to model motion cues over the exposure duration. However, these methods directly fuse event features with image, overlooking the intrinsic heterogeneity of events. In this paper, we identify that the event modality contains two conflicting types of information: edge features and motion cues. Events accumulated over a short exposure period capture sharp edge details but lose motion information, while those accumulated over a long exposure period blur edge details due to motion. To address this issue, we propose a simple yet effective approach to disentangle these two cues from event features and employ an edge-aware sharpening module along with motion-driven scale-adaptive deblurring module to fully leverage both. Specifically, the first module aids in restoring sharp edges by leveraging the clear edge features provided by events, while the second module leverages motion cues to learn diverse blur kernels, adaptively adjusting the receptive field for optimal deblurring. Extensive experiments on synthetic and real-world datasets validate the effectiveness of our approach and yield a substantial improvement over state-of-the-art single-frame methods and surpasses most multi-frame-based methods. Code will be publicly available." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.827, + 0.224, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Motion blur is often caused by camera shake or the movement of objects. The task of motion deblurring [1, 8, 19-21]" + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.402, + 0.555, + 0.42 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.4, + 0.67, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.4, + 0.778, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.4, + 0.883, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.499, + 0.555, + 0.516 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.508, + 0.648, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.508, + 0.773, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.517, + 0.893, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.621, + 0.554, + 0.636 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.618, + 0.648, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.62, + 0.774, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.65, + 0.888, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.742, + 0.908, + 0.798 + ], + "angle": 0, + "content": "Figure 1. (a) Visualization of SCER(Symmetric Cumulative Event Representation) [35]. (b) Previous fusion methods between events and image. (c) Ours. Separate events into edge features and motion cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "aims to recover sharp images from blurred ones by restoring clear edges and capturing fine details, ultimately enhancing the image quality. However, since image features lack the ability to model motion cues over the exposure time, they cannot accurately estimate the variant blur ker" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "nels present in real-world scenarios. This limitation often results in subpar performance of current methods on real-world datasets. Recently, event cameras [2, 9], bio-inspired sensors that asynchronously capture changes in pixel intensities, have been leveraged to aid motion deblurring thanks to their microsecond-level latency and low power consumption." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.201, + 0.485, + 0.792 + ], + "angle": 0, + "content": "Several studies [6, 11, 35, 47, 48] aim to integrate image and event features to enhance deblurring performance. For example, Sun et al. [35] introduced EFNet, which considers the inherent differences between event and image modalities and utilizes a cross-attention mechanism, improving upon previous methods that relied on simple concatenation or addition for feature fusion. Kim et al. [12] discovered that event streams contain both clear edge and motion cues and proposed a novel attention-based fusion method, allowing them to leverage these event characteristics to extract discriminative channel and spatial information from the blurred image features. Their approaches directly combine the heterogeneous cues contained in events (edge and motion) with RGB features indiscriminately. However, we find that sharp edge features and detailed motion cues are inherently conflicting in their representation and previous one-step fusion approaches [12, 35, 36] tend to bias the event modality toward capturing either edge or motion information, leading to suboptimal results. Specifically, as shown in Figure 1 (a), it is obvious that events provide sharp edge features within short exposure times(the \\(2^{nd}\\) column), while accumulating the motion trajectory of objects over the entire exposure period(the \\(3^{rd}\\) column). Therefore, sharp edges typically lack motion information, whereas representations rich in motion cues often compromise edge clarity. As a result, previous one-step fusion methods make it challenging for the model to accurately learn both edge and motion features from events, leading to coarse edges and blurred textures(Figure 1 (b)). Our insight is that edge and motion should be modeled separately and then integrated with RGB to respect both perspectives and avoid conflicts. To restore sharp edges and clear textures, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to independently harness the sharp edge features and rich motion cues provided by the event stream (Figure 1 (c)). Contributing to much better leverage of edge and motion cues, our model shows significant improvements in deblurring on both edges and textures compared to previous methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "A straightforward approach to restore sharp edges is to utilize event features accumulated during a short exposure (the second column in Figure 1 (a)) as edge information and fuse them with image features, but sacrificing the complete event features captured over the full exposure duration. Besides, relying on a fixed time interval to capture sharp edge features is unreliable, as varying motion speeds" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "result in different event trigger frequencies: excessive event accumulation can obscure edge details, while insufficient events may fail to capture edges fully. Therefore, our Edge-aware Sharpening module combines both short-exposure event features with complete event cues, using the former as an edge prior to extract complete edge cues by performing cross-attention with the latter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.905, + 0.641 + ], + "angle": 0, + "content": "As for motion cues, previous work [11, 14-16, 44] has typically utilized events as a substitute for optical flow to align frames, achieving significant advancements in video motion deblurring. Some works [11, 33] have used event to predict optical flow for frame-to-frame alignment. FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\" [13, 39, 40], directly using RGB images and events to achieve frame-to-frame alignment without relying on intermediate steps like optical flow prediction. However, these approaches overlook the crucial role that motion cues play in motion deblurring: different motion patterns yield diverse blur kernels. By leveraging motion cues, we can explicitly estimate blur kernels for various motion patterns, enhancing the model's ability to adapt to a wide range of blurring scenarios. To effectively learn blur kernels of varying scales, previous studies [34, 37, 46] have attempted to expand the receptive field by adding more convolutional layers or using larger convolutional kernel sizes. However, traditional convolutional kernels have a fixed size, which limits their robustness when handling diverse blur kernels, reducing the model's adaptability in real-world scenarios. In our work, inspired by [22, 38, 42], we propose a Motion-driven Scale-adaptive Deblurring module that leverages the rich motion cues provided by the event stream to explicitly estimate different blur kernels, enabling adaptive adjustments to the receptive field to better accommodate a variety of blur patterns. By leveraging motion cues effectively, our single-frame motion deblurring model achieves performance that surpasses most video-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.646, + 0.835, + 0.66 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.665, + 0.905, + 0.756 + ], + "angle": 0, + "content": "(1) We find that the edge and motion cues provided by the event stream exhibit a conflict in their representation which is unexplored in the event community. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to effectively utilize these two distinct features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.905, + 0.821 + ], + "angle": 0, + "content": "(2) We find that motion cues are instrumental in learning variant blur kernels. Building on this insight, we utilize motion cues to achieve adaptive adjustments to the receptive field to accommodate variant blur kernels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "(3) Conducting extensive experiments on both synthetic and real-world datasets, our method demonstrates superior performance, significantly surpassing existing single-frame-based approaches and even most multi-frame-based methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.665, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.234, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.455, + 0.132 + ], + "angle": 0, + "content": "2.1. Frame-based Motion Deblurring on Image" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.485, + 0.442 + ], + "angle": 0, + "content": "With the advancement of deep learning, convolutional neural networks (CNNs) have been increasingly applied to the motion deblurring. To better handle non-uniform and varying blur kernels, researchers have primarily explored three technical approaches: multi-scale architecture designs [7, 10, 18, 34, 37, 46], more sophisticated decoders [26, 29], and novel network layer designs [22, 30]. For example, SRN(Scale-Recurrent Network) [37] achieves deblurring of high-resolution images by guiding the process with low-resolution images through a multi-scale input-output approach. However, traditional convolutional kernels have spatial uniformity and a fixed receptive field, which makes them suboptimal for modeling dynamic scene deblurring. Although Li et al. [22] introduced deformable convolution to fit blur kernels of varying sizes, their approach still faces challenges in fitting non-uniform and varying blur kernels due to limited motion cues. In our work, we incorporate the event modality, leveraging its high temporal resolution and utilize deformable convolution to address this issue effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.373, + 0.47 + ], + "angle": 0, + "content": "2.2. Event-based Motion Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.567 + ], + "angle": 0, + "content": "Early work, based on the imaging principles of event cameras and the process of how blurred images are generated, proposed the Event Double Integral (EDI) model [28, 47, 48]. However, due to the uncertain event-triggering thresholds and abundant noise, these methods have not yielded satisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Other approaches [6, 12, 35, 36] focus on fusing event and image features, training a deblurring network in an end-to-end fashion for improved performance. For example, EFNet [35] utilizes cross-attention mechanism to fuse features from both modalities. REFID [36] employs bidirectional event recurrent (EVR) blocks to extract features from the event streams and introduces an Event-Guided Adaptive Channel Attention module to uncover discriminative channels within the event features. However, these approaches integrate event and image features in a single step, failing to fully account for the edge and motion cues inherent in the event features. In our research, we highlight that event features contain two inherently conflicting aspects: distinct edge information and abundant motion cues. Specifically, distinct edge features help restore the object's sharp boundaries, while abundant motion cues assist the model in learning diverse blur kernels, leading to better recovery of the object's texture details. To leverage each optimally, we introduce an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To utilize the rich motion cues offered by events, several studies [11, 14, 15, 33] focus on employing the event" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.302 + ], + "angle": 0, + "content": "modality to handle video motion deblurring tasks. For example, FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\", utilizing bidirectional recurrent networks and cascade networks respectively, to achieve frame-to-frame alignment. However, these works use events for frame-to-frame alignment, overlooking the role of motion cues in learning variant blur kernels. In this paper, our Motion-Driven Scale-Adaptive Deblurring module effectively leverages motion cues to guide the adjustment of the model's receptive field, enabling it to adapt to diverse blur kernels. This approach ensures robust handling of various motion patterns and achieves competitive performance, even compared to state-of-the-art video-based methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.605, + 0.334 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Firstly, we introduce the main architecture of our model in Sec. 3.1. In Sec. 3.2, an Edge-aware Sharpening module is proposed to utilize sharp edges captured during short exposure times from event to restore the clear boundaries of deblurring objects. In Sec. 3.3, to adaptively learn variant blur kernels, in the Motion-driven Scale-adaptive Deblurring module, we propose using motion cues to explicitly estimate the blur kernels associated with different motion patterns, enabling the model to adaptively adjust its receptive field for improved deblurring performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.702, + 0.521 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.633 + ], + "angle": 0, + "content": "Event-based image motion deblurring aims to recover sharp images from a blurred image \\(\\mathbf{B}\\) and an event stream \\(\\mathcal{E}\\). We use SCER [35] to represent the event \\((\\mathcal{E} = \\{E_i\\}_{i=0}^{2N-1})\\). As shown in Figure 2, based on EFNet [35], our model consists of two U-Net [32] networks. During the first stage, an event encoding branch extracts multi-scale features from the event stream \\(\\mathcal{E}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.634, + 0.907, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\{F _ {i} ^ {E} \\} _ {i = 0} ^ {3} = R e s (\\mathcal {E}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.873 + ], + "angle": 0, + "content": "Event features encompass both object edge information and motion cues. Previous approaches [12, 14, 35] often fused these event features directly, which can cause edge information and motion cues to conflict, leading to a model that neither fully utilizes the events to sharpen object edges nor effectively exploits the motion cues for deblurring. Specifically, event stream contains clear edge features from short exposure durations and motion cues accumulated over longer exposure times. This one-step fusion approach can easily cause the model to neglect one aspect of the event information. Therefore, an Edge-aware Sharpening module and a Motion-driven Scale-adaptive Deblurring module are proposed to fully leverage the features contained in the event stream respectively." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\{F _ {i} ^ {I} \\} _ {i = 1} ^ {3} = f _ {2} (f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}), F _ {i} ^ {m}) \\qquad (2)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14734" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.891, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.496, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 2. The main architecture of our model. Based on EFNet [35], we propose a two-stage U-Net architecture to achieve coarse-to-fine motion deblurring. During the first stage, we separate event into edge features and motion cues to guide the model: edge features enhance the recovery of sharp image edges with an Edge-aware Sharpening module, while motion cues help the model learn diverse blur kernels through Motion-driven Scale-adaptive Deblurring module. In the second stage, similar to EFNet, the output from the first stage is concatenated with the original blurred image. An EMGC (Event Mask Gated Connection) [35] module is then applied to enhance features specifically in the regions events triggered." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.698 + ], + "angle": 0, + "content": "where \\( f_{1} \\) means Edge-aware Sharpening module and \\( f_{2} \\) means Motion-driven Scale-adaptive Deblurring module. \\( F_{i}^{e} \\) and \\( F_{i}^{m} \\) represent edge features and motion cues from the event stream \\( \\mathcal{E} \\) respectively. Next, a standard U-Net decoder layer with skip connections is employed to predict the deblurred image \\( I^{initial} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.708, + 0.483, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\left\\{D _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(D _ {i - 1} ^ {I}, F _ {i} ^ {I}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.741, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\nI ^ {i n i t i a l} = \\operatorname {c o n v} _ {3 \\times 3} \\left(D _ {3} ^ {I}\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.769, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\nL _ {1} = P S N R L o s s (I ^ {\\text {i n i t i a l}}, I) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Considering that event-triggered areas often suffer from more severe blur, in the second stage, the deblurred image \\( I^{initial} \\) is concatenated with the blurred image B and then processed through another U-Net network to focus on restoring the event-triggered areas. EMGC(Event Mask Gated Connection) [35] module generates masks from events to create skip connections from the first stage to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.652 + ], + "angle": 0, + "content": "encoder of the second stage, guiding the model to concentrate on event-triggered areas. The encoder layer can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.655, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) = c o n v \\left(F _ {i} ^ {I}\\right) + c o n v \\left(F _ {i} ^ {I} \\times (1 - \\mathcal {M})\\right) \\\\ + \\operatorname {c o n v} \\left(D _ {i} ^ {I}\\right) + \\operatorname {c o n v} \\left(D _ {i} ^ {I} \\times \\mathcal {M}\\right) \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.7, + 0.905, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\hat {F} _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(\\hat {F} _ {i - 1} ^ {I}\\right) + E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.907, + 0.858 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) represents the mask generated from event stream \\(\\mathcal{E}\\). Pixels where events triggered are set to 1, while others are set to 0. As a result, feature maps from encoder overlook the event-triggered areas, which may contain severe blur, while feature maps from the decoder enhance these regions, reflecting the restored blurred content. Finally, the final deblurred result \\(I^{final}\\) is obtained after processing through a decoder identical to the one used in the first stage. Loss function is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.862, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\nL _ {2} = P S N R L o s s \\left(I ^ {f i n a l}, I\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\omega \\times L _ {1} + (1 - \\omega) \\times L _ {2} \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14735" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.376, + 0.108 + ], + "angle": 0, + "content": "3.2. Edge-aware Sharpening Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.43 + ], + "angle": 0, + "content": "As shown in the second column in Figure 1 (a), the accumulation of events over a short period often provides clear edge features of objects, which help restore sharp boundaries for blurred objects. Consequently, FEVD [15] employs fast Fourier transform to simultaneously capture the high-frequency signals, which typically contain detailed edge cues, from both events and images. However, event features often contain abundant motion cues, which can blur the sharp edge features and make it difficult to extract clear edge information from events. On the other hand, even objects undergoing rapid motion can accumulate clear edge features within short exposure times. This characteristic is particularly valuable for preserving fine details, as edges play a crucial role in maintaining the integrity of object boundaries and structures in the image. However, when combined with motion cues, which are more prominent in longer exposure times, these edge features can become diluted or lost. Therefore, it is essential to selectively enhance the edge information from the event stream to ensure that sharp boundaries are preserved, while still utilizing the motion information to handle blur effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.597 + ], + "angle": 0, + "content": "To address this issue, we propose an Edge-aware Sharpening module that firstly combines the event features accumulated over a short period with the image to enhance the edges of moving objects. These enhanced features serve as queries, allowing the attention mechanism to effectively integrate the edge features from the event stream. Specifically, SCER [35] uses the midpoint as a reference and accumulates \\(2 \\times N\\) frames of events symmetrically from this point. We select the channels closest to the midpoint as edge priors to enhance the image features. Thus, our Edge-aware Sharpening module can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.607, + 0.482, + 0.623 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {e} = \\operatorname {c o n v} \\left(E _ {N - 1} \\| E _ {N}\\right) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.626, + 0.338, + 0.644 + ], + "angle": 0, + "content": "\\[\nf _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.647, + 0.482, + 0.665 + ], + "angle": 0, + "content": "\\[\nA t t \\left(\\operatorname {c o n v} \\left(R e s \\left(F _ {i - 1} ^ {I}\\right) \\| F _ {i} ^ {e}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.485, + 0.72 + ], + "angle": 0, + "content": "where \\( \\parallel \\) means connection between features and \\( Att(x,y,z) \\) denotes an attention mechanism where \\( x \\) serves as the query, \\( y \\) as the key, and \\( z \\) as the value." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.484, + 0.758 + ], + "angle": 0, + "content": "3.3. Motion-driven Scale-adaptive Deblurring Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.869 + ], + "angle": 0, + "content": "Previous approaches have often concentrated on utilizing event streams for frame-to-frame alignment to achieve video motion deblurring. However, these methods frequently overlook the importance of motion features in the context of static image deblurring. The motion cues of objects during exposure time help us explicitly learn variant blur kernels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in the third column in Figure 1 (a), the girl exhibits a larger blur kernel, causing the edges accumulated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.363 + ], + "angle": 0, + "content": "in the event stream to appear broader and more diffused. This effect is in stark contrast to the background crowd, where the blur kernel is smaller due to less motion, resulting in sharper and more defined edges in the event accumulation. These varying blur kernels indicate differences in the degree of motion across the scene, which are essential for accurately restoring fine textures in the deblurring process. To effectively handle such diverse blur kernels, we introduce a Motion-driven Scale-adaptive Deblurring module that adaptively adjusts the convolutional receptive field based on the degree of motion captured in the event stream. This is achieved through the use of deformable convolutions, which enable the network to learn dynamic and spatially flexible convolutional kernels. This approach allows the model to handle both large, motion-induced blur (such as that of the girl) and smaller, subtle blur (such as that of the background crowd) simultaneously, improving the overall deblurring performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.909, + 0.469 + ], + "angle": 0, + "content": "As mentioned earlier, the event stream contains both edge and motion features. To highlight the motion cues of objects, it is essential to separate the motion cues from the event features. Specifically, we use events from both sides of SCER as motion priors \\((E_0\\) and \\(E_{2N - 1})\\) and apply cross-attention to capture detailed motion cues from the event features." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.481, + 0.907, + 0.501 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {m} = \\operatorname {A t t} \\left(\\operatorname {c o n v} \\left(E _ {0} \\| E _ {2 N - 1}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.573 + ], + "angle": 0, + "content": "Then, we calculate offsets with the motion cues to explicitly estimate variant blur kernels with different scales, and employ deformable convolutions to capture contextual features, enhancing the ability to express blur effectively." + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.586, + 0.905, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\Theta = \\operatorname {c o n v} \\left(F _ {i} ^ {m}\\right) \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.616, + 0.905, + 0.633 + ], + "angle": 0, + "content": "\\[\nf _ {2} \\left(f, F _ {i} ^ {m}\\right) = D C N (f, \\Theta) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.641, + 0.874, + 0.655 + ], + "angle": 0, + "content": "where \\(\\Theta\\) means the offsets of deformable convolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.64, + 0.689 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.807, + 0.714 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Synthetic Dataset. To make a fair comparison with previous methods based on events and images, we selected the GoPro dataset [27], which is widely used for deblurring tasks. We utilized ESIM [31] to generate the event stream, applying the same parameters as other studies, specifically using a Gaussian normal distribution with parameters \\(\\mathcal{N}(\\mu = 0.2,\\sigma = 0.03)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Real-world Datasets. To validate the deblurring capabilities of event data in real-world scenarios, REBlur [35] collects events from Dynamic and Active Pixel Vision Sensor (DAVIS) [2] in real environments and uses an electronically controlled stepping motor to relocate the slide-rail" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14736" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "to the coordinate corresponding to the timestamp t. However, DAVIS-346 camera has low image quality and resolution. Therefore, high resolution and quality event-based real-world deblurring datasets REVD [15] and EVRB [14] are proposed with Triple-axis Hybrid Camera System." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Evaluation Metrics. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) [43] are proposed to assess image quality. PSNR calculates the ratio of signal power to noise, with higher values indicating closer similarity to the reference image, though it mainly reflects pixel-level fidelity. SSIM, on the other hand, measures structural, luminance, and contrast similarities, providing a more perceptually relevant quality assessment, with values closer to 1 indicating better similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.424 + ], + "angle": 0, + "content": "Implementation Details. Our model supports end-to-end training with a batch size of 16. The N of SCER [35] is set to 8. For data augmentation, we randomly crop images to \\(256 \\times 256\\) pixels and apply random horizontal and vertical flips. We chose Adam [17] as the optimizer. Training runs for 400,000 iterations, starting with a learning rate of 2e-4, which is progressively reduced to a minimum of 1e-7 using CosineAnnealingLR [25]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.438, + 0.482, + 0.533 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the GoPro [27] dataset. \\(\\uparrow\\) indicates that higher values correspond to better performance. The numbers in “()” indicate the publication year of the referenced papers. In the second column, “Single” in black means frame-based method, and red represents methods based on single-frame and event while green means multi-frame event-based methods. The best results are indicated in bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.546, + 0.479, + 0.825 + ], + "angle": 0, + "content": "
MethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
Restormer [45] (22')Single32.920.961
NAFNet [4] (22')Single33.690.967
FFTformer [18] (23')Single34.210.969
GRL-B [23] (23')Single33.930.968
MLWNet [10] (24')Single33.830.968
Jianget al. [11] (20')Single31.790.949
REDNet [44] (21')Single35.070.971
EFNet [35] (22')Single35.460.973
SpkNet [5] (23')Single36.120.971
D2Nets [33] (21')3 frames31.760.943
UEVD [12] (22')2 frames35.480.971
REFID [36] (23')2 frames35.910.973
FEVD [15] (24')Multi-frames36.700.978
CMTA-5 [14] (24')5 frames36.550.977
CMTA-7 [14] (24')7 frames36.780.977
OursSingle36.700.977
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.423, + 0.864 + ], + "angle": 0, + "content": "4.2. Compared to State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Synthetic Dataset. We evaluate our proposed network against state-of-the-art motion deblurring methods on the" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.09, + 0.878, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative results on three real-world datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.114, + 0.898, + 0.378 + ], + "angle": 0, + "content": "
DatasetMethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
REBlurSNR+ [37]Single36.870.970
HINet+ [3]Single37.680.973
EFNet [35]Single38.120.975
OursSingle38.530.977
EVRBUEVD [12]2 frames30.550.915
EFNet [35]Single30.980.927
REFID [36]2 frames30.330.918
CMTA [14]Multi-frames31.380.927
OursSingle32.050.936
REVDeSL-Net [41]Single26.990.787
EFNet [35]Single31.750.920
REDNet [44]Single31.900.920
UEVD [12]2 frames31.970.921
FEVD [15]Multi-frames32.990.933
OursSingle32.140.922
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.726 + ], + "angle": 0, + "content": "GoPro [27] dataset. The comparison covers three categories: frame-based image deblurring, event(spike)-based image deblurring and event-based video deblurring. As shown in Table 1, in comparison to SpkNet [5], the existing state-of-the-art spike-based image motion deblurring method, our model achieves an impressive performance boost of \\(+0.58\\) dB. Moreover, compared to CMTA [14], the leading video motion deblurring method, our single-frame model outperforms CMTA-5 (which uses five frames for inference) by \\(+0.15\\) dB and is only \\(-0.08\\) dB behind CMTA-7. Thus, our model demonstrates the remarkable effectiveness of event features in the task of image motion deblurring. We show qualitative results on GoPro in Figure 3, RVRT [24], a video-based motion deblurring method, models long-term temporal dependencies to provide contextual references for deblurring the current frame. MLWNet [10] stands as the leading single-frame motion deblurring approach, while EFNet [35] is an event-based image deblurring method. It's evident that our approach achieves superior detail restoration compared to these previous methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Real-world Datasets. To assess our model's effectiveness in real-world settings, we utilized three datasets: REBlur [35], REVD [15], and EVRB [14]. For REBlur, given its limited dataset size, we first train the model on the GoPro [27] dataset and then finetune it on REBlur training set. For FEVD and EVRB, we retrained the model according to the training and testing set splits provided in the official guidelines. As shown in Table 2, our method surpasses event-based image motion deblurring approaches across all three datasets and even outperforms the current top video-based methods on EVRB [14]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14737" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.128, + 0.09, + 0.845, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.698, + 0.746, + 0.713 + ], + "angle": 0, + "content": "Figure 3. Visual results with our model and other SOTA methods on GoPro dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.245, + 0.756 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Our baseline model \\((B)\\) is a standard two-stage U-Net that does not utilize any event features. As shown in Table 3, it is evident that the integration of event modalities results in significant performance enhancements, as demonstrated by both attention-based event fusion and event-driven deformable convolutions, which yield impressive outcomes. As shown in Figure 4, the baseline performs poorly in deblurring, failing to effectively restore either sharp edges or clear textures. And with the introduction of the event" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.74, + 0.84, + 0.756 + ], + "angle": 0, + "content": "modality, the visual results improve significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As shown in the first part in Table 3, simply applying cross-attention to fuse event and image modalities \\((B + Event^{Att})\\), yields promising results. However, this straightforward fusion approach overlooks the inherent conflict between motion and edge features in event streams, which prevents the model from accurately capturing sharp edge details. As shown in the first example of Figure 4, when image features are fused with events, the text in the yellow box becomes recognizable but still those in blue" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14738" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.882, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.383, + 0.873, + 0.399 + ], + "angle": 0, + "content": "Figure 4. Visual results of ablation study on EVRB [14] dataset. It is recommended to zoom in for a clearer view of the details." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Table 3. Ablation study on EVRB [14] dataset. \\( B \\) represents our baseline, which does not utilize event streams to enhance the image. \\( Event^{Att} \\) denotes applying cross-attention to integrate image with event features. \\( Edge^{Att} \\) means our Edge-aware Sharpening module. \\( Img^{DCN} \\) and \\( Event^{DCN} \\) represent utilizing image features and event features to calculate offsets for the deformable convolution respectively. \\( Motion^{DCN} \\) is our Motion-driven Scale-adaptive Deblurring module." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.544, + 0.472, + 0.701 + ], + "angle": 0, + "content": "
MethodMetrics
PSNRSSIM
B28.650.900
B+EventAtt30.710.923
B+EdgeAtt31.130.929
B+ImgDCN28.800.894
B+EventDCN30.510.921
B+MotionDCN31.630.930
Ours(B+EdgeAtt + MotionDCN)32.050.936
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.793 + ], + "angle": 0, + "content": "box retain significant blurriness. Instead, our Edge-aware Sharpening module \\((B + Edge^{Att})\\) strengthens the representation of edge features within the events, producing a more refined result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The second part of the Table 3 demonstrates the value of motion cues in learning diverse blur kernels. Initially, we utilize image features to determine offsets for deformable convolutions \\((B + Img^{DCN})\\); however, because image lacks the motion cues, they are unable to accurately capture the diverse blur kernels that correspond to different motion characteristics. As shown in the second example in Fig-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.59 + ], + "angle": 0, + "content": "ure 4, using image features to estimate the blur kernel still results in artifacts around the \"No Parking Sign\". Conversely, training deformable convolutions with event features \\((B + Event^{DCN})\\) leads to a better performance, but it still results in problematic artifacts in some local regions (in the yellow box). To further improve the capture of motion information, our proposed Motion-driven Scale-adaptive Deblurring module \\((B + Motion^{DCN})\\) accurately captures object motion cues and estimates non-uniform and varying blur kernels, significantly minimizing artifacts both globally and locally." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Finally, by effectively fusing both motion and edge cues, our model demonstrates substantial improvements in the restoration of both sharp edges and fine texture details." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we observe that event features inherently contain conflicting edge and motion cues: edge features accumulate over short exposure times lack sufficient motion information, while motion cues accumulate over long exposure times compromise edge clarity. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to separately leverage these two aspects of cues. By incorporating edge and motion priors, we enhance the corresponding cues in event features: edge cues restore sharp object edges, while motion cues guide deformable convolutions to adaptively estimate blur kernels, adjusting the receptive field for diverse motion patterns. Extensive experiments on synthetic and real-world datasets demonstrate the superior performance of our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14739" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.253 + ], + "angle": 0, + "content": "The work is jointly supported by the National Natural Science Foundation of China (NSFC) under Grant 62261160576 and Grant 62203024, the Beijing Natural Science Foundation (4252026), the Research and Development Program of Beijing Municipal Education Commission (KM202310005027), and the Fundamental Research Funds for the Central Universities of China. This research work is also supported by the Big Data Computing Center of Southeast University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.267, + 0.188, + 0.282 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.291, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[1] Yuval Bahat, Netelee Efrat, and Michal Irani. Non-uniform blind deblurring by reblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3286-3294, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.349, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.406, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[5] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems, 36:70279-70292, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[6] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[7] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[8] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In ACM SIG International Conference on Computer Graphics and Interactive Techniques, pages 787-794. 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Xin Gao, Tianheng Qiu, Xinyu Zhang, Hanlin Bai, Kang Liu, Xuan Huang, Hu Wei, Guoying Zhang, and Huaping Liu." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.291, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2733-2742, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[11] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cross-modal temporal alignment for event-guided video deblurring. European Conference on Computer Vision, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.561 + ], + "angle": 0, + "content": "[17] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[18] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.908, + 0.717 + ], + "angle": 0, + "content": "[19] Jan Kotera, Filip Šroubek, and Peyman Milanfar. Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In Computer Analysis of Images and Patterns: 15th International Conference, CAIP 2013, York, UK, August 27-29, 2013, Proceedings, Part II 15, pages 59-66. Springer, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[20] Dilip Krishnan, Terence Tay, and Rob Fergus. Blind deconvolution using a normalized sparsity measure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 233-240. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[21] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1971. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Jichun Li, Weimin Tan, and Bo Yan. Perceptual variousness motion deblurring with light global context refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4116-4125, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14740" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[26] Xintian Mao, Qingli Li, and Yan Wang. Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25681-25690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[27] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[28] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[30] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurringli2021 perceptual. In Association for the Advancement of Artificial Intelligence, pages 11882-11889, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. pages 969-982. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[33] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606–3615, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[38] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[39] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[40] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[41] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In European Conference on Computer Vision, pages 155-171. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[42] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[44] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[47] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[48] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10734-10744, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14742" + } + ] +] \ No newline at end of file diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_origin.pdf b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a20c98b5efae77a6c5e1f43f1a88c1c842dc019e --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/9f65b86f-9c52-4842-ae57-85222dbb0f4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb77caffd4cdf3761250682892db4a582092352038eff839baa312b6b1fa2a1d +size 7020064 diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/full.md b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93dd88961eacde27ca7a3b8670a76c14799b2e9c --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/full.md @@ -0,0 +1,301 @@ +# Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring + +Yufei Zhu $^{1,2}$ , Hao Chen $^{*1,2}$ , Yongjian Deng $^{3}$ , and Wei You $^{4}$ + +$^{1}$ School of Computer Science and Engineering, Southeast University, Nanjing, China + $^{2}$ Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China. + +$^{3}$ College of Computer Science, Beijing University of Technology, Beijing, China + +$^{4}$ Advanced Computing and Storage Lab, Huawei Technologies, Beijing, China + +{220232213，haochen303}@seu.edu.cn，yjdeng@bjut.edu.cn，youwei22@huawei.com + +# Abstract + +Traditional motion deblurring methods struggle to effectively model motion information within the exposure time. Recently, event cameras have attracted significant research interest for its ability to model motion cues over the exposure duration. However, these methods directly fuse event features with image, overlooking the intrinsic heterogeneity of events. In this paper, we identify that the event modality contains two conflicting types of information: edge features and motion cues. Events accumulated over a short exposure period capture sharp edge details but lose motion information, while those accumulated over a long exposure period blur edge details due to motion. To address this issue, we propose a simple yet effective approach to disentangle these two cues from event features and employ an edge-aware sharpening module along with motion-driven scale-adaptive deblurring module to fully leverage both. Specifically, the first module aids in restoring sharp edges by leveraging the clear edge features provided by events, while the second module leverages motion cues to learn diverse blur kernels, adaptively adjusting the receptive field for optimal deblurring. Extensive experiments on synthetic and real-world datasets validate the effectiveness of our approach and yield a substantial improvement over state-of-the-art single-frame methods and surpasses most multi-frame-based methods. Code will be publicly available. + +# 1. Introduction + +Motion blur is often caused by camera shake or the movement of objects. The task of motion deblurring [1, 8, 19-21] + +![](images/537ef25ec2eeb135a1513e1918b7d7a517e1184f6acb2db4e786ec498a4dd313.jpg) +(a) + +![](images/8d6ce0b261fb260abaeed3c13e82b96ed1d1e634fc74184112f47dc1d19701bf.jpg) + +![](images/01edaff7caadb0444c9efbe55b4360bbd9c88c5a9c849a6042c84ea33bf1a23d.jpg) + +![](images/51b98486babf9b2e163501a268bcd4e4761a412c5ce47b26c0a32c20b391b277.jpg) +(b) +(c) + +![](images/b0176bf9fbee85f14b9d91bb0407d5dedf163889b5e642216958739b6c5ca652.jpg) + +![](images/e9f549a2886fcb970dd8f9d58d1efebf670624d83f0d4f1e94184f3a3fcee2d0.jpg) + +![](images/94d73839f2a813d7498e818a336bd249e0f01371f9acfbe41df402ae1b8b4e26.jpg) + +![](images/da57136c57da64e49f60922a322a191a5afb3791bab5ba6904feaf328901fc0f.jpg) +Figure 1. (a) Visualization of SCER(Symmetric Cumulative Event Representation) [35]. (b) Previous fusion methods between events and image. (c) Ours. Separate events into edge features and motion cues. + +![](images/3c0fa5a681be0fb37001a6bd936300fc6f08cfd83b8ebac0084ecce4bd7970aa.jpg) + +aims to recover sharp images from blurred ones by restoring clear edges and capturing fine details, ultimately enhancing the image quality. However, since image features lack the ability to model motion cues over the exposure time, they cannot accurately estimate the variant blur ker + +nels present in real-world scenarios. This limitation often results in subpar performance of current methods on real-world datasets. Recently, event cameras [2, 9], bio-inspired sensors that asynchronously capture changes in pixel intensities, have been leveraged to aid motion deblurring thanks to their microsecond-level latency and low power consumption. + +Several studies [6, 11, 35, 47, 48] aim to integrate image and event features to enhance deblurring performance. For example, Sun et al. [35] introduced EFNet, which considers the inherent differences between event and image modalities and utilizes a cross-attention mechanism, improving upon previous methods that relied on simple concatenation or addition for feature fusion. Kim et al. [12] discovered that event streams contain both clear edge and motion cues and proposed a novel attention-based fusion method, allowing them to leverage these event characteristics to extract discriminative channel and spatial information from the blurred image features. Their approaches directly combine the heterogeneous cues contained in events (edge and motion) with RGB features indiscriminately. However, we find that sharp edge features and detailed motion cues are inherently conflicting in their representation and previous one-step fusion approaches [12, 35, 36] tend to bias the event modality toward capturing either edge or motion information, leading to suboptimal results. Specifically, as shown in Figure 1 (a), it is obvious that events provide sharp edge features within short exposure times(the $2^{nd}$ column), while accumulating the motion trajectory of objects over the entire exposure period(the $3^{rd}$ column). Therefore, sharp edges typically lack motion information, whereas representations rich in motion cues often compromise edge clarity. As a result, previous one-step fusion methods make it challenging for the model to accurately learn both edge and motion features from events, leading to coarse edges and blurred textures(Figure 1 (b)). Our insight is that edge and motion should be modeled separately and then integrated with RGB to respect both perspectives and avoid conflicts. To restore sharp edges and clear textures, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to independently harness the sharp edge features and rich motion cues provided by the event stream (Figure 1 (c)). Contributing to much better leverage of edge and motion cues, our model shows significant improvements in deblurring on both edges and textures compared to previous methods. + +A straightforward approach to restore sharp edges is to utilize event features accumulated during a short exposure (the second column in Figure 1 (a)) as edge information and fuse them with image features, but sacrificing the complete event features captured over the full exposure duration. Besides, relying on a fixed time interval to capture sharp edge features is unreliable, as varying motion speeds + +result in different event trigger frequencies: excessive event accumulation can obscure edge details, while insufficient events may fail to capture edges fully. Therefore, our Edge-aware Sharpening module combines both short-exposure event features with complete event cues, using the former as an edge prior to extract complete edge cues by performing cross-attention with the latter. + +As for motion cues, previous work [11, 14-16, 44] has typically utilized events as a substitute for optical flow to align frames, achieving significant advancements in video motion deblurring. Some works [11, 33] have used event to predict optical flow for frame-to-frame alignment. FEVD [15] and CMTA [14] adopt the concept of "synthesis-based alignment" [13, 39, 40], directly using RGB images and events to achieve frame-to-frame alignment without relying on intermediate steps like optical flow prediction. However, these approaches overlook the crucial role that motion cues play in motion deblurring: different motion patterns yield diverse blur kernels. By leveraging motion cues, we can explicitly estimate blur kernels for various motion patterns, enhancing the model's ability to adapt to a wide range of blurring scenarios. To effectively learn blur kernels of varying scales, previous studies [34, 37, 46] have attempted to expand the receptive field by adding more convolutional layers or using larger convolutional kernel sizes. However, traditional convolutional kernels have a fixed size, which limits their robustness when handling diverse blur kernels, reducing the model's adaptability in real-world scenarios. In our work, inspired by [22, 38, 42], we propose a Motion-driven Scale-adaptive Deblurring module that leverages the rich motion cues provided by the event stream to explicitly estimate different blur kernels, enabling adaptive adjustments to the receptive field to better accommodate a variety of blur patterns. By leveraging motion cues effectively, our single-frame motion deblurring model achieves performance that surpasses most video-based approaches. + +In summary, our contributions are as follows: + +(1) We find that the edge and motion cues provided by the event stream exhibit a conflict in their representation which is unexplored in the event community. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to effectively utilize these two distinct features. +(2) We find that motion cues are instrumental in learning variant blur kernels. Building on this insight, we utilize motion cues to achieve adaptive adjustments to the receptive field to accommodate variant blur kernels. +(3) Conducting extensive experiments on both synthetic and real-world datasets, our method demonstrates superior performance, significantly surpassing existing single-frame-based approaches and even most multi-frame-based methods. + +# 2. Related Work + +# 2.1. Frame-based Motion Deblurring on Image + +With the advancement of deep learning, convolutional neural networks (CNNs) have been increasingly applied to the motion deblurring. To better handle non-uniform and varying blur kernels, researchers have primarily explored three technical approaches: multi-scale architecture designs [7, 10, 18, 34, 37, 46], more sophisticated decoders [26, 29], and novel network layer designs [22, 30]. For example, SRN(Scale-Recurrent Network) [37] achieves deblurring of high-resolution images by guiding the process with low-resolution images through a multi-scale input-output approach. However, traditional convolutional kernels have spatial uniformity and a fixed receptive field, which makes them suboptimal for modeling dynamic scene deblurring. Although Li et al. [22] introduced deformable convolution to fit blur kernels of varying sizes, their approach still faces challenges in fitting non-uniform and varying blur kernels due to limited motion cues. In our work, we incorporate the event modality, leveraging its high temporal resolution and utilize deformable convolution to address this issue effectively. + +# 2.2. Event-based Motion Deblurring + +Early work, based on the imaging principles of event cameras and the process of how blurred images are generated, proposed the Event Double Integral (EDI) model [28, 47, 48]. However, due to the uncertain event-triggering thresholds and abundant noise, these methods have not yielded satisfactory results. + +Other approaches [6, 12, 35, 36] focus on fusing event and image features, training a deblurring network in an end-to-end fashion for improved performance. For example, EFNet [35] utilizes cross-attention mechanism to fuse features from both modalities. REFID [36] employs bidirectional event recurrent (EVR) blocks to extract features from the event streams and introduces an Event-Guided Adaptive Channel Attention module to uncover discriminative channels within the event features. However, these approaches integrate event and image features in a single step, failing to fully account for the edge and motion cues inherent in the event features. In our research, we highlight that event features contain two inherently conflicting aspects: distinct edge information and abundant motion cues. Specifically, distinct edge features help restore the object's sharp boundaries, while abundant motion cues assist the model in learning diverse blur kernels, leading to better recovery of the object's texture details. To leverage each optimally, we introduce an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module. + +To utilize the rich motion cues offered by events, several studies [11, 14, 15, 33] focus on employing the event + +modality to handle video motion deblurring tasks. For example, FEVD [15] and CMTA [14] adopt the concept of "synthesis-based alignment", utilizing bidirectional recurrent networks and cascade networks respectively, to achieve frame-to-frame alignment. However, these works use events for frame-to-frame alignment, overlooking the role of motion cues in learning variant blur kernels. In this paper, our Motion-Driven Scale-Adaptive Deblurring module effectively leverages motion cues to guide the adjustment of the model's receptive field, enabling it to adapt to diverse blur kernels. This approach ensures robust handling of various motion patterns and achieves competitive performance, even compared to state-of-the-art video-based methods. + +# 3. Method + +Firstly, we introduce the main architecture of our model in Sec. 3.1. In Sec. 3.2, an Edge-aware Sharpening module is proposed to utilize sharp edges captured during short exposure times from event to restore the clear boundaries of deblurring objects. In Sec. 3.3, to adaptively learn variant blur kernels, in the Motion-driven Scale-adaptive Deblurring module, we propose using motion cues to explicitly estimate the blur kernels associated with different motion patterns, enabling the model to adaptively adjust its receptive field for improved deblurring performance. + +# 3.1. Model Architecture + +Event-based image motion deblurring aims to recover sharp images from a blurred image $\mathbf{B}$ and an event stream $\mathcal{E}$ . We use SCER [35] to represent the event $(\mathcal{E} = \{E_i\}_{i=0}^{2N-1})$ . As shown in Figure 2, based on EFNet [35], our model consists of two U-Net [32] networks. During the first stage, an event encoding branch extracts multi-scale features from the event stream $\mathcal{E}$ . + +$$ +\{F _ {i} ^ {E} \} _ {i = 0} ^ {3} = R e s (\mathcal {E}) \tag {1} +$$ + +Event features encompass both object edge information and motion cues. Previous approaches [12, 14, 35] often fused these event features directly, which can cause edge information and motion cues to conflict, leading to a model that neither fully utilizes the events to sharpen object edges nor effectively exploits the motion cues for deblurring. Specifically, event stream contains clear edge features from short exposure durations and motion cues accumulated over longer exposure times. This one-step fusion approach can easily cause the model to neglect one aspect of the event information. Therefore, an Edge-aware Sharpening module and a Motion-driven Scale-adaptive Deblurring module are proposed to fully leverage the features contained in the event stream respectively. + +$$ +\{F _ {i} ^ {I} \} _ {i = 1} ^ {3} = f _ {2} (f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}), F _ {i} ^ {m}) \qquad (2) +$$ + +![](images/f4262f2a4b8b4f113522ba28a561d03b020def9f764ef3f2a43ed9094a7dfa96.jpg) +Figure 2. The main architecture of our model. Based on EFNet [35], we propose a two-stage U-Net architecture to achieve coarse-to-fine motion deblurring. During the first stage, we separate event into edge features and motion cues to guide the model: edge features enhance the recovery of sharp image edges with an Edge-aware Sharpening module, while motion cues help the model learn diverse blur kernels through Motion-driven Scale-adaptive Deblurring module. In the second stage, similar to EFNet, the output from the first stage is concatenated with the original blurred image. An EMGC (Event Mask Gated Connection) [35] module is then applied to enhance features specifically in the regions events triggered. + +where $f_{1}$ means Edge-aware Sharpening module and $f_{2}$ means Motion-driven Scale-adaptive Deblurring module. $F_{i}^{e}$ and $F_{i}^{m}$ represent edge features and motion cues from the event stream $\mathcal{E}$ respectively. Next, a standard U-Net decoder layer with skip connections is employed to predict the deblurred image $I^{initial}$ . + +$$ +\left\{D _ {i} ^ {I} \right\} _ {i = 1} ^ {3} = R e s \left(D _ {i - 1} ^ {I}, F _ {i} ^ {I}\right) \tag {3} +$$ + +$$ +I ^ {i n i t i a l} = \operatorname {c o n v} _ {3 \times 3} \left(D _ {3} ^ {I}\right) \tag {4} +$$ + +$$ +L _ {1} = P S N R L o s s (I ^ {\text {i n i t i a l}}, I) \tag {5} +$$ + +Considering that event-triggered areas often suffer from more severe blur, in the second stage, the deblurred image $I^{initial}$ is concatenated with the blurred image B and then processed through another U-Net network to focus on restoring the event-triggered areas. EMGC(Event Mask Gated Connection) [35] module generates masks from events to create skip connections from the first stage to the + +encoder of the second stage, guiding the model to concentrate on event-triggered areas. The encoder layer can be represented as: + +$$ +\begin{array}{l} E M G C \left(F _ {i} ^ {I}, D _ {i} ^ {I}\right) = c o n v \left(F _ {i} ^ {I}\right) + c o n v \left(F _ {i} ^ {I} \times (1 - \mathcal {M})\right) \\ + \operatorname {c o n v} \left(D _ {i} ^ {I}\right) + \operatorname {c o n v} \left(D _ {i} ^ {I} \times \mathcal {M}\right) \tag {6} \\ \end{array} +$$ + +$$ +\left\{\hat {F} _ {i} ^ {I} \right\} _ {i = 1} ^ {3} = R e s \left(\hat {F} _ {i - 1} ^ {I}\right) + E M G C \left(F _ {i} ^ {I}, D _ {i} ^ {I}\right) \tag {7} +$$ + +where $\mathcal{M}$ represents the mask generated from event stream $\mathcal{E}$ . Pixels where events triggered are set to 1, while others are set to 0. As a result, feature maps from encoder overlook the event-triggered areas, which may contain severe blur, while feature maps from the decoder enhance these regions, reflecting the restored blurred content. Finally, the final deblurred result $I^{final}$ is obtained after processing through a decoder identical to the one used in the first stage. Loss function is expressed as follows: + +$$ +L _ {2} = P S N R L o s s \left(I ^ {f i n a l}, I\right) \tag {8} +$$ + +$$ +\mathcal {L} = \omega \times L _ {1} + (1 - \omega) \times L _ {2} \tag {9} +$$ + +# 3.2. Edge-aware Sharpening Module + +As shown in the second column in Figure 1 (a), the accumulation of events over a short period often provides clear edge features of objects, which help restore sharp boundaries for blurred objects. Consequently, FEVD [15] employs fast Fourier transform to simultaneously capture the high-frequency signals, which typically contain detailed edge cues, from both events and images. However, event features often contain abundant motion cues, which can blur the sharp edge features and make it difficult to extract clear edge information from events. On the other hand, even objects undergoing rapid motion can accumulate clear edge features within short exposure times. This characteristic is particularly valuable for preserving fine details, as edges play a crucial role in maintaining the integrity of object boundaries and structures in the image. However, when combined with motion cues, which are more prominent in longer exposure times, these edge features can become diluted or lost. Therefore, it is essential to selectively enhance the edge information from the event stream to ensure that sharp boundaries are preserved, while still utilizing the motion information to handle blur effectively. + +To address this issue, we propose an Edge-aware Sharpening module that firstly combines the event features accumulated over a short period with the image to enhance the edges of moving objects. These enhanced features serve as queries, allowing the attention mechanism to effectively integrate the edge features from the event stream. Specifically, SCER [35] uses the midpoint as a reference and accumulates $2 \times N$ frames of events symmetrically from this point. We select the channels closest to the midpoint as edge priors to enhance the image features. Thus, our Edge-aware Sharpening module can be represented as: + +$$ +F _ {i} ^ {e} = \operatorname {c o n v} \left(E _ {N - 1} \| E _ {N}\right) \tag {10} +$$ + +$$ +f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}) = +$$ + +$$ +A t t \left(\operatorname {c o n v} \left(R e s \left(F _ {i - 1} ^ {I}\right) \| F _ {i} ^ {e}\right), F _ {i} ^ {E}, F _ {i} ^ {E}\right) \tag {11} +$$ + +where $\parallel$ means connection between features and $Att(x,y,z)$ denotes an attention mechanism where $x$ serves as the query, $y$ as the key, and $z$ as the value. + +# 3.3. Motion-driven Scale-adaptive Deblurring Module + +Previous approaches have often concentrated on utilizing event streams for frame-to-frame alignment to achieve video motion deblurring. However, these methods frequently overlook the importance of motion features in the context of static image deblurring. The motion cues of objects during exposure time help us explicitly learn variant blur kernels. + +As shown in the third column in Figure 1 (a), the girl exhibits a larger blur kernel, causing the edges accumulated + +in the event stream to appear broader and more diffused. This effect is in stark contrast to the background crowd, where the blur kernel is smaller due to less motion, resulting in sharper and more defined edges in the event accumulation. These varying blur kernels indicate differences in the degree of motion across the scene, which are essential for accurately restoring fine textures in the deblurring process. To effectively handle such diverse blur kernels, we introduce a Motion-driven Scale-adaptive Deblurring module that adaptively adjusts the convolutional receptive field based on the degree of motion captured in the event stream. This is achieved through the use of deformable convolutions, which enable the network to learn dynamic and spatially flexible convolutional kernels. This approach allows the model to handle both large, motion-induced blur (such as that of the girl) and smaller, subtle blur (such as that of the background crowd) simultaneously, improving the overall deblurring performance. + +As mentioned earlier, the event stream contains both edge and motion features. To highlight the motion cues of objects, it is essential to separate the motion cues from the event features. Specifically, we use events from both sides of SCER as motion priors $(E_0$ and $E_{2N - 1})$ and apply cross-attention to capture detailed motion cues from the event features. + +$$ +F _ {i} ^ {m} = \operatorname {A t t} \left(\operatorname {c o n v} \left(E _ {0} \| E _ {2 N - 1}\right), F _ {i} ^ {E}, F _ {i} ^ {E}\right) \tag {12} +$$ + +Then, we calculate offsets with the motion cues to explicitly estimate variant blur kernels with different scales, and employ deformable convolutions to capture contextual features, enhancing the ability to express blur effectively. + +$$ +\Theta = \operatorname {c o n v} \left(F _ {i} ^ {m}\right) \tag {13} +$$ + +$$ +f _ {2} \left(f, F _ {i} ^ {m}\right) = D C N (f, \Theta) \tag {14} +$$ + +where $\Theta$ means the offsets of deformable convolution. + +# 4. Experiment + +# 4.1. Datasets and Experimental Setup + +Synthetic Dataset. To make a fair comparison with previous methods based on events and images, we selected the GoPro dataset [27], which is widely used for deblurring tasks. We utilized ESIM [31] to generate the event stream, applying the same parameters as other studies, specifically using a Gaussian normal distribution with parameters $\mathcal{N}(\mu = 0.2,\sigma = 0.03)$ . + +Real-world Datasets. To validate the deblurring capabilities of event data in real-world scenarios, REBlur [35] collects events from Dynamic and Active Pixel Vision Sensor (DAVIS) [2] in real environments and uses an electronically controlled stepping motor to relocate the slide-rail + +to the coordinate corresponding to the timestamp t. However, DAVIS-346 camera has low image quality and resolution. Therefore, high resolution and quality event-based real-world deblurring datasets REVD [15] and EVRB [14] are proposed with Triple-axis Hybrid Camera System. + +Evaluation Metrics. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) [43] are proposed to assess image quality. PSNR calculates the ratio of signal power to noise, with higher values indicating closer similarity to the reference image, though it mainly reflects pixel-level fidelity. SSIM, on the other hand, measures structural, luminance, and contrast similarities, providing a more perceptually relevant quality assessment, with values closer to 1 indicating better similarity. + +Implementation Details. Our model supports end-to-end training with a batch size of 16. The N of SCER [35] is set to 8. For data augmentation, we randomly crop images to $256 \times 256$ pixels and apply random horizontal and vertical flips. We chose Adam [17] as the optimizer. Training runs for 400,000 iterations, starting with a learning rate of 2e-4, which is progressively reduced to a minimum of 1e-7 using CosineAnnealingLR [25]. + +Table 1. Quantitative results on the GoPro [27] dataset. $\uparrow$ indicates that higher values correspond to better performance. The numbers in “()” indicate the publication year of the referenced papers. In the second column, “Single” in black means frame-based method, and red represents methods based on single-frame and event while green means multi-frame event-based methods. The best results are indicated in bold. + +
MethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
Restormer [45] (22')Single32.920.961
NAFNet [4] (22')Single33.690.967
FFTformer [18] (23')Single34.210.969
GRL-B [23] (23')Single33.930.968
MLWNet [10] (24')Single33.830.968
Jianget al. [11] (20')Single31.790.949
REDNet [44] (21')Single35.070.971
EFNet [35] (22')Single35.460.973
SpkNet [5] (23')Single36.120.971
D2Nets [33] (21')3 frames31.760.943
UEVD [12] (22')2 frames35.480.971
REFID [36] (23')2 frames35.910.973
FEVD [15] (24')Multi-frames36.700.978
CMTA-5 [14] (24')5 frames36.550.977
CMTA-7 [14] (24')7 frames36.780.977
OursSingle36.700.977
+ +# 4.2. Compared to State-of-the-art Methods + +Synthetic Dataset. We evaluate our proposed network against state-of-the-art motion deblurring methods on the + +Table 2. Quantitative results on three real-world datasets. + +
DatasetMethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
REBlurSNR+ [37]Single36.870.970
HINet+ [3]Single37.680.973
EFNet [35]Single38.120.975
OursSingle38.530.977
EVRBUEVD [12]2 frames30.550.915
EFNet [35]Single30.980.927
REFID [36]2 frames30.330.918
CMTA [14]Multi-frames31.380.927
OursSingle32.050.936
REVDeSL-Net [41]Single26.990.787
EFNet [35]Single31.750.920
REDNet [44]Single31.900.920
UEVD [12]2 frames31.970.921
FEVD [15]Multi-frames32.990.933
OursSingle32.140.922
+ +GoPro [27] dataset. The comparison covers three categories: frame-based image deblurring, event(spike)-based image deblurring and event-based video deblurring. As shown in Table 1, in comparison to SpkNet [5], the existing state-of-the-art spike-based image motion deblurring method, our model achieves an impressive performance boost of $+0.58$ dB. Moreover, compared to CMTA [14], the leading video motion deblurring method, our single-frame model outperforms CMTA-5 (which uses five frames for inference) by $+0.15$ dB and is only $-0.08$ dB behind CMTA-7. Thus, our model demonstrates the remarkable effectiveness of event features in the task of image motion deblurring. We show qualitative results on GoPro in Figure 3, RVRT [24], a video-based motion deblurring method, models long-term temporal dependencies to provide contextual references for deblurring the current frame. MLWNet [10] stands as the leading single-frame motion deblurring approach, while EFNet [35] is an event-based image deblurring method. It's evident that our approach achieves superior detail restoration compared to these previous methods. + +Real-world Datasets. To assess our model's effectiveness in real-world settings, we utilized three datasets: REBlur [35], REVD [15], and EVRB [14]. For REBlur, given its limited dataset size, we first train the model on the GoPro [27] dataset and then finetune it on REBlur training set. For FEVD and EVRB, we retrained the model according to the training and testing set splits provided in the official guidelines. As shown in Table 2, our method surpasses event-based image motion deblurring approaches across all three datasets and even outperforms the current top video-based methods on EVRB [14]. + +![](images/9798852e165c9a5c59a105ed6da9051949ce718f0e7f77f36d7842b7f645dc59.jpg) +Figure 3. Visual results with our model and other SOTA methods on GoPro dataset. + +# 4.3. Ablation Study + +Our baseline model $(B)$ is a standard two-stage U-Net that does not utilize any event features. As shown in Table 3, it is evident that the integration of event modalities results in significant performance enhancements, as demonstrated by both attention-based event fusion and event-driven deformable convolutions, which yield impressive outcomes. As shown in Figure 4, the baseline performs poorly in deblurring, failing to effectively restore either sharp edges or clear textures. And with the introduction of the event + +modality, the visual results improve significantly. + +As shown in the first part in Table 3, simply applying cross-attention to fuse event and image modalities $(B + Event^{Att})$ , yields promising results. However, this straightforward fusion approach overlooks the inherent conflict between motion and edge features in event streams, which prevents the model from accurately capturing sharp edge details. As shown in the first example of Figure 4, when image features are fused with events, the text in the yellow box becomes recognizable but still those in blue + +![](images/25730c6f6a6328d98c9c5ac4f2c1a08f32b20dac26e4efb9c46d28ac8dc3b1d1.jpg) +Figure 4. Visual results of ablation study on EVRB [14] dataset. It is recommended to zoom in for a clearer view of the details. + +Table 3. Ablation study on EVRB [14] dataset. $B$ represents our baseline, which does not utilize event streams to enhance the image. $Event^{Att}$ denotes applying cross-attention to integrate image with event features. $Edge^{Att}$ means our Edge-aware Sharpening module. $Img^{DCN}$ and $Event^{DCN}$ represent utilizing image features and event features to calculate offsets for the deformable convolution respectively. $Motion^{DCN}$ is our Motion-driven Scale-adaptive Deblurring module. + +
MethodMetrics
PSNRSSIM
B28.650.900
B+EventAtt30.710.923
B+EdgeAtt31.130.929
B+ImgDCN28.800.894
B+EventDCN30.510.921
B+MotionDCN31.630.930
Ours(B+EdgeAtt + MotionDCN)32.050.936
+ +box retain significant blurriness. Instead, our Edge-aware Sharpening module $(B + Edge^{Att})$ strengthens the representation of edge features within the events, producing a more refined result. + +The second part of the Table 3 demonstrates the value of motion cues in learning diverse blur kernels. Initially, we utilize image features to determine offsets for deformable convolutions $(B + Img^{DCN})$ ; however, because image lacks the motion cues, they are unable to accurately capture the diverse blur kernels that correspond to different motion characteristics. As shown in the second example in Fig- + +ure 4, using image features to estimate the blur kernel still results in artifacts around the "No Parking Sign". Conversely, training deformable convolutions with event features $(B + Event^{DCN})$ leads to a better performance, but it still results in problematic artifacts in some local regions (in the yellow box). To further improve the capture of motion information, our proposed Motion-driven Scale-adaptive Deblurring module $(B + Motion^{DCN})$ accurately captures object motion cues and estimates non-uniform and varying blur kernels, significantly minimizing artifacts both globally and locally. + +Finally, by effectively fusing both motion and edge cues, our model demonstrates substantial improvements in the restoration of both sharp edges and fine texture details. + +# 5. Conclusion + +In this paper, we observe that event features inherently contain conflicting edge and motion cues: edge features accumulate over short exposure times lack sufficient motion information, while motion cues accumulate over long exposure times compromise edge clarity. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to separately leverage these two aspects of cues. By incorporating edge and motion priors, we enhance the corresponding cues in event features: edge cues restore sharp object edges, while motion cues guide deformable convolutions to adaptively estimate blur kernels, adjusting the receptive field for diverse motion patterns. Extensive experiments on synthetic and real-world datasets demonstrate the superior performance of our approach. + +# 6. Acknowledgement + +The work is jointly supported by the National Natural Science Foundation of China (NSFC) under Grant 62261160576 and Grant 62203024, the Beijing Natural Science Foundation (4252026), the Research and Development Program of Beijing Municipal Education Commission (KM202310005027), and the Fundamental Research Funds for the Central Universities of China. This research work is also supported by the Big Data Computing Center of Southeast University. + +# References + +[1] Yuval Bahat, Netelee Efrat, and Michal Irani. Non-uniform blind deblurring by reblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3286-3294, 2017. 1 +[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 5 +[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6 +[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 6 +[5] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems, 36:70279-70292, 2023. 6 +[6] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2, 3 +[7] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 3 +[8] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In ACM SIG International Conference on Computer Graphics and Interactive Techniques, pages 787-794. 2006. 1 +[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 2 +[10] Xin Gao, Tianheng Qiu, Xinyu Zhang, Hanlin Bai, Kang Liu, Xuan Huang, Hu Wei, Guoying Zhang, and Huaping Liu. + +Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2733-2742, 2024. 3, 6 +[11] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2, 3, 6 +[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 3, 6 +[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2 +[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cross-modal temporal alignment for event-guided video deblurring. European Conference on Computer Vision, 2024. 2, 3, 6, 8 +[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2, 3, 5, 6 +[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 2 +[17] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 6 +[19] Jan Kotera, Filip Šroubek, and Peyman Milanfar. Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In Computer Analysis of Images and Patterns: 15th International Conference, CAIP 2013, York, UK, August 27-29, 2013, Proceedings, Part II 15, pages 59-66. Springer, 2013. 1 +[20] Dilip Krishnan, Terence Tay, and Rob Fergus. Blind deconvolution using a normalized sparsity measure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 233-240. IEEE, 2011. +[21] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1971. IEEE, 2009. 1 +[22] Jichun Li, Weimin Tan, and Bo Yan. Perceptual variousness motion deblurring with light global context refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4116-4125, 2021. 2, 3 + +[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 6 +[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6 +[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[26] Xintian Mao, Qingli Li, and Yan Wang. Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25681-25690, 2024. 3 +[27] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5, 6 +[28] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3 +[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 3 +[30] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurringli2021 perceptual. In Association for the Advancement of Artificial Intelligence, pages 11882-11889, 2020. 3 +[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. pages 969-982. PMLR, 2018. 5 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241. Springer, 2015. 3 +[33] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2021. 2, 3, 6 +[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606–3615, 2020. 2, 3 +[35] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van + +Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 1, 2, 3, 4, 5, 6 +[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 2, 3, 6 +[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 3, 6 +[38] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 2 +[39] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 2 +[40] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2 +[41] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In European Conference on Computer Vision, pages 155-171. Springer, 2020. 6 +[42] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 0–0, 2019. 2 +[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6 +[44] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2, 6 +[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6 +[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 3 +[47] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2, 3 +[48] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10734-10744, 2023. 2, 3 \ No newline at end of file diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/images.zip b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f45ff028da40a13c3485c4f0884a79fa969a8bec --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:978ab175be072d427de39c9d9023573fc435b7a7bf43423be31b90503168fef6 +size 830513 diff --git a/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/layout.json b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..10ae20839a181233042f7c71bb3eb8e566663f4e --- /dev/null +++ b/2025/Separation for Better Integration_ Disentangling Edge and Motion in Event-based Deblurring/layout.json @@ -0,0 +1,7270 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 518, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 518, + 140 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 518, + 140 + ], + "type": "text", + "content": "Separation for Better Integration: Disentangling Edge and Motion in Event-based Deblurring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "spans": [ + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "text", + "content": "Yufei Zhu" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "text", + "content": ", Hao Chen" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "inline_equation", + "content": "^{*1,2}" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "text", + "content": ", Yongjian Deng" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "text", + "content": ", and Wei You" + }, + { + "bbox": [ + 162, + 170, + 448, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "spans": [ + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Southeast University, Nanjing, China \n" + }, + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 194, + 534, + 237 + ], + "type": "text", + "content": "Key Laboratory of New Generation Artificial Intelligence Technology and Its Interdisciplinary Applications (Southeast University), Ministry of Education, China." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 237, + 501, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 237, + 501, + 251 + ], + "spans": [ + { + "bbox": [ + 110, + 237, + 501, + 251 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 237, + 501, + 251 + ], + "type": "text", + "content": "College of Computer Science, Beijing University of Technology, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 95, + 255, + 475, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 255, + 475, + 270 + ], + "spans": [ + { + "bbox": [ + 95, + 255, + 475, + 270 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 255, + 475, + 270 + ], + "type": "text", + "content": "Advanced Computing and Storage Lab, Huawei Technologies, Beijing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 269, + 507, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 269, + 507, + 281 + ], + "spans": [ + { + "bbox": [ + 102, + 269, + 507, + 281 + ], + "type": "text", + "content": "{220232213，haochen303}@seu.edu.cn，yjdeng@bjut.edu.cn，youwei22@huawei.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 316, + 200, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 316, + 200, + 328 + ], + "spans": [ + { + "bbox": [ + 152, + 316, + 200, + 328 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 341, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 296, + 641 + ], + "type": "text", + "content": "Traditional motion deblurring methods struggle to effectively model motion information within the exposure time. Recently, event cameras have attracted significant research interest for its ability to model motion cues over the exposure duration. However, these methods directly fuse event features with image, overlooking the intrinsic heterogeneity of events. In this paper, we identify that the event modality contains two conflicting types of information: edge features and motion cues. Events accumulated over a short exposure period capture sharp edge details but lose motion information, while those accumulated over a long exposure period blur edge details due to motion. To address this issue, we propose a simple yet effective approach to disentangle these two cues from event features and employ an edge-aware sharpening module along with motion-driven scale-adaptive deblurring module to fully leverage both. Specifically, the first module aids in restoring sharp edges by leveraging the clear edge features provided by events, while the second module leverages motion cues to learn diverse blur kernels, adaptively adjusting the receptive field for optimal deblurring. Extensive experiments on synthetic and real-world datasets validate the effectiveness of our approach and yield a substantial improvement over state-of-the-art single-frame methods and surpasses most multi-frame-based methods. Code will be publicly available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 137, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 137, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 137, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "content": "Motion blur is often caused by camera shake or the movement of objects. The task of motion deblurring [1, 8, 19-21]" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 352, + 316, + 410, + 392 + ], + "blocks": [ + { + "bbox": [ + 324, + 318, + 339, + 332 + ], + "lines": [ + { + "bbox": [ + 324, + 318, + 339, + 332 + ], + "spans": [ + { + "bbox": [ + 324, + 318, + 339, + 332 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 316, + 410, + 392 + ], + "lines": [ + { + "bbox": [ + 352, + 316, + 410, + 392 + ], + "spans": [ + { + "bbox": [ + 352, + 316, + 410, + 392 + ], + "type": "image", + "image_path": "537ef25ec2eeb135a1513e1918b7d7a517e1184f6acb2db4e786ec498a4dd313.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 413, + 316, + 476, + 392 + ], + "blocks": [ + { + "bbox": [ + 413, + 316, + 476, + 392 + ], + "lines": [ + { + "bbox": [ + 413, + 316, + 476, + 392 + ], + "spans": [ + { + "bbox": [ + 413, + 316, + 476, + 392 + ], + "type": "image", + "image_path": "8d6ce0b261fb260abaeed3c13e82b96ed1d1e634fc74184112f47dc1d19701bf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 477, + 316, + 540, + 393 + ], + "blocks": [ + { + "bbox": [ + 477, + 316, + 540, + 393 + ], + "lines": [ + { + "bbox": [ + 477, + 316, + 540, + 393 + ], + "spans": [ + { + "bbox": [ + 477, + 316, + 540, + 393 + ], + "type": "image", + "image_path": "01edaff7caadb0444c9efbe55b4360bbd9c88c5a9c849a6042c84ea33bf1a23d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 331, + 402, + 396, + 495 + ], + "blocks": [ + { + "bbox": [ + 324, + 395, + 339, + 408 + ], + "lines": [ + { + "bbox": [ + 324, + 395, + 339, + 408 + ], + "spans": [ + { + "bbox": [ + 324, + 395, + 339, + 408 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 402, + 396, + 495 + ], + "lines": [ + { + "bbox": [ + 331, + 402, + 396, + 495 + ], + "spans": [ + { + "bbox": [ + 331, + 402, + 396, + 495 + ], + "type": "image", + "image_path": "51b98486babf9b2e163501a268bcd4e4761a412c5ce47b26c0a32c20b391b277.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 491, + 339, + 503 + ], + "lines": [ + { + "bbox": [ + 324, + 491, + 339, + 503 + ], + "spans": [ + { + "bbox": [ + 324, + 491, + 339, + 503 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 402, + 402, + 473, + 467 + ], + "blocks": [ + { + "bbox": [ + 402, + 402, + 473, + 467 + ], + "lines": [ + { + "bbox": [ + 402, + 402, + 473, + 467 + ], + "spans": [ + { + "bbox": [ + 402, + 402, + 473, + 467 + ], + "type": "image", + "image_path": "b0176bf9fbee85f14b9d91bb0407d5dedf163889b5e642216958739b6c5ca652.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 481, + 409, + 546, + 464 + ], + "blocks": [ + { + "bbox": [ + 481, + 409, + 546, + 464 + ], + "lines": [ + { + "bbox": [ + 481, + 409, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 481, + 409, + 546, + 464 + ], + "type": "image", + "image_path": "e9f549a2886fcb970dd8f9d58d1efebf670624d83f0d4f1e94184f3a3fcee2d0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 331, + 489, + 396, + 575 + ], + "blocks": [ + { + "bbox": [ + 331, + 489, + 396, + 575 + ], + "lines": [ + { + "bbox": [ + 331, + 489, + 396, + 575 + ], + "spans": [ + { + "bbox": [ + 331, + 489, + 396, + 575 + ], + "type": "image", + "image_path": "94d73839f2a813d7498e818a336bd249e0f01371f9acfbe41df402ae1b8b4e26.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 402, + 491, + 473, + 578 + ], + "blocks": [ + { + "bbox": [ + 402, + 491, + 473, + 578 + ], + "lines": [ + { + "bbox": [ + 402, + 491, + 473, + 578 + ], + "spans": [ + { + "bbox": [ + 402, + 491, + 473, + 578 + ], + "type": "image", + "image_path": "da57136c57da64e49f60922a322a191a5afb3791bab5ba6904feaf328901fc0f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 587, + 555, + 632 + ], + "lines": [ + { + "bbox": [ + 313, + 587, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 555, + 632 + ], + "type": "text", + "content": "Figure 1. (a) Visualization of SCER(Symmetric Cumulative Event Representation) [35]. (b) Previous fusion methods between events and image. (c) Ours. Separate events into edge features and motion cues." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 477, + 514, + 543, + 571 + ], + "blocks": [ + { + "bbox": [ + 477, + 514, + 543, + 571 + ], + "lines": [ + { + "bbox": [ + 477, + 514, + 543, + 571 + ], + "spans": [ + { + "bbox": [ + 477, + 514, + 543, + 571 + ], + "type": "image", + "image_path": "3c0fa5a681be0fb37001a6bd936300fc6f08cfd83b8ebac0084ecce4bd7970aa.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "aims to recover sharp images from blurred ones by restoring clear edges and capturing fine details, ultimately enhancing the image quality. However, since image features lack the ability to model motion cues over the exposure time, they cannot accurately estimate the variant blur ker" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14732" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "nels present in real-world scenarios. This limitation often results in subpar performance of current methods on real-world datasets. Recently, event cameras [2, 9], bio-inspired sensors that asynchronously capture changes in pixel intensities, have been leveraged to aid motion deblurring thanks to their microsecond-level latency and low power consumption." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "text", + "content": "Several studies [6, 11, 35, 47, 48] aim to integrate image and event features to enhance deblurring performance. For example, Sun et al. [35] introduced EFNet, which considers the inherent differences between event and image modalities and utilizes a cross-attention mechanism, improving upon previous methods that relied on simple concatenation or addition for feature fusion. Kim et al. [12] discovered that event streams contain both clear edge and motion cues and proposed a novel attention-based fusion method, allowing them to leverage these event characteristics to extract discriminative channel and spatial information from the blurred image features. Their approaches directly combine the heterogeneous cues contained in events (edge and motion) with RGB features indiscriminately. However, we find that sharp edge features and detailed motion cues are inherently conflicting in their representation and previous one-step fusion approaches [12, 35, 36] tend to bias the event modality toward capturing either edge or motion information, leading to suboptimal results. Specifically, as shown in Figure 1 (a), it is obvious that events provide sharp edge features within short exposure times(the " + }, + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "text", + "content": " column), while accumulating the motion trajectory of objects over the entire exposure period(the " + }, + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 56, + 159, + 296, + 627 + ], + "type": "text", + "content": " column). Therefore, sharp edges typically lack motion information, whereas representations rich in motion cues often compromise edge clarity. As a result, previous one-step fusion methods make it challenging for the model to accurately learn both edge and motion features from events, leading to coarse edges and blurred textures(Figure 1 (b)). Our insight is that edge and motion should be modeled separately and then integrated with RGB to respect both perspectives and avoid conflicts. To restore sharp edges and clear textures, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to independently harness the sharp edge features and rich motion cues provided by the event stream (Figure 1 (c)). Contributing to much better leverage of edge and motion cues, our model shows significant improvements in deblurring on both edges and textures compared to previous methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "A straightforward approach to restore sharp edges is to utilize event features accumulated during a short exposure (the second column in Figure 1 (a)) as edge information and fuse them with image features, but sacrificing the complete event features captured over the full exposure duration. Besides, relying on a fixed time interval to capture sharp edge features is unreliable, as varying motion speeds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "result in different event trigger frequencies: excessive event accumulation can obscure edge details, while insufficient events may fail to capture edges fully. Therefore, our Edge-aware Sharpening module combines both short-exposure event features with complete event cues, using the former as an edge prior to extract complete edge cues by performing cross-attention with the latter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 159, + 553, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 507 + ], + "type": "text", + "content": "As for motion cues, previous work [11, 14-16, 44] has typically utilized events as a substitute for optical flow to align frames, achieving significant advancements in video motion deblurring. Some works [11, 33] have used event to predict optical flow for frame-to-frame alignment. FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\" [13, 39, 40], directly using RGB images and events to achieve frame-to-frame alignment without relying on intermediate steps like optical flow prediction. However, these approaches overlook the crucial role that motion cues play in motion deblurring: different motion patterns yield diverse blur kernels. By leveraging motion cues, we can explicitly estimate blur kernels for various motion patterns, enhancing the model's ability to adapt to a wide range of blurring scenarios. To effectively learn blur kernels of varying scales, previous studies [34, 37, 46] have attempted to expand the receptive field by adding more convolutional layers or using larger convolutional kernel sizes. However, traditional convolutional kernels have a fixed size, which limits their robustness when handling diverse blur kernels, reducing the model's adaptability in real-world scenarios. In our work, inspired by [22, 38, 42], we propose a Motion-driven Scale-adaptive Deblurring module that leverages the rich motion cues provided by the event stream to explicitly estimate different blur kernels, enabling adaptive adjustments to the receptive field to better accommodate a variety of blur patterns. By leveraging motion cues effectively, our single-frame motion deblurring model achieves performance that surpasses most video-based approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 511, + 511, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 511, + 511, + 522 + ], + "spans": [ + { + "bbox": [ + 326, + 511, + 511, + 522 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 526, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 526, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 553, + 598 + ], + "type": "text", + "content": "(1) We find that the edge and motion cues provided by the event stream exhibit a conflict in their representation which is unexplored in the event community. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to effectively utilize these two distinct features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 602, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 553, + 650 + ], + "type": "text", + "content": "(2) We find that motion cues are instrumental in learning variant blur kernels. Building on this insight, we utilize motion cues to achieve adaptive adjustments to the receptive field to accommodate variant blur kernels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "(3) Conducting extensive experiments on both synthetic and real-world datasets, our method demonstrates superior performance, significantly surpassing existing single-frame-based approaches and even most multi-frame-based methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14733" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 278, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 278, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 278, + 104 + ], + "type": "text", + "content": "2.1. Frame-based Motion Deblurring on Image" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 296, + 350 + ], + "type": "text", + "content": "With the advancement of deep learning, convolutional neural networks (CNNs) have been increasingly applied to the motion deblurring. To better handle non-uniform and varying blur kernels, researchers have primarily explored three technical approaches: multi-scale architecture designs [7, 10, 18, 34, 37, 46], more sophisticated decoders [26, 29], and novel network layer designs [22, 30]. For example, SRN(Scale-Recurrent Network) [37] achieves deblurring of high-resolution images by guiding the process with low-resolution images through a multi-scale input-output approach. However, traditional convolutional kernels have spatial uniformity and a fixed receptive field, which makes them suboptimal for modeling dynamic scene deblurring. Although Li et al. [22] introduced deformable convolution to fit blur kernels of varying sizes, their approach still faces challenges in fitting non-uniform and varying blur kernels due to limited motion cues. In our work, we incorporate the event modality, leveraging its high temporal resolution and utilize deformable convolution to address this issue effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 228, + 372 + ], + "type": "text", + "content": "2.2. Event-based Motion Deblurring" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 449 + ], + "type": "text", + "content": "Early work, based on the imaging principles of event cameras and the process of how blurred images are generated, proposed the Event Double Integral (EDI) model [28, 47, 48]. However, due to the uncertain event-triggering thresholds and abundant noise, these methods have not yielded satisfactory results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 688 + ], + "type": "text", + "content": "Other approaches [6, 12, 35, 36] focus on fusing event and image features, training a deblurring network in an end-to-end fashion for improved performance. For example, EFNet [35] utilizes cross-attention mechanism to fuse features from both modalities. REFID [36] employs bidirectional event recurrent (EVR) blocks to extract features from the event streams and introduces an Event-Guided Adaptive Channel Attention module to uncover discriminative channels within the event features. However, these approaches integrate event and image features in a single step, failing to fully account for the edge and motion cues inherent in the event features. In our research, we highlight that event features contain two inherently conflicting aspects: distinct edge information and abundant motion cues. Specifically, distinct edge features help restore the object's sharp boundaries, while abundant motion cues assist the model in learning diverse blur kernels, leading to better recovery of the object's texture details. To leverage each optimally, we introduce an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To utilize the rich motion cues offered by events, several studies [11, 14, 15, 33] focus on employing the event" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 239 + ], + "type": "text", + "content": "modality to handle video motion deblurring tasks. For example, FEVD [15] and CMTA [14] adopt the concept of \"synthesis-based alignment\", utilizing bidirectional recurrent networks and cascade networks respectively, to achieve frame-to-frame alignment. However, these works use events for frame-to-frame alignment, overlooking the role of motion cues in learning variant blur kernels. In this paper, our Motion-Driven Scale-Adaptive Deblurring module effectively leverages motion cues to guide the adjustment of the model's receptive field, enabling it to adapt to diverse blur kernels. This approach ensures robust handling of various motion patterns and achieves competitive performance, even compared to state-of-the-art video-based methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 370, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 370, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 370, + 264 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 272, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 392 + ], + "type": "text", + "content": "Firstly, we introduce the main architecture of our model in Sec. 3.1. In Sec. 3.2, an Edge-aware Sharpening module is proposed to utilize sharp edges captured during short exposure times from event to restore the clear boundaries of deblurring objects. In Sec. 3.3, to adaptively learn variant blur kernels, in the Motion-driven Scale-adaptive Deblurring module, we propose using motion cues to explicitly estimate the blur kernels associated with different motion patterns, enabling the model to adaptively adjust its receptive field for improved deblurring performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 429, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 429, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 429, + 412 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": "Event-based image motion deblurring aims to recover sharp images from a blurred image " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": " and an event stream " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": ". We use SCER [35] to represent the event " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "(\\mathcal{E} = \\{E_i\\}_{i=0}^{2N-1})" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": ". As shown in Figure 2, based on EFNet [35], our model consists of two U-Net [32] networks. During the first stage, an event encoding branch extracts multi-scale features from the event stream " + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 418, + 554, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 502, + 555, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 502, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 392, + 502, + 555, + 517 + ], + "type": "interline_equation", + "content": "\\{F _ {i} ^ {E} \\} _ {i = 0} ^ {3} = R e s (\\mathcal {E}) \\tag {1}", + "image_path": "b8f700ea92aa606ab7a20ac13b0e4fee6160478b584c1f92321ecbab30bc355a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 523, + 555, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 691 + ], + "type": "text", + "content": "Event features encompass both object edge information and motion cues. Previous approaches [12, 14, 35] often fused these event features directly, which can cause edge information and motion cues to conflict, leading to a model that neither fully utilizes the events to sharpen object edges nor effectively exploits the motion cues for deblurring. Specifically, event stream contains clear edge features from short exposure durations and motion cues accumulated over longer exposure times. This one-step fusion approach can easily cause the model to neglect one aspect of the event information. Therefore, an Edge-aware Sharpening module and a Motion-driven Scale-adaptive Deblurring module are proposed to fully leverage the features contained in the event stream respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 341, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\{F _ {i} ^ {I} \\} _ {i = 1} ^ {3} = f _ {2} (f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}), F _ {i} ^ {m}) \\qquad (2)", + "image_path": "046c8541df21fae986de2dda6d9c1227cec0e3f6c66447457829ed8ac6f5ee81.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14734" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 545, + 381 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 545, + 381 + ], + "type": "image", + "image_path": "f4262f2a4b8b4f113522ba28a561d03b020def9f764ef3f2a43ed9094a7dfa96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 392, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 392, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 555, + 460 + ], + "type": "text", + "content": "Figure 2. The main architecture of our model. Based on EFNet [35], we propose a two-stage U-Net architecture to achieve coarse-to-fine motion deblurring. During the first stage, we separate event into edge features and motion cues to guide the model: edge features enhance the recovery of sharp image edges with an Edge-aware Sharpening module, while motion cues help the model learn diverse blur kernels through Motion-driven Scale-adaptive Deblurring module. In the second stage, similar to EFNet, the output from the first stage is concatenated with the original blurred image. An EMGC (Event Mask Gated Connection) [35] module is then applied to enhance features specifically in the regions events triggered." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "f_{1}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": " means Edge-aware Sharpening module and " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "f_{2}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": " means Motion-driven Scale-adaptive Deblurring module. " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "F_{i}^{e}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "F_{i}^{m}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": " represent edge features and motion cues from the event stream " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": " respectively. Next, a standard U-Net decoder layer with skip connections is employed to predict the deblurred image " + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "inline_equation", + "content": "I^{initial}" + }, + { + "bbox": [ + 55, + 479, + 296, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 560, + 295, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 560, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 117, + 560, + 295, + 577 + ], + "type": "interline_equation", + "content": "\\left\\{D _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(D _ {i - 1} ^ {I}, F _ {i} ^ {I}\\right) \\tag {3}", + "image_path": "5aab4431562d4a6e203ad86c7e33fe975e134a1ee56a0a45dcc5ee4f030b8d63.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 586, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 586, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 125, + 586, + 295, + 602 + ], + "type": "interline_equation", + "content": "I ^ {i n i t i a l} = \\operatorname {c o n v} _ {3 \\times 3} \\left(D _ {3} ^ {I}\\right) \\tag {4}", + "image_path": "220475fe98ed8176eb14fd62c9ed423d00ceeca882425898282afc3f2eb54cf0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 609, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 609, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 113, + 609, + 295, + 623 + ], + "type": "interline_equation", + "content": "L _ {1} = P S N R L o s s (I ^ {\\text {i n i t i a l}}, I) \\tag {5}", + "image_path": "de5efd5fc08f73829b8e0f1bfbf4cd70bfb41ca034d40c56590688286b937192.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Considering that event-triggered areas often suffer from more severe blur, in the second stage, the deblurred image " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "I^{initial}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " is concatenated with the blurred image B and then processed through another U-Net network to focus on restoring the event-triggered areas. EMGC(Event Mask Gated Connection) [35] module generates masks from events to create skip connections from the first stage to the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 516 + ], + "type": "text", + "content": "encoder of the second stage, guiding the model to concentrate on event-triggered areas. The encoder layer can be represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 518, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 518, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 318, + 518, + 553, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) = c o n v \\left(F _ {i} ^ {I}\\right) + c o n v \\left(F _ {i} ^ {I} \\times (1 - \\mathcal {M})\\right) \\\\ + \\operatorname {c o n v} \\left(D _ {i} ^ {I}\\right) + \\operatorname {c o n v} \\left(D _ {i} ^ {I} \\times \\mathcal {M}\\right) \\tag {6} \\\\ \\end{array}", + "image_path": "50f1016e5408768452aa04e341f28ca6284a08305bec9308601471c58ebbc6a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 554, + 553, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 554, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 344, + 554, + 553, + 569 + ], + "type": "interline_equation", + "content": "\\left\\{\\hat {F} _ {i} ^ {I} \\right\\} _ {i = 1} ^ {3} = R e s \\left(\\hat {F} _ {i - 1} ^ {I}\\right) + E M G C \\left(F _ {i} ^ {I}, D _ {i} ^ {I}\\right) \\tag {7}", + "image_path": "fffba684b8618e9d1f3f8eb509576f9e857a2822c1f3ae0a816c380df9f0c75b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "text", + "content": " represents the mask generated from event stream " + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "text", + "content": ". Pixels where events triggered are set to 1, while others are set to 0. As a result, feature maps from encoder overlook the event-triggered areas, which may contain severe blur, while feature maps from the decoder enhance these regions, reflecting the restored blurred content. Finally, the final deblurred result " + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "inline_equation", + "content": "I^{final}" + }, + { + "bbox": [ + 313, + 571, + 555, + 679 + ], + "type": "text", + "content": " is obtained after processing through a decoder identical to the one used in the first stage. Loss function is expressed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 373, + 682, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 682, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 373, + 682, + 553, + 697 + ], + "type": "interline_equation", + "content": "L _ {2} = P S N R L o s s \\left(I ^ {f i n a l}, I\\right) \\tag {8}", + "image_path": "acd6c68d0d18b7e029e6c62b2443deeb8269a2ee94558ef48bd16d92f694c23d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 374, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\omega \\times L _ {1} + (1 - \\omega) \\times L _ {2} \\tag {9}", + "image_path": "fb9a055b413ff80ab027e3653ff6eb92a44e5d3777df0be6dd0e8080649751bb.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "14735" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 230, + 85 + ], + "type": "text", + "content": "3.2. Edge-aware Sharpening Module" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 340 + ], + "type": "text", + "content": "As shown in the second column in Figure 1 (a), the accumulation of events over a short period often provides clear edge features of objects, which help restore sharp boundaries for blurred objects. Consequently, FEVD [15] employs fast Fourier transform to simultaneously capture the high-frequency signals, which typically contain detailed edge cues, from both events and images. However, event features often contain abundant motion cues, which can blur the sharp edge features and make it difficult to extract clear edge information from events. On the other hand, even objects undergoing rapid motion can accumulate clear edge features within short exposure times. This characteristic is particularly valuable for preserving fine details, as edges play a crucial role in maintaining the integrity of object boundaries and structures in the image. However, when combined with motion cues, which are more prominent in longer exposure times, these edge features can become diluted or lost. Therefore, it is essential to selectively enhance the edge information from the event stream to ensure that sharp boundaries are preserved, while still utilizing the motion information to handle blur effectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 341, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 472 + ], + "type": "text", + "content": "To address this issue, we propose an Edge-aware Sharpening module that firstly combines the event features accumulated over a short period with the image to enhance the edges of moving objects. These enhanced features serve as queries, allowing the attention mechanism to effectively integrate the edge features from the event stream. Specifically, SCER [35] uses the midpoint as a reference and accumulates " + }, + { + "bbox": [ + 55, + 341, + 296, + 472 + ], + "type": "inline_equation", + "content": "2 \\times N" + }, + { + "bbox": [ + 55, + 341, + 296, + 472 + ], + "type": "text", + "content": " frames of events symmetrically from this point. We select the channels closest to the midpoint as edge priors to enhance the image features. Thus, our Edge-aware Sharpening module can be represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 480, + 294, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 480, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 107, + 480, + 294, + 493 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {e} = \\operatorname {c o n v} \\left(E _ {N - 1} \\| E _ {N}\\right) \\tag {10}", + "image_path": "8944f3a2530cc32ae85728e5a3ff1a18e57c798e1412df1ae5e1ef3ee538e6fd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 495, + 206, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 495, + 206, + 510 + ], + "spans": [ + { + "bbox": [ + 97, + 495, + 206, + 510 + ], + "type": "interline_equation", + "content": "f _ {1} (R e s (F _ {i - 1} ^ {I}), F _ {i} ^ {e}, F _ {i} ^ {E}) =", + "image_path": "7ba4d1164de3ae1d20c7f730d33357f74ab9caa9a2fc0da1f143163a6af8472e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 512, + 294, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 512, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 107, + 512, + 294, + 526 + ], + "type": "interline_equation", + "content": "A t t \\left(\\operatorname {c o n v} \\left(R e s \\left(F _ {i - 1} ^ {I}\\right) \\| F _ {i} ^ {e}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {11}", + "image_path": "e3ca2cf764ea124dd8151e9317e1fcfc0ccc2bf2d27469af090c18779e7fba9a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": " means connection between features and " + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "inline_equation", + "content": "Att(x,y,z)" + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": " denotes an attention mechanism where " + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": " serves as the query, " + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": " as the key, and " + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 533, + 296, + 570 + ], + "type": "text", + "content": " as the value." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 296, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 600 + ], + "type": "text", + "content": "3.3. Motion-driven Scale-adaptive Deblurring Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 688 + ], + "type": "text", + "content": "Previous approaches have often concentrated on utilizing event streams for frame-to-frame alignment to achieve video motion deblurring. However, these methods frequently overlook the importance of motion features in the context of static image deblurring. The motion cues of objects during exposure time help us explicitly learn variant blur kernels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "As shown in the third column in Figure 1 (a), the girl exhibits a larger blur kernel, causing the edges accumulated" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "type": "text", + "content": "in the event stream to appear broader and more diffused. This effect is in stark contrast to the background crowd, where the blur kernel is smaller due to less motion, resulting in sharper and more defined edges in the event accumulation. These varying blur kernels indicate differences in the degree of motion across the scene, which are essential for accurately restoring fine textures in the deblurring process. To effectively handle such diverse blur kernels, we introduce a Motion-driven Scale-adaptive Deblurring module that adaptively adjusts the convolutional receptive field based on the degree of motion captured in the event stream. This is achieved through the use of deformable convolutions, which enable the network to learn dynamic and spatially flexible convolutional kernels. This approach allows the model to handle both large, motion-induced blur (such as that of the girl) and smaller, subtle blur (such as that of the background crowd) simultaneously, improving the overall deblurring performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "text", + "content": "As mentioned earlier, the event stream contains both edge and motion features. To highlight the motion cues of objects, it is essential to separate the motion cues from the event features. Specifically, we use events from both sides of SCER as motion priors " + }, + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "inline_equation", + "content": "(E_0" + }, + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "inline_equation", + "content": "E_{2N - 1})" + }, + { + "bbox": [ + 313, + 289, + 556, + 371 + ], + "type": "text", + "content": " and apply cross-attention to capture detailed motion cues from the event features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 351, + 380, + 555, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 380, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 351, + 380, + 555, + 396 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {m} = \\operatorname {A t t} \\left(\\operatorname {c o n v} \\left(E _ {0} \\| E _ {2 N - 1}\\right), F _ {i} ^ {E}, F _ {i} ^ {E}\\right) \\tag {12}", + "image_path": "785ccc37ca12d6bbcb44cd84476ab2a7c004d07e81bb908ed78c582b2499cfb3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 405, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 453 + ], + "type": "text", + "content": "Then, we calculate offsets with the motion cues to explicitly estimate variant blur kernels with different scales, and employ deformable convolutions to capture contextual features, enhancing the ability to express blur effectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 400, + 464, + 553, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 464, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 400, + 464, + 553, + 476 + ], + "type": "interline_equation", + "content": "\\Theta = \\operatorname {c o n v} \\left(F _ {i} ^ {m}\\right) \\tag {13}", + "image_path": "4b6837855f37c2a306c694d06e56138e467522e98537cd4197bca5c83fec46a5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 487, + 553, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 487, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 380, + 487, + 553, + 501 + ], + "type": "interline_equation", + "content": "f _ {2} \\left(f, F _ {i} ^ {m}\\right) = D C N (f, \\Theta) \\tag {14}", + "image_path": "e0b597af9085464c65b5e8b15261657088710337de0f0be22e67cc501c3ac8b1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 507, + 534, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 534, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 534, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 507, + 534, + 518 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 507, + 534, + 518 + ], + "type": "text", + "content": " means the offsets of deformable convolution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 531, + 391, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 391, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 391, + 545 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 552, + 493, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 493, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 493, + 565 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "content": "Synthetic Dataset. To make a fair comparison with previous methods based on events and images, we selected the GoPro dataset [27], which is widely used for deblurring tasks. We utilized ESIM [31] to generate the event stream, applying the same parameters as other studies, specifically using a Gaussian normal distribution with parameters " + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu = 0.2,\\sigma = 0.03)" + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Real-world Datasets. To validate the deblurring capabilities of event data in real-world scenarios, REBlur [35] collects events from Dynamic and Active Pixel Vision Sensor (DAVIS) [2] in real environments and uses an electronically controlled stepping motor to relocate the slide-rail" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14736" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "to the coordinate corresponding to the timestamp t. However, DAVIS-346 camera has low image quality and resolution. Therefore, high resolution and quality event-based real-world deblurring datasets REVD [15] and EVRB [14] are proposed with Triple-axis Hybrid Camera System." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 239 + ], + "type": "text", + "content": "Evaluation Metrics. PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure) [43] are proposed to assess image quality. PSNR calculates the ratio of signal power to noise, with higher values indicating closer similarity to the reference image, though it mainly reflects pixel-level fidelity. SSIM, on the other hand, measures structural, luminance, and contrast similarities, providing a more perceptually relevant quality assessment, with values closer to 1 indicating better similarity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "text", + "content": "Implementation Details. Our model supports end-to-end training with a batch size of 16. The N of SCER [35] is set to 8. For data augmentation, we randomly crop images to " + }, + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "text", + "content": " pixels and apply random horizontal and vertical flips. We chose Adam [17] as the optimizer. Training runs for 400,000 iterations, starting with a learning rate of 2e-4, which is progressively reduced to a minimum of 1e-7 using CosineAnnealingLR [25]." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 432, + 293, + 653 + ], + "blocks": [ + { + "bbox": [ + 55, + 346, + 294, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 294, + 422 + ], + "type": "text", + "content": "Table 1. Quantitative results on the GoPro [27] dataset. " + }, + { + "bbox": [ + 55, + 346, + 294, + 422 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 346, + 294, + 422 + ], + "type": "text", + "content": " indicates that higher values correspond to better performance. The numbers in “()” indicate the publication year of the referenced papers. In the second column, “Single” in black means frame-based method, and red represents methods based on single-frame and event while green means multi-frame event-based methods. The best results are indicated in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 432, + 293, + 653 + ], + "lines": [ + { + "bbox": [ + 57, + 432, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 293, + 653 + ], + "type": "table", + "html": "
MethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
Restormer [45] (22')Single32.920.961
NAFNet [4] (22')Single33.690.967
FFTformer [18] (23')Single34.210.969
GRL-B [23] (23')Single33.930.968
MLWNet [10] (24')Single33.830.968
Jianget al. [11] (20')Single31.790.949
REDNet [44] (21')Single35.070.971
EFNet [35] (22')Single35.460.973
SpkNet [5] (23')Single36.120.971
D2Nets [33] (21')3 frames31.760.943
UEVD [12] (22')2 frames35.480.971
REFID [36] (23')2 frames35.910.973
FEVD [15] (24')Multi-frames36.700.978
CMTA-5 [14] (24')5 frames36.550.977
CMTA-7 [14] (24')7 frames36.780.977
OursSingle36.700.977
", + "image_path": "f238a818858af0e03e75d285c3b83dd7f3603cb5a81bddf7d364324bfdbd98a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 258, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 258, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 258, + 684 + ], + "type": "text", + "content": "4.2. Compared to State-of-the-art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "Synthetic Dataset. We evaluate our proposed network against state-of-the-art motion deblurring methods on the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 90, + 549, + 299 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 537, + 82 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 537, + 82 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 537, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative results on three real-world datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 90, + 549, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 90, + 549, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 90, + 549, + 299 + ], + "type": "table", + "html": "
DatasetMethodFrames for InferenceMetrics
PSNR ↑SSIM ↑
REBlurSNR+ [37]Single36.870.970
HINet+ [3]Single37.680.973
EFNet [35]Single38.120.975
OursSingle38.530.977
EVRBUEVD [12]2 frames30.550.915
EFNet [35]Single30.980.927
REFID [36]2 frames30.330.918
CMTA [14]Multi-frames31.380.927
OursSingle32.050.936
REVDeSL-Net [41]Single26.990.787
EFNet [35]Single31.750.920
REDNet [44]Single31.900.920
UEVD [12]2 frames31.970.921
FEVD [15]Multi-frames32.990.933
OursSingle32.140.922
", + "image_path": "2207bc5ecf85d9bff28b66ddd1868d6608c353a6d0450ed30d03f902c6456383.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "text", + "content": "GoPro [27] dataset. The comparison covers three categories: frame-based image deblurring, event(spike)-based image deblurring and event-based video deblurring. As shown in Table 1, in comparison to SpkNet [5], the existing state-of-the-art spike-based image motion deblurring method, our model achieves an impressive performance boost of " + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "inline_equation", + "content": "+0.58" + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "text", + "content": " dB. Moreover, compared to CMTA [14], the leading video motion deblurring method, our single-frame model outperforms CMTA-5 (which uses five frames for inference) by " + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "inline_equation", + "content": "+0.15" + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "text", + "content": " dB and is only " + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "inline_equation", + "content": "-0.08" + }, + { + "bbox": [ + 313, + 334, + 553, + 574 + ], + "type": "text", + "content": " dB behind CMTA-7. Thus, our model demonstrates the remarkable effectiveness of event features in the task of image motion deblurring. We show qualitative results on GoPro in Figure 3, RVRT [24], a video-based motion deblurring method, models long-term temporal dependencies to provide contextual references for deblurring the current frame. MLWNet [10] stands as the leading single-frame motion deblurring approach, while EFNet [35] is an event-based image deblurring method. It's evident that our approach achieves superior detail restoration compared to these previous methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Real-world Datasets. To assess our model's effectiveness in real-world settings, we utilized three datasets: REBlur [35], REVD [15], and EVRB [14]. For REBlur, given its limited dataset size, we first train the model on the GoPro [27] dataset and then finetune it on REBlur training set. For FEVD and EVRB, we retrained the model according to the training and testing set splits provided in the official guidelines. As shown in Table 2, our method surpasses event-based image motion deblurring approaches across all three datasets and even outperforms the current top video-based methods on EVRB [14]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14737" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 71, + 517, + 533 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 517, + 533 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 517, + 533 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 517, + 533 + ], + "type": "image", + "image_path": "9798852e165c9a5c59a105ed6da9051949ce718f0e7f77f36d7842b7f645dc59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 552, + 456, + 564 + ], + "lines": [ + { + "bbox": [ + 153, + 552, + 456, + 564 + ], + "spans": [ + { + "bbox": [ + 153, + 552, + 456, + 564 + ], + "type": "text", + "content": "Figure 3. Visual results with our model and other SOTA methods on GoPro dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 585, + 149, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 149, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 149, + 598 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": "Our baseline model " + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 55, + 605, + 297, + 713 + ], + "type": "text", + "content": " is a standard two-stage U-Net that does not utilize any event features. As shown in Table 3, it is evident that the integration of event modalities results in significant performance enhancements, as demonstrated by both attention-based event fusion and event-driven deformable convolutions, which yield impressive outcomes. As shown in Figure 4, the baseline performs poorly in deblurring, failing to effectively restore either sharp edges or clear textures. And with the introduction of the event" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 586, + 514, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 514, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 514, + 598 + ], + "type": "text", + "content": "modality, the visual results improve significantly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "As shown in the first part in Table 3, simply applying cross-attention to fuse event and image modalities " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "(B + Event^{Att})" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ", yields promising results. However, this straightforward fusion approach overlooks the inherent conflict between motion and edge features in event streams, which prevents the model from accurately capturing sharp edge details. As shown in the first example of Figure 4, when image features are fused with events, the text in the yellow box becomes recognizable but still those in blue" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14738" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 539, + 293 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 539, + 293 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 539, + 293 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 539, + 293 + ], + "type": "image", + "image_path": "25730c6f6a6328d98c9c5ac4f2c1a08f32b20dac26e4efb9c46d28ac8dc3b1d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 303, + 534, + 316 + ], + "lines": [ + { + "bbox": [ + 74, + 303, + 534, + 316 + ], + "spans": [ + { + "bbox": [ + 74, + 303, + 534, + 316 + ], + "type": "text", + "content": "Figure 4. Visual results of ablation study on EVRB [14] dataset. It is recommended to zoom in for a clearer view of the details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 430, + 288, + 555 + ], + "blocks": [ + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": "Table 3. Ablation study on EVRB [14] dataset. " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " represents our baseline, which does not utilize event streams to enhance the image. " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "Event^{Att}" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " denotes applying cross-attention to integrate image with event features. " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "Edge^{Att}" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " means our Edge-aware Sharpening module. " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "Img^{DCN}" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "Event^{DCN}" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " represent utilizing image features and event features to calculate offsets for the deformable convolution respectively. " + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "inline_equation", + "content": "Motion^{DCN}" + }, + { + "bbox": [ + 55, + 334, + 296, + 422 + ], + "type": "text", + "content": " is our Motion-driven Scale-adaptive Deblurring module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 430, + 288, + 555 + ], + "lines": [ + { + "bbox": [ + 61, + 430, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 61, + 430, + 288, + 555 + ], + "type": "table", + "html": "
MethodMetrics
PSNRSSIM
B28.650.900
B+EventAtt30.710.923
B+EdgeAtt31.130.929
B+ImgDCN28.800.894
B+EventDCN30.510.921
B+MotionDCN31.630.930
Ours(B+EdgeAtt + MotionDCN)32.050.936
", + "image_path": "9b64544ab5e4973a85ee202be68d18bd295688d2dcbf7c1f679c2b4ee799db00.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": "box retain significant blurriness. Instead, our Edge-aware Sharpening module " + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "inline_equation", + "content": "(B + Edge^{Att})" + }, + { + "bbox": [ + 55, + 578, + 296, + 628 + ], + "type": "text", + "content": " strengthens the representation of edge features within the events, producing a more refined result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "The second part of the Table 3 demonstrates the value of motion cues in learning diverse blur kernels. Initially, we utilize image features to determine offsets for deformable convolutions " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "(B + Img^{DCN})" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "; however, because image lacks the motion cues, they are unable to accurately capture the diverse blur kernels that correspond to different motion characteristics. As shown in the second example in Fig-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "content": "ure 4, using image features to estimate the blur kernel still results in artifacts around the \"No Parking Sign\". Conversely, training deformable convolutions with event features " + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "inline_equation", + "content": "(B + Event^{DCN})" + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "content": " leads to a better performance, but it still results in problematic artifacts in some local regions (in the yellow box). To further improve the capture of motion information, our proposed Motion-driven Scale-adaptive Deblurring module " + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "inline_equation", + "content": "(B + Motion^{DCN})" + }, + { + "bbox": [ + 313, + 335, + 555, + 467 + ], + "type": "text", + "content": " accurately captures object motion cues and estimates non-uniform and varying blur kernels, significantly minimizing artifacts both globally and locally." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 467, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 504 + ], + "type": "text", + "content": "Finally, by effectively fusing both motion and edge cues, our model demonstrates substantial improvements in the restoration of both sharp edges and fine texture details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we observe that event features inherently contain conflicting edge and motion cues: edge features accumulate over short exposure times lack sufficient motion information, while motion cues accumulate over long exposure times compromise edge clarity. To address this, we propose an Edge-Aware Sharpening module and a Motion-Driven Scale-Adaptive Deblurring module to separately leverage these two aspects of cues. By incorporating edge and motion priors, we enhance the corresponding cues in event features: edge cues restore sharp object edges, while motion cues guide deformable convolutions to adaptively estimate blur kernels, adjusting the receptive field for diverse motion patterns. Extensive experiments on synthetic and real-world datasets demonstrate the superior performance of our approach." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14739" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "type": "text", + "content": "The work is jointly supported by the National Natural Science Foundation of China (NSFC) under Grant 62261160576 and Grant 62203024, the Beijing Natural Science Foundation (4252026), the Research and Development Program of Beijing Municipal Education Commission (KM202310005027), and the Fundamental Research Funds for the Central Universities of China. This research work is also supported by the Big Data Computing Center of Southeast University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 274 + ], + "type": "text", + "content": "[1] Yuval Bahat, Netelee Efrat, and Michal Irani. Non-uniform blind deblurring by reblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3286-3294, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "text", + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 61, + 276, + 296, + 319 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 321, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 296, + 374 + ], + "type": "text", + "content": "[3] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 420 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 465 + ], + "type": "text", + "content": "[5] Shiyan Chen, Jiyuan Zhang, Yajing Zheng, Tiejun Huang, and Zhaofei Yu. Enhancing motion deblurring in high-speed scenes with spike streams. Advances in Neural Information Processing Systems, 36:70279-70292, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "type": "text", + "content": "[6] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12492-12503, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 578 + ], + "type": "text", + "content": "[7] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 633 + ], + "type": "text", + "content": "[8] Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T Roweis, and William T Freeman. Removing camera shake from a single photograph. In ACM SIG International Conference on Computer Graphics and Interactive Techniques, pages 787-794. 2006. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Xin Gao, Tianheng Qiu, Xinyu Zhang, Hanlin Bai, Kang Liu, Xuan Huang, Hu Wei, Guoying Zhang, and Huaping Liu." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Efficient multi-scale network with learnable discrete wavelet transform for blind motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2733-2742, 2024. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "text", + "content": "[11] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18032-18042, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 319 + ], + "type": "text", + "content": "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cross-modal temporal alignment for event-guided video deblurring. European Conference on Computer Vision, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 554, + 374 + ], + "type": "text", + "content": "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 420 + ], + "type": "text", + "content": "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 444 + ], + "type": "text", + "content": "[17] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 499 + ], + "type": "text", + "content": "[18] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 567 + ], + "type": "text", + "content": "[19] Jan Kotera, Filip Šroubek, and Peyman Milanfar. Blind deconvolution using alternating maximum a posteriori estimation with heavy-tailed priors. In Computer Analysis of Images and Patterns: 15th International Conference, CAIP 2013, York, UK, August 27-29, 2013, Proceedings, Part II 15, pages 59-66. Springer, 2013. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 555, + 612 + ], + "type": "text", + "content": "[20] Dilip Krishnan, Terence Tay, and Rob Fergus. Blind deconvolution using a normalized sparsity measure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 233-240. IEEE, 2011." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 667 + ], + "type": "text", + "content": "[21] Anat Levin, Yair Weiss, Fredo Durand, and William T Freeman. Understanding and evaluating blind deconvolution algorithms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1964-1971. IEEE, 2009. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[22] Jichun Li, Weimin Tan, and Bo Yan. Perceptual variousness motion deblurring with light global context refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4116-4125, 2021. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14740" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 194 + ], + "type": "text", + "content": "[24] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 226 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[26] Xintian Mao, Qingli Li, and Yan Wang. Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25681-25690, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "type": "text", + "content": "[27] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3883-3891, 2017. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[28] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "text", + "content": "[29] Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. Multi-temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European Conference on Computer Vision, pages 327-343. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[30] Kuldeep Purohit and AN Rajagopalan. Region-adaptive dense network for efficient motion deblurringli2021 perceptual. In Association for the Advancement of Artificial Intelligence, pages 11882-11889, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 524 + ], + "type": "text", + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. pages 969-982. PMLR, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer Assisted Intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[33] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2021. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "text", + "content": "[34] Maitreya Suin, Kuldeep Purohit, and AN Rajagopalan. Spatially-attentive patch-hierarchical network for adaptive motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3606–3615, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[35] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[36] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18043-18052, 2023. 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 226 + ], + "type": "text", + "content": "[37] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[38] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 347 + ], + "type": "text", + "content": "[39] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "type": "text", + "content": "[40] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17755-17764, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[41] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In European Conference on Computer Vision, pages 155-171. Springer, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[42] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 558 + ], + "type": "text", + "content": "[43] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[44] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion deblurring with real events. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2583-2592, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "type": "text", + "content": "[45] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[46] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14741" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 5978-5986, 2019. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[47] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17765-17774, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[48] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing event-based motion deblurring in real-world scenarios. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10734-10744, 2023. 2, 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14742" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_content_list.json b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3eb4f2952e84994db809aa9d160fd013a5bf0be3 --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions", + "text_level": 1, + "bbox": [ + 120, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengwei Xie $^{1}$ , Shuang Zeng $^{1,2}$ , Xinyuan Chang $^{1}$ , Xinran Liu $^{1}$ , Zheng Pan $^{1}$ , Mu Xu $^{1}$ , Xing Wei $^{2\\dagger}$", + "bbox": [ + 222, + 179, + 774, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Amap, Alibaba Group 2Xi'an Jiaotong University", + "bbox": [ + 285, + 215, + 696, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xiemengwei.xmw, changxinyuan.cxy, tom.lxr, panzheng.pan, xumu.xm}@alibaba-inc.com,", + "bbox": [ + 132, + 236, + 861, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zengshuang@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn", + "bbox": [ + 263, + 253, + 723, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate lane topology is essential for autonomous driving, yet traditional methods struggle to model the complex, non-linear structures—such as loops and bidirectional lanes—prevalent in real-world road structure. We present SeqGrowGraph, a novel framework that learns lane topology as a chain of graph expansions, inspired by human map-drawing processes. Representing the lane graph as a directed graph $G = (V,E)$ , with intersections $(V)$ and centerlines $(E)$ , SeqGrowGraph incrementally constructs this graph by introducing one vertex at a time. At each step, an adjacency matrix $(A)$ expands from $n \\times n$ to $(n + 1) \\times (n + 1)$ to encode connectivity, while a geometric matrix $(M)$ captures centerline shapes as quadratic Bézier curves. The graph is serialized into sequences, enabling a transformer model to autoregressively predict the chain of expansions, guided by a depth-first search ordering. Evaluated on nuScenes and Argoverse 2 datasets, SeqGrowGraph achieves state-of-the-art performance.", + "bbox": [ + 89, + 335, + 485, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 638, + 220, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rise of autonomous driving has heightened the need for precise road structure modeling to ensure safe and efficient navigation [8, 23, 32]. Road structures encompass explicit features such as centerlines, road edges, and traffic signs, as well as implicit topological relationships, notably the connectivity between centerlines. Constructing lane graphs is crucial for representing these complex structures, facilitating vehicle decision-making and path planning in dynamic environments.", + "bbox": [ + 89, + 662, + 483, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional lane graph generation methods fall into detection-based and generation-based approaches, each with notable limitations. Detection-based methods can be categorized into three types: 1) Pixel-level approaches, such as HDMapNet [9], segment lane structures in bird's-eye view", + "bbox": [ + 89, + 800, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2560af0b17c3d36e390afeaf88b6f8b20e0a32232999df17d4a092966ec9d5f.jpg", + "image_caption": [ + "Figure 1. Describing the lane graph step by step. Starting from an initial point, it is necessary to determine the location of the next point and establish whether there is a \"from\" or \"to\" relationship with the existing points. By repetitively carrying out this process, a complete description of the lane graph can be obtained. The abbreviation \"pos\" in the diagram stands for \"position\"." + ], + "image_footnote": [], + "bbox": [ + 535, + 303, + 885, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(BEV) space on a per-pixel basis but struggle to infer global topology. 2) Piece-level methods [4, 10, 18] detect centerlines in fragmented segments based on topological points, often resulting in discontinuities and misalignments (Figure 2). 3) Path-level methods, such as LaneGAP [15], aim to mitigate discontinuities by detecting lane paths but introduce redundancies and alignment errors due to heavy reliance on post-processing.", + "bbox": [ + 511, + 696, + 906, + 818 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Within the widely applied paradigm of sequence generation [2, 6, 12, 26], generative models are regarded as offering a more direct method for inferring lane graphs by capturing hidden structural patterns.", + "bbox": [ + 511, + 821, + 906, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, formulating lane graphs as sequences poses", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 91, + 886, + 222, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27166", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenges and existing generative models have certain limitations. For instance, RNTR [17] converts lane graph into a Directed Acyclic Graph (DAG), which brings sequence redundancy in pre-processing and post-processing and fails to express loops and bidirectional lanes. LaneGraph2Seq [21], another generation-based method, generates the junction positions and shapes of centerlines separately, missing a holistic understanding of the network topology and struggling with integrated representation. And DAG-based methods fail to express loops and bidirectional lanes. Existing methodologies for graph-to-sequence conversion tend to excessively deconstruct the intrinsic graph structure, leading to transformations that disrupt the original topology.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Wait—how do humans construct lane graphs? Instead of processing an entire graph at once, people naturally start from a single node and progressively add new elements, continuously expanding the graph by establishing connections with previously defined nodes (Figure 1). This intuitive approach suggests a novel formulation: 1) Modeling lane graph construction as a chain of stepwise expansion process. 2) Using sequence modeling to capture incremental adjacency matrix updates.", + "bbox": [ + 89, + 289, + 483, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this observation, we propose SeqGrowGraph, a novel framework for incrementally constructing a lane Graph via Sequence Growth. SeqGrowGraph represents a lane graph as a directed graph $G = (V,E)$ , where $V$ denotes intersections or key topological nodes, and $E$ represents centerlines, defining connectivity between nodes. SeqGrowGraph models lane topology as a chain of graph expansions:", + "bbox": [ + 89, + 426, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new node is introduced with its spatial position.", + "- The adjacency matrix $A$ expands from $n \\times n$ to $(n + 1) \\times (n + 1)$ to encode connectivity, with the upper triangle representing incoming edges (\"from\") and the lower triangle representing outgoing edges (\"to\").", + "- A geometric matrix $M$ updates the centerline shape, represented as quadratic Bézier curves.", + "- A transformer-based generative model autoregressively predicts these expansions using a depth-first search ordering." + ], + "bbox": [ + 89, + 549, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike DAG-based approaches, SeqGrowGraph flexibly models complex, real-world road structures, including loops, bidirectional lanes, and non-trivial topologies, without excessive pre-processing or post-processing. Our key contributions are as follows:", + "bbox": [ + 89, + 702, + 483, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce SeqGrowGraph, a novel framework that reformulates lane graph generation as a sequential expansion process, leveraging adjacency matrix updates and geometric modeling.", + "- We develop an autoregressive transformer-based model that progressively constructs lane graphs, overcoming the limitations of DAG-based methods.", + "- We achieve state-of-the-art performance on large-scale" + ], + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/be1ea35d80df5457e87ea0355480a07424a6d106e1e9e662aceed79bef43e1c6.jpg", + "image_caption": [ + "Figure 2. Visualization comparison of lane graph results from different methods. This figure illustrates the inference results using the official models of TopoNet [10] and LaneGAP [15]. It can be observed that TopoNet, by detecting centerlines separately, results in discontinuities between the centerlines. In contrast, LaneGAP, which detects at the path level, tends to confuse multiple lanes and duplicate paths. Both methods rely heavily on extensive post-processing, failing to generate continuous centerline results in a single step." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 890, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets (nuScenes, Argoverse 2), demonstrating superior topological accuracy and network completeness.", + "bbox": [ + 526, + 412, + 906, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SeqGrowGraph offers a practical and effective solution for lane topology modeling, aligning closely with the natural human approach to map construction.", + "bbox": [ + 511, + 443, + 905, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 503, + 650, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Online HD map construction", + "text_level": 1, + "bbox": [ + 513, + 530, + 772, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Online high-definition map construction enables accurate road geometry modeling for autonomous navigation, with contemporary approaches [9, 13, 14, 16, 30, 33] focusing on BEV-space localization of 2D map elements. Initial approaches conceptualized this task as a semantic segmentation problem, as demonstrated in foundational works like HDMapNet [9], which necessitated computationally intensive post-processing procedures to convert pixel-level segmentation outputs into usable vector formats. To address this limitation, VectorMapNet [16] has focused on developing end-to-end frameworks for vectorized map learning. This approach employs a two-phase hierarchical architecture utilizing topological keypoint detection, implementing an initial coarse prediction stage followed by geometric refinement processes to achieve precise vectorized outputs. Notably, MapTR [13] achieves competitive performance using only PV image inputs, eliminating LiDAR dependency. 3D lane detection in autonomous driving follows two primary paradigms: BEV-based and BEV-free approaches. BEV-based methods [11, 23, 28] transform perspective view (PV) images into bird's-eye view (BEV) representations using camera intrinsics and ground coordinates to extract height-aware BEV features for accurate 3D lane modeling.", + "bbox": [ + 511, + 551, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27167", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09e7b0799a769dda9df849901e8e4e4575170b869d803d2b210e69a1d54f46f3.jpg", + "image_caption": [ + "Figure 3. Expansion process of the adjacency matrix. Each node corresponding to an intersections or key topological node and edges representing centerlines. At each step, we describe a new node along with the its connected centerlines. We focus only on the topological relationships between the newly introduced node and the existing nodes. To facilitate better formatting into a sequence, we use the new node as the subject to describe its topology. The terms \"from\" and \"to\" respectively indicate which nodes the new node can come \"from\" and go \"to\". For example, when describing a new node $n$ we specify its position and connections to existing nodes, corresponding to the non-zero elements in the $n$ -th row and $n$ -th column of the adjacency matrix. By expanding the adjacency matrix row by row and column by column, we generate subgraphs that are incrementally combined to form the complete graph." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BEV-free methods [1, 19, 32] eliminates view transformation by operating directly in 3D space, leveraging perspective view information for direct 3D lane detection and reconstruction.", + "bbox": [ + 89, + 508, + 485, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Language models for graph", + "text_level": 1, + "bbox": [ + 89, + 577, + 339, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language models (LMs) have exhibited impressive capabilities not only in processing pure text but also across diverse application scenarios for autonomous driving [5, 31]. A significant challenge in graph fields is leveraging LMs to effectively understand and reason with graph information. How to encode graphs into learnable sequences according to specific rules is essential. Researchers [7, 20, 29] have demonstrated that language models possess substantial potential for learning from graphs. Lane graph, which are inherently graph-like with their interdependent nodes and edges, can significantly benefit from the contextual understanding and pattern recognition capabilities of language models. The integration of these models with graph structures shows great promise for advancing the comprehension of lane graph.", + "bbox": [ + 89, + 599, + 485, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Lane connectivity topology", + "text_level": 1, + "bbox": [ + 89, + 818, + 333, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lane topology understanding is a crucial aspect of comprehensive road comprehension. STSU [4] is the first to introduce this task, predicting lane centerlines and their connectivity. TopoNet [10] views the lane topology as a graph,", + "bbox": [ + 89, + 840, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "employing an end-to-end framework with a graph neural network architecture to learn the complete road structure. These methods treat the task of extracting centerlines and topology as two separate components, which can lead to a lack of coherence and misalignment between centerlines. LaneGAP [15] tackles these challenges by directly recognizing paths as units, accounting for existing overlapping centerlines between paths. However, this approach still requires stringent processing to remove duplicates. RNTR [17] and LaneGraph2Seq [21] view the lane graph as a DAG and establish a bidirectional transformation between the DAG and sequences, leveraging language models to generate sequences from PV images. RNTR [17] requires converting the DAG into a directed forest first, which involves complex preprocessing and post-processing steps, and the sequence information is redundant. On the other hand, LaneGraph2Seq [21] separates the learning of nodes and edges in the DAG by learning nodes first, followed by learning edges, which does not align with the language model's step-by-step reasoning approach. Addressing these challenges through a holistic approach could enhance the coherence and efficiency of lane graph modeling.", + "bbox": [ + 511, + 508, + 908, + 843 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 89, + 181, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we will introduce the mathematical modeling of the lane graph in Section 3.1, the method of serializing the lane graph in Section 3.2, and introduce the model architecture and objective in Section 3.3.", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Mathematical modeling of lane graph", + "text_level": 1, + "bbox": [ + 89, + 185, + 415, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The lane graph formed by centerlines can be modeled as a directed graph $G = (V, E)$ , where $V$ denotes a set of vertices and $E$ represents a set of edges. The vertices $V = \\{v_1, v_2, v_3, \\ldots\\}$ correspond to intersections or key topological nodes $\\{1, 2, 3, \\ldots\\}$ , where each vertex $v_n = (x_n, y_n)$ specifies its geographic position. The edges $E$ signify the lane lines connecting these intersections. To represent $E$ , we employ an adjacency matrix $A$ alongside an additional matrix $M$ . The adjacency matrix $A$ indicates the connectivity between the vertices and the matrix $M$ stores the shape of the centerlines that connect the intersections represented by the vertices.", + "bbox": [ + 89, + 207, + 483, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- $A$ : The adjacency matrix $A$ indicates the connectivity between the vertices. An entry $A[i][j]$ typically denotes whether there is a directed edge (directed centerline) from vertex $v_{i}$ to vertex $v_{j}$ without passing through other edges. Specifically,", + "bbox": [ + 89, + 392, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f t h e r e i s a e d g e b e t w e e n v _ {i} a n d v _ {j}} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 479, + 465, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since lane graph is directed, $A(i,j)\\neq A(j,i)$", + "bbox": [ + 102, + 530, + 410, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- $M$ : The matrix $M$ stores the shape of the centerlines that connect the intersections represented by the vertices. Specifically, $M(i,j)$ represents the centerline position from intersection $i$ to intersection $j$ , expressed using a quadratic Bézier function. This function includes the starting point $v_{i}$ and the endpoint $v_{j}$ , whose positions are already stored in $V$ , along with a middle control point at position $(\\sigma_x^{ij},\\sigma_y^{ij})$ . Therefore, we define", + "bbox": [ + 89, + 546, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM (i, j) = \\left\\{ \\begin{array}{l l} (\\sigma_ {x} ^ {i j}, \\sigma_ {y} ^ {i j}), & \\text {i f} A (i, j) = 1 \\\\ \\emptyset , & \\text {i f} A (i, j) = 0 \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 680, + 424, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Together, $G = (V, A, M)$ provide a comprehensive representation of the lane graph, encapsulating both the connectivity between intersections and the geometric details of the centerlines.", + "bbox": [ + 102, + 731, + 485, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Serializing the lane graph", + "text_level": 1, + "bbox": [ + 89, + 801, + 323, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the method of modeling lane graph as sequences. Previous research has primarily considered the lane graph as a Directed Acyclic Graph (DAG), taking into account that most roads are one-way and acyclic. However, real-world road infrastructure is complex, even", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the local map surrounding a vehicle, the road graph may contain loops and bidirectional single-lane roads, which are issues not considered by previous work. We propose a novel modeling approach called SeqGrowGraph.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously defined, $v_{n} = (x_{n},y_{n})$ denotes the position of intersection node $n$ , and $M$ stores the positions of the centerlines, with $M(i,j)$ representing the control point of the Bezier curve for the centerline between nodes $i$ and $j$ .", + "bbox": [ + 511, + 152, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define $S_{n}$ as the sequence that represents the subgraph formed by the first $n$ nodes. Let $F_{n}$ denote the set of all centerlines from which node $n$ can originate, and $T_{n}$ the set of all centerlines to which node $n$ can proceed. Here, the + symbol indicates sequence concatenation. Our approach grows as follows: $S_{n}$ is the concatenation of the preceding sequence and the new subsequences from node $n$ . The addition operation denotes sequence concatenation.", + "bbox": [ + 511, + 212, + 908, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {n} = S _ {n - 1} + \\left(v _ {n} + F _ {n} + T _ {n}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 358, + 812, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{n} = (x_{n},y_{n})$", + "bbox": [ + 633, + 378, + 785, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sets $F_{n}$ and $T_{n}$ , which represent the centerlines connecting the $n$ -th node to the preceding $n - 1$ nodes, are expressed as:", + "bbox": [ + 511, + 405, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n} = \\sum_ {k = 0} ^ {n - 1} M (k, n)\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 473, + 772, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {n} = \\sum_ {k = 0} ^ {n - 1} M (n, k)\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 517, + 772, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming there are a total of $N$ topological points, the final $S_N$ represents the ultimate sequence, which forms a complete description of the lane graph.", + "bbox": [ + 511, + 568, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 illustrates our step-by-step process for describing a graph as a sequence. This progressive expansion of the adjacency matrix and subgraph results in a comprehensive description of the lane graph. The left side of Figure 3 presents the incremental depiction of subgraphs, comprising six steps and corresponding subgraphs. In each step, we describe one node. As for the order in which the nodes should be described at each step, our method utilizes depth-first search.", + "bbox": [ + 511, + 614, + 906, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The right side of Figure 3 uses the introduction of $④$ as an example to demonstrate this incremental subgraph description.", + "bbox": [ + 511, + 750, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Before introducing ④ the adjacency matrix $A$ includes only the first 3 rows and 3 columns, capturing the existing topological information.", + "2. After specifying the position $x_{4}, y_{4}$ of node ④ we describe its topology by considering the first 4 rows and first 4 columns of matrix A. The non-zero elements in the 4-th row indicate the nodes to which ④ can directly go \"to\"; for" + ], + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27169", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/479f9785b9ecdcc1b27acde069e60e9bda8557194383b54836762e1a0a012b4d.jpg", + "image_caption": [ + "Figure 4. The demonstration of graph serialization by SeqGrowGraph. Each node corresponds to a sequence, and the sequences are connected by $$ , which together represent final lane graph." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 460, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "instance, $A(4,2) = 1$ signifies that ④ can go to ② without passing any other node. We use the notation 4_to_2 to represent the corresponding centerline. Conversely, the non-zero elements in the 4-th column indicate the nodes that ④ can come \"from\"; for instance, $A(2,4) = 1$ means that ④ can come directly from ② . We use 4_from_2 to denote the corresponding centerline.", + "bbox": [ + 89, + 410, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. The new sequence, formed by concatenating the previous sequence with the information of the current node, represents the subgraph of all nodes present so far. This process is repeated iteratively, ultimately yielding a final sequence that represents the complete lane graph.", + "bbox": [ + 89, + 517, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 illustrates the sequence structure of SeqGrowGraph. Each node's sequence includes its shape, connected node indices, and the Bezier control points of the connecting centerlines. By concatenating these sequences, we generate the target sequence. This serialized representation enables a more flexible expression of the topology, while maintaining clarity and conciseness and preserving essential graph-level information.", + "bbox": [ + 89, + 593, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the complete sequential representation for the lane graph has been generated, the next step is to express the positions and Bézier curve control points in a discretized form, excluding the special tokens and indices. This involves dividing the coordinates within the ego vehicle's system into discrete bins, allowing us to assign a corresponding integer number to each position. It is important to note that the range of special tokens is greater than the total number of bins.", + "bbox": [ + 89, + 714, + 483, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Model architecture and objective", + "text_level": 1, + "bbox": [ + 89, + 847, + 380, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture Following prior studies [17, 21], our method commences with the use of BEV encoder LSS [22] to project", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8609ec1b483d3fb1d1b2983030d72f574ffb13b96d99347c540b4fadfd62340c.jpg", + "image_caption": [ + "Figure 5. Model architecture and inference process. We employ a BEV-encoder to transform surrounding camera perspective view (PV) images into BEV features. Subsequently, a Transformer decoder generates tokens of the target sequence based on the BEV features and the beginning of sequence $< BOS>$ . All the tokens generated represent a complete lane graph." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 906, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features extracted from surrounding images onto the BEV plane. Following the conversion of the lane graph into a sequence, a transformer decoder is employed to generate the sequence token by token. The comprehensive architecture of our approach is depicted in Figure 5.", + "bbox": [ + 511, + 354, + 906, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objective We define the final target sequence $S = (x_{1}, x_{2}, \\ldots, x_{T})$ for each sample and have $\\hat{x}$ represent the predicted value. During the training phase, the loss function is expressed using maximum likelihood estimation as follows:", + "bbox": [ + 511, + 449, + 908, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log p (\\hat {x} _ {t} \\mid x _ {1}, x _ {2}, \\dots , x _ {t - 1})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 547, + 859, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss function measures the model's performance by taking the negative logarithm of the probability of predicting each subsequent step given the previous true token.", + "bbox": [ + 511, + 599, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 657, + 645, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 681, + 609, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method using two widely-used autonomous driving research datasets: nuScenes [3] and Argoverse 2 [27]. The nuScenes dataset comprises 1,000 scenes with six cameras offering a $360^{\\circ}$ field of view. Due to overlapping lanes in the default nuScenes split, PON [24] redistributes the train/validation sequences to eliminate overlap, allowing us to test our approach on both the default and PON splits. Consistent with past studies [17, 21], we input all surrounding camera images and Bird's Eye View (BEV) target ranges from $-48\\mathrm{m}$ to $48\\mathrm{m}$ on the x-axis and $-30\\mathrm{m}$ to $30\\mathrm{m}$ on the y-axis. The Argoverse 2 dataset contains 1,000 sequences and includes seven cameras, also providing a $360^{\\circ}$ field of view. For this dataset, we use camera images as input, with", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8cadc0d2971e68817db525f931b837b8cd5f410cdf37ff4baee49089d9ca85d5.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of our method with leading approaches. This evaluates Landmark Precision-Recall and Reachability Precision-Recall on the default and PON [24] split of nuScenes. Results marked with * are reproduced using the official code, while other results are obtained from their respective papers or released checkpoints." + ], + "table_body": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
DefaultTopoNet [10]52.547.149.646.910.817.5
LaneGAP [15]49.957.053.274.134.947.5
RNTR [17]57.142.748.963.745.252.8
LaneGraph2Seq* [21]46.943.745.263.736.346.2
SeqGrowGraph63.650.856.475.561.467.8
PONRNTR* [17]39.931.034.963.324.935.7
LaneGraph2Seq* [21]21.939.928.246.311.017.8
SeqGrowGraph43.533.337.763.636.946.7
", + "bbox": [ + 91, + 88, + 486, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the BEV target range from $-30\\mathrm{m}$ to $30\\mathrm{m}$ on the x-axis and $-15\\mathrm{m}$ to $15\\mathrm{m}$ on the y-axis.", + "bbox": [ + 88, + 375, + 482, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 89, + 419, + 250, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [21] We employ an LSS encoder [22] based on ResNet50 as the BEV encoder, initially training this encoder on a centerline segmentation task. For the sequence decoder, we use a 6-layer transformer. The training process is conducted on 4 NVIDIA A100 GPUs with batch size of $4 \\times 18$ . Specifically, the nuScenes dataset is trained for 400 epochs and the Argoverse 2 dataset is trained for 50 epochs.", + "bbox": [ + 89, + 443, + 483, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our sequence comprises the positional coordinates of nodes, node indices, and the positional coordinates of Bézier control points. We employ a discretization resolution of $0.5\\mathrm{m}$ with the node coordinates range set between 0 and 200. To distinguish the meanings of different tokens and to assign varying levels of importance, we introduce a shift to the node indices and Bézier control point coordinates. Consequently, the range for node indices extends from 200 to 350, while the range for Bézier point coordinates spans from 350 to 570. For the special tokens introduced in Figure 4, we omitted the representations for $< \\text{None}>$ and $< \\text{From}>$ to make the sequence more concise. And $< \\text{to}>$ , is denoted as 571, and $< \\text{SEP}>$ is represented as 572. Other special tokens, such as $< \\text{EOS}>$ (end token of sequence), $< \\text{BOS}>$ (beginning token of sequence), and $< \\text{N/A}>$ (padding token), are represented as 573, 574, and 575, respectively.", + "bbox": [ + 89, + 551, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We believe that the accuracy of edges relies on the correct prediction of points, so the positions of points should carry more weight than those of edges. Therefore, we designed our model such that the tokens representing point positions, which fall within the range of 0 to 200, have a weight of 2 in the cross-entropy loss. Meanwhile, the loss weight for the remaining tokens is set to 1.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ad140fcf610cd3d8733128da247588fd8ccc90e294b2d248bed33b67d67fbe97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
RNTR* [17]50.729.437.268.129.641.3
LaneGraph2Seq* [21]60.648.753.975.537.850.4
SeqGrowGraph64.150.456.477.743.455.7
", + "bbox": [ + 517, + 89, + 903, + 174 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a468edb501b44752fcb11dafb18b1ba5d390215b4fc73a540c0711b25c8e20b0.jpg", + "table_caption": [ + "Table 2. Comparison of our method with leading approaches on Argoverse 2, based on Landmark Precision-Recall and Reachability Precision-Recall. Results marked with * are reproduced using the official code." + ], + "table_footnote": [ + "Table 3. Comparison of our method with leading approaches on nuScenes, based on metrics including mean precision (M-P), recall (M-R), F1-score (M-F), and connectivity precision (C-P), recall (C-R), F1-score (C-F), and detection ratio. Results are obtained from the original paper." + ], + "table_body": "
MethodsM-PM-RM-FC-PC-RC-FDet.
LaneGraph2Seq [21]64.663.764.169.458.063.264.5
SeqGrowGraph71.270.270.676.258.065.867.2
", + "bbox": [ + 514, + 255, + 906, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Metric", + "text_level": 1, + "bbox": [ + 513, + 420, + 602, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure a fair comparison, we utilize the metrics proposed by RNTR [17], which encompass Landmark Precision-Recall and Reachability Precision-Recall. Landmark Precision-Recall evaluates the algorithm's accuracy in predicting the locations of key centerline points, such as segmented, fork, and merge points on lane lines. Meanwhile, Reachability Precision-Recall assesses the algorithm's capability to predict connectivity between these landmark nodes, indicating whether roads exist between them. Together, these metrics offer a comprehensive evaluation of the lane graph.", + "bbox": [ + 511, + 443, + 908, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we employ metrics introduced by STSU [4], including Mean Precision, Mean Recall, and Mean F1-score of centerlines, derived from precision-threshold and recall-threshold curves, alongside their F1-scores. These metrics focus on assessing the accuracy of centerline position matching, without accounting for missing predictions, thereby requiring the detection ratio to evaluate the presence of centerlines. Moreover, Connectivity Precision, Recall, and F1-score ratios evaluate the connectivity between centerlines, primarily focusing on centerline positions and not considering crucial landmark nodes, like junction points.", + "bbox": [ + 511, + 595, + 910, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison", + "text_level": 1, + "bbox": [ + 511, + 771, + 645, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach with state-of-the-art methods using the aforementioned metrics.", + "bbox": [ + 511, + 794, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares the results of our method with state-of-the-art methods on nuScenes using both the default and PON [24] splits. Our method outperforms existing algorithms in terms of F1-score for both landmarks and reachability. When there is no lane overlap between the training and", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/515278a6cd1ff0940d9ed5024b345aa43f739e16b743322c994e20fae652c86f.jpg", + "image_caption": [ + "Figure 6. Various definitions of centerline split have been considered in this study. We have re-segmented the original centerline from the nuScenes dataset to enhance its reasonableness and learning accessibility. Fixed-Length : The centerline between intersection points is partitioned at consistent intervals along the driving direction. In the provided example, these intervals are set at 20 meters and 30 meters. By refining the split using these methods, the learning process is facilitated, thereby yielding more accurate and interpretable results of lane graphs." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 460, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "test sets, as is the case with the PON split, the performance of all methods significantly decreases, but SeqGrowGraph remains in a leading position. Table 2 compares our method with leading methods on the larger dataset AV2 , demonstrating the superiority of our approach. Table 3 presents a comparison between our method and LaneGraph2Path concerning centerline-related metrics on nuScenes. It is evident that our method still outperforms the comparison method. This proves that modeling sequences in a manner similar to human language effectively harnesses the seq2seq model's capabilities in understanding and reasoning.", + "bbox": [ + 89, + 580, + 486, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Centerline re-segmentation", + "text_level": 1, + "bbox": [ + 89, + 757, + 336, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the original dataset, the division of map elements is based on map rules, but these division points often lack distinct visual features. For example, even a straight road might have multiple points that split a centerline into several segments, and these division points may not have any visual features that can be easily identified. Such data can disrupt the model's learning process and impact its ability to accurately interpret the rules. This issue is also noted in [25],", + "bbox": [ + 89, + 780, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0f76cc445c1029e97fa4c556f9d86cb19d2463357b5eb5af207f64fbdd3416d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Segment LengthJunction LandmarkJunction Reachability
L-PL-RL-FR-PR-RR-F
Original68.652.659.569.456.462.2
20m70.552.560.270.657.663.4
30m69.352.359.670.858.163.8
40m69.753.360.471.958.664.6
", + "bbox": [ + 514, + 89, + 911, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Training results under different segmentation ways. We resegment the centerline split to accurately and comprehensively represent the lane graph information by re-segmenting the centerlines between junctions based on length.", + "bbox": [ + 511, + 198, + 906, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and our tasks observe instability in the position of nodes as well. To address this, when a single topological node is referred to as a continuous node that only connects two centelines, we remove it and merge the centerlines. Our method fundamentally ensures that centerlines are segmented exclusively at junctions, including forks, and merge points, which constitute the topological nodes.", + "bbox": [ + 511, + 280, + 906, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When topological points are restricted to intersections, forks, and merges, the centerline may not be accurately represented by a quadratic Bezier function. To address this, we introduce additional topological points along each centerline and subsequently segment it. This segmentation is performed by dividing the centerline based on length. Figure 6 shows the examples when the intervals are set at 20 meters and 30 meters.", + "bbox": [ + 511, + 386, + 908, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance of our method under specific settings by adjusting the metrics to accommodate different ground truth standards. The evaluation framework remains aligned with Landmark Precision-Recall and Reachability Precision-Recall, while incorporating data post-processing. For both the ground truth and predictions, we remove continuous nodes and merge connected centerlines. This revised metric, termed Junction Landmark Precision-Recall and Junction Reachability Precision-Recall, emphasizes the accuracy of junction locations and the connectivity between junctions.", + "bbox": [ + 511, + 507, + 908, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents a comparison between the re-segmented centerline and the original ground truth. It can be observed that the metrics retrained after re-segmentation are all higher than those of the original split way. This suggests that we should adopt a similar standard when constructing ground truth for lane graphs, which could provide insights for research on lane graphs.", + "bbox": [ + 511, + 672, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Visualization", + "text_level": 1, + "bbox": [ + 511, + 787, + 651, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a visual comparison of the final results produced by various algorithms on the nuScenes dataset, as illustrated in Figure 7. This comparison reveals several limitations inherent in previous methods: TopoNet [10] often produces centerlines that are misaligned with each other, failing to provide a continuous centerline representation. LaneGAP [15]", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27172", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1d2e27242362a2dc773ce73e5ef3accb9fa0db10608b2efbb873bfcd8e57db4c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WeightLandmarkReachability
L-PL-RL-FR-PR-RR-F
162.648.654.774.460.266.5
263.650.856.475.561.467.8
362.549.655.375.461.367.6
463.050.255.975.361.267.5
", + "bbox": [ + 91, + 88, + 478, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tends to generate duplicate centerlines and may erroneously merge adjacent centerlines, leading to inefficiencies and inaccuracies. RNTR [17] demonstrates instability in its modeling capabilities, compromising the reliability of the resulting lane graphs. Furthermore, LaneGraph2Path [21] may result in a decoupling between points and edges, undermining the structural integrity of the graphs. In contrast, our algorithm not only provides stable and robust modeling but also effectively merges redundant points present in the ground truth data, thereby achieving superior results in constructing lane graphs. This advancement highlights the effectiveness and reliability of our approach in overcoming these challenges.", + "bbox": [ + 88, + 267, + 483, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 458, + 241, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation experiments on the nuScenes dataset, which includes examining the impact of loss weight on point positions and the order of subgraphs.", + "bbox": [ + 89, + 479, + 483, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Higher positional weight is applied to tokens ranging from 0 to 200 because these correspond to the positions of intersection points. Accurate descriptions of these intersection positions are vital for correctly depicting the topology and centerline locations between intersections. Table 5 illustrates that the modeling method remains relatively stable across different weights, with a slight improvement when the weight is set to 2, highlighting the robustness of the algorithm.", + "bbox": [ + 89, + 525, + 483, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The potential order in which each intersection point can be traversed to construct the corresponding subgraph is also investigated. In this context, Center refers to starting from the lane graph's center and proceeding from the closest to the farthest intersection based on their distance from the center. Coord denotes traversal based on the coordinates of the intersection points, specifically according to the x and y values. BFS stands for breadth-first search traversal, and DFS indicates depth-first search traversal. It can be observed that DFS performs the best, which suggests that depth-first search helps capture the hierarchical relationships between intersections.", + "bbox": [ + 89, + 661, + 483, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 845, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SeqGrowGraph, an innovative framework for autoregressive modeling of lane graphs that", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8d52ce30608c410f36bdf39cf0adecbbedb2e73816a6c5ac3d0808c584027d2.jpg", + "image_caption": [ + "Figure 7. We visualize and compare the final results of various algorithms on the nuScenes dataset." + ], + "image_footnote": [], + "bbox": [ + 539, + 90, + 888, + 415 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/129c13b057c3345029c4e5764a7223050fa85cc914ccb8e79a1feaf983c774b0.jpg", + "table_caption": [ + "Table 5. Ablation study on different loss weights assigned to token of point coordinates conducted using the nuScenes dataset. The weights for the remaining tokens are set to 1." + ], + "table_footnote": [], + "table_body": "
OrderLandmarkReachability
L-PL-RL-FR-PR-RR-F
Center59.047.152.473.356.363.7
Coord60.148.653.773.757.264.4
BFS62.649.255.175.361.467.7
DFS63.650.856.475.561.467.8
", + "bbox": [ + 514, + 472, + 903, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the ordering of subgraph descriptions conducted using the nuScenes dataset. Center pertains to initiating from the central point of the lane graph and navigating towards intersections in order of their increasing distance from this center. Coord involves traversing based on the coordinates of intersection points, specifically ordered by their x and y values. $BFS$ and $DFS$ refers to breadth-first traversal and depth-first traversal.", + "bbox": [ + 511, + 575, + 906, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "converts lane graphs into sequences via incremental local growth, facilitating a comprehensive global understanding. Our method enables flexible representation of complex road structures, including loops and bidirectional lanes, which pose challenges for traditional DAG-based approaches. Extensive experiments conducted on the nuScenes and Argoverse 2 datasets demonstrate superior performance in topological accuracy and graph completeness compared to existing detection-based and generation-based methods. SeqGrowGraph advances lane graph construction by bringing human-like structural reasoning and sequence modeling, thereby providing accurate centerlines for autonomous driving in dynamic environments.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27173", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In ICRA, 2023. 3", + "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 1", + "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5", + "[4] Yigit Baran Can, Alexander Liniger, Danda Pani Paudel, and Luc Van Gool. Structured bird's-eye-view traffic scene understanding from onboard images. In ICCV, 2021. 1, 3, 6", + "[5] Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map. In CVPR, 2025. 3", + "[6] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 1", + "[7] Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. Can llms effectively leverage graph structural information through prompts, and why? Trans. Mach. Learn. Res., 2024. 3", + "[8] Chunliang Li, Wencheng Han, Jun Yin, Sanyuan Zhao, and Jianbing Shen. Repvf: A unified vector fields representation for multi-task 3d perception. In ECCV, 2024. 1", + "[9] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 1, 2", + "[10] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based topology reasoning for driving scenes. arXiv preprint arXiv:2304.05277, 2023. 1, 2, 3, 6, 7", + "[11] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Grouplane: End-to-end 3d lane detection with channel-wise grouping. Robotics Autom. Lett., 2024. 2", + "[12] Shiyi Liang, Yifan Bai, Yihong Gong, and Xing Wei. Autoregressive sequential pretraining for visual tracking. In CVPR, 2025. 1", + "[13] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2", + "[14] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized HD map construction. CoRR, 2023. 2", + "[15] Benchcheng Liao, Shaoyu Chen, Bo Jiang, Tianheng Cheng, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Lane graph as path: Continuity-preserving path-wise modeling for online lane graph construction. In ECCV, 2024. 1, 2, 3, 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2", + "[17] Jiachen Lu, Hongyang Li, Renyuan Peng, Feng Wen, Xinyue Cai, Wei Zhang, Hang Xu, and Li Zhang. Translating images to road network: A non-autoregressive sequence to-sequence approach. In ICCV, 2023. 2, 3, 5, 6, 8", + "[18] Katie Z. Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In ICRA, 2024. 1", + "[19] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. LATR: 3d lane detection from monocular images with transformer. In ICCV, 2023. 3", + "[20] Yuxian Meng, Shi Zong, Xiaoya Li, Xiaofei Sun, Tianwei Zhang, Fei Wu, and Jiwei Li. GNN-LM: language modeling based on global contexts via GNN. In ICLR, 2022. 3", + "[21] Renyuan Peng, Xinyue Cai, Hang Xu, Jiachen Lu, Feng Wen, Wei Zhang, and Li Zhang. Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement. In AAAI, 2024. 2, 3, 5, 6, 8", + "[22] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 5, 6", + "[23] Maximilian Pittner, Joel Janai, and Alexandru Paul Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In CVPR, 2024. 1, 2", + "[24] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In CVPR, 2020. 5, 6", + "[25] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 7", + "[26] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1", + "[27] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 5", + "[28] Chengtang Yao, Lidong Yu, Yuwei Wu, and Yunde Jia. Sparse point guided 3d lane detection. In ICCV, 2023. 2", + "[29] Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In *Findings of the Association for Computational Linguistics: EACL 2024*, St. Julian's, Malta, March 17-22, 2024, 2024. 3", + "[30] Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2", + "[31] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27174", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In AAAI, 2024. 1, 3", + "[33] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized HD map construction. In CVPR, 2024. 2" + ], + "bbox": [ + 89, + 92, + 485, + 204 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27175", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_model.json b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2543cf4df94eb88d3d971d488e4d6b1529d19e73 --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.18, + 0.775, + 0.217 + ], + "angle": 0, + "content": "Mengwei Xie\\(^{1}\\), Shuang Zeng\\(^{1,2}\\), Xinyuan Chang\\(^{1}\\), Xinran Liu\\(^{1}\\), Zheng Pan\\(^{1}\\), Mu Xu\\(^{1}\\), Xing Wei\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.217, + 0.697, + 0.235 + ], + "angle": 0, + "content": "1Amap, Alibaba Group 2Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.237, + 0.862, + 0.253 + ], + "angle": 0, + "content": "{xiemengwei.xmw, changxinyuan.cxy, tom.lxr, panzheng.pan, xumu.xm}@alibaba-inc.com," + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.255, + 0.725, + 0.269 + ], + "angle": 0, + "content": "zengshuang@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.61 + ], + "angle": 0, + "content": "Accurate lane topology is essential for autonomous driving, yet traditional methods struggle to model the complex, non-linear structures—such as loops and bidirectional lanes—prevalent in real-world road structure. We present SeqGrowGraph, a novel framework that learns lane topology as a chain of graph expansions, inspired by human map-drawing processes. Representing the lane graph as a directed graph \\( G = (V,E) \\), with intersections \\( (V) \\) and centerlines \\( (E) \\), SeqGrowGraph incrementally constructs this graph by introducing one vertex at a time. At each step, an adjacency matrix \\( (A) \\) expands from \\( n \\times n \\) to \\( (n + 1) \\times (n + 1) \\) to encode connectivity, while a geometric matrix \\( (M) \\) captures centerline shapes as quadratic Bézier curves. The graph is serialized into sequences, enabling a transformer model to autoregressively predict the chain of expansions, guided by a depth-first search ordering. Evaluated on nuScenes and Argoverse 2 datasets, SeqGrowGraph achieves state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.639, + 0.222, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.799 + ], + "angle": 0, + "content": "The rise of autonomous driving has heightened the need for precise road structure modeling to ensure safe and efficient navigation [8, 23, 32]. Road structures encompass explicit features such as centerlines, road edges, and traffic signs, as well as implicit topological relationships, notably the connectivity between centerlines. Constructing lane graphs is crucial for representing these complex structures, facilitating vehicle decision-making and path planning in dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Traditional lane graph generation methods fall into detection-based and generation-based approaches, each with notable limitations. Detection-based methods can be categorized into three types: 1) Pixel-level approaches, such as HDMapNet [9], segment lane structures in bird's-eye view" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.304, + 0.887, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.582, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Figure 1. Describing the lane graph step by step. Starting from an initial point, it is necessary to determine the location of the next point and establish whether there is a \"from\" or \"to\" relationship with the existing points. By repetitively carrying out this process, a complete description of the lane graph can be obtained. The abbreviation \"pos\" in the diagram stands for \"position\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.907, + 0.819 + ], + "angle": 0, + "content": "(BEV) space on a per-pixel basis but struggle to infer global topology. 2) Piece-level methods [4, 10, 18] detect centerlines in fragmented segments based on topological points, often resulting in discontinuities and misalignments (Figure 2). 3) Path-level methods, such as LaneGAP [15], aim to mitigate discontinuities by detecting lane paths but introduce redundancies and alignment errors due to heavy reliance on post-processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Within the widely applied paradigm of sequence generation [2, 6, 12, 26], generative models are regarded as offering a more direct method for inferring lane graphs by capturing hidden structural patterns." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "However, formulating lane graphs as sequences poses" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.887, + 0.223, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "challenges and existing generative models have certain limitations. For instance, RNTR [17] converts lane graph into a Directed Acyclic Graph (DAG), which brings sequence redundancy in pre-processing and post-processing and fails to express loops and bidirectional lanes. LaneGraph2Seq [21], another generation-based method, generates the junction positions and shapes of centerlines separately, missing a holistic understanding of the network topology and struggling with integrated representation. And DAG-based methods fail to express loops and bidirectional lanes. Existing methodologies for graph-to-sequence conversion tend to excessively deconstruct the intrinsic graph structure, leading to transformations that disrupt the original topology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.426 + ], + "angle": 0, + "content": "Wait—how do humans construct lane graphs? Instead of processing an entire graph at once, people naturally start from a single node and progressively add new elements, continuously expanding the graph by establishing connections with previously defined nodes (Figure 1). This intuitive approach suggests a novel formulation: 1) Modeling lane graph construction as a chain of stepwise expansion process. 2) Using sequence modeling to capture incremental adjacency matrix updates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.485, + 0.548 + ], + "angle": 0, + "content": "Inspired by this observation, we propose SeqGrowGraph, a novel framework for incrementally constructing a lane Graph via Sequence Growth. SeqGrowGraph represents a lane graph as a directed graph \\( G = (V,E) \\), where \\( V \\) denotes intersections or key topological nodes, and \\( E \\) represents centerlines, defining connectivity between nodes. SeqGrowGraph models lane topology as a chain of graph expansions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.55, + 0.435, + 0.565 + ], + "angle": 0, + "content": "- A new node is introduced with its spatial position." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.565, + 0.483, + 0.626 + ], + "angle": 0, + "content": "- The adjacency matrix \\( A \\) expands from \\( n \\times n \\) to \\( (n + 1) \\times (n + 1) \\) to encode connectivity, with the upper triangle representing incoming edges (\"from\") and the lower triangle representing outgoing edges (\"to\")." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.626, + 0.484, + 0.655 + ], + "angle": 0, + "content": "- A geometric matrix \\(M\\) updates the centerline shape, represented as quadratic Bézier curves." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.656, + 0.484, + 0.701 + ], + "angle": 0, + "content": "- A transformer-based generative model autoregressively predicts these expansions using a depth-first search ordering." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.55, + 0.484, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.485, + 0.778 + ], + "angle": 0, + "content": "Unlike DAG-based approaches, SeqGrowGraph flexibly models complex, real-world road structures, including loops, bidirectional lanes, and non-trivial topologies, without excessive pre-processing or post-processing. Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.78, + 0.484, + 0.84 + ], + "angle": 0, + "content": "- We introduce SeqGrowGraph, a novel framework that reformulates lane graph generation as a sequential expansion process, leveraging adjacency matrix updates and geometric modeling." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We develop an autoregressive transformer-based model that progressively constructs lane graphs, overcoming the limitations of DAG-based methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We achieve state-of-the-art performance on large-scale" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.089, + 0.891, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Visualization comparison of lane graph results from different methods. This figure illustrates the inference results using the official models of TopoNet [10] and LaneGAP [15]. It can be observed that TopoNet, by detecting centerlines separately, results in discontinuities between the centerlines. In contrast, LaneGAP, which detects at the path level, tends to confuse multiple lanes and duplicate paths. Both methods rely heavily on extensive post-processing, failing to generate continuous centerline results in a single step." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.413, + 0.907, + 0.444 + ], + "angle": 0, + "content": "datasets (nuScenes, Argoverse 2), demonstrating superior topological accuracy and network completeness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.444, + 0.906, + 0.489 + ], + "angle": 0, + "content": "SeqGrowGraph offers a practical and effective solution for lane topology modeling, aligning closely with the natural human approach to map construction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.651, + 0.521 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.531, + 0.773, + 0.547 + ], + "angle": 0, + "content": "2.1. Online HD map construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Online high-definition map construction enables accurate road geometry modeling for autonomous navigation, with contemporary approaches [9, 13, 14, 16, 30, 33] focusing on BEV-space localization of 2D map elements. Initial approaches conceptualized this task as a semantic segmentation problem, as demonstrated in foundational works like HDMapNet [9], which necessitated computationally intensive post-processing procedures to convert pixel-level segmentation outputs into usable vector formats. To address this limitation, VectorMapNet [16] has focused on developing end-to-end frameworks for vectorized map learning. This approach employs a two-phase hierarchical architecture utilizing topological keypoint detection, implementing an initial coarse prediction stage followed by geometric refinement processes to achieve precise vectorized outputs. Notably, MapTR [13] achieves competitive performance using only PV image inputs, eliminating LiDAR dependency. 3D lane detection in autonomous driving follows two primary paradigms: BEV-based and BEV-free approaches. BEV-based methods [11, 23, 28] transform perspective view (PV) images into bird's-eye view (BEV) representations using camera intrinsics and ground coordinates to extract height-aware BEV features for accurate 3D lane modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27167" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.904, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.377, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Figure 3. Expansion process of the adjacency matrix. Each node corresponding to an intersections or key topological node and edges representing centerlines. At each step, we describe a new node along with the its connected centerlines. We focus only on the topological relationships between the newly introduced node and the existing nodes. To facilitate better formatting into a sequence, we use the new node as the subject to describe its topology. The terms \"from\" and \"to\" respectively indicate which nodes the new node can come \"from\" and go \"to\". For example, when describing a new node \\( n \\) we specify its position and connections to existing nodes, corresponding to the non-zero elements in the \\( n \\)-th row and \\( n \\)-th column of the adjacency matrix. By expanding the adjacency matrix row by row and column by column, we generate subgraphs that are incrementally combined to form the complete graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.486, + 0.569 + ], + "angle": 0, + "content": "BEV-free methods [1, 19, 32] eliminates view transformation by operating directly in 3D space, leveraging perspective view information for direct 3D lane detection and reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.34, + 0.594 + ], + "angle": 0, + "content": "2.2. Language models for graph" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.812 + ], + "angle": 0, + "content": "Language models (LMs) have exhibited impressive capabilities not only in processing pure text but also across diverse application scenarios for autonomous driving [5, 31]. A significant challenge in graph fields is leveraging LMs to effectively understand and reason with graph information. How to encode graphs into learnable sequences according to specific rules is essential. Researchers [7, 20, 29] have demonstrated that language models possess substantial potential for learning from graphs. Lane graph, which are inherently graph-like with their interdependent nodes and edges, can significantly benefit from the contextual understanding and pattern recognition capabilities of language models. The integration of these models with graph structures shows great promise for advancing the comprehension of lane graph." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.334, + 0.835 + ], + "angle": 0, + "content": "2.3. Lane connectivity topology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Lane topology understanding is a crucial aspect of comprehensive road comprehension. STSU [4] is the first to introduce this task, predicting lane centerlines and their connectivity. TopoNet [10] views the lane topology as a graph," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.844 + ], + "angle": 0, + "content": "employing an end-to-end framework with a graph neural network architecture to learn the complete road structure. These methods treat the task of extracting centerlines and topology as two separate components, which can lead to a lack of coherence and misalignment between centerlines. LaneGAP [15] tackles these challenges by directly recognizing paths as units, accounting for existing overlapping centerlines between paths. However, this approach still requires stringent processing to remove duplicates. RNTR [17] and LaneGraph2Seq [21] view the lane graph as a DAG and establish a bidirectional transformation between the DAG and sequences, leveraging language models to generate sequences from PV images. RNTR [17] requires converting the DAG into a directed forest first, which involves complex preprocessing and post-processing steps, and the sequence information is redundant. On the other hand, LaneGraph2Seq [21] separates the learning of nodes and edges in the DAG by learning nodes first, followed by learning edges, which does not align with the language model's step-by-step reasoning approach. Addressing these challenges through a holistic approach could enhance the coherence and efficiency of lane graph modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27168" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.182, + 0.106 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.485, + 0.176 + ], + "angle": 0, + "content": "In this section, we will introduce the mathematical modeling of the lane graph in Section 3.1, the method of serializing the lane graph in Section 3.2, and introduce the model architecture and objective in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.186, + 0.416, + 0.203 + ], + "angle": 0, + "content": "3.1. Mathematical modeling of lane graph" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.388 + ], + "angle": 0, + "content": "The lane graph formed by centerlines can be modeled as a directed graph \\( G = (V, E) \\), where \\( V \\) denotes a set of vertices and \\( E \\) represents a set of edges. The vertices \\( V = \\{v_1, v_2, v_3, \\ldots\\} \\) correspond to intersections or key topological nodes \\(\\{1, 2, 3, \\ldots\\}\\), where each vertex \\( v_n = (x_n, y_n) \\) specifies its geographic position. The edges \\( E \\) signify the lane lines connecting these intersections. To represent \\( E \\), we employ an adjacency matrix \\( A \\) alongside an additional matrix \\( M \\). The adjacency matrix \\( A \\) indicates the connectivity between the vertices and the matrix \\( M \\) stores the shape of the centerlines that connect the intersections represented by the vertices." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.393, + 0.484, + 0.469 + ], + "angle": 0, + "content": "- \\(A\\): The adjacency matrix \\(A\\) indicates the connectivity between the vertices. An entry \\(A[i][j]\\) typically denotes whether there is a directed edge (directed centerline) from vertex \\(v_{i}\\) to vertex \\(v_{j}\\) without passing through other edges. Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.48, + 0.466, + 0.52 + ], + "angle": 0, + "content": "\\[\nA (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f t h e r e i s a e d g e b e t w e e n v _ {i} a n d v _ {j}} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.531, + 0.411, + 0.547 + ], + "angle": 0, + "content": "Since lane graph is directed, \\(A(i,j)\\neq A(j,i)\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.484, + 0.669 + ], + "angle": 0, + "content": "- \\(M\\): The matrix \\(M\\) stores the shape of the centerlines that connect the intersections represented by the vertices. Specifically, \\(M(i,j)\\) represents the centerline position from intersection \\(i\\) to intersection \\(j\\), expressed using a quadratic Bézier function. This function includes the starting point \\(v_{i}\\) and the endpoint \\(v_{j}\\), whose positions are already stored in \\(V\\), along with a middle control point at position \\((\\sigma_x^{ij},\\sigma_y^{ij})\\). Therefore, we define" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.681, + 0.426, + 0.722 + ], + "angle": 0, + "content": "\\[\nM (i, j) = \\left\\{ \\begin{array}{l l} (\\sigma_ {x} ^ {i j}, \\sigma_ {y} ^ {i j}), & \\text {i f} A (i, j) = 1 \\\\ \\emptyset , & \\text {i f} A (i, j) = 0 \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.732, + 0.486, + 0.793 + ], + "angle": 0, + "content": "Together, \\( G = (V, A, M) \\) provide a comprehensive representation of the lane graph, encapsulating both the connectivity between intersections and the geometric details of the centerlines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.325, + 0.82 + ], + "angle": 0, + "content": "3.2. Serializing the lane graph" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the method of modeling lane graph as sequences. Previous research has primarily considered the lane graph as a Directed Acyclic Graph (DAG), taking into account that most roads are one-way and acyclic. However, real-world road infrastructure is complex, even" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "in the local map surrounding a vehicle, the road graph may contain loops and bidirectional single-lane roads, which are issues not considered by previous work. We propose a novel modeling approach called SeqGrowGraph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.212 + ], + "angle": 0, + "content": "As previously defined, \\( v_{n} = (x_{n},y_{n}) \\) denotes the position of intersection node \\( n \\), and \\( M \\) stores the positions of the centerlines, with \\( M(i,j) \\) representing the control point of the Bezier curve for the centerline between nodes \\( i \\) and \\( j \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.909, + 0.333 + ], + "angle": 0, + "content": "We define \\( S_{n} \\) as the sequence that represents the subgraph formed by the first \\( n \\) nodes. Let \\( F_{n} \\) denote the set of all centerlines from which node \\( n \\) can originate, and \\( T_{n} \\) the set of all centerlines to which node \\( n \\) can proceed. Here, the + symbol indicates sequence concatenation. Our approach grows as follows: \\( S_{n} \\) is the concatenation of the preceding sequence and the new subsequences from node \\( n \\). The addition operation denotes sequence concatenation." + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.359, + 0.813, + 0.376 + ], + "angle": 0, + "content": "\\[\nS _ {n} = S _ {n - 1} + \\left(v _ {n} + F _ {n} + T _ {n}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.634, + 0.379, + 0.786, + 0.395 + ], + "angle": 0, + "content": "where \\(v_{n} = (x_{n},y_{n})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.908, + 0.451 + ], + "angle": 0, + "content": "The sets \\( F_{n} \\) and \\( T_{n} \\), which represent the centerlines connecting the \\( n \\)-th node to the preceding \\( n - 1 \\) nodes, are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.474, + 0.773, + 0.515 + ], + "angle": 0, + "content": "\\[\nF _ {n} = \\sum_ {k = 0} ^ {n - 1} M (k, n)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.518, + 0.774, + 0.559 + ], + "angle": 0, + "content": "\\[\nT _ {n} = \\sum_ {k = 0} ^ {n - 1} M (n, k)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.614 + ], + "angle": 0, + "content": "Assuming there are a total of \\(N\\) topological points, the final \\(S_N\\) represents the ultimate sequence, which forms a complete description of the lane graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Figure 3 illustrates our step-by-step process for describing a graph as a sequence. This progressive expansion of the adjacency matrix and subgraph results in a comprehensive description of the lane graph. The left side of Figure 3 presents the incremental depiction of subgraphs, comprising six steps and corresponding subgraphs. In each step, we describe one node. As for the order in which the nodes should be described at each step, our method utilizes depth-first search." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.795 + ], + "angle": 0, + "content": "The right side of Figure 3 uses the introduction of \\(④\\) as an example to demonstrate this incremental subgraph description." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.84 + ], + "angle": 0, + "content": "1. Before introducing ④ the adjacency matrix \\( A \\) includes only the first 3 rows and 3 columns, capturing the existing topological information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "2. After specifying the position \\( x_{4}, y_{4} \\) of node ④ we describe its topology by considering the first 4 rows and first 4 columns of matrix A. The non-zero elements in the 4-th row indicate the nodes to which ④ can directly go \"to\"; for" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.462, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.485, + 0.384 + ], + "angle": 0, + "content": "Figure 4. The demonstration of graph serialization by SeqGrowGraph. Each node corresponds to a sequence, and the sequences are connected by \\(\\), which together represent final lane graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.517 + ], + "angle": 0, + "content": "instance, \\( A(4,2) = 1 \\) signifies that ④ can go to ② without passing any other node. We use the notation 4_to_2 to represent the corresponding centerline. Conversely, the non-zero elements in the 4-th column indicate the nodes that ④ can come \"from\"; for instance, \\( A(2,4) = 1 \\) means that ④ can come directly from ② . We use 4_from_2 to denote the corresponding centerline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.594 + ], + "angle": 0, + "content": "3. The new sequence, formed by concatenating the previous sequence with the information of the current node, represents the subgraph of all nodes present so far. This process is repeated iteratively, ultimately yielding a final sequence that represents the complete lane graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Figure 4 illustrates the sequence structure of SeqGrowGraph. Each node's sequence includes its shape, connected node indices, and the Bezier control points of the connecting centerlines. By concatenating these sequences, we generate the target sequence. This serialized representation enables a more flexible expression of the topology, while maintaining clarity and conciseness and preserving essential graph-level information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Once the complete sequential representation for the lane graph has been generated, the next step is to express the positions and Bézier curve control points in a discretized form, excluding the special tokens and indices. This involves dividing the coordinates within the ego vehicle's system into discrete bins, allowing us to assign a corresponding integer number to each position. It is important to note that the range of special tokens is greater than the total number of bins." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.382, + 0.865 + ], + "angle": 0, + "content": "3.3. Model architecture and objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture Following prior studies [17, 21], our method commences with the use of BEV encoder LSS [22] to project" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.907, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.247, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 5. Model architecture and inference process. We employ a BEV-encoder to transform surrounding camera perspective view (PV) images into BEV features. Subsequently, a Transformer decoder generates tokens of the target sequence based on the BEV features and the beginning of sequence \\(< BOS>\\). All the tokens generated represent a complete lane graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.433 + ], + "angle": 0, + "content": "features extracted from surrounding images onto the BEV plane. Following the conversion of the lane graph into a sequence, a transformer decoder is employed to generate the sequence token by token. The comprehensive architecture of our approach is depicted in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.525 + ], + "angle": 0, + "content": "Objective We define the final target sequence \\( S = (x_{1}, x_{2}, \\ldots, x_{T}) \\) for each sample and have \\( \\hat{x} \\) represent the predicted value. During the training phase, the loss function is expressed using maximum likelihood estimation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.548, + 0.861, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log p (\\hat {x} _ {t} \\mid x _ {1}, x _ {2}, \\dots , x _ {t - 1})\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.645 + ], + "angle": 0, + "content": "This loss function measures the model's performance by taking the negative logarithm of the probability of predicting each subsequent step given the previous true token." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.646, + 0.675 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.61, + 0.697 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method using two widely-used autonomous driving research datasets: nuScenes [3] and Argoverse 2 [27]. The nuScenes dataset comprises 1,000 scenes with six cameras offering a \\(360^{\\circ}\\) field of view. Due to overlapping lanes in the default nuScenes split, PON [24] redistributes the train/validation sequences to eliminate overlap, allowing us to test our approach on both the default and PON splits. Consistent with past studies [17, 21], we input all surrounding camera images and Bird's Eye View (BEV) target ranges from \\(-48\\mathrm{m}\\) to \\(48\\mathrm{m}\\) on the x-axis and \\(-30\\mathrm{m}\\) to \\(30\\mathrm{m}\\) on the y-axis. The Argoverse 2 dataset contains 1,000 sequences and includes seven cameras, also providing a \\(360^{\\circ}\\) field of view. For this dataset, we use camera images as input, with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27170" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.487, + 0.251 + ], + "angle": 0, + "content": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
DefaultTopoNet [10]52.547.149.646.910.817.5
LaneGAP [15]49.957.053.274.134.947.5
RNTR [17]57.142.748.963.745.252.8
LaneGraph2Seq* [21]46.943.745.263.736.346.2
SeqGrowGraph63.650.856.475.561.467.8
PONRNTR* [17]39.931.034.963.324.935.7
LaneGraph2Seq* [21]21.939.928.246.311.017.8
SeqGrowGraph43.533.337.763.636.946.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.345 + ], + "angle": 0, + "content": "Table 1. Comparison of our method with leading approaches. This evaluates Landmark Precision-Recall and Reachability Precision-Recall on the default and PON [24] split of nuScenes. Results marked with * are reproduced using the official code, while other results are obtained from their respective papers or released checkpoints." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.483, + 0.406 + ], + "angle": 0, + "content": "the BEV target range from \\(-30\\mathrm{m}\\) to \\(30\\mathrm{m}\\) on the x-axis and \\(-15\\mathrm{m}\\) to \\(15\\mathrm{m}\\) on the y-axis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.42, + 0.251, + 0.436 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.55 + ], + "angle": 0, + "content": "Following [21] We employ an LSS encoder [22] based on ResNet50 as the BEV encoder, initially training this encoder on a centerline segmentation task. For the sequence decoder, we use a 6-layer transformer. The training process is conducted on 4 NVIDIA A100 GPUs with batch size of \\(4 \\times 18\\). Specifically, the nuScenes dataset is trained for 400 epochs and the Argoverse 2 dataset is trained for 50 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Our sequence comprises the positional coordinates of nodes, node indices, and the positional coordinates of Bézier control points. We employ a discretization resolution of \\(0.5\\mathrm{m}\\) with the node coordinates range set between 0 and 200. To distinguish the meanings of different tokens and to assign varying levels of importance, we introduce a shift to the node indices and Bézier control point coordinates. Consequently, the range for node indices extends from 200 to 350, while the range for Bézier point coordinates spans from 350 to 570. For the special tokens introduced in Figure 4, we omitted the representations for \\(< \\text{None}>\\) and \\(< \\text{From}>\\) to make the sequence more concise. And \\(< \\text{to}>\\), is denoted as 571, and \\(< \\text{SEP}>\\) is represented as 572. Other special tokens, such as \\(< \\text{EOS}>\\) (end token of sequence), \\(< \\text{BOS}>\\) (beginning token of sequence), and \\(< \\text{N/A}>\\) (padding token), are represented as 573, 574, and 575, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We believe that the accuracy of edges relies on the correct prediction of points, so the positions of points should carry more weight than those of edges. Therefore, we designed our model such that the tokens representing point positions, which fall within the range of 0 to 200, have a weight of 2 in the cross-entropy loss. Meanwhile, the loss weight for the remaining tokens is set to 1." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.175 + ], + "angle": 0, + "content": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
RNTR* [17]50.729.437.268.129.641.3
LaneGraph2Seq* [21]60.648.753.975.537.850.4
SeqGrowGraph64.150.456.477.743.455.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.185, + 0.907, + 0.241 + ], + "angle": 0, + "content": "Table 2. Comparison of our method with leading approaches on Argoverse 2, based on Landmark Precision-Recall and Reachability Precision-Recall. Results marked with * are reproduced using the official code." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.256, + 0.907, + 0.315 + ], + "angle": 0, + "content": "
MethodsM-PM-RM-FC-PC-RC-FDet.
LaneGraph2Seq [21]64.663.764.169.458.063.264.5
SeqGrowGraph71.270.270.676.258.065.867.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.325, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Table 3. Comparison of our method with leading approaches on nuScenes, based on metrics including mean precision (M-P), recall (M-R), F1-score (M-F), and connectivity precision (C-P), recall (C-R), F1-score (C-F), and detection ratio. Results are obtained from the original paper." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.421, + 0.604, + 0.436 + ], + "angle": 0, + "content": "4.3. Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.909, + 0.595 + ], + "angle": 0, + "content": "To ensure a fair comparison, we utilize the metrics proposed by RNTR [17], which encompass Landmark Precision-Recall and Reachability Precision-Recall. Landmark Precision-Recall evaluates the algorithm's accuracy in predicting the locations of key centerline points, such as segmented, fork, and merge points on lane lines. Meanwhile, Reachability Precision-Recall assesses the algorithm's capability to predict connectivity between these landmark nodes, indicating whether roads exist between them. Together, these metrics offer a comprehensive evaluation of the lane graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.911, + 0.763 + ], + "angle": 0, + "content": "In addition, we employ metrics introduced by STSU [4], including Mean Precision, Mean Recall, and Mean F1-score of centerlines, derived from precision-threshold and recall-threshold curves, alongside their F1-scores. These metrics focus on assessing the accuracy of centerline position matching, without accounting for missing predictions, thereby requiring the detection ratio to evaluate the presence of centerlines. Moreover, Connectivity Precision, Recall, and F1-score ratios evaluate the connectivity between centerlines, primarily focusing on centerline positions and not considering crucial landmark nodes, like junction points." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.646, + 0.789 + ], + "angle": 0, + "content": "4.4. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.825 + ], + "angle": 0, + "content": "We compare our approach with state-of-the-art methods using the aforementioned metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Table 1 compares the results of our method with state-of-the-art methods on nuScenes using both the default and PON [24] splits. Our method outperforms existing algorithms in terms of F1-score for both landmarks and reachability. When there is no lane overlap between the training and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27171" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.09, + 0.462, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.43, + 0.487, + 0.556 + ], + "angle": 0, + "content": "Figure 6. Various definitions of centerline split have been considered in this study. We have re-segmented the original centerline from the nuScenes dataset to enhance its reasonableness and learning accessibility. Fixed-Length : The centerline between intersection points is partitioned at consistent intervals along the driving direction. In the provided example, these intervals are set at 20 meters and 30 meters. By refining the split using these methods, the learning process is facilitated, thereby yielding more accurate and interpretable results of lane graphs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.487, + 0.75 + ], + "angle": 0, + "content": "test sets, as is the case with the PON split, the performance of all methods significantly decreases, but SeqGrowGraph remains in a leading position. Table 2 compares our method with leading methods on the larger dataset AV2 , demonstrating the superiority of our approach. Table 3 presents a comparison between our method and LaneGraph2Path concerning centerline-related metrics on nuScenes. It is evident that our method still outperforms the comparison method. This proves that modeling sequences in a manner similar to human language effectively harnesses the seq2seq model's capabilities in understanding and reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.338, + 0.775 + ], + "angle": 0, + "content": "4.5. Centerline re-segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In the original dataset, the division of map elements is based on map rules, but these division points often lack distinct visual features. For example, even a straight road might have multiple points that split a centerline into several segments, and these division points may not have any visual features that can be easily identified. Such data can disrupt the model's learning process and impact its ability to accurately interpret the rules. This issue is also noted in [25]," + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.912, + 0.189 + ], + "angle": 0, + "content": "
Segment LengthJunction LandmarkJunction Reachability
L-PL-RL-FR-PR-RR-F
Original68.652.659.569.456.462.2
20m70.552.560.270.657.663.4
30m69.352.359.670.858.163.8
40m69.753.360.471.958.664.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.199, + 0.908, + 0.256 + ], + "angle": 0, + "content": "Table 4. Training results under different segmentation ways. We resegment the centerline split to accurately and comprehensively represent the lane graph information by re-segmenting the centerlines between junctions based on length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.386 + ], + "angle": 0, + "content": "and our tasks observe instability in the position of nodes as well. To address this, when a single topological node is referred to as a continuous node that only connects two centelines, we remove it and merge the centerlines. Our method fundamentally ensures that centerlines are segmented exclusively at junctions, including forks, and merge points, which constitute the topological nodes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.506 + ], + "angle": 0, + "content": "When topological points are restricted to intersections, forks, and merges, the centerline may not be accurately represented by a quadratic Bezier function. To address this, we introduce additional topological points along each centerline and subsequently segment it. This segmentation is performed by dividing the centerline based on length. Figure 6 shows the examples when the intervals are set at 20 meters and 30 meters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.674 + ], + "angle": 0, + "content": "We evaluate the performance of our method under specific settings by adjusting the metrics to accommodate different ground truth standards. The evaluation framework remains aligned with Landmark Precision-Recall and Reachability Precision-Recall, while incorporating data post-processing. For both the ground truth and predictions, we remove continuous nodes and merge connected centerlines. This revised metric, termed Junction Landmark Precision-Recall and Junction Reachability Precision-Recall, emphasizes the accuracy of junction locations and the connectivity between junctions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.91, + 0.78 + ], + "angle": 0, + "content": "Table 4 presents a comparison between the re-segmented centerline and the original ground truth. It can be observed that the metrics retrained after re-segmentation are all higher than those of the original split way. This suggests that we should adopt a similar standard when constructing ground truth for lane graphs, which could provide insights for research on lane graphs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.652, + 0.803 + ], + "angle": 0, + "content": "4.6. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We conduct a visual comparison of the final results produced by various algorithms on the nuScenes dataset, as illustrated in Figure 7. This comparison reveals several limitations inherent in previous methods: TopoNet [10] often produces centerlines that are misaligned with each other, failing to provide a continuous centerline representation. LaneGAP [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27172" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.48, + 0.189 + ], + "angle": 0, + "content": "
WeightLandmarkReachability
L-PL-RL-FR-PR-RR-F
162.648.654.774.460.266.5
263.650.856.475.561.467.8
362.549.655.375.461.367.6
463.050.255.975.361.267.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 5. Ablation study on different loss weights assigned to token of point coordinates conducted using the nuScenes dataset. The weights for the remaining tokens are set to 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.268, + 0.485, + 0.451 + ], + "angle": 0, + "content": "tends to generate duplicate centerlines and may erroneously merge adjacent centerlines, leading to inefficiencies and inaccuracies. RNTR [17] demonstrates instability in its modeling capabilities, compromising the reliability of the resulting lane graphs. Furthermore, LaneGraph2Path [21] may result in a decoupling between points and edges, undermining the structural integrity of the graphs. In contrast, our algorithm not only provides stable and robust modeling but also effectively merges redundant points present in the ground truth data, thereby achieving superior results in constructing lane graphs. This advancement highlights the effectiveness and reliability of our approach in overcoming these challenges." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.459, + 0.242, + 0.475 + ], + "angle": 0, + "content": "4.7. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.484, + 0.526 + ], + "angle": 0, + "content": "We conduct ablation experiments on the nuScenes dataset, which includes examining the impact of loss weight on point positions and the order of subgraphs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Higher positional weight is applied to tokens ranging from 0 to 200 because these correspond to the positions of intersection points. Accurate descriptions of these intersection positions are vital for correctly depicting the topology and centerline locations between intersections. Table 5 illustrates that the modeling method remains relatively stable across different weights, with a slight improvement when the weight is set to 2, highlighting the robustness of the algorithm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.484, + 0.842 + ], + "angle": 0, + "content": "The potential order in which each intersection point can be traversed to construct the corresponding subgraph is also investigated. In this context, Center refers to starting from the lane graph's center and proceeding from the closest to the farthest intersection based on their distance from the center. Coord denotes traversal based on the coordinates of the intersection points, specifically according to the x and y values. BFS stands for breadth-first search traversal, and DFS indicates depth-first search traversal. It can be observed that DFS performs the best, which suggests that depth-first search helps capture the hierarchical relationships between intersections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.21, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce SeqGrowGraph, an innovative framework for autoregressive modeling of lane graphs that" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.092, + 0.889, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 7. We visualize and compare the final results of various algorithms on the nuScenes dataset." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.473, + 0.905, + 0.567 + ], + "angle": 0, + "content": "
OrderLandmarkReachability
L-PL-RL-FR-PR-RR-F
Center59.047.152.473.356.363.7
Coord60.148.653.773.757.264.4
BFS62.649.255.175.361.467.7
DFS63.650.856.475.561.467.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.675 + ], + "angle": 0, + "content": "Table 6. Ablation study on the ordering of subgraph descriptions conducted using the nuScenes dataset. Center pertains to initiating from the central point of the lane graph and navigating towards intersections in order of their increasing distance from this center. Coord involves traversing based on the coordinates of intersection points, specifically ordered by their x and y values. \\(BFS\\) and \\(DFS\\) refers to breadth-first traversal and depth-first traversal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "converts lane graphs into sequences via incremental local growth, facilitating a comprehensive global understanding. Our method enables flexible representation of complex road structures, including loops and bidirectional lanes, which pose challenges for traditional DAG-based approaches. Extensive experiments conducted on the nuScenes and Argoverse 2 datasets demonstrate superior performance in topological accuracy and graph completeness compared to existing detection-based and generation-based methods. SeqGrowGraph advances lane graph construction by bringing human-like structural reasoning and sequence modeling, thereby providing accurate centerlines for autonomous driving in dynamic environments." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In ICRA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.211 + ], + "angle": 0, + "content": "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[4] Yigit Baran Can, Alexander Liniger, Danda Pani Paudel, and Luc Van Gool. Structured bird's-eye-view traffic scene understanding from onboard images. In ICCV, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[5] Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[6] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[7] Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. Can llms effectively leverage graph structural information through prompts, and why? Trans. Mach. Learn. Res., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[8] Chunliang Li, Wencheng Han, Jun Yin, Sanyuan Zhao, and Jianbing Shen. Repvf: A unified vector fields representation for multi-task 3d perception. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[9] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.552, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[10] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based topology reasoning for driving scenes. arXiv preprint arXiv:2304.05277, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[11] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Grouplane: End-to-end 3d lane detection with channel-wise grouping. Robotics Autom. Lett., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[12] Shiyi Liang, Yifan Bai, Yihong Gong, and Xing Wei. Autoregressive sequential pretraining for visual tracking. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[13] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[14] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized HD map construction. CoRR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[15] Benchcheng Liao, Shaoyu Chen, Bo Jiang, Tianheng Cheng, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Lane graph as path: Continuity-preserving path-wise modeling for online lane graph construction. In ECCV, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[17] Jiachen Lu, Hongyang Li, Renyuan Peng, Feng Wen, Xinyue Cai, Wei Zhang, Hang Xu, and Li Zhang. Translating images to road network: A non-autoregressive sequence to-sequence approach. In ICCV, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[18] Katie Z. Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In ICRA, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[19] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. LATR: 3d lane detection from monocular images with transformer. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[20] Yuxian Meng, Shi Zong, Xiaoya Li, Xiaofei Sun, Tianwei Zhang, Fei Wu, and Jiwei Li. GNN-LM: language modeling based on global contexts via GNN. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[21] Renyuan Peng, Xinyue Cai, Hang Xu, Jiachen Lu, Feng Wen, Wei Zhang, and Li Zhang. Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement. In AAAI, 2024. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.386, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[22] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[23] Maximilian Pittner, Joel Janai, and Alexandru Paul Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[24] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In CVPR, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[25] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[26] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[27] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[28] Chengtang Yao, Lidong Yu, Yuwei Wu, and Yunde Jia. Sparse point guided 3d lane detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[29] Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In *Findings of the Association for Computational Linguistics: EACL 2024*, St. Julian's, Malta, March 17-22, 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[31] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[32] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In AAAI, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[33] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized HD map construction. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.486, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27175" + } + ] +] \ No newline at end of file diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_origin.pdf b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6472a96c4a884c9731037404a298cf53f4da2db4 --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/f66ab5d4-0d27-477f-96dd-6cf261bca1fa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8dfc165348b81d539ac443d54dcbd02c9d2e708d3248d2e802fb6c3e325e7f0 +size 4859001 diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/full.md b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4897624a7d0a93e4128dec66f53aebb1598e5f23 --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/full.md @@ -0,0 +1,290 @@ +# SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions + +Mengwei Xie $^{1}$ , Shuang Zeng $^{1,2}$ , Xinyuan Chang $^{1}$ , Xinran Liu $^{1}$ , Zheng Pan $^{1}$ , Mu Xu $^{1}$ , Xing Wei $^{2\dagger}$ + +1Amap, Alibaba Group 2Xi'an Jiaotong University + +{xiemengwei.xmw, changxinyuan.cxy, tom.lxr, panzheng.pan, xumu.xm}@alibaba-inc.com, + +zengshuang@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn + +# Abstract + +Accurate lane topology is essential for autonomous driving, yet traditional methods struggle to model the complex, non-linear structures—such as loops and bidirectional lanes—prevalent in real-world road structure. We present SeqGrowGraph, a novel framework that learns lane topology as a chain of graph expansions, inspired by human map-drawing processes. Representing the lane graph as a directed graph $G = (V,E)$ , with intersections $(V)$ and centerlines $(E)$ , SeqGrowGraph incrementally constructs this graph by introducing one vertex at a time. At each step, an adjacency matrix $(A)$ expands from $n \times n$ to $(n + 1) \times (n + 1)$ to encode connectivity, while a geometric matrix $(M)$ captures centerline shapes as quadratic Bézier curves. The graph is serialized into sequences, enabling a transformer model to autoregressively predict the chain of expansions, guided by a depth-first search ordering. Evaluated on nuScenes and Argoverse 2 datasets, SeqGrowGraph achieves state-of-the-art performance. + +# 1. Introduction + +The rise of autonomous driving has heightened the need for precise road structure modeling to ensure safe and efficient navigation [8, 23, 32]. Road structures encompass explicit features such as centerlines, road edges, and traffic signs, as well as implicit topological relationships, notably the connectivity between centerlines. Constructing lane graphs is crucial for representing these complex structures, facilitating vehicle decision-making and path planning in dynamic environments. + +Traditional lane graph generation methods fall into detection-based and generation-based approaches, each with notable limitations. Detection-based methods can be categorized into three types: 1) Pixel-level approaches, such as HDMapNet [9], segment lane structures in bird's-eye view + +![](images/f2560af0b17c3d36e390afeaf88b6f8b20e0a32232999df17d4a092966ec9d5f.jpg) +Figure 1. Describing the lane graph step by step. Starting from an initial point, it is necessary to determine the location of the next point and establish whether there is a "from" or "to" relationship with the existing points. By repetitively carrying out this process, a complete description of the lane graph can be obtained. The abbreviation "pos" in the diagram stands for "position". + +(BEV) space on a per-pixel basis but struggle to infer global topology. 2) Piece-level methods [4, 10, 18] detect centerlines in fragmented segments based on topological points, often resulting in discontinuities and misalignments (Figure 2). 3) Path-level methods, such as LaneGAP [15], aim to mitigate discontinuities by detecting lane paths but introduce redundancies and alignment errors due to heavy reliance on post-processing. + +Within the widely applied paradigm of sequence generation [2, 6, 12, 26], generative models are regarded as offering a more direct method for inferring lane graphs by capturing hidden structural patterns. + +However, formulating lane graphs as sequences poses + +challenges and existing generative models have certain limitations. For instance, RNTR [17] converts lane graph into a Directed Acyclic Graph (DAG), which brings sequence redundancy in pre-processing and post-processing and fails to express loops and bidirectional lanes. LaneGraph2Seq [21], another generation-based method, generates the junction positions and shapes of centerlines separately, missing a holistic understanding of the network topology and struggling with integrated representation. And DAG-based methods fail to express loops and bidirectional lanes. Existing methodologies for graph-to-sequence conversion tend to excessively deconstruct the intrinsic graph structure, leading to transformations that disrupt the original topology. + +Wait—how do humans construct lane graphs? Instead of processing an entire graph at once, people naturally start from a single node and progressively add new elements, continuously expanding the graph by establishing connections with previously defined nodes (Figure 1). This intuitive approach suggests a novel formulation: 1) Modeling lane graph construction as a chain of stepwise expansion process. 2) Using sequence modeling to capture incremental adjacency matrix updates. + +Inspired by this observation, we propose SeqGrowGraph, a novel framework for incrementally constructing a lane Graph via Sequence Growth. SeqGrowGraph represents a lane graph as a directed graph $G = (V,E)$ , where $V$ denotes intersections or key topological nodes, and $E$ represents centerlines, defining connectivity between nodes. SeqGrowGraph models lane topology as a chain of graph expansions: + +- A new node is introduced with its spatial position. +- The adjacency matrix $A$ expands from $n \times n$ to $(n + 1) \times (n + 1)$ to encode connectivity, with the upper triangle representing incoming edges ("from") and the lower triangle representing outgoing edges ("to"). +- A geometric matrix $M$ updates the centerline shape, represented as quadratic Bézier curves. +- A transformer-based generative model autoregressively predicts these expansions using a depth-first search ordering. + +Unlike DAG-based approaches, SeqGrowGraph flexibly models complex, real-world road structures, including loops, bidirectional lanes, and non-trivial topologies, without excessive pre-processing or post-processing. Our key contributions are as follows: + +- We introduce SeqGrowGraph, a novel framework that reformulates lane graph generation as a sequential expansion process, leveraging adjacency matrix updates and geometric modeling. +- We develop an autoregressive transformer-based model that progressively constructs lane graphs, overcoming the limitations of DAG-based methods. +- We achieve state-of-the-art performance on large-scale + +![](images/be1ea35d80df5457e87ea0355480a07424a6d106e1e9e662aceed79bef43e1c6.jpg) +Figure 2. Visualization comparison of lane graph results from different methods. This figure illustrates the inference results using the official models of TopoNet [10] and LaneGAP [15]. It can be observed that TopoNet, by detecting centerlines separately, results in discontinuities between the centerlines. In contrast, LaneGAP, which detects at the path level, tends to confuse multiple lanes and duplicate paths. Both methods rely heavily on extensive post-processing, failing to generate continuous centerline results in a single step. + +datasets (nuScenes, Argoverse 2), demonstrating superior topological accuracy and network completeness. + +SeqGrowGraph offers a practical and effective solution for lane topology modeling, aligning closely with the natural human approach to map construction. + +# 2. Related work + +# 2.1. Online HD map construction + +Online high-definition map construction enables accurate road geometry modeling for autonomous navigation, with contemporary approaches [9, 13, 14, 16, 30, 33] focusing on BEV-space localization of 2D map elements. Initial approaches conceptualized this task as a semantic segmentation problem, as demonstrated in foundational works like HDMapNet [9], which necessitated computationally intensive post-processing procedures to convert pixel-level segmentation outputs into usable vector formats. To address this limitation, VectorMapNet [16] has focused on developing end-to-end frameworks for vectorized map learning. This approach employs a two-phase hierarchical architecture utilizing topological keypoint detection, implementing an initial coarse prediction stage followed by geometric refinement processes to achieve precise vectorized outputs. Notably, MapTR [13] achieves competitive performance using only PV image inputs, eliminating LiDAR dependency. 3D lane detection in autonomous driving follows two primary paradigms: BEV-based and BEV-free approaches. BEV-based methods [11, 23, 28] transform perspective view (PV) images into bird's-eye view (BEV) representations using camera intrinsics and ground coordinates to extract height-aware BEV features for accurate 3D lane modeling. + +![](images/09e7b0799a769dda9df849901e8e4e4575170b869d803d2b210e69a1d54f46f3.jpg) +Figure 3. Expansion process of the adjacency matrix. Each node corresponding to an intersections or key topological node and edges representing centerlines. At each step, we describe a new node along with the its connected centerlines. We focus only on the topological relationships between the newly introduced node and the existing nodes. To facilitate better formatting into a sequence, we use the new node as the subject to describe its topology. The terms "from" and "to" respectively indicate which nodes the new node can come "from" and go "to". For example, when describing a new node $n$ we specify its position and connections to existing nodes, corresponding to the non-zero elements in the $n$ -th row and $n$ -th column of the adjacency matrix. By expanding the adjacency matrix row by row and column by column, we generate subgraphs that are incrementally combined to form the complete graph. + +BEV-free methods [1, 19, 32] eliminates view transformation by operating directly in 3D space, leveraging perspective view information for direct 3D lane detection and reconstruction. + +# 2.2. Language models for graph + +Language models (LMs) have exhibited impressive capabilities not only in processing pure text but also across diverse application scenarios for autonomous driving [5, 31]. A significant challenge in graph fields is leveraging LMs to effectively understand and reason with graph information. How to encode graphs into learnable sequences according to specific rules is essential. Researchers [7, 20, 29] have demonstrated that language models possess substantial potential for learning from graphs. Lane graph, which are inherently graph-like with their interdependent nodes and edges, can significantly benefit from the contextual understanding and pattern recognition capabilities of language models. The integration of these models with graph structures shows great promise for advancing the comprehension of lane graph. + +# 2.3. Lane connectivity topology + +Lane topology understanding is a crucial aspect of comprehensive road comprehension. STSU [4] is the first to introduce this task, predicting lane centerlines and their connectivity. TopoNet [10] views the lane topology as a graph, + +employing an end-to-end framework with a graph neural network architecture to learn the complete road structure. These methods treat the task of extracting centerlines and topology as two separate components, which can lead to a lack of coherence and misalignment between centerlines. LaneGAP [15] tackles these challenges by directly recognizing paths as units, accounting for existing overlapping centerlines between paths. However, this approach still requires stringent processing to remove duplicates. RNTR [17] and LaneGraph2Seq [21] view the lane graph as a DAG and establish a bidirectional transformation between the DAG and sequences, leveraging language models to generate sequences from PV images. RNTR [17] requires converting the DAG into a directed forest first, which involves complex preprocessing and post-processing steps, and the sequence information is redundant. On the other hand, LaneGraph2Seq [21] separates the learning of nodes and edges in the DAG by learning nodes first, followed by learning edges, which does not align with the language model's step-by-step reasoning approach. Addressing these challenges through a holistic approach could enhance the coherence and efficiency of lane graph modeling. + +# 3. Method + +In this section, we will introduce the mathematical modeling of the lane graph in Section 3.1, the method of serializing the lane graph in Section 3.2, and introduce the model architecture and objective in Section 3.3. + +# 3.1. Mathematical modeling of lane graph + +The lane graph formed by centerlines can be modeled as a directed graph $G = (V, E)$ , where $V$ denotes a set of vertices and $E$ represents a set of edges. The vertices $V = \{v_1, v_2, v_3, \ldots\}$ correspond to intersections or key topological nodes $\{1, 2, 3, \ldots\}$ , where each vertex $v_n = (x_n, y_n)$ specifies its geographic position. The edges $E$ signify the lane lines connecting these intersections. To represent $E$ , we employ an adjacency matrix $A$ alongside an additional matrix $M$ . The adjacency matrix $A$ indicates the connectivity between the vertices and the matrix $M$ stores the shape of the centerlines that connect the intersections represented by the vertices. + +- $A$ : The adjacency matrix $A$ indicates the connectivity between the vertices. An entry $A[i][j]$ typically denotes whether there is a directed edge (directed centerline) from vertex $v_{i}$ to vertex $v_{j}$ without passing through other edges. Specifically, + +$$ +A (i, j) = \left\{ \begin{array}{l l} 1, & \text {i f t h e r e i s a e d g e b e t w e e n v _ {i} a n d v _ {j}} \\ 0, & \text {o t h e r w i s e} \end{array} \right. +$$ + +Since lane graph is directed, $A(i,j)\neq A(j,i)$ + +- $M$ : The matrix $M$ stores the shape of the centerlines that connect the intersections represented by the vertices. Specifically, $M(i,j)$ represents the centerline position from intersection $i$ to intersection $j$ , expressed using a quadratic Bézier function. This function includes the starting point $v_{i}$ and the endpoint $v_{j}$ , whose positions are already stored in $V$ , along with a middle control point at position $(\sigma_x^{ij},\sigma_y^{ij})$ . Therefore, we define + +$$ +M (i, j) = \left\{ \begin{array}{l l} (\sigma_ {x} ^ {i j}, \sigma_ {y} ^ {i j}), & \text {i f} A (i, j) = 1 \\ \emptyset , & \text {i f} A (i, j) = 0 \end{array} \right. +$$ + +Together, $G = (V, A, M)$ provide a comprehensive representation of the lane graph, encapsulating both the connectivity between intersections and the geometric details of the centerlines. + +# 3.2. Serializing the lane graph + +In this section, we introduce the method of modeling lane graph as sequences. Previous research has primarily considered the lane graph as a Directed Acyclic Graph (DAG), taking into account that most roads are one-way and acyclic. However, real-world road infrastructure is complex, even + +in the local map surrounding a vehicle, the road graph may contain loops and bidirectional single-lane roads, which are issues not considered by previous work. We propose a novel modeling approach called SeqGrowGraph. + +As previously defined, $v_{n} = (x_{n},y_{n})$ denotes the position of intersection node $n$ , and $M$ stores the positions of the centerlines, with $M(i,j)$ representing the control point of the Bezier curve for the centerline between nodes $i$ and $j$ . + +We define $S_{n}$ as the sequence that represents the subgraph formed by the first $n$ nodes. Let $F_{n}$ denote the set of all centerlines from which node $n$ can originate, and $T_{n}$ the set of all centerlines to which node $n$ can proceed. Here, the + symbol indicates sequence concatenation. Our approach grows as follows: $S_{n}$ is the concatenation of the preceding sequence and the new subsequences from node $n$ . The addition operation denotes sequence concatenation. + +$$ +S _ {n} = S _ {n - 1} + \left(v _ {n} + F _ {n} + T _ {n}\right) +$$ + +where $v_{n} = (x_{n},y_{n})$ + +The sets $F_{n}$ and $T_{n}$ , which represent the centerlines connecting the $n$ -th node to the preceding $n - 1$ nodes, are expressed as: + +$$ +F _ {n} = \sum_ {k = 0} ^ {n - 1} M (k, n) +$$ + +$$ +T _ {n} = \sum_ {k = 0} ^ {n - 1} M (n, k) +$$ + +Assuming there are a total of $N$ topological points, the final $S_N$ represents the ultimate sequence, which forms a complete description of the lane graph. + +Figure 3 illustrates our step-by-step process for describing a graph as a sequence. This progressive expansion of the adjacency matrix and subgraph results in a comprehensive description of the lane graph. The left side of Figure 3 presents the incremental depiction of subgraphs, comprising six steps and corresponding subgraphs. In each step, we describe one node. As for the order in which the nodes should be described at each step, our method utilizes depth-first search. + +The right side of Figure 3 uses the introduction of $④$ as an example to demonstrate this incremental subgraph description. + +1. Before introducing ④ the adjacency matrix $A$ includes only the first 3 rows and 3 columns, capturing the existing topological information. +2. After specifying the position $x_{4}, y_{4}$ of node ④ we describe its topology by considering the first 4 rows and first 4 columns of matrix A. The non-zero elements in the 4-th row indicate the nodes to which ④ can directly go "to"; for + +![](images/479f9785b9ecdcc1b27acde069e60e9bda8557194383b54836762e1a0a012b4d.jpg) +Figure 4. The demonstration of graph serialization by SeqGrowGraph. Each node corresponds to a sequence, and the sequences are connected by $$ , which together represent final lane graph. + +instance, $A(4,2) = 1$ signifies that ④ can go to ② without passing any other node. We use the notation 4_to_2 to represent the corresponding centerline. Conversely, the non-zero elements in the 4-th column indicate the nodes that ④ can come "from"; for instance, $A(2,4) = 1$ means that ④ can come directly from ② . We use 4_from_2 to denote the corresponding centerline. + +3. The new sequence, formed by concatenating the previous sequence with the information of the current node, represents the subgraph of all nodes present so far. This process is repeated iteratively, ultimately yielding a final sequence that represents the complete lane graph. + +Figure 4 illustrates the sequence structure of SeqGrowGraph. Each node's sequence includes its shape, connected node indices, and the Bezier control points of the connecting centerlines. By concatenating these sequences, we generate the target sequence. This serialized representation enables a more flexible expression of the topology, while maintaining clarity and conciseness and preserving essential graph-level information. + +Once the complete sequential representation for the lane graph has been generated, the next step is to express the positions and Bézier curve control points in a discretized form, excluding the special tokens and indices. This involves dividing the coordinates within the ego vehicle's system into discrete bins, allowing us to assign a corresponding integer number to each position. It is important to note that the range of special tokens is greater than the total number of bins. + +# 3.3. Model architecture and objective + +Architecture Following prior studies [17, 21], our method commences with the use of BEV encoder LSS [22] to project + +![](images/8609ec1b483d3fb1d1b2983030d72f574ffb13b96d99347c540b4fadfd62340c.jpg) +Figure 5. Model architecture and inference process. We employ a BEV-encoder to transform surrounding camera perspective view (PV) images into BEV features. Subsequently, a Transformer decoder generates tokens of the target sequence based on the BEV features and the beginning of sequence $< BOS>$ . All the tokens generated represent a complete lane graph. + +features extracted from surrounding images onto the BEV plane. Following the conversion of the lane graph into a sequence, a transformer decoder is employed to generate the sequence token by token. The comprehensive architecture of our approach is depicted in Figure 5. + +Objective We define the final target sequence $S = (x_{1}, x_{2}, \ldots, x_{T})$ for each sample and have $\hat{x}$ represent the predicted value. During the training phase, the loss function is expressed using maximum likelihood estimation as follows: + +$$ +\mathcal {L} = - \left(\frac {1}{T} \sum_ {t = 1} ^ {T} \log p (\hat {x} _ {t} \mid x _ {1}, x _ {2}, \dots , x _ {t - 1})\right) +$$ + +This loss function measures the model's performance by taking the negative logarithm of the probability of predicting each subsequent step given the previous true token. + +# 4. Experiments + +# 4.1. Dataset + +We evaluate our method using two widely-used autonomous driving research datasets: nuScenes [3] and Argoverse 2 [27]. The nuScenes dataset comprises 1,000 scenes with six cameras offering a $360^{\circ}$ field of view. Due to overlapping lanes in the default nuScenes split, PON [24] redistributes the train/validation sequences to eliminate overlap, allowing us to test our approach on both the default and PON splits. Consistent with past studies [17, 21], we input all surrounding camera images and Bird's Eye View (BEV) target ranges from $-48\mathrm{m}$ to $48\mathrm{m}$ on the x-axis and $-30\mathrm{m}$ to $30\mathrm{m}$ on the y-axis. The Argoverse 2 dataset contains 1,000 sequences and includes seven cameras, also providing a $360^{\circ}$ field of view. For this dataset, we use camera images as input, with + +
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
DefaultTopoNet [10]52.547.149.646.910.817.5
LaneGAP [15]49.957.053.274.134.947.5
RNTR [17]57.142.748.963.745.252.8
LaneGraph2Seq* [21]46.943.745.263.736.346.2
SeqGrowGraph63.650.856.475.561.467.8
PONRNTR* [17]39.931.034.963.324.935.7
LaneGraph2Seq* [21]21.939.928.246.311.017.8
SeqGrowGraph43.533.337.763.636.946.7
+ +Table 1. Comparison of our method with leading approaches. This evaluates Landmark Precision-Recall and Reachability Precision-Recall on the default and PON [24] split of nuScenes. Results marked with * are reproduced using the official code, while other results are obtained from their respective papers or released checkpoints. + +the BEV target range from $-30\mathrm{m}$ to $30\mathrm{m}$ on the x-axis and $-15\mathrm{m}$ to $15\mathrm{m}$ on the y-axis. + +# 4.2. Implementation + +Following [21] We employ an LSS encoder [22] based on ResNet50 as the BEV encoder, initially training this encoder on a centerline segmentation task. For the sequence decoder, we use a 6-layer transformer. The training process is conducted on 4 NVIDIA A100 GPUs with batch size of $4 \times 18$ . Specifically, the nuScenes dataset is trained for 400 epochs and the Argoverse 2 dataset is trained for 50 epochs. + +Our sequence comprises the positional coordinates of nodes, node indices, and the positional coordinates of Bézier control points. We employ a discretization resolution of $0.5\mathrm{m}$ with the node coordinates range set between 0 and 200. To distinguish the meanings of different tokens and to assign varying levels of importance, we introduce a shift to the node indices and Bézier control point coordinates. Consequently, the range for node indices extends from 200 to 350, while the range for Bézier point coordinates spans from 350 to 570. For the special tokens introduced in Figure 4, we omitted the representations for $< \text{None}>$ and $< \text{From}>$ to make the sequence more concise. And $< \text{to}>$ , is denoted as 571, and $< \text{SEP}>$ is represented as 572. Other special tokens, such as $< \text{EOS}>$ (end token of sequence), $< \text{BOS}>$ (beginning token of sequence), and $< \text{N/A}>$ (padding token), are represented as 573, 574, and 575, respectively. + +We believe that the accuracy of edges relies on the correct prediction of points, so the positions of points should carry more weight than those of edges. Therefore, we designed our model such that the tokens representing point positions, which fall within the range of 0 to 200, have a weight of 2 in the cross-entropy loss. Meanwhile, the loss weight for the remaining tokens is set to 1. + +
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
RNTR* [17]50.729.437.268.129.641.3
LaneGraph2Seq* [21]60.648.753.975.537.850.4
SeqGrowGraph64.150.456.477.743.455.7
+ +Table 2. Comparison of our method with leading approaches on Argoverse 2, based on Landmark Precision-Recall and Reachability Precision-Recall. Results marked with * are reproduced using the official code. + +
MethodsM-PM-RM-FC-PC-RC-FDet.
LaneGraph2Seq [21]64.663.764.169.458.063.264.5
SeqGrowGraph71.270.270.676.258.065.867.2
+ +Table 3. Comparison of our method with leading approaches on nuScenes, based on metrics including mean precision (M-P), recall (M-R), F1-score (M-F), and connectivity precision (C-P), recall (C-R), F1-score (C-F), and detection ratio. Results are obtained from the original paper. + +# 4.3. Metric + +To ensure a fair comparison, we utilize the metrics proposed by RNTR [17], which encompass Landmark Precision-Recall and Reachability Precision-Recall. Landmark Precision-Recall evaluates the algorithm's accuracy in predicting the locations of key centerline points, such as segmented, fork, and merge points on lane lines. Meanwhile, Reachability Precision-Recall assesses the algorithm's capability to predict connectivity between these landmark nodes, indicating whether roads exist between them. Together, these metrics offer a comprehensive evaluation of the lane graph. + +In addition, we employ metrics introduced by STSU [4], including Mean Precision, Mean Recall, and Mean F1-score of centerlines, derived from precision-threshold and recall-threshold curves, alongside their F1-scores. These metrics focus on assessing the accuracy of centerline position matching, without accounting for missing predictions, thereby requiring the detection ratio to evaluate the presence of centerlines. Moreover, Connectivity Precision, Recall, and F1-score ratios evaluate the connectivity between centerlines, primarily focusing on centerline positions and not considering crucial landmark nodes, like junction points. + +# 4.4. Comparison + +We compare our approach with state-of-the-art methods using the aforementioned metrics. + +Table 1 compares the results of our method with state-of-the-art methods on nuScenes using both the default and PON [24] splits. Our method outperforms existing algorithms in terms of F1-score for both landmarks and reachability. When there is no lane overlap between the training and + +![](images/515278a6cd1ff0940d9ed5024b345aa43f739e16b743322c994e20fae652c86f.jpg) +Figure 6. Various definitions of centerline split have been considered in this study. We have re-segmented the original centerline from the nuScenes dataset to enhance its reasonableness and learning accessibility. Fixed-Length : The centerline between intersection points is partitioned at consistent intervals along the driving direction. In the provided example, these intervals are set at 20 meters and 30 meters. By refining the split using these methods, the learning process is facilitated, thereby yielding more accurate and interpretable results of lane graphs. + +test sets, as is the case with the PON split, the performance of all methods significantly decreases, but SeqGrowGraph remains in a leading position. Table 2 compares our method with leading methods on the larger dataset AV2 , demonstrating the superiority of our approach. Table 3 presents a comparison between our method and LaneGraph2Path concerning centerline-related metrics on nuScenes. It is evident that our method still outperforms the comparison method. This proves that modeling sequences in a manner similar to human language effectively harnesses the seq2seq model's capabilities in understanding and reasoning. + +# 4.5. Centerline re-segmentation + +In the original dataset, the division of map elements is based on map rules, but these division points often lack distinct visual features. For example, even a straight road might have multiple points that split a centerline into several segments, and these division points may not have any visual features that can be easily identified. Such data can disrupt the model's learning process and impact its ability to accurately interpret the rules. This issue is also noted in [25], + +
Segment LengthJunction LandmarkJunction Reachability
L-PL-RL-FR-PR-RR-F
Original68.652.659.569.456.462.2
20m70.552.560.270.657.663.4
30m69.352.359.670.858.163.8
40m69.753.360.471.958.664.6
+ +Table 4. Training results under different segmentation ways. We resegment the centerline split to accurately and comprehensively represent the lane graph information by re-segmenting the centerlines between junctions based on length. + +and our tasks observe instability in the position of nodes as well. To address this, when a single topological node is referred to as a continuous node that only connects two centelines, we remove it and merge the centerlines. Our method fundamentally ensures that centerlines are segmented exclusively at junctions, including forks, and merge points, which constitute the topological nodes. + +When topological points are restricted to intersections, forks, and merges, the centerline may not be accurately represented by a quadratic Bezier function. To address this, we introduce additional topological points along each centerline and subsequently segment it. This segmentation is performed by dividing the centerline based on length. Figure 6 shows the examples when the intervals are set at 20 meters and 30 meters. + +We evaluate the performance of our method under specific settings by adjusting the metrics to accommodate different ground truth standards. The evaluation framework remains aligned with Landmark Precision-Recall and Reachability Precision-Recall, while incorporating data post-processing. For both the ground truth and predictions, we remove continuous nodes and merge connected centerlines. This revised metric, termed Junction Landmark Precision-Recall and Junction Reachability Precision-Recall, emphasizes the accuracy of junction locations and the connectivity between junctions. + +Table 4 presents a comparison between the re-segmented centerline and the original ground truth. It can be observed that the metrics retrained after re-segmentation are all higher than those of the original split way. This suggests that we should adopt a similar standard when constructing ground truth for lane graphs, which could provide insights for research on lane graphs. + +# 4.6. Visualization + +We conduct a visual comparison of the final results produced by various algorithms on the nuScenes dataset, as illustrated in Figure 7. This comparison reveals several limitations inherent in previous methods: TopoNet [10] often produces centerlines that are misaligned with each other, failing to provide a continuous centerline representation. LaneGAP [15] + +
WeightLandmarkReachability
L-PL-RL-FR-PR-RR-F
162.648.654.774.460.266.5
263.650.856.475.561.467.8
362.549.655.375.461.367.6
463.050.255.975.361.267.5
+ +tends to generate duplicate centerlines and may erroneously merge adjacent centerlines, leading to inefficiencies and inaccuracies. RNTR [17] demonstrates instability in its modeling capabilities, compromising the reliability of the resulting lane graphs. Furthermore, LaneGraph2Path [21] may result in a decoupling between points and edges, undermining the structural integrity of the graphs. In contrast, our algorithm not only provides stable and robust modeling but also effectively merges redundant points present in the ground truth data, thereby achieving superior results in constructing lane graphs. This advancement highlights the effectiveness and reliability of our approach in overcoming these challenges. + +# 4.7. Ablation study + +We conduct ablation experiments on the nuScenes dataset, which includes examining the impact of loss weight on point positions and the order of subgraphs. + +Higher positional weight is applied to tokens ranging from 0 to 200 because these correspond to the positions of intersection points. Accurate descriptions of these intersection positions are vital for correctly depicting the topology and centerline locations between intersections. Table 5 illustrates that the modeling method remains relatively stable across different weights, with a slight improvement when the weight is set to 2, highlighting the robustness of the algorithm. + +The potential order in which each intersection point can be traversed to construct the corresponding subgraph is also investigated. In this context, Center refers to starting from the lane graph's center and proceeding from the closest to the farthest intersection based on their distance from the center. Coord denotes traversal based on the coordinates of the intersection points, specifically according to the x and y values. BFS stands for breadth-first search traversal, and DFS indicates depth-first search traversal. It can be observed that DFS performs the best, which suggests that depth-first search helps capture the hierarchical relationships between intersections. + +# 5. Conclusion + +In this paper, we introduce SeqGrowGraph, an innovative framework for autoregressive modeling of lane graphs that + +![](images/b8d52ce30608c410f36bdf39cf0adecbbedb2e73816a6c5ac3d0808c584027d2.jpg) +Figure 7. We visualize and compare the final results of various algorithms on the nuScenes dataset. + +Table 5. Ablation study on different loss weights assigned to token of point coordinates conducted using the nuScenes dataset. The weights for the remaining tokens are set to 1. + +
OrderLandmarkReachability
L-PL-RL-FR-PR-RR-F
Center59.047.152.473.356.363.7
Coord60.148.653.773.757.264.4
BFS62.649.255.175.361.467.7
DFS63.650.856.475.561.467.8
+ +Table 6. Ablation study on the ordering of subgraph descriptions conducted using the nuScenes dataset. Center pertains to initiating from the central point of the lane graph and navigating towards intersections in order of their increasing distance from this center. Coord involves traversing based on the coordinates of intersection points, specifically ordered by their x and y values. $BFS$ and $DFS$ refers to breadth-first traversal and depth-first traversal. + +converts lane graphs into sequences via incremental local growth, facilitating a comprehensive global understanding. Our method enables flexible representation of complex road structures, including loops and bidirectional lanes, which pose challenges for traditional DAG-based approaches. Extensive experiments conducted on the nuScenes and Argoverse 2 datasets demonstrate superior performance in topological accuracy and graph completeness compared to existing detection-based and generation-based methods. SeqGrowGraph advances lane graph construction by bringing human-like structural reasoning and sequence modeling, thereby providing accurate centerlines for autonomous driving in dynamic environments. + +# References + +[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In ICRA, 2023. 3 +[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 1 +[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5 +[4] Yigit Baran Can, Alexander Liniger, Danda Pani Paudel, and Luc Van Gool. Structured bird's-eye-view traffic scene understanding from onboard images. In ICCV, 2021. 1, 3, 6 +[5] Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map. In CVPR, 2025. 3 +[6] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 1 +[7] Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. Can llms effectively leverage graph structural information through prompts, and why? Trans. Mach. Learn. Res., 2024. 3 +[8] Chunliang Li, Wencheng Han, Jun Yin, Sanyuan Zhao, and Jianbing Shen. Repvf: A unified vector fields representation for multi-task 3d perception. In ECCV, 2024. 1 +[9] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 1, 2 +[10] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based topology reasoning for driving scenes. arXiv preprint arXiv:2304.05277, 2023. 1, 2, 3, 6, 7 +[11] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Grouplane: End-to-end 3d lane detection with channel-wise grouping. Robotics Autom. Lett., 2024. 2 +[12] Shiyi Liang, Yifan Bai, Yihong Gong, and Xing Wei. Autoregressive sequential pretraining for visual tracking. In CVPR, 2025. 1 +[13] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2 +[14] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized HD map construction. CoRR, 2023. 2 +[15] Benchcheng Liao, Shaoyu Chen, Bo Jiang, Tianheng Cheng, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Lane graph as path: Continuity-preserving path-wise modeling for online lane graph construction. In ECCV, 2024. 1, 2, 3, 6, 7 + +[16] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2 +[17] Jiachen Lu, Hongyang Li, Renyuan Peng, Feng Wen, Xinyue Cai, Wei Zhang, Hang Xu, and Li Zhang. Translating images to road network: A non-autoregressive sequence to-sequence approach. In ICCV, 2023. 2, 3, 5, 6, 8 +[18] Katie Z. Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In ICRA, 2024. 1 +[19] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. LATR: 3d lane detection from monocular images with transformer. In ICCV, 2023. 3 +[20] Yuxian Meng, Shi Zong, Xiaoya Li, Xiaofei Sun, Tianwei Zhang, Fei Wu, and Jiwei Li. GNN-LM: language modeling based on global contexts via GNN. In ICLR, 2022. 3 +[21] Renyuan Peng, Xinyue Cai, Hang Xu, Jiachen Lu, Feng Wen, Wei Zhang, and Li Zhang. Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement. In AAAI, 2024. 2, 3, 5, 6, 8 +[22] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 5, 6 +[23] Maximilian Pittner, Joel Janai, and Alexandru Paul Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In CVPR, 2024. 1, 2 +[24] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In CVPR, 2020. 5, 6 +[25] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 7 +[26] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1 +[27] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 5 +[28] Chengtang Yao, Lidong Yu, Yuwei Wu, and Yunde Jia. Sparse point guided 3d lane detection. In ICCV, 2023. 2 +[29] Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In *Findings of the Association for Computational Linguistics: EACL 2024*, St. Julian's, Malta, March 17-22, 2024, 2024. 3 +[30] Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2 +[31] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025. 3 + +[32] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In AAAI, 2024. 1, 3 +[33] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized HD map construction. In CVPR, 2024. 2 \ No newline at end of file diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/images.zip b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a553e8083daec55ce6801d2c32c688871e34b33 --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b62d14a4451b6006275f0faeb5ce7c6ee99065b3db9cf60a403bfff3de210fa4 +size 530578 diff --git a/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/layout.json b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34b76b6ba247d76e680cef2713173a8b66903b3d --- /dev/null +++ b/2025/SeqGrowGraph_ Learning Lane Topology as a Chain of Graph Expansions/layout.json @@ -0,0 +1,8140 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "type": "text", + "content": "SeqGrowGraph: Learning Lane Topology as a Chain of Graph Expansions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "spans": [ + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": "Mengwei Xie" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Shuang Zeng" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Xinyuan Chang" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Xinran Liu" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Zheng Pan" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Mu Xu" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "text", + "content": ", Xing Wei" + }, + { + "bbox": [ + 136, + 142, + 474, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 171, + 426, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 171, + 426, + 186 + ], + "spans": [ + { + "bbox": [ + 175, + 171, + 426, + 186 + ], + "type": "text", + "content": "1Amap, Alibaba Group 2Xi'an Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 81, + 187, + 527, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 187, + 527, + 200 + ], + "spans": [ + { + "bbox": [ + 81, + 187, + 527, + 200 + ], + "type": "text", + "content": "{xiemengwei.xmw, changxinyuan.cxy, tom.lxr, panzheng.pan, xumu.xm}@alibaba-inc.com," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 161, + 201, + 443, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 201, + 443, + 213 + ], + "spans": [ + { + "bbox": [ + 161, + 201, + 443, + 213 + ], + "type": "text", + "content": "zengshuang@stu.xjtu.edu.cn, weixing@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": "Accurate lane topology is essential for autonomous driving, yet traditional methods struggle to model the complex, non-linear structures—such as loops and bidirectional lanes—prevalent in real-world road structure. We present SeqGrowGraph, a novel framework that learns lane topology as a chain of graph expansions, inspired by human map-drawing processes. Representing the lane graph as a directed graph " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "G = (V,E)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": ", with intersections " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": " and centerlines " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "(E)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": ", SeqGrowGraph incrementally constructs this graph by introducing one vertex at a time. At each step, an adjacency matrix " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "(A)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": " expands from " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "(n + 1) \\times (n + 1)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": " to encode connectivity, while a geometric matrix " + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 55, + 266, + 297, + 483 + ], + "type": "text", + "content": " captures centerline shapes as quadratic Bézier curves. The graph is serialized into sequences, enabling a transformer model to autoregressively predict the chain of expansions, guided by a depth-first search ordering. Evaluated on nuScenes and Argoverse 2 datasets, SeqGrowGraph achieves state-of-the-art performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 506, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 135, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 632 + ], + "type": "text", + "content": "The rise of autonomous driving has heightened the need for precise road structure modeling to ensure safe and efficient navigation [8, 23, 32]. Road structures encompass explicit features such as centerlines, road edges, and traffic signs, as well as implicit topological relationships, notably the connectivity between centerlines. Constructing lane graphs is crucial for representing these complex structures, facilitating vehicle decision-making and path planning in dynamic environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "content": "Traditional lane graph generation methods fall into detection-based and generation-based approaches, each with notable limitations. Detection-based methods can be categorized into three types: 1) Pixel-level approaches, such as HDMapNet [9], segment lane structures in bird's-eye view" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 240, + 542, + 452 + ], + "blocks": [ + { + "bbox": [ + 328, + 240, + 542, + 452 + ], + "lines": [ + { + "bbox": [ + 328, + 240, + 542, + 452 + ], + "spans": [ + { + "bbox": [ + 328, + 240, + 542, + 452 + ], + "type": "image", + "image_path": "f2560af0b17c3d36e390afeaf88b6f8b20e0a32232999df17d4a092966ec9d5f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 460, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 526 + ], + "type": "text", + "content": "Figure 1. Describing the lane graph step by step. Starting from an initial point, it is necessary to determine the location of the next point and establish whether there is a \"from\" or \"to\" relationship with the existing points. By repetitively carrying out this process, a complete description of the lane graph can be obtained. The abbreviation \"pos\" in the diagram stands for \"position\"." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 648 + ], + "type": "text", + "content": "(BEV) space on a per-pixel basis but struggle to infer global topology. 2) Piece-level methods [4, 10, 18] detect centerlines in fragmented segments based on topological points, often resulting in discontinuities and misalignments (Figure 2). 3) Path-level methods, such as LaneGAP [15], aim to mitigate discontinuities by detecting lane paths but introduce redundancies and alignment errors due to heavy reliance on post-processing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 651, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 699 + ], + "type": "text", + "content": "Within the widely applied paradigm of sequence generation [2, 6, 12, 26], generative models are regarded as offering a more direct method for inferring lane graphs by capturing hidden structural patterns." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "However, formulating lane graphs as sequences poses" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 702, + 136, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 136, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 136, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27166" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "challenges and existing generative models have certain limitations. For instance, RNTR [17] converts lane graph into a Directed Acyclic Graph (DAG), which brings sequence redundancy in pre-processing and post-processing and fails to express loops and bidirectional lanes. LaneGraph2Seq [21], another generation-based method, generates the junction positions and shapes of centerlines separately, missing a holistic understanding of the network topology and struggling with integrated representation. And DAG-based methods fail to express loops and bidirectional lanes. Existing methodologies for graph-to-sequence conversion tend to excessively deconstruct the intrinsic graph structure, leading to transformations that disrupt the original topology." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": "Wait—how do humans construct lane graphs? Instead of processing an entire graph at once, people naturally start from a single node and progressively add new elements, continuously expanding the graph by establishing connections with previously defined nodes (Figure 1). This intuitive approach suggests a novel formulation: 1) Modeling lane graph construction as a chain of stepwise expansion process. 2) Using sequence modeling to capture incremental adjacency matrix updates." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": "Inspired by this observation, we propose SeqGrowGraph, a novel framework for incrementally constructing a lane Graph via Sequence Growth. SeqGrowGraph represents a lane graph as a directed graph " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "G = (V,E)" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " denotes intersections or key topological nodes, and " + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 338, + 296, + 434 + ], + "type": "text", + "content": " represents centerlines, defining connectivity between nodes. SeqGrowGraph models lane topology as a chain of graph expansions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 296, + 555 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 435, + 266, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 266, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 266, + 447 + ], + "type": "text", + "content": "- A new node is introduced with its spatial position." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": "- The adjacency matrix " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " expands from " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "inline_equation", + "content": "(n + 1) \\times (n + 1)" + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": " to encode connectivity, with the upper triangle representing incoming edges (\"from\") and the lower triangle representing outgoing edges (\"to\")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 518 + ], + "type": "text", + "content": "- A geometric matrix " + }, + { + "bbox": [ + 55, + 495, + 296, + 518 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 495, + 296, + 518 + ], + "type": "text", + "content": " updates the centerline shape, represented as quadratic Bézier curves." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "type": "text", + "content": "- A transformer-based generative model autoregressively predicts these expansions using a depth-first search ordering." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": "Unlike DAG-based approaches, SeqGrowGraph flexibly models complex, real-world road structures, including loops, bidirectional lanes, and non-trivial topologies, without excessive pre-processing or post-processing. Our key contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": "- We introduce SeqGrowGraph, a novel framework that reformulates lane graph generation as a sequential expansion process, leveraging adjacency matrix updates and geometric modeling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 700 + ], + "type": "text", + "content": "- We develop an autoregressive transformer-based model that progressively constructs lane graphs, overcoming the limitations of DAG-based methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- We achieve state-of-the-art performance on large-scale" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 545, + 199 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 545, + 199 + ], + "type": "image", + "image_path": "be1ea35d80df5457e87ea0355480a07424a6d106e1e9e662aceed79bef43e1c6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Visualization comparison of lane graph results from different methods. This figure illustrates the inference results using the official models of TopoNet [10] and LaneGAP [15]. It can be observed that TopoNet, by detecting centerlines separately, results in discontinuities between the centerlines. In contrast, LaneGAP, which detects at the path level, tends to confuse multiple lanes and duplicate paths. Both methods rely heavily on extensive post-processing, failing to generate continuous centerline results in a single step." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 327, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 327, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 322, + 327, + 555, + 351 + ], + "type": "text", + "content": "datasets (nuScenes, Argoverse 2), demonstrating superior topological accuracy and network completeness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 351, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 387 + ], + "type": "text", + "content": "SeqGrowGraph offers a practical and effective solution for lane topology modeling, aligning closely with the natural human approach to map construction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 399, + 398, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 398, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 398, + 412 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 420, + 473, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 473, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 473, + 433 + ], + "type": "text", + "content": "2.1. Online HD map construction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 715 + ], + "type": "text", + "content": "Online high-definition map construction enables accurate road geometry modeling for autonomous navigation, with contemporary approaches [9, 13, 14, 16, 30, 33] focusing on BEV-space localization of 2D map elements. Initial approaches conceptualized this task as a semantic segmentation problem, as demonstrated in foundational works like HDMapNet [9], which necessitated computationally intensive post-processing procedures to convert pixel-level segmentation outputs into usable vector formats. To address this limitation, VectorMapNet [16] has focused on developing end-to-end frameworks for vectorized map learning. This approach employs a two-phase hierarchical architecture utilizing topological keypoint detection, implementing an initial coarse prediction stage followed by geometric refinement processes to achieve precise vectorized outputs. Notably, MapTR [13] achieves competitive performance using only PV image inputs, eliminating LiDAR dependency. 3D lane detection in autonomous driving follows two primary paradigms: BEV-based and BEV-free approaches. BEV-based methods [11, 23, 28] transform perspective view (PV) images into bird's-eye view (BEV) representations using camera intrinsics and ground coordinates to extract height-aware BEV features for accurate 3D lane modeling." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27167" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "type": "image", + "image_path": "09e7b0799a769dda9df849901e8e4e4575170b869d803d2b210e69a1d54f46f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "text", + "content": "Figure 3. Expansion process of the adjacency matrix. Each node corresponding to an intersections or key topological node and edges representing centerlines. At each step, we describe a new node along with the its connected centerlines. We focus only on the topological relationships between the newly introduced node and the existing nodes. To facilitate better formatting into a sequence, we use the new node as the subject to describe its topology. The terms \"from\" and \"to\" respectively indicate which nodes the new node can come \"from\" and go \"to\". For example, when describing a new node " + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "text", + "content": " we specify its position and connections to existing nodes, corresponding to the non-zero elements in the " + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 298, + 555, + 382 + ], + "type": "text", + "content": "-th column of the adjacency matrix. By expanding the adjacency matrix row by row and column by column, we generate subgraphs that are incrementally combined to form the complete graph." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 403, + 297, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 297, + 450 + ], + "type": "text", + "content": "BEV-free methods [1, 19, 32] eliminates view transformation by operating directly in 3D space, leveraging perspective view information for direct 3D lane detection and reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 457, + 208, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 208, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 208, + 470 + ], + "type": "text", + "content": "2.2. Language models for graph" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 643 + ], + "type": "text", + "content": "Language models (LMs) have exhibited impressive capabilities not only in processing pure text but also across diverse application scenarios for autonomous driving [5, 31]. A significant challenge in graph fields is leveraging LMs to effectively understand and reason with graph information. How to encode graphs into learnable sequences according to specific rules is essential. Researchers [7, 20, 29] have demonstrated that language models possess substantial potential for learning from graphs. Lane graph, which are inherently graph-like with their interdependent nodes and edges, can significantly benefit from the contextual understanding and pattern recognition capabilities of language models. The integration of these models with graph structures shows great promise for advancing the comprehension of lane graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 204, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 204, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 204, + 661 + ], + "type": "text", + "content": "2.3. Lane connectivity topology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "content": "Lane topology understanding is a crucial aspect of comprehensive road comprehension. STSU [4] is the first to introduce this task, predicting lane centerlines and their connectivity. TopoNet [10] views the lane topology as a graph," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 403, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 668 + ], + "type": "text", + "content": "employing an end-to-end framework with a graph neural network architecture to learn the complete road structure. These methods treat the task of extracting centerlines and topology as two separate components, which can lead to a lack of coherence and misalignment between centerlines. LaneGAP [15] tackles these challenges by directly recognizing paths as units, accounting for existing overlapping centerlines between paths. However, this approach still requires stringent processing to remove duplicates. RNTR [17] and LaneGraph2Seq [21] view the lane graph as a DAG and establish a bidirectional transformation between the DAG and sequences, leveraging language models to generate sequences from PV images. RNTR [17] requires converting the DAG into a directed forest first, which involves complex preprocessing and post-processing steps, and the sequence information is redundant. On the other hand, LaneGraph2Seq [21] separates the learning of nodes and edges in the DAG by learning nodes first, followed by learning edges, which does not align with the language model's step-by-step reasoning approach. Addressing these challenges through a holistic approach could enhance the coherence and efficiency of lane graph modeling." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27168" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "In this section, we will introduce the mathematical modeling of the lane graph in Section 3.1, the method of serializing the lane graph in Section 3.2, and introduce the model architecture and objective in Section 3.3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 254, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 254, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 254, + 160 + ], + "type": "text", + "content": "3.1. Mathematical modeling of lane graph" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": "The lane graph formed by centerlines can be modeled as a directed graph " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " denotes a set of vertices and " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " represents a set of edges. The vertices " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "V = \\{v_1, v_2, v_3, \\ldots\\}" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " correspond to intersections or key topological nodes " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\{1, 2, 3, \\ldots\\}" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": ", where each vertex " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "v_n = (x_n, y_n)" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " specifies its geographic position. The edges " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " signify the lane lines connecting these intersections. To represent " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": ", we employ an adjacency matrix " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " alongside an additional matrix " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": ". The adjacency matrix " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " indicates the connectivity between the vertices and the matrix " + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 164, + 296, + 307 + ], + "type": "text", + "content": " stores the shape of the centerlines that connect the intersections represented by the vertices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": ": The adjacency matrix " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " indicates the connectivity between the vertices. An entry " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "A[i][j]" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " typically denotes whether there is a directed edge (directed centerline) from vertex " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " to vertex " + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 311, + 296, + 371 + ], + "type": "text", + "content": " without passing through other edges. Specifically," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 380, + 285, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 380, + 285, + 411 + ], + "spans": [ + { + "bbox": [ + 73, + 380, + 285, + 411 + ], + "type": "interline_equation", + "content": "A (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f t h e r e i s a e d g e b e t w e e n v _ {i} a n d v _ {j}} \\\\ 0, & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "dd20e6b4354cd50d0bb4ab08f24dee185e156239e3c8dc87061bedadff37207b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 420, + 251, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 420, + 251, + 433 + ], + "spans": [ + { + "bbox": [ + 63, + 420, + 251, + 433 + ], + "type": "text", + "content": "Since lane graph is directed, " + }, + { + "bbox": [ + 63, + 420, + 251, + 433 + ], + "type": "inline_equation", + "content": "A(i,j)\\neq A(j,i)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": ": The matrix " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": " stores the shape of the centerlines that connect the intersections represented by the vertices. Specifically, " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "M(i,j)" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": " represents the centerline position from intersection " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": " to intersection " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": ", expressed using a quadratic Bézier function. This function includes the starting point " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": " and the endpoint " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "v_{j}" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": ", whose positions are already stored in " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": ", along with a middle control point at position " + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "inline_equation", + "content": "(\\sigma_x^{ij},\\sigma_y^{ij})" + }, + { + "bbox": [ + 55, + 433, + 296, + 529 + ], + "type": "text", + "content": ". Therefore, we define" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 539, + 260, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 539, + 260, + 571 + ], + "spans": [ + { + "bbox": [ + 98, + 539, + 260, + 571 + ], + "type": "interline_equation", + "content": "M (i, j) = \\left\\{ \\begin{array}{l l} (\\sigma_ {x} ^ {i j}, \\sigma_ {y} ^ {i j}), & \\text {i f} A (i, j) = 1 \\\\ \\emptyset , & \\text {i f} A (i, j) = 0 \\end{array} \\right.", + "image_path": "e6f17e2c3df91d20a8bfe68b8138081f0daf5cf6ee0d540810772b83ceb52a2e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 579, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 579, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 63, + 579, + 297, + 628 + ], + "type": "text", + "content": "Together, " + }, + { + "bbox": [ + 63, + 579, + 297, + 628 + ], + "type": "inline_equation", + "content": "G = (V, A, M)" + }, + { + "bbox": [ + 63, + 579, + 297, + 628 + ], + "type": "text", + "content": " provide a comprehensive representation of the lane graph, encapsulating both the connectivity between intersections and the geometric details of the centerlines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 198, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 198, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 198, + 649 + ], + "type": "text", + "content": "3.2. Serializing the lane graph" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In this section, we introduce the method of modeling lane graph as sequences. Previous research has primarily considered the lane graph as a Directed Acyclic Graph (DAG), taking into account that most roads are one-way and acyclic. However, real-world road infrastructure is complex, even" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "in the local map surrounding a vehicle, the road graph may contain loops and bidirectional single-lane roads, which are issues not considered by previous work. We propose a novel modeling approach called SeqGrowGraph." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": "As previously defined, " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "v_{n} = (x_{n},y_{n})" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": " denotes the position of intersection node " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": " stores the positions of the centerlines, with " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "M(i,j)" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": " representing the control point of the Bezier curve for the centerline between nodes " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 121, + 555, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " as the sequence that represents the subgraph formed by the first " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " nodes. Let " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " denote the set of all centerlines from which node " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " can originate, and " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " the set of all centerlines to which node " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " can proceed. Here, the + symbol indicates sequence concatenation. Our approach grows as follows: " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": " is the concatenation of the preceding sequence and the new subsequences from node " + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 168, + 556, + 263 + ], + "type": "text", + "content": ". The addition operation denotes sequence concatenation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 284, + 497, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 284, + 497, + 297 + ], + "spans": [ + { + "bbox": [ + 369, + 284, + 497, + 297 + ], + "type": "interline_equation", + "content": "S _ {n} = S _ {n - 1} + \\left(v _ {n} + F _ {n} + T _ {n}\\right)", + "image_path": "f54a398822421b40f3cc8adb0720eb71d2b52a1dc133bbce869ce953e24a9496.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 388, + 300, + 481, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 300, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 388, + 300, + 481, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 388, + 300, + 481, + 312 + ], + "type": "inline_equation", + "content": "v_{n} = (x_{n},y_{n})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": "The sets " + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "inline_equation", + "content": "F_{n}" + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": ", which represent the centerlines connecting the " + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": "-th node to the preceding " + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": " nodes, are expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 394, + 375, + 473, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 375, + 473, + 407 + ], + "spans": [ + { + "bbox": [ + 394, + 375, + 473, + 407 + ], + "type": "interline_equation", + "content": "F _ {n} = \\sum_ {k = 0} ^ {n - 1} M (k, n)", + "image_path": "128975dd7fda360f85530f9fbb773d91bd44267086a177987cec34e15d3ce489.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 394, + 410, + 473, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 410, + 473, + 442 + ], + "spans": [ + { + "bbox": [ + 394, + 410, + 473, + 442 + ], + "type": "interline_equation", + "content": "T _ {n} = \\sum_ {k = 0} ^ {n - 1} M (n, k)", + "image_path": "dc023a5a7fde14b327458c77eaca124ed59186cb4893b45ba20d654d28f68289.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "text", + "content": "Assuming there are a total of " + }, + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "text", + "content": " topological points, the final " + }, + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "inline_equation", + "content": "S_N" + }, + { + "bbox": [ + 313, + 450, + 554, + 486 + ], + "type": "text", + "content": " represents the ultimate sequence, which forms a complete description of the lane graph." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 487, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 593 + ], + "type": "text", + "content": "Figure 3 illustrates our step-by-step process for describing a graph as a sequence. This progressive expansion of the adjacency matrix and subgraph results in a comprehensive description of the lane graph. The left side of Figure 3 presents the incremental depiction of subgraphs, comprising six steps and corresponding subgraphs. In each step, we describe one node. As for the order in which the nodes should be described at each step, our method utilizes depth-first search." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 629 + ], + "type": "text", + "content": "The right side of Figure 3 uses the introduction of " + }, + { + "bbox": [ + 313, + 594, + 554, + 629 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 313, + 594, + 554, + 629 + ], + "type": "text", + "content": " as an example to demonstrate this incremental subgraph description." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 313, + 630, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 665 + ], + "type": "text", + "content": "1. Before introducing ④ the adjacency matrix " + }, + { + "bbox": [ + 313, + 630, + 553, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 630, + 553, + 665 + ], + "type": "text", + "content": " includes only the first 3 rows and 3 columns, capturing the existing topological information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "2. After specifying the position " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_{4}, y_{4}" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " of node ④ we describe its topology by considering the first 4 rows and first 4 columns of matrix A. The non-zero elements in the 4-th row indicate the nodes to which ④ can directly go \"to\"; for" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27169" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 282, + 250 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 282, + 250 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 282, + 250 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 282, + 250 + ], + "type": "image", + "image_path": "479f9785b9ecdcc1b27acde069e60e9bda8557194383b54836762e1a0a012b4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": "Figure 4. The demonstration of graph serialization by SeqGrowGraph. Each node corresponds to a sequence, and the sequences are connected by " + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": ", which together represent final lane graph." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "text", + "content": "instance, " + }, + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "inline_equation", + "content": "A(4,2) = 1" + }, + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "text", + "content": " signifies that ④ can go to ② without passing any other node. We use the notation 4_to_2 to represent the corresponding centerline. Conversely, the non-zero elements in the 4-th column indicate the nodes that ④ can come \"from\"; for instance, " + }, + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "inline_equation", + "content": "A(2,4) = 1" + }, + { + "bbox": [ + 55, + 325, + 296, + 409 + ], + "type": "text", + "content": " means that ④ can come directly from ② . We use 4_from_2 to denote the corresponding centerline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 470 + ], + "type": "text", + "content": "3. The new sequence, formed by concatenating the previous sequence with the information of the current node, represents the subgraph of all nodes present so far. This process is repeated iteratively, ultimately yielding a final sequence that represents the complete lane graph." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 564 + ], + "type": "text", + "content": "Figure 4 illustrates the sequence structure of SeqGrowGraph. Each node's sequence includes its shape, connected node indices, and the Bezier control points of the connecting centerlines. By concatenating these sequences, we generate the target sequence. This serialized representation enables a more flexible expression of the topology, while maintaining clarity and conciseness and preserving essential graph-level information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "content": "Once the complete sequential representation for the lane graph has been generated, the next step is to express the positions and Bézier curve control points in a discretized form, excluding the special tokens and indices. This involves dividing the coordinates within the ego vehicle's system into discrete bins, allowing us to assign a corresponding integer number to each position. It is important to note that the range of special tokens is greater than the total number of bins." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 233, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 233, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 233, + 685 + ], + "type": "text", + "content": "3.3. Model architecture and objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Architecture Following prior studies [17, 21], our method commences with the use of BEV encoder LSS [22] to project" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 555, + 187 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 555, + 187 + ], + "type": "image", + "image_path": "8609ec1b483d3fb1d1b2983030d72f574ffb13b96d99347c540b4fadfd62340c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 195, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 262 + ], + "type": "text", + "content": "Figure 5. Model architecture and inference process. We employ a BEV-encoder to transform surrounding camera perspective view (PV) images into BEV features. Subsequently, a Transformer decoder generates tokens of the target sequence based on the BEV features and the beginning of sequence " + }, + { + "bbox": [ + 313, + 195, + 555, + 262 + ], + "type": "inline_equation", + "content": "< BOS>" + }, + { + "bbox": [ + 313, + 195, + 555, + 262 + ], + "type": "text", + "content": ". All the tokens generated represent a complete lane graph." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 342 + ], + "type": "text", + "content": "features extracted from surrounding images onto the BEV plane. Following the conversion of the lane graph into a sequence, a transformer decoder is employed to generate the sequence token by token. The comprehensive architecture of our approach is depicted in Figure 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "text", + "content": "Objective We define the final target sequence " + }, + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "inline_equation", + "content": "S = (x_{1}, x_{2}, \\ldots, x_{T})" + }, + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "text", + "content": " for each sample and have " + }, + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 356, + 556, + 415 + ], + "type": "text", + "content": " represent the predicted value. During the training phase, the loss function is expressed using maximum likelihood estimation as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 434, + 526, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 434, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 340, + 434, + 526, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\left(\\frac {1}{T} \\sum_ {t = 1} ^ {T} \\log p (\\hat {x} _ {t} \\mid x _ {1}, x _ {2}, \\dots , x _ {t - 1})\\right)", + "image_path": "8539d09e7ff42b979855d94100a7dfa2105e7c5ba92c9bbccf65ff0d122f3df8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 510 + ], + "type": "text", + "content": "This loss function measures the model's performance by taking the negative logarithm of the probability of predicting each subsequent step given the previous true token." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 395, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 395, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 395, + 534 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 373, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 373, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 373, + 552 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "We evaluate our method using two widely-used autonomous driving research datasets: nuScenes [3] and Argoverse 2 [27]. The nuScenes dataset comprises 1,000 scenes with six cameras offering a " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " field of view. Due to overlapping lanes in the default nuScenes split, PON [24] redistributes the train/validation sequences to eliminate overlap, allowing us to test our approach on both the default and PON splits. Consistent with past studies [17, 21], we input all surrounding camera images and Bird's Eye View (BEV) target ranges from " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "-48\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "48\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " on the x-axis and " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "-30\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " on the y-axis. The Argoverse 2 dataset contains 1,000 sequences and includes seven cameras, also providing a " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " field of view. For this dataset, we use camera images as input, with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27170" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 198 + ], + "type": "table", + "html": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
DefaultTopoNet [10]52.547.149.646.910.817.5
LaneGAP [15]49.957.053.274.134.947.5
RNTR [17]57.142.748.963.745.252.8
LaneGraph2Seq* [21]46.943.745.263.736.346.2
SeqGrowGraph63.650.856.475.561.467.8
PONRNTR* [17]39.931.034.963.324.935.7
LaneGraph2Seq* [21]21.939.928.246.311.017.8
SeqGrowGraph43.533.337.763.636.946.7
", + "image_path": "8cadc0d2971e68817db525f931b837b8cd5f410cdf37ff4baee49089d9ca85d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 206, + 296, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 273 + ], + "type": "text", + "content": "Table 1. Comparison of our method with leading approaches. This evaluates Landmark Precision-Recall and Reachability Precision-Recall on the default and PON [24] split of nuScenes. Results marked with * are reproduced using the official code, while other results are obtained from their respective papers or released checkpoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "content": "the BEV target range from " + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "inline_equation", + "content": "-30\\mathrm{m}" + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "inline_equation", + "content": "30\\mathrm{m}" + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "content": " on the x-axis and " + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "inline_equation", + "content": "-15\\mathrm{m}" + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "inline_equation", + "content": "15\\mathrm{m}" + }, + { + "bbox": [ + 54, + 297, + 295, + 321 + ], + "type": "text", + "content": " on the y-axis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 332, + 153, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 153, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 153, + 345 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": "Following [21] We employ an LSS encoder [22] based on ResNet50 as the BEV encoder, initially training this encoder on a centerline segmentation task. For the sequence decoder, we use a 6-layer transformer. The training process is conducted on 4 NVIDIA A100 GPUs with batch size of " + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "inline_equation", + "content": "4 \\times 18" + }, + { + "bbox": [ + 55, + 351, + 296, + 435 + ], + "type": "text", + "content": ". Specifically, the nuScenes dataset is trained for 400 epochs and the Argoverse 2 dataset is trained for 50 epochs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": "Our sequence comprises the positional coordinates of nodes, node indices, and the positional coordinates of Bézier control points. We employ a discretization resolution of " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " with the node coordinates range set between 0 and 200. To distinguish the meanings of different tokens and to assign varying levels of importance, we introduce a shift to the node indices and Bézier control point coordinates. Consequently, the range for node indices extends from 200 to 350, while the range for Bézier point coordinates spans from 350 to 570. For the special tokens introduced in Figure 4, we omitted the representations for " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{None}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{From}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " to make the sequence more concise. And " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{to}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": ", is denoted as 571, and " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{SEP}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " is represented as 572. Other special tokens, such as " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{EOS}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " (end token of sequence), " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{BOS}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " (beginning token of sequence), and " + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "inline_equation", + "content": "< \\text{N/A}>" + }, + { + "bbox": [ + 55, + 437, + 296, + 628 + ], + "type": "text", + "content": " (padding token), are represented as 573, 574, and 575, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "We believe that the accuracy of edges relies on the correct prediction of points, so the positions of points should carry more weight than those of edges. Therefore, we designed our model such that the tokens representing point positions, which fall within the range of 0 to 200, have a weight of 2 in the cross-entropy loss. Meanwhile, the loss weight for the remaining tokens is set to 1." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 138 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 138 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 138 + ], + "type": "table", + "html": "
MethodsLandmarkReachability
L-PL-RL-FR-PR-RR-F
RNTR* [17]50.729.437.268.129.641.3
LaneGraph2Seq* [21]60.648.753.975.537.850.4
SeqGrowGraph64.150.456.477.743.455.7
", + "image_path": "ad140fcf610cd3d8733128da247588fd8ccc90e294b2d248bed33b67d67fbe97.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 202, + 555, + 249 + ], + "blocks": [ + { + "bbox": [ + 313, + 146, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 190 + ], + "type": "text", + "content": "Table 2. Comparison of our method with leading approaches on Argoverse 2, based on Landmark Precision-Recall and Reachability Precision-Recall. Results marked with * are reproduced using the official code." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 202, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 315, + 202, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 202, + 555, + 249 + ], + "type": "table", + "html": "
MethodsM-PM-RM-FC-PC-RC-FDet.
LaneGraph2Seq [21]64.663.764.169.458.063.264.5
SeqGrowGraph71.270.270.676.258.065.867.2
", + "image_path": "a468edb501b44752fcb11dafb18b1ba5d390215b4fc73a540c0711b25c8e20b0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 257, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 313 + ], + "type": "text", + "content": "Table 3. Comparison of our method with leading approaches on nuScenes, based on metrics including mean precision (M-P), recall (M-R), F1-score (M-F), and connectivity precision (C-P), recall (C-R), F1-score (C-F), and detection ratio. Results are obtained from the original paper." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 333, + 369, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 333, + 369, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 369, + 345 + ], + "type": "text", + "content": "4.3. Metric" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 351, + 556, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 556, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 556, + 471 + ], + "type": "text", + "content": "To ensure a fair comparison, we utilize the metrics proposed by RNTR [17], which encompass Landmark Precision-Recall and Reachability Precision-Recall. Landmark Precision-Recall evaluates the algorithm's accuracy in predicting the locations of key centerline points, such as segmented, fork, and merge points on lane lines. Meanwhile, Reachability Precision-Recall assesses the algorithm's capability to predict connectivity between these landmark nodes, indicating whether roads exist between them. Together, these metrics offer a comprehensive evaluation of the lane graph." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 557, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 557, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 557, + 604 + ], + "type": "text", + "content": "In addition, we employ metrics introduced by STSU [4], including Mean Precision, Mean Recall, and Mean F1-score of centerlines, derived from precision-threshold and recall-threshold curves, alongside their F1-scores. These metrics focus on assessing the accuracy of centerline position matching, without accounting for missing predictions, thereby requiring the detection ratio to evaluate the presence of centerlines. Moreover, Connectivity Precision, Recall, and F1-score ratios evaluate the connectivity between centerlines, primarily focusing on centerline positions and not considering crucial landmark nodes, like junction points." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "type": "text", + "content": "4.4. Comparison" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 653 + ], + "type": "text", + "content": "We compare our approach with state-of-the-art methods using the aforementioned metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "Table 1 compares the results of our method with state-of-the-art methods on nuScenes using both the default and PON [24] splits. Our method outperforms existing algorithms in terms of F1-score for both landmarks and reachability. When there is no lane overlap between the training and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27171" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 282, + 334 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 282, + 334 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 282, + 334 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 282, + 334 + ], + "type": "image", + "image_path": "515278a6cd1ff0940d9ed5024b345aa43f739e16b743322c994e20fae652c86f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 340, + 298, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 340, + 298, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 298, + 440 + ], + "type": "text", + "content": "Figure 6. Various definitions of centerline split have been considered in this study. We have re-segmented the original centerline from the nuScenes dataset to enhance its reasonableness and learning accessibility. Fixed-Length : The centerline between intersection points is partitioned at consistent intervals along the driving direction. In the provided example, these intervals are set at 20 meters and 30 meters. By refining the split using these methods, the learning process is facilitated, thereby yielding more accurate and interpretable results of lane graphs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 460, + 298, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 298, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 298, + 594 + ], + "type": "text", + "content": "test sets, as is the case with the PON split, the performance of all methods significantly decreases, but SeqGrowGraph remains in a leading position. Table 2 compares our method with leading methods on the larger dataset AV2 , demonstrating the superiority of our approach. Table 3 presents a comparison between our method and LaneGraph2Path concerning centerline-related metrics on nuScenes. It is evident that our method still outperforms the comparison method. This proves that modeling sequences in a manner similar to human language effectively harnesses the seq2seq model's capabilities in understanding and reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 600, + 206, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 206, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 206, + 613 + ], + "type": "text", + "content": "4.5. Centerline re-segmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "type": "text", + "content": "In the original dataset, the division of map elements is based on map rules, but these division points often lack distinct visual features. For example, even a straight road might have multiple points that split a centerline into several segments, and these division points may not have any visual features that can be easily identified. Such data can disrupt the model's learning process and impact its ability to accurately interpret the rules. This issue is also noted in [25]," + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 558, + 149 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 558, + 149 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 558, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 558, + 149 + ], + "type": "table", + "html": "
Segment LengthJunction LandmarkJunction Reachability
L-PL-RL-FR-PR-RR-F
Original68.652.659.569.456.462.2
20m70.552.560.270.657.663.4
30m69.352.359.670.858.163.8
40m69.753.360.471.958.664.6
", + "image_path": "c0f76cc445c1029e97fa4c556f9d86cb19d2463357b5eb5af207f64fbdd3416d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 157, + 555, + 202 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 202 + ], + "type": "text", + "content": "Table 4. Training results under different segmentation ways. We resegment the centerline split to accurately and comprehensively represent the lane graph information by re-segmenting the centerlines between junctions based on length." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 222, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 305 + ], + "type": "text", + "content": "and our tasks observe instability in the position of nodes as well. To address this, when a single topological node is referred to as a continuous node that only connects two centelines, we remove it and merge the centerlines. Our method fundamentally ensures that centerlines are segmented exclusively at junctions, including forks, and merge points, which constitute the topological nodes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 556, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 400 + ], + "type": "text", + "content": "When topological points are restricted to intersections, forks, and merges, the centerline may not be accurately represented by a quadratic Bezier function. To address this, we introduce additional topological points along each centerline and subsequently segment it. This segmentation is performed by dividing the centerline based on length. Figure 6 shows the examples when the intervals are set at 20 meters and 30 meters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 533 + ], + "type": "text", + "content": "We evaluate the performance of our method under specific settings by adjusting the metrics to accommodate different ground truth standards. The evaluation framework remains aligned with Landmark Precision-Recall and Reachability Precision-Recall, while incorporating data post-processing. For both the ground truth and predictions, we remove continuous nodes and merge connected centerlines. This revised metric, termed Junction Landmark Precision-Recall and Junction Reachability Precision-Recall, emphasizes the accuracy of junction locations and the connectivity between junctions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": "Table 4 presents a comparison between the re-segmented centerline and the original ground truth. It can be observed that the metrics retrained after re-segmentation are all higher than those of the original split way. This suggests that we should adopt a similar standard when constructing ground truth for lane graphs, which could provide insights for research on lane graphs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "text", + "content": "4.6. Visualization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We conduct a visual comparison of the final results produced by various algorithms on the nuScenes dataset, as illustrated in Figure 7. This comparison reveals several limitations inherent in previous methods: TopoNet [10] often produces centerlines that are misaligned with each other, failing to provide a continuous centerline representation. LaneGAP [15]" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27172" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 293, + 149 + ], + "type": "table", + "html": "
WeightLandmarkReachability
L-PL-RL-FR-PR-RR-F
162.648.654.774.460.266.5
263.650.856.475.561.467.8
362.549.655.375.461.367.6
463.050.255.975.361.267.5
", + "image_path": "1d2e27242362a2dc773ce73e5ef3accb9fa0db10608b2efbb873bfcd8e57db4c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 212, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 212, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 296, + 357 + ], + "type": "text", + "content": "tends to generate duplicate centerlines and may erroneously merge adjacent centerlines, leading to inefficiencies and inaccuracies. RNTR [17] demonstrates instability in its modeling capabilities, compromising the reliability of the resulting lane graphs. Furthermore, LaneGraph2Path [21] may result in a decoupling between points and edges, undermining the structural integrity of the graphs. In contrast, our algorithm not only provides stable and robust modeling but also effectively merges redundant points present in the ground truth data, thereby achieving superior results in constructing lane graphs. This advancement highlights the effectiveness and reliability of our approach in overcoming these challenges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 148, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 148, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 148, + 376 + ], + "type": "text", + "content": "4.7. Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 416 + ], + "type": "text", + "content": "We conduct ablation experiments on the nuScenes dataset, which includes examining the impact of loss weight on point positions and the order of subgraphs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 523 + ], + "type": "text", + "content": "Higher positional weight is applied to tokens ranging from 0 to 200 because these correspond to the positions of intersection points. Accurate descriptions of these intersection positions are vital for correctly depicting the topology and centerline locations between intersections. Table 5 illustrates that the modeling method remains relatively stable across different weights, with a slight improvement when the weight is set to 2, highlighting the robustness of the algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 524, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 666 + ], + "type": "text", + "content": "The potential order in which each intersection point can be traversed to construct the corresponding subgraph is also investigated. In this context, Center refers to starting from the lane graph's center and proceeding from the closest to the farthest intersection based on their distance from the center. Coord denotes traversal based on the coordinates of the intersection points, specifically according to the x and y values. BFS stands for breadth-first search traversal, and DFS indicates depth-first search traversal. It can be observed that DFS performs the best, which suggests that depth-first search helps capture the hierarchical relationships between intersections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we introduce SeqGrowGraph, an innovative framework for autoregressive modeling of lane graphs that" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 330, + 72, + 544, + 329 + ], + "blocks": [ + { + "bbox": [ + 330, + 72, + 544, + 329 + ], + "lines": [ + { + "bbox": [ + 330, + 72, + 544, + 329 + ], + "spans": [ + { + "bbox": [ + 330, + 72, + 544, + 329 + ], + "type": "image", + "image_path": "b8d52ce30608c410f36bdf39cf0adecbbedb2e73816a6c5ac3d0808c584027d2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 7. We visualize and compare the final results of various algorithms on the nuScenes dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 374, + 553, + 449 + ], + "blocks": [ + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "type": "text", + "content": "Table 5. Ablation study on different loss weights assigned to token of point coordinates conducted using the nuScenes dataset. The weights for the remaining tokens are set to 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 374, + 553, + 449 + ], + "lines": [ + { + "bbox": [ + 315, + 374, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 315, + 374, + 553, + 449 + ], + "type": "table", + "html": "
OrderLandmarkReachability
L-PL-RL-FR-PR-RR-F
Center59.047.152.473.356.363.7
Coord60.148.653.773.757.264.4
BFS62.649.255.175.361.467.7
DFS63.650.856.475.561.467.8
", + "image_path": "129c13b057c3345029c4e5764a7223050fa85cc914ccb8e79a1feaf983c774b0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "text", + "content": "Table 6. Ablation study on the ordering of subgraph descriptions conducted using the nuScenes dataset. Center pertains to initiating from the central point of the lane graph and navigating towards intersections in order of their increasing distance from this center. Coord involves traversing based on the coordinates of intersection points, specifically ordered by their x and y values. " + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "inline_equation", + "content": "BFS" + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "inline_equation", + "content": "DFS" + }, + { + "bbox": [ + 313, + 456, + 555, + 534 + ], + "type": "text", + "content": " refers to breadth-first traversal and depth-first traversal." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "converts lane graphs into sequences via incremental local growth, facilitating a comprehensive global understanding. Our method enables flexible representation of complex road structures, including loops and bidirectional lanes, which pose challenges for traditional DAG-based approaches. Extensive experiments conducted on the nuScenes and Argoverse 2 datasets demonstrate superior performance in topological accuracy and graph completeness compared to existing detection-based and generation-based methods. SeqGrowGraph advances lane graph construction by bringing human-like structural reasoning and sequence modeling, thereby providing accurate centerlines for autonomous driving in dynamic environments." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27173" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In ICRA, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Yifan Bai, Zeyang Zhao, Yihong Gong, and Xing Wei. Artrackv2: Prompting autoregressive tracker where to look and how to describe. In CVPR, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 255 + ], + "type": "text", + "content": "[4] Yigit Baran Can, Alexander Liniger, Danda Pani Paudel, and Luc Van Gool. Structured bird's-eye-view traffic scene understanding from onboard images. In ICCV, 2021. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[5] Xinyuan Chang, Maixuan Xue, Xinran Liu, Zheng Pan, and Xing Wei. Driving by the rules: A benchmark for integrating traffic sign regulations into vectorized hd map. In CVPR, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "type": "text", + "content": "[6] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "text", + "content": "[7] Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. Can llms effectively leverage graph structural information through prompts, and why? Trans. Mach. Learn. Res., 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 402 + ], + "type": "text", + "content": "[8] Chunliang Li, Wencheng Han, Jun Yin, Sanyuan Zhao, and Jianbing Shen. Repvf: A unified vector fields representation for multi-task 3d perception. In ECCV, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 434 + ], + "type": "text", + "content": "[9] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In ICRA, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 437, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 296, + 490 + ], + "type": "text", + "content": "[10] Tianyu Li, Li Chen, Huijie Wang, Yang Li, Jiazhi Yang, Xiangwei Geng, Shengyin Jiang, Yuting Wang, Hang Xu, Chunjing Xu, Junchi Yan, Ping Luo, and Hongyang Li. Graph-based topology reasoning for driving scenes. arXiv preprint arXiv:2304.05277, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 296, + 534 + ], + "type": "text", + "content": "[11] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Grouplane: End-to-end 3d lane detection with channel-wise grouping. Robotics Autom. Lett., 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 296, + 567 + ], + "type": "text", + "content": "[12] Shiyi Liang, Yifan Bai, Yihong Gong, and Xing Wei. Autoregressive sequential pretraining for visual tracking. In CVPR, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 296, + 613 + ], + "type": "text", + "content": "[13] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In ICLR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 615, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 296, + 656 + ], + "type": "text", + "content": "[14] Benchcheng Liao, Shaoyu Chen, Yunchi Zhang, Bo Jiang, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Maptrv2: An end-to-end framework for online vectorized HD map construction. CoRR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[15] Benchcheng Liao, Shaoyu Chen, Bo Jiang, Tianheng Cheng, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Lane graph as path: Continuity-preserving path-wise modeling for online lane graph construction. In ECCV, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Vectormapnet: End-to-end vectorized HD map learning. In ICML, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[17] Jiachen Lu, Hongyang Li, Renyuan Peng, Feng Wen, Xinyue Cai, Wei Zhang, Hang Xu, and Li Zhang. Translating images to road network: A non-autoregressive sequence to-sequence approach. In ICCV, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[18] Katie Z. Luo, Xinshuo Weng, Yan Wang, Shuang Wu, Jie Li, Kilian Q. Weinberger, Yue Wang, and Marco Pavone. Augmenting lane perception and topology understanding with standard definition navigation maps. In ICRA, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 227 + ], + "type": "text", + "content": "[19] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. LATR: 3d lane detection from monocular images with transformer. In ICCV, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 259 + ], + "type": "text", + "content": "[20] Yuxian Meng, Shi Zong, Xiaoya Li, Xiaofei Sun, Tianwei Zhang, Fei Wu, and Jiwei Li. GNN-LM: language modeling based on global contexts via GNN. In ICLR, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 262, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 553, + 304 + ], + "type": "text", + "content": "[21] Renyuan Peng, Xinyue Cai, Hang Xu, Jiachen Lu, Feng Wen, Wei Zhang, and Li Zhang. Lanegraph2seq: Lane topology extraction with language model via vertex-edge encoding and connectivity enhancement. In AAAI, 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 305, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 305, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 305, + 553, + 337 + ], + "type": "text", + "content": "[22] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 370 + ], + "type": "text", + "content": "[23] Maximilian Pittner, Joel Janai, and Alexandru Paul Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "type": "text", + "content": "[24] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In CVPR, 2020. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 405, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 468 + ], + "type": "text", + "content": "[25] Huijie Wang, Tianyu Li, Yang Li, Li Chen, Chonghao Sima, Zhenbo Liu, Bangjun Wang, Peijin Jia, Yuting Wang, Shengyin Jiang, Feng Wen, Hang Xu, Ping Luo, Junchi Yan, Wei Zhang, and Hongyang Li. Openlane-v2: A topology reasoning benchmark for unified 3d HD mapping. In NeurIPS, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 491 + ], + "type": "text", + "content": "[26] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 557 + ], + "type": "text", + "content": "[27] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In NeurIPS Datasets and Benchmarks, 2021. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 559, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 580 + ], + "type": "text", + "content": "[28] Chengtang Yao, Lidong Yu, Yuwei Wu, and Yunde Jia. Sparse point guided 3d lane detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 624 + ], + "type": "text", + "content": "[29] Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. Language is all a graph needs. In *Findings of the Association for Computational Linguistics: EACL 2024*, St. Julian's, Malta, March 17-22, 2024, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "text", + "content": "[30] Shuang Zeng, Xinyuan Chang, Xinran Liu, Zheng Pan, and Xing Wei. Driving with prior maps: Unified vector prior encoding for autonomous vehicle mapping. arXiv preprint arXiv:2409.05352, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27174" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[32] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In AAAI, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 162 + ], + "type": "text", + "content": "[33] Yi Zhou, Hui Zhang, Jiaqian Yu, Yifan Yang, Sangil Jung, Seung-In Park, and ByungIn Yoo. Himap: Hybrid representation learning for end-to-end vectorized HD map construction. In CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27175" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_content_list.json b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d78399cc372aa41b68013a7884396d36966302e2 --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_content_list.json @@ -0,0 +1,1901 @@ +[ + { + "type": "text", + "text": "Sequential Gaussian Avatars with Hierarchical Motion Context", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wangze Xu $^{1*}$ Yifan Zhan $^{1,2}$ Zhihang Zhong $^{1\\dagger}$ Xiao Sun $^{1\\dagger}$ $^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ The University of Tokyo", + "bbox": [ + 212, + 179, + 785, + 219 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a382d1fa5228c3c14aeda352e1bade0b89ee9dfcbe2eb901aa8326f6391aeb9.jpg", + "image_caption": [ + "Figure 1. Illustration of the Hierarchical Motion Context Design. We model the motion-dependent appearance variations with both coarse skeleton condition $\\Delta P$ and fine-grained point-wise velocity condition $\\mathcal{V}$ . $\\Delta P$ describes the overall human skeleton motion, which is derived from the difference between the human poses at adjacent frames. $\\mathcal{V}$ is the point-wise velocity that indicates finer-grained motion in local regions. To achieve robust deformation, we propose Spatio-temporal Multi-scale Sampling, which samples the overall motion trend and inter-frame details via diverse time intervals $s$ for $\\Delta P$ and $\\mathcal{V}$ ." + ], + "image_footnote": [], + "bbox": [ + 116, + 234, + 883, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 569, + 326, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of neural rendering has significantly advanced the rendering quality of 3D human avatars, with the recently popular 3DGS technique enabling real-time performance. However, SMPL-driven 3DGS human avatars still struggle to capture fine appearance details due to the complex mapping from pose to appearance during fitting. In this paper, we propose SeqAvatar, which excavates the explicit 3DGS representation to better model human avatars based on a hierarchical motion context. Specifically, we utilize a coarse-to-fine motion conditions that incorporate both the overall human skeleton and fine-grained vertex motions for non-rigid deformation. To enhance the robustness of the proposed motion conditions, we adopt a spatiotemporal multi-scale sampling strategy to hierarchically integrate more motion clues to model human avatars. Extensive experiments demonstrate that our method significantly outperforms 3DGS-based approaches and renders human", + "bbox": [ + 88, + 603, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "avatars orders of magnitude faster than the latest NeRF-based models that incorporate temporal context, all while delivering performance that is at least comparable or even superior. Project page: https://zezeaaa.github.io/projects/SeqAvatar/", + "bbox": [ + 511, + 571, + 906, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 676, + 645, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research on digital humans has highlighted the efficiency of 3D Gaussian Splatting (3DGS) [27], demonstrating its capability for high-quality and real-time rendering. By defining T-posed human Gaussians in a canonical space and employing specific warping techniques, we can render human avatars from diverse perspectives and in any pose. The canonical Gaussian primitives and warping weights are then jointly optimized under supervision from video inputs.", + "bbox": [ + 509, + 703, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animatable human avatar reconstruction is primarily hindered by limited modeling of non-rigid warping, forcing a trade-off between rendering quality and animation capability. Current methods [22, 56] incorporate SMPL(-X) pose priors [43, 51] to guide motion for each observation", + "bbox": [ + 509, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during the author's internship at Shanghai Artificial Intelligence Laboratory. † denotes co-corresponding authors.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13592", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and rely on Linear Blend Skinning (LBS) for warping. Although these approaches are effective for pose-driven animation, they often struggle with per-frame non-rigid warping in scenarios involving complex garments.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We experimentally find that current human pose conditions [22, 56] do not fully capture the complex one-to-many mapping from pose to appearance, particularly for 3DGS-based approaches [27]. These methods [56, 68] rely on the spatial information of a template body in each frame's human pose to predict non-rigid deformations. However, they often overlook local details, such as garment deformations far from the skeleton, and cannot resolve cases where the same pose corresponds to different appearances during complex motions. Furthermore, although previous NeRF-based [44] method [8] has attempted to model motion sequences using human pose residuals, the inherently global nature of the pose sequence limits the ability to capture finer motion details. Naive pose sequence modeling does not fully account for the explicit characteristics of 3DGS.", + "bbox": [ + 89, + 152, + 482, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a hierarchical motion context condition for 3DGS-based human avatar modeling to address the limitations of relying solely on human pose, which provides only limited global skeletal information. Our approach improves the ability of 3DGS-based methods to accurately capture the complex relationship between human pose and appearance in challenging scenarios. Specifically, we design a coarse-to-fine motion condition that incorporates both overall skeletal movements and fine-grained point-wise motions. Leveraging the explicit nature of Gaussian primitives, this condition seamlessly integrates into 3DGS-based methods, enabling more precise predictions of complex non-rigid deformations. To further enhance robustness, we propose a spatio-temporal multi-scale sampling strategy for constructing the motion context with a larger receptive field. Spatially, we consider the motion states of neighboring points within the local region of each Gaussian primitive to obtain more stable motion embeddings. Temporally, we capture human motion patterns across multiple time scales, combining long-term trends with fine-grained inter-frame motion details. This improves the model's generalization to complex human movements.", + "bbox": [ + 89, + 381, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our key contributions are as follows:", + "bbox": [ + 109, + 715, + 452, + 729 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose a novel hierarchical motion condition that integrates coarse-to-fine human motion, combining global skeletal poses with localized vertex residuals to enhance non-rigid deformation prediction.", + "2) We introduce a spatiotemporal multiscale sampling strategy that expands the receptive field of hierarchical motion context, improving generalization to complex motions.", + "3) Experiments on the I3D-Human [8], DNA-Rendering [9], and ZJU-MoCap [53] datasets show the effectiveness of the proposed SeqAvatar, which is capable of modeling details of the human body in complex motions." + ], + "bbox": [ + 83, + 731, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering. Neural rendering techniques have brought significant progress to human reconstruction and rendering. In particular, Neural Radiance Fields (NeRF) [44] introduces an implicit scene representation that models color and density using multilayer perceptrons, delivering photorealistic rendering results. More recently, point-based 3D Gaussian Splatting [27] utilizes 3D Gaussians to represent scenes explicitly, achieving real-time high-quality rendering. Building upon NeRF and 3DGS, subsequent works have advanced neural rendering across various dimensions, e.g., enhancing visual fidelity [1, 2], improving sparse-view reconstruction quality [10, 52, 63, 71, 80], enabling pose-free optimization [37, 58, 61, 67], modeling dynamic scenes [13, 34, 49, 55, 69, 74], and accelerating training and inference [4, 12, 18, 39, 46, 60]. These developments also benefit human avatar modeling, which demands high-quality and efficient rendering.", + "bbox": [ + 511, + 114, + 903, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SMPL(-X)-Based Neural Human Modeling. The SMPL(-X) family [43, 51] provides a parametric representation of the human body by decomposing it into pose-related and shape-related components using 3D mesh scans and principal component analysis (PCA). Its pose blend shapes enables body deformation through joint-wise pose blending, offering an efficient and compact way for animation. SMPL(-X) has thus become a cornerstone in human body modeling and animation, with many methods [11, 59] estimating its parameters directly from 2D inputs. Recent neural human reconstruction methods integrate SMPL(-X) with implicit representations such as NeRF [44] and 3DGS [27] to enable animatable avatars with high rendering quality. Different from purely 2D image animation methods [48, 62, 65, 73], these methods typically register the input data to a canonical T-pose space and use linear blend skinning (LBS) to transform 3D points into observation space based on SMPL(-X) poses. NeRF-based approaches [5-8, 14, 17, 19, 20, 30, 53, 64, 68, 75] focus on monocular or multi-view reconstruction. More recently, 3DGS-based methods [21-23, 25, 26, 29, 31-33, 36, 40, 41, 45, 47, 56, 76, 78, 79, 81] have gained popularity due to their real-time rendering and high fidelity. Some works [8, 23] further incorporate pose sequences as temporal context, but they suffer from slow rendering and lack fine details of motion caused by simply integrating coarse human pose embedding with NeRF.", + "bbox": [ + 511, + 372, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Rendering with Temporal Embeddings The earliest modeling of dynamic radiance field [3, 13, 15, 16, 24, 34, 35, 42, 50, 57, 70] can be traced back to temporal embeddings in general scenes, where each frame's observation is obtained by constructing a canonical space and a time-conditioned deformation field. Motivated by these, a stream of research [38, 72] focuses on pure rendering quality, employing temporal embeddings instead of human pose to en", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "code each frame of human videos. Although these methods achieve high-quality, temporally continuous rendering, they struggle with the lack of geometric constraints from human pose, making parameterized animation unfeasible.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 165, + 217, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMPL(-X) Series [43, 51] adopt a parameterized framework to represent human bodies across diverse shapes and poses. For each frame, the human mesh is derived by deforming a canonical template mesh based on shape and pose parameters. Specifically, a 3D point $\\mathbf{x}$ on the canonical mesh is warped to the corresponding point on the deformed mesh in the observation space, following", + "bbox": [ + 89, + 190, + 483, + 297 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L B S} (\\mathbf {x}, \\mathbf {B} _ {k}, \\omega_ {k} (\\mathbf {x})) = \\sum_ {k = 1} ^ {K} \\omega_ {k} (\\mathbf {x}) \\mathbf {B} _ {k} \\mathbf {x}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 308, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $K$ is the total bone number and $\\mathbf{B}_k$ is the transformation matrix for each bone. Specifically, $\\mathbf{B}_k$ consists of a rotation and translation matrix $\\mathbf{R}_k$ and $\\mathbf{b}_k$ . Here, $\\mathbf{R}_k$ represents the global rotation of each joint, influenced by the local joint rotations $\\boldsymbol{r}$ , while $\\mathbf{b}_k$ denotes each joint's translation, determined by the joint positions $\\boldsymbol{j}$ and the human shape parameters $\\beta$ . The linear blending weight $\\omega^k$ depends on $\\mathbf{x}$ and is regressed from comprehensive human meshes.", + "bbox": [ + 89, + 359, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting [27] utilizes a set of point-based Gaussian primitives to explicitly represent scenes, enabling real-time and high-quality rendering. Each Gaussian primitive is defined by its center position $\\mathbf{x} \\in \\mathbb{R}^3$ and covariance matrix $\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ . To ensure positive semi-definiteness and simplify optimization, the covariance matrix $\\boldsymbol{\\Sigma}$ is decomposed into a rotation matrix $\\mathbf{R}$ and a scaling matrix $\\mathbf{S}$ :", + "bbox": [ + 89, + 481, + 483, + 587 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 597, + 480, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{R}$ and $\\mathbf{S}$ are derived by a scaling vector $\\mathbf{s} \\in \\mathbb{R}^3$ and a quaternion vector $\\mathbf{r} \\in \\mathbb{R}^4$ in practice. Additionally, each Gaussian primitive is also assigned a color feature $sh \\in \\mathbb{R}^k$ represented by spherical harmonics (SH) and an opacity $\\alpha$ for rendering. During the rendering process, a splatting technique [27] is used to project each Gaussian primitive onto the 2D image space with a viewing transform $\\mathbf{W}$ and the Jacobian $\\mathbf{J}$ of the projective transformation's affine approximation. The transformed covariance $\\boldsymbol{\\Sigma}'$ in the camera's coordinates is", + "bbox": [ + 89, + 627, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 787, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After projection, the pixel color $C$ is computed by blending $N$ ordered Gaussian primitives overlapping at the pixel:", + "bbox": [ + 89, + 818, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 861, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ is computed from SH feature $sh$ and $\\alpha_{i}^{\\prime}$ is the product of $\\alpha_{i}$ and probability density of $i$ -th 2D Gaussian.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For optimization, a photometric loss is defined by a combination of $L_{1}$ and SSIM [66] losses:", + "bbox": [ + 511, + 121, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p h o t o}} = \\lambda L _ {1} (\\hat {I}, I) + (1 - \\lambda) (1 - S S I M (\\hat {I}, I)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 161, + 903, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{I}$ and $I$ denote the rendered and ground-truth images, and $\\lambda$ controls the balance between the two terms.", + "bbox": [ + 511, + 191, + 903, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 234, + 604, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the framework of our method. We use the explicit point-based 3DGS as the representation of the human body. Given a collection of input cameras and images, we optimize a set of Gaussian primitives $\\{\\mathcal{G}_i\\}_{i=1}^{n}$ to fit the body's shape and appearance. Each Gaussian primitive $\\mathcal{G}_i$ includes the center position $\\mathbf{x}$ , scaling vector $\\mathbf{s}$ , rotation quaternion vector $\\mathbf{r}$ , color feature $sh$ and opacity $\\alpha$ , where $\\mathbf{x}$ is initialized from the SMPL template vertices $\\{\\mathbf{T}_i\\}_{i=1}^{i=N}$ . During the optimization process, we first apply a coarse-to-fine motion context to capture more accurate and fine-grained details (Sec. 4.1). To mitigate overfitting, we propose Spatio-Temporal Multi-Scale Sampling to obtain more robust point-wise Gaussian motion, which serves as the embedding for non-rigid deformation (Sec.4.2). Next, we adopt Linear Blend Skinning (LBS) to map canonical Gaussian primitives $\\mathcal{G}$ to observation space and render images via differentiable splitting (Sec.4.3).", + "bbox": [ + 511, + 261, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Non-Rigid Deformation with Coarse-to-fine Motion Context", + "text_level": 1, + "bbox": [ + 511, + 527, + 903, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse Skeleton Motion. As illustrated in Fig. 1, we introduce a coarse-to-fine motion context to excavate additional conditions for non-rigid deformation. Typically, most human 3D avatar methods use the current frame pose as a condition to predict non-rigid deformation [22, 56]. While this provides spatial information to distinguish the motion of different body parts, it fails to capture temporal motion changes, as discussed in [8]. Therefore, given a frame at time $t$ , we consider a sequence of regularly interval-sampled frames $\\mathcal{T}$ to model inter-frame body motion variations:", + "bbox": [ + 511, + 565, + 906, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} = \\{t - s, t - 2 s, \\dots , t - L s \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 728, + 903, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L$ is the sequence length and $s$ is the time interval. The coarse motion of body skeletons at each time step $t \\in \\mathcal{T}$ can be derived by calculating the difference of poses between adjacent frames as illustrated in Fig. 1, following", + "bbox": [ + 511, + 755, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathcal {P} = \\left\\{\\Delta P ^ {t} = \\delta \\left(P ^ {t}, P ^ {t - s}\\right) | t \\in \\mathcal {T} \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 825, + 903, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{R}^{K \\times 3}$ is the body pose and $\\delta$ denotes the difference $\\Delta \\mathcal{P}$ between two poses in axis-angle form. We employ an MLP $\\mathcal{E}_{\\Delta \\mathcal{P}}$ to encode the sequential skeleton mo", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13594", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bec0715e77cfb6f8898d1bbbc367c58618f4d3b2c0dee2f2110e75d4162cb622.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. We first initialize canonical Gaussian positions $\\mathbf{x}$ with SMPL template vertexes. For each Gaussian, we derive both coarse skeleton motion condition $f_{\\Delta \\mathcal{P}}$ and fine-grained vertex motion condition $f_{\\nu}$ that sampled from the vertex motion template $\\mathcal{F}_{\\nu}$ (points in different colors represent different motions). Based on such hierarchical motion information, we utilize an MLP $\\mathcal{E}_{non-rigid}$ to better predict each Gaussian's non-rigid deformation prediction. The non-rigid deformed Gaussians are then warped into observation space via the standard LBS transformation for rendering." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 883, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion $\\Delta \\mathcal{P}\\in \\mathbb{R}^{L\\times K\\times 3}$ , flattened along the temporal dimension, into a skeleton motion embedding $f_{\\Delta \\mathcal{P}}\\in \\mathbb{R}^{32}$", + "bbox": [ + 88, + 402, + 480, + 433 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} (\\Delta \\mathcal {P}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 441, + 480, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which serves as a condition for the subsequent non-rigid deformation. Further details are provided in the supplementary material.", + "bbox": [ + 89, + 465, + 482, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine Vertex Motion. Compared to previous implicit NeRF-based methods, the point-based 3DGS representation enables us to explore more fine-grained temporal motion information. We derive a point-wise velocity vector $v_{i}$ for each Gaussian primitive $\\mathcal{G}_i$ to model the fine-grained localized body motion, which is beyond the scope of pose skeleton motion $\\Delta \\mathcal{P}$ . By measuring the position variations across adjacent time steps, we calculate per-frame velocity", + "bbox": [ + 89, + 510, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {t} = \\frac {\\mathbf {x} _ {\\mathbf {o}} ^ {t} - \\mathbf {x} _ {\\mathbf {o}} ^ {t - s}}{s}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 637, + 480, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x_o}$ denotes the warped coordinates in observation space. However, since the positions of Gaussian primitives are updated during optimization, it is not stable to obtain $v_{i}$ based on such a dynamic variable. Moreover, transforming the canonical Gaussians into the observation space requires non-rigid deformation first, which leads to a circular dependency and conflicts with the current velocity design. To this end, we introduce a motion template field $\\mathcal{F}_{\\mathrm{V}} = \\{\\mathbf{V}_i\\}_{i = 1}^{i = N}$ that stores the velocity of each SMPL vertex, and $v_{i}$ for each Gaussian primitive is sampled from $\\mathcal{F}_{\\mathrm{V}}$ based on the distance between query points and vertexes in this template. Specifically, given a time step $t$ and the corresponding body pose $P^t = \\{(\\mathbf{R}_k^t,\\mathbf{b}_k^t)|k\\in K\\}$ , the SMPL template vertexes in $\\mathbf{T}$ can be warped into observation space with the template skinning weights $\\mathbf{W}$ and", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "standard linear blend skinning, following", + "bbox": [ + 513, + 402, + 787, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\mathbf {o}} ^ {t} = \\mathbf {L B S} (\\mathbf {T}, \\mathbf {B} ^ {t}, \\mathbf {W}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 426, + 903, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}^t$ is the transformation matrix derived from rotation and translation matrix $\\mathbf{R}^t$ and $\\mathbf{b}^t$ . $\\mathbf{W}$ is the template SMPL LBS weights. The velocity $\\mathbf{V}$ of each template vertexes $\\mathbf{T}$ can be further derived as", + "bbox": [ + 511, + 450, + 906, + 511 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} ^ {t} = \\frac {\\mathbf {T} _ {\\mathbf {o}} ^ {t} - \\mathbf {T} _ {\\mathbf {o}} ^ {t - s}}{s}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 516, + 903, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then sample each Gaussian's velocity from this motion template field $\\mathcal{F}_{\\mathbf{V}}$ , which is discussed in Sec. 4.2.", + "bbox": [ + 511, + 555, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Spatio-temporal Multi-scale Sampling (STMS)", + "text_level": 1, + "bbox": [ + 511, + 593, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the robustness of human motion conditions, we propose modeling both local region motion information of the human body and motion patterns across different temporal windows to mitigate overfitting and improve generalization by capturing more comprehensive motion dynamics.", + "bbox": [ + 511, + 614, + 905, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the spatial dimension, we sample the $\\tau$ nearest template to model the body's local region motion more robustly. Specifically, for each canonical Gaussian primitive $\\mathcal{G}_i$ , the $\\tau$ nearest vertexes' velocities are sampled as input to an MLP $\\mathcal{E}_{knn}$ to learn a local point-wise motion embedding", + "bbox": [ + 511, + 691, + 905, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {t} = \\mathcal {E} _ {k n n} \\left(\\left\\{\\mathbf {V} _ {j} ^ {t} \\right\\}\\right), \\quad j \\in \\operatorname {K N N} (\\mathbf {T}, \\mathbf {x} _ {i}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 773, + 903, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{KNN}(\\mathbf{T},\\mathbf{x}_i)$ denotes the $\\tau$ nearest SMPL template vertexes of the canonical Gaussian position $\\mathbf{x}_i$ . Similar to Eq. 7, we then apply an MLP $\\mathcal{E}_{\\mathcal{V}}$ to encode each Gaussian primitive's sequential motion embedding $\\mathcal{V} = \\{e^{t}|t\\in \\mathcal{T}\\}$ into a point-wise sequential condition $f_{\\mathcal{V}}\\in \\mathbb{R}^{96}$", + "bbox": [ + 511, + 800, + 905, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {V}). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the temporal dimension, to capture both the overall motion trend and the inter-frame motion details, we adopt a multi-scale sequence sampling strategy. In detail, we sample a series of sequences at several progressively increasing intervals to get human motion information across different temporal windows", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{s = s _ {0} + i \\Delta s \\right\\} _ {i = 0} ^ {i = m}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 191, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta s$ is the increasing rate of sampling interval $s$ . We then input the multi-scale sampled sequence motions into the skeleton motion encoder $\\mathcal{E}_{\\Delta P}$ and the point-wise motion encoder $\\mathcal{E}_{V}$ to obtain hierarchical temporal motion embeddings. Therefore, Eq. 7 and Eq. 13 can be revised as", + "bbox": [ + 89, + 217, + 483, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} \\left(\\left\\{\\Delta \\mathcal {P} _ {s} \\right\\}\\right), \\quad s \\in \\mathcal {S} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 303, + 482, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\{\\mathcal {V} _ {s} \\}), \\quad s \\in \\mathcal {S}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 321, + 482, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we concatenate all skeleton motion conditions, $\\Delta \\mathcal{P}_s$ for $s\\in S$ , and localized vertex motion conditions, $\\nu_{s}$ for $s\\in S$ , across different sampling scales. These are then input to $\\varepsilon_{\\Delta P}$ and $\\varepsilon_{\\nu}$ , respectively. More details are shown in the supplementary material.", + "bbox": [ + 89, + 348, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Non-Rigid Deformation. Given the coarse skeleton motion $f_{\\Delta \\mathcal{P}} = \\mathcal{E}_{\\Delta \\mathcal{P}}(\\{\\Delta \\mathcal{P}_s\\})$ condition and fine point-wise velocity condition $f_{\\mathcal{V}} = \\mathcal{E}_{\\mathcal{V}}(\\{\\mathcal{V}_s\\})$ , we utilize an MLP to predict each Gaussian's non-rigid deformation as", + "bbox": [ + 89, + 422, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\mathbf {x}, \\delta \\mathbf {s}, \\delta \\mathbf {r} = \\mathcal {E} _ {\\text {n o n - r i g i d}} (\\mathbf {x}, P, f _ {\\Delta \\mathcal {P}}, f _ {v}). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 493, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The deformed canonical Gaussian $\\mathcal{G}'$ is", + "bbox": [ + 89, + 518, + 352, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = \\mathbf {x} + \\delta \\mathbf {x}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 544, + 482, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} ^ {\\prime} = \\mathbf {s} + \\delta \\mathbf {s}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 564, + 480, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} ^ {\\prime} = \\mathbf {r} \\cdot \\delta \\mathbf {r}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 582, + 480, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cdot$ denotes the multiplication of two quaternions.", + "bbox": [ + 89, + 608, + 447, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Optimization", + "text_level": 1, + "bbox": [ + 89, + 632, + 230, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rigid Deformation. We utilize the standard LBS operation to map the non-rigid deformed Gaussians $\\mathcal{G}_o$ into the observation space, following", + "bbox": [ + 89, + 654, + 483, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathbf {o}} = \\mathbf {L B S} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {B}, \\omega\\right), \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 709, + 482, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {\\mathbf {o}} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\mathbf {x} ^ {\\prime}\\right) \\mathbf {B} _ {k} \\mathbf {R} ^ {\\prime}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 729, + 482, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{R}'$ is the rotation matrix derived from the non-rigid deformed Gaussian's rotation quaternion $\\mathbf{r}'$ , and $\\mathbf{R_o}$ represents the rotation matrix in observation space. Following the previous method [22], we utilize an MLP $\\mathcal{E}_{lbs}$ to predict the LBS weight offsets for each query canonical point and update the sampled weights from the nearest SMPL vertex as", + "bbox": [ + 89, + 779, + 483, + 883 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {k} (\\mathbf {x}) = \\omega_ {k} ^ {S M P L} (\\mathbf {x}) + \\mathcal {E} _ {l b s} (\\mathbf {x}). \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to [8, 22, 68], we introduce a pose refinement MLP $\\mathcal{E}_{\\text {pose }}$ to refine the pose estimate from SMPL for a better fit to the human body.", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. With the transformed Gaussian primitives $\\mathcal{G}_o$ in observation space, we apply the standard splitting [27] to render images as", + "bbox": [ + 511, + 136, + 906, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI = \\mathbf {S p l a t t i n g} \\left(\\mathbf {x} _ {\\mathbf {o}}, \\mathbf {R} _ {\\mathbf {o}}, s ^ {\\prime}, \\alpha , s h\\right). \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 194, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the optimization process, we employ a combination of loss functions as supervision, summarized as", + "bbox": [ + 511, + 222, + 905, + 252 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {l p i p s}} + \\mathcal {L} _ {\\text {m a s k}}, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 265, + 906, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is used to balance the weight of different losses. $\\mathcal{L}_{mask}$ [22] is an $L_{2}$ loss between the rendered $\\alpha$ and human body mask. Similar to [56], we apply $\\mathcal{L}_{isopos}$ and $\\mathcal{L}_{isocov}$ to control the Gaussian primitive's position and covariance. Details are shown in the supplementary material.", + "bbox": [ + 511, + 292, + 905, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 382, + 645, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 407, + 614, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DNA-Rendering [9] is a challenging human-centric rendering dataset. It contains diverse scenes from everyday life to professional occasions. We use 6 sequences with loose-fitting garments (1_0206_04, 2_0007_04, 2_0019_10, 2_0044_11, 2_0051_09, 2_0813_05) for experiments. We adopt 24 views for training and 6 views for testing.", + "bbox": [ + 511, + 429, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "13D-Human [8] contains scenes closer to daily life, which comprises multi-view video frames of humans with loose clothing and complex movements. We conduct experiments on 4 sequences and use 4-5 views for training and the rest for testing. All training and testing data are retained to be the same in comparisons following Dyco's [8] data split.", + "bbox": [ + 511, + 520, + 905, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Baselines and Metrics", + "text_level": 1, + "bbox": [ + 513, + 619, + 718, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare against the state-of-the-art human modeling methods including Dyco [8], 3DGS-Avatar [56], GART [31], and GauHuman [22]. Dyco is an implicit NeRF-based method that adopts pose sequence information for temporal modeling. We compare performance with Dyco to showcase the better efficacy of our fine-grained motion condition design. 3DGS-Avatar, GART, and GauHuman are explicit 3DGS-based methods, and all of them are originally designed for monocular inputs. For fairness, we extend them to multi-view input under the same settings. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [66], and learned perceptual image patch similarity (LPIPS) [77].", + "bbox": [ + 511, + 642, + 906, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Comparison", + "text_level": 1, + "bbox": [ + 513, + 848, + 645, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison on DNA-Rendering. Tab. 1 shows the quantitative results on DNA-Rendering dataset. The proposed", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8b2e7d7e3bd8c797a92ae9a148e925229f3a6de81cbdbcc5c324a702206cc45b.jpg", + "table_caption": [ + "Table 1. Quantitative Results on DNA-Rendering Dataset. We report the performance of novel view rendering. Our method outperforms previous state-of-the-art human modeling methods. We mark the best and the second best methods in cells. LPIPS*=LPIPS × 10^3." + ], + "table_footnote": [], + "table_body": "
Subject:1_0206_042_0007_042_0019_102_0044_112_0051_092_0813_05
Metric:PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]26.640.943950.0928.060.949256.7032.060.971031.1228.230.951237.1525.040.953442.2231.750.970131.30
GART [31]27.190.945456.4728.220.952556.4631.130.968635.1628.900.956141.2726.470.962842.2132.060.972935.74
GauHuman [22]27.960.950050.4227.930.949656.9331.630.969832.8230.380.964134.1925.990.959942.0033.430.976429.42
Ours30.810.964939.4729.630.956643.7234.970.977723.5332.620.974723.8328.450.968534.0635.810.984320.83
", + "bbox": [ + 94, + 128, + 903, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9054b106fd64f5802227d0a6ead54f294f28323c1bd85cce0257cb682b76d4c7.jpg", + "image_caption": [ + "Figure 3. Novel View Qualitative Results on DNA-Rendering. We zoom into the local region and compute the error maps compared with ground truth images. The results show that our method achieves competitive results on both the overall and local region qualities." + ], + "image_footnote": [], + "bbox": [ + 140, + 207, + 854, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method outperforms previous state-of-the-art human modeling methods across all metrics. To demonstrate the visual improvement, we compare the quality of rendering images in Fig. 3. Previous methods, which lack the temporal motion information for non-rigid deformations, are unable to predict the movement of Gaussian primitives properly, resulting in blurred renderings in regions with complex tex", + "bbox": [ + 89, + 789, + 485, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tures (scene 1_0206_04 in Fig. 3). Besides, the results in Fig. 3 show that the compared methods fail to capture the proper shape variation caused by movement (e.g., the white dress in scene 2_0051_09). Although these methods achieve a smooth appearance in these areas, the error maps show that they fail to match the actual shape closely. We leverage the hierarchical motion context to predict Gaussian defor", + "bbox": [ + 511, + 789, + 908, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13597", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2a6a4f54f86aadd0d3ac9c932cff8889dd3c76bb174f9060aadf46058cd4f3a.jpg", + "table_caption": [ + "Table 2. Novel View Rendering Quantitative Results on I3D-Human." + ], + "table_footnote": [], + "table_body": "
Subject: Metric:ID1_1ID1_2ID2_1ID3_1
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]29.780.960031.4031.290.962728.0929.630.960437.3332.730.960239.47
Dyco [8]30.810.961728.2931.450.962825.7529.410.961833.2632.500.961235.46
GauHuman [22]29.470.957139.4630.420.958538.8728.510.954649.4632.100.954653.67
Ours31.810.965927.0331.980.966027.8431.530.968530.2333.620.965134.02
", + "bbox": [ + 133, + 114, + 859, + 193 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be622cd66f8e05671fbb193d1ca59bb47ff345a8f7d0fbda373a3ec6015a235f.jpg", + "image_caption": [ + "Figure 4. Novel Pose Qualitative Results on I3D-Human. We compare the proposed method with previous SOTA approaches. The rendered images and error maps demonstrate our robustness on novel pose rendering." + ], + "image_footnote": [], + "bbox": [ + 114, + 207, + 883, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mations, allowing our method to capture detailed appearance variations caused by human motions more accurately.", + "bbox": [ + 89, + 642, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on I3D-Human. The quantitative results in Tab. 2 show that our method surpasses the previous SOTA method on most metrics, demonstrating our competitive performance on the I3D-Human dataset. Previous NeRF-based Dyco [8] utilizes pose variation as a condition to model human motions. However, it is limited to capturing only overall body motions and lacks the capacity for finer-grained modeling in local regions. For example, in ID1_1 and ID2_1 shown in Fig. 8 (in the Supp. Mat.), our method renders results that accurately reflect the true motion, while Dyco fails to model the information in regions far from the human body skeleton. Tab. 3 reports the quantitative results of novel pose rendering on the I3D-Human dataset. Our method achieves best performance among 3DGS-based human modeling methods and also enables real-time render-", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing ( $\\sim$ 45 FPS on I3D-Human) compared to NeRF-based Dyco ( $\\sim$ 0.7 FPS). Fig. 4 shows the qualitative comparisons of novel pose rendering. Different from Dyco [8], which relies solely on joint motions as conditions, our method leverages the motion state of each Gaussian primitive to predict deformations, enabling more flexible modeling of motions in regions distant from the human body.", + "bbox": [ + 511, + 642, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalizability, we also conduct experiments on ZJU-MoCap [53]. Please refer to the supplementary materials for more details.", + "bbox": [ + 511, + 750, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Out-of-Distribution Poses. To evaluate the generalization ability to novel poses, we conduct experiments using animations with large pose variations. Specifically, we use a model trained on one sequence to render target poses sampled from other unseen sequences, as different sequences in DNA-Rendering and I3D-Human contain distinct motion patterns. As illustrated in Fig. 6, the Target Pose is", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d7b24e057c3c73a0d3f38b70399850617ea366f5d266a48053f7b27dbbbb948.jpg", + "image_caption": [ + "(a). Baseline", + "(b). (a) + Vanilla Non-Rigid MLP.", + "Figure 5. Qualitative Ablation Results. We compare the visual influence of different components on ID2_1 scene of I3D-Human dataset." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 248, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77cb632348fd966e2ec4cdd19e16e83b4a1031ec1909fadc4a307f618a192f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 88, + 390, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e66811fefeb560e3053e1d6e714909949085b0110c6b098902c07685cc5c841.jpg", + "image_caption": [ + "(c). $(\\mathbf{b}) + \\Delta \\mathcal{P}$ Conditions" + ], + "image_footnote": [], + "bbox": [ + 395, + 88, + 529, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f6aca3ad615cebc7b57dd9d9f561befdb91306a9d44655b7650007386329569.jpg", + "image_caption": [ + "(d).(c) $^+$ V Conditions" + ], + "image_footnote": [], + "bbox": [ + 534, + 88, + 669, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7131523b57af1d890fc681715c8dfb5f5dd9eb7d70bbfc7fbbc72fbeb68e00af.jpg", + "image_caption": [ + "(e). (d) + STMS." + ], + "image_footnote": [], + "bbox": [ + 674, + 88, + 810, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07bf9af8d75eff5f0e90590ce72132077306b8b223db98bc4613292a359e5204.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 816, + 88, + 885, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/013197955407564466371049b0137826d77b0e8ab6c9e346bb82deaa4104c121.jpg", + "image_caption": [ + "ID3" + ], + "image_footnote": [], + "bbox": [ + 93, + 244, + 225, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c84b07fde1f292219b4f95bd820a564cdcb49ad21bd55453e2cbd3782f059b85.jpg", + "image_caption": [ + "Figure 6. Out-of-distribution Pose Animation." + ], + "image_footnote": [], + "bbox": [ + 93, + 337, + 223, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5dded8ec12a23f0aa1b9f42bdb1eb89c0300143f6228cefc56cf797a2179a1c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 244, + 276, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df8e7eaa707ddfadf0b196a0e5141ce265c2e9d66ffae5be89fe41e5f06acc5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 232, + 330, + 276, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0395ae9202ab5eaaad448ae0c2e726e5d38009f73fbe0a955aec1d1409f1cb08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 244, + 418, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb60a18e0b5b2e55e5c8ba3ec0a12a4c2a0526fbd32960f026f08c9eaa49fb6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 330, + 416, + 422 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c79bd81aa1de4777ea49ab439030998b14d44aad444f6d206f1e125ae47e8d9c.jpg", + "image_caption": [ + "2_0051_09" + ], + "image_footnote": [], + "bbox": [ + 426, + 244, + 470, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41ce47e0e9dd79d7bf75c0f64c38a965b2b6dd6ff37eedf4b12354d590ac3321.jpg", + "image_caption": [ + "1_0206_04" + ], + "image_footnote": [], + "bbox": [ + 424, + 339, + 475, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "drawn from an unseen sequence exhibiting different human motions. The results suggest that our method is capable of generalizing to these significant pose changes, even though such poses were not observed during training.", + "bbox": [ + 89, + 476, + 483, + 537 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec627ef9db7f027ab73918fed83e48b8c93a859166f985ba3cd27b5ac98ad9e6.jpg", + "table_caption": [ + "Table 3. Metrics of Novel Pose Renderings on I3D-Human." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
Dyco [8]30.090.956935.32
3DGS-Avatar [56]29.540.955538.59
GauHuman [22]29.310.952648.29
Ours30.280.958336.07
", + "bbox": [ + 150, + 575, + 419, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 666, + 243, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we ablate the influence of various components in our methods on I3D-Human dataset. Tab. 4 shows the average metrics of 4 sequences under different settings. Skeleton motion condition $\\Delta \\mathcal{P}$ denotes the overall human body movement. Compared to predicting non-rigid deformation only with the Gaussian position, incorporating this additional global temporal information allows our method to better model the overall human shape and achieve higher rendering quality, as demonstrated in Tab. 4 (c).", + "bbox": [ + 89, + 688, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-grained point-wise motion condition $\\mathcal{V}$ describes the local region's variation in a more detailed manner. Such point-based motion information captures the relationship between motion and appearance variations in areas beyond the human skeleton. Tab. 4 (d) and Fig. 5 (d) show that our", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method achieves higher rendering details in local regions. Spatio-temporal Multi-scale Sampling (STMS) is designed to enhance the robustness of human motion conditions by incorporating local region motion information and motion patterns across different temporal windows. With such a hierarchical motion context embedding design, our method is able to predict each Gaussian's deformation more robustly, leading to better rendering quality as shown in Tab. 4 (e) and Fig. 5 (e). Please refer to the supplementary materials for more detailed ablation experiments.", + "bbox": [ + 511, + 247, + 903, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2793141f88d497f19b980417c142c33ee67bf7b558a74b70228e5922db6af7e8.jpg", + "table_caption": [ + "Table 4. Quantitative Ablation Results on I3D-Human. (a) denotes experiments without non-rigid deformation. (b) only utilizes Gaussian positions $\\mathbf{x}$ and the current frame pose $P$ as the conditions for non-rigid MLP." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
(a) Baseline29.760.956938.35
(b) (a)+Vanilla Non-Rigid MLP+P Conditions.31.050.961734.35
(c) (b)+ΔP Conditions.31.890.964532.17
(d) (c)+V Conditions.32.010.965131.23
(e) (d)+STMS. (Ours)32.240.966429.78
", + "bbox": [ + 517, + 477, + 901, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 769, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The Gaussian-based representation used in our method may introduce slight blur artifacts during rendering, whereas NeRF's ray-based volumetric integration tends to produce sharper results. Moreover, our local velocity cues are derived from the coarse SMPL model rather than dense surface tracking [79], which may limit the accuracy of fine-scale garment deformation. Addressing these limitations remains an open challenge for future work.", + "bbox": [ + 511, + 598, + 905, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose a 3DGS-based framework that integrates hierarchical motion context for 3D human modeling. Specifically, the non-rigid deformation for the Gaussian primitive is learned based on the global human skeleton variations and fine-grained Gaussian's point-wise motions. To capture each Gaussian primitive's motion information more robustly, we introduce a spatio-temporal multi-scale sampling strategy to incorporate both local region motion features and motion patterns across different temporal intervals. Through the above design, the proposed method achieves state-of-the-art rendering of human avatars with complex garments and motions.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13599", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 114, + 480, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 186, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 2", + "[3] Ang Cao and Justin Johnson. Hexplane: A Fast Representation for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[5] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Lin chao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 2", + "[6] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021.", + "[7] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023.", + "[8] Yutong Chen, Yifan Zhan, Zhihang Zhong, Wei Wang, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Within the dynamic context: Inertia-aware 3d human modeling with pose sequence. arXiv preprint arXiv:2403.19160, 2024. 2, 3, 5, 7, 8, 1", + "[9] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 5, 1, 3", + "[10] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2", + "[11] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose" + ], + "bbox": [ + 93, + 186, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation and tracking from multiple views. In T-PAMI, 2021. 2", + "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2", + "[14] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2", + "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic View Synthesis from Dynamic Monocular Video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[17] Qingzhe Gao, Yiming Wang, Libin Liu, Lingjie Liu, Christian Theobalt, and Baoquan Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[18] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, pages 14346-14355, 2021. 2", + "[19] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 2", + "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2", + "[21] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2", + "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 1, 2, 3, 5, 6, 7, 8", + "[23] Tao Hu, Fangzhou Hong, and Ziwei Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13600", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375, 2022. 2", + "[25] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2", + "[26] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 2", + "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5", + "[28] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In ACM SIGGRAPH 2007 papers, pages 64-es. 2007. 1", + "[29] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 2", + "[30] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2", + "[31] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 5, 6", + "[32] Mingwei Li, Jiachen Tao, Zongxin Yang, and Yi Yang. Human101: Training 100+ fps human gaussians in 100s from 1 view. arXiv preprint arXiv:2312.15258, 2023.", + "[33] Mengtian Li, Shengxiang Yao, Zhifeng Xie, Keyu Chen, and Yu-Gang Jiang. Gaussianbody: Clothed human reconstruction via 3d gaussian splatting. arXiv preprint arXiv:2401.09720, 2024. 2", + "[34] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3D Video Synthesis from Multi-view Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2", + "[35] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2", + "[36] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, pages 5741-5751, 2021. 2", + "[38] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2", + "[39] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. pages 15651-15663, 2020. 2", + "[40] Xinqi Liu, Chenming Wu, Xing Liu, Jialun Liu, Jinbo Wu, Chen Zhao, Haocheng Feng, Errui Ding, and Jingdong Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 2", + "[41] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 2", + "[42] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2", + "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Spl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3", + "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2", + "[45] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 2", + "[46] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[47] Muyao Niu, Yifan Zhan, Qingtian Zhu, Zhuoxiao Li, Wei Wang, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Bundle adjusted gaussian avatars deblurring. arXiv preprint arXiv:2411.16758, 2024. 2", + "[48] Muyao Niu, Mingdeng Cao, Yifan Zhan, Qingtian Zhu, Mingze Ma, Jiancheng Zhao, Yanhong Zeng, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025. 2", + "[49] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 2", + "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3", + "[52] Rui Peng, Wangze Xu, Luyang Tang, Jianbo Jiao, Ronggang Wang, et al. Structure consistent gaussian splatting with matching prior for few-shot novel view synthesis. Advances in Neural Information Processing Systems, 37:97328-97352, 2024. 2", + "[53] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 7, 1", + "[54] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7964-7976, 2023. 1", + "[55] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-neRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2", + "[56] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 1, 2, 3, 5, 6, 7, 8", + "[57] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2", + "[58] Shihe Shen, Huachen Gao, Wangze Xu, Rui Peng, Luyang Tang, Kaiqiang Xiong, Jianbo Jiao, and Ronggang Wang. Disentangled generation and aggregation for robust radiance fields. In European Conference on Computer Vision, pages 218-236. Springer, 2024. 2", + "[59] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2", + "[60] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2", + "[61] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21096-21106, 2025. 2", + "[63] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2", + "[64] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, pages 1-19. Springer, 2022. 2", + "[65] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2", + "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3, 5", + "[67] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2", + "[68] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2, 5", + "[69] Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. Swift4d: Adaptive divide-and-conquer gaussian splatting for compact and efficient reconstruction of dynamic scene. arXiv preprint arXiv:2503.12307, 2025. 2", + "[70] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time Neural Irradiance Fields for Free-viewpoint Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2", + "[71] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 2", + "[72] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In CVPR, 2024. 2", + "[73] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13602", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Jinbo Yan, Rui Peng, Zhiyan Wang, Luyang Tang, Jiayu Yang, Jie Liang, Jiahao Wu, and Ronggang Wang. Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16520-16531, 2025. 2", + "[75] Yifan Zhan, Qingtian Zhu, Muyao Niu, Mingze Ma, Jiancheng Zhao, Zhihang Zhong, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Tomie: Towards modular growth in enhanced smpl skeleton for 3d human with animatable garments. arXiv preprint arXiv:2410.08082, 2024. 2", + "[76] Yifan Zhan, Wangze Xu, Qingtian Zhu, Muyao Niu, Mingze Ma, Yifei Liu, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. R3-athlete: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 2", + "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[78] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2", + "[79] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 2, 8", + "[80] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, pages 145-163. Springer, 2024. 2", + "[81] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 654 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13603", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_model.json b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2310a639110e4e9274cc4628c1e08bf85ec43558 --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_model.json @@ -0,0 +1,2842 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "Sequential Gaussian Avatars with Hierarchical Motion Context" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.18, + 0.787, + 0.22 + ], + "angle": 0, + "content": "Wangze Xu\\(^{1*}\\) Yifan Zhan\\(^{1,2}\\) Zhihang Zhong\\(^{1\\dagger}\\) Xiao Sun\\(^{1\\dagger}\\) \n\\(^{1}\\)Shanghai Artificial Intelligence Laboratory \\(^{2}\\)The University of Tokyo" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.236, + 0.885, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.483, + 0.908, + 0.555 + ], + "angle": 0, + "content": "Figure 1. Illustration of the Hierarchical Motion Context Design. We model the motion-dependent appearance variations with both coarse skeleton condition \\(\\Delta P\\) and fine-grained point-wise velocity condition \\(\\mathcal{V}\\). \\(\\Delta P\\) describes the overall human skeleton motion, which is derived from the difference between the human poses at adjacent frames. \\(\\mathcal{V}\\) is the point-wise velocity that indicates finer-grained motion in local regions. To achieve robust deformation, we propose Spatio-temporal Multi-scale Sampling, which samples the overall motion trend and inter-frame details via diverse time intervals \\(s\\) for \\(\\Delta P\\) and \\(\\mathcal{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.57, + 0.328, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.604, + 0.486, + 0.862 + ], + "angle": 0, + "content": "The emergence of neural rendering has significantly advanced the rendering quality of 3D human avatars, with the recently popular 3DGS technique enabling real-time performance. However, SMPL-driven 3DGS human avatars still struggle to capture fine appearance details due to the complex mapping from pose to appearance during fitting. In this paper, we propose SeqAvatar, which excavates the explicit 3DGS representation to better model human avatars based on a hierarchical motion context. Specifically, we utilize a coarse-to-fine motion conditions that incorporate both the overall human skeleton and fine-grained vertex motions for non-rigid deformation. To enhance the robustness of the proposed motion conditions, we adopt a spatiotemporal multi-scale sampling strategy to hierarchically integrate more motion clues to model human avatars. Extensive experiments demonstrate that our method significantly outperforms 3DGS-based approaches and renders human" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.907, + 0.648 + ], + "angle": 0, + "content": "avatars orders of magnitude faster than the latest NeRF-based models that incorporate temporal context, all while delivering performance that is at least comparable or even superior. Project page: https://zezeaaa.github.io/projects/SeqAvatar/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.678, + 0.646, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Recent research on digital humans has highlighted the efficiency of 3D Gaussian Splatting (3DGS) [27], demonstrating its capability for high-quality and real-time rendering. By defining T-posed human Gaussians in a canonical space and employing specific warping techniques, we can render human avatars from diverse perspectives and in any pose. The canonical Gaussian primitives and warping weights are then jointly optimized under supervision from video inputs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Animatable human avatar reconstruction is primarily hindered by limited modeling of non-rigid warping, forcing a trade-off between rendering quality and animation capability. Current methods [22, 56] incorporate SMPL(-X) pose priors [43, 51] to guide motion for each observation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done during the author's internship at Shanghai Artificial Intelligence Laboratory. † denotes co-corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "and rely on Linear Blend Skinning (LBS) for warping. Although these approaches are effective for pose-driven animation, they often struggle with per-frame non-rigid warping in scenarios involving complex garments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.38 + ], + "angle": 0, + "content": "We experimentally find that current human pose conditions [22, 56] do not fully capture the complex one-to-many mapping from pose to appearance, particularly for 3DGS-based approaches [27]. These methods [56, 68] rely on the spatial information of a template body in each frame's human pose to predict non-rigid deformations. However, they often overlook local details, such as garment deformations far from the skeleton, and cannot resolve cases where the same pose corresponds to different appearances during complex motions. Furthermore, although previous NeRF-based [44] method [8] has attempted to model motion sequences using human pose residuals, the inherently global nature of the pose sequence limits the ability to capture finer motion details. Naive pose sequence modeling does not fully account for the explicit characteristics of 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.713 + ], + "angle": 0, + "content": "In this paper, we introduce a hierarchical motion context condition for 3DGS-based human avatar modeling to address the limitations of relying solely on human pose, which provides only limited global skeletal information. Our approach improves the ability of 3DGS-based methods to accurately capture the complex relationship between human pose and appearance in challenging scenarios. Specifically, we design a coarse-to-fine motion condition that incorporates both overall skeletal movements and fine-grained point-wise motions. Leveraging the explicit nature of Gaussian primitives, this condition seamlessly integrates into 3DGS-based methods, enabling more precise predictions of complex non-rigid deformations. To further enhance robustness, we propose a spatio-temporal multi-scale sampling strategy for constructing the motion context with a larger receptive field. Spatially, we consider the motion states of neighboring points within the local region of each Gaussian primitive to obtain more stable motion embeddings. Temporally, we capture human motion patterns across multiple time scales, combining long-term trends with fine-grained inter-frame motion details. This improves the model's generalization to complex human movements." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.716, + 0.453, + 0.73 + ], + "angle": 0, + "content": "To summarize, our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.732, + 0.483, + 0.793 + ], + "angle": 0, + "content": "1) We propose a novel hierarchical motion condition that integrates coarse-to-fine human motion, combining global skeletal poses with localized vertex residuals to enhance non-rigid deformation prediction." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.794, + 0.483, + 0.837 + ], + "angle": 0, + "content": "2) We introduce a spatiotemporal multiscale sampling strategy that expands the receptive field of hierarchical motion context, improving generalization to complex motions." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.838, + 0.483, + 0.898 + ], + "angle": 0, + "content": "3) Experiments on the I3D-Human [8], DNA-Rendering [9], and ZJU-MoCap [53] datasets show the effectiveness of the proposed SeqAvatar, which is capable of modeling details of the human body in complex motions." + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.732, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Neural Rendering. Neural rendering techniques have brought significant progress to human reconstruction and rendering. In particular, Neural Radiance Fields (NeRF) [44] introduces an implicit scene representation that models color and density using multilayer perceptrons, delivering photorealistic rendering results. More recently, point-based 3D Gaussian Splatting [27] utilizes 3D Gaussians to represent scenes explicitly, achieving real-time high-quality rendering. Building upon NeRF and 3DGS, subsequent works have advanced neural rendering across various dimensions, e.g., enhancing visual fidelity [1, 2], improving sparse-view reconstruction quality [10, 52, 63, 71, 80], enabling pose-free optimization [37, 58, 61, 67], modeling dynamic scenes [13, 34, 49, 55, 69, 74], and accelerating training and inference [4, 12, 18, 39, 46, 60]. These developments also benefit human avatar modeling, which demands high-quality and efficient rendering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.905, + 0.78 + ], + "angle": 0, + "content": "SMPL(-X)-Based Neural Human Modeling. The SMPL(-X) family [43, 51] provides a parametric representation of the human body by decomposing it into pose-related and shape-related components using 3D mesh scans and principal component analysis (PCA). Its pose blend shapes enables body deformation through joint-wise pose blending, offering an efficient and compact way for animation. SMPL(-X) has thus become a cornerstone in human body modeling and animation, with many methods [11, 59] estimating its parameters directly from 2D inputs. Recent neural human reconstruction methods integrate SMPL(-X) with implicit representations such as NeRF [44] and 3DGS [27] to enable animatable avatars with high rendering quality. Different from purely 2D image animation methods [48, 62, 65, 73], these methods typically register the input data to a canonical T-pose space and use linear blend skinning (LBS) to transform 3D points into observation space based on SMPL(-X) poses. NeRF-based approaches [5-8, 14, 17, 19, 20, 30, 53, 64, 68, 75] focus on monocular or multi-view reconstruction. More recently, 3DGS-based methods [21-23, 25, 26, 29, 31-33, 36, 40, 41, 45, 47, 56, 76, 78, 79, 81] have gained popularity due to their real-time rendering and high fidelity. Some works [8, 23] further incorporate pose sequences as temporal context, but they suffer from slow rendering and lack fine details of motion caused by simply integrating coarse human pose embedding with NeRF." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Human Rendering with Temporal Embeddings The earliest modeling of dynamic radiance field [3, 13, 15, 16, 24, 34, 35, 42, 50, 57, 70] can be traced back to temporal embeddings in general scenes, where each frame's observation is obtained by constructing a canonical space and a time-conditioned deformation field. Motivated by these, a stream of research [38, 72] focuses on pure rendering quality, employing temporal embeddings instead of human pose to en" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "code each frame of human videos. Although these methods achieve high-quality, temporally continuous rendering, they struggle with the lack of geometric constraints from human pose, making parameterized animation unfeasible." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.218, + 0.183 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.485, + 0.299 + ], + "angle": 0, + "content": "SMPL(-X) Series [43, 51] adopt a parameterized framework to represent human bodies across diverse shapes and poses. For each frame, the human mesh is derived by deforming a canonical template mesh based on shape and pose parameters. Specifically, a 3D point \\(\\mathbf{x}\\) on the canonical mesh is warped to the corresponding point on the deformed mesh in the observation space, following" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.309, + 0.484, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L B S} (\\mathbf {x}, \\mathbf {B} _ {k}, \\omega_ {k} (\\mathbf {x})) = \\sum_ {k = 1} ^ {K} \\omega_ {k} (\\mathbf {x}) \\mathbf {B} _ {k} \\mathbf {x}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.481 + ], + "angle": 0, + "content": "where \\(K\\) is the total bone number and \\(\\mathbf{B}_k\\) is the transformation matrix for each bone. Specifically, \\(\\mathbf{B}_k\\) consists of a rotation and translation matrix \\(\\mathbf{R}_k\\) and \\(\\mathbf{b}_k\\). Here, \\(\\mathbf{R}_k\\) represents the global rotation of each joint, influenced by the local joint rotations \\(\\boldsymbol{r}\\), while \\(\\mathbf{b}_k\\) denotes each joint's translation, determined by the joint positions \\(\\boldsymbol{j}\\) and the human shape parameters \\(\\beta\\). The linear blending weight \\(\\omega^k\\) depends on \\(\\mathbf{x}\\) and is regressed from comprehensive human meshes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.588 + ], + "angle": 0, + "content": "3D Gaussian Splatting [27] utilizes a set of point-based Gaussian primitives to explicitly represent scenes, enabling real-time and high-quality rendering. Each Gaussian primitive is defined by its center position \\(\\mathbf{x} \\in \\mathbb{R}^3\\) and covariance matrix \\(\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\). To ensure positive semi-definiteness and simplify optimization, the covariance matrix \\(\\boldsymbol{\\Sigma}\\) is decomposed into a rotation matrix \\(\\mathbf{R}\\) and a scaling matrix \\(\\mathbf{S}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.598, + 0.482, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}\\) and \\(\\mathbf{S}\\) are derived by a scaling vector \\(\\mathbf{s} \\in \\mathbb{R}^3\\) and a quaternion vector \\(\\mathbf{r} \\in \\mathbb{R}^4\\) in practice. Additionally, each Gaussian primitive is also assigned a color feature \\(sh \\in \\mathbb{R}^k\\) represented by spherical harmonics (SH) and an opacity \\(\\alpha\\) for rendering. During the rendering process, a splatting technique [27] is used to project each Gaussian primitive onto the 2D image space with a viewing transform \\(\\mathbf{W}\\) and the Jacobian \\(\\mathbf{J}\\) of the projective transformation's affine approximation. The transformed covariance \\(\\boldsymbol{\\Sigma}'\\) in the camera's coordinates is" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.789, + 0.483, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.484, + 0.851 + ], + "angle": 0, + "content": "After projection, the pixel color \\(C\\) is computed by blending \\(N\\) ordered Gaussian primitives overlapping at the pixel:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.862, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "where \\(c_{i}\\) is computed from SH feature \\(sh\\) and \\(\\alpha_{i}^{\\prime}\\) is the product of \\(\\alpha_{i}\\) and probability density of \\(i\\)-th 2D Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.152 + ], + "angle": 0, + "content": "For optimization, a photometric loss is defined by a combination of \\( L_{1} \\) and SSIM [66] losses:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.162, + 0.905, + 0.181 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p h o t o}} = \\lambda L _ {1} (\\hat {I}, I) + (1 - \\lambda) (1 - S S I M (\\hat {I}, I)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.193, + 0.905, + 0.223 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) and \\(I\\) denote the rendered and ground-truth images, and \\(\\lambda\\) controls the balance between the two terms." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.236, + 0.606, + 0.252 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.262, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Fig. 2 shows the framework of our method. We use the explicit point-based 3DGS as the representation of the human body. Given a collection of input cameras and images, we optimize a set of Gaussian primitives \\(\\{\\mathcal{G}_i\\}_{i=1}^{n}\\) to fit the body's shape and appearance. Each Gaussian primitive \\(\\mathcal{G}_i\\) includes the center position \\(\\mathbf{x}\\), scaling vector \\(\\mathbf{s}\\), rotation quaternion vector \\(\\mathbf{r}\\), color feature \\(sh\\) and opacity \\(\\alpha\\), where \\(\\mathbf{x}\\) is initialized from the SMPL template vertices \\(\\{\\mathbf{T}_i\\}_{i=1}^{i=N}\\). During the optimization process, we first apply a coarse-to-fine motion context to capture more accurate and fine-grained details (Sec. 4.1). To mitigate overfitting, we propose Spatio-Temporal Multi-Scale Sampling to obtain more robust point-wise Gaussian motion, which serves as the embedding for non-rigid deformation (Sec.4.2). Next, we adopt Linear Blend Skinning (LBS) to map canonical Gaussian primitives \\(\\mathcal{G}\\) to observation space and render images via differentiable splitting (Sec.4.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.905, + 0.558 + ], + "angle": 0, + "content": "4.1. Non-Rigid Deformation with Coarse-to-fine Motion Context" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Coarse Skeleton Motion. As illustrated in Fig. 1, we introduce a coarse-to-fine motion context to excavate additional conditions for non-rigid deformation. Typically, most human 3D avatar methods use the current frame pose as a condition to predict non-rigid deformation [22, 56]. While this provides spatial information to distinguish the motion of different body parts, it fails to capture temporal motion changes, as discussed in [8]. Therefore, given a frame at time \\( t \\), we consider a sequence of regularly interval-sampled frames \\( \\mathcal{T} \\) to model inter-frame body motion variations:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.729, + 0.905, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} = \\{t - s, t - 2 s, \\dots , t - L s \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.756, + 0.907, + 0.817 + ], + "angle": 0, + "content": "where \\(L\\) is the sequence length and \\(s\\) is the time interval. The coarse motion of body skeletons at each time step \\(t \\in \\mathcal{T}\\) can be derived by calculating the difference of poses between adjacent frames as illustrated in Fig. 1, following" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.827, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathcal {P} = \\left\\{\\Delta P ^ {t} = \\delta \\left(P ^ {t}, P ^ {t - s}\\right) | t \\in \\mathcal {T} \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.855, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(P \\in \\mathbb{R}^{K \\times 3}\\) is the body pose and \\(\\delta\\) denotes the difference \\(\\Delta \\mathcal{P}\\) between two poses in axis-angle form. We employ an MLP \\(\\mathcal{E}_{\\Delta \\mathcal{P}}\\) to encode the sequential skeleton mo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.885, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.307, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. We first initialize canonical Gaussian positions \\(\\mathbf{x}\\) with SMPL template vertexes. For each Gaussian, we derive both coarse skeleton motion condition \\(f_{\\Delta \\mathcal{P}}\\) and fine-grained vertex motion condition \\(f_{\\nu}\\) that sampled from the vertex motion template \\(\\mathcal{F}_{\\nu}\\) (points in different colors represent different motions). Based on such hierarchical motion information, we utilize an MLP \\(\\mathcal{E}_{non-rigid}\\) to better predict each Gaussian's non-rigid deformation prediction. The non-rigid deformed Gaussians are then warped into observation space via the standard LBS transformation for rendering." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.482, + 0.434 + ], + "angle": 0, + "content": "tion \\(\\Delta \\mathcal{P}\\in \\mathbb{R}^{L\\times K\\times 3}\\), flattened along the temporal dimension, into a skeleton motion embedding \\(f_{\\Delta \\mathcal{P}}\\in \\mathbb{R}^{32}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.442, + 0.482, + 0.459 + ], + "angle": 0, + "content": "\\[\nf _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} (\\Delta \\mathcal {P}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.483, + 0.51 + ], + "angle": 0, + "content": "which serves as a condition for the subsequent non-rigid deformation. Further details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Fine Vertex Motion. Compared to previous implicit NeRF-based methods, the point-based 3DGS representation enables us to explore more fine-grained temporal motion information. We derive a point-wise velocity vector \\( v_{i} \\) for each Gaussian primitive \\( \\mathcal{G}_i \\) to model the fine-grained localized body motion, which is beyond the scope of pose skeleton motion \\( \\Delta \\mathcal{P} \\). By measuring the position variations across adjacent time steps, we calculate per-frame velocity" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.638, + 0.482, + 0.669 + ], + "angle": 0, + "content": "\\[\nv ^ {t} = \\frac {\\mathbf {x} _ {\\mathbf {o}} ^ {t} - \\mathbf {x} _ {\\mathbf {o}} ^ {t - s}}{s}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{x_o}\\) denotes the warped coordinates in observation space. However, since the positions of Gaussian primitives are updated during optimization, it is not stable to obtain \\(v_{i}\\) based on such a dynamic variable. Moreover, transforming the canonical Gaussians into the observation space requires non-rigid deformation first, which leads to a circular dependency and conflicts with the current velocity design. To this end, we introduce a motion template field \\(\\mathcal{F}_{\\mathrm{V}} = \\{\\mathbf{V}_i\\}_{i = 1}^{i = N}\\) that stores the velocity of each SMPL vertex, and \\(v_{i}\\) for each Gaussian primitive is sampled from \\(\\mathcal{F}_{\\mathrm{V}}\\) based on the distance between query points and vertexes in this template. Specifically, given a time step \\(t\\) and the corresponding body pose \\(P^t = \\{(\\mathbf{R}_k^t,\\mathbf{b}_k^t)|k\\in K\\}\\), the SMPL template vertexes in \\(\\mathbf{T}\\) can be warped into observation space with the template skinning weights \\(\\mathbf{W}\\) and" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.404, + 0.788, + 0.42 + ], + "angle": 0, + "content": "standard linear blend skinning, following" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.427, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\mathbf {o}} ^ {t} = \\mathbf {L B S} (\\mathbf {T}, \\mathbf {B} ^ {t}, \\mathbf {W}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.512 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}^t\\) is the transformation matrix derived from rotation and translation matrix \\(\\mathbf{R}^t\\) and \\(\\mathbf{b}^t\\). \\(\\mathbf{W}\\) is the template SMPL LBS weights. The velocity \\(\\mathbf{V}\\) of each template vertexes \\(\\mathbf{T}\\) can be further derived as" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.517, + 0.905, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} ^ {t} = \\frac {\\mathbf {T} _ {\\mathbf {o}} ^ {t} - \\mathbf {T} _ {\\mathbf {o}} ^ {t - s}}{s}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.556, + 0.906, + 0.587 + ], + "angle": 0, + "content": "We then sample each Gaussian's velocity from this motion template field \\(\\mathcal{F}_{\\mathbf{V}}\\), which is discussed in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.594, + 0.906, + 0.612 + ], + "angle": 0, + "content": "4.2. Spatio-temporal Multi-scale Sampling (STMS)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.692 + ], + "angle": 0, + "content": "To enhance the robustness of human motion conditions, we propose modeling both local region motion information of the human body and motion patterns across different temporal windows to mitigate overfitting and improve generalization by capturing more comprehensive motion dynamics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.768 + ], + "angle": 0, + "content": "In the spatial dimension, we sample the \\(\\tau\\) nearest template to model the body's local region motion more robustly. Specifically, for each canonical Gaussian primitive \\(\\mathcal{G}_i\\), the \\(\\tau\\) nearest vertexes' velocities are sampled as input to an MLP \\(\\mathcal{E}_{knn}\\) to learn a local point-wise motion embedding" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.775, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {t} = \\mathcal {E} _ {k n n} \\left(\\left\\{\\mathbf {V} _ {j} ^ {t} \\right\\}\\right), \\quad j \\in \\operatorname {K N N} (\\mathbf {T}, \\mathbf {x} _ {i}), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.906, + 0.877 + ], + "angle": 0, + "content": "where \\(\\mathbf{KNN}(\\mathbf{T},\\mathbf{x}_i)\\) denotes the \\(\\tau\\) nearest SMPL template vertexes of the canonical Gaussian position \\(\\mathbf{x}_i\\). Similar to Eq. 7, we then apply an MLP \\(\\mathcal{E}_{\\mathcal{V}}\\) to encode each Gaussian primitive's sequential motion embedding \\(\\mathcal{V} = \\{e^{t}|t\\in \\mathcal{T}\\}\\) into a point-wise sequential condition \\(f_{\\mathcal{V}}\\in \\mathbb{R}^{96}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.664, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {V}). \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "In the temporal dimension, to capture both the overall motion trend and the inter-frame motion details, we adopt a multi-scale sequence sampling strategy. In detail, we sample a series of sequences at several progressively increasing intervals to get human motion information across different temporal windows" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.192, + 0.483, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{s = s _ {0} + i \\Delta s \\right\\} _ {i = 0} ^ {i = m}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.485, + 0.295 + ], + "angle": 0, + "content": "where \\(\\Delta s\\) is the increasing rate of sampling interval \\(s\\). We then input the multi-scale sampled sequence motions into the skeleton motion encoder \\(\\mathcal{E}_{\\Delta P}\\) and the point-wise motion encoder \\(\\mathcal{E}_{V}\\) to obtain hierarchical temporal motion embeddings. Therefore, Eq. 7 and Eq. 13 can be revised as" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.304, + 0.483, + 0.32 + ], + "angle": 0, + "content": "\\[\nf _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} \\left(\\left\\{\\Delta \\mathcal {P} _ {s} \\right\\}\\right), \\quad s \\in \\mathcal {S} \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.323, + 0.483, + 0.34 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\{\\mathcal {V} _ {s} \\}), \\quad s \\in \\mathcal {S}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.424 + ], + "angle": 0, + "content": "In practice, we concatenate all skeleton motion conditions, \\(\\Delta \\mathcal{P}_s\\) for \\(s\\in S\\), and localized vertex motion conditions, \\(\\nu_{s}\\) for \\(s\\in S\\), across different sampling scales. These are then input to \\(\\varepsilon_{\\Delta P}\\) and \\(\\varepsilon_{\\nu}\\), respectively. More details are shown in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Non-Rigid Deformation. Given the coarse skeleton motion \\( f_{\\Delta \\mathcal{P}} = \\mathcal{E}_{\\Delta \\mathcal{P}}(\\{\\Delta \\mathcal{P}_s\\}) \\) condition and fine point-wise velocity condition \\( f_{\\mathcal{V}} = \\mathcal{E}_{\\mathcal{V}}(\\{\\mathcal{V}_s\\}) \\), we utilize an MLP to predict each Gaussian's non-rigid deformation as" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.494, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\delta \\mathbf {x}, \\delta \\mathbf {s}, \\delta \\mathbf {r} = \\mathcal {E} _ {\\text {n o n - r i g i d}} (\\mathbf {x}, P, f _ {\\Delta \\mathcal {P}}, f _ {v}). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.354, + 0.535 + ], + "angle": 0, + "content": "The deformed canonical Gaussian \\(\\mathcal{G}'\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.545, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = \\mathbf {x} + \\delta \\mathbf {x}, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.565, + 0.482, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} ^ {\\prime} = \\mathbf {s} + \\delta \\mathbf {s}, \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.583, + 0.482, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} ^ {\\prime} = \\mathbf {r} \\cdot \\delta \\mathbf {r}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.609, + 0.449, + 0.624 + ], + "angle": 0, + "content": "where \\(\\cdot\\) denotes the multiplication of two quaternions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.231, + 0.649 + ], + "angle": 0, + "content": "4.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Rigid Deformation. We utilize the standard LBS operation to map the non-rigid deformed Gaussians \\(\\mathcal{G}_o\\) into the observation space, following" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.71, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathbf {o}} = \\mathbf {L B S} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {B}, \\omega\\right), \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.73, + 0.483, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {\\mathbf {o}} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\mathbf {x} ^ {\\prime}\\right) \\mathbf {B} _ {k} \\mathbf {R} ^ {\\prime}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.884 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}'\\) is the rotation matrix derived from the non-rigid deformed Gaussian's rotation quaternion \\(\\mathbf{r}'\\), and \\(\\mathbf{R_o}\\) represents the rotation matrix in observation space. Following the previous method [22], we utilize an MLP \\(\\mathcal{E}_{lbs}\\) to predict the LBS weight offsets for each query canonical point and update the sampled weights from the nearest SMPL vertex as" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\omega_ {k} (\\mathbf {x}) = \\omega_ {k} ^ {S M P L} (\\mathbf {x}) + \\mathcal {E} _ {l b s} (\\mathbf {x}). \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "Similar to [8, 22, 68], we introduce a pose refinement MLP \\(\\mathcal{E}_{\\text {pose }}\\) to refine the pose estimate from SMPL for a better fit to the human body." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.183 + ], + "angle": 0, + "content": "Loss Function. With the transformed Gaussian primitives \\(\\mathcal{G}_o\\) in observation space, we apply the standard splitting [27] to render images as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.195, + 0.907, + 0.212 + ], + "angle": 0, + "content": "\\[\nI = \\mathbf {S p l a t t i n g} \\left(\\mathbf {x} _ {\\mathbf {o}}, \\mathbf {R} _ {\\mathbf {o}}, s ^ {\\prime}, \\alpha , s h\\right). \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.223, + 0.906, + 0.253 + ], + "angle": 0, + "content": "During the optimization process, we employ a combination of loss functions as supervision, summarized as" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.266, + 0.907, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {l p i p s}} + \\mathcal {L} _ {\\text {m a s k}}, \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.294, + 0.906, + 0.37 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is used to balance the weight of different losses. \\(\\mathcal{L}_{mask}\\) [22] is an \\(L_{2}\\) loss between the rendered \\(\\alpha\\) and human body mask. Similar to [56], we apply \\(\\mathcal{L}_{isopos}\\) and \\(\\mathcal{L}_{isocov}\\) to control the Gaussian primitive's position and covariance. Details are shown in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.383, + 0.646, + 0.4 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.615, + 0.422 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.906, + 0.52 + ], + "angle": 0, + "content": "DNA-Rendering [9] is a challenging human-centric rendering dataset. It contains diverse scenes from everyday life to professional occasions. We use 6 sequences with loose-fitting garments (1_0206_04, 2_0007_04, 2_0019_10, 2_0044_11, 2_0051_09, 2_0813_05) for experiments. We adopt 24 views for training and 6 views for testing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.906, + 0.612 + ], + "angle": 0, + "content": "13D-Human [8] contains scenes closer to daily life, which comprises multi-view video frames of humans with loose clothing and complex movements. We conduct experiments on 4 sequences and use 4-5 views for training and the rest for testing. All training and testing data are retained to be the same in comparisons following Dyco's [8] data split." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.719, + 0.635 + ], + "angle": 0, + "content": "5.2. Baselines and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.841 + ], + "angle": 0, + "content": "We compare against the state-of-the-art human modeling methods including Dyco [8], 3DGS-Avatar [56], GART [31], and GauHuman [22]. Dyco is an implicit NeRF-based method that adopts pose sequence information for temporal modeling. We compare performance with Dyco to showcase the better efficacy of our fine-grained motion condition design. 3DGS-Avatar, GART, and GauHuman are explicit 3DGS-based methods, and all of them are originally designed for monocular inputs. For fairness, we extend them to multi-view input under the same settings. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [66], and learned perceptual image patch similarity (LPIPS) [77]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.849, + 0.646, + 0.865 + ], + "angle": 0, + "content": "5.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Comparison on DNA-Rendering. Tab. 1 shows the quantitative results on DNA-Rendering dataset. The proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13596" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. Quantitative Results on DNA-Rendering Dataset. We report the performance of novel view rendering. Our method outperforms previous state-of-the-art human modeling methods. We mark the best and the second best methods in cells. LPIPS*=LPIPS × 10^3." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.193 + ], + "angle": 0, + "content": "
Subject:1_0206_042_0007_042_0019_102_0044_112_0051_092_0813_05
Metric:PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]26.640.943950.0928.060.949256.7032.060.971031.1228.230.951237.1525.040.953442.2231.750.970131.30
GART [31]27.190.945456.4728.220.952556.4631.130.968635.1628.900.956141.2726.470.962842.2132.060.972935.74
GauHuman [22]27.960.950050.4227.930.949656.9331.630.969832.8230.380.964134.1925.990.959942.0033.430.976429.42
Ours30.810.964939.4729.630.956643.7234.970.977723.5332.620.974723.8328.450.968534.0635.810.984320.83
" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.208, + 0.856, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.735, + 0.908, + 0.765 + ], + "angle": 0, + "content": "Figure 3. Novel View Qualitative Results on DNA-Rendering. We zoom into the local region and compute the error maps compared with ground truth images. The results show that our method achieves competitive results on both the overall and local region qualities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.486, + 0.898 + ], + "angle": 0, + "content": "method outperforms previous state-of-the-art human modeling methods across all metrics. To demonstrate the visual improvement, we compare the quality of rendering images in Fig. 3. Previous methods, which lack the temporal motion information for non-rigid deformations, are unable to predict the movement of Gaussian primitives properly, resulting in blurred renderings in regions with complex tex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.909, + 0.898 + ], + "angle": 0, + "content": "tures (scene 1_0206_04 in Fig. 3). Besides, the results in Fig. 3 show that the compared methods fail to capture the proper shape variation caused by movement (e.g., the white dress in scene 2_0051_09). Although these methods achieve a smooth appearance in these areas, the error maps show that they fail to match the actual shape closely. We leverage the hierarchical motion context to predict Gaussian defor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13597" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.283, + 0.089, + 0.713, + 0.104 + ], + "angle": 0, + "content": "Table 2. Novel View Rendering Quantitative Results on I3D-Human." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.115, + 0.861, + 0.194 + ], + "angle": 0, + "content": "
Subject: Metric:ID1_1ID1_2ID2_1ID3_1
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]29.780.960031.4031.290.962728.0929.630.960437.3332.730.960239.47
Dyco [8]30.810.961728.2931.450.962825.7529.410.961833.2632.500.961235.46
GauHuman [22]29.470.957139.4630.420.958538.8728.510.954649.4632.100.954653.67
Ours31.810.965927.0331.980.966027.8431.530.968530.2333.620.965134.02
" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.208, + 0.885, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.588, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 4. Novel Pose Qualitative Results on I3D-Human. We compare the proposed method with previous SOTA approaches. The rendered images and error maps demonstrate our robustness on novel pose rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.674 + ], + "angle": 0, + "content": "mations, allowing our method to capture detailed appearance variations caused by human motions more accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison on I3D-Human. The quantitative results in Tab. 2 show that our method surpasses the previous SOTA method on most metrics, demonstrating our competitive performance on the I3D-Human dataset. Previous NeRF-based Dyco [8] utilizes pose variation as a condition to model human motions. However, it is limited to capturing only overall body motions and lacks the capacity for finer-grained modeling in local regions. For example, in ID1_1 and ID2_1 shown in Fig. 8 (in the Supp. Mat.), our method renders results that accurately reflect the true motion, while Dyco fails to model the information in regions far from the human body skeleton. Tab. 3 reports the quantitative results of novel pose rendering on the I3D-Human dataset. Our method achieves best performance among 3DGS-based human modeling methods and also enables real-time render-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.748 + ], + "angle": 0, + "content": "ing (\\(\\sim\\) 45 FPS on I3D-Human) compared to NeRF-based Dyco (\\(\\sim\\) 0.7 FPS). Fig. 4 shows the qualitative comparisons of novel pose rendering. Different from Dyco [8], which relies solely on joint motions as conditions, our method leverages the motion state of each Gaussian primitive to predict deformations, enabling more flexible modeling of motions in regions distant from the human body." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.794 + ], + "angle": 0, + "content": "To evaluate the generalizability, we also conduct experiments on ZJU-MoCap [53]. Please refer to the supplementary materials for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Out-of-Distribution Poses. To evaluate the generalization ability to novel poses, we conduct experiments using animations with large pose variations. Specifically, we use a model trained on one sequence to render target poses sampled from other unseen sequences, as different sequences in DNA-Rendering and I3D-Human contain distinct motion patterns. As illustrated in Fig. 6, the Target Pose is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.249, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.187, + 0.206, + 0.196 + ], + "angle": 0, + "content": "(a). Baseline" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.089, + 0.391, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.186, + 0.388, + 0.196 + ], + "angle": 0, + "content": "(b). (a) + Vanilla Non-Rigid MLP." + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.089, + 0.531, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.186, + 0.509, + 0.196 + ], + "angle": 0, + "content": "(c). \\((\\mathbf{b}) + \\Delta \\mathcal{P}\\) Conditions" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.089, + 0.671, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.186, + 0.647, + 0.196 + ], + "angle": 0, + "content": "(d).(c) \\(^+\\) V Conditions" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.089, + 0.812, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.186, + 0.777, + 0.196 + ], + "angle": 0, + "content": "(e). (d) + STMS." + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.089, + 0.886, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.186, + 0.858, + 0.195 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.207, + 0.904, + 0.222 + ], + "angle": 0, + "content": "Figure 5. Qualitative Ablation Results. We compare the visual influence of different components on ID2_1 scene of I3D-Human dataset." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.246, + 0.226, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.332, + 0.168, + 0.337 + ], + "angle": 0, + "content": "ID3" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.338, + 0.225, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.246, + 0.277, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.233, + 0.331, + 0.277, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.246, + 0.419, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.332, + 0.418, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.246, + 0.472, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.331, + 0.475, + 0.339 + ], + "angle": 0, + "content": "2_0051_09" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.34, + 0.476, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.415, + 0.468, + 0.424 + ], + "angle": 0, + "content": "1_0206_04" + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.436, + 0.431, + 0.45 + ], + "angle": 0, + "content": "Figure 6. Out-of-distribution Pose Animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.539 + ], + "angle": 0, + "content": "drawn from an unseen sequence exhibiting different human motions. The results suggest that our method is capable of generalizing to these significant pose changes, even though such poses were not observed during training." + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.551, + 0.472, + 0.566 + ], + "angle": 0, + "content": "Table 3. Metrics of Novel Pose Renderings on I3D-Human." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.577, + 0.421, + 0.641 + ], + "angle": 0, + "content": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
Dyco [8]30.090.956935.32
3DGS-Avatar [56]29.540.955538.59
GauHuman [22]29.310.952648.29
Ours30.280.958336.07
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.244, + 0.683 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In this section, we ablate the influence of various components in our methods on I3D-Human dataset. Tab. 4 shows the average metrics of 4 sequences under different settings. Skeleton motion condition \\(\\Delta \\mathcal{P}\\) denotes the overall human body movement. Compared to predicting non-rigid deformation only with the Gaussian position, incorporating this additional global temporal information allows our method to better model the overall human shape and achieve higher rendering quality, as demonstrated in Tab. 4 (c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Fine-grained point-wise motion condition \\(\\mathcal{V}\\) describes the local region's variation in a more detailed manner. Such point-based motion information captures the relationship between motion and appearance variations in areas beyond the human skeleton. Tab. 4 (d) and Fig. 5 (d) show that our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.905, + 0.4 + ], + "angle": 0, + "content": "method achieves higher rendering details in local regions. Spatio-temporal Multi-scale Sampling (STMS) is designed to enhance the robustness of human motion conditions by incorporating local region motion information and motion patterns across different temporal windows. With such a hierarchical motion context embedding design, our method is able to predict each Gaussian's deformation more robustly, leading to better rendering quality as shown in Tab. 4 (e) and Fig. 5 (e). Please refer to the supplementary materials for more detailed ablation experiments." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.41, + 0.906, + 0.466 + ], + "angle": 0, + "content": "Table 4. Quantitative Ablation Results on I3D-Human. (a) denotes experiments without non-rigid deformation. (b) only utilizes Gaussian positions \\(\\mathbf{x}\\) and the current frame pose \\(P\\) as the conditions for non-rigid MLP." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.478, + 0.903, + 0.55 + ], + "angle": 0, + "content": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
(a) Baseline29.760.956938.35
(b) (a)+Vanilla Non-Rigid MLP+P Conditions.31.050.961734.35
(c) (b)+ΔP Conditions.31.890.964532.17
(d) (c)+V Conditions.32.010.965131.23
(e) (d)+STMS. (Ours)32.240.966429.78
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.771, + 0.589 + ], + "angle": 0, + "content": "6. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Limitations. The Gaussian-based representation used in our method may introduce slight blur artifacts during rendering, whereas NeRF's ray-based volumetric integration tends to produce sharper results. Moreover, our local velocity cues are derived from the coarse SMPL model rather than dense surface tracking [79], which may limit the accuracy of fine-scale garment deformation. Addressing these limitations remains an open challenge for future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose a 3DGS-based framework that integrates hierarchical motion context for 3D human modeling. Specifically, the non-rigid deformation for the Gaussian primitive is learned based on the global human skeleton variations and fine-grained Gaussian's point-wise motions. To capture each Gaussian primitive's motion information more robustly, we introduce a spatio-temporal multi-scale sampling strategy to incorporate both local region motion features and motion patterns across different temporal intervals. Through the above design, the proposed method achieves state-of-the-art rendering of human avatars with complex garments and motions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13599" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.482, + 0.148 + ], + "angle": 0, + "content": "This research was supported by Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.187, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.273, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[3] Ang Cao and Justin Johnson. Hexplane: A Fast Representation for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[5] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Lin chao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[6] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[8] Yutong Chen, Yifan Zhan, Zhihang Zhong, Wei Wang, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Within the dynamic context: Inertia-aware 3d human modeling with pose sequence. arXiv preprint arXiv:2403.19160, 2024. 2, 3, 5, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 5, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[10] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "estimation and tracking from multiple views. In T-PAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic View Synthesis from Dynamic Monocular Video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[17] Qingzhe Gao, Yiming Wang, Libin Liu, Lingjie Liu, Christian Theobalt, and Baoquan Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[18] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[19] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[21] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Tao Hu, Fangzhou Hong, and Ziwei Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13600" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[24] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[25] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[26] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[28] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In ACM SIGGRAPH 2007 papers, pages 64-es. 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.439 + ], + "angle": 0, + "content": "[29] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[30] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[31] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[32] Mingwei Li, Jiachen Tao, Zongxin Yang, and Yi Yang. Human101: Training 100+ fps human gaussians in 100s from 1 view. arXiv preprint arXiv:2312.15258, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[33] Mengtian Li, Shengxiang Yao, Zhifeng Xie, Keyu Chen, and Yu-Gang Jiang. Gaussianbody: Clothed human reconstruction via 3d gaussian splatting. arXiv preprint arXiv:2401.09720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[34] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3D Video Synthesis from Multi-view Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[35] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, pages 5741-5751, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. pages 15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[40] Xinqi Liu, Chenming Wu, Xing Liu, Jialun Liu, Jinbo Wu, Chen Zhao, Haocheng Feng, Errui Ding, and Jingdong Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[41] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[42] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Spl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[45] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[46] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[47] Muyao Niu, Yifan Zhan, Qingtian Zhu, Zhuoxiao Li, Wei Wang, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Bundle adjusted gaussian avatars deblurring. arXiv preprint arXiv:2411.16758, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[48] Muyao Niu, Mingdeng Cao, Yifan Zhan, Qingtian Zhu, Mingze Ma, Jiancheng Zhao, Yanhong Zeng, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13601" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[50] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.245 + ], + "angle": 0, + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[52] Rui Peng, Wangze Xu, Luyang Tang, Jianbo Jiao, Ronggang Wang, et al. Structure consistent gaussian splatting with matching prior for few-shot novel view synthesis. Advances in Neural Information Processing Systems, 37:97328-97352, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[53] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[54] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7964-7976, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[55] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-neRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[56] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[57] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[58] Shihe Shen, Huachen Gao, Wangze Xu, Rui Peng, Luyang Tang, Kaiqiang Xiong, Jianbo Jiao, and Ronggang Wang. Disentangled generation and aggregation for robust radiance fields. In European Conference on Computer Vision, pages 218-236. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.79 + ], + "angle": 0, + "content": "[59] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.832 + ], + "angle": 0, + "content": "[60] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[61] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21096-21106, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[63] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[64] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, pages 1-19. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[65] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[67] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[68] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[69] Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. Swift4d: Adaptive divide-and-conquer gaussian splatting for compact and efficient reconstruction of dynamic scene. arXiv preprint arXiv:2503.12307, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[70] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time Neural Irradiance Fields for Free-viewpoint Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.762 + ], + "angle": 0, + "content": "[71] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[72] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[73] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13602" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[74] Jinbo Yan, Rui Peng, Zhiyan Wang, Luyang Tang, Jiayu Yang, Jie Liang, Jiahao Wu, and Ronggang Wang. Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16520-16531, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[75] Yifan Zhan, Qingtian Zhu, Muyao Niu, Mingze Ma, Jiancheng Zhao, Zhihang Zhong, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Tomie: Towards modular growth in enhanced smpl skeleton for 3d human with animatable garments. arXiv preprint arXiv:2410.08082, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[76] Yifan Zhan, Wangze Xu, Qingtian Zhu, Muyao Niu, Mingze Ma, Yifei Liu, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. R3-athlete: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[78] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[79] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[80] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, pages 145-163. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[81] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13603" + } + ] +] \ No newline at end of file diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_origin.pdf b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7506863347f9cffa2129ee5667bc8afe7ad1fd39 --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/ce271790-0752-41a2-ac20-e12f48492a27_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d76edf4d81408d600127808abcf3e68b869529b626dacdbdff86e9d9280f7c0 +size 4791399 diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/full.md b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5063037e5c3f6a18584e12d120c74e311d9aaae --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/full.md @@ -0,0 +1,423 @@ +# Sequential Gaussian Avatars with Hierarchical Motion Context + +Wangze Xu $^{1*}$ Yifan Zhan $^{1,2}$ Zhihang Zhong $^{1\dagger}$ Xiao Sun $^{1\dagger}$ $^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ The University of Tokyo + +![](images/6a382d1fa5228c3c14aeda352e1bade0b89ee9dfcbe2eb901aa8326f6391aeb9.jpg) +Figure 1. Illustration of the Hierarchical Motion Context Design. We model the motion-dependent appearance variations with both coarse skeleton condition $\Delta P$ and fine-grained point-wise velocity condition $\mathcal{V}$ . $\Delta P$ describes the overall human skeleton motion, which is derived from the difference between the human poses at adjacent frames. $\mathcal{V}$ is the point-wise velocity that indicates finer-grained motion in local regions. To achieve robust deformation, we propose Spatio-temporal Multi-scale Sampling, which samples the overall motion trend and inter-frame details via diverse time intervals $s$ for $\Delta P$ and $\mathcal{V}$ . + +# Abstract + +The emergence of neural rendering has significantly advanced the rendering quality of 3D human avatars, with the recently popular 3DGS technique enabling real-time performance. However, SMPL-driven 3DGS human avatars still struggle to capture fine appearance details due to the complex mapping from pose to appearance during fitting. In this paper, we propose SeqAvatar, which excavates the explicit 3DGS representation to better model human avatars based on a hierarchical motion context. Specifically, we utilize a coarse-to-fine motion conditions that incorporate both the overall human skeleton and fine-grained vertex motions for non-rigid deformation. To enhance the robustness of the proposed motion conditions, we adopt a spatiotemporal multi-scale sampling strategy to hierarchically integrate more motion clues to model human avatars. Extensive experiments demonstrate that our method significantly outperforms 3DGS-based approaches and renders human + +avatars orders of magnitude faster than the latest NeRF-based models that incorporate temporal context, all while delivering performance that is at least comparable or even superior. Project page: https://zezeaaa.github.io/projects/SeqAvatar/ + +# 1. Introduction + +Recent research on digital humans has highlighted the efficiency of 3D Gaussian Splatting (3DGS) [27], demonstrating its capability for high-quality and real-time rendering. By defining T-posed human Gaussians in a canonical space and employing specific warping techniques, we can render human avatars from diverse perspectives and in any pose. The canonical Gaussian primitives and warping weights are then jointly optimized under supervision from video inputs. + +Animatable human avatar reconstruction is primarily hindered by limited modeling of non-rigid warping, forcing a trade-off between rendering quality and animation capability. Current methods [22, 56] incorporate SMPL(-X) pose priors [43, 51] to guide motion for each observation + +and rely on Linear Blend Skinning (LBS) for warping. Although these approaches are effective for pose-driven animation, they often struggle with per-frame non-rigid warping in scenarios involving complex garments. + +We experimentally find that current human pose conditions [22, 56] do not fully capture the complex one-to-many mapping from pose to appearance, particularly for 3DGS-based approaches [27]. These methods [56, 68] rely on the spatial information of a template body in each frame's human pose to predict non-rigid deformations. However, they often overlook local details, such as garment deformations far from the skeleton, and cannot resolve cases where the same pose corresponds to different appearances during complex motions. Furthermore, although previous NeRF-based [44] method [8] has attempted to model motion sequences using human pose residuals, the inherently global nature of the pose sequence limits the ability to capture finer motion details. Naive pose sequence modeling does not fully account for the explicit characteristics of 3DGS. + +In this paper, we introduce a hierarchical motion context condition for 3DGS-based human avatar modeling to address the limitations of relying solely on human pose, which provides only limited global skeletal information. Our approach improves the ability of 3DGS-based methods to accurately capture the complex relationship between human pose and appearance in challenging scenarios. Specifically, we design a coarse-to-fine motion condition that incorporates both overall skeletal movements and fine-grained point-wise motions. Leveraging the explicit nature of Gaussian primitives, this condition seamlessly integrates into 3DGS-based methods, enabling more precise predictions of complex non-rigid deformations. To further enhance robustness, we propose a spatio-temporal multi-scale sampling strategy for constructing the motion context with a larger receptive field. Spatially, we consider the motion states of neighboring points within the local region of each Gaussian primitive to obtain more stable motion embeddings. Temporally, we capture human motion patterns across multiple time scales, combining long-term trends with fine-grained inter-frame motion details. This improves the model's generalization to complex human movements. + +To summarize, our key contributions are as follows: + +1) We propose a novel hierarchical motion condition that integrates coarse-to-fine human motion, combining global skeletal poses with localized vertex residuals to enhance non-rigid deformation prediction. +2) We introduce a spatiotemporal multiscale sampling strategy that expands the receptive field of hierarchical motion context, improving generalization to complex motions. +3) Experiments on the I3D-Human [8], DNA-Rendering [9], and ZJU-MoCap [53] datasets show the effectiveness of the proposed SeqAvatar, which is capable of modeling details of the human body in complex motions. + +# 2. Related Work + +Neural Rendering. Neural rendering techniques have brought significant progress to human reconstruction and rendering. In particular, Neural Radiance Fields (NeRF) [44] introduces an implicit scene representation that models color and density using multilayer perceptrons, delivering photorealistic rendering results. More recently, point-based 3D Gaussian Splatting [27] utilizes 3D Gaussians to represent scenes explicitly, achieving real-time high-quality rendering. Building upon NeRF and 3DGS, subsequent works have advanced neural rendering across various dimensions, e.g., enhancing visual fidelity [1, 2], improving sparse-view reconstruction quality [10, 52, 63, 71, 80], enabling pose-free optimization [37, 58, 61, 67], modeling dynamic scenes [13, 34, 49, 55, 69, 74], and accelerating training and inference [4, 12, 18, 39, 46, 60]. These developments also benefit human avatar modeling, which demands high-quality and efficient rendering. + +SMPL(-X)-Based Neural Human Modeling. The SMPL(-X) family [43, 51] provides a parametric representation of the human body by decomposing it into pose-related and shape-related components using 3D mesh scans and principal component analysis (PCA). Its pose blend shapes enables body deformation through joint-wise pose blending, offering an efficient and compact way for animation. SMPL(-X) has thus become a cornerstone in human body modeling and animation, with many methods [11, 59] estimating its parameters directly from 2D inputs. Recent neural human reconstruction methods integrate SMPL(-X) with implicit representations such as NeRF [44] and 3DGS [27] to enable animatable avatars with high rendering quality. Different from purely 2D image animation methods [48, 62, 65, 73], these methods typically register the input data to a canonical T-pose space and use linear blend skinning (LBS) to transform 3D points into observation space based on SMPL(-X) poses. NeRF-based approaches [5-8, 14, 17, 19, 20, 30, 53, 64, 68, 75] focus on monocular or multi-view reconstruction. More recently, 3DGS-based methods [21-23, 25, 26, 29, 31-33, 36, 40, 41, 45, 47, 56, 76, 78, 79, 81] have gained popularity due to their real-time rendering and high fidelity. Some works [8, 23] further incorporate pose sequences as temporal context, but they suffer from slow rendering and lack fine details of motion caused by simply integrating coarse human pose embedding with NeRF. + +Human Rendering with Temporal Embeddings The earliest modeling of dynamic radiance field [3, 13, 15, 16, 24, 34, 35, 42, 50, 57, 70] can be traced back to temporal embeddings in general scenes, where each frame's observation is obtained by constructing a canonical space and a time-conditioned deformation field. Motivated by these, a stream of research [38, 72] focuses on pure rendering quality, employing temporal embeddings instead of human pose to en + +code each frame of human videos. Although these methods achieve high-quality, temporally continuous rendering, they struggle with the lack of geometric constraints from human pose, making parameterized animation unfeasible. + +# 3. Preliminary + +SMPL(-X) Series [43, 51] adopt a parameterized framework to represent human bodies across diverse shapes and poses. For each frame, the human mesh is derived by deforming a canonical template mesh based on shape and pose parameters. Specifically, a 3D point $\mathbf{x}$ on the canonical mesh is warped to the corresponding point on the deformed mesh in the observation space, following + +$$ +\mathbf {L B S} (\mathbf {x}, \mathbf {B} _ {k}, \omega_ {k} (\mathbf {x})) = \sum_ {k = 1} ^ {K} \omega_ {k} (\mathbf {x}) \mathbf {B} _ {k} \mathbf {x}, \tag {1} +$$ + +where $K$ is the total bone number and $\mathbf{B}_k$ is the transformation matrix for each bone. Specifically, $\mathbf{B}_k$ consists of a rotation and translation matrix $\mathbf{R}_k$ and $\mathbf{b}_k$ . Here, $\mathbf{R}_k$ represents the global rotation of each joint, influenced by the local joint rotations $\boldsymbol{r}$ , while $\mathbf{b}_k$ denotes each joint's translation, determined by the joint positions $\boldsymbol{j}$ and the human shape parameters $\beta$ . The linear blending weight $\omega^k$ depends on $\mathbf{x}$ and is regressed from comprehensive human meshes. + +3D Gaussian Splatting [27] utilizes a set of point-based Gaussian primitives to explicitly represent scenes, enabling real-time and high-quality rendering. Each Gaussian primitive is defined by its center position $\mathbf{x} \in \mathbb{R}^3$ and covariance matrix $\boldsymbol{\Sigma} \in \mathbb{R}^{3 \times 3}$ . To ensure positive semi-definiteness and simplify optimization, the covariance matrix $\boldsymbol{\Sigma}$ is decomposed into a rotation matrix $\mathbf{R}$ and a scaling matrix $\mathbf{S}$ : + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {T} \mathbf {R} ^ {T}, \tag {2} +$$ + +where $\mathbf{R}$ and $\mathbf{S}$ are derived by a scaling vector $\mathbf{s} \in \mathbb{R}^3$ and a quaternion vector $\mathbf{r} \in \mathbb{R}^4$ in practice. Additionally, each Gaussian primitive is also assigned a color feature $sh \in \mathbb{R}^k$ represented by spherical harmonics (SH) and an opacity $\alpha$ for rendering. During the rendering process, a splatting technique [27] is used to project each Gaussian primitive onto the 2D image space with a viewing transform $\mathbf{W}$ and the Jacobian $\mathbf{J}$ of the projective transformation's affine approximation. The transformed covariance $\boldsymbol{\Sigma}'$ in the camera's coordinates is + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {T} \mathbf {J} ^ {T}. \tag {3} +$$ + +After projection, the pixel color $C$ is computed by blending $N$ ordered Gaussian primitives overlapping at the pixel: + +$$ +C = \sum_ {i \in N} c _ {i} \alpha_ {i} ^ {\prime} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} ^ {\prime}\right), \tag {4} +$$ + +where $c_{i}$ is computed from SH feature $sh$ and $\alpha_{i}^{\prime}$ is the product of $\alpha_{i}$ and probability density of $i$ -th 2D Gaussian. + +For optimization, a photometric loss is defined by a combination of $L_{1}$ and SSIM [66] losses: + +$$ +L _ {\text {p h o t o}} = \lambda L _ {1} (\hat {I}, I) + (1 - \lambda) (1 - S S I M (\hat {I}, I)), \tag {5} +$$ + +where $\hat{I}$ and $I$ denote the rendered and ground-truth images, and $\lambda$ controls the balance between the two terms. + +# 4. Method + +Fig. 2 shows the framework of our method. We use the explicit point-based 3DGS as the representation of the human body. Given a collection of input cameras and images, we optimize a set of Gaussian primitives $\{\mathcal{G}_i\}_{i=1}^{n}$ to fit the body's shape and appearance. Each Gaussian primitive $\mathcal{G}_i$ includes the center position $\mathbf{x}$ , scaling vector $\mathbf{s}$ , rotation quaternion vector $\mathbf{r}$ , color feature $sh$ and opacity $\alpha$ , where $\mathbf{x}$ is initialized from the SMPL template vertices $\{\mathbf{T}_i\}_{i=1}^{i=N}$ . During the optimization process, we first apply a coarse-to-fine motion context to capture more accurate and fine-grained details (Sec. 4.1). To mitigate overfitting, we propose Spatio-Temporal Multi-Scale Sampling to obtain more robust point-wise Gaussian motion, which serves as the embedding for non-rigid deformation (Sec.4.2). Next, we adopt Linear Blend Skinning (LBS) to map canonical Gaussian primitives $\mathcal{G}$ to observation space and render images via differentiable splitting (Sec.4.3). + +# 4.1. Non-Rigid Deformation with Coarse-to-fine Motion Context + +Coarse Skeleton Motion. As illustrated in Fig. 1, we introduce a coarse-to-fine motion context to excavate additional conditions for non-rigid deformation. Typically, most human 3D avatar methods use the current frame pose as a condition to predict non-rigid deformation [22, 56]. While this provides spatial information to distinguish the motion of different body parts, it fails to capture temporal motion changes, as discussed in [8]. Therefore, given a frame at time $t$ , we consider a sequence of regularly interval-sampled frames $\mathcal{T}$ to model inter-frame body motion variations: + +$$ +\mathcal {T} = \{t - s, t - 2 s, \dots , t - L s \}, \tag {6} +$$ + +where $L$ is the sequence length and $s$ is the time interval. The coarse motion of body skeletons at each time step $t \in \mathcal{T}$ can be derived by calculating the difference of poses between adjacent frames as illustrated in Fig. 1, following + +$$ +\Delta \mathcal {P} = \left\{\Delta P ^ {t} = \delta \left(P ^ {t}, P ^ {t - s}\right) | t \in \mathcal {T} \right\}, \tag {7} +$$ + +where $P \in \mathbb{R}^{K \times 3}$ is the body pose and $\delta$ denotes the difference $\Delta \mathcal{P}$ between two poses in axis-angle form. We employ an MLP $\mathcal{E}_{\Delta \mathcal{P}}$ to encode the sequential skeleton mo + +![](images/bec0715e77cfb6f8898d1bbbc367c58618f4d3b2c0dee2f2110e75d4162cb622.jpg) +Figure 2. Overview of the proposed method. We first initialize canonical Gaussian positions $\mathbf{x}$ with SMPL template vertexes. For each Gaussian, we derive both coarse skeleton motion condition $f_{\Delta \mathcal{P}}$ and fine-grained vertex motion condition $f_{\nu}$ that sampled from the vertex motion template $\mathcal{F}_{\nu}$ (points in different colors represent different motions). Based on such hierarchical motion information, we utilize an MLP $\mathcal{E}_{non-rigid}$ to better predict each Gaussian's non-rigid deformation prediction. The non-rigid deformed Gaussians are then warped into observation space via the standard LBS transformation for rendering. + +tion $\Delta \mathcal{P}\in \mathbb{R}^{L\times K\times 3}$ , flattened along the temporal dimension, into a skeleton motion embedding $f_{\Delta \mathcal{P}}\in \mathbb{R}^{32}$ + +$$ +f _ {\Delta \mathcal {P}} = \mathcal {E} _ {\Delta \mathcal {P}} (\Delta \mathcal {P}), \tag {8} +$$ + +which serves as a condition for the subsequent non-rigid deformation. Further details are provided in the supplementary material. + +Fine Vertex Motion. Compared to previous implicit NeRF-based methods, the point-based 3DGS representation enables us to explore more fine-grained temporal motion information. We derive a point-wise velocity vector $v_{i}$ for each Gaussian primitive $\mathcal{G}_i$ to model the fine-grained localized body motion, which is beyond the scope of pose skeleton motion $\Delta \mathcal{P}$ . By measuring the position variations across adjacent time steps, we calculate per-frame velocity + +$$ +v ^ {t} = \frac {\mathbf {x} _ {\mathbf {o}} ^ {t} - \mathbf {x} _ {\mathbf {o}} ^ {t - s}}{s}, \tag {9} +$$ + +where $\mathbf{x_o}$ denotes the warped coordinates in observation space. However, since the positions of Gaussian primitives are updated during optimization, it is not stable to obtain $v_{i}$ based on such a dynamic variable. Moreover, transforming the canonical Gaussians into the observation space requires non-rigid deformation first, which leads to a circular dependency and conflicts with the current velocity design. To this end, we introduce a motion template field $\mathcal{F}_{\mathrm{V}} = \{\mathbf{V}_i\}_{i = 1}^{i = N}$ that stores the velocity of each SMPL vertex, and $v_{i}$ for each Gaussian primitive is sampled from $\mathcal{F}_{\mathrm{V}}$ based on the distance between query points and vertexes in this template. Specifically, given a time step $t$ and the corresponding body pose $P^t = \{(\mathbf{R}_k^t,\mathbf{b}_k^t)|k\in K\}$ , the SMPL template vertexes in $\mathbf{T}$ can be warped into observation space with the template skinning weights $\mathbf{W}$ and + +standard linear blend skinning, following + +$$ +\mathbf {T} _ {\mathbf {o}} ^ {t} = \mathbf {L B S} (\mathbf {T}, \mathbf {B} ^ {t}, \mathbf {W}), \tag {10} +$$ + +where $\mathbf{B}^t$ is the transformation matrix derived from rotation and translation matrix $\mathbf{R}^t$ and $\mathbf{b}^t$ . $\mathbf{W}$ is the template SMPL LBS weights. The velocity $\mathbf{V}$ of each template vertexes $\mathbf{T}$ can be further derived as + +$$ +\mathbf {V} ^ {t} = \frac {\mathbf {T} _ {\mathbf {o}} ^ {t} - \mathbf {T} _ {\mathbf {o}} ^ {t - s}}{s}. \tag {11} +$$ + +We then sample each Gaussian's velocity from this motion template field $\mathcal{F}_{\mathbf{V}}$ , which is discussed in Sec. 4.2. + +# 4.2. Spatio-temporal Multi-scale Sampling (STMS) + +To enhance the robustness of human motion conditions, we propose modeling both local region motion information of the human body and motion patterns across different temporal windows to mitigate overfitting and improve generalization by capturing more comprehensive motion dynamics. + +In the spatial dimension, we sample the $\tau$ nearest template to model the body's local region motion more robustly. Specifically, for each canonical Gaussian primitive $\mathcal{G}_i$ , the $\tau$ nearest vertexes' velocities are sampled as input to an MLP $\mathcal{E}_{knn}$ to learn a local point-wise motion embedding + +$$ +e _ {i} ^ {t} = \mathcal {E} _ {k n n} \left(\left\{\mathbf {V} _ {j} ^ {t} \right\}\right), \quad j \in \operatorname {K N N} (\mathbf {T}, \mathbf {x} _ {i}), \tag {12} +$$ + +where $\mathbf{KNN}(\mathbf{T},\mathbf{x}_i)$ denotes the $\tau$ nearest SMPL template vertexes of the canonical Gaussian position $\mathbf{x}_i$ . Similar to Eq. 7, we then apply an MLP $\mathcal{E}_{\mathcal{V}}$ to encode each Gaussian primitive's sequential motion embedding $\mathcal{V} = \{e^{t}|t\in \mathcal{T}\}$ into a point-wise sequential condition $f_{\mathcal{V}}\in \mathbb{R}^{96}$ + +$$ +f _ {\mathcal {V}} = \mathcal {E} _ {\mathcal {V}} (\mathcal {V}). \tag {13} +$$ + +In the temporal dimension, to capture both the overall motion trend and the inter-frame motion details, we adopt a multi-scale sequence sampling strategy. In detail, we sample a series of sequences at several progressively increasing intervals to get human motion information across different temporal windows + +$$ +\mathcal {S} = \left\{s = s _ {0} + i \Delta s \right\} _ {i = 0} ^ {i = m}, \tag {14} +$$ + +where $\Delta s$ is the increasing rate of sampling interval $s$ . We then input the multi-scale sampled sequence motions into the skeleton motion encoder $\mathcal{E}_{\Delta P}$ and the point-wise motion encoder $\mathcal{E}_{V}$ to obtain hierarchical temporal motion embeddings. Therefore, Eq. 7 and Eq. 13 can be revised as + +$$ +f _ {\Delta \mathcal {P}} = \mathcal {E} _ {\Delta \mathcal {P}} \left(\left\{\Delta \mathcal {P} _ {s} \right\}\right), \quad s \in \mathcal {S} \tag {15} +$$ + +$$ +f _ {\mathcal {V}} = \mathcal {E} _ {\mathcal {V}} (\{\mathcal {V} _ {s} \}), \quad s \in \mathcal {S}. \tag {16} +$$ + +In practice, we concatenate all skeleton motion conditions, $\Delta \mathcal{P}_s$ for $s\in S$ , and localized vertex motion conditions, $\nu_{s}$ for $s\in S$ , across different sampling scales. These are then input to $\varepsilon_{\Delta P}$ and $\varepsilon_{\nu}$ , respectively. More details are shown in the supplementary material. + +Non-Rigid Deformation. Given the coarse skeleton motion $f_{\Delta \mathcal{P}} = \mathcal{E}_{\Delta \mathcal{P}}(\{\Delta \mathcal{P}_s\})$ condition and fine point-wise velocity condition $f_{\mathcal{V}} = \mathcal{E}_{\mathcal{V}}(\{\mathcal{V}_s\})$ , we utilize an MLP to predict each Gaussian's non-rigid deformation as + +$$ +\delta \mathbf {x}, \delta \mathbf {s}, \delta \mathbf {r} = \mathcal {E} _ {\text {n o n - r i g i d}} (\mathbf {x}, P, f _ {\Delta \mathcal {P}}, f _ {v}). \tag {17} +$$ + +The deformed canonical Gaussian $\mathcal{G}'$ is + +$$ +\mathbf {x} ^ {\prime} = \mathbf {x} + \delta \mathbf {x}, \tag {18} +$$ + +$$ +\mathbf {s} ^ {\prime} = \mathbf {s} + \delta \mathbf {s}, \tag {19} +$$ + +$$ +\mathbf {r} ^ {\prime} = \mathbf {r} \cdot \delta \mathbf {r}, \tag {20} +$$ + +where $\cdot$ denotes the multiplication of two quaternions. + +# 4.3. Optimization + +Rigid Deformation. We utilize the standard LBS operation to map the non-rigid deformed Gaussians $\mathcal{G}_o$ into the observation space, following + +$$ +\mathbf {x} _ {\mathbf {o}} = \mathbf {L B S} \left(\mathbf {x} ^ {\prime}, \mathbf {B}, \omega\right), \tag {21} +$$ + +$$ +\mathbf {R} _ {\mathbf {o}} = \sum_ {k = 1} ^ {K} \omega_ {k} \left(\mathbf {x} ^ {\prime}\right) \mathbf {B} _ {k} \mathbf {R} ^ {\prime}, \tag {22} +$$ + +where $\mathbf{R}'$ is the rotation matrix derived from the non-rigid deformed Gaussian's rotation quaternion $\mathbf{r}'$ , and $\mathbf{R_o}$ represents the rotation matrix in observation space. Following the previous method [22], we utilize an MLP $\mathcal{E}_{lbs}$ to predict the LBS weight offsets for each query canonical point and update the sampled weights from the nearest SMPL vertex as + +$$ +\omega_ {k} (\mathbf {x}) = \omega_ {k} ^ {S M P L} (\mathbf {x}) + \mathcal {E} _ {l b s} (\mathbf {x}). \tag {23} +$$ + +Similar to [8, 22, 68], we introduce a pose refinement MLP $\mathcal{E}_{\text {pose }}$ to refine the pose estimate from SMPL for a better fit to the human body. + +Loss Function. With the transformed Gaussian primitives $\mathcal{G}_o$ in observation space, we apply the standard splitting [27] to render images as + +$$ +I = \mathbf {S p l a t t i n g} \left(\mathbf {x} _ {\mathbf {o}}, \mathbf {R} _ {\mathbf {o}}, s ^ {\prime}, \alpha , s h\right). \tag {24} +$$ + +During the optimization process, we employ a combination of loss functions as supervision, summarized as + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {c o l o r}} + \lambda_ {2} \mathcal {L} _ {\text {s s i m}} + \lambda_ {3} \mathcal {L} _ {\text {l p i p s}} + \mathcal {L} _ {\text {m a s k}}, \tag {25} +$$ + +where $\lambda$ is used to balance the weight of different losses. $\mathcal{L}_{mask}$ [22] is an $L_{2}$ loss between the rendered $\alpha$ and human body mask. Similar to [56], we apply $\mathcal{L}_{isopos}$ and $\mathcal{L}_{isocov}$ to control the Gaussian primitive's position and covariance. Details are shown in the supplementary material. + +# 5. Experiments + +# 5.1. Datasets + +DNA-Rendering [9] is a challenging human-centric rendering dataset. It contains diverse scenes from everyday life to professional occasions. We use 6 sequences with loose-fitting garments (1_0206_04, 2_0007_04, 2_0019_10, 2_0044_11, 2_0051_09, 2_0813_05) for experiments. We adopt 24 views for training and 6 views for testing. + +13D-Human [8] contains scenes closer to daily life, which comprises multi-view video frames of humans with loose clothing and complex movements. We conduct experiments on 4 sequences and use 4-5 views for training and the rest for testing. All training and testing data are retained to be the same in comparisons following Dyco's [8] data split. + +# 5.2. Baselines and Metrics + +We compare against the state-of-the-art human modeling methods including Dyco [8], 3DGS-Avatar [56], GART [31], and GauHuman [22]. Dyco is an implicit NeRF-based method that adopts pose sequence information for temporal modeling. We compare performance with Dyco to showcase the better efficacy of our fine-grained motion condition design. 3DGS-Avatar, GART, and GauHuman are explicit 3DGS-based methods, and all of them are originally designed for monocular inputs. For fairness, we extend them to multi-view input under the same settings. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [66], and learned perceptual image patch similarity (LPIPS) [77]. + +# 5.3. Comparison + +Comparison on DNA-Rendering. Tab. 1 shows the quantitative results on DNA-Rendering dataset. The proposed + +Table 1. Quantitative Results on DNA-Rendering Dataset. We report the performance of novel view rendering. Our method outperforms previous state-of-the-art human modeling methods. We mark the best and the second best methods in cells. LPIPS*=LPIPS × 10^3. + +
Subject:1_0206_042_0007_042_0019_102_0044_112_0051_092_0813_05
Metric:PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]26.640.943950.0928.060.949256.7032.060.971031.1228.230.951237.1525.040.953442.2231.750.970131.30
GART [31]27.190.945456.4728.220.952556.4631.130.968635.1628.900.956141.2726.470.962842.2132.060.972935.74
GauHuman [22]27.960.950050.4227.930.949656.9331.630.969832.8230.380.964134.1925.990.959942.0033.430.976429.42
Ours30.810.964939.4729.630.956643.7234.970.977723.5332.620.974723.8328.450.968534.0635.810.984320.83
+ +![](images/9054b106fd64f5802227d0a6ead54f294f28323c1bd85cce0257cb682b76d4c7.jpg) +Figure 3. Novel View Qualitative Results on DNA-Rendering. We zoom into the local region and compute the error maps compared with ground truth images. The results show that our method achieves competitive results on both the overall and local region qualities. + +method outperforms previous state-of-the-art human modeling methods across all metrics. To demonstrate the visual improvement, we compare the quality of rendering images in Fig. 3. Previous methods, which lack the temporal motion information for non-rigid deformations, are unable to predict the movement of Gaussian primitives properly, resulting in blurred renderings in regions with complex tex + +tures (scene 1_0206_04 in Fig. 3). Besides, the results in Fig. 3 show that the compared methods fail to capture the proper shape variation caused by movement (e.g., the white dress in scene 2_0051_09). Although these methods achieve a smooth appearance in these areas, the error maps show that they fail to match the actual shape closely. We leverage the hierarchical motion context to predict Gaussian defor + +Table 2. Novel View Rendering Quantitative Results on I3D-Human. + +
Subject: Metric:ID1_1ID1_2ID2_1ID3_1
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]29.780.960031.4031.290.962728.0929.630.960437.3332.730.960239.47
Dyco [8]30.810.961728.2931.450.962825.7529.410.961833.2632.500.961235.46
GauHuman [22]29.470.957139.4630.420.958538.8728.510.954649.4632.100.954653.67
Ours31.810.965927.0331.980.966027.8431.530.968530.2333.620.965134.02
+ +![](images/be622cd66f8e05671fbb193d1ca59bb47ff345a8f7d0fbda373a3ec6015a235f.jpg) +Figure 4. Novel Pose Qualitative Results on I3D-Human. We compare the proposed method with previous SOTA approaches. The rendered images and error maps demonstrate our robustness on novel pose rendering. + +mations, allowing our method to capture detailed appearance variations caused by human motions more accurately. + +Comparison on I3D-Human. The quantitative results in Tab. 2 show that our method surpasses the previous SOTA method on most metrics, demonstrating our competitive performance on the I3D-Human dataset. Previous NeRF-based Dyco [8] utilizes pose variation as a condition to model human motions. However, it is limited to capturing only overall body motions and lacks the capacity for finer-grained modeling in local regions. For example, in ID1_1 and ID2_1 shown in Fig. 8 (in the Supp. Mat.), our method renders results that accurately reflect the true motion, while Dyco fails to model the information in regions far from the human body skeleton. Tab. 3 reports the quantitative results of novel pose rendering on the I3D-Human dataset. Our method achieves best performance among 3DGS-based human modeling methods and also enables real-time render- + +ing ( $\sim$ 45 FPS on I3D-Human) compared to NeRF-based Dyco ( $\sim$ 0.7 FPS). Fig. 4 shows the qualitative comparisons of novel pose rendering. Different from Dyco [8], which relies solely on joint motions as conditions, our method leverages the motion state of each Gaussian primitive to predict deformations, enabling more flexible modeling of motions in regions distant from the human body. + +To evaluate the generalizability, we also conduct experiments on ZJU-MoCap [53]. Please refer to the supplementary materials for more details. + +Out-of-Distribution Poses. To evaluate the generalization ability to novel poses, we conduct experiments using animations with large pose variations. Specifically, we use a model trained on one sequence to render target poses sampled from other unseen sequences, as different sequences in DNA-Rendering and I3D-Human contain distinct motion patterns. As illustrated in Fig. 6, the Target Pose is + +![](images/3d7b24e057c3c73a0d3f38b70399850617ea366f5d266a48053f7b27dbbbb948.jpg) +(a). Baseline +(b). (a) + Vanilla Non-Rigid MLP. +Figure 5. Qualitative Ablation Results. We compare the visual influence of different components on ID2_1 scene of I3D-Human dataset. + +![](images/77cb632348fd966e2ec4cdd19e16e83b4a1031ec1909fadc4a307f618a192f5d.jpg) + +![](images/1e66811fefeb560e3053e1d6e714909949085b0110c6b098902c07685cc5c841.jpg) +(c). $(\mathbf{b}) + \Delta \mathcal{P}$ Conditions + +![](images/9f6aca3ad615cebc7b57dd9d9f561befdb91306a9d44655b7650007386329569.jpg) +(d).(c) $^+$ V Conditions + +![](images/7131523b57af1d890fc681715c8dfb5f5dd9eb7d70bbfc7fbbc72fbeb68e00af.jpg) +(e). (d) + STMS. + +![](images/07bf9af8d75eff5f0e90590ce72132077306b8b223db98bc4613292a359e5204.jpg) +GT + +![](images/013197955407564466371049b0137826d77b0e8ab6c9e346bb82deaa4104c121.jpg) +ID3 + +![](images/c84b07fde1f292219b4f95bd820a564cdcb49ad21bd55453e2cbd3782f059b85.jpg) +Figure 6. Out-of-distribution Pose Animation. + +![](images/5dded8ec12a23f0aa1b9f42bdb1eb89c0300143f6228cefc56cf797a2179a1c1.jpg) + +![](images/df8e7eaa707ddfadf0b196a0e5141ce265c2e9d66ffae5be89fe41e5f06acc5d.jpg) + +![](images/0395ae9202ab5eaaad448ae0c2e726e5d38009f73fbe0a955aec1d1409f1cb08.jpg) + +![](images/fb60a18e0b5b2e55e5c8ba3ec0a12a4c2a0526fbd32960f026f08c9eaa49fb6d.jpg) + +![](images/c79bd81aa1de4777ea49ab439030998b14d44aad444f6d206f1e125ae47e8d9c.jpg) +2_0051_09 + +![](images/41ce47e0e9dd79d7bf75c0f64c38a965b2b6dd6ff37eedf4b12354d590ac3321.jpg) +1_0206_04 + +drawn from an unseen sequence exhibiting different human motions. The results suggest that our method is capable of generalizing to these significant pose changes, even though such poses were not observed during training. + +Table 3. Metrics of Novel Pose Renderings on I3D-Human. + +
MethodsPSNR ↑SSIM ↑LPIPS* ↓
Dyco [8]30.090.956935.32
3DGS-Avatar [56]29.540.955538.59
GauHuman [22]29.310.952648.29
Ours30.280.958336.07
+ +# 5.4. Ablation Study + +In this section, we ablate the influence of various components in our methods on I3D-Human dataset. Tab. 4 shows the average metrics of 4 sequences under different settings. Skeleton motion condition $\Delta \mathcal{P}$ denotes the overall human body movement. Compared to predicting non-rigid deformation only with the Gaussian position, incorporating this additional global temporal information allows our method to better model the overall human shape and achieve higher rendering quality, as demonstrated in Tab. 4 (c). + +Fine-grained point-wise motion condition $\mathcal{V}$ describes the local region's variation in a more detailed manner. Such point-based motion information captures the relationship between motion and appearance variations in areas beyond the human skeleton. Tab. 4 (d) and Fig. 5 (d) show that our + +method achieves higher rendering details in local regions. Spatio-temporal Multi-scale Sampling (STMS) is designed to enhance the robustness of human motion conditions by incorporating local region motion information and motion patterns across different temporal windows. With such a hierarchical motion context embedding design, our method is able to predict each Gaussian's deformation more robustly, leading to better rendering quality as shown in Tab. 4 (e) and Fig. 5 (e). Please refer to the supplementary materials for more detailed ablation experiments. + +Table 4. Quantitative Ablation Results on I3D-Human. (a) denotes experiments without non-rigid deformation. (b) only utilizes Gaussian positions $\mathbf{x}$ and the current frame pose $P$ as the conditions for non-rigid MLP. + +
MethodsPSNR ↑SSIM ↑LPIPS* ↓
(a) Baseline29.760.956938.35
(b) (a)+Vanilla Non-Rigid MLP+P Conditions.31.050.961734.35
(c) (b)+ΔP Conditions.31.890.964532.17
(d) (c)+V Conditions.32.010.965131.23
(e) (d)+STMS. (Ours)32.240.966429.78
+ +# 6. Limitations and Conclusion + +Limitations. The Gaussian-based representation used in our method may introduce slight blur artifacts during rendering, whereas NeRF's ray-based volumetric integration tends to produce sharper results. Moreover, our local velocity cues are derived from the coarse SMPL model rather than dense surface tracking [79], which may limit the accuracy of fine-scale garment deformation. Addressing these limitations remains an open challenge for future work. + +Conclusion. In this paper, we propose a 3DGS-based framework that integrates hierarchical motion context for 3D human modeling. Specifically, the non-rigid deformation for the Gaussian primitive is learned based on the global human skeleton variations and fine-grained Gaussian's point-wise motions. To capture each Gaussian primitive's motion information more robustly, we introduce a spatio-temporal multi-scale sampling strategy to incorporate both local region motion features and motion patterns across different temporal intervals. Through the above design, the proposed method achieves state-of-the-art rendering of human avatars with complex garments and motions. + +# Acknowledgment + +This research was supported by Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 2 +[3] Ang Cao and Justin Johnson. Hexplane: A Fast Representation for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[5] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Lin chao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 2 +[6] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. +[7] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023. +[8] Yutong Chen, Yifan Zhan, Zhihang Zhong, Wei Wang, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Within the dynamic context: Inertia-aware 3d human modeling with pose sequence. arXiv preprint arXiv:2403.19160, 2024. 2, 3, 5, 7, 8, 1 +[9] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 5, 1, 3 +[10] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2 +[11] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose + +estimation and tracking from multiple views. In T-PAMI, 2021. 2 +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2 +[14] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2 +[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic View Synthesis from Dynamic Monocular Video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[17] Qingzhe Gao, Yiming Wang, Libin Liu, Lingjie Liu, Christian Theobalt, and Baoquan Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[18] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, pages 14346-14355, 2021. 2 +[19] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 2 +[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2 +[21] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2 +[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 1, 2, 3, 5, 6, 7, 8 +[23] Tao Hu, Fangzhou Hong, and Ziwei Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 2 + +[24] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375, 2022. 2 +[25] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2 +[26] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 2 +[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5 +[28] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In ACM SIGGRAPH 2007 papers, pages 64-es. 2007. 1 +[29] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 2 +[30] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2 +[31] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 5, 6 +[32] Mingwei Li, Jiachen Tao, Zongxin Yang, and Yi Yang. Human101: Training 100+ fps human gaussians in 100s from 1 view. arXiv preprint arXiv:2312.15258, 2023. +[33] Mengtian Li, Shengxiang Yao, Zhifeng Xie, Keyu Chen, and Yu-Gang Jiang. Gaussianbody: Clothed human reconstruction via 3d gaussian splatting. arXiv preprint arXiv:2401.09720, 2024. 2 +[34] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3D Video Synthesis from Multi-view Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2 +[35] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2 +[36] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2 + +[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, pages 5741-5751, 2021. 2 +[38] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2 +[39] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. pages 15651-15663, 2020. 2 +[40] Xinqi Liu, Chenming Wu, Xing Liu, Jialun Liu, Jinbo Wu, Chen Zhao, Haocheng Feng, Errui Ding, and Jingdong Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 2 +[41] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 2 +[42] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2 +[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Spl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2 +[45] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 2 +[46] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[47] Muyao Niu, Yifan Zhan, Qingtian Zhu, Zhuoxiao Li, Wei Wang, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Bundle adjusted gaussian avatars deblurring. arXiv preprint arXiv:2411.16758, 2024. 2 +[48] Muyao Niu, Mingdeng Cao, Yifan Zhan, Qingtian Zhu, Mingze Ma, Jiancheng Zhao, Yanhong Zeng, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025. 2 +[49] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2 + +[50] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 2 +[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3 +[52] Rui Peng, Wangze Xu, Luyang Tang, Jianbo Jiao, Ronggang Wang, et al. Structure consistent gaussian splatting with matching prior for few-shot novel view synthesis. Advances in Neural Information Processing Systems, 37:97328-97352, 2024. 2 +[53] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 7, 1 +[54] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7964-7976, 2023. 1 +[55] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-neRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2 +[56] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 1, 2, 3, 5, 6, 7, 8 +[57] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2 +[58] Shihe Shen, Huachen Gao, Wangze Xu, Rui Peng, Luyang Tang, Kaiqiang Xiong, Jianbo Jiao, and Ronggang Wang. Disentangled generation and aggregation for robust radiance fields. In European Conference on Computer Vision, pages 218-236. Springer, 2024. 2 +[59] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2 +[60] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2 +[61] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2 + +[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21096-21106, 2025. 2 +[63] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2 +[64] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, pages 1-19. Springer, 2022. 2 +[65] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2 +[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3, 5 +[67] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2 +[68] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2, 5 +[69] Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. Swift4d: Adaptive divide-and-conquer gaussian splatting for compact and efficient reconstruction of dynamic scene. arXiv preprint arXiv:2503.12307, 2025. 2 +[70] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time Neural Irradiance Fields for Free-viewpoint Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2 +[71] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 2 +[72] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In CVPR, 2024. 2 +[73] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2 + +[74] Jinbo Yan, Rui Peng, Zhiyan Wang, Luyang Tang, Jiayu Yang, Jie Liang, Jiahao Wu, and Ronggang Wang. Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16520-16531, 2025. 2 +[75] Yifan Zhan, Qingtian Zhu, Muyao Niu, Mingze Ma, Jiancheng Zhao, Zhihang Zhong, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Tomie: Towards modular growth in enhanced smpl skeleton for 3d human with animatable garments. arXiv preprint arXiv:2410.08082, 2024. 2 +[76] Yifan Zhan, Wangze Xu, Qingtian Zhu, Muyao Niu, Mingze Ma, Yifei Liu, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. R3-athlete: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 2 +[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[78] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2 +[79] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 2, 8 +[80] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, pages 145-163. Springer, 2024. 2 +[81] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2 \ No newline at end of file diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/images.zip b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25860dd7ebc7b46718ec5c4353b9519cd34c7038 --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0654b3265b9b6c663adf6f467143d873f1de9ba8856449143c71c4340403879b +size 705518 diff --git a/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/layout.json b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b0567c1802a86d8711987faab5c100c10eb85a0b --- /dev/null +++ b/2025/Sequential Gaussian Avatars with Hierarchical Motion Context/layout.json @@ -0,0 +1,11517 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "text", + "content": "Sequential Gaussian Avatars with Hierarchical Motion Context" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": "Wangze Xu" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": " Yifan Zhan" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": " Zhihang Zhong" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": " Xiao Sun" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 481, + 174 + ], + "type": "text", + "content": "The University of Tokyo" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 71, + 186, + 541, + 373 + ], + "blocks": [ + { + "bbox": [ + 71, + 186, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 71, + 186, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 71, + 186, + 541, + 373 + ], + "type": "image", + "image_path": "6a382d1fa5228c3c14aeda352e1bade0b89ee9dfcbe2eb901aa8326f6391aeb9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": "Figure 1. Illustration of the Hierarchical Motion Context Design. We model the motion-dependent appearance variations with both coarse skeleton condition " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " and fine-grained point-wise velocity condition " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " describes the overall human skeleton motion, which is derived from the difference between the human poses at adjacent frames. " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " is the point-wise velocity that indicates finer-grained motion in local regions. To achieve robust deformation, we propose Spatio-temporal Multi-scale Sampling, which samples the overall motion trend and inter-frame details via diverse time intervals " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 382, + 555, + 439 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "spans": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 478, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 478, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 297, + 682 + ], + "type": "text", + "content": "The emergence of neural rendering has significantly advanced the rendering quality of 3D human avatars, with the recently popular 3DGS technique enabling real-time performance. However, SMPL-driven 3DGS human avatars still struggle to capture fine appearance details due to the complex mapping from pose to appearance during fitting. In this paper, we propose SeqAvatar, which excavates the explicit 3DGS representation to better model human avatars based on a hierarchical motion context. Specifically, we utilize a coarse-to-fine motion conditions that incorporate both the overall human skeleton and fine-grained vertex motions for non-rigid deformation. To enhance the robustness of the proposed motion conditions, we adopt a spatiotemporal multi-scale sampling strategy to hierarchically integrate more motion clues to model human avatars. Extensive experiments demonstrate that our method significantly outperforms 3DGS-based approaches and renders human" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 513 + ], + "type": "text", + "content": "avatars orders of magnitude faster than the latest NeRF-based models that incorporate temporal context, all while delivering performance that is at least comparable or even superior. Project page: https://zezeaaa.github.io/projects/SeqAvatar/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 536, + 395, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 395, + 549 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 395, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 653 + ], + "type": "text", + "content": "Recent research on digital humans has highlighted the efficiency of 3D Gaussian Splatting (3DGS) [27], demonstrating its capability for high-quality and real-time rendering. By defining T-posed human Gaussians in a canonical space and employing specific warping techniques, we can render human avatars from diverse perspectives and in any pose. The canonical Gaussian primitives and warping weights are then jointly optimized under supervision from video inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "type": "text", + "content": "Animatable human avatar reconstruction is primarily hindered by limited modeling of non-rigid warping, forcing a trade-off between rendering quality and animation capability. Current methods [22, 56] incorporate SMPL(-X) pose priors [43, 51] to guide motion for each observation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*This work was done during the author's internship at Shanghai Artificial Intelligence Laboratory. † denotes co-corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13592" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "and rely on Linear Blend Skinning (LBS) for warping. Although these approaches are effective for pose-driven animation, they often struggle with per-frame non-rigid warping in scenarios involving complex garments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "type": "text", + "content": "We experimentally find that current human pose conditions [22, 56] do not fully capture the complex one-to-many mapping from pose to appearance, particularly for 3DGS-based approaches [27]. These methods [56, 68] rely on the spatial information of a template body in each frame's human pose to predict non-rigid deformations. However, they often overlook local details, such as garment deformations far from the skeleton, and cannot resolve cases where the same pose corresponds to different appearances during complex motions. Furthermore, although previous NeRF-based [44] method [8] has attempted to model motion sequences using human pose residuals, the inherently global nature of the pose sequence limits the ability to capture finer motion details. Naive pose sequence modeling does not fully account for the explicit characteristics of 3DGS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 564 + ], + "type": "text", + "content": "In this paper, we introduce a hierarchical motion context condition for 3DGS-based human avatar modeling to address the limitations of relying solely on human pose, which provides only limited global skeletal information. Our approach improves the ability of 3DGS-based methods to accurately capture the complex relationship between human pose and appearance in challenging scenarios. Specifically, we design a coarse-to-fine motion condition that incorporates both overall skeletal movements and fine-grained point-wise motions. Leveraging the explicit nature of Gaussian primitives, this condition seamlessly integrates into 3DGS-based methods, enabling more precise predictions of complex non-rigid deformations. To further enhance robustness, we propose a spatio-temporal multi-scale sampling strategy for constructing the motion context with a larger receptive field. Spatially, we consider the motion states of neighboring points within the local region of each Gaussian primitive to obtain more stable motion embeddings. Temporally, we capture human motion patterns across multiple time scales, combining long-term trends with fine-grained inter-frame motion details. This improves the model's generalization to complex human movements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 567, + 277, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 277, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 277, + 578 + ], + "type": "text", + "content": "To summarize, our key contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 579, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 51, + 579, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 579, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 51, + 579, + 295, + 628 + ], + "type": "text", + "content": "1) We propose a novel hierarchical motion condition that integrates coarse-to-fine human motion, combining global skeletal poses with localized vertex residuals to enhance non-rigid deformation prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 628, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 628, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 51, + 628, + 295, + 662 + ], + "type": "text", + "content": "2) We introduce a spatiotemporal multiscale sampling strategy that expands the receptive field of hierarchical motion context, improving generalization to complex motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 51, + 663, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 663, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 51, + 663, + 295, + 711 + ], + "type": "text", + "content": "3) Experiments on the I3D-Human [8], DNA-Rendering [9], and ZJU-MoCap [53] datasets show the effectiveness of the proposed SeqAvatar, which is capable of modeling details of the human body in complex motions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 294 + ], + "type": "text", + "content": "Neural Rendering. Neural rendering techniques have brought significant progress to human reconstruction and rendering. In particular, Neural Radiance Fields (NeRF) [44] introduces an implicit scene representation that models color and density using multilayer perceptrons, delivering photorealistic rendering results. More recently, point-based 3D Gaussian Splatting [27] utilizes 3D Gaussians to represent scenes explicitly, achieving real-time high-quality rendering. Building upon NeRF and 3DGS, subsequent works have advanced neural rendering across various dimensions, e.g., enhancing visual fidelity [1, 2], improving sparse-view reconstruction quality [10, 52, 63, 71, 80], enabling pose-free optimization [37, 58, 61, 67], modeling dynamic scenes [13, 34, 49, 55, 69, 74], and accelerating training and inference [4, 12, 18, 39, 46, 60]. These developments also benefit human avatar modeling, which demands high-quality and efficient rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 617 + ], + "type": "text", + "content": "SMPL(-X)-Based Neural Human Modeling. The SMPL(-X) family [43, 51] provides a parametric representation of the human body by decomposing it into pose-related and shape-related components using 3D mesh scans and principal component analysis (PCA). Its pose blend shapes enables body deformation through joint-wise pose blending, offering an efficient and compact way for animation. SMPL(-X) has thus become a cornerstone in human body modeling and animation, with many methods [11, 59] estimating its parameters directly from 2D inputs. Recent neural human reconstruction methods integrate SMPL(-X) with implicit representations such as NeRF [44] and 3DGS [27] to enable animatable avatars with high rendering quality. Different from purely 2D image animation methods [48, 62, 65, 73], these methods typically register the input data to a canonical T-pose space and use linear blend skinning (LBS) to transform 3D points into observation space based on SMPL(-X) poses. NeRF-based approaches [5-8, 14, 17, 19, 20, 30, 53, 64, 68, 75] focus on monocular or multi-view reconstruction. More recently, 3DGS-based methods [21-23, 25, 26, 29, 31-33, 36, 40, 41, 45, 47, 56, 76, 78, 79, 81] have gained popularity due to their real-time rendering and high fidelity. Some works [8, 23] further incorporate pose sequences as temporal context, but they suffer from slow rendering and lack fine details of motion caused by simply integrating coarse human pose embedding with NeRF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Human Rendering with Temporal Embeddings The earliest modeling of dynamic radiance field [3, 13, 15, 16, 24, 34, 35, 42, 50, 57, 70] can be traced back to temporal embeddings in general scenes, where each frame's observation is obtained by constructing a canonical space and a time-conditioned deformation field. Motivated by these, a stream of research [38, 72] focuses on pure rendering quality, employing temporal embeddings instead of human pose to en" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13593" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "code each frame of human videos. Although these methods achieve high-quality, temporally continuous rendering, they struggle with the lack of geometric constraints from human pose, making parameterized animation unfeasible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 133, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 133, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 133, + 144 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": "SMPL(-X) Series [43, 51] adopt a parameterized framework to represent human bodies across diverse shapes and poses. For each frame, the human mesh is derived by deforming a canonical template mesh based on shape and pose parameters. Specifically, a 3D point " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": " on the canonical mesh is warped to the corresponding point on the deformed mesh in the observation space, following" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 244, + 296, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 244, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 94, + 244, + 296, + 277 + ], + "type": "interline_equation", + "content": "\\mathbf {L B S} (\\mathbf {x}, \\mathbf {B} _ {k}, \\omega_ {k} (\\mathbf {x})) = \\sum_ {k = 1} ^ {K} \\omega_ {k} (\\mathbf {x}) \\mathbf {B} _ {k} \\mathbf {x}, \\tag {1}", + "image_path": "02c51a07880cad13077fef3db79af26b57cf390353de47f2d61da2d4bdd793d9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " is the total bone number and " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " is the transformation matrix for each bone. Specifically, " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " consists of a rotation and translation matrix " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " represents the global rotation of each joint, influenced by the local joint rotations " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " denotes each joint's translation, determined by the joint positions " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\boldsymbol{j}" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " and the human shape parameters " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": ". The linear blending weight " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\omega^k" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " depends on " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " and is regressed from comprehensive human meshes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": "3D Gaussian Splatting [27] utilizes a set of point-based Gaussian primitives to explicitly represent scenes, enabling real-time and high-quality rendering. Each Gaussian primitive is defined by its center position " + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": ". To ensure positive semi-definiteness and simplify optimization, the covariance matrix " + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": " is decomposed into a rotation matrix " + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": " and a scaling matrix " + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 381, + 296, + 465 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 473, + 294, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 473, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 140, + 473, + 294, + 488 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {T} \\mathbf {R} ^ {T}, \\tag {2}", + "image_path": "f167bae737c297eea5b00c5ed7bd991a519587549a018b378a5abd0d8e9b59e5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " are derived by a scaling vector " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " and a quaternion vector " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " in practice. Additionally, each Gaussian primitive is also assigned a color feature " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "sh \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " represented by spherical harmonics (SH) and an opacity " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " for rendering. During the rendering process, a splatting technique [27] is used to project each Gaussian primitive onto the 2D image space with a viewing transform " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " and the Jacobian " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{J}" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " of the projective transformation's affine approximation. The transformed covariance " + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}'" + }, + { + "bbox": [ + 55, + 497, + 296, + 616 + ], + "type": "text", + "content": " in the camera's coordinates is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 624, + 295, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 624, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 132, + 624, + 295, + 640 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {3}", + "image_path": "6c043b2f5c808c6eebf782f44ff57deed9e1c0f91c8a128daa77fdb5620646e7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "text", + "content": "After projection, the pixel color " + }, + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "text", + "content": " is computed by blending " + }, + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 648, + 296, + 673 + ], + "type": "text", + "content": " ordered Gaussian primitives overlapping at the pixel:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 118, + 682, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 682, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 118, + 682, + 295, + 715 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} ^ {\\prime} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {4}", + "image_path": "3c978aa4418931cf59570050b894fb0b3c175aba2af2ed217a0a5f1aef8fcc27.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is computed from SH feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "sh" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is the product of " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and probability density of " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "-th 2D Gaussian." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "text", + "content": "For optimization, a photometric loss is defined by a combination of " + }, + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 96, + 553, + 120 + ], + "type": "text", + "content": " and SSIM [66] losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 128, + 553, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 128, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 325, + 128, + 553, + 143 + ], + "type": "interline_equation", + "content": "L _ {\\text {p h o t o}} = \\lambda L _ {1} (\\hat {I}, I) + (1 - \\lambda) (1 - S S I M (\\hat {I}, I)), \\tag {5}", + "image_path": "a5280ef89457ea3ba7b09dfa3360b98f9b5b5635788573ad861a19476c13be77.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "content": " denote the rendered and ground-truth images, and " + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 152, + 553, + 176 + ], + "type": "text", + "content": " controls the balance between the two terms." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 186, + 370, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 186, + 370, + 199 + ], + "spans": [ + { + "bbox": [ + 314, + 186, + 370, + 199 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": "Fig. 2 shows the framework of our method. We use the explicit point-based 3DGS as the representation of the human body. Given a collection of input cameras and images, we optimize a set of Gaussian primitives " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i\\}_{i=1}^{n}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": " to fit the body's shape and appearance. Each Gaussian primitive " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": " includes the center position " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": ", scaling vector " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": ", rotation quaternion vector " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": ", color feature " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "sh" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": " is initialized from the SMPL template vertices " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_i\\}_{i=1}^{i=N}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": ". During the optimization process, we first apply a coarse-to-fine motion context to capture more accurate and fine-grained details (Sec. 4.1). To mitigate overfitting, we propose Spatio-Temporal Multi-Scale Sampling to obtain more robust point-wise Gaussian motion, which serves as the embedding for non-rigid deformation (Sec.4.2). Next, we adopt Linear Blend Skinning (LBS) to map canonical Gaussian primitives " + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 207, + 555, + 411 + ], + "type": "text", + "content": " to observation space and render images via differentiable splitting (Sec.4.3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 418, + 553, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 441 + ], + "type": "text", + "content": "4.1. Non-Rigid Deformation with Coarse-to-fine Motion Context" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "text", + "content": "Coarse Skeleton Motion. As illustrated in Fig. 1, we introduce a coarse-to-fine motion context to excavate additional conditions for non-rigid deformation. Typically, most human 3D avatar methods use the current frame pose as a condition to predict non-rigid deformation [22, 56]. While this provides spatial information to distinguish the motion of different body parts, it fails to capture temporal motion changes, as discussed in [8]. Therefore, given a frame at time " + }, + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "text", + "content": ", we consider a sequence of regularly interval-sampled frames " + }, + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 448, + 555, + 567 + ], + "type": "text", + "content": " to model inter-frame body motion variations:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 577, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 577, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 367, + 577, + 553, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {T} = \\{t - s, t - 2 s, \\dots , t - L s \\}, \\tag {6}", + "image_path": "2831e486837e344f596002be6c46bb456930ad744229cbb2c44b97048b7bbe41.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "text", + "content": " is the time interval. The coarse motion of body skeletons at each time step " + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{T}" + }, + { + "bbox": [ + 313, + 598, + 555, + 647 + ], + "type": "text", + "content": " can be derived by calculating the difference of poses between adjacent frames as illustrated in Fig. 1, following" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 357, + 654, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 654, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 357, + 654, + 553, + 669 + ], + "type": "interline_equation", + "content": "\\Delta \\mathcal {P} = \\left\\{\\Delta P ^ {t} = \\delta \\left(P ^ {t}, P ^ {t - s}\\right) | t \\in \\mathcal {T} \\right\\}, \\tag {7}", + "image_path": "f34aa3ec0ca5f7af90ed3d3325f6da66d965396e1f3f4108545bc6688dcb3142.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " is the body pose and " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " denotes the difference " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{P}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " between two poses in axis-angle form. We employ an MLP " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\Delta \\mathcal{P}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " to encode the sequential skeleton mo" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13594" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 541, + 235 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 541, + 235 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 541, + 235 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 541, + 235 + ], + "type": "image", + "image_path": "bec0715e77cfb6f8898d1bbbc367c58618f4d3b2c0dee2f2110e75d4162cb622.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. We first initialize canonical Gaussian positions " + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": " with SMPL template vertexes. For each Gaussian, we derive both coarse skeleton motion condition " + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "inline_equation", + "content": "f_{\\Delta \\mathcal{P}}" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": " and fine-grained vertex motion condition " + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "inline_equation", + "content": "f_{\\nu}" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": " that sampled from the vertex motion template " + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\nu}" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": " (points in different colors represent different motions). Based on such hierarchical motion information, we utilize an MLP " + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{non-rigid}" + }, + { + "bbox": [ + 54, + 243, + 555, + 299 + ], + "type": "text", + "content": " to better predict each Gaussian's non-rigid deformation prediction. The non-rigid deformed Gaussians are then warped into observation space via the standard LBS transformation for rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "type": "text", + "content": "tion " + }, + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{P}\\in \\mathbb{R}^{L\\times K\\times 3}" + }, + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "type": "text", + "content": ", flattened along the temporal dimension, into a skeleton motion embedding " + }, + { + "bbox": [ + 54, + 319, + 294, + 343 + ], + "type": "inline_equation", + "content": "f_{\\Delta \\mathcal{P}}\\in \\mathbb{R}^{32}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 350, + 294, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 350, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 135, + 350, + 294, + 363 + ], + "type": "interline_equation", + "content": "f _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} (\\Delta \\mathcal {P}), \\tag {8}", + "image_path": "1005eadcc03fda1370457004b772be873e1fe71a87a4897b641e4f3d7bfd9624.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 403 + ], + "type": "text", + "content": "which serves as a condition for the subsequent non-rigid deformation. Further details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "text", + "content": "Fine Vertex Motion. Compared to previous implicit NeRF-based methods, the point-based 3DGS representation enables us to explore more fine-grained temporal motion information. We derive a point-wise velocity vector " + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "text", + "content": " for each Gaussian primitive " + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "text", + "content": " to model the fine-grained localized body motion, which is beyond the scope of pose skeleton motion " + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{P}" + }, + { + "bbox": [ + 55, + 404, + 296, + 501 + ], + "type": "text", + "content": ". By measuring the position variations across adjacent time steps, we calculate per-frame velocity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 505, + 294, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 505, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 134, + 505, + 294, + 529 + ], + "type": "interline_equation", + "content": "v ^ {t} = \\frac {\\mathbf {x} _ {\\mathbf {o}} ^ {t} - \\mathbf {x} _ {\\mathbf {o}} ^ {t - s}}{s}, \\tag {9}", + "image_path": "2fc3eba92f3b337f83afdcae188630b0826a0ed03358c67a35cfc7160a8c97b6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x_o}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " denotes the warped coordinates in observation space. However, since the positions of Gaussian primitives are updated during optimization, it is not stable to obtain " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " based on such a dynamic variable. Moreover, transforming the canonical Gaussians into the observation space requires non-rigid deformation first, which leads to a circular dependency and conflicts with the current velocity design. To this end, we introduce a motion template field " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{V}} = \\{\\mathbf{V}_i\\}_{i = 1}^{i = N}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " that stores the velocity of each SMPL vertex, and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " for each Gaussian primitive is sampled from " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{V}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " based on the distance between query points and vertexes in this template. Specifically, given a time step " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and the corresponding body pose " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "P^t = \\{(\\mathbf{R}_k^t,\\mathbf{b}_k^t)|k\\in K\\}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", the SMPL template vertexes in " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " can be warped into observation space with the template skinning weights " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 319, + 482, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 482, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 482, + 332 + ], + "type": "text", + "content": "standard linear blend skinning, following" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 338, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 338, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 381, + 338, + 553, + 352 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\mathbf {o}} ^ {t} = \\mathbf {L B S} (\\mathbf {T}, \\mathbf {B} ^ {t}, \\mathbf {W}), \\tag {10}", + "image_path": "0ac709ec417c0b08115e74d9fea908de383678c5134296d2001cc24d14e1bd40.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^t" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": " is the transformation matrix derived from rotation and translation matrix " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^t" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": " is the template SMPL LBS weights. The velocity " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": " of each template vertexes " + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 357, + 555, + 405 + ], + "type": "text", + "content": " can be further derived as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 389, + 409, + 553, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 409, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 389, + 409, + 553, + 435 + ], + "type": "interline_equation", + "content": "\\mathbf {V} ^ {t} = \\frac {\\mathbf {T} _ {\\mathbf {o}} ^ {t} - \\mathbf {T} _ {\\mathbf {o}} ^ {t - s}}{s}. \\tag {11}", + "image_path": "44fa2da0e4cf882bf30d1d49eefe0906929eb9d63162fae3349e6869c9f73ed5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 440, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 464 + ], + "type": "text", + "content": "We then sample each Gaussian's velocity from this motion template field " + }, + { + "bbox": [ + 313, + 440, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathbf{V}}" + }, + { + "bbox": [ + 313, + 440, + 554, + 464 + ], + "type": "text", + "content": ", which is discussed in Sec. 4.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 470, + 554, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 484 + ], + "type": "text", + "content": "4.2. Spatio-temporal Multi-scale Sampling (STMS)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 548 + ], + "type": "text", + "content": "To enhance the robustness of human motion conditions, we propose modeling both local region motion information of the human body and motion patterns across different temporal windows to mitigate overfitting and improve generalization by capturing more comprehensive motion dynamics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": "In the spatial dimension, we sample the " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": " nearest template to model the body's local region motion more robustly. Specifically, for each canonical Gaussian primitive " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": " nearest vertexes' velocities are sampled as input to an MLP " + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{knn}" + }, + { + "bbox": [ + 313, + 548, + 554, + 608 + ], + "type": "text", + "content": " to learn a local point-wise motion embedding" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 351, + 613, + 553, + 628 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {t} = \\mathcal {E} _ {k n n} \\left(\\left\\{\\mathbf {V} _ {j} ^ {t} \\right\\}\\right), \\quad j \\in \\operatorname {K N N} (\\mathbf {T}, \\mathbf {x} _ {i}), \\tag {12}", + "image_path": "b3b89071dd9418663ca2e6a33aa3322ec2b9de0aa13499583375f7aa0f9203c5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{KNN}(\\mathbf{T},\\mathbf{x}_i)" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": " nearest SMPL template vertexes of the canonical Gaussian position " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": ". Similar to Eq. 7, we then apply an MLP " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{V}}" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": " to encode each Gaussian primitive's sequential motion embedding " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{e^{t}|t\\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "text", + "content": " into a point-wise sequential condition " + }, + { + "bbox": [ + 313, + 634, + 554, + 694 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{V}}\\in \\mathbb{R}^{96}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 406, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 406, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\mathcal {V}). \\tag {13}", + "image_path": "ca891215303bd367be09f0aae895c5468b2ba0c1e58be62a4b6d22bafaa73005.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13595" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "In the temporal dimension, to capture both the overall motion trend and the inter-frame motion details, we adopt a multi-scale sequence sampling strategy. In detail, we sample a series of sequences at several progressively increasing intervals to get human motion information across different temporal windows" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 152, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 152, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 121, + 152, + 295, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{s = s _ {0} + i \\Delta s \\right\\} _ {i = 0} ^ {i = m}, \\tag {14}", + "image_path": "41b065d9ca8f3ed2876f753e9fc55903a0d38a095ff77875554039c362ad459e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\Delta s" + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "content": " is the increasing rate of sampling interval " + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "content": ". We then input the multi-scale sampled sequence motions into the skeleton motion encoder " + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\Delta P}" + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "content": " and the point-wise motion encoder " + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{V}" + }, + { + "bbox": [ + 55, + 172, + 296, + 233 + ], + "type": "text", + "content": " to obtain hierarchical temporal motion embeddings. Therefore, Eq. 7 and Eq. 13 can be revised as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 240, + 295, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 240, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 111, + 240, + 295, + 253 + ], + "type": "interline_equation", + "content": "f _ {\\Delta \\mathcal {P}} = \\mathcal {E} _ {\\Delta \\mathcal {P}} \\left(\\left\\{\\Delta \\mathcal {P} _ {s} \\right\\}\\right), \\quad s \\in \\mathcal {S} \\tag {15}", + "image_path": "516cf2833c6feb42ce168f95fdd60713b08d4cbbb3c3c73bebd2396565d58315.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 255, + 295, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 255, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 119, + 255, + 295, + 269 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {V}} = \\mathcal {E} _ {\\mathcal {V}} (\\{\\mathcal {V} _ {s} \\}), \\quad s \\in \\mathcal {S}. \\tag {16}", + "image_path": "8e3e8ac440222a586ba6900ddb8143dc98905ccd1ba074cab0d3c644ecf969bd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": "In practice, we concatenate all skeleton motion conditions, " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{P}_s" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "s\\in S" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": ", and localized vertex motion conditions, " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\nu_{s}" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "s\\in S" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": ", across different sampling scales. These are then input to " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\Delta P}" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\nu}" + }, + { + "bbox": [ + 55, + 276, + 296, + 335 + ], + "type": "text", + "content": ", respectively. More details are shown in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": "Non-Rigid Deformation. Given the coarse skeleton motion " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "f_{\\Delta \\mathcal{P}} = \\mathcal{E}_{\\Delta \\mathcal{P}}(\\{\\Delta \\mathcal{P}_s\\})" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " condition and fine point-wise velocity condition " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{V}} = \\mathcal{E}_{\\mathcal{V}}(\\{\\mathcal{V}_s\\})" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": ", we utilize an MLP to predict each Gaussian's non-rigid deformation as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 391, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 391, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 93, + 391, + 295, + 405 + ], + "type": "interline_equation", + "content": "\\delta \\mathbf {x}, \\delta \\mathbf {s}, \\delta \\mathbf {r} = \\mathcal {E} _ {\\text {n o n - r i g i d}} (\\mathbf {x}, P, f _ {\\Delta \\mathcal {P}}, f _ {v}). \\tag {17}", + "image_path": "2ea5d49cc2fd5e25e35383ce22c476e8a0c2a8a32540bf2c6ba7b2e97bec4771.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 411, + 216, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 216, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 216, + 423 + ], + "type": "text", + "content": "The deformed canonical Gaussian " + }, + { + "bbox": [ + 55, + 411, + 216, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{G}'" + }, + { + "bbox": [ + 55, + 411, + 216, + 423 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 431, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 431, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 146, + 431, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = \\mathbf {x} + \\delta \\mathbf {x}, \\tag {18}", + "image_path": "ffdbff1c233cb27fb5c3ce3f4ec3920b87334182ab977b127842111fbb8d9296.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 148, + 447, + 294, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 447, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 148, + 447, + 294, + 459 + ], + "type": "interline_equation", + "content": "\\mathbf {s} ^ {\\prime} = \\mathbf {s} + \\delta \\mathbf {s}, \\tag {19}", + "image_path": "2ce3d6570b9ef43bf6374de8c3ddf27c25d8e27c8dfbe25b7e2b2e08da89dd2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 148, + 461, + 294, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 461, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 148, + 461, + 294, + 473 + ], + "type": "interline_equation", + "content": "\\mathbf {r} ^ {\\prime} = \\mathbf {r} \\cdot \\delta \\mathbf {r}, \\tag {20}", + "image_path": "aebf586addb60aa80feed33a4076d20065ed5f599061b3bc708f32f1edfde91a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 482, + 274, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 274, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 274, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 482, + 274, + 494 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 482, + 274, + 494 + ], + "type": "text", + "content": " denotes the multiplication of two quaternions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 501, + 141, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 141, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 141, + 514 + ], + "type": "text", + "content": "4.3. Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 518, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 554 + ], + "type": "text", + "content": "Rigid Deformation. We utilize the standard LBS operation to map the non-rigid deformed Gaussians " + }, + { + "bbox": [ + 55, + 518, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 55, + 518, + 296, + 554 + ], + "type": "text", + "content": " into the observation space, following" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 562, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 562, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 127, + 562, + 295, + 575 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathbf {o}} = \\mathbf {L B S} \\left(\\mathbf {x} ^ {\\prime}, \\mathbf {B}, \\omega\\right), \\tag {21}", + "image_path": "e1e7dc084d8dfbbfd373128534dd08c3d653458303fb596b9a948c4447e18bdc.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 578, + 295, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 578, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 125, + 578, + 295, + 610 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {\\mathbf {o}} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\mathbf {x} ^ {\\prime}\\right) \\mathbf {B} _ {k} \\mathbf {R} ^ {\\prime}, \\tag {22}", + "image_path": "fd509fbec1cb0f1b9e6b40dd70ea4979cd65a2106cc4b9cc3570271e97232141.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R}'" + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "content": " is the rotation matrix derived from the non-rigid deformed Gaussian's rotation quaternion " + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{r}'" + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R_o}" + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "content": " represents the rotation matrix in observation space. Following the previous method [22], we utilize an MLP " + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{lbs}" + }, + { + "bbox": [ + 55, + 617, + 296, + 700 + ], + "type": "text", + "content": " to predict the LBS weight offsets for each query canonical point and update the sampled weights from the nearest SMPL vertex as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\omega_ {k} (\\mathbf {x}) = \\omega_ {k} ^ {S M P L} (\\mathbf {x}) + \\mathcal {E} _ {l b s} (\\mathbf {x}). \\tag {23}", + "image_path": "23a2ccf0a3b582927c712496bdcd29fd63f6d6673efc0ee2e26d7ece98aaad7a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "Similar to [8, 22, 68], we introduce a pose refinement MLP " + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\text {pose }}" + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": " to refine the pose estimate from SMPL for a better fit to the human body." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 108, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 144 + ], + "type": "text", + "content": "Loss Function. With the transformed Gaussian primitives " + }, + { + "bbox": [ + 313, + 108, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_o" + }, + { + "bbox": [ + 313, + 108, + 555, + 144 + ], + "type": "text", + "content": " in observation space, we apply the standard splitting [27] to render images as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 154, + 555, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 154, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 361, + 154, + 555, + 167 + ], + "type": "interline_equation", + "content": "I = \\mathbf {S p l a t t i n g} \\left(\\mathbf {x} _ {\\mathbf {o}}, \\mathbf {R} _ {\\mathbf {o}}, s ^ {\\prime}, \\alpha , s h\\right). \\tag {24}", + "image_path": "5065e9af7a31f84363f3d91bd4c54bc712dfc5b8b90a18bea4dd988d43c93bb2.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 176, + 554, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 554, + 200 + ], + "type": "text", + "content": "During the optimization process, we employ a combination of loss functions as supervision, summarized as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 328, + 210, + 555, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 210, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 328, + 210, + 555, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {s s i m}} + \\lambda_ {3} \\mathcal {L} _ {\\text {l p i p s}} + \\mathcal {L} _ {\\text {m a s k}}, \\tag {25}", + "image_path": "a7c9f809000ca97f2aa8246005a64d5efb98f2f18e97a550e28ba900461684d8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " is used to balance the weight of different losses. " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " [22] is an " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " loss between the rendered " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " and human body mask. Similar to [56], we apply " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{isopos}" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{isocov}" + }, + { + "bbox": [ + 313, + 232, + 554, + 293 + ], + "type": "text", + "content": " to control the Gaussian primitive's position and covariance. Details are shown in the supplementary material." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 303, + 395, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 303, + 395, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 303, + 395, + 316 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 323, + 376, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 376, + 334 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 376, + 334 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 411 + ], + "type": "text", + "content": "DNA-Rendering [9] is a challenging human-centric rendering dataset. It contains diverse scenes from everyday life to professional occasions. We use 6 sequences with loose-fitting garments (1_0206_04, 2_0007_04, 2_0019_10, 2_0044_11, 2_0051_09, 2_0813_05) for experiments. We adopt 24 views for training and 6 views for testing." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 412, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 484 + ], + "type": "text", + "content": "13D-Human [8] contains scenes closer to daily life, which comprises multi-view video frames of humans with loose clothing and complex movements. We conduct experiments on 4 sequences and use 4-5 views for training and the rest for testing. All training and testing data are retained to be the same in comparisons following Dyco's [8] data split." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 491, + 440, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 440, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 440, + 502 + ], + "type": "text", + "content": "5.2. Baselines and Metrics" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": "We compare against the state-of-the-art human modeling methods including Dyco [8], 3DGS-Avatar [56], GART [31], and GauHuman [22]. Dyco is an implicit NeRF-based method that adopts pose sequence information for temporal modeling. We compare performance with Dyco to showcase the better efficacy of our fine-grained motion condition design. 3DGS-Avatar, GART, and GauHuman are explicit 3DGS-based methods, and all of them are originally designed for monocular inputs. For fairness, we extend them to multi-view input under the same settings. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [66], and learned perceptual image patch similarity (LPIPS) [77]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 672, + 395, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 672, + 395, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 672, + 395, + 685 + ], + "type": "text", + "content": "5.3. Comparison" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Comparison on DNA-Rendering. Tab. 1 shows the quantitative results on DNA-Rendering dataset. The proposed" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13596" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative Results on DNA-Rendering Dataset. We report the performance of novel view rendering. Our method outperforms previous state-of-the-art human modeling methods. We mark the best and the second best methods in cells. LPIPS*=LPIPS × 10^3." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 152 + ], + "type": "table", + "html": "
Subject:1_0206_042_0007_042_0019_102_0044_112_0051_092_0813_05
Metric:PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]26.640.943950.0928.060.949256.7032.060.971031.1228.230.951237.1525.040.953442.2231.750.970131.30
GART [31]27.190.945456.4728.220.952556.4631.130.968635.1628.900.956141.2726.470.962842.2132.060.972935.74
GauHuman [22]27.960.950050.4227.930.949656.9331.630.969832.8230.380.964134.1925.990.959942.0033.430.976429.42
Ours30.810.964939.4729.630.956643.7234.970.977723.5332.620.974723.8328.450.968534.0635.810.984320.83
", + "image_path": "8b2e7d7e3bd8c797a92ae9a148e925229f3a6de81cbdbcc5c324a702206cc45b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 164, + 523, + 571 + ], + "blocks": [ + { + "bbox": [ + 86, + 164, + 523, + 571 + ], + "lines": [ + { + "bbox": [ + 86, + 164, + 523, + 571 + ], + "spans": [ + { + "bbox": [ + 86, + 164, + 523, + 571 + ], + "type": "image", + "image_path": "9054b106fd64f5802227d0a6ead54f294f28323c1bd85cce0257cb682b76d4c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 582, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 582, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 555, + 605 + ], + "type": "text", + "content": "Figure 3. Novel View Qualitative Results on DNA-Rendering. We zoom into the local region and compute the error maps compared with ground truth images. The results show that our method achieves competitive results on both the overall and local region qualities." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 625, + 297, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 297, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 297, + 711 + ], + "type": "text", + "content": "method outperforms previous state-of-the-art human modeling methods across all metrics. To demonstrate the visual improvement, we compare the quality of rendering images in Fig. 3. Previous methods, which lack the temporal motion information for non-rigid deformations, are unable to predict the movement of Gaussian primitives properly, resulting in blurred renderings in regions with complex tex" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 625, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 556, + 711 + ], + "type": "text", + "content": "tures (scene 1_0206_04 in Fig. 3). Besides, the results in Fig. 3 show that the compared methods fail to capture the proper shape variation caused by movement (e.g., the white dress in scene 2_0051_09). Although these methods achieve a smooth appearance in these areas, the error maps show that they fail to match the actual shape closely. We leverage the hierarchical motion context to predict Gaussian defor" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13597" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 91, + 526, + 153 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 436, + 82 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 436, + 82 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 436, + 82 + ], + "type": "text", + "content": "Table 2. Novel View Rendering Quantitative Results on I3D-Human." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 91, + 526, + 153 + ], + "lines": [ + { + "bbox": [ + 82, + 91, + 526, + 153 + ], + "spans": [ + { + "bbox": [ + 82, + 91, + 526, + 153 + ], + "type": "table", + "html": "
Subject: Metric:ID1_1ID1_2ID2_1ID3_1
PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓PSNR↑SSIM↑LPIPS*↓
3DGS-Avatar [56]29.780.960031.4031.290.962728.0929.630.960437.3332.730.960239.47
Dyco [8]30.810.961728.2931.450.962825.7529.410.961833.2632.500.961235.46
GauHuman [22]29.470.957139.4630.420.958538.8728.510.954649.4632.100.954653.67
Ours31.810.965927.0331.980.966027.8431.530.968530.2333.620.965134.02
", + "image_path": "b2a6a4f54f86aadd0d3ac9c932cff8889dd3c76bb174f9060aadf46058cd4f3a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 164, + 541, + 457 + ], + "blocks": [ + { + "bbox": [ + 70, + 164, + 541, + 457 + ], + "lines": [ + { + "bbox": [ + 70, + 164, + 541, + 457 + ], + "spans": [ + { + "bbox": [ + 70, + 164, + 541, + 457 + ], + "type": "image", + "image_path": "be622cd66f8e05671fbb193d1ca59bb47ff345a8f7d0fbda373a3ec6015a235f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 555, + 489 + ], + "type": "text", + "content": "Figure 4. Novel Pose Qualitative Results on I3D-Human. We compare the proposed method with previous SOTA approaches. The rendered images and error maps demonstrate our robustness on novel pose rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 533 + ], + "type": "text", + "content": "mations, allowing our method to capture detailed appearance variations caused by human motions more accurately." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Comparison on I3D-Human. The quantitative results in Tab. 2 show that our method surpasses the previous SOTA method on most metrics, demonstrating our competitive performance on the I3D-Human dataset. Previous NeRF-based Dyco [8] utilizes pose variation as a condition to model human motions. However, it is limited to capturing only overall body motions and lacks the capacity for finer-grained modeling in local regions. For example, in ID1_1 and ID2_1 shown in Fig. 8 (in the Supp. Mat.), our method renders results that accurately reflect the true motion, while Dyco fails to model the information in regions far from the human body skeleton. Tab. 3 reports the quantitative results of novel pose rendering on the I3D-Human dataset. Our method achieves best performance among 3DGS-based human modeling methods and also enables real-time render-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "content": "ing (" + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "content": " 45 FPS on I3D-Human) compared to NeRF-based Dyco (" + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "content": " 0.7 FPS). Fig. 4 shows the qualitative comparisons of novel pose rendering. Different from Dyco [8], which relies solely on joint motions as conditions, our method leverages the motion state of each Gaussian primitive to predict deformations, enabling more flexible modeling of motions in regions distant from the human body." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 628 + ], + "type": "text", + "content": "To evaluate the generalizability, we also conduct experiments on ZJU-MoCap [53]. Please refer to the supplementary materials for more details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Out-of-Distribution Poses. To evaluate the generalization ability to novel poses, we conduct experiments using animations with large pose variations. Specifically, we use a model trained on one sequence to render target poses sampled from other unseen sequences, as different sequences in DNA-Rendering and I3D-Human contain distinct motion patterns. As illustrated in Fig. 6, the Target Pose is" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13598" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 152, + 145 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 152, + 145 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 152, + 145 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 152, + 145 + ], + "type": "image", + "image_path": "3d7b24e057c3c73a0d3f38b70399850617ea366f5d266a48053f7b27dbbbb948.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 148, + 126, + 155 + ], + "lines": [ + { + "bbox": [ + 94, + 148, + 126, + 155 + ], + "spans": [ + { + "bbox": [ + 94, + 148, + 126, + 155 + ], + "type": "text", + "content": "(a). Baseline" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 147, + 237, + 155 + ], + "lines": [ + { + "bbox": [ + 154, + 147, + 237, + 155 + ], + "spans": [ + { + "bbox": [ + 154, + 147, + 237, + 155 + ], + "type": "text", + "content": "(b). (a) + Vanilla Non-Rigid MLP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 163, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 553, + 175 + ], + "type": "text", + "content": "Figure 5. Qualitative Ablation Results. We compare the visual influence of different components on ID2_1 scene of I3D-Human dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 154, + 70, + 239, + 145 + ], + "blocks": [ + { + "bbox": [ + 154, + 70, + 239, + 145 + ], + "lines": [ + { + "bbox": [ + 154, + 70, + 239, + 145 + ], + "spans": [ + { + "bbox": [ + 154, + 70, + 239, + 145 + ], + "type": "image", + "image_path": "77cb632348fd966e2ec4cdd19e16e83b4a1031ec1909fadc4a307f618a192f5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 242, + 70, + 324, + 145 + ], + "blocks": [ + { + "bbox": [ + 242, + 70, + 324, + 145 + ], + "lines": [ + { + "bbox": [ + 242, + 70, + 324, + 145 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 324, + 145 + ], + "type": "image", + "image_path": "1e66811fefeb560e3053e1d6e714909949085b0110c6b098902c07685cc5c841.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 147, + 311, + 155 + ], + "lines": [ + { + "bbox": [ + 251, + 147, + 311, + 155 + ], + "spans": [ + { + "bbox": [ + 251, + 147, + 311, + 155 + ], + "type": "text", + "content": "(c). " + }, + { + "bbox": [ + 251, + 147, + 311, + 155 + ], + "type": "inline_equation", + "content": "(\\mathbf{b}) + \\Delta \\mathcal{P}" + }, + { + "bbox": [ + 251, + 147, + 311, + 155 + ], + "type": "text", + "content": " Conditions" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 327, + 70, + 410, + 145 + ], + "blocks": [ + { + "bbox": [ + 327, + 70, + 410, + 145 + ], + "lines": [ + { + "bbox": [ + 327, + 70, + 410, + 145 + ], + "spans": [ + { + "bbox": [ + 327, + 70, + 410, + 145 + ], + "type": "image", + "image_path": "9f6aca3ad615cebc7b57dd9d9f561befdb91306a9d44655b7650007386329569.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 147, + 395, + 155 + ], + "lines": [ + { + "bbox": [ + 340, + 147, + 395, + 155 + ], + "spans": [ + { + "bbox": [ + 340, + 147, + 395, + 155 + ], + "type": "text", + "content": "(d).(c) " + }, + { + "bbox": [ + 340, + 147, + 395, + 155 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 340, + 147, + 395, + 155 + ], + "type": "text", + "content": " V Conditions" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 70, + 496, + 145 + ], + "blocks": [ + { + "bbox": [ + 413, + 70, + 496, + 145 + ], + "lines": [ + { + "bbox": [ + 413, + 70, + 496, + 145 + ], + "spans": [ + { + "bbox": [ + 413, + 70, + 496, + 145 + ], + "type": "image", + "image_path": "7131523b57af1d890fc681715c8dfb5f5dd9eb7d70bbfc7fbbc72fbeb68e00af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 147, + 475, + 155 + ], + "lines": [ + { + "bbox": [ + 435, + 147, + 475, + 155 + ], + "spans": [ + { + "bbox": [ + 435, + 147, + 475, + 155 + ], + "type": "text", + "content": "(e). (d) + STMS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 500, + 70, + 542, + 145 + ], + "blocks": [ + { + "bbox": [ + 500, + 70, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 500, + 70, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 500, + 70, + 542, + 145 + ], + "type": "image", + "image_path": "07bf9af8d75eff5f0e90590ce72132077306b8b223db98bc4613292a359e5204.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 147, + 525, + 154 + ], + "lines": [ + { + "bbox": [ + 515, + 147, + 525, + 154 + ], + "spans": [ + { + "bbox": [ + 515, + 147, + 525, + 154 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 57, + 194, + 138, + 262 + ], + "blocks": [ + { + "bbox": [ + 57, + 194, + 138, + 262 + ], + "lines": [ + { + "bbox": [ + 57, + 194, + 138, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 194, + 138, + 262 + ], + "type": "image", + "image_path": "013197955407564466371049b0137826d77b0e8ab6c9e346bb82deaa4104c121.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 262, + 102, + 266 + ], + "lines": [ + { + "bbox": [ + 90, + 262, + 102, + 266 + ], + "spans": [ + { + "bbox": [ + 90, + 262, + 102, + 266 + ], + "type": "text", + "content": "ID3" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 57, + 267, + 137, + 335 + ], + "blocks": [ + { + "bbox": [ + 57, + 267, + 137, + 335 + ], + "lines": [ + { + "bbox": [ + 57, + 267, + 137, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 137, + 335 + ], + "type": "image", + "image_path": "c84b07fde1f292219b4f95bd820a564cdcb49ad21bd55453e2cbd3782f059b85.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 345, + 263, + 356 + ], + "lines": [ + { + "bbox": [ + 87, + 345, + 263, + 356 + ], + "spans": [ + { + "bbox": [ + 87, + 345, + 263, + 356 + ], + "type": "text", + "content": "Figure 6. Out-of-distribution Pose Animation." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 142, + 194, + 169, + 261 + ], + "blocks": [ + { + "bbox": [ + 142, + 194, + 169, + 261 + ], + "lines": [ + { + "bbox": [ + 142, + 194, + 169, + 261 + ], + "spans": [ + { + "bbox": [ + 142, + 194, + 169, + 261 + ], + "type": "image", + "image_path": "5dded8ec12a23f0aa1b9f42bdb1eb89c0300143f6228cefc56cf797a2179a1c1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 142, + 262, + 169, + 335 + ], + "blocks": [ + { + "bbox": [ + 142, + 262, + 169, + 335 + ], + "lines": [ + { + "bbox": [ + 142, + 262, + 169, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 262, + 169, + 335 + ], + "type": "image", + "image_path": "df8e7eaa707ddfadf0b196a0e5141ce265c2e9d66ffae5be89fe41e5f06acc5d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 175, + 194, + 256, + 262 + ], + "blocks": [ + { + "bbox": [ + 175, + 194, + 256, + 262 + ], + "lines": [ + { + "bbox": [ + 175, + 194, + 256, + 262 + ], + "spans": [ + { + "bbox": [ + 175, + 194, + 256, + 262 + ], + "type": "image", + "image_path": "0395ae9202ab5eaaad448ae0c2e726e5d38009f73fbe0a955aec1d1409f1cb08.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 175, + 262, + 255, + 335 + ], + "blocks": [ + { + "bbox": [ + 175, + 262, + 255, + 335 + ], + "lines": [ + { + "bbox": [ + 175, + 262, + 255, + 335 + ], + "spans": [ + { + "bbox": [ + 175, + 262, + 255, + 335 + ], + "type": "image", + "image_path": "fb60a18e0b5b2e55e5c8ba3ec0a12a4c2a0526fbd32960f026f08c9eaa49fb6d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 261, + 194, + 288, + 261 + ], + "blocks": [ + { + "bbox": [ + 261, + 194, + 288, + 261 + ], + "lines": [ + { + "bbox": [ + 261, + 194, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 261, + 194, + 288, + 261 + ], + "type": "image", + "image_path": "c79bd81aa1de4777ea49ab439030998b14d44aad444f6d206f1e125ae47e8d9c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 262, + 290, + 268 + ], + "lines": [ + { + "bbox": [ + 261, + 262, + 290, + 268 + ], + "spans": [ + { + "bbox": [ + 261, + 262, + 290, + 268 + ], + "type": "text", + "content": "2_0051_09" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 260, + 269, + 291, + 327 + ], + "blocks": [ + { + "bbox": [ + 260, + 269, + 291, + 327 + ], + "lines": [ + { + "bbox": [ + 260, + 269, + 291, + 327 + ], + "spans": [ + { + "bbox": [ + 260, + 269, + 291, + 327 + ], + "type": "image", + "image_path": "41ce47e0e9dd79d7bf75c0f64c38a965b2b6dd6ff37eedf4b12354d590ac3321.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 328, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 263, + 328, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 263, + 328, + 286, + 335 + ], + "type": "text", + "content": "1_0206_04" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "type": "text", + "content": "drawn from an unseen sequence exhibiting different human motions. The results suggest that our method is capable of generalizing to these significant pose changes, even though such poses were not observed during training." + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 92, + 456, + 257, + 507 + ], + "blocks": [ + { + "bbox": [ + 62, + 436, + 288, + 448 + ], + "lines": [ + { + "bbox": [ + 62, + 436, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 288, + 448 + ], + "type": "text", + "content": "Table 3. Metrics of Novel Pose Renderings on I3D-Human." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 456, + 257, + 507 + ], + "lines": [ + { + "bbox": [ + 92, + 456, + 257, + 507 + ], + "spans": [ + { + "bbox": [ + 92, + 456, + 257, + 507 + ], + "type": "table", + "html": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
Dyco [8]30.090.956935.32
3DGS-Avatar [56]29.540.955538.59
GauHuman [22]29.310.952648.29
Ours30.280.958336.07
", + "image_path": "ec627ef9db7f027ab73918fed83e48b8c93a859166f985ba3cd27b5ac98ad9e6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 149, + 540 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": "In this section, we ablate the influence of various components in our methods on I3D-Human dataset. Tab. 4 shows the average metrics of 4 sequences under different settings. Skeleton motion condition " + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\Delta \\mathcal{P}" + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": " denotes the overall human body movement. Compared to predicting non-rigid deformation only with the Gaussian position, incorporating this additional global temporal information allows our method to better model the overall human shape and achieve higher rendering quality, as demonstrated in Tab. 4 (c)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Fine-grained point-wise motion condition " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " describes the local region's variation in a more detailed manner. Such point-based motion information captures the relationship between motion and appearance variations in areas beyond the human skeleton. Tab. 4 (d) and Fig. 5 (d) show that our" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 196, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 553, + 316 + ], + "type": "text", + "content": "method achieves higher rendering details in local regions. Spatio-temporal Multi-scale Sampling (STMS) is designed to enhance the robustness of human motion conditions by incorporating local region motion information and motion patterns across different temporal windows. With such a hierarchical motion context embedding design, our method is able to predict each Gaussian's deformation more robustly, leading to better rendering quality as shown in Tab. 4 (e) and Fig. 5 (e). Please refer to the supplementary materials for more detailed ablation experiments." + } + ] + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 317, + 378, + 552, + 435 + ], + "blocks": [ + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "type": "text", + "content": "Table 4. Quantitative Ablation Results on I3D-Human. (a) denotes experiments without non-rigid deformation. (b) only utilizes Gaussian positions " + }, + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "type": "text", + "content": " and the current frame pose " + }, + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 324, + 554, + 369 + ], + "type": "text", + "content": " as the conditions for non-rigid MLP." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 378, + 552, + 435 + ], + "lines": [ + { + "bbox": [ + 317, + 378, + 552, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 552, + 435 + ], + "type": "table", + "html": "
MethodsPSNR ↑SSIM ↑LPIPS* ↓
(a) Baseline29.760.956938.35
(b) (a)+Vanilla Non-Rigid MLP+P Conditions.31.050.961734.35
(c) (b)+ΔP Conditions.31.890.964532.17
(d) (c)+V Conditions.32.010.965131.23
(e) (d)+STMS. (Ours)32.240.966429.78
", + "image_path": "2793141f88d497f19b980417c142c33ee67bf7b558a74b70228e5922db6af7e8.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 454, + 471, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 471, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 471, + 466 + ], + "type": "text", + "content": "6. Limitations and Conclusion" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": "Limitations. The Gaussian-based representation used in our method may introduce slight blur artifacts during rendering, whereas NeRF's ray-based volumetric integration tends to produce sharper results. Moreover, our local velocity cues are derived from the coarse SMPL model rather than dense surface tracking [79], which may limit the accuracy of fine-scale garment deformation. Addressing these limitations remains an open challenge for future work." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose a 3DGS-based framework that integrates hierarchical motion context for 3D human modeling. Specifically, the non-rigid deformation for the Gaussian primitive is learned based on the global human skeleton variations and fine-grained Gaussian's point-wise motions. To capture each Gaussian primitive's motion information more robustly, we introduce a spatio-temporal multi-scale sampling strategy to incorporate both local region motion features and motion patterns across different temporal intervals. Through the above design, the proposed method achieves state-of-the-art rendering of human avatars with complex garments and motions." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13599" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 117 + ], + "type": "text", + "content": "This research was supported by Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 114, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 114, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 114, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 270 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "type": "text", + "content": "[3] Ang Cao and Justin Johnson. Hexplane: A Fast Representation for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 407 + ], + "type": "text", + "content": "[5] Jianchuan Chen, Ying Zhang, Di Kang, Xuefei Zhe, Lin chao Bao, Xu Jia, and Huchuan Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 464 + ], + "type": "text", + "content": "[6] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 520 + ], + "type": "text", + "content": "[7] Xu Chen, Tianjian Jiang, Jie Song, Max Rietmann, Andreas Geiger, Michael J Black, and Otmar Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 11796-11809, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[8] Yutong Chen, Yifan Zhan, Zhihang Zhong, Wei Wang, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Within the dynamic context: Inertia-aware 3d human modeling with pose sequence. arXiv preprint arXiv:2403.19160, 2024. 2, 3, 5, 7, 8, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "type": "text", + "content": "[9] Wei Cheng, Ruixiang Chen, Siming Fan, Wanqi Yin, Keyu Chen, Zhongang Cai, Jingbo Wang, Yang Gao, Zhengming Yu, Zhengyu Lin, et al. Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 5, 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[10] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and robust multi-person 3d pose" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "estimation and tracking from multiple views. In T-PAMI, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit Radiance Fields in Space, Time, and Appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "text", + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[15] Wanshui Gan, Hongbin Xu, Yi Huang, Shifeng Chen, and Naoto Yokoya. V4D: Voxel for 4D Novel View Synthesis. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[16] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic View Synthesis from Dynamic Monocular Video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 399 + ], + "type": "text", + "content": "[17] Qingzhe Gao, Yiming Wang, Libin Liu, Lingjie Liu, Christian Theobalt, and Baoquan Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 433 + ], + "type": "text", + "content": "[18] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 489 + ], + "type": "text", + "content": "[19] Chen Geng, Sida Peng, Zhen Xu, Hujun Bao, and Xiaowei Zhou. Learning neural volumetric representations of dynamic humans in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 545 + ], + "type": "text", + "content": "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 546, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 612 + ], + "type": "text", + "content": "[21] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "text", + "content": "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[23] Tao Hu, Fangzhou Hong, and Ziwei Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13600" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[24] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial Radiance Fields for Dynamic Scenes. arXiv preprint arXiv:2212.02375, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "text", + "content": "[25] Rohit Jena, Ganesh Subramanian Iyer, Siddharth Choudhary, Brandon Smith, Pratik Chaudhari, and James Gee. Splatarmor: Articulated gaussian splatting for animatable humans from monocular rgb videos. arXiv preprint arXiv:2311.10812, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 296, + 215 + ], + "type": "text", + "content": "[26] HyunJun Jung, Nikolas Brasch, Jifei Song, Eduardo Perez-Pellitero, Yiren Zhou, Zhihao Li, Nassir Navab, and Benjamin Busam. Deformable 3d gaussian splatting for animatable human avatars. arXiv preprint arXiv:2312.15059, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 259 + ], + "type": "text", + "content": "[27] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 294 + ], + "type": "text", + "content": "[28] Martin Kilian, Niloy J Mitra, and Helmut Pottmann. Geometric modeling in shape space. In ACM SIGGRAPH 2007 papers, pages 64-es. 2007. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "type": "text", + "content": "[29] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[30] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[31] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 492 + ], + "type": "text", + "content": "[32] Mingwei Li, Jiachen Tao, Zongxin Yang, and Yi Yang. Human101: Training 100+ fps human gaussians in 100s from 1 view. arXiv preprint arXiv:2312.15258, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "text", + "content": "[33] Mengtian Li, Shengxiang Yao, Zhifeng Xie, Keyu Chen, and Yu-Gang Jiang. Gaussianbody: Clothed human reconstruction via 3d gaussian splatting. arXiv preprint arXiv:2401.09720, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 603 + ], + "type": "text", + "content": "[34] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3D Video Synthesis from Multi-view Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 658 + ], + "type": "text", + "content": "[35] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural Scene Flow Fields for Space-time View Synthesis of Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[36] Zhe Li, Zerong Zheng, Lizhen Wang, and Yebin Liu. Animatable gaussians: Learning pose-dependent gaussian maps for high-fidelity human avatar modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, pages 5741-5751, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[38] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[39] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. pages 15651-15663, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[40] Xinqi Liu, Chenming Wu, Xing Liu, Jialun Liu, Jinbo Wu, Chen Zhao, Haocheng Feng, Errui Ding, and Jingdong Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[41] Yang Liu, Xiang Huang, Minghan Qin, Qinwei Lin, and Haoqian Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 348 + ], + "type": "text", + "content": "[42] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust Dynamic Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Spl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "text", + "content": "[45] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[46] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[47] Muyao Niu, Yifan Zhan, Qingtian Zhu, Zhuoxiao Li, Wei Wang, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Bundle adjusted gaussian avatars deblurring. arXiv preprint arXiv:2411.16758, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 658 + ], + "type": "text", + "content": "[48] Muyao Niu, Mingdeng Cao, Yifan Zhan, Qingtian Zhu, Mingze Ma, Jiancheng Zhao, Yanhong Zeng, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable Neural Radiance Fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865–5874, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13601" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[50] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiyup Kang. Temporal Interpolation Is All You Need for Dynamic Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 194 + ], + "type": "text", + "content": "[51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 248 + ], + "type": "text", + "content": "[52] Rui Peng, Wangze Xu, Luyang Tang, Jianbo Jiao, Ronggang Wang, et al. Structure consistent gaussian splatting with matching prior for few-shot novel view synthesis. Advances in Neural Information Processing Systems, 37:97328-97352, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "type": "text", + "content": "[53] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 7, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "type": "text", + "content": "[54] Sergey Prokudin, Qianli Ma, Maxime Raafat, Julien Valentin, and Siyu Tang. Dynamic point fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7964-7976, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[55] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-neRF: Neural Radiance Fields for Dynamic Scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[56] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-avatar: Animatable avatars via deformable 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "type": "text", + "content": "[57] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient Neural 4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "text", + "content": "[58] Shihe Shen, Huachen Gao, Wangze Xu, Rui Peng, Luyang Tang, Kaiqiang Xiong, Jianbo Jiao, and Ronggang Wang. Disentangled generation and aggregation for robust radiance fields. In European Conference on Computer Vision, pages 218-236. Springer, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "type": "text", + "content": "[59] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "text", + "content": "[60] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[61] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21096-21106, 2025. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[63] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9065-9076, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 183, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 554, + 228 + ], + "type": "text", + "content": "[64] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: Animatable volume rendering of articulated human sdfs. In European conference on computer vision, pages 1-19. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "type": "text", + "content": "[65] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 554, + 327 + ], + "type": "text", + "content": "[66] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "type": "text", + "content": "[67] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 437 + ], + "type": "text", + "content": "[68] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 491 + ], + "type": "text", + "content": "[69] Jiahao Wu, Rui Peng, Zhiyan Wang, Lu Xiao, Luyang Tang, Jinbo Yan, Kaiqiang Xiong, and Ronggang Wang. Swift4d: Adaptive divide-and-conquer gaussian splatting for compact and efficient reconstruction of dynamic scene. arXiv preprint arXiv:2503.12307, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "text", + "content": "[70] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time Neural Irradiance Fields for Free-viewpoint Video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 554, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 603 + ], + "type": "text", + "content": "[71] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In European Conference on Computer Vision, pages 203-220. Springer, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "text", + "content": "[72] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In CVPR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[73] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13602" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 518 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[74] Jinbo Yan, Rui Peng, Zhiyan Wang, Luyang Tang, Jiayu Yang, Jie Liang, Jiahao Wu, and Ronggang Wang. Instant gaussian stream: Fast and generalizable streaming of dynamic scene reconstruction via gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16520-16531, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[75] Yifan Zhan, Qingtian Zhu, Muyao Niu, Mingze Ma, Jiancheng Zhao, Zhihang Zhong, Xiao Sun, Yu Qiao, and Yinqiang Zheng. Tomie: Towards modular growth in enhanced smpl skeleton for 3d human with animatable garments. arXiv preprint arXiv:2410.08082, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[76] Yifan Zhan, Wangze Xu, Qingtian Zhu, Muyao Niu, Mingze Ma, Yifei Liu, Zhihang Zhong, Xiao Sun, and Yinqiang Zheng. R3-athlete: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 373 + ], + "type": "text", + "content": "[78] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 430 + ], + "type": "text", + "content": "[79] Yang Zheng, Qingqing Zhao, Guandao Yang, Wang Yifan, Donglai Xiang, Florian Dubost, Dmitry Lagun, Thabo Beeler, Federico Tombari, Leonidas Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 430, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 474 + ], + "type": "text", + "content": "[80] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European conference on computer vision, pages 145-163. Springer, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[81] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13603" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_content_list.json b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9845950840c55a16d7d0965f5973b717a9095f1b --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_content_list.json @@ -0,0 +1,1530 @@ +[ + { + "type": "text", + "text": "Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anja Delić, Matej Grcić, Siniša Šegvic \nUniversity of Zagreb, Faculty of Electrical Engineering and Computing Zagreb, Croatia", + "bbox": [ + 217, + 202, + 781, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "name.surname@fer.hr", + "bbox": [ + 410, + 260, + 581, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting anomalous human behaviour is an important visual task in safety-critical applications such as healthcare monitoring, workplace safety, or public surveillance. In these contexts, abnormalities are often reflected with unusual human poses. Thus, we propose SeeKer, a method for detecting anomalies in sequences of human skeletons. Our method formulates the skeleton sequence density through autoregressive factorization at the keypoint level. The corresponding conditional distributions represent probable keypoint locations given prior skeletal motion. We formulate the joint distribution of the considered skeleton as causal prediction of conditional Gaussians across its constituent keypoints. A skeleton is flagged as anomalous if its keypoint locations surprise our model (i.e. receive a low density). In practice, our anomaly score is a weighted sum of per-keypoint log-conditionals, where the weights account for the confidence of the underlying keypoint detector. Despite its conceptual simplicity, SeeKer surpasses all previous methods on the UBnormal and MSAD-HR datasets while delivering competitive performance on the ShanghaiTech dataset.", + "bbox": [ + 88, + 342, + 485, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 691, + 220, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection in videos is a safety-critical computer vision task that focuses on identifying events that deviate from regular patterns within observed video sequences [30, 43]. This task is particularly relevant for applications such as public institution monitoring, workplace safety, or crowd surveillance [34, 48], where abnormal frames may indicate emergencies such as strokes, accidents or acts of violence [3]. Designing more accurate human-related anomaly detectors may improve the security of shared spaces and contribute to overall public safety.", + "bbox": [ + 89, + 717, + 482, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Capturing normality in raw RGB video sequences is challenging due to high resolution, complex back", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2363106981f0e5e0c4bf12a0452b6d277a1f79c9dc53df88eb69814244e9ab5e.jpg", + "image_caption": [ + "Figure 1. Given preceding skeleton keypoints (colored dots), SeeKer estimates the multivariate normal distribution of probable locations for the subsequent keypoint. Observing keypoints at improbable locations indicates anomalous behaviour (e.g. fainting)." + ], + "image_footnote": [], + "bbox": [ + 519, + 308, + 903, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "grounds, and spatio-temporal correlations between consecutive frames [34]. Consequently, previous anomaly detection approaches turn to feature representations extracted with deep models [2, 55] and optical flow predictions [36] as well as to combinations of multiple modalities [42]. Although these approaches achieve varying degrees of success, video anomaly detection remains an open challenge. In the context of human-related anomalies, promising approaches rely solely on skeletal representations of human poses [16, 23]. These are particularly relevant as problematic human behaviour can often be associated with irregular body positions [3]. Moreover, skeleton sequences provide a low-dimensional and anonymous representation that reduces susceptibility to biases such as variations in clothing, body shape, or ethnicity and ensures that anomaly detection is based solely on motion dynamics. Previous approaches to skeleton-based anomaly detection [17, 39, 55] consider skeletons as graphs and train graph-oriented models [56] to maximize the likelihood of the skeleton sequence [23]. However, such monolithic approaches ignore the causality in skeleton sequences inherent in the temporal domain and overlook the compositional nature of skeletons. Moreover, anomaly scores of these approaches cannot account for the uncertainty of skeleton keypoint detectors.", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11579", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this manuscript, we propose SeeKer (Sequential Keypoint Density Estimator), a method for detecting anomalous human behaviour by predicting the conditional density of skeleton keypoints, as visualized in Figure 1. Given the preceding keypoints, SeeKer predicts a multivariate normal distribution over possible locations for the subsequent keypoint. Applying this approach across the entire skeleton sequence results in an autoregressive factorization of the joint sequence density at the keypoint level. Consequently, SeeKer training corresponds to maximizing sequence likelihood. During inference, SeeKer identifies skeletons as anomalous if their constituent keypoints deviate from the predicted distributions. The corresponding skeleton anomaly score is a sum of keypoint likelihoods weighted by the keypoint extraction confidence from an off-the-shelf skeleton extractor. Thus, our final decision accounts for uncertainties across all pipeline components. Overall, SeeKer presents the following key novelties to skeleton-based anomaly detection: autoregressive density factorization at the keypoint level, interpretable compositional anomaly scoring, and principled inclusion of keypoint detection uncertainty. Experimental evaluation shows that SeeKer achieves six percentage points absolute improvement over the best previous method on the UBnormal dataset and five percentage points absolute improvement on the MSAD-HR dataset in terms of AUROC.", + "bbox": [ + 89, + 90, + 483, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 497, + 228, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video anomaly detection. Video anomaly detection [30, 43] focuses on identifying anomalous frames within video sequences. This task is particularly relevant in surveillance [48] where anomalies may indicate accidents or emergencies [34]. Early approaches to frame-based video anomaly detection map input frames into a shared feature space [2, 14] and apply decision rules that differentiate between normal and abnormal patterns [25]. More recent approaches [8, 16, 19] focus on bounding box detections and recognize anomalies using self-supervised features [20, 53]. Other methods leverage failed optical flow predictions [36], errors in denoising of diffusion models [55], and predictions from cooperatively trained deep models [59]. In addition, there are exciting density estimation techniques such as denoising score matching [42]. However, these RGB-based methods may raise ethical concerns due to potential appearance-based biases and difficult anonymization. SeeKer builds atop skeleton sequences, and thus avoids such concerns.", + "bbox": [ + 89, + 522, + 483, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-based video anomaly detection. Video anomaly detection based on human skeleton sequences [11, 41] is particularly effective due to anonymity, low computational requirements, and domain invariance [23]. Existing skeleton-based methods capture spatio-temporal dependencies within skeleton sequences using recurrent models [11] or convolutional variational autoencoders [26, 27]. Alterna-", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tively, viewing skeletons as graphs [17, 23, 39, 41] allows for processing with graph-based architectures [56, 57]. The trained models then detect anomalies as failed pose predictions [16, 26] or unlikely sequences [23]. Unlike previous approaches, SeeKer leverages the compositional nature of skeletons and applies autoregressive sequence factorization at the keypoint level. This results in improved alignment, principled inclusion of keypoint detection uncertainty, and interpretable scoring in the sense that we can easily identify the keypoint(s) that triggered the anomaly detector.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human pose estimation. Human pose estimation [47, 49] aims to detect key body joints in an image and assemble them into skeletal structures. Recent pose estimators [15, 61] allow for accurate real-time detection of all skeletons in a given image, facilitating skeleton-based recognition. SeeKer builds on efficient pose estimators [15] to enable real-time anomaly detection. An overview of human pose estimation can be found in surveys [10, 29].", + "bbox": [ + 511, + 243, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive data likelihood factorization. Autoregressive factorization of data likelihood emerged as a foundational approach across numerous machine learning applications with a dominant temporal domain [5, 12]. Such factorization is conveniently modeled using sliding window based approaches [6, 52], recurrent neural networks [24, 46] or transformer decoders [9, 51] with causal self-attention layers [32, 50]. SeeKer relies on autoregressive factorization of skeleton sequence density.", + "bbox": [ + 511, + 366, + 905, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Skeleton anomaly detection by predicting probable locations of subsequent keypoint", + "text_level": 1, + "bbox": [ + 511, + 520, + 903, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem setup. Let $\\mathcal{D} = \\{\\mathbf{X}^i = (X_1^i, X_2^i, \\ldots, X_{T_i}^i)\\}_{i=1}^M$ be a dataset of $M$ skeleton sequences $\\mathbf{X}^i$ that track human activity throughout $T_i$ time steps. Each skeleton $X_t^i \\in \\mathbf{X}^i$ is extracted from the corresponding image and tracked across multiple frames with AlphaPose [15], an off-the-shelf detector and tracker. Each skeleton $X_t$ is an $N \\times D$ matrix with rows corresponding to skeleton joints and columns representing the spatial coordinates of the detected keypoint. In practice, $N = 18$ (number of keypoints) and $D = 2$ (coordinates of each keypoint). Thus, each sequence $\\mathbf{X}^i$ forms a third-order tensor of shape $T_i \\times N \\times D$ . Given the dataset $\\mathcal{D}$ of normal skeleton sequences, our objective is to develop an anomaly detector $\\tau: \\mathbf{X}^i \\to \\{0,1\\}_{T_i}^{T_i}$ capable of distinguishing between regular and anomalous skeletons in each time instant of a given sequence.", + "bbox": [ + 511, + 564, + 905, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core approach behind SeeKer is to predict a conditional distribution of probable keypoint locations based on the preceding keypoints in the skeleton sequence, as visualized in Figure 2. Leveraging this design in conjunction with the compositional nature of skeletons provides a highly interpretable skeleton anomaly score that aggregates likelihoods estimated at the keypoint level.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11580", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad66c28b84dfc99668a59007e82cb28a5b239dda5ae16615338160c78ba4aff7.jpg", + "image_caption": [ + "Figure 2. SeeKer training starts by flattening the target skeleton $X_{t}$ and the preceding skeletons $X_{ \\delta]]$ , where $[[\\cdot]]$ represents Iverson brackets. Note that the proposed anomaly score readily discloses perkeypoint contributions to the decision, which makes it much more interpretable than the current state of the art [23, 42].", + "bbox": [ + 511, + 419, + 905, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case when multiple humans are present in a scene $\\mathcal{I}_t$ , we simply determine the per-frame anomaly score as the maximum score among all skeletons within that frame. Consequently, our per-frame anomaly score $s_F$ equals:", + "bbox": [ + 511, + 541, + 905, + 601 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {F} \\left(\\mathcal {I} _ {t} \\mid \\mathcal {I} _ {< t}\\right) = \\max _ {X _ {t} \\in \\mathcal {I} _ {t}} s \\left(X _ {t} \\mid \\mathbf {X} _ {\\Delta}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 613, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we experimentally validate SeeKer utility on the standard human-related anomaly detection benchmarks.", + "bbox": [ + 511, + 648, + 905, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 693, + 702, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We consider three publicly available datasets for unsupervised video anomaly detection in skeleton sequences. UBnormal [3] is a synthetic dataset with 268 training, 64 validation, and 211 test video sequences. The inlier sequences contain actions such as walking while texting and talking to others. Anomalous actions include unexpected behaviour, such as stealing or fighting. This is the most relevant dataset for our task due to highly accurate labels. ShanghaiTech [34] consists of 330 training and 107 test videos of real-world scenes. Walking, standing, or sitting are considered as inlier actions, while cycling, skateboarding, and fighting are anomalous. Furthermore,", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11582", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "both datasets contain anomalous events that are not human-related (e.g. car accident or fire). Consequently, we also consider dataset subsets that contain only human-related anomalies as suggested in [16, 45] referred to as UBnormalHR and ShanghaiTech-HR. The MSAD [62] dataset includes videos captured across diverse locations and various viewpoints, ensuring a wide range of perspectives and conditions, making it the most challenging benchmark to date. MSAD contains 720 videos in total, some extending up to 6K frames, and features 55 distinct abnormal events. We focus on the human-related subset with 20 distinct abnormal actions, MSAD-HR, as the full dataset track includes scenarios and events related to industrial automation and manufacturing without any human involvement. Note that nonhuman related anomalies on UBnormal and ShanghaiTech still involve humans, making them detectable through human reactions to events like fire or unusual vehicle behavior. Previous work [23] noticed that contemporary skeleton extractors [15] and trackers [54] fail to produce satisfactory outputs on some datasets (UCFCrime [48] and Avenue [37]). Hence, we skip these datasets in our experiments.", + "bbox": [ + 89, + 90, + 480, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We report the standard anomaly detection metrics AUROC and AP, as well as human-related anomaly detection metrics RBDC and TBDC. Notably, we are the first to introduce the measurement of AP in this context. We compute AUROC and AP jointly over all test frames, which therefore corresponds to micro-averaging, e.g. micro-AUROC. We also report the region-based and track-based detection criteria (RBDC and TBDC) [44] to evaluate the spatial and temporal localization of detected anomalies on UBnormal and ShanghaiTech since they offer location annotations. RBDC assesses each detected region by calculating its Intersection-over-Union (IoU) with the corresponding ground truth region, labeling it as a true positive if the overlap exceeds a specified threshold $\\alpha$ . On the other hand, TBDC evaluates the accuracy of tracking abnormal regions over time, identifying a detected track as a true positive if the number of detections within the track exceeds a threshold $\\beta$ . Following the previous work [8], we set both $\\alpha$ and $\\beta$ at 0.1. We calculate the area under the ROC curve for both metrics, taking into account the tradeoff between detection rate and false-positive rate. For this evaluation, we rely on the ground truths provided by [19]. Some methods are not suitable for calculating RBDC and TBDC since they cannot pair features to their temporal or spatial locations in the original videos [42].", + "bbox": [ + 91, + 412, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We organize the considered baselines with respect to the modality they receive on the input. Methods that operate on deep features [22, 34] are denoted with (D). Methods that utilize optical flow velocities [36, 57] are denoted with (V), while skeletons-based approaches [23, 42] are denoted with (S). We briefly discuss the most relevant skeleton-based methods. STG-NF [23] proposes to estimate", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the likelihood of skeleton sequences with a graph-oriented normalizing flow with space and time separable graph convolutions in the coupling layers. MoCoDAD [16] proposes to model both normality and abnormality as multimodal by forecasting future skeletons with a diffusion model conditioned on the past. MULDE [42] is a recent method that leverages an energy-based model trained by a modification of denoising score matching that requires the model to have two times differentiable layers. For a fair comparison, we evaluate the AP of the most relevant baseline, STG-NF, with models provided by the authors.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our skeleton extractor procedure closely follows STG-NF [23]. We extract human skeletons with AlphaPose [15] within the YOLOX bounding boxes [18] and track skeletons with reID [60] across the video. The output skeletons consist of 18 keypoints in the COCO pose format [33]. We divide the extracted skeleton trajectories into time segments of $T = 24$ frames and normalize them to have zero mean and unit variance.", + "bbox": [ + 511, + 257, + 903, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We predict the per-keypoint distribution parameters $\\mu$ and $\\Sigma$ with deep autoregressive models that we configure for skeleton sequences. We consider causally masked fully connected models [21], transformers [51], and recurrent models. We train the causal fully connected model for 20 epochs on UBnormal and MSAD-HR and 10 epochs on ShanghaiTech with the Adam optimizer, learning rate $10^{-3}$ and batch size 256. Since UBnormal is the only dataset with a validation split, we validate all hyperparameters on the UBnormal validation subset and reuse them for ShanghaiTech and MSAD-HR. During inference, SeeKer considers $T$ preceding frames in a sliding-window manner. At the beginning of a sequence, we backtrack the score of the $T$ -th frame to the first $T - 1$ frames. Following [16, 23, 42], we finally apply Gaussian smoothing of the computed anomaly scores. However, we apply smoothing for each person separately before aggregating the per-frame anomaly score with $\\sigma = 10$ . All reported results are averaged over three runs, and variance is consistently below 0.05 pp. Our code is publicly available1.", + "bbox": [ + 511, + 378, + 903, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental results", + "text_level": 1, + "bbox": [ + 511, + 694, + 712, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Skeleton-based video anomaly detection", + "text_level": 1, + "bbox": [ + 511, + 719, + 854, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares SeeKer with previous methods on the UBnormal dataset. Our method substantially improves over all previous approaches in terms of AUROC, AP, and RBDC. Compared to the best previous skeleton-based method STGNF [23], SeeKer attains substantial increase of over 6 percentage points (pp) AUROC, 17 pp improvement in terms of AP, and 10 pp improvement in terms of RBDC, all while maintaining competitive TBDC. Moreover, SeeKer achieves over 15 pp AUROC improvement compared to", + "bbox": [ + 511, + 739, + 903, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/adelic99/seeker", + "bbox": [ + 531, + 886, + 807, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11583", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the leading method based on deep features Ssmtl++ [20]. SeeKer also outperforms the multimodal baseline MULDE [42], despite relying solely on skeletons. For human-related anomalies, SeeKer achieves over 7 pp gain in terms of AUROC and over 12 pp AP improvement over the best performing baseline STG-NF [23]. Thus, SeeKer is a new best model for anomaly detection on the UBnormal dataset.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/229cdeb8d94222c1ff881561f4833f9a9c8c357d6903ad6f75f09ab1a650db8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓BAF [19]59.3---21.953.4
✓Jigsaw [53]56.4---11.836.8
✓Sglm++ [20]62.1---25.663.5
✓ ✓FPDM[55]62.7-----
✓ ✓ ✓MULDE [42]72.8-----
BiPOCO [27]50.752.3----
GEPC [41]53.455.2----
MPED-RNN [11]60.661.2----
COSKAD [17]65.065.5----
MoCoDAD [16]68.368.4----
STG-NF [23]71.871.562.767.231.762.3
SeeKer77.978.980.379.842.058.5
", + "bbox": [ + 107, + 208, + 470, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 compares SeeKer performance with all previous approaches on the ShanghaiTech dataset. Among skeleton-based methods, our approach attains the highest AP and RBDC, while ranking second in AUROC and TBDC. Notably, the improvement upon the top-performing skeleton baseline STG-NF is over 9pp in terms of RBDC. In the case of human-related anomalies, SeeKer achieves the best AP and the second-best AUROC among skeleton-based approaches and continues to outperform methods that depend on deep features and optical flow velocities. Overall, SeeKer is a competitive model on ShanghaiTech dataset.", + "bbox": [ + 89, + 460, + 483, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 compares SeeKer with relevant skeleton-based methods on the MSAD-HR dataset [62]. As the first to report results on MSAD-HR, we reproduce the most relevant baselines [16, 23]. SeeKer surpasses these baselines by a substantial margin of $5.4\\mathrm{pp}$ in AUROC and more than 3 pp in terms of AP. The MSAD-HR dataset presents significant challenges due to crowded scenes, where frequent human trajectory intersections impact skeleton estimation and tracking accuracy. Consequently, the baseline methods only marginally exceed random guessing. Next, we qualitatively analyse SeeKer performance to gain deeper insights into the obtained quantitative results.", + "bbox": [ + 89, + 628, + 483, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Qualitative performance", + "text_level": 1, + "bbox": [ + 89, + 818, + 316, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SeeKer anomaly score (8) is highly interpretable, as it consolidates keypoint likelihoods into skeleton-level anomaly scores while operating in the original 2D space. Figure 5 shows the difference in predictions during nor", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2b32255e227ca69df7bdea50df8a1fb47c2306a5576c38002bbc4e920601ca75.jpg", + "table_caption": [ + "Table 1. Experimental evaluation of unsupervised anomaly detection on UBnormal [3]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + ], + "table_footnote": [], + "table_body": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓sRNN [38]68.0-----
✓Conv-AE [22]70.469.8----
✓LSA [2]72.5-----
✓ ✓FFP [35]72.872.7----
✓GCL [59]79.6-----
✓ ✓CAE-SVM [25]78.7-----
✓ ✓VEC [58]74.8-----
✓ ✓HF2 [36]76.2-----
✓ ✓FPDM[55]78.6-----
✓SSMTL [20]82.7-----
✓ ✓BA-AED [19]82.7---41.378.8
✓SSMTL++ [8]83.8---47.185.6
✓Jigsaw [53]84.284.7--22.460.8
✓ ✓ ✓MULDE [42]86.7---52.783.6
✓BiPOCO [27]-74.9----
✓MPED-RNN [11]73.475.4----
✓MTP [45]76.077.0----
✓GEPC [41]76.174.8----
✓PoseCVAE [26]-75.5----
✓Normal Graph [39]-76.5----
✓COSKAD [17]-77.1----
✓GCAE-LSTM [31]-77.2----
✓MoCoDAD [16]-77.6----
✓MULDE [42]78.5-----
✓STG-NF [23]85.987.477.681.452.182.4
✓SeeKer85.586.980.081.562.581.8
", + "bbox": [ + 526, + 87, + 895, + 441 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f915b2ec09db9105ea309c6ad2d681e1c54765c6fa80f0c8f3adc82b91d8a89.jpg", + "table_caption": [ + "Table 2. Experimental evaluation of anomaly detection on ShanghaiTech [38]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + ], + "table_footnote": [], + "table_body": "
ModalityMethodAUROCAP
DVS
✓53.951.7
✓55.756.5
✓61.160.1
", + "bbox": [ + 576, + 510, + 844, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Experimental evaluation of unsupervised anomaly detection on the MSAD-HR [62]. All baseline evaluations were reproduced by us.", + "bbox": [ + 511, + 583, + 906, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mal and abnormal behavior. The observations align closely with predictions of possible keypoint locations at normal timesteps. At abnormal timesteps, the predictions tend to follow normal motion patterns and deviate significantly from actual observations, resulting in high anomaly scores. In addition, Figure 6 emphasizes the interpretability of perkeypoint contributions. At each timestep, keypoints contribute with different intensity to the abnormality of the skeleton, as visualized through color variation. Notably, this level of analysis is not feasible with previous methods.", + "bbox": [ + 511, + 656, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 7 compares SeeKer and the best skeleton-oriented baseline STG-NF [23] on an anomalous sequence from ShanghaiTech (bike riding in the pedestrian area). We plot both the raw anomaly scores and the anomaly scores smoothed with a 1D Gaussian filter. The x- and y-axes represent time and anomaly score, respectively, and red-shaded", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11584", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4236ed622456be311c4aa6f0dd5d93e41bc096411534fa338b4577966a4f5f5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 90, + 478, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35e719f9c705375bd25caf53af45046140be39472e795b760ebfbc88d797ebce.jpg", + "image_caption": [ + "Figure 5. Qualitative performance of SeeKer on an example from UBnormal. During normal behavior, the observed keypoints emerge at anticipated locations. Abnormal behavior causes significant deviation from the expected keypoint locations, which enables seamless anomaly detection." + ], + "image_footnote": [], + "bbox": [ + 94, + 188, + 483, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d710bbae1c2333f81ba1734c704798359607c442a0ce755774211e99e9d54988.jpg", + "image_caption": [ + "Figure 6. Interpretability of SeeKer shown on an anomalous action from UBnormal (running). SeeKer anomaly score aggregates perkeypoint likelihoods denoted with varying color intensities. Each keypoint contribution to the final score changes over time." + ], + "image_footnote": [], + "bbox": [ + 94, + 484, + 483, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "regions mark frames labeled as anomalous. SeeKer accurately identifies the anomalous frames, while STG-NF fails to detect anomalies at the temporal borders of the abnormal action. Prior work [42] notes that ShanghaiTech contains noisy labels for skeleton-based anomaly detection. For instance, a bicycle wheel may appear before the rider, delaying detection by skeleton-based methods.", + "bbox": [ + 89, + 670, + 482, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Discussion", + "text_level": 1, + "bbox": [ + 89, + 786, + 209, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For all experiments in the discussion, we report AUROC on the UBnormal and ShanghaiTech test subsets.", + "bbox": [ + 89, + 809, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On predefined or learned covariance. SeeKer predicts parameters of multivariate normal distribution, $\\mu$ and $\\Sigma$ . We validate the importance of the learned covariance matrix compared to the predefined alternative. We consider prede", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91bc72656f2f16461d0781e124726c412c04c9175df21e34ad9e036bb1a31c06.jpg", + "image_caption": [ + "Figure 7. SeeKer accurately signals anomalous frames at the temporal border of the abnormal action, as shown by the blue curves, even without Gaussian smoothing (example from ShanghaiTech)." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 906, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fined covariance (identity matrix), learned diagonal covariance $(\\Sigma_{\\theta} = \\sigma^{2}I)$ , and learned full covariance. In the latter case, we model the inverse covariance matrix with Cholesky decomposition $\\Sigma_{\\theta}^{-1} = L_{\\theta}L_{\\theta}^{\\top}$ where $L_{\\theta}$ is a lower triangular matrix outputted by the model [13]. The decomposition ensures numerically stable full covariance implementation. Note that the case with the identity matrix as covariance makes the second term of objective (6) constant.", + "bbox": [ + 511, + 332, + 906, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 compares the performance of learned covariance, either full or diagonal, and baseline with the fixed covariance matrix. The obtained results indicate over $10\\%$ absolute improvement with respect to the fixed covariance, emphasising the importance of the learned covariance in the SeeKer framework.", + "bbox": [ + 511, + 453, + 905, + 542 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/55ec3c719d8499977b1e98758843ca7129cbef50de747c78ea9c1ba70a873530.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Covariance ΣUBnormalShanghaiTech
FullHRFullHR
identity61.463.474.174.4
full77.878.184.385.8
diagonal77.978.985.586.9
", + "bbox": [ + 527, + 553, + 890, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Learned covariance significantly outperforms the predefined identity matrix covariance.", + "bbox": [ + 511, + 627, + 905, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On prediction granularity. Autoregressive factorization can be spanned at different granularity levels, i.e. at the keypoint or skeleton level. Each of these options requires the adjustment in masking strategy, as detailed in the Appendix B. Table 5 validates the performance of each of these granularities with the same hyperparameters. We observe that modeling skeleton sequences at the keypoint level consistently outperforms its skeleton-level counterpart across the datasets. Interestingly, both granularities from Table 5 outperform the current state of the art on UBnormal. This outcome underscores the effectiveness of our autoregressive modelling of skeleton sequences.", + "bbox": [ + 511, + 674, + 906, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the impact of keypoint uncertainty. Our anomaly score weights conditional log-likelihoods of each keypoint with its keypoint extractor uncertainty from the skeleton", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11585", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d1cb88b34c181a1425ad4a2bde3122135d68b15a72e44db81169363585442a32.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GranularityUBnormalShanghaiTech
FullHRFullHR
Keypoint77.978.985.586.9
Skeleton75.076.879.082.1
", + "bbox": [ + 143, + 88, + 433, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "detector. Table 6 shows that incorporating the keypoint uncertainty into the final anomaly score substantially enhances anomaly detection performance across all datasets. These results emphasise that aggregating uncertainties of all pipeline components contribute to the end decision. Note that previous best methods [23, 42] cannot easily incorporate keypoint extraction uncertainty in their anomaly score. We also integrate the detector uncertainty in the training process as discussed in Appendix C.", + "bbox": [ + 89, + 191, + 483, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/be4bfaa3bf6f1a6a6b077b882192774b8bb3b64146dbfb79f3c14e0539728e22.jpg", + "table_caption": [ + "Table 5. Per-keypoint vs. per-skeleton prediction granularity." + ], + "table_footnote": [], + "table_body": "
Confidence weightingAnomaly scoreUBnormalShanghaiTech
FullHRFullHR
XEq. (7)76.877.783.785.2
✓Eq. (8)77.978.985.586.9
", + "bbox": [ + 130, + 340, + 444, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the heuristic smoothing. Post-process smoothing is a common technique for reducing outliers in time-sequence data. Previous works [16, 23, 42] apply Gaussian kernel smoothing in a sliding window to refine per-frame anomaly scores. This postprocessing strategy yields around three percentage points performance improvement for the STGNF baseline. SeeKer, on the other hand, accurately detects action boundaries by construction, as shown in Figure 7. Thus, smoothing contributes less to the end score. Furthermore, if we remove postprocessing from both methods, SeeKer consistently outperforms STG-NF on all datasets.", + "bbox": [ + 89, + 449, + 483, + 614 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b0f7382c287cabd46586e15445e4e79b3c9ebe3a7bbe74e9e34892b1185f7415.jpg", + "table_caption": [ + "Table 6. Integrating keypoint detection uncertainty into the anomaly score improves performance." + ], + "table_footnote": [], + "table_body": "
MethodSmoothingUBnormalShanghaiTech
FullHRFullHR
STG-NF✗68.268.083.084.4
SeeKer✗76.877.783.484.9
STG-NF✓71.873.985.987.4
SeeKer✓77.978.985.586.9
", + "bbox": [ + 133, + 627, + 442, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the keypoint ordering. SeeKer assumes a fixed ordering of skeleton keypoints. Ablations with different keypoint orderings incur a variance of less than 0.1, demonstrating SeeKer invariance to predefined keypoint order. Thus, we rely on the standard keypoint ordering [33] by default.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On autoregressive model architectures. We analyse the performance of SeeKer depending on the specific model architecture used to predict parameters $\\mu$ and $\\Sigma$ . In addition", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the causally masked fully-connected model, we also test transformer architecture [51] that use causal self-attention layers. Each transformer token is a 2D skeleton keypoint location projected into a higher dimensional space. Further, we encode timestep and keypoint IDs to positional encoding of each token, as detailed in Appendix B.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8 shows that the masked fully connected model significantly outperforms its transformer-based counterpart. We attribute this gap to the low variability in skeleton sequences, as each person's gait differs only subtly. Additionally, skeleton sequences are much shorter than text paragraphs, and keypoint tokens have a much lower dimensionality, which influences the nature of the task and applicability of simpler models.", + "bbox": [ + 511, + 183, + 906, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6fa04c77717c4a51369905d54bfcdcf72439db4eb5087290e1f78e54397f7426.jpg", + "table_caption": [ + "Table 7. Post-process smoothing only slightly boosts the performance of SeeKer. Importantly, SeeKer without smoothing outperforms STG-NF with smoothing." + ], + "table_footnote": [], + "table_body": "
ArchitectureUBnormalShanghaiTech
FullHRFullHR
Transformer decoder [51]71.171.679.180.0
Masked fully connected [21]77.978.985.586.9
", + "bbox": [ + 521, + 320, + 898, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparison of the proposed masked fully-connected model and the causal transformer decoder.", + "bbox": [ + 511, + 382, + 905, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 450, + 633, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SeeKer, a detector for anomalous human behaviour based on generative modeling of skeletal sequences. Unlike previous approaches of this kind, SeeKer considers each skeleton as a collection of keypoints and models sequence density as a product of Gaussian conditional keypoint densities. The parameters of the conditionals $(\\mu, \\Sigma)$ are regressed by a causal deep model that ensures that each prediction can only observe the locations of the precedent keypoints. The proposed architecture infers all per-keypoint distributions with a single forward pass. We implement the corresponding causal architecture with masked fully connected layers. Interestingly, this formulation outperforms its counterpart based on a transformer decoder, probably due to the specifics of our data. During inference, we evaluate the per-keypoint distributions of the considered skeleton given the preceding keypoints in a sliding window, and summarize them into a skeleton-level anomaly score. Here, we take advantage of the composite structure of the proposed anomaly score by weighting the per-keypoint log-densities with the corresponding confidence values of the keypoint detector. The resulting anomaly score identifies abnormal behaviour by assessing the surprise of our model. Furthermore, our compositional anomaly score is highly interpretable since we can track keypoint contributions to the final decision. The conducted experiments indicate that SeeKer surpasses the state of the art on UBnormal and MSAD-HR by a wide margin, while delivering a close runner-up performance on ShanghaiTech.", + "bbox": [ + 511, + 477, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11586", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work has been supported by Croatian Recovery and Resilience Fund - NextGenerationEU (grant C1.4 R5-I2.01.0001). The authors declare no competing interests.", + "bbox": [ + 89, + 113, + 485, + 157 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 184, + 187, + 200 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reprint of: Mahalanobis, p.c. (1936) “on the generalised distance in statistics.” Sankhya Ser. A, 2018. 4", + "[2] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1, 2, 6", + "[3] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Abnormal: New benchmark for supervised open-set video anomaly detection. In Computer Vision and Pattern Recognition, 2022. 1, 4, 6, 13, 14", + "[4] Yoshua Bengio and Samy Bengio. Modeling high-dimensional discrete data with multi-layer neural networks. In Advances in Neural Information Processing Systems. MIT Press, 1999. 12", + "[5] Maria A. Berezina, Daniel Rudoy, and Patrick J. Wolfe. Autoregressive modeling of voiced speech. In IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP, pages 5042-5045, 2010. 2", + "[6] Vladimir Braverman. Sliding Window Algorithms. 2015. 2", + "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3, 13", + "[8] Antonio Bärbalău, Radu Tudor Ionescu, Mariana-Juliana Georgescu, Jacob Velling Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas Baltzer Moeslund, and Mubarak Shah. Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Comput. Vis. Image Underst., page 103656, 2022. 2, 5, 6", + "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. 2", + "[10] Yucheng Chen, Yingli Tian, and Mingyi He. Monocular human pose estimation: A survey of deep learning-based methods. Comput. Vis. Image Underst., 192:102897, 2020. 2" + ], + "bbox": [ + 93, + 209, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2, 6", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, 2019. 2", + "[13] Garoe Dorta, Sara Vicente, Lourdes Agapito, Neill D. F. Campbell, and Ivor Simpson. Structured uncertainty prediction networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 7", + "[14] Fernando Pereira dos Santos, Leonardo Sampaio Ferraz Ribeiro, and Moacir A. Ponti. Generalization of feature embeddings transferred from different video anomaly detection domains. J. Vis. Commun. Image Represent., 2019. 2", + "[15] Haoshu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 2, 4, 5, 15", + "[16] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 5, 6, 8, 14", + "[17] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Marco Aurelio Sterpa, Alessio Sampieri, and Fabio Galasso. Contracting skeletal kinematics for human-related video anomaly detection. Pattern Recognit., 2024. 1, 2, 6", + "[18] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. ArXiv, 2021. 5", + "[19] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 2, 5, 6", + "[20] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Claudi Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6", + "[21] Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for distribution estimation. In International Conference on Machine Learning, 2015. 3, 4, 5, 8, 12", + "[22] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11587", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 733-742. IEEE Computer Society, 2016. 5, 6", + "[23] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In International Conference on Computer Vision, 2023. 1, 2, 4, 5, 6, 8", + "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 1997. 2, 4", + "[25] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 2, 6", + "[26] Yashswi Jain, Ashvini Kumar Sharma, Rajbabu Velmurugan, and Biplab Banerjee. Posecvae: Anomalous human activity detection. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, 2020. 2, 6", + "[27] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. Bipoco: Bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection. CoRR, abs/2207.02281, 2022. 2, 6", + "[28] Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances in Neural Information Processing Systems, 2016. 4", + "[29] Gongjin Lan, Yu Wu, Fei Hu, and Qi Hao. Vision-based human pose estimation via deep learning: A survey. IEEE Trans. Hum. Mach. Syst., 2023. 2", + "[30] Roberto Leyva, Victor Sanchez, and Chang-Tsun Li. Video anomaly detection with compact feature sets for online performance. IEEE Trans. Image Process., 2017. 1, 2", + "[31] Nanjun Li, Faliang Chang, and Chunsheng Liu. Human-related anomalous event detection via spatial-temporal graph convolutional autoencoder with embedded long short-term memory network. Neurocomputing, 2021. 6", + "[32] Yingcong Li, Yixiao Huang, Muhammed Emrullah Ildiz, Ankit Singh Rawat, and Samet Oymak. Mechanics of next token prediction with self-attention. In International Conference on Artificial Intelligence and Statistics, 2024. 2", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5, 8", + "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 14", + "[35] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6536-6545. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6", + "[37] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 5, 15", + "[38] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In International Conference on Computer Vision, 2017. 6", + "[39] Weixin Luo, Wen Liu, and Shenghua Gao. Normal graph: Spatial temporal graph convolutional networks based prediction network for skeleton based video anomaly detection. Neurocomputing, 2020. 1, 2, 6", + "[40] Eran Malach. Auto-regressive next-token predictors are universal learners. In International Conference on Machine Learning, 2024. 3", + "[41] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 2, 6", + "[42] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: multiscale log-density estimation via denoising score matching for video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 2024. 1, 2, 4, 5, 6, 7, 8", + "[43] Xuan Mo, Vishal Monga, and Raja Bala. Simultaneous sparsity model for multi-perspective video anomaly detection. In International Conference on Image Processing, 2014. 1, 2", + "[44] Bharathkumar Ramachandra and Michael J. Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5", + "[45] Royston Rodrigues, Neha Bhargava, Rajbabu Velmurugan, and Subhasis Chaudhuri. Multi-timescale trajectory prediction for abnormal human activity detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5, 6, 14", + "[46] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 1986. 2", + "[47] Benjamin Sapp, Alexander Toshev, and Ben Taskar. Cascaded models for articulated pose estimation. In European Conference on Computer Vision, 2010. 2", + "[48] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 15", + "[49] Min Sun, Pushmeet Kohli, and Jamie Shotton. Conditional regression forests for human pose estimation. In Computer Vision and Pattern Recognition, 2012. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11588", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Davoud Ataee Tarzanagh, Yingcong Li, Xuechen Zhang, and Samet Oymak. Max-margin token selection in attention mechanism. In Neural Information Processing Systems, 2023. 2", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 2, 4, 5, 8, 13", + "[52] Alexander Waibel, Toshiyuki Hanazawa, Geoffrey E. Hinton, Kiyohiro Shikano, and Kevin J. Lang. Phoneme recognition using time-delay neural networks. IEEE Trans. Acoust. Speech Signal Process., 1989. 2", + "[53] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, 2022. 2, 6", + "[54] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. In British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018, 2018. 5, 15", + "[55] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature prediction diffusion model for video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 6", + "[56] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, 2018. 1, 2", + "[57] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, July 13-19, 2018, Stockholm, Sweden, 2018. 2, 5", + "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In MM '20: The 28th ACM International Conference on Multimedia, Virtual Event / Seattle, WA, USA, October 12-16, 2020. ACM, 2020. 6", + "[59] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14724-14734. IEEE, 2022. 2, 6", + "[60] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In 2019 IEEE/CVF International Conference" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, 2019. 5", + "[61] Leyan Zhu, Zitian Wang, Si Liu, Xuecheng Nie, Luoqi Liu, and Bo Li. Multi-person pose regression with distribution-aware single-stage models. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2", + "[62] Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise review and a new dataset. Advances in Neural Information Processing Systems, 37:89943-89977, 2025. 5, 6, 14" + ], + "bbox": [ + 516, + 92, + 903, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11589", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_model.json b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e42dff458d85cbad785b3604176f0d15d6e178e3 --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.782, + 0.257 + ], + "angle": 0, + "content": "Anja Delić, Matej Grcić, Siniša Šegvic \nUniversity of Zagreb, Faculty of Electrical Engineering and Computing Zagreb, Croatia" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.261, + 0.583, + 0.273 + ], + "angle": 0, + "content": "name.surname@fer.hr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.328, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.486, + 0.66 + ], + "angle": 0, + "content": "Detecting anomalous human behaviour is an important visual task in safety-critical applications such as healthcare monitoring, workplace safety, or public surveillance. In these contexts, abnormalities are often reflected with unusual human poses. Thus, we propose SeeKer, a method for detecting anomalies in sequences of human skeletons. Our method formulates the skeleton sequence density through autoregressive factorization at the keypoint level. The corresponding conditional distributions represent probable keypoint locations given prior skeletal motion. We formulate the joint distribution of the considered skeleton as causal prediction of conditional Gaussians across its constituent keypoints. A skeleton is flagged as anomalous if its keypoint locations surprise our model (i.e. receive a low density). In practice, our anomaly score is a weighted sum of per-keypoint log-conditionals, where the weights account for the confidence of the underlying keypoint detector. Despite its conceptual simplicity, SeeKer surpasses all previous methods on the UBnormal and MSAD-HR datasets while delivering competitive performance on the ShanghaiTech dataset." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.692, + 0.222, + 0.708 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Anomaly detection in videos is a safety-critical computer vision task that focuses on identifying events that deviate from regular patterns within observed video sequences [30, 43]. This task is particularly relevant for applications such as public institution monitoring, workplace safety, or crowd surveillance [34, 48], where abnormal frames may indicate emergencies such as strokes, accidents or acts of violence [3]. Designing more accurate human-related anomaly detectors may improve the security of shared spaces and contribute to overall public safety." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Capturing normality in raw RGB video sequences is challenging due to high resolution, complex back" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.309, + 0.905, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.456, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Figure 1. Given preceding skeleton keypoints (colored dots), SeeKer estimates the multivariate normal distribution of probable locations for the subsequent keypoint. Observing keypoints at improbable locations indicates anomalous behaviour (e.g. fainting)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "grounds, and spatio-temporal correlations between consecutive frames [34]. Consequently, previous anomaly detection approaches turn to feature representations extracted with deep models [2, 55] and optical flow predictions [36] as well as to combinations of multiple modalities [42]. Although these approaches achieve varying degrees of success, video anomaly detection remains an open challenge. In the context of human-related anomalies, promising approaches rely solely on skeletal representations of human poses [16, 23]. These are particularly relevant as problematic human behaviour can often be associated with irregular body positions [3]. Moreover, skeleton sequences provide a low-dimensional and anonymous representation that reduces susceptibility to biases such as variations in clothing, body shape, or ethnicity and ensures that anomaly detection is based solely on motion dynamics. Previous approaches to skeleton-based anomaly detection [17, 39, 55] consider skeletons as graphs and train graph-oriented models [56] to maximize the likelihood of the skeleton sequence [23]. However, such monolithic approaches ignore the causality in skeleton sequences inherent in the temporal domain and overlook the compositional nature of skeletons. Moreover, anomaly scores of these approaches cannot account for the uncertainty of skeleton keypoint detectors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.485 + ], + "angle": 0, + "content": "In this manuscript, we propose SeeKer (Sequential Keypoint Density Estimator), a method for detecting anomalous human behaviour by predicting the conditional density of skeleton keypoints, as visualized in Figure 1. Given the preceding keypoints, SeeKer predicts a multivariate normal distribution over possible locations for the subsequent keypoint. Applying this approach across the entire skeleton sequence results in an autoregressive factorization of the joint sequence density at the keypoint level. Consequently, SeeKer training corresponds to maximizing sequence likelihood. During inference, SeeKer identifies skeletons as anomalous if their constituent keypoints deviate from the predicted distributions. The corresponding skeleton anomaly score is a sum of keypoint likelihoods weighted by the keypoint extraction confidence from an off-the-shelf skeleton extractor. Thus, our final decision accounts for uncertainties across all pipeline components. Overall, SeeKer presents the following key novelties to skeleton-based anomaly detection: autoregressive density factorization at the keypoint level, interpretable compositional anomaly scoring, and principled inclusion of keypoint detection uncertainty. Experimental evaluation shows that SeeKer achieves six percentage points absolute improvement over the best previous method on the UBnormal dataset and five percentage points absolute improvement on the MSAD-HR dataset in terms of AUROC." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.229, + 0.513 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Video anomaly detection. Video anomaly detection [30, 43] focuses on identifying anomalous frames within video sequences. This task is particularly relevant in surveillance [48] where anomalies may indicate accidents or emergencies [34]. Early approaches to frame-based video anomaly detection map input frames into a shared feature space [2, 14] and apply decision rules that differentiate between normal and abnormal patterns [25]. More recent approaches [8, 16, 19] focus on bounding box detections and recognize anomalies using self-supervised features [20, 53]. Other methods leverage failed optical flow predictions [36], errors in denoising of diffusion models [55], and predictions from cooperatively trained deep models [59]. In addition, there are exciting density estimation techniques such as denoising score matching [42]. However, these RGB-based methods may raise ethical concerns due to potential appearance-based biases and difficult anonymization. SeeKer builds atop skeleton sequences, and thus avoids such concerns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Skeleton-based video anomaly detection. Video anomaly detection based on human skeleton sequences [11, 41] is particularly effective due to anonymity, low computational requirements, and domain invariance [23]. Existing skeleton-based methods capture spatio-temporal dependencies within skeleton sequences using recurrent models [11] or convolutional variational autoencoders [26, 27]. Alterna-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "tively, viewing skeletons as graphs [17, 23, 39, 41] allows for processing with graph-based architectures [56, 57]. The trained models then detect anomalies as failed pose predictions [16, 26] or unlikely sequences [23]. Unlike previous approaches, SeeKer leverages the compositional nature of skeletons and applies autoregressive sequence factorization at the keypoint level. This results in improved alignment, principled inclusion of keypoint detection uncertainty, and interpretable scoring in the sense that we can easily identify the keypoint(s) that triggered the anomaly detector." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.366 + ], + "angle": 0, + "content": "Human pose estimation. Human pose estimation [47, 49] aims to detect key body joints in an image and assemble them into skeletal structures. Recent pose estimators [15, 61] allow for accurate real-time detection of all skeletons in a given image, facilitating skeleton-based recognition. SeeKer builds on efficient pose estimators [15] to enable real-time anomaly detection. An overview of human pose estimation can be found in surveys [10, 29]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.906, + 0.503 + ], + "angle": 0, + "content": "Autoregressive data likelihood factorization. Autoregressive factorization of data likelihood emerged as a foundational approach across numerous machine learning applications with a dominant temporal domain [5, 12]. Such factorization is conveniently modeled using sliding window based approaches [6, 52], recurrent neural networks [24, 46] or transformer decoders [9, 51] with causal self-attention layers [32, 50]. SeeKer relies on autoregressive factorization of skeleton sequence density." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.521, + 0.905, + 0.557 + ], + "angle": 0, + "content": "3. Skeleton anomaly detection by predicting probable locations of subsequent keypoint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Problem setup. Let \\(\\mathcal{D} = \\{\\mathbf{X}^i = (X_1^i, X_2^i, \\ldots, X_{T_i}^i)\\}_{i=1}^M\\) be a dataset of \\(M\\) skeleton sequences \\(\\mathbf{X}^i\\) that track human activity throughout \\(T_i\\) time steps. Each skeleton \\(X_t^i \\in \\mathbf{X}^i\\) is extracted from the corresponding image and tracked across multiple frames with AlphaPose [15], an off-the-shelf detector and tracker. Each skeleton \\(X_t\\) is an \\(N \\times D\\) matrix with rows corresponding to skeleton joints and columns representing the spatial coordinates of the detected keypoint. In practice, \\(N = 18\\) (number of keypoints) and \\(D = 2\\) (coordinates of each keypoint). Thus, each sequence \\(\\mathbf{X}^i\\) forms a third-order tensor of shape \\(T_i \\times N \\times D\\). Given the dataset \\(\\mathcal{D}\\) of normal skeleton sequences, our objective is to develop an anomaly detector \\(\\tau: \\mathbf{X}^i \\to \\{0,1\\}_{T_i}^{T_i}\\) capable of distinguishing between regular and anomalous skeletons in each time instant of a given sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The core approach behind SeeKer is to predict a conditional distribution of probable keypoint locations based on the preceding keypoints in the skeleton sequence, as visualized in Figure 2. Leveraging this design in conjunction with the compositional nature of skeletons provides a highly interpretable skeleton anomaly score that aggregates likelihoods estimated at the keypoint level." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.281, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. SeeKer training starts by flattening the target skeleton \\( X_{t} \\) and the preceding skeletons \\( X_{ \\delta]] \\), where \\( [[\\cdot]] \\) represents Iverson brackets. Note that the proposed anomaly score readily discloses perkeypoint contributions to the decision, which makes it much more interpretable than the current state of the art [23, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.906, + 0.602 + ], + "angle": 0, + "content": "In the case when multiple humans are present in a scene \\(\\mathcal{I}_t\\), we simply determine the per-frame anomaly score as the maximum score among all skeletons within that frame. Consequently, our per-frame anomaly score \\(s_F\\) equals:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.614, + 0.906, + 0.639 + ], + "angle": 0, + "content": "\\[\ns _ {F} \\left(\\mathcal {I} _ {t} \\mid \\mathcal {I} _ {< t}\\right) = \\max _ {X _ {t} \\in \\mathcal {I} _ {t}} s \\left(X _ {t} \\mid \\mathbf {X} _ {\\Delta}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.906, + 0.679 + ], + "angle": 0, + "content": "Next, we experimentally validate SeeKer utility on the standard human-related anomaly detection benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.704, + 0.712 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We consider three publicly available datasets for unsupervised video anomaly detection in skeleton sequences. UBnormal [3] is a synthetic dataset with 268 training, 64 validation, and 211 test video sequences. The inlier sequences contain actions such as walking while texting and talking to others. Anomalous actions include unexpected behaviour, such as stealing or fighting. This is the most relevant dataset for our task due to highly accurate labels. ShanghaiTech [34] consists of 330 training and 107 test videos of real-world scenes. Walking, standing, or sitting are considered as inlier actions, while cycling, skateboarding, and fighting are anomalous. Furthermore," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.41 + ], + "angle": 0, + "content": "both datasets contain anomalous events that are not human-related (e.g. car accident or fire). Consequently, we also consider dataset subsets that contain only human-related anomalies as suggested in [16, 45] referred to as UBnormalHR and ShanghaiTech-HR. The MSAD [62] dataset includes videos captured across diverse locations and various viewpoints, ensuring a wide range of perspectives and conditions, making it the most challenging benchmark to date. MSAD contains 720 videos in total, some extending up to 6K frames, and features 55 distinct abnormal events. We focus on the human-related subset with 20 distinct abnormal actions, MSAD-HR, as the full dataset track includes scenarios and events related to industrial automation and manufacturing without any human involvement. Note that nonhuman related anomalies on UBnormal and ShanghaiTech still involve humans, making them detectable through human reactions to events like fire or unusual vehicle behavior. Previous work [23] noticed that contemporary skeleton extractors [15] and trackers [54] fail to produce satisfactory outputs on some datasets (UCFCrime [48] and Avenue [37]). Hence, we skip these datasets in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Metrics. We report the standard anomaly detection metrics AUROC and AP, as well as human-related anomaly detection metrics RBDC and TBDC. Notably, we are the first to introduce the measurement of AP in this context. We compute AUROC and AP jointly over all test frames, which therefore corresponds to micro-averaging, e.g. micro-AUROC. We also report the region-based and track-based detection criteria (RBDC and TBDC) [44] to evaluate the spatial and temporal localization of detected anomalies on UBnormal and ShanghaiTech since they offer location annotations. RBDC assesses each detected region by calculating its Intersection-over-Union (IoU) with the corresponding ground truth region, labeling it as a true positive if the overlap exceeds a specified threshold \\(\\alpha\\). On the other hand, TBDC evaluates the accuracy of tracking abnormal regions over time, identifying a detected track as a true positive if the number of detections within the track exceeds a threshold \\(\\beta\\). Following the previous work [8], we set both \\(\\alpha\\) and \\(\\beta\\) at 0.1. We calculate the area under the ROC curve for both metrics, taking into account the tradeoff between detection rate and false-positive rate. For this evaluation, we rely on the ground truths provided by [19]. Some methods are not suitable for calculating RBDC and TBDC since they cannot pair features to their temporal or spatial locations in the original videos [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Baselines. We organize the considered baselines with respect to the modality they receive on the input. Methods that operate on deep features [22, 34] are denoted with (D). Methods that utilize optical flow velocities [36, 57] are denoted with (V), while skeletons-based approaches [23, 42] are denoted with (S). We briefly discuss the most relevant skeleton-based methods. STG-NF [23] proposes to estimate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "the likelihood of skeleton sequences with a graph-oriented normalizing flow with space and time separable graph convolutions in the coupling layers. MoCoDAD [16] proposes to model both normality and abnormality as multimodal by forecasting future skeletons with a diffusion model conditioned on the past. MULDE [42] is a recent method that leverages an energy-based model trained by a modification of denoising score matching that requires the model to have two times differentiable layers. For a fair comparison, we evaluate the AP of the most relevant baseline, STG-NF, with models provided by the authors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Implementation details. Our skeleton extractor procedure closely follows STG-NF [23]. We extract human skeletons with AlphaPose [15] within the YOLOX bounding boxes [18] and track skeletons with reID [60] across the video. The output skeletons consist of 18 keypoints in the COCO pose format [33]. We divide the extracted skeleton trajectories into time segments of \\( T = 24 \\) frames and normalize them to have zero mean and unit variance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.68 + ], + "angle": 0, + "content": "We predict the per-keypoint distribution parameters \\(\\mu\\) and \\(\\Sigma\\) with deep autoregressive models that we configure for skeleton sequences. We consider causally masked fully connected models [21], transformers [51], and recurrent models. We train the causal fully connected model for 20 epochs on UBnormal and MSAD-HR and 10 epochs on ShanghaiTech with the Adam optimizer, learning rate \\(10^{-3}\\) and batch size 256. Since UBnormal is the only dataset with a validation split, we validate all hyperparameters on the UBnormal validation subset and reuse them for ShanghaiTech and MSAD-HR. During inference, SeeKer considers \\(T\\) preceding frames in a sliding-window manner. At the beginning of a sequence, we backtrack the score of the \\(T\\)-th frame to the first \\(T - 1\\) frames. Following [16, 23, 42], we finally apply Gaussian smoothing of the computed anomaly scores. However, we apply smoothing for each person separately before aggregating the per-frame anomaly score with \\(\\sigma = 10\\). All reported results are averaged over three runs, and variance is consistently below 0.05 pp. Our code is publicly available1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.714, + 0.712 + ], + "angle": 0, + "content": "5. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.72, + 0.855, + 0.735 + ], + "angle": 0, + "content": "5.1. Skeleton-based video anomaly detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.905, + 0.879 + ], + "angle": 0, + "content": "Table 1 compares SeeKer with previous methods on the UBnormal dataset. Our method substantially improves over all previous approaches in terms of AUROC, AP, and RBDC. Compared to the best previous skeleton-based method STGNF [23], SeeKer attains substantial increase of over 6 percentage points (pp) AUROC, 17 pp improvement in terms of AP, and 10 pp improvement in terms of RBDC, all while maintaining competitive TBDC. Moreover, SeeKer achieves over 15 pp AUROC improvement compared to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.808, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/adelic99/seeker" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "the leading method based on deep features Ssmtl++ [20]. SeeKer also outperforms the multimodal baseline MULDE [42], despite relying solely on skeletons. For human-related anomalies, SeeKer achieves over 7 pp gain in terms of AUROC and over 12 pp AP improvement over the best performing baseline STG-NF [23]. Thus, SeeKer is a new best model for anomaly detection on the UBnormal dataset." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.209, + 0.472, + 0.395 + ], + "angle": 0, + "content": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓BAF [19]59.3---21.953.4
✓Jigsaw [53]56.4---11.836.8
✓Sglm++ [20]62.1---25.663.5
✓ ✓FPDM[55]62.7-----
✓ ✓ ✓MULDE [42]72.8-----
BiPOCO [27]50.752.3----
GEPC [41]53.455.2----
MPED-RNN [11]60.661.2----
COSKAD [17]65.065.5----
MoCoDAD [16]68.368.4----
STG-NF [23]71.871.562.767.231.762.3
SeeKer77.978.980.379.842.058.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Table 1. Experimental evaluation of unsupervised anomaly detection on UBnormal [3]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.628 + ], + "angle": 0, + "content": "Table 2 compares SeeKer performance with all previous approaches on the ShanghaiTech dataset. Among skeleton-based methods, our approach attains the highest AP and RBDC, while ranking second in AUROC and TBDC. Notably, the improvement upon the top-performing skeleton baseline STG-NF is over 9pp in terms of RBDC. In the case of human-related anomalies, SeeKer achieves the best AP and the second-best AUROC among skeleton-based approaches and continues to outperform methods that depend on deep features and optical flow velocities. Overall, SeeKer is a competitive model on ShanghaiTech dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Table 3 compares SeeKer with relevant skeleton-based methods on the MSAD-HR dataset [62]. As the first to report results on MSAD-HR, we reproduce the most relevant baselines [16, 23]. SeeKer surpasses these baselines by a substantial margin of \\(5.4\\mathrm{pp}\\) in AUROC and more than 3 pp in terms of AP. The MSAD-HR dataset presents significant challenges due to crowded scenes, where frequent human trajectory intersections impact skeleton estimation and tracking accuracy. Consequently, the baseline methods only marginally exceed random guessing. Next, we qualitatively analyse SeeKer performance to gain deeper insights into the obtained quantitative results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.317, + 0.835 + ], + "angle": 0, + "content": "5.2. Qualitative performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The SeeKer anomaly score (8) is highly interpretable, as it consolidates keypoint likelihoods into skeleton-level anomaly scores while operating in the original 2D space. Figure 5 shows the difference in predictions during nor" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.088, + 0.897, + 0.442 + ], + "angle": 0, + "content": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓sRNN [38]68.0-----
✓Conv-AE [22]70.469.8----
✓LSA [2]72.5-----
✓ ✓FFP [35]72.872.7----
✓GCL [59]79.6-----
✓ ✓CAE-SVM [25]78.7-----
✓ ✓VEC [58]74.8-----
✓ ✓HF2 [36]76.2-----
✓ ✓FPDM[55]78.6-----
✓SSMTL [20]82.7-----
✓ ✓BA-AED [19]82.7---41.378.8
✓SSMTL++ [8]83.8---47.185.6
✓Jigsaw [53]84.284.7--22.460.8
✓ ✓ ✓MULDE [42]86.7---52.783.6
✓BiPOCO [27]-74.9----
✓MPED-RNN [11]73.475.4----
✓MTP [45]76.077.0----
✓GEPC [41]76.174.8----
✓PoseCVAE [26]-75.5----
✓Normal Graph [39]-76.5----
✓COSKAD [17]-77.1----
✓GCAE-LSTM [31]-77.2----
✓MoCoDAD [16]-77.6----
✓MULDE [42]78.5-----
✓STG-NF [23]85.987.477.681.452.182.4
✓SeeKer85.586.980.081.562.581.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Table 2. Experimental evaluation of anomaly detection on ShanghaiTech [38]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.511, + 0.846, + 0.574 + ], + "angle": 0, + "content": "
ModalityMethodAUROCAP
DVS
✓53.951.7
✓55.756.5
✓61.160.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.584, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Table 3. Experimental evaluation of unsupervised anomaly detection on the MSAD-HR [62]. All baseline evaluations were reproduced by us." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.809 + ], + "angle": 0, + "content": "mal and abnormal behavior. The observations align closely with predictions of possible keypoint locations at normal timesteps. At abnormal timesteps, the predictions tend to follow normal motion patterns and deviate significantly from actual observations, resulting in high anomaly scores. In addition, Figure 6 emphasizes the interpretability of perkeypoint contributions. At each timestep, keypoints contribute with different intensity to the abnormality of the skeleton, as visualized through color variation. Notably, this level of analysis is not feasible with previous methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Figure 7 compares SeeKer and the best skeleton-oriented baseline STG-NF [23] on an anomalous sequence from ShanghaiTech (bike riding in the pedestrian area). We plot both the raw anomaly scores and the anomaly scores smoothed with a 1D Gaussian filter. The x- and y-axes represent time and anomaly score, respectively, and red-shaded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.48, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.189, + 0.484, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.471 + ], + "angle": 0, + "content": "Figure 5. Qualitative performance of SeeKer on an example from UBnormal. During normal behavior, the observed keypoints emerge at anticipated locations. Abnormal behavior causes significant deviation from the expected keypoint locations, which enables seamless anomaly detection." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.486, + 0.484, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Figure 6. Interpretability of SeeKer shown on an anomalous action from UBnormal (running). SeeKer anomaly score aggregates perkeypoint likelihoods denoted with varying color intensities. Each keypoint contribution to the final score changes over time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.779 + ], + "angle": 0, + "content": "regions mark frames labeled as anomalous. SeeKer accurately identifies the anomalous frames, while STG-NF fails to detect anomalies at the temporal borders of the abnormal action. Prior work [42] notes that ShanghaiTech contains noisy labels for skeleton-based anomaly detection. For instance, a bicycle wheel may appear before the rider, delaying detection by skeleton-based methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "5.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.84 + ], + "angle": 0, + "content": "For all experiments in the discussion, we report AUROC on the UBnormal and ShanghaiTech test subsets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "On predefined or learned covariance. SeeKer predicts parameters of multivariate normal distribution, \\(\\mu\\) and \\(\\Sigma\\). We validate the importance of the learned covariance matrix compared to the predefined alternative. We consider prede" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.907, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Figure 7. SeeKer accurately signals anomalous frames at the temporal border of the abnormal action, as shown by the blue curves, even without Gaussian smoothing (example from ShanghaiTech)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.453 + ], + "angle": 0, + "content": "fined covariance (identity matrix), learned diagonal covariance \\((\\Sigma_{\\theta} = \\sigma^{2}I)\\), and learned full covariance. In the latter case, we model the inverse covariance matrix with Cholesky decomposition \\(\\Sigma_{\\theta}^{-1} = L_{\\theta}L_{\\theta}^{\\top}\\) where \\(L_{\\theta}\\) is a lower triangular matrix outputted by the model [13]. The decomposition ensures numerically stable full covariance implementation. Note that the case with the identity matrix as covariance makes the second term of objective (6) constant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Table 4 compares the performance of learned covariance, either full or diagonal, and baseline with the fixed covariance matrix. The obtained results indicate over \\(10\\%\\) absolute improvement with respect to the fixed covariance, emphasising the importance of the learned covariance in the SeeKer framework." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.554, + 0.892, + 0.617 + ], + "angle": 0, + "content": "
Covariance ΣUBnormalShanghaiTech
FullHRFullHR
identity61.463.474.174.4
full77.878.184.385.8
diagonal77.978.985.586.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.628, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Table 4. Learned covariance significantly outperforms the predefined identity matrix covariance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.854 + ], + "angle": 0, + "content": "On prediction granularity. Autoregressive factorization can be spanned at different granularity levels, i.e. at the keypoint or skeleton level. Each of these options requires the adjustment in masking strategy, as detailed in the Appendix B. Table 5 validates the performance of each of these granularities with the same hyperparameters. We observe that modeling skeleton sequences at the keypoint level consistently outperforms its skeleton-level counterpart across the datasets. Interestingly, both granularities from Table 5 outperform the current state of the art on UBnormal. This outcome underscores the effectiveness of our autoregressive modelling of skeleton sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "On the impact of keypoint uncertainty. Our anomaly score weights conditional log-likelihoods of each keypoint with its keypoint extractor uncertainty from the skeleton" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11585" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.434, + 0.14 + ], + "angle": 0, + "content": "
GranularityUBnormalShanghaiTech
FullHRFullHR
Keypoint77.978.985.586.9
Skeleton75.076.879.082.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.15, + 0.468, + 0.166 + ], + "angle": 0, + "content": "Table 5. Per-keypoint vs. per-skeleton prediction granularity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.484, + 0.33 + ], + "angle": 0, + "content": "detector. Table 6 shows that incorporating the keypoint uncertainty into the final anomaly score substantially enhances anomaly detection performance across all datasets. These results emphasise that aggregating uncertainties of all pipeline components contribute to the end decision. Note that previous best methods [23, 42] cannot easily incorporate keypoint extraction uncertainty in their anomaly score. We also integrate the detector uncertainty in the training process as discussed in Appendix C." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.341, + 0.445, + 0.393 + ], + "angle": 0, + "content": "
Confidence weightingAnomaly scoreUBnormalShanghaiTech
FullHRFullHR
XEq. (7)76.877.783.785.2
✓Eq. (8)77.978.985.586.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Table 6. Integrating keypoint detection uncertainty into the anomaly score improves performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.616 + ], + "angle": 0, + "content": "On the heuristic smoothing. Post-process smoothing is a common technique for reducing outliers in time-sequence data. Previous works [16, 23, 42] apply Gaussian kernel smoothing in a sliding window to refine per-frame anomaly scores. This postprocessing strategy yields around three percentage points performance improvement for the STGNF baseline. SeeKer, on the other hand, accurately detects action boundaries by construction, as shown in Figure 7. Thus, smoothing contributes less to the end score. Furthermore, if we remove postprocessing from both methods, SeeKer consistently outperforms STG-NF on all datasets." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.628, + 0.443, + 0.711 + ], + "angle": 0, + "content": "
MethodSmoothingUBnormalShanghaiTech
FullHRFullHR
STG-NF✗68.268.083.084.4
SeeKer✗76.877.783.484.9
STG-NF✓71.873.985.987.4
SeeKer✓77.978.985.586.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Table 7. Post-process smoothing only slightly boosts the performance of SeeKer. Importantly, SeeKer without smoothing outperforms STG-NF with smoothing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "On the keypoint ordering. SeeKer assumes a fixed ordering of skeleton keypoints. Ablations with different keypoint orderings incur a variance of less than 0.1, demonstrating SeeKer invariance to predefined keypoint order. Thus, we rely on the standard keypoint ordering [33] by default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "On autoregressive model architectures. We analyse the performance of SeeKer depending on the specific model architecture used to predict parameters \\(\\mu\\) and \\(\\Sigma\\). In addition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "to the causally masked fully-connected model, we also test transformer architecture [51] that use causal self-attention layers. Each transformer token is a 2D skeleton keypoint location projected into a higher dimensional space. Further, we encode timestep and keypoint IDs to positional encoding of each token, as detailed in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Table 8 shows that the masked fully connected model significantly outperforms its transformer-based counterpart. We attribute this gap to the low variability in skeleton sequences, as each person's gait differs only subtly. Additionally, skeleton sequences are much shorter than text paragraphs, and keypoint tokens have a much lower dimensionality, which influences the nature of the task and applicability of simpler models." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.321, + 0.9, + 0.373 + ], + "angle": 0, + "content": "
ArchitectureUBnormalShanghaiTech
FullHRFullHR
Transformer decoder [51]71.171.679.180.0
Masked fully connected [21]77.978.985.586.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.383, + 0.906, + 0.41 + ], + "angle": 0, + "content": "Table 8. Comparison of the proposed masked fully-connected model and the causal transformer decoder." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.451, + 0.634, + 0.467 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduced SeeKer, a detector for anomalous human behaviour based on generative modeling of skeletal sequences. Unlike previous approaches of this kind, SeeKer considers each skeleton as a collection of keypoints and models sequence density as a product of Gaussian conditional keypoint densities. The parameters of the conditionals \\((\\mu, \\Sigma)\\) are regressed by a causal deep model that ensures that each prediction can only observe the locations of the precedent keypoints. The proposed architecture infers all per-keypoint distributions with a single forward pass. We implement the corresponding causal architecture with masked fully connected layers. Interestingly, this formulation outperforms its counterpart based on a transformer decoder, probably due to the specifics of our data. During inference, we evaluate the per-keypoint distributions of the considered skeleton given the preceding keypoints in a sliding window, and summarize them into a skeleton-level anomaly score. Here, we take advantage of the composite structure of the proposed anomaly score by weighting the per-keypoint log-densities with the corresponding confidence values of the keypoint detector. The resulting anomaly score identifies abnormal behaviour by assessing the surprise of our model. Furthermore, our compositional anomaly score is highly interpretable since we can track keypoint contributions to the final decision. The conducted experiments indicate that SeeKer surpasses the state of the art on UBnormal and MSAD-HR by a wide margin, while delivering a close runner-up performance on ShanghaiTech." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11586" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.158 + ], + "angle": 0, + "content": "This work has been supported by Croatian Recovery and Resilience Fund - NextGenerationEU (grant C1.4 R5-I2.01.0001). The authors declare no competing interests." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.185, + 0.188, + 0.201 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.21, + 0.484, + 0.238 + ], + "angle": 0, + "content": "[1] Reprint of: Mahalanobis, p.c. (1936) “on the generalised distance in statistics.” Sankhya Ser. A, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.24, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[2] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.484, + 0.392 + ], + "angle": 0, + "content": "[3] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Abnormal: New benchmark for supervised open-set video anomaly detection. In Computer Vision and Pattern Recognition, 2022. 1, 4, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.394, + 0.484, + 0.449 + ], + "angle": 0, + "content": "[4] Yoshua Bengio and Samy Bengio. Modeling high-dimensional discrete data with multi-layer neural networks. In Advances in Neural Information Processing Systems. MIT Press, 1999. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[5] Maria A. Berezina, Daniel Rudoy, and Patrick J. Wolfe. Autoregressive modeling of voiced speech. In IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP, pages 5042-5045, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.478, + 0.522 + ], + "angle": 0, + "content": "[6] Vladimir Braverman. Sliding Window Algorithms. 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[8] Antonio Bärbalău, Radu Tudor Ionescu, Mariana-Juliana Georgescu, Jacob Velling Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas Baltzer Moeslund, and Mubarak Shah. Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Comput. Vis. Image Underst., page 103656, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Yucheng Chen, Yingli Tian, and Mingyi He. Monocular human pose estimation: A survey of deep learning-based methods. Comput. Vis. Image Underst., 192:102897, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.21, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[11] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.262, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[13] Garoe Dorta, Sara Vicente, Lourdes Agapito, Neill D. F. Campbell, and Ivor Simpson. Structured uncertainty prediction networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[14] Fernando Pereira dos Santos, Leonardo Sampaio Ferraz Ribeiro, and Moacir A. Ponti. Generalization of feature embeddings transferred from different video anomaly detection domains. J. Vis. Commun. Image Represent., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[15] Haoshu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 2, 4, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[16] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 5, 6, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.56, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[17] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Marco Aurelio Sterpa, Alessio Sampieri, and Fabio Galasso. Contracting skeletal kinematics for human-related video anomaly detection. Pattern Recognit., 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[18] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. ArXiv, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[19] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[20] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Claudi Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for distribution estimation. In International Conference on Machine Learning, 2015. 3, 4, 5, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 733-742. IEEE Computer Society, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[23] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In International Conference on Computer Vision, 2023. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[25] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[26] Yashswi Jain, Ashvini Kumar Sharma, Rajbabu Velmurugan, and Biplab Banerjee. Posecvae: Anomalous human activity detection. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[27] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. Bipoco: Bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection. CoRR, abs/2207.02281, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[28] Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances in Neural Information Processing Systems, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[29] Gongjin Lan, Yu Wu, Fei Hu, and Qi Hao. Vision-based human pose estimation via deep learning: A survey. IEEE Trans. Hum. Mach. Syst., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[30] Roberto Leyva, Victor Sanchez, and Chang-Tsun Li. Video anomaly detection with compact feature sets for online performance. IEEE Trans. Image Process., 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[31] Nanjun Li, Faliang Chang, and Chunsheng Liu. Human-related anomalous event detection via spatial-temporal graph convolutional autoencoder with embedded long short-term memory network. Neurocomputing, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.624, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[32] Yingcong Li, Yixiao Huang, Muhammed Emrullah Ildiz, Ankit Singh Rawat, and Samet Oymak. Mechanics of next token prediction with self-attention. In International Conference on Artificial Intelligence and Statistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[35] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6536-6545. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[36] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[37] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[38] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In International Conference on Computer Vision, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[39] Weixin Luo, Wen Liu, and Shenghua Gao. Normal graph: Spatial temporal graph convolutional networks based prediction network for skeleton based video anomaly detection. Neurocomputing, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[40] Eran Malach. Auto-regressive next-token predictors are universal learners. In International Conference on Machine Learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[41] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[42] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: multiscale log-density estimation via denoising score matching for video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[43] Xuan Mo, Vishal Monga, and Raja Bala. Simultaneous sparsity model for multi-perspective video anomaly detection. In International Conference on Image Processing, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.905, + 0.638 + ], + "angle": 0, + "content": "[44] Bharathkumar Ramachandra and Michael J. Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.906, + 0.724 + ], + "angle": 0, + "content": "[45] Royston Rodrigues, Neha Bhargava, Rajbabu Velmurugan, and Subhasis Chaudhuri. Multi-timescale trajectory prediction for abnormal human activity detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.727, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[46] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.905, + 0.812 + ], + "angle": 0, + "content": "[47] Benjamin Sapp, Alexander Toshev, and Ben Taskar. Cascaded models for articulated pose estimation. In European Conference on Computer Vision, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Min Sun, Pushmeet Kohli, and Jamie Shotton. Conditional regression forests for human pose estimation. In Computer Vision and Pattern Recognition, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[50] Davoud Ataee Tarzanagh, Yingcong Li, Xuechen Zhang, and Samet Oymak. Max-margin token selection in attention mechanism. In Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 2, 4, 5, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[52] Alexander Waibel, Toshiyuki Hanazawa, Geoffrey E. Hinton, Kiyohiro Shikano, and Kevin J. Lang. Phoneme recognition using time-delay neural networks. IEEE Trans. Acoust. Speech Signal Process., 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[53] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[54] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. In British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018, 2018. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[55] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature prediction diffusion model for video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[56] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[57] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, July 13-19, 2018, Stockholm, Sweden, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In MM '20: The 28th ACM International Conference on Multimedia, Virtual Event / Seattle, WA, USA, October 12-16, 2020. ACM, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[59] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14724-14734. IEEE, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In 2019 IEEE/CVF International Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[61] Leyan Zhu, Zitian Wang, Si Liu, Xuecheng Nie, Luoqi Liu, and Bo Li. Multi-person pose regression with distribution-aware single-stage models. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[62] Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise review and a new dataset. Advances in Neural Information Processing Systems, 37:89943-89977, 2025. 5, 6, 14" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11589" + } + ] +] \ No newline at end of file diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_origin.pdf b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d71ff8148dd15986a27e2757d01b7909e51fcc4d --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/a7b634d4-2fd6-469f-a18b-58cbc0beefcb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a61dced6fa367186b7e7bf8363f1a0b419d94fc31479d19546a8e578b95f0b9 +size 4589315 diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/full.md b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95650d2ea9db2acc0ee60f59af2fe1c36efd346d --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/full.md @@ -0,0 +1,313 @@ +# Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection + +Anja Delić, Matej Grcić, Siniša Šegvic +University of Zagreb, Faculty of Electrical Engineering and Computing Zagreb, Croatia + +name.surname@fer.hr + +# Abstract + +Detecting anomalous human behaviour is an important visual task in safety-critical applications such as healthcare monitoring, workplace safety, or public surveillance. In these contexts, abnormalities are often reflected with unusual human poses. Thus, we propose SeeKer, a method for detecting anomalies in sequences of human skeletons. Our method formulates the skeleton sequence density through autoregressive factorization at the keypoint level. The corresponding conditional distributions represent probable keypoint locations given prior skeletal motion. We formulate the joint distribution of the considered skeleton as causal prediction of conditional Gaussians across its constituent keypoints. A skeleton is flagged as anomalous if its keypoint locations surprise our model (i.e. receive a low density). In practice, our anomaly score is a weighted sum of per-keypoint log-conditionals, where the weights account for the confidence of the underlying keypoint detector. Despite its conceptual simplicity, SeeKer surpasses all previous methods on the UBnormal and MSAD-HR datasets while delivering competitive performance on the ShanghaiTech dataset. + +# 1. Introduction + +Anomaly detection in videos is a safety-critical computer vision task that focuses on identifying events that deviate from regular patterns within observed video sequences [30, 43]. This task is particularly relevant for applications such as public institution monitoring, workplace safety, or crowd surveillance [34, 48], where abnormal frames may indicate emergencies such as strokes, accidents or acts of violence [3]. Designing more accurate human-related anomaly detectors may improve the security of shared spaces and contribute to overall public safety. + +Capturing normality in raw RGB video sequences is challenging due to high resolution, complex back + +![](images/2363106981f0e5e0c4bf12a0452b6d277a1f79c9dc53df88eb69814244e9ab5e.jpg) +Figure 1. Given preceding skeleton keypoints (colored dots), SeeKer estimates the multivariate normal distribution of probable locations for the subsequent keypoint. Observing keypoints at improbable locations indicates anomalous behaviour (e.g. fainting). + +grounds, and spatio-temporal correlations between consecutive frames [34]. Consequently, previous anomaly detection approaches turn to feature representations extracted with deep models [2, 55] and optical flow predictions [36] as well as to combinations of multiple modalities [42]. Although these approaches achieve varying degrees of success, video anomaly detection remains an open challenge. In the context of human-related anomalies, promising approaches rely solely on skeletal representations of human poses [16, 23]. These are particularly relevant as problematic human behaviour can often be associated with irregular body positions [3]. Moreover, skeleton sequences provide a low-dimensional and anonymous representation that reduces susceptibility to biases such as variations in clothing, body shape, or ethnicity and ensures that anomaly detection is based solely on motion dynamics. Previous approaches to skeleton-based anomaly detection [17, 39, 55] consider skeletons as graphs and train graph-oriented models [56] to maximize the likelihood of the skeleton sequence [23]. However, such monolithic approaches ignore the causality in skeleton sequences inherent in the temporal domain and overlook the compositional nature of skeletons. Moreover, anomaly scores of these approaches cannot account for the uncertainty of skeleton keypoint detectors. + +In this manuscript, we propose SeeKer (Sequential Keypoint Density Estimator), a method for detecting anomalous human behaviour by predicting the conditional density of skeleton keypoints, as visualized in Figure 1. Given the preceding keypoints, SeeKer predicts a multivariate normal distribution over possible locations for the subsequent keypoint. Applying this approach across the entire skeleton sequence results in an autoregressive factorization of the joint sequence density at the keypoint level. Consequently, SeeKer training corresponds to maximizing sequence likelihood. During inference, SeeKer identifies skeletons as anomalous if their constituent keypoints deviate from the predicted distributions. The corresponding skeleton anomaly score is a sum of keypoint likelihoods weighted by the keypoint extraction confidence from an off-the-shelf skeleton extractor. Thus, our final decision accounts for uncertainties across all pipeline components. Overall, SeeKer presents the following key novelties to skeleton-based anomaly detection: autoregressive density factorization at the keypoint level, interpretable compositional anomaly scoring, and principled inclusion of keypoint detection uncertainty. Experimental evaluation shows that SeeKer achieves six percentage points absolute improvement over the best previous method on the UBnormal dataset and five percentage points absolute improvement on the MSAD-HR dataset in terms of AUROC. + +# 2. Related work + +Video anomaly detection. Video anomaly detection [30, 43] focuses on identifying anomalous frames within video sequences. This task is particularly relevant in surveillance [48] where anomalies may indicate accidents or emergencies [34]. Early approaches to frame-based video anomaly detection map input frames into a shared feature space [2, 14] and apply decision rules that differentiate between normal and abnormal patterns [25]. More recent approaches [8, 16, 19] focus on bounding box detections and recognize anomalies using self-supervised features [20, 53]. Other methods leverage failed optical flow predictions [36], errors in denoising of diffusion models [55], and predictions from cooperatively trained deep models [59]. In addition, there are exciting density estimation techniques such as denoising score matching [42]. However, these RGB-based methods may raise ethical concerns due to potential appearance-based biases and difficult anonymization. SeeKer builds atop skeleton sequences, and thus avoids such concerns. + +Skeleton-based video anomaly detection. Video anomaly detection based on human skeleton sequences [11, 41] is particularly effective due to anonymity, low computational requirements, and domain invariance [23]. Existing skeleton-based methods capture spatio-temporal dependencies within skeleton sequences using recurrent models [11] or convolutional variational autoencoders [26, 27]. Alterna- + +tively, viewing skeletons as graphs [17, 23, 39, 41] allows for processing with graph-based architectures [56, 57]. The trained models then detect anomalies as failed pose predictions [16, 26] or unlikely sequences [23]. Unlike previous approaches, SeeKer leverages the compositional nature of skeletons and applies autoregressive sequence factorization at the keypoint level. This results in improved alignment, principled inclusion of keypoint detection uncertainty, and interpretable scoring in the sense that we can easily identify the keypoint(s) that triggered the anomaly detector. + +Human pose estimation. Human pose estimation [47, 49] aims to detect key body joints in an image and assemble them into skeletal structures. Recent pose estimators [15, 61] allow for accurate real-time detection of all skeletons in a given image, facilitating skeleton-based recognition. SeeKer builds on efficient pose estimators [15] to enable real-time anomaly detection. An overview of human pose estimation can be found in surveys [10, 29]. + +Autoregressive data likelihood factorization. Autoregressive factorization of data likelihood emerged as a foundational approach across numerous machine learning applications with a dominant temporal domain [5, 12]. Such factorization is conveniently modeled using sliding window based approaches [6, 52], recurrent neural networks [24, 46] or transformer decoders [9, 51] with causal self-attention layers [32, 50]. SeeKer relies on autoregressive factorization of skeleton sequence density. + +# 3. Skeleton anomaly detection by predicting probable locations of subsequent keypoint + +Problem setup. Let $\mathcal{D} = \{\mathbf{X}^i = (X_1^i, X_2^i, \ldots, X_{T_i}^i)\}_{i=1}^M$ be a dataset of $M$ skeleton sequences $\mathbf{X}^i$ that track human activity throughout $T_i$ time steps. Each skeleton $X_t^i \in \mathbf{X}^i$ is extracted from the corresponding image and tracked across multiple frames with AlphaPose [15], an off-the-shelf detector and tracker. Each skeleton $X_t$ is an $N \times D$ matrix with rows corresponding to skeleton joints and columns representing the spatial coordinates of the detected keypoint. In practice, $N = 18$ (number of keypoints) and $D = 2$ (coordinates of each keypoint). Thus, each sequence $\mathbf{X}^i$ forms a third-order tensor of shape $T_i \times N \times D$ . Given the dataset $\mathcal{D}$ of normal skeleton sequences, our objective is to develop an anomaly detector $\tau: \mathbf{X}^i \to \{0,1\}_{T_i}^{T_i}$ capable of distinguishing between regular and anomalous skeletons in each time instant of a given sequence. + +The core approach behind SeeKer is to predict a conditional distribution of probable keypoint locations based on the preceding keypoints in the skeleton sequence, as visualized in Figure 2. Leveraging this design in conjunction with the compositional nature of skeletons provides a highly interpretable skeleton anomaly score that aggregates likelihoods estimated at the keypoint level. + +![](images/ad66c28b84dfc99668a59007e82cb28a5b239dda5ae16615338160c78ba4aff7.jpg) +Figure 2. SeeKer training starts by flattening the target skeleton $X_{t}$ and the preceding skeletons $X_{ \delta]]$ , where $[[\cdot]]$ represents Iverson brackets. Note that the proposed anomaly score readily discloses perkeypoint contributions to the decision, which makes it much more interpretable than the current state of the art [23, 42]. + +In the case when multiple humans are present in a scene $\mathcal{I}_t$ , we simply determine the per-frame anomaly score as the maximum score among all skeletons within that frame. Consequently, our per-frame anomaly score $s_F$ equals: + +$$ +s _ {F} \left(\mathcal {I} _ {t} \mid \mathcal {I} _ {< t}\right) = \max _ {X _ {t} \in \mathcal {I} _ {t}} s \left(X _ {t} \mid \mathbf {X} _ {\Delta}\right). \tag {9} +$$ + +Next, we experimentally validate SeeKer utility on the standard human-related anomaly detection benchmarks. + +# 4. Experimental setup + +Datasets. We consider three publicly available datasets for unsupervised video anomaly detection in skeleton sequences. UBnormal [3] is a synthetic dataset with 268 training, 64 validation, and 211 test video sequences. The inlier sequences contain actions such as walking while texting and talking to others. Anomalous actions include unexpected behaviour, such as stealing or fighting. This is the most relevant dataset for our task due to highly accurate labels. ShanghaiTech [34] consists of 330 training and 107 test videos of real-world scenes. Walking, standing, or sitting are considered as inlier actions, while cycling, skateboarding, and fighting are anomalous. Furthermore, + +both datasets contain anomalous events that are not human-related (e.g. car accident or fire). Consequently, we also consider dataset subsets that contain only human-related anomalies as suggested in [16, 45] referred to as UBnormalHR and ShanghaiTech-HR. The MSAD [62] dataset includes videos captured across diverse locations and various viewpoints, ensuring a wide range of perspectives and conditions, making it the most challenging benchmark to date. MSAD contains 720 videos in total, some extending up to 6K frames, and features 55 distinct abnormal events. We focus on the human-related subset with 20 distinct abnormal actions, MSAD-HR, as the full dataset track includes scenarios and events related to industrial automation and manufacturing without any human involvement. Note that nonhuman related anomalies on UBnormal and ShanghaiTech still involve humans, making them detectable through human reactions to events like fire or unusual vehicle behavior. Previous work [23] noticed that contemporary skeleton extractors [15] and trackers [54] fail to produce satisfactory outputs on some datasets (UCFCrime [48] and Avenue [37]). Hence, we skip these datasets in our experiments. + +Metrics. We report the standard anomaly detection metrics AUROC and AP, as well as human-related anomaly detection metrics RBDC and TBDC. Notably, we are the first to introduce the measurement of AP in this context. We compute AUROC and AP jointly over all test frames, which therefore corresponds to micro-averaging, e.g. micro-AUROC. We also report the region-based and track-based detection criteria (RBDC and TBDC) [44] to evaluate the spatial and temporal localization of detected anomalies on UBnormal and ShanghaiTech since they offer location annotations. RBDC assesses each detected region by calculating its Intersection-over-Union (IoU) with the corresponding ground truth region, labeling it as a true positive if the overlap exceeds a specified threshold $\alpha$ . On the other hand, TBDC evaluates the accuracy of tracking abnormal regions over time, identifying a detected track as a true positive if the number of detections within the track exceeds a threshold $\beta$ . Following the previous work [8], we set both $\alpha$ and $\beta$ at 0.1. We calculate the area under the ROC curve for both metrics, taking into account the tradeoff between detection rate and false-positive rate. For this evaluation, we rely on the ground truths provided by [19]. Some methods are not suitable for calculating RBDC and TBDC since they cannot pair features to their temporal or spatial locations in the original videos [42]. + +Baselines. We organize the considered baselines with respect to the modality they receive on the input. Methods that operate on deep features [22, 34] are denoted with (D). Methods that utilize optical flow velocities [36, 57] are denoted with (V), while skeletons-based approaches [23, 42] are denoted with (S). We briefly discuss the most relevant skeleton-based methods. STG-NF [23] proposes to estimate + +the likelihood of skeleton sequences with a graph-oriented normalizing flow with space and time separable graph convolutions in the coupling layers. MoCoDAD [16] proposes to model both normality and abnormality as multimodal by forecasting future skeletons with a diffusion model conditioned on the past. MULDE [42] is a recent method that leverages an energy-based model trained by a modification of denoising score matching that requires the model to have two times differentiable layers. For a fair comparison, we evaluate the AP of the most relevant baseline, STG-NF, with models provided by the authors. + +Implementation details. Our skeleton extractor procedure closely follows STG-NF [23]. We extract human skeletons with AlphaPose [15] within the YOLOX bounding boxes [18] and track skeletons with reID [60] across the video. The output skeletons consist of 18 keypoints in the COCO pose format [33]. We divide the extracted skeleton trajectories into time segments of $T = 24$ frames and normalize them to have zero mean and unit variance. + +We predict the per-keypoint distribution parameters $\mu$ and $\Sigma$ with deep autoregressive models that we configure for skeleton sequences. We consider causally masked fully connected models [21], transformers [51], and recurrent models. We train the causal fully connected model for 20 epochs on UBnormal and MSAD-HR and 10 epochs on ShanghaiTech with the Adam optimizer, learning rate $10^{-3}$ and batch size 256. Since UBnormal is the only dataset with a validation split, we validate all hyperparameters on the UBnormal validation subset and reuse them for ShanghaiTech and MSAD-HR. During inference, SeeKer considers $T$ preceding frames in a sliding-window manner. At the beginning of a sequence, we backtrack the score of the $T$ -th frame to the first $T - 1$ frames. Following [16, 23, 42], we finally apply Gaussian smoothing of the computed anomaly scores. However, we apply smoothing for each person separately before aggregating the per-frame anomaly score with $\sigma = 10$ . All reported results are averaged over three runs, and variance is consistently below 0.05 pp. Our code is publicly available1. + +# 5. Experimental results + +# 5.1. Skeleton-based video anomaly detection + +Table 1 compares SeeKer with previous methods on the UBnormal dataset. Our method substantially improves over all previous approaches in terms of AUROC, AP, and RBDC. Compared to the best previous skeleton-based method STGNF [23], SeeKer attains substantial increase of over 6 percentage points (pp) AUROC, 17 pp improvement in terms of AP, and 10 pp improvement in terms of RBDC, all while maintaining competitive TBDC. Moreover, SeeKer achieves over 15 pp AUROC improvement compared to + +the leading method based on deep features Ssmtl++ [20]. SeeKer also outperforms the multimodal baseline MULDE [42], despite relying solely on skeletons. For human-related anomalies, SeeKer achieves over 7 pp gain in terms of AUROC and over 12 pp AP improvement over the best performing baseline STG-NF [23]. Thus, SeeKer is a new best model for anomaly detection on the UBnormal dataset. + +
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓BAF [19]59.3---21.953.4
✓Jigsaw [53]56.4---11.836.8
✓Sglm++ [20]62.1---25.663.5
✓ ✓FPDM[55]62.7-----
✓ ✓ ✓MULDE [42]72.8-----
BiPOCO [27]50.752.3----
GEPC [41]53.455.2----
MPED-RNN [11]60.661.2----
COSKAD [17]65.065.5----
MoCoDAD [16]68.368.4----
STG-NF [23]71.871.562.767.231.762.3
SeeKer77.978.980.379.842.058.5
+ +Table 2 compares SeeKer performance with all previous approaches on the ShanghaiTech dataset. Among skeleton-based methods, our approach attains the highest AP and RBDC, while ranking second in AUROC and TBDC. Notably, the improvement upon the top-performing skeleton baseline STG-NF is over 9pp in terms of RBDC. In the case of human-related anomalies, SeeKer achieves the best AP and the second-best AUROC among skeleton-based approaches and continues to outperform methods that depend on deep features and optical flow velocities. Overall, SeeKer is a competitive model on ShanghaiTech dataset. + +Table 3 compares SeeKer with relevant skeleton-based methods on the MSAD-HR dataset [62]. As the first to report results on MSAD-HR, we reproduce the most relevant baselines [16, 23]. SeeKer surpasses these baselines by a substantial margin of $5.4\mathrm{pp}$ in AUROC and more than 3 pp in terms of AP. The MSAD-HR dataset presents significant challenges due to crowded scenes, where frequent human trajectory intersections impact skeleton estimation and tracking accuracy. Consequently, the baseline methods only marginally exceed random guessing. Next, we qualitatively analyse SeeKer performance to gain deeper insights into the obtained quantitative results. + +# 5.2. Qualitative performance + +The SeeKer anomaly score (8) is highly interpretable, as it consolidates keypoint likelihoods into skeleton-level anomaly scores while operating in the original 2D space. Figure 5 shows the difference in predictions during nor + +Table 1. Experimental evaluation of unsupervised anomaly detection on UBnormal [3]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S). + +
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓sRNN [38]68.0-----
✓Conv-AE [22]70.469.8----
✓LSA [2]72.5-----
✓ ✓FFP [35]72.872.7----
✓GCL [59]79.6-----
✓ ✓CAE-SVM [25]78.7-----
✓ ✓VEC [58]74.8-----
✓ ✓HF2 [36]76.2-----
✓ ✓FPDM[55]78.6-----
✓SSMTL [20]82.7-----
✓ ✓BA-AED [19]82.7---41.378.8
✓SSMTL++ [8]83.8---47.185.6
✓Jigsaw [53]84.284.7--22.460.8
✓ ✓ ✓MULDE [42]86.7---52.783.6
✓BiPOCO [27]-74.9----
✓MPED-RNN [11]73.475.4----
✓MTP [45]76.077.0----
✓GEPC [41]76.174.8----
✓PoseCVAE [26]-75.5----
✓Normal Graph [39]-76.5----
✓COSKAD [17]-77.1----
✓GCAE-LSTM [31]-77.2----
✓MoCoDAD [16]-77.6----
✓MULDE [42]78.5-----
✓STG-NF [23]85.987.477.681.452.182.4
✓SeeKer85.586.980.081.562.581.8
+ +Table 2. Experimental evaluation of anomaly detection on ShanghaiTech [38]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S). + +
ModalityMethodAUROCAP
DVS
✓53.951.7
✓55.756.5
✓61.160.1
+ +Table 3. Experimental evaluation of unsupervised anomaly detection on the MSAD-HR [62]. All baseline evaluations were reproduced by us. + +mal and abnormal behavior. The observations align closely with predictions of possible keypoint locations at normal timesteps. At abnormal timesteps, the predictions tend to follow normal motion patterns and deviate significantly from actual observations, resulting in high anomaly scores. In addition, Figure 6 emphasizes the interpretability of perkeypoint contributions. At each timestep, keypoints contribute with different intensity to the abnormality of the skeleton, as visualized through color variation. Notably, this level of analysis is not feasible with previous methods. + +Figure 7 compares SeeKer and the best skeleton-oriented baseline STG-NF [23] on an anomalous sequence from ShanghaiTech (bike riding in the pedestrian area). We plot both the raw anomaly scores and the anomaly scores smoothed with a 1D Gaussian filter. The x- and y-axes represent time and anomaly score, respectively, and red-shaded + +![](images/4236ed622456be311c4aa6f0dd5d93e41bc096411534fa338b4577966a4f5f5c.jpg) + +![](images/35e719f9c705375bd25caf53af45046140be39472e795b760ebfbc88d797ebce.jpg) +Figure 5. Qualitative performance of SeeKer on an example from UBnormal. During normal behavior, the observed keypoints emerge at anticipated locations. Abnormal behavior causes significant deviation from the expected keypoint locations, which enables seamless anomaly detection. + +![](images/d710bbae1c2333f81ba1734c704798359607c442a0ce755774211e99e9d54988.jpg) +Figure 6. Interpretability of SeeKer shown on an anomalous action from UBnormal (running). SeeKer anomaly score aggregates perkeypoint likelihoods denoted with varying color intensities. Each keypoint contribution to the final score changes over time. + +regions mark frames labeled as anomalous. SeeKer accurately identifies the anomalous frames, while STG-NF fails to detect anomalies at the temporal borders of the abnormal action. Prior work [42] notes that ShanghaiTech contains noisy labels for skeleton-based anomaly detection. For instance, a bicycle wheel may appear before the rider, delaying detection by skeleton-based methods. + +# 5.3. Discussion + +For all experiments in the discussion, we report AUROC on the UBnormal and ShanghaiTech test subsets. + +On predefined or learned covariance. SeeKer predicts parameters of multivariate normal distribution, $\mu$ and $\Sigma$ . We validate the importance of the learned covariance matrix compared to the predefined alternative. We consider prede + +![](images/91bc72656f2f16461d0781e124726c412c04c9175df21e34ad9e036bb1a31c06.jpg) +Figure 7. SeeKer accurately signals anomalous frames at the temporal border of the abnormal action, as shown by the blue curves, even without Gaussian smoothing (example from ShanghaiTech). + +fined covariance (identity matrix), learned diagonal covariance $(\Sigma_{\theta} = \sigma^{2}I)$ , and learned full covariance. In the latter case, we model the inverse covariance matrix with Cholesky decomposition $\Sigma_{\theta}^{-1} = L_{\theta}L_{\theta}^{\top}$ where $L_{\theta}$ is a lower triangular matrix outputted by the model [13]. The decomposition ensures numerically stable full covariance implementation. Note that the case with the identity matrix as covariance makes the second term of objective (6) constant. + +Table 4 compares the performance of learned covariance, either full or diagonal, and baseline with the fixed covariance matrix. The obtained results indicate over $10\%$ absolute improvement with respect to the fixed covariance, emphasising the importance of the learned covariance in the SeeKer framework. + +
Covariance ΣUBnormalShanghaiTech
FullHRFullHR
identity61.463.474.174.4
full77.878.184.385.8
diagonal77.978.985.586.9
+ +Table 4. Learned covariance significantly outperforms the predefined identity matrix covariance. + +On prediction granularity. Autoregressive factorization can be spanned at different granularity levels, i.e. at the keypoint or skeleton level. Each of these options requires the adjustment in masking strategy, as detailed in the Appendix B. Table 5 validates the performance of each of these granularities with the same hyperparameters. We observe that modeling skeleton sequences at the keypoint level consistently outperforms its skeleton-level counterpart across the datasets. Interestingly, both granularities from Table 5 outperform the current state of the art on UBnormal. This outcome underscores the effectiveness of our autoregressive modelling of skeleton sequences. + +On the impact of keypoint uncertainty. Our anomaly score weights conditional log-likelihoods of each keypoint with its keypoint extractor uncertainty from the skeleton + +
GranularityUBnormalShanghaiTech
FullHRFullHR
Keypoint77.978.985.586.9
Skeleton75.076.879.082.1
+ +detector. Table 6 shows that incorporating the keypoint uncertainty into the final anomaly score substantially enhances anomaly detection performance across all datasets. These results emphasise that aggregating uncertainties of all pipeline components contribute to the end decision. Note that previous best methods [23, 42] cannot easily incorporate keypoint extraction uncertainty in their anomaly score. We also integrate the detector uncertainty in the training process as discussed in Appendix C. + +Table 5. Per-keypoint vs. per-skeleton prediction granularity. + +
Confidence weightingAnomaly scoreUBnormalShanghaiTech
FullHRFullHR
XEq. (7)76.877.783.785.2
✓Eq. (8)77.978.985.586.9
+ +On the heuristic smoothing. Post-process smoothing is a common technique for reducing outliers in time-sequence data. Previous works [16, 23, 42] apply Gaussian kernel smoothing in a sliding window to refine per-frame anomaly scores. This postprocessing strategy yields around three percentage points performance improvement for the STGNF baseline. SeeKer, on the other hand, accurately detects action boundaries by construction, as shown in Figure 7. Thus, smoothing contributes less to the end score. Furthermore, if we remove postprocessing from both methods, SeeKer consistently outperforms STG-NF on all datasets. + +Table 6. Integrating keypoint detection uncertainty into the anomaly score improves performance. + +
MethodSmoothingUBnormalShanghaiTech
FullHRFullHR
STG-NF✗68.268.083.084.4
SeeKer✗76.877.783.484.9
STG-NF✓71.873.985.987.4
SeeKer✓77.978.985.586.9
+ +On the keypoint ordering. SeeKer assumes a fixed ordering of skeleton keypoints. Ablations with different keypoint orderings incur a variance of less than 0.1, demonstrating SeeKer invariance to predefined keypoint order. Thus, we rely on the standard keypoint ordering [33] by default. + +On autoregressive model architectures. We analyse the performance of SeeKer depending on the specific model architecture used to predict parameters $\mu$ and $\Sigma$ . In addition + +to the causally masked fully-connected model, we also test transformer architecture [51] that use causal self-attention layers. Each transformer token is a 2D skeleton keypoint location projected into a higher dimensional space. Further, we encode timestep and keypoint IDs to positional encoding of each token, as detailed in Appendix B. + +Table 8 shows that the masked fully connected model significantly outperforms its transformer-based counterpart. We attribute this gap to the low variability in skeleton sequences, as each person's gait differs only subtly. Additionally, skeleton sequences are much shorter than text paragraphs, and keypoint tokens have a much lower dimensionality, which influences the nature of the task and applicability of simpler models. + +Table 7. Post-process smoothing only slightly boosts the performance of SeeKer. Importantly, SeeKer without smoothing outperforms STG-NF with smoothing. + +
ArchitectureUBnormalShanghaiTech
FullHRFullHR
Transformer decoder [51]71.171.679.180.0
Masked fully connected [21]77.978.985.586.9
+ +Table 8. Comparison of the proposed masked fully-connected model and the causal transformer decoder. + +# 6. Conclusion + +We introduced SeeKer, a detector for anomalous human behaviour based on generative modeling of skeletal sequences. Unlike previous approaches of this kind, SeeKer considers each skeleton as a collection of keypoints and models sequence density as a product of Gaussian conditional keypoint densities. The parameters of the conditionals $(\mu, \Sigma)$ are regressed by a causal deep model that ensures that each prediction can only observe the locations of the precedent keypoints. The proposed architecture infers all per-keypoint distributions with a single forward pass. We implement the corresponding causal architecture with masked fully connected layers. Interestingly, this formulation outperforms its counterpart based on a transformer decoder, probably due to the specifics of our data. During inference, we evaluate the per-keypoint distributions of the considered skeleton given the preceding keypoints in a sliding window, and summarize them into a skeleton-level anomaly score. Here, we take advantage of the composite structure of the proposed anomaly score by weighting the per-keypoint log-densities with the corresponding confidence values of the keypoint detector. The resulting anomaly score identifies abnormal behaviour by assessing the surprise of our model. Furthermore, our compositional anomaly score is highly interpretable since we can track keypoint contributions to the final decision. The conducted experiments indicate that SeeKer surpasses the state of the art on UBnormal and MSAD-HR by a wide margin, while delivering a close runner-up performance on ShanghaiTech. + +# 7. Acknowledgements + +This work has been supported by Croatian Recovery and Resilience Fund - NextGenerationEU (grant C1.4 R5-I2.01.0001). The authors declare no competing interests. + +# References + +[1] Reprint of: Mahalanobis, p.c. (1936) “on the generalised distance in statistics.” Sankhya Ser. A, 2018. 4 +[2] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1, 2, 6 +[3] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Abnormal: New benchmark for supervised open-set video anomaly detection. In Computer Vision and Pattern Recognition, 2022. 1, 4, 6, 13, 14 +[4] Yoshua Bengio and Samy Bengio. Modeling high-dimensional discrete data with multi-layer neural networks. In Advances in Neural Information Processing Systems. MIT Press, 1999. 12 +[5] Maria A. Berezina, Daniel Rudoy, and Patrick J. Wolfe. Autoregressive modeling of voiced speech. In IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP, pages 5042-5045, 2010. 2 +[6] Vladimir Braverman. Sliding Window Algorithms. 2015. 2 +[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3, 13 +[8] Antonio Bärbalău, Radu Tudor Ionescu, Mariana-Juliana Georgescu, Jacob Velling Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas Baltzer Moeslund, and Mubarak Shah. Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Comput. Vis. Image Underst., page 103656, 2022. 2, 5, 6 +[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. 2 +[10] Yucheng Chen, Yingli Tian, and Mingyi He. Monocular human pose estimation: A survey of deep learning-based methods. Comput. Vis. Image Underst., 192:102897, 2020. 2 + +[11] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2, 6 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, 2019. 2 +[13] Garoe Dorta, Sara Vicente, Lourdes Agapito, Neill D. F. Campbell, and Ivor Simpson. Structured uncertainty prediction networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 7 +[14] Fernando Pereira dos Santos, Leonardo Sampaio Ferraz Ribeiro, and Moacir A. Ponti. Generalization of feature embeddings transferred from different video anomaly detection domains. J. Vis. Commun. Image Represent., 2019. 2 +[15] Haoshu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 2, 4, 5, 15 +[16] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 5, 6, 8, 14 +[17] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Marco Aurelio Sterpa, Alessio Sampieri, and Fabio Galasso. Contracting skeletal kinematics for human-related video anomaly detection. Pattern Recognit., 2024. 1, 2, 6 +[18] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. ArXiv, 2021. 5 +[19] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 2, 5, 6 +[20] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Claudi Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6 +[21] Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for distribution estimation. In International Conference on Machine Learning, 2015. 3, 4, 5, 8, 12 +[22] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal + +regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 733-742. IEEE Computer Society, 2016. 5, 6 +[23] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In International Conference on Computer Vision, 2023. 1, 2, 4, 5, 6, 8 +[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 1997. 2, 4 +[25] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 2, 6 +[26] Yashswi Jain, Ashvini Kumar Sharma, Rajbabu Velmurugan, and Biplab Banerjee. Posecvae: Anomalous human activity detection. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, 2020. 2, 6 +[27] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. Bipoco: Bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection. CoRR, abs/2207.02281, 2022. 2, 6 +[28] Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances in Neural Information Processing Systems, 2016. 4 +[29] Gongjin Lan, Yu Wu, Fei Hu, and Qi Hao. Vision-based human pose estimation via deep learning: A survey. IEEE Trans. Hum. Mach. Syst., 2023. 2 +[30] Roberto Leyva, Victor Sanchez, and Chang-Tsun Li. Video anomaly detection with compact feature sets for online performance. IEEE Trans. Image Process., 2017. 1, 2 +[31] Nanjun Li, Faliang Chang, and Chunsheng Liu. Human-related anomalous event detection via spatial-temporal graph convolutional autoencoder with embedded long short-term memory network. Neurocomputing, 2021. 6 +[32] Yingcong Li, Yixiao Huang, Muhammed Emrullah Ildiz, Ankit Singh Rawat, and Samet Oymak. Mechanics of next token prediction with self-attention. In International Conference on Artificial Intelligence and Statistics, 2024. 2 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5, 8 +[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 14 +[35] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6536-6545. Computer Vision Foundation / IEEE Computer Society, 2018. 6 + +[36] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6 +[37] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 5, 15 +[38] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In International Conference on Computer Vision, 2017. 6 +[39] Weixin Luo, Wen Liu, and Shenghua Gao. Normal graph: Spatial temporal graph convolutional networks based prediction network for skeleton based video anomaly detection. Neurocomputing, 2020. 1, 2, 6 +[40] Eran Malach. Auto-regressive next-token predictors are universal learners. In International Conference on Machine Learning, 2024. 3 +[41] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 2, 6 +[42] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: multiscale log-density estimation via denoising score matching for video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 2024. 1, 2, 4, 5, 6, 7, 8 +[43] Xuan Mo, Vishal Monga, and Raja Bala. Simultaneous sparsity model for multi-perspective video anomaly detection. In International Conference on Image Processing, 2014. 1, 2 +[44] Bharathkumar Ramachandra and Michael J. Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5 +[45] Royston Rodrigues, Neha Bhargava, Rajbabu Velmurugan, and Subhasis Chaudhuri. Multi-timescale trajectory prediction for abnormal human activity detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5, 6, 14 +[46] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 1986. 2 +[47] Benjamin Sapp, Alexander Toshev, and Ben Taskar. Cascaded models for articulated pose estimation. In European Conference on Computer Vision, 2010. 2 +[48] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 15 +[49] Min Sun, Pushmeet Kohli, and Jamie Shotton. Conditional regression forests for human pose estimation. In Computer Vision and Pattern Recognition, 2012. 2 + +[50] Davoud Ataee Tarzanagh, Yingcong Li, Xuechen Zhang, and Samet Oymak. Max-margin token selection in attention mechanism. In Neural Information Processing Systems, 2023. 2 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 2, 4, 5, 8, 13 +[52] Alexander Waibel, Toshiyuki Hanazawa, Geoffrey E. Hinton, Kiyohiro Shikano, and Kevin J. Lang. Phoneme recognition using time-delay neural networks. IEEE Trans. Acoust. Speech Signal Process., 1989. 2 +[53] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, 2022. 2, 6 +[54] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. In British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018, 2018. 5, 15 +[55] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature prediction diffusion model for video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 6 +[56] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, 2018. 1, 2 +[57] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, July 13-19, 2018, Stockholm, Sweden, 2018. 2, 5 +[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In MM '20: The 28th ACM International Conference on Multimedia, Virtual Event / Seattle, WA, USA, October 12-16, 2020. ACM, 2020. 6 +[59] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14724-14734. IEEE, 2022. 2, 6 +[60] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In 2019 IEEE/CVF International Conference + +on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, 2019. 5 +[61] Leyan Zhu, Zitian Wang, Si Liu, Xuecheng Nie, Luoqi Liu, and Bo Li. Multi-person pose regression with distribution-aware single-stage models. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2 +[62] Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise review and a new dataset. Advances in Neural Information Processing Systems, 37:89943-89977, 2025. 5, 6, 14 \ No newline at end of file diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/images.zip b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc5d8f2bb0eb35042830d1288650ff95a4e6bd45 --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d6dac70cd705a713f209d10613498849a6b865e7c9ddfaa902a12ef1b75c105 +size 544137 diff --git a/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/layout.json b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f21feb0bf8e44b11409a15c5d45457f8f5c23b33 --- /dev/null +++ b/2025/Sequential keypoint density estimator_ an overlooked baseline of skeleton-based video anomaly detection/layout.json @@ -0,0 +1,8562 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "text", + "content": "Sequential keypoint density estimator: an overlooked baseline of skeleton-based video anomaly detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 478, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 478, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 478, + 203 + ], + "type": "text", + "content": "Anja Delić, Matej Grcić, Siniša Šegvic \nUniversity of Zagreb, Faculty of Electrical Engineering and Computing Zagreb, Croatia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 206, + 356, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 206, + 356, + 216 + ], + "spans": [ + { + "bbox": [ + 251, + 206, + 356, + 216 + ], + "type": "text", + "content": "name.surname@fer.hr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 271, + 297, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 297, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 297, + 522 + ], + "type": "text", + "content": "Detecting anomalous human behaviour is an important visual task in safety-critical applications such as healthcare monitoring, workplace safety, or public surveillance. In these contexts, abnormalities are often reflected with unusual human poses. Thus, we propose SeeKer, a method for detecting anomalies in sequences of human skeletons. Our method formulates the skeleton sequence density through autoregressive factorization at the keypoint level. The corresponding conditional distributions represent probable keypoint locations given prior skeletal motion. We formulate the joint distribution of the considered skeleton as causal prediction of conditional Gaussians across its constituent keypoints. A skeleton is flagged as anomalous if its keypoint locations surprise our model (i.e. receive a low density). In practice, our anomaly score is a weighted sum of per-keypoint log-conditionals, where the weights account for the confidence of the underlying keypoint detector. Despite its conceptual simplicity, SeeKer surpasses all previous methods on the UBnormal and MSAD-HR datasets while delivering competitive performance on the ShanghaiTech dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 135, + 560 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 568, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 688 + ], + "type": "text", + "content": "Anomaly detection in videos is a safety-critical computer vision task that focuses on identifying events that deviate from regular patterns within observed video sequences [30, 43]. This task is particularly relevant for applications such as public institution monitoring, workplace safety, or crowd surveillance [34, 48], where abnormal frames may indicate emergencies such as strokes, accidents or acts of violence [3]. Designing more accurate human-related anomaly detectors may improve the security of shared spaces and contribute to overall public safety." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Capturing normality in raw RGB video sequences is challenging due to high resolution, complex back" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 244, + 553, + 353 + ], + "blocks": [ + { + "bbox": [ + 318, + 244, + 553, + 353 + ], + "lines": [ + { + "bbox": [ + 318, + 244, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 318, + 244, + 553, + 353 + ], + "type": "image", + "image_path": "2363106981f0e5e0c4bf12a0452b6d277a1f79c9dc53df88eb69814244e9ab5e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 361, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. Given preceding skeleton keypoints (colored dots), SeeKer estimates the multivariate normal distribution of probable locations for the subsequent keypoint. Observing keypoints at improbable locations indicates anomalous behaviour (e.g. fainting)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "grounds, and spatio-temporal correlations between consecutive frames [34]. Consequently, previous anomaly detection approaches turn to feature representations extracted with deep models [2, 55] and optical flow predictions [36] as well as to combinations of multiple modalities [42]. Although these approaches achieve varying degrees of success, video anomaly detection remains an open challenge. In the context of human-related anomalies, promising approaches rely solely on skeletal representations of human poses [16, 23]. These are particularly relevant as problematic human behaviour can often be associated with irregular body positions [3]. Moreover, skeleton sequences provide a low-dimensional and anonymous representation that reduces susceptibility to biases such as variations in clothing, body shape, or ethnicity and ensures that anomaly detection is based solely on motion dynamics. Previous approaches to skeleton-based anomaly detection [17, 39, 55] consider skeletons as graphs and train graph-oriented models [56] to maximize the likelihood of the skeleton sequence [23]. However, such monolithic approaches ignore the causality in skeleton sequences inherent in the temporal domain and overlook the compositional nature of skeletons. Moreover, anomaly scores of these approaches cannot account for the uncertainty of skeleton keypoint detectors." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11579" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": "In this manuscript, we propose SeeKer (Sequential Keypoint Density Estimator), a method for detecting anomalous human behaviour by predicting the conditional density of skeleton keypoints, as visualized in Figure 1. Given the preceding keypoints, SeeKer predicts a multivariate normal distribution over possible locations for the subsequent keypoint. Applying this approach across the entire skeleton sequence results in an autoregressive factorization of the joint sequence density at the keypoint level. Consequently, SeeKer training corresponds to maximizing sequence likelihood. During inference, SeeKer identifies skeletons as anomalous if their constituent keypoints deviate from the predicted distributions. The corresponding skeleton anomaly score is a sum of keypoint likelihoods weighted by the keypoint extraction confidence from an off-the-shelf skeleton extractor. Thus, our final decision accounts for uncertainties across all pipeline components. Overall, SeeKer presents the following key novelties to skeleton-based anomaly detection: autoregressive density factorization at the keypoint level, interpretable compositional anomaly scoring, and principled inclusion of keypoint detection uncertainty. Experimental evaluation shows that SeeKer achieves six percentage points absolute improvement over the best previous method on the UBnormal dataset and five percentage points absolute improvement on the MSAD-HR dataset in terms of AUROC." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 394, + 140, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 140, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 140, + 406 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 414, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 628 + ], + "type": "text", + "content": "Video anomaly detection. Video anomaly detection [30, 43] focuses on identifying anomalous frames within video sequences. This task is particularly relevant in surveillance [48] where anomalies may indicate accidents or emergencies [34]. Early approaches to frame-based video anomaly detection map input frames into a shared feature space [2, 14] and apply decision rules that differentiate between normal and abnormal patterns [25]. More recent approaches [8, 16, 19] focus on bounding box detections and recognize anomalies using self-supervised features [20, 53]. Other methods leverage failed optical flow predictions [36], errors in denoising of diffusion models [55], and predictions from cooperatively trained deep models [59]. In addition, there are exciting density estimation techniques such as denoising score matching [42]. However, these RGB-based methods may raise ethical concerns due to potential appearance-based biases and difficult anonymization. SeeKer builds atop skeleton sequences, and thus avoids such concerns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Skeleton-based video anomaly detection. Video anomaly detection based on human skeleton sequences [11, 41] is particularly effective due to anonymity, low computational requirements, and domain invariance [23]. Existing skeleton-based methods capture spatio-temporal dependencies within skeleton sequences using recurrent models [11] or convolutional variational autoencoders [26, 27]. Alterna-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "tively, viewing skeletons as graphs [17, 23, 39, 41] allows for processing with graph-based architectures [56, 57]. The trained models then detect anomalies as failed pose predictions [16, 26] or unlikely sequences [23]. Unlike previous approaches, SeeKer leverages the compositional nature of skeletons and applies autoregressive sequence factorization at the keypoint level. This results in improved alignment, principled inclusion of keypoint detection uncertainty, and interpretable scoring in the sense that we can easily identify the keypoint(s) that triggered the anomaly detector." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 193, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 289 + ], + "type": "text", + "content": "Human pose estimation. Human pose estimation [47, 49] aims to detect key body joints in an image and assemble them into skeletal structures. Recent pose estimators [15, 61] allow for accurate real-time detection of all skeletons in a given image, facilitating skeleton-based recognition. SeeKer builds on efficient pose estimators [15] to enable real-time anomaly detection. An overview of human pose estimation can be found in surveys [10, 29]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "type": "text", + "content": "Autoregressive data likelihood factorization. Autoregressive factorization of data likelihood emerged as a foundational approach across numerous machine learning applications with a dominant temporal domain [5, 12]. Such factorization is conveniently modeled using sliding window based approaches [6, 52], recurrent neural networks [24, 46] or transformer decoders [9, 51] with causal self-attention layers [32, 50]. SeeKer relies on autoregressive factorization of skeleton sequence density." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 412, + 553, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 441 + ], + "type": "text", + "content": "3. Skeleton anomaly detection by predicting probable locations of subsequent keypoint" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": "Problem setup. Let " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{X}^i = (X_1^i, X_2^i, \\ldots, X_{T_i}^i)\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " be a dataset of " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " skeleton sequences " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " that track human activity throughout " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " time steps. Each skeleton " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "X_t^i \\in \\mathbf{X}^i" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " is extracted from the corresponding image and tracked across multiple frames with AlphaPose [15], an off-the-shelf detector and tracker. Each skeleton " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "X_t" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " is an " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "N \\times D" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " matrix with rows corresponding to skeleton joints and columns representing the spatial coordinates of the detected keypoint. In practice, " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "N = 18" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " (number of keypoints) and " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "D = 2" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " (coordinates of each keypoint). Thus, each sequence " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " forms a third-order tensor of shape " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "T_i \\times N \\times D" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": ". Given the dataset " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " of normal skeleton sequences, our objective is to develop an anomaly detector " + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\tau: \\mathbf{X}^i \\to \\{0,1\\}_{T_i}^{T_i}" + }, + { + "bbox": [ + 313, + 447, + 554, + 628 + ], + "type": "text", + "content": " capable of distinguishing between regular and anomalous skeletons in each time instant of a given sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "The core approach behind SeeKer is to predict a conditional distribution of probable keypoint locations based on the preceding keypoints in the skeleton sequence, as visualized in Figure 2. Leveraging this design in conjunction with the compositional nature of skeletons provides a highly interpretable skeleton anomaly score that aggregates likelihoods estimated at the keypoint level." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11580" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 556, + 216 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 216 + ], + "type": "image", + "image_path": "ad66c28b84dfc99668a59007e82cb28a5b239dda5ae16615338160c78ba4aff7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. SeeKer training starts by flattening the target skeleton " + }, + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "type": "text", + "content": " and the preceding skeletons " + }, + { + "bbox": [ + 54, + 222, + 555, + 257 + ], + "type": "inline_equation", + "content": "X_{ \\delta]]" + }, + { + "bbox": [ + 313, + 332, + 554, + 427 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 332, + 554, + 427 + ], + "type": "inline_equation", + "content": "[[\\cdot]]" + }, + { + "bbox": [ + 313, + 332, + 554, + 427 + ], + "type": "text", + "content": " represents Iverson brackets. Note that the proposed anomaly score readily discloses perkeypoint contributions to the decision, which makes it much more interpretable than the current state of the art [23, 42]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "text", + "content": "In the case when multiple humans are present in a scene " + }, + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "text", + "content": ", we simply determine the per-frame anomaly score as the maximum score among all skeletons within that frame. Consequently, our per-frame anomaly score " + }, + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "inline_equation", + "content": "s_F" + }, + { + "bbox": [ + 313, + 429, + 554, + 476 + ], + "type": "text", + "content": " equals:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 367, + 486, + 554, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 486, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 367, + 486, + 554, + 506 + ], + "type": "interline_equation", + "content": "s _ {F} \\left(\\mathcal {I} _ {t} \\mid \\mathcal {I} _ {< t}\\right) = \\max _ {X _ {t} \\in \\mathcal {I} _ {t}} s \\left(X _ {t} \\mid \\mathbf {X} _ {\\Delta}\\right). \\tag {9}", + "image_path": "2be0a242141cd8701e51664bdae10d0f0589d578d32817ae2630c0e2444501df.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 514, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 537 + ], + "type": "text", + "content": "Next, we experimentally validate SeeKer utility on the standard human-related anomaly detection benchmarks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 549, + 430, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 430, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 430, + 563 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We consider three publicly available datasets for unsupervised video anomaly detection in skeleton sequences. UBnormal [3] is a synthetic dataset with 268 training, 64 validation, and 211 test video sequences. The inlier sequences contain actions such as walking while texting and talking to others. Anomalous actions include unexpected behaviour, such as stealing or fighting. This is the most relevant dataset for our task due to highly accurate labels. ShanghaiTech [34] consists of 330 training and 107 test videos of real-world scenes. Walking, standing, or sitting are considered as inlier actions, while cycling, skateboarding, and fighting are anomalous. Furthermore," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11582" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 324 + ], + "type": "text", + "content": "both datasets contain anomalous events that are not human-related (e.g. car accident or fire). Consequently, we also consider dataset subsets that contain only human-related anomalies as suggested in [16, 45] referred to as UBnormalHR and ShanghaiTech-HR. The MSAD [62] dataset includes videos captured across diverse locations and various viewpoints, ensuring a wide range of perspectives and conditions, making it the most challenging benchmark to date. MSAD contains 720 videos in total, some extending up to 6K frames, and features 55 distinct abnormal events. We focus on the human-related subset with 20 distinct abnormal actions, MSAD-HR, as the full dataset track includes scenarios and events related to industrial automation and manufacturing without any human involvement. Note that nonhuman related anomalies on UBnormal and ShanghaiTech still involve humans, making them detectable through human reactions to events like fire or unusual vehicle behavior. Previous work [23] noticed that contemporary skeleton extractors [15] and trackers [54] fail to produce satisfactory outputs on some datasets (UCFCrime [48] and Avenue [37]). Hence, we skip these datasets in our experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "content": "Metrics. We report the standard anomaly detection metrics AUROC and AP, as well as human-related anomaly detection metrics RBDC and TBDC. Notably, we are the first to introduce the measurement of AP in this context. We compute AUROC and AP jointly over all test frames, which therefore corresponds to micro-averaging, e.g. micro-AUROC. We also report the region-based and track-based detection criteria (RBDC and TBDC) [44] to evaluate the spatial and temporal localization of detected anomalies on UBnormal and ShanghaiTech since they offer location annotations. RBDC assesses each detected region by calculating its Intersection-over-Union (IoU) with the corresponding ground truth region, labeling it as a true positive if the overlap exceeds a specified threshold " + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "content": ". On the other hand, TBDC evaluates the accuracy of tracking abnormal regions over time, identifying a detected track as a true positive if the number of detections within the track exceeds a threshold " + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "content": ". Following the previous work [8], we set both " + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 327, + 295, + 625 + ], + "type": "text", + "content": " at 0.1. We calculate the area under the ROC curve for both metrics, taking into account the tradeoff between detection rate and false-positive rate. For this evaluation, we rely on the ground truths provided by [19]. Some methods are not suitable for calculating RBDC and TBDC since they cannot pair features to their temporal or spatial locations in the original videos [42]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Baselines. We organize the considered baselines with respect to the modality they receive on the input. Methods that operate on deep features [22, 34] are denoted with (D). Methods that utilize optical flow velocities [36, 57] are denoted with (V), while skeletons-based approaches [23, 42] are denoted with (S). We briefly discuss the most relevant skeleton-based methods. STG-NF [23] proposes to estimate" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "the likelihood of skeleton sequences with a graph-oriented normalizing flow with space and time separable graph convolutions in the coupling layers. MoCoDAD [16] proposes to model both normality and abnormality as multimodal by forecasting future skeletons with a diffusion model conditioned on the past. MULDE [42] is a recent method that leverages an energy-based model trained by a modification of denoising score matching that requires the model to have two times differentiable layers. For a fair comparison, we evaluate the AP of the most relevant baseline, STG-NF, with models provided by the authors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 204, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 299 + ], + "type": "text", + "content": "Implementation details. Our skeleton extractor procedure closely follows STG-NF [23]. We extract human skeletons with AlphaPose [15] within the YOLOX bounding boxes [18] and track skeletons with reID [60] across the video. The output skeletons consist of 18 keypoints in the COCO pose format [33]. We divide the extracted skeleton trajectories into time segments of " + }, + { + "bbox": [ + 313, + 204, + 553, + 299 + ], + "type": "inline_equation", + "content": "T = 24" + }, + { + "bbox": [ + 313, + 204, + 553, + 299 + ], + "type": "text", + "content": " frames and normalize them to have zero mean and unit variance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": "We predict the per-keypoint distribution parameters " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": " with deep autoregressive models that we configure for skeleton sequences. We consider causally masked fully connected models [21], transformers [51], and recurrent models. We train the causal fully connected model for 20 epochs on UBnormal and MSAD-HR and 10 epochs on ShanghaiTech with the Adam optimizer, learning rate " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": " and batch size 256. Since UBnormal is the only dataset with a validation split, we validate all hyperparameters on the UBnormal validation subset and reuse them for ShanghaiTech and MSAD-HR. During inference, SeeKer considers " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": " preceding frames in a sliding-window manner. At the beginning of a sequence, we backtrack the score of the " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": "-th frame to the first " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": " frames. Following [16, 23, 42], we finally apply Gaussian smoothing of the computed anomaly scores. However, we apply smoothing for each person separately before aggregating the per-frame anomaly score with " + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "inline_equation", + "content": "\\sigma = 10" + }, + { + "bbox": [ + 313, + 300, + 553, + 538 + ], + "type": "text", + "content": ". All reported results are averaged over three runs, and variance is consistently below 0.05 pp. Our code is publicly available1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 550, + 436, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 436, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 436, + 563 + ], + "type": "text", + "content": "5. Experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 523, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 523, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 523, + 582 + ], + "type": "text", + "content": "5.1. Skeleton-based video anomaly detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 586, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 553, + 696 + ], + "type": "text", + "content": "Table 1 compares SeeKer with previous methods on the UBnormal dataset. Our method substantially improves over all previous approaches in terms of AUROC, AP, and RBDC. Compared to the best previous skeleton-based method STGNF [23], SeeKer attains substantial increase of over 6 percentage points (pp) AUROC, 17 pp improvement in terms of AP, and 10 pp improvement in terms of RBDC, all while maintaining competitive TBDC. Moreover, SeeKer achieves over 15 pp AUROC improvement compared to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 494, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 494, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 494, + 712 + ], + "type": "text", + "content": "1https://github.com/adelic99/seeker" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11583" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "the leading method based on deep features Ssmtl++ [20]. SeeKer also outperforms the multimodal baseline MULDE [42], despite relying solely on skeletons. For human-related anomalies, SeeKer achieves over 7 pp gain in terms of AUROC and over 12 pp AP improvement over the best performing baseline STG-NF [23]. Thus, SeeKer is a new best model for anomaly detection on the UBnormal dataset." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 165, + 288, + 312 + ], + "blocks": [ + { + "bbox": [ + 66, + 165, + 288, + 312 + ], + "lines": [ + { + "bbox": [ + 66, + 165, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 165, + 288, + 312 + ], + "type": "table", + "html": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓BAF [19]59.3---21.953.4
✓Jigsaw [53]56.4---11.836.8
✓Sglm++ [20]62.1---25.663.5
✓ ✓FPDM[55]62.7-----
✓ ✓ ✓MULDE [42]72.8-----
BiPOCO [27]50.752.3----
GEPC [41]53.455.2----
MPED-RNN [11]60.661.2----
COSKAD [17]65.065.5----
MoCoDAD [16]68.368.4----
STG-NF [23]71.871.562.767.231.762.3
SeeKer77.978.980.379.842.058.5
", + "image_path": "229cdeb8d94222c1ff881561f4833f9a9c8c357d6903ad6f75f09ab1a650db8b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 497 + ], + "type": "text", + "content": "Table 2 compares SeeKer performance with all previous approaches on the ShanghaiTech dataset. Among skeleton-based methods, our approach attains the highest AP and RBDC, while ranking second in AUROC and TBDC. Notably, the improvement upon the top-performing skeleton baseline STG-NF is over 9pp in terms of RBDC. In the case of human-related anomalies, SeeKer achieves the best AP and the second-best AUROC among skeleton-based approaches and continues to outperform methods that depend on deep features and optical flow velocities. Overall, SeeKer is a competitive model on ShanghaiTech dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": "Table 3 compares SeeKer with relevant skeleton-based methods on the MSAD-HR dataset [62]. As the first to report results on MSAD-HR, we reproduce the most relevant baselines [16, 23]. SeeKer surpasses these baselines by a substantial margin of " + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "inline_equation", + "content": "5.4\\mathrm{pp}" + }, + { + "bbox": [ + 55, + 498, + 296, + 641 + ], + "type": "text", + "content": " in AUROC and more than 3 pp in terms of AP. The MSAD-HR dataset presents significant challenges due to crowded scenes, where frequent human trajectory intersections impact skeleton estimation and tracking accuracy. Consequently, the baseline methods only marginally exceed random guessing. Next, we qualitatively analyse SeeKer performance to gain deeper insights into the obtained quantitative results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 194, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 194, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 194, + 661 + ], + "type": "text", + "content": "5.2. Qualitative performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The SeeKer anomaly score (8) is highly interpretable, as it consolidates keypoint likelihoods into skeleton-level anomaly scores while operating in the original 2D space. Figure 5 shows the difference in predictions during nor" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 69, + 548, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 296, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 353 + ], + "type": "text", + "content": "Table 1. Experimental evaluation of unsupervised anomaly detection on UBnormal [3]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 69, + 548, + 350 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 548, + 350 + ], + "type": "table", + "html": "
Modality D V SMethodAUROCAPRBDCTBDC
FullHRFullHR
✓sRNN [38]68.0-----
✓Conv-AE [22]70.469.8----
✓LSA [2]72.5-----
✓ ✓FFP [35]72.872.7----
✓GCL [59]79.6-----
✓ ✓CAE-SVM [25]78.7-----
✓ ✓VEC [58]74.8-----
✓ ✓HF2 [36]76.2-----
✓ ✓FPDM[55]78.6-----
✓SSMTL [20]82.7-----
✓ ✓BA-AED [19]82.7---41.378.8
✓SSMTL++ [8]83.8---47.185.6
✓Jigsaw [53]84.284.7--22.460.8
✓ ✓ ✓MULDE [42]86.7---52.783.6
✓BiPOCO [27]-74.9----
✓MPED-RNN [11]73.475.4----
✓MTP [45]76.077.0----
✓GEPC [41]76.174.8----
✓PoseCVAE [26]-75.5----
✓Normal Graph [39]-76.5----
✓COSKAD [17]-77.1----
✓GCAE-LSTM [31]-77.2----
✓MoCoDAD [16]-77.6----
✓MULDE [42]78.5-----
✓STG-NF [23]85.987.477.681.452.182.4
✓SeeKer85.586.980.081.562.581.8
", + "image_path": "2b32255e227ca69df7bdea50df8a1fb47c2306a5576c38002bbc4e920601ca75.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 353, + 404, + 517, + 454 + ], + "blocks": [ + { + "bbox": [ + 313, + 356, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 390 + ], + "type": "text", + "content": "Table 2. Experimental evaluation of anomaly detection on ShanghaiTech [38]. Input modality acronyms denote deep features (D), optical flow (V), and skeletons (S)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 404, + 517, + 454 + ], + "lines": [ + { + "bbox": [ + 353, + 404, + 517, + 454 + ], + "spans": [ + { + "bbox": [ + 353, + 404, + 517, + 454 + ], + "type": "table", + "html": "
ModalityMethodAUROCAP
DVS
✓53.951.7
✓55.756.5
✓61.160.1
", + "image_path": "3f915b2ec09db9105ea309c6ad2d681e1c54765c6fa80f0c8f3adc82b91d8a89.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 496 + ], + "type": "text", + "content": "Table 3. Experimental evaluation of unsupervised anomaly detection on the MSAD-HR [62]. All baseline evaluations were reproduced by us." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "content": "mal and abnormal behavior. The observations align closely with predictions of possible keypoint locations at normal timesteps. At abnormal timesteps, the predictions tend to follow normal motion patterns and deviate significantly from actual observations, resulting in high anomaly scores. In addition, Figure 6 emphasizes the interpretability of perkeypoint contributions. At each timestep, keypoints contribute with different intensity to the abnormality of the skeleton, as visualized through color variation. Notably, this level of analysis is not feasible with previous methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Figure 7 compares SeeKer and the best skeleton-oriented baseline STG-NF [23] on an anomalous sequence from ShanghaiTech (bike riding in the pedestrian area). We plot both the raw anomaly scores and the anomaly scores smoothed with a 1D Gaussian filter. The x- and y-axes represent time and anomaly score, respectively, and red-shaded" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11584" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 293, + 148 + ], + "type": "image", + "image_path": "4236ed622456be311c4aa6f0dd5d93e41bc096411534fa338b4577966a4f5f5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 149, + 296, + 308 + ], + "blocks": [ + { + "bbox": [ + 58, + 149, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 58, + 149, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 58, + 149, + 296, + 308 + ], + "type": "image", + "image_path": "35e719f9c705375bd25caf53af45046140be39472e795b760ebfbc88d797ebce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 296, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 373 + ], + "type": "text", + "content": "Figure 5. Qualitative performance of SeeKer on an example from UBnormal. During normal behavior, the observed keypoints emerge at anticipated locations. Abnormal behavior causes significant deviation from the expected keypoint locations, which enables seamless anomaly detection." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 384, + 296, + 458 + ], + "blocks": [ + { + "bbox": [ + 58, + 384, + 296, + 458 + ], + "lines": [ + { + "bbox": [ + 58, + 384, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 58, + 384, + 296, + 458 + ], + "type": "image", + "image_path": "d710bbae1c2333f81ba1734c704798359607c442a0ce755774211e99e9d54988.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 296, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 510 + ], + "type": "text", + "content": "Figure 6. Interpretability of SeeKer shown on an anomalous action from UBnormal (running). SeeKer anomaly score aggregates perkeypoint likelihoods denoted with varying color intensities. Each keypoint contribution to the final score changes over time." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 531, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 616 + ], + "type": "text", + "content": "regions mark frames labeled as anomalous. SeeKer accurately identifies the anomalous frames, while STG-NF fails to detect anomalies at the temporal borders of the abnormal action. Prior work [42] notes that ShanghaiTech contains noisy labels for skeleton-based anomaly detection. For instance, a bicycle wheel may appear before the rider, delaying detection by skeleton-based methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 128, + 635 + ], + "type": "text", + "content": "5.3. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 665 + ], + "type": "text", + "content": "For all experiments in the discussion, we report AUROC on the UBnormal and ShanghaiTech test subsets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "On predefined or learned covariance. SeeKer predicts parameters of multivariate normal distribution, " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ". We validate the importance of the learned covariance matrix compared to the predefined alternative. We consider prede" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 555, + 198 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 555, + 198 + ], + "type": "image", + "image_path": "91bc72656f2f16461d0781e124726c412c04c9175df21e34ad9e036bb1a31c06.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": "Figure 7. SeeKer accurately signals anomalous frames at the temporal border of the abnormal action, as shown by the blue curves, even without Gaussian smoothing (example from ShanghaiTech)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "content": "fined covariance (identity matrix), learned diagonal covariance " + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "inline_equation", + "content": "(\\Sigma_{\\theta} = \\sigma^{2}I)" + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "content": ", and learned full covariance. In the latter case, we model the inverse covariance matrix with Cholesky decomposition " + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}^{-1} = L_{\\theta}L_{\\theta}^{\\top}" + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "inline_equation", + "content": "L_{\\theta}" + }, + { + "bbox": [ + 313, + 263, + 555, + 358 + ], + "type": "text", + "content": " is a lower triangular matrix outputted by the model [13]. The decomposition ensures numerically stable full covariance implementation. Note that the case with the identity matrix as covariance makes the second term of objective (6) constant." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 359, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 554, + 430 + ], + "type": "text", + "content": "Table 4 compares the performance of learned covariance, either full or diagonal, and baseline with the fixed covariance matrix. The obtained results indicate over " + }, + { + "bbox": [ + 313, + 359, + 554, + 430 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 359, + 554, + 430 + ], + "type": "text", + "content": " absolute improvement with respect to the fixed covariance, emphasising the importance of the learned covariance in the SeeKer framework." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 323, + 438, + 545, + 488 + ], + "blocks": [ + { + "bbox": [ + 323, + 438, + 545, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 438, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 438, + 545, + 488 + ], + "type": "table", + "html": "
Covariance ΣUBnormalShanghaiTech
FullHRFullHR
identity61.463.474.174.4
full77.878.184.385.8
diagonal77.978.985.586.9
", + "image_path": "55ec3c719d8499977b1e98758843ca7129cbef50de747c78ea9c1ba70a873530.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 554, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 519 + ], + "type": "text", + "content": "Table 4. Learned covariance significantly outperforms the predefined identity matrix covariance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 534, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 676 + ], + "type": "text", + "content": "On prediction granularity. Autoregressive factorization can be spanned at different granularity levels, i.e. at the keypoint or skeleton level. Each of these options requires the adjustment in masking strategy, as detailed in the Appendix B. Table 5 validates the performance of each of these granularities with the same hyperparameters. We observe that modeling skeleton sequences at the keypoint level consistently outperforms its skeleton-level counterpart across the datasets. Interestingly, both granularities from Table 5 outperform the current state of the art on UBnormal. This outcome underscores the effectiveness of our autoregressive modelling of skeleton sequences." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "On the impact of keypoint uncertainty. Our anomaly score weights conditional log-likelihoods of each keypoint with its keypoint extractor uncertainty from the skeleton" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11585" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 265, + 110 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 265, + 110 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 265, + 110 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 265, + 110 + ], + "type": "table", + "html": "
GranularityUBnormalShanghaiTech
FullHRFullHR
Keypoint77.978.985.586.9
Skeleton75.076.879.082.1
", + "image_path": "d1cb88b34c181a1425ad4a2bde3122135d68b15a72e44db81169363585442a32.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 261 + ], + "type": "text", + "content": "detector. Table 6 shows that incorporating the keypoint uncertainty into the final anomaly score substantially enhances anomaly detection performance across all datasets. These results emphasise that aggregating uncertainties of all pipeline components contribute to the end decision. Note that previous best methods [23, 42] cannot easily incorporate keypoint extraction uncertainty in their anomaly score. We also integrate the detector uncertainty in the training process as discussed in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 80, + 270, + 272, + 311 + ], + "blocks": [ + { + "bbox": [ + 64, + 118, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 64, + 118, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 64, + 118, + 286, + 131 + ], + "type": "text", + "content": "Table 5. Per-keypoint vs. per-skeleton prediction granularity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 270, + 272, + 311 + ], + "lines": [ + { + "bbox": [ + 80, + 270, + 272, + 311 + ], + "spans": [ + { + "bbox": [ + 80, + 270, + 272, + 311 + ], + "type": "table", + "html": "
Confidence weightingAnomaly scoreUBnormalShanghaiTech
FullHRFullHR
XEq. (7)76.877.783.785.2
✓Eq. (8)77.978.985.586.9
", + "image_path": "be4bfaa3bf6f1a6a6b077b882192774b8bb3b64146dbfb79f3c14e0539728e22.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "content": "On the heuristic smoothing. Post-process smoothing is a common technique for reducing outliers in time-sequence data. Previous works [16, 23, 42] apply Gaussian kernel smoothing in a sliding window to refine per-frame anomaly scores. This postprocessing strategy yields around three percentage points performance improvement for the STGNF baseline. SeeKer, on the other hand, accurately detects action boundaries by construction, as shown in Figure 7. Thus, smoothing contributes less to the end score. Furthermore, if we remove postprocessing from both methods, SeeKer consistently outperforms STG-NF on all datasets." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 82, + 497, + 271, + 563 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 296, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 342 + ], + "type": "text", + "content": "Table 6. Integrating keypoint detection uncertainty into the anomaly score improves performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 497, + 271, + 563 + ], + "lines": [ + { + "bbox": [ + 82, + 497, + 271, + 563 + ], + "spans": [ + { + "bbox": [ + 82, + 497, + 271, + 563 + ], + "type": "table", + "html": "
MethodSmoothingUBnormalShanghaiTech
FullHRFullHR
STG-NF✗68.268.083.084.4
SeeKer✗76.877.783.484.9
STG-NF✓71.873.985.987.4
SeeKer✓77.978.985.586.9
", + "image_path": "b0f7382c287cabd46586e15445e4e79b3c9ebe3a7bbe74e9e34892b1185f7415.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "On the keypoint ordering. SeeKer assumes a fixed ordering of skeleton keypoints. Ablations with different keypoint orderings incur a variance of less than 0.1, demonstrating SeeKer invariance to predefined keypoint order. Thus, we rely on the standard keypoint ordering [33] by default." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "On autoregressive model architectures. We analyse the performance of SeeKer depending on the specific model architecture used to predict parameters " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". In addition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "to the causally masked fully-connected model, we also test transformer architecture [51] that use causal self-attention layers. Each transformer token is a 2D skeleton keypoint location projected into a higher dimensional space. Further, we encode timestep and keypoint IDs to positional encoding of each token, as detailed in Appendix B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 145, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 242 + ], + "type": "text", + "content": "Table 8 shows that the masked fully connected model significantly outperforms its transformer-based counterpart. We attribute this gap to the low variability in skeleton sequences, as each person's gait differs only subtly. Additionally, skeleton sequences are much shorter than text paragraphs, and keypoint tokens have a much lower dimensionality, which influences the nature of the task and applicability of simpler models." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 254, + 550, + 295 + ], + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 604 + ], + "type": "text", + "content": "Table 7. Post-process smoothing only slightly boosts the performance of SeeKer. Importantly, SeeKer without smoothing outperforms STG-NF with smoothing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 254, + 550, + 295 + ], + "lines": [ + { + "bbox": [ + 319, + 254, + 550, + 295 + ], + "spans": [ + { + "bbox": [ + 319, + 254, + 550, + 295 + ], + "type": "table", + "html": "
ArchitectureUBnormalShanghaiTech
FullHRFullHR
Transformer decoder [51]71.171.679.180.0
Masked fully connected [21]77.978.985.586.9
", + "image_path": "6fa04c77717c4a51369905d54bfcdcf72439db4eb5087290e1f78e54397f7426.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 303, + 554, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 324 + ], + "type": "text", + "content": "Table 8. Comparison of the proposed masked fully-connected model and the causal transformer decoder." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 357, + 388, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 388, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 388, + 369 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "text", + "content": "We introduced SeeKer, a detector for anomalous human behaviour based on generative modeling of skeletal sequences. Unlike previous approaches of this kind, SeeKer considers each skeleton as a collection of keypoints and models sequence density as a product of Gaussian conditional keypoint densities. The parameters of the conditionals " + }, + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "inline_equation", + "content": "(\\mu, \\Sigma)" + }, + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "text", + "content": " are regressed by a causal deep model that ensures that each prediction can only observe the locations of the precedent keypoints. The proposed architecture infers all per-keypoint distributions with a single forward pass. We implement the corresponding causal architecture with masked fully connected layers. Interestingly, this formulation outperforms its counterpart based on a transformer decoder, probably due to the specifics of our data. During inference, we evaluate the per-keypoint distributions of the considered skeleton given the preceding keypoints in a sliding window, and summarize them into a skeleton-level anomaly score. Here, we take advantage of the composite structure of the proposed anomaly score by weighting the per-keypoint log-densities with the corresponding confidence values of the keypoint detector. The resulting anomaly score identifies abnormal behaviour by assessing the surprise of our model. Furthermore, our compositional anomaly score is highly interpretable since we can track keypoint contributions to the final decision. The conducted experiments indicate that SeeKer surpasses the state of the art on UBnormal and MSAD-HR by a wide margin, while delivering a close runner-up performance on ShanghaiTech." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11586" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 125 + ], + "type": "text", + "content": "This work has been supported by Croatian Recovery and Resilience Fund - NextGenerationEU (grant C1.4 R5-I2.01.0001). The authors declare no competing interests." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 166, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 166, + 296, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 296, + 188 + ], + "type": "text", + "content": "[1] Reprint of: Mahalanobis, p.c. (1936) “on the generalised distance in statistics.” Sankhya Ser. A, 2018. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 190, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 296, + 243 + ], + "type": "text", + "content": "[2] Davide Abati, Angelo Porrello, Simone Calderara, and Rita Cucchiara. Latent space autoregression for novelty detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 245, + 296, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 296, + 310 + ], + "type": "text", + "content": "[3] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, and Mubarak Shah. Abnormal: New benchmark for supervised open-set video anomaly detection. In Computer Vision and Pattern Recognition, 2022. 1, 4, 6, 13, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 312, + 296, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 296, + 355 + ], + "type": "text", + "content": "[4] Yoshua Bengio and Samy Bengio. Modeling high-dimensional discrete data with multi-layer neural networks. In Advances in Neural Information Processing Systems. MIT Press, 1999. 12" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 400 + ], + "type": "text", + "content": "[5] Maria A. Berezina, Daniel Rudoy, and Patrick J. Wolfe. Autoregressive modeling of voiced speech. In IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP, pages 5042-5045, 2010. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 292, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 292, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 292, + 413 + ], + "type": "text", + "content": "[6] Vladimir Braverman. Sliding Window Algorithms. 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 555 + ], + "type": "text", + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 3, 13" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 634 + ], + "type": "text", + "content": "[8] Antonio Bärbalău, Radu Tudor Ionescu, Mariana-Juliana Georgescu, Jacob Velling Dueholm, Bharathkumar Ramachandra, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas Baltzer Moeslund, and Mubarak Shah. Ssmtl++: Revisiting self-supervised multi-task learning for video anomaly detection. Comput. Vis. Image Underst., page 103656, 2022. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 679 + ], + "type": "text", + "content": "[9] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[10] Yucheng Chen, Yingli Tian, and Mingyi He. Monocular human pose estimation: A survey of deep learning-based methods. Comput. Vis. Image Underst., 192:102897, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 138 + ], + "type": "text", + "content": "[11] Romero F. A. B. de Morais, Vuong Le, Truyen Tran, Budhaditya Saha, Moussa Reda Mansour, and Svetha Venkatesh. Learning regularity in skeleton trajectories for anomaly detection in videos. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 205 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 207, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 207, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 207, + 553, + 261 + ], + "type": "text", + "content": "[13] Garoe Dorta, Sara Vicente, Lourdes Agapito, Neill D. F. Campbell, and Ivor Simpson. Structured uncertainty prediction networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "text", + "content": "[14] Fernando Pereira dos Santos, Leonardo Sampaio Ferraz Ribeiro, and Moacir A. Ponti. Generalization of feature embeddings transferred from different video anomaly detection domains. J. Vis. Commun. Image Represent., 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 363 + ], + "type": "text", + "content": "[15] Haoshu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 2, 4, 5, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 365, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 441 + ], + "type": "text", + "content": "[16] Alessandro Flaborea, Luca Collorone, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Bardh Prenkaj, and Fabio Galasso. Multimodal motion conditioned diffusion model for skeleton-based video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 5, 6, 8, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 443, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 553, + 497 + ], + "type": "text", + "content": "[17] Alessandro Flaborea, Guido Maria D'Amely di Melendugno, Stefano D'Arrigo, Marco Aurelio Sterpa, Alessio Sampieri, and Fabio Galasso. Contracting skeletal kinematics for human-related video anomaly detection. Pattern Recognit., 2024. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 520 + ], + "type": "text", + "content": "[18] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. ArXiv, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 577 + ], + "type": "text", + "content": "[19] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 643 + ], + "type": "text", + "content": "[20] Mariana-Iuliana Georgescu, Antonio Bärbalău, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Claudi Popescu, and Mubarak Shah. Anomaly detection in video via self-supervised and multi-task learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[21] Mathieu Germain, Karol Gregor, Iain Murray, and Hugo Larochelle. MADE: masked autoencoder for distribution estimation. In International Conference on Machine Learning, 2015. 3, 4, 5, 8, 12" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11587" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 733-742. IEEE Computer Society, 2016. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 149 + ], + "type": "text", + "content": "[23] Or Hirschorn and Shai Avidan. Normalizing flows for human pose anomaly detection. In International Conference on Computer Vision, 2023. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 295, + 172 + ], + "type": "text", + "content": "[24] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 1997. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 227 + ], + "type": "text", + "content": "[25] Radu Tudor Ionescu, Sorina Smeureanu, Bogdan Alexe, and Marius Popescu. Unmasking the abnormal events in video. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "type": "text", + "content": "[26] Yashswi Jain, Ashvini Kumar Sharma, Rajbabu Velmurugan, and Biplab Banerjee. Posecvae: Anomalous human activity detection. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 295, + 326 + ], + "type": "text", + "content": "[27] Asiegbu Miracle Kanu-Asiegbu, Ram Vasudevan, and Xiaoxiao Du. Bipoco: Bi-directional trajectory prediction with pose constraints for pedestrian anomaly detection. CoRR, abs/2207.02281, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[28] Alex M Lamb, Anirudh Goyal ALIAS PARTH GOYAL, Ying Zhang, Saizheng Zhang, Aaron C Courville, and Yoshua Bengio. Professor forcing: A new algorithm for training recurrent networks. In Advances in Neural Information Processing Systems, 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "type": "text", + "content": "[29] Gongjin Lan, Yu Wu, Fei Hu, and Qi Hao. Vision-based human pose estimation via deep learning: A survey. IEEE Trans. Hum. Mach. Syst., 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 448 + ], + "type": "text", + "content": "[30] Roberto Leyva, Victor Sanchez, and Chang-Tsun Li. Video anomaly detection with compact feature sets for online performance. IEEE Trans. Image Process., 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 492 + ], + "type": "text", + "content": "[31] Nanjun Li, Faliang Chang, and Chunsheng Liu. Human-related anomalous event detection via spatial-temporal graph convolutional autoencoder with embedded long short-term memory network. Neurocomputing, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 494, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 295, + 536 + ], + "type": "text", + "content": "[32] Yingcong Li, Yixiao Huang, Muhammed Emrullah Ildiz, Ankit Singh Rawat, and Samet Oymak. Mechanics of next token prediction with self-attention. In International Conference on Artificial Intelligence and Statistics, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 602 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 295, + 646 + ], + "type": "text", + "content": "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 4, 5, 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 712 + ], + "type": "text", + "content": "[35] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6536-6545. Computer Vision Foundation / IEEE Computer Society, 2018. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[36] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[37] Cewu Lu, Jianping Shi, and Jiaya Jia. Abnormal event detection at 150 FPS in MATLAB. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 5, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "text", + "content": "[38] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked RNN framework. In International Conference on Computer Vision, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 210, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 553, + 255 + ], + "type": "text", + "content": "[39] Weixin Luo, Wen Liu, and Shenghua Gao. Normal graph: Spatial temporal graph convolutional networks based prediction network for skeleton based video anomaly detection. Neurocomputing, 2020. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 289 + ], + "type": "text", + "content": "[40] Eran Malach. Auto-regressive next-token predictors are universal learners. In International Conference on Machine Learning, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 291, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 346 + ], + "type": "text", + "content": "[41] Amir Markovitz, Gilad Sharir, Itamar Friedman, Lihi Zelnik-Manor, and Shai Avidan. Graph embedded pose clustering for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 348, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 553, + 413 + ], + "type": "text", + "content": "[42] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. MULDE: multiscale log-density estimation via denoising score matching for video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "text", + "content": "[43] Xuan Mo, Vishal Monga, and Raja Bala. Simultaneous sparsity model for multi-perspective video anomaly detection. In International Conference on Image Processing, 2014. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 451, + 553, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 553, + 505 + ], + "type": "text", + "content": "[44] Bharathkumar Ramachandra and Michael J. Jones. Street scene: A new dataset and evaluation protocol for video anomaly detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 507, + 554, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 554, + 573 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 554, + 573 + ], + "type": "text", + "content": "[45] Royston Rodrigues, Neha Bhargava, Rajbabu Velmurugan, and Subhasis Chaudhuri. Multi-timescale trajectory prediction for abnormal human activity detection. In IEEE Winter Conference on Applications of Computer Vision, WACV 2020, Snowmass Village, CO, USA, March 1-5, 2020, 2020. 5, 6, 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 575, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 575, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 575, + 553, + 608 + ], + "type": "text", + "content": "[46] David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams. Learning representations by back-propagating errors. Nature, 1986. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 610, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 553, + 643 + ], + "type": "text", + "content": "[47] Benjamin Sapp, Alexander Toshev, and Ben Taskar. Cascaded models for articulated pose estimation. In European Conference on Computer Vision, 2010. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "type": "text", + "content": "[48] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Conference on Computer Vision and Pattern Recognition, 2018. 1, 2, 5, 15" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[49] Min Sun, Pushmeet Kohli, and Jamie Shotton. Conditional regression forests for human pose estimation. In Computer Vision and Pattern Recognition, 2012. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11588" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[50] Davoud Ataee Tarzanagh, Yingcong Li, Xuechen Zhang, and Samet Oymak. Max-margin token selection in attention mechanism. In Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 2, 4, 5, 8, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 207 + ], + "type": "text", + "content": "[52] Alexander Waibel, Toshiyuki Hanazawa, Geoffrey E. Hinton, Kiyohiro Shikano, and Kevin J. Lang. Phoneme recognition using time-delay neural networks. IEEE Trans. Acoust. Speech Signal Process., 1989. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 274 + ], + "type": "text", + "content": "[53] Guodong Wang, Yunhong Wang, Jie Qin, Dongming Zhang, Xiuguo Bao, and Di Huang. Video anomaly detection by solving decoupled spatio-temporal jigsaw puzzles. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part X, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "text", + "content": "[54] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. In British Machine Vision Conference 2018, BMVC 2018, Newcastle, UK, September 3-6, 2018, 2018. 5, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 375 + ], + "type": "text", + "content": "[55] Cheng Yan, Shiyu Zhang, Yang Liu, Guansong Pang, and Wenjun Wang. Feature prediction diffusion model for video anomaly detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 378, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 465 + ], + "type": "text", + "content": "[56] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 468, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 531 + ], + "type": "text", + "content": "[57] Bing Yu, Haoteng Yin, and Zhanxing Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJ-CAI 2018, July 13-19, 2018, Stockholm, Sweden, 2018. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 535, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 599 + ], + "type": "text", + "content": "[58] Guang Yu, Siqi Wang, Zhiping Cai, En Zhu, Chuanfu Xu, Jianping Yin, and Marius Kloft. Cloze test helps: Effective video anomaly detection via learning to complete video events. In MM '20: The 28th ACM International Conference on Multimedia, Virtual Event / Seattle, WA, USA, October 12-16, 2020. ACM, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 602, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 677 + ], + "type": "text", + "content": "[59] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segù, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14724-14734. IEEE, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Yongxin Yang, Andrea Cavallaro, and Tao Xiang. Omni-scale feature learning for person re-identification. In 2019 IEEE/CVF International Conference" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 184 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[61] Leyan Zhu, Zitian Wang, Si Liu, Xuecheng Nie, Luoqi Liu, and Bo Li. Multi-person pose regression with distribution-aware single-stage models. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[62] Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise review and a new dataset. Advances in Neural Information Processing Systems, 37:89943-89977, 2025. 5, 6, 14" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11589" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_content_list.json b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..360908cc442231c21154fb6c4c05126667be8548 --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_content_list.json @@ -0,0 +1,1543 @@ +[ + { + "type": "text", + "text": "Serialization based Point Cloud Oversegmentation", + "text_level": 1, + "bbox": [ + 243, + 130, + 754, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenghui Lu $^{1,2}$ , Jianlong Kwan $^{1,2}$ , Dilong Li $^{1,2*}$ , Ziyi Chen $^{1,2}$ , Haiyan Guan $^{3}$ $^{1}$ Department of Computer Science & Xiamen CVPR Key Laboratory, Huaqiao University \n $^{2}$ Fujian Key Lab. of Big Data Intell. and Security, Huaqiao University \n $^{3}$ Nanjing University of Information Science and Technology", + "bbox": [ + 143, + 179, + 851, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{23013083012, jianlongkwan}@stu.hqu.edu.cn, {scholar.dll, chenziyihq}@hqu.edu.cn", + "bbox": [ + 143, + 253, + 846, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3c467364161f46c290c1b5ef66f8c23f4391459826ef532d4ad5edaa5eafa3e.jpg", + "image_caption": [ + "Figure 1. Our method efficiently generates superpoints from serialized point clouds by optimizing boundaries through reallocating points from each segment to the most appropriate surrounding segments (shown in green and red)." + ], + "image_footnote": [], + "bbox": [ + 94, + 304, + 344, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85afddb3e8aa4b56c4856d653b2d0b8337afbc9152978dfd116f8e3ece568fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 304, + 506, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d05036b13904f0b04924e72e7f0bb618979d7aa84786014eac8c191494dbaaf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 304, + 622, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0157efd9819ccf3663eaae73f4541ffb220b738e3fd44c49a054840cd7378601.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 304, + 763, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df7619d1151017d0ae3a45d7782ec24fdbb52dd80514bf26339af008ab9110d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 304, + 901, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 467, + 326, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud oversegmentation, as a fundamental preprocessing step for 3D understanding, is a challenging task due to its spatial proximity and semantic similarity requirements. Most existing works struggle to efficiently group semantically consistent points into superpoints while maintaining spatial proximity. In this paper, we propose a novel serialization based point cloud oversegmentation method, which leverages serialization to avoid complex spatial queries, directly accessing neighboring points through sequence locality for similarity matching and superpoint clustering. Specifically, we first serialize point clouds into a Hilbert curve and spatially-continuously partition them into initial segments. Then, to guarantee the internal semantic consistency of superpoints, we design an adaptive update algorithm that clusters superpoints by matching feature similarities between neighboring segments and refines segment features via Cross-Attention. Experiments on large-scale indoor and outdoor datasets demonstrate state-of-the-art performance in point cloud oversegmentation. Moreover, it is also adaptable to semantic segmentation and achieves promising performance. The code is available at https://github.com/CHL-glitch/SPCNet.", + "bbox": [ + 88, + 491, + 485, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud oversegmentation partitions point clouds into multiple spatially and semantically homogeneous regions,", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "termed superpoints, which adaptively represent a set of points. Due to their representativeness, superpoint-based methods can significantly reduce computational complexity and improve semantic consistency. Therefore, it is becoming increasingly important in the field of point cloud processing and has attracted considerable attention from many researchers.", + "bbox": [ + 511, + 469, + 906, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early point cloud oversegmentation approaches primarily combined hand-crafted features with clustering or graph partitioning to generate superpoints. Papon et al. [26] introduced VCCS, which employed Fast Point Feature Histograms, selected seed points via octree structures, and performed K-means clustering. Lin et al. [22] reformulated the task as a subset selection problem while still using FPFH descriptors. Guinard et al. [12] simplified the process into a graph cut problem by extracting geometric features before applying a greedy algorithm. Similarly, Landrieu et al. [19] partitioned superpoints by solving an energy minimization problem with graph-based simplicity penalties. However, handcrafted features have limited representation, and the non-differentiable superpoint assignments prevent end-to-end training.", + "bbox": [ + 511, + 579, + 908, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent point cloud oversegmentation methods generate superpoints using deep networks. Landrieu et al. [18] first used deep networks to extract point embeddings instead of handcrafted features, but their framework is not end-to-end. SPNet [15] proposed an end-to-end superpoint network that achieved clustering by iteratively constructing a", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25831", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "point-superpoint association map. However, their Farthest Point Sampling (FPS) based k-nearest neighbor (KNN) association map faced efficiency bottlenecks, and the unlimited range of original points being clustered resulted in irregular superpoint boundaries and even introduced noise. SuperLiDAR [16] combined Breadth-First Search (BFS) grouping with local discriminative loss to generate compact superpoints, and while it achieved $O(V + E)$ complexity on sparse LiDAR data, it may approach $O(V^2)$ in dense scenarios, limiting scalability. Those point-based methods design complex algorithms to cluster superpoints and/or inefficiently balance semantic consistency and spatial proximity, resulting in high computational burden.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose an efficient serialization based point cloud oversegmentation method. Inspired by PTv3 [38], we adopt a serialization approach instead of traditional point-based methods to avoid high computational complexity spatial queries such as FPS and KNN, as shown in Fig. 1. We find that one of the representative serialization strategies, the Hilbert curve [14], which has excellent locality-preserving properties, is crucial for spatial proximity in oversegmentation tasks. Thus, we apply the Hilbert curve to serialize point clouds and initialize superpoints. To achieve internal semantic consistency for the superpoints, we design an adaptive superpoint update algorithm that clusters semantically similar points within prepartitioned segments to update superpoints. Specifically, we use cosine similarity to evaluate the semantic similarity between points and nearby segments, which serves as the metric for associating points and superpoints to reallocate points, then apply Cross-Attention to update superpoint features. By combining these two steps and iterating them, the algorithm can obtain superpoints with both spatial proximity and semantic consistency.", + "bbox": [ + 91, + 289, + 482, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to optimize superpoint feature representations, we employ a stacked Graph Transformer [42] and Graph Convolutional Network [42] to capture the global context between superpoints. Similar to prior work, SPCNet is designed as an end-to-end oversegmentation network and can also be integrated with downstream tasks such as semantic segmentation. To fully evaluate SPCNet, we conduct experiments on four large-scale datasets: S3DIS [1], ScanNet [9], nuScenes [3], and SemanticKITTI [2]. Experimental results demonstrate that SPCNet achieves state-of-the-art performance in all core oversegmentation metrics (BR, BP, and F1) across all datasets. The contributions of this paper are as follows:", + "bbox": [ + 89, + 608, + 482, + 803 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our best knowledge, SPCNet is the first serialization based network for point cloud oversegmentation.", + "- We introduce a serialization strategy to initialize superpoints, which can efficiently produce superpoints with spontaneous spatial proximity.", + "- We propose an adaptive superpoint update algorithm to" + ], + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "strengthen internal semantic consistency of superpoints.", + "bbox": [ + 527, + 90, + 895, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method achieves remarkable performance in point cloud oversegmentation across multiple datasets. It can be integrated with semantic segmentation networks and shows promising performance.", + "bbox": [ + 513, + 106, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 181, + 653, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Learning in Point Cloud", + "text_level": 1, + "bbox": [ + 513, + 208, + 779, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning approaches for 3D point cloud primarily follow three paradigms: projection-based, voxel-based, and point-based methods. Projection-based methods project unordered points onto 2D planes for CNN processing [6, 20, 21, 32], sacrificing 3D geometry for computational efficiency. Voxel-based methods discretize space into grids for 3D convolution [25, 31], improved by sparse convolution techniques [8, 10, 35] to mitigate memory costs, though kernel size limitations remain a constraint. Point-based methods [24, 27, 28, 33, 43] process raw point clouds directly and have evolved toward transformer architectures [13, 29, 37, 40, 44]. These methods offer powerful representation capabilities while struggling with scalability. The recently proposed SPT [29] leverages hierarchical superpoints and sparse self-attention to efficiently process large-scale scenes, achieving superior semantic segmentation with reduced model size and computational cost compared to traditional approaches.", + "bbox": [ + 511, + 229, + 906, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Point Cloud Oversegmentation", + "text_level": 1, + "bbox": [ + 513, + 513, + 785, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing point cloud oversegmentation methods can be broadly categorized into optimization-based methods and deep learning-based approaches. Early methods such as Papon et al. [26] proposed the VCCS method, which used FPFH as input features, selected seed points using an octree, and performed point cloud superpoint segmentation using the K-means clustering algorithm. Lin et al. [22] viewed the superpoint segmentation problem as a subset selection problem and employed subset selection methods to generate superpoints, still using FPFH as point cloud feature descriptors. Guinard et al. [12] simplified the superpoint segmentation problem into a graph cut problem, extracting features such as local linearity, flatness, divergence, and normal vectors from the point cloud, and applied a greedy graph cut algorithm to generate superpoints. Xiao et al. [39] defined the superpoint segmentation problem as an energy minimization problem and proposed a merging-exchange optimization framework to generate supervoxels. These methods often require complex and computationally expensive preprocessing steps, and their performance is limited by the insufficient expressiveness of handcrafted features. Recent deep learning methods overcome these limitations.", + "bbox": [ + 511, + 537, + 906, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Superpoint (SSP) method [18] pioneered deep network-based point cloud embedding, although it", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25832", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c2b7bdd95d3c04ae8a3346b5d94122776995e0d362401c86b53c9a4031e675c.jpg", + "image_caption": [ + "(a) Point Feature Extractor", + "(b) Hierarchical Segment Initialization", + "(c) Adaptive Superpoint Update", + "Figure 2. The architecture includes three stages. First, point embeddings are extracted using a U-Net backbone with sparse convolutions in shallow layers and a Mamba bottleneck enhanced by Conditional Position Encoding (CPE). Second, the point cloud is serialized via a Hilbert curve to initialize segments for superpoint aggregation, with granularity at each level controlled by the hyperparameter $m$ . Finally, points are reassigned across segments to group semantically similar points into coherent superpoints, completing the clustering process." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 903, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "still depended on optimization techniques from [12]. SP-Net [15] further advanced the field by introducing learnable point-to-superpoint associations using FPS for center selection, enabling joint optimization in both coordinate and feature spaces. Furthermore, SuperLiDAR [16] improved efficiency by combining a BFS-based grouping algorithm with a local discriminative loss to generate compact superpoints. However, this method results in high computational complexity in spatial query operations during superpoint generation, especially for dense point cloud oversegmentation.", + "bbox": [ + 88, + 435, + 482, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Localization-based Method", + "text_level": 1, + "bbox": [ + 89, + 595, + 338, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works [5, 23, 34, 38] have introduced serialization-based approaches that differ from traditional point cloud processing by transforming unstructured point clouds into ordered sequences to preserve spatial proximity. For example, OctFormer [34] uses octree-based ordering similar to $z$ -ordering, offering scalability but limited by octree constraints. FlatFormer [23] employs window-based sorting to group point pillars, though its scalability in receptive field is limited, making it suitable for pillar-based 3D object detection. Building on these methods, PTv3 [38] extends attention scales by replacing KNN-based query with serialized neighborhoods, enhancing efficiency and scalability.", + "bbox": [ + 89, + 618, + 483, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 814, + 181, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our oversegmentation network comprises three key steps: point embedding extraction (Sec. 3.1), hierarchical segment initialization (Sec. 3.2), and adaptive superpoint update (Sec. 3.3). The overall process is shown in Fig. 2.", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Backbone", + "text_level": 1, + "bbox": [ + 513, + 434, + 625, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization-based oversegmentation methods are constrained by the quality of handcrafted features, while learning based approaches similarly encounter limitations in deep feature representation.", + "bbox": [ + 511, + 457, + 906, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse 3D U-Net. Let the input point cloud be $\\mathcal{P} \\in \\mathbb{R}^{N \\times (C + 3)}$ , where each of the $N$ points is represented by its features (e.g. normal vectors, RGB) and spatial coordinates $(x, y, z)$ . Following [10], we first voxelize $\\mathcal{P}$ using the operator $\\mathcal{V}$ , and then process the discretized data with a U-Net architecture based on sparse convolutions (SpConv). In the shallow part of the U-Net [30], SpConv layers capture fine-grained local geometric information. These operations produce feature representations that maintain high spatial resolution and extract detailed geometric structures crucial for accurate object boundary delineation.", + "bbox": [ + 511, + 518, + 908, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bottleneck Reinforcement. Pure SpConv restricts over-segmentation effectiveness due to their inherently limited receptive field. We enhance our U-Net backbone by integrating Mamba [11, 45] at the bottleneck of the U-Net, as shown in Fig. 2 (a). After the shallow SpConv layers have abstracted local features, the Mamba module is applied to perform global context interaction. Mamba, with its linear-complexity global modeling capability, compensates for the local inductive bias of convolution and promotes semantic consistency across superpoints.", + "bbox": [ + 511, + 685, + 908, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Segment Initialization", + "text_level": 1, + "bbox": [ + 511, + 847, + 820, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methodology Justification. To efficiently partition point clouds into multiple segments, we propose a serialization-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25833", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based partitioning method that utilizes the spatial locality preservation property of the Hilbert curve [38, 41] to divide the same point cloud into hierarchical structures of different granularities. Specifically, the Hilbert curve maps adjacent point clusters in 3D Euclidean space to contiguous positions in a 1D sequence, thereby simplifying the spatial proximity-based 3D partitioning task into an iterative partitioning of a 1D sequence. This transformation enables us to efficiently divide point clouds into compact segments in 3D space, significantly enhancing partitioning efficiency. This enables (1) efficient 1D clustering that avoids expensive 3D queries and clustering; and (2) a simplified oversegmentation process that decouples spatial locality from semantic similarity by performing only semantic comparisons within already spatially adjacent segments to complete clustering.", + "bbox": [ + 88, + 90, + 482, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Partitioning Strategy. Due to the variable receptive field of differently sized superpoints, smaller superpoints preserve more detailed local geometric information, whereas larger superpoints capture more comprehensive global semantic context. Therefore, we propose to construct a hierarchical partitioning structure to fully utilize multi-scale information. Given an input point cloud $\\mathcal{P}$ and its corresponding features $F = \\{f_i\\}_{i=1}^N$ , where $f_i \\in \\mathbb{R}^d$ is generated by our backbone, we construct a hierarchical partitioning structure $\\mathcal{H} = \\{L_k\\}_{k=1}^K$ . Each layer $L_k$ contains the complete serialized point cloud and is recursively quaternary partitioned into a set of segments $L_k = \\{S_i^k\\}_{i=0}^{M_k-1}$ where $S_i^k$ represents the $i$ -th segment, as shown in Fig. 2 (b). Specifically, the 3D point cloud is first serialized into a 1D sequence via Hilbert space-filling curve mapping [38], denoted as $\\mathcal{P}_{seq} \\in \\mathbb{R}^{N \\times d}$ . The sequence is then padded to $\\mathcal{P}_{pad} \\in \\mathbb{R}^{N_{pad} \\times d}$ where $N_{pad} = \\lceil N / M_K \\rceil \\cdot M_K$ , ensuring divisibility by the finest segment count $M_K$ . The serialized point cloud $\\mathcal{P}_{pad}$ is recursively partitioned using 1D quaternary splitting, such that at each layer $L_k$ , it is uniformly divided into $M_k = 2^{m - 2k + 2}$ segments. For each layer $L_k$ , $\\mathcal{S}_i^k$ can be represented as:", + "bbox": [ + 93, + 324, + 485, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} ^ {k} = \\left\\{\\mathcal {P} _ {p a d} [ j ] \\mid j \\in [ i \\cdot l _ {k}, (i + 1) \\cdot l _ {k} - 1 ] \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 662, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l_{k} = N_{pad} / M_{k}$ represents the segment length for each segment in the $k$ -th layer. Recursive partitioning establishes fixed parent-child indexing between layers. For any inter-layer gap $l \\geq 1$ , the hierarchical inclusion relationship is defined as:", + "bbox": [ + 89, + 686, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} ^ {k} = \\bigcup_ {j \\in [ i \\cdot 4 ^ {l}, (i + 1) \\cdot 4 ^ {l} - 1 ]} \\mathcal {S} _ {j} ^ {k + l}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 770, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the finest layer $L_{K}$ , segment features are generated through permutation-invariant aggregation of constituent points:", + "bbox": [ + 89, + 801, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {S _ {i} ^ {K}} = \\mathcal {A} \\left(f _ {j} \\mid p _ {j} \\in S _ {i} ^ {K}\\right) \\in \\mathbb {R} ^ {C}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 845, + 480, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{A}:\\mathbb{R}^{n_i\\times C}\\to \\mathbb{R}^C$ denotes a permutation-invariant operator (e.g. max-pooling) with $n_i = |\\mathcal{S}_i^K | = N_{pad} / M_K$ .", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Superpoint Update", + "text_level": 1, + "bbox": [ + 511, + 90, + 769, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the structure $\\mathcal{H}$ , the goal is to refine the partition $L_{k} = \\{S_{1}^{k},\\ldots ,S_{M_{k}}^{k}\\}$ with features $f_{S_i^k}\\in \\mathbb{R}^d$ into semantically consistent superpoints $L_k^\\prime = \\{S_1'^k,\\dots,S_M'^k\\}$ with precise geometric boundaries and updated features $f_{S_i^k}^\\prime \\in \\mathbb{R}^d$ . This is achieved by dynamically allocating raw points within coarse-grained segments to their appropriate sub-segments based on semantic similarity, a process that redistributes raw points within segments to fit object boundaries. This process can be transformed into a boundary fitting problem between adjacent segments in 1D sequence point clouds, as shown in Fig. 2 (c).", + "bbox": [ + 511, + 112, + 903, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Similarity based Affiliation. To establish the affiliation relationship between raw points within each segment and nearby segments, we consider two adjacent levels in $\\mathcal{H}$ : the finest $L_{k} = \\{S_{1}^{k}, S_{2}^{k}, \\ldots, S_{M_{k}}^{k}\\}$ and its coarser predecessor $L_{k-1} = \\{S_{1}^{k-1}, S_{2}^{k-1}, \\ldots, S_{M_{k-1}}^{k-1}\\}$ (we only use $l = 1$ defined in Eq. (2)). The hierarchical relationship between these levels is given by (1) $M_{k-1} = \\frac{M_{k}}{4}$ , meaning each coarser level has one-fourth the segments of the finer level and (2) Each segment $S_{i}^{k-1}$ at level $L_{k-1}$ contains four consecutive segments from level $L_{k}$ : $S_{i}^{k-1} = S_{4i}^{k} \\cup S_{4i+1}^{k} \\cup S_{4i+2}^{k} \\cup S_{4i+3}^{k}$ .", + "bbox": [ + 511, + 291, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each raw point $p_i \\in S_i^{k-1}$ , we identify the most similar initialized segment feature from its four containing segments $S_{4i}^k, S_{4i+1}^k, S_{4i+2}^k$ , and $S_{4i+3}^k$ . Meanwhile, to expand the receptive field of raw points relative to their potential superpoints, we incorporate segment features from eight sub-segments at $L_k$ belonging to the left and right adjacent segments $S_{i-1}^{k-1}$ and $S_{i+1}^{k-1}$ , forming a local feature set $\\{f_{S_{4i+j}^k} \\mid j \\in \\{-4, -3, \\dots, 7\\}\\}$ , resulting in a set of 12 initial segment features. For each segment $S_i^{k-1}$ with $N_{k-1} = 4N_{pad} / M_k$ points, we compute a similarity matrix $\\mathbf{A}_i \\in \\mathbb{R}^{N_{k-1} \\times 12}$ . Each entry $a_{j,t}$ measures the similarity between the $j$ -th point and the $t$ -th segment feature in the above set of segment features:", + "bbox": [ + 511, + 460, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {j, t} = \\frac {f _ {p _ {j}} \\cdot f _ {S _ {t} ^ {k}}}{\\| f _ {p _ {j}} \\| \\| f _ {S _ {t} ^ {k}} \\|}, \\quad j \\in [ 1, N _ {k - 1} ], \\quad t \\in [ 1, \\tau ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 667, + 903, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This process generates similarity matrices $\\{\\mathbf{A}_i\\}_{i=1}^{M_{k-1}}$ for the entire level $L_{k-1}$ . Point-to-superpoint assignments are determined by selecting the superpoint with the highest similarity for each point, which is then mapped to global superpoint indices $\\phi(p_j) \\in [1, M_k]$ . The resulting assignment index set $Ass_i = \\{\\phi(p_j)\\}_{j=1}^n$ provides a key reference for updating superpoint features in subsequent steps.", + "bbox": [ + 511, + 710, + 903, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Attention based Update. Cross-Attention [4] enables dynamic and adaptive relationships between superpoints and their constituent points. Through feature similarity based affiliation, precise similarity relationships are established, guiding accurate superpoint feature updates. For", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25834", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0db175f23efedc3fe081828f140afd4176b41311856bb7bb2c483578fd4f6e5b.jpg", + "image_caption": [ + "Figure 3. The end-to-end semantic segmentation pipeline constructs a graph over superpoints, captures global context using graph transformers, and integrates point-level embeddings for final semantic prediction, where $\\oplus$ denotes simple feature addition." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the assignment indices $A s s i = \\{\\phi(p_{j})\\}_{j=1}^{n}$ , the inverse mapping $\\phi^{-1}(i) = \\{p_{j} \\mid \\phi(p_{j}) = i\\}$ represents the original points contained within superpoint $S_{i}^{\\prime k}$ . Based on $A s s i$ , each superpoint's feature is updated solely from its member points. Using superpoint feature $f_{S_{i}^{\\prime k}}^{\\prime}$ as the Query and its member point features $\\{f_{p_{j}} \\mid p_{j} \\in \\phi^{-1}(i)\\}$ as Keys and Values, the update formula is:", + "bbox": [ + 89, + 369, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {S _ {i} ^ {k}} ^ {\\prime} = \\text {C r o s s A t t n} \\left(\\underbrace {f _ {S _ {i} ^ {k}}} _ {\\text {Q u e r y}}, \\underbrace {\\left\\{f _ {p _ {j}} \\mid p _ {j} \\in \\phi^ {- 1} (i) \\right\\}} _ {\\text {K e y / V a l u e}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 493, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention weights are computed through scaled dot product, allowing each superpoint to focus on the most informative points within its local neighborhood:", + "bbox": [ + 89, + 564, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C r o s s A t t n} \\left(f _ {S _ {i} ^ {k}}, \\left\\{f _ {p _ {j}} \\right\\}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 618, + 482, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_i = f_{S_i^k}W_K\\in \\mathbb{R}^{1\\times d}$ , $\\mathbf{K}_i = \\mathrm{Concat}(\\{f_{p_j}W_K\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}$ , $\\mathbf{V}_i = \\mathrm{Concat}(\\{f_{p_j}W_V\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}$ , $n_i^k = |\\phi^{-1}(i)|$ is the number of points contained in each superpoint, and $W_{Q},W_{K},W_{V}\\in \\mathbb{R}^{d\\times d}$ are learnable projection matrices. For coarse-grained superpoints, we apply the same update logic, with interactions between fine-grained features and their coarse-grained initializations.", + "bbox": [ + 89, + 662, + 483, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Superpoint Interaction", + "text_level": 1, + "bbox": [ + 89, + 800, + 303, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The superpoint features $f_{S_i^k}^\\prime$ encode geometric-semantic information for locally homogeneous points. To leverage these features for semantic segmentation, inspired by SPT [29], we integrate a superpoint interaction module with our oversegmentation network, forming an end-to-end", + "bbox": [ + 89, + 821, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "framework, as shown in Fig. 3. In our hierarchical design, fine-grained superpoints capture local details while coarse grained ones represent object-level semantics. We construct multi-level superpoint graphs and employ a cascaded Transformer-GCN [42] to facilitate cross-superpoint interactions and global context modeling. In the Graph Transformer, a superpoint graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ is built by linking each superpoint with its $k$ nearest neighbors in feature space, where $\\mathcal{V} = \\{f_{S_i^k}'\\}_{i=1}^{M_k}$ . Edge features are defined via relative position encoding as $\\mathbf{e}_{ij} = g(p_j - p_i)$ , with $p_i$ and $p_j$ representing node coordinates and $g$ being a linear mapping. A subsequent GCN processes this graph to enhance local feature interactions through message passing among adjacent superpoints. Finally, we fuse backbone point features with their corresponding multi-level superpoint before feeding them into the segmentation head.", + "bbox": [ + 511, + 90, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Superpoint Aggregation Loss", + "text_level": 1, + "bbox": [ + 511, + 344, + 774, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a superpoint aggregation loss to optimize superpoint generation, inspired by LMNN [36]. This comprehensive loss consists of three components. First, a compactness loss $\\ell_{\\mathrm{compact}}$ enforces intra-superpoint feature cohesion by minimizing point-to-superpoint distances. Second, a distinction loss $\\ell_{\\mathrm{dist}}$ encourages inter-class separability by maximizing the distances between superpoints of different classes, where each superpoint's label is determined by majority voting among its points. Third, a purity loss $\\ell_{\\mathrm{purity}}$ maintains label consistency within each superpoint by penalizing high entropy in its class distribution. The overall loss function is defined as:", + "bbox": [ + 511, + 366, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\ell_ {s p} = \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\frac {\\sum_ {p \\in S _ {i}} \\| f _ {i} - f _ {S _ {i} ^ {k}} ^ {\\prime} \\| ^ {2}}{| S _ {i} |} + \\\\ \\frac {1}{N _ {c}} \\sum_ {i, j \\in c, i \\neq j} \\max \\left(0, \\delta_ {\\text {d i s t}} - \\| f _ {S _ {i} ^ {k}} ^ {\\prime} - f _ {S _ {j} ^ {k}} ^ {\\prime} \\|\\right) ^ {2} + \\tag {7} \\\\ \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {k = 1} ^ {K} p _ {i, k} \\log (p _ {i, k}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 553, + 906, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{i}$ denotes the feature of point $p$ , $f_{S_i^k}^\\prime$ represents the feature of superpoint $S_{i}^{\\prime k}$ , $\\delta_{\\mathrm{dist}}$ is a predefined separation margin, $p_{i,k}$ is the probability of superpoint $i$ belonging to class $k$ , $N_{s}$ is the number of superpoints, and $N_{c}$ is the number of superpoint pairs involved in $\\ell_{\\mathrm{dist}}$ .", + "bbox": [ + 511, + 689, + 905, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 784, + 645, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 511, + 816, + 712, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate SPCNet on four datasets with diverse characteristics. S3DIS [1] contains $274M$ points from six office areas, with Area 5 as the test set. ScanNet [9] includes 1,513 indoor scans from 707 scenes with", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68583646c2c86e7d1defb0fe1e6c589c815ca02d12e9b76b343157e2cfa2954a.jpg", + "image_caption": [ + "Figure 4. The visualization of superpoints generated on the ScanNet (first row) and S3DIS (second row) datasets shows a two-level hierarchy: $L_{1}$ with 1024 superpoints and $L_{2}$ with 256 superpoints. Finer $L_{1}$ superpoints merge based on semantic similarity to form larger $L_{2}$ superpoints. For example, in S3DIS, $L_{1}$ superpoints of table parts merge into a single table superpoint, and $L_{1}$ superpoints of whiteboard sections form a complete whiteboard superpoint. Our superpoints are regular and orderly, with nearly rectangular shapes in semantically consistent areas (e.g. floors, tables), maintaining continuity, precise object boundaries, and demonstrating superior regularity, compactness, and purity." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "20 categories, averaging 148k points/scan. nuScenes [3] comprises 1,000 scenes from 32-beam LiDAR, split 750/150/150 for train/val/test, with 17 categories. SemanticKITTI [2] is an outdoor self-driving benchmark with 22 sequences and 20 categories, using sequences 00-10 (densely annotated) for training and 11-22 for testing.", + "bbox": [ + 88, + 381, + 480, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following SPNet [15] and SSP [18], we use Oracle Overall Accuracy (OOA), Boundary Recall (BR), Boundary Precision (BP), and F1 score to evaluate superpoints. BR and BP assess boundary quality, while OOA represents the upper bound of semantic segmentation accuracy using superpoints. The F1 score balances recall and precision, defined as $F1 = \\frac{2\\cdot BP\\cdot BR}{BP + BR}$ . For semantic segmentation in indoor and outdoor scenarios, we adopt mean intersection over union (mIoU).", + "bbox": [ + 88, + 472, + 482, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Model Configuration", + "text_level": 1, + "bbox": [ + 89, + 619, + 290, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We maintain a consistent architecture across all experiments with minimal dataset-specific adjustments. Our backbone features a four-stage symmetric encoder-decoder structure, each with a single block depth. The encoder has embedding dimensions [32, 64, 128, 256] and the decoder [128, 64, 64], with Mamba blocks using conditional positional encoding at the 256 stage. We implement a three-level hierarchy $(m = 8, K = 2, l = 1)$ , organizing points into n, 1024, 256 segments. Training is performed on dual NVIDIA RTX 4090 GPUs with the Adam optimizer.", + "bbox": [ + 88, + 641, + 482, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Point Cloud Oversegmentation", + "text_level": 1, + "bbox": [ + 89, + 801, + 364, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Tab. 1, we evaluate SPCNet on three datasets, reserving ScanNet for ablation studies, ensuring fair comparison with consistent superpoint counts. For example, in S3DIS Area 5, methods like VCCS, Lin et al., SPG, SSP, and SPNet generate around 1050 su", + "bbox": [ + 88, + 824, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a918902f1be89a6699ae439b17f03fc6ab984fbc29b3a8042936b2b9caa23b3a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBRBPF1OOA
S3DIS Area 5 [1]
VCCS [26]62.0610.8618.4895.22
SPG [19]52.0613.1120.9495.84
SSP [18]80.7313.0222.4297.04
SPNet [15]84.7513.1422.6596.50
SPCNet(Ours)89.5026.7840.5096.81
nuScenes [3]
SPG [19]28.2017.2621.4189.22
SSP [18]22.0415.6318.2892.01
SPNet [15]68.9218.3428.9787.67
SuperLiDAR [16]74.7225.1237.5996.31
SPCNet(Ours)88.7732.8247.5296.33
SemanticKITTI [2]
SPG [19]25.6415.8219.5686.86
SSP [18]18.7410.6713.5992.27
SPNet [15]56.5214.7823.4392.24
SuperLiDAR [16]65.5220.5231.2596.21
SPCNet(Ours)73.1830.1542.5796.84
", + "bbox": [ + 526, + 378, + 893, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison results of generated superpoints on the S3DIS, nuScenes, and SemanticKITTI datasets.", + "bbox": [ + 511, + 742, + 906, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perpoints, matching our 1024 Hilbert partitions. Our deep learning-based approach outperforms traditional methods, benefiting from the specialized backbone network's ability to extract rich geometric features that capture complex local structures. Unlike methods such as SPNet [15], which rely on FPS+KNN for presampling and soft assignment map-", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25836", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ping without limiting the scope of raw points being clustered, resulting in blurry boundaries, our approach leverages the Hilbert curve's locality, limiting the range of raw point clustering (preventing clustering of semantically similar but spatially distant points), and finally refining superpoint through the Cross-Attention, minimizing noise and maintaining precise, simple boundaries, as shown in Fig. 5 (d) and (e). Visualization results in Fig. 4 show that our method tends to aggregate superpoints within object interiors while maintaining distinct object boundaries, especially in challenging cases like wall-whiteboard interfaces, resulting in higher BR and BP. Additionally, our method excels at merging semantically continuous regions (e.g. floors, tabletops), leading to fewer incorrect superpoint boundaries and substantially higher BP metrics. Fig. 6 shows the performance of different methods on SemanticKITTI.", + "bbox": [ + 89, + 90, + 483, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4136bf7b35d0725ded018c521b9631c6c78ffadba11ac19071c062bccd725c65.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. The inference time of different oversegmentation methods on SemanticKITTI. The inference time consists of data preprocessing/backbone (red/blue), superpoint initialization (green), and superpoint generation (purple)." + ], + "table_body": "
MethodDeviceInference Time (ms) data proc. / backbone + SP ini. + SP gen.
SPG [19]CPU13152
SSP [18]RTX 309016115 (11812 + 0 + 13426)
SPNet [15]RTX 309012182 (11812 + 0 + 1154)
SuperLiDAR [16]RTX 309072(0 + 0 + 72)
SPCNet (ours)RTX 4090, RTX 306012 (6 + 0.4 + 6), 61
", + "bbox": [ + 91, + 342, + 482, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Costs. We evaluated oversegmentation network efficiency by measuring inference time per scan on SemanticKITTI validation (batch size=1) in Tab. 2. For optimization based SPG [19], we followed SuperLiDAR's [16] configuration using Core i5 CPU. For learning-based methods (SSP [18], SPNet [15], SuperLiDAR), we used SuperLiDAR's results (RTX 3090). Our SPCNet was benchmarked on RTX 4090 and 3060 GPUs. Our method achieved total inference times of 12 ms (RTX 4090) and 61 ms (RTX 3060), achieving $100 \\times$ speedup versus SPG/SSP/SPNet, and $5 \\times$ versus SuperLiDAR. Even on RTX3060, our inference is 10ms faster than SuperLiDAR, indicating at least $3 \\times$ efficiency accounting for hardware differences. This stems from: (1) efficient superpoint initialization (0.4 ms) that compresses FPS+KNN sampling time in SPNet; (2) the application of a localized attention mechanism within the superpoint refinement (6 ms). Our backbone architecture allows further acceleration using lightweight alternatives like PointNet[27]. Traditional methods face key bottlenecks: SPG/SSP rely on optimization-based generation (e.g. graph cuts), SSP/SPNet suffer from redundant processing due to mixed handcrafted/learned features, and SuperLiDAR's BFS-based grouping has complexity dependent on scene density, potentially reaching $O(n^{2})$ in dense scenes, versus our serialization based approach $(O(n))$ .", + "bbox": [ + 89, + 503, + 483, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Studies. Mamba mitigates the local inductive bias", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c95b7aa84caa670107ecd0e17bf52de6eeb5ec47fad76ad32dfd6f262f22ce0.jpg", + "image_caption": [ + "(a) 3D-UNet only with $k = 2$" + ], + "image_footnote": [], + "bbox": [ + 517, + 94, + 614, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f89d9e30935e1b0368a2e56f53d8e28ffdfc4fcad1d6f2ecff826a3e30b4715d.jpg", + "image_caption": [ + "(b) Bottleneck reinforcement with $k = 2$" + ], + "image_footnote": [], + "bbox": [ + 616, + 94, + 712, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8f9808ff53a2c30e00548eac0e07af490314ce6ba87be1c2ee4af92c4efca42.jpg", + "image_caption": [ + "(d) PointNet with $k = 2$", + "Figure 5. Oversegmentation comparisons. (a) Sparse 3D U-Net with neighbor expansion of 2. (b) Adding the Mamba module to the bottleneck. (c) Replacing the backbone with a simplified PointNet, as in SPNet. (d) SPNet results. Notably, while (b) configuration does not fully aggregate semantically consistent objects into complete superpoints, it notably enhances consistency. SPC-Net also avoids clustering semantically similar but spatially distant points, ensuring clearer superpoint boundaries." + ], + "image_footnote": [], + "bbox": [ + 714, + 94, + 803, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28a5561d25924c26d780d6c78c4fe37fc4e4b55d5c8510c0150c8779447fba24.jpg", + "image_caption": [ + "(e) SPNet" + ], + "image_footnote": [], + "bbox": [ + 805, + 94, + 901, + 166 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80297249ba28dbd24c35b5a30214dfff8658d4652eb49da1a076300473c7711f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Performance comparison of different backbone architectures on the ScanNet validation set." + ], + "table_body": "
Backbone SettingBR(%)BP(%)OOA(%)mIou(%)
3D U-Net76.33 ↓7.9028.45 ↓1.7090.93 ↓1.1974.10 ↓2.86
+ Mamba fully78.37 ↓5.8629.13 ↓1.0291.24 ↓0.8876.18 ↓0.78
+ Mamba bottleneck (default)84.2330.1592.1276.96
", + "bbox": [ + 514, + 334, + 903, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6240a39bcc8b9ae4536efffe9a3d404efd8abe43e535473480ffa4fecbdf036f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. The effect of expanding neighbor segment range on superpoint performance on the ScanNet validation set." + ], + "table_body": "
Neighbor ScopeBR(%)BP(%)OOA(%)
No expansion79.86 ↓4.3728.57 ↓1.5891.22 ↓0.90
2 expansion (default)84.2330.1592.12
4 expansion83.67 ↓0.5629.85 ↓0.3091.24 ↓0.88
6 expansion79.93 ↓4.3028.65 ↓1.5089.93 ↓2.19
", + "bbox": [ + 514, + 445, + 901, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of convolutions in oversegmentation, improving superpoint continuity and semantic consistency. As shown in Fig. 5 (a) and (b), this bias causes superpoints in the same semantic region to align with boundaries but fragment due to internal inconsistency. Thus, we retain convolutions in U-Net's shallow layers for local geometric cues to ensure boundary precision, and introduce Mamba at the bottleneck to capture long range semantics. Tables 3 and 4 analyze backbone design and neighbor expansion effects on superpoint quality. We evaluated three backbone configurations: (1) baseline 3D sparse U-Net only, (2) fully Mamba-based U-Net, and (3) our default hybrid design with Mamba modules in the bottleneck layers. Results demonstrate that the baseline 3D sparse U-Net reduces BR, BP, and OOA by $7.90\\%$ , $1.70\\%$ , and $1.19\\%$ respectively compared to our default. Similarly, the fully Mamba-based U-Net underperforms, as Mamba's global context modeling in shallow networks with low-dimensional features smooths out high-frequency boundary information. Our hybrid \"shallow convolution + deep Mamba\" architecture achieves optimal per", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25837", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8de302899f4f95dfd852cc00ec70e2f91c49f12bc3ceeae25552303e13887394.jpg", + "image_caption": [ + "Figure 6. Performance of different methods on the SemanticKITTI validation set under varying numbers of superpoints. It is worth noting that our method can achieve better results than other methods with fewer superpoints (256, 512, and 1024)." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 297, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bafd1e0708c5c25e62a082e0384fdac660e6d5ea006be9e022455b37d4fc07b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 92, + 496, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a03200b16ee743058f2e9b76a44c63aa914a8d384d7a320257d3e5de66549db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 700, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1836726bac8ed4d909e7809700771b8b5243b4cb9e1cf101822911af447133d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 90, + 901, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0965269f4f330e05a171c42559a4516a34a592096cb9a23840584889b134f8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ExperimentScanNet [9]S3DIS [1]nuScenes [3]Sem.KITTI [2]
Best ModelValTestArea5ValTestValTest
†SPG [19]--58.0----
MinkUNet [8]72.273.665.473.3-63.8-
†SPG + SSP [19]--61.7----
ST [17]74.373.772.0----
†SPT [29]--68.9----
PointNeXt [28]71.571.270.5----
OctFormer [34]75.776.6-76.177.264.367.8
Swin3D [40]76.4-72.5----
AF2S3Net [7]---62.278.074.270.8
†SuperLiDAR [16]----78.5-69.6
PTv2 [37]75.474.271.680.282.670.372.6
PTv3 [38]77.577.973.480.482.770.874.2
Backbone76.9-69.375.4-67.6-
†SPCNet(Ours)77.878.072.177.680.269.571.9
", + "bbox": [ + 91, + 270, + 483, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance by leveraging convolutions to preserve local details in shallow layers while Mamba modules handle global context modeling in deeper layers.", + "bbox": [ + 89, + 522, + 482, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the neighborhood expansion experiments (analogous to k-parameter tuning in KNN), we observed that moderately expanding the candidate superpoint pool of original points during the feature similarity based affiliation process significantly improves sampling accuracy. The model achieves optimal performance when $\\mathrm{K} = 2$ ; excessive expansion $(\\mathrm{K} = 4,6)$ leads to performance degradation. This may be because larger receptive fields introduce noise, and the hard assignment approach introduces more uncertainty as the number of candidate targets increases, causing early assignment errors to propagate to later processing stages.", + "bbox": [ + 88, + 568, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 742, + 307, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 5 evaluates our approach on four indoor and outdoor datasets, comparing it with point-based and superpoint-based methods. Our optimized oversegmentation strategy shows significant gains: on S3DIS Area 5, we achieve a $3.2\\%$ mIoU improvement over SPT. Compared to state-of-the-art point-based methods, our approach remains competitive, outperforming PTv3 by $0.3\\%$ on ScanNet $(77.8\\%)$ and PTv2 by $0.5\\%$ on S3DIS Area 5. Additionally, we achieve $71.9\\%$ and $80.2\\%$ on Sem.KITTI and nuScenes, improving", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/766e54e96e3c1d3c71a41f7b418a6ae5d5e80532ee0d85943dd887067ae1332e.jpg", + "table_caption": [ + "Table 5. Semantic segmentation results on the ScanNet, S3DIS, nuScenes, and SemanticKITTI datasets. †Superpoint-based." + ], + "table_footnote": [], + "table_body": "
SettingComponentsModel Size (m)mIoU(%)Δ(%)
Backbone + SP gen. + SP int.
backbone (default)-18.876.96-
Hierarchical Enhancement
1 partition level+ Level1 Superpoint0.24 (0.218 + 0.022)77.50+0.54
2 partition level+ Level2 Superpoint0.02 (0.014 + 0.006)77.80+0.30
", + "bbox": [ + 514, + 270, + 903, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Semantic segmentation improvement with hierarchical superpoint features in ScanNet validation set. SP gen. means superpoint generation; SP int. means superpoint interaction.", + "bbox": [ + 511, + 351, + 906, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SuperLiDAR by $1.3\\%$ and $1.7\\%$ , respectively. These results confirm that our superpoints capture local and global semantic information effectively, while the simple backbone, enhanced by superpoint features, matches the performance of more complex point-based methods, validating the superior quality of the generated superpoints.", + "bbox": [ + 511, + 421, + 903, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 6, integrating hierarchical superpoint features with per-point features extracted by the backbone improved segmentation. First-level superpoints aggregated local semantics, boosting intra-class consistency with a $0.54\\%$ mIoU gain. Second-level features captured long-range dependencies, adding another $0.30\\%$ improvement. These hierarchical superpoints acted as adaptive receptive fields, enhancing performance with only $0.24M$ parameters.", + "bbox": [ + 511, + 512, + 906, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel serialization based oversegmentation method to address the challenge of efficiently grouping semantically consistent points into superpoints while maintaining spatial proximity. SPCNet achieves state-of-the-art performance in point cloud oversegmentation while being $3 \\times$ faster than existing methods in inference speed. Extensive experimental results demonstrate the effectiveness and efficiency of SPCNet. Additionally, SPCNet can be flexibly applied to semantic segmentation tasks, effectively improving the accuracy of semantic segmentation.", + "bbox": [ + 511, + 688, + 903, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported in part by the National Natural Science Foundation of China (Grant No. 42201475).", + "bbox": [ + 511, + 854, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25838", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 8", + "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2, 6, 8", + "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 8", + "[4] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4", + "[5] Wanli Chen, Xinge Zhu, Guojin Chen, and Bei Yu. Efficient point cloud analysis using hilbert curve. In European Conference on Computer Vision, pages 730-747. Springer, 2022. 3", + "[6] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2", + "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 8", + "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 8", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 8", + "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2, 3", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3", + "[12] Stéphane Guinard and Loic Landrieu. Weakly supervised segmentation-aided classification of urban scenes from 3d" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "lidar point clouds. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42:151-157, 2017. 1, 2, 3", + "bbox": [ + 544, + 92, + 903, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2", + "[14] David Hilbert and David Hilbert. Über die stetige abbildung einer linie auf ein flächenstück. Dritter Band: Analysis-Grundlagen der Mathematik-Physik Verschiedenes: Nebst Einer Lebensgeschichte, pages 1–2, 1935. 2", + "[15] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 1, 3, 6, 7", + "[16] Le Hui, Linghua Tang, Yuchao Dai, Jin Xie, and Jian Yang. Efficient lidar point cloud oversegmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18003-18012, 2023. 2, 3, 6, 7, 8", + "[17] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8500-8509, 2022. 8", + "[18] Loic Landrieu and Mohamed Boussaha. Point cloud over-segmentation with graph-structured deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7440–7449, 2019. 1, 2, 6, 7", + "[19] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 1, 6, 7, 8", + "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2", + "[21] Bo Li, T Zhang, and T Xia. Vehicle detection from 3d lidar using fully convolutional network. arxiv 2016. arXiv preprint arXiv:1608.07916. 2", + "[22] Yangbin Lin, Cheng Wang, Dawei Zhai, Wei Li, and Jonathan Li. Toward better boundary preserved supervoxel segmentation for 3d point clouds. *ISPRS journal of photogrammetry and remote sensing*, 143:39–47, 2018. 1, 2", + "[23] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1200-1211, 2023. 3", + "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 2", + "[25] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition." + ], + "bbox": [ + 93, + 114, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25839", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2", + "[26] Jeremie Papon, Alexey Abramov, Markus Schoeler, and Florentin Worgotter. Voxel cloud connectivity segmentation-supervoxels for point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2027-2034, 2013. 1, 2, 6", + "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 7", + "[28] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 2, 8", + "[29] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 2, 5, 8", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3", + "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2", + "[32] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 2", + "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2", + "[34] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023. 3, 8", + "[35] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2", + "[36] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. Advances in neural information processing systems, 18, 2005. 5", + "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 8" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 2, 3, 4, 8", + "[39] Yanyang Xiao, Zhonggui Chen, Zhengtao Lin, Juan Cao, Yongjie Jessica Zhang, Yangbin Lin, and Cheng Wang. Merge-swap optimization framework for supervoxel generation from three-dimensional point clouds. Remote Sensing, 12(3):473, 2020. 2", + "[40] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 2, 8", + "[41] Guowen Zhang, Lue Fan, Chenhang He, Zhen Lei, Zhaoxiang Zhang, and Lei Zhang. Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700, 2024. 4", + "[42] Peiyan Zhang, Yuchen Yan, Xi Zhang, Chaozhuo Li, Senzhang Wang, Feiran Huang, and Sunghun Kim. Transgnn: Harnessing the collaborative power of transformers and graph neural networks for recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1285-1295, 2024. 2, 5", + "[43] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2", + "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2", + "[45] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3" + ], + "bbox": [ + 516, + 90, + 906, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25840", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_model.json b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..58c24daecac840a3ed9a788992291d12b3076f73 --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.756, + 0.154 + ], + "angle": 0, + "content": "Serialization based Point Cloud Oversegmentation" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.18, + 0.852, + 0.252 + ], + "angle": 0, + "content": "Chenghui Lu\\(^{1,2}\\), Jianlong Kwan\\(^{1,2}\\), Dilong Li\\(^{1,2*}\\), Ziyi Chen\\(^{1,2}\\), Haiyan Guan\\(^{3}\\) \n\\(^{1}\\)Department of Computer Science & Xiamen CVPR Key Laboratory, Huaqiao University \n\\(^{2}\\)Fujian Key Lab. of Big Data Intell. and Security, Huaqiao University \n\\(^{3}\\)Nanjing University of Information Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.255, + 0.847, + 0.27 + ], + "angle": 0, + "content": "{23013083012, jianlongkwan}@stu.hqu.edu.cn, {scholar.dll, chenziyihq}@hqu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.305, + 0.346, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.305, + 0.508, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.305, + 0.623, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.305, + 0.764, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.305, + 0.903, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 1. Our method efficiently generates superpoints from serialized point clouds by optimizing boundaries through reallocating points from each segment to the most appropriate surrounding segments (shown in green and red)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.468, + 0.327, + 0.484 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.486, + 0.825 + ], + "angle": 0, + "content": "Point cloud oversegmentation, as a fundamental preprocessing step for 3D understanding, is a challenging task due to its spatial proximity and semantic similarity requirements. Most existing works struggle to efficiently group semantically consistent points into superpoints while maintaining spatial proximity. In this paper, we propose a novel serialization based point cloud oversegmentation method, which leverages serialization to avoid complex spatial queries, directly accessing neighboring points through sequence locality for similarity matching and superpoint clustering. Specifically, we first serialize point clouds into a Hilbert curve and spatially-continuously partition them into initial segments. Then, to guarantee the internal semantic consistency of superpoints, we design an adaptive update algorithm that clusters superpoints by matching feature similarities between neighboring segments and refines segment features via Cross-Attention. Experiments on large-scale indoor and outdoor datasets demonstrate state-of-the-art performance in point cloud oversegmentation. Moreover, it is also adaptable to semantic segmentation and achieves promising performance. The code is available at https://github.com/CHL-glitch/SPCNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Point cloud oversegmentation partitions point clouds into multiple spatially and semantically homogeneous regions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.574 + ], + "angle": 0, + "content": "termed superpoints, which adaptively represent a set of points. Due to their representativeness, superpoint-based methods can significantly reduce computational complexity and improve semantic consistency. Therefore, it is becoming increasingly important in the field of point cloud processing and has attracted considerable attention from many researchers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.806 + ], + "angle": 0, + "content": "Early point cloud oversegmentation approaches primarily combined hand-crafted features with clustering or graph partitioning to generate superpoints. Papon et al. [26] introduced VCCS, which employed Fast Point Feature Histograms, selected seed points via octree structures, and performed K-means clustering. Lin et al. [22] reformulated the task as a subset selection problem while still using FPFH descriptors. Guinard et al. [12] simplified the process into a graph cut problem by extracting geometric features before applying a greedy algorithm. Similarly, Landrieu et al. [19] partitioned superpoints by solving an energy minimization problem with graph-based simplicity penalties. However, handcrafted features have limited representation, and the non-differentiable superpoint assignments prevent end-to-end training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Recent point cloud oversegmentation methods generate superpoints using deep networks. Landrieu et al. [18] first used deep networks to extract point embeddings instead of handcrafted features, but their framework is not end-to-end. SPNet [15] proposed an end-to-end superpoint network that achieved clustering by iteratively constructing a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "point-superpoint association map. However, their Farthest Point Sampling (FPS) based k-nearest neighbor (KNN) association map faced efficiency bottlenecks, and the unlimited range of original points being clustered resulted in irregular superpoint boundaries and even introduced noise. SuperLiDAR [16] combined Breadth-First Search (BFS) grouping with local discriminative loss to generate compact superpoints, and while it achieved \\( O(V + E) \\) complexity on sparse LiDAR data, it may approach \\( O(V^2) \\) in dense scenarios, limiting scalability. Those point-based methods design complex algorithms to cluster superpoints and/or inefficiently balance semantic consistency and spatial proximity, resulting in high computational burden." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.483, + 0.606 + ], + "angle": 0, + "content": "To address this issue, we propose an efficient serialization based point cloud oversegmentation method. Inspired by PTv3 [38], we adopt a serialization approach instead of traditional point-based methods to avoid high computational complexity spatial queries such as FPS and KNN, as shown in Fig. 1. We find that one of the representative serialization strategies, the Hilbert curve [14], which has excellent locality-preserving properties, is crucial for spatial proximity in oversegmentation tasks. Thus, we apply the Hilbert curve to serialize point clouds and initialize superpoints. To achieve internal semantic consistency for the superpoints, we design an adaptive superpoint update algorithm that clusters semantically similar points within prepartitioned segments to update superpoints. Specifically, we use cosine similarity to evaluate the semantic similarity between points and nearby segments, which serves as the metric for associating points and superpoints to reallocate points, then apply Cross-Attention to update superpoint features. By combining these two steps and iterating them, the algorithm can obtain superpoints with both spatial proximity and semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Furthermore, to optimize superpoint feature representations, we employ a stacked Graph Transformer [42] and Graph Convolutional Network [42] to capture the global context between superpoints. Similar to prior work, SPCNet is designed as an end-to-end oversegmentation network and can also be integrated with downstream tasks such as semantic segmentation. To fully evaluate SPCNet, we conduct experiments on four large-scale datasets: S3DIS [1], ScanNet [9], nuScenes [3], and SemanticKITTI [2]. Experimental results demonstrate that SPCNet achieves state-of-the-art performance in all core oversegmentation metrics (BR, BP, and F1) across all datasets. The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.483, + 0.841 + ], + "angle": 0, + "content": "- To our best knowledge, SPCNet is the first serialization based network for point cloud oversegmentation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- We introduce a serialization strategy to initialize superpoints, which can efficiently produce superpoints with spontaneous spatial proximity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose an adaptive superpoint update algorithm to" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.897, + 0.107 + ], + "angle": 0, + "content": "strengthen internal semantic consistency of superpoints." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.168 + ], + "angle": 0, + "content": "- Our method achieves remarkable performance in point cloud oversegmentation across multiple datasets. It can be integrated with semantic segmentation networks and shows promising performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.183, + 0.655, + 0.199 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.209, + 0.78, + 0.224 + ], + "angle": 0, + "content": "2.1. Deep Learning in Point Cloud" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Deep learning approaches for 3D point cloud primarily follow three paradigms: projection-based, voxel-based, and point-based methods. Projection-based methods project unordered points onto 2D planes for CNN processing [6, 20, 21, 32], sacrificing 3D geometry for computational efficiency. Voxel-based methods discretize space into grids for 3D convolution [25, 31], improved by sparse convolution techniques [8, 10, 35] to mitigate memory costs, though kernel size limitations remain a constraint. Point-based methods [24, 27, 28, 33, 43] process raw point clouds directly and have evolved toward transformer architectures [13, 29, 37, 40, 44]. These methods offer powerful representation capabilities while struggling with scalability. The recently proposed SPT [29] leverages hierarchical superpoints and sparse self-attention to efficiently process large-scale scenes, achieving superior semantic segmentation with reduced model size and computational cost compared to traditional approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.515, + 0.787, + 0.531 + ], + "angle": 0, + "content": "2.2. Point Cloud Oversegmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Existing point cloud oversegmentation methods can be broadly categorized into optimization-based methods and deep learning-based approaches. Early methods such as Papon et al. [26] proposed the VCCS method, which used FPFH as input features, selected seed points using an octree, and performed point cloud superpoint segmentation using the K-means clustering algorithm. Lin et al. [22] viewed the superpoint segmentation problem as a subset selection problem and employed subset selection methods to generate superpoints, still using FPFH as point cloud feature descriptors. Guinard et al. [12] simplified the superpoint segmentation problem into a graph cut problem, extracting features such as local linearity, flatness, divergence, and normal vectors from the point cloud, and applied a greedy graph cut algorithm to generate superpoints. Xiao et al. [39] defined the superpoint segmentation problem as an energy minimization problem and proposed a merging-exchange optimization framework to generate supervoxels. These methods often require complex and computationally expensive preprocessing steps, and their performance is limited by the insufficient expressiveness of handcrafted features. Recent deep learning methods overcome these limitations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Supervised Superpoint (SSP) method [18] pioneered deep network-based point cloud embedding, although it" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25832" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.904, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.326, + 0.316, + 0.338 + ], + "angle": 0, + "content": "(a) Point Feature Extractor" + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.327, + 0.629, + 0.339 + ], + "angle": 0, + "content": "(b) Hierarchical Segment Initialization" + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.326, + 0.884, + 0.34 + ], + "angle": 0, + "content": "(c) Adaptive Superpoint Update" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Figure 2. The architecture includes three stages. First, point embeddings are extracted using a U-Net backbone with sparse convolutions in shallow layers and a Mamba bottleneck enhanced by Conditional Position Encoding (CPE). Second, the point cloud is serialized via a Hilbert curve to initialize segments for superpoint aggregation, with granularity at each level controlled by the hyperparameter \\( m \\). Finally, points are reassigned across segments to group semantically similar points into coherent superpoints, completing the clustering process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.436, + 0.483, + 0.587 + ], + "angle": 0, + "content": "still depended on optimization techniques from [12]. SP-Net [15] further advanced the field by introducing learnable point-to-superpoint associations using FPS for center selection, enabling joint optimization in both coordinate and feature spaces. Furthermore, SuperLiDAR [16] improved efficiency by combining a BFS-based grouping algorithm with a local discriminative loss to generate compact superpoints. However, this method results in high computational complexity in spatial query operations during superpoint generation, especially for dense point cloud oversegmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.597, + 0.339, + 0.612 + ], + "angle": 0, + "content": "2.3. Localization-based Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.801 + ], + "angle": 0, + "content": "Recent works [5, 23, 34, 38] have introduced serialization-based approaches that differ from traditional point cloud processing by transforming unstructured point clouds into ordered sequences to preserve spatial proximity. For example, OctFormer [34] uses octree-based ordering similar to \\(z\\)-ordering, offering scalability but limited by octree constraints. FlatFormer [23] employs window-based sorting to group point pillars, though its scalability in receptive field is limited, making it suitable for pillar-based 3D object detection. Building on these methods, PTv3 [38] extends attention scales by replacing KNN-based query with serialized neighborhoods, enhancing efficiency and scalability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.182, + 0.83 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our oversegmentation network comprises three key steps: point embedding extraction (Sec. 3.1), hierarchical segment initialization (Sec. 3.2), and adaptive superpoint update (Sec. 3.3). The overall process is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.435, + 0.627, + 0.45 + ], + "angle": 0, + "content": "3.1. Backbone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Optimization-based oversegmentation methods are constrained by the quality of handcrafted features, while learning based approaches similarly encounter limitations in deep feature representation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.519, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Sparse 3D U-Net. Let the input point cloud be \\(\\mathcal{P} \\in \\mathbb{R}^{N \\times (C + 3)}\\), where each of the \\(N\\) points is represented by its features (e.g. normal vectors, RGB) and spatial coordinates \\((x, y, z)\\). Following [10], we first voxelize \\(\\mathcal{P}\\) using the operator \\(\\mathcal{V}\\), and then process the discretized data with a U-Net architecture based on sparse convolutions (SpConv). In the shallow part of the U-Net [30], SpConv layers capture fine-grained local geometric information. These operations produce feature representations that maintain high spatial resolution and extract detailed geometric structures crucial for accurate object boundary delineation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Bottleneck Reinforcement. Pure SpConv restricts over-segmentation effectiveness due to their inherently limited receptive field. We enhance our U-Net backbone by integrating Mamba [11, 45] at the bottleneck of the U-Net, as shown in Fig. 2 (a). After the shallow SpConv layers have abstracted local features, the Mamba module is applied to perform global context interaction. Mamba, with its linear-complexity global modeling capability, compensates for the local inductive bias of convolution and promotes semantic consistency across superpoints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.821, + 0.864 + ], + "angle": 0, + "content": "3.2. Hierarchical Segment Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Methodology Justification. To efficiently partition point clouds into multiple segments, we propose a serialization-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25833" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.319 + ], + "angle": 0, + "content": "based partitioning method that utilizes the spatial locality preservation property of the Hilbert curve [38, 41] to divide the same point cloud into hierarchical structures of different granularities. Specifically, the Hilbert curve maps adjacent point clusters in 3D Euclidean space to contiguous positions in a 1D sequence, thereby simplifying the spatial proximity-based 3D partitioning task into an iterative partitioning of a 1D sequence. This transformation enables us to efficiently divide point clouds into compact segments in 3D space, significantly enhancing partitioning efficiency. This enables (1) efficient 1D clustering that avoids expensive 3D queries and clustering; and (2) a simplified oversegmentation process that decouples spatial locality from semantic similarity by performing only semantic comparisons within already spatially adjacent segments to complete clustering." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.325, + 0.486, + 0.657 + ], + "angle": 0, + "content": "Partitioning Strategy. Due to the variable receptive field of differently sized superpoints, smaller superpoints preserve more detailed local geometric information, whereas larger superpoints capture more comprehensive global semantic context. Therefore, we propose to construct a hierarchical partitioning structure to fully utilize multi-scale information. Given an input point cloud \\(\\mathcal{P}\\) and its corresponding features \\(F = \\{f_i\\}_{i=1}^N\\), where \\(f_i \\in \\mathbb{R}^d\\) is generated by our backbone, we construct a hierarchical partitioning structure \\(\\mathcal{H} = \\{L_k\\}_{k=1}^K\\). Each layer \\(L_k\\) contains the complete serialized point cloud and is recursively quaternary partitioned into a set of segments \\(L_k = \\{S_i^k\\}_{i=0}^{M_k-1}\\) where \\(S_i^k\\) represents the \\(i\\)-th segment, as shown in Fig. 2 (b). Specifically, the 3D point cloud is first serialized into a 1D sequence via Hilbert space-filling curve mapping [38], denoted as \\(\\mathcal{P}_{seq} \\in \\mathbb{R}^{N \\times d}\\). The sequence is then padded to \\(\\mathcal{P}_{pad} \\in \\mathbb{R}^{N_{pad} \\times d}\\) where \\(N_{pad} = \\lceil N / M_K \\rceil \\cdot M_K\\), ensuring divisibility by the finest segment count \\(M_K\\). The serialized point cloud \\(\\mathcal{P}_{pad}\\) is recursively partitioned using 1D quaternary splitting, such that at each layer \\(L_k\\), it is uniformly divided into \\(M_k = 2^{m - 2k + 2}\\) segments. For each layer \\(L_k\\), \\(\\mathcal{S}_i^k\\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.664, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} ^ {k} = \\left\\{\\mathcal {P} _ {p a d} [ j ] \\mid j \\in [ i \\cdot l _ {k}, (i + 1) \\cdot l _ {k} - 1 ] \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.764 + ], + "angle": 0, + "content": "where \\( l_{k} = N_{pad} / M_{k} \\) represents the segment length for each segment in the \\( k \\)-th layer. Recursive partitioning establishes fixed parent-child indexing between layers. For any inter-layer gap \\( l \\geq 1 \\), the hierarchical inclusion relationship is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.771, + 0.483, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} ^ {k} = \\bigcup_ {j \\in [ i \\cdot 4 ^ {l}, (i + 1) \\cdot 4 ^ {l} - 1 ]} \\mathcal {S} _ {j} ^ {k + l}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.483, + 0.846 + ], + "angle": 0, + "content": "At the finest layer \\(L_{K}\\), segment features are generated through permutation-invariant aggregation of constituent points:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.846, + 0.482, + 0.866 + ], + "angle": 0, + "content": "\\[\nf _ {S _ {i} ^ {K}} = \\mathcal {A} \\left(f _ {j} \\mid p _ {j} \\in S _ {i} ^ {K}\\right) \\in \\mathbb {R} ^ {C}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}:\\mathbb{R}^{n_i\\times C}\\to \\mathbb{R}^C\\) denotes a permutation-invariant operator (e.g. max-pooling) with \\(n_i = |\\mathcal{S}_i^K | = N_{pad} / M_K\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.771, + 0.108 + ], + "angle": 0, + "content": "3.3. Adaptive Superpoint Update" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.286 + ], + "angle": 0, + "content": "Given the structure \\(\\mathcal{H}\\), the goal is to refine the partition \\(L_{k} = \\{S_{1}^{k},\\ldots ,S_{M_{k}}^{k}\\}\\) with features \\(f_{S_i^k}\\in \\mathbb{R}^d\\) into semantically consistent superpoints \\(L_k^\\prime = \\{S_1'^k,\\dots,S_M'^k\\}\\) with precise geometric boundaries and updated features \\(f_{S_i^k}^\\prime \\in \\mathbb{R}^d\\). This is achieved by dynamically allocating raw points within coarse-grained segments to their appropriate sub-segments based on semantic similarity, a process that redistributes raw points within segments to fit object boundaries. This process can be transformed into a boundary fitting problem between adjacent segments in 1D sequence point clouds, as shown in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Feature Similarity based Affiliation. To establish the affiliation relationship between raw points within each segment and nearby segments, we consider two adjacent levels in \\(\\mathcal{H}\\): the finest \\(L_{k} = \\{S_{1}^{k}, S_{2}^{k}, \\ldots, S_{M_{k}}^{k}\\}\\) and its coarser predecessor \\(L_{k-1} = \\{S_{1}^{k-1}, S_{2}^{k-1}, \\ldots, S_{M_{k-1}}^{k-1}\\}\\) (we only use \\(l = 1\\) defined in Eq. (2)). The hierarchical relationship between these levels is given by (1) \\(M_{k-1} = \\frac{M_{k}}{4}\\), meaning each coarser level has one-fourth the segments of the finer level and (2) Each segment \\(S_{i}^{k-1}\\) at level \\(L_{k-1}\\) contains four consecutive segments from level \\(L_{k}\\): \\(S_{i}^{k-1} = S_{4i}^{k} \\cup S_{4i+1}^{k} \\cup S_{4i+2}^{k} \\cup S_{4i+3}^{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.907, + 0.663 + ], + "angle": 0, + "content": "For each raw point \\( p_i \\in S_i^{k-1} \\), we identify the most similar initialized segment feature from its four containing segments \\( S_{4i}^k, S_{4i+1}^k, S_{4i+2}^k \\), and \\( S_{4i+3}^k \\). Meanwhile, to expand the receptive field of raw points relative to their potential superpoints, we incorporate segment features from eight sub-segments at \\( L_k \\) belonging to the left and right adjacent segments \\( S_{i-1}^{k-1} \\) and \\( S_{i+1}^{k-1} \\), forming a local feature set \\( \\{f_{S_{4i+j}^k} \\mid j \\in \\{-4, -3, \\dots, 7\\}\\} \\), resulting in a set of 12 initial segment features. For each segment \\( S_i^{k-1} \\) with \\( N_{k-1} = 4N_{pad} / M_k \\) points, we compute a similarity matrix \\( \\mathbf{A}_i \\in \\mathbb{R}^{N_{k-1} \\times 12} \\). Each entry \\( a_{j,t} \\) measures the similarity between the \\( j \\)-th point and the \\( t \\)-th segment feature in the above set of segment features:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.668, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\na _ {j, t} = \\frac {f _ {p _ {j}} \\cdot f _ {S _ {t} ^ {k}}}{\\| f _ {p _ {j}} \\| \\| f _ {S _ {t} ^ {k}} \\|}, \\quad j \\in [ 1, N _ {k - 1} ], \\quad t \\in [ 1, \\tau ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.905, + 0.82 + ], + "angle": 0, + "content": "This process generates similarity matrices \\(\\{\\mathbf{A}_i\\}_{i=1}^{M_{k-1}}\\) for the entire level \\(L_{k-1}\\). Point-to-superpoint assignments are determined by selecting the superpoint with the highest similarity for each point, which is then mapped to global superpoint indices \\(\\phi(p_j) \\in [1, M_k]\\). The resulting assignment index set \\(Ass_i = \\{\\phi(p_j)\\}_{j=1}^n\\) provides a key reference for updating superpoint features in subsequent steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Cross-Attention based Update. Cross-Attention [4] enables dynamic and adaptive relationships between superpoints and their constituent points. Through feature similarity based affiliation, precise similarity relationships are established, guiding accurate superpoint feature updates. For" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.485, + 0.348 + ], + "angle": 0, + "content": "Figure 3. The end-to-end semantic segmentation pipeline constructs a graph over superpoints, captures global context using graph transformers, and integrates point-level embeddings for final semantic prediction, where \\(\\oplus\\) denotes simple feature addition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.485 + ], + "angle": 0, + "content": "the assignment indices \\( A s s i = \\{\\phi(p_{j})\\}_{j=1}^{n} \\), the inverse mapping \\( \\phi^{-1}(i) = \\{p_{j} \\mid \\phi(p_{j}) = i\\} \\) represents the original points contained within superpoint \\( S_{i}^{\\prime k} \\). Based on \\( A s s i \\), each superpoint's feature is updated solely from its member points. Using superpoint feature \\( f_{S_{i}^{\\prime k}}^{\\prime} \\) as the Query and its member point features \\( \\{f_{p_{j}} \\mid p_{j} \\in \\phi^{-1}(i)\\} \\) as Keys and Values, the update formula is:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.494, + 0.484, + 0.557 + ], + "angle": 0, + "content": "\\[\nf _ {S _ {i} ^ {k}} ^ {\\prime} = \\text {C r o s s A t t n} \\left(\\underbrace {f _ {S _ {i} ^ {k}}} _ {\\text {Q u e r y}}, \\underbrace {\\left\\{f _ {p _ {j}} \\mid p _ {j} \\in \\phi^ {- 1} (i) \\right\\}} _ {\\text {K e y / V a l u e}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Attention weights are computed through scaled dot product, allowing each superpoint to focus on the most informative points within its local neighborhood:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.619, + 0.483, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C r o s s A t t n} \\left(f _ {S _ {i} ^ {k}}, \\left\\{f _ {p _ {j}} \\right\\}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {i}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.791 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_i = f_{S_i^k}W_K\\in \\mathbb{R}^{1\\times d}\\), \\(\\mathbf{K}_i = \\mathrm{Concat}(\\{f_{p_j}W_K\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}\\), \\(\\mathbf{V}_i = \\mathrm{Concat}(\\{f_{p_j}W_V\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}\\), \\(n_i^k = |\\phi^{-1}(i)|\\) is the number of points contained in each superpoint, and \\(W_{Q},W_{K},W_{V}\\in \\mathbb{R}^{d\\times d}\\) are learnable projection matrices. For coarse-grained superpoints, we apply the same update logic, with interactions between fine-grained features and their coarse-grained initializations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.305, + 0.817 + ], + "angle": 0, + "content": "3.4. Superpoint Interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The superpoint features \\( f_{S_i^k}^\\prime \\) encode geometric-semantic information for locally homogeneous points. To leverage these features for semantic segmentation, inspired by SPT [29], we integrate a superpoint interaction module with our oversegmentation network, forming an end-to-end" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.337 + ], + "angle": 0, + "content": "framework, as shown in Fig. 3. In our hierarchical design, fine-grained superpoints capture local details while coarse grained ones represent object-level semantics. We construct multi-level superpoint graphs and employ a cascaded Transformer-GCN [42] to facilitate cross-superpoint interactions and global context modeling. In the Graph Transformer, a superpoint graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\) is built by linking each superpoint with its \\(k\\) nearest neighbors in feature space, where \\(\\mathcal{V} = \\{f_{S_i^k}'\\}_{i=1}^{M_k}\\). Edge features are defined via relative position encoding as \\(\\mathbf{e}_{ij} = g(p_j - p_i)\\), with \\(p_i\\) and \\(p_j\\) representing node coordinates and \\(g\\) being a linear mapping. A subsequent GCN processes this graph to enhance local feature interactions through message passing among adjacent superpoints. Finally, we fuse backbone point features with their corresponding multi-level superpoint before feeding them into the segmentation head." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.776, + 0.362 + ], + "angle": 0, + "content": "3.5. Superpoint Aggregation Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.548 + ], + "angle": 0, + "content": "We propose a superpoint aggregation loss to optimize superpoint generation, inspired by LMNN [36]. This comprehensive loss consists of three components. First, a compactness loss \\(\\ell_{\\mathrm{compact}}\\) enforces intra-superpoint feature cohesion by minimizing point-to-superpoint distances. Second, a distinction loss \\(\\ell_{\\mathrm{dist}}\\) encourages inter-class separability by maximizing the distances between superpoints of different classes, where each superpoint's label is determined by majority voting among its points. Third, a purity loss \\(\\ell_{\\mathrm{purity}}\\) maintains label consistency within each superpoint by penalizing high entropy in its class distribution. The overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.554, + 0.907, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\ell_ {s p} = \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\frac {\\sum_ {p \\in S _ {i}} \\| f _ {i} - f _ {S _ {i} ^ {k}} ^ {\\prime} \\| ^ {2}}{| S _ {i} |} + \\\\ \\frac {1}{N _ {c}} \\sum_ {i, j \\in c, i \\neq j} \\max \\left(0, \\delta_ {\\text {d i s t}} - \\| f _ {S _ {i} ^ {k}} ^ {\\prime} - f _ {S _ {j} ^ {k}} ^ {\\prime} \\|\\right) ^ {2} + \\tag {7} \\\\ \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {k = 1} ^ {K} p _ {i, k} \\log (p _ {i, k}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.773 + ], + "angle": 0, + "content": "where \\( f_{i} \\) denotes the feature of point \\( p \\), \\( f_{S_i^k}^\\prime \\) represents the feature of superpoint \\( S_{i}^{\\prime k} \\), \\( \\delta_{\\mathrm{dist}} \\) is a predefined separation margin, \\( p_{i,k} \\) is the probability of superpoint \\( i \\) belonging to class \\( k \\), \\( N_{s} \\) is the number of superpoints, and \\( N_{c} \\) is the number of superpoint pairs involved in \\( \\ell_{\\mathrm{dist}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.646, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.714, + 0.833 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate SPCNet on four datasets with diverse characteristics. S3DIS [1] contains \\(274M\\) points from six office areas, with Area 5 as the test set. ScanNet [9] includes 1,513 indoor scans from 707 scenes with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25835" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.271, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 4. The visualization of superpoints generated on the ScanNet (first row) and S3DIS (second row) datasets shows a two-level hierarchy: \\( L_{1} \\) with 1024 superpoints and \\( L_{2} \\) with 256 superpoints. Finer \\( L_{1} \\) superpoints merge based on semantic similarity to form larger \\( L_{2} \\) superpoints. For example, in S3DIS, \\( L_{1} \\) superpoints of table parts merge into a single table superpoint, and \\( L_{1} \\) superpoints of whiteboard sections form a complete whiteboard superpoint. Our superpoints are regular and orderly, with nearly rectangular shapes in semantically consistent areas (e.g. floors, tables), maintaining continuity, precise object boundaries, and demonstrating superior regularity, compactness, and purity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.382, + 0.482, + 0.473 + ], + "angle": 0, + "content": "20 categories, averaging 148k points/scan. nuScenes [3] comprises 1,000 scenes from 32-beam LiDAR, split 750/150/150 for train/val/test, with 17 categories. SemanticKITTI [2] is an outdoor self-driving benchmark with 22 sequences and 20 categories, using sequences 00-10 (densely annotated) for training and 11-22 for testing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.609 + ], + "angle": 0, + "content": "Evaluation Metrics. Following SPNet [15] and SSP [18], we use Oracle Overall Accuracy (OOA), Boundary Recall (BR), Boundary Precision (BP), and F1 score to evaluate superpoints. BR and BP assess boundary quality, while OOA represents the upper bound of semantic segmentation accuracy using superpoints. The F1 score balances recall and precision, defined as \\( F1 = \\frac{2\\cdot BP\\cdot BR}{BP + BR} \\). For semantic segmentation in indoor and outdoor scenarios, we adopt mean intersection over union (mIoU)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.62, + 0.291, + 0.636 + ], + "angle": 0, + "content": "4.2. Model Configuration" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.483, + 0.792 + ], + "angle": 0, + "content": "We maintain a consistent architecture across all experiments with minimal dataset-specific adjustments. Our backbone features a four-stage symmetric encoder-decoder structure, each with a single block depth. The encoder has embedding dimensions [32, 64, 128, 256] and the decoder [128, 64, 64], with Mamba blocks using conditional positional encoding at the 256 stage. We implement a three-level hierarchy \\((m = 8, K = 2, l = 1)\\), organizing points into n, 1024, 256 segments. Training is performed on dual NVIDIA RTX 4090 GPUs with the Adam optimizer." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.803, + 0.365, + 0.819 + ], + "angle": 0, + "content": "4.3. Point Cloud Oversegmentation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Tab. 1, we evaluate SPCNet on three datasets, reserving ScanNet for ablation studies, ensuring fair comparison with consistent superpoint counts. For example, in S3DIS Area 5, methods like VCCS, Lin et al., SPG, SSP, and SPNet generate around 1050 su" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.379, + 0.895, + 0.735 + ], + "angle": 0, + "content": "
MethodBRBPF1OOA
S3DIS Area 5 [1]
VCCS [26]62.0610.8618.4895.22
SPG [19]52.0613.1120.9495.84
SSP [18]80.7313.0222.4297.04
SPNet [15]84.7513.1422.6596.50
SPCNet(Ours)89.5026.7840.5096.81
nuScenes [3]
SPG [19]28.2017.2621.4189.22
SSP [18]22.0415.6318.2892.01
SPNet [15]68.9218.3428.9787.67
SuperLiDAR [16]74.7225.1237.5996.31
SPCNet(Ours)88.7732.8247.5296.33
SemanticKITTI [2]
SPG [19]25.6415.8219.5686.86
SSP [18]18.7410.6713.5992.27
SPNet [15]56.5214.7823.4392.24
SuperLiDAR [16]65.5220.5231.2596.21
SPCNet(Ours)73.1830.1542.5796.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.743, + 0.907, + 0.772 + ], + "angle": 0, + "content": "Table 1. Comparison results of generated superpoints on the S3DIS, nuScenes, and SemanticKITTI datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "perpoints, matching our 1024 Hilbert partitions. Our deep learning-based approach outperforms traditional methods, benefiting from the specialized backbone network's ability to extract rich geometric features that capture complex local structures. Unlike methods such as SPNet [15], which rely on FPS+KNN for presampling and soft assignment map-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.332 + ], + "angle": 0, + "content": "ping without limiting the scope of raw points being clustered, resulting in blurry boundaries, our approach leverages the Hilbert curve's locality, limiting the range of raw point clustering (preventing clustering of semantically similar but spatially distant points), and finally refining superpoint through the Cross-Attention, minimizing noise and maintaining precise, simple boundaries, as shown in Fig. 5 (d) and (e). Visualization results in Fig. 4 show that our method tends to aggregate superpoints within object interiors while maintaining distinct object boundaries, especially in challenging cases like wall-whiteboard interfaces, resulting in higher BR and BP. Additionally, our method excels at merging semantically continuous regions (e.g. floors, tabletops), leading to fewer incorrect superpoint boundaries and substantially higher BP metrics. Fig. 6 shows the performance of different methods on SemanticKITTI." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.343, + 0.483, + 0.426 + ], + "angle": 0, + "content": "
MethodDeviceInference Time (ms) data proc. / backbone + SP ini. + SP gen.
SPG [19]CPU13152
SSP [18]RTX 309016115 (11812 + 0 + 13426)
SPNet [15]RTX 309012182 (11812 + 0 + 1154)
SuperLiDAR [16]RTX 309072(0 + 0 + 72)
SPCNet (ours)RTX 4090, RTX 306012 (6 + 0.4 + 6), 61
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Table 2. The inference time of different oversegmentation methods on SemanticKITTI. The inference time consists of data preprocessing/backbone (red/blue), superpoint initialization (green), and superpoint generation (purple)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.882 + ], + "angle": 0, + "content": "Time Costs. We evaluated oversegmentation network efficiency by measuring inference time per scan on SemanticKITTI validation (batch size=1) in Tab. 2. For optimization based SPG [19], we followed SuperLiDAR's [16] configuration using Core i5 CPU. For learning-based methods (SSP [18], SPNet [15], SuperLiDAR), we used SuperLiDAR's results (RTX 3090). Our SPCNet was benchmarked on RTX 4090 and 3060 GPUs. Our method achieved total inference times of 12 ms (RTX 4090) and 61 ms (RTX 3060), achieving \\(100 \\times\\) speedup versus SPG/SSP/SPNet, and \\(5 \\times\\) versus SuperLiDAR. Even on RTX3060, our inference is 10ms faster than SuperLiDAR, indicating at least \\(3 \\times\\) efficiency accounting for hardware differences. This stems from: (1) efficient superpoint initialization (0.4 ms) that compresses FPS+KNN sampling time in SPNet; (2) the application of a localized attention mechanism within the superpoint refinement (6 ms). Our backbone architecture allows further acceleration using lightweight alternatives like PointNet[27]. Traditional methods face key bottlenecks: SPG/SSP rely on optimization-based generation (e.g. graph cuts), SSP/SPNet suffer from redundant processing due to mixed handcrafted/learned features, and SuperLiDAR's BFS-based grouping has complexity dependent on scene density, potentially reaching \\(O(n^{2})\\) in dense scenes, versus our serialization based approach \\((O(n))\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation Studies. Mamba mitigates the local inductive bias" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.095, + 0.616, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.171, + 0.605, + 0.191 + ], + "angle": 0, + "content": "(a) 3D-UNet only with \\(k = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.095, + 0.714, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.171, + 0.712, + 0.191 + ], + "angle": 0, + "content": "(b) Bottleneck reinforcement with \\( k = 2 \\)" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.095, + 0.805, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.171, + 0.786, + 0.191 + ], + "angle": 0, + "content": "(d) PointNet with \\( k = 2 \\)" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.095, + 0.902, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.171, + 0.875, + 0.182 + ], + "angle": 0, + "content": "(e) SPNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.206, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Oversegmentation comparisons. (a) Sparse 3D U-Net with neighbor expansion of 2. (b) Adding the Mamba module to the bottleneck. (c) Replacing the backbone with a simplified PointNet, as in SPNet. (d) SPNet results. Notably, while (b) configuration does not fully aggregate semantically consistent objects into complete superpoints, it notably enhances consistency. SPC-Net also avoids clustering semantically similar but spatially distant points, ensuring clearer superpoint boundaries." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "
Backbone SettingBR(%)BP(%)OOA(%)mIou(%)
3D U-Net76.33 ↓7.9028.45 ↓1.7090.93 ↓1.1974.10 ↓2.86
+ Mamba fully78.37 ↓5.8629.13 ↓1.0291.24 ↓0.8876.18 ↓0.78
+ Mamba bottleneck (default)84.2330.1592.1276.96
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.399, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Table 3. Performance comparison of different backbone architectures on the ScanNet validation set." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.446, + 0.903, + 0.53 + ], + "angle": 0, + "content": "
Neighbor ScopeBR(%)BP(%)OOA(%)
No expansion79.86 ↓4.3728.57 ↓1.5891.22 ↓0.90
2 expansion (default)84.2330.1592.12
4 expansion83.67 ↓0.5629.85 ↓0.3091.24 ↓0.88
6 expansion79.93 ↓4.3028.65 ↓1.5089.93 ↓2.19
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.54, + 0.905, + 0.569 + ], + "angle": 0, + "content": "Table 4. The effect of expanding neighbor segment range on superpoint performance on the ScanNet validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "of convolutions in oversegmentation, improving superpoint continuity and semantic consistency. As shown in Fig. 5 (a) and (b), this bias causes superpoints in the same semantic region to align with boundaries but fragment due to internal inconsistency. Thus, we retain convolutions in U-Net's shallow layers for local geometric cues to ensure boundary precision, and introduce Mamba at the bottleneck to capture long range semantics. Tables 3 and 4 analyze backbone design and neighbor expansion effects on superpoint quality. We evaluated three backbone configurations: (1) baseline 3D sparse U-Net only, (2) fully Mamba-based U-Net, and (3) our default hybrid design with Mamba modules in the bottleneck layers. Results demonstrate that the baseline 3D sparse U-Net reduces BR, BP, and OOA by \\(7.90\\%\\), \\(1.70\\%\\), and \\(1.19\\%\\) respectively compared to our default. Similarly, the fully Mamba-based U-Net underperforms, as Mamba's global context modeling in shallow networks with low-dimensional features smooths out high-frequency boundary information. Our hybrid \"shallow convolution + deep Mamba\" architecture achieves optimal per" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.298, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.093, + 0.498, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.092, + 0.701, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.092, + 0.902, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.908, + 0.248 + ], + "angle": 0, + "content": "Figure 6. Performance of different methods on the SemanticKITTI validation set under varying numbers of superpoints. It is worth noting that our method can achieve better results than other methods with fewer superpoints (256, 512, and 1024)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.271, + 0.485, + 0.458 + ], + "angle": 0, + "content": "
ExperimentScanNet [9]S3DIS [1]nuScenes [3]Sem.KITTI [2]
Best ModelValTestArea5ValTestValTest
†SPG [19]--58.0----
MinkUNet [8]72.273.665.473.3-63.8-
†SPG + SSP [19]--61.7----
ST [17]74.373.772.0----
†SPT [29]--68.9----
PointNeXt [28]71.571.270.5----
OctFormer [34]75.776.6-76.177.264.367.8
Swin3D [40]76.4-72.5----
AF2S3Net [7]---62.278.074.270.8
†SuperLiDAR [16]----78.5-69.6
PTv2 [37]75.474.271.680.282.670.372.6
PTv3 [38]77.577.973.480.482.770.874.2
Backbone76.9-69.375.4-67.6-
†SPCNet(Ours)77.878.072.177.680.269.571.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.497 + ], + "angle": 0, + "content": "Table 5. Semantic segmentation results on the ScanNet, S3DIS, nuScenes, and SemanticKITTI datasets. †Superpoint-based." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.569 + ], + "angle": 0, + "content": "formance by leveraging convolutions to preserve local details in shallow layers while Mamba modules handle global context modeling in deeper layers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.483, + 0.734 + ], + "angle": 0, + "content": "In the neighborhood expansion experiments (analogous to k-parameter tuning in KNN), we observed that moderately expanding the candidate superpoint pool of original points during the feature similarity based affiliation process significantly improves sampling accuracy. The model achieves optimal performance when \\( \\mathrm{K} = 2 \\); excessive expansion \\( (\\mathrm{K} = 4,6) \\) leads to performance degradation. This may be because larger receptive fields introduce noise, and the hard assignment approach introduces more uncertainty as the number of candidate targets increases, causing early assignment errors to propagate to later processing stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.308, + 0.759 + ], + "angle": 0, + "content": "4.4. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tab. 5 evaluates our approach on four indoor and outdoor datasets, comparing it with point-based and superpoint-based methods. Our optimized oversegmentation strategy shows significant gains: on S3DIS Area 5, we achieve a \\(3.2\\%\\) mIoU improvement over SPT. Compared to state-of-the-art point-based methods, our approach remains competitive, outperforming PTv3 by \\(0.3\\%\\) on ScanNet \\((77.8\\%)\\) and PTv2 by \\(0.5\\%\\) on S3DIS Area 5. Additionally, we achieve \\(71.9\\%\\) and \\(80.2\\%\\) on Sem.KITTI and nuScenes, improving" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.271, + 0.905, + 0.343 + ], + "angle": 0, + "content": "
SettingComponentsModel Size (m)mIoU(%)Δ(%)
Backbone + SP gen. + SP int.
backbone (default)-18.876.96-
Hierarchical Enhancement
1 partition level+ Level1 Superpoint0.24 (0.218 + 0.022)77.50+0.54
2 partition level+ Level2 Superpoint0.02 (0.014 + 0.006)77.80+0.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.352, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Table 6. Semantic segmentation improvement with hierarchical superpoint features in ScanNet validation set. SP gen. means superpoint generation; SP int. means superpoint interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.513 + ], + "angle": 0, + "content": "SuperLiDAR by \\(1.3\\%\\) and \\(1.7\\%\\), respectively. These results confirm that our superpoints capture local and global semantic information effectively, while the simple backbone, enhanced by superpoint features, matches the performance of more complex point-based methods, validating the superior quality of the generated superpoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.648 + ], + "angle": 0, + "content": "As shown in Tab. 6, integrating hierarchical superpoint features with per-point features extracted by the backbone improved segmentation. First-level superpoints aggregated local semantics, boosting intra-class consistency with a \\(0.54\\%\\) mIoU gain. Second-level features captured long-range dependencies, adding another \\(0.30\\%\\) improvement. These hierarchical superpoints acted as adaptive receptive fields, enhancing performance with only \\(0.24M\\) parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.853 + ], + "angle": 0, + "content": "In this paper, we propose a novel serialization based oversegmentation method to address the challenge of efficiently grouping semantically consistent points into superpoints while maintaining spatial proximity. SPCNet achieves state-of-the-art performance in point cloud oversegmentation while being \\(3 \\times\\) faster than existing methods in inference speed. Extensive experimental results demonstrate the effectiveness and efficiency of SPCNet. Additionally, SPCNet can be flexibly applied to semantic segmentation tasks, effectively improving the accuracy of semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments This work was supported in part by the National Natural Science Foundation of China (Grant No. 42201475)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25838" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.353 + ], + "angle": 0, + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[4] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[5] Wanli Chen, Xinge Zhu, Guojin Chen, and Bei Yu. Efficient point cloud analysis using hilbert curve. In European Conference on Computer Vision, pages 730-747. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Stéphane Guinard and Loic Landrieu. Weakly supervised segmentation-aided classification of urban scenes from 3d" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "lidar point clouds. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42:151-157, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] David Hilbert and David Hilbert. Über die stetige abbildung einer linie auf ein flächenstück. Dritter Band: Analysis-Grundlagen der Mathematik-Physik Verschiedenes: Nebst Einer Lebensgeschichte, pages 1–2, 1935. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Le Hui, Linghua Tang, Yuchao Dai, Jin Xie, and Jian Yang. Efficient lidar point cloud oversegmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18003-18012, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[17] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8500-8509, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[18] Loic Landrieu and Mohamed Boussaha. Point cloud over-segmentation with graph-structured deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7440–7449, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[19] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[21] Bo Li, T Zhang, and T Xia. Vehicle detection from 3d lidar using fully convolutional network. arxiv 2016. arXiv preprint arXiv:1608.07916. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Yangbin Lin, Cheng Wang, Dawei Zhai, Wei Li, and Jonathan Li. Toward better boundary preserved supervoxel segmentation for 3d point clouds. *ISPRS journal of photogrammetry and remote sensing*, 143:39–47, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1200-1211, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25839" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[26] Jeremie Papon, Alexey Abramov, Markus Schoeler, and Florentin Worgotter. Voxel cloud connectivity segmentation-supervoxels for point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2027-2034, 2013. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[28] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[29] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[32] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[34] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[35] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[36] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. Advances in neural information processing systems, 18, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Yanyang Xiao, Zhonggui Chen, Zhengtao Lin, Juan Cao, Yongjie Jessica Zhang, Yangbin Lin, and Cheng Wang. Merge-swap optimization framework for supervoxel generation from three-dimensional point clouds. Remote Sensing, 12(3):473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[40] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[41] Guowen Zhang, Lue Fan, Chenhang He, Zhen Lei, Zhaoxiang Zhang, and Lei Zhang. Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[42] Peiyan Zhang, Yuchen Yan, Xi Zhang, Chaozhuo Li, Senzhang Wang, Feiran Huang, and Sunghun Kim. Transgnn: Harnessing the collaborative power of transformers and graph neural networks for recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1285-1295, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.527 + ], + "angle": 0, + "content": "[43] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.529, + 0.907, + 0.584 + ], + "angle": 0, + "content": "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[45] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25840" + } + ] +] \ No newline at end of file diff --git a/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_origin.pdf b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4dbaf90f37361d169d066722d5def572c015978f --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/02d08649-2762-474d-94e0-d2ae2cc700de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efd4e0212c8bfbcc7ee183caf7166951d6e970329fafafd6d7809c7fb0668480 +size 6802589 diff --git a/2025/Serialization based Point Cloud Oversegmentation/full.md b/2025/Serialization based Point Cloud Oversegmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d96e4e1035de7ee56da1bce1749769aac97179c0 --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/full.md @@ -0,0 +1,306 @@ +# Serialization based Point Cloud Oversegmentation + +Chenghui Lu $^{1,2}$ , Jianlong Kwan $^{1,2}$ , Dilong Li $^{1,2*}$ , Ziyi Chen $^{1,2}$ , Haiyan Guan $^{3}$ $^{1}$ Department of Computer Science & Xiamen CVPR Key Laboratory, Huaqiao University + $^{2}$ Fujian Key Lab. of Big Data Intell. and Security, Huaqiao University + $^{3}$ Nanjing University of Information Science and Technology + +{23013083012, jianlongkwan}@stu.hqu.edu.cn, {scholar.dll, chenziyihq}@hqu.edu.cn + +![](images/b3c467364161f46c290c1b5ef66f8c23f4391459826ef532d4ad5edaa5eafa3e.jpg) +Figure 1. Our method efficiently generates superpoints from serialized point clouds by optimizing boundaries through reallocating points from each segment to the most appropriate surrounding segments (shown in green and red). + +![](images/85afddb3e8aa4b56c4856d653b2d0b8337afbc9152978dfd116f8e3ece568fc0.jpg) + +![](images/d05036b13904f0b04924e72e7f0bb618979d7aa84786014eac8c191494dbaaf8.jpg) + +![](images/0157efd9819ccf3663eaae73f4541ffb220b738e3fd44c49a054840cd7378601.jpg) + +![](images/df7619d1151017d0ae3a45d7782ec24fdbb52dd80514bf26339af008ab9110d8.jpg) + +# Abstract + +Point cloud oversegmentation, as a fundamental preprocessing step for 3D understanding, is a challenging task due to its spatial proximity and semantic similarity requirements. Most existing works struggle to efficiently group semantically consistent points into superpoints while maintaining spatial proximity. In this paper, we propose a novel serialization based point cloud oversegmentation method, which leverages serialization to avoid complex spatial queries, directly accessing neighboring points through sequence locality for similarity matching and superpoint clustering. Specifically, we first serialize point clouds into a Hilbert curve and spatially-continuously partition them into initial segments. Then, to guarantee the internal semantic consistency of superpoints, we design an adaptive update algorithm that clusters superpoints by matching feature similarities between neighboring segments and refines segment features via Cross-Attention. Experiments on large-scale indoor and outdoor datasets demonstrate state-of-the-art performance in point cloud oversegmentation. Moreover, it is also adaptable to semantic segmentation and achieves promising performance. The code is available at https://github.com/CHL-glitch/SPCNet. + +# 1. Introduction + +Point cloud oversegmentation partitions point clouds into multiple spatially and semantically homogeneous regions, + +termed superpoints, which adaptively represent a set of points. Due to their representativeness, superpoint-based methods can significantly reduce computational complexity and improve semantic consistency. Therefore, it is becoming increasingly important in the field of point cloud processing and has attracted considerable attention from many researchers. + +Early point cloud oversegmentation approaches primarily combined hand-crafted features with clustering or graph partitioning to generate superpoints. Papon et al. [26] introduced VCCS, which employed Fast Point Feature Histograms, selected seed points via octree structures, and performed K-means clustering. Lin et al. [22] reformulated the task as a subset selection problem while still using FPFH descriptors. Guinard et al. [12] simplified the process into a graph cut problem by extracting geometric features before applying a greedy algorithm. Similarly, Landrieu et al. [19] partitioned superpoints by solving an energy minimization problem with graph-based simplicity penalties. However, handcrafted features have limited representation, and the non-differentiable superpoint assignments prevent end-to-end training. + +Recent point cloud oversegmentation methods generate superpoints using deep networks. Landrieu et al. [18] first used deep networks to extract point embeddings instead of handcrafted features, but their framework is not end-to-end. SPNet [15] proposed an end-to-end superpoint network that achieved clustering by iteratively constructing a + +point-superpoint association map. However, their Farthest Point Sampling (FPS) based k-nearest neighbor (KNN) association map faced efficiency bottlenecks, and the unlimited range of original points being clustered resulted in irregular superpoint boundaries and even introduced noise. SuperLiDAR [16] combined Breadth-First Search (BFS) grouping with local discriminative loss to generate compact superpoints, and while it achieved $O(V + E)$ complexity on sparse LiDAR data, it may approach $O(V^2)$ in dense scenarios, limiting scalability. Those point-based methods design complex algorithms to cluster superpoints and/or inefficiently balance semantic consistency and spatial proximity, resulting in high computational burden. + +To address this issue, we propose an efficient serialization based point cloud oversegmentation method. Inspired by PTv3 [38], we adopt a serialization approach instead of traditional point-based methods to avoid high computational complexity spatial queries such as FPS and KNN, as shown in Fig. 1. We find that one of the representative serialization strategies, the Hilbert curve [14], which has excellent locality-preserving properties, is crucial for spatial proximity in oversegmentation tasks. Thus, we apply the Hilbert curve to serialize point clouds and initialize superpoints. To achieve internal semantic consistency for the superpoints, we design an adaptive superpoint update algorithm that clusters semantically similar points within prepartitioned segments to update superpoints. Specifically, we use cosine similarity to evaluate the semantic similarity between points and nearby segments, which serves as the metric for associating points and superpoints to reallocate points, then apply Cross-Attention to update superpoint features. By combining these two steps and iterating them, the algorithm can obtain superpoints with both spatial proximity and semantic consistency. + +Furthermore, to optimize superpoint feature representations, we employ a stacked Graph Transformer [42] and Graph Convolutional Network [42] to capture the global context between superpoints. Similar to prior work, SPCNet is designed as an end-to-end oversegmentation network and can also be integrated with downstream tasks such as semantic segmentation. To fully evaluate SPCNet, we conduct experiments on four large-scale datasets: S3DIS [1], ScanNet [9], nuScenes [3], and SemanticKITTI [2]. Experimental results demonstrate that SPCNet achieves state-of-the-art performance in all core oversegmentation metrics (BR, BP, and F1) across all datasets. The contributions of this paper are as follows: + +- To our best knowledge, SPCNet is the first serialization based network for point cloud oversegmentation. +- We introduce a serialization strategy to initialize superpoints, which can efficiently produce superpoints with spontaneous spatial proximity. +- We propose an adaptive superpoint update algorithm to + +strengthen internal semantic consistency of superpoints. + +- Our method achieves remarkable performance in point cloud oversegmentation across multiple datasets. It can be integrated with semantic segmentation networks and shows promising performance. + +# 2. Related Work + +# 2.1. Deep Learning in Point Cloud + +Deep learning approaches for 3D point cloud primarily follow three paradigms: projection-based, voxel-based, and point-based methods. Projection-based methods project unordered points onto 2D planes for CNN processing [6, 20, 21, 32], sacrificing 3D geometry for computational efficiency. Voxel-based methods discretize space into grids for 3D convolution [25, 31], improved by sparse convolution techniques [8, 10, 35] to mitigate memory costs, though kernel size limitations remain a constraint. Point-based methods [24, 27, 28, 33, 43] process raw point clouds directly and have evolved toward transformer architectures [13, 29, 37, 40, 44]. These methods offer powerful representation capabilities while struggling with scalability. The recently proposed SPT [29] leverages hierarchical superpoints and sparse self-attention to efficiently process large-scale scenes, achieving superior semantic segmentation with reduced model size and computational cost compared to traditional approaches. + +# 2.2. Point Cloud Oversegmentation + +Existing point cloud oversegmentation methods can be broadly categorized into optimization-based methods and deep learning-based approaches. Early methods such as Papon et al. [26] proposed the VCCS method, which used FPFH as input features, selected seed points using an octree, and performed point cloud superpoint segmentation using the K-means clustering algorithm. Lin et al. [22] viewed the superpoint segmentation problem as a subset selection problem and employed subset selection methods to generate superpoints, still using FPFH as point cloud feature descriptors. Guinard et al. [12] simplified the superpoint segmentation problem into a graph cut problem, extracting features such as local linearity, flatness, divergence, and normal vectors from the point cloud, and applied a greedy graph cut algorithm to generate superpoints. Xiao et al. [39] defined the superpoint segmentation problem as an energy minimization problem and proposed a merging-exchange optimization framework to generate supervoxels. These methods often require complex and computationally expensive preprocessing steps, and their performance is limited by the insufficient expressiveness of handcrafted features. Recent deep learning methods overcome these limitations. + +Supervised Superpoint (SSP) method [18] pioneered deep network-based point cloud embedding, although it + +![](images/0c2b7bdd95d3c04ae8a3346b5d94122776995e0d362401c86b53c9a4031e675c.jpg) +(a) Point Feature Extractor +(b) Hierarchical Segment Initialization +(c) Adaptive Superpoint Update +Figure 2. The architecture includes three stages. First, point embeddings are extracted using a U-Net backbone with sparse convolutions in shallow layers and a Mamba bottleneck enhanced by Conditional Position Encoding (CPE). Second, the point cloud is serialized via a Hilbert curve to initialize segments for superpoint aggregation, with granularity at each level controlled by the hyperparameter $m$ . Finally, points are reassigned across segments to group semantically similar points into coherent superpoints, completing the clustering process. + +still depended on optimization techniques from [12]. SP-Net [15] further advanced the field by introducing learnable point-to-superpoint associations using FPS for center selection, enabling joint optimization in both coordinate and feature spaces. Furthermore, SuperLiDAR [16] improved efficiency by combining a BFS-based grouping algorithm with a local discriminative loss to generate compact superpoints. However, this method results in high computational complexity in spatial query operations during superpoint generation, especially for dense point cloud oversegmentation. + +# 2.3. Localization-based Method + +Recent works [5, 23, 34, 38] have introduced serialization-based approaches that differ from traditional point cloud processing by transforming unstructured point clouds into ordered sequences to preserve spatial proximity. For example, OctFormer [34] uses octree-based ordering similar to $z$ -ordering, offering scalability but limited by octree constraints. FlatFormer [23] employs window-based sorting to group point pillars, though its scalability in receptive field is limited, making it suitable for pillar-based 3D object detection. Building on these methods, PTv3 [38] extends attention scales by replacing KNN-based query with serialized neighborhoods, enhancing efficiency and scalability. + +# 3. Method + +Our oversegmentation network comprises three key steps: point embedding extraction (Sec. 3.1), hierarchical segment initialization (Sec. 3.2), and adaptive superpoint update (Sec. 3.3). The overall process is shown in Fig. 2. + +# 3.1. Backbone + +Optimization-based oversegmentation methods are constrained by the quality of handcrafted features, while learning based approaches similarly encounter limitations in deep feature representation. + +Sparse 3D U-Net. Let the input point cloud be $\mathcal{P} \in \mathbb{R}^{N \times (C + 3)}$ , where each of the $N$ points is represented by its features (e.g. normal vectors, RGB) and spatial coordinates $(x, y, z)$ . Following [10], we first voxelize $\mathcal{P}$ using the operator $\mathcal{V}$ , and then process the discretized data with a U-Net architecture based on sparse convolutions (SpConv). In the shallow part of the U-Net [30], SpConv layers capture fine-grained local geometric information. These operations produce feature representations that maintain high spatial resolution and extract detailed geometric structures crucial for accurate object boundary delineation. + +Bottleneck Reinforcement. Pure SpConv restricts over-segmentation effectiveness due to their inherently limited receptive field. We enhance our U-Net backbone by integrating Mamba [11, 45] at the bottleneck of the U-Net, as shown in Fig. 2 (a). After the shallow SpConv layers have abstracted local features, the Mamba module is applied to perform global context interaction. Mamba, with its linear-complexity global modeling capability, compensates for the local inductive bias of convolution and promotes semantic consistency across superpoints. + +# 3.2. Hierarchical Segment Initialization + +Methodology Justification. To efficiently partition point clouds into multiple segments, we propose a serialization- + +based partitioning method that utilizes the spatial locality preservation property of the Hilbert curve [38, 41] to divide the same point cloud into hierarchical structures of different granularities. Specifically, the Hilbert curve maps adjacent point clusters in 3D Euclidean space to contiguous positions in a 1D sequence, thereby simplifying the spatial proximity-based 3D partitioning task into an iterative partitioning of a 1D sequence. This transformation enables us to efficiently divide point clouds into compact segments in 3D space, significantly enhancing partitioning efficiency. This enables (1) efficient 1D clustering that avoids expensive 3D queries and clustering; and (2) a simplified oversegmentation process that decouples spatial locality from semantic similarity by performing only semantic comparisons within already spatially adjacent segments to complete clustering. + +Partitioning Strategy. Due to the variable receptive field of differently sized superpoints, smaller superpoints preserve more detailed local geometric information, whereas larger superpoints capture more comprehensive global semantic context. Therefore, we propose to construct a hierarchical partitioning structure to fully utilize multi-scale information. Given an input point cloud $\mathcal{P}$ and its corresponding features $F = \{f_i\}_{i=1}^N$ , where $f_i \in \mathbb{R}^d$ is generated by our backbone, we construct a hierarchical partitioning structure $\mathcal{H} = \{L_k\}_{k=1}^K$ . Each layer $L_k$ contains the complete serialized point cloud and is recursively quaternary partitioned into a set of segments $L_k = \{S_i^k\}_{i=0}^{M_k-1}$ where $S_i^k$ represents the $i$ -th segment, as shown in Fig. 2 (b). Specifically, the 3D point cloud is first serialized into a 1D sequence via Hilbert space-filling curve mapping [38], denoted as $\mathcal{P}_{seq} \in \mathbb{R}^{N \times d}$ . The sequence is then padded to $\mathcal{P}_{pad} \in \mathbb{R}^{N_{pad} \times d}$ where $N_{pad} = \lceil N / M_K \rceil \cdot M_K$ , ensuring divisibility by the finest segment count $M_K$ . The serialized point cloud $\mathcal{P}_{pad}$ is recursively partitioned using 1D quaternary splitting, such that at each layer $L_k$ , it is uniformly divided into $M_k = 2^{m - 2k + 2}$ segments. For each layer $L_k$ , $\mathcal{S}_i^k$ can be represented as: + +$$ +\mathcal {S} _ {i} ^ {k} = \left\{\mathcal {P} _ {p a d} [ j ] \mid j \in [ i \cdot l _ {k}, (i + 1) \cdot l _ {k} - 1 ] \right\}, \tag {1} +$$ + +where $l_{k} = N_{pad} / M_{k}$ represents the segment length for each segment in the $k$ -th layer. Recursive partitioning establishes fixed parent-child indexing between layers. For any inter-layer gap $l \geq 1$ , the hierarchical inclusion relationship is defined as: + +$$ +\mathcal {S} _ {i} ^ {k} = \bigcup_ {j \in [ i \cdot 4 ^ {l}, (i + 1) \cdot 4 ^ {l} - 1 ]} \mathcal {S} _ {j} ^ {k + l}. \tag {2} +$$ + +At the finest layer $L_{K}$ , segment features are generated through permutation-invariant aggregation of constituent points: + +$$ +f _ {S _ {i} ^ {K}} = \mathcal {A} \left(f _ {j} \mid p _ {j} \in S _ {i} ^ {K}\right) \in \mathbb {R} ^ {C}, \tag {3} +$$ + +where $\mathcal{A}:\mathbb{R}^{n_i\times C}\to \mathbb{R}^C$ denotes a permutation-invariant operator (e.g. max-pooling) with $n_i = |\mathcal{S}_i^K | = N_{pad} / M_K$ . + +# 3.3. Adaptive Superpoint Update + +Given the structure $\mathcal{H}$ , the goal is to refine the partition $L_{k} = \{S_{1}^{k},\ldots ,S_{M_{k}}^{k}\}$ with features $f_{S_i^k}\in \mathbb{R}^d$ into semantically consistent superpoints $L_k^\prime = \{S_1'^k,\dots,S_M'^k\}$ with precise geometric boundaries and updated features $f_{S_i^k}^\prime \in \mathbb{R}^d$ . This is achieved by dynamically allocating raw points within coarse-grained segments to their appropriate sub-segments based on semantic similarity, a process that redistributes raw points within segments to fit object boundaries. This process can be transformed into a boundary fitting problem between adjacent segments in 1D sequence point clouds, as shown in Fig. 2 (c). + +Feature Similarity based Affiliation. To establish the affiliation relationship between raw points within each segment and nearby segments, we consider two adjacent levels in $\mathcal{H}$ : the finest $L_{k} = \{S_{1}^{k}, S_{2}^{k}, \ldots, S_{M_{k}}^{k}\}$ and its coarser predecessor $L_{k-1} = \{S_{1}^{k-1}, S_{2}^{k-1}, \ldots, S_{M_{k-1}}^{k-1}\}$ (we only use $l = 1$ defined in Eq. (2)). The hierarchical relationship between these levels is given by (1) $M_{k-1} = \frac{M_{k}}{4}$ , meaning each coarser level has one-fourth the segments of the finer level and (2) Each segment $S_{i}^{k-1}$ at level $L_{k-1}$ contains four consecutive segments from level $L_{k}$ : $S_{i}^{k-1} = S_{4i}^{k} \cup S_{4i+1}^{k} \cup S_{4i+2}^{k} \cup S_{4i+3}^{k}$ . + +For each raw point $p_i \in S_i^{k-1}$ , we identify the most similar initialized segment feature from its four containing segments $S_{4i}^k, S_{4i+1}^k, S_{4i+2}^k$ , and $S_{4i+3}^k$ . Meanwhile, to expand the receptive field of raw points relative to their potential superpoints, we incorporate segment features from eight sub-segments at $L_k$ belonging to the left and right adjacent segments $S_{i-1}^{k-1}$ and $S_{i+1}^{k-1}$ , forming a local feature set $\{f_{S_{4i+j}^k} \mid j \in \{-4, -3, \dots, 7\}\}$ , resulting in a set of 12 initial segment features. For each segment $S_i^{k-1}$ with $N_{k-1} = 4N_{pad} / M_k$ points, we compute a similarity matrix $\mathbf{A}_i \in \mathbb{R}^{N_{k-1} \times 12}$ . Each entry $a_{j,t}$ measures the similarity between the $j$ -th point and the $t$ -th segment feature in the above set of segment features: + +$$ +a _ {j, t} = \frac {f _ {p _ {j}} \cdot f _ {S _ {t} ^ {k}}}{\| f _ {p _ {j}} \| \| f _ {S _ {t} ^ {k}} \|}, \quad j \in [ 1, N _ {k - 1} ], \quad t \in [ 1, \tau ], \tag {4} +$$ + +This process generates similarity matrices $\{\mathbf{A}_i\}_{i=1}^{M_{k-1}}$ for the entire level $L_{k-1}$ . Point-to-superpoint assignments are determined by selecting the superpoint with the highest similarity for each point, which is then mapped to global superpoint indices $\phi(p_j) \in [1, M_k]$ . The resulting assignment index set $Ass_i = \{\phi(p_j)\}_{j=1}^n$ provides a key reference for updating superpoint features in subsequent steps. + +Cross-Attention based Update. Cross-Attention [4] enables dynamic and adaptive relationships between superpoints and their constituent points. Through feature similarity based affiliation, precise similarity relationships are established, guiding accurate superpoint feature updates. For + +![](images/0db175f23efedc3fe081828f140afd4176b41311856bb7bb2c483578fd4f6e5b.jpg) +Figure 3. The end-to-end semantic segmentation pipeline constructs a graph over superpoints, captures global context using graph transformers, and integrates point-level embeddings for final semantic prediction, where $\oplus$ denotes simple feature addition. + +the assignment indices $A s s i = \{\phi(p_{j})\}_{j=1}^{n}$ , the inverse mapping $\phi^{-1}(i) = \{p_{j} \mid \phi(p_{j}) = i\}$ represents the original points contained within superpoint $S_{i}^{\prime k}$ . Based on $A s s i$ , each superpoint's feature is updated solely from its member points. Using superpoint feature $f_{S_{i}^{\prime k}}^{\prime}$ as the Query and its member point features $\{f_{p_{j}} \mid p_{j} \in \phi^{-1}(i)\}$ as Keys and Values, the update formula is: + +$$ +f _ {S _ {i} ^ {k}} ^ {\prime} = \text {C r o s s A t t n} \left(\underbrace {f _ {S _ {i} ^ {k}}} _ {\text {Q u e r y}}, \underbrace {\left\{f _ {p _ {j}} \mid p _ {j} \in \phi^ {- 1} (i) \right\}} _ {\text {K e y / V a l u e}}\right), \tag {5} +$$ + +Attention weights are computed through scaled dot product, allowing each superpoint to focus on the most informative points within its local neighborhood: + +$$ +\operatorname {C r o s s A t t n} \left(f _ {S _ {i} ^ {k}}, \left\{f _ {p _ {j}} \right\}\right) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {i} \mathbf {K} _ {i} ^ {\top}}{\sqrt {d}}\right) \mathbf {V} _ {i}, \tag {6} +$$ + +where $\mathbf{Q}_i = f_{S_i^k}W_K\in \mathbb{R}^{1\times d}$ , $\mathbf{K}_i = \mathrm{Concat}(\{f_{p_j}W_K\mid p_j\in \phi^{-1}(i)\})\in \mathbb{R}^{n_i^k\times d}$ , $\mathbf{V}_i = \mathrm{Concat}(\{f_{p_j}W_V\mid p_j\in \phi^{-1}(i)\})\in \mathbb{R}^{n_i^k\times d}$ , $n_i^k = |\phi^{-1}(i)|$ is the number of points contained in each superpoint, and $W_{Q},W_{K},W_{V}\in \mathbb{R}^{d\times d}$ are learnable projection matrices. For coarse-grained superpoints, we apply the same update logic, with interactions between fine-grained features and their coarse-grained initializations. + +# 3.4. Superpoint Interaction + +The superpoint features $f_{S_i^k}^\prime$ encode geometric-semantic information for locally homogeneous points. To leverage these features for semantic segmentation, inspired by SPT [29], we integrate a superpoint interaction module with our oversegmentation network, forming an end-to-end + +framework, as shown in Fig. 3. In our hierarchical design, fine-grained superpoints capture local details while coarse grained ones represent object-level semantics. We construct multi-level superpoint graphs and employ a cascaded Transformer-GCN [42] to facilitate cross-superpoint interactions and global context modeling. In the Graph Transformer, a superpoint graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ is built by linking each superpoint with its $k$ nearest neighbors in feature space, where $\mathcal{V} = \{f_{S_i^k}'\}_{i=1}^{M_k}$ . Edge features are defined via relative position encoding as $\mathbf{e}_{ij} = g(p_j - p_i)$ , with $p_i$ and $p_j$ representing node coordinates and $g$ being a linear mapping. A subsequent GCN processes this graph to enhance local feature interactions through message passing among adjacent superpoints. Finally, we fuse backbone point features with their corresponding multi-level superpoint before feeding them into the segmentation head. + +# 3.5. Superpoint Aggregation Loss + +We propose a superpoint aggregation loss to optimize superpoint generation, inspired by LMNN [36]. This comprehensive loss consists of three components. First, a compactness loss $\ell_{\mathrm{compact}}$ enforces intra-superpoint feature cohesion by minimizing point-to-superpoint distances. Second, a distinction loss $\ell_{\mathrm{dist}}$ encourages inter-class separability by maximizing the distances between superpoints of different classes, where each superpoint's label is determined by majority voting among its points. Third, a purity loss $\ell_{\mathrm{purity}}$ maintains label consistency within each superpoint by penalizing high entropy in its class distribution. The overall loss function is defined as: + +$$ +\begin{array}{l} \ell_ {s p} = \frac {1}{N _ {s}} \sum_ {i = 1} ^ {N _ {s}} \frac {\sum_ {p \in S _ {i}} \| f _ {i} - f _ {S _ {i} ^ {k}} ^ {\prime} \| ^ {2}}{| S _ {i} |} + \\ \frac {1}{N _ {c}} \sum_ {i, j \in c, i \neq j} \max \left(0, \delta_ {\text {d i s t}} - \| f _ {S _ {i} ^ {k}} ^ {\prime} - f _ {S _ {j} ^ {k}} ^ {\prime} \|\right) ^ {2} + \tag {7} \\ \frac {1}{N _ {s}} \sum_ {i = 1} ^ {N _ {s}} \sum_ {k = 1} ^ {K} p _ {i, k} \log (p _ {i, k}), \\ \end{array} +$$ + +where $f_{i}$ denotes the feature of point $p$ , $f_{S_i^k}^\prime$ represents the feature of superpoint $S_{i}^{\prime k}$ , $\delta_{\mathrm{dist}}$ is a predefined separation margin, $p_{i,k}$ is the probability of superpoint $i$ belonging to class $k$ , $N_{s}$ is the number of superpoints, and $N_{c}$ is the number of superpoint pairs involved in $\ell_{\mathrm{dist}}$ . + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. We evaluate SPCNet on four datasets with diverse characteristics. S3DIS [1] contains $274M$ points from six office areas, with Area 5 as the test set. ScanNet [9] includes 1,513 indoor scans from 707 scenes with + +![](images/68583646c2c86e7d1defb0fe1e6c589c815ca02d12e9b76b343157e2cfa2954a.jpg) +Figure 4. The visualization of superpoints generated on the ScanNet (first row) and S3DIS (second row) datasets shows a two-level hierarchy: $L_{1}$ with 1024 superpoints and $L_{2}$ with 256 superpoints. Finer $L_{1}$ superpoints merge based on semantic similarity to form larger $L_{2}$ superpoints. For example, in S3DIS, $L_{1}$ superpoints of table parts merge into a single table superpoint, and $L_{1}$ superpoints of whiteboard sections form a complete whiteboard superpoint. Our superpoints are regular and orderly, with nearly rectangular shapes in semantically consistent areas (e.g. floors, tables), maintaining continuity, precise object boundaries, and demonstrating superior regularity, compactness, and purity. + +20 categories, averaging 148k points/scan. nuScenes [3] comprises 1,000 scenes from 32-beam LiDAR, split 750/150/150 for train/val/test, with 17 categories. SemanticKITTI [2] is an outdoor self-driving benchmark with 22 sequences and 20 categories, using sequences 00-10 (densely annotated) for training and 11-22 for testing. + +Evaluation Metrics. Following SPNet [15] and SSP [18], we use Oracle Overall Accuracy (OOA), Boundary Recall (BR), Boundary Precision (BP), and F1 score to evaluate superpoints. BR and BP assess boundary quality, while OOA represents the upper bound of semantic segmentation accuracy using superpoints. The F1 score balances recall and precision, defined as $F1 = \frac{2\cdot BP\cdot BR}{BP + BR}$ . For semantic segmentation in indoor and outdoor scenarios, we adopt mean intersection over union (mIoU). + +# 4.2. Model Configuration + +We maintain a consistent architecture across all experiments with minimal dataset-specific adjustments. Our backbone features a four-stage symmetric encoder-decoder structure, each with a single block depth. The encoder has embedding dimensions [32, 64, 128, 256] and the decoder [128, 64, 64], with Mamba blocks using conditional positional encoding at the 256 stage. We implement a three-level hierarchy $(m = 8, K = 2, l = 1)$ , organizing points into n, 1024, 256 segments. Training is performed on dual NVIDIA RTX 4090 GPUs with the Adam optimizer. + +# 4.3. Point Cloud Oversegmentation + +Quantitative Results. As shown in Tab. 1, we evaluate SPCNet on three datasets, reserving ScanNet for ablation studies, ensuring fair comparison with consistent superpoint counts. For example, in S3DIS Area 5, methods like VCCS, Lin et al., SPG, SSP, and SPNet generate around 1050 su + +
MethodBRBPF1OOA
S3DIS Area 5 [1]
VCCS [26]62.0610.8618.4895.22
SPG [19]52.0613.1120.9495.84
SSP [18]80.7313.0222.4297.04
SPNet [15]84.7513.1422.6596.50
SPCNet(Ours)89.5026.7840.5096.81
nuScenes [3]
SPG [19]28.2017.2621.4189.22
SSP [18]22.0415.6318.2892.01
SPNet [15]68.9218.3428.9787.67
SuperLiDAR [16]74.7225.1237.5996.31
SPCNet(Ours)88.7732.8247.5296.33
SemanticKITTI [2]
SPG [19]25.6415.8219.5686.86
SSP [18]18.7410.6713.5992.27
SPNet [15]56.5214.7823.4392.24
SuperLiDAR [16]65.5220.5231.2596.21
SPCNet(Ours)73.1830.1542.5796.84
+ +Table 1. Comparison results of generated superpoints on the S3DIS, nuScenes, and SemanticKITTI datasets. + +perpoints, matching our 1024 Hilbert partitions. Our deep learning-based approach outperforms traditional methods, benefiting from the specialized backbone network's ability to extract rich geometric features that capture complex local structures. Unlike methods such as SPNet [15], which rely on FPS+KNN for presampling and soft assignment map- + +ping without limiting the scope of raw points being clustered, resulting in blurry boundaries, our approach leverages the Hilbert curve's locality, limiting the range of raw point clustering (preventing clustering of semantically similar but spatially distant points), and finally refining superpoint through the Cross-Attention, minimizing noise and maintaining precise, simple boundaries, as shown in Fig. 5 (d) and (e). Visualization results in Fig. 4 show that our method tends to aggregate superpoints within object interiors while maintaining distinct object boundaries, especially in challenging cases like wall-whiteboard interfaces, resulting in higher BR and BP. Additionally, our method excels at merging semantically continuous regions (e.g. floors, tabletops), leading to fewer incorrect superpoint boundaries and substantially higher BP metrics. Fig. 6 shows the performance of different methods on SemanticKITTI. + +
MethodDeviceInference Time (ms) data proc. / backbone + SP ini. + SP gen.
SPG [19]CPU13152
SSP [18]RTX 309016115 (11812 + 0 + 13426)
SPNet [15]RTX 309012182 (11812 + 0 + 1154)
SuperLiDAR [16]RTX 309072(0 + 0 + 72)
SPCNet (ours)RTX 4090, RTX 306012 (6 + 0.4 + 6), 61
+ +Table 2. The inference time of different oversegmentation methods on SemanticKITTI. The inference time consists of data preprocessing/backbone (red/blue), superpoint initialization (green), and superpoint generation (purple). + +Time Costs. We evaluated oversegmentation network efficiency by measuring inference time per scan on SemanticKITTI validation (batch size=1) in Tab. 2. For optimization based SPG [19], we followed SuperLiDAR's [16] configuration using Core i5 CPU. For learning-based methods (SSP [18], SPNet [15], SuperLiDAR), we used SuperLiDAR's results (RTX 3090). Our SPCNet was benchmarked on RTX 4090 and 3060 GPUs. Our method achieved total inference times of 12 ms (RTX 4090) and 61 ms (RTX 3060), achieving $100 \times$ speedup versus SPG/SSP/SPNet, and $5 \times$ versus SuperLiDAR. Even on RTX3060, our inference is 10ms faster than SuperLiDAR, indicating at least $3 \times$ efficiency accounting for hardware differences. This stems from: (1) efficient superpoint initialization (0.4 ms) that compresses FPS+KNN sampling time in SPNet; (2) the application of a localized attention mechanism within the superpoint refinement (6 ms). Our backbone architecture allows further acceleration using lightweight alternatives like PointNet[27]. Traditional methods face key bottlenecks: SPG/SSP rely on optimization-based generation (e.g. graph cuts), SSP/SPNet suffer from redundant processing due to mixed handcrafted/learned features, and SuperLiDAR's BFS-based grouping has complexity dependent on scene density, potentially reaching $O(n^{2})$ in dense scenes, versus our serialization based approach $(O(n))$ . + +Ablation Studies. Mamba mitigates the local inductive bias + +![](images/7c95b7aa84caa670107ecd0e17bf52de6eeb5ec47fad76ad32dfd6f262f22ce0.jpg) +(a) 3D-UNet only with $k = 2$ + +![](images/f89d9e30935e1b0368a2e56f53d8e28ffdfc4fcad1d6f2ecff826a3e30b4715d.jpg) +(b) Bottleneck reinforcement with $k = 2$ + +![](images/f8f9808ff53a2c30e00548eac0e07af490314ce6ba87be1c2ee4af92c4efca42.jpg) +(d) PointNet with $k = 2$ +Figure 5. Oversegmentation comparisons. (a) Sparse 3D U-Net with neighbor expansion of 2. (b) Adding the Mamba module to the bottleneck. (c) Replacing the backbone with a simplified PointNet, as in SPNet. (d) SPNet results. Notably, while (b) configuration does not fully aggregate semantically consistent objects into complete superpoints, it notably enhances consistency. SPC-Net also avoids clustering semantically similar but spatially distant points, ensuring clearer superpoint boundaries. + +![](images/28a5561d25924c26d780d6c78c4fe37fc4e4b55d5c8510c0150c8779447fba24.jpg) +(e) SPNet + +
Backbone SettingBR(%)BP(%)OOA(%)mIou(%)
3D U-Net76.33 ↓7.9028.45 ↓1.7090.93 ↓1.1974.10 ↓2.86
+ Mamba fully78.37 ↓5.8629.13 ↓1.0291.24 ↓0.8876.18 ↓0.78
+ Mamba bottleneck (default)84.2330.1592.1276.96
+ +Table 3. Performance comparison of different backbone architectures on the ScanNet validation set. + +
Neighbor ScopeBR(%)BP(%)OOA(%)
No expansion79.86 ↓4.3728.57 ↓1.5891.22 ↓0.90
2 expansion (default)84.2330.1592.12
4 expansion83.67 ↓0.5629.85 ↓0.3091.24 ↓0.88
6 expansion79.93 ↓4.3028.65 ↓1.5089.93 ↓2.19
+ +Table 4. The effect of expanding neighbor segment range on superpoint performance on the ScanNet validation set. + +of convolutions in oversegmentation, improving superpoint continuity and semantic consistency. As shown in Fig. 5 (a) and (b), this bias causes superpoints in the same semantic region to align with boundaries but fragment due to internal inconsistency. Thus, we retain convolutions in U-Net's shallow layers for local geometric cues to ensure boundary precision, and introduce Mamba at the bottleneck to capture long range semantics. Tables 3 and 4 analyze backbone design and neighbor expansion effects on superpoint quality. We evaluated three backbone configurations: (1) baseline 3D sparse U-Net only, (2) fully Mamba-based U-Net, and (3) our default hybrid design with Mamba modules in the bottleneck layers. Results demonstrate that the baseline 3D sparse U-Net reduces BR, BP, and OOA by $7.90\%$ , $1.70\%$ , and $1.19\%$ respectively compared to our default. Similarly, the fully Mamba-based U-Net underperforms, as Mamba's global context modeling in shallow networks with low-dimensional features smooths out high-frequency boundary information. Our hybrid "shallow convolution + deep Mamba" architecture achieves optimal per + +![](images/8de302899f4f95dfd852cc00ec70e2f91c49f12bc3ceeae25552303e13887394.jpg) +Figure 6. Performance of different methods on the SemanticKITTI validation set under varying numbers of superpoints. It is worth noting that our method can achieve better results than other methods with fewer superpoints (256, 512, and 1024). + +![](images/4bafd1e0708c5c25e62a082e0384fdac660e6d5ea006be9e022455b37d4fc07b.jpg) + +![](images/9a03200b16ee743058f2e9b76a44c63aa914a8d384d7a320257d3e5de66549db.jpg) + +![](images/1836726bac8ed4d909e7809700771b8b5243b4cb9e1cf101822911af447133d4.jpg) + +
ExperimentScanNet [9]S3DIS [1]nuScenes [3]Sem.KITTI [2]
Best ModelValTestArea5ValTestValTest
†SPG [19]--58.0----
MinkUNet [8]72.273.665.473.3-63.8-
†SPG + SSP [19]--61.7----
ST [17]74.373.772.0----
†SPT [29]--68.9----
PointNeXt [28]71.571.270.5----
OctFormer [34]75.776.6-76.177.264.367.8
Swin3D [40]76.4-72.5----
AF2S3Net [7]---62.278.074.270.8
†SuperLiDAR [16]----78.5-69.6
PTv2 [37]75.474.271.680.282.670.372.6
PTv3 [38]77.577.973.480.482.770.874.2
Backbone76.9-69.375.4-67.6-
†SPCNet(Ours)77.878.072.177.680.269.571.9
+ +formance by leveraging convolutions to preserve local details in shallow layers while Mamba modules handle global context modeling in deeper layers. + +In the neighborhood expansion experiments (analogous to k-parameter tuning in KNN), we observed that moderately expanding the candidate superpoint pool of original points during the feature similarity based affiliation process significantly improves sampling accuracy. The model achieves optimal performance when $\mathrm{K} = 2$ ; excessive expansion $(\mathrm{K} = 4,6)$ leads to performance degradation. This may be because larger receptive fields introduce noise, and the hard assignment approach introduces more uncertainty as the number of candidate targets increases, causing early assignment errors to propagate to later processing stages. + +# 4.4. Semantic Segmentation + +Tab. 5 evaluates our approach on four indoor and outdoor datasets, comparing it with point-based and superpoint-based methods. Our optimized oversegmentation strategy shows significant gains: on S3DIS Area 5, we achieve a $3.2\%$ mIoU improvement over SPT. Compared to state-of-the-art point-based methods, our approach remains competitive, outperforming PTv3 by $0.3\%$ on ScanNet $(77.8\%)$ and PTv2 by $0.5\%$ on S3DIS Area 5. Additionally, we achieve $71.9\%$ and $80.2\%$ on Sem.KITTI and nuScenes, improving + +Table 5. Semantic segmentation results on the ScanNet, S3DIS, nuScenes, and SemanticKITTI datasets. †Superpoint-based. + +
SettingComponentsModel Size (m)mIoU(%)Δ(%)
Backbone + SP gen. + SP int.
backbone (default)-18.876.96-
Hierarchical Enhancement
1 partition level+ Level1 Superpoint0.24 (0.218 + 0.022)77.50+0.54
2 partition level+ Level2 Superpoint0.02 (0.014 + 0.006)77.80+0.30
+ +Table 6. Semantic segmentation improvement with hierarchical superpoint features in ScanNet validation set. SP gen. means superpoint generation; SP int. means superpoint interaction. + +SuperLiDAR by $1.3\%$ and $1.7\%$ , respectively. These results confirm that our superpoints capture local and global semantic information effectively, while the simple backbone, enhanced by superpoint features, matches the performance of more complex point-based methods, validating the superior quality of the generated superpoints. + +As shown in Tab. 6, integrating hierarchical superpoint features with per-point features extracted by the backbone improved segmentation. First-level superpoints aggregated local semantics, boosting intra-class consistency with a $0.54\%$ mIoU gain. Second-level features captured long-range dependencies, adding another $0.30\%$ improvement. These hierarchical superpoints acted as adaptive receptive fields, enhancing performance with only $0.24M$ parameters. + +# 5. Conclusion + +In this paper, we propose a novel serialization based oversegmentation method to address the challenge of efficiently grouping semantically consistent points into superpoints while maintaining spatial proximity. SPCNet achieves state-of-the-art performance in point cloud oversegmentation while being $3 \times$ faster than existing methods in inference speed. Extensive experimental results demonstrate the effectiveness and efficiency of SPCNet. Additionally, SPCNet can be flexibly applied to semantic segmentation tasks, effectively improving the accuracy of semantic segmentation. + +Acknowledgments This work was supported in part by the National Natural Science Foundation of China (Grant No. 42201475). + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 8 +[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2, 6, 8 +[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 8 +[4] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4 +[5] Wanli Chen, Xinge Zhu, Guojin Chen, and Bei Yu. Efficient point cloud analysis using hilbert curve. In European Conference on Computer Vision, pages 730-747. Springer, 2022. 3 +[6] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2 +[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 8 +[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 8 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 8 +[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2, 3 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3 +[12] Stéphane Guinard and Loic Landrieu. Weakly supervised segmentation-aided classification of urban scenes from 3d + +lidar point clouds. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42:151-157, 2017. 1, 2, 3 + +[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2 +[14] David Hilbert and David Hilbert. Über die stetige abbildung einer linie auf ein flächenstück. Dritter Band: Analysis-Grundlagen der Mathematik-Physik Verschiedenes: Nebst Einer Lebensgeschichte, pages 1–2, 1935. 2 +[15] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 1, 3, 6, 7 +[16] Le Hui, Linghua Tang, Yuchao Dai, Jin Xie, and Jian Yang. Efficient lidar point cloud oversegmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18003-18012, 2023. 2, 3, 6, 7, 8 +[17] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8500-8509, 2022. 8 +[18] Loic Landrieu and Mohamed Boussaha. Point cloud over-segmentation with graph-structured deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7440–7449, 2019. 1, 2, 6, 7 +[19] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 1, 6, 7, 8 +[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2 +[21] Bo Li, T Zhang, and T Xia. Vehicle detection from 3d lidar using fully convolutional network. arxiv 2016. arXiv preprint arXiv:1608.07916. 2 +[22] Yangbin Lin, Cheng Wang, Dawei Zhai, Wei Li, and Jonathan Li. Toward better boundary preserved supervoxel segmentation for 3d point clouds. *ISPRS journal of photogrammetry and remote sensing*, 143:39–47, 2018. 1, 2 +[23] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1200-1211, 2023. 3 +[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 2 +[25] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. + +In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2 +[26] Jeremie Papon, Alexey Abramov, Markus Schoeler, and Florentin Worgotter. Voxel cloud connectivity segmentation-supervoxels for point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2027-2034, 2013. 1, 2, 6 +[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 7 +[28] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 2, 8 +[29] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 2, 5, 8 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 +[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2 +[32] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 2 +[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2 +[34] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023. 3, 8 +[35] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2 +[36] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. Advances in neural information processing systems, 18, 2005. 5 +[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 8 + +[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 2, 3, 4, 8 +[39] Yanyang Xiao, Zhonggui Chen, Zhengtao Lin, Juan Cao, Yongjie Jessica Zhang, Yangbin Lin, and Cheng Wang. Merge-swap optimization framework for supervoxel generation from three-dimensional point clouds. Remote Sensing, 12(3):473, 2020. 2 +[40] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 2, 8 +[41] Guowen Zhang, Lue Fan, Chenhang He, Zhen Lei, Zhaoxiang Zhang, and Lei Zhang. Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700, 2024. 4 +[42] Peiyan Zhang, Yuchen Yan, Xi Zhang, Chaozhuo Li, Senzhang Wang, Feiran Huang, and Sunghun Kim. Transgnn: Harnessing the collaborative power of transformers and graph neural networks for recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1285-1295, 2024. 2, 5 +[43] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2 +[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2 +[45] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3 \ No newline at end of file diff --git a/2025/Serialization based Point Cloud Oversegmentation/images.zip b/2025/Serialization based Point Cloud Oversegmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10c42ee48d7ad9a3ac6de7efbeefb33c2e578746 --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7bc7be2ff3ea49e3335878327a716940c331be856cb550189f11d5d7fed25cf +size 652021 diff --git a/2025/Serialization based Point Cloud Oversegmentation/layout.json b/2025/Serialization based Point Cloud Oversegmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..347f5d698d3135067eea9eb0c55f4ac3b566b4cd --- /dev/null +++ b/2025/Serialization based Point Cloud Oversegmentation/layout.json @@ -0,0 +1,9341 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 462, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 462, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 462, + 121 + ], + "type": "text", + "content": "Serialization based Point Cloud Oversegmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "spans": [ + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": "Chenghui Lu" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": ", Jianlong Kwan" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": ", Dilong Li" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": ", Ziyi Chen" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": ", Haiyan Guan" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": "Department of Computer Science & Xiamen CVPR Key Laboratory, Huaqiao University \n" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": "Fujian Key Lab. of Big Data Intell. and Security, Huaqiao University \n" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 142, + 521, + 199 + ], + "type": "text", + "content": "Nanjing University of Information Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 201, + 518, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 201, + 518, + 213 + ], + "spans": [ + { + "bbox": [ + 88, + 201, + 518, + 213 + ], + "type": "text", + "content": "{23013083012, jianlongkwan}@stu.hqu.edu.cn, {scholar.dll, chenziyihq}@hqu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 241, + 211, + 330 + ], + "blocks": [ + { + "bbox": [ + 58, + 241, + 211, + 330 + ], + "lines": [ + { + "bbox": [ + 58, + 241, + 211, + 330 + ], + "spans": [ + { + "bbox": [ + 58, + 241, + 211, + 330 + ], + "type": "image", + "image_path": "b3c467364161f46c290c1b5ef66f8c23f4391459826ef532d4ad5edaa5eafa3e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. Our method efficiently generates superpoints from serialized point clouds by optimizing boundaries through reallocating points from each segment to the most appropriate surrounding segments (shown in green and red)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 211, + 241, + 310, + 329 + ], + "blocks": [ + { + "bbox": [ + 211, + 241, + 310, + 329 + ], + "lines": [ + { + "bbox": [ + 211, + 241, + 310, + 329 + ], + "spans": [ + { + "bbox": [ + 211, + 241, + 310, + 329 + ], + "type": "image", + "image_path": "85afddb3e8aa4b56c4856d653b2d0b8337afbc9152978dfd116f8e3ece568fc0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 241, + 381, + 329 + ], + "blocks": [ + { + "bbox": [ + 310, + 241, + 381, + 329 + ], + "lines": [ + { + "bbox": [ + 310, + 241, + 381, + 329 + ], + "spans": [ + { + "bbox": [ + 310, + 241, + 381, + 329 + ], + "type": "image", + "image_path": "d05036b13904f0b04924e72e7f0bb618979d7aa84786014eac8c191494dbaaf8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 381, + 241, + 467, + 329 + ], + "blocks": [ + { + "bbox": [ + 381, + 241, + 467, + 329 + ], + "lines": [ + { + "bbox": [ + 381, + 241, + 467, + 329 + ], + "spans": [ + { + "bbox": [ + 381, + 241, + 467, + 329 + ], + "type": "image", + "image_path": "0157efd9819ccf3663eaae73f4541ffb220b738e3fd44c49a054840cd7378601.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 241, + 552, + 329 + ], + "blocks": [ + { + "bbox": [ + 467, + 241, + 552, + 329 + ], + "lines": [ + { + "bbox": [ + 467, + 241, + 552, + 329 + ], + "spans": [ + { + "bbox": [ + 467, + 241, + 552, + 329 + ], + "type": "image", + "image_path": "df7619d1151017d0ae3a45d7782ec24fdbb52dd80514bf26339af008ab9110d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 370, + 200, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 370, + 200, + 383 + ], + "spans": [ + { + "bbox": [ + 151, + 370, + 200, + 383 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 389, + 297, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 297, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 297, + 653 + ], + "type": "text", + "content": "Point cloud oversegmentation, as a fundamental preprocessing step for 3D understanding, is a challenging task due to its spatial proximity and semantic similarity requirements. Most existing works struggle to efficiently group semantically consistent points into superpoints while maintaining spatial proximity. In this paper, we propose a novel serialization based point cloud oversegmentation method, which leverages serialization to avoid complex spatial queries, directly accessing neighboring points through sequence locality for similarity matching and superpoint clustering. Specifically, we first serialize point clouds into a Hilbert curve and spatially-continuously partition them into initial segments. Then, to guarantee the internal semantic consistency of superpoints, we design an adaptive update algorithm that clusters superpoints by matching feature similarities between neighboring segments and refines segment features via Cross-Attention. Experiments on large-scale indoor and outdoor datasets demonstrate state-of-the-art performance in point cloud oversegmentation. Moreover, it is also adaptable to semantic segmentation and achieves promising performance. The code is available at https://github.com/CHL-glitch/SPCNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Point cloud oversegmentation partitions point clouds into multiple spatially and semantically homogeneous regions," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 372, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 454 + ], + "type": "text", + "content": "termed superpoints, which adaptively represent a set of points. Due to their representativeness, superpoint-based methods can significantly reduce computational complexity and improve semantic consistency. Therefore, it is becoming increasingly important in the field of point cloud processing and has attracted considerable attention from many researchers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 459, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 638 + ], + "type": "text", + "content": "Early point cloud oversegmentation approaches primarily combined hand-crafted features with clustering or graph partitioning to generate superpoints. Papon et al. [26] introduced VCCS, which employed Fast Point Feature Histograms, selected seed points via octree structures, and performed K-means clustering. Lin et al. [22] reformulated the task as a subset selection problem while still using FPFH descriptors. Guinard et al. [12] simplified the process into a graph cut problem by extracting geometric features before applying a greedy algorithm. Similarly, Landrieu et al. [19] partitioned superpoints by solving an energy minimization problem with graph-based simplicity penalties. However, handcrafted features have limited representation, and the non-differentiable superpoint assignments prevent end-to-end training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "Recent point cloud oversegmentation methods generate superpoints using deep networks. Landrieu et al. [18] first used deep networks to extract point embeddings instead of handcrafted features, but their framework is not end-to-end. SPNet [15] proposed an end-to-end superpoint network that achieved clustering by iteratively constructing a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25831" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "point-superpoint association map. However, their Farthest Point Sampling (FPS) based k-nearest neighbor (KNN) association map faced efficiency bottlenecks, and the unlimited range of original points being clustered resulted in irregular superpoint boundaries and even introduced noise. SuperLiDAR [16] combined Breadth-First Search (BFS) grouping with local discriminative loss to generate compact superpoints, and while it achieved " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "O(V + E)" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " complexity on sparse LiDAR data, it may approach " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "O(V^2)" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " in dense scenarios, limiting scalability. Those point-based methods design complex algorithms to cluster superpoints and/or inefficiently balance semantic consistency and spatial proximity, resulting in high computational burden." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 229, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 479 + ], + "type": "text", + "content": "To address this issue, we propose an efficient serialization based point cloud oversegmentation method. Inspired by PTv3 [38], we adopt a serialization approach instead of traditional point-based methods to avoid high computational complexity spatial queries such as FPS and KNN, as shown in Fig. 1. We find that one of the representative serialization strategies, the Hilbert curve [14], which has excellent locality-preserving properties, is crucial for spatial proximity in oversegmentation tasks. Thus, we apply the Hilbert curve to serialize point clouds and initialize superpoints. To achieve internal semantic consistency for the superpoints, we design an adaptive superpoint update algorithm that clusters semantically similar points within prepartitioned segments to update superpoints. Specifically, we use cosine similarity to evaluate the semantic similarity between points and nearby segments, which serves as the metric for associating points and superpoints to reallocate points, then apply Cross-Attention to update superpoint features. By combining these two steps and iterating them, the algorithm can obtain superpoints with both spatial proximity and semantic consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 482, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 636 + ], + "type": "text", + "content": "Furthermore, to optimize superpoint feature representations, we employ a stacked Graph Transformer [42] and Graph Convolutional Network [42] to capture the global context between superpoints. Similar to prior work, SPCNet is designed as an end-to-end oversegmentation network and can also be integrated with downstream tasks such as semantic segmentation. To fully evaluate SPCNet, we conduct experiments on four large-scale datasets: S3DIS [1], ScanNet [9], nuScenes [3], and SemanticKITTI [2]. Experimental results demonstrate that SPCNet achieves state-of-the-art performance in all core oversegmentation metrics (BR, BP, and F1) across all datasets. The contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "content": "- To our best knowledge, SPCNet is the first serialization based network for point cloud oversegmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 701 + ], + "type": "text", + "content": "- We introduce a serialization strategy to initialize superpoints, which can efficiently produce superpoints with spontaneous spatial proximity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- We propose an adaptive superpoint update algorithm to" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 548, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 548, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 548, + 84 + ], + "type": "text", + "content": "strengthen internal semantic consistency of superpoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 84, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 133 + ], + "type": "text", + "content": "- Our method achieves remarkable performance in point cloud oversegmentation across multiple datasets. It can be integrated with semantic segmentation networks and shows promising performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 144, + 400, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 400, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 400, + 157 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 165, + 477, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 165, + 477, + 177 + ], + "spans": [ + { + "bbox": [ + 314, + 165, + 477, + 177 + ], + "type": "text", + "content": "2.1. Deep Learning in Point Cloud" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 399 + ], + "type": "text", + "content": "Deep learning approaches for 3D point cloud primarily follow three paradigms: projection-based, voxel-based, and point-based methods. Projection-based methods project unordered points onto 2D planes for CNN processing [6, 20, 21, 32], sacrificing 3D geometry for computational efficiency. Voxel-based methods discretize space into grids for 3D convolution [25, 31], improved by sparse convolution techniques [8, 10, 35] to mitigate memory costs, though kernel size limitations remain a constraint. Point-based methods [24, 27, 28, 33, 43] process raw point clouds directly and have evolved toward transformer architectures [13, 29, 37, 40, 44]. These methods offer powerful representation capabilities while struggling with scalability. The recently proposed SPT [29] leverages hierarchical superpoints and sparse self-attention to efficiently process large-scale scenes, achieving superior semantic segmentation with reduced model size and computational cost compared to traditional approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 407, + 481, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 481, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 481, + 420 + ], + "type": "text", + "content": "2.2. Point Cloud Oversegmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 688 + ], + "type": "text", + "content": "Existing point cloud oversegmentation methods can be broadly categorized into optimization-based methods and deep learning-based approaches. Early methods such as Papon et al. [26] proposed the VCCS method, which used FPFH as input features, selected seed points using an octree, and performed point cloud superpoint segmentation using the K-means clustering algorithm. Lin et al. [22] viewed the superpoint segmentation problem as a subset selection problem and employed subset selection methods to generate superpoints, still using FPFH as point cloud feature descriptors. Guinard et al. [12] simplified the superpoint segmentation problem into a graph cut problem, extracting features such as local linearity, flatness, divergence, and normal vectors from the point cloud, and applied a greedy graph cut algorithm to generate superpoints. Xiao et al. [39] defined the superpoint segmentation problem as an energy minimization problem and proposed a merging-exchange optimization framework to generate supervoxels. These methods often require complex and computationally expensive preprocessing steps, and their performance is limited by the insufficient expressiveness of handcrafted features. Recent deep learning methods overcome these limitations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Supervised Superpoint (SSP) method [18] pioneered deep network-based point cloud embedding, although it" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25832" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 553, + 256 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 553, + 256 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 553, + 256 + ], + "type": "image", + "image_path": "0c2b7bdd95d3c04ae8a3346b5d94122776995e0d362401c86b53c9a4031e675c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 258, + 193, + 267 + ], + "lines": [ + { + "bbox": [ + 104, + 258, + 193, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 258, + 193, + 267 + ], + "type": "text", + "content": "(a) Point Feature Extractor" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 258, + 384, + 268 + ], + "lines": [ + { + "bbox": [ + 257, + 258, + 384, + 268 + ], + "spans": [ + { + "bbox": [ + 257, + 258, + 384, + 268 + ], + "type": "text", + "content": "(b) Hierarchical Segment Initialization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 258, + 541, + 269 + ], + "lines": [ + { + "bbox": [ + 435, + 258, + 541, + 269 + ], + "spans": [ + { + "bbox": [ + 435, + 258, + 541, + 269 + ], + "type": "text", + "content": "(c) Adaptive Superpoint Update" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 279, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 324 + ], + "type": "text", + "content": "Figure 2. The architecture includes three stages. First, point embeddings are extracted using a U-Net backbone with sparse convolutions in shallow layers and a Mamba bottleneck enhanced by Conditional Position Encoding (CPE). Second, the point cloud is serialized via a Hilbert curve to initialize segments for superpoint aggregation, with granularity at each level controlled by the hyperparameter " + }, + { + "bbox": [ + 55, + 279, + 555, + 324 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 279, + 555, + 324 + ], + "type": "text", + "content": ". Finally, points are reassigned across segments to group semantically similar points into coherent superpoints, completing the clustering process." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 345, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 345, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 295, + 464 + ], + "type": "text", + "content": "still depended on optimization techniques from [12]. SP-Net [15] further advanced the field by introducing learnable point-to-superpoint associations using FPS for center selection, enabling joint optimization in both coordinate and feature spaces. Furthermore, SuperLiDAR [16] improved efficiency by combining a BFS-based grouping algorithm with a local discriminative loss to generate compact superpoints. However, this method results in high computational complexity in spatial query operations during superpoint generation, especially for dense point cloud oversegmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 207, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 207, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 207, + 484 + ], + "type": "text", + "content": "2.3. Localization-based Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 490, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 634 + ], + "type": "text", + "content": "Recent works [5, 23, 34, 38] have introduced serialization-based approaches that differ from traditional point cloud processing by transforming unstructured point clouds into ordered sequences to preserve spatial proximity. For example, OctFormer [34] uses octree-based ordering similar to " + }, + { + "bbox": [ + 55, + 490, + 296, + 634 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 490, + 296, + 634 + ], + "type": "text", + "content": "-ordering, offering scalability but limited by octree constraints. FlatFormer [23] employs window-based sorting to group point pillars, though its scalability in receptive field is limited, making it suitable for pillar-based 3D object detection. Building on these methods, PTv3 [38] extends attention scales by replacing KNN-based query with serialized neighborhoods, enhancing efficiency and scalability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 111, + 657 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Our oversegmentation network comprises three key steps: point embedding extraction (Sec. 3.1), hierarchical segment initialization (Sec. 3.2), and adaptive superpoint update (Sec. 3.3). The overall process is shown in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 344, + 383, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 344, + 383, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 344, + 383, + 356 + ], + "type": "text", + "content": "3.1. Backbone" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 362, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 409 + ], + "type": "text", + "content": "Optimization-based oversegmentation methods are constrained by the quality of handcrafted features, while learning based approaches similarly encounter limitations in deep feature representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": "Sparse 3D U-Net. Let the input point cloud be " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\mathbb{R}^{N \\times (C + 3)}" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": ", where each of the " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": " points is represented by its features (e.g. normal vectors, RGB) and spatial coordinates " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": ". Following [10], we first voxelize " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": " using the operator " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": ", and then process the discretized data with a U-Net architecture based on sparse convolutions (SpConv). In the shallow part of the U-Net [30], SpConv layers capture fine-grained local geometric information. These operations produce feature representations that maintain high spatial resolution and extract detailed geometric structures crucial for accurate object boundary delineation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 543, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 662 + ], + "type": "text", + "content": "Bottleneck Reinforcement. Pure SpConv restricts over-segmentation effectiveness due to their inherently limited receptive field. We enhance our U-Net backbone by integrating Mamba [11, 45] at the bottleneck of the U-Net, as shown in Fig. 2 (a). After the shallow SpConv layers have abstracted local features, the Mamba module is applied to perform global context interaction. Mamba, with its linear-complexity global modeling capability, compensates for the local inductive bias of convolution and promotes semantic consistency across superpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 671, + 502, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 502, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 502, + 684 + ], + "type": "text", + "content": "3.2. Hierarchical Segment Initialization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Methodology Justification. To efficiently partition point clouds into multiple segments, we propose a serialization-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25833" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 252 + ], + "type": "text", + "content": "based partitioning method that utilizes the spatial locality preservation property of the Hilbert curve [38, 41] to divide the same point cloud into hierarchical structures of different granularities. Specifically, the Hilbert curve maps adjacent point clusters in 3D Euclidean space to contiguous positions in a 1D sequence, thereby simplifying the spatial proximity-based 3D partitioning task into an iterative partitioning of a 1D sequence. This transformation enables us to efficiently divide point clouds into compact segments in 3D space, significantly enhancing partitioning efficiency. This enables (1) efficient 1D clustering that avoids expensive 3D queries and clustering; and (2) a simplified oversegmentation process that decouples spatial locality from semantic similarity by performing only semantic comparisons within already spatially adjacent segments to complete clustering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "spans": [ + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": "Partitioning Strategy. Due to the variable receptive field of differently sized superpoints, smaller superpoints preserve more detailed local geometric information, whereas larger superpoints capture more comprehensive global semantic context. Therefore, we propose to construct a hierarchical partitioning structure to fully utilize multi-scale information. Given an input point cloud " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " and its corresponding features " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "F = \\{f_i\\}_{i=1}^N" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "f_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " is generated by our backbone, we construct a hierarchical partitioning structure " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{H} = \\{L_k\\}_{k=1}^K" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ". Each layer " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "L_k" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " contains the complete serialized point cloud and is recursively quaternary partitioned into a set of segments " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "L_k = \\{S_i^k\\}_{i=0}^{M_k-1}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "S_i^k" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": "-th segment, as shown in Fig. 2 (b). Specifically, the 3D point cloud is first serialized into a 1D sequence via Hilbert space-filling curve mapping [38], denoted as " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{seq} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ". The sequence is then padded to " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{pad} \\in \\mathbb{R}^{N_{pad} \\times d}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "N_{pad} = \\lceil N / M_K \\rceil \\cdot M_K" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ", ensuring divisibility by the finest segment count " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "M_K" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ". The serialized point cloud " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{pad}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " is recursively partitioned using 1D quaternary splitting, such that at each layer " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "L_k" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ", it is uniformly divided into " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "M_k = 2^{m - 2k + 2}" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " segments. For each layer " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "L_k" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i^k" + }, + { + "bbox": [ + 57, + 257, + 297, + 520 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 525, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 525, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 82, + 525, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} ^ {k} = \\left\\{\\mathcal {P} _ {p a d} [ j ] \\mid j \\in [ i \\cdot l _ {k}, (i + 1) \\cdot l _ {k} - 1 ] \\right\\}, \\tag {1}", + "image_path": "0f0ac2a88c221410fbedaabdf09fd21f08f37f9c605b5d69f2354d1760e54790.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "l_{k} = N_{pad} / M_{k}" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": " represents the segment length for each segment in the " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": "-th layer. Recursive partitioning establishes fixed parent-child indexing between layers. For any inter-layer gap " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "l \\geq 1" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": ", the hierarchical inclusion relationship is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 610, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 610, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 108, + 610, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} ^ {k} = \\bigcup_ {j \\in [ i \\cdot 4 ^ {l}, (i + 1) \\cdot 4 ^ {l} - 1 ]} \\mathcal {S} _ {j} ^ {k + l}. \\tag {2}", + "image_path": "b8cfcfe5bcf9e0b0fcc701919ee342816ab6315cc2d1863adb454123d2b4e7b8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 670 + ], + "type": "text", + "content": "At the finest layer " + }, + { + "bbox": [ + 55, + 635, + 295, + 670 + ], + "type": "inline_equation", + "content": "L_{K}" + }, + { + "bbox": [ + 55, + 635, + 295, + 670 + ], + "type": "text", + "content": ", segment features are generated through permutation-invariant aggregation of constituent points:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 670, + 294, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 670, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 108, + 670, + 294, + 685 + ], + "type": "interline_equation", + "content": "f _ {S _ {i} ^ {K}} = \\mathcal {A} \\left(f _ {j} \\mid p _ {j} \\in S _ {i} ^ {K}\\right) \\in \\mathbb {R} ^ {C}, \\tag {3}", + "image_path": "e89cd95e1f900a18fe1787c00f8cc9757bb51eb1f5a6cd81eed9eded007635b2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}:\\mathbb{R}^{n_i\\times C}\\to \\mathbb{R}^C" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " denotes a permutation-invariant operator (e.g. max-pooling) with " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "n_i = |\\mathcal{S}_i^K | = N_{pad} / M_K" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 471, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 471, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 471, + 85 + ], + "type": "text", + "content": "3.3. Adaptive Superpoint Update" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": "Given the structure " + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": ", the goal is to refine the partition " + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "inline_equation", + "content": "L_{k} = \\{S_{1}^{k},\\ldots ,S_{M_{k}}^{k}\\}" + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": " with features " + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "inline_equation", + "content": "f_{S_i^k}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": " into semantically consistent superpoints " + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "inline_equation", + "content": "L_k^\\prime = \\{S_1'^k,\\dots,S_M'^k\\}" + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": " with precise geometric boundaries and updated features " + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "inline_equation", + "content": "f_{S_i^k}^\\prime \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 89, + 553, + 226 + ], + "type": "text", + "content": ". This is achieved by dynamically allocating raw points within coarse-grained segments to their appropriate sub-segments based on semantic similarity, a process that redistributes raw points within segments to fit object boundaries. This process can be transformed into a boundary fitting problem between adjacent segments in 1D sequence point clouds, as shown in Fig. 2 (c)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": "Feature Similarity based Affiliation. To establish the affiliation relationship between raw points within each segment and nearby segments, we consider two adjacent levels in " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": ": the finest " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "L_{k} = \\{S_{1}^{k}, S_{2}^{k}, \\ldots, S_{M_{k}}^{k}\\}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": " and its coarser predecessor " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "L_{k-1} = \\{S_{1}^{k-1}, S_{2}^{k-1}, \\ldots, S_{M_{k-1}}^{k-1}\\}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": " (we only use " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "l = 1" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": " defined in Eq. (2)). The hierarchical relationship between these levels is given by (1) " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "M_{k-1} = \\frac{M_{k}}{4}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": ", meaning each coarser level has one-fourth the segments of the finer level and (2) Each segment " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "S_{i}^{k-1}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": " at level " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "L_{k-1}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": " contains four consecutive segments from level " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "L_{k}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "inline_equation", + "content": "S_{i}^{k-1} = S_{4i}^{k} \\cup S_{4i+1}^{k} \\cup S_{4i+2}^{k} \\cup S_{4i+3}^{k}" + }, + { + "bbox": [ + 313, + 231, + 553, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": "For each raw point " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "p_i \\in S_i^{k-1}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ", we identify the most similar initialized segment feature from its four containing segments " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "S_{4i}^k, S_{4i+1}^k, S_{4i+2}^k" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "S_{4i+3}^k" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ". Meanwhile, to expand the receptive field of raw points relative to their potential superpoints, we incorporate segment features from eight sub-segments at " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "L_k" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": " belonging to the left and right adjacent segments " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "S_{i-1}^{k-1}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "S_{i+1}^{k-1}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ", forming a local feature set " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\{f_{S_{4i+j}^k} \\mid j \\in \\{-4, -3, \\dots, 7\\}\\}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ", resulting in a set of 12 initial segment features. For each segment " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "S_i^{k-1}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "N_{k-1} = 4N_{pad} / M_k" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": " points, we compute a similarity matrix " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i \\in \\mathbb{R}^{N_{k-1} \\times 12}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": ". Each entry " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "a_{j,t}" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": " measures the similarity between the " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": "-th point and the " + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 365, + 555, + 525 + ], + "type": "text", + "content": "-th segment feature in the above set of segment features:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 529, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 529, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 325, + 529, + 553, + 558 + ], + "type": "interline_equation", + "content": "a _ {j, t} = \\frac {f _ {p _ {j}} \\cdot f _ {S _ {t} ^ {k}}}{\\| f _ {p _ {j}} \\| \\| f _ {S _ {t} ^ {k}} \\|}, \\quad j \\in [ 1, N _ {k - 1} ], \\quad t \\in [ 1, \\tau ], \\tag {4}", + "image_path": "537ff29c169c856627ce024bcaa7b9d30d30eba079eaf3ab35965c75db4d192b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "content": "This process generates similarity matrices " + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_i\\}_{i=1}^{M_{k-1}}" + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "content": " for the entire level " + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "inline_equation", + "content": "L_{k-1}" + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "content": ". Point-to-superpoint assignments are determined by selecting the superpoint with the highest similarity for each point, which is then mapped to global superpoint indices " + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "inline_equation", + "content": "\\phi(p_j) \\in [1, M_k]" + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "content": ". The resulting assignment index set " + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "inline_equation", + "content": "Ass_i = \\{\\phi(p_j)\\}_{j=1}^n" + }, + { + "bbox": [ + 313, + 563, + 553, + 649 + ], + "type": "text", + "content": " provides a key reference for updating superpoint features in subsequent steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Cross-Attention based Update. Cross-Attention [4] enables dynamic and adaptive relationships between superpoints and their constituent points. Through feature similarity based affiliation, precise similarity relationships are established, guiding accurate superpoint feature updates. For" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25834" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 219 + ], + "type": "image", + "image_path": "0db175f23efedc3fe081828f140afd4176b41311856bb7bb2c483578fd4f6e5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 275 + ], + "type": "text", + "content": "Figure 3. The end-to-end semantic segmentation pipeline constructs a graph over superpoints, captures global context using graph transformers, and integrates point-level embeddings for final semantic prediction, where " + }, + { + "bbox": [ + 55, + 229, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 229, + 296, + 275 + ], + "type": "text", + "content": " denotes simple feature addition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": "the assignment indices " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "A s s i = \\{\\phi(p_{j})\\}_{j=1}^{n}" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": ", the inverse mapping " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\phi^{-1}(i) = \\{p_{j} \\mid \\phi(p_{j}) = i\\}" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": " represents the original points contained within superpoint " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "S_{i}^{\\prime k}" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "A s s i" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": ", each superpoint's feature is updated solely from its member points. Using superpoint feature " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "f_{S_{i}^{\\prime k}}^{\\prime}" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": " as the Query and its member point features " + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\{f_{p_{j}} \\mid p_{j} \\in \\phi^{-1}(i)\\}" + }, + { + "bbox": [ + 55, + 293, + 296, + 384 + ], + "type": "text", + "content": " as Keys and Values, the update formula is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 391, + 296, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 391, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 69, + 391, + 296, + 441 + ], + "type": "interline_equation", + "content": "f _ {S _ {i} ^ {k}} ^ {\\prime} = \\text {C r o s s A t t n} \\left(\\underbrace {f _ {S _ {i} ^ {k}}} _ {\\text {Q u e r y}}, \\underbrace {\\left\\{f _ {p _ {j}} \\mid p _ {j} \\in \\phi^ {- 1} (i) \\right\\}} _ {\\text {K e y / V a l u e}}\\right), \\tag {5}", + "image_path": "95d9fe68ffe202006b493185dcf60cfaf7e19684aaaef6e439db1852fed04bbd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 483 + ], + "type": "text", + "content": "Attention weights are computed through scaled dot product, allowing each superpoint to focus on the most informative points within its local neighborhood:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 490, + 295, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 490, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 67, + 490, + 295, + 518 + ], + "type": "interline_equation", + "content": "\\operatorname {C r o s s A t t n} \\left(f _ {S _ {i} ^ {k}}, \\left\\{f _ {p _ {j}} \\right\\}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {\\top}}{\\sqrt {d}}\\right) \\mathbf {V} _ {i}, \\tag {6}", + "image_path": "2a3796f33ad57450bb17d644280022b59234715439d8c2b25d6f088146b90ef8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i = f_{S_i^k}W_K\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_i = \\mathrm{Concat}(\\{f_{p_j}W_K\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}" + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i = \\mathrm{Concat}(\\{f_{p_j}W_V\\mid p_j\\in \\phi^{-1}(i)\\})\\in \\mathbb{R}^{n_i^k\\times d}" + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "inline_equation", + "content": "n_i^k = |\\phi^{-1}(i)|" + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": " is the number of points contained in each superpoint, and " + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "inline_equation", + "content": "W_{Q},W_{K},W_{V}\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 55, + 525, + 296, + 626 + ], + "type": "text", + "content": " are learnable projection matrices. For coarse-grained superpoints, we apply the same update logic, with interactions between fine-grained features and their coarse-grained initializations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 186, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 186, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 186, + 647 + ], + "type": "text", + "content": "3.4. Superpoint Interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": "The superpoint features " + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{S_i^k}^\\prime" + }, + { + "bbox": [ + 55, + 651, + 296, + 714 + ], + "type": "text", + "content": " encode geometric-semantic information for locally homogeneous points. To leverage these features for semantic segmentation, inspired by SPT [29], we integrate a superpoint interaction module with our oversegmentation network, forming an end-to-end" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": "framework, as shown in Fig. 3. In our hierarchical design, fine-grained superpoints capture local details while coarse grained ones represent object-level semantics. We construct multi-level superpoint graphs and employ a cascaded Transformer-GCN [42] to facilitate cross-superpoint interactions and global context modeling. In the Graph Transformer, a superpoint graph " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": " is built by linking each superpoint with its " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": " nearest neighbors in feature space, where " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{f_{S_i^k}'\\}_{i=1}^{M_k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": ". Edge features are defined via relative position encoding as " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{ij} = g(p_j - p_i)" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": " representing node coordinates and " + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 72, + 555, + 266 + ], + "type": "text", + "content": " being a linear mapping. A subsequent GCN processes this graph to enhance local feature interactions through message passing among adjacent superpoints. Finally, we fuse backbone point features with their corresponding multi-level superpoint before feeding them into the segmentation head." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 273, + 474, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 474, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 474, + 286 + ], + "type": "text", + "content": "3.5. Superpoint Aggregation Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "text", + "content": "We propose a superpoint aggregation loss to optimize superpoint generation, inspired by LMNN [36]. This comprehensive loss consists of three components. First, a compactness loss " + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{compact}}" + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "text", + "content": " enforces intra-superpoint feature cohesion by minimizing point-to-superpoint distances. Second, a distinction loss " + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "text", + "content": " encourages inter-class separability by maximizing the distances between superpoints of different classes, where each superpoint's label is determined by majority voting among its points. Third, a purity loss " + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{purity}}" + }, + { + "bbox": [ + 313, + 290, + 555, + 434 + ], + "type": "text", + "content": " maintains label consistency within each superpoint by penalizing high entropy in its class distribution. The overall loss function is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 438, + 555, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 438, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 326, + 438, + 555, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\ell_ {s p} = \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\frac {\\sum_ {p \\in S _ {i}} \\| f _ {i} - f _ {S _ {i} ^ {k}} ^ {\\prime} \\| ^ {2}}{| S _ {i} |} + \\\\ \\frac {1}{N _ {c}} \\sum_ {i, j \\in c, i \\neq j} \\max \\left(0, \\delta_ {\\text {d i s t}} - \\| f _ {S _ {i} ^ {k}} ^ {\\prime} - f _ {S _ {j} ^ {k}} ^ {\\prime} \\|\\right) ^ {2} + \\tag {7} \\\\ \\frac {1}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\sum_ {k = 1} ^ {K} p _ {i, k} \\log (p _ {i, k}), \\\\ \\end{array}", + "image_path": "4e61d12308147d3fa3a9f6adfa0412b3a10b4a209a3dc5f9e40e1fa3ce531c3a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " denotes the feature of point " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "f_{S_i^k}^\\prime" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " represents the feature of superpoint " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "S_{i}^{\\prime k}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " is a predefined separation margin, " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "p_{i,k}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " is the probability of superpoint " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " is the number of superpoints, and " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": " is the number of superpoint pairs involved in " + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{dist}}" + }, + { + "bbox": [ + 313, + 546, + 554, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 621, + 395, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 395, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 395, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 436, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 436, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 436, + 659 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate SPCNet on four datasets with diverse characteristics. S3DIS [1] contains " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "274M" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " points from six office areas, with Area 5 as the test set. ScanNet [9] includes 1,513 indoor scans from 707 scenes with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25835" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 206 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 206 + ], + "type": "image", + "image_path": "68583646c2c86e7d1defb0fe1e6c589c815ca02d12e9b76b343157e2cfa2954a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": "Figure 4. The visualization of superpoints generated on the ScanNet (first row) and S3DIS (second row) datasets shows a two-level hierarchy: " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " with 1024 superpoints and " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " with 256 superpoints. Finer " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " superpoints merge based on semantic similarity to form larger " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " superpoints. For example, in S3DIS, " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " superpoints of table parts merge into a single table superpoint, and " + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 54, + 214, + 555, + 282 + ], + "type": "text", + "content": " superpoints of whiteboard sections form a complete whiteboard superpoint. Our superpoints are regular and orderly, with nearly rectangular shapes in semantically consistent areas (e.g. floors, tables), maintaining continuity, precise object boundaries, and demonstrating superior regularity, compactness, and purity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 302, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 302, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 54, + 302, + 294, + 374 + ], + "type": "text", + "content": "20 categories, averaging 148k points/scan. nuScenes [3] comprises 1,000 scenes from 32-beam LiDAR, split 750/150/150 for train/val/test, with 17 categories. SemanticKITTI [2] is an outdoor self-driving benchmark with 22 sequences and 20 categories, using sequences 00-10 (densely annotated) for training and 11-22 for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 374, + 295, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 482 + ], + "type": "text", + "content": "Evaluation Metrics. Following SPNet [15] and SSP [18], we use Oracle Overall Accuracy (OOA), Boundary Recall (BR), Boundary Precision (BP), and F1 score to evaluate superpoints. BR and BP assess boundary quality, while OOA represents the upper bound of semantic segmentation accuracy using superpoints. The F1 score balances recall and precision, defined as " + }, + { + "bbox": [ + 54, + 374, + 295, + 482 + ], + "type": "inline_equation", + "content": "F1 = \\frac{2\\cdot BP\\cdot BR}{BP + BR}" + }, + { + "bbox": [ + 54, + 374, + 295, + 482 + ], + "type": "text", + "content": ". For semantic segmentation in indoor and outdoor scenarios, we adopt mean intersection over union (mIoU)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 178, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 178, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 178, + 503 + ], + "type": "text", + "content": "4.2. Model Configuration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 508, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 295, + 627 + ], + "type": "text", + "content": "We maintain a consistent architecture across all experiments with minimal dataset-specific adjustments. Our backbone features a four-stage symmetric encoder-decoder structure, each with a single block depth. The encoder has embedding dimensions [32, 64, 128, 256] and the decoder [128, 64, 64], with Mamba blocks using conditional positional encoding at the 256 stage. We implement a three-level hierarchy " + }, + { + "bbox": [ + 54, + 508, + 295, + 627 + ], + "type": "inline_equation", + "content": "(m = 8, K = 2, l = 1)" + }, + { + "bbox": [ + 54, + 508, + 295, + 627 + ], + "type": "text", + "content": ", organizing points into n, 1024, 256 segments. Training is performed on dual NVIDIA RTX 4090 GPUs with the Adam optimizer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 223, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 223, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 223, + 648 + ], + "type": "text", + "content": "4.3. Point Cloud Oversegmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. As shown in Tab. 1, we evaluate SPCNet on three datasets, reserving ScanNet for ablation studies, ensuring fair comparison with consistent superpoint counts. For example, in S3DIS Area 5, methods like VCCS, Lin et al., SPG, SSP, and SPNet generate around 1050 su" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 300, + 547, + 582 + ], + "blocks": [ + { + "bbox": [ + 322, + 300, + 547, + 582 + ], + "lines": [ + { + "bbox": [ + 322, + 300, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 322, + 300, + 547, + 582 + ], + "type": "table", + "html": "
MethodBRBPF1OOA
S3DIS Area 5 [1]
VCCS [26]62.0610.8618.4895.22
SPG [19]52.0613.1120.9495.84
SSP [18]80.7313.0222.4297.04
SPNet [15]84.7513.1422.6596.50
SPCNet(Ours)89.5026.7840.5096.81
nuScenes [3]
SPG [19]28.2017.2621.4189.22
SSP [18]22.0415.6318.2892.01
SPNet [15]68.9218.3428.9787.67
SuperLiDAR [16]74.7225.1237.5996.31
SPCNet(Ours)88.7732.8247.5296.33
SemanticKITTI [2]
SPG [19]25.6415.8219.5686.86
SSP [18]18.7410.6713.5992.27
SPNet [15]56.5214.7823.4392.24
SuperLiDAR [16]65.5220.5231.2596.21
SPCNet(Ours)73.1830.1542.5796.84
", + "image_path": "a918902f1be89a6699ae439b17f03fc6ab984fbc29b3a8042936b2b9caa23b3a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 588, + 555, + 611 + ], + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 611 + ], + "type": "text", + "content": "Table 1. Comparison results of generated superpoints on the S3DIS, nuScenes, and SemanticKITTI datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "perpoints, matching our 1024 Hilbert partitions. Our deep learning-based approach outperforms traditional methods, benefiting from the specialized backbone network's ability to extract rich geometric features that capture complex local structures. Unlike methods such as SPNet [15], which rely on FPS+KNN for presampling and soft assignment map-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25836" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": "ping without limiting the scope of raw points being clustered, resulting in blurry boundaries, our approach leverages the Hilbert curve's locality, limiting the range of raw point clustering (preventing clustering of semantically similar but spatially distant points), and finally refining superpoint through the Cross-Attention, minimizing noise and maintaining precise, simple boundaries, as shown in Fig. 5 (d) and (e). Visualization results in Fig. 4 show that our method tends to aggregate superpoints within object interiors while maintaining distinct object boundaries, especially in challenging cases like wall-whiteboard interfaces, resulting in higher BR and BP. Additionally, our method excels at merging semantically continuous regions (e.g. floors, tabletops), leading to fewer incorrect superpoint boundaries and substantially higher BP metrics. Fig. 6 shows the performance of different methods on SemanticKITTI." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 271, + 295, + 337 + ], + "blocks": [ + { + "bbox": [ + 56, + 271, + 295, + 337 + ], + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 337 + ], + "type": "table", + "html": "
MethodDeviceInference Time (ms) data proc. / backbone + SP ini. + SP gen.
SPG [19]CPU13152
SSP [18]RTX 309016115 (11812 + 0 + 13426)
SPNet [15]RTX 309012182 (11812 + 0 + 1154)
SuperLiDAR [16]RTX 309072(0 + 0 + 72)
SPCNet (ours)RTX 4090, RTX 306012 (6 + 0.4 + 6), 61
", + "image_path": "4136bf7b35d0725ded018c521b9631c6c78ffadba11ac19071c062bccd725c65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 345, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 390 + ], + "type": "text", + "content": "Table 2. The inference time of different oversegmentation methods on SemanticKITTI. The inference time consists of data preprocessing/backbone (red/blue), superpoint initialization (green), and superpoint generation (purple)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": "Time Costs. We evaluated oversegmentation network efficiency by measuring inference time per scan on SemanticKITTI validation (batch size=1) in Tab. 2. For optimization based SPG [19], we followed SuperLiDAR's [16] configuration using Core i5 CPU. For learning-based methods (SSP [18], SPNet [15], SuperLiDAR), we used SuperLiDAR's results (RTX 3090). Our SPCNet was benchmarked on RTX 4090 and 3060 GPUs. Our method achieved total inference times of 12 ms (RTX 4090) and 61 ms (RTX 3060), achieving " + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": " speedup versus SPG/SSP/SPNet, and " + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": " versus SuperLiDAR. Even on RTX3060, our inference is 10ms faster than SuperLiDAR, indicating at least " + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": " efficiency accounting for hardware differences. This stems from: (1) efficient superpoint initialization (0.4 ms) that compresses FPS+KNN sampling time in SPNet; (2) the application of a localized attention mechanism within the superpoint refinement (6 ms). Our backbone architecture allows further acceleration using lightweight alternatives like PointNet[27]. Traditional methods face key bottlenecks: SPG/SSP rely on optimization-based generation (e.g. graph cuts), SSP/SPNet suffer from redundant processing due to mixed handcrafted/learned features, and SuperLiDAR's BFS-based grouping has complexity dependent on scene density, potentially reaching " + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": " in dense scenes, versus our serialization based approach " + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "inline_equation", + "content": "(O(n))" + }, + { + "bbox": [ + 55, + 399, + 296, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Ablation Studies. Mamba mitigates the local inductive bias" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 75, + 376, + 133 + ], + "blocks": [ + { + "bbox": [ + 317, + 75, + 376, + 133 + ], + "lines": [ + { + "bbox": [ + 317, + 75, + 376, + 133 + ], + "spans": [ + { + "bbox": [ + 317, + 75, + 376, + 133 + ], + "type": "image", + "image_path": "7c95b7aa84caa670107ecd0e17bf52de6eeb5ec47fad76ad32dfd6f262f22ce0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 135, + 370, + 151 + ], + "lines": [ + { + "bbox": [ + 324, + 135, + 370, + 151 + ], + "spans": [ + { + "bbox": [ + 324, + 135, + 370, + 151 + ], + "type": "text", + "content": "(a) 3D-UNet only with " + }, + { + "bbox": [ + 324, + 135, + 370, + 151 + ], + "type": "inline_equation", + "content": "k = 2" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 377, + 75, + 436, + 132 + ], + "blocks": [ + { + "bbox": [ + 377, + 75, + 436, + 132 + ], + "lines": [ + { + "bbox": [ + 377, + 75, + 436, + 132 + ], + "spans": [ + { + "bbox": [ + 377, + 75, + 436, + 132 + ], + "type": "image", + "image_path": "f89d9e30935e1b0368a2e56f53d8e28ffdfc4fcad1d6f2ecff826a3e30b4715d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 135, + 435, + 151 + ], + "lines": [ + { + "bbox": [ + 375, + 135, + 435, + 151 + ], + "spans": [ + { + "bbox": [ + 375, + 135, + 435, + 151 + ], + "type": "text", + "content": "(b) Bottleneck reinforcement with " + }, + { + "bbox": [ + 375, + 135, + 435, + 151 + ], + "type": "inline_equation", + "content": "k = 2" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 75, + 492, + 132 + ], + "blocks": [ + { + "bbox": [ + 437, + 75, + 492, + 132 + ], + "lines": [ + { + "bbox": [ + 437, + 75, + 492, + 132 + ], + "spans": [ + { + "bbox": [ + 437, + 75, + 492, + 132 + ], + "type": "image", + "image_path": "f8f9808ff53a2c30e00548eac0e07af490314ce6ba87be1c2ee4af92c4efca42.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 135, + 481, + 151 + ], + "lines": [ + { + "bbox": [ + 447, + 135, + 481, + 151 + ], + "spans": [ + { + "bbox": [ + 447, + 135, + 481, + 151 + ], + "type": "text", + "content": "(d) PointNet with " + }, + { + "bbox": [ + 447, + 135, + 481, + 151 + ], + "type": "inline_equation", + "content": "k = 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 251 + ], + "type": "text", + "content": "Figure 5. Oversegmentation comparisons. (a) Sparse 3D U-Net with neighbor expansion of 2. (b) Adding the Mamba module to the bottleneck. (c) Replacing the backbone with a simplified PointNet, as in SPNet. (d) SPNet results. Notably, while (b) configuration does not fully aggregate semantically consistent objects into complete superpoints, it notably enhances consistency. SPC-Net also avoids clustering semantically similar but spatially distant points, ensuring clearer superpoint boundaries." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 493, + 75, + 552, + 132 + ], + "blocks": [ + { + "bbox": [ + 493, + 75, + 552, + 132 + ], + "lines": [ + { + "bbox": [ + 493, + 75, + 552, + 132 + ], + "spans": [ + { + "bbox": [ + 493, + 75, + 552, + 132 + ], + "type": "image", + "image_path": "28a5561d25924c26d780d6c78c4fe37fc4e4b55d5c8510c0150c8779447fba24.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 135, + 535, + 144 + ], + "lines": [ + { + "bbox": [ + 509, + 135, + 535, + 144 + ], + "spans": [ + { + "bbox": [ + 509, + 135, + 535, + 144 + ], + "type": "text", + "content": "(e) SPNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 265, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 315, + 265, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 553, + 308 + ], + "type": "table", + "html": "
Backbone SettingBR(%)BP(%)OOA(%)mIou(%)
3D U-Net76.33 ↓7.9028.45 ↓1.7090.93 ↓1.1974.10 ↓2.86
+ Mamba fully78.37 ↓5.8629.13 ↓1.0291.24 ↓0.8876.18 ↓0.78
+ Mamba bottleneck (default)84.2330.1592.1276.96
", + "image_path": "80297249ba28dbd24c35b5a30214dfff8658d4652eb49da1a076300473c7711f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "type": "text", + "content": "Table 3. Performance comparison of different backbone architectures on the ScanNet validation set." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 353, + 552, + 419 + ], + "blocks": [ + { + "bbox": [ + 315, + 353, + 552, + 419 + ], + "lines": [ + { + "bbox": [ + 315, + 353, + 552, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 552, + 419 + ], + "type": "table", + "html": "
Neighbor ScopeBR(%)BP(%)OOA(%)
No expansion79.86 ↓4.3728.57 ↓1.5891.22 ↓0.90
2 expansion (default)84.2330.1592.12
4 expansion83.67 ↓0.5629.85 ↓0.3091.24 ↓0.88
6 expansion79.93 ↓4.3028.65 ↓1.5089.93 ↓2.19
", + "image_path": "6240a39bcc8b9ae4536efffe9a3d404efd8abe43e535473480ffa4fecbdf036f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 427, + 553, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 427, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 553, + 450 + ], + "type": "text", + "content": "Table 4. The effect of expanding neighbor segment range on superpoint performance on the ScanNet validation set." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "of convolutions in oversegmentation, improving superpoint continuity and semantic consistency. As shown in Fig. 5 (a) and (b), this bias causes superpoints in the same semantic region to align with boundaries but fragment due to internal inconsistency. Thus, we retain convolutions in U-Net's shallow layers for local geometric cues to ensure boundary precision, and introduce Mamba at the bottleneck to capture long range semantics. Tables 3 and 4 analyze backbone design and neighbor expansion effects on superpoint quality. We evaluated three backbone configurations: (1) baseline 3D sparse U-Net only, (2) fully Mamba-based U-Net, and (3) our default hybrid design with Mamba modules in the bottleneck layers. Results demonstrate that the baseline 3D sparse U-Net reduces BR, BP, and OOA by " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "7.90\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.70\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " respectively compared to our default. Similarly, the fully Mamba-based U-Net underperforms, as Mamba's global context modeling in shallow networks with low-dimensional features smooths out high-frequency boundary information. Our hybrid \"shallow convolution + deep Mamba\" architecture achieves optimal per" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25837" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 182, + 163 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 182, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 182, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 182, + 163 + ], + "type": "image", + "image_path": "8de302899f4f95dfd852cc00ec70e2f91c49f12bc3ceeae25552303e13887394.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 196 + ], + "type": "text", + "content": "Figure 6. Performance of different methods on the SemanticKITTI validation set under varying numbers of superpoints. It is worth noting that our method can achieve better results than other methods with fewer superpoints (256, 512, and 1024)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 73, + 304, + 163 + ], + "blocks": [ + { + "bbox": [ + 182, + 73, + 304, + 163 + ], + "lines": [ + { + "bbox": [ + 182, + 73, + 304, + 163 + ], + "spans": [ + { + "bbox": [ + 182, + 73, + 304, + 163 + ], + "type": "image", + "image_path": "4bafd1e0708c5c25e62a082e0384fdac660e6d5ea006be9e022455b37d4fc07b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 429, + 163 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 429, + 163 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 429, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 429, + 163 + ], + "type": "image", + "image_path": "9a03200b16ee743058f2e9b76a44c63aa914a8d384d7a320257d3e5de66549db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 552, + 163 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 552, + 163 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 552, + 163 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 552, + 163 + ], + "type": "image", + "image_path": "1836726bac8ed4d909e7809700771b8b5243b4cb9e1cf101822911af447133d4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 214, + 296, + 362 + ], + "blocks": [ + { + "bbox": [ + 56, + 214, + 296, + 362 + ], + "lines": [ + { + "bbox": [ + 56, + 214, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 214, + 296, + 362 + ], + "type": "table", + "html": "
ExperimentScanNet [9]S3DIS [1]nuScenes [3]Sem.KITTI [2]
Best ModelValTestArea5ValTestValTest
†SPG [19]--58.0----
MinkUNet [8]72.273.665.473.3-63.8-
†SPG + SSP [19]--61.7----
ST [17]74.373.772.0----
†SPT [29]--68.9----
PointNeXt [28]71.571.270.5----
OctFormer [34]75.776.6-76.177.264.367.8
Swin3D [40]76.4-72.5----
AF2S3Net [7]---62.278.074.270.8
†SuperLiDAR [16]----78.5-69.6
PTv2 [37]75.474.271.680.282.670.372.6
PTv3 [38]77.577.973.480.482.770.874.2
Backbone76.9-69.375.4-67.6-
†SPCNet(Ours)77.878.072.177.680.269.571.9
", + "image_path": "e0965269f4f330e05a171c42559a4516a34a592096cb9a23840584889b134f8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 414, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 450 + ], + "type": "text", + "content": "formance by leveraging convolutions to preserve local details in shallow layers while Mamba modules handle global context modeling in deeper layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "text", + "content": "In the neighborhood expansion experiments (analogous to k-parameter tuning in KNN), we observed that moderately expanding the candidate superpoint pool of original points during the feature similarity based affiliation process significantly improves sampling accuracy. The model achieves optimal performance when " + }, + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 2" + }, + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "text", + "content": "; excessive expansion " + }, + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4,6)" + }, + { + "bbox": [ + 54, + 450, + 295, + 581 + ], + "type": "text", + "content": " leads to performance degradation. This may be because larger receptive fields introduce noise, and the hard assignment approach introduces more uncertainty as the number of candidate targets increases, causing early assignment errors to propagate to later processing stages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "type": "text", + "content": "4.4. Semantic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Tab. 5 evaluates our approach on four indoor and outdoor datasets, comparing it with point-based and superpoint-based methods. Our optimized oversegmentation strategy shows significant gains: on S3DIS Area 5, we achieve a " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " mIoU improvement over SPT. Compared to state-of-the-art point-based methods, our approach remains competitive, outperforming PTv3 by " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " on ScanNet " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "(77.8\\%)" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and PTv2 by " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " on S3DIS Area 5. Additionally, we achieve " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "71.9\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "80.2\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " on Sem.KITTI and nuScenes, improving" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 214, + 553, + 271 + ], + "blocks": [ + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 393 + ], + "type": "text", + "content": "Table 5. Semantic segmentation results on the ScanNet, S3DIS, nuScenes, and SemanticKITTI datasets. †Superpoint-based." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 214, + 553, + 271 + ], + "lines": [ + { + "bbox": [ + 315, + 214, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 214, + 553, + 271 + ], + "type": "table", + "html": "
SettingComponentsModel Size (m)mIoU(%)Δ(%)
Backbone + SP gen. + SP int.
backbone (default)-18.876.96-
Hierarchical Enhancement
1 partition level+ Level1 Superpoint0.24 (0.218 + 0.022)77.50+0.54
2 partition level+ Level2 Superpoint0.02 (0.014 + 0.006)77.80+0.30
", + "image_path": "766e54e96e3c1d3c71a41f7b418a6ae5d5e80532ee0d85943dd887067ae1332e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": "Table 6. Semantic segmentation improvement with hierarchical superpoint features in ScanNet validation set. SP gen. means superpoint generation; SP int. means superpoint interaction." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "content": "SuperLiDAR by " + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 334, + 553, + 406 + ], + "type": "text", + "content": ", respectively. These results confirm that our superpoints capture local and global semantic information effectively, while the simple backbone, enhanced by superpoint features, matches the performance of more complex point-based methods, validating the superior quality of the generated superpoints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "text", + "content": "As shown in Tab. 6, integrating hierarchical superpoint features with per-point features extracted by the backbone improved segmentation. First-level superpoints aggregated local semantics, boosting intra-class consistency with a " + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "inline_equation", + "content": "0.54\\%" + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "text", + "content": " mIoU gain. Second-level features captured long-range dependencies, adding another " + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "inline_equation", + "content": "0.30\\%" + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "text", + "content": " improvement. These hierarchical superpoints acted as adaptive receptive fields, enhancing performance with only " + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "inline_equation", + "content": "0.24M" + }, + { + "bbox": [ + 313, + 406, + 555, + 513 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "text", + "content": "In this paper, we propose a novel serialization based oversegmentation method to address the challenge of efficiently grouping semantically consistent points into superpoints while maintaining spatial proximity. SPCNet achieves state-of-the-art performance in point cloud oversegmentation while being " + }, + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "text", + "content": " faster than existing methods in inference speed. Extensive experimental results demonstrate the effectiveness and efficiency of SPCNet. Additionally, SPCNet can be flexibly applied to semantic segmentation tasks, effectively improving the accuracy of semantic segmentation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "content": "Acknowledgments This work was supported in part by the National Natural Science Foundation of China (Grant No. 42201475)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25838" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 296, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 279 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 279 + ], + "type": "text", + "content": "[3] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 334 + ], + "type": "text", + "content": "[4] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 376 + ], + "type": "text", + "content": "[5] Wanli Chen, Xinge Zhu, Guojin Chen, and Bei Yu. Efficient point cloud analysis using hilbert curve. In European Conference on Computer Vision, pages 730-747. Springer, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 489 + ], + "type": "text", + "content": "[7] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. 2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12547-12556, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 544 + ], + "type": "text", + "content": "[8] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 601 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[10] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 689 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Stéphane Guinard and Loic Landrieu. Weakly supervised segmentation-aided classification of urban scenes from 3d" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "lidar point clouds. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42:151-157, 2017. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 91, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[14] David Hilbert and David Hilbert. Über die stetige abbildung einer linie auf ein flächenstück. Dritter Band: Analysis-Grundlagen der Mathematik-Physik Verschiedenes: Nebst Einer Lebensgeschichte, pages 1–2, 1935. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[15] Le Hui, Jia Yuan, Mingmei Cheng, Jin Xie, Xiaoya Zhang, and Jian Yang. Superpoint network for point cloud oversegmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5510-5519, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Le Hui, Linghua Tang, Yuchao Dai, Jin Xie, and Jian Yang. Efficient lidar point cloud oversegmentation network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18003-18012, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 555, + 353 + ], + "type": "text", + "content": "[17] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8500-8509, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "text", + "content": "[18] Loic Landrieu and Mohamed Boussaha. Point cloud over-segmentation with graph-structured deep metric learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7440–7449, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[19] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 1, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "type": "text", + "content": "[20] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "text", + "content": "[21] Bo Li, T Zhang, and T Xia. Vehicle detection from 3d lidar using fully convolutional network. arxiv 2016. arXiv preprint arXiv:1608.07916. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[22] Yangbin Lin, Cheng Wang, Dawei Zhai, Wei Li, and Jonathan Li. Toward better boundary preserved supervoxel segmentation for 3d point clouds. *ISPRS journal of photogrammetry and remote sensing*, 143:39–47, 2018. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 644 + ], + "type": "text", + "content": "[23] Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for efficient point cloud transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1200-1211, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25839" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[26] Jeremie Papon, Alexey Abramov, Markus Schoeler, and Florentin Worgotter. Voxel cloud connectivity segmentation-supervoxels for point clouds. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2027-2034, 2013. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "text", + "content": "[28] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems, 35:23192-23204, 2022. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[29] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17195-17204, 2023. 2, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 376 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 432 + ], + "type": "text", + "content": "[31] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "text", + "content": "[32] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, pages 945-953, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 544 + ], + "type": "text", + "content": "[33] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 577 + ], + "type": "text", + "content": "[34] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. ACM Transactions on Graphics (TOG), 42(4):1-11, 2023. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[35] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[36] Kilian Q Weinberger, John Blitzer, and Lawrence Saul. Distance metric learning for large margin nearest neighbor classification. Advances in neural information processing systems, 18, 2005. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[37] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 507 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[38] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4840-4851, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[39] Yanyang Xiao, Zhonggui Chen, Zhengtao Lin, Juan Cao, Yongjie Jessica Zhang, Yangbin Lin, and Cheng Wang. Merge-swap optimization framework for supervoxel generation from three-dimensional point clouds. Remote Sensing, 12(3):473, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[40] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "text", + "content": "[41] Guowen Zhang, Lue Fan, Chenhang He, Zhen Lei, Zhaoxiang Zhang, and Lei Zhang. Voxel mamba: Group-free state space models for point cloud based 3d object detection. arXiv preprint arXiv:2406.10700, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 361 + ], + "type": "text", + "content": "[42] Peiyan Zhang, Yuchen Yan, Xi Zhang, Chaozhuo Li, Senzhang Wang, Feiran Huang, and Sunghun Kim. Transgnn: Harnessing the collaborative power of transformers and graph neural networks for recommender systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1285-1295, 2024. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 553, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 417 + ], + "type": "text", + "content": "[43] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 418, + 555, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 555, + 462 + ], + "type": "text", + "content": "[44] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[45] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25840" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_content_list.json b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e8206e006d02a4b3c4dec892a6a812ea560d781d --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_content_list.json @@ -0,0 +1,1976 @@ +[ + { + "type": "text", + "text": "ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer", + "text_level": 1, + "bbox": [ + 117, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin Hu* Mingjia Li* Xiaojie Guo† Tianjin University", + "bbox": [ + 349, + 180, + 651, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jinhu,mingjiali}@tju.edu.cn,xj.max.guo@gmail.com", + "bbox": [ + 269, + 219, + 720, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shadows introduce challenges such as reduced brightness, texture deterioration, and color distortion in images, complicating a holistic solution. This study presents ShadowHack, a divide-and-conquer strategy that tackles these complexities by decomposing the original task into luminance recovery and color remedy. To brighten shadow regions and repair the corrupted textures in the luminance space, we customize LRNet, a U-shaped network with a rectified attention module, to enhance information interaction and recalibrate contaminated attention maps. With luminance recovered, CRNet then leverages cross-attention mechanisms to revive vibrant colors, producing visually compelling results. Extensive experiments on multiple datasets are conducted to demonstrate the superiority of ShadowHack over existing state-of-the-art solutions both quantitatively and qualitatively, highlighting the effectiveness of our design. Our code is available at https://github.com/lime-j/ShadowHack.", + "bbox": [ + 89, + 300, + 485, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shadow is a color as light is, but less brilliant; light and shadow are only the relation of two tones.", + "bbox": [ + 104, + 589, + 467, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paul Cezanne", + "bbox": [ + 349, + 619, + 462, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 648, + 220, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shadows are the absence of light, an everyday phenomenon that often shapes the aesthetic beauty of nature. However, shadows obscure the actual shape of instances, diminishing their visibility. Very likely, they deteriorate the performance of various machine vision tasks, such as object detection [21, 42], face recognition [62], and semantic segmentation [63], to name just a few. Consequently, shadow removal has become a critical research topic.", + "bbox": [ + 89, + 674, + 483, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have greatly improved the efficiency and accuracy of shadow removal. Specifically, DeshadowNet [43] marked a pioneering step with its end-to-end Convolutional Neural Network (CNN)", + "bbox": [ + 89, + 795, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b30d9a990595dcfae1f4171d5175f8f36c2153bb5f291532432ae46c6726f55.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 516, + 266, + 624, + 375 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a9cb6cd8ada88587366c5818631f8fae929cecd51810af087ce46845ea2370e.jpg", + "image_caption": [ + "(b) SG [47]" + ], + "image_footnote": [], + "bbox": [ + 625, + 267, + 717, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3aa9e851605414b79f8a90731ba1bb6904f6a30181ba428593d833e899c18569.jpg", + "image_caption": [ + "(c) HF [53]", + "Figure 1. Visual comparison between our method and state-of-the-art methods in both Luminance and RGB spaces." + ], + "image_footnote": [], + "bbox": [ + 718, + 267, + 808, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd5c55a873d199a5d7fbad2a127472d670dad0b5d9ea58fd1b849158d612fcb6.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 812, + 267, + 901, + 373 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "approach, while subsequent GAN-based methods [22, 36, 48, 61] further advanced the field by adopting adversarial learning frameworks, enabling effective unpaired training. Additionally, Wan et al. [47] introduced style transfer, and Li et al. [30] incorporated pre-trained inpainting priors to better restore textural information. Transformer-based architectures, such as ShadowFormer [13], HomoFormer [53], and RASM [32], have since been developed to restore shadow images more effectively, while diffusion-based approaches [14, 24, 41, 49] have also demonstrated promising results, particularly in complex scenarios where previous generative models often struggled.", + "bbox": [ + 511, + 445, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, the odyssey of removing shadows is still far from its end. As evidenced in Fig. 1, even state-of-the-art models fail to recover color and underlying texture accurately. The difficulty stems from the complex and intertwined nature of shadow degradation, which involves not only reduced brightness in shadow (umbra) regions [40], but also texture deterioration from the imaging process [30, 34], and color distortion from surface material properties and environmental color influences [7, 28, 47], an example of which is in Fig. 2. This complexity naturally suggests breaking down the original task, with its entangled interferences, into isolated and simpler sub-problems, each targeting a specific aspect and addressed independently.", + "bbox": [ + 511, + 628, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Following the above line of thought, some works [16, 27, 28] attempted to estimate light and color biases as scalars. Insights can also be drawn from the low-light enhancement literature, where the Retinex decomposition [1, 26] is commonly employed [2, 17, 18, 54]. Under the Retinex theory,", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 862, + 222, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. This work was supported by National Natural Science Foundation of China under Grant nos. 62372251 and 62072327.", + "bbox": [ + 91, + 876, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ec9418c33c512cb4125e6438f08416b5f97c607ef5d5be82c37ae6fa08b6e88.jpg", + "image_caption": [ + "(a) Color Shift" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 318, + 315 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0ffca9b6ead9a3ca2a7f9bd9937be4ff358346cb3ba574aa40c01eae3bddab6.jpg", + "image_caption": [ + "(b) Texture Distortion" + ], + "image_footnote": [], + "bbox": [ + 328, + 88, + 478, + 205 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/290f65f281a42d0fdb019ff16c0687a8a92e6b82359364d5157d761603a9d7f5.jpg", + "image_caption": [ + "(c) Illumination Underlit", + "Figure 2. Example of the complex degradation in a shadow image. a) Color shift in the umbra region; (b) Structural texture deterioration in the penumbra transition area; (c) Attenuated illumination levels across the shadow (Blue indicates lower illumination)." + ], + "image_footnote": [], + "bbox": [ + 330, + 224, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "an image can be decomposed into a structured illumination and a colored texture-rich counterpart. However, these approaches encounter challenges due to the inherent inhomogeneity of shadows. In other words, shadow removal extends beyond low-light enhancement and brings additional color biases to the reflectance map. This realization underscores the need for more sophisticated solutions to effectively disentangle the complex degradations introduced by shadows.", + "bbox": [ + 89, + 426, + 485, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work invents ShadowHack, an effective shadow removal framework that individually tackles illumination and color components. ShadowHack starts with a simple yet highly effective hack for shadow decomposition. Unlike conventional methods that decouple illumination from color and texture, we alternatively separate color from illumination and texture. Once the input images are decomposed into luminance and color components, two dedicated networks are established, i.e., the Luminance Restoration Network (LRNet) and the Color Regeneration Network (CRNet). LRNet, a lightweight transformer-based restoration model, focuses on recovering illumination and texture. To enable an effective reference for shadow region refinement, we introduce the RecTified Attention (RTA) Module, which employs a rectified mechanism on window-based attentions to support high-quality luminance restoration.", + "bbox": [ + 89, + 550, + 483, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Afterward, the processed luminance serves as a guide for querying accurate color information in shadow regions from the non-shadow areas. We inject color information via an ImageNet-pre-trained color encoder, which integrates color data into the intermediate features of the skip connections of our CRNet. This is facilitated by cross-attention modules that align the color in the shadow region with patches", + "bbox": [ + 89, + 795, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of similar luminance. Through this sequential processing, ShadowHack achieves effective shadow removal, supported by extensive experimental results. The primary contributions of this study are summarized as follows:", + "bbox": [ + 511, + 90, + 906, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Based on an in-depth analysis of shadow properties and the limitations in current end-to-end decomposition methods, we introduce ShadowHack, a novel shadow removal framework that breaks shadow removal into two distinct subtasks: luminance recovery and color regeneration.", + "- We identified the distraction phenomenon in the attention matrix of shadow removal networks, and propose RecTified Attention (RTA) to regain focus in attention layers with marginal computational overhead, enabling precise restoration of irregular texture and illumination.", + "- We design two specialized networks, LRNet and CRNet, tailored for luminance refinement and color regeneration. Together, these networks allow our model to achieve accurate and robust results, setting new state-of-the-art performance across benchmarking datasets." + ], + "bbox": [ + 513, + 152, + 908, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 392, + 655, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shadow Removal. Traditional methods, such as [57] and [45], leveraged principles of physics and mathematical statistics on factors like illumination [10, 60], color [25, 51], region [16], and gradient [11, 12]. However, these methods often struggle with complex real-world scenes, where shadows interact in nuanced ways with surrounding colors and textures. Deep learning has recently made significant strides in the field [27, 43, 48]. For instance, Qu et al. [43] proposed to extract both global and semantic features from images, marking an early attempt to integrate deep learning into shadow removal. Later, Wang et al. [48] proposed a multitask framework that performs shadow detection and removal concurrently, enhancing overall accuracy. Generative Adversarial Networks (GANs) have also been applied to this task. As a representative, RIS-GAN [61] incorporates residual and illumination features to refine shadow regions, while other works [22, 36] use unsupervised GANs to generate shadows for achieving cycle consistency. Li et al. [30] adopted the concept of inpainting, simultaneously using a shadow image encoder and a generative-pretrained image encoder to extract features. Zhu et al. [65] employed normalizing flow for a reversible link between shadow generation and removal. Additionally, Le et al. [28] proposed a linear model that adjusts the global lightness of shadow regions through physical modeling. More recently, diffusion models [46] have shown even greater generative capabilities than GANs. Guo et al. [14] pioneered the use of diffusion models for shadow removal, refining shadow masks throughout the process. Liu et al. [34] applied Retinex decomposition [18, 26] and used diffusion models to refine textures under the guidance of the obtained illumination. However, a primary drawback of these diffusion models is their intensive computational cost.", + "bbox": [ + 511, + 417, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11404", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ViT-based Restoration. Vision Transformers (ViTs) have demonstrated impressive performance across vision tasks, owing to their powerful attention mechanisms that can effectively capture long-range dependencies [3, 9, 19, 35, 52]. In image restoration, IPT [4] was among the first ViT-based solutions. Unfortunately, its dependence on large-scale paired datasets for training restricts its applicability when data is scarce. Subsequent models have sought to mitigate this limitation with more efficient architectures. Uformer [50] combined UNet with a window attention, yielding strong results, while Restormer [59] introduced transposed attention for efficient channel-wise processing. Further, X-Restormer [5] extended Restormer by integrating the OCA [6] module, providing a generalized framework for image restoration through rigorous analysis. Specific to the domain of shadow removal, ShadowFormer [13] introduced an attention mechanism that facilitates interaction between shadow and non-shadow regions, while HomoFormer [53] improved the transformer by integrating a shuffling mechanism into its transformer-based encoder. RASM [32] refreshed the window-based attention with a sliding-window mechanism, achieving more accurate results. More recently, OmniSR [55] leveraged DINO features and depth maps to incorporate semantic and geometric attention into self-attention, expanding its perception of image content and ability to distinguish shadows. Even though, as discussed in Sec. 1, there remains significant room for the community to improve the effectiveness of shadow removal.", + "bbox": [ + 89, + 90, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Analysis", + "text_level": 1, + "bbox": [ + 89, + 525, + 264, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Degradation in umbra primarily encompasses reduced brightness, color distortion, and texture degradation (e.g., noise and blurriness), as illustrated in Fig. 2. To better understand the types of involved degradation, we begin by formulating shadow effects via a modified image formation model as proposed in [1, 28, 45]. In this model, the observed pixel intensity $I(x,\\lambda)$ at a given spatial coordinate $x$ and wavelength $\\lambda$ is expressed as the product of shading $S(x,\\lambda)$ and albedo $A(x,\\lambda)$ as follows:", + "bbox": [ + 89, + 550, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, \\lambda) = S (x, \\lambda) \\cdot A (x, \\lambda). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 694, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In scenes with a single dominant light source (e.g., sunlight in outdoor settings), the shading for a pixel $x$ can be decomposed into direct $S_{d}(x, \\lambda)$ and ambient $S_{a}(x, \\lambda)$ components:", + "bbox": [ + 89, + 717, + 483, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, \\lambda) = \\left(S _ {d} (x, \\lambda) + S _ {a} (x, \\lambda)\\right) \\cdot A (x, \\lambda). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 785, + 483, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, when an object occludes the primary light source, shadow regions occur, and the pixel intensity becomes:", + "bbox": [ + 89, + 809, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, \\lambda) = a (x) \\cdot S _ {a} (x, \\lambda) \\cdot A (x, \\lambda), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 845, + 483, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $a(x) \\in [0,1]$ designates a spatially varying attenuation function. This non-uniform attenuation of light forms", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bd8ea2d5b45784a606493168347da4db9d4564b56f686c8a0965ee10bef0183c.jpg", + "image_caption": [ + "(a)ISTD+training set" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 705, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b26a5997b48d572781fb6515455d27baa78d670433afdf66f41a402fb59d314.jpg", + "image_caption": [ + "(b) $\\mathrm{ISTD + }$ testing set" + ], + "image_footnote": [], + "bbox": [ + 712, + 90, + 903, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0b480c837dc7342b80d92abe57dde87203ab4c23a3342a66382c841b206b364c.jpg", + "image_caption": [ + "(c) SRD training set" + ], + "image_footnote": [], + "bbox": [ + 514, + 195, + 707, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f67cd43ea73fb250a71d11e122431b896debc62479d76d50600abf26f57000c.jpg", + "image_caption": [ + "(d) SRD testing set", + "Figure 3. Color bias in shadow regions between the reflectances of shadow and shadow-free images by Retinex decomposition [18]." + ], + "image_footnote": [], + "bbox": [ + 712, + 195, + 903, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shadow. This involves not only illumination variance but also challenging inconsistency in texture[31], an example of which is illustrated in Fig. 2 (b) and (c). To obtain an RGB image $^{1}$ $I_{RGB} \\coloneqq \\{I_R, I_G, I_B\\}$ , the observed color intensity at each pixel is integrated across wavelengths as follows:", + "bbox": [ + 511, + 349, + 906, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} (x) = \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} S (x, \\lambda) \\cdot A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda , k \\in \\{R, G, B \\} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 436, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P_{k}(\\lambda)$ is the spectral sensitivity function of the camera for the $k$ -channel and $[\\lambda_{\\mathrm{min}}, \\lambda_{\\mathrm{max}}]$ defines the visible spectrum range. Previous shadow removal and low-light enhancement methods [15, 17, 18, 34] suggest that it would be neat to decompose lightness from the shadow image following the Retinex theory. In these works, the image is assumed to be decomposed into a one-channel illumination map $L(x)$ and an illumination-invariant counterpart (reflectance) $R_{k}(x)$ . For sRGB space and white-balanced images, the operation can be formulated as:", + "bbox": [ + 511, + 484, + 908, + 636 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k} (x) \\approx L (x) \\cdot R _ {k} (x) = L (x) \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda . \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 647, + 906, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We argue that shadow posits not just illumination map changes, but shifts that variants across the color. To quantify this, we plot the per-channel color difference between the input and ground-truth images in the reflectance map derived from a Retinex-based illumination estimator (Fig.3). The reflectance decomposition exhibits substantial color bias, particularly in the red and blue channels, violating Retinex's assumption of illumination-invariant reflectance. This color shift cannot be adequately addressed by previous Retinex-inspired decomposition methods [15, 34], as their core premise which reflectance remains uncontaminated by illumination effects are broken down in shadow regions.", + "bbox": [ + 511, + 696, + 908, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Here, we omit the impact from the ISP pipeline.", + "bbox": [ + 531, + 886, + 792, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31724fe29116adeed49ae31780ed0afe19912db570cde5388f8aa3a38f7eeba6.jpg", + "image_caption": [ + "(a) Luminance Restoration Network", + "Figure 4. The overall architecture of ShadowHack comprises two core components: (a) a luminance restoration network, which focuses on recovering illumination and texture, and (b) a color regeneration network dedicated to recalibrating harmonized colors for natural outputs." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 424, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ff812b5335c13c5202b750e44775930642917100f75250bcee5ac6b644c9020.jpg", + "image_caption": [ + "(b) Color Regeneration Network" + ], + "image_footnote": [], + "bbox": [ + 426, + 97, + 901, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel approach to address the triple degradation of reduced brightness, texture deterioration, and color distortion in a divide-and-conquer manner. The key insight is identifying a decomposition scheme that cleanly separates brightness and texture information from color information. An ideal decomposition would isolate a grayscale component capturing illumination and textural details with minimum color interference, alongside a pure color component that can be independently corrected. This separation would enable a targeted restoration process, which first recovers accurate textural details and then uses this restored foundation as a reliable reference for precise color regeneration.", + "bbox": [ + 88, + 448, + 485, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "YCbCr color space provides an optimal lossless decomposition for this task. The Y channel, which predominantly weights the green channel (0.587G compared to 0.299R and 0.114B), is significantly less contaminated by the color bias that predominantly affects red and blue channels in shadow regions. Additionally, the Y channel is specifically designed to preserve perceptually important texture details that align with human visual sensitivity. By first restoring this luminance component, we can recover accurate illumination and texture information without interference from color distortions. The restored luminance then serves as a geometric and semantic guide for subsequent color regeneration, ensuring spatial coherence between brightness and chrominance adjustments. This sequential enables dedicated networks to specialize in luminance refinement (LRNet) and chroma harmonization (CRNet) without mutual interference.", + "bbox": [ + 89, + 659, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 446, + 650, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A viable solution involves separating an image into texture $I_{t}$ and color $I_{c}$ components for texture restoration and color regeneration, respectively. We achieve this through an invertible decoupling function $D(\\cdot)$ , which splits an RGB image $I$ into a luminance component $I_{y}$ and a chrominance component $I_{c}$ :", + "bbox": [ + 511, + 472, + 908, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(I _ {t}, I _ {c}\\right) = \\mathcal {D} (I). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 563, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the conversion from RGB to YCbCr as $\\mathcal{D}(I)$ , where $I_{t}$ corresponds to the Y (luminance) channel, and $I_{c}$ comprises the Cb and Cr (chrominance) channels. This decomposition isolates brightness and texture degradation in $I_{t}$ while retaining color distortions in $I_{c}$ . The inverse function $\\mathcal{D}^{-1}$ reconstructs the final RGB image from processed components. This manner enables a divide-and-conquer strategy for shadow removal:", + "bbox": [ + 511, + 585, + 908, + 707 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} = \\mathcal {D} ^ {- 1} \\mathcal {C} (\\mathcal {R} (I _ {t}), I _ {c}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 719, + 906, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}(\\cdot)$ denotes the luminance restoration network (LRNet) and, $\\mathcal{C}(\\cdot)$ represents the color regeneration network (CRNet). $\\hat{I}$ is the shadow-free output, and $I_{c}$ serves as a conditioning variable for color regeneration. Additionally, $\\mathcal{D}^{-1}(\\cdot)$ is the inverse procedure of $\\mathcal{D}(\\cdot)$ .", + "bbox": [ + 511, + 750, + 908, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As in Fig. 4, ShadowHack operates in three stages: 1) Decomposition. The RGB image is decomposed via $\\mathcal{D}(\\cdot)$ into luminance $I_{t}$ and color $I_{c}$ components; 2) Luminance Restoration. LRNet processes $I_{t}$ to recover illumination and texture, yield in $\\hat{I}_t = \\mathcal{R}(I_t)$ ; 3) Color Regeneration.", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11406", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d21659e7b04d4d660d5c69b80a0ee1a655841de9a3c58fa3b961cd359adf5565.jpg", + "image_caption": [ + "Figure 5. The rectified attention module. We compute two sets of attention. In $Att_1$ , both texture $(I_t)$ and color $(I_c)$ are involved while in $Att_2$ , we calculate attention with only color components. The difference $Att_1 - \\lambda Att_2$ delivers more accurate interactions" + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 486, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CRNet harmonizes colors using $I_{c}$ and $\\hat{I}_{t}$ . Finally, the estimated luminance and color components are recombined using $\\mathcal{D}^{-1}(\\cdot)$ to produce the final shadow-free RGB result. Below, we detail LRNet and CRNet.", + "bbox": [ + 89, + 361, + 485, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Luminance Restoration Network", + "text_level": 1, + "bbox": [ + 89, + 436, + 380, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To eliminate shadows in $I_{t}$ , enhancing the brightness of the shadow areas and restoring their textures is essential. As illustrated in Fig. 4 (a), the shallow two stages and the refinement module in the luminance restoration network (LRNet) are constructed using a Local Range block (LRB), a Multi-head Transpose Attention (MTA) module, and a Feed-Forward Network (FFN), while a RecTified Attention (RTA) module replaces LRB in the deeper stages. The LRB is designed to extract local texture information through a depth-wise convolution block, capturing fine-grained details. The MTA, which lies in the channel attention block (CAB), following previous works [5, 59], reallocates the weights of channels to enhance the relevant features. The FFN further processes these extracted features to improve texture restoration. Finally, the RTA module, introduced in the deeper layers, captures more global information by focusing on regions close to shadow boundaries, enabling the network to refine luminance restoration in complex shadow regions.", + "bbox": [ + 88, + 460, + 485, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rectified Attention Module. In the shadow removal process, achieving robust enhancement in shadow regions requires restoration efforts to reference better surrounding, related non-shadow areas. Therefore, we replace the Local Range Block (LRB) with self-attention modules [20, 32] in the deeper layers to capture more distant information. Since the goal of LRNet is to cure shadow-contaminated texture and lightness, the feature from the deep layer (i.e., blocks with SAB equipped) becomes less and less discriminative to shadow. As a result, the attention matrix in deep layers, suffers from a distraction. In the first column of Fig. 6, the", + "bbox": [ + 89, + 734, + 486, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b776d461501287bc8fb4a296417a8d390c80c3f039f2155fb3ff5f8133c8f233.jpg", + "image_caption": [ + "Figure 6. An example of our rectified attention module. The star indicates the position of the query token. We exhibit the attention map from the first head of the encoder, latent, and decoder layers. The final attention is obtained by subtracting $\\mathrm{Att}_2$ from $\\mathrm{Att}_1$ . The attention matrix without rectify, $\\mathrm{Att}_1$ and $\\mathrm{Att}_2$ are normalized to [0, 1], while their differential result is in [-1, 1]." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model tends to focus on only a limited number of tokens which may be inaccurate for shadow area refinement. We address this via a RecTified Attention (RTA) mechanism to rectify the original attention weight that is re-injecting contaminated chrominance channels $I_{c}$ , by combining them into $Q$ and $K$ calculations (Please notice that the $V$ is still clear without chromance), resulting in $Att_{1}$ . Since this operation re-integrates color bias to the calculation, inspired by [58], we calculate a second set with only $I_{c}$ as $Att_{2}$ . By subtracting the two sets, LRNet has the flexibility to get rid of chrominance information if needed.", + "bbox": [ + 511, + 356, + 908, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be specific, we obtain the features $F_{t}$ corresponding to $I_{t}$ through convolution with downsampling, and compute the softmax attention map for $F_{t}$ and $F_{c}$ , as shown in Fig. 5. The query, key, and value matrices are formed via:", + "bbox": [ + 511, + 523, + 908, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {Q}, K _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {K}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 590, + 905, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {2} = F _ {c} W _ {2} ^ {Q}, K _ {2} = F _ {c} W _ {2} ^ {K}, V = F _ {t} W ^ {V},\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 613, + 854, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W$ matrices are learnable projection parameters and \"; refers to the concatenation of the two features. Next, the above two sets of queries and keys are used to compute their respective attentions simply by:", + "bbox": [ + 511, + 638, + 906, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t} _ {i} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {i} K _ {i} ^ {\\mathrm {T}}}{\\sqrt {d}} + B\\right), i \\in \\{1, 2 \\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 705, + 906, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ denotes the relative position bias. This allows us to apply the rectified attention mechanism as follows:", + "bbox": [ + 511, + 746, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R T A} \\left(F _ {t}, F _ {c}\\right) = \\left(\\operatorname {A t t} _ {1} - \\lambda \\cdot \\operatorname {A t t} _ {2}\\right) V, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 785, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a reweighting factor. As illustrated in Fig. 6, such a rectified attention mechanism creates a strong affinity with well-lit reference, urging the model to focus on relevant content. As validated in Sec. 5.3, our rectified attention can focus on more non-shadow area information to transfer robust clues for shadow removal.", + "bbox": [ + 511, + 809, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11407", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/62d0813550cacd0b8a947580368a6a27d8866ca560854e2ecc8e239adb1ab5a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsMethodShadow Region (S)Non-Shadow Region (NS)All Image (ALL)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
ISTD+DHAN [8] (AAAI'20)33.080.9889.4927.280.9727.3925.780.9587.74
G2R [36] (CVPR'21)33.880.9788.7135.940.9772.8130.850.9463.78
BMNet [65] (CVPR'22)38.170.9915.7237.950.9862.4234.340.9742.93
SG-ShadowNet [47] (ECCV'22)36.800.9905.9335.570.9782.9232.460.9623.41
ShadowFormer [13] (AAAI'23)39.670.9925.2138.820.9832.3035.460.9732.80
ShadowDiffusion [14] (CVPR'23)39.82-4.9038.90-2.3035.72-2.70
Li et al. [30] (ICCV'23)38.460.9895.9337.270.9772.9034.140.9603.39
Liu et al. [34] (AAAI'24)38.040.9905.6939.150.9842.3134.960.9682.87
RASM [32] (MM'24)40.730.9934.4139.230.9852.1736.160.9762.53
Homofformer [53] (CVPR'24)39.470.9934.7238.730.9842.2335.340.9752.64
OmniSR+GM [55] (AAAI'25)------34.200.973-
ShadowHack (Ours)40.560.9934.4639.660.9872.0936.310.9772.48
SRDDHAN [8] (AAAI'20)33.670.9788.9434.790.9794.8030.510.9495.67
BMNet [65] (CVPR'22)35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNet [47] (ECCV'22)--7.53--2.97--4.23
ShadowFormer [13] (AAAI'23)36.910.9895.9036.220.9893.4432.900.9584.04
ShadowDiffusion [14] (CVPR'23)38.720.9874.9837.780.9853.4434.730.9703.63
Li et al. [30] (ICCV'23)39.330.9856.0935.610.9672.9733.170.9403.83
Liu et al. [34] (AAAI'24)36.510.9835.4937.710.9863.0033.480.9673.66
DeS3 [24] (AAAI'24)37.450.9845.8838.120.9882.8334.110.9683.72
RASM [32] (MM'24)37.910.9885.0238.700.9922.7234.460.9763.37
Homofformer [53] (CVPR'24)38.810.9874.2539.450.9882.8535.370.9723.33
OmniSR+GM [55] (AAAI'25)------34.560.977-
ShadowHack (Ours)39.470.9914.1840.060.9942.3835.940.9822.90
", + "bbox": [ + 94, + 89, + 903, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on the ISTD+ and SRD datasets. The best results are in bold, while the second-best ones are underlined.", + "bbox": [ + 102, + 429, + 890, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Color Regeneration Network", + "text_level": 1, + "bbox": [ + 89, + 463, + 349, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After restoring texture and lighting information in the luminance channel, the next task is to accurately regenerate color. Inspired by exemplar-based colorization models [23, 29, 39], we leverage contextual cues and condition-based guidance to improve color fidelity and ensure alignment with the semantic content of the images. Concretely, we propose a dual-encoder network, termed CRNet, which integrates a multiscale color feature extractor into the U-shaped architecture of LRNet. The color encoder is a ConvNext-v2utto model [37] pre-trained on ImageNet-21k with 2M parameters, which takes the Cb and Cr channels as input and extracts color features across multiple scales, as illustrated in Fig. 4 (b). To achieve precise color restoration, the network shall reference regions with similar/identical material properties and accurate color, using them as reliable guides for color regeneration. The outputs generated by LRNet provide qualified indexing information to facilitate this purpose. Therefore, we introduce an attention-based mechanism where luminance features serve as the query and key to compute similarity, while color features act as the value, aggregating essential color information. This approach injects color features into the skip connections of the U-shaped backbone, remarkably enriching the network's color regeneration capability.", + "bbox": [ + 89, + 488, + 485, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiple Checkpoint Training. We observe that the luminance network often yields better performance on the training set, which suggests the color regeneration network faces challenges in adapting to imperfect luminance out", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "puts encountered during testing. To address this issue, we employ a simple yet effective way to overcome this issue, where multiple early-stage checkpoints from the luminance network training are randomly selected as candidates. This technique, applied only during training, strengthens the robustness of the color network without incurring additional computational costs at inference time.", + "bbox": [ + 511, + 464, + 908, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimental Validation", + "text_level": 1, + "bbox": [ + 511, + 580, + 743, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 606, + 705, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our models are implemented in PyTorch, and the entire training process utilizes the AdamW optimizer [38] with momentum set to (0.9, 0.999) and a weight decay of $10^{-2}$ . The initial learning rate is set to $2 \\times 10^{-4}$ . During training, we randomly crop the training data into patches of size $384 \\times 384$ . For the base network of each stage, we use a four-layer encoder-decoder structure (from L1 to L4), consisting of three downsampling and three upsampling operations. The dimension of the first layer is 32. Our RTA modules are integrated into L3 and L4, with an overlapping ratio set to 0.5 and a rectified mechanism $\\lambda_0$ value of 0.7. For training, we use L1 and VGG loss terms as employed in previous works [32, 34]. For detailed information on the implementation of each stage and training setup, please refer to the supplementary material.", + "bbox": [ + 511, + 628, + 906, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments on the following benchmark datasets for shadow removal: 1) Adjusted ISTD (ISTD+) dataset [27], which refines the ISTD dataset [48]", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11408", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f37dc20ab13f136ec8a4e765509da6d2bc32e6352818be0a953c6938d0ef5881.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 197, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38461c8262550ce788585328fe55192401919a22ea5397c1b931647994569d90.jpg", + "image_caption": [ + "(b) BMNet [65]" + ], + "image_footnote": [], + "bbox": [ + 200, + 88, + 300, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee2b8fb14f897d8daadf792f0865955b25926c4636d26b4d45620a5d1d08fd1c.jpg", + "image_caption": [ + "(c) SG [47]" + ], + "image_footnote": [], + "bbox": [ + 300, + 88, + 398, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4faf09f58d3a56a94f4fdcf7105442932506e3e0f65d83d7b826d20ee6d948b6.jpg", + "image_caption": [ + "(d) SF [13]" + ], + "image_footnote": [], + "bbox": [ + 400, + 88, + 498, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7c77ea1bc50baad530eb6a49cb24b724358bc9538983d25c1f5b8795db5fea9.jpg", + "image_caption": [ + "(e) HF [53]" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 596, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b98aed533779497ca16e14fe5cec61d7bda9d81cd4718725c57641e5fe7d07f2.jpg", + "image_caption": [ + "(f) RASM [32]" + ], + "image_footnote": [], + "bbox": [ + 598, + 88, + 696, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55218b3e115f37398494d61e6b0b5b564ce2fe972f5bc4b1c7bb44afc06877f2.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 88, + 797, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6282831f4b3114f3622fd3f992e37aa01ba49b160feb674bd24de4b07d445c9d.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 799, + 88, + 895, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d528eeb5c62f287eb153184cf74fb402d938a8d6c3b73fb05e01ddc364e6add3.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 99, + 232, + 199, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35696371e640e1c04de01e03304e5a359a2d7b046b26bcb47b15a8255c03fba6.jpg", + "image_caption": [ + "(b) BMNet [65]" + ], + "image_footnote": [], + "bbox": [ + 200, + 232, + 299, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e01c5c66741d828e27dc00b9be5b3471555d30ed2f3609827ad098366862834c.jpg", + "image_caption": [ + "Figure 7. Visual comparison on a sample from the $\\mathrm{ISTD + }$ dataset.", + "(c) SG [47]" + ], + "image_footnote": [], + "bbox": [ + 300, + 232, + 398, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b3c3d3751b4ab75ba8d7c82a2747fc50a8eaee09d463f3e0b563466f60a56c0.jpg", + "image_caption": [ + "(d) DMTN [33]" + ], + "image_footnote": [], + "bbox": [ + 400, + 232, + 496, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5a8f8c7e679f52a9d15485a87be9bb71a004a3aaf63737a5bcaecf264773fae.jpg", + "image_caption": [ + "(e) DeS3 [24]" + ], + "image_footnote": [], + "bbox": [ + 500, + 232, + 596, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b1b901d86cb562b501a1a42add838f85c882f88bb9e97103d34f213759541c3.jpg", + "image_caption": [ + "(f) HF [53]" + ], + "image_footnote": [], + "bbox": [ + 599, + 232, + 697, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5434a03aec5755c8c6d1aff0929e048b27f43b9bc62d2b8765be8d21b40969ee.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 232, + 795, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1efbff3d2ff6e82cc67f06b84527da8df2878520b9b6bb7a8fd343d6f04a5b75.jpg", + "image_caption": [ + "(h) GT" + ], + "image_footnote": [], + "bbox": [ + 799, + 232, + 897, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7eb3515a06431e012162bc8850ba940723b5a37493ea99107ce8077c346f503a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricRGB InputYCbCr InputRetinex-basedOurs
PSNR↑36.1635.6235.8536.31
RMSE↓2.532.622.632.48
", + "bbox": [ + 94, + 383, + 482, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study of the proposed decoupling framework on the ISTD+ dataset.", + "bbox": [ + 89, + 443, + 482, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by correcting the illumination discrepancies between the ground-truth and input images, includes 1330 training and 540 testing triplets (shadow image, shadow-free image, and shadow mask); 2) SRD dataset [43] consists of 2680 training and 408 testing pairs. We adopt the predicted masks provided by DHAN [8] for training and testing following previous works [32, 53, 65]; 3) UIUC dataset [16] provides 76 images with shadow mask for testing; and 4) UCF dataset [64] provided by Jin et al. [24] contains shadow images without shadow-free references.", + "bbox": [ + 89, + 487, + 483, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. To evaluate the performance of our proposed framework, we utilize the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE). PSNR and SSIM are both calculated in the RGB space, with higher values indicating better results. For RMSE, which is calculated in the LAB color space, the lower values indicate better results. To be consistent with previous works [32, 53, 65], we inference the full-resolution image and then evaluate the images at a resolution of $256 \\times 256$ .", + "bbox": [ + 89, + 638, + 483, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with Prior Arts", + "text_level": 1, + "bbox": [ + 89, + 801, + 344, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with the state-of-the-art shadow removal approaches on benchmark datasets, including DHAN [8], G2R [36], BMNet [65], SG-ShadowNet [47], ShadowFormer [13], ShadowDiffusion [14], Li et al. [30], Liu et al. [34], DeS3 [24], Homoformer [53], RASM [32]", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a26935eb81525fa360f2b4d877dcb2328eeb68f2080d82f581234a5e694066c8.jpg", + "image_caption": [ + "Figure 8. Visual comparison on a sample from the SRD dataset.", + "(a) Input", + "Figure 9. Ablation study on rectified attention. The results are displayed in the luminance space." + ], + "image_footnote": [], + "bbox": [ + 517, + 383, + 612, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/049b438133538ce0dcba1405a127464561a4b5c0848076ef30153af7a15d3d19.jpg", + "image_caption": [ + "(b) RGB Input" + ], + "image_footnote": [], + "bbox": [ + 614, + 383, + 709, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d02916a5f0aa03a8de21bba3c1d49a866486b11c0e458eb8b3e99409fce5bda.jpg", + "image_caption": [ + "(c) w/o rectify" + ], + "image_footnote": [], + "bbox": [ + 710, + 383, + 805, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80c2e9c2bec6ce90bc368a60eb8796fb9ce8f43c6dac4ebcd090640bd4b636cd.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 807, + 383, + 903, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and OmniSR+GM [55] (GM means the version with shadow mask), as shown in Tab. 1. For a fair comparison, we use results published by the authors. Regarding missing data, we utilize available images or checkpoints provided by the authors and evaluate them via the same metrics.", + "bbox": [ + 511, + 534, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed method achieves the highest scores in most metrics on the ISTD+ dataset compared to other state-of-the-art methods. Specifically, our model attains a PSNR of 36.31 for all image regions, surpassing the second-highest result of 36.16 by 0.15. The high PSNR highlights our model's ability to recover texture details effectively, demonstrating superior reconstruction quality. Additionally, the RMSE value of 2.48 is the lowest, which indicates that our recoloring module plays a crucial role in accurately restoring color information and reducing errors across the entire image. From the magnified areas in Figs. 7 and 8, it can be observed that our method not only effectively corrects color bias but also provides superior edge correction compared to other methods. The second part of Tab. 1 presents a comparison of our results on the SRD dataset against other state-of-the-art methods. It can be observed that our method achieves a significant improvement in RMSE (-0.43), further demonstrating the substantial benefit of the decoupling approach for color recovery.", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11409", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/994225a5e08c4ceb92284091c0edad6794da017c42ce13a2c9ecc345309783ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodShadow PSNR ↑Non-Shadow PSNR ↑All PSNR ↑
[Ft]39.9641.2736.60
[Ft; Fc]40.1741.2936.76
[Ft]&[Ft]40.0241.1836.58
[Ft; Fc]&[Ft; Fc]40.3541.3736.87
[Ft; Fc]&[Fc]40.3641.4036.90
", + "bbox": [ + 96, + 88, + 482, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/660c8b6c69f8a13b2ab8bdbce36d0fcf7f5d64a4b2dc20a6045fe1cb4a34b2d7.jpg", + "image_caption": [ + "(a) Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 169, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d372ecc92047120dceb1a5e38398083872bf22326e754e8c86e1eb42befd0790.jpg", + "image_caption": [ + "(b) SG [47]" + ], + "image_footnote": [], + "bbox": [ + 171, + 253, + 248, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8f5ad926afbd21af8d9e7c10cfaf1ecaa1d4a9aa0326af6ce87e054a392a0b5.jpg", + "image_caption": [ + "(c) SF [13]" + ], + "image_footnote": [], + "bbox": [ + 250, + 253, + 325, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d3c0c83baef97b872903b7eeab7379ad3b0f0bdcba79be90af533bc8c764f62.jpg", + "image_caption": [ + "(d) HF [53]" + ], + "image_footnote": [], + "bbox": [ + 326, + 253, + 401, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66369a419170a8557032737d5f81ce7599e47d9b7f1f2b045be7758c0d41a4bd.jpg", + "image_caption": [ + "(e) Ours", + "Figure 10. Visual comparisons on the UIUC and UCF datasets." + ], + "image_footnote": [], + "bbox": [ + 403, + 253, + 480, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 438, + 243, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the decoupling framework. To validate the effectiveness of our decoupling framework, we fine-tune our baseline model structure to adapt to RGB and YCbCr input/output for end-to-end shadow removal and compare its performance with our decoupling framework. As shown in Tab. 2, our decoupling approach significantly improves both PSNR and RMSE. Moreover, based on Retinex decomposition, we separately recover L and R, and compare it with our decoupling method. Detailed architectures are provided in the supplementary material. This improvement indicates that this strategy benefits texture recovery and color correction.", + "bbox": [ + 89, + 460, + 483, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of rectified attention module. We demonstrate the effectiveness of the rectified attention module in Tab. 3, by conducting experiments on rectification mechanisms. Furthermore, extra experiments are conducted on different feature combinations of rectified approaches to explore more suitable configurations. $[X]$ & $[Y]$ in Tab. 3 means performing a subtraction calculation of corresponding maps for rectification between $X$ and $Y$ . $[X]$ alone denotes only one pair of $Q$ and $K$ is used, where both inputs are $X$ . Clearly, leveraging color information, even with inherent bias, can allow shadow regions to better reference surrounding areas, resulting in significant performance improvements. As for potential color bias, rectification refines the attention map for greater precision. Furthermore, we find that $[F_t; F_c] \\& [F_c]$ slightly outperforms $[F_t; F_c] \\& [F_t; F_c]$ , and thus we ultimately choose $[F_t; F_c] \\& [F_c]$ . Fig. 9 visualizes the effectiveness of rectification, which has a more precise result with less edge remaining.", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8d8d0dc622090f4aaed4fbaad745f01ceebecd55357b6c4431d136e7c27a5e53.jpg", + "table_caption": [ + "Table 3. Ablation study of the proposed rectified attention module on the SRD dataset. We calculate PSNRs on the luminance." + ], + "table_footnote": [], + "table_body": "
MethodShadow RMSE ↓Non-Shadow RMSE ↓All RMSE ↓
w/o cross-att.5.782.643.39
w/o multi checkpt.5.012.593.28
Ours4.182.382.90
", + "bbox": [ + 519, + 88, + 903, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study of the CRNet on the SRD dataset.", + "bbox": [ + 535, + 178, + 880, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e4f5100a07a7580a95cde3e5005b6eb7de21db683f26ef18cd16c6fa2e0e0c6.jpg", + "image_caption": [ + "(a) dirty mask" + ], + "image_footnote": [], + "bbox": [ + 516, + 198, + 612, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f663130dbcec884f9580ae68d11e253a7b5d8676b2dfd9a0683e27955168d0dc.jpg", + "image_caption": [ + "Figure 11. Robustness analysis under corrupted mask. The accurate mask is highlighted in red. For fairness, we only involve mask refinement network here, and it is not involved in any comparison." + ], + "image_footnote": [], + "bbox": [ + 614, + 198, + 709, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/125cdfd79baba77ff3c038f7d95951382cace7ac25bf281c2f8e2522b1ad9138.jpg", + "image_caption": [ + "(b) refined mask (c) w/o refinement" + ], + "image_footnote": [], + "bbox": [ + 710, + 198, + 805, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f8bd731dc0ee73f9930104a2041d3114a066b98006a5530a43e9bc458b54b2a6.jpg", + "image_caption": [ + "(d) w refinement" + ], + "image_footnote": [], + "bbox": [ + 807, + 198, + 901, + 248 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the color restoration network. To validate the effectiveness of the key component and design in the CRNet, we replace the cross-attention-based color feature injection with concatenation along the feature channel. As demonstrated in Tab. 4, if cross-attention is replaced, the accuracy has dropped by a large margin, indicating the effectiveness of the cross-attention. Furthermore, the model suffers from a misalignment between training and testing luminance results if no previous checkpoints are involved.", + "bbox": [ + 511, + 316, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Robustness Analysis", + "text_level": 1, + "bbox": [ + 511, + 463, + 705, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to other datasets. To demonstrate the generalization capability of ShadowHack, we evaluate it on the UIUC [16] and UCF [64] datasets with our $\\mathrm{ISTD + }$ pretrained model. The results are presented in Fig. 10, where ShadowHack achieves the cleanest shadow removal with minimal color bias. More comparisons are available in the supplementary material.", + "bbox": [ + 511, + 484, + 906, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to inaccurate masks. To get rid of the performance degradation brought by inaccurate masks, we propose a mask refine network, which corrects the mask for better generalization. After our remediation, the result is recovered from inaccurate enlightening, as in Fig. 11 shows. More details are available in the supplementary material.", + "bbox": [ + 511, + 590, + 906, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposed a novel framework, ShadowHack, which decouples images into luminance and color components, enabling targeted processing for more effective shadow removal. This approach can well address limitations in existing end-to-end models by enhancing the use of color information to improve both texture and luminance recovery. To further boost shadow removal, a rectified attention mechanism has been designed to boost the reference ability of self-attention. Additionally, the color regeneration network has demonstrated its critical role in accurately correcting image colors. Extensive experimental results have underscored the superior performance and robustness of our model.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11410", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2(3-26):2, 1978. 1, 3", + "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In ICCV, pages 12470-12479, 2023. 1", + "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640, 2021. 3", + "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 3", + "[5] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. arXiv preprint arXiv:2310.11881, 2023. 3, 5", + "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In CVPR, pages 22367-22377, 2023. 3", + "[7] Michael E Churma. Blue shadows: physical, physiological, and psychological causes. Applied optics, 33(21):4719-4722, 1994. 1", + "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting GAN. In AAAI, pages 10680-10687, 2020. 6, 7", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[10] Graham D. Finlayson, Steven D. Hordley, and Mark S. Drew. Removing shadows from images. In ECCV (4), pages 823-836, 2002. 2", + "[11] Graham D. Finlayson, Steven D. Hordley, Cheng Lu, and Mark S. Drew. On the removal of shadows from images. IEEE Trans. Pattern Anal. Mach. Intell., 28(1):59-68, 2006. 2", + "[12] Graham D. Finlayson, Mark S. Drew, and Cheng Lu. Entropy minimization for shadow removal. Int. J. Comput. Vis., 85(1): 35-57, 2009. 2", + "[13] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 1, 3, 6, 7, 8, 2, 4", + "[14] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 6, 7", + "[15] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In ICCV, pages 12999-13008. IEEE, 2023. 3" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2956-2967, 2013. 1, 2, 7, 8", + "[17] Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness. Int. J. Comput. Vis., 131(1): 48-66, 2023. 1, 3, 2", + "[18] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: low-light image enhancement via illumination map estimation. IEEE Trans. Image Process., 26(2):982-993, 2017. 1, 2, 3", + "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 3", + "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 5", + "[21] Jun-Wei Hsieh, Wen-Fong Hu, Chia-Jung Chang, and Yung-Sheng Chen. Shadow elimination for effective moving object detection by gaussian shadow modeling. Image Vis. Comput., 21(6):505-516, 2003. 1", + "[22] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2", + "[23] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Trans. Graph., 41(6):205:1-205:16, 2022. 6", + "[24] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T. Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 1, 6, 7, 2, 5", + "[25] Salman Hameed Khan, Mohammed Bennamoun, Ferdous Ahmed Sohel, and Roberto Togneri. Automatic shadow detection and removal from a single image. IEEE Trans. Pattern Anal. Mach. Intell., 38(3):431-446, 2016. 2", + "[26] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 1, 2", + "[27] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8577-8586, 2019. 1, 2, 6", + "[28] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9088-9101, 2022. 1, 2, 3", + "[29] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809, 2020. 6", + "[30] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor W. Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13009-13018, 2023. 1, 2, 6, 7", + "[31] Feng Liu and Michael Gleicher. Texture-consistent shadow removal. In ECCV (4), pages 437-450. Springer, 2008. 3", + "[32] Hengxing Liu, Mingjia Li, and Xiaojie Guo. Regional attention for shadow removal. In ACM Multimedia, 2024. 1, 3, 5, 6, 7, 2, 4", + "[33] Jiawei Liu, Qiang Wang, Huijie Fan, Wentao Li, Liangqiong Qu, and Yandong Tang. A decoupled multi-task network for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shadow removal. IEEE Trans. Multim., 25:9449-9463, 2023. 7, 1, 5, 6", + "[34] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson W. H. Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 1, 2, 3, 6, 7", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3", + "[36] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 1, 2, 6, 7", + "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11966-11976, 2022. 6, 2", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[39] Peng Lu, Jinbei Yu, Xujun Peng, Zhaoran Zhao, and Xiaojie Wang. Gray2colornet: Transfer more colors from reference image. In ACM Multimedia, pages 3210-3218, 2020. 6", + "[40] David K. Lynch. Shadows. Appl. Opt., 54(4):B154-B164, 2015. 1", + "[41] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In WACV, pages 4313-4322, 2024. 1", + "[42] Sohail Nadimi and Bir Bhanu. Physical models for moving shadow and object detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 26(8):1079-1087, 2004. 1", + "[43] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson W. H. Lau. Desshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 2308-2316, 2017. 1, 2, 7", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 2", + "[45] Yael Shor and Dani Lischinski. The shadow meets the mask: Pyramid-based shadow removal. Comput. Graph. Forum, 27 (2):577-586, 2008. 2, 3", + "[46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 2", + "[47] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV (19), pages 361-378, 2022. 1, 6, 7, 8, 3, 4, 5", + "[48] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 1, 2, 6", + "[49] Tianyu Wang, Jianming Zhang, Haitian Zheng, Zhihong Ding, Scott Cohen, Zhe Lin, Wei Xiong, Chi-Wing Fu, Luis Figueroa, and Soo Ye Kim. Metashadow: Object-centered shadow detection, removal, and synthesis. In CVPR, pages 28252-28262, 2025. 1" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17662-17672, 2022. 3", + "[51] Tai-Pang Wu and Chi-Keung Tang. A bayesian approach for shadow extraction from a single image. In ICCV, pages 480–487, 2005. 2", + "[52] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4784-4793, 2022. 3", + "[53] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 1, 3, 6, 7, 8, 2, 4, 5", + "[54] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. Int. J. Comput. Vis., 132(9):3610-3632, 2024. 1", + "[55] Jiamin Xu, Zelong Li, Yuxin Zheng, Chenyu Huang, Renshu Gu, Weiwei Xu, and Gang Xu. Omnisr: Shadow removal under direct and indirect lighting. arXiv preprint arXiv:2410.01719, 2024. 3, 6, 7, 2", + "[56] Han Yang, Tianyu Wang, Xiaowei Hu, and Chi-Wing Fu. Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12687-12698, 2023. 2", + "[57] Qingxiong Yang, Kar-Han Tan, and Narendra Ahuja. Shadow removal using bilateral filtering. IEEE Trans. Image Process., 21(10):4361-4368, 2012. 2", + "[58] Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. arXiv preprint arXiv:2410.05258, 2024. 5, 1", + "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729, 2022. 3, 5", + "[60] Ling Zhang, Qing Zhang, and Chunxia Xiao. Shadow remover: Image shadow removal based on illumination recovering optimization. IEEE Trans. Image Process., 24(11): 4623-4636, 2015. 2", + "[61] Ling Zhang, Chengjiang Long, Xiaolong Zhang, and Chunxia Xiao. RIS-GAN: explore residual and illumination with generative adversarial networks for shadow removal. In AAAI, pages 12829-12836, 2020. 1, 2", + "[62] Wuming Zhang, Xi Zhao, Jean-Marie Morvan, and Liming Chen. Improving shadow suppression for illumination robust face recognition. IEEE Trans. Pattern Anal. Mach. Intell., 41 (3):611-624, 2019. 1", + "[63] Wei Zhou, Julie Stephany Berrio, Stewart Worrall, and Eduardo M. Nebot. Automated evaluation of semantic segmentation robustness for autonomous driving. IEEE Trans. Intell. Transp. Syst., 21(5):1951-1963, 2020. 1", + "[64] Jiejie Zhu, Kegan G. G. Samuel, Syed Zain Masood, and Marshall F. Tappen. Learning to recognize shadows in monochromatic natural images. In CVPR, pages 223-230, 2010. 7, 8" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11412", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[65] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5617-5626, 2022. 2, 6, 7, 1, 3, 4, 5", + "bbox": [ + 91, + 90, + 486, + 136 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11413", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_model.json b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e250c3408810950ba4a21f810038e339fc022bc0 --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_model.json @@ -0,0 +1,3038 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.181, + 0.652, + 0.22 + ], + "angle": 0, + "content": "Jin Hu* Mingjia Li* Xiaojie Guo† Tianjin University" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.22, + 0.722, + 0.234 + ], + "angle": 0, + "content": "{jinhu,mingjiali}@tju.edu.cn,xj.max.guo@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.558 + ], + "angle": 0, + "content": "Shadows introduce challenges such as reduced brightness, texture deterioration, and color distortion in images, complicating a holistic solution. This study presents ShadowHack, a divide-and-conquer strategy that tackles these complexities by decomposing the original task into luminance recovery and color remedy. To brighten shadow regions and repair the corrupted textures in the luminance space, we customize LRNet, a U-shaped network with a rectified attention module, to enhance information interaction and recalibrate contaminated attention maps. With luminance recovered, CRNet then leverages cross-attention mechanisms to revive vibrant colors, producing visually compelling results. Extensive experiments on multiple datasets are conducted to demonstrate the superiority of ShadowHack over existing state-of-the-art solutions both quantitatively and qualitatively, highlighting the effectiveness of our design. Our code is available at https://github.com/lime-j/ShadowHack." + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.59, + 0.468, + 0.62 + ], + "angle": 0, + "content": "Shadow is a color as light is, but less brilliant; light and shadow are only the relation of two tones." + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.621, + 0.463, + 0.635 + ], + "angle": 0, + "content": "Paul Cezanne" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.649, + 0.222, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Shadows are the absence of light, an everyday phenomenon that often shapes the aesthetic beauty of nature. However, shadows obscure the actual shape of instances, diminishing their visibility. Very likely, they deteriorate the performance of various machine vision tasks, such as object detection [21, 42], face recognition [62], and semantic segmentation [63], to name just a few. Consequently, shadow removal has become a critical research topic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Recent advancements in deep learning have greatly improved the efficiency and accuracy of shadow removal. Specifically, DeshadowNet [43] marked a pioneering step with its end-to-end Convolutional Neural Network (CNN)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.267, + 0.625, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.378, + 0.597, + 0.391 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.268, + 0.718, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.378, + 0.697, + 0.39 + ], + "angle": 0, + "content": "(b) SG [47]" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.268, + 0.81, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.378, + 0.79, + 0.39 + ], + "angle": 0, + "content": "(c) HF [53]" + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.268, + 0.903, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.829, + 0.378, + 0.875, + 0.389 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.402, + 0.908, + 0.431 + ], + "angle": 0, + "content": "Figure 1. Visual comparison between our method and state-of-the-art methods in both Luminance and RGB spaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.909, + 0.627 + ], + "angle": 0, + "content": "approach, while subsequent GAN-based methods [22, 36, 48, 61] further advanced the field by adopting adversarial learning frameworks, enabling effective unpaired training. Additionally, Wan et al. [47] introduced style transfer, and Li et al. [30] incorporated pre-trained inpainting priors to better restore textural information. Transformer-based architectures, such as ShadowFormer [13], HomoFormer [53], and RASM [32], have since been developed to restore shadow images more effectively, while diffusion-based approaches [14, 24, 41, 49] have also demonstrated promising results, particularly in complex scenarios where previous generative models often struggled." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Despite these advancements, the odyssey of removing shadows is still far from its end. As evidenced in Fig. 1, even state-of-the-art models fail to recover color and underlying texture accurately. The difficulty stems from the complex and intertwined nature of shadow degradation, which involves not only reduced brightness in shadow (umbra) regions [40], but also texture deterioration from the imaging process [30, 34], and color distortion from surface material properties and environmental color influences [7, 28, 47], an example of which is in Fig. 2. This complexity naturally suggests breaking down the original task, with its entangled interferences, into isolated and simpler sub-problems, each targeting a specific aspect and addressed independently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Following the above line of thought, some works [16, 27, 28] attempted to estimate light and color biases as scalars. Insights can also be drawn from the low-light enhancement literature, where the Retinex decomposition [1, 26] is commonly employed [2, 17, 18, 54]. Under the Retinex theory," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.223, + 0.876 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.482, + 0.899 + ], + "angle": 0, + "content": "†Corresponding author. This work was supported by National Natural Science Foundation of China under Grant nos. 62372251 and 62072327." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.319, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.319, + 0.246, + 0.332 + ], + "angle": 0, + "content": "(a) Color Shift" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.089, + 0.48, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.209, + 0.468, + 0.223 + ], + "angle": 0, + "content": "(b) Texture Distortion" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.226, + 0.481, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.319, + 0.481, + 0.333 + ], + "angle": 0, + "content": "(c) Illumination Underlit" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.486, + 0.402 + ], + "angle": 0, + "content": "Figure 2. Example of the complex degradation in a shadow image. a) Color shift in the umbra region; (b) Structural texture deterioration in the penumbra transition area; (c) Attenuated illumination levels across the shadow (Blue indicates lower illumination)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.486, + 0.548 + ], + "angle": 0, + "content": "an image can be decomposed into a structured illumination and a colored texture-rich counterpart. However, these approaches encounter challenges due to the inherent inhomogeneity of shadows. In other words, shadow removal extends beyond low-light enhancement and brings additional color biases to the reflectance map. This realization underscores the need for more sophisticated solutions to effectively disentangle the complex degradations introduced by shadows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.485, + 0.793 + ], + "angle": 0, + "content": "This work invents ShadowHack, an effective shadow removal framework that individually tackles illumination and color components. ShadowHack starts with a simple yet highly effective hack for shadow decomposition. Unlike conventional methods that decouple illumination from color and texture, we alternatively separate color from illumination and texture. Once the input images are decomposed into luminance and color components, two dedicated networks are established, i.e., the Luminance Restoration Network (LRNet) and the Color Regeneration Network (CRNet). LRNet, a lightweight transformer-based restoration model, focuses on recovering illumination and texture. To enable an effective reference for shadow region refinement, we introduce the RecTified Attention (RTA) Module, which employs a rectified mechanism on window-based attentions to support high-quality luminance restoration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Afterward, the processed luminance serves as a guide for querying accurate color information in shadow regions from the non-shadow areas. We inject color information via an ImageNet-pre-trained color encoder, which integrates color data into the intermediate features of the skip connections of our CRNet. This is facilitated by cross-attention modules that align the color in the shadow region with patches" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.151 + ], + "angle": 0, + "content": "of similar luminance. Through this sequential processing, ShadowHack achieves effective shadow removal, supported by extensive experimental results. The primary contributions of this study are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.228 + ], + "angle": 0, + "content": "- Based on an in-depth analysis of shadow properties and the limitations in current end-to-end decomposition methods, we introduce ShadowHack, a novel shadow removal framework that breaks shadow removal into two distinct subtasks: luminance recovery and color regeneration." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.907, + 0.303 + ], + "angle": 0, + "content": "- We identified the distraction phenomenon in the attention matrix of shadow removal networks, and propose RecTified Attention (RTA) to regain focus in attention layers with marginal computational overhead, enabling precise restoration of irregular texture and illumination." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.909, + 0.379 + ], + "angle": 0, + "content": "- We design two specialized networks, LRNet and CRNet, tailored for luminance refinement and color regeneration. Together, these networks allow our model to achieve accurate and robust results, setting new state-of-the-art performance across benchmarking datasets." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.153, + 0.909, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.393, + 0.656, + 0.408 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Shadow Removal. Traditional methods, such as [57] and [45], leveraged principles of physics and mathematical statistics on factors like illumination [10, 60], color [25, 51], region [16], and gradient [11, 12]. However, these methods often struggle with complex real-world scenes, where shadows interact in nuanced ways with surrounding colors and textures. Deep learning has recently made significant strides in the field [27, 43, 48]. For instance, Qu et al. [43] proposed to extract both global and semantic features from images, marking an early attempt to integrate deep learning into shadow removal. Later, Wang et al. [48] proposed a multitask framework that performs shadow detection and removal concurrently, enhancing overall accuracy. Generative Adversarial Networks (GANs) have also been applied to this task. As a representative, RIS-GAN [61] incorporates residual and illumination features to refine shadow regions, while other works [22, 36] use unsupervised GANs to generate shadows for achieving cycle consistency. Li et al. [30] adopted the concept of inpainting, simultaneously using a shadow image encoder and a generative-pretrained image encoder to extract features. Zhu et al. [65] employed normalizing flow for a reversible link between shadow generation and removal. Additionally, Le et al. [28] proposed a linear model that adjusts the global lightness of shadow regions through physical modeling. More recently, diffusion models [46] have shown even greater generative capabilities than GANs. Guo et al. [14] pioneered the use of diffusion models for shadow removal, refining shadow masks throughout the process. Liu et al. [34] applied Retinex decomposition [18, 26] and used diffusion models to refine textures under the guidance of the obtained illumination. However, a primary drawback of these diffusion models is their intensive computational cost." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.514 + ], + "angle": 0, + "content": "ViT-based Restoration. Vision Transformers (ViTs) have demonstrated impressive performance across vision tasks, owing to their powerful attention mechanisms that can effectively capture long-range dependencies [3, 9, 19, 35, 52]. In image restoration, IPT [4] was among the first ViT-based solutions. Unfortunately, its dependence on large-scale paired datasets for training restricts its applicability when data is scarce. Subsequent models have sought to mitigate this limitation with more efficient architectures. Uformer [50] combined UNet with a window attention, yielding strong results, while Restormer [59] introduced transposed attention for efficient channel-wise processing. Further, X-Restormer [5] extended Restormer by integrating the OCA [6] module, providing a generalized framework for image restoration through rigorous analysis. Specific to the domain of shadow removal, ShadowFormer [13] introduced an attention mechanism that facilitates interaction between shadow and non-shadow regions, while HomoFormer [53] improved the transformer by integrating a shuffling mechanism into its transformer-based encoder. RASM [32] refreshed the window-based attention with a sliding-window mechanism, achieving more accurate results. More recently, OmniSR [55] leveraged DINO features and depth maps to incorporate semantic and geometric attention into self-attention, expanding its perception of image content and ability to distinguish shadows. Even though, as discussed in Sec. 1, there remains significant room for the community to improve the effectiveness of shadow removal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.526, + 0.265, + 0.543 + ], + "angle": 0, + "content": "3. Problem Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Degradation in umbra primarily encompasses reduced brightness, color distortion, and texture degradation (e.g., noise and blurriness), as illustrated in Fig. 2. To better understand the types of involved degradation, we begin by formulating shadow effects via a modified image formation model as proposed in [1, 28, 45]. In this model, the observed pixel intensity \\( I(x,\\lambda) \\) at a given spatial coordinate \\( x \\) and wavelength \\( \\lambda \\) is expressed as the product of shading \\( S(x,\\lambda) \\) and albedo \\( A(x,\\lambda) \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.695, + 0.484, + 0.712 + ], + "angle": 0, + "content": "\\[\nI (x, \\lambda) = S (x, \\lambda) \\cdot A (x, \\lambda). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.485, + 0.779 + ], + "angle": 0, + "content": "In scenes with a single dominant light source (e.g., sunlight in outdoor settings), the shading for a pixel \\( x \\) can be decomposed into direct \\( S_{d}(x, \\lambda) \\) and ambient \\( S_{a}(x, \\lambda) \\) components:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.786, + 0.484, + 0.803 + ], + "angle": 0, + "content": "\\[\nI (x, \\lambda) = \\left(S _ {d} (x, \\lambda) + S _ {a} (x, \\lambda)\\right) \\cdot A (x, \\lambda). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.84 + ], + "angle": 0, + "content": "However, when an object occludes the primary light source, shadow regions occur, and the pixel intensity becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.847, + 0.484, + 0.865 + ], + "angle": 0, + "content": "\\[\nI (x, \\lambda) = a (x) \\cdot S _ {a} (x, \\lambda) \\cdot A (x, \\lambda), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(a(x) \\in [0,1]\\) designates a spatially varying attenuation function. This non-uniform attenuation of light forms" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.707, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.173, + 0.672, + 0.185 + ], + "angle": 0, + "content": "(a)ISTD+training set" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.091, + 0.905, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.173, + 0.866, + 0.185 + ], + "angle": 0, + "content": "(b) \\(\\mathrm{ISTD + }\\) testing set" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.196, + 0.708, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.278, + 0.666, + 0.29 + ], + "angle": 0, + "content": "(c) SRD training set" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.196, + 0.905, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.278, + 0.86, + 0.29 + ], + "angle": 0, + "content": "(d) SRD testing set" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 3. Color bias in shadow regions between the reflectances of shadow and shadow-free images by Retinex decomposition [18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.427 + ], + "angle": 0, + "content": "shadow. This involves not only illumination variance but also challenging inconsistency in texture[31], an example of which is illustrated in Fig. 2 (b) and (c). To obtain an RGB image\\(^{1}\\) \\( I_{RGB} \\coloneqq \\{I_R, I_G, I_B\\} \\), the observed color intensity at each pixel is integrated across wavelengths as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.438, + 0.907, + 0.486 + ], + "angle": 0, + "content": "\\[\nI _ {k} (x) = \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} S (x, \\lambda) \\cdot A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda , k \\in \\{R, G, B \\} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.637 + ], + "angle": 0, + "content": "where \\( P_{k}(\\lambda) \\) is the spectral sensitivity function of the camera for the \\( k \\)-channel and \\( [\\lambda_{\\mathrm{min}}, \\lambda_{\\mathrm{max}}] \\) defines the visible spectrum range. Previous shadow removal and low-light enhancement methods [15, 17, 18, 34] suggest that it would be neat to decompose lightness from the shadow image following the Retinex theory. In these works, the image is assumed to be decomposed into a one-channel illumination map \\( L(x) \\) and an illumination-invariant counterpart (reflectance) \\( R_{k}(x) \\). For sRGB space and white-balanced images, the operation can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.648, + 0.907, + 0.696 + ], + "angle": 0, + "content": "\\[\nI _ {k} (x) \\approx L (x) \\cdot R _ {k} (x) = L (x) \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda . \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.879 + ], + "angle": 0, + "content": "We argue that shadow posits not just illumination map changes, but shifts that variants across the color. To quantify this, we plot the per-channel color difference between the input and ground-truth images in the reflectance map derived from a Retinex-based illumination estimator (Fig.3). The reflectance decomposition exhibits substantial color bias, particularly in the red and blue channels, violating Retinex's assumption of illumination-invariant reflectance. This color shift cannot be adequately addressed by previous Retinex-inspired decomposition methods [15, 34], as their core premise which reflectance remains uncontaminated by illumination effects are broken down in shadow regions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.794, + 0.901 + ], + "angle": 0, + "content": "Here, we omit the impact from the ISP pipeline." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.425, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.378, + 0.345, + 0.393 + ], + "angle": 0, + "content": "(a) Luminance Restoration Network" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.098, + 0.903, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.378, + 0.648, + 0.393 + ], + "angle": 0, + "content": "(b) Color Regeneration Network" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.407, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of ShadowHack comprises two core components: (a) a luminance restoration network, which focuses on recovering illumination and texture, and (b) a color regeneration network dedicated to recalibrating harmonized colors for natural outputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.486, + 0.631 + ], + "angle": 0, + "content": "We propose a novel approach to address the triple degradation of reduced brightness, texture deterioration, and color distortion in a divide-and-conquer manner. The key insight is identifying a decomposition scheme that cleanly separates brightness and texture information from color information. An ideal decomposition would isolate a grayscale component capturing illumination and textural details with minimum color interference, alongside a pure color component that can be independently corrected. This separation would enable a targeted restoration process, which first recovers accurate textural details and then uses this restored foundation as a reliable reference for precise color regeneration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.901 + ], + "angle": 0, + "content": "YCbCr color space provides an optimal lossless decomposition for this task. The Y channel, which predominantly weights the green channel (0.587G compared to 0.299R and 0.114B), is significantly less contaminated by the color bias that predominantly affects red and blue channels in shadow regions. Additionally, the Y channel is specifically designed to preserve perceptually important texture details that align with human visual sensitivity. By first restoring this luminance component, we can recover accurate illumination and texture information without interference from color distortions. The restored luminance then serves as a geometric and semantic guide for subsequent color regeneration, ensuring spatial coherence between brightness and chrominance adjustments. This sequential enables dedicated networks to specialize in luminance refinement (LRNet) and chroma harmonization (CRNet) without mutual interference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.448, + 0.651, + 0.466 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.563 + ], + "angle": 0, + "content": "A viable solution involves separating an image into texture \\( I_{t} \\) and color \\( I_{c} \\) components for texture restoration and color regeneration, respectively. We achieve this through an invertible decoupling function \\( D(\\cdot) \\), which splits an RGB image \\( I \\) into a luminance component \\( I_{y} \\) and a chrominance component \\( I_{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.564, + 0.907, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\left(I _ {t}, I _ {c}\\right) = \\mathcal {D} (I). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.708 + ], + "angle": 0, + "content": "We adopt the conversion from RGB to YCbCr as \\(\\mathcal{D}(I)\\), where \\(I_{t}\\) corresponds to the Y (luminance) channel, and \\(I_{c}\\) comprises the Cb and Cr (chrominance) channels. This decomposition isolates brightness and texture degradation in \\(I_{t}\\) while retaining color distortions in \\(I_{c}\\). The inverse function \\(\\mathcal{D}^{-1}\\) reconstructs the final RGB image from processed components. This manner enables a divide-and-conquer strategy for shadow removal:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.72, + 0.907, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\hat {I} = \\mathcal {D} ^ {- 1} \\mathcal {C} (\\mathcal {R} (I _ {t}), I _ {c}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}(\\cdot)\\) denotes the luminance restoration network (LRNet) and, \\(\\mathcal{C}(\\cdot)\\) represents the color regeneration network (CRNet). \\(\\hat{I}\\) is the shadow-free output, and \\(I_{c}\\) serves as a conditioning variable for color regeneration. Additionally, \\(\\mathcal{D}^{-1}(\\cdot)\\) is the inverse procedure of \\(\\mathcal{D}(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As in Fig. 4, ShadowHack operates in three stages: 1) Decomposition. The RGB image is decomposed via \\(\\mathcal{D}(\\cdot)\\) into luminance \\(I_{t}\\) and color \\(I_{c}\\) components; 2) Luminance Restoration. LRNet processes \\(I_{t}\\) to recover illumination and texture, yield in \\(\\hat{I}_t = \\mathcal{R}(I_t)\\); 3) Color Regeneration." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.087, + 0.487, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.487, + 0.355 + ], + "angle": 0, + "content": "Figure 5. The rectified attention module. We compute two sets of attention. In \\( Att_1 \\), both texture \\( (I_t) \\) and color \\( (I_c) \\) are involved while in \\( Att_2 \\), we calculate attention with only color components. The difference \\( Att_1 - \\lambda Att_2 \\) delivers more accurate interactions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.486, + 0.424 + ], + "angle": 0, + "content": "CRNet harmonizes colors using \\( I_{c} \\) and \\( \\hat{I}_{t} \\). Finally, the estimated luminance and color components are recombined using \\( \\mathcal{D}^{-1}(\\cdot) \\) to produce the final shadow-free RGB result. Below, we detail LRNet and CRNet." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.438, + 0.382, + 0.453 + ], + "angle": 0, + "content": "4.1. Luminance Restoration Network" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.486, + 0.734 + ], + "angle": 0, + "content": "To eliminate shadows in \\( I_{t} \\), enhancing the brightness of the shadow areas and restoring their textures is essential. As illustrated in Fig. 4 (a), the shallow two stages and the refinement module in the luminance restoration network (LRNet) are constructed using a Local Range block (LRB), a Multi-head Transpose Attention (MTA) module, and a Feed-Forward Network (FFN), while a RecTified Attention (RTA) module replaces LRB in the deeper stages. The LRB is designed to extract local texture information through a depth-wise convolution block, capturing fine-grained details. The MTA, which lies in the channel attention block (CAB), following previous works [5, 59], reallocates the weights of channels to enhance the relevant features. The FFN further processes these extracted features to improve texture restoration. Finally, the RTA module, introduced in the deeper layers, captures more global information by focusing on regions close to shadow boundaries, enabling the network to refine luminance restoration in complex shadow regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Rectified Attention Module. In the shadow removal process, achieving robust enhancement in shadow regions requires restoration efforts to reference better surrounding, related non-shadow areas. Therefore, we replace the Local Range Block (LRB) with self-attention modules [20, 32] in the deeper layers to capture more distant information. Since the goal of LRNet is to cure shadow-contaminated texture and lightness, the feature from the deep layer (i.e., blocks with SAB equipped) becomes less and less discriminative to shadow. As a result, the attention matrix in deep layers, suffers from a distraction. In the first column of Fig. 6, the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.263, + 0.909, + 0.348 + ], + "angle": 0, + "content": "Figure 6. An example of our rectified attention module. The star indicates the position of the query token. We exhibit the attention map from the first head of the encoder, latent, and decoder layers. The final attention is obtained by subtracting \\(\\mathrm{Att}_2\\) from \\(\\mathrm{Att}_1\\). The attention matrix without rectify, \\(\\mathrm{Att}_1\\) and \\(\\mathrm{Att}_2\\) are normalized to [0, 1], while their differential result is in [-1, 1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.909, + 0.522 + ], + "angle": 0, + "content": "model tends to focus on only a limited number of tokens which may be inaccurate for shadow area refinement. We address this via a RecTified Attention (RTA) mechanism to rectify the original attention weight that is re-injecting contaminated chrominance channels \\( I_{c} \\), by combining them into \\( Q \\) and \\( K \\) calculations (Please notice that the \\( V \\) is still clear without chromance), resulting in \\( Att_{1} \\). Since this operation re-integrates color bias to the calculation, inspired by [58], we calculate a second set with only \\( I_{c} \\) as \\( Att_{2} \\). By subtracting the two sets, LRNet has the flexibility to get rid of chrominance information if needed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.584 + ], + "angle": 0, + "content": "To be specific, we obtain the features \\( F_{t} \\) corresponding to \\( I_{t} \\) through convolution with downsampling, and compute the softmax attention map for \\( F_{t} \\) and \\( F_{c} \\), as shown in Fig. 5. The query, key, and value matrices are formed via:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.591, + 0.906, + 0.618 + ], + "angle": 0, + "content": "\\[\nQ _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {Q}, K _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {K}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.614, + 0.855, + 0.632 + ], + "angle": 0, + "content": "\\[\nQ _ {2} = F _ {c} W _ {2} ^ {Q}, K _ {2} = F _ {c} W _ {2} ^ {K}, V = F _ {t} W ^ {V},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.908, + 0.7 + ], + "angle": 0, + "content": "where \\( W \\) matrices are learnable projection parameters and \"; refers to the concatenation of the two features. Next, the above two sets of queries and keys are used to compute their respective attentions simply by:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.707, + 0.907, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t} _ {i} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {i} K _ {i} ^ {\\mathrm {T}}}{\\sqrt {d}} + B\\right), i \\in \\{1, 2 \\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.908, + 0.779 + ], + "angle": 0, + "content": "where \\( B \\) denotes the relative position bias. This allows us to apply the rectified attention mechanism as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.786, + 0.907, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R T A} \\left(F _ {t}, F _ {c}\\right) = \\left(\\operatorname {A t t} _ {1} - \\lambda \\cdot \\operatorname {A t t} _ {2}\\right) V, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.9 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a reweighting factor. As illustrated in Fig. 6, such a rectified attention mechanism creates a strong affinity with well-lit reference, urging the model to focus on relevant content. As validated in Sec. 5.3, our rectified attention can focus on more non-shadow area information to transfer robust clues for shadow removal." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11407" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.425 + ], + "angle": 0, + "content": "
DatasetsMethodShadow Region (S)Non-Shadow Region (NS)All Image (ALL)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
ISTD+DHAN [8] (AAAI'20)33.080.9889.4927.280.9727.3925.780.9587.74
G2R [36] (CVPR'21)33.880.9788.7135.940.9772.8130.850.9463.78
BMNet [65] (CVPR'22)38.170.9915.7237.950.9862.4234.340.9742.93
SG-ShadowNet [47] (ECCV'22)36.800.9905.9335.570.9782.9232.460.9623.41
ShadowFormer [13] (AAAI'23)39.670.9925.2138.820.9832.3035.460.9732.80
ShadowDiffusion [14] (CVPR'23)39.82-4.9038.90-2.3035.72-2.70
Li et al. [30] (ICCV'23)38.460.9895.9337.270.9772.9034.140.9603.39
Liu et al. [34] (AAAI'24)38.040.9905.6939.150.9842.3134.960.9682.87
RASM [32] (MM'24)40.730.9934.4139.230.9852.1736.160.9762.53
Homofformer [53] (CVPR'24)39.470.9934.7238.730.9842.2335.340.9752.64
OmniSR+GM [55] (AAAI'25)------34.200.973-
ShadowHack (Ours)40.560.9934.4639.660.9872.0936.310.9772.48
SRDDHAN [8] (AAAI'20)33.670.9788.9434.790.9794.8030.510.9495.67
BMNet [65] (CVPR'22)35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNet [47] (ECCV'22)--7.53--2.97--4.23
ShadowFormer [13] (AAAI'23)36.910.9895.9036.220.9893.4432.900.9584.04
ShadowDiffusion [14] (CVPR'23)38.720.9874.9837.780.9853.4434.730.9703.63
Li et al. [30] (ICCV'23)39.330.9856.0935.610.9672.9733.170.9403.83
Liu et al. [34] (AAAI'24)36.510.9835.4937.710.9863.0033.480.9673.66
DeS3 [24] (AAAI'24)37.450.9845.8838.120.9882.8334.110.9683.72
RASM [32] (MM'24)37.910.9885.0238.700.9922.7234.460.9763.37
Homofformer [53] (CVPR'24)38.810.9874.2539.450.9882.8535.370.9723.33
OmniSR+GM [55] (AAAI'25)------34.560.977-
ShadowHack (Ours)39.470.9914.1840.060.9942.3835.940.9822.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.43, + 0.892, + 0.444 + ], + "angle": 0, + "content": "Table 1. Quantitative results on the ISTD+ and SRD datasets. The best results are in bold, while the second-best ones are underlined." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.35, + 0.48 + ], + "angle": 0, + "content": "4.2. Color Regeneration Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.486, + 0.837 + ], + "angle": 0, + "content": "After restoring texture and lighting information in the luminance channel, the next task is to accurately regenerate color. Inspired by exemplar-based colorization models [23, 29, 39], we leverage contextual cues and condition-based guidance to improve color fidelity and ensure alignment with the semantic content of the images. Concretely, we propose a dual-encoder network, termed CRNet, which integrates a multiscale color feature extractor into the U-shaped architecture of LRNet. The color encoder is a ConvNext-v2utto model [37] pre-trained on ImageNet-21k with 2M parameters, which takes the Cb and Cr channels as input and extracts color features across multiple scales, as illustrated in Fig. 4 (b). To achieve precise color restoration, the network shall reference regions with similar/identical material properties and accurate color, using them as reliable guides for color regeneration. The outputs generated by LRNet provide qualified indexing information to facilitate this purpose. Therefore, we introduce an attention-based mechanism where luminance features serve as the query and key to compute similarity, while color features act as the value, aggregating essential color information. This approach injects color features into the skip connections of the U-shaped backbone, remarkably enriching the network's color regeneration capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Multiple Checkpoint Training. We observe that the luminance network often yields better performance on the training set, which suggests the color regeneration network faces challenges in adapting to imperfect luminance out" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.909, + 0.571 + ], + "angle": 0, + "content": "puts encountered during testing. To address this issue, we employ a simple yet effective way to overcome this issue, where multiple early-stage checkpoints from the luminance network training are randomly selected as candidates. This technique, applied only during training, strengthens the robustness of the color network without incurring additional computational costs at inference time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.744, + 0.6 + ], + "angle": 0, + "content": "5. Experimental Validation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.707, + 0.624 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Implementation details. Our models are implemented in PyTorch, and the entire training process utilizes the AdamW optimizer [38] with momentum set to (0.9, 0.999) and a weight decay of \\(10^{-2}\\). The initial learning rate is set to \\(2 \\times 10^{-4}\\). During training, we randomly crop the training data into patches of size \\(384 \\times 384\\). For the base network of each stage, we use a four-layer encoder-decoder structure (from L1 to L4), consisting of three downsampling and three upsampling operations. The dimension of the first layer is 32. Our RTA modules are integrated into L3 and L4, with an overlapping ratio set to 0.5 and a rectified mechanism \\(\\lambda_0\\) value of 0.7. For training, we use L1 and VGG loss terms as employed in previous works [32, 34]. For detailed information on the implementation of each stage and training setup, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments on the following benchmark datasets for shadow removal: 1) Adjusted ISTD (ISTD+) dataset [27], which refines the ISTD dataset [48]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.199, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.193, + 0.172, + 0.205 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.089, + 0.301, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.193, + 0.29, + 0.205 + ], + "angle": 0, + "content": "(b) BMNet [65]" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.4, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.193, + 0.378, + 0.205 + ], + "angle": 0, + "content": "(c) SG [47]" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.089, + 0.499, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.193, + 0.477, + 0.205 + ], + "angle": 0, + "content": "(d) SF [13]" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.598, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.193, + 0.577, + 0.205 + ], + "angle": 0, + "content": "(e) HF [53]" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.089, + 0.697, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.193, + 0.686, + 0.205 + ], + "angle": 0, + "content": "(f) RASM [32]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.798, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.193, + 0.769, + 0.205 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.089, + 0.897, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.193, + 0.865, + 0.205 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.211, + 0.694, + 0.226 + ], + "angle": 0, + "content": "Figure 7. Visual comparison on a sample from the \\(\\mathrm{ISTD + }\\) dataset." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.233, + 0.2, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.335, + 0.172, + 0.346 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.233, + 0.3, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.334, + 0.29, + 0.346 + ], + "angle": 0, + "content": "(b) BMNet [65]" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.233, + 0.399, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.334, + 0.378, + 0.346 + ], + "angle": 0, + "content": "(c) SG [47]" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.233, + 0.498, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.334, + 0.489, + 0.346 + ], + "angle": 0, + "content": "(d) DMTN [33]" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.233, + 0.598, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.334, + 0.583, + 0.346 + ], + "angle": 0, + "content": "(e) DeS3 [24]" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.233, + 0.698, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.334, + 0.676, + 0.346 + ], + "angle": 0, + "content": "(f) HF [53]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.233, + 0.797, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.334, + 0.768, + 0.346 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.233, + 0.898, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.827, + 0.334, + 0.864, + 0.346 + ], + "angle": 0, + "content": "(h) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.352, + 0.687, + 0.367 + ], + "angle": 0, + "content": "Figure 8. Visual comparison on a sample from the SRD dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.384, + 0.483, + 0.438 + ], + "angle": 0, + "content": "
MetricRGB InputYCbCr InputRetinex-basedOurs
PSNR↑36.1635.6235.8536.31
RMSE↓2.532.622.632.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Table 2. Ablation study of the proposed decoupling framework on the ISTD+ dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.638 + ], + "angle": 0, + "content": "by correcting the illumination discrepancies between the ground-truth and input images, includes 1330 training and 540 testing triplets (shadow image, shadow-free image, and shadow mask); 2) SRD dataset [43] consists of 2680 training and 408 testing pairs. We adopt the predicted masks provided by DHAN [8] for training and testing following previous works [32, 53, 65]; 3) UIUC dataset [16] provides 76 images with shadow mask for testing; and 4) UCF dataset [64] provided by Jin et al. [24] contains shadow images without shadow-free references." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.791 + ], + "angle": 0, + "content": "Evaluation metrics. To evaluate the performance of our proposed framework, we utilize the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE). PSNR and SSIM are both calculated in the RGB space, with higher values indicating better results. For RMSE, which is calculated in the LAB color space, the lower values indicate better results. To be consistent with previous works [32, 53, 65], we inference the full-resolution image and then evaluate the images at a resolution of \\(256 \\times 256\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.345, + 0.818 + ], + "angle": 0, + "content": "5.2. Comparison with Prior Arts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our method with the state-of-the-art shadow removal approaches on benchmark datasets, including DHAN [8], G2R [36], BMNet [65], SG-ShadowNet [47], ShadowFormer [13], ShadowDiffusion [14], Li et al. [30], Liu et al. [34], DeS3 [24], Homoformer [53], RASM [32]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.384, + 0.613, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.462, + 0.589, + 0.474 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.384, + 0.71, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.462, + 0.701, + 0.474 + ], + "angle": 0, + "content": "(b) RGB Input" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.384, + 0.807, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.462, + 0.797, + 0.474 + ], + "angle": 0, + "content": "(c) w/o rectify" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.384, + 0.904, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.462, + 0.879, + 0.474 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.48, + 0.906, + 0.508 + ], + "angle": 0, + "content": "Figure 9. Ablation study on rectified attention. The results are displayed in the luminance space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.61 + ], + "angle": 0, + "content": "and OmniSR+GM [55] (GM means the version with shadow mask), as shown in Tab. 1. For a fair comparison, we use results published by the authors. Regarding missing data, we utilize available images or checkpoints provided by the authors and evaluate them via the same metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Our proposed method achieves the highest scores in most metrics on the ISTD+ dataset compared to other state-of-the-art methods. Specifically, our model attains a PSNR of 36.31 for all image regions, surpassing the second-highest result of 36.16 by 0.15. The high PSNR highlights our model's ability to recover texture details effectively, demonstrating superior reconstruction quality. Additionally, the RMSE value of 2.48 is the lowest, which indicates that our recoloring module plays a crucial role in accurately restoring color information and reducing errors across the entire image. From the magnified areas in Figs. 7 and 8, it can be observed that our method not only effectively corrects color bias but also provides superior edge correction compared to other methods. The second part of Tab. 1 presents a comparison of our results on the SRD dataset against other state-of-the-art methods. It can be observed that our method achieves a significant improvement in RMSE (-0.43), further demonstrating the substantial benefit of the decoupling approach for color recovery." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11409" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.483, + 0.212 + ], + "angle": 0, + "content": "
MethodShadow PSNR ↑Non-Shadow PSNR ↑All PSNR ↑
[Ft]39.9641.2736.60
[Ft; Fc]40.1741.2936.76
[Ft]&[Ft]40.0241.1836.58
[Ft; Fc]&[Ft; Fc]40.3541.3736.87
[Ft; Fc]&[Fc]40.3641.4036.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.217, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Table 3. Ablation study of the proposed rectified attention module on the SRD dataset. We calculate PSNRs on the luminance." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.254, + 0.171, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.397, + 0.154, + 0.41 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.255, + 0.249, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.397, + 0.24, + 0.41 + ], + "angle": 0, + "content": "(b) SG [47]" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.255, + 0.326, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.397, + 0.316, + 0.41 + ], + "angle": 0, + "content": "(c) SF [13]" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.255, + 0.403, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.397, + 0.394, + 0.41 + ], + "angle": 0, + "content": "(d) HF [53]" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.255, + 0.481, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.397, + 0.463, + 0.41 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.415, + 0.473, + 0.43 + ], + "angle": 0, + "content": "Figure 10. Visual comparisons on the UIUC and UCF datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.244, + 0.455 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.485, + 0.629 + ], + "angle": 0, + "content": "Effect of the decoupling framework. To validate the effectiveness of our decoupling framework, we fine-tune our baseline model structure to adapt to RGB and YCbCr input/output for end-to-end shadow removal and compare its performance with our decoupling framework. As shown in Tab. 2, our decoupling approach significantly improves both PSNR and RMSE. Moreover, based on Retinex decomposition, we separately recover L and R, and compare it with our decoupling method. Detailed architectures are provided in the supplementary material. This improvement indicates that this strategy benefits texture recovery and color correction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Effect of rectified attention module. We demonstrate the effectiveness of the rectified attention module in Tab. 3, by conducting experiments on rectification mechanisms. Furthermore, extra experiments are conducted on different feature combinations of rectified approaches to explore more suitable configurations. \\([X]\\) & \\([Y]\\) in Tab. 3 means performing a subtraction calculation of corresponding maps for rectification between \\(X\\) and \\(Y\\). \\([X]\\) alone denotes only one pair of \\(Q\\) and \\(K\\) is used, where both inputs are \\(X\\). Clearly, leveraging color information, even with inherent bias, can allow shadow regions to better reference surrounding areas, resulting in significant performance improvements. As for potential color bias, rectification refines the attention map for greater precision. Furthermore, we find that \\([F_t; F_c] \\& [F_c]\\) slightly outperforms \\([F_t; F_c] \\& [F_t; F_c]\\), and thus we ultimately choose \\([F_t; F_c] \\& [F_c]\\). Fig. 9 visualizes the effectiveness of rectification, which has a more precise result with less edge remaining." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.177 + ], + "angle": 0, + "content": "
MethodShadow RMSE ↓Non-Shadow RMSE ↓All RMSE ↓
w/o cross-att.5.782.643.39
w/o multi checkpt.5.012.593.28
Ours4.182.382.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.179, + 0.882, + 0.193 + ], + "angle": 0, + "content": "Table 4. Ablation study of the CRNet on the SRD dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.199, + 0.614, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.199, + 0.71, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.199, + 0.806, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.199, + 0.903, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.252, + 0.604, + 0.265 + ], + "angle": 0, + "content": "(a) dirty mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.252, + 0.806, + 0.265 + ], + "angle": 0, + "content": "(b) refined mask (c) w/o refinement" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.252, + 0.901, + 0.265 + ], + "angle": 0, + "content": "(d) w refinement" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.906, + 0.309 + ], + "angle": 0, + "content": "Figure 11. Robustness analysis under corrupted mask. The accurate mask is highlighted in red. For fairness, we only involve mask refinement network here, and it is not involved in any comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Effect of the color restoration network. To validate the effectiveness of the key component and design in the CRNet, we replace the cross-attention-based color feature injection with concatenation along the feature channel. As demonstrated in Tab. 4, if cross-attention is replaced, the accuracy has dropped by a large margin, indicating the effectiveness of the cross-attention. Furthermore, the model suffers from a misalignment between training and testing luminance results if no previous checkpoints are involved." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.706, + 0.48 + ], + "angle": 0, + "content": "5.4. Robustness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Generalization to other datasets. To demonstrate the generalization capability of ShadowHack, we evaluate it on the UIUC [16] and UCF [64] datasets with our \\(\\mathrm{ISTD + }\\) pretrained model. The results are presented in Fig. 10, where ShadowHack achieves the cleanest shadow removal with minimal color bias. More comparisons are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Robustness to inaccurate masks. To get rid of the performance degradation brought by inaccurate masks, we propose a mask refine network, which corrects the mask for better generalization. After our remediation, the result is recovered from inaccurate enlightening, as in Fig. 11 shows. More details are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This work proposed a novel framework, ShadowHack, which decouples images into luminance and color components, enabling targeted processing for more effective shadow removal. This approach can well address limitations in existing end-to-end models by enhancing the use of color information to improve both texture and luminance recovery. To further boost shadow removal, a rectified attention mechanism has been designed to boost the reference ability of self-attention. Additionally, the color regeneration network has demonstrated its critical role in accurately correcting image colors. Extensive experimental results have underscored the superior performance and robustness of our model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11410" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.156 + ], + "angle": 0, + "content": "[1] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2(3-26):2, 1978. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.485, + 0.213 + ], + "angle": 0, + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In ICCV, pages 12470-12479, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[5] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. arXiv preprint arXiv:2310.11881, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In CVPR, pages 22367-22377, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[7] Michael E Churma. Blue shadows: physical, physiological, and psychological causes. Applied optics, 33(21):4719-4722, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.485, + 0.521 + ], + "angle": 0, + "content": "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting GAN. In AAAI, pages 10680-10687, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.485, + 0.647 + ], + "angle": 0, + "content": "[10] Graham D. Finlayson, Steven D. Hordley, and Mark S. Drew. Removing shadows from images. In ECCV (4), pages 823-836, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[11] Graham D. Finlayson, Steven D. Hordley, Cheng Lu, and Mark S. Drew. On the removal of shadows from images. IEEE Trans. Pattern Anal. Mach. Intell., 28(1):59-68, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[12] Graham D. Finlayson, Mark S. Drew, and Cheng Lu. Entropy minimization for shadow removal. Int. J. Comput. Vis., 85(1): 35-57, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[13] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 1, 3, 6, 7, 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[14] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In ICCV, pages 12999-13008. IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[16] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2956-2967, 2013. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[17] Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness. Int. J. Comput. Vis., 131(1): 48-66, 2023. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[18] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: low-light image enhancement via illumination map estimation. IEEE Trans. Image Process., 26(2):982-993, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[21] Jun-Wei Hsieh, Wen-Fong Hu, Chia-Jung Chang, and Yung-Sheng Chen. Shadow elimination for effective moving object detection by gaussian shadow modeling. Image Vis. Comput., 21(6):505-516, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[22] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[23] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Trans. Graph., 41(6):205:1-205:16, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.502 + ], + "angle": 0, + "content": "[24] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T. Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 1, 6, 7, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[25] Salman Hameed Khan, Mohammed Bennamoun, Ferdous Ahmed Sohel, and Roberto Togneri. Automatic shadow detection and removal from a single image. IEEE Trans. Pattern Anal. Mach. Intell., 38(3):431-446, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.588 + ], + "angle": 0, + "content": "[26] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[27] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8577-8586, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[28] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9088-9101, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[29] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[30] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor W. Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13009-13018, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[31] Feng Liu and Michael Gleicher. Texture-consistent shadow removal. In ECCV (4), pages 437-450. Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[32] Hengxing Liu, Mingjia Li, and Xiaojie Guo. Regional attention for shadow removal. In ACM Multimedia, 2024. 1, 3, 5, 6, 7, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[33] Jiawei Liu, Qiang Wang, Huijie Fan, Wentao Li, Liangqiong Qu, and Yandong Tang. A decoupled multi-task network for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11411" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "shadow removal. IEEE Trans. Multim., 25:9449-9463, 2023. 7, 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[34] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson W. H. Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.237, + 0.482, + 0.279 + ], + "angle": 0, + "content": "[36] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11966-11976, 2022. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[39] Peng Lu, Jinbei Yu, Xujun Peng, Zhaoran Zhao, and Xiaojie Wang. Gray2colornet: Transfer more colors from reference image. In ACM Multimedia, pages 3210-3218, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[40] David K. Lynch. Shadows. Appl. Opt., 54(4):B154-B164, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[41] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In WACV, pages 4313-4322, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[42] Sohail Nadimi and Bir Bhanu. Physical models for moving shadow and object detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 26(8):1079-1087, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.529, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[43] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson W. H. Lau. Desshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 2308-2316, 2017. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[45] Yael Shor and Dani Lischinski. The shadow meets the mask: Pyramid-based shadow removal. Comput. Graph. Forum, 27 (2):577-586, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[47] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV (19), pages 361-378, 2022. 1, 6, 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[48] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[49] Tianyu Wang, Jianming Zhang, Haitian Zheng, Zhihong Ding, Scott Cohen, Zhe Lin, Wei Xiong, Chi-Wing Fu, Luis Figueroa, and Soo Ye Kim. Metashadow: Object-centered shadow detection, removal, and synthesis. In CVPR, pages 28252-28262, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17662-17672, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[51] Tai-Pang Wu and Chi-Keung Tang. A bayesian approach for shadow extraction from a single image. In ICCV, pages 480–487, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[52] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4784-4793, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[53] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 1, 3, 6, 7, 8, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.345 + ], + "angle": 0, + "content": "[54] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. Int. J. Comput. Vis., 132(9):3610-3632, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[55] Jiamin Xu, Zelong Li, Yuxin Zheng, Chenyu Huang, Renshu Gu, Weiwei Xu, and Gang Xu. Omnisr: Shadow removal under direct and indirect lighting. arXiv preprint arXiv:2410.01719, 2024. 3, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[56] Han Yang, Tianyu Wang, Xiaowei Hu, and Chi-Wing Fu. Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12687-12698, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[57] Qingxiong Yang, Kar-Han Tan, and Narendra Ahuja. Shadow removal using bilateral filtering. IEEE Trans. Image Process., 21(10):4361-4368, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[58] Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. arXiv preprint arXiv:2410.05258, 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[60] Ling Zhang, Qing Zhang, and Chunxia Xiao. Shadow remover: Image shadow removal based on illumination recovering optimization. IEEE Trans. Image Process., 24(11): 4623-4636, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[61] Ling Zhang, Chengjiang Long, Xiaolong Zhang, and Chunxia Xiao. RIS-GAN: explore residual and illumination with generative adversarial networks for shadow removal. In AAAI, pages 12829-12836, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[62] Wuming Zhang, Xi Zhao, Jean-Marie Morvan, and Liming Chen. Improving shadow suppression for illumination robust face recognition. IEEE Trans. Pattern Anal. Mach. Intell., 41 (3):611-624, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[63] Wei Zhou, Julie Stephany Berrio, Stewart Worrall, and Eduardo M. Nebot. Automated evaluation of semantic segmentation robustness for autonomous driving. IEEE Trans. Intell. Transp. Syst., 21(5):1951-1963, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[64] Jiejie Zhu, Kegan G. G. Samuel, Syed Zain Masood, and Marshall F. Tappen. Learning to recognize shadows in monochromatic natural images. In CVPR, pages 223-230, 2010. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11412" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.137 + ], + "angle": 0, + "content": "[65] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5617-5626, 2022. 2, 6, 7, 1, 3, 4, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11413" + } + ] +] \ No newline at end of file diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_origin.pdf b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..83c2acf05c26444606b1c355c7e6b20a65527028 --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/d86f4be3-9e03-4b38-b628-22e777dbb589_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a6354f1643e7528b13736960a25eac9378097066cffe96c10b674843ccd475f +size 22422562 diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/full.md b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dc4ca759d1577f5b1e468a7b05374769876af611 --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/full.md @@ -0,0 +1,410 @@ +# ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer + +Jin Hu* Mingjia Li* Xiaojie Guo† Tianjin University + +{jinhu,mingjiali}@tju.edu.cn,xj.max.guo@gmail.com + +# Abstract + +Shadows introduce challenges such as reduced brightness, texture deterioration, and color distortion in images, complicating a holistic solution. This study presents ShadowHack, a divide-and-conquer strategy that tackles these complexities by decomposing the original task into luminance recovery and color remedy. To brighten shadow regions and repair the corrupted textures in the luminance space, we customize LRNet, a U-shaped network with a rectified attention module, to enhance information interaction and recalibrate contaminated attention maps. With luminance recovered, CRNet then leverages cross-attention mechanisms to revive vibrant colors, producing visually compelling results. Extensive experiments on multiple datasets are conducted to demonstrate the superiority of ShadowHack over existing state-of-the-art solutions both quantitatively and qualitatively, highlighting the effectiveness of our design. Our code is available at https://github.com/lime-j/ShadowHack. + +Shadow is a color as light is, but less brilliant; light and shadow are only the relation of two tones. + +Paul Cezanne + +# 1. Introduction + +Shadows are the absence of light, an everyday phenomenon that often shapes the aesthetic beauty of nature. However, shadows obscure the actual shape of instances, diminishing their visibility. Very likely, they deteriorate the performance of various machine vision tasks, such as object detection [21, 42], face recognition [62], and semantic segmentation [63], to name just a few. Consequently, shadow removal has become a critical research topic. + +Recent advancements in deep learning have greatly improved the efficiency and accuracy of shadow removal. Specifically, DeshadowNet [43] marked a pioneering step with its end-to-end Convolutional Neural Network (CNN) + +![](images/6b30d9a990595dcfae1f4171d5175f8f36c2153bb5f291532432ae46c6726f55.jpg) +(a) Input + +![](images/4a9cb6cd8ada88587366c5818631f8fae929cecd51810af087ce46845ea2370e.jpg) +(b) SG [47] + +![](images/3aa9e851605414b79f8a90731ba1bb6904f6a30181ba428593d833e899c18569.jpg) +(c) HF [53] +Figure 1. Visual comparison between our method and state-of-the-art methods in both Luminance and RGB spaces. + +![](images/dd5c55a873d199a5d7fbad2a127472d670dad0b5d9ea58fd1b849158d612fcb6.jpg) +(d) Ours + +approach, while subsequent GAN-based methods [22, 36, 48, 61] further advanced the field by adopting adversarial learning frameworks, enabling effective unpaired training. Additionally, Wan et al. [47] introduced style transfer, and Li et al. [30] incorporated pre-trained inpainting priors to better restore textural information. Transformer-based architectures, such as ShadowFormer [13], HomoFormer [53], and RASM [32], have since been developed to restore shadow images more effectively, while diffusion-based approaches [14, 24, 41, 49] have also demonstrated promising results, particularly in complex scenarios where previous generative models often struggled. + +Despite these advancements, the odyssey of removing shadows is still far from its end. As evidenced in Fig. 1, even state-of-the-art models fail to recover color and underlying texture accurately. The difficulty stems from the complex and intertwined nature of shadow degradation, which involves not only reduced brightness in shadow (umbra) regions [40], but also texture deterioration from the imaging process [30, 34], and color distortion from surface material properties and environmental color influences [7, 28, 47], an example of which is in Fig. 2. This complexity naturally suggests breaking down the original task, with its entangled interferences, into isolated and simpler sub-problems, each targeting a specific aspect and addressed independently. + +Following the above line of thought, some works [16, 27, 28] attempted to estimate light and color biases as scalars. Insights can also be drawn from the low-light enhancement literature, where the Retinex decomposition [1, 26] is commonly employed [2, 17, 18, 54]. Under the Retinex theory, + +![](images/2ec9418c33c512cb4125e6438f08416b5f97c607ef5d5be82c37ae6fa08b6e88.jpg) +(a) Color Shift + +![](images/e0ffca9b6ead9a3ca2a7f9bd9937be4ff358346cb3ba574aa40c01eae3bddab6.jpg) +(b) Texture Distortion + +![](images/290f65f281a42d0fdb019ff16c0687a8a92e6b82359364d5157d761603a9d7f5.jpg) +(c) Illumination Underlit +Figure 2. Example of the complex degradation in a shadow image. a) Color shift in the umbra region; (b) Structural texture deterioration in the penumbra transition area; (c) Attenuated illumination levels across the shadow (Blue indicates lower illumination). + +an image can be decomposed into a structured illumination and a colored texture-rich counterpart. However, these approaches encounter challenges due to the inherent inhomogeneity of shadows. In other words, shadow removal extends beyond low-light enhancement and brings additional color biases to the reflectance map. This realization underscores the need for more sophisticated solutions to effectively disentangle the complex degradations introduced by shadows. + +This work invents ShadowHack, an effective shadow removal framework that individually tackles illumination and color components. ShadowHack starts with a simple yet highly effective hack for shadow decomposition. Unlike conventional methods that decouple illumination from color and texture, we alternatively separate color from illumination and texture. Once the input images are decomposed into luminance and color components, two dedicated networks are established, i.e., the Luminance Restoration Network (LRNet) and the Color Regeneration Network (CRNet). LRNet, a lightweight transformer-based restoration model, focuses on recovering illumination and texture. To enable an effective reference for shadow region refinement, we introduce the RecTified Attention (RTA) Module, which employs a rectified mechanism on window-based attentions to support high-quality luminance restoration. + +Afterward, the processed luminance serves as a guide for querying accurate color information in shadow regions from the non-shadow areas. We inject color information via an ImageNet-pre-trained color encoder, which integrates color data into the intermediate features of the skip connections of our CRNet. This is facilitated by cross-attention modules that align the color in the shadow region with patches + +of similar luminance. Through this sequential processing, ShadowHack achieves effective shadow removal, supported by extensive experimental results. The primary contributions of this study are summarized as follows: + +- Based on an in-depth analysis of shadow properties and the limitations in current end-to-end decomposition methods, we introduce ShadowHack, a novel shadow removal framework that breaks shadow removal into two distinct subtasks: luminance recovery and color regeneration. +- We identified the distraction phenomenon in the attention matrix of shadow removal networks, and propose RecTified Attention (RTA) to regain focus in attention layers with marginal computational overhead, enabling precise restoration of irregular texture and illumination. +- We design two specialized networks, LRNet and CRNet, tailored for luminance refinement and color regeneration. Together, these networks allow our model to achieve accurate and robust results, setting new state-of-the-art performance across benchmarking datasets. + +# 2. Related Work + +Shadow Removal. Traditional methods, such as [57] and [45], leveraged principles of physics and mathematical statistics on factors like illumination [10, 60], color [25, 51], region [16], and gradient [11, 12]. However, these methods often struggle with complex real-world scenes, where shadows interact in nuanced ways with surrounding colors and textures. Deep learning has recently made significant strides in the field [27, 43, 48]. For instance, Qu et al. [43] proposed to extract both global and semantic features from images, marking an early attempt to integrate deep learning into shadow removal. Later, Wang et al. [48] proposed a multitask framework that performs shadow detection and removal concurrently, enhancing overall accuracy. Generative Adversarial Networks (GANs) have also been applied to this task. As a representative, RIS-GAN [61] incorporates residual and illumination features to refine shadow regions, while other works [22, 36] use unsupervised GANs to generate shadows for achieving cycle consistency. Li et al. [30] adopted the concept of inpainting, simultaneously using a shadow image encoder and a generative-pretrained image encoder to extract features. Zhu et al. [65] employed normalizing flow for a reversible link between shadow generation and removal. Additionally, Le et al. [28] proposed a linear model that adjusts the global lightness of shadow regions through physical modeling. More recently, diffusion models [46] have shown even greater generative capabilities than GANs. Guo et al. [14] pioneered the use of diffusion models for shadow removal, refining shadow masks throughout the process. Liu et al. [34] applied Retinex decomposition [18, 26] and used diffusion models to refine textures under the guidance of the obtained illumination. However, a primary drawback of these diffusion models is their intensive computational cost. + +ViT-based Restoration. Vision Transformers (ViTs) have demonstrated impressive performance across vision tasks, owing to their powerful attention mechanisms that can effectively capture long-range dependencies [3, 9, 19, 35, 52]. In image restoration, IPT [4] was among the first ViT-based solutions. Unfortunately, its dependence on large-scale paired datasets for training restricts its applicability when data is scarce. Subsequent models have sought to mitigate this limitation with more efficient architectures. Uformer [50] combined UNet with a window attention, yielding strong results, while Restormer [59] introduced transposed attention for efficient channel-wise processing. Further, X-Restormer [5] extended Restormer by integrating the OCA [6] module, providing a generalized framework for image restoration through rigorous analysis. Specific to the domain of shadow removal, ShadowFormer [13] introduced an attention mechanism that facilitates interaction between shadow and non-shadow regions, while HomoFormer [53] improved the transformer by integrating a shuffling mechanism into its transformer-based encoder. RASM [32] refreshed the window-based attention with a sliding-window mechanism, achieving more accurate results. More recently, OmniSR [55] leveraged DINO features and depth maps to incorporate semantic and geometric attention into self-attention, expanding its perception of image content and ability to distinguish shadows. Even though, as discussed in Sec. 1, there remains significant room for the community to improve the effectiveness of shadow removal. + +# 3. Problem Analysis + +Degradation in umbra primarily encompasses reduced brightness, color distortion, and texture degradation (e.g., noise and blurriness), as illustrated in Fig. 2. To better understand the types of involved degradation, we begin by formulating shadow effects via a modified image formation model as proposed in [1, 28, 45]. In this model, the observed pixel intensity $I(x,\lambda)$ at a given spatial coordinate $x$ and wavelength $\lambda$ is expressed as the product of shading $S(x,\lambda)$ and albedo $A(x,\lambda)$ as follows: + +$$ +I (x, \lambda) = S (x, \lambda) \cdot A (x, \lambda). \tag {1} +$$ + +In scenes with a single dominant light source (e.g., sunlight in outdoor settings), the shading for a pixel $x$ can be decomposed into direct $S_{d}(x, \lambda)$ and ambient $S_{a}(x, \lambda)$ components: + +$$ +I (x, \lambda) = \left(S _ {d} (x, \lambda) + S _ {a} (x, \lambda)\right) \cdot A (x, \lambda). \tag {2} +$$ + +However, when an object occludes the primary light source, shadow regions occur, and the pixel intensity becomes: + +$$ +I (x, \lambda) = a (x) \cdot S _ {a} (x, \lambda) \cdot A (x, \lambda), \tag {3} +$$ + +where $a(x) \in [0,1]$ designates a spatially varying attenuation function. This non-uniform attenuation of light forms + +![](images/bd8ea2d5b45784a606493168347da4db9d4564b56f686c8a0965ee10bef0183c.jpg) +(a)ISTD+training set + +![](images/0b26a5997b48d572781fb6515455d27baa78d670433afdf66f41a402fb59d314.jpg) +(b) $\mathrm{ISTD + }$ testing set + +![](images/0b480c837dc7342b80d92abe57dde87203ab4c23a3342a66382c841b206b364c.jpg) +(c) SRD training set + +![](images/3f67cd43ea73fb250a71d11e122431b896debc62479d76d50600abf26f57000c.jpg) +(d) SRD testing set +Figure 3. Color bias in shadow regions between the reflectances of shadow and shadow-free images by Retinex decomposition [18]. + +shadow. This involves not only illumination variance but also challenging inconsistency in texture[31], an example of which is illustrated in Fig. 2 (b) and (c). To obtain an RGB image $^{1}$ $I_{RGB} \coloneqq \{I_R, I_G, I_B\}$ , the observed color intensity at each pixel is integrated across wavelengths as follows: + +$$ +I _ {k} (x) = \int_ {\lambda_ {\min }} ^ {\lambda_ {\max }} S (x, \lambda) \cdot A (x, \lambda) \cdot P _ {k} (\lambda) d \lambda , k \in \{R, G, B \} \tag {4} +$$ + +where $P_{k}(\lambda)$ is the spectral sensitivity function of the camera for the $k$ -channel and $[\lambda_{\mathrm{min}}, \lambda_{\mathrm{max}}]$ defines the visible spectrum range. Previous shadow removal and low-light enhancement methods [15, 17, 18, 34] suggest that it would be neat to decompose lightness from the shadow image following the Retinex theory. In these works, the image is assumed to be decomposed into a one-channel illumination map $L(x)$ and an illumination-invariant counterpart (reflectance) $R_{k}(x)$ . For sRGB space and white-balanced images, the operation can be formulated as: + +$$ +I _ {k} (x) \approx L (x) \cdot R _ {k} (x) = L (x) \int_ {\lambda_ {\min }} ^ {\lambda_ {\max }} A (x, \lambda) \cdot P _ {k} (\lambda) d \lambda . \tag {5} +$$ + +We argue that shadow posits not just illumination map changes, but shifts that variants across the color. To quantify this, we plot the per-channel color difference between the input and ground-truth images in the reflectance map derived from a Retinex-based illumination estimator (Fig.3). The reflectance decomposition exhibits substantial color bias, particularly in the red and blue channels, violating Retinex's assumption of illumination-invariant reflectance. This color shift cannot be adequately addressed by previous Retinex-inspired decomposition methods [15, 34], as their core premise which reflectance remains uncontaminated by illumination effects are broken down in shadow regions. + +![](images/31724fe29116adeed49ae31780ed0afe19912db570cde5388f8aa3a38f7eeba6.jpg) +(a) Luminance Restoration Network +Figure 4. The overall architecture of ShadowHack comprises two core components: (a) a luminance restoration network, which focuses on recovering illumination and texture, and (b) a color regeneration network dedicated to recalibrating harmonized colors for natural outputs. + +![](images/1ff812b5335c13c5202b750e44775930642917100f75250bcee5ac6b644c9020.jpg) +(b) Color Regeneration Network + +We propose a novel approach to address the triple degradation of reduced brightness, texture deterioration, and color distortion in a divide-and-conquer manner. The key insight is identifying a decomposition scheme that cleanly separates brightness and texture information from color information. An ideal decomposition would isolate a grayscale component capturing illumination and textural details with minimum color interference, alongside a pure color component that can be independently corrected. This separation would enable a targeted restoration process, which first recovers accurate textural details and then uses this restored foundation as a reliable reference for precise color regeneration. + +YCbCr color space provides an optimal lossless decomposition for this task. The Y channel, which predominantly weights the green channel (0.587G compared to 0.299R and 0.114B), is significantly less contaminated by the color bias that predominantly affects red and blue channels in shadow regions. Additionally, the Y channel is specifically designed to preserve perceptually important texture details that align with human visual sensitivity. By first restoring this luminance component, we can recover accurate illumination and texture information without interference from color distortions. The restored luminance then serves as a geometric and semantic guide for subsequent color regeneration, ensuring spatial coherence between brightness and chrominance adjustments. This sequential enables dedicated networks to specialize in luminance refinement (LRNet) and chroma harmonization (CRNet) without mutual interference. + +# 4. Methodology + +A viable solution involves separating an image into texture $I_{t}$ and color $I_{c}$ components for texture restoration and color regeneration, respectively. We achieve this through an invertible decoupling function $D(\cdot)$ , which splits an RGB image $I$ into a luminance component $I_{y}$ and a chrominance component $I_{c}$ : + +$$ +\left(I _ {t}, I _ {c}\right) = \mathcal {D} (I). \tag {6} +$$ + +We adopt the conversion from RGB to YCbCr as $\mathcal{D}(I)$ , where $I_{t}$ corresponds to the Y (luminance) channel, and $I_{c}$ comprises the Cb and Cr (chrominance) channels. This decomposition isolates brightness and texture degradation in $I_{t}$ while retaining color distortions in $I_{c}$ . The inverse function $\mathcal{D}^{-1}$ reconstructs the final RGB image from processed components. This manner enables a divide-and-conquer strategy for shadow removal: + +$$ +\hat {I} = \mathcal {D} ^ {- 1} \mathcal {C} (\mathcal {R} (I _ {t}), I _ {c}), \tag {7} +$$ + +where $\mathcal{R}(\cdot)$ denotes the luminance restoration network (LRNet) and, $\mathcal{C}(\cdot)$ represents the color regeneration network (CRNet). $\hat{I}$ is the shadow-free output, and $I_{c}$ serves as a conditioning variable for color regeneration. Additionally, $\mathcal{D}^{-1}(\cdot)$ is the inverse procedure of $\mathcal{D}(\cdot)$ . + +As in Fig. 4, ShadowHack operates in three stages: 1) Decomposition. The RGB image is decomposed via $\mathcal{D}(\cdot)$ into luminance $I_{t}$ and color $I_{c}$ components; 2) Luminance Restoration. LRNet processes $I_{t}$ to recover illumination and texture, yield in $\hat{I}_t = \mathcal{R}(I_t)$ ; 3) Color Regeneration. + +![](images/d21659e7b04d4d660d5c69b80a0ee1a655841de9a3c58fa3b961cd359adf5565.jpg) +Figure 5. The rectified attention module. We compute two sets of attention. In $Att_1$ , both texture $(I_t)$ and color $(I_c)$ are involved while in $Att_2$ , we calculate attention with only color components. The difference $Att_1 - \lambda Att_2$ delivers more accurate interactions + +CRNet harmonizes colors using $I_{c}$ and $\hat{I}_{t}$ . Finally, the estimated luminance and color components are recombined using $\mathcal{D}^{-1}(\cdot)$ to produce the final shadow-free RGB result. Below, we detail LRNet and CRNet. + +# 4.1. Luminance Restoration Network + +To eliminate shadows in $I_{t}$ , enhancing the brightness of the shadow areas and restoring their textures is essential. As illustrated in Fig. 4 (a), the shallow two stages and the refinement module in the luminance restoration network (LRNet) are constructed using a Local Range block (LRB), a Multi-head Transpose Attention (MTA) module, and a Feed-Forward Network (FFN), while a RecTified Attention (RTA) module replaces LRB in the deeper stages. The LRB is designed to extract local texture information through a depth-wise convolution block, capturing fine-grained details. The MTA, which lies in the channel attention block (CAB), following previous works [5, 59], reallocates the weights of channels to enhance the relevant features. The FFN further processes these extracted features to improve texture restoration. Finally, the RTA module, introduced in the deeper layers, captures more global information by focusing on regions close to shadow boundaries, enabling the network to refine luminance restoration in complex shadow regions. + +Rectified Attention Module. In the shadow removal process, achieving robust enhancement in shadow regions requires restoration efforts to reference better surrounding, related non-shadow areas. Therefore, we replace the Local Range Block (LRB) with self-attention modules [20, 32] in the deeper layers to capture more distant information. Since the goal of LRNet is to cure shadow-contaminated texture and lightness, the feature from the deep layer (i.e., blocks with SAB equipped) becomes less and less discriminative to shadow. As a result, the attention matrix in deep layers, suffers from a distraction. In the first column of Fig. 6, the + +![](images/b776d461501287bc8fb4a296417a8d390c80c3f039f2155fb3ff5f8133c8f233.jpg) +Figure 6. An example of our rectified attention module. The star indicates the position of the query token. We exhibit the attention map from the first head of the encoder, latent, and decoder layers. The final attention is obtained by subtracting $\mathrm{Att}_2$ from $\mathrm{Att}_1$ . The attention matrix without rectify, $\mathrm{Att}_1$ and $\mathrm{Att}_2$ are normalized to [0, 1], while their differential result is in [-1, 1]. + +model tends to focus on only a limited number of tokens which may be inaccurate for shadow area refinement. We address this via a RecTified Attention (RTA) mechanism to rectify the original attention weight that is re-injecting contaminated chrominance channels $I_{c}$ , by combining them into $Q$ and $K$ calculations (Please notice that the $V$ is still clear without chromance), resulting in $Att_{1}$ . Since this operation re-integrates color bias to the calculation, inspired by [58], we calculate a second set with only $I_{c}$ as $Att_{2}$ . By subtracting the two sets, LRNet has the flexibility to get rid of chrominance information if needed. + +To be specific, we obtain the features $F_{t}$ corresponding to $I_{t}$ through convolution with downsampling, and compute the softmax attention map for $F_{t}$ and $F_{c}$ , as shown in Fig. 5. The query, key, and value matrices are formed via: + +$$ +Q _ {1} = \left[ F _ {t}; F _ {c} \right] W _ {1} ^ {Q}, K _ {1} = \left[ F _ {t}; F _ {c} \right] W _ {1} ^ {K}, \tag {8} +$$ + +$$ +Q _ {2} = F _ {c} W _ {2} ^ {Q}, K _ {2} = F _ {c} W _ {2} ^ {K}, V = F _ {t} W ^ {V}, +$$ + +where $W$ matrices are learnable projection parameters and "; refers to the concatenation of the two features. Next, the above two sets of queries and keys are used to compute their respective attentions simply by: + +$$ +\operatorname {A t t} _ {i} = \operatorname {S o f t m a x} \left(\frac {Q _ {i} K _ {i} ^ {\mathrm {T}}}{\sqrt {d}} + B\right), i \in \{1, 2 \}, \tag {9} +$$ + +where $B$ denotes the relative position bias. This allows us to apply the rectified attention mechanism as follows: + +$$ +\operatorname {R T A} \left(F _ {t}, F _ {c}\right) = \left(\operatorname {A t t} _ {1} - \lambda \cdot \operatorname {A t t} _ {2}\right) V, \tag {10} +$$ + +where $\lambda$ is a reweighting factor. As illustrated in Fig. 6, such a rectified attention mechanism creates a strong affinity with well-lit reference, urging the model to focus on relevant content. As validated in Sec. 5.3, our rectified attention can focus on more non-shadow area information to transfer robust clues for shadow removal. + +
DatasetsMethodShadow Region (S)Non-Shadow Region (NS)All Image (ALL)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
ISTD+DHAN [8] (AAAI'20)33.080.9889.4927.280.9727.3925.780.9587.74
G2R [36] (CVPR'21)33.880.9788.7135.940.9772.8130.850.9463.78
BMNet [65] (CVPR'22)38.170.9915.7237.950.9862.4234.340.9742.93
SG-ShadowNet [47] (ECCV'22)36.800.9905.9335.570.9782.9232.460.9623.41
ShadowFormer [13] (AAAI'23)39.670.9925.2138.820.9832.3035.460.9732.80
ShadowDiffusion [14] (CVPR'23)39.82-4.9038.90-2.3035.72-2.70
Li et al. [30] (ICCV'23)38.460.9895.9337.270.9772.9034.140.9603.39
Liu et al. [34] (AAAI'24)38.040.9905.6939.150.9842.3134.960.9682.87
RASM [32] (MM'24)40.730.9934.4139.230.9852.1736.160.9762.53
Homofformer [53] (CVPR'24)39.470.9934.7238.730.9842.2335.340.9752.64
OmniSR+GM [55] (AAAI'25)------34.200.973-
ShadowHack (Ours)40.560.9934.4639.660.9872.0936.310.9772.48
SRDDHAN [8] (AAAI'20)33.670.9788.9434.790.9794.8030.510.9495.67
BMNet [65] (CVPR'22)35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNet [47] (ECCV'22)--7.53--2.97--4.23
ShadowFormer [13] (AAAI'23)36.910.9895.9036.220.9893.4432.900.9584.04
ShadowDiffusion [14] (CVPR'23)38.720.9874.9837.780.9853.4434.730.9703.63
Li et al. [30] (ICCV'23)39.330.9856.0935.610.9672.9733.170.9403.83
Liu et al. [34] (AAAI'24)36.510.9835.4937.710.9863.0033.480.9673.66
DeS3 [24] (AAAI'24)37.450.9845.8838.120.9882.8334.110.9683.72
RASM [32] (MM'24)37.910.9885.0238.700.9922.7234.460.9763.37
Homofformer [53] (CVPR'24)38.810.9874.2539.450.9882.8535.370.9723.33
OmniSR+GM [55] (AAAI'25)------34.560.977-
ShadowHack (Ours)39.470.9914.1840.060.9942.3835.940.9822.90
+ +Table 1. Quantitative results on the ISTD+ and SRD datasets. The best results are in bold, while the second-best ones are underlined. + +# 4.2. Color Regeneration Network + +After restoring texture and lighting information in the luminance channel, the next task is to accurately regenerate color. Inspired by exemplar-based colorization models [23, 29, 39], we leverage contextual cues and condition-based guidance to improve color fidelity and ensure alignment with the semantic content of the images. Concretely, we propose a dual-encoder network, termed CRNet, which integrates a multiscale color feature extractor into the U-shaped architecture of LRNet. The color encoder is a ConvNext-v2utto model [37] pre-trained on ImageNet-21k with 2M parameters, which takes the Cb and Cr channels as input and extracts color features across multiple scales, as illustrated in Fig. 4 (b). To achieve precise color restoration, the network shall reference regions with similar/identical material properties and accurate color, using them as reliable guides for color regeneration. The outputs generated by LRNet provide qualified indexing information to facilitate this purpose. Therefore, we introduce an attention-based mechanism where luminance features serve as the query and key to compute similarity, while color features act as the value, aggregating essential color information. This approach injects color features into the skip connections of the U-shaped backbone, remarkably enriching the network's color regeneration capability. + +Multiple Checkpoint Training. We observe that the luminance network often yields better performance on the training set, which suggests the color regeneration network faces challenges in adapting to imperfect luminance out + +puts encountered during testing. To address this issue, we employ a simple yet effective way to overcome this issue, where multiple early-stage checkpoints from the luminance network training are randomly selected as candidates. This technique, applied only during training, strengthens the robustness of the color network without incurring additional computational costs at inference time. + +# 5. Experimental Validation + +# 5.1. Experimental Setup + +Implementation details. Our models are implemented in PyTorch, and the entire training process utilizes the AdamW optimizer [38] with momentum set to (0.9, 0.999) and a weight decay of $10^{-2}$ . The initial learning rate is set to $2 \times 10^{-4}$ . During training, we randomly crop the training data into patches of size $384 \times 384$ . For the base network of each stage, we use a four-layer encoder-decoder structure (from L1 to L4), consisting of three downsampling and three upsampling operations. The dimension of the first layer is 32. Our RTA modules are integrated into L3 and L4, with an overlapping ratio set to 0.5 and a rectified mechanism $\lambda_0$ value of 0.7. For training, we use L1 and VGG loss terms as employed in previous works [32, 34]. For detailed information on the implementation of each stage and training setup, please refer to the supplementary material. + +Datasets. We conduct our experiments on the following benchmark datasets for shadow removal: 1) Adjusted ISTD (ISTD+) dataset [27], which refines the ISTD dataset [48] + +![](images/f37dc20ab13f136ec8a4e765509da6d2bc32e6352818be0a953c6938d0ef5881.jpg) +(a) Input + +![](images/38461c8262550ce788585328fe55192401919a22ea5397c1b931647994569d90.jpg) +(b) BMNet [65] + +![](images/ee2b8fb14f897d8daadf792f0865955b25926c4636d26b4d45620a5d1d08fd1c.jpg) +(c) SG [47] + +![](images/4faf09f58d3a56a94f4fdcf7105442932506e3e0f65d83d7b826d20ee6d948b6.jpg) +(d) SF [13] + +![](images/b7c77ea1bc50baad530eb6a49cb24b724358bc9538983d25c1f5b8795db5fea9.jpg) +(e) HF [53] + +![](images/b98aed533779497ca16e14fe5cec61d7bda9d81cd4718725c57641e5fe7d07f2.jpg) +(f) RASM [32] + +![](images/55218b3e115f37398494d61e6b0b5b564ce2fe972f5bc4b1c7bb44afc06877f2.jpg) +(g) Ours + +![](images/6282831f4b3114f3622fd3f992e37aa01ba49b160feb674bd24de4b07d445c9d.jpg) +(h) GT + +![](images/d528eeb5c62f287eb153184cf74fb402d938a8d6c3b73fb05e01ddc364e6add3.jpg) +(a) Input + +![](images/35696371e640e1c04de01e03304e5a359a2d7b046b26bcb47b15a8255c03fba6.jpg) +(b) BMNet [65] + +![](images/e01c5c66741d828e27dc00b9be5b3471555d30ed2f3609827ad098366862834c.jpg) +Figure 7. Visual comparison on a sample from the $\mathrm{ISTD + }$ dataset. +(c) SG [47] + +![](images/5b3c3d3751b4ab75ba8d7c82a2747fc50a8eaee09d463f3e0b563466f60a56c0.jpg) +(d) DMTN [33] + +![](images/f5a8f8c7e679f52a9d15485a87be9bb71a004a3aaf63737a5bcaecf264773fae.jpg) +(e) DeS3 [24] + +![](images/3b1b901d86cb562b501a1a42add838f85c882f88bb9e97103d34f213759541c3.jpg) +(f) HF [53] + +![](images/5434a03aec5755c8c6d1aff0929e048b27f43b9bc62d2b8765be8d21b40969ee.jpg) +(g) Ours + +![](images/1efbff3d2ff6e82cc67f06b84527da8df2878520b9b6bb7a8fd343d6f04a5b75.jpg) +(h) GT + +
MetricRGB InputYCbCr InputRetinex-basedOurs
PSNR↑36.1635.6235.8536.31
RMSE↓2.532.622.632.48
+ +Table 2. Ablation study of the proposed decoupling framework on the ISTD+ dataset. + +by correcting the illumination discrepancies between the ground-truth and input images, includes 1330 training and 540 testing triplets (shadow image, shadow-free image, and shadow mask); 2) SRD dataset [43] consists of 2680 training and 408 testing pairs. We adopt the predicted masks provided by DHAN [8] for training and testing following previous works [32, 53, 65]; 3) UIUC dataset [16] provides 76 images with shadow mask for testing; and 4) UCF dataset [64] provided by Jin et al. [24] contains shadow images without shadow-free references. + +Evaluation metrics. To evaluate the performance of our proposed framework, we utilize the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE). PSNR and SSIM are both calculated in the RGB space, with higher values indicating better results. For RMSE, which is calculated in the LAB color space, the lower values indicate better results. To be consistent with previous works [32, 53, 65], we inference the full-resolution image and then evaluate the images at a resolution of $256 \times 256$ . + +# 5.2. Comparison with Prior Arts + +We compare our method with the state-of-the-art shadow removal approaches on benchmark datasets, including DHAN [8], G2R [36], BMNet [65], SG-ShadowNet [47], ShadowFormer [13], ShadowDiffusion [14], Li et al. [30], Liu et al. [34], DeS3 [24], Homoformer [53], RASM [32] + +![](images/a26935eb81525fa360f2b4d877dcb2328eeb68f2080d82f581234a5e694066c8.jpg) +Figure 8. Visual comparison on a sample from the SRD dataset. +(a) Input +Figure 9. Ablation study on rectified attention. The results are displayed in the luminance space. + +![](images/049b438133538ce0dcba1405a127464561a4b5c0848076ef30153af7a15d3d19.jpg) +(b) RGB Input + +![](images/2d02916a5f0aa03a8de21bba3c1d49a866486b11c0e458eb8b3e99409fce5bda.jpg) +(c) w/o rectify + +![](images/80c2e9c2bec6ce90bc368a60eb8796fb9ce8f43c6dac4ebcd090640bd4b636cd.jpg) +(d) Ours + +and OmniSR+GM [55] (GM means the version with shadow mask), as shown in Tab. 1. For a fair comparison, we use results published by the authors. Regarding missing data, we utilize available images or checkpoints provided by the authors and evaluate them via the same metrics. + +Our proposed method achieves the highest scores in most metrics on the ISTD+ dataset compared to other state-of-the-art methods. Specifically, our model attains a PSNR of 36.31 for all image regions, surpassing the second-highest result of 36.16 by 0.15. The high PSNR highlights our model's ability to recover texture details effectively, demonstrating superior reconstruction quality. Additionally, the RMSE value of 2.48 is the lowest, which indicates that our recoloring module plays a crucial role in accurately restoring color information and reducing errors across the entire image. From the magnified areas in Figs. 7 and 8, it can be observed that our method not only effectively corrects color bias but also provides superior edge correction compared to other methods. The second part of Tab. 1 presents a comparison of our results on the SRD dataset against other state-of-the-art methods. It can be observed that our method achieves a significant improvement in RMSE (-0.43), further demonstrating the substantial benefit of the decoupling approach for color recovery. + +
MethodShadow PSNR ↑Non-Shadow PSNR ↑All PSNR ↑
[Ft]39.9641.2736.60
[Ft; Fc]40.1741.2936.76
[Ft]&[Ft]40.0241.1836.58
[Ft; Fc]&[Ft; Fc]40.3541.3736.87
[Ft; Fc]&[Fc]40.3641.4036.90
+ +![](images/660c8b6c69f8a13b2ab8bdbce36d0fcf7f5d64a4b2dc20a6045fe1cb4a34b2d7.jpg) +(a) Input + +![](images/d372ecc92047120dceb1a5e38398083872bf22326e754e8c86e1eb42befd0790.jpg) +(b) SG [47] + +![](images/c8f5ad926afbd21af8d9e7c10cfaf1ecaa1d4a9aa0326af6ce87e054a392a0b5.jpg) +(c) SF [13] + +![](images/2d3c0c83baef97b872903b7eeab7379ad3b0f0bdcba79be90af533bc8c764f62.jpg) +(d) HF [53] + +![](images/66369a419170a8557032737d5f81ce7599e47d9b7f1f2b045be7758c0d41a4bd.jpg) +(e) Ours +Figure 10. Visual comparisons on the UIUC and UCF datasets. + +# 5.3. Ablation Study + +Effect of the decoupling framework. To validate the effectiveness of our decoupling framework, we fine-tune our baseline model structure to adapt to RGB and YCbCr input/output for end-to-end shadow removal and compare its performance with our decoupling framework. As shown in Tab. 2, our decoupling approach significantly improves both PSNR and RMSE. Moreover, based on Retinex decomposition, we separately recover L and R, and compare it with our decoupling method. Detailed architectures are provided in the supplementary material. This improvement indicates that this strategy benefits texture recovery and color correction. + +Effect of rectified attention module. We demonstrate the effectiveness of the rectified attention module in Tab. 3, by conducting experiments on rectification mechanisms. Furthermore, extra experiments are conducted on different feature combinations of rectified approaches to explore more suitable configurations. $[X]$ & $[Y]$ in Tab. 3 means performing a subtraction calculation of corresponding maps for rectification between $X$ and $Y$ . $[X]$ alone denotes only one pair of $Q$ and $K$ is used, where both inputs are $X$ . Clearly, leveraging color information, even with inherent bias, can allow shadow regions to better reference surrounding areas, resulting in significant performance improvements. As for potential color bias, rectification refines the attention map for greater precision. Furthermore, we find that $[F_t; F_c] \& [F_c]$ slightly outperforms $[F_t; F_c] \& [F_t; F_c]$ , and thus we ultimately choose $[F_t; F_c] \& [F_c]$ . Fig. 9 visualizes the effectiveness of rectification, which has a more precise result with less edge remaining. + +Table 3. Ablation study of the proposed rectified attention module on the SRD dataset. We calculate PSNRs on the luminance. + +
MethodShadow RMSE ↓Non-Shadow RMSE ↓All RMSE ↓
w/o cross-att.5.782.643.39
w/o multi checkpt.5.012.593.28
Ours4.182.382.90
+ +Table 4. Ablation study of the CRNet on the SRD dataset. + +![](images/9e4f5100a07a7580a95cde3e5005b6eb7de21db683f26ef18cd16c6fa2e0e0c6.jpg) +(a) dirty mask + +![](images/f663130dbcec884f9580ae68d11e253a7b5d8676b2dfd9a0683e27955168d0dc.jpg) +Figure 11. Robustness analysis under corrupted mask. The accurate mask is highlighted in red. For fairness, we only involve mask refinement network here, and it is not involved in any comparison. + +![](images/125cdfd79baba77ff3c038f7d95951382cace7ac25bf281c2f8e2522b1ad9138.jpg) +(b) refined mask (c) w/o refinement + +![](images/f8bd731dc0ee73f9930104a2041d3114a066b98006a5530a43e9bc458b54b2a6.jpg) +(d) w refinement + +Effect of the color restoration network. To validate the effectiveness of the key component and design in the CRNet, we replace the cross-attention-based color feature injection with concatenation along the feature channel. As demonstrated in Tab. 4, if cross-attention is replaced, the accuracy has dropped by a large margin, indicating the effectiveness of the cross-attention. Furthermore, the model suffers from a misalignment between training and testing luminance results if no previous checkpoints are involved. + +# 5.4. Robustness Analysis + +Generalization to other datasets. To demonstrate the generalization capability of ShadowHack, we evaluate it on the UIUC [16] and UCF [64] datasets with our $\mathrm{ISTD + }$ pretrained model. The results are presented in Fig. 10, where ShadowHack achieves the cleanest shadow removal with minimal color bias. More comparisons are available in the supplementary material. + +Robustness to inaccurate masks. To get rid of the performance degradation brought by inaccurate masks, we propose a mask refine network, which corrects the mask for better generalization. After our remediation, the result is recovered from inaccurate enlightening, as in Fig. 11 shows. More details are available in the supplementary material. + +# 6. Conclusion + +This work proposed a novel framework, ShadowHack, which decouples images into luminance and color components, enabling targeted processing for more effective shadow removal. This approach can well address limitations in existing end-to-end models by enhancing the use of color information to improve both texture and luminance recovery. To further boost shadow removal, a rectified attention mechanism has been designed to boost the reference ability of self-attention. Additionally, the color regeneration network has demonstrated its critical role in accurately correcting image colors. Extensive experimental results have underscored the superior performance and robustness of our model. + +# References + +[1] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2(3-26):2, 1978. 1, 3 +[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In ICCV, pages 12470-12479, 2023. 1 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640, 2021. 3 +[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 3 +[5] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. arXiv preprint arXiv:2310.11881, 2023. 3, 5 +[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In CVPR, pages 22367-22377, 2023. 3 +[7] Michael E Churma. Blue shadows: physical, physiological, and psychological causes. Applied optics, 33(21):4719-4722, 1994. 1 +[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting GAN. In AAAI, pages 10680-10687, 2020. 6, 7 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[10] Graham D. Finlayson, Steven D. Hordley, and Mark S. Drew. Removing shadows from images. In ECCV (4), pages 823-836, 2002. 2 +[11] Graham D. Finlayson, Steven D. Hordley, Cheng Lu, and Mark S. Drew. On the removal of shadows from images. IEEE Trans. Pattern Anal. Mach. Intell., 28(1):59-68, 2006. 2 +[12] Graham D. Finlayson, Mark S. Drew, and Cheng Lu. Entropy minimization for shadow removal. Int. J. Comput. Vis., 85(1): 35-57, 2009. 2 +[13] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 1, 3, 6, 7, 8, 2, 4 +[14] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 6, 7 +[15] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In ICCV, pages 12999-13008. IEEE, 2023. 3 + +[16] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2956-2967, 2013. 1, 2, 7, 8 +[17] Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness. Int. J. Comput. Vis., 131(1): 48-66, 2023. 1, 3, 2 +[18] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: low-light image enhancement via illumination map estimation. IEEE Trans. Image Process., 26(2):982-993, 2017. 1, 2, 3 +[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 3 +[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 5 +[21] Jun-Wei Hsieh, Wen-Fong Hu, Chia-Jung Chang, and Yung-Sheng Chen. Shadow elimination for effective moving object detection by gaussian shadow modeling. Image Vis. Comput., 21(6):505-516, 2003. 1 +[22] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2 +[23] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Trans. Graph., 41(6):205:1-205:16, 2022. 6 +[24] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T. Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 1, 6, 7, 2, 5 +[25] Salman Hameed Khan, Mohammed Bennamoun, Ferdous Ahmed Sohel, and Roberto Togneri. Automatic shadow detection and removal from a single image. IEEE Trans. Pattern Anal. Mach. Intell., 38(3):431-446, 2016. 2 +[26] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 1, 2 +[27] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8577-8586, 2019. 1, 2, 6 +[28] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9088-9101, 2022. 1, 2, 3 +[29] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809, 2020. 6 +[30] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor W. Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13009-13018, 2023. 1, 2, 6, 7 +[31] Feng Liu and Michael Gleicher. Texture-consistent shadow removal. In ECCV (4), pages 437-450. Springer, 2008. 3 +[32] Hengxing Liu, Mingjia Li, and Xiaojie Guo. Regional attention for shadow removal. In ACM Multimedia, 2024. 1, 3, 5, 6, 7, 2, 4 +[33] Jiawei Liu, Qiang Wang, Huijie Fan, Wentao Li, Liangqiong Qu, and Yandong Tang. A decoupled multi-task network for + +shadow removal. IEEE Trans. Multim., 25:9449-9463, 2023. 7, 1, 5, 6 +[34] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson W. H. Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 1, 2, 3, 6, 7 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3 +[36] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 1, 2, 6, 7 +[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11966-11976, 2022. 6, 2 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[39] Peng Lu, Jinbei Yu, Xujun Peng, Zhaoran Zhao, and Xiaojie Wang. Gray2colornet: Transfer more colors from reference image. In ACM Multimedia, pages 3210-3218, 2020. 6 +[40] David K. Lynch. Shadows. Appl. Opt., 54(4):B154-B164, 2015. 1 +[41] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In WACV, pages 4313-4322, 2024. 1 +[42] Sohail Nadimi and Bir Bhanu. Physical models for moving shadow and object detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 26(8):1079-1087, 2004. 1 +[43] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson W. H. Lau. Desshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 2308-2316, 2017. 1, 2, 7 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 2 +[45] Yael Shor and Dani Lischinski. The shadow meets the mask: Pyramid-based shadow removal. Comput. Graph. Forum, 27 (2):577-586, 2008. 2, 3 +[46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 2 +[47] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV (19), pages 361-378, 2022. 1, 6, 7, 8, 3, 4, 5 +[48] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 1, 2, 6 +[49] Tianyu Wang, Jianming Zhang, Haitian Zheng, Zhihong Ding, Scott Cohen, Zhe Lin, Wei Xiong, Chi-Wing Fu, Luis Figueroa, and Soo Ye Kim. Metashadow: Object-centered shadow detection, removal, and synthesis. In CVPR, pages 28252-28262, 2025. 1 + +[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17662-17672, 2022. 3 +[51] Tai-Pang Wu and Chi-Keung Tang. A bayesian approach for shadow extraction from a single image. In ICCV, pages 480–487, 2005. 2 +[52] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4784-4793, 2022. 3 +[53] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 1, 3, 6, 7, 8, 2, 4, 5 +[54] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. Int. J. Comput. Vis., 132(9):3610-3632, 2024. 1 +[55] Jiamin Xu, Zelong Li, Yuxin Zheng, Chenyu Huang, Renshu Gu, Weiwei Xu, and Gang Xu. Omnisr: Shadow removal under direct and indirect lighting. arXiv preprint arXiv:2410.01719, 2024. 3, 6, 7, 2 +[56] Han Yang, Tianyu Wang, Xiaowei Hu, and Chi-Wing Fu. Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12687-12698, 2023. 2 +[57] Qingxiong Yang, Kar-Han Tan, and Narendra Ahuja. Shadow removal using bilateral filtering. IEEE Trans. Image Process., 21(10):4361-4368, 2012. 2 +[58] Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. arXiv preprint arXiv:2410.05258, 2024. 5, 1 +[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729, 2022. 3, 5 +[60] Ling Zhang, Qing Zhang, and Chunxia Xiao. Shadow remover: Image shadow removal based on illumination recovering optimization. IEEE Trans. Image Process., 24(11): 4623-4636, 2015. 2 +[61] Ling Zhang, Chengjiang Long, Xiaolong Zhang, and Chunxia Xiao. RIS-GAN: explore residual and illumination with generative adversarial networks for shadow removal. In AAAI, pages 12829-12836, 2020. 1, 2 +[62] Wuming Zhang, Xi Zhao, Jean-Marie Morvan, and Liming Chen. Improving shadow suppression for illumination robust face recognition. IEEE Trans. Pattern Anal. Mach. Intell., 41 (3):611-624, 2019. 1 +[63] Wei Zhou, Julie Stephany Berrio, Stewart Worrall, and Eduardo M. Nebot. Automated evaluation of semantic segmentation robustness for autonomous driving. IEEE Trans. Intell. Transp. Syst., 21(5):1951-1963, 2020. 1 +[64] Jiejie Zhu, Kegan G. G. Samuel, Syed Zain Masood, and Marshall F. Tappen. Learning to recognize shadows in monochromatic natural images. In CVPR, pages 223-230, 2010. 7, 8 + +[65] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5617-5626, 2022. 2, 6, 7, 1, 3, 4, 5 \ No newline at end of file diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/images.zip b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..29b8f2de1fdb6266fc7505de4351e655124516a4 --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a96369713987cdf9a7e44b82a9807c6c9e302ac9c46ea46cc798c13ae9d97c62 +size 739914 diff --git a/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/layout.json b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..403eb64ed787ca95892986dfbd65f6553050f1e9 --- /dev/null +++ b/2025/ShadowHack_ Hacking Shadows via Luminance-Color Divide and Conquer/layout.json @@ -0,0 +1,11322 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 538, + 121 + ], + "type": "text", + "content": "ShadowHack: Hacking Shadows via Luminance-Color Divide and Conquer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 143, + 399, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 143, + 399, + 174 + ], + "spans": [ + { + "bbox": [ + 214, + 143, + 399, + 174 + ], + "type": "text", + "content": "Jin Hu* Mingjia Li* Xiaojie Guo† Tianjin University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "spans": [ + { + "bbox": [ + 165, + 174, + 441, + 185 + ], + "type": "text", + "content": "{jinhu,mingjiali}@tju.edu.cn,xj.max.guo@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 441 + ], + "type": "text", + "content": "Shadows introduce challenges such as reduced brightness, texture deterioration, and color distortion in images, complicating a holistic solution. This study presents ShadowHack, a divide-and-conquer strategy that tackles these complexities by decomposing the original task into luminance recovery and color remedy. To brighten shadow regions and repair the corrupted textures in the luminance space, we customize LRNet, a U-shaped network with a rectified attention module, to enhance information interaction and recalibrate contaminated attention maps. With luminance recovered, CRNet then leverages cross-attention mechanisms to revive vibrant colors, producing visually compelling results. Extensive experiments on multiple datasets are conducted to demonstrate the superiority of ShadowHack over existing state-of-the-art solutions both quantitatively and qualitatively, highlighting the effectiveness of our design. Our code is available at https://github.com/lime-j/ShadowHack." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 467, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 467, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 64, + 467, + 286, + 491 + ], + "type": "text", + "content": "Shadow is a color as light is, but less brilliant; light and shadow are only the relation of two tones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 214, + 491, + 283, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 491, + 283, + 502 + ], + "spans": [ + { + "bbox": [ + 214, + 491, + 283, + 502 + ], + "type": "text", + "content": "Paul Cezanne" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 629 + ], + "type": "text", + "content": "Shadows are the absence of light, an everyday phenomenon that often shapes the aesthetic beauty of nature. However, shadows obscure the actual shape of instances, diminishing their visibility. Very likely, they deteriorate the performance of various machine vision tasks, such as object detection [21, 42], face recognition [62], and semantic segmentation [63], to name just a few. Consequently, shadow removal has become a critical research topic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 677 + ], + "type": "text", + "content": "Recent advancements in deep learning have greatly improved the efficiency and accuracy of shadow removal. Specifically, DeshadowNet [43] marked a pioneering step with its end-to-end Convolutional Neural Network (CNN)" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 211, + 382, + 297 + ], + "blocks": [ + { + "bbox": [ + 316, + 211, + 382, + 297 + ], + "lines": [ + { + "bbox": [ + 316, + 211, + 382, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 382, + 297 + ], + "type": "image", + "image_path": "6b30d9a990595dcfae1f4171d5175f8f36c2153bb5f291532432ae46c6726f55.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 299, + 365, + 309 + ], + "lines": [ + { + "bbox": [ + 335, + 299, + 365, + 309 + ], + "spans": [ + { + "bbox": [ + 335, + 299, + 365, + 309 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 383, + 212, + 439, + 296 + ], + "blocks": [ + { + "bbox": [ + 383, + 212, + 439, + 296 + ], + "lines": [ + { + "bbox": [ + 383, + 212, + 439, + 296 + ], + "spans": [ + { + "bbox": [ + 383, + 212, + 439, + 296 + ], + "type": "image", + "image_path": "4a9cb6cd8ada88587366c5818631f8fae929cecd51810af087ce46845ea2370e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 299, + 426, + 308 + ], + "lines": [ + { + "bbox": [ + 389, + 299, + 426, + 308 + ], + "spans": [ + { + "bbox": [ + 389, + 299, + 426, + 308 + ], + "type": "text", + "content": "(b) SG [47]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 440, + 212, + 495, + 296 + ], + "blocks": [ + { + "bbox": [ + 440, + 212, + 495, + 296 + ], + "lines": [ + { + "bbox": [ + 440, + 212, + 495, + 296 + ], + "spans": [ + { + "bbox": [ + 440, + 212, + 495, + 296 + ], + "type": "image", + "image_path": "3aa9e851605414b79f8a90731ba1bb6904f6a30181ba428593d833e899c18569.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 299, + 483, + 308 + ], + "lines": [ + { + "bbox": [ + 445, + 299, + 483, + 308 + ], + "spans": [ + { + "bbox": [ + 445, + 299, + 483, + 308 + ], + "type": "text", + "content": "(c) HF [53]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 318, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 341 + ], + "type": "text", + "content": "Figure 1. Visual comparison between our method and state-of-the-art methods in both Luminance and RGB spaces." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 497, + 212, + 552, + 296 + ], + "blocks": [ + { + "bbox": [ + 497, + 212, + 552, + 296 + ], + "lines": [ + { + "bbox": [ + 497, + 212, + 552, + 296 + ], + "spans": [ + { + "bbox": [ + 497, + 212, + 552, + 296 + ], + "type": "image", + "image_path": "dd5c55a873d199a5d7fbad2a127472d670dad0b5d9ea58fd1b849158d612fcb6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 299, + 535, + 308 + ], + "lines": [ + { + "bbox": [ + 507, + 299, + 535, + 308 + ], + "spans": [ + { + "bbox": [ + 507, + 299, + 535, + 308 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 353, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 556, + 496 + ], + "type": "text", + "content": "approach, while subsequent GAN-based methods [22, 36, 48, 61] further advanced the field by adopting adversarial learning frameworks, enabling effective unpaired training. Additionally, Wan et al. [47] introduced style transfer, and Li et al. [30] incorporated pre-trained inpainting priors to better restore textural information. Transformer-based architectures, such as ShadowFormer [13], HomoFormer [53], and RASM [32], have since been developed to restore shadow images more effectively, while diffusion-based approaches [14, 24, 41, 49] have also demonstrated promising results, particularly in complex scenarios where previous generative models often struggled." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "type": "text", + "content": "Despite these advancements, the odyssey of removing shadows is still far from its end. As evidenced in Fig. 1, even state-of-the-art models fail to recover color and underlying texture accurately. The difficulty stems from the complex and intertwined nature of shadow degradation, which involves not only reduced brightness in shadow (umbra) regions [40], but also texture deterioration from the imaging process [30, 34], and color distortion from surface material properties and environmental color influences [7, 28, 47], an example of which is in Fig. 2. This complexity naturally suggests breaking down the original task, with its entangled interferences, into isolated and simpler sub-problems, each targeting a specific aspect and addressed independently." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "Following the above line of thought, some works [16, 27, 28] attempted to estimate light and color biases as scalars. Insights can also be drawn from the low-light enhancement literature, where the Retinex decomposition [1, 26] is commonly employed [2, 17, 18, 54]. Under the Retinex theory," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 136, + 693 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 294, + 712 + ], + "type": "text", + "content": "†Corresponding author. This work was supported by National Natural Science Foundation of China under Grant nos. 62372251 and 62072327." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11403" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 195, + 250 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 195, + 250 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 195, + 250 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 195, + 250 + ], + "type": "image", + "image_path": "2ec9418c33c512cb4125e6438f08416b5f97c607ef5d5be82c37ae6fa08b6e88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 252, + 150, + 262 + ], + "lines": [ + { + "bbox": [ + 96, + 252, + 150, + 262 + ], + "spans": [ + { + "bbox": [ + 96, + 252, + 150, + 262 + ], + "type": "text", + "content": "(a) Color Shift" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 70, + 293, + 163 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 293, + 163 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 293, + 163 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 293, + 163 + ], + "type": "image", + "image_path": "e0ffca9b6ead9a3ca2a7f9bd9937be4ff358346cb3ba574aa40c01eae3bddab6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 165, + 286, + 176 + ], + "lines": [ + { + "bbox": [ + 205, + 165, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 205, + 165, + 286, + 176 + ], + "type": "text", + "content": "(b) Texture Distortion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 202, + 178, + 294, + 251 + ], + "blocks": [ + { + "bbox": [ + 202, + 178, + 294, + 251 + ], + "lines": [ + { + "bbox": [ + 202, + 178, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 202, + 178, + 294, + 251 + ], + "type": "image", + "image_path": "290f65f281a42d0fdb019ff16c0687a8a92e6b82359364d5157d761603a9d7f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 252, + 294, + 263 + ], + "lines": [ + { + "bbox": [ + 203, + 252, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 203, + 252, + 294, + 263 + ], + "type": "text", + "content": "(c) Illumination Underlit" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 273, + 297, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 297, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 297, + 318 + ], + "type": "text", + "content": "Figure 2. Example of the complex degradation in a shadow image. a) Color shift in the umbra region; (b) Structural texture deterioration in the penumbra transition area; (c) Attenuated illumination levels across the shadow (Blue indicates lower illumination)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 338, + 297, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 297, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 297, + 434 + ], + "type": "text", + "content": "an image can be decomposed into a structured illumination and a colored texture-rich counterpart. However, these approaches encounter challenges due to the inherent inhomogeneity of shadows. In other words, shadow removal extends beyond low-light enhancement and brings additional color biases to the reflectance map. This realization underscores the need for more sophisticated solutions to effectively disentangle the complex degradations introduced by shadows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 436, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 628 + ], + "type": "text", + "content": "This work invents ShadowHack, an effective shadow removal framework that individually tackles illumination and color components. ShadowHack starts with a simple yet highly effective hack for shadow decomposition. Unlike conventional methods that decouple illumination from color and texture, we alternatively separate color from illumination and texture. Once the input images are decomposed into luminance and color components, two dedicated networks are established, i.e., the Luminance Restoration Network (LRNet) and the Color Regeneration Network (CRNet). LRNet, a lightweight transformer-based restoration model, focuses on recovering illumination and texture. To enable an effective reference for shadow region refinement, we introduce the RecTified Attention (RTA) Module, which employs a rectified mechanism on window-based attentions to support high-quality luminance restoration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 715 + ], + "type": "text", + "content": "Afterward, the processed luminance serves as a guide for querying accurate color information in shadow regions from the non-shadow areas. We inject color information via an ImageNet-pre-trained color encoder, which integrates color data into the intermediate features of the skip connections of our CRNet. This is facilitated by cross-attention modules that align the color in the shadow region with patches" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": "of similar luminance. Through this sequential processing, ShadowHack achieves effective shadow removal, supported by extensive experimental results. The primary contributions of this study are summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 121, + 556, + 300 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "type": "text", + "content": "- Based on an in-depth analysis of shadow properties and the limitations in current end-to-end decomposition methods, we introduce ShadowHack, a novel shadow removal framework that breaks shadow removal into two distinct subtasks: luminance recovery and color regeneration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 180, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 555, + 239 + ], + "type": "text", + "content": "- We identified the distraction phenomenon in the attention matrix of shadow removal networks, and propose RecTified Attention (RTA) to regain focus in attention layers with marginal computational overhead, enabling precise restoration of irregular texture and illumination." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 240, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 556, + 300 + ], + "type": "text", + "content": "- We design two specialized networks, LRNet and CRNet, tailored for luminance refinement and color regeneration. Together, these networks allow our model to achieve accurate and robust results, setting new state-of-the-art performance across benchmarking datasets." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 311, + 401, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 401, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 401, + 323 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 331, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 713 + ], + "type": "text", + "content": "Shadow Removal. Traditional methods, such as [57] and [45], leveraged principles of physics and mathematical statistics on factors like illumination [10, 60], color [25, 51], region [16], and gradient [11, 12]. However, these methods often struggle with complex real-world scenes, where shadows interact in nuanced ways with surrounding colors and textures. Deep learning has recently made significant strides in the field [27, 43, 48]. For instance, Qu et al. [43] proposed to extract both global and semantic features from images, marking an early attempt to integrate deep learning into shadow removal. Later, Wang et al. [48] proposed a multitask framework that performs shadow detection and removal concurrently, enhancing overall accuracy. Generative Adversarial Networks (GANs) have also been applied to this task. As a representative, RIS-GAN [61] incorporates residual and illumination features to refine shadow regions, while other works [22, 36] use unsupervised GANs to generate shadows for achieving cycle consistency. Li et al. [30] adopted the concept of inpainting, simultaneously using a shadow image encoder and a generative-pretrained image encoder to extract features. Zhu et al. [65] employed normalizing flow for a reversible link between shadow generation and removal. Additionally, Le et al. [28] proposed a linear model that adjusts the global lightness of shadow regions through physical modeling. More recently, diffusion models [46] have shown even greater generative capabilities than GANs. Guo et al. [14] pioneered the use of diffusion models for shadow removal, refining shadow masks throughout the process. Liu et al. [34] applied Retinex decomposition [18, 26] and used diffusion models to refine textures under the guidance of the obtained illumination. However, a primary drawback of these diffusion models is their intensive computational cost." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11404" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": "ViT-based Restoration. Vision Transformers (ViTs) have demonstrated impressive performance across vision tasks, owing to their powerful attention mechanisms that can effectively capture long-range dependencies [3, 9, 19, 35, 52]. In image restoration, IPT [4] was among the first ViT-based solutions. Unfortunately, its dependence on large-scale paired datasets for training restricts its applicability when data is scarce. Subsequent models have sought to mitigate this limitation with more efficient architectures. Uformer [50] combined UNet with a window attention, yielding strong results, while Restormer [59] introduced transposed attention for efficient channel-wise processing. Further, X-Restormer [5] extended Restormer by integrating the OCA [6] module, providing a generalized framework for image restoration through rigorous analysis. Specific to the domain of shadow removal, ShadowFormer [13] introduced an attention mechanism that facilitates interaction between shadow and non-shadow regions, while HomoFormer [53] improved the transformer by integrating a shuffling mechanism into its transformer-based encoder. RASM [32] refreshed the window-based attention with a sliding-window mechanism, achieving more accurate results. More recently, OmniSR [55] leveraged DINO features and depth maps to incorporate semantic and geometric attention into self-attention, expanding its perception of image content and ability to distinguish shadows. Even though, as discussed in Sec. 1, there remains significant room for the community to improve the effectiveness of shadow removal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 416, + 162, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 162, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 162, + 430 + ], + "type": "text", + "content": "3. Problem Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": "Degradation in umbra primarily encompasses reduced brightness, color distortion, and texture degradation (e.g., noise and blurriness), as illustrated in Fig. 2. To better understand the types of involved degradation, we begin by formulating shadow effects via a modified image formation model as proposed in [1, 28, 45]. In this model, the observed pixel intensity " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "I(x,\\lambda)" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " at a given spatial coordinate " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " and wavelength " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " is expressed as the product of shading " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "S(x,\\lambda)" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " and albedo " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "A(x,\\lambda)" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 550, + 296, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 550, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 117, + 550, + 296, + 563 + ], + "type": "interline_equation", + "content": "I (x, \\lambda) = S (x, \\lambda) \\cdot A (x, \\lambda). \\tag {1}", + "image_path": "49bcdcef7c2d29518bee57185367787a056e9193e8a2a7e8a55efcd86db89b57.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "text", + "content": "In scenes with a single dominant light source (e.g., sunlight in outdoor settings), the shading for a pixel " + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "text", + "content": " can be decomposed into direct " + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "inline_equation", + "content": "S_{d}(x, \\lambda)" + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "text", + "content": " and ambient " + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "inline_equation", + "content": "S_{a}(x, \\lambda)" + }, + { + "bbox": [ + 55, + 568, + 296, + 616 + ], + "type": "text", + "content": " components:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 622, + 296, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 622, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 88, + 622, + 296, + 635 + ], + "type": "interline_equation", + "content": "I (x, \\lambda) = \\left(S _ {d} (x, \\lambda) + S _ {a} (x, \\lambda)\\right) \\cdot A (x, \\lambda). \\tag {2}", + "image_path": "a23ea3a81f6a4fd5f18f9c0be65387f941e1ed9bb6628f0ec4631deebfb743bb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 665 + ], + "type": "text", + "content": "However, when an object occludes the primary light source, shadow regions occur, and the pixel intensity becomes:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 670, + 296, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 670, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 102, + 670, + 296, + 685 + ], + "type": "interline_equation", + "content": "I (x, \\lambda) = a (x) \\cdot S _ {a} (x, \\lambda) \\cdot A (x, \\lambda), \\tag {3}", + "image_path": "798ab92fa782fded566a77d58351fe8184b1085366bec10624587bcfb6dcc0aa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "a(x) \\in [0,1]" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " designates a spatially varying attenuation function. This non-uniform attenuation of light forms" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 432, + 135 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 432, + 135 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 432, + 135 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 432, + 135 + ], + "type": "image", + "image_path": "bd8ea2d5b45784a606493168347da4db9d4564b56f686c8a0965ee10bef0183c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 137, + 411, + 146 + ], + "lines": [ + { + "bbox": [ + 337, + 137, + 411, + 146 + ], + "spans": [ + { + "bbox": [ + 337, + 137, + 411, + 146 + ], + "type": "text", + "content": "(a)ISTD+training set" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 553, + 135 + ], + "type": "image", + "image_path": "0b26a5997b48d572781fb6515455d27baa78d670433afdf66f41a402fb59d314.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 137, + 529, + 146 + ], + "lines": [ + { + "bbox": [ + 460, + 137, + 529, + 146 + ], + "spans": [ + { + "bbox": [ + 460, + 137, + 529, + 146 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 460, + 137, + 529, + 146 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 460, + 137, + 529, + 146 + ], + "type": "text", + "content": " testing set" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 155, + 433, + 218 + ], + "blocks": [ + { + "bbox": [ + 315, + 155, + 433, + 218 + ], + "lines": [ + { + "bbox": [ + 315, + 155, + 433, + 218 + ], + "spans": [ + { + "bbox": [ + 315, + 155, + 433, + 218 + ], + "type": "image", + "image_path": "0b480c837dc7342b80d92abe57dde87203ab4c23a3342a66382c841b206b364c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 220, + 407, + 229 + ], + "lines": [ + { + "bbox": [ + 342, + 220, + 407, + 229 + ], + "spans": [ + { + "bbox": [ + 342, + 220, + 407, + 229 + ], + "type": "text", + "content": "(c) SRD training set" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 155, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 436, + 155, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 436, + 155, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 436, + 155, + 553, + 218 + ], + "type": "image", + "image_path": "3f67cd43ea73fb250a71d11e122431b896debc62479d76d50600abf26f57000c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 220, + 526, + 229 + ], + "lines": [ + { + "bbox": [ + 463, + 220, + 526, + 229 + ], + "spans": [ + { + "bbox": [ + 463, + 220, + 526, + 229 + ], + "type": "text", + "content": "(d) SRD testing set" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 239, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 262 + ], + "type": "text", + "content": "Figure 3. Color bias in shadow regions between the reflectances of shadow and shadow-free images by Retinex decomposition [18]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "type": "text", + "content": "shadow. This involves not only illumination variance but also challenging inconsistency in texture[31], an example of which is illustrated in Fig. 2 (b) and (c). To obtain an RGB image" + }, + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "type": "inline_equation", + "content": "I_{RGB} \\coloneqq \\{I_R, I_G, I_B\\}" + }, + { + "bbox": [ + 313, + 277, + 555, + 338 + ], + "type": "text", + "content": ", the observed color intensity at each pixel is integrated across wavelengths as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 346, + 555, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 384 + ], + "type": "interline_equation", + "content": "I _ {k} (x) = \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} S (x, \\lambda) \\cdot A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda , k \\in \\{R, G, B \\} \\tag {4}", + "image_path": "4f06eb49676efdc6cd81d910b4f4d26360bec5188b1d3d98219ddf5dd8755bb9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "P_{k}(\\lambda)" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": " is the spectral sensitivity function of the camera for the " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": "-channel and " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "[\\lambda_{\\mathrm{min}}, \\lambda_{\\mathrm{max}}]" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": " defines the visible spectrum range. Previous shadow removal and low-light enhancement methods [15, 17, 18, 34] suggest that it would be neat to decompose lightness from the shadow image following the Retinex theory. In these works, the image is assumed to be decomposed into a one-channel illumination map " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "L(x)" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": " and an illumination-invariant counterpart (reflectance) " + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "inline_equation", + "content": "R_{k}(x)" + }, + { + "bbox": [ + 313, + 384, + 556, + 504 + ], + "type": "text", + "content": ". For sRGB space and white-balanced images, the operation can be formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 513, + 555, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 513, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 318, + 513, + 555, + 551 + ], + "type": "interline_equation", + "content": "I _ {k} (x) \\approx L (x) \\cdot R _ {k} (x) = L (x) \\int_ {\\lambda_ {\\min }} ^ {\\lambda_ {\\max }} A (x, \\lambda) \\cdot P _ {k} (\\lambda) d \\lambda . \\tag {5}", + "image_path": "2e23a0da6b0a47d3ebfb37626c00f0af63104ade4e62b9b8980d3196e81f6935.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 552, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 696 + ], + "type": "text", + "content": "We argue that shadow posits not just illumination map changes, but shifts that variants across the color. To quantify this, we plot the per-channel color difference between the input and ground-truth images in the reflectance map derived from a Retinex-based illumination estimator (Fig.3). The reflectance decomposition exhibits substantial color bias, particularly in the red and blue channels, violating Retinex's assumption of illumination-invariant reflectance. This color shift cannot be adequately addressed by previous Retinex-inspired decomposition methods [15, 34], as their core premise which reflectance remains uncontaminated by illumination effects are broken down in shadow regions." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 485, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 485, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 485, + 713 + ], + "type": "text", + "content": "Here, we omit the impact from the ISP pipeline." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11405" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 260, + 297 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 260, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 260, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 260, + 297 + ], + "type": "image", + "image_path": "31724fe29116adeed49ae31780ed0afe19912db570cde5388f8aa3a38f7eeba6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 299, + 211, + 311 + ], + "lines": [ + { + "bbox": [ + 62, + 299, + 211, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 211, + 311 + ], + "type": "text", + "content": "(a) Luminance Restoration Network" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 322, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 555, + 344 + ], + "type": "text", + "content": "Figure 4. The overall architecture of ShadowHack comprises two core components: (a) a luminance restoration network, which focuses on recovering illumination and texture, and (b) a color regeneration network dedicated to recalibrating harmonized colors for natural outputs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 261, + 77, + 552, + 295 + ], + "blocks": [ + { + "bbox": [ + 261, + 77, + 552, + 295 + ], + "lines": [ + { + "bbox": [ + 261, + 77, + 552, + 295 + ], + "spans": [ + { + "bbox": [ + 261, + 77, + 552, + 295 + ], + "type": "image", + "image_path": "1ff812b5335c13c5202b750e44775930642917100f75250bcee5ac6b644c9020.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 299, + 396, + 311 + ], + "lines": [ + { + "bbox": [ + 264, + 299, + 396, + 311 + ], + "spans": [ + { + "bbox": [ + 264, + 299, + 396, + 311 + ], + "type": "text", + "content": "(b) Color Regeneration Network" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 355, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 297, + 499 + ], + "type": "text", + "content": "We propose a novel approach to address the triple degradation of reduced brightness, texture deterioration, and color distortion in a divide-and-conquer manner. The key insight is identifying a decomposition scheme that cleanly separates brightness and texture information from color information. An ideal decomposition would isolate a grayscale component capturing illumination and textural details with minimum color interference, alongside a pure color component that can be independently corrected. This separation would enable a targeted restoration process, which first recovers accurate textural details and then uses this restored foundation as a reliable reference for precise color regeneration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "content": "YCbCr color space provides an optimal lossless decomposition for this task. The Y channel, which predominantly weights the green channel (0.587G compared to 0.299R and 0.114B), is significantly less contaminated by the color bias that predominantly affects red and blue channels in shadow regions. Additionally, the Y channel is specifically designed to preserve perceptually important texture details that align with human visual sensitivity. By first restoring this luminance component, we can recover accurate illumination and texture information without interference from color distortions. The restored luminance then serves as a geometric and semantic guide for subsequent color regeneration, ensuring spatial coherence between brightness and chrominance adjustments. This sequential enables dedicated networks to specialize in luminance refinement (LRNet) and chroma harmonization (CRNet) without mutual interference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 354, + 398, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 398, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 398, + 369 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": "A viable solution involves separating an image into texture " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": " components for texture restoration and color regeneration, respectively. We achieve this through an invertible decoupling function " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": ", which splits an RGB image " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": " into a luminance component " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "I_{y}" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": " and a chrominance component " + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 374, + 556, + 445 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 400, + 446, + 555, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 446, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 400, + 446, + 555, + 460 + ], + "type": "interline_equation", + "content": "\\left(I _ {t}, I _ {c}\\right) = \\mathcal {D} (I). \\tag {6}", + "image_path": "586e214bf152e3c78a4d73d379219d4a4a40b65df7a97f415a6392419794d609.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": "We adopt the conversion from RGB to YCbCr as " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(I)" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": " corresponds to the Y (luminance) channel, and " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": " comprises the Cb and Cr (chrominance) channels. This decomposition isolates brightness and texture degradation in " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": " while retaining color distortions in " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": ". The inverse function " + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{-1}" + }, + { + "bbox": [ + 313, + 464, + 556, + 560 + ], + "type": "text", + "content": " reconstructs the final RGB image from processed components. This manner enables a divide-and-conquer strategy for shadow removal:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 387, + 570, + 555, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 570, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 387, + 570, + 555, + 585 + ], + "type": "interline_equation", + "content": "\\hat {I} = \\mathcal {D} ^ {- 1} \\mathcal {C} (\\mathcal {R} (I _ {t}), I _ {c}), \\tag {7}", + "image_path": "6b3cc1d996c895ebbe7a1ca7984e709ea738ffad9205af1e83faa980485a714a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": " denotes the luminance restoration network (LRNet) and, " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": " represents the color regeneration network (CRNet). " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": " is the shadow-free output, and " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": " serves as a conditioning variable for color regeneration. Additionally, " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": " is the inverse procedure of " + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 556, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "As in Fig. 4, ShadowHack operates in three stages: 1) Decomposition. The RGB image is decomposed via " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " into luminance " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " components; 2) Luminance Restoration. LRNet processes " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " to recover illumination and texture, yield in " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_t = \\mathcal{R}(I_t)" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "; 3) Color Regeneration." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11406" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 298, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 298, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 298, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 298, + 228 + ], + "type": "image", + "image_path": "d21659e7b04d4d660d5c69b80a0ee1a655841de9a3c58fa3b961cd359adf5565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": "Figure 5. The rectified attention module. We compute two sets of attention. In " + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "inline_equation", + "content": "Att_1" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": ", both texture " + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "inline_equation", + "content": "(I_t)" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "inline_equation", + "content": "(I_c)" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": " are involved while in " + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "inline_equation", + "content": "Att_2" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": ", we calculate attention with only color components. The difference " + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "inline_equation", + "content": "Att_1 - \\lambda Att_2" + }, + { + "bbox": [ + 55, + 236, + 298, + 281 + ], + "type": "text", + "content": " delivers more accurate interactions" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "text", + "content": "CRNet harmonizes colors using " + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t}" + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "text", + "content": ". Finally, the estimated luminance and color components are recombined using " + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{-1}(\\cdot)" + }, + { + "bbox": [ + 55, + 286, + 297, + 335 + ], + "type": "text", + "content": " to produce the final shadow-free RGB result. Below, we detail LRNet and CRNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 233, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 233, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 233, + 358 + ], + "type": "text", + "content": "4.1. Luminance Restoration Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 297, + 581 + ], + "type": "text", + "content": "To eliminate shadows in " + }, + { + "bbox": [ + 54, + 365, + 297, + 581 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 54, + 365, + 297, + 581 + ], + "type": "text", + "content": ", enhancing the brightness of the shadow areas and restoring their textures is essential. As illustrated in Fig. 4 (a), the shallow two stages and the refinement module in the luminance restoration network (LRNet) are constructed using a Local Range block (LRB), a Multi-head Transpose Attention (MTA) module, and a Feed-Forward Network (FFN), while a RecTified Attention (RTA) module replaces LRB in the deeper stages. The LRB is designed to extract local texture information through a depth-wise convolution block, capturing fine-grained details. The MTA, which lies in the channel attention block (CAB), following previous works [5, 59], reallocates the weights of channels to enhance the relevant features. The FFN further processes these extracted features to improve texture restoration. Finally, the RTA module, introduced in the deeper layers, captures more global information by focusing on regions close to shadow boundaries, enabling the network to refine luminance restoration in complex shadow regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": "Rectified Attention Module. In the shadow removal process, achieving robust enhancement in shadow regions requires restoration efforts to reference better surrounding, related non-shadow areas. Therefore, we replace the Local Range Block (LRB) with self-attention modules [20, 32] in the deeper layers to capture more distant information. Since the goal of LRNet is to cure shadow-contaminated texture and lightness, the feature from the deep layer (i.e., blocks with SAB equipped) becomes less and less discriminative to shadow. As a result, the attention matrix in deep layers, suffers from a distraction. In the first column of Fig. 6, the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 200 + ], + "type": "image", + "image_path": "b776d461501287bc8fb4a296417a8d390c80c3f039f2155fb3ff5f8133c8f233.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "text", + "content": "Figure 6. An example of our rectified attention module. The star indicates the position of the query token. We exhibit the attention map from the first head of the encoder, latent, and decoder layers. The final attention is obtained by subtracting " + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}_2" + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}_1" + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "text", + "content": ". The attention matrix without rectify, " + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}_1" + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{Att}_2" + }, + { + "bbox": [ + 313, + 208, + 556, + 275 + ], + "type": "text", + "content": " are normalized to [0, 1], while their differential result is in [-1, 1]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": "model tends to focus on only a limited number of tokens which may be inaccurate for shadow area refinement. We address this via a RecTified Attention (RTA) mechanism to rectify the original attention weight that is re-injecting contaminated chrominance channels " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": ", by combining them into " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": " calculations (Please notice that the " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": " is still clear without chromance), resulting in " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "Att_{1}" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": ". Since this operation re-integrates color bias to the calculation, inspired by [58], we calculate a second set with only " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "inline_equation", + "content": "Att_{2}" + }, + { + "bbox": [ + 313, + 282, + 556, + 413 + ], + "type": "text", + "content": ". By subtracting the two sets, LRNet has the flexibility to get rid of chrominance information if needed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "content": "To be specific, we obtain the features " + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "content": " through convolution with downsampling, and compute the softmax attention map for " + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 313, + 415, + 556, + 462 + ], + "type": "text", + "content": ", as shown in Fig. 5. The query, key, and value matrices are formed via:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 468, + 554, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 468, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 344, + 468, + 554, + 489 + ], + "type": "interline_equation", + "content": "Q _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {Q}, K _ {1} = \\left[ F _ {t}; F _ {c} \\right] W _ {1} ^ {K}, \\tag {8}", + "image_path": "8a79844f5867d1ed4743e51b0da475422e99bcc4048320b447b78bd5a2058139.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 486, + 523, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 486, + 523, + 500 + ], + "spans": [ + { + "bbox": [ + 345, + 486, + 523, + 500 + ], + "type": "interline_equation", + "content": "Q _ {2} = F _ {c} W _ {2} ^ {Q}, K _ {2} = F _ {c} W _ {2} ^ {K}, V = F _ {t} W ^ {V},", + "image_path": "b3ed5ec2633e99c0353ac952bd4fb3d2b3ce96e84258ad9fb9d6cb572c7f876c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "content": " matrices are learnable projection parameters and \"; refers to the concatenation of the two features. Next, the above two sets of queries and keys are used to compute their respective attentions simply by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 559, + 555, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 559, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 342, + 559, + 555, + 587 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t} _ {i} = \\operatorname {S o f t m a x} \\left(\\frac {Q _ {i} K _ {i} ^ {\\mathrm {T}}}{\\sqrt {d}} + B\\right), i \\in \\{1, 2 \\}, \\tag {9}", + "image_path": "bbbcfa3b26411c1f6b996114f09241e9e3a6e6eb4304f70e5acc41c972083499.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 591, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 591, + 555, + 616 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 591, + 555, + 616 + ], + "type": "text", + "content": " denotes the relative position bias. This allows us to apply the rectified attention mechanism as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 622, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 622, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 362, + 622, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\operatorname {R T A} \\left(F _ {t}, F _ {c}\\right) = \\left(\\operatorname {A t t} _ {1} - \\lambda \\cdot \\operatorname {A t t} _ {2}\\right) V, \\tag {10}", + "image_path": "a392418e21b561b7a8f51fdf80611f30f96e8cab3104ff2867f19c8d9a9a69d9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "text", + "content": " is a reweighting factor. As illustrated in Fig. 6, such a rectified attention mechanism creates a strong affinity with well-lit reference, urging the model to focus on relevant content. As validated in Sec. 5.3, our rectified attention can focus on more non-shadow area information to transfer robust clues for shadow removal." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11407" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 336 + ], + "type": "table", + "html": "
DatasetsMethodShadow Region (S)Non-Shadow Region (NS)All Image (ALL)
PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓PSNR↑SSIM↑RMSE↓
ISTD+DHAN [8] (AAAI'20)33.080.9889.4927.280.9727.3925.780.9587.74
G2R [36] (CVPR'21)33.880.9788.7135.940.9772.8130.850.9463.78
BMNet [65] (CVPR'22)38.170.9915.7237.950.9862.4234.340.9742.93
SG-ShadowNet [47] (ECCV'22)36.800.9905.9335.570.9782.9232.460.9623.41
ShadowFormer [13] (AAAI'23)39.670.9925.2138.820.9832.3035.460.9732.80
ShadowDiffusion [14] (CVPR'23)39.82-4.9038.90-2.3035.72-2.70
Li et al. [30] (ICCV'23)38.460.9895.9337.270.9772.9034.140.9603.39
Liu et al. [34] (AAAI'24)38.040.9905.6939.150.9842.3134.960.9682.87
RASM [32] (MM'24)40.730.9934.4139.230.9852.1736.160.9762.53
Homofformer [53] (CVPR'24)39.470.9934.7238.730.9842.2335.340.9752.64
OmniSR+GM [55] (AAAI'25)------34.200.973-
ShadowHack (Ours)40.560.9934.4639.660.9872.0936.310.9772.48
SRDDHAN [8] (AAAI'20)33.670.9788.9434.790.9794.8030.510.9495.67
BMNet [65] (CVPR'22)35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNet [47] (ECCV'22)--7.53--2.97--4.23
ShadowFormer [13] (AAAI'23)36.910.9895.9036.220.9893.4432.900.9584.04
ShadowDiffusion [14] (CVPR'23)38.720.9874.9837.780.9853.4434.730.9703.63
Li et al. [30] (ICCV'23)39.330.9856.0935.610.9672.9733.170.9403.83
Liu et al. [34] (AAAI'24)36.510.9835.4937.710.9863.0033.480.9673.66
DeS3 [24] (AAAI'24)37.450.9845.8838.120.9882.8334.110.9683.72
RASM [32] (MM'24)37.910.9885.0238.700.9922.7234.460.9763.37
Homofformer [53] (CVPR'24)38.810.9874.2539.450.9882.8535.370.9723.33
OmniSR+GM [55] (AAAI'25)------34.560.977-
ShadowHack (Ours)39.470.9914.1840.060.9942.3835.940.9822.90
", + "image_path": "62d0813550cacd0b8a947580368a6a27d8866ca560854e2ecc8e239adb1ab5a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 340, + 545, + 351 + ], + "lines": [ + { + "bbox": [ + 63, + 340, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 63, + 340, + 545, + 351 + ], + "type": "text", + "content": "Table 1. Quantitative results on the ISTD+ and SRD datasets. The best results are in bold, while the second-best ones are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 367, + 214, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 214, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 214, + 380 + ], + "type": "text", + "content": "4.2. Color Regeneration Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 387, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 297, + 662 + ], + "type": "text", + "content": "After restoring texture and lighting information in the luminance channel, the next task is to accurately regenerate color. Inspired by exemplar-based colorization models [23, 29, 39], we leverage contextual cues and condition-based guidance to improve color fidelity and ensure alignment with the semantic content of the images. Concretely, we propose a dual-encoder network, termed CRNet, which integrates a multiscale color feature extractor into the U-shaped architecture of LRNet. The color encoder is a ConvNext-v2utto model [37] pre-trained on ImageNet-21k with 2M parameters, which takes the Cb and Cr channels as input and extracts color features across multiple scales, as illustrated in Fig. 4 (b). To achieve precise color restoration, the network shall reference regions with similar/identical material properties and accurate color, using them as reliable guides for color regeneration. The outputs generated by LRNet provide qualified indexing information to facilitate this purpose. Therefore, we introduce an attention-based mechanism where luminance features serve as the query and key to compute similarity, while color features act as the value, aggregating essential color information. This approach injects color features into the skip connections of the U-shaped backbone, remarkably enriching the network's color regeneration capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Multiple Checkpoint Training. We observe that the luminance network often yields better performance on the training set, which suggests the color regeneration network faces challenges in adapting to imperfect luminance out" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 556, + 452 + ], + "type": "text", + "content": "puts encountered during testing. To address this issue, we employ a simple yet effective way to overcome this issue, where multiple early-stage checkpoints from the luminance network training are randomly selected as candidates. This technique, applied only during training, strengthens the robustness of the color network without incurring additional computational costs at inference time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 460, + 455, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 455, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 455, + 475 + ], + "type": "text", + "content": "5. Experimental Validation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 432, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 432, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 432, + 494 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "content": "Implementation details. Our models are implemented in PyTorch, and the entire training process utilizes the AdamW optimizer [38] with momentum set to (0.9, 0.999) and a weight decay of " + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "content": ". The initial learning rate is set to " + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "content": ". During training, we randomly crop the training data into patches of size " + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "content": ". For the base network of each stage, we use a four-layer encoder-decoder structure (from L1 to L4), consisting of three downsampling and three upsampling operations. The dimension of the first layer is 32. Our RTA modules are integrated into L3 and L4, with an overlapping ratio set to 0.5 and a rectified mechanism " + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\lambda_0" + }, + { + "bbox": [ + 313, + 498, + 555, + 677 + ], + "type": "text", + "content": " value of 0.7. For training, we use L1 and VGG loss terms as employed in previous works [32, 34]. For detailed information on the implementation of each stage and training setup, please refer to the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Datasets. We conduct our experiments on the following benchmark datasets for shadow removal: 1) Adjusted ISTD (ISTD+) dataset [27], which refines the ISTD dataset [48]" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11408" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 121, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 121, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 121, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 121, + 150 + ], + "type": "image", + "image_path": "f37dc20ab13f136ec8a4e765509da6d2bc32e6352818be0a953c6938d0ef5881.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 152, + 105, + 162 + ], + "lines": [ + { + "bbox": [ + 76, + 152, + 105, + 162 + ], + "spans": [ + { + "bbox": [ + 76, + 152, + 105, + 162 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 70, + 184, + 150 + ], + "blocks": [ + { + "bbox": [ + 123, + 70, + 184, + 150 + ], + "lines": [ + { + "bbox": [ + 123, + 70, + 184, + 150 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 184, + 150 + ], + "type": "image", + "image_path": "38461c8262550ce788585328fe55192401919a22ea5397c1b931647994569d90.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 152, + 177, + 162 + ], + "lines": [ + { + "bbox": [ + 125, + 152, + 177, + 162 + ], + "spans": [ + { + "bbox": [ + 125, + 152, + 177, + 162 + ], + "type": "text", + "content": "(b) BMNet [65]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 244, + 150 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 244, + 150 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 244, + 150 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 244, + 150 + ], + "type": "image", + "image_path": "ee2b8fb14f897d8daadf792f0865955b25926c4636d26b4d45620a5d1d08fd1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 152, + 231, + 162 + ], + "lines": [ + { + "bbox": [ + 194, + 152, + 231, + 162 + ], + "spans": [ + { + "bbox": [ + 194, + 152, + 231, + 162 + ], + "type": "text", + "content": "(c) SG [47]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 245, + 70, + 305, + 150 + ], + "blocks": [ + { + "bbox": [ + 245, + 70, + 305, + 150 + ], + "lines": [ + { + "bbox": [ + 245, + 70, + 305, + 150 + ], + "spans": [ + { + "bbox": [ + 245, + 70, + 305, + 150 + ], + "type": "image", + "image_path": "4faf09f58d3a56a94f4fdcf7105442932506e3e0f65d83d7b826d20ee6d948b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 152, + 291, + 162 + ], + "lines": [ + { + "bbox": [ + 255, + 152, + 291, + 162 + ], + "spans": [ + { + "bbox": [ + 255, + 152, + 291, + 162 + ], + "type": "text", + "content": "(d) SF [13]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 365, + 150 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 365, + 150 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 365, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 365, + 150 + ], + "type": "image", + "image_path": "b7c77ea1bc50baad530eb6a49cb24b724358bc9538983d25c1f5b8795db5fea9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 152, + 353, + 162 + ], + "lines": [ + { + "bbox": [ + 315, + 152, + 353, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 353, + 162 + ], + "type": "text", + "content": "(e) HF [53]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 366, + 70, + 426, + 150 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 426, + 150 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 426, + 150 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 426, + 150 + ], + "type": "image", + "image_path": "b98aed533779497ca16e14fe5cec61d7bda9d81cd4718725c57641e5fe7d07f2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 152, + 419, + 162 + ], + "lines": [ + { + "bbox": [ + 371, + 152, + 419, + 162 + ], + "spans": [ + { + "bbox": [ + 371, + 152, + 419, + 162 + ], + "type": "text", + "content": "(f) RASM [32]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 488, + 150 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 488, + 150 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 488, + 150 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 488, + 150 + ], + "type": "image", + "image_path": "55218b3e115f37398494d61e6b0b5b564ce2fe972f5bc4b1c7bb44afc06877f2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 152, + 470, + 162 + ], + "lines": [ + { + "bbox": [ + 443, + 152, + 470, + 162 + ], + "spans": [ + { + "bbox": [ + 443, + 152, + 470, + 162 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 489, + 70, + 548, + 150 + ], + "blocks": [ + { + "bbox": [ + 489, + 70, + 548, + 150 + ], + "lines": [ + { + "bbox": [ + 489, + 70, + 548, + 150 + ], + "spans": [ + { + "bbox": [ + 489, + 70, + 548, + 150 + ], + "type": "image", + "image_path": "6282831f4b3114f3622fd3f992e37aa01ba49b160feb674bd24de4b07d445c9d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 152, + 529, + 162 + ], + "lines": [ + { + "bbox": [ + 505, + 152, + 529, + 162 + ], + "spans": [ + { + "bbox": [ + 505, + 152, + 529, + 162 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 61, + 184, + 122, + 262 + ], + "blocks": [ + { + "bbox": [ + 61, + 184, + 122, + 262 + ], + "lines": [ + { + "bbox": [ + 61, + 184, + 122, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 122, + 262 + ], + "type": "image", + "image_path": "d528eeb5c62f287eb153184cf74fb402d938a8d6c3b73fb05e01ddc364e6add3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 265, + 105, + 274 + ], + "lines": [ + { + "bbox": [ + 76, + 265, + 105, + 274 + ], + "spans": [ + { + "bbox": [ + 76, + 265, + 105, + 274 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 123, + 184, + 183, + 262 + ], + "blocks": [ + { + "bbox": [ + 123, + 184, + 183, + 262 + ], + "lines": [ + { + "bbox": [ + 123, + 184, + 183, + 262 + ], + "spans": [ + { + "bbox": [ + 123, + 184, + 183, + 262 + ], + "type": "image", + "image_path": "35696371e640e1c04de01e03304e5a359a2d7b046b26bcb47b15a8255c03fba6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 264, + 177, + 274 + ], + "lines": [ + { + "bbox": [ + 126, + 264, + 177, + 274 + ], + "spans": [ + { + "bbox": [ + 126, + 264, + 177, + 274 + ], + "type": "text", + "content": "(b) BMNet [65]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 184, + 184, + 244, + 262 + ], + "blocks": [ + { + "bbox": [ + 184, + 167, + 424, + 178 + ], + "lines": [ + { + "bbox": [ + 184, + 167, + 424, + 178 + ], + "spans": [ + { + "bbox": [ + 184, + 167, + 424, + 178 + ], + "type": "text", + "content": "Figure 7. Visual comparison on a sample from the " + }, + { + "bbox": [ + 184, + 167, + 424, + 178 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 184, + 167, + 424, + 178 + ], + "type": "text", + "content": " dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 184, + 244, + 262 + ], + "lines": [ + { + "bbox": [ + 184, + 184, + 244, + 262 + ], + "spans": [ + { + "bbox": [ + 184, + 184, + 244, + 262 + ], + "type": "image", + "image_path": "e01c5c66741d828e27dc00b9be5b3471555d30ed2f3609827ad098366862834c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 264, + 231, + 274 + ], + "lines": [ + { + "bbox": [ + 194, + 264, + 231, + 274 + ], + "spans": [ + { + "bbox": [ + 194, + 264, + 231, + 274 + ], + "type": "text", + "content": "(c) SG [47]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 245, + 184, + 304, + 262 + ], + "blocks": [ + { + "bbox": [ + 245, + 184, + 304, + 262 + ], + "lines": [ + { + "bbox": [ + 245, + 184, + 304, + 262 + ], + "spans": [ + { + "bbox": [ + 245, + 184, + 304, + 262 + ], + "type": "image", + "image_path": "5b3c3d3751b4ab75ba8d7c82a2747fc50a8eaee09d463f3e0b563466f60a56c0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 264, + 299, + 274 + ], + "lines": [ + { + "bbox": [ + 247, + 264, + 299, + 274 + ], + "spans": [ + { + "bbox": [ + 247, + 264, + 299, + 274 + ], + "type": "text", + "content": "(d) DMTN [33]" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 306, + 184, + 365, + 262 + ], + "blocks": [ + { + "bbox": [ + 306, + 184, + 365, + 262 + ], + "lines": [ + { + "bbox": [ + 306, + 184, + 365, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 365, + 262 + ], + "type": "image", + "image_path": "f5a8f8c7e679f52a9d15485a87be9bb71a004a3aaf63737a5bcaecf264773fae.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 264, + 356, + 274 + ], + "lines": [ + { + "bbox": [ + 312, + 264, + 356, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 356, + 274 + ], + "type": "text", + "content": "(e) DeS3 [24]" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 367, + 184, + 427, + 262 + ], + "blocks": [ + { + "bbox": [ + 367, + 184, + 427, + 262 + ], + "lines": [ + { + "bbox": [ + 367, + 184, + 427, + 262 + ], + "spans": [ + { + "bbox": [ + 367, + 184, + 427, + 262 + ], + "type": "image", + "image_path": "3b1b901d86cb562b501a1a42add838f85c882f88bb9e97103d34f213759541c3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 264, + 413, + 274 + ], + "lines": [ + { + "bbox": [ + 377, + 264, + 413, + 274 + ], + "spans": [ + { + "bbox": [ + 377, + 264, + 413, + 274 + ], + "type": "text", + "content": "(f) HF [53]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 427, + 184, + 487, + 262 + ], + "blocks": [ + { + "bbox": [ + 427, + 184, + 487, + 262 + ], + "lines": [ + { + "bbox": [ + 427, + 184, + 487, + 262 + ], + "spans": [ + { + "bbox": [ + 427, + 184, + 487, + 262 + ], + "type": "image", + "image_path": "5434a03aec5755c8c6d1aff0929e048b27f43b9bc62d2b8765be8d21b40969ee.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 264, + 470, + 274 + ], + "lines": [ + { + "bbox": [ + 443, + 264, + 470, + 274 + ], + "spans": [ + { + "bbox": [ + 443, + 264, + 470, + 274 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 489, + 184, + 549, + 262 + ], + "blocks": [ + { + "bbox": [ + 489, + 184, + 549, + 262 + ], + "lines": [ + { + "bbox": [ + 489, + 184, + 549, + 262 + ], + "spans": [ + { + "bbox": [ + 489, + 184, + 549, + 262 + ], + "type": "image", + "image_path": "1efbff3d2ff6e82cc67f06b84527da8df2878520b9b6bb7a8fd343d6f04a5b75.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 264, + 528, + 274 + ], + "lines": [ + { + "bbox": [ + 506, + 264, + 528, + 274 + ], + "spans": [ + { + "bbox": [ + 506, + 264, + 528, + 274 + ], + "type": "text", + "content": "(h) GT" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 58, + 304, + 295, + 346 + ], + "blocks": [ + { + "bbox": [ + 58, + 304, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 304, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 304, + 295, + 346 + ], + "type": "table", + "html": "
MetricRGB InputYCbCr InputRetinex-basedOurs
PSNR↑36.1635.6235.8536.31
RMSE↓2.532.622.632.48
", + "image_path": "7eb3515a06431e012162bc8850ba940723b5a37493ea99107ce8077c346f503a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 373 + ], + "type": "text", + "content": "Table 2. Ablation study of the proposed decoupling framework on the ISTD+ dataset." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 386, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 505 + ], + "type": "text", + "content": "by correcting the illumination discrepancies between the ground-truth and input images, includes 1330 training and 540 testing triplets (shadow image, shadow-free image, and shadow mask); 2) SRD dataset [43] consists of 2680 training and 408 testing pairs. We adopt the predicted masks provided by DHAN [8] for training and testing following previous works [32, 53, 65]; 3) UIUC dataset [16] provides 76 images with shadow mask for testing; and 4) UCF dataset [64] provided by Jin et al. [24] contains shadow images without shadow-free references." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 506, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 626 + ], + "type": "text", + "content": "Evaluation metrics. To evaluate the performance of our proposed framework, we utilize the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Root Mean Square Error (RMSE). PSNR and SSIM are both calculated in the RGB space, with higher values indicating better results. For RMSE, which is calculated in the LAB color space, the lower values indicate better results. To be consistent with previous works [32, 53, 65], we inference the full-resolution image and then evaluate the images at a resolution of " + }, + { + "bbox": [ + 55, + 506, + 296, + 626 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 506, + 296, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 635, + 211, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 211, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 211, + 647 + ], + "type": "text", + "content": "5.2. Comparison with Prior Arts" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "We compare our method with the state-of-the-art shadow removal approaches on benchmark datasets, including DHAN [8], G2R [36], BMNet [65], SG-ShadowNet [47], ShadowFormer [13], ShadowDiffusion [14], Li et al. [30], Liu et al. [34], DeS3 [24], Homoformer [53], RASM [32]" + } + ] + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 317, + 304, + 375, + 363 + ], + "blocks": [ + { + "bbox": [ + 188, + 278, + 420, + 290 + ], + "lines": [ + { + "bbox": [ + 188, + 278, + 420, + 290 + ], + "spans": [ + { + "bbox": [ + 188, + 278, + 420, + 290 + ], + "type": "text", + "content": "Figure 8. Visual comparison on a sample from the SRD dataset." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 304, + 375, + 363 + ], + "lines": [ + { + "bbox": [ + 317, + 304, + 375, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 375, + 363 + ], + "type": "image", + "image_path": "a26935eb81525fa360f2b4d877dcb2328eeb68f2080d82f581234a5e694066c8.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 365, + 360, + 375 + ], + "lines": [ + { + "bbox": [ + 331, + 365, + 360, + 375 + ], + "spans": [ + { + "bbox": [ + 331, + 365, + 360, + 375 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 380, + 554, + 402 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 402 + ], + "type": "text", + "content": "Figure 9. Ablation study on rectified attention. The results are displayed in the luminance space." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 376, + 304, + 434, + 363 + ], + "blocks": [ + { + "bbox": [ + 376, + 304, + 434, + 363 + ], + "lines": [ + { + "bbox": [ + 376, + 304, + 434, + 363 + ], + "spans": [ + { + "bbox": [ + 376, + 304, + 434, + 363 + ], + "type": "image", + "image_path": "049b438133538ce0dcba1405a127464561a4b5c0848076ef30153af7a15d3d19.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 365, + 429, + 375 + ], + "lines": [ + { + "bbox": [ + 381, + 365, + 429, + 375 + ], + "spans": [ + { + "bbox": [ + 381, + 365, + 429, + 375 + ], + "type": "text", + "content": "(b) RGB Input" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 435, + 304, + 493, + 363 + ], + "blocks": [ + { + "bbox": [ + 435, + 304, + 493, + 363 + ], + "lines": [ + { + "bbox": [ + 435, + 304, + 493, + 363 + ], + "spans": [ + { + "bbox": [ + 435, + 304, + 493, + 363 + ], + "type": "image", + "image_path": "2d02916a5f0aa03a8de21bba3c1d49a866486b11c0e458eb8b3e99409fce5bda.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 365, + 487, + 375 + ], + "lines": [ + { + "bbox": [ + 441, + 365, + 487, + 375 + ], + "spans": [ + { + "bbox": [ + 441, + 365, + 487, + 375 + ], + "type": "text", + "content": "(c) w/o rectify" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 494, + 304, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 494, + 304, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 494, + 304, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 494, + 304, + 553, + 363 + ], + "type": "image", + "image_path": "80c2e9c2bec6ce90bc368a60eb8796fb9ce8f43c6dac4ebcd090640bd4b636cd.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 365, + 537, + 375 + ], + "lines": [ + { + "bbox": [ + 509, + 365, + 537, + 375 + ], + "spans": [ + { + "bbox": [ + 509, + 365, + 537, + 375 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "content": "and OmniSR+GM [55] (GM means the version with shadow mask), as shown in Tab. 1. For a fair comparison, we use results published by the authors. Regarding missing data, we utilize available images or checkpoints provided by the authors and evaluate them via the same metrics." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Our proposed method achieves the highest scores in most metrics on the ISTD+ dataset compared to other state-of-the-art methods. Specifically, our model attains a PSNR of 36.31 for all image regions, surpassing the second-highest result of 36.16 by 0.15. The high PSNR highlights our model's ability to recover texture details effectively, demonstrating superior reconstruction quality. Additionally, the RMSE value of 2.48 is the lowest, which indicates that our recoloring module plays a crucial role in accurately restoring color information and reducing errors across the entire image. From the magnified areas in Figs. 7 and 8, it can be observed that our method not only effectively corrects color bias but also provides superior edge correction compared to other methods. The second part of Tab. 1 presents a comparison of our results on the SRD dataset against other state-of-the-art methods. It can be observed that our method achieves a significant improvement in RMSE (-0.43), further demonstrating the substantial benefit of the decoupling approach for color recovery." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11409" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 295, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 295, + 167 + ], + "type": "table", + "html": "
MethodShadow PSNR ↑Non-Shadow PSNR ↑All PSNR ↑
[Ft]39.9641.2736.60
[Ft; Fc]40.1741.2936.76
[Ft]&[Ft]40.0241.1836.58
[Ft; Fc]&[Ft; Fc]40.3541.3736.87
[Ft; Fc]&[Fc]40.3641.4036.90
", + "image_path": "994225a5e08c4ceb92284091c0edad6794da017c42ce13a2c9ecc345309783ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 104, + 312 + ], + "blocks": [ + { + "bbox": [ + 58, + 201, + 104, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 104, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 104, + 312 + ], + "type": "image", + "image_path": "660c8b6c69f8a13b2ab8bdbce36d0fcf7f5d64a4b2dc20a6045fe1cb4a34b2d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 314, + 94, + 324 + ], + "lines": [ + { + "bbox": [ + 65, + 314, + 94, + 324 + ], + "spans": [ + { + "bbox": [ + 65, + 314, + 94, + 324 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 105, + 201, + 152, + 312 + ], + "blocks": [ + { + "bbox": [ + 105, + 201, + 152, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 152, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 152, + 312 + ], + "type": "image", + "image_path": "d372ecc92047120dceb1a5e38398083872bf22326e754e8c86e1eb42befd0790.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 314, + 146, + 324 + ], + "lines": [ + { + "bbox": [ + 108, + 314, + 146, + 324 + ], + "spans": [ + { + "bbox": [ + 108, + 314, + 146, + 324 + ], + "type": "text", + "content": "(b) SG [47]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 153, + 201, + 199, + 312 + ], + "blocks": [ + { + "bbox": [ + 153, + 201, + 199, + 312 + ], + "lines": [ + { + "bbox": [ + 153, + 201, + 199, + 312 + ], + "spans": [ + { + "bbox": [ + 153, + 201, + 199, + 312 + ], + "type": "image", + "image_path": "c8f5ad926afbd21af8d9e7c10cfaf1ecaa1d4a9aa0326af6ce87e054a392a0b5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 314, + 193, + 324 + ], + "lines": [ + { + "bbox": [ + 156, + 314, + 193, + 324 + ], + "spans": [ + { + "bbox": [ + 156, + 314, + 193, + 324 + ], + "type": "text", + "content": "(c) SF [13]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 201, + 246, + 312 + ], + "blocks": [ + { + "bbox": [ + 200, + 201, + 246, + 312 + ], + "lines": [ + { + "bbox": [ + 200, + 201, + 246, + 312 + ], + "spans": [ + { + "bbox": [ + 200, + 201, + 246, + 312 + ], + "type": "image", + "image_path": "2d3c0c83baef97b872903b7eeab7379ad3b0f0bdcba79be90af533bc8c764f62.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 314, + 241, + 324 + ], + "lines": [ + { + "bbox": [ + 203, + 314, + 241, + 324 + ], + "spans": [ + { + "bbox": [ + 203, + 314, + 241, + 324 + ], + "type": "text", + "content": "(d) HF [53]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 247, + 201, + 294, + 312 + ], + "blocks": [ + { + "bbox": [ + 247, + 201, + 294, + 312 + ], + "lines": [ + { + "bbox": [ + 247, + 201, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 247, + 201, + 294, + 312 + ], + "type": "image", + "image_path": "66369a419170a8557032737d5f81ce7599e47d9b7f1f2b045be7758c0d41a4bd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 314, + 283, + 324 + ], + "lines": [ + { + "bbox": [ + 255, + 314, + 283, + 324 + ], + "spans": [ + { + "bbox": [ + 255, + 314, + 283, + 324 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 328, + 289, + 340 + ], + "lines": [ + { + "bbox": [ + 59, + 328, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 59, + 328, + 289, + 340 + ], + "type": "text", + "content": "Figure 10. Visual comparisons on the UIUC and UCF datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "content": "Effect of the decoupling framework. To validate the effectiveness of our decoupling framework, we fine-tune our baseline model structure to adapt to RGB and YCbCr input/output for end-to-end shadow removal and compare its performance with our decoupling framework. As shown in Tab. 2, our decoupling approach significantly improves both PSNR and RMSE. Moreover, based on Retinex decomposition, we separately recover L and R, and compare it with our decoupling method. Detailed architectures are provided in the supplementary material. This improvement indicates that this strategy benefits texture recovery and color correction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "Effect of rectified attention module. We demonstrate the effectiveness of the rectified attention module in Tab. 3, by conducting experiments on rectification mechanisms. Furthermore, extra experiments are conducted on different feature combinations of rectified approaches to explore more suitable configurations. " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[X]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[Y]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " in Tab. 3 means performing a subtraction calculation of corresponding maps for rectification between " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[X]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " alone denotes only one pair of " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " is used, where both inputs are " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ". Clearly, leveraging color information, even with inherent bias, can allow shadow regions to better reference surrounding areas, resulting in significant performance improvements. As for potential color bias, rectification refines the attention map for greater precision. Furthermore, we find that " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[F_t; F_c] \\& [F_c]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": " slightly outperforms " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[F_t; F_c] \\& [F_t; F_c]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ", and thus we ultimately choose " + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "inline_equation", + "content": "[F_t; F_c] \\& [F_c]" + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": ". Fig. 9 visualizes the effectiveness of rectification, which has a more precise result with less edge remaining." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 194 + ], + "type": "text", + "content": "Table 3. Ablation study of the proposed rectified attention module on the SRD dataset. We calculate PSNRs on the luminance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 140 + ], + "type": "table", + "html": "
MethodShadow RMSE ↓Non-Shadow RMSE ↓All RMSE ↓
w/o cross-att.5.782.643.39
w/o multi checkpt.5.012.593.28
Ours4.182.382.90
", + "image_path": "8d8d0dc622090f4aaed4fbaad745f01ceebecd55357b6c4431d136e7c27a5e53.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 141, + 539, + 152 + ], + "lines": [ + { + "bbox": [ + 328, + 141, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 328, + 141, + 539, + 152 + ], + "type": "text", + "content": "Table 4. Ablation study of the CRNet on the SRD dataset." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 157, + 375, + 197 + ], + "blocks": [ + { + "bbox": [ + 316, + 157, + 375, + 197 + ], + "lines": [ + { + "bbox": [ + 316, + 157, + 375, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 157, + 375, + 197 + ], + "type": "image", + "image_path": "9e4f5100a07a7580a95cde3e5005b6eb7de21db683f26ef18cd16c6fa2e0e0c6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 199, + 369, + 209 + ], + "lines": [ + { + "bbox": [ + 321, + 199, + 369, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 199, + 369, + 209 + ], + "type": "text", + "content": "(a) dirty mask" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 376, + 157, + 434, + 197 + ], + "blocks": [ + { + "bbox": [ + 376, + 157, + 434, + 197 + ], + "lines": [ + { + "bbox": [ + 376, + 157, + 434, + 197 + ], + "spans": [ + { + "bbox": [ + 376, + 157, + 434, + 197 + ], + "type": "image", + "image_path": "f663130dbcec884f9580ae68d11e253a7b5d8676b2dfd9a0683e27955168d0dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 554, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 554, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 554, + 244 + ], + "type": "text", + "content": "Figure 11. Robustness analysis under corrupted mask. The accurate mask is highlighted in red. For fairness, we only involve mask refinement network here, and it is not involved in any comparison." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 435, + 157, + 493, + 197 + ], + "blocks": [ + { + "bbox": [ + 435, + 157, + 493, + 197 + ], + "lines": [ + { + "bbox": [ + 435, + 157, + 493, + 197 + ], + "spans": [ + { + "bbox": [ + 435, + 157, + 493, + 197 + ], + "type": "image", + "image_path": "125cdfd79baba77ff3c038f7d95951382cace7ac25bf281c2f8e2522b1ad9138.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 199, + 493, + 209 + ], + "lines": [ + { + "bbox": [ + 378, + 199, + 493, + 209 + ], + "spans": [ + { + "bbox": [ + 378, + 199, + 493, + 209 + ], + "type": "text", + "content": "(b) refined mask (c) w/o refinement" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 494, + 157, + 552, + 197 + ], + "blocks": [ + { + "bbox": [ + 494, + 157, + 552, + 197 + ], + "lines": [ + { + "bbox": [ + 494, + 157, + 552, + 197 + ], + "spans": [ + { + "bbox": [ + 494, + 157, + 552, + 197 + ], + "type": "image", + "image_path": "f8bd731dc0ee73f9930104a2041d3114a066b98006a5530a43e9bc458b54b2a6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 199, + 551, + 209 + ], + "lines": [ + { + "bbox": [ + 496, + 199, + 551, + 209 + ], + "spans": [ + { + "bbox": [ + 496, + 199, + 551, + 209 + ], + "type": "text", + "content": "(d) w refinement" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 361 + ], + "type": "text", + "content": "Effect of the color restoration network. To validate the effectiveness of the key component and design in the CRNet, we replace the cross-attention-based color feature injection with concatenation along the feature channel. As demonstrated in Tab. 4, if cross-attention is replaced, the accuracy has dropped by a large margin, indicating the effectiveness of the cross-attention. Furthermore, the model suffers from a misalignment between training and testing luminance results if no previous checkpoints are involved." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 367, + 432, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 432, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 432, + 380 + ], + "type": "text", + "content": "5.4. Robustness Analysis" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": "Generalization to other datasets. To demonstrate the generalization capability of ShadowHack, we evaluate it on the UIUC [16] and UCF [64] datasets with our " + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": " pretrained model. The results are presented in Fig. 10, where ShadowHack achieves the cleanest shadow removal with minimal color bias. More comparisons are available in the supplementary material." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "content": "Robustness to inaccurate masks. To get rid of the performance degradation brought by inaccurate masks, we propose a mask refine network, which corrects the mask for better generalization. After our remediation, the result is recovered from inaccurate enlightening, as in Fig. 11 shows. More details are available in the supplementary material." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "This work proposed a novel framework, ShadowHack, which decouples images into luminance and color components, enabling targeted processing for more effective shadow removal. This approach can well address limitations in existing end-to-end models by enhancing the use of color information to improve both texture and luminance recovery. To further boost shadow removal, a rectified attention mechanism has been designed to boost the reference ability of self-attention. Additionally, the color regeneration network has demonstrated its critical role in accurately correcting image colors. Extensive experimental results have underscored the superior performance and robustness of our model." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11410" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "type": "text", + "content": "[1] Harry Barrow, J Tenenbaum, A Hanson, and E Riseman. Recovering intrinsic scene characteristics. Comput. vis. syst, 2(3-26):2, 1978. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 296, + 168 + ], + "type": "text", + "content": "[2] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In ICCV, pages 12470-12479, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "text", + "content": "[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9630-9640, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[4] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[5] Xiangyu Chen, Zheyuan Li, Yuandong Pu, Yihao Liu, Jiantao Zhou, Yu Qiao, and Chao Dong. A comparative study of image restoration networks for general backbone network design. arXiv preprint arXiv:2310.11881, 2023. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "text", + "content": "[6] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super-resolution transformer. In CVPR, pages 22367-22377, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 367 + ], + "type": "text", + "content": "[7] Michael E Churma. Blue shadows: physical, physiological, and psychological causes. Applied optics, 33(21):4719-4722, 1994. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 412 + ], + "type": "text", + "content": "[8] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting GAN. In AAAI, pages 10680-10687, 2020. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 479 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 296, + 512 + ], + "type": "text", + "content": "[10] Graham D. Finlayson, Steven D. Hordley, and Mark S. Drew. Removing shadows from images. In ECCV (4), pages 823-836, 2002. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 296, + 555 + ], + "type": "text", + "content": "[11] Graham D. Finlayson, Steven D. Hordley, Cheng Lu, and Mark S. Drew. On the removal of shadows from images. IEEE Trans. Pattern Anal. Mach. Intell., 28(1):59-68, 2006. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "type": "text", + "content": "[12] Graham D. Finlayson, Mark S. Drew, and Cheng Lu. Entropy minimization for shadow removal. Int. J. Comput. Vis., 85(1): 35-57, 2009. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 296, + 624 + ], + "type": "text", + "content": "[13] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In AAAI, pages 710-718, 2023. 1, 3, 6, 7, 8, 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 669 + ], + "type": "text", + "content": "[14] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In CVPR, pages 14049-14058, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[15] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In ICCV, pages 12999-13008. IEEE, 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[16] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE Trans. Pattern Anal. Mach. Intell., 35(12):2956-2967, 2013. 1, 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 139 + ], + "type": "text", + "content": "[17] Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness. Int. J. Comput. Vis., 131(1): 48-66, 2023. 1, 3, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "type": "text", + "content": "[18] Xiaojie Guo, Yu Li, and Haibin Ling. LIME: low-light image enhancement via illumination map estimation. IEEE Trans. Image Process., 26(2):982-993, 2017. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[19] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 209, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 241 + ], + "type": "text", + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, pages 6185-6194, 2023. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 285 + ], + "type": "text", + "content": "[21] Jun-Wei Hsieh, Wen-Fong Hu, Chia-Jung Chang, and Yung-Sheng Chen. Shadow elimination for effective moving object detection by gaussian shadow modeling. Image Vis. Comput., 21(6):505-516, 2003. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 287, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 319 + ], + "type": "text", + "content": "[22] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In ICCV, pages 2472–2481, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 321, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 354 + ], + "type": "text", + "content": "[23] Zhitong Huang, Nanxuan Zhao, and Jing Liao. Unicolor: A unified framework for multi-modal colorization with transformer. ACM Trans. Graph., 41(6):205:1-205:16, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "type": "text", + "content": "[24] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T. Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In AAAI, pages 2634-2642, 2024. 1, 6, 7, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "text", + "content": "[25] Salman Hameed Khan, Mohammed Bennamoun, Ferdous Ahmed Sohel, and Roberto Togneri. Automatic shadow detection and removal from a single image. IEEE Trans. Pattern Anal. Mach. Intell., 38(3):431-446, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 444, + 554, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 465 + ], + "type": "text", + "content": "[26] Edwin H Land. The retina theory of color vision. Scientific american, 237(6):108-129, 1977. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "type": "text", + "content": "[27] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In ICCV, pages 8577-8586, 2019. 1, 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 533 + ], + "type": "text", + "content": "[28] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9088-9101, 2022. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 535, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 588 + ], + "type": "text", + "content": "[29] Junsoo Lee, Eungyeup Kim, Yunsung Lee, Dongjun Kim, Jaehyuk Chang, and Jaegul Choo. Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence. In CVPR, pages 5800-5809, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 633 + ], + "type": "text", + "content": "[30] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor W. Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In ICCV, pages 13009-13018, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 657 + ], + "type": "text", + "content": "[31] Feng Liu and Michael Gleicher. Texture-consistent shadow removal. In ECCV (4), pages 437-450. Springer, 2008. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "type": "text", + "content": "[32] Hengxing Liu, Mingjia Li, and Xiaojie Guo. Regional attention for shadow removal. In ACM Multimedia, 2024. 1, 3, 5, 6, 7, 2, 4" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[33] Jiawei Liu, Qiang Wang, Huijie Fan, Wentao Li, Liangqiong Qu, and Yandong Tang. A decoupled multi-task network for" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11411" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "text", + "content": "shadow removal. IEEE Trans. Multim., 25:9449-9463, 2023. 7, 1, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[34] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson W. H. Lau. Recasting regional lighting for shadow removal. In AAAI, pages 3810-3818, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 186 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 9992-10002, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 187, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 294, + 220 + ], + "type": "text", + "content": "[36] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In CVPR, pages 4927-4936, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 222, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 255 + ], + "type": "text", + "content": "[37] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11966-11976, 2022. 6, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 257, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 278 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 280, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 313 + ], + "type": "text", + "content": "[39] Peng Lu, Jinbei Yu, Xujun Peng, Zhaoran Zhao, and Xiaojie Wang. Gray2colornet: Transfer more colors from reference image. In ACM Multimedia, pages 3210-3218, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 315, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 335 + ], + "type": "text", + "content": "[40] David K. Lynch. Shadows. Appl. Opt., 54(4):B154-B164, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 380 + ], + "type": "text", + "content": "[41] Kangfu Mei, Luis Figueroa, Zhe Lin, Zhihong Ding, Scott Cohen, and Vishal M Patel. Latent feature-guided diffusion models for shadow removal. In WACV, pages 4313-4322, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "type": "text", + "content": "[42] Sohail Nadimi and Bir Bhanu. Physical models for moving shadow and object detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 26(8):1079-1087, 2004. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "text", + "content": "[43] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson W. H. Lau. Desshadownet: A multi-context embedding deep network for shadow removal. In CVPR, pages 2308-2316, 2017. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 497 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3), pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 531 + ], + "type": "text", + "content": "[45] Yael Shor and Dani Lischinski. The shadow meets the mask: Pyramid-based shadow removal. Comput. Graph. Forum, 27 (2):577-586, 2008. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "type": "text", + "content": "[46] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "type": "text", + "content": "[47] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In ECCV (19), pages 361-378, 2022. 1, 6, 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "text", + "content": "[48] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In CVPR, pages 1788-1797, 2018. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[49] Tianyu Wang, Jianming Zhang, Haitian Zheng, Zhihong Ding, Scott Cohen, Zhe Lin, Wei Xiong, Chi-Wing Fu, Luis Figueroa, and Soo Ye Kim. Metashadow: Object-centered shadow detection, removal, and synthesis. In CVPR, pages 28252-28262, 2025. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[50] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, pages 17662-17672, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[51] Tai-Pang Wu and Chi-Keung Tang. A bayesian approach for shadow extraction from a single image. In ICCV, pages 480–487, 2005. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[52] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, pages 4784-4793, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "text", + "content": "[53] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In CVPR, pages 25617-25626, 2024. 1, 3, 6, 7, 8, 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 273 + ], + "type": "text", + "content": "[54] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. Int. J. Comput. Vis., 132(9):3610-3632, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 275, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 555, + 319 + ], + "type": "text", + "content": "[55] Jiamin Xu, Zelong Li, Yuxin Zheng, Chenyu Huang, Renshu Gu, Weiwei Xu, and Gang Xu. Omnisr: Shadow removal under direct and indirect lighting. arXiv preprint arXiv:2410.01719, 2024. 3, 6, 7, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 320, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 374 + ], + "type": "text", + "content": "[56] Han Yang, Tianyu Wang, Xiaowei Hu, and Chi-Wing Fu. Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12687-12698, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[57] Qingxiong Yang, Kar-Han Tan, and Narendra Ahuja. Shadow removal using bilateral filtering. IEEE Trans. Image Process., 21(10):4361-4368, 2012. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "text", + "content": "[58] Tianzhu Ye, Li Dong, Yuqing Xia, Yutao Sun, Yi Zhu, Gao Huang, and Furu Wei. Differential transformer. arXiv preprint arXiv:2410.05258, 2024. 5, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, pages 5718-5729, 2022. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 533 + ], + "type": "text", + "content": "[60] Ling Zhang, Qing Zhang, and Chunxia Xiao. Shadow remover: Image shadow removal based on illumination recovering optimization. IEEE Trans. Image Process., 24(11): 4623-4636, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "text", + "content": "[61] Ling Zhang, Chengjiang Long, Xiaolong Zhang, and Chunxia Xiao. RIS-GAN: explore residual and illumination with generative adversarial networks for shadow removal. In AAAI, pages 12829-12836, 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "type": "text", + "content": "[62] Wuming Zhang, Xi Zhao, Jean-Marie Morvan, and Liming Chen. Improving shadow suppression for illumination robust face recognition. IEEE Trans. Pattern Anal. Mach. Intell., 41 (3):611-624, 2019. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 668 + ], + "type": "text", + "content": "[63] Wei Zhou, Julie Stephany Berrio, Stewart Worrall, and Eduardo M. Nebot. Automated evaluation of semantic segmentation robustness for autonomous driving. IEEE Trans. Intell. Transp. Syst., 21(5):1951-1963, 2020. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[64] Jiejie Zhu, Kegan G. G. Samuel, Syed Zain Masood, and Marshall F. Tappen. Learning to recognize shadows in monochromatic natural images. In CVPR, pages 223-230, 2010. 7, 8" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11412" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 108 + ], + "type": "text", + "content": "[65] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In CVPR, pages 5617-5626, 2022. 2, 6, 7, 1, 3, 4, 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11413" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_content_list.json b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd2e1ca16a897b1a07310a3d9f5fea475a386910 --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "Shape of Motion: 4D Reconstruction from a Single Video", + "text_level": 1, + "bbox": [ + 209, + 130, + 787, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qianqian Wang $^{1,2*}$ , Vickie Ye $^{1*}$ , Hang Gao $^{1*}$ , Weijia Zeng $^{3*}$ , Jake Austin $^{1}$ , Zhengqi Li $^{4}$ , Angjoo Kanazawa $^{1}$", + "bbox": [ + 256, + 179, + 740, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UC Berkeley $^{2}$ Google DeepMind $^{3}$ UC San Diego $^{4}$ Adobe Research", + "bbox": [ + 196, + 224, + 795, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e4a5d046a457279c9ea0ce38be0ac15921887a07fc6f4ecae3ad61d9e044311.jpg", + "image_caption": [ + "Figure 1. Shape of Motion. Our method enables joint long-range 3D tracking and novel view synthesis from a monocular video of a complex dynamic scene. We render moving elements at a fixed viewpoint across time and visualize estimated 3D motion as colorful trajectories. These trajectories reveal distinct geometric patterns, which leads to the term \"Shape of Motion\"." + ], + "image_footnote": [], + "bbox": [ + 93, + 277, + 495, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3ce98b04f76bf54f69821528a0ff68dd5219d77e9ab4f88e60757429348fef9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 277, + 905, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 497, + 326, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular dynamic reconstruction is a challenging and long-standing vision problem due to the highly ill-posed nature of the task. Existing approaches depend on templates, are effective only in quasi-static scenes, or fail to model 3D motion explicitly. We introduce a method for reconstructing generic dynamic scenes, featuring explicit, persistent 3D motion trajectories in the world coordinate frame, from casually captured monocular videos. We tackle the problem with two key insights: First, we exploit the low-dimensional structure of 3D motion by representing scene motion with a compact set of $\\mathbb{S}\\mathbb{E}(3)$ motion bases. Each point's motion is expressed as a linear combination of these bases, facilitating soft decomposition of the scene into multiple rigidly-moving groups. Second, we take advantage of off-the-shelf data-driven priors such as monocular depth maps and long-range 2D tracks, and devise a method to effectively consolidate these noisy supervisory signals, resulting in a globally consistent representation of the dynamic scene. Experiments show that our method achieves state-of-the-art performance for both long-range 3D/2D motion estimation and novel view synthesis on dynamic scenes. Project Page: https://shape-of-motion.github.io/", + "bbox": [ + 88, + 529, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 497, + 643, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing the persistent geometry and their 3D motion across a video is crucial for understanding and interacting with the underlying physical world. While recent years have seen impressive progress in modeling static 3D scenes [45, 69], recovering the geometry and motion of complex dynamic 3D scenes, especially from a single video, remains an open challenge. A number of prior dynamic reconstruction and novel view synthesis approaches have attempted to tackle this problem. However, most methods rely on synchronized multi-view videos [9, 23, 23, 67, 108] or additional LIDAR/depth sensors [29, 63, 77, 98, 106]. Recent monocular approaches can operate on regular dynamic videos, but they typically model 3D scene motion as short-range scene flow between consecutive times [24, 57, 58] or deformation fields that map between canonical and view space [74, 75, 119], failing to capture 3D motion trajectories persistent over a video.", + "bbox": [ + 511, + 522, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The longstanding challenge for general in-the-wild videos lies in the poorly constrained nature of the 4D reconstruction problem. In this work, we tackle this challenge with two key insights. The first is that, while the image space dynamics can be complex and discontinuous, the underlying 3D motion is a composition of continuous simple rigid motions. Our second insight is that data-driven priors provide complementary, though noisy cues, that aggregate", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 91, + 887, + 200, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9660", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "well into a globally coherent representation of the 3D scene geometry and motion.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these two insights, we represent the dynamic scene as a set of persistent 3D Gaussians, and represent their motion across the video in terms of a compact set of shared $\\mathbb{S}\\mathbb{E}(3)$ motion bases. Unlike traditional scene flow, which computes 3D correspondences between consecutive frames, our representation recovers a persistent 3D trajectory over the whole video, enabling long-range 3D tracking across the entire video. As the 3D trajectories produced by our method capture the geometric patterns that trace each point's movement through 3D space and time as shown in Figure 1, we refer to our approach as \"Shape of Motion\". We show how to fit our explicit scene representation to a general video in-the-wild, by fusing together complementary cues from two main sources: monocular depth estimates per-frame, and 2D track estimates across frames. We conduct extensive evaluations on both synthetic and real-world dynamic video datasets, and show that our proposed approach significantly outperforms prior methods in both long-range 2D/3D tracking and novel view synthesis.", + "bbox": [ + 91, + 122, + 482, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are: (1) A 4D scene representation enabling both real-time novel view synthesis and globally consistent 3D tracking for any point at any time. (2) An optimization framework that fuses learned motion and geometry priors into a unified 4D representation from a single, casually captured monocular video.", + "bbox": [ + 89, + 409, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 512, + 232, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Correspondences and Tracking. While monocular 3D long range tracking remains largely unexplored, numerous approaches track in 2D image space, typically using optical flow for point correspondences. This involves estimating dense motion fields between image pairs [4, 7, 8, 18, 32, 35, 35, 39, 41-43, 66, 79, 87, 93, 94, 112]. Sparse keypoint matching methods can enable long trajectory generation [2, 14, 61, 64, 81], but these methods are primarily intended for sparse 3D reconstruction. Long-range 2D trajectory estimation for arbitrary points has been explored in earlier works, which relied on hand-crafted priors to generate motion trajectories [3, 80, 83, 84, 88, 99]. Recently, there has been a resurgence of interest in this problem, with several works showcasing impressive long-range 2D tracking results on challenging, in-the-wild videos. These approaches employ either test-time optimization where models consolidate noisy short-range motion estimates into long-term correspondences [70, 90, 101], or data-driven strategies [16, 17, 30, 44], where neural networks learn long-term correspondence estimates from synthetic data [15, 126]. While these methods effectively track any 2D point throughout a video, they lack the knowledge of underlying 3D scene geometry and motions.", + "bbox": [ + 91, + 537, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model 3D scene motion, recent work in monocu", + "bbox": [ + 109, + 886, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lar settings has explored self-supervised learning and test-time optimization [37, 57, 58, 114-116]. More recent approaches estimate 3D trajectories in general scenes in a feedforward manner [50, 72, 110], but the predicted motion remains in the frame space, entangling object and camera motion. In contrast, our method recovers persistent, long-range 3D trajectories in the world coordinate space.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Reconstruction and View Synthesis. Our work also relates to dynamic 3D scene reconstruction and novel view synthesis. In non-rigid reconstruction, early methods often required RGB-D sensors [5, 19, 40, 71, 127] or strong hand-crafted priors [52, 78, 82]. A family of works [6, 13, 73] apply low-rank assumptions to the underlying motion model. Recent work has demonstrated progress toward general dynamic reconstruction either by integrating monocular depth priors [49, 56, 60, 68, 124, 125], or directly learning to perform dynamic reconstruction [65, 102, 103, 123] from data. However, these methods do not perform long-range 3D tracking, and do not focus on novel view synthesis.", + "bbox": [ + 511, + 199, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [69] and Gaussian Splating [46] have shown strong performance in novel view synthesis. For dynamic scenes, many methods [1, 9, 10, 23, 23, 55, 59, 89, 92, 100] still require simultaneous multi-view video observations or predefined templates [38, 54, 107]. Template-free monocular approaches model dynamic scenes with different types of representations such as video depth maps [122], time-dependent NeRFs [20, 57, 58, 74-76, 97, 109], and dynamic 3D Gaussians [21, 108, 119, 120]. Although significant progress has been made, as DyCheck [25] pointed out, many approaches focus on scenarios with camera teleportation [51, 119] or quasi-static scenes [74, 75], which are effectively multiview and do not represent real-world monocular videos. In this work, we focus on casually-captured monocular videos, a more practical and challenging setup. Recent and concurrent efforts [12, 53, 62, 91, 104] have also investigated similar settings, using strong off-the-shelf data-driven priors.", + "bbox": [ + 511, + 380, + 903, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 664, + 604, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method takes as input a sequence of $T$ video frames $\\{I_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}$ of a dynamic scene, the camera intrinsics $\\mathbf{K}_t\\in \\mathbb{R}^{3\\times 3}$ , and world-to-camera extrinsics $\\mathbf{E}_t\\in \\mathbb{S}\\mathbb{E}(3)$ of each input frame $I_{t}$ . From these inputs, we aim to recover the geometry of the entire dynamic scene and the full-length 3D motion trajectory of every point in the scene. Unlike most prior dynamic NeRFs methods [25, 56, 58, 97, 109] which render the scene contents through volumetric ray casting and represent the motion implicitly at fixed 3D locations, we model the dense scene elements as a set of canonical 3D Gaussians, that translate and rotate over entire video as persistent motion trajectories. We adopt explicit point-based representation because it simultaneously allows for both (1) high-fidelity rendering in real-time and (2) full", + "bbox": [ + 511, + 688, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9661", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6acad7d745ddf3a9d6f7a829e41a90cfdcd4442515a6531f79b06bbad3d31478.jpg", + "image_caption": [ + "Figure 2. System Overview. Given a single RGB video sequence with known camera poses, monocular depth maps and 2D tracks computed from off-the-shelf models [16, 118] as input, we optimize a dynamic scene representation as a set of persistent 3D Gaussians that translate and rotate over time. To capture the low-dimensional nature of scene motion, we model the motion with a set of compact $\\mathbb{SE}(3)$ motion bases shared across all scene elements. Each 3D Gaussian's motion is represented as a linear combination of these motion bases, weighted by motion coefficients specific to each Gaussian. We supervise our scene representation (canonical 3D Gaussians, per-Gaussian motion coefficients, and global motion bases) by comparing the rendered outputs (RGB, depths and 2D tracks) with the corresponding input signals. This results in a globally coherent dynamic 3D scene representation in the world coordinate with long-range 3D trajectories." + ], + "image_footnote": [], + "bbox": [ + 132, + 94, + 867, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "length 3D tracking of any surface point from any input time.", + "bbox": [ + 89, + 487, + 482, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimizing an explicit representation of dynamic 3D Gaussians from a single video is severely ill-posed — at each point in time, the moving subjects in the scene are observed from only a single viewpoint. In order to overcome this ambiguity, we make two insights: First, while the projected 2D dynamics might be complex in the video, the underlying 3D motion in the scene is low-dimensional, and composed of simpler units of rigid motion. Second, powerful data-driven priors, namely monocular depth estimates and long-range 2D tracks, provide complementary but noisy signals of the underlying 3D scene. We propose a system that fuses these noisy estimates together into a globally coherent representation of both the scene geometry and motion. We show a schematic of our pipeline in Figure 2.", + "bbox": [ + 89, + 503, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 89, + 726, + 415, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent the appearance and geometry of a dynamic scene with a global set of 3D Gaussians, an explicit and expressive differentiable scene representation [45] for efficient optimization and rendering. We define parameters of each 3D Gaussian in the canonical frame $t_0$ as $\\vec{g}_0 \\equiv (\\pmb{\\mu}_0, \\mathbf{R}_0, \\mathbf{s}, o, \\mathbf{c})$ , where $\\pmb{\\mu}_0 \\in \\mathbb{R}^3$ , $\\mathbf{R}_0 \\in \\mathbb{SO}(3)$ are the 3D mean and orientation in the canonical frame, and $\\mathbf{s} \\in \\mathbb{R}^3$ the scale, $o \\in \\mathbb{R}$ the opacity, and $\\mathbf{c} \\in \\mathbb{R}^3$ the color. $\\Sigma_0$ is the canonical frame covariance matrix. These properties are persistent and shared across time. To render a set of 3D", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussians from a camera with world-to-camera extrinsics $\\mathbf{E}$ and intrinsics $\\mathbf{K}$ , the projections of the 3D Gaussians in the image plane are obtained by 2D Gaussians parameterized as $\\pmb{\\mu}_0^{\\prime} \\in \\mathbb{R}^{2}$ and $\\pmb{\\Sigma}_0^{\\prime} \\in \\mathbb{R}^{2 \\times 2}$ via affine approximation", + "bbox": [ + 511, + 487, + 906, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\Pi (\\mathbf {K E} \\boldsymbol {\\mu} _ {0}) \\in \\mathbb {R} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 573, + 799, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\mathbf {J} _ {\\mathbf {K E}} \\boldsymbol {\\Sigma} _ {0} \\mathbf {J} _ {\\mathbf {K E}} ^ {\\top} \\in \\mathbb {R} ^ {2 \\times 2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 593, + 906, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Pi$ is perspective projection, and $\\mathbf{J}_{\\mathbf{KE}}$ is the Jacobian of perspective projection with $\\mathbf{K}$ and $\\mathbf{E}$ at the point $\\mu_0$ . The 2D Gaussians can then be efficiently rasterized into RGB image and depth map via alpha compositing as", + "bbox": [ + 511, + 625, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {I}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {D}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {d} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 696, + 905, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i} = o_{i}\\cdot \\exp \\left(-\\frac{1}{2} (\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})^{T}\\boldsymbol{\\Sigma}_{0}^{\\prime}(\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})\\right)$ , and $T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})$ . $H(\\mathbf{p})$ is the set of Gaussians that intersect the ray shoot from the pixel $\\mathbf{p}$ . This process is fully differentiable, and enables direct optimization of the 3D Gaussian parameters.", + "bbox": [ + 511, + 743, + 906, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dynamic Scene Representation", + "text_level": 1, + "bbox": [ + 511, + 832, + 787, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Motion Parameterization. To model a dynamic 3D scene, we keep track of $N$ canonical 3D Gaussians and vary their positions and orientations over time with a", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9662", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "per frame rigid transformation. In particular, for a moving 3D Gaussian at time $t$ , its pose parameters $(\\pmb{\\mu}_t,\\mathbf{R}_t)$ are rigidly transformed from the canonical frame $t_0$ to $t$ via $\\mathbf{T}_{0\\rightarrow t} = \\left[\\mathbf{R}_{0\\rightarrow t}\\quad \\mathbf{t}_{0\\rightarrow t}\\right]\\in \\mathbb{S}\\mathbb{E}(3)$ :", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\boldsymbol {\\mu} _ {0} + \\mathbf {t} _ {0 \\rightarrow t}, \\quad \\mathbf {R} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\mathbf {R} _ {0}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 161, + 482, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than modeling the 3D motion trajectories independently for each Gaussian, we define a set of $B \\ll N$ learnable basis trajectories $\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^{B}$ that are globally shared across all Gaussians [51]. The transformation $\\mathbf{T}_{0\\rightarrow t}$ at each time $t$ is then computed by a weighted combination of this global set of basis trajectories through per-point basis coefficients $\\mathbf{w}^{(b)}$ via", + "bbox": [ + 89, + 185, + 483, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {0 \\rightarrow t} = \\sum_ {b = 0} ^ {B} \\mathbf {w} ^ {(b)} \\mathbf {T} _ {0 \\rightarrow t} ^ {(b)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 299, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| \\mathbf{w}^{(b)}\\| = 1$ . In our implementation, we parameterize $\\mathbf{T}_{0\\rightarrow t}^{(b)}$ as 6D rotation [31] and translation, and perform the weighted combination separately on each with the same weight $\\mathbf{w}^{(b)}$ . During optimization, we jointly learn the set of global motion bases and motion coefficients of each 3D Gaussian. These compact motion bases explicitly regularize the trajectories to be low-dimensional, encouraging the 3D Gaussians that move similarly to each other to be represented by similar motion coefficients.", + "bbox": [ + 89, + 349, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rasterizing 3D Trajectories. Given this representation, we now describe how we obtain pixelwise 3D motion trajectory at any query frame $I_{t}$ . We take a similar approach to Wang et al. [101] and rasterize the motion trajectories of 3D Gaussians into query frame $I_{t}$ . Namely, for a query camera at time $t$ with intrinsics $\\mathbf{K}_t$ and extrinsics $\\mathbf{E}_t$ , we perform rasterization to obtain a map ${}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'} \\in \\mathcal{R}^{H\\times W\\times 3}$ that contains the expected 3D world coordinates of the surface points corresponding to each pixel at target time $t'$", + "bbox": [ + 89, + 494, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { \\mathrm { w } } \\hat { \\mathbf { X } } _ { t \\rightarrow t ^ { \\prime } } ( \\mathbf { p } ) = \\sum _ { i \\in H ( \\mathbf { p } ) } T _ { i } \\alpha _ { i } \\boldsymbol { \\mu } _ { i , t ^ { \\prime } } , ( 5 )\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 638, + 482, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H(\\mathbf{p})$ is the set of Gaussians that intersect the pixel $\\mathbf{p}$ at query time $t$ . The 2D correspondence location at time $t'$ for a given pixel $\\mathbf{p}$ , $\\hat{\\mathbf{U}}_{t\\rightarrow t'}(\\mathbf{p})$ , and the corresponding depth value at time $t'$ , $\\hat{\\mathbf{D}}_{t\\rightarrow t'}(\\mathbf{p})$ can then be written as", + "bbox": [ + 89, + 681, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\Pi \\left(\\mathbf {K} _ {t ^ {\\prime}} ^ {\\mathrm {c}} \\hat {\\mathbf {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p})\\right), \\hat {\\mathbf {D}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\binom {\\mathrm {c}} {\\mathrm {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 752, + 506, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{^c}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p}) = \\mathbf{E}_{t'}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p}),$ $\\Pi$ is a perspective projection operation, and $(\\cdot)_{[3]}$ is the third element of a vector.", + "bbox": [ + 89, + 792, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Optimization", + "text_level": 1, + "bbox": [ + 89, + 833, + 230, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We prepare the following estimates using off-the-shelf methods in our optimization: 1) masks for the moving objects for each frame $\\{\\mathbf{M}_t\\}$ , which can be easily obtained", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "using Track-Anything [48, 117] with a few user clicks, 2) monocular depth maps $\\{\\mathbf{D}_t\\}$ computed using state-of-the-art relative depth estimation method such as Depth Anything [118] and 3) long-range 2D tracks $\\{\\mathbf{U}_{t\\rightarrow t^{\\prime}}\\}$ for foreground pixels from state-of-the-art point tracking method TAPIR [16]. We align the relative depth maps with the metric depth maps by computing a per-frame global scale and shift and use them for our optimization, as we found relative depth maps tend to contain finer details. We treat the lifted 2D tracks unprojected with the aligned depth maps as noisy initial 3D track observations $\\{\\mathbf{X}_t\\}$ for the moving objects. For the static part of the scene, we model them using standard static 3D Gaussians and initialize their 3D locations by unprojecting them into the 3D space using the aligned depth maps. The static and dynamic Gaussians are jointly optimized and rasterized together to form an image. We describe the optimization process below.", + "bbox": [ + 511, + 90, + 906, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initialization. We first select the canonical frame $t_0$ to be the frame in which the most 3D tracks are visible, and initialize the Gaussian means in the canonical frame $\\mu_0$ as $N$ randomly sampled 3D tracks locations from this set of initial observations. We then perform k-means clustering on the vectorized velocities of the noisy 3D tracks from every frame $\\{\\mathbf{X}_t\\}$ , and initialize the motion bases $\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B$ from these $B$ clusters of tracks. Specifically, for the set of trajectories $\\{\\mathbf{X}_t\\}_b$ belonging to cluster $b$ , we initialize the basis transform $\\mathbf{T}_{0\\rightarrow \\tau}^{(b)}$ using weighted Procrustes alignment between the point sets $\\{\\mathbf{X}_0\\}_b$ and $\\{\\mathbf{X}_{\\tau}\\}_b$ for all $\\tau = 0,\\dots ,T$ , where the weights are computed using uncertainty and visibility scores from TAPIR predictions. We initialize $\\mathbf{w}^{(b)}$ of each Gaussian to decay exponentially with its distance from the center of cluster $b$ in the canonical frame. We then optimize $\\mu_0$ , $\\mathbf{w}^{(b)}$ , and set of basis functions $\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B$ to fit the observed 3D tracks with an $\\ell_1$ -loss under temporal smoothness constraints.", + "bbox": [ + 511, + 359, + 906, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. We supervise the dynamic Gaussians with two sets of losses. The first set comprise our reconstruction loss to match the per-frame pixelwise color, depth, and masks inputs. The second set enforce consistency of correspondences across time. Specifically, during each training step, we render the image $\\hat{\\mathbf{I}}_t$ , depth $\\hat{\\mathbf{D}}_t$ , and mask $\\hat{\\mathbf{M}}_t$ from their corresponding training cameras $(\\mathbf{K}_t,\\mathbf{E}_t)$ according to Equation 2. We supervise these predictions with a reconstruction loss applied independently per-frame", + "bbox": [ + 511, + 646, + 908, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e c o n}} = \\left\\| \\hat {\\mathbf {I}} - \\mathbf {I} \\right\\| _ {1} + \\lambda_ {\\text {d e p t h}} \\left\\| \\hat {\\mathbf {D}} - \\mathbf {D} \\right\\| _ {1} + \\lambda_ {\\text {m a s k}} \\left\\| \\hat {\\mathbf {M}} - \\mathbf {M} \\right\\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 792, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second set of losses supervises the motion of the Gaussians between frames. Specifically, we additionally render the 2D tracks $\\hat{\\mathbf{u}}_{t\\rightarrow t^{\\prime}}$ and reprojected depths $\\hat{\\mathbf{D}}_{t\\rightarrow t^{\\prime}}$ for a pair of randomly sampled query time $t$ and target time $t^\\prime$ . We supervise these rendered correspondences with the", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9663", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "long-range 2D track estimates via", + "bbox": [ + 89, + 90, + 316, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t r a c k - 2 d}} = \\left\\| \\mathbf {U} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} \\right\\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 122, + 480, + 140 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t r a c k - d e p t h}} = \\left\\| \\hat {\\mathbf {d}} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {D}} \\left(\\mathbf {U} _ {t \\rightarrow t ^ {\\prime}}\\right)\\right\\| _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 143, + 482, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we leverage physics motion prior by enforcing a distance-preserving loss between randomly sampled dynamic Gaussians and their k-nearest neighbors. Let $\\hat{\\mathbf{X}}_t$ and $\\hat{\\mathbf{X}}_{t'}$ denote the location of a Gaussian at time $t$ and $t'$ , and $\\mathcal{C}_k(\\hat{\\mathbf{X}}_t)$ denote the set of k-nearest neighbors of $\\hat{\\mathbf{X}}_t$ , we define", + "bbox": [ + 89, + 181, + 483, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r i g i d i t y}} = \\left\\| \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t}\\right)\\right) - \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 285, + 482, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{dist}(\\cdot ,\\cdot)$ measures Euclidean distance.", + "bbox": [ + 89, + 325, + 387, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For in-the-wild videos, multiple methods [60, 95] are available to estimate camera parameters; further details are provided in the supplement. During reconstruction, camera extrinsics are refined as learnable parameters, update via camera gradient function in gsplat [121] CUDA kernel. For evaluating our methods on public benchmark, we use the camera annotations that come with the datasets (e.g., from COLMAP [86] or simulation).", + "bbox": [ + 89, + 349, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimize our model using Adam Optimizer [47]. We perform 1000 iterations of optimization for the initialization phase and 500 epochs for training phase, respectively. The number of $\\mathbb{SE}(3)$ bases $B$ is set to 10 for all of our experiments. We initialize 40k dynamic Gaussians for the dynamic part and 100k static Gaussians for the static part of the scene, respectively. We perform the same adaptive Gaussian controls for dynamic and static Gaussians as per 3D-GS [46]. Training on a sequence of 300 frames of $960 \\times 720$ resolution takes about 2 hours to finish on an A100 GPU. Our rendering FPS is around 140 fps. We use the same set of hyper-parameters for all experiments.", + "bbox": [ + 89, + 473, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 676, + 215, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our performance quantitatively and qualitatively on a broad range of tasks: long range 3D point tracking, long-range 2D point tracking, and novel view synthesis. We focus our evaluation on datasets that exhibit substantial scene motion. In particular, the iPhone dataset [25] features casual captures of real-world scenes that closely match our target scenarios. It provides comprehensive annotations, including simultaneous validation views, lidar depth, sparse 2D point correspondences across the entire video, and can be used to evaluate our performance on all three tasks. Given the challenge of obtaining precise 3D track annotations for real data, we also evaluate performance using the scenes from the synthetic MOVi-F Kubric dataset [28].", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/523415722d16eb2f471f166f3a94f9630ee5a1cb76301ab9406bdd05f0515d90.jpg", + "image_caption": [ + "Figure 3. 3D Track visualization on iPhone dataset. We render novel views for each method and overlay their predicted 3D tracks on top of the rendered images. For clarity, we only display a segment of the trails spanning 50 frames for a specific set of grid query points. However, it is important to note that our method can generate dense, full-length 3D tracks. *Note that TAPIR + DA cannot produce novel view rendering, and we overlay their tracking results with our rendering to make it easier to interpret." + ], + "image_footnote": [], + "bbox": [ + 526, + 87, + 908, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Task Specification", + "text_level": 1, + "bbox": [ + 511, + 484, + 689, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Long-Range 3D Tracking. Our primary task is estimating 3D scene motion for any pixel over the entire video. For this purpose, we extend the metrics for scene flow evaluation introduced in RAFT-3D [96] to evaluate long-range 3D tracking. Specifically, we report the 3D end-point-error (EPE), which measures the Euclidean distance between the ground truth 3D tracks and predicted tracks at each target time step. In addition, we report the percentage of points that fall within a given threshold of the ground truth 3D location $\\delta_{3D}^{0.05} = 5\\mathrm{cm}$ and $\\delta_{3D}^{10} = 10\\mathrm{cm}$ in metric scale.", + "bbox": [ + 511, + 512, + 906, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Long-Range 2D Tracking. Our 3D motion representation can be easily projected onto image plane to get long-range 2D tracks. Thus we also evaluate 2D tracking performance in terms of both position and occlusion accuracy. Following the metrics introduced in the TAP-Vid benchmark [15], we report the Average Jaccard (AJ), average position accuracy $(<\\delta_{\\mathrm{avg}})$ , and Occlusion Accuracy (OA).", + "bbox": [ + 511, + 676, + 908, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Novel View Synthesis. We measure our method's novel view synthesis quality as a comprehensive assessment for geometry, appearance, and motion. We evaluate on the iPhone dataset [25] which provides validation views and report co-visibility masked image metrics [25]: mPSNR, mSSIM [105] and mLPIPS [26, 36]. Additionally, we evaluate on NVIDIA dataset for more comparison.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9664", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/015a08db001ce24520d1e4368e3b2ae671425e9b1490b59f39609c5c73729107.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3D Tracking2D TrackingView Synthesis
EPE ↓δ053D↑δ103D↑AJ ↑<δavg↑OA ↑PSNR ↑SSIM ↑LPIPS ↓
T-NeRF [25]------15.600.550.55
HyperNeRF [75]0.18228.445.810.119.352.015.990.590.51
DynIBaR [58]0.25211.424.65.48.737.713.410.480.55
Deformable-3D-GS [119]0.15133.455.314.020.963.911.920.490.66
DynMF [51]0.18822.953.85.59.560.516.540.590.49
CoTracker [44]+DA [118]0.20234.357.924.133.973.0---
TAPIR [16]+DA [118]0.11438.163.227.841.567.4---
DELTA (world) [72]0.15932.555.324.734.168.9---
SpatialTracker (world) [110]0.12537.763.924.936.973.5---
Ours0.08243.073.334.447.086.616.720.630.45
Ours + 2DGS[33]0.09747.371.335.847.087.316.750.650.40
", + "bbox": [ + 236, + 88, + 767, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation on iPhone dataset. Our method achieves SOTA performance all tasks of 3D point tracking, 2D point tracking, and novel view synthesis. The baselines that perform best on 2D and 3D tracking (TAPIR [16] + DA [118], CoTracker [44] + DA [118], DELTA [72], SpatialTracker [110]) are unable to synthesize novel views of the scene, while the methods that perform best in novel view synthesis struggle with or fail to produce 2D and 3D tracks. Our method achieves a significant boost in all three tasks above baselines. We include training details about “Ours + 2DGS [33]” in the supplement.", + "bbox": [ + 88, + 267, + 906, + 337 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2efc3ec5fb7ff822b96f8f2939d5645a8cb570a8d97cf93df172b407b85c48f.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of novel view synthesis on iPhone dataset. The leftmost image in each row shows the training view at the same time step as the validation view. The regions highlighted in green indicate areas excluded from evaluation due to the lack of co-visibility between training and validation views. Please see the supplemental video for more qualitative results and comparisons." + ], + "image_footnote": [], + "bbox": [ + 101, + 339, + 906, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 89, + 760, + 199, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method represents dynamic 3D scene comprehensively with explicit long-range 3D scene motion estimation, which also allows for novel view synthesis. While no existing method achieves the exact same goals as ours, there are methods that focus on sub-tasks related to our problem, such as dynamic novel view synthesis, 2D tracking, or monocular depth estimation. We therefore adapt existing", + "bbox": [ + 88, + 794, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods as our baselines which are introduced below.", + "bbox": [ + 511, + 761, + 870, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While dynamic novel view synthesis approaches focus primarily on the photometric reconstruction quality and do not explicitly output 3D point tracks, we can adapt some of their representations to derive 3D point tracks for our evaluation. For HyperNeRF [75], we compose the learned inverse mapping (from view space to canonical space) and a forward mapping solved via root-finding [11, 25] to produce 3D tracks at query points. DynIBaR [58] produces", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9665", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8dc3a3c551e55532cedb75d38e0fc8590cff27f2aaf0113bc5db7f86ad662df.jpg", + "image_caption": [ + "Figure 5. Visualization of motion coefficients after PCA and predicted depth maps on iPhone dataset. The motion coefficients encode information regarding rigid moving components. For instance, our motion coefficient PCA produces constant color for the block in the second example which exhibits rigid motion." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 478, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "short-range view-to-view scene flow, which we chain into long-range 3D tracks for our evaluation. Deformable-3D-GS (D-3DGS) [119] represents dynamic scenes using 3D Gaussians [45] in the canonical space and a deformation MLP network that deforms the canonical 3D Gaussians into each view, which naturally allows 3D motion computation. T-NeRF [25] models dynamic scenes using time as MLP input in addition to 3D locations, which does not provide a method for extracting 3D motion, and hence they are not considered for 2D/3D tracking evaluation. Finally, we implemented our own version of DynMF [51] with neural network motion bases. Implementation details are provided in the supplemental material. We evaluate the 3D tracks of these methods only on the iPhone dataset, because we found none can handle the Kubric scene motion.", + "bbox": [ + 88, + 277, + 482, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alongside dynamic view synthesis baselines, we include 3D tracking baselines. Specifically, we take state-of-the-art long-range 2D tracks from TAPIR [16] and CoTracker [44] and lift them into 3D scene motion using monocular depth maps produced by Depth Anything (DA) [118]. We compute the correct scale and shift for each relative depth map from Depth Anything to align them with the scene. The two resulting baselines are called \"CoTracker $[44] + DA[118]$ \" and \"TAPIR $[16] + DA[118]$ \". Note that these baselines can only produce 3D tracks for visible regions, as the depth values for occluded points are unknown from such a 2.5D representation. In contrast, our global representation allows for modeling 3D motion through occlusions. In addition, we include frame-space 3D tracking methods such as DELTA [72] and SpatialTracker [110], and transform their trajectories into world coordinates using annotated camera poses for evaluation.", + "bbox": [ + 89, + 505, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on iPhone Dataset", + "text_level": 1, + "bbox": [ + 89, + 771, + 354, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iPhone dataset [25] contains 14 sequences of 200-500 frames featuring various types of challenging real world scene motion. All sequences are recorded using a handheld moving camera in a casual manner, with 7 of them additionally featuring two synchronized static cameras with a large baseline for novel view synthesis evaluation. For 3D tracking evaluation, we generate the groundtruth 3D tracks", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by lifting the 2D keypoint annotations into 3D using lidar depth, and masking out points that are occluded or with invalid lidar depth values. All experiments are conducted with the original instead of half resolution as in [25] given that our method can handle high-res video input. We also discard scenes SPACE-OUT and WHEEL due to camera and lidar error. We find that the original camera annotations from ARKit is not accurate enough, and refine them using global bundle adjustment from COLMAP [85]. The refined poses are used for all methods.", + "bbox": [ + 511, + 90, + 903, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the quantitative comparison in Tab. 1 which shows that our method outperforms all baselines on all tasks by a substantial margin. For tracking, we observe significant improvements over naive baselines that combine 2D tracks with depth maps, especially \"TAPIR+DA\", which also serves as input to our system, highlighting the effectiveness of our consolidation process. The improvement stems from three factors: 1) a low-dimensional $\\mathbb{SE}(3)$ motion representation, which allows inferring occluded motion via nearby visible regions; 2) 3D regularizations, such as low acceleration constraints that promote coherent trajectories; 3) the consolidation of all pairwise tracks into a single 4D representation with persistent 3D geometry, which serves as a structural prior to correct noisy 2D tracks. Our method also achieves the best novel view synthesis quality across dynamic NeRF and 3D-GS-based baselines.", + "bbox": [ + 511, + 242, + 906, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3 shows qualitative comparison of the 3D tracking results. We render the novel views and plot the predicted 3D tracks of the given query points onto the novel views. Since \"TAPIR + DA\" cannot perform novel view synthesis, we overlay their track predictions onto our renderings to aid interpretation. D-3DGS [119] and HyperNeRF [75] fail to capture the significant scene motion in PAPER-WINDMILL and SPIN, resulting in structure degradation and blurry rendering. \"TAPIR + DA\" can track large motions, but their 3D tracks tend to be noisy and erroneous. In contrast, our method not only generates high-quality novel views but also the most smooth and accurate 3D tracks. Fig. 4 provides additional novel view synthesis comparison on the validation views. In Fig. 5, we visualize the rendering of the first three components of PCA decomposition for the motion coefficients, which correlates well with the rigid groups of the moving objects.", + "bbox": [ + 511, + 484, + 906, + 741 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29b746b8d0113c44b9453dc9f40c531baee4e1942dd0b0cd6f4843107c1542e1.jpg", + "table_caption": [ + "4.4. Evaluation on Kubric Dataset" + ], + "table_footnote": [], + "table_body": "
MethodEPE↓δ0.053D↑δ103D↑
CoTracker [44]+DA [118]0.1934.456.5
TAPIR [16]+DA [118]0.2034.056.2
Ours0.1639.862.2
", + "bbox": [ + 540, + 781, + 857, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. 3D Tracking evaluation on Kubric dataset.", + "bbox": [ + 535, + 857, + 859, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Kubric MOVi-F dataset contains short 24-frame videos", + "bbox": [ + 513, + 885, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9666", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of scenes of 10-20 objects, rendered with linear camera movement and motion blur. Multiple rigid objects are tossed onto the scene, at a speed that far exceeds the speed of the moving camera, making it a similar capture scenario to in-the-wild capture scenarios. It provides dense comprehensive annotations, including ground truth depth maps, camera parameters, segmentation masks, and point correspondences that are dense across time. We use 30 scenes from the MOVi-F validation set to evaluate the accuracy of our long-range 3D point tracks for all points in time.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate our method on the synthetic Kubric MOVi-F dataset in long-range 3D point tracking across time. Of the above baselines, only \"CoTracker [44] + DA [118]\" and \"TAPIR [44] + DA [118]\" yield 3D tracks for these scenes. For all baselines, we provide the ground truth camera intrinsics and extrinsics, and monocular depth estimates that have been aligned to the ground truth depth map. We obtain point tracks for all non-background pixels for each method.", + "bbox": [ + 89, + 242, + 480, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report our quantitative 3D point tracking metrics in Table 2, and find that across all metrics, our method outperforms the baselines. Please see the supplement for figures of the optimized motion coefficients, which coherently group moving objects.", + "bbox": [ + 89, + 378, + 480, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Evaluation on NVIDIA Dataset", + "text_level": 1, + "bbox": [ + 89, + 463, + 366, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on seven scenes from the NVIDIA dataset [122], following the evaluation protocol of Dynamic Gaussian Marbles [91], which evaluates the task from a single static camera instead of interleaving different camera as discussed in [25]. We adopt the same image resolution, camera parameters and depths maps as input. Please see the supplement for more details. For each scene, we compute covisibility masks between each evaluation view and the training view at each time step to exclude unobservable regions during evaluation. Results are reported in Tab. 3.", + "bbox": [ + 89, + 484, + 480, + 636 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70e8a78032ebdcf58034e91b069e73527e3c079c56222b977c6f6ba411cf6674.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBalloon1Ballon2JumpingTruck
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]23.830.810.06623.620.810.08820.020.730.13427.310.860.055
Ours23.900.790.05823.910.830.07420.080.690.15127.650.880.048
MethodPlaygroundUmbrellaSkatingMean
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]16.680.600.17724.990.700.08327.790.900.05223.460.770.093
Ours16.870.550.16823.270.570.16727.910.920.04723.370.750.102
", + "bbox": [ + 91, + 648, + 480, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 787, + 250, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate various components of our method on the iPhone dataset in 3D tracking in Tab. 4. We first validate our choices of motion representation, namely the $\\mathbb{SE}(3)$ motion bases parameterization, with three ablations: 1) \"Per-Gaussian Transl\": replacing our motion representation with naive per-Gaussian translational motion trajectories,", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1a10bfde66aab45cb30daeb57087b0bc242bb2728c59dee01aed1b007d5f259.jpg", + "table_caption": [ + "Table 3. Evaluation on NVIDIA dataset. Our method is comparable with Dynamic Gaussian Marbles (DGM) [91]." + ], + "table_footnote": [], + "table_body": "
MethodsSE(3)Motion Basis2D tracksInitializationEPE↓δ0.053D↑δ1.103D↑
Ours (Full)✓✓✓✓0.08243.073.3
Transl. Bases✓✓✓0.09342.369.9
Per-Gaussian SE(3)✓✓✓0.08343.670.2
Per-Gaussian Transl.✓✓0.08741.269.2
No SE(3) Init.✓✓✓0.11139.365.7
No 2D Tracks✓✓0.14130.457.8
", + "bbox": [ + 514, + 89, + 903, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Studies on iPhone dataset.", + "bbox": [ + 568, + 175, + 846, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2) \"Per-Gaussian $\\mathbb{SE}(3)$ : replacing our motion representation with naive per-Gaussian $\\mathbb{SE}(3)$ motion trajectories, and 2) \"Transl. Bases\": keeping the motion bases representation but only using translational bases instead of $\\mathbb{SE}(3)$ . We find that $\\mathbb{SE}(3)$ bases significantly improve 3D tracking over both translational bases and per-Gaussian translational motion. It also outperforms per-Gaussian $\\mathbb{SE}(3)$ in overall accuracy and offers noticeably better visual quality, as the latter exhibits significantly more popping and jitter artifacts.", + "bbox": [ + 511, + 191, + 903, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we ablate our training strategies including initialization and supervision signal. We conduct an ablation of \"No $\\mathbb{SE}(3)$ Init.\", where instead of performing our initial $\\mathbb{SE}(3)$ fitting stage, we initialize the translational part of the motion bases with randomly selected noisy 3D tracks formed by directly lifting input 2D tracks using depth maps into 3D, and the rotation part as identity. We find that skipping this initialization noticeably hurts performance. Lastly, we remove the 2D track supervision entirely (\"No 2D Tracks\") and find it to lead to significant drop in performance, which verifies the importance of the 2D track supervision for 3D tracking.", + "bbox": [ + 511, + 329, + 906, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 513, + 525, + 625, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Like most prior monocular view synthesis methods, it still requires per-scene test-time optimization, hindering streamable applications. Recent feed-forward methods for joint reconstruction and tracking [22, 111, 113] offer a promising alternative, showing encouraging progress toward overcoming this limitation. Our method relies on off-the-shelf predictions for camera poses, geometry, and motion, which may degrade in textureless regions or under large motions. However, it also naturally benefits from advances in these components. Finally, it also requires user input to mask moving objects. Recent advances in moving object segmentation [27, 34] offer promising directions beyond our current manual solution and could be integrated into our pipeline.", + "bbox": [ + 511, + 551, + 906, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. We introduce a method for joint long-range 3D tracking and novel view synthesis from monocular video, representing dynamic elements via global 3D Gaussians with time-varying translations and rotations. Motion trajectories are regularized using low-dimensional rigid motion bases. By consolidating noisy observations into globally consistent scene estimates, our approach outperforms state-of-the-art methods in 2D/3D tracking and novel view synthesis across synthetic and real benchmarks.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9667", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank Ruilong Li, Noah Snavely, Brent Yi and Aleksander Holynski for helpful discussion. We are in memory of our beloved cat Sriracha, who will always be missed and loved. This project is supported in part by DARPA No. HR001123C0021. and IARPA DOI/IBC No. 140D0423C0035. The views and conclusions contained herein are those of the authors and do not represent the official policies or endorsements of these institutions.", + "bbox": [ + 89, + 90, + 485, + 203 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, 2006. 2", + "[3] Stanley T Birchfield and Shrinivas J Pundlik. Joint tracking of features and edges. In 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2008. 2", + "[4] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, 1993. 2", + "[5] Aljaz Bozic, Michael Zollhöfer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[6] C. Bregler, A. Hertzmann, and H. Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No.PR00662), 2000. 2", + "[7] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In European Conference on Computer Vision, 2004. 2", + "[8] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2", + "[9] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4), 2020. 1, 2", + "[10] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[11] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6", + "[12] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki." + ], + "bbox": [ + 101, + 257, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dreamscene4d: Dynamic multi-object scene generation from monocular videos. NeurIPS, 2024. 2", + "[13] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. Int. J. Comput. Vision, 107(2), 2014. 2", + "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2", + "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 2022. 2, 5", + "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 6, 7", + "[17] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Joao Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, 2024. 2", + "[18] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Häusser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), 2015. 2", + "[19] Mingsong Dou, S. Khamis, Yu.G. Degtyarev, Philip L. Davidson, S. Fanello, Adarsh Kowdle, Sergio Orts, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4d. ACM Transactions on Graphics (TOG), 35, 2016. 2", + "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2021. 2", + "[21] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2", + "[22] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J Black, Trevor Darrell, and Angjoo Kanazawa. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. arXiv preprint arXiv:2504.13152, 2025.8", + "[23] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. Kplanes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9668", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1", + "[25] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Dynamic novel-view synthesis: A reality check. In NeurIPS, 2022. 2, 5, 6, 7, 8", + "[26] Leon A Gatys, Alexander S Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 5", + "[27] Lily Goli, Sara Sabour, Mark Matthews, Marcus Brubaker, Dmitry Lagun, Alec Jacobson, David J Fleet, Saurabh Saxena, and Andrea Tagliasacchi. Romo: Robust motion segmentation improves structure from motion. arXiv preprint arXiv:2411.18650, 2024. 8", + "[28] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5", + "[29] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1", + "[30] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, 2022. 2", + "[31] Thorsten Hempel, Ahmed A Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In 2022 IEEE International Conference on Image Processing (ICIP), 2022. 4", + "[32] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. In Other Conferences, 1981. 2", + "[33] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 6", + "[34] Nan Huang, Wenzhao Zheng, Chenfeng Xu, Kurt Keutzer, Shanghang Zhang, Angjoo Kanazawa, and Qianqian Wang. Segment any motion in videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025. 8", + "[35] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. In European Conference on Computer Vision. Springer, 2022. 2", + "[36] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, 2020. 5" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[38] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 2023. 2", + "[39] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2", + "[40] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European Conference on Computer Vision, 2016. 2", + "[41] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), 2018. 2", + "[42] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[43] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2", + "[44] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European conference on computer vision, 2024. 2, 6, 7, 8", + "[45] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 1, 3, 7", + "[46] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 2, 5", + "[47] Diederik P Kingma, J Adam Ba, and J Adam. A method for stochastic optimization. arxiv 2014. arXiv preprint arXiv:1412.6980, 2020. 5", + "[48] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4", + "[49] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[50] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, Joao Carreira, Andrew Zisserman, Gabriel Brostow, and" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9669", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. Advances in Neural Information Processing Systems, 2024. 2", + "[51] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. ECCV, 2024. 2, 4, 6, 7", + "[52] Suryansh Kumar, Yuchao Dai, and Hongdong Li. Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2", + "[53] Jiahui Lei, Yijia Weng, Adam W Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2", + "[54] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision, 2022. 2", + "[55] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[56] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2", + "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2", + "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 6", + "[59] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[60] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2, 5", + "[61] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 2", + "[62] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy" + ], + "bbox": [ + 99, + 92, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "namic gaussian splatting from causally-captured monocular videos. ICLR, 2025. 2", + "[63] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 1", + "[64] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004. 2", + "[65] Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2", + "[66] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, 1981. 2", + "[67] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), 2024. 1", + "[68] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4), 2020. 2", + "[69] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[70] Michal Neoral, Jonás Šerych, and Jiří Matas. Mft: Long-term tracking of every pixel. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 2", + "[71] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2", + "[72] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 2, 6, 7", + "[73] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In ICCV, 2019. 2", + "[74] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2", + "[75] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. SIGGRAPH Asia, 2021. 1, 2, 6, 7", + "[76] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + ], + "bbox": [ + 522, + 92, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9670", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[77] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision, 2020. 1", + "[78] Rene Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[79] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multi-frame optical flow estimation. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 2", + "[80] Michael Rubinstein and Ce Liu. Towards longer long-range motion trajectories. In *British Machine Vision Conference*, 2012. 2", + "[81] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R. Bradski. Orb: An efficient alternative to sift or surf. 2011 International Conference on Computer Vision, 2011. 2", + "[82] Chris Russell, Rui Yu, and Lourdes Agapito. Video popu- up: Monocular 3d reconstruction of dynamic scenes. In European conference on computer vision. Springer, 2014. 2", + "[83] P Sand. Long-range video motion estimation using point trajectories. PhD thesis, Ph.D. dissertation, Cambridge, MA, USA, 2006, adviser-Teller, Seth, 2006. 2", + "[84] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 2008. 2", + "[85] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7", + "[86] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 5", + "[87] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In ICCV, 2023. 2", + "[88] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. International Journal of Computer Vision, 2004. 2", + "[89] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5), 2023. 2", + "[90] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision, 2024. 2", + "[91] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. Siggraph Asia, 2024. 2, 8", + "[92] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum. Wiley Online Library, 2008. 2", + "[93] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[94] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. Springer, 2020. 2", + "[95] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 2021. 5", + "[96] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5", + "[97] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2", + "[98] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[99] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. 2013 IEEE International Conference on Computer Vision, 2013. 2", + "[100] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[101] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In International Conference on Computer Vision, 2023. 2, 4", + "[102] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2", + "[103] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2", + "[104] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2", + "[105] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9671", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[106] Zirui Wang, Shuda Li, Henry Howard-Jenkins, Victor Adrian Prisacariu, and Min Chen. Flownet3d++: Geometric losses for deep scene flow estimation. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 1", + "[107] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, 2022. 2", + "[108] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2023. 1, 2", + "[109] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[110] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7", + "[111] Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Yuri Makarov, Bingyi Kang, Xing Zhu, Hujun Bao, Yujun Shen, and Xiaowei Zhou. Spatialtrackerv2: 3d point tracking made easy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 8", + "[112] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2", + "[113] Zhen Xu, Zhengqin Li, Zhao Dong, Xiaowei Zhou, Richard Newcombe, and Zhaoyang Lv. 4dgt: Learning a 4d gaussian transformer using real-world monocular videos. arXiv preprint arXiv:2506.08015, 2025. 8", + "[114] Gengshan Yang, Deqing Sun, V. Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[115] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021.", + "[116] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[117] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4", + "[118] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Un" + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "leashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 4, 6, 7, 8", + "[119] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 1, 2, 6, 7", + "[120] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. *ICLR*, 2024. 2", + "[121] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research, 26(34), 2025. 5", + "[122] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 8", + "[123] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arxiv:2410.03825, 2024. 2", + "[124] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4), 2021. 2", + "[125] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision. Springer, 2022. 2", + "[126] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2", + "[127] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew William Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (TOG), 33, 2014. 2" + ], + "bbox": [ + 516, + 92, + 903, + 739 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "9672", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_model.json b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b40827cf6e059c6170283dbb1c65bbb65302d8fd --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_model.json @@ -0,0 +1,2965 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.131, + 0.788, + 0.154 + ], + "angle": 0, + "content": "Shape of Motion: 4D Reconstruction from a Single Video" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.18, + 0.741, + 0.218 + ], + "angle": 0, + "content": "Qianqian Wang\\(^{1,2*}\\), Vickie Ye\\(^{1*}\\), Hang Gao\\(^{1*}\\), Weijia Zeng\\(^{3*}\\), Jake Austin\\(^{1}\\), Zhengqi Li\\(^{4}\\), Angjoo Kanazawa\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.226, + 0.796, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)UC Berkeley \\(^{2}\\)Google DeepMind \\(^{3}\\)UC San Diego \\(^{4}\\)Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.278, + 0.496, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.278, + 0.906, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.443, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 1. Shape of Motion. Our method enables joint long-range 3D tracking and novel view synthesis from a monocular video of a complex dynamic scene. We render moving elements at a fixed viewpoint across time and visualize estimated 3D motion as colorful trajectories. These trajectories reveal distinct geometric patterns, which leads to the term \"Shape of Motion\"." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.498, + 0.327, + 0.513 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Monocular dynamic reconstruction is a challenging and long-standing vision problem due to the highly ill-posed nature of the task. Existing approaches depend on templates, are effective only in quasi-static scenes, or fail to model 3D motion explicitly. We introduce a method for reconstructing generic dynamic scenes, featuring explicit, persistent 3D motion trajectories in the world coordinate frame, from casually captured monocular videos. We tackle the problem with two key insights: First, we exploit the low-dimensional structure of 3D motion by representing scene motion with a compact set of \\(\\mathbb{S}\\mathbb{E}(3)\\) motion bases. Each point's motion is expressed as a linear combination of these bases, facilitating soft decomposition of the scene into multiple rigidly-moving groups. Second, we take advantage of off-the-shelf data-driven priors such as monocular depth maps and long-range 2D tracks, and devise a method to effectively consolidate these noisy supervisory signals, resulting in a globally consistent representation of the dynamic scene. Experiments show that our method achieves state-of-the-art performance for both long-range 3D/2D motion estimation and novel view synthesis on dynamic scenes. Project Page: https://shape-of-motion.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.498, + 0.645, + 0.513 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Reconstructing the persistent geometry and their 3D motion across a video is crucial for understanding and interacting with the underlying physical world. While recent years have seen impressive progress in modeling static 3D scenes [45, 69], recovering the geometry and motion of complex dynamic 3D scenes, especially from a single video, remains an open challenge. A number of prior dynamic reconstruction and novel view synthesis approaches have attempted to tackle this problem. However, most methods rely on synchronized multi-view videos [9, 23, 23, 67, 108] or additional LIDAR/depth sensors [29, 63, 77, 98, 106]. Recent monocular approaches can operate on regular dynamic videos, but they typically model 3D scene motion as short-range scene flow between consecutive times [24, 57, 58] or deformation fields that map between canonical and view space [74, 75, 119], failing to capture 3D motion trajectories persistent over a video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The longstanding challenge for general in-the-wild videos lies in the poorly constrained nature of the 4D reconstruction problem. In this work, we tackle this challenge with two key insights. The first is that, while the image space dynamics can be complex and discontinuous, the underlying 3D motion is a composition of continuous simple rigid motions. Our second insight is that data-driven priors provide complementary, though noisy cues, that aggregate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.202, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9660" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "well into a globally coherent representation of the 3D scene geometry and motion." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.408 + ], + "angle": 0, + "content": "Motivated by these two insights, we represent the dynamic scene as a set of persistent 3D Gaussians, and represent their motion across the video in terms of a compact set of shared \\(\\mathbb{S}\\mathbb{E}(3)\\) motion bases. Unlike traditional scene flow, which computes 3D correspondences between consecutive frames, our representation recovers a persistent 3D trajectory over the whole video, enabling long-range 3D tracking across the entire video. As the 3D trajectories produced by our method capture the geometric patterns that trace each point's movement through 3D space and time as shown in Figure 1, we refer to our approach as \"Shape of Motion\". We show how to fit our explicit scene representation to a general video in-the-wild, by fusing together complementary cues from two main sources: monocular depth estimates per-frame, and 2D track estimates across frames. We conduct extensive evaluations on both synthetic and real-world dynamic video datasets, and show that our proposed approach significantly outperforms prior methods in both long-range 2D/3D tracking and novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.5 + ], + "angle": 0, + "content": "In summary, our key contributions are: (1) A 4D scene representation enabling both real-time novel view synthesis and globally consistent 3D tracking for any point at any time. (2) An optimization framework that fuses learned motion and geometry priors into a unified 4D representation from a single, casually captured monocular video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.233, + 0.529 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Correspondences and Tracking. While monocular 3D long range tracking remains largely unexplored, numerous approaches track in 2D image space, typically using optical flow for point correspondences. This involves estimating dense motion fields between image pairs [4, 7, 8, 18, 32, 35, 35, 39, 41-43, 66, 79, 87, 93, 94, 112]. Sparse keypoint matching methods can enable long trajectory generation [2, 14, 61, 64, 81], but these methods are primarily intended for sparse 3D reconstruction. Long-range 2D trajectory estimation for arbitrary points has been explored in earlier works, which relied on hand-crafted priors to generate motion trajectories [3, 80, 83, 84, 88, 99]. Recently, there has been a resurgence of interest in this problem, with several works showcasing impressive long-range 2D tracking results on challenging, in-the-wild videos. These approaches employ either test-time optimization where models consolidate noisy short-range motion estimates into long-term correspondences [70, 90, 101], or data-driven strategies [16, 17, 30, 44], where neural networks learn long-term correspondence estimates from synthetic data [15, 126]. While these methods effectively track any 2D point throughout a video, they lack the knowledge of underlying 3D scene geometry and motions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To model 3D scene motion, recent work in monocu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "lar settings has explored self-supervised learning and test-time optimization [37, 57, 58, 114-116]. More recent approaches estimate 3D trajectories in general scenes in a feedforward manner [50, 72, 110], but the predicted motion remains in the frame space, entangling object and camera motion. In contrast, our method recovers persistent, long-range 3D trajectories in the world coordinate space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.905, + 0.38 + ], + "angle": 0, + "content": "Dynamic Reconstruction and View Synthesis. Our work also relates to dynamic 3D scene reconstruction and novel view synthesis. In non-rigid reconstruction, early methods often required RGB-D sensors [5, 19, 40, 71, 127] or strong hand-crafted priors [52, 78, 82]. A family of works [6, 13, 73] apply low-rank assumptions to the underlying motion model. Recent work has demonstrated progress toward general dynamic reconstruction either by integrating monocular depth priors [49, 56, 60, 68, 124, 125], or directly learning to perform dynamic reconstruction [65, 102, 103, 123] from data. However, these methods do not perform long-range 3D tracking, and do not focus on novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.905, + 0.653 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [69] and Gaussian Splating [46] have shown strong performance in novel view synthesis. For dynamic scenes, many methods [1, 9, 10, 23, 23, 55, 59, 89, 92, 100] still require simultaneous multi-view video observations or predefined templates [38, 54, 107]. Template-free monocular approaches model dynamic scenes with different types of representations such as video depth maps [122], time-dependent NeRFs [20, 57, 58, 74-76, 97, 109], and dynamic 3D Gaussians [21, 108, 119, 120]. Although significant progress has been made, as DyCheck [25] pointed out, many approaches focus on scenarios with camera teleportation [51, 119] or quasi-static scenes [74, 75], which are effectively multiview and do not represent real-world monocular videos. In this work, we focus on casually-captured monocular videos, a more practical and challenging setup. Recent and concurrent efforts [12, 53, 62, 91, 104] have also investigated similar settings, using strong off-the-shelf data-driven priors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.605, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our method takes as input a sequence of \\(T\\) video frames \\(\\{I_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}\\) of a dynamic scene, the camera intrinsics \\(\\mathbf{K}_t\\in \\mathbb{R}^{3\\times 3}\\), and world-to-camera extrinsics \\(\\mathbf{E}_t\\in \\mathbb{S}\\mathbb{E}(3)\\) of each input frame \\(I_{t}\\). From these inputs, we aim to recover the geometry of the entire dynamic scene and the full-length 3D motion trajectory of every point in the scene. Unlike most prior dynamic NeRFs methods [25, 56, 58, 97, 109] which render the scene contents through volumetric ray casting and represent the motion implicitly at fixed 3D locations, we model the dense scene elements as a set of canonical 3D Gaussians, that translate and rotate over entire video as persistent motion trajectories. We adopt explicit point-based representation because it simultaneously allows for both (1) high-fidelity rendering in real-time and (2) full" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9661" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.095, + 0.868, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.382, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 2. System Overview. Given a single RGB video sequence with known camera poses, monocular depth maps and 2D tracks computed from off-the-shelf models [16, 118] as input, we optimize a dynamic scene representation as a set of persistent 3D Gaussians that translate and rotate over time. To capture the low-dimensional nature of scene motion, we model the motion with a set of compact \\(\\mathbb{SE}(3)\\) motion bases shared across all scene elements. Each 3D Gaussian's motion is represented as a linear combination of these motion bases, weighted by motion coefficients specific to each Gaussian. We supervise our scene representation (canonical 3D Gaussians, per-Gaussian motion coefficients, and global motion bases) by comparing the rendered outputs (RGB, depths and 2D tracks) with the corresponding input signals. This results in a globally coherent dynamic 3D scene representation in the world coordinate with long-range 3D trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.504 + ], + "angle": 0, + "content": "length 3D tracking of any surface point from any input time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Optimizing an explicit representation of dynamic 3D Gaussians from a single video is severely ill-posed — at each point in time, the moving subjects in the scene are observed from only a single viewpoint. In order to overcome this ambiguity, we make two insights: First, while the projected 2D dynamics might be complex in the video, the underlying 3D motion in the scene is low-dimensional, and composed of simpler units of rigid motion. Second, powerful data-driven priors, namely monocular depth estimates and long-range 2D tracks, provide complementary but noisy signals of the underlying 3D scene. We propose a system that fuses these noisy estimates together into a globally coherent representation of both the scene geometry and motion. We show a schematic of our pipeline in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.416, + 0.743 + ], + "angle": 0, + "content": "3.1. Preliminaries: 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We represent the appearance and geometry of a dynamic scene with a global set of 3D Gaussians, an explicit and expressive differentiable scene representation [45] for efficient optimization and rendering. We define parameters of each 3D Gaussian in the canonical frame \\( t_0 \\) as \\( \\vec{g}_0 \\equiv (\\pmb{\\mu}_0, \\mathbf{R}_0, \\mathbf{s}, o, \\mathbf{c}) \\), where \\( \\pmb{\\mu}_0 \\in \\mathbb{R}^3 \\), \\( \\mathbf{R}_0 \\in \\mathbb{SO}(3) \\) are the 3D mean and orientation in the canonical frame, and \\( \\mathbf{s} \\in \\mathbb{R}^3 \\) the scale, \\( o \\in \\mathbb{R} \\) the opacity, and \\( \\mathbf{c} \\in \\mathbb{R}^3 \\) the color. \\( \\Sigma_0 \\) is the canonical frame covariance matrix. These properties are persistent and shared across time. To render a set of 3D" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Gaussians from a camera with world-to-camera extrinsics \\(\\mathbf{E}\\) and intrinsics \\(\\mathbf{K}\\), the projections of the 3D Gaussians in the image plane are obtained by 2D Gaussians parameterized as \\(\\pmb{\\mu}_0^{\\prime} \\in \\mathbb{R}^{2}\\) and \\(\\pmb{\\Sigma}_0^{\\prime} \\in \\mathbb{R}^{2 \\times 2}\\) via affine approximation" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.574, + 0.8, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\Pi (\\mathbf {K E} \\boldsymbol {\\mu} _ {0}) \\in \\mathbb {R} ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.594, + 0.907, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\mathbf {J} _ {\\mathbf {K E}} \\boldsymbol {\\Sigma} _ {0} \\mathbf {J} _ {\\mathbf {K E}} ^ {\\top} \\in \\mathbb {R} ^ {2 \\times 2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.687 + ], + "angle": 0, + "content": "where \\(\\Pi\\) is perspective projection, and \\(\\mathbf{J}_{\\mathbf{KE}}\\) is the Jacobian of perspective projection with \\(\\mathbf{K}\\) and \\(\\mathbf{E}\\) at the point \\(\\mu_0\\). The 2D Gaussians can then be efficiently rasterized into RGB image and depth map via alpha compositing as" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.697, + 0.906, + 0.733 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {I}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {D}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {d} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.824 + ], + "angle": 0, + "content": "where \\(\\alpha_{i} = o_{i}\\cdot \\exp \\left(-\\frac{1}{2} (\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})^{T}\\boldsymbol{\\Sigma}_{0}^{\\prime}(\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})\\right)\\), and \\(T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})\\). \\(H(\\mathbf{p})\\) is the set of Gaussians that intersect the ray shoot from the pixel \\(\\mathbf{p}\\). This process is fully differentiable, and enables direct optimization of the 3D Gaussian parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.788, + 0.85 + ], + "angle": 0, + "content": "3.2. Dynamic Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Scene Motion Parameterization. To model a dynamic 3D scene, we keep track of \\( N \\) canonical 3D Gaussians and vary their positions and orientations over time with a" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9662" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "per frame rigid transformation. In particular, for a moving 3D Gaussian at time \\(t\\), its pose parameters \\((\\pmb{\\mu}_t,\\mathbf{R}_t)\\) are rigidly transformed from the canonical frame \\(t_0\\) to \\(t\\) via \\(\\mathbf{T}_{0\\rightarrow t} = \\left[\\mathbf{R}_{0\\rightarrow t}\\quad \\mathbf{t}_{0\\rightarrow t}\\right]\\in \\mathbb{S}\\mathbb{E}(3)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.162, + 0.483, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\boldsymbol {\\mu} _ {0} + \\mathbf {t} _ {0 \\rightarrow t}, \\quad \\mathbf {R} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\mathbf {R} _ {0}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Rather than modeling the 3D motion trajectories independently for each Gaussian, we define a set of \\(B \\ll N\\) learnable basis trajectories \\(\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^{B}\\) that are globally shared across all Gaussians [51]. The transformation \\(\\mathbf{T}_{0\\rightarrow t}\\) at each time \\(t\\) is then computed by a weighted combination of this global set of basis trajectories through per-point basis coefficients \\(\\mathbf{w}^{(b)}\\) via" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.3, + 0.483, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {0 \\rightarrow t} = \\sum_ {b = 0} ^ {B} \\mathbf {w} ^ {(b)} \\mathbf {T} _ {0 \\rightarrow t} ^ {(b)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.485, + 0.49 + ], + "angle": 0, + "content": "where \\(\\| \\mathbf{w}^{(b)}\\| = 1\\). In our implementation, we parameterize \\(\\mathbf{T}_{0\\rightarrow t}^{(b)}\\) as 6D rotation [31] and translation, and perform the weighted combination separately on each with the same weight \\(\\mathbf{w}^{(b)}\\). During optimization, we jointly learn the set of global motion bases and motion coefficients of each 3D Gaussian. These compact motion bases explicitly regularize the trajectories to be low-dimensional, encouraging the 3D Gaussians that move similarly to each other to be represented by similar motion coefficients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Rasterizing 3D Trajectories. Given this representation, we now describe how we obtain pixelwise 3D motion trajectory at any query frame \\(I_{t}\\). We take a similar approach to Wang et al. [101] and rasterize the motion trajectories of 3D Gaussians into query frame \\(I_{t}\\). Namely, for a query camera at time \\(t\\) with intrinsics \\(\\mathbf{K}_t\\) and extrinsics \\(\\mathbf{E}_t\\), we perform rasterization to obtain a map \\({}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'} \\in \\mathcal{R}^{H\\times W\\times 3}\\) that contains the expected 3D world coordinates of the surface points corresponding to each pixel at target time \\(t'\\)" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.64, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\n{ } ^ { \\mathrm { w } } \\hat { \\mathbf { X } } _ { t \\rightarrow t ^ { \\prime } } ( \\mathbf { p } ) = \\sum _ { i \\in H ( \\mathbf { p } ) } T _ { i } \\alpha _ { i } \\boldsymbol { \\mu } _ { i , t ^ { \\prime } } , ( 5 )\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.484, + 0.746 + ], + "angle": 0, + "content": "where \\( H(\\mathbf{p}) \\) is the set of Gaussians that intersect the pixel \\( \\mathbf{p} \\) at query time \\( t \\). The 2D correspondence location at time \\( t' \\) for a given pixel \\( \\mathbf{p} \\), \\( \\hat{\\mathbf{U}}_{t\\rightarrow t'}(\\mathbf{p}) \\), and the corresponding depth value at time \\( t' \\), \\( \\hat{\\mathbf{D}}_{t\\rightarrow t'}(\\mathbf{p}) \\) can then be written as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.753, + 0.508, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\Pi \\left(\\mathbf {K} _ {t ^ {\\prime}} ^ {\\mathrm {c}} \\hat {\\mathbf {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p})\\right), \\hat {\\mathbf {D}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\binom {\\mathrm {c}} {\\mathrm {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.484, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathrm{^c}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p}) = \\mathbf{E}_{t'}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p}),\\) \\(\\Pi\\) is a perspective projection operation, and \\((\\cdot)_{[3]}\\) is the third element of a vector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.231, + 0.85 + ], + "angle": 0, + "content": "3.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We prepare the following estimates using off-the-shelf methods in our optimization: 1) masks for the moving objects for each frame \\(\\{\\mathbf{M}_t\\}\\), which can be easily obtained" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.35 + ], + "angle": 0, + "content": "using Track-Anything [48, 117] with a few user clicks, 2) monocular depth maps \\(\\{\\mathbf{D}_t\\}\\) computed using state-of-the-art relative depth estimation method such as Depth Anything [118] and 3) long-range 2D tracks \\(\\{\\mathbf{U}_{t\\rightarrow t^{\\prime}}\\}\\) for foreground pixels from state-of-the-art point tracking method TAPIR [16]. We align the relative depth maps with the metric depth maps by computing a per-frame global scale and shift and use them for our optimization, as we found relative depth maps tend to contain finer details. We treat the lifted 2D tracks unprojected with the aligned depth maps as noisy initial 3D track observations \\(\\{\\mathbf{X}_t\\}\\) for the moving objects. For the static part of the scene, we model them using standard static 3D Gaussians and initialize their 3D locations by unprojecting them into the 3D space using the aligned depth maps. The static and dynamic Gaussians are jointly optimized and rasterized together to form an image. We describe the optimization process below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.908, + 0.642 + ], + "angle": 0, + "content": "Initialization. We first select the canonical frame \\( t_0 \\) to be the frame in which the most 3D tracks are visible, and initialize the Gaussian means in the canonical frame \\( \\mu_0 \\) as \\( N \\) randomly sampled 3D tracks locations from this set of initial observations. We then perform k-means clustering on the vectorized velocities of the noisy 3D tracks from every frame \\( \\{\\mathbf{X}_t\\} \\), and initialize the motion bases \\( \\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B \\) from these \\( B \\) clusters of tracks. Specifically, for the set of trajectories \\( \\{\\mathbf{X}_t\\}_b \\) belonging to cluster \\( b \\), we initialize the basis transform \\( \\mathbf{T}_{0\\rightarrow \\tau}^{(b)} \\) using weighted Procrustes alignment between the point sets \\( \\{\\mathbf{X}_0\\}_b \\) and \\( \\{\\mathbf{X}_{\\tau}\\}_b \\) for all \\( \\tau = 0,\\dots ,T \\), where the weights are computed using uncertainty and visibility scores from TAPIR predictions. We initialize \\( \\mathbf{w}^{(b)} \\) of each Gaussian to decay exponentially with its distance from the center of cluster \\( b \\) in the canonical frame. We then optimize \\( \\mu_0 \\), \\( \\mathbf{w}^{(b)} \\), and set of basis functions \\( \\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B \\) to fit the observed 3D tracks with an \\( \\ell_1 \\)-loss under temporal smoothness constraints." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.647, + 0.909, + 0.785 + ], + "angle": 0, + "content": "Training. We supervise the dynamic Gaussians with two sets of losses. The first set comprise our reconstruction loss to match the per-frame pixelwise color, depth, and masks inputs. The second set enforce consistency of correspondences across time. Specifically, during each training step, we render the image \\(\\hat{\\mathbf{I}}_t\\), depth \\(\\hat{\\mathbf{D}}_t\\), and mask \\(\\hat{\\mathbf{M}}_t\\) from their corresponding training cameras \\((\\mathbf{K}_t,\\mathbf{E}_t)\\) according to Equation 2. We supervise these predictions with a reconstruction loss applied independently per-frame" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.794, + 0.907, + 0.813 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e c o n}} = \\left\\| \\hat {\\mathbf {I}} - \\mathbf {I} \\right\\| _ {1} + \\lambda_ {\\text {d e p t h}} \\left\\| \\hat {\\mathbf {D}} - \\mathbf {D} \\right\\| _ {1} + \\lambda_ {\\text {m a s k}} \\left\\| \\hat {\\mathbf {M}} - \\mathbf {M} \\right\\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The second set of losses supervises the motion of the Gaussians between frames. Specifically, we additionally render the 2D tracks \\(\\hat{\\mathbf{u}}_{t\\rightarrow t^{\\prime}}\\) and reprojected depths \\(\\hat{\\mathbf{D}}_{t\\rightarrow t^{\\prime}}\\) for a pair of randomly sampled query time \\(t\\) and target time \\(t^\\prime\\). We supervise these rendered correspondences with the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9663" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.318, + 0.106 + ], + "angle": 0, + "content": "long-range 2D track estimates via" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.123, + 0.482, + 0.141 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t r a c k - 2 d}} = \\left\\| \\mathbf {U} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} \\right\\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.144, + 0.483, + 0.163 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t r a c k - d e p t h}} = \\left\\| \\hat {\\mathbf {d}} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {D}} \\left(\\mathbf {U} _ {t \\rightarrow t ^ {\\prime}}\\right)\\right\\| _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.484, + 0.271 + ], + "angle": 0, + "content": "Finally, we leverage physics motion prior by enforcing a distance-preserving loss between randomly sampled dynamic Gaussians and their k-nearest neighbors. Let \\(\\hat{\\mathbf{X}}_t\\) and \\(\\hat{\\mathbf{X}}_{t'}\\) denote the location of a Gaussian at time \\(t\\) and \\(t'\\), and \\(\\mathcal{C}_k(\\hat{\\mathbf{X}}_t)\\) denote the set of k-nearest neighbors of \\(\\hat{\\mathbf{X}}_t\\), we define" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.286, + 0.483, + 0.327 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r i g i d i t y}} = \\left\\| \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t}\\right)\\right) - \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.327, + 0.388, + 0.342 + ], + "angle": 0, + "content": "where \\(\\mathrm{dist}(\\cdot ,\\cdot)\\) measures Euclidean distance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.484, + 0.472 + ], + "angle": 0, + "content": "Implementation Details. For in-the-wild videos, multiple methods [60, 95] are available to estimate camera parameters; further details are provided in the supplement. During reconstruction, camera extrinsics are refined as learnable parameters, update via camera gradient function in gsplat [121] CUDA kernel. For evaluating our methods on public benchmark, we use the camera annotations that come with the datasets (e.g., from COLMAP [86] or simulation)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.656 + ], + "angle": 0, + "content": "We optimize our model using Adam Optimizer [47]. We perform 1000 iterations of optimization for the initialization phase and 500 epochs for training phase, respectively. The number of \\(\\mathbb{SE}(3)\\) bases \\(B\\) is set to 10 for all of our experiments. We initialize 40k dynamic Gaussians for the dynamic part and 100k static Gaussians for the static part of the scene, respectively. We perform the same adaptive Gaussian controls for dynamic and static Gaussians as per 3D-GS [46]. Training on a sequence of 300 frames of \\(960 \\times 720\\) resolution takes about 2 hours to finish on an A100 GPU. Our rendering FPS is around 140 fps. We use the same set of hyper-parameters for all experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.677, + 0.216, + 0.695 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate our performance quantitatively and qualitatively on a broad range of tasks: long range 3D point tracking, long-range 2D point tracking, and novel view synthesis. We focus our evaluation on datasets that exhibit substantial scene motion. In particular, the iPhone dataset [25] features casual captures of real-world scenes that closely match our target scenarios. It provides comprehensive annotations, including simultaneous validation views, lidar depth, sparse 2D point correspondences across the entire video, and can be used to evaluate our performance on all three tasks. Given the challenge of obtaining precise 3D track annotations for real data, we also evaluate performance using the scenes from the synthetic MOVi-F Kubric dataset [28]." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.088, + 0.91, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Figure 3. 3D Track visualization on iPhone dataset. We render novel views for each method and overlay their predicted 3D tracks on top of the rendered images. For clarity, we only display a segment of the trails spanning 50 frames for a specific set of grid query points. However, it is important to note that our method can generate dense, full-length 3D tracks. *Note that TAPIR + DA cannot produce novel view rendering, and we overlay their tracking results with our rendering to make it easier to interpret." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.69, + 0.501 + ], + "angle": 0, + "content": "4.1. Task Specification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.667 + ], + "angle": 0, + "content": "Long-Range 3D Tracking. Our primary task is estimating 3D scene motion for any pixel over the entire video. For this purpose, we extend the metrics for scene flow evaluation introduced in RAFT-3D [96] to evaluate long-range 3D tracking. Specifically, we report the 3D end-point-error (EPE), which measures the Euclidean distance between the ground truth 3D tracks and predicted tracks at each target time step. In addition, we report the percentage of points that fall within a given threshold of the ground truth 3D location \\(\\delta_{3D}^{0.05} = 5\\mathrm{cm}\\) and \\(\\delta_{3D}^{10} = 10\\mathrm{cm}\\) in metric scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.909, + 0.784 + ], + "angle": 0, + "content": "Long-Range 2D Tracking. Our 3D motion representation can be easily projected onto image plane to get long-range 2D tracks. Thus we also evaluate 2D tracking performance in terms of both position and occlusion accuracy. Following the metrics introduced in the TAP-Vid benchmark [15], we report the Average Jaccard (AJ), average position accuracy \\((<\\delta_{\\mathrm{avg}})\\), and Occlusion Accuracy (OA)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Novel View Synthesis. We measure our method's novel view synthesis quality as a comprehensive assessment for geometry, appearance, and motion. We evaluate on the iPhone dataset [25] which provides validation views and report co-visibility masked image metrics [25]: mPSNR, mSSIM [105] and mLPIPS [26, 36]. Additionally, we evaluate on NVIDIA dataset for more comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9664" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.237, + 0.089, + 0.768, + 0.26 + ], + "angle": 0, + "content": "
Method3D Tracking2D TrackingView Synthesis
EPE ↓δ053D↑δ103D↑AJ ↑<δavg↑OA ↑PSNR ↑SSIM ↑LPIPS ↓
T-NeRF [25]------15.600.550.55
HyperNeRF [75]0.18228.445.810.119.352.015.990.590.51
DynIBaR [58]0.25211.424.65.48.737.713.410.480.55
Deformable-3D-GS [119]0.15133.455.314.020.963.911.920.490.66
DynMF [51]0.18822.953.85.59.560.516.540.590.49
CoTracker [44]+DA [118]0.20234.357.924.133.973.0---
TAPIR [16]+DA [118]0.11438.163.227.841.567.4---
DELTA (world) [72]0.15932.555.324.734.168.9---
SpatialTracker (world) [110]0.12537.763.924.936.973.5---
Ours0.08243.073.334.447.086.616.720.630.45
Ours + 2DGS[33]0.09747.371.335.847.087.316.750.650.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.268, + 0.908, + 0.338 + ], + "angle": 0, + "content": "Table 1. Evaluation on iPhone dataset. Our method achieves SOTA performance all tasks of 3D point tracking, 2D point tracking, and novel view synthesis. The baselines that perform best on 2D and 3D tracking (TAPIR [16] + DA [118], CoTracker [44] + DA [118], DELTA [72], SpatialTracker [110]) are unable to synthesize novel views of the scene, while the methods that perform best in novel view synthesis struggle with or fail to produce 2D and 3D tracks. Our method achieves a significant boost in all three tasks above baselines. We include training details about “Ours + 2DGS [33]” in the supplement." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.34, + 0.907, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.706, + 0.912, + 0.751 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of novel view synthesis on iPhone dataset. The leftmost image in each row shows the training view at the same time step as the validation view. The regions highlighted in green indicate areas excluded from evaluation due to the lack of co-visibility between training and validation views. Please see the supplemental video for more qualitative results and comparisons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.2, + 0.777 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Our method represents dynamic 3D scene comprehensively with explicit long-range 3D scene motion estimation, which also allows for novel view synthesis. While no existing method achieves the exact same goals as ours, there are methods that focus on sub-tasks related to our problem, such as dynamic novel view synthesis, 2D tracking, or monocular depth estimation. We therefore adapt existing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.762, + 0.871, + 0.776 + ], + "angle": 0, + "content": "methods as our baselines which are introduced below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While dynamic novel view synthesis approaches focus primarily on the photometric reconstruction quality and do not explicitly output 3D point tracks, we can adapt some of their representations to derive 3D point tracks for our evaluation. For HyperNeRF [75], we compose the learned inverse mapping (from view space to canonical space) and a forward mapping solved via root-finding [11, 25] to produce 3D tracks at query points. DynIBaR [58] produces" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9665" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.48, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.193, + 0.483, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Visualization of motion coefficients after PCA and predicted depth maps on iPhone dataset. The motion coefficients encode information regarding rigid moving components. For instance, our motion coefficient PCA produces constant color for the block in the second example which exhibits rigid motion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.278, + 0.483, + 0.502 + ], + "angle": 0, + "content": "short-range view-to-view scene flow, which we chain into long-range 3D tracks for our evaluation. Deformable-3D-GS (D-3DGS) [119] represents dynamic scenes using 3D Gaussians [45] in the canonical space and a deformation MLP network that deforms the canonical 3D Gaussians into each view, which naturally allows 3D motion computation. T-NeRF [25] models dynamic scenes using time as MLP input in addition to 3D locations, which does not provide a method for extracting 3D motion, and hence they are not considered for 2D/3D tracking evaluation. Finally, we implemented our own version of DynMF [51] with neural network motion bases. Implementation details are provided in the supplemental material. We evaluate the 3D tracks of these methods only on the iPhone dataset, because we found none can handle the Kubric scene motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Alongside dynamic view synthesis baselines, we include 3D tracking baselines. Specifically, we take state-of-the-art long-range 2D tracks from TAPIR [16] and CoTracker [44] and lift them into 3D scene motion using monocular depth maps produced by Depth Anything (DA) [118]. We compute the correct scale and shift for each relative depth map from Depth Anything to align them with the scene. The two resulting baselines are called \"CoTracker \\([44] + DA[118]\\)\" and \"TAPIR \\([16] + DA[118]\\)\". Note that these baselines can only produce 3D tracks for visible regions, as the depth values for occluded points are unknown from such a 2.5D representation. In contrast, our global representation allows for modeling 3D motion through occlusions. In addition, we include frame-space 3D tracking methods such as DELTA [72] and SpatialTracker [110], and transform their trajectories into world coordinates using annotated camera poses for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.355, + 0.787 + ], + "angle": 0, + "content": "4.3. Evaluation on iPhone Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "iPhone dataset [25] contains 14 sequences of 200-500 frames featuring various types of challenging real world scene motion. All sequences are recorded using a handheld moving camera in a casual manner, with 7 of them additionally featuring two synchronized static cameras with a large baseline for novel view synthesis evaluation. For 3D tracking evaluation, we generate the groundtruth 3D tracks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.241 + ], + "angle": 0, + "content": "by lifting the 2D keypoint annotations into 3D using lidar depth, and masking out points that are occluded or with invalid lidar depth values. All experiments are conducted with the original instead of half resolution as in [25] given that our method can handle high-res video input. We also discard scenes SPACE-OUT and WHEEL due to camera and lidar error. We find that the original camera annotations from ARKit is not accurate enough, and refine them using global bundle adjustment from COLMAP [85]. The refined poses are used for all methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.483 + ], + "angle": 0, + "content": "We report the quantitative comparison in Tab. 1 which shows that our method outperforms all baselines on all tasks by a substantial margin. For tracking, we observe significant improvements over naive baselines that combine 2D tracks with depth maps, especially \"TAPIR+DA\", which also serves as input to our system, highlighting the effectiveness of our consolidation process. The improvement stems from three factors: 1) a low-dimensional \\(\\mathbb{SE}(3)\\) motion representation, which allows inferring occluded motion via nearby visible regions; 2) 3D regularizations, such as low acceleration constraints that promote coherent trajectories; 3) the consolidation of all pairwise tracks into a single 4D representation with persistent 3D geometry, which serves as a structural prior to correct noisy 2D tracks. Our method also achieves the best novel view synthesis quality across dynamic NeRF and 3D-GS-based baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Fig. 3 shows qualitative comparison of the 3D tracking results. We render the novel views and plot the predicted 3D tracks of the given query points onto the novel views. Since \"TAPIR + DA\" cannot perform novel view synthesis, we overlay their track predictions onto our renderings to aid interpretation. D-3DGS [119] and HyperNeRF [75] fail to capture the significant scene motion in PAPER-WINDMILL and SPIN, resulting in structure degradation and blurry rendering. \"TAPIR + DA\" can track large motions, but their 3D tracks tend to be noisy and erroneous. In contrast, our method not only generates high-quality novel views but also the most smooth and accurate 3D tracks. Fig. 4 provides additional novel view synthesis comparison on the validation views. In Fig. 5, we visualize the rendering of the first three components of PCA decomposition for the motion coefficients, which correlates well with the rigid groups of the moving objects." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.75, + 0.78, + 0.765 + ], + "angle": 0, + "content": "4.4. Evaluation on Kubric Dataset" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.782, + 0.859, + 0.848 + ], + "angle": 0, + "content": "
MethodEPE↓δ0.053D↑δ103D↑
CoTracker [44]+DA [118]0.1934.456.5
TAPIR [16]+DA [118]0.2034.056.2
Ours0.1639.862.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.537, + 0.858, + 0.86, + 0.871 + ], + "angle": 0, + "content": "Table 2. 3D Tracking evaluation on Kubric dataset." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The Kubric MOVi-F dataset contains short 24-frame videos" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "of scenes of 10-20 objects, rendered with linear camera movement and motion blur. Multiple rigid objects are tossed onto the scene, at a speed that far exceeds the speed of the moving camera, making it a similar capture scenario to in-the-wild capture scenarios. It provides dense comprehensive annotations, including ground truth depth maps, camera parameters, segmentation masks, and point correspondences that are dense across time. We use 30 scenes from the MOVi-F validation set to evaluate the accuracy of our long-range 3D point tracks for all points in time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.378 + ], + "angle": 0, + "content": "We demonstrate our method on the synthetic Kubric MOVi-F dataset in long-range 3D point tracking across time. Of the above baselines, only \"CoTracker [44] + DA [118]\" and \"TAPIR [44] + DA [118]\" yield 3D tracks for these scenes. For all baselines, we provide the ground truth camera intrinsics and extrinsics, and monocular depth estimates that have been aligned to the ground truth depth map. We obtain point tracks for all non-background pixels for each method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.482, + 0.454 + ], + "angle": 0, + "content": "We report our quantitative 3D point tracking metrics in Table 2, and find that across all metrics, our method outperforms the baselines. Please see the supplement for figures of the optimized motion coefficients, which coherently group moving objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.367, + 0.479 + ], + "angle": 0, + "content": "4.5. Evaluation on NVIDIA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.482, + 0.637 + ], + "angle": 0, + "content": "We conduct experiments on seven scenes from the NVIDIA dataset [122], following the evaluation protocol of Dynamic Gaussian Marbles [91], which evaluates the task from a single static camera instead of interleaving different camera as discussed in [25]. We adopt the same image resolution, camera parameters and depths maps as input. Please see the supplement for more details. For each scene, we compute covisibility masks between each evaluation view and the training view at each time step to exclude unobservable regions during evaluation. Results are reported in Tab. 3." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.65, + 0.482, + 0.724 + ], + "angle": 0, + "content": "
MethodBalloon1Ballon2JumpingTruck
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]23.830.810.06623.620.810.08820.020.730.13427.310.860.055
Ours23.900.790.05823.910.830.07420.080.690.15127.650.880.048
MethodPlaygroundUmbrellaSkatingMean
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]16.680.600.17724.990.700.08327.790.900.05223.460.770.093
Ours16.870.550.16823.270.570.16727.910.920.04723.370.750.102
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.733, + 0.482, + 0.762 + ], + "angle": 0, + "content": "Table 3. Evaluation on NVIDIA dataset. Our method is comparable with Dynamic Gaussian Marbles (DGM) [91]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.251, + 0.803 + ], + "angle": 0, + "content": "4.6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We ablate various components of our method on the iPhone dataset in 3D tracking in Tab. 4. We first validate our choices of motion representation, namely the \\(\\mathbb{SE}(3)\\) motion bases parameterization, with three ablations: 1) \"Per-Gaussian Transl\": replacing our motion representation with naive per-Gaussian translational motion trajectories," + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.905, + 0.165 + ], + "angle": 0, + "content": "
MethodsSE(3)Motion Basis2D tracksInitializationEPE↓δ0.053D↑δ1.103D↑
Ours (Full)✓✓✓✓0.08243.073.3
Transl. Bases✓✓✓0.09342.369.9
Per-Gaussian SE(3)✓✓✓0.08343.670.2
Per-Gaussian Transl.✓✓0.08741.269.2
No SE(3) Init.✓✓✓0.11139.365.7
No 2D Tracks✓✓0.14130.457.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.176, + 0.848, + 0.189 + ], + "angle": 0, + "content": "Table 4. Ablation Studies on iPhone dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.905, + 0.329 + ], + "angle": 0, + "content": "2) \"Per-Gaussian \\(\\mathbb{SE}(3)\\): replacing our motion representation with naive per-Gaussian \\(\\mathbb{SE}(3)\\) motion trajectories, and 2) \"Transl. Bases\": keeping the motion bases representation but only using translational bases instead of \\(\\mathbb{SE}(3)\\). We find that \\(\\mathbb{SE}(3)\\) bases significantly improve 3D tracking over both translational bases and per-Gaussian translational motion. It also outperforms per-Gaussian \\(\\mathbb{SE}(3)\\) in overall accuracy and offers noticeably better visual quality, as the latter exhibits significantly more popping and jitter artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Next, we ablate our training strategies including initialization and supervision signal. We conduct an ablation of \"No \\(\\mathbb{SE}(3)\\) Init.\", where instead of performing our initial \\(\\mathbb{SE}(3)\\) fitting stage, we initialize the translational part of the motion bases with randomly selected noisy 3D tracks formed by directly lifting input 2D tracks using depth maps into 3D, and the rotation part as identity. We find that skipping this initialization noticeably hurts performance. Lastly, we remove the 2D track supervision entirely (\"No 2D Tracks\") and find it to lead to significant drop in performance, which verifies the importance of the 2D track supervision for 3D tracking." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.526, + 0.627, + 0.541 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Limitations. Like most prior monocular view synthesis methods, it still requires per-scene test-time optimization, hindering streamable applications. Recent feed-forward methods for joint reconstruction and tracking [22, 111, 113] offer a promising alternative, showing encouraging progress toward overcoming this limitation. Our method relies on off-the-shelf predictions for camera poses, geometry, and motion, which may degrade in textureless regions or under large motions. However, it also naturally benefits from advances in these components. Finally, it also requires user input to mask moving objects. Recent advances in moving object segmentation [27, 34] offer promising directions beyond our current manual solution and could be integrated into our pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Conclusion. We introduce a method for joint long-range 3D tracking and novel view synthesis from monocular video, representing dynamic elements via global 3D Gaussians with time-varying translations and rotations. Motion trajectories are regularized using low-dimensional rigid motion bases. By consolidating noisy observations into globally consistent scene estimates, our approach outperforms state-of-the-art methods in 2D/3D tracking and novel view synthesis across synthetic and real benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.204 + ], + "angle": 0, + "content": "Acknowledgement. We thank Ruilong Li, Noah Snavely, Brent Yi and Aleksander Holynski for helpful discussion. We are in memory of our beloved cat Sriracha, who will always be missed and loved. This project is supported in part by DARPA No. HR001123C0021. and IARPA DOI/IBC No. 140D0423C0035. The views and conclusions contained herein are those of the authors and do not represent the official policies or endorsements of these institutions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.258, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[1] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.329, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.373, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[3] Stanley T Birchfield and Shrinivas J Pundlik. Joint tracking of features and edges. In 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.416, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[4] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.46, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[5] Aljaz Bozic, Michael Zollhöfer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[6] C. Bregler, A. Hertzmann, and H. Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No.PR00662), 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[7] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In European Conference on Computer Vision, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.645, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[8] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.688, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[9] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[10] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[11] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.888, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Dreamscene4d: Dynamic multi-object scene generation from monocular videos. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. Int. J. Comput. Vision, 107(2), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.164, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.304, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.387, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[17] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Joao Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.457, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Häusser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.54, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[19] Mingsong Dou, S. Khamis, Yu.G. Degtyarev, Philip L. Davidson, S. Fanello, Adarsh Kowdle, Sergio Orts, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4d. ACM Transactions on Graphics (TOG), 35, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.624, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.694, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[21] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J Black, Trevor Darrell, and Angjoo Kanazawa. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. arXiv preprint arXiv:2504.13152, 2025.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. Kplanes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9668" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[25] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Dynamic novel-view synthesis: A reality check. In NeurIPS, 2022. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.191, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[26] Leon A Gatys, Alexander S Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[27] Lily Goli, Sara Sabour, Mark Matthews, Marcus Brubaker, Dmitry Lagun, Alec Jacobson, David J Fleet, Saurabh Saxena, and Andrea Tagliasacchi. Romo: Robust motion segmentation improves structure from motion. arXiv preprint arXiv:2411.18650, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[28] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[29] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[30] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[31] Thorsten Hempel, Ahmed A Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In 2022 IEEE International Conference on Image Processing (ICIP), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.596, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[32] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. In Other Conferences, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.624, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[33] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[34] Nan Huang, Wenzhao Zheng, Chenfeng Xu, Kurt Keutzer, Shanghang Zhang, Angjoo Kanazawa, and Qianqian Wang. Segment any motion in videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[35] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. In European Conference on Computer Vision. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.819, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[36] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[37] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[40] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European Conference on Computer Vision, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[41] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[43] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[44] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European conference on computer vision, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[45] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.633, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[46] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.689, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[47] Diederik P Kingma, J Adam Ba, and J Adam. A method for stochastic optimization. arxiv 2014. arXiv preprint arXiv:1412.6980, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.733, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[48] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[49] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, Joao Carreira, Andrew Zisserman, Gabriel Brostow, and" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9669" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[51] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. ECCV, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.193, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[52] Suryansh Kumar, Yuchao Dai, and Hongdong Li. Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[53] Jiahui Lei, Yijia Weng, Adam W Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[54] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[55] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[56] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[59] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[60] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[61] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[62] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "namic gaussian splatting from causally-captured monocular videos. ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[63] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[64] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.222, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[65] Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[66] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[67] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[68] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[69] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[70] Michal Neoral, Jonás Šerych, and Jiří Matas. Mft: Long-term tracking of every pixel. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[71] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.619, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[72] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[73] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[74] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[75] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. SIGGRAPH Asia, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[76] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9670" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[77] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[78] Rene Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[79] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multi-frame optical flow estimation. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[80] Michael Rubinstein and Ce Liu. Towards longer long-range motion trajectories. In *British Machine Vision Conference*, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[81] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R. Bradski. Orb: An efficient alternative to sift or surf. 2011 International Conference on Computer Vision, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[82] Chris Russell, Rui Yu, and Lourdes Agapito. Video popu- up: Monocular 3d reconstruction of dynamic scenes. In European conference on computer vision. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[83] P Sand. Long-range video motion estimation using point trajectories. PhD thesis, Ph.D. dissertation, Cambridge, MA, USA, 2006, adviser-Teller, Seth, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[84] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[85] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[86] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[87] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[88] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. International Journal of Computer Vision, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[89] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[90] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[91] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. Siggraph Asia, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.135, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[92] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum. Wiley Online Library, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[93] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[94] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.303, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[95] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.345, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[96] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.401, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[97] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.442, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[98] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.498, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[99] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. 2013 IEEE International Conference on Computer Vision, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[100] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[101] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In International Conference on Computer Vision, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[102] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[103] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[104] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[105] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9671" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[106] Zirui Wang, Shuda Li, Henry Howard-Jenkins, Victor Adrian Prisacariu, and Min Chen. Flownet3d++: Geometric losses for deep scene flow estimation. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.244 + ], + "angle": 0, + "content": "[107] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[108] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[109] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.485, + 0.426 + ], + "angle": 0, + "content": "[110] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[111] Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Yuri Makarov, Bingyi Kang, Xing Zhu, Hujun Bao, Yujun Shen, and Xiaowei Zhou. Spatialtrackerv2: 3d point tracking made easy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[112] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[113] Zhen Xu, Zhengqin Li, Zhao Dong, Xiaowei Zhou, Richard Newcombe, and Zhaoyang Lv. 4dgt: Learning a 4d gaussian transformer using real-world monocular videos. arXiv preprint arXiv:2506.08015, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[114] Gengshan Yang, Deqing Sun, V. Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[115] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[116] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[117] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[118] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Un" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "leashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[119] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[120] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. *ICLR*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[121] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research, 26(34), 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[122] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[123] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arxiv:2410.03825, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[124] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[125] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[126] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[127] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew William Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (TOG), 33, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9672" + } + ] +] \ No newline at end of file diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_origin.pdf b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02b39cb4bbc1c04ea8a995829f471ff8ade95ed9 --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/a825c689-ac4d-430b-9fb3-d2a95e2ffbee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a841023dfb7608617372415cc699e6a94732c347dc468f83930adc78d77b94de +size 20561122 diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/full.md b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..13cc3aa2a0df119001ca47cee4103fef11e48967 --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/full.md @@ -0,0 +1,375 @@ +# Shape of Motion: 4D Reconstruction from a Single Video + +Qianqian Wang $^{1,2*}$ , Vickie Ye $^{1*}$ , Hang Gao $^{1*}$ , Weijia Zeng $^{3*}$ , Jake Austin $^{1}$ , Zhengqi Li $^{4}$ , Angjoo Kanazawa $^{1}$ + +$^{1}$ UC Berkeley $^{2}$ Google DeepMind $^{3}$ UC San Diego $^{4}$ Adobe Research + +![](images/6e4a5d046a457279c9ea0ce38be0ac15921887a07fc6f4ecae3ad61d9e044311.jpg) +Figure 1. Shape of Motion. Our method enables joint long-range 3D tracking and novel view synthesis from a monocular video of a complex dynamic scene. We render moving elements at a fixed viewpoint across time and visualize estimated 3D motion as colorful trajectories. These trajectories reveal distinct geometric patterns, which leads to the term "Shape of Motion". + +![](images/e3ce98b04f76bf54f69821528a0ff68dd5219d77e9ab4f88e60757429348fef9.jpg) + +# Abstract + +Monocular dynamic reconstruction is a challenging and long-standing vision problem due to the highly ill-posed nature of the task. Existing approaches depend on templates, are effective only in quasi-static scenes, or fail to model 3D motion explicitly. We introduce a method for reconstructing generic dynamic scenes, featuring explicit, persistent 3D motion trajectories in the world coordinate frame, from casually captured monocular videos. We tackle the problem with two key insights: First, we exploit the low-dimensional structure of 3D motion by representing scene motion with a compact set of $\mathbb{S}\mathbb{E}(3)$ motion bases. Each point's motion is expressed as a linear combination of these bases, facilitating soft decomposition of the scene into multiple rigidly-moving groups. Second, we take advantage of off-the-shelf data-driven priors such as monocular depth maps and long-range 2D tracks, and devise a method to effectively consolidate these noisy supervisory signals, resulting in a globally consistent representation of the dynamic scene. Experiments show that our method achieves state-of-the-art performance for both long-range 3D/2D motion estimation and novel view synthesis on dynamic scenes. Project Page: https://shape-of-motion.github.io/ + +# 1. Introduction + +Reconstructing the persistent geometry and their 3D motion across a video is crucial for understanding and interacting with the underlying physical world. While recent years have seen impressive progress in modeling static 3D scenes [45, 69], recovering the geometry and motion of complex dynamic 3D scenes, especially from a single video, remains an open challenge. A number of prior dynamic reconstruction and novel view synthesis approaches have attempted to tackle this problem. However, most methods rely on synchronized multi-view videos [9, 23, 23, 67, 108] or additional LIDAR/depth sensors [29, 63, 77, 98, 106]. Recent monocular approaches can operate on regular dynamic videos, but they typically model 3D scene motion as short-range scene flow between consecutive times [24, 57, 58] or deformation fields that map between canonical and view space [74, 75, 119], failing to capture 3D motion trajectories persistent over a video. + +The longstanding challenge for general in-the-wild videos lies in the poorly constrained nature of the 4D reconstruction problem. In this work, we tackle this challenge with two key insights. The first is that, while the image space dynamics can be complex and discontinuous, the underlying 3D motion is a composition of continuous simple rigid motions. Our second insight is that data-driven priors provide complementary, though noisy cues, that aggregate + +well into a globally coherent representation of the 3D scene geometry and motion. + +Motivated by these two insights, we represent the dynamic scene as a set of persistent 3D Gaussians, and represent their motion across the video in terms of a compact set of shared $\mathbb{S}\mathbb{E}(3)$ motion bases. Unlike traditional scene flow, which computes 3D correspondences between consecutive frames, our representation recovers a persistent 3D trajectory over the whole video, enabling long-range 3D tracking across the entire video. As the 3D trajectories produced by our method capture the geometric patterns that trace each point's movement through 3D space and time as shown in Figure 1, we refer to our approach as "Shape of Motion". We show how to fit our explicit scene representation to a general video in-the-wild, by fusing together complementary cues from two main sources: monocular depth estimates per-frame, and 2D track estimates across frames. We conduct extensive evaluations on both synthetic and real-world dynamic video datasets, and show that our proposed approach significantly outperforms prior methods in both long-range 2D/3D tracking and novel view synthesis. + +In summary, our key contributions are: (1) A 4D scene representation enabling both real-time novel view synthesis and globally consistent 3D tracking for any point at any time. (2) An optimization framework that fuses learned motion and geometry priors into a unified 4D representation from a single, casually captured monocular video. + +# 2. Related Work + +Correspondences and Tracking. While monocular 3D long range tracking remains largely unexplored, numerous approaches track in 2D image space, typically using optical flow for point correspondences. This involves estimating dense motion fields between image pairs [4, 7, 8, 18, 32, 35, 35, 39, 41-43, 66, 79, 87, 93, 94, 112]. Sparse keypoint matching methods can enable long trajectory generation [2, 14, 61, 64, 81], but these methods are primarily intended for sparse 3D reconstruction. Long-range 2D trajectory estimation for arbitrary points has been explored in earlier works, which relied on hand-crafted priors to generate motion trajectories [3, 80, 83, 84, 88, 99]. Recently, there has been a resurgence of interest in this problem, with several works showcasing impressive long-range 2D tracking results on challenging, in-the-wild videos. These approaches employ either test-time optimization where models consolidate noisy short-range motion estimates into long-term correspondences [70, 90, 101], or data-driven strategies [16, 17, 30, 44], where neural networks learn long-term correspondence estimates from synthetic data [15, 126]. While these methods effectively track any 2D point throughout a video, they lack the knowledge of underlying 3D scene geometry and motions. + +To model 3D scene motion, recent work in monocu + +lar settings has explored self-supervised learning and test-time optimization [37, 57, 58, 114-116]. More recent approaches estimate 3D trajectories in general scenes in a feedforward manner [50, 72, 110], but the predicted motion remains in the frame space, entangling object and camera motion. In contrast, our method recovers persistent, long-range 3D trajectories in the world coordinate space. + +Dynamic Reconstruction and View Synthesis. Our work also relates to dynamic 3D scene reconstruction and novel view synthesis. In non-rigid reconstruction, early methods often required RGB-D sensors [5, 19, 40, 71, 127] or strong hand-crafted priors [52, 78, 82]. A family of works [6, 13, 73] apply low-rank assumptions to the underlying motion model. Recent work has demonstrated progress toward general dynamic reconstruction either by integrating monocular depth priors [49, 56, 60, 68, 124, 125], or directly learning to perform dynamic reconstruction [65, 102, 103, 123] from data. However, these methods do not perform long-range 3D tracking, and do not focus on novel view synthesis. + +Neural Radiance Fields (NeRF) [69] and Gaussian Splating [46] have shown strong performance in novel view synthesis. For dynamic scenes, many methods [1, 9, 10, 23, 23, 55, 59, 89, 92, 100] still require simultaneous multi-view video observations or predefined templates [38, 54, 107]. Template-free monocular approaches model dynamic scenes with different types of representations such as video depth maps [122], time-dependent NeRFs [20, 57, 58, 74-76, 97, 109], and dynamic 3D Gaussians [21, 108, 119, 120]. Although significant progress has been made, as DyCheck [25] pointed out, many approaches focus on scenarios with camera teleportation [51, 119] or quasi-static scenes [74, 75], which are effectively multiview and do not represent real-world monocular videos. In this work, we focus on casually-captured monocular videos, a more practical and challenging setup. Recent and concurrent efforts [12, 53, 62, 91, 104] have also investigated similar settings, using strong off-the-shelf data-driven priors. + +# 3. Method + +Our method takes as input a sequence of $T$ video frames $\{I_t\in \mathbb{R}^{H\times W\times 3}\}$ of a dynamic scene, the camera intrinsics $\mathbf{K}_t\in \mathbb{R}^{3\times 3}$ , and world-to-camera extrinsics $\mathbf{E}_t\in \mathbb{S}\mathbb{E}(3)$ of each input frame $I_{t}$ . From these inputs, we aim to recover the geometry of the entire dynamic scene and the full-length 3D motion trajectory of every point in the scene. Unlike most prior dynamic NeRFs methods [25, 56, 58, 97, 109] which render the scene contents through volumetric ray casting and represent the motion implicitly at fixed 3D locations, we model the dense scene elements as a set of canonical 3D Gaussians, that translate and rotate over entire video as persistent motion trajectories. We adopt explicit point-based representation because it simultaneously allows for both (1) high-fidelity rendering in real-time and (2) full + +![](images/6acad7d745ddf3a9d6f7a829e41a90cfdcd4442515a6531f79b06bbad3d31478.jpg) +Figure 2. System Overview. Given a single RGB video sequence with known camera poses, monocular depth maps and 2D tracks computed from off-the-shelf models [16, 118] as input, we optimize a dynamic scene representation as a set of persistent 3D Gaussians that translate and rotate over time. To capture the low-dimensional nature of scene motion, we model the motion with a set of compact $\mathbb{SE}(3)$ motion bases shared across all scene elements. Each 3D Gaussian's motion is represented as a linear combination of these motion bases, weighted by motion coefficients specific to each Gaussian. We supervise our scene representation (canonical 3D Gaussians, per-Gaussian motion coefficients, and global motion bases) by comparing the rendered outputs (RGB, depths and 2D tracks) with the corresponding input signals. This results in a globally coherent dynamic 3D scene representation in the world coordinate with long-range 3D trajectories. + +length 3D tracking of any surface point from any input time. + +Optimizing an explicit representation of dynamic 3D Gaussians from a single video is severely ill-posed — at each point in time, the moving subjects in the scene are observed from only a single viewpoint. In order to overcome this ambiguity, we make two insights: First, while the projected 2D dynamics might be complex in the video, the underlying 3D motion in the scene is low-dimensional, and composed of simpler units of rigid motion. Second, powerful data-driven priors, namely monocular depth estimates and long-range 2D tracks, provide complementary but noisy signals of the underlying 3D scene. We propose a system that fuses these noisy estimates together into a globally coherent representation of both the scene geometry and motion. We show a schematic of our pipeline in Figure 2. + +# 3.1. Preliminaries: 3D Gaussian Splatting + +We represent the appearance and geometry of a dynamic scene with a global set of 3D Gaussians, an explicit and expressive differentiable scene representation [45] for efficient optimization and rendering. We define parameters of each 3D Gaussian in the canonical frame $t_0$ as $\vec{g}_0 \equiv (\pmb{\mu}_0, \mathbf{R}_0, \mathbf{s}, o, \mathbf{c})$ , where $\pmb{\mu}_0 \in \mathbb{R}^3$ , $\mathbf{R}_0 \in \mathbb{SO}(3)$ are the 3D mean and orientation in the canonical frame, and $\mathbf{s} \in \mathbb{R}^3$ the scale, $o \in \mathbb{R}$ the opacity, and $\mathbf{c} \in \mathbb{R}^3$ the color. $\Sigma_0$ is the canonical frame covariance matrix. These properties are persistent and shared across time. To render a set of 3D + +Gaussians from a camera with world-to-camera extrinsics $\mathbf{E}$ and intrinsics $\mathbf{K}$ , the projections of the 3D Gaussians in the image plane are obtained by 2D Gaussians parameterized as $\pmb{\mu}_0^{\prime} \in \mathbb{R}^{2}$ and $\pmb{\Sigma}_0^{\prime} \in \mathbb{R}^{2 \times 2}$ via affine approximation + +$$ +\boldsymbol {\mu} _ {0} ^ {\prime} (\mathbf {K}, \mathbf {E}) = \Pi (\mathbf {K E} \boldsymbol {\mu} _ {0}) \in \mathbb {R} ^ {2}, +$$ + +$$ +\boldsymbol {\Sigma} _ {0} ^ {\prime} (\mathbf {K}, \mathbf {E}) = \mathbf {J} _ {\mathbf {K E}} \boldsymbol {\Sigma} _ {0} \mathbf {J} _ {\mathbf {K E}} ^ {\top} \in \mathbb {R} ^ {2 \times 2}, \tag {1} +$$ + +where $\Pi$ is perspective projection, and $\mathbf{J}_{\mathbf{KE}}$ is the Jacobian of perspective projection with $\mathbf{K}$ and $\mathbf{E}$ at the point $\mu_0$ . The 2D Gaussians can then be efficiently rasterized into RGB image and depth map via alpha compositing as + +$$ +\hat {\mathbf {I}} (\mathbf {p}) = \sum_ {i \in H (\mathbf {p})} T _ {i} \alpha_ {i} \mathbf {c} _ {i}, \quad \hat {\mathbf {D}} (\mathbf {p}) = \sum_ {i \in H (\mathbf {p})} T _ {i} \alpha_ {i} \mathbf {d} _ {i}, \tag {2} +$$ + +where $\alpha_{i} = o_{i}\cdot \exp \left(-\frac{1}{2} (\mathbf{p} - \boldsymbol{\mu}_{0}^{\prime})^{T}\boldsymbol{\Sigma}_{0}^{\prime}(\mathbf{p} - \boldsymbol{\mu}_{0}^{\prime})\right)$ , and $T_{i} = \prod_{j = 1}^{i - 1}(1 - \alpha_{j})$ . $H(\mathbf{p})$ is the set of Gaussians that intersect the ray shoot from the pixel $\mathbf{p}$ . This process is fully differentiable, and enables direct optimization of the 3D Gaussian parameters. + +# 3.2. Dynamic Scene Representation + +Scene Motion Parameterization. To model a dynamic 3D scene, we keep track of $N$ canonical 3D Gaussians and vary their positions and orientations over time with a + +per frame rigid transformation. In particular, for a moving 3D Gaussian at time $t$ , its pose parameters $(\pmb{\mu}_t,\mathbf{R}_t)$ are rigidly transformed from the canonical frame $t_0$ to $t$ via $\mathbf{T}_{0\rightarrow t} = \left[\mathbf{R}_{0\rightarrow t}\quad \mathbf{t}_{0\rightarrow t}\right]\in \mathbb{S}\mathbb{E}(3)$ : + +$$ +\boldsymbol {\mu} _ {t} = \mathbf {R} _ {0 \rightarrow t} \boldsymbol {\mu} _ {0} + \mathbf {t} _ {0 \rightarrow t}, \quad \mathbf {R} _ {t} = \mathbf {R} _ {0 \rightarrow t} \mathbf {R} _ {0}. \tag {3} +$$ + +Rather than modeling the 3D motion trajectories independently for each Gaussian, we define a set of $B \ll N$ learnable basis trajectories $\{\mathbf{T}_{0\rightarrow t}^{(b)}\}_{b = 1}^{B}$ that are globally shared across all Gaussians [51]. The transformation $\mathbf{T}_{0\rightarrow t}$ at each time $t$ is then computed by a weighted combination of this global set of basis trajectories through per-point basis coefficients $\mathbf{w}^{(b)}$ via + +$$ +\mathbf {T} _ {0 \rightarrow t} = \sum_ {b = 0} ^ {B} \mathbf {w} ^ {(b)} \mathbf {T} _ {0 \rightarrow t} ^ {(b)}, \tag {4} +$$ + +where $\| \mathbf{w}^{(b)}\| = 1$ . In our implementation, we parameterize $\mathbf{T}_{0\rightarrow t}^{(b)}$ as 6D rotation [31] and translation, and perform the weighted combination separately on each with the same weight $\mathbf{w}^{(b)}$ . During optimization, we jointly learn the set of global motion bases and motion coefficients of each 3D Gaussian. These compact motion bases explicitly regularize the trajectories to be low-dimensional, encouraging the 3D Gaussians that move similarly to each other to be represented by similar motion coefficients. + +Rasterizing 3D Trajectories. Given this representation, we now describe how we obtain pixelwise 3D motion trajectory at any query frame $I_{t}$ . We take a similar approach to Wang et al. [101] and rasterize the motion trajectories of 3D Gaussians into query frame $I_{t}$ . Namely, for a query camera at time $t$ with intrinsics $\mathbf{K}_t$ and extrinsics $\mathbf{E}_t$ , we perform rasterization to obtain a map ${}^{\mathrm{w}}\hat{\mathbf{X}}_{t\rightarrow t'} \in \mathcal{R}^{H\times W\times 3}$ that contains the expected 3D world coordinates of the surface points corresponding to each pixel at target time $t'$ + +$$ +{ } ^ { \mathrm { w } } \hat { \mathbf { X } } _ { t \rightarrow t ^ { \prime } } ( \mathbf { p } ) = \sum _ { i \in H ( \mathbf { p } ) } T _ { i } \alpha _ { i } \boldsymbol { \mu } _ { i , t ^ { \prime } } , ( 5 ) +$$ + +where $H(\mathbf{p})$ is the set of Gaussians that intersect the pixel $\mathbf{p}$ at query time $t$ . The 2D correspondence location at time $t'$ for a given pixel $\mathbf{p}$ , $\hat{\mathbf{U}}_{t\rightarrow t'}(\mathbf{p})$ , and the corresponding depth value at time $t'$ , $\hat{\mathbf{D}}_{t\rightarrow t'}(\mathbf{p})$ can then be written as + +$$ +\hat {\mathbf {U}} _ {t \rightarrow t ^ {\prime}} (\mathbf {p}) = \Pi \left(\mathbf {K} _ {t ^ {\prime}} ^ {\mathrm {c}} \hat {\mathbf {X}} _ {t \rightarrow t ^ {\prime}} (\mathbf {p})\right), \hat {\mathbf {D}} _ {t \rightarrow t ^ {\prime}} (\mathbf {p}) = \binom {\mathrm {c}} {\mathrm {X}} _ {t \rightarrow t ^ {\prime}} (\mathbf {p}) \tag {6} +$$ + +where $\mathrm{^c}\hat{\mathbf{X}}_{t\rightarrow t'}(\mathbf{p}) = \mathbf{E}_{t'}^{\mathrm{w}}\hat{\mathbf{X}}_{t\rightarrow t'}(\mathbf{p}),$ $\Pi$ is a perspective projection operation, and $(\cdot)_{[3]}$ is the third element of a vector. + +# 3.3. Optimization + +We prepare the following estimates using off-the-shelf methods in our optimization: 1) masks for the moving objects for each frame $\{\mathbf{M}_t\}$ , which can be easily obtained + +using Track-Anything [48, 117] with a few user clicks, 2) monocular depth maps $\{\mathbf{D}_t\}$ computed using state-of-the-art relative depth estimation method such as Depth Anything [118] and 3) long-range 2D tracks $\{\mathbf{U}_{t\rightarrow t^{\prime}}\}$ for foreground pixels from state-of-the-art point tracking method TAPIR [16]. We align the relative depth maps with the metric depth maps by computing a per-frame global scale and shift and use them for our optimization, as we found relative depth maps tend to contain finer details. We treat the lifted 2D tracks unprojected with the aligned depth maps as noisy initial 3D track observations $\{\mathbf{X}_t\}$ for the moving objects. For the static part of the scene, we model them using standard static 3D Gaussians and initialize their 3D locations by unprojecting them into the 3D space using the aligned depth maps. The static and dynamic Gaussians are jointly optimized and rasterized together to form an image. We describe the optimization process below. + +Initialization. We first select the canonical frame $t_0$ to be the frame in which the most 3D tracks are visible, and initialize the Gaussian means in the canonical frame $\mu_0$ as $N$ randomly sampled 3D tracks locations from this set of initial observations. We then perform k-means clustering on the vectorized velocities of the noisy 3D tracks from every frame $\{\mathbf{X}_t\}$ , and initialize the motion bases $\{\mathbf{T}_{0\rightarrow t}^{(b)}\}_{b = 1}^B$ from these $B$ clusters of tracks. Specifically, for the set of trajectories $\{\mathbf{X}_t\}_b$ belonging to cluster $b$ , we initialize the basis transform $\mathbf{T}_{0\rightarrow \tau}^{(b)}$ using weighted Procrustes alignment between the point sets $\{\mathbf{X}_0\}_b$ and $\{\mathbf{X}_{\tau}\}_b$ for all $\tau = 0,\dots ,T$ , where the weights are computed using uncertainty and visibility scores from TAPIR predictions. We initialize $\mathbf{w}^{(b)}$ of each Gaussian to decay exponentially with its distance from the center of cluster $b$ in the canonical frame. We then optimize $\mu_0$ , $\mathbf{w}^{(b)}$ , and set of basis functions $\{\mathbf{T}_{0\rightarrow t}^{(b)}\}_{b = 1}^B$ to fit the observed 3D tracks with an $\ell_1$ -loss under temporal smoothness constraints. + +Training. We supervise the dynamic Gaussians with two sets of losses. The first set comprise our reconstruction loss to match the per-frame pixelwise color, depth, and masks inputs. The second set enforce consistency of correspondences across time. Specifically, during each training step, we render the image $\hat{\mathbf{I}}_t$ , depth $\hat{\mathbf{D}}_t$ , and mask $\hat{\mathbf{M}}_t$ from their corresponding training cameras $(\mathbf{K}_t,\mathbf{E}_t)$ according to Equation 2. We supervise these predictions with a reconstruction loss applied independently per-frame + +$$ +L _ {\text {r e c o n}} = \left\| \hat {\mathbf {I}} - \mathbf {I} \right\| _ {1} + \lambda_ {\text {d e p t h}} \left\| \hat {\mathbf {D}} - \mathbf {D} \right\| _ {1} + \lambda_ {\text {m a s k}} \left\| \hat {\mathbf {M}} - \mathbf {M} \right\| _ {1}. \tag {7} +$$ + +The second set of losses supervises the motion of the Gaussians between frames. Specifically, we additionally render the 2D tracks $\hat{\mathbf{u}}_{t\rightarrow t^{\prime}}$ and reprojected depths $\hat{\mathbf{D}}_{t\rightarrow t^{\prime}}$ for a pair of randomly sampled query time $t$ and target time $t^\prime$ . We supervise these rendered correspondences with the + +long-range 2D track estimates via + +$$ +L _ {\text {t r a c k - 2 d}} = \left\| \mathbf {U} _ {t \rightarrow t ^ {\prime}} - \hat {\mathbf {U}} _ {t \rightarrow t ^ {\prime}} \right\| _ {1}, \tag {8} +$$ + +$$ +L _ {\text {t r a c k - d e p t h}} = \left\| \hat {\mathbf {d}} _ {t \rightarrow t ^ {\prime}} - \hat {\mathbf {D}} \left(\mathbf {U} _ {t \rightarrow t ^ {\prime}}\right)\right\| _ {1}. \tag {9} +$$ + +Finally, we leverage physics motion prior by enforcing a distance-preserving loss between randomly sampled dynamic Gaussians and their k-nearest neighbors. Let $\hat{\mathbf{X}}_t$ and $\hat{\mathbf{X}}_{t'}$ denote the location of a Gaussian at time $t$ and $t'$ , and $\mathcal{C}_k(\hat{\mathbf{X}}_t)$ denote the set of k-nearest neighbors of $\hat{\mathbf{X}}_t$ , we define + +$$ +L _ {\text {r i g i d i t y}} = \left\| \operatorname {d i s t} \left(\hat {\mathbf {X}} _ {t}, \mathcal {C} _ {k} \left(\hat {\mathbf {X}} _ {t}\right)\right) - \operatorname {d i s t} \left(\hat {\mathbf {X}} _ {t ^ {\prime}}, \mathcal {C} _ {k} \left(\hat {\mathbf {X}} _ {t ^ {\prime}}\right)\right) \right\| _ {2} ^ {2}, \tag {10} +$$ + +where $\mathrm{dist}(\cdot ,\cdot)$ measures Euclidean distance. + +Implementation Details. For in-the-wild videos, multiple methods [60, 95] are available to estimate camera parameters; further details are provided in the supplement. During reconstruction, camera extrinsics are refined as learnable parameters, update via camera gradient function in gsplat [121] CUDA kernel. For evaluating our methods on public benchmark, we use the camera annotations that come with the datasets (e.g., from COLMAP [86] or simulation). + +We optimize our model using Adam Optimizer [47]. We perform 1000 iterations of optimization for the initialization phase and 500 epochs for training phase, respectively. The number of $\mathbb{SE}(3)$ bases $B$ is set to 10 for all of our experiments. We initialize 40k dynamic Gaussians for the dynamic part and 100k static Gaussians for the static part of the scene, respectively. We perform the same adaptive Gaussian controls for dynamic and static Gaussians as per 3D-GS [46]. Training on a sequence of 300 frames of $960 \times 720$ resolution takes about 2 hours to finish on an A100 GPU. Our rendering FPS is around 140 fps. We use the same set of hyper-parameters for all experiments. + +# 4. Experiment + +We evaluate our performance quantitatively and qualitatively on a broad range of tasks: long range 3D point tracking, long-range 2D point tracking, and novel view synthesis. We focus our evaluation on datasets that exhibit substantial scene motion. In particular, the iPhone dataset [25] features casual captures of real-world scenes that closely match our target scenarios. It provides comprehensive annotations, including simultaneous validation views, lidar depth, sparse 2D point correspondences across the entire video, and can be used to evaluate our performance on all three tasks. Given the challenge of obtaining precise 3D track annotations for real data, we also evaluate performance using the scenes from the synthetic MOVi-F Kubric dataset [28]. + +![](images/523415722d16eb2f471f166f3a94f9630ee5a1cb76301ab9406bdd05f0515d90.jpg) +Figure 3. 3D Track visualization on iPhone dataset. We render novel views for each method and overlay their predicted 3D tracks on top of the rendered images. For clarity, we only display a segment of the trails spanning 50 frames for a specific set of grid query points. However, it is important to note that our method can generate dense, full-length 3D tracks. *Note that TAPIR + DA cannot produce novel view rendering, and we overlay their tracking results with our rendering to make it easier to interpret. + +# 4.1. Task Specification + +Long-Range 3D Tracking. Our primary task is estimating 3D scene motion for any pixel over the entire video. For this purpose, we extend the metrics for scene flow evaluation introduced in RAFT-3D [96] to evaluate long-range 3D tracking. Specifically, we report the 3D end-point-error (EPE), which measures the Euclidean distance between the ground truth 3D tracks and predicted tracks at each target time step. In addition, we report the percentage of points that fall within a given threshold of the ground truth 3D location $\delta_{3D}^{0.05} = 5\mathrm{cm}$ and $\delta_{3D}^{10} = 10\mathrm{cm}$ in metric scale. + +Long-Range 2D Tracking. Our 3D motion representation can be easily projected onto image plane to get long-range 2D tracks. Thus we also evaluate 2D tracking performance in terms of both position and occlusion accuracy. Following the metrics introduced in the TAP-Vid benchmark [15], we report the Average Jaccard (AJ), average position accuracy $(<\delta_{\mathrm{avg}})$ , and Occlusion Accuracy (OA). + +Novel View Synthesis. We measure our method's novel view synthesis quality as a comprehensive assessment for geometry, appearance, and motion. We evaluate on the iPhone dataset [25] which provides validation views and report co-visibility masked image metrics [25]: mPSNR, mSSIM [105] and mLPIPS [26, 36]. Additionally, we evaluate on NVIDIA dataset for more comparison. + +
Method3D Tracking2D TrackingView Synthesis
EPE ↓δ053D↑δ103D↑AJ ↑<δavg↑OA ↑PSNR ↑SSIM ↑LPIPS ↓
T-NeRF [25]------15.600.550.55
HyperNeRF [75]0.18228.445.810.119.352.015.990.590.51
DynIBaR [58]0.25211.424.65.48.737.713.410.480.55
Deformable-3D-GS [119]0.15133.455.314.020.963.911.920.490.66
DynMF [51]0.18822.953.85.59.560.516.540.590.49
CoTracker [44]+DA [118]0.20234.357.924.133.973.0---
TAPIR [16]+DA [118]0.11438.163.227.841.567.4---
DELTA (world) [72]0.15932.555.324.734.168.9---
SpatialTracker (world) [110]0.12537.763.924.936.973.5---
Ours0.08243.073.334.447.086.616.720.630.45
Ours + 2DGS[33]0.09747.371.335.847.087.316.750.650.40
+ +Table 1. Evaluation on iPhone dataset. Our method achieves SOTA performance all tasks of 3D point tracking, 2D point tracking, and novel view synthesis. The baselines that perform best on 2D and 3D tracking (TAPIR [16] + DA [118], CoTracker [44] + DA [118], DELTA [72], SpatialTracker [110]) are unable to synthesize novel views of the scene, while the methods that perform best in novel view synthesis struggle with or fail to produce 2D and 3D tracks. Our method achieves a significant boost in all three tasks above baselines. We include training details about “Ours + 2DGS [33]” in the supplement. + +![](images/c2efc3ec5fb7ff822b96f8f2939d5645a8cb570a8d97cf93df172b407b85c48f.jpg) +Figure 4. Qualitative comparison of novel view synthesis on iPhone dataset. The leftmost image in each row shows the training view at the same time step as the validation view. The regions highlighted in green indicate areas excluded from evaluation due to the lack of co-visibility between training and validation views. Please see the supplemental video for more qualitative results and comparisons. + +# 4.2. Baselines + +Our method represents dynamic 3D scene comprehensively with explicit long-range 3D scene motion estimation, which also allows for novel view synthesis. While no existing method achieves the exact same goals as ours, there are methods that focus on sub-tasks related to our problem, such as dynamic novel view synthesis, 2D tracking, or monocular depth estimation. We therefore adapt existing + +methods as our baselines which are introduced below. + +While dynamic novel view synthesis approaches focus primarily on the photometric reconstruction quality and do not explicitly output 3D point tracks, we can adapt some of their representations to derive 3D point tracks for our evaluation. For HyperNeRF [75], we compose the learned inverse mapping (from view space to canonical space) and a forward mapping solved via root-finding [11, 25] to produce 3D tracks at query points. DynIBaR [58] produces + +![](images/b8dc3a3c551e55532cedb75d38e0fc8590cff27f2aaf0113bc5db7f86ad662df.jpg) +Figure 5. Visualization of motion coefficients after PCA and predicted depth maps on iPhone dataset. The motion coefficients encode information regarding rigid moving components. For instance, our motion coefficient PCA produces constant color for the block in the second example which exhibits rigid motion. + +short-range view-to-view scene flow, which we chain into long-range 3D tracks for our evaluation. Deformable-3D-GS (D-3DGS) [119] represents dynamic scenes using 3D Gaussians [45] in the canonical space and a deformation MLP network that deforms the canonical 3D Gaussians into each view, which naturally allows 3D motion computation. T-NeRF [25] models dynamic scenes using time as MLP input in addition to 3D locations, which does not provide a method for extracting 3D motion, and hence they are not considered for 2D/3D tracking evaluation. Finally, we implemented our own version of DynMF [51] with neural network motion bases. Implementation details are provided in the supplemental material. We evaluate the 3D tracks of these methods only on the iPhone dataset, because we found none can handle the Kubric scene motion. + +Alongside dynamic view synthesis baselines, we include 3D tracking baselines. Specifically, we take state-of-the-art long-range 2D tracks from TAPIR [16] and CoTracker [44] and lift them into 3D scene motion using monocular depth maps produced by Depth Anything (DA) [118]. We compute the correct scale and shift for each relative depth map from Depth Anything to align them with the scene. The two resulting baselines are called "CoTracker $[44] + DA[118]$ " and "TAPIR $[16] + DA[118]$ ". Note that these baselines can only produce 3D tracks for visible regions, as the depth values for occluded points are unknown from such a 2.5D representation. In contrast, our global representation allows for modeling 3D motion through occlusions. In addition, we include frame-space 3D tracking methods such as DELTA [72] and SpatialTracker [110], and transform their trajectories into world coordinates using annotated camera poses for evaluation. + +# 4.3. Evaluation on iPhone Dataset + +iPhone dataset [25] contains 14 sequences of 200-500 frames featuring various types of challenging real world scene motion. All sequences are recorded using a handheld moving camera in a casual manner, with 7 of them additionally featuring two synchronized static cameras with a large baseline for novel view synthesis evaluation. For 3D tracking evaluation, we generate the groundtruth 3D tracks + +by lifting the 2D keypoint annotations into 3D using lidar depth, and masking out points that are occluded or with invalid lidar depth values. All experiments are conducted with the original instead of half resolution as in [25] given that our method can handle high-res video input. We also discard scenes SPACE-OUT and WHEEL due to camera and lidar error. We find that the original camera annotations from ARKit is not accurate enough, and refine them using global bundle adjustment from COLMAP [85]. The refined poses are used for all methods. + +We report the quantitative comparison in Tab. 1 which shows that our method outperforms all baselines on all tasks by a substantial margin. For tracking, we observe significant improvements over naive baselines that combine 2D tracks with depth maps, especially "TAPIR+DA", which also serves as input to our system, highlighting the effectiveness of our consolidation process. The improvement stems from three factors: 1) a low-dimensional $\mathbb{SE}(3)$ motion representation, which allows inferring occluded motion via nearby visible regions; 2) 3D regularizations, such as low acceleration constraints that promote coherent trajectories; 3) the consolidation of all pairwise tracks into a single 4D representation with persistent 3D geometry, which serves as a structural prior to correct noisy 2D tracks. Our method also achieves the best novel view synthesis quality across dynamic NeRF and 3D-GS-based baselines. + +Fig. 3 shows qualitative comparison of the 3D tracking results. We render the novel views and plot the predicted 3D tracks of the given query points onto the novel views. Since "TAPIR + DA" cannot perform novel view synthesis, we overlay their track predictions onto our renderings to aid interpretation. D-3DGS [119] and HyperNeRF [75] fail to capture the significant scene motion in PAPER-WINDMILL and SPIN, resulting in structure degradation and blurry rendering. "TAPIR + DA" can track large motions, but their 3D tracks tend to be noisy and erroneous. In contrast, our method not only generates high-quality novel views but also the most smooth and accurate 3D tracks. Fig. 4 provides additional novel view synthesis comparison on the validation views. In Fig. 5, we visualize the rendering of the first three components of PCA decomposition for the motion coefficients, which correlates well with the rigid groups of the moving objects. + +4.4. Evaluation on Kubric Dataset + +
MethodEPE↓δ0.053D↑δ103D↑
CoTracker [44]+DA [118]0.1934.456.5
TAPIR [16]+DA [118]0.2034.056.2
Ours0.1639.862.2
+ +Table 2. 3D Tracking evaluation on Kubric dataset. + +The Kubric MOVi-F dataset contains short 24-frame videos + +of scenes of 10-20 objects, rendered with linear camera movement and motion blur. Multiple rigid objects are tossed onto the scene, at a speed that far exceeds the speed of the moving camera, making it a similar capture scenario to in-the-wild capture scenarios. It provides dense comprehensive annotations, including ground truth depth maps, camera parameters, segmentation masks, and point correspondences that are dense across time. We use 30 scenes from the MOVi-F validation set to evaluate the accuracy of our long-range 3D point tracks for all points in time. + +We demonstrate our method on the synthetic Kubric MOVi-F dataset in long-range 3D point tracking across time. Of the above baselines, only "CoTracker [44] + DA [118]" and "TAPIR [44] + DA [118]" yield 3D tracks for these scenes. For all baselines, we provide the ground truth camera intrinsics and extrinsics, and monocular depth estimates that have been aligned to the ground truth depth map. We obtain point tracks for all non-background pixels for each method. + +We report our quantitative 3D point tracking metrics in Table 2, and find that across all metrics, our method outperforms the baselines. Please see the supplement for figures of the optimized motion coefficients, which coherently group moving objects. + +# 4.5. Evaluation on NVIDIA Dataset + +We conduct experiments on seven scenes from the NVIDIA dataset [122], following the evaluation protocol of Dynamic Gaussian Marbles [91], which evaluates the task from a single static camera instead of interleaving different camera as discussed in [25]. We adopt the same image resolution, camera parameters and depths maps as input. Please see the supplement for more details. For each scene, we compute covisibility masks between each evaluation view and the training view at each time step to exclude unobservable regions during evaluation. Results are reported in Tab. 3. + +
MethodBalloon1Ballon2JumpingTruck
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]23.830.810.06623.620.810.08820.020.730.13427.310.860.055
Ours23.900.790.05823.910.830.07420.080.690.15127.650.880.048
MethodPlaygroundUmbrellaSkatingMean
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]16.680.600.17724.990.700.08327.790.900.05223.460.770.093
Ours16.870.550.16823.270.570.16727.910.920.04723.370.750.102
+ +# 4.6. Ablation studies + +We ablate various components of our method on the iPhone dataset in 3D tracking in Tab. 4. We first validate our choices of motion representation, namely the $\mathbb{SE}(3)$ motion bases parameterization, with three ablations: 1) "Per-Gaussian Transl": replacing our motion representation with naive per-Gaussian translational motion trajectories, + +Table 3. Evaluation on NVIDIA dataset. Our method is comparable with Dynamic Gaussian Marbles (DGM) [91]. + +
MethodsSE(3)Motion Basis2D tracksInitializationEPE↓δ0.053D↑δ1.103D↑
Ours (Full)✓✓✓✓0.08243.073.3
Transl. Bases✓✓✓0.09342.369.9
Per-Gaussian SE(3)✓✓✓0.08343.670.2
Per-Gaussian Transl.✓✓0.08741.269.2
No SE(3) Init.✓✓✓0.11139.365.7
No 2D Tracks✓✓0.14130.457.8
+ +Table 4. Ablation Studies on iPhone dataset. + +2) "Per-Gaussian $\mathbb{SE}(3)$ : replacing our motion representation with naive per-Gaussian $\mathbb{SE}(3)$ motion trajectories, and 2) "Transl. Bases": keeping the motion bases representation but only using translational bases instead of $\mathbb{SE}(3)$ . We find that $\mathbb{SE}(3)$ bases significantly improve 3D tracking over both translational bases and per-Gaussian translational motion. It also outperforms per-Gaussian $\mathbb{SE}(3)$ in overall accuracy and offers noticeably better visual quality, as the latter exhibits significantly more popping and jitter artifacts. + +Next, we ablate our training strategies including initialization and supervision signal. We conduct an ablation of "No $\mathbb{SE}(3)$ Init.", where instead of performing our initial $\mathbb{SE}(3)$ fitting stage, we initialize the translational part of the motion bases with randomly selected noisy 3D tracks formed by directly lifting input 2D tracks using depth maps into 3D, and the rotation part as identity. We find that skipping this initialization noticeably hurts performance. Lastly, we remove the 2D track supervision entirely ("No 2D Tracks") and find it to lead to significant drop in performance, which verifies the importance of the 2D track supervision for 3D tracking. + +# 5. Discussion + +Limitations. Like most prior monocular view synthesis methods, it still requires per-scene test-time optimization, hindering streamable applications. Recent feed-forward methods for joint reconstruction and tracking [22, 111, 113] offer a promising alternative, showing encouraging progress toward overcoming this limitation. Our method relies on off-the-shelf predictions for camera poses, geometry, and motion, which may degrade in textureless regions or under large motions. However, it also naturally benefits from advances in these components. Finally, it also requires user input to mask moving objects. Recent advances in moving object segmentation [27, 34] offer promising directions beyond our current manual solution and could be integrated into our pipeline. + +Conclusion. We introduce a method for joint long-range 3D tracking and novel view synthesis from monocular video, representing dynamic elements via global 3D Gaussians with time-varying translations and rotations. Motion trajectories are regularized using low-dimensional rigid motion bases. By consolidating noisy observations into globally consistent scene estimates, our approach outperforms state-of-the-art methods in 2D/3D tracking and novel view synthesis across synthetic and real benchmarks. + +Acknowledgement. We thank Ruilong Li, Noah Snavely, Brent Yi and Aleksander Holynski for helpful discussion. We are in memory of our beloved cat Sriracha, who will always be missed and loved. This project is supported in part by DARPA No. HR001123C0021. and IARPA DOI/IBC No. 140D0423C0035. The views and conclusions contained herein are those of the authors and do not represent the official policies or endorsements of these institutions. + +# References + +[1] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, 2006. 2 +[3] Stanley T Birchfield and Shrinivas J Pundlik. Joint tracking of features and edges. In 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2008. 2 +[4] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, 1993. 2 +[5] Aljaz Bozic, Michael Zollhöfer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[6] C. Bregler, A. Hertzmann, and H. Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No.PR00662), 2000. 2 +[7] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In European Conference on Computer Vision, 2004. 2 +[8] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2 +[9] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4), 2020. 1, 2 +[10] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[11] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6 +[12] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. + +Dreamscene4d: Dynamic multi-object scene generation from monocular videos. NeurIPS, 2024. 2 +[13] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. Int. J. Comput. Vision, 107(2), 2014. 2 +[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2 +[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 2022. 2, 5 +[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 6, 7 +[17] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Joao Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, 2024. 2 +[18] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Häusser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), 2015. 2 +[19] Mingsong Dou, S. Khamis, Yu.G. Degtyarev, Philip L. Davidson, S. Fanello, Adarsh Kowdle, Sergio Orts, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4d. ACM Transactions on Graphics (TOG), 35, 2016. 2 +[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2021. 2 +[21] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2 +[22] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J Black, Trevor Darrell, and Angjoo Kanazawa. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. arXiv preprint arXiv:2504.13152, 2025.8 +[23] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. Kplanes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2 + +[24] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1 +[25] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Dynamic novel-view synthesis: A reality check. In NeurIPS, 2022. 2, 5, 6, 7, 8 +[26] Leon A Gatys, Alexander S Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 5 +[27] Lily Goli, Sara Sabour, Mark Matthews, Marcus Brubaker, Dmitry Lagun, Alec Jacobson, David J Fleet, Saurabh Saxena, and Andrea Tagliasacchi. Romo: Robust motion segmentation improves structure from motion. arXiv preprint arXiv:2411.18650, 2024. 8 +[28] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5 +[29] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1 +[30] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, 2022. 2 +[31] Thorsten Hempel, Ahmed A Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In 2022 IEEE International Conference on Image Processing (ICIP), 2022. 4 +[32] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. In Other Conferences, 1981. 2 +[33] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 6 +[34] Nan Huang, Wenzhao Zheng, Chenfeng Xu, Kurt Keutzer, Shanghang Zhang, Angjoo Kanazawa, and Qianqian Wang. Segment any motion in videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025. 8 +[35] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. In European Conference on Computer Vision. Springer, 2022. 2 +[36] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, 2020. 5 + +[37] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[38] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 2023. 2 +[39] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2 +[40] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European Conference on Computer Vision, 2016. 2 +[41] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), 2018. 2 +[42] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[43] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2 +[44] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European conference on computer vision, 2024. 2, 6, 7, 8 +[45] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 1, 3, 7 +[46] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 2, 5 +[47] Diederik P Kingma, J Adam Ba, and J Adam. A method for stochastic optimization. arxiv 2014. arXiv preprint arXiv:1412.6980, 2020. 5 +[48] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4 +[49] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[50] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, Joao Carreira, Andrew Zisserman, Gabriel Brostow, and + +Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. Advances in Neural Information Processing Systems, 2024. 2 +[51] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. ECCV, 2024. 2, 4, 6, 7 +[52] Suryansh Kumar, Yuchao Dai, and Hongdong Li. Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2 +[53] Jiahui Lei, Yijia Weng, Adam W Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[54] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision, 2022. 2 +[55] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[56] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2 +[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2 +[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 6 +[59] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[60] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2, 5 +[61] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 2 +[62] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy + +namic gaussian splatting from causally-captured monocular videos. ICLR, 2025. 2 +[63] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 1 +[64] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004. 2 +[65] Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[66] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, 1981. 2 +[67] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), 2024. 1 +[68] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4), 2020. 2 +[69] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[70] Michal Neoral, Jonás Šerych, and Jiří Matas. Mft: Long-term tracking of every pixel. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 2 +[71] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2 +[72] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 2, 6, 7 +[73] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In ICCV, 2019. 2 +[74] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2 +[75] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. SIGGRAPH Asia, 2021. 1, 2, 6, 7 +[76] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields + +for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[77] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision, 2020. 1 +[78] Rene Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[79] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multi-frame optical flow estimation. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 2 +[80] Michael Rubinstein and Ce Liu. Towards longer long-range motion trajectories. In *British Machine Vision Conference*, 2012. 2 +[81] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R. Bradski. Orb: An efficient alternative to sift or surf. 2011 International Conference on Computer Vision, 2011. 2 +[82] Chris Russell, Rui Yu, and Lourdes Agapito. Video popu- up: Monocular 3d reconstruction of dynamic scenes. In European conference on computer vision. Springer, 2014. 2 +[83] P Sand. Long-range video motion estimation using point trajectories. PhD thesis, Ph.D. dissertation, Cambridge, MA, USA, 2006, adviser-Teller, Seth, 2006. 2 +[84] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 2008. 2 +[85] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7 +[86] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 5 +[87] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In ICCV, 2023. 2 +[88] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. International Journal of Computer Vision, 2004. 2 +[89] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5), 2023. 2 +[90] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision, 2024. 2 +[91] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas + +Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. Siggraph Asia, 2024. 2, 8 +[92] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum. Wiley Online Library, 2008. 2 +[93] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2 +[94] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. Springer, 2020. 2 +[95] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 2021. 5 +[96] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5 +[97] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2 +[98] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[99] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. 2013 IEEE International Conference on Computer Vision, 2013. 2 +[100] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[101] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In International Conference on Computer Vision, 2023. 2, 4 +[102] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[103] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2 +[104] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2 +[105] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5 + +[106] Zirui Wang, Shuda Li, Henry Howard-Jenkins, Victor Adrian Prisacariu, and Min Chen. Flownet3d++: Geometric losses for deep scene flow estimation. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 1 +[107] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, 2022. 2 +[108] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2023. 1, 2 +[109] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[110] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7 +[111] Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Yuri Makarov, Bingyi Kang, Xing Zhu, Hujun Bao, Yujun Shen, and Xiaowei Zhou. Spatialtrackerv2: 3d point tracking made easy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 8 +[112] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2 +[113] Zhen Xu, Zhengqin Li, Zhao Dong, Xiaowei Zhou, Richard Newcombe, and Zhaoyang Lv. 4dgt: Learning a 4d gaussian transformer using real-world monocular videos. arXiv preprint arXiv:2506.08015, 2025. 8 +[114] Gengshan Yang, Deqing Sun, V. Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[115] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021. +[116] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[117] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4 +[118] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Un + +leashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 4, 6, 7, 8 +[119] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 1, 2, 6, 7 +[120] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. *ICLR*, 2024. 2 +[121] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research, 26(34), 2025. 5 +[122] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 8 +[123] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arxiv:2410.03825, 2024. 2 +[124] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4), 2021. 2 +[125] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision. Springer, 2022. 2 +[126] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2 +[127] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew William Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (TOG), 33, 2014. 2 \ No newline at end of file diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/images.zip b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..74075bb1e8aacf42119d2e63960f25545011c8a9 --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50c02bc1d58f6040bbfa49f4b9a0b3b59ff8ea686978602c9a26abf93af5dc22 +size 699572 diff --git a/2025/Shape of Motion_ 4D Reconstruction from a Single Video/layout.json b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38771b4037491e9b5b3b7aa28f760100e01150fb --- /dev/null +++ b/2025/Shape of Motion_ 4D Reconstruction from a Single Video/layout.json @@ -0,0 +1,11385 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "spans": [ + { + "bbox": [ + 128, + 103, + 482, + 121 + ], + "type": "text", + "content": "Shape of Motion: 4D Reconstruction from a Single Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": "Qianqian Wang" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Vickie Ye" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Hang Gao" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Weijia Zeng" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Jake Austin" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Zhengqi Li" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Angjoo Kanazawa" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "spans": [ + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "text", + "content": "UC Berkeley " + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "text", + "content": "Google DeepMind " + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "text", + "content": "UC San Diego " + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 120, + 178, + 487, + 194 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 220, + 303, + 342 + ], + "blocks": [ + { + "bbox": [ + 57, + 220, + 303, + 342 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 303, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 303, + 342 + ], + "type": "image", + "image_path": "6e4a5d046a457279c9ea0ce38be0ac15921887a07fc6f4ecae3ad61d9e044311.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 555, + 384 + ], + "type": "text", + "content": "Figure 1. Shape of Motion. Our method enables joint long-range 3D tracking and novel view synthesis from a monocular video of a complex dynamic scene. We render moving elements at a fixed viewpoint across time and visualize estimated 3D motion as colorful trajectories. These trajectories reveal distinct geometric patterns, which leads to the term \"Shape of Motion\"." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 220, + 554, + 342 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 554, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 554, + 342 + ], + "type": "image", + "image_path": "e3ce98b04f76bf54f69821528a0ff68dd5219d77e9ab4f88e60757429348fef9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "spans": [ + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 419, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 297, + 684 + ], + "type": "text", + "content": "Monocular dynamic reconstruction is a challenging and long-standing vision problem due to the highly ill-posed nature of the task. Existing approaches depend on templates, are effective only in quasi-static scenes, or fail to model 3D motion explicitly. We introduce a method for reconstructing generic dynamic scenes, featuring explicit, persistent 3D motion trajectories in the world coordinate frame, from casually captured monocular videos. We tackle the problem with two key insights: First, we exploit the low-dimensional structure of 3D motion by representing scene motion with a compact set of " + }, + { + "bbox": [ + 54, + 419, + 297, + 684 + ], + "type": "inline_equation", + "content": "\\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 54, + 419, + 297, + 684 + ], + "type": "text", + "content": " motion bases. Each point's motion is expressed as a linear combination of these bases, facilitating soft decomposition of the scene into multiple rigidly-moving groups. Second, we take advantage of off-the-shelf data-driven priors such as monocular depth maps and long-range 2D tracks, and devise a method to effectively consolidate these noisy supervisory signals, resulting in a globally consistent representation of the dynamic scene. Experiments show that our method achieves state-of-the-art performance for both long-range 3D/2D motion estimation and novel view synthesis on dynamic scenes. Project Page: https://shape-of-motion.github.io/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 394, + 394, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 394, + 394, + 406 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 394, + 406 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 414, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 617 + ], + "type": "text", + "content": "Reconstructing the persistent geometry and their 3D motion across a video is crucial for understanding and interacting with the underlying physical world. While recent years have seen impressive progress in modeling static 3D scenes [45, 69], recovering the geometry and motion of complex dynamic 3D scenes, especially from a single video, remains an open challenge. A number of prior dynamic reconstruction and novel view synthesis approaches have attempted to tackle this problem. However, most methods rely on synchronized multi-view videos [9, 23, 23, 67, 108] or additional LIDAR/depth sensors [29, 63, 77, 98, 106]. Recent monocular approaches can operate on regular dynamic videos, but they typically model 3D scene motion as short-range scene flow between consecutive times [24, 57, 58] or deformation fields that map between canonical and view space [74, 75, 119], failing to capture 3D motion trajectories persistent over a video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "The longstanding challenge for general in-the-wild videos lies in the poorly constrained nature of the 4D reconstruction problem. In this work, we tackle this challenge with two key insights. The first is that, while the image space dynamics can be complex and discontinuous, the underlying 3D motion is a composition of continuous simple rigid motions. Our second insight is that data-driven priors provide complementary, though noisy cues, that aggregate" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 703, + 123, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 123, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 123, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9660" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "well into a globally coherent representation of the 3D scene geometry and motion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 323 + ], + "type": "text", + "content": "Motivated by these two insights, we represent the dynamic scene as a set of persistent 3D Gaussians, and represent their motion across the video in terms of a compact set of shared " + }, + { + "bbox": [ + 56, + 97, + 295, + 323 + ], + "type": "inline_equation", + "content": "\\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 56, + 97, + 295, + 323 + ], + "type": "text", + "content": " motion bases. Unlike traditional scene flow, which computes 3D correspondences between consecutive frames, our representation recovers a persistent 3D trajectory over the whole video, enabling long-range 3D tracking across the entire video. As the 3D trajectories produced by our method capture the geometric patterns that trace each point's movement through 3D space and time as shown in Figure 1, we refer to our approach as \"Shape of Motion\". We show how to fit our explicit scene representation to a general video in-the-wild, by fusing together complementary cues from two main sources: monocular depth estimates per-frame, and 2D track estimates across frames. We conduct extensive evaluations on both synthetic and real-world dynamic video datasets, and show that our proposed approach significantly outperforms prior methods in both long-range 2D/3D tracking and novel view synthesis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 396 + ], + "type": "text", + "content": "In summary, our key contributions are: (1) A 4D scene representation enabling both real-time novel view synthesis and globally consistent 3D tracking for any point at any time. (2) An optimization framework that fuses learned motion and geometry priors into a unified 4D representation from a single, casually captured monocular video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 406, + 142, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 142, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 142, + 418 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 426, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 700 + ], + "type": "text", + "content": "Correspondences and Tracking. While monocular 3D long range tracking remains largely unexplored, numerous approaches track in 2D image space, typically using optical flow for point correspondences. This involves estimating dense motion fields between image pairs [4, 7, 8, 18, 32, 35, 35, 39, 41-43, 66, 79, 87, 93, 94, 112]. Sparse keypoint matching methods can enable long trajectory generation [2, 14, 61, 64, 81], but these methods are primarily intended for sparse 3D reconstruction. Long-range 2D trajectory estimation for arbitrary points has been explored in earlier works, which relied on hand-crafted priors to generate motion trajectories [3, 80, 83, 84, 88, 99]. Recently, there has been a resurgence of interest in this problem, with several works showcasing impressive long-range 2D tracking results on challenging, in-the-wild videos. These approaches employ either test-time optimization where models consolidate noisy short-range motion estimates into long-term correspondences [70, 90, 101], or data-driven strategies [16, 17, 30, 44], where neural networks learn long-term correspondence estimates from synthetic data [15, 126]. While these methods effectively track any 2D point throughout a video, they lack the knowledge of underlying 3D scene geometry and motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 702, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 294, + 712 + ], + "type": "text", + "content": "To model 3D scene motion, recent work in monocu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "lar settings has explored self-supervised learning and test-time optimization [37, 57, 58, 114-116]. More recent approaches estimate 3D trajectories in general scenes in a feedforward manner [50, 72, 110], but the predicted motion remains in the frame space, entangling object and camera motion. In contrast, our method recovers persistent, long-range 3D trajectories in the world coordinate space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 158, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 300 + ], + "type": "text", + "content": "Dynamic Reconstruction and View Synthesis. Our work also relates to dynamic 3D scene reconstruction and novel view synthesis. In non-rigid reconstruction, early methods often required RGB-D sensors [5, 19, 40, 71, 127] or strong hand-crafted priors [52, 78, 82]. A family of works [6, 13, 73] apply low-rank assumptions to the underlying motion model. Recent work has demonstrated progress toward general dynamic reconstruction either by integrating monocular depth priors [49, 56, 60, 68, 124, 125], or directly learning to perform dynamic reconstruction [65, 102, 103, 123] from data. However, these methods do not perform long-range 3D tracking, and do not focus on novel view synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 301, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 517 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [69] and Gaussian Splating [46] have shown strong performance in novel view synthesis. For dynamic scenes, many methods [1, 9, 10, 23, 23, 55, 59, 89, 92, 100] still require simultaneous multi-view video observations or predefined templates [38, 54, 107]. Template-free monocular approaches model dynamic scenes with different types of representations such as video depth maps [122], time-dependent NeRFs [20, 57, 58, 74-76, 97, 109], and dynamic 3D Gaussians [21, 108, 119, 120]. Although significant progress has been made, as DyCheck [25] pointed out, many approaches focus on scenarios with camera teleportation [51, 119] or quasi-static scenes [74, 75], which are effectively multiview and do not represent real-world monocular videos. In this work, we focus on casually-captured monocular videos, a more practical and challenging setup. Recent and concurrent efforts [12, 53, 62, 91, 104] have also investigated similar settings, using strong off-the-shelf data-driven priors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 526, + 370, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 370, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 370, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": "Our method takes as input a sequence of " + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": " video frames " + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\{I_t\\in \\mathbb{R}^{H\\times W\\times 3}\\}" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": " of a dynamic scene, the camera intrinsics " + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": ", and world-to-camera extrinsics " + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t\\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": " of each input frame " + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": ". From these inputs, we aim to recover the geometry of the entire dynamic scene and the full-length 3D motion trajectory of every point in the scene. Unlike most prior dynamic NeRFs methods [25, 56, 58, 97, 109] which render the scene contents through volumetric ray casting and represent the motion implicitly at fixed 3D locations, we model the dense scene elements as a set of canonical 3D Gaussians, that translate and rotate over entire video as persistent motion trajectories. We adopt explicit point-based representation because it simultaneously allows for both (1) high-fidelity rendering in real-time and (2) full" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9661" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 75, + 531, + 293 + ], + "blocks": [ + { + "bbox": [ + 81, + 75, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 81, + 75, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 81, + 75, + 531, + 293 + ], + "type": "image", + "image_path": "6acad7d745ddf3a9d6f7a829e41a90cfdcd4442515a6531f79b06bbad3d31478.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 302, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 54, + 302, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 302, + 555, + 380 + ], + "type": "text", + "content": "Figure 2. System Overview. Given a single RGB video sequence with known camera poses, monocular depth maps and 2D tracks computed from off-the-shelf models [16, 118] as input, we optimize a dynamic scene representation as a set of persistent 3D Gaussians that translate and rotate over time. To capture the low-dimensional nature of scene motion, we model the motion with a set of compact " + }, + { + "bbox": [ + 54, + 302, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 54, + 302, + 555, + 380 + ], + "type": "text", + "content": " motion bases shared across all scene elements. Each 3D Gaussian's motion is represented as a linear combination of these motion bases, weighted by motion coefficients specific to each Gaussian. We supervise our scene representation (canonical 3D Gaussians, per-Gaussian motion coefficients, and global motion bases) by comparing the rendered outputs (RGB, depths and 2D tracks) with the corresponding input signals. This results in a globally coherent dynamic 3D scene representation in the world coordinate with long-range 3D trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 386, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 399 + ], + "type": "text", + "content": "length 3D tracking of any surface point from any input time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 567 + ], + "type": "text", + "content": "Optimizing an explicit representation of dynamic 3D Gaussians from a single video is severely ill-posed — at each point in time, the moving subjects in the scene are observed from only a single viewpoint. In order to overcome this ambiguity, we make two insights: First, while the projected 2D dynamics might be complex in the video, the underlying 3D motion in the scene is low-dimensional, and composed of simpler units of rigid motion. Second, powerful data-driven priors, namely monocular depth estimates and long-range 2D tracks, provide complementary but noisy signals of the underlying 3D scene. We propose a system that fuses these noisy estimates together into a globally coherent representation of both the scene geometry and motion. We show a schematic of our pipeline in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 254, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 254, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 254, + 588 + ], + "type": "text", + "content": "3.1. Preliminaries: 3D Gaussian Splatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "We represent the appearance and geometry of a dynamic scene with a global set of 3D Gaussians, an explicit and expressive differentiable scene representation [45] for efficient optimization and rendering. We define parameters of each 3D Gaussian in the canonical frame " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\vec{g}_0 \\equiv (\\pmb{\\mu}_0, \\mathbf{R}_0, \\mathbf{s}, o, \\mathbf{c})" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_0 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_0 \\in \\mathbb{SO}(3)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " are the 3D mean and orientation in the canonical frame, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " the scale, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " the opacity, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " the color. " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the canonical frame covariance matrix. These properties are persistent and shared across time. To render a set of 3D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": "Gaussians from a camera with world-to-camera extrinsics " + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": " and intrinsics " + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": ", the projections of the 3D Gaussians in the image plane are obtained by 2D Gaussians parameterized as " + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_0^{\\prime} \\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_0^{\\prime} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 386, + 555, + 436 + ], + "type": "text", + "content": " via affine approximation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 361, + 454, + 489, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 454, + 489, + 469 + ], + "spans": [ + { + "bbox": [ + 361, + 454, + 489, + 469 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\Pi (\\mathbf {K E} \\boldsymbol {\\mu} _ {0}) \\in \\mathbb {R} ^ {2},", + "image_path": "964cbcfa5c875b0b459ad352d75f9d5c0e908f89f9cf6423c3df08a7e13162aa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 361, + 470, + 555, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 470, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 361, + 470, + 555, + 486 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {0} ^ {\\prime} (\\mathbf {K}, \\mathbf {E}) = \\mathbf {J} _ {\\mathbf {K E}} \\boldsymbol {\\Sigma} _ {0} \\mathbf {J} _ {\\mathbf {K E}} ^ {\\top} \\in \\mathbb {R} ^ {2 \\times 2}, \\tag {1}", + "image_path": "49b587a34cbdc41607215babba4bba10b42bb876386f3fa3643a59acfacf0dd1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": " is perspective projection, and " + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{KE}}" + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": " is the Jacobian of perspective projection with " + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": " at the point " + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mu_0" + }, + { + "bbox": [ + 313, + 495, + 555, + 544 + ], + "type": "text", + "content": ". The 2D Gaussians can then be efficiently rasterized into RGB image and depth map via alpha compositing as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 552, + 554, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 552, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 321, + 552, + 554, + 580 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {I}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {c} _ {i}, \\quad \\hat {\\mathbf {D}} (\\mathbf {p}) = \\sum_ {i \\in H (\\mathbf {p})} T _ {i} \\alpha_ {i} \\mathbf {d} _ {i}, \\tag {2}", + "image_path": "469ba9689abe93eb08cfdd56ace798e98fb8ab62394d21d0740b6e9889e13cdc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = o_{i}\\cdot \\exp \\left(-\\frac{1}{2} (\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})^{T}\\boldsymbol{\\Sigma}_{0}^{\\prime}(\\mathbf{p} - \\boldsymbol{\\mu}_{0}^{\\prime})\\right)" + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j = 1}^{i - 1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "inline_equation", + "content": "H(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "content": " is the set of Gaussians that intersect the ray shoot from the pixel " + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 589, + 555, + 652 + ], + "type": "text", + "content": ". This process is fully differentiable, and enables direct optimization of the 3D Gaussian parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 482, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 482, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 482, + 673 + ], + "type": "text", + "content": "3.2. Dynamic Scene Representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Scene Motion Parameterization. To model a dynamic 3D scene, we keep track of " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " canonical 3D Gaussians and vary their positions and orientations over time with a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9662" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "per frame rigid transformation. In particular, for a moving 3D Gaussian at time " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ", its pose parameters " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "(\\pmb{\\mu}_t,\\mathbf{R}_t)" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " are rigidly transformed from the canonical frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0\\rightarrow t} = \\left[\\mathbf{R}_{0\\rightarrow t}\\quad \\mathbf{t}_{0\\rightarrow t}\\right]\\in \\mathbb{S}\\mathbb{E}(3)" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 128, + 295, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 128, + 295, + 141 + ], + "spans": [ + { + "bbox": [ + 88, + 128, + 295, + 141 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\boldsymbol {\\mu} _ {0} + \\mathbf {t} _ {0 \\rightarrow t}, \\quad \\mathbf {R} _ {t} = \\mathbf {R} _ {0 \\rightarrow t} \\mathbf {R} _ {0}. \\tag {3}", + "image_path": "bb5777a5f0ecb0cbe7b29887802f27ec3edb70716002cc24ffff8ba1879aa5e4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": "Rather than modeling the 3D motion trajectories independently for each Gaussian, we define a set of " + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "inline_equation", + "content": "B \\ll N" + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": " learnable basis trajectories " + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^{B}" + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": " that are globally shared across all Gaussians [51]. The transformation " + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0\\rightarrow t}" + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": " at each time " + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": " is then computed by a weighted combination of this global set of basis trajectories through per-point basis coefficients " + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{(b)}" + }, + { + "bbox": [ + 55, + 147, + 296, + 233 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 237, + 295, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 237, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 123, + 237, + 295, + 270 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {0 \\rightarrow t} = \\sum_ {b = 0} ^ {B} \\mathbf {w} ^ {(b)} \\mathbf {T} _ {0 \\rightarrow t} ^ {(b)}, \\tag {4}", + "image_path": "7d2a9cfcc908358d8b3f1115de19b2a03bf0d427d18d325da86bcf2a3de3e67c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{w}^{(b)}\\| = 1" + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "text", + "content": ". In our implementation, we parameterize " + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0\\rightarrow t}^{(b)}" + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "text", + "content": " as 6D rotation [31] and translation, and perform the weighted combination separately on each with the same weight " + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{(b)}" + }, + { + "bbox": [ + 55, + 277, + 296, + 388 + ], + "type": "text", + "content": ". During optimization, we jointly learn the set of global motion bases and motion coefficients of each 3D Gaussian. These compact motion bases explicitly regularize the trajectories to be low-dimensional, encouraging the 3D Gaussians that move similarly to each other to be represented by similar motion coefficients." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": "Rasterizing 3D Trajectories. Given this representation, we now describe how we obtain pixelwise 3D motion trajectory at any query frame " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": ". We take a similar approach to Wang et al. [101] and rasterize the motion trajectories of 3D Gaussians into query frame " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": ". Namely, for a query camera at time " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": " with intrinsics " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": " and extrinsics " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_t" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": ", we perform rasterization to obtain a map " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "{}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'} \\in \\mathcal{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "text", + "content": " that contains the expected 3D world coordinates of the surface points corresponding to each pixel at target time " + }, + { + "bbox": [ + 55, + 392, + 296, + 501 + ], + "type": "inline_equation", + "content": "t'" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 506, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 506, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 110, + 506, + 295, + 534 + ], + "type": "interline_equation", + "content": "{ } ^ { \\mathrm { w } } \\hat { \\mathbf { X } } _ { t \\rightarrow t ^ { \\prime } } ( \\mathbf { p } ) = \\sum _ { i \\in H ( \\mathbf { p } ) } T _ { i } \\alpha _ { i } \\boldsymbol { \\mu } _ { i , t ^ { \\prime } } , ( 5 )", + "image_path": "7dc69f8a31df475bd4de0f51ba5097f3a0927ff0eacc77ef204f5a86f0dadcf7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "H(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": " is the set of Gaussians that intersect the pixel " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": " at query time " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": ". The 2D correspondence location at time " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": " for a given pixel " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{U}}_{t\\rightarrow t'}(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": ", and the corresponding depth value at time " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}_{t\\rightarrow t'}(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 540, + 296, + 590 + ], + "type": "text", + "content": " can then be written as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 596, + 310, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 310, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 310, + 628 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\Pi \\left(\\mathbf {K} _ {t ^ {\\prime}} ^ {\\mathrm {c}} \\hat {\\mathbf {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p})\\right), \\hat {\\mathbf {D}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) = \\binom {\\mathrm {c}} {\\mathrm {X}} _ {t \\rightarrow t ^ {\\prime}} (\\mathbf {p}) \\tag {6}", + "image_path": "e9e5710425e8bf07bd95bdf196a6ea241aec1959d8314448fc1a77248b644490.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\mathrm{^c}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p}) = \\mathbf{E}_{t'}^{\\mathrm{w}}\\hat{\\mathbf{X}}_{t\\rightarrow t'}(\\mathbf{p})," + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "text", + "content": " is a perspective projection operation, and " + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "inline_equation", + "content": "(\\cdot)_{[3]}" + }, + { + "bbox": [ + 55, + 628, + 296, + 654 + ], + "type": "text", + "content": " is the third element of a vector." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "type": "text", + "content": "3.3. Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We prepare the following estimates using off-the-shelf methods in our optimization: 1) masks for the moving objects for each frame " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_t\\}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", which can be easily obtained" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": "using Track-Anything [48, 117] with a few user clicks, 2) monocular depth maps " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_t\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " computed using state-of-the-art relative depth estimation method such as Depth Anything [118] and 3) long-range 2D tracks " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}_{t\\rightarrow t^{\\prime}}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " for foreground pixels from state-of-the-art point tracking method TAPIR [16]. We align the relative depth maps with the metric depth maps by computing a per-frame global scale and shift and use them for our optimization, as we found relative depth maps tend to contain finer details. We treat the lifted 2D tracks unprojected with the aligned depth maps as noisy initial 3D track observations " + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 277 + ], + "type": "text", + "content": " for the moving objects. For the static part of the scene, we model them using standard static 3D Gaussians and initialize their 3D locations by unprojecting them into the 3D space using the aligned depth maps. The static and dynamic Gaussians are jointly optimized and rasterized together to form an image. We describe the optimization process below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": "Initialization. We first select the canonical frame " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " to be the frame in which the most 3D tracks are visible, and initialize the Gaussian means in the canonical frame " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mu_0" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " randomly sampled 3D tracks locations from this set of initial observations. We then perform k-means clustering on the vectorized velocities of the noisy 3D tracks from every frame " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t\\}" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": ", and initialize the motion bases " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " from these " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " clusters of tracks. Specifically, for the set of trajectories " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_t\\}_b" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " belonging to cluster " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": ", we initialize the basis transform " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{0\\rightarrow \\tau}^{(b)}" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " using weighted Procrustes alignment between the point sets " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_0\\}_b" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_{\\tau}\\}_b" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\tau = 0,\\dots ,T" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": ", where the weights are computed using uncertainty and visibility scores from TAPIR predictions. We initialize " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{(b)}" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " of each Gaussian to decay exponentially with its distance from the center of cluster " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " in the canonical frame. We then optimize " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mu_0" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{(b)}" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": ", and set of basis functions " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{0\\rightarrow t}^{(b)}\\}_{b = 1}^B" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": " to fit the observed 3D tracks with an " + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 285, + 555, + 508 + ], + "type": "text", + "content": "-loss under temporal smoothness constraints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "content": "Training. We supervise the dynamic Gaussians with two sets of losses. The first set comprise our reconstruction loss to match the per-frame pixelwise color, depth, and masks inputs. The second set enforce consistency of correspondences across time. Specifically, during each training step, we render the image " + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_t" + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "content": ", depth " + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}_t" + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "content": ", and mask " + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_t" + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "content": " from their corresponding training cameras " + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_t,\\mathbf{E}_t)" + }, + { + "bbox": [ + 313, + 512, + 556, + 621 + ], + "type": "text", + "content": " according to Equation 2. We supervise these predictions with a reconstruction loss applied independently per-frame" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 628, + 555, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 628, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 319, + 628, + 555, + 643 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e c o n}} = \\left\\| \\hat {\\mathbf {I}} - \\mathbf {I} \\right\\| _ {1} + \\lambda_ {\\text {d e p t h}} \\left\\| \\hat {\\mathbf {D}} - \\mathbf {D} \\right\\| _ {1} + \\lambda_ {\\text {m a s k}} \\left\\| \\hat {\\mathbf {M}} - \\mathbf {M} \\right\\| _ {1}. \\tag {7}", + "image_path": "a8e8d2f2336958389dbe740f16db062e3d5f0d26dd8bbdc92aa5dd5d8ab358e3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "The second set of losses supervises the motion of the Gaussians between frames. Specifically, we additionally render the 2D tracks " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{u}}_{t\\rightarrow t^{\\prime}}" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " and reprojected depths " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}_{t\\rightarrow t^{\\prime}}" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " for a pair of randomly sampled query time " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " and target time " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ". We supervise these rendered correspondences with the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9663" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 194, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 194, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 194, + 83 + ], + "type": "text", + "content": "long-range 2D track estimates via" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 97, + 294, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 97, + 294, + 111 + ], + "spans": [ + { + "bbox": [ + 123, + 97, + 294, + 111 + ], + "type": "interline_equation", + "content": "L _ {\\text {t r a c k - 2 d}} = \\left\\| \\mathbf {U} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {U}} _ {t \\rightarrow t ^ {\\prime}} \\right\\| _ {1}, \\tag {8}", + "image_path": "0117be24a89161b3926d7c52687f1b2ba1fdf3c9d0fc24d4cb5fd973ed904c88.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 114, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 114, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 100, + 114, + 295, + 129 + ], + "type": "interline_equation", + "content": "L _ {\\text {t r a c k - d e p t h}} = \\left\\| \\hat {\\mathbf {d}} _ {t \\rightarrow t ^ {\\prime}} - \\hat {\\mathbf {D}} \\left(\\mathbf {U} _ {t \\rightarrow t ^ {\\prime}}\\right)\\right\\| _ {1}. \\tag {9}", + "image_path": "3dc58adbeb484c8ed21aa09f609a7965c514cf3b5695ef83bc7fc50ccfcb43d1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": "Finally, we leverage physics motion prior by enforcing a distance-preserving loss between randomly sampled dynamic Gaussians and their k-nearest neighbors. Let " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{t'}" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": " denote the location of a Gaussian at time " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k(\\hat{\\mathbf{X}}_t)" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": " denote the set of k-nearest neighbors of " + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_t" + }, + { + "bbox": [ + 55, + 144, + 296, + 214 + ], + "type": "text", + "content": ", we define" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 226, + 295, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 226, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 66, + 226, + 295, + 258 + ], + "type": "interline_equation", + "content": "L _ {\\text {r i g i d i t y}} = \\left\\| \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t}\\right)\\right) - \\operatorname {d i s t} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}, \\mathcal {C} _ {k} \\left(\\hat {\\mathbf {X}} _ {t ^ {\\prime}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {10}", + "image_path": "ea5bfd2785a4369ae9fbd66d536f905d1739980b9cd67cf70ae07d52c84f33f5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 258, + 237, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 237, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 237, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 258, + 237, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 258, + 237, + 270 + ], + "type": "text", + "content": " measures Euclidean distance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 373 + ], + "type": "text", + "content": "Implementation Details. For in-the-wild videos, multiple methods [60, 95] are available to estimate camera parameters; further details are provided in the supplement. During reconstruction, camera extrinsics are refined as learnable parameters, update via camera gradient function in gsplat [121] CUDA kernel. For evaluating our methods on public benchmark, we use the camera annotations that come with the datasets (e.g., from COLMAP [86] or simulation)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "text", + "content": "We optimize our model using Adam Optimizer [47]. We perform 1000 iterations of optimization for the initialization phase and 500 epochs for training phase, respectively. The number of " + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "text", + "content": " bases " + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "text", + "content": " is set to 10 for all of our experiments. We initialize 40k dynamic Gaussians for the dynamic part and 100k static Gaussians for the static part of the scene, respectively. We perform the same adaptive Gaussian controls for dynamic and static Gaussians as per 3D-GS [46]. Training on a sequence of 300 frames of " + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "inline_equation", + "content": "960 \\times 720" + }, + { + "bbox": [ + 55, + 375, + 296, + 519 + ], + "type": "text", + "content": " resolution takes about 2 hours to finish on an A100 GPU. Our rendering FPS is around 140 fps. We use the same set of hyper-parameters for all experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 536, + 132, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 132, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 132, + 550 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our performance quantitatively and qualitatively on a broad range of tasks: long range 3D point tracking, long-range 2D point tracking, and novel view synthesis. We focus our evaluation on datasets that exhibit substantial scene motion. In particular, the iPhone dataset [25] features casual captures of real-world scenes that closely match our target scenarios. It provides comprehensive annotations, including simultaneous validation views, lidar depth, sparse 2D point correspondences across the entire video, and can be used to evaluate our performance on all three tasks. Given the challenge of obtaining precise 3D track annotations for real data, we also evaluate performance using the scenes from the synthetic MOVi-F Kubric dataset [28]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 69, + 556, + 277 + ], + "blocks": [ + { + "bbox": [ + 322, + 69, + 556, + 277 + ], + "lines": [ + { + "bbox": [ + 322, + 69, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 322, + 69, + 556, + 277 + ], + "type": "image", + "image_path": "523415722d16eb2f471f166f3a94f9630ee5a1cb76301ab9406bdd05f0515d90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 373 + ], + "type": "text", + "content": "Figure 3. 3D Track visualization on iPhone dataset. We render novel views for each method and overlay their predicted 3D tracks on top of the rendered images. For clarity, we only display a segment of the trails spanning 50 frames for a specific set of grid query points. However, it is important to note that our method can generate dense, full-length 3D tracks. *Note that TAPIR + DA cannot produce novel view rendering, and we overlay their tracking results with our rendering to make it easier to interpret." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 422, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 422, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 422, + 396 + ], + "type": "text", + "content": "4.1. Task Specification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "text", + "content": "Long-Range 3D Tracking. Our primary task is estimating 3D scene motion for any pixel over the entire video. For this purpose, we extend the metrics for scene flow evaluation introduced in RAFT-3D [96] to evaluate long-range 3D tracking. Specifically, we report the 3D end-point-error (EPE), which measures the Euclidean distance between the ground truth 3D tracks and predicted tracks at each target time step. In addition, we report the percentage of points that fall within a given threshold of the ground truth 3D location " + }, + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\delta_{3D}^{0.05} = 5\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\delta_{3D}^{10} = 10\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 406, + 555, + 528 + ], + "type": "text", + "content": " in metric scale." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 536, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 620 + ], + "type": "text", + "content": "Long-Range 2D Tracking. Our 3D motion representation can be easily projected onto image plane to get long-range 2D tracks. Thus we also evaluate 2D tracking performance in terms of both position and occlusion accuracy. Following the metrics introduced in the TAP-Vid benchmark [15], we report the Average Jaccard (AJ), average position accuracy " + }, + { + "bbox": [ + 313, + 536, + 556, + 620 + ], + "type": "inline_equation", + "content": "(<\\delta_{\\mathrm{avg}})" + }, + { + "bbox": [ + 313, + 536, + 556, + 620 + ], + "type": "text", + "content": ", and Occlusion Accuracy (OA)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Novel View Synthesis. We measure our method's novel view synthesis quality as a comprehensive assessment for geometry, appearance, and motion. We evaluate on the iPhone dataset [25] which provides validation views and report co-visibility masked image metrics [25]: mPSNR, mSSIM [105] and mLPIPS [26, 36]. Additionally, we evaluate on NVIDIA dataset for more comparison." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9664" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 70, + 470, + 205 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 470, + 205 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 470, + 205 + ], + "type": "table", + "html": "
Method3D Tracking2D TrackingView Synthesis
EPE ↓δ053D↑δ103D↑AJ ↑<δavg↑OA ↑PSNR ↑SSIM ↑LPIPS ↓
T-NeRF [25]------15.600.550.55
HyperNeRF [75]0.18228.445.810.119.352.015.990.590.51
DynIBaR [58]0.25211.424.65.48.737.713.410.480.55
Deformable-3D-GS [119]0.15133.455.314.020.963.911.920.490.66
DynMF [51]0.18822.953.85.59.560.516.540.590.49
CoTracker [44]+DA [118]0.20234.357.924.133.973.0---
TAPIR [16]+DA [118]0.11438.163.227.841.567.4---
DELTA (world) [72]0.15932.555.324.734.168.9---
SpatialTracker (world) [110]0.12537.763.924.936.973.5---
Ours0.08243.073.334.447.086.616.720.630.45
Ours + 2DGS[33]0.09747.371.335.847.087.316.750.650.40
", + "image_path": "015a08db001ce24520d1e4368e3b2ae671425e9b1490b59f39609c5c73729107.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 212, + 555, + 267 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 267 + ], + "type": "text", + "content": "Table 1. Evaluation on iPhone dataset. Our method achieves SOTA performance all tasks of 3D point tracking, 2D point tracking, and novel view synthesis. The baselines that perform best on 2D and 3D tracking (TAPIR [16] + DA [118], CoTracker [44] + DA [118], DELTA [72], SpatialTracker [110]) are unable to synthesize novel views of the scene, while the methods that perform best in novel view synthesis struggle with or fail to produce 2D and 3D tracks. Our method achieves a significant boost in all three tasks above baselines. We include training details about “Ours + 2DGS [33]” in the supplement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 269, + 555, + 554 + ], + "blocks": [ + { + "bbox": [ + 62, + 269, + 555, + 554 + ], + "lines": [ + { + "bbox": [ + 62, + 269, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 555, + 554 + ], + "type": "image", + "image_path": "c2efc3ec5fb7ff822b96f8f2939d5645a8cb570a8d97cf93df172b407b85c48f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 559, + 558, + 594 + ], + "lines": [ + { + "bbox": [ + 54, + 559, + 558, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 558, + 594 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of novel view synthesis on iPhone dataset. The leftmost image in each row shows the training view at the same time step as the validation view. The regions highlighted in green indicate areas excluded from evaluation due to the lack of co-visibility between training and validation views. Please see the supplemental video for more qualitative results and comparisons." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 602, + 122, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 122, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 122, + 615 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 297, + 715 + ], + "type": "text", + "content": "Our method represents dynamic 3D scene comprehensively with explicit long-range 3D scene motion estimation, which also allows for novel view synthesis. While no existing method achieves the exact same goals as ours, there are methods that focus on sub-tasks related to our problem, such as dynamic novel view synthesis, 2D tracking, or monocular depth estimation. We therefore adapt existing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 603, + 533, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 533, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 533, + 614 + ], + "type": "text", + "content": "methods as our baselines which are introduced below." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "While dynamic novel view synthesis approaches focus primarily on the photometric reconstruction quality and do not explicitly output 3D point tracks, we can adapt some of their representations to derive 3D point tracks for our evaluation. For HyperNeRF [75], we compose the learned inverse mapping (from view space to canonical space) and a forward mapping solved via root-finding [11, 25] to produce 3D tracks at query points. DynIBaR [58] produces" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "9665" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 293, + 145 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 293, + 145 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 293, + 145 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 293, + 145 + ], + "type": "image", + "image_path": "b8dc3a3c551e55532cedb75d38e0fc8590cff27f2aaf0113bc5db7f86ad662df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 152, + 295, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 208 + ], + "type": "text", + "content": "Figure 5. Visualization of motion coefficients after PCA and predicted depth maps on iPhone dataset. The motion coefficients encode information regarding rigid moving components. For instance, our motion coefficient PCA produces constant color for the block in the second example which exhibits rigid motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 220, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 295, + 397 + ], + "type": "text", + "content": "short-range view-to-view scene flow, which we chain into long-range 3D tracks for our evaluation. Deformable-3D-GS (D-3DGS) [119] represents dynamic scenes using 3D Gaussians [45] in the canonical space and a deformation MLP network that deforms the canonical 3D Gaussians into each view, which naturally allows 3D motion computation. T-NeRF [25] models dynamic scenes using time as MLP input in addition to 3D locations, which does not provide a method for extracting 3D motion, and hence they are not considered for 2D/3D tracking evaluation. Finally, we implemented our own version of DynMF [51] with neural network motion bases. Implementation details are provided in the supplemental material. We evaluate the 3D tracks of these methods only on the iPhone dataset, because we found none can handle the Kubric scene motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "text", + "content": "Alongside dynamic view synthesis baselines, we include 3D tracking baselines. Specifically, we take state-of-the-art long-range 2D tracks from TAPIR [16] and CoTracker [44] and lift them into 3D scene motion using monocular depth maps produced by Depth Anything (DA) [118]. We compute the correct scale and shift for each relative depth map from Depth Anything to align them with the scene. The two resulting baselines are called \"CoTracker " + }, + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "inline_equation", + "content": "[44] + DA[118]" + }, + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "text", + "content": "\" and \"TAPIR " + }, + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "inline_equation", + "content": "[16] + DA[118]" + }, + { + "bbox": [ + 55, + 400, + 295, + 604 + ], + "type": "text", + "content": "\". Note that these baselines can only produce 3D tracks for visible regions, as the depth values for occluded points are unknown from such a 2.5D representation. In contrast, our global representation allows for modeling 3D motion through occlusions. In addition, we include frame-space 3D tracking methods such as DELTA [72] and SpatialTracker [110], and transform their trajectories into world coordinates using annotated camera poses for evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 611, + 217, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 217, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 217, + 623 + ], + "type": "text", + "content": "4.3. Evaluation on iPhone Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "iPhone dataset [25] contains 14 sequences of 200-500 frames featuring various types of challenging real world scene motion. All sequences are recorded using a handheld moving camera in a casual manner, with 7 of them additionally featuring two synchronized static cameras with a large baseline for novel view synthesis evaluation. For 3D tracking evaluation, we generate the groundtruth 3D tracks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "type": "text", + "content": "by lifting the 2D keypoint annotations into 3D using lidar depth, and masking out points that are occluded or with invalid lidar depth values. All experiments are conducted with the original instead of half resolution as in [25] given that our method can handle high-res video input. We also discard scenes SPACE-OUT and WHEEL due to camera and lidar error. We find that the original camera annotations from ARKit is not accurate enough, and refine them using global bundle adjustment from COLMAP [85]. The refined poses are used for all methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 382 + ], + "type": "text", + "content": "We report the quantitative comparison in Tab. 1 which shows that our method outperforms all baselines on all tasks by a substantial margin. For tracking, we observe significant improvements over naive baselines that combine 2D tracks with depth maps, especially \"TAPIR+DA\", which also serves as input to our system, highlighting the effectiveness of our consolidation process. The improvement stems from three factors: 1) a low-dimensional " + }, + { + "bbox": [ + 313, + 192, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 192, + 555, + 382 + ], + "type": "text", + "content": " motion representation, which allows inferring occluded motion via nearby visible regions; 2) 3D regularizations, such as low acceleration constraints that promote coherent trajectories; 3) the consolidation of all pairwise tracks into a single 4D representation with persistent 3D geometry, which serves as a structural prior to correct noisy 2D tracks. Our method also achieves the best novel view synthesis quality across dynamic NeRF and 3D-GS-based baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 384, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 587 + ], + "type": "text", + "content": "Fig. 3 shows qualitative comparison of the 3D tracking results. We render the novel views and plot the predicted 3D tracks of the given query points onto the novel views. Since \"TAPIR + DA\" cannot perform novel view synthesis, we overlay their track predictions onto our renderings to aid interpretation. D-3DGS [119] and HyperNeRF [75] fail to capture the significant scene motion in PAPER-WINDMILL and SPIN, resulting in structure degradation and blurry rendering. \"TAPIR + DA\" can track large motions, but their 3D tracks tend to be noisy and erroneous. In contrast, our method not only generates high-quality novel views but also the most smooth and accurate 3D tracks. Fig. 4 provides additional novel view synthesis comparison on the validation views. In Fig. 5, we visualize the rendering of the first three components of PCA decomposition for the motion coefficients, which correlates well with the rigid groups of the moving objects." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 331, + 619, + 525, + 671 + ], + "blocks": [ + { + "bbox": [ + 314, + 594, + 477, + 605 + ], + "lines": [ + { + "bbox": [ + 314, + 594, + 477, + 605 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 477, + 605 + ], + "type": "text", + "content": "4.4. Evaluation on Kubric Dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 619, + 525, + 671 + ], + "lines": [ + { + "bbox": [ + 331, + 619, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 331, + 619, + 525, + 671 + ], + "type": "table", + "html": "
MethodEPE↓δ0.053D↑δ103D↑
CoTracker [44]+DA [118]0.1934.456.5
TAPIR [16]+DA [118]0.2034.056.2
Ours0.1639.862.2
", + "image_path": "29b746b8d0113c44b9453dc9f40c531baee4e1942dd0b0cd6f4843107c1542e1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 679, + 526, + 689 + ], + "lines": [ + { + "bbox": [ + 328, + 679, + 526, + 689 + ], + "spans": [ + { + "bbox": [ + 328, + 679, + 526, + 689 + ], + "type": "text", + "content": "Table 2. 3D Tracking evaluation on Kubric dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 701, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 712 + ], + "type": "text", + "content": "The Kubric MOVi-F dataset contains short 24-frame videos" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9666" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "of scenes of 10-20 objects, rendered with linear camera movement and motion blur. Multiple rigid objects are tossed onto the scene, at a speed that far exceeds the speed of the moving camera, making it a similar capture scenario to in-the-wild capture scenarios. It provides dense comprehensive annotations, including ground truth depth maps, camera parameters, segmentation masks, and point correspondences that are dense across time. We use 30 scenes from the MOVi-F validation set to evaluate the accuracy of our long-range 3D point tracks for all points in time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 299 + ], + "type": "text", + "content": "We demonstrate our method on the synthetic Kubric MOVi-F dataset in long-range 3D point tracking across time. Of the above baselines, only \"CoTracker [44] + DA [118]\" and \"TAPIR [44] + DA [118]\" yield 3D tracks for these scenes. For all baselines, we provide the ground truth camera intrinsics and extrinsics, and monocular depth estimates that have been aligned to the ground truth depth map. We obtain point tracks for all non-background pixels for each method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 294, + 359 + ], + "type": "text", + "content": "We report our quantitative 3D point tracking metrics in Table 2, and find that across all metrics, our method outperforms the baselines. Please see the supplement for figures of the optimized motion coefficients, which coherently group moving objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 224, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 224, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 224, + 379 + ], + "type": "text", + "content": "4.5. Evaluation on NVIDIA Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 294, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 504 + ], + "type": "text", + "content": "We conduct experiments on seven scenes from the NVIDIA dataset [122], following the evaluation protocol of Dynamic Gaussian Marbles [91], which evaluates the task from a single static camera instead of interleaving different camera as discussed in [25]. We adopt the same image resolution, camera parameters and depths maps as input. Please see the supplement for more details. For each scene, we compute covisibility masks between each evaluation view and the training view at each time step to exclude unobservable regions during evaluation. Results are reported in Tab. 3." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 514, + 294, + 573 + ], + "blocks": [ + { + "bbox": [ + 56, + 514, + 294, + 573 + ], + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 573 + ], + "type": "table", + "html": "
MethodBalloon1Ballon2JumpingTruck
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]23.830.810.06623.620.810.08820.020.730.13427.310.860.055
Ours23.900.790.05823.910.830.07420.080.690.15127.650.880.048
MethodPlaygroundUmbrellaSkatingMean
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
DGM [91]16.680.600.17724.990.700.08327.790.900.05223.460.770.093
Ours16.870.550.16823.270.570.16727.910.920.04723.370.750.102
", + "image_path": "70e8a78032ebdcf58034e91b069e73527e3c079c56222b977c6f6ba411cf6674.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 153, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 153, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 153, + 635 + ], + "type": "text", + "content": "4.6. Ablation studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "We ablate various components of our method on the iPhone dataset in 3D tracking in Tab. 4. We first validate our choices of motion representation, namely the " + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": " motion bases parameterization, with three ablations: 1) \"Per-Gaussian Transl\": replacing our motion representation with naive per-Gaussian translational motion trajectories," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 553, + 130 + ], + "blocks": [ + { + "bbox": [ + 55, + 580, + 294, + 603 + ], + "lines": [ + { + "bbox": [ + 55, + 580, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 294, + 603 + ], + "type": "text", + "content": "Table 3. Evaluation on NVIDIA dataset. Our method is comparable with Dynamic Gaussian Marbles (DGM) [91]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 553, + 130 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 130 + ], + "type": "table", + "html": "
MethodsSE(3)Motion Basis2D tracksInitializationEPE↓δ0.053D↑δ1.103D↑
Ours (Full)✓✓✓✓0.08243.073.3
Transl. Bases✓✓✓0.09342.369.9
Per-Gaussian SE(3)✓✓✓0.08343.670.2
Per-Gaussian Transl.✓✓0.08741.269.2
No SE(3) Init.✓✓✓0.11139.365.7
No 2D Tracks✓✓0.14130.457.8
", + "image_path": "f1a10bfde66aab45cb30daeb57087b0bc242bb2728c59dee01aed1b007d5f259.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 139, + 518, + 149 + ], + "lines": [ + { + "bbox": [ + 348, + 139, + 518, + 149 + ], + "spans": [ + { + "bbox": [ + 348, + 139, + 518, + 149 + ], + "type": "text", + "content": "Table 4. Ablation Studies on iPhone dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": "2) \"Per-Gaussian " + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": ": replacing our motion representation with naive per-Gaussian " + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": " motion trajectories, and 2) \"Transl. Bases\": keeping the motion bases representation but only using translational bases instead of " + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": ". We find that " + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": " bases significantly improve 3D tracking over both translational bases and per-Gaussian translational motion. It also outperforms per-Gaussian " + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 152, + 553, + 260 + ], + "type": "text", + "content": " in overall accuracy and offers noticeably better visual quality, as the latter exhibits significantly more popping and jitter artifacts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "text", + "content": "Next, we ablate our training strategies including initialization and supervision signal. We conduct an ablation of \"No " + }, + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "text", + "content": " Init.\", where instead of performing our initial " + }, + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "inline_equation", + "content": "\\mathbb{SE}(3)" + }, + { + "bbox": [ + 313, + 261, + 555, + 404 + ], + "type": "text", + "content": " fitting stage, we initialize the translational part of the motion bases with randomly selected noisy 3D tracks formed by directly lifting input 2D tracks using depth maps into 3D, and the rotation part as identity. We find that skipping this initialization noticeably hurts performance. Lastly, we remove the 2D track supervision entirely (\"No 2D Tracks\") and find it to lead to significant drop in performance, which verifies the importance of the 2D track supervision for 3D tracking." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 416, + 383, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 383, + 428 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 383, + 428 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 604 + ], + "type": "text", + "content": "Limitations. Like most prior monocular view synthesis methods, it still requires per-scene test-time optimization, hindering streamable applications. Recent feed-forward methods for joint reconstruction and tracking [22, 111, 113] offer a promising alternative, showing encouraging progress toward overcoming this limitation. Our method relies on off-the-shelf predictions for camera poses, geometry, and motion, which may degrade in textureless regions or under large motions. However, it also naturally benefits from advances in these components. Finally, it also requires user input to mask moving objects. Recent advances in moving object segmentation [27, 34] offer promising directions beyond our current manual solution and could be integrated into our pipeline." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Conclusion. We introduce a method for joint long-range 3D tracking and novel view synthesis from monocular video, representing dynamic elements via global 3D Gaussians with time-varying translations and rotations. Motion trajectories are regularized using low-dimensional rigid motion bases. By consolidating noisy observations into globally consistent scene estimates, our approach outperforms state-of-the-art methods in 2D/3D tracking and novel view synthesis across synthetic and real benchmarks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9667" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 161 + ], + "type": "text", + "content": "Acknowledgement. We thank Ruilong Li, Noah Snavely, Brent Yi and Aleksander Holynski for helpful discussion. We are in memory of our beloved cat Sriracha, who will always be missed and loved. This project is supported in part by DARPA No. HR001123C0021. and IARPA DOI/IBC No. 140D0423C0035. The views and conclusions contained herein are those of the authors and do not represent the official policies or endorsements of these institutions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 204, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 204, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 204, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 66, + 204, + 295, + 258 + ], + "type": "text", + "content": "[1] Aayush Bansal, Minh Vo, Yaser Sheikh, Deva Ramanan, and Srinivasa Narasimhan. 4d visualization of dynamic events from unconstrained multi-view videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 260, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 260, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 66, + 260, + 295, + 293 + ], + "type": "text", + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In European Conference on Computer Vision, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 295, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 295, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 66, + 295, + 294, + 327 + ], + "type": "text", + "content": "[3] Stanley T Birchfield and Shrinivas J Pundlik. Joint tracking of features and edges. In 2008 IEEE Conference on Computer Vision and Pattern Recognition. IEEE, 2008. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 329, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 329, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 66, + 329, + 294, + 361 + ], + "type": "text", + "content": "[4] Michael J Black and Padmanabhan Anandan. A framework for the robust estimation of optical flow. In 1993 (4th) International Conference on Computer Vision, 1993. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 364, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 364, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 66, + 364, + 295, + 418 + ], + "type": "text", + "content": "[5] Aljaz Bozic, Michael Zollhöfer, Christian Theobalt, and Matthias Nießner. Deepdeform: Learning non-rigid rgb-d reconstruction with semi-supervised data. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 66, + 420, + 294, + 463 + ], + "type": "text", + "content": "[6] C. Bregler, A. Hertzmann, and H. Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No.PR00662), 2000. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 66, + 465, + 295, + 508 + ], + "type": "text", + "content": "[7] Thomas Brox, Andrés Bruhn, Nils Papenberg, and Joachim Weickert. High accuracy optical flow estimation based on a theory for warping. In European Conference on Computer Vision, 2004. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 510, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 510, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 66, + 510, + 295, + 543 + ], + "type": "text", + "content": "[8] Thomas Brox, Christoph Bregler, and Jitendra Malik. Large displacement optical flow. 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 66, + 544, + 295, + 599 + ], + "type": "text", + "content": "[9] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. ACM Transactions on Graphics (TOG), 39(4), 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "type": "text", + "content": "[10] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 700 + ], + "type": "text", + "content": "[11] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skimming for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 703, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 294, + 713 + ], + "type": "text", + "content": "[12] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Dreamscene4d: Dynamic multi-object scene generation from monocular videos. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 128 + ], + "type": "text", + "content": "[13] Yuchao Dai, Hongdong Li, and Mingyi He. A simple prior-free method for non-rigid structure-from-motion factorization. Int. J. Comput. Vision, 107(2), 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 129, + 553, + 182 + ], + "type": "text", + "content": "[14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 553, + 239 + ], + "type": "text", + "content": "[15] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 2022. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 240, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 240, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 240, + 553, + 304 + ], + "type": "text", + "content": "[16] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 306, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 306, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 321, + 306, + 553, + 360 + ], + "type": "text", + "content": "[17] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, Joao Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 361, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 361, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 361, + 553, + 426 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Philipp Fischer, Eddy Ilg, Philip Häusser, Caner Hazirbas, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. 2015 IEEE International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 321, + 427, + 553, + 491 + ], + "type": "text", + "content": "[19] Mingsong Dou, S. Khamis, Yu.G. Degtyarev, Philip L. Davidson, S. Fanello, Adarsh Kowdle, Sergio Orts, Christoph Rhemann, David Kim, Jonathan Taylor, Pushmeet Kohli, Vladimir Tankovich, and Shahram Izadi. Fusion4d. ACM Transactions on Graphics (TOG), 35, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 321, + 494, + 553, + 548 + ], + "type": "text", + "content": "[20] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE Computer Society, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 549, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 549, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 549, + 553, + 601 + ], + "type": "text", + "content": "[21] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes. In ACM SIGGRAPH 2024 Conference Papers, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 321, + 604, + 553, + 657 + ], + "type": "text", + "content": "[22] Haiwen Feng, Junyi Zhang, Qianqian Wang, Yufei Ye, Pengcheng Yu, Michael J Black, Trevor Darrell, and Angjoo Kanazawa. St4rtrack: Simultaneous 4d reconstruction and tracking in the world. arXiv preprint arXiv:2504.13152, 2025.8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 659, + 553, + 713 + ], + "type": "text", + "content": "[23] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. Kplanes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9668" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 149 + ], + "type": "text", + "content": "[25] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Dynamic novel-view synthesis: A reality check. In NeurIPS, 2022. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 151, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 151, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 62, + 151, + 294, + 204 + ], + "type": "text", + "content": "[26] Leon A Gatys, Alexander S Ecker, Matthias Bethge, Aaron Hertzmann, and Eli Shechtman. Controlling perceptual factors in neural style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 294, + 259 + ], + "type": "text", + "content": "[27] Lily Goli, Sara Sabour, Mark Matthews, Marcus Brubaker, Dmitry Lagun, Alec Jacobson, David J Fleet, Saurabh Saxena, and Andrea Tagliasacchi. Romo: Robust motion segmentation improves structure from motion. arXiv preprint arXiv:2411.18650, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 262, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 325 + ], + "type": "text", + "content": "[28] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "type": "text", + "content": "[29] Xiuye Gu, Yijie Wang, Chongruo Wu, Yong Jae Lee, and Panqu Wang. Hplflownet: Hierarchical permutohedral lattice flownet for scene flow estimation on large-scale point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 425 + ], + "type": "text", + "content": "[30] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 427, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 470 + ], + "type": "text", + "content": "[31] Thorsten Hempel, Ahmed A Abdelrahman, and Ayoub Al-Hamadi. 6d rotation representation for unconstrained head pose estimation. In 2022 IEEE International Conference on Image Processing (ICIP), 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 472, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 492 + ], + "type": "text", + "content": "[32] Berthold K. P. Horn and Brian G. Schunck. Determining optical flow. In Other Conferences, 1981. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 494, + 294, + 536 + ], + "type": "text", + "content": "[33] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 538, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 294, + 590 + ], + "type": "text", + "content": "[34] Nan Huang, Wenzhao Zheng, Chenfeng Xu, Kurt Keutzer, Shanghang Zhang, Angjoo Kanazawa, and Qianqian Wang. Segment any motion in videos. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2025. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "type": "text", + "content": "[35] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. In European Conference on Computer Vision. Springer, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 648, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 648, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 648, + 294, + 712 + ], + "type": "text", + "content": "[36] Minyoung Huh, Richard Zhang, Jun-Yan Zhu, Sylvain Paris, and Aaron Hertzmann. Transforming and projecting images into class-conditional generative networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, 2020. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Junhwa Hur and Stefan Roth. Self-supervised monocular scene flow estimation. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 161 + ], + "type": "text", + "content": "[38] Mustafa Işik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 553, + 262 + ], + "type": "text", + "content": "[40] Matthias Innmann, Michael Zollhöfer, Matthias Nießner, Christian Theobalt, and Marc Stamminger. Volumedeform: Real-time volumetric non-rigid reconstruction. In European Conference on Computer Vision, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 553, + 308 + ], + "type": "text", + "content": "[41] Joel Janai, Fatma Guney, Anurag Ranjan, Michael Black, and Andreas Geiger. Unsupervised learning of multi-frame optical flow with occlusions. In Proceedings of the European conference on computer vision (ECCV), 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 309, + 553, + 363 + ], + "type": "text", + "content": "[42] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 408 + ], + "type": "text", + "content": "[43] Shihao Jiang, Yao Lu, Hongdong Li, and Richard Hartley. Learning optical flow from a few matches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 553, + 453 + ], + "type": "text", + "content": "[44] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European conference on computer vision, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 498 + ], + "type": "text", + "content": "[45] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 1, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 501, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 501, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 321, + 501, + 553, + 543 + ], + "type": "text", + "content": "[46] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 2023. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 321, + 545, + 553, + 577 + ], + "type": "text", + "content": "[47] Diederik P Kingma, J Adam Ba, and J Adam. A method for stochastic optimization. arxiv 2014. arXiv preprint arXiv:1412.6980, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 580, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 580, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 321, + 580, + 553, + 643 + ], + "type": "text", + "content": "[48] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 555, + 689 + ], + "type": "text", + "content": "[49] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 712 + ], + "type": "text", + "content": "[50] Skanda Koppula, Ignacio Rocco, Yi Yang, Joe Heyward, Joao Carreira, Andrew Zisserman, Gabriel Brostow, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9669" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "text", + "content": "[51] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. ECCV, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 295, + 196 + ], + "type": "text", + "content": "[52] Suryansh Kumar, Yuchao Dai, and Hongdong Li. Monocular dense 3d reconstruction of a complex dynamic scene from two perspective frames. 2017 IEEE International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 252 + ], + "type": "text", + "content": "[53] Jiahui Lei, Yijia Weng, Adam W Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 295, + 297 + ], + "type": "text", + "content": "[54] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhöfer, Jürgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. In European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 295, + 364 + ], + "type": "text", + "content": "[55] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 421 + ], + "type": "text", + "content": "[56] Zhengqi Li, Tali Dekel, Forrester Cole, Richard Tucker, Noah Snavely, Ce Liu, and William T Freeman. Learning the depths of moving people by watching frozen people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 476 + ], + "type": "text", + "content": "[57] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 533 + ], + "type": "text", + "content": "[58] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 578 + ], + "type": "text", + "content": "[59] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 295, + 645 + ], + "type": "text", + "content": "[60] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "text", + "content": "[61] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 714 + ], + "type": "text", + "content": "[62] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dy" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "namic gaussian splatting from causally-captured monocular videos. ICLR, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "text", + "content": "[63] Xingyu Liu, Charles R Qi, and Leonidas J Guibas. Flownet3d: Learning scene flow in 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 173 + ], + "type": "text", + "content": "[64] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 2004. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 175, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 230 + ], + "type": "text", + "content": "[65] Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "type": "text", + "content": "[66] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In International Joint Conference on Artificial Intelligence, 1981. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 553, + 319 + ], + "type": "text", + "content": "[67] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 354 + ], + "type": "text", + "content": "[68] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4), 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 320, + 355, + 553, + 399 + ], + "type": "text", + "content": "[69] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 553, + 443 + ], + "type": "text", + "content": "[70] Michal Neoral, Jonás Šerych, and Jiří Matas. Mft: Long-term tracking of every pixel. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 553, + 488 + ], + "type": "text", + "content": "[71] Richard A. Newcombe, Dieter Fox, and Steven M. Seitz. Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 533 + ], + "type": "text", + "content": "[72] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 578 + ], + "type": "text", + "content": "[73] David Novotny, Nikhila Ravi, Benjamin Graham, Natalia Neverova, and Andrea Vedaldi. C3dpo: Canonical 3d pose networks for non-rigid structure from motion. In ICCV, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 579, + 553, + 634 + ], + "type": "text", + "content": "[74] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 689 + ], + "type": "text", + "content": "[75] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. SIGGRAPH Asia, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 714 + ], + "type": "text", + "content": "[76] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9670" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 80, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 294, + 105 + ], + "type": "text", + "content": "for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 294, + 140 + ], + "type": "text", + "content": "[77] Gilles Puy, Alexandre Boulch, and Renaud Marlet. Flot: Scene flow on point clouds guided by optimal transport. In European Conference on Computer Vision, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 294, + 184 + ], + "type": "text", + "content": "[78] Rene Ranftl, Vibhav Vineet, Qifeng Chen, and Vladlen Koltun. Dense monocular depth estimation in complex dynamic scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 239 + ], + "type": "text", + "content": "[79] Zhile Ren, Orazio Gallo, Deqing Sun, Ming-Hsuan Yang, Erik B Sudderth, and Jan Kautz. A fusion approach for multi-frame optical flow estimation. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 274 + ], + "type": "text", + "content": "[80] Michael Rubinstein and Ce Liu. Towards longer long-range motion trajectories. In *British Machine Vision Conference*, 2012. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "text", + "content": "[81] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R. Bradski. Orb: An efficient alternative to sift or surf. 2011 International Conference on Computer Vision, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 310, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 352 + ], + "type": "text", + "content": "[82] Chris Russell, Rui Yu, and Lourdes Agapito. Video popu- up: Monocular 3d reconstruction of dynamic scenes. In European conference on computer vision. Springer, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 387 + ], + "type": "text", + "content": "[83] P Sand. Long-range video motion estimation using point trajectories. PhD thesis, Ph.D. dissertation, Cambridge, MA, USA, 2006, adviser-Teller, Seth, 2006. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 421 + ], + "type": "text", + "content": "[84] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 455 + ], + "type": "text", + "content": "[85] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 456, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 499 + ], + "type": "text", + "content": "[86] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 555 + ], + "type": "text", + "content": "[87] Xiaoyu Shi, Zhaoyang Huang, Weikang Bian, Dasong Li, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Videoflow: Exploiting temporal cues for multi-frame optical flow estimation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 589 + ], + "type": "text", + "content": "[88] Josef Sivic, Frederik Schaffalitzky, and Andrew Zisserman. Object level grouping for video shots. International Journal of Computer Vision, 2004. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 294, + 645 + ], + "type": "text", + "content": "[89] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5), 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[90] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[91] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 105 + ], + "type": "text", + "content": "Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. Siggraph Asia, 2024. 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 106, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 106, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 106, + 553, + 148 + ], + "type": "text", + "content": "[92] Timo Stich, Christian Linz, Georgia Albuquerque, and Marcus Magnor. View and time interpolation in image space. In Computer Graphics Forum. Wiley Online Library, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 150, + 553, + 194 + ], + "type": "text", + "content": "[93] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 553, + 238 + ], + "type": "text", + "content": "[94] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 239, + 553, + 271 + ], + "type": "text", + "content": "[95] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. Advances in neural information processing systems, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 273, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 273, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 273, + 553, + 316 + ], + "type": "text", + "content": "[96] Zachary Teed and Jia Deng. Raft-3d: Scene flow using rigid-motion embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 317, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 317, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 317, + 553, + 348 + ], + "type": "text", + "content": "[97] Chaoyang Wang, Ben Eckart, Simon Lucey, and Orazio Gallo. Neural trajectory fields for dynamic novel view synthesis. arXiv preprint arXiv:2105.05994, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 392 + ], + "type": "text", + "content": "[98] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 394, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 553, + 426 + ], + "type": "text", + "content": "[99] Heng Wang and Cordelia Schmid. Action recognition with improved trajectories. 2013 IEEE International Conference on Computer Vision, 2013. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "text", + "content": "[100] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "type": "text", + "content": "[101] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In International Conference on Computer Vision, 2023. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "type": "text", + "content": "[102] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "text", + "content": "[103] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[104] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[105] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9671" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[106] Zirui Wang, Shuda Li, Henry Howard-Jenkins, Victor Adrian Prisacariu, and Min Chen. Flownet3d++: Geometric losses for deep scene flow estimation. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "type": "text", + "content": "[107] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 239 + ], + "type": "text", + "content": "[108] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "text", + "content": "[109] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 337 + ], + "type": "text", + "content": "[110] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 392 + ], + "type": "text", + "content": "[111] Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Yuri Makarov, Bingyi Kang, Xing Zhu, Hujun Bao, Yujun Shen, and Xiaowei Zhou. Spatialtrackerv2: 3d point tracking made easy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "text", + "content": "[112] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 481 + ], + "type": "text", + "content": "[113] Zhen Xu, Zhengqin Li, Zhao Dong, Xiaowei Zhou, Richard Newcombe, and Zhaoyang Lv. 4dgt: Learning a 4d gaussian transformer using real-world monocular videos. arXiv preprint arXiv:2506.08015, 2025. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 547 + ], + "type": "text", + "content": "[114] Gengshan Yang, Deqing Sun, V. Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T. Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 603 + ], + "type": "text", + "content": "[115] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "text", + "content": "[116] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 691 + ], + "type": "text", + "content": "[117] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 715 + ], + "type": "text", + "content": "[118] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Un" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 586 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "text", + "content": "leashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[119] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[120] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. *ICLR*, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 262 + ], + "type": "text", + "content": "[121] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. Journal of Machine Learning Research, 26(34), 2025. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 319 + ], + "type": "text", + "content": "[122] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[123] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arxiv:2410.03825, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "type": "text", + "content": "[124] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "text", + "content": "[125] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "text", + "content": "[126] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 586 + ], + "type": "text", + "content": "[127] Michael Zollhöfer, Matthias Nießner, Shahram Izadi, Christoph Rhemann, Christopher Zach, Matthew Fisher, Chenglei Wu, Andrew William Fitzgibbon, Charles T. Loop, Christian Theobalt, and Marc Stamminger. Real-time non-rigid reconstruction using an rgb-d camera. ACM Transactions on Graphics (TOG), 33, 2014. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9672" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_content_list.json b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b447cd1a22f6d8755f5a1ec54ce356dc65c6b433 --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_content_list.json @@ -0,0 +1,1496 @@ +[ + { + "type": "text", + "text": "SHORTFT: Diffusion Model Alignment via Shortcut-based Fine-Tuning", + "text_level": 1, + "bbox": [ + 137, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiefan Guo $^{1,2}$ Miaomiao Cui Liefeng Bo Di Huang $^{1,2*}$", + "bbox": [ + 254, + 179, + 746, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing 100191, China $^{2}$ School of Computer Science and Engineering, Beihang University, Beijing 100191, China", + "bbox": [ + 122, + 205, + 875, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xfquo, dhuang}@buaa.edu.cn", + "bbox": [ + 382, + 242, + 612, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project: https://xiefan-guo.github.io/shortft", + "bbox": [ + 289, + 258, + 700, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb45afea23bbd1f306fb7e363ff0e5c51de2ac14cd178b9320bcc7344838c899.jpg", + "image_caption": [ + "Figure 1. Example results synthesized by SHORTFT. SHORTFT endeavors to achieve the alignment of diffusion models with reward functions by facilitating the end-to-end backpropagation of the targeted reward gradient throughout the denoising chain. Our method has exhibited remarkable efficacy, particularly evident in the realms of text-image alignment and the overall enhancement of image quality (Top). Moreover, its versatility is underscored by the successful application across diverse reward functions, substantially amplifying alignment performance (Bottom). Combined reward is a weighted combination of rewards: PickScore = 10, HPS v2 = 2, Aesthetic = 0.05." + ], + "image_footnote": [], + "bbox": [ + 93, + 305, + 903, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 710, + 326, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backpropagation-based approaches aim to align diffusion models with reward functions through end-to-end backpropagation of the reward gradient within the denoising chain, offering a promising perspective. However, due to the computational costs and the risk of gradient explosion associated with the lengthy denoising chain, existing approaches struggle to achieve complete gradient backpropagation, leading to suboptimal results. In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), an effi", + "bbox": [ + 88, + 742, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cient fine-tuning strategy that utilizes the shorter denoising chain. More specifically, we employ the recently researched trajectory-preserving few-step diffusion model, which enables a shortcut over the original denoising chain, and construct a shortcut-based denoising chain of shorter length. The optimization on this chain notably enhances the efficiency and effectiveness of fine-tuning the foundational model. Our method has been rigorously tested and can be effectively applied to various reward functions, significantly improving alignment performance and surpassing state-of-the-art alternatives.", + "bbox": [ + 511, + 712, + 908, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "678", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df8fd5b04a30b4043b3b9758e4f2bfb39986e993be1df3cd6902731f41652753.jpg", + "image_caption": [ + "(a) Vanilla backpropagation", + "(c) AlignProp, DRTune: Truncated backpropagation within the denoising chain" + ], + "image_footnote": [], + "bbox": [ + 101, + 99, + 493, + 143 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0d1e554b1ad63cd4487ec725e8f4a088c6379518f9ff17683997e197df3b58bc.jpg", + "image_caption": [], + "image_footnote": [ + "Sample $(\\theta_{\\mathrm{diffusion}},\\mathbf{c},\\mathbf{z})$", + "- Vanilla backpropagation" + ], + "bbox": [ + 101, + 162, + 493, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a6c79e18b1d266b7ad2c0c2fd328436f4ebea9138349db98ca970384c88a391.jpg", + "image_caption": [ + "(b) DRaFT-K, ReFL: Truncating backpropagation at the end of the denoising chain" + ], + "image_footnote": [], + "bbox": [ + 503, + 99, + 895, + 143 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d1f5902ba3aba90b26d987238069734b662def7ef6b8a67da292b6efafd741c8.jpg", + "image_caption": [ + "(d) SHORTFT (Ours): Shortcut-based backpropagation", + "Figure 2. Comparison of fine-tuning strategies. (a) The vanilla backpropagation-based fine-tuning strategy, which suffers from lengthy backpropagation chains. (b) DRaFT-K [6] and ReFL [52] truncate the backpropagation chain, focusing on the latter half of the denoising chain, they ignore the direct supervision at the early stage, resulting in suboptimal alignment with text prompts. (c) AlignProp [34] and DRTune [51] truncate part of the backpropagation within the denoising chain by disabling some gradients. Specifically, given the denoising operation: $\\mathbf{x}_{t-1} = \\alpha_t \\mathbf{x}_t + \\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t) + c_t \\epsilon$ , partial backpropagation only utilizes the gradients of the red part ( $\\alpha_t \\mathbf{x}_t$ ), truncating the green part ( $\\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t)$ ). This inevitably introduces gradient errors, leading to unstable optimization. (d) Our method, using the few-step diffusion model to construct denoising shortcuts, facilitates complete gradient backpropagation through the entire denoising chain." + ], + "image_footnote": [ + "←- Partial backpropagation", + "Sample( $\\theta_{\\mathrm{shortcut}}$ ,c,z)", + "←- Shortcut-based backpropagation" + ], + "bbox": [ + 503, + 159, + 895, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 351, + 220, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [1, 9, 17, 18, 22, 32, 38, 44, 45] have established themselves as a pioneering approach in generative modeling, demonstrating exceptional prowess in applications such as photo-realistic text-to-image synthesis. However, the maximum likelihood training objective of these models, which aims to model the training data distribution accurately, can often conflict with downstream goals like aesthetics, fairness, safety, and text-to-image alignment. Therefore, aligning text-to-image diffusion models with human preferences has emerged as a pivotal and practical task.", + "bbox": [ + 89, + 378, + 482, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Directly supervised fine-tuning on small-scale, human-curated datasets, e.g., LAION Aesthetics [42], presents a straightforward solution. However, the prohibitive cost of data collection and the rapid obsolescence of datasets, particularly in terms of resolution compatibility with the latest text-to-image models, make this approach impractical.", + "bbox": [ + 89, + 532, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Emulating the successful application of Reinforcement Learning from Human Feedback (RLHF) [5, 13, 33, 47, 59] in Large Language Models (LLMs), several studies [3, 10, 12, 25, 50] have experimented with Reinforcement Learning (RL) techniques to align diffusion models with a reward function. Despite promising performance enhancements in specific domains, RL-based methods are notorious for their high-variance gradients, leading to inefficiencies and limited adaptability to diverse prompts.", + "bbox": [ + 89, + 626, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, backpropagation-based methods [6, 34, 51, 52] have sought to align diffusion models with reward functions using end-to-end backpropagation of the reward gradient through the denoising chain, showing potential. However, these strategies face challenges arising from the lengthy denoising chain, which demands considerable computational resources and is susceptible to gradient explosion. As illustrated in Fig. 2 (b), [6, 52] have made strides by truncating backpropagation to focus on the latter part of the de", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "noising chain. However, they overlook direct supervision in the early stages, leading to suboptimal alignment with text prompts. As illustrated in Fig. 2 (c), [34, 51] truncate a portion of backpropagation within the denoising chain by deactivating some gradients and introduce gradient checkpointing, enabling gradient backpropagation to the early stage of the denoising chain. However, they are time-consuming and introduce gradient bias, leading to optimization instability.", + "bbox": [ + 511, + 352, + 906, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper revisits the issue of excessively long denoising chain and propose an alternative approach, employing the shorter denoising chain to facilitate full gradient backpropagation throughout the entire denoising chain.", + "bbox": [ + 511, + 474, + 905, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), which leverages the few-step diffusion model to construct a denoising shortcut to fine-tune the foundational model (e.g., SD 1.5), inspired by recent trajectory-preserving diffusion distillation methods [23, 37, 39, 46]. This technique enables us to bypass the original lengthy denoising chain and complete the inference, creating a shortcut-based denoising chain of shorter length. In addition, we construct a timestep-aware LoRA as an expert LoRA ensemble, based on the intriguing temporal dynamics exhibited by the text-to-image diffusion model during the denoising process. This approach increases the number of trainable parameters without increasing cost in the inference phase, enabling faster convergence and improved performance. We also devise a custom progressive training strategy to mitigate training inference bias introduced by using denoising shortcuts during the training stage.", + "bbox": [ + 511, + 537, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive quantitative and qualitative analyses demonstrate that SHORTFT can be effectively applied to various reward functions and architectures, significantly enhancing alignment performance. Furthermore, SHORTFT, benefiting from the short denoising chain and without the need for gradient checkpointing, is particularly efficient, learning faster than DRTune [51], the current most efficient method.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "679", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e14e44833c570bbebe4aeb01441cec2603f28e7aef93b3ce283f64f7f6b1bdc2.jpg", + "image_caption": [ + "Figure 3. Denoising shortcut. The trajectory-preserving few-step diffusion model naturally introduces denoising shortcut, allowing for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. The 4-step Hyper-SD distilled from SD 1.5 is used in our experiments. SHORTcut(i) denotes completing the denoising process from timestep $i$ to 0 using the few-step diffusion model. In addition, we also provide the well-known one-step denoising results $\\mathrm{DDIM}(i)$ , where $\\mathrm{DDIM}(i)$ denotes performing a one-step denoising operation from timestep $i$ to 0. Intuitively, the output of $\\mathrm{DDIM}(i)$ is more blurred, lacking accurate structure and texture details, while the output of SHORTcut(i) is closer to the original output of SD 1.5. Furthermore, we provide their corresponding HPS v2 scores, with the absolute value of the deviation from the score corresponding to SD 1.5 provided in parentheses, SHORTcut(i) exhibits smaller deviations. These observations collectively indicate the reliability and validity of the denoising shortcut." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 903, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 323, + 230, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Alignment of diffusion models", + "text_level": 1, + "bbox": [ + 89, + 347, + 357, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [9, 18, 44, 45] have become a dominant force in generative modeling, showing exceptional performance in diverse applications [1, 14-17, 22, 32, 38, 53, 57]. However, certain misalignments with human intentions can arise. Recent research, fueled by the successful alignment of large language models, has sought to align diffusion models with human expectations and preferences.", + "bbox": [ + 89, + 369, + 482, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning via data augmentation. Several studies [7, 10, 20, 25, 50] have explored altering the training data distribution for fine-tuning diffusion models on visually compelling and textually coherent data, which has led to improved results. Other methods [2, 43] involve re-captioning pre-collected web images to enhance textual precision.", + "bbox": [ + 89, + 477, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning via reward models. Reward models [24, 25, 42, 49, 50, 52] are employed to emulate human preferences given an input prompt and generated images. Several approaches have attempted to integrate these signals to augment text-to-image generation. A significant direction is the utilization of reinforcement learning-based algorithms [3, 4, 8, 11, 58] for fine-tuning text-to-image diffusion models in alignment with these rewards, [29, 30, 36, 48, 54, 55] bypass it entirely with Direct Preference Optimization. However, these methods are costly and have high gradient variance, leading to inefficiency and limited adaptability to diverse prompts. Consequently, backpropagation-based techniques [6, 34, 51, 52] have been explored, which directly fine-tune diffusion models using differentiable rewards [24, 42, 49, 50, 52].", + "bbox": [ + 89, + 566, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The challenge of backpropagation-based strategies stem from the lengthy denoising chain, which often requires numerous denoising operations (e.g., 50 for DDIM), corresponding to a long backpropagation chain. This process incurs substantial time and memory costs and is prone to gradient explosion. To mitigate this issue, [6, 52] truncate backpropagation by concentrating on the latter part", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the denoising chain. While these approaches yield some improvements, they neglect direct supervision in the early stage of the denoising chain, leading to less precise alignment with text prompts. [34, 51] truncate part of the backpropagation within the denoising chain by deactivating some gradients. By employing gradient checkpointing, they allows for the propagation of gradients to the early stages of the denoising chain. Despite their merits, these techniques can be computationally demanding and induce gradient bias, resulting in optimization instability.", + "bbox": [ + 511, + 324, + 903, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from previous methods, this paper revisits the fundamental challenge of excessively long denoising chain and proposes an alternative approach: leveraging the shorter denoising chain to facilitate full gradient backpropagation throughout the denoising chain.", + "bbox": [ + 511, + 474, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion distillation", + "text_level": 1, + "bbox": [ + 511, + 560, + 705, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing methods for the distillation of diffusion models can be primarily classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]. The former aims to preserve the original denoising trajectory dictated by an ordinary differential equation (ODE), while the latter focuses on leveraging the denoising endpoint as the main supervision, disregarding the intermediate trajectory steps. This paper focuses on trajectory-preserving distillation, supporting the establishment of a denoising shortcut within the complete denoising chain.", + "bbox": [ + 511, + 582, + 905, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Fine-tuning few-step diffusion models", + "text_level": 1, + "bbox": [ + 511, + 757, + 836, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing works [26-28] have successfully explored finetuning few-step diffusion models. While they share similarities with our research in utilizing few-step diffusion models, a crucial difference is that we focus on fine-tuning the foundational models. In contrast to the foundational models, few-step diffusion models suffer from performance degradation and reduced capacity caused by the distillation process, leading fine-tuning them suboptimal (see Sec. 4.4).", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "680", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a28ef26760df40c1dfbc6a55a089b37d55d8ec57dfeb47d4d19ad7d13e8cb83b.jpg", + "image_caption": [ + "Figure 4. Illustration of SHORTFT. The core of the proposed method is the Shortcut-based Fine-Tuning (SHORTFT), which leverages the trajectory-preserving few-step diffusion model as the shortcut (identified as blue arrow) to achieve direct end-to-end backpropagation through the diffusion sampling process, fine-tuning the parameters of the pre-trained diffusion model to align it with the reward function." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 898, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. SHORTFT", + "text_level": 1, + "bbox": [ + 89, + 246, + 202, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method, Shortcut-based Fine-Tuning (SHORTFT), capitalizes on the trajectory-preserving few-step diffusion model as a shortcut to achieve direct end-to-end backpropagation through the diffusion sampling process. This approach fine-tunes the parameters of the pre-trained diffusion model to align it with the reward function.", + "bbox": [ + 89, + 272, + 483, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 375, + 290, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In line with [6, 34, 51], SHORTFT focuses on fine-tuning the parameters $\\theta$ of pre-trained diffusion models to maximize the differentiable reward function $\\mathcal{R}(\\cdot)$ for generated images. This can be formally represented as in Eq. 1:", + "bbox": [ + 89, + 397, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {T} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} [ \\mathcal {R} (\\text {S a m p l e} (\\boldsymbol {\\theta}, \\mathbf {c}, \\mathbf {x} _ {T}), \\mathbf {c}) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 470, + 482, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Sample}(\\theta, \\mathbf{c}, \\mathbf{x}_T)$ represents the denoising process for the timestep $t = T \\to 0$ with prompt condition $\\mathbf{c}$ .", + "bbox": [ + 89, + 501, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consistent with [6, 34, 51], Eq. 1 is resolved by calculating $\\nabla \\mathcal{R}$ (Sample $(\\theta ,\\mathbf{c},\\mathbf{x}_T)$ , c) and employing gradient ascent. The computation of this gradient necessitates backpropagation through multiple diffusion models in the denoising chain, akin to backpropagation through time in recurrent neural networks.", + "bbox": [ + 89, + 532, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Denoising shortcut", + "text_level": 1, + "bbox": [ + 89, + 635, + 272, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The recent emergence of a series of diffusion-aware distillation algorithms [23, 37, 39, 46] has been instrumental in mitigating the computational burden associated with the multi-step inference process of diffusion models. These algorithms can be roughly classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56].", + "bbox": [ + 89, + 657, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among these, trajectory-preserving few-step diffusion models naturally introduce a denoising shortcut, which allows for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. As demonstrated in Fig. 3, the integration of these shortcuts into the denoising chain significantly reduces the total number of denoising steps, thereby shortening the length of the denoising chain. This key insight paves the way for efficient and effective end-to-end backpropagation.", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bff8c66218c5a096b0eac7fe9122ad0c44ca29d3563e12531ff9f6c7d3bee0a5.jpg", + "image_caption": [ + "(a) SD denoising chain" + ], + "image_footnote": [], + "bbox": [ + 524, + 256, + 893, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9cc46b97ab6bdf01e9754f9c0d251fde482e8da6eb24cfe99606f103cfbf0d29.jpg", + "image_caption": [ + "(b) Shortcut-based denoising chain", + "Figure 5. Illustration of timestep-aware LoRA. (a) Vanilla SD denoising chain; (b) Shortcut-based denoising chain. In accordance with the interesting time dynamics in the text-to-image diffusion model denoising process revealed by [1], different from existing methods that share the same LoRA parameters at all timesteps, we introduce time-step aware LoRA, which effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing the computational cost during the inference stage." + ], + "image_footnote": [], + "bbox": [ + 524, + 314, + 893, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Shortcut-based fine-tuning", + "text_level": 1, + "bbox": [ + 513, + 522, + 754, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Insight. The naive optimization of Eq. 1 through backpropagation necessitates the construction of the complete denoising chain: $\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\dots ,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\dots ,$ $\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}$ . This process involves the storage of intermediate activations linked to each neural layer and each denoising timestep within GPU VRAM, which is not feasible due to memory constraints. Furthermore, the typical length of a denoising chain is approximately 50, which results in an overly long backpropagation chain that can lead to issues of gradient explosion.", + "bbox": [ + 511, + 546, + 906, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key insight of SHORTFT is rooted in the utilization of the trajectory-preserving few-step diffusion model to construct the denoising shortcut, significantly reducing the length of the denoising chain. As elaborated in Sec. 3.2, this denoising shortcut enables us to bypass a substantial number of denoising timesteps, leading to create a more streamlined and efficient denoising chain: $\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}$ .", + "bbox": [ + 511, + 699, + 906, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 4, such a design allows for the direct implementation of reward supervision at the early stages of the denoising chain, and facilitates full gradient backpropagation throughout the denoising chain.", + "bbox": [ + 511, + 823, + 905, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shortcut-based denoising chain. As illustrated in Fig. 5,", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "681", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7a8a5ada6bc548c7d540ab54d0d17c5f164f2cdddf68b1f9207b9139d6bf506.jpg", + "image_caption": [ + "Figure 6. Illustration of progressive training strategy. Corresponding to the time-step aware LoRA design, we develop the progressive training strategy, which eliminates the training-inference gap introduced by the shortcut-based fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 473, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a vanilla SD denoising chain comprises a sequence of step-by-step denoising operations that transform the input noise $\\mathbf{z}_T$ into the output image $\\mathbf{z}_0$ . This process typically necessitates numerous denoising steps. By harnessing the denoising shortcut, we are able to construct a shortcut-based denoising chain and fine-tune the diffusion model through end-to-end backpropagation.", + "bbox": [ + 89, + 363, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Timestep-aware LoRA. Instead of fine-tuning the full weights of the original diffusion model, Low-Rank Adaptation (LoRA) [19] preserves the weights of the pre-trained model and introduces new low-rank weight matrices alongside the original model weights. The contributions of these matrices are summed to generate the final outputs. Specifically, each linear layer of the UNet of SD is modified from $\\mathbf{h} = \\mathbf{W}\\mathbf{x}$ to $\\mathbf{h} = \\mathbf{W}\\mathbf{x} + \\mathbf{B}\\mathbf{A}\\mathbf{x}$ , where $\\mathbf{W} \\in \\mathbb{R}^{d \\times d}$ , $\\mathbf{B} \\in \\mathbb{R}^{d \\times k}$ , $\\mathbf{A} \\in \\mathbb{R}^{k \\times d}$ , and $k \\ll d$ . LoRA considerably reduces the number of parameters to be optimized, thereby decreasing the memory requirements for fine-tuning.", + "bbox": [ + 89, + 472, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, [1] uncovers intriguing temporal dynamics during the denoising process of the text-to-image diffusion model. In the initial sampling stage, the model largely depends on the text prompts to guide the sampling process. As the generation progresses, the model gradually leans on visual features to denoise the image. This indicates that sharing LoRA parameters (standard practice in the existing methods) throughout the entire denoising process may not be optimal and may fail to capture the distinct patterns that emerge during denoising. Therefore, in contrast to [6, 34, 51, 52] which share the same LoRA parameters for all timesteps, we introduce timestep-aware LoRA. This design effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing computational cost in the inference phase.", + "bbox": [ + 89, + 641, + 482, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as depicted in Fig. 5, we initially divide the entire denoising chain into $k$ segments, $\\Delta t = \\left\\lfloor \\frac{T}{k}\\right\\rfloor$ . Ex-", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cept for the first segment, we assign a corresponding LoRA to the last timestep of each subsequent segment. For the first segment, we adopt the methodology of [6] and share the same LoRA parameters across all timesteps. Notably, a sequence of continuous timesteps lacking LoRA exists in later segments, supporting the denoising shortcuts.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Progressive training strategy. Although the few-step diffusion model is capable of creating the denoising shortcut, bypassing the entire denoising chain, it inherently introduces errors in the output, meaning it still cannot be fully consistent with the output of SD, particularly in fine details, as shown in Fig. 3. This incongruity can lead to a training-inference gap, resulting in suboptimal results. To mitigate this, we design a progressive training strategy.", + "bbox": [ + 511, + 181, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 6, in accordance with the design of timestep-aware LoRA, we divide the SHORTFT training process into $k$ stages. For the $i$ -th training stage, we optimize the weights of LoRA $i$ to LoRA $k$ . For the $i$ -th segment and preceding denoising processes, we retain the original denoising chain, while for the denoising processes post the $i$ -th segment, we introduce the denoising shortcut, thereby shortening the depth of the backpropagation chain. Moreover, in line with [6], we also employ the truncated backpropagation technique at the $i$ -th segment.", + "bbox": [ + 511, + 304, + 905, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, as displayed in Fig. 6, the denoising shortcut is bypassed, and the original denoising chain is used to generate the final output image.", + "bbox": [ + 511, + 455, + 905, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 518, + 645, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 511, + 542, + 718, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, Stable Diffusion 1.5 serves as the foundational diffusion model. The DDIM schedule [44] is employed to execute 50 steps of denoising, with a classifier-free guidance scale of 7.5.", + "bbox": [ + 511, + 566, + 905, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shortcut. SHORTFT shortens the denoising chain by bypassing certain steps within the chain. Hence, the selection of few-step diffusion models is focused predominantly on methods that employ trajectory-preserving distillation algorithms. To accommodate the proposed time-aware LoRA, one-step diffusion models are avoided, whose output image quality is also relatively inferior. Specifically, 4-step Hyper-SD [37], distilled from SD 1.5, is utilized to construct the denoising shortcut. The value of $k$ is set to 4, and the timesteps configured for LoRA are {761, 501, 261, 1}. Consequently, the denoising shortcuts are executed separately between timesteps 741 to 501, timesteps 481 to 261, and timesteps 241 to 1.", + "bbox": [ + 511, + 628, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Timestep-aware LoRA. As suggested by [6], LoRA is applied to both the feedforward and attention layers in the UNet. The LoRA rank is set to 128. Furthermore, we adopt a stepwise stacking approach, where for LoRA $i$ , we introduce a new LoRA branch on top of LoRA $i - 1$ .", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "682", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b04d6475d0b1c018e64c0032ddea6c5039d1fee22ac6d01f524b4a78ae778b9.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on PickScore and HPS v2. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 874, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reward functions. The proposed method is evaluated using three reward functions: Human Preference Score v2 (HPS v2) [49], PickScore [24], and Symmetry [51]. HPS v2 and PickScore capture human preference for images based on input prompts, while Symmetry encourages images to have horizontal symmetry features. Different from [51], which utilizes CLIPScore [35] as a regularization term, our experiment amalgamates HPS v2 and PickScore in a ratio of 1:10 to function as a joint regularization term. This approach has demonstrated superior performance in terms of text-image alignment and overall image quality.", + "bbox": [ + 88, + 507, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental details. All experiments are conducted using 2 A800 GPUs and the AdamW optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and a weight decay of 0.1. SHORTFT is performed with a batch size of 128 and a constant learning rate of $5 \\times 10^{-5}$ . During training, the pre-trained SD parameters are converted to bfloat16 to reduce memory usage, while the LoRA parameters under training remain in float32. Gradient checkpointing is not required.", + "bbox": [ + 89, + 674, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We compare SHORTFT-finetuned diffusion models to those of the state-of-the-art counterparts on the Human Preference Score v2 dataset (HPDv2) [49]. The final reward is computed on the 400 prompts from the test split. Following [51], for fair comparison, we evaluate all methods using the same computational budget. Specifically, all methods are trained for six hours on 2 A800 GPUs.", + "bbox": [ + 89, + 795, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d90aaa1213146e7cc2555afe4ddf7049811dbe0e58005012d4d9b9544836fd9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHPS v2↑PickScore↑Symmetry↓
SD 1.5 [38]26.9120.460.853
DRaFT-LV [6]33.1323.350.418
DRTune [51]32.7923.220.207
SHORTFT33.8824.160.138
", + "bbox": [ + 537, + 505, + 885, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Objective evaluation. Our method performs over other counterparts, under the same computational cost.", + "bbox": [ + 511, + 601, + 906, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative comparison", + "text_level": 1, + "bbox": [ + 511, + 648, + 728, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 7 and 8 present the quantitative comparison of our results against those of representative methods, including current state-of-the-art techniques, under same text prompts and random seeds. SD 1.5 suffers from low-quality image generation. DRaFT [6], while proficient at managing local image details, struggles to effectively handle global layouts and the optimization of the symmetry reward function. Similarly, DRTune [51] is constrained by gradient inaccuracies, which contribute to instability during training and a deficiency in managing complex semantics. One particular area of weakness is the synthesis of images with specific counts of objects, such as accurately depicting a given number of cows. In contrast, SHORTFT exhibits enhanced capabilities in generating images that are both visually realistic and semantically faithful, outperforming over other counterparts.", + "bbox": [ + 509, + 674, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "683", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/066169bad1c7248978d8e5c9b23bd445a964eb7a40b25ed4795988fdb52eaa76.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison on HPS v2 and Symmetry. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + ], + "image_footnote": [], + "bbox": [ + 120, + 89, + 874, + 460 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/401a6f5f6634bf4694c4774099fe5225bb898ee0bf0213aec23bfc81132ecd67.jpg", + "image_caption": [ + "Figure 9. Human and AI preference evaluation against current methods. SHORTFT performs over other counterparts." + ], + "image_footnote": [], + "bbox": [ + 109, + 508, + 460, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative comparison", + "text_level": 1, + "bbox": [ + 89, + 619, + 318, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Objective evaluation. Table 1 shows the quantitative results achieved on the Human Preference Score v2 benchmark [49], where the proposed method outperforms the other approaches, clearly demonstrating its effectiveness. It is important to highlight that both our method and DRTune [51] strategically employ backpropagation of the reward gradient to the initial stages of the denoising chain. This intentional design choice significantly enhances the Symmetry score performance when compared to alternative methods. Despite this, DRTune continues to grapple with the challenge of gradient bias. In contrast, our method significantly mitigates this issue, delivering superior performance. User study. A subjective user study comprising 11 volunteers is conducted, with five possessing expertise in image processing and the remaining participants having no background in computer vision or graph. Participants are tasked to select the most visually attractive and semantically ac", + "bbox": [ + 88, + 642, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "curate image among those generated by our method and current state-of-the-art techniques. Each participant has 10 questions for each pair of comparisons. Furthermore, an MLLM-assisted evaluation is employed using GPT-4V. We make 20 queries to GPT-4V for each pair of comparisons. More details are provided in the Appendix. As depicted in Fig. 9, the results exhibit a significant inclination towards SHORTFT in comparison to other techniques.", + "bbox": [ + 511, + 511, + 906, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. More results", + "text_level": 1, + "bbox": [ + 511, + 642, + 648, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SHORTFT, 10k training step. Due to the shorter denoising and backpropagation chains, our method achieves superior performance under the same computational budget. Furthermore, to validate the upper bound of our approach, following the protocol in [6], we conduct the full training process using HPS v2 reward function on HPDv2, comprising 10k training steps, and evaluate it on the corresponding benchmark. The obtained HPS v2 score of 35.97 surpasses the reported score for DRaFT-LV in [6].", + "bbox": [ + 511, + 664, + 906, + 800 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b831730447c47c2ae34f4474d16b44a618525060396a1996a9e276ee7492805e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTuning Hyper-SDTuning SD 1.5
HPS v2↑32.9235.97
", + "bbox": [ + 568, + 813, + 854, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Objective evaluation on tuning SD 1.5 and Hyper-SD.", + "bbox": [ + 517, + 859, + 898, + 876 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-tuning SD vs. Hyper-SD. As mentioned in Sec. 2.3,", + "bbox": [ + 511, + 885, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "684", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9793504cf4bae9a8204d59ee35176efc78f931fa676aa14438c9dbedeadb2649.jpg", + "image_caption": [ + "Figure 10. Qualitative comparison on tuning SD 1.5 (SHORTFT) and Hyper-SD. Fine-tuning SD 1.5 significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[26-28] explore fine-tuning few-step diffusion models and have achieved certain successes. However, compared to the foundational model, few-step diffusion models face performance degradation caused by the distillation process. Finetuning the few-step diffusion model is actually suboptimal compared to fine-tuning the foundational model. Furthermore, following two strategies, we separately conduct the training processes on the HPDv2 using the HPS v2 reward function. As shown in Table 2 and Fig. 10, fine-tuning SD 1.5 (SHORTFT) significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details.", + "bbox": [ + 89, + 204, + 483, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a3dcd691a9729e6d9aab4d09235fad431334f3739ed423ddd16735b1fb4f841.jpg", + "image_caption": [ + "Figure 11. Example results synthesized by SHORTFT on SD 3." + ], + "image_footnote": [], + "bbox": [ + 94, + 382, + 480, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning SD 3. SHORTFT is an architecture-agnostic fine-tuning strategy, applicable to both UNet-based (SD 1.5) and Transformer-based (SD 3) architectures. As illustrated in Fig. 11, SHORTFT is also capable of mastering SD 3, where SD 3 aligns with HPS v2.", + "bbox": [ + 89, + 521, + 483, + 595 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0e476c64b57777984eb424274a223b9f13ee8d140e3d1a974cc426f7ec67399.jpg", + "image_caption": [ + "Figure 12. Generalization to wild text prompts from Sora. Our method is capable of effectively handling the wild text prompts." + ], + "image_footnote": [], + "bbox": [ + 96, + 607, + 478, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other reward functions. SHORTFT exhibits remarkable versatility, demonstrating efficacy across a spectrum of reward functions, significantly improving the alignment performance and the quality and fidelity of the output. As shown in Fig. 1 and 12, SHORTFT not only accommodates HPS v2, PickScore, and Symmetry, but also exhibits profi", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c16d52c13f0f3fc8e0bd921ad084060b2b59e1c6e5af983531dfbd317d0fe83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHPS v2↑PickScore↑Symmetry↓
w/o T-LoRA33.4623.820.187
w/o P-Training33.2723.970.146
SHORTFT33.8824.160.138
", + "bbox": [ + 545, + 88, + 877, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on timestep-aware LoRA.", + "bbox": [ + 555, + 167, + 861, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00c11bb04059c56e6b37cbe69496c8d16538b52d331a4014e8a64b8fd3a8bd3c.jpg", + "image_caption": [ + "Figure 13. Ablation study on progressive training strategy. The red circle marks the incoherent local details, i.e., unsmooth hair." + ], + "image_footnote": [], + "bbox": [ + 555, + 186, + 867, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ciency in managing Compressibility and Combined reward. Generalization to wild text prompts. As shown in Fig. 12, we present the qualitative results of prompt generalization. We found that using SHORTFT to fine-tune the model on HPDv2 still enables effective handling of wild text prompts, enhancing the overall quality of the generated images.", + "bbox": [ + 511, + 325, + 905, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 424, + 663, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On timestep-aware LoRA. As shown in Table 3, timestep-aware LoRA effectively increases the capacity of the diffusion model and accelerates the convergence of training. Under the same computational cost, the time-aware LoRA achieves better performance.", + "bbox": [ + 511, + 446, + 905, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On progressive training strategy. The progressive training strategy is committed to eliminating the training-inference gap. As shown in Table 3 and Fig. 13, directly integrating the LoRA parameters obtained from training stage 1 into the pre-trained diffusion model results in incoherent local details, resulting in worse results, which can be effectively handled by progressive training strategy.", + "bbox": [ + 511, + 522, + 905, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 640, + 633, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel Shortcut-based Fine-Tuning (SHORTFT), an advanced technique for aligning diffusion models with reward functions through end-to-end backpropagation in the denoising chain. While existing methods struggle with computational costs and the risk of gradient explosion, SHORTFT leverages shorter denoising chains, markedly improving fine-tuning efficiency and effectiveness. Rigorous evaluations demonstrate that our method can be effectively applied to various reward functions, significantly enhancing alignment performance and surpassing state-of-the-art alternative solutions.", + "bbox": [ + 511, + 666, + 906, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 845, + 665, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Plan (2024YFB3309302).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "685", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3, 4, 5", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3", + "[3] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 2, 3", + "[4] Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffusion models with text-encoder reinforcement learning. In ECCV, 2024.3", + "[5] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 2", + "[6] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR, 2024. 2, 3, 4, 5, 6, 7", + "[7] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3", + "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3", + "[10] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023. 2, 3", + "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 3", + "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 2", + "[13] Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L Isbell, and Andrea L Thomaz. Policy shaping: Integrating human feedback with reinforcement learning. In NeurIPS, 2013. 2", + "[14] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh Susskind, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Navdeep Jaitly. Matryoshka diffusion models. arXiv preprint arXiv:2310.15111, 2023. 3", + "[15] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Liefeng Bo, and Di Huang. I4vgen: Image as free stepping stone for text-to-video generation. arXiv preprint arXiv:2406.02230, 2024.", + "[16] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024.", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3", + "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 5", + "[20] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In NeurIPS, 2023. 3", + "[21] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In ECCV, 2024. 3, 4", + "[22] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 2, 3", + "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 2, 3, 4", + "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023. 3, 6", + "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3", + "[26] Jiachen Li, Weixi Feng, Wenhu Chen, and William Yang Wang. Reward guided latent consistency distillation. arXiv preprint arXiv:2403.11027, 2024. 3, 8", + "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750, 2024.", + "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3, 8", + "[29] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. In NeurIPS, 2024. 3" + ], + "bbox": [ + 517, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "686", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. arXiv preprint arXiv:2406.04314, 2024.3", + "[31] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3, 4", + "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 3", + "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2", + "[34] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2, 3, 4, 5", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 3", + "[37] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024. 2, 3, 4, 5", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6", + "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 2, 3, 4", + "[40] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 4", + "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 3, 4", + "[42] Christoph Schuhmann and Romain Beaumont. Laoin aesthetic predictor. https://laion.ai/blog/laion-aesthetics/, 2022. 2, 3", + "[43] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 3", + "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 5", + "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 4", + "[47] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In NeurIPS, 2020. 2", + "[48] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3", + "[49] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 3, 6, 7", + "[50] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In ICCV, 2023. 2, 3", + "[51] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7", + "[52] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, 2023. 2, 3, 5", + "[53] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. In NeurIPS, 2023. 3", + "[54] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3", + "[55] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265, 2024. 3", + "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 3, 4", + "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[58] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In ECCV, 2024. 3", + "[59] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + ], + "bbox": [ + 516, + 90, + 903, + 842 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "687", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_model.json b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de6a0909430f6ff3f7e9447539881d71599b87e0 --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "SHORTFT: Diffusion Model Alignment via Shortcut-based Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.18, + 0.748, + 0.2 + ], + "angle": 0, + "content": "Xiefan Guo\\(^{1,2}\\) Miaomiao Cui Liefeng Bo Di Huang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.206, + 0.877, + 0.241 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing 100191, China \\(^{2}\\)School of Computer Science and Engineering, Beihang University, Beijing 100191, China" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.243, + 0.614, + 0.259 + ], + "angle": 0, + "content": "{xfquo, dhuang}@buaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.26, + 0.702, + 0.276 + ], + "angle": 0, + "content": "Project: https://xiefan-guo.github.io/shortft" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.306, + 0.904, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.629, + 0.908, + 0.7 + ], + "angle": 0, + "content": "Figure 1. Example results synthesized by SHORTFT. SHORTFT endeavors to achieve the alignment of diffusion models with reward functions by facilitating the end-to-end backpropagation of the targeted reward gradient throughout the denoising chain. Our method has exhibited remarkable efficacy, particularly evident in the realms of text-image alignment and the overall enhancement of image quality (Top). Moreover, its versatility is underscored by the successful application across diverse reward functions, substantially amplifying alignment performance (Bottom). Combined reward is a weighted combination of rewards: PickScore = 10, HPS v2 = 2, Aesthetic = 0.05." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.711, + 0.327, + 0.726 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.743, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Backpropagation-based approaches aim to align diffusion models with reward functions through end-to-end backpropagation of the reward gradient within the denoising chain, offering a promising perspective. However, due to the computational costs and the risk of gradient explosion associated with the lengthy denoising chain, existing approaches struggle to achieve complete gradient backpropagation, leading to suboptimal results. In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), an effi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.909, + 0.879 + ], + "angle": 0, + "content": "cient fine-tuning strategy that utilizes the shorter denoising chain. More specifically, we employ the recently researched trajectory-preserving few-step diffusion model, which enables a shortcut over the original denoising chain, and construct a shortcut-based denoising chain of shorter length. The optimization on this chain notably enhances the efficiency and effectiveness of fine-tuning the foundational model. Our method has been rigorously tested and can be effectively applied to various reward functions, significantly improving alignment performance and surpassing state-of-the-art alternatives." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "678" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.09, + 0.228, + 0.1 + ], + "angle": 0, + "content": "(a) Vanilla backpropagation" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.101, + 0.495, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.148, + 0.456, + 0.159 + ], + "angle": 0, + "content": "(c) AlignProp, DRTune: Truncated backpropagation within the denoising chain" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.164, + 0.495, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.128, + 0.216, + 0.245, + 0.226 + ], + "angle": 0, + "content": "Sample \\((\\theta_{\\mathrm{diffusion}},\\mathbf{c},\\mathbf{z})\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.274, + 0.217, + 0.394, + 0.226 + ], + "angle": 0, + "content": "- Vanilla backpropagation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.424, + 0.217, + 0.543, + 0.226 + ], + "angle": 0, + "content": "←- Partial backpropagation" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.09, + 0.871, + 0.101 + ], + "angle": 0, + "content": "(b) DRaFT-K, ReFL: Truncating backpropagation at the end of the denoising chain" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.101, + 0.896, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.148, + 0.743, + 0.159 + ], + "angle": 0, + "content": "(d) SHORTFT (Ours): Shortcut-based backpropagation" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.16, + 0.896, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.574, + 0.216, + 0.689, + 0.226 + ], + "angle": 0, + "content": "Sample( \\(\\theta_{\\mathrm{shortcut}}\\) ,c,z)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.718, + 0.216, + 0.872, + 0.226 + ], + "angle": 0, + "content": "←- Shortcut-based backpropagation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 2. Comparison of fine-tuning strategies. (a) The vanilla backpropagation-based fine-tuning strategy, which suffers from lengthy backpropagation chains. (b) DRaFT-K [6] and ReFL [52] truncate the backpropagation chain, focusing on the latter half of the denoising chain, they ignore the direct supervision at the early stage, resulting in suboptimal alignment with text prompts. (c) AlignProp [34] and DRTune [51] truncate part of the backpropagation within the denoising chain by disabling some gradients. Specifically, given the denoising operation: \\(\\mathbf{x}_{t-1} = \\alpha_t \\mathbf{x}_t + \\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t) + c_t \\epsilon\\), partial backpropagation only utilizes the gradients of the red part (\\(\\alpha_t \\mathbf{x}_t\\)), truncating the green part (\\(\\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t)\\)). This inevitably introduces gradient errors, leading to unstable optimization. (d) Our method, using the few-step diffusion model to construct denoising shortcuts, facilitates complete gradient backpropagation through the entire denoising chain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.222, + 0.367 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Diffusion models [1, 9, 17, 18, 22, 32, 38, 44, 45] have established themselves as a pioneering approach in generative modeling, demonstrating exceptional prowess in applications such as photo-realistic text-to-image synthesis. However, the maximum likelihood training objective of these models, which aims to model the training data distribution accurately, can often conflict with downstream goals like aesthetics, fairness, safety, and text-to-image alignment. Therefore, aligning text-to-image diffusion models with human preferences has emerged as a pivotal and practical task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Directly supervised fine-tuning on small-scale, human-curated datasets, e.g., LAION Aesthetics [42], presents a straightforward solution. However, the prohibitive cost of data collection and the rapid obsolescence of datasets, particularly in terms of resolution compatibility with the latest text-to-image models, make this approach impractical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Emulating the successful application of Reinforcement Learning from Human Feedback (RLHF) [5, 13, 33, 47, 59] in Large Language Models (LLMs), several studies [3, 10, 12, 25, 50] have experimented with Reinforcement Learning (RL) techniques to align diffusion models with a reward function. Despite promising performance enhancements in specific domains, RL-based methods are notorious for their high-variance gradients, leading to inefficiencies and limited adaptability to diverse prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Recently, backpropagation-based methods [6, 34, 51, 52] have sought to align diffusion models with reward functions using end-to-end backpropagation of the reward gradient through the denoising chain, showing potential. However, these strategies face challenges arising from the lengthy denoising chain, which demands considerable computational resources and is susceptible to gradient explosion. As illustrated in Fig. 2 (b), [6, 52] have made strides by truncating backpropagation to focus on the latter part of the de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.475 + ], + "angle": 0, + "content": "noising chain. However, they overlook direct supervision in the early stages, leading to suboptimal alignment with text prompts. As illustrated in Fig. 2 (c), [34, 51] truncate a portion of backpropagation within the denoising chain by deactivating some gradients and introduce gradient checkpointing, enabling gradient backpropagation to the early stage of the denoising chain. However, they are time-consuming and introduce gradient bias, leading to optimization instability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.537 + ], + "angle": 0, + "content": "This paper revisits the issue of excessively long denoising chain and propose an alternative approach, employing the shorter denoising chain to facilitate full gradient backpropagation throughout the entire denoising chain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.794 + ], + "angle": 0, + "content": "In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), which leverages the few-step diffusion model to construct a denoising shortcut to fine-tune the foundational model (e.g., SD 1.5), inspired by recent trajectory-preserving diffusion distillation methods [23, 37, 39, 46]. This technique enables us to bypass the original lengthy denoising chain and complete the inference, creating a shortcut-based denoising chain of shorter length. In addition, we construct a timestep-aware LoRA as an expert LoRA ensemble, based on the intriguing temporal dynamics exhibited by the text-to-image diffusion model during the denoising process. This approach increases the number of trainable parameters without increasing cost in the inference phase, enabling faster convergence and improved performance. We also devise a custom progressive training strategy to mitigate training inference bias introduced by using denoising shortcuts during the training stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Extensive quantitative and qualitative analyses demonstrate that SHORTFT can be effectively applied to various reward functions and architectures, significantly enhancing alignment performance. Furthermore, SHORTFT, benefiting from the short denoising chain and without the need for gradient checkpointing, is particularly efficient, learning faster than DRTune [51], the current most efficient method." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.198, + 0.907, + 0.309 + ], + "angle": 0, + "content": "Figure 3. Denoising shortcut. The trajectory-preserving few-step diffusion model naturally introduces denoising shortcut, allowing for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. The 4-step Hyper-SD distilled from SD 1.5 is used in our experiments. SHORTcut(i) denotes completing the denoising process from timestep \\(i\\) to 0 using the few-step diffusion model. In addition, we also provide the well-known one-step denoising results \\(\\mathrm{DDIM}(i)\\), where \\(\\mathrm{DDIM}(i)\\) denotes performing a one-step denoising operation from timestep \\(i\\) to 0. Intuitively, the output of \\(\\mathrm{DDIM}(i)\\) is more blurred, lacking accurate structure and texture details, while the output of SHORTcut(i) is closer to the original output of SD 1.5. Furthermore, we provide their corresponding HPS v2 scores, with the absolute value of the deviation from the score corresponding to SD 1.5 provided in parentheses, SHORTcut(i) exhibits smaller deviations. These observations collectively indicate the reliability and validity of the denoising shortcut." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.324, + 0.232, + 0.339 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.348, + 0.358, + 0.364 + ], + "angle": 0, + "content": "2.1. Alignment of diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.477 + ], + "angle": 0, + "content": "Diffusion models [9, 18, 44, 45] have become a dominant force in generative modeling, showing exceptional performance in diverse applications [1, 14-17, 22, 32, 38, 53, 57]. However, certain misalignments with human intentions can arise. Recent research, fueled by the successful alignment of large language models, has sought to align diffusion models with human expectations and preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.568 + ], + "angle": 0, + "content": "Fine-tuning via data augmentation. Several studies [7, 10, 20, 25, 50] have explored altering the training data distribution for fine-tuning diffusion models on visually compelling and textually coherent data, which has led to improved results. Other methods [2, 43] involve re-captioning pre-collected web images to enhance textual precision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Fine-tuning via reward models. Reward models [24, 25, 42, 49, 50, 52] are employed to emulate human preferences given an input prompt and generated images. Several approaches have attempted to integrate these signals to augment text-to-image generation. A significant direction is the utilization of reinforcement learning-based algorithms [3, 4, 8, 11, 58] for fine-tuning text-to-image diffusion models in alignment with these rewards, [29, 30, 36, 48, 54, 55] bypass it entirely with Direct Preference Optimization. However, these methods are costly and have high gradient variance, leading to inefficiency and limited adaptability to diverse prompts. Consequently, backpropagation-based techniques [6, 34, 51, 52] have been explored, which directly fine-tune diffusion models using differentiable rewards [24, 42, 49, 50, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The challenge of backpropagation-based strategies stem from the lengthy denoising chain, which often requires numerous denoising operations (e.g., 50 for DDIM), corresponding to a long backpropagation chain. This process incurs substantial time and memory costs and is prone to gradient explosion. To mitigate this issue, [6, 52] truncate backpropagation by concentrating on the latter part" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.476 + ], + "angle": 0, + "content": "of the denoising chain. While these approaches yield some improvements, they neglect direct supervision in the early stage of the denoising chain, leading to less precise alignment with text prompts. [34, 51] truncate part of the backpropagation within the denoising chain by deactivating some gradients. By employing gradient checkpointing, they allows for the propagation of gradients to the early stages of the denoising chain. Despite their merits, these techniques can be computationally demanding and induce gradient bias, resulting in optimization instability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Different from previous methods, this paper revisits the fundamental challenge of excessively long denoising chain and proposes an alternative approach: leveraging the shorter denoising chain to facilitate full gradient backpropagation throughout the denoising chain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.707, + 0.575 + ], + "angle": 0, + "content": "2.2. Diffusion distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Existing methods for the distillation of diffusion models can be primarily classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]. The former aims to preserve the original denoising trajectory dictated by an ordinary differential equation (ODE), while the latter focuses on leveraging the denoising endpoint as the main supervision, disregarding the intermediate trajectory steps. This paper focuses on trajectory-preserving distillation, supporting the establishment of a denoising shortcut within the complete denoising chain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.838, + 0.774 + ], + "angle": 0, + "content": "2.3. Fine-tuning few-step diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Existing works [26-28] have successfully explored finetuning few-step diffusion models. While they share similarities with our research in utilizing few-step diffusion models, a crucial difference is that we focus on fine-tuning the foundational models. In contrast to the foundational models, few-step diffusion models suffer from performance degradation and reduced capacity caused by the distillation process, leading fine-tuning them suboptimal (see Sec. 4.4)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.9, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.194, + 0.907, + 0.237 + ], + "angle": 0, + "content": "Figure 4. Illustration of SHORTFT. The core of the proposed method is the Shortcut-based Fine-Tuning (SHORTFT), which leverages the trajectory-preserving few-step diffusion model as the shortcut (identified as blue arrow) to achieve direct end-to-end backpropagation through the diffusion sampling process, fine-tuning the parameters of the pre-trained diffusion model to align it with the reward function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.247, + 0.204, + 0.263 + ], + "angle": 0, + "content": "3. SHORTFT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.364 + ], + "angle": 0, + "content": "Our method, Shortcut-based Fine-Tuning (SHORTFT), capitalizes on the trajectory-preserving few-step diffusion model as a shortcut to achieve direct end-to-end backpropagation through the diffusion sampling process. This approach fine-tunes the parameters of the pre-trained diffusion model to align it with the reward function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.291, + 0.39 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.46 + ], + "angle": 0, + "content": "In line with [6, 34, 51], SHORTFT focuses on fine-tuning the parameters \\(\\theta\\) of pre-trained diffusion models to maximize the differentiable reward function \\(\\mathcal{R}(\\cdot)\\) for generated images. This can be formally represented as in Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.472, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\nJ (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {T} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} [ \\mathcal {R} (\\text {S a m p l e} (\\boldsymbol {\\theta}, \\mathbf {c}, \\mathbf {x} _ {T}), \\mathbf {c}) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.533 + ], + "angle": 0, + "content": "where \\( \\mathrm{Sample}(\\theta, \\mathbf{c}, \\mathbf{x}_T) \\) represents the denoising process for the timestep \\( t = T \\to 0 \\) with prompt condition \\( \\mathbf{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Consistent with [6, 34, 51], Eq. 1 is resolved by calculating \\(\\nabla \\mathcal{R}\\) (Sample \\((\\theta ,\\mathbf{c},\\mathbf{x}_T)\\), c) and employing gradient ascent. The computation of this gradient necessitates backpropagation through multiple diffusion models in the denoising chain, akin to backpropagation through time in recurrent neural networks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.273, + 0.652 + ], + "angle": 0, + "content": "3.2. Denoising shortcut" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.764 + ], + "angle": 0, + "content": "The recent emergence of a series of diffusion-aware distillation algorithms [23, 37, 39, 46] has been instrumental in mitigating the computational burden associated with the multi-step inference process of diffusion models. These algorithms can be roughly classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Among these, trajectory-preserving few-step diffusion models naturally introduce a denoising shortcut, which allows for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. As demonstrated in Fig. 3, the integration of these shortcuts into the denoising chain significantly reduces the total number of denoising steps, thereby shortening the length of the denoising chain. This key insight paves the way for efficient and effective end-to-end backpropagation." + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.245, + 0.623, + 0.256 + ], + "angle": 0, + "content": "(a) SD denoising chain" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.257, + 0.895, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.304, + 0.678, + 0.314 + ], + "angle": 0, + "content": "(b) Shortcut-based denoising chain" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.315, + 0.895, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Figure 5. Illustration of timestep-aware LoRA. (a) Vanilla SD denoising chain; (b) Shortcut-based denoising chain. In accordance with the interesting time dynamics in the text-to-image diffusion model denoising process revealed by [1], different from existing methods that share the same LoRA parameters at all timesteps, we introduce time-step aware LoRA, which effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing the computational cost during the inference stage." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.523, + 0.756, + 0.539 + ], + "angle": 0, + "content": "3.3. Shortcut-based fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Insight. The naive optimization of Eq. 1 through backpropagation necessitates the construction of the complete denoising chain: \\(\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\dots ,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\dots ,\\) \\(\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}\\). This process involves the storage of intermediate activations linked to each neural layer and each denoising timestep within GPU VRAM, which is not feasible due to memory constraints. Furthermore, the typical length of a denoising chain is approximately 50, which results in an overly long backpropagation chain that can lead to issues of gradient explosion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.822 + ], + "angle": 0, + "content": "The key insight of SHORTFT is rooted in the utilization of the trajectory-preserving few-step diffusion model to construct the denoising shortcut, significantly reducing the length of the denoising chain. As elaborated in Sec. 3.2, this denoising shortcut enables us to bypass a substantial number of denoising timesteps, leading to create a more streamlined and efficient denoising chain: \\(\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.906, + 0.884 + ], + "angle": 0, + "content": "As depicted in Fig. 4, such a design allows for the direct implementation of reward supervision at the early stages of the denoising chain, and facilitates full gradient backpropagation throughout the denoising chain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Shortcut-based denoising chain. As illustrated in Fig. 5," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "681" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.475, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Figure 6. Illustration of progressive training strategy. Corresponding to the time-step aware LoRA design, we develop the progressive training strategy, which eliminates the training-inference gap introduced by the shortcut-based fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.472 + ], + "angle": 0, + "content": "a vanilla SD denoising chain comprises a sequence of step-by-step denoising operations that transform the input noise \\(\\mathbf{z}_T\\) into the output image \\(\\mathbf{z}_0\\). This process typically necessitates numerous denoising steps. By harnessing the denoising shortcut, we are able to construct a shortcut-based denoising chain and fine-tune the diffusion model through end-to-end backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Timestep-aware LoRA. Instead of fine-tuning the full weights of the original diffusion model, Low-Rank Adaptation (LoRA) [19] preserves the weights of the pre-trained model and introduces new low-rank weight matrices alongside the original model weights. The contributions of these matrices are summed to generate the final outputs. Specifically, each linear layer of the UNet of SD is modified from \\(\\mathbf{h} = \\mathbf{W}\\mathbf{x}\\) to \\(\\mathbf{h} = \\mathbf{W}\\mathbf{x} + \\mathbf{B}\\mathbf{A}\\mathbf{x}\\), where \\(\\mathbf{W} \\in \\mathbb{R}^{d \\times d}\\), \\(\\mathbf{B} \\in \\mathbb{R}^{d \\times k}\\), \\(\\mathbf{A} \\in \\mathbb{R}^{k \\times d}\\), and \\(k \\ll d\\). LoRA considerably reduces the number of parameters to be optimized, thereby decreasing the memory requirements for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Moreover, [1] uncovers intriguing temporal dynamics during the denoising process of the text-to-image diffusion model. In the initial sampling stage, the model largely depends on the text prompts to guide the sampling process. As the generation progresses, the model gradually leans on visual features to denoise the image. This indicates that sharing LoRA parameters (standard practice in the existing methods) throughout the entire denoising process may not be optimal and may fail to capture the distinct patterns that emerge during denoising. Therefore, in contrast to [6, 34, 51, 52] which share the same LoRA parameters for all timesteps, we introduce timestep-aware LoRA. This design effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing computational cost in the inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Specifically, as depicted in Fig. 5, we initially divide the entire denoising chain into \\(k\\) segments, \\(\\Delta t = \\left\\lfloor \\frac{T}{k}\\right\\rfloor\\). Ex-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "cept for the first segment, we assign a corresponding LoRA to the last timestep of each subsequent segment. For the first segment, we adopt the methodology of [6] and share the same LoRA parameters across all timesteps. Notably, a sequence of continuous timesteps lacking LoRA exists in later segments, supporting the denoising shortcuts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Progressive training strategy. Although the few-step diffusion model is capable of creating the denoising shortcut, bypassing the entire denoising chain, it inherently introduces errors in the output, meaning it still cannot be fully consistent with the output of SD, particularly in fine details, as shown in Fig. 3. This incongruity can lead to a training-inference gap, resulting in suboptimal results. To mitigate this, we design a progressive training strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.456 + ], + "angle": 0, + "content": "As illustrated in Fig. 6, in accordance with the design of timestep-aware LoRA, we divide the SHORTFT training process into \\(k\\) stages. For the \\(i\\)-th training stage, we optimize the weights of LoRA \\(i\\) to LoRA \\(k\\). For the \\(i\\)-th segment and preceding denoising processes, we retain the original denoising chain, while for the denoising processes post the \\(i\\)-th segment, we introduce the denoising shortcut, thereby shortening the depth of the backpropagation chain. Moreover, in line with [6], we also employ the truncated backpropagation technique at the \\(i\\)-th segment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.906, + 0.503 + ], + "angle": 0, + "content": "During inference, as displayed in Fig. 6, the denoising shortcut is bypassed, and the original denoising chain is used to generate the final output image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.646, + 0.535 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.719, + 0.56 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.627 + ], + "angle": 0, + "content": "In our experiments, Stable Diffusion 1.5 serves as the foundational diffusion model. The DDIM schedule [44] is employed to execute 50 steps of denoising, with a classifier-free guidance scale of 7.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Shortcut. SHORTFT shortens the denoising chain by bypassing certain steps within the chain. Hence, the selection of few-step diffusion models is focused predominantly on methods that employ trajectory-preserving distillation algorithms. To accommodate the proposed time-aware LoRA, one-step diffusion models are avoided, whose output image quality is also relatively inferior. Specifically, 4-step Hyper-SD [37], distilled from SD 1.5, is utilized to construct the denoising shortcut. The value of \\( k \\) is set to 4, and the timesteps configured for LoRA are {761, 501, 261, 1}. Consequently, the denoising shortcuts are executed separately between timesteps 741 to 501, timesteps 481 to 261, and timesteps 241 to 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Timestep-aware LoRA. As suggested by [6], LoRA is applied to both the feedforward and attention layers in the UNet. The LoRA rank is set to 128. Furthermore, we adopt a stepwise stacking approach, where for LoRA \\( i \\), we introduce a new LoRA branch on top of LoRA \\( i - 1 \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.875, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.908, + 0.497 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on PickScore and HPS v2. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Reward functions. The proposed method is evaluated using three reward functions: Human Preference Score v2 (HPS v2) [49], PickScore [24], and Symmetry [51]. HPS v2 and PickScore capture human preference for images based on input prompts, while Symmetry encourages images to have horizontal symmetry features. Different from [51], which utilizes CLIPScore [35] as a regularization term, our experiment amalgamates HPS v2 and PickScore in a ratio of 1:10 to function as a joint regularization term. This approach has demonstrated superior performance in terms of text-image alignment and overall image quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Experimental details. All experiments are conducted using 2 A800 GPUs and the AdamW optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and a weight decay of 0.1. SHORTFT is performed with a batch size of 128 and a constant learning rate of \\(5 \\times 10^{-5}\\). During training, the pre-trained SD parameters are converted to bfloat16 to reduce memory usage, while the LoRA parameters under training remain in float32. Gradient checkpointing is not required." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Datasets. We compare SHORTFT-finetuned diffusion models to those of the state-of-the-art counterparts on the Human Preference Score v2 dataset (HPDv2) [49]. The final reward is computed on the 400 prompts from the test split. Following [51], for fair comparison, we evaluate all methods using the same computational budget. Specifically, all methods are trained for six hours on 2 A800 GPUs." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.506, + 0.887, + 0.597 + ], + "angle": 0, + "content": "
MethodHPS v2↑PickScore↑Symmetry↓
SD 1.5 [38]26.9120.460.853
DRaFT-LV [6]33.1323.350.418
DRTune [51]32.7923.220.207
SHORTFT33.8824.160.138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.602, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Table 1. Objective evaluation. Our method performs over other counterparts, under the same computational cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.73, + 0.667 + ], + "angle": 0, + "content": "4.2. Qualitative comparison" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Fig. 7 and 8 present the quantitative comparison of our results against those of representative methods, including current state-of-the-art techniques, under same text prompts and random seeds. SD 1.5 suffers from low-quality image generation. DRaFT [6], while proficient at managing local image details, struggles to effectively handle global layouts and the optimization of the symmetry reward function. Similarly, DRTune [51] is constrained by gradient inaccuracies, which contribute to instability during training and a deficiency in managing complex semantics. One particular area of weakness is the synthesis of images with specific counts of objects, such as accurately depicting a given number of cows. In contrast, SHORTFT exhibits enhanced capabilities in generating images that are both visually realistic and semantically faithful, outperforming over other counterparts." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.09, + 0.875, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.908, + 0.497 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison on HPS v2 and Symmetry. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.509, + 0.462, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.616 + ], + "angle": 0, + "content": "Figure 9. Human and AI preference evaluation against current methods. SHORTFT performs over other counterparts." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.621, + 0.319, + 0.637 + ], + "angle": 0, + "content": "4.3.Quantitative comparison" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Objective evaluation. Table 1 shows the quantitative results achieved on the Human Preference Score v2 benchmark [49], where the proposed method outperforms the other approaches, clearly demonstrating its effectiveness. It is important to highlight that both our method and DRTune [51] strategically employ backpropagation of the reward gradient to the initial stages of the denoising chain. This intentional design choice significantly enhances the Symmetry score performance when compared to alternative methods. Despite this, DRTune continues to grapple with the challenge of gradient bias. In contrast, our method significantly mitigates this issue, delivering superior performance. User study. A subjective user study comprising 11 volunteers is conducted, with five possessing expertise in image processing and the remaining participants having no background in computer vision or graph. Participants are tasked to select the most visually attractive and semantically ac" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.633 + ], + "angle": 0, + "content": "curate image among those generated by our method and current state-of-the-art techniques. Each participant has 10 questions for each pair of comparisons. Furthermore, an MLLM-assisted evaluation is employed using GPT-4V. We make 20 queries to GPT-4V for each pair of comparisons. More details are provided in the Appendix. As depicted in Fig. 9, the results exhibit a significant inclination towards SHORTFT in comparison to other techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.643, + 0.649, + 0.658 + ], + "angle": 0, + "content": "4.4. More results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.801 + ], + "angle": 0, + "content": "SHORTFT, 10k training step. Due to the shorter denoising and backpropagation chains, our method achieves superior performance under the same computational budget. Furthermore, to validate the upper bound of our approach, following the protocol in [6], we conduct the full training process using HPS v2 reward function on HPDv2, comprising 10k training steps, and evaluate it on the corresponding benchmark. The obtained HPS v2 score of 35.97 surpasses the reported score for DRaFT-LV in [6]." + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.814, + 0.855, + 0.857 + ], + "angle": 0, + "content": "
MethodTuning Hyper-SDTuning SD 1.5
HPS v2↑32.9235.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.861, + 0.9, + 0.877 + ], + "angle": 0, + "content": "Table 2. Objective evaluation on tuning SD 1.5 and Hyper-SD." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Fine-tuning SD vs. Hyper-SD. As mentioned in Sec. 2.3," + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.197 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison on tuning SD 1.5 (SHORTFT) and Hyper-SD. Fine-tuning SD 1.5 significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.205, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[26-28] explore fine-tuning few-step diffusion models and have achieved certain successes. However, compared to the foundational model, few-step diffusion models face performance degradation caused by the distillation process. Finetuning the few-step diffusion model is actually suboptimal compared to fine-tuning the foundational model. Furthermore, following two strategies, we separately conduct the training processes on the HPDv2 using the HPS v2 reward function. As shown in Table 2 and Fig. 10, fine-tuning SD 1.5 (SHORTFT) significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.383, + 0.482, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.498, + 0.478, + 0.512 + ], + "angle": 0, + "content": "Figure 11. Example results synthesized by SHORTFT on SD 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.596 + ], + "angle": 0, + "content": "Fine-tuning SD 3. SHORTFT is an architecture-agnostic fine-tuning strategy, applicable to both UNet-based (SD 1.5) and Transformer-based (SD 3) architectures. As illustrated in Fig. 11, SHORTFT is also capable of mastering SD 3, where SD 3 aligns with HPS v2." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.608, + 0.48, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Figure 12. Generalization to wild text prompts from Sora. Our method is capable of effectively handling the wild text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Other reward functions. SHORTFT exhibits remarkable versatility, demonstrating efficacy across a spectrum of reward functions, significantly improving the alignment performance and the quality and fidelity of the output. As shown in Fig. 1 and 12, SHORTFT not only accommodates HPS v2, PickScore, and Symmetry, but also exhibits profi" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.089, + 0.878, + 0.164 + ], + "angle": 0, + "content": "
MethodHPS v2↑PickScore↑Symmetry↓
w/o T-LoRA33.4623.820.187
w/o P-Training33.2723.970.146
SHORTFT33.8824.160.138
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.168, + 0.862, + 0.182 + ], + "angle": 0, + "content": "Table 3. Ablation study on timestep-aware LoRA." + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.187, + 0.868, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.906, + 0.316 + ], + "angle": 0, + "content": "Figure 13. Ablation study on progressive training strategy. The red circle marks the incoherent local details, i.e., unsmooth hair." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.906, + 0.418 + ], + "angle": 0, + "content": "ciency in managing Compressibility and Combined reward. Generalization to wild text prompts. As shown in Fig. 12, we present the qualitative results of prompt generalization. We found that using SHORTFT to fine-tune the model on HPDv2 still enables effective handling of wild text prompts, enhancing the overall quality of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.665, + 0.442 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.522 + ], + "angle": 0, + "content": "On timestep-aware LoRA. As shown in Table 3, timestep-aware LoRA effectively increases the capacity of the diffusion model and accelerates the convergence of training. Under the same computational cost, the time-aware LoRA achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.629 + ], + "angle": 0, + "content": "On progressive training strategy. The progressive training strategy is committed to eliminating the training-inference gap. As shown in Table 3 and Fig. 13, directly integrating the LoRA parameters obtained from training stage 1 into the pre-trained diffusion model results in incoherent local details, resulting in worse results, which can be effectively handled by progressive training strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.634, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.833 + ], + "angle": 0, + "content": "In this paper, we propose a novel Shortcut-based Fine-Tuning (SHORTFT), an advanced technique for aligning diffusion models with reward functions through end-to-end backpropagation in the denoising chain. While existing methods struggle with computational costs and the risk of gradient explosion, SHORTFT leverages shorter denoising chains, markedly improving fine-tuning efficiency and effectiveness. Rigorous evaluations demonstrate that our method can be effectively applied to various reward functions, significantly enhancing alignment performance and surpassing state-of-the-art alternative solutions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.666, + 0.863 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Plan (2024YFB3309302)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "685" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.315, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[4] Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffusion models with text-encoder reinforcement learning. In ECCV, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[5] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[6] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[10] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.756 + ], + "angle": 0, + "content": "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.482, + 0.884 + ], + "angle": 0, + "content": "[13] Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L Isbell, and Andrea L Thomaz. Policy shaping: Integrating human feedback with reinforcement learning. In NeurIPS, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh Susskind, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Navdeep Jaitly. Matryoshka diffusion models. arXiv preprint arXiv:2310.15111, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[15] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Liefeng Bo, and Di Huang. I4vgen: Image as free stepping stone for text-to-video generation. arXiv preprint arXiv:2406.02230, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.906, + 0.207 + ], + "angle": 0, + "content": "[16] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.235 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.266, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[20] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.906, + 0.434 + ], + "angle": 0, + "content": "[21] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In ECCV, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.436, + 0.906, + 0.476 + ], + "angle": 0, + "content": "[22] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.906, + 0.547 + ], + "angle": 0, + "content": "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.606, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[26] Jiachen Li, Weixi Feng, Wenhu Chen, and William Yang Wang. Reward guided latent consistency distillation. arXiv preprint arXiv:2403.11027, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[29] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. In NeurIPS, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[30] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. arXiv preprint arXiv:2406.04314, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[31] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[34] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[37] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[40] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann and Romain Beaumont. Laoin aesthetic predictor. https://laion.ai/blog/laion-aesthetics/, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[43] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[47] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[48] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[49] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[50] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[51] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[52] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[53] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[54] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[55] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[58] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[59] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "687" + } + ] +] \ No newline at end of file diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_origin.pdf b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..23fa07a26f1c77e924a34ecbdec18eae85de2e5b --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/447f852c-3aba-4c21-a78d-15886a337f7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce572c7c5c0f78313f2e835169279d3790fe898b0803bab12d290898ff264288 +size 7082627 diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/full.md b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6f46abdba95824509ef46118b3c8fc778d79730d --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/full.md @@ -0,0 +1,295 @@ +# SHORTFT: Diffusion Model Alignment via Shortcut-based Fine-Tuning + +Xiefan Guo $^{1,2}$ Miaomiao Cui Liefeng Bo Di Huang $^{1,2*}$ + +$^{1}$ State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing 100191, China $^{2}$ School of Computer Science and Engineering, Beihang University, Beijing 100191, China + +{xfquo, dhuang}@buaa.edu.cn + +Project: https://xiefan-guo.github.io/shortft + +![](images/eb45afea23bbd1f306fb7e363ff0e5c51de2ac14cd178b9320bcc7344838c899.jpg) +Figure 1. Example results synthesized by SHORTFT. SHORTFT endeavors to achieve the alignment of diffusion models with reward functions by facilitating the end-to-end backpropagation of the targeted reward gradient throughout the denoising chain. Our method has exhibited remarkable efficacy, particularly evident in the realms of text-image alignment and the overall enhancement of image quality (Top). Moreover, its versatility is underscored by the successful application across diverse reward functions, substantially amplifying alignment performance (Bottom). Combined reward is a weighted combination of rewards: PickScore = 10, HPS v2 = 2, Aesthetic = 0.05. + +# Abstract + +Backpropagation-based approaches aim to align diffusion models with reward functions through end-to-end backpropagation of the reward gradient within the denoising chain, offering a promising perspective. However, due to the computational costs and the risk of gradient explosion associated with the lengthy denoising chain, existing approaches struggle to achieve complete gradient backpropagation, leading to suboptimal results. In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), an effi + +cient fine-tuning strategy that utilizes the shorter denoising chain. More specifically, we employ the recently researched trajectory-preserving few-step diffusion model, which enables a shortcut over the original denoising chain, and construct a shortcut-based denoising chain of shorter length. The optimization on this chain notably enhances the efficiency and effectiveness of fine-tuning the foundational model. Our method has been rigorously tested and can be effectively applied to various reward functions, significantly improving alignment performance and surpassing state-of-the-art alternatives. + +![](images/df8fd5b04a30b4043b3b9758e4f2bfb39986e993be1df3cd6902731f41652753.jpg) +(a) Vanilla backpropagation +(c) AlignProp, DRTune: Truncated backpropagation within the denoising chain + +![](images/0d1e554b1ad63cd4487ec725e8f4a088c6379518f9ff17683997e197df3b58bc.jpg) +Sample $(\theta_{\mathrm{diffusion}},\mathbf{c},\mathbf{z})$ +- Vanilla backpropagation + +![](images/4a6c79e18b1d266b7ad2c0c2fd328436f4ebea9138349db98ca970384c88a391.jpg) +(b) DRaFT-K, ReFL: Truncating backpropagation at the end of the denoising chain + +(d) SHORTFT (Ours): Shortcut-based backpropagation +Figure 2. Comparison of fine-tuning strategies. (a) The vanilla backpropagation-based fine-tuning strategy, which suffers from lengthy backpropagation chains. (b) DRaFT-K [6] and ReFL [52] truncate the backpropagation chain, focusing on the latter half of the denoising chain, they ignore the direct supervision at the early stage, resulting in suboptimal alignment with text prompts. (c) AlignProp [34] and DRTune [51] truncate part of the backpropagation within the denoising chain by disabling some gradients. Specifically, given the denoising operation: $\mathbf{x}_{t-1} = \alpha_t \mathbf{x}_t + \beta_t \epsilon_\theta(\mathbf{x}_t, t) + c_t \epsilon$ , partial backpropagation only utilizes the gradients of the red part ( $\alpha_t \mathbf{x}_t$ ), truncating the green part ( $\beta_t \epsilon_\theta(\mathbf{x}_t, t)$ ). This inevitably introduces gradient errors, leading to unstable optimization. (d) Our method, using the few-step diffusion model to construct denoising shortcuts, facilitates complete gradient backpropagation through the entire denoising chain. +![](images/d1f5902ba3aba90b26d987238069734b662def7ef6b8a67da292b6efafd741c8.jpg) +←- Partial backpropagation +Sample( $\theta_{\mathrm{shortcut}}$ ,c,z) +←- Shortcut-based backpropagation + +# 1. Introduction + +Diffusion models [1, 9, 17, 18, 22, 32, 38, 44, 45] have established themselves as a pioneering approach in generative modeling, demonstrating exceptional prowess in applications such as photo-realistic text-to-image synthesis. However, the maximum likelihood training objective of these models, which aims to model the training data distribution accurately, can often conflict with downstream goals like aesthetics, fairness, safety, and text-to-image alignment. Therefore, aligning text-to-image diffusion models with human preferences has emerged as a pivotal and practical task. + +Directly supervised fine-tuning on small-scale, human-curated datasets, e.g., LAION Aesthetics [42], presents a straightforward solution. However, the prohibitive cost of data collection and the rapid obsolescence of datasets, particularly in terms of resolution compatibility with the latest text-to-image models, make this approach impractical. + +Emulating the successful application of Reinforcement Learning from Human Feedback (RLHF) [5, 13, 33, 47, 59] in Large Language Models (LLMs), several studies [3, 10, 12, 25, 50] have experimented with Reinforcement Learning (RL) techniques to align diffusion models with a reward function. Despite promising performance enhancements in specific domains, RL-based methods are notorious for their high-variance gradients, leading to inefficiencies and limited adaptability to diverse prompts. + +Recently, backpropagation-based methods [6, 34, 51, 52] have sought to align diffusion models with reward functions using end-to-end backpropagation of the reward gradient through the denoising chain, showing potential. However, these strategies face challenges arising from the lengthy denoising chain, which demands considerable computational resources and is susceptible to gradient explosion. As illustrated in Fig. 2 (b), [6, 52] have made strides by truncating backpropagation to focus on the latter part of the de + +noising chain. However, they overlook direct supervision in the early stages, leading to suboptimal alignment with text prompts. As illustrated in Fig. 2 (c), [34, 51] truncate a portion of backpropagation within the denoising chain by deactivating some gradients and introduce gradient checkpointing, enabling gradient backpropagation to the early stage of the denoising chain. However, they are time-consuming and introduce gradient bias, leading to optimization instability. + +This paper revisits the issue of excessively long denoising chain and propose an alternative approach, employing the shorter denoising chain to facilitate full gradient backpropagation throughout the entire denoising chain. + +In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), which leverages the few-step diffusion model to construct a denoising shortcut to fine-tune the foundational model (e.g., SD 1.5), inspired by recent trajectory-preserving diffusion distillation methods [23, 37, 39, 46]. This technique enables us to bypass the original lengthy denoising chain and complete the inference, creating a shortcut-based denoising chain of shorter length. In addition, we construct a timestep-aware LoRA as an expert LoRA ensemble, based on the intriguing temporal dynamics exhibited by the text-to-image diffusion model during the denoising process. This approach increases the number of trainable parameters without increasing cost in the inference phase, enabling faster convergence and improved performance. We also devise a custom progressive training strategy to mitigate training inference bias introduced by using denoising shortcuts during the training stage. + +Extensive quantitative and qualitative analyses demonstrate that SHORTFT can be effectively applied to various reward functions and architectures, significantly enhancing alignment performance. Furthermore, SHORTFT, benefiting from the short denoising chain and without the need for gradient checkpointing, is particularly efficient, learning faster than DRTune [51], the current most efficient method. + +![](images/e14e44833c570bbebe4aeb01441cec2603f28e7aef93b3ce283f64f7f6b1bdc2.jpg) +Figure 3. Denoising shortcut. The trajectory-preserving few-step diffusion model naturally introduces denoising shortcut, allowing for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. The 4-step Hyper-SD distilled from SD 1.5 is used in our experiments. SHORTcut(i) denotes completing the denoising process from timestep $i$ to 0 using the few-step diffusion model. In addition, we also provide the well-known one-step denoising results $\mathrm{DDIM}(i)$ , where $\mathrm{DDIM}(i)$ denotes performing a one-step denoising operation from timestep $i$ to 0. Intuitively, the output of $\mathrm{DDIM}(i)$ is more blurred, lacking accurate structure and texture details, while the output of SHORTcut(i) is closer to the original output of SD 1.5. Furthermore, we provide their corresponding HPS v2 scores, with the absolute value of the deviation from the score corresponding to SD 1.5 provided in parentheses, SHORTcut(i) exhibits smaller deviations. These observations collectively indicate the reliability and validity of the denoising shortcut. + +# 2. Related Work + +# 2.1. Alignment of diffusion models + +Diffusion models [9, 18, 44, 45] have become a dominant force in generative modeling, showing exceptional performance in diverse applications [1, 14-17, 22, 32, 38, 53, 57]. However, certain misalignments with human intentions can arise. Recent research, fueled by the successful alignment of large language models, has sought to align diffusion models with human expectations and preferences. + +Fine-tuning via data augmentation. Several studies [7, 10, 20, 25, 50] have explored altering the training data distribution for fine-tuning diffusion models on visually compelling and textually coherent data, which has led to improved results. Other methods [2, 43] involve re-captioning pre-collected web images to enhance textual precision. + +Fine-tuning via reward models. Reward models [24, 25, 42, 49, 50, 52] are employed to emulate human preferences given an input prompt and generated images. Several approaches have attempted to integrate these signals to augment text-to-image generation. A significant direction is the utilization of reinforcement learning-based algorithms [3, 4, 8, 11, 58] for fine-tuning text-to-image diffusion models in alignment with these rewards, [29, 30, 36, 48, 54, 55] bypass it entirely with Direct Preference Optimization. However, these methods are costly and have high gradient variance, leading to inefficiency and limited adaptability to diverse prompts. Consequently, backpropagation-based techniques [6, 34, 51, 52] have been explored, which directly fine-tune diffusion models using differentiable rewards [24, 42, 49, 50, 52]. + +The challenge of backpropagation-based strategies stem from the lengthy denoising chain, which often requires numerous denoising operations (e.g., 50 for DDIM), corresponding to a long backpropagation chain. This process incurs substantial time and memory costs and is prone to gradient explosion. To mitigate this issue, [6, 52] truncate backpropagation by concentrating on the latter part + +of the denoising chain. While these approaches yield some improvements, they neglect direct supervision in the early stage of the denoising chain, leading to less precise alignment with text prompts. [34, 51] truncate part of the backpropagation within the denoising chain by deactivating some gradients. By employing gradient checkpointing, they allows for the propagation of gradients to the early stages of the denoising chain. Despite their merits, these techniques can be computationally demanding and induce gradient bias, resulting in optimization instability. + +Different from previous methods, this paper revisits the fundamental challenge of excessively long denoising chain and proposes an alternative approach: leveraging the shorter denoising chain to facilitate full gradient backpropagation throughout the denoising chain. + +# 2.2. Diffusion distillation + +Existing methods for the distillation of diffusion models can be primarily classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]. The former aims to preserve the original denoising trajectory dictated by an ordinary differential equation (ODE), while the latter focuses on leveraging the denoising endpoint as the main supervision, disregarding the intermediate trajectory steps. This paper focuses on trajectory-preserving distillation, supporting the establishment of a denoising shortcut within the complete denoising chain. + +# 2.3. Fine-tuning few-step diffusion models + +Existing works [26-28] have successfully explored finetuning few-step diffusion models. While they share similarities with our research in utilizing few-step diffusion models, a crucial difference is that we focus on fine-tuning the foundational models. In contrast to the foundational models, few-step diffusion models suffer from performance degradation and reduced capacity caused by the distillation process, leading fine-tuning them suboptimal (see Sec. 4.4). + +![](images/a28ef26760df40c1dfbc6a55a089b37d55d8ec57dfeb47d4d19ad7d13e8cb83b.jpg) +Figure 4. Illustration of SHORTFT. The core of the proposed method is the Shortcut-based Fine-Tuning (SHORTFT), which leverages the trajectory-preserving few-step diffusion model as the shortcut (identified as blue arrow) to achieve direct end-to-end backpropagation through the diffusion sampling process, fine-tuning the parameters of the pre-trained diffusion model to align it with the reward function. + +# 3. SHORTFT + +Our method, Shortcut-based Fine-Tuning (SHORTFT), capitalizes on the trajectory-preserving few-step diffusion model as a shortcut to achieve direct end-to-end backpropagation through the diffusion sampling process. This approach fine-tunes the parameters of the pre-trained diffusion model to align it with the reward function. + +# 3.1. Problem formulation + +In line with [6, 34, 51], SHORTFT focuses on fine-tuning the parameters $\theta$ of pre-trained diffusion models to maximize the differentiable reward function $\mathcal{R}(\cdot)$ for generated images. This can be formally represented as in Eq. 1: + +$$ +J (\boldsymbol {\theta}) = \mathbb {E} _ {\mathbf {c}, \mathbf {x} _ {T} \sim \mathcal {N} (\mathbf {0}, \mathbf {1})} [ \mathcal {R} (\text {S a m p l e} (\boldsymbol {\theta}, \mathbf {c}, \mathbf {x} _ {T}), \mathbf {c}) ], \tag {1} +$$ + +where $\mathrm{Sample}(\theta, \mathbf{c}, \mathbf{x}_T)$ represents the denoising process for the timestep $t = T \to 0$ with prompt condition $\mathbf{c}$ . + +Consistent with [6, 34, 51], Eq. 1 is resolved by calculating $\nabla \mathcal{R}$ (Sample $(\theta ,\mathbf{c},\mathbf{x}_T)$ , c) and employing gradient ascent. The computation of this gradient necessitates backpropagation through multiple diffusion models in the denoising chain, akin to backpropagation through time in recurrent neural networks. + +# 3.2. Denoising shortcut + +The recent emergence of a series of diffusion-aware distillation algorithms [23, 37, 39, 46] has been instrumental in mitigating the computational burden associated with the multi-step inference process of diffusion models. These algorithms can be roughly classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]. + +Among these, trajectory-preserving few-step diffusion models naturally introduce a denoising shortcut, which allows for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. As demonstrated in Fig. 3, the integration of these shortcuts into the denoising chain significantly reduces the total number of denoising steps, thereby shortening the length of the denoising chain. This key insight paves the way for efficient and effective end-to-end backpropagation. + +![](images/bff8c66218c5a096b0eac7fe9122ad0c44ca29d3563e12531ff9f6c7d3bee0a5.jpg) +(a) SD denoising chain + +![](images/9cc46b97ab6bdf01e9754f9c0d251fde482e8da6eb24cfe99606f103cfbf0d29.jpg) +(b) Shortcut-based denoising chain +Figure 5. Illustration of timestep-aware LoRA. (a) Vanilla SD denoising chain; (b) Shortcut-based denoising chain. In accordance with the interesting time dynamics in the text-to-image diffusion model denoising process revealed by [1], different from existing methods that share the same LoRA parameters at all timesteps, we introduce time-step aware LoRA, which effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing the computational cost during the inference stage. + +# 3.3. Shortcut-based fine-tuning + +Insight. The naive optimization of Eq. 1 through backpropagation necessitates the construction of the complete denoising chain: $\{\mathbf{x}_T,\dots ,\mathbf{x}_t,\dots ,\mathbf{x}_{t - \Delta t},\mathbf{x}_{t - \Delta t - 1},\dots ,$ $\mathbf{x}_{t - 2\Delta t - 1},\dots ,\mathbf{x}_0\}$ . This process involves the storage of intermediate activations linked to each neural layer and each denoising timestep within GPU VRAM, which is not feasible due to memory constraints. Furthermore, the typical length of a denoising chain is approximately 50, which results in an overly long backpropagation chain that can lead to issues of gradient explosion. + +The key insight of SHORTFT is rooted in the utilization of the trajectory-preserving few-step diffusion model to construct the denoising shortcut, significantly reducing the length of the denoising chain. As elaborated in Sec. 3.2, this denoising shortcut enables us to bypass a substantial number of denoising timesteps, leading to create a more streamlined and efficient denoising chain: $\{\mathbf{x}_T,\dots ,\mathbf{x}_t,\mathbf{x}_{t - \Delta t},\mathbf{x}_{t - \Delta t - 1},\mathbf{x}_{t - 2\Delta t - 1},\dots ,\mathbf{x}_0\}$ . + +As depicted in Fig. 4, such a design allows for the direct implementation of reward supervision at the early stages of the denoising chain, and facilitates full gradient backpropagation throughout the denoising chain. + +Shortcut-based denoising chain. As illustrated in Fig. 5, + +![](images/f7a8a5ada6bc548c7d540ab54d0d17c5f164f2cdddf68b1f9207b9139d6bf506.jpg) +Figure 6. Illustration of progressive training strategy. Corresponding to the time-step aware LoRA design, we develop the progressive training strategy, which eliminates the training-inference gap introduced by the shortcut-based fine-tuning. + +a vanilla SD denoising chain comprises a sequence of step-by-step denoising operations that transform the input noise $\mathbf{z}_T$ into the output image $\mathbf{z}_0$ . This process typically necessitates numerous denoising steps. By harnessing the denoising shortcut, we are able to construct a shortcut-based denoising chain and fine-tune the diffusion model through end-to-end backpropagation. + +Timestep-aware LoRA. Instead of fine-tuning the full weights of the original diffusion model, Low-Rank Adaptation (LoRA) [19] preserves the weights of the pre-trained model and introduces new low-rank weight matrices alongside the original model weights. The contributions of these matrices are summed to generate the final outputs. Specifically, each linear layer of the UNet of SD is modified from $\mathbf{h} = \mathbf{W}\mathbf{x}$ to $\mathbf{h} = \mathbf{W}\mathbf{x} + \mathbf{B}\mathbf{A}\mathbf{x}$ , where $\mathbf{W} \in \mathbb{R}^{d \times d}$ , $\mathbf{B} \in \mathbb{R}^{d \times k}$ , $\mathbf{A} \in \mathbb{R}^{k \times d}$ , and $k \ll d$ . LoRA considerably reduces the number of parameters to be optimized, thereby decreasing the memory requirements for fine-tuning. + +Moreover, [1] uncovers intriguing temporal dynamics during the denoising process of the text-to-image diffusion model. In the initial sampling stage, the model largely depends on the text prompts to guide the sampling process. As the generation progresses, the model gradually leans on visual features to denoise the image. This indicates that sharing LoRA parameters (standard practice in the existing methods) throughout the entire denoising process may not be optimal and may fail to capture the distinct patterns that emerge during denoising. Therefore, in contrast to [6, 34, 51, 52] which share the same LoRA parameters for all timesteps, we introduce timestep-aware LoRA. This design effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing computational cost in the inference phase. + +Specifically, as depicted in Fig. 5, we initially divide the entire denoising chain into $k$ segments, $\Delta t = \left\lfloor \frac{T}{k}\right\rfloor$ . Ex- + +cept for the first segment, we assign a corresponding LoRA to the last timestep of each subsequent segment. For the first segment, we adopt the methodology of [6] and share the same LoRA parameters across all timesteps. Notably, a sequence of continuous timesteps lacking LoRA exists in later segments, supporting the denoising shortcuts. + +Progressive training strategy. Although the few-step diffusion model is capable of creating the denoising shortcut, bypassing the entire denoising chain, it inherently introduces errors in the output, meaning it still cannot be fully consistent with the output of SD, particularly in fine details, as shown in Fig. 3. This incongruity can lead to a training-inference gap, resulting in suboptimal results. To mitigate this, we design a progressive training strategy. + +As illustrated in Fig. 6, in accordance with the design of timestep-aware LoRA, we divide the SHORTFT training process into $k$ stages. For the $i$ -th training stage, we optimize the weights of LoRA $i$ to LoRA $k$ . For the $i$ -th segment and preceding denoising processes, we retain the original denoising chain, while for the denoising processes post the $i$ -th segment, we introduce the denoising shortcut, thereby shortening the depth of the backpropagation chain. Moreover, in line with [6], we also employ the truncated backpropagation technique at the $i$ -th segment. + +During inference, as displayed in Fig. 6, the denoising shortcut is bypassed, and the original denoising chain is used to generate the final output image. + +# 4. Experiments + +# 4.1. Experimental settings + +In our experiments, Stable Diffusion 1.5 serves as the foundational diffusion model. The DDIM schedule [44] is employed to execute 50 steps of denoising, with a classifier-free guidance scale of 7.5. + +Shortcut. SHORTFT shortens the denoising chain by bypassing certain steps within the chain. Hence, the selection of few-step diffusion models is focused predominantly on methods that employ trajectory-preserving distillation algorithms. To accommodate the proposed time-aware LoRA, one-step diffusion models are avoided, whose output image quality is also relatively inferior. Specifically, 4-step Hyper-SD [37], distilled from SD 1.5, is utilized to construct the denoising shortcut. The value of $k$ is set to 4, and the timesteps configured for LoRA are {761, 501, 261, 1}. Consequently, the denoising shortcuts are executed separately between timesteps 741 to 501, timesteps 481 to 261, and timesteps 241 to 1. + +Timestep-aware LoRA. As suggested by [6], LoRA is applied to both the feedforward and attention layers in the UNet. The LoRA rank is set to 128. Furthermore, we adopt a stepwise stacking approach, where for LoRA $i$ , we introduce a new LoRA branch on top of LoRA $i - 1$ . + +![](images/3b04d6475d0b1c018e64c0032ddea6c5039d1fee22ac6d01f524b4a78ae778b9.jpg) +Figure 7. Qualitative comparison on PickScore and HPS v2. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality. + +Reward functions. The proposed method is evaluated using three reward functions: Human Preference Score v2 (HPS v2) [49], PickScore [24], and Symmetry [51]. HPS v2 and PickScore capture human preference for images based on input prompts, while Symmetry encourages images to have horizontal symmetry features. Different from [51], which utilizes CLIPScore [35] as a regularization term, our experiment amalgamates HPS v2 and PickScore in a ratio of 1:10 to function as a joint regularization term. This approach has demonstrated superior performance in terms of text-image alignment and overall image quality. + +Experimental details. All experiments are conducted using 2 A800 GPUs and the AdamW optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and a weight decay of 0.1. SHORTFT is performed with a batch size of 128 and a constant learning rate of $5 \times 10^{-5}$ . During training, the pre-trained SD parameters are converted to bfloat16 to reduce memory usage, while the LoRA parameters under training remain in float32. Gradient checkpointing is not required. + +Datasets. We compare SHORTFT-finetuned diffusion models to those of the state-of-the-art counterparts on the Human Preference Score v2 dataset (HPDv2) [49]. The final reward is computed on the 400 prompts from the test split. Following [51], for fair comparison, we evaluate all methods using the same computational budget. Specifically, all methods are trained for six hours on 2 A800 GPUs. + +
MethodHPS v2↑PickScore↑Symmetry↓
SD 1.5 [38]26.9120.460.853
DRaFT-LV [6]33.1323.350.418
DRTune [51]32.7923.220.207
SHORTFT33.8824.160.138
+ +Table 1. Objective evaluation. Our method performs over other counterparts, under the same computational cost. + +# 4.2. Qualitative comparison + +Fig. 7 and 8 present the quantitative comparison of our results against those of representative methods, including current state-of-the-art techniques, under same text prompts and random seeds. SD 1.5 suffers from low-quality image generation. DRaFT [6], while proficient at managing local image details, struggles to effectively handle global layouts and the optimization of the symmetry reward function. Similarly, DRTune [51] is constrained by gradient inaccuracies, which contribute to instability during training and a deficiency in managing complex semantics. One particular area of weakness is the synthesis of images with specific counts of objects, such as accurately depicting a given number of cows. In contrast, SHORTFT exhibits enhanced capabilities in generating images that are both visually realistic and semantically faithful, outperforming over other counterparts. + +![](images/066169bad1c7248978d8e5c9b23bd445a964eb7a40b25ed4795988fdb52eaa76.jpg) +Figure 8. Qualitative comparison on HPS v2 and Symmetry. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality. + +![](images/401a6f5f6634bf4694c4774099fe5225bb898ee0bf0213aec23bfc81132ecd67.jpg) +Figure 9. Human and AI preference evaluation against current methods. SHORTFT performs over other counterparts. + +# 4.3.Quantitative comparison + +Objective evaluation. Table 1 shows the quantitative results achieved on the Human Preference Score v2 benchmark [49], where the proposed method outperforms the other approaches, clearly demonstrating its effectiveness. It is important to highlight that both our method and DRTune [51] strategically employ backpropagation of the reward gradient to the initial stages of the denoising chain. This intentional design choice significantly enhances the Symmetry score performance when compared to alternative methods. Despite this, DRTune continues to grapple with the challenge of gradient bias. In contrast, our method significantly mitigates this issue, delivering superior performance. User study. A subjective user study comprising 11 volunteers is conducted, with five possessing expertise in image processing and the remaining participants having no background in computer vision or graph. Participants are tasked to select the most visually attractive and semantically ac + +curate image among those generated by our method and current state-of-the-art techniques. Each participant has 10 questions for each pair of comparisons. Furthermore, an MLLM-assisted evaluation is employed using GPT-4V. We make 20 queries to GPT-4V for each pair of comparisons. More details are provided in the Appendix. As depicted in Fig. 9, the results exhibit a significant inclination towards SHORTFT in comparison to other techniques. + +# 4.4. More results + +SHORTFT, 10k training step. Due to the shorter denoising and backpropagation chains, our method achieves superior performance under the same computational budget. Furthermore, to validate the upper bound of our approach, following the protocol in [6], we conduct the full training process using HPS v2 reward function on HPDv2, comprising 10k training steps, and evaluate it on the corresponding benchmark. The obtained HPS v2 score of 35.97 surpasses the reported score for DRaFT-LV in [6]. + +
MethodTuning Hyper-SDTuning SD 1.5
HPS v2↑32.9235.97
+ +Table 2. Objective evaluation on tuning SD 1.5 and Hyper-SD. + +Fine-tuning SD vs. Hyper-SD. As mentioned in Sec. 2.3, + +![](images/9793504cf4bae9a8204d59ee35176efc78f931fa676aa14438c9dbedeadb2649.jpg) +Figure 10. Qualitative comparison on tuning SD 1.5 (SHORTFT) and Hyper-SD. Fine-tuning SD 1.5 significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details. + +[26-28] explore fine-tuning few-step diffusion models and have achieved certain successes. However, compared to the foundational model, few-step diffusion models face performance degradation caused by the distillation process. Finetuning the few-step diffusion model is actually suboptimal compared to fine-tuning the foundational model. Furthermore, following two strategies, we separately conduct the training processes on the HPDv2 using the HPS v2 reward function. As shown in Table 2 and Fig. 10, fine-tuning SD 1.5 (SHORTFT) significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details. + +![](images/3a3dcd691a9729e6d9aab4d09235fad431334f3739ed423ddd16735b1fb4f841.jpg) +Figure 11. Example results synthesized by SHORTFT on SD 3. + +Fine-tuning SD 3. SHORTFT is an architecture-agnostic fine-tuning strategy, applicable to both UNet-based (SD 1.5) and Transformer-based (SD 3) architectures. As illustrated in Fig. 11, SHORTFT is also capable of mastering SD 3, where SD 3 aligns with HPS v2. + +![](images/b0e476c64b57777984eb424274a223b9f13ee8d140e3d1a974cc426f7ec67399.jpg) +Figure 12. Generalization to wild text prompts from Sora. Our method is capable of effectively handling the wild text prompts. + +Other reward functions. SHORTFT exhibits remarkable versatility, demonstrating efficacy across a spectrum of reward functions, significantly improving the alignment performance and the quality and fidelity of the output. As shown in Fig. 1 and 12, SHORTFT not only accommodates HPS v2, PickScore, and Symmetry, but also exhibits profi + +
MethodHPS v2↑PickScore↑Symmetry↓
w/o T-LoRA33.4623.820.187
w/o P-Training33.2723.970.146
SHORTFT33.8824.160.138
+ +Table 3. Ablation study on timestep-aware LoRA. + +![](images/00c11bb04059c56e6b37cbe69496c8d16538b52d331a4014e8a64b8fd3a8bd3c.jpg) +Figure 13. Ablation study on progressive training strategy. The red circle marks the incoherent local details, i.e., unsmooth hair. + +ciency in managing Compressibility and Combined reward. Generalization to wild text prompts. As shown in Fig. 12, we present the qualitative results of prompt generalization. We found that using SHORTFT to fine-tune the model on HPDv2 still enables effective handling of wild text prompts, enhancing the overall quality of the generated images. + +# 4.5. Ablation study + +On timestep-aware LoRA. As shown in Table 3, timestep-aware LoRA effectively increases the capacity of the diffusion model and accelerates the convergence of training. Under the same computational cost, the time-aware LoRA achieves better performance. + +On progressive training strategy. The progressive training strategy is committed to eliminating the training-inference gap. As shown in Table 3 and Fig. 13, directly integrating the LoRA parameters obtained from training stage 1 into the pre-trained diffusion model results in incoherent local details, resulting in worse results, which can be effectively handled by progressive training strategy. + +# 5. Conclusion + +In this paper, we propose a novel Shortcut-based Fine-Tuning (SHORTFT), an advanced technique for aligning diffusion models with reward functions through end-to-end backpropagation in the denoising chain. While existing methods struggle with computational costs and the risk of gradient explosion, SHORTFT leverages shorter denoising chains, markedly improving fine-tuning efficiency and effectiveness. Rigorous evaluations demonstrate that our method can be effectively applied to various reward functions, significantly enhancing alignment performance and surpassing state-of-the-art alternative solutions. + +# Acknowledgment + +This work is supported by the National Key Research and Development Plan (2024YFB3309302). + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3, 4, 5 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3 +[3] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 2, 3 +[4] Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffusion models with text-encoder reinforcement learning. In ECCV, 2024.3 +[5] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 2 +[6] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR, 2024. 2, 3, 4, 5, 6, 7 +[7] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3 +[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3 +[10] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023. 2, 3 +[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 3 +[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 2 +[13] Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L Isbell, and Andrea L Thomaz. Policy shaping: Integrating human feedback with reinforcement learning. In NeurIPS, 2013. 2 +[14] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh Susskind, and + +Navdeep Jaitly. Matryoshka diffusion models. arXiv preprint arXiv:2310.15111, 2023. 3 +[15] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Liefeng Bo, and Di Huang. I4vgen: Image as free stepping stone for text-to-video generation. arXiv preprint arXiv:2406.02230, 2024. +[16] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024. +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 +[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 5 +[20] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In NeurIPS, 2023. 3 +[21] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In ECCV, 2024. 3, 4 +[22] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 2, 3 +[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 2, 3, 4 +[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023. 3, 6 +[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3 +[26] Jiachen Li, Weixi Feng, Wenhu Chen, and William Yang Wang. Reward guided latent consistency distillation. arXiv preprint arXiv:2403.11027, 2024. 3, 8 +[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750, 2024. +[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3, 8 +[29] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. In NeurIPS, 2024. 3 + +[30] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. arXiv preprint arXiv:2406.04314, 2024.3 +[31] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3, 4 +[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 3 +[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2 +[34] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2, 3, 4, 5 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 3 +[37] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024. 2, 3, 4, 5 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6 +[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 2, 3, 4 +[40] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 4 +[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 3, 4 +[42] Christoph Schuhmann and Romain Beaumont. Laoin aesthetic predictor. https://laion.ai/blog/laion-aesthetics/, 2022. 2, 3 +[43] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 3 +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 5 +[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3 + +[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 4 +[47] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In NeurIPS, 2020. 2 +[48] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3 +[49] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 3, 6, 7 +[50] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In ICCV, 2023. 2, 3 +[51] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7 +[52] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, 2023. 2, 3, 5 +[53] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. In NeurIPS, 2023. 3 +[54] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3 +[55] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265, 2024. 3 +[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 3, 4 +[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[58] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In ECCV, 2024. 3 +[59] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2 \ No newline at end of file diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/images.zip b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6098974e8fe1da1bc08dc743f4abc34a40d28a71 --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31be54a0cdf47a1aa759395853c5abecae425b1d795caa8866897be209384c5a +size 1100937 diff --git a/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/layout.json b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d87326467aa078698c71b18f7663f3c23f2f66bd --- /dev/null +++ b/2025/ShortFT_ Diffusion Model Alignment via Shortcut-based Fine-Tuning/layout.json @@ -0,0 +1,7916 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "text", + "content": "SHORTFT: Diffusion Model Alignment via Shortcut-based Fine-Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "spans": [ + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "type": "text", + "content": "Xiefan Guo" + }, + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "type": "text", + "content": " Miaomiao Cui Liefeng Bo Di Huang" + }, + { + "bbox": [ + 156, + 142, + 457, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "spans": [ + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "type": "text", + "content": "State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing 100191, China " + }, + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 163, + 536, + 190 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Beihang University, Beijing 100191, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 192, + 375, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 192, + 375, + 205 + ], + "spans": [ + { + "bbox": [ + 234, + 192, + 375, + 205 + ], + "type": "text", + "content": "{xfquo, dhuang}@buaa.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 205, + 429, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 205, + 429, + 218 + ], + "spans": [ + { + "bbox": [ + 177, + 205, + 429, + 218 + ], + "type": "text", + "content": "Project: https://xiefan-guo.github.io/shortft" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 242, + 553, + 495 + ], + "blocks": [ + { + "bbox": [ + 57, + 242, + 553, + 495 + ], + "lines": [ + { + "bbox": [ + 57, + 242, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 242, + 553, + 495 + ], + "type": "image", + "image_path": "eb45afea23bbd1f306fb7e363ff0e5c51de2ac14cd178b9320bcc7344838c899.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 498, + 555, + 554 + ], + "lines": [ + { + "bbox": [ + 54, + 498, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 555, + 554 + ], + "type": "text", + "content": "Figure 1. Example results synthesized by SHORTFT. SHORTFT endeavors to achieve the alignment of diffusion models with reward functions by facilitating the end-to-end backpropagation of the targeted reward gradient throughout the denoising chain. Our method has exhibited remarkable efficacy, particularly evident in the realms of text-image alignment and the overall enhancement of image quality (Top). Moreover, its versatility is underscored by the successful application across diverse reward functions, substantially amplifying alignment performance (Bottom). Combined reward is a weighted combination of rewards: PickScore = 10, HPS v2 = 2, Aesthetic = 0.05." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 563, + 200, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 563, + 200, + 574 + ], + "spans": [ + { + "bbox": [ + 151, + 563, + 200, + 574 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 588, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 588, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 588, + 297, + 696 + ], + "type": "text", + "content": "Backpropagation-based approaches aim to align diffusion models with reward functions through end-to-end backpropagation of the reward gradient within the denoising chain, offering a promising perspective. However, due to the computational costs and the risk of gradient explosion associated with the lengthy denoising chain, existing approaches struggle to achieve complete gradient backpropagation, leading to suboptimal results. In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), an effi" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 556, + 696 + ], + "type": "text", + "content": "cient fine-tuning strategy that utilizes the shorter denoising chain. More specifically, we employ the recently researched trajectory-preserving few-step diffusion model, which enables a shortcut over the original denoising chain, and construct a shortcut-based denoising chain of shorter length. The optimization on this chain notably enhances the efficiency and effectiveness of fine-tuning the foundational model. Our method has been rigorously tested and can be effectively applied to various reward functions, significantly improving alignment performance and surpassing state-of-the-art alternatives." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "678" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 79, + 302, + 114 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 139, + 79 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 139, + 79 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 139, + 79 + ], + "type": "text", + "content": "(a) Vanilla backpropagation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 79, + 302, + 114 + ], + "lines": [ + { + "bbox": [ + 62, + 79, + 302, + 114 + ], + "spans": [ + { + "bbox": [ + 62, + 79, + 302, + 114 + ], + "type": "image", + "image_path": "df8fd5b04a30b4043b3b9758e4f2bfb39986e993be1df3cd6902731f41652753.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 117, + 279, + 125 + ], + "lines": [ + { + "bbox": [ + 63, + 117, + 279, + 125 + ], + "spans": [ + { + "bbox": [ + 63, + 117, + 279, + 125 + ], + "type": "text", + "content": "(c) AlignProp, DRTune: Truncated backpropagation within the denoising chain" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 129, + 302, + 167 + ], + "blocks": [ + { + "bbox": [ + 62, + 129, + 302, + 167 + ], + "lines": [ + { + "bbox": [ + 62, + 129, + 302, + 167 + ], + "spans": [ + { + "bbox": [ + 62, + 129, + 302, + 167 + ], + "type": "image", + "image_path": "0d1e554b1ad63cd4487ec725e8f4a088c6379518f9ff17683997e197df3b58bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 171, + 149, + 178 + ], + "lines": [ + { + "bbox": [ + 78, + 171, + 149, + 178 + ], + "spans": [ + { + "bbox": [ + 78, + 171, + 149, + 178 + ], + "type": "text", + "content": "Sample " + }, + { + "bbox": [ + 78, + 171, + 149, + 178 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mathrm{diffusion}},\\mathbf{c},\\mathbf{z})" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 167, + 171, + 241, + 178 + ], + "lines": [ + { + "bbox": [ + 167, + 171, + 241, + 178 + ], + "spans": [ + { + "bbox": [ + 167, + 171, + 241, + 178 + ], + "type": "text", + "content": "- Vanilla backpropagation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 79, + 548, + 114 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 533, + 79 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 533, + 79 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 533, + 79 + ], + "type": "text", + "content": "(b) DRaFT-K, ReFL: Truncating backpropagation at the end of the denoising chain" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 79, + 548, + 114 + ], + "lines": [ + { + "bbox": [ + 308, + 79, + 548, + 114 + ], + "spans": [ + { + "bbox": [ + 308, + 79, + 548, + 114 + ], + "type": "image", + "image_path": "4a6c79e18b1d266b7ad2c0c2fd328436f4ebea9138349db98ca970384c88a391.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 126, + 548, + 169 + ], + "blocks": [ + { + "bbox": [ + 259, + 171, + 332, + 178 + ], + "lines": [ + { + "bbox": [ + 259, + 171, + 332, + 178 + ], + "spans": [ + { + "bbox": [ + 259, + 171, + 332, + 178 + ], + "type": "text", + "content": "←- Partial backpropagation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 309, + 117, + 454, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 117, + 454, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 117, + 454, + 125 + ], + "type": "text", + "content": "(d) SHORTFT (Ours): Shortcut-based backpropagation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 126, + 548, + 169 + ], + "lines": [ + { + "bbox": [ + 308, + 126, + 548, + 169 + ], + "spans": [ + { + "bbox": [ + 308, + 126, + 548, + 169 + ], + "type": "image", + "image_path": "d1f5902ba3aba90b26d987238069734b662def7ef6b8a67da292b6efafd741c8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 171, + 421, + 178 + ], + "lines": [ + { + "bbox": [ + 351, + 171, + 421, + 178 + ], + "spans": [ + { + "bbox": [ + 351, + 171, + 421, + 178 + ], + "type": "text", + "content": "Sample( " + }, + { + "bbox": [ + 351, + 171, + 421, + 178 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{shortcut}}" + }, + { + "bbox": [ + 351, + 171, + 421, + 178 + ], + "type": "text", + "content": " ,c,z)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 439, + 171, + 533, + 178 + ], + "lines": [ + { + "bbox": [ + 439, + 171, + 533, + 178 + ], + "spans": [ + { + "bbox": [ + 439, + 171, + 533, + 178 + ], + "type": "text", + "content": "←- Shortcut-based backpropagation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "text", + "content": "Figure 2. Comparison of fine-tuning strategies. (a) The vanilla backpropagation-based fine-tuning strategy, which suffers from lengthy backpropagation chains. (b) DRaFT-K [6] and ReFL [52] truncate the backpropagation chain, focusing on the latter half of the denoising chain, they ignore the direct supervision at the early stage, resulting in suboptimal alignment with text prompts. (c) AlignProp [34] and DRTune [51] truncate part of the backpropagation within the denoising chain by disabling some gradients. Specifically, given the denoising operation: " + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t-1} = \\alpha_t \\mathbf{x}_t + \\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t) + c_t \\epsilon" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "text", + "content": ", partial backpropagation only utilizes the gradients of the red part (" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\alpha_t \\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "text", + "content": "), truncating the green part (" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\beta_t \\epsilon_\\theta(\\mathbf{x}_t, t)" + }, + { + "bbox": [ + 55, + 184, + 555, + 262 + ], + "type": "text", + "content": "). This inevitably introduces gradient errors, leading to unstable optimization. (d) Our method, using the few-step diffusion model to construct denoising shortcuts, facilitates complete gradient backpropagation through the entire denoising chain." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 278, + 135, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 135, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 135, + 290 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 420 + ], + "type": "text", + "content": "Diffusion models [1, 9, 17, 18, 22, 32, 38, 44, 45] have established themselves as a pioneering approach in generative modeling, demonstrating exceptional prowess in applications such as photo-realistic text-to-image synthesis. However, the maximum likelihood training objective of these models, which aims to model the training data distribution accurately, can often conflict with downstream goals like aesthetics, fairness, safety, and text-to-image alignment. Therefore, aligning text-to-image diffusion models with human preferences has emerged as a pivotal and practical task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": "Directly supervised fine-tuning on small-scale, human-curated datasets, e.g., LAION Aesthetics [42], presents a straightforward solution. However, the prohibitive cost of data collection and the rapid obsolescence of datasets, particularly in terms of resolution compatibility with the latest text-to-image models, make this approach impractical." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 604 + ], + "type": "text", + "content": "Emulating the successful application of Reinforcement Learning from Human Feedback (RLHF) [5, 13, 33, 47, 59] in Large Language Models (LLMs), several studies [3, 10, 12, 25, 50] have experimented with Reinforcement Learning (RL) techniques to align diffusion models with a reward function. Despite promising performance enhancements in specific domains, RL-based methods are notorious for their high-variance gradients, leading to inefficiencies and limited adaptability to diverse prompts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Recently, backpropagation-based methods [6, 34, 51, 52] have sought to align diffusion models with reward functions using end-to-end backpropagation of the reward gradient through the denoising chain, showing potential. However, these strategies face challenges arising from the lengthy denoising chain, which demands considerable computational resources and is susceptible to gradient explosion. As illustrated in Fig. 2 (b), [6, 52] have made strides by truncating backpropagation to focus on the latter part of the de" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 279, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 376 + ], + "type": "text", + "content": "noising chain. However, they overlook direct supervision in the early stages, leading to suboptimal alignment with text prompts. As illustrated in Fig. 2 (c), [34, 51] truncate a portion of backpropagation within the denoising chain by deactivating some gradients and introduce gradient checkpointing, enabling gradient backpropagation to the early stage of the denoising chain. However, they are time-consuming and introduce gradient bias, leading to optimization instability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 376, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 425 + ], + "type": "text", + "content": "This paper revisits the issue of excessively long denoising chain and propose an alternative approach, employing the shorter denoising chain to facilitate full gradient backpropagation throughout the entire denoising chain." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "type": "text", + "content": "In this paper, we introduce Shortcut-based Fine-Tuning (SHORTFT), which leverages the few-step diffusion model to construct a denoising shortcut to fine-tune the foundational model (e.g., SD 1.5), inspired by recent trajectory-preserving diffusion distillation methods [23, 37, 39, 46]. This technique enables us to bypass the original lengthy denoising chain and complete the inference, creating a shortcut-based denoising chain of shorter length. In addition, we construct a timestep-aware LoRA as an expert LoRA ensemble, based on the intriguing temporal dynamics exhibited by the text-to-image diffusion model during the denoising process. This approach increases the number of trainable parameters without increasing cost in the inference phase, enabling faster convergence and improved performance. We also devise a custom progressive training strategy to mitigate training inference bias introduced by using denoising shortcuts during the training stage." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Extensive quantitative and qualitative analyses demonstrate that SHORTFT can be effectively applied to various reward functions and architectures, significantly enhancing alignment performance. Furthermore, SHORTFT, benefiting from the short denoising chain and without the need for gradient checkpointing, is particularly efficient, learning faster than DRTune [51], the current most efficient method." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "679" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 553, + 152 + ], + "type": "image", + "image_path": "e14e44833c570bbebe4aeb01441cec2603f28e7aef93b3ce283f64f7f6b1bdc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": "Figure 3. Denoising shortcut. The trajectory-preserving few-step diffusion model naturally introduces denoising shortcut, allowing for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. The 4-step Hyper-SD distilled from SD 1.5 is used in our experiments. SHORTcut(i) denotes completing the denoising process from timestep " + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": " to 0 using the few-step diffusion model. In addition, we also provide the well-known one-step denoising results " + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "inline_equation", + "content": "\\mathrm{DDIM}(i)" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "inline_equation", + "content": "\\mathrm{DDIM}(i)" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": " denotes performing a one-step denoising operation from timestep " + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": " to 0. Intuitively, the output of " + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "inline_equation", + "content": "\\mathrm{DDIM}(i)" + }, + { + "bbox": [ + 55, + 156, + 555, + 244 + ], + "type": "text", + "content": " is more blurred, lacking accurate structure and texture details, while the output of SHORTcut(i) is closer to the original output of SD 1.5. Furthermore, we provide their corresponding HPS v2 scores, with the absolute value of the deviation from the score corresponding to SD 1.5 provided in parentheses, SHORTcut(i) exhibits smaller deviations. These observations collectively indicate the reliability and validity of the denoising shortcut." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 141, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 141, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 141, + 268 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 219, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 219, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 219, + 288 + ], + "type": "text", + "content": "2.1. Alignment of diffusion models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 295, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 377 + ], + "type": "text", + "content": "Diffusion models [9, 18, 44, 45] have become a dominant force in generative modeling, showing exceptional performance in diverse applications [1, 14-17, 22, 32, 38, 53, 57]. However, certain misalignments with human intentions can arise. Recent research, fueled by the successful alignment of large language models, has sought to align diffusion models with human expectations and preferences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 449 + ], + "type": "text", + "content": "Fine-tuning via data augmentation. Several studies [7, 10, 20, 25, 50] have explored altering the training data distribution for fine-tuning diffusion models on visually compelling and textually coherent data, which has led to improved results. Other methods [2, 43] involve re-captioning pre-collected web images to enhance textual precision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 628 + ], + "type": "text", + "content": "Fine-tuning via reward models. Reward models [24, 25, 42, 49, 50, 52] are employed to emulate human preferences given an input prompt and generated images. Several approaches have attempted to integrate these signals to augment text-to-image generation. A significant direction is the utilization of reinforcement learning-based algorithms [3, 4, 8, 11, 58] for fine-tuning text-to-image diffusion models in alignment with these rewards, [29, 30, 36, 48, 54, 55] bypass it entirely with Direct Preference Optimization. However, these methods are costly and have high gradient variance, leading to inefficiency and limited adaptability to diverse prompts. Consequently, backpropagation-based techniques [6, 34, 51, 52] have been explored, which directly fine-tune diffusion models using differentiable rewards [24, 42, 49, 50, 52]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "The challenge of backpropagation-based strategies stem from the lengthy denoising chain, which often requires numerous denoising operations (e.g., 50 for DDIM), corresponding to a long backpropagation chain. This process incurs substantial time and memory costs and is prone to gradient explosion. To mitigate this issue, [6, 52] truncate backpropagation by concentrating on the latter part" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 257, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 376 + ], + "type": "text", + "content": "of the denoising chain. While these approaches yield some improvements, they neglect direct supervision in the early stage of the denoising chain, leading to less precise alignment with text prompts. [34, 51] truncate part of the backpropagation within the denoising chain by deactivating some gradients. By employing gradient checkpointing, they allows for the propagation of gradients to the early stages of the denoising chain. Despite their merits, these techniques can be computationally demanding and induce gradient bias, resulting in optimization instability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": "Different from previous methods, this paper revisits the fundamental challenge of excessively long denoising chain and proposes an alternative approach: leveraging the shorter denoising chain to facilitate full gradient backpropagation throughout the denoising chain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 432, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 432, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 432, + 455 + ], + "type": "text", + "content": "2.2. Diffusion distillation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 593 + ], + "type": "text", + "content": "Existing methods for the distillation of diffusion models can be primarily classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]. The former aims to preserve the original denoising trajectory dictated by an ordinary differential equation (ODE), while the latter focuses on leveraging the denoising endpoint as the main supervision, disregarding the intermediate trajectory steps. This paper focuses on trajectory-preserving distillation, supporting the establishment of a denoising shortcut within the complete denoising chain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 512, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 512, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 512, + 613 + ], + "type": "text", + "content": "2.3. Fine-tuning few-step diffusion models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Existing works [26-28] have successfully explored finetuning few-step diffusion models. While they share similarities with our research in utilizing few-step diffusion models, a crucial difference is that we focus on fine-tuning the foundational models. In contrast to the foundational models, few-step diffusion models suffer from performance degradation and reduced capacity caused by the distillation process, leading fine-tuning them suboptimal (see Sec. 4.4)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "680" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 550, + 148 + ], + "type": "image", + "image_path": "a28ef26760df40c1dfbc6a55a089b37d55d8ec57dfeb47d4d19ad7d13e8cb83b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 153, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 555, + 187 + ], + "type": "text", + "content": "Figure 4. Illustration of SHORTFT. The core of the proposed method is the Shortcut-based Fine-Tuning (SHORTFT), which leverages the trajectory-preserving few-step diffusion model as the shortcut (identified as blue arrow) to achieve direct end-to-end backpropagation through the diffusion sampling process, fine-tuning the parameters of the pre-trained diffusion model to align it with the reward function." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 195, + 124, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 124, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 124, + 208 + ], + "type": "text", + "content": "3. SHORTFT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": "Our method, Shortcut-based Fine-Tuning (SHORTFT), capitalizes on the trajectory-preserving few-step diffusion model as a shortcut to achieve direct end-to-end backpropagation through the diffusion sampling process. This approach fine-tunes the parameters of the pre-trained diffusion model to align it with the reward function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 178, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 178, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 178, + 308 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "text", + "content": "In line with [6, 34, 51], SHORTFT focuses on fine-tuning the parameters " + }, + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "text", + "content": " of pre-trained diffusion models to maximize the differentiable reward function " + }, + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot)" + }, + { + "bbox": [ + 55, + 315, + 296, + 364 + ], + "type": "text", + "content": " for generated images. This can be formally represented as in Eq. 1:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 373, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 373, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 66, + 373, + 295, + 388 + ], + "type": "interline_equation", + "content": "J (\\boldsymbol {\\theta}) = \\mathbb {E} _ {\\mathbf {c}, \\mathbf {x} _ {T} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} [ \\mathcal {R} (\\text {S a m p l e} (\\boldsymbol {\\theta}, \\mathbf {c}, \\mathbf {x} _ {T}), \\mathbf {c}) ], \\tag {1}", + "image_path": "92f6b1462fc442052f8ee785b9280add9f7c11457a4e1c427092525389717997.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathrm{Sample}(\\theta, \\mathbf{c}, \\mathbf{x}_T)" + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "text", + "content": " represents the denoising process for the timestep " + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "inline_equation", + "content": "t = T \\to 0" + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "text", + "content": " with prompt condition " + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 397, + 295, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "text", + "content": "Consistent with [6, 34, 51], Eq. 1 is resolved by calculating " + }, + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{R}" + }, + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "text", + "content": " (Sample " + }, + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "inline_equation", + "content": "(\\theta ,\\mathbf{c},\\mathbf{x}_T)" + }, + { + "bbox": [ + 55, + 422, + 296, + 493 + ], + "type": "text", + "content": ", c) and employing gradient ascent. The computation of this gradient necessitates backpropagation through multiple diffusion models in the denoising chain, akin to backpropagation through time in recurrent neural networks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 503, + 167, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 167, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 167, + 516 + ], + "type": "text", + "content": "3.2. Denoising shortcut" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "text", + "content": "The recent emergence of a series of diffusion-aware distillation algorithms [23, 37, 39, 46] has been instrumental in mitigating the computational burden associated with the multi-step inference process of diffusion models. These algorithms can be roughly classified into two categories: trajectory-preserving distillation [23, 37, 39, 46] and trajectory-reformulating distillation [21, 31, 40, 41, 56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "Among these, trajectory-preserving few-step diffusion models naturally introduce a denoising shortcut, which allows for flexible skipping within the denoising chain while still ensuring high-quality and consistent image synthesis. As demonstrated in Fig. 3, the integration of these shortcuts into the denoising chain significantly reduces the total number of denoising steps, thereby shortening the length of the denoising chain. This key insight paves the way for efficient and effective end-to-end backpropagation." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 203, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 321, + 194, + 381, + 202 + ], + "lines": [ + { + "bbox": [ + 321, + 194, + 381, + 202 + ], + "spans": [ + { + "bbox": [ + 321, + 194, + 381, + 202 + ], + "type": "text", + "content": "(a) SD denoising chain" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 203, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 321, + 203, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 321, + 203, + 547, + 236 + ], + "type": "image", + "image_path": "bff8c66218c5a096b0eac7fe9122ad0c44ca29d3563e12531ff9f6c7d3bee0a5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 321, + 249, + 547, + 291 + ], + "blocks": [ + { + "bbox": [ + 321, + 240, + 414, + 248 + ], + "lines": [ + { + "bbox": [ + 321, + 240, + 414, + 248 + ], + "spans": [ + { + "bbox": [ + 321, + 240, + 414, + 248 + ], + "type": "text", + "content": "(b) Shortcut-based denoising chain" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 249, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 321, + 249, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 321, + 249, + 547, + 291 + ], + "type": "image", + "image_path": "9cc46b97ab6bdf01e9754f9c0d251fde482e8da6eb24cfe99606f103cfbf0d29.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 296, + 554, + 395 + ], + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 395 + ], + "type": "text", + "content": "Figure 5. Illustration of timestep-aware LoRA. (a) Vanilla SD denoising chain; (b) Shortcut-based denoising chain. In accordance with the interesting time dynamics in the text-to-image diffusion model denoising process revealed by [1], different from existing methods that share the same LoRA parameters at all timesteps, we introduce time-step aware LoRA, which effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing the computational cost during the inference stage." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 414, + 462, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 414, + 462, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 414, + 462, + 426 + ], + "type": "text", + "content": "3.3. Shortcut-based fine-tuning" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "type": "text", + "content": "Insight. The naive optimization of Eq. 1 through backpropagation necessitates the construction of the complete denoising chain: " + }, + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\dots ,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\dots ," + }, + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}" + }, + { + "bbox": [ + 313, + 433, + 555, + 553 + ], + "type": "text", + "content": ". This process involves the storage of intermediate activations linked to each neural layer and each denoising timestep within GPU VRAM, which is not feasible due to memory constraints. Furthermore, the typical length of a denoising chain is approximately 50, which results in an overly long backpropagation chain that can lead to issues of gradient explosion." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 554, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 651 + ], + "type": "text", + "content": "The key insight of SHORTFT is rooted in the utilization of the trajectory-preserving few-step diffusion model to construct the denoising shortcut, significantly reducing the length of the denoising chain. As elaborated in Sec. 3.2, this denoising shortcut enables us to bypass a substantial number of denoising timesteps, leading to create a more streamlined and efficient denoising chain: " + }, + { + "bbox": [ + 313, + 554, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_T,\\dots ,\\mathbf{x}_t,\\mathbf{x}_{t - \\Delta t},\\mathbf{x}_{t - \\Delta t - 1},\\mathbf{x}_{t - 2\\Delta t - 1},\\dots ,\\mathbf{x}_0\\}" + }, + { + "bbox": [ + 313, + 554, + 555, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 652, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 554, + 700 + ], + "type": "text", + "content": "As depicted in Fig. 4, such a design allows for the direct implementation of reward supervision at the early stages of the denoising chain, and facilitates full gradient backpropagation throughout the denoising chain." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Shortcut-based denoising chain. As illustrated in Fig. 5," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "681" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 290, + 220 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 290, + 220 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 290, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 290, + 220 + ], + "type": "image", + "image_path": "f7a8a5ada6bc548c7d540ab54d0d17c5f164f2cdddf68b1f9207b9139d6bf506.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 269 + ], + "type": "text", + "content": "Figure 6. Illustration of progressive training strategy. Corresponding to the time-step aware LoRA design, we develop the progressive training strategy, which eliminates the training-inference gap introduced by the shortcut-based fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": "a vanilla SD denoising chain comprises a sequence of step-by-step denoising operations that transform the input noise " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_T" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": " into the output image " + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 288, + 296, + 373 + ], + "type": "text", + "content": ". This process typically necessitates numerous denoising steps. By harnessing the denoising shortcut, we are able to construct a shortcut-based denoising chain and fine-tune the diffusion model through end-to-end backpropagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": "Timestep-aware LoRA. Instead of fine-tuning the full weights of the original diffusion model, Low-Rank Adaptation (LoRA) [19] preserves the weights of the pre-trained model and introduces new low-rank weight matrices alongside the original model weights. The contributions of these matrices are summed to generate the final outputs. Specifically, each linear layer of the UNet of SD is modified from " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = \\mathbf{W}\\mathbf{x}" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = \\mathbf{W}\\mathbf{x} + \\mathbf{B}\\mathbf{A}\\mathbf{x}" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "inline_equation", + "content": "k \\ll d" + }, + { + "bbox": [ + 55, + 374, + 296, + 506 + ], + "type": "text", + "content": ". LoRA considerably reduces the number of parameters to be optimized, thereby decreasing the memory requirements for fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "content": "Moreover, [1] uncovers intriguing temporal dynamics during the denoising process of the text-to-image diffusion model. In the initial sampling stage, the model largely depends on the text prompts to guide the sampling process. As the generation progresses, the model gradually leans on visual features to denoise the image. This indicates that sharing LoRA parameters (standard practice in the existing methods) throughout the entire denoising process may not be optimal and may fail to capture the distinct patterns that emerge during denoising. Therefore, in contrast to [6, 34, 51, 52] which share the same LoRA parameters for all timesteps, we introduce timestep-aware LoRA. This design effectively increases the capacity of the diffusion model and accelerates the convergence of training, without increasing computational cost in the inference phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Specifically, as depicted in Fig. 5, we initially divide the entire denoising chain into " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " segments, " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta t = \\left\\lfloor \\frac{T}{k}\\right\\rfloor" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". Ex-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "cept for the first segment, we assign a corresponding LoRA to the last timestep of each subsequent segment. For the first segment, we adopt the methodology of [6] and share the same LoRA parameters across all timesteps. Notably, a sequence of continuous timesteps lacking LoRA exists in later segments, supporting the denoising shortcuts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 240 + ], + "type": "text", + "content": "Progressive training strategy. Although the few-step diffusion model is capable of creating the denoising shortcut, bypassing the entire denoising chain, it inherently introduces errors in the output, meaning it still cannot be fully consistent with the output of SD, particularly in fine details, as shown in Fig. 3. This incongruity can lead to a training-inference gap, resulting in suboptimal results. To mitigate this, we design a progressive training strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": "As illustrated in Fig. 6, in accordance with the design of timestep-aware LoRA, we divide the SHORTFT training process into " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": " stages. For the " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": "-th training stage, we optimize the weights of LoRA " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": " to LoRA " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": "-th segment and preceding denoising processes, we retain the original denoising chain, while for the denoising processes post the " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": "-th segment, we introduce the denoising shortcut, thereby shortening the depth of the backpropagation chain. Moreover, in line with [6], we also employ the truncated backpropagation technique at the " + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 241, + 554, + 361 + ], + "type": "text", + "content": "-th segment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 398 + ], + "type": "text", + "content": "During inference, as displayed in Fig. 6, the denoising shortcut is bypassed, and the original denoising chain is used to generate the final output image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 395, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 395, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 395, + 423 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 440, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 440, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 440, + 443 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 496 + ], + "type": "text", + "content": "In our experiments, Stable Diffusion 1.5 serves as the foundational diffusion model. The DDIM schedule [44] is employed to execute 50 steps of denoising, with a classifier-free guidance scale of 7.5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 652 + ], + "type": "text", + "content": "Shortcut. SHORTFT shortens the denoising chain by bypassing certain steps within the chain. Hence, the selection of few-step diffusion models is focused predominantly on methods that employ trajectory-preserving distillation algorithms. To accommodate the proposed time-aware LoRA, one-step diffusion models are avoided, whose output image quality is also relatively inferior. Specifically, 4-step Hyper-SD [37], distilled from SD 1.5, is utilized to construct the denoising shortcut. The value of " + }, + { + "bbox": [ + 313, + 498, + 554, + 652 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 498, + 554, + 652 + ], + "type": "text", + "content": " is set to 4, and the timesteps configured for LoRA are {761, 501, 261, 1}. Consequently, the denoising shortcuts are executed separately between timesteps 741 to 501, timesteps 481 to 261, and timesteps 241 to 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Timestep-aware LoRA. As suggested by [6], LoRA is applied to both the feedforward and attention layers in the UNet. The LoRA rank is set to 128. Furthermore, we adopt a stepwise stacking approach, where for LoRA " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", we introduce a new LoRA branch on top of LoRA " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "i - 1" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "682" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 535, + 365 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "type": "image", + "image_path": "3b04d6475d0b1c018e64c0032ddea6c5039d1fee22ac6d01f524b4a78ae778b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on PickScore and HPS v2. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 402, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 533 + ], + "type": "text", + "content": "Reward functions. The proposed method is evaluated using three reward functions: Human Preference Score v2 (HPS v2) [49], PickScore [24], and Symmetry [51]. HPS v2 and PickScore capture human preference for images based on input prompts, while Symmetry encourages images to have horizontal symmetry features. Different from [51], which utilizes CLIPScore [35] as a regularization term, our experiment amalgamates HPS v2 and PickScore in a ratio of 1:10 to function as a joint regularization term. This approach has demonstrated superior performance in terms of text-image alignment and overall image quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": "Experimental details. All experiments are conducted using 2 A800 GPUs and the AdamW optimizer with " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": ", and a weight decay of 0.1. SHORTFT is performed with a batch size of 128 and a constant learning rate of " + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": ". During training, the pre-trained SD parameters are converted to bfloat16 to reduce memory usage, while the LoRA parameters under training remain in float32. Gradient checkpointing is not required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 712 + ], + "type": "text", + "content": "Datasets. We compare SHORTFT-finetuned diffusion models to those of the state-of-the-art counterparts on the Human Preference Score v2 dataset (HPDv2) [49]. The final reward is computed on the 400 prompts from the test split. Following [51], for fair comparison, we evaluate all methods using the same computational budget. Specifically, all methods are trained for six hours on 2 A800 GPUs." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 329, + 400, + 542, + 472 + ], + "blocks": [ + { + "bbox": [ + 329, + 400, + 542, + 472 + ], + "lines": [ + { + "bbox": [ + 329, + 400, + 542, + 472 + ], + "spans": [ + { + "bbox": [ + 329, + 400, + 542, + 472 + ], + "type": "table", + "html": "
MethodHPS v2↑PickScore↑Symmetry↓
SD 1.5 [38]26.9120.460.853
DRaFT-LV [6]33.1323.350.418
DRTune [51]32.7923.220.207
SHORTFT33.8824.160.138
", + "image_path": "2d90aaa1213146e7cc2555afe4ddf7049811dbe0e58005012d4d9b9544836fd9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 476, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 499 + ], + "type": "text", + "content": "Table 1. Objective evaluation. Our method performs over other counterparts, under the same computational cost." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 514, + 446, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 446, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 446, + 528 + ], + "type": "text", + "content": "4.2. Qualitative comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "content": "Fig. 7 and 8 present the quantitative comparison of our results against those of representative methods, including current state-of-the-art techniques, under same text prompts and random seeds. SD 1.5 suffers from low-quality image generation. DRaFT [6], while proficient at managing local image details, struggles to effectively handle global layouts and the optimization of the symmetry reward function. Similarly, DRTune [51] is constrained by gradient inaccuracies, which contribute to instability during training and a deficiency in managing complex semantics. One particular area of weakness is the synthesis of images with specific counts of objects, such as accurately depicting a given number of cows. In contrast, SHORTFT exhibits enhanced capabilities in generating images that are both visually realistic and semantically faithful, outperforming over other counterparts." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "683" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 71, + 535, + 365 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 535, + 365 + ], + "type": "image", + "image_path": "066169bad1c7248978d8e5c9b23bd445a964eb7a40b25ed4795988fdb52eaa76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 393 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison on HPS v2 and Symmetry. Each image is generated with the same text prompt and random seed for all methods. Our method outperforms existing methods in both text-image alignment and image quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 403, + 282, + 460 + ], + "blocks": [ + { + "bbox": [ + 67, + 403, + 282, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 403, + 282, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 282, + 460 + ], + "type": "image", + "image_path": "401a6f5f6634bf4694c4774099fe5225bb898ee0bf0213aec23bfc81132ecd67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 487 + ], + "type": "text", + "content": "Figure 9. Human and AI preference evaluation against current methods. SHORTFT performs over other counterparts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 195, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 195, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 195, + 504 + ], + "type": "text", + "content": "4.3.Quantitative comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 509, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 713 + ], + "type": "text", + "content": "Objective evaluation. Table 1 shows the quantitative results achieved on the Human Preference Score v2 benchmark [49], where the proposed method outperforms the other approaches, clearly demonstrating its effectiveness. It is important to highlight that both our method and DRTune [51] strategically employ backpropagation of the reward gradient to the initial stages of the denoising chain. This intentional design choice significantly enhances the Symmetry score performance when compared to alternative methods. Despite this, DRTune continues to grapple with the challenge of gradient bias. In contrast, our method significantly mitigates this issue, delivering superior performance. User study. A subjective user study comprising 11 volunteers is conducted, with five possessing expertise in image processing and the remaining participants having no background in computer vision or graph. Participants are tasked to select the most visually attractive and semantically ac" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 405, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 501 + ], + "type": "text", + "content": "curate image among those generated by our method and current state-of-the-art techniques. Each participant has 10 questions for each pair of comparisons. Furthermore, an MLLM-assisted evaluation is employed using GPT-4V. We make 20 queries to GPT-4V for each pair of comparisons. More details are provided in the Appendix. As depicted in Fig. 9, the results exhibit a significant inclination towards SHORTFT in comparison to other techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 509, + 397, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 397, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 397, + 521 + ], + "type": "text", + "content": "4.4. More results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 526, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 634 + ], + "type": "text", + "content": "SHORTFT, 10k training step. Due to the shorter denoising and backpropagation chains, our method achieves superior performance under the same computational budget. Furthermore, to validate the upper bound of our approach, following the protocol in [6], we conduct the full training process using HPS v2 reward function on HPDv2, comprising 10k training steps, and evaluate it on the corresponding benchmark. The obtained HPS v2 score of 35.97 surpasses the reported score for DRaFT-LV in [6]." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 348, + 644, + 523, + 678 + ], + "blocks": [ + { + "bbox": [ + 348, + 644, + 523, + 678 + ], + "lines": [ + { + "bbox": [ + 348, + 644, + 523, + 678 + ], + "spans": [ + { + "bbox": [ + 348, + 644, + 523, + 678 + ], + "type": "table", + "html": "
MethodTuning Hyper-SDTuning SD 1.5
HPS v2↑32.9235.97
", + "image_path": "b831730447c47c2ae34f4474d16b44a618525060396a1996a9e276ee7492805e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 681, + 550, + 694 + ], + "lines": [ + { + "bbox": [ + 317, + 681, + 550, + 694 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 550, + 694 + ], + "type": "text", + "content": "Table 2. Objective evaluation on tuning SD 1.5 and Hyper-SD." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 714 + ], + "type": "text", + "content": "Fine-tuning SD vs. Hyper-SD. As mentioned in Sec. 2.3," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "684" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 118 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 118 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 118 + ], + "type": "image", + "image_path": "9793504cf4bae9a8204d59ee35176efc78f931fa676aa14438c9dbedeadb2649.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 156 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison on tuning SD 1.5 (SHORTFT) and Hyper-SD. Fine-tuning SD 1.5 significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 162, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 162, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 296, + 293 + ], + "type": "text", + "content": "[26-28] explore fine-tuning few-step diffusion models and have achieved certain successes. However, compared to the foundational model, few-step diffusion models face performance degradation caused by the distillation process. Finetuning the few-step diffusion model is actually suboptimal compared to fine-tuning the foundational model. Furthermore, following two strategies, we separately conduct the training processes on the HPDv2 using the HPS v2 reward function. As shown in Table 2 and Fig. 10, fine-tuning SD 1.5 (SHORTFT) significantly outperforms fine-tuning Hyper-SD, with the former enjoying more exquisite details." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 303, + 294, + 389 + ], + "blocks": [ + { + "bbox": [ + 58, + 303, + 294, + 389 + ], + "lines": [ + { + "bbox": [ + 58, + 303, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 294, + 389 + ], + "type": "image", + "image_path": "3a3dcd691a9729e6d9aab4d09235fad431334f3739ed423ddd16735b1fb4f841.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 394, + 292, + 405 + ], + "lines": [ + { + "bbox": [ + 58, + 394, + 292, + 405 + ], + "spans": [ + { + "bbox": [ + 58, + 394, + 292, + 405 + ], + "type": "text", + "content": "Figure 11. Example results synthesized by SHORTFT on SD 3." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 472 + ], + "type": "text", + "content": "Fine-tuning SD 3. SHORTFT is an architecture-agnostic fine-tuning strategy, applicable to both UNet-based (SD 1.5) and Transformer-based (SD 3) architectures. As illustrated in Fig. 11, SHORTFT is also capable of mastering SD 3, where SD 3 aligns with HPS v2." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 481, + 293, + 608 + ], + "blocks": [ + { + "bbox": [ + 59, + 481, + 293, + 608 + ], + "lines": [ + { + "bbox": [ + 59, + 481, + 293, + 608 + ], + "spans": [ + { + "bbox": [ + 59, + 481, + 293, + 608 + ], + "type": "image", + "image_path": "b0e476c64b57777984eb424274a223b9f13ee8d140e3d1a974cc426f7ec67399.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 635 + ], + "type": "text", + "content": "Figure 12. Generalization to wild text prompts from Sora. Our method is capable of effectively handling the wild text prompts." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Other reward functions. SHORTFT exhibits remarkable versatility, demonstrating efficacy across a spectrum of reward functions, significantly improving the alignment performance and the quality and fidelity of the output. As shown in Fig. 1 and 12, SHORTFT not only accommodates HPS v2, PickScore, and Symmetry, but also exhibits profi" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 537, + 129 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 537, + 129 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 537, + 129 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 537, + 129 + ], + "type": "table", + "html": "
MethodHPS v2↑PickScore↑Symmetry↓
w/o T-LoRA33.4623.820.187
w/o P-Training33.2723.970.146
SHORTFT33.8824.160.138
", + "image_path": "0c16d52c13f0f3fc8e0bd921ad084060b2b59e1c6e5af983531dfbd317d0fe83.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 133, + 527, + 144 + ], + "lines": [ + { + "bbox": [ + 340, + 133, + 527, + 144 + ], + "spans": [ + { + "bbox": [ + 340, + 133, + 527, + 144 + ], + "type": "text", + "content": "Table 3. Ablation study on timestep-aware LoRA." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 340, + 148, + 531, + 223 + ], + "blocks": [ + { + "bbox": [ + 340, + 148, + 531, + 223 + ], + "lines": [ + { + "bbox": [ + 340, + 148, + 531, + 223 + ], + "spans": [ + { + "bbox": [ + 340, + 148, + 531, + 223 + ], + "type": "image", + "image_path": "00c11bb04059c56e6b37cbe69496c8d16538b52d331a4014e8a64b8fd3a8bd3c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 554, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 554, + 250 + ], + "type": "text", + "content": "Figure 13. Ablation study on progressive training strategy. The red circle marks the incoherent local details, i.e., unsmooth hair." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 331 + ], + "type": "text", + "content": "ciency in managing Compressibility and Combined reward. Generalization to wild text prompts. As shown in Fig. 12, we present the qualitative results of prompt generalization. We found that using SHORTFT to fine-tune the model on HPDv2 still enables effective handling of wild text prompts, enhancing the overall quality of the generated images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 406, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 406, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 406, + 350 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 413 + ], + "type": "text", + "content": "On timestep-aware LoRA. As shown in Table 3, timestep-aware LoRA effectively increases the capacity of the diffusion model and accelerates the convergence of training. Under the same computational cost, the time-aware LoRA achieves better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": "On progressive training strategy. The progressive training strategy is committed to eliminating the training-inference gap. As shown in Table 3 and Fig. 13, directly integrating the LoRA parameters obtained from training stage 1 into the pre-trained diffusion model results in incoherent local details, resulting in worse results, which can be effectively handled by progressive training strategy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 659 + ], + "type": "text", + "content": "In this paper, we propose a novel Shortcut-based Fine-Tuning (SHORTFT), an advanced technique for aligning diffusion models with reward functions through end-to-end backpropagation in the denoising chain. While existing methods struggle with computational costs and the risk of gradient explosion, SHORTFT leverages shorter denoising chains, markedly improving fine-tuning efficiency and effectiveness. Rigorous evaluations demonstrate that our method can be effectively applied to various reward functions, significantly enhancing alignment performance and surpassing state-of-the-art alternative solutions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 670, + 407, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 407, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 407, + 683 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Plan (2024YFB3309302)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "685" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 246 + ], + "type": "text", + "content": "[3] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 249, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 249, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 295, + 292 + ], + "type": "text", + "content": "[4] Chaofeng Chen, Annan Wang, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Enhancing diffusion models with text-encoder reinforcement learning. In ECCV, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 327 + ], + "type": "text", + "content": "[5] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 361 + ], + "type": "text", + "content": "[6] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. In ICLR, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 418 + ], + "type": "text", + "content": "[7] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 462 + ], + "type": "text", + "content": "[8] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 486 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 542 + ], + "type": "text", + "content": "[10] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 294, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 598 + ], + "type": "text", + "content": "[11] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "text", + "content": "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 294, + 700 + ], + "type": "text", + "content": "[13] Shane Griffith, Kaushik Subramanian, Jonathan Scholz, Charles L Isbell, and Andrea L Thomaz. Policy shaping: Integrating human feedback with reinforcement learning. In NeurIPS, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "type": "text", + "content": "[14] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh Susskind, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "Navdeep Jaitly. Matryoshka diffusion models. arXiv preprint arXiv:2310.15111, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 129 + ], + "type": "text", + "content": "[15] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Liefeng Bo, and Di Huang. I4vgen: Image as free stepping stone for text-to-video generation. arXiv preprint arXiv:2406.02230, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 130, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 554, + 163 + ], + "type": "text", + "content": "[16] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 554, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 186 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 186, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 208 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 210, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 210, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 210, + 554, + 252 + ], + "type": "text", + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 255, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 554, + 297 + ], + "type": "text", + "content": "[20] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 300, + 554, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 554, + 343 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 554, + 343 + ], + "type": "text", + "content": "[21] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. In ECCV, 2024. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 345, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 554, + 376 + ], + "type": "text", + "content": "[22] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 378, + 554, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 554, + 433 + ], + "type": "text", + "content": "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 554, + 477 + ], + "type": "text", + "content": "[24] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 479, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 554, + 533 + ], + "type": "text", + "content": "[25] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 535, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 567 + ], + "type": "text", + "content": "[26] Jiachen Li, Weixi Feng, Wenhu Chen, and William Yang Wang. Reward guided latent consistency distillation. arXiv preprint arXiv:2403.11027, 2024. 3, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 569, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 554, + 623 + ], + "type": "text", + "content": "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. arXiv preprint arXiv:2405.18750, 2024." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 679 + ], + "type": "text", + "content": "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. arXiv preprint arXiv:2410.05677, 2024. 3, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 554, + 713 + ], + "type": "text", + "content": "[29] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "686" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[30] Zhanhao Liang, Yuhui Yuan, Shuyang Gu, Bohan Chen, Tiankai Hang, Ji Li, and Liang Zheng. Step-aware preference optimization: Aligning preference with denoising performance at each step. arXiv preprint arXiv:2406.04314, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "text", + "content": "[31] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "text", + "content": "[32] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "text", + "content": "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[34] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[37] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. arXiv preprint arXiv:2404.13686, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "type": "text", + "content": "[39] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[40] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "text", + "content": "[41] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[42] Christoph Schuhmann and Romain Beaumont. Laoin aesthetic predictor. https://laion.ai/blog/laion-aesthetics/, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "text", + "content": "[43] Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation. arXiv preprint arXiv:2310.16656, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "text", + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 667 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[46] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[47] Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[48] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[49] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "type": "text", + "content": "[50] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[51] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. In ECCV, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[52] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, 2023. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[53] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 475 + ], + "type": "text", + "content": "[54] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 509 + ], + "type": "text", + "content": "[55] Shentao Yang, Tianqi Chen, and Mingyuan Zhou. A dense reward view on aligning text-to-image diffusion with preference. arXiv preprint arXiv:2402.08265, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[56] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "type": "text", + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[58] Yinan Zhang, Eric Tzeng, Yilun Du, and Dmitry Kislyuk. Large-scale reinforcement learning for diffusion models. In ECCV, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 623, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 667 + ], + "type": "text", + "content": "[59] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "687" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_content_list.json b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3e4848dd536be554cb28483b9eccf862806b16ed --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_content_list.json @@ -0,0 +1,1525 @@ +[ + { + "type": "text", + "text": "ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers", + "text_level": 1, + "bbox": [ + 127, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qianhao Yuan $^{1,2}$ , Qingyu Zhang $^{1,2}$ , Yanjiang Liu $^{1,2}$ , Jiawei Chen $^{1,2}$ , Yaojie Lu $^{1,*}$ , Hongyu Lin $^{1}$ , Jia Zheng $^{1,*}$ , Xianpei Han $^{1}$ , Le Sun $^{1}$", + "bbox": [ + 214, + 202, + 781, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Software, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences", + "bbox": [ + 290, + 238, + 709, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yuanqianhao2024,zhangqingyu2024,liuyanjiang2021,chenjiawei2024, luyaojie,hongyu,zhengjia,xianpei,sunle}@iscas.ac.cn", + "bbox": [ + 122, + 273, + 870, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) suffer from high computational costs due to their massive size and the large number of visual tokens. In this paper, we investigate layer-wise redundancy in MLLMs by introducing a novel metric, Layer Contribution (LC), which quantifies the impact of a layer's transformations on visual and text tokens, respectively. The calculation of LC involves measuring the divergence in model output that results from removing the layer's transformations on the specified tokens. Our pilot experiment reveals that many layers of MLLMs exhibit minimal contribution during the processing of visual tokens. Motivated by this observation, we propose ShortV, a training-free method that leverages LC to identify ineffective layers, and freezes visual token updates in these layers. Experiments show that ShortV can freeze visual token in approximately $60\\%$ of the MLLM layers, thereby dramatically reducing computational costs related to updating visual tokens. For example, it achieves a $50\\%$ reduction in FLOPs on LLaVA-NeXT-13B while maintaining superior performance. The code is publicly available at https://github.com/icip-cas/ShortV.", + "bbox": [ + 88, + 377, + 485, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have achieved remarkable performance in natural language tasks [1, 12, 37, 47, 49]. Building upon LLMs, Multimodal Large Language Models (MLLMs) [31, 32, 38, 39] take a significant step towards understanding the real physical world by incorporating visual information into their processes. Typically, an MLLM consists of a visual encoder, a projector, and an LLM backbone. Most of them preprocess visual information through", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66b6621944e119e1e75bbc22db6ecbb29532ea43e4bb6380c238b633a9345249.jpg", + "image_caption": [ + "(a) ShortV" + ], + "image_footnote": [], + "bbox": [ + 540, + 339, + 880, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/635f725d3c85e7cae0aab88e6316eec731d238d26f4ac361155a5e70a2063179.jpg", + "image_caption": [ + "(b) Performance vs. the Number of ShortV Layers", + "Figure 1. (a) Illustration of ShortV. We identify ineffective layers for visual tokens and replace these layers with sparse ShortV layers. In ShortV layers, we freeze visual tokens, and eliminate computations related to updating them. ShortV improves MLLM efficiency in a training-free manner and involves no parameter updates. Notably, ShortV is compatible with token pruning methods, e.g. FastV. (b) Performance vs. the number of ShortV layers. Average Performance means a normalized average score on multiple benchmarks. ShortV can freeze visual tokens in approximately $60\\%$ of the MLLM layers with nearly no performance degradation." + ], + "image_footnote": [], + "bbox": [ + 517, + 463, + 901, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a visual encoder, e.g. a CLIP-ViT [14, 41], and project the patch-level visual features into visual tokens through a projector. Then they concatenate visual and text tokens and feed them into the LLM backbone.", + "bbox": [ + 511, + 821, + 906, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, MLLMs face a substantial increase in compu", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "329", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tational overhead. This burden primarily stems from the large scale of the LLM backbones and the significantly extended length of the concatenated visual-text token sequences. To address this issue, Chen et al. [10] discovers significant token-wise redundancy in MLLMs. Based on this redundancy, they propose FastV, which identifies and prunes unimportant visual tokens in MLLMs to improve their efficiency.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other than token-wise redundancy, in this paper, we reveal that MLLMs also exhibit significant layer-wise redundancy in processing visual tokens. Specifically, we propose Layer Contribution (LC), a metric that quantifies how much a layer's transformations on certain tokens contribute to the model's output. In the LC calculation of a layer, we freeze certain tokens in this layer, i.e. keep the hidden states of the tokens unchanged, and then compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original model. This metric provides a direct measure of a layer's importance for certain tokens. By comparing LC scores on visual tokens and those on text tokens, we discover that MLLM layers are ineffective for visual tokens, and their transformations on visual tokens contribute minimal to the model's output.", + "bbox": [ + 89, + 214, + 482, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This phenomenon inspires us to propose ShortV, a simple but effective method to improve the efficiency of MLLMs. In ShortV, we first utilize the LC metric to identify layers least effective at transforming visual tokens, and then replace these layers with sparse ShortV layers. Within these sparse layers, visual tokens remain frozen, and the corresponding computations for updating them are eliminated, as shown in Figure 1a.", + "bbox": [ + 89, + 441, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of ShortV, we conduct evaluations across multiple benchmarks, including MME [15], MMBench [33], MMMU [55], MMStar [9], SEED-Bench [25], GQA [19], and Flickr30K [40]. Figure 1b illustrates the correlation between the normalized average performance on these benchmarks and the number of replaced layers. As observed, ShortV can replace approximately $60\\%$ of MLLM layers without performance degradation. Unlike FastV and other token pruning methods, ShortV reduces computations of per visual token rather than reducing the number of visual tokens. Therefore, ShortV and token pruning methods are orthogonal and compatible. Furthermore, we demonstrate that combining ShortV and FastV can further enhance MLLM efficiency.", + "bbox": [ + 89, + 564, + 482, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contribution as follows.", + "bbox": [ + 109, + 777, + 393, + 791 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Layer Contribution (LC), a metric to quantify how much a layer's transformations on specific tokens contribute to the model's output.", + "- Leveraging LC, we reveal significant redundancy in MLLM layers for visual tokens. Transformations on visual tokens in many layers contribute minimally and are thus ineffective." + ], + "bbox": [ + 91, + 795, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Based on the observation above, we propose ShortV, which improves MLLM efficiency by freezing visual tokens in ineffective layers. ShortV can freeze visual tokens in approximately $60\\%$ of MLLM layers without performance degradation. Extensive experiments and ablation studies demonstrate ShortV's effectiveness.", + "bbox": [ + 513, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Layer Redundancy in MLLMs", + "text_level": 1, + "bbox": [ + 511, + 194, + 795, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the background of layer redundancy in text-only LLMs, and then propose a metric to measure MLLM layer redundancy for different types of tokens. Next, we conduct a pilot experiment to investigate layer redundancy in MLLMs.", + "bbox": [ + 511, + 220, + 903, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Background", + "text_level": 1, + "bbox": [ + 511, + 304, + 643, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typically, MLLMs are built on text-only LLMs. MLLMs employ pre-trained visual encoders, such as CLIP-ViT [14, 41], to convert images into visual features, and then use projectors to project them into visual tokens in the text token embedding space. The visual tokens are concatenated with text tokens and fed into the LLM backbones.", + "bbox": [ + 511, + 325, + 903, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For text-only LLMs, Men et al. [36] identify notable redundancy across their layer. Some layers' transformations on the hidden states of text token contribute minimally to the overall model functionality. Consequently, these layers are considered ineffective. Removing these transformations in approximately $25\\%$ of LLM layers has minimal impact on model outputs. Such redundancy mainly occurs in middle-to-deeper layers, whereas initial layers and the last layer remain critical to the model functionality. However, this pattern may not hold for MLLMs. Huang et al. [18] demonstrate a clear modality gap in the embedding space of current MLLMs, where visual and text tokens exhibit a uniform distribution within each modality but a significant distribution gap between modalities. Such a modality gap implies that MLLMs might adopt distinct computational patterns or strategies for processing visual and text tokens, potentially affecting how redundancy is distributed across layers. This raises several key questions: Are MLLM layers as ineffective for visual tokens as LLM layers are for text tokens? To what extent does layer redundancy exist in MLLMs? How is this redundancy distributed across different MLLM layers?", + "bbox": [ + 511, + 417, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Layer Contribution Metric", + "text_level": 1, + "bbox": [ + 511, + 757, + 756, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To investigate layer redundancy for certain tokens, we freeze these tokens within the investigated layer, i.e. keep hidden states of these tokens unchanged. To achieve this, we introduce sparse layers shown in Figure 2 for visual and text tokens, respectively. Based on these designs, we propose the Layer Contribution (LC) metric, which evaluates how much a layer's transformations on certain tokens contribute to the model's overall functionality. In the calcu", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "330", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44a32d4ecee0764438b174d8aa73406d7836394a5427d968f4a082df6e8b35ef.jpg", + "image_caption": [ + "(a) Sparse layer where visual tokens are frozen, used to investigate layer redundancy for visual tokens. Only text tokens function as queries and are passed through the FFN. We also denote this layer as ShortV layer." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 483, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/605d72357eb987a089e455802200c69111b29563f3df670ac362d840f83f468e.jpg", + "image_caption": [ + "(b) Sparse layer where text tokens are frozen, used to investigate layer redundancy for text tokens. Only visual tokens function as queries and are passed through the FFN." + ], + "image_footnote": [], + "bbox": [ + 96, + 233, + 370, + 311 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9525adcce18eef1a74c836efa284244c85b3cfd37fd441d513d3309fb0e6a00f.jpg", + "image_caption": [ + "Figure 2. Sparse layers used to investigate layer redundancy for different tokens. To investigate layer redundancy for certain tokens, we freeze these tokens within the layer, i.e. keep hidden states of these tokens unchanged, and measure the divergence between the model's output logits and those of the original model. We gray out the attention that does not need calculation." + ], + "image_footnote": [], + "bbox": [ + 380, + 232, + 473, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lation of LC, we replace the investigated dense layer with the corresponding sparse layer, and compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original dense model.", + "bbox": [ + 89, + 474, + 482, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, We assume the LLM backbone $M$ has $L$ layers, and each of them consists of a self-attention block and a feed-forward network (FFN). The input of the $i$ -th layer at the $j$ -th token position is $H_{i}^{j}$ , and the corresponding output is $H_{i+1}^{j}$ . The output of the last layer, i.e. the $L$ -th layer, at the same position is $H_{L+1}^{j}$ . The model $M$ utilizes $H_{L+1}^{-1}$ at the last token position to compute logits for next-token prediction through the language model head $LM_{head}$ :", + "bbox": [ + 89, + 536, + 483, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {l o g i t s} (M) = L M _ {\\text {h e a d}} \\left(H _ {L + 1} ^ {- 1}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 678, + 482, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate how much the $i$ -th layer's transformations on certain tokens $X$ contribute to the model functionality, we replace the $i$ -th layer with a sparse layer where $X$ are frozen, i.e. $X$ 's hidden states remain unchanged in this sparse layer. The resulting model is denoted as $\\mathcal{M}_i^X$ . In practice, $X$ can be visual tokens $V$ or text tokens $T$ . As shown in Figure 2, we introduce sparse layers where $V$ and $T$ are frozen, respectively. In Figure 2a, we freeze the visual tokens in the sparse layer. Within the self-attention block of this layer, the visual tokens do not attend to other tokens, and only the text tokens function as queries. For the FFN of this layer, we simply do not pass the visual tokens through it. In Figure 2b, we freeze the text tokens in another sparse", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "layer with similar designs. Based on these, we define the $i$ -th layer's Layer Contribution (LC) score for certain tokens $X$ as the KL divergence between the output logits of the original model $M$ and those of the model $\\mathcal{M}_i^X$ where $X$ are frozen in the $i$ -th layer:", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL C _ {i} ^ {X} = K L \\left(\\operatorname {l o g i t s} (M), \\operatorname {l o g i t s} \\left(\\mathcal {M} _ {i} ^ {X}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 178, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "here $KL(\\cdot)$ denotes KL divergence. A lower LC score implies that the layer's transformations on the tokens exhibit minimal contribution to the model's output, suggesting that these transformations are ineffective.", + "bbox": [ + 511, + 208, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion: Why not use perplexity or cosine similarity as the metric to measure the importance of layers? Some work in text-only LLMs utilizes perplexity [23, 46] or cosine similarity [11, 17, 36] as metrics to measure the importance of each layer. For the former metric, instead of KL divergence, they measure the change in perplexity of the models, and the layers causing minimal perplexity changes are deemed ineffective. This metric, however, is inadequate when measuring layer redundancy for visual tokens. We find that even if we do not feed visual tokens into the MLLMs, they can still generate reasonable responses, and the changes in perplexity of the MLLMs is relatively low. Nevertheless, they face significant performance degradation in vision language tasks when the visual information is absent. Thus, perplexity is not a reliable measure when evaluating layer redundancy for visual tokens.", + "bbox": [ + 511, + 277, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the latter metric, the cosine similarity between the input and output of a certain layer is calculated. The hypothesis here is that the ineffective layers have less transformations on the hidden states of tokens, and therefore their inputs and outputs demonstrate higher similarities. However, in our evaluation on LLaVA-1.5 [31] models, the cosine similarity metric and LC differ in their measurement of the redundancy distribution across different layers. Compared with the LC metric, which directly measures the logits divergence between model outputs, cosine similarity consistently overestimates the redundancy of the shallow layers and underestimates the redundancy of the deep layers. We believe the reason behind this difference is that cosine similarity neglects the position of a layer within the model. Specifically, minor transformations of hidden states in the shallow layers can influence all subsequent layers, whereas transformations with similar extent in deeper layers tend to have less impact on overall model functionality.", + "bbox": [ + 511, + 520, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Ineffective MLLM Layers for Visual Tokens", + "text_level": 1, + "bbox": [ + 511, + 801, + 888, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We conduct a pilot experiment to investigate layer redundancy in LLaVA-1.5-7B and LLaVA-1.5-13B [31]. We first randomly sample 2,000 cases from a combination of two major vision language tasks, including caption (Flickr30K [40]) and visual question answering", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "331", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/50ae17a45224c197973a813145c51a42819f3f150e8116e07b9580967f191d5b.jpg", + "image_caption": [ + "(a) LLaVA-1.5-7B" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 491, + 250 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fb94ad90f8c6e4bbae719f773e9f78a260782cdc9741d3a5273d4714abcf1b61.jpg", + "image_caption": [ + "(b) LLaVA-1.5-13B" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 903, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(GQA [19]). Then we utilize these samples to calculate each MLLM layer's average LC score for visual and text tokens, respectively. Figure 3 shows the results. We summarize our findings as follows.", + "bbox": [ + 88, + 345, + 482, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, for text tokens, middle to deeper layers are more ineffective, while the initial and last layers make more contributions to the MLLM functionality. These observations align with the layer redundancy distribution of text-only LLMs found in Men et al. [36], indicating that visual instruction tuning [30, 31] does not significantly alter the manner LLMs process text tokens.", + "bbox": [ + 88, + 407, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, for visual tokens, the initial and the deep layers, including the last one, exhibit higher redundancy than other layers, which is different from the distribution for text token. Notably, since the last layer's transformations on visual tokens do not contribute to the model's output, its LC score for visual tokens is always 0.", + "bbox": [ + 88, + 515, + 482, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Third, layer redundancy shows an imbalance between visual and text tokens. Each layer's LC score on visual tokens are lower than that on text tokens, which means that many layers' transformations on visual tokens are ineffective, and freezing visual token in these ineffective layers results in minimal impact on the models' output.", + "bbox": [ + 88, + 607, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We attribute the different layer redundancy patterns for different modalities to the modality gap. The clear distribution gap of visual and text tokens result in the difference in how MLLMs process them. We hope these findings can provide insights into how MLLMs process visual and text tokens in different layers.", + "bbox": [ + 88, + 699, + 482, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. ShortV", + "text_level": 1, + "bbox": [ + 89, + 806, + 176, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Freezing Visual Tokens in Ineffective Layers", + "text_level": 1, + "bbox": [ + 89, + 830, + 467, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As demonstrated in the previous section, we identify significant layer redundancy for visual tokens in MLLMs. Most layers' transformations on visual tokens are ineffective for", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1c34b6b57f6727d65239d91bf26949f6695bdd211480c8d59101295abf334f00.jpg", + "image_caption": [ + "Figure 3. The Layer Contribution (LC) scores of LLaVA-1.5-7B and LLaVA-1.5-13B. A lower LC score implies that the layer's transformations on the specified tokens are more ineffective. Layers are more ineffective for visual tokens than for text tokens, and freezing visual tokens in ineffective layers results in minimal output divergence from the original model.", + "Figure 4. Details of ShortV layer. In this layer, only text tokens pass through the $W_{Q}$ and $W_{O}$ matrices and the FFN. The attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries." + ], + "image_footnote": [], + "bbox": [ + 558, + 347, + 861, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the model functionality. Based on this observation, we propose a direct method to enhance MLLM efficiency in a training-free manner: freezing visual tokens in ineffective layers. We denote this method as ShortV.", + "bbox": [ + 511, + 607, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In ShortV, we replace the ineffective dense layers with sparse ShortV layers, where visual tokens are frozen. The ShortV layers are same as the sparse layer shown in 2a, and we illustrate their detailed architecture in Figure 4. In each ShortV layer, only text tokens are passed through the FFN and the $W_{Q}$ and $W_{O}$ matrices of the attention block, and the attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries.", + "bbox": [ + 511, + 671, + 906, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ShortV has one parameter: the number of replaced layers, which we denote as $N$ . First, we construct a tiny dataset, which contains a small number of samples from vision language tasks. Then we use this dataset to calculate each layer's average LC score for visual tokens. Next, we sort the layers in ascending order according to LC scores,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "332", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/ebe88fb54e577822306932ecbba5a453d7b3bb79a44182e3f7ea588ea3821c63.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTFLOPsFLOPs RatioVQAv2GQASEED-BenchMMMU (val)MMEMMBench ENMMStarFlickr30K CIDEr
LLaVA-1.5-7B
Vanilla8.5100%76.561.966.136.31510.764.133.774.9
FastV (K=2, R=50%)4.958%73.560.265.435.81475.664.332.467.5
VTW (K=16)4.755%66.355.166.236.11497.064.032.844.5
ShortV (Ours, N=19)4.755%75.760.966.236.21503.164.833.371.3
LLaVA-1.5-13B
Vanilla16.6100%78.063.368.235.41531.368.936.179.6
FastV (K=2, R=50%)9.457%76.759.467.834.61506.668.335.973.4
VTW (K=20)9.155%75.360.668.234.91533.068.536.165.9
ShortV (Ours, N=24)9.155%77.262.068.035.81535.968.637.176.4
LLaVA-NeXT-7B
Vanilla42.7100%80.064.170.236.41519.067.137.169.7
FastV (K=2, R=50%)22.052%79.563.069.635.11482.066.336.566.7
VTW (K=16)21.851%75.655.870.235.71518.267.137.638.4
ShortV (Ours, N=19)21.651%78.863.470.436.01525.167.237.865.7
LLaVA-NeXT-13B
Vanilla81.8100%80.965.771.935.91570.069.339.966.7
FastV (K=2, R=50%)42.151%76.862.971.535.91546.468.539.666.0
VTW (K=20)41.751%77.061.571.834.81569.469.139.856.6
ShortV (Ours, N=24)41.050%79.763.671.836.21553.070.239.967.5
", + "bbox": [ + 94, + 88, + 906, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of various training-free methods for MLLM efficiency. FLOPs Ratio denotes the proportion of FLOPs retained after applying the corresponding method to improve MLLM efficiency, compared with the vanilla model.", + "bbox": [ + 89, + 455, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the layers with lower LC scores are more ineffective for visual tokens. Finally, we freeze visual tokens in the $N$ layers with the lowest LC scores by replacing them with sparse ShortV layers, while keeping their original parameters.", + "bbox": [ + 88, + 510, + 480, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ShortV is training-free and involves no parameter updates. It can be applied to various MLLMs for different vision language tasks. Notably, ShortV is orthogonal to and compatible with popular visual token pruning methods, e.g. FastV [10]. Visual token pruning directly reduces the number of visual tokens, while ShortV mitigates the computational overhead related to each visual token. This means that we can apply ShortV and token pruning at the same time to further improve MLLM efficiency.", + "bbox": [ + 88, + 571, + 482, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Computational Cost", + "text_level": 1, + "bbox": [ + 89, + 715, + 282, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the computations of self-attention blocks and feed-forward networks (FFNs) in layers of the LLM backbone. Assume $t$ is the number of text tokens, $v$ is the number of visual tokens, $h$ is the hidden state size, $m$ is the intermediate size of the FFNs, the total FLOPs of one dense Transformer layer can be estimated as:", + "bbox": [ + 89, + 737, + 483, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF L O P s = 2 (t + v) (4 h + 3 m) h + 4 (t + v) ^ {2} h. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 835, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For one ShortV layer, the FLOPs can be calculated as:", + "bbox": [ + 89, + 861, + 449, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF L O P s ^ {*} = 2 t (4 h + 3 m) h + 4 v h ^ {2} + 4 t (t + v) h. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 883, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the whole model, assume the LLM has $L$ layers in total, ShortV selects $N$ ineffective dense layers and replaces them with ShortV layers. The FLOPs of the original dense model are $L \\times FLOPs$ , and the FLOPs of ShortV are calculated as $(L - N) \\times FLOPs + N \\times FLOPs^{*}$ . The FLOPs ratio of ShortV and the original model is computed as:", + "bbox": [ + 511, + 510, + 906, + 602 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr = \\frac {(L - N) \\times F L O P s + N \\times F L O P s ^ {*}}{L \\times F L O P s}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 612, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 656, + 648, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 511, + 681, + 710, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. To validate the effectiveness of ShortV, we conduct experiments on popular open-source MLLMs, such as LLaVA-1.5-7B [31], LLaVA-1.5-13B, LLaVA-NeXT-7B [32] and LLaVA-NeXT-13B. LLaVA-1.5 models process images with a $336 \\times 336$ resolution and treat each image as 576 tokens. LLaVA-NeXT splits high-resolution images into subimages, and encode the subimages and down-sampled original images independently. This allows the models to scale the input to any arbitrary resolution, without performing positional embedding interpolation for ViTs [14]. LLaVA-NeXT scales the input image resolution to $4 \\times$ and visual token number up to $5 \\times$ compared with LLaVA-1.5, i.e. 2880 tokens for each image.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "333", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2fb67566c27355a598e0643213fbcc9b969b7db87c518b9c1af64617ac945486.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# ShortV Layers (N)TFLOPsFLOPs RatioMMEMMBench ENMMMU (val)MMStarSEED-BenchGQAFlickr30K CIDErAvg.Per.
LLaVA-1.5-7B (32 layers)
08.5100%1510.764.136.333.766.161.974.958.9100.0
86.981%1508.664.336.033.866.261.474.558.899.8
165.362%1487.064.936.133.365.761.072.858.399.0
194.755%1503.164.836.233.366.260.971.358.399.0
243.744%1341.760.734.133.462.558.364.254.392.2
LLaVA-1.5-13B (40 layers)
016.6100%1531.368.935.436.168.263.379.661.2100.0
814.185%1521.968.635.636.068.263.079.060.999.5
1611.670%1534.968.636.336.268.062.978.561.099.7
249.155%1535.968.635.837.168.062.076.460.799.2
326.640%1298.864.533.636.063.259.368.455.791.0
LLaVA-NeXT-7B (32 layers)
042.7100%1519.067.136.437.170.264.169.760.1100.0
833.879%1515.167.236.636.970.264.170.060.1100.0
1624.958%1476.867.236.237.370.263.567.859.498.8
1921.651%1525.167.236.037.870.463.465.759.599.0
2416.037%1504.165.436.436.068.160.564.958.196.7
LLaVA-NeXT-13B (40 layers)
081.8100%1570.070.535.939.971.965.766.761.3100.0
868.283%1552.470.635.039.671.965.166.961.099.5
1654.667%1561.070.135.039.771.964.866.960.999.3
2441.050%1553.070.236.239.971.863.667.561.099.5
3227.534%1468.465.835.238.969.360.558.557.493.6
", + "bbox": [ + 135, + 88, + 867, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance vs. Efficiency Balance of ShortV under different configurations. # ShortV Layers $(N)$ : the number of ShorV layers, Avg.: a normalized average score on the benchmarks, Per.: the relative performance retention compared with the vanilla models.", + "bbox": [ + 89, + 460, + 906, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To evaluate the effectiveness of ShortV, which improves MLLM efficiency in a training-free manner, we compare it with popular training-free methods for MLLM efficiency, such as FastV [10] and VTW [29]. FastV drops visual tokens by a percentage of $R$ after the $K$ -th layer in the forward process of input tokens. It computes the average attention score one token received from all other tokens as the importance criterion to select pruned tokens. VTW drops all visual tokens after the $K$ -th layer, enabling only text tokens to engage in the subsequent layers. We use the default settings for the baselines as in their original papers. Specifically, for FastV, $K = 2$ and $R = 50\\%$ . For VTW, $K = 16$ for 7B models and $K = 20$ for 13B models.", + "bbox": [ + 88, + 515, + 480, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 726, + 232, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to compare ShortV with the baselines. The results are shown in Table 1. We perform evaluation on multiple popular vision language benchmarks, including MME [15], MMBench [33], MMMU (val) [55], MMStar [9], SEED-Bench [25], VQAv2 [16], and GQA [19]. We manually choose the number of ShortV layers $N$ to maintain a similar or lower FLOPs ratio compared with the baselines. The FLOPs ratio of ShortV is calculated according to Equation 5, where we set the number of text tokens to 64. Specifically, we", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "choose $N = 19$ for the 7B models, and $N = 24$ for the 13B models. As shown in Table 1, our ShortV achieves comparable or superior performance across multiple benchmarks compared with the baselines.", + "bbox": [ + 511, + 515, + 906, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Balance between Efficiency and Performance", + "text_level": 1, + "bbox": [ + 511, + 589, + 895, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct an experiment to investigate the impact of ShortV's parameter $N$ , which denotes the number of ShortV layers. The experimental results are presented in Table 2. To facilitate intuitive comprehension, we plot the correlation between the normalized average score on benchmarks and the number of ShortV layers in Figure 1b.", + "bbox": [ + 511, + 612, + 905, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe that ShortV can freeze visual tokens in approximately $60\\%$ of the MLLM layers while preserving superior performance. As $N$ continues to increase, both 7B and 13B models can maintain more than $90\\%$ performance when the hidden states of visual tokens remain unchanged in about $80\\%$ of the layers. These results are significantly different from those on text-only LLMs. For LLMs, Men et al. [36] remove transformations on text tokens in approximately $25\\%$ of the LLM layers, and this results in about $10\\%$ performance degradation on language benchmarks. As the number of layers increases, the performance of LLMs rapidly declines. These differences align with our observation in Section 2.3 that layers are more ineffective for visual", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "334", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/138d2b7ca9a58584623247a9bd6f74bc4d3b113c7bfc64a0b81f02bab0f974b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# ShortV Layers (N)08161924
LLaVA-1.5-7B1.00×1.13×1.23×1.30×1.40×
LLaVA-NeXT-7B1.00×1.15×1.35×1.44×1.64×
", + "bbox": [ + 94, + 88, + 478, + 140 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49e39406ce903f452cfbd5f98f406577985315613f5643718061cbfd6c54d72e.jpg", + "table_caption": [ + "Table 3. Inference speedups over the vanilla models, based on the 7B models. We conduct this test on a single GPU." + ], + "table_footnote": [], + "table_body": "
# ShortV Layers (N)08162432
LLaVA-1.5-13B1.00×1.13×1.24×1.39×1.50×
LLaVA-NeXT-13B1.00×1.13×1.30×1.52×1.84×
", + "bbox": [ + 94, + 200, + 480, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/70cfe4b23769aa0153c4ce72efda1fe1e56b363a61580d5beea6607f3530f099.jpg", + "table_caption": [ + "Table 4. Inference speedups over the vanilla models, based on the 13B models. We conduct this test on a single GPU." + ], + "table_footnote": [], + "table_body": "
MethodFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
FastV58%64.335.865.460.2
ShortV55%64.836.266.260.9
ShortV+FastV29%64.237.165.159.3
", + "bbox": [ + 94, + 315, + 480, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tokens than for text tokens.", + "bbox": [ + 91, + 500, + 272, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the theoretical FLOPs ratios, we provide the speedups on real hardware using different settings, as shown in Table 3 and Table 4. To get rid of the influence of different output sequence lengths, we use the first token latency to calculate the speedups. We utilize the MMMU dataset for the latency test. For comparison, we note that FastV [10] with its default setting, i.e. $K = 2$ and $R = 50\\%$ , achieves a $1.31 \\times$ speedup over the vanilla LLaVA-1.5-13B model. In contrast, our ShortV with its default parameter, i.e. $N = 24$ , achieves a greater speedup of $1.39 \\times$ .", + "bbox": [ + 89, + 518, + 483, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Orthogonal to Token Pruning", + "text_level": 1, + "bbox": [ + 89, + 691, + 354, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we demonstrate that ShortV is orthogonal to and compatible with visual token pruning, e.g. FastV [10]. FastV identifies $R\\%$ unimportant visual tokens and drops them after the $K$ -th layer in the forward process of input tokens. We apply FastV to ShortV, which already replaces $N$ ineffective layers for visual tokens with ShortV layers. We use the default settings for FastV and ShortV in this experiment, i.e. $K = 2$ and $R = 50\\%$ for FastV, and $N = 19$ for ShortV. We employ LLaVA-1.5-7B as the vanilla model. The experimental results in Table 5 demonstrate that ShortV is compatible with FastV and that the application of both can further improve MLLM efficiency.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dae87ed262657bb97eda50243d7710185adf36e8ef0c90f8ddcb840d6c283175.jpg", + "table_caption": [ + "Table 5. ShortV is compatible with FastV, and applying both at the same time can further enhance MLLM efficiency. This experiment is based on LLaVA-1.5-7B." + ], + "table_footnote": [], + "table_body": "
StrategyFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
Random55%58.433.660.556.1
Cosine Sim.55%60.834.262.759.5
LC (Ours)55%64.836.266.260.9
", + "bbox": [ + 517, + 88, + 903, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation on strategies to select replaced layers, based on LLaVA-1.5-7B. \"Random\" denotes randomly selecting 19 layers and freezing visual tokens in them. \"Cosine Sim.\" denotes using cosine similarity to select ineffective layers for visual tokens.", + "bbox": [ + 511, + 196, + 906, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 280, + 678, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on strategies to select replaced layers. In this paragraph, we perform an ablation experiment on LLaVA-1.5-7B to investigate the impact of strategies for selecting which layers to replace. ShortV selects ineffective layers for visual tokens, and replace them with ShortV layers. To identify which layers are ineffective, we utilize the LC metric introduced in Section 2. In contrast, previous work [36] on text-only LLMs uses a metric based on cosine similarity. It calculates the average cosine similarity between the inputs and outputs of each layer. The layers with higher cosine similarities are deemed more ineffective. To make a comparison between this cosine similarity metric and our LC metric, we calculate each layer's cosine similarity between the input hidden states and output hidden states of visual tokens, and select the same number of layers with the highest cosine similarities. We show the comparison in Table 6. We also include the results of another baseline, ShortV (Random), where visual tokens are frozen in the same number of randomly selected layers. These results clearly demonstrate that our LC metric performs better than cosine similarity in identifying ineffective MLLM layers for visual tokens, and ShortV based on the LC metric achieves performance comparable to the vanilla model. In contrast, ShortV based on the cosine similarity metric cannot match the performance of the vanilla model, although it outperforms the baseline with randomly selected layers.", + "bbox": [ + 511, + 301, + 906, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on frozen tokens. In this paragraph, we conduct an ablation study on LLaVA-1.5-7B to investigate the impact of freezing different types of tokens. In Section 2, we demonstrate that MLLM layers are ineffective for visual tokens, as measured by the LC metric. Motivated by this observation, ShortV freezes visual tokens in ineffective layers. In Table 7, we compare our method with the strategies of freezing other tokens. In the experiment detailed in line (a), we utilize the LC metric to identify 19 ineffective layers for text tokens, and freeze text tokens in them. Despite having fewer frozen tokens, we can observe that this strategy results in significant performance declines compared with our method, which freezes visual tokens rather than text to", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "335", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/106e4dbd13532c8635ff48687f165b7c0154c6919609d4cb9f13163d284001a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Frozen TokensMMBench ENMMMU (val)SEED-BenchGQA
None (Vanilla)64.036.366.161.9
(a) Text2.123.78.92.9
(b) Text+Visual1.326.60.80.0
(c) Random1.522.95.52.3
(d) Visual (Ours)64.836.266.260.9
", + "bbox": [ + 119, + 88, + 455, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation on frozen tokens, based on LLaVA-1.5-7B. (a) identifying 19 ineffective layers for text tokens and freezing text tokens in them. In lines (b) and (c), we select ineffective layers for all tokens. line (b) involves freezing all input tokens in them, whereas line (c) denotes randomly freezing the same number of tokens as the visual tokens.", + "bbox": [ + 89, + 202, + 482, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "kens. These experimental results align with our findings in Section 2 that MLLM layers are more ineffective for visual tokens than for text tokens. In lines (b) and (c), we first calculate each layer's average LC score for all tokens, including visual and text tokens, and then select 19 ineffective layers. Next, in the experiment corresponding to line (b), we freeze all input tokens in these layers. In line (c), we freeze random input tokens, and the number of frozen tokens matches that of the visual tokens. As a result, the computational overhead associated with line (c) is the same as that of our method in line (d). We can find that freezing tokens other than visual tokens leads to substantial performance degradation in vision-language tasks. These ablations demonstrate the effectiveness of our ShortV.", + "bbox": [ + 88, + 311, + 482, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 536, + 232, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 561, + 408, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Built upon Large Language Models (LLMs) [1, 12, 37, 47, 49], Multimodal Large Language Models (MLLMs) [4, 7, 35, 48, 52, 57] have made significant progress in processing and understanding the visual world. Typically, they use a decoder-only architecture. Specifically, they utilize visual encoders [14, 21, 34, 41, 56] to convert input visual information into visual features and then use projectors to project these visual features into visual tokens. These visual tokens are then concatenated with text tokens and fed into the LLM backbones. Current MLLMs use hundreds to thousands of visual tokens to represent a single image, significantly increasing the length of the token sequences. For instance, the LLaVA-1.5 models [31] transform each image with $336 \\times 336$ resolution into 576 tokens. For images with higher resolutions, the LLaVA-NeXT models [32] process images into up to 2,880 visual tokens, and SPHINX-2k [28] divides one image into nine subimages, resulting in 2,890 visual tokens. Applying LLMs to such large numbers of visual tokens incurs substantial computational costs. In this paper, we introduce ShortV to enhance the efficiency of MLLMs by reducing the computational overhead associ", + "bbox": [ + 91, + 583, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ated with visual tokens.", + "bbox": [ + 513, + 90, + 671, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Efficient LLMs and MLLMs", + "text_level": 1, + "bbox": [ + 511, + 119, + 769, + 133 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For LLMs, previous studies [36, 46] find that layers in LLMs are ineffective for text tokens. They remove computations in about $25\\%$ of the layers, while preserving approximately $90\\%$ of the performance. LaCo [50] utilizes layer merging for efficient LLMs.", + "bbox": [ + 511, + 143, + 903, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the computational inefficiency of MLLMs, previous methods [10, 29, 42, 44, 45, 54] have primarily focused on two aspects: efficient model architecture and visual token compression. Among efficient model architectures, cross-attention-based models [2, 3, 8, 20] insert gated cross-attention layers within LLM layers for visual perception, but previous studies [13, 24] demonstrate that this architecture performs worse than the decoder-only architecture in the same settings. Instead of inserting cross-attention layers, mPLUG-Owl3 [51] and Vamba [42] introduce cross-attention operations in parallel with self-attention. In contrast, SAISA [54] introduces NAAViT self-attention blocks, which incorporate multimodal cross-attention into the original self-attention operations of the LLMs, and reuse the parameters of self-attention blocks. The design of ShortV layers is inspired by NAAViT. Differently, in ShortV layers, visual tokens also skip their FFNs.", + "bbox": [ + 511, + 220, + 903, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual token compression methods improve MLLM efficiency in both training-based [5, 6, 22, 26, 27] and training-free [10, 43, 53] manners. FastV [10] reveals token-wise redundancy, and it removes unimportant tokens during inference. In this paper, we reveal layer-wise redundancy in MLLMs. Layers in MLLMs are much more ineffective for visual tokens than for text token. Therefore, we can freeze visual tokens in approximately $60\\%$ of the MLLM layers with minimal performance degradation. Unlike previous methods for MLLM efficiency, ShortV does not reduce the number of visual tokens but instead decreases the computational costs of processing each token. ShortV is training-free and orthogonal to token compression. We demonstrate that ShortV is compatible with FastV, allowing for simultaneous application to further enhance MLLM efficiency.", + "bbox": [ + 511, + 479, + 903, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore the layer-wise redundancy in MLLMs. We discover that layers in MLLMs are more ineffective for visual tokens than for text tokens. MLLM layers' transformations on visual tokens have a minimal impact on the MLLM output. Motivated by this observation, we propose ShortV, a training-free method to enhance MLLM efficiency. ShortV utilizes our proposed LC metric to select ineffective layers for visual tokens, and freezes visual tokens in these layers. It can freeze visual tokens in about $60\\%$ of the layers while preserving superior performance.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "336", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 261, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265) and the Basic Research Program of ISCAS (Grant No. ISCAS-ZD-202401).", + "bbox": [ + 89, + 113, + 483, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 8", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8", + "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8", + "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models, 2023. 8", + "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8", + "[6] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 8", + "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8", + "[8] Kaibing Chen, Dong Shen, Hanwen Zhong, Huasong Zhong, Kui Xia, Di Xu, Wei Yuan, Yifei Hu, Bin Wen, Tianke Zhang, et al. Evlm: An efficient vision-language model for visual understanding. arXiv preprint arXiv:2407.14177, 2024. 8", + "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 6" + ], + "bbox": [ + 99, + 262, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5, 6, 7, 8", + "[11] Xiaodong Chen, Yuxuan Hu, Jing Zhang, Yanling Wang, Cuiping Li, and Hong Chen. Streamlining redundant layers to compress large language models. arXiv preprint arXiv:2403.19135, 2024. 3", + "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 8", + "[13] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024.8", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2, 5, 8", + "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 2, 6", + "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6", + "[17] Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. What matters in transformers? not all attention is needed. arXiv preprint arXiv:2406.15786, 2024. 3", + "[18] Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Deciphering cross-modal alignment in large vision-language models with modality integration rate. arXiv preprint arXiv:2410.07167, 2024. 2", + "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 4, 6", + "[20] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.8", + "[21] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "337", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 8", + "[23] Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim, Thibault Castells, Shinkook Choi, Junho Shin, and Hyoung-Kyu Song. Shortened llama: Depth pruning for large language models with comparison of retraining methods. arXiv preprint arXiv:2402.02834, 2024. 3", + "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2025. 8", + "[25] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 6", + "[26] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 8", + "[27] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 8", + "[28] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 8", + "[29] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 6, 8", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 4", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 4, 5, 8", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,5, 8", + "[33] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 2, 6", + "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 8" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024.8", + "[36] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect, 2024. URL https://arxiv.org/abs/2403.03853, 2024. 2, 3, 4, 6, 7, 8", + "[37] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2(5):6, 2024. 1, 8", + "[38] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1", + "[39] OpenAI. Introducing gpt-4o: our fastest and most affordable flagship model. https://platform.openai.com/docs/guides/vision, 2024. Accessed: 2024-05-26. 1", + "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 2, 3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 8", + "[42] Weiming Ren, Wentao Ma, Huan Yang, Cong Wei, Ge Zhang, and Wenhu Chen. Vamba: Understanding hour-long videos with hybrid mamba-transformers. arXiv preprint arXiv:2503.11579, 2025. 8", + "[43] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 8", + "[44] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8", + "[45] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 8", + "[46] Jiwon Song, Kyungseok Oh, Taesu Kim, Hyungjun Kim, Yulhwa Kim, and Jae-Joon Kim. Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks. arXiv preprint arXiv:2402.09025, 2024. 3, 8", + "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 8" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "338", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 8", + "[49] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 1, 8", + "[50] Yifei Yang, Zouying Cao, and Hai Zhao. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187, 2024. 8", + "[51] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. MPLUG-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8", + "[52] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 8", + "[53] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 8", + "[54] Qianhao Yuan, Yanjiang Liu, Yaojie Lu, Hongyu Lin, Ben He, Xianpei Han, and Le Sun. Saisa: Towards multimodal large language models with both training and inference efficiency. arXiv preprint arXiv:2502.02458, 2025. 8", + "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 6", + "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 8", + "[57] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + ], + "bbox": [ + 91, + 90, + 480, + 739 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "339", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_model.json b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ff2e9fa8b5555d1baa5472ec718303501e5b8fe1 --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.203, + 0.782, + 0.24 + ], + "angle": 0, + "content": "Qianhao Yuan\\(^{1,2}\\), Qingyu Zhang\\(^{1,2}\\), Yanjiang Liu\\(^{1,2}\\), Jiawei Chen\\(^{1,2}\\), Yaojie Lu\\(^{1,*}\\), Hongyu Lin\\(^{1}\\), Jia Zheng\\(^{1,*}\\), Xianpei Han\\(^{1}\\), Le Sun\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.239, + 0.71, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Software, Chinese Academy of Sciences \\(^{2}\\)University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.275, + 0.872, + 0.311 + ], + "angle": 0, + "content": "{yuanqianhao2024,zhangqingyu2024,liuyanjiang2021,chenjiawei2024, luyaojie,hongyu,zhengjia,xianpei,sunle}@iscas.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.378, + 0.486, + 0.697 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) suffer from high computational costs due to their massive size and the large number of visual tokens. In this paper, we investigate layer-wise redundancy in MLLMs by introducing a novel metric, Layer Contribution (LC), which quantifies the impact of a layer's transformations on visual and text tokens, respectively. The calculation of LC involves measuring the divergence in model output that results from removing the layer's transformations on the specified tokens. Our pilot experiment reveals that many layers of MLLMs exhibit minimal contribution during the processing of visual tokens. Motivated by this observation, we propose ShortV, a training-free method that leverages LC to identify ineffective layers, and freezes visual token updates in these layers. Experiments show that ShortV can freeze visual token in approximately \\(60\\%\\) of the MLLM layers, thereby dramatically reducing computational costs related to updating visual tokens. For example, it achieves a \\(50\\%\\) reduction in FLOPs on LLaVA-NeXT-13B while maintaining superior performance. The code is publicly available at https://github.com/icip-cas/ShortV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Large language models (LLMs) have achieved remarkable performance in natural language tasks [1, 12, 37, 47, 49]. Building upon LLMs, Multimodal Large Language Models (MLLMs) [31, 32, 38, 39] take a significant step towards understanding the real physical world by incorporating visual information into their processes. Typically, an MLLM consists of a visual encoder, a projector, and an LLM backbone. Most of them preprocess visual information through" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.34, + 0.882, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.45, + 0.74, + 0.462 + ], + "angle": 0, + "content": "(a) ShortV" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.464, + 0.903, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.624, + 0.844, + 0.637 + ], + "angle": 0, + "content": "(b) Performance vs. the Number of ShortV Layers" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.788 + ], + "angle": 0, + "content": "Figure 1. (a) Illustration of ShortV. We identify ineffective layers for visual tokens and replace these layers with sparse ShortV layers. In ShortV layers, we freeze visual tokens, and eliminate computations related to updating them. ShortV improves MLLM efficiency in a training-free manner and involves no parameter updates. Notably, ShortV is compatible with token pruning methods, e.g. FastV. (b) Performance vs. the number of ShortV layers. Average Performance means a normalized average score on multiple benchmarks. ShortV can freeze visual tokens in approximately \\(60\\%\\) of the MLLM layers with nearly no performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.907, + 0.881 + ], + "angle": 0, + "content": "a visual encoder, e.g. a CLIP-ViT [14, 41], and project the patch-level visual features into visual tokens through a projector. Then they concatenate visual and text tokens and feed them into the LLM backbone." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "However, MLLMs face a substantial increase in compu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "329" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tational overhead. This burden primarily stems from the large scale of the LLM backbones and the significantly extended length of the concatenated visual-text token sequences. To address this issue, Chen et al. [10] discovers significant token-wise redundancy in MLLMs. Based on this redundancy, they propose FastV, which identifies and prunes unimportant visual tokens in MLLMs to improve their efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.44 + ], + "angle": 0, + "content": "Other than token-wise redundancy, in this paper, we reveal that MLLMs also exhibit significant layer-wise redundancy in processing visual tokens. Specifically, we propose Layer Contribution (LC), a metric that quantifies how much a layer's transformations on certain tokens contribute to the model's output. In the LC calculation of a layer, we freeze certain tokens in this layer, i.e. keep the hidden states of the tokens unchanged, and then compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original model. This metric provides a direct measure of a layer's importance for certain tokens. By comparing LC scores on visual tokens and those on text tokens, we discover that MLLM layers are ineffective for visual tokens, and their transformations on visual tokens contribute minimal to the model's output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.563 + ], + "angle": 0, + "content": "This phenomenon inspires us to propose ShortV, a simple but effective method to improve the efficiency of MLLMs. In ShortV, we first utilize the LC metric to identify layers least effective at transforming visual tokens, and then replace these layers with sparse ShortV layers. Within these sparse layers, visual tokens remain frozen, and the corresponding computations for updating them are eliminated, as shown in Figure 1a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.776 + ], + "angle": 0, + "content": "To validate the effectiveness of ShortV, we conduct evaluations across multiple benchmarks, including MME [15], MMBench [33], MMMU [55], MMStar [9], SEED-Bench [25], GQA [19], and Flickr30K [40]. Figure 1b illustrates the correlation between the normalized average performance on these benchmarks and the number of replaced layers. As observed, ShortV can replace approximately \\(60\\%\\) of MLLM layers without performance degradation. Unlike FastV and other token pruning methods, ShortV reduces computations of per visual token rather than reducing the number of visual tokens. Therefore, ShortV and token pruning methods are orthogonal and compatible. Furthermore, we demonstrate that combining ShortV and FastV can further enhance MLLM efficiency." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.779, + 0.395, + 0.792 + ], + "angle": 0, + "content": "We summarize our contribution as follows." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- We propose Layer Contribution (LC), a metric to quantify how much a layer's transformations on specific tokens contribute to the model's output." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- Leveraging LC, we reveal significant redundancy in MLLM layers for visual tokens. Transformations on visual tokens in many layers contribute minimally and are thus ineffective." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.796, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Based on the observation above, we propose ShortV, which improves MLLM efficiency by freezing visual tokens in ineffective layers. ShortV can freeze visual tokens in approximately \\(60\\%\\) of MLLM layers without performance degradation. Extensive experiments and ablation studies demonstrate ShortV's effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.796, + 0.212 + ], + "angle": 0, + "content": "2. Layer Redundancy in MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.905, + 0.296 + ], + "angle": 0, + "content": "In this section, we first introduce the background of layer redundancy in text-only LLMs, and then propose a metric to measure MLLM layer redundancy for different types of tokens. Next, we conduct a pilot experiment to investigate layer redundancy in MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.644, + 0.32 + ], + "angle": 0, + "content": "2.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.416 + ], + "angle": 0, + "content": "Typically, MLLMs are built on text-only LLMs. MLLMs employ pre-trained visual encoders, such as CLIP-ViT [14, 41], to convert images into visual features, and then use projectors to project them into visual tokens in the text token embedding space. The visual tokens are concatenated with text tokens and fed into the LLM backbones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.75 + ], + "angle": 0, + "content": "For text-only LLMs, Men et al. [36] identify notable redundancy across their layer. Some layers' transformations on the hidden states of text token contribute minimally to the overall model functionality. Consequently, these layers are considered ineffective. Removing these transformations in approximately \\(25\\%\\) of LLM layers has minimal impact on model outputs. Such redundancy mainly occurs in middle-to-deeper layers, whereas initial layers and the last layer remain critical to the model functionality. However, this pattern may not hold for MLLMs. Huang et al. [18] demonstrate a clear modality gap in the embedding space of current MLLMs, where visual and text tokens exhibit a uniform distribution within each modality but a significant distribution gap between modalities. Such a modality gap implies that MLLMs might adopt distinct computational patterns or strategies for processing visual and text tokens, potentially affecting how redundancy is distributed across layers. This raises several key questions: Are MLLM layers as ineffective for visual tokens as LLM layers are for text tokens? To what extent does layer redundancy exist in MLLMs? How is this redundancy distributed across different MLLM layers?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.757, + 0.773 + ], + "angle": 0, + "content": "2.2. Layer Contribution Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To investigate layer redundancy for certain tokens, we freeze these tokens within the investigated layer, i.e. keep hidden states of these tokens unchanged. To achieve this, we introduce sparse layers shown in Figure 2 for visual and text tokens, respectively. Based on these designs, we propose the Layer Contribution (LC) metric, which evaluates how much a layer's transformations on certain tokens contribute to the model's overall functionality. In the calcu" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "330" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.484, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.195, + 0.483, + 0.232 + ], + "angle": 0, + "content": "(a) Sparse layer where visual tokens are frozen, used to investigate layer redundancy for visual tokens. Only text tokens function as queries and are passed through the FFN. We also denote this layer as ShortV layer." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.234, + 0.372, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.233, + 0.474, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.483, + 0.355 + ], + "angle": 0, + "content": "(b) Sparse layer where text tokens are frozen, used to investigate layer redundancy for text tokens. Only visual tokens function as queries and are passed through the FFN." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.449 + ], + "angle": 0, + "content": "Figure 2. Sparse layers used to investigate layer redundancy for different tokens. To investigate layer redundancy for certain tokens, we freeze these tokens within the layer, i.e. keep hidden states of these tokens unchanged, and measure the divergence between the model's output logits and those of the original model. We gray out the attention that does not need calculation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.536 + ], + "angle": 0, + "content": "lation of LC, we replace the investigated dense layer with the corresponding sparse layer, and compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original dense model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Specifically, We assume the LLM backbone \\(M\\) has \\(L\\) layers, and each of them consists of a self-attention block and a feed-forward network (FFN). The input of the \\(i\\)-th layer at the \\(j\\)-th token position is \\(H_{i}^{j}\\), and the corresponding output is \\(H_{i+1}^{j}\\). The output of the last layer, i.e. the \\(L\\)-th layer, at the same position is \\(H_{L+1}^{j}\\). The model \\(M\\) utilizes \\(H_{L+1}^{-1}\\) at the last token position to compute logits for next-token prediction through the language model head \\(LM_{head}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.679, + 0.483, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\operatorname {l o g i t s} (M) = L M _ {\\text {h e a d}} \\left(H _ {L + 1} ^ {- 1}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To investigate how much the \\(i\\)-th layer's transformations on certain tokens \\(X\\) contribute to the model functionality, we replace the \\(i\\)-th layer with a sparse layer where \\(X\\) are frozen, i.e. \\(X\\)'s hidden states remain unchanged in this sparse layer. The resulting model is denoted as \\(\\mathcal{M}_i^X\\). In practice, \\(X\\) can be visual tokens \\(V\\) or text tokens \\(T\\). As shown in Figure 2, we introduce sparse layers where \\(V\\) and \\(T\\) are frozen, respectively. In Figure 2a, we freeze the visual tokens in the sparse layer. Within the self-attention block of this layer, the visual tokens do not attend to other tokens, and only the text tokens function as queries. For the FFN of this layer, we simply do not pass the visual tokens through it. In Figure 2b, we freeze the text tokens in another sparse" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "layer with similar designs. Based on these, we define the \\(i\\)-th layer's Layer Contribution (LC) score for certain tokens \\(X\\) as the KL divergence between the output logits of the original model \\(M\\) and those of the model \\(\\mathcal{M}_i^X\\) where \\(X\\) are frozen in the \\(i\\)-th layer:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.179, + 0.905, + 0.198 + ], + "angle": 0, + "content": "\\[\nL C _ {i} ^ {X} = K L \\left(\\operatorname {l o g i t s} (M), \\operatorname {l o g i t s} \\left(\\mathcal {M} _ {i} ^ {X}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.905, + 0.269 + ], + "angle": 0, + "content": "here \\(KL(\\cdot)\\) denotes KL divergence. A lower LC score implies that the layer's transformations on the tokens exhibit minimal contribution to the model's output, suggesting that these transformations are ineffective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Discussion: Why not use perplexity or cosine similarity as the metric to measure the importance of layers? Some work in text-only LLMs utilizes perplexity [23, 46] or cosine similarity [11, 17, 36] as metrics to measure the importance of each layer. For the former metric, instead of KL divergence, they measure the change in perplexity of the models, and the layers causing minimal perplexity changes are deemed ineffective. This metric, however, is inadequate when measuring layer redundancy for visual tokens. We find that even if we do not feed visual tokens into the MLLMs, they can still generate reasonable responses, and the changes in perplexity of the MLLMs is relatively low. Nevertheless, they face significant performance degradation in vision language tasks when the visual information is absent. Thus, perplexity is not a reliable measure when evaluating layer redundancy for visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.793 + ], + "angle": 0, + "content": "For the latter metric, the cosine similarity between the input and output of a certain layer is calculated. The hypothesis here is that the ineffective layers have less transformations on the hidden states of tokens, and therefore their inputs and outputs demonstrate higher similarities. However, in our evaluation on LLaVA-1.5 [31] models, the cosine similarity metric and LC differ in their measurement of the redundancy distribution across different layers. Compared with the LC metric, which directly measures the logits divergence between model outputs, cosine similarity consistently overestimates the redundancy of the shallow layers and underestimates the redundancy of the deep layers. We believe the reason behind this difference is that cosine similarity neglects the position of a layer within the model. Specifically, minor transformations of hidden states in the shallow layers can influence all subsequent layers, whereas transformations with similar extent in deeper layers tend to have less impact on overall model functionality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.889, + 0.819 + ], + "angle": 0, + "content": "2.3. Ineffective MLLM Layers for Visual Tokens" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We conduct a pilot experiment to investigate layer redundancy in LLaVA-1.5-7B and LLaVA-1.5-13B [31]. We first randomly sample 2,000 cases from a combination of two major vision language tasks, including caption (Flickr30K [40]) and visual question answering" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.493, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.254, + 0.343, + 0.266 + ], + "angle": 0, + "content": "(a) LLaVA-1.5-7B" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.904, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.254, + 0.758, + 0.266 + ], + "angle": 0, + "content": "(b) LLaVA-1.5-13B" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 3. The Layer Contribution (LC) scores of LLaVA-1.5-7B and LLaVA-1.5-13B. A lower LC score implies that the layer's transformations on the specified tokens are more ineffective. Layers are more ineffective for visual tokens than for text tokens, and freezing visual tokens in ineffective layers results in minimal output divergence from the original model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.347, + 0.483, + 0.408 + ], + "angle": 0, + "content": "(GQA [19]). Then we utilize these samples to calculate each MLLM layer's average LC score for visual and text tokens, respectively. Figure 3 shows the results. We summarize our findings as follows." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.409, + 0.483, + 0.514 + ], + "angle": 0, + "content": "First, for text tokens, middle to deeper layers are more ineffective, while the initial and last layers make more contributions to the MLLM functionality. These observations align with the layer redundancy distribution of text-only LLMs found in Men et al. [36], indicating that visual instruction tuning [30, 31] does not significantly alter the manner LLMs process text tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Second, for visual tokens, the initial and the deep layers, including the last one, exhibit higher redundancy than other layers, which is different from the distribution for text token. Notably, since the last layer's transformations on visual tokens do not contribute to the model's output, its LC score for visual tokens is always 0." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.608, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Third, layer redundancy shows an imbalance between visual and text tokens. Each layer's LC score on visual tokens are lower than that on text tokens, which means that many layers' transformations on visual tokens are ineffective, and freezing visual token in these ineffective layers results in minimal impact on the models' output." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.7, + 0.483, + 0.79 + ], + "angle": 0, + "content": "We attribute the different layer redundancy patterns for different modalities to the modality gap. The clear distribution gap of visual and text tokens result in the difference in how MLLMs process them. We hope these findings can provide insights into how MLLMs process visual and text tokens in different layers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.178, + 0.822 + ], + "angle": 0, + "content": "3. ShortV" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.468, + 0.849 + ], + "angle": 0, + "content": "3.1. Freezing Visual Tokens in Ineffective Layers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As demonstrated in the previous section, we identify significant layer redundancy for visual tokens in MLLMs. Most layers' transformations on visual tokens are ineffective for" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.348, + 0.862, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 4. Details of ShortV layer. In this layer, only text tokens pass through the \\( W_{Q} \\) and \\( W_{O} \\) matrices and the FFN. The attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.669 + ], + "angle": 0, + "content": "the model functionality. Based on this observation, we propose a direct method to enhance MLLM efficiency in a training-free manner: freezing visual tokens in ineffective layers. We denote this method as ShortV." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.807 + ], + "angle": 0, + "content": "In ShortV, we replace the ineffective dense layers with sparse ShortV layers, where visual tokens are frozen. The ShortV layers are same as the sparse layer shown in 2a, and we illustrate their detailed architecture in Figure 4. In each ShortV layer, only text tokens are passed through the FFN and the \\( W_{Q} \\) and \\( W_{O} \\) matrices of the attention block, and the attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "ShortV has one parameter: the number of replaced layers, which we denote as \\( N \\). First, we construct a tiny dataset, which contains a small number of samples from vision language tasks. Then we use this dataset to calculate each layer's average LC score for visual tokens. Next, we sort the layers in ascending order according to LC scores," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "332" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.446 + ], + "angle": 0, + "content": "
MethodTFLOPsFLOPs RatioVQAv2GQASEED-BenchMMMU (val)MMEMMBench ENMMStarFlickr30K CIDEr
LLaVA-1.5-7B
Vanilla8.5100%76.561.966.136.31510.764.133.774.9
FastV (K=2, R=50%)4.958%73.560.265.435.81475.664.332.467.5
VTW (K=16)4.755%66.355.166.236.11497.064.032.844.5
ShortV (Ours, N=19)4.755%75.760.966.236.21503.164.833.371.3
LLaVA-1.5-13B
Vanilla16.6100%78.063.368.235.41531.368.936.179.6
FastV (K=2, R=50%)9.457%76.759.467.834.61506.668.335.973.4
VTW (K=20)9.155%75.360.668.234.91533.068.536.165.9
ShortV (Ours, N=24)9.155%77.262.068.035.81535.968.637.176.4
LLaVA-NeXT-7B
Vanilla42.7100%80.064.170.236.41519.067.137.169.7
FastV (K=2, R=50%)22.052%79.563.069.635.11482.066.336.566.7
VTW (K=16)21.851%75.655.870.235.71518.267.137.638.4
ShortV (Ours, N=19)21.651%78.863.470.436.01525.167.237.865.7
LLaVA-NeXT-13B
Vanilla81.8100%80.965.771.935.91570.069.339.966.7
FastV (K=2, R=50%)42.151%76.862.971.535.91546.468.539.666.0
VTW (K=20)41.751%77.061.571.834.81569.469.139.856.6
ShortV (Ours, N=24)41.050%79.763.671.836.21553.070.239.967.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.456, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Table 1. Comparison of various training-free methods for MLLM efficiency. FLOPs Ratio denotes the proportion of FLOPs retained after applying the corresponding method to improve MLLM efficiency, compared with the vanilla model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.511, + 0.482, + 0.572 + ], + "angle": 0, + "content": "and the layers with lower LC scores are more ineffective for visual tokens. Finally, we freeze visual tokens in the \\(N\\) layers with the lowest LC scores by replacing them with sparse ShortV layers, while keeping their original parameters." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.572, + 0.483, + 0.709 + ], + "angle": 0, + "content": "ShortV is training-free and involves no parameter updates. It can be applied to various MLLMs for different vision language tasks. Notably, ShortV is orthogonal to and compatible with popular visual token pruning methods, e.g. FastV [10]. Visual token pruning directly reduces the number of visual tokens, while ShortV mitigates the computational overhead related to each visual token. This means that we can apply ShortV and token pruning at the same time to further improve MLLM efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.283, + 0.731 + ], + "angle": 0, + "content": "3.2. Computational Cost" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.828 + ], + "angle": 0, + "content": "We consider the computations of self-attention blocks and feed-forward networks (FFNs) in layers of the LLM backbone. Assume \\( t \\) is the number of text tokens, \\( v \\) is the number of visual tokens, \\( h \\) is the hidden state size, \\( m \\) is the intermediate size of the FFNs, the total FLOPs of one dense Transformer layer can be estimated as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.836, + 0.483, + 0.854 + ], + "angle": 0, + "content": "\\[\nF L O P s = 2 (t + v) (4 h + 3 m) h + 4 (t + v) ^ {2} h. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.862, + 0.45, + 0.876 + ], + "angle": 0, + "content": "For one ShortV layer, the FLOPs can be calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nF L O P s ^ {*} = 2 t (4 h + 3 m) h + 4 v h ^ {2} + 4 t (t + v) h. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.907, + 0.603 + ], + "angle": 0, + "content": "For the whole model, assume the LLM has \\(L\\) layers in total, ShortV selects \\(N\\) ineffective dense layers and replaces them with ShortV layers. The FLOPs of the original dense model are \\(L \\times FLOPs\\), and the FLOPs of ShortV are calculated as \\((L - N) \\times FLOPs + N \\times FLOPs^{*}\\). The FLOPs ratio of ShortV and the original model is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.613, + 0.907, + 0.645 + ], + "angle": 0, + "content": "\\[\nr = \\frac {(L - N) \\times F L O P s + N \\times F L O P s ^ {*}}{L \\times F L O P s}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.649, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.712, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Models. To validate the effectiveness of ShortV, we conduct experiments on popular open-source MLLMs, such as LLaVA-1.5-7B [31], LLaVA-1.5-13B, LLaVA-NeXT-7B [32] and LLaVA-NeXT-13B. LLaVA-1.5 models process images with a \\(336 \\times 336\\) resolution and treat each image as 576 tokens. LLaVA-NeXT splits high-resolution images into subimages, and encode the subimages and down-sampled original images independently. This allows the models to scale the input to any arbitrary resolution, without performing positional embedding interpolation for ViTs [14]. LLaVA-NeXT scales the input image resolution to \\(4 \\times\\) and visual token number up to \\(5 \\times\\) compared with LLaVA-1.5, i.e. 2880 tokens for each image." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "333" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.868, + 0.452 + ], + "angle": 0, + "content": "
# ShortV Layers (N)TFLOPsFLOPs RatioMMEMMBench ENMMMU (val)MMStarSEED-BenchGQAFlickr30K CIDErAvg.Per.
LLaVA-1.5-7B (32 layers)
08.5100%1510.764.136.333.766.161.974.958.9100.0
86.981%1508.664.336.033.866.261.474.558.899.8
165.362%1487.064.936.133.365.761.072.858.399.0
194.755%1503.164.836.233.366.260.971.358.399.0
243.744%1341.760.734.133.462.558.364.254.392.2
LLaVA-1.5-13B (40 layers)
016.6100%1531.368.935.436.168.263.379.661.2100.0
814.185%1521.968.635.636.068.263.079.060.999.5
1611.670%1534.968.636.336.268.062.978.561.099.7
249.155%1535.968.635.837.168.062.076.460.799.2
326.640%1298.864.533.636.063.259.368.455.791.0
LLaVA-NeXT-7B (32 layers)
042.7100%1519.067.136.437.170.264.169.760.1100.0
833.879%1515.167.236.636.970.264.170.060.1100.0
1624.958%1476.867.236.237.370.263.567.859.498.8
1921.651%1525.167.236.037.870.463.465.759.599.0
2416.037%1504.165.436.436.068.160.564.958.196.7
LLaVA-NeXT-13B (40 layers)
081.8100%1570.070.535.939.971.965.766.761.3100.0
868.283%1552.470.635.039.671.965.166.961.099.5
1654.667%1561.070.135.039.771.964.866.960.999.3
2441.050%1553.070.236.239.971.863.667.561.099.5
3227.534%1468.465.835.238.969.360.558.557.493.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.462, + 0.907, + 0.491 + ], + "angle": 0, + "content": "Table 2. Performance vs. Efficiency Balance of ShortV under different configurations. # ShortV Layers \\((N)\\) : the number of ShorV layers, Avg.: a normalized average score on the benchmarks, Per.: the relative performance retention compared with the vanilla models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.482, + 0.713 + ], + "angle": 0, + "content": "Baselines. To evaluate the effectiveness of ShortV, which improves MLLM efficiency in a training-free manner, we compare it with popular training-free methods for MLLM efficiency, such as FastV [10] and VTW [29]. FastV drops visual tokens by a percentage of \\( R \\) after the \\( K \\)-th layer in the forward process of input tokens. It computes the average attention score one token received from all other tokens as the importance criterion to select pruned tokens. VTW drops all visual tokens after the \\( K \\)-th layer, enabling only text tokens to engage in the subsequent layers. We use the default settings for the baselines as in their original papers. Specifically, for FastV, \\( K = 2 \\) and \\( R = 50\\% \\). For VTW, \\( K = 16 \\) for 7B models and \\( K = 20 \\) for 13B models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.727, + 0.233, + 0.743 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we conduct experiments to compare ShortV with the baselines. The results are shown in Table 1. We perform evaluation on multiple popular vision language benchmarks, including MME [15], MMBench [33], MMMU (val) [55], MMStar [9], SEED-Bench [25], VQAv2 [16], and GQA [19]. We manually choose the number of ShortV layers \\(N\\) to maintain a similar or lower FLOPs ratio compared with the baselines. The FLOPs ratio of ShortV is calculated according to Equation 5, where we set the number of text tokens to 64. Specifically, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.577 + ], + "angle": 0, + "content": "choose \\(N = 19\\) for the 7B models, and \\(N = 24\\) for the 13B models. As shown in Table 1, our ShortV achieves comparable or superior performance across multiple benchmarks compared with the baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.896, + 0.607 + ], + "angle": 0, + "content": "4.3. Balance between Efficiency and Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.704 + ], + "angle": 0, + "content": "In this section, we conduct an experiment to investigate the impact of ShortV's parameter \\( N \\), which denotes the number of ShortV layers. The experimental results are presented in Table 2. To facilitate intuitive comprehension, we plot the correlation between the normalized average score on benchmarks and the number of ShortV layers in Figure 1b." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We observe that ShortV can freeze visual tokens in approximately \\(60\\%\\) of the MLLM layers while preserving superior performance. As \\(N\\) continues to increase, both 7B and 13B models can maintain more than \\(90\\%\\) performance when the hidden states of visual tokens remain unchanged in about \\(80\\%\\) of the layers. These results are significantly different from those on text-only LLMs. For LLMs, Men et al. [36] remove transformations on text tokens in approximately \\(25\\%\\) of the LLM layers, and this results in about \\(10\\%\\) performance degradation on language benchmarks. As the number of layers increases, the performance of LLMs rapidly declines. These differences align with our observation in Section 2.3 that layers are more ineffective for visual" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "334" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.141 + ], + "angle": 0, + "content": "
# ShortV Layers (N)08161924
LLaVA-1.5-7B1.00×1.13×1.23×1.30×1.40×
LLaVA-NeXT-7B1.00×1.15×1.35×1.44×1.64×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.151, + 0.483, + 0.18 + ], + "angle": 0, + "content": "Table 3. Inference speedups over the vanilla models, based on the 7B models. We conduct this test on a single GPU." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.202, + 0.481, + 0.255 + ], + "angle": 0, + "content": "
# ShortV Layers (N)08162432
LLaVA-1.5-13B1.00×1.13×1.24×1.39×1.50×
LLaVA-NeXT-13B1.00×1.13×1.30×1.52×1.84×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.265, + 0.483, + 0.294 + ], + "angle": 0, + "content": "Table 4. Inference speedups over the vanilla models, based on the 13B models. We conduct this test on a single GPU." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.316, + 0.481, + 0.411 + ], + "angle": 0, + "content": "
MethodFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
FastV58%64.335.865.460.2
ShortV55%64.836.266.260.9
ShortV+FastV29%64.237.165.159.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.421, + 0.483, + 0.462 + ], + "angle": 0, + "content": "Table 5. ShortV is compatible with FastV, and applying both at the same time can further enhance MLLM efficiency. This experiment is based on LLaVA-1.5-7B." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.5, + 0.273, + 0.513 + ], + "angle": 0, + "content": "tokens than for text tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.671 + ], + "angle": 0, + "content": "In addition to the theoretical FLOPs ratios, we provide the speedups on real hardware using different settings, as shown in Table 3 and Table 4. To get rid of the influence of different output sequence lengths, we use the first token latency to calculate the speedups. We utilize the MMMU dataset for the latency test. For comparison, we note that FastV [10] with its default setting, i.e. \\( K = 2 \\) and \\( R = 50\\% \\), achieves a \\( 1.31 \\times \\) speedup over the vanilla LLaVA-1.5-13B model. In contrast, our ShortV with its default parameter, i.e. \\( N = 24 \\), achieves a greater speedup of \\( 1.39 \\times \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.356, + 0.71 + ], + "angle": 0, + "content": "4.4. Orthogonal to Token Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we demonstrate that ShortV is orthogonal to and compatible with visual token pruning, e.g. FastV [10]. FastV identifies \\( R\\% \\) unimportant visual tokens and drops them after the \\( K \\)-th layer in the forward process of input tokens. We apply FastV to ShortV, which already replaces \\( N \\) ineffective layers for visual tokens with ShortV layers. We use the default settings for FastV and ShortV in this experiment, i.e. \\( K = 2 \\) and \\( R = 50\\% \\) for FastV, and \\( N = 19 \\) for ShortV. We employ LLaVA-1.5-7B as the vanilla model. The experimental results in Table 5 demonstrate that ShortV is compatible with FastV and that the application of both can further improve MLLM efficiency." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
StrategyFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
Random55%58.433.660.556.1
Cosine Sim.55%60.834.262.759.5
LC (Ours)55%64.836.266.260.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.197, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Table 6. Ablation on strategies to select replaced layers, based on LLaVA-1.5-7B. \"Random\" denotes randomly selecting 19 layers and freezing visual tokens in them. \"Cosine Sim.\" denotes using cosine similarity to select ineffective layers for visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.281, + 0.679, + 0.295 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.907, + 0.695 + ], + "angle": 0, + "content": "Ablation on strategies to select replaced layers. In this paragraph, we perform an ablation experiment on LLaVA-1.5-7B to investigate the impact of strategies for selecting which layers to replace. ShortV selects ineffective layers for visual tokens, and replace them with ShortV layers. To identify which layers are ineffective, we utilize the LC metric introduced in Section 2. In contrast, previous work [36] on text-only LLMs uses a metric based on cosine similarity. It calculates the average cosine similarity between the inputs and outputs of each layer. The layers with higher cosine similarities are deemed more ineffective. To make a comparison between this cosine similarity metric and our LC metric, we calculate each layer's cosine similarity between the input hidden states and output hidden states of visual tokens, and select the same number of layers with the highest cosine similarities. We show the comparison in Table 6. We also include the results of another baseline, ShortV (Random), where visual tokens are frozen in the same number of randomly selected layers. These results clearly demonstrate that our LC metric performs better than cosine similarity in identifying ineffective MLLM layers for visual tokens, and ShortV based on the LC metric achieves performance comparable to the vanilla model. In contrast, ShortV based on the cosine similarity metric cannot match the performance of the vanilla model, although it outperforms the baseline with randomly selected layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Ablation on frozen tokens. In this paragraph, we conduct an ablation study on LLaVA-1.5-7B to investigate the impact of freezing different types of tokens. In Section 2, we demonstrate that MLLM layers are ineffective for visual tokens, as measured by the LC metric. Motivated by this observation, ShortV freezes visual tokens in ineffective layers. In Table 7, we compare our method with the strategies of freezing other tokens. In the experiment detailed in line (a), we utilize the LC metric to identify 19 ineffective layers for text tokens, and freeze text tokens in them. Despite having fewer frozen tokens, we can observe that this strategy results in significant performance declines compared with our method, which freezes visual tokens rather than text to" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "335" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.457, + 0.193 + ], + "angle": 0, + "content": "
Frozen TokensMMBench ENMMMU (val)SEED-BenchGQA
None (Vanilla)64.036.366.161.9
(a) Text2.123.78.92.9
(b) Text+Visual1.326.60.80.0
(c) Random1.522.95.52.3
(d) Visual (Ours)64.836.266.260.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.203, + 0.483, + 0.286 + ], + "angle": 0, + "content": "Table 7. Ablation on frozen tokens, based on LLaVA-1.5-7B. (a) identifying 19 ineffective layers for text tokens and freezing text tokens in them. In lines (b) and (c), we select ineffective layers for all tokens. line (b) involves freezing all input tokens in them, whereas line (c) denotes randomly freezing the same number of tokens as the visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.483, + 0.523 + ], + "angle": 0, + "content": "kens. These experimental results align with our findings in Section 2 that MLLM layers are more ineffective for visual tokens than for text tokens. In lines (b) and (c), we first calculate each layer's average LC score for all tokens, including visual and text tokens, and then select 19 ineffective layers. Next, in the experiment corresponding to line (b), we freeze all input tokens in these layers. In line (c), we freeze random input tokens, and the number of frozen tokens matches that of the visual tokens. As a result, the computational overhead associated with line (c) is the same as that of our method in line (d). We can find that freezing tokens other than visual tokens leads to substantial performance degradation in vision-language tasks. These ablations demonstrate the effectiveness of our ShortV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.233, + 0.552 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.409, + 0.578 + ], + "angle": 0, + "content": "5.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Built upon Large Language Models (LLMs) [1, 12, 37, 47, 49], Multimodal Large Language Models (MLLMs) [4, 7, 35, 48, 52, 57] have made significant progress in processing and understanding the visual world. Typically, they use a decoder-only architecture. Specifically, they utilize visual encoders [14, 21, 34, 41, 56] to convert input visual information into visual features and then use projectors to project these visual features into visual tokens. These visual tokens are then concatenated with text tokens and fed into the LLM backbones. Current MLLMs use hundreds to thousands of visual tokens to represent a single image, significantly increasing the length of the token sequences. For instance, the LLaVA-1.5 models [31] transform each image with \\(336 \\times 336\\) resolution into 576 tokens. For images with higher resolutions, the LLaVA-NeXT models [32] process images into up to 2,880 visual tokens, and SPHINX-2k [28] divides one image into nine subimages, resulting in 2,890 visual tokens. Applying LLMs to such large numbers of visual tokens incurs substantial computational costs. In this paper, we introduce ShortV to enhance the efficiency of MLLMs by reducing the computational overhead associ" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.672, + 0.106 + ], + "angle": 0, + "content": "ated with visual tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.12, + 0.771, + 0.135 + ], + "angle": 0, + "content": "5.2. Efficient LLMs and MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.144, + 0.905, + 0.219 + ], + "angle": 0, + "content": "For LLMs, previous studies [36, 46] find that layers in LLMs are ineffective for text tokens. They remove computations in about \\(25\\%\\) of the layers, while preserving approximately \\(90\\%\\) of the performance. LaCo [50] utilizes layer merging for efficient LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.905, + 0.476 + ], + "angle": 0, + "content": "To address the computational inefficiency of MLLMs, previous methods [10, 29, 42, 44, 45, 54] have primarily focused on two aspects: efficient model architecture and visual token compression. Among efficient model architectures, cross-attention-based models [2, 3, 8, 20] insert gated cross-attention layers within LLM layers for visual perception, but previous studies [13, 24] demonstrate that this architecture performs worse than the decoder-only architecture in the same settings. Instead of inserting cross-attention layers, mPLUG-Owl3 [51] and Vamba [42] introduce cross-attention operations in parallel with self-attention. In contrast, SAISA [54] introduces NAAViT self-attention blocks, which incorporate multimodal cross-attention into the original self-attention operations of the LLMs, and reuse the parameters of self-attention blocks. The design of ShortV layers is inspired by NAAViT. Differently, in ShortV layers, visual tokens also skip their FFNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.905, + 0.706 + ], + "angle": 0, + "content": "Visual token compression methods improve MLLM efficiency in both training-based [5, 6, 22, 26, 27] and training-free [10, 43, 53] manners. FastV [10] reveals token-wise redundancy, and it removes unimportant tokens during inference. In this paper, we reveal layer-wise redundancy in MLLMs. Layers in MLLMs are much more ineffective for visual tokens than for text token. Therefore, we can freeze visual tokens in approximately \\(60\\%\\) of the MLLM layers with minimal performance degradation. Unlike previous methods for MLLM efficiency, ShortV does not reduce the number of visual tokens but instead decreases the computational costs of processing each token. ShortV is training-free and orthogonal to token compression. We demonstrate that ShortV is compatible with FastV, allowing for simultaneous application to further enhance MLLM efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we explore the layer-wise redundancy in MLLMs. We discover that layers in MLLMs are more ineffective for visual tokens than for text tokens. MLLM layers' transformations on visual tokens have a minimal impact on the MLLM output. Motivated by this observation, we propose ShortV, a training-free method to enhance MLLM efficiency. ShortV utilizes our proposed LC metric to select ineffective layers for visual tokens, and freezes visual tokens in these layers. It can freeze visual tokens in about \\(60\\%\\) of the layers while preserving superior performance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "336" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.262, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.213 + ], + "angle": 0, + "content": "We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265) and the Basic Research Program of ISCAS (Grant No. ISCAS-ZD-202401)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.238, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.263, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.611, + 0.483, + 0.679 + ], + "angle": 0, + "content": "[6] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[8] Kaibing Chen, Dong Shen, Hanwen Zhong, Huasong Zhong, Kui Xia, Di Xu, Wei Yuan, Yifei Hu, Bin Wen, Tianke Zhang, et al. Evlm: An efficient vision-language model for visual understanding. arXiv preprint arXiv:2407.14177, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.263, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[11] Xiaodong Chen, Yuxuan Hu, Jing Zhang, Yanling Wang, Cuiping Li, and Hong Chen. Streamlining redundant layers to compress large language models. arXiv preprint arXiv:2403.19135, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[13] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.904, + 0.53 + ], + "angle": 0, + "content": "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[17] Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. What matters in transformers? not all attention is needed. arXiv preprint arXiv:2406.15786, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[18] Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Deciphering cross-modal alignment in large vision-language models with modality integration rate. arXiv preprint arXiv:2410.07167, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[20] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "337" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[22] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[23] Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim, Thibault Castells, Shinkook Choi, Junho Shin, and Hyoung-Kyu Song. Shortened llama: Depth pruning for large language models with comparison of retraining methods. arXiv preprint arXiv:2402.02834, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[25] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[26] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[27] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[28] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[29] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[33] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[35] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[36] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect, 2024. URL https://arxiv.org/abs/2403.03853, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[37] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2(5):6, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[38] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[39] OpenAI. Introducing gpt-4o: our fastest and most affordable flagship model. https://platform.openai.com/docs/guides/vision, 2024. Accessed: 2024-05-26. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[42] Weiming Ren, Wentao Ma, Huan Yang, Cong Wei, Ge Zhang, and Wenhu Chen. Vamba: Understanding hour-long videos with hybrid mamba-transformers. arXiv preprint arXiv:2503.11579, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[43] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[44] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[45] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[46] Jiwon Song, Kyungseok Oh, Taesu Kim, Hyungjun Kim, Yulhwa Kim, and Jae-Joon Kim. Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks. arXiv preprint arXiv:2402.09025, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "338" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[49] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[50] Yifei Yang, Zouying Cao, and Hai Zhao. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[51] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. MPLUG-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[52] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[53] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[54] Qianhao Yuan, Yanjiang Liu, Yaojie Lu, Hongyu Lin, Ben He, Xianpei Han, and Le Sun. Saisa: Towards multimodal large language models with both training and inference efficiency. arXiv preprint arXiv:2502.02458, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.685, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[57] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "339" + } + ] +] \ No newline at end of file diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_origin.pdf b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5cab64f44b93201dcc536a8c2a73e81c65149d4c --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/213fbd78-7b79-438a-a9fe-b3e13e7f59de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed513f5295680b3797959e821f125df5da58e4af1ff3200abdb5b170caba2568 +size 1625978 diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/full.md b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b40467b47379b1aa5114fd8eacecfc45a509b47f --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/full.md @@ -0,0 +1,298 @@ +# ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers + +Qianhao Yuan $^{1,2}$ , Qingyu Zhang $^{1,2}$ , Yanjiang Liu $^{1,2}$ , Jiawei Chen $^{1,2}$ , Yaojie Lu $^{1,*}$ , Hongyu Lin $^{1}$ , Jia Zheng $^{1,*}$ , Xianpei Han $^{1}$ , Le Sun $^{1}$ + +$^{1}$ Institute of Software, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences + +{yuanqianhao2024,zhangqingyu2024,liuyanjiang2021,chenjiawei2024, luyaojie,hongyu,zhengjia,xianpei,sunle}@iscas.ac.cn + +# Abstract + +Multimodal Large Language Models (MLLMs) suffer from high computational costs due to their massive size and the large number of visual tokens. In this paper, we investigate layer-wise redundancy in MLLMs by introducing a novel metric, Layer Contribution (LC), which quantifies the impact of a layer's transformations on visual and text tokens, respectively. The calculation of LC involves measuring the divergence in model output that results from removing the layer's transformations on the specified tokens. Our pilot experiment reveals that many layers of MLLMs exhibit minimal contribution during the processing of visual tokens. Motivated by this observation, we propose ShortV, a training-free method that leverages LC to identify ineffective layers, and freezes visual token updates in these layers. Experiments show that ShortV can freeze visual token in approximately $60\%$ of the MLLM layers, thereby dramatically reducing computational costs related to updating visual tokens. For example, it achieves a $50\%$ reduction in FLOPs on LLaVA-NeXT-13B while maintaining superior performance. The code is publicly available at https://github.com/icip-cas/ShortV. + +# 1. Introduction + +Large language models (LLMs) have achieved remarkable performance in natural language tasks [1, 12, 37, 47, 49]. Building upon LLMs, Multimodal Large Language Models (MLLMs) [31, 32, 38, 39] take a significant step towards understanding the real physical world by incorporating visual information into their processes. Typically, an MLLM consists of a visual encoder, a projector, and an LLM backbone. Most of them preprocess visual information through + +![](images/66b6621944e119e1e75bbc22db6ecbb29532ea43e4bb6380c238b633a9345249.jpg) +(a) ShortV + +![](images/635f725d3c85e7cae0aab88e6316eec731d238d26f4ac361155a5e70a2063179.jpg) +(b) Performance vs. the Number of ShortV Layers +Figure 1. (a) Illustration of ShortV. We identify ineffective layers for visual tokens and replace these layers with sparse ShortV layers. In ShortV layers, we freeze visual tokens, and eliminate computations related to updating them. ShortV improves MLLM efficiency in a training-free manner and involves no parameter updates. Notably, ShortV is compatible with token pruning methods, e.g. FastV. (b) Performance vs. the number of ShortV layers. Average Performance means a normalized average score on multiple benchmarks. ShortV can freeze visual tokens in approximately $60\%$ of the MLLM layers with nearly no performance degradation. + +a visual encoder, e.g. a CLIP-ViT [14, 41], and project the patch-level visual features into visual tokens through a projector. Then they concatenate visual and text tokens and feed them into the LLM backbone. + +However, MLLMs face a substantial increase in compu + +tational overhead. This burden primarily stems from the large scale of the LLM backbones and the significantly extended length of the concatenated visual-text token sequences. To address this issue, Chen et al. [10] discovers significant token-wise redundancy in MLLMs. Based on this redundancy, they propose FastV, which identifies and prunes unimportant visual tokens in MLLMs to improve their efficiency. + +Other than token-wise redundancy, in this paper, we reveal that MLLMs also exhibit significant layer-wise redundancy in processing visual tokens. Specifically, we propose Layer Contribution (LC), a metric that quantifies how much a layer's transformations on certain tokens contribute to the model's output. In the LC calculation of a layer, we freeze certain tokens in this layer, i.e. keep the hidden states of the tokens unchanged, and then compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original model. This metric provides a direct measure of a layer's importance for certain tokens. By comparing LC scores on visual tokens and those on text tokens, we discover that MLLM layers are ineffective for visual tokens, and their transformations on visual tokens contribute minimal to the model's output. + +This phenomenon inspires us to propose ShortV, a simple but effective method to improve the efficiency of MLLMs. In ShortV, we first utilize the LC metric to identify layers least effective at transforming visual tokens, and then replace these layers with sparse ShortV layers. Within these sparse layers, visual tokens remain frozen, and the corresponding computations for updating them are eliminated, as shown in Figure 1a. + +To validate the effectiveness of ShortV, we conduct evaluations across multiple benchmarks, including MME [15], MMBench [33], MMMU [55], MMStar [9], SEED-Bench [25], GQA [19], and Flickr30K [40]. Figure 1b illustrates the correlation between the normalized average performance on these benchmarks and the number of replaced layers. As observed, ShortV can replace approximately $60\%$ of MLLM layers without performance degradation. Unlike FastV and other token pruning methods, ShortV reduces computations of per visual token rather than reducing the number of visual tokens. Therefore, ShortV and token pruning methods are orthogonal and compatible. Furthermore, we demonstrate that combining ShortV and FastV can further enhance MLLM efficiency. + +We summarize our contribution as follows. + +- We propose Layer Contribution (LC), a metric to quantify how much a layer's transformations on specific tokens contribute to the model's output. +- Leveraging LC, we reveal significant redundancy in MLLM layers for visual tokens. Transformations on visual tokens in many layers contribute minimally and are thus ineffective. + +- Based on the observation above, we propose ShortV, which improves MLLM efficiency by freezing visual tokens in ineffective layers. ShortV can freeze visual tokens in approximately $60\%$ of MLLM layers without performance degradation. Extensive experiments and ablation studies demonstrate ShortV's effectiveness. + +# 2. Layer Redundancy in MLLMs + +In this section, we first introduce the background of layer redundancy in text-only LLMs, and then propose a metric to measure MLLM layer redundancy for different types of tokens. Next, we conduct a pilot experiment to investigate layer redundancy in MLLMs. + +# 2.1. Background + +Typically, MLLMs are built on text-only LLMs. MLLMs employ pre-trained visual encoders, such as CLIP-ViT [14, 41], to convert images into visual features, and then use projectors to project them into visual tokens in the text token embedding space. The visual tokens are concatenated with text tokens and fed into the LLM backbones. + +For text-only LLMs, Men et al. [36] identify notable redundancy across their layer. Some layers' transformations on the hidden states of text token contribute minimally to the overall model functionality. Consequently, these layers are considered ineffective. Removing these transformations in approximately $25\%$ of LLM layers has minimal impact on model outputs. Such redundancy mainly occurs in middle-to-deeper layers, whereas initial layers and the last layer remain critical to the model functionality. However, this pattern may not hold for MLLMs. Huang et al. [18] demonstrate a clear modality gap in the embedding space of current MLLMs, where visual and text tokens exhibit a uniform distribution within each modality but a significant distribution gap between modalities. Such a modality gap implies that MLLMs might adopt distinct computational patterns or strategies for processing visual and text tokens, potentially affecting how redundancy is distributed across layers. This raises several key questions: Are MLLM layers as ineffective for visual tokens as LLM layers are for text tokens? To what extent does layer redundancy exist in MLLMs? How is this redundancy distributed across different MLLM layers? + +# 2.2. Layer Contribution Metric + +To investigate layer redundancy for certain tokens, we freeze these tokens within the investigated layer, i.e. keep hidden states of these tokens unchanged. To achieve this, we introduce sparse layers shown in Figure 2 for visual and text tokens, respectively. Based on these designs, we propose the Layer Contribution (LC) metric, which evaluates how much a layer's transformations on certain tokens contribute to the model's overall functionality. In the calcu + +![](images/44a32d4ecee0764438b174d8aa73406d7836394a5427d968f4a082df6e8b35ef.jpg) +(a) Sparse layer where visual tokens are frozen, used to investigate layer redundancy for visual tokens. Only text tokens function as queries and are passed through the FFN. We also denote this layer as ShortV layer. + +![](images/605d72357eb987a089e455802200c69111b29563f3df670ac362d840f83f468e.jpg) +(b) Sparse layer where text tokens are frozen, used to investigate layer redundancy for text tokens. Only visual tokens function as queries and are passed through the FFN. + +![](images/9525adcce18eef1a74c836efa284244c85b3cfd37fd441d513d3309fb0e6a00f.jpg) +Figure 2. Sparse layers used to investigate layer redundancy for different tokens. To investigate layer redundancy for certain tokens, we freeze these tokens within the layer, i.e. keep hidden states of these tokens unchanged, and measure the divergence between the model's output logits and those of the original model. We gray out the attention that does not need calculation. + +lation of LC, we replace the investigated dense layer with the corresponding sparse layer, and compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original dense model. + +Specifically, We assume the LLM backbone $M$ has $L$ layers, and each of them consists of a self-attention block and a feed-forward network (FFN). The input of the $i$ -th layer at the $j$ -th token position is $H_{i}^{j}$ , and the corresponding output is $H_{i+1}^{j}$ . The output of the last layer, i.e. the $L$ -th layer, at the same position is $H_{L+1}^{j}$ . The model $M$ utilizes $H_{L+1}^{-1}$ at the last token position to compute logits for next-token prediction through the language model head $LM_{head}$ : + +$$ +\operatorname {l o g i t s} (M) = L M _ {\text {h e a d}} \left(H _ {L + 1} ^ {- 1}\right). \tag {1} +$$ + +To investigate how much the $i$ -th layer's transformations on certain tokens $X$ contribute to the model functionality, we replace the $i$ -th layer with a sparse layer where $X$ are frozen, i.e. $X$ 's hidden states remain unchanged in this sparse layer. The resulting model is denoted as $\mathcal{M}_i^X$ . In practice, $X$ can be visual tokens $V$ or text tokens $T$ . As shown in Figure 2, we introduce sparse layers where $V$ and $T$ are frozen, respectively. In Figure 2a, we freeze the visual tokens in the sparse layer. Within the self-attention block of this layer, the visual tokens do not attend to other tokens, and only the text tokens function as queries. For the FFN of this layer, we simply do not pass the visual tokens through it. In Figure 2b, we freeze the text tokens in another sparse + +layer with similar designs. Based on these, we define the $i$ -th layer's Layer Contribution (LC) score for certain tokens $X$ as the KL divergence between the output logits of the original model $M$ and those of the model $\mathcal{M}_i^X$ where $X$ are frozen in the $i$ -th layer: + +$$ +L C _ {i} ^ {X} = K L \left(\operatorname {l o g i t s} (M), \operatorname {l o g i t s} \left(\mathcal {M} _ {i} ^ {X}\right)\right), \tag {2} +$$ + +here $KL(\cdot)$ denotes KL divergence. A lower LC score implies that the layer's transformations on the tokens exhibit minimal contribution to the model's output, suggesting that these transformations are ineffective. + +Discussion: Why not use perplexity or cosine similarity as the metric to measure the importance of layers? Some work in text-only LLMs utilizes perplexity [23, 46] or cosine similarity [11, 17, 36] as metrics to measure the importance of each layer. For the former metric, instead of KL divergence, they measure the change in perplexity of the models, and the layers causing minimal perplexity changes are deemed ineffective. This metric, however, is inadequate when measuring layer redundancy for visual tokens. We find that even if we do not feed visual tokens into the MLLMs, they can still generate reasonable responses, and the changes in perplexity of the MLLMs is relatively low. Nevertheless, they face significant performance degradation in vision language tasks when the visual information is absent. Thus, perplexity is not a reliable measure when evaluating layer redundancy for visual tokens. + +For the latter metric, the cosine similarity between the input and output of a certain layer is calculated. The hypothesis here is that the ineffective layers have less transformations on the hidden states of tokens, and therefore their inputs and outputs demonstrate higher similarities. However, in our evaluation on LLaVA-1.5 [31] models, the cosine similarity metric and LC differ in their measurement of the redundancy distribution across different layers. Compared with the LC metric, which directly measures the logits divergence between model outputs, cosine similarity consistently overestimates the redundancy of the shallow layers and underestimates the redundancy of the deep layers. We believe the reason behind this difference is that cosine similarity neglects the position of a layer within the model. Specifically, minor transformations of hidden states in the shallow layers can influence all subsequent layers, whereas transformations with similar extent in deeper layers tend to have less impact on overall model functionality. + +# 2.3. Ineffective MLLM Layers for Visual Tokens + +We conduct a pilot experiment to investigate layer redundancy in LLaVA-1.5-7B and LLaVA-1.5-13B [31]. We first randomly sample 2,000 cases from a combination of two major vision language tasks, including caption (Flickr30K [40]) and visual question answering + +![](images/50ae17a45224c197973a813145c51a42819f3f150e8116e07b9580967f191d5b.jpg) +(a) LLaVA-1.5-7B + +![](images/fb94ad90f8c6e4bbae719f773e9f78a260782cdc9741d3a5273d4714abcf1b61.jpg) +(b) LLaVA-1.5-13B + +(GQA [19]). Then we utilize these samples to calculate each MLLM layer's average LC score for visual and text tokens, respectively. Figure 3 shows the results. We summarize our findings as follows. + +First, for text tokens, middle to deeper layers are more ineffective, while the initial and last layers make more contributions to the MLLM functionality. These observations align with the layer redundancy distribution of text-only LLMs found in Men et al. [36], indicating that visual instruction tuning [30, 31] does not significantly alter the manner LLMs process text tokens. + +Second, for visual tokens, the initial and the deep layers, including the last one, exhibit higher redundancy than other layers, which is different from the distribution for text token. Notably, since the last layer's transformations on visual tokens do not contribute to the model's output, its LC score for visual tokens is always 0. + +Third, layer redundancy shows an imbalance between visual and text tokens. Each layer's LC score on visual tokens are lower than that on text tokens, which means that many layers' transformations on visual tokens are ineffective, and freezing visual token in these ineffective layers results in minimal impact on the models' output. + +We attribute the different layer redundancy patterns for different modalities to the modality gap. The clear distribution gap of visual and text tokens result in the difference in how MLLMs process them. We hope these findings can provide insights into how MLLMs process visual and text tokens in different layers. + +# 3. ShortV + +# 3.1. Freezing Visual Tokens in Ineffective Layers + +As demonstrated in the previous section, we identify significant layer redundancy for visual tokens in MLLMs. Most layers' transformations on visual tokens are ineffective for + +![](images/1c34b6b57f6727d65239d91bf26949f6695bdd211480c8d59101295abf334f00.jpg) +Figure 3. The Layer Contribution (LC) scores of LLaVA-1.5-7B and LLaVA-1.5-13B. A lower LC score implies that the layer's transformations on the specified tokens are more ineffective. Layers are more ineffective for visual tokens than for text tokens, and freezing visual tokens in ineffective layers results in minimal output divergence from the original model. +Figure 4. Details of ShortV layer. In this layer, only text tokens pass through the $W_{Q}$ and $W_{O}$ matrices and the FFN. The attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries. + +the model functionality. Based on this observation, we propose a direct method to enhance MLLM efficiency in a training-free manner: freezing visual tokens in ineffective layers. We denote this method as ShortV. + +In ShortV, we replace the ineffective dense layers with sparse ShortV layers, where visual tokens are frozen. The ShortV layers are same as the sparse layer shown in 2a, and we illustrate their detailed architecture in Figure 4. In each ShortV layer, only text tokens are passed through the FFN and the $W_{Q}$ and $W_{O}$ matrices of the attention block, and the attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries. + +ShortV has one parameter: the number of replaced layers, which we denote as $N$ . First, we construct a tiny dataset, which contains a small number of samples from vision language tasks. Then we use this dataset to calculate each layer's average LC score for visual tokens. Next, we sort the layers in ascending order according to LC scores, + +
MethodTFLOPsFLOPs RatioVQAv2GQASEED-BenchMMMU (val)MMEMMBench ENMMStarFlickr30K CIDEr
LLaVA-1.5-7B
Vanilla8.5100%76.561.966.136.31510.764.133.774.9
FastV (K=2, R=50%)4.958%73.560.265.435.81475.664.332.467.5
VTW (K=16)4.755%66.355.166.236.11497.064.032.844.5
ShortV (Ours, N=19)4.755%75.760.966.236.21503.164.833.371.3
LLaVA-1.5-13B
Vanilla16.6100%78.063.368.235.41531.368.936.179.6
FastV (K=2, R=50%)9.457%76.759.467.834.61506.668.335.973.4
VTW (K=20)9.155%75.360.668.234.91533.068.536.165.9
ShortV (Ours, N=24)9.155%77.262.068.035.81535.968.637.176.4
LLaVA-NeXT-7B
Vanilla42.7100%80.064.170.236.41519.067.137.169.7
FastV (K=2, R=50%)22.052%79.563.069.635.11482.066.336.566.7
VTW (K=16)21.851%75.655.870.235.71518.267.137.638.4
ShortV (Ours, N=19)21.651%78.863.470.436.01525.167.237.865.7
LLaVA-NeXT-13B
Vanilla81.8100%80.965.771.935.91570.069.339.966.7
FastV (K=2, R=50%)42.151%76.862.971.535.91546.468.539.666.0
VTW (K=20)41.751%77.061.571.834.81569.469.139.856.6
ShortV (Ours, N=24)41.050%79.763.671.836.21553.070.239.967.5
+ +Table 1. Comparison of various training-free methods for MLLM efficiency. FLOPs Ratio denotes the proportion of FLOPs retained after applying the corresponding method to improve MLLM efficiency, compared with the vanilla model. + +and the layers with lower LC scores are more ineffective for visual tokens. Finally, we freeze visual tokens in the $N$ layers with the lowest LC scores by replacing them with sparse ShortV layers, while keeping their original parameters. + +ShortV is training-free and involves no parameter updates. It can be applied to various MLLMs for different vision language tasks. Notably, ShortV is orthogonal to and compatible with popular visual token pruning methods, e.g. FastV [10]. Visual token pruning directly reduces the number of visual tokens, while ShortV mitigates the computational overhead related to each visual token. This means that we can apply ShortV and token pruning at the same time to further improve MLLM efficiency. + +# 3.2. Computational Cost + +We consider the computations of self-attention blocks and feed-forward networks (FFNs) in layers of the LLM backbone. Assume $t$ is the number of text tokens, $v$ is the number of visual tokens, $h$ is the hidden state size, $m$ is the intermediate size of the FFNs, the total FLOPs of one dense Transformer layer can be estimated as: + +$$ +F L O P s = 2 (t + v) (4 h + 3 m) h + 4 (t + v) ^ {2} h. \tag {3} +$$ + +For one ShortV layer, the FLOPs can be calculated as: + +$$ +F L O P s ^ {*} = 2 t (4 h + 3 m) h + 4 v h ^ {2} + 4 t (t + v) h. \tag {4} +$$ + +For the whole model, assume the LLM has $L$ layers in total, ShortV selects $N$ ineffective dense layers and replaces them with ShortV layers. The FLOPs of the original dense model are $L \times FLOPs$ , and the FLOPs of ShortV are calculated as $(L - N) \times FLOPs + N \times FLOPs^{*}$ . The FLOPs ratio of ShortV and the original model is computed as: + +$$ +r = \frac {(L - N) \times F L O P s + N \times F L O P s ^ {*}}{L \times F L O P s}. \tag {5} +$$ + +# 4. Experiments + +# 4.1. Experimental Setups + +Models. To validate the effectiveness of ShortV, we conduct experiments on popular open-source MLLMs, such as LLaVA-1.5-7B [31], LLaVA-1.5-13B, LLaVA-NeXT-7B [32] and LLaVA-NeXT-13B. LLaVA-1.5 models process images with a $336 \times 336$ resolution and treat each image as 576 tokens. LLaVA-NeXT splits high-resolution images into subimages, and encode the subimages and down-sampled original images independently. This allows the models to scale the input to any arbitrary resolution, without performing positional embedding interpolation for ViTs [14]. LLaVA-NeXT scales the input image resolution to $4 \times$ and visual token number up to $5 \times$ compared with LLaVA-1.5, i.e. 2880 tokens for each image. + +
# ShortV Layers (N)TFLOPsFLOPs RatioMMEMMBench ENMMMU (val)MMStarSEED-BenchGQAFlickr30K CIDErAvg.Per.
LLaVA-1.5-7B (32 layers)
08.5100%1510.764.136.333.766.161.974.958.9100.0
86.981%1508.664.336.033.866.261.474.558.899.8
165.362%1487.064.936.133.365.761.072.858.399.0
194.755%1503.164.836.233.366.260.971.358.399.0
243.744%1341.760.734.133.462.558.364.254.392.2
LLaVA-1.5-13B (40 layers)
016.6100%1531.368.935.436.168.263.379.661.2100.0
814.185%1521.968.635.636.068.263.079.060.999.5
1611.670%1534.968.636.336.268.062.978.561.099.7
249.155%1535.968.635.837.168.062.076.460.799.2
326.640%1298.864.533.636.063.259.368.455.791.0
LLaVA-NeXT-7B (32 layers)
042.7100%1519.067.136.437.170.264.169.760.1100.0
833.879%1515.167.236.636.970.264.170.060.1100.0
1624.958%1476.867.236.237.370.263.567.859.498.8
1921.651%1525.167.236.037.870.463.465.759.599.0
2416.037%1504.165.436.436.068.160.564.958.196.7
LLaVA-NeXT-13B (40 layers)
081.8100%1570.070.535.939.971.965.766.761.3100.0
868.283%1552.470.635.039.671.965.166.961.099.5
1654.667%1561.070.135.039.771.964.866.960.999.3
2441.050%1553.070.236.239.971.863.667.561.099.5
3227.534%1468.465.835.238.969.360.558.557.493.6
+ +Table 2. Performance vs. Efficiency Balance of ShortV under different configurations. # ShortV Layers $(N)$ : the number of ShorV layers, Avg.: a normalized average score on the benchmarks, Per.: the relative performance retention compared with the vanilla models. + +Baselines. To evaluate the effectiveness of ShortV, which improves MLLM efficiency in a training-free manner, we compare it with popular training-free methods for MLLM efficiency, such as FastV [10] and VTW [29]. FastV drops visual tokens by a percentage of $R$ after the $K$ -th layer in the forward process of input tokens. It computes the average attention score one token received from all other tokens as the importance criterion to select pruned tokens. VTW drops all visual tokens after the $K$ -th layer, enabling only text tokens to engage in the subsequent layers. We use the default settings for the baselines as in their original papers. Specifically, for FastV, $K = 2$ and $R = 50\%$ . For VTW, $K = 16$ for 7B models and $K = 20$ for 13B models. + +# 4.2. Main Results + +In this section, we conduct experiments to compare ShortV with the baselines. The results are shown in Table 1. We perform evaluation on multiple popular vision language benchmarks, including MME [15], MMBench [33], MMMU (val) [55], MMStar [9], SEED-Bench [25], VQAv2 [16], and GQA [19]. We manually choose the number of ShortV layers $N$ to maintain a similar or lower FLOPs ratio compared with the baselines. The FLOPs ratio of ShortV is calculated according to Equation 5, where we set the number of text tokens to 64. Specifically, we + +choose $N = 19$ for the 7B models, and $N = 24$ for the 13B models. As shown in Table 1, our ShortV achieves comparable or superior performance across multiple benchmarks compared with the baselines. + +# 4.3. Balance between Efficiency and Performance + +In this section, we conduct an experiment to investigate the impact of ShortV's parameter $N$ , which denotes the number of ShortV layers. The experimental results are presented in Table 2. To facilitate intuitive comprehension, we plot the correlation between the normalized average score on benchmarks and the number of ShortV layers in Figure 1b. + +We observe that ShortV can freeze visual tokens in approximately $60\%$ of the MLLM layers while preserving superior performance. As $N$ continues to increase, both 7B and 13B models can maintain more than $90\%$ performance when the hidden states of visual tokens remain unchanged in about $80\%$ of the layers. These results are significantly different from those on text-only LLMs. For LLMs, Men et al. [36] remove transformations on text tokens in approximately $25\%$ of the LLM layers, and this results in about $10\%$ performance degradation on language benchmarks. As the number of layers increases, the performance of LLMs rapidly declines. These differences align with our observation in Section 2.3 that layers are more ineffective for visual + +
# ShortV Layers (N)08161924
LLaVA-1.5-7B1.00×1.13×1.23×1.30×1.40×
LLaVA-NeXT-7B1.00×1.15×1.35×1.44×1.64×
+ +Table 3. Inference speedups over the vanilla models, based on the 7B models. We conduct this test on a single GPU. + +
# ShortV Layers (N)08162432
LLaVA-1.5-13B1.00×1.13×1.24×1.39×1.50×
LLaVA-NeXT-13B1.00×1.13×1.30×1.52×1.84×
+ +Table 4. Inference speedups over the vanilla models, based on the 13B models. We conduct this test on a single GPU. + +
MethodFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
FastV58%64.335.865.460.2
ShortV55%64.836.266.260.9
ShortV+FastV29%64.237.165.159.3
+ +tokens than for text tokens. + +In addition to the theoretical FLOPs ratios, we provide the speedups on real hardware using different settings, as shown in Table 3 and Table 4. To get rid of the influence of different output sequence lengths, we use the first token latency to calculate the speedups. We utilize the MMMU dataset for the latency test. For comparison, we note that FastV [10] with its default setting, i.e. $K = 2$ and $R = 50\%$ , achieves a $1.31 \times$ speedup over the vanilla LLaVA-1.5-13B model. In contrast, our ShortV with its default parameter, i.e. $N = 24$ , achieves a greater speedup of $1.39 \times$ . + +# 4.4. Orthogonal to Token Pruning + +In this section, we demonstrate that ShortV is orthogonal to and compatible with visual token pruning, e.g. FastV [10]. FastV identifies $R\%$ unimportant visual tokens and drops them after the $K$ -th layer in the forward process of input tokens. We apply FastV to ShortV, which already replaces $N$ ineffective layers for visual tokens with ShortV layers. We use the default settings for FastV and ShortV in this experiment, i.e. $K = 2$ and $R = 50\%$ for FastV, and $N = 19$ for ShortV. We employ LLaVA-1.5-7B as the vanilla model. The experimental results in Table 5 demonstrate that ShortV is compatible with FastV and that the application of both can further improve MLLM efficiency. + +Table 5. ShortV is compatible with FastV, and applying both at the same time can further enhance MLLM efficiency. This experiment is based on LLaVA-1.5-7B. + +
StrategyFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
Random55%58.433.660.556.1
Cosine Sim.55%60.834.262.759.5
LC (Ours)55%64.836.266.260.9
+ +Table 6. Ablation on strategies to select replaced layers, based on LLaVA-1.5-7B. "Random" denotes randomly selecting 19 layers and freezing visual tokens in them. "Cosine Sim." denotes using cosine similarity to select ineffective layers for visual tokens. + +# 4.5. Ablation Studies + +Ablation on strategies to select replaced layers. In this paragraph, we perform an ablation experiment on LLaVA-1.5-7B to investigate the impact of strategies for selecting which layers to replace. ShortV selects ineffective layers for visual tokens, and replace them with ShortV layers. To identify which layers are ineffective, we utilize the LC metric introduced in Section 2. In contrast, previous work [36] on text-only LLMs uses a metric based on cosine similarity. It calculates the average cosine similarity between the inputs and outputs of each layer. The layers with higher cosine similarities are deemed more ineffective. To make a comparison between this cosine similarity metric and our LC metric, we calculate each layer's cosine similarity between the input hidden states and output hidden states of visual tokens, and select the same number of layers with the highest cosine similarities. We show the comparison in Table 6. We also include the results of another baseline, ShortV (Random), where visual tokens are frozen in the same number of randomly selected layers. These results clearly demonstrate that our LC metric performs better than cosine similarity in identifying ineffective MLLM layers for visual tokens, and ShortV based on the LC metric achieves performance comparable to the vanilla model. In contrast, ShortV based on the cosine similarity metric cannot match the performance of the vanilla model, although it outperforms the baseline with randomly selected layers. + +Ablation on frozen tokens. In this paragraph, we conduct an ablation study on LLaVA-1.5-7B to investigate the impact of freezing different types of tokens. In Section 2, we demonstrate that MLLM layers are ineffective for visual tokens, as measured by the LC metric. Motivated by this observation, ShortV freezes visual tokens in ineffective layers. In Table 7, we compare our method with the strategies of freezing other tokens. In the experiment detailed in line (a), we utilize the LC metric to identify 19 ineffective layers for text tokens, and freeze text tokens in them. Despite having fewer frozen tokens, we can observe that this strategy results in significant performance declines compared with our method, which freezes visual tokens rather than text to + +
Frozen TokensMMBench ENMMMU (val)SEED-BenchGQA
None (Vanilla)64.036.366.161.9
(a) Text2.123.78.92.9
(b) Text+Visual1.326.60.80.0
(c) Random1.522.95.52.3
(d) Visual (Ours)64.836.266.260.9
+ +Table 7. Ablation on frozen tokens, based on LLaVA-1.5-7B. (a) identifying 19 ineffective layers for text tokens and freezing text tokens in them. In lines (b) and (c), we select ineffective layers for all tokens. line (b) involves freezing all input tokens in them, whereas line (c) denotes randomly freezing the same number of tokens as the visual tokens. + +kens. These experimental results align with our findings in Section 2 that MLLM layers are more ineffective for visual tokens than for text tokens. In lines (b) and (c), we first calculate each layer's average LC score for all tokens, including visual and text tokens, and then select 19 ineffective layers. Next, in the experiment corresponding to line (b), we freeze all input tokens in these layers. In line (c), we freeze random input tokens, and the number of frozen tokens matches that of the visual tokens. As a result, the computational overhead associated with line (c) is the same as that of our method in line (d). We can find that freezing tokens other than visual tokens leads to substantial performance degradation in vision-language tasks. These ablations demonstrate the effectiveness of our ShortV. + +# 5. Related Work + +# 5.1. Multimodal Large Language Models + +Built upon Large Language Models (LLMs) [1, 12, 37, 47, 49], Multimodal Large Language Models (MLLMs) [4, 7, 35, 48, 52, 57] have made significant progress in processing and understanding the visual world. Typically, they use a decoder-only architecture. Specifically, they utilize visual encoders [14, 21, 34, 41, 56] to convert input visual information into visual features and then use projectors to project these visual features into visual tokens. These visual tokens are then concatenated with text tokens and fed into the LLM backbones. Current MLLMs use hundreds to thousands of visual tokens to represent a single image, significantly increasing the length of the token sequences. For instance, the LLaVA-1.5 models [31] transform each image with $336 \times 336$ resolution into 576 tokens. For images with higher resolutions, the LLaVA-NeXT models [32] process images into up to 2,880 visual tokens, and SPHINX-2k [28] divides one image into nine subimages, resulting in 2,890 visual tokens. Applying LLMs to such large numbers of visual tokens incurs substantial computational costs. In this paper, we introduce ShortV to enhance the efficiency of MLLMs by reducing the computational overhead associ + +ated with visual tokens. + +# 5.2. Efficient LLMs and MLLMs + +For LLMs, previous studies [36, 46] find that layers in LLMs are ineffective for text tokens. They remove computations in about $25\%$ of the layers, while preserving approximately $90\%$ of the performance. LaCo [50] utilizes layer merging for efficient LLMs. + +To address the computational inefficiency of MLLMs, previous methods [10, 29, 42, 44, 45, 54] have primarily focused on two aspects: efficient model architecture and visual token compression. Among efficient model architectures, cross-attention-based models [2, 3, 8, 20] insert gated cross-attention layers within LLM layers for visual perception, but previous studies [13, 24] demonstrate that this architecture performs worse than the decoder-only architecture in the same settings. Instead of inserting cross-attention layers, mPLUG-Owl3 [51] and Vamba [42] introduce cross-attention operations in parallel with self-attention. In contrast, SAISA [54] introduces NAAViT self-attention blocks, which incorporate multimodal cross-attention into the original self-attention operations of the LLMs, and reuse the parameters of self-attention blocks. The design of ShortV layers is inspired by NAAViT. Differently, in ShortV layers, visual tokens also skip their FFNs. + +Visual token compression methods improve MLLM efficiency in both training-based [5, 6, 22, 26, 27] and training-free [10, 43, 53] manners. FastV [10] reveals token-wise redundancy, and it removes unimportant tokens during inference. In this paper, we reveal layer-wise redundancy in MLLMs. Layers in MLLMs are much more ineffective for visual tokens than for text token. Therefore, we can freeze visual tokens in approximately $60\%$ of the MLLM layers with minimal performance degradation. Unlike previous methods for MLLM efficiency, ShortV does not reduce the number of visual tokens but instead decreases the computational costs of processing each token. ShortV is training-free and orthogonal to token compression. We demonstrate that ShortV is compatible with FastV, allowing for simultaneous application to further enhance MLLM efficiency. + +# 6. Conclusion + +In this paper, we explore the layer-wise redundancy in MLLMs. We discover that layers in MLLMs are more ineffective for visual tokens than for text tokens. MLLM layers' transformations on visual tokens have a minimal impact on the MLLM output. Motivated by this observation, we propose ShortV, a training-free method to enhance MLLM efficiency. ShortV utilizes our proposed LC metric to select ineffective layers for visual tokens, and freezes visual tokens in these layers. It can freeze visual tokens in about $60\%$ of the layers while preserving superior performance. + +# 7. Acknowledgment + +We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265) and the Basic Research Program of ISCAS (Grant No. ISCAS-ZD-202401). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 8 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8 +[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8 +[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models, 2023. 8 +[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8 +[6] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 8 +[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8 +[8] Kaibing Chen, Dong Shen, Hanwen Zhong, Huasong Zhong, Kui Xia, Di Xu, Wei Yuan, Yifei Hu, Bin Wen, Tianke Zhang, et al. Evlm: An efficient vision-language model for visual understanding. arXiv preprint arXiv:2407.14177, 2024. 8 +[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 6 + +[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5, 6, 7, 8 +[11] Xiaodong Chen, Yuxuan Hu, Jing Zhang, Yanling Wang, Cuiping Li, and Hong Chen. Streamlining redundant layers to compress large language models. arXiv preprint arXiv:2403.19135, 2024. 3 +[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 8 +[13] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024.8 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2, 5, 8 +[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 2, 6 +[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6 +[17] Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. What matters in transformers? not all attention is needed. arXiv preprint arXiv:2406.15786, 2024. 3 +[18] Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Deciphering cross-modal alignment in large vision-language models with modality integration rate. arXiv preprint arXiv:2410.07167, 2024. 2 +[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 4, 6 +[20] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.8 +[21] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 8 + +[22] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 8 +[23] Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim, Thibault Castells, Shinkook Choi, Junho Shin, and Hyoung-Kyu Song. Shortened llama: Depth pruning for large language models with comparison of retraining methods. arXiv preprint arXiv:2402.02834, 2024. 3 +[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2025. 8 +[25] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 6 +[26] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 8 +[27] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 8 +[28] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 8 +[29] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 6, 8 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 4 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 4, 5, 8 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,5, 8 +[33] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 2, 6 +[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 8 + +[35] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024.8 +[36] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect, 2024. URL https://arxiv.org/abs/2403.03853, 2024. 2, 3, 4, 6, 7, 8 +[37] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2(5):6, 2024. 1, 8 +[38] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1 +[39] OpenAI. Introducing gpt-4o: our fastest and most affordable flagship model. https://platform.openai.com/docs/guides/vision, 2024. Accessed: 2024-05-26. 1 +[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 2, 3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 8 +[42] Weiming Ren, Wentao Ma, Huan Yang, Cong Wei, Ge Zhang, and Wenhu Chen. Vamba: Understanding hour-long videos with hybrid mamba-transformers. arXiv preprint arXiv:2503.11579, 2025. 8 +[43] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 8 +[44] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8 +[45] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 8 +[46] Jiwon Song, Kyungseok Oh, Taesu Kim, Hyungjun Kim, Yulhwa Kim, and Jae-Joon Kim. Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks. arXiv preprint arXiv:2402.09025, 2024. 3, 8 +[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 8 + +[48] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 8 +[49] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 1, 8 +[50] Yifei Yang, Zouying Cao, and Hai Zhao. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187, 2024. 8 +[51] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. MPLUG-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8 +[52] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 8 +[53] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 8 +[54] Qianhao Yuan, Yanjiang Liu, Yaojie Lu, Hongyu Lin, Ben He, Xianpei Han, and Le Sun. Saisa: Towards multimodal large language models with both training and inference efficiency. arXiv preprint arXiv:2502.02458, 2025. 8 +[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 6 +[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 8 +[57] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8 \ No newline at end of file diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/images.zip b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dcfb1647753fb90b50bca44098666dd979f9e895 --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cbe3588ee8cc30edc0f5c32a1828dc2e6262dcdcf4901c27a4eb6293ade1fd0 +size 543974 diff --git a/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/layout.json b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cdf00a5a46773a2445914c4b7eb3520c230534b6 --- /dev/null +++ b/2025/ShortV_ Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers/layout.json @@ -0,0 +1,8613 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "text", + "content": "ShortV: Efficient Multimodal Large Language Models by Freezing Visual Tokens in Ineffective Layers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "spans": [ + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": "Qianhao Yuan" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Qingyu Zhang" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Yanjiang Liu" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Jiawei Chen" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Yaojie Lu" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Hongyu Lin" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Jia Zheng" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Xianpei Han" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "text", + "content": ", Le Sun" + }, + { + "bbox": [ + 131, + 160, + 478, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "spans": [ + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "type": "text", + "content": "Institute of Software, Chinese Academy of Sciences " + }, + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 189, + 434, + 217 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 217, + 533, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 217, + 533, + 246 + ], + "spans": [ + { + "bbox": [ + 75, + 217, + 533, + 246 + ], + "type": "text", + "content": "{yuanqianhao2024,zhangqingyu2024,liuyanjiang2021,chenjiawei2024, luyaojie,hongyu,zhengjia,xianpei,sunle}@iscas.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) suffer from high computational costs due to their massive size and the large number of visual tokens. In this paper, we investigate layer-wise redundancy in MLLMs by introducing a novel metric, Layer Contribution (LC), which quantifies the impact of a layer's transformations on visual and text tokens, respectively. The calculation of LC involves measuring the divergence in model output that results from removing the layer's transformations on the specified tokens. Our pilot experiment reveals that many layers of MLLMs exhibit minimal contribution during the processing of visual tokens. Motivated by this observation, we propose ShortV, a training-free method that leverages LC to identify ineffective layers, and freezes visual token updates in these layers. Experiments show that ShortV can freeze visual token in approximately " + }, + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "text", + "content": " of the MLLM layers, thereby dramatically reducing computational costs related to updating visual tokens. For example, it achieves a " + }, + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 299, + 297, + 552 + ], + "type": "text", + "content": " reduction in FLOPs on LLaVA-NeXT-13B while maintaining superior performance. The code is publicly available at https://github.com/icip-cas/ShortV." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Large language models (LLMs) have achieved remarkable performance in natural language tasks [1, 12, 37, 47, 49]. Building upon LLMs, Multimodal Large Language Models (MLLMs) [31, 32, 38, 39] take a significant step towards understanding the real physical world by incorporating visual information into their processes. Typically, an MLLM consists of a visual encoder, a projector, and an LLM backbone. Most of them preprocess visual information through" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 269, + 539, + 355 + ], + "blocks": [ + { + "bbox": [ + 331, + 269, + 539, + 355 + ], + "lines": [ + { + "bbox": [ + 331, + 269, + 539, + 355 + ], + "spans": [ + { + "bbox": [ + 331, + 269, + 539, + 355 + ], + "type": "image", + "image_path": "66b6621944e119e1e75bbc22db6ecbb29532ea43e4bb6380c238b633a9345249.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 356, + 452, + 365 + ], + "lines": [ + { + "bbox": [ + 417, + 356, + 452, + 365 + ], + "spans": [ + { + "bbox": [ + 417, + 356, + 452, + 365 + ], + "type": "text", + "content": "(a) ShortV" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 367, + 552, + 491 + ], + "blocks": [ + { + "bbox": [ + 317, + 367, + 552, + 491 + ], + "lines": [ + { + "bbox": [ + 317, + 367, + 552, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 552, + 491 + ], + "type": "image", + "image_path": "635f725d3c85e7cae0aab88e6316eec731d238d26f4ac361155a5e70a2063179.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 494, + 516, + 504 + ], + "lines": [ + { + "bbox": [ + 353, + 494, + 516, + 504 + ], + "spans": [ + { + "bbox": [ + 353, + 494, + 516, + 504 + ], + "type": "text", + "content": "(b) Performance vs. the Number of ShortV Layers" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 513, + 555, + 624 + ], + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 624 + ], + "type": "text", + "content": "Figure 1. (a) Illustration of ShortV. We identify ineffective layers for visual tokens and replace these layers with sparse ShortV layers. In ShortV layers, we freeze visual tokens, and eliminate computations related to updating them. ShortV improves MLLM efficiency in a training-free manner and involves no parameter updates. Notably, ShortV is compatible with token pruning methods, e.g. FastV. (b) Performance vs. the number of ShortV layers. Average Performance means a normalized average score on multiple benchmarks. ShortV can freeze visual tokens in approximately " + }, + { + "bbox": [ + 313, + 513, + 555, + 624 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 513, + 555, + 624 + ], + "type": "text", + "content": " of the MLLM layers with nearly no performance degradation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 651, + 555, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 697 + ], + "type": "text", + "content": "a visual encoder, e.g. a CLIP-ViT [14, 41], and project the patch-level visual features into visual tokens through a projector. Then they concatenate visual and text tokens and feed them into the LLM backbone." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "However, MLLMs face a substantial increase in compu" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "329" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tational overhead. This burden primarily stems from the large scale of the LLM backbones and the significantly extended length of the concatenated visual-text token sequences. To address this issue, Chen et al. [10] discovers significant token-wise redundancy in MLLMs. Based on this redundancy, they propose FastV, which identifies and prunes unimportant visual tokens in MLLMs to improve their efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 170, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 348 + ], + "type": "text", + "content": "Other than token-wise redundancy, in this paper, we reveal that MLLMs also exhibit significant layer-wise redundancy in processing visual tokens. Specifically, we propose Layer Contribution (LC), a metric that quantifies how much a layer's transformations on certain tokens contribute to the model's output. In the LC calculation of a layer, we freeze certain tokens in this layer, i.e. keep the hidden states of the tokens unchanged, and then compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original model. This metric provides a direct measure of a layer's importance for certain tokens. By comparing LC scores on visual tokens and those on text tokens, we discover that MLLM layers are ineffective for visual tokens, and their transformations on visual tokens contribute minimal to the model's output." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 350, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 445 + ], + "type": "text", + "content": "This phenomenon inspires us to propose ShortV, a simple but effective method to improve the efficiency of MLLMs. In ShortV, we first utilize the LC metric to identify layers least effective at transforming visual tokens, and then replace these layers with sparse ShortV layers. Within these sparse layers, visual tokens remain frozen, and the corresponding computations for updating them are eliminated, as shown in Figure 1a." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 614 + ], + "type": "text", + "content": "To validate the effectiveness of ShortV, we conduct evaluations across multiple benchmarks, including MME [15], MMBench [33], MMMU [55], MMStar [9], SEED-Bench [25], GQA [19], and Flickr30K [40]. Figure 1b illustrates the correlation between the normalized average performance on these benchmarks and the number of replaced layers. As observed, ShortV can replace approximately " + }, + { + "bbox": [ + 55, + 447, + 295, + 614 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 447, + 295, + 614 + ], + "type": "text", + "content": " of MLLM layers without performance degradation. Unlike FastV and other token pruning methods, ShortV reduces computations of per visual token rather than reducing the number of visual tokens. Therefore, ShortV and token pruning methods are orthogonal and compatible. Furthermore, we demonstrate that combining ShortV and FastV can further enhance MLLM efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 616, + 241, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 241, + 627 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 241, + 627 + ], + "type": "text", + "content": "We summarize our contribution as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 630, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 665 + ], + "type": "text", + "content": "- We propose Layer Contribution (LC), a metric to quantify how much a layer's transformations on specific tokens contribute to the model's output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "type": "text", + "content": "- Leveraging LC, we reveal significant redundancy in MLLM layers for visual tokens. Transformations on visual tokens in many layers contribute minimally and are thus ineffective." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "text", + "content": "- Based on the observation above, we propose ShortV, which improves MLLM efficiency by freezing visual tokens in ineffective layers. ShortV can freeze visual tokens in approximately " + }, + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 314, + 72, + 553, + 144 + ], + "type": "text", + "content": " of MLLM layers without performance degradation. Extensive experiments and ablation studies demonstrate ShortV's effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 154, + 487, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 487, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 487, + 167 + ], + "type": "text", + "content": "2. Layer Redundancy in MLLMs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 553, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 234 + ], + "type": "text", + "content": "In this section, we first introduce the background of layer redundancy in text-only LLMs, and then propose a metric to measure MLLM layer redundancy for different types of tokens. Next, we conduct a pilot experiment to investigate layer redundancy in MLLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 241, + 394, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 394, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 394, + 253 + ], + "type": "text", + "content": "2.1. Background" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 329 + ], + "type": "text", + "content": "Typically, MLLMs are built on text-only LLMs. MLLMs employ pre-trained visual encoders, such as CLIP-ViT [14, 41], to convert images into visual features, and then use projectors to project them into visual tokens in the text token embedding space. The visual tokens are concatenated with text tokens and fed into the LLM backbones." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 331, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 594 + ], + "type": "text", + "content": "For text-only LLMs, Men et al. [36] identify notable redundancy across their layer. Some layers' transformations on the hidden states of text token contribute minimally to the overall model functionality. Consequently, these layers are considered ineffective. Removing these transformations in approximately " + }, + { + "bbox": [ + 313, + 331, + 553, + 594 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 331, + 553, + 594 + ], + "type": "text", + "content": " of LLM layers has minimal impact on model outputs. Such redundancy mainly occurs in middle-to-deeper layers, whereas initial layers and the last layer remain critical to the model functionality. However, this pattern may not hold for MLLMs. Huang et al. [18] demonstrate a clear modality gap in the embedding space of current MLLMs, where visual and text tokens exhibit a uniform distribution within each modality but a significant distribution gap between modalities. Such a modality gap implies that MLLMs might adopt distinct computational patterns or strategies for processing visual and text tokens, potentially affecting how redundancy is distributed across layers. This raises several key questions: Are MLLM layers as ineffective for visual tokens as LLM layers are for text tokens? To what extent does layer redundancy exist in MLLMs? How is this redundancy distributed across different MLLM layers?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 463, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 463, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 463, + 612 + ], + "type": "text", + "content": "2.2. Layer Contribution Metric" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "To investigate layer redundancy for certain tokens, we freeze these tokens within the investigated layer, i.e. keep hidden states of these tokens unchanged. To achieve this, we introduce sparse layers shown in Figure 2 for visual and text tokens, respectively. Based on these designs, we propose the Layer Contribution (LC) metric, which evaluates how much a layer's transformations on certain tokens contribute to the model's overall functionality. In the calcu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "330" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 296, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 296, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 296, + 152 + ], + "type": "image", + "image_path": "44a32d4ecee0764438b174d8aa73406d7836394a5427d968f4a082df6e8b35ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 154, + 295, + 183 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 295, + 183 + ], + "type": "text", + "content": "(a) Sparse layer where visual tokens are frozen, used to investigate layer redundancy for visual tokens. Only text tokens function as queries and are passed through the FFN. We also denote this layer as ShortV layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 185, + 227, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 185, + 227, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 185, + 227, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 185, + 227, + 247 + ], + "type": "image", + "image_path": "605d72357eb987a089e455802200c69111b29563f3df670ac362d840f83f468e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 295, + 281 + ], + "type": "text", + "content": "(b) Sparse layer where text tokens are frozen, used to investigate layer redundancy for text tokens. Only visual tokens function as queries and are passed through the FFN." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 184, + 290, + 248 + ], + "blocks": [ + { + "bbox": [ + 233, + 184, + 290, + 248 + ], + "lines": [ + { + "bbox": [ + 233, + 184, + 290, + 248 + ], + "spans": [ + { + "bbox": [ + 233, + 184, + 290, + 248 + ], + "type": "image", + "image_path": "9525adcce18eef1a74c836efa284244c85b3cfd37fd441d513d3309fb0e6a00f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 296, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 355 + ], + "type": "text", + "content": "Figure 2. Sparse layers used to investigate layer redundancy for different tokens. To investigate layer redundancy for certain tokens, we freeze these tokens within the layer, i.e. keep hidden states of these tokens unchanged, and measure the divergence between the model's output logits and those of the original model. We gray out the attention that does not need calculation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "type": "text", + "content": "lation of LC, we replace the investigated dense layer with the corresponding sparse layer, and compute the Kullback-Leibler (KL) divergence between the resulting model's output logits and those of the original dense model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": "Specifically, We assume the LLM backbone " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": " layers, and each of them consists of a self-attention block and a feed-forward network (FFN). The input of the " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": "-th layer at the " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": "-th token position is " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "H_{i}^{j}" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": ", and the corresponding output is " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "H_{i+1}^{j}" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": ". The output of the last layer, i.e. the " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": "-th layer, at the same position is " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "H_{L+1}^{j}" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": ". The model " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": " utilizes " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "H_{L+1}^{-1}" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": " at the last token position to compute logits for next-token prediction through the language model head " + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "inline_equation", + "content": "LM_{head}" + }, + { + "bbox": [ + 55, + 425, + 296, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 537, + 295, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 537, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 110, + 537, + 295, + 553 + ], + "type": "interline_equation", + "content": "\\operatorname {l o g i t s} (M) = L M _ {\\text {h e a d}} \\left(H _ {L + 1} ^ {- 1}\\right). \\tag {1}", + "image_path": "6a41b025a97801b37ac9ad13ac981b002018884047f64fa14aae5090cfb51d63.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "To investigate how much the " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "-th layer's transformations on certain tokens " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " contribute to the model functionality, we replace the " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "-th layer with a sparse layer where " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " are frozen, i.e. " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "'s hidden states remain unchanged in this sparse layer. The resulting model is denoted as " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i^X" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ". In practice, " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " can be visual tokens " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " or text tokens " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ". As shown in Figure 2, we introduce sparse layers where " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " are frozen, respectively. In Figure 2a, we freeze the visual tokens in the sparse layer. Within the self-attention block of this layer, the visual tokens do not attend to other tokens, and only the text tokens function as queries. For the FFN of this layer, we simply do not pass the visual tokens through it. In Figure 2b, we freeze the text tokens in another sparse" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "layer with similar designs. Based on these, we define the " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "-th layer's Layer Contribution (LC) score for certain tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " as the KL divergence between the output logits of the original model " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " and those of the model " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i^X" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " are frozen in the " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "-th layer:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 141, + 553, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 141, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 351, + 141, + 553, + 156 + ], + "type": "interline_equation", + "content": "L C _ {i} ^ {X} = K L \\left(\\operatorname {l o g i t s} (M), \\operatorname {l o g i t s} \\left(\\mathcal {M} _ {i} ^ {X}\\right)\\right), \\tag {2}", + "image_path": "d034900e0eff994f9fd6705939036e83d8d45c583e6fe7b0be8fee60daf2d9aa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "inline_equation", + "content": "KL(\\cdot)" + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": " denotes KL divergence. A lower LC score implies that the layer's transformations on the tokens exhibit minimal contribution to the model's output, suggesting that these transformations are ineffective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 411 + ], + "type": "text", + "content": "Discussion: Why not use perplexity or cosine similarity as the metric to measure the importance of layers? Some work in text-only LLMs utilizes perplexity [23, 46] or cosine similarity [11, 17, 36] as metrics to measure the importance of each layer. For the former metric, instead of KL divergence, they measure the change in perplexity of the models, and the layers causing minimal perplexity changes are deemed ineffective. This metric, however, is inadequate when measuring layer redundancy for visual tokens. We find that even if we do not feed visual tokens into the MLLMs, they can still generate reasonable responses, and the changes in perplexity of the MLLMs is relatively low. Nevertheless, they face significant performance degradation in vision language tasks when the visual information is absent. Thus, perplexity is not a reliable measure when evaluating layer redundancy for visual tokens." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 412, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 628 + ], + "type": "text", + "content": "For the latter metric, the cosine similarity between the input and output of a certain layer is calculated. The hypothesis here is that the ineffective layers have less transformations on the hidden states of tokens, and therefore their inputs and outputs demonstrate higher similarities. However, in our evaluation on LLaVA-1.5 [31] models, the cosine similarity metric and LC differ in their measurement of the redundancy distribution across different layers. Compared with the LC metric, which directly measures the logits divergence between model outputs, cosine similarity consistently overestimates the redundancy of the shallow layers and underestimates the redundancy of the deep layers. We believe the reason behind this difference is that cosine similarity neglects the position of a layer within the model. Specifically, minor transformations of hidden states in the shallow layers can influence all subsequent layers, whereas transformations with similar extent in deeper layers tend to have less impact on overall model functionality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 544, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 544, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 544, + 648 + ], + "type": "text", + "content": "2.3. Ineffective MLLM Layers for Visual Tokens" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "We conduct a pilot experiment to investigate layer redundancy in LLaVA-1.5-7B and LLaVA-1.5-13B [31]. We first randomly sample 2,000 cases from a combination of two major vision language tasks, including caption (Flickr30K [40]) and visual question answering" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "331" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 301, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 301, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 301, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 301, + 198 + ], + "type": "image", + "image_path": "50ae17a45224c197973a813145c51a42819f3f150e8116e07b9580967f191d5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 201, + 209, + 210 + ], + "lines": [ + { + "bbox": [ + 148, + 201, + 209, + 210 + ], + "spans": [ + { + "bbox": [ + 148, + 201, + 209, + 210 + ], + "type": "text", + "content": "(a) LLaVA-1.5-7B" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 553, + 198 + ], + "type": "image", + "image_path": "fb94ad90f8c6e4bbae719f773e9f78a260782cdc9741d3a5273d4714abcf1b61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 201, + 463, + 210 + ], + "lines": [ + { + "bbox": [ + 399, + 201, + 463, + 210 + ], + "spans": [ + { + "bbox": [ + 399, + 201, + 463, + 210 + ], + "type": "text", + "content": "(b) LLaVA-1.5-13B" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 274, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 295, + 323 + ], + "type": "text", + "content": "(GQA [19]). Then we utilize these samples to calculate each MLLM layer's average LC score for visual and text tokens, respectively. Figure 3 shows the results. We summarize our findings as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 323, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 295, + 407 + ], + "type": "text", + "content": "First, for text tokens, middle to deeper layers are more ineffective, while the initial and last layers make more contributions to the MLLM functionality. These observations align with the layer redundancy distribution of text-only LLMs found in Men et al. [36], indicating that visual instruction tuning [30, 31] does not significantly alter the manner LLMs process text tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 408, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 480 + ], + "type": "text", + "content": "Second, for visual tokens, the initial and the deep layers, including the last one, exhibit higher redundancy than other layers, which is different from the distribution for text token. Notably, since the last layer's transformations on visual tokens do not contribute to the model's output, its LC score for visual tokens is always 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 481, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 481, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 481, + 295, + 552 + ], + "type": "text", + "content": "Third, layer redundancy shows an imbalance between visual and text tokens. Each layer's LC score on visual tokens are lower than that on text tokens, which means that many layers' transformations on visual tokens are ineffective, and freezing visual token in these ineffective layers results in minimal impact on the models' output." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 554, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 554, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 54, + 554, + 295, + 625 + ], + "type": "text", + "content": "We attribute the different layer redundancy patterns for different modalities to the modality gap. The clear distribution gap of visual and text tokens result in the difference in how MLLMs process them. We hope these findings can provide insights into how MLLMs process visual and text tokens in different layers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 108, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 108, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 108, + 651 + ], + "type": "text", + "content": "3. ShortV" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 658, + 286, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 286, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 286, + 672 + ], + "type": "text", + "content": "3.1. Freezing Visual Tokens in Ineffective Layers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As demonstrated in the previous section, we identify significant layer redundancy for visual tokens in MLLMs. Most layers' transformations on visual tokens are ineffective for" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 342, + 275, + 527, + 403 + ], + "blocks": [ + { + "bbox": [ + 54, + 220, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 255 + ], + "type": "text", + "content": "Figure 3. The Layer Contribution (LC) scores of LLaVA-1.5-7B and LLaVA-1.5-13B. A lower LC score implies that the layer's transformations on the specified tokens are more ineffective. Layers are more ineffective for visual tokens than for text tokens, and freezing visual tokens in ineffective layers results in minimal output divergence from the original model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 275, + 527, + 403 + ], + "lines": [ + { + "bbox": [ + 342, + 275, + 527, + 403 + ], + "spans": [ + { + "bbox": [ + 342, + 275, + 527, + 403 + ], + "type": "image", + "image_path": "1c34b6b57f6727d65239d91bf26949f6695bdd211480c8d59101295abf334f00.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "text", + "content": "Figure 4. Details of ShortV layer. In this layer, only text tokens pass through the " + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "inline_equation", + "content": "W_{Q}" + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "inline_equation", + "content": "W_{O}" + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "text", + "content": " matrices and the FFN. The attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 529 + ], + "type": "text", + "content": "the model functionality. Based on this observation, we propose a direct method to enhance MLLM efficiency in a training-free manner: freezing visual tokens in ineffective layers. We denote this method as ShortV." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "text", + "content": "In ShortV, we replace the ineffective dense layers with sparse ShortV layers, where visual tokens are frozen. The ShortV layers are same as the sparse layer shown in 2a, and we illustrate their detailed architecture in Figure 4. In each ShortV layer, only text tokens are passed through the FFN and the " + }, + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "inline_equation", + "content": "W_{Q}" + }, + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "inline_equation", + "content": "W_{O}" + }, + { + "bbox": [ + 313, + 532, + 555, + 639 + ], + "type": "text", + "content": " matrices of the attention block, and the attention mask is same as that in Figure 2a, where visual tokens do not attend to other tokens, and only text tokens function as queries." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "ShortV has one parameter: the number of replaced layers, which we denote as " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". First, we construct a tiny dataset, which contains a small number of samples from vision language tasks. Then we use this dataset to calculate each layer's average LC score for visual tokens. Next, we sort the layers in ascending order according to LC scores," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "332" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 353 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 353 + ], + "type": "table", + "html": "
MethodTFLOPsFLOPs RatioVQAv2GQASEED-BenchMMMU (val)MMEMMBench ENMMStarFlickr30K CIDEr
LLaVA-1.5-7B
Vanilla8.5100%76.561.966.136.31510.764.133.774.9
FastV (K=2, R=50%)4.958%73.560.265.435.81475.664.332.467.5
VTW (K=16)4.755%66.355.166.236.11497.064.032.844.5
ShortV (Ours, N=19)4.755%75.760.966.236.21503.164.833.371.3
LLaVA-1.5-13B
Vanilla16.6100%78.063.368.235.41531.368.936.179.6
FastV (K=2, R=50%)9.457%76.759.467.834.61506.668.335.973.4
VTW (K=20)9.155%75.360.668.234.91533.068.536.165.9
ShortV (Ours, N=24)9.155%77.262.068.035.81535.968.637.176.4
LLaVA-NeXT-7B
Vanilla42.7100%80.064.170.236.41519.067.137.169.7
FastV (K=2, R=50%)22.052%79.563.069.635.11482.066.336.566.7
VTW (K=16)21.851%75.655.870.235.71518.267.137.638.4
ShortV (Ours, N=19)21.651%78.863.470.436.01525.167.237.865.7
LLaVA-NeXT-13B
Vanilla81.8100%80.965.771.935.91570.069.339.966.7
FastV (K=2, R=50%)42.151%76.862.971.535.91546.468.539.666.0
VTW (K=20)41.751%77.061.571.834.81569.469.139.856.6
ShortV (Ours, N=24)41.050%79.763.671.836.21553.070.239.967.5
", + "image_path": "ebe88fb54e577822306932ecbba5a453d7b3bb79a44182e3f7ea588ea3821c63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 384 + ], + "type": "text", + "content": "Table 1. Comparison of various training-free methods for MLLM efficiency. FLOPs Ratio denotes the proportion of FLOPs retained after applying the corresponding method to improve MLLM efficiency, compared with the vanilla model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 404, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 404, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 404, + 294, + 453 + ], + "type": "text", + "content": "and the layers with lower LC scores are more ineffective for visual tokens. Finally, we freeze visual tokens in the " + }, + { + "bbox": [ + 54, + 404, + 294, + 453 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 404, + 294, + 453 + ], + "type": "text", + "content": " layers with the lowest LC scores by replacing them with sparse ShortV layers, while keeping their original parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 453, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 295, + 561 + ], + "type": "text", + "content": "ShortV is training-free and involves no parameter updates. It can be applied to various MLLMs for different vision language tasks. Notably, ShortV is orthogonal to and compatible with popular visual token pruning methods, e.g. FastV [10]. Visual token pruning directly reduces the number of visual tokens, while ShortV mitigates the computational overhead related to each visual token. This means that we can apply ShortV and token pruning at the same time to further improve MLLM efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 567, + 173, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 173, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 173, + 578 + ], + "type": "text", + "content": "3.2. Computational Cost" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "content": "We consider the computations of self-attention blocks and feed-forward networks (FFNs) in layers of the LLM backbone. Assume " + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "content": " is the number of text tokens, " + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "content": " is the number of visual tokens, " + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "content": " is the hidden state size, " + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 584, + 296, + 655 + ], + "type": "text", + "content": " is the intermediate size of the FFNs, the total FLOPs of one dense Transformer layer can be estimated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 662, + 295, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 662, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 73, + 662, + 295, + 676 + ], + "type": "interline_equation", + "content": "F L O P s = 2 (t + v) (4 h + 3 m) h + 4 (t + v) ^ {2} h. \\tag {3}", + "image_path": "3e16388b368123b88afe147634afa1054e6dc6d8adc8d004c9675fdcae3efcc7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 682, + 275, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 682, + 275, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 682, + 275, + 693 + ], + "type": "text", + "content": "For one ShortV layer, the FLOPs can be calculated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "F L O P s ^ {*} = 2 t (4 h + 3 m) h + 4 v h ^ {2} + 4 t (t + v) h. \\tag {4}", + "image_path": "54288465b51906a672435c3bd570cdbffa67d6d0d76d482ee2aeae6e33a09047.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "content": "For the whole model, assume the LLM has " + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "content": " layers in total, ShortV selects " + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "content": " ineffective dense layers and replaces them with ShortV layers. The FLOPs of the original dense model are " + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "inline_equation", + "content": "L \\times FLOPs" + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "content": ", and the FLOPs of ShortV are calculated as " + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "inline_equation", + "content": "(L - N) \\times FLOPs + N \\times FLOPs^{*}" + }, + { + "bbox": [ + 313, + 404, + 555, + 477 + ], + "type": "text", + "content": ". The FLOPs ratio of ShortV and the original model is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 485, + 555, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 485, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 342, + 485, + 555, + 510 + ], + "type": "interline_equation", + "content": "r = \\frac {(L - N) \\times F L O P s + N \\times F L O P s ^ {*}}{L \\times F L O P s}. \\tag {5}", + "image_path": "42dc89a2fc26e845415289e31360bef082437dc732bf613addc9c3591a9885fa.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 520, + 397, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 397, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 397, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "Models. To validate the effectiveness of ShortV, we conduct experiments on popular open-source MLLMs, such as LLaVA-1.5-7B [31], LLaVA-1.5-13B, LLaVA-NeXT-7B [32] and LLaVA-NeXT-13B. LLaVA-1.5 models process images with a " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " resolution and treat each image as 576 tokens. LLaVA-NeXT splits high-resolution images into subimages, and encode the subimages and down-sampled original images independently. This allows the models to scale the input to any arbitrary resolution, without performing positional embedding interpolation for ViTs [14]. LLaVA-NeXT scales the input image resolution to " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " and visual token number up to " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " compared with LLaVA-1.5, i.e. 2880 tokens for each image." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "333" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 531, + 357 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 531, + 357 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 531, + 357 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 531, + 357 + ], + "type": "table", + "html": "
# ShortV Layers (N)TFLOPsFLOPs RatioMMEMMBench ENMMMU (val)MMStarSEED-BenchGQAFlickr30K CIDErAvg.Per.
LLaVA-1.5-7B (32 layers)
08.5100%1510.764.136.333.766.161.974.958.9100.0
86.981%1508.664.336.033.866.261.474.558.899.8
165.362%1487.064.936.133.365.761.072.858.399.0
194.755%1503.164.836.233.366.260.971.358.399.0
243.744%1341.760.734.133.462.558.364.254.392.2
LLaVA-1.5-13B (40 layers)
016.6100%1531.368.935.436.168.263.379.661.2100.0
814.185%1521.968.635.636.068.263.079.060.999.5
1611.670%1534.968.636.336.268.062.978.561.099.7
249.155%1535.968.635.837.168.062.076.460.799.2
326.640%1298.864.533.636.063.259.368.455.791.0
LLaVA-NeXT-7B (32 layers)
042.7100%1519.067.136.437.170.264.169.760.1100.0
833.879%1515.167.236.636.970.264.170.060.1100.0
1624.958%1476.867.236.237.370.263.567.859.498.8
1921.651%1525.167.236.037.870.463.465.759.599.0
2416.037%1504.165.436.436.068.160.564.958.196.7
LLaVA-NeXT-13B (40 layers)
081.8100%1570.070.535.939.971.965.766.761.3100.0
868.283%1552.470.635.039.671.965.166.961.099.5
1654.667%1561.070.135.039.771.964.866.960.999.3
2441.050%1553.070.236.239.971.863.667.561.099.5
3227.534%1468.465.835.238.969.360.558.557.493.6
", + "image_path": "2fb67566c27355a598e0643213fbcc9b969b7db87c518b9c1af64617ac945486.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "text", + "content": "Table 2. Performance vs. Efficiency Balance of ShortV under different configurations. # ShortV Layers " + }, + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "text", + "content": " : the number of ShorV layers, Avg.: a normalized average score on the benchmarks, Per.: the relative performance retention compared with the vanilla models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": "Baselines. To evaluate the effectiveness of ShortV, which improves MLLM efficiency in a training-free manner, we compare it with popular training-free methods for MLLM efficiency, such as FastV [10] and VTW [29]. FastV drops visual tokens by a percentage of " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": " after the " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": "-th layer in the forward process of input tokens. It computes the average attention score one token received from all other tokens as the importance criterion to select pruned tokens. VTW drops all visual tokens after the " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": "-th layer, enabling only text tokens to engage in the subsequent layers. We use the default settings for the baselines as in their original papers. Specifically, for FastV, " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "R = 50\\%" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": ". For VTW, " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": " for 7B models and " + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 54, + 408, + 294, + 564 + ], + "type": "text", + "content": " for 13B models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 575, + 142, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 142, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 142, + 588 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "In this section, we conduct experiments to compare ShortV with the baselines. The results are shown in Table 1. We perform evaluation on multiple popular vision language benchmarks, including MME [15], MMBench [33], MMMU (val) [55], MMStar [9], SEED-Bench [25], VQAv2 [16], and GQA [19]. We manually choose the number of ShortV layers " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " to maintain a similar or lower FLOPs ratio compared with the baselines. The FLOPs ratio of ShortV is calculated according to Equation 5, where we set the number of text tokens to 64. Specifically, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "content": "choose " + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "inline_equation", + "content": "N = 19" + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "content": " for the 7B models, and " + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "inline_equation", + "content": "N = 24" + }, + { + "bbox": [ + 313, + 408, + 555, + 456 + ], + "type": "text", + "content": " for the 13B models. As shown in Table 1, our ShortV achieves comparable or superior performance across multiple benchmarks compared with the baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 467, + 548, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 548, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 548, + 480 + ], + "type": "text", + "content": "4.3. Balance between Efficiency and Performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 485, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 557 + ], + "type": "text", + "content": "In this section, we conduct an experiment to investigate the impact of ShortV's parameter " + }, + { + "bbox": [ + 313, + 485, + 554, + 557 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 485, + 554, + 557 + ], + "type": "text", + "content": ", which denotes the number of ShortV layers. The experimental results are presented in Table 2. To facilitate intuitive comprehension, we plot the correlation between the normalized average score on benchmarks and the number of ShortV layers in Figure 1b." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "We observe that ShortV can freeze visual tokens in approximately " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the MLLM layers while preserving superior performance. As " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " continues to increase, both 7B and 13B models can maintain more than " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " performance when the hidden states of visual tokens remain unchanged in about " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the layers. These results are significantly different from those on text-only LLMs. For LLMs, Men et al. [36] remove transformations on text tokens in approximately " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " of the LLM layers, and this results in about " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " performance degradation on language benchmarks. As the number of layers increases, the performance of LLMs rapidly declines. These differences align with our observation in Section 2.3 that layers are more ineffective for visual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "334" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 111 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 111 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 111 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 111 + ], + "type": "table", + "html": "
# ShortV Layers (N)08161924
LLaVA-1.5-7B1.00×1.13×1.23×1.30×1.40×
LLaVA-NeXT-7B1.00×1.15×1.35×1.44×1.64×
", + "image_path": "138d2b7ca9a58584623247a9bd6f74bc4d3b113c7bfc64a0b81f02bab0f974b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 159, + 294, + 201 + ], + "blocks": [ + { + "bbox": [ + 56, + 119, + 295, + 142 + ], + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 142 + ], + "type": "text", + "content": "Table 3. Inference speedups over the vanilla models, based on the 7B models. We conduct this test on a single GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 159, + 294, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 159, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 294, + 201 + ], + "type": "table", + "html": "
# ShortV Layers (N)08162432
LLaVA-1.5-13B1.00×1.13×1.24×1.39×1.50×
LLaVA-NeXT-13B1.00×1.13×1.30×1.52×1.84×
", + "image_path": "49e39406ce903f452cfbd5f98f406577985315613f5643718061cbfd6c54d72e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 250, + 294, + 325 + ], + "blocks": [ + { + "bbox": [ + 56, + 209, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 232 + ], + "type": "text", + "content": "Table 4. Inference speedups over the vanilla models, based on the 13B models. We conduct this test on a single GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 250, + 294, + 325 + ], + "lines": [ + { + "bbox": [ + 58, + 250, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 58, + 250, + 294, + 325 + ], + "type": "table", + "html": "
MethodFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
FastV58%64.335.865.460.2
ShortV55%64.836.266.260.9
ShortV+FastV29%64.237.165.159.3
", + "image_path": "70cfe4b23769aa0153c4ce72efda1fe1e56b363a61580d5beea6607f3530f099.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 396, + 167, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 167, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 167, + 406 + ], + "type": "text", + "content": "tokens than for text tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": "In addition to the theoretical FLOPs ratios, we provide the speedups on real hardware using different settings, as shown in Table 3 and Table 4. To get rid of the influence of different output sequence lengths, we use the first token latency to calculate the speedups. We utilize the MMMU dataset for the latency test. For comparison, we note that FastV [10] with its default setting, i.e. " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "R = 50\\%" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": ", achieves a " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "1.31 \\times" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": " speedup over the vanilla LLaVA-1.5-13B model. In contrast, our ShortV with its default parameter, i.e. " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "N = 24" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": ", achieves a greater speedup of " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "1.39 \\times" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 217, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 217, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 217, + 562 + ], + "type": "text", + "content": "4.4. Orthogonal to Token Pruning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this section, we demonstrate that ShortV is orthogonal to and compatible with visual token pruning, e.g. FastV [10]. FastV identifies " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "R\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " unimportant visual tokens and drops them after the " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "-th layer in the forward process of input tokens. We apply FastV to ShortV, which already replaces " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " ineffective layers for visual tokens with ShortV layers. We use the default settings for FastV and ShortV in this experiment, i.e. " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "R = 50\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " for FastV, and " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "N = 19" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " for ShortV. We employ LLaVA-1.5-7B as the vanilla model. The experimental results in Table 5 demonstrate that ShortV is compatible with FastV and that the application of both can further improve MLLM efficiency." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 365 + ], + "type": "text", + "content": "Table 5. ShortV is compatible with FastV, and applying both at the same time can further enhance MLLM efficiency. This experiment is based on LLaVA-1.5-7B." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "type": "table", + "html": "
StrategyFLOP RatioMMBench ENMMMU (val)SEED-BenchGQA
Vanilla100%64.036.366.161.9
Random55%58.433.660.556.1
Cosine Sim.55%60.834.262.759.5
LC (Ours)55%64.836.266.260.9
", + "image_path": "dae87ed262657bb97eda50243d7710185adf36e8ef0c90f8ddcb840d6c283175.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "type": "text", + "content": "Table 6. Ablation on strategies to select replaced layers, based on LLaVA-1.5-7B. \"Random\" denotes randomly selecting 19 layers and freezing visual tokens in them. \"Cosine Sim.\" denotes using cosine similarity to select ineffective layers for visual tokens." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 222, + 415, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 222, + 415, + 233 + ], + "spans": [ + { + "bbox": [ + 314, + 222, + 415, + 233 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 550 + ], + "type": "text", + "content": "Ablation on strategies to select replaced layers. In this paragraph, we perform an ablation experiment on LLaVA-1.5-7B to investigate the impact of strategies for selecting which layers to replace. ShortV selects ineffective layers for visual tokens, and replace them with ShortV layers. To identify which layers are ineffective, we utilize the LC metric introduced in Section 2. In contrast, previous work [36] on text-only LLMs uses a metric based on cosine similarity. It calculates the average cosine similarity between the inputs and outputs of each layer. The layers with higher cosine similarities are deemed more ineffective. To make a comparison between this cosine similarity metric and our LC metric, we calculate each layer's cosine similarity between the input hidden states and output hidden states of visual tokens, and select the same number of layers with the highest cosine similarities. We show the comparison in Table 6. We also include the results of another baseline, ShortV (Random), where visual tokens are frozen in the same number of randomly selected layers. These results clearly demonstrate that our LC metric performs better than cosine similarity in identifying ineffective MLLM layers for visual tokens, and ShortV based on the LC metric achieves performance comparable to the vanilla model. In contrast, ShortV based on the cosine similarity metric cannot match the performance of the vanilla model, although it outperforms the baseline with randomly selected layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Ablation on frozen tokens. In this paragraph, we conduct an ablation study on LLaVA-1.5-7B to investigate the impact of freezing different types of tokens. In Section 2, we demonstrate that MLLM layers are ineffective for visual tokens, as measured by the LC metric. Motivated by this observation, ShortV freezes visual tokens in ineffective layers. In Table 7, we compare our method with the strategies of freezing other tokens. In the experiment detailed in line (a), we utilize the LC metric to identify 19 ineffective layers for text tokens, and freeze text tokens in them. Despite having fewer frozen tokens, we can observe that this strategy results in significant performance declines compared with our method, which freezes visual tokens rather than text to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "335" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 279, + 152 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 279, + 152 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 279, + 152 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 279, + 152 + ], + "type": "table", + "html": "
Frozen TokensMMBench ENMMMU (val)SEED-BenchGQA
None (Vanilla)64.036.366.161.9
(a) Text2.123.78.92.9
(b) Text+Visual1.326.60.80.0
(c) Random1.522.95.52.3
(d) Visual (Ours)64.836.266.260.9
", + "image_path": "106e4dbd13532c8635ff48687f165b7c0154c6919609d4cb9f13163d284001a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 160, + 295, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 295, + 226 + ], + "type": "text", + "content": "Table 7. Ablation on frozen tokens, based on LLaVA-1.5-7B. (a) identifying 19 ineffective layers for text tokens and freezing text tokens in them. In lines (b) and (c), we select ineffective layers for all tokens. line (b) involves freezing all input tokens in them, whereas line (c) denotes randomly freezing the same number of tokens as the visual tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 247, + 295, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 295, + 414 + ], + "type": "text", + "content": "kens. These experimental results align with our findings in Section 2 that MLLM layers are more ineffective for visual tokens than for text tokens. In lines (b) and (c), we first calculate each layer's average LC score for all tokens, including visual and text tokens, and then select 19 ineffective layers. Next, in the experiment corresponding to line (b), we freeze all input tokens in these layers. In line (c), we freeze random input tokens, and the number of frozen tokens matches that of the visual tokens. As a result, the computational overhead associated with line (c) is the same as that of our method in line (d). We can find that freezing tokens other than visual tokens leads to substantial performance degradation in vision-language tasks. These ablations demonstrate the effectiveness of our ShortV." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 142, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 142, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 142, + 437 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 250, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 250, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 250, + 457 + ], + "type": "text", + "content": "5.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "text", + "content": "Built upon Large Language Models (LLMs) [1, 12, 37, 47, 49], Multimodal Large Language Models (MLLMs) [4, 7, 35, 48, 52, 57] have made significant progress in processing and understanding the visual world. Typically, they use a decoder-only architecture. Specifically, they utilize visual encoders [14, 21, 34, 41, 56] to convert input visual information into visual features and then use projectors to project these visual features into visual tokens. These visual tokens are then concatenated with text tokens and fed into the LLM backbones. Current MLLMs use hundreds to thousands of visual tokens to represent a single image, significantly increasing the length of the token sequences. For instance, the LLaVA-1.5 models [31] transform each image with " + }, + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "336 \\times 336" + }, + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "text", + "content": " resolution into 576 tokens. For images with higher resolutions, the LLaVA-NeXT models [32] process images into up to 2,880 visual tokens, and SPHINX-2k [28] divides one image into nine subimages, resulting in 2,890 visual tokens. Applying LLMs to such large numbers of visual tokens incurs substantial computational costs. In this paper, we introduce ShortV to enhance the efficiency of MLLMs by reducing the computational overhead associ" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 411, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 411, + 83 + ], + "type": "text", + "content": "ated with visual tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 95, + 471, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 471, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 471, + 106 + ], + "type": "text", + "content": "5.2. Efficient LLMs and MLLMs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "text", + "content": "For LLMs, previous studies [36, 46] find that layers in LLMs are ineffective for text tokens. They remove computations in about " + }, + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "text", + "content": " of the layers, while preserving approximately " + }, + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 114, + 553, + 173 + ], + "type": "text", + "content": " of the performance. LaCo [50] utilizes layer merging for efficient LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 376 + ], + "type": "text", + "content": "To address the computational inefficiency of MLLMs, previous methods [10, 29, 42, 44, 45, 54] have primarily focused on two aspects: efficient model architecture and visual token compression. Among efficient model architectures, cross-attention-based models [2, 3, 8, 20] insert gated cross-attention layers within LLM layers for visual perception, but previous studies [13, 24] demonstrate that this architecture performs worse than the decoder-only architecture in the same settings. Instead of inserting cross-attention layers, mPLUG-Owl3 [51] and Vamba [42] introduce cross-attention operations in parallel with self-attention. In contrast, SAISA [54] introduces NAAViT self-attention blocks, which incorporate multimodal cross-attention into the original self-attention operations of the LLMs, and reuse the parameters of self-attention blocks. The design of ShortV layers is inspired by NAAViT. Differently, in ShortV layers, visual tokens also skip their FFNs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 559 + ], + "type": "text", + "content": "Visual token compression methods improve MLLM efficiency in both training-based [5, 6, 22, 26, 27] and training-free [10, 43, 53] manners. FastV [10] reveals token-wise redundancy, and it removes unimportant tokens during inference. In this paper, we reveal layer-wise redundancy in MLLMs. Layers in MLLMs are much more ineffective for visual tokens than for text token. Therefore, we can freeze visual tokens in approximately " + }, + { + "bbox": [ + 313, + 380, + 553, + 559 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 380, + 553, + 559 + ], + "type": "text", + "content": " of the MLLM layers with minimal performance degradation. Unlike previous methods for MLLM efficiency, ShortV does not reduce the number of visual tokens but instead decreases the computational costs of processing each token. ShortV is training-free and orthogonal to token compression. We demonstrate that ShortV is compatible with FastV, allowing for simultaneous application to further enhance MLLM efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we explore the layer-wise redundancy in MLLMs. We discover that layers in MLLMs are more ineffective for visual tokens than for text tokens. MLLM layers' transformations on visual tokens have a minimal impact on the MLLM output. Motivated by this observation, we propose ShortV, a training-free method to enhance MLLM efficiency. ShortV utilizes our proposed LC metric to select ineffective layers for visual tokens, and freezes visual tokens in these layers. It can freeze visual tokens in about " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " of the layers while preserving superior performance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "336" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "content": "We sincerely thank the reviewers for their insightful comments and valuable suggestions. This work was supported by Beijing Natural Science Foundation (L243006), Beijing Municipal Science and Technology Project (Nos. Z231100010323002), the Natural Science Foundation of China (No. 62306303, 62476265) and the Basic Research Program of ISCAS (Grant No. ISCAS-ZD-202401)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 208, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 208, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 208, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 208, + 295, + 262 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 295, + 327 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "type": "text", + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 294, + 427 + ], + "type": "text", + "content": "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Tasirlar. Introducing our multimodal models, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 295, + 482 + ], + "type": "text", + "content": "[5] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 483, + 295, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 295, + 537 + ], + "type": "text", + "content": "[6] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jeng-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 538, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 295, + 602 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 295, + 657 + ], + "type": "text", + "content": "[8] Kaibing Chen, Dong Shen, Hanwen Zhong, Huasong Zhong, Kui Xia, Di Xu, Wei Yuan, Yifei Hu, Bin Wen, Tianke Zhang, et al. Evlm: An efficient vision-language model for visual understanding. arXiv preprint arXiv:2407.14177, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "type": "text", + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[11] Xiaodong Chen, Yuxuan Hu, Jing Zhang, Yanling Wang, Cuiping Li, and Hong Chen. Streamlining redundant layers to compress large language models. arXiv preprint arXiv:2403.19135, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "text", + "content": "[12] Wei-Lin Chiang, Zhuohan Li, Ziqing Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 1, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 296 + ], + "type": "text", + "content": "[13] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024.8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 298, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 363 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2, 5, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 419 + ], + "type": "text", + "content": "[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 476 + ], + "type": "text", + "content": "[16] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 510 + ], + "type": "text", + "content": "[17] Shwai He, Guoheng Sun, Zheyu Shen, and Ang Li. What matters in transformers? not all attention is needed. arXiv preprint arXiv:2406.15786, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[18] Qidong Huang, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Deciphering cross-modal alignment in large vision-language models with modality integration rate. arXiv preprint arXiv:2410.07167, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 622 + ], + "type": "text", + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[20] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https://huggingface.co/blog/idefics, 2023.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[21] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 8" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "337" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[22] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[23] Bo-Kyeong Kim, Geonmin Kim, Tae-Ho Kim, Thibault Castells, Shinkook Choi, Junho Shin, and Hyoung-Kyu Song. Shortened llama: Depth pruning for large language models with comparison of retraining methods. arXiv preprint arXiv:2402.02834, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[24] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? Advances in Neural Information Processing Systems, 37:87874-87907, 2025. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[25] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[26] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[27] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 421 + ], + "type": "text", + "content": "[28] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "text", + "content": "[29] Zhihang Lin, Mingbao Lin, Luxi Lin, and Rongrong Ji. Boosting multimodal large language models with visual tokens withdrawal for rapid inference. arXiv preprint arXiv:2405.05803, 2024. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 4, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[33] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[35] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024.8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[36] Xin Men, Mingyu Xu, Qingyu Zhang, Bingning Wang, Hongyu Lin, Yaojie Lu, Xianpei Han, and Weipeng Chen. Shortgpt: Layers in large language models are more redundant than you expect, 2024. URL https://arxiv.org/abs/2403.03853, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "text", + "content": "[37] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2(5):6, 2024. 1, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[38] OpenAI. Gpt-4v(ison) system card. https://cdn.openai.com/papers/GPTV_System/Card.pdf, 2023.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[39] OpenAI. Introducing gpt-4o: our fastest and most affordable flagship model. https://platform.openai.com/docs/guides/vision, 2024. Accessed: 2024-05-26. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 338 + ], + "type": "text", + "content": "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[42] Weiming Ren, Wentao Ma, Huan Yang, Cong Wei, Ge Zhang, and Wenhu Chen. Vamba: Understanding hour-long videos with hybrid mamba-transformers. arXiv preprint arXiv:2503.11579, 2025. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[43] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "type": "text", + "content": "[44] Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18221-18232, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[45] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 656 + ], + "type": "text", + "content": "[46] Jiwon Song, Kyungseok Oh, Taesu Kim, Hyungjun Kim, Yulhwa Kim, and Jae-Joon Kim. Sleb: Streamlining llms through redundancy verification and elimination of transformer blocks. arXiv preprint arXiv:2402.09025, 2024. 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[47] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "338" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 586 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[48] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Song XiXuan, et al. Cogvlm: Visual expert for pretrained language models. Advances in Neural Information Processing Systems, 37:121475-121499, 2025. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[49] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 206 + ], + "type": "text", + "content": "[50] Yifei Yang, Zouying Cao, and Hai Zhao. Laco: Large language model pruning via layer collapse. arXiv preprint arXiv:2402.11187, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "type": "text", + "content": "[51] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. MPLUG-owl3: Towards long image-sequence understanding in multimodal large language models. In The Thirteenth International Conference on Learning Representations, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 329 + ], + "type": "text", + "content": "[52] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[53] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 374, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 418 + ], + "type": "text", + "content": "[54] Qianhao Yuan, Yanjiang Liu, Yaojie Lu, Hongyu Lin, Ben He, Xianpei Han, and Le Sun. Saisa: Towards multimodal large language models with both training and inference efficiency. arXiv preprint arXiv:2502.02458, 2025. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 496 + ], + "type": "text", + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "text", + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 294, + 586 + ], + "type": "text", + "content": "[57] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "339" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_content_list.json b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54935e3afb6c9fc4d87182ca68ac7fdc757d242a --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_content_list.json @@ -0,0 +1,1890 @@ +[ + { + "type": "text", + "text": "Shot-by-Shot: Film-Grammar-Aware Training-Free Audio Description Generation", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 151 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e2ecb7a89bd739996aaa3094b2113e7a0ecdfd23db37b1591318beec45e157a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 176, + 823, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/460a0df0f9a1e2a515e76ad3232416449c89f32d6817766767c05d63396397fd.jpg", + "image_caption": [ + "Figure 1. Overview of our training-free Audio Description (AD) generation framework. The input video clip (corresponding to the AD interval) is extended to include adjacent shots, providing richer temporal visual context. Each video frame is labeled with its corresponding shot number. The extended clip then undergoes a film grammar prediction process, where the thread structure and shot scales are estimated. The AD is generated in two stages: Stage I utilises the predicted cinematic information as training-free prompt guidance to produce dense video descriptions. Stage II then employs an LLM to generate a summarised AD output." + ], + "image_footnote": [], + "bbox": [ + 122, + 280, + 870, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 527, + 325, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages \"shots\" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure - an action score - specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection.", + "bbox": [ + 88, + 554, + 485, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 834, + 215, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In movies and TV series, Audio Descriptions (ADs) are narrations provided for the visually impaired, conveying visual information to complement the original soundtrack. Their", + "bbox": [ + 89, + 856, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "purpose is to ensure a continuous and coherent narrative flow, enabling audiences to follow the plot effectively. Unlike video captions, ADs are constrained by length, prioritising the most visually salient and story-centric information, such as character dynamics and significant objects, whilst omitting redundant details like background figures or unchanging locations. Additionally, ADs are typically produced by professional narrators in a specific style and format, ensuring coherence while not interfering with the original audio.", + "bbox": [ + 511, + 529, + 906, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advent of Visual-Language Models (VLMs), there has been a growth of interest in automatically generating ADs for both movies [16, 23-25, 45, 72, 86, 89] and TV material [19, 82]. However, as anyone who has ever watched a movie or read a book about film editing knows – the fundamental unit of edited video material is the shot, not the frame [37, 50, 52]. Shots are used to structure the video material, defining the granularity and temporal context through choices in its scale (close-up, long shot, etc.) and its movement (panning, tracking, etc.). The \"film grammar\" is then used to convey meaning through specific editing choices on shot transitions (cuts, fades, etc.), durations, and composition (threads, montage, etc.).", + "bbox": [ + 511, + 662, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches to automating AD generation, and video large language models (VideoLLMs) in general, are not shot-aware, hindering interpretation of edited video material that has frequent shot transitions [68]. In this paper, our principal", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16503", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objective is to incorporate shot information and editing structure into the AD generation process. To this end, we consider temporal context in terms of shots, and take account of the two key properties of thread structure and shot scale.", + "bbox": [ + 89, + 90, + 480, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thread structure identifies the sequence of shots captured with the same camera. An example of a scene with multiple simultaneous threads in space-time is shown in Fig. 3. We develop a robust thread clustering method and use its predictions to guide VLMs in understanding shot-wise relationships.", + "bbox": [ + 89, + 150, + 480, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Shot scale typically implies the type of content of the frame. For instance, close-up shots often highlight characters' facial expressions or gaze interactions, while long shots tend to depict the overall environment and ambience of the scene, as shown in Fig. 4 (top). We leverage this property by developing an off-the-shelf shot scale classifier and employing a scale-dependent prompting strategy to improve the contextual relevance of VLM-generated descriptions.", + "bbox": [ + 88, + 224, + 480, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we develop a training-free AD generation approach. Inspired by AutoAD-Zero [82], as illustrated in Fig. 1, we adopt a two-stage pipeline in which a VideoLLM generates dense text descriptions in the first stage, followed by an LLM that produces the final AD outputs from this text. We improve the effectiveness of training-free approaches by incorporating two key improvements based on: (i) shot-based temporal context; and (ii) the shot scale and thread structure, that are both crucial to cinematic composition and understanding. Previous methods could describe human-object interactions or human-human interactions, such as looks [47], within a frame or shot. With the improvements we introduce, the generated AD also includes such interactions when they are implied by the shot structure.", + "bbox": [ + 89, + 343, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As well as the challenge of generating ADs, another challenge is how to evaluate the predicted ADs. Apart from conventional metrics [8, 44, 55, 71], several new AD metrics [23, 25, 89] have been proposed, including CRITIC [25] measuring who is mentioned in the AD. However, these metrics fail to emphasise character actions – one of the most critical aspects of ADs. To address this, we introduce an action score, assessing whether the predictions accurately capture the actions described in GT AD, independent of character information.", + "bbox": [ + 89, + 534, + 482, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, as ADs are time-limited, there is always a choice on which visual aspect (characters, actions, objects, etc.) should be included. Consequently, a single video clip can have multiple equally valid ADs, each highlighting a different aspect. This observation is supported by the inter-rater agreement experiments in AutoAD-III [25] and previous user studies [78]. Therefore, beyond evaluating single AD performance, we assess our framework's capability as an assistant to generate multiple AD candidates, and report the performance of the selected one AD.", + "bbox": [ + 89, + 666, + 482, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose an enhanced training-free AD generation framework, with the following contributions: (i) We incorporate shot-based temporal context into AD generation via training-free prompting techniques including shot number referral and dynamic frame sampling; (ii) We develop state-of-the-art methods for thread structure and shot scale predictions, and demonstrate that incorporating predicted film grammar knowl", + "bbox": [ + 89, + 799, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge enhances AD generation; (iii) We improve the current AD evaluation by introducing the character-free action score, and a new assistance-oriented evaluation protocol; (iv) Our approach achieves state-of-the-art performance in training-free AD generation, and furthermore, surpasses fine-tuned models on multiple benchmarks. This is the first time a training-free approach has achieved superior performance to fine-tuned methods.", + "bbox": [ + 511, + 90, + 903, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 209, + 643, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio Description generation. Efforts have been made to curate Audio Description (AD) datasets for both movies [25, 67] and TV series [82], with human annotations sourced from platforms such as AudioVault [2].", + "bbox": [ + 511, + 234, + 903, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For automatic AD generation, prior works [19, 23-25, 45, 72] fine-tune pre-trained VLMs on AD annotations to produce descriptions in an end-to-end manner. However, these methods face challenges due to limited high-quality AD annotations and the high computational cost of fine-tuning each new backbone. Alternatively, training-free approaches [16, 82, 86, 89] have gained traction for their scalability and flexibility, allowing customised AD output based on official guidelines [1] or specific needs. Yet, these methods still lag in performance, while our approach is the first to achieve results on par with fine-tuned methods.", + "bbox": [ + 511, + 294, + 905, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of limiting the video input to each AD interval, UniAD [72] and DistinctAD [19] fine-tune VLMs with multiple AD clip inputs to incorporate broader temporal context. In contrast, our method systematically extends AD clip to adjacent shots and introduces a training-free approach that enables pretrained VideoLLMs to better capture localised temporal context.", + "bbox": [ + 511, + 455, + 905, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Film grammar analysis. Prior research has sought to understand film grammar from two major perspectives: (i) intra-shot properties, (ii) shot-wise relationships.", + "bbox": [ + 511, + 545, + 905, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For individual shots, several datasets [5, 31, 36, 62, 65] categorise their characteristics based on camera setups, including shot scales (examples shown in Fig. 4) and camera movements. Correspondingly, various models [15, 42, 46, 62, 70] have been developed for shot type classification.", + "bbox": [ + 511, + 590, + 905, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding shot-wise relationships, a few datasets [5, 10, 56] have been proposed to explore transitions (i.e. cuts) between shots, which have also been leveraged in video content generation [20, 57, 63, 92]. Beyond pairwise shot transitions, research has also investigated longer temporal contexts with thread-based editing structures. Notably, Hoai et al. introduced the Thread-Safe [27] dataset, demonstrating that thread information can enhance action recognition. These structures have also been utilised in video-based face and people clustering [9, 69].", + "bbox": [ + 511, + 664, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While prior work on film grammar has mainly focused on classification and generation tasks, we specifically utilise shot scale and thread structure information to enhance AD generation in movies and TV series.", + "bbox": [ + 511, + 795, + 903, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Dense video captioning. Dense video captioning is closely related to AD generation. Early works [32, 33, 39, 74, 76] in video captioning typically treat event localisation and", + "bbox": [ + 511, + 856, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16504", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "captioning as independent stages, whereas more recent approaches [12, 18, 34, 38, 51, 58, 61, 77, 80, 85, 96, 97] integrate these tasks in an end-to-end manner. Video captioning benchmarks cover a range of domains, including cooking [95], actions [39], movies [64], TV series [41], and open-domain settings [11, 14, 26, 49, 66, 83, 84, 88, 93].", + "bbox": [ + 89, + 90, + 482, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal grounding in VLM. To equip conventional VLMs with temporal grounding capability, some studies [29, 43, 79] generate additional data with enhanced temporal information for fine-tuning. A more common approach explicitly incorporates temporal information into inputs, either by inserting temporal tokens [13, 17, 21, 22, 28, 30, 59, 73] or embedding time information within visual tokens [29, 43, 79].", + "bbox": [ + 89, + 181, + 482, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-free approaches [60, 81, 94] have also been explored for temporally grounded understanding. Notably, a recent work [81] achieves temporal grounding by overlaying frame numbers as visual prompts. Instead of using uniformly sampled timestamps, we leverage the natural shot structures in movies and TV series, adopting them as fundamental units for training-free temporal referral.", + "bbox": [ + 89, + 285, + 483, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Training-free AD generation framework", + "text_level": 1, + "bbox": [ + 89, + 402, + 434, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video clip $\\mathcal{V} = \\{\\mathcal{I}_0,\\dots,\\mathcal{I}_T\\}$ , the task of audio description is to generate a concise narration $\\mathcal{N}$ describing what happens around a given AD interval $[t_A,t_B]$ . In this work, we propose a two-stage framework that leverages VideoLLMs and LLMs to predict ADs in a training-free manner.", + "bbox": [ + 89, + 429, + 483, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Stage I, we employ a VideoLLM that takes a sequence of frames (from multiple shots) as input and generates a dense description $\\mathcal{D}$ , guided by instructions $\\mathcal{P}_{\\text{VideoLLM}}$ :", + "bbox": [ + 89, + 503, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = \\operatorname {V i d e o L L M} (\\mathcal {V}, [ t _ {A}, t _ {B} ], \\mathcal {P} _ {\\text {V i d e o L L M}}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 553, + 482, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Stage II, we then prompt an LLM (with instructions $\\mathcal{P}_{\\mathrm{LLM}}$ ) to extract key information from the dense Stage I description and format it into an AD-style narration $\\mathcal{N}$ :", + "bbox": [ + 89, + 574, + 483, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} = \\operatorname {L L M} \\left(\\mathcal {D}, \\left[ t _ {A}, t _ {B} \\right], \\mathcal {P} _ {\\text {L L M}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 625, + 482, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we focus on enhancing the visual understanding of the Stage I VideoLLM for edited video material, and make three innovations: In Sec. 3.1, we incorporate shot-based temporal context into Stage I visual inputs. In Sec. 3.2, we leverage the thread structure to enrich cross-shot understanding. Finally, in Sec. 3.3, we incorporate shot-scale awareness into Stage I prompt formulation.", + "bbox": [ + 89, + 646, + 483, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Leveraging shot-based temporal context", + "text_level": 1, + "bbox": [ + 89, + 761, + 434, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regarding the visual inputs to VideoLLM, prior works [25, 45] often sample frames $\\{\\mathcal{I}_{t_A},\\dots,\\mathcal{I}_{t_B}\\}$ that directly correspond to the AD interval $[t_A,t_B]$ . However, this approach can be problematic due to (i) misalignment between the AD interval and the actual timestamps when the action occurs, and (ii) the lack of contextual information from adjacent shots. Therefore, we investigate how incorporating temporal context can enhance the understanding of the video clip.", + "bbox": [ + 89, + 782, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/66f272fa2845c90db118b889236117301ec8842351092366ffc8e8c1f940bba1.jpg", + "image_caption": [ + "Figure 2. Shot-based temporal context, where current shots are defined as those temporally overlapping with the AD interval. Past shots and future shots provide extended contextual information. Shot numbers are visually overlaid on the top-left of each frame, and frames within the AD interval are sampled more densely than context frames." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Structuring temporal context with shots. To obtain more visual context information, instead of simply extending the AD interval by fixed timestamps, we explore a more structured approach that treats shots as the fundamental units.", + "bbox": [ + 511, + 297, + 903, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we apply an off-the-shelf shot segmentation model to partition the entire video clip into individual shots. For each AD interval, we first identify the shots that (partially) overlap with it, referred to as \"current shots\", as illustrated in Fig. 2. We then consider at most two \"past shots\" and two \"future shots\" adjacent to current shots as temporal context. For all shots included, we label them sequentially from past to future with a number starting from \"Shot $O$ \".", + "bbox": [ + 511, + 356, + 906, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emphasising the targeted AD interval. To ensure that the VideoLLM focuses on describing visual content within the AD interval, we propose two strategies: dynamic frame sampling and shot number referral.", + "bbox": [ + 511, + 474, + 905, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic frame sampling. As shown in Fig. 2, to emphasise the frames of interest, we adopt denser sampling within the AD interval (red region) and sparser sampling for the surrounding context frames (yellow region). In practice, we specify fixed numbers of frames to be sampled within and outside the AD interval and apply uniform sampling according to these constraints.", + "bbox": [ + 511, + 535, + 906, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shot number referral. To further enhance the attention towards the current shot content, we label each sampled frame with its shot number (e.g. \"Shot 0\") at the top-left. During the formulation of the text prompt, instead of prompting the VideoLLM to \"describe what happened in the video clip\", we ask it to \"describe what happened in [Shot 2, Shot 3]\" (i.e. current shots). Through this visual-textual prompting strategy, we found that the VideoLLM could successfully interpret the meaning of shot numbers and refer to the correct shots.", + "bbox": [ + 511, + 623, + 908, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Leveraging thread structure", + "text_level": 1, + "bbox": [ + 511, + 762, + 767, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Movies are generally edited such that viewpoints from two or more cameras are intertwined in shot threads, as illustrated in Fig. 3 (top). These interleaved arrangements of shot threads often imply relationships between objects and characters (e.g. gaze interactions) and their 3D arrangement. To leverage this information for Stage I description, we first determine the thread structure using a separate module, then incorporate it into the VideoLLM through prompt guidance.", + "bbox": [ + 511, + 784, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16505", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/415d9bcd947817223b459873e0102686dfe077fe8c3128a52cb332ab845236b8.jpg", + "image_caption": [ + "Figure 3. Thread structure. Top: Example of a thread structure with interleaving shots. Bottom: A training-free approach for thread clustering, where shots are pairwise compared using dense feature matching to construct an adjacency matrix, which is then used to predict the thread structure." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thread structure prediction. We develop a training-free method to predict thread clustering. The problem here is to determine if two shots correspond to the same viewpoint or not. Specifically, given two shots (Shot $i$ and a later Shot $j$ ), we compare the last frame of Shot $i$ $(\\mathcal{I}_{T_i}^i)$ with the first frame of Shot $j$ $(\\mathcal{I}_0^j)$ , extracting their DINoV2 [54] features as $f_{T_i}^i$ and $f_0^j \\in \\mathbb{R}^{h \\times w \\times c}$ .", + "bbox": [ + 89, + 446, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [35], we assess frame-wise dense correlations by computing a cost volume between their feature maps:", + "bbox": [ + 89, + 556, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {p, q} ^ {i, j} = m (p, q) \\circ \\left(\\hat {f} _ {T _ {i}; p} ^ {i} \\cdot \\hat {f} _ {0; q} ^ {j}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 585, + 482, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{f}_{T_i;p}^i$ and $\\hat{f}_{0;q}^{j}\\in \\mathbb{R}^{c}$ are the normalised $p$ -th and $q$ -th spatial elements in the respective feature maps. The binary attention mask $m(p,q)$ is set to 1 only if the spatial position of the $q$ -th element is within an $n\\times n$ neighbourhood of the $p$ -th element.", + "bbox": [ + 89, + 609, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then apply a softmax operation along the last dimension $(q)$ and find the maximum similarity for each $p$ , followed by averaging over all $p$ -th elements to obtain a matching score between Shot $i$ and Shot $j$ :", + "bbox": [ + 89, + 671, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {i, j} = \\frac {1}{N} \\sum_ {p} ^ {N} \\max _ {q} \\left(\\frac {\\exp \\left(\\mathcal {C} _ {p , q} ^ {i , j} / \\tau\\right)}{\\sum_ {l} ^ {N} \\exp \\left(\\mathcal {C} _ {p , l} ^ {i , j} / \\tau\\right)}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 734, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ denotes the number of feature patches, and $\\tau$ is the softmax temperature.", + "bbox": [ + 89, + 780, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, as shown in Fig. 3, this process effectively checks whether each patch $(p)$ in one shot frame matches with a patch $(q)$ in the other shot frame at a roughly similar spatial position (i.e., within an $n\\times n$ neighbourhood).", + "bbox": [ + 89, + 811, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we construct an adjacency matrix based on the predicted scores $s^{i,j}$ for all possible pairs of Shot $i$ and Shot", + "bbox": [ + 89, + 871, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ff0827dcc9824848c65dfe9348fba3c429bc6c7facff7938dfc0ca8a9ab5193.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 895, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac35de672e581b6d0d470dc8ba4e94bcc1817e4ea72e254c4f0725162a56dd7d.jpg", + "image_caption": [ + "Figure 4. Shot scales. Top: Examples of five different shot scales. Bottom: Stage I factor composition based on shot scale classification. The scales of the current shots in the clip are predicted and averaged. The resulting effective shot scale then guides the formulation of the Stage I prompt, incorporating additional factors such as facial expressions, etc." + ], + "image_footnote": [], + "bbox": [ + 516, + 157, + 895, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$j$ . By thresholding the adjacency matrix (threshold $\\epsilon$ ) and identifying the largest connected components, we cluster the set of shots into multiple threads.", + "bbox": [ + 511, + 369, + 906, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thread structure injection. Once the thread information is obtained, we inject it into the text prompt $\\mathcal{P}_{\\mathrm{VideoLLM}}$ for the Stage I VideoLLM. In practice, we conduct this information injection only to clips that exhibit thread structures (i.e. $N_{\\mathrm{thread}} < N_{\\mathrm{shot}}$ ). For each given thread [Shot $i,\\ldots ,$ Shot $j$ ], we formulate the prompt as: \"[Shot $i,\\ldots ,$ Shot $j$ ] share the same camera setup\". This statement implies that the cameras in these shots maintain consistent angles and scales.", + "bbox": [ + 511, + 415, + 905, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than simply providing this information, we further engage the VideoLLM by asking it to explain why the given thread structure is correct. This effectively corresponds to a Chain-of-Thought (CoT) process, enhancing its understanding of these repetitive thread structures.", + "bbox": [ + 511, + 531, + 905, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Leveraging shot scale information", + "text_level": 1, + "bbox": [ + 511, + 612, + 808, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In movies and TV series, shot scales are often carefully designed during filming or post-editing to implicitly convey information to the audience. Our objective is to use the shot scale to choose what should be included in the Stage I prompts. Shot scale classification. We first build a shot scale classification network by fine-tuning a pre-trained DINOv2 [54] model. For all current shots, we classify their shot scales $\\{S_i\\}$ into one of five classes, represented by values 0-4, as shown in Fig. 4 (top). We then compute their average to obtain the effective shot scale $S_{\\mathrm{eff}}$ .", + "bbox": [ + 511, + 633, + 906, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage I factor composition. We then leverage the predicted shot scale to determine the factors to include in Stage I instructions. We first consider three fixed factors that form the basis of ADs, namely characters, actions, and interactions. The additional factor can be determined through applying a set of thresholds to $S_{\\mathrm{eff}}$ , as detailed in Fig. 4 (bottom). For example, for close-up shots ( $S_{\\mathrm{eff}} \\leq 1.5$ ), we would ask the VideoLLM to additionally describe the \"facial expression\", whereas for long shots", + "bbox": [ + 511, + 784, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16506", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6d87d3d749e1e68e37f4e0747526850f2c1a3d038f803923cdedaeaf3de4d644.jpg", + "image_caption": [ + "Figure 5. Action score. Left: Hierarchical prediction parsing. Given a predicted paragraph or sentence, it is first divided into a set of sentences (light blue) and action phrases (dark blue). Each action phrase is further processed to obtain its verb lemma (light orange). Middle: Scoring based on semantic similarity. Sentence embeddings for the GT action sentence and the set of predicted sentences and phrases are extracted, with the maximum cosine similarity defined as the similarity score. Right: Scoring based on verb matching. The predicted verb lemma is compared with the GT lemma and multiplied by the corresponding semantic-based similarity. The highest resultant score is defined as the verb score." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 450, + 213 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ffd3453db4326325f47c85581c0dba5a5bc2449893575bff8be90fdc09fc33be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 88, + 857, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$(S_{\\mathrm{eff}} \\geq 3.5)$ , a description on the \"environment\" will be included.", + "bbox": [ + 89, + 297, + 485, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Action score", + "text_level": 1, + "bbox": [ + 89, + 327, + 212, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce a new metric, termed \"action score\", which focuses on whether a specific ground truth (GT) action is captured within a predicted Stage I description paragraph or a Stage II AD output. For instance, for the GT action \"He dials the phone\", we want the metric to measure the action of \"dial the phone\", but not be sensitive to character names and other predicted content. Therefore, this metric is designed to possess two key properties: (i) it is character-free, meaning that the presence of character names has minimal impact on the evaluation, and (ii) it is recall-oriented, without penalising additional action information in the prediction.", + "bbox": [ + 88, + 351, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preprocessing of GT actions. For each GT AD, we extract the character-free GT actions by (i) replacing character names with pronouns, and (ii) splitting the AD into subsentences, each containing one action (verb). For example, given a GT AD \"Chandler dials the phone, then hurriedly hangs up,\" the extracted GT actions are \"He dials the phone\" and \"He hurriedly hangs up.\"", + "bbox": [ + 89, + 513, + 485, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical parsing of predictions. To process the predicted paragraph during evaluation, we first decompose it into individual sentences (light blue), as shown in the left column of Fig. 5. For each predicted sentence, we then perform rule-based dependency parsing to obtain action phrases (deep blue). Then, all extracted sentences and action phrases are collected to form a prediction set. Additionally, for each action phrase, we extract the corresponding verb lemma (i.e. the root form of verbs).", + "bbox": [ + 89, + 604, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Action score computation. The action score features a combination of semantic-based and verb-matching-based components, as illustrated in the middle and right columns of Fig. 5, respectively. Semantic-based evaluation. For each GT action sentence, we assess its semantic similarity with the set of predicted sentences and phrases. Specifically, we employ a general text embedding (GTE) model to compute sentence-level embeddings for the GT action sentence $e_{\\mathrm{GT}}$ and each element in the predicted set $\\{e_{\\mathrm{pred};i}\\}$ . By computing cosine similarities and taking the maximum value, the similarity score is defined as $s_{\\mathrm{sim}} = \\max_i(s_{\\mathrm{sim};i}) = \\max_i[(e_{\\mathrm{GT}} \\cdot e_{\\mathrm{pred};i}) / (|e_{\\mathrm{GT}}||e_{\\mathrm{pred};i}|)]$ .", + "bbox": [ + 89, + 723, + 485, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Verb-matching-based evaluation. In addition to semantic", + "bbox": [ + 89, + 886, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matching, we credit predictions that contain the same verbs as those in GT actions. Practically, we compare the predicted verb lemmas with GT verb lemma and compute binary matching scores $\\{m_j\\}$ . These matching scores are further weighted by the corresponding similarity scores $\\{s_{\\mathrm{sim},j}\\}$ through element-wise multiplication. Finally, we take the maximum of the regulated scores to compute the verb score, i.e. $s_{\\mathrm{verb}} = \\max_j(s_{\\mathrm{sim};j} \\circ m_j)$ .", + "bbox": [ + 511, + 297, + 906, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the final action score, we combine the scores from both sources using a weighted average $s_{\\text{action}} = \\alpha_{\\text{sim}} \\circ s_{\\text{sim}} + \\alpha_{\\text{verb}} \\circ s_{\\text{verb}}$ , with $\\alpha_{\\text{sim}}$ and $\\alpha_{\\text{verb}}$ denoting the weighted factors.", + "bbox": [ + 511, + 400, + 906, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 453, + 638, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section begins with datasets and metrics for AD generation in Sec. 5.1, followed by implementation details in Sec. 5.2. Results on film grammar predictions are presented in Secs. 5.3 and 5.4, while Sec. 5.5 analyses human alignment with action scores. AD generation results are detailed in Secs. 5.6 to 5.8.", + "bbox": [ + 511, + 478, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. AD generation datasets and metrics", + "text_level": 1, + "bbox": [ + 511, + 558, + 823, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our framework on AD generation datasets for both movies (CMD-AD [25], MAD-Eval [23]) and TV series (TV-AD [82]). In more detail, CMD-AD is constructed by aligning ground truth ADs with the Condensed Movie Dataset (CMD) [7], comprising 101k ADs (94k for training and 7k for testing) from 1.4k movies. MAD-Eval consists of 6.5k ADs sampled from 10 movies within LSMDC [64]. On the other hand, TV-AD features 34k AD annotations from 13 TV series, with its test set sourced from TVQA [40], consisting of 3k ADs.", + "bbox": [ + 511, + 579, + 906, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For AD evaluation, we follow the prior works [19, 25] to assess general prediction quality using CIDEr [71], Recall@k/N [24], and LLM-AD-Eval [25]. We also consider character recognition accuracy using CRITIC [25] and character-free action evaluation using action scores (described in Sec. 4).", + "bbox": [ + 511, + 710, + 906, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 789, + 728, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide key implementation details for AD generation and action score computation. For additional information on film grammar predictions and other details, please refer to the Supp. Mat.", + "bbox": [ + 511, + 810, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shot detection. To segment the video clip into shots, we use PySceneDetect [3] with the \"Adaptive Detection\" method,", + "bbox": [ + 511, + 871, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16507", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d80fc5e62182020e6096a5a717b58b384de2b55826c2ab7eadb1627a6f779de3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FeatureFrame setupPrecisionRecallAPWCP
CLIP-L14 CLSSide0.6910.6350.7050.922
DINOv2-L14 CLSSide0.7590.6830.7880.933
DINOv2-g14 CLSSide0.7610.6750.7860.936
DINOv2-g14 spatialMid0.8080.7170.8220.953
DINOv2-g14 spatialAll0.8700.7950.8960.964
DINOv2-g14 spatialSide0.8780.7990.9020.965
", + "bbox": [ + 94, + 88, + 480, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/43a256f9c707d074117de6aff5d021d42f15a739aee0923a616c2f62314bafee.jpg", + "table_caption": [ + "Table 1. Thread structure prediction on Thread-Safe. \"Side\" refers to comparisons between the temporally nearest frames in two shots; \"Mid\" refers to comparisons between the middle frames of each shot; \"All\" refers to comparisons between all frame pairs from the two shots." + ], + "table_footnote": [], + "table_body": "
MetricInputAccuracyMacro-F1
ViViT [6]RGB0.7470.751
SGNet [62]RGB + Flow0.875-
Lu et al. [46]RGB + mask0.892-
Li et al. [42]RGB0.8950.897
OursRGB0.8970.899
", + "bbox": [ + 125, + 251, + 444, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which compares the ratio of pixel changes with the neighbouring frames. On average, each shot spans 3.5s, 3.3s, 3.8s in CMD-AD, TV-AD, and MAD-Eval, respectively.", + "bbox": [ + 89, + 367, + 482, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AD generation setting. During dynamic frame sampling, we select a total of 32 frames, with 16 frames uniformly sampled within and outside the AD interval, respectively. For character recognition, we adopt the same visual-textual prompting method proposed by [82], which applies coloured circles around faces for visual character indication. For simplicity, the visualisations in this paper do not display these circle labels.", + "bbox": [ + 89, + 412, + 482, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding the base models, we use Qwen2-VL-7B [75] as the VideoLLM in Stage I and LLaMA3-8B [48] as the LLM in Stage II. This setup is used as the default unless stated otherwise. Additionally, we explore the framework with the proprietary GPT-4o [53] model for both stages.", + "bbox": [ + 89, + 516, + 482, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action score evaluation setting. For action score computation, we set the weight factors as $\\alpha_{\\mathrm{sim}} = 0.8$ and $\\alpha_{\\mathrm{verb}} = 0.2$ . When aggregating the action score results, we first average over multiple GT actions within each GT AD, and then perform global averaging across all AD samples. Moreover, in practice, we find that most action scores are clustered within the range of $0.25 - 0.75$ . To improve clarity, we apply further rescaling $f(x) = (x - 0.25) \\times 2$ as post-processing. For evaluations in this paper, unless otherwise specified, we use the action score to assess the Stage II AD outputs.", + "bbox": [ + 89, + 590, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Thread structure prediction", + "text_level": 1, + "bbox": [ + 89, + 747, + 341, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate thread structure prediction on Thread-Safe [27], which consists of approximately 4.7k video clips collected from 15 TV series. Each video clip contains a multi-shot scene with corresponding thread clusters manually annotated.", + "bbox": [ + 89, + 768, + 482, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluation, we first construct an adjacency matrix from the GT clusters and extract binary GT labels $\\hat{s}_{i,j}$ from the off-diagonal entries, where each label indicates whether a given pair of shots belong to the same cluster. We then compute the Average Precision (AP) between the predicted", + "bbox": [ + 89, + 827, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/01f3ec615666b10400b9d61a1f6e0a226058b40445a260baf60825424dbd7757.jpg", + "table_caption": [ + "Table 2. Shot scale classification on MovieShots." + ], + "table_footnote": [], + "table_body": "
MetricParagraphSentence
PearsonSpearmanPearsonSpearman
CIDEr [71]0.2050.2640.4120.528
ROUGE-L [44]0.3050.2800.5260.512
SPICE [4]0.0220.0480.0310.012
BERTScore [91]0.3770.3930.5080.507
LLM-based (GPT-4o [53])0.7420.6780.7970.807
Action Score (w/o verb matching)0.7350.7280.7650.790
Action Score (w verb matching)0.7490.7290.8060.820
", + "bbox": [ + 517, + 88, + 903, + 208 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of action score with other metrics. The listed metrics measure the similarity between predicted paragraphs/sentences and ground truth actions. The reported values indicate the correlation (i.e. alignment) between these metrics and human-annotated scores.", + "bbox": [ + 511, + 209, + 903, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "relationships $\\{s_{i,j}\\}$ and the ground truth $\\{\\hat{s}_{i,j}\\}$ , as well as report the precision and recall values. Additionally, we directly compare the predicted clusters with ground truth clusters by reporting the weighted clustering purity (WCP) [69].", + "bbox": [ + 511, + 277, + 903, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 verifies the choice of DINOv2-g14 [54] as the feature extractor for frame pair comparison. Compared to abstract CLS tokens, dense spatial matching achieves higher AP in frame pair relationship prediction and higher WCP in thread clustering. Additionally, we observe that using the temporally closest frames from two shots (\"Side\") leads to a noticeable performance improvement. This can be attributed to the continuous story flow across shots within the same thread.", + "bbox": [ + 511, + 338, + 905, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Shot scale classification", + "text_level": 1, + "bbox": [ + 511, + 465, + 728, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following prior work on shot scale classification [42, 46], we use the MovieShots [62] dataset, which consists of 46k shots (train:val:test = 7:1:2) collected from over 7k movie trailers. We follow its definition of shot scales, categorising shots into five classes ranging from extreme close-up to long shots, as illustrated in Fig. 4 (top). To evaluate the model performance, we report classification accuracy and Macro-F1 [42] scores on the MovieShots test set.", + "bbox": [ + 511, + 486, + 905, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since previous state-of-the-art methods on shot scale classification are not open-sourced, we develop a new network by fine-tuning DINOv2 [54], achieving superior performance over prior approaches that rely on additional optical flow or SAM-based mask inputs, as demonstrated in Tab. 2.", + "bbox": [ + 511, + 603, + 905, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Human alignment with action scores", + "text_level": 1, + "bbox": [ + 511, + 686, + 830, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action scores aim to evaluate whether a GT action is captured within a predicted description, making them recall-oriented. Such descriptions can be in the form of paragraphs (Stage I descriptions) or single sentences (Stage II ADs). To assess whether action scores align with human judgments, we create a dataset containing pairs of predicted descriptions and GT actions. For each GT action, human annotators manually annotate the quality of predictions into $\\{0,1,2,3\\}$ based on the relevance towards GT action ranging from \"unrelated\" (0) to \"exact matching\" (3).", + "bbox": [ + 511, + 708, + 906, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with other metrics. Next, we use the human-annotated scores as a reference to compare different metrics in terms of human agreement (measured by correlations), as reported in Tab. 3. Additionally, we consider an LLM-based", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16508", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec1e0e97f8f07974e88e4b67ae75612ae6f8e3d937dddc95914cf7ce833deb96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Exp.Temporal contextFrame samplingShot labelCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
A---22.445.727.026.042.222.3
B1 shotDyn.Shot num.24.746.827.828.941.623.1
C2 shotsDyn.-24.546.627.427.842.322.3
D2 shotsDyn.Frame box24.846.427.425.041.822.8
E2 shotsUni.Shot num.24.147.526.226.542.222.0
F2 shotsDyn.Shot num.25.147.527.828.942.123.0
", + "bbox": [ + 93, + 88, + 482, + 193 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f4fe17fc7c107b99458cfe3f74a3cacb14168313a0faf6465f08bc359da5358d.jpg", + "table_caption": [ + "Table 4. Leveraging shot-based temporal context. Key changes relative to the default setting (Exp. F) are highlighted in light blue. \"Temporal context\" indicates the number of context (past & future) shots. \"Shot num.\" refers to overlaying the shot number at the top-left of each current shot frame, while \"frame box\" represents highlighting the boundary of each current shot frame with a red box." + ], + "table_footnote": [], + "table_body": "
Thread structureCMD-AD subsetTV-AD subset
CIDErCRITICActionCIDErCRITICAction
X29.947.527.428.842.022.6
✓30.7↑0.848.9↑1.427.7↑0.330.7↑1.942.7↑0.722.9↑0.3
", + "bbox": [ + 94, + 285, + 478, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "metric to predict scores, following the same scoring criteria as human annotations. In general, the action score achieves the best correlations with human annotations. Note, it is also more efficient than LLM-based metrics, with 0.15s compared to 6s per prediction evaluation. For more details regarding this human agreement study, please refer to the Supp. Mat.", + "bbox": [ + 89, + 405, + 483, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. AD generation - evaluation of components", + "text_level": 1, + "bbox": [ + 89, + 505, + 452, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shot-based temporal context. We investigate different setups for leveraging temporal context in AD generation, as shown in Tab. 4, leading to the following observations: (i) Expanding the temporal context range noticeably boosts the performance, with gains saturating around “2 shots” (Exp. A, B, and F); (ii) “Shot number referral” is the most effective strategy for outlining the current shot. (Exp. C, D, and F); (iii) Dynamically sampling the current shots at a higher frame rate boosts AD generation (Exp. E and F). The latter two improvements can be attributed to more efficient focus on the visual content around the targeted AD interval.", + "bbox": [ + 89, + 530, + 483, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thread structure injection. After extending the context information with neighbouring shots, we further enhance the VideoLLM's understanding by incorporating thread structures. Note that this guidance is applied only to video sequences exhibiting thread structures. As shown in Tab. 5, this additional information improves AD generation performance across both datasets.", + "bbox": [ + 89, + 693, + 483, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scale-dependent Stage I factors. Tab. 6 explores the impact of different Stage I factors on final AD performance. Using shot scales as guidance for Stage I factor formulation (i.e. scale-dependent) not only outperforms configurations with single fixed factors but also surpasses the case where all factors are included in Stage I. This could be attributed to that the scale-dependent description contains more relevant and less redundant information, enabling more efficient AD extraction in Stage II.", + "bbox": [ + 89, + 784, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e7d2508a9f862be3e8f451eda0354800591fb312753e5aa8b096767fb115f0a8.jpg", + "table_caption": [ + "Table 5. Thread structure injection. Thread structure information is injected only into subsets predicted to exhibit thread structures ( $\\sim 30\\%$ in CMD-AD and $\\sim 60\\%$ in TV-AD)." + ], + "table_footnote": [], + "table_body": "
Stage I factorsCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
Base25.447.427.729.242.023.0
Base + Face (AutoAD-Zero)25.246.827.830.042.322.5
Base + Obj.26.145.827.930.442.922.8
Base + Env.25.246.727.429.840.722.2
Base + Face + Obj. + Env.26.047.227.430.142.722.9
Scale-dependent (Ours)26.347.828.431.142.223.9
", + "bbox": [ + 517, + 88, + 903, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Factors included in Stage I description. Base: character + action + interaction; Face: facial expression; Env.: environment; Obj.: object. \"Scale-dependent\" refers to our approach, which leverages shot scale predictions to determine the relevant factors for each clip.", + "bbox": [ + 511, + 199, + 906, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.7. AD generation - comparison with SotA", + "text_level": 1, + "bbox": [ + 511, + 268, + 849, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 7 provides a comprehensive summary of AD generation performance on CMD-AD and TV-AD, comparing across training-free methods with and without proprietary models, as well as models fine-tuned on human-annotated ADs. Notably, with the same base model setup (Qwen2-VL-7B + LLaMA3-8B), our training-free framework significantly outperforms AutoAD-Zero, primarily due to the usage of temporal context and film grammar information. By incorporating the more powerful GPT-4o models, our performance scales up further, surpassing even existing fine-tuned models. Additionally, we also report our performance on the MAD-Eval benchmark in Supp. Mat.", + "bbox": [ + 511, + 291, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative visualisations. Fig. 6 presents several qualitative examples, where the top two cases illustrate how temporal context information aids in identifying key objects.", + "bbox": [ + 511, + 455, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the bottom-left example, prior methods fail to associate characters, leading to the omission of the man (Alonso). In contrast, our method recognises the thread structure (i.e. [Shot 0, Shot 2], [Shot 1, Shot 3]), which guides the correct prediction of the man's gaze direction towards the lying woman.", + "bbox": [ + 511, + 500, + 905, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The bottom-right example highlights the effectiveness of scale-dependent Stage I factor formulation. AutoAD-Zero, designed to query characters, actions, interactions, and facial expressions, sometimes overlooks environmental details. Our method, in contrast, correctly identifies the shot as a long shot and instructs the VideoLLM to incorporate environmental context, resulting in more accurate scene descriptions. For more visualisations, please refer to the Supp. Mat. and Supp. Videos.", + "bbox": [ + 511, + 573, + 908, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.8. Assisted AD generation", + "text_level": 1, + "bbox": [ + 511, + 702, + 728, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The subjective nature of AD sets a practical limit on metric scores, lower than the theoretical maximum, because human annotators often provide different but valid descriptions. Therefore, beyond enforcing generating a single AD sentence, we also consider employing our framework as an assistant to produce multiple candidate AD sentences.", + "bbox": [ + 511, + 724, + 906, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To standardise such a protocol, we consider five candidate ADs generated by an assistant and employ an expert to select the best one. To effectively benchmark performance against existing GT ADs, we define the \"expert\" as an automatic selection mechanism that chooses the candidate with the highest average CIDEr and action score.", + "bbox": [ + 511, + 813, + 905, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16509", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a6a63a4608aee770187ddf47428a39193bc0f572b92e7621a9ba4a47424de17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVLMLLMTraining -freePropriet. modelCMD-ADTV-AD
CIDErCRITICActionR@1/5LLM-AD-EvalCIDErCRITICActionR@1/5LLM-AD-Eval
AutoAD-II [24]CLIP-B32GPT-2XX13.58.2-26.12.08|------
AutoAD-III [25]EVA-CLIPLLaMA2-7BXX25.032.731.531.22.89|2.0126.128.826.430.12.78|1.99
DistinctAD [19]CLIPAD-B16LLaMA3-8BXX22.7--33.02.88|2.0327.4--32.12.89|2.00
Video-LLaMA [90]Video-LLaMA-7B-✓X4.80.0-22.01.89|------
VideoBLIP [87]VideoBLIP-✓X5.20.0-23.61.91|------
AutoAD-Zero [82]VideoLLaMA2-7BLLaMA3-8B✓X17.743.725.526.92.83|1.9622.639.421.727.42.94|2.00
AutoAD-Zero [82]Qwen2-VL-7BLLaMA3-8B✓X21.944.326.930.83.00|2.2026.441.622.130.43.05|2.27
OursQwen2-VL-7BLLaMA3-8B✓X26.347.828.433.03.15|2.4231.142.223.933.13.09|2.35
AutoAD-Zero [82]GPT-4oGPT-4o✓✓22.445.130.732.93.08|2.4930.944.426.834.73.12|2.57
OursGPT-4oGPT-4o✓✓26.149.132.536.53.17|2.6634.246.527.436.63.12|2.59
", + "bbox": [ + 119, + 88, + 877, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71ca775083e5ae89d3985eab88d55be7abf93be590ebffd34c975c33a71d0780.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 266, + 491, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb346c808e3b929cf4e6eddd051180befeac426aebc5d869f4d791584ce59a41.jpg", + "image_caption": [ + "Figure 6. Qualitative visualisations. Current shots (corresponding to AD intervals) are outlined by red boxes for illustration purposes only. For simplicity, not all context shots are shown. Training-free methods adopt Qwen2-VL + LLaMA3. Examples are taken from The Big Bang Theory (S2E14) (top left), Inferno (2016) (top right), The Asphalt Jungle (1950) (bottom left), and Signs (2002) (bottom right). The top-left example demonstrates the benefits of shot-based temporal context, where the objects (i.e. bills) in Penny's hands are not clearly visible within the AD interval (Shot 1), leading to ambiguous or incorrect predictions by AutoAD-Zero. In contrast, our method successfully identifies the objects from the context shot (Shot 2). The top-right example describing the action of Harry Sims similarly verifies the effectiveness of incorporating context shots." + ], + "image_footnote": [], + "bbox": [ + 127, + 362, + 488, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdb92f3ea35988d1b9d02d4fc9587ec7874849dab8cd92bbe7fb319f4d024979.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 267, + 867, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a01d0a72273ec5c5256f8d69acf912d78237083f84543dd2cd0f1dd9cbcbd63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 362, + 867, + 463 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/04cbc1fcdcc579e4e1a7e5d77d8f69fb36932880e1793209020e606a79b8dd57.jpg", + "table_caption": [ + "Table 7. Quantitative comparison on CMD-AD and TV-AD. For training-free methods, \"VLM\" and \"LLM\" refer to the models used in separate stages, while for fine-tuned models, they denote the pre-trained components within an end-to-end model." + ], + "table_footnote": [], + "table_body": "
MethodCandidate samplingCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
OursSingle AD (Ref.)26.347.828.431.142.223.9
OursIndep. output (p=0.90;τp=0.6)33.349.632.541.244.328.6
OursIndep. output (p=0.95;τp=1.5)37.050.335.145.546.431.5
AutoAD-Zero [82]Joint output31.646.433.843.246.830.1
OursJoint output38.449.235.751.347.331.8
", + "bbox": [ + 93, + 556, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Assisted AD generation results. All methods adopt Qwen2-VL + LLaMA3-8B as base models. The first row provides single AD generation results as references (labelled in gray), the rest rows report the performance of one selected AD out of five candidates. \"Indep. output\" denotes five random independent Stage II runs, with $p$ as the hyperparameter for top-p (nucleus) sampling and ${\\tau }_{p}$ as the sampling temperature. \"Joint output\" generates five ADs simultaneously in a single run.", + "bbox": [ + 89, + 674, + 485, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To develop an AD generation assistant, we fix the Stage I dense descriptions and explore generating multiple candidates in Stage II. This can be achieved by either running Stage II independently five times (termed the \"independent output\" setup) or generating five AD outputs simultaneously within a single run (termed the \"joint output\" setup). As observed in Tab. 8, the assistant-based setup significantly improves upon the single AD performance, highlighting the potential", + "bbox": [ + 89, + 784, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of training-free methods in effectively capturing the desired content for AD generation. Within the \"independent output\" setup, increasing the randomness of sampling (i.e. higher $p$ and $\\tau_p$ ) enhances the quality of the selected AD, owing to greater candidate diversity. Meanwhile, the \"joint output\" setup achieves superior performance, which could be attributed to reduced information redundancy across the simultaneously generated ADs. For additional visualisations, discussions, and detailed text prompts, please refer to the Supp. Mat.", + "bbox": [ + 511, + 560, + 908, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion - summary and limitations", + "text_level": 1, + "bbox": [ + 511, + 702, + 841, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have demonstrated the benefit of shot-based context and film grammar awareness in AD generation – our training-free two-stage framework achieves state-of-the-art performance among all training-free counterparts, even surpassing fine-tuned models on multiple benchmarks.", + "bbox": [ + 511, + 724, + 906, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The current framework has two main limitations: (i) the performance depends on the base VideoLLM, which may occasionally hallucinate details inconsistent with the visual content; and (ii) story-level context is not incorporated into the AD generation process. These limitations could potentially be addressed in future work by improving visual grounding and extending the visual and textual context to include the plot.", + "bbox": [ + 511, + 797, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16510", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research is supported by the UK EPSRC Programme Grant Visual AI (EP/T028572/1), a Royal Society Research Professorship RSRP\\R\\241003, a Clarendon Scholarship, and ANR-21-CE23-0003-01 CorVis.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 162, + 183, + 178 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guidelines for audio descriptors. https://adp.acb.org/guidelines.html.2", + "[2] Audiovault. https://audiovault.net/. 2", + "[3] Pyscenedetect. https://www.scenedetect.com/. 5", + "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 6", + "[5] Dawit Mureja Argaw, Fabian Caba Heilbron, Joon-Young Lee, Markus Woodson, and In So Kweon. The anatomy of video editing: A dataset and benchmark suite for ai-assisted video editing. In ECCV, 2022. 2", + "[6] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 6", + "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5", + "[8] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 2", + "[9] Andrew Brown, Vicky Kalogeiton, and Andrew Zisserman. Face, body, voice: Video person-clustering with multiple modalities. In ICCV 2021 Workshop on AI for Creative Video Editing and Understanding, 2021. 2", + "[10] Boris Chen, Amir Ziai, Rebecca S. Tucker, and Yuchen Xie. Match cutting: Finding cuts with smooth visual transitions. In WACV, 2023. 2", + "[11] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, pages 190-200, 2011. 3", + "[12] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, 2021. 3", + "[13] Shimin Chen, Xiaohan Lan, Yitian Yuan, Zequn Jie, and Lin Ma. Timemarker: A versatile video-llm for long and short video understanding with superior temporal localization ability. arXiv preprint arXiv:2411.18211, 2024. 3", + "[14] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 3", + "[15] Zeyu Chen, Yana Zhang, Suya Zhang, and Cheng Yang. Study on location bias of cnn for shot scale classification. Multimedia Tools and Applications, 2022. 2", + "[16] Peng Chu, Jiang Wang, and Andre Abrantes. LLM-AD: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2", + "[17] Andong Deng, Zhongpai Gao, Anwesa Choudhuri, Benjamin Planche, Meng Zheng, Bin Wang, Terrence Chen, Chen Chen, and Ziyan Wu. Seq2time: Sequential knowledge transfer for" + ], + "bbox": [ + 93, + 186, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video llm temporal grounding. arXiv preprint arXiv:2411.16932, 2024.3", + "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, 2021. 3", + "[19] Bo Fang, Wenhao Wu, Qiangqiang Wu, Yuxin Song, and Antoni B. Chan. DistinctAD: Distinctive audio description generation in contexts. arXiv preprint arXiv:2411.18180, 2024. 1, 2, 5, 8", + "[20] Dennis Fedorishin, Lie Lu, Srirangaraj Setlur, and Venu Govindaraju. Audio match cutting: Finding and creating matching audio transitions in movies and videos. In ICASSP, 2024. 2", + "[21] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 3", + "[22] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. arXiv preprint arXiv:2410.05643, 2024. 3", + "[23] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, 2023. 1, 2, 5", + "[24] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD II: The sequel – who, when, and what in movie audio description. In ICCV, 2023. 5, 8", + "[25] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD III: The prequel - back to the pixels. In CVPR, 2024. 1, 2, 3, 5, 8", + "[26] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In EMNLP, 2018. 3", + "[27] Minh Hoai and Andrew Zisserman. Thread-safe: Towards recognizing human actions across shot boundaries. In ACCV, 2014. 2, 6", + "[28] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing-song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 3", + "[29] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 3", + "[30] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3", + "[31] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2", + "[32] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2", + "[33] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops on Multimodal Learning, 2020. 2", + "[34] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In CVPR, 2024. 3", + "[35] Allan Jabri, Andrew Owens, and Alexei A Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, 2020. 4", + "[36] Xuekun Jiang, Libiao Jin, Anyi Rao, Linning Xu, and Dahua Lin. Jointly learning the attributes and composition of shots for boundary detection in videos. Trans. Multi., 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Steve D. Katz. *Film Directing Shot by Shot: Visualizing from Concept to Screen*. Michael Wiese Productions, ISBN: 1615932976, 2019. 1", + "[38] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In CVPR, 2024. 3", + "[39] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 3", + "[40] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, compositional video question answering. In EMNLP, 2018. 5", + "[41] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 3", + "[42] Yuzhi Li, Feng Tian, Haojun Xu, and Tianfeng Lu. Toward unified and quantitative cinematic shot attribute analysis. *Electronics*, 2023. 2, 6", + "[43] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, 2024. 3", + "[44] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 2, 6", + "[45] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. In ECCV, 2024. 1, 2, 3", + "[46] Fengtian Lu, Yuzhi Li, and Feng Tian. Exploring challenge and explainable shot type classification using sam-guided approaches. Signal, Image and Video Processing, 2024. 2, 6", + "[47] Manuel Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net++: revisiting people looking at each other in videos. IEEE TPAMI, pages 1-1, 2020. 2", + "[48] Meta. The llama 3 herd of models. arXiv preprint arXiv: 2407.21783, 2024. 6", + "[49] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3", + "[50] James Monaco. How to Read a Film: Movies, Media, and Beyond. OUP, ISBN: 0571168973, 2009. 1", + "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, 2019. 3", + "[52] Walter Murch. In the Blink of an Eye: A Perspective on Film Editing. Silman-James Press, ISBN: 1879505622, 2021. 1", + "[53] OpenAI. Gpt-4 technical report. arXiv preprint arXiv: 2303.08774, 2024. 6", + "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4, 6", + "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 2" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2", + "[57] Alejandro Pardo, Fabio Pizzati, Tong Zhang, Alexander Pondaven, Philip Torr, Juan Camilo Perez, and Bernard Ghanem. Matchdiffusion: Training-free generation of match-cuts. arXiv preprint arXiv:2411.18677, 2024. 2", + "[58] Toby Perrett, Tengda Han, Dima Damen, and Andrew Zisserman. It's just another day: Unique video captioning by discriminative prompting. In ACCV, 2024. 3", + "[59] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 3", + "[60] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPRW, 2024. 3", + "[61] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, 2019. 3", + "[62] Anyi Rao, Jiaze Wang, Linning Xu, Xuekun Jiang, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A unified framework for shot type classification based on subject centric lens. In ECCV, 2020. 2, 6", + "[63] Anyi Rao, Xuekun Jiang, Sichen Wang, Yuwei Guo, Zihao Liu, Bo Dai, Long Pang, Xiaoyu Wu, Dahua Lin, and Libiao Jin. Temporal and contextual transformer for multi-camera editing of tv shows. In Proceedings of the CVEU Workshop at ICCV, 2022. 2", + "[64] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 3, 5", + "[65] Mattia Savardi, András Bálint Kovács, Alberto Signoroni, and Sergio Benini. Recognition of camera angle and camera level in movies from single frames. In Proceedings of the ACM International Conference on Interactive Media Experiences Workshops, 2023. 2", + "[66] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting LLMs to transform video annotations at scale. arXiv:2310.04900, 2023. 3", + "[67] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 2", + "[68] Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, and Chenliang Xu. Vidcomposition: Can mllms analyze compositions in compiled videos? arXiv preprint arXiv:2411.10979, 2024. 1", + "[69] Makarand Tapaswi, Omkar M. Parkhi, Esa Rahtu, Eric Sommerlade, Rainer Stiefelhagen, and Andrew Zisserman. Total cluster: A person agnostic clustering method for broadcast videos. In Proceedings of the Indian Conference on Computer Vision Graphics and Image Processing, 2014. 2, 6", + "[70] Bartolomeo Vacchetti and Tania Cerquitelli. Cinematographic shot classification with deep ensemble learning. *Electronics*, 2022, 2", + "[71] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 2, 5, 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16512", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2", + "[73] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videoollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3", + "[74] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, 2018. 2", + "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6", + "[76] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 2", + "[77] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 3", + "[78] Yujia Wang, Wei Liang, Haikun Huang, Yongqi Zhang, Dingzeyu Li, and Lap-Fai Yu. Toward automatic audio description generation for accessible videos. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2", + "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3", + "[80] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, 2024. 3", + "[81] Yongliang Wu, Xinting Hu, Yuyang Sun, Yizhou Zhou, Wenbo Zhu, Fengyun Rao, Bernt Schiele, and Xu Yang. Number it: Temporal grounding videos like flipping manga. In CVPR, 2025. 3", + "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. In ACCV, 2024. 1, 2, 5, 6, 8", + "[83] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 3", + "[84] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3", + "[85] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. arXiv preprint arXiv:2302.14115, 2023. 3", + "[86] Xiaojun Ye, Junhao Chen, Xiang Li, Haidong Xin, Chao Li, Sheng Zhou, and Jiajun Bu. MMAD: Multi-modal movie audio description. In LREC-COLING, 2024. 1, 2", + "[87] Keunwoo Peter Yu. Videoblip, 2023. 8", + "[88] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Farhadi, and Yejin Choi. MERLOT reserve: Multimodal neural script knowledge through vision and language and sound. In CVPR, 2022. 3", + "[89] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. MM-Narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, 2024. 1, 2", + "[90] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In ENNLP, 2023. 8", + "[91] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In *ICLR*, 2020. 6", + "[92] Yuqi Zhang, Bin Guo, Nuo Li, Ying Zhang, Qianru Wang, and Zhiwen Yu. Representation learning of next shot selection for vlog editing. In Proceedings of the CVEU Workshop at ICCV, 2023. 2", + "[93] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 3", + "[94] Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding using large-scale pre-trained models. arXiv preprint arXiv:2408.16219, 2024.3", + "[95] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos In AAAI, 2018. 3", + "[96] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 3", + "[97] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 542 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_model.json b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7bbdac1776fa3685ab7c262f8aba2e9f39cab711 --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_model.json @@ -0,0 +1,2917 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.13, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Shot-by-Shot: Film-Grammar-Aware Training-Free Audio Description Generation" + }, + { + "type": "image", + "bbox": [ + 0.171, + 0.177, + 0.825, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.281, + 0.871, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.443, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 1. Overview of our training-free Audio Description (AD) generation framework. The input video clip (corresponding to the AD interval) is extended to include adjacent shots, providing richer temporal visual context. Each video frame is labeled with its corresponding shot number. The extended clip then undergoes a film grammar prediction process, where the thread structure and shot scales are estimated. The AD is generated in two stages: Stage I utilises the predicted cinematic information as training-free prompt guidance to produce dense video descriptions. Stage II then employs an LLM to generate a summarised AD output." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.528, + 0.326, + 0.543 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.555, + 0.486, + 0.817 + ], + "angle": 0, + "content": "Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages \"shots\" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure - an action score - specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.835, + 0.216, + 0.85 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In movies and TV series, Audio Descriptions (ADs) are narrations provided for the visually impaired, conveying visual information to complement the original soundtrack. Their" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.661 + ], + "angle": 0, + "content": "purpose is to ensure a continuous and coherent narrative flow, enabling audiences to follow the plot effectively. Unlike video captions, ADs are constrained by length, prioritising the most visually salient and story-centric information, such as character dynamics and significant objects, whilst omitting redundant details like background figures or unchanging locations. Additionally, ADs are typically produced by professional narrators in a specific style and format, ensuring coherence while not interfering with the original audio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.909, + 0.839 + ], + "angle": 0, + "content": "With the advent of Visual-Language Models (VLMs), there has been a growth of interest in automatically generating ADs for both movies [16, 23-25, 45, 72, 86, 89] and TV material [19, 82]. However, as anyone who has ever watched a movie or read a book about film editing knows – the fundamental unit of edited video material is the shot, not the frame [37, 50, 52]. Shots are used to structure the video material, defining the granularity and temporal context through choices in its scale (close-up, long shot, etc.) and its movement (panning, tracking, etc.). The \"film grammar\" is then used to convey meaning through specific editing choices on shot transitions (cuts, fades, etc.), durations, and composition (threads, montage, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Current approaches to automating AD generation, and video large language models (VideoLLMs) in general, are not shot-aware, hindering interpretation of edited video material that has frequent shot transitions [68]. In this paper, our principal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16503" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.15 + ], + "angle": 0, + "content": "objective is to incorporate shot information and editing structure into the AD generation process. To this end, we consider temporal context in terms of shots, and take account of the two key properties of thread structure and shot scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.151, + 0.482, + 0.224 + ], + "angle": 0, + "content": "Thread structure identifies the sequence of shots captured with the same camera. An example of a scene with multiple simultaneous threads in space-time is shown in Fig. 3. We develop a robust thread clustering method and use its predictions to guide VLMs in understanding shot-wise relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.226, + 0.482, + 0.342 + ], + "angle": 0, + "content": "Shot scale typically implies the type of content of the frame. For instance, close-up shots often highlight characters' facial expressions or gaze interactions, while long shots tend to depict the overall environment and ambience of the scene, as shown in Fig. 4 (top). We leverage this property by developing an off-the-shelf shot scale classifier and employing a scale-dependent prompting strategy to improve the contextual relevance of VLM-generated descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.483, + 0.533 + ], + "angle": 0, + "content": "In this paper, we develop a training-free AD generation approach. Inspired by AutoAD-Zero [82], as illustrated in Fig. 1, we adopt a two-stage pipeline in which a VideoLLM generates dense text descriptions in the first stage, followed by an LLM that produces the final AD outputs from this text. We improve the effectiveness of training-free approaches by incorporating two key improvements based on: (i) shot-based temporal context; and (ii) the shot scale and thread structure, that are both crucial to cinematic composition and understanding. Previous methods could describe human-object interactions or human-human interactions, such as looks [47], within a frame or shot. With the improvements we introduce, the generated AD also includes such interactions when they are implied by the shot structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.665 + ], + "angle": 0, + "content": "As well as the challenge of generating ADs, another challenge is how to evaluate the predicted ADs. Apart from conventional metrics [8, 44, 55, 71], several new AD metrics [23, 25, 89] have been proposed, including CRITIC [25] measuring who is mentioned in the AD. However, these metrics fail to emphasise character actions – one of the most critical aspects of ADs. To address this, we introduce an action score, assessing whether the predictions accurately capture the actions described in GT AD, independent of character information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.798 + ], + "angle": 0, + "content": "Moreover, as ADs are time-limited, there is always a choice on which visual aspect (characters, actions, objects, etc.) should be included. Consequently, a single video clip can have multiple equally valid ADs, each highlighting a different aspect. This observation is supported by the inter-rater agreement experiments in AutoAD-III [25] and previous user studies [78]. Therefore, beyond evaluating single AD performance, we assess our framework's capability as an assistant to generate multiple AD candidates, and report the performance of the selected one AD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, we propose an enhanced training-free AD generation framework, with the following contributions: (i) We incorporate shot-based temporal context into AD generation via training-free prompting techniques including shot number referral and dynamic frame sampling; (ii) We develop state-of-the-art methods for thread structure and shot scale predictions, and demonstrate that incorporating predicted film grammar knowl" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.193 + ], + "angle": 0, + "content": "edge enhances AD generation; (iii) We improve the current AD evaluation by introducing the character-free action score, and a new assistance-oriented evaluation protocol; (iv) Our approach achieves state-of-the-art performance in training-free AD generation, and furthermore, surpasses fine-tuned models on multiple benchmarks. This is the first time a training-free approach has achieved superior performance to fine-tuned methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.21, + 0.644, + 0.225 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Audio Description generation. Efforts have been made to curate Audio Description (AD) datasets for both movies [25, 67] and TV series [82], with human annotations sourced from platforms such as AudioVault [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.906, + 0.455 + ], + "angle": 0, + "content": "For automatic AD generation, prior works [19, 23-25, 45, 72] fine-tune pre-trained VLMs on AD annotations to produce descriptions in an end-to-end manner. However, these methods face challenges due to limited high-quality AD annotations and the high computational cost of fine-tuning each new backbone. Alternatively, training-free approaches [16, 82, 86, 89] have gained traction for their scalability and flexibility, allowing customised AD output based on official guidelines [1] or specific needs. Yet, these methods still lag in performance, while our approach is the first to achieve results on par with fine-tuned methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Instead of limiting the video input to each AD interval, UniAD [72] and DistinctAD [19] fine-tune VLMs with multiple AD clip inputs to incorporate broader temporal context. In contrast, our method systematically extends AD clip to adjacent shots and introduces a training-free approach that enables pretrained VideoLLMs to better capture localised temporal context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.589 + ], + "angle": 0, + "content": "Film grammar analysis. Prior research has sought to understand film grammar from two major perspectives: (i) intra-shot properties, (ii) shot-wise relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.663 + ], + "angle": 0, + "content": "For individual shots, several datasets [5, 31, 36, 62, 65] categorise their characteristics based on camera setups, including shot scales (examples shown in Fig. 4) and camera movements. Correspondingly, various models [15, 42, 46, 62, 70] have been developed for shot type classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Regarding shot-wise relationships, a few datasets [5, 10, 56] have been proposed to explore transitions (i.e. cuts) between shots, which have also been leveraged in video content generation [20, 57, 63, 92]. Beyond pairwise shot transitions, research has also investigated longer temporal contexts with thread-based editing structures. Notably, Hoai et al. introduced the Thread-Safe [27] dataset, demonstrating that thread information can enhance action recognition. These structures have also been utilised in video-based face and people clustering [9, 69]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.853 + ], + "angle": 0, + "content": "While prior work on film grammar has mainly focused on classification and generation tasks, we specifically utilise shot scale and thread structure information to enhance AD generation in movies and TV series." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Dense video captioning. Dense video captioning is closely related to AD generation. Early works [32, 33, 39, 74, 76] in video captioning typically treat event localisation and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16504" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.179 + ], + "angle": 0, + "content": "captioning as independent stages, whereas more recent approaches [12, 18, 34, 38, 51, 58, 61, 77, 80, 85, 96, 97] integrate these tasks in an end-to-end manner. Video captioning benchmarks cover a range of domains, including cooking [95], actions [39], movies [64], TV series [41], and open-domain settings [11, 14, 26, 49, 66, 83, 84, 88, 93]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.284 + ], + "angle": 0, + "content": "Temporal grounding in VLM. To equip conventional VLMs with temporal grounding capability, some studies [29, 43, 79] generate additional data with enhanced temporal information for fine-tuning. A more common approach explicitly incorporates temporal information into inputs, either by inserting temporal tokens [13, 17, 21, 22, 28, 30, 59, 73] or embedding time information within visual tokens [29, 43, 79]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Training-free approaches [60, 81, 94] have also been explored for temporally grounded understanding. Notably, a recent work [81] achieves temporal grounding by overlaying frame numbers as visual prompts. Instead of using uniformly sampled timestamps, we leverage the natural shot structures in movies and TV series, adopting them as fundamental units for training-free temporal referral." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.404, + 0.436, + 0.423 + ], + "angle": 0, + "content": "3. Training-free AD generation framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.503 + ], + "angle": 0, + "content": "Given a video clip \\(\\mathcal{V} = \\{\\mathcal{I}_0,\\dots,\\mathcal{I}_T\\}\\), the task of audio description is to generate a concise narration \\(\\mathcal{N}\\) describing what happens around a given AD interval \\([t_A,t_B]\\). In this work, we propose a two-stage framework that leverages VideoLLMs and LLMs to predict ADs in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.549 + ], + "angle": 0, + "content": "In Stage I, we employ a VideoLLM that takes a sequence of frames (from multiple shots) as input and generates a dense description \\(\\mathcal{D}\\), guided by instructions \\(\\mathcal{P}_{\\text{VideoLLM}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.554, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = \\operatorname {V i d e o L L M} (\\mathcal {V}, [ t _ {A}, t _ {B} ], \\mathcal {P} _ {\\text {V i d e o L L M}}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.575, + 0.484, + 0.618 + ], + "angle": 0, + "content": "In Stage II, we then prompt an LLM (with instructions \\(\\mathcal{P}_{\\mathrm{LLM}}\\)) to extract key information from the dense Stage I description and format it into an AD-style narration \\(\\mathcal{N}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.626, + 0.483, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} = \\operatorname {L L M} \\left(\\mathcal {D}, \\left[ t _ {A}, t _ {B} \\right], \\mathcal {P} _ {\\text {L L M}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.749 + ], + "angle": 0, + "content": "In this section, we focus on enhancing the visual understanding of the Stage I VideoLLM for edited video material, and make three innovations: In Sec. 3.1, we incorporate shot-based temporal context into Stage I visual inputs. In Sec. 3.2, we leverage the thread structure to enrich cross-shot understanding. Finally, in Sec. 3.3, we incorporate shot-scale awareness into Stage I prompt formulation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.436, + 0.779 + ], + "angle": 0, + "content": "3.1. Leveraging shot-based temporal context" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Regarding the visual inputs to VideoLLM, prior works [25, 45] often sample frames \\(\\{\\mathcal{I}_{t_A},\\dots,\\mathcal{I}_{t_B}\\}\\) that directly correspond to the AD interval \\([t_A,t_B]\\). However, this approach can be problematic due to (i) misalignment between the AD interval and the actual timestamps when the action occurs, and (ii) the lack of contextual information from adjacent shots. Therefore, we investigate how incorporating temporal context can enhance the understanding of the video clip." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.903, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.218, + 0.907, + 0.285 + ], + "angle": 0, + "content": "Figure 2. Shot-based temporal context, where current shots are defined as those temporally overlapping with the AD interval. Past shots and future shots provide extended contextual information. Shot numbers are visually overlaid on the top-left of each frame, and frames within the AD interval are sampled more densely than context frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.905, + 0.357 + ], + "angle": 0, + "content": "Structuring temporal context with shots. To obtain more visual context information, instead of simply extending the AD interval by fixed timestamps, we explore a more structured approach that treats shots as the fundamental units." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Specifically, we apply an off-the-shelf shot segmentation model to partition the entire video clip into individual shots. For each AD interval, we first identify the shots that (partially) overlap with it, referred to as \"current shots\", as illustrated in Fig. 2. We then consider at most two \"past shots\" and two \"future shots\" adjacent to current shots as temporal context. For all shots included, we label them sequentially from past to future with a number starting from \"Shot \\(O\\)\"." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.906, + 0.533 + ], + "angle": 0, + "content": "Emphasising the targeted AD interval. To ensure that the VideoLLM focuses on describing visual content within the AD interval, we propose two strategies: dynamic frame sampling and shot number referral." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.623 + ], + "angle": 0, + "content": "Dynamic frame sampling. As shown in Fig. 2, to emphasise the frames of interest, we adopt denser sampling within the AD interval (red region) and sparser sampling for the surrounding context frames (yellow region). In practice, we specify fixed numbers of frames to be sampled within and outside the AD interval and apply uniform sampling according to these constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Shot number referral. To further enhance the attention towards the current shot content, we label each sampled frame with its shot number (e.g. \"Shot 0\") at the top-left. During the formulation of the text prompt, instead of prompting the VideoLLM to \"describe what happened in the video clip\", we ask it to \"describe what happened in [Shot 2, Shot 3]\" (i.e. current shots). Through this visual-textual prompting strategy, we found that the VideoLLM could successfully interpret the meaning of shot numbers and refer to the correct shots." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.768, + 0.779 + ], + "angle": 0, + "content": "3.2. Leveraging thread structure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Movies are generally edited such that viewpoints from two or more cameras are intertwined in shot threads, as illustrated in Fig. 3 (top). These interleaved arrangements of shot threads often imply relationships between objects and characters (e.g. gaze interactions) and their 3D arrangement. To leverage this information for Stage I description, we first determine the thread structure using a separate module, then incorporate it into the VideoLLM through prompt guidance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16505" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Thread structure. Top: Example of a thread structure with interleaving shots. Bottom: A training-free approach for thread clustering, where shots are pairwise compared using dense feature matching to construct an adjacency matrix, which is then used to predict the thread structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.557 + ], + "angle": 0, + "content": "Thread structure prediction. We develop a training-free method to predict thread clustering. The problem here is to determine if two shots correspond to the same viewpoint or not. Specifically, given two shots (Shot \\(i\\) and a later Shot \\(j\\)), we compare the last frame of Shot \\(i\\) \\((\\mathcal{I}_{T_i}^i)\\) with the first frame of Shot \\(j\\) \\((\\mathcal{I}_0^j)\\), extracting their DINoV2 [54] features as \\(f_{T_i}^i\\) and \\(f_0^j \\in \\mathbb{R}^{h \\times w \\times c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Inspired by [35], we assess frame-wise dense correlations by computing a cost volume between their feature maps:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.587, + 0.483, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {p, q} ^ {i, j} = m (p, q) \\circ \\left(\\hat {f} _ {T _ {i}; p} ^ {i} \\cdot \\hat {f} _ {0; q} ^ {j}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.671 + ], + "angle": 0, + "content": "where \\(\\hat{f}_{T_i;p}^i\\) and \\(\\hat{f}_{0;q}^{j}\\in \\mathbb{R}^{c}\\) are the normalised \\(p\\)-th and \\(q\\)-th spatial elements in the respective feature maps. The binary attention mask \\(m(p,q)\\) is set to 1 only if the spatial position of the \\(q\\)-th element is within an \\(n\\times n\\) neighbourhood of the \\(p\\)-th element." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.732 + ], + "angle": 0, + "content": "We then apply a softmax operation along the last dimension \\((q)\\) and find the maximum similarity for each \\(p\\), followed by averaging over all \\(p\\)-th elements to obtain a matching score between Shot \\(i\\) and Shot \\(j\\):" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.735, + 0.483, + 0.778 + ], + "angle": 0, + "content": "\\[\ns ^ {i, j} = \\frac {1}{N} \\sum_ {p} ^ {N} \\max _ {q} \\left(\\frac {\\exp \\left(\\mathcal {C} _ {p , q} ^ {i , j} / \\tau\\right)}{\\sum_ {l} ^ {N} \\exp \\left(\\mathcal {C} _ {p , l} ^ {i , j} / \\tau\\right)}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.81 + ], + "angle": 0, + "content": "where \\(N\\) denotes the number of feature patches, and \\(\\tau\\) is the softmax temperature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.812, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Intuitively, as shown in Fig. 3, this process effectively checks whether each patch \\((p)\\) in one shot frame matches with a patch \\((q)\\) in the other shot frame at a roughly similar spatial position (i.e., within an \\(n\\times n\\) neighbourhood)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finally, we construct an adjacency matrix based on the predicted scores \\( s^{i,j} \\) for all possible pairs of Shot \\( i \\) and Shot" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.897, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.159, + 0.897, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.285, + 0.908, + 0.352 + ], + "angle": 0, + "content": "Figure 4. Shot scales. Top: Examples of five different shot scales. Bottom: Stage I factor composition based on shot scale classification. The scales of the current shots in the clip are predicted and averaged. The resulting effective shot scale then guides the formulation of the Stage I prompt, incorporating additional factors such as facial expressions, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.414 + ], + "angle": 0, + "content": "\\(j\\). By thresholding the adjacency matrix (threshold \\(\\epsilon\\)) and identifying the largest connected components, we cluster the set of shots into multiple threads." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.532 + ], + "angle": 0, + "content": "Thread structure injection. Once the thread information is obtained, we inject it into the text prompt \\(\\mathcal{P}_{\\mathrm{VideoLLM}}\\) for the Stage I VideoLLM. In practice, we conduct this information injection only to clips that exhibit thread structures (i.e. \\(N_{\\mathrm{thread}} < N_{\\mathrm{shot}}\\)). For each given thread [Shot \\(i,\\ldots ,\\) Shot \\(j\\)], we formulate the prompt as: \"[Shot \\(i,\\ldots ,\\) Shot \\(j\\)] share the same camera setup\". This statement implies that the cameras in these shots maintain consistent angles and scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.605 + ], + "angle": 0, + "content": "Rather than simply providing this information, we further engage the VideoLLM by asking it to explain why the given thread structure is correct. This effectively corresponds to a Chain-of-Thought (CoT) process, enhancing its understanding of these repetitive thread structures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.81, + 0.629 + ], + "angle": 0, + "content": "3.3. Leveraging shot scale information" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.782 + ], + "angle": 0, + "content": "In movies and TV series, shot scales are often carefully designed during filming or post-editing to implicitly convey information to the audience. Our objective is to use the shot scale to choose what should be included in the Stage I prompts. Shot scale classification. We first build a shot scale classification network by fine-tuning a pre-trained DINOv2 [54] model. For all current shots, we classify their shot scales \\(\\{S_i\\}\\) into one of five classes, represented by values 0-4, as shown in Fig. 4 (top). We then compute their average to obtain the effective shot scale \\(S_{\\mathrm{eff}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Stage I factor composition. We then leverage the predicted shot scale to determine the factors to include in Stage I instructions. We first consider three fixed factors that form the basis of ADs, namely characters, actions, and interactions. The additional factor can be determined through applying a set of thresholds to \\( S_{\\mathrm{eff}} \\), as detailed in Fig. 4 (bottom). For example, for close-up shots (\\( S_{\\mathrm{eff}} \\leq 1.5 \\)), we would ask the VideoLLM to additionally describe the \"facial expression\", whereas for long shots" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16506" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.089, + 0.452, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.858, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.217, + 0.907, + 0.286 + ], + "angle": 0, + "content": "Figure 5. Action score. Left: Hierarchical prediction parsing. Given a predicted paragraph or sentence, it is first divided into a set of sentences (light blue) and action phrases (dark blue). Each action phrase is further processed to obtain its verb lemma (light orange). Middle: Scoring based on semantic similarity. Sentence embeddings for the GT action sentence and the set of predicted sentences and phrases are extracted, with the maximum cosine similarity defined as the similarity score. Right: Scoring based on verb matching. The predicted verb lemma is compared with the GT lemma and multiplied by the corresponding semantic-based similarity. The highest resultant score is defined as the verb score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.486, + 0.314 + ], + "angle": 0, + "content": "\\((S_{\\mathrm{eff}} \\geq 3.5)\\), a description on the \"environment\" will be included." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.328, + 0.214, + 0.343 + ], + "angle": 0, + "content": "4. Action score" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.485, + 0.513 + ], + "angle": 0, + "content": "In this section, we introduce a new metric, termed \"action score\", which focuses on whether a specific ground truth (GT) action is captured within a predicted Stage I description paragraph or a Stage II AD output. For instance, for the GT action \"He dials the phone\", we want the metric to measure the action of \"dial the phone\", but not be sensitive to character names and other predicted content. Therefore, this metric is designed to possess two key properties: (i) it is character-free, meaning that the presence of character names has minimal impact on the evaluation, and (ii) it is recall-oriented, without penalising additional action information in the prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.486, + 0.603 + ], + "angle": 0, + "content": "Preprocessing of GT actions. For each GT AD, we extract the character-free GT actions by (i) replacing character names with pronouns, and (ii) splitting the AD into subsentences, each containing one action (verb). For example, given a GT AD \"Chandler dials the phone, then hurriedly hangs up,\" the extracted GT actions are \"He dials the phone\" and \"He hurriedly hangs up.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.485, + 0.721 + ], + "angle": 0, + "content": "Hierarchical parsing of predictions. To process the predicted paragraph during evaluation, we first decompose it into individual sentences (light blue), as shown in the left column of Fig. 5. For each predicted sentence, we then perform rule-based dependency parsing to obtain action phrases (deep blue). Then, all extracted sentences and action phrases are collected to form a prediction set. Additionally, for each action phrase, we extract the corresponding verb lemma (i.e. the root form of verbs)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.486, + 0.887 + ], + "angle": 0, + "content": "Action score computation. The action score features a combination of semantic-based and verb-matching-based components, as illustrated in the middle and right columns of Fig. 5, respectively. Semantic-based evaluation. For each GT action sentence, we assess its semantic similarity with the set of predicted sentences and phrases. Specifically, we employ a general text embedding (GTE) model to compute sentence-level embeddings for the GT action sentence \\( e_{\\mathrm{GT}} \\) and each element in the predicted set \\( \\{e_{\\mathrm{pred};i}\\} \\). By computing cosine similarities and taking the maximum value, the similarity score is defined as \\( s_{\\mathrm{sim}} = \\max_i(s_{\\mathrm{sim};i}) = \\max_i[(e_{\\mathrm{GT}} \\cdot e_{\\mathrm{pred};i}) / (|e_{\\mathrm{GT}}||e_{\\mathrm{pred};i}|)] \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.887, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Verb-matching-based evaluation. In addition to semantic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.401 + ], + "angle": 0, + "content": "matching, we credit predictions that contain the same verbs as those in GT actions. Practically, we compare the predicted verb lemmas with GT verb lemma and compute binary matching scores \\(\\{m_j\\}\\). These matching scores are further weighted by the corresponding similarity scores \\(\\{s_{\\mathrm{sim},j}\\}\\) through element-wise multiplication. Finally, we take the maximum of the regulated scores to compute the verb score, i.e. \\(s_{\\mathrm{verb}} = \\max_j(s_{\\mathrm{sim};j} \\circ m_j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.445 + ], + "angle": 0, + "content": "To obtain the final action score, we combine the scores from both sources using a weighted average \\( s_{\\text{action}} = \\alpha_{\\text{sim}} \\circ s_{\\text{sim}} + \\alpha_{\\text{verb}} \\circ s_{\\text{verb}} \\), with \\( \\alpha_{\\text{sim}} \\) and \\( \\alpha_{\\text{verb}} \\) denoting the weighted factors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.64, + 0.471 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.553 + ], + "angle": 0, + "content": "This section begins with datasets and metrics for AD generation in Sec. 5.1, followed by implementation details in Sec. 5.2. Results on film grammar predictions are presented in Secs. 5.3 and 5.4, while Sec. 5.5 analyses human alignment with action scores. AD generation results are detailed in Secs. 5.6 to 5.8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.824, + 0.575 + ], + "angle": 0, + "content": "5.1. AD generation datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.711 + ], + "angle": 0, + "content": "We evaluate our framework on AD generation datasets for both movies (CMD-AD [25], MAD-Eval [23]) and TV series (TV-AD [82]). In more detail, CMD-AD is constructed by aligning ground truth ADs with the Condensed Movie Dataset (CMD) [7], comprising 101k ADs (94k for training and 7k for testing) from 1.4k movies. MAD-Eval consists of 6.5k ADs sampled from 10 movies within LSMDC [64]. On the other hand, TV-AD features 34k AD annotations from 13 TV series, with its test set sourced from TVQA [40], consisting of 3k ADs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.784 + ], + "angle": 0, + "content": "For AD evaluation, we follow the prior works [19, 25] to assess general prediction quality using CIDEr [71], Recall@k/N [24], and LLM-AD-Eval [25]. We also consider character recognition accuracy using CRITIC [25] and character-free action evaluation using action scores (described in Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.79, + 0.729, + 0.805 + ], + "angle": 0, + "content": "5.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In this section, we provide key implementation details for AD generation and action score computation. For additional information on film grammar predictions and other details, please refer to the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.872, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Shot detection. To segment the video clip into shots, we use PySceneDetect [3] with the \"Adaptive Detection\" method," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16507" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.188 + ], + "angle": 0, + "content": "
FeatureFrame setupPrecisionRecallAPWCP
CLIP-L14 CLSSide0.6910.6350.7050.922
DINOv2-L14 CLSSide0.7590.6830.7880.933
DINOv2-g14 CLSSide0.7610.6750.7860.936
DINOv2-g14 spatialMid0.8080.7170.8220.953
DINOv2-g14 spatialAll0.8700.7950.8960.964
DINOv2-g14 spatialSide0.8780.7990.9020.965
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.19, + 0.483, + 0.244 + ], + "angle": 0, + "content": "Table 1. Thread structure prediction on Thread-Safe. \"Side\" refers to comparisons between the temporally nearest frames in two shots; \"Mid\" refers to comparisons between the middle frames of each shot; \"All\" refers to comparisons between all frame pairs from the two shots." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.252, + 0.446, + 0.338 + ], + "angle": 0, + "content": "
MetricInputAccuracyMacro-F1
ViViT [6]RGB0.7470.751
SGNet [62]RGB + Flow0.875-
Lu et al. [46]RGB + mask0.892-
Li et al. [42]RGB0.8950.897
OursRGB0.8970.899
" + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.34, + 0.431, + 0.352 + ], + "angle": 0, + "content": "Table 2. Shot scale classification on MovieShots." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.413 + ], + "angle": 0, + "content": "which compares the ratio of pixel changes with the neighbouring frames. On average, each shot spans 3.5s, 3.3s, 3.8s in CMD-AD, TV-AD, and MAD-Eval, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.517 + ], + "angle": 0, + "content": "AD generation setting. During dynamic frame sampling, we select a total of 32 frames, with 16 frames uniformly sampled within and outside the AD interval, respectively. For character recognition, we adopt the same visual-textual prompting method proposed by [82], which applies coloured circles around faces for visual character indication. For simplicity, the visualisations in this paper do not display these circle labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Regarding the base models, we use Qwen2-VL-7B [75] as the VideoLLM in Stage I and LLaMA3-8B [48] as the LLM in Stage II. This setup is used as the default unless stated otherwise. Additionally, we explore the framework with the proprietary GPT-4o [53] model for both stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Action score evaluation setting. For action score computation, we set the weight factors as \\(\\alpha_{\\mathrm{sim}} = 0.8\\) and \\(\\alpha_{\\mathrm{verb}} = 0.2\\). When aggregating the action score results, we first average over multiple GT actions within each GT AD, and then perform global averaging across all AD samples. Moreover, in practice, we find that most action scores are clustered within the range of \\(0.25 - 0.75\\). To improve clarity, we apply further rescaling \\(f(x) = (x - 0.25) \\times 2\\) as post-processing. For evaluations in this paper, unless otherwise specified, we use the action score to assess the Stage II AD outputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.748, + 0.342, + 0.765 + ], + "angle": 0, + "content": "5.3. Thread structure prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.828 + ], + "angle": 0, + "content": "We evaluate thread structure prediction on Thread-Safe [27], which consists of approximately 4.7k video clips collected from 15 TV series. Each video clip contains a multi-shot scene with corresponding thread clusters manually annotated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For evaluation, we first construct an adjacency matrix from the GT clusters and extract binary GT labels \\(\\hat{s}_{i,j}\\) from the off-diagonal entries, where each label indicates whether a given pair of shots belong to the same cluster. We then compute the Average Precision (AP) between the predicted" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.209 + ], + "angle": 0, + "content": "
MetricParagraphSentence
PearsonSpearmanPearsonSpearman
CIDEr [71]0.2050.2640.4120.528
ROUGE-L [44]0.3050.2800.5260.512
SPICE [4]0.0220.0480.0310.012
BERTScore [91]0.3770.3930.5080.507
LLM-based (GPT-4o [53])0.7420.6780.7970.807
Action Score (w/o verb matching)0.7350.7280.7650.790
Action Score (w verb matching)0.7490.7290.8060.820
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.21, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Table 3. Comparison of action score with other metrics. The listed metrics measure the similarity between predicted paragraphs/sentences and ground truth actions. The reported values indicate the correlation (i.e. alignment) between these metrics and human-annotated scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.338 + ], + "angle": 0, + "content": "relationships \\(\\{s_{i,j}\\}\\) and the ground truth \\(\\{\\hat{s}_{i,j}\\}\\), as well as report the precision and recall values. Additionally, we directly compare the predicted clusters with ground truth clusters by reporting the weighted clustering purity (WCP) [69]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.906, + 0.455 + ], + "angle": 0, + "content": "Tab. 1 verifies the choice of DINOv2-g14 [54] as the feature extractor for frame pair comparison. Compared to abstract CLS tokens, dense spatial matching achieves higher AP in frame pair relationship prediction and higher WCP in thread clustering. Additionally, we observe that using the temporally closest frames from two shots (\"Side\") leads to a noticeable performance improvement. This can be attributed to the continuous story flow across shots within the same thread." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.466, + 0.729, + 0.48 + ], + "angle": 0, + "content": "5.4. Shot scale classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.603 + ], + "angle": 0, + "content": "Following prior work on shot scale classification [42, 46], we use the MovieShots [62] dataset, which consists of 46k shots (train:val:test = 7:1:2) collected from over 7k movie trailers. We follow its definition of shot scales, categorising shots into five classes ranging from extreme close-up to long shots, as illustrated in Fig. 4 (top). To evaluate the model performance, we report classification accuracy and Macro-F1 [42] scores on the MovieShots test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.906, + 0.677 + ], + "angle": 0, + "content": "Since previous state-of-the-art methods on shot scale classification are not open-sourced, we develop a new network by fine-tuning DINOv2 [54], achieving superior performance over prior approaches that rely on additional optical flow or SAM-based mask inputs, as demonstrated in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.831, + 0.703 + ], + "angle": 0, + "content": "5.5. Human alignment with action scores" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Action scores aim to evaluate whether a GT action is captured within a predicted description, making them recall-oriented. Such descriptions can be in the form of paragraphs (Stage I descriptions) or single sentences (Stage II ADs). To assess whether action scores align with human judgments, we create a dataset containing pairs of predicted descriptions and GT actions. For each GT action, human annotators manually annotate the quality of predictions into \\(\\{0,1,2,3\\}\\) based on the relevance towards GT action ranging from \"unrelated\" (0) to \"exact matching\" (3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison with other metrics. Next, we use the human-annotated scores as a reference to compare different metrics in terms of human agreement (measured by correlations), as reported in Tab. 3. Additionally, we consider an LLM-based" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16508" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.194 + ], + "angle": 0, + "content": "
Exp.Temporal contextFrame samplingShot labelCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
A---22.445.727.026.042.222.3
B1 shotDyn.Shot num.24.746.827.828.941.623.1
C2 shotsDyn.-24.546.627.427.842.322.3
D2 shotsDyn.Frame box24.846.427.425.041.822.8
E2 shotsUni.Shot num.24.147.526.226.542.222.0
F2 shotsDyn.Shot num.25.147.527.828.942.123.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.197, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Table 4. Leveraging shot-based temporal context. Key changes relative to the default setting (Exp. F) are highlighted in light blue. \"Temporal context\" indicates the number of context (past & future) shots. \"Shot num.\" refers to overlaying the shot number at the top-left of each current shot frame, while \"frame box\" represents highlighting the boundary of each current shot frame with a red box." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.286, + 0.48, + 0.348 + ], + "angle": 0, + "content": "
Thread structureCMD-AD subsetTV-AD subset
CIDErCRITICActionCIDErCRITICAction
X29.947.527.428.842.022.6
✓30.7↑0.848.9↑1.427.7↑0.330.7↑1.942.7↑0.722.9↑0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.391 + ], + "angle": 0, + "content": "Table 5. Thread structure injection. Thread structure information is injected only into subsets predicted to exhibit thread structures (\\(\\sim 30\\%\\) in CMD-AD and \\(\\sim 60\\%\\) in TV-AD)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.495 + ], + "angle": 0, + "content": "metric to predict scores, following the same scoring criteria as human annotations. In general, the action score achieves the best correlations with human annotations. Note, it is also more efficient than LLM-based metrics, with 0.15s compared to 6s per prediction evaluation. For more details regarding this human agreement study, please refer to the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.506, + 0.453, + 0.523 + ], + "angle": 0, + "content": "5.6. AD generation - evaluation of components" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.69 + ], + "angle": 0, + "content": "Shot-based temporal context. We investigate different setups for leveraging temporal context in AD generation, as shown in Tab. 4, leading to the following observations: (i) Expanding the temporal context range noticeably boosts the performance, with gains saturating around “2 shots” (Exp. A, B, and F); (ii) “Shot number referral” is the most effective strategy for outlining the current shot. (Exp. C, D, and F); (iii) Dynamically sampling the current shots at a higher frame rate boosts AD generation (Exp. E and F). The latter two improvements can be attributed to more efficient focus on the visual content around the targeted AD interval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.484, + 0.782 + ], + "angle": 0, + "content": "Thread structure injection. After extending the context information with neighbouring shots, we further enhance the VideoLLM's understanding by incorporating thread structures. Note that this guidance is applied only to video sequences exhibiting thread structures. As shown in Tab. 5, this additional information improves AD generation performance across both datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Scale-dependent Stage I factors. Tab. 6 explores the impact of different Stage I factors on final AD performance. Using shot scales as guidance for Stage I factor formulation (i.e. scale-dependent) not only outperforms configurations with single fixed factors but also surpasses the case where all factors are included in Stage I. This could be attributed to that the scale-dependent description contains more relevant and less redundant information, enabling more efficient AD extraction in Stage II." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.198 + ], + "angle": 0, + "content": "
Stage I factorsCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
Base25.447.427.729.242.023.0
Base + Face (AutoAD-Zero)25.246.827.830.042.322.5
Base + Obj.26.145.827.930.442.922.8
Base + Env.25.246.727.429.840.722.2
Base + Face + Obj. + Env.26.047.227.430.142.722.9
Scale-dependent (Ours)26.347.828.431.142.223.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Table 6. Factors included in Stage I description. Base: character + action + interaction; Face: facial expression; Env.: environment; Obj.: object. \"Scale-dependent\" refers to our approach, which leverages shot scale predictions to determine the relevant factors for each clip." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.27, + 0.851, + 0.286 + ], + "angle": 0, + "content": "5.7. AD generation - comparison with SotA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Tab. 7 provides a comprehensive summary of AD generation performance on CMD-AD and TV-AD, comparing across training-free methods with and without proprietary models, as well as models fine-tuned on human-annotated ADs. Notably, with the same base model setup (Qwen2-VL-7B + LLaMA3-8B), our training-free framework significantly outperforms AutoAD-Zero, primarily due to the usage of temporal context and film grammar information. By incorporating the more powerful GPT-4o models, our performance scales up further, surpassing even existing fine-tuned models. Additionally, we also report our performance on the MAD-Eval benchmark in Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.5 + ], + "angle": 0, + "content": "Qualitative visualisations. Fig. 6 presents several qualitative examples, where the top two cases illustrate how temporal context information aids in identifying key objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.573 + ], + "angle": 0, + "content": "In the bottom-left example, prior methods fail to associate characters, leading to the omission of the man (Alonso). In contrast, our method recognises the thread structure (i.e. [Shot 0, Shot 2], [Shot 1, Shot 3]), which guides the correct prediction of the man's gaze direction towards the lying woman." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.909, + 0.691 + ], + "angle": 0, + "content": "The bottom-right example highlights the effectiveness of scale-dependent Stage I factor formulation. AutoAD-Zero, designed to query characters, actions, interactions, and facial expressions, sometimes overlooks environmental details. Our method, in contrast, correctly identifies the shot as a long shot and instructs the VideoLLM to incorporate environmental context, resulting in more accurate scene descriptions. For more visualisations, please refer to the Supp. Mat. and Supp. Videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.73, + 0.719 + ], + "angle": 0, + "content": "5.8. Assisted AD generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.812 + ], + "angle": 0, + "content": "The subjective nature of AD sets a practical limit on metric scores, lower than the theoretical maximum, because human annotators often provide different but valid descriptions. Therefore, beyond enforcing generating a single AD sentence, we also consider employing our framework as an assistant to produce multiple candidate AD sentences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.814, + 0.906, + 0.9 + ], + "angle": 0, + "content": "To standardise such a protocol, we consider five candidate ADs generated by an assistant and employ an expert to select the best one. To effectively benchmark performance against existing GT ADs, we define the \"expert\" as an automatic selection mechanism that chooses the candidate with the highest average CIDEr and action score." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16509" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.878, + 0.226 + ], + "angle": 0, + "content": "
MethodVLMLLMTraining -freePropriet. modelCMD-ADTV-AD
CIDErCRITICActionR@1/5LLM-AD-EvalCIDErCRITICActionR@1/5LLM-AD-Eval
AutoAD-II [24]CLIP-B32GPT-2XX13.58.2-26.12.08|------
AutoAD-III [25]EVA-CLIPLLaMA2-7BXX25.032.731.531.22.89|2.0126.128.826.430.12.78|1.99
DistinctAD [19]CLIPAD-B16LLaMA3-8BXX22.7--33.02.88|2.0327.4--32.12.89|2.00
Video-LLaMA [90]Video-LLaMA-7B-✓X4.80.0-22.01.89|------
VideoBLIP [87]VideoBLIP-✓X5.20.0-23.61.91|------
AutoAD-Zero [82]VideoLLaMA2-7BLLaMA3-8B✓X17.743.725.526.92.83|1.9622.639.421.727.42.94|2.00
AutoAD-Zero [82]Qwen2-VL-7BLLaMA3-8B✓X21.944.326.930.83.00|2.2026.441.622.130.43.05|2.27
OursQwen2-VL-7BLLaMA3-8B✓X26.347.828.433.03.15|2.4231.142.223.933.13.09|2.35
AutoAD-Zero [82]GPT-4oGPT-4o✓✓22.445.130.732.93.08|2.4930.944.426.834.73.12|2.57
OursGPT-4oGPT-4o✓✓26.149.132.536.53.17|2.6634.246.527.436.63.12|2.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.227, + 0.908, + 0.255 + ], + "angle": 0, + "content": "Table 7. Quantitative comparison on CMD-AD and TV-AD. For training-free methods, \"VLM\" and \"LLM\" refer to the models used in separate stages, while for fine-tuned models, they denote the pre-trained components within an end-to-end model." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.267, + 0.492, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.363, + 0.49, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.268, + 0.869, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.363, + 0.868, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.468, + 0.908, + 0.551 + ], + "angle": 0, + "content": "Figure 6. Qualitative visualisations. Current shots (corresponding to AD intervals) are outlined by red boxes for illustration purposes only. For simplicity, not all context shots are shown. Training-free methods adopt Qwen2-VL + LLaMA3. Examples are taken from The Big Bang Theory (S2E14) (top left), Inferno (2016) (top right), The Asphalt Jungle (1950) (bottom left), and Signs (2002) (bottom right). The top-left example demonstrates the benefits of shot-based temporal context, where the objects (i.e. bills) in Penny's hands are not clearly visible within the AD interval (Shot 1), leading to ambiguous or incorrect predictions by AutoAD-Zero. In contrast, our method successfully identifies the objects from the context shot (Shot 2). The top-right example describing the action of Harry Sims similarly verifies the effectiveness of incorporating context shots." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.558, + 0.485, + 0.674 + ], + "angle": 0, + "content": "
MethodCandidate samplingCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
OursSingle AD (Ref.)26.347.828.431.142.223.9
OursIndep. output (p=0.90;τp=0.6)33.349.632.541.244.328.6
OursIndep. output (p=0.95;τp=1.5)37.050.335.145.546.431.5
AutoAD-Zero [82]Joint output31.646.433.843.246.830.1
OursJoint output38.449.235.751.347.331.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.771 + ], + "angle": 0, + "content": "Table 8. Assisted AD generation results. All methods adopt Qwen2-VL + LLaMA3-8B as base models. The first row provides single AD generation results as references (labelled in gray), the rest rows report the performance of one selected AD out of five candidates. \"Indep. output\" denotes five random independent Stage II runs, with \\( p \\) as the hyperparameter for top-p (nucleus) sampling and \\( {\\tau }_{p} \\) as the sampling temperature. \"Joint output\" generates five ADs simultaneously in a single run." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To develop an AD generation assistant, we fix the Stage I dense descriptions and explore generating multiple candidates in Stage II. This can be achieved by either running Stage II independently five times (termed the \"independent output\" setup) or generating five AD outputs simultaneously within a single run (termed the \"joint output\" setup). As observed in Tab. 8, the assistant-based setup significantly improves upon the single AD performance, highlighting the potential" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.693 + ], + "angle": 0, + "content": "of training-free methods in effectively capturing the desired content for AD generation. Within the \"independent output\" setup, increasing the randomness of sampling (i.e. higher \\( p \\) and \\( \\tau_p \\)) enhances the quality of the selected AD, owing to greater candidate diversity. Meanwhile, the \"joint output\" setup achieves superior performance, which could be attributed to reduced information redundancy across the simultaneously generated ADs. For additional visualisations, discussions, and detailed text prompts, please refer to the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.842, + 0.72 + ], + "angle": 0, + "content": "6. Discussion - summary and limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.799 + ], + "angle": 0, + "content": "We have demonstrated the benefit of shot-based context and film grammar awareness in AD generation – our training-free two-stage framework achieves state-of-the-art performance among all training-free counterparts, even surpassing fine-tuned models on multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The current framework has two main limitations: (i) the performance depends on the base VideoLLM, which may occasionally hallucinate details inconsistent with the visual content; and (ii) story-level context is not incorporated into the AD generation process. These limitations could potentially be addressed in future work by improving visual grounding and extending the visual and textual context to include the plot." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgements. This research is supported by the UK EPSRC Programme Grant Visual AI (EP/T028572/1), a Royal Society Research Professorship RSRP\\R\\241003, a Clarendon Scholarship, and ANR-21-CE23-0003-01 CorVis." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.163, + 0.184, + 0.179 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.188, + 0.484, + 0.216 + ], + "angle": 0, + "content": "[1] Guidelines for audio descriptors. https://adp.acb.org/guidelines.html.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.217, + 0.375, + 0.23 + ], + "angle": 0, + "content": "[2] Audiovault. https://audiovault.net/. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.231, + 0.425, + 0.244 + ], + "angle": 0, + "content": "[3] Pyscenedetect. https://www.scenedetect.com/. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.287, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[5] Dawit Mureja Argaw, Fabian Caba Heilbron, Joon-Young Lee, Markus Woodson, and In So Kweon. The anatomy of video editing: A dataset and benchmark suite for ai-assisted video editing. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[6] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.382, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[8] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.491, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[9] Andrew Brown, Vicky Kalogeiton, and Andrew Zisserman. Face, body, voice: Video person-clustering with multiple modalities. In ICCV 2021 Workshop on AI for Creative Video Editing and Understanding, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.483, + 0.584 + ], + "angle": 0, + "content": "[10] Boris Chen, Amir Ziai, Rebecca S. Tucker, and Yuchen Xie. Match cutting: Finding cuts with smooth visual transitions. In WACV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.587, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[11] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, pages 190-200, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[12] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.656, + 0.483, + 0.709 + ], + "angle": 0, + "content": "[13] Shimin Chen, Xiaohan Lan, Yitian Yuan, Zequn Jie, and Lin Ma. Timemarker: A versatile video-llm for long and short video understanding with superior temporal localization ability. arXiv preprint arXiv:2411.18211, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.711, + 0.483, + 0.777 + ], + "angle": 0, + "content": "[14] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.778, + 0.483, + 0.819 + ], + "angle": 0, + "content": "[15] Zeyu Chen, Yana Zhang, Suya Zhang, and Cheng Yang. Study on location bias of cnn for shot scale classification. Multimedia Tools and Applications, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.82, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[16] Peng Chu, Jiang Wang, and Andre Abrantes. LLM-AD: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.861, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[17] Andong Deng, Zhongpai Gao, Anwesa Choudhuri, Benjamin Planche, Meng Zheng, Bin Wang, Terrence Chen, Chen Chen, and Ziyan Wu. Seq2time: Sequential knowledge transfer for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.119 + ], + "angle": 0, + "content": "video llm temporal grounding. arXiv preprint arXiv:2411.16932, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.161, + 0.908, + 0.202 + ], + "angle": 0, + "content": "[19] Bo Fang, Wenhao Wu, Qiangqiang Wu, Yuxin Song, and Antoni B. Chan. DistinctAD: Distinctive audio description generation in contexts. arXiv preprint arXiv:2411.18180, 2024. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.203, + 0.907, + 0.242 + ], + "angle": 0, + "content": "[20] Dennis Fedorishin, Lie Lu, Srirangaraj Setlur, and Venu Govindaraju. Audio match cutting: Finding and creating matching audio transitions in movies and videos. In ICASSP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.244, + 0.907, + 0.297 + ], + "angle": 0, + "content": "[21] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.298, + 0.907, + 0.339 + ], + "angle": 0, + "content": "[22] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. arXiv preprint arXiv:2410.05643, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.907, + 0.379 + ], + "angle": 0, + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.907, + 0.42 + ], + "angle": 0, + "content": "[24] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD II: The sequel – who, when, and what in movie audio description. In ICCV, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.907, + 0.462 + ], + "angle": 0, + "content": "[25] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD III: The prequel - back to the pixels. In CVPR, 2024. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[26] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In EMNLP, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[27] Minh Hoai and Andrew Zisserman. Thread-safe: Towards recognizing human actions across shot boundaries. In ACCV, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.907, + 0.585 + ], + "angle": 0, + "content": "[28] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing-song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.906, + 0.625 + ], + "angle": 0, + "content": "[29] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.907, + 0.679 + ], + "angle": 0, + "content": "[30] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.681, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[31] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.722, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[32] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.907, + 0.791 + ], + "angle": 0, + "content": "[33] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops on Multimodal Learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.792, + 0.907, + 0.832 + ], + "angle": 0, + "content": "[34] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.907, + 0.86 + ], + "angle": 0, + "content": "[35] Allan Jabri, Andrew Owens, and Alexei A Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[36] Xuekun Jiang, Libiao Jin, Anyi Rao, Linning Xu, and Dahua Lin. Jointly learning the attributes and composition of shots for boundary detection in videos. Trans. Multi., 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16511" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.131 + ], + "angle": 0, + "content": "[37] Steve D. Katz. *Film Directing Shot by Shot: Visualizing from Concept to Screen*. Michael Wiese Productions, ISBN: 1615932976, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.173 + ], + "angle": 0, + "content": "[38] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.175, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[39] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[40] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, compositional video question answering. In EMNLP, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.255, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[41] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.484, + 0.334 + ], + "angle": 0, + "content": "[42] Yuzhi Li, Feng Tian, Haojun Xu, and Tianfeng Lu. Toward unified and quantitative cinematic shot attribute analysis. *Electronics*, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[43] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[44] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[45] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. In ECCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.472, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[46] Fengtian Lu, Yuzhi Li, and Feng Tian. Exploring challenge and explainable shot type classification using sam-guided approaches. Signal, Image and Video Processing, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[47] Manuel Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net++: revisiting people looking at each other in videos. IEEE TPAMI, pages 1-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[48] Meta. The llama 3 herd of models. arXiv preprint arXiv: 2407.21783, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.579, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[49] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[50] James Monaco. How to Read a Film: Movies, Media, and Beyond. OUP, ISBN: 0571168973, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.484, + 0.712 + ], + "angle": 0, + "content": "[52] Walter Murch. In the Blink of an Eye: A Perspective on Film Editing. Silman-James Press, ISBN: 1879505622, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.484, + 0.739 + ], + "angle": 0, + "content": "[53] OpenAI. Gpt-4 technical report. arXiv preprint arXiv: 2303.08774, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.741, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.862, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[56] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.187 + ], + "angle": 0, + "content": "[57] Alejandro Pardo, Fabio Pizzati, Tong Zhang, Alexander Pondaven, Philip Torr, Juan Camilo Perez, and Bernard Ghanem. Matchdiffusion: Training-free generation of match-cuts. arXiv preprint arXiv:2411.18677, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.229 + ], + "angle": 0, + "content": "[58] Toby Perrett, Tengda Han, Dima Damen, and Andrew Zisserman. It's just another day: Unique video captioning by discriminative prompting. In ACCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.231, + 0.907, + 0.283 + ], + "angle": 0, + "content": "[59] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.286, + 0.907, + 0.325 + ], + "angle": 0, + "content": "[60] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPRW, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.328, + 0.907, + 0.365 + ], + "angle": 0, + "content": "[61] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.907, + 0.42 + ], + "angle": 0, + "content": "[62] Anyi Rao, Jiaze Wang, Linning Xu, Xuekun Jiang, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A unified framework for shot type classification based on subject centric lens. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.423, + 0.907, + 0.476 + ], + "angle": 0, + "content": "[63] Anyi Rao, Xuekun Jiang, Sichen Wang, Yuwei Guo, Zihao Liu, Bo Dai, Long Pang, Xiaoyu Wu, Dahua Lin, and Libiao Jin. Temporal and contextual transformer for multi-camera editing of tv shows. In Proceedings of the CVEU Workshop at ICCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[64] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[65] Mattia Savardi, András Bálint Kovács, Alberto Signoroni, and Sergio Benini. Recognition of camera angle and camera level in movies from single frames. In Proceedings of the ACM International Conference on Interactive Media Experiences Workshops, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.907, + 0.625 + ], + "angle": 0, + "content": "[66] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting LLMs to transform video annotations at scale. arXiv:2310.04900, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.907, + 0.681 + ], + "angle": 0, + "content": "[67] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[68] Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, and Chenliang Xu. Vidcomposition: Can mllms analyze compositions in compiled videos? arXiv preprint arXiv:2411.10979, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.752, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[69] Makarand Tapaswi, Omkar M. Parkhi, Esa Rahtu, Eric Sommerlade, Rainer Stiefelhagen, and Andrew Zisserman. Total cluster: A person agnostic clustering method for broadcast videos. In Proceedings of the Indian Conference on Computer Vision Graphics and Image Processing, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.819, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[70] Bartolomeo Vacchetti and Tania Cerquitelli. Cinematographic shot classification with deep ensemble learning. *Electronics*, 2022, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.861, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[71] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16512" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.133 + ], + "angle": 0, + "content": "[72] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.187 + ], + "angle": 0, + "content": "[73] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videoollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.188, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[74] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.231, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.484, + 0.365 + ], + "angle": 0, + "content": "[76] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.367, + 0.484, + 0.407 + ], + "angle": 0, + "content": "[77] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.408, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[78] Yujia Wang, Wei Liang, Haikun Huang, Yongqi Zhang, Dingzeyu Li, and Lap-Fai Yu. Toward automatic audio description generation for accessible videos. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[80] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[81] Yongliang Wu, Xinting Hu, Yuyang Sun, Yizhou Zhou, Wenbo Zhu, Fengyun Rao, Bernt Schiele, and Xu Yang. Number it: Temporal grounding videos like flipping manga. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.484, + 0.652 + ], + "angle": 0, + "content": "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. In ACCV, 2024. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.654, + 0.484, + 0.693 + ], + "angle": 0, + "content": "[83] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.695, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[84] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.75, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[85] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. arXiv preprint arXiv:2302.14115, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[86] Xiaojun Ye, Junhao Chen, Xiang Li, Haidong Xin, Chao Li, Sheng Zhou, and Jiajun Bu. MMAD: Multi-modal movie audio description. In LREC-COLING, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.346, + 0.873 + ], + "angle": 0, + "content": "[87] Keunwoo Peter Yu. Videoblip, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[88] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Farhadi, and Yejin Choi. MERLOT reserve: Multimodal neural script knowledge through vision and language and sound. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.907, + 0.187 + ], + "angle": 0, + "content": "[89] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. MM-Narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.188, + 0.905, + 0.228 + ], + "angle": 0, + "content": "[90] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In ENNLP, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.229, + 0.905, + 0.269 + ], + "angle": 0, + "content": "[91] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In *ICLR*, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.271, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[92] Yuqi Zhang, Bin Guo, Nuo Li, Ying Zhang, Qianru Wang, and Zhiwen Yu. Representation learning of next shot selection for vlog editing. In Proceedings of the CVEU Workshop at ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.325, + 0.905, + 0.365 + ], + "angle": 0, + "content": "[93] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.367, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[94] Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding using large-scale pre-trained models. arXiv preprint arXiv:2408.16219, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[95] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos In AAAI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[96] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[97] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16513" + } + ] +] \ No newline at end of file diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_origin.pdf b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d47214b6c31050bd517c1460c11055cb7b43fd92 --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/b2b3829b-f130-4214-bc1a-c2339869b439_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dbf060e1419b032e9049af019261d69cfb69cbd0349598f5b1065cba279de10 +size 14171854 diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/full.md b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b7b641d71bf1d3714d60cd0cc6580f842ef4c27 --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/full.md @@ -0,0 +1,396 @@ +# Shot-by-Shot: Film-Grammar-Aware Training-Free Audio Description Generation + +![](images/7e2ecb7a89bd739996aaa3094b2113e7a0ecdfd23db37b1591318beec45e157a.jpg) + +![](images/460a0df0f9a1e2a515e76ad3232416449c89f32d6817766767c05d63396397fd.jpg) +Figure 1. Overview of our training-free Audio Description (AD) generation framework. The input video clip (corresponding to the AD interval) is extended to include adjacent shots, providing richer temporal visual context. Each video frame is labeled with its corresponding shot number. The extended clip then undergoes a film grammar prediction process, where the thread structure and shot scales are estimated. The AD is generated in two stages: Stage I utilises the predicted cinematic information as training-free prompt guidance to produce dense video descriptions. Stage II then employs an LLM to generate a summarised AD output. + +# Abstract + +Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages "shots" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure - an action score - specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection. + +# 1. Introduction + +In movies and TV series, Audio Descriptions (ADs) are narrations provided for the visually impaired, conveying visual information to complement the original soundtrack. Their + +purpose is to ensure a continuous and coherent narrative flow, enabling audiences to follow the plot effectively. Unlike video captions, ADs are constrained by length, prioritising the most visually salient and story-centric information, such as character dynamics and significant objects, whilst omitting redundant details like background figures or unchanging locations. Additionally, ADs are typically produced by professional narrators in a specific style and format, ensuring coherence while not interfering with the original audio. + +With the advent of Visual-Language Models (VLMs), there has been a growth of interest in automatically generating ADs for both movies [16, 23-25, 45, 72, 86, 89] and TV material [19, 82]. However, as anyone who has ever watched a movie or read a book about film editing knows – the fundamental unit of edited video material is the shot, not the frame [37, 50, 52]. Shots are used to structure the video material, defining the granularity and temporal context through choices in its scale (close-up, long shot, etc.) and its movement (panning, tracking, etc.). The "film grammar" is then used to convey meaning through specific editing choices on shot transitions (cuts, fades, etc.), durations, and composition (threads, montage, etc.). + +Current approaches to automating AD generation, and video large language models (VideoLLMs) in general, are not shot-aware, hindering interpretation of edited video material that has frequent shot transitions [68]. In this paper, our principal + +objective is to incorporate shot information and editing structure into the AD generation process. To this end, we consider temporal context in terms of shots, and take account of the two key properties of thread structure and shot scale. + +Thread structure identifies the sequence of shots captured with the same camera. An example of a scene with multiple simultaneous threads in space-time is shown in Fig. 3. We develop a robust thread clustering method and use its predictions to guide VLMs in understanding shot-wise relationships. + +Shot scale typically implies the type of content of the frame. For instance, close-up shots often highlight characters' facial expressions or gaze interactions, while long shots tend to depict the overall environment and ambience of the scene, as shown in Fig. 4 (top). We leverage this property by developing an off-the-shelf shot scale classifier and employing a scale-dependent prompting strategy to improve the contextual relevance of VLM-generated descriptions. + +In this paper, we develop a training-free AD generation approach. Inspired by AutoAD-Zero [82], as illustrated in Fig. 1, we adopt a two-stage pipeline in which a VideoLLM generates dense text descriptions in the first stage, followed by an LLM that produces the final AD outputs from this text. We improve the effectiveness of training-free approaches by incorporating two key improvements based on: (i) shot-based temporal context; and (ii) the shot scale and thread structure, that are both crucial to cinematic composition and understanding. Previous methods could describe human-object interactions or human-human interactions, such as looks [47], within a frame or shot. With the improvements we introduce, the generated AD also includes such interactions when they are implied by the shot structure. + +As well as the challenge of generating ADs, another challenge is how to evaluate the predicted ADs. Apart from conventional metrics [8, 44, 55, 71], several new AD metrics [23, 25, 89] have been proposed, including CRITIC [25] measuring who is mentioned in the AD. However, these metrics fail to emphasise character actions – one of the most critical aspects of ADs. To address this, we introduce an action score, assessing whether the predictions accurately capture the actions described in GT AD, independent of character information. + +Moreover, as ADs are time-limited, there is always a choice on which visual aspect (characters, actions, objects, etc.) should be included. Consequently, a single video clip can have multiple equally valid ADs, each highlighting a different aspect. This observation is supported by the inter-rater agreement experiments in AutoAD-III [25] and previous user studies [78]. Therefore, beyond evaluating single AD performance, we assess our framework's capability as an assistant to generate multiple AD candidates, and report the performance of the selected one AD. + +In summary, we propose an enhanced training-free AD generation framework, with the following contributions: (i) We incorporate shot-based temporal context into AD generation via training-free prompting techniques including shot number referral and dynamic frame sampling; (ii) We develop state-of-the-art methods for thread structure and shot scale predictions, and demonstrate that incorporating predicted film grammar knowl + +edge enhances AD generation; (iii) We improve the current AD evaluation by introducing the character-free action score, and a new assistance-oriented evaluation protocol; (iv) Our approach achieves state-of-the-art performance in training-free AD generation, and furthermore, surpasses fine-tuned models on multiple benchmarks. This is the first time a training-free approach has achieved superior performance to fine-tuned methods. + +# 2. Related work + +Audio Description generation. Efforts have been made to curate Audio Description (AD) datasets for both movies [25, 67] and TV series [82], with human annotations sourced from platforms such as AudioVault [2]. + +For automatic AD generation, prior works [19, 23-25, 45, 72] fine-tune pre-trained VLMs on AD annotations to produce descriptions in an end-to-end manner. However, these methods face challenges due to limited high-quality AD annotations and the high computational cost of fine-tuning each new backbone. Alternatively, training-free approaches [16, 82, 86, 89] have gained traction for their scalability and flexibility, allowing customised AD output based on official guidelines [1] or specific needs. Yet, these methods still lag in performance, while our approach is the first to achieve results on par with fine-tuned methods. + +Instead of limiting the video input to each AD interval, UniAD [72] and DistinctAD [19] fine-tune VLMs with multiple AD clip inputs to incorporate broader temporal context. In contrast, our method systematically extends AD clip to adjacent shots and introduces a training-free approach that enables pretrained VideoLLMs to better capture localised temporal context. + +Film grammar analysis. Prior research has sought to understand film grammar from two major perspectives: (i) intra-shot properties, (ii) shot-wise relationships. + +For individual shots, several datasets [5, 31, 36, 62, 65] categorise their characteristics based on camera setups, including shot scales (examples shown in Fig. 4) and camera movements. Correspondingly, various models [15, 42, 46, 62, 70] have been developed for shot type classification. + +Regarding shot-wise relationships, a few datasets [5, 10, 56] have been proposed to explore transitions (i.e. cuts) between shots, which have also been leveraged in video content generation [20, 57, 63, 92]. Beyond pairwise shot transitions, research has also investigated longer temporal contexts with thread-based editing structures. Notably, Hoai et al. introduced the Thread-Safe [27] dataset, demonstrating that thread information can enhance action recognition. These structures have also been utilised in video-based face and people clustering [9, 69]. + +While prior work on film grammar has mainly focused on classification and generation tasks, we specifically utilise shot scale and thread structure information to enhance AD generation in movies and TV series. + +- Dense video captioning. Dense video captioning is closely related to AD generation. Early works [32, 33, 39, 74, 76] in video captioning typically treat event localisation and + +captioning as independent stages, whereas more recent approaches [12, 18, 34, 38, 51, 58, 61, 77, 80, 85, 96, 97] integrate these tasks in an end-to-end manner. Video captioning benchmarks cover a range of domains, including cooking [95], actions [39], movies [64], TV series [41], and open-domain settings [11, 14, 26, 49, 66, 83, 84, 88, 93]. + +Temporal grounding in VLM. To equip conventional VLMs with temporal grounding capability, some studies [29, 43, 79] generate additional data with enhanced temporal information for fine-tuning. A more common approach explicitly incorporates temporal information into inputs, either by inserting temporal tokens [13, 17, 21, 22, 28, 30, 59, 73] or embedding time information within visual tokens [29, 43, 79]. + +Training-free approaches [60, 81, 94] have also been explored for temporally grounded understanding. Notably, a recent work [81] achieves temporal grounding by overlaying frame numbers as visual prompts. Instead of using uniformly sampled timestamps, we leverage the natural shot structures in movies and TV series, adopting them as fundamental units for training-free temporal referral. + +# 3. Training-free AD generation framework + +Given a video clip $\mathcal{V} = \{\mathcal{I}_0,\dots,\mathcal{I}_T\}$ , the task of audio description is to generate a concise narration $\mathcal{N}$ describing what happens around a given AD interval $[t_A,t_B]$ . In this work, we propose a two-stage framework that leverages VideoLLMs and LLMs to predict ADs in a training-free manner. + +In Stage I, we employ a VideoLLM that takes a sequence of frames (from multiple shots) as input and generates a dense description $\mathcal{D}$ , guided by instructions $\mathcal{P}_{\text{VideoLLM}}$ : + +$$ +\mathcal {D} = \operatorname {V i d e o L L M} (\mathcal {V}, [ t _ {A}, t _ {B} ], \mathcal {P} _ {\text {V i d e o L L M}}) \tag {1} +$$ + +In Stage II, we then prompt an LLM (with instructions $\mathcal{P}_{\mathrm{LLM}}$ ) to extract key information from the dense Stage I description and format it into an AD-style narration $\mathcal{N}$ : + +$$ +\mathcal {N} = \operatorname {L L M} \left(\mathcal {D}, \left[ t _ {A}, t _ {B} \right], \mathcal {P} _ {\text {L L M}}\right) \tag {2} +$$ + +In this section, we focus on enhancing the visual understanding of the Stage I VideoLLM for edited video material, and make three innovations: In Sec. 3.1, we incorporate shot-based temporal context into Stage I visual inputs. In Sec. 3.2, we leverage the thread structure to enrich cross-shot understanding. Finally, in Sec. 3.3, we incorporate shot-scale awareness into Stage I prompt formulation. + +# 3.1. Leveraging shot-based temporal context + +Regarding the visual inputs to VideoLLM, prior works [25, 45] often sample frames $\{\mathcal{I}_{t_A},\dots,\mathcal{I}_{t_B}\}$ that directly correspond to the AD interval $[t_A,t_B]$ . However, this approach can be problematic due to (i) misalignment between the AD interval and the actual timestamps when the action occurs, and (ii) the lack of contextual information from adjacent shots. Therefore, we investigate how incorporating temporal context can enhance the understanding of the video clip. + +![](images/66f272fa2845c90db118b889236117301ec8842351092366ffc8e8c1f940bba1.jpg) +Figure 2. Shot-based temporal context, where current shots are defined as those temporally overlapping with the AD interval. Past shots and future shots provide extended contextual information. Shot numbers are visually overlaid on the top-left of each frame, and frames within the AD interval are sampled more densely than context frames. + +Structuring temporal context with shots. To obtain more visual context information, instead of simply extending the AD interval by fixed timestamps, we explore a more structured approach that treats shots as the fundamental units. + +Specifically, we apply an off-the-shelf shot segmentation model to partition the entire video clip into individual shots. For each AD interval, we first identify the shots that (partially) overlap with it, referred to as "current shots", as illustrated in Fig. 2. We then consider at most two "past shots" and two "future shots" adjacent to current shots as temporal context. For all shots included, we label them sequentially from past to future with a number starting from "Shot $O$ ". + +Emphasising the targeted AD interval. To ensure that the VideoLLM focuses on describing visual content within the AD interval, we propose two strategies: dynamic frame sampling and shot number referral. + +Dynamic frame sampling. As shown in Fig. 2, to emphasise the frames of interest, we adopt denser sampling within the AD interval (red region) and sparser sampling for the surrounding context frames (yellow region). In practice, we specify fixed numbers of frames to be sampled within and outside the AD interval and apply uniform sampling according to these constraints. + +Shot number referral. To further enhance the attention towards the current shot content, we label each sampled frame with its shot number (e.g. "Shot 0") at the top-left. During the formulation of the text prompt, instead of prompting the VideoLLM to "describe what happened in the video clip", we ask it to "describe what happened in [Shot 2, Shot 3]" (i.e. current shots). Through this visual-textual prompting strategy, we found that the VideoLLM could successfully interpret the meaning of shot numbers and refer to the correct shots. + +# 3.2. Leveraging thread structure + +Movies are generally edited such that viewpoints from two or more cameras are intertwined in shot threads, as illustrated in Fig. 3 (top). These interleaved arrangements of shot threads often imply relationships between objects and characters (e.g. gaze interactions) and their 3D arrangement. To leverage this information for Stage I description, we first determine the thread structure using a separate module, then incorporate it into the VideoLLM through prompt guidance. + +![](images/415d9bcd947817223b459873e0102686dfe077fe8c3128a52cb332ab845236b8.jpg) +Figure 3. Thread structure. Top: Example of a thread structure with interleaving shots. Bottom: A training-free approach for thread clustering, where shots are pairwise compared using dense feature matching to construct an adjacency matrix, which is then used to predict the thread structure. + +Thread structure prediction. We develop a training-free method to predict thread clustering. The problem here is to determine if two shots correspond to the same viewpoint or not. Specifically, given two shots (Shot $i$ and a later Shot $j$ ), we compare the last frame of Shot $i$ $(\mathcal{I}_{T_i}^i)$ with the first frame of Shot $j$ $(\mathcal{I}_0^j)$ , extracting their DINoV2 [54] features as $f_{T_i}^i$ and $f_0^j \in \mathbb{R}^{h \times w \times c}$ . + +Inspired by [35], we assess frame-wise dense correlations by computing a cost volume between their feature maps: + +$$ +\mathcal {C} _ {p, q} ^ {i, j} = m (p, q) \circ \left(\hat {f} _ {T _ {i}; p} ^ {i} \cdot \hat {f} _ {0; q} ^ {j}\right) \tag {3} +$$ + +where $\hat{f}_{T_i;p}^i$ and $\hat{f}_{0;q}^{j}\in \mathbb{R}^{c}$ are the normalised $p$ -th and $q$ -th spatial elements in the respective feature maps. The binary attention mask $m(p,q)$ is set to 1 only if the spatial position of the $q$ -th element is within an $n\times n$ neighbourhood of the $p$ -th element. + +We then apply a softmax operation along the last dimension $(q)$ and find the maximum similarity for each $p$ , followed by averaging over all $p$ -th elements to obtain a matching score between Shot $i$ and Shot $j$ : + +$$ +s ^ {i, j} = \frac {1}{N} \sum_ {p} ^ {N} \max _ {q} \left(\frac {\exp \left(\mathcal {C} _ {p , q} ^ {i , j} / \tau\right)}{\sum_ {l} ^ {N} \exp \left(\mathcal {C} _ {p , l} ^ {i , j} / \tau\right)}\right) \tag {4} +$$ + +where $N$ denotes the number of feature patches, and $\tau$ is the softmax temperature. + +Intuitively, as shown in Fig. 3, this process effectively checks whether each patch $(p)$ in one shot frame matches with a patch $(q)$ in the other shot frame at a roughly similar spatial position (i.e., within an $n\times n$ neighbourhood). + +Finally, we construct an adjacency matrix based on the predicted scores $s^{i,j}$ for all possible pairs of Shot $i$ and Shot + +![](images/3ff0827dcc9824848c65dfe9348fba3c429bc6c7facff7938dfc0ca8a9ab5193.jpg) + +![](images/ac35de672e581b6d0d470dc8ba4e94bcc1817e4ea72e254c4f0725162a56dd7d.jpg) +Figure 4. Shot scales. Top: Examples of five different shot scales. Bottom: Stage I factor composition based on shot scale classification. The scales of the current shots in the clip are predicted and averaged. The resulting effective shot scale then guides the formulation of the Stage I prompt, incorporating additional factors such as facial expressions, etc. + +$j$ . By thresholding the adjacency matrix (threshold $\epsilon$ ) and identifying the largest connected components, we cluster the set of shots into multiple threads. + +Thread structure injection. Once the thread information is obtained, we inject it into the text prompt $\mathcal{P}_{\mathrm{VideoLLM}}$ for the Stage I VideoLLM. In practice, we conduct this information injection only to clips that exhibit thread structures (i.e. $N_{\mathrm{thread}} < N_{\mathrm{shot}}$ ). For each given thread [Shot $i,\ldots ,$ Shot $j$ ], we formulate the prompt as: "[Shot $i,\ldots ,$ Shot $j$ ] share the same camera setup". This statement implies that the cameras in these shots maintain consistent angles and scales. + +Rather than simply providing this information, we further engage the VideoLLM by asking it to explain why the given thread structure is correct. This effectively corresponds to a Chain-of-Thought (CoT) process, enhancing its understanding of these repetitive thread structures. + +# 3.3. Leveraging shot scale information + +In movies and TV series, shot scales are often carefully designed during filming or post-editing to implicitly convey information to the audience. Our objective is to use the shot scale to choose what should be included in the Stage I prompts. Shot scale classification. We first build a shot scale classification network by fine-tuning a pre-trained DINOv2 [54] model. For all current shots, we classify their shot scales $\{S_i\}$ into one of five classes, represented by values 0-4, as shown in Fig. 4 (top). We then compute their average to obtain the effective shot scale $S_{\mathrm{eff}}$ . + +Stage I factor composition. We then leverage the predicted shot scale to determine the factors to include in Stage I instructions. We first consider three fixed factors that form the basis of ADs, namely characters, actions, and interactions. The additional factor can be determined through applying a set of thresholds to $S_{\mathrm{eff}}$ , as detailed in Fig. 4 (bottom). For example, for close-up shots ( $S_{\mathrm{eff}} \leq 1.5$ ), we would ask the VideoLLM to additionally describe the "facial expression", whereas for long shots + +![](images/6d87d3d749e1e68e37f4e0747526850f2c1a3d038f803923cdedaeaf3de4d644.jpg) +Figure 5. Action score. Left: Hierarchical prediction parsing. Given a predicted paragraph or sentence, it is first divided into a set of sentences (light blue) and action phrases (dark blue). Each action phrase is further processed to obtain its verb lemma (light orange). Middle: Scoring based on semantic similarity. Sentence embeddings for the GT action sentence and the set of predicted sentences and phrases are extracted, with the maximum cosine similarity defined as the similarity score. Right: Scoring based on verb matching. The predicted verb lemma is compared with the GT lemma and multiplied by the corresponding semantic-based similarity. The highest resultant score is defined as the verb score. + +![](images/ffd3453db4326325f47c85581c0dba5a5bc2449893575bff8be90fdc09fc33be.jpg) + +$(S_{\mathrm{eff}} \geq 3.5)$ , a description on the "environment" will be included. + +# 4. Action score + +In this section, we introduce a new metric, termed "action score", which focuses on whether a specific ground truth (GT) action is captured within a predicted Stage I description paragraph or a Stage II AD output. For instance, for the GT action "He dials the phone", we want the metric to measure the action of "dial the phone", but not be sensitive to character names and other predicted content. Therefore, this metric is designed to possess two key properties: (i) it is character-free, meaning that the presence of character names has minimal impact on the evaluation, and (ii) it is recall-oriented, without penalising additional action information in the prediction. + +Preprocessing of GT actions. For each GT AD, we extract the character-free GT actions by (i) replacing character names with pronouns, and (ii) splitting the AD into subsentences, each containing one action (verb). For example, given a GT AD "Chandler dials the phone, then hurriedly hangs up," the extracted GT actions are "He dials the phone" and "He hurriedly hangs up." + +Hierarchical parsing of predictions. To process the predicted paragraph during evaluation, we first decompose it into individual sentences (light blue), as shown in the left column of Fig. 5. For each predicted sentence, we then perform rule-based dependency parsing to obtain action phrases (deep blue). Then, all extracted sentences and action phrases are collected to form a prediction set. Additionally, for each action phrase, we extract the corresponding verb lemma (i.e. the root form of verbs). + +Action score computation. The action score features a combination of semantic-based and verb-matching-based components, as illustrated in the middle and right columns of Fig. 5, respectively. Semantic-based evaluation. For each GT action sentence, we assess its semantic similarity with the set of predicted sentences and phrases. Specifically, we employ a general text embedding (GTE) model to compute sentence-level embeddings for the GT action sentence $e_{\mathrm{GT}}$ and each element in the predicted set $\{e_{\mathrm{pred};i}\}$ . By computing cosine similarities and taking the maximum value, the similarity score is defined as $s_{\mathrm{sim}} = \max_i(s_{\mathrm{sim};i}) = \max_i[(e_{\mathrm{GT}} \cdot e_{\mathrm{pred};i}) / (|e_{\mathrm{GT}}||e_{\mathrm{pred};i}|)]$ . + +Verb-matching-based evaluation. In addition to semantic + +matching, we credit predictions that contain the same verbs as those in GT actions. Practically, we compare the predicted verb lemmas with GT verb lemma and compute binary matching scores $\{m_j\}$ . These matching scores are further weighted by the corresponding similarity scores $\{s_{\mathrm{sim},j}\}$ through element-wise multiplication. Finally, we take the maximum of the regulated scores to compute the verb score, i.e. $s_{\mathrm{verb}} = \max_j(s_{\mathrm{sim};j} \circ m_j)$ . + +To obtain the final action score, we combine the scores from both sources using a weighted average $s_{\text{action}} = \alpha_{\text{sim}} \circ s_{\text{sim}} + \alpha_{\text{verb}} \circ s_{\text{verb}}$ , with $\alpha_{\text{sim}}$ and $\alpha_{\text{verb}}$ denoting the weighted factors. + +# 5. Experiments + +This section begins with datasets and metrics for AD generation in Sec. 5.1, followed by implementation details in Sec. 5.2. Results on film grammar predictions are presented in Secs. 5.3 and 5.4, while Sec. 5.5 analyses human alignment with action scores. AD generation results are detailed in Secs. 5.6 to 5.8. + +# 5.1. AD generation datasets and metrics + +We evaluate our framework on AD generation datasets for both movies (CMD-AD [25], MAD-Eval [23]) and TV series (TV-AD [82]). In more detail, CMD-AD is constructed by aligning ground truth ADs with the Condensed Movie Dataset (CMD) [7], comprising 101k ADs (94k for training and 7k for testing) from 1.4k movies. MAD-Eval consists of 6.5k ADs sampled from 10 movies within LSMDC [64]. On the other hand, TV-AD features 34k AD annotations from 13 TV series, with its test set sourced from TVQA [40], consisting of 3k ADs. + +For AD evaluation, we follow the prior works [19, 25] to assess general prediction quality using CIDEr [71], Recall@k/N [24], and LLM-AD-Eval [25]. We also consider character recognition accuracy using CRITIC [25] and character-free action evaluation using action scores (described in Sec. 4). + +# 5.2. Implementation details + +In this section, we provide key implementation details for AD generation and action score computation. For additional information on film grammar predictions and other details, please refer to the Supp. Mat. + +Shot detection. To segment the video clip into shots, we use PySceneDetect [3] with the "Adaptive Detection" method, + +
FeatureFrame setupPrecisionRecallAPWCP
CLIP-L14 CLSSide0.6910.6350.7050.922
DINOv2-L14 CLSSide0.7590.6830.7880.933
DINOv2-g14 CLSSide0.7610.6750.7860.936
DINOv2-g14 spatialMid0.8080.7170.8220.953
DINOv2-g14 spatialAll0.8700.7950.8960.964
DINOv2-g14 spatialSide0.8780.7990.9020.965
+ +Table 1. Thread structure prediction on Thread-Safe. "Side" refers to comparisons between the temporally nearest frames in two shots; "Mid" refers to comparisons between the middle frames of each shot; "All" refers to comparisons between all frame pairs from the two shots. + +
MetricInputAccuracyMacro-F1
ViViT [6]RGB0.7470.751
SGNet [62]RGB + Flow0.875-
Lu et al. [46]RGB + mask0.892-
Li et al. [42]RGB0.8950.897
OursRGB0.8970.899
+ +which compares the ratio of pixel changes with the neighbouring frames. On average, each shot spans 3.5s, 3.3s, 3.8s in CMD-AD, TV-AD, and MAD-Eval, respectively. + +AD generation setting. During dynamic frame sampling, we select a total of 32 frames, with 16 frames uniformly sampled within and outside the AD interval, respectively. For character recognition, we adopt the same visual-textual prompting method proposed by [82], which applies coloured circles around faces for visual character indication. For simplicity, the visualisations in this paper do not display these circle labels. + +Regarding the base models, we use Qwen2-VL-7B [75] as the VideoLLM in Stage I and LLaMA3-8B [48] as the LLM in Stage II. This setup is used as the default unless stated otherwise. Additionally, we explore the framework with the proprietary GPT-4o [53] model for both stages. + +Action score evaluation setting. For action score computation, we set the weight factors as $\alpha_{\mathrm{sim}} = 0.8$ and $\alpha_{\mathrm{verb}} = 0.2$ . When aggregating the action score results, we first average over multiple GT actions within each GT AD, and then perform global averaging across all AD samples. Moreover, in practice, we find that most action scores are clustered within the range of $0.25 - 0.75$ . To improve clarity, we apply further rescaling $f(x) = (x - 0.25) \times 2$ as post-processing. For evaluations in this paper, unless otherwise specified, we use the action score to assess the Stage II AD outputs. + +# 5.3. Thread structure prediction + +We evaluate thread structure prediction on Thread-Safe [27], which consists of approximately 4.7k video clips collected from 15 TV series. Each video clip contains a multi-shot scene with corresponding thread clusters manually annotated. + +For evaluation, we first construct an adjacency matrix from the GT clusters and extract binary GT labels $\hat{s}_{i,j}$ from the off-diagonal entries, where each label indicates whether a given pair of shots belong to the same cluster. We then compute the Average Precision (AP) between the predicted + +Table 2. Shot scale classification on MovieShots. + +
MetricParagraphSentence
PearsonSpearmanPearsonSpearman
CIDEr [71]0.2050.2640.4120.528
ROUGE-L [44]0.3050.2800.5260.512
SPICE [4]0.0220.0480.0310.012
BERTScore [91]0.3770.3930.5080.507
LLM-based (GPT-4o [53])0.7420.6780.7970.807
Action Score (w/o verb matching)0.7350.7280.7650.790
Action Score (w verb matching)0.7490.7290.8060.820
+ +Table 3. Comparison of action score with other metrics. The listed metrics measure the similarity between predicted paragraphs/sentences and ground truth actions. The reported values indicate the correlation (i.e. alignment) between these metrics and human-annotated scores. + +relationships $\{s_{i,j}\}$ and the ground truth $\{\hat{s}_{i,j}\}$ , as well as report the precision and recall values. Additionally, we directly compare the predicted clusters with ground truth clusters by reporting the weighted clustering purity (WCP) [69]. + +Tab. 1 verifies the choice of DINOv2-g14 [54] as the feature extractor for frame pair comparison. Compared to abstract CLS tokens, dense spatial matching achieves higher AP in frame pair relationship prediction and higher WCP in thread clustering. Additionally, we observe that using the temporally closest frames from two shots ("Side") leads to a noticeable performance improvement. This can be attributed to the continuous story flow across shots within the same thread. + +# 5.4. Shot scale classification + +Following prior work on shot scale classification [42, 46], we use the MovieShots [62] dataset, which consists of 46k shots (train:val:test = 7:1:2) collected from over 7k movie trailers. We follow its definition of shot scales, categorising shots into five classes ranging from extreme close-up to long shots, as illustrated in Fig. 4 (top). To evaluate the model performance, we report classification accuracy and Macro-F1 [42] scores on the MovieShots test set. + +Since previous state-of-the-art methods on shot scale classification are not open-sourced, we develop a new network by fine-tuning DINOv2 [54], achieving superior performance over prior approaches that rely on additional optical flow or SAM-based mask inputs, as demonstrated in Tab. 2. + +# 5.5. Human alignment with action scores + +Action scores aim to evaluate whether a GT action is captured within a predicted description, making them recall-oriented. Such descriptions can be in the form of paragraphs (Stage I descriptions) or single sentences (Stage II ADs). To assess whether action scores align with human judgments, we create a dataset containing pairs of predicted descriptions and GT actions. For each GT action, human annotators manually annotate the quality of predictions into $\{0,1,2,3\}$ based on the relevance towards GT action ranging from "unrelated" (0) to "exact matching" (3). + +Comparison with other metrics. Next, we use the human-annotated scores as a reference to compare different metrics in terms of human agreement (measured by correlations), as reported in Tab. 3. Additionally, we consider an LLM-based + +
Exp.Temporal contextFrame samplingShot labelCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
A---22.445.727.026.042.222.3
B1 shotDyn.Shot num.24.746.827.828.941.623.1
C2 shotsDyn.-24.546.627.427.842.322.3
D2 shotsDyn.Frame box24.846.427.425.041.822.8
E2 shotsUni.Shot num.24.147.526.226.542.222.0
F2 shotsDyn.Shot num.25.147.527.828.942.123.0
+ +Table 4. Leveraging shot-based temporal context. Key changes relative to the default setting (Exp. F) are highlighted in light blue. "Temporal context" indicates the number of context (past & future) shots. "Shot num." refers to overlaying the shot number at the top-left of each current shot frame, while "frame box" represents highlighting the boundary of each current shot frame with a red box. + +
Thread structureCMD-AD subsetTV-AD subset
CIDErCRITICActionCIDErCRITICAction
X29.947.527.428.842.022.6
✓30.7↑0.848.9↑1.427.7↑0.330.7↑1.942.7↑0.722.9↑0.3
+ +metric to predict scores, following the same scoring criteria as human annotations. In general, the action score achieves the best correlations with human annotations. Note, it is also more efficient than LLM-based metrics, with 0.15s compared to 6s per prediction evaluation. For more details regarding this human agreement study, please refer to the Supp. Mat. + +# 5.6. AD generation - evaluation of components + +Shot-based temporal context. We investigate different setups for leveraging temporal context in AD generation, as shown in Tab. 4, leading to the following observations: (i) Expanding the temporal context range noticeably boosts the performance, with gains saturating around “2 shots” (Exp. A, B, and F); (ii) “Shot number referral” is the most effective strategy for outlining the current shot. (Exp. C, D, and F); (iii) Dynamically sampling the current shots at a higher frame rate boosts AD generation (Exp. E and F). The latter two improvements can be attributed to more efficient focus on the visual content around the targeted AD interval. + +Thread structure injection. After extending the context information with neighbouring shots, we further enhance the VideoLLM's understanding by incorporating thread structures. Note that this guidance is applied only to video sequences exhibiting thread structures. As shown in Tab. 5, this additional information improves AD generation performance across both datasets. + +Scale-dependent Stage I factors. Tab. 6 explores the impact of different Stage I factors on final AD performance. Using shot scales as guidance for Stage I factor formulation (i.e. scale-dependent) not only outperforms configurations with single fixed factors but also surpasses the case where all factors are included in Stage I. This could be attributed to that the scale-dependent description contains more relevant and less redundant information, enabling more efficient AD extraction in Stage II. + +Table 5. Thread structure injection. Thread structure information is injected only into subsets predicted to exhibit thread structures ( $\sim 30\%$ in CMD-AD and $\sim 60\%$ in TV-AD). + +
Stage I factorsCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
Base25.447.427.729.242.023.0
Base + Face (AutoAD-Zero)25.246.827.830.042.322.5
Base + Obj.26.145.827.930.442.922.8
Base + Env.25.246.727.429.840.722.2
Base + Face + Obj. + Env.26.047.227.430.142.722.9
Scale-dependent (Ours)26.347.828.431.142.223.9
+ +Table 6. Factors included in Stage I description. Base: character + action + interaction; Face: facial expression; Env.: environment; Obj.: object. "Scale-dependent" refers to our approach, which leverages shot scale predictions to determine the relevant factors for each clip. + +# 5.7. AD generation - comparison with SotA + +Tab. 7 provides a comprehensive summary of AD generation performance on CMD-AD and TV-AD, comparing across training-free methods with and without proprietary models, as well as models fine-tuned on human-annotated ADs. Notably, with the same base model setup (Qwen2-VL-7B + LLaMA3-8B), our training-free framework significantly outperforms AutoAD-Zero, primarily due to the usage of temporal context and film grammar information. By incorporating the more powerful GPT-4o models, our performance scales up further, surpassing even existing fine-tuned models. Additionally, we also report our performance on the MAD-Eval benchmark in Supp. Mat. + +Qualitative visualisations. Fig. 6 presents several qualitative examples, where the top two cases illustrate how temporal context information aids in identifying key objects. + +In the bottom-left example, prior methods fail to associate characters, leading to the omission of the man (Alonso). In contrast, our method recognises the thread structure (i.e. [Shot 0, Shot 2], [Shot 1, Shot 3]), which guides the correct prediction of the man's gaze direction towards the lying woman. + +The bottom-right example highlights the effectiveness of scale-dependent Stage I factor formulation. AutoAD-Zero, designed to query characters, actions, interactions, and facial expressions, sometimes overlooks environmental details. Our method, in contrast, correctly identifies the shot as a long shot and instructs the VideoLLM to incorporate environmental context, resulting in more accurate scene descriptions. For more visualisations, please refer to the Supp. Mat. and Supp. Videos. + +# 5.8. Assisted AD generation + +The subjective nature of AD sets a practical limit on metric scores, lower than the theoretical maximum, because human annotators often provide different but valid descriptions. Therefore, beyond enforcing generating a single AD sentence, we also consider employing our framework as an assistant to produce multiple candidate AD sentences. + +To standardise such a protocol, we consider five candidate ADs generated by an assistant and employ an expert to select the best one. To effectively benchmark performance against existing GT ADs, we define the "expert" as an automatic selection mechanism that chooses the candidate with the highest average CIDEr and action score. + +
MethodVLMLLMTraining -freePropriet. modelCMD-ADTV-AD
CIDErCRITICActionR@1/5LLM-AD-EvalCIDErCRITICActionR@1/5LLM-AD-Eval
AutoAD-II [24]CLIP-B32GPT-2XX13.58.2-26.12.08|------
AutoAD-III [25]EVA-CLIPLLaMA2-7BXX25.032.731.531.22.89|2.0126.128.826.430.12.78|1.99
DistinctAD [19]CLIPAD-B16LLaMA3-8BXX22.7--33.02.88|2.0327.4--32.12.89|2.00
Video-LLaMA [90]Video-LLaMA-7B-✓X4.80.0-22.01.89|------
VideoBLIP [87]VideoBLIP-✓X5.20.0-23.61.91|------
AutoAD-Zero [82]VideoLLaMA2-7BLLaMA3-8B✓X17.743.725.526.92.83|1.9622.639.421.727.42.94|2.00
AutoAD-Zero [82]Qwen2-VL-7BLLaMA3-8B✓X21.944.326.930.83.00|2.2026.441.622.130.43.05|2.27
OursQwen2-VL-7BLLaMA3-8B✓X26.347.828.433.03.15|2.4231.142.223.933.13.09|2.35
AutoAD-Zero [82]GPT-4oGPT-4o✓✓22.445.130.732.93.08|2.4930.944.426.834.73.12|2.57
OursGPT-4oGPT-4o✓✓26.149.132.536.53.17|2.6634.246.527.436.63.12|2.59
+ +![](images/71ca775083e5ae89d3985eab88d55be7abf93be590ebffd34c975c33a71d0780.jpg) + +![](images/eb346c808e3b929cf4e6eddd051180befeac426aebc5d869f4d791584ce59a41.jpg) +Figure 6. Qualitative visualisations. Current shots (corresponding to AD intervals) are outlined by red boxes for illustration purposes only. For simplicity, not all context shots are shown. Training-free methods adopt Qwen2-VL + LLaMA3. Examples are taken from The Big Bang Theory (S2E14) (top left), Inferno (2016) (top right), The Asphalt Jungle (1950) (bottom left), and Signs (2002) (bottom right). The top-left example demonstrates the benefits of shot-based temporal context, where the objects (i.e. bills) in Penny's hands are not clearly visible within the AD interval (Shot 1), leading to ambiguous or incorrect predictions by AutoAD-Zero. In contrast, our method successfully identifies the objects from the context shot (Shot 2). The top-right example describing the action of Harry Sims similarly verifies the effectiveness of incorporating context shots. + +![](images/fdb92f3ea35988d1b9d02d4fc9587ec7874849dab8cd92bbe7fb319f4d024979.jpg) + +![](images/5a01d0a72273ec5c5256f8d69acf912d78237083f84543dd2cd0f1dd9cbcbd63.jpg) + +Table 7. Quantitative comparison on CMD-AD and TV-AD. For training-free methods, "VLM" and "LLM" refer to the models used in separate stages, while for fine-tuned models, they denote the pre-trained components within an end-to-end model. + +
MethodCandidate samplingCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
OursSingle AD (Ref.)26.347.828.431.142.223.9
OursIndep. output (p=0.90;τp=0.6)33.349.632.541.244.328.6
OursIndep. output (p=0.95;τp=1.5)37.050.335.145.546.431.5
AutoAD-Zero [82]Joint output31.646.433.843.246.830.1
OursJoint output38.449.235.751.347.331.8
+ +Table 8. Assisted AD generation results. All methods adopt Qwen2-VL + LLaMA3-8B as base models. The first row provides single AD generation results as references (labelled in gray), the rest rows report the performance of one selected AD out of five candidates. "Indep. output" denotes five random independent Stage II runs, with $p$ as the hyperparameter for top-p (nucleus) sampling and ${\tau }_{p}$ as the sampling temperature. "Joint output" generates five ADs simultaneously in a single run. + +To develop an AD generation assistant, we fix the Stage I dense descriptions and explore generating multiple candidates in Stage II. This can be achieved by either running Stage II independently five times (termed the "independent output" setup) or generating five AD outputs simultaneously within a single run (termed the "joint output" setup). As observed in Tab. 8, the assistant-based setup significantly improves upon the single AD performance, highlighting the potential + +of training-free methods in effectively capturing the desired content for AD generation. Within the "independent output" setup, increasing the randomness of sampling (i.e. higher $p$ and $\tau_p$ ) enhances the quality of the selected AD, owing to greater candidate diversity. Meanwhile, the "joint output" setup achieves superior performance, which could be attributed to reduced information redundancy across the simultaneously generated ADs. For additional visualisations, discussions, and detailed text prompts, please refer to the Supp. Mat. + +# 6. Discussion - summary and limitations + +We have demonstrated the benefit of shot-based context and film grammar awareness in AD generation – our training-free two-stage framework achieves state-of-the-art performance among all training-free counterparts, even surpassing fine-tuned models on multiple benchmarks. + +The current framework has two main limitations: (i) the performance depends on the base VideoLLM, which may occasionally hallucinate details inconsistent with the visual content; and (ii) story-level context is not incorporated into the AD generation process. These limitations could potentially be addressed in future work by improving visual grounding and extending the visual and textual context to include the plot. + +Acknowledgements. This research is supported by the UK EPSRC Programme Grant Visual AI (EP/T028572/1), a Royal Society Research Professorship RSRP\R\241003, a Clarendon Scholarship, and ANR-21-CE23-0003-01 CorVis. + +# References + +[1] Guidelines for audio descriptors. https://adp.acb.org/guidelines.html.2 +[2] Audiovault. https://audiovault.net/. 2 +[3] Pyscenedetect. https://www.scenedetect.com/. 5 +[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 6 +[5] Dawit Mureja Argaw, Fabian Caba Heilbron, Joon-Young Lee, Markus Woodson, and In So Kweon. The anatomy of video editing: A dataset and benchmark suite for ai-assisted video editing. In ECCV, 2022. 2 +[6] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 6 +[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5 +[8] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 2 +[9] Andrew Brown, Vicky Kalogeiton, and Andrew Zisserman. Face, body, voice: Video person-clustering with multiple modalities. In ICCV 2021 Workshop on AI for Creative Video Editing and Understanding, 2021. 2 +[10] Boris Chen, Amir Ziai, Rebecca S. Tucker, and Yuchen Xie. Match cutting: Finding cuts with smooth visual transitions. In WACV, 2023. 2 +[11] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, pages 190-200, 2011. 3 +[12] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, 2021. 3 +[13] Shimin Chen, Xiaohan Lan, Yitian Yuan, Zequn Jie, and Lin Ma. Timemarker: A versatile video-llm for long and short video understanding with superior temporal localization ability. arXiv preprint arXiv:2411.18211, 2024. 3 +[14] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 3 +[15] Zeyu Chen, Yana Zhang, Suya Zhang, and Cheng Yang. Study on location bias of cnn for shot scale classification. Multimedia Tools and Applications, 2022. 2 +[16] Peng Chu, Jiang Wang, and Andre Abrantes. LLM-AD: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2 +[17] Andong Deng, Zhongpai Gao, Anwesa Choudhuri, Benjamin Planche, Meng Zheng, Bin Wang, Terrence Chen, Chen Chen, and Ziyan Wu. Seq2time: Sequential knowledge transfer for + +video llm temporal grounding. arXiv preprint arXiv:2411.16932, 2024.3 +[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, 2021. 3 +[19] Bo Fang, Wenhao Wu, Qiangqiang Wu, Yuxin Song, and Antoni B. Chan. DistinctAD: Distinctive audio description generation in contexts. arXiv preprint arXiv:2411.18180, 2024. 1, 2, 5, 8 +[20] Dennis Fedorishin, Lie Lu, Srirangaraj Setlur, and Venu Govindaraju. Audio match cutting: Finding and creating matching audio transitions in movies and videos. In ICASSP, 2024. 2 +[21] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 3 +[22] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. arXiv preprint arXiv:2410.05643, 2024. 3 +[23] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, 2023. 1, 2, 5 +[24] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD II: The sequel – who, when, and what in movie audio description. In ICCV, 2023. 5, 8 +[25] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD III: The prequel - back to the pixels. In CVPR, 2024. 1, 2, 3, 5, 8 +[26] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In EMNLP, 2018. 3 +[27] Minh Hoai and Andrew Zisserman. Thread-safe: Towards recognizing human actions across shot boundaries. In ACCV, 2014. 2, 6 +[28] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing-song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 3 +[29] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 3 +[30] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3 +[31] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2 +[32] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2 +[33] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops on Multimodal Learning, 2020. 2 +[34] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In CVPR, 2024. 3 +[35] Allan Jabri, Andrew Owens, and Alexei A Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, 2020. 4 +[36] Xuekun Jiang, Libiao Jin, Anyi Rao, Linning Xu, and Dahua Lin. Jointly learning the attributes and composition of shots for boundary detection in videos. Trans. Multi., 2022. 2 + +[37] Steve D. Katz. *Film Directing Shot by Shot: Visualizing from Concept to Screen*. Michael Wiese Productions, ISBN: 1615932976, 2019. 1 +[38] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In CVPR, 2024. 3 +[39] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 3 +[40] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, compositional video question answering. In EMNLP, 2018. 5 +[41] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 3 +[42] Yuzhi Li, Feng Tian, Haojun Xu, and Tianfeng Lu. Toward unified and quantitative cinematic shot attribute analysis. *Electronics*, 2023. 2, 6 +[43] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, 2024. 3 +[44] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 2, 6 +[45] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. In ECCV, 2024. 1, 2, 3 +[46] Fengtian Lu, Yuzhi Li, and Feng Tian. Exploring challenge and explainable shot type classification using sam-guided approaches. Signal, Image and Video Processing, 2024. 2, 6 +[47] Manuel Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net++: revisiting people looking at each other in videos. IEEE TPAMI, pages 1-1, 2020. 2 +[48] Meta. The llama 3 herd of models. arXiv preprint arXiv: 2407.21783, 2024. 6 +[49] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3 +[50] James Monaco. How to Read a Film: Movies, Media, and Beyond. OUP, ISBN: 0571168973, 2009. 1 +[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, 2019. 3 +[52] Walter Murch. In the Blink of an Eye: A Perspective on Film Editing. Silman-James Press, ISBN: 1879505622, 2021. 1 +[53] OpenAI. Gpt-4 technical report. arXiv preprint arXiv: 2303.08774, 2024. 6 +[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4, 6 +[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 2 + +[56] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2 +[57] Alejandro Pardo, Fabio Pizzati, Tong Zhang, Alexander Pondaven, Philip Torr, Juan Camilo Perez, and Bernard Ghanem. Matchdiffusion: Training-free generation of match-cuts. arXiv preprint arXiv:2411.18677, 2024. 2 +[58] Toby Perrett, Tengda Han, Dima Damen, and Andrew Zisserman. It's just another day: Unique video captioning by discriminative prompting. In ACCV, 2024. 3 +[59] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 3 +[60] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPRW, 2024. 3 +[61] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, 2019. 3 +[62] Anyi Rao, Jiaze Wang, Linning Xu, Xuekun Jiang, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A unified framework for shot type classification based on subject centric lens. In ECCV, 2020. 2, 6 +[63] Anyi Rao, Xuekun Jiang, Sichen Wang, Yuwei Guo, Zihao Liu, Bo Dai, Long Pang, Xiaoyu Wu, Dahua Lin, and Libiao Jin. Temporal and contextual transformer for multi-camera editing of tv shows. In Proceedings of the CVEU Workshop at ICCV, 2022. 2 +[64] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 3, 5 +[65] Mattia Savardi, András Bálint Kovács, Alberto Signoroni, and Sergio Benini. Recognition of camera angle and camera level in movies from single frames. In Proceedings of the ACM International Conference on Interactive Media Experiences Workshops, 2023. 2 +[66] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting LLMs to transform video annotations at scale. arXiv:2310.04900, 2023. 3 +[67] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 2 +[68] Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, and Chenliang Xu. Vidcomposition: Can mllms analyze compositions in compiled videos? arXiv preprint arXiv:2411.10979, 2024. 1 +[69] Makarand Tapaswi, Omkar M. Parkhi, Esa Rahtu, Eric Sommerlade, Rainer Stiefelhagen, and Andrew Zisserman. Total cluster: A person agnostic clustering method for broadcast videos. In Proceedings of the Indian Conference on Computer Vision Graphics and Image Processing, 2014. 2, 6 +[70] Bartolomeo Vacchetti and Tania Cerquitelli. Cinematographic shot classification with deep ensemble learning. *Electronics*, 2022, 2 +[71] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 2, 5, 6 + +[72] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2 +[73] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videoollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3 +[74] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, 2018. 2 +[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6 +[76] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 2 +[77] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 3 +[78] Yujia Wang, Wei Liang, Haikun Huang, Yongqi Zhang, Dingzeyu Li, and Lap-Fai Yu. Toward automatic audio description generation for accessible videos. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2 +[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3 +[80] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, 2024. 3 +[81] Yongliang Wu, Xinting Hu, Yuyang Sun, Yizhou Zhou, Wenbo Zhu, Fengyun Rao, Bernt Schiele, and Xu Yang. Number it: Temporal grounding videos like flipping manga. In CVPR, 2025. 3 +[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. In ACCV, 2024. 1, 2, 5, 6, 8 +[83] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 3 +[84] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3 +[85] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. arXiv preprint arXiv:2302.14115, 2023. 3 +[86] Xiaojun Ye, Junhao Chen, Xiang Li, Haidong Xin, Chao Li, Sheng Zhou, and Jiajun Bu. MMAD: Multi-modal movie audio description. In LREC-COLING, 2024. 1, 2 +[87] Keunwoo Peter Yu. Videoblip, 2023. 8 +[88] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali + +Farhadi, and Yejin Choi. MERLOT reserve: Multimodal neural script knowledge through vision and language and sound. In CVPR, 2022. 3 +[89] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. MM-Narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, 2024. 1, 2 +[90] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In ENNLP, 2023. 8 +[91] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In *ICLR*, 2020. 6 +[92] Yuqi Zhang, Bin Guo, Nuo Li, Ying Zhang, Qianru Wang, and Zhiwen Yu. Representation learning of next shot selection for vlog editing. In Proceedings of the CVEU Workshop at ICCV, 2023. 2 +[93] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 3 +[94] Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding using large-scale pre-trained models. arXiv preprint arXiv:2408.16219, 2024.3 +[95] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos In AAAI, 2018. 3 +[96] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 3 +[97] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 3 \ No newline at end of file diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/images.zip b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..389f176f2039875fb65ffeb40e8853444a81b1a5 --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4e78d05254a41e5baedfc121c2470bb0e8043156669d38d2f288e7dc7a1cf8d +size 700805 diff --git a/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/layout.json b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2c390b6ad151b79efd155983405695bed87986d --- /dev/null +++ b/2025/Shot-by-Shot_ Film-Grammar-Aware Training-Free Audio Description Generation/layout.json @@ -0,0 +1,10384 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 120 + ], + "type": "text", + "content": "Shot-by-Shot: Film-Grammar-Aware Training-Free Audio Description Generation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 104, + 140, + 504, + 209 + ], + "blocks": [ + { + "bbox": [ + 104, + 140, + 504, + 209 + ], + "lines": [ + { + "bbox": [ + 104, + 140, + 504, + 209 + ], + "spans": [ + { + "bbox": [ + 104, + 140, + 504, + 209 + ], + "type": "image", + "image_path": "7e2ecb7a89bd739996aaa3094b2113e7a0ecdfd23db37b1591318beec45e157a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 75, + 222, + 533, + 348 + ], + "blocks": [ + { + "bbox": [ + 75, + 222, + 533, + 348 + ], + "lines": [ + { + "bbox": [ + 75, + 222, + 533, + 348 + ], + "spans": [ + { + "bbox": [ + 75, + 222, + 533, + 348 + ], + "type": "image", + "image_path": "460a0df0f9a1e2a515e76ad3232416449c89f32d6817766767c05d63396397fd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 350, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 54, + 350, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 555, + 404 + ], + "type": "text", + "content": "Figure 1. Overview of our training-free Audio Description (AD) generation framework. The input video clip (corresponding to the AD interval) is extended to include adjacent shots, providing richer temporal visual context. Each video frame is labeled with its corresponding shot number. The extended clip then undergoes a film grammar prediction process, where the thread structure and shot scales are estimated. The AD is generated in two stages: Stage I utilises the predicted cinematic information as training-free prompt guidance to produce dense video descriptions. Stage II then employs an LLM to generate a summarised AD output." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 418, + 199, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 418, + 199, + 430 + ], + "spans": [ + { + "bbox": [ + 153, + 418, + 199, + 430 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 439, + 297, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 439, + 297, + 647 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 297, + 647 + ], + "type": "text", + "content": "Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages \"shots\" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure - an action score - specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 661, + 132, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 132, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 132, + 673 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "content": "In movies and TV series, Audio Descriptions (ADs) are narrations provided for the visually impaired, conveying visual information to complement the original soundtrack. Their" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 419, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 523 + ], + "type": "text", + "content": "purpose is to ensure a continuous and coherent narrative flow, enabling audiences to follow the plot effectively. Unlike video captions, ADs are constrained by length, prioritising the most visually salient and story-centric information, such as character dynamics and significant objects, whilst omitting redundant details like background figures or unchanging locations. Additionally, ADs are typically produced by professional narrators in a specific style and format, ensuring coherence while not interfering with the original audio." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 556, + 664 + ], + "type": "text", + "content": "With the advent of Visual-Language Models (VLMs), there has been a growth of interest in automatically generating ADs for both movies [16, 23-25, 45, 72, 86, 89] and TV material [19, 82]. However, as anyone who has ever watched a movie or read a book about film editing knows – the fundamental unit of edited video material is the shot, not the frame [37, 50, 52]. Shots are used to structure the video material, defining the granularity and temporal context through choices in its scale (close-up, long shot, etc.) and its movement (panning, tracking, etc.). The \"film grammar\" is then used to convey meaning through specific editing choices on shot transitions (cuts, fades, etc.), durations, and composition (threads, montage, etc.)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Current approaches to automating AD generation, and video large language models (VideoLLMs) in general, are not shot-aware, hindering interpretation of edited video material that has frequent shot transitions [68]. In this paper, our principal" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16503" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 118 + ], + "type": "text", + "content": "objective is to incorporate shot information and editing structure into the AD generation process. To this end, we consider temporal context in terms of shots, and take account of the two key properties of thread structure and shot scale." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 294, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 294, + 177 + ], + "type": "text", + "content": "Thread structure identifies the sequence of shots captured with the same camera. An example of a scene with multiple simultaneous threads in space-time is shown in Fig. 3. We develop a robust thread clustering method and use its predictions to guide VLMs in understanding shot-wise relationships." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 178, + 294, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 178, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 178, + 294, + 270 + ], + "type": "text", + "content": "Shot scale typically implies the type of content of the frame. For instance, close-up shots often highlight characters' facial expressions or gaze interactions, while long shots tend to depict the overall environment and ambience of the scene, as shown in Fig. 4 (top). We leverage this property by developing an off-the-shelf shot scale classifier and employing a scale-dependent prompting strategy to improve the contextual relevance of VLM-generated descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 295, + 422 + ], + "type": "text", + "content": "In this paper, we develop a training-free AD generation approach. Inspired by AutoAD-Zero [82], as illustrated in Fig. 1, we adopt a two-stage pipeline in which a VideoLLM generates dense text descriptions in the first stage, followed by an LLM that produces the final AD outputs from this text. We improve the effectiveness of training-free approaches by incorporating two key improvements based on: (i) shot-based temporal context; and (ii) the shot scale and thread structure, that are both crucial to cinematic composition and understanding. Previous methods could describe human-object interactions or human-human interactions, such as looks [47], within a frame or shot. With the improvements we introduce, the generated AD also includes such interactions when they are implied by the shot structure." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 526 + ], + "type": "text", + "content": "As well as the challenge of generating ADs, another challenge is how to evaluate the predicted ADs. Apart from conventional metrics [8, 44, 55, 71], several new AD metrics [23, 25, 89] have been proposed, including CRITIC [25] measuring who is mentioned in the AD. However, these metrics fail to emphasise character actions – one of the most critical aspects of ADs. To address this, we introduce an action score, assessing whether the predictions accurately capture the actions described in GT AD, independent of character information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 632 + ], + "type": "text", + "content": "Moreover, as ADs are time-limited, there is always a choice on which visual aspect (characters, actions, objects, etc.) should be included. Consequently, a single video clip can have multiple equally valid ADs, each highlighting a different aspect. This observation is supported by the inter-rater agreement experiments in AutoAD-III [25] and previous user studies [78]. Therefore, beyond evaluating single AD performance, we assess our framework's capability as an assistant to generate multiple AD candidates, and report the performance of the selected one AD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 633, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 713 + ], + "type": "text", + "content": "In summary, we propose an enhanced training-free AD generation framework, with the following contributions: (i) We incorporate shot-based temporal context into AD generation via training-free prompting techniques including shot number referral and dynamic frame sampling; (ii) We develop state-of-the-art methods for thread structure and shot scale predictions, and demonstrate that incorporating predicted film grammar knowl" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 152 + ], + "type": "text", + "content": "edge enhances AD generation; (iii) We improve the current AD evaluation by introducing the character-free action score, and a new assistance-oriented evaluation protocol; (iv) Our approach achieves state-of-the-art performance in training-free AD generation, and furthermore, surpasses fine-tuned models on multiple benchmarks. This is the first time a training-free approach has achieved superior performance to fine-tuned methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 166, + 394, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 394, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 394, + 178 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 186, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 232 + ], + "type": "text", + "content": "Audio Description generation. Efforts have been made to curate Audio Description (AD) datasets for both movies [25, 67] and TV series [82], with human annotations sourced from platforms such as AudioVault [2]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 554, + 360 + ], + "type": "text", + "content": "For automatic AD generation, prior works [19, 23-25, 45, 72] fine-tune pre-trained VLMs on AD annotations to produce descriptions in an end-to-end manner. However, these methods face challenges due to limited high-quality AD annotations and the high computational cost of fine-tuning each new backbone. Alternatively, training-free approaches [16, 82, 86, 89] have gained traction for their scalability and flexibility, allowing customised AD output based on official guidelines [1] or specific needs. Yet, these methods still lag in performance, while our approach is the first to achieve results on par with fine-tuned methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 430 + ], + "type": "text", + "content": "Instead of limiting the video input to each AD interval, UniAD [72] and DistinctAD [19] fine-tune VLMs with multiple AD clip inputs to incorporate broader temporal context. In contrast, our method systematically extends AD clip to adjacent shots and introduces a training-free approach that enables pretrained VideoLLMs to better capture localised temporal context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 466 + ], + "type": "text", + "content": "Film grammar analysis. Prior research has sought to understand film grammar from two major perspectives: (i) intra-shot properties, (ii) shot-wise relationships." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 468, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 525 + ], + "type": "text", + "content": "For individual shots, several datasets [5, 31, 36, 62, 65] categorise their characteristics based on camera setups, including shot scales (examples shown in Fig. 4) and camera movements. Correspondingly, various models [15, 42, 46, 62, 70] have been developed for shot type classification." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 629 + ], + "type": "text", + "content": "Regarding shot-wise relationships, a few datasets [5, 10, 56] have been proposed to explore transitions (i.e. cuts) between shots, which have also been leveraged in video content generation [20, 57, 63, 92]. Beyond pairwise shot transitions, research has also investigated longer temporal contexts with thread-based editing structures. Notably, Hoai et al. introduced the Thread-Safe [27] dataset, demonstrating that thread information can enhance action recognition. These structures have also been utilised in video-based face and people clustering [9, 69]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 675 + ], + "type": "text", + "content": "While prior work on film grammar has mainly focused on classification and generation tasks, we specifically utilise shot scale and thread structure information to enhance AD generation in movies and TV series." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "content": "- Dense video captioning. Dense video captioning is closely related to AD generation. Early works [32, 33, 39, 74, 76] in video captioning typically treat event localisation and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16504" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 141 + ], + "type": "text", + "content": "captioning as independent stages, whereas more recent approaches [12, 18, 34, 38, 51, 58, 61, 77, 80, 85, 96, 97] integrate these tasks in an end-to-end manner. Video captioning benchmarks cover a range of domains, including cooking [95], actions [39], movies [64], TV series [41], and open-domain settings [11, 14, 26, 49, 66, 83, 84, 88, 93]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 224 + ], + "type": "text", + "content": "Temporal grounding in VLM. To equip conventional VLMs with temporal grounding capability, some studies [29, 43, 79] generate additional data with enhanced temporal information for fine-tuning. A more common approach explicitly incorporates temporal information into inputs, either by inserting temporal tokens [13, 17, 21, 22, 28, 30, 59, 73] or embedding time information within visual tokens [29, 43, 79]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 307 + ], + "type": "text", + "content": "Training-free approaches [60, 81, 94] have also been explored for temporally grounded understanding. Notably, a recent work [81] achieves temporal grounding by overlaying frame numbers as visual prompts. Instead of using uniformly sampled timestamps, we leverage the natural shot structures in movies and TV series, adopting them as fundamental units for training-free temporal referral." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 266, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 266, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 266, + 335 + ], + "type": "text", + "content": "3. Training-free AD generation framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "text", + "content": "Given a video clip " + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathcal{I}_0,\\dots,\\mathcal{I}_T\\}" + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "text", + "content": ", the task of audio description is to generate a concise narration " + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "text", + "content": " describing what happens around a given AD interval " + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "inline_equation", + "content": "[t_A,t_B]" + }, + { + "bbox": [ + 55, + 340, + 296, + 398 + ], + "type": "text", + "content": ". In this work, we propose a two-stage framework that leverages VideoLLMs and LLMs to predict ADs in a training-free manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "text", + "content": "In Stage I, we employ a VideoLLM that takes a sequence of frames (from multiple shots) as input and generates a dense description " + }, + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "text", + "content": ", guided by instructions " + }, + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{VideoLLM}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 438, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 438, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 102, + 438, + 295, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = \\operatorname {V i d e o L L M} (\\mathcal {V}, [ t _ {A}, t _ {B} ], \\mathcal {P} _ {\\text {V i d e o L L M}}) \\tag {1}", + "image_path": "154e5409810320cbe56890e9f662d0f14aaed884137a5f9bcb0bc76343978a9b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "text", + "content": "In Stage II, we then prompt an LLM (with instructions " + }, + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "text", + "content": ") to extract key information from the dense Stage I description and format it into an AD-style narration " + }, + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 455, + 296, + 489 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 495, + 295, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 495, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 119, + 495, + 295, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {N} = \\operatorname {L L M} \\left(\\mathcal {D}, \\left[ t _ {A}, t _ {B} \\right], \\mathcal {P} _ {\\text {L L M}}\\right) \\tag {2}", + "image_path": "7eaec4d933b3d40a63626cb24f150fb8d8a2bc9a604cc3dd5307df4b05fd5b0d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 593 + ], + "type": "text", + "content": "In this section, we focus on enhancing the visual understanding of the Stage I VideoLLM for edited video material, and make three innovations: In Sec. 3.1, we incorporate shot-based temporal context into Stage I visual inputs. In Sec. 3.2, we leverage the thread structure to enrich cross-shot understanding. Finally, in Sec. 3.3, we incorporate shot-scale awareness into Stage I prompt formulation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 603, + 266, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 266, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 266, + 616 + ], + "type": "text", + "content": "3.1. Leveraging shot-based temporal context" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "content": "Regarding the visual inputs to VideoLLM, prior works [25, 45] often sample frames " + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_{t_A},\\dots,\\mathcal{I}_{t_B}\\}" + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "content": " that directly correspond to the AD interval " + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "inline_equation", + "content": "[t_A,t_B]" + }, + { + "bbox": [ + 55, + 620, + 296, + 713 + ], + "type": "text", + "content": ". However, this approach can be problematic due to (i) misalignment between the AD interval and the actual timestamps when the action occurs, and (ii) the lack of contextual information from adjacent shots. Therefore, we investigate how incorporating temporal context can enhance the understanding of the video clip." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 172 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 172 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 172 + ], + "type": "image", + "image_path": "66f272fa2845c90db118b889236117301ec8842351092366ffc8e8c1f940bba1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 172, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 225 + ], + "type": "text", + "content": "Figure 2. Shot-based temporal context, where current shots are defined as those temporally overlapping with the AD interval. Past shots and future shots provide extended contextual information. Shot numbers are visually overlaid on the top-left of each frame, and frames within the AD interval are sampled more densely than context frames." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 282 + ], + "type": "text", + "content": "Structuring temporal context with shots. To obtain more visual context information, instead of simply extending the AD interval by fixed timestamps, we explore a more structured approach that treats shots as the fundamental units." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 282, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 374 + ], + "type": "text", + "content": "Specifically, we apply an off-the-shelf shot segmentation model to partition the entire video clip into individual shots. For each AD interval, we first identify the shots that (partially) overlap with it, referred to as \"current shots\", as illustrated in Fig. 2. We then consider at most two \"past shots\" and two \"future shots\" adjacent to current shots as temporal context. For all shots included, we label them sequentially from past to future with a number starting from \"Shot " + }, + { + "bbox": [ + 313, + 282, + 555, + 374 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 282, + 555, + 374 + ], + "type": "text", + "content": "\"." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 376, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 422 + ], + "type": "text", + "content": "Emphasising the targeted AD interval. To ensure that the VideoLLM focuses on describing visual content within the AD interval, we propose two strategies: dynamic frame sampling and shot number referral." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 424, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 493 + ], + "type": "text", + "content": "Dynamic frame sampling. As shown in Fig. 2, to emphasise the frames of interest, we adopt denser sampling within the AD interval (red region) and sparser sampling for the surrounding context frames (yellow region). In practice, we specify fixed numbers of frames to be sampled within and outside the AD interval and apply uniform sampling according to these constraints." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 494, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 598 + ], + "type": "text", + "content": "Shot number referral. To further enhance the attention towards the current shot content, we label each sampled frame with its shot number (e.g. \"Shot 0\") at the top-left. During the formulation of the text prompt, instead of prompting the VideoLLM to \"describe what happened in the video clip\", we ask it to \"describe what happened in [Shot 2, Shot 3]\" (i.e. current shots). Through this visual-textual prompting strategy, we found that the VideoLLM could successfully interpret the meaning of shot numbers and refer to the correct shots." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 604, + 470, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 470, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 470, + 616 + ], + "type": "text", + "content": "3.2. Leveraging thread structure" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 621, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 555, + 714 + ], + "type": "text", + "content": "Movies are generally edited such that viewpoints from two or more cameras are intertwined in shot threads, as illustrated in Fig. 3 (top). These interleaved arrangements of shot threads often imply relationships between objects and characters (e.g. gaze interactions) and their 3D arrangement. To leverage this information for Stage I description, we first determine the thread structure using a separate module, then incorporate it into the VideoLLM through prompt guidance." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16505" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 284 + ], + "type": "image", + "image_path": "415d9bcd947817223b459873e0102686dfe077fe8c3128a52cb332ab845236b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 286, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 340 + ], + "type": "text", + "content": "Figure 3. Thread structure. Top: Example of a thread structure with interleaving shots. Bottom: A training-free approach for thread clustering, where shots are pairwise compared using dense feature matching to construct an adjacency matrix, which is then used to predict the thread structure." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": "Thread structure prediction. We develop a training-free method to predict thread clustering. The problem here is to determine if two shots correspond to the same viewpoint or not. Specifically, given two shots (Shot " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " and a later Shot " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": "), we compare the last frame of Shot " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_{T_i}^i)" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " with the first frame of Shot " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_0^j)" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": ", extracting their DINoV2 [54] features as " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "f_{T_i}^i" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "f_0^j \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "content": "Inspired by [35], we assess frame-wise dense correlations by computing a cost volume between their feature maps:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 464, + 295, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 464, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 121, + 464, + 295, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {p, q} ^ {i, j} = m (p, q) \\circ \\left(\\hat {f} _ {T _ {i}; p} ^ {i} \\cdot \\hat {f} _ {0; q} ^ {j}\\right) \\tag {3}", + "image_path": "f304f39464604dca41bc8bda83e54db6c257be8b2423413532ef6cbca0a93dc0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\hat{f}_{T_i;p}^i" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\hat{f}_{0;q}^{j}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": " are the normalised " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "-th spatial elements in the respective feature maps. The binary attention mask " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "m(p,q)" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": " is set to 1 only if the spatial position of the " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "-th element is within an " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "n\\times n" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": " neighbourhood of the " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "-th element." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": "We then apply a softmax operation along the last dimension " + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "inline_equation", + "content": "(q)" + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": " and find the maximum similarity for each " + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": ", followed by averaging over all " + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": "-th elements to obtain a matching score between Shot " + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": " and Shot " + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 532, + 296, + 579 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 582, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 582, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 99, + 582, + 295, + 616 + ], + "type": "interline_equation", + "content": "s ^ {i, j} = \\frac {1}{N} \\sum_ {p} ^ {N} \\max _ {q} \\left(\\frac {\\exp \\left(\\mathcal {C} _ {p , q} ^ {i , j} / \\tau\\right)}{\\sum_ {l} ^ {N} \\exp \\left(\\mathcal {C} _ {p , l} ^ {i , j} / \\tau\\right)}\\right) \\tag {4}", + "image_path": "ba860aaca40cf0815adba28a3f22edf95c328d4f7fe2a34c8333eff888175e80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "content": " denotes the number of feature patches, and " + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 618, + 295, + 641 + ], + "type": "text", + "content": " is the softmax temperature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": "Intuitively, as shown in Fig. 3, this process effectively checks whether each patch " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " in one shot frame matches with a patch " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "(q)" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " in the other shot frame at a roughly similar spatial position (i.e., within an " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "n\\times n" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " neighbourhood)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": "Finally, we construct an adjacency matrix based on the predicted scores " + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "inline_equation", + "content": "s^{i,j}" + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": " for all possible pairs of Shot " + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": " and Shot" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 548, + 121 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 548, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 548, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 548, + 121 + ], + "type": "image", + "image_path": "3ff0827dcc9824848c65dfe9348fba3c429bc6c7facff7938dfc0ca8a9ab5193.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 125, + 548, + 220 + ], + "blocks": [ + { + "bbox": [ + 316, + 125, + 548, + 220 + ], + "lines": [ + { + "bbox": [ + 316, + 125, + 548, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 125, + 548, + 220 + ], + "type": "image", + "image_path": "ac35de672e581b6d0d470dc8ba4e94bcc1817e4ea72e254c4f0725162a56dd7d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 225, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 278 + ], + "type": "text", + "content": "Figure 4. Shot scales. Top: Examples of five different shot scales. Bottom: Stage I factor composition based on shot scale classification. The scales of the current shots in the clip are predicted and averaged. The resulting effective shot scale then guides the formulation of the Stage I prompt, incorporating additional factors such as facial expressions, etc." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "text", + "content": ". By thresholding the adjacency matrix (threshold " + }, + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 293, + 555, + 327 + ], + "type": "text", + "content": ") and identifying the largest connected components, we cluster the set of shots into multiple threads." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": "Thread structure injection. Once the thread information is obtained, we inject it into the text prompt " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{VideoLLM}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": " for the Stage I VideoLLM. In practice, we conduct this information injection only to clips that exhibit thread structures (i.e. " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{thread}} < N_{\\mathrm{shot}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": "). For each given thread [Shot " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "i,\\ldots ," + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": " Shot " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": "], we formulate the prompt as: \"[Shot " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "i,\\ldots ," + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": " Shot " + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 329, + 554, + 421 + ], + "type": "text", + "content": "] share the same camera setup\". This statement implies that the cameras in these shots maintain consistent angles and scales." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 421, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 479 + ], + "type": "text", + "content": "Rather than simply providing this information, we further engage the VideoLLM by asking it to explain why the given thread structure is correct. This effectively corresponds to a Chain-of-Thought (CoT) process, enhancing its understanding of these repetitive thread structures." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 495, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 495, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 495, + 498 + ], + "type": "text", + "content": "3.3. Leveraging shot scale information" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "text", + "content": "In movies and TV series, shot scales are often carefully designed during filming or post-editing to implicitly convey information to the audience. Our objective is to use the shot scale to choose what should be included in the Stage I prompts. Shot scale classification. We first build a shot scale classification network by fine-tuning a pre-trained DINOv2 [54] model. For all current shots, we classify their shot scales " + }, + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\{S_i\\}" + }, + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "text", + "content": " into one of five classes, represented by values 0-4, as shown in Fig. 4 (top). We then compute their average to obtain the effective shot scale " + }, + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{eff}}" + }, + { + "bbox": [ + 313, + 502, + 555, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "content": "Stage I factor composition. We then leverage the predicted shot scale to determine the factors to include in Stage I instructions. We first consider three fixed factors that form the basis of ADs, namely characters, actions, and interactions. The additional factor can be determined through applying a set of thresholds to " + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{eff}}" + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "content": ", as detailed in Fig. 4 (bottom). For example, for close-up shots (" + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{eff}} \\leq 1.5" + }, + { + "bbox": [ + 313, + 621, + 556, + 714 + ], + "type": "text", + "content": "), we would ask the VideoLLM to additionally describe the \"facial expression\", whereas for long shots" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16506" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 276, + 169 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 276, + 169 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 276, + 169 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 276, + 169 + ], + "type": "image", + "image_path": "6d87d3d749e1e68e37f4e0747526850f2c1a3d038f803923cdedaeaf3de4d644.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 171, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 54, + 171, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 555, + 226 + ], + "type": "text", + "content": "Figure 5. Action score. Left: Hierarchical prediction parsing. Given a predicted paragraph or sentence, it is first divided into a set of sentences (light blue) and action phrases (dark blue). Each action phrase is further processed to obtain its verb lemma (light orange). Middle: Scoring based on semantic similarity. Sentence embeddings for the GT action sentence and the set of predicted sentences and phrases are extracted, with the maximum cosine similarity defined as the similarity score. Right: Scoring based on verb matching. The predicted verb lemma is compared with the GT lemma and multiplied by the corresponding semantic-based similarity. The highest resultant score is defined as the verb score." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 525, + 169 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 525, + 169 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 525, + 169 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 525, + 169 + ], + "type": "image", + "image_path": "ffd3453db4326325f47c85581c0dba5a5bc2449893575bff8be90fdc09fc33be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 236, + 297, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 297, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 297, + 248 + ], + "type": "inline_equation", + "content": "(S_{\\mathrm{eff}} \\geq 3.5)" + }, + { + "bbox": [ + 55, + 236, + 297, + 248 + ], + "type": "text", + "content": ", a description on the \"environment\" will be included." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 259, + 130, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 130, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 130, + 271 + ], + "type": "text", + "content": "4. Action score" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 278, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 296, + 406 + ], + "type": "text", + "content": "In this section, we introduce a new metric, termed \"action score\", which focuses on whether a specific ground truth (GT) action is captured within a predicted Stage I description paragraph or a Stage II AD output. For instance, for the GT action \"He dials the phone\", we want the metric to measure the action of \"dial the phone\", but not be sensitive to character names and other predicted content. Therefore, this metric is designed to possess two key properties: (i) it is character-free, meaning that the presence of character names has minimal impact on the evaluation, and (ii) it is recall-oriented, without penalising additional action information in the prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 297, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 297, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 297, + 477 + ], + "type": "text", + "content": "Preprocessing of GT actions. For each GT AD, we extract the character-free GT actions by (i) replacing character names with pronouns, and (ii) splitting the AD into subsentences, each containing one action (verb). For example, given a GT AD \"Chandler dials the phone, then hurriedly hangs up,\" the extracted GT actions are \"He dials the phone\" and \"He hurriedly hangs up.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 571 + ], + "type": "text", + "content": "Hierarchical parsing of predictions. To process the predicted paragraph during evaluation, we first decompose it into individual sentences (light blue), as shown in the left column of Fig. 5. For each predicted sentence, we then perform rule-based dependency parsing to obtain action phrases (deep blue). Then, all extracted sentences and action phrases are collected to form a prediction set. Additionally, for each action phrase, we extract the corresponding verb lemma (i.e. the root form of verbs)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "text", + "content": "Action score computation. The action score features a combination of semantic-based and verb-matching-based components, as illustrated in the middle and right columns of Fig. 5, respectively. Semantic-based evaluation. For each GT action sentence, we assess its semantic similarity with the set of predicted sentences and phrases. Specifically, we employ a general text embedding (GTE) model to compute sentence-level embeddings for the GT action sentence " + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{GT}}" + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "text", + "content": " and each element in the predicted set " + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "inline_equation", + "content": "\\{e_{\\mathrm{pred};i}\\}" + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "text", + "content": ". By computing cosine similarities and taking the maximum value, the similarity score is defined as " + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{sim}} = \\max_i(s_{\\mathrm{sim};i}) = \\max_i[(e_{\\mathrm{GT}} \\cdot e_{\\mathrm{pred};i}) / (|e_{\\mathrm{GT}}||e_{\\mathrm{pred};i}|)]" + }, + { + "bbox": [ + 55, + 573, + 297, + 702 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 702, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 296, + 713 + ], + "type": "text", + "content": "Verb-matching-based evaluation. In addition to semantic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "text", + "content": "matching, we credit predictions that contain the same verbs as those in GT actions. Practically, we compare the predicted verb lemmas with GT verb lemma and compute binary matching scores " + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "inline_equation", + "content": "\\{m_j\\}" + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "text", + "content": ". These matching scores are further weighted by the corresponding similarity scores " + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "inline_equation", + "content": "\\{s_{\\mathrm{sim},j}\\}" + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "text", + "content": " through element-wise multiplication. Finally, we take the maximum of the regulated scores to compute the verb score, i.e. " + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{verb}} = \\max_j(s_{\\mathrm{sim};j} \\circ m_j)" + }, + { + "bbox": [ + 313, + 236, + 555, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "text", + "content": "To obtain the final action score, we combine the scores from both sources using a weighted average " + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "inline_equation", + "content": "s_{\\text{action}} = \\alpha_{\\text{sim}} \\circ s_{\\text{sim}} + \\alpha_{\\text{verb}} \\circ s_{\\text{verb}}" + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\alpha_{\\text{sim}}" + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\alpha_{\\text{verb}}" + }, + { + "bbox": [ + 313, + 317, + 555, + 352 + ], + "type": "text", + "content": " denoting the weighted factors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 359, + 391, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 391, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 391, + 373 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 437 + ], + "type": "text", + "content": "This section begins with datasets and metrics for AD generation in Sec. 5.1, followed by implementation details in Sec. 5.2. Results on film grammar predictions are presented in Secs. 5.3 and 5.4, while Sec. 5.5 analyses human alignment with action scores. AD generation results are detailed in Secs. 5.6 to 5.8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 442, + 504, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 504, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 504, + 455 + ], + "type": "text", + "content": "5.1. AD generation datasets and metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 459, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 563 + ], + "type": "text", + "content": "We evaluate our framework on AD generation datasets for both movies (CMD-AD [25], MAD-Eval [23]) and TV series (TV-AD [82]). In more detail, CMD-AD is constructed by aligning ground truth ADs with the Condensed Movie Dataset (CMD) [7], comprising 101k ADs (94k for training and 7k for testing) from 1.4k movies. MAD-Eval consists of 6.5k ADs sampled from 10 movies within LSMDC [64]. On the other hand, TV-AD features 34k AD annotations from 13 TV series, with its test set sourced from TVQA [40], consisting of 3k ADs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 620 + ], + "type": "text", + "content": "For AD evaluation, we follow the prior works [19, 25] to assess general prediction quality using CIDEr [71], Recall@k/N [24], and LLM-AD-Eval [25]. We also consider character recognition accuracy using CRITIC [25] and character-free action evaluation using action scores (described in Sec. 4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 625, + 446, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 446, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 446, + 637 + ], + "type": "text", + "content": "5.2. Implementation details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 689 + ], + "type": "text", + "content": "In this section, we provide key implementation details for AD generation and action score computation. For additional information on film grammar predictions and other details, please refer to the Supp. Mat." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 714 + ], + "type": "text", + "content": "Shot detection. To segment the video clip into shots, we use PySceneDetect [3] with the \"Adaptive Detection\" method," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16507" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 148 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 148 + ], + "type": "table", + "html": "
FeatureFrame setupPrecisionRecallAPWCP
CLIP-L14 CLSSide0.6910.6350.7050.922
DINOv2-L14 CLSSide0.7590.6830.7880.933
DINOv2-g14 CLSSide0.7610.6750.7860.936
DINOv2-g14 spatialMid0.8080.7170.8220.953
DINOv2-g14 spatialAll0.8700.7950.8960.964
DINOv2-g14 spatialSide0.8780.7990.9020.965
", + "image_path": "d80fc5e62182020e6096a5a717b58b384de2b55826c2ab7eadb1627a6f779de3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 199, + 272, + 267 + ], + "blocks": [ + { + "bbox": [ + 55, + 150, + 295, + 193 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 295, + 193 + ], + "type": "text", + "content": "Table 1. Thread structure prediction on Thread-Safe. \"Side\" refers to comparisons between the temporally nearest frames in two shots; \"Mid\" refers to comparisons between the middle frames of each shot; \"All\" refers to comparisons between all frame pairs from the two shots." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 199, + 272, + 267 + ], + "lines": [ + { + "bbox": [ + 77, + 199, + 272, + 267 + ], + "spans": [ + { + "bbox": [ + 77, + 199, + 272, + 267 + ], + "type": "table", + "html": "
MetricInputAccuracyMacro-F1
ViViT [6]RGB0.7470.751
SGNet [62]RGB + Flow0.875-
Lu et al. [46]RGB + mask0.892-
Li et al. [42]RGB0.8950.897
OursRGB0.8970.899
", + "image_path": "43a256f9c707d074117de6aff5d021d42f15a739aee0923a616c2f62314bafee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "content": "which compares the ratio of pixel changes with the neighbouring frames. On average, each shot spans 3.5s, 3.3s, 3.8s in CMD-AD, TV-AD, and MAD-Eval, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 327, + 295, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 409 + ], + "type": "text", + "content": "AD generation setting. During dynamic frame sampling, we select a total of 32 frames, with 16 frames uniformly sampled within and outside the AD interval, respectively. For character recognition, we adopt the same visual-textual prompting method proposed by [82], which applies coloured circles around faces for visual character indication. For simplicity, the visualisations in this paper do not display these circle labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 409, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 467 + ], + "type": "text", + "content": "Regarding the base models, we use Qwen2-VL-7B [75] as the VideoLLM in Stage I and LLaMA3-8B [48] as the LLM in Stage II. This setup is used as the default unless stated otherwise. Additionally, we explore the framework with the proprietary GPT-4o [53] model for both stages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "content": "Action score evaluation setting. For action score computation, we set the weight factors as " + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{sim}} = 0.8" + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{verb}} = 0.2" + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "content": ". When aggregating the action score results, we first average over multiple GT actions within each GT AD, and then perform global averaging across all AD samples. Moreover, in practice, we find that most action scores are clustered within the range of " + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "inline_equation", + "content": "0.25 - 0.75" + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "content": ". To improve clarity, we apply further rescaling " + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "inline_equation", + "content": "f(x) = (x - 0.25) \\times 2" + }, + { + "bbox": [ + 55, + 468, + 295, + 583 + ], + "type": "text", + "content": " as post-processing. For evaluations in this paper, unless otherwise specified, we use the action score to assess the Stage II AD outputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 592, + 209, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 209, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 209, + 605 + ], + "type": "text", + "content": "5.3. Thread structure prediction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 655 + ], + "type": "text", + "content": "We evaluate thread structure prediction on Thread-Safe [27], which consists of approximately 4.7k video clips collected from 15 TV series. Each video clip contains a multi-shot scene with corresponding thread clusters manually annotated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 655, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 295, + 713 + ], + "type": "text", + "content": "For evaluation, we first construct an adjacency matrix from the GT clusters and extract binary GT labels " + }, + { + "bbox": [ + 55, + 655, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{s}_{i,j}" + }, + { + "bbox": [ + 55, + 655, + 295, + 713 + ], + "type": "text", + "content": " from the off-diagonal entries, where each label indicates whether a given pair of shots belong to the same cluster. We then compute the Average Precision (AP) between the predicted" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 165 + ], + "blocks": [ + { + "bbox": [ + 88, + 269, + 263, + 278 + ], + "lines": [ + { + "bbox": [ + 88, + 269, + 263, + 278 + ], + "spans": [ + { + "bbox": [ + 88, + 269, + 263, + 278 + ], + "type": "text", + "content": "Table 2. Shot scale classification on MovieShots." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 165 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 165 + ], + "type": "table", + "html": "
MetricParagraphSentence
PearsonSpearmanPearsonSpearman
CIDEr [71]0.2050.2640.4120.528
ROUGE-L [44]0.3050.2800.5260.512
SPICE [4]0.0220.0480.0310.012
BERTScore [91]0.3770.3930.5080.507
LLM-based (GPT-4o [53])0.7420.6780.7970.807
Action Score (w/o verb matching)0.7350.7280.7650.790
Action Score (w verb matching)0.7490.7290.8060.820
", + "image_path": "01f3ec615666b10400b9d61a1f6e0a226058b40445a260baf60825424dbd7757.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 166, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 166, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 553, + 209 + ], + "type": "text", + "content": "Table 3. Comparison of action score with other metrics. The listed metrics measure the similarity between predicted paragraphs/sentences and ground truth actions. The reported values indicate the correlation (i.e. alignment) between these metrics and human-annotated scores." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "content": "relationships " + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "inline_equation", + "content": "\\{s_{i,j}\\}" + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_{i,j}\\}" + }, + { + "bbox": [ + 313, + 220, + 553, + 267 + ], + "type": "text", + "content": ", as well as report the precision and recall values. Additionally, we directly compare the predicted clusters with ground truth clusters by reporting the weighted clustering purity (WCP) [69]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 268, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 554, + 360 + ], + "type": "text", + "content": "Tab. 1 verifies the choice of DINOv2-g14 [54] as the feature extractor for frame pair comparison. Compared to abstract CLS tokens, dense spatial matching achieves higher AP in frame pair relationship prediction and higher WCP in thread clustering. Additionally, we observe that using the temporally closest frames from two shots (\"Side\") leads to a noticeable performance improvement. This can be attributed to the continuous story flow across shots within the same thread." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 369, + 446, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 446, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 446, + 380 + ], + "type": "text", + "content": "5.4. Shot scale classification" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 385, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 477 + ], + "type": "text", + "content": "Following prior work on shot scale classification [42, 46], we use the MovieShots [62] dataset, which consists of 46k shots (train:val:test = 7:1:2) collected from over 7k movie trailers. We follow its definition of shot scales, categorising shots into five classes ranging from extreme close-up to long shots, as illustrated in Fig. 4 (top). To evaluate the model performance, we report classification accuracy and Macro-F1 [42] scores on the MovieShots test set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 478, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 536 + ], + "type": "text", + "content": "Since previous state-of-the-art methods on shot scale classification are not open-sourced, we develop a new network by fine-tuning DINOv2 [54], achieving superior performance over prior approaches that rely on additional optical flow or SAM-based mask inputs, as demonstrated in Tab. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 508, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 508, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 508, + 556 + ], + "type": "text", + "content": "5.5. Human alignment with action scores" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 561, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 666 + ], + "type": "text", + "content": "Action scores aim to evaluate whether a GT action is captured within a predicted description, making them recall-oriented. Such descriptions can be in the form of paragraphs (Stage I descriptions) or single sentences (Stage II ADs). To assess whether action scores align with human judgments, we create a dataset containing pairs of predicted descriptions and GT actions. For each GT action, human annotators manually annotate the quality of predictions into " + }, + { + "bbox": [ + 313, + 561, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{0,1,2,3\\}" + }, + { + "bbox": [ + 313, + 561, + 555, + 666 + ], + "type": "text", + "content": " based on the relevance towards GT action ranging from \"unrelated\" (0) to \"exact matching\" (3)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Comparison with other metrics. Next, we use the human-annotated scores as a reference to compare different metrics in terms of human agreement (measured by correlations), as reported in Tab. 3. Additionally, we consider an LLM-based" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16508" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 295, + 153 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 153 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 153 + ], + "type": "table", + "html": "
Exp.Temporal contextFrame samplingShot labelCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
A---22.445.727.026.042.222.3
B1 shotDyn.Shot num.24.746.827.828.941.623.1
C2 shotsDyn.-24.546.627.427.842.322.3
D2 shotsDyn.Frame box24.846.427.425.041.822.8
E2 shotsUni.Shot num.24.147.526.226.542.222.0
F2 shotsDyn.Shot num.25.147.527.828.942.123.0
", + "image_path": "ec1e0e97f8f07974e88e4b67ae75612ae6f8e3d937dddc95914cf7ce833deb96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 226, + 293, + 275 + ], + "blocks": [ + { + "bbox": [ + 55, + 156, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 219 + ], + "type": "text", + "content": "Table 4. Leveraging shot-based temporal context. Key changes relative to the default setting (Exp. F) are highlighted in light blue. \"Temporal context\" indicates the number of context (past & future) shots. \"Shot num.\" refers to overlaying the shot number at the top-left of each current shot frame, while \"frame box\" represents highlighting the boundary of each current shot frame with a red box." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 226, + 293, + 275 + ], + "lines": [ + { + "bbox": [ + 58, + 226, + 293, + 275 + ], + "spans": [ + { + "bbox": [ + 58, + 226, + 293, + 275 + ], + "type": "table", + "html": "
Thread structureCMD-AD subsetTV-AD subset
CIDErCRITICActionCIDErCRITICAction
X29.947.527.428.842.022.6
✓30.7↑0.848.9↑1.427.7↑0.330.7↑1.942.7↑0.722.9↑0.3
", + "image_path": "f4fe17fc7c107b99458cfe3f74a3cacb14168313a0faf6465f08bc359da5358d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 321, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 392 + ], + "type": "text", + "content": "metric to predict scores, following the same scoring criteria as human annotations. In general, the action score achieves the best correlations with human annotations. Note, it is also more efficient than LLM-based metrics, with 0.15s compared to 6s per prediction evaluation. For more details regarding this human agreement study, please refer to the Supp. Mat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 400, + 277, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 277, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 277, + 414 + ], + "type": "text", + "content": "5.6. AD generation - evaluation of components" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 420, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 546 + ], + "type": "text", + "content": "Shot-based temporal context. We investigate different setups for leveraging temporal context in AD generation, as shown in Tab. 4, leading to the following observations: (i) Expanding the temporal context range noticeably boosts the performance, with gains saturating around “2 shots” (Exp. A, B, and F); (ii) “Shot number referral” is the most effective strategy for outlining the current shot. (Exp. C, D, and F); (iii) Dynamically sampling the current shots at a higher frame rate boosts AD generation (Exp. E and F). The latter two improvements can be attributed to more efficient focus on the visual content around the targeted AD interval." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 619 + ], + "type": "text", + "content": "Thread structure injection. After extending the context information with neighbouring shots, we further enhance the VideoLLM's understanding by incorporating thread structures. Note that this guidance is applied only to video sequences exhibiting thread structures. As shown in Tab. 5, this additional information improves AD generation performance across both datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 621, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 713 + ], + "type": "text", + "content": "Scale-dependent Stage I factors. Tab. 6 explores the impact of different Stage I factors on final AD performance. Using shot scales as guidance for Stage I factor formulation (i.e. scale-dependent) not only outperforms configurations with single fixed factors but also surpasses the case where all factors are included in Stage I. This could be attributed to that the scale-dependent description contains more relevant and less redundant information, enabling more efficient AD extraction in Stage II." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "type": "text", + "content": "Table 5. Thread structure injection. Thread structure information is injected only into subsets predicted to exhibit thread structures (" + }, + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "type": "text", + "content": " in CMD-AD and " + }, + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "type": "inline_equation", + "content": "\\sim 60\\%" + }, + { + "bbox": [ + 55, + 277, + 295, + 309 + ], + "type": "text", + "content": " in TV-AD)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 156 + ], + "type": "table", + "html": "
Stage I factorsCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
Base25.447.427.729.242.023.0
Base + Face (AutoAD-Zero)25.246.827.830.042.322.5
Base + Obj.26.145.827.930.442.922.8
Base + Env.25.246.727.429.840.722.2
Base + Face + Obj. + Env.26.047.227.430.142.722.9
Scale-dependent (Ours)26.347.828.431.142.223.9
", + "image_path": "e7d2508a9f862be3e8f451eda0354800591fb312753e5aa8b096767fb115f0a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 158, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 201 + ], + "type": "text", + "content": "Table 6. Factors included in Stage I description. Base: character + action + interaction; Face: facial expression; Env.: environment; Obj.: object. \"Scale-dependent\" refers to our approach, which leverages shot scale predictions to determine the relevant factors for each clip." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 213, + 520, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 520, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 520, + 226 + ], + "type": "text", + "content": "5.7. AD generation - comparison with SotA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 231, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 357 + ], + "type": "text", + "content": "Tab. 7 provides a comprehensive summary of AD generation performance on CMD-AD and TV-AD, comparing across training-free methods with and without proprietary models, as well as models fine-tuned on human-annotated ADs. Notably, with the same base model setup (Qwen2-VL-7B + LLaMA3-8B), our training-free framework significantly outperforms AutoAD-Zero, primarily due to the usage of temporal context and film grammar information. By incorporating the more powerful GPT-4o models, our performance scales up further, surpassing even existing fine-tuned models. Additionally, we also report our performance on the MAD-Eval benchmark in Supp. Mat." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 396 + ], + "type": "text", + "content": "Qualitative visualisations. Fig. 6 presents several qualitative examples, where the top two cases illustrate how temporal context information aids in identifying key objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 396, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 453 + ], + "type": "text", + "content": "In the bottom-left example, prior methods fail to associate characters, leading to the omission of the man (Alonso). In contrast, our method recognises the thread structure (i.e. [Shot 0, Shot 2], [Shot 1, Shot 3]), which guides the correct prediction of the man's gaze direction towards the lying woman." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 454, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 547 + ], + "type": "text", + "content": "The bottom-right example highlights the effectiveness of scale-dependent Stage I factor formulation. AutoAD-Zero, designed to query characters, actions, interactions, and facial expressions, sometimes overlooks environmental details. Our method, in contrast, correctly identifies the shot as a long shot and instructs the VideoLLM to incorporate environmental context, resulting in more accurate scene descriptions. For more visualisations, please refer to the Supp. Mat. and Supp. Videos." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 556, + 446, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 446, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 446, + 569 + ], + "type": "text", + "content": "5.8. Assisted AD generation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 574, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 643 + ], + "type": "text", + "content": "The subjective nature of AD sets a practical limit on metric scores, lower than the theoretical maximum, because human annotators often provide different but valid descriptions. Therefore, beyond enforcing generating a single AD sentence, we also consider employing our framework as an assistant to produce multiple candidate AD sentences." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 554, + 712 + ], + "type": "text", + "content": "To standardise such a protocol, we consider five candidate ADs generated by an assistant and employ an expert to select the best one. To effectively benchmark performance against existing GT ADs, we define the \"expert\" as an automatic selection mechanism that chooses the candidate with the highest average CIDEr and action score." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16509" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 537, + 178 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 537, + 178 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 537, + 178 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 537, + 178 + ], + "type": "table", + "html": "
MethodVLMLLMTraining -freePropriet. modelCMD-ADTV-AD
CIDErCRITICActionR@1/5LLM-AD-EvalCIDErCRITICActionR@1/5LLM-AD-Eval
AutoAD-II [24]CLIP-B32GPT-2XX13.58.2-26.12.08|------
AutoAD-III [25]EVA-CLIPLLaMA2-7BXX25.032.731.531.22.89|2.0126.128.826.430.12.78|1.99
DistinctAD [19]CLIPAD-B16LLaMA3-8BXX22.7--33.02.88|2.0327.4--32.12.89|2.00
Video-LLaMA [90]Video-LLaMA-7B-✓X4.80.0-22.01.89|------
VideoBLIP [87]VideoBLIP-✓X5.20.0-23.61.91|------
AutoAD-Zero [82]VideoLLaMA2-7BLLaMA3-8B✓X17.743.725.526.92.83|1.9622.639.421.727.42.94|2.00
AutoAD-Zero [82]Qwen2-VL-7BLLaMA3-8B✓X21.944.326.930.83.00|2.2026.441.622.130.43.05|2.27
OursQwen2-VL-7BLLaMA3-8B✓X26.347.828.433.03.15|2.4231.142.223.933.13.09|2.35
AutoAD-Zero [82]GPT-4oGPT-4o✓✓22.445.130.732.93.08|2.4930.944.426.834.73.12|2.57
OursGPT-4oGPT-4o✓✓26.149.132.536.53.17|2.6634.246.527.436.63.12|2.59
", + "image_path": "0a6a63a4608aee770187ddf47428a39193bc0f572b92e7621a9ba4a47424de17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 211, + 301, + 285 + ], + "blocks": [ + { + "bbox": [ + 77, + 211, + 301, + 285 + ], + "lines": [ + { + "bbox": [ + 77, + 211, + 301, + 285 + ], + "spans": [ + { + "bbox": [ + 77, + 211, + 301, + 285 + ], + "type": "image", + "image_path": "71ca775083e5ae89d3985eab88d55be7abf93be590ebffd34c975c33a71d0780.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 78, + 287, + 299, + 369 + ], + "blocks": [ + { + "bbox": [ + 78, + 287, + 299, + 369 + ], + "lines": [ + { + "bbox": [ + 78, + 287, + 299, + 369 + ], + "spans": [ + { + "bbox": [ + 78, + 287, + 299, + 369 + ], + "type": "image", + "image_path": "eb346c808e3b929cf4e6eddd051180befeac426aebc5d869f4d791584ce59a41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 370, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 555, + 436 + ], + "type": "text", + "content": "Figure 6. Qualitative visualisations. Current shots (corresponding to AD intervals) are outlined by red boxes for illustration purposes only. For simplicity, not all context shots are shown. Training-free methods adopt Qwen2-VL + LLaMA3. Examples are taken from The Big Bang Theory (S2E14) (top left), Inferno (2016) (top right), The Asphalt Jungle (1950) (bottom left), and Signs (2002) (bottom right). The top-left example demonstrates the benefits of shot-based temporal context, where the objects (i.e. bills) in Penny's hands are not clearly visible within the AD interval (Shot 1), leading to ambiguous or incorrect predictions by AutoAD-Zero. In contrast, our method successfully identifies the objects from the context shot (Shot 2). The top-right example describing the action of Harry Sims similarly verifies the effectiveness of incorporating context shots." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 531, + 285 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 531, + 285 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 531, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 531, + 285 + ], + "type": "image", + "image_path": "fdb92f3ea35988d1b9d02d4fc9587ec7874849dab8cd92bbe7fb319f4d024979.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 287, + 531, + 367 + ], + "blocks": [ + { + "bbox": [ + 306, + 287, + 531, + 367 + ], + "lines": [ + { + "bbox": [ + 306, + 287, + 531, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 531, + 367 + ], + "type": "image", + "image_path": "5a01d0a72273ec5c5256f8d69acf912d78237083f84543dd2cd0f1dd9cbcbd63.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 57, + 441, + 296, + 533 + ], + "blocks": [ + { + "bbox": [ + 55, + 179, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 555, + 201 + ], + "type": "text", + "content": "Table 7. Quantitative comparison on CMD-AD and TV-AD. For training-free methods, \"VLM\" and \"LLM\" refer to the models used in separate stages, while for fine-tuned models, they denote the pre-trained components within an end-to-end model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 441, + 296, + 533 + ], + "lines": [ + { + "bbox": [ + 57, + 441, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 441, + 296, + 533 + ], + "type": "table", + "html": "
MethodCandidate samplingCMD-ADTV-AD
CIDErCRITICActionCIDErCRITICAction
OursSingle AD (Ref.)26.347.828.431.142.223.9
OursIndep. output (p=0.90;τp=0.6)33.349.632.541.244.328.6
OursIndep. output (p=0.95;τp=1.5)37.050.335.145.546.431.5
AutoAD-Zero [82]Joint output31.646.433.843.246.830.1
OursJoint output38.449.235.751.347.331.8
", + "image_path": "04cbc1fcdcc579e4e1a7e5d77d8f69fb36932880e1793209020e606a79b8dd57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "type": "text", + "content": "Table 8. Assisted AD generation results. All methods adopt Qwen2-VL + LLaMA3-8B as base models. The first row provides single AD generation results as references (labelled in gray), the rest rows report the performance of one selected AD out of five candidates. \"Indep. output\" denotes five random independent Stage II runs, with " + }, + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "type": "text", + "content": " as the hyperparameter for top-p (nucleus) sampling and " + }, + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "type": "inline_equation", + "content": "{\\tau }_{p}" + }, + { + "bbox": [ + 55, + 534, + 297, + 610 + ], + "type": "text", + "content": " as the sampling temperature. \"Joint output\" generates five ADs simultaneously in a single run." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 621, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 297, + 715 + ], + "type": "text", + "content": "To develop an AD generation assistant, we fix the Stage I dense descriptions and explore generating multiple candidates in Stage II. This can be achieved by either running Stage II independently five times (termed the \"independent output\" setup) or generating five AD outputs simultaneously within a single run (termed the \"joint output\" setup). As observed in Tab. 8, the assistant-based setup significantly improves upon the single AD performance, highlighting the potential" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "text", + "content": "of training-free methods in effectively capturing the desired content for AD generation. Within the \"independent output\" setup, increasing the randomness of sampling (i.e. higher " + }, + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "inline_equation", + "content": "\\tau_p" + }, + { + "bbox": [ + 313, + 444, + 556, + 548 + ], + "type": "text", + "content": ") enhances the quality of the selected AD, owing to greater candidate diversity. Meanwhile, the \"joint output\" setup achieves superior performance, which could be attributed to reduced information redundancy across the simultaneously generated ADs. For additional visualisations, discussions, and detailed text prompts, please refer to the Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 556, + 515, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 515, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 515, + 570 + ], + "type": "text", + "content": "6. Discussion - summary and limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 632 + ], + "type": "text", + "content": "We have demonstrated the benefit of shot-based context and film grammar awareness in AD generation – our training-free two-stage framework achieves state-of-the-art performance among all training-free counterparts, even surpassing fine-tuned models on multiple benchmarks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 632, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 556, + 713 + ], + "type": "text", + "content": "The current framework has two main limitations: (i) the performance depends on the base VideoLLM, which may occasionally hallucinate details inconsistent with the visual content; and (ii) story-level context is not incorporated into the AD generation process. These limitations could potentially be addressed in future work by improving visual grounding and extending the visual and textual context to include the plot." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16510" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgements. This research is supported by the UK EPSRC Programme Grant Visual AI (EP/T028572/1), a Royal Society Research Professorship RSRP\\R\\241003, a Clarendon Scholarship, and ANR-21-CE23-0003-01 CorVis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 112, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 112, + 141 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 112, + 141 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 148, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 61, + 148, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 171 + ], + "type": "text", + "content": "[1] Guidelines for audio descriptors. https://adp.acb.org/guidelines.html.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 171, + 229, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 229, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 229, + 182 + ], + "type": "text", + "content": "[2] Audiovault. https://audiovault.net/. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 182, + 260, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 260, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 260, + 193 + ], + "type": "text", + "content": "[3] Pyscenedetect. https://www.scenedetect.com/. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 194, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 224 + ], + "type": "text", + "content": "[4] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 227, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 227, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 296, + 269 + ], + "type": "text", + "content": "[5] Dawit Mureja Argaw, Fabian Caba Heilbron, Joon-Young Lee, Markus Woodson, and In So Kweon. The anatomy of video editing: A dataset and benchmark suite for ai-assisted video editing. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 270, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 300 + ], + "type": "text", + "content": "[6] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 302, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 302, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 302, + 295, + 334 + ], + "type": "text", + "content": "[7] Max Bain, Arsha Nagrani, Andrew Brown, and Andrew Zisserman. Condensed movies: Story based retrieval with contextual embeddings. In ACCV, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 335, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 295, + 388 + ], + "type": "text", + "content": "[8] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 61, + 388, + 296, + 431 + ], + "type": "text", + "content": "[9] Andrew Brown, Vicky Kalogeiton, and Andrew Zisserman. Face, body, voice: Video person-clustering with multiple modalities. In ICCV 2021 Workshop on AI for Creative Video Editing and Understanding, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 432, + 295, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 295, + 462 + ], + "type": "text", + "content": "[10] Boris Chen, Amir Ziai, Rebecca S. Tucker, and Yuchen Xie. Match cutting: Finding cuts with smooth visual transitions. In WACV, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 464, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 295, + 486 + ], + "type": "text", + "content": "[11] David Chen and William B Dolan. Collecting highly parallel data for paraphrase evaluation. In ACL, pages 190-200, 2011. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 487, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 295, + 518 + ], + "type": "text", + "content": "[12] Shaoxiang Chen and Yu-Gang Jiang. Towards bridging event captioner and sentence localizer for weakly supervised dense event captioning. In CVPR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 519, + 295, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 519, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 57, + 519, + 295, + 561 + ], + "type": "text", + "content": "[13] Shimin Chen, Xiaohan Lan, Yitian Yuan, Zequn Jie, and Lin Ma. Timemarker: A versatile video-llm for long and short video understanding with superior temporal localization ability. arXiv preprint arXiv:2411.18211, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 563, + 295, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 563, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 57, + 563, + 295, + 615 + ], + "type": "text", + "content": "[14] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, and Sergey Tulyakov. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In CVPR, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 616, + 295, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 616, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 57, + 616, + 295, + 648 + ], + "type": "text", + "content": "[15] Zeyu Chen, Yana Zhang, Suya Zhang, and Cheng Yang. Study on location bias of cnn for shot scale classification. Multimedia Tools and Applications, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 649, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 649, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 649, + 295, + 680 + ], + "type": "text", + "content": "[16] Peng Chu, Jiang Wang, and Andre Abrantes. LLM-AD: Large language model based audio description system. arXiv preprint arXiv:2405.00983, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[17] Andong Deng, Zhongpai Gao, Anwesa Choudhuri, Benjamin Planche, Meng Zheng, Bin Wang, Terrence Chen, Chen Chen, and Ziyan Wu. Seq2time: Sequential knowledge transfer for" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "text", + "content": "video llm temporal grounding. arXiv preprint arXiv:2411.16932, 2024.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 95, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 126 + ], + "type": "text", + "content": "[18] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 127, + 555, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 127, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 127, + 555, + 159 + ], + "type": "text", + "content": "[19] Bo Fang, Wenhao Wu, Qiangqiang Wu, Yuxin Song, and Antoni B. Chan. DistinctAD: Distinctive audio description generation in contexts. arXiv preprint arXiv:2411.18180, 2024. 1, 2, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 160, + 555, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 160, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 160, + 555, + 191 + ], + "type": "text", + "content": "[20] Dennis Fedorishin, Lie Lu, Srirangaraj Setlur, and Venu Govindaraju. Audio match cutting: Finding and creating matching audio transitions in movies and videos. In ICASSP, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 193, + 555, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 193, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 316, + 193, + 555, + 235 + ], + "type": "text", + "content": "[21] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Xi Chen, and Bo Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. arXiv preprint arXiv:2405.13382, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 236, + 555, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 555, + 268 + ], + "type": "text", + "content": "[22] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. arXiv preprint arXiv:2410.05643, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 555, + 300 + ], + "type": "text", + "content": "[23] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. Autoad: Movie description in context. In CVPR, 2023. 1, 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 301, + 555, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 555, + 332 + ], + "type": "text", + "content": "[24] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD II: The sequel – who, when, and what in movie audio description. In ICCV, 2023. 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 334, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 555, + 365 + ], + "type": "text", + "content": "[25] Tengda Han, Max Bain, Arsha Nagrani, Güll Varol, Weidi Xie, and Andrew Zisserman. AutoAD III: The prequel - back to the pixels. In CVPR, 2024. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 366, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 555, + 398 + ], + "type": "text", + "content": "[26] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with temporal language. In EMNLP, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 420 + ], + "type": "text", + "content": "[27] Minh Hoai and Andrew Zisserman. Thread-safe: Towards recognizing human actions across shot boundaries. In ACCV, 2014. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 421, + 555, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 463 + ], + "type": "text", + "content": "[28] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qing-song Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 464, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 495 + ], + "type": "text", + "content": "[29] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, 2024. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 496, + 555, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 555, + 537 + ], + "type": "text", + "content": "[30] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 539, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 539, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 539, + 554, + 570 + ], + "type": "text", + "content": "[31] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 571, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 555, + 602 + ], + "type": "text", + "content": "[32] Vladimir Iashin and Esa Rahtu. A better use of audio-visual cues: Dense video captioning with bi-modal transformer. In BMVC, 2020. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 604, + 555, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 626 + ], + "type": "text", + "content": "[33] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops on Multimodal Learning, 2020. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 316, + 627, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 627, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 627, + 555, + 658 + ], + "type": "text", + "content": "[34] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video recap: Recursive captioning of hour-long videos. In CVPR, 2024. 3" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 316, + 659, + 555, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 555, + 681 + ], + "type": "text", + "content": "[35] Allan Jabri, Andrew Owens, and Alexei A Efros. Space-time correspondence as a contrastive random walk. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[36] Xuekun Jiang, Libiao Jin, Anyi Rao, Linning Xu, and Dahua Lin. Jointly learning the attributes and composition of shots for boundary detection in videos. Trans. Multi., 2022. 2" + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16511" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 103 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 103 + ], + "type": "text", + "content": "[37] Steve D. Katz. *Film Directing Shot by Shot: Visualizing from Concept to Screen*. Michael Wiese Productions, ISBN: 1615932976, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 137 + ], + "type": "text", + "content": "[38] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In CVPR, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 138, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 138, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 138, + 296, + 167 + ], + "type": "text", + "content": "[39] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 170, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 200 + ], + "type": "text", + "content": "[40] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, compositional video question answering. In EMNLP, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 201, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 296, + 232 + ], + "type": "text", + "content": "[41] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 234, + 296, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 264 + ], + "type": "text", + "content": "[42] Yuzhi Li, Feng Tian, Haojun Xu, and Tianfeng Lu. Toward unified and quantitative cinematic shot attribute analysis. *Electronics*, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 266, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 306 + ], + "type": "text", + "content": "[43] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 308, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 329 + ], + "type": "text", + "content": "[44] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, 2004. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 331, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 371 + ], + "type": "text", + "content": "[45] Kevin Qinghong Lin, Pengchuan Zhang, Difei Gao, Xide Xia, Joya Chen, Ziteng Gao, Jinheng Xie, Xuhong Xiao, and Mike Zheng Shou. Learning video context as interleaved multimodal sequences. In ECCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 373, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 403 + ], + "type": "text", + "content": "[46] Fengtian Lu, Yuzhi Li, and Feng Tian. Exploring challenge and explainable shot type classification using sam-guided approaches. Signal, Image and Video Processing, 2024. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 404, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 435 + ], + "type": "text", + "content": "[47] Manuel Marin-Jimenez, Vicky Kalogeiton, Pablo Medina-Suarez, and Andrew Zisserman. Laeo-net++: revisiting people looking at each other in videos. IEEE TPAMI, pages 1-1, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 437, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 456 + ], + "type": "text", + "content": "[48] Meta. The llama 3 herd of models. arXiv preprint arXiv: 2407.21783, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 458, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 458, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 458, + 296, + 499 + ], + "type": "text", + "content": "[49] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 501, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 520 + ], + "type": "text", + "content": "[50] James Monaco. How to Read a Film: Movies, Media, and Beyond. OUP, ISBN: 0571168973, 2009. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 522, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 296, + 542 + ], + "type": "text", + "content": "[51] Jonghwan Mun, Linjie Yang, Zhou Ren, Ning Xu, and Bohyung Han. Streamlined dense video captioning. In CVPR, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 544, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 563 + ], + "type": "text", + "content": "[52] Walter Murch. In the Blink of an Eye: A Perspective on Film Editing. Silman-James Press, ISBN: 1879505622, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 566, + 296, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 296, + 585 + ], + "type": "text", + "content": "[53] OpenAI. Gpt-4 technical report. arXiv preprint arXiv: 2303.08774, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 586, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 296, + 680 + ], + "type": "text", + "content": "[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 682, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 682, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 682, + 296, + 712 + ], + "type": "text", + "content": "[55] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[56] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "type": "text", + "content": "[57] Alejandro Pardo, Fabio Pizzati, Tong Zhang, Alexander Pondaven, Philip Torr, Juan Camilo Perez, and Bernard Ghanem. Matchdiffusion: Training-free generation of match-cuts. arXiv preprint arXiv:2411.18677, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 150, + 555, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 181 + ], + "type": "text", + "content": "[58] Toby Perrett, Tengda Han, Dima Damen, and Andrew Zisserman. It's just another day: Unique video captioning by discriminative prompting. In ACCV, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 182, + 555, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 555, + 224 + ], + "type": "text", + "content": "[59] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 226, + 555, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 226, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 555, + 257 + ], + "type": "text", + "content": "[60] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPRW, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 259, + 555, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 259, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 555, + 289 + ], + "type": "text", + "content": "[61] Tanzila Rahman, Bicheng Xu, and Leonid Sigal. Watch, listen and tell: Multi-modal weakly supervised dense event captioning. In ICCV, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 292, + 555, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 555, + 332 + ], + "type": "text", + "content": "[62] Anyi Rao, Jiaze Wang, Linning Xu, Xuekun Jiang, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A unified framework for shot type classification based on subject centric lens. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 335, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 555, + 376 + ], + "type": "text", + "content": "[63] Anyi Rao, Xuekun Jiang, Sichen Wang, Yuwei Guo, Zihao Liu, Bo Dai, Long Pang, Xiaoyu Wu, Dahua Lin, and Libiao Jin. Temporal and contextual transformer for multi-camera editing of tv shows. In Proceedings of the CVEU Workshop at ICCV, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 399 + ], + "type": "text", + "content": "[64] Anna Rohrbach, Marcus Rohrbach, Niket Tandon, and Bernt Schiele. A dataset for movie description. In CVPR, 2015. 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "text", + "content": "[65] Mattia Savardi, András Bálint Kovács, Alberto Signoroni, and Sergio Benini. Recognition of camera angle and camera level in movies from single frames. In Proceedings of the ACM International Conference on Interactive Media Experiences Workshops, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 454, + 555, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 555, + 495 + ], + "type": "text", + "content": "[66] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting LLMs to transform video annotations at scale. arXiv:2310.04900, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 498, + 555, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 555, + 539 + ], + "type": "text", + "content": "[67] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 541, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 553, + 592 + ], + "type": "text", + "content": "[68] Yunlong Tang, Junjia Guo, Hang Hua, Susan Liang, Mingqian Feng, Xinyang Li, Rui Mao, Chao Huang, Jing Bi, Zeliang Zhang, Pooyan Fazli, and Chenliang Xu. Vidcomposition: Can mllms analyze compositions in compiled videos? arXiv preprint arXiv:2411.10979, 2024. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 595, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 595, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 595, + 553, + 647 + ], + "type": "text", + "content": "[69] Makarand Tapaswi, Omkar M. Parkhi, Esa Rahtu, Eric Sommerlade, Rainer Stiefelhagen, and Andrew Zisserman. Total cluster: A person agnostic clustering method for broadcast videos. In Proceedings of the Indian Conference on Computer Vision Graphics and Image Processing, 2014. 2, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 648, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 648, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 648, + 553, + 678 + ], + "type": "text", + "content": "[70] Bartolomeo Vacchetti and Tania Cerquitelli. Cinematographic shot classification with deep ensemble learning. *Electronics*, 2022, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 712 + ], + "type": "text", + "content": "[71] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. CIDEr: Consensus-based image description evaluation. In CVPR, 2015. 2, 5, 6" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16512" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 105 + ], + "type": "text", + "content": "[72] Hanlin Wang, Zhan Tong, Kecheng Zheng, Yujun Shen, and Limin Wang. Contextual ad narration with interleaved multimodal sequence. arXiv preprint arXiv:2403.12922, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "text", + "content": "[73] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videoollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 296, + 180 + ], + "type": "text", + "content": "[74] Jingwen Wang, Wenhao Jiang, Lin Ma, Wei Liu, and Yong Xu. Bidirectional attentive fusion with context gating for dense video captioning. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 182, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 296, + 245 + ], + "type": "text", + "content": "[75] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 296, + 289 + ], + "type": "text", + "content": "[76] Teng Wang, Huicheng Zheng, Mingjing Yu, Qian Tian, and Haifeng Hu. Event-centric hierarchical representation for dense video captioning. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 290, + 296, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 296, + 322 + ], + "type": "text", + "content": "[77] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 323, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 296, + 365 + ], + "type": "text", + "content": "[78] Yujia Wang, Wei Liang, Haikun Huang, Yongqi Zhang, Dingzeyu Li, and Lap-Fai Yu. Toward automatic audio description generation for accessible videos. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 407 + ], + "type": "text", + "content": "[79] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "type": "text", + "content": "[80] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 441, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 296, + 474 + ], + "type": "text", + "content": "[81] Yongliang Wu, Xinting Hu, Yuyang Sun, Yizhou Zhou, Wenbo Zhu, Fengyun Rao, Bernt Schiele, and Xu Yang. Number it: Temporal grounding videos like flipping manga. In CVPR, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 296, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 516 + ], + "type": "text", + "content": "[82] Junyu Xie, Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. Autoad-zero: A training-free framework for zero-shot audio description. In ACCV, 2024. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 517, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 296, + 548 + ], + "type": "text", + "content": "[83] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 550, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 550, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 550, + 296, + 592 + ], + "type": "text", + "content": "[84] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 594, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 296, + 645 + ], + "type": "text", + "content": "[85] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. arXiv preprint arXiv:2302.14115, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 679 + ], + "type": "text", + "content": "[86] Xiaojun Ye, Junhao Chen, Xiang Li, Haidong Xin, Chao Li, Sheng Zhou, and Jiajun Bu. MMAD: Multi-modal movie audio description. In LREC-COLING, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 211, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 211, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 211, + 691 + ], + "type": "text", + "content": "[87] Keunwoo Peter Yu. Videoblip, 2023. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[88] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 430 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Farhadi, and Yejin Choi. MERLOT reserve: Multimodal neural script knowledge through vision and language and sound. In CVPR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 148 + ], + "type": "text", + "content": "[89] Chaoyi Zhang, Kevin Lin, Zhengyuan Yang, Jianfeng Wang, Linjie Li, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. MM-Narrator: Narrating long-form videos with multimodal in-context learning. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 553, + 180 + ], + "type": "text", + "content": "[90] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In ENNLP, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 181, + 553, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 181, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 181, + 553, + 213 + ], + "type": "text", + "content": "[91] Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In *ICLR*, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 214, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 214, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 214, + 553, + 255 + ], + "type": "text", + "content": "[92] Yuqi Zhang, Bin Guo, Nuo Li, Ying Zhang, Qianru Wang, and Zhiwen Yu. Representation learning of next shot selection for vlog editing. In Proceedings of the CVEU Workshop at ICCV, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 553, + 289 + ], + "type": "text", + "content": "[93] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 290, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 290, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 553, + 331 + ], + "type": "text", + "content": "[94] Minghang Zheng, Xinhao Cai, Qingchao Chen, Yuxin Peng, and Yang Liu. Training-free video temporal grounding using large-scale pre-trained models. arXiv preprint arXiv:2408.16219, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 364 + ], + "type": "text", + "content": "[95] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos In AAAI, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 396 + ], + "type": "text", + "content": "[96] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 430 + ], + "type": "text", + "content": "[97] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16513" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_content_list.json b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d4d70ad5833cc5a16a0e893ca440719e0e65ef1a --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_content_list.json @@ -0,0 +1,1578 @@ +[ + { + "type": "text", + "text": "SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alex Costanzino1 Pierluigi Zama Ramirez1 Luigi Lella1 Matteo Ragaglia2 Alessandro Oliva2 Giuseppe Lisanti1 Luigi Di Stefano1 \n1CVLab, University of Bologna, Italy 2SACMI Imola, Italy https://alex-costanzino.github.io/SiM3D/", + "bbox": [ + 99, + 184, + 908, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose SiM3D, the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS), where the task is to produce a voxel-based Anomaly Volume. Moreover, SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. In this respect, SiM3D stands out as the first ADS benchmark that addresses the challenge of generalising from synthetic training data to real test data. SiM3D includes a novel multimodal multiview dataset acquired using top-tier industrial sensors and robots. The dataset features multiview high-resolution images (12 Mpx) and point clouds ( $\\sim 7M$ points) for 333 instances of eight types of objects, alongside a CAD model for each type. We also provide manually annotated 3D segmentation GTs for anomalous test samples. To establish reference baselines for the proposed multiview 3D ADS task, we adapt prominent singleview methods and assess their performance using novel metrics that operate on Anomaly Volumes.", + "bbox": [ + 88, + 304, + 485, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 633, + 223, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In computer vision, anomaly detection and segmentation (ADS) involves identifying anomalous samples and localising their defects. This task is challenging in industrial contexts due to the variability and rarity of defects. Typically, ADS in the industry follows a cold-start approach, where training is unsupervised and uses only defect-free (nominal) samples. Before the release of the MVTec AD dataset in 2019 [4], only a limited number of ADS methods were proposed, with just 9 publications in major computer vision conferences (e.g., CVPR, ECCV, ICCV, WACV) over the five years before. The introduction of MVTec AD marked a pivotal moment in the field, fostering the publication of roughly 172 papers in these venues over the next 5 years. This dataset not only stimulated the development of several ADS methods [14, 30, 36] but also promoted the creation of novel and more challenging benchmarks. For instance,", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ced18cd97e949e4bffadc3884fba9fc61b90ce90b5a86edc90653bbcdc06c372.jpg", + "image_caption": [ + "Figure 1. SiM3D dataset overview. From top to bottom: the single-instance real and synthetic training samples for object Plastic Stool, one of the anomalous samples from the test set." + ], + "image_footnote": [], + "bbox": [ + 545, + 271, + 861, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MVTec 3D-AD and [6] and Eyecandies [7] introduced the first benchmarks deploying multimodal inputs to enhance ADS performance. Then, PAD [45] and Real-IAD [35] introduced novel benchmarks that leverage multiple RGB views of objects. Although these benchmarks tackle various settings, two real industrial issues remain unaddressed.", + "bbox": [ + 511, + 776, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, existing benchmarks have primarily addressed the problem of 2D ADS. Consequently, current methods gen", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20944", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/1f1b52d32266d9aa1f3a6d001edefd1b6857ed86f2d3617363c88e98b462b831.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetVenue & Year2DNo. Pixels3DNo. PointsScenarioTrainTestSetupTaskNo. Classes
MVTec ADCVPR 2019RGB Images490k - 1M——Multi-instanceSingleviewSingleviewreal2real2D15
MVTec 3D-ADVISAPP 2021RGB Images160k - 810kXYZ Images10k - 195kMulti-instanceSingleviewSingleviewreal2real2D10
EyecandiesACCV 2022RGB Images262kDepths + Normals262kMulti-instanceSingleviewSingleviewsynth2synth2D10
MVTec LOCO ADIJCV 2022RGB Images1.28M - 2.05M——Multi-instanceSingleviewSingleviewreal2real2D5
VisAECCV 2022RGB Images1.5M——Multi-instanceSingleviewSingleviewreal2real2D12
PADNeurIPS 2023RGB Images12M——Single-instanceMultiviewSingleviewreal2real + synth2synth2D20
Real3D-ADNeurIPS 2023——Point Clouds43k - 2.68MMulti-instanceMultiviewSingleviewreal2real3D12
Real-IADCVPR 2024RGB Images10M——Multi-instanceSingleviewSingleviewreal2real2D30
SiM3D2025Grayscale Images12MPoint Clouds + Mesh5M - 7MSingle-instanceMultiviewMultiviewreal2real + synth2real3D8
", + "bbox": [ + 94, + 88, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Popular ADS benchmarks. Chronologically sorted, the sequence hints at a growing interest in multimodal and multiview data.", + "bbox": [ + 93, + 190, + 898, + 204 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d25cb59f2febfc8be8552ba1c6a8346ef4dad0420ec11fa46f90073270c39a42.jpg", + "image_caption": [ + "Figure 2. 2D vs. 3D anomaly detection and segmentation. Typical ADS benchmarks rely on 2D Anomaly Maps and 2D ground-truths (left) to assess segmentation performance. SiM3D proposes to do so by 3D Anomaly Volumes and 3D ground-truths (right)." + ], + "image_footnote": [], + "bbox": [ + 91, + 229, + 480, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erate Anomaly Maps, i.e., images that encode the likelihood of each pixel belonging to a defect, as shown in Fig. 2 (left). However, many industrial applications require precise localisation of defects in the 3D space to allow automatic intervention, reducing waste and optimising production. Benchmarks relying on multimodal input data, like those proposed in MVTec 3D-AD [6] and Eyecandies [7], represent progress toward this goal by providing 3D information for each pixel within an image. However, in these benchmarks, ADS methods are still expected to provide as output a 2D Anomaly Map, and objects are inspected from a single viewpoint. For comprehensive 3D anomaly detection, we must observe an object in its entirety, utilising captures from multiple viewpoints to detect defects across the entire 3D space. This calls for methods capable of integrating multiview information to estimate the defect probability at each 3D location within an Anomaly Volume, as shown in Fig. 2 (right). Furthermore, as demonstrated by MVTec 3D-AD and Eyecandies, these methods should process multimodal information, i.e., RGB images as well as 3D information, to achieve more reliable anomaly detection.", + "bbox": [ + 88, + 415, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, current benchmarks typically provide multiple training samples per object class to capture the variability across nominal objects. However, collecting many samples - even nominal ones - can be costly and challenging. For example, line changeovers would demand acquiring new datasets and retraining models, both highly time-intensive processes. Additionally, collecting large datasets increases the risk of errors, such as misjudged nominal samples, which may degrade model performance. Moreover, manufactured objects typically closely resemble one another as they replicate the structure and materials of a syn", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thetic CAD prototype. Thus, the variability across nominal samples of manufactured objects is frequently minimal, with a single nominal item subsuming the information required to spot anomalies in other objects.", + "bbox": [ + 511, + 232, + 903, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These considerations highlight the need for ADS techniques that may be effectively trained by a single nominal sample or even a synthetic prototype (e.g., an object's CAD model) and then be able to generalise to other manufactured objects to identify anomalous ones effectively. The ability to perform training by a single object instance, either real or synthetic, would avoid the need for extensive data collection, vastly facilitating the creation and adaptation of ADS models. Yet, most current benchmarks focus on multiple training instances setups, and none of the prior works consider challenging synthetic (training) to real (test) setups.", + "bbox": [ + 511, + 292, + 905, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fill all the gaps highlighted by our considerations and stimulate interest in the development of multimodal and multiview 3D ADS methods, we propose the Single-instance Multiview Multimodal Multisetup dataset, dubbed SiM3D. The dataset contains a series of high-resolution scans of manufactured objects, which can be nominal or anomalous. Each scan consists of multiple views, each featuring a greyscale image, its corresponding point cloud (i.e., multimodal data), and the ground-truth acquisition pose. In addition, an integrated 3D mesh is provided with each scan. For each object, a single nominal sample has been retained to be employed as a single-instance training set, while the remaining samples, both nominal and anomalous, compose the test set. Moreover, to foster research on novel approaches that may address the domain gap between synthetic training data and real test samples in ADS, for each object of our dataset, we also provide a synthetic scan, i.e., we deploy the available CAD model to obtain the same kind of data as in our high-resolution real scans, with images and point clouds rendered from the same poses. An overview of the proposed dataset is depicted in Fig. 1.", + "bbox": [ + 511, + 460, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on our novel dataset, we create the first benchmark for 3D anomaly detection, where the task is to predict anomaly scores within a voxelized 3D Anomaly Volume rather than a 2D Anomaly Map (Fig. 2). Peculiarly, our benchmark includes two setups based on the domain of the data available for training: in the real2real setup, a model is given access to a single nominal instance of an object (Fig. 1, top), while in synth2real the same kind", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20945", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of training data are obtained from the CAD model of the object (Fig. 1, centre). We adapt prominent ADS methods – both unimodal and multimodal – for the proposed task and evaluate them on our benchmark. To assess segmentation performance, we provide accurate 3D ground-truths (Fig. 2) and extend ADS metrics to operate on anomaly scores and ground-truths specified on 3D voxel grids. Our experimental findings highlight the open challenges in the field and offer insights into possible future directions.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 239, + 232, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The MVTec AD [4] dataset significantly boosted anomaly detection research, inspiring new techniques and benchmarks for various challenges. MVTec AD provides a training set containing only nominal images and a test set that includes nominal and anomalous ones for several object categories. This unsupervised setup exemplifies many real-world applications in which anomalous objects are challenging to collect, and thus, it has also been adopted by more recent ADS benchmarks. Among them, MVTec LOCO [5] presents a challenging scenario with both structural anomalies (e.g., dents, holes) and logical anomalies (e.g., incorrect object combination). VisA [46] introduced high-resolution images of complex scenes containing multiple instances of the same object. However, certain defects may be visible only by analysing objects' 3D geometry. For this reason, multimodal datasets, such as MVTec 3D-AD [6] and Eyecandies [7] include 3D data alongside images. For each sample in the dataset, MVTec 3D-AD provides images and pixel-aligned XYZ information captured by a 3D sensor, while Eyecandies contains synthetic images with pixel-aligned depths and normals. Recently, PAD [45] addressed pose-agnostic ADS with the first multiview dataset (MAD), featuring images of objects from various viewpoints. In PAD, training samples are multiview RGB images with known poses, while test samples are singleview RGB images with unknown poses. MAD includes both synthetic and real objects, but its benchmarks do not explore synthetic-to-real generalisation. Real-IAD [35] introduced a large-scale multiview dataset with RGB images covering various defect sizes. Eventually, Real3D-AD [21] is the first point cloud anomaly detection benchmark where the task is to predict an anomaly score for each point. It uses front-and-back scans of nominal objects for training and a singleview point cloud for testing.", + "bbox": [ + 89, + 265, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Tab. 1, the proposed SiM3D benchmark offers several outstanding features, including very-high-resolution images (12 Mpx), large point clouds/meshes containing 5-7 million elements captured with high-precision 3D sensors. Moreover, it is designed for a highly relevant manufacturing scenario: single-instance anomaly detection, where only one object is available for model training. To promote research toward even more cost-effective scenar", + "bbox": [ + 89, + 780, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ios, ours is also the only benchmark that addresses the challenge of generalising from synthetic training data to real test data. Finally, SiM3D stands out as the first benchmark that considers integrating multiview information at test time for comprehensive 3D anomaly detection, where the task consists of outputting a voxel-based Anomaly Volume.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ADS Methods. Following the setup defined in the MVTec-AD benchmark, most ADS methods [22] are designed to estimate 2D anomaly maps from a single image [2, 3, 9, 10, 13, 15, 16, 19, 20, 23, 26, 27, 27-29, 31, 34, 37-41, 43, 44]. One of the most popular methods is PatchCore [30], which uses feature extractors pre-trained on large datasets [8, 17, 24] to gather features from nominal samples into a memory bank. At test time, input sample features are compared to this bank to detect anomalies. Another more recent yet popular approach, EfficientAD [1], uses a lightweight feature extractor and a student-teacher model to identify anomalies based on feature prediction discrepancy. The introduction of the MVTec 3D-AD benchmark has led to the development of several multimodal approaches that use images and 3D data to generate 2D anomaly maps [14, 18, 32, 36]. Among them, BTF [18] is a memory bank approach inspired by PatchCore, in which each element is the concatenation of 2D and 3D features extracted by a pre-trained image backbone and a traditional point cloud descriptor, i.e., FPFH [33], respectively. M3DM [36] improves BTF using powerful transformer-based pre-trained backbones, such as DINO-v1[8] and Point-MAE [25], to extract features from RGB images and point clouds. It also learns to fuse 2D and 3D features into multimodal features. During inference, a trained One-Class SVM learns to aggregate modality-specific scores. AST [32] follows a Teacher-Student paradigm, in which the Teacher leverages Normalizing Flows to model the multimodal features distribution, while the Student employs a feed-forward network to model such distribution. This asymmetry in the architectures exacerbates the discrepancies between nominal and anomalous features at inference time. Recently, starting from the same transformer-based backbones as in [36], CFM [14] learns to map features from one modality to the other by nominal samples and then, at inference time, detects anomalies by pinpointing inconsistencies between observed and predicted features. In this paper, we employ the above-mentioned methods to investigate the challenges of SiM3D by adapting them to work in a multiview setup.", + "bbox": [ + 511, + 181, + 903, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Creation", + "text_level": 1, + "bbox": [ + 513, + 784, + 679, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material preparation. We scanned eight types of manufactured objects spanning different sizes, shapes, textures and materials, as shown in Fig. 3 (1.i). For each object type, between 20 and 100 instances were purchased from a store. We selected objects suitable for single-instance scenarios – instances are manufactured according to a unique", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20946", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ff934651d17324fe9b5aa23751bafe5d029740d434cd7fc88f931420a69846c.jpg", + "image_caption": [ + "(1) Material preparation" + ], + "image_footnote": [], + "bbox": [ + 106, + 113, + 151, + 148 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4f1f8ee071598b7c23b6deaa553ee65f938240487e0bc84a408e515c28eb0d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 154, + 151, + 176 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5b7bdb6910b4eefc26a97782b5754dafb421bb835adeaac43ff366ce7d78ab29.jpg", + "image_caption": [ + "(i) Real objects purchase" + ], + "image_footnote": [], + "bbox": [ + 153, + 113, + 202, + 179 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e45da2c68f71f6644c49964ab4dac5970ffc875760a72c14c35fa6db6ae0bc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 113, + 254, + 148 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e11ccfd440b9929915ea1741545982a25ca45086c92e3529f6be5e0dbe540465.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 148, + 254, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f23dd7682ba3295026491b4676d9e3ddcfcc8a7ad8248066e0e62abb7d01618b.jpg", + "image_caption": [ + "(2) Acquisition setup" + ], + "image_footnote": [], + "bbox": [ + 264, + 113, + 303, + 180 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/68d138f94156275eddb1246fa993dd2a655fb34b12979bf4a4f3f77c9b7dbc52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 113, + 473, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e75cea0664333c347b6a442fd81cd16cbbe14183f9521011ec893709d1da3031.jpg", + "image_caption": [ + "(ii) CAD models collection" + ], + "image_footnote": [], + "bbox": [ + 106, + 200, + 156, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94ac0ef751c1158402051d318db7692407fa45096236e47ccd96390f6c340b6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 233, + 156, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/129621fc3da7cc9ae5d5ba7cd65cfa2ad4c3eb37afab97fa42472eb3c708e3e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 200, + 207, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f447c1e84704c700eb2bc7c1f6612b56f8d761da53a895692c7195a285176bba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 156, + 233, + 207, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b42dfe144bf9eda909a083b7c16f5a57ca291e7ed48e1339d60393f3324fde33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 200, + 256, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f7f7e34f26bb7aa2a32e98ebe4930be31cd99f3d293d4384fbcdfe52f84d551.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 233, + 256, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efc0722f016b683c0b42077e36695c29d0c07cf3635ea7aa90ca660b190ff892.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 200, + 310, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/874b0a10fcfaf19188adc278c8d6710ad5d2cc9ea8fd419f10d22092ac09f4aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 233, + 310, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7195463304f6052647ece2b6c83c1c9c8dbe6042e79ebc3274030437453fe7dc.jpg", + "image_caption": [ + "High-precision calibration pattern \nUnder 1mm precise corners" + ], + "image_footnote": [], + "bbox": [ + 334, + 210, + 421, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ac3d3865a19941ae2db2a0f602279d7d60b12c2786fcc8708b6ed2adeb029e26.jpg", + "image_caption": [ + "Top-tier robot arm Precise poses" + ], + "image_footnote": [], + "bbox": [ + 385, + 268, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a2f3f8f307dd9a819dcc73db2d4228a24f502f46f38d2d362578ebf0f1a3aa7.jpg", + "image_caption": [ + "(3) Data gathering", + "(i) Real data acquisition" + ], + "image_footnote": [], + "bbox": [ + 521, + 113, + 660, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a02b50d078fe3edafe71d0ab3638bc5c5bddd3542a3b9a2437eb34e09a5b26d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 172, + 617, + 198 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e3a0e084876fbf53d19b7af1fc8eed4d51a70b6a38e4a1c13de2bba21ec05bd.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 517, + 200, + 553, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a49ce4f839b822f2dc49b37e32f3383ca19abf576328ad915a9ceaae867d796a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 553, + 200, + 586, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/157ea8aba7041065075008065c61a237319e4a144353b0fe90c9b39e9e2b493a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 208, + 612, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/edd9316479c3794334faca86caa59d0c0844e9915c29d60895499be4c73f9a33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 239, + 557, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17ece7b18ebb80ba195c10661735b761b558fa0a34f53240737470146282244c.jpg", + "image_caption": [ + "View N", + "#" + ], + "image_footnote": [], + "bbox": [ + 558, + 239, + 614, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/785e121e9804db3f8fe7b2ad9cc7a685db19d0e4ec2a939b5a46758b963bfe07.jpg", + "image_caption": [ + "Integrated Mesh" + ], + "image_footnote": [], + "bbox": [ + 627, + 213, + 671, + 248 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7fcd01438e7915a9e95b8ff928d5bc8a751a14845db799f154d18b3e556e89f9.jpg", + "image_caption": [ + "Prototype registration", + "Figure 3. SiM3D dataset creation pipeline. (1) We purchased eight types of objects with different properties, each associated with a CAD model. We altered part of the samples to create realistic anomalies; (2) We designed and calibrated an acquisition setup consisting of a high-precision 3D sensor with high-resolution cameras on top an industrial robot arm; (3) We acquired 360-degree scans of each object instance to populate the real split of the training and test sets, and rendered images and point clouds from the CAD models to create the synthetic split of the training sets; (4) We created voxel-based segmentation GTs by a mixed 2D-3D labelling pipeline." + ], + "image_footnote": [], + "bbox": [ + 514, + 297, + 589, + 353 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ffcc203adb553e9a5da42024b7426e08604f984c3933c6532e21c127b3017161.jpg", + "image_caption": [ + "Realistic rendering" + ], + "image_footnote": [], + "bbox": [ + 596, + 297, + 683, + 323 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27c6758f9ce2ca253503440d10d7f341a116d9ab404029ce0ba3d0241a1af103.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 330, + 676, + 352 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a33eb270d948decff7ffefb30c1862c185a613a16e0a24d8d262f8017e05ea4.jpg", + "image_caption": [ + "(4) Labelling", + "Annotated Views" + ], + "image_footnote": [], + "bbox": [ + 712, + 114, + 784, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1995090b562a5f2e5c76f151b3759663eb18097678d533d91a63cebd8ba3563e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 114, + 883, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/647779311375c4cefad78cf33c92de0c96a9cd09be881fef60e40a1ddd159f59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 172, + 784, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31974e1192941ca1c61071e5a989a01b40a9803c71e3f86f5f4ef5e422933518.jpg", + "image_caption": [ + "↓" + ], + "image_footnote": [], + "bbox": [ + 785, + 172, + 851, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df2c12c470d93ed3141e164a1fdb26b8e6082bac41a12813d4c4af90b42f8d30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 191, + 769, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8bab64fbf22ed1ec5b96317f64fd6a8bebaf358b1326810148b2550ddac2f00f.jpg", + "image_caption": [ + "rojection" + ], + "image_footnote": [], + "bbox": [ + 709, + 203, + 784, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/13202d5d5ccfe76283addf07a3390fa08ffef800eecbb72c8b6f2bb1068c4ff0.jpg", + "image_caption": [ + "Integrated Mesh" + ], + "image_footnote": [], + "bbox": [ + 712, + 222, + 784, + 282 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03f8bf14dd4285b025192edb6e01b2393473513b8107aa48b250a003d31aad10.jpg", + "image_caption": [ + "Annotated Mesh" + ], + "image_footnote": [], + "bbox": [ + 785, + 224, + 880, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/75a3bd599272a1aa50be8b8b95cadd241ecf6f5e41f06e9098ca07e2b37a9f5b.jpg", + "image_caption": [ + "3D Segmentation GT", + "Manual Refinement & Voxelization" + ], + "image_footnote": [], + "bbox": [ + 712, + 295, + 784, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prototype model so that all nominal ones should be identical – and for which an official CAD prototype was available from the manufacturer. The prototype CAD models are depicted in Fig. 3 (1.ii). To simulate various anomalies, we manually introduced different types of defects that represent realistic manufacturing flaws, as illustrated in Fig. 3 (1.iii). We altered the objects' appearance (e.g., paint modifications, Fig. 3 (1.iii) – bottom left), geometry (e.g., dents, Fig. 3 (1.iii) – bottom right), or both (e.g., contaminations, Fig. 3 (1.iii) – top left). It is important to point out that only half of the instances within each object type are modified, while the remaining half are kept in their original state to serve as nominal data. For each object type, one nominal instance will be used as the training set in the real2real experimental setup, while all other ones, both nominal and modified (i.e., anomalous), will make up the test set. This very same test set is also used in the synth2real setup, though the training data comes from CAD prototypes.", + "bbox": [ + 91, + 473, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sensor. We acquired our dataset with a top-tier 3D scanner, the Atos Q from ZEISS (Fig. 3 (2.i)). It consists of a stereo pair with two $12\\mathrm{Mpx}$ greyscale cameras and a light projector. We consider the left camera to be our reference. The images have a resolution of $4096\\times 3000$ pixels. The point clouds can feature up to 12 million elements, with a point-to-point distance between 0.04 and $0.15\\mathrm{mm}$ . The sensor is mounted on an industrial robot arm with high repeatability and precision (Fig. 3 (2.iii)).", + "bbox": [ + 89, + 747, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Calibration. We calibrate the greyscale cameras with a", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "standard checkerboard pattern, estimating the intrinsic parameters matrix, $A$ , and the radial and tangential distortion coefficients, $\\delta = [k_{1}, k_{2}, k_{3}, p_{1}, p_{2}]$ . We also calibrate the rotation, $R_{pc}$ , and translation, $T_{pc}$ , from the point cloud to the left camera reference frame, with a precise dotted calibration pattern built by ZEISS (Fig. 3 (2.ii)) that provides centre coordinates with precision $< 1$ mm. More details are in the supplementary.", + "bbox": [ + 511, + 473, + 906, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real data acquisition. The acquisition process consists of 360-degree scanning operations around all object instances of our dataset (see Fig. 3 (3.i)). We define specific viewpoints to capture data from multiple angles around the object. These trajectories are sampled from concentric hemispheres that depend on the object's size. We provide poses relative to a reference view, i.e., $v_{1}$ . Thus, the pose for the view $i$ , $v_{i}$ , consists of a rotation $R_{i}$ and a translation $T_{i}$ that aligns $v_{i}$ to $v_{1}$ , with $R_{1}$ the identity matrix and $T_{1}$ null. Since in real industrial setups object instances would be acquired from as similar as possible viewpoints, we place markers in the scene to facilitate consistent positioning of objects of the same type. Thus, thanks to the high repeatability of the robot arm, images, $I_{i}$ , and point clouds, $P_{i}$ , from the same viewpoint $v_{i}$ turn out similar across acquisitions of different instances of the same object type. However, since the items were positioned manually with the help of the markers, slight displacements cannot be avoided. This introduces some variability across the scans of the instances of an object, as it also typically occurs in real indus", + "bbox": [ + 511, + 598, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20947", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trial pipelines. For each object, after collecting data from all viewpoints, the ATOS Q sensor also returns an integrated mesh, $M$ , obtained by registering all point clouds from the different views using a proprietary algorithm. We collect and release the mesh to foster the development of methodologies that may also exploit this 3D representation. The 3D reference frame is aligned to the one of the first point cloud $P_{1}$ taken from the first viewpoint $v_{1}$ .", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic data generation. We employ the Blender [11] Python API to render synthetic data that simulates real acquisitions. To ensure the synthetic and real data are aligned, we select a reference mesh from the collected real instances for each object type, load it in Blender, and align it to the prototype CAD model, as shown in (Fig. 3 (3.ii) - left). Then, we set the Blender camera parameters to the intrinsics, $A$ , of the left camera of our Atos Q sensor and render images from the same viewpoints $v_{i}$ employed during the real acquisitions. The RGB renders are converted into greyscale images to match the real images provided by our acquisition setup (Fig. 3 (3.ii) - right). Renderings have been obtained by exploiting Cycles Printerer, a path-tracing renderer able to provide realistic results with reflections, emissive surfaces, and physically-based lighting. We also render depth maps (Fig. 3 (3.ii) - right), which we project into 3D point clouds (see Fig. 1, centre) using the intrinsic parameters, $A$ .", + "bbox": [ + 91, + 215, + 486, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Labelling. Since some anomalies, such as paintings and minor scratches, are visible only in greyscale images, while others, such as dents, are better visible in 3D, we must consider both sensing modalities to obtain precise 3D segmentation ground-truths for anomalous samples. Therefore, we develop a two-step labelling strategy, illustrated in Fig. 3 (4), that operates on both 2D images and 3D data. First, we manually annotate all the view images, $I_{i}$ , where anomalies are clearly visible, obtaining dense 2D segmentation masks. We employ a different ID and colour for every defect to ease the voxel AUPRO computation (see Sec. 4). Subsequently, for each annotated image, $I_{i}$ , we leverage the associated integrated mesh, $M$ , the intrinsic parameters $A$ , the point cloud to camera reference frame transformation, $(R_{\\mathrm{pc}}, T_{\\mathrm{pc}})$ , and the transformation between $v_{i}$ and $v_{1}$ , $(R_{i}, T_{i})$ , to project the 2D annotations onto the integrated mesh. Then, since the projection may not be precise due to occlusions and slanted surfaces, each ground-truth mesh is thoroughly inspected and manually refined using a 3D visualisation and editing software, i.e., CloudCompare [12]. The labelling process was conducted by a team of 4 experts, with a final cross-check of the results. Finally, the annotated mesh is converted into a voxel grid with a voxel size of 2 mm. The choice of the size is determined by the smallest defect size discernible in the meshes. In the conversion process, a voxel is labelled as anomalous if it intersects at least one triangle labelled as such in the annotated mesh.", + "bbox": [ + 91, + 493, + 486, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/298d2f195d133bd06d679291d4951fec2abbfac0bc724df73e05fd2274192a9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeDimensions [cm]Dimensions [voxels]Total InstancesTrain Instances NominalTest Instances NominalAnomalousViews per Instance
Plastic Stool\\( {35} \\times {35} \\times {30} \\)\\( {330} \\times {326} \\times {317} \\)221 real or 1 synth101012
Rubbish Bin\\( {26} \\times {21} \\times {33} \\)\\( {333} \\times {339} \\times {326} \\)421 real or 1 synth202012
Wicker Vase\\( {17} \\times {17} \\times {15} \\)\\( {329} \\times {309} \\times {290} \\)221 real or 1 synth101012
Bathroom Furniture\\( {33} \\times {33} \\times {50} \\)\\( {371} \\times {392} \\times {363} \\)201 real or 1 synth81036
Container\\( {20} \\times {25} \\times {10} \\)\\( {327} \\times {308} \\times {286} \\)941 real or 1 synth464612
Plastic Vase\\( {12} \\times {12} \\times 9 \\)\\( {326} \\times {308} \\times {286} \\)991 real or 1 synth484912
Wooden Stool\\( {48} \\times {42} \\times {45} \\)\\( {331} \\times {367} \\times {327} \\)151 real or 1 synth6712
Sink Cabinet\\( {44} \\times {25} \\times {50} \\)\\( {342} \\times {349} \\times {357} \\)191 real or 1 synth9836
", + "bbox": [ + 516, + 88, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Dataset statistics.", + "bbox": [ + 627, + 172, + 790, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. SiM3D Benchmark", + "text_level": 1, + "bbox": [ + 511, + 215, + 699, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset, Tasks, and Metrics", + "text_level": 1, + "bbox": [ + 511, + 241, + 761, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Setups. Using the procedure described in Sec. 3, we collected 333 object instances belonging to the eight object types, as shown in Tab. 2. We captured from 12 to 36 views for each instance, depending on the type. We split the data into training and test sets. For each type, the training set comprises a single instance, either real in the real2real setup or synthetic in the synth2real setup. The test set includes all other instances and is balanced between nominal and anomalous samples. The test set is the same for both setups.", + "bbox": [ + 511, + 265, + 906, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks. Our benchmark focuses on the tasks of detecting and segmenting objects' anomalies using multiview and multimodal data. Specifically, the inputs to the methods consist of a set of images $\\{I_i\\}_{i=1}^n$ captured from different viewpoints alongside 3D data, $\\mathbf{X}$ , that capture an object entirely. The 3D data that a method can use are a set of point clouds $\\{P_i\\}_{i=1}^n$ or depth maps $\\{D_i\\}_{i=1}^n$ taken from the same vantage points as the images, an integrated mesh $M$ , or also a merged point cloud $P$ . A method capable of multiview multimodal 3D anomaly detection should take these inputs, integrate the 2D and 3D information from multiple views, and produce a global anomaly score for the test sample under inspection. On the other hand, a method pursuing multiview multimodal 3D anomaly segmentation should yield a 3D output, where each 3D position carries a score representing the likelihood of belonging to an anomaly. In our benchmark, a method for this task takes as input all images and 3D data, $(\\{I_i\\}_{i=1}^n, \\mathbf{X})$ to output an Anomaly Volume, i.e., a voxel grid, $V \\in \\mathbb{R}^{X \\times Y \\times Z}$ , where $X, Y$ , and $Z$ are the grid dimensions and each voxel carries an anomaly score. For several reasons, we chose voxel grids as the 3D output representation of our benchmark. Firstly, as their grid structure resembles 2D Anomaly Maps, existing 2D segmentation methods and metrics can be easily extended. Moreover, voxel grids are suitable for representing every type of 3D defect, including missing parts. One major challenge with voxel grids pertains to choosing the voxel size. Ideally, we would want a size that allows accurate segmentation of even the smallest defects. However, the memory occupancy of voxel grids scales cubically with the inverse of the voxel size. Thus, we create ground-truth volumes using a voxel resolution of $2\\mathrm{mm}$ , which enables precise localisation of", + "bbox": [ + 511, + 417, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20948", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e924cd3aa3e0550b385feea033f18ba92ed8d53f140ce5a8bde5fb8367a108fb.jpg", + "image_caption": [ + "Figure 4. Extending singleview methods. Anomaly Maps obtained by processing singleview inputs with standard ADS methods are projected into the 3D and aggregated into a voxel grid, with each voxel containing a list of scores from multiple views. The Anomaly Volume is obtained by voxel-wise max pooling." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "even the smallest defects present in our dataset while keeping memory occupancy affordable.", + "bbox": [ + 88, + 351, + 482, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Akin to standard practice in 2D ADS [4], we evaluate detection performance by the Area Under the Receiver Operator Curve (AUROC) of the global anomaly scores predicted by the methods. As we compute a global score for each object instance, we name it instance-level AUROC (I-AUROC). Regarding segmentation performance, we evaluate methods with an adaptation of the popular Area Under the Per-Region Overlap (AUPRO) curve. First of all, we define the voxel-based Per-Region Overlap (PRO) metric as $\\mathrm{PRO}_t = \\frac{1}{N}\\sum_{n = 1}^N\\frac{|\\overline{\\Psi}_t\\cap V_n|}{|V_n|}$ , where $\\overline{\\Psi}_t$ is an Anomaly Volume binarised using a threshold, $t$ , $N$ is the number of blobs, i.e., defects, in the ground-truth, and $V_{n}$ are the voxels belonging to that blob. Since during labelling, we annotated each defect in our dataset with a different ID (see Sec. 3), the $N$ defect blobs in a ground-truth are easily identifiable. This metric is computed across multiple thresholds, $t$ , estimated at different False Positive Rates (FPRs). We calculate the FPR using only non-empty voxels. We sample FPRs from 0 to the maximum value uniformly, we estimate the corresponding threshold $t$ , and we binarise the score grid accordingly. In this way, we construct a curve that plots the PRO values against the corresponding FPRs. The segmentation performance is obtained by integrating such a curve up to a specific False Positive Rate. As proposed in [14], we use $1\\%$ as integration bound for the PRO curve to set a challenging bar representative of the requirements of real industrial applications. Finally, we normalise the resulting area to the interval [0, 1]. We dub this metric voxel-level AUPRO (V-AUPRO).", + "bbox": [ + 91, + 382, + 483, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 89, + 833, + 199, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Extending singleview methods to Multiview 3D ADS. Current ADS methods predict a 2D anomaly map from a singleview input, either unimodal, e.g., images, or multi-", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modal, e.g., images and point clouds. To obtain a set of baselines for our novel multiview 3D ADS task, we extend popular and effective singleview methods by means of the 3D aggregation strategy illustrated in Fig. 4. Given the considered method, we process the data corresponding to an input view, $v_{i}$ , so as to obtain a 2D Anomaly Map. Similarly to Sec. 3, given the mesh, $M$ , intrinsic parameters $A$ , point cloud to camera reference frame transformation, $(R_{\\mathrm{pc}}, T_{\\mathrm{pc}})$ and view pose $v_{0}$ , $(R_{i}, T_{i})$ , we can project each pixel of the 2D Anomaly Map to its corresponding 3D position. We discretise the 3D position, finding the corresponding voxel index. We associate that voxel with the corresponding anomaly score. We repeat the process for each view, keeping track of the scores that project onto the same voxels. Eventually, we compute the maximum among all scores projected into each voxel, obtaining the final Anomaly Volume. To extract the global score required by the I-AUROC, we compute the maximum across all the scores within the Anomaly Volume.", + "bbox": [ + 511, + 90, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Selected Methods. We extend some of the most popular and effective unimodal and multimodal singleview ADS methods, starting from their official code, and evaluate them on SiM3D. Regarding unimodal methods, we select PatchCore [30], with either WideResNet101 [42] or DINOv2 [24] as backbone, to extract features from images of all views, $\\{I_i\\}_{i=1}^N$ of the single-instance training set to create memory banks. Moreover, we use EfficientAD [1] using all views as their training set. As for multimodal ADS methods, we use BTF [18] employing WideResNet101 [42] as the image encoder and FPFH [33] as the algorithm to extract point cloud features. We create memory banks using features extracted from images, $\\{I_i\\}_{i=1}^N$ , and point clouds, $\\{P_i\\}_{i=1}^N$ , of all views of the single-instance training set. We use the same strategy with M3DM [36]. As 2D feature extractor, we exploit DINO-v2. As point cloud feature extractor, instead of using a transformer backbone, i.e., PointMAE [17], we utilise FPFH to extract 3D features, similarly to BTF. Indeed, we would need a severe downsample (from $\\sim 5 - 7\\mathrm{M}$ to $\\sim 8\\mathrm{k}$ points) to work with Point-MAE, and we verified that at such low resolution, most of the defects no longer show up in the point clouds. Finally, we select CFM [14], using all views and clouds as a single training set. Based on the same consideration as for M3DM, we use DINO-v2 and FPFH to produce 2D and 3D features. Due to computational constraints, all methods process images padded and downsampled to $1540 \\times 1540$ . Moreover, for compatibility with the employed 2D feature extractors, which accept RGB images, we create 3-channel inputs by stacking our greyscale images thrice.", + "bbox": [ + 511, + 382, + 908, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Data pre-processing. To obtain less noisy data and improve the results of the above-mentioned baselines, we remove the background from 3D data. We estimate a 3D plane using a RANSAC-based plane segmentation algo", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20949", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7534f9153be095bf7fd2844e007c4f09313d5a09d32f83ba7fcf455b80669f4a.jpg", + "image_caption": [ + "Figure 5. Qualitatives. Selected views with defects (left) and Anomaly Volumes for several methods (right) downsampled for visualisation." + ], + "image_footnote": [], + "bbox": [ + 99, + 97, + 890, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b84e61275dbf91d8a6cf57acf80d31dabaeeb6f8f480cb60648dc2b9bb8253c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityreal2realsynth2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.MeanPl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
PatchCore w/ WRN-101RGB0.7400.9870.6360.7770.7740.5511.0000.5660.7540.4620.2350.5370.3530.6780.5760.5170.2500.451
PatchCore w/DINO-v2RGB0.5000.9580.6360.6220.5780.5631.0000.5630.6780.9580.8970.4130.4640.2070.6840.6070.0870.540
EfficientADRGB0.2800.7320.0000.8780.4240.7300.9280.7120.5850.1230.6730.0000.8480.0080.4870.7500.6620.443
ASTRGB0.5370.5660.6110.5150.5730.4960.8390.5370.5841.0000.3441.0000.5250.4820.3880.4280.1870.544
BTFRGB + Point Cloud0.4210.2170.5040.5650.5450.4710.6780.4240.4780.4620.2940.5200.4440.4640.4240.4820.2870.422
CFM w/DINO-v2 + FPFHRGB + Point Cloud0.1980.3010.0740.5150.4830.4560.7320.8250.4480.0080.0000.1900.4240.3130.4590.4280.3620.273
M3DM w/DINO-v2 + FPFHRGB + Point Cloud0.7020.9880.6610.5450.5560.6490.3920.4750.6210.1070.1170.7350.5650.3810.5860.2140.5120.402
ASTRGB + Depth0.9500.9270.7850.4740.5420.4700.4280.9250.6870.6360.0020.5040.4740.4620.5690.4280.8870.495
", + "bbox": [ + 96, + 387, + 901, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Anomaly detection results (I-AUROC). Best results in bold, runner-ups underlined.", + "bbox": [ + 217, + 488, + 774, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rithm, which, iteratively, randomly selects a small subset of points (e.g., 10) to fit a candidate plane model. For the candidate planes, we compute the consensus set with threshold $\\tau$ on the point-to-plane distance. After 1000 iterations, we return the best-fitting plane. To account for spurious background artefacts, we shift the plane by an offset $\\alpha$ and finally remove all points below the shifted plane. Given the different sizes of the objects, for each object type, we select different optimal thresholds $\\tau$ and offsets $\\alpha$ , as detailed in the supplementary material. We apply this background removal technique to the integrated mesh, $M$ , due to the larger number of background points, which facilitates the RANSAC algorithm. Then, from the background-filtered integrated mesh, we leverage ray-tracing techniques and known camera parameters to render, for each view, depth maps, $\\{D_i\\}_{i=1}^N$ , pixel-aligned to the corresponding images, $\\{I_i\\}_{i=1}^N$ at the sensor resolution $(4096 \\times 3000)$ . Finally, these depth maps are padded and downsampled to $1540 \\times 1540$ , and the depth is lifted in 3D via the camera parameters to obtain dense and clean point clouds. The depth maps and point clouds obtained by rendering the preprocessed meshes serve as the input data for all the experiments presented in the subsequent section and will be available in the SiM3D dataset.", + "bbox": [ + 88, + 529, + 485, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Benchmark Results", + "text_level": 1, + "bbox": [ + 511, + 527, + 702, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Main Results. In Tab. 3 and Tab. 4, we report the results obtained on the SiM3D benchmark by extending the considered methods as Multiview 3D ADS baselines. By looking at the mean detection performance (Tab. 3), we notice how methods that natively process only RGB images outperform natively multimodal ones in both the real2real and synth2real setups. With the notable exception of AST, this also stands for mean segmentation performance (Tab. 4). We posit that this is due to the main multimodal methods having been designed and tuned to process point clouds of the resolution featured by existing multimodal benchmarks, which is much lower than that of SiM3D, i.e. $\\sim 8\\mathrm{K}$ vs $\\sim 2.3\\mathrm{M}$ points (after downsampling). Indeed, AST is the only one designed to work natively with depth maps. Hence, our experiments suggest that current multimodal ADS methods do not scale to significantly higher resolutions. In turn, we could not even deploy the native 3D feature extractors of M3DM and CFM, the methods currently excelling on [6, 7]. Furthermore, we notice how methods relying on memory banks, such as PatchCore and M3DM, tend to attain slightly better mean detection and segmentation performance than their counterparts", + "bbox": [ + 509, + 568, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20950", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/446c24ed812763c89421346d05157bf3eac35a72457f50ac6a2ac9314621475e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModality
PatchCore w/ WRN-101RGB
PatchCore w/ DINO-v2RGB
EfficientADRGB
ASTRGB
BTFRGB + Point Cloud
CFM w/ DINO-v2 + FPFHRGB + Point Cloud
M3DM w/ DINO-v2 + FPFHRGB + Point Cloud
ASTRGB + Depth
", + "bbox": [ + 96, + 89, + 269, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66ab6a75f887ca4f2f6096465ddf3f78e9606c0961d4280b818bf799b8266692.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
real2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
0.7100.4610.7610.6750.6940.7470.3800.6090.630
0.7450.4690.7750.7920.7090.7530.4350.6900.671
0.6820.4620.7630.5340.6800.7430.4070.4880.594
0.7200.5020.7890.8030.7160.7640.4460.7580.687
0.5510.4020.7500.3770.6140.7410.0920.0300.444
0.5970.4150.7680.5050.6400.7430.3150.3210.538
0.7330.4520.7670.7020.7020.7520.2880.1260.565
0.7500.5030.7920.8070.7160.7640.4670.7980.699
", + "bbox": [ + 272, + 90, + 584, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c4b37ea7a5d3af939ad3519c6a8402c13c2770ce1b4e9ed517dd34a5797f7d39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
pl.Stoolrub.BinW.VaseB.Furn.Cont.Pl.VaseW.StoolSink Cab.Mean
0.7340.4370.7510.4540.6780.7410.3860.6180.600
0.7010.4570.7720.5630.6480.7340.3210.5750.596
0.6800.4490.7290.5230.6690.7470.4310.3610.573
0.7260.4880.7910.8060.7050.7640.4610.6950.679
0.5470.4020.7560.3690.6100.7410.1030.0400.446
0.5230.4130.7530.5420.6210.7410.2220.3140.516
0.6900.4540.7700.4610.6450.7340.3180.1320.526
0.7510.5020.7930.8050.7170.7640.4500.7850.695
", + "bbox": [ + 586, + 90, + 900, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa82f769d9445f8b5ec4d0a6318ba3cba498766183684382f0f6b53be8c60d04.jpg", + "table_caption": [ + "Table 4. Anomaly segmentation results (V-AUPRO@1%). Best results in bold, runner-ups underlined." + ], + "table_footnote": [], + "table_body": "
MethodModalityFeatures
2D3D
PatchCoreRGBWRN-101-
RGBDINO-v2-
Depth-WRN-101
Depth-DINO-v2
Point Cloud-FPFH
BTFRGB + Point CloudWRN-101FPFH
RGB + DepthWRN-101WRN-101
CFMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
M3DMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
ASTRGBEffNet-B5-
RGB + DepthEffNet-B5EffNet-B5
", + "bbox": [ + 96, + 229, + 292, + 361 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f1564c5a2133011f8b5473cf28ebd96bdddeb217748d3d3f47437808b3a71e7e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
real2realsynth2real
DetectionSegmentationDetectionSegmentation
0.7540.6300.4510.600
0.6780.6710.5400.596
0.5820.4860.5070.484
0.6000.4960.2460.471
0.4150.4640.3130.460
0.4780.4440.4220.446
0.7070.6090.4480.543
0.4480.5380.2730.516
0.5480.6630.3110.620
0.6210.5650.4020.526
0.6590.5030.2540.476
0.5840.6870.5440.679
0.6870.6990.4950.695
", + "bbox": [ + 295, + 229, + 478, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Anomaly detection and segmentation mean results with different input modalities. Best results per method in bold.", + "bbox": [ + 89, + 373, + 482, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "within the same modality that require substantial training, i.e., EfficientAD and CFM. We argue that this may be ascribed to the single-instance setup, which renders the latter category of approaches harder to optimise due to the limited number of training images, which are as many as the views. The methods most effectively extended to our multiview 3D task are Patchcore, which, equipped with either WRN-101 or DINO-v2, can achieve the best detection in both setups, and AST, achieving the best segmentation performance in both setups. Indeed, Patchcore yields I-AUROCs of 0.754 (WRN-101) and 0.540 (DINO-v2), while AST yields a V-AUPRO@1% of 0.699 and 0.695, in the real2real and synth2real setups, respectively. It is also worth pointing out how, due to the domain shift, we can notice a gap between the real2real and synth2real setups, especially in detection (0.754 vs. 0.540, I-AUROC) compared to segmentation performance (0.699 vs. 0.695, V-AUPRO@1%). The qualitative results, shown in Fig. 5, visually support the above observations. Overall, the relatively low, definitely far from saturated, performance figures provided by baselines designed by straightforward extension of state-of-the-art methods delivering impressive results in singleview 2D benchmarks hint at the novel and challenging nature of the task set forth by SiM3D. Thus, our experimental findings by baseline approaches reveal the need for more advanced and specific methods and paradigms, both in the real2real and, even more so, the synth2real setups.", + "bbox": [ + 89, + 431, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative 3D Representations. Given the lack of feature extractors amenable to processing high-resolution point clouds, we explore the use of depth maps to capture 3D", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cues. The 2D grid structure of depth maps enables the deployment of popular backbones pre-trained on RGB data that support high-resolution inputs, such as DINOv2. Thus, we create and evaluate additional baselines using depth maps and report the results in Tab. 5. Similarly to greyscale images, we tile depth maps along the channel axis to ensure compatibility with the employed backbones. The results show that the use of depth maps rather than point clouds in multimodal methods tends to improve detection and segmentation performance by sizable margins, e.g., in the real2real setup, BTF yields 0.707 I-AUROC and 0.609 V-AUPRO when processing depths compared to 0.478 I-AUROC and 0.444 V-AUPRO when fed with point clouds (row 6 vs. row 7) while relying on the same WRN101 image backbone. Indeed, all multimodal methods but M3DM achieve their best performance in both tasks and setups by processing depth maps instead of point clouds. Yet, as vouched by the results achieved by PatchCore, RGB images compare favourably to depth maps in terms of detection performance. We ascribe this to the feature extractors used to compute depth features being pre-trained on RGB images, which may be conducive to yielding weaker features when fed with depth inputs. This points out an intriguing research question: would it be possible to develop a foundation model to process depth maps effectively?", + "bbox": [ + 511, + 231, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Final Discussion", + "text_level": 1, + "bbox": [ + 511, + 621, + 674, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced SiM3D, the first benchmark and dataset focused on integrating multiview and multimodal information for comprehensive 3D ADS. Our experiments highlight several open challenges. First, a need for methods that can process multiview inputs more effectively, e.g., by taking all views as input rather than integrating singleview outputs, and that can be trained with a very limited number of nominal samples, e.g., a single instance. Second, training with synthetic data requires more robust techniques to address domain-shift challenges. Finally, the current literature lacks 3D backbones capable of processing high-resolution point clouds. Employing depth maps seems a promising strategy to facilitate this processing, yet a foundational model for depth maps is still missing. We believe that SiM3D may serve as a strong foundation for tackling these challenges and fostering future research toward 3D ADS.", + "bbox": [ + 511, + 647, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20951", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We acknowledge financial support from the National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.1, Call No. 1409 published on 14 September 2022 by the Italian Ministry of University and Research (MUR), funded by the European Union.", + "bbox": [ + 89, + 114, + 480, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was also partially supported by the FSE+ 2021-2027 program under Article 24, paragraph 3, letter a) of Law 240/2010 and its subsequent amendments, as well as by Regional Resolution No. 693/2023 (Project Code: 2023-20090/ER), Grant CUP: J19J23000730002.", + "bbox": [ + 89, + 190, + 480, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank Lucia Gasperini for her contribution to dataset acquisition and labelling. We also extend our gratitude to Manuel Turrini for his support in setting up the acquisition process.", + "bbox": [ + 89, + 266, + 480, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 340, + 186, + 354 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficient-tad: Accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 128-138, 2024. 3, 6", + "[2] Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection. arXiv preprint arXiv:2002.10445, 2020.3", + "[3] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3", + "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6", + "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130, 2022. 3", + "[6] Paul Bergmann, Jin Xin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 202-213, 2022. 1, 2, 3, 7", + "[7] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the 16th Asian Conference on Computer Vision (ACCV2022), 2022. ACCV. 1, 2, 3, 7", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + ], + "bbox": [ + 99, + 364, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Li-Ling Chiu and Shang-Hong Lai. Self-supervised normalizing flows for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2926-2935, 2023. 3", + "[10] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. ArXiv, 2020. 3", + "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2024. 5", + "[12] CloudCompare Online Community. CloudCompare - 3D point cloud and mesh processing software Open Source Project. EDF, 2024. 5", + "[13] Alex Costanzino, Pierluigi Zama Ramirez, Mirko Del Moro, Agostino Aiezzo, Giuseppe Lisanti, Samuele Salti, and Luigi Di Stefano. Test time training for industrial anomaly segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3910-3920, 2024. 3", + "[14] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. CVPR. 1, 3, 6", + "[15] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 3", + "[16] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 98-107, 2022. 3", + "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 6", + "[18] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2976, 2023. 3, 6", + "[19] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 3", + "[20] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 3", + "[21] Jiaqi Liu, Guoyang Xie, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, Feng Zheng, et al. Real3d-ad: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20952", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv preprint arXiv:2301.11514, 2, 2023. 3", + "[23] Fabio Valerio Massoli, Fabrizio Falchi, Alperen Kantarci, Şeymanur Akti, Hazim Kemal Ekenel, and Giuseppe Amato. Cocca: Multilayer one-class classification for anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 33(6):2313-2323, 2021. 3", + "[24] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6", + "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 604-621. Springer, 2022. 3", + "[26] Jonathan Pinnay and Keng Chai. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing, pages 394-406. Springer, 2022. 3", + "[27] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 3", + "[28] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733. IEEE, 2021.", + "[29] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13576-13586, 2022. 3", + "[30] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of 2022 IEEE Conference on Computer Vision and Pattern Recognition, pages 14298-14308, 2022. 1, 3, 6", + "[31] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907-1916, 2021. 3", + "[32] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE International Conference on Robotics and Automation, pages 3212-3217, 2009. 3, 6", + "[34] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. In International Conference on Learning Representations, 2021. 3", + "[35] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22883-22892, 2024. 1, 3", + "[36] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 1, 3, 6", + "[37] Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650-656, 2022. 3", + "[38] Minghui Yang, Peng Wu, and Hui Feng. Memseg: A semi-supervised method for image surface defect detection using differences and commonalities. Engineering Applications of Artificial Intelligence, 119:105835, 2023.", + "[39] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision, 2020.", + "[40] Seungdong Yoa, Seungjun Lee, Chiyoon Kim, and Hyunwoo J Kim. Self-supervised learning for anomaly detection with dynamic local augmentation. IEEE Access, 9:147201-147211, 2021.", + "[41] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 3", + "[42] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6", + "[43] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3", + "[44] Zheng Zhang and Xiaogang Deng. Anomaly detection using improved deep svdd model with data structure preservation. Pattern Recognition Letters, 148:1-6, 2021. 3", + "[45] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection. arXiv preprint arXiv:2310.07716, 2023. 1, 3", + "[46] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20953", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_model.json b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..02ceebc1b3e9408223b0f9735b067301067d3e48 --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.819, + 0.177 + ], + "angle": 0, + "content": "SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.185, + 0.91, + 0.258 + ], + "angle": 0, + "content": "Alex Costanzino1 Pierluigi Zama Ramirez1 Luigi Lella1 Matteo Ragaglia2 Alessandro Oliva2 Giuseppe Lisanti1 Luigi Di Stefano1 \n1CVLab, University of Bologna, Italy 2SACMI Imola, Italy https://alex-costanzino.github.io/SiM3D/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.486, + 0.609 + ], + "angle": 0, + "content": "We propose SiM3D, the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS), where the task is to produce a voxel-based Anomaly Volume. Moreover, SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. In this respect, SiM3D stands out as the first ADS benchmark that addresses the challenge of generalising from synthetic training data to real test data. SiM3D includes a novel multimodal multiview dataset acquired using top-tier industrial sensors and robots. The dataset features multiview high-resolution images (12 Mpx) and point clouds (\\(\\sim 7M\\) points) for 333 instances of eight types of objects, alongside a CAD model for each type. We also provide manually annotated 3D segmentation GTs for anomalous test samples. To establish reference baselines for the proposed multiview 3D ADS task, we adapt prominent singleview methods and assess their performance using novel metrics that operate on Anomaly Volumes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.224, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In computer vision, anomaly detection and segmentation (ADS) involves identifying anomalous samples and localising their defects. This task is challenging in industrial contexts due to the variability and rarity of defects. Typically, ADS in the industry follows a cold-start approach, where training is unsupervised and uses only defect-free (nominal) samples. Before the release of the MVTec AD dataset in 2019 [4], only a limited number of ADS methods were proposed, with just 9 publications in major computer vision conferences (e.g., CVPR, ECCV, ICCV, WACV) over the five years before. The introduction of MVTec AD marked a pivotal moment in the field, fostering the publication of roughly 172 papers in these venues over the next 5 years. This dataset not only stimulated the development of several ADS methods [14, 30, 36] but also promoted the creation of novel and more challenging benchmarks. For instance," + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.272, + 0.862, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.703, + 0.908, + 0.745 + ], + "angle": 0, + "content": "Figure 1. SiM3D dataset overview. From top to bottom: the single-instance real and synthetic training samples for object Plastic Stool, one of the anomalous samples from the test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.868 + ], + "angle": 0, + "content": "MVTec 3D-AD and [6] and Eyecandies [7] introduced the first benchmarks deploying multimodal inputs to enhance ADS performance. Then, PAD [45] and Real-IAD [35] introduced novel benchmarks that leverage multiple RGB views of objects. Although these benchmarks tackle various settings, two real industrial issues remain unaddressed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "First, existing benchmarks have primarily addressed the problem of 2D ADS. Consequently, current methods gen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20944" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.182 + ], + "angle": 0, + "content": "
DatasetVenue & Year2DNo. Pixels3DNo. PointsScenarioTrainTestSetupTaskNo. Classes
MVTec ADCVPR 2019RGB Images490k - 1M——Multi-instanceSingleviewSingleviewreal2real2D15
MVTec 3D-ADVISAPP 2021RGB Images160k - 810kXYZ Images10k - 195kMulti-instanceSingleviewSingleviewreal2real2D10
EyecandiesACCV 2022RGB Images262kDepths + Normals262kMulti-instanceSingleviewSingleviewsynth2synth2D10
MVTec LOCO ADIJCV 2022RGB Images1.28M - 2.05M——Multi-instanceSingleviewSingleviewreal2real2D5
VisAECCV 2022RGB Images1.5M——Multi-instanceSingleviewSingleviewreal2real2D12
PADNeurIPS 2023RGB Images12M——Single-instanceMultiviewSingleviewreal2real + synth2synth2D20
Real3D-ADNeurIPS 2023——Point Clouds43k - 2.68MMulti-instanceMultiviewSingleviewreal2real3D12
Real-IADCVPR 2024RGB Images10M——Multi-instanceSingleviewSingleviewreal2real2D30
SiM3D2025Grayscale Images12MPoint Clouds + Mesh5M - 7MSingle-instanceMultiviewMultiviewreal2real + synth2real3D8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.191, + 0.9, + 0.205 + ], + "angle": 0, + "content": "Table 1. Popular ADS benchmarks. Chronologically sorted, the sequence hints at a growing interest in multimodal and multiview data." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.23, + 0.482, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.483, + 0.388 + ], + "angle": 0, + "content": "Figure 2. 2D vs. 3D anomaly detection and segmentation. Typical ADS benchmarks rely on 2D Anomaly Maps and 2D ground-truths (left) to assess segmentation performance. SiM3D proposes to do so by 3D Anomaly Volumes and 3D ground-truths (right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.734 + ], + "angle": 0, + "content": "erate Anomaly Maps, i.e., images that encode the likelihood of each pixel belonging to a defect, as shown in Fig. 2 (left). However, many industrial applications require precise localisation of defects in the 3D space to allow automatic intervention, reducing waste and optimising production. Benchmarks relying on multimodal input data, like those proposed in MVTec 3D-AD [6] and Eyecandies [7], represent progress toward this goal by providing 3D information for each pixel within an image. However, in these benchmarks, ADS methods are still expected to provide as output a 2D Anomaly Map, and objects are inspected from a single viewpoint. For comprehensive 3D anomaly detection, we must observe an object in its entirety, utilising captures from multiple viewpoints to detect defects across the entire 3D space. This calls for methods capable of integrating multiview information to estimate the defect probability at each 3D location within an Anomaly Volume, as shown in Fig. 2 (right). Furthermore, as demonstrated by MVTec 3D-AD and Eyecandies, these methods should process multimodal information, i.e., RGB images as well as 3D information, to achieve more reliable anomaly detection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Secondly, current benchmarks typically provide multiple training samples per object class to capture the variability across nominal objects. However, collecting many samples - even nominal ones - can be costly and challenging. For example, line changeovers would demand acquiring new datasets and retraining models, both highly time-intensive processes. Additionally, collecting large datasets increases the risk of errors, such as misjudged nominal samples, which may degrade model performance. Moreover, manufactured objects typically closely resemble one another as they replicate the structure and materials of a syn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.293 + ], + "angle": 0, + "content": "thetic CAD prototype. Thus, the variability across nominal samples of manufactured objects is frequently minimal, with a single nominal item subsuming the information required to spot anomalies in other objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.906, + 0.461 + ], + "angle": 0, + "content": "These considerations highlight the need for ADS techniques that may be effectively trained by a single nominal sample or even a synthetic prototype (e.g., an object's CAD model) and then be able to generalise to other manufactured objects to identify anomalous ones effectively. The ability to perform training by a single object instance, either real or synthetic, would avoid the need for extensive data collection, vastly facilitating the creation and adaptation of ADS models. Yet, most current benchmarks focus on multiple training instances setups, and none of the prior works consider challenging synthetic (training) to real (test) setups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.78 + ], + "angle": 0, + "content": "To fill all the gaps highlighted by our considerations and stimulate interest in the development of multimodal and multiview 3D ADS methods, we propose the Single-instance Multiview Multimodal Multisetup dataset, dubbed SiM3D. The dataset contains a series of high-resolution scans of manufactured objects, which can be nominal or anomalous. Each scan consists of multiple views, each featuring a greyscale image, its corresponding point cloud (i.e., multimodal data), and the ground-truth acquisition pose. In addition, an integrated 3D mesh is provided with each scan. For each object, a single nominal sample has been retained to be employed as a single-instance training set, while the remaining samples, both nominal and anomalous, compose the test set. Moreover, to foster research on novel approaches that may address the domain gap between synthetic training data and real test samples in ADS, for each object of our dataset, we also provide a synthetic scan, i.e., we deploy the available CAD model to obtain the same kind of data as in our high-resolution real scans, with images and point clouds rendered from the same poses. An overview of the proposed dataset is depicted in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on our novel dataset, we create the first benchmark for 3D anomaly detection, where the task is to predict anomaly scores within a voxelized 3D Anomaly Volume rather than a 2D Anomaly Map (Fig. 2). Peculiarly, our benchmark includes two setups based on the domain of the data available for training: in the real2real setup, a model is given access to a single nominal instance of an object (Fig. 1, top), while in synth2real the same kind" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "of training data are obtained from the CAD model of the object (Fig. 1, centre). We adapt prominent ADS methods – both unimodal and multimodal – for the proposed task and evaluate them on our benchmark. To assess segmentation performance, we provide accurate 3D ground-truths (Fig. 2) and extend ADS metrics to operate on anomaly scores and ground-truths specified on 3D voxel grids. Our experimental findings highlight the open challenges in the field and offer insights into possible future directions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.241, + 0.233, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.483, + 0.78 + ], + "angle": 0, + "content": "The MVTec AD [4] dataset significantly boosted anomaly detection research, inspiring new techniques and benchmarks for various challenges. MVTec AD provides a training set containing only nominal images and a test set that includes nominal and anomalous ones for several object categories. This unsupervised setup exemplifies many real-world applications in which anomalous objects are challenging to collect, and thus, it has also been adopted by more recent ADS benchmarks. Among them, MVTec LOCO [5] presents a challenging scenario with both structural anomalies (e.g., dents, holes) and logical anomalies (e.g., incorrect object combination). VisA [46] introduced high-resolution images of complex scenes containing multiple instances of the same object. However, certain defects may be visible only by analysing objects' 3D geometry. For this reason, multimodal datasets, such as MVTec 3D-AD [6] and Eyecandies [7] include 3D data alongside images. For each sample in the dataset, MVTec 3D-AD provides images and pixel-aligned XYZ information captured by a 3D sensor, while Eyecandies contains synthetic images with pixel-aligned depths and normals. Recently, PAD [45] addressed pose-agnostic ADS with the first multiview dataset (MAD), featuring images of objects from various viewpoints. In PAD, training samples are multiview RGB images with known poses, while test samples are singleview RGB images with unknown poses. MAD includes both synthetic and real objects, but its benchmarks do not explore synthetic-to-real generalisation. Real-IAD [35] introduced a large-scale multiview dataset with RGB images covering various defect sizes. Eventually, Real3D-AD [21] is the first point cloud anomaly detection benchmark where the task is to predict an anomaly score for each point. It uses front-and-back scans of nominal objects for training and a singleview point cloud for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1, the proposed SiM3D benchmark offers several outstanding features, including very-high-resolution images (12 Mpx), large point clouds/meshes containing 5-7 million elements captured with high-precision 3D sensors. Moreover, it is designed for a highly relevant manufacturing scenario: single-instance anomaly detection, where only one object is available for model training. To promote research toward even more cost-effective scenar" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "ios, ours is also the only benchmark that addresses the challenge of generalising from synthetic training data to real test data. Finally, SiM3D stands out as the first benchmark that considers integrating multiview information at test time for comprehensive 3D anomaly detection, where the task consists of outputting a voxel-based Anomaly Volume." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.905, + 0.771 + ], + "angle": 0, + "content": "ADS Methods. Following the setup defined in the MVTec-AD benchmark, most ADS methods [22] are designed to estimate 2D anomaly maps from a single image [2, 3, 9, 10, 13, 15, 16, 19, 20, 23, 26, 27, 27-29, 31, 34, 37-41, 43, 44]. One of the most popular methods is PatchCore [30], which uses feature extractors pre-trained on large datasets [8, 17, 24] to gather features from nominal samples into a memory bank. At test time, input sample features are compared to this bank to detect anomalies. Another more recent yet popular approach, EfficientAD [1], uses a lightweight feature extractor and a student-teacher model to identify anomalies based on feature prediction discrepancy. The introduction of the MVTec 3D-AD benchmark has led to the development of several multimodal approaches that use images and 3D data to generate 2D anomaly maps [14, 18, 32, 36]. Among them, BTF [18] is a memory bank approach inspired by PatchCore, in which each element is the concatenation of 2D and 3D features extracted by a pre-trained image backbone and a traditional point cloud descriptor, i.e., FPFH [33], respectively. M3DM [36] improves BTF using powerful transformer-based pre-trained backbones, such as DINO-v1[8] and Point-MAE [25], to extract features from RGB images and point clouds. It also learns to fuse 2D and 3D features into multimodal features. During inference, a trained One-Class SVM learns to aggregate modality-specific scores. AST [32] follows a Teacher-Student paradigm, in which the Teacher leverages Normalizing Flows to model the multimodal features distribution, while the Student employs a feed-forward network to model such distribution. This asymmetry in the architectures exacerbates the discrepancies between nominal and anomalous features at inference time. Recently, starting from the same transformer-based backbones as in [36], CFM [14] learns to map features from one modality to the other by nominal samples and then, at inference time, detects anomalies by pinpointing inconsistencies between observed and predicted features. In this paper, we employ the above-mentioned methods to investigate the challenges of SiM3D by adapting them to work in a multiview setup." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.681, + 0.801 + ], + "angle": 0, + "content": "3. Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Material preparation. We scanned eight types of manufactured objects spanning different sizes, shapes, textures and materials, as shown in Fig. 3 (1.i). For each object type, between 20 and 100 instances were purchased from a store. We selected objects suitable for single-instance scenarios – instances are manufactured according to a unique" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20946" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.091, + 0.2, + 0.101 + ], + "angle": 0, + "content": "(1) Material preparation" + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.104, + 0.187, + 0.113 + ], + "angle": 0, + "content": "(i) Real objects purchase" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.114, + 0.152, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.155, + 0.152, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.155, + 0.114, + 0.204, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.114, + 0.255, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.15, + 0.255, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.114, + 0.304, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.092, + 0.413, + 0.101 + ], + "angle": 0, + "content": "(2) Acquisition setup" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.114, + 0.475, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.192, + 0.195, + 0.199 + ], + "angle": 0, + "content": "(ii) CAD models collection" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.201, + 0.157, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.234, + 0.157, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.201, + 0.208, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.234, + 0.208, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.201, + 0.258, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.234, + 0.258, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.201, + 0.311, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.234, + 0.311, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.212, + 0.422, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.225, + 0.485, + 0.251 + ], + "angle": 0, + "content": "High-precision calibration pattern \nUnder 1mm precise corners" + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.302, + 0.377, + 0.324 + ], + "angle": 0, + "content": "Top-tier robot arm Precise poses" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.269, + 0.469, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.092, + 0.581, + 0.101 + ], + "angle": 0, + "content": "(3) Data gathering" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.104, + 0.59, + 0.111 + ], + "angle": 0, + "content": "(i) Real data acquisition" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.114, + 0.661, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.174, + 0.618, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.202, + 0.554, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.202, + 0.587, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.209, + 0.614, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.233, + 0.612, + 0.24 + ], + "angle": 0, + "content": "View N" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.233, + 0.546, + 0.239 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.24, + 0.558, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.24, + 0.616, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.214, + 0.672, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.251, + 0.674, + 0.256 + ], + "angle": 0, + "content": "Integrated Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.27, + 0.607, + 0.277 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.29, + 0.584, + 0.3 + ], + "angle": 0, + "content": "Prototype registration" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.299, + 0.591, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.289, + 0.667, + 0.296 + ], + "angle": 0, + "content": "Realistic rendering" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.299, + 0.684, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.331, + 0.678, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.092, + 0.756, + 0.101 + ], + "angle": 0, + "content": "(4) Labelling" + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.108, + 0.823, + 0.114 + ], + "angle": 0, + "content": "Annotated Views" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.115, + 0.785, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.116, + 0.884, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.173, + 0.785, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.173, + 0.852, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.192, + 0.771, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.192, + 0.817, + 0.2 + ], + "angle": 0, + "content": "↓" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.204, + 0.785, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.208, + 0.816, + 0.215 + ], + "angle": 0, + "content": "rojection" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.223, + 0.785, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.225, + 0.882, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.285, + 0.773, + 0.292 + ], + "angle": 0, + "content": "Integrated Mesh" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.296, + 0.785, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.35, + 0.78, + 0.357 + ], + "angle": 0, + "content": "3D Segmentation GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.286, + 0.871, + 0.292 + ], + "angle": 0, + "content": "Annotated Mesh" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.316, + 0.879, + 0.332 + ], + "angle": 0, + "content": "Manual Refinement & Voxelization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 3. SiM3D dataset creation pipeline. (1) We purchased eight types of objects with different properties, each associated with a CAD model. We altered part of the samples to create realistic anomalies; (2) We designed and calibrated an acquisition setup consisting of a high-precision 3D sensor with high-resolution cameras on top an industrial robot arm; (3) We acquired 360-degree scans of each object instance to populate the real split of the training and test sets, and rendered images and point clouds from the CAD models to create the synthetic split of the training sets; (4) We created voxel-based segmentation GTs by a mixed 2D-3D labelling pipeline." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.746 + ], + "angle": 0, + "content": "prototype model so that all nominal ones should be identical – and for which an official CAD prototype was available from the manufacturer. The prototype CAD models are depicted in Fig. 3 (1.ii). To simulate various anomalies, we manually introduced different types of defects that represent realistic manufacturing flaws, as illustrated in Fig. 3 (1.iii). We altered the objects' appearance (e.g., paint modifications, Fig. 3 (1.iii) – bottom left), geometry (e.g., dents, Fig. 3 (1.iii) – bottom right), or both (e.g., contaminations, Fig. 3 (1.iii) – top left). It is important to point out that only half of the instances within each object type are modified, while the remaining half are kept in their original state to serve as nominal data. For each object type, one nominal instance will be used as the training set in the real2real experimental setup, while all other ones, both nominal and modified (i.e., anomalous), will make up the test set. This very same test set is also used in the synth2real setup, though the training data comes from CAD prototypes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Sensor. We acquired our dataset with a top-tier 3D scanner, the Atos Q from ZEISS (Fig. 3 (2.i)). It consists of a stereo pair with two \\(12\\mathrm{Mpx}\\) greyscale cameras and a light projector. We consider the left camera to be our reference. The images have a resolution of \\(4096\\times 3000\\) pixels. The point clouds can feature up to 12 million elements, with a point-to-point distance between 0.04 and \\(0.15\\mathrm{mm}\\). The sensor is mounted on an industrial robot arm with high repeatability and precision (Fig. 3 (2.iii))." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Calibration. We calibrate the greyscale cameras with a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.595 + ], + "angle": 0, + "content": "standard checkerboard pattern, estimating the intrinsic parameters matrix, \\( A \\), and the radial and tangential distortion coefficients, \\( \\delta = [k_{1}, k_{2}, k_{3}, p_{1}, p_{2}] \\). We also calibrate the rotation, \\( R_{pc} \\), and translation, \\( T_{pc} \\), from the point cloud to the left camera reference frame, with a precise dotted calibration pattern built by ZEISS (Fig. 3 (2.ii)) that provides centre coordinates with precision \\( < 1 \\) mm. More details are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Real data acquisition. The acquisition process consists of 360-degree scanning operations around all object instances of our dataset (see Fig. 3 (3.i)). We define specific viewpoints to capture data from multiple angles around the object. These trajectories are sampled from concentric hemispheres that depend on the object's size. We provide poses relative to a reference view, i.e., \\( v_{1} \\). Thus, the pose for the view \\( i \\), \\( v_{i} \\), consists of a rotation \\( R_{i} \\) and a translation \\( T_{i} \\) that aligns \\( v_{i} \\) to \\( v_{1} \\), with \\( R_{1} \\) the identity matrix and \\( T_{1} \\) null. Since in real industrial setups object instances would be acquired from as similar as possible viewpoints, we place markers in the scene to facilitate consistent positioning of objects of the same type. Thus, thanks to the high repeatability of the robot arm, images, \\( I_{i} \\), and point clouds, \\( P_{i} \\), from the same viewpoint \\( v_{i} \\) turn out similar across acquisitions of different instances of the same object type. However, since the items were positioned manually with the help of the markers, slight displacements cannot be avoided. This introduces some variability across the scans of the instances of an object, as it also typically occurs in real indus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "trial pipelines. For each object, after collecting data from all viewpoints, the ATOS Q sensor also returns an integrated mesh, \\( M \\), obtained by registering all point clouds from the different views using a proprietary algorithm. We collect and release the mesh to foster the development of methodologies that may also exploit this 3D representation. The 3D reference frame is aligned to the one of the first point cloud \\( P_{1} \\) taken from the first viewpoint \\( v_{1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.217, + 0.487, + 0.489 + ], + "angle": 0, + "content": "Synthetic data generation. We employ the Blender [11] Python API to render synthetic data that simulates real acquisitions. To ensure the synthetic and real data are aligned, we select a reference mesh from the collected real instances for each object type, load it in Blender, and align it to the prototype CAD model, as shown in (Fig. 3 (3.ii) - left). Then, we set the Blender camera parameters to the intrinsics, \\( A \\), of the left camera of our Atos Q sensor and render images from the same viewpoints \\( v_{i} \\) employed during the real acquisitions. The RGB renders are converted into greyscale images to match the real images provided by our acquisition setup (Fig. 3 (3.ii) - right). Renderings have been obtained by exploiting Cycles Printerer, a path-tracing renderer able to provide realistic results with reflections, emissive surfaces, and physically-based lighting. We also render depth maps (Fig. 3 (3.ii) - right), which we project into 3D point clouds (see Fig. 1, centre) using the intrinsic parameters, \\( A \\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.494, + 0.487, + 0.901 + ], + "angle": 0, + "content": "Labelling. Since some anomalies, such as paintings and minor scratches, are visible only in greyscale images, while others, such as dents, are better visible in 3D, we must consider both sensing modalities to obtain precise 3D segmentation ground-truths for anomalous samples. Therefore, we develop a two-step labelling strategy, illustrated in Fig. 3 (4), that operates on both 2D images and 3D data. First, we manually annotate all the view images, \\( I_{i} \\), where anomalies are clearly visible, obtaining dense 2D segmentation masks. We employ a different ID and colour for every defect to ease the voxel AUPRO computation (see Sec. 4). Subsequently, for each annotated image, \\( I_{i} \\), we leverage the associated integrated mesh, \\( M \\), the intrinsic parameters \\( A \\), the point cloud to camera reference frame transformation, \\( (R_{\\mathrm{pc}}, T_{\\mathrm{pc}}) \\), and the transformation between \\( v_{i} \\) and \\( v_{1} \\), \\( (R_{i}, T_{i}) \\), to project the 2D annotations onto the integrated mesh. Then, since the projection may not be precise due to occlusions and slanted surfaces, each ground-truth mesh is thoroughly inspected and manually refined using a 3D visualisation and editing software, i.e., CloudCompare [12]. The labelling process was conducted by a team of 4 experts, with a final cross-check of the results. Finally, the annotated mesh is converted into a voxel grid with a voxel size of 2 mm. The choice of the size is determined by the smallest defect size discernible in the meshes. In the conversion process, a voxel is labelled as anomalous if it intersects at least one triangle labelled as such in the annotated mesh." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.163 + ], + "angle": 0, + "content": "
TypeDimensions [cm]Dimensions [voxels]Total InstancesTrain Instances NominalTest Instances NominalAnomalousViews per Instance
Plastic Stool\\( {35} \\times {35} \\times {30} \\)\\( {330} \\times {326} \\times {317} \\)221 real or 1 synth101012
Rubbish Bin\\( {26} \\times {21} \\times {33} \\)\\( {333} \\times {339} \\times {326} \\)421 real or 1 synth202012
Wicker Vase\\( {17} \\times {17} \\times {15} \\)\\( {329} \\times {309} \\times {290} \\)221 real or 1 synth101012
Bathroom Furniture\\( {33} \\times {33} \\times {50} \\)\\( {371} \\times {392} \\times {363} \\)201 real or 1 synth81036
Container\\( {20} \\times {25} \\times {10} \\)\\( {327} \\times {308} \\times {286} \\)941 real or 1 synth464612
Plastic Vase\\( {12} \\times {12} \\times 9 \\)\\( {326} \\times {308} \\times {286} \\)991 real or 1 synth484912
Wooden Stool\\( {48} \\times {42} \\times {45} \\)\\( {331} \\times {367} \\times {327} \\)151 real or 1 synth6712
Sink Cabinet\\( {44} \\times {25} \\times {50} \\)\\( {342} \\times {349} \\times {357} \\)191 real or 1 synth9836
" + }, + { + "type": "table_caption", + "bbox": [ + 0.628, + 0.173, + 0.792, + 0.186 + ], + "angle": 0, + "content": "Table 2. Dataset statistics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.216, + 0.7, + 0.231 + ], + "angle": 0, + "content": "4. SiM3D Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.242, + 0.762, + 0.258 + ], + "angle": 0, + "content": "4.1. Dataset, Tasks, and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.908, + 0.417 + ], + "angle": 0, + "content": "Dataset and Setups. Using the procedure described in Sec. 3, we collected 333 object instances belonging to the eight object types, as shown in Tab. 2. We captured from 12 to 36 views for each instance, depending on the type. We split the data into training and test sets. For each type, the training set comprises a single instance, either real in the real2real setup or synthetic in the synth2real setup. The test set includes all other instances and is balanced between nominal and anomalous samples. The test set is the same for both setups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Tasks. Our benchmark focuses on the tasks of detecting and segmenting objects' anomalies using multiview and multimodal data. Specifically, the inputs to the methods consist of a set of images \\(\\{I_i\\}_{i=1}^n\\) captured from different viewpoints alongside 3D data, \\(\\mathbf{X}\\), that capture an object entirely. The 3D data that a method can use are a set of point clouds \\(\\{P_i\\}_{i=1}^n\\) or depth maps \\(\\{D_i\\}_{i=1}^n\\) taken from the same vantage points as the images, an integrated mesh \\(M\\), or also a merged point cloud \\(P\\). A method capable of multiview multimodal 3D anomaly detection should take these inputs, integrate the 2D and 3D information from multiple views, and produce a global anomaly score for the test sample under inspection. On the other hand, a method pursuing multiview multimodal 3D anomaly segmentation should yield a 3D output, where each 3D position carries a score representing the likelihood of belonging to an anomaly. In our benchmark, a method for this task takes as input all images and 3D data, \\((\\{I_i\\}_{i=1}^n, \\mathbf{X})\\) to output an Anomaly Volume, i.e., a voxel grid, \\(V \\in \\mathbb{R}^{X \\times Y \\times Z}\\), where \\(X, Y\\), and \\(Z\\) are the grid dimensions and each voxel carries an anomaly score. For several reasons, we chose voxel grids as the 3D output representation of our benchmark. Firstly, as their grid structure resembles 2D Anomaly Maps, existing 2D segmentation methods and metrics can be easily extended. Moreover, voxel grids are suitable for representing every type of 3D defect, including missing parts. One major challenge with voxel grids pertains to choosing the voxel size. Ideally, we would want a size that allows accurate segmentation of even the smallest defects. However, the memory occupancy of voxel grids scales cubically with the inverse of the voxel size. Thus, we create ground-truth volumes using a voxel resolution of \\(2\\mathrm{mm}\\), which enables precise localisation of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Extending singleview methods. Anomaly Maps obtained by processing singleview inputs with standard ADS methods are projected into the 3D and aggregated into a voxel grid, with each voxel containing a list of scores from multiple views. The Anomaly Volume is obtained by voxel-wise max pooling." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.483, + 0.381 + ], + "angle": 0, + "content": "even the smallest defects present in our dataset while keeping memory occupancy affordable." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.383, + 0.485, + 0.825 + ], + "angle": 0, + "content": "Metrics. Akin to standard practice in 2D ADS [4], we evaluate detection performance by the Area Under the Receiver Operator Curve (AUROC) of the global anomaly scores predicted by the methods. As we compute a global score for each object instance, we name it instance-level AUROC (I-AUROC). Regarding segmentation performance, we evaluate methods with an adaptation of the popular Area Under the Per-Region Overlap (AUPRO) curve. First of all, we define the voxel-based Per-Region Overlap (PRO) metric as \\(\\mathrm{PRO}_t = \\frac{1}{N}\\sum_{n = 1}^N\\frac{|\\overline{\\Psi}_t\\cap V_n|}{|V_n|}\\), where \\(\\overline{\\Psi}_t\\) is an Anomaly Volume binarised using a threshold, \\(t\\), \\(N\\) is the number of blobs, i.e., defects, in the ground-truth, and \\(V_{n}\\) are the voxels belonging to that blob. Since during labelling, we annotated each defect in our dataset with a different ID (see Sec. 3), the \\(N\\) defect blobs in a ground-truth are easily identifiable. This metric is computed across multiple thresholds, \\(t\\), estimated at different False Positive Rates (FPRs). We calculate the FPR using only non-empty voxels. We sample FPRs from 0 to the maximum value uniformly, we estimate the corresponding threshold \\(t\\), and we binarise the score grid accordingly. In this way, we construct a curve that plots the PRO values against the corresponding FPRs. The segmentation performance is obtained by integrating such a curve up to a specific False Positive Rate. As proposed in [14], we use \\(1\\%\\) as integration bound for the PRO curve to set a challenging bar representative of the requirements of real industrial applications. Finally, we normalise the resulting area to the interval [0, 1]. We dub this metric voxel-level AUPRO (V-AUPRO)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.2, + 0.849 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Extending singleview methods to Multiview 3D ADS. Current ADS methods predict a 2D anomaly map from a singleview input, either unimodal, e.g., images, or multi-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.38 + ], + "angle": 0, + "content": "modal, e.g., images and point clouds. To obtain a set of baselines for our novel multiview 3D ADS task, we extend popular and effective singleview methods by means of the 3D aggregation strategy illustrated in Fig. 4. Given the considered method, we process the data corresponding to an input view, \\( v_{i} \\), so as to obtain a 2D Anomaly Map. Similarly to Sec. 3, given the mesh, \\( M \\), intrinsic parameters \\( A \\), point cloud to camera reference frame transformation, \\( (R_{\\mathrm{pc}}, T_{\\mathrm{pc}}) \\) and view pose \\( v_{0} \\), \\( (R_{i}, T_{i}) \\), we can project each pixel of the 2D Anomaly Map to its corresponding 3D position. We discretise the 3D position, finding the corresponding voxel index. We associate that voxel with the corresponding anomaly score. We repeat the process for each view, keeping track of the scores that project onto the same voxels. Eventually, we compute the maximum among all scores projected into each voxel, obtaining the final Anomaly Volume. To extract the global score required by the I-AUROC, we compute the maximum across all the scores within the Anomaly Volume." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Selected Methods. We extend some of the most popular and effective unimodal and multimodal singleview ADS methods, starting from their official code, and evaluate them on SiM3D. Regarding unimodal methods, we select PatchCore [30], with either WideResNet101 [42] or DINOv2 [24] as backbone, to extract features from images of all views, \\(\\{I_i\\}_{i=1}^N\\) of the single-instance training set to create memory banks. Moreover, we use EfficientAD [1] using all views as their training set. As for multimodal ADS methods, we use BTF [18] employing WideResNet101 [42] as the image encoder and FPFH [33] as the algorithm to extract point cloud features. We create memory banks using features extracted from images, \\(\\{I_i\\}_{i=1}^N\\), and point clouds, \\(\\{P_i\\}_{i=1}^N\\), of all views of the single-instance training set. We use the same strategy with M3DM [36]. As 2D feature extractor, we exploit DINO-v2. As point cloud feature extractor, instead of using a transformer backbone, i.e., PointMAE [17], we utilise FPFH to extract 3D features, similarly to BTF. Indeed, we would need a severe downsample (from \\(\\sim 5 - 7\\mathrm{M}\\) to \\(\\sim 8\\mathrm{k}\\) points) to work with Point-MAE, and we verified that at such low resolution, most of the defects no longer show up in the point clouds. Finally, we select CFM [14], using all views and clouds as a single training set. Based on the same consideration as for M3DM, we use DINO-v2 and FPFH to produce 2D and 3D features. Due to computational constraints, all methods process images padded and downsampled to \\(1540 \\times 1540\\). Moreover, for compatibility with the employed 2D feature extractors, which accept RGB images, we create 3-channel inputs by stacking our greyscale images thrice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Data pre-processing. To obtain less noisy data and improve the results of the above-mentioned baselines, we remove the background from 3D data. We estimate a 3D plane using a RANSAC-based plane segmentation algo" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.098, + 0.892, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.359, + 0.904, + 0.374 + ], + "angle": 0, + "content": "Figure 5. Qualitatives. Selected views with defects (left) and Anomaly Volumes for several methods (right) downsampled for visualisation." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.388, + 0.902, + 0.478 + ], + "angle": 0, + "content": "
MethodModalityreal2realsynth2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.MeanPl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
PatchCore w/ WRN-101RGB0.7400.9870.6360.7770.7740.5511.0000.5660.7540.4620.2350.5370.3530.6780.5760.5170.2500.451
PatchCore w/DINO-v2RGB0.5000.9580.6360.6220.5780.5631.0000.5630.6780.9580.8970.4130.4640.2070.6840.6070.0870.540
EfficientADRGB0.2800.7320.0000.8780.4240.7300.9280.7120.5850.1230.6730.0000.8480.0080.4870.7500.6620.443
ASTRGB0.5370.5660.6110.5150.5730.4960.8390.5370.5841.0000.3441.0000.5250.4820.3880.4280.1870.544
BTFRGB + Point Cloud0.4210.2170.5040.5650.5450.4710.6780.4240.4780.4620.2940.5200.4440.4640.4240.4820.2870.422
CFM w/DINO-v2 + FPFHRGB + Point Cloud0.1980.3010.0740.5150.4830.4560.7320.8250.4480.0080.0000.1900.4240.3130.4590.4280.3620.273
M3DM w/DINO-v2 + FPFHRGB + Point Cloud0.7020.9880.6610.5450.5560.6490.3920.4750.6210.1070.1170.7350.5650.3810.5860.2140.5120.402
ASTRGB + Depth0.9500.9270.7850.4740.5420.4700.4280.9250.6870.6360.0020.5040.4740.4620.5690.4280.8870.495
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.489, + 0.776, + 0.503 + ], + "angle": 0, + "content": "Table 3. Anomaly detection results (I-AUROC). Best results in bold, runner-ups underlined." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.486, + 0.892 + ], + "angle": 0, + "content": "rithm, which, iteratively, randomly selects a small subset of points (e.g., 10) to fit a candidate plane model. For the candidate planes, we compute the consensus set with threshold \\(\\tau\\) on the point-to-plane distance. After 1000 iterations, we return the best-fitting plane. To account for spurious background artefacts, we shift the plane by an offset \\(\\alpha\\) and finally remove all points below the shifted plane. Given the different sizes of the objects, for each object type, we select different optimal thresholds \\(\\tau\\) and offsets \\(\\alpha\\), as detailed in the supplementary material. We apply this background removal technique to the integrated mesh, \\(M\\), due to the larger number of background points, which facilitates the RANSAC algorithm. Then, from the background-filtered integrated mesh, we leverage ray-tracing techniques and known camera parameters to render, for each view, depth maps, \\(\\{D_i\\}_{i=1}^N\\), pixel-aligned to the corresponding images, \\(\\{I_i\\}_{i=1}^N\\) at the sensor resolution \\((4096 \\times 3000)\\). Finally, these depth maps are padded and downsampled to \\(1540 \\times 1540\\), and the depth is lifted in 3D via the camera parameters to obtain dense and clean point clouds. The depth maps and point clouds obtained by rendering the preprocessed meshes serve as the input data for all the experiments presented in the subsequent section and will be available in the SiM3D dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.704, + 0.544 + ], + "angle": 0, + "content": "5. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Main Results. In Tab. 3 and Tab. 4, we report the results obtained on the SiM3D benchmark by extending the considered methods as Multiview 3D ADS baselines. By looking at the mean detection performance (Tab. 3), we notice how methods that natively process only RGB images outperform natively multimodal ones in both the real2real and synth2real setups. With the notable exception of AST, this also stands for mean segmentation performance (Tab. 4). We posit that this is due to the main multimodal methods having been designed and tuned to process point clouds of the resolution featured by existing multimodal benchmarks, which is much lower than that of SiM3D, i.e. \\(\\sim 8\\mathrm{K}\\) vs \\(\\sim 2.3\\mathrm{M}\\) points (after downsampling). Indeed, AST is the only one designed to work natively with depth maps. Hence, our experiments suggest that current multimodal ADS methods do not scale to significantly higher resolutions. In turn, we could not even deploy the native 3D feature extractors of M3DM and CFM, the methods currently excelling on [6, 7]. Furthermore, we notice how methods relying on memory banks, such as PatchCore and M3DM, tend to attain slightly better mean detection and segmentation performance than their counterparts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20950" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.09, + 0.27, + 0.178 + ], + "angle": 0, + "content": "
MethodModality
PatchCore w/ WRN-101RGB
PatchCore w/ DINO-v2RGB
EfficientADRGB
ASTRGB
BTFRGB + Point Cloud
CFM w/ DINO-v2 + FPFHRGB + Point Cloud
M3DM w/ DINO-v2 + FPFHRGB + Point Cloud
ASTRGB + Depth
" + }, + { + "type": "table", + "bbox": [ + 0.274, + 0.091, + 0.585, + 0.178 + ], + "angle": 0, + "content": "
real2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
0.7100.4610.7610.6750.6940.7470.3800.6090.630
0.7450.4690.7750.7920.7090.7530.4350.6900.671
0.6820.4620.7630.5340.6800.7430.4070.4880.594
0.7200.5020.7890.8030.7160.7640.4460.7580.687
0.5510.4020.7500.3770.6140.7410.0920.0300.444
0.5970.4150.7680.5050.6400.7430.3150.3210.538
0.7330.4520.7670.7020.7020.7520.2880.1260.565
0.7500.5030.7920.8070.7160.7640.4670.7980.699
" + }, + { + "type": "table", + "bbox": [ + 0.587, + 0.091, + 0.901, + 0.178 + ], + "angle": 0, + "content": "
pl.Stoolrub.BinW.VaseB.Furn.Cont.Pl.VaseW.StoolSink Cab.Mean
0.7340.4370.7510.4540.6780.7410.3860.6180.600
0.7010.4570.7720.5630.6480.7340.3210.5750.596
0.6800.4490.7290.5230.6690.7470.4310.3610.573
0.7260.4880.7910.8060.7050.7640.4610.6950.679
0.5470.4020.7560.3690.6100.7410.1030.0400.446
0.5230.4130.7530.5420.6210.7410.2220.3140.516
0.6900.4540.7700.4610.6450.7340.3180.1320.526
0.7510.5020.7930.8050.7170.7640.4500.7850.695
" + }, + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.19, + 0.809, + 0.205 + ], + "angle": 0, + "content": "Table 4. Anomaly segmentation results (V-AUPRO@1%). Best results in bold, runner-ups underlined." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.23, + 0.293, + 0.362 + ], + "angle": 0, + "content": "
MethodModalityFeatures
2D3D
PatchCoreRGBWRN-101-
RGBDINO-v2-
Depth-WRN-101
Depth-DINO-v2
Point Cloud-FPFH
BTFRGB + Point CloudWRN-101FPFH
RGB + DepthWRN-101WRN-101
CFMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
M3DMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
ASTRGBEffNet-B5-
RGB + DepthEffNet-B5EffNet-B5
" + }, + { + "type": "table", + "bbox": [ + 0.296, + 0.23, + 0.479, + 0.362 + ], + "angle": 0, + "content": "
real2realsynth2real
DetectionSegmentationDetectionSegmentation
0.7540.6300.4510.600
0.6780.6710.5400.596
0.5820.4860.5070.484
0.6000.4960.2460.471
0.4150.4640.3130.460
0.4780.4440.4220.446
0.7070.6090.4480.543
0.4480.5380.2730.516
0.5480.6630.3110.620
0.6210.5650.4020.526
0.6590.5030.2540.476
0.5840.6870.5440.679
0.6870.6990.4950.695
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.374, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Table 5. Anomaly detection and segmentation mean results with different input modalities. Best results per method in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.854 + ], + "angle": 0, + "content": "within the same modality that require substantial training, i.e., EfficientAD and CFM. We argue that this may be ascribed to the single-instance setup, which renders the latter category of approaches harder to optimise due to the limited number of training images, which are as many as the views. The methods most effectively extended to our multiview 3D task are Patchcore, which, equipped with either WRN-101 or DINO-v2, can achieve the best detection in both setups, and AST, achieving the best segmentation performance in both setups. Indeed, Patchcore yields I-AUROCs of 0.754 (WRN-101) and 0.540 (DINO-v2), while AST yields a V-AUPRO@1% of 0.699 and 0.695, in the real2real and synth2real setups, respectively. It is also worth pointing out how, due to the domain shift, we can notice a gap between the real2real and synth2real setups, especially in detection (0.754 vs. 0.540, I-AUROC) compared to segmentation performance (0.699 vs. 0.695, V-AUPRO@1%). The qualitative results, shown in Fig. 5, visually support the above observations. Overall, the relatively low, definitely far from saturated, performance figures provided by baselines designed by straightforward extension of state-of-the-art methods delivering impressive results in singleview 2D benchmarks hint at the novel and challenging nature of the task set forth by SiM3D. Thus, our experimental findings by baseline approaches reveal the need for more advanced and specific methods and paradigms, both in the real2real and, even more so, the synth2real setups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Alternative 3D Representations. Given the lack of feature extractors amenable to processing high-resolution point clouds, we explore the use of depth maps to capture 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.908, + 0.61 + ], + "angle": 0, + "content": "cues. The 2D grid structure of depth maps enables the deployment of popular backbones pre-trained on RGB data that support high-resolution inputs, such as DINOv2. Thus, we create and evaluate additional baselines using depth maps and report the results in Tab. 5. Similarly to greyscale images, we tile depth maps along the channel axis to ensure compatibility with the employed backbones. The results show that the use of depth maps rather than point clouds in multimodal methods tends to improve detection and segmentation performance by sizable margins, e.g., in the real2real setup, BTF yields 0.707 I-AUROC and 0.609 V-AUPRO when processing depths compared to 0.478 I-AUROC and 0.444 V-AUPRO when fed with point clouds (row 6 vs. row 7) while relying on the same WRN101 image backbone. Indeed, all multimodal methods but M3DM achieve their best performance in both tasks and setups by processing depth maps instead of point clouds. Yet, as vouched by the results achieved by PatchCore, RGB images compare favourably to depth maps in terms of detection performance. We ascribe this to the feature extractors used to compute depth features being pre-trained on RGB images, which may be conducive to yielding weaker features when fed with depth inputs. This points out an intriguing research question: would it be possible to develop a foundation model to process depth maps effectively?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.676, + 0.639 + ], + "angle": 0, + "content": "6. Final Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.89 + ], + "angle": 0, + "content": "We introduced SiM3D, the first benchmark and dataset focused on integrating multiview and multimodal information for comprehensive 3D ADS. Our experiments highlight several open challenges. First, a need for methods that can process multiview inputs more effectively, e.g., by taking all views as input rather than integrating singleview outputs, and that can be trained with a very limited number of nominal samples, e.g., a single instance. Second, training with synthetic data requires more robust techniques to address domain-shift challenges. Finally, the current literature lacks 3D backbones capable of processing high-resolution point clouds. Employing depth maps seems a promising strategy to facilitate this processing, yet a foundational model for depth maps is still missing. We believe that SiM3D may serve as a strong foundation for tackling these challenges and fostering future research toward 3D ADS." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20951" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.191 + ], + "angle": 0, + "content": "We acknowledge financial support from the National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.1, Call No. 1409 published on 14 September 2022 by the Italian Ministry of University and Research (MUR), funded by the European Union." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.482, + 0.266 + ], + "angle": 0, + "content": "This work was also partially supported by the FSE+ 2021-2027 program under Article 24, paragraph 3, letter a) of Law 240/2010 and its subsequent amendments, as well as by Regional Resolution No. 693/2023 (Project Code: 2023-20090/ER), Grant CUP: J19J23000730002." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.482, + 0.327 + ], + "angle": 0, + "content": "We sincerely thank Lucia Gasperini for her contribution to dataset acquisition and labelling. We also extend our gratitude to Manuel Turrini for his support in setting up the acquisition process." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.341, + 0.187, + 0.356 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.366, + 0.482, + 0.435 + ], + "angle": 0, + "content": "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficient-tad: Accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 128-138, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.437, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[2] Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection. arXiv preprint arXiv:2002.10445, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.48, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[3] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[6] Paul Bergmann, Jin Xin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 202-213, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[7] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the 16th Asian Conference on Computer Vision (ACCV2022), 2022. ACCV. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.366, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[9] Li-Ling Chiu and Shang-Hong Lai. Self-supervised normalizing flows for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2926-2935, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[10] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. ArXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[12] CloudCompare Online Community. CloudCompare - 3D point cloud and mesh processing software Open Source Project. EDF, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[13] Alex Costanzino, Pierluigi Zama Ramirez, Mirko Del Moro, Agostino Aiezzo, Giuseppe Lisanti, Samuele Salti, and Luigi Di Stefano. Test time training for industrial anomaly segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3910-3920, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[14] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. CVPR. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[15] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[16] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 98-107, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[18] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2976, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[19] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[20] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Jiaqi Liu, Guoyang Xie, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, Feng Zheng, et al. Real3d-ad: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv preprint arXiv:2301.11514, 2, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[23] Fabio Valerio Massoli, Fabrizio Falchi, Alperen Kantarci, Şeymanur Akti, Hazim Kemal Ekenel, and Giuseppe Amato. Cocca: Multilayer one-class classification for anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 33(6):2313-2323, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[24] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 604-621. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[26] Jonathan Pinnay and Keng Chai. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing, pages 394-406. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[27] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[28] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[29] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13576-13586, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[30] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of 2022 IEEE Conference on Computer Vision and Pattern Recognition, pages 14298-14308, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[31] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907-1916, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[32] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[33] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE International Conference on Robotics and Automation, pages 3212-3217, 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[34] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[35] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22883-22892, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[36] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[37] Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650-656, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[38] Minghui Yang, Peng Wu, and Hui Feng. Memseg: A semi-supervised method for image surface defect detection using differences and commonalities. Engineering Applications of Artificial Intelligence, 119:105835, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[39] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[40] Seungdong Yoa, Seungjun Lee, Chiyoon Kim, and Hyunwoo J Kim. Self-supervised learning for anomaly detection with dynamic local augmentation. IEEE Access, 9:147201-147211, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[41] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[42] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[43] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[44] Zheng Zhang and Xiaogang Deng. Anomaly detection using improved deep svdd model with data structure preservation. Pattern Recognition Letters, 148:1-6, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[45] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection. arXiv preprint arXiv:2310.07716, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20953" + } + ] +] \ No newline at end of file diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_origin.pdf b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e67010f2d4c8965d6e4332061b78f5dd7ae8de4 --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/fc999caf-1202-46bd-84ea-94a9faab06d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe50a6a6493d106bd410382b83992bb68b5a92df1a4bfe7fa793cf1e7338b6bb +size 27131231 diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/full.md b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/full.md new file mode 100644 index 0000000000000000000000000000000000000000..596b1675c54c4fb0bd466d099a3ee0d5fd04bac8 --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/full.md @@ -0,0 +1,294 @@ +# SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark + +Alex Costanzino1 Pierluigi Zama Ramirez1 Luigi Lella1 Matteo Ragaglia2 Alessandro Oliva2 Giuseppe Lisanti1 Luigi Di Stefano1 +1CVLab, University of Bologna, Italy 2SACMI Imola, Italy https://alex-costanzino.github.io/SiM3D/ + +# Abstract + +We propose SiM3D, the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS), where the task is to produce a voxel-based Anomaly Volume. Moreover, SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. In this respect, SiM3D stands out as the first ADS benchmark that addresses the challenge of generalising from synthetic training data to real test data. SiM3D includes a novel multimodal multiview dataset acquired using top-tier industrial sensors and robots. The dataset features multiview high-resolution images (12 Mpx) and point clouds ( $\sim 7M$ points) for 333 instances of eight types of objects, alongside a CAD model for each type. We also provide manually annotated 3D segmentation GTs for anomalous test samples. To establish reference baselines for the proposed multiview 3D ADS task, we adapt prominent singleview methods and assess their performance using novel metrics that operate on Anomaly Volumes. + +# 1. Introduction + +In computer vision, anomaly detection and segmentation (ADS) involves identifying anomalous samples and localising their defects. This task is challenging in industrial contexts due to the variability and rarity of defects. Typically, ADS in the industry follows a cold-start approach, where training is unsupervised and uses only defect-free (nominal) samples. Before the release of the MVTec AD dataset in 2019 [4], only a limited number of ADS methods were proposed, with just 9 publications in major computer vision conferences (e.g., CVPR, ECCV, ICCV, WACV) over the five years before. The introduction of MVTec AD marked a pivotal moment in the field, fostering the publication of roughly 172 papers in these venues over the next 5 years. This dataset not only stimulated the development of several ADS methods [14, 30, 36] but also promoted the creation of novel and more challenging benchmarks. For instance, + +![](images/ced18cd97e949e4bffadc3884fba9fc61b90ce90b5a86edc90653bbcdc06c372.jpg) +Figure 1. SiM3D dataset overview. From top to bottom: the single-instance real and synthetic training samples for object Plastic Stool, one of the anomalous samples from the test set. + +MVTec 3D-AD and [6] and Eyecandies [7] introduced the first benchmarks deploying multimodal inputs to enhance ADS performance. Then, PAD [45] and Real-IAD [35] introduced novel benchmarks that leverage multiple RGB views of objects. Although these benchmarks tackle various settings, two real industrial issues remain unaddressed. + +First, existing benchmarks have primarily addressed the problem of 2D ADS. Consequently, current methods gen + +
DatasetVenue & Year2DNo. Pixels3DNo. PointsScenarioTrainTestSetupTaskNo. Classes
MVTec ADCVPR 2019RGB Images490k - 1M——Multi-instanceSingleviewSingleviewreal2real2D15
MVTec 3D-ADVISAPP 2021RGB Images160k - 810kXYZ Images10k - 195kMulti-instanceSingleviewSingleviewreal2real2D10
EyecandiesACCV 2022RGB Images262kDepths + Normals262kMulti-instanceSingleviewSingleviewsynth2synth2D10
MVTec LOCO ADIJCV 2022RGB Images1.28M - 2.05M——Multi-instanceSingleviewSingleviewreal2real2D5
VisAECCV 2022RGB Images1.5M——Multi-instanceSingleviewSingleviewreal2real2D12
PADNeurIPS 2023RGB Images12M——Single-instanceMultiviewSingleviewreal2real + synth2synth2D20
Real3D-ADNeurIPS 2023——Point Clouds43k - 2.68MMulti-instanceMultiviewSingleviewreal2real3D12
Real-IADCVPR 2024RGB Images10M——Multi-instanceSingleviewSingleviewreal2real2D30
SiM3D2025Grayscale Images12MPoint Clouds + Mesh5M - 7MSingle-instanceMultiviewMultiviewreal2real + synth2real3D8
+ +Table 1. Popular ADS benchmarks. Chronologically sorted, the sequence hints at a growing interest in multimodal and multiview data. + +![](images/d25cb59f2febfc8be8552ba1c6a8346ef4dad0420ec11fa46f90073270c39a42.jpg) +Figure 2. 2D vs. 3D anomaly detection and segmentation. Typical ADS benchmarks rely on 2D Anomaly Maps and 2D ground-truths (left) to assess segmentation performance. SiM3D proposes to do so by 3D Anomaly Volumes and 3D ground-truths (right). + +erate Anomaly Maps, i.e., images that encode the likelihood of each pixel belonging to a defect, as shown in Fig. 2 (left). However, many industrial applications require precise localisation of defects in the 3D space to allow automatic intervention, reducing waste and optimising production. Benchmarks relying on multimodal input data, like those proposed in MVTec 3D-AD [6] and Eyecandies [7], represent progress toward this goal by providing 3D information for each pixel within an image. However, in these benchmarks, ADS methods are still expected to provide as output a 2D Anomaly Map, and objects are inspected from a single viewpoint. For comprehensive 3D anomaly detection, we must observe an object in its entirety, utilising captures from multiple viewpoints to detect defects across the entire 3D space. This calls for methods capable of integrating multiview information to estimate the defect probability at each 3D location within an Anomaly Volume, as shown in Fig. 2 (right). Furthermore, as demonstrated by MVTec 3D-AD and Eyecandies, these methods should process multimodal information, i.e., RGB images as well as 3D information, to achieve more reliable anomaly detection. + +Secondly, current benchmarks typically provide multiple training samples per object class to capture the variability across nominal objects. However, collecting many samples - even nominal ones - can be costly and challenging. For example, line changeovers would demand acquiring new datasets and retraining models, both highly time-intensive processes. Additionally, collecting large datasets increases the risk of errors, such as misjudged nominal samples, which may degrade model performance. Moreover, manufactured objects typically closely resemble one another as they replicate the structure and materials of a syn + +thetic CAD prototype. Thus, the variability across nominal samples of manufactured objects is frequently minimal, with a single nominal item subsuming the information required to spot anomalies in other objects. + +These considerations highlight the need for ADS techniques that may be effectively trained by a single nominal sample or even a synthetic prototype (e.g., an object's CAD model) and then be able to generalise to other manufactured objects to identify anomalous ones effectively. The ability to perform training by a single object instance, either real or synthetic, would avoid the need for extensive data collection, vastly facilitating the creation and adaptation of ADS models. Yet, most current benchmarks focus on multiple training instances setups, and none of the prior works consider challenging synthetic (training) to real (test) setups. + +To fill all the gaps highlighted by our considerations and stimulate interest in the development of multimodal and multiview 3D ADS methods, we propose the Single-instance Multiview Multimodal Multisetup dataset, dubbed SiM3D. The dataset contains a series of high-resolution scans of manufactured objects, which can be nominal or anomalous. Each scan consists of multiple views, each featuring a greyscale image, its corresponding point cloud (i.e., multimodal data), and the ground-truth acquisition pose. In addition, an integrated 3D mesh is provided with each scan. For each object, a single nominal sample has been retained to be employed as a single-instance training set, while the remaining samples, both nominal and anomalous, compose the test set. Moreover, to foster research on novel approaches that may address the domain gap between synthetic training data and real test samples in ADS, for each object of our dataset, we also provide a synthetic scan, i.e., we deploy the available CAD model to obtain the same kind of data as in our high-resolution real scans, with images and point clouds rendered from the same poses. An overview of the proposed dataset is depicted in Fig. 1. + +Based on our novel dataset, we create the first benchmark for 3D anomaly detection, where the task is to predict anomaly scores within a voxelized 3D Anomaly Volume rather than a 2D Anomaly Map (Fig. 2). Peculiarly, our benchmark includes two setups based on the domain of the data available for training: in the real2real setup, a model is given access to a single nominal instance of an object (Fig. 1, top), while in synth2real the same kind + +of training data are obtained from the CAD model of the object (Fig. 1, centre). We adapt prominent ADS methods – both unimodal and multimodal – for the proposed task and evaluate them on our benchmark. To assess segmentation performance, we provide accurate 3D ground-truths (Fig. 2) and extend ADS metrics to operate on anomaly scores and ground-truths specified on 3D voxel grids. Our experimental findings highlight the open challenges in the field and offer insights into possible future directions. + +# 2. Related Work + +The MVTec AD [4] dataset significantly boosted anomaly detection research, inspiring new techniques and benchmarks for various challenges. MVTec AD provides a training set containing only nominal images and a test set that includes nominal and anomalous ones for several object categories. This unsupervised setup exemplifies many real-world applications in which anomalous objects are challenging to collect, and thus, it has also been adopted by more recent ADS benchmarks. Among them, MVTec LOCO [5] presents a challenging scenario with both structural anomalies (e.g., dents, holes) and logical anomalies (e.g., incorrect object combination). VisA [46] introduced high-resolution images of complex scenes containing multiple instances of the same object. However, certain defects may be visible only by analysing objects' 3D geometry. For this reason, multimodal datasets, such as MVTec 3D-AD [6] and Eyecandies [7] include 3D data alongside images. For each sample in the dataset, MVTec 3D-AD provides images and pixel-aligned XYZ information captured by a 3D sensor, while Eyecandies contains synthetic images with pixel-aligned depths and normals. Recently, PAD [45] addressed pose-agnostic ADS with the first multiview dataset (MAD), featuring images of objects from various viewpoints. In PAD, training samples are multiview RGB images with known poses, while test samples are singleview RGB images with unknown poses. MAD includes both synthetic and real objects, but its benchmarks do not explore synthetic-to-real generalisation. Real-IAD [35] introduced a large-scale multiview dataset with RGB images covering various defect sizes. Eventually, Real3D-AD [21] is the first point cloud anomaly detection benchmark where the task is to predict an anomaly score for each point. It uses front-and-back scans of nominal objects for training and a singleview point cloud for testing. + +As shown in Tab. 1, the proposed SiM3D benchmark offers several outstanding features, including very-high-resolution images (12 Mpx), large point clouds/meshes containing 5-7 million elements captured with high-precision 3D sensors. Moreover, it is designed for a highly relevant manufacturing scenario: single-instance anomaly detection, where only one object is available for model training. To promote research toward even more cost-effective scenar + +ios, ours is also the only benchmark that addresses the challenge of generalising from synthetic training data to real test data. Finally, SiM3D stands out as the first benchmark that considers integrating multiview information at test time for comprehensive 3D anomaly detection, where the task consists of outputting a voxel-based Anomaly Volume. + +ADS Methods. Following the setup defined in the MVTec-AD benchmark, most ADS methods [22] are designed to estimate 2D anomaly maps from a single image [2, 3, 9, 10, 13, 15, 16, 19, 20, 23, 26, 27, 27-29, 31, 34, 37-41, 43, 44]. One of the most popular methods is PatchCore [30], which uses feature extractors pre-trained on large datasets [8, 17, 24] to gather features from nominal samples into a memory bank. At test time, input sample features are compared to this bank to detect anomalies. Another more recent yet popular approach, EfficientAD [1], uses a lightweight feature extractor and a student-teacher model to identify anomalies based on feature prediction discrepancy. The introduction of the MVTec 3D-AD benchmark has led to the development of several multimodal approaches that use images and 3D data to generate 2D anomaly maps [14, 18, 32, 36]. Among them, BTF [18] is a memory bank approach inspired by PatchCore, in which each element is the concatenation of 2D and 3D features extracted by a pre-trained image backbone and a traditional point cloud descriptor, i.e., FPFH [33], respectively. M3DM [36] improves BTF using powerful transformer-based pre-trained backbones, such as DINO-v1[8] and Point-MAE [25], to extract features from RGB images and point clouds. It also learns to fuse 2D and 3D features into multimodal features. During inference, a trained One-Class SVM learns to aggregate modality-specific scores. AST [32] follows a Teacher-Student paradigm, in which the Teacher leverages Normalizing Flows to model the multimodal features distribution, while the Student employs a feed-forward network to model such distribution. This asymmetry in the architectures exacerbates the discrepancies between nominal and anomalous features at inference time. Recently, starting from the same transformer-based backbones as in [36], CFM [14] learns to map features from one modality to the other by nominal samples and then, at inference time, detects anomalies by pinpointing inconsistencies between observed and predicted features. In this paper, we employ the above-mentioned methods to investigate the challenges of SiM3D by adapting them to work in a multiview setup. + +# 3. Dataset Creation + +Material preparation. We scanned eight types of manufactured objects spanning different sizes, shapes, textures and materials, as shown in Fig. 3 (1.i). For each object type, between 20 and 100 instances were purchased from a store. We selected objects suitable for single-instance scenarios – instances are manufactured according to a unique + +![](images/1ff934651d17324fe9b5aa23751bafe5d029740d434cd7fc88f931420a69846c.jpg) +(1) Material preparation + +![](images/e4f1f8ee071598b7c23b6deaa553ee65f938240487e0bc84a408e515c28eb0d1.jpg) + +![](images/5b7bdb6910b4eefc26a97782b5754dafb421bb835adeaac43ff366ce7d78ab29.jpg) +(i) Real objects purchase + +![](images/7e45da2c68f71f6644c49964ab4dac5970ffc875760a72c14c35fa6db6ae0bc4.jpg) + +![](images/e11ccfd440b9929915ea1741545982a25ca45086c92e3529f6be5e0dbe540465.jpg) + +![](images/f23dd7682ba3295026491b4676d9e3ddcfcc8a7ad8248066e0e62abb7d01618b.jpg) +(2) Acquisition setup + +![](images/68d138f94156275eddb1246fa993dd2a655fb34b12979bf4a4f3f77c9b7dbc52.jpg) + +![](images/e75cea0664333c347b6a442fd81cd16cbbe14183f9521011ec893709d1da3031.jpg) +(ii) CAD models collection + +![](images/94ac0ef751c1158402051d318db7692407fa45096236e47ccd96390f6c340b6a.jpg) + +![](images/129621fc3da7cc9ae5d5ba7cd65cfa2ad4c3eb37afab97fa42472eb3c708e3e2.jpg) + +![](images/f447c1e84704c700eb2bc7c1f6612b56f8d761da53a895692c7195a285176bba.jpg) + +![](images/b42dfe144bf9eda909a083b7c16f5a57ca291e7ed48e1339d60393f3324fde33.jpg) + +![](images/0f7f7e34f26bb7aa2a32e98ebe4930be31cd99f3d293d4384fbcdfe52f84d551.jpg) + +![](images/efc0722f016b683c0b42077e36695c29d0c07cf3635ea7aa90ca660b190ff892.jpg) + +![](images/874b0a10fcfaf19188adc278c8d6710ad5d2cc9ea8fd419f10d22092ac09f4aa.jpg) + +![](images/7195463304f6052647ece2b6c83c1c9c8dbe6042e79ebc3274030437453fe7dc.jpg) +High-precision calibration pattern +Under 1mm precise corners + +![](images/ac3d3865a19941ae2db2a0f602279d7d60b12c2786fcc8708b6ed2adeb029e26.jpg) +Top-tier robot arm Precise poses + +![](images/9a2f3f8f307dd9a819dcc73db2d4228a24f502f46f38d2d362578ebf0f1a3aa7.jpg) +(3) Data gathering +(i) Real data acquisition + +![](images/a02b50d078fe3edafe71d0ab3638bc5c5bddd3542a3b9a2437eb34e09a5b26d5.jpg) + +![](images/9e3a0e084876fbf53d19b7af1fc8eed4d51a70b6a38e4a1c13de2bba21ec05bd.jpg) +View 1 + +![](images/a49ce4f839b822f2dc49b37e32f3383ca19abf576328ad915a9ceaae867d796a.jpg) + +![](images/157ea8aba7041065075008065c61a237319e4a144353b0fe90c9b39e9e2b493a.jpg) + +![](images/edd9316479c3794334faca86caa59d0c0844e9915c29d60895499be4c73f9a33.jpg) + +![](images/17ece7b18ebb80ba195c10661735b761b558fa0a34f53240737470146282244c.jpg) +View N +# + +![](images/785e121e9804db3f8fe7b2ad9cc7a685db19d0e4ec2a939b5a46758b963bfe07.jpg) +Integrated Mesh + +![](images/7fcd01438e7915a9e95b8ff928d5bc8a751a14845db799f154d18b3e556e89f9.jpg) +Prototype registration +Figure 3. SiM3D dataset creation pipeline. (1) We purchased eight types of objects with different properties, each associated with a CAD model. We altered part of the samples to create realistic anomalies; (2) We designed and calibrated an acquisition setup consisting of a high-precision 3D sensor with high-resolution cameras on top an industrial robot arm; (3) We acquired 360-degree scans of each object instance to populate the real split of the training and test sets, and rendered images and point clouds from the CAD models to create the synthetic split of the training sets; (4) We created voxel-based segmentation GTs by a mixed 2D-3D labelling pipeline. + +![](images/ffcc203adb553e9a5da42024b7426e08604f984c3933c6532e21c127b3017161.jpg) +Realistic rendering + +![](images/27c6758f9ce2ca253503440d10d7f341a116d9ab404029ce0ba3d0241a1af103.jpg) + +![](images/4a33eb270d948decff7ffefb30c1862c185a613a16e0a24d8d262f8017e05ea4.jpg) +(4) Labelling +Annotated Views + +![](images/1995090b562a5f2e5c76f151b3759663eb18097678d533d91a63cebd8ba3563e.jpg) + +![](images/647779311375c4cefad78cf33c92de0c96a9cd09be881fef60e40a1ddd159f59.jpg) + +![](images/31974e1192941ca1c61071e5a989a01b40a9803c71e3f86f5f4ef5e422933518.jpg) +↓ + +![](images/df2c12c470d93ed3141e164a1fdb26b8e6082bac41a12813d4c4af90b42f8d30.jpg) + +![](images/8bab64fbf22ed1ec5b96317f64fd6a8bebaf358b1326810148b2550ddac2f00f.jpg) +rojection + +![](images/13202d5d5ccfe76283addf07a3390fa08ffef800eecbb72c8b6f2bb1068c4ff0.jpg) +Integrated Mesh + +![](images/03f8bf14dd4285b025192edb6e01b2393473513b8107aa48b250a003d31aad10.jpg) +Annotated Mesh + +![](images/75a3bd599272a1aa50be8b8b95cadd241ecf6f5e41f06e9098ca07e2b37a9f5b.jpg) +3D Segmentation GT +Manual Refinement & Voxelization + +prototype model so that all nominal ones should be identical – and for which an official CAD prototype was available from the manufacturer. The prototype CAD models are depicted in Fig. 3 (1.ii). To simulate various anomalies, we manually introduced different types of defects that represent realistic manufacturing flaws, as illustrated in Fig. 3 (1.iii). We altered the objects' appearance (e.g., paint modifications, Fig. 3 (1.iii) – bottom left), geometry (e.g., dents, Fig. 3 (1.iii) – bottom right), or both (e.g., contaminations, Fig. 3 (1.iii) – top left). It is important to point out that only half of the instances within each object type are modified, while the remaining half are kept in their original state to serve as nominal data. For each object type, one nominal instance will be used as the training set in the real2real experimental setup, while all other ones, both nominal and modified (i.e., anomalous), will make up the test set. This very same test set is also used in the synth2real setup, though the training data comes from CAD prototypes. + +Sensor. We acquired our dataset with a top-tier 3D scanner, the Atos Q from ZEISS (Fig. 3 (2.i)). It consists of a stereo pair with two $12\mathrm{Mpx}$ greyscale cameras and a light projector. We consider the left camera to be our reference. The images have a resolution of $4096\times 3000$ pixels. The point clouds can feature up to 12 million elements, with a point-to-point distance between 0.04 and $0.15\mathrm{mm}$ . The sensor is mounted on an industrial robot arm with high repeatability and precision (Fig. 3 (2.iii)). + +Calibration. We calibrate the greyscale cameras with a + +standard checkerboard pattern, estimating the intrinsic parameters matrix, $A$ , and the radial and tangential distortion coefficients, $\delta = [k_{1}, k_{2}, k_{3}, p_{1}, p_{2}]$ . We also calibrate the rotation, $R_{pc}$ , and translation, $T_{pc}$ , from the point cloud to the left camera reference frame, with a precise dotted calibration pattern built by ZEISS (Fig. 3 (2.ii)) that provides centre coordinates with precision $< 1$ mm. More details are in the supplementary. + +Real data acquisition. The acquisition process consists of 360-degree scanning operations around all object instances of our dataset (see Fig. 3 (3.i)). We define specific viewpoints to capture data from multiple angles around the object. These trajectories are sampled from concentric hemispheres that depend on the object's size. We provide poses relative to a reference view, i.e., $v_{1}$ . Thus, the pose for the view $i$ , $v_{i}$ , consists of a rotation $R_{i}$ and a translation $T_{i}$ that aligns $v_{i}$ to $v_{1}$ , with $R_{1}$ the identity matrix and $T_{1}$ null. Since in real industrial setups object instances would be acquired from as similar as possible viewpoints, we place markers in the scene to facilitate consistent positioning of objects of the same type. Thus, thanks to the high repeatability of the robot arm, images, $I_{i}$ , and point clouds, $P_{i}$ , from the same viewpoint $v_{i}$ turn out similar across acquisitions of different instances of the same object type. However, since the items were positioned manually with the help of the markers, slight displacements cannot be avoided. This introduces some variability across the scans of the instances of an object, as it also typically occurs in real indus + +trial pipelines. For each object, after collecting data from all viewpoints, the ATOS Q sensor also returns an integrated mesh, $M$ , obtained by registering all point clouds from the different views using a proprietary algorithm. We collect and release the mesh to foster the development of methodologies that may also exploit this 3D representation. The 3D reference frame is aligned to the one of the first point cloud $P_{1}$ taken from the first viewpoint $v_{1}$ . + +Synthetic data generation. We employ the Blender [11] Python API to render synthetic data that simulates real acquisitions. To ensure the synthetic and real data are aligned, we select a reference mesh from the collected real instances for each object type, load it in Blender, and align it to the prototype CAD model, as shown in (Fig. 3 (3.ii) - left). Then, we set the Blender camera parameters to the intrinsics, $A$ , of the left camera of our Atos Q sensor and render images from the same viewpoints $v_{i}$ employed during the real acquisitions. The RGB renders are converted into greyscale images to match the real images provided by our acquisition setup (Fig. 3 (3.ii) - right). Renderings have been obtained by exploiting Cycles Printerer, a path-tracing renderer able to provide realistic results with reflections, emissive surfaces, and physically-based lighting. We also render depth maps (Fig. 3 (3.ii) - right), which we project into 3D point clouds (see Fig. 1, centre) using the intrinsic parameters, $A$ . + +Labelling. Since some anomalies, such as paintings and minor scratches, are visible only in greyscale images, while others, such as dents, are better visible in 3D, we must consider both sensing modalities to obtain precise 3D segmentation ground-truths for anomalous samples. Therefore, we develop a two-step labelling strategy, illustrated in Fig. 3 (4), that operates on both 2D images and 3D data. First, we manually annotate all the view images, $I_{i}$ , where anomalies are clearly visible, obtaining dense 2D segmentation masks. We employ a different ID and colour for every defect to ease the voxel AUPRO computation (see Sec. 4). Subsequently, for each annotated image, $I_{i}$ , we leverage the associated integrated mesh, $M$ , the intrinsic parameters $A$ , the point cloud to camera reference frame transformation, $(R_{\mathrm{pc}}, T_{\mathrm{pc}})$ , and the transformation between $v_{i}$ and $v_{1}$ , $(R_{i}, T_{i})$ , to project the 2D annotations onto the integrated mesh. Then, since the projection may not be precise due to occlusions and slanted surfaces, each ground-truth mesh is thoroughly inspected and manually refined using a 3D visualisation and editing software, i.e., CloudCompare [12]. The labelling process was conducted by a team of 4 experts, with a final cross-check of the results. Finally, the annotated mesh is converted into a voxel grid with a voxel size of 2 mm. The choice of the size is determined by the smallest defect size discernible in the meshes. In the conversion process, a voxel is labelled as anomalous if it intersects at least one triangle labelled as such in the annotated mesh. + +
TypeDimensions [cm]Dimensions [voxels]Total InstancesTrain Instances NominalTest Instances NominalAnomalousViews per Instance
Plastic Stool\( {35} \times {35} \times {30} \)\( {330} \times {326} \times {317} \)221 real or 1 synth101012
Rubbish Bin\( {26} \times {21} \times {33} \)\( {333} \times {339} \times {326} \)421 real or 1 synth202012
Wicker Vase\( {17} \times {17} \times {15} \)\( {329} \times {309} \times {290} \)221 real or 1 synth101012
Bathroom Furniture\( {33} \times {33} \times {50} \)\( {371} \times {392} \times {363} \)201 real or 1 synth81036
Container\( {20} \times {25} \times {10} \)\( {327} \times {308} \times {286} \)941 real or 1 synth464612
Plastic Vase\( {12} \times {12} \times 9 \)\( {326} \times {308} \times {286} \)991 real or 1 synth484912
Wooden Stool\( {48} \times {42} \times {45} \)\( {331} \times {367} \times {327} \)151 real or 1 synth6712
Sink Cabinet\( {44} \times {25} \times {50} \)\( {342} \times {349} \times {357} \)191 real or 1 synth9836
+ +Table 2. Dataset statistics. + +# 4. SiM3D Benchmark + +# 4.1. Dataset, Tasks, and Metrics + +Dataset and Setups. Using the procedure described in Sec. 3, we collected 333 object instances belonging to the eight object types, as shown in Tab. 2. We captured from 12 to 36 views for each instance, depending on the type. We split the data into training and test sets. For each type, the training set comprises a single instance, either real in the real2real setup or synthetic in the synth2real setup. The test set includes all other instances and is balanced between nominal and anomalous samples. The test set is the same for both setups. + +Tasks. Our benchmark focuses on the tasks of detecting and segmenting objects' anomalies using multiview and multimodal data. Specifically, the inputs to the methods consist of a set of images $\{I_i\}_{i=1}^n$ captured from different viewpoints alongside 3D data, $\mathbf{X}$ , that capture an object entirely. The 3D data that a method can use are a set of point clouds $\{P_i\}_{i=1}^n$ or depth maps $\{D_i\}_{i=1}^n$ taken from the same vantage points as the images, an integrated mesh $M$ , or also a merged point cloud $P$ . A method capable of multiview multimodal 3D anomaly detection should take these inputs, integrate the 2D and 3D information from multiple views, and produce a global anomaly score for the test sample under inspection. On the other hand, a method pursuing multiview multimodal 3D anomaly segmentation should yield a 3D output, where each 3D position carries a score representing the likelihood of belonging to an anomaly. In our benchmark, a method for this task takes as input all images and 3D data, $(\{I_i\}_{i=1}^n, \mathbf{X})$ to output an Anomaly Volume, i.e., a voxel grid, $V \in \mathbb{R}^{X \times Y \times Z}$ , where $X, Y$ , and $Z$ are the grid dimensions and each voxel carries an anomaly score. For several reasons, we chose voxel grids as the 3D output representation of our benchmark. Firstly, as their grid structure resembles 2D Anomaly Maps, existing 2D segmentation methods and metrics can be easily extended. Moreover, voxel grids are suitable for representing every type of 3D defect, including missing parts. One major challenge with voxel grids pertains to choosing the voxel size. Ideally, we would want a size that allows accurate segmentation of even the smallest defects. However, the memory occupancy of voxel grids scales cubically with the inverse of the voxel size. Thus, we create ground-truth volumes using a voxel resolution of $2\mathrm{mm}$ , which enables precise localisation of + +![](images/e924cd3aa3e0550b385feea033f18ba92ed8d53f140ce5a8bde5fb8367a108fb.jpg) +Figure 4. Extending singleview methods. Anomaly Maps obtained by processing singleview inputs with standard ADS methods are projected into the 3D and aggregated into a voxel grid, with each voxel containing a list of scores from multiple views. The Anomaly Volume is obtained by voxel-wise max pooling. + +even the smallest defects present in our dataset while keeping memory occupancy affordable. + +Metrics. Akin to standard practice in 2D ADS [4], we evaluate detection performance by the Area Under the Receiver Operator Curve (AUROC) of the global anomaly scores predicted by the methods. As we compute a global score for each object instance, we name it instance-level AUROC (I-AUROC). Regarding segmentation performance, we evaluate methods with an adaptation of the popular Area Under the Per-Region Overlap (AUPRO) curve. First of all, we define the voxel-based Per-Region Overlap (PRO) metric as $\mathrm{PRO}_t = \frac{1}{N}\sum_{n = 1}^N\frac{|\overline{\Psi}_t\cap V_n|}{|V_n|}$ , where $\overline{\Psi}_t$ is an Anomaly Volume binarised using a threshold, $t$ , $N$ is the number of blobs, i.e., defects, in the ground-truth, and $V_{n}$ are the voxels belonging to that blob. Since during labelling, we annotated each defect in our dataset with a different ID (see Sec. 3), the $N$ defect blobs in a ground-truth are easily identifiable. This metric is computed across multiple thresholds, $t$ , estimated at different False Positive Rates (FPRs). We calculate the FPR using only non-empty voxels. We sample FPRs from 0 to the maximum value uniformly, we estimate the corresponding threshold $t$ , and we binarise the score grid accordingly. In this way, we construct a curve that plots the PRO values against the corresponding FPRs. The segmentation performance is obtained by integrating such a curve up to a specific False Positive Rate. As proposed in [14], we use $1\%$ as integration bound for the PRO curve to set a challenging bar representative of the requirements of real industrial applications. Finally, we normalise the resulting area to the interval [0, 1]. We dub this metric voxel-level AUPRO (V-AUPRO). + +# 4.2. Baselines + +Extending singleview methods to Multiview 3D ADS. Current ADS methods predict a 2D anomaly map from a singleview input, either unimodal, e.g., images, or multi- + +modal, e.g., images and point clouds. To obtain a set of baselines for our novel multiview 3D ADS task, we extend popular and effective singleview methods by means of the 3D aggregation strategy illustrated in Fig. 4. Given the considered method, we process the data corresponding to an input view, $v_{i}$ , so as to obtain a 2D Anomaly Map. Similarly to Sec. 3, given the mesh, $M$ , intrinsic parameters $A$ , point cloud to camera reference frame transformation, $(R_{\mathrm{pc}}, T_{\mathrm{pc}})$ and view pose $v_{0}$ , $(R_{i}, T_{i})$ , we can project each pixel of the 2D Anomaly Map to its corresponding 3D position. We discretise the 3D position, finding the corresponding voxel index. We associate that voxel with the corresponding anomaly score. We repeat the process for each view, keeping track of the scores that project onto the same voxels. Eventually, we compute the maximum among all scores projected into each voxel, obtaining the final Anomaly Volume. To extract the global score required by the I-AUROC, we compute the maximum across all the scores within the Anomaly Volume. + +Selected Methods. We extend some of the most popular and effective unimodal and multimodal singleview ADS methods, starting from their official code, and evaluate them on SiM3D. Regarding unimodal methods, we select PatchCore [30], with either WideResNet101 [42] or DINOv2 [24] as backbone, to extract features from images of all views, $\{I_i\}_{i=1}^N$ of the single-instance training set to create memory banks. Moreover, we use EfficientAD [1] using all views as their training set. As for multimodal ADS methods, we use BTF [18] employing WideResNet101 [42] as the image encoder and FPFH [33] as the algorithm to extract point cloud features. We create memory banks using features extracted from images, $\{I_i\}_{i=1}^N$ , and point clouds, $\{P_i\}_{i=1}^N$ , of all views of the single-instance training set. We use the same strategy with M3DM [36]. As 2D feature extractor, we exploit DINO-v2. As point cloud feature extractor, instead of using a transformer backbone, i.e., PointMAE [17], we utilise FPFH to extract 3D features, similarly to BTF. Indeed, we would need a severe downsample (from $\sim 5 - 7\mathrm{M}$ to $\sim 8\mathrm{k}$ points) to work with Point-MAE, and we verified that at such low resolution, most of the defects no longer show up in the point clouds. Finally, we select CFM [14], using all views and clouds as a single training set. Based on the same consideration as for M3DM, we use DINO-v2 and FPFH to produce 2D and 3D features. Due to computational constraints, all methods process images padded and downsampled to $1540 \times 1540$ . Moreover, for compatibility with the employed 2D feature extractors, which accept RGB images, we create 3-channel inputs by stacking our greyscale images thrice. + +3D Data pre-processing. To obtain less noisy data and improve the results of the above-mentioned baselines, we remove the background from 3D data. We estimate a 3D plane using a RANSAC-based plane segmentation algo + +![](images/7534f9153be095bf7fd2844e007c4f09313d5a09d32f83ba7fcf455b80669f4a.jpg) +Figure 5. Qualitatives. Selected views with defects (left) and Anomaly Volumes for several methods (right) downsampled for visualisation. + +
MethodModalityreal2realsynth2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.MeanPl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
PatchCore w/ WRN-101RGB0.7400.9870.6360.7770.7740.5511.0000.5660.7540.4620.2350.5370.3530.6780.5760.5170.2500.451
PatchCore w/DINO-v2RGB0.5000.9580.6360.6220.5780.5631.0000.5630.6780.9580.8970.4130.4640.2070.6840.6070.0870.540
EfficientADRGB0.2800.7320.0000.8780.4240.7300.9280.7120.5850.1230.6730.0000.8480.0080.4870.7500.6620.443
ASTRGB0.5370.5660.6110.5150.5730.4960.8390.5370.5841.0000.3441.0000.5250.4820.3880.4280.1870.544
BTFRGB + Point Cloud0.4210.2170.5040.5650.5450.4710.6780.4240.4780.4620.2940.5200.4440.4640.4240.4820.2870.422
CFM w/DINO-v2 + FPFHRGB + Point Cloud0.1980.3010.0740.5150.4830.4560.7320.8250.4480.0080.0000.1900.4240.3130.4590.4280.3620.273
M3DM w/DINO-v2 + FPFHRGB + Point Cloud0.7020.9880.6610.5450.5560.6490.3920.4750.6210.1070.1170.7350.5650.3810.5860.2140.5120.402
ASTRGB + Depth0.9500.9270.7850.4740.5420.4700.4280.9250.6870.6360.0020.5040.4740.4620.5690.4280.8870.495
+ +Table 3. Anomaly detection results (I-AUROC). Best results in bold, runner-ups underlined. + +rithm, which, iteratively, randomly selects a small subset of points (e.g., 10) to fit a candidate plane model. For the candidate planes, we compute the consensus set with threshold $\tau$ on the point-to-plane distance. After 1000 iterations, we return the best-fitting plane. To account for spurious background artefacts, we shift the plane by an offset $\alpha$ and finally remove all points below the shifted plane. Given the different sizes of the objects, for each object type, we select different optimal thresholds $\tau$ and offsets $\alpha$ , as detailed in the supplementary material. We apply this background removal technique to the integrated mesh, $M$ , due to the larger number of background points, which facilitates the RANSAC algorithm. Then, from the background-filtered integrated mesh, we leverage ray-tracing techniques and known camera parameters to render, for each view, depth maps, $\{D_i\}_{i=1}^N$ , pixel-aligned to the corresponding images, $\{I_i\}_{i=1}^N$ at the sensor resolution $(4096 \times 3000)$ . Finally, these depth maps are padded and downsampled to $1540 \times 1540$ , and the depth is lifted in 3D via the camera parameters to obtain dense and clean point clouds. The depth maps and point clouds obtained by rendering the preprocessed meshes serve as the input data for all the experiments presented in the subsequent section and will be available in the SiM3D dataset. + +# 5. Benchmark Results + +Main Results. In Tab. 3 and Tab. 4, we report the results obtained on the SiM3D benchmark by extending the considered methods as Multiview 3D ADS baselines. By looking at the mean detection performance (Tab. 3), we notice how methods that natively process only RGB images outperform natively multimodal ones in both the real2real and synth2real setups. With the notable exception of AST, this also stands for mean segmentation performance (Tab. 4). We posit that this is due to the main multimodal methods having been designed and tuned to process point clouds of the resolution featured by existing multimodal benchmarks, which is much lower than that of SiM3D, i.e. $\sim 8\mathrm{K}$ vs $\sim 2.3\mathrm{M}$ points (after downsampling). Indeed, AST is the only one designed to work natively with depth maps. Hence, our experiments suggest that current multimodal ADS methods do not scale to significantly higher resolutions. In turn, we could not even deploy the native 3D feature extractors of M3DM and CFM, the methods currently excelling on [6, 7]. Furthermore, we notice how methods relying on memory banks, such as PatchCore and M3DM, tend to attain slightly better mean detection and segmentation performance than their counterparts + +
MethodModality
PatchCore w/ WRN-101RGB
PatchCore w/ DINO-v2RGB
EfficientADRGB
ASTRGB
BTFRGB + Point Cloud
CFM w/ DINO-v2 + FPFHRGB + Point Cloud
M3DM w/ DINO-v2 + FPFHRGB + Point Cloud
ASTRGB + Depth
+ +
real2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
0.7100.4610.7610.6750.6940.7470.3800.6090.630
0.7450.4690.7750.7920.7090.7530.4350.6900.671
0.6820.4620.7630.5340.6800.7430.4070.4880.594
0.7200.5020.7890.8030.7160.7640.4460.7580.687
0.5510.4020.7500.3770.6140.7410.0920.0300.444
0.5970.4150.7680.5050.6400.7430.3150.3210.538
0.7330.4520.7670.7020.7020.7520.2880.1260.565
0.7500.5030.7920.8070.7160.7640.4670.7980.699
+ +
pl.Stoolrub.BinW.VaseB.Furn.Cont.Pl.VaseW.StoolSink Cab.Mean
0.7340.4370.7510.4540.6780.7410.3860.6180.600
0.7010.4570.7720.5630.6480.7340.3210.5750.596
0.6800.4490.7290.5230.6690.7470.4310.3610.573
0.7260.4880.7910.8060.7050.7640.4610.6950.679
0.5470.4020.7560.3690.6100.7410.1030.0400.446
0.5230.4130.7530.5420.6210.7410.2220.3140.516
0.6900.4540.7700.4610.6450.7340.3180.1320.526
0.7510.5020.7930.8050.7170.7640.4500.7850.695
+ +Table 4. Anomaly segmentation results (V-AUPRO@1%). Best results in bold, runner-ups underlined. + +
MethodModalityFeatures
2D3D
PatchCoreRGBWRN-101-
RGBDINO-v2-
Depth-WRN-101
Depth-DINO-v2
Point Cloud-FPFH
BTFRGB + Point CloudWRN-101FPFH
RGB + DepthWRN-101WRN-101
CFMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
M3DMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
ASTRGBEffNet-B5-
RGB + DepthEffNet-B5EffNet-B5
+ +
real2realsynth2real
DetectionSegmentationDetectionSegmentation
0.7540.6300.4510.600
0.6780.6710.5400.596
0.5820.4860.5070.484
0.6000.4960.2460.471
0.4150.4640.3130.460
0.4780.4440.4220.446
0.7070.6090.4480.543
0.4480.5380.2730.516
0.5480.6630.3110.620
0.6210.5650.4020.526
0.6590.5030.2540.476
0.5840.6870.5440.679
0.6870.6990.4950.695
+ +Table 5. Anomaly detection and segmentation mean results with different input modalities. Best results per method in bold. + +within the same modality that require substantial training, i.e., EfficientAD and CFM. We argue that this may be ascribed to the single-instance setup, which renders the latter category of approaches harder to optimise due to the limited number of training images, which are as many as the views. The methods most effectively extended to our multiview 3D task are Patchcore, which, equipped with either WRN-101 or DINO-v2, can achieve the best detection in both setups, and AST, achieving the best segmentation performance in both setups. Indeed, Patchcore yields I-AUROCs of 0.754 (WRN-101) and 0.540 (DINO-v2), while AST yields a V-AUPRO@1% of 0.699 and 0.695, in the real2real and synth2real setups, respectively. It is also worth pointing out how, due to the domain shift, we can notice a gap between the real2real and synth2real setups, especially in detection (0.754 vs. 0.540, I-AUROC) compared to segmentation performance (0.699 vs. 0.695, V-AUPRO@1%). The qualitative results, shown in Fig. 5, visually support the above observations. Overall, the relatively low, definitely far from saturated, performance figures provided by baselines designed by straightforward extension of state-of-the-art methods delivering impressive results in singleview 2D benchmarks hint at the novel and challenging nature of the task set forth by SiM3D. Thus, our experimental findings by baseline approaches reveal the need for more advanced and specific methods and paradigms, both in the real2real and, even more so, the synth2real setups. + +Alternative 3D Representations. Given the lack of feature extractors amenable to processing high-resolution point clouds, we explore the use of depth maps to capture 3D + +cues. The 2D grid structure of depth maps enables the deployment of popular backbones pre-trained on RGB data that support high-resolution inputs, such as DINOv2. Thus, we create and evaluate additional baselines using depth maps and report the results in Tab. 5. Similarly to greyscale images, we tile depth maps along the channel axis to ensure compatibility with the employed backbones. The results show that the use of depth maps rather than point clouds in multimodal methods tends to improve detection and segmentation performance by sizable margins, e.g., in the real2real setup, BTF yields 0.707 I-AUROC and 0.609 V-AUPRO when processing depths compared to 0.478 I-AUROC and 0.444 V-AUPRO when fed with point clouds (row 6 vs. row 7) while relying on the same WRN101 image backbone. Indeed, all multimodal methods but M3DM achieve their best performance in both tasks and setups by processing depth maps instead of point clouds. Yet, as vouched by the results achieved by PatchCore, RGB images compare favourably to depth maps in terms of detection performance. We ascribe this to the feature extractors used to compute depth features being pre-trained on RGB images, which may be conducive to yielding weaker features when fed with depth inputs. This points out an intriguing research question: would it be possible to develop a foundation model to process depth maps effectively? + +# 6. Final Discussion + +We introduced SiM3D, the first benchmark and dataset focused on integrating multiview and multimodal information for comprehensive 3D ADS. Our experiments highlight several open challenges. First, a need for methods that can process multiview inputs more effectively, e.g., by taking all views as input rather than integrating singleview outputs, and that can be trained with a very limited number of nominal samples, e.g., a single instance. Second, training with synthetic data requires more robust techniques to address domain-shift challenges. Finally, the current literature lacks 3D backbones capable of processing high-resolution point clouds. Employing depth maps seems a promising strategy to facilitate this processing, yet a foundational model for depth maps is still missing. We believe that SiM3D may serve as a strong foundation for tackling these challenges and fostering future research toward 3D ADS. + +# Acknowledgment + +We acknowledge financial support from the National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.1, Call No. 1409 published on 14 September 2022 by the Italian Ministry of University and Research (MUR), funded by the European Union. + +This work was also partially supported by the FSE+ 2021-2027 program under Article 24, paragraph 3, letter a) of Law 240/2010 and its subsequent amendments, as well as by Regional Resolution No. 693/2023 (Project Code: 2023-20090/ER), Grant CUP: J19J23000730002. + +We sincerely thank Lucia Gasperini for her contribution to dataset acquisition and labelling. We also extend our gratitude to Manuel Turrini for his support in setting up the acquisition process. + +# References + +[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficient-tad: Accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 128-138, 2024. 3, 6 +[2] Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection. arXiv preprint arXiv:2002.10445, 2020.3 +[3] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3 +[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6 +[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130, 2022. 3 +[6] Paul Bergmann, Jin Xin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 202-213, 2022. 1, 2, 3, 7 +[7] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the 16th Asian Conference on Computer Vision (ACCV2022), 2022. ACCV. 1, 2, 3, 7 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3 + +[9] Li-Ling Chiu and Shang-Hong Lai. Self-supervised normalizing flows for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2926-2935, 2023. 3 +[10] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. ArXiv, 2020. 3 +[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2024. 5 +[12] CloudCompare Online Community. CloudCompare - 3D point cloud and mesh processing software Open Source Project. EDF, 2024. 5 +[13] Alex Costanzino, Pierluigi Zama Ramirez, Mirko Del Moro, Agostino Aiezzo, Giuseppe Lisanti, Samuele Salti, and Luigi Di Stefano. Test time training for industrial anomaly segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3910-3920, 2024. 3 +[14] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. CVPR. 1, 3, 6 +[15] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 3 +[16] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 98-107, 2022. 3 +[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 6 +[18] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2976, 2023. 3, 6 +[19] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 3 +[20] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 3 +[21] Jiaqi Liu, Guoyang Xie, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, Feng Zheng, et al. Real3d-ad: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3 + +[22] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv preprint arXiv:2301.11514, 2, 2023. 3 +[23] Fabio Valerio Massoli, Fabrizio Falchi, Alperen Kantarci, Şeymanur Akti, Hazim Kemal Ekenel, and Giuseppe Amato. Cocca: Multilayer one-class classification for anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 33(6):2313-2323, 2021. 3 +[24] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6 +[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 604-621. Springer, 2022. 3 +[26] Jonathan Pinnay and Keng Chai. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing, pages 394-406. Springer, 2022. 3 +[27] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 3 +[28] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733. IEEE, 2021. +[29] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13576-13586, 2022. 3 +[30] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of 2022 IEEE Conference on Computer Vision and Pattern Recognition, pages 14298-14308, 2022. 1, 3, 6 +[31] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907-1916, 2021. 3 +[32] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3 + +[33] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE International Conference on Robotics and Automation, pages 3212-3217, 2009. 3, 6 +[34] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. In International Conference on Learning Representations, 2021. 3 +[35] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22883-22892, 2024. 1, 3 +[36] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 1, 3, 6 +[37] Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650-656, 2022. 3 +[38] Minghui Yang, Peng Wu, and Hui Feng. Memseg: A semi-supervised method for image surface defect detection using differences and commonalities. Engineering Applications of Artificial Intelligence, 119:105835, 2023. +[39] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision, 2020. +[40] Seungdong Yoa, Seungjun Lee, Chiyoon Kim, and Hyunwoo J Kim. Self-supervised learning for anomaly detection with dynamic local augmentation. IEEE Access, 9:147201-147211, 2021. +[41] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 3 +[42] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6 +[43] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3 +[44] Zheng Zhang and Xiaogang Deng. Anomaly detection using improved deep svdd model with data structure preservation. Pattern Recognition Letters, 148:1-6, 2021. 3 +[45] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection. arXiv preprint arXiv:2310.07716, 2023. 1, 3 +[46] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 3 \ No newline at end of file diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/images.zip b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21a03389c5b450c90c9f07c11478a02d7c040e1b --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:725a4c0180945f385c2d9a60734ad5963a3b87e52297d3c4ac0db11804552ac6 +size 589656 diff --git a/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/layout.json b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4895eb1f2a5a29586e68e7a128946c13083dc7f6 --- /dev/null +++ b/2025/SiM3D_ Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark/layout.json @@ -0,0 +1,8697 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "SiM3D: Single-instance Multiview Multimodal and Multisetup 3D Anomaly Detection Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 146, + 556, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 556, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 556, + 204 + ], + "type": "text", + "content": "Alex Costanzino1 Pierluigi Zama Ramirez1 Luigi Lella1 Matteo Ragaglia2 Alessandro Oliva2 Giuseppe Lisanti1 Luigi Di Stefano1 \n1CVLab, University of Bologna, Italy 2SACMI Imola, Italy https://alex-costanzino.github.io/SiM3D/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 228 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 241, + 297, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 297, + 482 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 297, + 482 + ], + "type": "text", + "content": "We propose SiM3D, the first benchmark considering the integration of multiview and multimodal information for comprehensive 3D anomaly detection and segmentation (ADS), where the task is to produce a voxel-based Anomaly Volume. Moreover, SiM3D focuses on a scenario of high interest in manufacturing: single-instance anomaly detection, where only one object, either real or synthetic, is available for training. In this respect, SiM3D stands out as the first ADS benchmark that addresses the challenge of generalising from synthetic training data to real test data. SiM3D includes a novel multimodal multiview dataset acquired using top-tier industrial sensors and robots. The dataset features multiview high-resolution images (12 Mpx) and point clouds (" + }, + { + "bbox": [ + 54, + 241, + 297, + 482 + ], + "type": "inline_equation", + "content": "\\sim 7M" + }, + { + "bbox": [ + 54, + 241, + 297, + 482 + ], + "type": "text", + "content": " points) for 333 instances of eight types of objects, alongside a CAD model for each type. We also provide manually annotated 3D segmentation GTs for anomalous test samples. To establish reference baselines for the proposed multiview 3D ADS task, we adapt prominent singleview methods and assess their performance using novel metrics that operate on Anomaly Volumes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "In computer vision, anomaly detection and segmentation (ADS) involves identifying anomalous samples and localising their defects. This task is challenging in industrial contexts due to the variability and rarity of defects. Typically, ADS in the industry follows a cold-start approach, where training is unsupervised and uses only defect-free (nominal) samples. Before the release of the MVTec AD dataset in 2019 [4], only a limited number of ADS methods were proposed, with just 9 publications in major computer vision conferences (e.g., CVPR, ECCV, ICCV, WACV) over the five years before. The introduction of MVTec AD marked a pivotal moment in the field, fostering the publication of roughly 172 papers in these venues over the next 5 years. This dataset not only stimulated the development of several ADS methods [14, 30, 36] but also promoted the creation of novel and more challenging benchmarks. For instance," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 334, + 215, + 527, + 550 + ], + "blocks": [ + { + "bbox": [ + 334, + 215, + 527, + 550 + ], + "lines": [ + { + "bbox": [ + 334, + 215, + 527, + 550 + ], + "spans": [ + { + "bbox": [ + 334, + 215, + 527, + 550 + ], + "type": "image", + "image_path": "ced18cd97e949e4bffadc3884fba9fc61b90ce90b5a86edc90653bbcdc06c372.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 556, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 590 + ], + "type": "text", + "content": "Figure 1. SiM3D dataset overview. From top to bottom: the single-instance real and synthetic training samples for object Plastic Stool, one of the anomalous samples from the test set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 687 + ], + "type": "text", + "content": "MVTec 3D-AD and [6] and Eyecandies [7] introduced the first benchmarks deploying multimodal inputs to enhance ADS performance. Then, PAD [45] and Real-IAD [35] introduced novel benchmarks that leverage multiple RGB views of objects. Although these benchmarks tackle various settings, two real industrial issues remain unaddressed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "First, existing benchmarks have primarily addressed the problem of 2D ADS. Consequently, current methods gen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20944" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "type": "table", + "html": "
DatasetVenue & Year2DNo. Pixels3DNo. PointsScenarioTrainTestSetupTaskNo. Classes
MVTec ADCVPR 2019RGB Images490k - 1M——Multi-instanceSingleviewSingleviewreal2real2D15
MVTec 3D-ADVISAPP 2021RGB Images160k - 810kXYZ Images10k - 195kMulti-instanceSingleviewSingleviewreal2real2D10
EyecandiesACCV 2022RGB Images262kDepths + Normals262kMulti-instanceSingleviewSingleviewsynth2synth2D10
MVTec LOCO ADIJCV 2022RGB Images1.28M - 2.05M——Multi-instanceSingleviewSingleviewreal2real2D5
VisAECCV 2022RGB Images1.5M——Multi-instanceSingleviewSingleviewreal2real2D12
PADNeurIPS 2023RGB Images12M——Single-instanceMultiviewSingleviewreal2real + synth2synth2D20
Real3D-ADNeurIPS 2023——Point Clouds43k - 2.68MMulti-instanceMultiviewSingleviewreal2real3D12
Real-IADCVPR 2024RGB Images10M——Multi-instanceSingleviewSingleviewreal2real2D30
SiM3D2025Grayscale Images12MPoint Clouds + Mesh5M - 7MSingle-instanceMultiviewMultiviewreal2real + synth2real3D8
", + "image_path": "1f1b52d32266d9aa1f3a6d001edefd1b6857ed86f2d3617363c88e98b462b831.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 151, + 550, + 162 + ], + "lines": [ + { + "bbox": [ + 57, + 151, + 550, + 162 + ], + "spans": [ + { + "bbox": [ + 57, + 151, + 550, + 162 + ], + "type": "text", + "content": "Table 1. Popular ADS benchmarks. Chronologically sorted, the sequence hints at a growing interest in multimodal and multiview data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 182, + 294, + 255 + ], + "blocks": [ + { + "bbox": [ + 56, + 182, + 294, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 182, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 294, + 255 + ], + "type": "image", + "image_path": "d25cb59f2febfc8be8552ba1c6a8346ef4dad0420ec11fa46f90073270c39a42.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "type": "text", + "content": "Figure 2. 2D vs. 3D anomaly detection and segmentation. Typical ADS benchmarks rely on 2D Anomaly Maps and 2D ground-truths (left) to assess segmentation performance. SiM3D proposes to do so by 3D Anomaly Volumes and 3D ground-truths (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 329, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 295, + 581 + ], + "type": "text", + "content": "erate Anomaly Maps, i.e., images that encode the likelihood of each pixel belonging to a defect, as shown in Fig. 2 (left). However, many industrial applications require precise localisation of defects in the 3D space to allow automatic intervention, reducing waste and optimising production. Benchmarks relying on multimodal input data, like those proposed in MVTec 3D-AD [6] and Eyecandies [7], represent progress toward this goal by providing 3D information for each pixel within an image. However, in these benchmarks, ADS methods are still expected to provide as output a 2D Anomaly Map, and objects are inspected from a single viewpoint. For comprehensive 3D anomaly detection, we must observe an object in its entirety, utilising captures from multiple viewpoints to detect defects across the entire 3D space. This calls for methods capable of integrating multiview information to estimate the defect probability at each 3D location within an Anomaly Volume, as shown in Fig. 2 (right). Furthermore, as demonstrated by MVTec 3D-AD and Eyecandies, these methods should process multimodal information, i.e., RGB images as well as 3D information, to achieve more reliable anomaly detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Secondly, current benchmarks typically provide multiple training samples per object class to capture the variability across nominal objects. However, collecting many samples - even nominal ones - can be costly and challenging. For example, line changeovers would demand acquiring new datasets and retraining models, both highly time-intensive processes. Additionally, collecting large datasets increases the risk of errors, such as misjudged nominal samples, which may degrade model performance. Moreover, manufactured objects typically closely resemble one another as they replicate the structure and materials of a syn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 184, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 232 + ], + "type": "text", + "content": "thetic CAD prototype. Thus, the variability across nominal samples of manufactured objects is frequently minimal, with a single nominal item subsuming the information required to spot anomalies in other objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 232, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 365 + ], + "type": "text", + "content": "These considerations highlight the need for ADS techniques that may be effectively trained by a single nominal sample or even a synthetic prototype (e.g., an object's CAD model) and then be able to generalise to other manufactured objects to identify anomalous ones effectively. The ability to perform training by a single object instance, either real or synthetic, would avoid the need for extensive data collection, vastly facilitating the creation and adaptation of ADS models. Yet, most current benchmarks focus on multiple training instances setups, and none of the prior works consider challenging synthetic (training) to real (test) setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 365, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 617 + ], + "type": "text", + "content": "To fill all the gaps highlighted by our considerations and stimulate interest in the development of multimodal and multiview 3D ADS methods, we propose the Single-instance Multiview Multimodal Multisetup dataset, dubbed SiM3D. The dataset contains a series of high-resolution scans of manufactured objects, which can be nominal or anomalous. Each scan consists of multiple views, each featuring a greyscale image, its corresponding point cloud (i.e., multimodal data), and the ground-truth acquisition pose. In addition, an integrated 3D mesh is provided with each scan. For each object, a single nominal sample has been retained to be employed as a single-instance training set, while the remaining samples, both nominal and anomalous, compose the test set. Moreover, to foster research on novel approaches that may address the domain gap between synthetic training data and real test samples in ADS, for each object of our dataset, we also provide a synthetic scan, i.e., we deploy the available CAD model to obtain the same kind of data as in our high-resolution real scans, with images and point clouds rendered from the same poses. An overview of the proposed dataset is depicted in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Based on our novel dataset, we create the first benchmark for 3D anomaly detection, where the task is to predict anomaly scores within a voxelized 3D Anomaly Volume rather than a 2D Anomaly Map (Fig. 2). Peculiarly, our benchmark includes two setups based on the domain of the data available for training: in the real2real setup, a model is given access to a single nominal instance of an object (Fig. 1, top), while in synth2real the same kind" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20945" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "of training data are obtained from the CAD model of the object (Fig. 1, centre). We adapt prominent ADS methods – both unimodal and multimodal – for the proposed task and evaluate them on our benchmark. To assess segmentation performance, we provide accurate 3D ground-truths (Fig. 2) and extend ADS metrics to operate on anomaly scores and ground-truths specified on 3D voxel grids. Our experimental findings highlight the open challenges in the field and offer insights into possible future directions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 142, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 295, + 617 + ], + "type": "text", + "content": "The MVTec AD [4] dataset significantly boosted anomaly detection research, inspiring new techniques and benchmarks for various challenges. MVTec AD provides a training set containing only nominal images and a test set that includes nominal and anomalous ones for several object categories. This unsupervised setup exemplifies many real-world applications in which anomalous objects are challenging to collect, and thus, it has also been adopted by more recent ADS benchmarks. Among them, MVTec LOCO [5] presents a challenging scenario with both structural anomalies (e.g., dents, holes) and logical anomalies (e.g., incorrect object combination). VisA [46] introduced high-resolution images of complex scenes containing multiple instances of the same object. However, certain defects may be visible only by analysing objects' 3D geometry. For this reason, multimodal datasets, such as MVTec 3D-AD [6] and Eyecandies [7] include 3D data alongside images. For each sample in the dataset, MVTec 3D-AD provides images and pixel-aligned XYZ information captured by a 3D sensor, while Eyecandies contains synthetic images with pixel-aligned depths and normals. Recently, PAD [45] addressed pose-agnostic ADS with the first multiview dataset (MAD), featuring images of objects from various viewpoints. In PAD, training samples are multiview RGB images with known poses, while test samples are singleview RGB images with unknown poses. MAD includes both synthetic and real objects, but its benchmarks do not explore synthetic-to-real generalisation. Real-IAD [35] introduced a large-scale multiview dataset with RGB images covering various defect sizes. Eventually, Real3D-AD [21] is the first point cloud anomaly detection benchmark where the task is to predict an anomaly score for each point. It uses front-and-back scans of nominal objects for training and a singleview point cloud for testing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1, the proposed SiM3D benchmark offers several outstanding features, including very-high-resolution images (12 Mpx), large point clouds/meshes containing 5-7 million elements captured with high-precision 3D sensors. Moreover, it is designed for a highly relevant manufacturing scenario: single-instance anomaly detection, where only one object is available for model training. To promote research toward even more cost-effective scenar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "ios, ours is also the only benchmark that addresses the challenge of generalising from synthetic training data to real test data. Finally, SiM3D stands out as the first benchmark that considers integrating multiview information at test time for comprehensive 3D anomaly detection, where the task consists of outputting a voxel-based Anomaly Volume." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 553, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 610 + ], + "type": "text", + "content": "ADS Methods. Following the setup defined in the MVTec-AD benchmark, most ADS methods [22] are designed to estimate 2D anomaly maps from a single image [2, 3, 9, 10, 13, 15, 16, 19, 20, 23, 26, 27, 27-29, 31, 34, 37-41, 43, 44]. One of the most popular methods is PatchCore [30], which uses feature extractors pre-trained on large datasets [8, 17, 24] to gather features from nominal samples into a memory bank. At test time, input sample features are compared to this bank to detect anomalies. Another more recent yet popular approach, EfficientAD [1], uses a lightweight feature extractor and a student-teacher model to identify anomalies based on feature prediction discrepancy. The introduction of the MVTec 3D-AD benchmark has led to the development of several multimodal approaches that use images and 3D data to generate 2D anomaly maps [14, 18, 32, 36]. Among them, BTF [18] is a memory bank approach inspired by PatchCore, in which each element is the concatenation of 2D and 3D features extracted by a pre-trained image backbone and a traditional point cloud descriptor, i.e., FPFH [33], respectively. M3DM [36] improves BTF using powerful transformer-based pre-trained backbones, such as DINO-v1[8] and Point-MAE [25], to extract features from RGB images and point clouds. It also learns to fuse 2D and 3D features into multimodal features. During inference, a trained One-Class SVM learns to aggregate modality-specific scores. AST [32] follows a Teacher-Student paradigm, in which the Teacher leverages Normalizing Flows to model the multimodal features distribution, while the Student employs a feed-forward network to model such distribution. This asymmetry in the architectures exacerbates the discrepancies between nominal and anomalous features at inference time. Recently, starting from the same transformer-based backbones as in [36], CFM [14] learns to map features from one modality to the other by nominal samples and then, at inference time, detects anomalies by pinpointing inconsistencies between observed and predicted features. In this paper, we employ the above-mentioned methods to investigate the challenges of SiM3D by adapting them to work in a multiview setup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 621, + 416, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 416, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 416, + 634 + ], + "type": "text", + "content": "3. Dataset Creation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Material preparation. We scanned eight types of manufactured objects spanning different sizes, shapes, textures and materials, as shown in Fig. 3 (1.i). For each object type, between 20 and 100 instances were purchased from a store. We selected objects suitable for single-instance scenarios – instances are manufactured according to a unique" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20946" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 90, + 93, + 118 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 122, + 79 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 122, + 79 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 122, + 79 + ], + "type": "text", + "content": "(1) Material preparation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 90, + 93, + 118 + ], + "lines": [ + { + "bbox": [ + 65, + 90, + 93, + 118 + ], + "spans": [ + { + "bbox": [ + 65, + 90, + 93, + 118 + ], + "type": "image", + "image_path": "1ff934651d17324fe9b5aa23751bafe5d029740d434cd7fc88f931420a69846c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 65, + 122, + 93, + 140 + ], + "blocks": [ + { + "bbox": [ + 65, + 122, + 93, + 140 + ], + "lines": [ + { + "bbox": [ + 65, + 122, + 93, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 122, + 93, + 140 + ], + "type": "image", + "image_path": "e4f1f8ee071598b7c23b6deaa553ee65f938240487e0bc84a408e515c28eb0d1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 94, + 90, + 124, + 142 + ], + "blocks": [ + { + "bbox": [ + 65, + 82, + 114, + 89 + ], + "lines": [ + { + "bbox": [ + 65, + 82, + 114, + 89 + ], + "spans": [ + { + "bbox": [ + 65, + 82, + 114, + 89 + ], + "type": "text", + "content": "(i) Real objects purchase" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 90, + 124, + 142 + ], + "lines": [ + { + "bbox": [ + 94, + 90, + 124, + 142 + ], + "spans": [ + { + "bbox": [ + 94, + 90, + 124, + 142 + ], + "type": "image", + "image_path": "5b7bdb6910b4eefc26a97782b5754dafb421bb835adeaac43ff366ce7d78ab29.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 129, + 90, + 156, + 118 + ], + "blocks": [ + { + "bbox": [ + 129, + 90, + 156, + 118 + ], + "lines": [ + { + "bbox": [ + 129, + 90, + 156, + 118 + ], + "spans": [ + { + "bbox": [ + 129, + 90, + 156, + 118 + ], + "type": "image", + "image_path": "7e45da2c68f71f6644c49964ab4dac5970ffc875760a72c14c35fa6db6ae0bc4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 129, + 118, + 156, + 144 + ], + "blocks": [ + { + "bbox": [ + 129, + 118, + 156, + 144 + ], + "lines": [ + { + "bbox": [ + 129, + 118, + 156, + 144 + ], + "spans": [ + { + "bbox": [ + 129, + 118, + 156, + 144 + ], + "type": "image", + "image_path": "e11ccfd440b9929915ea1741545982a25ca45086c92e3529f6be5e0dbe540465.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 162, + 90, + 186, + 143 + ], + "blocks": [ + { + "bbox": [ + 162, + 90, + 186, + 143 + ], + "lines": [ + { + "bbox": [ + 162, + 90, + 186, + 143 + ], + "spans": [ + { + "bbox": [ + 162, + 90, + 186, + 143 + ], + "type": "image", + "image_path": "f23dd7682ba3295026491b4676d9e3ddcfcc8a7ad8248066e0e62abb7d01618b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 72, + 252, + 79 + ], + "lines": [ + { + "bbox": [ + 200, + 72, + 252, + 79 + ], + "spans": [ + { + "bbox": [ + 200, + 72, + 252, + 79 + ], + "type": "text", + "content": "(2) Acquisition setup" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 212, + 90, + 290, + 159 + ], + "blocks": [ + { + "bbox": [ + 212, + 90, + 290, + 159 + ], + "lines": [ + { + "bbox": [ + 212, + 90, + 290, + 159 + ], + "spans": [ + { + "bbox": [ + 212, + 90, + 290, + 159 + ], + "type": "image", + "image_path": "68d138f94156275eddb1246fa993dd2a655fb34b12979bf4a4f3f77c9b7dbc52.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 65, + 159, + 96, + 185 + ], + "blocks": [ + { + "bbox": [ + 65, + 152, + 119, + 157 + ], + "lines": [ + { + "bbox": [ + 65, + 152, + 119, + 157 + ], + "spans": [ + { + "bbox": [ + 65, + 152, + 119, + 157 + ], + "type": "text", + "content": "(ii) CAD models collection" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 159, + 96, + 185 + ], + "lines": [ + { + "bbox": [ + 65, + 159, + 96, + 185 + ], + "spans": [ + { + "bbox": [ + 65, + 159, + 96, + 185 + ], + "type": "image", + "image_path": "e75cea0664333c347b6a442fd81cd16cbbe14183f9521011ec893709d1da3031.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 65, + 185, + 96, + 213 + ], + "blocks": [ + { + "bbox": [ + 65, + 185, + 96, + 213 + ], + "lines": [ + { + "bbox": [ + 65, + 185, + 96, + 213 + ], + "spans": [ + { + "bbox": [ + 65, + 185, + 96, + 213 + ], + "type": "image", + "image_path": "94ac0ef751c1158402051d318db7692407fa45096236e47ccd96390f6c340b6a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 96, + 159, + 127, + 185 + ], + "blocks": [ + { + "bbox": [ + 96, + 159, + 127, + 185 + ], + "lines": [ + { + "bbox": [ + 96, + 159, + 127, + 185 + ], + "spans": [ + { + "bbox": [ + 96, + 159, + 127, + 185 + ], + "type": "image", + "image_path": "129621fc3da7cc9ae5d5ba7cd65cfa2ad4c3eb37afab97fa42472eb3c708e3e2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 96, + 185, + 127, + 213 + ], + "blocks": [ + { + "bbox": [ + 96, + 185, + 127, + 213 + ], + "lines": [ + { + "bbox": [ + 96, + 185, + 127, + 213 + ], + "spans": [ + { + "bbox": [ + 96, + 185, + 127, + 213 + ], + "type": "image", + "image_path": "f447c1e84704c700eb2bc7c1f6612b56f8d761da53a895692c7195a285176bba.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 128, + 159, + 157, + 185 + ], + "blocks": [ + { + "bbox": [ + 128, + 159, + 157, + 185 + ], + "lines": [ + { + "bbox": [ + 128, + 159, + 157, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 159, + 157, + 185 + ], + "type": "image", + "image_path": "b42dfe144bf9eda909a083b7c16f5a57ca291e7ed48e1339d60393f3324fde33.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 128, + 185, + 157, + 213 + ], + "blocks": [ + { + "bbox": [ + 128, + 185, + 157, + 213 + ], + "lines": [ + { + "bbox": [ + 128, + 185, + 157, + 213 + ], + "spans": [ + { + "bbox": [ + 128, + 185, + 157, + 213 + ], + "type": "image", + "image_path": "0f7f7e34f26bb7aa2a32e98ebe4930be31cd99f3d293d4384fbcdfe52f84d551.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 159, + 159, + 190, + 185 + ], + "blocks": [ + { + "bbox": [ + 159, + 159, + 190, + 185 + ], + "lines": [ + { + "bbox": [ + 159, + 159, + 190, + 185 + ], + "spans": [ + { + "bbox": [ + 159, + 159, + 190, + 185 + ], + "type": "image", + "image_path": "efc0722f016b683c0b42077e36695c29d0c07cf3635ea7aa90ca660b190ff892.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 159, + 185, + 190, + 213 + ], + "blocks": [ + { + "bbox": [ + 159, + 185, + 190, + 213 + ], + "lines": [ + { + "bbox": [ + 159, + 185, + 190, + 213 + ], + "spans": [ + { + "bbox": [ + 159, + 185, + 190, + 213 + ], + "type": "image", + "image_path": "874b0a10fcfaf19188adc278c8d6710ad5d2cc9ea8fd419f10d22092ac09f4aa.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 205, + 167, + 258, + 208 + ], + "blocks": [ + { + "bbox": [ + 205, + 167, + 258, + 208 + ], + "lines": [ + { + "bbox": [ + 205, + 167, + 258, + 208 + ], + "spans": [ + { + "bbox": [ + 205, + 167, + 258, + 208 + ], + "type": "image", + "image_path": "7195463304f6052647ece2b6c83c1c9c8dbe6042e79ebc3274030437453fe7dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 178, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 261, + 178, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 261, + 178, + 296, + 198 + ], + "type": "text", + "content": "High-precision calibration pattern \nUnder 1mm precise corners" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 236, + 213, + 287, + 281 + ], + "blocks": [ + { + "bbox": [ + 206, + 239, + 230, + 256 + ], + "lines": [ + { + "bbox": [ + 206, + 239, + 230, + 256 + ], + "spans": [ + { + "bbox": [ + 206, + 239, + 230, + 256 + ], + "type": "text", + "content": "Top-tier robot arm Precise poses" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 213, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 236, + 213, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 236, + 213, + 287, + 281 + ], + "type": "image", + "image_path": "ac3d3865a19941ae2db2a0f602279d7d60b12c2786fcc8708b6ed2adeb029e26.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 319, + 90, + 404, + 137 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 355, + 79 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 355, + 79 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 355, + 79 + ], + "type": "text", + "content": "(3) Data gathering" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 82, + 361, + 87 + ], + "lines": [ + { + "bbox": [ + 314, + 82, + 361, + 87 + ], + "spans": [ + { + "bbox": [ + 314, + 82, + 361, + 87 + ], + "type": "text", + "content": "(i) Real data acquisition" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 90, + 404, + 137 + ], + "lines": [ + { + "bbox": [ + 319, + 90, + 404, + 137 + ], + "spans": [ + { + "bbox": [ + 319, + 90, + 404, + 137 + ], + "type": "image", + "image_path": "9a2f3f8f307dd9a819dcc73db2d4228a24f502f46f38d2d362578ebf0f1a3aa7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 353, + 137, + 378, + 157 + ], + "blocks": [ + { + "bbox": [ + 353, + 137, + 378, + 157 + ], + "lines": [ + { + "bbox": [ + 353, + 137, + 378, + 157 + ], + "spans": [ + { + "bbox": [ + 353, + 137, + 378, + 157 + ], + "type": "image", + "image_path": "a02b50d078fe3edafe71d0ab3638bc5c5bddd3542a3b9a2437eb34e09a5b26d5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 317, + 159, + 339, + 183 + ], + "blocks": [ + { + "bbox": [ + 317, + 159, + 339, + 183 + ], + "lines": [ + { + "bbox": [ + 317, + 159, + 339, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 339, + 183 + ], + "type": "image", + "image_path": "9e3a0e084876fbf53d19b7af1fc8eed4d51a70b6a38e4a1c13de2bba21ec05bd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 184, + 334, + 189 + ], + "lines": [ + { + "bbox": [ + 318, + 184, + 334, + 189 + ], + "spans": [ + { + "bbox": [ + 318, + 184, + 334, + 189 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 339, + 159, + 359, + 183 + ], + "blocks": [ + { + "bbox": [ + 339, + 159, + 359, + 183 + ], + "lines": [ + { + "bbox": [ + 339, + 159, + 359, + 183 + ], + "spans": [ + { + "bbox": [ + 339, + 159, + 359, + 183 + ], + "type": "image", + "image_path": "a49ce4f839b822f2dc49b37e32f3383ca19abf576328ad915a9ceaae867d796a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 359, + 165, + 375, + 183 + ], + "blocks": [ + { + "bbox": [ + 359, + 165, + 375, + 183 + ], + "lines": [ + { + "bbox": [ + 359, + 165, + 375, + 183 + ], + "spans": [ + { + "bbox": [ + 359, + 165, + 375, + 183 + ], + "type": "image", + "image_path": "157ea8aba7041065075008065c61a237319e4a144353b0fe90c9b39e9e2b493a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 318, + 190, + 341, + 210 + ], + "blocks": [ + { + "bbox": [ + 318, + 190, + 341, + 210 + ], + "lines": [ + { + "bbox": [ + 318, + 190, + 341, + 210 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 341, + 210 + ], + "type": "image", + "image_path": "edd9316479c3794334faca86caa59d0c0844e9915c29d60895499be4c73f9a33.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 342, + 190, + 376, + 211 + ], + "blocks": [ + { + "bbox": [ + 359, + 184, + 374, + 190 + ], + "lines": [ + { + "bbox": [ + 359, + 184, + 374, + 190 + ], + "spans": [ + { + "bbox": [ + 359, + 184, + 374, + 190 + ], + "type": "text", + "content": "View N" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 190, + 376, + 211 + ], + "lines": [ + { + "bbox": [ + 342, + 190, + 376, + 211 + ], + "spans": [ + { + "bbox": [ + 342, + 190, + 376, + 211 + ], + "type": "image", + "image_path": "17ece7b18ebb80ba195c10661735b761b558fa0a34f53240737470146282244c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 213, + 371, + 219 + ], + "lines": [ + { + "bbox": [ + 310, + 213, + 371, + 219 + ], + "spans": [ + { + "bbox": [ + 310, + 213, + 371, + 219 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 384, + 169, + 411, + 197 + ], + "blocks": [ + { + "bbox": [ + 384, + 169, + 411, + 197 + ], + "lines": [ + { + "bbox": [ + 384, + 169, + 411, + 197 + ], + "spans": [ + { + "bbox": [ + 384, + 169, + 411, + 197 + ], + "type": "image", + "image_path": "785e121e9804db3f8fe7b2ad9cc7a685db19d0e4ec2a939b5a46758b963bfe07.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 198, + 412, + 202 + ], + "lines": [ + { + "bbox": [ + 383, + 198, + 412, + 202 + ], + "spans": [ + { + "bbox": [ + 383, + 198, + 412, + 202 + ], + "type": "text", + "content": "Integrated Mesh" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 315, + 236, + 361, + 280 + ], + "blocks": [ + { + "bbox": [ + 318, + 229, + 357, + 237 + ], + "lines": [ + { + "bbox": [ + 318, + 229, + 357, + 237 + ], + "spans": [ + { + "bbox": [ + 318, + 229, + 357, + 237 + ], + "type": "text", + "content": "Prototype registration" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 236, + 361, + 280 + ], + "lines": [ + { + "bbox": [ + 315, + 236, + 361, + 280 + ], + "spans": [ + { + "bbox": [ + 315, + 236, + 361, + 280 + ], + "type": "image", + "image_path": "7fcd01438e7915a9e95b8ff928d5bc8a751a14845db799f154d18b3e556e89f9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 354 + ], + "type": "text", + "content": "Figure 3. SiM3D dataset creation pipeline. (1) We purchased eight types of objects with different properties, each associated with a CAD model. We altered part of the samples to create realistic anomalies; (2) We designed and calibrated an acquisition setup consisting of a high-precision 3D sensor with high-resolution cameras on top an industrial robot arm; (3) We acquired 360-degree scans of each object instance to populate the real split of the training and test sets, and rendered images and point clouds from the CAD models to create the synthetic split of the training sets; (4) We created voxel-based segmentation GTs by a mixed 2D-3D labelling pipeline." + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 365, + 236, + 418, + 256 + ], + "blocks": [ + { + "bbox": [ + 375, + 228, + 408, + 234 + ], + "lines": [ + { + "bbox": [ + 375, + 228, + 408, + 234 + ], + "spans": [ + { + "bbox": [ + 375, + 228, + 408, + 234 + ], + "type": "text", + "content": "Realistic rendering" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 236, + 418, + 256 + ], + "lines": [ + { + "bbox": [ + 365, + 236, + 418, + 256 + ], + "spans": [ + { + "bbox": [ + 365, + 236, + 418, + 256 + ], + "type": "image", + "image_path": "ffcc203adb553e9a5da42024b7426e08604f984c3933c6532e21c127b3017161.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 368, + 262, + 414, + 279 + ], + "blocks": [ + { + "bbox": [ + 368, + 262, + 414, + 279 + ], + "lines": [ + { + "bbox": [ + 368, + 262, + 414, + 279 + ], + "spans": [ + { + "bbox": [ + 368, + 262, + 414, + 279 + ], + "type": "image", + "image_path": "27c6758f9ce2ca253503440d10d7f341a116d9ab404029ce0ba3d0241a1af103.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 436, + 91, + 480, + 135 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 462, + 79 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 462, + 79 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 462, + 79 + ], + "type": "text", + "content": "(4) Labelling" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 85, + 503, + 90 + ], + "lines": [ + { + "bbox": [ + 474, + 85, + 503, + 90 + ], + "spans": [ + { + "bbox": [ + 474, + 85, + 503, + 90 + ], + "type": "text", + "content": "Annotated Views" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 91, + 480, + 135 + ], + "lines": [ + { + "bbox": [ + 436, + 91, + 480, + 135 + ], + "spans": [ + { + "bbox": [ + 436, + 91, + 480, + 135 + ], + "type": "image", + "image_path": "4a33eb270d948decff7ffefb30c1862c185a613a16e0a24d8d262f8017e05ea4.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 481, + 91, + 541, + 135 + ], + "blocks": [ + { + "bbox": [ + 481, + 91, + 541, + 135 + ], + "lines": [ + { + "bbox": [ + 481, + 91, + 541, + 135 + ], + "spans": [ + { + "bbox": [ + 481, + 91, + 541, + 135 + ], + "type": "image", + "image_path": "1995090b562a5f2e5c76f151b3759663eb18097678d533d91a63cebd8ba3563e.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 434, + 137, + 480, + 150 + ], + "blocks": [ + { + "bbox": [ + 434, + 137, + 480, + 150 + ], + "lines": [ + { + "bbox": [ + 434, + 137, + 480, + 150 + ], + "spans": [ + { + "bbox": [ + 434, + 137, + 480, + 150 + ], + "type": "image", + "image_path": "647779311375c4cefad78cf33c92de0c96a9cd09be881fef60e40a1ddd159f59.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 481, + 137, + 521, + 150 + ], + "blocks": [ + { + "bbox": [ + 481, + 137, + 521, + 150 + ], + "lines": [ + { + "bbox": [ + 481, + 137, + 521, + 150 + ], + "spans": [ + { + "bbox": [ + 481, + 137, + 521, + 150 + ], + "type": "image", + "image_path": "31974e1192941ca1c61071e5a989a01b40a9803c71e3f86f5f4ef5e422933518.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 152, + 500, + 158 + ], + "lines": [ + { + "bbox": [ + 481, + 152, + 500, + 158 + ], + "spans": [ + { + "bbox": [ + 481, + 152, + 500, + 158 + ], + "type": "text", + "content": "↓" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 434, + 152, + 471, + 159 + ], + "blocks": [ + { + "bbox": [ + 434, + 152, + 471, + 159 + ], + "lines": [ + { + "bbox": [ + 434, + 152, + 471, + 159 + ], + "spans": [ + { + "bbox": [ + 434, + 152, + 471, + 159 + ], + "type": "image", + "image_path": "df2c12c470d93ed3141e164a1fdb26b8e6082bac41a12813d4c4af90b42f8d30.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 434, + 161, + 480, + 175 + ], + "blocks": [ + { + "bbox": [ + 434, + 161, + 480, + 175 + ], + "lines": [ + { + "bbox": [ + 434, + 161, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 434, + 161, + 480, + 175 + ], + "type": "image", + "image_path": "8bab64fbf22ed1ec5b96317f64fd6a8bebaf358b1326810148b2550ddac2f00f.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 164, + 499, + 170 + ], + "lines": [ + { + "bbox": [ + 481, + 164, + 499, + 170 + ], + "spans": [ + { + "bbox": [ + 481, + 164, + 499, + 170 + ], + "type": "text", + "content": "rojection" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 436, + 176, + 480, + 224 + ], + "blocks": [ + { + "bbox": [ + 436, + 176, + 480, + 224 + ], + "lines": [ + { + "bbox": [ + 436, + 176, + 480, + 224 + ], + "spans": [ + { + "bbox": [ + 436, + 176, + 480, + 224 + ], + "type": "image", + "image_path": "13202d5d5ccfe76283addf07a3390fa08ffef800eecbb72c8b6f2bb1068c4ff0.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 225, + 473, + 231 + ], + "lines": [ + { + "bbox": [ + 444, + 225, + 473, + 231 + ], + "spans": [ + { + "bbox": [ + 444, + 225, + 473, + 231 + ], + "type": "text", + "content": "Integrated Mesh" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 481, + 178, + 539, + 223 + ], + "blocks": [ + { + "bbox": [ + 481, + 178, + 539, + 223 + ], + "lines": [ + { + "bbox": [ + 481, + 178, + 539, + 223 + ], + "spans": [ + { + "bbox": [ + 481, + 178, + 539, + 223 + ], + "type": "image", + "image_path": "03f8bf14dd4285b025192edb6e01b2393473513b8107aa48b250a003d31aad10.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 226, + 533, + 231 + ], + "lines": [ + { + "bbox": [ + 499, + 226, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 499, + 226, + 533, + 231 + ], + "type": "text", + "content": "Annotated Mesh" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 436, + 234, + 480, + 276 + ], + "blocks": [ + { + "bbox": [ + 436, + 234, + 480, + 276 + ], + "lines": [ + { + "bbox": [ + 436, + 234, + 480, + 276 + ], + "spans": [ + { + "bbox": [ + 436, + 234, + 480, + 276 + ], + "type": "image", + "image_path": "75a3bd599272a1aa50be8b8b95cadd241ecf6f5e41f06e9098ca07e2b37a9f5b.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 277, + 477, + 282 + ], + "lines": [ + { + "bbox": [ + 440, + 277, + 477, + 282 + ], + "spans": [ + { + "bbox": [ + 440, + 277, + 477, + 282 + ], + "type": "text", + "content": "3D Segmentation GT" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 250, + 537, + 262 + ], + "lines": [ + { + "bbox": [ + 498, + 250, + 537, + 262 + ], + "spans": [ + { + "bbox": [ + 498, + 250, + 537, + 262 + ], + "type": "text", + "content": "Manual Refinement & Voxelization" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "bbox": [ + 56, + 375, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 590 + ], + "type": "text", + "content": "prototype model so that all nominal ones should be identical – and for which an official CAD prototype was available from the manufacturer. The prototype CAD models are depicted in Fig. 3 (1.ii). To simulate various anomalies, we manually introduced different types of defects that represent realistic manufacturing flaws, as illustrated in Fig. 3 (1.iii). We altered the objects' appearance (e.g., paint modifications, Fig. 3 (1.iii) – bottom left), geometry (e.g., dents, Fig. 3 (1.iii) – bottom right), or both (e.g., contaminations, Fig. 3 (1.iii) – top left). It is important to point out that only half of the instances within each object type are modified, while the remaining half are kept in their original state to serve as nominal data. For each object type, one nominal instance will be used as the training set in the real2real experimental setup, while all other ones, both nominal and modified (i.e., anomalous), will make up the test set. This very same test set is also used in the synth2real setup, though the training data comes from CAD prototypes." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "text", + "content": "Sensor. We acquired our dataset with a top-tier 3D scanner, the Atos Q from ZEISS (Fig. 3 (2.i)). It consists of a stereo pair with two " + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "inline_equation", + "content": "12\\mathrm{Mpx}" + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "text", + "content": " greyscale cameras and a light projector. We consider the left camera to be our reference. The images have a resolution of " + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "inline_equation", + "content": "4096\\times 3000" + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "text", + "content": " pixels. The point clouds can feature up to 12 million elements, with a point-to-point distance between 0.04 and " + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "inline_equation", + "content": "0.15\\mathrm{mm}" + }, + { + "bbox": [ + 55, + 592, + 295, + 700 + ], + "type": "text", + "content": ". The sensor is mounted on an industrial robot arm with high repeatability and precision (Fig. 3 (2.iii))." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Calibration. We calibrate the greyscale cameras with a" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": "standard checkerboard pattern, estimating the intrinsic parameters matrix, " + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": ", and the radial and tangential distortion coefficients, " + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\delta = [k_{1}, k_{2}, k_{3}, p_{1}, p_{2}]" + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": ". We also calibrate the rotation, " + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "inline_equation", + "content": "R_{pc}" + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": ", and translation, " + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "inline_equation", + "content": "T_{pc}" + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": ", from the point cloud to the left camera reference frame, with a precise dotted calibration pattern built by ZEISS (Fig. 3 (2.ii)) that provides centre coordinates with precision " + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "inline_equation", + "content": "< 1" + }, + { + "bbox": [ + 313, + 375, + 555, + 471 + ], + "type": "text", + "content": " mm. More details are in the supplementary." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": "Real data acquisition. The acquisition process consists of 360-degree scanning operations around all object instances of our dataset (see Fig. 3 (3.i)). We define specific viewpoints to capture data from multiple angles around the object. These trajectories are sampled from concentric hemispheres that depend on the object's size. We provide poses relative to a reference view, i.e., " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ". Thus, the pose for the view " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ", consists of a rotation " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " and a translation " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " that aligns " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " the identity matrix and " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " null. Since in real industrial setups object instances would be acquired from as similar as possible viewpoints, we place markers in the scene to facilitate consistent positioning of objects of the same type. Thus, thanks to the high repeatability of the robot arm, images, " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ", and point clouds, " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": ", from the same viewpoint " + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 474, + 556, + 714 + ], + "type": "text", + "content": " turn out similar across acquisitions of different instances of the same object type. However, since the items were positioned manually with the help of the markers, slight displacements cannot be avoided. This introduces some variability across the scans of the instances of an object, as it also typically occurs in real indus" + } + ] + } + ], + "index": 64 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20947" + } + ] + } + ], + "index": 65 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "trial pipelines. For each object, after collecting data from all viewpoints, the ATOS Q sensor also returns an integrated mesh, " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": ", obtained by registering all point clouds from the different views using a proprietary algorithm. We collect and release the mesh to foster the development of methodologies that may also exploit this 3D representation. The 3D reference frame is aligned to the one of the first point cloud " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": " taken from the first viewpoint " + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "text", + "content": "Synthetic data generation. We employ the Blender [11] Python API to render synthetic data that simulates real acquisitions. To ensure the synthetic and real data are aligned, we select a reference mesh from the collected real instances for each object type, load it in Blender, and align it to the prototype CAD model, as shown in (Fig. 3 (3.ii) - left). Then, we set the Blender camera parameters to the intrinsics, " + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "text", + "content": ", of the left camera of our Atos Q sensor and render images from the same viewpoints " + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "text", + "content": " employed during the real acquisitions. The RGB renders are converted into greyscale images to match the real images provided by our acquisition setup (Fig. 3 (3.ii) - right). Renderings have been obtained by exploiting Cycles Printerer, a path-tracing renderer able to provide realistic results with reflections, emissive surfaces, and physically-based lighting. We also render depth maps (Fig. 3 (3.ii) - right), which we project into 3D point clouds (see Fig. 1, centre) using the intrinsic parameters, " + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 56, + 171, + 298, + 387 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": "Labelling. Since some anomalies, such as paintings and minor scratches, are visible only in greyscale images, while others, such as dents, are better visible in 3D, we must consider both sensing modalities to obtain precise 3D segmentation ground-truths for anomalous samples. Therefore, we develop a two-step labelling strategy, illustrated in Fig. 3 (4), that operates on both 2D images and 3D data. First, we manually annotate all the view images, " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", where anomalies are clearly visible, obtaining dense 2D segmentation masks. We employ a different ID and colour for every defect to ease the voxel AUPRO computation (see Sec. 4). Subsequently, for each annotated image, " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", we leverage the associated integrated mesh, " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", the intrinsic parameters " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", the point cloud to camera reference frame transformation, " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{pc}}, T_{\\mathrm{pc}})" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", and the transformation between " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "inline_equation", + "content": "(R_{i}, T_{i})" + }, + { + "bbox": [ + 56, + 391, + 298, + 713 + ], + "type": "text", + "content": ", to project the 2D annotations onto the integrated mesh. Then, since the projection may not be precise due to occlusions and slanted surfaces, each ground-truth mesh is thoroughly inspected and manually refined using a 3D visualisation and editing software, i.e., CloudCompare [12]. The labelling process was conducted by a team of 4 experts, with a final cross-check of the results. Finally, the annotated mesh is converted into a voxel grid with a voxel size of 2 mm. The choice of the size is determined by the smallest defect size discernible in the meshes. In the conversion process, a voxel is labelled as anomalous if it intersects at least one triangle labelled as such in the annotated mesh." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 129 + ], + "type": "table", + "html": "
TypeDimensions [cm]Dimensions [voxels]Total InstancesTrain Instances NominalTest Instances NominalAnomalousViews per Instance
Plastic Stool\\( {35} \\times {35} \\times {30} \\)\\( {330} \\times {326} \\times {317} \\)221 real or 1 synth101012
Rubbish Bin\\( {26} \\times {21} \\times {33} \\)\\( {333} \\times {339} \\times {326} \\)421 real or 1 synth202012
Wicker Vase\\( {17} \\times {17} \\times {15} \\)\\( {329} \\times {309} \\times {290} \\)221 real or 1 synth101012
Bathroom Furniture\\( {33} \\times {33} \\times {50} \\)\\( {371} \\times {392} \\times {363} \\)201 real or 1 synth81036
Container\\( {20} \\times {25} \\times {10} \\)\\( {327} \\times {308} \\times {286} \\)941 real or 1 synth464612
Plastic Vase\\( {12} \\times {12} \\times 9 \\)\\( {326} \\times {308} \\times {286} \\)991 real or 1 synth484912
Wooden Stool\\( {48} \\times {42} \\times {45} \\)\\( {331} \\times {367} \\times {327} \\)151 real or 1 synth6712
Sink Cabinet\\( {44} \\times {25} \\times {50} \\)\\( {342} \\times {349} \\times {357} \\)191 real or 1 synth9836
", + "image_path": "298d2f195d133bd06d679291d4951fec2abbfac0bc724df73e05fd2274192a9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 384, + 137, + 484, + 147 + ], + "lines": [ + { + "bbox": [ + 384, + 137, + 484, + 147 + ], + "spans": [ + { + "bbox": [ + 384, + 137, + 484, + 147 + ], + "type": "text", + "content": "Table 2. Dataset statistics." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 171, + 428, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 428, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 428, + 182 + ], + "type": "text", + "content": "4. SiM3D Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 191, + 466, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 466, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 466, + 204 + ], + "type": "text", + "content": "4.1. Dataset, Tasks, and Metrics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 210, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 330 + ], + "type": "text", + "content": "Dataset and Setups. Using the procedure described in Sec. 3, we collected 333 object instances belonging to the eight object types, as shown in Tab. 2. We captured from 12 to 36 views for each instance, depending on the type. We split the data into training and test sets. For each type, the training set comprises a single instance, either real in the real2real setup or synthetic in the synth2real setup. The test set includes all other instances and is balanced between nominal and anomalous samples. The test set is the same for both setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": "Tasks. Our benchmark focuses on the tasks of detecting and segmenting objects' anomalies using multiview and multimodal data. Specifically, the inputs to the methods consist of a set of images " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": " captured from different viewpoints alongside 3D data, " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ", that capture an object entirely. The 3D data that a method can use are a set of point clouds " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": " or depth maps " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": " taken from the same vantage points as the images, an integrated mesh " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ", or also a merged point cloud " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ". A method capable of multiview multimodal 3D anomaly detection should take these inputs, integrate the 2D and 3D information from multiple views, and produce a global anomaly score for the test sample under inspection. On the other hand, a method pursuing multiview multimodal 3D anomaly segmentation should yield a 3D output, where each 3D position carries a score representing the likelihood of belonging to an anomaly. In our benchmark, a method for this task takes as input all images and 3D data, " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "(\\{I_i\\}_{i=1}^n, \\mathbf{X})" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": " to output an Anomaly Volume, i.e., a voxel grid, " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{X \\times Y \\times Z}" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "X, Y" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": " are the grid dimensions and each voxel carries an anomaly score. For several reasons, we chose voxel grids as the 3D output representation of our benchmark. Firstly, as their grid structure resembles 2D Anomaly Maps, existing 2D segmentation methods and metrics can be easily extended. Moreover, voxel grids are suitable for representing every type of 3D defect, including missing parts. One major challenge with voxel grids pertains to choosing the voxel size. Ideally, we would want a size that allows accurate segmentation of even the smallest defects. However, the memory occupancy of voxel grids scales cubically with the inverse of the voxel size. Thus, we create ground-truth volumes using a voxel resolution of " + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "inline_equation", + "content": "2\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 331, + 556, + 714 + ], + "type": "text", + "content": ", which enables precise localisation of" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20948" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 194 + ], + "type": "image", + "image_path": "e924cd3aa3e0550b385feea033f18ba92ed8d53f140ce5a8bde5fb8367a108fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 202, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 258 + ], + "type": "text", + "content": "Figure 4. Extending singleview methods. Anomaly Maps obtained by processing singleview inputs with standard ADS methods are projected into the 3D and aggregated into a voxel grid, with each voxel containing a list of scores from multiple views. The Anomaly Volume is obtained by voxel-wise max pooling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 278, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 295, + 301 + ], + "type": "text", + "content": "even the smallest defects present in our dataset while keeping memory occupancy affordable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": "Metrics. Akin to standard practice in 2D ADS [4], we evaluate detection performance by the Area Under the Receiver Operator Curve (AUROC) of the global anomaly scores predicted by the methods. As we compute a global score for each object instance, we name it instance-level AUROC (I-AUROC). Regarding segmentation performance, we evaluate methods with an adaptation of the popular Area Under the Per-Region Overlap (AUPRO) curve. First of all, we define the voxel-based Per-Region Overlap (PRO) metric as " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{PRO}_t = \\frac{1}{N}\\sum_{n = 1}^N\\frac{|\\overline{\\Psi}_t\\cap V_n|}{|V_n|}" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\overline{\\Psi}_t" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": " is an Anomaly Volume binarised using a threshold, " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": " is the number of blobs, i.e., defects, in the ground-truth, and " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "V_{n}" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": " are the voxels belonging to that blob. Since during labelling, we annotated each defect in our dataset with a different ID (see Sec. 3), the " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": " defect blobs in a ground-truth are easily identifiable. This metric is computed across multiple thresholds, " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": ", estimated at different False Positive Rates (FPRs). We calculate the FPR using only non-empty voxels. We sample FPRs from 0 to the maximum value uniformly, we estimate the corresponding threshold " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": ", and we binarise the score grid accordingly. In this way, we construct a curve that plots the PRO values against the corresponding FPRs. The segmentation performance is obtained by integrating such a curve up to a specific False Positive Rate. As proposed in [14], we use " + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 56, + 303, + 296, + 653 + ], + "type": "text", + "content": " as integration bound for the PRO curve to set a challenging bar representative of the requirements of real industrial applications. Finally, we normalise the resulting area to the interval [0, 1]. We dub this metric voxel-level AUPRO (V-AUPRO)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 122, + 672 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Extending singleview methods to Multiview 3D ADS. Current ADS methods predict a 2D anomaly map from a singleview input, either unimodal, e.g., images, or multi-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": "modal, e.g., images and point clouds. To obtain a set of baselines for our novel multiview 3D ADS task, we extend popular and effective singleview methods by means of the 3D aggregation strategy illustrated in Fig. 4. Given the considered method, we process the data corresponding to an input view, " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", so as to obtain a 2D Anomaly Map. Similarly to Sec. 3, given the mesh, " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", intrinsic parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", point cloud to camera reference frame transformation, " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{pc}}, T_{\\mathrm{pc}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": " and view pose " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "v_{0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "inline_equation", + "content": "(R_{i}, T_{i})" + }, + { + "bbox": [ + 313, + 72, + 555, + 300 + ], + "type": "text", + "content": ", we can project each pixel of the 2D Anomaly Map to its corresponding 3D position. We discretise the 3D position, finding the corresponding voxel index. We associate that voxel with the corresponding anomaly score. We repeat the process for each view, keeping track of the scores that project onto the same voxels. Eventually, we compute the maximum among all scores projected into each voxel, obtaining the final Anomaly Volume. To extract the global score required by the I-AUROC, we compute the maximum across all the scores within the Anomaly Volume." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": "Selected Methods. We extend some of the most popular and effective unimodal and multimodal singleview ADS methods, starting from their official code, and evaluate them on SiM3D. Regarding unimodal methods, we select PatchCore [30], with either WideResNet101 [42] or DINOv2 [24] as backbone, to extract features from images of all views, " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": " of the single-instance training set to create memory banks. Moreover, we use EfficientAD [1] using all views as their training set. As for multimodal ADS methods, we use BTF [18] employing WideResNet101 [42] as the image encoder and FPFH [33] as the algorithm to extract point cloud features. We create memory banks using features extracted from images, " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": ", and point clouds, " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": ", of all views of the single-instance training set. We use the same strategy with M3DM [36]. As 2D feature extractor, we exploit DINO-v2. As point cloud feature extractor, instead of using a transformer backbone, i.e., PointMAE [17], we utilise FPFH to extract 3D features, similarly to BTF. Indeed, we would need a severe downsample (from " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\sim 5 - 7\\mathrm{M}" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\sim 8\\mathrm{k}" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": " points) to work with Point-MAE, and we verified that at such low resolution, most of the defects no longer show up in the point clouds. Finally, we select CFM [14], using all views and clouds as a single training set. Based on the same consideration as for M3DM, we use DINO-v2 and FPFH to produce 2D and 3D features. Due to computational constraints, all methods process images padded and downsampled to " + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "inline_equation", + "content": "1540 \\times 1540" + }, + { + "bbox": [ + 313, + 303, + 556, + 662 + ], + "type": "text", + "content": ". Moreover, for compatibility with the employed 2D feature extractors, which accept RGB images, we create 3-channel inputs by stacking our greyscale images thrice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "3D Data pre-processing. To obtain less noisy data and improve the results of the above-mentioned baselines, we remove the background from 3D data. We estimate a 3D plane using a RANSAC-based plane segmentation algo" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20949" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 545, + 274 + ], + "type": "image", + "image_path": "7534f9153be095bf7fd2844e007c4f09313d5a09d32f83ba7fcf455b80669f4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 284, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 553, + 296 + ], + "type": "text", + "content": "Figure 5. Qualitatives. Selected views with defects (left) and Anomaly Volumes for several methods (right) downsampled for visualisation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 307, + 552, + 378 + ], + "blocks": [ + { + "bbox": [ + 59, + 307, + 552, + 378 + ], + "lines": [ + { + "bbox": [ + 59, + 307, + 552, + 378 + ], + "spans": [ + { + "bbox": [ + 59, + 307, + 552, + 378 + ], + "type": "table", + "html": "
MethodModalityreal2realsynth2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.MeanPl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
PatchCore w/ WRN-101RGB0.7400.9870.6360.7770.7740.5511.0000.5660.7540.4620.2350.5370.3530.6780.5760.5170.2500.451
PatchCore w/DINO-v2RGB0.5000.9580.6360.6220.5780.5631.0000.5630.6780.9580.8970.4130.4640.2070.6840.6070.0870.540
EfficientADRGB0.2800.7320.0000.8780.4240.7300.9280.7120.5850.1230.6730.0000.8480.0080.4870.7500.6620.443
ASTRGB0.5370.5660.6110.5150.5730.4960.8390.5370.5841.0000.3441.0000.5250.4820.3880.4280.1870.544
BTFRGB + Point Cloud0.4210.2170.5040.5650.5450.4710.6780.4240.4780.4620.2940.5200.4440.4640.4240.4820.2870.422
CFM w/DINO-v2 + FPFHRGB + Point Cloud0.1980.3010.0740.5150.4830.4560.7320.8250.4480.0080.0000.1900.4240.3130.4590.4280.3620.273
M3DM w/DINO-v2 + FPFHRGB + Point Cloud0.7020.9880.6610.5450.5560.6490.3920.4750.6210.1070.1170.7350.5650.3810.5860.2140.5120.402
ASTRGB + Depth0.9500.9270.7850.4740.5420.4700.4280.9250.6870.6360.0020.5040.4740.4620.5690.4280.8870.495
", + "image_path": "b84e61275dbf91d8a6cf57acf80d31dabaeeb6f8f480cb60648dc2b9bb8253c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 387, + 474, + 398 + ], + "lines": [ + { + "bbox": [ + 133, + 387, + 474, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 387, + 474, + 398 + ], + "type": "text", + "content": "Table 3. Anomaly detection results (I-AUROC). Best results in bold, runner-ups underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": "rithm, which, iteratively, randomly selects a small subset of points (e.g., 10) to fit a candidate plane model. For the candidate planes, we compute the consensus set with threshold " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": " on the point-to-plane distance. After 1000 iterations, we return the best-fitting plane. To account for spurious background artefacts, we shift the plane by an offset " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": " and finally remove all points below the shifted plane. Given the different sizes of the objects, for each object type, we select different optimal thresholds " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": " and offsets " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": ", as detailed in the supplementary material. We apply this background removal technique to the integrated mesh, " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": ", due to the larger number of background points, which facilitates the RANSAC algorithm. Then, from the background-filtered integrated mesh, we leverage ray-tracing techniques and known camera parameters to render, for each view, depth maps, " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\{D_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": ", pixel-aligned to the corresponding images, " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": " at the sensor resolution " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "(4096 \\times 3000)" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": ". Finally, these depth maps are padded and downsampled to " + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "inline_equation", + "content": "1540 \\times 1540" + }, + { + "bbox": [ + 54, + 419, + 297, + 706 + ], + "type": "text", + "content": ", and the depth is lifted in 3D via the camera parameters to obtain dense and clean point clouds. The depth maps and point clouds obtained by rendering the preprocessed meshes serve as the input data for all the experiments presented in the subsequent section and will be available in the SiM3D dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 418, + 430, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 430, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 430, + 430 + ], + "type": "text", + "content": "5. Benchmark Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "text", + "content": "Main Results. In Tab. 3 and Tab. 4, we report the results obtained on the SiM3D benchmark by extending the considered methods as Multiview 3D ADS baselines. By looking at the mean detection performance (Tab. 3), we notice how methods that natively process only RGB images outperform natively multimodal ones in both the real2real and synth2real setups. With the notable exception of AST, this also stands for mean segmentation performance (Tab. 4). We posit that this is due to the main multimodal methods having been designed and tuned to process point clouds of the resolution featured by existing multimodal benchmarks, which is much lower than that of SiM3D, i.e. " + }, + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\sim 8\\mathrm{K}" + }, + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\sim 2.3\\mathrm{M}" + }, + { + "bbox": [ + 312, + 450, + 555, + 715 + ], + "type": "text", + "content": " points (after downsampling). Indeed, AST is the only one designed to work natively with depth maps. Hence, our experiments suggest that current multimodal ADS methods do not scale to significantly higher resolutions. In turn, we could not even deploy the native 3D feature extractors of M3DM and CFM, the methods currently excelling on [6, 7]. Furthermore, we notice how methods relying on memory banks, such as PatchCore and M3DM, tend to attain slightly better mean detection and segmentation performance than their counterparts" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20950" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 165, + 140 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 165, + 140 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 165, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 165, + 140 + ], + "type": "table", + "html": "
MethodModality
PatchCore w/ WRN-101RGB
PatchCore w/ DINO-v2RGB
EfficientADRGB
ASTRGB
BTFRGB + Point Cloud
CFM w/ DINO-v2 + FPFHRGB + Point Cloud
M3DM w/ DINO-v2 + FPFHRGB + Point Cloud
ASTRGB + Depth
", + "image_path": "446c24ed812763c89421346d05157bf3eac35a72457f50ac6a2ac9314621475e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 167, + 72, + 358, + 140 + ], + "blocks": [ + { + "bbox": [ + 167, + 72, + 358, + 140 + ], + "lines": [ + { + "bbox": [ + 167, + 72, + 358, + 140 + ], + "spans": [ + { + "bbox": [ + 167, + 72, + 358, + 140 + ], + "type": "table", + "html": "
real2real
Pl. StoolRub. BinW. VaseB. Furn.Cont.Pl. VaseW. StoolSink Cab.Mean
0.7100.4610.7610.6750.6940.7470.3800.6090.630
0.7450.4690.7750.7920.7090.7530.4350.6900.671
0.6820.4620.7630.5340.6800.7430.4070.4880.594
0.7200.5020.7890.8030.7160.7640.4460.7580.687
0.5510.4020.7500.3770.6140.7410.0920.0300.444
0.5970.4150.7680.5050.6400.7430.3150.3210.538
0.7330.4520.7670.7020.7020.7520.2880.1260.565
0.7500.5030.7920.8070.7160.7640.4670.7980.699
", + "image_path": "66ab6a75f887ca4f2f6096465ddf3f78e9606c0961d4280b818bf799b8266692.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 359, + 72, + 551, + 140 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 551, + 140 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 551, + 140 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 551, + 140 + ], + "type": "table", + "html": "
pl.Stoolrub.BinW.VaseB.Furn.Cont.Pl.VaseW.StoolSink Cab.Mean
0.7340.4370.7510.4540.6780.7410.3860.6180.600
0.7010.4570.7720.5630.6480.7340.3210.5750.596
0.6800.4490.7290.5230.6690.7470.4310.3610.573
0.7260.4880.7910.8060.7050.7640.4610.6950.679
0.5470.4020.7560.3690.6100.7410.1030.0400.446
0.5230.4130.7530.5420.6210.7410.2220.3140.516
0.6900.4540.7700.4610.6450.7340.3180.1320.526
0.7510.5020.7930.8050.7170.7640.4500.7850.695
", + "image_path": "c4b37ea7a5d3af939ad3519c6a8402c13c2770ce1b4e9ed517dd34a5797f7d39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 182, + 179, + 286 + ], + "blocks": [ + { + "bbox": [ + 113, + 150, + 495, + 162 + ], + "lines": [ + { + "bbox": [ + 113, + 150, + 495, + 162 + ], + "spans": [ + { + "bbox": [ + 113, + 150, + 495, + 162 + ], + "type": "text", + "content": "Table 4. Anomaly segmentation results (V-AUPRO@1%). Best results in bold, runner-ups underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 182, + 179, + 286 + ], + "lines": [ + { + "bbox": [ + 59, + 182, + 179, + 286 + ], + "spans": [ + { + "bbox": [ + 59, + 182, + 179, + 286 + ], + "type": "table", + "html": "
MethodModalityFeatures
2D3D
PatchCoreRGBWRN-101-
RGBDINO-v2-
Depth-WRN-101
Depth-DINO-v2
Point Cloud-FPFH
BTFRGB + Point CloudWRN-101FPFH
RGB + DepthWRN-101WRN-101
CFMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
M3DMRGB + Point CloudDINO-v2FPFH
RGB + DepthDINO-v2DINO-v2
ASTRGBEffNet-B5-
RGB + DepthEffNet-B5EffNet-B5
", + "image_path": "fa82f769d9445f8b5ec4d0a6318ba3cba498766183684382f0f6b53be8c60d04.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 181, + 182, + 293, + 286 + ], + "blocks": [ + { + "bbox": [ + 181, + 182, + 293, + 286 + ], + "lines": [ + { + "bbox": [ + 181, + 182, + 293, + 286 + ], + "spans": [ + { + "bbox": [ + 181, + 182, + 293, + 286 + ], + "type": "table", + "html": "
real2realsynth2real
DetectionSegmentationDetectionSegmentation
0.7540.6300.4510.600
0.6780.6710.5400.596
0.5820.4860.5070.484
0.6000.4960.2460.471
0.4150.4640.3130.460
0.4780.4440.4220.446
0.7070.6090.4480.543
0.4480.5380.2730.516
0.5480.6630.3110.620
0.6210.5650.4020.526
0.6590.5030.2540.476
0.5840.6870.5440.679
0.6870.6990.4950.695
", + "image_path": "f1564c5a2133011f8b5473cf28ebd96bdddeb217748d3d3f47437808b3a71e7e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "type": "text", + "content": "Table 5. Anomaly detection and segmentation mean results with different input modalities. Best results per method in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 342, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 676 + ], + "type": "text", + "content": "within the same modality that require substantial training, i.e., EfficientAD and CFM. We argue that this may be ascribed to the single-instance setup, which renders the latter category of approaches harder to optimise due to the limited number of training images, which are as many as the views. The methods most effectively extended to our multiview 3D task are Patchcore, which, equipped with either WRN-101 or DINO-v2, can achieve the best detection in both setups, and AST, achieving the best segmentation performance in both setups. Indeed, Patchcore yields I-AUROCs of 0.754 (WRN-101) and 0.540 (DINO-v2), while AST yields a V-AUPRO@1% of 0.699 and 0.695, in the real2real and synth2real setups, respectively. It is also worth pointing out how, due to the domain shift, we can notice a gap between the real2real and synth2real setups, especially in detection (0.754 vs. 0.540, I-AUROC) compared to segmentation performance (0.699 vs. 0.695, V-AUPRO@1%). The qualitative results, shown in Fig. 5, visually support the above observations. Overall, the relatively low, definitely far from saturated, performance figures provided by baselines designed by straightforward extension of state-of-the-art methods delivering impressive results in singleview 2D benchmarks hint at the novel and challenging nature of the task set forth by SiM3D. Thus, our experimental findings by baseline approaches reveal the need for more advanced and specific methods and paradigms, both in the real2real and, even more so, the synth2real setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Alternative 3D Representations. Given the lack of feature extractors amenable to processing high-resolution point clouds, we explore the use of depth maps to capture 3D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 183, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 555, + 483 + ], + "type": "text", + "content": "cues. The 2D grid structure of depth maps enables the deployment of popular backbones pre-trained on RGB data that support high-resolution inputs, such as DINOv2. Thus, we create and evaluate additional baselines using depth maps and report the results in Tab. 5. Similarly to greyscale images, we tile depth maps along the channel axis to ensure compatibility with the employed backbones. The results show that the use of depth maps rather than point clouds in multimodal methods tends to improve detection and segmentation performance by sizable margins, e.g., in the real2real setup, BTF yields 0.707 I-AUROC and 0.609 V-AUPRO when processing depths compared to 0.478 I-AUROC and 0.444 V-AUPRO when fed with point clouds (row 6 vs. row 7) while relying on the same WRN101 image backbone. Indeed, all multimodal methods but M3DM achieve their best performance in both tasks and setups by processing depth maps instead of point clouds. Yet, as vouched by the results achieved by PatchCore, RGB images compare favourably to depth maps in terms of detection performance. We ascribe this to the feature extractors used to compute depth features being pre-trained on RGB images, which may be conducive to yielding weaker features when fed with depth inputs. This points out an intriguing research question: would it be possible to develop a foundation model to process depth maps effectively?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 492, + 413, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 413, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 413, + 506 + ], + "type": "text", + "content": "6. Final Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 704 + ], + "type": "text", + "content": "We introduced SiM3D, the first benchmark and dataset focused on integrating multiview and multimodal information for comprehensive 3D ADS. Our experiments highlight several open challenges. First, a need for methods that can process multiview inputs more effectively, e.g., by taking all views as input rather than integrating singleview outputs, and that can be trained with a very limited number of nominal samples, e.g., a single instance. Second, training with synthetic data requires more robust techniques to address domain-shift challenges. Finally, the current literature lacks 3D backbones capable of processing high-resolution point clouds. Employing depth maps seems a promising strategy to facilitate this processing, yet a foundational model for depth maps is still missing. We believe that SiM3D may serve as a strong foundation for tackling these challenges and fostering future research toward 3D ADS." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20951" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "content": "We acknowledge financial support from the National Recovery and Resilience Plan (NRRP), Mission 4, Component 2, Investment 1.1, Call No. 1409 published on 14 September 2022 by the Italian Ministry of University and Research (MUR), funded by the European Union." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 294, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 294, + 210 + ], + "type": "text", + "content": "This work was also partially supported by the FSE+ 2021-2027 program under Article 24, paragraph 3, letter a) of Law 240/2010 and its subsequent amendments, as well as by Regional Resolution No. 693/2023 (Project Code: 2023-20090/ER), Grant CUP: J19J23000730002." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 294, + 258 + ], + "type": "text", + "content": "We sincerely thank Lucia Gasperini for her contribution to dataset acquisition and labelling. We also extend our gratitude to Manuel Turrini for his support in setting up the acquisition process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 270, + 114, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 114, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 114, + 281 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 289, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 289, + 294, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 294, + 344 + ], + "type": "text", + "content": "[1] Kilian Batzner, Lars Heckler, and Rebecca König. Efficient-tad: Accurate visual anomaly detection at millisecond-level latencies. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 128-138, 2024. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 61, + 346, + 295, + 378 + ], + "type": "text", + "content": "[2] Liron Bergman, Niv Cohen, and Yedid Hoshen. Deep nearest neighbor anomaly detection. arXiv preprint arXiv:2002.10445, 2020.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 380, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 380, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 294, + 422 + ], + "type": "text", + "content": "[3] Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 479 + ], + "type": "text", + "content": "[4] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad - a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 533 + ], + "type": "text", + "content": "[5] Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "text", + "content": "[6] Paul Bergmann, Jin Xin, David Sattlegger, and Carsten Steger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, pages 202-213, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 603, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 657 + ], + "type": "text", + "content": "[7] Luca Bonfiglioli, Marco Toschi, Davide Silvestri, Nicola Fioraio, and Daniele De Gregorio. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. In Proceedings of the 16th Asian Conference on Computer Vision (ACCV2022), 2022. ACCV. 1, 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 712 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 117 + ], + "type": "text", + "content": "[9] Li-Ling Chiu and Shang-Hong Lai. Self-supervised normalizing flows for image anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2926-2935, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "text", + "content": "[10] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. ArXiv, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[11] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[12] CloudCompare Online Community. CloudCompare - 3D point cloud and mesh processing software Open Source Project. EDF, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 271 + ], + "type": "text", + "content": "[13] Alex Costanzino, Pierluigi Zama Ramirez, Mirko Del Moro, Agostino Aiezzo, Giuseppe Lisanti, Samuele Salti, and Luigi Di Stefano. Test time training for industrial anomaly segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 3910-3920, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "text", + "content": "[14] Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, and Luigi Di Stefano. Multimodal industrial anomaly detection by crossmodal feature mapping. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. CVPR. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "type": "text", + "content": "[15] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. In International Conference on Pattern Recognition, pages 475-489. Springer, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 437 + ], + "type": "text", + "content": "[16] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 98-107, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 491 + ], + "type": "text", + "content": "[17] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "text", + "content": "[18] Eliahu Horwitz and Yedid Hoshen. Back to the feature: classical 3d features are (almost) all you need for 3d anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967-2976, 2023. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "type": "text", + "content": "[19] Jinlei Hou, Yingying Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, and Hong Zhou. Divide-and-assemble: Learning block-wise memory for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8791-8800, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 658 + ], + "type": "text", + "content": "[20] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664-9674, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[21] Jiaqi Liu, Guoyang Xie, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, Feng Zheng, et al. Real3d-ad: A dataset of point cloud anomaly detection. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20952" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[22] Jiaqi Liu, Guoyang Xie, Jingbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey. arXiv preprint arXiv:2301.11514, 2, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[23] Fabio Valerio Massoli, Fabrizio Falchi, Alperen Kantarci, Şeymanur Akti, Hazim Kemal Ekenel, and Giuseppe Amato. Cocca: Multilayer one-class classification for anomaly detection. IEEE Transactions on Neural Networks and Learning Systems, 33(6):2313-2323, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 273 + ], + "type": "text", + "content": "[24] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 340 + ], + "type": "text", + "content": "[25] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 604-621. Springer, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "text", + "content": "[26] Jonathan Pinnay and Keng Chai. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing, pages 394-406. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 431 + ], + "type": "text", + "content": "[27] Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2806-2814, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[28] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733. IEEE, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 555 + ], + "type": "text", + "content": "[29] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13576-13586, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[30] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of 2022 IEEE Conference on Computer Vision and Pattern Recognition, pages 14298-14308, 2022. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[31] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907-1916, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[32] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmetric student-teacher networks for industrial anomaly detection. In Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[33] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE International Conference on Robotics and Automation, pages 3212-3217, 2009. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[34] Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learning and evaluating representations for deep one-class classification. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 237 + ], + "type": "text", + "content": "[35] Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22883-22892, 2024. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[36] Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032-8041, 2023. 1, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[37] Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650-656, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[38] Minghui Yang, Peng Wu, and Hui Feng. Memseg: A semi-supervised method for image surface defect detection using differences and commonalities. Engineering Applications of Artificial Intelligence, 119:105835, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 425 + ], + "type": "text", + "content": "[39] Jihun Yi and Sungroh Yoon. Patch svdd: Patch-level svdd for anomaly detection and segmentation. In Proceedings of the Asian conference on computer vision, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[40] Seungdong Yoa, Seungjun Lee, Chiyoon Kim, and Hyunwoo J Kim. Self-supervised learning for anomaly detection with dynamic local augmentation. IEEE Access, 9:147201-147211, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[41] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 536 + ], + "type": "text", + "content": "[42] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[43] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8330-8339, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 624 + ], + "type": "text", + "content": "[44] Zheng Zhang and Xiaogang Deng. Anomaly detection using improved deep svdd model with data structure preservation. Pattern Recognition Letters, 148:1-6, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[45] Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection. arXiv preprint arXiv:2310.07716, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[46] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. arXiv preprint arXiv:2207.14315, 2022. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20953" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_content_list.json b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5dbb5124a51acb3c1257c108e22c69a8493f10ec --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_content_list.json @@ -0,0 +1,1458 @@ +[ + { + "type": "text", + "text": "Sibai: A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning", + "text_level": 1, + "bbox": [ + 101, + 140, + 893, + 161 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Melanie Gotoz, Torsten Krauß, Alexandra Dmitrienko", + "bbox": [ + 295, + 189, + 700, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Wurzburg", + "bbox": [ + 405, + 207, + 589, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanderring 2, 97070 Wurzburg, Germany", + "bbox": [ + 339, + 224, + 656, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "melanie.goetz@uni-wuerzburg.de, torsten.krauss@uni-wuerzburg.de,", + "bbox": [ + 222, + 244, + 771, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "alexandra.dmitrienko@uni-wuerzburg.de", + "bbox": [ + 333, + 261, + 653, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 309, + 330, + 323 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) enables collaborative machine learning across decentralized clients without sharing raw data, which offers enhanced privacy and improved performance. However, FL is vulnerable to poisoning attacks, compromising model integrity through both untargeted performance degradation and targeted backdoor attacks. Detecting backdoors in FL is challenging due to their stealthy nature and variability in local datasets. Existing defenses struggle against adaptive adversaries and distinguishing between poisoning and genuine dataset anomalies. This paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning defense for FL. Leveraging the staple few-shot learning technique of Siamese networks, Sibai effectively detects malicious contributions in various scenarios, including settings with strong variations between clients' datasets and encounters with adaptive adversaries. Sibai achieves high detection rates, prevents backdoors, minimizes performance impact, and outperforms eight recent defenses regarding F1 score, poisoning prevention, and consistency across complex scenarios.", + "bbox": [ + 99, + 339, + 485, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 102, + 659, + 230, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Learning (FL) [59] enables collaborative deep neural network (DNN) training across decentralized clients by sharing model updates instead of raw data, preserving privacy. While FL offers significant benefits, it is susceptible to poisoning attacks by untrustworthy or compromised clients who manipulate training data, labels, hyperparameters, or model parameters. These attacks can be untargeted [15, 31, 55, 57], degrading global model performance, or targeted [2-4, 11, 18, 19, 33, 40, 42, 45, 52, 54, 56], introducing backdoors that misclassify inputs with specific triggers while maintaining normal performance otherwise. Backdoors are particularly dangerous due to their stealth and severe risks in security-critical applications like autonomous vehicles [32, 41, 62]. FL poisoning defenses include in", + "bbox": [ + 99, + 683, + 485, + 890 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fluence reduction (IR) [1, 3, 39, 50], robust aggregation (RA) [36, 61], and the more common detection and filtering (DF) [5, 8, 17, 27, 43, 47, 48, 63], which excludes poisoned models before aggregation. Although IR and RA can leverage information from all local models, their methods, such as noising and clipping, often degrade benign performance when applied broadly.", + "bbox": [ + 509, + 310, + 895, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Challenges. Most DF-based defenses rely on predefined metrics to detect poisoning, but this approach faces two key challenges. First, adaptive adversaries can tailor poisoned models to evade detection by appearing normal within these metrics [27, 29]. Expanding the metric sets [27] partially addresses this, but still relies on rigid criteria which are vulnerable to adaptation. Second, the natural diversity of non-identically and independently distributed (non-i.i.d.) datasets in real-world scenarios makes distinguishing poisoned models from those reflecting unusual data distributions difficult.", + "bbox": [ + 509, + 414, + 895, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meta-Classifiers. This paper introduces a dynamic meta-classifier approach to FL poisoning defense, which avoids rigid metrics and instead learns to differentiate malicious directly from observing local models. Recent methods [9, 21, 22, 25, 58] show that meta-classifiers can effectively detect poisoned models in an offline scenario. However, applying this to FL poses challenges: training requires a diverse set of so-called shadow models, including both benign and poisoned models, and existing methods depend on time-consuming preprocessing.", + "bbox": [ + 509, + 575, + 895, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contributions. To tackle the aforementioned challenges, this paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning detection method specifically tailored to function as a DF-based defense for FL. Sibai is based on the staple few-shot learning technique of Siamese networks. Our contributions can be summarized as follows:", + "bbox": [ + 509, + 724, + 895, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- Sibai is the first application of a meta-classifier in DF-based poisoning defenses for FL. Unlike conventional approaches that are reliant on predefined fixed metrics for detection, Sibai dynamically learns the criteria to distin", + "bbox": [ + 511, + 830, + 895, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 54 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3787", + "bbox": [ + 482, + 926, + 514, + 938 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guish between benign and malicious local models. Using a dynamic approach based on ML, Sibai poses a challenge for sophisticated attackers that try to adapt to predefined metrics, as it utilizes a diverse set of learned, complex model characteristics for its decision-making.", + "bbox": [ + 112, + 102, + 482, + 175 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We utilize Siamese Networks in the context of a poisoning detection meta-classifier. These networks are designed to compare and find dissimilarities between two inputs, in the case of Sibai DNN models.", + "- Sibai operates directly on flattened DNN representations using a novel layer-wise dimensionality reduction pipeline, using Principal Component Analysis (PCA) enhanced with engineered features. This feature-rich approach improves Sibai's ability to learn complex criteria, enhancing robustness against adaptation. Additionally, it enables Sibai to efficiently inspect all local models in a federation without requiring extensive time-consuming preprocessing steps.", + "- We extensively evaluate Sibai in various FL scenarios against five different poisoning attacks, assessing its effectiveness against stealthy and adaptive attackers and in scenarios with non-i.i.d. data. To conduct our evaluation, we created a substantial dataset of shadow models, including 5720 benign and 4145 malicious models. Sibai consistently achieves high detection rates across all tested scenarios without notable degradation of the global model's performance. Notably, Sibai successfully prevents the establishment of effective backdoors in all experiments, and outperforms the eight existing FL poisoning defenses, demonstrating higher F1 scores and consistently maintaining the lowest backdoor accuracy across all experiments." + ], + "bbox": [ + 101, + 176, + 482, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, Sibai introduces a robust ML-based FL poisoning defense to effectively address adaptive adversaries and non-i.i.d. scenarios.", + "bbox": [ + 101, + 546, + 483, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 102, + 606, + 225, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section introduces concepts relevant to this paper.", + "bbox": [ + 101, + 631, + 452, + 646 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Learning FL [26, 36, 59] enables collaborative ML by training on local data while preserving privacy. In each round, the server sends the global model to selected clients, who train it locally and return updates. The server aggregates updates, often using FedAVG [36], a weighted average with a global learning rate $\\delta$ . The algorithm is explained in more detail in Appendix A.", + "bbox": [ + 101, + 650, + 482, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "IID vs. non-IID Data Distributions ML models capture the characteristics of their datasets, emphasizing patterns in the training data. In an ideal identically and independently distributed (IID) scenario, all clients have similar data distributions, making outlier detection an effective strategy to detect poisoning. However, the real world rarely resembles those conditions; instead, clients usually exhibit non-i.i.d. data distributions which complicate the distinction between anomalies and poisoning. For classification tasks, '1-class", + "bbox": [ + 101, + 757, + 482, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "non-i.i.d.' describes datasets dominated by a single label, with the disproportionality degree regulated by a parameter $q \\in [0,1]$ . Inter-client non-i.i.d. [27] however, is even more complex and allows arbitrary and disjoint distributions, which resemble real-world conditions.", + "bbox": [ + 511, + 102, + 893, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Poisoning Attacks Poisoning attacks [51, 54] on DNNs involve attackers altering the training process to manipulate the model. Untargeted attacks lower main task accuracy (MA) by reducing local model accuracy, while targeted (backdoor) attacks embed triggers causing specific misclassifications with minimal impact on MA [3, 18, 34]. Poisoning involves data (e.g., malicious labels or triggers [49]) or model modifications (e.g., altered gradients [28]). This paper mainly considers five specific poisoning types: pixel trigger backdoors [19], blend backdoors [10], edge-case attacks [54], semantic backdoors [3], and random label flipping [27], which are discussed in more detail in Appendix D", + "bbox": [ + 511, + 179, + 893, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptive Attackers Adaptive attackers can bypass FL defenses that focus on fixed metrics by manipulating their local models to evade detection. Bagdasaryan et al. [3] introduced a method that uses an additional loss component that balances backdoor accuracy (BA) and stealthiness to the metrics the defense considers, as shown in Equation (1) [3].", + "bbox": [ + 511, + 359, + 893, + 446 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = \\alpha * L o s s ^ {M A \\& B A} + (1 - \\alpha) * L o s s ^ {A d a p t a t i o n} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 465, + 893, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Siamese Neural Networks Siamese networks [7, 24] are few-shot classifiers that compare embedding vectors from twin networks with shared parameters. Using a distance metric like Euclidean distance [13], they assess similarity between inputs. The Triplet loss [20], shown in Equation (2), is used to minimize distances between the feature vectors of same-class pairs and maximize them for different-class pairs.", + "bbox": [ + 511, + 493, + 893, + 611 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL (a, p, n) = \\max \\left(\\left\\| f (a) - f (p) \\right\\| ^ {2} - \\left\\| f (a) - f (n) \\right\\| ^ {2} + \\gamma , 0\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 631, + 893, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Sibai", + "text_level": 1, + "bbox": [ + 513, + 672, + 578, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present Siamiese Backdoor Inspector (Sibai), a novel defense mechanism against poisoning attacks in FL. Sibai addresses non-i.i.d. scenarios and adaptive attackers by dynamically learning discriminative features of malicious local models using a meta-classifier. This approach avoids the reliance on predefined metrics. By directly operating on flattened model representations, Sibai does not require time-consuming preprocessing and enables detailed analysis that hinders adversarial adaptation. Sibai quickly adapts to federation behavior and new attack strategies with minimal training samples. The neural network architecture adds a layer of opacity, complicating the task for attackers to even discern the specific criteria they must adapt to.", + "bbox": [ + 511, + 698, + 893, + 888 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3788", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Threat Model", + "text_level": 1, + "bbox": [ + 101, + 101, + 240, + 116 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider an FL scenario with a centralized server using FedAVG [36] and a fixed global learning rate. Consistent with prior defenses [1, 5, 6, 27, 35, 37] we assume adversaries control fewer than half of the clients, however, they can fully manipulate their data, training process, and model parameters. Adversaries are aware of the algorithms and defenses deployed on the server. Importantly, Sibai's design, training, and evaluation make no assumptions about local dataset distributions. We evaluate Sibai across i.i.d., 1-class non-i.i.d., and inter-client non-i.i.d. scenarios, the latter closely reflecting real-world FL conditions as described in Sec. 2.", + "bbox": [ + 99, + 125, + 485, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Design", + "text_level": 1, + "bbox": [ + 101, + 311, + 189, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sibai operates on the central server to detect and filter suspicious local models before aggregation. It employs a Siamese network [16, 24] meta-classifier trained to identify dissimilarities between malicious and benign client updates. Sibai consists of two phases: (I) Offline training, where the Siamese network learns from shadow models, including both benign and malicious examples; (II) Online inference, where Sibai evaluates local models against the global model during federation to identify and exclude poisoned models based on learned criteria.", + "bbox": [ + 99, + 334, + 483, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. As shown in Fig. 1, Sibai takes the global model and a local model as input, processing them through four components: Dimensionality Reduction, Twin Networks, Distance Function, and Decision Network. The output is a binary decision indicating the likelihood of a local model being malicious. Detailed descriptions of each component follow in subsequent sections.", + "bbox": [ + 99, + 484, + 485, + 589 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/91ff721698c8f93b3b5ec1f36c6256140e16cda3a97b533f7f90f17c4367aa44.jpg", + "image_caption": [ + "Figure 1. Sibai Architecture" + ], + "image_footnote": [], + "bbox": [ + 130, + 609, + 406, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dimensionality Reduction. Sibai processes models of varying sizes by condensing input data while retaining key features. Local models, often containing hundreds of thousands of parameters, are broken into slices, and Principal Component Analysis (PCA) extracts the most significant components for each slice. PCA identifies and ranks features by variance, ensuring compact yet informative representations. Additional metrics, specifically Cosine and Euclidean distances to the global model, that were shown to be effective by prior research [5, 17, 27, 43], are also incorporated into the input vector. Experiments showed that these metrics consistently improve detection accuracy by up to $10\\%$ . More details about the implementation of the dimensionality reduction pipeline are provided in Appendix C.2.", + "bbox": [ + 511, + 102, + 897, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Twin Networks. The Twin Networks consist of two identical 1D Convolutional Neural Networks (CNNs), serving as the core of the Siamese network. Each CNN processes one of the two pre-processed data samples, acting as a feature extractor. They map the samples to fixed-length embedding vectors using the same filters. These two vectors can then be compared to assess similarity. A high dissimilarity between embedding vectors indicates potential maliciousness, as the twins are explicitly trained to focus on the distinguishing features of malicious local models. Appendix C.4 provides further details about the architecture of the twin models.", + "bbox": [ + 511, + 313, + 893, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distance Function. The embedding vectors generated by the Twin Networks are compared using the $L_{2}$ -distance, which demonstrated superior performance compared to other metrics like $L_{1}$ -distance or cosine distance. This distance is used in the triplet loss function, which is utilized to train the twin networks.", + "bbox": [ + 511, + 481, + 893, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decision Network. To avoid using a fixed distance threshold, a fully connected neural network, called a decision network, is used to classify the output of the distance function. The last layer features a sigmoid activation function [38] that produces a value between 0 and 1, with values closer to 1 indicating a higher probability of maliciousness. We classify local models that when compared to the global model produce an output $>0.5$ as malicious.", + "bbox": [ + 511, + 575, + 895, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Shadow Model Dataset.** Sibai's performance depends on a shadow model dataset used during offline training. This dataset should cover a range of attacks and scenarios to ensure effective generalization by the Siamese Network. To create the shadow model set we simulate FL across various federations with different non-i.i.d. degrees attacks and PDRs. Incorporating stealthy models from adaptive adversaries with low PDR allows Sibai to focus on nuanced criteria while including poisoned models with high PDR enhances generalization. It's crucial to maintain consistent local model architectures in the shadow set, and federations with varied input or output layer sizes cannot be mixed. Further details are provided in Appendix C.3.", + "bbox": [ + 511, + 698, + 895, + 890 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3789", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 101, + 101, + 215, + 116 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section evaluates Sibai during online inference against five poisoning attacks in challenging scenarios, including complex non-i.i.d. settings, low PDRs, and adaptive adversaries.", + "bbox": [ + 99, + 126, + 485, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hardware. Training and evaluation were performed on a Linux server with an Intel Xeon Gold 6342 CPU, 251GB RAM, CUDA [44], and an NVIDIA A30 GPU with 24GB memory.", + "bbox": [ + 101, + 190, + 483, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference Defenses. We compare Sibai with eight defenses: HDBSCAN [35] which represents naive clustering via cosine distance, Krum and M-Krum [5] which select the new global model based on Euclidean distance, MESAS [27] which tries to counter adaptive adversary by using six metrics, FLTrust [8] which uses trusted reference-based weighting, FoolsGold [17] which analyzes the final layer cosine distance, and the robust aggregation method Trimmed Mean/Median [60]. Hyperparameters are detailed in Appendix E. However, it is important to note that a direct comparison of methods with fundamentally different defense paradigms, such as a DF-based method like ours and IR-based defenses like FLTrust, cannot capture the whole picture since these paradigms approach poisoning defense with fundamentally different philosophies.", + "bbox": [ + 99, + 253, + 485, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attacks. Sibai is tested against five attacks: Random label flipping [27], Pixel trigger backdoor [19], Blend backdoor [10], Edge case backdoor [54], and Semantic backdoor [3]. These attacks are introduced in more detail in Appendix D", + "bbox": [ + 99, + 479, + 485, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and Models. We evaluate federations collaboratively training Squeezezenet [23] on CIFAR-10 [30]. Additionally, we conducted experiments with a custom DNN and MNIST, which are provided in Sec. 4.5.", + "bbox": [ + 101, + 556, + 483, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Shadow Model Dataset.** Sibai's offline training uses 520 simulated federations, each with 20 clients (11 benign, 9 malicious), totaling 5720 benign and 4145 malicious models. Scenarios include i.i.d. and non-i.i.d. distributions with varying degrees (e.g., 0.3, 0.6, 0.9 1-class non-i.i.d. and inter-client non-i.i.d. [27]). Realistic and exaggerated settings enhance the distinction between benign and malicious models. Each federation experiences one of the five attacks. Federations with local backdoor accuracies below $45\\%$ were excluded to ensure poisoned models exhibit detectable traits. Distribution details are in Tab. 1 and a detailed breakdown is listed in Appendix H.", + "bbox": [ + 99, + 619, + 485, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Default Federation Parameters. We evaluate Sibai in simulated federations collaboratively training Squeezenet [23] with SGD (learning rate: 0.01, momentum: 0.9, decay: 0.005) and a batch size of 64. Each client trains locally for 10 epochs on i.i.d. datasets with 2560 samples. The global model is initialized with pre-trained weights from", + "bbox": [ + 99, + 801, + 485, + 891 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/aa1c08dad49d9a5bf23598a7f97a157c54ccbca4d45a965e4793b74e60d4385c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryNumber of Federations
IID/non-i.i.d.i.i.d.225
0.3 1-class non-i.i.d.66
0.6 1-class non-i.i.d.71
0.9 1-class non-i.i.d.96
inter-client non-i.i.d. [27]62
AttacksRandom [27]90
Pixel [19]91
Blend [10]119
Edge [54]121
Semantic [3]99
Total Number of Federations520
", + "bbox": [ + 566, + 99, + 843, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Composition of the shadow model dataset", + "bbox": [ + 555, + 224, + 849, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PyTorch [46] and undergoes 49 benign rounds to mitigate early backdoor persistence [3], followed by poisoning in the 50th round.", + "bbox": [ + 511, + 265, + 893, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sibai Parameters. During offline training, the twin and decision networks are trained separately for 25 epochs using SGD. The twin networks use a learning rate of 0.001, while the decision network uses 0.005. Model slices contain up to 20,000 parameters, with larger layers partitioned. PCA retains 200 components per slice. Siamese network training does not use batch processing due to memory constraints, while the decision network processes batches of 16 distances derived from 8 Siamese network triples.", + "bbox": [ + 511, + 311, + 895, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Metrics. We evaluate Sibai with precision, recall, and F1 score. True positives are correctly flagged malicious models, and true negatives are benign models. Precision measures the ratio of correctly identified malicious models to all flagged models, while recall assesses the proportion of flagged malicious models among all actual malicious models. The F1 score balances these metrics. Additionally, we track the global model's BA to evaluate attack mitigation and MA changes to assess benign scenario impact. The MA diff. represents the difference between the MA after Sibai application and perfect benign aggregation.", + "bbox": [ + 511, + 448, + 895, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. IID Scenarios", + "text_level": 1, + "bbox": [ + 511, + 619, + 651, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section evaluates Sibai's online performance after training on the shadow model set. We test Sibai on 12 unseen federation scenarios1 of increasing difficulty and adversary stealthiness. In the baseline scenario (Tab. 2), federations training Squeezenet [23] on CIFAR-10 [30] in an i.i.d. setting were attacked with the five poisoning attacks. Adversaries employed a PDR of 0.5, which is a relatively overt attack. Sibai flawlessly detected all malicious clients without false positives. Scenario (1) increased stealthiness by minimizing PDRs while maintaining local BAs above $45\\%$ . The blend [10] and semantic [3] backdoors used a PDR of 0.04 due to their stronger effects, while the edge case backdoor [54] required a PDR of 0.16 for consistent backdoor introduction. These adjustments mirror real-world attacker strategies. Sibai achieved perfect recall and F1 scores for", + "bbox": [ + 511, + 641, + 895, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Each scenario is designated by a circled number", + "bbox": [ + 529, + 875, + 785, + 888 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3790", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7c222052bd28178d3012fa30b32023427b74f05a8fd3e5f515f48a9985424ea0.jpg", + "table_caption": [], + "table_footnote": [ + "$\\boxed{0} =$ Default + high PDRs", + "$\\widetilde{\\mathbf{1}} =$ Default and low PDRs" + ], + "table_body": "
ScenarioRandomPixelBlendEdgeSemantic
①PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.09%0.20%18.87%0.0%
①PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall0.88881.00001.00001.00000.7777
F10.94111.00001.00001.00000.8750
MA diff.-0.88%0.0%0.0%0.0%+4.12%
BA-10.43%0.0%18.87%0.0%
", + "bbox": [ + 148, + 99, + 436, + 218 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d153e02133e41716479ea73f65d9636deb0b53a3f3968d196bfd7ba9bc287b1.jpg", + "table_caption": [ + "Table 2. Detection performance of Sibai in i.i.d. scenarios" + ], + "table_footnote": [ + "$(2) =$ Default $+0.61$ -class non-i.i.d. $^+$ high PDR", + "$\\widehat{(3)} =$ Default $+0.61$ -class non-i.i.d. $^+$ low PDR", + "$(4) =$ Default $+0.9$ 1-class non-i.i.d. $^+$ high PDR", + "$(5) =$ Default + 0.91-class non-i.i.d. + low PDR" + ], + "table_body": "
qScenarioRandomPixelBlendEdgeSemantic
0.6②PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-2.65%0.20%6.12%0.0%
③PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00000.88881.0000
F11.00001.00001.00000.94111.0000
MA diff.0.0%0.0%0.0%-6.71%0.0%
BA-2.65%0.21%8.67%0.0%
0.9④PDR0.50.50.50.50.5
Precision0.87501.00001.00000.90001.0000
Recall0.77770.88881.00001.00001.0000
F10.82350.94111.00000.94731.0000
MA diff.+4.80%+1.08%0.0%+2.44%0.0%
BA-0.11%0.0%0.0%0.0%
⑤PDR0.40.080.040.160.04
Precision0.81810.75001.00000.81811.0000
Recall1.00001.00001.00001.00000.8888
F10.90000.85711.00000.90000.9411
MA diff.+0.48%-4.26%0.0%-1.65%+0.48%
BA-0.0%0.21%0.0%0.0%
", + "bbox": [ + 135, + 282, + 450, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pixel trigger and edge case backdoors, identifying all malicious clients. In the random label flip and semantic backdoor attacks, one and two malicious clients, respectively, evaded detection. However, Sibai identified over $77\\%$ of malicious clients in all cases, preventing effective poisoning. Across all i.i.d. scenarios, Sibai maintained BA below $19\\%$ , ensuring attacks were unsuccessful. Interestingly, in the semantic backdoor scenario, Sibai mistakenly aggregated two malicious models, resulting in a slight MA improvement over perfect aggregation. This occurred because the aggregated models provided additional knowledge, but the low PDR and limited bypassed defenses still prevented backdoor establishment.", + "bbox": [ + 99, + 619, + 485, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Non-IID Scenarios", + "text_level": 1, + "bbox": [ + 101, + 821, + 277, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate Sibai in challenging non-i.i.d. scenarios, including varying degrees of 1-class non-i.i.d. and inter-client non-i.i.d. setups as described in Sect. 2 and [27].", + "bbox": [ + 99, + 844, + 485, + 890 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ccdaa6f881e7556ae921d17fa79c60f81fb17c8d4d7b27961f87afc07df1e868.jpg", + "table_caption": [ + "Table 3. Detection performance of Sibai in 1-class non-i.i.d. scenarios, with different level of disproportionality $q$" + ], + "table_footnote": [ + "$\\widehat{(6)} =$ Default $^+$ inter-client non-i.i.d. $^+$ high PDR", + "$\\widetilde{7} =$ Default + inter-client non-i.i.d. + low PDR" + ], + "table_body": "
ScenarioRandomPixelBlendEdgeSemantic
6PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00000.9000
Recall0.77770.88881.00001.00001.0000
F10.87500.94111.00001.00000.9473
MA diff.+2.61%+0.86%0.0%0.0%0.0%
BA-4.03%0.0%0.0%0.0%
7PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.66661.00000.8888
F11.00000.94110.80001.00000.9411
MA diff.0.0%+1.38%-1.43%0.0%-1.48%
BA-13.57%0.0%0.0%0.0%
", + "bbox": [ + 560, + 99, + 848, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4. Detection performance of Sibai in inter-client non-i.i.d. scenarios [27]", + "bbox": [ + 511, + 253, + 895, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "0.6 1-class non-i.i.d. For the scenarios $\\boxed{2}$ and $\\boxed{3} q = 0.6$ was used. Sibai achieved perfect detection in the 0.5 PDR case $\\boxed{2}$ . For the low PDRs scenario $\\boxed{3}$ , F1 scores were 1.0 for all attacks except the edge case backdoor, where one malicious client evaded detection. Despite this, the MA dropped by less than $7\\%$ , and BA stayed below $9\\%$ , ensuring no effective backdoor was introduced.", + "bbox": [ + 511, + 306, + 893, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "0.9 1-class non-i.i.d. In the more extreme 0.9 1-class non-i.i.d. scenarios (4) and (5) in Tab. 10, Sibai occasionally flagged benign models, with three false positives in the worst-case pixel trigger backdoor scenario (5). However, these errors had minimal impact on MA. Sibai consistently eliminated enough malicious models to prevent effective backdoors, as no attack achieved a BA higher than $1\\%$ .", + "bbox": [ + 511, + 419, + 895, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inter-client non-i.i.d. Inter-client non-i.i.d. features erratic and highly unbalanced data distributions, reflecting real-world scenarios. In the 0.5 PDR scenario (6), Sibai achieved perfect detection for blend and edge case backdoors, perfect precision for untargeted and pixel attacks, and perfect recall for the semantic attack. Only one false positive occurred during the semantic attack. The lowest F1 score of 0.8750 occurred during the untargeted attack) In the attacker-adapted low PDR scenario (7), Sibai maintained perfect precision for all attacks despite high non-i.i.d.-ness. Recall was lower, yielding F1 scores of 0.9411 for pixel and semantic backdoors and 0.8000 for the blend attack. Nonetheless, MA was minimally affected, and BA remained low across all attacks. Overall, Sibai demonstrated robust performance in 1-class and inter-client non-i.i.d. scenarios, with minimal false positives and no effective backdoors.", + "bbox": [ + 511, + 527, + 895, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Adaptive Adversaries", + "text_level": 1, + "bbox": [ + 511, + 779, + 709, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sibai incorporates slice-wise Euclidean and cosine distances to the global model into its input vectors to enhance detection. While effective, these metrics may be vulnerable to adaptive attackers. To test Sibai's robustness, we evaluated four scenarios involving attackers using Bagdasaryan et al.'s [3] adaptation method with an $\\alpha$ parameter balancing adaptation", + "bbox": [ + 511, + 800, + 895, + 890 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3791", + "bbox": [ + 483, + 926, + 511, + 938 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4d56aadb72bfaa074f10b5bab66d7f9d8e6bbb10b19c050cf8c4d117ffd762b6.jpg", + "table_caption": [], + "table_footnote": [ + "$\\boxed{8} =$ Default $^+$ cosine distance adaption with $0.3\\alpha$ by mal. clients $^+$ high PDR", + "9 = Default + cosine distance adaption with 0.3 α by mal. clients + low PDR", + "$\\boxed{10} =$ Default $^+$ Euclidean distance adaption with $0.3\\alpha$ by mal. clients $^+$ high PDR", + "$\\widehat{(11)} =$ Default $^+$ Euclidean distance adaption with $0.3\\alpha$ by mal. clients $^+$ low PDR" + ], + "table_body": "
Form of AdaptationScenarioRandomPixelBlendEdgeSemantic
Cosine dist.⑧PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18.87%0.0%
⑨PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001,00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.14%18.87%0.0%
Euclidian dist.⑩PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18. 87%0.0%
⑪PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.88881.00001.0000
F11.00000.94110.94111.00001.0000
MA diff.0.0%+3.1%-6.3%0.0%0.0%
BA-5.98%0.0%18.87%0.0%
", + "bbox": [ + 106, + 99, + 482, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and task losses. Results for $\\alpha = 0.3$ , shown in Tab. 5, highlight Sibai's resilience. In cosine-adapted scenarios, Sibai achieved perfect F1 scores (1.0), showing immunity to such attacks. Similarly, Sibai maintained perfect performance in Euclidean-adapted Scenario 12. In low-PDR Euclidean-adapted scenarios, Sibai showed flawless precision, with one false negative for pixel and blend backdoors. However, no attack established an effective backdoor, with BA remaining below $19\\%$ . Overall, Sibai demonstrates strong resilience against adaptive attacks, proving it does not overly rely on engineered features. Its ability to internalize specific patterns makes it impractical for attackers to adapt using Bagdasaryan et al.'s method.", + "bbox": [ + 99, + 443, + 483, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison to Existing Defenses", + "text_level": 1, + "bbox": [ + 101, + 652, + 383, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the effectiveness of Sibai compared to metric-based poisoning detection methods in FL, we also evaluate existing defenses in some of our previous scenarios. Due to space limitations, we focus on a selection of the most representative scenarios, including scenario 1 as the i.i.d. baseline, scenario 4 as the highest degree of 1-class non-i.i.d., and scenario 6 to showcase inter-client non-i.i.d. We also include scenario 0 as the overall simplest scenario and scenario 11 to cover adaptive adversaries. Since some reference defenses rely on modified aggregation methods instead of strictly flagging and filtering out malicious models, calculating F1 scores for all reference defenses faced challenges. Further details are provided in Appendix F.", + "bbox": [ + 99, + 675, + 485, + 890 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec105ded48e714f308d9479917108288dc229f4bb675aedf979e026607c15543.jpg", + "table_caption": [ + "Table 5. Detection performance of Sibai against adaptive attackers in i.i.d. scenarios" + ], + "table_footnote": [], + "table_body": "
ScenarioDefenseRandomPixelBlendEdgeSemantic
0PDR0.50.50.50.50.5
MESAS0.90000.90000.90001.00000.9473
Krum0.57140.57140.57890.57140.5714
M-Krum0.26080.26080.64280.26080.2608
HDBSCAN1.00000.00000.11111.00001.0000
FLTrust (0.09)0.00000.26080.33330.00000.2608
Sibai (ours)1.00001.00001.00001.00001.0000
1PDR0.40.080.040.160.04
MESAS0.94730.00000.18180.16660.3076
Krum0.57140.64280.64280.57140.5714
M-Krum0.26080.60860.60860.26080.4347
HDBSCAN0.00000.55550.44440.00000.1176
FLTrust (0.09)0.09520.66660.57140.51850.5714
Sibai (ours)0.94111.00001.00001.00000.8750
4PDR0.50.50.50.50.5
MESAS0.00000.16660.80000.94730.9000
Krum0.57140.57140.57140.64280.6428
M-Krum0.26080.69560.69560.78260.7826
HDBSCAN0.00000.40000.77770.82350.8235
FLTrust (0.09)0.57140.88880.88880.88880.8571
Sibai (ours)0.82350.94111.00000.94731.0000
6PDR0.50.50.50.50.5
MESAS0.64280.47610.21050.66660.3157
Krum0.57140.57140.57140.57140.5714
M-Krum0.26080.26080.26080.52170.3478
HDBSCAN0.00000.00000.14280.30760.1428
FLTrust (0.09)0.00000.57140.57140.57140.6206
Sibai (ours)0.87500.94111.00001.00000.9473
11PDR0.40.080.040.160.04
MESAS1.00000.66660.90000.61530.9473
Krum0.57140.57140.64280.64280.6428
M-Krum0.26080.26080.78260.60860.6956
HDBSCAN0.00001.00001.00001.00000.2000
FLTrust (0.09)0.57140.62060.66660.62060.6206
Sibai (ours)1.00000.94110.94111.00001.0000
", + "bbox": [ + 517, + 99, + 839, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6. F1 score comparison between Sibai and current FL poisoning defenses", + "bbox": [ + 511, + 431, + 895, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 6, Sibai achieved superior F1 scores across most scenarios. In high PDR i.i.d. scenario $\\boxed{0}$ , it matched or outperformed all defenses. In low PDR i.i.d. scenario $\\boxed{1}$ , Sibai achieved perfect F1 scores for pixel, blend, and edge case backdoors, outperforming MESAS for these attacks. Even when MESAS maintained low BA, it highlighted the fragility of backdoor attacks in low PDR settings. For 1-class non-i.i.d. scenario $\\boxed{4}$ , Sibai significantly outperformed other methods, and in inter-client non-i.i.d. scenario $\\boxed{7}$ , its worst-case F1 score of 0.8750 surpassed the best scores of other defenses. In adaptive scenario $\\boxed{11}$ , Sibai slightly trailed HDBSCAN for pixel and blend backdoors however, it excelled in the detection of the semantic and untargeted attacks, achieving perfect F1 scores. Overall, Sibai consistently outperformed existing methods in identifying malicious models across diverse scenarios and attacks.", + "bbox": [ + 511, + 488, + 895, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 7, also considered methods for which the F1 was unavailable, specifically FoolsGold [17] and trimmed mean/median [60]. Sibai showed performance close to perfect benign aggregation, effectively preventing poisoning without significant MA reduction. Notably, Sibai and MESAS achieved low BA in challenging scenarios (4) and (6), where other defenses failed to stop backdoors. Although FLTrust [8], produced higher MA in non-i.i.d. scenarios by using a weighted aggregation that particularly emphasizes the old global model, it also often allowed higher BA and", + "bbox": [ + 511, + 739, + 895, + 890 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3792", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e546f7c4adcca3f0f93a0056a8fd94c666db7c028bac4390b559939e9c0f692.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScenarioDefenseRandom MA diff.Pixel MA diff.BABlend MA diff.BAEdge MA diff.BASemantic MA diff.BA
①PDR0.40.080.040.160.04
benign agg.0.00.010.40.00.00.018.80.00.0
MESAS [27]-0.1+5.19.50.020.4+0.27.10.00.0
Krum [5]-35.0-43.20.0-14.10.0-35.50.0-21.1100.0
M-Krum [5]-24.5-8.512.5-9.52.9-16.30.0-11.40.0
FoolsGold [17]-6.3+3.011.0-1.718.8-8.36.1+0.10.0
HDBSCAN [35]-16.0-1.910.9-6.511.4-13.80.0-5.80.0
Trimmed mean [60]-8.3+3.510.2-2.512.8-9.06.6-0.60.0
Trimmed median [60]-22.3-2.65.2-11.60.1-14.72.5-6.40.0
FLTrust [8]-22.3-0.117.2-5.213.0+1.516.8-4.80.0
Sibai (ours)-0.80.010.40.00.00.018.80.00.0
④PDR0.40.50.50.50.5
benign agg.0.00.00.00.00.00.00.00.00.0
MESAS [27]-5.7+0.334.3+5.00.6+2.40.0+10.60.0
Krum [5]-7.7-6.1100.0-0.2100.0-4.80.0-5.30.0
M-Krum [5]-2.1+5.51.7+1.0100.0-5.10.0+4.00.0
FoolsGold [17]-2.0-5.030.2-0.2100.0-10.80.0-7.30.0
HDBSCAN [35]-7.5+1.013.2+1.40.0-8.70.0-0.920.0
Trimmed mean [60]-5.3-0.428.9-0.2100.0-11.10.0-7.840.0
Trimmed median [60]-6.6-5.10.0-0.4100.0-10.40.0-6.50.0
FLTrust [8]+19.5+14.753.1+10.898.2+4.80.0+11.120.0
Sibai (ours)+4.8+1.00.00.00.0+2.40.00.00.0
⑥PDR0.50.50.50.50.5
benign agg.0.013.92.90.02.30.02.40.00.0
MESAS [27]+11.9-0.31.6-0.132.7+9.70.00.00.0
Krum [5]+4.8-3.9100.0-5.1100.0+16.885.7+18.780.0
M-Krum [5]-1.6-3.085.3-0.3100.0+2.90.00.00.0
FoolsGold [17]-1.6+1.57.0-5.0100.0-2.00.00.00.0
HDBSCAN [35]-5.8-0.719.0+3.899.9-0.30.0-0.40.0
Trimmed mean [60]-0.5+1.17.3-3.8.1100.0-3.00.00.00.0
Trimmed median [60]-2.5-3.50.0-4.9100.0-2.30.0-0.30.0
FLTrust [8]+14.7+14.347.4+13.099.8+14.00.0+8.320.0
Sibai (ours)+2.6+0.84.00.00.00.00.00.00.0
", + "bbox": [ + 248, + 99, + 750, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "therefore failed to prevent effective backdoors. In contrast, Sibai excelled at malicious model detection and consistently maintained low BA while achieving MAs aligned with benign aggregation, making Sibai an overall reliable defense.", + "bbox": [ + 101, + 474, + 483, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Alternative Dataset and DNN Architecture", + "text_level": 1, + "bbox": [ + 101, + 546, + 455, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate that Sibai is not limited to SqueezeNET or the CIFAR-10 dataset, we conducted additional experiments using a custom CNN and the MNIST dataset [12]. The CNN model consists of 2 convolutional layers, 2 max-pooling layers, and 3 linear layers. We excluded the dataset-specific semantic and edge case backdoors and instead included a label swap attack [27], which targets samples of one label and swaps them to another to reduce the global model's accuracy. The shadow-model dataset for these experiments comprised 1401 federations. The F1 scores in Tab. 8 indicate that Sibai effectively detected malicious models, and therefore that the approach is not specific to SqueezeNET or CIFAR-10.", + "bbox": [ + 99, + 569, + 485, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Multiple Poisonings in one Model", + "text_level": 1, + "bbox": [ + 101, + 761, + 398, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate Sibai's effectiveness in detecting models poisoned with multiple backdoors simultaneously, we conducted experiments using federations where each malicious local model contained both a blend backdoor [10] and a pixel trigger backdoor [19] simultaneously. For those local models, the PDR quantifies the ratio of the total poisoned samples, i.e. the sum used for both backdoors, to the com", + "bbox": [ + 99, + 786, + 485, + 890 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/04e20f685edff62e5865dcb50d33b225fc367c6df622c5354d9fb5bf95068070.jpg", + "table_caption": [ + "Table 7. Comparison of MA diff. and BA (both in percentage) between Sibai and existing FL poisoning defenses. The best value for each attack in each scenario is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
PDRDistributionRandom [27]Pixel [19]Blend [10]Label Swap [27]
0.4i.i.d.1.00001.00001.00001.0000
0.40.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.40.6 l-class non-i.i.d.0.69230.72000.72000.7200
0.40.9 l-class non-i.i.d.0.87500.94111.00001.0000
0.4inter-client non-i.i.d.0.90000.94110.94730.9000
0.5i.i.d.1.00001.00001.00001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.50.6 l-class non-i.i.d.0.72000.66660.72000.7200
0.50.9 l-class non-i.i.d.1.00000.94111.00001.0000
0.5inter-client non-i.i.d.0.94730.94110.94730.9000
", + "bbox": [ + 521, + 472, + 885, + 575 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d00c14f182a6e1c8da69f9b630ba02a00401c2fd8eeef081cbfb99cca2bd3734.jpg", + "table_caption": [ + "Table 8. F1 score of Sibai in federations training a small CNN with the MNIST dataset" + ], + "table_footnote": [], + "table_body": "
PDRDistributionPrecisionRecallF1
0.2i.i.d.1.00001.0001.0000
0.20.3 l-class non-i.i.d.1.00001.00001.0000
0.20.6 l-class non-i.i.d.1.00001.00001.0000
0.20.9 l-class non-i.i.d.1.00000.88880.9411
0.2inter-client non-i.i.d.1.00000.88880.9411
0.5i.i.d.1.00001.0001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.0000
0.50.6 l-class non-i.i.d.1.00001.00001.0000
0.50.9 l-class non-i.i.d.1.00001.00001.0000
0.5inter-client non-i.i.d.1.00001.00001.0000
", + "bbox": [ + 576, + 631, + 831, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. Detection performance of Sibai against local models poisoned with two backdoors (pixel trigger backdoor [19] and blend backdoor [10]) simultaneously in different scenarios", + "bbox": [ + 511, + 744, + 895, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plete local model dataset. In our implementation of this multi-backdoor half of the poisoned samples were used to introduce the pixel trigger backdoor, while the other half was used to apply the blend backdoor. Importantly, our shadow model dataset, used to train Sibai, did not include", + "bbox": [ + 511, + 816, + 895, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3793", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "any models with such multi-backdoors. As illustrated in Tab. 9, Sibai consistently identified local models containing multiple backdoors across all tested scenarios.", + "bbox": [ + 99, + 102, + 485, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related Work", + "text_level": 1, + "bbox": [ + 101, + 159, + 241, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section provides a brief overview of relevant literature. First, Sect. 6.1 outlines poisoning defenses tailored for FL. Then, Sect. 6.2 discusses machine learning methods for general offline poisoning analysis.", + "bbox": [ + 99, + 183, + 485, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Defenses against Poisoning Attacks in Federated Learning", + "text_level": 1, + "bbox": [ + 101, + 251, + 483, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There exists a diverse range of techniques that employ fixed metrics to detect poisoning attacks in FL. FoolsGold [17] uses a client-specific update history which is compared to the update histories of other clients using the cosine similarity. Krum [5] selects a new global model based on the Euclidean distance to other local models. M-Krum [5] extends this method to select multiple models simultaneously. Trimmed-Mean/Median [60] are robust aggregation methods that replace FedAVG [36] and calculate the updated global model by using coordinate-wise means or medians from local models. FLAME [43] clusters models using HDBSCAN [35] and filters out smaller clusters. It also applies differential privacy techniques, weight clipping, and noisng. DeepSight [47] combines filtering with differential privacy. It relies on two metrics: The cosine distance between models and values extracted from the output layer of the model. Auror [48] uses K-Means clustering on model parameters, however, unfortunately, it exhibited an extreme performance overhead in our experiments. FLTrust [8] uses a weighted aggregation system for which it assigns weights to local models based on their cosine and Euclidean distances to a benign reference model, which is trained server-side on a small trusted dataset. FLDetector [63] is a historical update-based defense that mathematically predicts client updates and compares the predictions to the actual updates. MESAS [27] was specifically designed to operate in non-i.i.d. scenario and counter adaptive adversaries and employs six metrics. These approaches differ from Sibai as they rely on fixed metrics or criteria, making them susceptible to adaptive attacks [27]. Defenses focusing on specific layers, like FoolsGold [17], are prone to adversaries exploiting advanced model poisoning techniques specifically targeting those layers. Similarly, methods that heavily focus on cosine or Euclidean distance are often vulnerable to adaptation methods using additional losses [3, 27], as detailed in Sect. 2. These rigid metrics also face challenges in non-i.i.d. scenarios, especially in inter-client non-i.i.d. settings, where it has been demonstrated that most defenses, except MESAS [27] which was specifically designed to handle such scenarios, face difficulties. In contrast, Sibai dynamically learns criteria based on its shadow model dataset, making it challenging for attackers", + "bbox": [ + 99, + 287, + 485, + 891 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to adapt without access to the Siamese Network. Sibai can also differentiate between variations in local models due to data distribution and those caused by poisoning.", + "bbox": [ + 511, + 102, + 895, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Non-FL Specific Offline Poisoning Detection", + "text_level": 1, + "bbox": [ + 511, + 160, + 877, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section discusses several machine-learning-based detection methods against poisoning attacks, not specifically designed for FL. Xu et al. [58] propose Meta Neural Trojan Detection (MNTD), a black-box backdoor detection approach utilizing a meta-classifier that assesses if a model's output behavior on a given input aligns more closely with a set of known benign models or a set of known malicious models. Huang et al. [21] introduce the one-pixel signature concept for DNNs, consisting of a set of single-pixel adversarial perturbations that have the highest probability of triggering misclassifications on a specific benign image dataset. Subsequently, a meta-classifier was trained on these signatures to discern between poisoned and benign models. Kolouri et al. [25] use universal litmus patterns and a meta-classifier to detect backdoors across different network architectures. Chen et al. [9] present DeepInspect, a black-box method utilizing model inversion and anomaly detection for backdoor detection. Huang et al. [22] also introduced Neuroninspect, employing explanation heatmaps and outlier detection for identifying abnormal neurons. Unlike these methods, Sibai is tailored to FL, avoiding intensive preprocessing steps. DeepInspect [9] involves extensive reverse engineering of potential triggers for every label of each local model, and Huang et al. [21] requires testing every local model with a large set of single-pixel adversarial perturbations. Conducting such intricate analyses for every model in a federation during each round is impractical, as these methods were originally designed for detailed offline analysis of individual models. Sibai instead only uses PCA as the sole dimensionality reduction mechanism, facilitating straightforward inference.", + "bbox": [ + 511, + 183, + 898, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 657, + 629, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To counter non-i.i.d. data distribution and adaptive adversaries in FL, we introduce Siamese Backdoor Inspector (Sibai), a novel FL poisoning defense mechanism utilizing a Siamese network meta-classifier. Inspired by successful offline backdoor detection [21, 25, 58], Sibai dynamically learns discriminative features directly from PCA decompositions of local models. In diverse evaluations, Sibai consistently achieves high detection rates without significantly impacting global model performance. It effectively prevents backdoor establishment and outperforms current defenses in terms of f1 score while maintaining low false positive rates. Future work may explore selectively applying influence reduction to models identified by Sibai for targeted defense.", + "bbox": [ + 511, + 683, + 898, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3794", + "bbox": [ + 483, + 926, + 514, + 939 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgment", + "text_level": 1, + "bbox": [ + 101, + 101, + 269, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research has been funded by the Federal Ministry of Education and Research of Germany (BMBF) within the program Digital. Sicher. Souverän. \"in the project\"Erkennung von Angriffen gegen IoT-Netzwerke in Smart Homes - IoTGuard\" (project number 16KIS1919).", + "bbox": [ + 99, + 126, + 485, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 218, + 196, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sebastien Andreina, Giorgia Azzurra Marson, Helen Möllering, and Ghassan Karame. Baffle: Backdoor detection via feedback-based federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), pages 852-863. IEEE, 2021. 1, 3", + "[2] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. In 30th USENIX Security Symposium (USENIX Security 21), pages 1505-1521, 2021. 1", + "[3] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 4, 5, 8, 13, 14, 15, 16, 17, 18, 19", + "[4] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International Conference on Machine Learning, pages 634-643. PMLR, 2019. 1", + "[5] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 3, 4, 7, 8, 15, 16", + "[6] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 3", + "[7] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in neural information processing systems, 6, 1993. 2", + "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. Network and Distributed Systems Security (NDSS) Symposium 2021, 2021. 1, 4, 6, 7, 8, 15, 16", + "[9] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In *IJCAI*, page 8, 2019. 1, 8", + "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 4, 7, 13, 14, 17, 18, 19", + "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual computer security applications conference, pages 554-569, 2021. 1" + ], + "bbox": [ + 104, + 243, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 7", + "[13] Elena Deza, Michel Marie Deza, Michel Marie Deza, and Elena Deza. Encyclopedia of distances. Springer, 2009. 2", + "[14] Cynthia Dwork et al. The mathematics of information coding, extraction, and distribution. The IMA Volumes in Mathematics and its applications, 107:31-47, 1999. 12", + "[15] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 1", + "[16] Hossein Fereidooni, Jan König, Phillip Rieger, Marco Chilese, Bora Gökbakan, Moritz Finke, Alexandra Dmitrienko, and Ahmad-Reza Sadeghi. Authentisense: A scalable behavioral biometrics authentication scheme using few-shot learning for mobile platforms. arXiv preprint arXiv:2302.02740, 2023. 3", + "[17] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2020), pages 301-316, 2020. 1, 3, 4, 6, 7, 8, 15, 16", + "[18] Yansong Gao, Bao Gia Doan, Zhi Zhang, Siqi Ma, Jiliang Zhang, Anmin Fu, Surya Nepal, and Hyoungshick Kim. Backdoor attacks and countermeasures on deep learning: A comprehensive review. arXiv preprint arXiv:2007.10760, 2020. 1, 2", + "[19] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 2, 4, 7, 13, 17, 18, 19", + "[20] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-Based Pattern Recognition: Third International Workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 2", + "[21] Shanjiaoyang Huang, Weiqi Peng, Zhiwei Jia, and Zhuowen Tu. One-pixel signature: Characterizing cnn models for backdoor detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 326-341. Springer, 2020. 1, 8", + "[22] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 1, 8", + "[23] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and i 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 4", + "[24] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop. Lille, 2015. 2, 3", + "[25] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 1, 8" + ], + "bbox": [ + 516, + 103, + 895, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3795", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jakub Konečný, H Brendan McMahan, Daniel Ramage, and Peter Richtárik. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint arXiv:1610.02527, 2016. 2", + "[27] Torsten Krauß and Alexandra Dmitrienko. Mesas: Poisoning defense for federated learning resilient against adaptive attackers. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, page 1526-1540, New York, NY, USA, 2023. Association for Computing Machinery. 1, 2, 3, 4, 5, 7, 8, 13, 15, 16, 17, 18, 19", + "[28] Torsten Krauß, Raphael Götz, and Alexandra Dmitrienko. Security of nvme offloaded data in large-scale machine learning. In Computer Security - ESORICS 2023, pages 143-163, Cham, 2024. Springer Nature Switzerland. 2", + "[29] Torsten Krauß, Jan König, Alexandra Dmitrienko, and Christian Kanzow. Automatic Adversarial Adaption for Stealthy Poisoning Attacks in Federated Learning. Network and Distributed Systems Security (NDSS) Symposium 2024, 2024. 1", + "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 13, 14, 15", + "[31] Liping Li, Wei Xu, Tianyi Chen, Georgios B Giannakis, and Qing Ling. Rsa: Byzantine-robust stochastic aggregation methods for distributed learning from heterogeneous datasets. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1544–1551, 2019. 1", + "[32] Yijing Li, Xiaofeng Tao, Xuefei Zhang, Junjie Liu, and Jin Xu. Privacy-preserved federated learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(7):8423-8434, 2021. 1", + "[33] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1", + "[34] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 2", + "[35] Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017. 3, 4, 7, 8, 15, 16", + "[36] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 8, 15", + "[37] Luis Muñoz-González, Kenneth T Co, and Emil C Lupu. Byzantine-robust federated machine learning through adaptive model averaging. arXiv preprint arXiv:1909.05125, 2019. 3", + "[38] Sridhar Narayan. The generalized sigmoid activation function: Competitive supervised learning. Information sciences, 99 (1-2):69-82, 1997. 3, 13", + "[39] Mohammad Naseri, Jamie Hayes, and Emiliano De Cristofaro. Local and central differential privacy for robustness and privacy in federated learning. arXiv preprint arXiv:2009.03561, 2020. 1" + ], + "bbox": [ + 102, + 103, + 485, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Blaine Nelson, Marco Barreno, Fuching Jack Chi, Anthony D Joseph, Benjamin IP Rubinstein, Udam Saini, Charles Sutton, J Doug Tygar, and Kai Xia. Exploiting machine learning to subvert your spam filter. LEET, 8(1-9):16-17, 2008. 1", + "[41] Anh Nguyen, Tuong Do, Minh Tran, Binh X Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, and Quang D Tran. Deep federated learning for autonomous driving. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1824-1830. IEEE, 2022. 1", + "[42] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), pages 1-7, 2020. 1", + "[43] Thien Duc Nguyen, Phillip Rieger, Roberta De Viti, Huili Chen, Björn B Brandenburg, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, et al. {FLAME}: Taming backdoors in federated learning. In 31st USENIX Security Symposium (USENIX Security 22), pages 1415–1432, 2022. 1, 3, 8", + "[44] NVIDIA, Péter Vingelmann, and Frank H.P. Fitzek. Cuda, release: 10.2.89, 2020. 4", + "[45] Xudong Pan, Mi Zhang, Beina Sheng, Jiaming Zhu, and Min Yang. Hidden trigger backdoor attack on {NLP} models via linguistic style manipulation. In 31st USENIX Security Symposium (USENIX Security 22), pages 3611-3628, 2022. 1", + "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 4, 12", + "[47] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. NNetwork and Distributed Systems Security (NDSS) Symposium 2022, 2022. 1, 8", + "[48] Shiqi Shen, Shruti Tople, and Prateek Saxena. Auror: Defending against poisoning attacks in collaborative deep learning systems. In Proceedings of the 32nd Annual Conference on Computer Security Applications, pages 508-519, 2016. 1, 8, 16", + "[49] Gan Sun, Yang Cong, Jiahua Dong, Qiang Wang, Lingjuan Lyu, and Ji Liu. Data poisoning attacks on federated machine learning. IEEE Internet of Things Journal, 9(13):11365-11375, 2021. 2", + "[50] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963, 2019. 1", + "[51] Zhiyi Tian, Lei Cui, Jie Liang, and Shui Yu. A comprehensive survey on poisoning attacks and countermeasures in machine learning. ACM Computing Surveys, 55(8):1-35, 2022. 2", + "[52] Alexander Turner, Dimitris Tsipras, and Aleksander" + ], + "bbox": [ + 514, + 103, + 895, + 888 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3796", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1", + "[53] Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual. CreateSpace, Scotts Valley, CA, 2009. 12", + "[54] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. Advances in Neural Information Processing Systems, 33:16070-16084, 2020. 1, 2, 4, 13, 14, 17, 18, 19", + "[55] Zhaoxian Wu, Qing Ling, Tianyi Chen, and Georgios B Giannakis. Federated variance-reduced stochastic gradient descent with robustness to byzantine attacks. IEEE Transactions on Signal Processing, 68:4583-4596, 2020. 1", + "[56] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations, 2019. 1", + "[57] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1", + "[58] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. In 2021 IEEE Symposium on Security and Privacy (SP), pages 103-120. IEEE, 2021. 1, 8", + "[59] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 2", + "[60] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 4, 6, 7, 8, 15, 16", + "[61] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 1", + "[62] Hongyi Zhang, Jan Bosch, and Helena Holmström Olsson. End-to-end federated learning for autonomous driving vehicles. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 1", + "[63] Zaixi Zhang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Fl detector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2545-2555, 2022. 1, 8" + ], + "bbox": [ + 104, + 103, + 483, + 760 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3797", + "bbox": [ + 483, + 928, + 513, + 938 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_model.json b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6885c17a38adb019e5b0473a588b97d0311bae37 --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.055 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.141, + 0.895, + 0.162 + ], + "angle": 0, + "content": "Sibai: A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.19, + 0.701, + 0.206 + ], + "angle": 0, + "content": "Melanie Gotoz, Torsten Krauß, Alexandra Dmitrienko" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.208, + 0.59, + 0.224 + ], + "angle": 0, + "content": "University of Wurzburg" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.226, + 0.658, + 0.242 + ], + "angle": 0, + "content": "Sanderring 2, 97070 Wurzburg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.245, + 0.772, + 0.259 + ], + "angle": 0, + "content": "melanie.goetz@uni-wuerzburg.de, torsten.krauss@uni-wuerzburg.de," + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.262, + 0.654, + 0.275 + ], + "angle": 0, + "content": "alexandra.dmitrienko@uni-wuerzburg.de" + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.31, + 0.331, + 0.324 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.34, + 0.486, + 0.635 + ], + "angle": 0, + "content": "Federated Learning (FL) enables collaborative machine learning across decentralized clients without sharing raw data, which offers enhanced privacy and improved performance. However, FL is vulnerable to poisoning attacks, compromising model integrity through both untargeted performance degradation and targeted backdoor attacks. Detecting backdoors in FL is challenging due to their stealthy nature and variability in local datasets. Existing defenses struggle against adaptive adversaries and distinguishing between poisoning and genuine dataset anomalies. This paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning defense for FL. Leveraging the staple few-shot learning technique of Siamese networks, Sibai effectively detects malicious contributions in various scenarios, including settings with strong variations between clients' datasets and encounters with adaptive adversaries. Sibai achieves high detection rates, prevents backdoors, minimizes performance impact, and outperforms eight recent defenses regarding F1 score, poisoning prevention, and consistency across complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.66, + 0.231, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.684, + 0.486, + 0.891 + ], + "angle": 0, + "content": "Federated Learning (FL) [59] enables collaborative deep neural network (DNN) training across decentralized clients by sharing model updates instead of raw data, preserving privacy. While FL offers significant benefits, it is susceptible to poisoning attacks by untrustworthy or compromised clients who manipulate training data, labels, hyperparameters, or model parameters. These attacks can be untargeted [15, 31, 55, 57], degrading global model performance, or targeted [2-4, 11, 18, 19, 33, 40, 42, 45, 52, 54, 56], introducing backdoors that misclassify inputs with specific triggers while maintaining normal performance otherwise. Backdoors are particularly dangerous due to their stealth and severe risks in security-critical applications like autonomous vehicles [32, 41, 62]. FL poisoning defenses include in" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.311, + 0.896, + 0.414 + ], + "angle": 0, + "content": "fluence reduction (IR) [1, 3, 39, 50], robust aggregation (RA) [36, 61], and the more common detection and filtering (DF) [5, 8, 17, 27, 43, 47, 48, 63], which excludes poisoned models before aggregation. Although IR and RA can leverage information from all local models, their methods, such as noising and clipping, often degrade benign performance when applied broadly." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.415, + 0.897, + 0.575 + ], + "angle": 0, + "content": "Challenges. Most DF-based defenses rely on predefined metrics to detect poisoning, but this approach faces two key challenges. First, adaptive adversaries can tailor poisoned models to evade detection by appearing normal within these metrics [27, 29]. Expanding the metric sets [27] partially addresses this, but still relies on rigid criteria which are vulnerable to adaptation. Second, the natural diversity of non-identically and independently distributed (non-i.i.d.) datasets in real-world scenarios makes distinguishing poisoned models from those reflecting unusual data distributions difficult." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.577, + 0.897, + 0.723 + ], + "angle": 0, + "content": "Meta-Classifiers. This paper introduces a dynamic meta-classifier approach to FL poisoning defense, which avoids rigid metrics and instead learns to differentiate malicious directly from observing local models. Recent methods [9, 21, 22, 25, 58] show that meta-classifiers can effectively detect poisoned models in an offline scenario. However, applying this to FL poses challenges: training requires a diverse set of so-called shadow models, including both benign and poisoned models, and existing methods depend on time-consuming preprocessing." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.725, + 0.896, + 0.827 + ], + "angle": 0, + "content": "Contributions. To tackle the aforementioned challenges, this paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning detection method specifically tailored to function as a DF-based defense for FL. Sibai is based on the staple few-shot learning technique of Siamese networks. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.831, + 0.897, + 0.89 + ], + "angle": 0, + "content": "- Sibai is the first application of a meta-classifier in DF-based poisoning defenses for FL. Unlike conventional approaches that are reliant on predefined fixed metrics for detection, Sibai dynamically learns the criteria to distin" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.103, + 0.483, + 0.176 + ], + "angle": 0, + "content": "guish between benign and malicious local models. Using a dynamic approach based on ML, Sibai poses a challenge for sophisticated attackers that try to adapt to predefined metrics, as it utilizes a diverse set of learned, complex model characteristics for its decision-making." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.177, + 0.483, + 0.234 + ], + "angle": 0, + "content": "- We utilize Siamese Networks in the context of a poisoning detection meta-classifier. These networks are designed to compare and find dissimilarities between two inputs, in the case of Sibai DNN models." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.235, + 0.483, + 0.352 + ], + "angle": 0, + "content": "- Sibai operates directly on flattened DNN representations using a novel layer-wise dimensionality reduction pipeline, using Principal Component Analysis (PCA) enhanced with engineered features. This feature-rich approach improves Sibai's ability to learn complex criteria, enhancing robustness against adaptation. Additionally, it enables Sibai to efficiently inspect all local models in a federation without requiring extensive time-consuming preprocessing steps." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.353, + 0.483, + 0.543 + ], + "angle": 0, + "content": "- We extensively evaluate Sibai in various FL scenarios against five different poisoning attacks, assessing its effectiveness against stealthy and adaptive attackers and in scenarios with non-i.i.d. data. To conduct our evaluation, we created a substantial dataset of shadow models, including 5720 benign and 4145 malicious models. Sibai consistently achieves high detection rates across all tested scenarios without notable degradation of the global model's performance. Notably, Sibai successfully prevents the establishment of effective backdoors in all experiments, and outperforms the eight existing FL poisoning defenses, demonstrating higher F1 scores and consistently maintaining the lowest backdoor accuracy across all experiments." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.177, + 0.483, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.547, + 0.485, + 0.591 + ], + "angle": 0, + "content": "In summary, Sibai introduces a robust ML-based FL poisoning defense to effectively address adaptive adversaries and non-i.i.d. scenarios." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.607, + 0.227, + 0.623 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.632, + 0.453, + 0.647 + ], + "angle": 0, + "content": "This section introduces concepts relevant to this paper." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.651, + 0.483, + 0.754 + ], + "angle": 0, + "content": "Federated Learning FL [26, 36, 59] enables collaborative ML by training on local data while preserving privacy. In each round, the server sends the global model to selected clients, who train it locally and return updates. The server aggregates updates, often using FedAVG [36], a weighted average with a global learning rate \\(\\delta\\). The algorithm is explained in more detail in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.758, + 0.483, + 0.89 + ], + "angle": 0, + "content": "IID vs. non-IID Data Distributions ML models capture the characteristics of their datasets, emphasizing patterns in the training data. In an ideal identically and independently distributed (IID) scenario, all clients have similar data distributions, making outlier detection an effective strategy to detect poisoning. However, the real world rarely resembles those conditions; instead, clients usually exhibit non-i.i.d. data distributions which complicate the distinction between anomalies and poisoning. For classification tasks, '1-class" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.176 + ], + "angle": 0, + "content": "non-i.i.d.' describes datasets dominated by a single label, with the disproportionality degree regulated by a parameter \\( q \\in [0,1] \\). Inter-client non-i.i.d. [27] however, is even more complex and allows arbitrary and disjoint distributions, which resemble real-world conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.18, + 0.895, + 0.356 + ], + "angle": 0, + "content": "Poisoning Attacks Poisoning attacks [51, 54] on DNNs involve attackers altering the training process to manipulate the model. Untargeted attacks lower main task accuracy (MA) by reducing local model accuracy, while targeted (backdoor) attacks embed triggers causing specific misclassifications with minimal impact on MA [3, 18, 34]. Poisoning involves data (e.g., malicious labels or triggers [49]) or model modifications (e.g., altered gradients [28]). This paper mainly considers five specific poisoning types: pixel trigger backdoors [19], blend backdoors [10], edge-case attacks [54], semantic backdoors [3], and random label flipping [27], which are discussed in more detail in Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Adaptive Attackers Adaptive attackers can bypass FL defenses that focus on fixed metrics by manipulating their local models to evade detection. Bagdasaryan et al. [3] introduced a method that uses an additional loss component that balances backdoor accuracy (BA) and stealthiness to the metrics the defense considers, as shown in Equation (1) [3]." + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.467, + 0.895, + 0.484 + ], + "angle": 0, + "content": "\\[\nL o s s = \\alpha * L o s s ^ {M A \\& B A} + (1 - \\alpha) * L o s s ^ {A d a p t a t i o n} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.894, + 0.612 + ], + "angle": 0, + "content": "Siamese Neural Networks Siamese networks [7, 24] are few-shot classifiers that compare embedding vectors from twin networks with shared parameters. Using a distance metric like Euclidean distance [13], they assess similarity between inputs. The Triplet loss [20], shown in Equation (2), is used to minimize distances between the feature vectors of same-class pairs and maximize them for different-class pairs." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.632, + 0.894, + 0.663 + ], + "angle": 0, + "content": "\\[\nL (a, p, n) = \\max \\left(\\left\\| f (a) - f (p) \\right\\| ^ {2} - \\left\\| f (a) - f (n) \\right\\| ^ {2} + \\gamma , 0\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.674, + 0.58, + 0.689 + ], + "angle": 0, + "content": "3. Sibai" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.895, + 0.89 + ], + "angle": 0, + "content": "In this section, we present Siamiese Backdoor Inspector (Sibai), a novel defense mechanism against poisoning attacks in FL. Sibai addresses non-i.i.d. scenarios and adaptive attackers by dynamically learning discriminative features of malicious local models using a meta-classifier. This approach avoids the reliance on predefined metrics. By directly operating on flattened model representations, Sibai does not require time-consuming preprocessing and enables detailed analysis that hinders adversarial adaptation. Sibai quickly adapts to federation behavior and new attack strategies with minimal training samples. The neural network architecture adds a layer of opacity, complicating the task for attackers to even discern the specific criteria they must adapt to." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3788" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.102, + 0.102, + 0.241, + 0.117 + ], + "angle": 0, + "content": "3.1. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.125, + 0.486, + 0.3 + ], + "angle": 0, + "content": "We consider an FL scenario with a centralized server using FedAVG [36] and a fixed global learning rate. Consistent with prior defenses [1, 5, 6, 27, 35, 37] we assume adversaries control fewer than half of the clients, however, they can fully manipulate their data, training process, and model parameters. Adversaries are aware of the algorithms and defenses deployed on the server. Importantly, Sibai's design, training, and evaluation make no assumptions about local dataset distributions. We evaluate Sibai across i.i.d., 1-class non-i.i.d., and inter-client non-i.i.d. scenarios, the latter closely reflecting real-world FL conditions as described in Sec. 2." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.312, + 0.19, + 0.328 + ], + "angle": 0, + "content": "3.2. Design" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.335, + 0.485, + 0.48 + ], + "angle": 0, + "content": "Sibai operates on the central server to detect and filter suspicious local models before aggregation. It employs a Siamese network [16, 24] meta-classifier trained to identify dissimilarities between malicious and benign client updates. Sibai consists of two phases: (I) Offline training, where the Siamese network learns from shadow models, including both benign and malicious examples; (II) Online inference, where Sibai evaluates local models against the global model during federation to identify and exclude poisoned models based on learned criteria." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.486, + 0.486, + 0.59 + ], + "angle": 0, + "content": "Architecture. As shown in Fig. 1, Sibai takes the global model and a local model as input, processing them through four components: Dimensionality Reduction, Twin Networks, Distance Function, and Decision Network. The output is a binary decision indicating the likelihood of a local model being malicious. Detailed descriptions of each component follow in subsequent sections." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.611, + 0.408, + 0.854 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.872, + 0.376, + 0.886 + ], + "angle": 0, + "content": "Figure 1. Sibai Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.898, + 0.309 + ], + "angle": 0, + "content": "Dimensionality Reduction. Sibai processes models of varying sizes by condensing input data while retaining key features. Local models, often containing hundreds of thousands of parameters, are broken into slices, and Principal Component Analysis (PCA) extracts the most significant components for each slice. PCA identifies and ranks features by variance, ensuring compact yet informative representations. Additional metrics, specifically Cosine and Euclidean distances to the global model, that were shown to be effective by prior research [5, 17, 27, 43], are also incorporated into the input vector. Experiments showed that these metrics consistently improve detection accuracy by up to \\(10\\%\\). More details about the implementation of the dimensionality reduction pipeline are provided in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.895, + 0.476 + ], + "angle": 0, + "content": "Twin Networks. The Twin Networks consist of two identical 1D Convolutional Neural Networks (CNNs), serving as the core of the Siamese network. Each CNN processes one of the two pre-processed data samples, acting as a feature extractor. They map the samples to fixed-length embedding vectors using the same filters. These two vectors can then be compared to assess similarity. A high dissimilarity between embedding vectors indicates potential maliciousness, as the twins are explicitly trained to focus on the distinguishing features of malicious local models. Appendix C.4 provides further details about the architecture of the twin models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.895, + 0.569 + ], + "angle": 0, + "content": "Distance Function. The embedding vectors generated by the Twin Networks are compared using the \\(L_{2}\\)-distance, which demonstrated superior performance compared to other metrics like \\(L_{1}\\)-distance or cosine distance. This distance is used in the triplet loss function, which is utilized to train the twin networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.897, + 0.693 + ], + "angle": 0, + "content": "Decision Network. To avoid using a fixed distance threshold, a fully connected neural network, called a decision network, is used to classify the output of the distance function. The last layer features a sigmoid activation function [38] that produces a value between 0 and 1, with values closer to 1 indicating a higher probability of maliciousness. We classify local models that when compared to the global model produce an output \\(>0.5\\) as malicious." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.897, + 0.891 + ], + "angle": 0, + "content": "**Shadow Model Dataset.** Sibai's performance depends on a shadow model dataset used during offline training. This dataset should cover a range of attacks and scenarios to ensure effective generalization by the Siamese Network. To create the shadow model set we simulate FL across various federations with different non-i.i.d. degrees attacks and PDRs. Incorporating stealthy models from adaptive adversaries with low PDR allows Sibai to focus on nuanced criteria while including poisoned models with high PDR enhances generalization. It's crucial to maintain consistent local model architectures in the shadow set, and federations with varied input or output layer sizes cannot be mixed. Further details are provided in Appendix C.3." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3789" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.102, + 0.102, + 0.216, + 0.117 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.127, + 0.486, + 0.186 + ], + "angle": 0, + "content": "This section evaluates Sibai during online inference against five poisoning attacks in challenging scenarios, including complex non-i.i.d. settings, low PDRs, and adaptive adversaries." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.191, + 0.485, + 0.25 + ], + "angle": 0, + "content": "Hardware. Training and evaluation were performed on a Linux server with an Intel Xeon Gold 6342 CPU, 251GB RAM, CUDA [44], and an NVIDIA A30 GPU with 24GB memory." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.254, + 0.486, + 0.475 + ], + "angle": 0, + "content": "Reference Defenses. We compare Sibai with eight defenses: HDBSCAN [35] which represents naive clustering via cosine distance, Krum and M-Krum [5] which select the new global model based on Euclidean distance, MESAS [27] which tries to counter adaptive adversary by using six metrics, FLTrust [8] which uses trusted reference-based weighting, FoolsGold [17] which analyzes the final layer cosine distance, and the robust aggregation method Trimmed Mean/Median [60]. Hyperparameters are detailed in Appendix E. However, it is important to note that a direct comparison of methods with fundamentally different defense paradigms, such as a DF-based method like ours and IR-based defenses like FLTrust, cannot capture the whole picture since these paradigms approach poisoning defense with fundamentally different philosophies." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.48, + 0.486, + 0.554 + ], + "angle": 0, + "content": "Attacks. Sibai is tested against five attacks: Random label flipping [27], Pixel trigger backdoor [19], Blend backdoor [10], Edge case backdoor [54], and Semantic backdoor [3]. These attacks are introduced in more detail in Appendix D" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.558, + 0.485, + 0.617 + ], + "angle": 0, + "content": "Datasets and Models. We evaluate federations collaboratively training Squeezezenet [23] on CIFAR-10 [30]. Additionally, we conducted experiments with a custom DNN and MNIST, which are provided in Sec. 4.5." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.621, + 0.486, + 0.798 + ], + "angle": 0, + "content": "**Shadow Model Dataset.** Sibai's offline training uses 520 simulated federations, each with 20 clients (11 benign, 9 malicious), totaling 5720 benign and 4145 malicious models. Scenarios include i.i.d. and non-i.i.d. distributions with varying degrees (e.g., 0.3, 0.6, 0.9 1-class non-i.i.d. and inter-client non-i.i.d. [27]). Realistic and exaggerated settings enhance the distinction between benign and malicious models. Each federation experiences one of the five attacks. Federations with local backdoor accuracies below \\(45\\%\\) were excluded to ensure poisoned models exhibit detectable traits. Distribution details are in Tab. 1 and a detailed breakdown is listed in Appendix H." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.802, + 0.486, + 0.892 + ], + "angle": 0, + "content": "Default Federation Parameters. We evaluate Sibai in simulated federations collaboratively training Squeezenet [23] with SGD (learning rate: 0.01, momentum: 0.9, decay: 0.005) and a batch size of 64. Each client trains locally for 10 epochs on i.i.d. datasets with 2560 samples. The global model is initialized with pre-trained weights from" + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.1, + 0.844, + 0.216 + ], + "angle": 0, + "content": "
CategoryNumber of Federations
IID/non-i.i.d.i.i.d.225
0.3 1-class non-i.i.d.66
0.6 1-class non-i.i.d.71
0.9 1-class non-i.i.d.96
inter-client non-i.i.d. [27]62
AttacksRandom [27]90
Pixel [19]91
Blend [10]119
Edge [54]121
Semantic [3]99
Total Number of Federations520
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.226, + 0.851, + 0.24 + ], + "angle": 0, + "content": "Table 1. Composition of the shadow model dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.895, + 0.309 + ], + "angle": 0, + "content": "PyTorch [46] and undergoes 49 benign rounds to mitigate early backdoor persistence [3], followed by poisoning in the 50th round." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.897, + 0.446 + ], + "angle": 0, + "content": "Sibai Parameters. During offline training, the twin and decision networks are trained separately for 25 epochs using SGD. The twin networks use a learning rate of 0.001, while the decision network uses 0.005. Model slices contain up to 20,000 parameters, with larger layers partitioned. PCA retains 200 components per slice. Siamese network training does not use batch processing due to memory constraints, while the decision network processes batches of 16 distances derived from 8 Siamese network triples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.897, + 0.612 + ], + "angle": 0, + "content": "Metrics. We evaluate Sibai with precision, recall, and F1 score. True positives are correctly flagged malicious models, and true negatives are benign models. Precision measures the ratio of correctly identified malicious models to all flagged models, while recall assesses the proportion of flagged malicious models among all actual malicious models. The F1 score balances these metrics. Additionally, we track the global model's BA to evaluate attack mitigation and MA changes to assess benign scenario impact. The MA diff. represents the difference between the MA after Sibai application and perfect benign aggregation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.652, + 0.635 + ], + "angle": 0, + "content": "4.1. IID Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.897, + 0.868 + ], + "angle": 0, + "content": "This section evaluates Sibai's online performance after training on the shadow model set. We test Sibai on 12 unseen federation scenarios1 of increasing difficulty and adversary stealthiness. In the baseline scenario (Tab. 2), federations training Squeezenet [23] on CIFAR-10 [30] in an i.i.d. setting were attacked with the five poisoning attacks. Adversaries employed a PDR of 0.5, which is a relatively overt attack. Sibai flawlessly detected all malicious clients without false positives. Scenario (1) increased stealthiness by minimizing PDRs while maintaining local BAs above \\(45\\%\\). The blend [10] and semantic [3] backdoors used a PDR of 0.04 due to their stronger effects, while the edge case backdoor [54] required a PDR of 0.16 for consistent backdoor introduction. These adjustments mirror real-world attacker strategies. Sibai achieved perfect recall and F1 scores for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.876, + 0.787, + 0.89 + ], + "angle": 0, + "content": "Each scenario is designated by a circled number" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3790" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.1, + 0.437, + 0.219 + ], + "angle": 0, + "content": "
ScenarioRandomPixelBlendEdgeSemantic
①PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.09%0.20%18.87%0.0%
①PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall0.88881.00001.00001.00000.7777
F10.94111.00001.00001.00000.8750
MA diff.-0.88%0.0%0.0%0.0%+4.12%
BA-10.43%0.0%18.87%0.0%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.237, + 0.22, + 0.35, + 0.231 + ], + "angle": 0, + "content": "\\(\\boxed{0} =\\) Default + high PDRs" + }, + { + "type": "table_footnote", + "bbox": [ + 0.237, + 0.231, + 0.347, + 0.245 + ], + "angle": 0, + "content": "\\(\\widetilde{\\mathbf{1}} =\\) Default and low PDRs" + }, + { + "type": "list", + "bbox": [ + 0.237, + 0.22, + 0.35, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.255, + 0.46, + 0.269 + ], + "angle": 0, + "content": "Table 2. Detection performance of Sibai in i.i.d. scenarios" + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.284, + 0.451, + 0.508 + ], + "angle": 0, + "content": "
qScenarioRandomPixelBlendEdgeSemantic
0.6②PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-2.65%0.20%6.12%0.0%
③PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00000.88881.0000
F11.00001.00001.00000.94111.0000
MA diff.0.0%0.0%0.0%-6.71%0.0%
BA-2.65%0.21%8.67%0.0%
0.9④PDR0.50.50.50.50.5
Precision0.87501.00001.00000.90001.0000
Recall0.77770.88881.00001.00001.0000
F10.82350.94111.00000.94731.0000
MA diff.+4.80%+1.08%0.0%+2.44%0.0%
BA-0.11%0.0%0.0%0.0%
⑤PDR0.40.080.040.160.04
Precision0.81810.75001.00000.81811.0000
Recall1.00001.00001.00001.00000.8888
F10.90000.85711.00000.90000.9411
MA diff.+0.48%-4.26%0.0%-1.65%+0.48%
BA-0.0%0.21%0.0%0.0%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.194, + 0.509, + 0.395, + 0.52 + ], + "angle": 0, + "content": "\\((2) =\\) Default \\(+0.61\\) -class non-i.i.d. \\(^+\\) high PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.194, + 0.52, + 0.393, + 0.53 + ], + "angle": 0, + "content": "\\(\\widehat{(3)} =\\) Default \\(+0.61\\) -class non-i.i.d. \\(^+\\) low PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.194, + 0.53, + 0.394, + 0.542 + ], + "angle": 0, + "content": "\\((4) =\\) Default \\(+0.9\\) 1-class non-i.i.d. \\(^+\\) high PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.194, + 0.542, + 0.392, + 0.556 + ], + "angle": 0, + "content": "\\((5) =\\) Default + 0.91-class non-i.i.d. + low PDR" + }, + { + "type": "list", + "bbox": [ + 0.194, + 0.509, + 0.395, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.567, + 0.486, + 0.595 + ], + "angle": 0, + "content": "Table 3. Detection performance of Sibai in 1-class non-i.i.d. scenarios, with different level of disproportionality \\( q \\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.621, + 0.486, + 0.812 + ], + "angle": 0, + "content": "pixel trigger and edge case backdoors, identifying all malicious clients. In the random label flip and semantic backdoor attacks, one and two malicious clients, respectively, evaded detection. However, Sibai identified over \\(77\\%\\) of malicious clients in all cases, preventing effective poisoning. Across all i.i.d. scenarios, Sibai maintained BA below \\(19\\%\\), ensuring attacks were unsuccessful. Interestingly, in the semantic backdoor scenario, Sibai mistakenly aggregated two malicious models, resulting in a slight MA improvement over perfect aggregation. This occurred because the aggregated models provided additional knowledge, but the low PDR and limited bypassed defenses still prevented backdoor establishment." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.823, + 0.279, + 0.838 + ], + "angle": 0, + "content": "4.2. Non-IID Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.845, + 0.486, + 0.891 + ], + "angle": 0, + "content": "In this section, we evaluate Sibai in challenging non-i.i.d. scenarios, including varying degrees of 1-class non-i.i.d. and inter-client non-i.i.d. setups as described in Sect. 2 and [27]." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.1, + 0.849, + 0.219 + ], + "angle": 0, + "content": "
ScenarioRandomPixelBlendEdgeSemantic
6PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00000.9000
Recall0.77770.88881.00001.00001.0000
F10.87500.94111.00001.00000.9473
MA diff.+2.61%+0.86%0.0%0.0%0.0%
BA-4.03%0.0%0.0%0.0%
7PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.66661.00000.8888
F11.00000.94110.80001.00000.9411
MA diff.0.0%+1.38%-1.43%0.0%-1.48%
BA-13.57%0.0%0.0%0.0%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.604, + 0.22, + 0.806, + 0.231 + ], + "angle": 0, + "content": "\\(\\widehat{(6)} =\\) Default \\(^+\\) inter-client non-i.i.d. \\(^+\\) high PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.605, + 0.231, + 0.804, + 0.244 + ], + "angle": 0, + "content": "\\(\\widetilde{7} =\\) Default + inter-client non-i.i.d. + low PDR" + }, + { + "type": "list", + "bbox": [ + 0.604, + 0.22, + 0.806, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.255, + 0.896, + 0.282 + ], + "angle": 0, + "content": "Table 4. Detection performance of Sibai in inter-client non-i.i.d. scenarios [27]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.895, + 0.415 + ], + "angle": 0, + "content": "0.6 1-class non-i.i.d. For the scenarios \\(\\boxed{2}\\) and \\(\\boxed{3} q = 0.6\\) was used. Sibai achieved perfect detection in the 0.5 PDR case \\(\\boxed{2}\\). For the low PDRs scenario \\(\\boxed{3}\\), F1 scores were 1.0 for all attacks except the edge case backdoor, where one malicious client evaded detection. Despite this, the MA dropped by less than \\(7\\%\\), and BA stayed below \\(9\\%\\), ensuring no effective backdoor was introduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.897, + 0.525 + ], + "angle": 0, + "content": "0.9 1-class non-i.i.d. In the more extreme 0.9 1-class non-i.i.d. scenarios (4) and (5) in Tab. 10, Sibai occasionally flagged benign models, with three false positives in the worst-case pixel trigger backdoor scenario (5). However, these errors had minimal impact on MA. Sibai consistently eliminated enough malicious models to prevent effective backdoors, as no attack achieved a BA higher than \\(1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.897, + 0.768 + ], + "angle": 0, + "content": "Inter-client non-i.i.d. Inter-client non-i.i.d. features erratic and highly unbalanced data distributions, reflecting real-world scenarios. In the 0.5 PDR scenario (6), Sibai achieved perfect detection for blend and edge case backdoors, perfect precision for untargeted and pixel attacks, and perfect recall for the semantic attack. Only one false positive occurred during the semantic attack. The lowest F1 score of 0.8750 occurred during the untargeted attack) In the attacker-adapted low PDR scenario (7), Sibai maintained perfect precision for all attacks despite high non-i.i.d.-ness. Recall was lower, yielding F1 scores of 0.9411 for pixel and semantic backdoors and 0.8000 for the blend attack. Nonetheless, MA was minimally affected, and BA remained low across all attacks. Overall, Sibai demonstrated robust performance in 1-class and inter-client non-i.i.d. scenarios, with minimal false positives and no effective backdoors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.78, + 0.71, + 0.795 + ], + "angle": 0, + "content": "4.3. Adaptive Adversaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.896, + 0.891 + ], + "angle": 0, + "content": "Sibai incorporates slice-wise Euclidean and cosine distances to the global model into its input vectors to enhance detection. While effective, these metrics may be vulnerable to adaptive attackers. To test Sibai's robustness, we evaluated four scenarios involving attackers using Bagdasaryan et al.'s [3] adaptation method with an \\(\\alpha\\) parameter balancing adaptation" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.513, + 0.939 + ], + "angle": 0, + "content": "3791" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.1, + 0.483, + 0.324 + ], + "angle": 0, + "content": "
Form of AdaptationScenarioRandomPixelBlendEdgeSemantic
Cosine dist.⑧PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18.87%0.0%
⑨PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001,00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.14%18.87%0.0%
Euclidian dist.⑩PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18. 87%0.0%
⑪PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.88881.00001.0000
F11.00000.94110.94111.00001.0000
MA diff.0.0%+3.1%-6.3%0.0%0.0%
BA-5.98%0.0%18.87%0.0%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.123, + 0.325, + 0.446, + 0.336 + ], + "angle": 0, + "content": "\\(\\boxed{8} =\\) Default \\(^+\\) cosine distance adaption with \\(0.3\\alpha\\) by mal. clients \\(^+\\) high PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.123, + 0.337, + 0.443, + 0.348 + ], + "angle": 0, + "content": "9 = Default + cosine distance adaption with 0.3 α by mal. clients + low PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.124, + 0.349, + 0.464, + 0.362 + ], + "angle": 0, + "content": "\\(\\boxed{10} =\\) Default \\(^+\\) Euclidean distance adaption with \\(0.3\\alpha\\) by mal. clients \\(^+\\) high PDR" + }, + { + "type": "table_footnote", + "bbox": [ + 0.124, + 0.363, + 0.461, + 0.377 + ], + "angle": 0, + "content": "\\(\\widehat{(11)} =\\) Default \\(^+\\) Euclidean distance adaption with \\(0.3\\alpha\\) by mal. clients \\(^+\\) low PDR" + }, + { + "type": "list", + "bbox": [ + 0.123, + 0.325, + 0.464, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Table 5. Detection performance of Sibai against adaptive attackers in i.i.d. scenarios" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.444, + 0.485, + 0.641 + ], + "angle": 0, + "content": "and task losses. Results for \\(\\alpha = 0.3\\), shown in Tab. 5, highlight Sibai's resilience. In cosine-adapted scenarios, Sibai achieved perfect F1 scores (1.0), showing immunity to such attacks. Similarly, Sibai maintained perfect performance in Euclidean-adapted Scenario 12. In low-PDR Euclidean-adapted scenarios, Sibai showed flawless precision, with one false negative for pixel and blend backdoors. However, no attack established an effective backdoor, with BA remaining below \\(19\\%\\). Overall, Sibai demonstrates strong resilience against adaptive attacks, proving it does not overly rely on engineered features. Its ability to internalize specific patterns makes it impractical for attackers to adapt using Bagdasaryan et al.'s method." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.653, + 0.384, + 0.669 + ], + "angle": 0, + "content": "4.4. Comparison to Existing Defenses" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.676, + 0.486, + 0.891 + ], + "angle": 0, + "content": "To assess the effectiveness of Sibai compared to metric-based poisoning detection methods in FL, we also evaluate existing defenses in some of our previous scenarios. Due to space limitations, we focus on a selection of the most representative scenarios, including scenario 1 as the i.i.d. baseline, scenario 4 as the highest degree of 1-class non-i.i.d., and scenario 6 to showcase inter-client non-i.i.d. We also include scenario 0 as the overall simplest scenario and scenario 11 to cover adaptive adversaries. Since some reference defenses rely on modified aggregation methods instead of strictly flagging and filtering out malicious models, calculating F1 scores for all reference defenses faced challenges. Further details are provided in Appendix F." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.1, + 0.84, + 0.422 + ], + "angle": 0, + "content": "
ScenarioDefenseRandomPixelBlendEdgeSemantic
0PDR0.50.50.50.50.5
MESAS0.90000.90000.90001.00000.9473
Krum0.57140.57140.57890.57140.5714
M-Krum0.26080.26080.64280.26080.2608
HDBSCAN1.00000.00000.11111.00001.0000
FLTrust (0.09)0.00000.26080.33330.00000.2608
Sibai (ours)1.00001.00001.00001.00001.0000
1PDR0.40.080.040.160.04
MESAS0.94730.00000.18180.16660.3076
Krum0.57140.64280.64280.57140.5714
M-Krum0.26080.60860.60860.26080.4347
HDBSCAN0.00000.55550.44440.00000.1176
FLTrust (0.09)0.09520.66660.57140.51850.5714
Sibai (ours)0.94111.00001.00001.00000.8750
4PDR0.50.50.50.50.5
MESAS0.00000.16660.80000.94730.9000
Krum0.57140.57140.57140.64280.6428
M-Krum0.26080.69560.69560.78260.7826
HDBSCAN0.00000.40000.77770.82350.8235
FLTrust (0.09)0.57140.88880.88880.88880.8571
Sibai (ours)0.82350.94111.00000.94731.0000
6PDR0.50.50.50.50.5
MESAS0.64280.47610.21050.66660.3157
Krum0.57140.57140.57140.57140.5714
M-Krum0.26080.26080.26080.52170.3478
HDBSCAN0.00000.00000.14280.30760.1428
FLTrust (0.09)0.00000.57140.57140.57140.6206
Sibai (ours)0.87500.94111.00001.00000.9473
11PDR0.40.080.040.160.04
MESAS1.00000.66660.90000.61530.9473
Krum0.57140.57140.64280.64280.6428
M-Krum0.26080.26080.78260.60860.6956
HDBSCAN0.00001.00001.00001.00000.2000
FLTrust (0.09)0.57140.62060.66660.62060.6206
Sibai (ours)1.00000.94110.94111.00001.0000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.432, + 0.896, + 0.46 + ], + "angle": 0, + "content": "Table 6. F1 score comparison between Sibai and current FL poisoning defenses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.896, + 0.739 + ], + "angle": 0, + "content": "As shown in Tab. 6, Sibai achieved superior F1 scores across most scenarios. In high PDR i.i.d. scenario \\(\\boxed{0}\\), it matched or outperformed all defenses. In low PDR i.i.d. scenario \\(\\boxed{1}\\), Sibai achieved perfect F1 scores for pixel, blend, and edge case backdoors, outperforming MESAS for these attacks. Even when MESAS maintained low BA, it highlighted the fragility of backdoor attacks in low PDR settings. For 1-class non-i.i.d. scenario \\(\\boxed{4}\\), Sibai significantly outperformed other methods, and in inter-client non-i.i.d. scenario \\(\\boxed{7}\\), its worst-case F1 score of 0.8750 surpassed the best scores of other defenses. In adaptive scenario \\(\\boxed{11}\\), Sibai slightly trailed HDBSCAN for pixel and blend backdoors however, it excelled in the detection of the semantic and untargeted attacks, achieving perfect F1 scores. Overall, Sibai consistently outperformed existing methods in identifying malicious models across diverse scenarios and attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.897, + 0.891 + ], + "angle": 0, + "content": "Table 7, also considered methods for which the F1 was unavailable, specifically FoolsGold [17] and trimmed mean/median [60]. Sibai showed performance close to perfect benign aggregation, effectively preventing poisoning without significant MA reduction. Notably, Sibai and MESAS achieved low BA in challenging scenarios (4) and (6), where other defenses failed to stop backdoors. Although FLTrust [8], produced higher MA in non-i.i.d. scenarios by using a weighted aggregation that particularly emphasizes the old global model, it also often allowed higher BA and" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3792" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.249, + 0.1, + 0.75, + 0.412 + ], + "angle": 0, + "content": "
ScenarioDefenseRandom MA diff.Pixel MA diff.BABlend MA diff.BAEdge MA diff.BASemantic MA diff.BA
①PDR0.40.080.040.160.04
benign agg.0.00.010.40.00.00.018.80.00.0
MESAS [27]-0.1+5.19.50.020.4+0.27.10.00.0
Krum [5]-35.0-43.20.0-14.10.0-35.50.0-21.1100.0
M-Krum [5]-24.5-8.512.5-9.52.9-16.30.0-11.40.0
FoolsGold [17]-6.3+3.011.0-1.718.8-8.36.1+0.10.0
HDBSCAN [35]-16.0-1.910.9-6.511.4-13.80.0-5.80.0
Trimmed mean [60]-8.3+3.510.2-2.512.8-9.06.6-0.60.0
Trimmed median [60]-22.3-2.65.2-11.60.1-14.72.5-6.40.0
FLTrust [8]-22.3-0.117.2-5.213.0+1.516.8-4.80.0
Sibai (ours)-0.80.010.40.00.00.018.80.00.0
④PDR0.40.50.50.50.5
benign agg.0.00.00.00.00.00.00.00.00.0
MESAS [27]-5.7+0.334.3+5.00.6+2.40.0+10.60.0
Krum [5]-7.7-6.1100.0-0.2100.0-4.80.0-5.30.0
M-Krum [5]-2.1+5.51.7+1.0100.0-5.10.0+4.00.0
FoolsGold [17]-2.0-5.030.2-0.2100.0-10.80.0-7.30.0
HDBSCAN [35]-7.5+1.013.2+1.40.0-8.70.0-0.920.0
Trimmed mean [60]-5.3-0.428.9-0.2100.0-11.10.0-7.840.0
Trimmed median [60]-6.6-5.10.0-0.4100.0-10.40.0-6.50.0
FLTrust [8]+19.5+14.753.1+10.898.2+4.80.0+11.120.0
Sibai (ours)+4.8+1.00.00.00.0+2.40.00.00.0
⑥PDR0.50.50.50.50.5
benign agg.0.013.92.90.02.30.02.40.00.0
MESAS [27]+11.9-0.31.6-0.132.7+9.70.00.00.0
Krum [5]+4.8-3.9100.0-5.1100.0+16.885.7+18.780.0
M-Krum [5]-1.6-3.085.3-0.3100.0+2.90.00.00.0
FoolsGold [17]-1.6+1.57.0-5.0100.0-2.00.00.00.0
HDBSCAN [35]-5.8-0.719.0+3.899.9-0.30.0-0.40.0
Trimmed mean [60]-0.5+1.17.3-3.8.1100.0-3.00.00.00.0
Trimmed median [60]-2.5-3.50.0-4.9100.0-2.30.0-0.30.0
FLTrust [8]+14.7+14.347.4+13.099.8+14.00.0+8.320.0
Sibai (ours)+2.6+0.84.00.00.00.00.00.00.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.423, + 0.897, + 0.451 + ], + "angle": 0, + "content": "Table 7. Comparison of MA diff. and BA (both in percentage) between Sibai and existing FL poisoning defenses. The best value for each attack in each scenario is highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.536 + ], + "angle": 0, + "content": "therefore failed to prevent effective backdoors. In contrast, Sibai excelled at malicious model detection and consistently maintained low BA while achieving MAs aligned with benign aggregation, making Sibai an overall reliable defense." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.547, + 0.456, + 0.563 + ], + "angle": 0, + "content": "4.5. Alternative Dataset and DNN Architecture" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.57, + 0.486, + 0.746 + ], + "angle": 0, + "content": "To demonstrate that Sibai is not limited to SqueezeNET or the CIFAR-10 dataset, we conducted additional experiments using a custom CNN and the MNIST dataset [12]. The CNN model consists of 2 convolutional layers, 2 max-pooling layers, and 3 linear layers. We excluded the dataset-specific semantic and edge case backdoors and instead included a label swap attack [27], which targets samples of one label and swaps them to another to reduce the global model's accuracy. The shadow-model dataset for these experiments comprised 1401 federations. The F1 scores in Tab. 8 indicate that Sibai effectively detected malicious models, and therefore that the approach is not specific to SqueezeNET or CIFAR-10." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.762, + 0.399, + 0.779 + ], + "angle": 0, + "content": "5. Multiple Poisonings in one Model" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.787, + 0.486, + 0.891 + ], + "angle": 0, + "content": "To evaluate Sibai's effectiveness in detecting models poisoned with multiple backdoors simultaneously, we conducted experiments using federations where each malicious local model contained both a blend backdoor [10] and a pixel trigger backdoor [19] simultaneously. For those local models, the PDR quantifies the ratio of the total poisoned samples, i.e. the sum used for both backdoors, to the com" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.473, + 0.887, + 0.576 + ], + "angle": 0, + "content": "
PDRDistributionRandom [27]Pixel [19]Blend [10]Label Swap [27]
0.4i.i.d.1.00001.00001.00001.0000
0.40.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.40.6 l-class non-i.i.d.0.69230.72000.72000.7200
0.40.9 l-class non-i.i.d.0.87500.94111.00001.0000
0.4inter-client non-i.i.d.0.90000.94110.94730.9000
0.5i.i.d.1.00001.00001.00001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.50.6 l-class non-i.i.d.0.72000.66660.72000.7200
0.50.9 l-class non-i.i.d.1.00000.94111.00001.0000
0.5inter-client non-i.i.d.0.94730.94110.94730.9000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.586, + 0.895, + 0.613 + ], + "angle": 0, + "content": "Table 8. F1 score of Sibai in federations training a small CNN with the MNIST dataset" + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.632, + 0.832, + 0.735 + ], + "angle": 0, + "content": "
PDRDistributionPrecisionRecallF1
0.2i.i.d.1.00001.0001.0000
0.20.3 l-class non-i.i.d.1.00001.00001.0000
0.20.6 l-class non-i.i.d.1.00001.00001.0000
0.20.9 l-class non-i.i.d.1.00000.88880.9411
0.2inter-client non-i.i.d.1.00000.88880.9411
0.5i.i.d.1.00001.0001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.0000
0.50.6 l-class non-i.i.d.1.00001.00001.0000
0.50.9 l-class non-i.i.d.1.00001.00001.0000
0.5inter-client non-i.i.d.1.00001.00001.0000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.745, + 0.896, + 0.786 + ], + "angle": 0, + "content": "Table 9. Detection performance of Sibai against local models poisoned with two backdoors (pixel trigger backdoor [19] and blend backdoor [10]) simultaneously in different scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.896, + 0.89 + ], + "angle": 0, + "content": "plete local model dataset. In our implementation of this multi-backdoor half of the poisoned samples were used to introduce the pixel trigger backdoor, while the other half was used to apply the blend backdoor. Importantly, our shadow model dataset, used to train Sibai, did not include" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3793" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.148 + ], + "angle": 0, + "content": "any models with such multi-backdoors. As illustrated in Tab. 9, Sibai consistently identified local models containing multiple backdoors across all tested scenarios." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.16, + 0.242, + 0.175 + ], + "angle": 0, + "content": "6. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.184, + 0.486, + 0.245 + ], + "angle": 0, + "content": "This section provides a brief overview of relevant literature. First, Sect. 6.1 outlines poisoning defenses tailored for FL. Then, Sect. 6.2 discusses machine learning methods for general offline poisoning analysis." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.252, + 0.485, + 0.283 + ], + "angle": 0, + "content": "6.1. Defenses against Poisoning Attacks in Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.288, + 0.486, + 0.892 + ], + "angle": 0, + "content": "There exists a diverse range of techniques that employ fixed metrics to detect poisoning attacks in FL. FoolsGold [17] uses a client-specific update history which is compared to the update histories of other clients using the cosine similarity. Krum [5] selects a new global model based on the Euclidean distance to other local models. M-Krum [5] extends this method to select multiple models simultaneously. Trimmed-Mean/Median [60] are robust aggregation methods that replace FedAVG [36] and calculate the updated global model by using coordinate-wise means or medians from local models. FLAME [43] clusters models using HDBSCAN [35] and filters out smaller clusters. It also applies differential privacy techniques, weight clipping, and noisng. DeepSight [47] combines filtering with differential privacy. It relies on two metrics: The cosine distance between models and values extracted from the output layer of the model. Auror [48] uses K-Means clustering on model parameters, however, unfortunately, it exhibited an extreme performance overhead in our experiments. FLTrust [8] uses a weighted aggregation system for which it assigns weights to local models based on their cosine and Euclidean distances to a benign reference model, which is trained server-side on a small trusted dataset. FLDetector [63] is a historical update-based defense that mathematically predicts client updates and compares the predictions to the actual updates. MESAS [27] was specifically designed to operate in non-i.i.d. scenario and counter adaptive adversaries and employs six metrics. These approaches differ from Sibai as they rely on fixed metrics or criteria, making them susceptible to adaptive attacks [27]. Defenses focusing on specific layers, like FoolsGold [17], are prone to adversaries exploiting advanced model poisoning techniques specifically targeting those layers. Similarly, methods that heavily focus on cosine or Euclidean distance are often vulnerable to adaptation methods using additional losses [3, 27], as detailed in Sect. 2. These rigid metrics also face challenges in non-i.i.d. scenarios, especially in inter-client non-i.i.d. settings, where it has been demonstrated that most defenses, except MESAS [27] which was specifically designed to handle such scenarios, face difficulties. In contrast, Sibai dynamically learns criteria based on its shadow model dataset, making it challenging for attackers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.896, + 0.149 + ], + "angle": 0, + "content": "to adapt without access to the Siamese Network. Sibai can also differentiate between variations in local models due to data distribution and those caused by poisoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.878, + 0.178 + ], + "angle": 0, + "content": "6.2. Non-FL Specific Offline Poisoning Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.899, + 0.641 + ], + "angle": 0, + "content": "This section discusses several machine-learning-based detection methods against poisoning attacks, not specifically designed for FL. Xu et al. [58] propose Meta Neural Trojan Detection (MNTD), a black-box backdoor detection approach utilizing a meta-classifier that assesses if a model's output behavior on a given input aligns more closely with a set of known benign models or a set of known malicious models. Huang et al. [21] introduce the one-pixel signature concept for DNNs, consisting of a set of single-pixel adversarial perturbations that have the highest probability of triggering misclassifications on a specific benign image dataset. Subsequently, a meta-classifier was trained on these signatures to discern between poisoned and benign models. Kolouri et al. [25] use universal litmus patterns and a meta-classifier to detect backdoors across different network architectures. Chen et al. [9] present DeepInspect, a black-box method utilizing model inversion and anomaly detection for backdoor detection. Huang et al. [22] also introduced Neuroninspect, employing explanation heatmaps and outlier detection for identifying abnormal neurons. Unlike these methods, Sibai is tailored to FL, avoiding intensive preprocessing steps. DeepInspect [9] involves extensive reverse engineering of potential triggers for every label of each local model, and Huang et al. [21] requires testing every local model with a large set of single-pixel adversarial perturbations. Conducting such intricate analyses for every model in a federation during each round is impractical, as these methods were originally designed for detailed offline analysis of individual models. Sibai instead only uses PCA as the sole dimensionality reduction mechanism, facilitating straightforward inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.63, + 0.673 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.899, + 0.89 + ], + "angle": 0, + "content": "To counter non-i.i.d. data distribution and adaptive adversaries in FL, we introduce Siamese Backdoor Inspector (Sibai), a novel FL poisoning defense mechanism utilizing a Siamese network meta-classifier. Inspired by successful offline backdoor detection [21, 25, 58], Sibai dynamically learns discriminative features directly from PCA decompositions of local models. In diverse evaluations, Sibai consistently achieves high detection rates without significantly impacting global model performance. It effectively prevents backdoor establishment and outperforms current defenses in terms of f1 score while maintaining low false positive rates. Future work may explore selectively applying influence reduction to models identified by Sibai for targeted defense." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.94 + ], + "angle": 0, + "content": "3794" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.102, + 0.102, + 0.27, + 0.119 + ], + "angle": 0, + "content": "8. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.127, + 0.486, + 0.202 + ], + "angle": 0, + "content": "This research has been funded by the Federal Ministry of Education and Research of Germany (BMBF) within the program Digital. Sicher. Souverän. \"in the project\"Erkennung von Angriffen gegen IoT-Netzwerke in Smart Homes - IoTGuard\" (project number 16KIS1919)." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.219, + 0.197, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.244, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[1] Sebastien Andreina, Giorgia Azzurra Marson, Helen Möllering, and Ghassan Karame. Baffle: Backdoor detection via feedback-based federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), pages 852-863. IEEE, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.314, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. In 30th USENIX Security Symposium (USENIX Security 21), pages 1505-1521, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.357, + 0.484, + 0.423 + ], + "angle": 0, + "content": "[3] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 4, 5, 8, 13, 14, 15, 16, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.426, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[4] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International Conference on Machine Learning, pages 634-643. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.483, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[5] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 3, 4, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.54, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[6] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.597, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[7] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in neural information processing systems, 6, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.653, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. Network and Distributed Systems Security (NDSS) Symposium 2021, 2021. 1, 4, 6, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.113, + 0.709, + 0.484, + 0.762 + ], + "angle": 0, + "content": "[9] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In *IJCAI*, page 8, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.766, + 0.484, + 0.82 + ], + "angle": 0, + "content": "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 4, 7, 13, 14, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.822, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual computer security applications conference, pages 554-569, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.244, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.104, + 0.896, + 0.143 + ], + "angle": 0, + "content": "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.146, + 0.895, + 0.172 + ], + "angle": 0, + "content": "[13] Elena Deza, Michel Marie Deza, Michel Marie Deza, and Elena Deza. Encyclopedia of distances. Springer, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.175, + 0.895, + 0.214 + ], + "angle": 0, + "content": "[14] Cynthia Dwork et al. The mathematics of information coding, extraction, and distribution. The IMA Volumes in Mathematics and its applications, 107:31-47, 1999. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.217, + 0.896, + 0.269 + ], + "angle": 0, + "content": "[15] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.272, + 0.896, + 0.338 + ], + "angle": 0, + "content": "[16] Hossein Fereidooni, Jan König, Phillip Rieger, Marco Chilese, Bora Gökbakan, Moritz Finke, Alexandra Dmitrienko, and Ahmad-Reza Sadeghi. Authentisense: A scalable behavioral biometrics authentication scheme using few-shot learning for mobile platforms. arXiv preprint arXiv:2302.02740, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.896, + 0.406 + ], + "angle": 0, + "content": "[17] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2020), pages 301-316, 2020. 1, 3, 4, 6, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.409, + 0.896, + 0.475 + ], + "angle": 0, + "content": "[18] Yansong Gao, Bao Gia Doan, Zhi Zhang, Siqi Ma, Jiliang Zhang, Anmin Fu, Surya Nepal, and Hyoungshick Kim. Backdoor attacks and countermeasures on deep learning: A comprehensive review. arXiv preprint arXiv:2007.10760, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.896, + 0.531 + ], + "angle": 0, + "content": "[19] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 2, 4, 7, 13, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.896, + 0.598 + ], + "angle": 0, + "content": "[20] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-Based Pattern Recognition: Third International Workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.896, + 0.68 + ], + "angle": 0, + "content": "[21] Shanjiaoyang Huang, Weiqi Peng, Zhiwei Jia, and Zhuowen Tu. One-pixel signature: Characterizing cnn models for backdoor detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 326-341. Springer, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.896, + 0.723 + ], + "angle": 0, + "content": "[22] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.726, + 0.896, + 0.779 + ], + "angle": 0, + "content": "[23] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and i 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.781, + 0.896, + 0.821 + ], + "angle": 0, + "content": "[24] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop. Lille, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.822, + 0.896, + 0.889 + ], + "angle": 0, + "content": "[25] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.104, + 0.896, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.515, + 0.939 + ], + "angle": 0, + "content": "3795" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.486, + 0.157 + ], + "angle": 0, + "content": "[26] Jakub Konečný, H Brendan McMahan, Daniel Ramage, and Peter Richtárik. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint arXiv:1610.02527, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.159, + 0.486, + 0.251 + ], + "angle": 0, + "content": "[27] Torsten Krauß and Alexandra Dmitrienko. Mesas: Poisoning defense for federated learning resilient against adaptive attackers. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, page 1526-1540, New York, NY, USA, 2023. Association for Computing Machinery. 1, 2, 3, 4, 5, 7, 8, 13, 15, 16, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.253, + 0.485, + 0.307 + ], + "angle": 0, + "content": "[28] Torsten Krauß, Raphael Götz, and Alexandra Dmitrienko. Security of nvme offloaded data in large-scale machine learning. In Computer Security - ESORICS 2023, pages 143-163, Cham, 2024. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.308, + 0.485, + 0.374 + ], + "angle": 0, + "content": "[29] Torsten Krauß, Jan König, Alexandra Dmitrienko, and Christian Kanzow. Automatic Adversarial Adaption for Stealthy Poisoning Attacks in Federated Learning. Network and Distributed Systems Security (NDSS) Symposium 2024, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.375, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 13, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.403, + 0.485, + 0.469 + ], + "angle": 0, + "content": "[31] Liping Li, Wei Xu, Tianyi Chen, Georgios B Giannakis, and Qing Ling. Rsa: Byzantine-robust stochastic aggregation methods for distributed learning from heterogeneous datasets. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1544–1551, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.47, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[32] Yijing Li, Xiaofeng Tao, Xuefei Zhang, Junjie Liu, and Jin Xu. Privacy-preserved federated learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(7):8423-8434, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.525, + 0.485, + 0.564 + ], + "angle": 0, + "content": "[33] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.565, + 0.485, + 0.631 + ], + "angle": 0, + "content": "[34] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.633, + 0.485, + 0.672 + ], + "angle": 0, + "content": "[35] Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017. 3, 4, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.673, + 0.485, + 0.739 + ], + "angle": 0, + "content": "[36] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.741, + 0.485, + 0.794 + ], + "angle": 0, + "content": "[37] Luis Muñoz-González, Kenneth T Co, and Emil C Lupu. Byzantine-robust federated machine learning through adaptive model averaging. arXiv preprint arXiv:1909.05125, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.796, + 0.485, + 0.835 + ], + "angle": 0, + "content": "[38] Sridhar Narayan. The generalized sigmoid activation function: Competitive supervised learning. Information sciences, 99 (1-2):69-82, 1997. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.836, + 0.485, + 0.889 + ], + "angle": 0, + "content": "[39] Mohammad Naseri, Jamie Hayes, and Emiliano De Cristofaro. Local and central differential privacy for robustness and privacy in federated learning. arXiv preprint arXiv:2009.03561, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.103, + 0.104, + 0.486, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.158 + ], + "angle": 0, + "content": "[40] Blaine Nelson, Marco Barreno, Fuching Jack Chi, Anthony D Joseph, Benjamin IP Rubinstein, Udam Saini, Charles Sutton, J Doug Tygar, and Kai Xia. Exploiting machine learning to subvert your spam filter. LEET, 8(1-9):16-17, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.159, + 0.895, + 0.225 + ], + "angle": 0, + "content": "[41] Anh Nguyen, Tuong Do, Minh Tran, Binh X Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, and Quang D Tran. Deep federated learning for autonomous driving. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1824-1830. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.228, + 0.896, + 0.294 + ], + "angle": 0, + "content": "[42] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), pages 1-7, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.297, + 0.895, + 0.378 + ], + "angle": 0, + "content": "[43] Thien Duc Nguyen, Phillip Rieger, Roberta De Viti, Huili Chen, Björn B Brandenburg, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, et al. {FLAME}: Taming backdoors in federated learning. In 31st USENIX Security Symposium (USENIX Security 22), pages 1415–1432, 2022. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.38, + 0.895, + 0.405 + ], + "angle": 0, + "content": "[44] NVIDIA, Péter Vingelmann, and Frank H.P. Fitzek. Cuda, release: 10.2.89, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.408, + 0.895, + 0.474 + ], + "angle": 0, + "content": "[45] Xudong Pan, Mi Zhang, Beina Sheng, Jiaming Zhu, and Min Yang. Hidden trigger backdoor attack on {NLP} models via linguistic style manipulation. In 31st USENIX Security Symposium (USENIX Security 22), pages 3611-3628, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.477, + 0.896, + 0.597 + ], + "angle": 0, + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.6, + 0.896, + 0.666 + ], + "angle": 0, + "content": "[47] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. NNetwork and Distributed Systems Security (NDSS) Symposium 2022, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.668, + 0.896, + 0.734 + ], + "angle": 0, + "content": "[48] Shiqi Shen, Shruti Tople, and Prateek Saxena. Auror: Defending against poisoning attacks in collaborative deep learning systems. In Proceedings of the 32nd Annual Conference on Computer Security Applications, pages 508-519, 2016. 1, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.737, + 0.896, + 0.789 + ], + "angle": 0, + "content": "[49] Gan Sun, Yang Cong, Jiahua Dong, Qiang Wang, Lingjuan Lyu, and Ji Liu. Data poisoning attacks on federated machine learning. IEEE Internet of Things Journal, 9(13):11365-11375, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.792, + 0.895, + 0.833 + ], + "angle": 0, + "content": "[50] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.835, + 0.895, + 0.875 + ], + "angle": 0, + "content": "[51] Zhiyi Tian, Lei Cui, Jie Liang, and Shui Yu. A comprehensive survey on poisoning attacks and countermeasures in machine learning. ACM Computing Surveys, 55(8):1-35, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.876, + 0.895, + 0.89 + ], + "angle": 0, + "content": "[52] Alexander Turner, Dimitris Tsipras, and Aleksander" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.896, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "3796" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.104, + 0.484, + 0.13 + ], + "angle": 0, + "content": "Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.132, + 0.484, + 0.159 + ], + "angle": 0, + "content": "[53] Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual. CreateSpace, Scotts Valley, CA, 2009. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.161, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[54] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. Advances in Neural Information Processing Systems, 33:16070-16084, 2020. 1, 2, 4, 13, 14, 17, 18, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.243, + 0.484, + 0.295 + ], + "angle": 0, + "content": "[55] Zhaoxian Wu, Qing Ling, Tianyi Chen, and Georgios B Giannakis. Federated variance-reduced stochastic gradient descent with robustness to byzantine attacks. IEEE Transactions on Signal Processing, 68:4583-4596, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.298, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[56] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.353, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[57] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.408, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[58] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. In 2021 IEEE Symposium on Security and Privacy (SP), pages 103-120. IEEE, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.463, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[59] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.518, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[60] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 4, 6, 7, 8, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.573, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[61] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.628, + 0.484, + 0.681 + ], + "angle": 0, + "content": "[62] Hongyi Zhang, Jan Bosch, and Helena Holmström Olsson. End-to-end federated learning for autonomous driving vehicles. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.683, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[63] Zaixi Zhang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Fl detector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2545-2555, 2022. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.484, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.514, + 0.939 + ], + "angle": 0, + "content": "3797" + } + ] +] \ No newline at end of file diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_origin.pdf b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..771d5325f06c9711e518955089caac59dad90c69 --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/c815ddc4-d777-4075-b792-39640cd474e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83d5ace7df5a5a63747b0aa14b81e63255e5f832a3e3570ec5c0a8d60624cc7b +size 178124 diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/full.md b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dee9ec299507b392bd57c268d469c0f8d144a173 --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/full.md @@ -0,0 +1,299 @@ +# Sibai: A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning + +Melanie Gotoz, Torsten Krauß, Alexandra Dmitrienko + +University of Wurzburg + +Sanderring 2, 97070 Wurzburg, Germany + +melanie.goetz@uni-wuerzburg.de, torsten.krauss@uni-wuerzburg.de, + +alexandra.dmitrienko@uni-wuerzburg.de + +# Abstract + +Federated Learning (FL) enables collaborative machine learning across decentralized clients without sharing raw data, which offers enhanced privacy and improved performance. However, FL is vulnerable to poisoning attacks, compromising model integrity through both untargeted performance degradation and targeted backdoor attacks. Detecting backdoors in FL is challenging due to their stealthy nature and variability in local datasets. Existing defenses struggle against adaptive adversaries and distinguishing between poisoning and genuine dataset anomalies. This paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning defense for FL. Leveraging the staple few-shot learning technique of Siamese networks, Sibai effectively detects malicious contributions in various scenarios, including settings with strong variations between clients' datasets and encounters with adaptive adversaries. Sibai achieves high detection rates, prevents backdoors, minimizes performance impact, and outperforms eight recent defenses regarding F1 score, poisoning prevention, and consistency across complex scenarios. + +# 1. Introduction + +Federated Learning (FL) [59] enables collaborative deep neural network (DNN) training across decentralized clients by sharing model updates instead of raw data, preserving privacy. While FL offers significant benefits, it is susceptible to poisoning attacks by untrustworthy or compromised clients who manipulate training data, labels, hyperparameters, or model parameters. These attacks can be untargeted [15, 31, 55, 57], degrading global model performance, or targeted [2-4, 11, 18, 19, 33, 40, 42, 45, 52, 54, 56], introducing backdoors that misclassify inputs with specific triggers while maintaining normal performance otherwise. Backdoors are particularly dangerous due to their stealth and severe risks in security-critical applications like autonomous vehicles [32, 41, 62]. FL poisoning defenses include in + +fluence reduction (IR) [1, 3, 39, 50], robust aggregation (RA) [36, 61], and the more common detection and filtering (DF) [5, 8, 17, 27, 43, 47, 48, 63], which excludes poisoned models before aggregation. Although IR and RA can leverage information from all local models, their methods, such as noising and clipping, often degrade benign performance when applied broadly. + +Challenges. Most DF-based defenses rely on predefined metrics to detect poisoning, but this approach faces two key challenges. First, adaptive adversaries can tailor poisoned models to evade detection by appearing normal within these metrics [27, 29]. Expanding the metric sets [27] partially addresses this, but still relies on rigid criteria which are vulnerable to adaptation. Second, the natural diversity of non-identically and independently distributed (non-i.i.d.) datasets in real-world scenarios makes distinguishing poisoned models from those reflecting unusual data distributions difficult. + +Meta-Classifiers. This paper introduces a dynamic meta-classifier approach to FL poisoning defense, which avoids rigid metrics and instead learns to differentiate malicious directly from observing local models. Recent methods [9, 21, 22, 25, 58] show that meta-classifiers can effectively detect poisoned models in an offline scenario. However, applying this to FL poses challenges: training requires a diverse set of so-called shadow models, including both benign and poisoned models, and existing methods depend on time-consuming preprocessing. + +Contributions. To tackle the aforementioned challenges, this paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning detection method specifically tailored to function as a DF-based defense for FL. Sibai is based on the staple few-shot learning technique of Siamese networks. Our contributions can be summarized as follows: + +- Sibai is the first application of a meta-classifier in DF-based poisoning defenses for FL. Unlike conventional approaches that are reliant on predefined fixed metrics for detection, Sibai dynamically learns the criteria to distin + +guish between benign and malicious local models. Using a dynamic approach based on ML, Sibai poses a challenge for sophisticated attackers that try to adapt to predefined metrics, as it utilizes a diverse set of learned, complex model characteristics for its decision-making. + +- We utilize Siamese Networks in the context of a poisoning detection meta-classifier. These networks are designed to compare and find dissimilarities between two inputs, in the case of Sibai DNN models. +- Sibai operates directly on flattened DNN representations using a novel layer-wise dimensionality reduction pipeline, using Principal Component Analysis (PCA) enhanced with engineered features. This feature-rich approach improves Sibai's ability to learn complex criteria, enhancing robustness against adaptation. Additionally, it enables Sibai to efficiently inspect all local models in a federation without requiring extensive time-consuming preprocessing steps. +- We extensively evaluate Sibai in various FL scenarios against five different poisoning attacks, assessing its effectiveness against stealthy and adaptive attackers and in scenarios with non-i.i.d. data. To conduct our evaluation, we created a substantial dataset of shadow models, including 5720 benign and 4145 malicious models. Sibai consistently achieves high detection rates across all tested scenarios without notable degradation of the global model's performance. Notably, Sibai successfully prevents the establishment of effective backdoors in all experiments, and outperforms the eight existing FL poisoning defenses, demonstrating higher F1 scores and consistently maintaining the lowest backdoor accuracy across all experiments. + +In summary, Sibai introduces a robust ML-based FL poisoning defense to effectively address adaptive adversaries and non-i.i.d. scenarios. + +# 2. Background + +This section introduces concepts relevant to this paper. + +Federated Learning FL [26, 36, 59] enables collaborative ML by training on local data while preserving privacy. In each round, the server sends the global model to selected clients, who train it locally and return updates. The server aggregates updates, often using FedAVG [36], a weighted average with a global learning rate $\delta$ . The algorithm is explained in more detail in Appendix A. + +IID vs. non-IID Data Distributions ML models capture the characteristics of their datasets, emphasizing patterns in the training data. In an ideal identically and independently distributed (IID) scenario, all clients have similar data distributions, making outlier detection an effective strategy to detect poisoning. However, the real world rarely resembles those conditions; instead, clients usually exhibit non-i.i.d. data distributions which complicate the distinction between anomalies and poisoning. For classification tasks, '1-class + +non-i.i.d.' describes datasets dominated by a single label, with the disproportionality degree regulated by a parameter $q \in [0,1]$ . Inter-client non-i.i.d. [27] however, is even more complex and allows arbitrary and disjoint distributions, which resemble real-world conditions. + +Poisoning Attacks Poisoning attacks [51, 54] on DNNs involve attackers altering the training process to manipulate the model. Untargeted attacks lower main task accuracy (MA) by reducing local model accuracy, while targeted (backdoor) attacks embed triggers causing specific misclassifications with minimal impact on MA [3, 18, 34]. Poisoning involves data (e.g., malicious labels or triggers [49]) or model modifications (e.g., altered gradients [28]). This paper mainly considers five specific poisoning types: pixel trigger backdoors [19], blend backdoors [10], edge-case attacks [54], semantic backdoors [3], and random label flipping [27], which are discussed in more detail in Appendix D + +Adaptive Attackers Adaptive attackers can bypass FL defenses that focus on fixed metrics by manipulating their local models to evade detection. Bagdasaryan et al. [3] introduced a method that uses an additional loss component that balances backdoor accuracy (BA) and stealthiness to the metrics the defense considers, as shown in Equation (1) [3]. + +$$ +L o s s = \alpha * L o s s ^ {M A \& B A} + (1 - \alpha) * L o s s ^ {A d a p t a t i o n} \tag {1} +$$ + +Siamese Neural Networks Siamese networks [7, 24] are few-shot classifiers that compare embedding vectors from twin networks with shared parameters. Using a distance metric like Euclidean distance [13], they assess similarity between inputs. The Triplet loss [20], shown in Equation (2), is used to minimize distances between the feature vectors of same-class pairs and maximize them for different-class pairs. + +$$ +L (a, p, n) = \max \left(\left\| f (a) - f (p) \right\| ^ {2} - \left\| f (a) - f (n) \right\| ^ {2} + \gamma , 0\right) \tag {2} +$$ + +# 3. Sibai + +In this section, we present Siamiese Backdoor Inspector (Sibai), a novel defense mechanism against poisoning attacks in FL. Sibai addresses non-i.i.d. scenarios and adaptive attackers by dynamically learning discriminative features of malicious local models using a meta-classifier. This approach avoids the reliance on predefined metrics. By directly operating on flattened model representations, Sibai does not require time-consuming preprocessing and enables detailed analysis that hinders adversarial adaptation. Sibai quickly adapts to federation behavior and new attack strategies with minimal training samples. The neural network architecture adds a layer of opacity, complicating the task for attackers to even discern the specific criteria they must adapt to. + +# 3.1. Threat Model + +We consider an FL scenario with a centralized server using FedAVG [36] and a fixed global learning rate. Consistent with prior defenses [1, 5, 6, 27, 35, 37] we assume adversaries control fewer than half of the clients, however, they can fully manipulate their data, training process, and model parameters. Adversaries are aware of the algorithms and defenses deployed on the server. Importantly, Sibai's design, training, and evaluation make no assumptions about local dataset distributions. We evaluate Sibai across i.i.d., 1-class non-i.i.d., and inter-client non-i.i.d. scenarios, the latter closely reflecting real-world FL conditions as described in Sec. 2. + +# 3.2. Design + +Sibai operates on the central server to detect and filter suspicious local models before aggregation. It employs a Siamese network [16, 24] meta-classifier trained to identify dissimilarities between malicious and benign client updates. Sibai consists of two phases: (I) Offline training, where the Siamese network learns from shadow models, including both benign and malicious examples; (II) Online inference, where Sibai evaluates local models against the global model during federation to identify and exclude poisoned models based on learned criteria. + +Architecture. As shown in Fig. 1, Sibai takes the global model and a local model as input, processing them through four components: Dimensionality Reduction, Twin Networks, Distance Function, and Decision Network. The output is a binary decision indicating the likelihood of a local model being malicious. Detailed descriptions of each component follow in subsequent sections. + +![](images/91ff721698c8f93b3b5ec1f36c6256140e16cda3a97b533f7f90f17c4367aa44.jpg) +Figure 1. Sibai Architecture + +Dimensionality Reduction. Sibai processes models of varying sizes by condensing input data while retaining key features. Local models, often containing hundreds of thousands of parameters, are broken into slices, and Principal Component Analysis (PCA) extracts the most significant components for each slice. PCA identifies and ranks features by variance, ensuring compact yet informative representations. Additional metrics, specifically Cosine and Euclidean distances to the global model, that were shown to be effective by prior research [5, 17, 27, 43], are also incorporated into the input vector. Experiments showed that these metrics consistently improve detection accuracy by up to $10\%$ . More details about the implementation of the dimensionality reduction pipeline are provided in Appendix C.2. + +Twin Networks. The Twin Networks consist of two identical 1D Convolutional Neural Networks (CNNs), serving as the core of the Siamese network. Each CNN processes one of the two pre-processed data samples, acting as a feature extractor. They map the samples to fixed-length embedding vectors using the same filters. These two vectors can then be compared to assess similarity. A high dissimilarity between embedding vectors indicates potential maliciousness, as the twins are explicitly trained to focus on the distinguishing features of malicious local models. Appendix C.4 provides further details about the architecture of the twin models. + +Distance Function. The embedding vectors generated by the Twin Networks are compared using the $L_{2}$ -distance, which demonstrated superior performance compared to other metrics like $L_{1}$ -distance or cosine distance. This distance is used in the triplet loss function, which is utilized to train the twin networks. + +Decision Network. To avoid using a fixed distance threshold, a fully connected neural network, called a decision network, is used to classify the output of the distance function. The last layer features a sigmoid activation function [38] that produces a value between 0 and 1, with values closer to 1 indicating a higher probability of maliciousness. We classify local models that when compared to the global model produce an output $>0.5$ as malicious. + +**Shadow Model Dataset.** Sibai's performance depends on a shadow model dataset used during offline training. This dataset should cover a range of attacks and scenarios to ensure effective generalization by the Siamese Network. To create the shadow model set we simulate FL across various federations with different non-i.i.d. degrees attacks and PDRs. Incorporating stealthy models from adaptive adversaries with low PDR allows Sibai to focus on nuanced criteria while including poisoned models with high PDR enhances generalization. It's crucial to maintain consistent local model architectures in the shadow set, and federations with varied input or output layer sizes cannot be mixed. Further details are provided in Appendix C.3. + +# 4. Evaluation + +This section evaluates Sibai during online inference against five poisoning attacks in challenging scenarios, including complex non-i.i.d. settings, low PDRs, and adaptive adversaries. + +Hardware. Training and evaluation were performed on a Linux server with an Intel Xeon Gold 6342 CPU, 251GB RAM, CUDA [44], and an NVIDIA A30 GPU with 24GB memory. + +Reference Defenses. We compare Sibai with eight defenses: HDBSCAN [35] which represents naive clustering via cosine distance, Krum and M-Krum [5] which select the new global model based on Euclidean distance, MESAS [27] which tries to counter adaptive adversary by using six metrics, FLTrust [8] which uses trusted reference-based weighting, FoolsGold [17] which analyzes the final layer cosine distance, and the robust aggregation method Trimmed Mean/Median [60]. Hyperparameters are detailed in Appendix E. However, it is important to note that a direct comparison of methods with fundamentally different defense paradigms, such as a DF-based method like ours and IR-based defenses like FLTrust, cannot capture the whole picture since these paradigms approach poisoning defense with fundamentally different philosophies. + +Attacks. Sibai is tested against five attacks: Random label flipping [27], Pixel trigger backdoor [19], Blend backdoor [10], Edge case backdoor [54], and Semantic backdoor [3]. These attacks are introduced in more detail in Appendix D + +Datasets and Models. We evaluate federations collaboratively training Squeezezenet [23] on CIFAR-10 [30]. Additionally, we conducted experiments with a custom DNN and MNIST, which are provided in Sec. 4.5. + +**Shadow Model Dataset.** Sibai's offline training uses 520 simulated federations, each with 20 clients (11 benign, 9 malicious), totaling 5720 benign and 4145 malicious models. Scenarios include i.i.d. and non-i.i.d. distributions with varying degrees (e.g., 0.3, 0.6, 0.9 1-class non-i.i.d. and inter-client non-i.i.d. [27]). Realistic and exaggerated settings enhance the distinction between benign and malicious models. Each federation experiences one of the five attacks. Federations with local backdoor accuracies below $45\%$ were excluded to ensure poisoned models exhibit detectable traits. Distribution details are in Tab. 1 and a detailed breakdown is listed in Appendix H. + +Default Federation Parameters. We evaluate Sibai in simulated federations collaboratively training Squeezenet [23] with SGD (learning rate: 0.01, momentum: 0.9, decay: 0.005) and a batch size of 64. Each client trains locally for 10 epochs on i.i.d. datasets with 2560 samples. The global model is initialized with pre-trained weights from + +
CategoryNumber of Federations
IID/non-i.i.d.i.i.d.225
0.3 1-class non-i.i.d.66
0.6 1-class non-i.i.d.71
0.9 1-class non-i.i.d.96
inter-client non-i.i.d. [27]62
AttacksRandom [27]90
Pixel [19]91
Blend [10]119
Edge [54]121
Semantic [3]99
Total Number of Federations520
+ +Table 1. Composition of the shadow model dataset + +PyTorch [46] and undergoes 49 benign rounds to mitigate early backdoor persistence [3], followed by poisoning in the 50th round. + +Sibai Parameters. During offline training, the twin and decision networks are trained separately for 25 epochs using SGD. The twin networks use a learning rate of 0.001, while the decision network uses 0.005. Model slices contain up to 20,000 parameters, with larger layers partitioned. PCA retains 200 components per slice. Siamese network training does not use batch processing due to memory constraints, while the decision network processes batches of 16 distances derived from 8 Siamese network triples. + +Metrics. We evaluate Sibai with precision, recall, and F1 score. True positives are correctly flagged malicious models, and true negatives are benign models. Precision measures the ratio of correctly identified malicious models to all flagged models, while recall assesses the proportion of flagged malicious models among all actual malicious models. The F1 score balances these metrics. Additionally, we track the global model's BA to evaluate attack mitigation and MA changes to assess benign scenario impact. The MA diff. represents the difference between the MA after Sibai application and perfect benign aggregation. + +# 4.1. IID Scenarios + +This section evaluates Sibai's online performance after training on the shadow model set. We test Sibai on 12 unseen federation scenarios1 of increasing difficulty and adversary stealthiness. In the baseline scenario (Tab. 2), federations training Squeezenet [23] on CIFAR-10 [30] in an i.i.d. setting were attacked with the five poisoning attacks. Adversaries employed a PDR of 0.5, which is a relatively overt attack. Sibai flawlessly detected all malicious clients without false positives. Scenario (1) increased stealthiness by minimizing PDRs while maintaining local BAs above $45\%$ . The blend [10] and semantic [3] backdoors used a PDR of 0.04 due to their stronger effects, while the edge case backdoor [54] required a PDR of 0.16 for consistent backdoor introduction. These adjustments mirror real-world attacker strategies. Sibai achieved perfect recall and F1 scores for + +
ScenarioRandomPixelBlendEdgeSemantic
①PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.09%0.20%18.87%0.0%
①PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall0.88881.00001.00001.00000.7777
F10.94111.00001.00001.00000.8750
MA diff.-0.88%0.0%0.0%0.0%+4.12%
BA-10.43%0.0%18.87%0.0%
+ +$\boxed{0} =$ Default + high PDRs +$\widetilde{\mathbf{1}} =$ Default and low PDRs + +Table 2. Detection performance of Sibai in i.i.d. scenarios + +
qScenarioRandomPixelBlendEdgeSemantic
0.6②PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-2.65%0.20%6.12%0.0%
③PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00000.88881.0000
F11.00001.00001.00000.94111.0000
MA diff.0.0%0.0%0.0%-6.71%0.0%
BA-2.65%0.21%8.67%0.0%
0.9④PDR0.50.50.50.50.5
Precision0.87501.00001.00000.90001.0000
Recall0.77770.88881.00001.00001.0000
F10.82350.94111.00000.94731.0000
MA diff.+4.80%+1.08%0.0%+2.44%0.0%
BA-0.11%0.0%0.0%0.0%
⑤PDR0.40.080.040.160.04
Precision0.81810.75001.00000.81811.0000
Recall1.00001.00001.00001.00000.8888
F10.90000.85711.00000.90000.9411
MA diff.+0.48%-4.26%0.0%-1.65%+0.48%
BA-0.0%0.21%0.0%0.0%
+ +$(2) =$ Default $+0.61$ -class non-i.i.d. $^+$ high PDR +$\widehat{(3)} =$ Default $+0.61$ -class non-i.i.d. $^+$ low PDR +$(4) =$ Default $+0.9$ 1-class non-i.i.d. $^+$ high PDR +$(5) =$ Default + 0.91-class non-i.i.d. + low PDR + +pixel trigger and edge case backdoors, identifying all malicious clients. In the random label flip and semantic backdoor attacks, one and two malicious clients, respectively, evaded detection. However, Sibai identified over $77\%$ of malicious clients in all cases, preventing effective poisoning. Across all i.i.d. scenarios, Sibai maintained BA below $19\%$ , ensuring attacks were unsuccessful. Interestingly, in the semantic backdoor scenario, Sibai mistakenly aggregated two malicious models, resulting in a slight MA improvement over perfect aggregation. This occurred because the aggregated models provided additional knowledge, but the low PDR and limited bypassed defenses still prevented backdoor establishment. + +# 4.2. Non-IID Scenarios + +In this section, we evaluate Sibai in challenging non-i.i.d. scenarios, including varying degrees of 1-class non-i.i.d. and inter-client non-i.i.d. setups as described in Sect. 2 and [27]. + +Table 3. Detection performance of Sibai in 1-class non-i.i.d. scenarios, with different level of disproportionality $q$ + +
ScenarioRandomPixelBlendEdgeSemantic
6PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00000.9000
Recall0.77770.88881.00001.00001.0000
F10.87500.94111.00001.00000.9473
MA diff.+2.61%+0.86%0.0%0.0%0.0%
BA-4.03%0.0%0.0%0.0%
7PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.66661.00000.8888
F11.00000.94110.80001.00000.9411
MA diff.0.0%+1.38%-1.43%0.0%-1.48%
BA-13.57%0.0%0.0%0.0%
+ +$\widehat{(6)} =$ Default $^+$ inter-client non-i.i.d. $^+$ high PDR +$\widetilde{7} =$ Default + inter-client non-i.i.d. + low PDR + +Table 4. Detection performance of Sibai in inter-client non-i.i.d. scenarios [27] + +0.6 1-class non-i.i.d. For the scenarios $\boxed{2}$ and $\boxed{3} q = 0.6$ was used. Sibai achieved perfect detection in the 0.5 PDR case $\boxed{2}$ . For the low PDRs scenario $\boxed{3}$ , F1 scores were 1.0 for all attacks except the edge case backdoor, where one malicious client evaded detection. Despite this, the MA dropped by less than $7\%$ , and BA stayed below $9\%$ , ensuring no effective backdoor was introduced. + +0.9 1-class non-i.i.d. In the more extreme 0.9 1-class non-i.i.d. scenarios (4) and (5) in Tab. 10, Sibai occasionally flagged benign models, with three false positives in the worst-case pixel trigger backdoor scenario (5). However, these errors had minimal impact on MA. Sibai consistently eliminated enough malicious models to prevent effective backdoors, as no attack achieved a BA higher than $1\%$ . + +Inter-client non-i.i.d. Inter-client non-i.i.d. features erratic and highly unbalanced data distributions, reflecting real-world scenarios. In the 0.5 PDR scenario (6), Sibai achieved perfect detection for blend and edge case backdoors, perfect precision for untargeted and pixel attacks, and perfect recall for the semantic attack. Only one false positive occurred during the semantic attack. The lowest F1 score of 0.8750 occurred during the untargeted attack) In the attacker-adapted low PDR scenario (7), Sibai maintained perfect precision for all attacks despite high non-i.i.d.-ness. Recall was lower, yielding F1 scores of 0.9411 for pixel and semantic backdoors and 0.8000 for the blend attack. Nonetheless, MA was minimally affected, and BA remained low across all attacks. Overall, Sibai demonstrated robust performance in 1-class and inter-client non-i.i.d. scenarios, with minimal false positives and no effective backdoors. + +# 4.3. Adaptive Adversaries + +Sibai incorporates slice-wise Euclidean and cosine distances to the global model into its input vectors to enhance detection. While effective, these metrics may be vulnerable to adaptive attackers. To test Sibai's robustness, we evaluated four scenarios involving attackers using Bagdasaryan et al.'s [3] adaptation method with an $\alpha$ parameter balancing adaptation + +
Form of AdaptationScenarioRandomPixelBlendEdgeSemantic
Cosine dist.⑧PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18.87%0.0%
⑨PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001,00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.14%18.87%0.0%
Euclidian dist.⑩PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18. 87%0.0%
⑪PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.88881.00001.0000
F11.00000.94110.94111.00001.0000
MA diff.0.0%+3.1%-6.3%0.0%0.0%
BA-5.98%0.0%18.87%0.0%
+ +$\boxed{8} =$ Default $^+$ cosine distance adaption with $0.3\alpha$ by mal. clients $^+$ high PDR +9 = Default + cosine distance adaption with 0.3 α by mal. clients + low PDR +$\boxed{10} =$ Default $^+$ Euclidean distance adaption with $0.3\alpha$ by mal. clients $^+$ high PDR +$\widehat{(11)} =$ Default $^+$ Euclidean distance adaption with $0.3\alpha$ by mal. clients $^+$ low PDR + +and task losses. Results for $\alpha = 0.3$ , shown in Tab. 5, highlight Sibai's resilience. In cosine-adapted scenarios, Sibai achieved perfect F1 scores (1.0), showing immunity to such attacks. Similarly, Sibai maintained perfect performance in Euclidean-adapted Scenario 12. In low-PDR Euclidean-adapted scenarios, Sibai showed flawless precision, with one false negative for pixel and blend backdoors. However, no attack established an effective backdoor, with BA remaining below $19\%$ . Overall, Sibai demonstrates strong resilience against adaptive attacks, proving it does not overly rely on engineered features. Its ability to internalize specific patterns makes it impractical for attackers to adapt using Bagdasaryan et al.'s method. + +# 4.4. Comparison to Existing Defenses + +To assess the effectiveness of Sibai compared to metric-based poisoning detection methods in FL, we also evaluate existing defenses in some of our previous scenarios. Due to space limitations, we focus on a selection of the most representative scenarios, including scenario 1 as the i.i.d. baseline, scenario 4 as the highest degree of 1-class non-i.i.d., and scenario 6 to showcase inter-client non-i.i.d. We also include scenario 0 as the overall simplest scenario and scenario 11 to cover adaptive adversaries. Since some reference defenses rely on modified aggregation methods instead of strictly flagging and filtering out malicious models, calculating F1 scores for all reference defenses faced challenges. Further details are provided in Appendix F. + +Table 5. Detection performance of Sibai against adaptive attackers in i.i.d. scenarios + +
ScenarioDefenseRandomPixelBlendEdgeSemantic
0PDR0.50.50.50.50.5
MESAS0.90000.90000.90001.00000.9473
Krum0.57140.57140.57890.57140.5714
M-Krum0.26080.26080.64280.26080.2608
HDBSCAN1.00000.00000.11111.00001.0000
FLTrust (0.09)0.00000.26080.33330.00000.2608
Sibai (ours)1.00001.00001.00001.00001.0000
1PDR0.40.080.040.160.04
MESAS0.94730.00000.18180.16660.3076
Krum0.57140.64280.64280.57140.5714
M-Krum0.26080.60860.60860.26080.4347
HDBSCAN0.00000.55550.44440.00000.1176
FLTrust (0.09)0.09520.66660.57140.51850.5714
Sibai (ours)0.94111.00001.00001.00000.8750
4PDR0.50.50.50.50.5
MESAS0.00000.16660.80000.94730.9000
Krum0.57140.57140.57140.64280.6428
M-Krum0.26080.69560.69560.78260.7826
HDBSCAN0.00000.40000.77770.82350.8235
FLTrust (0.09)0.57140.88880.88880.88880.8571
Sibai (ours)0.82350.94111.00000.94731.0000
6PDR0.50.50.50.50.5
MESAS0.64280.47610.21050.66660.3157
Krum0.57140.57140.57140.57140.5714
M-Krum0.26080.26080.26080.52170.3478
HDBSCAN0.00000.00000.14280.30760.1428
FLTrust (0.09)0.00000.57140.57140.57140.6206
Sibai (ours)0.87500.94111.00001.00000.9473
11PDR0.40.080.040.160.04
MESAS1.00000.66660.90000.61530.9473
Krum0.57140.57140.64280.64280.6428
M-Krum0.26080.26080.78260.60860.6956
HDBSCAN0.00001.00001.00001.00000.2000
FLTrust (0.09)0.57140.62060.66660.62060.6206
Sibai (ours)1.00000.94110.94111.00001.0000
+ +Table 6. F1 score comparison between Sibai and current FL poisoning defenses + +As shown in Tab. 6, Sibai achieved superior F1 scores across most scenarios. In high PDR i.i.d. scenario $\boxed{0}$ , it matched or outperformed all defenses. In low PDR i.i.d. scenario $\boxed{1}$ , Sibai achieved perfect F1 scores for pixel, blend, and edge case backdoors, outperforming MESAS for these attacks. Even when MESAS maintained low BA, it highlighted the fragility of backdoor attacks in low PDR settings. For 1-class non-i.i.d. scenario $\boxed{4}$ , Sibai significantly outperformed other methods, and in inter-client non-i.i.d. scenario $\boxed{7}$ , its worst-case F1 score of 0.8750 surpassed the best scores of other defenses. In adaptive scenario $\boxed{11}$ , Sibai slightly trailed HDBSCAN for pixel and blend backdoors however, it excelled in the detection of the semantic and untargeted attacks, achieving perfect F1 scores. Overall, Sibai consistently outperformed existing methods in identifying malicious models across diverse scenarios and attacks. + +Table 7, also considered methods for which the F1 was unavailable, specifically FoolsGold [17] and trimmed mean/median [60]. Sibai showed performance close to perfect benign aggregation, effectively preventing poisoning without significant MA reduction. Notably, Sibai and MESAS achieved low BA in challenging scenarios (4) and (6), where other defenses failed to stop backdoors. Although FLTrust [8], produced higher MA in non-i.i.d. scenarios by using a weighted aggregation that particularly emphasizes the old global model, it also often allowed higher BA and + +
ScenarioDefenseRandom MA diff.Pixel MA diff.BABlend MA diff.BAEdge MA diff.BASemantic MA diff.BA
①PDR0.40.080.040.160.04
benign agg.0.00.010.40.00.00.018.80.00.0
MESAS [27]-0.1+5.19.50.020.4+0.27.10.00.0
Krum [5]-35.0-43.20.0-14.10.0-35.50.0-21.1100.0
M-Krum [5]-24.5-8.512.5-9.52.9-16.30.0-11.40.0
FoolsGold [17]-6.3+3.011.0-1.718.8-8.36.1+0.10.0
HDBSCAN [35]-16.0-1.910.9-6.511.4-13.80.0-5.80.0
Trimmed mean [60]-8.3+3.510.2-2.512.8-9.06.6-0.60.0
Trimmed median [60]-22.3-2.65.2-11.60.1-14.72.5-6.40.0
FLTrust [8]-22.3-0.117.2-5.213.0+1.516.8-4.80.0
Sibai (ours)-0.80.010.40.00.00.018.80.00.0
④PDR0.40.50.50.50.5
benign agg.0.00.00.00.00.00.00.00.00.0
MESAS [27]-5.7+0.334.3+5.00.6+2.40.0+10.60.0
Krum [5]-7.7-6.1100.0-0.2100.0-4.80.0-5.30.0
M-Krum [5]-2.1+5.51.7+1.0100.0-5.10.0+4.00.0
FoolsGold [17]-2.0-5.030.2-0.2100.0-10.80.0-7.30.0
HDBSCAN [35]-7.5+1.013.2+1.40.0-8.70.0-0.920.0
Trimmed mean [60]-5.3-0.428.9-0.2100.0-11.10.0-7.840.0
Trimmed median [60]-6.6-5.10.0-0.4100.0-10.40.0-6.50.0
FLTrust [8]+19.5+14.753.1+10.898.2+4.80.0+11.120.0
Sibai (ours)+4.8+1.00.00.00.0+2.40.00.00.0
⑥PDR0.50.50.50.50.5
benign agg.0.013.92.90.02.30.02.40.00.0
MESAS [27]+11.9-0.31.6-0.132.7+9.70.00.00.0
Krum [5]+4.8-3.9100.0-5.1100.0+16.885.7+18.780.0
M-Krum [5]-1.6-3.085.3-0.3100.0+2.90.00.00.0
FoolsGold [17]-1.6+1.57.0-5.0100.0-2.00.00.00.0
HDBSCAN [35]-5.8-0.719.0+3.899.9-0.30.0-0.40.0
Trimmed mean [60]-0.5+1.17.3-3.8.1100.0-3.00.00.00.0
Trimmed median [60]-2.5-3.50.0-4.9100.0-2.30.0-0.30.0
FLTrust [8]+14.7+14.347.4+13.099.8+14.00.0+8.320.0
Sibai (ours)+2.6+0.84.00.00.00.00.00.00.0
+ +therefore failed to prevent effective backdoors. In contrast, Sibai excelled at malicious model detection and consistently maintained low BA while achieving MAs aligned with benign aggregation, making Sibai an overall reliable defense. + +# 4.5. Alternative Dataset and DNN Architecture + +To demonstrate that Sibai is not limited to SqueezeNET or the CIFAR-10 dataset, we conducted additional experiments using a custom CNN and the MNIST dataset [12]. The CNN model consists of 2 convolutional layers, 2 max-pooling layers, and 3 linear layers. We excluded the dataset-specific semantic and edge case backdoors and instead included a label swap attack [27], which targets samples of one label and swaps them to another to reduce the global model's accuracy. The shadow-model dataset for these experiments comprised 1401 federations. The F1 scores in Tab. 8 indicate that Sibai effectively detected malicious models, and therefore that the approach is not specific to SqueezeNET or CIFAR-10. + +# 5. Multiple Poisonings in one Model + +To evaluate Sibai's effectiveness in detecting models poisoned with multiple backdoors simultaneously, we conducted experiments using federations where each malicious local model contained both a blend backdoor [10] and a pixel trigger backdoor [19] simultaneously. For those local models, the PDR quantifies the ratio of the total poisoned samples, i.e. the sum used for both backdoors, to the com + +Table 7. Comparison of MA diff. and BA (both in percentage) between Sibai and existing FL poisoning defenses. The best value for each attack in each scenario is highlighted in bold. + +
PDRDistributionRandom [27]Pixel [19]Blend [10]Label Swap [27]
0.4i.i.d.1.00001.00001.00001.0000
0.40.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.40.6 l-class non-i.i.d.0.69230.72000.72000.7200
0.40.9 l-class non-i.i.d.0.87500.94111.00001.0000
0.4inter-client non-i.i.d.0.90000.94110.94730.9000
0.5i.i.d.1.00001.00001.00001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.50.6 l-class non-i.i.d.0.72000.66660.72000.7200
0.50.9 l-class non-i.i.d.1.00000.94111.00001.0000
0.5inter-client non-i.i.d.0.94730.94110.94730.9000
+ +Table 8. F1 score of Sibai in federations training a small CNN with the MNIST dataset + +
PDRDistributionPrecisionRecallF1
0.2i.i.d.1.00001.0001.0000
0.20.3 l-class non-i.i.d.1.00001.00001.0000
0.20.6 l-class non-i.i.d.1.00001.00001.0000
0.20.9 l-class non-i.i.d.1.00000.88880.9411
0.2inter-client non-i.i.d.1.00000.88880.9411
0.5i.i.d.1.00001.0001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.0000
0.50.6 l-class non-i.i.d.1.00001.00001.0000
0.50.9 l-class non-i.i.d.1.00001.00001.0000
0.5inter-client non-i.i.d.1.00001.00001.0000
+ +Table 9. Detection performance of Sibai against local models poisoned with two backdoors (pixel trigger backdoor [19] and blend backdoor [10]) simultaneously in different scenarios + +plete local model dataset. In our implementation of this multi-backdoor half of the poisoned samples were used to introduce the pixel trigger backdoor, while the other half was used to apply the blend backdoor. Importantly, our shadow model dataset, used to train Sibai, did not include + +any models with such multi-backdoors. As illustrated in Tab. 9, Sibai consistently identified local models containing multiple backdoors across all tested scenarios. + +# 6. Related Work + +This section provides a brief overview of relevant literature. First, Sect. 6.1 outlines poisoning defenses tailored for FL. Then, Sect. 6.2 discusses machine learning methods for general offline poisoning analysis. + +# 6.1. Defenses against Poisoning Attacks in Federated Learning + +There exists a diverse range of techniques that employ fixed metrics to detect poisoning attacks in FL. FoolsGold [17] uses a client-specific update history which is compared to the update histories of other clients using the cosine similarity. Krum [5] selects a new global model based on the Euclidean distance to other local models. M-Krum [5] extends this method to select multiple models simultaneously. Trimmed-Mean/Median [60] are robust aggregation methods that replace FedAVG [36] and calculate the updated global model by using coordinate-wise means or medians from local models. FLAME [43] clusters models using HDBSCAN [35] and filters out smaller clusters. It also applies differential privacy techniques, weight clipping, and noisng. DeepSight [47] combines filtering with differential privacy. It relies on two metrics: The cosine distance between models and values extracted from the output layer of the model. Auror [48] uses K-Means clustering on model parameters, however, unfortunately, it exhibited an extreme performance overhead in our experiments. FLTrust [8] uses a weighted aggregation system for which it assigns weights to local models based on their cosine and Euclidean distances to a benign reference model, which is trained server-side on a small trusted dataset. FLDetector [63] is a historical update-based defense that mathematically predicts client updates and compares the predictions to the actual updates. MESAS [27] was specifically designed to operate in non-i.i.d. scenario and counter adaptive adversaries and employs six metrics. These approaches differ from Sibai as they rely on fixed metrics or criteria, making them susceptible to adaptive attacks [27]. Defenses focusing on specific layers, like FoolsGold [17], are prone to adversaries exploiting advanced model poisoning techniques specifically targeting those layers. Similarly, methods that heavily focus on cosine or Euclidean distance are often vulnerable to adaptation methods using additional losses [3, 27], as detailed in Sect. 2. These rigid metrics also face challenges in non-i.i.d. scenarios, especially in inter-client non-i.i.d. settings, where it has been demonstrated that most defenses, except MESAS [27] which was specifically designed to handle such scenarios, face difficulties. In contrast, Sibai dynamically learns criteria based on its shadow model dataset, making it challenging for attackers + +to adapt without access to the Siamese Network. Sibai can also differentiate between variations in local models due to data distribution and those caused by poisoning. + +# 6.2. Non-FL Specific Offline Poisoning Detection + +This section discusses several machine-learning-based detection methods against poisoning attacks, not specifically designed for FL. Xu et al. [58] propose Meta Neural Trojan Detection (MNTD), a black-box backdoor detection approach utilizing a meta-classifier that assesses if a model's output behavior on a given input aligns more closely with a set of known benign models or a set of known malicious models. Huang et al. [21] introduce the one-pixel signature concept for DNNs, consisting of a set of single-pixel adversarial perturbations that have the highest probability of triggering misclassifications on a specific benign image dataset. Subsequently, a meta-classifier was trained on these signatures to discern between poisoned and benign models. Kolouri et al. [25] use universal litmus patterns and a meta-classifier to detect backdoors across different network architectures. Chen et al. [9] present DeepInspect, a black-box method utilizing model inversion and anomaly detection for backdoor detection. Huang et al. [22] also introduced Neuroninspect, employing explanation heatmaps and outlier detection for identifying abnormal neurons. Unlike these methods, Sibai is tailored to FL, avoiding intensive preprocessing steps. DeepInspect [9] involves extensive reverse engineering of potential triggers for every label of each local model, and Huang et al. [21] requires testing every local model with a large set of single-pixel adversarial perturbations. Conducting such intricate analyses for every model in a federation during each round is impractical, as these methods were originally designed for detailed offline analysis of individual models. Sibai instead only uses PCA as the sole dimensionality reduction mechanism, facilitating straightforward inference. + +# 7. Conclusion + +To counter non-i.i.d. data distribution and adaptive adversaries in FL, we introduce Siamese Backdoor Inspector (Sibai), a novel FL poisoning defense mechanism utilizing a Siamese network meta-classifier. Inspired by successful offline backdoor detection [21, 25, 58], Sibai dynamically learns discriminative features directly from PCA decompositions of local models. In diverse evaluations, Sibai consistently achieves high detection rates without significantly impacting global model performance. It effectively prevents backdoor establishment and outperforms current defenses in terms of f1 score while maintaining low false positive rates. Future work may explore selectively applying influence reduction to models identified by Sibai for targeted defense. + +# 8. Acknowledgment + +This research has been funded by the Federal Ministry of Education and Research of Germany (BMBF) within the program Digital. Sicher. Souverän. "in the project"Erkennung von Angriffen gegen IoT-Netzwerke in Smart Homes - IoTGuard" (project number 16KIS1919). + +# References + +[1] Sebastien Andreina, Giorgia Azzurra Marson, Helen Möllering, and Ghassan Karame. Baffle: Backdoor detection via feedback-based federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), pages 852-863. IEEE, 2021. 1, 3 +[2] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. In 30th USENIX Security Symposium (USENIX Security 21), pages 1505-1521, 2021. 1 +[3] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 4, 5, 8, 13, 14, 15, 16, 17, 18, 19 +[4] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International Conference on Machine Learning, pages 634-643. PMLR, 2019. 1 +[5] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 3, 4, 7, 8, 15, 16 +[6] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 3 +[7] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a" siamese" time delay neural network. Advances in neural information processing systems, 6, 1993. 2 +[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. Network and Distributed Systems Security (NDSS) Symposium 2021, 2021. 1, 4, 6, 7, 8, 15, 16 +[9] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In *IJCAI*, page 8, 2019. 1, 8 +[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 4, 7, 13, 14, 17, 18, 19 +[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual computer security applications conference, pages 554-569, 2021. 1 + +[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 7 +[13] Elena Deza, Michel Marie Deza, Michel Marie Deza, and Elena Deza. Encyclopedia of distances. Springer, 2009. 2 +[14] Cynthia Dwork et al. The mathematics of information coding, extraction, and distribution. The IMA Volumes in Mathematics and its applications, 107:31-47, 1999. 12 +[15] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 1 +[16] Hossein Fereidooni, Jan König, Phillip Rieger, Marco Chilese, Bora Gökbakan, Moritz Finke, Alexandra Dmitrienko, and Ahmad-Reza Sadeghi. Authentisense: A scalable behavioral biometrics authentication scheme using few-shot learning for mobile platforms. arXiv preprint arXiv:2302.02740, 2023. 3 +[17] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2020), pages 301-316, 2020. 1, 3, 4, 6, 7, 8, 15, 16 +[18] Yansong Gao, Bao Gia Doan, Zhi Zhang, Siqi Ma, Jiliang Zhang, Anmin Fu, Surya Nepal, and Hyoungshick Kim. Backdoor attacks and countermeasures on deep learning: A comprehensive review. arXiv preprint arXiv:2007.10760, 2020. 1, 2 +[19] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 2, 4, 7, 13, 17, 18, 19 +[20] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-Based Pattern Recognition: Third International Workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 2 +[21] Shanjiaoyang Huang, Weiqi Peng, Zhiwei Jia, and Zhuowen Tu. One-pixel signature: Characterizing cnn models for backdoor detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 326-341. Springer, 2020. 1, 8 +[22] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 1, 8 +[23] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and i 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 4 +[24] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop. Lille, 2015. 2, 3 +[25] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 1, 8 + +[26] Jakub Konečný, H Brendan McMahan, Daniel Ramage, and Peter Richtárik. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint arXiv:1610.02527, 2016. 2 +[27] Torsten Krauß and Alexandra Dmitrienko. Mesas: Poisoning defense for federated learning resilient against adaptive attackers. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, page 1526-1540, New York, NY, USA, 2023. Association for Computing Machinery. 1, 2, 3, 4, 5, 7, 8, 13, 15, 16, 17, 18, 19 +[28] Torsten Krauß, Raphael Götz, and Alexandra Dmitrienko. Security of nvme offloaded data in large-scale machine learning. In Computer Security - ESORICS 2023, pages 143-163, Cham, 2024. Springer Nature Switzerland. 2 +[29] Torsten Krauß, Jan König, Alexandra Dmitrienko, and Christian Kanzow. Automatic Adversarial Adaption for Stealthy Poisoning Attacks in Federated Learning. Network and Distributed Systems Security (NDSS) Symposium 2024, 2024. 1 +[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 13, 14, 15 +[31] Liping Li, Wei Xu, Tianyi Chen, Georgios B Giannakis, and Qing Ling. Rsa: Byzantine-robust stochastic aggregation methods for distributed learning from heterogeneous datasets. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1544–1551, 2019. 1 +[32] Yijing Li, Xiaofeng Tao, Xuefei Zhang, Junjie Liu, and Jin Xu. Privacy-preserved federated learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(7):8423-8434, 2021. 1 +[33] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1 +[34] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 2 +[35] Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017. 3, 4, 7, 8, 15, 16 +[36] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 8, 15 +[37] Luis Muñoz-González, Kenneth T Co, and Emil C Lupu. Byzantine-robust federated machine learning through adaptive model averaging. arXiv preprint arXiv:1909.05125, 2019. 3 +[38] Sridhar Narayan. The generalized sigmoid activation function: Competitive supervised learning. Information sciences, 99 (1-2):69-82, 1997. 3, 13 +[39] Mohammad Naseri, Jamie Hayes, and Emiliano De Cristofaro. Local and central differential privacy for robustness and privacy in federated learning. arXiv preprint arXiv:2009.03561, 2020. 1 + +[40] Blaine Nelson, Marco Barreno, Fuching Jack Chi, Anthony D Joseph, Benjamin IP Rubinstein, Udam Saini, Charles Sutton, J Doug Tygar, and Kai Xia. Exploiting machine learning to subvert your spam filter. LEET, 8(1-9):16-17, 2008. 1 +[41] Anh Nguyen, Tuong Do, Minh Tran, Binh X Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, and Quang D Tran. Deep federated learning for autonomous driving. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1824-1830. IEEE, 2022. 1 +[42] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), pages 1-7, 2020. 1 +[43] Thien Duc Nguyen, Phillip Rieger, Roberta De Viti, Huili Chen, Björn B Brandenburg, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, et al. {FLAME}: Taming backdoors in federated learning. In 31st USENIX Security Symposium (USENIX Security 22), pages 1415–1432, 2022. 1, 3, 8 +[44] NVIDIA, Péter Vingelmann, and Frank H.P. Fitzek. Cuda, release: 10.2.89, 2020. 4 +[45] Xudong Pan, Mi Zhang, Beina Sheng, Jiaming Zhu, and Min Yang. Hidden trigger backdoor attack on {NLP} models via linguistic style manipulation. In 31st USENIX Security Symposium (USENIX Security 22), pages 3611-3628, 2022. 1 +[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 4, 12 +[47] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. NNetwork and Distributed Systems Security (NDSS) Symposium 2022, 2022. 1, 8 +[48] Shiqi Shen, Shruti Tople, and Prateek Saxena. Auror: Defending against poisoning attacks in collaborative deep learning systems. In Proceedings of the 32nd Annual Conference on Computer Security Applications, pages 508-519, 2016. 1, 8, 16 +[49] Gan Sun, Yang Cong, Jiahua Dong, Qiang Wang, Lingjuan Lyu, and Ji Liu. Data poisoning attacks on federated machine learning. IEEE Internet of Things Journal, 9(13):11365-11375, 2021. 2 +[50] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963, 2019. 1 +[51] Zhiyi Tian, Lei Cui, Jie Liang, and Shui Yu. A comprehensive survey on poisoning attacks and countermeasures in machine learning. ACM Computing Surveys, 55(8):1-35, 2022. 2 +[52] Alexander Turner, Dimitris Tsipras, and Aleksander + +Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1 +[53] Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual. CreateSpace, Scotts Valley, CA, 2009. 12 +[54] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. Advances in Neural Information Processing Systems, 33:16070-16084, 2020. 1, 2, 4, 13, 14, 17, 18, 19 +[55] Zhaoxian Wu, Qing Ling, Tianyi Chen, and Georgios B Giannakis. Federated variance-reduced stochastic gradient descent with robustness to byzantine attacks. IEEE Transactions on Signal Processing, 68:4583-4596, 2020. 1 +[56] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations, 2019. 1 +[57] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1 +[58] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. In 2021 IEEE Symposium on Security and Privacy (SP), pages 103-120. IEEE, 2021. 1, 8 +[59] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 2 +[60] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 4, 6, 7, 8, 15, 16 +[61] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 1 +[62] Hongyi Zhang, Jan Bosch, and Helena Holmström Olsson. End-to-end federated learning for autonomous driving vehicles. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 1 +[63] Zaixi Zhang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Fl detector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2545-2555, 2022. 1, 8 \ No newline at end of file diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/images.zip b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ea27a7bb996cec12cfb42b68a4d3d0b901081da5 --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90dbfbe033d504940a3ae0445a979c861a0e0951a9623c49e76b624712d198bb +size 496092 diff --git a/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/layout.json b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76656580bf4f9b29fd57d75537111fdc027b1357 --- /dev/null +++ b/2025/Sibai_ A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning/layout.json @@ -0,0 +1,7803 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 62, + 111, + 547, + 128 + ], + "type": "text", + "content": "Sibai: A Few-Shot Meta-Classifier for Poisoning Detection in Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 150, + 429, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 150, + 429, + 163 + ], + "spans": [ + { + "bbox": [ + 181, + 150, + 429, + 163 + ], + "type": "text", + "content": "Melanie Gotoz, Torsten Krauß, Alexandra Dmitrienko" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 248, + 164, + 361, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 164, + 361, + 177 + ], + "spans": [ + { + "bbox": [ + 248, + 164, + 361, + 177 + ], + "type": "text", + "content": "University of Wurzburg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 178, + 402, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 178, + 402, + 191 + ], + "spans": [ + { + "bbox": [ + 208, + 178, + 402, + 191 + ], + "type": "text", + "content": "Sanderring 2, 97070 Wurzburg, Germany" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 194, + 472, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 194, + 472, + 205 + ], + "spans": [ + { + "bbox": [ + 136, + 194, + 472, + 205 + ], + "type": "text", + "content": "melanie.goetz@uni-wuerzburg.de, torsten.krauss@uni-wuerzburg.de," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 204, + 207, + 400, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 207, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 204, + 207, + 400, + 217 + ], + "type": "text", + "content": "alexandra.dmitrienko@uni-wuerzburg.de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 156, + 245, + 202, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 245, + 202, + 256 + ], + "spans": [ + { + "bbox": [ + 156, + 245, + 202, + 256 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 269, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 297, + 502 + ], + "type": "text", + "content": "Federated Learning (FL) enables collaborative machine learning across decentralized clients without sharing raw data, which offers enhanced privacy and improved performance. However, FL is vulnerable to poisoning attacks, compromising model integrity through both untargeted performance degradation and targeted backdoor attacks. Detecting backdoors in FL is challenging due to their stealthy nature and variability in local datasets. Existing defenses struggle against adaptive adversaries and distinguishing between poisoning and genuine dataset anomalies. This paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning defense for FL. Leveraging the staple few-shot learning technique of Siamese networks, Sibai effectively detects malicious contributions in various scenarios, including settings with strong variations between clients' datasets and encounters with adaptive adversaries. Sibai achieves high detection rates, prevents backdoors, minimizes performance impact, and outperforms eight recent defenses regarding F1 score, poisoning prevention, and consistency across complex scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 522, + 141, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 522, + 141, + 534 + ], + "spans": [ + { + "bbox": [ + 63, + 522, + 141, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 541, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 541, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 541, + 297, + 705 + ], + "type": "text", + "content": "Federated Learning (FL) [59] enables collaborative deep neural network (DNN) training across decentralized clients by sharing model updates instead of raw data, preserving privacy. While FL offers significant benefits, it is susceptible to poisoning attacks by untrustworthy or compromised clients who manipulate training data, labels, hyperparameters, or model parameters. These attacks can be untargeted [15, 31, 55, 57], degrading global model performance, or targeted [2-4, 11, 18, 19, 33, 40, 42, 45, 52, 54, 56], introducing backdoors that misclassify inputs with specific triggers while maintaining normal performance otherwise. Backdoors are particularly dangerous due to their stealth and severe risks in security-critical applications like autonomous vehicles [32, 41, 62]. FL poisoning defenses include in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 246, + 548, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 246, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 312, + 246, + 548, + 327 + ], + "type": "text", + "content": "fluence reduction (IR) [1, 3, 39, 50], robust aggregation (RA) [36, 61], and the more common detection and filtering (DF) [5, 8, 17, 27, 43, 47, 48, 63], which excludes poisoned models before aggregation. Although IR and RA can leverage information from all local models, their methods, such as noising and clipping, often degrade benign performance when applied broadly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 328, + 548, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 328, + 548, + 455 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 548, + 455 + ], + "type": "text", + "content": "Challenges. Most DF-based defenses rely on predefined metrics to detect poisoning, but this approach faces two key challenges. First, adaptive adversaries can tailor poisoned models to evade detection by appearing normal within these metrics [27, 29]. Expanding the metric sets [27] partially addresses this, but still relies on rigid criteria which are vulnerable to adaptation. Second, the natural diversity of non-identically and independently distributed (non-i.i.d.) datasets in real-world scenarios makes distinguishing poisoned models from those reflecting unusual data distributions difficult." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 456, + 548, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 548, + 572 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 548, + 572 + ], + "type": "text", + "content": "Meta-Classifiers. This paper introduces a dynamic meta-classifier approach to FL poisoning defense, which avoids rigid metrics and instead learns to differentiate malicious directly from observing local models. Recent methods [9, 21, 22, 25, 58] show that meta-classifiers can effectively detect poisoned models in an offline scenario. However, applying this to FL poses challenges: training requires a diverse set of so-called shadow models, including both benign and poisoned models, and existing methods depend on time-consuming preprocessing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 574, + 548, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 574, + 548, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 574, + 548, + 654 + ], + "type": "text", + "content": "Contributions. To tackle the aforementioned challenges, this paper introduces the Siamese Backdoor Inspector (Sibai), a novel meta-classifier-based poisoning detection method specifically tailored to function as a DF-based defense for FL. Sibai is based on the staple few-shot learning technique of Siamese networks. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 658, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 548, + 704 + ], + "type": "text", + "content": "- Sibai is the first application of a meta-classifier in DF-based poisoning defenses for FL. Unlike conventional approaches that are reliant on predefined fixed metrics for detection, Sibai dynamically learns the criteria to distin" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 295, + 734, + 315, + 743 + ], + "type": "text", + "content": "3787" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 81, + 295, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 81, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 81, + 295, + 139 + ], + "type": "text", + "content": "guish between benign and malicious local models. Using a dynamic approach based on ML, Sibai poses a challenge for sophisticated attackers that try to adapt to predefined metrics, as it utilizes a diverse set of learned, complex model characteristics for its decision-making." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 140, + 295, + 430 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 62, + 140, + 295, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 295, + 185 + ], + "type": "text", + "content": "- We utilize Siamese Networks in the context of a poisoning detection meta-classifier. These networks are designed to compare and find dissimilarities between two inputs, in the case of Sibai DNN models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 186, + 295, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 278 + ], + "type": "text", + "content": "- Sibai operates directly on flattened DNN representations using a novel layer-wise dimensionality reduction pipeline, using Principal Component Analysis (PCA) enhanced with engineered features. This feature-rich approach improves Sibai's ability to learn complex criteria, enhancing robustness against adaptation. Additionally, it enables Sibai to efficiently inspect all local models in a federation without requiring extensive time-consuming preprocessing steps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 279, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 279, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 279, + 295, + 430 + ], + "type": "text", + "content": "- We extensively evaluate Sibai in various FL scenarios against five different poisoning attacks, assessing its effectiveness against stealthy and adaptive attackers and in scenarios with non-i.i.d. data. To conduct our evaluation, we created a substantial dataset of shadow models, including 5720 benign and 4145 malicious models. Sibai consistently achieves high detection rates across all tested scenarios without notable degradation of the global model's performance. Notably, Sibai successfully prevents the establishment of effective backdoors in all experiments, and outperforms the eight existing FL poisoning defenses, demonstrating higher F1 scores and consistently maintaining the lowest backdoor accuracy across all experiments." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 62, + 433, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 296, + 468 + ], + "type": "text", + "content": "In summary, Sibai introduces a robust ML-based FL poisoning defense to effectively address adaptive adversaries and non-i.i.d. scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 480, + 138, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 480, + 138, + 493 + ], + "spans": [ + { + "bbox": [ + 63, + 480, + 138, + 493 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 500, + 277, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 277, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 277, + 512 + ], + "type": "text", + "content": "This section introduces concepts relevant to this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 515, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 295, + 597 + ], + "type": "text", + "content": "Federated Learning FL [26, 36, 59] enables collaborative ML by training on local data while preserving privacy. In each round, the server sends the global model to selected clients, who train it locally and return updates. The server aggregates updates, often using FedAVG [36], a weighted average with a global learning rate " + }, + { + "bbox": [ + 62, + 515, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 62, + 515, + 295, + 597 + ], + "type": "text", + "content": ". The algorithm is explained in more detail in Appendix A." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 600, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 295, + 704 + ], + "type": "text", + "content": "IID vs. non-IID Data Distributions ML models capture the characteristics of their datasets, emphasizing patterns in the training data. In an ideal identically and independently distributed (IID) scenario, all clients have similar data distributions, making outlier detection an effective strategy to detect poisoning. However, the real world rarely resembles those conditions; instead, clients usually exhibit non-i.i.d. data distributions which complicate the distinction between anomalies and poisoning. For classification tasks, '1-class" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "content": "non-i.i.d.' describes datasets dominated by a single label, with the disproportionality degree regulated by a parameter " + }, + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "inline_equation", + "content": "q \\in [0,1]" + }, + { + "bbox": [ + 313, + 81, + 547, + 139 + ], + "type": "text", + "content": ". Inter-client non-i.i.d. [27] however, is even more complex and allows arbitrary and disjoint distributions, which resemble real-world conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 142, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 547, + 281 + ], + "type": "text", + "content": "Poisoning Attacks Poisoning attacks [51, 54] on DNNs involve attackers altering the training process to manipulate the model. Untargeted attacks lower main task accuracy (MA) by reducing local model accuracy, while targeted (backdoor) attacks embed triggers causing specific misclassifications with minimal impact on MA [3, 18, 34]. Poisoning involves data (e.g., malicious labels or triggers [49]) or model modifications (e.g., altered gradients [28]). This paper mainly considers five specific poisoning types: pixel trigger backdoors [19], blend backdoors [10], edge-case attacks [54], semantic backdoors [3], and random label flipping [27], which are discussed in more detail in Appendix D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 547, + 354 + ], + "type": "text", + "content": "Adaptive Attackers Adaptive attackers can bypass FL defenses that focus on fixed metrics by manipulating their local models to evade detection. Bagdasaryan et al. [3] introduced a method that uses an additional loss component that balances backdoor accuracy (BA) and stealthiness to the metrics the defense considers, as shown in Equation (1) [3]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 369, + 547, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 369, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 319, + 369, + 547, + 383 + ], + "type": "interline_equation", + "content": "L o s s = \\alpha * L o s s ^ {M A \\& B A} + (1 - \\alpha) * L o s s ^ {A d a p t a t i o n} \\tag {1}", + "image_path": "1481568671f8b5ba81db1f645caabea47681d9b51e8e1ab64dade0ee4f7d6bc6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 391, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 547, + 484 + ], + "type": "text", + "content": "Siamese Neural Networks Siamese networks [7, 24] are few-shot classifiers that compare embedding vectors from twin networks with shared parameters. Using a distance metric like Euclidean distance [13], they assess similarity between inputs. The Triplet loss [20], shown in Equation (2), is used to minimize distances between the feature vectors of same-class pairs and maximize them for different-class pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 500, + 547, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 547, + 525 + ], + "type": "interline_equation", + "content": "L (a, p, n) = \\max \\left(\\left\\| f (a) - f (p) \\right\\| ^ {2} - \\left\\| f (a) - f (n) \\right\\| ^ {2} + \\gamma , 0\\right) \\tag {2}", + "image_path": "a321a241b0423b19cb5580dcb410d5a51b5d8402bdab562f1f548275dde92c0a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 533, + 354, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 354, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 354, + 545 + ], + "type": "text", + "content": "3. Sibai" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 553, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 547, + 704 + ], + "type": "text", + "content": "In this section, we present Siamiese Backdoor Inspector (Sibai), a novel defense mechanism against poisoning attacks in FL. Sibai addresses non-i.i.d. scenarios and adaptive attackers by dynamically learning discriminative features of malicious local models using a meta-classifier. This approach avoids the reliance on predefined metrics. By directly operating on flattened model representations, Sibai does not require time-consuming preprocessing and enables detailed analysis that hinders adversarial adaptation. Sibai quickly adapts to federation behavior and new attack strategies with minimal training samples. The neural network architecture adds a layer of opacity, complicating the task for attackers to even discern the specific criteria they must adapt to." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3788" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 80, + 147, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 80, + 147, + 92 + ], + "spans": [ + { + "bbox": [ + 62, + 80, + 147, + 92 + ], + "type": "text", + "content": "3.1. Threat Model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 99, + 297, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 99, + 297, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 99, + 297, + 237 + ], + "type": "text", + "content": "We consider an FL scenario with a centralized server using FedAVG [36] and a fixed global learning rate. Consistent with prior defenses [1, 5, 6, 27, 35, 37] we assume adversaries control fewer than half of the clients, however, they can fully manipulate their data, training process, and model parameters. Adversaries are aware of the algorithms and defenses deployed on the server. Importantly, Sibai's design, training, and evaluation make no assumptions about local dataset distributions. We evaluate Sibai across i.i.d., 1-class non-i.i.d., and inter-client non-i.i.d. scenarios, the latter closely reflecting real-world FL conditions as described in Sec. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 247, + 116, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 116, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 116, + 259 + ], + "type": "text", + "content": "3.2. Design" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 265, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 265, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 265, + 296, + 380 + ], + "type": "text", + "content": "Sibai operates on the central server to detect and filter suspicious local models before aggregation. It employs a Siamese network [16, 24] meta-classifier trained to identify dissimilarities between malicious and benign client updates. Sibai consists of two phases: (I) Offline training, where the Siamese network learns from shadow models, including both benign and malicious examples; (II) Online inference, where Sibai evaluates local models against the global model during federation to identify and exclude poisoned models based on learned criteria." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 384, + 297, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 384, + 297, + 467 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 297, + 467 + ], + "type": "text", + "content": "Architecture. As shown in Fig. 1, Sibai takes the global model and a local model as input, processing them through four components: Dimensionality Reduction, Twin Networks, Distance Function, and Decision Network. The output is a binary decision indicating the likelihood of a local model being malicious. Detailed descriptions of each component follow in subsequent sections." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 483, + 249, + 676 + ], + "blocks": [ + { + "bbox": [ + 80, + 483, + 249, + 676 + ], + "lines": [ + { + "bbox": [ + 80, + 483, + 249, + 676 + ], + "spans": [ + { + "bbox": [ + 80, + 483, + 249, + 676 + ], + "type": "image", + "image_path": "91ff721698c8f93b3b5ec1f36c6256140e16cda3a97b533f7f90f17c4367aa44.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 690, + 230, + 701 + ], + "lines": [ + { + "bbox": [ + 128, + 690, + 230, + 701 + ], + "spans": [ + { + "bbox": [ + 128, + 690, + 230, + 701 + ], + "type": "text", + "content": "Figure 1. Sibai Architecture" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 81, + 549, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 549, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 549, + 244 + ], + "type": "text", + "content": "Dimensionality Reduction. Sibai processes models of varying sizes by condensing input data while retaining key features. Local models, often containing hundreds of thousands of parameters, are broken into slices, and Principal Component Analysis (PCA) extracts the most significant components for each slice. PCA identifies and ranks features by variance, ensuring compact yet informative representations. Additional metrics, specifically Cosine and Euclidean distances to the global model, that were shown to be effective by prior research [5, 17, 27, 43], are also incorporated into the input vector. Experiments showed that these metrics consistently improve detection accuracy by up to " + }, + { + "bbox": [ + 313, + 81, + 549, + 244 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 81, + 549, + 244 + ], + "type": "text", + "content": ". More details about the implementation of the dimensionality reduction pipeline are provided in Appendix C.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 248, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 547, + 376 + ], + "type": "text", + "content": "Twin Networks. The Twin Networks consist of two identical 1D Convolutional Neural Networks (CNNs), serving as the core of the Siamese network. Each CNN processes one of the two pre-processed data samples, acting as a feature extractor. They map the samples to fixed-length embedding vectors using the same filters. These two vectors can then be compared to assess similarity. A high dissimilarity between embedding vectors indicates potential maliciousness, as the twins are explicitly trained to focus on the distinguishing features of malicious local models. Appendix C.4 provides further details about the architecture of the twin models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "text", + "content": "Distance Function. The embedding vectors generated by the Twin Networks are compared using the " + }, + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "text", + "content": "-distance, which demonstrated superior performance compared to other metrics like " + }, + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 381, + 547, + 450 + ], + "type": "text", + "content": "-distance or cosine distance. This distance is used in the triplet loss function, which is utilized to train the twin networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 548, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 548, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 548, + 548 + ], + "type": "text", + "content": "Decision Network. To avoid using a fixed distance threshold, a fully connected neural network, called a decision network, is used to classify the output of the distance function. The last layer features a sigmoid activation function [38] that produces a value between 0 and 1, with values closer to 1 indicating a higher probability of maliciousness. We classify local models that when compared to the global model produce an output " + }, + { + "bbox": [ + 313, + 456, + 548, + 548 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 313, + 456, + 548, + 548 + ], + "type": "text", + "content": " as malicious." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 553, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 548, + 705 + ], + "type": "text", + "content": "**Shadow Model Dataset.** Sibai's performance depends on a shadow model dataset used during offline training. This dataset should cover a range of attacks and scenarios to ensure effective generalization by the Siamese Network. To create the shadow model set we simulate FL across various federations with different non-i.i.d. degrees attacks and PDRs. Incorporating stealthy models from adaptive adversaries with low PDR allows Sibai to focus on nuanced criteria while including poisoned models with high PDR enhances generalization. It's crucial to maintain consistent local model architectures in the shadow set, and federations with varied input or output layer sizes cannot be mixed. Further details are provided in Appendix C.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3789" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 80, + 132, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 80, + 132, + 92 + ], + "spans": [ + { + "bbox": [ + 62, + 80, + 132, + 92 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 100, + 297, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 100, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 100, + 297, + 147 + ], + "type": "text", + "content": "This section evaluates Sibai during online inference against five poisoning attacks in challenging scenarios, including complex non-i.i.d. settings, low PDRs, and adaptive adversaries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 151, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 151, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 151, + 296, + 198 + ], + "type": "text", + "content": "Hardware. Training and evaluation were performed on a Linux server with an Intel Xeon Gold 6342 CPU, 251GB RAM, CUDA [44], and an NVIDIA A30 GPU with 24GB memory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 201, + 297, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 297, + 376 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 297, + 376 + ], + "type": "text", + "content": "Reference Defenses. We compare Sibai with eight defenses: HDBSCAN [35] which represents naive clustering via cosine distance, Krum and M-Krum [5] which select the new global model based on Euclidean distance, MESAS [27] which tries to counter adaptive adversary by using six metrics, FLTrust [8] which uses trusted reference-based weighting, FoolsGold [17] which analyzes the final layer cosine distance, and the robust aggregation method Trimmed Mean/Median [60]. Hyperparameters are detailed in Appendix E. However, it is important to note that a direct comparison of methods with fundamentally different defense paradigms, such as a DF-based method like ours and IR-based defenses like FLTrust, cannot capture the whole picture since these paradigms approach poisoning defense with fundamentally different philosophies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 380, + 297, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 380, + 297, + 438 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 297, + 438 + ], + "type": "text", + "content": "Attacks. Sibai is tested against five attacks: Random label flipping [27], Pixel trigger backdoor [19], Blend backdoor [10], Edge case backdoor [54], and Semantic backdoor [3]. These attacks are introduced in more detail in Appendix D" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 441, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 488 + ], + "type": "text", + "content": "Datasets and Models. We evaluate federations collaboratively training Squeezezenet [23] on CIFAR-10 [30]. Additionally, we conducted experiments with a custom DNN and MNIST, which are provided in Sec. 4.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 491, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 491, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 61, + 491, + 297, + 632 + ], + "type": "text", + "content": "**Shadow Model Dataset.** Sibai's offline training uses 520 simulated federations, each with 20 clients (11 benign, 9 malicious), totaling 5720 benign and 4145 malicious models. Scenarios include i.i.d. and non-i.i.d. distributions with varying degrees (e.g., 0.3, 0.6, 0.9 1-class non-i.i.d. and inter-client non-i.i.d. [27]). Realistic and exaggerated settings enhance the distinction between benign and malicious models. Each federation experiences one of the five attacks. Federations with local backdoor accuracies below " + }, + { + "bbox": [ + 61, + 491, + 297, + 632 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 61, + 491, + 297, + 632 + ], + "type": "text", + "content": " were excluded to ensure poisoned models exhibit detectable traits. Distribution details are in Tab. 1 and a detailed breakdown is listed in Appendix H." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 635, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 297, + 706 + ], + "type": "text", + "content": "Default Federation Parameters. We evaluate Sibai in simulated federations collaboratively training Squeezenet [23] with SGD (learning rate: 0.01, momentum: 0.9, decay: 0.005) and a batch size of 64. Each client trains locally for 10 epochs on i.i.d. datasets with 2560 samples. The global model is initialized with pre-trained weights from" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 79, + 516, + 171 + ], + "blocks": [ + { + "bbox": [ + 347, + 79, + 516, + 171 + ], + "lines": [ + { + "bbox": [ + 347, + 79, + 516, + 171 + ], + "spans": [ + { + "bbox": [ + 347, + 79, + 516, + 171 + ], + "type": "table", + "html": "
CategoryNumber of Federations
IID/non-i.i.d.i.i.d.225
0.3 1-class non-i.i.d.66
0.6 1-class non-i.i.d.71
0.9 1-class non-i.i.d.96
inter-client non-i.i.d. [27]62
AttacksRandom [27]90
Pixel [19]91
Blend [10]119
Edge [54]121
Semantic [3]99
Total Number of Federations520
", + "image_path": "aa1c08dad49d9a5bf23598a7f97a157c54ccbca4d45a965e4793b74e60d4385c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 178, + 520, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 178, + 520, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 178, + 520, + 190 + ], + "type": "text", + "content": "Table 1. Composition of the shadow model dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 210, + 547, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 547, + 244 + ], + "type": "text", + "content": "PyTorch [46] and undergoes 49 benign rounds to mitigate early backdoor persistence [3], followed by poisoning in the 50th round." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 548, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 548, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 548, + 353 + ], + "type": "text", + "content": "Sibai Parameters. During offline training, the twin and decision networks are trained separately for 25 epochs using SGD. The twin networks use a learning rate of 0.001, while the decision network uses 0.005. Model slices contain up to 20,000 parameters, with larger layers partitioned. PCA retains 200 components per slice. Siamese network training does not use batch processing due to memory constraints, while the decision network processes batches of 16 distances derived from 8 Siamese network triples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 355, + 548, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 548, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 548, + 484 + ], + "type": "text", + "content": "Metrics. We evaluate Sibai with precision, recall, and F1 score. True positives are correctly flagged malicious models, and true negatives are benign models. Precision measures the ratio of correctly identified malicious models to all flagged models, while recall assesses the proportion of flagged malicious models among all actual malicious models. The F1 score balances these metrics. Additionally, we track the global model's BA to evaluate attack mitigation and MA changes to assess benign scenario impact. The MA diff. represents the difference between the MA after Sibai application and perfect benign aggregation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 399, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 399, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 399, + 502 + ], + "type": "text", + "content": "4.1. IID Scenarios" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 508, + 548, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 548, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 548, + 687 + ], + "type": "text", + "content": "This section evaluates Sibai's online performance after training on the shadow model set. We test Sibai on 12 unseen federation scenarios1 of increasing difficulty and adversary stealthiness. In the baseline scenario (Tab. 2), federations training Squeezenet [23] on CIFAR-10 [30] in an i.i.d. setting were attacked with the five poisoning attacks. Adversaries employed a PDR of 0.5, which is a relatively overt attack. Sibai flawlessly detected all malicious clients without false positives. Scenario (1) increased stealthiness by minimizing PDRs while maintaining local BAs above " + }, + { + "bbox": [ + 313, + 508, + 548, + 687 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 508, + 548, + 687 + ], + "type": "text", + "content": ". The blend [10] and semantic [3] backdoors used a PDR of 0.04 due to their stronger effects, while the edge case backdoor [54] required a PDR of 0.16 for consistent backdoor introduction. These adjustments mirror real-world attacker strategies. Sibai achieved perfect recall and F1 scores for" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 693, + 481, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 481, + 704 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 481, + 704 + ], + "type": "text", + "content": "Each scenario is designated by a circled number" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3790" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 79, + 267, + 173 + ], + "blocks": [ + { + "bbox": [ + 91, + 79, + 267, + 173 + ], + "lines": [ + { + "bbox": [ + 91, + 79, + 267, + 173 + ], + "spans": [ + { + "bbox": [ + 91, + 79, + 267, + 173 + ], + "type": "table", + "html": "
ScenarioRandomPixelBlendEdgeSemantic
①PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.09%0.20%18.87%0.0%
①PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall0.88881.00001.00001.00000.7777
F10.94111.00001.00001.00000.8750
MA diff.-0.88%0.0%0.0%0.0%+4.12%
BA-10.43%0.0%18.87%0.0%
", + "image_path": "7c222052bd28178d3012fa30b32023427b74f05a8fd3e5f515f48a9985424ea0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 145, + 174, + 214, + 182 + ], + "lines": [ + { + "bbox": [ + 145, + 174, + 214, + 182 + ], + "spans": [ + { + "bbox": [ + 145, + 174, + 214, + 182 + ], + "type": "inline_equation", + "content": "\\boxed{0} =" + }, + { + "bbox": [ + 145, + 174, + 214, + 182 + ], + "type": "text", + "content": " Default + high PDRs" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 145, + 182, + 212, + 194 + ], + "lines": [ + { + "bbox": [ + 145, + 182, + 212, + 194 + ], + "spans": [ + { + "bbox": [ + 145, + 182, + 212, + 194 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{1}} =" + }, + { + "bbox": [ + 145, + 182, + 212, + 194 + ], + "type": "text", + "content": " Default and low PDRs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 224, + 276, + 402 + ], + "blocks": [ + { + "bbox": [ + 77, + 201, + 281, + 213 + ], + "lines": [ + { + "bbox": [ + 77, + 201, + 281, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 201, + 281, + 213 + ], + "type": "text", + "content": "Table 2. Detection performance of Sibai in i.i.d. scenarios" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 224, + 276, + 402 + ], + "lines": [ + { + "bbox": [ + 83, + 224, + 276, + 402 + ], + "spans": [ + { + "bbox": [ + 83, + 224, + 276, + 402 + ], + "type": "table", + "html": "
qScenarioRandomPixelBlendEdgeSemantic
0.6②PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-2.65%0.20%6.12%0.0%
③PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00000.88881.0000
F11.00001.00001.00000.94111.0000
MA diff.0.0%0.0%0.0%-6.71%0.0%
BA-2.65%0.21%8.67%0.0%
0.9④PDR0.50.50.50.50.5
Precision0.87501.00001.00000.90001.0000
Recall0.77770.88881.00001.00001.0000
F10.82350.94111.00000.94731.0000
MA diff.+4.80%+1.08%0.0%+2.44%0.0%
BA-0.11%0.0%0.0%0.0%
⑤PDR0.40.080.040.160.04
Precision0.81810.75001.00000.81811.0000
Recall1.00001.00001.00001.00000.8888
F10.90000.85711.00000.90000.9411
MA diff.+0.48%-4.26%0.0%-1.65%+0.48%
BA-0.0%0.21%0.0%0.0%
", + "image_path": "6d153e02133e41716479ea73f65d9636deb0b53a3f3968d196bfd7ba9bc287b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "lines": [ + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "spans": [ + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "inline_equation", + "content": "(2) =" + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "inline_equation", + "content": "+0.61" + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "text", + "content": " -class non-i.i.d. " + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 118, + 403, + 241, + 411 + ], + "type": "text", + "content": " high PDR" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "lines": [ + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "spans": [ + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "inline_equation", + "content": "\\widehat{(3)} =" + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "inline_equation", + "content": "+0.61" + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "text", + "content": " -class non-i.i.d. " + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 118, + 411, + 240, + 419 + ], + "type": "text", + "content": " low PDR" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "lines": [ + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "spans": [ + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "inline_equation", + "content": "(4) =" + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "text", + "content": " 1-class non-i.i.d. " + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 118, + 419, + 241, + 429 + ], + "type": "text", + "content": " high PDR" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 118, + 429, + 239, + 440 + ], + "lines": [ + { + "bbox": [ + 118, + 429, + 239, + 440 + ], + "spans": [ + { + "bbox": [ + 118, + 429, + 239, + 440 + ], + "type": "inline_equation", + "content": "(5) =" + }, + { + "bbox": [ + 118, + 429, + 239, + 440 + ], + "type": "text", + "content": " Default + 0.91-class non-i.i.d. + low PDR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "text", + "content": "pixel trigger and edge case backdoors, identifying all malicious clients. In the random label flip and semantic backdoor attacks, one and two malicious clients, respectively, evaded detection. However, Sibai identified over " + }, + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "text", + "content": " of malicious clients in all cases, preventing effective poisoning. Across all i.i.d. scenarios, Sibai maintained BA below " + }, + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 61, + 491, + 297, + 643 + ], + "type": "text", + "content": ", ensuring attacks were unsuccessful. Interestingly, in the semantic backdoor scenario, Sibai mistakenly aggregated two malicious models, resulting in a slight MA improvement over perfect aggregation. This occurred because the aggregated models provided additional knowledge, but the low PDR and limited bypassed defenses still prevented backdoor establishment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 651, + 170, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 651, + 170, + 663 + ], + "spans": [ + { + "bbox": [ + 62, + 651, + 170, + 663 + ], + "type": "text", + "content": "4.2. Non-IID Scenarios" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 669, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 669, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 669, + 297, + 705 + ], + "type": "text", + "content": "In this section, we evaluate Sibai in challenging non-i.i.d. scenarios, including varying degrees of 1-class non-i.i.d. and inter-client non-i.i.d. setups as described in Sect. 2 and [27]." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 343, + 79, + 519, + 173 + ], + "blocks": [ + { + "bbox": [ + 62, + 449, + 297, + 471 + ], + "lines": [ + { + "bbox": [ + 62, + 449, + 297, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 297, + 471 + ], + "type": "text", + "content": "Table 3. Detection performance of Sibai in 1-class non-i.i.d. scenarios, with different level of disproportionality " + }, + { + "bbox": [ + 62, + 449, + 297, + 471 + ], + "type": "inline_equation", + "content": "q" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 79, + 519, + 173 + ], + "lines": [ + { + "bbox": [ + 343, + 79, + 519, + 173 + ], + "spans": [ + { + "bbox": [ + 343, + 79, + 519, + 173 + ], + "type": "table", + "html": "
ScenarioRandomPixelBlendEdgeSemantic
6PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00000.9000
Recall0.77770.88881.00001.00001.0000
F10.87500.94111.00001.00000.9473
MA diff.+2.61%+0.86%0.0%0.0%0.0%
BA-4.03%0.0%0.0%0.0%
7PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.66661.00000.8888
F11.00000.94110.80001.00000.9411
MA diff.0.0%+1.38%-1.43%0.0%-1.48%
BA-13.57%0.0%0.0%0.0%
", + "image_path": "ccdaa6f881e7556ae921d17fa79c60f81fb17c8d4d7b27961f87afc07df1e868.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "lines": [ + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "spans": [ + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "inline_equation", + "content": "\\widehat{(6)} =" + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "text", + "content": " inter-client non-i.i.d. " + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 369, + 174, + 493, + 182 + ], + "type": "text", + "content": " high PDR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 370, + 182, + 492, + 193 + ], + "lines": [ + { + "bbox": [ + 370, + 182, + 492, + 193 + ], + "spans": [ + { + "bbox": [ + 370, + 182, + 492, + 193 + ], + "type": "inline_equation", + "content": "\\widetilde{7} =" + }, + { + "bbox": [ + 370, + 182, + 492, + 193 + ], + "type": "text", + "content": " Default + inter-client non-i.i.d. + low PDR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 201, + 548, + 223 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 548, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 548, + 223 + ], + "type": "text", + "content": "Table 4. Detection performance of Sibai in inter-client non-i.i.d. scenarios [27]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": "0.6 1-class non-i.i.d. For the scenarios " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\boxed{2}" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\boxed{3} q = 0.6" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": " was used. Sibai achieved perfect detection in the 0.5 PDR case " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\boxed{2}" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": ". For the low PDRs scenario " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "\\boxed{3}" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": ", F1 scores were 1.0 for all attacks except the edge case backdoor, where one malicious client evaded detection. Despite this, the MA dropped by less than " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": ", and BA stayed below " + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 243, + 547, + 328 + ], + "type": "text", + "content": ", ensuring no effective backdoor was introduced." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 332, + 548, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 548, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 548, + 415 + ], + "type": "text", + "content": "0.9 1-class non-i.i.d. In the more extreme 0.9 1-class non-i.i.d. scenarios (4) and (5) in Tab. 10, Sibai occasionally flagged benign models, with three false positives in the worst-case pixel trigger backdoor scenario (5). However, these errors had minimal impact on MA. Sibai consistently eliminated enough malicious models to prevent effective backdoors, as no attack achieved a BA higher than " + }, + { + "bbox": [ + 313, + 332, + 548, + 415 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 332, + 548, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 418, + 548, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 548, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 548, + 608 + ], + "type": "text", + "content": "Inter-client non-i.i.d. Inter-client non-i.i.d. features erratic and highly unbalanced data distributions, reflecting real-world scenarios. In the 0.5 PDR scenario (6), Sibai achieved perfect detection for blend and edge case backdoors, perfect precision for untargeted and pixel attacks, and perfect recall for the semantic attack. Only one false positive occurred during the semantic attack. The lowest F1 score of 0.8750 occurred during the untargeted attack) In the attacker-adapted low PDR scenario (7), Sibai maintained perfect precision for all attacks despite high non-i.i.d.-ness. Recall was lower, yielding F1 scores of 0.9411 for pixel and semantic backdoors and 0.8000 for the blend attack. Nonetheless, MA was minimally affected, and BA remained low across all attacks. Overall, Sibai demonstrated robust performance in 1-class and inter-client non-i.i.d. scenarios, with minimal false positives and no effective backdoors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 617, + 434, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 434, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 434, + 629 + ], + "type": "text", + "content": "4.3. Adaptive Adversaries" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 634, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 548, + 705 + ], + "type": "text", + "content": "Sibai incorporates slice-wise Euclidean and cosine distances to the global model into its input vectors to enhance detection. While effective, these metrics may be vulnerable to adaptive attackers. To test Sibai's robustness, we evaluated four scenarios involving attackers using Bagdasaryan et al.'s [3] adaptation method with an " + }, + { + "bbox": [ + 313, + 634, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 634, + 548, + 705 + ], + "type": "text", + "content": " parameter balancing adaptation" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 313, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 313, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 313, + 743 + ], + "type": "text", + "content": "3791" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 79, + 295, + 256 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 295, + 256 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 295, + 256 + ], + "type": "table", + "html": "
Form of AdaptationScenarioRandomPixelBlendEdgeSemantic
Cosine dist.⑧PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18.87%0.0%
⑨PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.00001.0000
F11.00001,00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.14%18.87%0.0%
Euclidian dist.⑩PDR0.50.50.50.50.5
Precision1.00001.00001.00001.00001.0000
Recall1.00001.00001.00001.0000
F11.00001.00001.00001.00001.0000
MA diff.0.0%0.0%0.0%0.0%0.0%
BA-1.08%0.20%18. 87%0.0%
⑪PDR0.40.080.040.160.04
Precision1.00001.00001.00001.00001.0000
Recall1.00000.88880.88881.00001.0000
F11.00000.94110.94111.00001.0000
MA diff.0.0%+3.1%-6.3%0.0%0.0%
BA-5.98%0.0%18.87%0.0%
", + "image_path": "4d56aadb72bfaa074f10b5bab66d7f9d8e6bbb10b19c050cf8c4d117ffd762b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "lines": [ + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "spans": [ + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "inline_equation", + "content": "\\boxed{8} =" + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "text", + "content": " cosine distance adaption with " + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "inline_equation", + "content": "0.3\\alpha" + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "text", + "content": " by mal. clients " + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 257, + 272, + 266 + ], + "type": "text", + "content": " high PDR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 75, + 266, + 271, + 275 + ], + "lines": [ + { + "bbox": [ + 75, + 266, + 271, + 275 + ], + "spans": [ + { + "bbox": [ + 75, + 266, + 271, + 275 + ], + "type": "text", + "content": "9 = Default + cosine distance adaption with 0.3 α by mal. clients + low PDR" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "lines": [ + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "spans": [ + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "inline_equation", + "content": "\\boxed{10} =" + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "text", + "content": " Euclidean distance adaption with " + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "inline_equation", + "content": "0.3\\alpha" + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "text", + "content": " by mal. clients " + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 276, + 283, + 286 + ], + "type": "text", + "content": " high PDR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "lines": [ + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "spans": [ + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "inline_equation", + "content": "\\widehat{(11)} =" + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "text", + "content": " Default " + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "text", + "content": " Euclidean distance adaption with " + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "inline_equation", + "content": "0.3\\alpha" + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "text", + "content": " by mal. clients " + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 75, + 287, + 282, + 298 + ], + "type": "text", + "content": " low PDR" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "text", + "content": "and task losses. Results for " + }, + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\alpha = 0.3" + }, + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "text", + "content": ", shown in Tab. 5, highlight Sibai's resilience. In cosine-adapted scenarios, Sibai achieved perfect F1 scores (1.0), showing immunity to such attacks. Similarly, Sibai maintained perfect performance in Euclidean-adapted Scenario 12. In low-PDR Euclidean-adapted scenarios, Sibai showed flawless precision, with one false negative for pixel and blend backdoors. However, no attack established an effective backdoor, with BA remaining below " + }, + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 61, + 351, + 296, + 507 + ], + "type": "text", + "content": ". Overall, Sibai demonstrates strong resilience against adaptive attacks, proving it does not overly rely on engineered features. Its ability to internalize specific patterns makes it impractical for attackers to adapt using Bagdasaryan et al.'s method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 517, + 235, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 517, + 235, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 517, + 235, + 529 + ], + "type": "text", + "content": "4.4. Comparison to Existing Defenses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 535, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 535, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 535, + 297, + 705 + ], + "type": "text", + "content": "To assess the effectiveness of Sibai compared to metric-based poisoning detection methods in FL, we also evaluate existing defenses in some of our previous scenarios. Due to space limitations, we focus on a selection of the most representative scenarios, including scenario 1 as the i.i.d. baseline, scenario 4 as the highest degree of 1-class non-i.i.d., and scenario 6 to showcase inter-client non-i.i.d. We also include scenario 0 as the overall simplest scenario and scenario 11 to cover adaptive adversaries. Since some reference defenses rely on modified aggregation methods instead of strictly flagging and filtering out malicious models, calculating F1 scores for all reference defenses faced challenges. Further details are provided in Appendix F." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 79, + 514, + 334 + ], + "blocks": [ + { + "bbox": [ + 62, + 308, + 295, + 328 + ], + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 328 + ], + "type": "text", + "content": "Table 5. Detection performance of Sibai against adaptive attackers in i.i.d. scenarios" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 79, + 514, + 334 + ], + "lines": [ + { + "bbox": [ + 317, + 79, + 514, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 79, + 514, + 334 + ], + "type": "table", + "html": "
ScenarioDefenseRandomPixelBlendEdgeSemantic
0PDR0.50.50.50.50.5
MESAS0.90000.90000.90001.00000.9473
Krum0.57140.57140.57890.57140.5714
M-Krum0.26080.26080.64280.26080.2608
HDBSCAN1.00000.00000.11111.00001.0000
FLTrust (0.09)0.00000.26080.33330.00000.2608
Sibai (ours)1.00001.00001.00001.00001.0000
1PDR0.40.080.040.160.04
MESAS0.94730.00000.18180.16660.3076
Krum0.57140.64280.64280.57140.5714
M-Krum0.26080.60860.60860.26080.4347
HDBSCAN0.00000.55550.44440.00000.1176
FLTrust (0.09)0.09520.66660.57140.51850.5714
Sibai (ours)0.94111.00001.00001.00000.8750
4PDR0.50.50.50.50.5
MESAS0.00000.16660.80000.94730.9000
Krum0.57140.57140.57140.64280.6428
M-Krum0.26080.69560.69560.78260.7826
HDBSCAN0.00000.40000.77770.82350.8235
FLTrust (0.09)0.57140.88880.88880.88880.8571
Sibai (ours)0.82350.94111.00000.94731.0000
6PDR0.50.50.50.50.5
MESAS0.64280.47610.21050.66660.3157
Krum0.57140.57140.57140.57140.5714
M-Krum0.26080.26080.26080.52170.3478
HDBSCAN0.00000.00000.14280.30760.1428
FLTrust (0.09)0.00000.57140.57140.57140.6206
Sibai (ours)0.87500.94111.00001.00000.9473
11PDR0.40.080.040.160.04
MESAS1.00000.66660.90000.61530.9473
Krum0.57140.57140.64280.64280.6428
M-Krum0.26080.26080.78260.60860.6956
HDBSCAN0.00001.00001.00001.00000.2000
FLTrust (0.09)0.57140.62060.66660.62060.6206
Sibai (ours)1.00000.94110.94111.00001.0000
", + "image_path": "ec105ded48e714f308d9479917108288dc229f4bb675aedf979e026607c15543.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 548, + 364 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 548, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 548, + 364 + ], + "type": "text", + "content": "Table 6. F1 score comparison between Sibai and current FL poisoning defenses" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": "As shown in Tab. 6, Sibai achieved superior F1 scores across most scenarios. In high PDR i.i.d. scenario " + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "inline_equation", + "content": "\\boxed{0}" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": ", it matched or outperformed all defenses. In low PDR i.i.d. scenario " + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "inline_equation", + "content": "\\boxed{1}" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": ", Sibai achieved perfect F1 scores for pixel, blend, and edge case backdoors, outperforming MESAS for these attacks. Even when MESAS maintained low BA, it highlighted the fragility of backdoor attacks in low PDR settings. For 1-class non-i.i.d. scenario " + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "inline_equation", + "content": "\\boxed{4}" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": ", Sibai significantly outperformed other methods, and in inter-client non-i.i.d. scenario " + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "inline_equation", + "content": "\\boxed{7}" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": ", its worst-case F1 score of 0.8750 surpassed the best scores of other defenses. In adaptive scenario " + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "inline_equation", + "content": "\\boxed{11}" + }, + { + "bbox": [ + 313, + 387, + 548, + 585 + ], + "type": "text", + "content": ", Sibai slightly trailed HDBSCAN for pixel and blend backdoors however, it excelled in the detection of the semantic and untargeted attacks, achieving perfect F1 scores. Overall, Sibai consistently outperformed existing methods in identifying malicious models across diverse scenarios and attacks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 548, + 705 + ], + "type": "text", + "content": "Table 7, also considered methods for which the F1 was unavailable, specifically FoolsGold [17] and trimmed mean/median [60]. Sibai showed performance close to perfect benign aggregation, effectively preventing poisoning without significant MA reduction. Notably, Sibai and MESAS achieved low BA in challenging scenarios (4) and (6), where other defenses failed to stop backdoors. Although FLTrust [8], produced higher MA in non-i.i.d. scenarios by using a weighted aggregation that particularly emphasizes the old global model, it also often allowed higher BA and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3792" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 79, + 459, + 326 + ], + "blocks": [ + { + "bbox": [ + 152, + 79, + 459, + 326 + ], + "lines": [ + { + "bbox": [ + 152, + 79, + 459, + 326 + ], + "spans": [ + { + "bbox": [ + 152, + 79, + 459, + 326 + ], + "type": "table", + "html": "
ScenarioDefenseRandom MA diff.Pixel MA diff.BABlend MA diff.BAEdge MA diff.BASemantic MA diff.BA
①PDR0.40.080.040.160.04
benign agg.0.00.010.40.00.00.018.80.00.0
MESAS [27]-0.1+5.19.50.020.4+0.27.10.00.0
Krum [5]-35.0-43.20.0-14.10.0-35.50.0-21.1100.0
M-Krum [5]-24.5-8.512.5-9.52.9-16.30.0-11.40.0
FoolsGold [17]-6.3+3.011.0-1.718.8-8.36.1+0.10.0
HDBSCAN [35]-16.0-1.910.9-6.511.4-13.80.0-5.80.0
Trimmed mean [60]-8.3+3.510.2-2.512.8-9.06.6-0.60.0
Trimmed median [60]-22.3-2.65.2-11.60.1-14.72.5-6.40.0
FLTrust [8]-22.3-0.117.2-5.213.0+1.516.8-4.80.0
Sibai (ours)-0.80.010.40.00.00.018.80.00.0
④PDR0.40.50.50.50.5
benign agg.0.00.00.00.00.00.00.00.00.0
MESAS [27]-5.7+0.334.3+5.00.6+2.40.0+10.60.0
Krum [5]-7.7-6.1100.0-0.2100.0-4.80.0-5.30.0
M-Krum [5]-2.1+5.51.7+1.0100.0-5.10.0+4.00.0
FoolsGold [17]-2.0-5.030.2-0.2100.0-10.80.0-7.30.0
HDBSCAN [35]-7.5+1.013.2+1.40.0-8.70.0-0.920.0
Trimmed mean [60]-5.3-0.428.9-0.2100.0-11.10.0-7.840.0
Trimmed median [60]-6.6-5.10.0-0.4100.0-10.40.0-6.50.0
FLTrust [8]+19.5+14.753.1+10.898.2+4.80.0+11.120.0
Sibai (ours)+4.8+1.00.00.00.0+2.40.00.00.0
⑥PDR0.50.50.50.50.5
benign agg.0.013.92.90.02.30.02.40.00.0
MESAS [27]+11.9-0.31.6-0.132.7+9.70.00.00.0
Krum [5]+4.8-3.9100.0-5.1100.0+16.885.7+18.780.0
M-Krum [5]-1.6-3.085.3-0.3100.0+2.90.00.00.0
FoolsGold [17]-1.6+1.57.0-5.0100.0-2.00.00.00.0
HDBSCAN [35]-5.8-0.719.0+3.899.9-0.30.0-0.40.0
Trimmed mean [60]-0.5+1.17.3-3.8.1100.0-3.00.00.00.0
Trimmed median [60]-2.5-3.50.0-4.9100.0-2.30.0-0.30.0
FLTrust [8]+14.7+14.347.4+13.099.8+14.00.0+8.320.0
Sibai (ours)+2.6+0.84.00.00.00.00.00.00.0
", + "image_path": "3e546f7c4adcca3f0f93a0056a8fd94c666db7c028bac4390b559939e9c0f692.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 376, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 424 + ], + "type": "text", + "content": "therefore failed to prevent effective backdoors. In contrast, Sibai excelled at malicious model detection and consistently maintained low BA while achieving MAs aligned with benign aggregation, making Sibai an overall reliable defense." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 433, + 279, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 279, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 279, + 445 + ], + "type": "text", + "content": "4.5. Alternative Dataset and DNN Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 451, + 297, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 451, + 297, + 590 + ], + "spans": [ + { + "bbox": [ + 61, + 451, + 297, + 590 + ], + "type": "text", + "content": "To demonstrate that Sibai is not limited to SqueezeNET or the CIFAR-10 dataset, we conducted additional experiments using a custom CNN and the MNIST dataset [12]. The CNN model consists of 2 convolutional layers, 2 max-pooling layers, and 3 linear layers. We excluded the dataset-specific semantic and edge case backdoors and instead included a label swap attack [27], which targets samples of one label and swaps them to another to reduce the global model's accuracy. The shadow-model dataset for these experiments comprised 1401 federations. The F1 scores in Tab. 8 indicate that Sibai effectively detected malicious models, and therefore that the approach is not specific to SqueezeNET or CIFAR-10." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 603, + 244, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 244, + 616 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 244, + 616 + ], + "type": "text", + "content": "5. Multiple Poisonings in one Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 623, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 623, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 623, + 297, + 705 + ], + "type": "text", + "content": "To evaluate Sibai's effectiveness in detecting models poisoned with multiple backdoors simultaneously, we conducted experiments using federations where each malicious local model contained both a blend backdoor [10] and a pixel trigger backdoor [19] simultaneously. For those local models, the PDR quantifies the ratio of the total poisoned samples, i.e. the sum used for both backdoors, to the com" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 319, + 374, + 542, + 456 + ], + "blocks": [ + { + "bbox": [ + 61, + 335, + 548, + 357 + ], + "lines": [ + { + "bbox": [ + 61, + 335, + 548, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 548, + 357 + ], + "type": "text", + "content": "Table 7. Comparison of MA diff. and BA (both in percentage) between Sibai and existing FL poisoning defenses. The best value for each attack in each scenario is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 374, + 542, + 456 + ], + "lines": [ + { + "bbox": [ + 319, + 374, + 542, + 456 + ], + "spans": [ + { + "bbox": [ + 319, + 374, + 542, + 456 + ], + "type": "table", + "html": "
PDRDistributionRandom [27]Pixel [19]Blend [10]Label Swap [27]
0.4i.i.d.1.00001.00001.00001.0000
0.40.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.40.6 l-class non-i.i.d.0.69230.72000.72000.7200
0.40.9 l-class non-i.i.d.0.87500.94111.00001.0000
0.4inter-client non-i.i.d.0.90000.94110.94730.9000
0.5i.i.d.1.00001.00001.00001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.00001.0000
0.50.6 l-class non-i.i.d.0.72000.66660.72000.7200
0.50.9 l-class non-i.i.d.1.00000.94111.00001.0000
0.5inter-client non-i.i.d.0.94730.94110.94730.9000
", + "image_path": "04e20f685edff62e5865dcb50d33b225fc367c6df622c5354d9fb5bf95068070.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 353, + 500, + 509, + 582 + ], + "blocks": [ + { + "bbox": [ + 313, + 464, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 547, + 485 + ], + "type": "text", + "content": "Table 8. F1 score of Sibai in federations training a small CNN with the MNIST dataset" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 500, + 509, + 582 + ], + "lines": [ + { + "bbox": [ + 353, + 500, + 509, + 582 + ], + "spans": [ + { + "bbox": [ + 353, + 500, + 509, + 582 + ], + "type": "table", + "html": "
PDRDistributionPrecisionRecallF1
0.2i.i.d.1.00001.0001.0000
0.20.3 l-class non-i.i.d.1.00001.00001.0000
0.20.6 l-class non-i.i.d.1.00001.00001.0000
0.20.9 l-class non-i.i.d.1.00000.88880.9411
0.2inter-client non-i.i.d.1.00000.88880.9411
0.5i.i.d.1.00001.0001.0000
0.50.3 l-class non-i.i.d.1.00001.00001.0000
0.50.6 l-class non-i.i.d.1.00001.00001.0000
0.50.9 l-class non-i.i.d.1.00001.00001.0000
0.5inter-client non-i.i.d.1.00001.00001.0000
", + "image_path": "d00c14f182a6e1c8da69f9b630ba02a00401c2fd8eeef081cbfb99cca2bd3734.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 590, + 548, + 622 + ], + "lines": [ + { + "bbox": [ + 313, + 590, + 548, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 548, + 622 + ], + "type": "text", + "content": "Table 9. Detection performance of Sibai against local models poisoned with two backdoors (pixel trigger backdoor [19] and blend backdoor [10]) simultaneously in different scenarios" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 647, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 548, + 704 + ], + "type": "text", + "content": "plete local model dataset. In our implementation of this multi-backdoor half of the poisoned samples were used to introduce the pixel trigger backdoor, while the other half was used to apply the blend backdoor. Importantly, our shadow model dataset, used to train Sibai, did not include" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3793" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 117 + ], + "type": "text", + "content": "any models with such multi-backdoors. As illustrated in Tab. 9, Sibai consistently identified local models containing multiple backdoors across all tested scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 126, + 148, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 126, + 148, + 138 + ], + "spans": [ + { + "bbox": [ + 62, + 126, + 148, + 138 + ], + "type": "text", + "content": "6. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 145, + 297, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 297, + 194 + ], + "type": "text", + "content": "This section provides a brief overview of relevant literature. First, Sect. 6.1 outlines poisoning defenses tailored for FL. Then, Sect. 6.2 discusses machine learning methods for general offline poisoning analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 296, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 296, + 224 + ], + "type": "text", + "content": "6.1. Defenses against Poisoning Attacks in Federated Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 297, + 706 + ], + "type": "text", + "content": "There exists a diverse range of techniques that employ fixed metrics to detect poisoning attacks in FL. FoolsGold [17] uses a client-specific update history which is compared to the update histories of other clients using the cosine similarity. Krum [5] selects a new global model based on the Euclidean distance to other local models. M-Krum [5] extends this method to select multiple models simultaneously. Trimmed-Mean/Median [60] are robust aggregation methods that replace FedAVG [36] and calculate the updated global model by using coordinate-wise means or medians from local models. FLAME [43] clusters models using HDBSCAN [35] and filters out smaller clusters. It also applies differential privacy techniques, weight clipping, and noisng. DeepSight [47] combines filtering with differential privacy. It relies on two metrics: The cosine distance between models and values extracted from the output layer of the model. Auror [48] uses K-Means clustering on model parameters, however, unfortunately, it exhibited an extreme performance overhead in our experiments. FLTrust [8] uses a weighted aggregation system for which it assigns weights to local models based on their cosine and Euclidean distances to a benign reference model, which is trained server-side on a small trusted dataset. FLDetector [63] is a historical update-based defense that mathematically predicts client updates and compares the predictions to the actual updates. MESAS [27] was specifically designed to operate in non-i.i.d. scenario and counter adaptive adversaries and employs six metrics. These approaches differ from Sibai as they rely on fixed metrics or criteria, making them susceptible to adaptive attacks [27]. Defenses focusing on specific layers, like FoolsGold [17], are prone to adversaries exploiting advanced model poisoning techniques specifically targeting those layers. Similarly, methods that heavily focus on cosine or Euclidean distance are often vulnerable to adaptation methods using additional losses [3, 27], as detailed in Sect. 2. These rigid metrics also face challenges in non-i.i.d. scenarios, especially in inter-client non-i.i.d. settings, where it has been demonstrated that most defenses, except MESAS [27] which was specifically designed to handle such scenarios, face difficulties. In contrast, Sibai dynamically learns criteria based on its shadow model dataset, making it challenging for attackers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 81, + 548, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 548, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 548, + 118 + ], + "type": "text", + "content": "to adapt without access to the Siamese Network. Sibai can also differentiate between variations in local models due to data distribution and those caused by poisoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 127, + 537, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 537, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 537, + 140 + ], + "type": "text", + "content": "6.2. Non-FL Specific Offline Poisoning Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 145, + 550, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 550, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 550, + 507 + ], + "type": "text", + "content": "This section discusses several machine-learning-based detection methods against poisoning attacks, not specifically designed for FL. Xu et al. [58] propose Meta Neural Trojan Detection (MNTD), a black-box backdoor detection approach utilizing a meta-classifier that assesses if a model's output behavior on a given input aligns more closely with a set of known benign models or a set of known malicious models. Huang et al. [21] introduce the one-pixel signature concept for DNNs, consisting of a set of single-pixel adversarial perturbations that have the highest probability of triggering misclassifications on a specific benign image dataset. Subsequently, a meta-classifier was trained on these signatures to discern between poisoned and benign models. Kolouri et al. [25] use universal litmus patterns and a meta-classifier to detect backdoors across different network architectures. Chen et al. [9] present DeepInspect, a black-box method utilizing model inversion and anomaly detection for backdoor detection. Huang et al. [22] also introduced Neuroninspect, employing explanation heatmaps and outlier detection for identifying abnormal neurons. Unlike these methods, Sibai is tailored to FL, avoiding intensive preprocessing steps. DeepInspect [9] involves extensive reverse engineering of potential triggers for every label of each local model, and Huang et al. [21] requires testing every local model with a large set of single-pixel adversarial perturbations. Conducting such intricate analyses for every model in a federation during each round is impractical, as these methods were originally designed for detailed offline analysis of individual models. Sibai instead only uses PCA as the sole dimensionality reduction mechanism, facilitating straightforward inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 385, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 385, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 385, + 533 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 541, + 550, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 550, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 550, + 704 + ], + "type": "text", + "content": "To counter non-i.i.d. data distribution and adaptive adversaries in FL, we introduce Siamese Backdoor Inspector (Sibai), a novel FL poisoning defense mechanism utilizing a Siamese network meta-classifier. Inspired by successful offline backdoor detection [21, 25, 58], Sibai dynamically learns discriminative features directly from PCA decompositions of local models. In diverse evaluations, Sibai consistently achieves high detection rates without significantly impacting global model performance. It effectively prevents backdoor establishment and outperforms current defenses in terms of f1 score while maintaining low false positive rates. Future work may explore selectively applying influence reduction to models identified by Sibai for targeted defense." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 744 + ], + "type": "text", + "content": "3794" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 80, + 165, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 80, + 165, + 94 + ], + "spans": [ + { + "bbox": [ + 62, + 80, + 165, + 94 + ], + "type": "text", + "content": "8. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 100, + 297, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 100, + 297, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 100, + 297, + 159 + ], + "type": "text", + "content": "This research has been funded by the Federal Ministry of Education and Research of Germany (BMBF) within the program Digital. Sicher. Souverän. \"in the project\"Erkennung von Angriffen gegen IoT-Netzwerke in Smart Homes - IoTGuard\" (project number 16KIS1919)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 173, + 120, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 173, + 120, + 185 + ], + "spans": [ + { + "bbox": [ + 63, + 173, + 120, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 193, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 69, + 193, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 193, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 69, + 193, + 296, + 246 + ], + "type": "text", + "content": "[1] Sebastien Andreina, Giorgia Azzurra Marson, Helen Möllering, and Ghassan Karame. Baffle: Backdoor detection via feedback-based federated learning. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS), pages 852-863. IEEE, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 248, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 248, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 296, + 280 + ], + "type": "text", + "content": "[2] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. In 30th USENIX Security Symposium (USENIX Security 21), pages 1505-1521, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 282, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 282, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 69, + 282, + 296, + 335 + ], + "type": "text", + "content": "[3] Eugene Bagdasaryan, Andreas Veit, Yiqing Hua, Deborah Estrin, and Vitaly Shmatikov. How to backdoor federated learning. In International conference on artificial intelligence and statistics, pages 2938-2948. PMLR, 2020. 1, 2, 4, 5, 8, 13, 14, 15, 16, 17, 18, 19" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 337, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 337, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 69, + 337, + 296, + 380 + ], + "type": "text", + "content": "[4] Arjun Nitin Bhagoji, Supriyo Chakraborty, Prateek Mittal, and Seraphin Calo. Analyzing federated learning through an adversarial lens. In International Conference on Machine Learning, pages 634-643. PMLR, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 382, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 382, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 69, + 382, + 296, + 425 + ], + "type": "text", + "content": "[5] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 1, 3, 4, 7, 8, 15, 16" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 427, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 427, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 69, + 427, + 296, + 470 + ], + "type": "text", + "content": "[6] Peva Blanchard, El Mahdi El Mhamdi, Rachid Guerraoui, and Julien Stainer. Machine learning with adversaries: Byzantine tolerant gradient descent. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 472, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 472, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 296, + 514 + ], + "type": "text", + "content": "[7] Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard Säckinger, and Roopak Shah. Signature verification using a\" siamese\" time delay neural network. Advances in neural information processing systems, 6, 1993. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 517, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 517, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 69, + 517, + 296, + 559 + ], + "type": "text", + "content": "[8] Xiaoyu Cao, Minghong Fang, Jia Liu, and Neil Zhenqiang Gong. Fltrust: Byzantine-robust federated learning via trust bootstrapping. Network and Distributed Systems Security (NDSS) Symposium 2021, 2021. 1, 4, 6, 7, 8, 15, 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 561, + 296, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 561, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 69, + 561, + 296, + 603 + ], + "type": "text", + "content": "[9] Huili Chen, Cheng Fu, Jishen Zhao, and Farinaz Koushanfar. Deepinspect: A black-box trojan detection and mitigation framework for deep neural networks. In *IJCAI*, page 8, 2019. 1, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 606, + 296, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 606, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 64, + 606, + 296, + 649 + ], + "type": "text", + "content": "[10] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 2, 4, 7, 13, 14, 17, 18, 19" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 651, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 651, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 64, + 651, + 296, + 704 + ], + "type": "text", + "content": "[11] Xiaoyi Chen, Ahmed Salem, Dingfan Chen, Michael Backes, Shiqing Ma, Qingni Shen, Zhonghai Wu, and Yang Zhang. Badnl: Backdoor attacks against nlp models with semantic-preserving improvements. In Annual computer security applications conference, pages 554-569, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 82, + 548, + 704 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 82, + 548, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 82, + 548, + 113 + ], + "spans": [ + { + "bbox": [ + 316, + 82, + 548, + 113 + ], + "type": "text", + "content": "[12] Li Deng. The mnist database of handwritten digit images for machine learning research. IEEE Signal Processing Magazine, 29(6):141-142, 2012. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 115, + 547, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 115, + 547, + 136 + ], + "spans": [ + { + "bbox": [ + 316, + 115, + 547, + 136 + ], + "type": "text", + "content": "[13] Elena Deza, Michel Marie Deza, Michel Marie Deza, and Elena Deza. Encyclopedia of distances. Springer, 2009. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 138, + 547, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 138, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 316, + 138, + 547, + 169 + ], + "type": "text", + "content": "[14] Cynthia Dwork et al. The mathematics of information coding, extraction, and distribution. The IMA Volumes in Mathematics and its applications, 107:31-47, 1999. 12" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 171, + 548, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 171, + 548, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 171, + 548, + 213 + ], + "type": "text", + "content": "[15] Minghong Fang, Xiaoyu Cao, Jinyuan Jia, and Neil Gong. Local model poisoning attacks to {Byzantine-Robust} federated learning. In 29th USENIX security symposium (USENIX Security 20), pages 1605-1622, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 215, + 548, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 215, + 548, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 215, + 548, + 267 + ], + "type": "text", + "content": "[16] Hossein Fereidooni, Jan König, Phillip Rieger, Marco Chilese, Bora Gökbakan, Moritz Finke, Alexandra Dmitrienko, and Ahmad-Reza Sadeghi. Authentisense: A scalable behavioral biometrics authentication scheme using few-shot learning for mobile platforms. arXiv preprint arXiv:2302.02740, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 270, + 548, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 548, + 321 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 548, + 321 + ], + "type": "text", + "content": "[17] Clement Fung, Chris JM Yoon, and Ivan Beschastnikh. The limitations of federated learning in sybil settings. In 23rd International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2020), pages 301-316, 2020. 1, 3, 4, 6, 7, 8, 15, 16" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 323, + 548, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 323, + 548, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 548, + 376 + ], + "type": "text", + "content": "[18] Yansong Gao, Bao Gia Doan, Zhi Zhang, Siqi Ma, Jiliang Zhang, Anmin Fu, Surya Nepal, and Hyoungshick Kim. Backdoor attacks and countermeasures on deep learning: A comprehensive review. arXiv preprint arXiv:2007.10760, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 378, + 548, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 548, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 548, + 420 + ], + "type": "text", + "content": "[19] Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Identifying vulnerabilities in the machine learning model supply chain. arXiv preprint arXiv:1708.06733, 2017. 1, 2, 4, 7, 13, 17, 18, 19" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 548, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 548, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 548, + 473 + ], + "type": "text", + "content": "[20] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In Similarity-Based Pattern Recognition: Third International Workshop, SIMBAD 2015, Copenhagen, Denmark, October 12-14, 2015. Proceedings 3, pages 84-92. Springer, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 476, + 548, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 548, + 538 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 548, + 538 + ], + "type": "text", + "content": "[21] Shanjiaoyang Huang, Weiqi Peng, Zhiwei Jia, and Zhuowen Tu. One-pixel signature: Characterizing cnn models for backdoor detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVII 16, pages 326-341. Springer, 2020. 1, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 541, + 548, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 548, + 572 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 548, + 572 + ], + "type": "text", + "content": "[22] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 1, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 574, + 548, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 574, + 548, + 616 + ], + "spans": [ + { + "bbox": [ + 316, + 574, + 548, + 616 + ], + "type": "text", + "content": "[23] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and i 0.5 mb model size. arXiv preprint arXiv:1602.07360, 2016. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 618, + 548, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 618, + 548, + 650 + ], + "spans": [ + { + "bbox": [ + 316, + 618, + 548, + 650 + ], + "type": "text", + "content": "[24] Gregory Koch, Richard Zemel, Ruslan Salakhutdinov, et al. Siamese neural networks for one-shot image recognition. In ICML deep learning workshop. Lille, 2015. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 651, + 548, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 651, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 316, + 651, + 548, + 704 + ], + "type": "text", + "content": "[25] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 1, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3795" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 297, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 63, + 82, + 297, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 297, + 124 + ], + "type": "text", + "content": "[26] Jakub Konečný, H Brendan McMahan, Daniel Ramage, and Peter Richtárik. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint arXiv:1610.02527, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 125, + 297, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 125, + 297, + 198 + ], + "spans": [ + { + "bbox": [ + 63, + 125, + 297, + 198 + ], + "type": "text", + "content": "[27] Torsten Krauß and Alexandra Dmitrienko. Mesas: Poisoning defense for federated learning resilient against adaptive attackers. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, page 1526-1540, New York, NY, USA, 2023. Association for Computing Machinery. 1, 2, 3, 4, 5, 7, 8, 13, 15, 16, 17, 18, 19" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 200, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 200, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 200, + 296, + 243 + ], + "type": "text", + "content": "[28] Torsten Krauß, Raphael Götz, and Alexandra Dmitrienko. Security of nvme offloaded data in large-scale machine learning. In Computer Security - ESORICS 2023, pages 143-163, Cham, 2024. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 243, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 243, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 63, + 243, + 296, + 296 + ], + "type": "text", + "content": "[29] Torsten Krauß, Jan König, Alexandra Dmitrienko, and Christian Kanzow. Automatic Adversarial Adaption for Stealthy Poisoning Attacks in Federated Learning. Network and Distributed Systems Security (NDSS) Symposium 2024, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 297, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 297, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 63, + 297, + 295, + 318 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 4, 13, 14, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 319, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 319, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 63, + 319, + 296, + 371 + ], + "type": "text", + "content": "[31] Liping Li, Wei Xu, Tianyi Chen, Georgios B Giannakis, and Qing Ling. Rsa: Byzantine-robust stochastic aggregation methods for distributed learning from heterogeneous datasets. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1544–1551, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 372, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 372, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 63, + 372, + 295, + 414 + ], + "type": "text", + "content": "[32] Yijing Li, Xiaofeng Tao, Xuefei Zhang, Junjie Liu, and Jin Xu. Privacy-preserved federated learning for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(7):8423-8434, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 415, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 415, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 63, + 415, + 296, + 446 + ], + "type": "text", + "content": "[33] Yiming Li, Yong Jiang, Zhifeng Li, and Shu-Tao Xia. Backdoor learning: A survey. IEEE Transactions on Neural Networks and Learning Systems, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 447, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 447, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 63, + 447, + 296, + 499 + ], + "type": "text", + "content": "[34] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In 25th Annual Network And Distributed System Security Symposium (NDSS 2018). Internet Soc, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 501, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 501, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 63, + 501, + 296, + 532 + ], + "type": "text", + "content": "[35] Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017. 3, 4, 7, 8, 15, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 533, + 296, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 533, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 63, + 533, + 296, + 585 + ], + "type": "text", + "content": "[36] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 2, 3, 8, 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 586, + 296, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 586, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 63, + 586, + 296, + 628 + ], + "type": "text", + "content": "[37] Luis Muñoz-González, Kenneth T Co, and Emil C Lupu. Byzantine-robust federated machine learning through adaptive model averaging. arXiv preprint arXiv:1909.05125, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 630, + 296, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 630, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 63, + 630, + 296, + 661 + ], + "type": "text", + "content": "[38] Sridhar Narayan. The generalized sigmoid activation function: Competitive supervised learning. Information sciences, 99 (1-2):69-82, 1997. 3, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 662, + 296, + 704 + ], + "type": "text", + "content": "[39] Mohammad Naseri, Jamie Hayes, and Emiliano De Cristofaro. Local and central differential privacy for robustness and privacy in federated learning. arXiv preprint arXiv:2009.03561, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 548, + 704 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 125 + ], + "type": "text", + "content": "[40] Blaine Nelson, Marco Barreno, Fuching Jack Chi, Anthony D Joseph, Benjamin IP Rubinstein, Udam Saini, Charles Sutton, J Doug Tygar, and Kai Xia. Exploiting machine learning to subvert your spam filter. LEET, 8(1-9):16-17, 2008. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 125, + 547, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 125, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 125, + 547, + 178 + ], + "type": "text", + "content": "[41] Anh Nguyen, Tuong Do, Minh Tran, Binh X Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, and Quang D Tran. Deep federated learning for autonomous driving. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 1824-1830. IEEE, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 180, + 548, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 180, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 548, + 232 + ], + "type": "text", + "content": "[42] Thien Duc Nguyen, Phillip Rieger, Markus Miettinen, and Ahmad-Reza Sadeghi. Poisoning attacks on federated learning-based IoT intrusion detection system. In Proc. Workshop Decentralized IoT Syst. Secur.(DISS), pages 1-7, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 235, + 547, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 235, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 235, + 547, + 299 + ], + "type": "text", + "content": "[43] Thien Duc Nguyen, Phillip Rieger, Roberta De Viti, Huili Chen, Björn B Brandenburg, Hossein Yalame, Helen Möllering, Hossein Fereidooni, Samuel Marchal, Markus Miettinen, et al. {FLAME}: Taming backdoors in federated learning. In 31st USENIX Security Symposium (USENIX Security 22), pages 1415–1432, 2022. 1, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 300, + 547, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 547, + 320 + ], + "type": "text", + "content": "[44] NVIDIA, Péter Vingelmann, and Frank H.P. Fitzek. Cuda, release: 10.2.89, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 323, + 547, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 323, + 547, + 375 + ], + "spans": [ + { + "bbox": [ + 315, + 323, + 547, + 375 + ], + "type": "text", + "content": "[45] Xudong Pan, Mi Zhang, Beina Sheng, Jiaming Zhu, and Min Yang. Hidden trigger backdoor attack on {NLP} models via linguistic style manipulation. In 31st USENIX Security Symposium (USENIX Security 22), pages 3611-3628, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 377, + 548, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 377, + 548, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 377, + 548, + 472 + ], + "type": "text", + "content": "[46] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 4, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 475, + 548, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 548, + 527 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 548, + 527 + ], + "type": "text", + "content": "[47] Phillip Rieger, Thien Duc Nguyen, Markus Miettinen, and Ahmad-Reza Sadeghi. Deepsight: Mitigating backdoor attacks in federated learning through deep model inspection. NNetwork and Distributed Systems Security (NDSS) Symposium 2022, 2022. 1, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 529, + 548, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 529, + 548, + 581 + ], + "spans": [ + { + "bbox": [ + 315, + 529, + 548, + 581 + ], + "type": "text", + "content": "[48] Shiqi Shen, Shruti Tople, and Prateek Saxena. Auror: Defending against poisoning attacks in collaborative deep learning systems. In Proceedings of the 32nd Annual Conference on Computer Security Applications, pages 508-519, 2016. 1, 8, 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 583, + 548, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 583, + 548, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 583, + 548, + 624 + ], + "type": "text", + "content": "[49] Gan Sun, Yang Cong, Jiahua Dong, Qiang Wang, Lingjuan Lyu, and Ji Liu. Data poisoning attacks on federated machine learning. IEEE Internet of Things Journal, 9(13):11365-11375, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 627, + 547, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 627, + 547, + 659 + ], + "spans": [ + { + "bbox": [ + 315, + 627, + 547, + 659 + ], + "type": "text", + "content": "[50] Ziteng Sun, Peter Kairouz, Ananda Theertha Suresh, and H Brendan McMahan. Can you really backdoor federated learning? arXiv preprint arXiv:1911.07963, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 661, + 547, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 661, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 315, + 661, + 547, + 693 + ], + "type": "text", + "content": "[51] Zhiyi Tian, Lei Cui, Jie Liang, and Shui Yu. A comprehensive survey on poisoning attacks and countermeasures in machine learning. ACM Computing Surveys, 55(8):1-35, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 693, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 693, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 693, + 547, + 704 + ], + "type": "text", + "content": "[52] Alexander Turner, Dimitris Tsipras, and Aleksander" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "3796" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 296, + 602 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 82, + 82, + 296, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 82, + 296, + 102 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 296, + 102 + ], + "type": "text", + "content": "Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 104, + 296, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 104, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 64, + 104, + 296, + 125 + ], + "type": "text", + "content": "[53] Guido Van Rossum and Fred L. Drake. Python 3 Reference Manual. CreateSpace, Scotts Valley, CA, 2009. 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 127, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 127, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 127, + 296, + 190 + ], + "type": "text", + "content": "[54] Hongyi Wang, Kartik Sreenivasan, Shashank Rajput, Harit Vishwakarma, Saurabh Agarwal, Jy-yong Sohn, Kangwook Lee, and Dimitris Papailiopoulos. Attack of the tails: Yes, you really can backdoor federated learning. Advances in Neural Information Processing Systems, 33:16070-16084, 2020. 1, 2, 4, 13, 14, 17, 18, 19" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 192, + 296, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 192, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 64, + 192, + 296, + 233 + ], + "type": "text", + "content": "[55] Zhaoxian Wu, Qing Ling, Tianyi Chen, and Georgios B Giannakis. Federated variance-reduced stochastic gradient descent with robustness to byzantine attacks. IEEE Transactions on Signal Processing, 68:4583-4596, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 236, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 236, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 64, + 236, + 296, + 276 + ], + "type": "text", + "content": "[56] Chulin Xie, Keli Huang, Pin-Yu Chen, and Bo Li. Dba: Distributed backdoor attacks against federated learning. In International conference on learning representations, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 279, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 279, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 64, + 279, + 296, + 320 + ], + "type": "text", + "content": "[57] Cong Xie, Oluwasanmi Koyejo, and Indranil Gupta. Fall of empires: Breaking byzantine-tolerant sgd by inner product manipulation. In Uncertainty in Artificial Intelligence, pages 261-270. PMLR, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 323, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 323, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 64, + 323, + 296, + 365 + ], + "type": "text", + "content": "[58] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. In 2021 IEEE Symposium on Security and Privacy (SP), pages 103-120. IEEE, 2021. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 366, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 366, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 64, + 366, + 296, + 407 + ], + "type": "text", + "content": "[59] Qiang Yang, Yang Liu, Tianjian Chen, and Yongxin Tong. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST), 10(2):1-19, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 410, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 410, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 64, + 410, + 296, + 452 + ], + "type": "text", + "content": "[60] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 4, 6, 7, 8, 15, 16" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 453, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 453, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 64, + 453, + 296, + 495 + ], + "type": "text", + "content": "[61] Dong Yin, Yudong Chen, Ramchandran Kannan, and Peter Bartlett. Byzantine-robust distributed learning: Towards optimal statistical rates. In International Conference on Machine Learning, pages 5650–5659. PMLR, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 497, + 296, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 497, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 64, + 497, + 296, + 539 + ], + "type": "text", + "content": "[62] Hongyi Zhang, Jan Bosch, and Helena Holmström Olsson. End-to-end federated learning for autonomous driving vehicles. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 540, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 540, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 64, + 540, + 296, + 602 + ], + "type": "text", + "content": "[63] Zaixi Zhang, Xiaoyu Cao, Jinyuan Jia, and Neil Zhenqiang Gong. Fl detector: Defending federated learning against model poisoning attacks via detecting malicious clients. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2545-2555, 2022. 1, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "text", + "content": "3797" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_content_list.json b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f277fdb363d876b4df040839d6f3a5cd036d7787 --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_content_list.json @@ -0,0 +1,1671 @@ +[ + { + "type": "text", + "text": "SignRep: Enhancing Self-Supervised Sign Representations", + "text_level": 1, + "bbox": [ + 202, + 130, + 794, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ryan Wong $^{1}$ , Necati Cihan Camgoz $^{2}$ , Richard Bowden $^{1}$ $^{1}$ University of Surrey, $^{2}$ Meta Reality Labs \n{r.wong, r.bowden}@surrey.ac.uk, neccam@meta.com", + "bbox": [ + 212, + 179, + 777, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language representation learning presents unique challenges due to the complex spatio-temporal nature of signs and the scarcity of labeled datasets. Existing methods often rely either on models pre-trained on general visual tasks, that lack sign-specific features, or use complex multimodal and multi-branch architectures. To bridge this gap, we introduce a scalable, self-supervised framework for sign representation learning. We leverage important inductive (sign) priors during the training of our RGB model. To do this, we leverage simple but important cues based on skeletons while pretraining a masked autoencoder. These sign specific priors alongside feature regularization and an adversarial style agnostic loss provide a powerful backbone. Notably, our model does not require skeletal keypoints during inference, avoiding the limitations of keypoint-based models during downstream tasks. When finetuned, we achieve state-of-the-art performance for sign recognition on the WLASL, ASL-Citizen and NMFs-CSL datasets, using a simpler architecture and with only a single-modality. Beyond recognition, our frozen model excels in sign dictionary retrieval and sign translation, surpassing standard MAE pretraining and skeletal-based representations in retrieval. It also reduces computational costs for training existing sign translation models while maintaining strong performance on Phoenix2014T, CSL-Daily and How2Sign.", + "bbox": [ + 89, + 301, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 708, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language is an important means of communication for millions of people worldwide. Sign languages have complex visual characteristics, which include intricate hand shapes, motions, body poses and facial expressions that models need to accurately interpret and process [5]. Furthermore, the computational demands of processing long video sequences add considerable challenges, making it difficult to scale these systems effectively. As a result, current approaches to sign recognition and translation often rely on general pretrained vision models [1, 7, 19, 22, 28, 49, 57]. Creating robust, label-free sign language representations", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that generalize across diverse datasets is challenging, yet essential for scalable sign language modeling.", + "bbox": [ + 511, + 268, + 903, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing methods for sign recognition rely on multimodal or an ensemble of specialized models to achieve state-of-the-art recognition results. These models often require multi-channel inputs (e.g. RGB, depth and skeleton data) or specialized architectures to capture the complex interactions of hand, body or facial expressions [17, 20, 22, 23, 52, 57]. With each country having its own sign language and linguistic study expensive, currently available labeled datasets are sparse with typically under 2000 unique signs [1, 19, 26, 28, 41]. Collecting annotated sign language data is costly and time-consuming, making it infeasible to rely solely on labeled datasets for sign representation learning. This highlights the need for methods that can learn sign language representations from large-scale, unlabeled data.", + "bbox": [ + 511, + 303, + 905, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Newer methods for pretraining sign language translation models often require sentence-aligned annotations [25, 37, 49, 53, 55]. While these prior approaches jointly learn individual sign representations and inter-sign relationships, they are not scalable to unlabeled datasets as they require sentence-aligned annotations. We therefore focus on building a scalable and generalizable individual sign representation framework using self-supervised learning. A strong foundation in individual sign embeddings is an important step before building effective inter-sign models, as it ensures robust feature representations before incorporating complex long range temporal information.", + "bbox": [ + 511, + 518, + 905, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our main contributions are as follows: (1) We propose a scalable self-supervised Masked Autoencoding (MAE) framework which leverages sign priors, adversarial loss and feature regularizations for sign representation learning. (2) We introduce a method for analyzing sign class similarities in unseen datasets, introducing an auxiliary class probability distribution loss, which enhances recognition performance. (3) Our single pretrained model achieves state-of-the-art sign recognition, surpassing complex architecture and multimodal models. (4) We demonstrate the effectiveness of our sign representations for sign dictionary retrieval, achieving strong performance without any downstream training. (5) We demonstrate that using our pre", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22804", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trained model as a frozen feature extractor, training is more tractable by reducing memory requirements and improving the performance of existing sign translation models.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 150, + 232, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Sign Recognition. Isolated Sign Recognition involves identifying a single sign within a given video. Sign recognition models often rely on pretrained spatiotemporal models from action recognition datasets such as Kinetics [8], fine-tuned for specific sign recognition tasks [1, 19, 22, 28, 57]. This transfer learning approach faces a domain shift problem, as sign videos have unique temporal dynamics and subtle gestures not well-represented in general action datasets, leading to performance degradation. To leverage invariance, some methods use skeletal keypoints instead of full-frame RGB models [10, 17, 20, 48, 52]. While keypoint-based models are more memory-efficient, they typically underperform relative to RGB models [2, 11, 36] and require complex architectural modifications to Graph Convolutional Networks (GCNs) [30] or transformers [46], to model the spatial-temporal relationships. Moreover, keypoints are prone to errors, such as missing or misdetected points [31, 32]. To solve these issues, recent state-of-the-art methods have combined keypoint and RGB modalities through branching or ensembling techniques [22, 23, 57]. While they achieve higher accuracy, they increase computational complexity due to the need for multiple models or support for multi-modal inputs.", + "bbox": [ + 89, + 176, + 483, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Sign Representation Learning. Self-supervised learning for vision [3, 33, 38, 44] and language [12, 35] has demonstrated substantial benefits by leveraging unlabeled datasets for various tasks. In sign language, approaches such as Skeleton [24], SignBERT [17], SignBERT+ [20] and BEST [52] focus on keypoint-based self-supervised learning, employing masked learning similar to BERT [12]. However, these methods depend on keypoints as inputs and therefore requiring ensembling with RGB models to achieve competitive performance, resulting in increasing complexity and limited pretraining benefits. This highlights the need for a simple and efficient model which capturing sign-specific information and retains the pretraining advantages.", + "bbox": [ + 89, + 537, + 483, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Paired Sign-Text Pretraining for Sign Translation. Supervised pretraining approaches for sign translation, such as GFSLT-VLP [53], Sign2GPT [49], SignHiera [37], MSLU [55] and VAP [25], use sign video-spoken language pairs to enhance sign translation. Sign2GPT employs a pretrained DinoV2 [33], which is effective for general vision tasks but requires LoRA [16] fine-tuning for sign translation to achieve strong performance, significantly increasing computational costs. This highlights the domain gap", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between pretrained foundation vision models and sign language. SignHiera [37] demonstrates success with large-scale training on YT-ASL [45] using Hiera MAE [38] with language-supervised pretraining. However, they acknowledge that SignHiera requires significant computational resources (64 A100 GPUs for two weeks), emphasizing the need for more accessible and cost-effective video pretraining on large-scale datasets. These methods also face scalability challenges due to their reliance on sentence-aligned annotations. In this paper, we introduce a scalable self-supervised pretraining framework that learns sign representations without paired sign-text annotations, aiming to significantly reduce computational costs for pretraining and improve efficiency in downstream sign tasks.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Sign Priors", + "text_level": 1, + "bbox": [ + 511, + 316, + 633, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign language communication relies on hand shapes, body posture and interactions. To guide the model toward meaningful sign representations, we introduce \"sign priors\", a set of cues that we know capture essential sign features. These priors $(\\mathcal{P})$ serve as the primary targets for our model, as detailed in Sec. 5.1. Using a human pose estimation model specialized for sign language videos [21], we extract joint angles and 3D keypoints. We categorize our priors into keypoint, angle, distance and signer activity.", + "bbox": [ + 511, + 340, + 908, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Keypoint Priors define the spatial structure of signing, we divide this into two priors:", + "bbox": [ + 511, + 492, + 906, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hand Keypoint Prior $(\\mathcal{P}^{\\{h,k\\}})$ : This prior is designed to capture hand shapes and orientations. We normalize the 3D coordinates of the 21 hand keypoints by setting the wrist as the origin, thereby eliminating positional variations in the body space. The resulting vector, $\\mathcal{P}^{\\{h,k\\}} \\in \\mathbb{R}^{21 \\times 3}$ , represents the configuration of the left or right hand, where $h \\in \\{\\mathrm{LH}, \\mathrm{RH}\\}$ denotes the left or right hands.", + "bbox": [ + 511, + 522, + 905, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Full Body Keypoint Prior $(\\mathcal{P}^{\\{b,k\\}})$ . This prior captures the positioning of the body and hands within the body space. We use all 61 3D keypoints (21 for each hand and 19 for the body), resulting in $\\mathcal{P}^{\\{b,k\\}} \\in \\mathbb{R}^{61 \\times 3}$ .", + "bbox": [ + 511, + 628, + 905, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Joint Angle Priors capture finger flexion, extension and encode upper-body posture variations:", + "bbox": [ + 511, + 704, + 905, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hand Joint Angle Prior $(\\mathcal{P}^{\\{h,a\\}})$ . This prior captures hand joint orientations. We extract 41 hand joint angles, yielding a vector $\\mathbb{R}^{41\\times 1}$ . To handle the continuous nature of angles, we apply sine and cosine transformations, resulting in the final angle prior $\\mathcal{P}^{\\{h,a\\}}\\in \\mathbb{R}^{41\\times 2}$ .", + "bbox": [ + 511, + 734, + 905, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Body Angle Pior $(\\mathcal{P}^{\\{b,a\\}})$ . To capture the body's joint orientations, we extract the 22 body joint angles. These angles are transformed using sine and cosine functions to handle their continuous nature, resulting in the body angle prior $\\mathcal{P}^{\\{b,a\\}} \\in \\mathbb{R}^{22 \\times 2}$ .", + "bbox": [ + 511, + 809, + 905, + 885 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22805", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7a3373601e0de8e37928fe062b76cf893af3837e9d4ae6654b993722199aa6b2.jpg", + "image_caption": [ + "Figure 1. (Left): The pretraining process for SignRep, which leverages masked representation learning to predict sign priors such as hand keypoints and joint angles. This is achieved through a Hiera encoder and a lightweight sign decoder. The representation is further refined with regularization losses, including variance, covariance and adversarial style loss. (Right): An example setup for the discriminator to obtain a representation pair to predict a style-representation match." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distance Priors capture fine-grained differences between signs and the interactions of hands in the signing space:", + "bbox": [ + 88, + 367, + 482, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fingertip Distance Prior $(\\mathcal{P}^{\\{h,d\\}})$ . Since fingertip interactions are a key component in sign language, we compute a distance matrix by measuring the distances from the fingertip keypoints to each of the knuckle, wrist and other fingertip keypoints. This produces the prior $\\mathcal{P}^{\\{h,d\\}} \\in \\mathbb{R}^{5 \\times 11 \\times 3}$ .", + "bbox": [ + 88, + 398, + 482, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand-Interaction Distance Prior $(\\mathcal{P}^{\\{b,d\\}})$ . Similar to the fingertip distance prior, we calculate a distance matrix using the wrist and five fingertip keypoints from each hand, resulting in a $6\\times 2$ set of keypoints. The distance between the hands, body and facial keypoints is computed to capture the interactions between the hands and other body parts, resulting in the final hand-interaction prior $\\mathcal{P}^{\\{b,d\\}}\\in \\mathbb{R}^{12\\times 22\\times 3}$", + "bbox": [ + 88, + 476, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Signer Activity. Some signs require only one hand or the signer may be in a resting pose. To enable the model to capture this information, we develop a simple heuristic. We define the active prior $\\mathcal{P}\\{h,\\mathrm{act}\\}$ , where a hand is considered inactive if it remains below the middle of the stomach and has not moved over the course of a video clip. This results in a prior $\\mathcal{P}\\{h,\\mathrm{act}\\} \\in [0,1]$ , indicating whether each hand is active (1) or inactive (0).", + "bbox": [ + 88, + 613, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. SignRep Architecture", + "text_level": 1, + "bbox": [ + 89, + 753, + 297, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn the sign priors, we introduce a novel adaptation of the Hierarchical Vision Transformer (Hiera) [38], designed with a pretraining task specifically tailored for sign representation learning without requiring labeled sign data. We choose the Hiera model as it aligns with our objective of being efficient, simple and effective. It has demonstrated strong performance in masked representation learning while eliminating the need for complex architectures and special", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ized modules. We aim to develop a single spatio-temporal sign model that enhances performance without relying on multi-branch or ensemble methods, instead improving the representation learning process.", + "bbox": [ + 511, + 367, + 903, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The standard Hiera MAE framework processes a video clip represented as $V = \\{v_{1}, v_{2}, \\dots, v_{t}, \\dots, v_{T}\\}$ , where $v_{t}$ denotes the frame at time step $t$ , and $T$ is the total number of frames. The video is first divided into spatiotemporal patches. A masking strategy is then applied, with a subset of patches randomly masked according to a masking ratio $M$ . The unmasked patches are fed into the encoder $f_{\\mathrm{enc}}$ , which processes them to generate an output representation $\\mathbf{z}^{\\mathrm{emb}} = \\{z_0^{\\mathrm{emb}}, z_1^{\\mathrm{emb}}, \\dots, z_K^{\\mathrm{emb}}\\}$ , where $K$ is the number of output tokens from the encoder. The representation $\\mathbf{z}^{\\mathrm{emb}}$ is then passed to a decoder, along with learnable masked tokens as input. The pretraining objective is to reconstruct the masked patches through pixel-level prediction, allowing the model to learn contextual relationships and enhance the visual representation.", + "bbox": [ + 511, + 429, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rather than pixel reconstruction, we use sign priors to learn meaningful sign representations, replacing the pixel reconstruction decoder with a lightweight sign decoder.", + "bbox": [ + 511, + 655, + 905, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sign Decoder $(f_{dec})$ . As shown in Fig. 1, instead of passing $\\mathbf{z}^{\\mathrm{emb}}$ to the decoder, we take the average across $\\mathbf{z}^{\\mathrm{emb}}$ tokens, which is then processed through layer normalization followed by a fully connected layer. This produces an output representation $z^{\\mathrm{avg}} \\in \\mathbb{R}^{1 \\times D}$ , where $D$ is the dimensionality of the representation. The representation $z^{\\mathrm{avg}}$ is then upsampled temporally to match the sequence length $T$ of the input video. The upsampling module is implemented using a lightweight network consisting of a 1D convolution with a kernel size of 1, followed by a GELU activation function, and then a transpose convolution with a kernel size of $T$ to match the number of input frames and", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22806", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an output dimension of $D^{\\prime}$ . This results in an upsampled vector $\\mathbf{z}^{\\mathrm{up}}\\in \\mathbb{R}^{T\\times D^{\\prime}}$ , where $\\mathbf{z}^{up} = [z_1^{up},z_2^{up},\\dots ,z_T^{up}]$", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can then incorporate prediction heads for each sign prior to the network as follows:", + "bbox": [ + 89, + 122, + 482, + 152 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} ^ {\\mathcal {P}} = g _ {\\mathcal {P}} \\left(z _ {t} ^ {\\mathrm {u p}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 162, + 482, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each sign prior, $\\mathcal{P} \\in \\{\\mathcal{P}\\{h,k\\}, \\mathcal{P}\\{h,a\\}, \\mathcal{P}\\{h,d\\}, \\mathcal{P}\\{b,k\\}, \\mathcal{P}\\{b,a\\}, \\mathcal{P}\\{b,d\\}\\}$ has a corresponding fully connected layer $(g_{\\mathcal{P}})$ with an associated output $\\hat{y}_t^{\\mathcal{P}}$ which matches the flattened dimension of the corresponding sign prior. For the activity prior we simply use an MLP layer using $z^{\\mathrm{avg}}$ as input which produces an output of 2 to match the dimension of $[\\mathcal{P}\\{\\mathrm{LH,act}\\}, \\mathcal{P}\\{\\mathrm{RH,act}\\}]$ .", + "bbox": [ + 89, + 191, + 483, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder is completely removed during downstream tasks and only the encoder is used as the sign representation model. As a result, the target sign priors are not required during downstream tasks, allowing the use of a single model that inherently captures sign knowledge. This approach ensures computational efficiency by leveraging masked representation learning while directly learning the essential sign priors.", + "bbox": [ + 89, + 299, + 483, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Sign Representation Objectives", + "text_level": 1, + "bbox": [ + 89, + 434, + 377, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Sign Priors Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 459, + 334, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the previous section, we identified the sign priors that our model needs to learn. Since the sign decoder applies upsampling to match the number of input frames, each frame's prior has an associated output prediction. For each predicted sign prior output, $\\hat{y}_t^{\\mathcal{P}}$ , we train the model to regress to the corresponding target value $y_t^{\\mathcal{P}}$ from prior $\\mathcal{P}$ at frame $t$ , using smooth L1 loss ( $L1$ ) across all sign priors for each frame in the input video sequence. The same loss is applied to the sign activity prior for consistency. We also mitigate the impact of low-quality keypoints in our objective function by only using keypoints with over $50\\%$ confidence from the pose estimation model and masking missing keypoints from the loss. The final reconstruction loss is as follows:", + "bbox": [ + 89, + 482, + 483, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {\\mathcal {P} \\in \\mathbf {P}} w _ {\\mathcal {P}} L 1 (\\hat {y} ^ {\\mathcal {P}}, y ^ {\\mathcal {P}}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 688, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{\\mathcal{P}}$ is the loss weighting for the prior $\\mathcal{P}$ .", + "bbox": [ + 89, + 731, + 403, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Representation Regularization", + "text_level": 1, + "bbox": [ + 89, + 756, + 362, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Regularization. Self-supervised methods have shown that regularizing features improves representation quality [4, 51]. To enhance our representations, we incorporate a variance and covariance loss into $z^{\\mathrm{avg}}$ .", + "bbox": [ + 89, + 779, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The variance loss, $\\mathcal{L}_{\\mathrm{var}}$ , encourages diversity in the learned representations by spreading them across the representation space. To compute the variance loss, we first calculate the standard deviation for each feature dimension", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(j)$ across a batch $(N)$ .", + "bbox": [ + 513, + 90, + 666, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {j} = \\sqrt {\\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left(z _ {i , j} ^ {\\mathrm {a v g}} - \\bar {z} _ {j} ^ {\\mathrm {a v g}}\\right) ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 116, + 903, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diversity of the features are then encouraged through a hinge function to ensure that the variance does not fall below a threshold of one.", + "bbox": [ + 511, + 175, + 906, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v a r}} = \\sum_ {j = 1} ^ {D} \\max (0, 1 - \\sigma_ {j}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 229, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The covariance loss, $\\mathcal{L}_{\\mathrm{cov}}$ , reduces correlations among features, helping to avoid redundancy. To compute the covariance loss, we first need to calculate the covariance matrix such that:", + "bbox": [ + 511, + 282, + 906, + 342 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {j, k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(z _ {i, j} ^ {\\text {a v g}} - \\bar {z} _ {j} ^ {\\text {a v g}}\\right) \\left(z _ {i, k} ^ {\\text {a v g}} - \\bar {z} _ {k} ^ {\\text {a v g}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 351, + 903, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_{j,k}$ represents the covariance between feature dimensions $j$ and $k$ and $\\bar{z}_j^{\\mathrm{avg}}$ is the mean of the $j$ -th feature across the batch. The loss then penalises off-diagonal values in the covariance matrix, encouraging different feature dimensions to be uncorrelated:", + "bbox": [ + 511, + 402, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o v}} = \\sum_ {j \\neq k} \\mathcal {C} _ {j, k} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 488, + 903, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Style Agnostic Representations. We also aim to encourage the encoder to capture robust and generalizable sign features, while filtering out irrelevant details such as background and person-specific appearance. We can explicitly learn this by introducing a discriminator that evaluates whether two representations share the same \"style\", defined here by background and appearance features.", + "bbox": [ + 511, + 537, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract style information, we calculate the gram matrix, commonly used in image style transfer [15], from the $\\mathbf{z}^{\\mathrm{emb}}$ tokens, averaging across the column dimension to produce the style representation $z^{\\mathrm{style}}\\in \\mathbb{R}^{D}$ . We then pair $z^{\\mathrm{avg}}$ with $z^{\\mathrm{style}}$ and pass them to the discriminator, which learns to output 1 for matching pairs and 0 for non-matching pairs.", + "bbox": [ + 511, + 643, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate these representations pairs, as shown in Fig. 1(right), we randomly crop a video sequence of length $L$ to create two segments, $A_{1}$ and $A_{2}$ , each of length $T$ . For each, we extract $z_{A_1}^{\\mathrm{avg}}$ and $z_{A_1}^{\\mathrm{style}}$ , as well as $z_{A_2}^{\\mathrm{avg}}$ and $z_{A_2}^{\\mathrm{style}}$ , assuming that they share background and appearance features due to originating from the same video. From a different video, we obtain segment $B$ with its style $z_B^{\\mathrm{style}}$ , which provides a contrasting background and appearance.", + "bbox": [ + 511, + 734, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The discriminator, $\\mathfrak{D}$ , is then trained to produce 0 output for mismatched styles and 1 for matched styles, such that $\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_B^{\\mathrm{style}}) = 0$ and $\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_{A_2}^{\\mathrm{style}}) = 1$ . This can be used", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22807", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to further enhance the generalization of the encoder, we add an adversarial loss during pretraining, designed to fool the discriminator which encourages the model to focus on sign-specific content over style-related features. The adversarial loss is formalized as follows:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p o s} ^ {\\left(A _ {1}, A _ {2}\\right)} = \\max \\left(0, \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {A _ {2}} ^ {\\text {s t y l e}}\\right) - \\mathbb {E} _ {q \\sim \\mathrm {U}} \\mathfrak {D} (q)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 175, + 482, + 194 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n e g}} ^ {(A _ {1}, B)} = \\max (0, \\mathbb {E} _ {q \\sim \\mathrm {M}} \\mathfrak {D} (q) - \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {B} ^ {\\text {s t y l e}}\\right)) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 196, + 480, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} ^ {A _ {1}} = \\left(\\mathcal {L} _ {p o s} ^ {(A _ {1}, A _ {2})}\\right) ^ {2} + \\left(\\mathcal {L} _ {n e g} ^ {(A _ {1}, B)}\\right) ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 219, + 480, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q)$ is the expected discriminator output of style for representation pairs that come from the same video sequence and $\\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)$ for when they come from different video sequence. We apply this loss only if $\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q) > \\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)$ , ensuring stability by restricting the adversarial loss to cases where the discriminator is likely to correctly identify the matches.", + "bbox": [ + 89, + 246, + 483, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our final loss for pretraining is defined as:", + "bbox": [ + 109, + 353, + 392, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {r e c o n}} + w _ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}} + w _ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + w _ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 377, + 480, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{\\mathrm{adv}}$ , $w_{\\mathrm{var}}$ and $w_{\\mathrm{cov}}$ are the respective weighting factors for the adversarial, variance and covariance losses.", + "bbox": [ + 89, + 404, + 483, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Representations for Dictionary Retrieval", + "text_level": 1, + "bbox": [ + 89, + 446, + 455, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sign dictionary retrieval is the task of identifying signs from a predefined database. This allows for quick lookup of related signs given a query sign which supports efficient dataset creation and promotes cross-linguistic understanding of signs [39]. The SignRep model can be used as a feature extractor for dictionary retrieval. Given a query sign video of length $L$ , we apply a sliding window to extract segment representations $\\mathbf{Z}^{\\mathrm{avg}} = \\{z_1^{\\mathrm{avg}}, z_2^{\\mathrm{avg}}, \\ldots, z_n^{\\mathrm{avg}}, \\ldots, z_N^{\\mathrm{avg}}\\}$ , where $N = \\left\\lfloor \\frac{L - T}{\\text{stride}} \\right\\rfloor + 1$ . Since isolated signs typically begin and end in a resting pose, we use the output from", + "bbox": [ + 89, + 470, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {n} ^ {\\mathrm {a c t}} = \\max \\left(\\mathcal {P} ^ {\\{L H, \\mathrm {a c t} \\}}, \\mathcal {P} ^ {\\{R H, \\mathrm {a c t} \\}}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 646, + 480, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to identify whether the signer is active. By checking hand activity, we compute a weighted average of the segment representations resulting in", + "bbox": [ + 89, + 674, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {\\text {o u t}} = \\frac {1}{\\sum \\gamma_ {n} ^ {\\text {a c t}}} \\sum_ {n = 1} ^ {N} \\gamma_ {n} ^ {\\text {a c t}} z _ {n} ^ {\\text {a v g}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 729, + 480, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z^{\\mathrm{out}}$ is the representation for the query video. Finally, we extract features for all dictionary videos and classify a query video by finding the closest match using cosine similarity of the normalized representations.", + "bbox": [ + 89, + 779, + 483, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class Probability Distribution. The retrieval approach enables the recognition of visually similar sign classes without additional training, which is useful for improving the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accuracy of downstream tasks such as recognition. We construct a class distribution by computing the cosine similarity across all samples in the training dataset, capturing the visual proximity between them. From the resulting cosine similarity matrix, we compute inter-class similarities using the sign-sample labels to generate a matrix of shape $\\mathbb{R}^{C\\times C}$ , where $C$ represents the number of classes. Applying a temperature-scaled softmax to each row of this matrix yields a probability distribution for each class relative to the others, forming a class probability distribution matrix $\\phi \\in \\mathbb{R}^{C\\times C}$ , where each element of $\\phi$ lies in the range [0, 1].", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can incorporate the class probability distribution $\\phi$ as a regularization term in the downstream recognition task using KL divergence. Given a sign video with target label $c$ , we can compute the KL divergence between the predicted class distribution $\\bar{y}$ (from the output classifier layer) and $\\phi_c$ which is the pre-computed class probability distribution of $c$ . This results in:", + "bbox": [ + 511, + 257, + 905, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\phi} = \\kappa \\phi_ {c} [ c ] \\mathrm {K L} \\left(\\phi_ {c} \\| \\bar {y}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 376, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa$ is the weighting factor for the loss and $\\mathrm{KL}(\\phi_c\\parallel \\bar{y})$ is the KL divergence. To emphasize the contribution of the target class distribution, we scale the KL divergence by $\\phi_c[c]$ , which is the value of the class probability for $\\phi_c$ at $c$ .", + "bbox": [ + 511, + 404, + 905, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7. Experiments", + "text_level": 1, + "bbox": [ + 513, + 479, + 645, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model on sign recognition and retrieval using three datasets: ASL-Citizen [11], WLASL2000 [28], and NMFs-CSL [19]. The ASL-Citizen dataset includes 2,731 isolated ASL signs recorded from a webcam. The WLASL2000 dataset, sourced from the web, contains 2,000 common ASL signs and presents challenges due to its noisy nature and limited samples per sign. The NMFs-CSL dataset consists of 1,067 Chinese Sign Language (CSL) signs, which require recognition of non-manual cues, such as facial expressions, to accurately identify signs.", + "bbox": [ + 511, + 505, + 905, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 511, + 666, + 702, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow standard evaluation protocols for sign recognition, measuring top-1 and top-5 per-instance and per-class accuracy on WLASL and NMFs-CSL. For ASL-Citizen, we use the benchmark metrics specified for this dataset, including discounted cumulative gain (DCG), mean reciprocal rank (MRR), and top-1 and top-5 instance accuracy.", + "bbox": [ + 511, + 688, + 903, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our model as a frozen feature extractor, we apply a retrieval protocol based on the dictionary-based retrieval approach from [11]. We choose a retrieval approach over a linear evaluation protocol, as dictionary retrieval is a common and practical task in sign language processing. This method also offers a better evaluation of the generalization of the features to unseen datasets, which is particularly important given the evolving nature of sign languages.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22808", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7.2. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 90, + 292, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining. We pretrain our model on the YouTube-SL-25 dataset [42], which consists of large-scale continuous sign videos from YouTube. We initialize the model with the base video Hiera-B architecture, pretrained on Kinetics using MAE. We randomly select 16 consecutive frames as inputs based on a co-articulated sign typically lasting for around 13 frames [1, 34, 47]. No changes are made to the original encoder architecture, while the decoder is replaced with our lightweight sign decoder.", + "bbox": [ + 89, + 112, + 483, + 248 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recognition. For downstream tasks, our approach leverages only the pretrained encoder, discarding the decoder and entirely eliminating the reliance on keypoint extraction. For recognition, we extend the pretrained model's input sequence to 64 frames. We preserve the pretrained weights by inflating the patch embeddings, which preserves the computational efficiency by avoiding increasing the number of tokens in the transformer, detailed in the supplementary material. A linear classifier is applied to the pooled features from the encoder model to predict the sign classes.", + "bbox": [ + 89, + 250, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retrieval. For dictionary retrieval evaluation, we follow the approach outlined in Sec. 6 using a stride of 2. For the label assignment, we compute the cosine similarity between the test video representation and those in the training dictionary. Each test query video is assigned the label of the training video that yields the highest similarity score.", + "bbox": [ + 89, + 400, + 483, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional implementation details are provided in the supplementary material.", + "bbox": [ + 89, + 489, + 483, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7.3. Evaluation on Sign Recognition", + "text_level": 1, + "bbox": [ + 89, + 531, + 370, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, our single-modality model demonstrates significant improvements over all existing single-modality approaches, achieving performance on par with complex multi-modal methods for WLASL2000. While ensembling with multi-modal approaches could further improve performance, we emphasize the importance of pretraining. Our Hiera encoder model with our pretraining, achieves these results without the extensive architectural modifications required by other methods. Instead of relying on additional modalities as input, our pretrained model has learned these features to create a robust sign representation model. Importantly, our pretraining strategy does not rely on any annotated sign data.", + "bbox": [ + 89, + 551, + 482, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effectiveness of our SignRep model on the NMFs-CSL dataset is further illustrated in Tab. 2, where we achieve a top-1 accuracy of $84.1\\%$ and a top-5 accuracy of $98.8\\%$ , outperforming all other methods, including multimodal approaches. Notably, the closest single-modality comparison is the StepNet RGB model, which achieves $77.2\\%$ in top-1 accuracy, nearly $7\\%$ lower than our model. While StepNet focuses on developing specialized architecture for sign, we demonstrate our framework can effectively learn sign features from large-scale datasets.", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/734691222fc8cf65e1546c7f828e7c5f7ade0810bcf32ba82201fb4414517533.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInstance Acc.Class Acc.
Top-1Top-5Top-1Top-5
Skeleton-based
ST-GCN [50]34.4066.5732.5365.45
SignBERT [17]39.4073.3536.7472.38
BEST [52]46.2579.3343.5277.65
SignBERT+[20]48.8582.4846.3781.33
Multi-modal
BEST (+R) [52]54.5988.0852.1287.28
SignBERT(+R) [17]54.6987.4952.0886.93
SignBERT+(+R) [20]55.5989.3753.3388.82
SAM (5ξ) [23]58.7391.4655.9390.94
SAM-v2 (5ξ) [22]59.3991.4856.6390.89
NLA-SLR [57]61.0591.4558.0590.70
NLA-SLR(3ξ) [57]61.2691.7758.3190.91
StepNet (R+F) [40]61.1791.9458.4391.43
RGB-based
I3D [8]32.4857.31--
I3D(BSL1K) [1]46.8279.3644.7278.47
StepNet [40]56.8988.6454.5487.97
SignRep (Ours)61.0590.2758.8989.44
", + "bbox": [ + 514, + 89, + 915, + 465 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2afc547537b682fb836f4d430a312f227034fc77c8cd81792fd724786c2e5dd.jpg", + "table_caption": [ + "Table 1. Comparison of downstream sign recognition results on WLASL2000. $\\xi$ , (R) and (F) denotes a multi-crop inference, RGB and Optical Flow modality respectively." + ], + "table_footnote": [], + "table_body": "
MethodTop-1Top-5
I3D [8]64.488.0
TSM [29]64.588.7
Slowfast [14]66.386.6
GLE-Net [19]69.088.1
HMA (◇) [18]75.695.3
StepNet [40]77.292.5
SignBERT (H+R) (◇) [17]78.497.3
BEST (◇) [52]79.297.1
NLA-SLR (◇) [57]83.498.3
StepNet (R+F) (◇) [40]83.697.0
NLA-SLR (◇,3ξ) [57]83.798.5
SignRep (Ours)84.198.8
", + "bbox": [ + 553, + 536, + 867, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of downstream sign recognition results on NMFs-CSL. $\\diamond$ denotes methods using multi-modality/ensemble models and $\\xi$ indicates a multi-crop inference.", + "bbox": [ + 511, + 766, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our results on the ASL-Citizen dataset, shown in Tab. 3, further demonstrate the effectiveness of our model where it surpasses the previous baseline I3D model by $18\\%$ in top-1 accuracy.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22809", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec294680473b8b8f5fdfa5f56d1490a0c345e7e586e7dc5a5bd153c8ccd62012.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDCGMRRRec@1Rec@5
ST-GCN* [50]76.3769.9759.5282.68
I3D* [8]79.1373.3263.1086.09
SignRep (Ours)90.8488.0581.3796.11
", + "bbox": [ + 93, + 88, + 480, + 171 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of downstream sign recognition results on ASL-Citizen. * denotes results produced by [11]", + "bbox": [ + 89, + 181, + 482, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.4. Evaluation on Sign Dictionary Retrieval", + "text_level": 1, + "bbox": [ + 89, + 227, + 434, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the effectiveness of our model as a feature extractor, we utilize a dictionary-based sign retrieval method that directly evaluates the quality of the learned representations without any fine-tuning. In Tab. 4, we compare our approach to the Hiera model pretrained with MAE on both the Kinetics and YT-SL datasets. We apply the same training settings outlined in Sec. 7.2 for pretraining on YT-SL. Our results demonstrate that our proposed pretraining strategy significantly outperforms standard MAE, even when trained on YT-SL, by optimizing specifically for sign language representation, whereas MAE's pixel-reconstruction objective does not directly support learning the sign specific features needed for sign retrieval.", + "bbox": [ + 88, + 250, + 482, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare our model's retrieval performance with the use of raw 3D keypoints and joint angles averaged over time. As shown in Tab. 4, our method achieves a substantial performance increase, with top-1 retrieval scores more than tripling those based on hand joint angles. These results underscore the strength of our pretraining approach in capturing spatio-temporal features specific to sign language, establishing our method as a highly effective framework for sign representation learning. We also validate the importance of incorporating hand activity awareness, observing that weighted averaging based on hand activity improves retrieval performance compared to averaging.", + "bbox": [ + 89, + 446, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 637, + 254, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Pretraining on Sign Retrieval. In Tab. 5, we examine the effects of various elements in our pretraining strategy on retrieval performance in WLASL2000. Our results show that combining all three types of sign priors that have angle, keypoint and distance yields the highest retrieval accuracy, underscoring the benefit of capturing multiple aspects of the sign features. We also analyze the role of masking during pretraining, removing it leads to a noticeable decline in retrieval performance, indicating that masking is essential for robust representation learning. While we observe lower reconstruction loss without masking, this does not translate to better retrieval scores, suggesting that precise reconstruction alone does not necessarily produce effective sign representations. Finally, the addition of both adversarial loss and variance-covariance regularization yields the strongest retrieval results, confirming that", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "these feature regularization techniques improve the quality and robustness of learned sign representations.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Pretraining for Recognition. In Tab. 6, we demonstrate that our pretrained model substantially outperforms both the MAE model pretrained on Kinetics and the model pretrained on YT-SL with pixel MAE. Additionally, the baseline Hiera model pretrained on Kinetics, achieving a top-1 accuracy of $51.5\\%$ , surpasses prior keypoint-based masked learning methods such as SignBERT, BEST and SignBERT+, as shown in Tab. 1. This result underscores the value of RGB input, revealing the limitations of using keypoints as a primary modality. Previous keypoint methods have required ensembling with RGB to remain competitive, illustrating the challenges of relying solely on skeletal information as an input modality for sign recognition. Finally, we observe that incorporating the sign class distribution loss further enhances recognition accuracy, improving top-1 accuracy from $59.9\\%$ to $61.0\\%$ with a $\\kappa$ of 0.2.", + "bbox": [ + 511, + 138, + 906, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "8. Feature Extractor for Sign Translation", + "text_level": 1, + "bbox": [ + 511, + 392, + 862, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate SignRep as a feature extractor for sign translation by integrating it into two open-source translation models: Sign2GPT [49] on RWTH-PHOENIX-Weather 2014T (Phoenix14T) [6] and CSL-Daily [54], as well as the approach from [43] on How2Sign [13]. We chose these models because they allow direct evaluation of SignRep as a frozen feature extractor without requiring sign-text pretraining or significant modifications to the architecture. We follow the same translation training hyperparameters as the original papers to ensure a fair comparison and isolate the impact of replacing the visual backbone.", + "bbox": [ + 511, + 417, + 906, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For Phoenix14T and CSL-Daily, we omit the pseudogloss pretraining used in [49] to ensure a fair evaluation of SignRep's learned features without additional linguistic supervision. In [49], they show that the DinoV2 backbone requires fine-tuning with LoRA to perform well on sign translation. Instead, we replace the original learnable DinoV2 backbone with pre-extracted SignRep features, obtained using our sliding window approach with a stride of 2. This setup places ours at a disadvantage, as we remove the trainable backbone weights and data augmentation, relying solely on pre-extracted features. Despite this, SignRep achieves competitive translation performance, as shown in Tab. 7, performing comparably on Phoenix14T and achieving substantial gains on CSL-Daily. By replacing the learnable DinoV2 model with pre-extracted SignRep features during training, we significantly reduce computational costs, making large-scale translation training more efficient and accessible without sacrificing performance. Our approach removes an additional 5.5GFLOP per frame required by the learnable DinoV2 model, which is important as translation models process tens to hundreds of frames.", + "bbox": [ + 511, + 584, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22810", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9baa74f6fc929f202a9ae3c9bb90268c70658ffc30f0dacfb3ed5cca97a47d11.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FeaturesASL-CitizenWLASL2000NMFs-CSL
DCGRec@1Rec@5DCGRec@1Rec@5DCGRec@1Rec@5
HieraMAE-Kinetic11.640.250.8413.212.083.4023.293.9612.18
HieraMAE-YTSL12.120.391.3414.062.574.4128.037.5718.38
All Joint Angles13.820.572.1717.922.546.5032.517.9325.50
All Keypoints14.290.982.7419.373.168.2336.6412.2630.99
Hand Keypoints25.917.9619.5828.117.5720.9241.7215.9142.62
Hand Joint Angle26.938.8121.4130.619.4224.3644.1718.1346.34
SignRep (avg)61.4037.4768.7753.5726.1660.9879.5158.4891.85
SignRep (weighted)71.2149.9580.0957.9329.9267.4183.0563.0495.63
", + "bbox": [ + 133, + 88, + 867, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4b2a6ad407adcc7d73567b258bf164334c012b73d25c74260915ae204cc52803.jpg", + "table_caption": [ + "Table 4. Comparison of generalization results on sign retrieval with no downstream training applied to SignRep." + ], + "table_footnote": [], + "table_body": "
Prior Comp.Reg.DCG
anglekptdistmaskvar+covadv
✓✓45.1
✓✓32.9
✓✓47.2
✓✓✓✓48.5
✓✓✓46.3
✓✓✓✓✓49.9
✓✓✓✓✓✓50.7
", + "bbox": [ + 102, + 305, + 470, + 463 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65d7d061e00e8c8279b0dd96c09c340ecfcf1e479a62d08bcac4b75ea9e79b30.jpg", + "table_caption": [ + "Table 5. Comparison of the impact of sign priors, masking and regularization on the impact of retrieval. Retrieval results are obtained on WLASL using a stride of 8." + ], + "table_footnote": [], + "table_body": "
MethodKInstance Acc.Class Acc.
top-1top-5top-1top-5
MAE (Kinetic)-51.583.448.482.0
MAE (YT-SL)-57.287.454.686.3
SignRep0.059.990.257.489.2
SignRep0.160.490.157.889.1
SignRep0.261.090.358.989.4
SignRep0.559.889.457.788.8
", + "bbox": [ + 94, + 525, + 480, + 669 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6c7c707af80015b91dbc02f164a28e574c17746e50a4f98c2a90064575456950.jpg", + "table_caption": [ + "Table 6. Comparisons of the impact of pretraining and weighting for the class distribution loss on WLASL." + ], + "table_footnote": [], + "table_body": "
Sign2GPT BackbonePhoenix14TCSL-Daily
B-4RB-4R
DinoV2(LoRA)19.4245.2312.9641.12
SignRep(extracted)20.3845.1716.3342.67
", + "bbox": [ + 91, + 718, + 485, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This leads to an $80\\%$ reduction in GPU memory usage during training, lowering requirements from 60GB to 11GB", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3ea10f46a16ff6a18b5dc8771badca425b4229293a24d25408c864e354c35059.jpg", + "table_caption": [ + "Table 7. Comparison of translation results using Sign2GPT [49] by replacing the learnable DinoV2 backbone with our extracted SignRep features. B-4 denotes BLEU4 and R denotes ROUGE." + ], + "table_footnote": [], + "table_body": "
Tarrés's BackboneHow2Sign
rBLEUBLEU
I3D(Supervised)2.218.03
SignRep(Self-Supervised)2.748.66
", + "bbox": [ + 547, + 304, + 874, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparison of translation results using Tarrés [43], replacing the I3D features with our SignRep features.", + "bbox": [ + 511, + 398, + 903, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with a batch size of 8. This reduction makes training on standard hardware more feasible, enabling broader accessibility for researchers working with sign translation models.", + "bbox": [ + 511, + 439, + 903, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For How2Sign, in Tab. 8, we replace the supervised I3D features with our self-supervised SignRep features on the translation framework from [43] and observe performance improvements. Unlike I3D features, which are pretrained with supervision on a labeled sign recognition data, SignRep learns features purely through self-supervision, highlighting our framework's effectiveness for learning sign-specific features without supervision.", + "bbox": [ + 511, + 484, + 905, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 618, + 640, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced a scalable, self-supervised framework for sign representation learning without labeled sign datasets. Using masked autoencoding with sign priors, adversarial loss and feature regularizations, our approach enhances generalization while eliminating the need for skeletal keypoints, multimodal or multi-branch architectures in downstream tasks. Our single model achieves state-of-the-art performance on multiple sign recognition benchmarks with a simpler, more efficient design. Beyond recognition, our representations demonstrate versatility in sign dictionary retrieval on unseen datasets and serve as an efficient feature extractor for existing sign translation systems, reducing computational costs while maintaining strong performance. Our findings highlight the potential of our self-supervised learning framework for scalable sign modeling and encourage further research into practical applications of sign representations.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22811", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 294 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European conference on computer vision, pages 35–53. Springer, 2020. 1, 2, 6", + "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bencie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford british sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2", + "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 2", + "[4] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 4", + "[5] P Boyes Braem and RL Sutton-Spence. The Hands Are The Head of The Mouth. The Mouth as Articulator in Sign Languages. Hamburg: Signum Press, 2001. 1", + "[6] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 7, 2", + "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1", + "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 6, 7", + "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. Electra: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations, 2020. 1", + "[10] Konstantinos M Dafnis. Bidirectional skeleton-based isolated sign recognition using graph convolution networks. In LREC proceedings, 2022. 2" + ], + "bbox": [ + 93, + 303, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by SNSF project 'SMILE II' (CRSII5 193686), European Union's Horizon2020 programme ('EASIER' grant agreement 101016982) and the Innosuisse IICT Flagship (PFFS-21-47). This work reflects only the authors view and the Commission is not responsible for any use that may be made of the information it contains. Neither Necati Cihan Camgoz nor Meta were involved in the model training, evaluation, or use of the datasets. The authors also thank Maksym Ivashechkin for their assistance in providing the 3D pose estimation model.", + "[11] Aashaka Desai, Lauren Berger, Fyodor Minakov, Nessa Milano, Chinmay Singh, Kriston Pumphrey, Richard Ladner, Hal Daume III, Alex X Lu, Naomi Caselli, et al. Asl citizen: a community-sourced dataset for advancing isolated sign language recognition. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 7", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2", + "[13] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 7, 2", + "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6", + "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 4", + "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2", + "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11087-11096, 2021. 1, 2, 6", + "[18] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 6", + "[19] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 1, 2, 5, 6", + "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11221-11239, 2023. 1, 2, 6", + "[21] Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. Improving 3d pose estimation for sign language. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 1-5. IEEE, 2023. 2, 1, 3", + "[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 1, 2, 6", + "[23] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22812", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3413-3423, 2021. 1, 2, 6", + "[24] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3394-3402, 2021. 2", + "[25] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In European Conference on Computer Vision, pages 349-367. Springer, 2025. 1, 2", + "[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 1", + "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1", + "[28] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 1, 2, 5", + "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 6", + "[30] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 2", + "[31] Amit Moryossef. Optimizing hand region detection in mediapipe holistic full-body pose estimation to improve accuracy and avoid downstream errors. arXiv preprint arXiv:2405.03545, 2024. 2", + "[32] Amit Moryossef, Ioannis Tsochantaridis, Joe Dinn, Necati Cihan Camgoz, Richard Bowden, Tao Jiang, Annette Rios, Mathias Muller, and Sarah Ebling. Evaluating the immediate applicability of pose estimation for sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3434-3440, 2021. 2", + "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2", + "[34] Tomas Pfister, James Charles, and Andrew Zisserman. Large-scale learning of sign language by watching tv. 2013. 6", + "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Charles Raude, KR Prajwal, Liliane Momeni, Hannah Bull, Samuel Albanie, Andrew Zisserman, and Gül Varol. A tale of two languages: Large-vocabulary continuous sign language recognition from spoken language supervision. arXiv preprint arXiv:2405.10266, 2024. 2", + "[37] Phillip Rust, Bowen Shi, Skyler Wang, Necati Cihan Camgoz, and Jean Maillard. Towards privacy-aware sign language translation at scale. arXiv preprint arXiv:2402.09611, 2024. 1, 2", + "[38] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 2, 3", + "[39] Marc Schulder, Sam Bigeard, Maria Kopf, Thomas Hanke, Anna Kuder, Joanna Wojcicka, Johanna Mesch, Thomas Björkstrand, Anna Vacalopoulou, Kyriaki Vasilaki, et al. Signs and synonymity: Continuing development of the multilingual sign language wordnet. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, pages 343-353, 2024. 5", + "[40] Xiaolong Shen, Zhedong Zheng, and Yi Yang. Step-net: Spatial-temporal part-aware network for isolated sign language recognition. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-19, 2024. 6", + "[41] Ozge Mercanoglu Sincan and Hacer Yalim Keles. Autsl: A large scale multi-modal turkish sign language dataset and baseline methods. IEEE access, 8:181340-181355, 2020. 1", + "[42] Garrett Tanzer and Biao Zhang. Youtube-sl-25: A large-scale, open-domain multilingual sign language parallel corpus. arXiv preprint arXiv:2407.11144, 2024. 6", + "[43] Laia Tarres, Gerard I Gallego, Amanda Duarte, Jordi Torres, and Xavier Giró-i Nieto. Sign language translation from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5625-5635, 2023. 7, 8, 2", + "[44] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 2", + "[45] Dave Uthus, Garrett Tanzer, and Manfred Georg. Youtube: A large-scale, open-domain american sign language- english parallel corpus. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[46] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[47] Ville Viitaniemi, Tommi Jantunen, Leena Savolainen, Matti Karppa, and Jorma Laaksonen. S-pot-a benchmark in spotting signs within continuous signing. In LREC, pages 4-1, 2014. 6", + "[48] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Learnt contrastive concept embeddings for sign recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1945-1954, 2023. 2" + ], + "bbox": [ + 516, + 90, + 905, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22813", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7, 8", + "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6, 7", + "[51] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 4", + "[52] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2, 6", + "[53] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20871-20881, 2023. 1, 2", + "[54] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316–1325, 2021. 7, 2", + "[55] Wengang Zhou, Weichao Zhao, Hezhen Hu, Zecheng Li, and Houqiang Li. Scaling up multimodal pre-training for sign language understanding. arXiv preprint arXiv:2408.08544, 2024. 1, 2", + "[56] Simone Zini, Alex Gomez-Villa, Marco Buzzelli, Bartlomiej Twardowski, Andrew D. Bagdanov, and Joost van de weijer. Planckian jitter: countering the color-crippling effects of color jitter on self-supervised training. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[57] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14890-14900, 2023. 1, 2, 6" + ], + "bbox": [ + 89, + 92, + 483, + 669 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22814", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_model.json b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_model.json new file mode 100644 index 0000000000000000000000000000000000000000..93f3370b3670a25de01ad47d795df34ca3fb2d03 --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.795, + 0.153 + ], + "angle": 0, + "content": "SignRep: Enhancing Self-Supervised Sign Representations" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.18, + 0.778, + 0.234 + ], + "angle": 0, + "content": "Ryan Wong\\(^{1}\\), Necati Cihan Camgoz\\(^{2}\\), Richard Bowden\\(^{1}\\) \n\\(^{1}\\)University of Surrey, \\(^{2}\\)Meta Reality Labs \n{r.wong, r.bowden}@surrey.ac.uk, neccam@meta.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Sign language representation learning presents unique challenges due to the complex spatio-temporal nature of signs and the scarcity of labeled datasets. Existing methods often rely either on models pre-trained on general visual tasks, that lack sign-specific features, or use complex multimodal and multi-branch architectures. To bridge this gap, we introduce a scalable, self-supervised framework for sign representation learning. We leverage important inductive (sign) priors during the training of our RGB model. To do this, we leverage simple but important cues based on skeletons while pretraining a masked autoencoder. These sign specific priors alongside feature regularization and an adversarial style agnostic loss provide a powerful backbone. Notably, our model does not require skeletal keypoints during inference, avoiding the limitations of keypoint-based models during downstream tasks. When finetuned, we achieve state-of-the-art performance for sign recognition on the WLASL, ASL-Citizen and NMFs-CSL datasets, using a simpler architecture and with only a single-modality. Beyond recognition, our frozen model excels in sign dictionary retrieval and sign translation, surpassing standard MAE pretraining and skeletal-based representations in retrieval. It also reduces computational costs for training existing sign translation models while maintaining strong performance on Phoenix2014T, CSL-Daily and How2Sign." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.709, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Sign language is an important means of communication for millions of people worldwide. Sign languages have complex visual characteristics, which include intricate hand shapes, motions, body poses and facial expressions that models need to accurately interpret and process [5]. Furthermore, the computational demands of processing long video sequences add considerable challenges, making it difficult to scale these systems effectively. As a result, current approaches to sign recognition and translation often rely on general pretrained vision models [1, 7, 19, 22, 28, 49, 57]. Creating robust, label-free sign language representations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.905, + 0.301 + ], + "angle": 0, + "content": "that generalize across diverse datasets is challenging, yet essential for scalable sign language modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Most existing methods for sign recognition rely on multimodal or an ensemble of specialized models to achieve state-of-the-art recognition results. These models often require multi-channel inputs (e.g. RGB, depth and skeleton data) or specialized architectures to capture the complex interactions of hand, body or facial expressions [17, 20, 22, 23, 52, 57]. With each country having its own sign language and linguistic study expensive, currently available labeled datasets are sparse with typically under 2000 unique signs [1, 19, 26, 28, 41]. Collecting annotated sign language data is costly and time-consuming, making it infeasible to rely solely on labeled datasets for sign representation learning. This highlights the need for methods that can learn sign language representations from large-scale, unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.702 + ], + "angle": 0, + "content": "Newer methods for pretraining sign language translation models often require sentence-aligned annotations [25, 37, 49, 53, 55]. While these prior approaches jointly learn individual sign representations and inter-sign relationships, they are not scalable to unlabeled datasets as they require sentence-aligned annotations. We therefore focus on building a scalable and generalizable individual sign representation framework using self-supervised learning. A strong foundation in individual sign embeddings is an important step before building effective inter-sign models, as it ensures robust feature representations before incorporating complex long range temporal information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our main contributions are as follows: (1) We propose a scalable self-supervised Masked Autoencoding (MAE) framework which leverages sign priors, adversarial loss and feature regularizations for sign representation learning. (2) We introduce a method for analyzing sign class similarities in unseen datasets, introducing an auxiliary class probability distribution loss, which enhances recognition performance. (3) Our single pretrained model achieves state-of-the-art sign recognition, surpassing complex architecture and multimodal models. (4) We demonstrate the effectiveness of our sign representations for sign dictionary retrieval, achieving strong performance without any downstream training. (5) We demonstrate that using our pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22804" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "trained model as a frozen feature extractor, training is more tractable by reducing memory requirements and improving the performance of existing sign translation models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.151, + 0.233, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.485, + 0.525 + ], + "angle": 0, + "content": "Supervised Sign Recognition. Isolated Sign Recognition involves identifying a single sign within a given video. Sign recognition models often rely on pretrained spatiotemporal models from action recognition datasets such as Kinetics [8], fine-tuned for specific sign recognition tasks [1, 19, 22, 28, 57]. This transfer learning approach faces a domain shift problem, as sign videos have unique temporal dynamics and subtle gestures not well-represented in general action datasets, leading to performance degradation. To leverage invariance, some methods use skeletal keypoints instead of full-frame RGB models [10, 17, 20, 48, 52]. While keypoint-based models are more memory-efficient, they typically underperform relative to RGB models [2, 11, 36] and require complex architectural modifications to Graph Convolutional Networks (GCNs) [30] or transformers [46], to model the spatial-temporal relationships. Moreover, keypoints are prone to errors, such as missing or misdetected points [31, 32]. To solve these issues, recent state-of-the-art methods have combined keypoint and RGB modalities through branching or ensembling techniques [22, 23, 57]. While they achieve higher accuracy, they increase computational complexity due to the need for multiple models or support for multi-modal inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Self-Supervised Sign Representation Learning. Self-supervised learning for vision [3, 33, 38, 44] and language [12, 35] has demonstrated substantial benefits by leveraging unlabeled datasets for various tasks. In sign language, approaches such as Skeleton [24], SignBERT [17], SignBERT+ [20] and BEST [52] focus on keypoint-based self-supervised learning, employing masked learning similar to BERT [12]. However, these methods depend on keypoints as inputs and therefore requiring ensembling with RGB models to achieve competitive performance, resulting in increasing complexity and limited pretraining benefits. This highlights the need for a simple and efficient model which capturing sign-specific information and retains the pretraining advantages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Paired Sign-Text Pretraining for Sign Translation. Supervised pretraining approaches for sign translation, such as GFSLT-VLP [53], Sign2GPT [49], SignHiera [37], MSLU [55] and VAP [25], use sign video-spoken language pairs to enhance sign translation. Sign2GPT employs a pretrained DinoV2 [33], which is effective for general vision tasks but requires LoRA [16] fine-tuning for sign translation to achieve strong performance, significantly increasing computational costs. This highlights the domain gap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.304 + ], + "angle": 0, + "content": "between pretrained foundation vision models and sign language. SignHiera [37] demonstrates success with large-scale training on YT-ASL [45] using Hiera MAE [38] with language-supervised pretraining. However, they acknowledge that SignHiera requires significant computational resources (64 A100 GPUs for two weeks), emphasizing the need for more accessible and cost-effective video pretraining on large-scale datasets. These methods also face scalability challenges due to their reliance on sentence-aligned annotations. In this paper, we introduce a scalable self-supervised pretraining framework that learns sign representations without paired sign-text annotations, aiming to significantly reduce computational costs for pretraining and improve efficiency in downstream sign tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.634, + 0.334 + ], + "angle": 0, + "content": "3. Sign Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.48 + ], + "angle": 0, + "content": "Sign language communication relies on hand shapes, body posture and interactions. To guide the model toward meaningful sign representations, we introduce \"sign priors\", a set of cues that we know capture essential sign features. These priors \\((\\mathcal{P})\\) serve as the primary targets for our model, as detailed in Sec. 5.1. Using a human pose estimation model specialized for sign language videos [21], we extract joint angles and 3D keypoints. We categorize our priors into keypoint, angle, distance and signer activity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Keypoint Priors define the spatial structure of signing, we divide this into two priors:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.628 + ], + "angle": 0, + "content": "Hand Keypoint Prior \\((\\mathcal{P}^{\\{h,k\\}})\\): This prior is designed to capture hand shapes and orientations. We normalize the 3D coordinates of the 21 hand keypoints by setting the wrist as the origin, thereby eliminating positional variations in the body space. The resulting vector, \\(\\mathcal{P}^{\\{h,k\\}} \\in \\mathbb{R}^{21 \\times 3}\\), represents the configuration of the left or right hand, where \\(h \\in \\{\\mathrm{LH}, \\mathrm{RH}\\}\\) denotes the left or right hands." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.69 + ], + "angle": 0, + "content": "Full Body Keypoint Prior \\((\\mathcal{P}^{\\{b,k\\}})\\). This prior captures the positioning of the body and hands within the body space. We use all 61 3D keypoints (21 for each hand and 19 for the body), resulting in \\(\\mathcal{P}^{\\{b,k\\}} \\in \\mathbb{R}^{61 \\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Joint Angle Priors capture finger flexion, extension and encode upper-body posture variations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Hand Joint Angle Prior \\((\\mathcal{P}^{\\{h,a\\}})\\). This prior captures hand joint orientations. We extract 41 hand joint angles, yielding a vector \\(\\mathbb{R}^{41\\times 1}\\). To handle the continuous nature of angles, we apply sine and cosine transformations, resulting in the final angle prior \\(\\mathcal{P}^{\\{h,a\\}}\\in \\mathbb{R}^{41\\times 2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.886 + ], + "angle": 0, + "content": "Body Angle Pior \\((\\mathcal{P}^{\\{b,a\\}})\\). To capture the body's joint orientations, we extract the 22 body joint angles. These angles are transformed using sine and cosine functions to handle their continuous nature, resulting in the body angle prior \\(\\mathcal{P}^{\\{b,a\\}} \\in \\mathbb{R}^{22 \\times 2}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.912, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.298, + 0.908, + 0.356 + ], + "angle": 0, + "content": "Figure 1. (Left): The pretraining process for SignRep, which leverages masked representation learning to predict sign priors such as hand keypoints and joint angles. This is achieved through a Hiera encoder and a lightweight sign decoder. The representation is further refined with regularization losses, including variance, covariance and adversarial style loss. (Right): An example setup for the discriminator to obtain a representation pair to predict a style-representation match." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Distance Priors capture fine-grained differences between signs and the interactions of hands in the signing space:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Fingertip Distance Prior \\((\\mathcal{P}^{\\{h,d\\}})\\). Since fingertip interactions are a key component in sign language, we compute a distance matrix by measuring the distances from the fingertip keypoints to each of the knuckle, wrist and other fingertip keypoints. This produces the prior \\(\\mathcal{P}^{\\{h,d\\}} \\in \\mathbb{R}^{5 \\times 11 \\times 3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Hand-Interaction Distance Prior \\((\\mathcal{P}^{\\{b,d\\}})\\). Similar to the fingertip distance prior, we calculate a distance matrix using the wrist and five fingertip keypoints from each hand, resulting in a \\(6\\times 2\\) set of keypoints. The distance between the hands, body and facial keypoints is computed to capture the interactions between the hands and other body parts, resulting in the final hand-interaction prior \\(\\mathcal{P}^{\\{b,d\\}}\\in \\mathbb{R}^{12\\times 22\\times 3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Signer Activity. Some signs require only one hand or the signer may be in a resting pose. To enable the model to capture this information, we develop a simple heuristic. We define the active prior \\(\\mathcal{P}\\{h,\\mathrm{act}\\}\\), where a hand is considered inactive if it remains below the middle of the stomach and has not moved over the course of a video clip. This results in a prior \\(\\mathcal{P}\\{h,\\mathrm{act}\\} \\in [0,1]\\), indicating whether each hand is active (1) or inactive (0)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.754, + 0.299, + 0.772 + ], + "angle": 0, + "content": "4. SignRep Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To learn the sign priors, we introduce a novel adaptation of the Hierarchical Vision Transformer (Hiera) [38], designed with a pretraining task specifically tailored for sign representation learning without requiring labeled sign data. We choose the Hiera model as it aligns with our objective of being efficient, simple and effective. It has demonstrated strong performance in masked representation learning while eliminating the need for complex architectures and special" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.429 + ], + "angle": 0, + "content": "ized modules. We aim to develop a single spatio-temporal sign model that enhances performance without relying on multi-branch or ensemble methods, instead improving the representation learning process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.906, + 0.655 + ], + "angle": 0, + "content": "The standard Hiera MAE framework processes a video clip represented as \\( V = \\{v_{1}, v_{2}, \\dots, v_{t}, \\dots, v_{T}\\} \\), where \\( v_{t} \\) denotes the frame at time step \\( t \\), and \\( T \\) is the total number of frames. The video is first divided into spatiotemporal patches. A masking strategy is then applied, with a subset of patches randomly masked according to a masking ratio \\( M \\). The unmasked patches are fed into the encoder \\( f_{\\mathrm{enc}} \\), which processes them to generate an output representation \\( \\mathbf{z}^{\\mathrm{emb}} = \\{z_0^{\\mathrm{emb}}, z_1^{\\mathrm{emb}}, \\dots, z_K^{\\mathrm{emb}}\\} \\), where \\( K \\) is the number of output tokens from the encoder. The representation \\( \\mathbf{z}^{\\mathrm{emb}} \\) is then passed to a decoder, along with learnable masked tokens as input. The pretraining objective is to reconstruct the masked patches through pixel-level prediction, allowing the model to learn contextual relationships and enhance the visual representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Rather than pixel reconstruction, we use sign priors to learn meaningful sign representations, replacing the pixel reconstruction decoder with a lightweight sign decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Sign Decoder \\((f_{dec})\\). As shown in Fig. 1, instead of passing \\(\\mathbf{z}^{\\mathrm{emb}}\\) to the decoder, we take the average across \\(\\mathbf{z}^{\\mathrm{emb}}\\) tokens, which is then processed through layer normalization followed by a fully connected layer. This produces an output representation \\(z^{\\mathrm{avg}} \\in \\mathbb{R}^{1 \\times D}\\), where \\(D\\) is the dimensionality of the representation. The representation \\(z^{\\mathrm{avg}}\\) is then upsampled temporally to match the sequence length \\(T\\) of the input video. The upsampling module is implemented using a lightweight network consisting of a 1D convolution with a kernel size of 1, followed by a GELU activation function, and then a transpose convolution with a kernel size of \\(T\\) to match the number of input frames and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "an output dimension of \\(D^{\\prime}\\). This results in an upsampled vector \\(\\mathbf{z}^{\\mathrm{up}}\\in \\mathbb{R}^{T\\times D^{\\prime}}\\), where \\(\\mathbf{z}^{up} = [z_1^{up},z_2^{up},\\dots ,z_T^{up}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.483, + 0.153 + ], + "angle": 0, + "content": "We can then incorporate prediction heads for each sign prior to the network as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.164, + 0.483, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} ^ {\\mathcal {P}} = g _ {\\mathcal {P}} \\left(z _ {t} ^ {\\mathrm {u p}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.484, + 0.3 + ], + "angle": 0, + "content": "where each sign prior, \\(\\mathcal{P} \\in \\{\\mathcal{P}\\{h,k\\}, \\mathcal{P}\\{h,a\\}, \\mathcal{P}\\{h,d\\}, \\mathcal{P}\\{b,k\\}, \\mathcal{P}\\{b,a\\}, \\mathcal{P}\\{b,d\\}\\}\\) has a corresponding fully connected layer \\((g_{\\mathcal{P}})\\) with an associated output \\(\\hat{y}_t^{\\mathcal{P}}\\) which matches the flattened dimension of the corresponding sign prior. For the activity prior we simply use an MLP layer using \\(z^{\\mathrm{avg}}\\) as input which produces an output of 2 to match the dimension of \\([\\mathcal{P}\\{\\mathrm{LH,act}\\}, \\mathcal{P}\\{\\mathrm{RH,act}\\}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.421 + ], + "angle": 0, + "content": "The decoder is completely removed during downstream tasks and only the encoder is used as the sign representation model. As a result, the target sign priors are not required during downstream tasks, allowing the use of a single model that inherently captures sign knowledge. This approach ensures computational efficiency by leveraging masked representation learning while directly learning the essential sign priors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.435, + 0.378, + 0.453 + ], + "angle": 0, + "content": "5. Sign Representation Objectives" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.46, + 0.336, + 0.476 + ], + "angle": 0, + "content": "5.1. Sign Priors Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.484, + 0.679 + ], + "angle": 0, + "content": "In the previous section, we identified the sign priors that our model needs to learn. Since the sign decoder applies upsampling to match the number of input frames, each frame's prior has an associated output prediction. For each predicted sign prior output, \\(\\hat{y}_t^{\\mathcal{P}}\\), we train the model to regress to the corresponding target value \\(y_t^{\\mathcal{P}}\\) from prior \\(\\mathcal{P}\\) at frame \\(t\\), using smooth L1 loss (\\(L1\\)) across all sign priors for each frame in the input video sequence. The same loss is applied to the sign activity prior for consistency. We also mitigate the impact of low-quality keypoints in our objective function by only using keypoints with over \\(50\\%\\) confidence from the pose estimation model and masking missing keypoints from the loss. The final reconstruction loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.689, + 0.483, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {\\mathcal {P} \\in \\mathbf {P}} w _ {\\mathcal {P}} L 1 (\\hat {y} ^ {\\mathcal {P}}, y ^ {\\mathcal {P}}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.732, + 0.405, + 0.749 + ], + "angle": 0, + "content": "where \\(w_{\\mathcal{P}}\\) is the loss weighting for the prior \\(\\mathcal{P}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.364, + 0.774 + ], + "angle": 0, + "content": "5.2. Representation Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Feature Regularization. Self-supervised methods have shown that regularizing features improves representation quality [4, 51]. To enhance our representations, we incorporate a variance and covariance loss into \\( z^{\\mathrm{avg}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The variance loss, \\(\\mathcal{L}_{\\mathrm{var}}\\), encourages diversity in the learned representations by spreading them across the representation space. To compute the variance loss, we first calculate the standard deviation for each feature dimension" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.668, + 0.107 + ], + "angle": 0, + "content": "\\((j)\\) across a batch \\((N)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.117, + 0.905, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {j} = \\sqrt {\\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left(z _ {i , j} ^ {\\mathrm {a v g}} - \\bar {z} _ {j} ^ {\\mathrm {a v g}}\\right) ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.176, + 0.907, + 0.221 + ], + "angle": 0, + "content": "The diversity of the features are then encouraged through a hinge function to ensure that the variance does not fall below a threshold of one." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.23, + 0.905, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v a r}} = \\sum_ {j = 1} ^ {D} \\max (0, 1 - \\sigma_ {j}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.284, + 0.907, + 0.343 + ], + "angle": 0, + "content": "The covariance loss, \\(\\mathcal{L}_{\\mathrm{cov}}\\), reduces correlations among features, helping to avoid redundancy. To compute the covariance loss, we first need to calculate the covariance matrix such that:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.352, + 0.905, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {j, k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(z _ {i, j} ^ {\\text {a v g}} - \\bar {z} _ {j} ^ {\\text {a v g}}\\right) \\left(z _ {i, k} ^ {\\text {a v g}} - \\bar {z} _ {k} ^ {\\text {a v g}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.404, + 0.906, + 0.48 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_{j,k}\\) represents the covariance between feature dimensions \\(j\\) and \\(k\\) and \\(\\bar{z}_j^{\\mathrm{avg}}\\) is the mean of the \\(j\\)-th feature across the batch. The loss then penalises off-diagonal values in the covariance matrix, encouraging different feature dimensions to be uncorrelated:" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.489, + 0.905, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o v}} = \\sum_ {j \\neq k} \\mathcal {C} _ {j, k} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Style Agnostic Representations. We also aim to encourage the encoder to capture robust and generalizable sign features, while filtering out irrelevant details such as background and person-specific appearance. We can explicitly learn this by introducing a discriminator that evaluates whether two representations share the same \"style\", defined here by background and appearance features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.735 + ], + "angle": 0, + "content": "To extract style information, we calculate the gram matrix, commonly used in image style transfer [15], from the \\(\\mathbf{z}^{\\mathrm{emb}}\\) tokens, averaging across the column dimension to produce the style representation \\(z^{\\mathrm{style}}\\in \\mathbb{R}^{D}\\). We then pair \\(z^{\\mathrm{avg}}\\) with \\(z^{\\mathrm{style}}\\) and pass them to the discriminator, which learns to output 1 for matching pairs and 0 for non-matching pairs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.906, + 0.856 + ], + "angle": 0, + "content": "To generate these representations pairs, as shown in Fig. 1(right), we randomly crop a video sequence of length \\( L \\) to create two segments, \\( A_{1} \\) and \\( A_{2} \\), each of length \\( T \\). For each, we extract \\( z_{A_1}^{\\mathrm{avg}} \\) and \\( z_{A_1}^{\\mathrm{style}} \\), as well as \\( z_{A_2}^{\\mathrm{avg}} \\) and \\( z_{A_2}^{\\mathrm{style}} \\), assuming that they share background and appearance features due to originating from the same video. From a different video, we obtain segment \\( B \\) with its style \\( z_B^{\\mathrm{style}} \\), which provides a contrasting background and appearance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.904 + ], + "angle": 0, + "content": "The discriminator, \\(\\mathfrak{D}\\), is then trained to produce 0 output for mismatched styles and 1 for matched styles, such that \\(\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_B^{\\mathrm{style}}) = 0\\) and \\(\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_{A_2}^{\\mathrm{style}}) = 1\\). This can be used" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22807" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "to further enhance the generalization of the encoder, we add an adversarial loss during pretraining, designed to fool the discriminator which encourages the model to focus on sign-specific content over style-related features. The adversarial loss is formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.176, + 0.483, + 0.195 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p o s} ^ {\\left(A _ {1}, A _ {2}\\right)} = \\max \\left(0, \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {A _ {2}} ^ {\\text {s t y l e}}\\right) - \\mathbb {E} _ {q \\sim \\mathrm {U}} \\mathfrak {D} (q)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.198, + 0.482, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n e g}} ^ {(A _ {1}, B)} = \\max (0, \\mathbb {E} _ {q \\sim \\mathrm {M}} \\mathfrak {D} (q) - \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {B} ^ {\\text {s t y l e}}\\right)) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.22, + 0.482, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} ^ {A _ {1}} = \\left(\\mathcal {L} _ {p o s} ^ {(A _ {1}, A _ {2})}\\right) ^ {2} + \\left(\\mathcal {L} _ {n e g} ^ {(A _ {1}, B)}\\right) ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.484, + 0.352 + ], + "angle": 0, + "content": "where \\(\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q)\\) is the expected discriminator output of style for representation pairs that come from the same video sequence and \\(\\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)\\) for when they come from different video sequence. We apply this loss only if \\(\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q) > \\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)\\), ensuring stability by restricting the adversarial loss to cases where the discriminator is likely to correctly identify the matches." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.354, + 0.393, + 0.368 + ], + "angle": 0, + "content": "Our final loss for pretraining is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.378, + 0.482, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {r e c o n}} + w _ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}} + w _ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + w _ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.434 + ], + "angle": 0, + "content": "where \\( w_{\\mathrm{adv}} \\), \\( w_{\\mathrm{var}} \\) and \\( w_{\\mathrm{cov}} \\) are the respective weighting factors for the adversarial, variance and covariance losses." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.448, + 0.456, + 0.466 + ], + "angle": 0, + "content": "6. Representations for Dictionary Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Sign dictionary retrieval is the task of identifying signs from a predefined database. This allows for quick lookup of related signs given a query sign which supports efficient dataset creation and promotes cross-linguistic understanding of signs [39]. The SignRep model can be used as a feature extractor for dictionary retrieval. Given a query sign video of length \\( L \\), we apply a sliding window to extract segment representations \\( \\mathbf{Z}^{\\mathrm{avg}} = \\{z_1^{\\mathrm{avg}}, z_2^{\\mathrm{avg}}, \\ldots, z_n^{\\mathrm{avg}}, \\ldots, z_N^{\\mathrm{avg}}\\} \\), where \\( N = \\left\\lfloor \\frac{L - T}{\\text{stride}} \\right\\rfloor + 1 \\). Since isolated signs typically begin and end in a resting pose, we use the output from" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.647, + 0.482, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {n} ^ {\\mathrm {a c t}} = \\max \\left(\\mathcal {P} ^ {\\{L H, \\mathrm {a c t} \\}}, \\mathcal {P} ^ {\\{R H, \\mathrm {a c t} \\}}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.721 + ], + "angle": 0, + "content": "to identify whether the signer is active. By checking hand activity, we compute a weighted average of the segment representations resulting in" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.73, + 0.482, + 0.771 + ], + "angle": 0, + "content": "\\[\nz ^ {\\text {o u t}} = \\frac {1}{\\sum \\gamma_ {n} ^ {\\text {a c t}}} \\sum_ {n = 1} ^ {N} \\gamma_ {n} ^ {\\text {a c t}} z _ {n} ^ {\\text {a v g}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.841 + ], + "angle": 0, + "content": "where \\( z^{\\mathrm{out}} \\) is the representation for the query video. Finally, we extract features for all dictionary videos and classify a query video by finding the closest match using cosine similarity of the normalized representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Class Probability Distribution. The retrieval approach enables the recognition of visually similar sign classes without additional training, which is useful for improving the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "accuracy of downstream tasks such as recognition. We construct a class distribution by computing the cosine similarity across all samples in the training dataset, capturing the visual proximity between them. From the resulting cosine similarity matrix, we compute inter-class similarities using the sign-sample labels to generate a matrix of shape \\(\\mathbb{R}^{C\\times C}\\), where \\(C\\) represents the number of classes. Applying a temperature-scaled softmax to each row of this matrix yields a probability distribution for each class relative to the others, forming a class probability distribution matrix \\(\\phi \\in \\mathbb{R}^{C\\times C}\\), where each element of \\(\\phi\\) lies in the range [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.906, + 0.363 + ], + "angle": 0, + "content": "We can incorporate the class probability distribution \\(\\phi\\) as a regularization term in the downstream recognition task using KL divergence. Given a sign video with target label \\(c\\), we can compute the KL divergence between the predicted class distribution \\(\\bar{y}\\) (from the output classifier layer) and \\(\\phi_c\\) which is the pre-computed class probability distribution of \\(c\\). This results in:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.377, + 0.905, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\phi} = \\kappa \\phi_ {c} [ c ] \\mathrm {K L} \\left(\\phi_ {c} \\| \\bar {y}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.405, + 0.906, + 0.467 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the weighting factor for the loss and \\(\\mathrm{KL}(\\phi_c\\parallel \\bar{y})\\) is the KL divergence. To emphasize the contribution of the target class distribution, we scale the KL divergence by \\(\\phi_c[c]\\), which is the value of the class probability for \\(\\phi_c\\) at \\(c\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.646, + 0.496 + ], + "angle": 0, + "content": "7. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.657 + ], + "angle": 0, + "content": "We evaluate our model on sign recognition and retrieval using three datasets: ASL-Citizen [11], WLASL2000 [28], and NMFs-CSL [19]. The ASL-Citizen dataset includes 2,731 isolated ASL signs recorded from a webcam. The WLASL2000 dataset, sourced from the web, contains 2,000 common ASL signs and presents challenges due to its noisy nature and limited samples per sign. The NMFs-CSL dataset consists of 1,067 Chinese Sign Language (CSL) signs, which require recognition of non-manual cues, such as facial expressions, to accurately identify signs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.704, + 0.682 + ], + "angle": 0, + "content": "7.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.78 + ], + "angle": 0, + "content": "We follow standard evaluation protocols for sign recognition, measuring top-1 and top-5 per-instance and per-class accuracy on WLASL and NMFs-CSL. For ASL-Citizen, we use the benchmark metrics specified for this dataset, including discounted cumulative gain (DCG), mean reciprocal rank (MRR), and top-1 and top-5 instance accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To evaluate our model as a frozen feature extractor, we apply a retrieval protocol based on the dictionary-based retrieval approach from [11]. We choose a retrieval approach over a linear evaluation protocol, as dictionary retrieval is a common and practical task in sign language processing. This method also offers a better evaluation of the generalization of the features to unseen datasets, which is particularly important given the evolving nature of sign languages." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22808" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.293, + 0.108 + ], + "angle": 0, + "content": "7.2. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.249 + ], + "angle": 0, + "content": "Pretraining. We pretrain our model on the YouTube-SL-25 dataset [42], which consists of large-scale continuous sign videos from YouTube. We initialize the model with the base video Hiera-B architecture, pretrained on Kinetics using MAE. We randomly select 16 consecutive frames as inputs based on a co-articulated sign typically lasting for around 13 frames [1, 34, 47]. No changes are made to the original encoder architecture, while the decoder is replaced with our lightweight sign decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.484, + 0.4 + ], + "angle": 0, + "content": "Recognition. For downstream tasks, our approach leverages only the pretrained encoder, discarding the decoder and entirely eliminating the reliance on keypoint extraction. For recognition, we extend the pretrained model's input sequence to 64 frames. We preserve the pretrained weights by inflating the patch embeddings, which preserves the computational efficiency by avoiding increasing the number of tokens in the transformer, detailed in the supplementary material. A linear classifier is applied to the pooled features from the encoder model to predict the sign classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.484, + 0.491 + ], + "angle": 0, + "content": "Retrieval. For dictionary retrieval evaluation, we follow the approach outlined in Sec. 6 using a stride of 2. For the label assignment, we compute the cosine similarity between the test video representation and those in the training dictionary. Each test query video is assigned the label of the training video that yields the highest similarity score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Additional implementation details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.371, + 0.548 + ], + "angle": 0, + "content": "7.3. Evaluation on Sign Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.75 + ], + "angle": 0, + "content": "In Tab. 1, our single-modality model demonstrates significant improvements over all existing single-modality approaches, achieving performance on par with complex multi-modal methods for WLASL2000. While ensembling with multi-modal approaches could further improve performance, we emphasize the importance of pretraining. Our Hiera encoder model with our pretraining, achieves these results without the extensive architectural modifications required by other methods. Instead of relying on additional modalities as input, our pretrained model has learned these features to create a robust sign representation model. Importantly, our pretraining strategy does not rely on any annotated sign data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The effectiveness of our SignRep model on the NMFs-CSL dataset is further illustrated in Tab. 2, where we achieve a top-1 accuracy of \\(84.1\\%\\) and a top-5 accuracy of \\(98.8\\%\\), outperforming all other methods, including multimodal approaches. Notably, the closest single-modality comparison is the StepNet RGB model, which achieves \\(77.2\\%\\) in top-1 accuracy, nearly \\(7\\%\\) lower than our model. While StepNet focuses on developing specialized architecture for sign, we demonstrate our framework can effectively learn sign features from large-scale datasets." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.916, + 0.467 + ], + "angle": 0, + "content": "
MethodInstance Acc.Class Acc.
Top-1Top-5Top-1Top-5
Skeleton-based
ST-GCN [50]34.4066.5732.5365.45
SignBERT [17]39.4073.3536.7472.38
BEST [52]46.2579.3343.5277.65
SignBERT+[20]48.8582.4846.3781.33
Multi-modal
BEST (+R) [52]54.5988.0852.1287.28
SignBERT(+R) [17]54.6987.4952.0886.93
SignBERT+(+R) [20]55.5989.3753.3388.82
SAM (5ξ) [23]58.7391.4655.9390.94
SAM-v2 (5ξ) [22]59.3991.4856.6390.89
NLA-SLR [57]61.0591.4558.0590.70
NLA-SLR(3ξ) [57]61.2691.7758.3190.91
StepNet (R+F) [40]61.1791.9458.4391.43
RGB-based
I3D [8]32.4857.31--
I3D(BSL1K) [1]46.8279.3644.7278.47
StepNet [40]56.8988.6454.5487.97
SignRep (Ours)61.0590.2758.8989.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Table 1. Comparison of downstream sign recognition results on WLASL2000. \\(\\xi\\), (R) and (F) denotes a multi-crop inference, RGB and Optical Flow modality respectively." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.537, + 0.869, + 0.757 + ], + "angle": 0, + "content": "
MethodTop-1Top-5
I3D [8]64.488.0
TSM [29]64.588.7
Slowfast [14]66.386.6
GLE-Net [19]69.088.1
HMA (◇) [18]75.695.3
StepNet [40]77.292.5
SignBERT (H+R) (◇) [17]78.497.3
BEST (◇) [52]79.297.1
NLA-SLR (◇) [57]83.498.3
StepNet (R+F) (◇) [40]83.697.0
NLA-SLR (◇,3ξ) [57]83.798.5
SignRep (Ours)84.198.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.767, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Table 2. Comparison of downstream sign recognition results on NMFs-CSL. \\(\\diamond\\) denotes methods using multi-modality/ensemble models and \\(\\xi\\) indicates a multi-crop inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our results on the ASL-Citizen dataset, shown in Tab. 3, further demonstrate the effectiveness of our model where it surpasses the previous baseline I3D model by \\(18\\%\\) in top-1 accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22809" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.172 + ], + "angle": 0, + "content": "
ModelDCGMRRRec@1Rec@5
ST-GCN* [50]76.3769.9759.5282.68
I3D* [8]79.1373.3263.1086.09
SignRep (Ours)90.8488.0581.3796.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.183, + 0.483, + 0.213 + ], + "angle": 0, + "content": "Table 3. Comparison of downstream sign recognition results on ASL-Citizen. * denotes results produced by [11]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.228, + 0.436, + 0.246 + ], + "angle": 0, + "content": "7.4. Evaluation on Sign Dictionary Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.251, + 0.483, + 0.446 + ], + "angle": 0, + "content": "To assess the effectiveness of our model as a feature extractor, we utilize a dictionary-based sign retrieval method that directly evaluates the quality of the learned representations without any fine-tuning. In Tab. 4, we compare our approach to the Hiera model pretrained with MAE on both the Kinetics and YT-SL datasets. We apply the same training settings outlined in Sec. 7.2 for pretraining on YT-SL. Our results demonstrate that our proposed pretraining strategy significantly outperforms standard MAE, even when trained on YT-SL, by optimizing specifically for sign language representation, whereas MAE's pixel-reconstruction objective does not directly support learning the sign specific features needed for sign retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.63 + ], + "angle": 0, + "content": "We also compare our model's retrieval performance with the use of raw 3D keypoints and joint angles averaged over time. As shown in Tab. 4, our method achieves a substantial performance increase, with top-1 retrieval scores more than tripling those based on hand joint angles. These results underscore the strength of our pretraining approach in capturing spatio-temporal features specific to sign language, establishing our method as a highly effective framework for sign representation learning. We also validate the importance of incorporating hand activity awareness, observing that weighted averaging based on hand activity improves retrieval performance compared to averaging." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.255, + 0.654 + ], + "angle": 0, + "content": "7.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Impact of Pretraining on Sign Retrieval. In Tab. 5, we examine the effects of various elements in our pretraining strategy on retrieval performance in WLASL2000. Our results show that combining all three types of sign priors that have angle, keypoint and distance yields the highest retrieval accuracy, underscoring the benefit of capturing multiple aspects of the sign features. We also analyze the role of masking during pretraining, removing it leads to a noticeable decline in retrieval performance, indicating that masking is essential for robust representation learning. While we observe lower reconstruction loss without masking, this does not translate to better retrieval scores, suggesting that precise reconstruction alone does not necessarily produce effective sign representations. Finally, the addition of both adversarial loss and variance-covariance regularization yields the strongest retrieval results, confirming that" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "these feature regularization techniques improve the quality and robustness of learned sign representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Impact of Pretraining for Recognition. In Tab. 6, we demonstrate that our pretrained model substantially outperforms both the MAE model pretrained on Kinetics and the model pretrained on YT-SL with pixel MAE. Additionally, the baseline Hiera model pretrained on Kinetics, achieving a top-1 accuracy of \\(51.5\\%\\), surpasses prior keypoint-based masked learning methods such as SignBERT, BEST and SignBERT+, as shown in Tab. 1. This result underscores the value of RGB input, revealing the limitations of using keypoints as a primary modality. Previous keypoint methods have required ensembling with RGB to remain competitive, illustrating the challenges of relying solely on skeletal information as an input modality for sign recognition. Finally, we observe that incorporating the sign class distribution loss further enhances recognition accuracy, improving top-1 accuracy from \\(59.9\\%\\) to \\(61.0\\%\\) with a \\(\\kappa\\) of 0.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.863, + 0.41 + ], + "angle": 0, + "content": "8. Feature Extractor for Sign Translation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.584 + ], + "angle": 0, + "content": "In this section, we evaluate SignRep as a feature extractor for sign translation by integrating it into two open-source translation models: Sign2GPT [49] on RWTH-PHOENIX-Weather 2014T (Phoenix14T) [6] and CSL-Daily [54], as well as the approach from [43] on How2Sign [13]. We chose these models because they allow direct evaluation of SignRep as a frozen feature extractor without requiring sign-text pretraining or significant modifications to the architecture. We follow the same translation training hyperparameters as the original papers to ensure a fair comparison and isolate the impact of replacing the visual backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.901 + ], + "angle": 0, + "content": "For Phoenix14T and CSL-Daily, we omit the pseudogloss pretraining used in [49] to ensure a fair evaluation of SignRep's learned features without additional linguistic supervision. In [49], they show that the DinoV2 backbone requires fine-tuning with LoRA to perform well on sign translation. Instead, we replace the original learnable DinoV2 backbone with pre-extracted SignRep features, obtained using our sliding window approach with a stride of 2. This setup places ours at a disadvantage, as we remove the trainable backbone weights and data augmentation, relying solely on pre-extracted features. Despite this, SignRep achieves competitive translation performance, as shown in Tab. 7, performing comparably on Phoenix14T and achieving substantial gains on CSL-Daily. By replacing the learnable DinoV2 model with pre-extracted SignRep features during training, we significantly reduce computational costs, making large-scale translation training more efficient and accessible without sacrificing performance. Our approach removes an additional 5.5GFLOP per frame required by the learnable DinoV2 model, which is important as translation models process tens to hundreds of frames." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22810" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.27 + ], + "angle": 0, + "content": "
FeaturesASL-CitizenWLASL2000NMFs-CSL
DCGRec@1Rec@5DCGRec@1Rec@5DCGRec@1Rec@5
HieraMAE-Kinetic11.640.250.8413.212.083.4023.293.9612.18
HieraMAE-YTSL12.120.391.3414.062.574.4128.037.5718.38
All Joint Angles13.820.572.1717.922.546.5032.517.9325.50
All Keypoints14.290.982.7419.373.168.2336.6412.2630.99
Hand Keypoints25.917.9619.5828.117.5720.9241.7215.9142.62
Hand Joint Angle26.938.8121.4130.619.4224.3644.1718.1346.34
SignRep (avg)61.4037.4768.7753.5726.1660.9879.5158.4891.85
SignRep (weighted)71.2149.9580.0957.9329.9267.4183.0563.0495.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.167, + 0.279, + 0.832, + 0.295 + ], + "angle": 0, + "content": "Table 4. Comparison of generalization results on sign retrieval with no downstream training applied to SignRep." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.306, + 0.472, + 0.464 + ], + "angle": 0, + "content": "
Prior Comp.Reg.DCG
anglekptdistmaskvar+covadv
✓✓45.1
✓✓32.9
✓✓47.2
✓✓✓✓48.5
✓✓✓46.3
✓✓✓✓✓49.9
✓✓✓✓✓✓50.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Table 5. Comparison of the impact of sign priors, masking and regularization on the impact of retrieval. Retrieval results are obtained on WLASL using a stride of 8." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.526, + 0.481, + 0.67 + ], + "angle": 0, + "content": "
MethodKInstance Acc.Class Acc.
top-1top-5top-1top-5
MAE (Kinetic)-51.583.448.482.0
MAE (YT-SL)-57.287.454.686.3
SignRep0.059.990.257.489.2
SignRep0.160.490.157.889.1
SignRep0.261.090.358.989.4
SignRep0.559.889.457.788.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Table 6. Comparisons of the impact of pretraining and weighting for the class distribution loss on WLASL." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.719, + 0.486, + 0.803 + ], + "angle": 0, + "content": "
Sign2GPT BackbonePhoenix14TCSL-Daily
B-4RB-4R
DinoV2(LoRA)19.4245.2312.9641.12
SignRep(extracted)20.3845.1716.3342.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.812, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Table 7. Comparison of translation results using Sign2GPT [49] by replacing the learnable DinoV2 backbone with our extracted SignRep features. B-4 denotes BLEU4 and R denotes ROUGE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "This leads to an \\(80\\%\\) reduction in GPU memory usage during training, lowering requirements from 60GB to 11GB" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.305, + 0.875, + 0.389 + ], + "angle": 0, + "content": "
Tarrés's BackboneHow2Sign
rBLEUBLEU
I3D(Supervised)2.218.03
SignRep(Self-Supervised)2.748.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.399, + 0.905, + 0.428 + ], + "angle": 0, + "content": "Table 8. Comparison of translation results using Tarrés [43], replacing the I3D features with our SignRep features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.905, + 0.485 + ], + "angle": 0, + "content": "with a batch size of 8. This reduction makes training on standard hardware more feasible, enabling broader accessibility for researchers working with sign translation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.606 + ], + "angle": 0, + "content": "For How2Sign, in Tab. 8, we replace the supervised I3D features with our self-supervised SignRep features on the translation framework from [43] and observe performance improvements. Unlike I3D features, which are pretrained with supervision on a labeled sign recognition data, SignRep learns features purely through self-supervision, highlighting our framework's effectiveness for learning sign-specific features without supervision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.641, + 0.635 + ], + "angle": 0, + "content": "9. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduced a scalable, self-supervised framework for sign representation learning without labeled sign datasets. Using masked autoencoding with sign priors, adversarial loss and feature regularizations, our approach enhances generalization while eliminating the need for skeletal keypoints, multimodal or multi-branch architectures in downstream tasks. Our single model achieves state-of-the-art performance on multiple sign recognition benchmarks with a simpler, more efficient design. Beyond recognition, our representations demonstrate versatility in sign dictionary retrieval on unseen datasets and serve as an efficient feature extractor for existing sign translation systems, reducing computational costs while maintaining strong performance. Our findings highlight the potential of our self-supervised learning framework for scalable sign modeling and encourage further research into practical applications of sign representations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22811" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.254 + ], + "angle": 0, + "content": "This work was supported by SNSF project 'SMILE II' (CRSII5 193686), European Union's Horizon2020 programme ('EASIER' grant agreement 101016982) and the Innosuisse IICT Flagship (PFFS-21-47). This work reflects only the authors view and the Commission is not responsible for any use that may be made of the information it contains. Neither Necati Cihan Camgoz nor Meta were involved in the model training, evaluation, or use of the datasets. The authors also thank Maksym Ivashechkin for their assistance in providing the 3D pose estimation model." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.279, + 0.188, + 0.295 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European conference on computer vision, pages 35–53. Springer, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.374, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bencie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford british sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.444, + 0.485, + 0.526 + ], + "angle": 0, + "content": "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.527, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[4] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.623 + ], + "angle": 0, + "content": "[5] P Boyes Braem and RL Sutton-Spence. The Hands Are The Head of The Mouth. The Mouth as Articulator in Sign Languages. Hamburg: Signum Press, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.624, + 0.484, + 0.679 + ], + "angle": 0, + "content": "[6] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.804 + ], + "angle": 0, + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. Electra: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Konstantinos M Dafnis. Bidirectional skeleton-based isolated sign recognition using graph convolution networks. In LREC proceedings, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.304, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[11] Aashaka Desai, Lauren Berger, Fyodor Minakov, Nessa Milano, Chinmay Singh, Kriston Pumphrey, Richard Ladner, Hal Daume III, Alex X Lu, Naomi Caselli, et al. Asl citizen: a community-sourced dataset for advancing isolated sign language recognition. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.244 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[13] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.906, + 0.566 + ], + "angle": 0, + "content": "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11087-11096, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[18] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[19] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11221-11239, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[21] Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. Improving 3d pose estimation for sign language. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 1-5. IEEE, 2023. 2, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22812" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3413-3423, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[24] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3394-3402, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[25] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In European Conference on Computer Vision, pages 349-367. Springer, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[28] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[30] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[31] Amit Moryossef. Optimizing hand region detection in mediapipe holistic full-body pose estimation to improve accuracy and avoid downstream errors. arXiv preprint arXiv:2405.03545, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[32] Amit Moryossef, Ioannis Tsochantaridis, Joe Dinn, Necati Cihan Camgoz, Richard Bowden, Tao Jiang, Annette Rios, Mathias Muller, and Sarah Ebling. Evaluating the immediate applicability of pose estimation for sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3434-3440, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[34] Tomas Pfister, James Charles, and Andrew Zisserman. Large-scale learning of sign language by watching tv. 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[36] Charles Raude, KR Prajwal, Liliane Momeni, Hannah Bull, Samuel Albanie, Andrew Zisserman, and Gül Varol. A tale of two languages: Large-vocabulary continuous sign language recognition from spoken language supervision. arXiv preprint arXiv:2405.10266, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[37] Phillip Rust, Bowen Shi, Skyler Wang, Necati Cihan Camgoz, and Jean Maillard. Towards privacy-aware sign language translation at scale. arXiv preprint arXiv:2402.09611, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[38] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[39] Marc Schulder, Sam Bigeard, Maria Kopf, Thomas Hanke, Anna Kuder, Joanna Wojcicka, Johanna Mesch, Thomas Björkstrand, Anna Vacalopoulou, Kyriaki Vasilaki, et al. Signs and synonymity: Continuing development of the multilingual sign language wordnet. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, pages 343-353, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.486 + ], + "angle": 0, + "content": "[40] Xiaolong Shen, Zhedong Zheng, and Yi Yang. Step-net: Spatial-temporal part-aware network for isolated sign language recognition. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-19, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.906, + 0.53 + ], + "angle": 0, + "content": "[41] Ozge Mercanoglu Sincan and Hacer Yalim Keles. Autsl: A large scale multi-modal turkish sign language dataset and baseline methods. IEEE access, 8:181340-181355, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.906, + 0.574 + ], + "angle": 0, + "content": "[42] Garrett Tanzer and Biao Zhang. Youtube-sl-25: A large-scale, open-domain multilingual sign language parallel corpus. arXiv preprint arXiv:2407.11144, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[43] Laia Tarres, Gerard I Gallego, Amanda Duarte, Jordi Torres, and Xavier Giró-i Nieto. Sign language translation from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5625-5635, 2023. 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[44] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[45] Dave Uthus, Garrett Tanzer, and Manfred Georg. Youtube: A large-scale, open-domain american sign language- english parallel corpus. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[46] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[47] Ville Viitaniemi, Tommi Jantunen, Leena Savolainen, Matti Karppa, and Jorma Laaksonen. S-pot-a benchmark in spotting signs within continuous signing. In LREC, pages 4-1, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[48] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Learnt contrastive concept embeddings for sign recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1945-1954, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[49] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.192, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[51] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.25, + 0.485, + 0.317 + ], + "angle": 0, + "content": "[52] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.32, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[53] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20871-20881, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.404, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[54] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316–1325, 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.475, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[55] Wengang Zhou, Weichao Zhao, Hezhen Hu, Zecheng Li, and Houqiang Li. Scaling up multimodal pre-training for sign language understanding. arXiv preprint arXiv:2408.08544, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.531, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[56] Simone Zini, Alex Gomez-Villa, Marco Buzzelli, Bartlomiej Twardowski, Andrew D. Bagdanov, and Joost van de weijer. Planckian jitter: countering the color-crippling effects of color jitter on self-supervised training. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[57] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14890-14900, 2023. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.485, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22814" + } + ] +] \ No newline at end of file diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_origin.pdf b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9be0d5f38cccb8a3ac24e9cf9c3f751dbb3f8f44 --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/b0873b30-3769-4f25-84b2-07088f813479_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e07e58aba4409c929b894fb699d9473a7ccf973dc8764bfba912a7d344b06e96 +size 2037786 diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/full.md b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b910de7f3670f9478e8adb3dbc6b0b25eb19d868 --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/full.md @@ -0,0 +1,343 @@ +# SignRep: Enhancing Self-Supervised Sign Representations + +Ryan Wong $^{1}$ , Necati Cihan Camgoz $^{2}$ , Richard Bowden $^{1}$ $^{1}$ University of Surrey, $^{2}$ Meta Reality Labs +{r.wong, r.bowden}@surrey.ac.uk, neccam@meta.com + +# Abstract + +Sign language representation learning presents unique challenges due to the complex spatio-temporal nature of signs and the scarcity of labeled datasets. Existing methods often rely either on models pre-trained on general visual tasks, that lack sign-specific features, or use complex multimodal and multi-branch architectures. To bridge this gap, we introduce a scalable, self-supervised framework for sign representation learning. We leverage important inductive (sign) priors during the training of our RGB model. To do this, we leverage simple but important cues based on skeletons while pretraining a masked autoencoder. These sign specific priors alongside feature regularization and an adversarial style agnostic loss provide a powerful backbone. Notably, our model does not require skeletal keypoints during inference, avoiding the limitations of keypoint-based models during downstream tasks. When finetuned, we achieve state-of-the-art performance for sign recognition on the WLASL, ASL-Citizen and NMFs-CSL datasets, using a simpler architecture and with only a single-modality. Beyond recognition, our frozen model excels in sign dictionary retrieval and sign translation, surpassing standard MAE pretraining and skeletal-based representations in retrieval. It also reduces computational costs for training existing sign translation models while maintaining strong performance on Phoenix2014T, CSL-Daily and How2Sign. + +# 1. Introduction + +Sign language is an important means of communication for millions of people worldwide. Sign languages have complex visual characteristics, which include intricate hand shapes, motions, body poses and facial expressions that models need to accurately interpret and process [5]. Furthermore, the computational demands of processing long video sequences add considerable challenges, making it difficult to scale these systems effectively. As a result, current approaches to sign recognition and translation often rely on general pretrained vision models [1, 7, 19, 22, 28, 49, 57]. Creating robust, label-free sign language representations + +that generalize across diverse datasets is challenging, yet essential for scalable sign language modeling. + +Most existing methods for sign recognition rely on multimodal or an ensemble of specialized models to achieve state-of-the-art recognition results. These models often require multi-channel inputs (e.g. RGB, depth and skeleton data) or specialized architectures to capture the complex interactions of hand, body or facial expressions [17, 20, 22, 23, 52, 57]. With each country having its own sign language and linguistic study expensive, currently available labeled datasets are sparse with typically under 2000 unique signs [1, 19, 26, 28, 41]. Collecting annotated sign language data is costly and time-consuming, making it infeasible to rely solely on labeled datasets for sign representation learning. This highlights the need for methods that can learn sign language representations from large-scale, unlabeled data. + +Newer methods for pretraining sign language translation models often require sentence-aligned annotations [25, 37, 49, 53, 55]. While these prior approaches jointly learn individual sign representations and inter-sign relationships, they are not scalable to unlabeled datasets as they require sentence-aligned annotations. We therefore focus on building a scalable and generalizable individual sign representation framework using self-supervised learning. A strong foundation in individual sign embeddings is an important step before building effective inter-sign models, as it ensures robust feature representations before incorporating complex long range temporal information. + +Our main contributions are as follows: (1) We propose a scalable self-supervised Masked Autoencoding (MAE) framework which leverages sign priors, adversarial loss and feature regularizations for sign representation learning. (2) We introduce a method for analyzing sign class similarities in unseen datasets, introducing an auxiliary class probability distribution loss, which enhances recognition performance. (3) Our single pretrained model achieves state-of-the-art sign recognition, surpassing complex architecture and multimodal models. (4) We demonstrate the effectiveness of our sign representations for sign dictionary retrieval, achieving strong performance without any downstream training. (5) We demonstrate that using our pre + +trained model as a frozen feature extractor, training is more tractable by reducing memory requirements and improving the performance of existing sign translation models. + +# 2. Related Work + +Supervised Sign Recognition. Isolated Sign Recognition involves identifying a single sign within a given video. Sign recognition models often rely on pretrained spatiotemporal models from action recognition datasets such as Kinetics [8], fine-tuned for specific sign recognition tasks [1, 19, 22, 28, 57]. This transfer learning approach faces a domain shift problem, as sign videos have unique temporal dynamics and subtle gestures not well-represented in general action datasets, leading to performance degradation. To leverage invariance, some methods use skeletal keypoints instead of full-frame RGB models [10, 17, 20, 48, 52]. While keypoint-based models are more memory-efficient, they typically underperform relative to RGB models [2, 11, 36] and require complex architectural modifications to Graph Convolutional Networks (GCNs) [30] or transformers [46], to model the spatial-temporal relationships. Moreover, keypoints are prone to errors, such as missing or misdetected points [31, 32]. To solve these issues, recent state-of-the-art methods have combined keypoint and RGB modalities through branching or ensembling techniques [22, 23, 57]. While they achieve higher accuracy, they increase computational complexity due to the need for multiple models or support for multi-modal inputs. + +Self-Supervised Sign Representation Learning. Self-supervised learning for vision [3, 33, 38, 44] and language [12, 35] has demonstrated substantial benefits by leveraging unlabeled datasets for various tasks. In sign language, approaches such as Skeleton [24], SignBERT [17], SignBERT+ [20] and BEST [52] focus on keypoint-based self-supervised learning, employing masked learning similar to BERT [12]. However, these methods depend on keypoints as inputs and therefore requiring ensembling with RGB models to achieve competitive performance, resulting in increasing complexity and limited pretraining benefits. This highlights the need for a simple and efficient model which capturing sign-specific information and retains the pretraining advantages. + +Paired Sign-Text Pretraining for Sign Translation. Supervised pretraining approaches for sign translation, such as GFSLT-VLP [53], Sign2GPT [49], SignHiera [37], MSLU [55] and VAP [25], use sign video-spoken language pairs to enhance sign translation. Sign2GPT employs a pretrained DinoV2 [33], which is effective for general vision tasks but requires LoRA [16] fine-tuning for sign translation to achieve strong performance, significantly increasing computational costs. This highlights the domain gap + +between pretrained foundation vision models and sign language. SignHiera [37] demonstrates success with large-scale training on YT-ASL [45] using Hiera MAE [38] with language-supervised pretraining. However, they acknowledge that SignHiera requires significant computational resources (64 A100 GPUs for two weeks), emphasizing the need for more accessible and cost-effective video pretraining on large-scale datasets. These methods also face scalability challenges due to their reliance on sentence-aligned annotations. In this paper, we introduce a scalable self-supervised pretraining framework that learns sign representations without paired sign-text annotations, aiming to significantly reduce computational costs for pretraining and improve efficiency in downstream sign tasks. + +# 3. Sign Priors + +Sign language communication relies on hand shapes, body posture and interactions. To guide the model toward meaningful sign representations, we introduce "sign priors", a set of cues that we know capture essential sign features. These priors $(\mathcal{P})$ serve as the primary targets for our model, as detailed in Sec. 5.1. Using a human pose estimation model specialized for sign language videos [21], we extract joint angles and 3D keypoints. We categorize our priors into keypoint, angle, distance and signer activity. + +Keypoint Priors define the spatial structure of signing, we divide this into two priors: + +Hand Keypoint Prior $(\mathcal{P}^{\{h,k\}})$ : This prior is designed to capture hand shapes and orientations. We normalize the 3D coordinates of the 21 hand keypoints by setting the wrist as the origin, thereby eliminating positional variations in the body space. The resulting vector, $\mathcal{P}^{\{h,k\}} \in \mathbb{R}^{21 \times 3}$ , represents the configuration of the left or right hand, where $h \in \{\mathrm{LH}, \mathrm{RH}\}$ denotes the left or right hands. + +Full Body Keypoint Prior $(\mathcal{P}^{\{b,k\}})$ . This prior captures the positioning of the body and hands within the body space. We use all 61 3D keypoints (21 for each hand and 19 for the body), resulting in $\mathcal{P}^{\{b,k\}} \in \mathbb{R}^{61 \times 3}$ . + +Joint Angle Priors capture finger flexion, extension and encode upper-body posture variations: + +Hand Joint Angle Prior $(\mathcal{P}^{\{h,a\}})$ . This prior captures hand joint orientations. We extract 41 hand joint angles, yielding a vector $\mathbb{R}^{41\times 1}$ . To handle the continuous nature of angles, we apply sine and cosine transformations, resulting in the final angle prior $\mathcal{P}^{\{h,a\}}\in \mathbb{R}^{41\times 2}$ . + +Body Angle Pior $(\mathcal{P}^{\{b,a\}})$ . To capture the body's joint orientations, we extract the 22 body joint angles. These angles are transformed using sine and cosine functions to handle their continuous nature, resulting in the body angle prior $\mathcal{P}^{\{b,a\}} \in \mathbb{R}^{22 \times 2}$ . + +![](images/7a3373601e0de8e37928fe062b76cf893af3837e9d4ae6654b993722199aa6b2.jpg) +Figure 1. (Left): The pretraining process for SignRep, which leverages masked representation learning to predict sign priors such as hand keypoints and joint angles. This is achieved through a Hiera encoder and a lightweight sign decoder. The representation is further refined with regularization losses, including variance, covariance and adversarial style loss. (Right): An example setup for the discriminator to obtain a representation pair to predict a style-representation match. + +Distance Priors capture fine-grained differences between signs and the interactions of hands in the signing space: + +Fingertip Distance Prior $(\mathcal{P}^{\{h,d\}})$ . Since fingertip interactions are a key component in sign language, we compute a distance matrix by measuring the distances from the fingertip keypoints to each of the knuckle, wrist and other fingertip keypoints. This produces the prior $\mathcal{P}^{\{h,d\}} \in \mathbb{R}^{5 \times 11 \times 3}$ . + +Hand-Interaction Distance Prior $(\mathcal{P}^{\{b,d\}})$ . Similar to the fingertip distance prior, we calculate a distance matrix using the wrist and five fingertip keypoints from each hand, resulting in a $6\times 2$ set of keypoints. The distance between the hands, body and facial keypoints is computed to capture the interactions between the hands and other body parts, resulting in the final hand-interaction prior $\mathcal{P}^{\{b,d\}}\in \mathbb{R}^{12\times 22\times 3}$ + +Signer Activity. Some signs require only one hand or the signer may be in a resting pose. To enable the model to capture this information, we develop a simple heuristic. We define the active prior $\mathcal{P}\{h,\mathrm{act}\}$ , where a hand is considered inactive if it remains below the middle of the stomach and has not moved over the course of a video clip. This results in a prior $\mathcal{P}\{h,\mathrm{act}\} \in [0,1]$ , indicating whether each hand is active (1) or inactive (0). + +# 4. SignRep Architecture + +To learn the sign priors, we introduce a novel adaptation of the Hierarchical Vision Transformer (Hiera) [38], designed with a pretraining task specifically tailored for sign representation learning without requiring labeled sign data. We choose the Hiera model as it aligns with our objective of being efficient, simple and effective. It has demonstrated strong performance in masked representation learning while eliminating the need for complex architectures and special + +ized modules. We aim to develop a single spatio-temporal sign model that enhances performance without relying on multi-branch or ensemble methods, instead improving the representation learning process. + +The standard Hiera MAE framework processes a video clip represented as $V = \{v_{1}, v_{2}, \dots, v_{t}, \dots, v_{T}\}$ , where $v_{t}$ denotes the frame at time step $t$ , and $T$ is the total number of frames. The video is first divided into spatiotemporal patches. A masking strategy is then applied, with a subset of patches randomly masked according to a masking ratio $M$ . The unmasked patches are fed into the encoder $f_{\mathrm{enc}}$ , which processes them to generate an output representation $\mathbf{z}^{\mathrm{emb}} = \{z_0^{\mathrm{emb}}, z_1^{\mathrm{emb}}, \dots, z_K^{\mathrm{emb}}\}$ , where $K$ is the number of output tokens from the encoder. The representation $\mathbf{z}^{\mathrm{emb}}$ is then passed to a decoder, along with learnable masked tokens as input. The pretraining objective is to reconstruct the masked patches through pixel-level prediction, allowing the model to learn contextual relationships and enhance the visual representation. + +Rather than pixel reconstruction, we use sign priors to learn meaningful sign representations, replacing the pixel reconstruction decoder with a lightweight sign decoder. + +Sign Decoder $(f_{dec})$ . As shown in Fig. 1, instead of passing $\mathbf{z}^{\mathrm{emb}}$ to the decoder, we take the average across $\mathbf{z}^{\mathrm{emb}}$ tokens, which is then processed through layer normalization followed by a fully connected layer. This produces an output representation $z^{\mathrm{avg}} \in \mathbb{R}^{1 \times D}$ , where $D$ is the dimensionality of the representation. The representation $z^{\mathrm{avg}}$ is then upsampled temporally to match the sequence length $T$ of the input video. The upsampling module is implemented using a lightweight network consisting of a 1D convolution with a kernel size of 1, followed by a GELU activation function, and then a transpose convolution with a kernel size of $T$ to match the number of input frames and + +an output dimension of $D^{\prime}$ . This results in an upsampled vector $\mathbf{z}^{\mathrm{up}}\in \mathbb{R}^{T\times D^{\prime}}$ , where $\mathbf{z}^{up} = [z_1^{up},z_2^{up},\dots ,z_T^{up}]$ + +We can then incorporate prediction heads for each sign prior to the network as follows: + +$$ +\hat {y} _ {t} ^ {\mathcal {P}} = g _ {\mathcal {P}} \left(z _ {t} ^ {\mathrm {u p}}\right) \tag {1} +$$ + +where each sign prior, $\mathcal{P} \in \{\mathcal{P}\{h,k\}, \mathcal{P}\{h,a\}, \mathcal{P}\{h,d\}, \mathcal{P}\{b,k\}, \mathcal{P}\{b,a\}, \mathcal{P}\{b,d\}\}$ has a corresponding fully connected layer $(g_{\mathcal{P}})$ with an associated output $\hat{y}_t^{\mathcal{P}}$ which matches the flattened dimension of the corresponding sign prior. For the activity prior we simply use an MLP layer using $z^{\mathrm{avg}}$ as input which produces an output of 2 to match the dimension of $[\mathcal{P}\{\mathrm{LH,act}\}, \mathcal{P}\{\mathrm{RH,act}\}]$ . + +The decoder is completely removed during downstream tasks and only the encoder is used as the sign representation model. As a result, the target sign priors are not required during downstream tasks, allowing the use of a single model that inherently captures sign knowledge. This approach ensures computational efficiency by leveraging masked representation learning while directly learning the essential sign priors. + +# 5. Sign Representation Objectives + +# 5.1. Sign Priors Reconstruction + +In the previous section, we identified the sign priors that our model needs to learn. Since the sign decoder applies upsampling to match the number of input frames, each frame's prior has an associated output prediction. For each predicted sign prior output, $\hat{y}_t^{\mathcal{P}}$ , we train the model to regress to the corresponding target value $y_t^{\mathcal{P}}$ from prior $\mathcal{P}$ at frame $t$ , using smooth L1 loss ( $L1$ ) across all sign priors for each frame in the input video sequence. The same loss is applied to the sign activity prior for consistency. We also mitigate the impact of low-quality keypoints in our objective function by only using keypoints with over $50\%$ confidence from the pose estimation model and masking missing keypoints from the loss. The final reconstruction loss is as follows: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \sum_ {\mathcal {P} \in \mathbf {P}} w _ {\mathcal {P}} L 1 (\hat {y} ^ {\mathcal {P}}, y ^ {\mathcal {P}}) \tag {2} +$$ + +where $w_{\mathcal{P}}$ is the loss weighting for the prior $\mathcal{P}$ . + +# 5.2. Representation Regularization + +Feature Regularization. Self-supervised methods have shown that regularizing features improves representation quality [4, 51]. To enhance our representations, we incorporate a variance and covariance loss into $z^{\mathrm{avg}}$ . + +The variance loss, $\mathcal{L}_{\mathrm{var}}$ , encourages diversity in the learned representations by spreading them across the representation space. To compute the variance loss, we first calculate the standard deviation for each feature dimension + +$(j)$ across a batch $(N)$ . + +$$ +\sigma_ {j} = \sqrt {\frac {1}{N - 1} \sum_ {i = 1} ^ {N} \left(z _ {i , j} ^ {\mathrm {a v g}} - \bar {z} _ {j} ^ {\mathrm {a v g}}\right) ^ {2}} \tag {3} +$$ + +The diversity of the features are then encouraged through a hinge function to ensure that the variance does not fall below a threshold of one. + +$$ +\mathcal {L} _ {\text {v a r}} = \sum_ {j = 1} ^ {D} \max (0, 1 - \sigma_ {j}) \tag {4} +$$ + +The covariance loss, $\mathcal{L}_{\mathrm{cov}}$ , reduces correlations among features, helping to avoid redundancy. To compute the covariance loss, we first need to calculate the covariance matrix such that: + +$$ +\mathcal {C} _ {j, k} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(z _ {i, j} ^ {\text {a v g}} - \bar {z} _ {j} ^ {\text {a v g}}\right) \left(z _ {i, k} ^ {\text {a v g}} - \bar {z} _ {k} ^ {\text {a v g}}\right) \tag {5} +$$ + +where $\mathcal{C}_{j,k}$ represents the covariance between feature dimensions $j$ and $k$ and $\bar{z}_j^{\mathrm{avg}}$ is the mean of the $j$ -th feature across the batch. The loss then penalises off-diagonal values in the covariance matrix, encouraging different feature dimensions to be uncorrelated: + +$$ +\mathcal {L} _ {\mathrm {c o v}} = \sum_ {j \neq k} \mathcal {C} _ {j, k} ^ {2} \tag {6} +$$ + +Style Agnostic Representations. We also aim to encourage the encoder to capture robust and generalizable sign features, while filtering out irrelevant details such as background and person-specific appearance. We can explicitly learn this by introducing a discriminator that evaluates whether two representations share the same "style", defined here by background and appearance features. + +To extract style information, we calculate the gram matrix, commonly used in image style transfer [15], from the $\mathbf{z}^{\mathrm{emb}}$ tokens, averaging across the column dimension to produce the style representation $z^{\mathrm{style}}\in \mathbb{R}^{D}$ . We then pair $z^{\mathrm{avg}}$ with $z^{\mathrm{style}}$ and pass them to the discriminator, which learns to output 1 for matching pairs and 0 for non-matching pairs. + +To generate these representations pairs, as shown in Fig. 1(right), we randomly crop a video sequence of length $L$ to create two segments, $A_{1}$ and $A_{2}$ , each of length $T$ . For each, we extract $z_{A_1}^{\mathrm{avg}}$ and $z_{A_1}^{\mathrm{style}}$ , as well as $z_{A_2}^{\mathrm{avg}}$ and $z_{A_2}^{\mathrm{style}}$ , assuming that they share background and appearance features due to originating from the same video. From a different video, we obtain segment $B$ with its style $z_B^{\mathrm{style}}$ , which provides a contrasting background and appearance. + +The discriminator, $\mathfrak{D}$ , is then trained to produce 0 output for mismatched styles and 1 for matched styles, such that $\mathfrak{D}(z_{A_1}^{\mathrm{avg}}, z_B^{\mathrm{style}}) = 0$ and $\mathfrak{D}(z_{A_1}^{\mathrm{avg}}, z_{A_2}^{\mathrm{style}}) = 1$ . This can be used + +to further enhance the generalization of the encoder, we add an adversarial loss during pretraining, designed to fool the discriminator which encourages the model to focus on sign-specific content over style-related features. The adversarial loss is formalized as follows: + +$$ +\mathcal {L} _ {p o s} ^ {\left(A _ {1}, A _ {2}\right)} = \max \left(0, \mathfrak {D} \left(z _ {A _ {1}} ^ {\text {a v g}}, z _ {A _ {2}} ^ {\text {s t y l e}}\right) - \mathbb {E} _ {q \sim \mathrm {U}} \mathfrak {D} (q)\right) \tag {7} +$$ + +$$ +\mathcal {L} _ {\text {n e g}} ^ {(A _ {1}, B)} = \max (0, \mathbb {E} _ {q \sim \mathrm {M}} \mathfrak {D} (q) - \mathfrak {D} \left(z _ {A _ {1}} ^ {\text {a v g}}, z _ {B} ^ {\text {s t y l e}}\right)) \tag {8} +$$ + +$$ +\mathcal {L} _ {\mathrm {a d v}} ^ {A _ {1}} = \left(\mathcal {L} _ {p o s} ^ {(A _ {1}, A _ {2})}\right) ^ {2} + \left(\mathcal {L} _ {n e g} ^ {(A _ {1}, B)}\right) ^ {2} \tag {9} +$$ + +where $\mathbb{E}_{q\sim \mathrm{M}}\mathfrak{D}(q)$ is the expected discriminator output of style for representation pairs that come from the same video sequence and $\mathbb{E}_{q\sim \mathrm{U}}\mathfrak{D}(q)$ for when they come from different video sequence. We apply this loss only if $\mathbb{E}_{q\sim \mathrm{M}}\mathfrak{D}(q) > \mathbb{E}_{q\sim \mathrm{U}}\mathfrak{D}(q)$ , ensuring stability by restricting the adversarial loss to cases where the discriminator is likely to correctly identify the matches. + +Our final loss for pretraining is defined as: + +$$ +\mathcal {L} _ {\text {f i n a l}} = \mathcal {L} _ {\text {r e c o n}} + w _ {\text {v a r}} \mathcal {L} _ {\text {v a r}} + w _ {\text {c o v}} \mathcal {L} _ {\text {c o v}} + w _ {\text {a d v}} \mathcal {L} _ {\text {a d v}} \tag {10} +$$ + +where $w_{\mathrm{adv}}$ , $w_{\mathrm{var}}$ and $w_{\mathrm{cov}}$ are the respective weighting factors for the adversarial, variance and covariance losses. + +# 6. Representations for Dictionary Retrieval + +Sign dictionary retrieval is the task of identifying signs from a predefined database. This allows for quick lookup of related signs given a query sign which supports efficient dataset creation and promotes cross-linguistic understanding of signs [39]. The SignRep model can be used as a feature extractor for dictionary retrieval. Given a query sign video of length $L$ , we apply a sliding window to extract segment representations $\mathbf{Z}^{\mathrm{avg}} = \{z_1^{\mathrm{avg}}, z_2^{\mathrm{avg}}, \ldots, z_n^{\mathrm{avg}}, \ldots, z_N^{\mathrm{avg}}\}$ , where $N = \left\lfloor \frac{L - T}{\text{stride}} \right\rfloor + 1$ . Since isolated signs typically begin and end in a resting pose, we use the output from + +$$ +\gamma_ {n} ^ {\mathrm {a c t}} = \max \left(\mathcal {P} ^ {\{L H, \mathrm {a c t} \}}, \mathcal {P} ^ {\{R H, \mathrm {a c t} \}}\right) \tag {11} +$$ + +to identify whether the signer is active. By checking hand activity, we compute a weighted average of the segment representations resulting in + +$$ +z ^ {\text {o u t}} = \frac {1}{\sum \gamma_ {n} ^ {\text {a c t}}} \sum_ {n = 1} ^ {N} \gamma_ {n} ^ {\text {a c t}} z _ {n} ^ {\text {a v g}} \tag {12} +$$ + +where $z^{\mathrm{out}}$ is the representation for the query video. Finally, we extract features for all dictionary videos and classify a query video by finding the closest match using cosine similarity of the normalized representations. + +Class Probability Distribution. The retrieval approach enables the recognition of visually similar sign classes without additional training, which is useful for improving the + +accuracy of downstream tasks such as recognition. We construct a class distribution by computing the cosine similarity across all samples in the training dataset, capturing the visual proximity between them. From the resulting cosine similarity matrix, we compute inter-class similarities using the sign-sample labels to generate a matrix of shape $\mathbb{R}^{C\times C}$ , where $C$ represents the number of classes. Applying a temperature-scaled softmax to each row of this matrix yields a probability distribution for each class relative to the others, forming a class probability distribution matrix $\phi \in \mathbb{R}^{C\times C}$ , where each element of $\phi$ lies in the range [0, 1]. + +We can incorporate the class probability distribution $\phi$ as a regularization term in the downstream recognition task using KL divergence. Given a sign video with target label $c$ , we can compute the KL divergence between the predicted class distribution $\bar{y}$ (from the output classifier layer) and $\phi_c$ which is the pre-computed class probability distribution of $c$ . This results in: + +$$ +\mathcal {L} _ {\phi} = \kappa \phi_ {c} [ c ] \mathrm {K L} \left(\phi_ {c} \| \bar {y}\right) \tag {13} +$$ + +where $\kappa$ is the weighting factor for the loss and $\mathrm{KL}(\phi_c\parallel \bar{y})$ is the KL divergence. To emphasize the contribution of the target class distribution, we scale the KL divergence by $\phi_c[c]$ , which is the value of the class probability for $\phi_c$ at $c$ . + +# 7. Experiments + +We evaluate our model on sign recognition and retrieval using three datasets: ASL-Citizen [11], WLASL2000 [28], and NMFs-CSL [19]. The ASL-Citizen dataset includes 2,731 isolated ASL signs recorded from a webcam. The WLASL2000 dataset, sourced from the web, contains 2,000 common ASL signs and presents challenges due to its noisy nature and limited samples per sign. The NMFs-CSL dataset consists of 1,067 Chinese Sign Language (CSL) signs, which require recognition of non-manual cues, such as facial expressions, to accurately identify signs. + +# 7.1. Evaluation Protocol + +We follow standard evaluation protocols for sign recognition, measuring top-1 and top-5 per-instance and per-class accuracy on WLASL and NMFs-CSL. For ASL-Citizen, we use the benchmark metrics specified for this dataset, including discounted cumulative gain (DCG), mean reciprocal rank (MRR), and top-1 and top-5 instance accuracy. + +To evaluate our model as a frozen feature extractor, we apply a retrieval protocol based on the dictionary-based retrieval approach from [11]. We choose a retrieval approach over a linear evaluation protocol, as dictionary retrieval is a common and practical task in sign language processing. This method also offers a better evaluation of the generalization of the features to unseen datasets, which is particularly important given the evolving nature of sign languages. + +# 7.2. Experimental Setting + +Pretraining. We pretrain our model on the YouTube-SL-25 dataset [42], which consists of large-scale continuous sign videos from YouTube. We initialize the model with the base video Hiera-B architecture, pretrained on Kinetics using MAE. We randomly select 16 consecutive frames as inputs based on a co-articulated sign typically lasting for around 13 frames [1, 34, 47]. No changes are made to the original encoder architecture, while the decoder is replaced with our lightweight sign decoder. + +Recognition. For downstream tasks, our approach leverages only the pretrained encoder, discarding the decoder and entirely eliminating the reliance on keypoint extraction. For recognition, we extend the pretrained model's input sequence to 64 frames. We preserve the pretrained weights by inflating the patch embeddings, which preserves the computational efficiency by avoiding increasing the number of tokens in the transformer, detailed in the supplementary material. A linear classifier is applied to the pooled features from the encoder model to predict the sign classes. + +Retrieval. For dictionary retrieval evaluation, we follow the approach outlined in Sec. 6 using a stride of 2. For the label assignment, we compute the cosine similarity between the test video representation and those in the training dictionary. Each test query video is assigned the label of the training video that yields the highest similarity score. + +Additional implementation details are provided in the supplementary material. + +# 7.3. Evaluation on Sign Recognition + +In Tab. 1, our single-modality model demonstrates significant improvements over all existing single-modality approaches, achieving performance on par with complex multi-modal methods for WLASL2000. While ensembling with multi-modal approaches could further improve performance, we emphasize the importance of pretraining. Our Hiera encoder model with our pretraining, achieves these results without the extensive architectural modifications required by other methods. Instead of relying on additional modalities as input, our pretrained model has learned these features to create a robust sign representation model. Importantly, our pretraining strategy does not rely on any annotated sign data. + +The effectiveness of our SignRep model on the NMFs-CSL dataset is further illustrated in Tab. 2, where we achieve a top-1 accuracy of $84.1\%$ and a top-5 accuracy of $98.8\%$ , outperforming all other methods, including multimodal approaches. Notably, the closest single-modality comparison is the StepNet RGB model, which achieves $77.2\%$ in top-1 accuracy, nearly $7\%$ lower than our model. While StepNet focuses on developing specialized architecture for sign, we demonstrate our framework can effectively learn sign features from large-scale datasets. + +
MethodInstance Acc.Class Acc.
Top-1Top-5Top-1Top-5
Skeleton-based
ST-GCN [50]34.4066.5732.5365.45
SignBERT [17]39.4073.3536.7472.38
BEST [52]46.2579.3343.5277.65
SignBERT+[20]48.8582.4846.3781.33
Multi-modal
BEST (+R) [52]54.5988.0852.1287.28
SignBERT(+R) [17]54.6987.4952.0886.93
SignBERT+(+R) [20]55.5989.3753.3388.82
SAM (5ξ) [23]58.7391.4655.9390.94
SAM-v2 (5ξ) [22]59.3991.4856.6390.89
NLA-SLR [57]61.0591.4558.0590.70
NLA-SLR(3ξ) [57]61.2691.7758.3190.91
StepNet (R+F) [40]61.1791.9458.4391.43
RGB-based
I3D [8]32.4857.31--
I3D(BSL1K) [1]46.8279.3644.7278.47
StepNet [40]56.8988.6454.5487.97
SignRep (Ours)61.0590.2758.8989.44
+ +Table 1. Comparison of downstream sign recognition results on WLASL2000. $\xi$ , (R) and (F) denotes a multi-crop inference, RGB and Optical Flow modality respectively. + +
MethodTop-1Top-5
I3D [8]64.488.0
TSM [29]64.588.7
Slowfast [14]66.386.6
GLE-Net [19]69.088.1
HMA (◇) [18]75.695.3
StepNet [40]77.292.5
SignBERT (H+R) (◇) [17]78.497.3
BEST (◇) [52]79.297.1
NLA-SLR (◇) [57]83.498.3
StepNet (R+F) (◇) [40]83.697.0
NLA-SLR (◇,3ξ) [57]83.798.5
SignRep (Ours)84.198.8
+ +Table 2. Comparison of downstream sign recognition results on NMFs-CSL. $\diamond$ denotes methods using multi-modality/ensemble models and $\xi$ indicates a multi-crop inference. + +Our results on the ASL-Citizen dataset, shown in Tab. 3, further demonstrate the effectiveness of our model where it surpasses the previous baseline I3D model by $18\%$ in top-1 accuracy. + +
ModelDCGMRRRec@1Rec@5
ST-GCN* [50]76.3769.9759.5282.68
I3D* [8]79.1373.3263.1086.09
SignRep (Ours)90.8488.0581.3796.11
+ +Table 3. Comparison of downstream sign recognition results on ASL-Citizen. * denotes results produced by [11] + +# 7.4. Evaluation on Sign Dictionary Retrieval + +To assess the effectiveness of our model as a feature extractor, we utilize a dictionary-based sign retrieval method that directly evaluates the quality of the learned representations without any fine-tuning. In Tab. 4, we compare our approach to the Hiera model pretrained with MAE on both the Kinetics and YT-SL datasets. We apply the same training settings outlined in Sec. 7.2 for pretraining on YT-SL. Our results demonstrate that our proposed pretraining strategy significantly outperforms standard MAE, even when trained on YT-SL, by optimizing specifically for sign language representation, whereas MAE's pixel-reconstruction objective does not directly support learning the sign specific features needed for sign retrieval. + +We also compare our model's retrieval performance with the use of raw 3D keypoints and joint angles averaged over time. As shown in Tab. 4, our method achieves a substantial performance increase, with top-1 retrieval scores more than tripling those based on hand joint angles. These results underscore the strength of our pretraining approach in capturing spatio-temporal features specific to sign language, establishing our method as a highly effective framework for sign representation learning. We also validate the importance of incorporating hand activity awareness, observing that weighted averaging based on hand activity improves retrieval performance compared to averaging. + +# 7.5. Ablation Studies + +Impact of Pretraining on Sign Retrieval. In Tab. 5, we examine the effects of various elements in our pretraining strategy on retrieval performance in WLASL2000. Our results show that combining all three types of sign priors that have angle, keypoint and distance yields the highest retrieval accuracy, underscoring the benefit of capturing multiple aspects of the sign features. We also analyze the role of masking during pretraining, removing it leads to a noticeable decline in retrieval performance, indicating that masking is essential for robust representation learning. While we observe lower reconstruction loss without masking, this does not translate to better retrieval scores, suggesting that precise reconstruction alone does not necessarily produce effective sign representations. Finally, the addition of both adversarial loss and variance-covariance regularization yields the strongest retrieval results, confirming that + +these feature regularization techniques improve the quality and robustness of learned sign representations. + +Impact of Pretraining for Recognition. In Tab. 6, we demonstrate that our pretrained model substantially outperforms both the MAE model pretrained on Kinetics and the model pretrained on YT-SL with pixel MAE. Additionally, the baseline Hiera model pretrained on Kinetics, achieving a top-1 accuracy of $51.5\%$ , surpasses prior keypoint-based masked learning methods such as SignBERT, BEST and SignBERT+, as shown in Tab. 1. This result underscores the value of RGB input, revealing the limitations of using keypoints as a primary modality. Previous keypoint methods have required ensembling with RGB to remain competitive, illustrating the challenges of relying solely on skeletal information as an input modality for sign recognition. Finally, we observe that incorporating the sign class distribution loss further enhances recognition accuracy, improving top-1 accuracy from $59.9\%$ to $61.0\%$ with a $\kappa$ of 0.2. + +# 8. Feature Extractor for Sign Translation + +In this section, we evaluate SignRep as a feature extractor for sign translation by integrating it into two open-source translation models: Sign2GPT [49] on RWTH-PHOENIX-Weather 2014T (Phoenix14T) [6] and CSL-Daily [54], as well as the approach from [43] on How2Sign [13]. We chose these models because they allow direct evaluation of SignRep as a frozen feature extractor without requiring sign-text pretraining or significant modifications to the architecture. We follow the same translation training hyperparameters as the original papers to ensure a fair comparison and isolate the impact of replacing the visual backbone. + +For Phoenix14T and CSL-Daily, we omit the pseudogloss pretraining used in [49] to ensure a fair evaluation of SignRep's learned features without additional linguistic supervision. In [49], they show that the DinoV2 backbone requires fine-tuning with LoRA to perform well on sign translation. Instead, we replace the original learnable DinoV2 backbone with pre-extracted SignRep features, obtained using our sliding window approach with a stride of 2. This setup places ours at a disadvantage, as we remove the trainable backbone weights and data augmentation, relying solely on pre-extracted features. Despite this, SignRep achieves competitive translation performance, as shown in Tab. 7, performing comparably on Phoenix14T and achieving substantial gains on CSL-Daily. By replacing the learnable DinoV2 model with pre-extracted SignRep features during training, we significantly reduce computational costs, making large-scale translation training more efficient and accessible without sacrificing performance. Our approach removes an additional 5.5GFLOP per frame required by the learnable DinoV2 model, which is important as translation models process tens to hundreds of frames. + +
FeaturesASL-CitizenWLASL2000NMFs-CSL
DCGRec@1Rec@5DCGRec@1Rec@5DCGRec@1Rec@5
HieraMAE-Kinetic11.640.250.8413.212.083.4023.293.9612.18
HieraMAE-YTSL12.120.391.3414.062.574.4128.037.5718.38
All Joint Angles13.820.572.1717.922.546.5032.517.9325.50
All Keypoints14.290.982.7419.373.168.2336.6412.2630.99
Hand Keypoints25.917.9619.5828.117.5720.9241.7215.9142.62
Hand Joint Angle26.938.8121.4130.619.4224.3644.1718.1346.34
SignRep (avg)61.4037.4768.7753.5726.1660.9879.5158.4891.85
SignRep (weighted)71.2149.9580.0957.9329.9267.4183.0563.0495.63
+ +Table 4. Comparison of generalization results on sign retrieval with no downstream training applied to SignRep. + +
Prior Comp.Reg.DCG
anglekptdistmaskvar+covadv
✓✓45.1
✓✓32.9
✓✓47.2
✓✓✓✓48.5
✓✓✓46.3
✓✓✓✓✓49.9
✓✓✓✓✓✓50.7
+ +Table 5. Comparison of the impact of sign priors, masking and regularization on the impact of retrieval. Retrieval results are obtained on WLASL using a stride of 8. + +
MethodKInstance Acc.Class Acc.
top-1top-5top-1top-5
MAE (Kinetic)-51.583.448.482.0
MAE (YT-SL)-57.287.454.686.3
SignRep0.059.990.257.489.2
SignRep0.160.490.157.889.1
SignRep0.261.090.358.989.4
SignRep0.559.889.457.788.8
+ +Table 6. Comparisons of the impact of pretraining and weighting for the class distribution loss on WLASL. + +
Sign2GPT BackbonePhoenix14TCSL-Daily
B-4RB-4R
DinoV2(LoRA)19.4245.2312.9641.12
SignRep(extracted)20.3845.1716.3342.67
+ +This leads to an $80\%$ reduction in GPU memory usage during training, lowering requirements from 60GB to 11GB + +Table 7. Comparison of translation results using Sign2GPT [49] by replacing the learnable DinoV2 backbone with our extracted SignRep features. B-4 denotes BLEU4 and R denotes ROUGE. + +
Tarrés's BackboneHow2Sign
rBLEUBLEU
I3D(Supervised)2.218.03
SignRep(Self-Supervised)2.748.66
+ +Table 8. Comparison of translation results using Tarrés [43], replacing the I3D features with our SignRep features. + +with a batch size of 8. This reduction makes training on standard hardware more feasible, enabling broader accessibility for researchers working with sign translation models. + +For How2Sign, in Tab. 8, we replace the supervised I3D features with our self-supervised SignRep features on the translation framework from [43] and observe performance improvements. Unlike I3D features, which are pretrained with supervision on a labeled sign recognition data, SignRep learns features purely through self-supervision, highlighting our framework's effectiveness for learning sign-specific features without supervision. + +# 9. Conclusions + +In this work, we introduced a scalable, self-supervised framework for sign representation learning without labeled sign datasets. Using masked autoencoding with sign priors, adversarial loss and feature regularizations, our approach enhances generalization while eliminating the need for skeletal keypoints, multimodal or multi-branch architectures in downstream tasks. Our single model achieves state-of-the-art performance on multiple sign recognition benchmarks with a simpler, more efficient design. Beyond recognition, our representations demonstrate versatility in sign dictionary retrieval on unseen datasets and serve as an efficient feature extractor for existing sign translation systems, reducing computational costs while maintaining strong performance. Our findings highlight the potential of our self-supervised learning framework for scalable sign modeling and encourage further research into practical applications of sign representations. + +# Acknowledgements + +# References + +[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European conference on computer vision, pages 35–53. Springer, 2020. 1, 2, 6 +[2] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bencie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford british sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2 +[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 2 +[4] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 4 +[5] P Boyes Braem and RL Sutton-Spence. The Hands Are The Head of The Mouth. The Mouth as Articulator in Sign Languages. Hamburg: Signum Press, 2001. 1 +[6] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 7, 2 +[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1 +[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 6, 7 +[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. Electra: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations, 2020. 1 +[10] Konstantinos M Dafnis. Bidirectional skeleton-based isolated sign recognition using graph convolution networks. In LREC proceedings, 2022. 2 + +This work was supported by SNSF project 'SMILE II' (CRSII5 193686), European Union's Horizon2020 programme ('EASIER' grant agreement 101016982) and the Innosuisse IICT Flagship (PFFS-21-47). This work reflects only the authors view and the Commission is not responsible for any use that may be made of the information it contains. Neither Necati Cihan Camgoz nor Meta were involved in the model training, evaluation, or use of the datasets. The authors also thank Maksym Ivashechkin for their assistance in providing the 3D pose estimation model. +[11] Aashaka Desai, Lauren Berger, Fyodor Minakov, Nessa Milano, Chinmay Singh, Kriston Pumphrey, Richard Ladner, Hal Daume III, Alex X Lu, Naomi Caselli, et al. Asl citizen: a community-sourced dataset for advancing isolated sign language recognition. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 7 +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2 +[13] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 7, 2 +[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6 +[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 4 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2 +[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11087-11096, 2021. 1, 2, 6 +[18] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 6 +[19] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 1, 2, 5, 6 +[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11221-11239, 2023. 1, 2, 6 +[21] Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. Improving 3d pose estimation for sign language. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 1-5. IEEE, 2023. 2, 1, 3 +[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 1, 2, 6 +[23] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language + +recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3413-3423, 2021. 1, 2, 6 +[24] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3394-3402, 2021. 2 +[25] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In European Conference on Computer Vision, pages 349-367. Springer, 2025. 1, 2 +[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 1 +[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1 +[28] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 1, 2, 5 +[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 6 +[30] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 2 +[31] Amit Moryossef. Optimizing hand region detection in mediapipe holistic full-body pose estimation to improve accuracy and avoid downstream errors. arXiv preprint arXiv:2405.03545, 2024. 2 +[32] Amit Moryossef, Ioannis Tsochantaridis, Joe Dinn, Necati Cihan Camgoz, Richard Bowden, Tao Jiang, Annette Rios, Mathias Muller, and Sarah Ebling. Evaluating the immediate applicability of pose estimation for sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3434-3440, 2021. 2 +[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2 +[34] Tomas Pfister, James Charles, and Andrew Zisserman. Large-scale learning of sign language by watching tv. 2013. 6 +[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 + +[36] Charles Raude, KR Prajwal, Liliane Momeni, Hannah Bull, Samuel Albanie, Andrew Zisserman, and Gül Varol. A tale of two languages: Large-vocabulary continuous sign language recognition from spoken language supervision. arXiv preprint arXiv:2405.10266, 2024. 2 +[37] Phillip Rust, Bowen Shi, Skyler Wang, Necati Cihan Camgoz, and Jean Maillard. Towards privacy-aware sign language translation at scale. arXiv preprint arXiv:2402.09611, 2024. 1, 2 +[38] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 2, 3 +[39] Marc Schulder, Sam Bigeard, Maria Kopf, Thomas Hanke, Anna Kuder, Joanna Wojcicka, Johanna Mesch, Thomas Björkstrand, Anna Vacalopoulou, Kyriaki Vasilaki, et al. Signs and synonymity: Continuing development of the multilingual sign language wordnet. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, pages 343-353, 2024. 5 +[40] Xiaolong Shen, Zhedong Zheng, and Yi Yang. Step-net: Spatial-temporal part-aware network for isolated sign language recognition. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-19, 2024. 6 +[41] Ozge Mercanoglu Sincan and Hacer Yalim Keles. Autsl: A large scale multi-modal turkish sign language dataset and baseline methods. IEEE access, 8:181340-181355, 2020. 1 +[42] Garrett Tanzer and Biao Zhang. Youtube-sl-25: A large-scale, open-domain multilingual sign language parallel corpus. arXiv preprint arXiv:2407.11144, 2024. 6 +[43] Laia Tarres, Gerard I Gallego, Amanda Duarte, Jordi Torres, and Xavier Giró-i Nieto. Sign language translation from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5625-5635, 2023. 7, 8, 2 +[44] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 2 +[45] Dave Uthus, Garrett Tanzer, and Manfred Georg. Youtube: A large-scale, open-domain american sign language- english parallel corpus. Advances in Neural Information Processing Systems, 36, 2024. 2 +[46] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[47] Ville Viitaniemi, Tommi Jantunen, Leena Savolainen, Matti Karppa, and Jorma Laaksonen. S-pot-a benchmark in spotting signs within continuous signing. In LREC, pages 4-1, 2014. 6 +[48] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Learnt contrastive concept embeddings for sign recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1945-1954, 2023. 2 + +[49] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7, 8 +[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6, 7 +[51] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 4 +[52] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2, 6 +[53] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20871-20881, 2023. 1, 2 +[54] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316–1325, 2021. 7, 2 +[55] Wengang Zhou, Weichao Zhao, Hezhen Hu, Zecheng Li, and Houqiang Li. Scaling up multimodal pre-training for sign language understanding. arXiv preprint arXiv:2408.08544, 2024. 1, 2 +[56] Simone Zini, Alex Gomez-Villa, Marco Buzzelli, Bartlomiej Twardowski, Andrew D. Bagdanov, and Joost van de weijer. Planckian jitter: countering the color-crippling effects of color jitter on self-supervised training. In The Eleventh International Conference on Learning Representations, 2023. 1 +[57] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14890-14900, 2023. 1, 2, 6 \ No newline at end of file diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/images.zip b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab16a5e99d0518828198bace6bf1aabdf88a79a4 --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b14a45999dcffe8ee7f4f2845d25d67b7e737b15ee2e1d28a319762187e9a97 +size 470789 diff --git a/2025/SignRep_ Enhancing Self-Supervised Sign Representations/layout.json b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e870050b3bb8cb76d4bfca950516e6d7f3152d9e --- /dev/null +++ b/2025/SignRep_ Enhancing Self-Supervised Sign Representations/layout.json @@ -0,0 +1,9253 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 486, + 121 + ], + "type": "text", + "content": "SignRep: Enhancing Self-Supervised Sign Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "content": "Ryan Wong" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "content": ", Necati Cihan Camgoz" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "content": ", Richard Bowden" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "content": "University of Surrey, " + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 142, + 476, + 185 + ], + "type": "text", + "content": "Meta Reality Labs \n{r.wong, r.bowden}@surrey.ac.uk, neccam@meta.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 539 + ], + "type": "text", + "content": "Sign language representation learning presents unique challenges due to the complex spatio-temporal nature of signs and the scarcity of labeled datasets. Existing methods often rely either on models pre-trained on general visual tasks, that lack sign-specific features, or use complex multimodal and multi-branch architectures. To bridge this gap, we introduce a scalable, self-supervised framework for sign representation learning. We leverage important inductive (sign) priors during the training of our RGB model. To do this, we leverage simple but important cues based on skeletons while pretraining a masked autoencoder. These sign specific priors alongside feature regularization and an adversarial style agnostic loss provide a powerful backbone. Notably, our model does not require skeletal keypoints during inference, avoiding the limitations of keypoint-based models during downstream tasks. When finetuned, we achieve state-of-the-art performance for sign recognition on the WLASL, ASL-Citizen and NMFs-CSL datasets, using a simpler architecture and with only a single-modality. Beyond recognition, our frozen model excels in sign dictionary retrieval and sign translation, surpassing standard MAE pretraining and skeletal-based representations in retrieval. It also reduces computational costs for training existing sign translation models while maintaining strong performance on Phoenix2014T, CSL-Daily and How2Sign." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 561, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Sign language is an important means of communication for millions of people worldwide. Sign languages have complex visual characteristics, which include intricate hand shapes, motions, body poses and facial expressions that models need to accurately interpret and process [5]. Furthermore, the computational demands of processing long video sequences add considerable challenges, making it difficult to scale these systems effectively. As a result, current approaches to sign recognition and translation often rely on general pretrained vision models [1, 7, 19, 22, 28, 49, 57]. Creating robust, label-free sign language representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 553, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 238 + ], + "type": "text", + "content": "that generalize across diverse datasets is challenging, yet essential for scalable sign language modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 409 + ], + "type": "text", + "content": "Most existing methods for sign recognition rely on multimodal or an ensemble of specialized models to achieve state-of-the-art recognition results. These models often require multi-channel inputs (e.g. RGB, depth and skeleton data) or specialized architectures to capture the complex interactions of hand, body or facial expressions [17, 20, 22, 23, 52, 57]. With each country having its own sign language and linguistic study expensive, currently available labeled datasets are sparse with typically under 2000 unique signs [1, 19, 26, 28, 41]. Collecting annotated sign language data is costly and time-consuming, making it infeasible to rely solely on labeled datasets for sign representation learning. This highlights the need for methods that can learn sign language representations from large-scale, unlabeled data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 555 + ], + "type": "text", + "content": "Newer methods for pretraining sign language translation models often require sentence-aligned annotations [25, 37, 49, 53, 55]. While these prior approaches jointly learn individual sign representations and inter-sign relationships, they are not scalable to unlabeled datasets as they require sentence-aligned annotations. We therefore focus on building a scalable and generalizable individual sign representation framework using self-supervised learning. A strong foundation in individual sign embeddings is an important step before building effective inter-sign models, as it ensures robust feature representations before incorporating complex long range temporal information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Our main contributions are as follows: (1) We propose a scalable self-supervised Masked Autoencoding (MAE) framework which leverages sign priors, adversarial loss and feature regularizations for sign representation learning. (2) We introduce a method for analyzing sign class similarities in unseen datasets, introducing an auxiliary class probability distribution loss, which enhances recognition performance. (3) Our single pretrained model achieves state-of-the-art sign recognition, surpassing complex architecture and multimodal models. (4) We demonstrate the effectiveness of our sign representations for sign dictionary retrieval, achieving strong performance without any downstream training. (5) We demonstrate that using our pre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22804" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "trained model as a frozen feature extractor, training is more tractable by reducing memory requirements and improving the performance of existing sign translation models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 142, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 142, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 142, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 415 + ], + "type": "text", + "content": "Supervised Sign Recognition. Isolated Sign Recognition involves identifying a single sign within a given video. Sign recognition models often rely on pretrained spatiotemporal models from action recognition datasets such as Kinetics [8], fine-tuned for specific sign recognition tasks [1, 19, 22, 28, 57]. This transfer learning approach faces a domain shift problem, as sign videos have unique temporal dynamics and subtle gestures not well-represented in general action datasets, leading to performance degradation. To leverage invariance, some methods use skeletal keypoints instead of full-frame RGB models [10, 17, 20, 48, 52]. While keypoint-based models are more memory-efficient, they typically underperform relative to RGB models [2, 11, 36] and require complex architectural modifications to Graph Convolutional Networks (GCNs) [30] or transformers [46], to model the spatial-temporal relationships. Moreover, keypoints are prone to errors, such as missing or misdetected points [31, 32]. To solve these issues, recent state-of-the-art methods have combined keypoint and RGB modalities through branching or ensembling techniques [22, 23, 57]. While they achieve higher accuracy, they increase computational complexity due to the need for multiple models or support for multi-modal inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 594 + ], + "type": "text", + "content": "Self-Supervised Sign Representation Learning. Self-supervised learning for vision [3, 33, 38, 44] and language [12, 35] has demonstrated substantial benefits by leveraging unlabeled datasets for various tasks. In sign language, approaches such as Skeleton [24], SignBERT [17], SignBERT+ [20] and BEST [52] focus on keypoint-based self-supervised learning, employing masked learning similar to BERT [12]. However, these methods depend on keypoints as inputs and therefore requiring ensembling with RGB models to achieve competitive performance, resulting in increasing complexity and limited pretraining benefits. This highlights the need for a simple and efficient model which capturing sign-specific information and retains the pretraining advantages." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Paired Sign-Text Pretraining for Sign Translation. Supervised pretraining approaches for sign translation, such as GFSLT-VLP [53], Sign2GPT [49], SignHiera [37], MSLU [55] and VAP [25], use sign video-spoken language pairs to enhance sign translation. Sign2GPT employs a pretrained DinoV2 [33], which is effective for general vision tasks but requires LoRA [16] fine-tuning for sign translation to achieve strong performance, significantly increasing computational costs. This highlights the domain gap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "between pretrained foundation vision models and sign language. SignHiera [37] demonstrates success with large-scale training on YT-ASL [45] using Hiera MAE [38] with language-supervised pretraining. However, they acknowledge that SignHiera requires significant computational resources (64 A100 GPUs for two weeks), emphasizing the need for more accessible and cost-effective video pretraining on large-scale datasets. These methods also face scalability challenges due to their reliance on sentence-aligned annotations. In this paper, we introduce a scalable self-supervised pretraining framework that learns sign representations without paired sign-text annotations, aiming to significantly reduce computational costs for pretraining and improve efficiency in downstream sign tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 388, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 388, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 388, + 264 + ], + "type": "text", + "content": "3. Sign Priors" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 270, + 556, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 380 + ], + "type": "text", + "content": "Sign language communication relies on hand shapes, body posture and interactions. To guide the model toward meaningful sign representations, we introduce \"sign priors\", a set of cues that we know capture essential sign features. These priors " + }, + { + "bbox": [ + 313, + 270, + 556, + 380 + ], + "type": "inline_equation", + "content": "(\\mathcal{P})" + }, + { + "bbox": [ + 313, + 270, + 556, + 380 + ], + "type": "text", + "content": " serve as the primary targets for our model, as detailed in Sec. 5.1. Using a human pose estimation model specialized for sign language videos [21], we extract joint angles and 3D keypoints. We categorize our priors into keypoint, angle, distance and signer activity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 414 + ], + "type": "text", + "content": "Keypoint Priors define the spatial structure of signing, we divide this into two priors:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "text", + "content": "Hand Keypoint Prior " + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{h,k\\}})" + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "text", + "content": ": This prior is designed to capture hand shapes and orientations. We normalize the 3D coordinates of the 21 hand keypoints by setting the wrist as the origin, thereby eliminating positional variations in the body space. The resulting vector, " + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{h,k\\}} \\in \\mathbb{R}^{21 \\times 3}" + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "text", + "content": ", represents the configuration of the left or right hand, where " + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "inline_equation", + "content": "h \\in \\{\\mathrm{LH}, \\mathrm{RH}\\}" + }, + { + "bbox": [ + 313, + 414, + 554, + 497 + ], + "type": "text", + "content": " denotes the left or right hands." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "content": "Full Body Keypoint Prior " + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{b,k\\}})" + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "content": ". This prior captures the positioning of the body and hands within the body space. We use all 61 3D keypoints (21 for each hand and 19 for the body), resulting in " + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{b,k\\}} \\in \\mathbb{R}^{61 \\times 3}" + }, + { + "bbox": [ + 313, + 498, + 554, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 581 + ], + "type": "text", + "content": "Joint Angle Priors capture finger flexion, extension and encode upper-body posture variations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "text", + "content": "Hand Joint Angle Prior " + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{h,a\\}})" + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "text", + "content": ". This prior captures hand joint orientations. We extract 41 hand joint angles, yielding a vector " + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{41\\times 1}" + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "text", + "content": ". To handle the continuous nature of angles, we apply sine and cosine transformations, resulting in the final angle prior " + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{h,a\\}}\\in \\mathbb{R}^{41\\times 2}" + }, + { + "bbox": [ + 313, + 582, + 554, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "text", + "content": "Body Angle Pior " + }, + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{b,a\\}})" + }, + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "text", + "content": ". To capture the body's joint orientations, we extract the 22 body joint angles. These angles are transformed using sine and cosine functions to handle their continuous nature, resulting in the body angle prior " + }, + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{b,a\\}} \\in \\mathbb{R}^{22 \\times 2}" + }, + { + "bbox": [ + 313, + 641, + 554, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22805" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 228 + ], + "type": "image", + "image_path": "7a3373601e0de8e37928fe062b76cf893af3837e9d4ae6654b993722199aa6b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 281 + ], + "type": "text", + "content": "Figure 1. (Left): The pretraining process for SignRep, which leverages masked representation learning to predict sign priors such as hand keypoints and joint angles. This is achieved through a Hiera encoder and a lightweight sign decoder. The representation is further refined with regularization losses, including variance, covariance and adversarial style loss. (Right): An example setup for the discriminator to obtain a representation pair to predict a style-representation match." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 291, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 295, + 316 + ], + "type": "text", + "content": "Distance Priors capture fine-grained differences between signs and the interactions of hands in the signing space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "text", + "content": "Fingertip Distance Prior " + }, + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{h,d\\}})" + }, + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "text", + "content": ". Since fingertip interactions are a key component in sign language, we compute a distance matrix by measuring the distances from the fingertip keypoints to each of the knuckle, wrist and other fingertip keypoints. This produces the prior " + }, + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{h,d\\}} \\in \\mathbb{R}^{5 \\times 11 \\times 3}" + }, + { + "bbox": [ + 54, + 316, + 295, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "text", + "content": "Hand-Interaction Distance Prior " + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}^{\\{b,d\\}})" + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "text", + "content": ". Similar to the fingertip distance prior, we calculate a distance matrix using the wrist and five fingertip keypoints from each hand, resulting in a " + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "inline_equation", + "content": "6\\times 2" + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "text", + "content": " set of keypoints. The distance between the hands, body and facial keypoints is computed to capture the interactions between the hands and other body parts, resulting in the final hand-interaction prior " + }, + { + "bbox": [ + 54, + 377, + 295, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\{b,d\\}}\\in \\mathbb{R}^{12\\times 22\\times 3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "text", + "content": "Signer Activity. Some signs require only one hand or the signer may be in a resting pose. To enable the model to capture this information, we develop a simple heuristic. We define the active prior " + }, + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\{h,\\mathrm{act}\\}" + }, + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "text", + "content": ", where a hand is considered inactive if it remains below the middle of the stomach and has not moved over the course of a video clip. This results in a prior " + }, + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{P}\\{h,\\mathrm{act}\\} \\in [0,1]" + }, + { + "bbox": [ + 54, + 486, + 295, + 581 + ], + "type": "text", + "content": ", indicating whether each hand is active (1) or inactive (0)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 182, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 182, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 182, + 611 + ], + "type": "text", + "content": "4. SignRep Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "To learn the sign priors, we introduce a novel adaptation of the Hierarchical Vision Transformer (Hiera) [38], designed with a pretraining task specifically tailored for sign representation learning without requiring labeled sign data. We choose the Hiera model as it aligns with our objective of being efficient, simple and effective. It has demonstrated strong performance in masked representation learning while eliminating the need for complex architectures and special" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 339 + ], + "type": "text", + "content": "ized modules. We aim to develop a single spatio-temporal sign model that enhances performance without relying on multi-branch or ensemble methods, instead improving the representation learning process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": "The standard Hiera MAE framework processes a video clip represented as " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "V = \\{v_{1}, v_{2}, \\dots, v_{t}, \\dots, v_{T}\\}" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": " denotes the frame at time step " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": " is the total number of frames. The video is first divided into spatiotemporal patches. A masking strategy is then applied, with a subset of patches randomly masked according to a masking ratio " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": ". The unmasked patches are fed into the encoder " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": ", which processes them to generate an output representation " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{emb}} = \\{z_0^{\\mathrm{emb}}, z_1^{\\mathrm{emb}}, \\dots, z_K^{\\mathrm{emb}}\\}" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": " is the number of output tokens from the encoder. The representation " + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 313, + 340, + 554, + 518 + ], + "type": "text", + "content": " is then passed to a decoder, along with learnable masked tokens as input. The pretraining objective is to reconstruct the masked patches through pixel-level prediction, allowing the model to learn contextual relationships and enhance the visual representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 519, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 555 + ], + "type": "text", + "content": "Rather than pixel reconstruction, we use sign priors to learn meaningful sign representations, replacing the pixel reconstruction decoder with a lightweight sign decoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Sign Decoder " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "(f_{dec})" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ". As shown in Fig. 1, instead of passing " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " to the decoder, we take the average across " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " tokens, which is then processed through layer normalization followed by a fully connected layer. This produces an output representation " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{avg}} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " is the dimensionality of the representation. The representation " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " is then upsampled temporally to match the sequence length " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " of the input video. The upsampling module is implemented using a lightweight network consisting of a 1D convolution with a kernel size of 1, followed by a GELU activation function, and then a transpose convolution with a kernel size of " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " to match the number of input frames and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22806" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "an output dimension of " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "D^{\\prime}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ". This results in an upsampled vector " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{up}}\\in \\mathbb{R}^{T\\times D^{\\prime}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{up} = [z_1^{up},z_2^{up},\\dots ,z_T^{up}]" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 121 + ], + "type": "text", + "content": "We can then incorporate prediction heads for each sign prior to the network as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 129, + 295, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 129, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 145, + 129, + 295, + 144 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} ^ {\\mathcal {P}} = g _ {\\mathcal {P}} \\left(z _ {t} ^ {\\mathrm {u p}}\\right) \\tag {1}", + "image_path": "d97e8b9262de6ce0c8939c80ac7e16189c93789382002a59a5e829cf27ffa97e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": "where each sign prior, " + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\in \\{\\mathcal{P}\\{h,k\\}, \\mathcal{P}\\{h,a\\}, \\mathcal{P}\\{h,d\\}, \\mathcal{P}\\{b,k\\}, \\mathcal{P}\\{b,a\\}, \\mathcal{P}\\{b,d\\}\\}" + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": " has a corresponding fully connected layer " + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "inline_equation", + "content": "(g_{\\mathcal{P}})" + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": " with an associated output " + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": " which matches the flattened dimension of the corresponding sign prior. For the activity prior we simply use an MLP layer using " + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": " as input which produces an output of 2 to match the dimension of " + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "inline_equation", + "content": "[\\mathcal{P}\\{\\mathrm{LH,act}\\}, \\mathcal{P}\\{\\mathrm{RH,act}\\}]" + }, + { + "bbox": [ + 55, + 152, + 296, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 237, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 333 + ], + "type": "text", + "content": "The decoder is completely removed during downstream tasks and only the encoder is used as the sign representation model. As a result, the target sign priors are not required during downstream tasks, allowing the use of a single model that inherently captures sign knowledge. This approach ensures computational efficiency by leveraging masked representation learning while directly learning the essential sign priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 231, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 231, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 231, + 358 + ], + "type": "text", + "content": "5. Sign Representation Objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 364, + 205, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 205, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 205, + 376 + ], + "type": "text", + "content": "5.1. Sign Priors Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": "In the previous section, we identified the sign priors that our model needs to learn. Since the sign decoder applies upsampling to match the number of input frames, each frame's prior has an associated output prediction. For each predicted sign prior output, " + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": ", we train the model to regress to the corresponding target value " + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "y_t^{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": " from prior " + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": ", using smooth L1 loss (" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": ") across all sign priors for each frame in the input video sequence. The same loss is applied to the sign activity prior for consistency. We also mitigate the impact of low-quality keypoints in our objective function by only using keypoints with over " + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 382, + 296, + 537 + ], + "type": "text", + "content": " confidence from the pose estimation model and masking missing keypoints from the loss. The final reconstruction loss is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 545, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 545, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 295, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\sum_ {\\mathcal {P} \\in \\mathbf {P}} w _ {\\mathcal {P}} L 1 (\\hat {y} ^ {\\mathcal {P}}, y ^ {\\mathcal {P}}) \\tag {2}", + "image_path": "aa08c129f9ae720ead05a5e04c1f8bdfbd3d934971133b5103dad3060b44595f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "inline_equation", + "content": "w_{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "text", + "content": " is the loss weighting for the prior " + }, + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 579, + 247, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 222, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 222, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 222, + 613 + ], + "type": "text", + "content": "5.2. Representation Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "text", + "content": "Feature Regularization. Self-supervised methods have shown that regularizing features improves representation quality [4, 51]. To enhance our representations, we incorporate a variance and covariance loss into " + }, + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "The variance loss, " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": ", encourages diversity in the learned representations by spreading them across the representation space. To compute the variance loss, we first calculate the standard deviation for each feature dimension" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "type": "inline_equation", + "content": "(j)" + }, + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "type": "text", + "content": " across a batch " + }, + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "type": "inline_equation", + "content": "(N)" + }, + { + "bbox": [ + 314, + 72, + 408, + 84 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 92, + 553, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 92, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 365, + 92, + 553, + 131 + ], + "type": "interline_equation", + "content": "\\sigma_ {j} = \\sqrt {\\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left(z _ {i , j} ^ {\\mathrm {a v g}} - \\bar {z} _ {j} ^ {\\mathrm {a v g}}\\right) ^ {2}} \\tag {3}", + "image_path": "c1f6ac27066c3e3be44a23b2c4fdf259d43f59f13a1cda38060b2f0e85412915.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 139, + 555, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 555, + 175 + ], + "type": "text", + "content": "The diversity of the features are then encouraged through a hinge function to ensure that the variance does not fall below a threshold of one." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 182, + 553, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 182, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 378, + 182, + 553, + 216 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v a r}} = \\sum_ {j = 1} ^ {D} \\max (0, 1 - \\sigma_ {j}) \\tag {4}", + "image_path": "4818cde8e66e9edd851306aca1f0cd6b1efc97175ecd6dedae6e3d594725df1c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 224, + 555, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 271 + ], + "type": "text", + "content": "The covariance loss, " + }, + { + "bbox": [ + 313, + 224, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cov}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 271 + ], + "type": "text", + "content": ", reduces correlations among features, helping to avoid redundancy. To compute the covariance loss, we first need to calculate the covariance matrix such that:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 352, + 278, + 553, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 278, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 352, + 278, + 553, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {j, k} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(z _ {i, j} ^ {\\text {a v g}} - \\bar {z} _ {j} ^ {\\text {a v g}}\\right) \\left(z _ {i, k} ^ {\\text {a v g}} - \\bar {z} _ {k} ^ {\\text {a v g}}\\right) \\tag {5}", + "image_path": "ff5d747f5cc3cd0cb270ff82e5d0f31453340baa01c404cf0d63e4381f4fb7a5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{j,k}" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": " represents the covariance between feature dimensions " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "\\bar{z}_j^{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": " is the mean of the " + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 319, + 554, + 380 + ], + "type": "text", + "content": "-th feature across the batch. The loss then penalises off-diagonal values in the covariance matrix, encouraging different feature dimensions to be uncorrelated:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 401, + 387, + 553, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 387, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 401, + 387, + 553, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o v}} = \\sum_ {j \\neq k} \\mathcal {C} _ {j, k} ^ {2} \\tag {6}", + "image_path": "ec2ca0ff26beb1d726c0c4550ae0fb13cd1221dae58e36e0b4e89ee20064eaaa.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "content": "Style Agnostic Representations. We also aim to encourage the encoder to capture robust and generalizable sign features, while filtering out irrelevant details such as background and person-specific appearance. We can explicitly learn this by introducing a discriminator that evaluates whether two representations share the same \"style\", defined here by background and appearance features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": "To extract style information, we calculate the gram matrix, commonly used in image style transfer [15], from the " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": " tokens, averaging across the column dimension to produce the style representation " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{style}}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": ". We then pair " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{style}}" + }, + { + "bbox": [ + 313, + 510, + 554, + 582 + ], + "type": "text", + "content": " and pass them to the discriminator, which learns to output 1 for matching pairs and 0 for non-matching pairs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": "To generate these representations pairs, as shown in Fig. 1(right), we randomly crop a video sequence of length " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": " to create two segments, " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "A_{2}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": ", each of length " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": ". For each, we extract " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "z_{A_1}^{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "z_{A_1}^{\\mathrm{style}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "z_{A_2}^{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "z_{A_2}^{\\mathrm{style}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": ", assuming that they share background and appearance features due to originating from the same video. From a different video, we obtain segment " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": " with its style " + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "inline_equation", + "content": "z_B^{\\mathrm{style}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": ", which provides a contrasting background and appearance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "The discriminator, " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ", is then trained to produce 0 output for mismatched styles and 1 for matched styles, such that " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_B^{\\mathrm{style}}) = 0" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}(z_{A_1}^{\\mathrm{avg}}, z_{A_2}^{\\mathrm{style}}) = 1" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ". This can be used" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22807" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "to further enhance the generalization of the encoder, we add an adversarial loss during pretraining, designed to fool the discriminator which encourages the model to focus on sign-specific content over style-related features. The adversarial loss is formalized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 139, + 295, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 139, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 72, + 139, + 295, + 154 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p o s} ^ {\\left(A _ {1}, A _ {2}\\right)} = \\max \\left(0, \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {A _ {2}} ^ {\\text {s t y l e}}\\right) - \\mathbb {E} _ {q \\sim \\mathrm {U}} \\mathfrak {D} (q)\\right) \\tag {7}", + "image_path": "f89c2639fb05eca6d1fb5910db8d02a10a2baa00736e0a3699a877879d235094.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 156, + 294, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 156, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 156, + 294, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n e g}} ^ {(A _ {1}, B)} = \\max (0, \\mathbb {E} _ {q \\sim \\mathrm {M}} \\mathfrak {D} (q) - \\mathfrak {D} \\left(z _ {A _ {1}} ^ {\\text {a v g}}, z _ {B} ^ {\\text {s t y l e}}\\right)) \\tag {8}", + "image_path": "e083de19026ce1b132684ac2b5e1628171feef7e594f477936ab8329cc28b06c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 174, + 294, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 174, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 90, + 174, + 294, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} ^ {A _ {1}} = \\left(\\mathcal {L} _ {p o s} ^ {(A _ {1}, A _ {2})}\\right) ^ {2} + \\left(\\mathcal {L} _ {n e g} ^ {(A _ {1}, B)}\\right) ^ {2} \\tag {9}", + "image_path": "42251240da1b433c142c09f37cd0bb308580b42665484f93cd341c38306c071d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q)" + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "text", + "content": " is the expected discriminator output of style for representation pairs that come from the same video sequence and " + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)" + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "text", + "content": " for when they come from different video sequence. We apply this loss only if " + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{q\\sim \\mathrm{M}}\\mathfrak{D}(q) > \\mathbb{E}_{q\\sim \\mathrm{U}}\\mathfrak{D}(q)" + }, + { + "bbox": [ + 55, + 195, + 296, + 278 + ], + "type": "text", + "content": ", ensuring stability by restricting the adversarial loss to cases where the discriminator is likely to correctly identify the matches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 280, + 240, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 280, + 240, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 240, + 291 + ], + "type": "text", + "content": "Our final loss for pretraining is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 299, + 294, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 299, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 299, + 294, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {r e c o n}} + w _ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}} + w _ {\\text {c o v}} \\mathcal {L} _ {\\text {c o v}} + w _ {\\text {a d v}} \\mathcal {L} _ {\\text {a d v}} \\tag {10}", + "image_path": "9821afaeaaec6cbaf6eb0934697dc86ef1815d2a26cd19194f350a20d5f5acf2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{cov}}" + }, + { + "bbox": [ + 55, + 320, + 296, + 343 + ], + "type": "text", + "content": " are the respective weighting factors for the adversarial, variance and covariance losses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 354, + 279, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 279, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 279, + 369 + ], + "type": "text", + "content": "6. Representations for Dictionary Retrieval" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "text", + "content": "Sign dictionary retrieval is the task of identifying signs from a predefined database. This allows for quick lookup of related signs given a query sign which supports efficient dataset creation and promotes cross-linguistic understanding of signs [39]. The SignRep model can be used as a feature extractor for dictionary retrieval. Given a query sign video of length " + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "text", + "content": ", we apply a sliding window to extract segment representations " + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{\\mathrm{avg}} = \\{z_1^{\\mathrm{avg}}, z_2^{\\mathrm{avg}}, \\ldots, z_n^{\\mathrm{avg}}, \\ldots, z_N^{\\mathrm{avg}}\\}" + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "inline_equation", + "content": "N = \\left\\lfloor \\frac{L - T}{\\text{stride}} \\right\\rfloor + 1" + }, + { + "bbox": [ + 55, + 373, + 296, + 506 + ], + "type": "text", + "content": ". Since isolated signs typically begin and end in a resting pose, we use the output from" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 512, + 294, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 512, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 294, + 528 + ], + "type": "interline_equation", + "content": "\\gamma_ {n} ^ {\\mathrm {a c t}} = \\max \\left(\\mathcal {P} ^ {\\{L H, \\mathrm {a c t} \\}}, \\mathcal {P} ^ {\\{R H, \\mathrm {a c t} \\}}\\right) \\tag {11}", + "image_path": "ee191fe2945847df618c64c9baed249310fa296bdfde26684e9cded22a7df4de.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "type": "text", + "content": "to identify whether the signer is active. By checking hand activity, we compute a weighted average of the segment representations resulting in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 121, + 578, + 294, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 121, + 578, + 294, + 610 + ], + "type": "interline_equation", + "content": "z ^ {\\text {o u t}} = \\frac {1}{\\sum \\gamma_ {n} ^ {\\text {a c t}}} \\sum_ {n = 1} ^ {N} \\gamma_ {n} ^ {\\text {a c t}} z _ {n} ^ {\\text {a v g}} \\tag {12}", + "image_path": "35bf8353e90705d85ffccdc52400a32174847b54f5000d41425c37bf41779807.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "z^{\\mathrm{out}}" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " is the representation for the query video. Finally, we extract features for all dictionary videos and classify a query video by finding the closest match using cosine similarity of the normalized representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Class Probability Distribution. The retrieval approach enables the recognition of visually similar sign classes without additional training, which is useful for improving the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "accuracy of downstream tasks such as recognition. We construct a class distribution by computing the cosine similarity across all samples in the training dataset, capturing the visual proximity between them. From the resulting cosine similarity matrix, we compute inter-class similarities using the sign-sample labels to generate a matrix of shape " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " represents the number of classes. Applying a temperature-scaled softmax to each row of this matrix yields a probability distribution for each class relative to the others, forming a class probability distribution matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", where each element of " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " lies in the range [0, 1]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": "We can incorporate the class probability distribution " + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": " as a regularization term in the downstream recognition task using KL divergence. Given a sign video with target label " + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": ", we can compute the KL divergence between the predicted class distribution " + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": " (from the output classifier layer) and " + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": " which is the pre-computed class probability distribution of " + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 204, + 554, + 287 + ], + "type": "text", + "content": ". This results in:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 382, + 298, + 553, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 298, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 382, + 298, + 553, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\phi} = \\kappa \\phi_ {c} [ c ] \\mathrm {K L} \\left(\\phi_ {c} \\| \\bar {y}\\right) \\tag {13}", + "image_path": "63d57a51d2cf27113fa61013e4da486a6fc4b192a9a004eb6edd1911660948b1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": " is the weighting factor for the loss and " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\phi_c\\parallel \\bar{y})" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": " is the KL divergence. To emphasize the contribution of the target class distribution, we scale the KL divergence by " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\phi_c[c]" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": ", which is the value of the class probability for " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 380, + 395, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 395, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 395, + 392 + ], + "type": "text", + "content": "7. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 400, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 520 + ], + "type": "text", + "content": "We evaluate our model on sign recognition and retrieval using three datasets: ASL-Citizen [11], WLASL2000 [28], and NMFs-CSL [19]. The ASL-Citizen dataset includes 2,731 isolated ASL signs recorded from a webcam. The WLASL2000 dataset, sourced from the web, contains 2,000 common ASL signs and presents challenges due to its noisy nature and limited samples per sign. The NMFs-CSL dataset consists of 1,067 Chinese Sign Language (CSL) signs, which require recognition of non-manual cues, such as facial expressions, to accurately identify signs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 528, + 430, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 430, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 430, + 540 + ], + "type": "text", + "content": "7.1. Evaluation Protocol" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 617 + ], + "type": "text", + "content": "We follow standard evaluation protocols for sign recognition, measuring top-1 and top-5 per-instance and per-class accuracy on WLASL and NMFs-CSL. For ASL-Citizen, we use the benchmark metrics specified for this dataset, including discounted cumulative gain (DCG), mean reciprocal rank (MRR), and top-1 and top-5 instance accuracy." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "To evaluate our model as a frozen feature extractor, we apply a retrieval protocol based on the dictionary-based retrieval approach from [11]. We choose a retrieval approach over a linear evaluation protocol, as dictionary retrieval is a common and practical task in sign language processing. This method also offers a better evaluation of the generalization of the features to unseen datasets, which is particularly important given the evolving nature of sign languages." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22808" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 179, + 85 + ], + "type": "text", + "content": "7.2. Experimental Setting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 197 + ], + "type": "text", + "content": "Pretraining. We pretrain our model on the YouTube-SL-25 dataset [42], which consists of large-scale continuous sign videos from YouTube. We initialize the model with the base video Hiera-B architecture, pretrained on Kinetics using MAE. We randomly select 16 consecutive frames as inputs based on a co-articulated sign typically lasting for around 13 frames [1, 34, 47]. No changes are made to the original encoder architecture, while the decoder is replaced with our lightweight sign decoder." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 316 + ], + "type": "text", + "content": "Recognition. For downstream tasks, our approach leverages only the pretrained encoder, discarding the decoder and entirely eliminating the reliance on keypoint extraction. For recognition, we extend the pretrained model's input sequence to 64 frames. We preserve the pretrained weights by inflating the patch embeddings, which preserves the computational efficiency by avoiding increasing the number of tokens in the transformer, detailed in the supplementary material. A linear classifier is applied to the pooled features from the encoder model to predict the sign classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 296, + 388 + ], + "type": "text", + "content": "Retrieval. For dictionary retrieval evaluation, we follow the approach outlined in Sec. 6 using a stride of 2. For the label assignment, we compute the cosine similarity between the test video representation and those in the training dictionary. Each test query video is assigned the label of the training video that yields the highest similarity score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 413 + ], + "type": "text", + "content": "Additional implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 227, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 227, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 227, + 434 + ], + "type": "text", + "content": "7.3. Evaluation on Sign Recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": "In Tab. 1, our single-modality model demonstrates significant improvements over all existing single-modality approaches, achieving performance on par with complex multi-modal methods for WLASL2000. While ensembling with multi-modal approaches could further improve performance, we emphasize the importance of pretraining. Our Hiera encoder model with our pretraining, achieves these results without the extensive architectural modifications required by other methods. Instead of relying on additional modalities as input, our pretrained model has learned these features to create a robust sign representation model. Importantly, our pretraining strategy does not rely on any annotated sign data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "The effectiveness of our SignRep model on the NMFs-CSL dataset is further illustrated in Tab. 2, where we achieve a top-1 accuracy of " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "84.1\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " and a top-5 accuracy of " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "98.8\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": ", outperforming all other methods, including multimodal approaches. Notably, the closest single-modality comparison is the StepNet RGB model, which achieves " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " in top-1 accuracy, nearly " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " lower than our model. While StepNet focuses on developing specialized architecture for sign, we demonstrate our framework can effectively learn sign features from large-scale datasets." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 560, + 369 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 560, + 369 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 560, + 369 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 560, + 369 + ], + "type": "table", + "html": "
MethodInstance Acc.Class Acc.
Top-1Top-5Top-1Top-5
Skeleton-based
ST-GCN [50]34.4066.5732.5365.45
SignBERT [17]39.4073.3536.7472.38
BEST [52]46.2579.3343.5277.65
SignBERT+[20]48.8582.4846.3781.33
Multi-modal
BEST (+R) [52]54.5988.0852.1287.28
SignBERT(+R) [17]54.6987.4952.0886.93
SignBERT+(+R) [20]55.5989.3753.3388.82
SAM (5ξ) [23]58.7391.4655.9390.94
SAM-v2 (5ξ) [22]59.3991.4856.6390.89
NLA-SLR [57]61.0591.4558.0590.70
NLA-SLR(3ξ) [57]61.2691.7758.3190.91
StepNet (R+F) [40]61.1791.9458.4391.43
RGB-based
I3D [8]32.4857.31--
I3D(BSL1K) [1]46.8279.3644.7278.47
StepNet [40]56.8988.6454.5487.97
SignRep (Ours)61.0590.2758.8989.44
", + "image_path": "734691222fc8cf65e1546c7f828e7c5f7ade0810bcf32ba82201fb4414517533.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 339, + 425, + 531, + 599 + ], + "blocks": [ + { + "bbox": [ + 313, + 377, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 411 + ], + "type": "text", + "content": "Table 1. Comparison of downstream sign recognition results on WLASL2000. " + }, + { + "bbox": [ + 313, + 377, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 377, + 555, + 411 + ], + "type": "text", + "content": ", (R) and (F) denotes a multi-crop inference, RGB and Optical Flow modality respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 425, + 531, + 599 + ], + "lines": [ + { + "bbox": [ + 339, + 425, + 531, + 599 + ], + "spans": [ + { + "bbox": [ + 339, + 425, + 531, + 599 + ], + "type": "table", + "html": "
MethodTop-1Top-5
I3D [8]64.488.0
TSM [29]64.588.7
Slowfast [14]66.386.6
GLE-Net [19]69.088.1
HMA (◇) [18]75.695.3
StepNet [40]77.292.5
SignBERT (H+R) (◇) [17]78.497.3
BEST (◇) [52]79.297.1
NLA-SLR (◇) [57]83.498.3
StepNet (R+F) (◇) [40]83.697.0
NLA-SLR (◇,3ξ) [57]83.798.5
SignRep (Ours)84.198.8
", + "image_path": "b2afc547537b682fb836f4d430a312f227034fc77c8cd81792fd724786c2e5dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "type": "text", + "content": "Table 2. Comparison of downstream sign recognition results on NMFs-CSL. " + }, + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "type": "text", + "content": " denotes methods using multi-modality/ensemble models and " + }, + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 607, + 555, + 641 + ], + "type": "text", + "content": " indicates a multi-crop inference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Our results on the ASL-Citizen dataset, shown in Tab. 3, further demonstrate the effectiveness of our model where it surpasses the previous baseline I3D model by " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " in top-1 accuracy." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22809" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 136 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 136 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 136 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 136 + ], + "type": "table", + "html": "
ModelDCGMRRRec@1Rec@5
ST-GCN* [50]76.3769.9759.5282.68
I3D* [8]79.1373.3263.1086.09
SignRep (Ours)90.8488.0581.3796.11
", + "image_path": "ec294680473b8b8f5fdfa5f56d1490a0c345e7e586e7dc5a5bd153c8ccd62012.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 168 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 168 + ], + "type": "text", + "content": "Table 3. Comparison of downstream sign recognition results on ASL-Citizen. * denotes results produced by [11]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 180, + 266, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 266, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 266, + 194 + ], + "type": "text", + "content": "7.4. Evaluation on Sign Dictionary Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 198, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 295, + 353 + ], + "type": "text", + "content": "To assess the effectiveness of our model as a feature extractor, we utilize a dictionary-based sign retrieval method that directly evaluates the quality of the learned representations without any fine-tuning. In Tab. 4, we compare our approach to the Hiera model pretrained with MAE on both the Kinetics and YT-SL datasets. We apply the same training settings outlined in Sec. 7.2 for pretraining on YT-SL. Our results demonstrate that our proposed pretraining strategy significantly outperforms standard MAE, even when trained on YT-SL, by optimizing specifically for sign language representation, whereas MAE's pixel-reconstruction objective does not directly support learning the sign specific features needed for sign retrieval." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 498 + ], + "type": "text", + "content": "We also compare our model's retrieval performance with the use of raw 3D keypoints and joint angles averaged over time. As shown in Tab. 4, our method achieves a substantial performance increase, with top-1 retrieval scores more than tripling those based on hand joint angles. These results underscore the strength of our pretraining approach in capturing spatio-temporal features specific to sign language, establishing our method as a highly effective framework for sign representation learning. We also validate the importance of incorporating hand activity awareness, observing that weighted averaging based on hand activity improves retrieval performance compared to averaging." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 156, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 156, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 156, + 517 + ], + "type": "text", + "content": "7.5. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "Impact of Pretraining on Sign Retrieval. In Tab. 5, we examine the effects of various elements in our pretraining strategy on retrieval performance in WLASL2000. Our results show that combining all three types of sign priors that have angle, keypoint and distance yields the highest retrieval accuracy, underscoring the benefit of capturing multiple aspects of the sign features. We also analyze the role of masking during pretraining, removing it leads to a noticeable decline in retrieval performance, indicating that masking is essential for robust representation learning. While we observe lower reconstruction loss without masking, this does not translate to better retrieval scores, suggesting that precise reconstruction alone does not necessarily produce effective sign representations. Finally, the addition of both adversarial loss and variance-covariance regularization yields the strongest retrieval results, confirming that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "these feature regularization techniques improve the quality and robustness of learned sign representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "content": "Impact of Pretraining for Recognition. In Tab. 6, we demonstrate that our pretrained model substantially outperforms both the MAE model pretrained on Kinetics and the model pretrained on YT-SL with pixel MAE. Additionally, the baseline Hiera model pretrained on Kinetics, achieving a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "inline_equation", + "content": "51.5\\%" + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "content": ", surpasses prior keypoint-based masked learning methods such as SignBERT, BEST and SignBERT+, as shown in Tab. 1. This result underscores the value of RGB input, revealing the limitations of using keypoints as a primary modality. Previous keypoint methods have required ensembling with RGB to remain competitive, illustrating the challenges of relying solely on skeletal information as an input modality for sign recognition. Finally, we observe that incorporating the sign class distribution loss further enhances recognition accuracy, improving top-1 accuracy from " + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "inline_equation", + "content": "59.9\\%" + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "inline_equation", + "content": "61.0\\%" + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 110, + 555, + 301 + ], + "type": "text", + "content": " of 0.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 311, + 528, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 528, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 528, + 324 + ], + "type": "text", + "content": "8. Feature Extractor for Sign Translation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 462 + ], + "type": "text", + "content": "In this section, we evaluate SignRep as a feature extractor for sign translation by integrating it into two open-source translation models: Sign2GPT [49] on RWTH-PHOENIX-Weather 2014T (Phoenix14T) [6] and CSL-Daily [54], as well as the approach from [43] on How2Sign [13]. We chose these models because they allow direct evaluation of SignRep as a frozen feature extractor without requiring sign-text pretraining or significant modifications to the architecture. We follow the same translation training hyperparameters as the original papers to ensure a fair comparison and isolate the impact of replacing the visual backbone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 463, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 556, + 713 + ], + "type": "text", + "content": "For Phoenix14T and CSL-Daily, we omit the pseudogloss pretraining used in [49] to ensure a fair evaluation of SignRep's learned features without additional linguistic supervision. In [49], they show that the DinoV2 backbone requires fine-tuning with LoRA to perform well on sign translation. Instead, we replace the original learnable DinoV2 backbone with pre-extracted SignRep features, obtained using our sliding window approach with a stride of 2. This setup places ours at a disadvantage, as we remove the trainable backbone weights and data augmentation, relying solely on pre-extracted features. Despite this, SignRep achieves competitive translation performance, as shown in Tab. 7, performing comparably on Phoenix14T and achieving substantial gains on CSL-Daily. By replacing the learnable DinoV2 model with pre-extracted SignRep features during training, we significantly reduce computational costs, making large-scale translation training more efficient and accessible without sacrificing performance. Our approach removes an additional 5.5GFLOP per frame required by the learnable DinoV2 model, which is important as translation models process tens to hundreds of frames." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22810" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 531, + 213 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "type": "table", + "html": "
FeaturesASL-CitizenWLASL2000NMFs-CSL
DCGRec@1Rec@5DCGRec@1Rec@5DCGRec@1Rec@5
HieraMAE-Kinetic11.640.250.8413.212.083.4023.293.9612.18
HieraMAE-YTSL12.120.391.3414.062.574.4128.037.5718.38
All Joint Angles13.820.572.1717.922.546.5032.517.9325.50
All Keypoints14.290.982.7419.373.168.2336.6412.2630.99
Hand Keypoints25.917.9619.5828.117.5720.9241.7215.9142.62
Hand Joint Angle26.938.8121.4130.619.4224.3644.1718.1346.34
SignRep (avg)61.4037.4768.7753.5726.1660.9879.5158.4891.85
SignRep (weighted)71.2149.9580.0957.9329.9267.4183.0563.0495.63
", + "image_path": "9baa74f6fc929f202a9ae3c9bb90268c70658ffc30f0dacfb3ed5cca97a47d11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 242, + 288, + 367 + ], + "blocks": [ + { + "bbox": [ + 102, + 220, + 509, + 233 + ], + "lines": [ + { + "bbox": [ + 102, + 220, + 509, + 233 + ], + "spans": [ + { + "bbox": [ + 102, + 220, + 509, + 233 + ], + "type": "text", + "content": "Table 4. Comparison of generalization results on sign retrieval with no downstream training applied to SignRep." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 242, + 288, + 367 + ], + "lines": [ + { + "bbox": [ + 63, + 242, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 63, + 242, + 288, + 367 + ], + "type": "table", + "html": "
Prior Comp.Reg.DCG
anglekptdistmaskvar+covadv
✓✓45.1
✓✓32.9
✓✓47.2
✓✓✓✓48.5
✓✓✓46.3
✓✓✓✓✓49.9
✓✓✓✓✓✓50.7
", + "image_path": "4b2a6ad407adcc7d73567b258bf164334c012b73d25c74260915ae204cc52803.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 416, + 294, + 530 + ], + "blocks": [ + { + "bbox": [ + 55, + 375, + 295, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 409 + ], + "type": "text", + "content": "Table 5. Comparison of the impact of sign priors, masking and regularization on the impact of retrieval. Retrieval results are obtained on WLASL using a stride of 8." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 416, + 294, + 530 + ], + "lines": [ + { + "bbox": [ + 58, + 416, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 58, + 416, + 294, + 530 + ], + "type": "table", + "html": "
MethodKInstance Acc.Class Acc.
top-1top-5top-1top-5
MAE (Kinetic)-51.583.448.482.0
MAE (YT-SL)-57.287.454.686.3
SignRep0.059.990.257.489.2
SignRep0.160.490.157.889.1
SignRep0.261.090.358.989.4
SignRep0.559.889.457.788.8
", + "image_path": "65d7d061e00e8c8279b0dd96c09c340ecfcf1e479a62d08bcac4b75ea9e79b30.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 569, + 297, + 635 + ], + "blocks": [ + { + "bbox": [ + 55, + 538, + 295, + 561 + ], + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 561 + ], + "type": "text", + "content": "Table 6. Comparisons of the impact of pretraining and weighting for the class distribution loss on WLASL." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 569, + 297, + 635 + ], + "lines": [ + { + "bbox": [ + 56, + 569, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 297, + 635 + ], + "type": "table", + "html": "
Sign2GPT BackbonePhoenix14TCSL-Daily
B-4RB-4R
DinoV2(LoRA)19.4245.2312.9641.12
SignRep(extracted)20.3845.1716.3342.67
", + "image_path": "6c7c707af80015b91dbc02f164a28e574c17746e50a4f98c2a90064575456950.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "This leads to an " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " reduction in GPU memory usage during training, lowering requirements from 60GB to 11GB" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 335, + 241, + 535, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "type": "text", + "content": "Table 7. Comparison of translation results using Sign2GPT [49] by replacing the learnable DinoV2 backbone with our extracted SignRep features. B-4 denotes BLEU4 and R denotes ROUGE." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 241, + 535, + 308 + ], + "lines": [ + { + "bbox": [ + 335, + 241, + 535, + 308 + ], + "spans": [ + { + "bbox": [ + 335, + 241, + 535, + 308 + ], + "type": "table", + "html": "
Tarrés's BackboneHow2Sign
rBLEUBLEU
I3D(Supervised)2.218.03
SignRep(Self-Supervised)2.748.66
", + "image_path": "3ea10f46a16ff6a18b5dc8771badca425b4229293a24d25408c864e354c35059.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 338 + ], + "type": "text", + "content": "Table 8. Comparison of translation results using Tarrés [43], replacing the I3D features with our SignRep features." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": "with a batch size of 8. This reduction makes training on standard hardware more feasible, enabling broader accessibility for researchers working with sign translation models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 384, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 479 + ], + "type": "text", + "content": "For How2Sign, in Tab. 8, we replace the supervised I3D features with our self-supervised SignRep features on the translation framework from [43] and observe performance improvements. Unlike I3D features, which are pretrained with supervision on a labeled sign recognition data, SignRep learns features purely through self-supervision, highlighting our framework's effectiveness for learning sign-specific features without supervision." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 490, + 392, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 392, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 392, + 502 + ], + "type": "text", + "content": "9. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduced a scalable, self-supervised framework for sign representation learning without labeled sign datasets. Using masked autoencoding with sign priors, adversarial loss and feature regularizations, our approach enhances generalization while eliminating the need for skeletal keypoints, multimodal or multi-branch architectures in downstream tasks. Our single model achieves state-of-the-art performance on multiple sign recognition benchmarks with a simpler, more efficient design. Beyond recognition, our representations demonstrate versatility in sign dictionary retrieval on unseen datasets and serve as an efficient feature extractor for existing sign translation systems, reducing computational costs while maintaining strong performance. Our findings highlight the potential of our self-supervised learning framework for scalable sign modeling and encourage further research into practical applications of sign representations." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22811" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 220, + 115, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 233 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 240, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "text", + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Triantafyllos Afouras, Joon Son Chung, Neil Fox, and Andrew Zisserman. Bsl-1k: Scaling up co-articulated sign language recognition using mouthing cues. In European conference on computer vision, pages 35–53. Springer, 2020. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 296, + 350 + ], + "type": "text", + "content": "[2] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bencie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford british sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 351, + 296, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 351, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 61, + 351, + 296, + 416 + ], + "type": "text", + "content": "[3] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 417, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 417, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 417, + 296, + 459 + ], + "type": "text", + "content": "[4] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 493 + ], + "type": "text", + "content": "[5] P Boyes Braem and RL Sutton-Spence. The Hands Are The Head of The Mouth. The Mouth as Articulator in Sign Languages. Hamburg: Signum Press, 2001. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 494, + 296, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 494, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 62, + 494, + 296, + 537 + ], + "type": "text", + "content": "[6] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 7, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 538, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 296, + 592 + ], + "type": "text", + "content": "[7] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 593, + 296, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 636 + ], + "type": "text", + "content": "[8] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 637, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 296, + 680 + ], + "type": "text", + "content": "[9] Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. Electra: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[10] Konstantinos M Dafnis. Bidirectional skeleton-based isolated sign recognition using graph convolution networks. In LREC proceedings, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 90, + 297, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 201 + ], + "type": "text", + "content": "This work was supported by SNSF project 'SMILE II' (CRSII5 193686), European Union's Horizon2020 programme ('EASIER' grant agreement 101016982) and the Innosuisse IICT Flagship (PFFS-21-47). This work reflects only the authors view and the Commission is not responsible for any use that may be made of the information it contains. Neither Necati Cihan Camgoz nor Meta were involved in the model training, evaluation, or use of the datasets. The authors also thank Maksym Ivashechkin for their assistance in providing the 3D pose estimation model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[11] Aashaka Desai, Lauren Berger, Fyodor Minakov, Nessa Milano, Chinmay Singh, Kriston Pumphrey, Richard Ladner, Hal Daume III, Alex X Lu, Naomi Caselli, et al. Asl citizen: a community-sourced dataset for advancing isolated sign language recognition. Advances in Neural Information Processing Systems, 36, 2024. 2, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 554, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 193 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 194, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 554, + 259 + ], + "type": "text", + "content": "[13] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 7, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 554, + 304 + ], + "type": "text", + "content": "[14] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 348 + ], + "type": "text", + "content": "[15] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "text", + "content": "[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 554, + 448 + ], + "type": "text", + "content": "[17] Hezhen Hu, Weichao Zhao, Wengang Zhou, Yuechen Wang, and Houqiang Li. Signbert: Pre-training of hand-model-aware representation for sign language recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11087-11096, 2021. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "text", + "content": "[18] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 536 + ], + "type": "text", + "content": "[19] Hezhen Hu, Wengang Zhou, Junfu Pu, and Houqiang Li. Global-local enhancement network for nmf-aware sign language recognition. ACM transactions on multimedia computing, communications, and applications (TOMM), 17(3): 1-19, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 537, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 591 + ], + "type": "text", + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. Signbert+: Hand-model-aware self-supervised pretraining for sign language understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):11221-11239, 2023. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 554, + 646 + ], + "type": "text", + "content": "[21] Maksym Ivashechkin, Oscar Mendez, and Richard Bowden. Improving 3d pose estimation for sign language. In 2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), pages 1-5. IEEE, 2023. 2, 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[22] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Sign language recognition via skeleton-aware multi-model ensemble. arXiv preprint arXiv:2110.06161, 2021. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[23] Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. Skeleton aware multi-modal sign language" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22812" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3413-3423, 2021. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[24] Tao Jiang, Necati Cihan Camgoz, and Richard Bowden. Skeleton: Skeletal transformers for robust body-pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3394-3402, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 205 + ], + "type": "text", + "content": "[25] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In European Conference on Computer Vision, pages 349-367. Springer, 2025. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "text", + "content": "[26] Hamid Reza Vaezi Joze and Oscar Koller. Ms-asl: A large-scale data set and benchmark for understanding american sign language. arXiv preprint arXiv:1812.01053, 2018. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 261 + ], + "type": "text", + "content": "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "text", + "content": "[28] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1459-1469, 2020. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 360 + ], + "type": "text", + "content": "[29] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7083-7093, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "text", + "content": "[30] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 143-152, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "type": "text", + "content": "[31] Amit Moryossef. Optimizing hand region detection in mediapipe holistic full-body pose estimation to improve accuracy and avoid downstream errors. arXiv preprint arXiv:2405.03545, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "type": "text", + "content": "[32] Amit Moryossef, Ioannis Tsochantaridis, Joe Dinn, Necati Cihan Camgoz, Richard Bowden, Tao Jiang, Annette Rios, Mathias Muller, and Sarah Ebling. Evaluating the immediate applicability of pose estimation for sign language recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3434-3440, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 646 + ], + "type": "text", + "content": "[33] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 679 + ], + "type": "text", + "content": "[34] Tomas Pfister, James Charles, and Andrew Zisserman. Large-scale learning of sign language by watching tv. 2013. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 714 + ], + "type": "text", + "content": "[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[36] Charles Raude, KR Prajwal, Liliane Momeni, Hannah Bull, Samuel Albanie, Andrew Zisserman, and Gül Varol. A tale of two languages: Large-vocabulary continuous sign language recognition from spoken language supervision. arXiv preprint arXiv:2405.10266, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[37] Phillip Rust, Bowen Shi, Skyler Wang, Necati Cihan Camgoz, and Jean Maillard. Towards privacy-aware sign language translation at scale. arXiv preprint arXiv:2402.09611, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 239 + ], + "type": "text", + "content": "[38] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 329 + ], + "type": "text", + "content": "[39] Marc Schulder, Sam Bigeard, Maria Kopf, Thomas Hanke, Anna Kuder, Joanna Wojcicka, Johanna Mesch, Thomas Björkstrand, Anna Vacalopoulou, Kyriaki Vasilaki, et al. Signs and synonymity: Continuing development of the multilingual sign language wordnet. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, pages 343-353, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "text", + "content": "[40] Xiaolong Shen, Zhedong Zheng, and Yi Yang. Step-net: Spatial-temporal part-aware network for isolated sign language recognition. ACM Transactions on Multimedia Computing, Communications and Applications, 20(7):1-19, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 386, + 554, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 554, + 419 + ], + "type": "text", + "content": "[41] Ozge Mercanoglu Sincan and Hacer Yalim Keles. Autsl: A large scale multi-modal turkish sign language dataset and baseline methods. IEEE access, 8:181340-181355, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 420, + 554, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 554, + 454 + ], + "type": "text", + "content": "[42] Garrett Tanzer and Biao Zhang. Youtube-sl-25: A large-scale, open-domain multilingual sign language parallel corpus. arXiv preprint arXiv:2407.11144, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "type": "text", + "content": "[43] Laia Tarres, Gerard I Gallego, Amanda Duarte, Jordi Torres, and Xavier Giró-i Nieto. Sign language translation from instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5625-5635, 2023. 7, 8, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 554, + 555 + ], + "type": "text", + "content": "[44] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[45] Dave Uthus, Garrett Tanzer, and Manfred Georg. Youtube: A large-scale, open-domain american sign language- english parallel corpus. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 623 + ], + "type": "text", + "content": "[46] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[47] Ville Viitaniemi, Tommi Jantunen, Leena Savolainen, Matti Karppa, and Jorma Laaksonen. S-pot-a benchmark in spotting signs within continuous signing. In LREC, pages 4-1, 2014. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 714 + ], + "type": "text", + "content": "[48] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Learnt contrastive concept embeddings for sign recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1945-1954, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22813" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 530 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[49] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 150 + ], + "type": "text", + "content": "[50] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 152, + 295, + 196 + ], + "type": "text", + "content": "[51] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pages 12310-12320. PMLR, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 198, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 296, + 251 + ], + "type": "text", + "content": "[52] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. Best: Bert pre-training for sign language recognition with coupling tokenization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3597-3605, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 253, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 253, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 253, + 295, + 317 + ], + "type": "text", + "content": "[53] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20871-20881, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 319, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 294, + 374 + ], + "type": "text", + "content": "[54] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316–1325, 2021. 7, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 294, + 418 + ], + "type": "text", + "content": "[55] Wengang Zhou, Weichao Zhao, Hezhen Hu, Zecheng Li, and Houqiang Li. Scaling up multimodal pre-training for sign language understanding. arXiv preprint arXiv:2408.08544, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 420, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 420, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 420, + 294, + 483 + ], + "type": "text", + "content": "[56] Simone Zini, Alex Gomez-Villa, Marco Buzzelli, Bartlomiej Twardowski, Andrew D. Bagdanov, and Joost van de weijer. Planckian jitter: countering the color-crippling effects of color jitter on self-supervised training. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 294, + 530 + ], + "type": "text", + "content": "[57] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14890-14900, 2023. 1, 2, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22814" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_content_list.json b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b697735e2ad33a8b936f0e01182056c971f83e4 --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_content_list.json @@ -0,0 +1,1368 @@ +[ + { + "type": "text", + "text": "Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ronglai Zuo, Rolandos Alexandros Potamias, Evangelos Ververas, Jiankang Deng, Stefanos Zafeiriou \nImperial College London", + "bbox": [ + 101, + 180, + 903, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://2000zrl.github.io/soke/", + "bbox": [ + 361, + 219, + 637, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d814b5c7743d60a34a033c47a1f9811edfb6a98285646d2e4d7a0bccb8355abd.jpg", + "image_caption": [ + "Figure 1. We propose Signs as Tokens (SOKE), a unified sign language generator that can produce multilingual signs from text inputs. (Left: American sign language; Middle: Chinese sign language; Right: German sign language.)" + ], + "image_footnote": [], + "bbox": [ + 112, + 255, + 883, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 489, + 326, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language is a visual language that encompasses all linguistic features of natural languages and serves as the primary communication method for the deaf and hard-of-hearing communities. Although many studies have successfully adapted pretrained language models (LMs) for sign language translation (sign-to-text), the reverse task—sign language generation (text-to-sign)—remains largely unexplored. In this work, we introduce a multilingual sign language model, Signs as Tokens (SOKE), which can generate 3D sign avatars autoregressively from text inputs using a pretrained LM. To align sign language with the LM, we leverage a decoupled tokenizer that discretizes continuous signs into token sequences representing various body parts. During decoding, unlike existing approaches that flatten all part-wise tokens into a single sequence and predict one token at a time, we propose a multi-head decoding method capable of predicting multiple tokens simultaneously. This approach improves inference efficiency while maintaining effective information fusion across different body parts. To further ease the generation process, we propose a retrieval-enhanced SLG approach, which incorporates external sign dictionaries to provide accurate word-level signs as auxiliary conditions, significantly improving the precision of generated signs. Extensive qualitative and quantitative evaluations demonstrate the effectiveness of SOKE.", + "bbox": [ + 86, + 523, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 489, + 643, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign language is the primary communication method utilized by the deaf and hard-of-hearing communities, embodying all linguistic properties found in spoken languages, such as discrete semantic units and grammatical structure [16, 71]. As an attempt to mitigate the barrier between sign and spoken languages, several methods have been developed to tackle sign language processing by unfolding the task into two research problems: sign language translation (SLT, sign-to-text) [20, 33, 56, 64, 67, 69, 80, 83] and sign language generation (SLG, text-to-sign) [2, 47, 53, 72, 88].", + "bbox": [ + 511, + 517, + 906, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast with SLT, the usage of language models (LMs) in SLG remains relatively unexplored. Most existing methods approach SLG as a visual content generation task (e.g., video, keypoint, or motion), using GANs [50, 53] or diffusion models [2, 12, 44, 58, 59]. However, these works neglect the linguistic nature of sign languages, leading to suboptimal performance and missing benefits of pretrained LMs such as generalizability and scalability [18, 29, 60]. In particular, sign languages share the fundamental linguistic properties and the discrete structure of spoken languages [16, 71], which necessitates processing using autoregressive language modeling techniques [70].", + "bbox": [ + 511, + 671, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To formulate SLG in the framework of language modeling, a crucial step is mapping signs into discrete tokens. A common approach to discretizing sign languages involves", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23806", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "using glosses—the written form of signs—as an intermediate representation [13, 22, 50, 88]. However, glosses require extensive annotation efforts and impose a predefined information bottleneck, which fail to fully capture the rich semantics of sign languages [39]. To address these limitations and better capture the multi-cue property of sign languages [26, 81, 85], we leverage a vector quantized-variational auto-encoder (VQ-VAE)-based tokenizer [37, 65, 68] that learns mappings between continuous sign motions and discrete tokens for each body part (upper body, left hand, and right hand). During decoding, existing autoregressive motion generation methods [37, 68] typically flatten tokens from different body parts into a single sequence, resulting in inefficient inference due to tripled decoding steps. To overcome this challenge, we propose a novel multi-head decoding method capable of predicting multiple tokens at a time while effectively fusing information from different body parts in real time. Furthermore, given the limited vocabulary size of continuous sign language datasets, we enhance the model's generalization by instantiating the LM as a multilingual LM [35], enabling it to handle a broader and more diverse domain of discourse. To support multilingual SLG in a unified framework, we train the proposed model on a curated multilingual sign language dataset encompassing American, Chinese, and German sign languages [3, 10, 82].", + "bbox": [ + 89, + 90, + 480, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, retrieval-augmented generation (RAG) [4, 11, 62] has proven effective in enhancing the interpretability and accuracy of content generated by large language models. Similarly, several SLG methods [50, 88] leverage glosses to retrieve sign dictionaries and directly incorporate retrieved signs into the outputs, improving the interpretability of generated signs. However, beyond the inherent limitations of glosses, these methods often face challenges in handling co-articulations between adjacent signs. This is because dictionary signs are typically recorded at the word level, whereas SLG models operate at the sentence level, leading to potential unnaturalness in the generated sequences. To address these issues, we propose a retrieval-enhanced SLG approach that utilizes motion tokens from retrieved dictionary signs as additional conditions for language modeling. This approach avoids the unnaturalness of directly copying signs into model outputs while still benefiting from the precision of dictionary signs. To sum up, the contributions of this paper can be summarized as:", + "bbox": [ + 89, + 472, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel SLG approach, Signs as Tokens (SOKE). With the aid of a pretrained LM, SOKE can handle multiple sign languages in a unified model.", + "- We propose a multi-head decoding method that reduces inference latency and efficiently fuses information from different body parts, enabling seamless integration of a decoupled tokenizer with the LM.", + "- By leveraging external sign dictionaries, we introduce a retrieval-enhanced SLG approach to improve the preci" + ], + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion of generated signs.", + "bbox": [ + 527, + 92, + 684, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We curate two sign language datasets [3, 82] with accurate SMPL-X [42] poses to facilitate multilingual SLG research. Experiments demonstrate that our SOKE achieves state-of-the-art performance on three challenging benchmarks, How2Sign [10], CSL-Daily [82], and Phoenix2014T [3], using a single unified model.", + "bbox": [ + 513, + 106, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 210, + 653, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign Language Generation. Early gloss-based methods [21, 22, 50, 57, 65, 73, 88] approach SLG under a text-to-gloss-to-sign framework. Several works [1, 47, 49] propose to directly regress 2D joints to represent sign poses, simplifying the modeling process, however, they suffer from mode collapse due to the limited scale of data. Recent state-of-the-art SLG works can be categorized into two classes: the first group of methods [2, 12] employs diffusion models to generate sign motions conditioned on text inputs; the second group of methods [9, 70, 72] adopts a tokenizer-LM two-stage autoregressive generation approach. Among these, [72] incorporates costly gloss and HamNoSys notations [1] to train a semantic adapter, while [9] focuses on generating signs from a single word, limiting their applicability. Moreover, a notable gap exists in the literature regarding unified models for multilingual SLG. In this work, we address these gaps by utilizing a decoupled tokenizer and a pretrained multilingual LM, integrated with our customized multi-head decoding method and retrieval-enhanced approach, to enable text-driven multilingual SLG. Text-to-Motion Generation. Text-to-motion generation (T2MG) is an important research area with potential applications in human-computer interaction and robotics. Recent T2MG methods can be categorized into two main types: 1) Diffusion-based approaches [5, 51, 75], which utilize latent diffusion [46] to generate motions conditioned on text inputs. 2) Autoregressive approaches [17, 24, 37, 74, 76, 84], which draw inspiration from text generation methods by first tokenizing motion sequences and then using LMs to perform autoregressive generation. When decoding with a decoupled tokenizer [37, 68], these autoregressive approaches simply flatten tokens from various body parts into one sequence, causing additional decoding steps. In this work, we propose a novel multi-head decoding method that can efficiently predict motion tokens for all body parts at one single step, significantly reducing inference latency.", + "bbox": [ + 511, + 236, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieval-Enhanced Generation. Using the concept of retrieval to enhance generation models has proven effective across various domains [4, 8, 11, 15, 62, 66]. As a pioneering effort, Piktus et al. [31] propose to retrieve relevant documents before generating answers for knowledge-intensive NLP tasks. FLARE [25] introduces selective retrieval, activating it only when low-confidence tokens appear in the generation process. ToG [55] leverages retrieved", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23807", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a38ba151181b6e68a215bb785e08ce9f460b432080e8340715bac1ed9c7182d5.jpg", + "image_caption": [ + "Figure 2. An overview of our proposed method, Signs as Tokens (SOKE). We begin by training a VQ-VAE-based decoupled tokenizer to map continuous sign motions into discrete tokens for various body parts (upper body, left hand, and right hand). These sign motion tokens are then integrated into the text vocabulary of a pretrained language model, which serves as the backbone of our autoregressive multilingual generator (AMG). Given a text input, the AMG first retrieves word-level signs from external dictionaries, appends their motion tokens to the text tokens, and feeds them into the language model encoder. During decoding, our novel multi-head decoding strategy generates motion tokens for all body parts simultaneously at each time step. Finally, the derived motion tokens are used to reconstruct sign avatars." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "knowledge to enhance the reasoning capabilities of large language models. In the field of image generation, retrieval has also been utilized to enable diffusion models to generate out-of-distribution images [52]. In the context of SLG, several methods [50, 88] adopt gloss-based retrieval to obtain dictionary signs and approach SLG in a retrieve-and-concatenate manner. However, glosses have inherent limitations, and such methods often suffer from unrealistic transitions between signs. In this paper, we propose a simple yet effective retrieval-enhanced SLG approach that uses word-level signs as auxiliary conditions, addressing these limitations and improving the overall quality of generations.", + "bbox": [ + 88, + 358, + 485, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 554, + 225, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method is shown in Figure 2. Our approach mainly consists of two parts: a decoupled tokenizer (DETO) and an autoregressive multilingual generator (AMG). In the first stage, DETO learns mappings between continuous sign motions and discrete tokens for each body part (Section 3.2). In the second stage, AMG is trained to autoregressively generate motion tokens with the proposed multi-head decoding strategy (Section 3.3) conditioned on text inputs and retrieved dictionary signs (Section 3.4).", + "bbox": [ + 88, + 580, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Preparation", + "text_level": 1, + "bbox": [ + 89, + 727, + 261, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although existing SLG works [2, 72] have made efforts in curating high-quality SMPL-X poses for an American sign language (ASL) dataset, How2Sign [10], there is a notable lack of 3D annotations for other continuous sign language datasets, impeding multilingual SLG research. In this work, we rely on a two-stage approach, leveraging state-of-the-art 3D pose estimation methods [34, 43], to extract high-fidelity SMPL-X poses for CSL-Daily [82] and Phoenix-2014T [3], two widely-adopted datasets of Chinese and German sign languages (CSL and DGS), respec", + "bbox": [ + 88, + 750, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tively. Specifically, we initially estimate a set of rough 3D body poses for each signer using OSX [34] along with accurate 3D hand poses from WiLoR [43]. Given that OSX often falls short in accurately capturing the arm poses, we follow [2] and refine the upper body joint rotations by minimizing a re-projection loss between the estimated joints and the detected 2D joint keypoints from Mediapipe [38]. Using this process we can ensure accurate whole-body annotations. We represent each sign motion sequence as $\\mathbf{S} \\in \\mathbb{R}^{T \\times d}$ , where $T$ is the sequence length and $d = 133$ denotes the number of SMPL-X parameters, including 11 upper body joints, 30 hand joints, and 10 expression parameters. Please check the supplementary materials for more details and quantitative evaluations of our pose fitting pipeline.", + "bbox": [ + 511, + 358, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Decoupled Tokenizer", + "text_level": 1, + "bbox": [ + 511, + 584, + 712, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sign languages exhibit a multi-cue property [19, 26, 77, 79, 81, 86], where the semantics of a sign are conveyed simultaneously through body and hand movements. Inspired by this, we leverage a decoupled tokenizer [37, 65, 68], comprising three VQ-VAEs [61], to independently model key regions: the upper body and both hands. An illustration is available in the supplementary materials.", + "bbox": [ + 511, + 606, + 905, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a $T$ -frame sign motion input $\\mathbf{S} \\in \\mathbb{R}^{T \\times d}$ , we first decompose it into three part-wise motion sequences based on the kinematic tree of SMPL-X: $\\mathbf{S}^p \\in \\mathbb{R}^{T \\times d_p}$ , where $p \\in \\{B, LH, RH\\}$ . Moreover, we build three distinct VQ-VAEs, where each of them consists of an encoder $\\mathcal{E}_p(\\cdot)$ , a decoder $\\mathcal{D}_p(\\cdot)$ , and a learnable codebook $\\mathbf{Z}^p \\in \\mathbb{R}^{N_Z^p \\times C}$ where $N_Z^p$ represents the number of codes and $C$ denotes the code dimension. For each motion sequence, the corresponding encoder first projects the sequence into a latent space: $\\mathbf{S}_f^p = \\mathcal{E}_p(\\mathbf{S}^p) = \\{s_{f,i}^p\\}_{i=1}^{T_f} \\in \\mathbb{R}^{T_f \\times C}$ , using a stack of 1D-CNN layers. Then, for each pose we can derive a set of discrete tokens $\\hat{\\mathbf{Z}}^p = \\{\\hat{z}_i^p\\}_{i=1}^{T_f}$ using a quantizer $Q(\\cdot)$ ,", + "bbox": [ + 511, + 713, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23808", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c523adcc341bf8773ce91323acf95e297bb6a697d2e6b77176915f82d8d6b61.jpg", + "image_caption": [ + "(a) Sequential Decoding $(O(3K))$" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 359, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0e7520f580bbb11ae4a354f4c4508f563b4f5f500a7bf4858ceeb6c221e29224.jpg", + "image_caption": [ + "(b) Parallel Decoding $(O(K))$", + "Figure 3. Comparison of various decoding methods. (a) Sequential decoding flattens all motion tokens into a single sequence, requiring an inefficient $3K$ decoding steps, where $K$ is the target motion sequence length. (b) Parallel decoding improves efficiency by decoding body parts in parallel but lacks information fusion, which may lead to suboptimal performance. (c) Our proposed multi-head decoding balances efficiency and fusion by predicting three tokens simultaneously via different heads, with input embeddings as a weighted average of part-wise token embeddings." + ], + "image_footnote": [], + "bbox": [ + 364, + 87, + 586, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51b151d1ffd557f7f08d9bbe831e08c7caaf1bb81f20b19c98f4696767510ba7.jpg", + "image_caption": [ + "(c) Multi-head Decoding (ours, $O(K)$ )" + ], + "image_footnote": [], + "bbox": [ + 586, + 89, + 905, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which searches the nearest neighbor from the codebook $\\mathbf{Z}^p$ :", + "bbox": [ + 89, + 332, + 482, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {i} ^ {p} = Q \\left(s _ {f, i} ^ {p}\\right) = \\arg \\min _ {z _ {j} \\in \\mathbf {Z} ^ {p}} \\| s _ {f, i} ^ {p} - z _ {j} \\| _ {2}, j \\in [ 1, N _ {Z} ^ {p} ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 363, + 482, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then feed the obtained token sequence to the corresponding part-decoder $\\mathcal{D}_p$ to reconstruct the input motions: $\\hat{\\mathbf{S}}^p = \\mathcal{D}_p(\\mathbf{Z}^p)$ . Following [24], the objective function of each VQ-VAE is defined as: $\\mathcal{L}_{vq}^p = \\mathcal{L}_{rec}^p +\\mathcal{L}_{emb}^p +\\mathcal{L}_{com}^p$ where $\\mathcal{L}_{rec}^p = \\| \\hat{\\mathbf{S}}^p -\\mathbf{S}^p\\| _2^2$ denotes the reconstruction loss, $\\mathcal{L}_{emb}^p$ is the embedding loss, and $\\mathcal{L}_{com}^p$ is the commitment loss, respectively.", + "bbox": [ + 89, + 407, + 483, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Autoregressive Multilingual Generator", + "text_level": 1, + "bbox": [ + 89, + 535, + 424, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using obtained discrete sign representations, we can respect the discrete properties of sign languages and approach SLG similarly to text generation with a pretrained LM [35].", + "bbox": [ + 89, + 560, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Integrated Vocabulary. We first gather the discrete tokens for different body parts to construct a part-wise motion vocabulary: $\\mathcal{V}_m^p = \\{\\langle p\\_ i\\rangle \\}_{i = 1}^{N_Z^p}$ , e.g., $\\langle B\\_ 1\\rangle$ and $\\langle \\mathrm{RH\\_2}\\rangle$ . The overall motion vocabulary is a combination of three part-wise vocabularies: $\\mathcal{V}_m = \\{\\mathcal{V}_m^B,\\mathcal{V}_m^{LH},\\mathcal{V}_m^{RH}\\}$ . Following [35], we further define a series of language identifier tokens: $\\mathcal{V}_l = \\{\\langle \\mathrm{ASL}\\rangle ,\\langle \\mathrm{CSL}\\rangle ,\\langle \\mathrm{DGS}\\rangle \\}$ , to prompt the LM with the information of target sign language. The integrated vocabulary for the generator is defined as: $\\mathcal{V} = \\{\\mathcal{V}_t,\\mathcal{V}_m,\\mathcal{V}_l\\}$ , which lays a foundation for the subsequent LM fine-tuning.", + "bbox": [ + 89, + 607, + 482, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequential Decoding. Following previous motion generation works [37, 68], a naive decoding method is simply flattening tokens from each body part into a single sequence with tripled length: $\\mathbf{Y} = \\{y_1^B, y_1^{LH}, y_1^{RH}, \\ldots, y_K^B, y_K^{LH}, y_K^{RH}\\}$ , where $y_k^p \\in \\mathcal{V}_m^p$ denotes the part-wise motion token, and $K$ denotes the length of the target sign motion sequence. The decoding process can then be formulated as:", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathbf {Y} \\mid \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {2} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {k} ^ {B}, y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {k} ^ {B, L H}, y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 329, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_{en}$ denotes the last encoder hidden states. For simplicity, we write $y_{$ when the model is required to generate upper-body motions for ASL. The decoding process can then be formulated as:", + "bbox": [ + 511, + 647, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = P (\\mathbf {Y} ^ {B}) P (\\mathbf {Y} ^ {L H}) P (\\mathbf {Y} ^ {R H}) \\tag {3} \\\\ = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B} \\mid y _ {< k} ^ {B}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {L H} \\mid y _ {< k} ^ {L H}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {R H} \\mid y _ {< k} ^ {R H}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 741, + 906, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparing Eq. 3 with Eq. 2, the parallel decoding strategy significantly reduces the dependence on previous tokens, as the decoding processes for each body part are completely separated. While this formulation enables efficient parallel execution, the overly independent assumption may result in suboptimal performance.", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23809", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1f2990349f912a1833ba2b86aa4ec4f89581f242e528731e8ba4dc3f38480cb2.jpg", + "image_caption": [ + "Figure 4. Illustration of retrieval-enhanced SLG. We utilize external sign dictionaries to retrieve accurate word-level signs, which serve as additional conditions to enhance the generation quality." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Head Decoding. To balance efficiency and information fusion across body parts, we propose a multi-head decoding strategy. As shown in Figure 3(c), we design three language modeling heads, implemented as fully connected layers, to predict motion tokens for each body part simultaneously at each step. Moreover, the decoder inputs at each step are modified to a weighted average of the token embeddings from different body parts. Specifically, denoting the token embeddings from upper body, left hand, and right hand as $\\mathbf{E}^B$ , $\\mathbf{E}^{LH}$ , and $\\mathbf{E}^{RH}$ , respectively, the input embeddings are defined as: $\\mathbf{E} = (1 - 2\\lambda)\\mathbf{E}^B + \\lambda \\mathbf{E}^{LH} + \\lambda \\mathbf{E}^{RH}$ where $\\lambda \\in (0,0.5)$ is a hyper-parameter to control the weight of hand embeddings. The decoding process can then be formulated as:", + "bbox": [ + 89, + 290, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {4} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 512, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which can be viewed as relaxing Eq. 2 under a conditional independence assumption.", + "bbox": [ + 89, + 609, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and Inference. The LM is trained with a standard cross-entropy loss: $\\mathcal{L}_{LM} = -\\log P(\\mathbf{Y}|\\mathbf{h_{en}})$ . During inference, we adopt a simple greedy decoding algorithm that the LM always outputs the token with the highest probability at each step. For parallel and multi-head decoding, the inference process terminates as soon as any head or decoder predicts an end-of-sentence token. The obtained token sequences will be fed into the corresponding part-decoder, $\\mathcal{D}_p(\\cdot)$ , to reconstruct sign motions.", + "bbox": [ + 89, + 640, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Retrieval-Enhanced SLG", + "text_level": 1, + "bbox": [ + 89, + 787, + 318, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the effectiveness of retrieval-augmented generation in NLP [4, 11, 62], we propose to leverage external sign dictionaries to improve the fidelity of generated signs.", + "bbox": [ + 89, + 809, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 4, we first build a word-level sign dictionary for each target sign language by utilizing resources from isolated sign language recognition datasets", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[28, 30, 32, 78, 87] and online sources. Through our pose fitting pipeline, each RGB video in these dictionaries is converted into SMPL-X poses. Next, we employ a well-trained decoupled tokenizer to map the poses into discrete tokens. The resulting dictionary is represented as a set of quadruples $\\{(w, m^B, m^{LH}, m^{RH})\\}$ , where $w$ denotes lemmatized words in the written language, and $m^p$ denotes motion tokens. If a word has multiple sign instances in the dictionary, we retain only the instance with the lowest reconstruction error after passing through the tokenizer.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a text input $\\mathbf{X} = \\{x_{i}\\}_{i = 1}^{L}$ , consisting of $L$ words $^1$ , we collect all words and corresponding tokens found in the dictionary. The final prompt is created by concatenating raw texts $\\mathbf{X}$ with the retrieved motion tokens $\\{m_j^p\\}$ , which is then fed into the LM encoder to generate hidden states $\\mathbf{h}_{en}$ . Several examples are shown in Figure 4.", + "bbox": [ + 511, + 242, + 903, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 349, + 645, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We construct a multilingual sign language dataset by merging three widely-adopted datasets: How2Sign [10], CSL-Daily [82], and Phoenix-2014T [3]. How2Sign is a large-scale ASL dataset, consisting of 35K video-text pairs. CSL-Daily is the largest existing continuous CSL dataset, containing 20K samples. Phoenix-2014T is a DGS dataset with 8K samples. To train SOKE, we use the SMPL-X poses of How2Sign dataset as provided in [2] along with the curated poses of CSL-Daily and Phoenix-2014T.", + "bbox": [ + 511, + 375, + 905, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since the length of generated signs may differ from the ground truth, we employ the well-established dynamic time warping (DTW) [1, 2] on both procrustes-aligned and original joint position errors (JPE) to measure sequence-level distances between the generated signs and ground truth. Following [2], we also report backtranslation (B-T) scores in BLEU-4 [41] to assess the interpretability of the generated signs.", + "bbox": [ + 511, + 511, + 905, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For DETO, we empirically set the code numbers to $N_Z^B = 96$ for the body and $N_Z^{LH} = N_Z^{RH} = 192$ for the left and right hands, respectively, with a code dimension of $C = 512$ . We train DETO on the combined multilingual sign language dataset and sign dictionaries with a batch size of 320 per GPU for 500 epochs, using the AdamW optimizer [36] and a cosine learning rate scheduler starting at 2e-4. For the LM, we use mBART-large-cc25 [35], which features 12 encoder-decoder layers with a model dimension of 1024 and is widely used in various sign language translation methods [6, 7, 27, 63, 89]. We empirically set $\\lambda = 1/3$ to balance part-wise token embeddings. The LM is fine-tuned with a batch size of 32 per GPU on the multilingual sign language dataset for 150 epochs, employing the same optimizer settings as DETO. All models are trained using 6 RTX 3090 GPUs.", + "bbox": [ + 511, + 633, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 Since Chinese sentences lack spaces between words, we use Jieba [14,", + "bbox": [ + 531, + 886, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23810", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82e67d16ec45b1a5e39c12388e13c2b566ac884f601d1fe8ab4dee46247fdcb9.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons of generated signs between our proposed method, SOKE, with the SOTA method, S-MotionGPT [24], on the test sets of How2Sign (left), CSL-Daily (middle), and Phoenix-2014T (right)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 753, + 460, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. We first reimplement a state-of-the-art (SOTA) open-sourced motion generation approach, MotionGPT [24], which utilizes a VQ-VAE to discretize whole-body motions and employs a monolingual language model [45], on sign language datasets. The resulting model, S-MotionGPT, achieves competitive results on sign", + "bbox": [ + 89, + 777, + 483, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "language benchmarks (Table 1). In Figure 5, we perform a qualitative comparison between our SOKE and S-MotionGPT. As can be easily seen, our proposed method produces significantly better results, with visual quality comparable to the ground truth sign videos across the ASL, CSL, and DGS datasets. Notably, our method can generate precise and informative hand movements, which play a key role in conveying semantics in sign languages. Additionally, the proposed method supports three sign languages", + "bbox": [ + 511, + 755, + 906, + 892 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "40] for word segmentation.", + "bbox": [ + 91, + 887, + 236, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23811", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2583ad48cfe47d4a6ef5da83c0ce5bea51938e7662706a4a6612539ba9d47381.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMulti-lingualHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑
BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4
NAR [23]×13.9411.80--5.75----------
Prog. Trans.* [47]×14.1511.5714.7430.172.7515.9812.9116.3032.633.0713.6711.9515.0131.774.94
Text2Mesh* [54]×13.9913.4715.5032.977.5113.4712.1013.7630.375.1113.4812.0614.0431.645.81
Adv. Train. [48]×13.7811.17--6.21----------
T2S-GPT* [70]×11.486.3912.6518.4411.2011.945.9312.3215.438.9410.386.4711.6519.099.06
NSA [2]×7.837.33--13.12----------
S-MotionGPT* [24]×11.234.3912.4113.7411.4510.813.7811.5811.318.829.453.4110.429.089.68
SOKE (ours)✓6.822.357.7510.0814.486.241.717.389.6811.304.771.386.047.7211.87
", + "bbox": [ + 94, + 88, + 903, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40f831b0f02126de3537f8463de29bc00403caebd40be9b0fc744594400ba123.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art sign language generation (text-to-sign) methods. * denotes reimplementation results." + ], + "table_footnote": [], + "table_body": "
Decoding MethodSign RetrievalHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓
AvgBodyHands/videoAvgBodyHands/videoAvgBodyHands/video
Sequential×4.648.553.533.263.748.142.493.153.266.972.203.23
Parallel×5.069.273.861.394.198.612.931.413.547.022.551.35
Multi-head×4.177.913.101.463.377.582.171.452.806.161.851.44
Multi-head✓3.346.822.351.552.726.241.711.522.134.771.381.51
", + "bbox": [ + 106, + 273, + 888, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study for the decoding method and sign retrieval. Latencies are measured using a single RTX 3090 GPU.", + "bbox": [ + 133, + 396, + 859, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "simultaneously, whereas S-MotionGPT is monolingual and requires separate models for each sign language dataset.", + "bbox": [ + 89, + 422, + 482, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. In addition to the above qualitative comparison, we also perform a quantitative evaluation against SOTA SLG methods, as detailed in Table 1. Our reimplemented S-MotionGPT achieves competitive results, with low DTW errors of 4.39 and 3.78 on How2Sign and CSL-Daily, respectively, demonstrating the effectiveness of formulating SLG as a language modeling task. The previous best SLG approach, Neural Sign Actors (NSA) [2], formulates SLG as a motion generation task and employs a diffusion model as its backbone. However, this approach overlooks the linguistic properties of sign languages, resulting in higher average DTW errors compared to S-MotionGPT. In contrast, our method sets a new state-of-the-art across all three benchmarks, achieving significantly lower DTW errors of 2.35, 1.71, and 1.38 on How2Sign, CSL-Daily, and Phoenix-2014T, respectively, using a single unified model.", + "bbox": [ + 88, + 454, + 482, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoding Method. We compare the performance of different decoding methods in Table 2. The traditional sequential decoding method is widely used in existing motion generation approaches [37, 68]. However, since tokens from different body parts are flattened into a single sequence, the number of decoding steps is tripled, leading to high inference latency (3.26s per video on How2Sign). By using three separate LM decoders, parallel decoding predicts token sequences for each body part simultaneously. This formulation significantly reduces latency (1.39s per video on How2Sign) but results in higher DTW errors due to the", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/685c6bd093b9d02bfc96fdac36184f478096bf7a4b1b7e62eacdb2628a96a28c.jpg", + "image_caption": [ + "Figure 6. Qualitative ablation study for retrieval-enhanced SLG. (Left: How2Sign; Middle: CSL-Daily; Right: Phoenix-2014T.)" + ], + "image_footnote": [], + "bbox": [ + 514, + 420, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lack of information fusion across part-wise motion tokens. Notably, our multi-head decoding method achieves a better trade-off between efficiency and information integration, maintaining comparable latency to parallel decoding while delivering even superior generation quality than sequential decoding. We attribute this to the weighting mechanism (soft gating): since the decoder inputs are a weighted average of all part-wise tokens, errors in a single token may", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23812", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0f3bb79bd908c9067cdba403d79362218c4cf81da94dce2f0529a5a80010b18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
H2SCSLPh-TGenerationReconstruction
H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)H2S (JPE↓)CSL (JPE↓)Ph-T (JPE↓)
BodyHandBodyHandBodyHandBodyHandBodyHandBodyHand
✓7.923.07----21.527.74----
✓✓7.112.636.792.14--19.886.9724.295.83--
✓✓✓6.822.356.241.714.771.3819.376.6523.525.1325.796.78
", + "bbox": [ + 114, + 88, + 880, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f769d1917485475ad48a5efbbe2d92ee0b0dc0042439f61ec0e37f3dca5e6c72.jpg", + "table_caption": [ + "Table 3. Study on the scalability of our approach. We use procrustes-aligned mean per joint position error (PA-MPJPE) [34, 72] to assess reconstruction performance of the decoupled tokenizer. (Ph-T: Phoenix-2014T.)" + ], + "table_footnote": [], + "table_body": "
VAE-1VAE-2VAE-3H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)
BodyHandBodyHandBodyHand
Whole-body××9.493.569.223.358.033.23
BodyHands×7.993.136.932.856.242.18
BodyL-handR-hand6.822.356.241.714.771.38
", + "bbox": [ + 94, + 227, + 480, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the decoupled tokenizer.", + "bbox": [ + 132, + 305, + 439, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "have a reduced impact on the overall performance.", + "bbox": [ + 89, + 337, + 424, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sign Retrieval. By leveraging external sign dictionaries, we develop a retrieval-enhanced SLG approach that incorporates retrieved word-level signs as additional generation conditions. As shown in Table 2, this retrieval mechanism significantly enhances the precision of generated signs, reducing the average DTW errors by $19.9\\% / 19.6\\% / 23.9\\%$ on How2Sign, CSL-Daily, and Phoenix-2014T, respectively. Furthermore, we perform a qualitative comparison in Figure 6. The figure demonstrates that the retrieved signs effectively guide the model's generation: similar pose patterns from the retrieved signs, particularly in hand shapes, appear in the generated outputs. This guidance substantially improves the precision of the generated signs.", + "bbox": [ + 88, + 352, + 482, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scalability with Multilingual Data. Results in Table 3 show that incorporating multilingual data significantly improves generation performance, demonstrating the scalability of our approach. This improvement stems from two key factors: 1) the tokenizer benefits from more pose training data (right half of Table 3), which subsequently enhances the generator's performance; 2) the presence of cross-lingual signs, as revealed in [63]. For example, the sign for \"cold\" in CSL and DGS shares similar body movements. Training on multilingual data helps the model better understand and generalize across these cross-lingual signs.", + "bbox": [ + 88, + 549, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decoupled Tokenizer. Inspired by the multi-cue characteristics of sign languages, we introduce a decoupled tokenizer that discretizes sign motion sequences for different body parts. In Table 4, we present an ablation study on various configurations: a single VQ-VAE for whole-body motions and one VAE for the body and another for both hands. The findings indicate that progressively separating the body and hands leads to consistently improved performance.", + "bbox": [ + 88, + 715, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 212, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the objective metrics presented in the previous tables, we conducted a user study involving 5 professional", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c35c6ee4f9fbc69d64e8273383c231c232e8706ef84d9a7677c68d5358a69fd0.jpg", + "image_caption": [ + "Figure 7. User study with professional ASL and CSL signers. We report average ratings of S-MotionGPT, our proposed SOKE, and ground truth motions." + ], + "image_footnote": [], + "bbox": [ + 517, + 229, + 901, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASL signers and 4 professional CSL signers. These signers were asked to rate the alignment between generated signs and text annotations on a scale from 1 to 10, with higher scores indicating better semantic conveyance of the texts. Ground truth sign motions were also included in the study. Specifically, we provided 15 generated signs from both the baseline, S-MotionGPT, and our proposed method, SOKE. The order of the signs was randomly shuffled to prevent potential bias.", + "bbox": [ + 511, + 398, + 906, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 7, our method achieved an encouraging average rating of 5.32/6.45, significantly surpassing the baseline, which received a low score of 1.36/2.70 on How2Sign and CSL-Daily, respectively. These results align with the above qualitative and quantitative evaluations, indicating the potential of our method in constituting a two-way communication system between the deaf and hearing.", + "bbox": [ + 511, + 535, + 908, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 652, + 633, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SOKE, a retrieval-enhanced multilingual sign language generator. Unlike existing methods that view sign language generation as a visual content generation task, we emphasize its linguistic nature, characterized by discrete semantic units. To capture the multi-cue characteristics of sign languages, we employ a decoupled tokenizer to map continuous motions into discrete tokens for the upper body and both hands. To balance efficiency and generation quality, we propose a novel multi-head decoding method that reduces decoding steps by two-thirds while dynamically integrating information from different body parts. Additionally, we present a simple yet effective retrieval-enhanced approach by leveraging external sign dictionaries. Extensive evaluations on a combined ASL-CSL-DGS dataset validate the effectiveness of our method.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23813", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1), EPSRC Project GNOMON (EP/X011364/1) and Turing AI Fellowship (EP/Z534699/1). R.A. Potamias and R. Zuo were supported by EPSRC Project GNOMON (EP/X011364/1). This research was also supported by the Sovereign AI AIRR programme. We would also like to thank anonymous signers who participated in the user study.", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 239, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rotem Shalev Arkushin, Amit Moryossef, and Ohad Fried. Ham2pose: Animating sign language notation into pose sequences. In CVPR, pages 21046-21056, 2023. 2, 5", + "[2] Vasileios Baltatzis, Rolandos Alexandros Potamias, Evangelos Ververas, Guanxiong Sun, Jiankang Deng, and Stefanos Zafeiriou. Neural sign actors: A diffusion model for 3d sign language production from text. In CVPR, pages 1985-1995, 2024. 1, 2, 3, 5, 7", + "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, 2018. 2, 3, 5", + "[4] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. In AAAI, pages 17754-17762, 2024. 2, 5", + "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2", + "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 5", + "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 5", + "[8] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. In ECCV, pages 93-110, 2024. 2", + "[9] Lu Dong, Lipisha Chaudhary, Fei Xu, Xiao Wang, Mason Lary, and Ifeoma Nwogu. Signavatar: Sign language 3d motion reconstruction and generation. In FG, 2024. 2", + "[10] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In CVPR, pages 2735-2744, 2021. 2, 3, 5", + "[11] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In SIGKDD, pages 6491-6501, 2024. 2, 5", + "[12] Sen Fang, Chunyu Sui, Xuedong Zhang, and Yapeng Tian. SignDiff: Learning diffusion models for American sign language production. arXiv preprint arXiv:2308.16082, 2023. 1, 2" + ], + "bbox": [ + 93, + 265, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Sen Fang, Lei Wang, Ce Zheng, Yapeng Tian, and Chen Chen. Signllm: Sign languages production large language models. arXiv preprint arXiv:2405.10718, 2024. 2", + "[14] fxsjy. https://github.com/fxsjy/jieba. In *Github*, 2012. 5", + "[15] Aditya Golatkar, Alessandro Achille, Luca Zancato, Yu-Xiang Wang, Ashwin Swaminathan, and Stefano Soatto. Cpr: Retrieval augmented generation for copyright protection. In CVPR, pages 12374-12384, 2024. 2", + "[16] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In CVPR, pages 18362-18372, 2024. 1", + "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2", + "[18] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 1", + "[19] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 3", + "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. SignBERT+: Hand-model-aware self-supervised pretraining for sign language understanding. TPAMI, 2023. 1", + "[21] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In MM, pages 3172-3181, 2021. 2", + "[22] Wencan Huang, Zhou Zhao, Jinzheng He, and Mingmin Zhang. Dualsign: Semi-supervised sign language production with balanced multi-modal multi-task dual transformation. In MM, pages 5486-5495, 2022. 2", + "[23] Euijun Hwang, Jung-Ho Kim, and Jong C Park. Nonautoregressive sign language production with gaussian space. In BMVC, 2021. 7", + "[24] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 36:20067-20079, 2023. 2, 4, 6, 7", + "[25] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In EMNLP, pages 7969-7992, 2023. 2", + "[26] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 3", + "[27] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In ECCV, pages 349-367, 2024. 5", + "[28] Hamid Reza Vaezi Joze and Oscar Koller. MS-ASL: A large-scale data set and benchmark for understanding American sign language. In BMVC, 2019. 5", + "[29] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23814", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Gabriele Langer, Anke Müller, Sabrina Wähl, Felicitas Otte, Lea Sepke, and Thomas Hanke. Introducing the DW-DGS – the digital dictionary of DGS. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, 2024. 5", + "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. NeurIPS, 33:9459-9474, 2020. 2", + "[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, pages 1459-1469, 2020. 5", + "[33] Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, and Houqiang Li. Uni-sign: Toward unified sign language understanding at scale. In ICLR, 2025. 1", + "[34] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 3, 8", + "[35] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. TACL, 8:726-742, 2020. 2, 4, 5", + "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5", + "[37] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, pages 32939-32977, 2024. 2, 3, 4, 7", + "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 3", + "[39] Mathias Müller, Zifan Jiang, Amit Moryossef, Annette Rios Gonzales, and Sarah Ebling. Considerations for meaningful sign language machine translation based on glosses. In ACL, pages 682-693, 2023. 2", + "[40] Zhe Niu, Ronglai Zuo, Brian Mak, and Fangyun Wei. A Hong Kong sign language corpus collected from sign-interpreted tv news. In LREC-COLING, pages 636-646, 2024. 6", + "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5", + "[42] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2", + "[43] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In CVPR, 2025. 3" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Fan Qi, Yu Duan, Changsheng Xu, and Huaiwen Zhang. Signgen: End-to-end sign language video generation with latent diffusion. In ECCV, 2024. 1", + "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 6", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2", + "[47] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In ECCV, pages 687-705, 2020. 1, 2, 7", + "[48] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Adversarial training for multi-channel sign language production. In BMVC, 2020. 7", + "[49] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3D multi-channel sign language production via progressive transformers and mixture density networks. IJCV, 129(7):2113-2135, 2021. 2", + "[50] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Signing at scale: Learning to co-articulate signs for large-scale photo-realistic sign language production. In CVPR, pages 5141-5151, 2022. 1, 2, 3", + "[51] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2", + "[52] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. knndiffusion: Image generation via large-scale retrieval. In ICLR, 2023. 3", + "[53] Tongkai Shi, Lianyu Hu, Fanhua Shang, Jichao Feng, Peidong Liu, and Wei Feng. Pose-guided fine-grained sign language video generation. In ECCV, pages 392-409, 2024. 1", + "[54] Stephanie Stoll, Armin Mustafa, and Jean-Yves Guillemaut. There and back again: 3d sign language generation from text using back-translation. In 3DV, pages 187–196, 2022. 7", + "[55] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In ICLR, 2024. 2", + "[56] Sihan Tan, Taro Miyazaki, and Kazuhiro Nakadai. Multilingual gloss-free sign language translation: Towards building a sign language foundation model. In ACL, 2025. 1", + "[57] Shengeng Tang, Richang Hong, Dan Guo, and Meng Wang. Gloss semantic-enhanced network with online backtranslation for sign language production. In MM, pages 5630-5638, 2022. 2", + "[58] Shengeng Tang, Jiayi He, Lechao Cheng, Jingjing Wu, Dan Guo, and Richang Hong. Discrete to continuous: Generating smooth transition poses from sign language observations. In CVPR, 2025. 1", + "[59] Shengeng Tang, Jiayi He, Dan Guo, Yanyan Wei, Feng Li, and Richang Hong. Sign-idd: Iconicity disentangled diffusion for sign language production. In AAAI, 2025. 1" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23815", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1", + "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3", + "[62] Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, et al. Searching for best practices in retrieval-augmented generation. In EMNLP, pages 17716-17736, 2024. 2, 5", + "[63] Fangyun Wei and Yutong Chen. Improving continuous sign language recognition with cross-lingual signs. In ICCV, pages 23612-23621, 2023. 5, 8", + "[64] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In ICLR, 2024. 1", + "[65] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI, pages 6234-6242, 2024. 2, 3", + "[66] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In CVPR, pages 13525-13536, 2024. 2", + "[67] Huijie Yao, Wengang Zhou, Hao Feng, Hezhen Hu, Hao Zhou, and Houqiang Li. Sign language translation with iterative prototype. In ICCV, pages 15592-15601, 2023. 1", + "[68] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2, 3, 4, 7", + "[69] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. MLSLT: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1", + "[70] Aoxiong Yin, Haoyuan Li, Kai Shen, Siliang Tang, and Yueting Zhuang. T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text. In ACL, 2024. 1, 2, 7", + "[71] Kayo Yin, Amit Moryossef, Julie Hochgesang, Yoav Goldberg, and Malihe Alikhani. Including signed languages in natural language processing. In ACL, pages 7347-7360, 2021. 1", + "[72] Zhengdi Yu, Shaoli Huang, Yongkang Cheng, and Tolga Birdal. Signavatars: A large-scale 3d sign language holistic motion dataset and benchmark. In ECCV, pages 1-19, 2024. 1, 2, 3, 8", + "[73] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In WACV, pages 3395-3403, 2020. 2", + "[74] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2", + "[75] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffeuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, pages 7368-7376, 2024. 2", + "[77] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. BEST: BERT pre-training for sign language recognition with coupling tokenization. In AAAI, 2023. 3", + "[78] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houqiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 5", + "[79] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. IEEE TIP, 2024. 3", + "[80] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In ICCV, pages 20871-20881, 2023. 1", + "[81] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, pages 13009-13016, 2020. 2, 3", + "[82] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 2, 3, 5", + "[83] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, pages 1-1, 2021. 1", + "[84] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In CVPR, pages 1357-1366, 2024. 2", + "[85] Ronglai Zuo and Brian Mak. C2SLR: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2", + "[86] Ronglai Zuo and Brian Mak. Improving continuous sign language recognition with consistency constraints and signer removal. ACM TOMM, 2024. 3", + "[87] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 5", + "[88] Ronglai Zuo, Fangyun Wei, Zenggui Chen, Brian Mak, Jiaolong Yang, and Xin Tong. A simple baseline for spoken language to sign language translation with 3d avatars. In ECCV, 2024. 1, 2, 3", + "[89] Ronglai Zuo, Fangyun Wei, and Brian Mak. Towards online continuous sign language recognition and translation. In EMNLP, 2024. 5" + ], + "bbox": [ + 516, + 90, + 903, + 799 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23816", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_model.json b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ec5b309cc8342f7a8a8a99658a26b67e8a9d5f5 --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.898, + 0.155 + ], + "angle": 0, + "content": "Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.181, + 0.905, + 0.217 + ], + "angle": 0, + "content": "Ronglai Zuo, Rolandos Alexandros Potamias, Evangelos Ververas, Jiankang Deng, Stefanos Zafeiriou \nImperial College London" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.22, + 0.638, + 0.234 + ], + "angle": 0, + "content": "https://2000zrl.github.io/soke/" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.256, + 0.885, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.449, + 0.908, + 0.48 + ], + "angle": 0, + "content": "Figure 1. We propose Signs as Tokens (SOKE), a unified sign language generator that can produce multilingual signs from text inputs. (Left: American sign language; Middle: Chinese sign language; Right: German sign language.)" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.49, + 0.328, + 0.506 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.524, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Sign language is a visual language that encompasses all linguistic features of natural languages and serves as the primary communication method for the deaf and hard-of-hearing communities. Although many studies have successfully adapted pretrained language models (LMs) for sign language translation (sign-to-text), the reverse task—sign language generation (text-to-sign)—remains largely unexplored. In this work, we introduce a multilingual sign language model, Signs as Tokens (SOKE), which can generate 3D sign avatars autoregressively from text inputs using a pretrained LM. To align sign language with the LM, we leverage a decoupled tokenizer that discretizes continuous signs into token sequences representing various body parts. During decoding, unlike existing approaches that flatten all part-wise tokens into a single sequence and predict one token at a time, we propose a multi-head decoding method capable of predicting multiple tokens simultaneously. This approach improves inference efficiency while maintaining effective information fusion across different body parts. To further ease the generation process, we propose a retrieval-enhanced SLG approach, which incorporates external sign dictionaries to provide accurate word-level signs as auxiliary conditions, significantly improving the precision of generated signs. Extensive qualitative and quantitative evaluations demonstrate the effectiveness of SOKE." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.49, + 0.645, + 0.505 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.67 + ], + "angle": 0, + "content": "Sign language is the primary communication method utilized by the deaf and hard-of-hearing communities, embodying all linguistic properties found in spoken languages, such as discrete semantic units and grammatical structure [16, 71]. As an attempt to mitigate the barrier between sign and spoken languages, several methods have been developed to tackle sign language processing by unfolding the task into two research problems: sign language translation (SLT, sign-to-text) [20, 33, 56, 64, 67, 69, 80, 83] and sign language generation (SLG, text-to-sign) [2, 47, 53, 72, 88]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.853 + ], + "angle": 0, + "content": "In contrast with SLT, the usage of language models (LMs) in SLG remains relatively unexplored. Most existing methods approach SLG as a visual content generation task (e.g., video, keypoint, or motion), using GANs [50, 53] or diffusion models [2, 12, 44, 58, 59]. However, these works neglect the linguistic nature of sign languages, leading to suboptimal performance and missing benefits of pretrained LMs such as generalizability and scalability [18, 29, 60]. In particular, sign languages share the fundamental linguistic properties and the discrete structure of spoken languages [16, 71], which necessitates processing using autoregressive language modeling techniques [70]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To formulate SLG in the framework of language modeling, a crucial step is mapping signs into discrete tokens. A common approach to discretizing sign languages involves" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23806" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.469 + ], + "angle": 0, + "content": "using glosses—the written form of signs—as an intermediate representation [13, 22, 50, 88]. However, glosses require extensive annotation efforts and impose a predefined information bottleneck, which fail to fully capture the rich semantics of sign languages [39]. To address these limitations and better capture the multi-cue property of sign languages [26, 81, 85], we leverage a vector quantized-variational auto-encoder (VQ-VAE)-based tokenizer [37, 65, 68] that learns mappings between continuous sign motions and discrete tokens for each body part (upper body, left hand, and right hand). During decoding, existing autoregressive motion generation methods [37, 68] typically flatten tokens from different body parts into a single sequence, resulting in inefficient inference due to tripled decoding steps. To overcome this challenge, we propose a novel multi-head decoding method capable of predicting multiple tokens at a time while effectively fusing information from different body parts in real time. Furthermore, given the limited vocabulary size of continuous sign language datasets, we enhance the model's generalization by instantiating the LM as a multilingual LM [35], enabling it to handle a broader and more diverse domain of discourse. To support multilingual SLG in a unified framework, we train the proposed model on a curated multilingual sign language dataset encompassing American, Chinese, and German sign languages [3, 10, 82]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Recently, retrieval-augmented generation (RAG) [4, 11, 62] has proven effective in enhancing the interpretability and accuracy of content generated by large language models. Similarly, several SLG methods [50, 88] leverage glosses to retrieve sign dictionaries and directly incorporate retrieved signs into the outputs, improving the interpretability of generated signs. However, beyond the inherent limitations of glosses, these methods often face challenges in handling co-articulations between adjacent signs. This is because dictionary signs are typically recorded at the word level, whereas SLG models operate at the sentence level, leading to potential unnaturalness in the generated sequences. To address these issues, we propose a retrieval-enhanced SLG approach that utilizes motion tokens from retrieved dictionary signs as additional conditions for language modeling. This approach avoids the unnaturalness of directly copying signs into model outputs while still benefiting from the precision of dictionary signs. To sum up, the contributions of this paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- We propose a novel SLG approach, Signs as Tokens (SOKE). With the aid of a pretrained LM, SOKE can handle multiple sign languages in a unified model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- We propose a multi-head decoding method that reduces inference latency and efficiently fuses information from different body parts, enabling seamless integration of a decoupled tokenizer with the LM." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- By leveraging external sign dictionaries, we introduce a retrieval-enhanced SLG approach to improve the preci" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.093, + 0.686, + 0.107 + ], + "angle": 0, + "content": "sion of generated signs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- We curate two sign language datasets [3, 82] with accurate SMPL-X [42] poses to facilitate multilingual SLG research. Experiments demonstrate that our SOKE achieves state-of-the-art performance on three challenging benchmarks, How2Sign [10], CSL-Daily [82], and Phoenix2014T [3], using a single unified model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.211, + 0.655, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.237, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Sign Language Generation. Early gloss-based methods [21, 22, 50, 57, 65, 73, 88] approach SLG under a text-to-gloss-to-sign framework. Several works [1, 47, 49] propose to directly regress 2D joints to represent sign poses, simplifying the modeling process, however, they suffer from mode collapse due to the limited scale of data. Recent state-of-the-art SLG works can be categorized into two classes: the first group of methods [2, 12] employs diffusion models to generate sign motions conditioned on text inputs; the second group of methods [9, 70, 72] adopts a tokenizer-LM two-stage autoregressive generation approach. Among these, [72] incorporates costly gloss and HamNoSys notations [1] to train a semantic adapter, while [9] focuses on generating signs from a single word, limiting their applicability. Moreover, a notable gap exists in the literature regarding unified models for multilingual SLG. In this work, we address these gaps by utilizing a decoupled tokenizer and a pretrained multilingual LM, integrated with our customized multi-head decoding method and retrieval-enhanced approach, to enable text-driven multilingual SLG. Text-to-Motion Generation. Text-to-motion generation (T2MG) is an important research area with potential applications in human-computer interaction and robotics. Recent T2MG methods can be categorized into two main types: 1) Diffusion-based approaches [5, 51, 75], which utilize latent diffusion [46] to generate motions conditioned on text inputs. 2) Autoregressive approaches [17, 24, 37, 74, 76, 84], which draw inspiration from text generation methods by first tokenizing motion sequences and then using LMs to perform autoregressive generation. When decoding with a decoupled tokenizer [37, 68], these autoregressive approaches simply flatten tokens from various body parts into one sequence, causing additional decoding steps. In this work, we propose a novel multi-head decoding method that can efficiently predict motion tokens for all body parts at one single step, significantly reducing inference latency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Retrieval-Enhanced Generation. Using the concept of retrieval to enhance generation models has proven effective across various domains [4, 8, 11, 15, 62, 66]. As a pioneering effort, Piktus et al. [31] propose to retrieve relevant documents before generating answers for knowledge-intensive NLP tasks. FLARE [25] introduces selective retrieval, activating it only when low-confidence tokens appear in the generation process. ToG [55] leverages retrieved" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.088, + 0.904, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. An overview of our proposed method, Signs as Tokens (SOKE). We begin by training a VQ-VAE-based decoupled tokenizer to map continuous sign motions into discrete tokens for various body parts (upper body, left hand, and right hand). These sign motion tokens are then integrated into the text vocabulary of a pretrained language model, which serves as the backbone of our autoregressive multilingual generator (AMG). Given a text input, the AMG first retrieves word-level signs from external dictionaries, appends their motion tokens to the text tokens, and feeds them into the language model encoder. During decoding, our novel multi-head decoding strategy generates motion tokens for all body parts simultaneously at each time step. Finally, the derived motion tokens are used to reconstruct sign avatars." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.542 + ], + "angle": 0, + "content": "knowledge to enhance the reasoning capabilities of large language models. In the field of image generation, retrieval has also been utilized to enable diffusion models to generate out-of-distribution images [52]. In the context of SLG, several methods [50, 88] adopt gloss-based retrieval to obtain dictionary signs and approach SLG in a retrieve-and-concatenate manner. However, glosses have inherent limitations, and such methods often suffer from unrealistic transitions between signs. In this paper, we propose a simple yet effective retrieval-enhanced SLG approach that uses word-level signs as auxiliary conditions, addressing these limitations and improving the overall quality of generations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.555, + 0.226, + 0.573 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.484, + 0.718 + ], + "angle": 0, + "content": "An overview of our method is shown in Figure 2. Our approach mainly consists of two parts: a decoupled tokenizer (DETO) and an autoregressive multilingual generator (AMG). In the first stage, DETO learns mappings between continuous sign motions and discrete tokens for each body part (Section 3.2). In the second stage, AMG is trained to autoregressively generate motion tokens with the proposed multi-head decoding strategy (Section 3.3) conditioned on text inputs and retrieved dictionary signs (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.262, + 0.743 + ], + "angle": 0, + "content": "3.1. Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Although existing SLG works [2, 72] have made efforts in curating high-quality SMPL-X poses for an American sign language (ASL) dataset, How2Sign [10], there is a notable lack of 3D annotations for other continuous sign language datasets, impeding multilingual SLG research. In this work, we rely on a two-stage approach, leveraging state-of-the-art 3D pose estimation methods [34, 43], to extract high-fidelity SMPL-X poses for CSL-Daily [82] and Phoenix-2014T [3], two widely-adopted datasets of Chinese and German sign languages (CSL and DGS), respec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.572 + ], + "angle": 0, + "content": "tively. Specifically, we initially estimate a set of rough 3D body poses for each signer using OSX [34] along with accurate 3D hand poses from WiLoR [43]. Given that OSX often falls short in accurately capturing the arm poses, we follow [2] and refine the upper body joint rotations by minimizing a re-projection loss between the estimated joints and the detected 2D joint keypoints from Mediapipe [38]. Using this process we can ensure accurate whole-body annotations. We represent each sign motion sequence as \\(\\mathbf{S} \\in \\mathbb{R}^{T \\times d}\\), where \\(T\\) is the sequence length and \\(d = 133\\) denotes the number of SMPL-X parameters, including 11 upper body joints, 30 hand joints, and 10 expression parameters. Please check the supplementary materials for more details and quantitative evaluations of our pose fitting pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.714, + 0.601 + ], + "angle": 0, + "content": "3.2. Decoupled Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.713 + ], + "angle": 0, + "content": "Sign languages exhibit a multi-cue property [19, 26, 77, 79, 81, 86], where the semantics of a sign are conveyed simultaneously through body and hand movements. Inspired by this, we leverage a decoupled tokenizer [37, 65, 68], comprising three VQ-VAEs [61], to independently model key regions: the upper body and both hands. An illustration is available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Given a \\(T\\)-frame sign motion input \\(\\mathbf{S} \\in \\mathbb{R}^{T \\times d}\\), we first decompose it into three part-wise motion sequences based on the kinematic tree of SMPL-X: \\(\\mathbf{S}^p \\in \\mathbb{R}^{T \\times d_p}\\), where \\(p \\in \\{B, LH, RH\\}\\). Moreover, we build three distinct VQ-VAEs, where each of them consists of an encoder \\(\\mathcal{E}_p(\\cdot)\\), a decoder \\(\\mathcal{D}_p(\\cdot)\\), and a learnable codebook \\(\\mathbf{Z}^p \\in \\mathbb{R}^{N_Z^p \\times C}\\) where \\(N_Z^p\\) represents the number of codes and \\(C\\) denotes the code dimension. For each motion sequence, the corresponding encoder first projects the sequence into a latent space: \\(\\mathbf{S}_f^p = \\mathcal{E}_p(\\mathbf{S}^p) = \\{s_{f,i}^p\\}_{i=1}^{T_f} \\in \\mathbb{R}^{T_f \\times C}\\), using a stack of 1D-CNN layers. Then, for each pose we can derive a set of discrete tokens \\(\\hat{\\mathbf{Z}}^p = \\{\\hat{z}_i^p\\}_{i=1}^{T_f}\\) using a quantizer \\(Q(\\cdot)\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23808" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.36, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.226, + 0.31, + 0.238 + ], + "angle": 0, + "content": "(a) Sequential Decoding \\((O(3K))\\)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.088, + 0.587, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.406, + 0.227, + 0.544, + 0.238 + ], + "angle": 0, + "content": "(b) Parallel Decoding \\((O(K))\\)" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.09, + 0.906, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.226, + 0.856, + 0.238 + ], + "angle": 0, + "content": "(c) Multi-head Decoding (ours, \\(O(K)\\))" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Comparison of various decoding methods. (a) Sequential decoding flattens all motion tokens into a single sequence, requiring an inefficient \\(3K\\) decoding steps, where \\(K\\) is the target motion sequence length. (b) Parallel decoding improves efficiency by decoding body parts in parallel but lacks information fusion, which may lead to suboptimal performance. (c) Our proposed multi-head decoding balances efficiency and fusion by predicting three tokens simultaneously via different heads, with input embeddings as a weighted average of part-wise token embeddings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.333, + 0.483, + 0.349 + ], + "angle": 0, + "content": "which searches the nearest neighbor from the codebook \\(\\mathbf{Z}^p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.364, + 0.483, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {i} ^ {p} = Q \\left(s _ {f, i} ^ {p}\\right) = \\arg \\min _ {z _ {j} \\in \\mathbf {Z} ^ {p}} \\| s _ {f, i} ^ {p} - z _ {j} \\| _ {2}, j \\in [ 1, N _ {Z} ^ {p} ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.518 + ], + "angle": 0, + "content": "We then feed the obtained token sequence to the corresponding part-decoder \\(\\mathcal{D}_p\\) to reconstruct the input motions: \\(\\hat{\\mathbf{S}}^p = \\mathcal{D}_p(\\mathbf{Z}^p)\\). Following [24], the objective function of each VQ-VAE is defined as: \\(\\mathcal{L}_{vq}^p = \\mathcal{L}_{rec}^p +\\mathcal{L}_{emb}^p +\\mathcal{L}_{com}^p\\) where \\(\\mathcal{L}_{rec}^p = \\| \\hat{\\mathbf{S}}^p -\\mathbf{S}^p\\| _2^2\\) denotes the reconstruction loss, \\(\\mathcal{L}_{emb}^p\\) is the embedding loss, and \\(\\mathcal{L}_{com}^p\\) is the commitment loss, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.425, + 0.552 + ], + "angle": 0, + "content": "3.3. Autoregressive Multilingual Generator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Using obtained discrete sign representations, we can respect the discrete properties of sign languages and approach SLG similarly to text generation with a pretrained LM [35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Integrated Vocabulary. We first gather the discrete tokens for different body parts to construct a part-wise motion vocabulary: \\(\\mathcal{V}_m^p = \\{\\langle p\\_ i\\rangle \\}_{i = 1}^{N_Z^p}\\), e.g., \\(\\langle B\\_ 1\\rangle\\) and \\(\\langle \\mathrm{RH\\_2}\\rangle\\). The overall motion vocabulary is a combination of three part-wise vocabularies: \\(\\mathcal{V}_m = \\{\\mathcal{V}_m^B,\\mathcal{V}_m^{LH},\\mathcal{V}_m^{RH}\\}\\). Following [35], we further define a series of language identifier tokens: \\(\\mathcal{V}_l = \\{\\langle \\mathrm{ASL}\\rangle ,\\langle \\mathrm{CSL}\\rangle ,\\langle \\mathrm{DGS}\\rangle \\}\\), to prompt the LM with the information of target sign language. The integrated vocabulary for the generator is defined as: \\(\\mathcal{V} = \\{\\mathcal{V}_t,\\mathcal{V}_m,\\mathcal{V}_l\\}\\), which lays a foundation for the subsequent LM fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Sequential Decoding. Following previous motion generation works [37, 68], a naive decoding method is simply flattening tokens from each body part into a single sequence with tripled length: \\(\\mathbf{Y} = \\{y_1^B, y_1^{LH}, y_1^{RH}, \\ldots, y_K^B, y_K^{LH}, y_K^{RH}\\}\\), where \\(y_k^p \\in \\mathcal{V}_m^p\\) denotes the part-wise motion token, and \\(K\\) denotes the length of the target sign motion sequence. The decoding process can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.33, + 0.908, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathbf {Y} \\mid \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {2} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {k} ^ {B}, y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {k} ^ {B, L H}, y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.906, + 0.496 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_{en}\\) denotes the last encoder hidden states. For simplicity, we write \\(y_{\\) when the model is required to generate upper-body motions for ASL. The decoding process can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.742, + 0.907, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = P (\\mathbf {Y} ^ {B}) P (\\mathbf {Y} ^ {L H}) P (\\mathbf {Y} ^ {R H}) \\tag {3} \\\\ = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B} \\mid y _ {< k} ^ {B}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {L H} \\mid y _ {< k} ^ {L H}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {R H} \\mid y _ {< k} ^ {R H}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Comparing Eq. 3 with Eq. 2, the parallel decoding strategy significantly reduces the dependence on previous tokens, as the decoding processes for each body part are completely separated. While this formulation enables efficient parallel execution, the overly independent assumption may result in suboptimal performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.482, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Figure 4. Illustration of retrieval-enhanced SLG. We utilize external sign dictionaries to retrieve accurate word-level signs, which serve as additional conditions to enhance the generation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Multi-Head Decoding. To balance efficiency and information fusion across body parts, we propose a multi-head decoding strategy. As shown in Figure 3(c), we design three language modeling heads, implemented as fully connected layers, to predict motion tokens for each body part simultaneously at each step. Moreover, the decoder inputs at each step are modified to a weighted average of the token embeddings from different body parts. Specifically, denoting the token embeddings from upper body, left hand, and right hand as \\(\\mathbf{E}^B\\), \\(\\mathbf{E}^{LH}\\), and \\(\\mathbf{E}^{RH}\\), respectively, the input embeddings are defined as: \\(\\mathbf{E} = (1 - 2\\lambda)\\mathbf{E}^B + \\lambda \\mathbf{E}^{LH} + \\lambda \\mathbf{E}^{RH}\\) where \\(\\lambda \\in (0,0.5)\\) is a hyper-parameter to control the weight of hand embeddings. The decoding process can then be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.513, + 0.483, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {4} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.64 + ], + "angle": 0, + "content": "which can be viewed as relaxing Eq. 2 under a conditional independence assumption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Training and Inference. The LM is trained with a standard cross-entropy loss: \\(\\mathcal{L}_{LM} = -\\log P(\\mathbf{Y}|\\mathbf{h_{en}})\\). During inference, we adopt a simple greedy decoding algorithm that the LM always outputs the token with the highest probability at each step. For parallel and multi-head decoding, the inference process terminates as soon as any head or decoder predicts an end-of-sentence token. The obtained token sequences will be fed into the corresponding part-decoder, \\(\\mathcal{D}_p(\\cdot)\\), to reconstruct sign motions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.32, + 0.803 + ], + "angle": 0, + "content": "3.4. Retrieval-Enhanced SLG" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Inspired by the effectiveness of retrieval-augmented generation in NLP [4, 11, 62], we propose to leverage external sign dictionaries to improve the fidelity of generated signs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 4, we first build a word-level sign dictionary for each target sign language by utilizing resources from isolated sign language recognition datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "[28, 30, 32, 78, 87] and online sources. Through our pose fitting pipeline, each RGB video in these dictionaries is converted into SMPL-X poses. Next, we employ a well-trained decoupled tokenizer to map the poses into discrete tokens. The resulting dictionary is represented as a set of quadruples \\(\\{(w, m^B, m^{LH}, m^{RH})\\}\\), where \\(w\\) denotes lemmatized words in the written language, and \\(m^p\\) denotes motion tokens. If a word has multiple sign instances in the dictionary, we retain only the instance with the lowest reconstruction error after passing through the tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Given a text input \\(\\mathbf{X} = \\{x_{i}\\}_{i = 1}^{L}\\), consisting of \\(L\\) words\\(^1\\), we collect all words and corresponding tokens found in the dictionary. The final prompt is created by concatenating raw texts \\(\\mathbf{X}\\) with the retrieved motion tokens \\(\\{m_j^p\\}\\), which is then fed into the LM encoder to generate hidden states \\(\\mathbf{h}_{en}\\). Several examples are shown in Figure 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.646, + 0.367 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.906, + 0.512 + ], + "angle": 0, + "content": "Datasets. We construct a multilingual sign language dataset by merging three widely-adopted datasets: How2Sign [10], CSL-Daily [82], and Phoenix-2014T [3]. How2Sign is a large-scale ASL dataset, consisting of 35K video-text pairs. CSL-Daily is the largest existing continuous CSL dataset, containing 20K samples. Phoenix-2014T is a DGS dataset with 8K samples. To train SOKE, we use the SMPL-X poses of How2Sign dataset as provided in [2] along with the curated poses of CSL-Daily and Phoenix-2014T." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.634 + ], + "angle": 0, + "content": "Evaluation Metrics. Since the length of generated signs may differ from the ground truth, we employ the well-established dynamic time warping (DTW) [1, 2] on both procrustes-aligned and original joint position errors (JPE) to measure sequence-level distances between the generated signs and ground truth. Following [2], we also report backtranslation (B-T) scores in BLEU-4 [41] to assess the interpretability of the generated signs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.634, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Implementation Details. For DETO, we empirically set the code numbers to \\( N_Z^B = 96 \\) for the body and \\( N_Z^{LH} = N_Z^{RH} = 192 \\) for the left and right hands, respectively, with a code dimension of \\( C = 512 \\). We train DETO on the combined multilingual sign language dataset and sign dictionaries with a batch size of 320 per GPU for 500 epochs, using the AdamW optimizer [36] and a cosine learning rate scheduler starting at 2e-4. For the LM, we use mBART-large-cc25 [35], which features 12 encoder-decoder layers with a model dimension of 1024 and is widely used in various sign language translation methods [6, 7, 27, 63, 89]. We empirically set \\( \\lambda = 1/3 \\) to balance part-wise token embeddings. The LM is fine-tuned with a batch size of 32 per GPU on the multilingual sign language dataset for 150 epochs, employing the same optimizer settings as DETO. All models are trained using 6 RTX 3090 GPUs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.905, + 0.9 + ], + "angle": 0, + "content": "1 Since Chinese sentences lack spaces between words, we use Jieba [14," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.715, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of generated signs between our proposed method, SOKE, with the SOTA method, S-MotionGPT [24], on the test sets of How2Sign (left), CSL-Daily (middle), and Phoenix-2014T (right)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.755, + 0.461, + 0.772 + ], + "angle": 0, + "content": "4.1. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Qualitative Comparison. We first reimplement a state-of-the-art (SOTA) open-sourced motion generation approach, MotionGPT [24], which utilizes a VQ-VAE to discretize whole-body motions and employs a monolingual language model [45], on sign language datasets. The resulting model, S-MotionGPT, achieves competitive results on sign" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.893 + ], + "angle": 0, + "content": "language benchmarks (Table 1). In Figure 5, we perform a qualitative comparison between our SOKE and S-MotionGPT. As can be easily seen, our proposed method produces significantly better results, with visual quality comparable to the ground truth sign videos across the ASL, CSL, and DGS datasets. Notably, our method can generate precise and informative hand movements, which play a key role in conveying semantics in sign languages. Additionally, the proposed method supports three sign languages" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.888, + 0.238, + 0.9 + ], + "angle": 0, + "content": "40] for word segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23811" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.248 + ], + "angle": 0, + "content": "
MethodMulti-lingualHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑
BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4
NAR [23]×13.9411.80--5.75----------
Prog. Trans.* [47]×14.1511.5714.7430.172.7515.9812.9116.3032.633.0713.6711.9515.0131.774.94
Text2Mesh* [54]×13.9913.4715.5032.977.5113.4712.1013.7630.375.1113.4812.0614.0431.645.81
Adv. Train. [48]×13.7811.17--6.21----------
T2S-GPT* [70]×11.486.3912.6518.4411.2011.945.9312.3215.438.9410.386.4711.6519.099.06
NSA [2]×7.837.33--13.12----------
S-MotionGPT* [24]×11.234.3912.4113.7411.4510.813.7811.5811.318.829.453.4110.429.089.68
SOKE (ours)✓6.822.357.7510.0814.486.241.717.389.6811.304.771.386.047.7211.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.253, + 0.871, + 0.268 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art sign language generation (text-to-sign) methods. * denotes reimplementation results." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.274, + 0.89, + 0.394 + ], + "angle": 0, + "content": "
Decoding MethodSign RetrievalHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓
AvgBodyHands/videoAvgBodyHands/videoAvgBodyHands/video
Sequential×4.648.553.533.263.748.142.493.153.266.972.203.23
Parallel×5.069.273.861.394.198.612.931.413.547.022.551.35
Multi-head×4.177.913.101.463.377.582.171.452.806.161.851.44
Multi-head✓3.346.822.351.552.726.241.711.522.134.771.381.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.397, + 0.861, + 0.412 + ], + "angle": 0, + "content": "Table 2. Ablation study for the decoding method and sign retrieval. Latencies are measured using a single RTX 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.454 + ], + "angle": 0, + "content": "simultaneously, whereas S-MotionGPT is monolingual and requires separate models for each sign language dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Quantitative Comparison. In addition to the above qualitative comparison, we also perform a quantitative evaluation against SOTA SLG methods, as detailed in Table 1. Our reimplemented S-MotionGPT achieves competitive results, with low DTW errors of 4.39 and 3.78 on How2Sign and CSL-Daily, respectively, demonstrating the effectiveness of formulating SLG as a language modeling task. The previous best SLG approach, Neural Sign Actors (NSA) [2], formulates SLG as a motion generation task and employs a diffusion model as its backbone. However, this approach overlooks the linguistic properties of sign languages, resulting in higher average DTW errors compared to S-MotionGPT. In contrast, our method sets a new state-of-the-art across all three benchmarks, achieving significantly lower DTW errors of 2.35, 1.71, and 1.38 on How2Sign, CSL-Daily, and Phoenix-2014T, respectively, using a single unified model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.245, + 0.727 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Decoding Method. We compare the performance of different decoding methods in Table 2. The traditional sequential decoding method is widely used in existing motion generation approaches [37, 68]. However, since tokens from different body parts are flattened into a single sequence, the number of decoding steps is tripled, leading to high inference latency (3.26s per video on How2Sign). By using three separate LM decoders, parallel decoding predicts token sequences for each body part simultaneously. This formulation significantly reduces latency (1.39s per video on How2Sign) but results in higher DTW errors due to the" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.421, + 0.907, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.731, + 0.905, + 0.759 + ], + "angle": 0, + "content": "Figure 6. Qualitative ablation study for retrieval-enhanced SLG. (Left: How2Sign; Middle: CSL-Daily; Right: Phoenix-2014T.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "lack of information fusion across part-wise motion tokens. Notably, our multi-head decoding method achieves a better trade-off between efficiency and information integration, maintaining comparable latency to parallel decoding while delivering even superior generation quality than sequential decoding. We attribute this to the weighting mechanism (soft gating): since the decoder inputs are a weighted average of all part-wise tokens, errors in a single token may" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23812" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.882, + 0.193 + ], + "angle": 0, + "content": "
H2SCSLPh-TGenerationReconstruction
H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)H2S (JPE↓)CSL (JPE↓)Ph-T (JPE↓)
BodyHandBodyHandBodyHandBodyHandBodyHandBodyHand
✓7.923.07----21.527.74----
✓✓7.112.636.792.14--19.886.9724.295.83--
✓✓✓6.822.356.241.714.771.3819.376.6523.525.1325.796.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.194, + 0.907, + 0.224 + ], + "angle": 0, + "content": "Table 3. Study on the scalability of our approach. We use procrustes-aligned mean per joint position error (PA-MPJPE) [34, 72] to assess reconstruction performance of the decoupled tokenizer. (Ph-T: Phoenix-2014T.)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.228, + 0.481, + 0.306 + ], + "angle": 0, + "content": "
VAE-1VAE-2VAE-3H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)
BodyHandBodyHandBodyHand
Whole-body××9.493.569.223.358.033.23
BodyHands×7.993.136.932.856.242.18
BodyL-handR-hand6.822.356.241.714.771.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.133, + 0.306, + 0.441, + 0.32 + ], + "angle": 0, + "content": "Table 4. Ablation study on the decoupled tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.425, + 0.352 + ], + "angle": 0, + "content": "have a reduced impact on the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.353, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Sign Retrieval. By leveraging external sign dictionaries, we develop a retrieval-enhanced SLG approach that incorporates retrieved word-level signs as additional generation conditions. As shown in Table 2, this retrieval mechanism significantly enhances the precision of generated signs, reducing the average DTW errors by \\(19.9\\% / 19.6\\% / 23.9\\%\\) on How2Sign, CSL-Daily, and Phoenix-2014T, respectively. Furthermore, we perform a qualitative comparison in Figure 6. The figure demonstrates that the retrieved signs effectively guide the model's generation: similar pose patterns from the retrieved signs, particularly in hand shapes, appear in the generated outputs. This guidance substantially improves the precision of the generated signs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Scalability with Multilingual Data. Results in Table 3 show that incorporating multilingual data significantly improves generation performance, demonstrating the scalability of our approach. This improvement stems from two key factors: 1) the tokenizer benefits from more pose training data (right half of Table 3), which subsequently enhances the generator's performance; 2) the presence of cross-lingual signs, as revealed in [63]. For example, the sign for \"cold\" in CSL and DGS shares similar body movements. Training on multilingual data helps the model better understand and generalize across these cross-lingual signs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.716, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Decoupled Tokenizer. Inspired by the multi-cue characteristics of sign languages, we introduce a decoupled tokenizer that discretizes sign motion sequences for different body parts. In Table 4, we present an ablation study on various configurations: a single VQ-VAE for whole-body motions and one VAE for the body and another for both hands. The findings indicate that progressively separating the body and hands leads to consistently improved performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.214, + 0.865 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In addition to the objective metrics presented in the previous tables, we conducted a user study involving 5 professional" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.231, + 0.903, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.344, + 0.907, + 0.387 + ], + "angle": 0, + "content": "Figure 7. User study with professional ASL and CSL signers. We report average ratings of S-MotionGPT, our proposed SOKE, and ground truth motions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.534 + ], + "angle": 0, + "content": "ASL signers and 4 professional CSL signers. These signers were asked to rate the alignment between generated signs and text annotations on a scale from 1 to 10, with higher scores indicating better semantic conveyance of the texts. Ground truth sign motions were also included in the study. Specifically, we provided 15 generated signs from both the baseline, S-MotionGPT, and our proposed method, SOKE. The order of the signs was randomly shuffled to prevent potential bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.909, + 0.643 + ], + "angle": 0, + "content": "As shown in Figure 7, our method achieved an encouraging average rating of 5.32/6.45, significantly surpassing the baseline, which received a low score of 1.36/2.70 on How2Sign and CSL-Daily, respectively. These results align with the above qualitative and quantitative evaluations, indicating the potential of our method in constituting a two-way communication system between the deaf and hearing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.634, + 0.669 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose SOKE, a retrieval-enhanced multilingual sign language generator. Unlike existing methods that view sign language generation as a visual content generation task, we emphasize its linguistic nature, characterized by discrete semantic units. To capture the multi-cue characteristics of sign languages, we employ a decoupled tokenizer to map continuous motions into discrete tokens for the upper body and both hands. To balance efficiency and generation quality, we propose a novel multi-head decoding method that reduces decoding steps by two-thirds while dynamically integrating information from different body parts. Additionally, we present a simple yet effective retrieval-enhanced approach by leveraging external sign dictionaries. Extensive evaluations on a combined ASL-CSL-DGS dataset validate the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgements. S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1), EPSRC Project GNOMON (EP/X011364/1) and Turing AI Fellowship (EP/Z534699/1). R.A. Potamias and R. Zuo were supported by EPSRC Project GNOMON (EP/X011364/1). This research was also supported by the Sovereign AI AIRR programme. We would also like to thank anonymous signers who participated in the user study." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.241, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.482, + 0.307 + ], + "angle": 0, + "content": "[1] Rotem Shalev Arkushin, Amit Moryossef, and Ohad Fried. Ham2pose: Animating sign language notation into pose sequences. In CVPR, pages 21046-21056, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.309, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[2] Vasileios Baltatzis, Rolandos Alexandros Potamias, Evangelos Ververas, Guanxiong Sun, Jiankang Deng, and Stefanos Zafeiriou. Neural sign actors: A diffusion model for 3d sign language production from text. In CVPR, pages 1985-1995, 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.482, + 0.42 + ], + "angle": 0, + "content": "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.482, + 0.463 + ], + "angle": 0, + "content": "[4] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. In AAAI, pages 17754-17762, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.465, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[8] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. In ECCV, pages 93-110, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[9] Lu Dong, Lipisha Chaudhary, Fei Xu, Xiao Wang, Mason Lary, and Ifeoma Nwogu. Signavatar: Sign language 3d motion reconstruction and generation. In FG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[10] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In CVPR, pages 2735-2744, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[11] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In SIGKDD, pages 6491-6501, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Sen Fang, Chunyu Sui, Xuedong Zhang, and Yapeng Tian. SignDiff: Learning diffusion models for American sign language production. arXiv preprint arXiv:2308.16082, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.266, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[13] Sen Fang, Lei Wang, Ce Zheng, Yapeng Tian, and Chen Chen. Signllm: Sign languages production large language models. arXiv preprint arXiv:2405.10718, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.88, + 0.148 + ], + "angle": 0, + "content": "[14] fxsjy. https://github.com/fxsjy/jieba. In *Github*, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Aditya Golatkar, Alessandro Achille, Luca Zancato, Yu-Xiang Wang, Ashwin Swaminathan, and Stefano Soatto. Cpr: Retrieval augmented generation for copyright protection. In CVPR, pages 12374-12384, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[16] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In CVPR, pages 18362-18372, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[18] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.904, + 0.412 + ], + "angle": 0, + "content": "[19] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.904, + 0.455 + ], + "angle": 0, + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. SignBERT+: Hand-model-aware self-supervised pretraining for sign language understanding. TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[21] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In MM, pages 3172-3181, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.904, + 0.552 + ], + "angle": 0, + "content": "[22] Wencan Huang, Zhou Zhao, Jinzheng He, and Mingmin Zhang. Dualsign: Semi-supervised sign language production with balanced multi-modal multi-task dual transformation. In MM, pages 5486-5495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.904, + 0.594 + ], + "angle": 0, + "content": "[23] Euijun Hwang, Jung-Ho Kim, and Jong C Park. Nonautoregressive sign language production with gaussian space. In BMVC, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[24] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 36:20067-20079, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[25] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In EMNLP, pages 7969-7992, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.694, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[26] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[27] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In ECCV, pages 349-367, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[28] Hamid Reza Vaezi Joze and Oscar Koller. MS-ASL: A large-scale data set and benchmark for understanding American sign language. In BMVC, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[29] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23814" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[30] Gabriele Langer, Anke Müller, Sabrina Wähl, Felicitas Otte, Lea Sepke, and Thomas Hanke. Introducing the DW-DGS – the digital dictionary of DGS. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. NeurIPS, 33:9459-9474, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, pages 1459-1469, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[33] Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, and Houqiang Li. Uni-sign: Toward unified sign language understanding at scale. In ICLR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.402 + ], + "angle": 0, + "content": "[34] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[35] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. TACL, 8:726-742, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[37] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, pages 32939-32977, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[39] Mathias Müller, Zifan Jiang, Amit Moryossef, Annette Rios Gonzales, and Sarah Ebling. Considerations for meaningful sign language machine translation based on glosses. In ACL, pages 682-693, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[40] Zhe Niu, Ronglai Zuo, Brian Mak, and Fangyun Wei. A Hong Kong sign language corpus collected from sign-interpreted tv news. In LREC-COLING, pages 636-646, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[42] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[43] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In CVPR, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[44] Fan Qi, Yu Duan, Changsheng Xu, and Huaiwen Zhang. Signgen: End-to-end sign language video generation with latent diffusion. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[47] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In ECCV, pages 687-705, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[48] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Adversarial training for multi-channel sign language production. In BMVC, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[49] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3D multi-channel sign language production via progressive transformers and mixture density networks. IJCV, 129(7):2113-2135, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[50] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Signing at scale: Learning to co-articulate signs for large-scale photo-realistic sign language production. In CVPR, pages 5141-5151, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[51] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[52] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. knndiffusion: Image generation via large-scale retrieval. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[53] Tongkai Shi, Lianyu Hu, Fanhua Shang, Jichao Feng, Peidong Liu, and Wei Feng. Pose-guided fine-grained sign language video generation. In ECCV, pages 392-409, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[54] Stephanie Stoll, Armin Mustafa, and Jean-Yves Guillemaut. There and back again: 3d sign language generation from text using back-translation. In 3DV, pages 187–196, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[55] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[56] Sihan Tan, Taro Miyazaki, and Kazuhiro Nakadai. Multilingual gloss-free sign language translation: Towards building a sign language foundation model. In ACL, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[57] Shengeng Tang, Richang Hong, Dan Guo, and Meng Wang. Gloss semantic-enhanced network with online backtranslation for sign language production. In MM, pages 5630-5638, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[58] Shengeng Tang, Jiayi He, Lechao Cheng, Jingjing Wu, Dan Guo, and Richang Hong. Discrete to continuous: Generating smooth transition poses from sign language observations. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[59] Shengeng Tang, Jiayi He, Dan Guo, Yanyan Wei, Feng Li, and Richang Hong. Sign-idd: Iconicity disentangled diffusion for sign language production. In AAAI, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23815" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[60] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.164 + ], + "angle": 0, + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[62] Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, et al. Searching for best practices in retrieval-augmented generation. In EMNLP, pages 17716-17736, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[63] Fangyun Wei and Yutong Chen. Improving continuous sign language recognition with cross-lingual signs. In ICCV, pages 23612-23621, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[64] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[65] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI, pages 6234-6242, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[66] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In CVPR, pages 13525-13536, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.463 + ], + "angle": 0, + "content": "[67] Huijie Yao, Wengang Zhou, Hao Feng, Hezhen Hu, Hao Zhou, and Houqiang Li. Sign language translation with iterative prototype. In ICCV, pages 15592-15601, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[68] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[69] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. MLSLT: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[70] Aoxiong Yin, Haoyuan Li, Kai Shen, Siliang Tang, and Yueting Zhuang. T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text. In ACL, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[71] Kayo Yin, Amit Moryossef, Julie Hochgesang, Yoav Goldberg, and Malihe Alikhani. Including signed languages in natural language processing. In ACL, pages 7347-7360, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[72] Zhengdi Yu, Shaoli Huang, Yongkang Cheng, and Tolga Birdal. Signavatars: A large-scale 3d sign language holistic motion dataset and benchmark. In ECCV, pages 1-19, 2024. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[73] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In WACV, pages 3395-3403, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[74] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[75] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffeuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[76] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, pages 7368-7376, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[77] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. BEST: BERT pre-training for sign language recognition with coupling tokenization. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[78] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houqiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[79] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. IEEE TIP, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[80] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In ICCV, pages 20871-20881, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[81] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, pages 13009-13016, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[82] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[83] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, pages 1-1, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[84] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In CVPR, pages 1357-1366, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[85] Ronglai Zuo and Brian Mak. C2SLR: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[86] Ronglai Zuo and Brian Mak. Improving continuous sign language recognition with consistency constraints and signer removal. ACM TOMM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[87] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[88] Ronglai Zuo, Fangyun Wei, Zenggui Chen, Brian Mak, Jiaolong Yang, and Xin Tong. A simple baseline for spoken language to sign language translation with 3d avatars. In ECCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[89] Ronglai Zuo, Fangyun Wei, and Brian Mak. Towards online continuous sign language recognition and translation. In EMNLP, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23816" + } + ] +] \ No newline at end of file diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_origin.pdf b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b16c1a6e58b61a4dda682187caeab560c83287fe --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/1cb6aa16-71d9-4a9a-a1f6-c59910d7b13b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c6bb30db41ae3fe93f4d30ce3c4655172ad2831d1d98dbcf8d39544420e3a65 +size 3160085 diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/full.md b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ed5011ae47b1f294528e53773632e99df58ac5e --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/full.md @@ -0,0 +1,293 @@ +# Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator + +Ronglai Zuo, Rolandos Alexandros Potamias, Evangelos Ververas, Jiankang Deng, Stefanos Zafeiriou +Imperial College London + +https://2000zrl.github.io/soke/ + +![](images/d814b5c7743d60a34a033c47a1f9811edfb6a98285646d2e4d7a0bccb8355abd.jpg) +Figure 1. We propose Signs as Tokens (SOKE), a unified sign language generator that can produce multilingual signs from text inputs. (Left: American sign language; Middle: Chinese sign language; Right: German sign language.) + +# Abstract + +Sign language is a visual language that encompasses all linguistic features of natural languages and serves as the primary communication method for the deaf and hard-of-hearing communities. Although many studies have successfully adapted pretrained language models (LMs) for sign language translation (sign-to-text), the reverse task—sign language generation (text-to-sign)—remains largely unexplored. In this work, we introduce a multilingual sign language model, Signs as Tokens (SOKE), which can generate 3D sign avatars autoregressively from text inputs using a pretrained LM. To align sign language with the LM, we leverage a decoupled tokenizer that discretizes continuous signs into token sequences representing various body parts. During decoding, unlike existing approaches that flatten all part-wise tokens into a single sequence and predict one token at a time, we propose a multi-head decoding method capable of predicting multiple tokens simultaneously. This approach improves inference efficiency while maintaining effective information fusion across different body parts. To further ease the generation process, we propose a retrieval-enhanced SLG approach, which incorporates external sign dictionaries to provide accurate word-level signs as auxiliary conditions, significantly improving the precision of generated signs. Extensive qualitative and quantitative evaluations demonstrate the effectiveness of SOKE. + +# 1. Introduction + +Sign language is the primary communication method utilized by the deaf and hard-of-hearing communities, embodying all linguistic properties found in spoken languages, such as discrete semantic units and grammatical structure [16, 71]. As an attempt to mitigate the barrier between sign and spoken languages, several methods have been developed to tackle sign language processing by unfolding the task into two research problems: sign language translation (SLT, sign-to-text) [20, 33, 56, 64, 67, 69, 80, 83] and sign language generation (SLG, text-to-sign) [2, 47, 53, 72, 88]. + +In contrast with SLT, the usage of language models (LMs) in SLG remains relatively unexplored. Most existing methods approach SLG as a visual content generation task (e.g., video, keypoint, or motion), using GANs [50, 53] or diffusion models [2, 12, 44, 58, 59]. However, these works neglect the linguistic nature of sign languages, leading to suboptimal performance and missing benefits of pretrained LMs such as generalizability and scalability [18, 29, 60]. In particular, sign languages share the fundamental linguistic properties and the discrete structure of spoken languages [16, 71], which necessitates processing using autoregressive language modeling techniques [70]. + +To formulate SLG in the framework of language modeling, a crucial step is mapping signs into discrete tokens. A common approach to discretizing sign languages involves + +using glosses—the written form of signs—as an intermediate representation [13, 22, 50, 88]. However, glosses require extensive annotation efforts and impose a predefined information bottleneck, which fail to fully capture the rich semantics of sign languages [39]. To address these limitations and better capture the multi-cue property of sign languages [26, 81, 85], we leverage a vector quantized-variational auto-encoder (VQ-VAE)-based tokenizer [37, 65, 68] that learns mappings between continuous sign motions and discrete tokens for each body part (upper body, left hand, and right hand). During decoding, existing autoregressive motion generation methods [37, 68] typically flatten tokens from different body parts into a single sequence, resulting in inefficient inference due to tripled decoding steps. To overcome this challenge, we propose a novel multi-head decoding method capable of predicting multiple tokens at a time while effectively fusing information from different body parts in real time. Furthermore, given the limited vocabulary size of continuous sign language datasets, we enhance the model's generalization by instantiating the LM as a multilingual LM [35], enabling it to handle a broader and more diverse domain of discourse. To support multilingual SLG in a unified framework, we train the proposed model on a curated multilingual sign language dataset encompassing American, Chinese, and German sign languages [3, 10, 82]. + +Recently, retrieval-augmented generation (RAG) [4, 11, 62] has proven effective in enhancing the interpretability and accuracy of content generated by large language models. Similarly, several SLG methods [50, 88] leverage glosses to retrieve sign dictionaries and directly incorporate retrieved signs into the outputs, improving the interpretability of generated signs. However, beyond the inherent limitations of glosses, these methods often face challenges in handling co-articulations between adjacent signs. This is because dictionary signs are typically recorded at the word level, whereas SLG models operate at the sentence level, leading to potential unnaturalness in the generated sequences. To address these issues, we propose a retrieval-enhanced SLG approach that utilizes motion tokens from retrieved dictionary signs as additional conditions for language modeling. This approach avoids the unnaturalness of directly copying signs into model outputs while still benefiting from the precision of dictionary signs. To sum up, the contributions of this paper can be summarized as: + +- We propose a novel SLG approach, Signs as Tokens (SOKE). With the aid of a pretrained LM, SOKE can handle multiple sign languages in a unified model. +- We propose a multi-head decoding method that reduces inference latency and efficiently fuses information from different body parts, enabling seamless integration of a decoupled tokenizer with the LM. +- By leveraging external sign dictionaries, we introduce a retrieval-enhanced SLG approach to improve the preci + +sion of generated signs. + +- We curate two sign language datasets [3, 82] with accurate SMPL-X [42] poses to facilitate multilingual SLG research. Experiments demonstrate that our SOKE achieves state-of-the-art performance on three challenging benchmarks, How2Sign [10], CSL-Daily [82], and Phoenix2014T [3], using a single unified model. + +# 2. Related Work + +Sign Language Generation. Early gloss-based methods [21, 22, 50, 57, 65, 73, 88] approach SLG under a text-to-gloss-to-sign framework. Several works [1, 47, 49] propose to directly regress 2D joints to represent sign poses, simplifying the modeling process, however, they suffer from mode collapse due to the limited scale of data. Recent state-of-the-art SLG works can be categorized into two classes: the first group of methods [2, 12] employs diffusion models to generate sign motions conditioned on text inputs; the second group of methods [9, 70, 72] adopts a tokenizer-LM two-stage autoregressive generation approach. Among these, [72] incorporates costly gloss and HamNoSys notations [1] to train a semantic adapter, while [9] focuses on generating signs from a single word, limiting their applicability. Moreover, a notable gap exists in the literature regarding unified models for multilingual SLG. In this work, we address these gaps by utilizing a decoupled tokenizer and a pretrained multilingual LM, integrated with our customized multi-head decoding method and retrieval-enhanced approach, to enable text-driven multilingual SLG. Text-to-Motion Generation. Text-to-motion generation (T2MG) is an important research area with potential applications in human-computer interaction and robotics. Recent T2MG methods can be categorized into two main types: 1) Diffusion-based approaches [5, 51, 75], which utilize latent diffusion [46] to generate motions conditioned on text inputs. 2) Autoregressive approaches [17, 24, 37, 74, 76, 84], which draw inspiration from text generation methods by first tokenizing motion sequences and then using LMs to perform autoregressive generation. When decoding with a decoupled tokenizer [37, 68], these autoregressive approaches simply flatten tokens from various body parts into one sequence, causing additional decoding steps. In this work, we propose a novel multi-head decoding method that can efficiently predict motion tokens for all body parts at one single step, significantly reducing inference latency. + +Retrieval-Enhanced Generation. Using the concept of retrieval to enhance generation models has proven effective across various domains [4, 8, 11, 15, 62, 66]. As a pioneering effort, Piktus et al. [31] propose to retrieve relevant documents before generating answers for knowledge-intensive NLP tasks. FLARE [25] introduces selective retrieval, activating it only when low-confidence tokens appear in the generation process. ToG [55] leverages retrieved + +![](images/a38ba151181b6e68a215bb785e08ce9f460b432080e8340715bac1ed9c7182d5.jpg) +Figure 2. An overview of our proposed method, Signs as Tokens (SOKE). We begin by training a VQ-VAE-based decoupled tokenizer to map continuous sign motions into discrete tokens for various body parts (upper body, left hand, and right hand). These sign motion tokens are then integrated into the text vocabulary of a pretrained language model, which serves as the backbone of our autoregressive multilingual generator (AMG). Given a text input, the AMG first retrieves word-level signs from external dictionaries, appends their motion tokens to the text tokens, and feeds them into the language model encoder. During decoding, our novel multi-head decoding strategy generates motion tokens for all body parts simultaneously at each time step. Finally, the derived motion tokens are used to reconstruct sign avatars. + +knowledge to enhance the reasoning capabilities of large language models. In the field of image generation, retrieval has also been utilized to enable diffusion models to generate out-of-distribution images [52]. In the context of SLG, several methods [50, 88] adopt gloss-based retrieval to obtain dictionary signs and approach SLG in a retrieve-and-concatenate manner. However, glosses have inherent limitations, and such methods often suffer from unrealistic transitions between signs. In this paper, we propose a simple yet effective retrieval-enhanced SLG approach that uses word-level signs as auxiliary conditions, addressing these limitations and improving the overall quality of generations. + +# 3. Methodology + +An overview of our method is shown in Figure 2. Our approach mainly consists of two parts: a decoupled tokenizer (DETO) and an autoregressive multilingual generator (AMG). In the first stage, DETO learns mappings between continuous sign motions and discrete tokens for each body part (Section 3.2). In the second stage, AMG is trained to autoregressively generate motion tokens with the proposed multi-head decoding strategy (Section 3.3) conditioned on text inputs and retrieved dictionary signs (Section 3.4). + +# 3.1. Data Preparation + +Although existing SLG works [2, 72] have made efforts in curating high-quality SMPL-X poses for an American sign language (ASL) dataset, How2Sign [10], there is a notable lack of 3D annotations for other continuous sign language datasets, impeding multilingual SLG research. In this work, we rely on a two-stage approach, leveraging state-of-the-art 3D pose estimation methods [34, 43], to extract high-fidelity SMPL-X poses for CSL-Daily [82] and Phoenix-2014T [3], two widely-adopted datasets of Chinese and German sign languages (CSL and DGS), respec + +tively. Specifically, we initially estimate a set of rough 3D body poses for each signer using OSX [34] along with accurate 3D hand poses from WiLoR [43]. Given that OSX often falls short in accurately capturing the arm poses, we follow [2] and refine the upper body joint rotations by minimizing a re-projection loss between the estimated joints and the detected 2D joint keypoints from Mediapipe [38]. Using this process we can ensure accurate whole-body annotations. We represent each sign motion sequence as $\mathbf{S} \in \mathbb{R}^{T \times d}$ , where $T$ is the sequence length and $d = 133$ denotes the number of SMPL-X parameters, including 11 upper body joints, 30 hand joints, and 10 expression parameters. Please check the supplementary materials for more details and quantitative evaluations of our pose fitting pipeline. + +# 3.2. Decoupled Tokenizer + +Sign languages exhibit a multi-cue property [19, 26, 77, 79, 81, 86], where the semantics of a sign are conveyed simultaneously through body and hand movements. Inspired by this, we leverage a decoupled tokenizer [37, 65, 68], comprising three VQ-VAEs [61], to independently model key regions: the upper body and both hands. An illustration is available in the supplementary materials. + +Given a $T$ -frame sign motion input $\mathbf{S} \in \mathbb{R}^{T \times d}$ , we first decompose it into three part-wise motion sequences based on the kinematic tree of SMPL-X: $\mathbf{S}^p \in \mathbb{R}^{T \times d_p}$ , where $p \in \{B, LH, RH\}$ . Moreover, we build three distinct VQ-VAEs, where each of them consists of an encoder $\mathcal{E}_p(\cdot)$ , a decoder $\mathcal{D}_p(\cdot)$ , and a learnable codebook $\mathbf{Z}^p \in \mathbb{R}^{N_Z^p \times C}$ where $N_Z^p$ represents the number of codes and $C$ denotes the code dimension. For each motion sequence, the corresponding encoder first projects the sequence into a latent space: $\mathbf{S}_f^p = \mathcal{E}_p(\mathbf{S}^p) = \{s_{f,i}^p\}_{i=1}^{T_f} \in \mathbb{R}^{T_f \times C}$ , using a stack of 1D-CNN layers. Then, for each pose we can derive a set of discrete tokens $\hat{\mathbf{Z}}^p = \{\hat{z}_i^p\}_{i=1}^{T_f}$ using a quantizer $Q(\cdot)$ , + +![](images/2c523adcc341bf8773ce91323acf95e297bb6a697d2e6b77176915f82d8d6b61.jpg) +(a) Sequential Decoding $(O(3K))$ + +![](images/0e7520f580bbb11ae4a354f4c4508f563b4f5f500a7bf4858ceeb6c221e29224.jpg) +(b) Parallel Decoding $(O(K))$ +Figure 3. Comparison of various decoding methods. (a) Sequential decoding flattens all motion tokens into a single sequence, requiring an inefficient $3K$ decoding steps, where $K$ is the target motion sequence length. (b) Parallel decoding improves efficiency by decoding body parts in parallel but lacks information fusion, which may lead to suboptimal performance. (c) Our proposed multi-head decoding balances efficiency and fusion by predicting three tokens simultaneously via different heads, with input embeddings as a weighted average of part-wise token embeddings. + +![](images/51b151d1ffd557f7f08d9bbe831e08c7caaf1bb81f20b19c98f4696767510ba7.jpg) +(c) Multi-head Decoding (ours, $O(K)$ ) + +which searches the nearest neighbor from the codebook $\mathbf{Z}^p$ : + +$$ +\hat {z} _ {i} ^ {p} = Q \left(s _ {f, i} ^ {p}\right) = \arg \min _ {z _ {j} \in \mathbf {Z} ^ {p}} \| s _ {f, i} ^ {p} - z _ {j} \| _ {2}, j \in [ 1, N _ {Z} ^ {p} ]. \tag {1} +$$ + +We then feed the obtained token sequence to the corresponding part-decoder $\mathcal{D}_p$ to reconstruct the input motions: $\hat{\mathbf{S}}^p = \mathcal{D}_p(\mathbf{Z}^p)$ . Following [24], the objective function of each VQ-VAE is defined as: $\mathcal{L}_{vq}^p = \mathcal{L}_{rec}^p +\mathcal{L}_{emb}^p +\mathcal{L}_{com}^p$ where $\mathcal{L}_{rec}^p = \| \hat{\mathbf{S}}^p -\mathbf{S}^p\| _2^2$ denotes the reconstruction loss, $\mathcal{L}_{emb}^p$ is the embedding loss, and $\mathcal{L}_{com}^p$ is the commitment loss, respectively. + +# 3.3. Autoregressive Multilingual Generator + +Using obtained discrete sign representations, we can respect the discrete properties of sign languages and approach SLG similarly to text generation with a pretrained LM [35]. + +Integrated Vocabulary. We first gather the discrete tokens for different body parts to construct a part-wise motion vocabulary: $\mathcal{V}_m^p = \{\langle p\_ i\rangle \}_{i = 1}^{N_Z^p}$ , e.g., $\langle B\_ 1\rangle$ and $\langle \mathrm{RH\_2}\rangle$ . The overall motion vocabulary is a combination of three part-wise vocabularies: $\mathcal{V}_m = \{\mathcal{V}_m^B,\mathcal{V}_m^{LH},\mathcal{V}_m^{RH}\}$ . Following [35], we further define a series of language identifier tokens: $\mathcal{V}_l = \{\langle \mathrm{ASL}\rangle ,\langle \mathrm{CSL}\rangle ,\langle \mathrm{DGS}\rangle \}$ , to prompt the LM with the information of target sign language. The integrated vocabulary for the generator is defined as: $\mathcal{V} = \{\mathcal{V}_t,\mathcal{V}_m,\mathcal{V}_l\}$ , which lays a foundation for the subsequent LM fine-tuning. + +Sequential Decoding. Following previous motion generation works [37, 68], a naive decoding method is simply flattening tokens from each body part into a single sequence with tripled length: $\mathbf{Y} = \{y_1^B, y_1^{LH}, y_1^{RH}, \ldots, y_K^B, y_K^{LH}, y_K^{RH}\}$ , where $y_k^p \in \mathcal{V}_m^p$ denotes the part-wise motion token, and $K$ denotes the length of the target sign motion sequence. The decoding process can then be formulated as: + +$$ +\begin{array}{l} P (\mathbf {Y} \mid \mathbf {h} _ {e n}) = \prod_ {k = 1} ^ {K} P \left(y _ {k} ^ {B, L H, R H} \mid y _ {< k} ^ {B, L H, R H}\right) \tag {2} \\ = \prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\{p \}}) P (y _ {k} ^ {L H} | y _ {k} ^ {B}, y _ {< k} ^ {\{p \}}) P (y _ {k} ^ {R H} | y _ {k} ^ {B, L H}, y _ {< k} ^ {\{p \}}), \\ \end{array} +$$ + +where $\mathbf{h}_{en}$ denotes the last encoder hidden states. For simplicity, we write $y_{$ when the model is required to generate upper-body motions for ASL. The decoding process can then be formulated as: + +$$ +\begin{array}{l} P (\mathbf {Y} | \mathbf {h} _ {e n}) = P (\mathbf {Y} ^ {B}) P (\mathbf {Y} ^ {L H}) P (\mathbf {Y} ^ {R H}) \tag {3} \\ = \prod_ {k = 1} ^ {K} P \left(y _ {k} ^ {B} \mid y _ {< k} ^ {B}\right) \prod_ {k = 1} ^ {K} P \left(y _ {k} ^ {L H} \mid y _ {< k} ^ {L H}\right) \prod_ {k = 1} ^ {K} P \left(y _ {k} ^ {R H} \mid y _ {< k} ^ {R H}\right). \\ \end{array} +$$ + +Comparing Eq. 3 with Eq. 2, the parallel decoding strategy significantly reduces the dependence on previous tokens, as the decoding processes for each body part are completely separated. While this formulation enables efficient parallel execution, the overly independent assumption may result in suboptimal performance. + +![](images/1f2990349f912a1833ba2b86aa4ec4f89581f242e528731e8ba4dc3f38480cb2.jpg) +Figure 4. Illustration of retrieval-enhanced SLG. We utilize external sign dictionaries to retrieve accurate word-level signs, which serve as additional conditions to enhance the generation quality. + +Multi-Head Decoding. To balance efficiency and information fusion across body parts, we propose a multi-head decoding strategy. As shown in Figure 3(c), we design three language modeling heads, implemented as fully connected layers, to predict motion tokens for each body part simultaneously at each step. Moreover, the decoder inputs at each step are modified to a weighted average of the token embeddings from different body parts. Specifically, denoting the token embeddings from upper body, left hand, and right hand as $\mathbf{E}^B$ , $\mathbf{E}^{LH}$ , and $\mathbf{E}^{RH}$ , respectively, the input embeddings are defined as: $\mathbf{E} = (1 - 2\lambda)\mathbf{E}^B + \lambda \mathbf{E}^{LH} + \lambda \mathbf{E}^{RH}$ where $\lambda \in (0,0.5)$ is a hyper-parameter to control the weight of hand embeddings. The decoding process can then be formulated as: + +$$ +\begin{array}{l} P (\mathbf {Y} | \mathbf {h} _ {e n}) = \prod_ {k = 1} ^ {K} P \left(y _ {k} ^ {B, L H, R H} \mid y _ {< k} ^ {B, L H, R H}\right) \tag {4} \\ = \prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\{p \}}) P (y _ {k} ^ {L H} | y _ {< k} ^ {\{p \}}) P (y _ {k} ^ {R H} | y _ {< k} ^ {\{p \}}), \\ \end{array} +$$ + +which can be viewed as relaxing Eq. 2 under a conditional independence assumption. + +Training and Inference. The LM is trained with a standard cross-entropy loss: $\mathcal{L}_{LM} = -\log P(\mathbf{Y}|\mathbf{h_{en}})$ . During inference, we adopt a simple greedy decoding algorithm that the LM always outputs the token with the highest probability at each step. For parallel and multi-head decoding, the inference process terminates as soon as any head or decoder predicts an end-of-sentence token. The obtained token sequences will be fed into the corresponding part-decoder, $\mathcal{D}_p(\cdot)$ , to reconstruct sign motions. + +# 3.4. Retrieval-Enhanced SLG + +Inspired by the effectiveness of retrieval-augmented generation in NLP [4, 11, 62], we propose to leverage external sign dictionaries to improve the fidelity of generated signs. + +As shown in Figure 4, we first build a word-level sign dictionary for each target sign language by utilizing resources from isolated sign language recognition datasets + +[28, 30, 32, 78, 87] and online sources. Through our pose fitting pipeline, each RGB video in these dictionaries is converted into SMPL-X poses. Next, we employ a well-trained decoupled tokenizer to map the poses into discrete tokens. The resulting dictionary is represented as a set of quadruples $\{(w, m^B, m^{LH}, m^{RH})\}$ , where $w$ denotes lemmatized words in the written language, and $m^p$ denotes motion tokens. If a word has multiple sign instances in the dictionary, we retain only the instance with the lowest reconstruction error after passing through the tokenizer. + +Given a text input $\mathbf{X} = \{x_{i}\}_{i = 1}^{L}$ , consisting of $L$ words $^1$ , we collect all words and corresponding tokens found in the dictionary. The final prompt is created by concatenating raw texts $\mathbf{X}$ with the retrieved motion tokens $\{m_j^p\}$ , which is then fed into the LM encoder to generate hidden states $\mathbf{h}_{en}$ . Several examples are shown in Figure 4. + +# 4. Experiments + +Datasets. We construct a multilingual sign language dataset by merging three widely-adopted datasets: How2Sign [10], CSL-Daily [82], and Phoenix-2014T [3]. How2Sign is a large-scale ASL dataset, consisting of 35K video-text pairs. CSL-Daily is the largest existing continuous CSL dataset, containing 20K samples. Phoenix-2014T is a DGS dataset with 8K samples. To train SOKE, we use the SMPL-X poses of How2Sign dataset as provided in [2] along with the curated poses of CSL-Daily and Phoenix-2014T. + +Evaluation Metrics. Since the length of generated signs may differ from the ground truth, we employ the well-established dynamic time warping (DTW) [1, 2] on both procrustes-aligned and original joint position errors (JPE) to measure sequence-level distances between the generated signs and ground truth. Following [2], we also report backtranslation (B-T) scores in BLEU-4 [41] to assess the interpretability of the generated signs. + +Implementation Details. For DETO, we empirically set the code numbers to $N_Z^B = 96$ for the body and $N_Z^{LH} = N_Z^{RH} = 192$ for the left and right hands, respectively, with a code dimension of $C = 512$ . We train DETO on the combined multilingual sign language dataset and sign dictionaries with a batch size of 320 per GPU for 500 epochs, using the AdamW optimizer [36] and a cosine learning rate scheduler starting at 2e-4. For the LM, we use mBART-large-cc25 [35], which features 12 encoder-decoder layers with a model dimension of 1024 and is widely used in various sign language translation methods [6, 7, 27, 63, 89]. We empirically set $\lambda = 1/3$ to balance part-wise token embeddings. The LM is fine-tuned with a batch size of 32 per GPU on the multilingual sign language dataset for 150 epochs, employing the same optimizer settings as DETO. All models are trained using 6 RTX 3090 GPUs. + +![](images/82e67d16ec45b1a5e39c12388e13c2b566ac884f601d1fe8ab4dee46247fdcb9.jpg) +Figure 5. Qualitative comparisons of generated signs between our proposed method, SOKE, with the SOTA method, S-MotionGPT [24], on the test sets of How2Sign (left), CSL-Daily (middle), and Phoenix-2014T (right). + +# 4.1. Comparison with State-of-the-Art Methods + +Qualitative Comparison. We first reimplement a state-of-the-art (SOTA) open-sourced motion generation approach, MotionGPT [24], which utilizes a VQ-VAE to discretize whole-body motions and employs a monolingual language model [45], on sign language datasets. The resulting model, S-MotionGPT, achieves competitive results on sign + +language benchmarks (Table 1). In Figure 5, we perform a qualitative comparison between our SOKE and S-MotionGPT. As can be easily seen, our proposed method produces significantly better results, with visual quality comparable to the ground truth sign videos across the ASL, CSL, and DGS datasets. Notably, our method can generate precise and informative hand movements, which play a key role in conveying semantics in sign languages. Additionally, the proposed method supports three sign languages + +
MethodMulti-lingualHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑
BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4
NAR [23]×13.9411.80--5.75----------
Prog. Trans.* [47]×14.1511.5714.7430.172.7515.9812.9116.3032.633.0713.6711.9515.0131.774.94
Text2Mesh* [54]×13.9913.4715.5032.977.5113.4712.1013.7630.375.1113.4812.0614.0431.645.81
Adv. Train. [48]×13.7811.17--6.21----------
T2S-GPT* [70]×11.486.3912.6518.4411.2011.945.9312.3215.438.9410.386.4711.6519.099.06
NSA [2]×7.837.33--13.12----------
S-MotionGPT* [24]×11.234.3912.4113.7411.4510.813.7811.5811.318.829.453.4110.429.089.68
SOKE (ours)✓6.822.357.7510.0814.486.241.717.389.6811.304.771.386.047.7211.87
+ +Table 1. Comparison with state-of-the-art sign language generation (text-to-sign) methods. * denotes reimplementation results. + +
Decoding MethodSign RetrievalHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓
AvgBodyHands/videoAvgBodyHands/videoAvgBodyHands/video
Sequential×4.648.553.533.263.748.142.493.153.266.972.203.23
Parallel×5.069.273.861.394.198.612.931.413.547.022.551.35
Multi-head×4.177.913.101.463.377.582.171.452.806.161.851.44
Multi-head✓3.346.822.351.552.726.241.711.522.134.771.381.51
+ +Table 2. Ablation study for the decoding method and sign retrieval. Latencies are measured using a single RTX 3090 GPU. + +simultaneously, whereas S-MotionGPT is monolingual and requires separate models for each sign language dataset. + +Quantitative Comparison. In addition to the above qualitative comparison, we also perform a quantitative evaluation against SOTA SLG methods, as detailed in Table 1. Our reimplemented S-MotionGPT achieves competitive results, with low DTW errors of 4.39 and 3.78 on How2Sign and CSL-Daily, respectively, demonstrating the effectiveness of formulating SLG as a language modeling task. The previous best SLG approach, Neural Sign Actors (NSA) [2], formulates SLG as a motion generation task and employs a diffusion model as its backbone. However, this approach overlooks the linguistic properties of sign languages, resulting in higher average DTW errors compared to S-MotionGPT. In contrast, our method sets a new state-of-the-art across all three benchmarks, achieving significantly lower DTW errors of 2.35, 1.71, and 1.38 on How2Sign, CSL-Daily, and Phoenix-2014T, respectively, using a single unified model. + +# 4.2. Ablation Study + +Decoding Method. We compare the performance of different decoding methods in Table 2. The traditional sequential decoding method is widely used in existing motion generation approaches [37, 68]. However, since tokens from different body parts are flattened into a single sequence, the number of decoding steps is tripled, leading to high inference latency (3.26s per video on How2Sign). By using three separate LM decoders, parallel decoding predicts token sequences for each body part simultaneously. This formulation significantly reduces latency (1.39s per video on How2Sign) but results in higher DTW errors due to the + +![](images/685c6bd093b9d02bfc96fdac36184f478096bf7a4b1b7e62eacdb2628a96a28c.jpg) +Figure 6. Qualitative ablation study for retrieval-enhanced SLG. (Left: How2Sign; Middle: CSL-Daily; Right: Phoenix-2014T.) + +lack of information fusion across part-wise motion tokens. Notably, our multi-head decoding method achieves a better trade-off between efficiency and information integration, maintaining comparable latency to parallel decoding while delivering even superior generation quality than sequential decoding. We attribute this to the weighting mechanism (soft gating): since the decoder inputs are a weighted average of all part-wise tokens, errors in a single token may + +
H2SCSLPh-TGenerationReconstruction
H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)H2S (JPE↓)CSL (JPE↓)Ph-T (JPE↓)
BodyHandBodyHandBodyHandBodyHandBodyHandBodyHand
✓7.923.07----21.527.74----
✓✓7.112.636.792.14--19.886.9724.295.83--
✓✓✓6.822.356.241.714.771.3819.376.6523.525.1325.796.78
+ +Table 3. Study on the scalability of our approach. We use procrustes-aligned mean per joint position error (PA-MPJPE) [34, 72] to assess reconstruction performance of the decoupled tokenizer. (Ph-T: Phoenix-2014T.) + +
VAE-1VAE-2VAE-3H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)
BodyHandBodyHandBodyHand
Whole-body××9.493.569.223.358.033.23
BodyHands×7.993.136.932.856.242.18
BodyL-handR-hand6.822.356.241.714.771.38
+ +Table 4. Ablation study on the decoupled tokenizer. + +have a reduced impact on the overall performance. + +Sign Retrieval. By leveraging external sign dictionaries, we develop a retrieval-enhanced SLG approach that incorporates retrieved word-level signs as additional generation conditions. As shown in Table 2, this retrieval mechanism significantly enhances the precision of generated signs, reducing the average DTW errors by $19.9\% / 19.6\% / 23.9\%$ on How2Sign, CSL-Daily, and Phoenix-2014T, respectively. Furthermore, we perform a qualitative comparison in Figure 6. The figure demonstrates that the retrieved signs effectively guide the model's generation: similar pose patterns from the retrieved signs, particularly in hand shapes, appear in the generated outputs. This guidance substantially improves the precision of the generated signs. + +Scalability with Multilingual Data. Results in Table 3 show that incorporating multilingual data significantly improves generation performance, demonstrating the scalability of our approach. This improvement stems from two key factors: 1) the tokenizer benefits from more pose training data (right half of Table 3), which subsequently enhances the generator's performance; 2) the presence of cross-lingual signs, as revealed in [63]. For example, the sign for "cold" in CSL and DGS shares similar body movements. Training on multilingual data helps the model better understand and generalize across these cross-lingual signs. + +Decoupled Tokenizer. Inspired by the multi-cue characteristics of sign languages, we introduce a decoupled tokenizer that discretizes sign motion sequences for different body parts. In Table 4, we present an ablation study on various configurations: a single VQ-VAE for whole-body motions and one VAE for the body and another for both hands. The findings indicate that progressively separating the body and hands leads to consistently improved performance. + +# 4.3. User Study + +In addition to the objective metrics presented in the previous tables, we conducted a user study involving 5 professional + +![](images/c35c6ee4f9fbc69d64e8273383c231c232e8706ef84d9a7677c68d5358a69fd0.jpg) +Figure 7. User study with professional ASL and CSL signers. We report average ratings of S-MotionGPT, our proposed SOKE, and ground truth motions. + +ASL signers and 4 professional CSL signers. These signers were asked to rate the alignment between generated signs and text annotations on a scale from 1 to 10, with higher scores indicating better semantic conveyance of the texts. Ground truth sign motions were also included in the study. Specifically, we provided 15 generated signs from both the baseline, S-MotionGPT, and our proposed method, SOKE. The order of the signs was randomly shuffled to prevent potential bias. + +As shown in Figure 7, our method achieved an encouraging average rating of 5.32/6.45, significantly surpassing the baseline, which received a low score of 1.36/2.70 on How2Sign and CSL-Daily, respectively. These results align with the above qualitative and quantitative evaluations, indicating the potential of our method in constituting a two-way communication system between the deaf and hearing. + +# 5. Conclusion + +In this paper, we propose SOKE, a retrieval-enhanced multilingual sign language generator. Unlike existing methods that view sign language generation as a visual content generation task, we emphasize its linguistic nature, characterized by discrete semantic units. To capture the multi-cue characteristics of sign languages, we employ a decoupled tokenizer to map continuous motions into discrete tokens for the upper body and both hands. To balance efficiency and generation quality, we propose a novel multi-head decoding method that reduces decoding steps by two-thirds while dynamically integrating information from different body parts. Additionally, we present a simple yet effective retrieval-enhanced approach by leveraging external sign dictionaries. Extensive evaluations on a combined ASL-CSL-DGS dataset validate the effectiveness of our method. + +Acknowledgements. S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1), EPSRC Project GNOMON (EP/X011364/1) and Turing AI Fellowship (EP/Z534699/1). R.A. Potamias and R. Zuo were supported by EPSRC Project GNOMON (EP/X011364/1). This research was also supported by the Sovereign AI AIRR programme. We would also like to thank anonymous signers who participated in the user study. + +# References + +[1] Rotem Shalev Arkushin, Amit Moryossef, and Ohad Fried. Ham2pose: Animating sign language notation into pose sequences. In CVPR, pages 21046-21056, 2023. 2, 5 +[2] Vasileios Baltatzis, Rolandos Alexandros Potamias, Evangelos Ververas, Guanxiong Sun, Jiankang Deng, and Stefanos Zafeiriou. Neural sign actors: A diffusion model for 3d sign language production from text. In CVPR, pages 1985-1995, 2024. 1, 2, 3, 5, 7 +[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, 2018. 2, 3, 5 +[4] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. In AAAI, pages 17754-17762, 2024. 2, 5 +[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2 +[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 5 +[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 5 +[8] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. In ECCV, pages 93-110, 2024. 2 +[9] Lu Dong, Lipisha Chaudhary, Fei Xu, Xiao Wang, Mason Lary, and Ifeoma Nwogu. Signavatar: Sign language 3d motion reconstruction and generation. In FG, 2024. 2 +[10] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In CVPR, pages 2735-2744, 2021. 2, 3, 5 +[11] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In SIGKDD, pages 6491-6501, 2024. 2, 5 +[12] Sen Fang, Chunyu Sui, Xuedong Zhang, and Yapeng Tian. SignDiff: Learning diffusion models for American sign language production. arXiv preprint arXiv:2308.16082, 2023. 1, 2 + +[13] Sen Fang, Lei Wang, Ce Zheng, Yapeng Tian, and Chen Chen. Signllm: Sign languages production large language models. arXiv preprint arXiv:2405.10718, 2024. 2 +[14] fxsjy. https://github.com/fxsjy/jieba. In *Github*, 2012. 5 +[15] Aditya Golatkar, Alessandro Achille, Luca Zancato, Yu-Xiang Wang, Ashwin Swaminathan, and Stefano Soatto. Cpr: Retrieval augmented generation for copyright protection. In CVPR, pages 12374-12384, 2024. 2 +[16] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In CVPR, pages 18362-18372, 2024. 1 +[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2 +[18] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 1 +[19] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 3 +[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. SignBERT+: Hand-model-aware self-supervised pretraining for sign language understanding. TPAMI, 2023. 1 +[21] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In MM, pages 3172-3181, 2021. 2 +[22] Wencan Huang, Zhou Zhao, Jinzheng He, and Mingmin Zhang. Dualsign: Semi-supervised sign language production with balanced multi-modal multi-task dual transformation. In MM, pages 5486-5495, 2022. 2 +[23] Euijun Hwang, Jung-Ho Kim, and Jong C Park. Nonautoregressive sign language production with gaussian space. In BMVC, 2021. 7 +[24] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 36:20067-20079, 2023. 2, 4, 6, 7 +[25] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In EMNLP, pages 7969-7992, 2023. 2 +[26] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 3 +[27] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In ECCV, pages 349-367, 2024. 5 +[28] Hamid Reza Vaezi Joze and Oscar Koller. MS-ASL: A large-scale data set and benchmark for understanding American sign language. In BMVC, 2019. 5 +[29] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 1 + +[30] Gabriele Langer, Anke Müller, Sabrina Wähl, Felicitas Otte, Lea Sepke, and Thomas Hanke. Introducing the DW-DGS – the digital dictionary of DGS. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, 2024. 5 +[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. NeurIPS, 33:9459-9474, 2020. 2 +[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, pages 1459-1469, 2020. 5 +[33] Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, and Houqiang Li. Uni-sign: Toward unified sign language understanding at scale. In ICLR, 2025. 1 +[34] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 3, 8 +[35] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. TACL, 8:726-742, 2020. 2, 4, 5 +[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5 +[37] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, pages 32939-32977, 2024. 2, 3, 4, 7 +[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 3 +[39] Mathias Müller, Zifan Jiang, Amit Moryossef, Annette Rios Gonzales, and Sarah Ebling. Considerations for meaningful sign language machine translation based on glosses. In ACL, pages 682-693, 2023. 2 +[40] Zhe Niu, Ronglai Zuo, Brian Mak, and Fangyun Wei. A Hong Kong sign language corpus collected from sign-interpreted tv news. In LREC-COLING, pages 636-646, 2024. 6 +[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5 +[42] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2 +[43] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In CVPR, 2025. 3 + +[44] Fan Qi, Yu Duan, Changsheng Xu, and Huaiwen Zhang. Signgen: End-to-end sign language video generation with latent diffusion. In ECCV, 2024. 1 +[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 6 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[47] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In ECCV, pages 687-705, 2020. 1, 2, 7 +[48] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Adversarial training for multi-channel sign language production. In BMVC, 2020. 7 +[49] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3D multi-channel sign language production via progressive transformers and mixture density networks. IJCV, 129(7):2113-2135, 2021. 2 +[50] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Signing at scale: Learning to co-articulate signs for large-scale photo-realistic sign language production. In CVPR, pages 5141-5151, 2022. 1, 2, 3 +[51] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2 +[52] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. knndiffusion: Image generation via large-scale retrieval. In ICLR, 2023. 3 +[53] Tongkai Shi, Lianyu Hu, Fanhua Shang, Jichao Feng, Peidong Liu, and Wei Feng. Pose-guided fine-grained sign language video generation. In ECCV, pages 392-409, 2024. 1 +[54] Stephanie Stoll, Armin Mustafa, and Jean-Yves Guillemaut. There and back again: 3d sign language generation from text using back-translation. In 3DV, pages 187–196, 2022. 7 +[55] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In ICLR, 2024. 2 +[56] Sihan Tan, Taro Miyazaki, and Kazuhiro Nakadai. Multilingual gloss-free sign language translation: Towards building a sign language foundation model. In ACL, 2025. 1 +[57] Shengeng Tang, Richang Hong, Dan Guo, and Meng Wang. Gloss semantic-enhanced network with online backtranslation for sign language production. In MM, pages 5630-5638, 2022. 2 +[58] Shengeng Tang, Jiayi He, Lechao Cheng, Jingjing Wu, Dan Guo, and Richang Hong. Discrete to continuous: Generating smooth transition poses from sign language observations. In CVPR, 2025. 1 +[59] Shengeng Tang, Jiayi He, Dan Guo, Yanyan Wei, Feng Li, and Richang Hong. Sign-idd: Iconicity disentangled diffusion for sign language production. In AAAI, 2025. 1 + +[60] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1 +[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3 +[62] Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, et al. Searching for best practices in retrieval-augmented generation. In EMNLP, pages 17716-17736, 2024. 2, 5 +[63] Fangyun Wei and Yutong Chen. Improving continuous sign language recognition with cross-lingual signs. In ICCV, pages 23612-23621, 2023. 5, 8 +[64] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In ICLR, 2024. 1 +[65] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI, pages 6234-6242, 2024. 2, 3 +[66] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In CVPR, pages 13525-13536, 2024. 2 +[67] Huijie Yao, Wengang Zhou, Hao Feng, Hezhen Hu, Hao Zhou, and Houqiang Li. Sign language translation with iterative prototype. In ICCV, pages 15592-15601, 2023. 1 +[68] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2, 3, 4, 7 +[69] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. MLSLT: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1 +[70] Aoxiong Yin, Haoyuan Li, Kai Shen, Siliang Tang, and Yueting Zhuang. T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text. In ACL, 2024. 1, 2, 7 +[71] Kayo Yin, Amit Moryossef, Julie Hochgesang, Yoav Goldberg, and Malihe Alikhani. Including signed languages in natural language processing. In ACL, pages 7347-7360, 2021. 1 +[72] Zhengdi Yu, Shaoli Huang, Yongkang Cheng, and Tolga Birdal. Signavatars: A large-scale 3d sign language holistic motion dataset and benchmark. In ECCV, pages 1-19, 2024. 1, 2, 3, 8 +[73] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In WACV, pages 3395-3403, 2020. 2 +[74] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2 +[75] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffeuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2 + +[76] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, pages 7368-7376, 2024. 2 +[77] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. BEST: BERT pre-training for sign language recognition with coupling tokenization. In AAAI, 2023. 3 +[78] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houqiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 5 +[79] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. IEEE TIP, 2024. 3 +[80] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In ICCV, pages 20871-20881, 2023. 1 +[81] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, pages 13009-13016, 2020. 2, 3 +[82] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 2, 3, 5 +[83] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, pages 1-1, 2021. 1 +[84] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In CVPR, pages 1357-1366, 2024. 2 +[85] Ronglai Zuo and Brian Mak. C2SLR: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2 +[86] Ronglai Zuo and Brian Mak. Improving continuous sign language recognition with consistency constraints and signer removal. ACM TOMM, 2024. 3 +[87] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 5 +[88] Ronglai Zuo, Fangyun Wei, Zenggui Chen, Brian Mak, Jiaolong Yang, and Xin Tong. A simple baseline for spoken language to sign language translation with 3d avatars. In ECCV, 2024. 1, 2, 3 +[89] Ronglai Zuo, Fangyun Wei, and Brian Mak. Towards online continuous sign language recognition and translation. In EMNLP, 2024. 5 \ No newline at end of file diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/images.zip b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e274d82f1e338f926ed1ebd3099ed02322278416 --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d2f23a545b6c5715de4d1fe3ec426fc752dd36dc8f787d0082c163fc28ba419 +size 830328 diff --git a/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/layout.json b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8910cb256c70e9cf633e2ba34ce2da998d7aac59 --- /dev/null +++ b/2025/Signs as Tokens_ A Retrieval-Enhanced Multilingual Sign Language Generator/layout.json @@ -0,0 +1,8454 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "text", + "content": "Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 143, + 553, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 143, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 143, + 553, + 171 + ], + "type": "text", + "content": "Ronglai Zuo, Rolandos Alexandros Potamias, Evangelos Ververas, Jiankang Deng, Stefanos Zafeiriou \nImperial College London" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 174, + 390, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 174, + 390, + 185 + ], + "spans": [ + { + "bbox": [ + 221, + 174, + 390, + 185 + ], + "type": "text", + "content": "https://2000zrl.github.io/soke/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 202, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 69, + 202, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 69, + 202, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 69, + 202, + 541, + 352 + ], + "type": "image", + "image_path": "d814b5c7743d60a34a033c47a1f9811edfb6a98285646d2e4d7a0bccb8355abd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 355, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 555, + 380 + ], + "type": "text", + "content": "Figure 1. We propose Signs as Tokens (SOKE), a unified sign language generator that can produce multilingual signs from text inputs. (Left: American sign language; Middle: Chinese sign language; Right: German sign language.)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 388, + 200, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 388, + 200, + 400 + ], + "spans": [ + { + "bbox": [ + 152, + 388, + 200, + 400 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 297, + 714 + ], + "type": "text", + "content": "Sign language is a visual language that encompasses all linguistic features of natural languages and serves as the primary communication method for the deaf and hard-of-hearing communities. Although many studies have successfully adapted pretrained language models (LMs) for sign language translation (sign-to-text), the reverse task—sign language generation (text-to-sign)—remains largely unexplored. In this work, we introduce a multilingual sign language model, Signs as Tokens (SOKE), which can generate 3D sign avatars autoregressively from text inputs using a pretrained LM. To align sign language with the LM, we leverage a decoupled tokenizer that discretizes continuous signs into token sequences representing various body parts. During decoding, unlike existing approaches that flatten all part-wise tokens into a single sequence and predict one token at a time, we propose a multi-head decoding method capable of predicting multiple tokens simultaneously. This approach improves inference efficiency while maintaining effective information fusion across different body parts. To further ease the generation process, we propose a retrieval-enhanced SLG approach, which incorporates external sign dictionaries to provide accurate word-level signs as auxiliary conditions, significantly improving the precision of generated signs. Extensive qualitative and quantitative evaluations demonstrate the effectiveness of SOKE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 394, + 399 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 530 + ], + "type": "text", + "content": "Sign language is the primary communication method utilized by the deaf and hard-of-hearing communities, embodying all linguistic properties found in spoken languages, such as discrete semantic units and grammatical structure [16, 71]. As an attempt to mitigate the barrier between sign and spoken languages, several methods have been developed to tackle sign language processing by unfolding the task into two research problems: sign language translation (SLT, sign-to-text) [20, 33, 56, 64, 67, 69, 80, 83] and sign language generation (SLG, text-to-sign) [2, 47, 53, 72, 88]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 532, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 675 + ], + "type": "text", + "content": "In contrast with SLT, the usage of language models (LMs) in SLG remains relatively unexplored. Most existing methods approach SLG as a visual content generation task (e.g., video, keypoint, or motion), using GANs [50, 53] or diffusion models [2, 12, 44, 58, 59]. However, these works neglect the linguistic nature of sign languages, leading to suboptimal performance and missing benefits of pretrained LMs such as generalizability and scalability [18, 29, 60]. In particular, sign languages share the fundamental linguistic properties and the discrete structure of spoken languages [16, 71], which necessitates processing using autoregressive language modeling techniques [70]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "To formulate SLG in the framework of language modeling, a crucial step is mapping signs into discrete tokens. A common approach to discretizing sign languages involves" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23806" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 371 + ], + "type": "text", + "content": "using glosses—the written form of signs—as an intermediate representation [13, 22, 50, 88]. However, glosses require extensive annotation efforts and impose a predefined information bottleneck, which fail to fully capture the rich semantics of sign languages [39]. To address these limitations and better capture the multi-cue property of sign languages [26, 81, 85], we leverage a vector quantized-variational auto-encoder (VQ-VAE)-based tokenizer [37, 65, 68] that learns mappings between continuous sign motions and discrete tokens for each body part (upper body, left hand, and right hand). During decoding, existing autoregressive motion generation methods [37, 68] typically flatten tokens from different body parts into a single sequence, resulting in inefficient inference due to tripled decoding steps. To overcome this challenge, we propose a novel multi-head decoding method capable of predicting multiple tokens at a time while effectively fusing information from different body parts in real time. Furthermore, given the limited vocabulary size of continuous sign language datasets, we enhance the model's generalization by instantiating the LM as a multilingual LM [35], enabling it to handle a broader and more diverse domain of discourse. To support multilingual SLG in a unified framework, we train the proposed model on a curated multilingual sign language dataset encompassing American, Chinese, and German sign languages [3, 10, 82]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 374, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 601 + ], + "type": "text", + "content": "Recently, retrieval-augmented generation (RAG) [4, 11, 62] has proven effective in enhancing the interpretability and accuracy of content generated by large language models. Similarly, several SLG methods [50, 88] leverage glosses to retrieve sign dictionaries and directly incorporate retrieved signs into the outputs, improving the interpretability of generated signs. However, beyond the inherent limitations of glosses, these methods often face challenges in handling co-articulations between adjacent signs. This is because dictionary signs are typically recorded at the word level, whereas SLG models operate at the sentence level, leading to potential unnaturalness in the generated sequences. To address these issues, we propose a retrieval-enhanced SLG approach that utilizes motion tokens from retrieved dictionary signs as additional conditions for language modeling. This approach avoids the unnaturalness of directly copying signs into model outputs while still benefiting from the precision of dictionary signs. To sum up, the contributions of this paper can be summarized as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": "- We propose a novel SLG approach, Signs as Tokens (SOKE). With the aid of a pretrained LM, SOKE can handle multiple sign languages in a unified model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": "- We propose a multi-head decoding method that reduces inference latency and efficiently fuses information from different body parts, enabling seamless integration of a decoupled tokenizer with the LM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- By leveraging external sign dictionaries, we introduce a retrieval-enhanced SLG approach to improve the preci" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 73, + 419, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 419, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 419, + 84 + ], + "type": "text", + "content": "sion of generated signs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 84, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 156 + ], + "type": "text", + "content": "- We curate two sign language datasets [3, 82] with accurate SMPL-X [42] poses to facilitate multilingual SLG research. Experiments demonstrate that our SOKE achieves state-of-the-art performance on three challenging benchmarks, How2Sign [10], CSL-Daily [82], and Phoenix2014T [3], using a single unified model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 400, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 187, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 617 + ], + "type": "text", + "content": "Sign Language Generation. Early gloss-based methods [21, 22, 50, 57, 65, 73, 88] approach SLG under a text-to-gloss-to-sign framework. Several works [1, 47, 49] propose to directly regress 2D joints to represent sign poses, simplifying the modeling process, however, they suffer from mode collapse due to the limited scale of data. Recent state-of-the-art SLG works can be categorized into two classes: the first group of methods [2, 12] employs diffusion models to generate sign motions conditioned on text inputs; the second group of methods [9, 70, 72] adopts a tokenizer-LM two-stage autoregressive generation approach. Among these, [72] incorporates costly gloss and HamNoSys notations [1] to train a semantic adapter, while [9] focuses on generating signs from a single word, limiting their applicability. Moreover, a notable gap exists in the literature regarding unified models for multilingual SLG. In this work, we address these gaps by utilizing a decoupled tokenizer and a pretrained multilingual LM, integrated with our customized multi-head decoding method and retrieval-enhanced approach, to enable text-driven multilingual SLG. Text-to-Motion Generation. Text-to-motion generation (T2MG) is an important research area with potential applications in human-computer interaction and robotics. Recent T2MG methods can be categorized into two main types: 1) Diffusion-based approaches [5, 51, 75], which utilize latent diffusion [46] to generate motions conditioned on text inputs. 2) Autoregressive approaches [17, 24, 37, 74, 76, 84], which draw inspiration from text generation methods by first tokenizing motion sequences and then using LMs to perform autoregressive generation. When decoding with a decoupled tokenizer [37, 68], these autoregressive approaches simply flatten tokens from various body parts into one sequence, causing additional decoding steps. In this work, we propose a novel multi-head decoding method that can efficiently predict motion tokens for all body parts at one single step, significantly reducing inference latency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Retrieval-Enhanced Generation. Using the concept of retrieval to enhance generation models has proven effective across various domains [4, 8, 11, 15, 62, 66]. As a pioneering effort, Piktus et al. [31] propose to retrieve relevant documents before generating answers for knowledge-intensive NLP tasks. FLARE [25] introduces selective retrieval, activating it only when low-confidence tokens appear in the generation process. ToG [55] leverages retrieved" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23807" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 199 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 199 + ], + "type": "image", + "image_path": "a38ba151181b6e68a215bb785e08ce9f460b432080e8340715bac1ed9c7182d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. An overview of our proposed method, Signs as Tokens (SOKE). We begin by training a VQ-VAE-based decoupled tokenizer to map continuous sign motions into discrete tokens for various body parts (upper body, left hand, and right hand). These sign motion tokens are then integrated into the text vocabulary of a pretrained language model, which serves as the backbone of our autoregressive multilingual generator (AMG). Given a text input, the AMG first retrieves word-level signs from external dictionaries, appends their motion tokens to the text tokens, and feeds them into the language model encoder. During decoding, our novel multi-head decoding strategy generates motion tokens for all body parts simultaneously at each time step. Finally, the derived motion tokens are used to reconstruct sign avatars." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 284, + 297, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 429 + ], + "type": "text", + "content": "knowledge to enhance the reasoning capabilities of large language models. In the field of image generation, retrieval has also been utilized to enable diffusion models to generate out-of-distribution images [52]. In the context of SLG, several methods [50, 88] adopt gloss-based retrieval to obtain dictionary signs and approach SLG in a retrieve-and-concatenate manner. However, glosses have inherent limitations, and such methods often suffer from unrealistic transitions between signs. In this paper, we propose a simple yet effective retrieval-enhanced SLG approach that uses word-level signs as auxiliary conditions, addressing these limitations and improving the overall quality of generations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 138, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 138, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 138, + 453 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 460, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 296, + 568 + ], + "type": "text", + "content": "An overview of our method is shown in Figure 2. Our approach mainly consists of two parts: a decoupled tokenizer (DETO) and an autoregressive multilingual generator (AMG). In the first stage, DETO learns mappings between continuous sign motions and discrete tokens for each body part (Section 3.2). In the second stage, AMG is trained to autoregressively generate motion tokens with the proposed multi-head decoding strategy (Section 3.3) conditioned on text inputs and retrieved dictionary signs (Section 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 160, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 160, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 160, + 588 + ], + "type": "text", + "content": "3.1. Data Preparation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 296, + 715 + ], + "type": "text", + "content": "Although existing SLG works [2, 72] have made efforts in curating high-quality SMPL-X poses for an American sign language (ASL) dataset, How2Sign [10], there is a notable lack of 3D annotations for other continuous sign language datasets, impeding multilingual SLG research. In this work, we rely on a two-stage approach, leveraging state-of-the-art 3D pose estimation methods [34, 43], to extract high-fidelity SMPL-X poses for CSL-Daily [82] and Phoenix-2014T [3], two widely-adopted datasets of Chinese and German sign languages (CSL and DGS), respec" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "text", + "content": "tively. Specifically, we initially estimate a set of rough 3D body poses for each signer using OSX [34] along with accurate 3D hand poses from WiLoR [43]. Given that OSX often falls short in accurately capturing the arm poses, we follow [2] and refine the upper body joint rotations by minimizing a re-projection loss between the estimated joints and the detected 2D joint keypoints from Mediapipe [38]. Using this process we can ensure accurate whole-body annotations. We represent each sign motion sequence as " + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "text", + "content": " is the sequence length and " + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "inline_equation", + "content": "d = 133" + }, + { + "bbox": [ + 313, + 284, + 555, + 453 + ], + "type": "text", + "content": " denotes the number of SMPL-X parameters, including 11 upper body joints, 30 hand joints, and 10 expression parameters. Please check the supplementary materials for more details and quantitative evaluations of our pose fitting pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 436, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 436, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 436, + 475 + ], + "type": "text", + "content": "3.2. Decoupled Tokenizer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 564 + ], + "type": "text", + "content": "Sign languages exhibit a multi-cue property [19, 26, 77, 79, 81, 86], where the semantics of a sign are conveyed simultaneously through body and hand movements. Inspired by this, we leverage a decoupled tokenizer [37, 65, 68], comprising three VQ-VAEs [61], to independently model key regions: the upper body and both hands. An illustration is available in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": "Given a " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": "-frame sign motion input " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ", we first decompose it into three part-wise motion sequences based on the kinematic tree of SMPL-X: " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^p \\in \\mathbb{R}^{T \\times d_p}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "p \\in \\{B, LH, RH\\}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ". Moreover, we build three distinct VQ-VAEs, where each of them consists of an encoder " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_p(\\cdot)" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ", a decoder " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p(\\cdot)" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ", and a learnable codebook " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^p \\in \\mathbb{R}^{N_Z^p \\times C}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "N_Z^p" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": " represents the number of codes and " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": " denotes the code dimension. For each motion sequence, the corresponding encoder first projects the sequence into a latent space: " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_f^p = \\mathcal{E}_p(\\mathbf{S}^p) = \\{s_{f,i}^p\\}_{i=1}^{T_f} \\in \\mathbb{R}^{T_f \\times C}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": ", using a stack of 1D-CNN layers. Then, for each pose we can derive a set of discrete tokens " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^p = \\{\\hat{z}_i^p\\}_{i=1}^{T_f}" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": " using a quantizer " + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 313, + 565, + 555, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23808" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 220, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 220, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 220, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 220, + 175 + ], + "type": "image", + "image_path": "2c523adcc341bf8773ce91323acf95e297bb6a697d2e6b77176915f82d8d6b61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 178, + 189, + 188 + ], + "lines": [ + { + "bbox": [ + 91, + 178, + 189, + 188 + ], + "spans": [ + { + "bbox": [ + 91, + 178, + 189, + 188 + ], + "type": "text", + "content": "(a) Sequential Decoding " + }, + { + "bbox": [ + 91, + 178, + 189, + 188 + ], + "type": "inline_equation", + "content": "(O(3K))" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 69, + 359, + 179 + ], + "blocks": [ + { + "bbox": [ + 223, + 69, + 359, + 179 + ], + "lines": [ + { + "bbox": [ + 223, + 69, + 359, + 179 + ], + "spans": [ + { + "bbox": [ + 223, + 69, + 359, + 179 + ], + "type": "image", + "image_path": "0e7520f580bbb11ae4a354f4c4508f563b4f5f500a7bf4858ceeb6c221e29224.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 179, + 332, + 188 + ], + "lines": [ + { + "bbox": [ + 248, + 179, + 332, + 188 + ], + "spans": [ + { + "bbox": [ + 248, + 179, + 332, + 188 + ], + "type": "text", + "content": "(b) Parallel Decoding " + }, + { + "bbox": [ + 248, + 179, + 332, + 188 + ], + "type": "inline_equation", + "content": "(O(K))" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. Comparison of various decoding methods. (a) Sequential decoding flattens all motion tokens into a single sequence, requiring an inefficient " + }, + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "type": "inline_equation", + "content": "3K" + }, + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "type": "text", + "content": " decoding steps, where " + }, + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 195, + 555, + 251 + ], + "type": "text", + "content": " is the target motion sequence length. (b) Parallel decoding improves efficiency by decoding body parts in parallel but lacks information fusion, which may lead to suboptimal performance. (c) Our proposed multi-head decoding balances efficiency and fusion by predicting three tokens simultaneously via different heads, with input embeddings as a weighted average of part-wise token embeddings." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 359, + 71, + 554, + 178 + ], + "blocks": [ + { + "bbox": [ + 359, + 71, + 554, + 178 + ], + "lines": [ + { + "bbox": [ + 359, + 71, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 359, + 71, + 554, + 178 + ], + "type": "image", + "image_path": "51b151d1ffd557f7f08d9bbe831e08c7caaf1bb81f20b19c98f4696767510ba7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 178, + 523, + 188 + ], + "lines": [ + { + "bbox": [ + 412, + 178, + 523, + 188 + ], + "spans": [ + { + "bbox": [ + 412, + 178, + 523, + 188 + ], + "type": "text", + "content": "(c) Multi-head Decoding (ours, " + }, + { + "bbox": [ + 412, + 178, + 523, + 188 + ], + "type": "inline_equation", + "content": "O(K)" + }, + { + "bbox": [ + 412, + 178, + 523, + 188 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 263, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 276 + ], + "type": "text", + "content": "which searches the nearest neighbor from the codebook " + }, + { + "bbox": [ + 55, + 263, + 295, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^p" + }, + { + "bbox": [ + 55, + 263, + 295, + 276 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 288, + 295, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 308 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {i} ^ {p} = Q \\left(s _ {f, i} ^ {p}\\right) = \\arg \\min _ {z _ {j} \\in \\mathbf {Z} ^ {p}} \\| s _ {f, i} ^ {p} - z _ {j} \\| _ {2}, j \\in [ 1, N _ {Z} ^ {p} ]. \\tag {1}", + "image_path": "155a1cd9f76fcd66a9b61acbc67e3360caed9eaf0434756d50ce8f3e52c92f30.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": "We then feed the obtained token sequence to the corresponding part-decoder " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": " to reconstruct the input motions: " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}^p = \\mathcal{D}_p(\\mathbf{Z}^p)" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": ". Following [24], the objective function of each VQ-VAE is defined as: " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vq}^p = \\mathcal{L}_{rec}^p +\\mathcal{L}_{emb}^p +\\mathcal{L}_{com}^p" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}^p = \\| \\hat{\\mathbf{S}}^p -\\mathbf{S}^p\\| _2^2" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": " denotes the reconstruction loss, " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{emb}^p" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": " is the embedding loss, and " + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{com}^p" + }, + { + "bbox": [ + 55, + 323, + 296, + 410 + ], + "type": "text", + "content": " is the commitment loss, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 424, + 260, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 260, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 260, + 437 + ], + "type": "text", + "content": "3.3. Autoregressive Multilingual Generator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 479 + ], + "type": "text", + "content": "Using obtained discrete sign representations, we can respect the discrete properties of sign languages and approach SLG similarly to text generation with a pretrained LM [35]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": "Integrated Vocabulary. We first gather the discrete tokens for different body parts to construct a part-wise motion vocabulary: " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_m^p = \\{\\langle p\\_ i\\rangle \\}_{i = 1}^{N_Z^p}" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\langle B\\_ 1\\rangle" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{RH\\_2}\\rangle" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": ". The overall motion vocabulary is a combination of three part-wise vocabularies: " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_m = \\{\\mathcal{V}_m^B,\\mathcal{V}_m^{LH},\\mathcal{V}_m^{RH}\\}" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": ". Following [35], we further define a series of language identifier tokens: " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_l = \\{\\langle \\mathrm{ASL}\\rangle ,\\langle \\mathrm{CSL}\\rangle ,\\langle \\mathrm{DGS}\\rangle \\}" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": ", to prompt the LM with the information of target sign language. The integrated vocabulary for the generator is defined as: " + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{\\mathcal{V}_t,\\mathcal{V}_m,\\mathcal{V}_l\\}" + }, + { + "bbox": [ + 55, + 481, + 295, + 616 + ], + "type": "text", + "content": ", which lays a foundation for the subsequent LM fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Sequential Decoding. Following previous motion generation works [37, 68], a naive decoding method is simply flattening tokens from each body part into a single sequence with tripled length: " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} = \\{y_1^B, y_1^{LH}, y_1^{RH}, \\ldots, y_K^B, y_K^{LH}, y_K^{RH}\\}" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "y_k^p \\in \\mathcal{V}_m^p" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " denotes the part-wise motion token, and " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " denotes the length of the target sign motion sequence. The decoding process can then be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 261, + 555, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 261, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 314, + 261, + 555, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathbf {Y} \\mid \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {2} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {k} ^ {B}, y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {k} ^ {B, L H}, y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}", + "image_path": "39d0c86e6718f9c56f9d92bf1c9113709dfc1da74a2778dd7586964e9d76b75e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{en}" + }, + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "type": "text", + "content": " denotes the last encoder hidden states. For simplicity, we write " + }, + { + "bbox": [ + 313, + 333, + 554, + 392 + ], + "type": "inline_equation", + "content": "y_{" + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": " when the model is required to generate upper-body motions for ASL. The decoding process can then be formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 587, + 555, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 587, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 319, + 587, + 555, + 638 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = P (\\mathbf {Y} ^ {B}) P (\\mathbf {Y} ^ {L H}) P (\\mathbf {Y} ^ {R H}) \\tag {3} \\\\ = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B} \\mid y _ {< k} ^ {B}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {L H} \\mid y _ {< k} ^ {L H}\\right) \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {R H} \\mid y _ {< k} ^ {R H}\\right). \\\\ \\end{array}", + "image_path": "d36014e823d0c7feda5c1a611960c1b967a85db04d06eae81fc8c79137185485.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Comparing Eq. 3 with Eq. 2, the parallel decoding strategy significantly reduces the dependence on previous tokens, as the decoding processes for each body part are completely separated. While this formulation enables efficient parallel execution, the overly independent assumption may result in suboptimal performance." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23809" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 177 + ], + "type": "image", + "image_path": "1f2990349f912a1833ba2b86aa4ec4f89581f242e528731e8ba4dc3f38480cb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 214 + ], + "type": "text", + "content": "Figure 4. Illustration of retrieval-enhanced SLG. We utilize external sign dictionaries to retrieve accurate word-level signs, which serve as additional conditions to enhance the generation quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": "Multi-Head Decoding. To balance efficiency and information fusion across body parts, we propose a multi-head decoding strategy. As shown in Figure 3(c), we design three language modeling heads, implemented as fully connected layers, to predict motion tokens for each body part simultaneously at each step. Moreover, the decoder inputs at each step are modified to a weighted average of the token embeddings from different body parts. Specifically, denoting the token embeddings from upper body, left hand, and right hand as " + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^B" + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{LH}" + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{RH}" + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": ", respectively, the input embeddings are defined as: " + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{E} = (1 - 2\\lambda)\\mathbf{E}^B + \\lambda \\mathbf{E}^{LH} + \\lambda \\mathbf{E}^{RH}" + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\lambda \\in (0,0.5)" + }, + { + "bbox": [ + 55, + 230, + 296, + 396 + ], + "type": "text", + "content": " is a hyper-parameter to control the weight of hand embeddings. The decoding process can then be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 406, + 295, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 406, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 80, + 406, + 295, + 474 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P (\\mathbf {Y} | \\mathbf {h} _ {e n}) = \\prod_ {k = 1} ^ {K} P \\left(y _ {k} ^ {B, L H, R H} \\mid y _ {< k} ^ {B, L H, R H}\\right) \\tag {4} \\\\ = \\prod_ {k = 1} ^ {K} P (y _ {k} ^ {B} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {L H} | y _ {< k} ^ {\\{p \\}}) P (y _ {k} ^ {R H} | y _ {< k} ^ {\\{p \\}}), \\\\ \\end{array}", + "image_path": "15a7fc2b199cd71bb2b364bed280ad0d5e1a7c48c04068b2c91c7201f12fa3dc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "type": "text", + "content": "which can be viewed as relaxing Eq. 2 under a conditional independence assumption." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "text", + "content": "Training and Inference. The LM is trained with a standard cross-entropy loss: " + }, + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LM} = -\\log P(\\mathbf{Y}|\\mathbf{h_{en}})" + }, + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "text", + "content": ". During inference, we adopt a simple greedy decoding algorithm that the LM always outputs the token with the highest probability at each step. For parallel and multi-head decoding, the inference process terminates as soon as any head or decoder predicts an end-of-sentence token. The obtained token sequences will be fed into the corresponding part-decoder, " + }, + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_p(\\cdot)" + }, + { + "bbox": [ + 55, + 507, + 296, + 616 + ], + "type": "text", + "content": ", to reconstruct sign motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 195, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 195, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 195, + 635 + ], + "type": "text", + "content": "3.4. Retrieval-Enhanced SLG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": "Inspired by the effectiveness of retrieval-augmented generation in NLP [4, 11, 62], we propose to leverage external sign dictionaries to improve the fidelity of generated signs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As shown in Figure 4, we first build a word-level sign dictionary for each target sign language by utilizing resources from isolated sign language recognition datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "[28, 30, 32, 78, 87] and online sources. Through our pose fitting pipeline, each RGB video in these dictionaries is converted into SMPL-X poses. Next, we employ a well-trained decoupled tokenizer to map the poses into discrete tokens. The resulting dictionary is represented as a set of quadruples " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "\\{(w, m^B, m^{LH}, m^{RH})\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " denotes lemmatized words in the written language, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "inline_equation", + "content": "m^p" + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": " denotes motion tokens. If a word has multiple sign instances in the dictionary, we retain only the instance with the lowest reconstruction error after passing through the tokenizer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": "Given a text input " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{x_{i}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " words" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": ", we collect all words and corresponding tokens found in the dictionary. The final prompt is created by concatenating raw texts " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": " with the retrieved motion tokens " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\{m_j^p\\}" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": ", which is then fed into the LM encoder to generate hidden states " + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{en}" + }, + { + "bbox": [ + 313, + 192, + 553, + 264 + ], + "type": "text", + "content": ". Several examples are shown in Figure 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 395, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 395, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 395, + 290 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 405 + ], + "type": "text", + "content": "Datasets. We construct a multilingual sign language dataset by merging three widely-adopted datasets: How2Sign [10], CSL-Daily [82], and Phoenix-2014T [3]. How2Sign is a large-scale ASL dataset, consisting of 35K video-text pairs. CSL-Daily is the largest existing continuous CSL dataset, containing 20K samples. Phoenix-2014T is a DGS dataset with 8K samples. To train SOKE, we use the SMPL-X poses of How2Sign dataset as provided in [2] along with the curated poses of CSL-Daily and Phoenix-2014T." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 405, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 502 + ], + "type": "text", + "content": "Evaluation Metrics. Since the length of generated signs may differ from the ground truth, we employ the well-established dynamic time warping (DTW) [1, 2] on both procrustes-aligned and original joint position errors (JPE) to measure sequence-level distances between the generated signs and ground truth. Following [2], we also report backtranslation (B-T) scores in BLEU-4 [41] to assess the interpretability of the generated signs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "content": "Implementation Details. For DETO, we empirically set the code numbers to " + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "inline_equation", + "content": "N_Z^B = 96" + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "content": " for the body and " + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "inline_equation", + "content": "N_Z^{LH} = N_Z^{RH} = 192" + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "content": " for the left and right hands, respectively, with a code dimension of " + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "inline_equation", + "content": "C = 512" + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "content": ". We train DETO on the combined multilingual sign language dataset and sign dictionaries with a batch size of 320 per GPU for 500 epochs, using the AdamW optimizer [36] and a cosine learning rate scheduler starting at 2e-4. For the LM, we use mBART-large-cc25 [35], which features 12 encoder-decoder layers with a model dimension of 1024 and is widely used in various sign language translation methods [6, 7, 27, 63, 89]. We empirically set " + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\lambda = 1/3" + }, + { + "bbox": [ + 313, + 502, + 554, + 693 + ], + "type": "text", + "content": " to balance part-wise token embeddings. The LM is fine-tuned with a batch size of 32 per GPU on the multilingual sign language dataset for 150 epochs, employing the same optimizer settings as DETO. All models are trained using 6 RTX 3090 GPUs." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "type": "text", + "content": "1 Since Chinese sentences lack spaces between words, we use Jieba [14," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23810" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 564 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 564 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 564 + ], + "type": "image", + "image_path": "82e67d16ec45b1a5e39c12388e13c2b566ac884f601d1fe8ab4dee46247fdcb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 566, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 55, + 566, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 555, + 589 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of generated signs between our proposed method, SOKE, with the SOTA method, S-MotionGPT [24], on the test sets of How2Sign (left), CSL-Daily (middle), and Phoenix-2014T (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 597, + 282, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 282, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 282, + 611 + ], + "type": "text", + "content": "4.1. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 616, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 689 + ], + "type": "text", + "content": "Qualitative Comparison. We first reimplement a state-of-the-art (SOTA) open-sourced motion generation approach, MotionGPT [24], which utilizes a VQ-VAE to discretize whole-body motions and employs a monolingual language model [45], on sign language datasets. The resulting model, S-MotionGPT, achieves competitive results on sign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 598, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 707 + ], + "type": "text", + "content": "language benchmarks (Table 1). In Figure 5, we perform a qualitative comparison between our SOKE and S-MotionGPT. As can be easily seen, our proposed method produces significantly better results, with visual quality comparable to the ground truth sign videos across the ASL, CSL, and DGS datasets. Notably, our method can generate precise and informative hand movements, which play a key role in conveying semantics in sign languages. Additionally, the proposed method supports three sign languages" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 145, + 712 + ], + "type": "text", + "content": "40] for word segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23811" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 196 + ], + "type": "table", + "html": "
MethodMulti-lingualHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑DTW-PA-JPE↓DTW-JPE↓B-T↑
BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4BodyHandBodyHandBLEU-4
NAR [23]×13.9411.80--5.75----------
Prog. Trans.* [47]×14.1511.5714.7430.172.7515.9812.9116.3032.633.0713.6711.9515.0131.774.94
Text2Mesh* [54]×13.9913.4715.5032.977.5113.4712.1013.7630.375.1113.4812.0614.0431.645.81
Adv. Train. [48]×13.7811.17--6.21----------
T2S-GPT* [70]×11.486.3912.6518.4411.2011.945.9312.3215.438.9410.386.4711.6519.099.06
NSA [2]×7.837.33--13.12----------
S-MotionGPT* [24]×11.234.3912.4113.7411.4510.813.7811.5811.318.829.453.4110.429.089.68
SOKE (ours)✓6.822.357.7510.0814.486.241.717.389.6811.304.771.386.047.7211.87
", + "image_path": "2583ad48cfe47d4a6ef5da83c0ce5bea51938e7662706a4a6612539ba9d47381.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 217, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 76, + 200, + 533, + 212 + ], + "lines": [ + { + "bbox": [ + 76, + 200, + 533, + 212 + ], + "spans": [ + { + "bbox": [ + 76, + 200, + 533, + 212 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art sign language generation (text-to-sign) methods. * denotes reimplementation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 217, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 65, + 217, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 65, + 217, + 544, + 312 + ], + "type": "table", + "html": "
Decoding MethodSign RetrievalHow2SignCSL-DailyPhoenix-2014T
DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓DTW-PA-JPE↓Latency↓
AvgBodyHands/videoAvgBodyHands/videoAvgBodyHands/video
Sequential×4.648.553.533.263.748.142.493.153.266.972.203.23
Parallel×5.069.273.861.394.198.612.931.413.547.022.551.35
Multi-head×4.177.913.101.463.377.582.171.452.806.161.851.44
Multi-head✓3.346.822.351.552.726.241.711.522.134.771.381.51
", + "image_path": "40f831b0f02126de3537f8463de29bc00403caebd40be9b0fc744594400ba123.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 314, + 526, + 326 + ], + "lines": [ + { + "bbox": [ + 82, + 314, + 526, + 326 + ], + "spans": [ + { + "bbox": [ + 82, + 314, + 526, + 326 + ], + "type": "text", + "content": "Table 2. Ablation study for the decoding method and sign retrieval. Latencies are measured using a single RTX 3090 GPU." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 335, + 295, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 359 + ], + "type": "text", + "content": "simultaneously, whereas S-MotionGPT is monolingual and requires separate models for each sign language dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 360, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 295, + 552 + ], + "type": "text", + "content": "Quantitative Comparison. In addition to the above qualitative comparison, we also perform a quantitative evaluation against SOTA SLG methods, as detailed in Table 1. Our reimplemented S-MotionGPT achieves competitive results, with low DTW errors of 4.39 and 3.78 on How2Sign and CSL-Daily, respectively, demonstrating the effectiveness of formulating SLG as a language modeling task. The previous best SLG approach, Neural Sign Actors (NSA) [2], formulates SLG as a motion generation task and employs a diffusion model as its backbone. However, this approach overlooks the linguistic properties of sign languages, resulting in higher average DTW errors compared to S-MotionGPT. In contrast, our method sets a new state-of-the-art across all three benchmarks, achieving significantly lower DTW errors of 2.35, 1.71, and 1.38 on How2Sign, CSL-Daily, and Phoenix-2014T, respectively, using a single unified model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Decoding Method. We compare the performance of different decoding methods in Table 2. The traditional sequential decoding method is widely used in existing motion generation approaches [37, 68]. However, since tokens from different body parts are flattened into a single sequence, the number of decoding steps is tripled, leading to high inference latency (3.26s per video on How2Sign). By using three separate LM decoders, parallel decoding predicts token sequences for each body part simultaneously. This formulation significantly reduces latency (1.39s per video on How2Sign) but results in higher DTW errors due to the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 333, + 555, + 578 + ], + "blocks": [ + { + "bbox": [ + 315, + 333, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 315, + 333, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 315, + 333, + 555, + 578 + ], + "type": "image", + "image_path": "685c6bd093b9d02bfc96fdac36184f478096bf7a4b1b7e62eacdb2628a96a28c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 578, + 553, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 578, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 553, + 601 + ], + "type": "text", + "content": "Figure 6. Qualitative ablation study for retrieval-enhanced SLG. (Left: How2Sign; Middle: CSL-Daily; Right: Phoenix-2014T.)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "lack of information fusion across part-wise motion tokens. Notably, our multi-head decoding method achieves a better trade-off between efficiency and information integration, maintaining comparable latency to parallel decoding while delivering even superior generation quality than sequential decoding. We attribute this to the weighting mechanism (soft gating): since the decoder inputs are a weighted average of all part-wise tokens, errors in a single token may" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23812" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 539, + 152 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 539, + 152 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 539, + 152 + ], + "type": "table", + "html": "
H2SCSLPh-TGenerationReconstruction
H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)H2S (JPE↓)CSL (JPE↓)Ph-T (JPE↓)
BodyHandBodyHandBodyHandBodyHandBodyHandBodyHand
✓7.923.07----21.527.74----
✓✓7.112.636.792.14--19.886.9724.295.83--
✓✓✓6.822.356.241.714.771.3819.376.6523.525.1325.796.78
", + "image_path": "c0f3bb79bd908c9067cdba403d79362218c4cf81da94dce2f0529a5a80010b18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 180, + 294, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 153, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 555, + 177 + ], + "type": "text", + "content": "Table 3. Study on the scalability of our approach. We use procrustes-aligned mean per joint position error (PA-MPJPE) [34, 72] to assess reconstruction performance of the decoupled tokenizer. (Ph-T: Phoenix-2014T.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 180, + 294, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 180, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 294, + 242 + ], + "type": "table", + "html": "
VAE-1VAE-2VAE-3H2S (DTW↓)CSL (DTW↓)Ph-T (DTW↓)
BodyHandBodyHandBodyHand
Whole-body××9.493.569.223.358.033.23
BodyHands×7.993.136.932.856.242.18
BodyL-handR-hand6.822.356.241.714.771.38
", + "image_path": "f769d1917485475ad48a5efbbe2d92ee0b0dc0042439f61ec0e37f3dca5e6c72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 242, + 269, + 253 + ], + "lines": [ + { + "bbox": [ + 81, + 242, + 269, + 253 + ], + "spans": [ + { + "bbox": [ + 81, + 242, + 269, + 253 + ], + "type": "text", + "content": "Table 4. Ablation study on the decoupled tokenizer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 267, + 260, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 260, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 260, + 278 + ], + "type": "text", + "content": "have a reduced impact on the overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 279, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 279, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 295, + 434 + ], + "type": "text", + "content": "Sign Retrieval. By leveraging external sign dictionaries, we develop a retrieval-enhanced SLG approach that incorporates retrieved word-level signs as additional generation conditions. As shown in Table 2, this retrieval mechanism significantly enhances the precision of generated signs, reducing the average DTW errors by " + }, + { + "bbox": [ + 54, + 279, + 295, + 434 + ], + "type": "inline_equation", + "content": "19.9\\% / 19.6\\% / 23.9\\%" + }, + { + "bbox": [ + 54, + 279, + 295, + 434 + ], + "type": "text", + "content": " on How2Sign, CSL-Daily, and Phoenix-2014T, respectively. Furthermore, we perform a qualitative comparison in Figure 6. The figure demonstrates that the retrieved signs effectively guide the model's generation: similar pose patterns from the retrieved signs, particularly in hand shapes, appear in the generated outputs. This guidance substantially improves the precision of the generated signs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 295, + 567 + ], + "type": "text", + "content": "Scalability with Multilingual Data. Results in Table 3 show that incorporating multilingual data significantly improves generation performance, demonstrating the scalability of our approach. This improvement stems from two key factors: 1) the tokenizer benefits from more pose training data (right half of Table 3), which subsequently enhances the generator's performance; 2) the presence of cross-lingual signs, as revealed in [63]. For example, the sign for \"cold\" in CSL and DGS shares similar body movements. Training on multilingual data helps the model better understand and generalize across these cross-lingual signs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 567, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 567, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 295, + 664 + ], + "type": "text", + "content": "Decoupled Tokenizer. Inspired by the multi-cue characteristics of sign languages, we introduce a decoupled tokenizer that discretizes sign motion sequences for different body parts. In Table 4, we present an ablation study on various configurations: a single VQ-VAE for whole-body motions and one VAE for the body and another for both hands. The findings indicate that progressively separating the body and hands leads to consistently improved performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In addition to the objective metrics presented in the previous tables, we conducted a user study involving 5 professional" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 182, + 552, + 270 + ], + "blocks": [ + { + "bbox": [ + 317, + 182, + 552, + 270 + ], + "lines": [ + { + "bbox": [ + 317, + 182, + 552, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 552, + 270 + ], + "type": "image", + "image_path": "c35c6ee4f9fbc69d64e8273383c231c232e8706ef84d9a7677c68d5358a69fd0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 272, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 306 + ], + "type": "text", + "content": "Figure 7. User study with professional ASL and CSL signers. We report average ratings of S-MotionGPT, our proposed SOKE, and ground truth motions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 422 + ], + "type": "text", + "content": "ASL signers and 4 professional CSL signers. These signers were asked to rate the alignment between generated signs and text annotations on a scale from 1 to 10, with higher scores indicating better semantic conveyance of the texts. Ground truth sign motions were also included in the study. Specifically, we provided 15 generated signs from both the baseline, S-MotionGPT, and our proposed method, SOKE. The order of the signs was randomly shuffled to prevent potential bias." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 556, + 509 + ], + "type": "text", + "content": "As shown in Figure 7, our method achieved an encouraging average rating of 5.32/6.45, significantly surpassing the baseline, which received a low score of 1.36/2.70 on How2Sign and CSL-Daily, respectively. These results align with the above qualitative and quantitative evaluations, indicating the potential of our method in constituting a two-way communication system between the deaf and hearing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we propose SOKE, a retrieval-enhanced multilingual sign language generator. Unlike existing methods that view sign language generation as a visual content generation task, we emphasize its linguistic nature, characterized by discrete semantic units. To capture the multi-cue characteristics of sign languages, we employ a decoupled tokenizer to map continuous motions into discrete tokens for the upper body and both hands. To balance efficiency and generation quality, we propose a novel multi-head decoding method that reduces decoding steps by two-thirds while dynamically integrating information from different body parts. Additionally, we present a simple yet effective retrieval-enhanced approach by leveraging external sign dictionaries. Extensive evaluations on a combined ASL-CSL-DGS dataset validate the effectiveness of our method." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23813" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgements. S. Zafeiriou and part of the research was funded by the EPSRC Fellowship DEFORM (EP/S010203/1), EPSRC Project GNOMON (EP/X011364/1) and Turing AI Fellowship (EP/Z534699/1). R.A. Potamias and R. Zuo were supported by EPSRC Project GNOMON (EP/X011364/1). This research was also supported by the Sovereign AI AIRR programme. We would also like to thank anonymous signers who participated in the user study." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 210, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 210, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 294, + 243 + ], + "type": "text", + "content": "[1] Rotem Shalev Arkushin, Amit Moryossef, and Ohad Fried. Ham2pose: Animating sign language notation into pose sequences. In CVPR, pages 21046-21056, 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 244, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 244, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 61, + 244, + 294, + 298 + ], + "type": "text", + "content": "[2] Vasileios Baltatzis, Rolandos Alexandros Potamias, Evangelos Ververas, Guanxiong Sun, Jiankang Deng, and Stefanos Zafeiriou. Neural sign actors: A diffusion model for 3d sign language production from text. In CVPR, pages 1985-1995, 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 300, + 294, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 294, + 332 + ], + "type": "text", + "content": "[3] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In CVPR, 2018. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 334, + 294, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 294, + 366 + ], + "type": "text", + "content": "[4] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. In AAAI, pages 17754-17762, 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 368, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 294, + 409 + ], + "type": "text", + "content": "[5] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In CVPR, pages 18000-18010, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 412, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 294, + 455 + ], + "type": "text", + "content": "[6] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In CVPR, pages 5120-5130, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "type": "text", + "content": "[7] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 533 + ], + "type": "text", + "content": "[8] Wenhao Ding, Yulong Cao, Ding Zhao, Chaowei Xiao, and Marco Pavone. Realgen: Retrieval augmented generation for controllable traffic scenarios. In ECCV, pages 93-110, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 567 + ], + "type": "text", + "content": "[9] Lu Dong, Lipisha Chaudhary, Fei Xu, Xiao Wang, Mason Lary, and Ifeoma Nwogu. Signavatar: Sign language 3d motion reconstruction and generation. In FG, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "text", + "content": "[10] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In CVPR, pages 2735-2744, 2021. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "type": "text", + "content": "[11] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In SIGKDD, pages 6491-6501, 2024. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] Sen Fang, Chunyu Sui, Xuedong Zhang, and Yapeng Tian. SignDiff: Learning diffusion models for American sign language production. arXiv preprint arXiv:2308.16082, 2023. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Sen Fang, Lei Wang, Ce Zheng, Yapeng Tian, and Chen Chen. Signllm: Sign languages production large language models. arXiv preprint arXiv:2405.10718, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 538, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 538, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 538, + 117 + ], + "type": "text", + "content": "[14] fxsjy. https://github.com/fxsjy/jieba. In *Github*, 2012. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[15] Aditya Golatkar, Alessandro Achille, Luca Zancato, Yu-Xiang Wang, Ashwin Swaminathan, and Stefano Soatto. Cpr: Retrieval augmented generation for copyright protection. In CVPR, pages 12374-12384, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[16] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In CVPR, pages 18362-18372, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, pages 580-597, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[18] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "text", + "content": "[19] Hezhen Hu, Wengang Zhou, and Houqiang Li. Hand-model-aware sign language recognition. In AAAI, pages 1558-1566, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[20] Hezhen Hu, Weichao Zhao, Wengang Zhou, and Houqiang Li. SignBERT+: Hand-model-aware self-supervised pretraining for sign language understanding. TPAMI, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[21] Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. Towards fast and high-quality sign language production. In MM, pages 3172-3181, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[22] Wencan Huang, Zhou Zhao, Jinzheng He, and Mingmin Zhang. Dualsign: Semi-supervised sign language production with balanced multi-modal multi-task dual transformation. In MM, pages 5486-5495, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[23] Euijun Hwang, Jung-Ho Kim, and Jong C Park. Nonautoregressive sign language production with gaussian space. In BMVC, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "text", + "content": "[24] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. NeurIPS, 36:20067-20079, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 505, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 548 + ], + "type": "text", + "content": "[25] Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. In EMNLP, pages 7969-7992, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 549, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 549, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 549, + 553, + 592 + ], + "type": "text", + "content": "[26] Peiqi Jiao, Yuecong Min, Yanan Li, Xiaotao Wang, Lei Lei, and Xilin Chen. Cosign: Exploring co-occurrence signals in skeleton-based continuous sign language recognition. In ICCV, pages 20676–20686, 2023. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "text", + "content": "[27] Peiqi Jiao, Yuecong Min, and Xilin Chen. Visual alignment pre-training for sign language translation. In ECCV, pages 349-367, 2024. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[28] Hamid Reza Vaezi Joze and Oscar Koller. MS-ASL: A large-scale data set and benchmark for understanding American sign language. In BMVC, 2019. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[29] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23814" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[30] Gabriele Langer, Anke Müller, Sabrina Wähl, Felicitas Otte, Lea Sepke, and Thomas Hanke. Introducing the DW-DGS – the digital dictionary of DGS. In Proceedings of the LREC-COLING 2024 11th Workshop on the Representation and Processing of Sign Languages: Evaluation of Sign Language Resources, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "text", + "content": "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. NeurIPS, 33:9459-9474, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 239 + ], + "type": "text", + "content": "[32] Dongxu Li, Cristian Rodriguez, Xin Yu, and Hongdong Li. Word-level deep sign language recognition from video: A new large-scale dataset and methods comparison. In WACV, pages 1459-1469, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "text", + "content": "[33] Zecheng Li, Wengang Zhou, Weichao Zhao, Kepeng Wu, Hezhen Hu, and Houqiang Li. Uni-sign: Toward unified sign language understanding at scale. In ICLR, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 318 + ], + "type": "text", + "content": "[34] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. In CVPR, pages 21159-21168, 2023. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "type": "text", + "content": "[35] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. TACL, 8:726-742, 2020. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 432 + ], + "type": "text", + "content": "[37] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In ICML, pages 32939-32977, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 488 + ], + "type": "text", + "content": "[38] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "text", + "content": "[39] Mathias Müller, Zifan Jiang, Amit Moryossef, Annette Rios Gonzales, and Sarah Ebling. Considerations for meaningful sign language machine translation based on glosses. In ACL, pages 682-693, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[40] Zhe Niu, Ronglai Zuo, Brian Mak, and Fangyun Wei. A Hong Kong sign language corpus collected from sign-interpreted tv news. In LREC-COLING, pages 636-646, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "type": "text", + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 667 + ], + "type": "text", + "content": "[42] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, pages 10975-10985, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[43] Rolandos Alexandros Potamias, Jinglei Zhang, Jiankang Deng, and Stefanos Zafeiriou. Wilor: End-to-end 3d hand localization and reconstruction in-the-wild. In CVPR, 2025. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[44] Fan Qi, Yu Duan, Changsheng Xu, and Huaiwen Zhang. Signgen: End-to-end sign language video generation with latent diffusion. In ECCV, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[47] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Progressive transformers for end-to-end sign language production. In ECCV, pages 687-705, 2020. 1, 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "type": "text", + "content": "[48] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Adversarial training for multi-channel sign language production. In BMVC, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "text", + "content": "[49] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Continuous 3D multi-channel sign language production via progressive transformers and mixture density networks. IJCV, 129(7):2113-2135, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 360 + ], + "type": "text", + "content": "[50] Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. Signing at scale: Learning to co-articulate signs for large-scale photo-realistic sign language production. In CVPR, pages 5141-5151, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[51] Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 436 + ], + "type": "text", + "content": "[52] Shelly Sheynin, Oron Ashual, Adam Polyak, Uriel Singer, Oran Gafni, Eliya Nachmani, and Yaniv Taigman. knndiffusion: Image generation via large-scale retrieval. In ICLR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[53] Tongkai Shi, Lianyu Hu, Fanhua Shang, Jichao Feng, Peidong Liu, and Wei Feng. Pose-guided fine-grained sign language video generation. In ECCV, pages 392-409, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 504 + ], + "type": "text", + "content": "[54] Stephanie Stoll, Armin Mustafa, and Jean-Yves Guillemaut. There and back again: 3d sign language generation from text using back-translation. In 3DV, pages 187–196, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 505, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 557 + ], + "type": "text", + "content": "[55] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel Ni, Heung-Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning of large language model on knowledge graph. In ICLR, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[56] Sihan Tan, Taro Miyazaki, and Kazuhiro Nakadai. Multilingual gloss-free sign language translation: Towards building a sign language foundation model. In ACL, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 635 + ], + "type": "text", + "content": "[57] Shengeng Tang, Richang Hong, Dan Guo, and Meng Wang. Gloss semantic-enhanced network with online backtranslation for sign language production. In MM, pages 5630-5638, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[58] Shengeng Tang, Jiayi He, Lechao Cheng, Jingjing Wu, Dan Guo, and Richang Hong. Discrete to continuous: Generating smooth transition poses from sign language observations. In CVPR, 2025. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[59] Shengeng Tang, Jiayi He, Dan Guo, Yanyan Wei, Feng Li, and Richang Hong. Sign-idd: Iconicity disentangled diffusion for sign language production. In AAAI, 2025. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23815" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[60] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 129 + ], + "type": "text", + "content": "[61] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "text", + "content": "[62] Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, et al. Searching for best practices in retrieval-augmented generation. In EMNLP, pages 17716-17736, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 219 + ], + "type": "text", + "content": "[63] Fangyun Wei and Yutong Chen. Improving continuous sign language recognition with cross-lingual signs. In ICCV, pages 23612-23621, 2023. 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 253 + ], + "type": "text", + "content": "[64] Ryan Wong, Necati Cihan Camgoz, and Richard Bowden. Sign2gpt: Leveraging large language models for gloss-free sign language translation. In ICLR, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[65] Pan Xie, Qipeng Zhang, Peng Taiying, Hao Tang, Yao Du, and Zexian Li. G2p-ddm: Generating sign pose sequence from gloss sequence with discrete diffusion model. In AAAI, pages 6234-6242, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[66] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In CVPR, pages 13525-13536, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 366 + ], + "type": "text", + "content": "[67] Huijie Yao, Wengang Zhou, Hao Feng, Hezhen Hu, Hao Zhou, and Houqiang Li. Sign language translation with iterative prototype. In ICCV, pages 15592-15601, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 410 + ], + "type": "text", + "content": "[68] Hongwei Yi, Hualin Liang, Yifei Liu, Qiong Cao, Yandong Wen, Timo Bolkart, Dacheng Tao, and Michael J Black. Generating holistic 3d human motion from speech. In CVPR, pages 469-480, 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 444 + ], + "type": "text", + "content": "[69] Aoxiong Yin, Zhou Zhao, Weike Jin, Meng Zhang, Xingshan Zeng, and Xiaofei He. MLSLT: Towards multilingual sign language translation. In CVPR, pages 5109-5119, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[70] Aoxiong Yin, Haoyuan Li, Kai Shen, Siliang Tang, and Yueting Zhuang. T2S-GPT: Dynamic vector quantization for autoregressive sign language production from text. In ACL, 2024. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[71] Kayo Yin, Amit Moryossef, Julie Hochgesang, Yoav Goldberg, and Malihe Alikhani. Including signed languages in natural language processing. In ACL, pages 7347-7360, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[72] Zhengdi Yu, Shaoli Huang, Yongkang Cheng, and Tolga Birdal. Signavatars: A large-scale 3d sign language holistic motion dataset and benchmark. In ECCV, pages 1-19, 2024. 1, 2, 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "type": "text", + "content": "[73] Jan Zelinka and Jakub Kanis. Neural sign language synthesis: Words are our glosses. In WACV, pages 3395-3403, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[74] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Yong Zhang, Hongwei Zhao, Hongtao Lu, Xi Shen, and Ying Shan. Generating human motion from textual descriptions with discrete representations. In CVPR, pages 14730-14740, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[75] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffeuse: Text-driven human motion generation with diffusion model. IEEE TPAMI, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 633 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[76] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Mottiongpt: Finetuned llms are general-purpose motion generators. In AAAI, pages 7368-7376, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[77] Weichao Zhao, Hezhen Hu, Wengang Zhou, Jiaxin Shi, and Houqiang Li. BEST: BERT pre-training for sign language recognition with coupling tokenization. In AAAI, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[78] Weichao Zhao, Hezhen Hu, Wengang Zhou, Yunyao Mao, Min Wang, and Houqiang Li. Masa: Motion-aware masked autoencoder with semantic alignment for sign language recognition. IEEE TCSVT, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[79] Weichao Zhao, Wengang Zhou, Hezhen Hu, Min Wang, and Houqiang Li. Self-supervised representation learning with spatial-temporal consistency for sign language recognition. IEEE TIP, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[80] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining. In ICCV, pages 20871-20881, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 340 + ], + "type": "text", + "content": "[81] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for continuous sign language recognition. In AAAI, pages 13009-13016, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[82] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In CVPR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 420 + ], + "type": "text", + "content": "[83] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. TMM, pages 1-1, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 454 + ], + "type": "text", + "content": "[84] Zixiang Zhou, Yu Wan, and Baoyuan Wang. Avatargpt: All-in-one framework for motion understanding planning generation and beyond. In CVPR, pages 1357-1366, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[85] Ronglai Zuo and Brian Mak. C2SLR: Consistency-enhanced continuous sign language recognition. In CVPR, pages 5131-5140, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "text", + "content": "[86] Ronglai Zuo and Brian Mak. Improving continuous sign language recognition with consistency constraints and signer removal. ACM TOMM, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[87] Ronglai Zuo, Fangyun Wei, and Brian Mak. Natural language-assisted sign language recognition. In CVPR, pages 14890-14900, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "text", + "content": "[88] Ronglai Zuo, Fangyun Wei, Zenggui Chen, Brian Mak, Jiaolong Yang, and Xin Tong. A simple baseline for spoken language to sign language translation with 3d avatars. In ECCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[89] Ronglai Zuo, Fangyun Wei, and Brian Mak. Towards online continuous sign language recognition and translation. In EMNLP, 2024. 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23816" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_content_list.json b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b0c15652732f5088e3242a70f206d9b60e4b24d --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "Sim-DETR: Unlock DETR for Temporal Sentence Grounding", + "text_level": 1, + "bbox": [ + 186, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiajin Tang $^{1*}$ , Zhengxuan Wei $^{1*}$ , Yuchen Zhu $^{1}$ , Cheng Shi $^{1}$ , Guanbin Li $^{2}$ , Liang Lin $^{2}$ , Sibei Yang $^{2\\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ School of Computer Science and Engineering, Sun Yat-sen University {tangjj, weizhx2022}@shanghaiitech.edu.cn yangsb3@mail.sysu.edu.cn", + "bbox": [ + 99, + 179, + 924, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal sentence grounding aims to identify exact moments in a video that correspond to a given textual query, typically addressed with detection transformer (DETR) solutions. However, we find that typical strategies designed to enhance DETR do not improve, and may even degrade, its performance in this task. We systematically analyze and identify the root causes of this abnormal behavior: (1) conflicts between queries from similar target moments and (2) internal query conflicts due to the tension between global semantics and local localization. Building on these insights, we propose a simple yet powerful baseline, Sim-DETR, which extends the standard DETR with two minor modifications in the decoder layers: (1) constraining self-attention between queries based on their semantic and positional overlap and (2) adding query-to-frame alignment to bridge the global and local contexts. Experiments demonstrate that Sim-DETR unlocks the full potential of DETR for temporal sentence grounding, offering a strong baseline for future research.", + "bbox": [ + 88, + 301, + 485, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 619, + 220, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video has become a dominant media on the internet, with short-form content rapidly expanding and achieving exponential growth in reach over recent years [18-20, 24]. Instead of passively watching entire videos, users now prefer to target specific segments of interest, with natural language descriptions serving as an intuitive and flexible way to convey intent [30, 61, 68]. This amplifies interest in the research topic of temporal sentence grounding [22, 25-27, 32, 40, 46, 47, 57, 71, 81], which aims to locate one or more semantically aligned segments within an untrimmed video according to a given natural language sentence, as shown in Figure 1a.", + "bbox": [ + 89, + 646, + 482, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Early methods either align sentences with predefined temporal proposals for selection [44, 70] or directly predict moment spans through cross-modal interactions between", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27ab39c0516a1dd467464753540dfc434879fd5381ee796ba941e0d44eab9188.jpg", + "image_caption": [ + "Figure 1. Illustration of (a) temporal sentence grounding task, (b) phenomenon of random matching across multiple layers, and (c) the challenge of achieving the balance between focus on boundary for span prediction and alignment with global semantics." + ], + "image_footnote": [], + "bbox": [ + 535, + 268, + 880, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "language and frames [36, 37, 72]. Recent advances center on integrating the Detection Transformer (DETR) [2] for object detection into the temporal sentence grounding, leveraging its query-based proposal detection framework to eliminate hand-crafted proposal components and deliver superior grounding performance with high efficiency [22, 25, 26, 46, 47, 57]. They leverage moment queries—often several times more numerous than ground truth segments—in the multi-layer decoder, which are initialized randomly [26, 46, 57], generated from sentences [22, 47], or extracted as event units from the video [25], to locate and search the most suitable matches for target segments. A one-to-one label assignment is applied in the training objective to facilitate non-redundant predictions for each ground-truth segment. The overall architecture of the DETR-based temporal sentence grounding framework is shown in Figure 6a.", + "bbox": [ + 511, + 506, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, we observe that strategies generally effective for DETR in other tasks, such as object detection, like increasing the number of queries or decoder layers, do not improve performance in temporal sentence grounding and may even degrade it (Sec 3.1). This motivates us to revisit the specific characteristics of temporal sentence grounding that give rise to the distinct challenges in applying DETR to this task. In a series of investigations on query-segment alignment (Sec 3.2), we verify that performance degradation primarily stems from the difficulty in identifying dis", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author is Sibei Yang.", + "bbox": [ + 98, + 886, + 429, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22760", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tinct local boundaries for target segments that share highly similar global semantics, leading to inherent conflicts both between and within queries. By diagnosing the similarities between queries and between queries and frames, we uncover two conflicts that result in learning puzzles for queries:", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Queries corresponding to distinct target segments exhibit high similarity, leading to random matching with various targets and creating learning puzzles for the queries. The presence of multiple semantically similar target segments results in the high similarity between their corresponding queries, making them difficult to clearly distinguish (Sec 3.3). This causes each query to potentially match well with multiple segments, leading to significant randomness in the one-to-one matching for each target segment (Figure 1b). We observe that across different decoder layers, the queries matched for optimization can vary considerably (Sec 3.2).", + "- Each query faces an inherent conflict between encoding the global semantics of its corresponding segment and decoding its local boundary. To align well with a segment, a query must encode the segment's global semantics from the start to the end frame, such as the entire process of a person holding a cup. However, the query also needs to directly predict the segment's boundary, requiring focus on the local frames near the boundary (Figure 1c). This creates a trade-off: the query either prioritizes global semantics or local boundaries, making it challenging to optimize both simultaneously (Sec 3.3)." + ], + "bbox": [ + 89, + 185, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these findings, we propose a simple yet strong baseline for video temporal grounding, Sim-DETR. The framework builds on the standard DETR-based temporal sentence grounding architecture, with two minor yet straightforward design modifications to the decoder layers to address the two conflicts between and within queries (Figure 6), respectively. First, we adjust the self-attention weights between queries based on their pairwise correlation, encouraging indistinguishable queries to focus on different contexts, reducing their similarity, and allowing the most suitable query to gather information from related queries to refine its prediction (Figure 6b). Second, we introduce a query-to-frame matching and loss term, accounting for the alignment between the query and each frame within the segment, as shown in Figure 6c. This ensures that all frames, not just those near the boundary, contribute to segment localization. The full sequence of all target frames, from start to finish, acts as a bridge connecting the global semantics to the local boundary.", + "bbox": [ + 89, + 534, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we indeed observe that our minor modifications (1) significantly reduce conflicts between queries corresponding to different target segments (Figure 3), (2) improve the alignment between the query's global semantic attention to frames and its boundary prediction (Figure 5),", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and (3) lead to more consistent query predictions across layers (Figure 4). Furthermore, (4) increasing the number of queries and decoder layers no longer causes performance degradation (Figure 2). Finally, due to the consistent matching and prediction of queries, (5) our Sim-DETR not only achieves substantial performance gains but also accelerates convergence (as shown in Appendix). All the above observations and advantages, including the minor modifications, are intended to unlock the potential of query-based frameworks for temporal sentence grounding and provide a simple yet strong baseline for future research.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 513, + 257, + 815, + 270 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We systematically analyze the root causes of abnormal behavior in the DETR-based temporal sentence grounding framework, identifying two key conflicts: (1) target segments with highly similar semantics but distinct temporal localization create learning puzzles between queries, and (2) a trade-off between relying on global semantics for matching and local boundaries for localization within queries.", + "- Based on our analysis, we propose two modifications to form Sim-DETR: (1) a simple adjustment to the self-attention to resolve conflicts between queries and (2) the introduction of a query-to-frame matching to reconcile global semantics and local localization within queries.", + "- Our Sim-DETR, a simple yet powerful baseline, achieves consistent and significant improvements across all benchmarks. More importantly, it eliminates observed anomalies and exhibits faster convergence." + ], + "bbox": [ + 513, + 272, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 542, + 655, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Sentence Grounding (TSG) aims to identify specific video moments that align with given language expressions. As computer vision advances across neural architectures, generation, detection, segmentation, and multimodal tasks [4, 8, 21, 33, 53-55, 73, 82], temporal understanding requires task-specific optimizations. Early methods fall into proposal-free and proposal-based categories. Proposal-free methods use end-to-end frameworks by directly regressing boundary coordinates [6, 43, 76] or predicting frame-level boundary likelihood [16, 78]. Proposal-based approaches densely sample candidate segments via sliding windows [1, 12, 15, 23, 38, 80], temporal anchors [5, 32, 41, 75, 77, 81], or multimodal feature similarity [7, 34, 35, 66, 69]. Recently, DETR has shown exceptional effectiveness in detection tasks. [27, 65] pioneered DETR for TSG. Subsequent studies enhance cross-modal representations [46, 47, 57, 67], improve decoder queries with semantic/spatial information [22, 25, 58], or explore joint training with other temporal tasks [57, 71].", + "bbox": [ + 511, + 566, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection Transformers (DETR) [2] is a pioneering end-to-end framework for object detection and segmentation that employs transformers to make direct set-based pre", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22761", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dictions [14, 52, 54, 55, 60, 61, 85]. However, DETR has notable limitations, especially its high computational complexity and slow convergence, impacting its practical efficiency. To address these issues, several derivative models have emerged. Some methods [51, 63, 74, 84] improve computational efficiency by reducing redundant calculations within the DETR architecture. Other models [9, 13, 59] refine the attention mechanism for more efficient information processing and resource utilization. Additionally [39, 45, 64] enhance convergence by embedding spatial information directly into the query design, which accelerates object localization. Recent models [28, 79] tackle optimization challenges in DETR's bipartite matching process by integrating denoising strategies during training, improving both convergence and performance. Despite recent advancements in enhancing DETR for object detection, there is a lack of systematic studies addressing the unique challenges of applying DETR to TSG and the potential difficulties in query optimization.", + "bbox": [ + 89, + 90, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Probing DETR's Inherent Behavior for TSG", + "text_level": 1, + "bbox": [ + 89, + 391, + 482, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This study begins with the observation that enhancements effective for object detection in DETR [2] do not apply to temporal sentence grounding (TSG), as discussed in Sec 3.1. To explore the reasons behind DETR's failure to enhance, this section presents preliminary studies to reveal the behavior of queries in the DETR decoder, as queries are central to predicting proposals. Specifically, we examine the similarity between queries and outputs in Sec 3.2, and the attention between queries and inputs (i.e., frames) in Sec 3.3. These help reveal the underlying cause, and our modification addressing it not only eliminates the abnormal phenomena but also has a positive side effect (Sec 3.4).", + "bbox": [ + 89, + 416, + 483, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental Settings. Unless otherwise specified, our default experimental setup involves conducting statistical analyses on the full validation set of QVHighlights [27]. CG-DETR [46] and TaskWeave [71], which serve as baselines for DETR-based TSG due to their strong performance while retaining the core vanilla DETR mechanism, are used with their standard configurations.", + "bbox": [ + 89, + 598, + 483, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Abnormal Phenomena", + "text_level": 1, + "bbox": [ + 89, + 712, + 302, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. DETR is originally proposed for object detection, and increasing the number of queries and decoder layers typically improves performance. We explore whether this trend also applies to TSG.", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setting. Both CG-DETR and TaskWeave default to 10 queries, with 3 and 2 decoder layers, respectively. This experiment specifically increases the number of queries and decoder layers to isolate their effects, keeping all other hyperparameters and training protocols constant.", + "bbox": [ + 89, + 795, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Result & Discussion. As shown in Figure 2, increasing the number of queries to 15 and 20 led to progressively", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33d6f4adc556afce73df5205366cfdbf662674600a9dffdec376d23e8e356225.jpg", + "image_caption": [ + "Figure 2. The impact of the number of queries and decoder layers." + ], + "image_footnote": [], + "bbox": [ + 519, + 84, + 895, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "larger performance declines, with a drop of over $-2\\%$ with TaskWeave [71] at 20 queries. Similarly, increasing the number of decoder layers caused a drop from $-1\\%$ to $-2.5\\%$ . In contrast, our Sim-DETR mitigates this phenomenon and achieves a slight improvement, demonstrating its robustness. This observation motivates an in-depth exploration of decoder layer, particularly the role of queries (see Sec 3.2).", + "bbox": [ + 511, + 252, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Similar Target Segments Cause Query Conflicts", + "text_level": 1, + "bbox": [ + 511, + 367, + 905, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. Initially, we hypothesize that the lack of performance improvement with increased queries is due to query redundancy. However, performance declined, suggesting that conflicts between queries, rather than redundancy, prevent compatibility with existing queries. Similarly, conflicts between queries across layers may explain the ineffectiveness of increasing the number of layers. This motivates us to explore (1) the relationships between queries to identify their potential conflicts and (2) their variations across layers to reveal the impact at different layers.", + "bbox": [ + 511, + 388, + 905, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setting. (1) To evaluate query relationships, we assess similarities between queries corresponding to the same segment (intra-segment similarity) and those from different segments (inter-segment similarity). Specifically, we identify the corresponding segment for each query based on the maximum IoU between the query's prediction and all ground-truth (GT) segments during inference. Queries whose IoU with the corresponding segment is 0.5 or greater are included in our analysis to ensure meaningful corresponding. We then compute feature similarities between queries associated with the same segment (intra) and those matched to different segments (inter). (2) To analyze variations across layers, we introduce the metric of cross-layer matching consistency. Specifically, during training optimization, each GT segment is uniquely matched to a query through bipartite matching at each decoding layer. We assess match consistency by measuring the proportion of segments that retain their corresponding queries across two consecutive decoder layers.", + "bbox": [ + 511, + 541, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Result & Discussion. (1) As shown in Figure 3, CGDETR fails to distinguish query similarities between inter- or intra-segment groups, indicating that queries are not clearly grouped with their corresponding segments. Consequently, the same query oscillates between segments,", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22762", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/12b42318272f8cd778de121aec07c3f67597410fffad46af97933919b8be2852.jpg", + "image_caption": [ + "Figure 3. Distribution of query similarity for intra-segment (blue) and inter-segment (red) pairs. Our method can effectively distinguish intra-segment and inter-segment queries, ensuring more stable query-segment associations and reducing conflicts in query assignments across different segments." + ], + "image_footnote": [], + "bbox": [ + 102, + 75, + 472, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/894e525cf266de01d9c4a5c26cfb550aa34da0aebdefda1a3efceba091c98937.jpg", + "image_caption": [ + "Figure 4. Measuring cross-layer matching consistency (y-axis, defined in Line-260) of segments across two consecutive decoding layers (x-axis). Our method achieves higher consistency." + ], + "image_footnote": [], + "bbox": [ + 109, + 287, + 460, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lacking stable and consistent corresponding. Figure 4 further supports our hypothesis, demonstrating that cross-layer matching consistency in CG-DETR is lower than ours. (2) We attribute this to multiple GT segments in a video sharing the same linguistic semantics. The high semantic similarity between GT segments results in high similarity between their corresponding queries. Figure 4 supports this, showing that the consistency for cases with multiple GT segments is lower than for those with a single segment.", + "bbox": [ + 89, + 503, + 483, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Global Matching vs. Local Localization", + "text_level": 1, + "bbox": [ + 89, + 650, + 431, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. We further question whether the inconsistency in matching is solely due to query conflicts from multiple GT segments. If so, why does the previous method still show roughly $25\\%$ mismatched queries in cases with a single target segment? Therefore, we shift our focus from conflicts between queries to conflicts within individual queries. A query serves dual roles: aligning with global linguistic semantics (global matching) and precisely localizing the segment, particularly its boundaries (local localization). This prompts us to explore whether conflicts arise from these roles within a single query.", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setting. (1) To evaluate global matching, we use the query's classification confidence score, as it reflects the confidence of the query being referenced by the global semantics of the linguistic description. (2) To evaluate local localization, we", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e21c353c2a08149735c3362c9cf8f98e645ea476b4c4879008433b2895113fb.jpg", + "image_caption": [ + "Figure 5. Global matching score (defined in Line-290) vs. local localization score (defined in Line-293). Compared to CGDETR, our method concentrates a query's attention significantly more within a single GT segment, rather than dispersing it across multiple segments, leading to improved local localization scores." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "extract attention scores between each query and frame from the cross-attention module in the last decoder layer, as they indicate query-based frame localization. We then compute the IoU score between this attention and the query's corresponding GT segment, reflecting the accuracy of the query's localization at the frame level.", + "bbox": [ + 511, + 343, + 905, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Result & Discussion. As shown in Figure 5, CG-DETR (marked in red) may yield low local localization scores despite high global matching scores, indicating that while queries align well with linguistic descriptions, they fail to achieve precise localization. Further analysis reveals that query attention may span frames in multiple GT segments (see Figure 5 \"Attention\"), leading to a low local score for the query's specific GT segment it corresponds to.", + "bbox": [ + 511, + 433, + 905, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Positive Side Effect", + "text_level": 1, + "bbox": [ + 511, + 564, + 697, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Convert Abnormal to Normal: Our Sim-DETR effectively (1) distinguishes queries within the same target segment and across different segments in Figure 3, (2) achieves consistent matching across decoder layers in Figure 4, and (3) aligns with global semantics while ensuring precise localization in Figure 5.", + "bbox": [ + 511, + 587, + 905, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positive Side Effect: (1) More importantly, Sim-DETR maintains stable and robust performance with increased queries and decoder layers (see Figure 2). (2) By eliminating the abnormal phenomena, it also achieves significantly faster convergence compared to previous methods, as shown in Appendix C.", + "bbox": [ + 511, + 679, + 905, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 784, + 604, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present a simple yet effective TSG baseline that resolves query conflicts in existing DETR-based TSG frameworks [22, 25, 27, 46, 47, 57, 67, 71] through two minor yet key modifications. First, we introduce the standard DETR-based [2] TSG architecture (Sec 4.1). Sec 4.2 introduces our first modification, which adjusts the self-attention", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22763", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b4b581d8f656c715184403c618e5cfacbb7a4a6b3c9e602f4743fcad84baa5f.jpg", + "image_caption": [ + "Figure 6. Framework of the proposed Sim-DETR, featuring two simple yet powerful modifications: Query Grouping and Ranking to mitigate cross-query conflicts and Global-Local Bridging to address global-local conflicts within each query." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mechanism between queries in the decoder to avoid conflicts among them, based on the observations presented in Sec 3.2. Next, Sec 4.3 presents another improvement that addresses the conflict between global matching and local localization within the query (discussed in Sec 3.3) by introducing query-to-frame matching loss, which serves as the bridge between the global and local perspectives.", + "bbox": [ + 89, + 412, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Standard DETR-based TSG Architecture", + "text_level": 1, + "bbox": [ + 89, + 529, + 444, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In TSG, the goal is to identify video segments with temporal spans that correspond to a given language expression.", + "bbox": [ + 89, + 551, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Extraction. Previous DETR-based works [25, 47, 57, 71] employ CLIP [49] and SlowFast [10] to extract video features, represented as $\\mathcal{T} \\in \\mathbb{R}^{N \\times C}$ , where $N$ and $C$ stands for the number of frames and the hidden dimension, respectively. For the language expression, typically a sentence with $L$ tokens, the CLIP text encoder is used to obtain token-level features, resulting in $\\mathcal{W} \\in \\mathbb{R}^{L \\times C}$ .", + "bbox": [ + 89, + 582, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Encoder. After feature extraction, DETR-based TSG methods [22, 46, 47, 57] employ a multimodal encoder to fuse these video and language features, producing multimodal video features, denoted as $\\hat{\\mathcal{T}}\\in \\mathbb{R}^{N\\times C}$ . The key component of the multimodal encoder is a cross-attention mechanism [62], where video features $\\mathcal{T}$ act as queries, while word features $\\mathcal{W}$ serve as keys and values.", + "bbox": [ + 89, + 688, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Sentence Decoding. In the decoder, the model initializes a set of query to gather global information from the video, which is then decoded into local spans in the span prediction head. Specifically, given initialized queries $\\mathcal{Q} \\in \\mathbb{R}^{M \\times C}$ , where $M$ is the query number, and the multimodal video features $\\hat{\\mathcal{T}}$ , the decoder facilitates interaction between queries and gathers global information through interleaved", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-attention and cross-attention modules, as follows:", + "bbox": [ + 513, + 412, + 870, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {k} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {k} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {v} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {v} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right). \\end{array} \\right. \\\\ \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {q} ^ {\\mathrm {C A}} (\\mathcal {Q} ^ {\\mathrm {S A}}), \\hat {\\mathcal {T}} _ {k} = \\mathrm {F C} _ {k} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}}), \\hat {\\mathcal {T}} _ {v} = \\mathrm {F C} _ {v} (\\hat {\\mathcal {T}}), \\\\ \\mathcal {Q} _ {o} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {C A}} \\left(\\mathrm {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}} _ {k}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\hat {\\mathcal {T}} _ {v}\\right). \\end{array} \\right. \\\\ \\hat {\\mathcal {Q}} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathcal {Q} ^ {\\mathrm {C A}}\\right) + \\mathcal {Q}\\right), \\\\ B = \\mathcal {H} _ {\\text {s p a n}} (\\hat {\\mathcal {Q}}), \\quad B = \\left\\{b _ {1}, \\dots , b _ {M} \\right\\}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 439, + 903, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{FC}(\\cdot)$ denotes the fully connected projection layer, with SA and CA representing self-attention and cross-attention, respectively. Subscripts $q$ , $k$ , and $v$ indicate query, key, and value computations. MLP( $\\cdot$ ) represents a nonlinear mapping with an activation function. $\\mathcal{H}_{\\mathrm{span}}(\\cdot)$ denotes the span prediction head, which takes the updated query features $\\hat{\\mathcal{Q}}$ to predict the span set $B$ , where each $b_i \\in \\mathbb{R}^2$ represents the (start, end) of a predicted span.", + "bbox": [ + 511, + 594, + 906, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Query Grouping and Ranking", + "text_level": 1, + "bbox": [ + 511, + 726, + 782, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we address the query conflicts identified in Sec 3.2: (1) the inability to distinguish between intra- and inter-segment queries (Figure 3), leading to query oscillation between segments without stable correspondence (Line-266), and (2) inconsistencies in segment-query matching across layers, where different queries may be matched to ground-truth (GT) segments at each layer for optimization, hindering effective learning (Figure 4).", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To differentiate queries across segments, the challenge arises from GT segments that inherently share identical lin", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22764", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a450f95a57d73051b75d765db4ca14f10e9081b2460426e601ecea0dec4deb5.jpg", + "image_caption": [ + "Figure 7. Overview of our Query Grouping and Ranking." + ], + "image_footnote": [], + "bbox": [ + 109, + 95, + 460, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "guistic semantics, leading to highly similar queries. Therefore, a natural solution is to directly group queries, assigning each group to a specific GT segment for training and prediction. Instead of hard grouping with predefined rules or extra hyperparameters, we learn a soft grouping based on predicted temporal spans of queries, where closer spans indicate potential corresponding for the same segment. Specifically, for any two queries $q_{i}$ and $q_{j}$ , we compute their span distance, which is then used in Eq. 4 to learn their grouping:", + "bbox": [ + 89, + 220, + 483, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i, j} ^ {\\text {i n t r a}} = \\| b _ {i} - b _ {j} \\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 364, + 480, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ denotes the L2 norm. Notably, we use L2 distance instead of the commonly used L1 because, when two spans are close (normalized distances $\\leq 1$ ), L2 distance decreases more sharply than L1, imposing a smaller penalty on minor differences (see Appendix G).", + "bbox": [ + 89, + 388, + 483, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Meanwhile, to enhance cross-layer matching consistency, we draw inspiration from EASE-DETR [14]'s query ranking strategy in object detection. By ranking queries within a segment at each layer, the leading query is prioritized for segment matching in subsequent layers. However, directly ranking queries by prediction scores, as in EASE-DETR, proves ineffective (see Appendix D). As analyzed in Sec. 3.3, high prediction/confidence scores do not guarantee precise predictions and may correspond to poor localization. Instead, we introduce an IoU head to assess localization accuracy and rank queries based on both factors, as follows:", + "bbox": [ + 89, + 464, + 483, + 641 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\text {r a n k}} \\left(q _ {i}, q _ {j}\\right) = \\left\\{ \\begin{array}{l l} + 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} \\geq \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}} \\\\ - 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} < \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}}, \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 638, + 480, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^{\\mathrm{cls}}$ and $\\mathcal{P}^{\\mathrm{IoU}}$ denote the query's confidence score and IoU prediction, respectively, with subscripts $i$ and $j$ indicating queries $q_{i}$ and $q_{j}$ . $R_{\\mathrm{rank}}(q_i,q_j) = 1$ indicates that query $q_{i}$ has a higher priority than query $q_{j}$ , considering both global alignment and local localization.", + "bbox": [ + 89, + 684, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we incorporate query grouping within segments and intra-segment ranking to infer relative relationships, then reshape query-to-query self-attention following [14] to mitigate conflicts, as follows:", + "bbox": [ + 89, + 760, + 483, + 821 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {S} ^ {\\text {a t t n}} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathcal {S} ^ {\\text {i n t r a}} \\circ \\mathcal {R} _ {\\text {r a n k}}\\right), \\right. \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\text {s o f t m a x} \\left(\\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} \\left(\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}\\right) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {S} ^ {\\mathrm {a t t n}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 825, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S^{\\mathrm{attn}}$ represents query relationships, with higher values promoting high-quality queries to dominate predictions by integrating similar queries within the same group.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Global-Local Bridging", + "text_level": 1, + "bbox": [ + 511, + 147, + 723, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we address the query conflicts identified in Sec 3.3, specifically the conflict between a query's global semantic matching and local localization. Since queries overly rely on global semantics while neglecting local localization, our core idea is to reinforce fine-grained frame-level localization within the segment. Specifically, we introduce a query-to-frame matching mechanism and loss term to enhance localization beyond segment boundary prediction, ensuring alignment with in-segment frames while avoiding matches with those from other segments. The full sequence of frames within the segment, from start to finish, serves as a bridge linking global semantics to local boundaries.", + "bbox": [ + 511, + 170, + 906, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we compute the semantic similarity between each query $q_{i}$ and all frames within its corresponding ground-truth span, maximizing similarity to ensure full alignment with in-segment frames while minimizing similarity to out-segment frames. The computation process is outlined as follows,", + "bbox": [ + 511, + 352, + 905, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z = \\operatorname {s i g m o i d} (\\tau \\cdot \\cos (q _ {i}, \\hat {\\mathcal {T}})), \\\\ \\mathcal {L} _ {\\mathrm {b r i d g e}} = \\lambda_ {\\mathrm {b r i d g e}} \\frac {- \\sum_ {j} z _ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}{\\sum_ {j} z _ {j} \\left(1 - \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}\\right) + \\sum_ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}, \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 445, + 906, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $z \\in \\mathbb{R}^N$ donates query-to-frame similarity, calculated using the cosine similarity between the query $q_i$ and the $N$ frame features $\\tilde{T}$ defined in Line-347, and $\\tau$ is a learnable scaling coefficient. $b_i^{\\mathrm{gt}}$ is the ground-truth span corresponding to query $q_i$ , while $\\mathbb{I}[b_i^{\\mathrm{gt}}] \\in \\mathbb{R}^N$ is an indicator function that assigns 1 to frames within the ground truth and 0 otherwise, with $j$ representing the frame index. Thus, the numerator in $\\mathcal{L}_{\\mathrm{bridge}}$ , i.e., $-\\sum_{j} z_j \\mathbb{I}[b_i^{\\mathrm{gt}}]_j$ , encourages the query to be similar to every frame within its corresponding segment. Meanwhile, the denominator's first term, $\\sum_{j} z_j (1 - \\mathbb{I}[b_i^{\\mathrm{gt}}]_j)$ , represents query $q_i$ 's similarity to out-segment frames, promoting its minimization. The denominator's second term, $\\sum_{j} \\mathbb{I}[b_i^{\\mathrm{gt}}]_j$ , represents the length of the ground-truth segment, serving as a normalization factor. And $\\lambda_{\\mathrm{bridge}}$ is a hyperparameter that controls the loss weight.", + "bbox": [ + 511, + 518, + 906, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Overall Loss", + "text_level": 1, + "bbox": [ + 511, + 771, + 648, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall loss is: $\\mathcal{L} = \\mathcal{L}_{\\mathrm{MD}} + \\lambda_{\\mathrm{bridge}}\\mathcal{L}_{\\mathrm{bridge}} + \\lambda_{\\mathrm{iou}}\\mathcal{L}_{\\mathrm{iou}}$ where $\\mathcal{L}_{\\mathrm{MD}}$ includes the L1, gIoU, classification, and saliency losses, as in Moment DETR [26]. $\\mathcal{L}_{\\mathrm{iou}}$ is the loss for the IoU head introduced in Line-406, designed to assist in ranking queries using $\\mathcal{P}_i^{\\mathrm{IoU}}$ and $\\mathcal{P}_j^{\\mathrm{IoU}}$ in Eq 3. The IoU head implementation follows [25, 58]. Ablations for hyperparameters $\\lambda_{\\mathrm{bridge}}$ and $\\lambda_{\\mathrm{iou}}$ are detailed in Appendix E.", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22765", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/00b8c3fc81e571754b6a029da9730fd32ce567cb58641bd41143254d1f61eeba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodtestval
R1mAPR1mAP
@0.5@0.7@0.5@0.75Avg.@0.5@0.7@0.5@0.75Avg.
M-DETR [27]NeurIPS'2152.8933.0254.8229.4030.7353.9434.84--32.20
UMT [40]CVPR'2256.2341.1853.8337.0136.1260.2644.26--38.59
QD-DETR [47]CVPR'2362.4044.9862.5239.8839.8662.6846.6662.2341.8241.22
UniVTG [32]ICCV'2358.8640.8657.6035.5935.4759.74---36.13
EaTR [22]ICCV'23-----61.3645.7961.8641.9141.74
MomentDiff [29]NeurIPS'2357.4239.6654.0235.7335.95-----
TR-DETR [57]AAAI'2464.6648.9663.9843.7342.6267.1051.4866.2746.4245.09
UVCOM [67]CVPR'2463.5547.4763.3742.6743.1865.1051.81--45.79
TaskWeave [71]CVPR'24-----64.2650.0665.3946.4745.38
BAM-DETR [25]ECCV'2462.7148.6464.5746.3345.3665.1051.6165.4148.5647.61
CG-DETR [46]Arxiv'2465.4348.3864.5142.7742.8667.3552.0665.5745.7344.93
SpikeMba [31]Arxiv'2464.1349.42-43.6743.7965.3251.33-44.9644.84
Sim-DETR(Ours)67.6450.9167.8147.5946.9369.4854.0669.7051.1149.50
", + "bbox": [ + 122, + 88, + 870, + 300 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4da9c5e7fe8791a0571d909483dc2142c28ab46ab3b2875cf1a3e915448db4b.jpg", + "table_caption": [ + "Table 1. Experimental results on the QVHighlights benchmark, comparing ours with state-of-the-art methods." + ], + "table_footnote": [], + "table_body": "
MethodTACoSCharades-STA
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [81]40.0127.9912.9227.2258.7646.0227.5041.25
M-DETR [27]37.9724.6711.9725.4965.8352.0730.5945.54
MomentDiff [29]44.7833.68---55.5732.42-
QD-DETR [47]-----57.3132.55-
UniVTG [32]51.4434.9717.3533.6070.8158.0135.6550.10
CG-DETR [46]52.2339.6122.2336.4870.4358.4436.3450.13
UVCOM [67]-36.3923.32--59.2536.64-
SpikeMba [31]51.9839.3422.8335.8171.2459.6536.1251.74
Sim-DETR(Ours)57.0642.7926.8239.4473.0961.3439.6252.56
", + "bbox": [ + 133, + 327, + 859, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 506, + 223, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 536, + 410, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. Following [27, 31, 32, 47, 67], we conduct comprehensive evaluation experiments across three datasets: QVHighlights [27], Charades-STA [12], and TACoS [50]. The QVHighlights dataset contains 10,310 text expressions and 18,367 distinct events, with most expressions corresponding to multiple events. Charades-STA is derived from a collection of 9,848 indoor scene videos, encompassing 16,128 events. TACoS includes only 273 videos, with an average length approaching five minutes per video. For a fair comparison, we adopt the same evaluation metrics as in previous methods. On the QVHighlights benchmark, we use Recall and Mean Average Precision (mAP) as primary metrics, while for Charades-STA and TACoS, we employ Recall and Mean Intersection over Union (mIoU).", + "bbox": [ + 89, + 563, + 482, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We set the input video resolution to $224 \\times 224$ and represent each video by concatenating the CLIP [49] [CLS] tokens with SlowFast [10] features. For the Charades-STA and TACoS datasets, we additionally constructed two experiments using I3D [3] and VGG [56] as visual encoders, with Glove [48] embeddings for text features. The model is trained for 200 epochs on a single NVIDIA A40 GPU, using the AdamW [42] optimizer with", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/631a2d318618f44d1e3ca3485136cc78eb3878a5c6267881672b5b8769289f72.jpg", + "table_caption": [ + "Table 2. Comparison results on the TACoS and Charades-STA benchmarks." + ], + "table_footnote": [], + "table_body": "
BackboneMethodR@0.5R@0.7
VGG [56]2D-TAN [81]40.9422.85
FVMR [11]42.3624.14
SSRN [83]46.7227.98
UMT [40]48.3129.25
MomentDiff [29]51.9428.25
QD-DETR [47]52.7731.13
TR-DETR [57]53.4730.81
CG-DETR [46]55.2234.19
I3D [3]Sim-DETR(Ours)55.9735.38
MAN [77]46.5322.72
VSLNet [78]47.3130.19
QD-DETR [47]50.6731.02
TR-DETR [57]55.5133.66
TaskWeave [71]53.3631.40
Sim-DETR(Ours)57.5535.73
", + "bbox": [ + 516, + 510, + 903, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Experimental results with VGG or I3D as backbone.", + "bbox": [ + 526, + 715, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an initial learning rate of 1e-4. Note that, for a fair comparison, the total number of layers in our Sim-DETR is set to 6, the same as recent works [22, 46, 67].", + "bbox": [ + 511, + 733, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 787, + 880, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tables 1-3, we report detailed evaluation results across the QVHighlights, Charades-STA, and TACoS benchmarks. Our Sim-DETR consistently outperforms all state-of-the-art methods (SOTAs) across all metrics.", + "bbox": [ + 511, + 810, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on QVHighlights are shown in Table 1. Compared to the latest published work, BAM-DETR [25], our Sim", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22766", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48108b3fd71e1a49e8dc368644797bf844529cbb1f5ca583a3b5514000fcecca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline65.4850.8465.3445.8244.97
+ Liou66.5851.9465.6846.3045.22
+ QGR68.7752.2667.7248.2847.03
+ GLB67.1652.7768.5848.8848.17
+ QGR,GLB69.4854.0669.7051.1149.50
", + "bbox": [ + 93, + 88, + 454, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c29056dd4cef746c80674b6a3317852702f0d94bfd3cca3376298f179bba9d9.jpg", + "table_caption": [ + "Table 4. Ablation study on components." + ], + "table_footnote": [], + "table_body": "
Inner RelevanceOuter GlobalOuter LocalmAP
(1)span border distconfidence predIoU pred49.50
(2)center distconfidence predIoU pred48.59
(3)span IoU48.94
(4)span border distw/oIoU pred48.93
(5)span border distconfidence predw/o48.66
(6)center pred48.97
", + "bbox": [ + 93, + 213, + 455, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Impact of different measures.", + "bbox": [ + 161, + 333, + 388, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DETR outperforms it by an average of $+2.88\\%$ across all metrics. More importantly, it achieves remarkable gains of $+4.93\\%$ and $+2.27\\%$ in R1@0.5 and R1@0.7, respectively. In comparison TR-DETR [57], which uses standard DETR [2] as its basic framework and is used as our baseline model, our Sim-DETR achieves an average improvement of $+4.31\\%$ and $+4.41\\%$ in mAP for test and validation, respectively. These improvements underscore the effectiveness of our Sim-DETR in fundamentally addressing conflicts within and between queries. We further conduct a comparison with SpikeMba [31], a contemporary method built on the Mamba [17] framework. Without any complex module designs, the consistent improvements across all metrics highlight the simplicity and effectiveness of our Sim-DETR. Sim-DETR establishes an efficient baseline for the task, significantly advancing research in TSG.", + "bbox": [ + 89, + 354, + 482, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Charades-STA and TACoS are presented in Table 2, where we consistently outperform all methods on both datasets. Compared to the state-of-the-art methods, we achieve a $+1.97\\%$ improvement on the Charades-STA dataset (with SpikeMba as the SOTA) and a $+3.89\\%$ improvement on the TACoS dataset (with CG-DETR [46] as the SOTA). These enhancements further demonstrate the outstanding capability of our method in temporal localization. More importantly, they confirm that our Sim-DETR exhibits significant performance advantages across various scenarios, types, and lengths of videos, highlighting its robustness and generalizability.", + "bbox": [ + 89, + 597, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results with Different Backbones. Following previous works, we conduct two additional experiments on the QVHighlights dataset. These experiments use VGG and I3D for visual feature extraction and use Glove [48]embedding as text features. Table 3 presents the detailed results of these two experiments, where we achieve state-of-the-art results in both cases. Notably, for the video feature extractor I3D, we surpass the latest pub", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lished method by $+4.26\\%$ . These results further demonstrate the excellent generalization capabilities of our SimDETR across different backbones.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 145, + 666, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Roadmap for building a simple yet effective TSG baseline is presented in Table 4. We employ TR-DETR without MR2HD module as our baseline, which achieves an mAP of $44.97\\%$ . Since our Query Grouping and Ranking (QGR) utilizes $\\mathcal{L}_{\\mathrm{iou}}$ proposed by [25, 58], we additionally conduct an ablation study to isolate the impact of $\\mathcal{L}_{\\mathrm{iou}}$ . (1) QGR: To mitigate conflicts and optimization challenges arising from overly similar semantics between queries, we introduce the QGR, which improves the baseline by $+2.07\\%$ on mAP, representing a subtle yet significant enhancement. (2) Global-Local Bridging (GLB): To address conflicts and ambiguities between global semantics and local localization within queries, we introduce GLB to enable seamless transformation through query-to-frame alignment. Experimental results show a performance boost to $48.17\\%$ on mAP, indicating the effectiveness of GLB in bridging global-to-local transformations. However, without QGR to address semantic conflicts in temporal DETR, the recall predictably decreased by $-3.1\\%$ . (3) Overall Framework: With these adjustments, our Sim-DETR achieves a remarkable SOTA performance: $49.50\\%$ on mAP and $69.48\\%$ on R1@0.5.", + "bbox": [ + 511, + 167, + 903, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Effects of Different Conflict Metrics in QGR. Table 5 summarizes five experiments on conflict metrics in QGR. (1) The first row displays our final implementation, achieving $49.50\\%$ mAP. (2) & (3) We evaluate L2 center distance and span IoU for inter-query relationships. Center distance alone yields $48.59\\%$ mAP, highlighting its inadequacy in segment representation. Span IoU outperforms center distance but lags by $0.56\\%$ compared to boundary distance, likely due to IoU's broader tolerance, which reduces precision in query relevance. (4) To assess outer global semantic alignment, we remove global semantics, resulting in a $-0.57\\%$ drop, emphasizing their role in query interaction and quality. (5) Finally, we test local location metrics in query prioritization. Excluding the outer local metric significantly lowers performance $(-0.84\\%$ mAP), indicating that without local information, queries encounter conflicts due to overly generalized global semantics, leading to random matches. Consistent with (2), center localization accuracy shows similar limitations, with a $0.53\\%$ decrease.", + "bbox": [ + 511, + 484, + 903, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 784, + 632, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a concise and efficient temporal sentence grounding framework Sim-DETR, introducing two simple modifications to the decoder for resolving conflicts both between and within queries. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods across all benchmarks.", + "bbox": [ + 511, + 810, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22767", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No.62206174).", + "bbox": [ + 89, + 90, + 485, + 119 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 148, + 187, + 165 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 8", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7", + "[4] Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[5] Jingyuan Chen, Xinpeng Chen, Lin Ma, Zequn Jie, and Tat-Seng Chua. Temporally grounding natural sentence in video. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 162-171, 2018. 2", + "[6] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10551-10558, 2020. 2", + "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 2", + "[8] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711-13722, 2024. 2", + "[9] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 3", + "[10] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 7", + "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 7", + "[12] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + ], + "bbox": [ + 93, + 174, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 7", + "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. arXiv preprint arXiv:2101.07448, 2021. 3", + "[14] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 3, 6", + "[15] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 245-253. IEEE, 2019. 2", + "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 2", + "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 8", + "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 1", + "[19] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7366–7375, 2018.", + "[20] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. Advances in Neural Information Processing Systems, 36:26135-26158, 2023. 1", + "[21] Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. Mvtokenflow: High-quality 4d content generation using multiview token flow. arXiv preprint arXiv:2502.11697, 2025. 2", + "[22] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 1, 2, 4, 5, 7", + "[23] Bin Jiang, Xin Huang, Chao Yang, and Junsong Yuan. Cross-modal video moment retrieval with spatial and language-temporal attention. In Proceedings of the 2019 on international conference on multimedia retrieval, pages 217-225, 2019. 2", + "[24] Manjin Kim, Heeseung Kwon, Chunyu Wang, Suha Kwak, and Minsu Cho. Relational self-attention: What's missing in attention for video understanding. Advances in Neural Information Processing Systems, 34:8046–8059, 2021. 1", + "[25] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 1, 2, 4, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22768", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 6", + "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Neurips, pages 11846-11858, 2021. 1, 2, 3, 4, 7", + "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3", + "[29] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In Neurips, 2023. 7", + "[30] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 1", + "[31] Wenrui Li, Xiaopeng Hong, Ruiqin Xiong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 7, 8", + "[32] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 7", + "[33] Liang Lin, Pengxiang Yan, Xiaoqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 2", + "[34] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4536-4545, 2022. 2", + "[35] Daizong Liu, Xiaoye Qu, Jianfeng Dong, and Pan Zhou. Adaptive proposal generation network for temporal sentence localization in videos. arXiv preprint arXiv:2109.06398, 2021. 2", + "[36] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1665-1673, 2022. 1", + "[37] Daizong Liu, Xiaoye Qu, and Wei Hu. Reducing the vision and language bias for temporal sentence grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4092-4101, 2022. 1", + "[38] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2", + "[39] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3", + "[40] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 1, 7", + "[41] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. $r^2$ -tuning: Efficient image-to-video transfer learning for video temporal grounding. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[42] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 2", + "[44] Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553, 2020. 1", + "[45] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3", + "[46] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 1, 2, 3, 4, 5, 7, 8", + "[47] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2, 4, 5, 7", + "[48] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 7, 8", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7", + "[50] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 7", + "[51] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021. 3", + "[52] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22769", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 3", + "[53] Cheng Shi and Sibei Yang. Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932-2941, 2023. 2", + "[54] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 3", + "[55] Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In European Conference on Computer Vision, pages 210-226. Springer, 2024. 2, 3", + "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7", + "[57] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4998-5007, 2024. 1, 2, 4, 5, 7, 8", + "[58] Xiaolong Sun, Liushuai Shi, Le Wang, Sanping Zhou, Kun Xia, Yabing Wang, and Gang Hua. Diversifying query: Region-guided transformer for temporal sentence grounding. arXiv preprint arXiv:2406.00143, 2024. 2, 6, 8", + "[59] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3", + "[60] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23570-23580, 2023. 3", + "[61] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1, 3", + "[62] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5", + "[63] Tao Wang, Li Yuan, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4661-4670, 2021. 3", + "[64] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 3", + "[65] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore-and-match: Bridging proposal-based and proposal-free with transformer for sentence grounding in videos. arXiv preprint arXiv:2201.10168, 2022. 2", + "[66] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2986-2994, 2021. 2", + "[67] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2, 4, 7", + "[68] Chen-Wei Xie, Siyang Sun, Xiong Xiong, Yun Zheng, Deli Zhao, and Jingren Zhou. Ra-clip: Retrieval augmented contrastive language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19265–19274, 2023. 1", + "[69] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9062–9069, 2019. 2", + "[70] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Unsupervised alignment-based iterative evidence retrieval for multihop question answering. arXiv preprint arXiv:2005.01218, 2020. 1", + "[71] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18308-18318, 2024. 1, 2, 3, 4, 5, 7", + "[72] Shuo Yang and Xinxiao Wu. Entity-aware and motion-aware transformers for language-driven action localization in videos. arXiv preprint arXiv:2205.05854, 2022. 1", + "[73] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 2", + "[74] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 3", + "[75] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[76] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2", + "[77] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2, 7", + "[78] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 2, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22770", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3", + "[80] Songyang Zhang, Jinsong Su, and Jiebo Luo. Exploiting temporal relationships in video moment localization with natural language. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1230-1238, 2019. 2", + "[81] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 7", + "[82] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 2", + "[83] Jiahao Zhu, Daizong Liu, Pan Zhou, Xing Di, Yu Cheng, Song Yang, Wenzheng Xu, Zichuan Xu, Yao Wan, Lichao Sun, et al. Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514, 2023. 7", + "[84] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3", + "[85] Yuchen Zhu, Cheng Shi, Dingyou Wang, Jiajin Tang, Zhengxuan Wei, Yu Wu, Guanbin Li, and Sibei Yang. Rethinking query-based transformer for continual image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4595-4606, 2025. 3" + ], + "bbox": [ + 91, + 92, + 480, + 542 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22771", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_model.json b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b542d961403b6d6bfce56f1d7b1ef74466c504e9 --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_model.json @@ -0,0 +1,2556 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Sim-DETR: Unlock DETR for Temporal Sentence Grounding" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.925, + 0.235 + ], + "angle": 0, + "content": "Jiajin Tang\\(^{1*}\\), Zhengxuan Wei\\(^{1*}\\), Yuchen Zhu\\(^{1}\\), Cheng Shi\\(^{1}\\), Guanbin Li\\(^{2}\\), Liang Lin\\(^{2}\\), Sibei Yang\\(^{2\\dagger}\\) \n\\(^{1}\\)ShanghaiTech University \\(^{2}\\)School of Computer Science and Engineering, Sun Yat-sen University {tangjj, weizhx2022}@shanghaiitech.edu.cn yangsb3@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.486, + 0.591 + ], + "angle": 0, + "content": "Temporal sentence grounding aims to identify exact moments in a video that correspond to a given textual query, typically addressed with detection transformer (DETR) solutions. However, we find that typical strategies designed to enhance DETR do not improve, and may even degrade, its performance in this task. We systematically analyze and identify the root causes of this abnormal behavior: (1) conflicts between queries from similar target moments and (2) internal query conflicts due to the tension between global semantics and local localization. Building on these insights, we propose a simple yet powerful baseline, Sim-DETR, which extends the standard DETR with two minor modifications in the decoder layers: (1) constraining self-attention between queries based on their semantic and positional overlap and (2) adding query-to-frame alignment to bridge the global and local contexts. Experiments demonstrate that Sim-DETR unlocks the full potential of DETR for temporal sentence grounding, offering a strong baseline for future research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.222, + 0.636 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.828 + ], + "angle": 0, + "content": "Video has become a dominant media on the internet, with short-form content rapidly expanding and achieving exponential growth in reach over recent years [18-20, 24]. Instead of passively watching entire videos, users now prefer to target specific segments of interest, with natural language descriptions serving as an intuitive and flexible way to convey intent [30, 61, 68]. This amplifies interest in the research topic of temporal sentence grounding [22, 25-27, 32, 40, 46, 47, 57, 71, 81], which aims to locate one or more semantically aligned segments within an untrimmed video according to a given natural language sentence, as shown in Figure 1a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Early methods either align sentences with predefined temporal proposals for selection [44, 70] or directly predict moment spans through cross-modal interactions between" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.269, + 0.882, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 1. Illustration of (a) temporal sentence grounding task, (b) phenomenon of random matching across multiple layers, and (c) the challenge of achieving the balance between focus on boundary for span prediction and alignment with global semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.749 + ], + "angle": 0, + "content": "language and frames [36, 37, 72]. Recent advances center on integrating the Detection Transformer (DETR) [2] for object detection into the temporal sentence grounding, leveraging its query-based proposal detection framework to eliminate hand-crafted proposal components and deliver superior grounding performance with high efficiency [22, 25, 26, 46, 47, 57]. They leverage moment queries—often several times more numerous than ground truth segments—in the multi-layer decoder, which are initialized randomly [26, 46, 57], generated from sentences [22, 47], or extracted as event units from the video [25], to locate and search the most suitable matches for target segments. A one-to-one label assignment is applied in the training objective to facilitate non-redundant predictions for each ground-truth segment. The overall architecture of the DETR-based temporal sentence grounding framework is shown in Figure 6a." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, we observe that strategies generally effective for DETR in other tasks, such as object detection, like increasing the number of queries or decoder layers, do not improve performance in temporal sentence grounding and may even degrade it (Sec 3.1). This motivates us to revisit the specific characteristics of temporal sentence grounding that give rise to the distinct challenges in applying DETR to this task. In a series of investigations on query-segment alignment (Sec 3.2), we verify that performance degradation primarily stems from the difficulty in identifying dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.887, + 0.431, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author is Sibei Yang." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "tinct local boundaries for target segments that share highly similar global semantics, leading to inherent conflicts both between and within queries. By diagnosing the similarities between queries and between queries and frames, we uncover two conflicts that result in learning puzzles for queries:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.186, + 0.483, + 0.366 + ], + "angle": 0, + "content": "- Queries corresponding to distinct target segments exhibit high similarity, leading to random matching with various targets and creating learning puzzles for the queries. The presence of multiple semantically similar target segments results in the high similarity between their corresponding queries, making them difficult to clearly distinguish (Sec 3.3). This causes each query to potentially match well with multiple segments, leading to significant randomness in the one-to-one matching for each target segment (Figure 1b). We observe that across different decoder layers, the queries matched for optimization can vary considerably (Sec 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.367, + 0.483, + 0.533 + ], + "angle": 0, + "content": "- Each query faces an inherent conflict between encoding the global semantics of its corresponding segment and decoding its local boundary. To align well with a segment, a query must encode the segment's global semantics from the start to the end frame, such as the entire process of a person holding a cup. However, the query also needs to directly predict the segment's boundary, requiring focus on the local frames near the boundary (Figure 1c). This creates a trade-off: the query either prioritizes global semantics or local boundaries, making it challenging to optimize both simultaneously (Sec 3.3)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.186, + 0.483, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Building on these findings, we propose a simple yet strong baseline for video temporal grounding, Sim-DETR. The framework builds on the standard DETR-based temporal sentence grounding architecture, with two minor yet straightforward design modifications to the decoder layers to address the two conflicts between and within queries (Figure 6), respectively. First, we adjust the self-attention weights between queries based on their pairwise correlation, encouraging indistinguishable queries to focus on different contexts, reducing their similarity, and allowing the most suitable query to gather information from related queries to refine its prediction (Figure 6b). Second, we introduce a query-to-frame matching and loss term, accounting for the alignment between the query and each frame within the segment, as shown in Figure 6c. This ensures that all frames, not just those near the boundary, contribute to segment localization. The full sequence of all target frames, from start to finish, acts as a bridge connecting the global semantics to the local boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Empirically, we indeed observe that our minor modifications (1) significantly reduce conflicts between queries corresponding to different target segments (Figure 3), (2) improve the alignment between the query's global semantic attention to frames and its boundary prediction (Figure 5)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "and (3) lead to more consistent query predictions across layers (Figure 4). Furthermore, (4) increasing the number of queries and decoder layers no longer causes performance degradation (Figure 2). Finally, due to the consistent matching and prediction of queries, (5) our Sim-DETR not only achieves substantial performance gains but also accelerates convergence (as shown in Appendix). All the above observations and advantages, including the minor modifications, are intended to unlock the potential of query-based frameworks for temporal sentence grounding and provide a simple yet strong baseline for future research." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.258, + 0.816, + 0.271 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We systematically analyze the root causes of abnormal behavior in the DETR-based temporal sentence grounding framework, identifying two key conflicts: (1) target segments with highly similar semantics but distinct temporal localization create learning puzzles between queries, and (2) a trade-off between relying on global semantics for matching and local boundaries for localization within queries." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.394, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- Based on our analysis, we propose two modifications to form Sim-DETR: (1) a simple adjustment to the self-attention to resolve conflicts between queries and (2) the introduction of a query-to-frame matching to reconcile global semantics and local localization within queries." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.47, + 0.905, + 0.53 + ], + "angle": 0, + "content": "- Our Sim-DETR, a simple yet powerful baseline, achieves consistent and significant improvements across all benchmarks. More importantly, it eliminates observed anomalies and exhibits faster convergence." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.543, + 0.656, + 0.559 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.856 + ], + "angle": 0, + "content": "Temporal Sentence Grounding (TSG) aims to identify specific video moments that align with given language expressions. As computer vision advances across neural architectures, generation, detection, segmentation, and multimodal tasks [4, 8, 21, 33, 53-55, 73, 82], temporal understanding requires task-specific optimizations. Early methods fall into proposal-free and proposal-based categories. Proposal-free methods use end-to-end frameworks by directly regressing boundary coordinates [6, 43, 76] or predicting frame-level boundary likelihood [16, 78]. Proposal-based approaches densely sample candidate segments via sliding windows [1, 12, 15, 23, 38, 80], temporal anchors [5, 32, 41, 75, 77, 81], or multimodal feature similarity [7, 34, 35, 66, 69]. Recently, DETR has shown exceptional effectiveness in detection tasks. [27, 65] pioneered DETR for TSG. Subsequent studies enhance cross-modal representations [46, 47, 57, 67], improve decoder queries with semantic/spatial information [22, 25, 58], or explore joint training with other temporal tasks [57, 71]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Detection Transformers (DETR) [2] is a pioneering end-to-end framework for object detection and segmentation that employs transformers to make direct set-based pre" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22761" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.381 + ], + "angle": 0, + "content": "dictions [14, 52, 54, 55, 60, 61, 85]. However, DETR has notable limitations, especially its high computational complexity and slow convergence, impacting its practical efficiency. To address these issues, several derivative models have emerged. Some methods [51, 63, 74, 84] improve computational efficiency by reducing redundant calculations within the DETR architecture. Other models [9, 13, 59] refine the attention mechanism for more efficient information processing and resource utilization. Additionally [39, 45, 64] enhance convergence by embedding spatial information directly into the query design, which accelerates object localization. Recent models [28, 79] tackle optimization challenges in DETR's bipartite matching process by integrating denoising strategies during training, improving both convergence and performance. Despite recent advancements in enhancing DETR for object detection, there is a lack of systematic studies addressing the unique challenges of applying DETR to TSG and the potential difficulties in query optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.392, + 0.483, + 0.409 + ], + "angle": 0, + "content": "3. Probing DETR's Inherent Behavior for TSG" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.598 + ], + "angle": 0, + "content": "This study begins with the observation that enhancements effective for object detection in DETR [2] do not apply to temporal sentence grounding (TSG), as discussed in Sec 3.1. To explore the reasons behind DETR's failure to enhance, this section presents preliminary studies to reveal the behavior of queries in the DETR decoder, as queries are central to predicting proposals. Specifically, we examine the similarity between queries and outputs in Sec 3.2, and the attention between queries and inputs (i.e., frames) in Sec 3.3. These help reveal the underlying cause, and our modification addressing it not only eliminates the abnormal phenomena but also has a positive side effect (Sec 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Experimental Settings. Unless otherwise specified, our default experimental setup involves conducting statistical analyses on the full validation set of QVHighlights [27]. CG-DETR [46] and TaskWeave [71], which serve as baselines for DETR-based TSG due to their strong performance while retaining the core vanilla DETR mechanism, are used with their standard configurations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.303, + 0.728 + ], + "angle": 0, + "content": "3.1. Abnormal Phenomena" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Motivation. DETR is originally proposed for object detection, and increasing the number of queries and decoder layers typically improves performance. We explore whether this trend also applies to TSG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Setting. Both CG-DETR and TaskWeave default to 10 queries, with 3 and 2 decoder layers, respectively. This experiment specifically increases the number of queries and decoder layers to isolate their effects, keeping all other hyperparameters and training protocols constant." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Result & Discussion. As shown in Figure 2, increasing the number of queries to 15 and 20 led to progressively" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.085, + 0.897, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.223, + 0.905, + 0.238 + ], + "angle": 0, + "content": "Figure 2. The impact of the number of queries and decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.907, + 0.36 + ], + "angle": 0, + "content": "larger performance declines, with a drop of over \\(-2\\%\\) with TaskWeave [71] at 20 queries. Similarly, increasing the number of decoder layers caused a drop from \\(-1\\%\\) to \\(-2.5\\%\\). In contrast, our Sim-DETR mitigates this phenomenon and achieves a slight improvement, demonstrating its robustness. This observation motivates an in-depth exploration of decoder layer, particularly the role of queries (see Sec 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.368, + 0.906, + 0.385 + ], + "angle": 0, + "content": "3.2. Similar Target Segments Cause Query Conflicts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.542 + ], + "angle": 0, + "content": "Motivation. Initially, we hypothesize that the lack of performance improvement with increased queries is due to query redundancy. However, performance declined, suggesting that conflicts between queries, rather than redundancy, prevent compatibility with existing queries. Similarly, conflicts between queries across layers may explain the ineffectiveness of increasing the number of layers. This motivates us to explore (1) the relationships between queries to identify their potential conflicts and (2) their variations across layers to reveal the impact at different layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.827 + ], + "angle": 0, + "content": "Setting. (1) To evaluate query relationships, we assess similarities between queries corresponding to the same segment (intra-segment similarity) and those from different segments (inter-segment similarity). Specifically, we identify the corresponding segment for each query based on the maximum IoU between the query's prediction and all ground-truth (GT) segments during inference. Queries whose IoU with the corresponding segment is 0.5 or greater are included in our analysis to ensure meaningful corresponding. We then compute feature similarities between queries associated with the same segment (intra) and those matched to different segments (inter). (2) To analyze variations across layers, we introduce the metric of cross-layer matching consistency. Specifically, during training optimization, each GT segment is uniquely matched to a query through bipartite matching at each decoding layer. We assess match consistency by measuring the proportion of segments that retain their corresponding queries across two consecutive decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Result & Discussion. (1) As shown in Figure 3, CGDETR fails to distinguish query similarities between inter- or intra-segment groups, indicating that queries are not clearly grouped with their corresponding segments. Consequently, the same query oscillates between segments," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22762" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.077, + 0.473, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.217, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 3. Distribution of query similarity for intra-segment (blue) and inter-segment (red) pairs. Our method can effectively distinguish intra-segment and inter-segment queries, ensuring more stable query-segment associations and reducing conflicts in query assignments across different segments." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.288, + 0.462, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.487 + ], + "angle": 0, + "content": "Figure 4. Measuring cross-layer matching consistency (y-axis, defined in Line-260) of segments across two consecutive decoding layers (x-axis). Our method achieves higher consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.641 + ], + "angle": 0, + "content": "lacking stable and consistent corresponding. Figure 4 further supports our hypothesis, demonstrating that cross-layer matching consistency in CG-DETR is lower than ours. (2) We attribute this to multiple GT segments in a video sharing the same linguistic semantics. The high semantic similarity between GT segments results in high similarity between their corresponding queries. Figure 4 supports this, showing that the consistency for cases with multiple GT segments is lower than for those with a single segment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.433, + 0.667 + ], + "angle": 0, + "content": "3.3. Global Matching vs. Local Localization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Motivation. We further question whether the inconsistency in matching is solely due to query conflicts from multiple GT segments. If so, why does the previous method still show roughly \\(25\\%\\) mismatched queries in cases with a single target segment? Therefore, we shift our focus from conflicts between queries to conflicts within individual queries. A query serves dual roles: aligning with global linguistic semantics (global matching) and precisely localizing the segment, particularly its boundaries (local localization). This prompts us to explore whether conflicts arise from these roles within a single query." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Setting. (1) To evaluate global matching, we use the query's classification confidence score, as it reflects the confidence of the query being referenced by the global semantics of the linguistic description. (2) To evaluate local localization, we" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.335 + ], + "angle": 0, + "content": "Figure 5. Global matching score (defined in Line-290) vs. local localization score (defined in Line-293). Compared to CGDETR, our method concentrates a query's attention significantly more within a single GT segment, rather than dispersing it across multiple segments, leading to improved local localization scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.433 + ], + "angle": 0, + "content": "extract attention scores between each query and frame from the cross-attention module in the last decoder layer, as they indicate query-based frame localization. We then compute the IoU score between this attention and the query's corresponding GT segment, reflecting the accuracy of the query's localization at the frame level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Result & Discussion. As shown in Figure 5, CG-DETR (marked in red) may yield low local localization scores despite high global matching scores, indicating that while queries align well with linguistic descriptions, they fail to achieve precise localization. Further analysis reveals that query attention may span frames in multiple GT segments (see Figure 5 \"Attention\"), leading to a low local score for the query's specific GT segment it corresponds to." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.698, + 0.58 + ], + "angle": 0, + "content": "3.4. Positive Side Effect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.679 + ], + "angle": 0, + "content": "Convert Abnormal to Normal: Our Sim-DETR effectively (1) distinguishes queries within the same target segment and across different segments in Figure 3, (2) achieves consistent matching across decoder layers in Figure 4, and (3) aligns with global semantics while ensuring precise localization in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.906, + 0.77 + ], + "angle": 0, + "content": "Positive Side Effect: (1) More importantly, Sim-DETR maintains stable and robust performance with increased queries and decoder layers (see Figure 2). (2) By eliminating the abnormal phenomena, it also achieves significantly faster convergence compared to previous methods, as shown in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.605, + 0.8 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We present a simple yet effective TSG baseline that resolves query conflicts in existing DETR-based TSG frameworks [22, 25, 27, 46, 47, 57, 67, 71] through two minor yet key modifications. First, we introduce the standard DETR-based [2] TSG architecture (Sec 4.1). Sec 4.2 introduces our first modification, which adjusts the self-attention" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.907, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 6. Framework of the proposed Sim-DETR, featuring two simple yet powerful modifications: Query Grouping and Ranking to mitigate cross-query conflicts and Global-Local Bridging to address global-local conflicts within each query." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.519 + ], + "angle": 0, + "content": "mechanism between queries in the decoder to avoid conflicts among them, based on the observations presented in Sec 3.2. Next, Sec 4.3 presents another improvement that addresses the conflict between global matching and local localization within the query (discussed in Sec 3.3) by introducing query-to-frame matching loss, which serves as the bridge between the global and local perspectives." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.53, + 0.446, + 0.545 + ], + "angle": 0, + "content": "4.1. Standard DETR-based TSG Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.582 + ], + "angle": 0, + "content": "In TSG, the goal is to identify video segments with temporal spans that correspond to a given language expression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Feature Extraction. Previous DETR-based works [25, 47, 57, 71] employ CLIP [49] and SlowFast [10] to extract video features, represented as \\(\\mathcal{T} \\in \\mathbb{R}^{N \\times C}\\), where \\(N\\) and \\(C\\) stands for the number of frames and the hidden dimension, respectively. For the language expression, typically a sentence with \\(L\\) tokens, the CLIP text encoder is used to obtain token-level features, resulting in \\(\\mathcal{W} \\in \\mathbb{R}^{L \\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Multimodal Encoder. After feature extraction, DETR-based TSG methods [22, 46, 47, 57] employ a multimodal encoder to fuse these video and language features, producing multimodal video features, denoted as \\(\\hat{\\mathcal{T}}\\in \\mathbb{R}^{N\\times C}\\). The key component of the multimodal encoder is a cross-attention mechanism [62], where video features \\(\\mathcal{T}\\) act as queries, while word features \\(\\mathcal{W}\\) serve as keys and values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Temporal Sentence Decoding. In the decoder, the model initializes a set of query to gather global information from the video, which is then decoded into local spans in the span prediction head. Specifically, given initialized queries \\(\\mathcal{Q} \\in \\mathbb{R}^{M \\times C}\\), where \\(M\\) is the query number, and the multimodal video features \\(\\hat{\\mathcal{T}}\\), the decoder facilitates interaction between queries and gathers global information through interleaved" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.413, + 0.871, + 0.427 + ], + "angle": 0, + "content": "self-attention and cross-attention modules, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.44, + 0.905, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {k} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {k} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {v} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {v} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right). \\end{array} \\right. \\\\ \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {q} ^ {\\mathrm {C A}} (\\mathcal {Q} ^ {\\mathrm {S A}}), \\hat {\\mathcal {T}} _ {k} = \\mathrm {F C} _ {k} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}}), \\hat {\\mathcal {T}} _ {v} = \\mathrm {F C} _ {v} (\\hat {\\mathcal {T}}), \\\\ \\mathcal {Q} _ {o} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {C A}} \\left(\\mathrm {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}} _ {k}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\hat {\\mathcal {T}} _ {v}\\right). \\end{array} \\right. \\\\ \\hat {\\mathcal {Q}} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathcal {Q} ^ {\\mathrm {C A}}\\right) + \\mathcal {Q}\\right), \\\\ B = \\mathcal {H} _ {\\text {s p a n}} (\\hat {\\mathcal {Q}}), \\quad B = \\left\\{b _ {1}, \\dots , b _ {M} \\right\\}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.716 + ], + "angle": 0, + "content": "where \\(\\mathrm{FC}(\\cdot)\\) denotes the fully connected projection layer, with SA and CA representing self-attention and cross-attention, respectively. Subscripts \\(q\\), \\(k\\), and \\(v\\) indicate query, key, and value computations. MLP(\\(\\cdot\\)) represents a nonlinear mapping with an activation function. \\(\\mathcal{H}_{\\mathrm{span}}(\\cdot)\\) denotes the span prediction head, which takes the updated query features \\(\\hat{\\mathcal{Q}}\\) to predict the span set \\(B\\), where each \\(b_i \\in \\mathbb{R}^2\\) represents the (start, end) of a predicted span." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.783, + 0.744 + ], + "angle": 0, + "content": "4.2. Query Grouping and Ranking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In this section, we address the query conflicts identified in Sec 3.2: (1) the inability to distinguish between intra- and inter-segment queries (Figure 3), leading to query oscillation between segments without stable correspondence (Line-266), and (2) inconsistencies in segment-query matching across layers, where different queries may be matched to ground-truth (GT) segments at each layer for optimization, hindering effective learning (Figure 4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To differentiate queries across segments, the challenge arises from GT segments that inherently share identical lin" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.096, + 0.461, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.204, + 0.457, + 0.219 + ], + "angle": 0, + "content": "Figure 7. Overview of our Query Grouping and Ranking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.372 + ], + "angle": 0, + "content": "guistic semantics, leading to highly similar queries. Therefore, a natural solution is to directly group queries, assigning each group to a specific GT segment for training and prediction. Instead of hard grouping with predefined rules or extra hyperparameters, we learn a soft grouping based on predicted temporal spans of queries, where closer spans indicate potential corresponding for the same segment. Specifically, for any two queries \\( q_{i} \\) and \\( q_{j} \\), we compute their span distance, which is then used in Eq. 4 to learn their grouping:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.365, + 0.482, + 0.384 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i, j} ^ {\\text {i n t r a}} = \\| b _ {i} - b _ {j} \\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.464 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2 \\) denotes the L2 norm. Notably, we use L2 distance instead of the commonly used L1 because, when two spans are close (normalized distances \\( \\leq 1 \\)), L2 distance decreases more sharply than L1, imposing a smaller penalty on minor differences (see Appendix G)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Meanwhile, to enhance cross-layer matching consistency, we draw inspiration from EASE-DETR [14]'s query ranking strategy in object detection. By ranking queries within a segment at each layer, the leading query is prioritized for segment matching in subsequent layers. However, directly ranking queries by prediction scores, as in EASE-DETR, proves ineffective (see Appendix D). As analyzed in Sec. 3.3, high prediction/confidence scores do not guarantee precise predictions and may correspond to poor localization. Instead, we introduce an IoU head to assess localization accuracy and rank queries based on both factors, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.64, + 0.482, + 0.679 + ], + "angle": 0, + "content": "\\[\nR _ {\\text {r a n k}} \\left(q _ {i}, q _ {j}\\right) = \\left\\{ \\begin{array}{l l} + 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} \\geq \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}} \\\\ - 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} < \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}}, \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.76 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^{\\mathrm{cls}}\\) and \\(\\mathcal{P}^{\\mathrm{IoU}}\\) denote the query's confidence score and IoU prediction, respectively, with subscripts \\(i\\) and \\(j\\) indicating queries \\(q_{i}\\) and \\(q_{j}\\). \\(R_{\\mathrm{rank}}(q_i,q_j) = 1\\) indicates that query \\(q_{i}\\) has a higher priority than query \\(q_{j}\\), considering both global alignment and local localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Finally, we incorporate query grouping within segments and intra-segment ranking to infer relative relationships, then reshape query-to-query self-attention following [14] to mitigate conflicts, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.827, + 0.482, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {S} ^ {\\text {a t t n}} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathcal {S} ^ {\\text {i n t r a}} \\circ \\mathcal {R} _ {\\text {r a n k}}\\right), \\right. \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\text {s o f t m a x} \\left(\\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} \\left(\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}\\right) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {S} ^ {\\mathrm {a t t n}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "where \\( S^{\\mathrm{attn}} \\) represents query relationships, with higher values promoting high-quality queries to dominate predictions by integrating similar queries within the same group." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.725, + 0.166 + ], + "angle": 0, + "content": "4.3. Global-Local Bridging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.907, + 0.352 + ], + "angle": 0, + "content": "In this section, we address the query conflicts identified in Sec 3.3, specifically the conflict between a query's global semantic matching and local localization. Since queries overly rely on global semantics while neglecting local localization, our core idea is to reinforce fine-grained frame-level localization within the segment. Specifically, we introduce a query-to-frame matching mechanism and loss term to enhance localization beyond segment boundary prediction, ensuring alignment with in-segment frames while avoiding matches with those from other segments. The full sequence of frames within the segment, from start to finish, serves as a bridge linking global semantics to local boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.442 + ], + "angle": 0, + "content": "During training, we compute the semantic similarity between each query \\( q_{i} \\) and all frames within its corresponding ground-truth span, maximizing similarity to ensure full alignment with in-segment frames while minimizing similarity to out-segment frames. The computation process is outlined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.446, + 0.907, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z = \\operatorname {s i g m o i d} (\\tau \\cdot \\cos (q _ {i}, \\hat {\\mathcal {T}})), \\\\ \\mathcal {L} _ {\\mathrm {b r i d g e}} = \\lambda_ {\\mathrm {b r i d g e}} \\frac {- \\sum_ {j} z _ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}{\\sum_ {j} z _ {j} \\left(1 - \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}\\right) + \\sum_ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}, \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.761 + ], + "angle": 0, + "content": "where \\(z \\in \\mathbb{R}^N\\) donates query-to-frame similarity, calculated using the cosine similarity between the query \\(q_i\\) and the \\(N\\) frame features \\(\\tilde{T}\\) defined in Line-347, and \\(\\tau\\) is a learnable scaling coefficient. \\(b_i^{\\mathrm{gt}}\\) is the ground-truth span corresponding to query \\(q_i\\), while \\(\\mathbb{I}[b_i^{\\mathrm{gt}}] \\in \\mathbb{R}^N\\) is an indicator function that assigns 1 to frames within the ground truth and 0 otherwise, with \\(j\\) representing the frame index. Thus, the numerator in \\(\\mathcal{L}_{\\mathrm{bridge}}\\), i.e., \\(-\\sum_{j} z_j \\mathbb{I}[b_i^{\\mathrm{gt}}]_j\\), encourages the query to be similar to every frame within its corresponding segment. Meanwhile, the denominator's first term, \\(\\sum_{j} z_j (1 - \\mathbb{I}[b_i^{\\mathrm{gt}}]_j)\\), represents query \\(q_i\\)'s similarity to out-segment frames, promoting its minimization. The denominator's second term, \\(\\sum_{j} \\mathbb{I}[b_i^{\\mathrm{gt}}]_j\\), represents the length of the ground-truth segment, serving as a normalization factor. And \\(\\lambda_{\\mathrm{bridge}}\\) is a hyperparameter that controls the loss weight." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.649, + 0.787 + ], + "angle": 0, + "content": "4.4. Overall Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The overall loss is: \\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{MD}} + \\lambda_{\\mathrm{bridge}}\\mathcal{L}_{\\mathrm{bridge}} + \\lambda_{\\mathrm{iou}}\\mathcal{L}_{\\mathrm{iou}}\\) where \\(\\mathcal{L}_{\\mathrm{MD}}\\) includes the L1, gIoU, classification, and saliency losses, as in Moment DETR [26]. \\(\\mathcal{L}_{\\mathrm{iou}}\\) is the loss for the IoU head introduced in Line-406, designed to assist in ranking queries using \\(\\mathcal{P}_i^{\\mathrm{IoU}}\\) and \\(\\mathcal{P}_j^{\\mathrm{IoU}}\\) in Eq 3. The IoU head implementation follows [25, 58]. Ablations for hyperparameters \\(\\lambda_{\\mathrm{bridge}}\\) and \\(\\lambda_{\\mathrm{iou}}\\) are detailed in Appendix E." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22765" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.872, + 0.301 + ], + "angle": 0, + "content": "
Methodtestval
R1mAPR1mAP
@0.5@0.7@0.5@0.75Avg.@0.5@0.7@0.5@0.75Avg.
M-DETR [27]NeurIPS'2152.8933.0254.8229.4030.7353.9434.84--32.20
UMT [40]CVPR'2256.2341.1853.8337.0136.1260.2644.26--38.59
QD-DETR [47]CVPR'2362.4044.9862.5239.8839.8662.6846.6662.2341.8241.22
UniVTG [32]ICCV'2358.8640.8657.6035.5935.4759.74---36.13
EaTR [22]ICCV'23-----61.3645.7961.8641.9141.74
MomentDiff [29]NeurIPS'2357.4239.6654.0235.7335.95-----
TR-DETR [57]AAAI'2464.6648.9663.9843.7342.6267.1051.4866.2746.4245.09
UVCOM [67]CVPR'2463.5547.4763.3742.6743.1865.1051.81--45.79
TaskWeave [71]CVPR'24-----64.2650.0665.3946.4745.38
BAM-DETR [25]ECCV'2462.7148.6464.5746.3345.3665.1051.6165.4148.5647.61
CG-DETR [46]Arxiv'2465.4348.3864.5142.7742.8667.3552.0665.5745.7344.93
SpikeMba [31]Arxiv'2464.1349.42-43.6743.7965.3251.33-44.9644.84
Sim-DETR(Ours)67.6450.9167.8147.5946.9369.4854.0669.7051.1149.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.173, + 0.305, + 0.822, + 0.32 + ], + "angle": 0, + "content": "Table 1. Experimental results on the QVHighlights benchmark, comparing ours with state-of-the-art methods." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.328, + 0.861, + 0.478 + ], + "angle": 0, + "content": "
MethodTACoSCharades-STA
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [81]40.0127.9912.9227.2258.7646.0227.5041.25
M-DETR [27]37.9724.6711.9725.4965.8352.0730.5945.54
MomentDiff [29]44.7833.68---55.5732.42-
QD-DETR [47]-----57.3132.55-
UniVTG [32]51.4434.9717.3533.6070.8158.0135.6550.10
CG-DETR [46]52.2339.6122.2336.4870.4358.4436.3450.13
UVCOM [67]-36.3923.32--59.2536.64-
SpikeMba [31]51.9839.3422.8335.8171.2459.6536.1251.74
Sim-DETR(Ours)57.0642.7926.8239.4473.0961.3439.6252.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.273, + 0.482, + 0.722, + 0.495 + ], + "angle": 0, + "content": "Table 2. Comparison results on the TACoS and Charades-STA benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.225, + 0.525 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.411, + 0.553 + ], + "angle": 0, + "content": "5.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.775 + ], + "angle": 0, + "content": "Datasets. Following [27, 31, 32, 47, 67], we conduct comprehensive evaluation experiments across three datasets: QVHighlights [27], Charades-STA [12], and TACoS [50]. The QVHighlights dataset contains 10,310 text expressions and 18,367 distinct events, with most expressions corresponding to multiple events. Charades-STA is derived from a collection of 9,848 indoor scene videos, encompassing 16,128 events. TACoS includes only 273 videos, with an average length approaching five minutes per video. For a fair comparison, we adopt the same evaluation metrics as in previous methods. On the QVHighlights benchmark, we use Recall and Mean Average Precision (mAP) as primary metrics, while for Charades-STA and TACoS, we employ Recall and Mean Intersection over Union (mIoU)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We set the input video resolution to \\(224 \\times 224\\) and represent each video by concatenating the CLIP [49] [CLS] tokens with SlowFast [10] features. For the Charades-STA and TACoS datasets, we additionally constructed two experiments using I3D [3] and VGG [56] as visual encoders, with Glove [48] embeddings for text features. The model is trained for 200 epochs on a single NVIDIA A40 GPU, using the AdamW [42] optimizer with" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.712 + ], + "angle": 0, + "content": "
BackboneMethodR@0.5R@0.7
VGG [56]2D-TAN [81]40.9422.85
FVMR [11]42.3624.14
SSRN [83]46.7227.98
UMT [40]48.3129.25
MomentDiff [29]51.9428.25
QD-DETR [47]52.7731.13
TR-DETR [57]53.4730.81
CG-DETR [46]55.2234.19
I3D [3]Sim-DETR(Ours)55.9735.38
MAN [77]46.5322.72
VSLNet [78]47.3130.19
QD-DETR [47]50.6731.02
TR-DETR [57]55.5133.66
TaskWeave [71]53.3631.40
Sim-DETR(Ours)57.5535.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.527, + 0.716, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Table 3. Experimental results with VGG or I3D as backbone." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.907, + 0.78 + ], + "angle": 0, + "content": "an initial learning rate of 1e-4. Note that, for a fair comparison, the total number of layers in our Sim-DETR is set to 6, the same as recent works [22, 46, 67]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.882, + 0.804 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.87 + ], + "angle": 0, + "content": "As shown in Tables 1-3, we report detailed evaluation results across the QVHighlights, Charades-STA, and TACoS benchmarks. Our Sim-DETR consistently outperforms all state-of-the-art methods (SOTAs) across all metrics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Results on QVHighlights are shown in Table 1. Compared to the latest published work, BAM-DETR [25], our Sim" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22766" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.455, + 0.189 + ], + "angle": 0, + "content": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline65.4850.8465.3445.8244.97
+ Liou66.5851.9465.6846.3045.22
+ QGR68.7752.2667.7248.2847.03
+ GLB67.1652.7768.5848.8848.17
+ QGR,GLB69.4854.0669.7051.1149.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.194, + 0.395, + 0.208 + ], + "angle": 0, + "content": "Table 4. Ablation study on components." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.214, + 0.456, + 0.33 + ], + "angle": 0, + "content": "
Inner RelevanceOuter GlobalOuter LocalmAP
(1)span border distconfidence predIoU pred49.50
(2)center distconfidence predIoU pred48.59
(3)span IoU48.94
(4)span border distw/oIoU pred48.93
(5)span border distconfidence predw/o48.66
(6)center pred48.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.334, + 0.389, + 0.348 + ], + "angle": 0, + "content": "Table 5. Impact of different measures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.597 + ], + "angle": 0, + "content": "DETR outperforms it by an average of \\(+2.88\\%\\) across all metrics. More importantly, it achieves remarkable gains of \\(+4.93\\%\\) and \\(+2.27\\%\\) in R1@0.5 and R1@0.7, respectively. In comparison TR-DETR [57], which uses standard DETR [2] as its basic framework and is used as our baseline model, our Sim-DETR achieves an average improvement of \\(+4.31\\%\\) and \\(+4.41\\%\\) in mAP for test and validation, respectively. These improvements underscore the effectiveness of our Sim-DETR in fundamentally addressing conflicts within and between queries. We further conduct a comparison with SpikeMba [31], a contemporary method built on the Mamba [17] framework. Without any complex module designs, the consistent improvements across all metrics highlight the simplicity and effectiveness of our Sim-DETR. Sim-DETR establishes an efficient baseline for the task, significantly advancing research in TSG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Results on Charades-STA and TACoS are presented in Table 2, where we consistently outperform all methods on both datasets. Compared to the state-of-the-art methods, we achieve a \\(+1.97\\%\\) improvement on the Charades-STA dataset (with SpikeMba as the SOTA) and a \\(+3.89\\%\\) improvement on the TACoS dataset (with CG-DETR [46] as the SOTA). These enhancements further demonstrate the outstanding capability of our method in temporal localization. More importantly, they confirm that our Sim-DETR exhibits significant performance advantages across various scenarios, types, and lengths of videos, highlighting its robustness and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results with Different Backbones. Following previous works, we conduct two additional experiments on the QVHighlights dataset. These experiments use VGG and I3D for visual feature extraction and use Glove [48]embedding as text features. Table 3 presents the detailed results of these two experiments, where we achieve state-of-the-art results in both cases. Notably, for the video feature extractor I3D, we surpass the latest pub" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "lished method by \\(+4.26\\%\\). These results further demonstrate the excellent generalization capabilities of our SimDETR across different backbones." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.146, + 0.667, + 0.163 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.485 + ], + "angle": 0, + "content": "Roadmap for building a simple yet effective TSG baseline is presented in Table 4. We employ TR-DETR without MR2HD module as our baseline, which achieves an mAP of \\(44.97\\%\\). Since our Query Grouping and Ranking (QGR) utilizes \\(\\mathcal{L}_{\\mathrm{iou}}\\) proposed by [25, 58], we additionally conduct an ablation study to isolate the impact of \\(\\mathcal{L}_{\\mathrm{iou}}\\). (1) QGR: To mitigate conflicts and optimization challenges arising from overly similar semantics between queries, we introduce the QGR, which improves the baseline by \\(+2.07\\%\\) on mAP, representing a subtle yet significant enhancement. (2) Global-Local Bridging (GLB): To address conflicts and ambiguities between global semantics and local localization within queries, we introduce GLB to enable seamless transformation through query-to-frame alignment. Experimental results show a performance boost to \\(48.17\\%\\) on mAP, indicating the effectiveness of GLB in bridging global-to-local transformations. However, without QGR to address semantic conflicts in temporal DETR, the recall predictably decreased by \\(-3.1\\%\\). (3) Overall Framework: With these adjustments, our Sim-DETR achieves a remarkable SOTA performance: \\(49.50\\%\\) on mAP and \\(69.48\\%\\) on R1@0.5." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.485, + 0.905, + 0.772 + ], + "angle": 0, + "content": "The Effects of Different Conflict Metrics in QGR. Table 5 summarizes five experiments on conflict metrics in QGR. (1) The first row displays our final implementation, achieving \\(49.50\\%\\) mAP. (2) & (3) We evaluate L2 center distance and span IoU for inter-query relationships. Center distance alone yields \\(48.59\\%\\) mAP, highlighting its inadequacy in segment representation. Span IoU outperforms center distance but lags by \\(0.56\\%\\) compared to boundary distance, likely due to IoU's broader tolerance, which reduces precision in query relevance. (4) To assess outer global semantic alignment, we remove global semantics, resulting in a \\(-0.57\\%\\) drop, emphasizing their role in query interaction and quality. (5) Finally, we test local location metrics in query prioritization. Excluding the outer local metric significantly lowers performance \\((-0.84\\%\\) mAP), indicating that without local information, queries encounter conflicts due to overly generalized global semantics, leading to random matches. Consistent with (2), center localization accuracy shows similar limitations, with a \\(0.53\\%\\) decrease." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.633, + 0.801 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this paper, we propose a concise and efficient temporal sentence grounding framework Sim-DETR, introducing two simple modifications to the decoder for resolving conflicts both between and within queries. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods across all benchmarks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22767" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.121 + ], + "angle": 0, + "content": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No.62206174)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.15, + 0.188, + 0.166 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.175, + 0.486, + 0.243 + ], + "angle": 0, + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[4] Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[5] Jingyuan Chen, Xinpeng Chen, Lin Ma, Zequn Jie, and Tat-Seng Chua. Temporally grounding natural sentence in video. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 162-171, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[6] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10551-10558, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[8] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711-13722, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[9] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[10] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.175, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. arXiv preprint arXiv:2101.07448, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[15] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 245-253. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.304, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[19] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7366–7375, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[20] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. Advances in Neural Information Processing Systems, 36:26135-26158, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[21] Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. Mvtokenflow: High-quality 4d content generation using multiview token flow. arXiv preprint arXiv:2502.11697, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[22] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[23] Bin Jiang, Xin Huang, Chao Yang, and Junsong Yuan. Cross-modal video moment retrieval with spatial and language-temporal attention. In Proceedings of the 2019 on international conference on multimedia retrieval, pages 217-225, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[24] Manjin Kim, Heeseung Kwon, Chunyu Wang, Suha Kwak, and Minsu Cho. Relational self-attention: What's missing in attention for video understanding. Advances in Neural Information Processing Systems, 34:8046–8059, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Neurips, pages 11846-11858, 2021. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[29] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In Neurips, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[30] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[31] Wenrui Li, Xiaopeng Hong, Ruiqin Xiong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[32] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[33] Liang Lin, Pengxiang Yan, Xiaoqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[34] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4536-4545, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[35] Daizong Liu, Xiaoye Qu, Jianfeng Dong, and Pan Zhou. Adaptive proposal generation network for temporal sentence localization in videos. arXiv preprint arXiv:2109.06398, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[36] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1665-1673, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[37] Daizong Liu, Xiaoye Qu, and Wei Hu. Reducing the vision and language bias for temporal sentence grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4092-4101, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[38] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[41] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. \\(r^2\\)-tuning: Efficient image-to-video transfer learning for video temporal grounding. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[42] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[44] Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[45] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[46] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[47] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[48] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[50] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[51] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22769" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[53] Cheng Shi and Sibei Yang. Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932-2941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[54] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.28 + ], + "angle": 0, + "content": "[55] Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In European Conference on Computer Vision, pages 210-226. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[57] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4998-5007, 2024. 1, 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[58] Xiaolong Sun, Liushuai Shi, Le Wang, Sanping Zhou, Kun Xia, Yabing Wang, and Gang Hua. Diversifying query: Region-guided transformer for temporal sentence grounding. arXiv preprint arXiv:2406.00143, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[59] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[60] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23570-23580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.64 + ], + "angle": 0, + "content": "[61] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[62] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[63] Tao Wang, Li Yuan, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4661-4670, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[64] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[65] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore-and-match: Bridging proposal-based and proposal-free with transformer for sentence grounding in videos. arXiv preprint arXiv:2201.10168, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2986-2994, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[67] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[68] Chen-Wei Xie, Siyang Sun, Xiong Xiong, Yun Zheng, Deli Zhao, and Jingren Zhou. Ra-clip: Retrieval augmented contrastive language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19265–19274, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.907, + 0.359 + ], + "angle": 0, + "content": "[69] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9062–9069, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[70] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Unsupervised alignment-based iterative evidence retrieval for multihop question answering. arXiv preprint arXiv:2005.01218, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[71] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18308-18318, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[72] Shuo Yang and Xinxiao Wu. Entity-aware and motion-aware transformers for language-driven action localization in videos. arXiv preprint arXiv:2205.05854, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[73] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[74] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[75] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[76] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[77] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[78] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22770" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[79] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[80] Songyang Zhang, Jinsong Su, and Jiebo Luo. Exploiting temporal relationships in video moment localization with natural language. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1230-1238, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[81] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[82] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[83] Jiahao Zhu, Daizong Liu, Pan Zhou, Xing Di, Yu Cheng, Song Yang, Wenzheng Xu, Zichuan Xu, Yao Wan, Lichao Sun, et al. Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[84] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[85] Yuchen Zhu, Cheng Shi, Dingyou Wang, Jiajin Tang, Zhengxuan Wei, Yu Wu, Guanbin Li, and Sibei Yang. Rethinking query-based transformer for continual image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4595-4606, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22771" + } + ] +] \ No newline at end of file diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_origin.pdf b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7699d2859e58514832fbeb76acdfa7e6563cda9 --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/461e9907-9af3-47f5-b610-f09041cded55_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6368c8406ff1dd2dab20874a8727fd8d831664a09775e9c782cda3eefd348329 +size 1538324 diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/full.md b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d358b4f0b3f1923497e1d2f179d597cdf38e15b0 --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/full.md @@ -0,0 +1,329 @@ +# Sim-DETR: Unlock DETR for Temporal Sentence Grounding + +Jiajin Tang $^{1*}$ , Zhengxuan Wei $^{1*}$ , Yuchen Zhu $^{1}$ , Cheng Shi $^{1}$ , Guanbin Li $^{2}$ , Liang Lin $^{2}$ , Sibei Yang $^{2\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ School of Computer Science and Engineering, Sun Yat-sen University {tangjj, weizhx2022}@shanghaiitech.edu.cn yangsb3@mail.sysu.edu.cn + +# Abstract + +Temporal sentence grounding aims to identify exact moments in a video that correspond to a given textual query, typically addressed with detection transformer (DETR) solutions. However, we find that typical strategies designed to enhance DETR do not improve, and may even degrade, its performance in this task. We systematically analyze and identify the root causes of this abnormal behavior: (1) conflicts between queries from similar target moments and (2) internal query conflicts due to the tension between global semantics and local localization. Building on these insights, we propose a simple yet powerful baseline, Sim-DETR, which extends the standard DETR with two minor modifications in the decoder layers: (1) constraining self-attention between queries based on their semantic and positional overlap and (2) adding query-to-frame alignment to bridge the global and local contexts. Experiments demonstrate that Sim-DETR unlocks the full potential of DETR for temporal sentence grounding, offering a strong baseline for future research. + +# 1. Introduction + +Video has become a dominant media on the internet, with short-form content rapidly expanding and achieving exponential growth in reach over recent years [18-20, 24]. Instead of passively watching entire videos, users now prefer to target specific segments of interest, with natural language descriptions serving as an intuitive and flexible way to convey intent [30, 61, 68]. This amplifies interest in the research topic of temporal sentence grounding [22, 25-27, 32, 40, 46, 47, 57, 71, 81], which aims to locate one or more semantically aligned segments within an untrimmed video according to a given natural language sentence, as shown in Figure 1a. + +Early methods either align sentences with predefined temporal proposals for selection [44, 70] or directly predict moment spans through cross-modal interactions between + +![](images/27ab39c0516a1dd467464753540dfc434879fd5381ee796ba941e0d44eab9188.jpg) +Figure 1. Illustration of (a) temporal sentence grounding task, (b) phenomenon of random matching across multiple layers, and (c) the challenge of achieving the balance between focus on boundary for span prediction and alignment with global semantics. + +language and frames [36, 37, 72]. Recent advances center on integrating the Detection Transformer (DETR) [2] for object detection into the temporal sentence grounding, leveraging its query-based proposal detection framework to eliminate hand-crafted proposal components and deliver superior grounding performance with high efficiency [22, 25, 26, 46, 47, 57]. They leverage moment queries—often several times more numerous than ground truth segments—in the multi-layer decoder, which are initialized randomly [26, 46, 57], generated from sentences [22, 47], or extracted as event units from the video [25], to locate and search the most suitable matches for target segments. A one-to-one label assignment is applied in the training objective to facilitate non-redundant predictions for each ground-truth segment. The overall architecture of the DETR-based temporal sentence grounding framework is shown in Figure 6a. + +However, we observe that strategies generally effective for DETR in other tasks, such as object detection, like increasing the number of queries or decoder layers, do not improve performance in temporal sentence grounding and may even degrade it (Sec 3.1). This motivates us to revisit the specific characteristics of temporal sentence grounding that give rise to the distinct challenges in applying DETR to this task. In a series of investigations on query-segment alignment (Sec 3.2), we verify that performance degradation primarily stems from the difficulty in identifying dis + +tinct local boundaries for target segments that share highly similar global semantics, leading to inherent conflicts both between and within queries. By diagnosing the similarities between queries and between queries and frames, we uncover two conflicts that result in learning puzzles for queries: + +- Queries corresponding to distinct target segments exhibit high similarity, leading to random matching with various targets and creating learning puzzles for the queries. The presence of multiple semantically similar target segments results in the high similarity between their corresponding queries, making them difficult to clearly distinguish (Sec 3.3). This causes each query to potentially match well with multiple segments, leading to significant randomness in the one-to-one matching for each target segment (Figure 1b). We observe that across different decoder layers, the queries matched for optimization can vary considerably (Sec 3.2). +- Each query faces an inherent conflict between encoding the global semantics of its corresponding segment and decoding its local boundary. To align well with a segment, a query must encode the segment's global semantics from the start to the end frame, such as the entire process of a person holding a cup. However, the query also needs to directly predict the segment's boundary, requiring focus on the local frames near the boundary (Figure 1c). This creates a trade-off: the query either prioritizes global semantics or local boundaries, making it challenging to optimize both simultaneously (Sec 3.3). + +Building on these findings, we propose a simple yet strong baseline for video temporal grounding, Sim-DETR. The framework builds on the standard DETR-based temporal sentence grounding architecture, with two minor yet straightforward design modifications to the decoder layers to address the two conflicts between and within queries (Figure 6), respectively. First, we adjust the self-attention weights between queries based on their pairwise correlation, encouraging indistinguishable queries to focus on different contexts, reducing their similarity, and allowing the most suitable query to gather information from related queries to refine its prediction (Figure 6b). Second, we introduce a query-to-frame matching and loss term, accounting for the alignment between the query and each frame within the segment, as shown in Figure 6c. This ensures that all frames, not just those near the boundary, contribute to segment localization. The full sequence of all target frames, from start to finish, acts as a bridge connecting the global semantics to the local boundary. + +Empirically, we indeed observe that our minor modifications (1) significantly reduce conflicts between queries corresponding to different target segments (Figure 3), (2) improve the alignment between the query's global semantic attention to frames and its boundary prediction (Figure 5), + +and (3) lead to more consistent query predictions across layers (Figure 4). Furthermore, (4) increasing the number of queries and decoder layers no longer causes performance degradation (Figure 2). Finally, due to the consistent matching and prediction of queries, (5) our Sim-DETR not only achieves substantial performance gains but also accelerates convergence (as shown in Appendix). All the above observations and advantages, including the minor modifications, are intended to unlock the potential of query-based frameworks for temporal sentence grounding and provide a simple yet strong baseline for future research. + +Our contributions are summarized as follows: + +- We systematically analyze the root causes of abnormal behavior in the DETR-based temporal sentence grounding framework, identifying two key conflicts: (1) target segments with highly similar semantics but distinct temporal localization create learning puzzles between queries, and (2) a trade-off between relying on global semantics for matching and local boundaries for localization within queries. +- Based on our analysis, we propose two modifications to form Sim-DETR: (1) a simple adjustment to the self-attention to resolve conflicts between queries and (2) the introduction of a query-to-frame matching to reconcile global semantics and local localization within queries. +- Our Sim-DETR, a simple yet powerful baseline, achieves consistent and significant improvements across all benchmarks. More importantly, it eliminates observed anomalies and exhibits faster convergence. + +# 2. Related Work + +Temporal Sentence Grounding (TSG) aims to identify specific video moments that align with given language expressions. As computer vision advances across neural architectures, generation, detection, segmentation, and multimodal tasks [4, 8, 21, 33, 53-55, 73, 82], temporal understanding requires task-specific optimizations. Early methods fall into proposal-free and proposal-based categories. Proposal-free methods use end-to-end frameworks by directly regressing boundary coordinates [6, 43, 76] or predicting frame-level boundary likelihood [16, 78]. Proposal-based approaches densely sample candidate segments via sliding windows [1, 12, 15, 23, 38, 80], temporal anchors [5, 32, 41, 75, 77, 81], or multimodal feature similarity [7, 34, 35, 66, 69]. Recently, DETR has shown exceptional effectiveness in detection tasks. [27, 65] pioneered DETR for TSG. Subsequent studies enhance cross-modal representations [46, 47, 57, 67], improve decoder queries with semantic/spatial information [22, 25, 58], or explore joint training with other temporal tasks [57, 71]. + +Detection Transformers (DETR) [2] is a pioneering end-to-end framework for object detection and segmentation that employs transformers to make direct set-based pre + +dictions [14, 52, 54, 55, 60, 61, 85]. However, DETR has notable limitations, especially its high computational complexity and slow convergence, impacting its practical efficiency. To address these issues, several derivative models have emerged. Some methods [51, 63, 74, 84] improve computational efficiency by reducing redundant calculations within the DETR architecture. Other models [9, 13, 59] refine the attention mechanism for more efficient information processing and resource utilization. Additionally [39, 45, 64] enhance convergence by embedding spatial information directly into the query design, which accelerates object localization. Recent models [28, 79] tackle optimization challenges in DETR's bipartite matching process by integrating denoising strategies during training, improving both convergence and performance. Despite recent advancements in enhancing DETR for object detection, there is a lack of systematic studies addressing the unique challenges of applying DETR to TSG and the potential difficulties in query optimization. + +# 3. Probing DETR's Inherent Behavior for TSG + +This study begins with the observation that enhancements effective for object detection in DETR [2] do not apply to temporal sentence grounding (TSG), as discussed in Sec 3.1. To explore the reasons behind DETR's failure to enhance, this section presents preliminary studies to reveal the behavior of queries in the DETR decoder, as queries are central to predicting proposals. Specifically, we examine the similarity between queries and outputs in Sec 3.2, and the attention between queries and inputs (i.e., frames) in Sec 3.3. These help reveal the underlying cause, and our modification addressing it not only eliminates the abnormal phenomena but also has a positive side effect (Sec 3.4). + +Experimental Settings. Unless otherwise specified, our default experimental setup involves conducting statistical analyses on the full validation set of QVHighlights [27]. CG-DETR [46] and TaskWeave [71], which serve as baselines for DETR-based TSG due to their strong performance while retaining the core vanilla DETR mechanism, are used with their standard configurations. + +# 3.1. Abnormal Phenomena + +Motivation. DETR is originally proposed for object detection, and increasing the number of queries and decoder layers typically improves performance. We explore whether this trend also applies to TSG. + +Setting. Both CG-DETR and TaskWeave default to 10 queries, with 3 and 2 decoder layers, respectively. This experiment specifically increases the number of queries and decoder layers to isolate their effects, keeping all other hyperparameters and training protocols constant. + +Result & Discussion. As shown in Figure 2, increasing the number of queries to 15 and 20 led to progressively + +![](images/33d6f4adc556afce73df5205366cfdbf662674600a9dffdec376d23e8e356225.jpg) +Figure 2. The impact of the number of queries and decoder layers. + +larger performance declines, with a drop of over $-2\%$ with TaskWeave [71] at 20 queries. Similarly, increasing the number of decoder layers caused a drop from $-1\%$ to $-2.5\%$ . In contrast, our Sim-DETR mitigates this phenomenon and achieves a slight improvement, demonstrating its robustness. This observation motivates an in-depth exploration of decoder layer, particularly the role of queries (see Sec 3.2). + +# 3.2. Similar Target Segments Cause Query Conflicts + +Motivation. Initially, we hypothesize that the lack of performance improvement with increased queries is due to query redundancy. However, performance declined, suggesting that conflicts between queries, rather than redundancy, prevent compatibility with existing queries. Similarly, conflicts between queries across layers may explain the ineffectiveness of increasing the number of layers. This motivates us to explore (1) the relationships between queries to identify their potential conflicts and (2) their variations across layers to reveal the impact at different layers. + +Setting. (1) To evaluate query relationships, we assess similarities between queries corresponding to the same segment (intra-segment similarity) and those from different segments (inter-segment similarity). Specifically, we identify the corresponding segment for each query based on the maximum IoU between the query's prediction and all ground-truth (GT) segments during inference. Queries whose IoU with the corresponding segment is 0.5 or greater are included in our analysis to ensure meaningful corresponding. We then compute feature similarities between queries associated with the same segment (intra) and those matched to different segments (inter). (2) To analyze variations across layers, we introduce the metric of cross-layer matching consistency. Specifically, during training optimization, each GT segment is uniquely matched to a query through bipartite matching at each decoding layer. We assess match consistency by measuring the proportion of segments that retain their corresponding queries across two consecutive decoder layers. + +Result & Discussion. (1) As shown in Figure 3, CGDETR fails to distinguish query similarities between inter- or intra-segment groups, indicating that queries are not clearly grouped with their corresponding segments. Consequently, the same query oscillates between segments, + +![](images/12b42318272f8cd778de121aec07c3f67597410fffad46af97933919b8be2852.jpg) +Figure 3. Distribution of query similarity for intra-segment (blue) and inter-segment (red) pairs. Our method can effectively distinguish intra-segment and inter-segment queries, ensuring more stable query-segment associations and reducing conflicts in query assignments across different segments. + +![](images/894e525cf266de01d9c4a5c26cfb550aa34da0aebdefda1a3efceba091c98937.jpg) +Figure 4. Measuring cross-layer matching consistency (y-axis, defined in Line-260) of segments across two consecutive decoding layers (x-axis). Our method achieves higher consistency. + +lacking stable and consistent corresponding. Figure 4 further supports our hypothesis, demonstrating that cross-layer matching consistency in CG-DETR is lower than ours. (2) We attribute this to multiple GT segments in a video sharing the same linguistic semantics. The high semantic similarity between GT segments results in high similarity between their corresponding queries. Figure 4 supports this, showing that the consistency for cases with multiple GT segments is lower than for those with a single segment. + +# 3.3. Global Matching vs. Local Localization + +Motivation. We further question whether the inconsistency in matching is solely due to query conflicts from multiple GT segments. If so, why does the previous method still show roughly $25\%$ mismatched queries in cases with a single target segment? Therefore, we shift our focus from conflicts between queries to conflicts within individual queries. A query serves dual roles: aligning with global linguistic semantics (global matching) and precisely localizing the segment, particularly its boundaries (local localization). This prompts us to explore whether conflicts arise from these roles within a single query. + +Setting. (1) To evaluate global matching, we use the query's classification confidence score, as it reflects the confidence of the query being referenced by the global semantics of the linguistic description. (2) To evaluate local localization, we + +![](images/8e21c353c2a08149735c3362c9cf8f98e645ea476b4c4879008433b2895113fb.jpg) +Figure 5. Global matching score (defined in Line-290) vs. local localization score (defined in Line-293). Compared to CGDETR, our method concentrates a query's attention significantly more within a single GT segment, rather than dispersing it across multiple segments, leading to improved local localization scores. + +extract attention scores between each query and frame from the cross-attention module in the last decoder layer, as they indicate query-based frame localization. We then compute the IoU score between this attention and the query's corresponding GT segment, reflecting the accuracy of the query's localization at the frame level. + +Result & Discussion. As shown in Figure 5, CG-DETR (marked in red) may yield low local localization scores despite high global matching scores, indicating that while queries align well with linguistic descriptions, they fail to achieve precise localization. Further analysis reveals that query attention may span frames in multiple GT segments (see Figure 5 "Attention"), leading to a low local score for the query's specific GT segment it corresponds to. + +# 3.4. Positive Side Effect + +Convert Abnormal to Normal: Our Sim-DETR effectively (1) distinguishes queries within the same target segment and across different segments in Figure 3, (2) achieves consistent matching across decoder layers in Figure 4, and (3) aligns with global semantics while ensuring precise localization in Figure 5. + +Positive Side Effect: (1) More importantly, Sim-DETR maintains stable and robust performance with increased queries and decoder layers (see Figure 2). (2) By eliminating the abnormal phenomena, it also achieves significantly faster convergence compared to previous methods, as shown in Appendix C. + +# 4. Method + +We present a simple yet effective TSG baseline that resolves query conflicts in existing DETR-based TSG frameworks [22, 25, 27, 46, 47, 57, 67, 71] through two minor yet key modifications. First, we introduce the standard DETR-based [2] TSG architecture (Sec 4.1). Sec 4.2 introduces our first modification, which adjusts the self-attention + +![](images/4b4b581d8f656c715184403c618e5cfacbb7a4a6b3c9e602f4743fcad84baa5f.jpg) +Figure 6. Framework of the proposed Sim-DETR, featuring two simple yet powerful modifications: Query Grouping and Ranking to mitigate cross-query conflicts and Global-Local Bridging to address global-local conflicts within each query. + +mechanism between queries in the decoder to avoid conflicts among them, based on the observations presented in Sec 3.2. Next, Sec 4.3 presents another improvement that addresses the conflict between global matching and local localization within the query (discussed in Sec 3.3) by introducing query-to-frame matching loss, which serves as the bridge between the global and local perspectives. + +# 4.1. Standard DETR-based TSG Architecture + +In TSG, the goal is to identify video segments with temporal spans that correspond to a given language expression. + +Feature Extraction. Previous DETR-based works [25, 47, 57, 71] employ CLIP [49] and SlowFast [10] to extract video features, represented as $\mathcal{T} \in \mathbb{R}^{N \times C}$ , where $N$ and $C$ stands for the number of frames and the hidden dimension, respectively. For the language expression, typically a sentence with $L$ tokens, the CLIP text encoder is used to obtain token-level features, resulting in $\mathcal{W} \in \mathbb{R}^{L \times C}$ . + +Multimodal Encoder. After feature extraction, DETR-based TSG methods [22, 46, 47, 57] employ a multimodal encoder to fuse these video and language features, producing multimodal video features, denoted as $\hat{\mathcal{T}}\in \mathbb{R}^{N\times C}$ . The key component of the multimodal encoder is a cross-attention mechanism [62], where video features $\mathcal{T}$ act as queries, while word features $\mathcal{W}$ serve as keys and values. + +Temporal Sentence Decoding. In the decoder, the model initializes a set of query to gather global information from the video, which is then decoded into local spans in the span prediction head. Specifically, given initialized queries $\mathcal{Q} \in \mathbb{R}^{M \times C}$ , where $M$ is the query number, and the multimodal video features $\hat{\mathcal{T}}$ , the decoder facilitates interaction between queries and gathers global information through interleaved + +self-attention and cross-attention modules, as follows: + +$$ +\begin{array}{l} \left\{ \begin{array}{l} \mathcal {Q} _ {q} ^ {\mathrm {S A}} = \operatorname {F C} _ {q} ^ {\mathrm {S A}} (\mathcal {Q}), \mathcal {Q} _ {k} ^ {\mathrm {S A}} = \operatorname {F C} _ {k} ^ {\mathrm {S A}} (\mathcal {Q}), \mathcal {Q} _ {v} ^ {\mathrm {S A}} = \operatorname {F C} _ {v} ^ {\mathrm {S A}} (\mathcal {Q}), \\ \mathcal {Q} ^ {\mathrm {S A}} = \operatorname {F C} _ {o} ^ {\mathrm {S A}} \left(\operatorname {s o f t m a x} \left(\frac {\mathcal {Q} _ {q} ^ {\mathrm {S A}} (\mathcal {Q} _ {k} ^ {\mathrm {S A}}) ^ {\intercal}}{\sqrt {C}}\right) \mathcal {Q} _ {v} ^ {\mathrm {S A}}\right). \end{array} \right. \\ \left\{ \begin{array}{l} \mathcal {Q} _ {q} ^ {\mathrm {C A}} = \mathrm {F C} _ {q} ^ {\mathrm {C A}} (\mathcal {Q} ^ {\mathrm {S A}}), \hat {\mathcal {T}} _ {k} = \mathrm {F C} _ {k} ^ {\mathrm {C A}} (\hat {\mathcal {T}}), \hat {\mathcal {T}} _ {v} = \mathrm {F C} _ {v} (\hat {\mathcal {T}}), \\ \mathcal {Q} _ {o} ^ {\mathrm {C A}} = \mathrm {F C} _ {o} ^ {\mathrm {C A}} \left(\mathrm {s o f t m a x} \left(\frac {\mathcal {Q} _ {q} ^ {\mathrm {C A}} (\hat {\mathcal {T}} _ {k}) ^ {\intercal}}{\sqrt {C}}\right) \hat {\mathcal {T}} _ {v}\right). \end{array} \right. \\ \hat {\mathcal {Q}} = \operatorname {M L P} \left(\operatorname {L a y e r N o r m} \left(\mathcal {Q} ^ {\mathrm {C A}}\right) + \mathcal {Q}\right), \\ B = \mathcal {H} _ {\text {s p a n}} (\hat {\mathcal {Q}}), \quad B = \left\{b _ {1}, \dots , b _ {M} \right\}, \tag {1} \\ \end{array} +$$ + +where $\mathrm{FC}(\cdot)$ denotes the fully connected projection layer, with SA and CA representing self-attention and cross-attention, respectively. Subscripts $q$ , $k$ , and $v$ indicate query, key, and value computations. MLP( $\cdot$ ) represents a nonlinear mapping with an activation function. $\mathcal{H}_{\mathrm{span}}(\cdot)$ denotes the span prediction head, which takes the updated query features $\hat{\mathcal{Q}}$ to predict the span set $B$ , where each $b_i \in \mathbb{R}^2$ represents the (start, end) of a predicted span. + +# 4.2. Query Grouping and Ranking + +In this section, we address the query conflicts identified in Sec 3.2: (1) the inability to distinguish between intra- and inter-segment queries (Figure 3), leading to query oscillation between segments without stable correspondence (Line-266), and (2) inconsistencies in segment-query matching across layers, where different queries may be matched to ground-truth (GT) segments at each layer for optimization, hindering effective learning (Figure 4). + +To differentiate queries across segments, the challenge arises from GT segments that inherently share identical lin + +![](images/4a450f95a57d73051b75d765db4ca14f10e9081b2460426e601ecea0dec4deb5.jpg) +Figure 7. Overview of our Query Grouping and Ranking. + +guistic semantics, leading to highly similar queries. Therefore, a natural solution is to directly group queries, assigning each group to a specific GT segment for training and prediction. Instead of hard grouping with predefined rules or extra hyperparameters, we learn a soft grouping based on predicted temporal spans of queries, where closer spans indicate potential corresponding for the same segment. Specifically, for any two queries $q_{i}$ and $q_{j}$ , we compute their span distance, which is then used in Eq. 4 to learn their grouping: + +$$ +\mathcal {S} _ {i, j} ^ {\text {i n t r a}} = \| b _ {i} - b _ {j} \| _ {2}, \tag {2} +$$ + +where $\| \cdot \|_2$ denotes the L2 norm. Notably, we use L2 distance instead of the commonly used L1 because, when two spans are close (normalized distances $\leq 1$ ), L2 distance decreases more sharply than L1, imposing a smaller penalty on minor differences (see Appendix G). + +Meanwhile, to enhance cross-layer matching consistency, we draw inspiration from EASE-DETR [14]'s query ranking strategy in object detection. By ranking queries within a segment at each layer, the leading query is prioritized for segment matching in subsequent layers. However, directly ranking queries by prediction scores, as in EASE-DETR, proves ineffective (see Appendix D). As analyzed in Sec. 3.3, high prediction/confidence scores do not guarantee precise predictions and may correspond to poor localization. Instead, we introduce an IoU head to assess localization accuracy and rank queries based on both factors, as follows: + +$$ +R _ {\text {r a n k}} \left(q _ {i}, q _ {j}\right) = \left\{ \begin{array}{l l} + 1 & \mathcal {P} _ {i} ^ {\text {c l s}} \circ \mathcal {P} _ {i} ^ {\text {I o U}} \geq \mathcal {P} _ {j} ^ {\text {c l s}} \circ \mathcal {P} _ {j} ^ {\text {I o U}} \\ - 1 & \mathcal {P} _ {i} ^ {\text {c l s}} \circ \mathcal {P} _ {i} ^ {\text {I o U}} < \mathcal {P} _ {j} ^ {\text {c l s}} \circ \mathcal {P} _ {j} ^ {\text {I o U}}, \end{array} \right. \tag {3} +$$ + +where $\mathcal{P}^{\mathrm{cls}}$ and $\mathcal{P}^{\mathrm{IoU}}$ denote the query's confidence score and IoU prediction, respectively, with subscripts $i$ and $j$ indicating queries $q_{i}$ and $q_{j}$ . $R_{\mathrm{rank}}(q_i,q_j) = 1$ indicates that query $q_{i}$ has a higher priority than query $q_{j}$ , considering both global alignment and local localization. + +Finally, we incorporate query grouping within segments and intra-segment ranking to infer relative relationships, then reshape query-to-query self-attention following [14] to mitigate conflicts, as follows: + +$$ +\begin{array}{l} \mathcal {S} ^ {\text {a t t n}} = \operatorname {s i g m o i d} \left(\mathrm {M L P} \left(\mathcal {S} ^ {\text {i n t r a}} \circ \mathcal {R} _ {\text {r a n k}}\right), \right. \\ \mathcal {Q} ^ {\mathrm {S A}} = \mathrm {F C} _ {o} ^ {\mathrm {S A}} \left(\text {s o f t m a x} \left(\left(\frac {\mathcal {Q} _ {q} ^ {\mathrm {S A}} \left(\mathcal {Q} _ {k} ^ {\mathrm {S A}}\right) ^ {\intercal}}{\sqrt {C}}\right) \mathcal {S} ^ {\mathrm {a t t n}}\right) \mathcal {Q} _ {v} ^ {\mathrm {S A}}\right) \tag {4} \\ \end{array} +$$ + +where $S^{\mathrm{attn}}$ represents query relationships, with higher values promoting high-quality queries to dominate predictions by integrating similar queries within the same group. + +# 4.3. Global-Local Bridging + +In this section, we address the query conflicts identified in Sec 3.3, specifically the conflict between a query's global semantic matching and local localization. Since queries overly rely on global semantics while neglecting local localization, our core idea is to reinforce fine-grained frame-level localization within the segment. Specifically, we introduce a query-to-frame matching mechanism and loss term to enhance localization beyond segment boundary prediction, ensuring alignment with in-segment frames while avoiding matches with those from other segments. The full sequence of frames within the segment, from start to finish, serves as a bridge linking global semantics to local boundaries. + +During training, we compute the semantic similarity between each query $q_{i}$ and all frames within its corresponding ground-truth span, maximizing similarity to ensure full alignment with in-segment frames while minimizing similarity to out-segment frames. The computation process is outlined as follows, + +$$ +\begin{array}{l} z = \operatorname {s i g m o i d} (\tau \cdot \cos (q _ {i}, \hat {\mathcal {T}})), \\ \mathcal {L} _ {\mathrm {b r i d g e}} = \lambda_ {\mathrm {b r i d g e}} \frac {- \sum_ {j} z _ {j} \mathbb {I} \left[ b _ {i} ^ {\mathrm {g t}} \right] _ {j}}{\sum_ {j} z _ {j} \left(1 - \mathbb {I} \left[ b _ {i} ^ {\mathrm {g t}} \right] _ {j}\right) + \sum_ {j} \mathbb {I} \left[ b _ {i} ^ {\mathrm {g t}} \right] _ {j}}, \tag {5} \\ \end{array} +$$ + +where $z \in \mathbb{R}^N$ donates query-to-frame similarity, calculated using the cosine similarity between the query $q_i$ and the $N$ frame features $\tilde{T}$ defined in Line-347, and $\tau$ is a learnable scaling coefficient. $b_i^{\mathrm{gt}}$ is the ground-truth span corresponding to query $q_i$ , while $\mathbb{I}[b_i^{\mathrm{gt}}] \in \mathbb{R}^N$ is an indicator function that assigns 1 to frames within the ground truth and 0 otherwise, with $j$ representing the frame index. Thus, the numerator in $\mathcal{L}_{\mathrm{bridge}}$ , i.e., $-\sum_{j} z_j \mathbb{I}[b_i^{\mathrm{gt}}]_j$ , encourages the query to be similar to every frame within its corresponding segment. Meanwhile, the denominator's first term, $\sum_{j} z_j (1 - \mathbb{I}[b_i^{\mathrm{gt}}]_j)$ , represents query $q_i$ 's similarity to out-segment frames, promoting its minimization. The denominator's second term, $\sum_{j} \mathbb{I}[b_i^{\mathrm{gt}}]_j$ , represents the length of the ground-truth segment, serving as a normalization factor. And $\lambda_{\mathrm{bridge}}$ is a hyperparameter that controls the loss weight. + +# 4.4. Overall Loss + +The overall loss is: $\mathcal{L} = \mathcal{L}_{\mathrm{MD}} + \lambda_{\mathrm{bridge}}\mathcal{L}_{\mathrm{bridge}} + \lambda_{\mathrm{iou}}\mathcal{L}_{\mathrm{iou}}$ where $\mathcal{L}_{\mathrm{MD}}$ includes the L1, gIoU, classification, and saliency losses, as in Moment DETR [26]. $\mathcal{L}_{\mathrm{iou}}$ is the loss for the IoU head introduced in Line-406, designed to assist in ranking queries using $\mathcal{P}_i^{\mathrm{IoU}}$ and $\mathcal{P}_j^{\mathrm{IoU}}$ in Eq 3. The IoU head implementation follows [25, 58]. Ablations for hyperparameters $\lambda_{\mathrm{bridge}}$ and $\lambda_{\mathrm{iou}}$ are detailed in Appendix E. + +
Methodtestval
R1mAPR1mAP
@0.5@0.7@0.5@0.75Avg.@0.5@0.7@0.5@0.75Avg.
M-DETR [27]NeurIPS'2152.8933.0254.8229.4030.7353.9434.84--32.20
UMT [40]CVPR'2256.2341.1853.8337.0136.1260.2644.26--38.59
QD-DETR [47]CVPR'2362.4044.9862.5239.8839.8662.6846.6662.2341.8241.22
UniVTG [32]ICCV'2358.8640.8657.6035.5935.4759.74---36.13
EaTR [22]ICCV'23-----61.3645.7961.8641.9141.74
MomentDiff [29]NeurIPS'2357.4239.6654.0235.7335.95-----
TR-DETR [57]AAAI'2464.6648.9663.9843.7342.6267.1051.4866.2746.4245.09
UVCOM [67]CVPR'2463.5547.4763.3742.6743.1865.1051.81--45.79
TaskWeave [71]CVPR'24-----64.2650.0665.3946.4745.38
BAM-DETR [25]ECCV'2462.7148.6464.5746.3345.3665.1051.6165.4148.5647.61
CG-DETR [46]Arxiv'2465.4348.3864.5142.7742.8667.3552.0665.5745.7344.93
SpikeMba [31]Arxiv'2464.1349.42-43.6743.7965.3251.33-44.9644.84
Sim-DETR(Ours)67.6450.9167.8147.5946.9369.4854.0669.7051.1149.50
+ +Table 1. Experimental results on the QVHighlights benchmark, comparing ours with state-of-the-art methods. + +
MethodTACoSCharades-STA
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [81]40.0127.9912.9227.2258.7646.0227.5041.25
M-DETR [27]37.9724.6711.9725.4965.8352.0730.5945.54
MomentDiff [29]44.7833.68---55.5732.42-
QD-DETR [47]-----57.3132.55-
UniVTG [32]51.4434.9717.3533.6070.8158.0135.6550.10
CG-DETR [46]52.2339.6122.2336.4870.4358.4436.3450.13
UVCOM [67]-36.3923.32--59.2536.64-
SpikeMba [31]51.9839.3422.8335.8171.2459.6536.1251.74
Sim-DETR(Ours)57.0642.7926.8239.4473.0961.3439.6252.56
+ +# 5. Experiments + +# 5.1. Datasets and Implementation Details + +Datasets. Following [27, 31, 32, 47, 67], we conduct comprehensive evaluation experiments across three datasets: QVHighlights [27], Charades-STA [12], and TACoS [50]. The QVHighlights dataset contains 10,310 text expressions and 18,367 distinct events, with most expressions corresponding to multiple events. Charades-STA is derived from a collection of 9,848 indoor scene videos, encompassing 16,128 events. TACoS includes only 273 videos, with an average length approaching five minutes per video. For a fair comparison, we adopt the same evaluation metrics as in previous methods. On the QVHighlights benchmark, we use Recall and Mean Average Precision (mAP) as primary metrics, while for Charades-STA and TACoS, we employ Recall and Mean Intersection over Union (mIoU). + +Implementation Details. We set the input video resolution to $224 \times 224$ and represent each video by concatenating the CLIP [49] [CLS] tokens with SlowFast [10] features. For the Charades-STA and TACoS datasets, we additionally constructed two experiments using I3D [3] and VGG [56] as visual encoders, with Glove [48] embeddings for text features. The model is trained for 200 epochs on a single NVIDIA A40 GPU, using the AdamW [42] optimizer with + +Table 2. Comparison results on the TACoS and Charades-STA benchmarks. + +
BackboneMethodR@0.5R@0.7
VGG [56]2D-TAN [81]40.9422.85
FVMR [11]42.3624.14
SSRN [83]46.7227.98
UMT [40]48.3129.25
MomentDiff [29]51.9428.25
QD-DETR [47]52.7731.13
TR-DETR [57]53.4730.81
CG-DETR [46]55.2234.19
I3D [3]Sim-DETR(Ours)55.9735.38
MAN [77]46.5322.72
VSLNet [78]47.3130.19
QD-DETR [47]50.6731.02
TR-DETR [57]55.5133.66
TaskWeave [71]53.3631.40
Sim-DETR(Ours)57.5535.73
+ +Table 3. Experimental results with VGG or I3D as backbone. + +an initial learning rate of 1e-4. Note that, for a fair comparison, the total number of layers in our Sim-DETR is set to 6, the same as recent works [22, 46, 67]. + +# 5.2. Comparison with State-of-the-Art Methods + +As shown in Tables 1-3, we report detailed evaluation results across the QVHighlights, Charades-STA, and TACoS benchmarks. Our Sim-DETR consistently outperforms all state-of-the-art methods (SOTAs) across all metrics. + +Results on QVHighlights are shown in Table 1. Compared to the latest published work, BAM-DETR [25], our Sim + +
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline65.4850.8465.3445.8244.97
+ Liou66.5851.9465.6846.3045.22
+ QGR68.7752.2667.7248.2847.03
+ GLB67.1652.7768.5848.8848.17
+ QGR,GLB69.4854.0669.7051.1149.50
+ +Table 4. Ablation study on components. + +
Inner RelevanceOuter GlobalOuter LocalmAP
(1)span border distconfidence predIoU pred49.50
(2)center distconfidence predIoU pred48.59
(3)span IoU48.94
(4)span border distw/oIoU pred48.93
(5)span border distconfidence predw/o48.66
(6)center pred48.97
+ +Table 5. Impact of different measures. + +DETR outperforms it by an average of $+2.88\%$ across all metrics. More importantly, it achieves remarkable gains of $+4.93\%$ and $+2.27\%$ in R1@0.5 and R1@0.7, respectively. In comparison TR-DETR [57], which uses standard DETR [2] as its basic framework and is used as our baseline model, our Sim-DETR achieves an average improvement of $+4.31\%$ and $+4.41\%$ in mAP for test and validation, respectively. These improvements underscore the effectiveness of our Sim-DETR in fundamentally addressing conflicts within and between queries. We further conduct a comparison with SpikeMba [31], a contemporary method built on the Mamba [17] framework. Without any complex module designs, the consistent improvements across all metrics highlight the simplicity and effectiveness of our Sim-DETR. Sim-DETR establishes an efficient baseline for the task, significantly advancing research in TSG. + +Results on Charades-STA and TACoS are presented in Table 2, where we consistently outperform all methods on both datasets. Compared to the state-of-the-art methods, we achieve a $+1.97\%$ improvement on the Charades-STA dataset (with SpikeMba as the SOTA) and a $+3.89\%$ improvement on the TACoS dataset (with CG-DETR [46] as the SOTA). These enhancements further demonstrate the outstanding capability of our method in temporal localization. More importantly, they confirm that our Sim-DETR exhibits significant performance advantages across various scenarios, types, and lengths of videos, highlighting its robustness and generalizability. + +Results with Different Backbones. Following previous works, we conduct two additional experiments on the QVHighlights dataset. These experiments use VGG and I3D for visual feature extraction and use Glove [48]embedding as text features. Table 3 presents the detailed results of these two experiments, where we achieve state-of-the-art results in both cases. Notably, for the video feature extractor I3D, we surpass the latest pub + +lished method by $+4.26\%$ . These results further demonstrate the excellent generalization capabilities of our SimDETR across different backbones. + +# 5.3. Ablation Study + +Roadmap for building a simple yet effective TSG baseline is presented in Table 4. We employ TR-DETR without MR2HD module as our baseline, which achieves an mAP of $44.97\%$ . Since our Query Grouping and Ranking (QGR) utilizes $\mathcal{L}_{\mathrm{iou}}$ proposed by [25, 58], we additionally conduct an ablation study to isolate the impact of $\mathcal{L}_{\mathrm{iou}}$ . (1) QGR: To mitigate conflicts and optimization challenges arising from overly similar semantics between queries, we introduce the QGR, which improves the baseline by $+2.07\%$ on mAP, representing a subtle yet significant enhancement. (2) Global-Local Bridging (GLB): To address conflicts and ambiguities between global semantics and local localization within queries, we introduce GLB to enable seamless transformation through query-to-frame alignment. Experimental results show a performance boost to $48.17\%$ on mAP, indicating the effectiveness of GLB in bridging global-to-local transformations. However, without QGR to address semantic conflicts in temporal DETR, the recall predictably decreased by $-3.1\%$ . (3) Overall Framework: With these adjustments, our Sim-DETR achieves a remarkable SOTA performance: $49.50\%$ on mAP and $69.48\%$ on R1@0.5. + +The Effects of Different Conflict Metrics in QGR. Table 5 summarizes five experiments on conflict metrics in QGR. (1) The first row displays our final implementation, achieving $49.50\%$ mAP. (2) & (3) We evaluate L2 center distance and span IoU for inter-query relationships. Center distance alone yields $48.59\%$ mAP, highlighting its inadequacy in segment representation. Span IoU outperforms center distance but lags by $0.56\%$ compared to boundary distance, likely due to IoU's broader tolerance, which reduces precision in query relevance. (4) To assess outer global semantic alignment, we remove global semantics, resulting in a $-0.57\%$ drop, emphasizing their role in query interaction and quality. (5) Finally, we test local location metrics in query prioritization. Excluding the outer local metric significantly lowers performance $(-0.84\%$ mAP), indicating that without local information, queries encounter conflicts due to overly generalized global semantics, leading to random matches. Consistent with (2), center localization accuracy shows similar limitations, with a $0.53\%$ decrease. + +# 6. Conclusion + +In this paper, we propose a concise and efficient temporal sentence grounding framework Sim-DETR, introducing two simple modifications to the decoder for resolving conflicts both between and within queries. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods across all benchmarks. + +Acknowledgment: This work is supported by the National Natural Science Foundation of China (No.62206174). + +# References + +[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 8 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7 +[4] Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[5] Jingyuan Chen, Xinpeng Chen, Lin Ma, Zequn Jie, and Tat-Seng Chua. Temporally grounding natural sentence in video. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 162-171, 2018. 2 +[6] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10551-10558, 2020. 2 +[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 2 +[8] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711-13722, 2024. 2 +[9] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 3 +[10] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 7 +[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 7 +[12] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In + +Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 7 +[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. arXiv preprint arXiv:2101.07448, 2021. 3 +[14] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 3, 6 +[15] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 245-253. IEEE, 2019. 2 +[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 2 +[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 8 +[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 1 +[19] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7366–7375, 2018. +[20] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. Advances in Neural Information Processing Systems, 36:26135-26158, 2023. 1 +[21] Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. Mvtokenflow: High-quality 4d content generation using multiview token flow. arXiv preprint arXiv:2502.11697, 2025. 2 +[22] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 1, 2, 4, 5, 7 +[23] Bin Jiang, Xin Huang, Chao Yang, and Junsong Yuan. Cross-modal video moment retrieval with spatial and language-temporal attention. In Proceedings of the 2019 on international conference on multimedia retrieval, pages 217-225, 2019. 2 +[24] Manjin Kim, Heeseung Kwon, Chunyu Wang, Suha Kwak, and Minsu Cho. Relational self-attention: What's missing in attention for video understanding. Advances in Neural Information Processing Systems, 34:8046–8059, 2021. 1 +[25] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 1, 2, 4, 5, 6, 7, 8 + +[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 6 +[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Neurips, pages 11846-11858, 2021. 1, 2, 3, 4, 7 +[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3 +[29] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In Neurips, 2023. 7 +[30] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 1 +[31] Wenrui Li, Xiaopeng Hong, Ruiqin Xiong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 7, 8 +[32] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 7 +[33] Liang Lin, Pengxiang Yan, Xiaoqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 2 +[34] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4536-4545, 2022. 2 +[35] Daizong Liu, Xiaoye Qu, Jianfeng Dong, and Pan Zhou. Adaptive proposal generation network for temporal sentence localization in videos. arXiv preprint arXiv:2109.06398, 2021. 2 +[36] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1665-1673, 2022. 1 +[37] Daizong Liu, Xiaoye Qu, and Wei Hu. Reducing the vision and language bias for temporal sentence grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4092-4101, 2022. 1 +[38] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2 +[39] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic + +anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3 +[40] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 1, 7 +[41] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. $r^2$ -tuning: Efficient image-to-video transfer learning for video temporal grounding. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[42] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 2 +[44] Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553, 2020. 1 +[45] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3 +[46] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 1, 2, 3, 4, 5, 7, 8 +[47] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2, 4, 5, 7 +[48] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 7, 8 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7 +[50] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 7 +[51] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021. 3 +[52] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 3 +[53] Cheng Shi and Sibei Yang. Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932-2941, 2023. 2 +[54] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 3 +[55] Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In European Conference on Computer Vision, pages 210-226. Springer, 2024. 2, 3 +[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7 +[57] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4998-5007, 2024. 1, 2, 4, 5, 7, 8 +[58] Xiaolong Sun, Liushuai Shi, Le Wang, Sanping Zhou, Kun Xia, Yabing Wang, and Gang Hua. Diversifying query: Region-guided transformer for temporal sentence grounding. arXiv preprint arXiv:2406.00143, 2024. 2, 6, 8 +[59] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3 +[60] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23570-23580, 2023. 3 +[61] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1, 3 +[62] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5 +[63] Tao Wang, Li Yuan, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4661-4670, 2021. 3 +[64] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 3 +[65] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore-and-match: Bridging proposal-based and proposal-free with transformer for sentence grounding in videos. arXiv preprint arXiv:2201.10168, 2022. 2 +[66] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for + +two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2986-2994, 2021. 2 +[67] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2, 4, 7 +[68] Chen-Wei Xie, Siyang Sun, Xiong Xiong, Yun Zheng, Deli Zhao, and Jingren Zhou. Ra-clip: Retrieval augmented contrastive language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19265–19274, 2023. 1 +[69] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9062–9069, 2019. 2 +[70] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Unsupervised alignment-based iterative evidence retrieval for multihop question answering. arXiv preprint arXiv:2005.01218, 2020. 1 +[71] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18308-18318, 2024. 1, 2, 3, 4, 5, 7 +[72] Shuo Yang and Xinxiao Wu. Entity-aware and motion-aware transformers for language-driven action localization in videos. arXiv preprint arXiv:2205.05854, 2022. 1 +[73] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 2 +[74] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 3 +[75] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. Advances in Neural Information Processing Systems, 32, 2019. 2 +[76] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2 +[77] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2, 7 +[78] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 2, 7 + +[79] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3 +[80] Songyang Zhang, Jinsong Su, and Jiebo Luo. Exploiting temporal relationships in video moment localization with natural language. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1230-1238, 2019. 2 +[81] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 7 +[82] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 2 +[83] Jiahao Zhu, Daizong Liu, Pan Zhou, Xing Di, Yu Cheng, Song Yang, Wenzheng Xu, Zichuan Xu, Yao Wan, Lichao Sun, et al. Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514, 2023. 7 +[84] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3 +[85] Yuchen Zhu, Cheng Shi, Dingyou Wang, Jiajin Tang, Zhengxuan Wei, Yu Wu, Guanbin Li, and Sibei Yang. Rethinking query-based transformer for continual image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4595-4606, 2025. 3 \ No newline at end of file diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/images.zip b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3a9d134a1f8adc5934126d6e1d4eae51b84fabc --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:888944cf2194cd28d535954af684be1e0ffe0ecc5d3f4e20056f14edd86ca832 +size 665216 diff --git a/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/layout.json b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0daa5735f5d1a9c88118c16d3d5424cb0c6c0d11 --- /dev/null +++ b/2025/Sim-DETR_ Unlock DETR for Temporal Sentence Grounding/layout.json @@ -0,0 +1,9581 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "type": "text", + "content": "Sim-DETR: Unlock DETR for Temporal Sentence Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": "Jiajin Tang" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Zhengxuan Wei" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Yuchen Zhu" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Cheng Shi" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Guanbin Li" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Liang Lin" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": ", Sibei Yang" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": "ShanghaiTech University " + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 566, + 186 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Sun Yat-sen University {tangjj, weizhx2022}@shanghaiitech.edu.cn yangsb3@mail.sysu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 239, + 297, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 468 + ], + "type": "text", + "content": "Temporal sentence grounding aims to identify exact moments in a video that correspond to a given textual query, typically addressed with detection transformer (DETR) solutions. However, we find that typical strategies designed to enhance DETR do not improve, and may even degrade, its performance in this task. We systematically analyze and identify the root causes of this abnormal behavior: (1) conflicts between queries from similar target moments and (2) internal query conflicts due to the tension between global semantics and local localization. Building on these insights, we propose a simple yet powerful baseline, Sim-DETR, which extends the standard DETR with two minor modifications in the decoder layers: (1) constraining self-attention between queries based on their semantic and positional overlap and (2) adding query-to-frame alignment to bridge the global and local contexts. Experiments demonstrate that Sim-DETR unlocks the full potential of DETR for temporal sentence grounding, offering a strong baseline for future research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 135, + 503 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 512, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 655 + ], + "type": "text", + "content": "Video has become a dominant media on the internet, with short-form content rapidly expanding and achieving exponential growth in reach over recent years [18-20, 24]. Instead of passively watching entire videos, users now prefer to target specific segments of interest, with natural language descriptions serving as an intuitive and flexible way to convey intent [30, 61, 68]. This amplifies interest in the research topic of temporal sentence grounding [22, 25-27, 32, 40, 46, 47, 57, 71, 81], which aims to locate one or more semantically aligned segments within an untrimmed video according to a given natural language sentence, as shown in Figure 1a." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "Early methods either align sentences with predefined temporal proposals for selection [44, 70] or directly predict moment spans through cross-modal interactions between" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 213, + 539, + 344 + ], + "blocks": [ + { + "bbox": [ + 328, + 213, + 539, + 344 + ], + "lines": [ + { + "bbox": [ + 328, + 213, + 539, + 344 + ], + "spans": [ + { + "bbox": [ + 328, + 213, + 539, + 344 + ], + "type": "image", + "image_path": "27ab39c0516a1dd467464753540dfc434879fd5381ee796ba941e0d44eab9188.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 348, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 392 + ], + "type": "text", + "content": "Figure 1. Illustration of (a) temporal sentence grounding task, (b) phenomenon of random matching across multiple layers, and (c) the challenge of achieving the balance between focus on boundary for span prediction and alignment with global semantics." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 593 + ], + "type": "text", + "content": "language and frames [36, 37, 72]. Recent advances center on integrating the Detection Transformer (DETR) [2] for object detection into the temporal sentence grounding, leveraging its query-based proposal detection framework to eliminate hand-crafted proposal components and deliver superior grounding performance with high efficiency [22, 25, 26, 46, 47, 57]. They leverage moment queries—often several times more numerous than ground truth segments—in the multi-layer decoder, which are initialized randomly [26, 46, 57], generated from sentences [22, 47], or extracted as event units from the video [25], to locate and search the most suitable matches for target segments. A one-to-one label assignment is applied in the training objective to facilitate non-redundant predictions for each ground-truth segment. The overall architecture of the DETR-based temporal sentence grounding framework is shown in Figure 6a." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "However, we observe that strategies generally effective for DETR in other tasks, such as object detection, like increasing the number of queries or decoder layers, do not improve performance in temporal sentence grounding and may even degrade it (Sec 3.1). This motivates us to revisit the specific characteristics of temporal sentence grounding that give rise to the distinct challenges in applying DETR to this task. In a series of investigations on query-segment alignment (Sec 3.2), we verify that performance degradation primarily stems from the difficulty in identifying dis" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 60, + 702, + 263, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 702, + 263, + 713 + ], + "spans": [ + { + "bbox": [ + 60, + 702, + 263, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author is Sibei Yang." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22760" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "tinct local boundaries for target segments that share highly similar global semantics, leading to inherent conflicts both between and within queries. By diagnosing the similarities between queries and between queries and frames, we uncover two conflicts that result in learning puzzles for queries:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 147, + 295, + 422 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 147, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 289 + ], + "type": "text", + "content": "- Queries corresponding to distinct target segments exhibit high similarity, leading to random matching with various targets and creating learning puzzles for the queries. The presence of multiple semantically similar target segments results in the high similarity between their corresponding queries, making them difficult to clearly distinguish (Sec 3.3). This causes each query to potentially match well with multiple segments, leading to significant randomness in the one-to-one matching for each target segment (Figure 1b). We observe that across different decoder layers, the queries matched for optimization can vary considerably (Sec 3.2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 422 + ], + "type": "text", + "content": "- Each query faces an inherent conflict between encoding the global semantics of its corresponding segment and decoding its local boundary. To align well with a segment, a query must encode the segment's global semantics from the start to the end frame, such as the entire process of a person holding a cup. However, the query also needs to directly predict the segment's boundary, requiring focus on the local frames near the boundary (Figure 1c). This creates a trade-off: the query either prioritizes global semantics or local boundaries, making it challenging to optimize both simultaneously (Sec 3.3)." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 423, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 651 + ], + "type": "text", + "content": "Building on these findings, we propose a simple yet strong baseline for video temporal grounding, Sim-DETR. The framework builds on the standard DETR-based temporal sentence grounding architecture, with two minor yet straightforward design modifications to the decoder layers to address the two conflicts between and within queries (Figure 6), respectively. First, we adjust the self-attention weights between queries based on their pairwise correlation, encouraging indistinguishable queries to focus on different contexts, reducing their similarity, and allowing the most suitable query to gather information from related queries to refine its prediction (Figure 6b). Second, we introduce a query-to-frame matching and loss term, accounting for the alignment between the query and each frame within the segment, as shown in Figure 6c. This ensures that all frames, not just those near the boundary, contribute to segment localization. The full sequence of all target frames, from start to finish, acts as a bridge connecting the global semantics to the local boundary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Empirically, we indeed observe that our minor modifications (1) significantly reduce conflicts between queries corresponding to different target segments (Figure 3), (2) improve the alignment between the query's global semantic attention to frames and its boundary prediction (Figure 5)," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "and (3) lead to more consistent query predictions across layers (Figure 4). Furthermore, (4) increasing the number of queries and decoder layers no longer causes performance degradation (Figure 2). Finally, due to the consistent matching and prediction of queries, (5) our Sim-DETR not only achieves substantial performance gains but also accelerates convergence (as shown in Appendix). All the above observations and advantages, including the minor modifications, are intended to unlock the potential of query-based frameworks for temporal sentence grounding and provide a simple yet strong baseline for future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 204, + 499, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 499, + 214 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 499, + 214 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 216, + 553, + 419 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 216, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 553, + 311 + ], + "type": "text", + "content": "- We systematically analyze the root causes of abnormal behavior in the DETR-based temporal sentence grounding framework, identifying two key conflicts: (1) target segments with highly similar semantics but distinct temporal localization create learning puzzles between queries, and (2) a trade-off between relying on global semantics for matching and local boundaries for localization within queries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 312, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 371 + ], + "type": "text", + "content": "- Based on our analysis, we propose two modifications to form Sim-DETR: (1) a simple adjustment to the self-attention to resolve conflicts between queries and (2) the introduction of a query-to-frame matching to reconcile global semantics and local localization within queries." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 372, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 553, + 419 + ], + "type": "text", + "content": "- Our Sim-DETR, a simple yet powerful baseline, achieves consistent and significant improvements across all benchmarks. More importantly, it eliminates observed anomalies and exhibits faster convergence." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 430, + 401, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 401, + 442 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 401, + 442 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 449, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 677 + ], + "type": "text", + "content": "Temporal Sentence Grounding (TSG) aims to identify specific video moments that align with given language expressions. As computer vision advances across neural architectures, generation, detection, segmentation, and multimodal tasks [4, 8, 21, 33, 53-55, 73, 82], temporal understanding requires task-specific optimizations. Early methods fall into proposal-free and proposal-based categories. Proposal-free methods use end-to-end frameworks by directly regressing boundary coordinates [6, 43, 76] or predicting frame-level boundary likelihood [16, 78]. Proposal-based approaches densely sample candidate segments via sliding windows [1, 12, 15, 23, 38, 80], temporal anchors [5, 32, 41, 75, 77, 81], or multimodal feature similarity [7, 34, 35, 66, 69]. Recently, DETR has shown exceptional effectiveness in detection tasks. [27, 65] pioneered DETR for TSG. Subsequent studies enhance cross-modal representations [46, 47, 57, 67], improve decoder queries with semantic/spatial information [22, 25, 58], or explore joint training with other temporal tasks [57, 71]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Detection Transformers (DETR) [2] is a pioneering end-to-end framework for object detection and segmentation that employs transformers to make direct set-based pre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22761" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "content": "dictions [14, 52, 54, 55, 60, 61, 85]. However, DETR has notable limitations, especially its high computational complexity and slow convergence, impacting its practical efficiency. To address these issues, several derivative models have emerged. Some methods [51, 63, 74, 84] improve computational efficiency by reducing redundant calculations within the DETR architecture. Other models [9, 13, 59] refine the attention mechanism for more efficient information processing and resource utilization. Additionally [39, 45, 64] enhance convergence by embedding spatial information directly into the query design, which accelerates object localization. Recent models [28, 79] tackle optimization challenges in DETR's bipartite matching process by integrating denoising strategies during training, improving both convergence and performance. Despite recent advancements in enhancing DETR for object detection, there is a lack of systematic studies addressing the unique challenges of applying DETR to TSG and the potential difficulties in query optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 310, + 295, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 323 + ], + "type": "text", + "content": "3. Probing DETR's Inherent Behavior for TSG" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 330, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 473 + ], + "type": "text", + "content": "This study begins with the observation that enhancements effective for object detection in DETR [2] do not apply to temporal sentence grounding (TSG), as discussed in Sec 3.1. To explore the reasons behind DETR's failure to enhance, this section presents preliminary studies to reveal the behavior of queries in the DETR decoder, as queries are central to predicting proposals. Specifically, we examine the similarity between queries and outputs in Sec 3.2, and the attention between queries and inputs (i.e., frames) in Sec 3.3. These help reveal the underlying cause, and our modification addressing it not only eliminates the abnormal phenomena but also has a positive side effect (Sec 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 558 + ], + "type": "text", + "content": "Experimental Settings. Unless otherwise specified, our default experimental setup involves conducting statistical analyses on the full validation set of QVHighlights [27]. CG-DETR [46] and TaskWeave [71], which serve as baselines for DETR-based TSG due to their strong performance while retaining the core vanilla DETR mechanism, are used with their standard configurations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 185, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 185, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 185, + 576 + ], + "type": "text", + "content": "3.1. Abnormal Phenomena" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "Motivation. DETR is originally proposed for object detection, and increasing the number of queries and decoder layers typically improves performance. We explore whether this trend also applies to TSG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "content": "Setting. Both CG-DETR and TaskWeave default to 10 queries, with 3 and 2 decoder layers, respectively. This experiment specifically increases the number of queries and decoder layers to isolate their effects, keeping all other hyperparameters and training protocols constant." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Result & Discussion. As shown in Figure 2, increasing the number of queries to 15 and 20 led to progressively" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 67, + 548, + 176 + ], + "blocks": [ + { + "bbox": [ + 318, + 67, + 548, + 176 + ], + "lines": [ + { + "bbox": [ + 318, + 67, + 548, + 176 + ], + "spans": [ + { + "bbox": [ + 318, + 67, + 548, + 176 + ], + "type": "image", + "image_path": "33d6f4adc556afce73df5205366cfdbf662674600a9dffdec376d23e8e356225.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 176, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 314, + 176, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 314, + 176, + 553, + 188 + ], + "type": "text", + "content": "Figure 2. The impact of the number of queries and decoder layers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "text", + "content": "larger performance declines, with a drop of over " + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "inline_equation", + "content": "-2\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "text", + "content": " with TaskWeave [71] at 20 queries. Similarly, increasing the number of decoder layers caused a drop from " + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "inline_equation", + "content": "-1\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "inline_equation", + "content": "-2.5\\%" + }, + { + "bbox": [ + 313, + 200, + 555, + 285 + ], + "type": "text", + "content": ". In contrast, our Sim-DETR mitigates this phenomenon and achieves a slight improvement, demonstrating its robustness. This observation motivates an in-depth exploration of decoder layer, particularly the role of queries (see Sec 3.2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 291, + 554, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 304 + ], + "type": "text", + "content": "3.2. Similar Target Segments Cause Query Conflicts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 308, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 429 + ], + "type": "text", + "content": "Motivation. Initially, we hypothesize that the lack of performance improvement with increased queries is due to query redundancy. However, performance declined, suggesting that conflicts between queries, rather than redundancy, prevent compatibility with existing queries. Similarly, conflicts between queries across layers may explain the ineffectiveness of increasing the number of layers. This motivates us to explore (1) the relationships between queries to identify their potential conflicts and (2) their variations across layers to reveal the impact at different layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 654 + ], + "type": "text", + "content": "Setting. (1) To evaluate query relationships, we assess similarities between queries corresponding to the same segment (intra-segment similarity) and those from different segments (inter-segment similarity). Specifically, we identify the corresponding segment for each query based on the maximum IoU between the query's prediction and all ground-truth (GT) segments during inference. Queries whose IoU with the corresponding segment is 0.5 or greater are included in our analysis to ensure meaningful corresponding. We then compute feature similarities between queries associated with the same segment (intra) and those matched to different segments (inter). (2) To analyze variations across layers, we introduce the metric of cross-layer matching consistency. Specifically, during training optimization, each GT segment is uniquely matched to a query through bipartite matching at each decoding layer. We assess match consistency by measuring the proportion of segments that retain their corresponding queries across two consecutive decoder layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Result & Discussion. (1) As shown in Figure 3, CGDETR fails to distinguish query similarities between inter- or intra-segment groups, indicating that queries are not clearly grouped with their corresponding segments. Consequently, the same query oscillates between segments," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22762" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 60, + 289, + 170 + ], + "blocks": [ + { + "bbox": [ + 63, + 60, + 289, + 170 + ], + "lines": [ + { + "bbox": [ + 63, + 60, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 63, + 60, + 289, + 170 + ], + "type": "image", + "image_path": "12b42318272f8cd778de121aec07c3f67597410fffad46af97933919b8be2852.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 227 + ], + "type": "text", + "content": "Figure 3. Distribution of query similarity for intra-segment (blue) and inter-segment (red) pairs. Our method can effectively distinguish intra-segment and inter-segment queries, ensuring more stable query-segment associations and reducing conflicts in query assignments across different segments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 228, + 282, + 351 + ], + "blocks": [ + { + "bbox": [ + 67, + 228, + 282, + 351 + ], + "lines": [ + { + "bbox": [ + 67, + 228, + 282, + 351 + ], + "spans": [ + { + "bbox": [ + 67, + 228, + 282, + 351 + ], + "type": "image", + "image_path": "894e525cf266de01d9c4a5c26cfb550aa34da0aebdefda1a3efceba091c98937.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 352, + 295, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 385 + ], + "type": "text", + "content": "Figure 4. Measuring cross-layer matching consistency (y-axis, defined in Line-260) of segments across two consecutive decoding layers (x-axis). Our method achieves higher consistency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 507 + ], + "type": "text", + "content": "lacking stable and consistent corresponding. Figure 4 further supports our hypothesis, demonstrating that cross-layer matching consistency in CG-DETR is lower than ours. (2) We attribute this to multiple GT segments in a video sharing the same linguistic semantics. The high semantic similarity between GT segments results in high similarity between their corresponding queries. Figure 4 supports this, showing that the consistency for cases with multiple GT segments is lower than for those with a single segment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "type": "text", + "content": "3.3. Global Matching vs. Local Localization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "Motivation. We further question whether the inconsistency in matching is solely due to query conflicts from multiple GT segments. If so, why does the previous method still show roughly " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " mismatched queries in cases with a single target segment? Therefore, we shift our focus from conflicts between queries to conflicts within individual queries. A query serves dual roles: aligning with global linguistic semantics (global matching) and precisely localizing the segment, particularly its boundaries (local localization). This prompts us to explore whether conflicts arise from these roles within a single query." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Setting. (1) To evaluate global matching, we use the query's classification confidence score, as it reflects the confidence of the query being referenced by the global semantics of the linguistic description. (2) To evaluate local localization, we" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 201 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 201 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 201 + ], + "type": "image", + "image_path": "8e21c353c2a08149735c3362c9cf8f98e645ea476b4c4879008433b2895113fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 265 + ], + "type": "text", + "content": "Figure 5. Global matching score (defined in Line-290) vs. local localization score (defined in Line-293). Compared to CGDETR, our method concentrates a query's attention significantly more within a single GT segment, rather than dispersing it across multiple segments, leading to improved local localization scores." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 272, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 342 + ], + "type": "text", + "content": "extract attention scores between each query and frame from the cross-attention module in the last decoder layer, as they indicate query-based frame localization. We then compute the IoU score between this attention and the query's corresponding GT segment, reflecting the accuracy of the query's localization at the frame level." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "type": "text", + "content": "Result & Discussion. As shown in Figure 5, CG-DETR (marked in red) may yield low local localization scores despite high global matching scores, indicating that while queries align well with linguistic descriptions, they fail to achieve precise localization. Further analysis reveals that query attention may span frames in multiple GT segments (see Figure 5 \"Attention\"), leading to a low local score for the query's specific GT segment it corresponds to." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 427, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 427, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 427, + 459 + ], + "type": "text", + "content": "3.4. Positive Side Effect" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 465, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 537 + ], + "type": "text", + "content": "Convert Abnormal to Normal: Our Sim-DETR effectively (1) distinguishes queries within the same target segment and across different segments in Figure 3, (2) achieves consistent matching across decoder layers in Figure 4, and (3) aligns with global semantics while ensuring precise localization in Figure 5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 554, + 609 + ], + "type": "text", + "content": "Positive Side Effect: (1) More importantly, Sim-DETR maintains stable and robust performance with increased queries and decoder layers (see Figure 2). (2) By eliminating the abnormal phenomena, it also achieves significantly faster convergence compared to previous methods, as shown in Appendix C." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 621, + 370, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 370, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 370, + 633 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "We present a simple yet effective TSG baseline that resolves query conflicts in existing DETR-based TSG frameworks [22, 25, 27, 46, 47, 57, 67, 71] through two minor yet key modifications. First, we introduce the standard DETR-based [2] TSG architecture (Sec 4.1). Sec 4.2 introduces our first modification, which adjusts the self-attention" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22763" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 555, + 277 + ], + "type": "image", + "image_path": "4b4b581d8f656c715184403c618e5cfacbb7a4a6b3c9e602f4743fcad84baa5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 294, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 319 + ], + "type": "text", + "content": "Figure 6. Framework of the proposed Sim-DETR, featuring two simple yet powerful modifications: Query Grouping and Ranking to mitigate cross-query conflicts and Global-Local Bridging to address global-local conflicts within each query." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 411 + ], + "type": "text", + "content": "mechanism between queries in the decoder to avoid conflicts among them, based on the observations presented in Sec 3.2. Next, Sec 4.3 presents another improvement that addresses the conflict between global matching and local localization within the query (discussed in Sec 3.3) by introducing query-to-frame matching loss, which serves as the bridge between the global and local perspectives." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 272, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 272, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 272, + 431 + ], + "type": "text", + "content": "4.1. Standard DETR-based TSG Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "content": "In TSG, the goal is to identify video segments with temporal spans that correspond to a given language expression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": "Feature Extraction. Previous DETR-based works [25, 47, 57, 71] employ CLIP [49] and SlowFast [10] to extract video features, represented as " + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": " stands for the number of frames and the hidden dimension, respectively. For the language expression, typically a sentence with " + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": " tokens, the CLIP text encoder is used to obtain token-level features, resulting in " + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 55, + 461, + 296, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "text", + "content": "Multimodal Encoder. After feature extraction, DETR-based TSG methods [22, 46, 47, 57] employ a multimodal encoder to fuse these video and language features, producing multimodal video features, denoted as " + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{T}}\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "text", + "content": ". The key component of the multimodal encoder is a cross-attention mechanism [62], where video features " + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "text", + "content": " act as queries, while word features " + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 545, + 296, + 628 + ], + "type": "text", + "content": " serve as keys and values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Temporal Sentence Decoding. In the decoder, the model initializes a set of query to gather global information from the video, which is then decoded into local spans in the span prediction head. Specifically, given initialized queries " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " is the query number, and the multimodal video features " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", the decoder facilitates interaction between queries and gathers global information through interleaved" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 327, + 533, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 533, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 533, + 338 + ], + "type": "text", + "content": "self-attention and cross-attention modules, as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 348, + 553, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 348, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 348, + 553, + 465 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {k} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {k} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\mathcal {Q} _ {v} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {v} ^ {\\mathrm {S A}} (\\mathcal {Q}), \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\operatorname {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\operatorname {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} (\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right). \\end{array} \\right. \\\\ \\left\\{ \\begin{array}{l} \\mathcal {Q} _ {q} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {q} ^ {\\mathrm {C A}} (\\mathcal {Q} ^ {\\mathrm {S A}}), \\hat {\\mathcal {T}} _ {k} = \\mathrm {F C} _ {k} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}}), \\hat {\\mathcal {T}} _ {v} = \\mathrm {F C} _ {v} (\\hat {\\mathcal {T}}), \\\\ \\mathcal {Q} _ {o} ^ {\\mathrm {C A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {C A}} \\left(\\mathrm {s o f t m a x} \\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {C A}} (\\hat {\\mathcal {T}} _ {k}) ^ {\\intercal}}{\\sqrt {C}}\\right) \\hat {\\mathcal {T}} _ {v}\\right). \\end{array} \\right. \\\\ \\hat {\\mathcal {Q}} = \\operatorname {M L P} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathcal {Q} ^ {\\mathrm {C A}}\\right) + \\mathcal {Q}\\right), \\\\ B = \\mathcal {H} _ {\\text {s p a n}} (\\hat {\\mathcal {Q}}), \\quad B = \\left\\{b _ {1}, \\dots , b _ {M} \\right\\}, \\tag {1} \\\\ \\end{array}", + "image_path": "6d6b02591ef7b0d883f955abffef99233e8db3ee7e51626fe19fc9861ab29de6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{FC}(\\cdot)" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": " denotes the fully connected projection layer, with SA and CA representing self-attention and cross-attention, respectively. Subscripts " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": " indicate query, key, and value computations. MLP(" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": ") represents a nonlinear mapping with an activation function. " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{span}}(\\cdot)" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": " denotes the span prediction head, which takes the updated query features " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": " to predict the span set " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "inline_equation", + "content": "b_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 471, + 555, + 567 + ], + "type": "text", + "content": " represents the (start, end) of a predicted span." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 479, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 479, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 479, + 589 + ], + "type": "text", + "content": "4.2. Query Grouping and Ranking" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "In this section, we address the query conflicts identified in Sec 3.2: (1) the inability to distinguish between intra- and inter-segment queries (Figure 3), leading to query oscillation between segments without stable correspondence (Line-266), and (2) inconsistencies in segment-query matching across layers, where different queries may be matched to ground-truth (GT) segments at each layer for optimization, hindering effective learning (Figure 4)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To differentiate queries across segments, the challenge arises from GT segments that inherently share identical lin" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22764" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 76, + 282, + 160 + ], + "blocks": [ + { + "bbox": [ + 67, + 76, + 282, + 160 + ], + "lines": [ + { + "bbox": [ + 67, + 76, + 282, + 160 + ], + "spans": [ + { + "bbox": [ + 67, + 76, + 282, + 160 + ], + "type": "image", + "image_path": "4a450f95a57d73051b75d765db4ca14f10e9081b2460426e601ecea0dec4deb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 161, + 279, + 173 + ], + "lines": [ + { + "bbox": [ + 70, + 161, + 279, + 173 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 279, + 173 + ], + "type": "text", + "content": "Figure 7. Overview of our Query Grouping and Ranking." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "content": "guistic semantics, leading to highly similar queries. Therefore, a natural solution is to directly group queries, assigning each group to a specific GT segment for training and prediction. Instead of hard grouping with predefined rules or extra hyperparameters, we learn a soft grouping based on predicted temporal spans of queries, where closer spans indicate potential corresponding for the same segment. Specifically, for any two queries " + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 55, + 175, + 296, + 294 + ], + "type": "text", + "content": ", we compute their span distance, which is then used in Eq. 4 to learn their grouping:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 289, + 294, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 289, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 134, + 289, + 294, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i, j} ^ {\\text {i n t r a}} = \\| b _ {i} - b _ {j} \\| _ {2}, \\tag {2}", + "image_path": "259c8c9eb3f4acfebdfc2e259dfa635e056006b51e3578b34f83762e226938e0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "text", + "content": " denotes the L2 norm. Notably, we use L2 distance instead of the commonly used L1 because, when two spans are close (normalized distances " + }, + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "inline_equation", + "content": "\\leq 1" + }, + { + "bbox": [ + 55, + 308, + 296, + 367 + ], + "type": "text", + "content": "), L2 distance decreases more sharply than L1, imposing a smaller penalty on minor differences (see Appendix G)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 368, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 508 + ], + "type": "text", + "content": "Meanwhile, to enhance cross-layer matching consistency, we draw inspiration from EASE-DETR [14]'s query ranking strategy in object detection. By ranking queries within a segment at each layer, the leading query is prioritized for segment matching in subsequent layers. However, directly ranking queries by prediction scores, as in EASE-DETR, proves ineffective (see Appendix D). As analyzed in Sec. 3.3, high prediction/confidence scores do not guarantee precise predictions and may correspond to poor localization. Instead, we introduce an IoU head to assess localization accuracy and rank queries based on both factors, as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 506, + 294, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 506, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 67, + 506, + 294, + 537 + ], + "type": "interline_equation", + "content": "R _ {\\text {r a n k}} \\left(q _ {i}, q _ {j}\\right) = \\left\\{ \\begin{array}{l l} + 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} \\geq \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}} \\\\ - 1 & \\mathcal {P} _ {i} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {i} ^ {\\text {I o U}} < \\mathcal {P} _ {j} ^ {\\text {c l s}} \\circ \\mathcal {P} _ {j} ^ {\\text {I o U}}, \\end{array} \\right. \\tag {3}", + "image_path": "8bb615dac9f1d381d4e7a673e7f0ad6ff02e989094e46651f50eecb490b0c4ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " denote the query's confidence score and IoU prediction, respectively, with subscripts " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " indicating queries " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{rank}}(q_i,q_j) = 1" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " indicates that query " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " has a higher priority than query " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": ", considering both global alignment and local localization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 651 + ], + "type": "text", + "content": "Finally, we incorporate query grouping within segments and intra-segment ranking to infer relative relationships, then reshape query-to-query self-attention following [14] to mitigate conflicts, as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 654, + 294, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 654, + 294, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {S} ^ {\\text {a t t n}} = \\operatorname {s i g m o i d} \\left(\\mathrm {M L P} \\left(\\mathcal {S} ^ {\\text {i n t r a}} \\circ \\mathcal {R} _ {\\text {r a n k}}\\right), \\right. \\\\ \\mathcal {Q} ^ {\\mathrm {S A}} = \\mathrm {F C} _ {o} ^ {\\mathrm {S A}} \\left(\\text {s o f t m a x} \\left(\\left(\\frac {\\mathcal {Q} _ {q} ^ {\\mathrm {S A}} \\left(\\mathcal {Q} _ {k} ^ {\\mathrm {S A}}\\right) ^ {\\intercal}}{\\sqrt {C}}\\right) \\mathcal {S} ^ {\\mathrm {a t t n}}\\right) \\mathcal {Q} _ {v} ^ {\\mathrm {S A}}\\right) \\tag {4} \\\\ \\end{array}", + "image_path": "85d45d256c89822963f8cc14859c72bc83f93cd08e93e4875df086d40eb9fc1b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "S^{\\mathrm{attn}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " represents query relationships, with higher values promoting high-quality queries to dominate predictions by integrating similar queries within the same group." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 117, + 443, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 443, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 443, + 131 + ], + "type": "text", + "content": "4.3. Global-Local Bridging" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 135, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 555, + 278 + ], + "type": "text", + "content": "In this section, we address the query conflicts identified in Sec 3.3, specifically the conflict between a query's global semantic matching and local localization. Since queries overly rely on global semantics while neglecting local localization, our core idea is to reinforce fine-grained frame-level localization within the segment. Specifically, we introduce a query-to-frame matching mechanism and loss term to enhance localization beyond segment boundary prediction, ensuring alignment with in-segment frames while avoiding matches with those from other segments. The full sequence of frames within the segment, from start to finish, serves as a bridge linking global semantics to local boundaries." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": "During training, we compute the semantic similarity between each query " + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 279, + 554, + 350 + ], + "type": "text", + "content": " and all frames within its corresponding ground-truth span, maximizing similarity to ensure full alignment with in-segment frames while minimizing similarity to out-segment frames. The computation process is outlined as follows," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 353, + 555, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 353, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 331, + 353, + 555, + 401 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z = \\operatorname {s i g m o i d} (\\tau \\cdot \\cos (q _ {i}, \\hat {\\mathcal {T}})), \\\\ \\mathcal {L} _ {\\mathrm {b r i d g e}} = \\lambda_ {\\mathrm {b r i d g e}} \\frac {- \\sum_ {j} z _ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}{\\sum_ {j} z _ {j} \\left(1 - \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}\\right) + \\sum_ {j} \\mathbb {I} \\left[ b _ {i} ^ {\\mathrm {g t}} \\right] _ {j}}, \\tag {5} \\\\ \\end{array}", + "image_path": "fac170c3d5aa409d418661db0b503ce0379e87cf901d10ec4012f619c4622b80.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " donates query-to-frame similarity, calculated using the cosine similarity between the query " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " frame features " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\tilde{T}" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " defined in Line-347, and " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " is a learnable scaling coefficient. " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "b_i^{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " is the ground-truth span corresponding to query " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[b_i^{\\mathrm{gt}}] \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " is an indicator function that assigns 1 to frames within the ground truth and 0 otherwise, with " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " representing the frame index. Thus, the numerator in " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bridge}}" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "-\\sum_{j} z_j \\mathbb{I}[b_i^{\\mathrm{gt}}]_j" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": ", encourages the query to be similar to every frame within its corresponding segment. Meanwhile, the denominator's first term, " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\sum_{j} z_j (1 - \\mathbb{I}[b_i^{\\mathrm{gt}}]_j)" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": ", represents query " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": "'s similarity to out-segment frames, promoting its minimization. The denominator's second term, " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\sum_{j} \\mathbb{I}[b_i^{\\mathrm{gt}}]_j" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": ", represents the length of the ground-truth segment, serving as a normalization factor. And " + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bridge}}" + }, + { + "bbox": [ + 313, + 411, + 555, + 602 + ], + "type": "text", + "content": " is a hyperparameter that controls the loss weight." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 611, + 397, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 397, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 397, + 623 + ], + "type": "text", + "content": "4.4. Overall Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "The overall loss is: " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{MD}} + \\lambda_{\\mathrm{bridge}}\\mathcal{L}_{\\mathrm{bridge}} + \\lambda_{\\mathrm{iou}}\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MD}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " includes the L1, gIoU, classification, and saliency losses, as in Moment DETR [26]. " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " is the loss for the IoU head introduced in Line-406, designed to assist in ranking queries using " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i^{\\mathrm{IoU}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_j^{\\mathrm{IoU}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " in Eq 3. The IoU head implementation follows [25, 58]. Ablations for hyperparameters " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{bridge}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " are detailed in Appendix E." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22765" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 533, + 238 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 533, + 238 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 533, + 238 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 533, + 238 + ], + "type": "table", + "html": "
Methodtestval
R1mAPR1mAP
@0.5@0.7@0.5@0.75Avg.@0.5@0.7@0.5@0.75Avg.
M-DETR [27]NeurIPS'2152.8933.0254.8229.4030.7353.9434.84--32.20
UMT [40]CVPR'2256.2341.1853.8337.0136.1260.2644.26--38.59
QD-DETR [47]CVPR'2362.4044.9862.5239.8839.8662.6846.6662.2341.8241.22
UniVTG [32]ICCV'2358.8640.8657.6035.5935.4759.74---36.13
EaTR [22]ICCV'23-----61.3645.7961.8641.9141.74
MomentDiff [29]NeurIPS'2357.4239.6654.0235.7335.95-----
TR-DETR [57]AAAI'2464.6648.9663.9843.7342.6267.1051.4866.2746.4245.09
UVCOM [67]CVPR'2463.5547.4763.3742.6743.1865.1051.81--45.79
TaskWeave [71]CVPR'24-----64.2650.0665.3946.4745.38
BAM-DETR [25]ECCV'2462.7148.6464.5746.3345.3665.1051.6165.4148.5647.61
CG-DETR [46]Arxiv'2465.4348.3864.5142.7742.8667.3552.0665.5745.7344.93
SpikeMba [31]Arxiv'2464.1349.42-43.6743.7965.3251.33-44.9644.84
Sim-DETR(Ours)67.6450.9167.8147.5946.9369.4854.0669.7051.1149.50
", + "image_path": "00b8c3fc81e571754b6a029da9730fd32ce567cb58641bd41143254d1f61eeba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 259, + 526, + 378 + ], + "blocks": [ + { + "bbox": [ + 105, + 241, + 503, + 253 + ], + "lines": [ + { + "bbox": [ + 105, + 241, + 503, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 503, + 253 + ], + "type": "text", + "content": "Table 1. Experimental results on the QVHighlights benchmark, comparing ours with state-of-the-art methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 259, + 526, + 378 + ], + "lines": [ + { + "bbox": [ + 82, + 259, + 526, + 378 + ], + "spans": [ + { + "bbox": [ + 82, + 259, + 526, + 378 + ], + "type": "table", + "html": "
MethodTACoSCharades-STA
R1@0.3R1@0.5R1@0.7mIoUR1@0.3R1@0.5R1@0.7mIoU
2D-TAN [81]40.0127.9912.9227.2258.7646.0227.5041.25
M-DETR [27]37.9724.6711.9725.4965.8352.0730.5945.54
MomentDiff [29]44.7833.68---55.5732.42-
QD-DETR [47]-----57.3132.55-
UniVTG [32]51.4434.9717.3533.6070.8158.0135.6550.10
CG-DETR [46]52.2339.6122.2336.4870.4358.4436.3450.13
UVCOM [67]-36.3923.32--59.2536.64-
SpikeMba [31]51.9839.3422.8335.8171.2459.6536.1251.74
Sim-DETR(Ours)57.0642.7926.8239.4473.0961.3439.6252.56
", + "image_path": "e4da9c5e7fe8791a0571d909483dc2142c28ab46ab3b2875cf1a3e915448db4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 137, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 137, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 137, + 415 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 251, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 251, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 251, + 437 + ], + "type": "text", + "content": "5.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "content": "Datasets. Following [27, 31, 32, 47, 67], we conduct comprehensive evaluation experiments across three datasets: QVHighlights [27], Charades-STA [12], and TACoS [50]. The QVHighlights dataset contains 10,310 text expressions and 18,367 distinct events, with most expressions corresponding to multiple events. Charades-STA is derived from a collection of 9,848 indoor scene videos, encompassing 16,128 events. TACoS includes only 273 videos, with an average length approaching five minutes per video. For a fair comparison, we adopt the same evaluation metrics as in previous methods. On the QVHighlights benchmark, we use Recall and Mean Average Precision (mAP) as primary metrics, while for Charades-STA and TACoS, we employ Recall and Mean Intersection over Union (mIoU)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Implementation Details. We set the input video resolution to " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " and represent each video by concatenating the CLIP [49] [CLS] tokens with SlowFast [10] features. For the Charades-STA and TACoS datasets, we additionally constructed two experiments using I3D [3] and VGG [56] as visual encoders, with Glove [48] embeddings for text features. The model is trained for 200 epochs on a single NVIDIA A40 GPU, using the AdamW [42] optimizer with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 404, + 553, + 563 + ], + "blocks": [ + { + "bbox": [ + 167, + 381, + 441, + 392 + ], + "lines": [ + { + "bbox": [ + 167, + 381, + 441, + 392 + ], + "spans": [ + { + "bbox": [ + 167, + 381, + 441, + 392 + ], + "type": "text", + "content": "Table 2. Comparison results on the TACoS and Charades-STA benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 404, + 553, + 563 + ], + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 563 + ], + "type": "table", + "html": "
BackboneMethodR@0.5R@0.7
VGG [56]2D-TAN [81]40.9422.85
FVMR [11]42.3624.14
SSRN [83]46.7227.98
UMT [40]48.3129.25
MomentDiff [29]51.9428.25
QD-DETR [47]52.7731.13
TR-DETR [57]53.4730.81
CG-DETR [46]55.2234.19
I3D [3]Sim-DETR(Ours)55.9735.38
MAN [77]46.5322.72
VSLNet [78]47.3130.19
QD-DETR [47]50.6731.02
TR-DETR [57]55.5133.66
TaskWeave [71]53.3631.40
Sim-DETR(Ours)57.5535.73
", + "image_path": "631a2d318618f44d1e3ca3485136cc78eb3878a5c6267881672b5b8769289f72.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 567, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 322, + 567, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 322, + 567, + 545, + 578 + ], + "type": "text", + "content": "Table 3. Experimental results with VGG or I3D as backbone." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 581, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 617 + ], + "type": "text", + "content": "an initial learning rate of 1e-4. Note that, for a fair comparison, the total number of layers in our Sim-DETR is set to 6, the same as recent works [22, 46, 67]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 624, + 539, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 539, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 539, + 636 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 689 + ], + "type": "text", + "content": "As shown in Tables 1-3, we report detailed evaluation results across the QVHighlights, Charades-STA, and TACoS benchmarks. Our Sim-DETR consistently outperforms all state-of-the-art methods (SOTAs) across all metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Results on QVHighlights are shown in Table 1. Compared to the latest published work, BAM-DETR [25], our Sim" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22766" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 278, + 149 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 278, + 149 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 278, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 278, + 149 + ], + "type": "table", + "html": "
MethodR1mAP
@0.5@0.7@0.5@0.75Avg.
Baseline65.4850.8465.3445.8244.97
+ Liou66.5851.9465.6846.3045.22
+ QGR68.7752.2667.7248.2847.03
+ GLB67.1652.7768.5848.8848.17
+ QGR,GLB69.4854.0669.7051.1149.50
", + "image_path": "48108b3fd71e1a49e8dc368644797bf844529cbb1f5ca583a3b5514000fcecca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 169, + 279, + 261 + ], + "blocks": [ + { + "bbox": [ + 96, + 153, + 241, + 164 + ], + "lines": [ + { + "bbox": [ + 96, + 153, + 241, + 164 + ], + "spans": [ + { + "bbox": [ + 96, + 153, + 241, + 164 + ], + "type": "text", + "content": "Table 4. Ablation study on components." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 169, + 279, + 261 + ], + "lines": [ + { + "bbox": [ + 57, + 169, + 279, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 169, + 279, + 261 + ], + "type": "table", + "html": "
Inner RelevanceOuter GlobalOuter LocalmAP
(1)span border distconfidence predIoU pred49.50
(2)center distconfidence predIoU pred48.59
(3)span IoU48.94
(4)span border distw/oIoU pred48.93
(5)span border distconfidence predw/o48.66
(6)center pred48.97
", + "image_path": "9c29056dd4cef746c80674b6a3317852702f0d94bfd3cca3376298f179bba9d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 264, + 238, + 275 + ], + "lines": [ + { + "bbox": [ + 99, + 264, + 238, + 275 + ], + "spans": [ + { + "bbox": [ + 99, + 264, + 238, + 275 + ], + "type": "text", + "content": "Table 5. Impact of different measures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": "DETR outperforms it by an average of " + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "inline_equation", + "content": "+2.88\\%" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": " across all metrics. More importantly, it achieves remarkable gains of " + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "inline_equation", + "content": "+4.93\\%" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "inline_equation", + "content": "+2.27\\%" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": " in R1@0.5 and R1@0.7, respectively. In comparison TR-DETR [57], which uses standard DETR [2] as its basic framework and is used as our baseline model, our Sim-DETR achieves an average improvement of " + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "inline_equation", + "content": "+4.31\\%" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "inline_equation", + "content": "+4.41\\%" + }, + { + "bbox": [ + 55, + 281, + 295, + 472 + ], + "type": "text", + "content": " in mAP for test and validation, respectively. These improvements underscore the effectiveness of our Sim-DETR in fundamentally addressing conflicts within and between queries. We further conduct a comparison with SpikeMba [31], a contemporary method built on the Mamba [17] framework. Without any complex module designs, the consistent improvements across all metrics highlight the simplicity and effectiveness of our Sim-DETR. Sim-DETR establishes an efficient baseline for the task, significantly advancing research in TSG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "text", + "content": "Results on Charades-STA and TACoS are presented in Table 2, where we consistently outperform all methods on both datasets. Compared to the state-of-the-art methods, we achieve a " + }, + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "inline_equation", + "content": "+1.97\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "text", + "content": " improvement on the Charades-STA dataset (with SpikeMba as the SOTA) and a " + }, + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "inline_equation", + "content": "+3.89\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 617 + ], + "type": "text", + "content": " improvement on the TACoS dataset (with CG-DETR [46] as the SOTA). These enhancements further demonstrate the outstanding capability of our method in temporal localization. More importantly, they confirm that our Sim-DETR exhibits significant performance advantages across various scenarios, types, and lengths of videos, highlighting its robustness and generalizability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Results with Different Backbones. Following previous works, we conduct two additional experiments on the QVHighlights dataset. These experiments use VGG and I3D for visual feature extraction and use Glove [48]embedding as text features. Table 3 presents the detailed results of these two experiments, where we achieve state-of-the-art results in both cases. Notably, for the video feature extractor I3D, we surpass the latest pub" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "lished method by " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "+4.26\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ". These results further demonstrate the excellent generalization capabilities of our SimDETR across different backbones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 115, + 408, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 115, + 408, + 129 + ], + "spans": [ + { + "bbox": [ + 314, + 115, + 408, + 129 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": "Roadmap for building a simple yet effective TSG baseline is presented in Table 4. We employ TR-DETR without MR2HD module as our baseline, which achieves an mAP of " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "44.97\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": ". Since our Query Grouping and Ranking (QGR) utilizes " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": " proposed by [25, 58], we additionally conduct an ablation study to isolate the impact of " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": ". (1) QGR: To mitigate conflicts and optimization challenges arising from overly similar semantics between queries, we introduce the QGR, which improves the baseline by " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "+2.07\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": " on mAP, representing a subtle yet significant enhancement. (2) Global-Local Bridging (GLB): To address conflicts and ambiguities between global semantics and local localization within queries, we introduce GLB to enable seamless transformation through query-to-frame alignment. Experimental results show a performance boost to " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "48.17\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": " on mAP, indicating the effectiveness of GLB in bridging global-to-local transformations. However, without QGR to address semantic conflicts in temporal DETR, the recall predictably decreased by " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "-3.1\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": ". (3) Overall Framework: With these adjustments, our Sim-DETR achieves a remarkable SOTA performance: " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "49.50\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": " on mAP and " + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "inline_equation", + "content": "69.48\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 384 + ], + "type": "text", + "content": " on R1@0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": "The Effects of Different Conflict Metrics in QGR. Table 5 summarizes five experiments on conflict metrics in QGR. (1) The first row displays our final implementation, achieving " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "49.50\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " mAP. (2) & (3) We evaluate L2 center distance and span IoU for inter-query relationships. Center distance alone yields " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "48.59\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " mAP, highlighting its inadequacy in segment representation. Span IoU outperforms center distance but lags by " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "0.56\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " compared to boundary distance, likely due to IoU's broader tolerance, which reduces precision in query relevance. (4) To assess outer global semantic alignment, we remove global semantics, resulting in a " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "-0.57\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " drop, emphasizing their role in query interaction and quality. (5) Finally, we test local location metrics in query prioritization. Excluding the outer local metric significantly lowers performance " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "(-0.84\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " mAP), indicating that without local information, queries encounter conflicts due to overly generalized global semantics, leading to random matches. Consistent with (2), center localization accuracy shows similar limitations, with a " + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "inline_equation", + "content": "0.53\\%" + }, + { + "bbox": [ + 313, + 384, + 553, + 611 + ], + "type": "text", + "content": " decrease." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 621, + 387, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 387, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 387, + 634 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "content": "In this paper, we propose a concise and efficient temporal sentence grounding framework Sim-DETR, introducing two simple modifications to the decoder for resolving conflicts both between and within queries. Experimental results demonstrate that our approach significantly outperforms state-of-the-art methods across all benchmarks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22767" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No.62206174)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 138, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 138, + 297, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 138, + 297, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 138, + 297, + 192 + ], + "type": "text", + "content": "[1] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In Proceedings of the IEEE international conference on computer vision, pages 5803-5812, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 295 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 360 + ], + "type": "text", + "content": "[4] Chaoqi Chen, Yushuang Wu, Qiyuan Dai, Hong-Yu Zhou, Mutian Xu, Sibei Yang, Xiaoguang Han, and Yizhou Yu. A survey on graph neural networks and graph transformers in computer vision: A task-oriented perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "type": "text", + "content": "[5] Jingyuan Chen, Xinpeng Chen, Lin Ma, Zequn Jie, and Tat-Seng Chua. Temporally grounding natural sentence in video. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 162-171, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "type": "text", + "content": "[6] Long Chen, Chujie Lu, Siliang Tang, Jun Xiao, Dong Zhang, Chilie Tan, and Xiaolin Li. Rethinking the bottom-up framework for query-based video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10551-10558, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 509 + ], + "type": "text", + "content": "[7] Shaoxiang Chen and Yu-Gang Jiang. Semantic proposal for activity localization in videos via sentence query. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8199-8206, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 554 + ], + "type": "text", + "content": "[8] Qiyuan Dai and Sibei Yang. Curriculum point prompting for weakly-supervised referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13711-13722, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 609 + ], + "type": "text", + "content": "[9] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 612, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 296, + 656 + ], + "type": "text", + "content": "[10] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 689 + ], + "type": "text", + "content": "[11] Junyu Gao and Changsheng Xu. Fast video moment retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1523-1532, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 714 + ], + "type": "text", + "content": "[12] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "text", + "content": "[13] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. arXiv preprint arXiv:2101.07448, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[14] Yulu Gao, Yifan Sun, Xudong Ding, Chuyang Zhao, and Si Liu. Ease-detr: Easing the competition among object queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17282-17291, 2024. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 555, + 239 + ], + "type": "text", + "content": "[15] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In 2019 IEEE winter conference on applications of computer vision (WACV), pages 245-253. IEEE, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 240, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 553, + 282 + ], + "type": "text", + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander Hauptmann. Excl: Extractive clip localization using natural language descriptions. arXiv preprint arXiv:1904.02755, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "text", + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "type": "text", + "content": "[18] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "text", + "content": "[19] De-An Huang, Vignesh Ramanathan, Dhruv Mahajan, Lorenzo Torresani, Manohar Paluri, Li Fei-Fei, and Juan Carlos Niebles. What makes a video a video: Analyzing temporal information in video understanding models and datasets. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7366–7375, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "type": "text", + "content": "[20] Hanzhuo Huang, Yufan Feng, Cheng Shi, Lan Xu, Jingyi Yu, and Sibei Yang. Free-bloom: Zero-shot text-to-video generator with llm director and ldm animator. Advances in Neural Information Processing Systems, 36:26135-26158, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 525 + ], + "type": "text", + "content": "[21] Hanzhuo Huang, Yuan Liu, Ge Zheng, Jiepeng Wang, Zhiyang Dou, and Sibei Yang. Mvtokenflow: High-quality 4d content generation using multiview token flow. arXiv preprint arXiv:2502.11697, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "text", + "content": "[22] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[23] Bin Jiang, Xin Huang, Chao Yang, and Junsong Yuan. Cross-modal video moment retrieval with spatial and language-temporal attention. In Proceedings of the 2019 on international conference on multimedia retrieval, pages 217-225, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[24] Manjin Kim, Heeseung Kwon, Chunyu Wang, Suha Kwak, and Minsu Cho. Relational self-attention: What's missing in attention for video understanding. Advances in Neural Information Processing Systems, 34:8046–8059, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 714 + ], + "type": "text", + "content": "[25] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22768" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34: 11846-11858, 2021. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[27] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Neurips, pages 11846-11858, 2021. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[28] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[29] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In Neurips, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 315 + ], + "type": "text", + "content": "[30] Pandeng Li, Chen-Wei Xie, Liming Zhao, Hongtao Xie, Jiannan Ge, Yun Zheng, Deli Zhao, and Yongdong Zhang. Progressive spatio-temporal prototype matching for text-video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4100–4110, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[31] Wenrui Li, Xiaopeng Hong, Ruiqin Xiong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "text", + "content": "[32] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[33] Liang Lin, Pengxiang Yan, Xiaoqian Xu, Sibei Yang, Kun Zeng, and Guanbin Li. Structured attention network for referring image segmentation. IEEE Transactions on Multimedia, 24:1922-1932, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[34] Daizong Liu and Wei Hu. Skimming, locating, then perusing: A human-like framework for natural language video localization. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4536-4545, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "text", + "content": "[35] Daizong Liu, Xiaoye Qu, Jianfeng Dong, and Pan Zhou. Adaptive proposal generation network for temporal sentence localization in videos. arXiv preprint arXiv:2109.06398, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "text", + "content": "[36] Daizong Liu, Xiaoye Qu, Xing Di, Yu Cheng, Zichuan Xu, and Pan Zhou. Memory-guided semantic learning network for temporal sentence grounding. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1665-1673, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 646 + ], + "type": "text", + "content": "[37] Daizong Liu, Xiaoye Qu, and Wei Hu. Reducing the vision and language bias for temporal sentence grounding. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4092-4101, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[38] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In Proceedings of the 26th ACM international conference on Multimedia, pages 843-851, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[39] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[40] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[41] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. " + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "inline_equation", + "content": "r^2" + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "-tuning: Efficient image-to-video transfer learning for video temporal grounding. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 216 + ], + "type": "text", + "content": "[42] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 293 + ], + "type": "text", + "content": "[43] Chujie Lu, Long Chen, Chilie Tan, Xiaolin Li, and Jun Xiao. Debug: A dense bottom-up grounding approach for natural language video localization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5144-5153, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[44] Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. Generation-augmented retrieval for open-domain question answering. arXiv preprint arXiv:2009.08553, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 392 + ], + "type": "text", + "content": "[45] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3651-3660, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[46] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[47] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 483, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 535 + ], + "type": "text", + "content": "[48] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 601 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[50] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. Transactions of the Association for Computational Linguistics, 1:25-36, 2013. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[51] Byungseok Roh, JaeWoong Shin, Wuhyun Shin, and Saehoon Kim. Sparse detr: Efficient end-to-end object detection with learnable sparsity. arXiv preprint arXiv:2111.14330, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[52] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22769" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 15724-15734, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[53] Cheng Shi and Sibei Yang. Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932-2941, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "text", + "content": "[54] Cheng Shi, Yulin Zhang, Bin Yang, Jiajin Tang, Yuexin Ma, and Sibei Yang. Part2object: Hierarchical unsupervised 3d instance segmentation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 221 + ], + "type": "text", + "content": "[55] Cheng Shi, Yuchen Zhu, and Sibei Yang. Plain-det: A plain multi-dataset object detector. In European Conference on Computer Vision, pages 210-226. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 255 + ], + "type": "text", + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 312 + ], + "type": "text", + "content": "[57] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4998-5007, 2024. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 315, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 358 + ], + "type": "text", + "content": "[58] Xiaolong Sun, Liushuai Shi, Le Wang, Sanping Zhou, Kun Xia, Yabing Wang, and Gang Hua. Diversifying query: Region-guided transformer for temporal sentence grounding. arXiv preprint arXiv:2406.00143, 2024. 2, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "text", + "content": "[59] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris M Kitani. Rethinking transformer-based set prediction for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3611-3620, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 460 + ], + "type": "text", + "content": "[60] Jiajin Tang, Ge Zheng, Cheng Shi, and Sibei Yang. Contrastive grouping with transformer for referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23570-23580, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 462, + 294, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 506 + ], + "type": "text", + "content": "[61] Jiajin Tang, Ge Zheng, and Sibei Yang. Temporal collection and distribution for referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15466-15476, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 509, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 529 + ], + "type": "text", + "content": "[62] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 586 + ], + "type": "text", + "content": "[63] Tao Wang, Li Yuan, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. Pnp-detr: Towards efficient visual analysis with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4661-4670, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "text", + "content": "[64] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In Proceedings of the AAAI conference on artificial intelligence, pages 2567-2575, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[65] Sangmin Woo, Jinyoung Park, Inyong Koo, Sumin Lee, Minki Jeong, and Changick Kim. Explore-and-match: Bridging proposal-based and proposal-free with transformer for sentence grounding in videos. arXiv preprint arXiv:2201.10168, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[66] Shaoning Xiao, Long Chen, Songyang Zhang, Wei Ji, Jian Shao, Lu Ye, and Jun Xiao. Boundary proposal network for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "two-stage natural language video localization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2986-2994, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 173 + ], + "type": "text", + "content": "[67] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "text", + "content": "[68] Chen-Wei Xie, Siyang Sun, Xiong Xiong, Yun Zheng, Deli Zhao, and Jingren Zhou. Ra-clip: Retrieval augmented contrastive language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19265–19274, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 231, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 555, + 284 + ], + "type": "text", + "content": "[69] Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9062–9069, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[70] Vikas Yadav, Steven Bethard, and Mihai Surdeanu. Unsupervised alignment-based iterative evidence retrieval for multihop question answering. arXiv preprint arXiv:2005.01218, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 386 + ], + "type": "text", + "content": "[71] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18308-18318, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[72] Shuo Yang and Xinxiao Wu. Entity-aware and motion-aware transformers for language-driven action localization in videos. arXiv preprint arXiv:2205.05854, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[73] Sibei Yang, Meng Xia, Guanbin Li, Hong-Yu Zhou, and Yizhou Yu. Bottom-up shift and reasoning for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11266-11275, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[74] Zhuyu Yao, Jiangbo Ai, Boxun Li, and Chi Zhang. Efficient detr: improving end-to-end object detector with dense prior. arXiv preprint arXiv:2104.01318, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 553, + 555 + ], + "type": "text", + "content": "[75] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[76] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 677 + ], + "type": "text", + "content": "[77] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1247-1257, 2019. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[78] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. arXiv preprint arXiv:2004.13931, 2020. 2, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22770" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 430 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[79] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[80] Songyang Zhang, Jinsong Su, and Jiebo Luo. Exploiting temporal relationships in video moment localization with natural language. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1230-1238, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[81] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12870-12877, 2020. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[82] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. Advances in Neural Information Processing Systems, 36:5168-5191, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "type": "text", + "content": "[83] Jiahao Zhu, Daizong Liu, Pan Zhou, Xing Di, Yu Cheng, Song Yang, Wenzheng Xu, Zichuan Xu, Yao Wan, Lichao Sun, et al. Rethinking the video sampling and reasoning strategies for temporal sentence grounding. arXiv preprint arXiv:2301.00514, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[84] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "text", + "content": "[85] Yuchen Zhu, Cheng Shi, Dingyou Wang, Jiajin Tang, Zhengxuan Wei, Yu Wu, Guanbin Li, and Sibei Yang. Rethinking query-based transformer for continual image segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4595-4606, 2025. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22771" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_content_list.json b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc588fcf697b3dbe30c9da17b6584385d9619cdd --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_content_list.json @@ -0,0 +1,1501 @@ +[ + { + "type": "text", + "text": "SimMLM: A Simple Framework for Multi-modal Learning with Missing Modality", + "text_level": 1, + "bbox": [ + 132, + 128, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sijie Li Chen Chen† Jungong Han*† \nSchool of Computer Science, University of Sheffield, UK \n{sli256, chen.chen2, jungong.han}@sheffield.ac.uk", + "bbox": [ + 269, + 202, + 725, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose SimMLM, a simple yet powerful framework for multimodal learning with missing modalities. Unlike existing approaches that rely on sophisticated network architectures or complex data imputation techniques, SimMLM provides a generic and effective solution that can adapt to various missing modality scenarios with improved accuracy and robustness. Specifically, SimMLM consists of a generic Dynamic Mixture of Modality Experts (DMoME) architecture, featuring a dynamic, learnable gating mechanism that automatically adjusts each modality's contribution in both full and partial modality settings. A key innovation of SimMLM is the proposed More vs. Fewer (MoFe) ranking loss, which ensures that task accuracy improves or remains stable as more modalities are made available. This aligns the model with an intuitive principle: removing one or more modalities should not increase accuracy. We validate SimMLM on multimodal medical image segmentation (BraTS 2018) and multimodal classification (UPMC Food-101, avMNIST) tasks, where it consistently surpasses competitive methods, demonstrating superior accuracy, interpretability, robustness, and reliability across both complete and missing modality scenarios at test time. Code is available at https://github.com/LezJ/SimMLM.", + "bbox": [ + 91, + 324, + 482, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal learning aims to build models that can process and integrate information from multiple, complementary modalities such as visual, textual and auditory information, which has been found useful for various tasks such as image-text food classification [4, 15], multimodal medical image segmentation [5, 24] with improved model accuracy. However, in real-world scenarios, it is common for one or more modalities to be missing during deployment due to hardware failures, environmental conditions, or constraints", + "bbox": [ + 89, + 727, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in data collection, budget, and resources. Addressing this test-time missing modality problem is crucial for building robust multimodal models.", + "bbox": [ + 511, + 292, + 903, + 337 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One mainstream to address the missing modality problem is data imputation, which focuses on synthesizing the missing modalities from available ones [18, 44] or approximating intermediate missing features [10, 35, 36, 40]. However, ensuring the factuality and fidelity of synthetic data remains a major challenge, as issues such as poor data quality [16, 27], data hallucination [30], and vulnerability to adversarial corruption [6, 29] have been reported. Moreover, the computational overhead of generative imputation limits its practicality. In parallel, sophisticated frameworks have been developed, which aim to tackle the missing modality by leveraging advanced, complex network architectures [1, 35, 41], and harmonize or align early or late features from different modalities into a unified, shared space. However, such approaches often introduce substantial resource demands. While some are not adaptable to other tasks at all, others constrain the choice of network architectures or require additional modules, such as projection layers, to accommodate diverse or heterogeneous modalities.", + "bbox": [ + 511, + 339, + 903, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce SimMLM, a simple yet effective framework for multi-modal learning with missing modalities, without relying on specific neural network architectures or complex data imputation. First, to effectively handle varying modality availability, we propose a Dynamic Mixture of Modality Experts (DMoME) architecture, where modality-specific expert networks process available modalities, and a learnable gating network dynamically adjusts their contributions. This design ensures that the model remains effective in both full and partial modality settings while being adaptable to different neural architectures and task domains.", + "bbox": [ + 511, + 642, + 903, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To further improve the model's robustness and reliability against missing modalities, we introduce a novel loss function, which is designed to optimize the loss landscape around the region of missing data. Inspired by an intuitive principle that the model should not degrade in per", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 107, + 875, + 233, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†† These authors contributed equally as senior authors.", + "bbox": [ + 109, + 887, + 395, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24068", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "formance when modalities are increased, this loss function, named More vs. Fewer (MoFe) ranking loss, ensures that as more modalities become available, the model's accuracy improves or at least remains stable. The MoFe loss encourages the model to compare performance between modality-rich and modality-poor pairs, promoting an inherent ranking that favors configurations with greater modality availability. This structured comparison in together with the dynamic weighting scheme in DMoME allows the model to learn consistent and reliable multimodal fusion strategies, improving robustness and adaptability across varying modality availability.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are summarized as follows:", + "bbox": [ + 109, + 273, + 434, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Dynamic Mixture of Modality Experts Architecture: SimMLM introduces the DMoME architecture with a learnable gating mechanism that adaptively adjusts modality contributions based on availability, improving performance in both complete and partial modality settings. This design enables seamless integration with various neural networks, ensuring flexibility and scalability across diverse tasks.", + "- More vs. Fewer ranking loss: We introduce the novel MoFe ranking loss, which ensures model performance improves or remains stable as additional modalities are incorporated. By enforcing a ranking comparison between modality-rich and modality-poor pairs, MoFe strengthens model robustness and reliability.", + "- Comprehensive Evaluation: We validate SimMLM on multimodal segmentation and classification across three datasets, demonstrating superior accuracy, interpretability, robustness, reliability, and computational efficiency over state-of-the-art methods. These results establish SimMLM as a practical and versatile solution for multimodal learning under missing modality conditions." + ], + "bbox": [ + 91, + 287, + 482, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 618, + 232, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve multi-modal learning with missing modalities, various approaches have been proposed, including a) input-level imputation, b) advanced representation learning; and c) mixture of expert strategies [37].", + "bbox": [ + 89, + 643, + 482, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Input-level imputation approaches leverage advanced generative networks such as auto-encoders [18, 20], generative adversarial networks [20] to reconstruct missing inputs. Such an approach however, can generate noisy and unreliable input, as discussed in the previous section.", + "bbox": [ + 89, + 704, + 482, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advanced representation learning focuses on learning improved representation either through advanced networks such as transformers [21, 41, 42], represented by mm-Former [41], or disentangling the latent features into modality-specific and modality-shared features for improving model accuracy, such as Robust-Mseg [1] and ShaSpec [35]. For instance, ShaSpec [35] employs a combination of modality-shared and modality-specific encoders", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with dedicated losses to promote feature disentanglement and perform feature fusion in the latent space before passing them to a task decoder. When a modality is missing, ShaSpec generates its features from the available modalities through a missing modality feature generation process enabled by the modality-shared encoder. However, this approach imposes a strict constraint that the shared encoder should be suitable for all input modalities and both shared and specific features must have the same dimensions, which can be less efficient and optimal when dealing with highly heterogeneous modalities.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture of Experts (MoE) has recently been adopted in both unimodal and multimodal learning, demonstrating superior performance in natural language processing [3, 14, 43] and computer vision tasks [28, 38, 39], while improving efficiency by reducing FLOPS. A very recent method for handling missing modalities is Mixture of Modality Knowledge Experts (MoMKE) [38], which uses MLP experts to generate unimodal and joint representations. Even when a modality is missing, MoMKE leverages existing modalities to generate joint representations by utilizing the other experts. While this approach has achieved promising results, it still incurs high computational costs, as each modality must pass through all modality experts, quadratically increasing complexity with the number of modalities. In contrast, the expert model in DMoME is dedicated to a specific modality. Additionally, unlike MoMKE, DMoME does not constrain the dimensionality of latent feature representations across modalities. Instead, it performs the weighting directly at the output space across experts. This flexibility makes DMoME more efficient and scalable, offering high performance in diverse multimodal tasks.", + "bbox": [ + 511, + 258, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 590, + 648, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 511, + 616, + 694, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assume we have a complete multimodal dataset with $M$ modalities, denoted as $\\mathcal{D} = \\{\\{\\mathbf{x}_i^m\\}_{m=1}^M, \\mathbf{y}_i\\}_{i=1}^N$ . where $\\mathbf{x}_i^m \\in \\mathcal{X}$ represents the $i$ -th data sample and $m$ indexes the modality. $\\mathbf{y}_i \\in \\mathcal{Y}$ is the label for the sample $i$ from the label space $\\mathcal{Y}$ , which can either be a classification label or segmentation map. In this paper, we are interested in finding a solution $f$ that learns a mapping from the input space to the label space: $f: \\mathcal{X} \\to \\mathcal{Y}$ , leveraging this complete modality dataset, so that $f$ can work robustly and reliably even when certain modalities $m$ are missing at inference time. For simplicity, we omit the sample index $i$ when not specified in the following sections.", + "bbox": [ + 511, + 638, + 903, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Overview of SimMLM", + "text_level": 1, + "bbox": [ + 511, + 832, + 723, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed SimMLM framework consists of two main components: a) a unified Dynamic Mixture of Modality Experts framework (DMoME) (Sec. 3.3), which can take a", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24069", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c4217bb8ce1735c44a6a9b7f4f45ea9f262483f95132e8c717a95a677284c699.jpg", + "image_caption": [ + "(a) Our DMoME" + ], + "image_footnote": [], + "bbox": [ + 151, + 89, + 442, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef5a9060e9ee5cc89056730ebd33f3dddb9216d85afc909c9de40a5f67fa5726.jpg", + "image_caption": [ + "(b) Computation flows of $\\mathcal{L}_{MoFe}$", + "Figure 1. Our proposed SimMLM consists of a) DMoME architecture: DMoME takes either partial or full modality inputs where each modality is processed by a corresponding expert network to produce its own output $(\\mathbf{o}^1, \\ldots, \\mathbf{o}^M)$ . A gating network (G) assesses the same inputs and produces weights to dynamically weigh the contributions of each expert's output in order to get the mixture of output $\\mathbf{o}$ ; b) $\\mathcal{L}_{\\mathrm{MoFe}}$ loss: At training time, in addition to standard task specific losses, DMoME is optimized with the More $(X^{+})$ vs. Fewer Modality $(X^{-})$ (MoFe) ranking loss, directly encouraging the network to distinguish between inputs with differing modality availability at the task performance level." + ], + "image_footnote": [], + "bbox": [ + 496, + 104, + 844, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "varied number of modalities as input and adaptively weight their contribution for robust performance at both training and inference time; b) a novel ranking loss called MoFe (Sec. 3.4), which is designed to calibrate the network's performance by differentiating network performance when the number of available modalities for the same sample varies.", + "bbox": [ + 89, + 446, + 482, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Dynamic Mixture of Modality Expert (DMoME)", + "text_level": 1, + "bbox": [ + 89, + 546, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 1 (a), Our DMoME consists of a set of modality-specific expert networks $\\{E^{m}(\\mathbf{x}_{m};\\theta_{m})\\}_{m = 1}^{M}$ each containing its own set of learnable parameters $\\theta_{m}$ and a gating network, $G(\\cdot ;\\phi)$ , with a set of learnable parameters $\\phi$ . Each expert network, $E^{m}(\\mathbf{x}_{m};\\theta_{m})$ , takes its corresponding modality as input and produces its output in logits $\\mathbf{o}^m$ , before applying the softmax function: $\\mathbf{o}^m = E^m (\\mathbf{x}^m;\\theta^m)\\in \\mathbb{R}^T$ . Here, $T$ denotes the dimensionality of the output space for a specific task, e.g., the number of predicted classes in a classification task.", + "bbox": [ + 89, + 583, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The gating network $G(\\cdot; \\phi)$ takes a set of different modalities as input and produces a set of gating values $\\{g^m\\}_{m=1}^M$ to calibrate different modality experts' output. The softmax function is then applied to computing non-negative weights $\\{w_1, w_2, \\ldots, w_M\\} \\in \\mathbb{R}$ , ensuring that the sum of the weights for existing modalities is equal to 1. The final output can be described as a reweighted combination of the outputs from the corresponding modality experts $\\{\\mathbf{o}^m\\}_{m=1}^M$ : $\\mathbf{o} = \\sum_{m=1}^M w^m E^m(\\mathbf{x}^m; \\theta_m)$ .", + "bbox": [ + 89, + 734, + 482, + 871 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, the input to the gating network can be an incomplete set of modalities. If certain modalities are miss", + "bbox": [ + 89, + 871, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing, those inputs are treated as 0, and the gating values for the corresponding missing modality experts' outputs are set to $-\\infty$ (which causes the corresponding weight values to be 0 after the softmax). Mathematically, the gating outputs are set as:", + "bbox": [ + 511, + 446, + 906, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {m} = \\left\\{ \\begin{array}{l l} G (\\mathbf {x}; \\phi) ^ {(m)}, & \\text {i f m o d a l i t y m i s p r e s e n t} \\\\ - \\infty , & \\text {i f m o d a l i t y m i s m i s s i n g} \\end{array} , \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 531, + 905, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ can be a complete set of modalities or an arbitrary incomplete set of modalities for a given sample. The final weights therefore can be simply computed as:", + "bbox": [ + 511, + 582, + 905, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw ^ {m} = \\frac {\\exp \\left(g ^ {m}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\exp \\left(g ^ {m ^ {\\prime}}\\right)} \\quad \\text {f o r} \\quad m = 1, 2, \\dots , M. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 626, + 905, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the weight vector $\\mathbf{w} = [w^{1}, w^{2}, \\ldots, w^{M}] \\in \\mathbb{R}^{M}$ is non-negative and sums to 1, with $w^{m} = 0$ for missing modalities, effectively excluding them from the final mixture. The above configuration can be easily extended to multi-task settings, where each task has its own task-specific weighting scheme. In this case, the gating network produces $\\mathbf{w} \\in \\mathbb{R}^{M \\times K}$ , with $K$ being the number of tasks.", + "bbox": [ + 511, + 661, + 905, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two-stage training", + "text_level": 1, + "bbox": [ + 513, + 776, + 650, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train the whole network, we adopt a two-stage training process, which consists of:", + "bbox": [ + 511, + 795, + 905, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. independent learning: we independently train a modality-specific expert $E_{m}$ for each modality $m$ , supervised by a task-specific loss $\\ell_{\\mathrm{task}}(\\mathbf{o}^{m},\\mathbf{y})$ . For example, for classification tasks, the loss function can be the cross-entropy (CE) loss, whereas for the segmentation,", + "bbox": [ + 514, + 825, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24070", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the loss can be Dice loss. This pretraining stage ensures that each modality expert independently learns specialized features, minimizing the risk of any single modality dominating the learning process. Different modalities can vary significantly in learning dynamics and task difficulty, which can lead to training instability, especially at the early stages of joint training.", + "bbox": [ + 107, + 90, + 482, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Cooperative learning: we co-train the expert networks alongside the gating network, ranking network performance under More vs. Fewer Modality conditions, see next section for details.", + "bbox": [ + 89, + 196, + 482, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This training scheme enables the expert networks to exploit modality-specific information without interference from other modalities before jointly optimizing with the gating network to exploit mutual benefits. It also improves framework flexibility with reduced training resources, e.g., a) first-stage training can be accelerated via parallel computing; b) when new modalities are introduced, existing pretrained expert networks can be reused for the second stage.", + "bbox": [ + 89, + 258, + 483, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. More vs. Fewer Modality Co-training", + "text_level": 1, + "bbox": [ + 89, + 392, + 416, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To boost model robustness against missing modalities, we simulate scenarios with missing modalities by randomly dropping out input modalities during training, which is a common practice. Yet, unlike existing approaches that focus on aligning outputs from missing modalities with those from the full set of modalities [7, 9, 17], we encourage the network to distinguish between modality-rich and modality-poor inputs.", + "bbox": [ + 89, + 414, + 483, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rational behind More vs. Fewer Modality Training: In many multi-modal learning tasks, all modalities are assumed to be predictive and complementary, contributing valuable information to the tasks that are of interest. Our approach is inspired by the intuition that richer multimodal inputs should yield more accurate results, as each modality complements the others. As the number of modalities decreases, the input becomes more limited, making it more challenging to capture complex patterns.", + "bbox": [ + 89, + 536, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As it is difficult to quantify the exact performance for different set-ups, we propose to formulate it as a ranking loss to encourage the network to compare the two scenarios. In cognitive science, it has been found that learners who engage in comparative tasks show an improved ability to transfer knowledge to new situations compared to those who learn through isolated examples [32]. Such a concept has also been applied in reinforcement learning for preference alignment for the applications like large language model training and finetuning with a reward function.", + "bbox": [ + 89, + 672, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our multimodal learning setting, at each training iteration, we sample two sets from a data point $\\mathbf{x}_{full}$ : $\\mathbf{x}^{+}$ and $\\mathbf{x}^{-}$ , where $\\mathbf{x}^{-} \\subset \\mathbf{x}^{+} \\subseteq \\mathbf{x}^{full}$ . The corresponding outputs from DMoME are denoted as $\\mathbf{o}^{+}$ and $\\mathbf{o}^{-}$ , respectively. We conjecture that the task-specific loss value for modality-", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rich setups with $\\mathbf{x}^{+}$ should be at least no higher than those in the modality $\\mathbf{x}^{-}$ . Mathematically, every pair of samples $(\\mathbf{x}^{+},\\mathbf{x}^{-})$ should satisfy the following relationship:", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) \\geq \\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) \\iff A c c \\left(\\mathbf {x} ^ {+}\\right) \\geq A c c \\left(\\mathbf {x} ^ {-}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 143, + 908, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition of $\\mathcal{L}_{MoFe}$ : To guide the network to learn the ordinary relationship, we introduce More vs. Fewer (MoFe) ranking loss, defined as follows:", + "bbox": [ + 511, + 167, + 906, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {M o F e}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\max \\left(0, \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) - \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 219, + 903, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which encourages the network to perform better when more modalities are available. It is worth mentioning that the proposed loss is general and thus can be applied to other existing multi-modal learning frameworks to better calibrate the network for improved robustness against missing modalities as well as improved reliability. The total training loss is thus defined as:", + "bbox": [ + 511, + 252, + 906, + 356 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {t o t a l}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) + \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) + \\lambda \\ell_ {\\mathrm {M o F e}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 364, + 903, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a coefficient to control the contribution of $\\ell_{\\mathrm{MOFE}}$ .", + "bbox": [ + 511, + 396, + 903, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We acknowledge the existence of prior research, such as the works of Noh et al. (2023) [26], and Ma et al. (2023) [19], which have investigated the application of ranking loss at the confidence levels between a pair of two inputs. However, our approach diverges significantly from these existing methods. Specifically, we propose the novel concept of applying ranking loss directly at the loss level, as opposed to the confidence level. This distinction is crucial, as our method aims to directly regularize the geometry of the loss landscape. In our experiments, we also showed extra benefit of our loss, which leads to reduced calibration errors, see Table. 4. Besides, our proposed method enables seamless integration with a range of loss functions, offering flexibility beyond classification and allowing for improved applicability in complex, diverse tasks. For example, our method can be applied for segmentation and regression tasks where the confidence scores are not straightforward in dense or continuous prediction settings. By focusing on the actual task error, our loss-level approach directly aligns with the downstream task performance, providing a more consistent measure across different tasks.", + "bbox": [ + 511, + 411, + 906, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 739, + 645, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate our approach on two types of tasks, multimodal medical segmentation and multimodal classification, spanning three datasets. For segmentation, we use BraTS 2018 [22] benchmark. For classification, we conduct experiments on UPMC Food-101 [4] and avMNIST dataset [34].", + "bbox": [ + 511, + 765, + 905, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 848, + 614, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BraTS 2018 (Brain Tumor Segmentation Challenge [22]) is a widely used benchmark for brain tumor segmenta", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24071", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion in 3D MRI scans. It has also been extensively employed to evaluate methods for handling missing modalities [7, 35, 41, 42]. The dataset includes four distinct MRI modalities: FLAIR, T1, T1 contrast-enhanced (T1ce), and T2. The segmentation task involves identifying three tumor subregions: the enhancing tumor (ET), the tumor core (TC), and the whole tumor (WT). The dataset consists of 285 manually labeled cases and 66 unlabeled cases for evaluation. In our experiments, we split the labeled cases into a 4:1 ratio for training and validation. Final evaluations are conducted on the 66 unlabeled cases via the official benchmarking platform for a consistent and fair accuracy comparison against existing methods [13, 41].", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UPMC Food-101 [4] is a large-scale food classification dataset with 90,704 image-text pairs spanning 101 classes. Collected from uncontrolled environments, it introduces inherent noise, posing challenges for accurate classification.", + "bbox": [ + 89, + 287, + 480, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "avMNIST (Audiovision MNIST [34]) is a digit classification dataset comprising 1,500 image-audio pairs. The visual modality consists of $28 \\times 28$ grayscale images of handwritten digits (0-9) from MNIST [12], while the audio modality is sourced from the Free Spoken Digits Dataset*.", + "bbox": [ + 89, + 347, + 480, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 433, + 307, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All the experiments were conducted on a single A100 GPU node. Detailed implementations are described below.", + "bbox": [ + 89, + 455, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BraTS 2018: We employed nnUNet [8], a widely used and powerful medical image segmentation backbone, as the expert network for each modality, which can be viewed as a 3D U-Net. The gating network $G$ in DMoME utilized lightweight CNNs to capture broad contextual information, followed by a linear layer generating $4 \\times 3$ weights for the four modalities and three segmentation tasks. We optimized both training stages using the Adam optimizer [11] ( $lr = 0.01$ ). The $\\mathcal{L}_{task}$ in Eq.4 combined Dice [23] and binary cross-entropy losses.", + "bbox": [ + 89, + 484, + 482, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UPMC Food-101: We employed Inception V3 [31] as the image expert and BERT [2] as the text expert, following [4]. To avoid redundant feature extraction, intermediate features from both were fed into an MLP gating network $G$ , which generated 2D weights for expert outputs. Optimization was performed using Adam ( $lr = 0.0001$ ) with cross-entropy loss as $\\mathcal{L}_{task}$ .", + "bbox": [ + 89, + 637, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "avMNIST: Following [20, 35], we adopted a LeNet-5 [12] backbone for both image and audio modalities, with audio converted into $20 \\times 20$ MFCC spectrograms [33]. The gating network $G$ followed a similar design as in BraTS 2018. Two-stage training was conducted using Adam ( $lr = 0.001$ ) with cross-entropy loss as $\\mathcal{L}_{task}$ .", + "bbox": [ + 89, + 741, + 482, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detailed gating network $G$ architectures for BraTS 2018 and avMNIST are provided in the Supplementary. For task accuracy evaluation, we follow common practice, using the", + "bbox": [ + 89, + 832, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dice score for segmentation (BraTS 2018) and Accuracy for classification (UPMC Food-101, avMNIST). The MoFe loss coefficient $\\lambda$ in Eq.4 is set to 0.1 across tasks; however, our method remains robust to variations in $\\lambda$ , as demonstrated in Sec. 5 Analysis.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Segmentation Results", + "text_level": 1, + "bbox": [ + 511, + 178, + 714, + 194 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarking segmentation accuracy on the official BraTS 2018 evaluation set. For the BraTS 2018 benchmark, we compare our approach against several state-of-the-art methods, including Robust-MSeg (RbSeg) [1], mmFormer (mmFm) [41], ShaSpec [35], and MoMKE [38] under 15 configurations of both missing and full modality setups in Table 1. Results show that SimMLM achieves the highest average accuracy in terms of the Dice scores, consistently ranking as the top or second-highest performer in each entry across 44 out of all 45 configurations (15 settings $\\times$ 3 subtasks), despite utilizing only $80\\%$ of the labeled data compared to the $100\\%$ used by ShaSpec and mmFormer. This consistent performance highlights our model's robustness and data efficiency in handling various incomplete modality inputs. Notably, when only the T1ce modality is available (Table 1, row 3), our model significantly improves the Dice score compared to the second-best methods, increasing from 73.29 to 78.58 for ET, 78.71 to 83.45 for TC, and 74.00 to 80.02 for WT. Such results underscore the advantage of our approach even in single-modality settings.", + "bbox": [ + 511, + 200, + 903, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, in both full and partial modality scenarios, SimMLM (DMoME with $\\mathcal{L}_{MoFe}$ ) achieves more accurate segmentation results with better model reliability compared to DMoME alone (A vs. B, C vs. D) thanks to the contribution of $\\mathcal{L}_{MoFe}$ . When T1ce is unavailable, DMoME with $\\mathcal{L}_{MoFe}$ can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce), which is aligned with clinical practice despite the network is not explicitly taught. Without MoFe loss, the network is biased to T2 (maybe due to higher intensity) and over-segment the tumor. In the following Sec. 5, we will provide a more detailed quantitative analysis for our method.", + "bbox": [ + 511, + 503, + 903, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Classification Results", + "text_level": 1, + "bbox": [ + 511, + 710, + 712, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarking classification accuracy on UPMC Food101 and avMNIST. For the classification tasks, we compared our model to the two most competitive methods: ShaSpec [35] and MoMKE [38]. To ensure fair comparisons, all models were trained under identical conditions. The results, presented in Table 2, report accuracy across different modality settings. Our method consistently surpasses SOTA methods across all settings, demonstrating its effectiveness in handling both missing and full modality scenarios. While clear improvements are observed on both datasets, our method achieves even greater gains on the", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "*Dataset link: https://github.com/Jakobovski/free-spoken-digit-dataset.", + "bbox": [ + 107, + 886, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24072", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/02c572e0df9a51b05f02259326cc6f537b42297455160933482a83083adf6fbc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsAvailable Modalities at Test TimeEnhancing tumor (ET)Tumor core (TC)Whole tumor (WT)
◇◇◇RbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]Ours
Missing1●○○25.6939.3343.5250.5351.3153.5761.2169.4475.0574.1285.6986.1088.6889.7889.24
2○●○17.2932.5341.0048.1248.4147.9056.5563.1868.6671.4870.1167.5273.4477.4580.28
3○○●67.0772.6073.2972.3078.5876.8375.4178.6578.7183.4573.3172.2273.8274.0080.02
4○○○28.9743.0546.3151.0250.6057.4964.2069.0372.3473.4982.2481.1583.9985.8286.01
5●●○32.1342.9644.7652.5253.6260.6865.9172.6775.0077.1488.2487.0689.7690.2389.98
6●○●70.3075.0775.6078.7580.2580.6277.8884.5084.6284.9688.5187.3090.0690.3490.35
7●○○33.8447.5247.2054.1254.6261.1669.7572.9376.9075.7688.2887.5990.0290.5390.28
8○●●69.0674.0475.7676.7878.7778.7278.5982.1081.7084.7977.1874.4278.7479.8682.25
9○●○32.0144.9946.8453.7453.4062.1969.4271.3873.7374.7384.7882.2086.0386.6087.42
10○○●69.7174.5175.9577.1178.6180.2078.6183.8283.6184.1985.1982.9985.4286.9288.11
11●●●70.7875.4776.4278.1980.4181.0679.8085.2382.8385.2588.7387.3390.2990.2190.63
12●●○36.4147.7046.5555.9355.8464.3871.5273.9777.0078.0288.8187.7590.3690.5990.62
13●○●70.8875.6775.9978.4480.2980.7279.5585.2683.5985.2389.2788.1490.7890.8590.99
14○●●70.1074.7576.3778.2780.3380.3380.3984.1882.3585.0886.0182.7186.4786.6287.78
Full15●●●71.1377.6178.0877.6580.4080.8685.7885.4582.6185.3989.4589.6490.8890.5491.07
Average51.0259.8561.5865.5667.1669.7872.9777.4578.5880.2084.3982.9485.9286.6987.67
", + "bbox": [ + 153, + 88, + 849, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Model accuracy (Dice score ↑ for the segmentation task, normalized to 100%) comparison of different methods for segmenting brain tumor subregions under various missing (○) and/or full modality settings. The best and second best performance of each subtask for each setting are highlighted in bold and underline respectively.", + "bbox": [ + 88, + 368, + 906, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d289ce2011664d25e89a2713d2558cf8757c63c5039de3c1496049502d016106.jpg", + "image_caption": [ + "Figure 2. Visualization of modality weight changes (bold numbers in the upper right corner) assigned by the gating network alongside the predicted ET segmentation and uncertainty (entropy) maps when full or partial modalities (T1ce is missing) are present. DMoME with $\\mathcal{L}_{MoFe}$ can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce) when T1ce is missing, which is aligned with clinical practice despite the network is not explicitly taught." + ], + "image_footnote": [], + "bbox": [ + 163, + 425, + 836, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "noisy UPMC Food-101 dataset, where classification is more challenging, underscoring the robustness of our SimMLM.", + "bbox": [ + 89, + 797, + 480, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 844, + 187, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of DMoME: To assess the unique contribution of DMoME, we trained the network w/o MoFe", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "loss: $\\mathcal{L}_{MoFe}$ and compared its performance with competitive models—mmFormer [41], ShaSpec [35], and MoMKE [38]. Experiments were conducted on the BraTS 2018 dataset, encompassing 15 different configurations for both missing-modality and full-modality setups. Table 3 reports the average Dice scores across 15 scenarios, alongside", + "bbox": [ + 511, + 797, + 906, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24073", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a17afc53a629f95fc46a65824e17241003bad66a11657853141d34574506cfe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModalitiesAccuracy (%) ↑
Modality 1Modality 2ShaSpec [35]MoMKE [38]SimMLM (Ours)
UPMC Food-101Image○69.2270.4672.20
○Text86.5586.5987.2
ImageText92.7392.7194.99
Average82.8383.2584.81
avMNISTImage○91.9092.6192.69
○Audio89.2891.1691.61
ImageAudio98.7198.6999.27
Average93.3094.1594.52
", + "bbox": [ + 109, + 88, + 467, + 204 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b2bcf25c785f3874312e4cc3ade489e58bd3537240b6536fcd08582dbcc931e.jpg", + "table_caption": [ + "Table 2. Accuracy $(\\%)$ ↑ comparison across missing & full modality settings on UPMC Food-101 and avMNIST datasets. The datasets have different modality pairs (Image-Text for UPMC Food-101, Image-Audio for avMNIST). o represents modality missingness." + ], + "table_footnote": [], + "table_body": "
Dice score (%) ↑Model efficiency
MethodETTCWT#ParamsFLOPS
mmFormer [41]59.8572.9782.94106M748G
ShaSpec [35]61.5877.4585.92187.7M713G
MoMKE [38]65.5678.5886.697.8M490G
DMOE w/o LMoFe66.0579.1486.717.8M123G
DMOE w LMoFe (SimMLM)67.1680.2087.677.8M123G
", + "bbox": [ + 109, + 300, + 465, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of model accuracy (Dice score $\\uparrow$ , normalized to $100\\%$ , averaged across all missing & full modality settings) and efficiency (# of parameters and FLOPS) on BraTS 2018 official-evaluated benchmark.", + "bbox": [ + 89, + 398, + 483, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "networks' parameters and floating-point operations per second (FLOPS). As a result, DMoME achieves the highest Dice scores across all tumor subregions, with particularly strong performance in the more challenging enhancing tumor (ET) and tumor core (TC) segmentation. Importantly, DMoME achieves a fourfold reduction in FLOPS compared to MoMKE, even with the same expert architecture, and requires over 13 times fewer parameters than mmFormer and ShaSpec, underscoring its efficiency and scalability. This reduction in FLOPS is due to a key difference in processing: unlike MoMKE, which feeds each modality's input to all experts [38], our DMoME processes each input with only its dedicated expert. This simple approach significantly reduces FLOPS without compromising representational quality. Further details on DMoME's micro design are provided in the Supplementary.", + "bbox": [ + 88, + 484, + 482, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretability of DMoME: Importantly, our network directly provides case-specific interpretability by dynamically revealing the importance of each modality, as shown in Fig. 2. In contrast, previous works such as mmFormer [41], ShaSpec [35], and MoMKE [38] lack this capability. We further performed population-wise study by plotting the average gating weights under different full- and missing-modality configurations on the official BraTS 2018 evaluation set in Fig. 3 across different tasks. It is evident that the gating network managed to dynamically adapt the weights to accommodate various scenarios, providing additional in", + "bbox": [ + 88, + 734, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26c4c60d7b0cb0da60794d819fd7123c11f1ae8ddbe76bd804d37aa1659f48f6.jpg", + "image_caption": [ + "Figure 3. Average gating weights analysis under both full and missing modality setups. Plotted values are the average gating weights on the BRATS 2018 official evaluation set." + ], + "image_footnote": [], + "bbox": [ + 532, + 87, + 888, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2ff5ffda69616017def2d3d5e84801cd0719721b7f3746c221ac24c1133147a.jpg", + "image_caption": [ + "Figure 4. Impact of the MoFe loss coefficient $\\lambda$ . The Dice score (\\%) $\\uparrow$ is reported with different values of coefficients $\\lambda$ for the MoFe loss across all missing & full modality settings on the BraTS 2018 validation set. The results include (1) enhancing tumor, (2) tumor core, (3) whole tumor, and (4) their average." + ], + "image_footnote": [], + "bbox": [ + 516, + 424, + 906, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sights for clinicians regarding modality importance in different subtasks. For instance, T1ce is prioritized for enhancing tumor (ET) segmentation, while Flair and T2 are more important for whole tumor (WT) segmentation, which is consistent with clinical findings [22].", + "bbox": [ + 511, + 742, + 906, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness and Sensitivity of MoFe loss: We further analyzed the impact of MoFe loss by varying its coefficient $\\lambda = \\{0, 0.01, 0.05, 0.1, 0.2, 0.5, 1\\}$ in Eq.4, where $\\lambda = 0$ corresponds to training without it. As shown in Figure 4, the model achieved the highest average Dice score across", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24074", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d213656548490e5cd164a1d44cd6da8a2ea599135b91a0b11133e8d880c85f5e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ECE (%) ↓SCE (%) ↓
MethodETTCWTETTCWT
MoMKE [38]3.464.104.0414.8211.357.69
DMoME3.343.863.9813.4011.017.11
DMoME w/ LConf [19]3.173.863.6114.8011.586.81
DMoME w/ LMoFe (SimMLM)3.153.753.5513.2110.835.65
", + "bbox": [ + 109, + 88, + 464, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the three subregions at $\\lambda = 0.1$ . Notably, most positive $\\lambda$ values resulted in improvements over the baseline $\\lambda = 0$ , underscoring the stability and effectiveness of MoFe loss. This improvement may be attributed to the calibration nature of the loss, which enables DMoME to adaptively distinguish between modality-rich and modality-poor conditions, guiding the network to assess each modality's contribution across various scenarios. As a result, the model learns a more casual, robust relationship between diverse inputs and the task output. In the following, we will verify how $\\mathcal{L}_{MoFe}$ can calibrate the network to achieve the goal.", + "bbox": [ + 88, + 268, + 482, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MoFe loss leads to better calibration: To evaluate the performance of $\\mathcal{L}_{MoFe}$ for model calibration, we compute the expected calibration error (ECE) $\\downarrow$ and static calibration error (SCE) $\\downarrow$ [25], both using 20 bins, and compare our method to the most competitive method: MoMKE [38], in addition to the state-of-the-art multimodal calibration approaches, leveraging missing modality information [19], where the task losses $\\mathcal{L}_{task}$ in $\\mathcal{L}_{MoFe}$ have been replaced with confidence scores obtained from model predictions using max probability: $\\mathcal{L}_{Conf} = \\max(0, Conf(\\mathbf{o}^{-}) - Conf(\\mathbf{o}^{+}))$ . Results in Table 4 are scores directly computed on our validation set, as BraTS does not release ground truth labels on the evaluation set. All experiments were conducted in the same settings. It is evident that our method with $\\mathcal{L}_{MoFe}$ achieves the best calibration performance in all settings, with the lowest ECEs and SCEs. In the context of this segmentation tasks, compared to $\\mathcal{L}_{Conf}$ , which is label-ignorant and performs comparisons at a low level (e.g., voxels), $\\mathcal{L}_{MoFe}$ operates at the sample level, and leverages task-specific loss with ground truth information, making it more task-informative. This encourages the network to focus on task-relevant areas across varying levels of missing data, reducing sensitivity to background noise, augmentation artifacts, and outliers.", + "bbox": [ + 91, + 448, + 482, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MoFe loss reduces model counterintuitive rate (CR): To directly quantify whether $\\mathcal{L}_{MoFe}$ can better align the model performance with the modality availability (i.e., reduces cases where the model performs better even with fewer input modalities), we introduce a novel metric called the", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f09563dfbbcb76ed3eec96902a0d2bb022075bcf1c27abc252559786580d1e5.jpg", + "table_caption": [ + "Table 4. Quantitative comparison of calibration errors (ECEs↓, SCEs↓) averaged across all 15 full and missing modality settings on the validation set of BraTS 2018 dataset." + ], + "table_footnote": [], + "table_body": "
Counterintuitive rate (%) ↓
MethodETTCWT
MoMKE [38]14.6032.479.86
DMoME w/o LMoFe10.4730.943.87
DMoME w/ LMoFe (SimMLM)7.2028.193.75
", + "bbox": [ + 534, + 88, + 883, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of counterintuitive rate $\\downarrow$ on the BraTS 2018 official evaluation set.", + "bbox": [ + 511, + 172, + 903, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Counterintuitive Rate (CR), which is defined as", + "bbox": [ + 511, + 224, + 828, + 239 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm{CR} = \\frac{1}{N}\\sum_{i = 1}^{N}\\frac{1}{|\\mathcal{S}_{i}|}\\sum_{(\\mathbf{x}_{i}^{-},\\mathbf{x}_{i}^{+})\\in \\mathcal{S}_{i}}\\left\\{ \\begin{array}{ll}1 & \\text{if} P(\\mathbf{x}_{i}^{-}) > P(\\mathbf{x}_{i}^{+})\\\\ 0 & \\text{otherwise} \\end{array} \\right..\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 244, + 895, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here $N$ is the total number of samples, $\\mathcal{S}_i$ is the set of all more vs. fewer modality pairs $(\\mathbf{x}_i^+ \\text{vs.} \\mathbf{x}_i^-)$ for sample $i$ . $P(\\mathbf{x}_i)$ denotes the model accuracy score (e.g., Dice score for the segmentation task) on the sample $\\mathbf{x}_i$ . An instance with $P(\\mathbf{x}_i^-) > P(\\mathbf{x}_i^+)$ is considered as counterintuitive.", + "bbox": [ + 511, + 297, + 903, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compared CR on BraTS 2018 across three settings: MoMKE [38] (previous SOTA), DMoME without $\\mathcal{L}_{MoFe}$ and DMoME with $\\mathcal{L}_{MoFe}$ (SimMLM) and report the performance in Table 5. Results show that $\\mathcal{L}_{MoFe}$ achieves the lowest CR, achieving $50.68\\%$ , $13.18\\%$ , and $61.97\\%$ reduction for ET, TC, and WT segmentation, respectively, when compared to MoMKE. Notably, DMoME alone also reduces CR significantly, thanks to the dynamic weighting mechanism to automatically identify and prioritize more relevant modalities.", + "bbox": [ + 511, + 375, + 905, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact: We have shown that SimMLM can provide direct importance values for each input modality and generate informative confidence maps for further analysis in the presence of incomplete data (Fig. 2). Such properties are particularly important to alert clinicians for further investigation in low-resource settings where certain modalities may be unavailable. Additionally, SimMLM's potential for handling varied levels of training-time missing modalities on avMNIST is also discussed in the Supplementary.", + "bbox": [ + 511, + 531, + 905, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced SimMLM, a simple, robust, and flexible framework for multimodal learning in scenarios with test-time missing modalities. By incorporating the Dynamic Mixture of Model Experts (DMoME) architecture and the novel More vs. Fewer (MoFe) ranking loss, SimMLM effectively addresses modality absence. Experimental results demonstrate that SimMLM achieves state-of-the-art performance across both segmentation and classification tasks, excelling in accuracy, efficiency, interpretability, robustness, and reliability. As a generic approach, future work will focus on extending SimMLM to other high-stakes real-world applications where trustworthiness matters and broader scenarios, including training-time missingness.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24075", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the UKRI funding (MR/Z505754/1), UK. Dr. Chen is supported by Royal Society (RGS/R2/242355).", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 447-456. Springer, 2019. 1, 2, 5, 6", + "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 5", + "[3] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2", + "[4] Ignazio Gallo, Gianmarco Ria, Nicola Landro, and Riccardo La Grassa. Image and text fusion for upmc food-101 using bert and cnns. In 2020 35th International Conference on Image and Vision Computing New Zealand (IVCNZ), pages 1-6. IEEE, 2020. 1, 4, 5", + "[5] Yunhe Gao. Training like a medical resident: Context-prior learning toward universal medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11194-11204, 2024. 1", + "[6] Shuang Hao, Wenfeng Han, Tao Jiang, Yiping Li, Haonan Wu, Chunlin Zhong, Zhangjun Zhou, and He Tang. Synthetic data in ai: Challenges, applications, and ethical implications. arXiv preprint arXiv:2401.01629, 2024. 1", + "[7] Minhao Hu, Matthias Maillard, Ya Zhang, Tommaso Ciceri, Giammarco La Barbera, Isabelle Bloch, and Pietro Gori. Knowledge distillation from multi-modal to mono-modal segmentation networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 772-781. Springer, 2020. 4, 5", + "[8] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 5", + "[9] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 4", + "[10] Donggeun Kim and Taesup Kim. Missing modality prediction for unpaired multimodal learning via joint embedding of unimodal models. arXiv preprint arXiv:2407.12616, 2024. 1" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[12] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5", + "[13] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 5", + "[14] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2", + "[15] Tao Liang, Guosheng Lin, Mingyang Wan, Tianrui Li, Guojun Ma, and Fengmao Lv. Expanding large pre-trained unimodal models with multimodal information injection for image-text multimodal classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15492-15501, 2022. 1", + "[16] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 1", + "[17] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotaining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 4", + "[18] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2", + "[19] Huan Ma, Qingyang Zhang, Changqing Zhang, Bingzhe Wu, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Calibrating multimodal learning. In International Conference on Machine Learning, pages 23429-23450. PMLR, 2023. 4, 8", + "[20] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 5", + "[21] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18177-18186, 2022. 2", + "[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24076", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 4, 7", + "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5", + "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11480-11491, 2024. 1", + "[25] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. In CVPR workshops, 2019. 8", + "[26] Jongyoun Noh, Hyekang Park, Junghyup Lee, and Bumsub Ham. Rankmixup: Ranking-based mixup training for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1358-1368, 2023. 4", + "[27] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems, 36:24829-24840, 2023. 1", + "[28] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2", + "[29] Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2515, 2024. 1", + "[30] Yujie Sun, Dongfang Sheng, Zihan Zhou, and Yifei Wu. AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content. *Humanities & social sciences communications*, 11(1): 1-14, 2024. 1", + "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5", + "[32] Jonathan G Tullis and Robert L Goldstone. Comparison versus reminding. Cognitive research: principles and implications, 1(1):20, 2016. 4", + "[33] George Tzanetakis and Perry Cook. Musical genre classification of audio signals. IEEE Transactions on speech and audio processing, 10(5):293-302, 2002. 5", + "[34] Valentin Vielzeuf, Alexis Lecheryv, Stéphane Pateux, and Frédéric Jurie. Centralnet: a multilayer approach for multimodal fusion. In Proceedings of the European Conference" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision (ECCV) Workshops, pages 0-0, 2018. 4, 5", + "[35] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 1, 2, 5, 6, 7", + "[36] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[37] Renjie Wu, Hu Wang, Hsiang-Ting Chen, and Gustavo Carneiro. Deep multimodal learning with missing modality: A survey, 2024. 2", + "[38] Wenxin Xu, Hexin Jiang, et al. Leveraging knowledge of modality experts for incomplete multimodal learning. In ACM Multimedia 2024, 2024. 2, 5, 6, 7, 8", + "[39] Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingxuan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tianlong Chen. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts. arXiv preprint arXiv:2410.08245, 2024. 2", + "[40] Zhilin Zeng, Zelin Peng, Xiaokang Yang, and Wei Shen. Missing as masking: Arbitrary cross-modal feature reconstruction for incomplete multimodal brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424-433. Springer, 2024. 1", + "[41] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 5, 6, 7", + "[42] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2, 5", + "[43] Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. Sparse moe with language guided routing for multilingual machine translation. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[44] Yunpeng Zhao, Cheng Chen, Qing You Pang, Quanzheng Li, Carol Tang, Beng-Ti Ang, and Yueming Jin. Dealing with all-stage missing modality: Towards a universal model with robust reconstruction and personalization. arXiv preprint arXiv:2406.01987, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 795 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_model.json b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..092792f3e0b457e4ae057743b4ea64f0e0d195f5 --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.13, + 0.868, + 0.178 + ], + "angle": 0, + "content": "SimMLM: A Simple Framework for Multi-modal Learning with Missing Modality" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.203, + 0.726, + 0.257 + ], + "angle": 0, + "content": "Sijie Li Chen Chen† Jungong Han*† \nSchool of Computer Science, University of Sheffield, UK \n{sli256, chen.chen2, jungong.han}@sheffield.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.325, + 0.483, + 0.671 + ], + "angle": 0, + "content": "In this paper, we propose SimMLM, a simple yet powerful framework for multimodal learning with missing modalities. Unlike existing approaches that rely on sophisticated network architectures or complex data imputation techniques, SimMLM provides a generic and effective solution that can adapt to various missing modality scenarios with improved accuracy and robustness. Specifically, SimMLM consists of a generic Dynamic Mixture of Modality Experts (DMoME) architecture, featuring a dynamic, learnable gating mechanism that automatically adjusts each modality's contribution in both full and partial modality settings. A key innovation of SimMLM is the proposed More vs. Fewer (MoFe) ranking loss, which ensures that task accuracy improves or remains stable as more modalities are made available. This aligns the model with an intuitive principle: removing one or more modalities should not increase accuracy. We validate SimMLM on multimodal medical image segmentation (BraTS 2018) and multimodal classification (UPMC Food-101, avMNIST) tasks, where it consistently surpasses competitive methods, demonstrating superior accuracy, interpretability, robustness, and reliability across both complete and missing modality scenarios at test time. Code is available at https://github.com/LezJ/SimMLM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Multimodal learning aims to build models that can process and integrate information from multiple, complementary modalities such as visual, textual and auditory information, which has been found useful for various tasks such as image-text food classification [4, 15], multimodal medical image segmentation [5, 24] with improved model accuracy. However, in real-world scenarios, it is common for one or more modalities to be missing during deployment due to hardware failures, environmental conditions, or constraints" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.338 + ], + "angle": 0, + "content": "in data collection, budget, and resources. Addressing this test-time missing modality problem is crucial for building robust multimodal models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.641 + ], + "angle": 0, + "content": "One mainstream to address the missing modality problem is data imputation, which focuses on synthesizing the missing modalities from available ones [18, 44] or approximating intermediate missing features [10, 35, 36, 40]. However, ensuring the factuality and fidelity of synthetic data remains a major challenge, as issues such as poor data quality [16, 27], data hallucination [30], and vulnerability to adversarial corruption [6, 29] have been reported. Moreover, the computational overhead of generative imputation limits its practicality. In parallel, sophisticated frameworks have been developed, which aim to tackle the missing modality by leveraging advanced, complex network architectures [1, 35, 41], and harmonize or align early or late features from different modalities into a unified, shared space. However, such approaches often introduce substantial resource demands. While some are not adaptable to other tasks at all, others constrain the choice of network architectures or require additional modules, such as projection layers, to accommodate diverse or heterogeneous modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.905, + 0.824 + ], + "angle": 0, + "content": "In this paper, we introduce SimMLM, a simple yet effective framework for multi-modal learning with missing modalities, without relying on specific neural network architectures or complex data imputation. First, to effectively handle varying modality availability, we propose a Dynamic Mixture of Modality Experts (DMoME) architecture, where modality-specific expert networks process available modalities, and a learnable gating network dynamically adjusts their contributions. This design ensures that the model remains effective in both full and partial modality settings while being adaptable to different neural architectures and task domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To further improve the model's robustness and reliability against missing modalities, we introduce a novel loss function, which is designed to optimize the loss landscape around the region of missing data. Inspired by an intuitive principle that the model should not degrade in per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.235, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.396, + 0.9 + ], + "angle": 0, + "content": "†† These authors contributed equally as senior authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "formance when modalities are increased, this loss function, named More vs. Fewer (MoFe) ranking loss, ensures that as more modalities become available, the model's accuracy improves or at least remains stable. The MoFe loss encourages the model to compare performance between modality-rich and modality-poor pairs, promoting an inherent ranking that favors configurations with greater modality availability. This structured comparison in together with the dynamic weighting scheme in DMoME allows the model to learn consistent and reliable multimodal fusion strategies, improving robustness and adaptability across varying modality availability." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.274, + 0.435, + 0.287 + ], + "angle": 0, + "content": "Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.289, + 0.483, + 0.408 + ], + "angle": 0, + "content": "- Dynamic Mixture of Modality Experts Architecture: SimMLM introduces the DMoME architecture with a learnable gating mechanism that adaptively adjusts modality contributions based on availability, improving performance in both complete and partial modality settings. This design enables seamless integration with various neural networks, ensuring flexibility and scalability across diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.409, + 0.483, + 0.5 + ], + "angle": 0, + "content": "- More vs. Fewer ranking loss: We introduce the novel MoFe ranking loss, which ensures model performance improves or remains stable as additional modalities are incorporated. By enforcing a ranking comparison between modality-rich and modality-poor pairs, MoFe strengthens model robustness and reliability." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.5, + 0.483, + 0.605 + ], + "angle": 0, + "content": "- Comprehensive Evaluation: We validate SimMLM on multimodal segmentation and classification across three datasets, demonstrating superior accuracy, interpretability, robustness, reliability, and computational efficiency over state-of-the-art methods. These results establish SimMLM as a practical and versatile solution for multimodal learning under missing modality conditions." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.289, + 0.483, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.233, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.705 + ], + "angle": 0, + "content": "To solve multi-modal learning with missing modalities, various approaches have been proposed, including a) input-level imputation, b) advanced representation learning; and c) mixture of expert strategies [37]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Input-level imputation approaches leverage advanced generative networks such as auto-encoders [18, 20], generative adversarial networks [20] to reconstruct missing inputs. Such an approach however, can generate noisy and unreliable input, as discussed in the previous section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Advanced representation learning focuses on learning improved representation either through advanced networks such as transformers [21, 41, 42], represented by mm-Former [41], or disentangling the latent features into modality-specific and modality-shared features for improving model accuracy, such as Robust-Mseg [1] and ShaSpec [35]. For instance, ShaSpec [35] employs a combination of modality-shared and modality-specific encoders" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "with dedicated losses to promote feature disentanglement and perform feature fusion in the latent space before passing them to a task decoder. When a modality is missing, ShaSpec generates its features from the available modalities through a missing modality feature generation process enabled by the modality-shared encoder. However, this approach imposes a strict constraint that the shared encoder should be suitable for all input modalities and both shared and specific features must have the same dimensions, which can be less efficient and optimal when dealing with highly heterogeneous modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Mixture of Experts (MoE) has recently been adopted in both unimodal and multimodal learning, demonstrating superior performance in natural language processing [3, 14, 43] and computer vision tasks [28, 38, 39], while improving efficiency by reducing FLOPS. A very recent method for handling missing modalities is Mixture of Modality Knowledge Experts (MoMKE) [38], which uses MLP experts to generate unimodal and joint representations. Even when a modality is missing, MoMKE leverages existing modalities to generate joint representations by utilizing the other experts. While this approach has achieved promising results, it still incurs high computational costs, as each modality must pass through all modality experts, quadratically increasing complexity with the number of modalities. In contrast, the expert model in DMoME is dedicated to a specific modality. Additionally, unlike MoMKE, DMoME does not constrain the dimensionality of latent feature representations across modalities. Instead, it performs the weighting directly at the output space across experts. This flexibility makes DMoME more efficient and scalable, offering high performance in diverse multimodal tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.592, + 0.649, + 0.609 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.696, + 0.632 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.821 + ], + "angle": 0, + "content": "Assume we have a complete multimodal dataset with \\(M\\) modalities, denoted as \\(\\mathcal{D} = \\{\\{\\mathbf{x}_i^m\\}_{m=1}^M, \\mathbf{y}_i\\}_{i=1}^N\\). where \\(\\mathbf{x}_i^m \\in \\mathcal{X}\\) represents the \\(i\\)-th data sample and \\(m\\) indexes the modality. \\(\\mathbf{y}_i \\in \\mathcal{Y}\\) is the label for the sample \\(i\\) from the label space \\(\\mathcal{Y}\\), which can either be a classification label or segmentation map. In this paper, we are interested in finding a solution \\(f\\) that learns a mapping from the input space to the label space: \\(f: \\mathcal{X} \\to \\mathcal{Y}\\), leveraging this complete modality dataset, so that \\(f\\) can work robustly and reliably even when certain modalities \\(m\\) are missing at inference time. For simplicity, we omit the sample index \\(i\\) when not specified in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.724, + 0.848 + ], + "angle": 0, + "content": "3.2. Overview of SimMLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our proposed SimMLM framework consists of two main components: a) a unified Dynamic Mixture of Modality Experts framework (DMoME) (Sec. 3.3), which can take a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24069" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.09, + 0.443, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.307, + 0.343, + 0.32 + ], + "angle": 0, + "content": "(a) Our DMoME" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.106, + 0.846, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.308, + 0.763, + 0.321 + ], + "angle": 0, + "content": "(b) Computation flows of \\(\\mathcal{L}_{MoFe}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.337, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Figure 1. Our proposed SimMLM consists of a) DMoME architecture: DMoME takes either partial or full modality inputs where each modality is processed by a corresponding expert network to produce its own output \\((\\mathbf{o}^1, \\ldots, \\mathbf{o}^M)\\). A gating network (G) assesses the same inputs and produces weights to dynamically weigh the contributions of each expert's output in order to get the mixture of output \\(\\mathbf{o}\\); b) \\(\\mathcal{L}_{\\mathrm{MoFe}}\\) loss: At training time, in addition to standard task specific losses, DMoME is optimized with the More \\((X^{+})\\) vs. Fewer Modality \\((X^{-})\\) (MoFe) ranking loss, directly encouraging the network to distinguish between inputs with differing modality availability at the task performance level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.54 + ], + "angle": 0, + "content": "varied number of modalities as input and adaptively weight their contribution for robust performance at both training and inference time; b) a novel ranking loss called MoFe (Sec. 3.4), which is designed to calibrate the network's performance by differentiating network performance when the number of available modalities for the same sample varies." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.577 + ], + "angle": 0, + "content": "3.3. Dynamic Mixture of Modality Expert (DMoME)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.734 + ], + "angle": 0, + "content": "As shown in Figure 1 (a), Our DMoME consists of a set of modality-specific expert networks \\(\\{E^{m}(\\mathbf{x}_{m};\\theta_{m})\\}_{m = 1}^{M}\\) each containing its own set of learnable parameters \\(\\theta_{m}\\) and a gating network, \\(G(\\cdot ;\\phi)\\), with a set of learnable parameters \\(\\phi\\). Each expert network, \\(E^{m}(\\mathbf{x}_{m};\\theta_{m})\\), takes its corresponding modality as input and produces its output in logits \\(\\mathbf{o}^m\\), before applying the softmax function: \\(\\mathbf{o}^m = E^m (\\mathbf{x}^m;\\theta^m)\\in \\mathbb{R}^T\\). Here, \\(T\\) denotes the dimensionality of the output space for a specific task, e.g., the number of predicted classes in a classification task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.872 + ], + "angle": 0, + "content": "The gating network \\( G(\\cdot; \\phi) \\) takes a set of different modalities as input and produces a set of gating values \\( \\{g^m\\}_{m=1}^M \\) to calibrate different modality experts' output. The softmax function is then applied to computing non-negative weights \\( \\{w_1, w_2, \\ldots, w_M\\} \\in \\mathbb{R} \\), ensuring that the sum of the weights for existing modalities is equal to 1. The final output can be described as a reweighted combination of the outputs from the corresponding modality experts \\( \\{\\mathbf{o}^m\\}_{m=1}^M \\): \\( \\mathbf{o} = \\sum_{m=1}^M w^m E^m(\\mathbf{x}^m; \\theta_m) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Notably, the input to the gating network can be an incomplete set of modalities. If certain modalities are miss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.522 + ], + "angle": 0, + "content": "ing, those inputs are treated as 0, and the gating values for the corresponding missing modality experts' outputs are set to \\(-\\infty\\) (which causes the corresponding weight values to be 0 after the softmax). Mathematically, the gating outputs are set as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.532, + 0.906, + 0.573 + ], + "angle": 0, + "content": "\\[\ng _ {m} = \\left\\{ \\begin{array}{l l} G (\\mathbf {x}; \\phi) ^ {(m)}, & \\text {i f m o d a l i t y m i s p r e s e n t} \\\\ - \\infty , & \\text {i f m o d a l i t y m i s m i s s i n g} \\end{array} , \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.622 + ], + "angle": 0, + "content": "where \\( \\mathbf{x} \\) can be a complete set of modalities or an arbitrary incomplete set of modalities for a given sample. The final weights therefore can be simply computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.627, + 0.906, + 0.657 + ], + "angle": 0, + "content": "\\[\nw ^ {m} = \\frac {\\exp \\left(g ^ {m}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\exp \\left(g ^ {m ^ {\\prime}}\\right)} \\quad \\text {f o r} \\quad m = 1, 2, \\dots , M. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.906, + 0.77 + ], + "angle": 0, + "content": "Here, the weight vector \\(\\mathbf{w} = [w^{1}, w^{2}, \\ldots, w^{M}] \\in \\mathbb{R}^{M}\\) is non-negative and sums to 1, with \\(w^{m} = 0\\) for missing modalities, effectively excluding them from the final mixture. The above configuration can be easily extended to multi-task settings, where each task has its own task-specific weighting scheme. In this case, the gating network produces \\(\\mathbf{w} \\in \\mathbb{R}^{M \\times K}\\), with \\(K\\) being the number of tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.777, + 0.651, + 0.792 + ], + "angle": 0, + "content": "Two-stage training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.906, + 0.825 + ], + "angle": 0, + "content": "To train the whole network, we adopt a two-stage training process, which consists of:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "1. independent learning: we independently train a modality-specific expert \\( E_{m} \\) for each modality \\( m \\), supervised by a task-specific loss \\( \\ell_{\\mathrm{task}}(\\mathbf{o}^{m},\\mathbf{y}) \\). For example, for classification tasks, the loss function can be the cross-entropy (CE) loss, whereas for the segmentation," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24070" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.483, + 0.198 + ], + "angle": 0, + "content": "the loss can be Dice loss. This pretraining stage ensures that each modality expert independently learns specialized features, minimizing the risk of any single modality dominating the learning process. Different modalities can vary significantly in learning dynamics and task difficulty, which can lead to training instability, especially at the early stages of joint training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.198, + 0.483, + 0.257 + ], + "angle": 0, + "content": "2. Cooperative learning: we co-train the expert networks alongside the gating network, ranking network performance under More vs. Fewer Modality conditions, see next section for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.381 + ], + "angle": 0, + "content": "This training scheme enables the expert networks to exploit modality-specific information without interference from other modalities before jointly optimizing with the gating network to exploit mutual benefits. It also improves framework flexibility with reduced training resources, e.g., a) first-stage training can be accelerated via parallel computing; b) when new modalities are introduced, existing pretrained expert networks can be reused for the second stage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.418, + 0.409 + ], + "angle": 0, + "content": "3.4. More vs. Fewer Modality Co-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.536 + ], + "angle": 0, + "content": "To boost model robustness against missing modalities, we simulate scenarios with missing modalities by randomly dropping out input modalities during training, which is a common practice. Yet, unlike existing approaches that focus on aligning outputs from missing modalities with those from the full set of modalities [7, 9, 17], we encourage the network to distinguish between modality-rich and modality-poor inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Rational behind More vs. Fewer Modality Training: In many multi-modal learning tasks, all modalities are assumed to be predictive and complementary, contributing valuable information to the tasks that are of interest. Our approach is inspired by the intuition that richer multimodal inputs should yield more accurate results, as each modality complements the others. As the number of modalities decreases, the input becomes more limited, making it more challenging to capture complex patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.824 + ], + "angle": 0, + "content": "As it is difficult to quantify the exact performance for different set-ups, we propose to formulate it as a ranking loss to encourage the network to compare the two scenarios. In cognitive science, it has been found that learners who engage in comparative tasks show an improved ability to transfer knowledge to new situations compared to those who learn through isolated examples [32]. Such a concept has also been applied in reinforcement learning for preference alignment for the applications like large language model training and finetuning with a reward function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our multimodal learning setting, at each training iteration, we sample two sets from a data point \\(\\mathbf{x}_{full}\\): \\(\\mathbf{x}^{+}\\) and \\(\\mathbf{x}^{-}\\), where \\(\\mathbf{x}^{-} \\subset \\mathbf{x}^{+} \\subseteq \\mathbf{x}^{full}\\). The corresponding outputs from DMoME are denoted as \\(\\mathbf{o}^{+}\\) and \\(\\mathbf{o}^{-}\\), respectively. We conjecture that the task-specific loss value for modality-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "rich setups with \\(\\mathbf{x}^{+}\\) should be at least no higher than those in the modality \\(\\mathbf{x}^{-}\\). Mathematically, every pair of samples \\((\\mathbf{x}^{+},\\mathbf{x}^{-})\\) should satisfy the following relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.144, + 0.91, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) \\geq \\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) \\iff A c c \\left(\\mathbf {x} ^ {+}\\right) \\geq A c c \\left(\\mathbf {x} ^ {-}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.214 + ], + "angle": 0, + "content": "Definition of \\(\\mathcal{L}_{MoFe}\\): To guide the network to learn the ordinary relationship, we introduce More vs. Fewer (MoFe) ranking loss, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {M o F e}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\max \\left(0, \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) - \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.907, + 0.357 + ], + "angle": 0, + "content": "which encourages the network to perform better when more modalities are available. It is worth mentioning that the proposed loss is general and thus can be applied to other existing multi-modal learning frameworks to better calibrate the network for improved robustness against missing modalities as well as improved reliability. The total training loss is thus defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.365, + 0.905, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {t o t a l}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) + \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) + \\lambda \\ell_ {\\mathrm {M o F e}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.904, + 0.411 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a coefficient to control the contribution of \\(\\ell_{\\mathrm{MOFE}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.729 + ], + "angle": 0, + "content": "We acknowledge the existence of prior research, such as the works of Noh et al. (2023) [26], and Ma et al. (2023) [19], which have investigated the application of ranking loss at the confidence levels between a pair of two inputs. However, our approach diverges significantly from these existing methods. Specifically, we propose the novel concept of applying ranking loss directly at the loss level, as opposed to the confidence level. This distinction is crucial, as our method aims to directly regularize the geometry of the loss landscape. In our experiments, we also showed extra benefit of our loss, which leads to reduced calibration errors, see Table. 4. Besides, our proposed method enables seamless integration with a range of loss functions, offering flexibility beyond classification and allowing for improved applicability in complex, diverse tasks. For example, our method can be applied for segmentation and regression tasks where the confidence scores are not straightforward in dense or continuous prediction settings. By focusing on the actual task error, our loss-level approach directly aligns with the downstream task performance, providing a more consistent measure across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.741, + 0.646, + 0.758 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.843 + ], + "angle": 0, + "content": "We evaluate our approach on two types of tasks, multimodal medical segmentation and multimodal classification, spanning three datasets. For segmentation, we use BraTS 2018 [22] benchmark. For classification, we conduct experiments on UPMC Food-101 [4] and avMNIST dataset [34]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.616, + 0.864 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "BraTS 2018 (Brain Tumor Segmentation Challenge [22]) is a widely used benchmark for brain tumor segmenta" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24071" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "tion in 3D MRI scans. It has also been extensively employed to evaluate methods for handling missing modalities [7, 35, 41, 42]. The dataset includes four distinct MRI modalities: FLAIR, T1, T1 contrast-enhanced (T1ce), and T2. The segmentation task involves identifying three tumor subregions: the enhancing tumor (ET), the tumor core (TC), and the whole tumor (WT). The dataset consists of 285 manually labeled cases and 66 unlabeled cases for evaluation. In our experiments, we split the labeled cases into a 4:1 ratio for training and validation. Final evaluations are conducted on the 66 unlabeled cases via the official benchmarking platform for a consistent and fair accuracy comparison against existing methods [13, 41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.348 + ], + "angle": 0, + "content": "UPMC Food-101 [4] is a large-scale food classification dataset with 90,704 image-text pairs spanning 101 classes. Collected from uncontrolled environments, it introduces inherent noise, posing challenges for accurate classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.482, + 0.424 + ], + "angle": 0, + "content": "avMNIST (Audiovision MNIST [34]) is a digit classification dataset comprising 1,500 image-audio pairs. The visual modality consists of \\(28 \\times 28\\) grayscale images of handwritten digits (0-9) from MNIST [12], while the audio modality is sourced from the Free Spoken Digits Dataset*." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.308, + 0.449 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.485 + ], + "angle": 0, + "content": "All the experiments were conducted on a single A100 GPU node. Detailed implementations are described below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.637 + ], + "angle": 0, + "content": "BraTS 2018: We employed nnUNet [8], a widely used and powerful medical image segmentation backbone, as the expert network for each modality, which can be viewed as a 3D U-Net. The gating network \\( G \\) in DMoME utilized lightweight CNNs to capture broad contextual information, followed by a linear layer generating \\( 4 \\times 3 \\) weights for the four modalities and three segmentation tasks. We optimized both training stages using the Adam optimizer [11] (\\( lr = 0.01 \\)). The \\( \\mathcal{L}_{task} \\) in Eq.4 combined Dice [23] and binary cross-entropy losses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.741 + ], + "angle": 0, + "content": "UPMC Food-101: We employed Inception V3 [31] as the image expert and BERT [2] as the text expert, following [4]. To avoid redundant feature extraction, intermediate features from both were fed into an MLP gating network \\( G \\), which generated 2D weights for expert outputs. Optimization was performed using Adam (\\( lr = 0.0001 \\)) with cross-entropy loss as \\( \\mathcal{L}_{task} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.833 + ], + "angle": 0, + "content": "avMNIST: Following [20, 35], we adopted a LeNet-5 [12] backbone for both image and audio modalities, with audio converted into \\(20 \\times 20\\) MFCC spectrograms [33]. The gating network \\(G\\) followed a similar design as in BraTS 2018. Two-stage training was conducted using Adam (\\(lr = 0.001\\)) with cross-entropy loss as \\(\\mathcal{L}_{task}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Detailed gating network \\( G \\) architectures for BraTS 2018 and avMNIST are provided in the Supplementary. For task accuracy evaluation, we follow common practice, using the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Dice score for segmentation (BraTS 2018) and Accuracy for classification (UPMC Food-101, avMNIST). The MoFe loss coefficient \\(\\lambda\\) in Eq.4 is set to 0.1 across tasks; however, our method remains robust to variations in \\(\\lambda\\), as demonstrated in Sec. 5 Analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.179, + 0.715, + 0.195 + ], + "angle": 0, + "content": "4.3. Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.905, + 0.504 + ], + "angle": 0, + "content": "Benchmarking segmentation accuracy on the official BraTS 2018 evaluation set. For the BraTS 2018 benchmark, we compare our approach against several state-of-the-art methods, including Robust-MSeg (RbSeg) [1], mmFormer (mmFm) [41], ShaSpec [35], and MoMKE [38] under 15 configurations of both missing and full modality setups in Table 1. Results show that SimMLM achieves the highest average accuracy in terms of the Dice scores, consistently ranking as the top or second-highest performer in each entry across 44 out of all 45 configurations (15 settings \\(\\times\\) 3 subtasks), despite utilizing only \\(80\\%\\) of the labeled data compared to the \\(100\\%\\) used by ShaSpec and mmFormer. This consistent performance highlights our model's robustness and data efficiency in handling various incomplete modality inputs. Notably, when only the T1ce modality is available (Table 1, row 3), our model significantly improves the Dice score compared to the second-best methods, increasing from 73.29 to 78.58 for ET, 78.71 to 83.45 for TC, and 74.00 to 80.02 for WT. Such results underscore the advantage of our approach even in single-modality settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.905, + 0.7 + ], + "angle": 0, + "content": "As shown in Figure 2, in both full and partial modality scenarios, SimMLM (DMoME with \\(\\mathcal{L}_{MoFe}\\)) achieves more accurate segmentation results with better model reliability compared to DMoME alone (A vs. B, C vs. D) thanks to the contribution of \\(\\mathcal{L}_{MoFe}\\). When T1ce is unavailable, DMoME with \\(\\mathcal{L}_{MoFe}\\) can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce), which is aligned with clinical practice despite the network is not explicitly taught. Without MoFe loss, the network is biased to T2 (maybe due to higher intensity) and over-segment the tumor. In the following Sec. 5, we will provide a more detailed quantitative analysis for our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.714, + 0.727 + ], + "angle": 0, + "content": "4.4. Classification Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Benchmarking classification accuracy on UPMC Food101 and avMNIST. For the classification tasks, we compared our model to the two most competitive methods: ShaSpec [35] and MoMKE [38]. To ensure fair comparisons, all models were trained under identical conditions. The results, presented in Table 2, report accuracy across different modality settings. Our method consistently surpasses SOTA methods across all settings, demonstrating its effectiveness in handling both missing and full modality scenarios. While clear improvements are observed on both datasets, our method achieves even greater gains on the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Dataset link: https://github.com/Jakobovski/free-spoken-digit-dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24072" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.154, + 0.089, + 0.85, + 0.359 + ], + "angle": 0, + "content": "
SettingsAvailable Modalities at Test TimeEnhancing tumor (ET)Tumor core (TC)Whole tumor (WT)
◇◇◇RbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]Ours
Missing1●○○25.6939.3343.5250.5351.3153.5761.2169.4475.0574.1285.6986.1088.6889.7889.24
2○●○17.2932.5341.0048.1248.4147.9056.5563.1868.6671.4870.1167.5273.4477.4580.28
3○○●67.0772.6073.2972.3078.5876.8375.4178.6578.7183.4573.3172.2273.8274.0080.02
4○○○28.9743.0546.3151.0250.6057.4964.2069.0372.3473.4982.2481.1583.9985.8286.01
5●●○32.1342.9644.7652.5253.6260.6865.9172.6775.0077.1488.2487.0689.7690.2389.98
6●○●70.3075.0775.6078.7580.2580.6277.8884.5084.6284.9688.5187.3090.0690.3490.35
7●○○33.8447.5247.2054.1254.6261.1669.7572.9376.9075.7688.2887.5990.0290.5390.28
8○●●69.0674.0475.7676.7878.7778.7278.5982.1081.7084.7977.1874.4278.7479.8682.25
9○●○32.0144.9946.8453.7453.4062.1969.4271.3873.7374.7384.7882.2086.0386.6087.42
10○○●69.7174.5175.9577.1178.6180.2078.6183.8283.6184.1985.1982.9985.4286.9288.11
11●●●70.7875.4776.4278.1980.4181.0679.8085.2382.8385.2588.7387.3390.2990.2190.63
12●●○36.4147.7046.5555.9355.8464.3871.5273.9777.0078.0288.8187.7590.3690.5990.62
13●○●70.8875.6775.9978.4480.2980.7279.5585.2683.5985.2389.2788.1490.7890.8590.99
14○●●70.1074.7576.3778.2780.3380.3380.3984.1882.3585.0886.0182.7186.4786.6287.78
Full15●●●71.1377.6178.0877.6580.4080.8685.7885.4582.6185.3989.4589.6490.8890.5491.07
Average51.0259.8561.5865.5667.1669.7872.9777.4578.5880.2084.3982.9485.9286.6987.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.369, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Table 1. Model accuracy (Dice score ↑ for the segmentation task, normalized to 100%) comparison of different methods for segmenting brain tumor subregions under various missing (○) and/or full modality settings. The best and second best performance of each subtask for each setting are highlighted in bold and underline respectively." + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.426, + 0.837, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.715, + 0.908, + 0.773 + ], + "angle": 0, + "content": "Figure 2. Visualization of modality weight changes (bold numbers in the upper right corner) assigned by the gating network alongside the predicted ET segmentation and uncertainty (entropy) maps when full or partial modalities (T1ce is missing) are present. DMoME with \\(\\mathcal{L}_{MoFe}\\) can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce) when T1ce is missing, which is aligned with clinical practice despite the network is not explicitly taught." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.482, + 0.829 + ], + "angle": 0, + "content": "noisy UPMC Food-101 dataset, where classification is more challenging, underscoring the robustness of our SimMLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.188, + 0.861 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of DMoME: To assess the unique contribution of DMoME, we trained the network w/o MoFe" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.907, + 0.89 + ], + "angle": 0, + "content": "loss: \\(\\mathcal{L}_{MoFe}\\) and compared its performance with competitive models—mmFormer [41], ShaSpec [35], and MoMKE [38]. Experiments were conducted on the BraTS 2018 dataset, encompassing 15 different configurations for both missing-modality and full-modality setups. Table 3 reports the average Dice scores across 15 scenarios, alongside" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24073" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.468, + 0.205 + ], + "angle": 0, + "content": "
DatasetModalitiesAccuracy (%) ↑
Modality 1Modality 2ShaSpec [35]MoMKE [38]SimMLM (Ours)
UPMC Food-101Image○69.2270.4672.20
○Text86.5586.5987.2
ImageText92.7392.7194.99
Average82.8383.2584.81
avMNISTImage○91.9092.6192.69
○Audio89.2891.1691.61
ImageAudio98.7198.6999.27
Average93.3094.1594.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.286 + ], + "angle": 0, + "content": "Table 2. Accuracy \\((\\%)\\) ↑ comparison across missing & full modality settings on UPMC Food-101 and avMNIST datasets. The datasets have different modality pairs (Image-Text for UPMC Food-101, Image-Audio for avMNIST). o represents modality missingness." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.301, + 0.466, + 0.39 + ], + "angle": 0, + "content": "
Dice score (%) ↑Model efficiency
MethodETTCWT#ParamsFLOPS
mmFormer [41]59.8572.9782.94106M748G
ShaSpec [35]61.5877.4585.92187.7M713G
MoMKE [38]65.5678.5886.697.8M490G
DMOE w/o LMoFe66.0579.1486.717.8M123G
DMOE w LMoFe (SimMLM)67.1680.2087.677.8M123G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Table 3. Comparison of model accuracy (Dice score \\(\\uparrow\\), normalized to \\(100\\%\\), averaged across all missing & full modality settings) and efficiency (# of parameters and FLOPS) on BraTS 2018 official-evaluated benchmark." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.485, + 0.483, + 0.729 + ], + "angle": 0, + "content": "networks' parameters and floating-point operations per second (FLOPS). As a result, DMoME achieves the highest Dice scores across all tumor subregions, with particularly strong performance in the more challenging enhancing tumor (ET) and tumor core (TC) segmentation. Importantly, DMoME achieves a fourfold reduction in FLOPS compared to MoMKE, even with the same expert architecture, and requires over 13 times fewer parameters than mmFormer and ShaSpec, underscoring its efficiency and scalability. This reduction in FLOPS is due to a key difference in processing: unlike MoMKE, which feeds each modality's input to all experts [38], our DMoME processes each input with only its dedicated expert. This simple approach significantly reduces FLOPS without compromising representational quality. Further details on DMoME's micro design are provided in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Interpretability of DMoME: Importantly, our network directly provides case-specific interpretability by dynamically revealing the importance of each modality, as shown in Fig. 2. In contrast, previous works such as mmFormer [41], ShaSpec [35], and MoMKE [38] lack this capability. We further performed population-wise study by plotting the average gating weights under different full- and missing-modality configurations on the official BraTS 2018 evaluation set in Fig. 3 across different tasks. It is evident that the gating network managed to dynamically adapt the weights to accommodate various scenarios, providing additional in" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.088, + 0.889, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.366, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Figure 3. Average gating weights analysis under both full and missing modality setups. Plotted values are the average gating weights on the BRATS 2018 official evaluation set." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.425, + 0.907, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.646, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Figure 4. Impact of the MoFe loss coefficient \\(\\lambda\\). The Dice score (\\%) \\(\\uparrow\\) is reported with different values of coefficients \\(\\lambda\\) for the MoFe loss across all missing & full modality settings on the BraTS 2018 validation set. The results include (1) enhancing tumor, (2) tumor core, (3) whole tumor, and (4) their average." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.907, + 0.82 + ], + "angle": 0, + "content": "sights for clinicians regarding modality importance in different subtasks. For instance, T1ce is prioritized for enhancing tumor (ET) segmentation, while Flair and T2 are more important for whole tumor (WT) segmentation, which is consistent with clinical findings [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness and Sensitivity of MoFe loss: We further analyzed the impact of MoFe loss by varying its coefficient \\(\\lambda = \\{0, 0.01, 0.05, 0.1, 0.2, 0.5, 1\\}\\) in Eq.4, where \\(\\lambda = 0\\) corresponds to training without it. As shown in Figure 4, the model achieved the highest average Dice score across" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24074" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.465, + 0.174 + ], + "angle": 0, + "content": "
ECE (%) ↓SCE (%) ↓
MethodETTCWTETTCWT
MoMKE [38]3.464.104.0414.8211.357.69
DMoME3.343.863.9813.4011.017.11
DMoME w/ LConf [19]3.173.863.6114.8011.586.81
DMoME w/ LMoFe (SimMLM)3.153.753.5513.2110.835.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.227 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of calibration errors (ECEs↓, SCEs↓) averaged across all 15 full and missing modality settings on the validation set of BraTS 2018 dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.269, + 0.483, + 0.435 + ], + "angle": 0, + "content": "the three subregions at \\(\\lambda = 0.1\\). Notably, most positive \\(\\lambda\\) values resulted in improvements over the baseline \\(\\lambda = 0\\), underscoring the stability and effectiveness of MoFe loss. This improvement may be attributed to the calibration nature of the loss, which enables DMoME to adaptively distinguish between modality-rich and modality-poor conditions, guiding the network to assess each modality's contribution across various scenarios. As a result, the model learns a more casual, robust relationship between diverse inputs and the task output. In the following, we will verify how \\(\\mathcal{L}_{MoFe}\\) can calibrate the network to achieve the goal." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.812 + ], + "angle": 0, + "content": "MoFe loss leads to better calibration: To evaluate the performance of \\(\\mathcal{L}_{MoFe}\\) for model calibration, we compute the expected calibration error (ECE) \\(\\downarrow\\) and static calibration error (SCE) \\(\\downarrow\\) [25], both using 20 bins, and compare our method to the most competitive method: MoMKE [38], in addition to the state-of-the-art multimodal calibration approaches, leveraging missing modality information [19], where the task losses \\(\\mathcal{L}_{task}\\) in \\(\\mathcal{L}_{MoFe}\\) have been replaced with confidence scores obtained from model predictions using max probability: \\(\\mathcal{L}_{Conf} = \\max(0, Conf(\\mathbf{o}^{-}) - Conf(\\mathbf{o}^{+}))\\). Results in Table 4 are scores directly computed on our validation set, as BraTS does not release ground truth labels on the evaluation set. All experiments were conducted in the same settings. It is evident that our method with \\(\\mathcal{L}_{MoFe}\\) achieves the best calibration performance in all settings, with the lowest ECEs and SCEs. In the context of this segmentation tasks, compared to \\(\\mathcal{L}_{Conf}\\), which is label-ignorant and performs comparisons at a low level (e.g., voxels), \\(\\mathcal{L}_{MoFe}\\) operates at the sample level, and leverages task-specific loss with ground truth information, making it more task-informative. This encourages the network to focus on task-relevant areas across varying levels of missing data, reducing sensitivity to background noise, augmentation artifacts, and outliers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "MoFe loss reduces model counterintuitive rate (CR): To directly quantify whether \\(\\mathcal{L}_{MoFe}\\) can better align the model performance with the modality availability (i.e., reduces cases where the model performs better even with fewer input modalities), we introduce a novel metric called the" + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.089, + 0.885, + 0.164 + ], + "angle": 0, + "content": "
Counterintuitive rate (%) ↓
MethodETTCWT
MoMKE [38]14.6032.479.86
DMoME w/o LMoFe10.4730.943.87
DMoME w/ LMoFe (SimMLM)7.2028.193.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.174, + 0.905, + 0.201 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of counterintuitive rate \\( \\downarrow \\) on the BraTS 2018 official evaluation set." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.226, + 0.829, + 0.24 + ], + "angle": 0, + "content": "Counterintuitive Rate (CR), which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.246, + 0.897, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\mathrm{CR} = \\frac{1}{N}\\sum_{i = 1}^{N}\\frac{1}{|\\mathcal{S}_{i}|}\\sum_{(\\mathbf{x}_{i}^{-},\\mathbf{x}_{i}^{+})\\in \\mathcal{S}_{i}}\\left\\{ \\begin{array}{ll}1 & \\text{if} P(\\mathbf{x}_{i}^{-}) > P(\\mathbf{x}_{i}^{+})\\\\ 0 & \\text{otherwise} \\end{array} \\right..\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.905, + 0.374 + ], + "angle": 0, + "content": "Here \\(N\\) is the total number of samples, \\(\\mathcal{S}_i\\) is the set of all more vs. fewer modality pairs \\((\\mathbf{x}_i^+ \\text{vs.} \\mathbf{x}_i^-)\\) for sample \\(i\\). \\(P(\\mathbf{x}_i)\\) denotes the model accuracy score (e.g., Dice score for the segmentation task) on the sample \\(\\mathbf{x}_i\\). An instance with \\(P(\\mathbf{x}_i^-) > P(\\mathbf{x}_i^+)\\) is considered as counterintuitive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.906, + 0.525 + ], + "angle": 0, + "content": "We compared CR on BraTS 2018 across three settings: MoMKE [38] (previous SOTA), DMoME without \\(\\mathcal{L}_{MoFe}\\) and DMoME with \\(\\mathcal{L}_{MoFe}\\) (SimMLM) and report the performance in Table 5. Results show that \\(\\mathcal{L}_{MoFe}\\) achieves the lowest CR, achieving \\(50.68\\%\\), \\(13.18\\%\\), and \\(61.97\\%\\) reduction for ET, TC, and WT segmentation, respectively, when compared to MoMKE. Notably, DMoME alone also reduces CR significantly, thanks to the dynamic weighting mechanism to automatically identify and prioritize more relevant modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Broader Impact: We have shown that SimMLM can provide direct importance values for each input modality and generate informative confidence maps for further analysis in the presence of incomplete data (Fig. 2). Such properties are particularly important to alert clinicians for further investigation in low-resource settings where certain modalities may be unavailable. Additionally, SimMLM's potential for handling varied levels of training-time missing modalities on avMNIST is also discussed in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced SimMLM, a simple, robust, and flexible framework for multimodal learning in scenarios with test-time missing modalities. By incorporating the Dynamic Mixture of Model Experts (DMoME) architecture and the novel More vs. Fewer (MoFe) ranking loss, SimMLM effectively addresses modality absence. Experimental results demonstrate that SimMLM achieves state-of-the-art performance across both segmentation and classification tasks, excelling in accuracy, efficiency, interpretability, robustness, and reliability. As a generic approach, future work will focus on extending SimMLM to other high-stakes real-world applications where trustworthiness matters and broader scenarios, including training-time missingness." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24075" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This research was supported by the UKRI funding (MR/Z505754/1), UK. Dr. Chen is supported by Royal Society (RGS/R2/242355)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.296 + ], + "angle": 0, + "content": "[1] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 447-456. Springer, 2019. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.297, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.395, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[3] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[4] Ignazio Gallo, Gianmarco Ria, Nicola Landro, and Riccardo La Grassa. Image and text fusion for upmc food-101 using bert and cnns. In 2020 35th International Conference on Image and Vision Computing New Zealand (IVCNZ), pages 1-6. IEEE, 2020. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Yunhe Gao. Training like a medical resident: Context-prior learning toward universal medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11194-11204, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[6] Shuang Hao, Wenfeng Han, Tao Jiang, Yiping Li, Haonan Wu, Chunlin Zhong, Zhangjun Zhou, and He Tang. Synthetic data in ai: Challenges, applications, and ethical implications. arXiv preprint arXiv:2401.01629, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[7] Minhao Hu, Matthias Maillard, Ya Zhang, Tommaso Ciceri, Giammarco La Barbera, Isabelle Bloch, and Pietro Gori. Knowledge distillation from multi-modal to mono-modal segmentation networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 772-781. Springer, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[8] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Donggeun Kim and Taesup Kim. Missing modality prediction for unpaired multimodal learning via joint embedding of unimodal models. arXiv preprint arXiv:2407.12616, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[11] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[13] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[14] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[15] Tao Liang, Guosheng Lin, Mingyang Wan, Tianrui Li, Guojun Ma, and Fengmao Lv. Expanding large pre-trained unimodal models with multimodal information injection for image-text multimodal classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15492-15501, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[16] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[17] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotaining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[18] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[19] Huan Ma, Qingyang Zhang, Changqing Zhang, Bingzhe Wu, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Calibrating multimodal learning. In International Conference on Machine Learning, pages 23429-23450. PMLR, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[20] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[21] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18177-18186, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24076" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11480-11491, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[25] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. In CVPR workshops, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[26] Jongyoun Noh, Hyekang Park, Junghyup Lee, and Bumsub Ham. Rankmixup: Ranking-based mixup training for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1358-1368, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[27] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems, 36:24829-24840, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[28] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[29] Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2515, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[30] Yujie Sun, Dongfang Sheng, Zihan Zhou, and Yifei Wu. AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content. *Humanities & social sciences communications*, 11(1): 1-14, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[32] Jonathan G Tullis and Robert L Goldstone. Comparison versus reminding. Cognitive research: principles and implications, 1(1):20, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] George Tzanetakis and Perry Cook. Musical genre classification of audio signals. IEEE Transactions on speech and audio processing, 10(5):293-302, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Valentin Vielzeuf, Alexis Lecheryv, Stéphane Pateux, and Frédéric Jurie. Centralnet: a multilayer approach for multimodal fusion. In Proceedings of the European Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.118 + ], + "angle": 0, + "content": "on Computer Vision (ECCV) Workshops, pages 0-0, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[35] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[36] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[37] Renjie Wu, Hu Wang, Hsiang-Ting Chen, and Gustavo Carneiro. Deep multimodal learning with missing modality: A survey, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[38] Wenxin Xu, Hexin Jiang, et al. Leveraging knowledge of modality experts for incomplete multimodal learning. In ACM Multimedia 2024, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[39] Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingxuan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tianlong Chen. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts. arXiv preprint arXiv:2410.08245, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[40] Zhilin Zeng, Zelin Peng, Xiaokang Yang, and Wei Shen. Missing as masking: Arbitrary cross-modal feature reconstruction for incomplete multimodal brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424-433. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[41] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[42] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[43] Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. Sparse moe with language guided routing for multilingual machine translation. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.796 + ], + "angle": 0, + "content": "[44] Yunpeng Zhao, Cheng Chen, Qing You Pang, Quanzheng Li, Carol Tang, Beng-Ti Ang, and Yueming Jin. Dealing with all-stage missing modality: Towards a universal model with robust reconstruction and personalization. arXiv preprint arXiv:2406.01987, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24077" + } + ] +] \ No newline at end of file diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_origin.pdf b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..530f73126b2af69b91e6cd2b258c88ea26f2e268 --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/398740ea-2adc-472d-af59-846c4c676c5a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:747c4845fd86a84305e533797dc11371edeb96638c9bff8de05ee8ff6b0d7682 +size 1325371 diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/full.md b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ec3518af07755065d154bf7cb118c5fd87fec20f --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/full.md @@ -0,0 +1,287 @@ +# SimMLM: A Simple Framework for Multi-modal Learning with Missing Modality + +Sijie Li Chen Chen† Jungong Han*† +School of Computer Science, University of Sheffield, UK +{sli256, chen.chen2, jungong.han}@sheffield.ac.uk + +# Abstract + +In this paper, we propose SimMLM, a simple yet powerful framework for multimodal learning with missing modalities. Unlike existing approaches that rely on sophisticated network architectures or complex data imputation techniques, SimMLM provides a generic and effective solution that can adapt to various missing modality scenarios with improved accuracy and robustness. Specifically, SimMLM consists of a generic Dynamic Mixture of Modality Experts (DMoME) architecture, featuring a dynamic, learnable gating mechanism that automatically adjusts each modality's contribution in both full and partial modality settings. A key innovation of SimMLM is the proposed More vs. Fewer (MoFe) ranking loss, which ensures that task accuracy improves or remains stable as more modalities are made available. This aligns the model with an intuitive principle: removing one or more modalities should not increase accuracy. We validate SimMLM on multimodal medical image segmentation (BraTS 2018) and multimodal classification (UPMC Food-101, avMNIST) tasks, where it consistently surpasses competitive methods, demonstrating superior accuracy, interpretability, robustness, and reliability across both complete and missing modality scenarios at test time. Code is available at https://github.com/LezJ/SimMLM. + +# 1. Introduction + +Multimodal learning aims to build models that can process and integrate information from multiple, complementary modalities such as visual, textual and auditory information, which has been found useful for various tasks such as image-text food classification [4, 15], multimodal medical image segmentation [5, 24] with improved model accuracy. However, in real-world scenarios, it is common for one or more modalities to be missing during deployment due to hardware failures, environmental conditions, or constraints + +in data collection, budget, and resources. Addressing this test-time missing modality problem is crucial for building robust multimodal models. + +One mainstream to address the missing modality problem is data imputation, which focuses on synthesizing the missing modalities from available ones [18, 44] or approximating intermediate missing features [10, 35, 36, 40]. However, ensuring the factuality and fidelity of synthetic data remains a major challenge, as issues such as poor data quality [16, 27], data hallucination [30], and vulnerability to adversarial corruption [6, 29] have been reported. Moreover, the computational overhead of generative imputation limits its practicality. In parallel, sophisticated frameworks have been developed, which aim to tackle the missing modality by leveraging advanced, complex network architectures [1, 35, 41], and harmonize or align early or late features from different modalities into a unified, shared space. However, such approaches often introduce substantial resource demands. While some are not adaptable to other tasks at all, others constrain the choice of network architectures or require additional modules, such as projection layers, to accommodate diverse or heterogeneous modalities. + +In this paper, we introduce SimMLM, a simple yet effective framework for multi-modal learning with missing modalities, without relying on specific neural network architectures or complex data imputation. First, to effectively handle varying modality availability, we propose a Dynamic Mixture of Modality Experts (DMoME) architecture, where modality-specific expert networks process available modalities, and a learnable gating network dynamically adjusts their contributions. This design ensures that the model remains effective in both full and partial modality settings while being adaptable to different neural architectures and task domains. + +To further improve the model's robustness and reliability against missing modalities, we introduce a novel loss function, which is designed to optimize the loss landscape around the region of missing data. Inspired by an intuitive principle that the model should not degrade in per + +formance when modalities are increased, this loss function, named More vs. Fewer (MoFe) ranking loss, ensures that as more modalities become available, the model's accuracy improves or at least remains stable. The MoFe loss encourages the model to compare performance between modality-rich and modality-poor pairs, promoting an inherent ranking that favors configurations with greater modality availability. This structured comparison in together with the dynamic weighting scheme in DMoME allows the model to learn consistent and reliable multimodal fusion strategies, improving robustness and adaptability across varying modality availability. + +Our key contributions are summarized as follows: + +- Dynamic Mixture of Modality Experts Architecture: SimMLM introduces the DMoME architecture with a learnable gating mechanism that adaptively adjusts modality contributions based on availability, improving performance in both complete and partial modality settings. This design enables seamless integration with various neural networks, ensuring flexibility and scalability across diverse tasks. +- More vs. Fewer ranking loss: We introduce the novel MoFe ranking loss, which ensures model performance improves or remains stable as additional modalities are incorporated. By enforcing a ranking comparison between modality-rich and modality-poor pairs, MoFe strengthens model robustness and reliability. +- Comprehensive Evaluation: We validate SimMLM on multimodal segmentation and classification across three datasets, demonstrating superior accuracy, interpretability, robustness, reliability, and computational efficiency over state-of-the-art methods. These results establish SimMLM as a practical and versatile solution for multimodal learning under missing modality conditions. + +# 2. Related Work + +To solve multi-modal learning with missing modalities, various approaches have been proposed, including a) input-level imputation, b) advanced representation learning; and c) mixture of expert strategies [37]. + +Input-level imputation approaches leverage advanced generative networks such as auto-encoders [18, 20], generative adversarial networks [20] to reconstruct missing inputs. Such an approach however, can generate noisy and unreliable input, as discussed in the previous section. + +Advanced representation learning focuses on learning improved representation either through advanced networks such as transformers [21, 41, 42], represented by mm-Former [41], or disentangling the latent features into modality-specific and modality-shared features for improving model accuracy, such as Robust-Mseg [1] and ShaSpec [35]. For instance, ShaSpec [35] employs a combination of modality-shared and modality-specific encoders + +with dedicated losses to promote feature disentanglement and perform feature fusion in the latent space before passing them to a task decoder. When a modality is missing, ShaSpec generates its features from the available modalities through a missing modality feature generation process enabled by the modality-shared encoder. However, this approach imposes a strict constraint that the shared encoder should be suitable for all input modalities and both shared and specific features must have the same dimensions, which can be less efficient and optimal when dealing with highly heterogeneous modalities. + +Mixture of Experts (MoE) has recently been adopted in both unimodal and multimodal learning, demonstrating superior performance in natural language processing [3, 14, 43] and computer vision tasks [28, 38, 39], while improving efficiency by reducing FLOPS. A very recent method for handling missing modalities is Mixture of Modality Knowledge Experts (MoMKE) [38], which uses MLP experts to generate unimodal and joint representations. Even when a modality is missing, MoMKE leverages existing modalities to generate joint representations by utilizing the other experts. While this approach has achieved promising results, it still incurs high computational costs, as each modality must pass through all modality experts, quadratically increasing complexity with the number of modalities. In contrast, the expert model in DMoME is dedicated to a specific modality. Additionally, unlike MoMKE, DMoME does not constrain the dimensionality of latent feature representations across modalities. Instead, it performs the weighting directly at the output space across experts. This flexibility makes DMoME more efficient and scalable, offering high performance in diverse multimodal tasks. + +# 3. Methodology + +# 3.1. Problem Definition + +Assume we have a complete multimodal dataset with $M$ modalities, denoted as $\mathcal{D} = \{\{\mathbf{x}_i^m\}_{m=1}^M, \mathbf{y}_i\}_{i=1}^N$ . where $\mathbf{x}_i^m \in \mathcal{X}$ represents the $i$ -th data sample and $m$ indexes the modality. $\mathbf{y}_i \in \mathcal{Y}$ is the label for the sample $i$ from the label space $\mathcal{Y}$ , which can either be a classification label or segmentation map. In this paper, we are interested in finding a solution $f$ that learns a mapping from the input space to the label space: $f: \mathcal{X} \to \mathcal{Y}$ , leveraging this complete modality dataset, so that $f$ can work robustly and reliably even when certain modalities $m$ are missing at inference time. For simplicity, we omit the sample index $i$ when not specified in the following sections. + +# 3.2. Overview of SimMLM + +Our proposed SimMLM framework consists of two main components: a) a unified Dynamic Mixture of Modality Experts framework (DMoME) (Sec. 3.3), which can take a + +![](images/c4217bb8ce1735c44a6a9b7f4f45ea9f262483f95132e8c717a95a677284c699.jpg) +(a) Our DMoME + +![](images/ef5a9060e9ee5cc89056730ebd33f3dddb9216d85afc909c9de40a5f67fa5726.jpg) +(b) Computation flows of $\mathcal{L}_{MoFe}$ +Figure 1. Our proposed SimMLM consists of a) DMoME architecture: DMoME takes either partial or full modality inputs where each modality is processed by a corresponding expert network to produce its own output $(\mathbf{o}^1, \ldots, \mathbf{o}^M)$ . A gating network (G) assesses the same inputs and produces weights to dynamically weigh the contributions of each expert's output in order to get the mixture of output $\mathbf{o}$ ; b) $\mathcal{L}_{\mathrm{MoFe}}$ loss: At training time, in addition to standard task specific losses, DMoME is optimized with the More $(X^{+})$ vs. Fewer Modality $(X^{-})$ (MoFe) ranking loss, directly encouraging the network to distinguish between inputs with differing modality availability at the task performance level. + +varied number of modalities as input and adaptively weight their contribution for robust performance at both training and inference time; b) a novel ranking loss called MoFe (Sec. 3.4), which is designed to calibrate the network's performance by differentiating network performance when the number of available modalities for the same sample varies. + +# 3.3. Dynamic Mixture of Modality Expert (DMoME) + +As shown in Figure 1 (a), Our DMoME consists of a set of modality-specific expert networks $\{E^{m}(\mathbf{x}_{m};\theta_{m})\}_{m = 1}^{M}$ each containing its own set of learnable parameters $\theta_{m}$ and a gating network, $G(\cdot ;\phi)$ , with a set of learnable parameters $\phi$ . Each expert network, $E^{m}(\mathbf{x}_{m};\theta_{m})$ , takes its corresponding modality as input and produces its output in logits $\mathbf{o}^m$ , before applying the softmax function: $\mathbf{o}^m = E^m (\mathbf{x}^m;\theta^m)\in \mathbb{R}^T$ . Here, $T$ denotes the dimensionality of the output space for a specific task, e.g., the number of predicted classes in a classification task. + +The gating network $G(\cdot; \phi)$ takes a set of different modalities as input and produces a set of gating values $\{g^m\}_{m=1}^M$ to calibrate different modality experts' output. The softmax function is then applied to computing non-negative weights $\{w_1, w_2, \ldots, w_M\} \in \mathbb{R}$ , ensuring that the sum of the weights for existing modalities is equal to 1. The final output can be described as a reweighted combination of the outputs from the corresponding modality experts $\{\mathbf{o}^m\}_{m=1}^M$ : $\mathbf{o} = \sum_{m=1}^M w^m E^m(\mathbf{x}^m; \theta_m)$ . + +Notably, the input to the gating network can be an incomplete set of modalities. If certain modalities are miss + +ing, those inputs are treated as 0, and the gating values for the corresponding missing modality experts' outputs are set to $-\infty$ (which causes the corresponding weight values to be 0 after the softmax). Mathematically, the gating outputs are set as: + +$$ +g _ {m} = \left\{ \begin{array}{l l} G (\mathbf {x}; \phi) ^ {(m)}, & \text {i f m o d a l i t y m i s p r e s e n t} \\ - \infty , & \text {i f m o d a l i t y m i s m i s s i n g} \end{array} , \right. \tag {1} +$$ + +where $\mathbf{x}$ can be a complete set of modalities or an arbitrary incomplete set of modalities for a given sample. The final weights therefore can be simply computed as: + +$$ +w ^ {m} = \frac {\exp \left(g ^ {m}\right)}{\sum_ {m ^ {\prime} = 1} ^ {M} \exp \left(g ^ {m ^ {\prime}}\right)} \quad \text {f o r} \quad m = 1, 2, \dots , M. \tag {2} +$$ + +Here, the weight vector $\mathbf{w} = [w^{1}, w^{2}, \ldots, w^{M}] \in \mathbb{R}^{M}$ is non-negative and sums to 1, with $w^{m} = 0$ for missing modalities, effectively excluding them from the final mixture. The above configuration can be easily extended to multi-task settings, where each task has its own task-specific weighting scheme. In this case, the gating network produces $\mathbf{w} \in \mathbb{R}^{M \times K}$ , with $K$ being the number of tasks. + +# Two-stage training + +To train the whole network, we adopt a two-stage training process, which consists of: + +1. independent learning: we independently train a modality-specific expert $E_{m}$ for each modality $m$ , supervised by a task-specific loss $\ell_{\mathrm{task}}(\mathbf{o}^{m},\mathbf{y})$ . For example, for classification tasks, the loss function can be the cross-entropy (CE) loss, whereas for the segmentation, + +the loss can be Dice loss. This pretraining stage ensures that each modality expert independently learns specialized features, minimizing the risk of any single modality dominating the learning process. Different modalities can vary significantly in learning dynamics and task difficulty, which can lead to training instability, especially at the early stages of joint training. + +2. Cooperative learning: we co-train the expert networks alongside the gating network, ranking network performance under More vs. Fewer Modality conditions, see next section for details. + +This training scheme enables the expert networks to exploit modality-specific information without interference from other modalities before jointly optimizing with the gating network to exploit mutual benefits. It also improves framework flexibility with reduced training resources, e.g., a) first-stage training can be accelerated via parallel computing; b) when new modalities are introduced, existing pretrained expert networks can be reused for the second stage. + +# 3.4. More vs. Fewer Modality Co-training + +To boost model robustness against missing modalities, we simulate scenarios with missing modalities by randomly dropping out input modalities during training, which is a common practice. Yet, unlike existing approaches that focus on aligning outputs from missing modalities with those from the full set of modalities [7, 9, 17], we encourage the network to distinguish between modality-rich and modality-poor inputs. + +Rational behind More vs. Fewer Modality Training: In many multi-modal learning tasks, all modalities are assumed to be predictive and complementary, contributing valuable information to the tasks that are of interest. Our approach is inspired by the intuition that richer multimodal inputs should yield more accurate results, as each modality complements the others. As the number of modalities decreases, the input becomes more limited, making it more challenging to capture complex patterns. + +As it is difficult to quantify the exact performance for different set-ups, we propose to formulate it as a ranking loss to encourage the network to compare the two scenarios. In cognitive science, it has been found that learners who engage in comparative tasks show an improved ability to transfer knowledge to new situations compared to those who learn through isolated examples [32]. Such a concept has also been applied in reinforcement learning for preference alignment for the applications like large language model training and finetuning with a reward function. + +In our multimodal learning setting, at each training iteration, we sample two sets from a data point $\mathbf{x}_{full}$ : $\mathbf{x}^{+}$ and $\mathbf{x}^{-}$ , where $\mathbf{x}^{-} \subset \mathbf{x}^{+} \subseteq \mathbf{x}^{full}$ . The corresponding outputs from DMoME are denoted as $\mathbf{o}^{+}$ and $\mathbf{o}^{-}$ , respectively. We conjecture that the task-specific loss value for modality- + +rich setups with $\mathbf{x}^{+}$ should be at least no higher than those in the modality $\mathbf{x}^{-}$ . Mathematically, every pair of samples $(\mathbf{x}^{+},\mathbf{x}^{-})$ should satisfy the following relationship: + +$$ +\mathcal {L} _ {\text {t a s k}} \left(\mathbf {o} ^ {-}, \mathbf {y}\right) \geq \mathcal {L} _ {\text {t a s k}} \left(\mathbf {o} ^ {+}, \mathbf {y}\right) \iff A c c \left(\mathbf {x} ^ {+}\right) \geq A c c \left(\mathbf {x} ^ {-}\right). +$$ + +Definition of $\mathcal{L}_{MoFe}$ : To guide the network to learn the ordinary relationship, we introduce More vs. Fewer (MoFe) ranking loss, defined as follows: + +$$ +\ell_ {\mathrm {M o F e}} \left(\mathbf {o} ^ {+}, \mathbf {o} ^ {-}, \mathbf {y}\right) = \max \left(0, \ell_ {\text {t a s k}} \left(\mathbf {o} ^ {+}, \mathbf {y}\right) - \ell_ {\text {t a s k}} \left(\mathbf {o} ^ {-}, \mathbf {y}\right)\right), \tag {3} +$$ + +which encourages the network to perform better when more modalities are available. It is worth mentioning that the proposed loss is general and thus can be applied to other existing multi-modal learning frameworks to better calibrate the network for improved robustness against missing modalities as well as improved reliability. The total training loss is thus defined as: + +$$ +\ell_ {\text {t o t a l}} \left(\mathbf {o} ^ {+}, \mathbf {o} ^ {-}, \mathbf {y}\right) = \ell_ {\text {t a s k}} \left(\mathbf {o} ^ {+}, \mathbf {y}\right) + \ell_ {\text {t a s k}} \left(\mathbf {o} ^ {-}, \mathbf {y}\right) + \lambda \ell_ {\mathrm {M o F e}}, \tag {4} +$$ + +where $\lambda$ is a coefficient to control the contribution of $\ell_{\mathrm{MOFE}}$ . + +We acknowledge the existence of prior research, such as the works of Noh et al. (2023) [26], and Ma et al. (2023) [19], which have investigated the application of ranking loss at the confidence levels between a pair of two inputs. However, our approach diverges significantly from these existing methods. Specifically, we propose the novel concept of applying ranking loss directly at the loss level, as opposed to the confidence level. This distinction is crucial, as our method aims to directly regularize the geometry of the loss landscape. In our experiments, we also showed extra benefit of our loss, which leads to reduced calibration errors, see Table. 4. Besides, our proposed method enables seamless integration with a range of loss functions, offering flexibility beyond classification and allowing for improved applicability in complex, diverse tasks. For example, our method can be applied for segmentation and regression tasks where the confidence scores are not straightforward in dense or continuous prediction settings. By focusing on the actual task error, our loss-level approach directly aligns with the downstream task performance, providing a more consistent measure across different tasks. + +# 4. Experiments + +We evaluate our approach on two types of tasks, multimodal medical segmentation and multimodal classification, spanning three datasets. For segmentation, we use BraTS 2018 [22] benchmark. For classification, we conduct experiments on UPMC Food-101 [4] and avMNIST dataset [34]. + +# 4.1. Datasets + +BraTS 2018 (Brain Tumor Segmentation Challenge [22]) is a widely used benchmark for brain tumor segmenta + +tion in 3D MRI scans. It has also been extensively employed to evaluate methods for handling missing modalities [7, 35, 41, 42]. The dataset includes four distinct MRI modalities: FLAIR, T1, T1 contrast-enhanced (T1ce), and T2. The segmentation task involves identifying three tumor subregions: the enhancing tumor (ET), the tumor core (TC), and the whole tumor (WT). The dataset consists of 285 manually labeled cases and 66 unlabeled cases for evaluation. In our experiments, we split the labeled cases into a 4:1 ratio for training and validation. Final evaluations are conducted on the 66 unlabeled cases via the official benchmarking platform for a consistent and fair accuracy comparison against existing methods [13, 41]. + +UPMC Food-101 [4] is a large-scale food classification dataset with 90,704 image-text pairs spanning 101 classes. Collected from uncontrolled environments, it introduces inherent noise, posing challenges for accurate classification. + +avMNIST (Audiovision MNIST [34]) is a digit classification dataset comprising 1,500 image-audio pairs. The visual modality consists of $28 \times 28$ grayscale images of handwritten digits (0-9) from MNIST [12], while the audio modality is sourced from the Free Spoken Digits Dataset*. + +# 4.2. Implementation Details + +All the experiments were conducted on a single A100 GPU node. Detailed implementations are described below. + +BraTS 2018: We employed nnUNet [8], a widely used and powerful medical image segmentation backbone, as the expert network for each modality, which can be viewed as a 3D U-Net. The gating network $G$ in DMoME utilized lightweight CNNs to capture broad contextual information, followed by a linear layer generating $4 \times 3$ weights for the four modalities and three segmentation tasks. We optimized both training stages using the Adam optimizer [11] ( $lr = 0.01$ ). The $\mathcal{L}_{task}$ in Eq.4 combined Dice [23] and binary cross-entropy losses. + +UPMC Food-101: We employed Inception V3 [31] as the image expert and BERT [2] as the text expert, following [4]. To avoid redundant feature extraction, intermediate features from both were fed into an MLP gating network $G$ , which generated 2D weights for expert outputs. Optimization was performed using Adam ( $lr = 0.0001$ ) with cross-entropy loss as $\mathcal{L}_{task}$ . + +avMNIST: Following [20, 35], we adopted a LeNet-5 [12] backbone for both image and audio modalities, with audio converted into $20 \times 20$ MFCC spectrograms [33]. The gating network $G$ followed a similar design as in BraTS 2018. Two-stage training was conducted using Adam ( $lr = 0.001$ ) with cross-entropy loss as $\mathcal{L}_{task}$ . + +Detailed gating network $G$ architectures for BraTS 2018 and avMNIST are provided in the Supplementary. For task accuracy evaluation, we follow common practice, using the + +Dice score for segmentation (BraTS 2018) and Accuracy for classification (UPMC Food-101, avMNIST). The MoFe loss coefficient $\lambda$ in Eq.4 is set to 0.1 across tasks; however, our method remains robust to variations in $\lambda$ , as demonstrated in Sec. 5 Analysis. + +# 4.3. Segmentation Results + +Benchmarking segmentation accuracy on the official BraTS 2018 evaluation set. For the BraTS 2018 benchmark, we compare our approach against several state-of-the-art methods, including Robust-MSeg (RbSeg) [1], mmFormer (mmFm) [41], ShaSpec [35], and MoMKE [38] under 15 configurations of both missing and full modality setups in Table 1. Results show that SimMLM achieves the highest average accuracy in terms of the Dice scores, consistently ranking as the top or second-highest performer in each entry across 44 out of all 45 configurations (15 settings $\times$ 3 subtasks), despite utilizing only $80\%$ of the labeled data compared to the $100\%$ used by ShaSpec and mmFormer. This consistent performance highlights our model's robustness and data efficiency in handling various incomplete modality inputs. Notably, when only the T1ce modality is available (Table 1, row 3), our model significantly improves the Dice score compared to the second-best methods, increasing from 73.29 to 78.58 for ET, 78.71 to 83.45 for TC, and 74.00 to 80.02 for WT. Such results underscore the advantage of our approach even in single-modality settings. + +As shown in Figure 2, in both full and partial modality scenarios, SimMLM (DMoME with $\mathcal{L}_{MoFe}$ ) achieves more accurate segmentation results with better model reliability compared to DMoME alone (A vs. B, C vs. D) thanks to the contribution of $\mathcal{L}_{MoFe}$ . When T1ce is unavailable, DMoME with $\mathcal{L}_{MoFe}$ can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce), which is aligned with clinical practice despite the network is not explicitly taught. Without MoFe loss, the network is biased to T2 (maybe due to higher intensity) and over-segment the tumor. In the following Sec. 5, we will provide a more detailed quantitative analysis for our method. + +# 4.4. Classification Results + +Benchmarking classification accuracy on UPMC Food101 and avMNIST. For the classification tasks, we compared our model to the two most competitive methods: ShaSpec [35] and MoMKE [38]. To ensure fair comparisons, all models were trained under identical conditions. The results, presented in Table 2, report accuracy across different modality settings. Our method consistently surpasses SOTA methods across all settings, demonstrating its effectiveness in handling both missing and full modality scenarios. While clear improvements are observed on both datasets, our method achieves even greater gains on the + +
SettingsAvailable Modalities at Test TimeEnhancing tumor (ET)Tumor core (TC)Whole tumor (WT)
◇◇◇RbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]Ours
Missing1●○○25.6939.3343.5250.5351.3153.5761.2169.4475.0574.1285.6986.1088.6889.7889.24
2○●○17.2932.5341.0048.1248.4147.9056.5563.1868.6671.4870.1167.5273.4477.4580.28
3○○●67.0772.6073.2972.3078.5876.8375.4178.6578.7183.4573.3172.2273.8274.0080.02
4○○○28.9743.0546.3151.0250.6057.4964.2069.0372.3473.4982.2481.1583.9985.8286.01
5●●○32.1342.9644.7652.5253.6260.6865.9172.6775.0077.1488.2487.0689.7690.2389.98
6●○●70.3075.0775.6078.7580.2580.6277.8884.5084.6284.9688.5187.3090.0690.3490.35
7●○○33.8447.5247.2054.1254.6261.1669.7572.9376.9075.7688.2887.5990.0290.5390.28
8○●●69.0674.0475.7676.7878.7778.7278.5982.1081.7084.7977.1874.4278.7479.8682.25
9○●○32.0144.9946.8453.7453.4062.1969.4271.3873.7374.7384.7882.2086.0386.6087.42
10○○●69.7174.5175.9577.1178.6180.2078.6183.8283.6184.1985.1982.9985.4286.9288.11
11●●●70.7875.4776.4278.1980.4181.0679.8085.2382.8385.2588.7387.3390.2990.2190.63
12●●○36.4147.7046.5555.9355.8464.3871.5273.9777.0078.0288.8187.7590.3690.5990.62
13●○●70.8875.6775.9978.4480.2980.7279.5585.2683.5985.2389.2788.1490.7890.8590.99
14○●●70.1074.7576.3778.2780.3380.3380.3984.1882.3585.0886.0182.7186.4786.6287.78
Full15●●●71.1377.6178.0877.6580.4080.8685.7885.4582.6185.3989.4589.6490.8890.5491.07
Average51.0259.8561.5865.5667.1669.7872.9777.4578.5880.2084.3982.9485.9286.6987.67
+ +Table 1. Model accuracy (Dice score ↑ for the segmentation task, normalized to 100%) comparison of different methods for segmenting brain tumor subregions under various missing (○) and/or full modality settings. The best and second best performance of each subtask for each setting are highlighted in bold and underline respectively. + +![](images/d289ce2011664d25e89a2713d2558cf8757c63c5039de3c1496049502d016106.jpg) +Figure 2. Visualization of modality weight changes (bold numbers in the upper right corner) assigned by the gating network alongside the predicted ET segmentation and uncertainty (entropy) maps when full or partial modalities (T1ce is missing) are present. DMoME with $\mathcal{L}_{MoFe}$ can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce) when T1ce is missing, which is aligned with clinical practice despite the network is not explicitly taught. + +noisy UPMC Food-101 dataset, where classification is more challenging, underscoring the robustness of our SimMLM. + +# 5. Analysis + +Effectiveness of DMoME: To assess the unique contribution of DMoME, we trained the network w/o MoFe + +loss: $\mathcal{L}_{MoFe}$ and compared its performance with competitive models—mmFormer [41], ShaSpec [35], and MoMKE [38]. Experiments were conducted on the BraTS 2018 dataset, encompassing 15 different configurations for both missing-modality and full-modality setups. Table 3 reports the average Dice scores across 15 scenarios, alongside + +
DatasetModalitiesAccuracy (%) ↑
Modality 1Modality 2ShaSpec [35]MoMKE [38]SimMLM (Ours)
UPMC Food-101Image○69.2270.4672.20
○Text86.5586.5987.2
ImageText92.7392.7194.99
Average82.8383.2584.81
avMNISTImage○91.9092.6192.69
○Audio89.2891.1691.61
ImageAudio98.7198.6999.27
Average93.3094.1594.52
+ +Table 2. Accuracy $(\%)$ ↑ comparison across missing & full modality settings on UPMC Food-101 and avMNIST datasets. The datasets have different modality pairs (Image-Text for UPMC Food-101, Image-Audio for avMNIST). o represents modality missingness. + +
Dice score (%) ↑Model efficiency
MethodETTCWT#ParamsFLOPS
mmFormer [41]59.8572.9782.94106M748G
ShaSpec [35]61.5877.4585.92187.7M713G
MoMKE [38]65.5678.5886.697.8M490G
DMOE w/o LMoFe66.0579.1486.717.8M123G
DMOE w LMoFe (SimMLM)67.1680.2087.677.8M123G
+ +Table 3. Comparison of model accuracy (Dice score $\uparrow$ , normalized to $100\%$ , averaged across all missing & full modality settings) and efficiency (# of parameters and FLOPS) on BraTS 2018 official-evaluated benchmark. + +networks' parameters and floating-point operations per second (FLOPS). As a result, DMoME achieves the highest Dice scores across all tumor subregions, with particularly strong performance in the more challenging enhancing tumor (ET) and tumor core (TC) segmentation. Importantly, DMoME achieves a fourfold reduction in FLOPS compared to MoMKE, even with the same expert architecture, and requires over 13 times fewer parameters than mmFormer and ShaSpec, underscoring its efficiency and scalability. This reduction in FLOPS is due to a key difference in processing: unlike MoMKE, which feeds each modality's input to all experts [38], our DMoME processes each input with only its dedicated expert. This simple approach significantly reduces FLOPS without compromising representational quality. Further details on DMoME's micro design are provided in the Supplementary. + +Interpretability of DMoME: Importantly, our network directly provides case-specific interpretability by dynamically revealing the importance of each modality, as shown in Fig. 2. In contrast, previous works such as mmFormer [41], ShaSpec [35], and MoMKE [38] lack this capability. We further performed population-wise study by plotting the average gating weights under different full- and missing-modality configurations on the official BraTS 2018 evaluation set in Fig. 3 across different tasks. It is evident that the gating network managed to dynamically adapt the weights to accommodate various scenarios, providing additional in + +![](images/26c4c60d7b0cb0da60794d819fd7123c11f1ae8ddbe76bd804d37aa1659f48f6.jpg) +Figure 3. Average gating weights analysis under both full and missing modality setups. Plotted values are the average gating weights on the BRATS 2018 official evaluation set. + +![](images/d2ff5ffda69616017def2d3d5e84801cd0719721b7f3746c221ac24c1133147a.jpg) +Figure 4. Impact of the MoFe loss coefficient $\lambda$ . The Dice score (\%) $\uparrow$ is reported with different values of coefficients $\lambda$ for the MoFe loss across all missing & full modality settings on the BraTS 2018 validation set. The results include (1) enhancing tumor, (2) tumor core, (3) whole tumor, and (4) their average. + +sights for clinicians regarding modality importance in different subtasks. For instance, T1ce is prioritized for enhancing tumor (ET) segmentation, while Flair and T2 are more important for whole tumor (WT) segmentation, which is consistent with clinical findings [22]. + +Effectiveness and Sensitivity of MoFe loss: We further analyzed the impact of MoFe loss by varying its coefficient $\lambda = \{0, 0.01, 0.05, 0.1, 0.2, 0.5, 1\}$ in Eq.4, where $\lambda = 0$ corresponds to training without it. As shown in Figure 4, the model achieved the highest average Dice score across + +
ECE (%) ↓SCE (%) ↓
MethodETTCWTETTCWT
MoMKE [38]3.464.104.0414.8211.357.69
DMoME3.343.863.9813.4011.017.11
DMoME w/ LConf [19]3.173.863.6114.8011.586.81
DMoME w/ LMoFe (SimMLM)3.153.753.5513.2110.835.65
+ +the three subregions at $\lambda = 0.1$ . Notably, most positive $\lambda$ values resulted in improvements over the baseline $\lambda = 0$ , underscoring the stability and effectiveness of MoFe loss. This improvement may be attributed to the calibration nature of the loss, which enables DMoME to adaptively distinguish between modality-rich and modality-poor conditions, guiding the network to assess each modality's contribution across various scenarios. As a result, the model learns a more casual, robust relationship between diverse inputs and the task output. In the following, we will verify how $\mathcal{L}_{MoFe}$ can calibrate the network to achieve the goal. + +MoFe loss leads to better calibration: To evaluate the performance of $\mathcal{L}_{MoFe}$ for model calibration, we compute the expected calibration error (ECE) $\downarrow$ and static calibration error (SCE) $\downarrow$ [25], both using 20 bins, and compare our method to the most competitive method: MoMKE [38], in addition to the state-of-the-art multimodal calibration approaches, leveraging missing modality information [19], where the task losses $\mathcal{L}_{task}$ in $\mathcal{L}_{MoFe}$ have been replaced with confidence scores obtained from model predictions using max probability: $\mathcal{L}_{Conf} = \max(0, Conf(\mathbf{o}^{-}) - Conf(\mathbf{o}^{+}))$ . Results in Table 4 are scores directly computed on our validation set, as BraTS does not release ground truth labels on the evaluation set. All experiments were conducted in the same settings. It is evident that our method with $\mathcal{L}_{MoFe}$ achieves the best calibration performance in all settings, with the lowest ECEs and SCEs. In the context of this segmentation tasks, compared to $\mathcal{L}_{Conf}$ , which is label-ignorant and performs comparisons at a low level (e.g., voxels), $\mathcal{L}_{MoFe}$ operates at the sample level, and leverages task-specific loss with ground truth information, making it more task-informative. This encourages the network to focus on task-relevant areas across varying levels of missing data, reducing sensitivity to background noise, augmentation artifacts, and outliers. + +MoFe loss reduces model counterintuitive rate (CR): To directly quantify whether $\mathcal{L}_{MoFe}$ can better align the model performance with the modality availability (i.e., reduces cases where the model performs better even with fewer input modalities), we introduce a novel metric called the + +Table 4. Quantitative comparison of calibration errors (ECEs↓, SCEs↓) averaged across all 15 full and missing modality settings on the validation set of BraTS 2018 dataset. + +
Counterintuitive rate (%) ↓
MethodETTCWT
MoMKE [38]14.6032.479.86
DMoME w/o LMoFe10.4730.943.87
DMoME w/ LMoFe (SimMLM)7.2028.193.75
+ +Table 5. Quantitative comparison of counterintuitive rate $\downarrow$ on the BraTS 2018 official evaluation set. + +Counterintuitive Rate (CR), which is defined as + +$$ +\mathrm{CR} = \frac{1}{N}\sum_{i = 1}^{N}\frac{1}{|\mathcal{S}_{i}|}\sum_{(\mathbf{x}_{i}^{-},\mathbf{x}_{i}^{+})\in \mathcal{S}_{i}}\left\{ \begin{array}{ll}1 & \text{if} P(\mathbf{x}_{i}^{-}) > P(\mathbf{x}_{i}^{+})\\ 0 & \text{otherwise} \end{array} \right.. +$$ + +Here $N$ is the total number of samples, $\mathcal{S}_i$ is the set of all more vs. fewer modality pairs $(\mathbf{x}_i^+ \text{vs.} \mathbf{x}_i^-)$ for sample $i$ . $P(\mathbf{x}_i)$ denotes the model accuracy score (e.g., Dice score for the segmentation task) on the sample $\mathbf{x}_i$ . An instance with $P(\mathbf{x}_i^-) > P(\mathbf{x}_i^+)$ is considered as counterintuitive. + +We compared CR on BraTS 2018 across three settings: MoMKE [38] (previous SOTA), DMoME without $\mathcal{L}_{MoFe}$ and DMoME with $\mathcal{L}_{MoFe}$ (SimMLM) and report the performance in Table 5. Results show that $\mathcal{L}_{MoFe}$ achieves the lowest CR, achieving $50.68\%$ , $13.18\%$ , and $61.97\%$ reduction for ET, TC, and WT segmentation, respectively, when compared to MoMKE. Notably, DMoME alone also reduces CR significantly, thanks to the dynamic weighting mechanism to automatically identify and prioritize more relevant modalities. + +Broader Impact: We have shown that SimMLM can provide direct importance values for each input modality and generate informative confidence maps for further analysis in the presence of incomplete data (Fig. 2). Such properties are particularly important to alert clinicians for further investigation in low-resource settings where certain modalities may be unavailable. Additionally, SimMLM's potential for handling varied levels of training-time missing modalities on avMNIST is also discussed in the Supplementary. + +# 6. Conclusion + +In this paper, we introduced SimMLM, a simple, robust, and flexible framework for multimodal learning in scenarios with test-time missing modalities. By incorporating the Dynamic Mixture of Model Experts (DMoME) architecture and the novel More vs. Fewer (MoFe) ranking loss, SimMLM effectively addresses modality absence. Experimental results demonstrate that SimMLM achieves state-of-the-art performance across both segmentation and classification tasks, excelling in accuracy, efficiency, interpretability, robustness, and reliability. As a generic approach, future work will focus on extending SimMLM to other high-stakes real-world applications where trustworthiness matters and broader scenarios, including training-time missingness. + +# Acknowledgements + +This research was supported by the UKRI funding (MR/Z505754/1), UK. Dr. Chen is supported by Royal Society (RGS/R2/242355). + +# References + +[1] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 447-456. Springer, 2019. 1, 2, 5, 6 +[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 5 +[3] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2 +[4] Ignazio Gallo, Gianmarco Ria, Nicola Landro, and Riccardo La Grassa. Image and text fusion for upmc food-101 using bert and cnns. In 2020 35th International Conference on Image and Vision Computing New Zealand (IVCNZ), pages 1-6. IEEE, 2020. 1, 4, 5 +[5] Yunhe Gao. Training like a medical resident: Context-prior learning toward universal medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11194-11204, 2024. 1 +[6] Shuang Hao, Wenfeng Han, Tao Jiang, Yiping Li, Haonan Wu, Chunlin Zhong, Zhangjun Zhou, and He Tang. Synthetic data in ai: Challenges, applications, and ethical implications. arXiv preprint arXiv:2401.01629, 2024. 1 +[7] Minhao Hu, Matthias Maillard, Ya Zhang, Tommaso Ciceri, Giammarco La Barbera, Isabelle Bloch, and Pietro Gori. Knowledge distillation from multi-modal to mono-modal segmentation networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 772-781. Springer, 2020. 4, 5 +[8] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 5 +[9] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 4 +[10] Donggeun Kim and Taesup Kim. Missing modality prediction for unpaired multimodal learning via joint embedding of unimodal models. arXiv preprint arXiv:2407.12616, 2024. 1 + +[11] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[12] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5 +[13] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 5 +[14] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2 +[15] Tao Liang, Guosheng Lin, Mingyang Wan, Tianrui Li, Guojun Ma, and Fengmao Lv. Expanding large pre-trained unimodal models with multimodal information injection for image-text multimodal classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15492-15501, 2022. 1 +[16] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 1 +[17] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotaining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 4 +[18] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2 +[19] Huan Ma, Qingyang Zhang, Changqing Zhang, Bingzhe Wu, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Calibrating multimodal learning. In International Conference on Machine Learning, pages 23429-23450. PMLR, 2023. 4, 8 +[20] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 5 +[21] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18177-18186, 2022. 2 +[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya + +Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 4, 7 +[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5 +[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11480-11491, 2024. 1 +[25] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. In CVPR workshops, 2019. 8 +[26] Jongyoun Noh, Hyekang Park, Junghyup Lee, and Bumsub Ham. Rankmixup: Ranking-based mixup training for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1358-1368, 2023. 4 +[27] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems, 36:24829-24840, 2023. 1 +[28] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2 +[29] Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2515, 2024. 1 +[30] Yujie Sun, Dongfang Sheng, Zihan Zhou, and Yifei Wu. AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content. *Humanities & social sciences communications*, 11(1): 1-14, 2024. 1 +[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5 +[32] Jonathan G Tullis and Robert L Goldstone. Comparison versus reminding. Cognitive research: principles and implications, 1(1):20, 2016. 4 +[33] George Tzanetakis and Perry Cook. Musical genre classification of audio signals. IEEE Transactions on speech and audio processing, 10(5):293-302, 2002. 5 +[34] Valentin Vielzeuf, Alexis Lecheryv, Stéphane Pateux, and Frédéric Jurie. Centralnet: a multilayer approach for multimodal fusion. In Proceedings of the European Conference + +on Computer Vision (ECCV) Workshops, pages 0-0, 2018. 4, 5 +[35] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 1, 2, 5, 6, 7 +[36] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36, 2024. 1 +[37] Renjie Wu, Hu Wang, Hsiang-Ting Chen, and Gustavo Carneiro. Deep multimodal learning with missing modality: A survey, 2024. 2 +[38] Wenxin Xu, Hexin Jiang, et al. Leveraging knowledge of modality experts for incomplete multimodal learning. In ACM Multimedia 2024, 2024. 2, 5, 6, 7, 8 +[39] Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingxuan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tianlong Chen. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts. arXiv preprint arXiv:2410.08245, 2024. 2 +[40] Zhilin Zeng, Zelin Peng, Xiaokang Yang, and Wei Shen. Missing as masking: Arbitrary cross-modal feature reconstruction for incomplete multimodal brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424-433. Springer, 2024. 1 +[41] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 5, 6, 7 +[42] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2, 5 +[43] Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. Sparse moe with language guided routing for multilingual machine translation. In The Twelfth International Conference on Learning Representations, 2023. 2 +[44] Yunpeng Zhao, Cheng Chen, Qing You Pang, Quanzheng Li, Carol Tang, Beng-Ti Ang, and Yueming Jin. Dealing with all-stage missing modality: Towards a universal model with robust reconstruction and personalization. arXiv preprint arXiv:2406.01987, 2024. 1 \ No newline at end of file diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/images.zip b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cd4ec2129fa494cf48539c5332b9cdfff3e3055a --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa30d23d38d6a1111ac14f92869be04898b3f596b35399e55c7f9518c52e4c03 +size 500836 diff --git a/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/layout.json b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..287f0a4e3131372339615956901fd60149f0eb9a --- /dev/null +++ b/2025/SimMLM_ A Simple Framework for Multi-modal Learning with Missing Modality/layout.json @@ -0,0 +1,8388 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 102, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 102, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 81, + 102, + 531, + 140 + ], + "type": "text", + "content": "SimMLM: A Simple Framework for Multi-modal Learning with Missing Modality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 160, + 444, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 160, + 444, + 203 + ], + "spans": [ + { + "bbox": [ + 165, + 160, + 444, + 203 + ], + "type": "text", + "content": "Sijie Li Chen Chen† Jungong Han*† \nSchool of Computer Science, University of Sheffield, UK \n{sli256, chen.chen2, jungong.han}@sheffield.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 257, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 295, + 531 + ], + "type": "text", + "content": "In this paper, we propose SimMLM, a simple yet powerful framework for multimodal learning with missing modalities. Unlike existing approaches that rely on sophisticated network architectures or complex data imputation techniques, SimMLM provides a generic and effective solution that can adapt to various missing modality scenarios with improved accuracy and robustness. Specifically, SimMLM consists of a generic Dynamic Mixture of Modality Experts (DMoME) architecture, featuring a dynamic, learnable gating mechanism that automatically adjusts each modality's contribution in both full and partial modality settings. A key innovation of SimMLM is the proposed More vs. Fewer (MoFe) ranking loss, which ensures that task accuracy improves or remains stable as more modalities are made available. This aligns the model with an intuitive principle: removing one or more modalities should not increase accuracy. We validate SimMLM on multimodal medical image segmentation (BraTS 2018) and multimodal classification (UPMC Food-101, avMNIST) tasks, where it consistently surpasses competitive methods, demonstrating superior accuracy, interpretability, robustness, and reliability across both complete and missing modality scenarios at test time. Code is available at https://github.com/LezJ/SimMLM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "content": "Multimodal learning aims to build models that can process and integrate information from multiple, complementary modalities such as visual, textual and auditory information, which has been found useful for various tasks such as image-text food classification [4, 15], multimodal medical image segmentation [5, 24] with improved model accuracy. However, in real-world scenarios, it is common for one or more modalities to be missing during deployment due to hardware failures, environmental conditions, or constraints" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "content": "in data collection, budget, and resources. Addressing this test-time missing modality problem is crucial for building robust multimodal models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 553, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 507 + ], + "type": "text", + "content": "One mainstream to address the missing modality problem is data imputation, which focuses on synthesizing the missing modalities from available ones [18, 44] or approximating intermediate missing features [10, 35, 36, 40]. However, ensuring the factuality and fidelity of synthetic data remains a major challenge, as issues such as poor data quality [16, 27], data hallucination [30], and vulnerability to adversarial corruption [6, 29] have been reported. Moreover, the computational overhead of generative imputation limits its practicality. In parallel, sophisticated frameworks have been developed, which aim to tackle the missing modality by leveraging advanced, complex network architectures [1, 35, 41], and harmonize or align early or late features from different modalities into a unified, shared space. However, such approaches often introduce substantial resource demands. While some are not adaptable to other tasks at all, others constrain the choice of network architectures or require additional modules, such as projection layers, to accommodate diverse or heterogeneous modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 652 + ], + "type": "text", + "content": "In this paper, we introduce SimMLM, a simple yet effective framework for multi-modal learning with missing modalities, without relying on specific neural network architectures or complex data imputation. First, to effectively handle varying modality availability, we propose a Dynamic Mixture of Modality Experts (DMoME) architecture, where modality-specific expert networks process available modalities, and a learnable gating network dynamically adjusts their contributions. This design ensures that the model remains effective in both full and partial modality settings while being adaptable to different neural architectures and task domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "To further improve the model's robustness and reliability against missing modalities, we introduce a novel loss function, which is designed to optimize the loss landscape around the region of missing data. Inspired by an intuitive principle that the model should not degrade in per" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 143, + 703 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 242, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 242, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 242, + 712 + ], + "type": "text", + "content": "†† These authors contributed equally as senior authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24068" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "formance when modalities are increased, this loss function, named More vs. Fewer (MoFe) ranking loss, ensures that as more modalities become available, the model's accuracy improves or at least remains stable. The MoFe loss encourages the model to compare performance between modality-rich and modality-poor pairs, promoting an inherent ranking that favors configurations with greater modality availability. This structured comparison in together with the dynamic weighting scheme in DMoME allows the model to learn consistent and reliable multimodal fusion strategies, improving robustness and adaptability across varying modality availability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 217, + 266, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 217, + 266, + 227 + ], + "spans": [ + { + "bbox": [ + 67, + 217, + 266, + 227 + ], + "type": "text", + "content": "Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 295, + 479 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 228, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 323 + ], + "type": "text", + "content": "- Dynamic Mixture of Modality Experts Architecture: SimMLM introduces the DMoME architecture with a learnable gating mechanism that adaptively adjusts modality contributions based on availability, improving performance in both complete and partial modality settings. This design enables seamless integration with various neural networks, ensuring flexibility and scalability across diverse tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 323, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 295, + 396 + ], + "type": "text", + "content": "- More vs. Fewer ranking loss: We introduce the novel MoFe ranking loss, which ensures model performance improves or remains stable as additional modalities are incorporated. By enforcing a ranking comparison between modality-rich and modality-poor pairs, MoFe strengthens model robustness and reliability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 396, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 295, + 479 + ], + "type": "text", + "content": "- Comprehensive Evaluation: We validate SimMLM on multimodal segmentation and classification across three datasets, demonstrating superior accuracy, interpretability, robustness, reliability, and computational efficiency over state-of-the-art methods. These results establish SimMLM as a practical and versatile solution for multimodal learning under missing modality conditions." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 490, + 142, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 142, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 142, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 558 + ], + "type": "text", + "content": "To solve multi-modal learning with missing modalities, various approaches have been proposed, including a) input-level imputation, b) advanced representation learning; and c) mixture of expert strategies [37]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 618 + ], + "type": "text", + "content": "Input-level imputation approaches leverage advanced generative networks such as auto-encoders [18, 20], generative adversarial networks [20] to reconstruct missing inputs. Such an approach however, can generate noisy and unreliable input, as discussed in the previous section." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Advanced representation learning focuses on learning improved representation either through advanced networks such as transformers [21, 41, 42], represented by mm-Former [41], or disentangling the latent features into modality-specific and modality-shared features for improving model accuracy, such as Robust-Mseg [1] and ShaSpec [35]. For instance, ShaSpec [35] employs a combination of modality-shared and modality-specific encoders" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "with dedicated losses to promote feature disentanglement and perform feature fusion in the latent space before passing them to a task decoder. When a modality is missing, ShaSpec generates its features from the available modalities through a missing modality feature generation process enabled by the modality-shared encoder. However, this approach imposes a strict constraint that the shared encoder should be suitable for all input modalities and both shared and specific features must have the same dimensions, which can be less efficient and optimal when dealing with highly heterogeneous modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 205, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 455 + ], + "type": "text", + "content": "Mixture of Experts (MoE) has recently been adopted in both unimodal and multimodal learning, demonstrating superior performance in natural language processing [3, 14, 43] and computer vision tasks [28, 38, 39], while improving efficiency by reducing FLOPS. A very recent method for handling missing modalities is Mixture of Modality Knowledge Experts (MoMKE) [38], which uses MLP experts to generate unimodal and joint representations. Even when a modality is missing, MoMKE leverages existing modalities to generate joint representations by utilizing the other experts. While this approach has achieved promising results, it still incurs high computational costs, as each modality must pass through all modality experts, quadratically increasing complexity with the number of modalities. In contrast, the expert model in DMoME is dedicated to a specific modality. Additionally, unlike MoMKE, DMoME does not constrain the dimensionality of latent feature representations across modalities. Instead, it performs the weighting directly at the output space across experts. This flexibility makes DMoME more efficient and scalable, offering high performance in diverse multimodal tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 468, + 397, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 397, + 482 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 397, + 482 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 488, + 425, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 425, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 425, + 500 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": "Assume we have a complete multimodal dataset with " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " modalities, denoted as " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\{\\mathbf{x}_i^m\\}_{m=1}^M, \\mathbf{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": ". where " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^m \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": "-th data sample and " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " indexes the modality. " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " is the label for the sample " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " from the label space " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": ", which can either be a classification label or segmentation map. In this paper, we are interested in finding a solution " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " that learns a mapping from the input space to the label space: " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": ", leveraging this complete modality dataset, so that " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " can work robustly and reliably even when certain modalities " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " are missing at inference time. For simplicity, we omit the sample index " + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 506, + 553, + 650 + ], + "type": "text", + "content": " when not specified in the following sections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 659, + 443, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 443, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 443, + 671 + ], + "type": "text", + "content": "3.2. Overview of SimMLM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Our proposed SimMLM framework consists of two main components: a) a unified Dynamic Mixture of Modality Experts framework (DMoME) (Sec. 3.3), which can take a" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24069" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 271, + 240 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 271, + 240 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 271, + 240 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 271, + 240 + ], + "type": "image", + "image_path": "c4217bb8ce1735c44a6a9b7f4f45ea9f262483f95132e8c717a95a677284c699.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 243, + 209, + 253 + ], + "lines": [ + { + "bbox": [ + 155, + 243, + 209, + 253 + ], + "spans": [ + { + "bbox": [ + 155, + 243, + 209, + 253 + ], + "type": "text", + "content": "(a) Our DMoME" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 83, + 517, + 229 + ], + "blocks": [ + { + "bbox": [ + 304, + 83, + 517, + 229 + ], + "lines": [ + { + "bbox": [ + 304, + 83, + 517, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 83, + 517, + 229 + ], + "type": "image", + "image_path": "ef5a9060e9ee5cc89056730ebd33f3dddb9216d85afc909c9de40a5f67fa5726.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 243, + 466, + 254 + ], + "lines": [ + { + "bbox": [ + 360, + 243, + 466, + 254 + ], + "spans": [ + { + "bbox": [ + 360, + 243, + 466, + 254 + ], + "type": "text", + "content": "(b) Computation flows of " + }, + { + "bbox": [ + 360, + 243, + 466, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": "Figure 1. Our proposed SimMLM consists of a) DMoME architecture: DMoME takes either partial or full modality inputs where each modality is processed by a corresponding expert network to produce its own output " + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "(\\mathbf{o}^1, \\ldots, \\mathbf{o}^M)" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": ". A gating network (G) assesses the same inputs and produces weights to dynamically weigh the contributions of each expert's output in order to get the mixture of output " + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": "; b) " + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MoFe}}" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": " loss: At training time, in addition to standard task specific losses, DMoME is optimized with the More " + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "(X^{+})" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": " vs. Fewer Modality " + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "inline_equation", + "content": "(X^{-})" + }, + { + "bbox": [ + 55, + 266, + 555, + 333 + ], + "type": "text", + "content": " (MoFe) ranking loss, directly encouraging the network to distinguish between inputs with differing modality availability at the task performance level." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 427 + ], + "type": "text", + "content": "varied number of modalities as input and adaptively weight their contribution for robust performance at both training and inference time; b) a novel ranking loss called MoFe (Sec. 3.4), which is designed to calibrate the network's performance by differentiating network performance when the number of available modalities for the same sample varies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 456 + ], + "type": "text", + "content": "3.3. Dynamic Mixture of Modality Expert (DMoME)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": "As shown in Figure 1 (a), Our DMoME consists of a set of modality-specific expert networks " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\{E^{m}(\\mathbf{x}_{m};\\theta_{m})\\}_{m = 1}^{M}" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": " each containing its own set of learnable parameters " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\theta_{m}" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": " and a gating network, " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "G(\\cdot ;\\phi)" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": ", with a set of learnable parameters " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": ". Each expert network, " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "E^{m}(\\mathbf{x}_{m};\\theta_{m})" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": ", takes its corresponding modality as input and produces its output in logits " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^m" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": ", before applying the softmax function: " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^m = E^m (\\mathbf{x}^m;\\theta^m)\\in \\mathbb{R}^T" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 462, + 295, + 581 + ], + "type": "text", + "content": " denotes the dimensionality of the output space for a specific task, e.g., the number of predicted classes in a classification task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": "The gating network " + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "inline_equation", + "content": "G(\\cdot; \\phi)" + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": " takes a set of different modalities as input and produces a set of gating values " + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\{g^m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": " to calibrate different modality experts' output. The softmax function is then applied to computing non-negative weights " + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\{w_1, w_2, \\ldots, w_M\\} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": ", ensuring that the sum of the weights for existing modalities is equal to 1. The final output can be described as a reweighted combination of the outputs from the corresponding modality experts " + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{o}^m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{o} = \\sum_{m=1}^M w^m E^m(\\mathbf{x}^m; \\theta_m)" + }, + { + "bbox": [ + 55, + 582, + 295, + 690 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 712 + ], + "type": "text", + "content": "Notably, the input to the gating network can be an incomplete set of modalities. If certain modalities are miss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "text", + "content": "ing, those inputs are treated as 0, and the gating values for the corresponding missing modality experts' outputs are set to " + }, + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 313, + 354, + 555, + 413 + ], + "type": "text", + "content": " (which causes the corresponding weight values to be 0 after the softmax). Mathematically, the gating outputs are set as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 421, + 554, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 421, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 328, + 421, + 554, + 453 + ], + "type": "interline_equation", + "content": "g _ {m} = \\left\\{ \\begin{array}{l l} G (\\mathbf {x}; \\phi) ^ {(m)}, & \\text {i f m o d a l i t y m i s p r e s e n t} \\\\ - \\infty , & \\text {i f m o d a l i t y m i s m i s s i n g} \\end{array} , \\right. \\tag {1}", + "image_path": "9f899a9dfd0308e8abb7398776bb9be63cc31461bd54f6600924906c7f7cbaa2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 492 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 461, + 554, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 461, + 554, + 492 + ], + "type": "text", + "content": " can be a complete set of modalities or an arbitrary incomplete set of modalities for a given sample. The final weights therefore can be simply computed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 496, + 554, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 496, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 344, + 496, + 554, + 520 + ], + "type": "interline_equation", + "content": "w ^ {m} = \\frac {\\exp \\left(g ^ {m}\\right)}{\\sum_ {m ^ {\\prime} = 1} ^ {M} \\exp \\left(g ^ {m ^ {\\prime}}\\right)} \\quad \\text {f o r} \\quad m = 1, 2, \\dots , M. \\tag {2}", + "image_path": "c875d06ae6f7eeb779d642370e3b35c03ba2ab1cc0cf7e536a5df7990134bee0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "content": "Here, the weight vector " + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = [w^{1}, w^{2}, \\ldots, w^{M}] \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "content": " is non-negative and sums to 1, with " + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "inline_equation", + "content": "w^{m} = 0" + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "content": " for missing modalities, effectively excluding them from the final mixture. The above configuration can be easily extended to multi-task settings, where each task has its own task-specific weighting scheme. In this case, the gating network produces " + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\mathbb{R}^{M \\times K}" + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 524, + 554, + 609 + ], + "type": "text", + "content": " being the number of tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 615, + 398, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 615, + 398, + 627 + ], + "spans": [ + { + "bbox": [ + 314, + 615, + 398, + 627 + ], + "type": "text", + "content": "Two-stage training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 653 + ], + "type": "text", + "content": "To train the whole network, we adopt a two-stage training process, which consists of:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "text", + "content": "1. independent learning: we independently train a modality-specific expert " + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "E_{m}" + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "text", + "content": " for each modality " + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "text", + "content": ", supervised by a task-specific loss " + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{task}}(\\mathbf{o}^{m},\\mathbf{y})" + }, + { + "bbox": [ + 315, + 654, + 555, + 714 + ], + "type": "text", + "content": ". For example, for classification tasks, the loss function can be the cross-entropy (CE) loss, whereas for the segmentation," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24070" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 295, + 156 + ], + "type": "text", + "content": "the loss can be Dice loss. This pretraining stage ensures that each modality expert independently learns specialized features, minimizing the risk of any single modality dominating the learning process. Different modalities can vary significantly in learning dynamics and task difficulty, which can lead to training instability, especially at the early stages of joint training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 295, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 203 + ], + "type": "text", + "content": "2. Cooperative learning: we co-train the expert networks alongside the gating network, ranking network performance under More vs. Fewer Modality conditions, see next section for details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 301 + ], + "type": "text", + "content": "This training scheme enables the expert networks to exploit modality-specific information without interference from other modalities before jointly optimizing with the gating network to exploit mutual benefits. It also improves framework flexibility with reduced training resources, e.g., a) first-stage training can be accelerated via parallel computing; b) when new modalities are introduced, existing pretrained expert networks can be reused for the second stage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 255, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 255, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 255, + 323 + ], + "type": "text", + "content": "3.4. More vs. Fewer Modality Co-training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 424 + ], + "type": "text", + "content": "To boost model robustness against missing modalities, we simulate scenarios with missing modalities by randomly dropping out input modalities during training, which is a common practice. Yet, unlike existing approaches that focus on aligning outputs from missing modalities with those from the full set of modalities [7, 9, 17], we encourage the network to distinguish between modality-rich and modality-poor inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 533 + ], + "type": "text", + "content": "Rational behind More vs. Fewer Modality Training: In many multi-modal learning tasks, all modalities are assumed to be predictive and complementary, contributing valuable information to the tasks that are of interest. Our approach is inspired by the intuition that richer multimodal inputs should yield more accurate results, as each modality complements the others. As the number of modalities decreases, the input becomes more limited, making it more challenging to capture complex patterns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": "As it is difficult to quantify the exact performance for different set-ups, we propose to formulate it as a ranking loss to encourage the network to compare the two scenarios. In cognitive science, it has been found that learners who engage in comparative tasks show an improved ability to transfer knowledge to new situations compared to those who learn through isolated examples [32]. Such a concept has also been applied in reinforcement learning for preference alignment for the applications like large language model training and finetuning with a reward function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In our multimodal learning setting, at each training iteration, we sample two sets from a data point " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{full}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{+}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{-}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{-} \\subset \\mathbf{x}^{+} \\subseteq \\mathbf{x}^{full}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". The corresponding outputs from DMoME are denoted as " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^{+}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{o}^{-}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", respectively. We conjecture that the task-specific loss value for modality-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "rich setups with " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{+}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " should be at least no higher than those in the modality " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{-}" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": ". Mathematically, every pair of samples " + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}^{+},\\mathbf{x}^{-})" + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": " should satisfy the following relationship:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 114, + 556, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 114, + 556, + 128 + ], + "spans": [ + { + "bbox": [ + 314, + 114, + 556, + 128 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) \\geq \\mathcal {L} _ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) \\iff A c c \\left(\\mathbf {x} ^ {+}\\right) \\geq A c c \\left(\\mathbf {x} ^ {-}\\right).", + "image_path": "cb1ea5f96ed2e4c3d52b13cdccf6c2b053c47d15e0c19dc0e0565e907f0ab994.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 555, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 169 + ], + "type": "text", + "content": "Definition of " + }, + { + "bbox": [ + 313, + 133, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 133, + 555, + 169 + ], + "type": "text", + "content": ": To guide the network to learn the ordinary relationship, we introduce More vs. Fewer (MoFe) ranking loss, defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 174, + 553, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 199 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {M o F e}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\max \\left(0, \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) - \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right)\\right), \\tag {3}", + "image_path": "04514c1a8ba1389cf083ab313403a9aca206e3d0f3913c0542d1939b56c4e575.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 200, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 282 + ], + "type": "text", + "content": "which encourages the network to perform better when more modalities are available. It is worth mentioning that the proposed loss is general and thus can be applied to other existing multi-modal learning frameworks to better calibrate the network for improved robustness against missing modalities as well as improved reliability. The total training loss is thus defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 289, + 553, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 289, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 319, + 289, + 553, + 313 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {t o t a l}} \\left(\\mathbf {o} ^ {+}, \\mathbf {o} ^ {-}, \\mathbf {y}\\right) = \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {+}, \\mathbf {y}\\right) + \\ell_ {\\text {t a s k}} \\left(\\mathbf {o} ^ {-}, \\mathbf {y}\\right) + \\lambda \\ell_ {\\mathrm {M o F e}}, \\tag {4}", + "image_path": "4666aacd5c7cbfd47f36c8752b984fe82f261551d899b30cb5a92d13a106d132.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "text", + "content": " is a coefficient to control the contribution of " + }, + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{MOFE}}" + }, + { + "bbox": [ + 313, + 314, + 553, + 325 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 326, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 577 + ], + "type": "text", + "content": "We acknowledge the existence of prior research, such as the works of Noh et al. (2023) [26], and Ma et al. (2023) [19], which have investigated the application of ranking loss at the confidence levels between a pair of two inputs. However, our approach diverges significantly from these existing methods. Specifically, we propose the novel concept of applying ranking loss directly at the loss level, as opposed to the confidence level. This distinction is crucial, as our method aims to directly regularize the geometry of the loss landscape. In our experiments, we also showed extra benefit of our loss, which leads to reduced calibration errors, see Table. 4. Besides, our proposed method enables seamless integration with a range of loss functions, offering flexibility beyond classification and allowing for improved applicability in complex, diverse tasks. For example, our method can be applied for segmentation and regression tasks where the confidence scores are not straightforward in dense or continuous prediction settings. By focusing on the actual task error, our loss-level approach directly aligns with the downstream task performance, providing a more consistent measure across different tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 586, + 395, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 395, + 600 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 395, + 600 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": "We evaluate our approach on two types of tasks, multimodal medical segmentation and multimodal classification, spanning three datasets. For segmentation, we use BraTS 2018 [22] benchmark. For classification, we conduct experiments on UPMC Food-101 [4] and avMNIST dataset [34]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 672, + 376, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 376, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 376, + 684 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "BraTS 2018 (Brain Tumor Segmentation Challenge [22]) is a widely used benchmark for brain tumor segmenta" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24071" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "tion in 3D MRI scans. It has also been extensively employed to evaluate methods for handling missing modalities [7, 35, 41, 42]. The dataset includes four distinct MRI modalities: FLAIR, T1, T1 contrast-enhanced (T1ce), and T2. The segmentation task involves identifying three tumor subregions: the enhancing tumor (ET), the tumor core (TC), and the whole tumor (WT). The dataset consists of 285 manually labeled cases and 66 unlabeled cases for evaluation. In our experiments, we split the labeled cases into a 4:1 ratio for training and validation. Final evaluations are conducted on the 66 unlabeled cases via the official benchmarking platform for a consistent and fair accuracy comparison against existing methods [13, 41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 275 + ], + "type": "text", + "content": "UPMC Food-101 [4] is a large-scale food classification dataset with 90,704 image-text pairs spanning 101 classes. Collected from uncontrolled environments, it introduces inherent noise, posing challenges for accurate classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 294, + 335 + ], + "type": "text", + "content": "avMNIST (Audiovision MNIST [34]) is a digit classification dataset comprising 1,500 image-audio pairs. The visual modality consists of " + }, + { + "bbox": [ + 55, + 275, + 294, + 335 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 55, + 275, + 294, + 335 + ], + "type": "text", + "content": " grayscale images of handwritten digits (0-9) from MNIST [12], while the audio modality is sourced from the Free Spoken Digits Dataset*." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 188, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 188, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 188, + 355 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 384 + ], + "type": "text", + "content": "All the experiments were conducted on a single A100 GPU node. Detailed implementations are described below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": "BraTS 2018: We employed nnUNet [8], a widely used and powerful medical image segmentation backbone, as the expert network for each modality, which can be viewed as a 3D U-Net. The gating network " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " in DMoME utilized lightweight CNNs to capture broad contextual information, followed by a linear layer generating " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "4 \\times 3" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " weights for the four modalities and three segmentation tasks. We optimized both training stages using the Adam optimizer [11] (" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "lr = 0.01" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": "). The " + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 55, + 384, + 295, + 504 + ], + "type": "text", + "content": " in Eq.4 combined Dice [23] and binary cross-entropy losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "text", + "content": "UPMC Food-101: We employed Inception V3 [31] as the image expert and BERT [2] as the text expert, following [4]. To avoid redundant feature extraction, intermediate features from both were fed into an MLP gating network " + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "text", + "content": ", which generated 2D weights for expert outputs. Optimization was performed using Adam (" + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "inline_equation", + "content": "lr = 0.0001" + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "text", + "content": ") with cross-entropy loss as " + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 55, + 505, + 295, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": "avMNIST: Following [20, 35], we adopted a LeNet-5 [12] backbone for both image and audio modalities, with audio converted into " + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "inline_equation", + "content": "20 \\times 20" + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": " MFCC spectrograms [33]. The gating network " + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": " followed a similar design as in BraTS 2018. Two-stage training was conducted using Adam (" + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "inline_equation", + "content": "lr = 0.001" + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": ") with cross-entropy loss as " + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 55, + 587, + 295, + 659 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": "Detailed gating network " + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": " architectures for BraTS 2018 and avMNIST are provided in the Supplementary. For task accuracy evaluation, we follow common practice, using the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "Dice score for segmentation (BraTS 2018) and Accuracy for classification (UPMC Food-101, avMNIST). The MoFe loss coefficient " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " in Eq.4 is set to 0.1 across tasks; however, our method remains robust to variations in " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", as demonstrated in Sec. 5 Analysis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 141, + 437, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 141, + 437, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 437, + 154 + ], + "type": "text", + "content": "4.3. Segmentation Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "text", + "content": "Benchmarking segmentation accuracy on the official BraTS 2018 evaluation set. For the BraTS 2018 benchmark, we compare our approach against several state-of-the-art methods, including Robust-MSeg (RbSeg) [1], mmFormer (mmFm) [41], ShaSpec [35], and MoMKE [38] under 15 configurations of both missing and full modality setups in Table 1. Results show that SimMLM achieves the highest average accuracy in terms of the Dice scores, consistently ranking as the top or second-highest performer in each entry across 44 out of all 45 configurations (15 settings " + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "text", + "content": " 3 subtasks), despite utilizing only " + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "text", + "content": " of the labeled data compared to the " + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 159, + 553, + 399 + ], + "type": "text", + "content": " used by ShaSpec and mmFormer. This consistent performance highlights our model's robustness and data efficiency in handling various incomplete modality inputs. Notably, when only the T1ce modality is available (Table 1, row 3), our model significantly improves the Dice score compared to the second-best methods, increasing from 73.29 to 78.58 for ET, 78.71 to 83.45 for TC, and 74.00 to 80.02 for WT. Such results underscore the advantage of our approach even in single-modality settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "text", + "content": "As shown in Figure 2, in both full and partial modality scenarios, SimMLM (DMoME with " + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "text", + "content": ") achieves more accurate segmentation results with better model reliability compared to DMoME alone (A vs. B, C vs. D) thanks to the contribution of " + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "text", + "content": ". When T1ce is unavailable, DMoME with " + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 399, + 553, + 554 + ], + "type": "text", + "content": " can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce), which is aligned with clinical practice despite the network is not explicitly taught. Without MoFe loss, the network is biased to T2 (maybe due to higher intensity) and over-segment the tumor. In the following Sec. 5, we will provide a more detailed quantitative analysis for our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 436, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 436, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 436, + 575 + ], + "type": "text", + "content": "4.4. Classification Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Benchmarking classification accuracy on UPMC Food101 and avMNIST. For the classification tasks, we compared our model to the two most competitive methods: ShaSpec [35] and MoMKE [38]. To ensure fair comparisons, all models were trained under identical conditions. The results, presented in Table 2, report accuracy across different modality settings. Our method consistently surpasses SOTA methods across all settings, demonstrating its effectiveness in handling both missing and full modality scenarios. While clear improvements are observed on both datasets, our method achieves even greater gains on the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 295, + 713 + ], + "type": "text", + "content": "*Dataset link: https://github.com/Jakobovski/free-spoken-digit-dataset." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24072" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 70, + 520, + 284 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 520, + 284 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 520, + 284 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 520, + 284 + ], + "type": "table", + "html": "
SettingsAvailable Modalities at Test TimeEnhancing tumor (ET)Tumor core (TC)Whole tumor (WT)
◇◇◇RbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]OursRbSeg [1]mmFm [41]ShaSpec [35]MoMKE [38]Ours
Missing1●○○25.6939.3343.5250.5351.3153.5761.2169.4475.0574.1285.6986.1088.6889.7889.24
2○●○17.2932.5341.0048.1248.4147.9056.5563.1868.6671.4870.1167.5273.4477.4580.28
3○○●67.0772.6073.2972.3078.5876.8375.4178.6578.7183.4573.3172.2273.8274.0080.02
4○○○28.9743.0546.3151.0250.6057.4964.2069.0372.3473.4982.2481.1583.9985.8286.01
5●●○32.1342.9644.7652.5253.6260.6865.9172.6775.0077.1488.2487.0689.7690.2389.98
6●○●70.3075.0775.6078.7580.2580.6277.8884.5084.6284.9688.5187.3090.0690.3490.35
7●○○33.8447.5247.2054.1254.6261.1669.7572.9376.9075.7688.2887.5990.0290.5390.28
8○●●69.0674.0475.7676.7878.7778.7278.5982.1081.7084.7977.1874.4278.7479.8682.25
9○●○32.0144.9946.8453.7453.4062.1969.4271.3873.7374.7384.7882.2086.0386.6087.42
10○○●69.7174.5175.9577.1178.6180.2078.6183.8283.6184.1985.1982.9985.4286.9288.11
11●●●70.7875.4776.4278.1980.4181.0679.8085.2382.8385.2588.7387.3390.2990.2190.63
12●●○36.4147.7046.5555.9355.8464.3871.5273.9777.0078.0288.8187.7590.3690.5990.62
13●○●70.8875.6775.9978.4480.2980.7279.5585.2683.5985.2389.2788.1490.7890.8590.99
14○●●70.1074.7576.3778.2780.3380.3380.3984.1882.3585.0886.0182.7186.4786.6287.78
Full15●●●71.1377.6178.0877.6580.4080.8685.7885.4582.6185.3989.4589.6490.8890.5491.07
Average51.0259.8561.5865.5667.1669.7872.9777.4578.5880.2084.3982.9485.9286.6987.67
", + "image_path": "02c572e0df9a51b05f02259326cc6f537b42297455160933482a83083adf6fbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 555, + 326 + ], + "type": "text", + "content": "Table 1. Model accuracy (Dice score ↑ for the segmentation task, normalized to 100%) comparison of different methods for segmenting brain tumor subregions under various missing (○) and/or full modality settings. The best and second best performance of each subtask for each setting are highlighted in bold and underline respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 100, + 337, + 512, + 555 + ], + "blocks": [ + { + "bbox": [ + 100, + 337, + 512, + 555 + ], + "lines": [ + { + "bbox": [ + 100, + 337, + 512, + 555 + ], + "spans": [ + { + "bbox": [ + 100, + 337, + 512, + 555 + ], + "type": "image", + "image_path": "d289ce2011664d25e89a2713d2558cf8757c63c5039de3c1496049502d016106.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 566, + 555, + 612 + ], + "lines": [ + { + "bbox": [ + 54, + 566, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 566, + 555, + 612 + ], + "type": "text", + "content": "Figure 2. Visualization of modality weight changes (bold numbers in the upper right corner) assigned by the gating network alongside the predicted ET segmentation and uncertainty (entropy) maps when full or partial modalities (T1ce is missing) are present. DMoME with " + }, + { + "bbox": [ + 54, + 566, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 54, + 566, + 555, + 612 + ], + "type": "text", + "content": " can intelligently shift the focus to the most relevant modality: T1 (the unenhanced counterpart of T1ce) when T1ce is missing, which is aligned with clinical practice despite the network is not explicitly taught." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 632, + 294, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 294, + 656 + ], + "type": "text", + "content": "noisy UPMC Food-101 dataset, where classification is more challenging, underscoring the robustness of our SimMLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 669, + 115, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 115, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 115, + 681 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "Effectiveness of DMoME: To assess the unique contribution of DMoME, we trained the network w/o MoFe" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 632, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 555, + 704 + ], + "type": "text", + "content": "loss: " + }, + { + "bbox": [ + 313, + 632, + 555, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 632, + 555, + 704 + ], + "type": "text", + "content": " and compared its performance with competitive models—mmFormer [41], ShaSpec [35], and MoMKE [38]. Experiments were conducted on the BraTS 2018 dataset, encompassing 15 different configurations for both missing-modality and full-modality setups. Table 3 reports the average Dice scores across 15 scenarios, alongside" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24073" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 286, + 162 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 286, + 162 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 286, + 162 + ], + "type": "table", + "html": "
DatasetModalitiesAccuracy (%) ↑
Modality 1Modality 2ShaSpec [35]MoMKE [38]SimMLM (Ours)
UPMC Food-101Image○69.2270.4672.20
○Text86.5586.5987.2
ImageText92.7392.7194.99
Average82.8383.2584.81
avMNISTImage○91.9092.6192.69
○Audio89.2891.1691.61
ImageAudio98.7198.6999.27
Average93.3094.1594.52
", + "image_path": "3a17afc53a629f95fc46a65824e17241003bad66a11657853141d34574506cfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 238, + 285, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 170, + 295, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 226 + ], + "type": "text", + "content": "Table 2. Accuracy " + }, + { + "bbox": [ + 55, + 170, + 295, + 226 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 170, + 295, + 226 + ], + "type": "text", + "content": " ↑ comparison across missing & full modality settings on UPMC Food-101 and avMNIST datasets. The datasets have different modality pairs (Image-Text for UPMC Food-101, Image-Audio for avMNIST). o represents modality missingness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 238, + 285, + 308 + ], + "lines": [ + { + "bbox": [ + 67, + 238, + 285, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 238, + 285, + 308 + ], + "type": "table", + "html": "
Dice score (%) ↑Model efficiency
MethodETTCWT#ParamsFLOPS
mmFormer [41]59.8572.9782.94106M748G
ShaSpec [35]61.5877.4585.92187.7M713G
MoMKE [38]65.5678.5886.697.8M490G
DMOE w/o LMoFe66.0579.1486.717.8M123G
DMOE w LMoFe (SimMLM)67.1680.2087.677.8M123G
", + "image_path": "3b2bcf25c785f3874312e4cc3ade489e58bd3537240b6536fcd08582dbcc931e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "text", + "content": "Table 3. Comparison of model accuracy (Dice score " + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "text", + "content": ", normalized to " + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "text", + "content": ", averaged across all missing & full modality settings) and efficiency (# of parameters and FLOPS) on BraTS 2018 official-evaluated benchmark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 384, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 577 + ], + "type": "text", + "content": "networks' parameters and floating-point operations per second (FLOPS). As a result, DMoME achieves the highest Dice scores across all tumor subregions, with particularly strong performance in the more challenging enhancing tumor (ET) and tumor core (TC) segmentation. Importantly, DMoME achieves a fourfold reduction in FLOPS compared to MoMKE, even with the same expert architecture, and requires over 13 times fewer parameters than mmFormer and ShaSpec, underscoring its efficiency and scalability. This reduction in FLOPS is due to a key difference in processing: unlike MoMKE, which feeds each modality's input to all experts [38], our DMoME processes each input with only its dedicated expert. This simple approach significantly reduces FLOPS without compromising representational quality. Further details on DMoME's micro design are provided in the Supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 715 + ], + "type": "text", + "content": "Interpretability of DMoME: Importantly, our network directly provides case-specific interpretability by dynamically revealing the importance of each modality, as shown in Fig. 2. In contrast, previous works such as mmFormer [41], ShaSpec [35], and MoMKE [38] lack this capability. We further performed population-wise study by plotting the average gating weights under different full- and missing-modality configurations on the official BraTS 2018 evaluation set in Fig. 3 across different tasks. It is evident that the gating network managed to dynamically adapt the weights to accommodate various scenarios, providing additional in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 544, + 281 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 544, + 281 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 544, + 281 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 544, + 281 + ], + "type": "image", + "image_path": "26c4c60d7b0cb0da60794d819fd7123c11f1ae8ddbe76bd804d37aa1659f48f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 289, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 323 + ], + "type": "text", + "content": "Figure 3. Average gating weights analysis under both full and missing modality setups. Plotted values are the average gating weights on the BRATS 2018 official evaluation set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 336, + 555, + 502 + ], + "blocks": [ + { + "bbox": [ + 316, + 336, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 316, + 336, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 555, + 502 + ], + "type": "image", + "image_path": "d2ff5ffda69616017def2d3d5e84801cd0719721b7f3746c221ac24c1133147a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "text", + "content": "Figure 4. Impact of the MoFe loss coefficient " + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "text", + "content": ". The Dice score (\\%) " + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "text", + "content": " is reported with different values of coefficients " + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 511, + 555, + 567 + ], + "type": "text", + "content": " for the MoFe loss across all missing & full modality settings on the BraTS 2018 validation set. The results include (1) enhancing tumor, (2) tumor core, (3) whole tumor, and (4) their average." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 588, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 649 + ], + "type": "text", + "content": "sights for clinicians regarding modality importance in different subtasks. For instance, T1ce is prioritized for enhancing tumor (ET) segmentation, while Flair and T2 are more important for whole tumor (WT) segmentation, which is consistent with clinical findings [22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness and Sensitivity of MoFe loss: We further analyzed the impact of MoFe loss by varying its coefficient " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = \\{0, 0.01, 0.05, 0.1, 0.2, 0.5, 1\\}" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " in Eq.4, where " + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": " corresponds to training without it. As shown in Figure 4, the model achieved the highest average Dice score across" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24074" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 284, + 137 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 284, + 137 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 284, + 137 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 284, + 137 + ], + "type": "table", + "html": "
ECE (%) ↓SCE (%) ↓
MethodETTCWTETTCWT
MoMKE [38]3.464.104.0414.8211.357.69
DMoME3.343.863.9813.4011.017.11
DMoME w/ LConf [19]3.173.863.6114.8011.586.81
DMoME w/ LMoFe (SimMLM)3.153.753.5513.2110.835.65
", + "image_path": "d213656548490e5cd164a1d44cd6da8a2ea599135b91a0b11133e8d880c85f5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": "the three subregions at " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": ". Notably, most positive " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": " values resulted in improvements over the baseline " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": ", underscoring the stability and effectiveness of MoFe loss. This improvement may be attributed to the calibration nature of the loss, which enables DMoME to adaptively distinguish between modality-rich and modality-poor conditions, guiding the network to assess each modality's contribution across various scenarios. As a result, the model learns a more casual, robust relationship between diverse inputs and the task output. In the following, we will verify how " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": " can calibrate the network to achieve the goal." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": "MoFe loss leads to better calibration: To evaluate the performance of " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " for model calibration, we compute the expected calibration error (ECE) " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " and static calibration error (SCE) " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " [25], both using 20 bins, and compare our method to the most competitive method: MoMKE [38], in addition to the state-of-the-art multimodal calibration approaches, leveraging missing modality information [19], where the task losses " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{task}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " have been replaced with confidence scores obtained from model predictions using max probability: " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Conf} = \\max(0, Conf(\\mathbf{o}^{-}) - Conf(\\mathbf{o}^{+}))" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": ". Results in Table 4 are scores directly computed on our validation set, as BraTS does not release ground truth labels on the evaluation set. All experiments were conducted in the same settings. It is evident that our method with " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " achieves the best calibration performance in all settings, with the lowest ECEs and SCEs. In the context of this segmentation tasks, compared to " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Conf}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": ", which is label-ignorant and performs comparisons at a low level (e.g., voxels), " + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 56, + 355, + 295, + 643 + ], + "type": "text", + "content": " operates at the sample level, and leverages task-specific loss with ground truth information, making it more task-informative. This encourages the network to focus on task-relevant areas across varying levels of missing data, reducing sensitivity to background noise, augmentation artifacts, and outliers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "MoFe loss reduces model counterintuitive rate (CR): To directly quantify whether " + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": " can better align the model performance with the modality availability (i.e., reduces cases where the model performs better even with fewer input modalities), we introduce a novel metric called the" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 327, + 70, + 541, + 129 + ], + "blocks": [ + { + "bbox": [ + 55, + 146, + 295, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 179 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of calibration errors (ECEs↓, SCEs↓) averaged across all 15 full and missing modality settings on the validation set of BraTS 2018 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 70, + 541, + 129 + ], + "lines": [ + { + "bbox": [ + 327, + 70, + 541, + 129 + ], + "spans": [ + { + "bbox": [ + 327, + 70, + 541, + 129 + ], + "type": "table", + "html": "
Counterintuitive rate (%) ↓
MethodETTCWT
MoMKE [38]14.6032.479.86
DMoME w/o LMoFe10.4730.943.87
DMoME w/ LMoFe (SimMLM)7.2028.193.75
", + "image_path": "6f09563dfbbcb76ed3eec96902a0d2bb022075bcf1c27abc252559786580d1e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 137, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 159 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of counterintuitive rate " + }, + { + "bbox": [ + 313, + 137, + 553, + 159 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 137, + 553, + 159 + ], + "type": "text", + "content": " on the BraTS 2018 official evaluation set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 507, + 190 + ], + "type": "text", + "content": "Counterintuitive Rate (CR), which is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 194, + 548, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 194, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 318, + 194, + 548, + 232 + ], + "type": "interline_equation", + "content": "\\mathrm{CR} = \\frac{1}{N}\\sum_{i = 1}^{N}\\frac{1}{|\\mathcal{S}_{i}|}\\sum_{(\\mathbf{x}_{i}^{-},\\mathbf{x}_{i}^{+})\\in \\mathcal{S}_{i}}\\left\\{ \\begin{array}{ll}1 & \\text{if} P(\\mathbf{x}_{i}^{-}) > P(\\mathbf{x}_{i}^{+})\\\\ 0 & \\text{otherwise} \\end{array} \\right..", + "image_path": "40059c80668cdc778f3abc1cf7590fe21545b04aeee468273116b47fb31a453b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": " is the total number of samples, " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_i" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": " is the set of all more vs. fewer modality pairs " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_i^+ \\text{vs.} \\mathbf{x}_i^-)" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": " for sample " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x}_i)" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": " denotes the model accuracy score (e.g., Dice score for the segmentation task) on the sample " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": ". An instance with " + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "inline_equation", + "content": "P(\\mathbf{x}_i^-) > P(\\mathbf{x}_i^+)" + }, + { + "bbox": [ + 313, + 236, + 553, + 296 + ], + "type": "text", + "content": " is considered as counterintuitive." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": "We compared CR on BraTS 2018 across three settings: MoMKE [38] (previous SOTA), DMoME without " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": " and DMoME with " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": " (SimMLM) and report the performance in Table 5. Results show that " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MoFe}" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": " achieves the lowest CR, achieving " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "50.68\\%" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "13.18\\%" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "inline_equation", + "content": "61.97\\%" + }, + { + "bbox": [ + 313, + 297, + 554, + 415 + ], + "type": "text", + "content": " reduction for ET, TC, and WT segmentation, respectively, when compared to MoMKE. Notably, DMoME alone also reduces CR significantly, thanks to the dynamic weighting mechanism to automatically identify and prioritize more relevant modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 421, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 529 + ], + "type": "text", + "content": "Broader Impact: We have shown that SimMLM can provide direct importance values for each input modality and generate informative confidence maps for further analysis in the presence of incomplete data (Fig. 2). Such properties are particularly important to alert clinicians for further investigation in low-resource settings where certain modalities may be unavailable. Additionally, SimMLM's potential for handling varied levels of training-time missing modalities on avMNIST is also discussed in the Supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced SimMLM, a simple, robust, and flexible framework for multimodal learning in scenarios with test-time missing modalities. By incorporating the Dynamic Mixture of Model Experts (DMoME) architecture and the novel More vs. Fewer (MoFe) ranking loss, SimMLM effectively addresses modality absence. Experimental results demonstrate that SimMLM achieves state-of-the-art performance across both segmentation and classification tasks, excelling in accuracy, efficiency, interpretability, robustness, and reliability. As a generic approach, future work will focus on extending SimMLM to other high-stakes real-world applications where trustworthiness matters and broader scenarios, including training-time missingness." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24075" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This research was supported by the UKRI funding (MR/Z505754/1), UK. Dr. Chen is supported by Royal Society (RGS/R2/242355)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 234 + ], + "type": "text", + "content": "[1] Cheng Chen, Qi Dou, Yueming Jin, Hao Chen, Jing Qin, and Pheng-Ann Heng. Robust multimodal brain tumor segmentation via feature disentanglement and gated fusion. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13-17, 2019, Proceedings, Part III 22, pages 447-456. Springer, 2019. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 235, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 296, + 311 + ], + "type": "text", + "content": "[2] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 312, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 296, + 356 + ], + "type": "text", + "content": "[3] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 411 + ], + "type": "text", + "content": "[4] Ignazio Gallo, Gianmarco Ria, Nicola Landro, and Riccardo La Grassa. Image and text fusion for upmc food-101 using bert and cnns. In 2020 35th International Conference on Image and Vision Computing New Zealand (IVCNZ), pages 1-6. IEEE, 2020. 1, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 296, + 456 + ], + "type": "text", + "content": "[5] Yunhe Gao. Training like a medical resident: Context-prior learning toward universal medical image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11194-11204, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "text", + "content": "[6] Shuang Hao, Wenfeng Han, Tao Jiang, Yiping Li, Haonan Wu, Chunlin Zhong, Zhangjun Zhou, and He Tang. Synthetic data in ai: Challenges, applications, and ethical implications. arXiv preprint arXiv:2401.01629, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 579 + ], + "type": "text", + "content": "[7] Minhao Hu, Matthias Maillard, Ya Zhang, Tommaso Ciceri, Giammarco La Barbera, Isabelle Bloch, and Pietro Gori. Knowledge distillation from multi-modal to mono-modal segmentation networks. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 772-781. Springer, 2020. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 624 + ], + "type": "text", + "content": "[8] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 678 + ], + "type": "text", + "content": "[9] Sanaz Karimijafarbigloo, Reza Azad, Amirhossein Kazerouni, Saeed Ebadollahi, and Dorit Merhof. Mmcformer: Missing modality compensation transformer for brain tumor segmentation. In Medical Imaging with Deep Learning, pages 1144-1162. PMLR, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[10] Donggeun Kim and Taesup Kim. Missing modality prediction for unpaired multimodal learning via joint embedding of unimodal models. arXiv preprint arXiv:2407.12616, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[11] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[12] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[13] Yi-Lun Lee, Yi-Hsuan Tsai, Wei-Chen Chiu, and Chen-Yu Lee. Multimodal prompting with missing modalities for visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14943-14952, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 274 + ], + "type": "text", + "content": "[14] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan First, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 342 + ], + "type": "text", + "content": "[15] Tao Liang, Guosheng Lin, Mingyang Wan, Tianrui Li, Guojun Ma, and Fengmao Lv. Expanding large pre-trained unimodal models with multimodal information injection for image-text multimodal classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15492-15501, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[16] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 475 + ], + "type": "text", + "content": "[17] Han Liu, Yubo Fan, Hao Li, Jiacheng Wang, Dewei Hu, Can Cui, Ho Hin Lee, Huahong Zhang, and Ipek Oguz. Moddrop++: A dynamic filter network with intra-subject cotaining for multiple sclerosis lesion segmentation with missing modalities. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 444-453. Springer, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "text", + "content": "[18] Hong Liu, Dong Wei, Donghuan Lu, Jinghan Sun, Liansheng Wang, and Yefeng Zheng. M3ae: multimodal representation learning for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1657-1665, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "text", + "content": "[19] Huan Ma, Qingyang Zhang, Changqing Zhang, Bingzhe Wu, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Calibrating multimodal learning. In International Conference on Machine Learning, pages 23429-23450. PMLR, 2023. 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 579, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 632 + ], + "type": "text", + "content": "[20] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2302-2310, 2021. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[21] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18177-18186, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[22] Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24076" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging, 34(10):1993-2024, 2014. 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "text", + "content": "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11480-11491, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[25] Jeremy Nixon, Michael W Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. Measuring calibration in deep learning. In CVPR workshops, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "text", + "content": "[26] Jongyoun Noh, Hyekang Park, Junghyup Lee, and Bumsub Ham. Rankmixup: Ranking-based mixup training for network calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1358-1368, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[27] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. Advances in Neural Information Processing Systems, 36:24829-24840, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[28] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In Advances in neural information processing systems, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "text", + "content": "[29] Krishnakant Singh, Thanush Navaratnam, Jannik Holmer, Simone Schaub-Meyer, and Stefan Roth. Is synthetic data all we need? benchmarking the robustness of models trained with synthetic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2505-2515, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "text", + "content": "[30] Yujie Sun, Dongfang Sheng, Zihan Zhou, and Yifei Wu. AI hallucination: towards a comprehensive classification of distorted information in artificial intelligence-generated content. *Humanities & social sciences communications*, 11(1): 1-14, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 609 + ], + "type": "text", + "content": "[31] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "text", + "content": "[32] Jonathan G Tullis and Robert L Goldstone. Comparison versus reminding. Cognitive research: principles and implications, 1(1):20, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 678 + ], + "type": "text", + "content": "[33] George Tzanetakis and Perry Cook. Musical genre classification of audio signals. IEEE Transactions on speech and audio processing, 10(5):293-302, 2002. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[34] Valentin Vielzeuf, Alexis Lecheryv, Stéphane Pateux, and Frédéric Jurie. Centralnet: a multilayer approach for multimodal fusion. In Proceedings of the European Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 630 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 93 + ], + "type": "text", + "content": "on Computer Vision (ECCV) Workshops, pages 0-0, 2018. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[35] Hu Wang, Yuanhong Chen, Congbo Ma, Jodie Avery, Louise Hull, and Gustavo Carneiro. Multi-modal learning with missing modality via shared-specific feature modelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15878-15887, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[36] Yuanzhi Wang, Yong Li, and Zhen Cui. Incomplete multimodality-diffused emotion recognition. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "text", + "content": "[37] Renjie Wu, Hu Wang, Hsiang-Ting Chen, and Gustavo Carneiro. Deep multimodal learning with missing modality: A survey, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[38] Wenxin Xu, Hexin Jiang, et al. Leveraging knowledge of modality experts for incomplete multimodal learning. In ACM Multimedia 2024, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[39] Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingxuan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tianlong Chen. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts. arXiv preprint arXiv:2410.08245, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "type": "text", + "content": "[40] Zhilin Zeng, Zelin Peng, Xiaokang Yang, and Wei Shen. Missing as masking: Arbitrary cross-modal feature reconstruction for incomplete multimodal brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 424-433. Springer, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 464 + ], + "type": "text", + "content": "[41] Yao Zhang, Nanjun He, Jiawei Yang, Yuexiang Li, Dong Wei, Yawen Huang, Yang Zhang, Zhiqiang He, and Yefeng Zheng. mmformer: Multimodal medical transformer for incomplete multimodal learning of brain tumor segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 107-117. Springer, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 465, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 529 + ], + "type": "text", + "content": "[42] Zheyu Zhang, Gang Yang, Yueyi Zhang, Huanjing Yue, Aiping Liu, Yunwei Ou, Jian Gong, and Xiaoyan Sun. Tmformer: Token merging transformer for brain tumor segmentation with missing modalities. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7414-7422, 2024. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[43] Xinyu Zhao, Xuxi Chen, Yu Cheng, and Tianlong Chen. Sparse moe with language guided routing for multilingual machine translation. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 576, + 553, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 630 + ], + "type": "text", + "content": "[44] Yunpeng Zhao, Cheng Chen, Qing You Pang, Quanzheng Li, Carol Tang, Beng-Ti Ang, and Yueming Jin. Dealing with all-stage missing modality: Towards a universal model with robust reconstruction and personalization. arXiv preprint arXiv:2406.01987, 2024. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24077" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_content_list.json b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dccbffac53dfef2c5febd546b697d546bdd5bf41 --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_content_list.json @@ -0,0 +1,1205 @@ +[ + { + "type": "text", + "text": "Similarity Memory Prior is All You Need for Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 133, + 140, + 864, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Tang $^{1}$ , Zhiqing Guo $^{1,2,*}$ , Liejun Wang $^{1,3,*}$ , Chao Liu $^{1}$", + "bbox": [ + 279, + 188, + 712, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Xinjiang University", + "bbox": [ + 243, + 205, + 754, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Xinjiang Multimodal Intelligent Processing and Information Security Engineering Technology Research Center", + "bbox": [ + 228, + 224, + 766, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Silk Road Multilingual Cognitive Computing International Cooperation Joint Laboratory, Xinjiang University", + "bbox": [ + 220, + 258, + 776, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{107552201330@stu.，guozhiqing@，wljxju@，107556522208@stu.}xju.edu.cn", + "bbox": [ + 196, + 295, + 795, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 343, + 330, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, it has been found that \"grandmother cells\" in the primary visual cortex (VI) of macaques can directly recognize visual input with complex shapes. This inspires us to examine the value of these cells in promoting the research of medical image segmentation. In this paper, we design a Similarity Memory Prior Network (Sim-MPNet) for medical image segmentation. Specifically, we propose a Dynamic Memory Weights-Loss Attention (DMW-LA), which matches and remembers the category features of specific lesions or organs in medical images through the similarity memory prior in the prototype memory bank, thus helping the network to learn subtle texture changes between categories. DMW-LA also dynamically updates the similarity memory prior in reverse through Weight-Loss Dynamic (W-LD) update strategy, effectively assisting the network directly extract category features. In addition, we propose the Double-Similarity Global Internal Enhancement Module (DS-GIM) to deeply explore the internal differences in the feature distribution of input data through cosine similarity and euclidean distance. Extensive experiments on four public datasets show that Sim-MPNet has better segmentation performance than other state-of-the-art methods. Our code is available on https://github.com/vpsg-research/Sim-MPNet.", + "bbox": [ + 99, + 375, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 102, + 753, + 230, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical image segmentation, as a basic and key technology, is increasingly highlighting its importance in disease diagnosis, treatment planning, and curative effect evaluation [9, 29, 50]. With the rapid development of artificial intelligence and computer vision, constructing automatic and accurate medical image segmentation methods has become a", + "bbox": [ + 99, + 777, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fd2a829bc5f7179d5eea0fa0163b9b8640b415d410288568140dec6c8169f0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 343, + 691, + 521 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f763ff7868dddc722125d4e0d6e3d5b37540d6f69bff7f87b8195fe39d58f90.jpg", + "image_caption": [ + "(a) primate visual system", + "(b) similarity memory prior directly memorizes target features", + "Figure 1. Motivation. (a) Schematic diagram of the primate visual system. (b) We use similarity memory priors stored in the prototype memory bank to imitate \"grandmother cells\" in V1, and utilize these features to directly match various organs in medical images." + ], + "image_footnote": [], + "bbox": [ + 696, + 342, + 893, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "research hotspot [19, 41]. These methods aim to accurately identify and segment the anatomical structures or pathological features of interest from medical images, to provide clinicians with more intuitive and quantitative medical information [21, 43].", + "bbox": [ + 509, + 681, + 895, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In existing medical image segmentation works, researchers attempt to use Convolutional Neural Network (CNN) or Vision Transformer (ViT) to enable networks to generate images with precise labels [3, 6, 8, 30, 48]. However, although the local receptive field design of CNNs conform to the translational invariance of images, its inductive bias is essentially a passive response to local textures (edges and corners), rather than actively modeling semantic objects (lesions and organs). ViTs require massive data to learn se", + "bbox": [ + 509, + 757, + 895, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 119, + 876, + 241, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23009", + "bbox": [ + 478, + 926, + 517, + 939 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/28fe4afcde9d2c5ece0cf7713f07e0b9c4f6058cd21f94a8b7dab42baae99501.jpg", + "image_caption": [ + "Figure 2. Compared with the existing works. CNN passively responds to local texture through convolution kernel. ViT only focuses on high-frequency textures through self-attention. DMW-LA uses similarity memory priors to directly identify and extract category features of organs." + ], + "image_footnote": [], + "bbox": [ + 104, + 99, + 893, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mantic associations of position encoding, while medical image annotation is scarce, leading to model degradation into \"pseudo-global modeling\" that focuses on high-frequency texture rather than anatomical structure. In addition, CNNs and ViTs belong to data-driven \"implicit learning\" and lack prior knowledge to quickly identify key features. Recently, researchers have used two-photon calcium imaging technology to discover the presence of curvature detecting cells in the V1 [33, 34]. These curvature detecting cells can be regarded as a type of \"grandmother cells\" with specific activation conditions, as shown in Figure 1(a). The \"grandmother cells\" have strong recognition abilities and only respond to specific input stimuli [14, 26, 32]. If the design of medical segmentation networks is not limited to traditional calculation mechanisms and directly focuses on lesions or organs (category features), it may be a promising design paradigm.", + "bbox": [ + 99, + 291, + 483, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this, we try to construct the prototype memory bank with different similarity memory priors to simulate the biological mechanism of \"grandmother cells\", as shown in Figure 1(b). Specifically, similarity memory priors are matched with organ features in medical images through similarity, and a multi-dimensional mapping relationship based on similarity is constructed to extract rich category features in medical images. Meanwhile, this mapping relationship is strengthened by updating the prototype memory bank. In this process, the prototype memory bank is equivalent to a dynamic prior knowledge repository. With the continuous input of new organ features, the similarity memory priors will be continuously optimized to match organ category features more accurately.", + "bbox": [ + 99, + 532, + 483, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we design the Dynamic Memory Weights-Loss Attention (DMW-LA) to directly extract rich category features from medical images. DMW-LA accurately matches the similarity memory priors with the category information in medical images through similarity. This process enables the network to directly capture the abstract category information in medical images, thus significantly enhancing the network's segmentation ability. In addition, to ensure that the similarity memory priors in the prototype memory bank can accurately identify lesions or organs,", + "bbox": [ + 99, + 742, + 483, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we design a novel Weight-Loss Dynamic (W-LD) updating strategy that dynamically updates the prototype memory bank through feature weights and training losses to maintain the effectiveness of similarity memory priors. To deeply explore the hidden patterns and structures in features, we design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand and distinguish subtle differences between different lesions or organs. Finally, we distill the above insights, and construct a Similarity Memory Prior Network (Sim-MPNet) with dual encoder structure by introducing MaxViT [35] to strengthen the learning of global context. In summary, the main contributions of this paper can be summarized as follows:", + "bbox": [ + 511, + 291, + 893, + 494 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel Similarity Memory Prior Network (Sim-MPNet), which can fully integrate and utilize category features and global context in medical images, thus improving the accuracy in segmentation tasks.", + "- We design Dynamic Memory Weights-Loss Attention (DMW-LA). DMW-LA utilizes prototype memory bank to recognize and extract category information, and dynamically adjusts the semantic cluster features through Weight-Loss Dynamic (W-LD) update strategy. This method is a novel feature extraction paradigm that provides a new perspective for the design of medical image segmentation networks.", + "- We design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand the content of medical images. DS-GIM utilizes cosine similarity and euclidean distance to finely distinguish and capture subtle differences between lesions or organs, thus enhancing the network representation ability." + ], + "bbox": [ + 511, + 497, + 893, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 776, + 651, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Medical Image Segmentation. In medical image segmentation, a common method is to segment the image with a full convolutional neural network (FCNN) [11, 15]. Especially after the landmark network U-Net [30] came out, researchers successively launched a series of similar U-shaped Convolutional Neural Network (CNN), such as U", + "bbox": [ + 511, + 801, + 893, + 888 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23010", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Net++ [48], U-Net3+ [20] and nnU-Net [22]. However, CNNs tend to explore the correlation between local elements, which results in poor performance in segmenting large lesions or organs. Therefore, some researchers use Vision Transformer (ViT) [12] for segmentation, such as MEDT [36], DAE-former [2] and DTMFormer [39]. The self-attention mechanism in ViT focuses on information interaction from a global perspective when processing image information, which lacks exploration of the relationships between local elements. Therefore, some scholars combine CNNs with ViTs to improve the local and global modeling ability of networks, such as TransUNet [8], MixTrans [38], UCTransNet [37] and Transfuse [47]. However, this strategy of integrating local details and global context may be difficult to capture useful category features directly, thus limiting the segmentation performance, as shown in Figure 2. Different from the traditional computing paradigm, our method directly extracts and remembers the representative category information from images, which realizes the accurate segmentation of the target contour.", + "bbox": [ + 99, + 102, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype Memory Mechanism. The memory bank can capture the cross-time features in network training and has been proven effective in other fields. Santoro et al. [31] used a memory bank to store new training data for few-shot learning, while Yang and Chan [44] employed it to memorize long-term object information for video tracking. Additionally, Caron et al. [7] utilized memory queues to accumulate negative samples for unsupervised learning, and Li et al. [23] proposed Memory Concept Attention (MOCA) to improve the quality of image generation. However, medical images contain various lesions and organs structures, and these structures exhibit significant differences in different individuals and different disease states. For example, the shape, size, position and surrounding tissues of tumors in medical images are ever-changing. The existing prototype memory mechanism lacks sufficient flexibility to update the representation and memory of different structural features in a timely manner. In our work, a novel dynamic memory prior mechanism is designed, which is called DMW-LA. DMW-LA utilizes prototype memory bank to effectively identify the category features of lesions or organs, and dynamically updates the memory bank through the W-LD update strategy. In addition, DS-GIM is also designed to enhance the difference of feature distribution between different lesions or organs.", + "bbox": [ + 99, + 396, + 486, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 101, + 781, + 233, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 101, + 806, + 209, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 3 shows the overall structure of the Sim-MPNet. Among them, the encoder 1 consists of four stages DMW- LA. We alternately stack DMW-LA and DS-GIM in the third stage. The encoder 2 consists of a four-stage visual", + "bbox": [ + 99, + 830, + 483, + 891 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/669c66d6f905afa256a36325d59af6aabb8392b39821b470d3fcfa68ca370a8d.jpg", + "image_caption": [ + "Figure 3. Overview of Sim-MPNet. The network consists of two encoders, one decoder and one skip connection, and it efficiently models intra-class and global dependencies of input features through the dual encoder structure." + ], + "image_footnote": [], + "bbox": [ + 506, + 99, + 900, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "backbone MaxViT [35]. Each encoder uses overlapping convolution operations to achieve down-sampling. The input image will be re-fused with the stage features of each encoder through bilinear interpolation to reduce information loss caused by resolution reduction. After the encoder completes feature extraction, the network fuses the corresponding stage features of the two encoders, as shown in Equation (1).", + "bbox": [ + 511, + 497, + 895, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {f _ {i}} = E _ {f _ {i}} ^ {1} + E _ {f _ {i}} ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 614, + 893, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $i = \\{1,2,3,4\\}$ , indicating the encoder stage. $E_{f_i}^1$ and $E_{f_i}^2$ represent the features of encoder 1 and encoder 2 in stage $i$ respectively. Then, the $E_{f_1}, E_{f_2}$ and $E_{f_3}$ is passed into the skip connection, $E_{f_4}$ as the input of the decoder. In skip connection, we connect channel attention (CA) and spatial attention (SA) [40] in series to fuse the corresponding features of the encoder and decoder, as shown in Equation (2).", + "bbox": [ + 511, + 641, + 895, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} = S A \\left(C A \\left(E _ {f 4 - i} + D _ {f i}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 762, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $i = \\{1,2,3\\}$ , $D_{f_i}$ represents the output of decoder in stage $i$ , and $O_i$ represents the output of skip connection. Then, $O_i + D_{f_i}$ is used as the input of the next stage. The decoder uses DMW-LA to extract deep semantic information and bilinear interpolation to restore the image to the same resolution as the input image to achieve pixel-level segmentation. Next, we will introduce each module in detail.", + "bbox": [ + 511, + 786, + 895, + 888 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23011", + "bbox": [ + 478, + 926, + 517, + 939 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1cdacf815e0318754cb4359f0e39b7c582bf8d3a6a16c6c4f4ac56aac1d89ec3.jpg", + "image_caption": [ + "(a) DMW-LA" + ], + "image_footnote": [], + "bbox": [ + 107, + 99, + 606, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf92095c259c9add8798ba5b7826bcdca2306d90843341251fc1bc72f8f63a07.jpg", + "image_caption": [ + "(b) W-LD", + "Figure 4. Overview of DMW-LA and W-LD update strategy. (a) DMW-LA consists of intra-cluster interaction and global interaction. (b) The W-LD update strategy adopts weight and loss dual guidance to update the prototype memory bank." + ], + "image_footnote": [], + "bbox": [ + 612, + 99, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic Memory Weight-Loss Attention", + "text_level": 1, + "bbox": [ + 101, + 415, + 444, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we design Dynamic Memory Weight-Loss Attention (DMW-LA) to extract key category features from medical images, as shown in Figure 4(a). The intra-cluster interaction is responsible for obtaining internal dependencies of category features, while global interaction is responsible for initializing prototype memory bank and modeling global dependencies. Assuming input image $X_{1} \\in \\mathbb{R}^{H \\times W \\times C}$ , where $H$ , $W$ , and $C$ refer to the height, width, and channel dimensions of the image, respectively. First, $X_{1}$ is linearly transformed with weights $W^{q}$ , $W^{k}$ and $W^{v}$ to generate Query, Key, and Value for global interaction. At the beginning of network training, we perform K-means on the result $X^{\\prime}$ of global interaction to fill the memory bank. Here, we call the generated semantic clusters as similarity memory prior, and the central point of each cluster is called the similarity core prior. Therefore, the similarity memory prior can be represented as the $\\{\\{E_{11}, E_{12}, \\dots, E_{1c}\\}, \\dots, \\{E_{k1}, E_{k2}, \\dots, E_{kc}\\}\\}$ , and the similarity core prior can be represented as the $\\{\\tilde{E}_1, \\tilde{E}_2, \\tilde{E}_3, \\dots, \\tilde{E}_k\\}$ . After the initialization of the memory bank is completed, we can supplement prior knowledge for network training. Here, we first calculate the cosine similarity between the Query and the similarity core prior to match the corresponding similarity memory prior. Assuming there is $\\{X^{i} \\in \\mathbb{R}^{N \\times C}, N \\leqslant H \\times W\\}$ in the Query that matches the similarity memory prior $\\{E_i = E_{i1}, E_{i2}, \\dots, E_{ic} \\mid E_i^T \\in \\mathbb{R}^{C \\times M}\\}$ , then $E_i$ will perform self-attention with $X^{i}$ , as shown in Equation (3).", + "bbox": [ + 99, + 438, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} = S M \\left(X ^ {i} \\cdot E _ {i} ^ {T}\\right) \\cdot E _ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 872, + 482, + 891 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $i = \\{1,2,3,\\ldots\\}$ , indicating the number of segmented categories, and $SM$ represents the softmax activation function. This step simulates the process of the \"grandmother cell\" recognizing specific visual stimuli, so the $E_{i}$ plays both the role of Key and Value. Then, we concatenate $A_{i}$ to obtain semantic category features $A$ . Finally, we aggregate semantic category features and global features, and then map to the original feature space through the convolution operation $\\psi(\\cdot)$ , as shown in Equation (4).", + "bbox": [ + 511, + 415, + 893, + 547 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {2} = \\psi \\left(A + X ^ {\\prime}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 556, + 893, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the initialization of the memory bank is completed, it still needs to continuously update it to ensure that it can adapt to complex medical images. At present, the common method is to use a random update strategy, that is, randomly insert a portion of the input features into the corresponding semantic clusters in the memory. This method is not suitable for complex and ever-changing medical images and may introduce interference information. Therefore, we propose a Weight-Loss Dynamic (W-LD) update strategy to dynamically update the memory bank, as shown in Figure 4(b). Specifically, assume that $X^{i} \\in \\mathbb{R}^{N \\times C}$ in the output $X_{2} \\in \\mathbb{R}^{H \\times W \\times C}$ of the DMW-LA matches the similarity memory prior $E_{i} \\in \\mathbb{R}^{M \\times C}$ . Then use the Softmax to calculate the weights of $X^{i}$ and $E_{i}$ separately, and sort the columns of $X^{i}$ and $E_{i}$ respectively according to the weight size. Note that the columns with large weight in $X^{i}$ contain more semantic information, while the columns with small weight in $E_{i}$ contain less semantic information. Therefore, we replace the $K$ columns in $E_{i}$ with the smallest weight with the $K$ columns in $X^{i}$ with the largest weight. In addition, we evaluate the quality of the current prototype mem", + "bbox": [ + 511, + 580, + 893, + 890 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23012", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3dfc5627fc4e2104e7a7c0f6f748a6f6960e3c779a8d7b8e02ae78a3a40ad75.jpg", + "image_caption": [ + "Figure 5. Overview of DS-GIM. DS-GIM uses cosine similarity and euclidean distance to comprehensively evaluate and highlight the differences between spatial element structures." + ], + "image_footnote": [], + "bbox": [ + 101, + 99, + 486, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ory bank by calculating the difference between the current epoch loss $(Loss^{\\Delta})$ and the previous epoch loss $(Loss^{\\nabla})$ . Specifically, when the value of $Loss^{\\nabla} - Loss^{\\Delta}$ is greater than 0, it shows that the prototype memory bank has a positive impact on network training, leading to the reduction of the current epoch loss. Therefore, the $K$ should be reduced to preserve these valuable features. Conversely, when the value of $Loss^{\\nabla} - Loss^{\\Delta}$ is less than 0, the $K$ should be increased to expand the range of updates and eliminate useless features as much as possible. In this way, the prototype memory bank can accumulate high-value semantic information to the maximum extent in each training epoch. In addition, a linear function is designed to link $K$ and $Loss^{\\nabla} - Loss^{\\Delta}$ , as shown in Equation (5).", + "bbox": [ + 99, + 311, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = (- \\alpha \\cdot (L o s s ^ {\\nabla} - L o s s ^ {\\Delta}) + \\beta) \\cdot \\theta \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 531, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the above formula, $\\alpha$ and $\\beta$ are decay factors, which are set to 0.5, and $\\theta = M$ . In addition, we set $\\frac{M}{4} \\leq K \\leq \\frac{3M}{4}$ to avoid $K$ falling into extreme situations and ensure stable training, see appendix A. In this way, we can not only minimize the loss of information but also ensure the semantic clarity and representation ability of the similarity memory prior.", + "bbox": [ + 99, + 554, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Double-Similarity Global Internal Enhancement Module", + "text_level": 1, + "bbox": [ + 101, + 666, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The structure of the Double Similarity Global Internal Enhancement Module (DS-GIM) is shown in Figure 5. Assuming the input size is $X \\in \\mathbb{R}^{H \\times W \\times C}$ , we first divide $X$ into $\\left\\{X_{1}, X_{2}, \\dots, X_{\\frac{H \\times W}{p^{2}}} \\mid X_{i} \\in \\mathbb{R}^{P \\times P \\times C}\\right\\}$ with the windows of $P \\times P$ , and calculate the internal cosine similarity of these windows. Weight different decay coefficients for each window based on similarity (from large to small). The purpose is to use the decay coefficient to increase the difference between windows. The formula for calculating the decay coefficient is as follows.", + "bbox": [ + 99, + 702, + 483, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\exp \\left(- \\left(0. 2 5 - 2 ^ {- 2. 5 - \\frac {5 \\cdot n}{l}}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 867, + 482, + 893 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the above formula, $l$ represents the total number of layers in the current stage, $n = \\{0,1,2,\\dots ,l - 1\\}$ . Next, concatenate these windows to form a new feature representation $X^{\\prime}$ . Map $X^{\\prime}$ to a new feature space for feature calculation to obtain $X_{I}\\in \\mathbb{R}^{HW\\times HW}$ . Then average the columns of $X_{I}$ to get $X_{II}\\in \\mathbb{R}^{HW\\times 1}$ , and then copy $X_{II}$ to get a new matrix $S\\in \\mathbb{R}^{HW\\times HW}$ . At this time, the elements in $S$ can be regarded as a set of spatial location elements, and we can calculate the euclidean distance between elements using the formula $Euc = |S - S^T|$ to measure their proximity in the feature space. Based on size, we can divide the internal elements of $Euc$ into $Ecu_{I}$ and $Ecu_{II}$ . Next, these two parts are calculated by the Softmax function and areadamard products with $X_{I}$ respectively. Finally, splice these two parts together to form a new matrix and multiply it with $X^{\\prime}$ , as shown in Equation (7).", + "bbox": [ + 511, + 102, + 893, + 337 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {D} = X ^ {\\prime} \\cdot \\left(S M (E c u _ {I}) * X _ {I} + S M (E c u _ {I I}) * X _ {I}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 342, + 893, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the above equation, $*$ represents hadamard product, and $SM$ represents softmax activation function. In this way, distance information between spatial elements will be embedded into feature encoding to enhance the model's understanding of the internal structure of input data.", + "bbox": [ + 511, + 364, + 893, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 449, + 640, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 472, + 612, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct extensive experiments on four publicly available datasets: Automated Cardiac Diagnosis Challenge (ACDC, 4 classes) [4], Segmentation of Multiple Myeloma Plasma Cells in Microscopic Images (SegPC-2021, 3 classes) [13, 16, 17], International Skin Imaging Collaboration 2018 Challenge (ISIC-2018, 2 classes) [10], and MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge (Synapse, 9 classes). (1) ACDC dataset consists of Magnetic Resonance Imaging (MRI) of the left ventricle (LV), right ventricle (RV) and myocardium (Myo). The division of training set and testing set is consistent with TransUNet++ [42]. (2) SegPC-2021 dataset consists of myeloma plasma cell images. The segmentation target is the cytoplasm and nucleus of the cells. We divide the training set and the testing set by 4:1 ratio. (3) ISIC-2018 dataset consists of dermoscopy images. We also divide the training set and the testing set by 4:1 ratio. (4) Synapse dataset consists of axial abdominal clinical CT images. Each image contains part or all of the eight abdominal organs (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach). We divide the training set and the testing set according to the TransUNet [8].", + "bbox": [ + 511, + 494, + 893, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 823, + 893, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all experiments, the hyperparameters of DS-GIM (decay coefficient) and DMW-LA (linear function) remained un", + "bbox": [ + 511, + 859, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23013", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/70b516c0227c791bea67404ad793781148bcf533a7c755bbd3d40404018635bc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodsACDCSegPC-2021ISIC-2018
DSC↑HD95↓RVMyoLVDSC↑HD95↓CytoplasmNucleusDSC↑HD95↓
CNNU-Net [30]90.141.4889.6186.5894.2380.5435.8980.7780.3089.494.66
Att-UNet [25]90.481.2489.1587.2995.0180.4036.0479.5681.3789.585.11
U-Net++ [48]89.541.2587.9786.6294.0580.0733.1879.4080.1289.194.88
EMCAD [28]91.431.1390.6588.7894.8680.3834.8479.4981.2790.034.33
MADGNet [24]91.131.1790.0188.3095.0780.0233.2179.5580.4889.774.30
SelfReg-UNet [49]91.431.1988.9289.4995.8877.6035.6577.4477.7588.995.74
DconnNet [45]89.961.8089.1686.1294.5879.8433.0779.0080.6890.014.10
ViTSwinUNet [6]89.492.0388.0186.0394.4578.2235.2477.3379.1187.325.47
TransDeepLab [1]89.222.0787.4885.7494.4478.8133.0277.3880.2490.114.07
DAE-former [2]89.781.9189.9184.3895.0478.5234.8677.1079.9389.854.22
SelfReg-SwinUNet [49]91.491.9789.4989.2795.7076.4032.4574.7078.0989.164.90
CNN&ViTTransUNet [8]89.371.4889.2085.2393.6776.6933.9578.6980.6889.924.34
MixTrans [38]90.982.2889.5288.6094.8280.7232.6279.7881.6686.467.34
TransClawUNet [46]91.291.1091.3187.9394.6479.7734.4579.0480.5089.624.81
HiFormer [18]90.122.1591.0684.5494.7779.5432.0877.7381.3589.794.58
TransCASCADE [27]91.481.1291.3788.3294.7680.3632.8779.1881.5389.764.42
PAG-TransYnet [5]91.331.1590.8088.2894.9279.4133.8878.6780.1689.654.53
Memory MechanismSim-MPNet (Ours)92.181.0892.0689.1995.2983.1229.6682.0084.2490.853.55
", + "bbox": [ + 102, + 99, + 893, + 426 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d8caade8590db4c9c095689a94be16d28d9b73f19b37261696cc77815ed02379.jpg", + "table_caption": [ + "Table 1. Performance on ACDC, SegPC-2021 and ISIC-2018. We report the DSC and HD95 scores of different networks on the ACDC, SegPC-2021 and ISIC-2018. We adopt publicly available SOTA methods for implementation, and all results are obtained under the same dataset partitioning conditions. DSC are reported for individual organs or lesions. $\\uparrow (\\downarrow)$ denotes the higher (lower) the better." + ], + "table_footnote": [], + "table_body": "
TypeMethodsP(M)F(G)DSC↑HD95↓AortaGBKLKRLiverPCSPSMTi
CNNU-Net [30]31.0450.1476.8539.7089.0769.7277.7768.6093.4353.9886.6775.58729
Att-UNet [25]34.8834.8877.7736.0289.5568.8877.9871.1193.5758.0487.3075.75761
U-Net++ [48]36.63107.0178.9830.6587.4565.8681.7574.2793.8564.1687.8276.651139
EMCAD [28]26.764.2983.6315.6888.1868.8788.0884.1095.2668.5192.1783.92752
MADGNet [24]32.167.7180.9827.4688.0165.1282.8179.0994.9267.2390.1580.501152
SelfReg-UNet [49]17.2630.7180.3426.9888.7471.7885.3280.7193.8062.2184.7875.39625
DconnNet [45]36.404.8081.5619.0187.4868.5683.7081.0494.6465.2490.6481.15849
ViTSwinUNet [6]27.175.9279.1321.5585.4766.5383.2879.6194.2956.5890.6676.60625
TransDeepLab [1]28.6117.0880.1621.2586.0469.1684.0879.8893.5361.1989.0078.40830
DAE-Former [2]48.0726.0782.4317.4688.9672.3086.0880.8894.9865.1291.9479.191187
SelfReg-SwinUNet [49]27.155.9280.5420.1086.0769.6585.1282.5894.1861.0987.4278.221014
CNN&ViTTransUNet [8]105.3224.6677.4831.6987.2363.1381.8777.0294.0855.8685.0875.62794
MixTrans [38]79.0744.7578.5926.5987.9264.9981.4777.2993.0659.4687.7576.81958
TransClawUNet [46]113.0238.0878.0926.3885.8761.3884.8379.3694.2857.6587.7473.551253
HiFormer [18]29.5211.8880.6919.1487.0368.6184.2378.3794.0760.7790.4482.03931
TransCASCIDE [27]123.4756.9382.6817.3486.6368.4887.6684.5694.4365.3390.7983.521472
PAG-TransYnet [5]144.2233.6483.4315.8289.6768.8986.7484.8895.8768.7592.0180.66940
Memory MechanismSim-MPNet (Ours)91.3043.7484.3414.8587.0673.4988.6585.3495.2869.2591.3684.30873
", + "bbox": [ + 102, + 491, + 895, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on Synapse. P and F represent Parameters and FLOPs respectively. FLOPs of all the methods are reported for ${224} \\times {224}$ inputs,except Sim-MPNet (dual encoder: ${224} \\times {224},{256} \\times {256}$ ). ${T}_{i}$ is the average inference efficiency of a single image in milliseconds. DSC are reported for individual organs. GB, KL, KR, PC, SP and SM represent gallbladder, left kidney, right kidney, pancreas, spleen and stomach respectively.", + "bbox": [ + 99, + 806, + 895, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23014", + "bbox": [ + 478, + 926, + 519, + 939 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75b6e15b9ee26c2ace8a24f9d3adda6328bb58cd2015f0a717e805d2d90d4aa4.jpg", + "image_caption": [ + "Figure 6. Multi-organ and cell segmentation. Comparison of visualization results between Sim-MPNet and other SOTA methods on Synapse and SegPC-2021. The different segmentation targets are represented by different colors." + ], + "image_footnote": [], + "bbox": [ + 106, + 99, + 890, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "changed. For different datasets, just set $\\mathrm{k}$ in $\\mathrm{K}$ -means as the number of categories in the dataset. The input resolutions of two encoders are set to $224 \\times 224$ and $256 \\times 256$ respectively. In addition, we train the model using an AdamW optimizer with a learning rate of 0.0001, and a weight decay of 1e-4. The model is trained on a single RTX 3090 GPU with a batch size 20 using Dice $(L_{dice})$ and Cross-entropy $(L_{ce})$ losses for 300 epochs. The specific proportion of these two losses is $Loss = 0.7 \\cdot L_{dice} + 0.3 \\cdot L_{ce}$ . Finally, all experiments are performed in the Python 3.10.0 and Torch 2.3.0 environment. Only use random rotation and flipping as data augmentation. We choose the Dice score (DSC) and $95\\%$ Hausdorff distance (HD95) as evaluation metrics. The HD95 represents the $95^{th}$ percentile distance between the boundaries of the model prediction graph and real labels.", + "bbox": [ + 99, + 436, + 483, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 101, + 674, + 387, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ACDC. ACDC datasets test the model's ability to explore local details, and ViT-based methods are often difficult to achieve better segmentation results. Table 1 shows that the DSC of SwinUNet, TransDeepLab and DAE-Former does not exceed $90\\%$ . Although the hybrid CNN-ViT methods partially compensate for the shortcomings of ViT, its highest DSC is only $91.48\\%$ . In contrast, Sim-MPNet can directly capture the category features of organs in ACDC by dynamic memory prior mechanism, thus achieving $92.18\\%$ DSC and acquiring better segmentation results.", + "bbox": [ + 99, + 698, + 482, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SegPC-2021. Table 1 shows that the Sim-MPNet outperforms existing methods and achieves a new SOTA on SegPC-2021. Because most of the myeloma plasma cells", + "bbox": [ + 99, + 845, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in SegPC-2021 are closely connected, this requires the network to have a strong ability to extract local details. Therefore, we can observe that the CNN-based methods and the hybrid CNN-ViT methods are better than the ViT-based methods in segmentation effect. However, compared with the optimal network MixTrans in the above methods, Sim-MPNet improved DSC by $2.40\\%$ and HD95 by $2.42\\mathrm{mm}$ . This proves that Sim-MPNet can effectively handle complex medical image.", + "bbox": [ + 511, + 436, + 893, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ISIC-2018. Table 1 shows that the Sim-MPNet also achieved SOTA results on the ISIC-2018 dataset. Since there are many large-scale lesions in the ISIC-2018, it verifies the global modeling ability of the network. Thus, the ViT-based TransDeepLab method has achieved the best segmentation results among the existing methods. However, compared with TransDeepLab method, Sim-MPNet improves DSC by $0.74\\%$ and HD95 by $0.52\\mathrm{mm}$ . This proves once again that Sim-MPNet has the advantage of capturing category dependencies.", + "bbox": [ + 511, + 569, + 895, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synapse. To further verify the performance of Sim-MPNet model in different complex scenarios, we introduce more diverse Synapse datasets for evaluation. Table 2 shows the comparison results on the Synapse. Firstly, CNN-based methods have significant advantages in small organ segmentation due to their excellent local modeling ability. For example, EMCAD achieved DSC of $88.08\\%$ and $84.10\\%$ on the left kidney (Kidney (L)) and right kidney (Kidney (R)), respectively. In contrast, Sim-MPNet performed better on these two organs, with DSC of $0.57\\%$ and $1.24\\%$ higher than EMCAD, respectively. However, on the Spleen", + "bbox": [ + 511, + 715, + 895, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23015", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4cc662e60cdb7ccd2ed545b1fedfcb074d76f195504652185611730f216576ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Encoder 1Encoder 2ACDCSegPC-2021ISIC-2018Synapse
DMW-LA (Ours)—91.5482.6589.5783.46
—MaxViT91.0681.9289.2383.27
MaxViTMaxViT91.3382.0790.0483.59
DMW-LA (Ours)MaxViT92.1883.1290.8584.34
", + "bbox": [ + 102, + 101, + 482, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with a complex contour, Sim-MPNet cannot fully calculate the correlation between local elements, and thus it cannot achieve the optimal segmentation result. Secondly, ViT-based methods have advantages in large-scale organ segmentation due to their global context exploration ability. DAE-Former achieved DSC of $72.30\\%$ and $94.98\\%$ on Gallbladder and Liver, respectively. Sim-MPNet has a DSC of $1.19\\%$ and $0.3\\%$ higher than DAE-Former in these two organs, respectively. Finally, the hybrid CNN-ViT methods combine the advantages of CNN and ViT, enabling the PAG-TransYnet network to explore both local and global features simultaneously, achieving optimal segmentation results in Aorta with the smallest area and Liver with the largest area. Although Sim-MPNet is slightly inferior to PAG-TransYnet in segmentation of the Aorta and Liver, its unique dynamic memory prior mechanism enables it to achieve the best average DSC and HD95 scores on the Synapse. This indicates that Sim-MPNet has excellent robustness in various scenarios. In addition, Sim-MPNet does not have significant advantages in terms of Params and FLOPs, but our method extracts medical image features from a new perspective and achieves excellent segmentation results. This shows the application potential of our method in the field of medical image segmentation and opens up a new path for future research.", + "bbox": [ + 99, + 316, + 482, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Comparison and Analysis. Figure 6 shows the visualizations of Sim-MPNet and other networks on Synapse and SegPC-2021. Firstly, the segmentation results of Sim-MPNet are very close to the ground truth in Synapse. Especially when dealing with organs like the Pancreas, which have complex contours and rich details, Sim-MPNet shows better segmentation performance compared to other networks. Secondly, Sim-MPNet still obtains the best segmentation results for images containing multiple cells in SegPC-2021. Although false positives and false negatives may also occur, the incidence of these errors is significantly reduced compared to other networks. In addition, the semantic clustering receptive fields of DMW-LA on Synapse, ACDC and SegPC-2021 see appendix B.", + "bbox": [ + 99, + 684, + 482, + 890 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd31a3f0d7a3a9ab3af7831cd05c43b8f9aba10463d83a33c278da596b6c7c8e.jpg", + "table_caption": [ + "Table 3. Ablation study of DMW-LA on four datasets. We use different encoders to verify their effectiveness separately. The input resolution of each encoder is the same as that of the comparative experiment, which is $224 \\times 224$ (Encoder 1) and $256 \\times 256$ (Encoder 2) respectively." + ], + "table_footnote": [], + "table_body": "
DatasetACDCSegPC-2021ISIC-2018Synapse
w/o W-LD91.5382.6390.2283.75
w/o DS-GIM91.6782.5890.4783.64
w/o W-LD&DS-GIM91.0282.2390.0783.39
Sim-MPNet92.1883.1290.8584.34
", + "bbox": [ + 513, + 101, + 898, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study of DS-GIM and W-LD update strategy on four datasets. In w/o W-LD, we use a random update strategy to replace the W-LD update strategy. The figures in the table are DSC.", + "bbox": [ + 511, + 210, + 893, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 291, + 661, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of DMW-LA. In this paper, Sim-MPNet adopts dual encoder structure, where Encoder 1 is composed of DMW-LA and Encoder 2 is composed of MaxViT. To evaluate the effectiveness of DMW-LA, we conducted experiments on each encoder. As shown in Table 3, the obvious performance difference indicates that DMW-LA can help the network obtain rich category features, thereby improving segmentation performance.", + "bbox": [ + 511, + 311, + 893, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of DS-GIM and W-LD Update Strategy. We conducted ablation experiments on DS-GIM and W-LD update strategy. It can be seen from Table 4 that compared with the random update strategy, our W-LD update strategy can minimize redundant information and retain useful features. In addition, the remarkable performance improvement in the experiment further confirms the effectiveness of DS-GIM in capturing the internal differences of features.", + "bbox": [ + 511, + 430, + 893, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 560, + 627, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we break the limitations of traditional methods and construct a novel medical image segmentation network, namely Sim-MPNet. Sim-MPNet directly extracts and locates the category features of lesions or organs. We develop two core modules: DMW-LA and DS-GIM. DMW-LA obtains category information through prototype memory bank and W-LD update strategy, while DS-GIM enhances the intrinsic differences between elements through a dual similarity mechanism. The comparative experiments show that Sim-MPNet outperforms the previous SOTA methods in segmenting multiple organs, cells, and lesions. We believe that Sim-MPNet brings new perspectives and inspirations to the design of medical image segmentation networks in future.", + "bbox": [ + 511, + 585, + 893, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 804, + 676, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Tianshan Talent Training Program under Grant 2022TSYCLJ0036, in part by the National Natural Science Foundation of China under Grant 62472368, Grant 62302427 and Grant 62462060.", + "bbox": [ + 511, + 828, + 893, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23016", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 101, + 194, + 116 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reza Azad, Moein Heidari, Moein Shariatnia, Ehsan Khodapanah Aghdam, Sanaz Karimijafarbigloo, Ehsan Adeli, and Dorit Merhof. Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation. In Int. Workshop PRedict. Intell. Med., pages 91-102. Springer, 2022. 6", + "[2] Reza Azad, René Arimond, Ehsan Khodapanah Aghdam, Amirhossein Kazerouni, and Dorit Merhof. Dae-former: Dual attention-guided efficient transformer for medical image segmentation. In In Int. Workshop PRedict. Intell. Med., pages 83-95. Springer, 2023. 3, 6", + "[3] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1", + "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans. Med. Imaging, 37(11):2514-2525, 2018. 5", + "[5] Fares Bougourzi, Fadi Dornaika, Abdelmalik Taleb-Ahmed, and Vinh Truong Hoang. Rethinking attention gated with hybrid dual pyramid transformer-cnn for generalized segmentation in medical imaging. In International Conference on Pattern Recognition, pages 243-258. Springer, 2024. 6", + "[6] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, pages 205-218. Springer, 2022. 1, 6", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Proc. Adv. Neural Inf. Process. Syst., 33:9912-9924, 2020. 3", + "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 3, 5, 6", + "[9] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE transactions on medical imaging, 37(11):2453-2462, 2018. 1", + "[10] Noel Codella, Veronica Rotemberg, Philipp Tschandi, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 5", + "[11] Shuang Cong and Yang Zhou. A review of convolutional neural network architectures and their optimizations. Artif. Intell. Rev., 56(3):1905-1969, 2023. 2" + ], + "bbox": [ + 102, + 125, + 482, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[13] Shiv Gehlot, Anubha Gupta, and Ritu Gupta. Ednfc-net: Convolutional neural network with nested feature concatenation for nuclei-instance segmentation. In ICASSP, pages 1389-1393. IEEE, 2020. 5", + "[14] Charles G Gross. Genealogy of the \"grandmother cell\". Neuroscientist., 8(5):512-518, 2002. 2", + "[15] Jieuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al. Recent advances in convolutional neural networks. Pattern Recognit., 77:354-377, 2018. 2", + "[16] Anubha Gupta, Pramit Mallick, Ojaswa Sharma, Ritu Gupta, and Rahul Duggal. Pcseg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma. PloS one, 13(12):e0207908, 2018. 5", + "[17] Anubha Gupta, Rahul Duggal, Shiv Gehlot, Ritu Gupta, Anvit Mangal, Lalit Kumar, Nisarg Thakkar, and Devprakash Satpathy. Gcti-sn: Geometry-inspired chemical and tissue invariant stain normalization of microscopic medical images. Med. Image Anal., 65:101788, 2020. 5", + "[18] Moein Heidari, Amirhossein Kazerouni, Milad Soltany, Reza Azad, Ehsan Khodapanah Aghdam, Julien Cohen-Adad, and Dorit Merhof. Hformer: Hierarchical multi-scale representations using transformers for medical image segmentation. In WACV, pages 6202–6212, 2023. 6", + "[19] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. J. Digital Imaging, 32:582-596, 2019. 1", + "[20] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet $3+$ : A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 3", + "[21] Ruobing Huang, Mingrong Lin, Haoran Dou, Zehui Lin, Qilong Ying, Xiaohong Jia, Wenwen Xu, Zihan Mei, Xin Yang, Yijie Dong, et al. Boundary-rendering network for breast lesion segmentation in ultrasound images. Med. Image Anal., 80:102478, 2022. 1", + "[22] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods, 18(2):203-211, 2021. 3", + "[23] Tianqin Li, Zijie Li, Harold Rockwell, Amir Farimani, and Tai Sing Lee. Prototype memory and attention mechanisms for few shot image generation. In ICLR, 2022. 3", + "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11480-11491, 2024. 6" + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23017", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 6", + "[26] Rodrigo Quian Quiroga, Itzhak Fried, and Christof Koch. Brain cells for grandmother. Sci. Am., 308(2):30-35, 2013. 2", + "[27] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In WACV, pages 6222-6231, 2023. 6", + "[28] Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. In WACV, pages 11769-11779, 2024. 6", + "[29] KKDSamesh, G Kiran Kumar, K Swapna, Debabrata Datta, and S Suman Rajest. A review of medical image segmentation algorithms. EAI Endorsed Transactions on Pervasive Health & Technology, 7(27), 2021. 1", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 1, 2, 6", + "[31] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In ICML, pages 1842-1850. PMLR, 2016. 3", + "[32] Tracey J Shors. From stem cells to grandmother cells: how neurogenesis relates to learning and memory. Cell Stem Cell, 3(3):253-258, 2008. 2", + "[33] Shiming Tang, Tai Sing Lee, Ming Li, Yimeng Zhang, Yue Xu, Fang Liu, Benjamin Teo, and Hongfei Jiang. Complex pattern selectivity in macaque primary visual cortex revealed by large-scale two-photon imaging. Curr. Biol., 28(1):38-48, 2018. 2", + "[34] Shiming Tang, Yimeng Zhang, Zhihao Li, Ming Li, Fang Liu, Hongfei Jiang, and Tai Sing Lee. Large-scale two-photon imaging revealed super-sparse population codes in the v1 superficial layer of awake monkeys. ELIFE, 7:e33370, 2018. 2", + "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, pages 459-479. Springer, 2022. 2, 3", + "[36] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In MICCAI, pages 36-46. Springer, 2021. 3", + "[37] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 3", + "[38] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390-2394. IEEE, 2022. 3, 6", + "[39] Zhehao Wang, Xian Lin, Nannan Wu, Li Yu, Kwang-Ting Cheng, and Zengqiang Yan. Dtmformer: Dynamic token" + ], + "bbox": [ + 102, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "merging for boosting transformer-based medical image segmentation. In AAAI, pages 5814-5822, 2024. 3", + "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 3", + "[41] Hanguang Xiao, Li Li, Qiyuan Liu, Xiuhong Zhu, and Qihang Zhang. Transformers in medical image segmentation: A review. Biomed. Signal Process. Control, 84:104791, 2023. 1", + "[42] Lianghui Xu, Liejun Wang, Yongming Li, and Anyu Du. Big model and small model: Remote modeling and local information extraction module for medical image segmentation. Appl. Soft Comput., 136:110128, 2023. 5", + "[43] Cheng Xue, Lei Zhu, Huazhu Fu, Xiaowei Hu, Xiaomeng Li, Hai Zhang, and Pheng-Ann Heng. Global guidance network for breast lesion segmentation in ultrasound images. Med. Image Anal., 70:101989, 2021. 1", + "[44] Tianyu Yang and Antoni B Chan. Learning dynamic memory networks for object tracking. In ECCV, pages 152-167, 2018. 3", + "[45] Ziyun Yang and Sina Farsiu. Directional connectivity-based segmentation of medical images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11535, 2023. 6", + "[46] Chang Yao, Menghan Hu, Qingli Li, Guangtao Zhai, and Xiao-Ping Zhang. Transclaw u-net: claw u-net with transformers for medical image segmentation. In ICASSP, pages 280-284. IEEE, 2022. 6", + "[47] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In MICCAI, pages 14-24. Springer, 2021. 3", + "[48] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Proc. Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis Support., pages 3-11. Springer, 2018. 1, 3, 6", + "[49] Wenhui Zhu, Xiwen Chen, Peijie Qiu, Mohammad Farazi, Aristeidis Sotiras, Abolfazl Razi, and Yalin Wang. Selffregunet: Self-regularized unet for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 601–611. Springer, 2024. 6", + "[50] Yanjie Zhu, Jing Cheng, Zhuo-Xu Cui, Qingyong Zhu, Leslie Ying, and Dong Liang. Physics-driven deep learning methods for fast quantitative magnetic resonance imaging: Performance improvements through integration with deep neural networks. IEEE Signal Process Mag., 40(2):116-128, 2023. 1" + ], + "bbox": [ + 514, + 103, + 893, + 760 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23018", + "bbox": [ + 478, + 926, + 517, + 938 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_model.json b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6930448ba1e018993f8259dfba0ba46c8d772ec2 --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_model.json @@ -0,0 +1,1815 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.141, + 0.865, + 0.164 + ], + "angle": 0, + "content": "Similarity Memory Prior is All You Need for Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.189, + 0.714, + 0.208 + ], + "angle": 0, + "content": "Hao Tang\\(^{1}\\), Zhiqing Guo\\(^{1,2,*}\\), Liejun Wang\\(^{1,3,*}\\), Chao Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.207, + 0.756, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Xinjiang University" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.225, + 0.767, + 0.259 + ], + "angle": 0, + "content": "\\(^{2}\\)Xinjiang Multimodal Intelligent Processing and Information Security Engineering Technology Research Center" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.259, + 0.777, + 0.294 + ], + "angle": 0, + "content": "\\(^{3}\\)Silk Road Multilingual Cognitive Computing International Cooperation Joint Laboratory, Xinjiang University" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.296, + 0.796, + 0.31 + ], + "angle": 0, + "content": "{107552201330@stu.，guozhiqing@，wljxju@，107556522208@stu.}xju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.344, + 0.331, + 0.359 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.375, + 0.485, + 0.726 + ], + "angle": 0, + "content": "In recent years, it has been found that \"grandmother cells\" in the primary visual cortex (VI) of macaques can directly recognize visual input with complex shapes. This inspires us to examine the value of these cells in promoting the research of medical image segmentation. In this paper, we design a Similarity Memory Prior Network (Sim-MPNet) for medical image segmentation. Specifically, we propose a Dynamic Memory Weights-Loss Attention (DMW-LA), which matches and remembers the category features of specific lesions or organs in medical images through the similarity memory prior in the prototype memory bank, thus helping the network to learn subtle texture changes between categories. DMW-LA also dynamically updates the similarity memory prior in reverse through Weight-Loss Dynamic (W-LD) update strategy, effectively assisting the network directly extract category features. In addition, we propose the Double-Similarity Global Internal Enhancement Module (DS-GIM) to deeply explore the internal differences in the feature distribution of input data through cosine similarity and euclidean distance. Extensive experiments on four public datasets show that Sim-MPNet has better segmentation performance than other state-of-the-art methods. Our code is available on https://github.com/vpsg-research/Sim-MPNet." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.755, + 0.231, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.779, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Medical image segmentation, as a basic and key technology, is increasingly highlighting its importance in disease diagnosis, treatment planning, and curative effect evaluation [9, 29, 50]. With the rapid development of artificial intelligence and computer vision, constructing automatic and accurate medical image segmentation methods has become a" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.344, + 0.692, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.55, + 0.674, + 0.563 + ], + "angle": 0, + "content": "(a) primate visual system" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.343, + 0.895, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.549, + 0.903, + 0.575 + ], + "angle": 0, + "content": "(b) similarity memory prior directly memorizes target features" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.586, + 0.896, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Motivation. (a) Schematic diagram of the primate visual system. (b) We use similarity memory priors stored in the prototype memory bank to imitate \"grandmother cells\" in V1, and utilize these features to directly match various organs in medical images." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.683, + 0.896, + 0.756 + ], + "angle": 0, + "content": "research hotspot [19, 41]. These methods aim to accurately identify and segment the anatomical structures or pathological features of interest from medical images, to provide clinicians with more intuitive and quantitative medical information [21, 43]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.758, + 0.897, + 0.891 + ], + "angle": 0, + "content": "In existing medical image segmentation works, researchers attempt to use Convolutional Neural Network (CNN) or Vision Transformer (ViT) to enable networks to generate images with precise labels [3, 6, 8, 30, 48]. However, although the local receptive field design of CNNs conform to the translational invariance of images, its inductive bias is essentially a passive response to local textures (edges and corners), rather than actively modeling semantic objects (lesions and organs). ViTs require massive data to learn se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.12, + 0.877, + 0.242, + 0.89 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.519, + 0.94 + ], + "angle": 0, + "content": "23009" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.105, + 0.1, + 0.895, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.224, + 0.896, + 0.266 + ], + "angle": 0, + "content": "Figure 2. Compared with the existing works. CNN passively responds to local texture through convolution kernel. ViT only focuses on high-frequency textures through self-attention. DMW-LA uses similarity memory priors to directly identify and extract category features of organs." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.292, + 0.484, + 0.527 + ], + "angle": 0, + "content": "mantic associations of position encoding, while medical image annotation is scarce, leading to model degradation into \"pseudo-global modeling\" that focuses on high-frequency texture rather than anatomical structure. In addition, CNNs and ViTs belong to data-driven \"implicit learning\" and lack prior knowledge to quickly identify key features. Recently, researchers have used two-photon calcium imaging technology to discover the presence of curvature detecting cells in the V1 [33, 34]. These curvature detecting cells can be regarded as a type of \"grandmother cells\" with specific activation conditions, as shown in Figure 1(a). The \"grandmother cells\" have strong recognition abilities and only respond to specific input stimuli [14, 26, 32]. If the design of medical segmentation networks is not limited to traditional calculation mechanisms and directly focuses on lesions or organs (category features), it may be a promising design paradigm." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.533, + 0.484, + 0.737 + ], + "angle": 0, + "content": "Inspired by this, we try to construct the prototype memory bank with different similarity memory priors to simulate the biological mechanism of \"grandmother cells\", as shown in Figure 1(b). Specifically, similarity memory priors are matched with organ features in medical images through similarity, and a multi-dimensional mapping relationship based on similarity is constructed to extract rich category features in medical images. Meanwhile, this mapping relationship is strengthened by updating the prototype memory bank. In this process, the prototype memory bank is equivalent to a dynamic prior knowledge repository. With the continuous input of new organ features, the similarity memory priors will be continuously optimized to match organ category features more accurately." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.484, + 0.891 + ], + "angle": 0, + "content": "In this paper, we design the Dynamic Memory Weights-Loss Attention (DMW-LA) to directly extract rich category features from medical images. DMW-LA accurately matches the similarity memory priors with the category information in medical images through similarity. This process enables the network to directly capture the abstract category information in medical images, thus significantly enhancing the network's segmentation ability. In addition, to ensure that the similarity memory priors in the prototype memory bank can accurately identify lesions or organs," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.895, + 0.495 + ], + "angle": 0, + "content": "we design a novel Weight-Loss Dynamic (W-LD) updating strategy that dynamically updates the prototype memory bank through feature weights and training losses to maintain the effectiveness of similarity memory priors. To deeply explore the hidden patterns and structures in features, we design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand and distinguish subtle differences between different lesions or organs. Finally, we distill the above insights, and construct a Similarity Memory Prior Network (Sim-MPNet) with dual encoder structure by introducing MaxViT [35] to strengthen the learning of global context. In summary, the main contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.498, + 0.894, + 0.556 + ], + "angle": 0, + "content": "- We design a novel Similarity Memory Prior Network (Sim-MPNet), which can fully integrate and utilize category features and global context in medical images, thus improving the accuracy in segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.895, + 0.673 + ], + "angle": 0, + "content": "- We design Dynamic Memory Weights-Loss Attention (DMW-LA). DMW-LA utilizes prototype memory bank to recognize and extract category information, and dynamically adjusts the semantic cluster features through Weight-Loss Dynamic (W-LD) update strategy. This method is a novel feature extraction paradigm that provides a new perspective for the design of medical image segmentation networks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.895, + 0.762 + ], + "angle": 0, + "content": "- We design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand the content of medical images. DS-GIM utilizes cosine similarity and euclidean distance to finely distinguish and capture subtle differences between lesions or organs, thus enhancing the network representation ability." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.498, + 0.895, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.652, + 0.792 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Medical Image Segmentation. In medical image segmentation, a common method is to segment the image with a full convolutional neural network (FCNN) [11, 15]. Especially after the landmark network U-Net [30] came out, researchers successively launched a series of similar U-shaped Convolutional Neural Network (CNN), such as U" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "23010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.485, + 0.395 + ], + "angle": 0, + "content": "Net++ [48], U-Net3+ [20] and nnU-Net [22]. However, CNNs tend to explore the correlation between local elements, which results in poor performance in segmenting large lesions or organs. Therefore, some researchers use Vision Transformer (ViT) [12] for segmentation, such as MEDT [36], DAE-former [2] and DTMFormer [39]. The self-attention mechanism in ViT focuses on information interaction from a global perspective when processing image information, which lacks exploration of the relationships between local elements. Therefore, some scholars combine CNNs with ViTs to improve the local and global modeling ability of networks, such as TransUNet [8], MixTrans [38], UCTransNet [37] and Transfuse [47]. However, this strategy of integrating local details and global context may be difficult to capture useful category features directly, thus limiting the segmentation performance, as shown in Figure 2. Different from the traditional computing paradigm, our method directly extracts and remembers the representative category information from images, which realizes the accurate segmentation of the target contour." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.397, + 0.487, + 0.765 + ], + "angle": 0, + "content": "Prototype Memory Mechanism. The memory bank can capture the cross-time features in network training and has been proven effective in other fields. Santoro et al. [31] used a memory bank to store new training data for few-shot learning, while Yang and Chan [44] employed it to memorize long-term object information for video tracking. Additionally, Caron et al. [7] utilized memory queues to accumulate negative samples for unsupervised learning, and Li et al. [23] proposed Memory Concept Attention (MOCA) to improve the quality of image generation. However, medical images contain various lesions and organs structures, and these structures exhibit significant differences in different individuals and different disease states. For example, the shape, size, position and surrounding tissues of tumors in medical images are ever-changing. The existing prototype memory mechanism lacks sufficient flexibility to update the representation and memory of different structural features in a timely manner. In our work, a novel dynamic memory prior mechanism is designed, which is called DMW-LA. DMW-LA utilizes prototype memory bank to effectively identify the category features of lesions or organs, and dynamically updates the memory bank through the W-LD update strategy. In addition, DS-GIM is also designed to enhance the difference of feature distribution between different lesions or organs." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.782, + 0.235, + 0.8 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.808, + 0.21, + 0.822 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.831, + 0.485, + 0.892 + ], + "angle": 0, + "content": "Figure 3 shows the overall structure of the Sim-MPNet. Among them, the encoder 1 consists of four stages DMW- LA. We alternately stack DMW-LA and DS-GIM in the third stage. The encoder 2 consists of a four-stage visual" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.101, + 0.901, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.417, + 0.897, + 0.472 + ], + "angle": 0, + "content": "Figure 3. Overview of Sim-MPNet. The network consists of two encoders, one decoder and one skip connection, and it efficiently models intra-class and global dependencies of input features through the dual encoder structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.896, + 0.615 + ], + "angle": 0, + "content": "backbone MaxViT [35]. Each encoder uses overlapping convolution operations to achieve down-sampling. The input image will be re-fused with the stage features of each encoder through bilinear interpolation to reduce information loss caused by resolution reduction. After the encoder completes feature extraction, the network fuses the corresponding stage features of the two encoders, as shown in Equation (1)." + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.616, + 0.895, + 0.636 + ], + "angle": 0, + "content": "\\[\nE _ {f _ {i}} = E _ {f _ {i}} ^ {1} + E _ {f _ {i}} ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.896, + 0.76 + ], + "angle": 0, + "content": "Where \\(i = \\{1,2,3,4\\}\\), indicating the encoder stage. \\(E_{f_i}^1\\) and \\(E_{f_i}^2\\) represent the features of encoder 1 and encoder 2 in stage \\(i\\) respectively. Then, the \\(E_{f_1}, E_{f_2}\\) and \\(E_{f_3}\\) is passed into the skip connection, \\(E_{f_4}\\) as the input of the decoder. In skip connection, we connect channel attention (CA) and spatial attention (SA) [40] in series to fuse the corresponding features of the encoder and decoder, as shown in Equation (2)." + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.763, + 0.895, + 0.78 + ], + "angle": 0, + "content": "\\[\nO _ {i} = S A \\left(C A \\left(E _ {f 4 - i} + D _ {f i}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.896, + 0.89 + ], + "angle": 0, + "content": "Where \\( i = \\{1,2,3\\} \\), \\( D_{f_i} \\) represents the output of decoder in stage \\( i \\), and \\( O_i \\) represents the output of skip connection. Then, \\( O_i + D_{f_i} \\) is used as the input of the next stage. The decoder uses DMW-LA to extract deep semantic information and bilinear interpolation to restore the image to the same resolution as the input image to achieve pixel-level segmentation. Next, we will introduce each module in detail." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.94 + ], + "angle": 0, + "content": "23011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.101, + 0.607, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.339, + 0.398, + 0.352 + ], + "angle": 0, + "content": "(a) DMW-LA" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.101, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.338, + 0.782, + 0.351 + ], + "angle": 0, + "content": "(b) W-LD" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.362, + 0.894, + 0.391 + ], + "angle": 0, + "content": "Figure 4. Overview of DMW-LA and W-LD update strategy. (a) DMW-LA consists of intra-cluster interaction and global interaction. (b) The W-LD update strategy adopts weight and loss dual guidance to update the prototype memory bank." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.416, + 0.445, + 0.431 + ], + "angle": 0, + "content": "3.2. Dynamic Memory Weight-Loss Attention" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.439, + 0.484, + 0.856 + ], + "angle": 0, + "content": "In this paper, we design Dynamic Memory Weight-Loss Attention (DMW-LA) to extract key category features from medical images, as shown in Figure 4(a). The intra-cluster interaction is responsible for obtaining internal dependencies of category features, while global interaction is responsible for initializing prototype memory bank and modeling global dependencies. Assuming input image \\(X_{1} \\in \\mathbb{R}^{H \\times W \\times C}\\), where \\(H\\), \\(W\\), and \\(C\\) refer to the height, width, and channel dimensions of the image, respectively. First, \\(X_{1}\\) is linearly transformed with weights \\(W^{q}\\), \\(W^{k}\\) and \\(W^{v}\\) to generate Query, Key, and Value for global interaction. At the beginning of network training, we perform K-means on the result \\(X^{\\prime}\\) of global interaction to fill the memory bank. Here, we call the generated semantic clusters as similarity memory prior, and the central point of each cluster is called the similarity core prior. Therefore, the similarity memory prior can be represented as the \\(\\{\\{E_{11}, E_{12}, \\dots, E_{1c}\\}, \\dots, \\{E_{k1}, E_{k2}, \\dots, E_{kc}\\}\\}\\), and the similarity core prior can be represented as the \\(\\{\\tilde{E}_1, \\tilde{E}_2, \\tilde{E}_3, \\dots, \\tilde{E}_k\\}\\). After the initialization of the memory bank is completed, we can supplement prior knowledge for network training. Here, we first calculate the cosine similarity between the Query and the similarity core prior to match the corresponding similarity memory prior. Assuming there is \\(\\{X^{i} \\in \\mathbb{R}^{N \\times C}, N \\leqslant H \\times W\\}\\) in the Query that matches the similarity memory prior \\(\\{E_i = E_{i1}, E_{i2}, \\dots, E_{ic} \\mid E_i^T \\in \\mathbb{R}^{C \\times M}\\}\\), then \\(E_i\\) will perform self-attention with \\(X^{i}\\), as shown in Equation (3)." + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.873, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\nA _ {i} = S M \\left(X ^ {i} \\cdot E _ {i} ^ {T}\\right) \\cdot E _ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.895, + 0.549 + ], + "angle": 0, + "content": "Where \\( i = \\{1,2,3,\\ldots\\} \\), indicating the number of segmented categories, and \\( SM \\) represents the softmax activation function. This step simulates the process of the \"grandmother cell\" recognizing specific visual stimuli, so the \\( E_{i} \\) plays both the role of Key and Value. Then, we concatenate \\( A_{i} \\) to obtain semantic category features \\( A \\). Finally, we aggregate semantic category features and global features, and then map to the original feature space through the convolution operation \\( \\psi(\\cdot) \\), as shown in Equation (4)." + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.557, + 0.894, + 0.574 + ], + "angle": 0, + "content": "\\[\nX _ {2} = \\psi \\left(A + X ^ {\\prime}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.895, + 0.891 + ], + "angle": 0, + "content": "After the initialization of the memory bank is completed, it still needs to continuously update it to ensure that it can adapt to complex medical images. At present, the common method is to use a random update strategy, that is, randomly insert a portion of the input features into the corresponding semantic clusters in the memory. This method is not suitable for complex and ever-changing medical images and may introduce interference information. Therefore, we propose a Weight-Loss Dynamic (W-LD) update strategy to dynamically update the memory bank, as shown in Figure 4(b). Specifically, assume that \\( X^{i} \\in \\mathbb{R}^{N \\times C} \\) in the output \\( X_{2} \\in \\mathbb{R}^{H \\times W \\times C} \\) of the DMW-LA matches the similarity memory prior \\( E_{i} \\in \\mathbb{R}^{M \\times C} \\). Then use the Softmax to calculate the weights of \\( X^{i} \\) and \\( E_{i} \\) separately, and sort the columns of \\( X^{i} \\) and \\( E_{i} \\) respectively according to the weight size. Note that the columns with large weight in \\( X^{i} \\) contain more semantic information, while the columns with small weight in \\( E_{i} \\) contain less semantic information. Therefore, we replace the \\( K \\) columns in \\( E_{i} \\) with the smallest weight with the \\( K \\) columns in \\( X^{i} \\) with the largest weight. In addition, we evaluate the quality of the current prototype mem" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "23012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.101, + 0.487, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.287 + ], + "angle": 0, + "content": "Figure 5. Overview of DS-GIM. DS-GIM uses cosine similarity and euclidean distance to comprehensively evaluate and highlight the differences between spatial element structures." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.313, + 0.484, + 0.52 + ], + "angle": 0, + "content": "ory bank by calculating the difference between the current epoch loss \\((Loss^{\\Delta})\\) and the previous epoch loss \\((Loss^{\\nabla})\\). Specifically, when the value of \\(Loss^{\\nabla} - Loss^{\\Delta}\\) is greater than 0, it shows that the prototype memory bank has a positive impact on network training, leading to the reduction of the current epoch loss. Therefore, the \\(K\\) should be reduced to preserve these valuable features. Conversely, when the value of \\(Loss^{\\nabla} - Loss^{\\Delta}\\) is less than 0, the \\(K\\) should be increased to expand the range of updates and eliminate useless features as much as possible. In this way, the prototype memory bank can accumulate high-value semantic information to the maximum extent in each training epoch. In addition, a linear function is designed to link \\(K\\) and \\(Loss^{\\nabla} - Loss^{\\Delta}\\), as shown in Equation (5)." + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.532, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\nK = (- \\alpha \\cdot (L o s s ^ {\\nabla} - L o s s ^ {\\Delta}) + \\beta) \\cdot \\theta \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.555, + 0.484, + 0.658 + ], + "angle": 0, + "content": "In the above formula, \\(\\alpha\\) and \\(\\beta\\) are decay factors, which are set to 0.5, and \\(\\theta = M\\). In addition, we set \\(\\frac{M}{4} \\leq K \\leq \\frac{3M}{4}\\) to avoid \\(K\\) falling into extreme situations and ensure stable training, see appendix A. In this way, we can not only minimize the loss of information but also ensure the semantic clarity and representation ability of the similarity memory prior." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.667, + 0.484, + 0.697 + ], + "angle": 0, + "content": "3.3. Double-Similarity Global Internal Enhancement Module" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.703, + 0.484, + 0.858 + ], + "angle": 0, + "content": "The structure of the Double Similarity Global Internal Enhancement Module (DS-GIM) is shown in Figure 5. Assuming the input size is \\( X \\in \\mathbb{R}^{H \\times W \\times C} \\), we first divide \\( X \\) into \\( \\left\\{X_{1}, X_{2}, \\dots, X_{\\frac{H \\times W}{p^{2}}} \\mid X_{i} \\in \\mathbb{R}^{P \\times P \\times C}\\right\\} \\) with the windows of \\( P \\times P \\), and calculate the internal cosine similarity of these windows. Weight different decay coefficients for each window based on similarity (from large to small). The purpose is to use the decay coefficient to increase the difference between windows. The formula for calculating the decay coefficient is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.868, + 0.483, + 0.894 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\exp \\left(- \\left(0. 2 5 - 2 ^ {- 2. 5 - \\frac {5 \\cdot n}{l}}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.338 + ], + "angle": 0, + "content": "In the above formula, \\(l\\) represents the total number of layers in the current stage, \\(n = \\{0,1,2,\\dots ,l - 1\\}\\). Next, concatenate these windows to form a new feature representation \\(X^{\\prime}\\). Map \\(X^{\\prime}\\) to a new feature space for feature calculation to obtain \\(X_{I}\\in \\mathbb{R}^{HW\\times HW}\\). Then average the columns of \\(X_{I}\\) to get \\(X_{II}\\in \\mathbb{R}^{HW\\times 1}\\), and then copy \\(X_{II}\\) to get a new matrix \\(S\\in \\mathbb{R}^{HW\\times HW}\\). At this time, the elements in \\(S\\) can be regarded as a set of spatial location elements, and we can calculate the euclidean distance between elements using the formula \\(Euc = |S - S^T|\\) to measure their proximity in the feature space. Based on size, we can divide the internal elements of \\(Euc\\) into \\(Ecu_{I}\\) and \\(Ecu_{II}\\). Next, these two parts are calculated by the Softmax function and areadamard products with \\(X_{I}\\) respectively. Finally, splice these two parts together to form a new matrix and multiply it with \\(X^{\\prime}\\), as shown in Equation (7)." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.343, + 0.894, + 0.36 + ], + "angle": 0, + "content": "\\[\nO _ {D} = X ^ {\\prime} \\cdot \\left(S M (E c u _ {I}) * X _ {I} + S M (E c u _ {I I}) * X _ {I}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.894, + 0.438 + ], + "angle": 0, + "content": "In the above equation, \\(*\\) represents hadamard product, and \\(SM\\) represents softmax activation function. In this way, distance information between spatial elements will be embedded into feature encoding to enhance the model's understanding of the internal structure of input data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.45, + 0.642, + 0.466 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.473, + 0.613, + 0.487 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.895, + 0.818 + ], + "angle": 0, + "content": "We conduct extensive experiments on four publicly available datasets: Automated Cardiac Diagnosis Challenge (ACDC, 4 classes) [4], Segmentation of Multiple Myeloma Plasma Cells in Microscopic Images (SegPC-2021, 3 classes) [13, 16, 17], International Skin Imaging Collaboration 2018 Challenge (ISIC-2018, 2 classes) [10], and MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge (Synapse, 9 classes). (1) ACDC dataset consists of Magnetic Resonance Imaging (MRI) of the left ventricle (LV), right ventricle (RV) and myocardium (Myo). The division of training set and testing set is consistent with TransUNet++ [42]. (2) SegPC-2021 dataset consists of myeloma plasma cell images. The segmentation target is the cytoplasm and nucleus of the cells. We divide the training set and the testing set by 4:1 ratio. (3) ISIC-2018 dataset consists of dermoscopy images. We also divide the training set and the testing set by 4:1 ratio. (4) Synapse dataset consists of axial abdominal clinical CT images. Each image contains part or all of the eight abdominal organs (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach). We divide the training set and the testing set according to the TransUNet [8]." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.824, + 0.894, + 0.854 + ], + "angle": 0, + "content": "4.2. Implementation Details and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.86, + 0.894, + 0.89 + ], + "angle": 0, + "content": "In all experiments, the hyperparameters of DS-GIM (decay coefficient) and DMW-LA (linear function) remained un" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "23013" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.101, + 0.894, + 0.428 + ], + "angle": 0, + "content": "
TypeMethodsACDCSegPC-2021ISIC-2018
DSC↑HD95↓RVMyoLVDSC↑HD95↓CytoplasmNucleusDSC↑HD95↓
CNNU-Net [30]90.141.4889.6186.5894.2380.5435.8980.7780.3089.494.66
Att-UNet [25]90.481.2489.1587.2995.0180.4036.0479.5681.3789.585.11
U-Net++ [48]89.541.2587.9786.6294.0580.0733.1879.4080.1289.194.88
EMCAD [28]91.431.1390.6588.7894.8680.3834.8479.4981.2790.034.33
MADGNet [24]91.131.1790.0188.3095.0780.0233.2179.5580.4889.774.30
SelfReg-UNet [49]91.431.1988.9289.4995.8877.6035.6577.4477.7588.995.74
DconnNet [45]89.961.8089.1686.1294.5879.8433.0779.0080.6890.014.10
ViTSwinUNet [6]89.492.0388.0186.0394.4578.2235.2477.3379.1187.325.47
TransDeepLab [1]89.222.0787.4885.7494.4478.8133.0277.3880.2490.114.07
DAE-former [2]89.781.9189.9184.3895.0478.5234.8677.1079.9389.854.22
SelfReg-SwinUNet [49]91.491.9789.4989.2795.7076.4032.4574.7078.0989.164.90
CNN&ViTTransUNet [8]89.371.4889.2085.2393.6776.6933.9578.6980.6889.924.34
MixTrans [38]90.982.2889.5288.6094.8280.7232.6279.7881.6686.467.34
TransClawUNet [46]91.291.1091.3187.9394.6479.7734.4579.0480.5089.624.81
HiFormer [18]90.122.1591.0684.5494.7779.5432.0877.7381.3589.794.58
TransCASCADE [27]91.481.1291.3788.3294.7680.3632.8779.1881.5389.764.42
PAG-TransYnet [5]91.331.1590.8088.2894.9279.4133.8878.6780.1689.654.53
Memory MechanismSim-MPNet (Ours)92.181.0892.0689.1995.2983.1229.6682.0084.2490.853.55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.438, + 0.897, + 0.48 + ], + "angle": 0, + "content": "Table 1. Performance on ACDC, SegPC-2021 and ISIC-2018. We report the DSC and HD95 scores of different networks on the ACDC, SegPC-2021 and ISIC-2018. We adopt publicly available SOTA methods for implementation, and all results are obtained under the same dataset partitioning conditions. DSC are reported for individual organs or lesions. \\(\\uparrow (\\downarrow)\\) denotes the higher (lower) the better." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.492, + 0.896, + 0.798 + ], + "angle": 0, + "content": "
TypeMethodsP(M)F(G)DSC↑HD95↓AortaGBKLKRLiverPCSPSMTi
CNNU-Net [30]31.0450.1476.8539.7089.0769.7277.7768.6093.4353.9886.6775.58729
Att-UNet [25]34.8834.8877.7736.0289.5568.8877.9871.1193.5758.0487.3075.75761
U-Net++ [48]36.63107.0178.9830.6587.4565.8681.7574.2793.8564.1687.8276.651139
EMCAD [28]26.764.2983.6315.6888.1868.8788.0884.1095.2668.5192.1783.92752
MADGNet [24]32.167.7180.9827.4688.0165.1282.8179.0994.9267.2390.1580.501152
SelfReg-UNet [49]17.2630.7180.3426.9888.7471.7885.3280.7193.8062.2184.7875.39625
DconnNet [45]36.404.8081.5619.0187.4868.5683.7081.0494.6465.2490.6481.15849
ViTSwinUNet [6]27.175.9279.1321.5585.4766.5383.2879.6194.2956.5890.6676.60625
TransDeepLab [1]28.6117.0880.1621.2586.0469.1684.0879.8893.5361.1989.0078.40830
DAE-Former [2]48.0726.0782.4317.4688.9672.3086.0880.8894.9865.1291.9479.191187
SelfReg-SwinUNet [49]27.155.9280.5420.1086.0769.6585.1282.5894.1861.0987.4278.221014
CNN&ViTTransUNet [8]105.3224.6677.4831.6987.2363.1381.8777.0294.0855.8685.0875.62794
MixTrans [38]79.0744.7578.5926.5987.9264.9981.4777.2993.0659.4687.7576.81958
TransClawUNet [46]113.0238.0878.0926.3885.8761.3884.8379.3694.2857.6587.7473.551253
HiFormer [18]29.5211.8880.6919.1487.0368.6184.2378.3794.0760.7790.4482.03931
TransCASCIDE [27]123.4756.9382.6817.3486.6368.4887.6684.5694.4365.3390.7983.521472
PAG-TransYnet [5]144.2233.6483.4315.8289.6768.8986.7484.8895.8768.7592.0180.66940
Memory MechanismSim-MPNet (Ours)91.3043.7484.3414.8587.0673.4988.6585.3495.2869.2591.3684.30873
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.808, + 0.897, + 0.864 + ], + "angle": 0, + "content": "Table 2. Performance on Synapse. P and F represent Parameters and FLOPs respectively. FLOPs of all the methods are reported for \\( {224} \\times {224} \\) inputs,except Sim-MPNet (dual encoder: \\( {224} \\times {224},{256} \\times {256} \\) ). \\( {T}_{i} \\) is the average inference efficiency of a single image in milliseconds. DSC are reported for individual organs. GB, KL, KR, PC, SP and SM represent gallbladder, left kidney, right kidney, pancreas, spleen and stomach respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.52, + 0.94 + ], + "angle": 0, + "content": "23014" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.101, + 0.891, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.385, + 0.897, + 0.414 + ], + "angle": 0, + "content": "Figure 6. Multi-organ and cell segmentation. Comparison of visualization results between Sim-MPNet and other SOTA methods on Synapse and SegPC-2021. The different segmentation targets are represented by different colors." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.438, + 0.484, + 0.659 + ], + "angle": 0, + "content": "changed. For different datasets, just set \\( \\mathrm{k} \\) in \\( \\mathrm{K} \\)-means as the number of categories in the dataset. The input resolutions of two encoders are set to \\( 224 \\times 224 \\) and \\( 256 \\times 256 \\) respectively. In addition, we train the model using an AdamW optimizer with a learning rate of 0.0001, and a weight decay of 1e-4. The model is trained on a single RTX 3090 GPU with a batch size 20 using Dice \\( (L_{dice}) \\) and Cross-entropy \\( (L_{ce}) \\) losses for 300 epochs. The specific proportion of these two losses is \\( Loss = 0.7 \\cdot L_{dice} + 0.3 \\cdot L_{ce} \\). Finally, all experiments are performed in the Python 3.10.0 and Torch 2.3.0 environment. Only use random rotation and flipping as data augmentation. We choose the Dice score (DSC) and \\( 95\\% \\) Hausdorff distance (HD95) as evaluation metrics. The HD95 represents the \\( 95^{th} \\) percentile distance between the boundaries of the model prediction graph and real labels." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.675, + 0.388, + 0.691 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.699, + 0.483, + 0.845 + ], + "angle": 0, + "content": "ACDC. ACDC datasets test the model's ability to explore local details, and ViT-based methods are often difficult to achieve better segmentation results. Table 1 shows that the DSC of SwinUNet, TransDeepLab and DAE-Former does not exceed \\(90\\%\\). Although the hybrid CNN-ViT methods partially compensate for the shortcomings of ViT, its highest DSC is only \\(91.48\\%\\). In contrast, Sim-MPNet can directly capture the category features of organs in ACDC by dynamic memory prior mechanism, thus achieving \\(92.18\\%\\) DSC and acquiring better segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.846, + 0.484, + 0.89 + ], + "angle": 0, + "content": "SegPC-2021. Table 1 shows that the Sim-MPNet outperforms existing methods and achieves a new SOTA on SegPC-2021. Because most of the myeloma plasma cells" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.895, + 0.569 + ], + "angle": 0, + "content": "in SegPC-2021 are closely connected, this requires the network to have a strong ability to extract local details. Therefore, we can observe that the CNN-based methods and the hybrid CNN-ViT methods are better than the ViT-based methods in segmentation effect. However, compared with the optimal network MixTrans in the above methods, Sim-MPNet improved DSC by \\(2.40\\%\\) and HD95 by \\(2.42\\mathrm{mm}\\). This proves that Sim-MPNet can effectively handle complex medical image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.896, + 0.716 + ], + "angle": 0, + "content": "ISIC-2018. Table 1 shows that the Sim-MPNet also achieved SOTA results on the ISIC-2018 dataset. Since there are many large-scale lesions in the ISIC-2018, it verifies the global modeling ability of the network. Thus, the ViT-based TransDeepLab method has achieved the best segmentation results among the existing methods. However, compared with TransDeepLab method, Sim-MPNet improves DSC by \\(0.74\\%\\) and HD95 by \\(0.52\\mathrm{mm}\\). This proves once again that Sim-MPNet has the advantage of capturing category dependencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.897, + 0.879 + ], + "angle": 0, + "content": "Synapse. To further verify the performance of Sim-MPNet model in different complex scenarios, we introduce more diverse Synapse datasets for evaluation. Table 2 shows the comparison results on the Synapse. Firstly, CNN-based methods have significant advantages in small organ segmentation due to their excellent local modeling ability. For example, EMCAD achieved DSC of \\(88.08\\%\\) and \\(84.10\\%\\) on the left kidney (Kidney (L)) and right kidney (Kidney (R)), respectively. In contrast, Sim-MPNet performed better on these two organs, with DSC of \\(0.57\\%\\) and \\(1.24\\%\\) higher than EMCAD, respectively. However, on the Spleen" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "23015" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.102, + 0.483, + 0.194 + ], + "angle": 0, + "content": "
Encoder 1Encoder 2ACDCSegPC-2021ISIC-2018Synapse
DMW-LA (Ours)—91.5482.6589.5783.46
—MaxViT91.0681.9289.2383.27
MaxViTMaxViT91.3382.0790.0483.59
DMW-LA (Ours)MaxViT92.1883.1290.8584.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.204, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Table 3. Ablation study of DMW-LA on four datasets. We use different encoders to verify their effectiveness separately. The input resolution of each encoder is the same as that of the comparative experiment, which is \\(224 \\times 224\\) (Encoder 1) and \\(256 \\times 256\\) (Encoder 2) respectively." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.317, + 0.483, + 0.683 + ], + "angle": 0, + "content": "with a complex contour, Sim-MPNet cannot fully calculate the correlation between local elements, and thus it cannot achieve the optimal segmentation result. Secondly, ViT-based methods have advantages in large-scale organ segmentation due to their global context exploration ability. DAE-Former achieved DSC of \\(72.30\\%\\) and \\(94.98\\%\\) on Gallbladder and Liver, respectively. Sim-MPNet has a DSC of \\(1.19\\%\\) and \\(0.3\\%\\) higher than DAE-Former in these two organs, respectively. Finally, the hybrid CNN-ViT methods combine the advantages of CNN and ViT, enabling the PAG-TransYnet network to explore both local and global features simultaneously, achieving optimal segmentation results in Aorta with the smallest area and Liver with the largest area. Although Sim-MPNet is slightly inferior to PAG-TransYnet in segmentation of the Aorta and Liver, its unique dynamic memory prior mechanism enables it to achieve the best average DSC and HD95 scores on the Synapse. This indicates that Sim-MPNet has excellent robustness in various scenarios. In addition, Sim-MPNet does not have significant advantages in terms of Params and FLOPs, but our method extracts medical image features from a new perspective and achieves excellent segmentation results. This shows the application potential of our method in the field of medical image segmentation and opens up a new path for future research." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.685, + 0.483, + 0.891 + ], + "angle": 0, + "content": "Visual Comparison and Analysis. Figure 6 shows the visualizations of Sim-MPNet and other networks on Synapse and SegPC-2021. Firstly, the segmentation results of Sim-MPNet are very close to the ground truth in Synapse. Especially when dealing with organs like the Pancreas, which have complex contours and rich details, Sim-MPNet shows better segmentation performance compared to other networks. Secondly, Sim-MPNet still obtains the best segmentation results for images containing multiple cells in SegPC-2021. Although false positives and false negatives may also occur, the incidence of these errors is significantly reduced compared to other networks. In addition, the semantic clustering receptive fields of DMW-LA on Synapse, ACDC and SegPC-2021 see appendix B." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.102, + 0.9, + 0.202 + ], + "angle": 0, + "content": "
DatasetACDCSegPC-2021ISIC-2018Synapse
w/o W-LD91.5382.6390.2283.75
w/o DS-GIM91.6782.5890.4783.64
w/o W-LD&DS-GIM91.0282.2390.0783.39
Sim-MPNet92.1883.1290.8584.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.212, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Table 4. Ablation study of DS-GIM and W-LD update strategy on four datasets. In w/o W-LD, we use a random update strategy to replace the W-LD update strategy. The figures in the table are DSC." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.292, + 0.663, + 0.308 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.895, + 0.431 + ], + "angle": 0, + "content": "Effectiveness of DMW-LA. In this paper, Sim-MPNet adopts dual encoder structure, where Encoder 1 is composed of DMW-LA and Encoder 2 is composed of MaxViT. To evaluate the effectiveness of DMW-LA, we conducted experiments on each encoder. As shown in Table 3, the obvious performance difference indicates that DMW-LA can help the network obtain rich category features, thereby improving segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Effectiveness of DS-GIM and W-LD Update Strategy. We conducted ablation experiments on DS-GIM and W-LD update strategy. It can be seen from Table 4 that compared with the random update strategy, our W-LD update strategy can minimize redundant information and retain useful features. In addition, the remarkable performance improvement in the experiment further confirms the effectiveness of DS-GIM in capturing the internal differences of features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.629, + 0.577 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.895, + 0.791 + ], + "angle": 0, + "content": "In this paper, we break the limitations of traditional methods and construct a novel medical image segmentation network, namely Sim-MPNet. Sim-MPNet directly extracts and locates the category features of lesions or organs. We develop two core modules: DMW-LA and DS-GIM. DMW-LA obtains category information through prototype memory bank and W-LD update strategy, while DS-GIM enhances the intrinsic differences between elements through a dual similarity mechanism. The comparative experiments show that Sim-MPNet outperforms the previous SOTA methods in segmenting multiple organs, cells, and lesions. We believe that Sim-MPNet brings new perspectives and inspirations to the design of medical image segmentation networks in future." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.805, + 0.677, + 0.822 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.829, + 0.895, + 0.888 + ], + "angle": 0, + "content": "This work was supported in part by the Tianshan Talent Training Program under Grant 2022TSYCLJ0036, in part by the National Natural Science Foundation of China under Grant 62472368, Grant 62302427 and Grant 62462060." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "23016" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.104, + 0.102, + 0.196, + 0.117 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.126, + 0.483, + 0.206 + ], + "angle": 0, + "content": "[1] Reza Azad, Moein Heidari, Moein Shariatnia, Ehsan Khodapanah Aghdam, Sanaz Karimijafarbigloo, Ehsan Adeli, and Dorit Merhof. Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation. In Int. Workshop PRedict. Intell. Med., pages 91-102. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.208, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[2] Reza Azad, René Arimond, Ehsan Khodapanah Aghdam, Amirhossein Kazerouni, and Dorit Merhof. Dae-former: Dual attention-guided efficient transformer for medical image segmentation. In In Int. Workshop PRedict. Intell. Med., pages 83-95. Springer, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.277, + 0.483, + 0.355 + ], + "angle": 0, + "content": "[3] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.358, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans. Med. Imaging, 37(11):2514-2525, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.453, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[5] Fares Bougourzi, Fadi Dornaika, Abdelmalik Taleb-Ahmed, and Vinh Truong Hoang. Rethinking attention gated with hybrid dual pyramid transformer-cnn for generalized segmentation in medical imaging. In International Conference on Pattern Recognition, pages 243-258. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.522, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[6] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, pages 205-218. Springer, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.577, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Proc. Adv. Neural Inf. Process. Syst., 33:9912-9924, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.632, + 0.483, + 0.697 + ], + "angle": 0, + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.7, + 0.483, + 0.753 + ], + "angle": 0, + "content": "[9] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE transactions on medical imaging, 37(11):2453-2462, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.755, + 0.483, + 0.848 + ], + "angle": 0, + "content": "[10] Noel Codella, Veronica Rotemberg, Philipp Tschandi, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.85, + 0.483, + 0.889 + ], + "angle": 0, + "content": "[11] Shuang Cong and Yang Zhou. A review of convolutional neural network architectures and their optimizations. Artif. Intell. Rev., 56(3):1905-1969, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.126, + 0.483, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.184 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.186, + 0.894, + 0.238 + ], + "angle": 0, + "content": "[13] Shiv Gehlot, Anubha Gupta, and Ritu Gupta. Ednfc-net: Convolutional neural network with nested feature concatenation for nuclei-instance segmentation. In ICASSP, pages 1389-1393. IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.24, + 0.894, + 0.266 + ], + "angle": 0, + "content": "[14] Charles G Gross. Genealogy of the \"grandmother cell\". Neuroscientist., 8(5):512-518, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.268, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[15] Jieuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al. Recent advances in convolutional neural networks. Pattern Recognit., 77:354-377, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.335, + 0.894, + 0.389 + ], + "angle": 0, + "content": "[16] Anubha Gupta, Pramit Mallick, Ojaswa Sharma, Ritu Gupta, and Rahul Duggal. Pcseg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma. PloS one, 13(12):e0207908, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.39, + 0.894, + 0.456 + ], + "angle": 0, + "content": "[17] Anubha Gupta, Rahul Duggal, Shiv Gehlot, Ritu Gupta, Anvit Mangal, Lalit Kumar, Nisarg Thakkar, and Devprakash Satpathy. Gcti-sn: Geometry-inspired chemical and tissue invariant stain normalization of microscopic medical images. Med. Image Anal., 65:101788, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.457, + 0.894, + 0.524 + ], + "angle": 0, + "content": "[18] Moein Heidari, Amirhossein Kazerouni, Milad Soltany, Reza Azad, Ehsan Khodapanah Aghdam, Julien Cohen-Adad, and Dorit Merhof. Hformer: Hierarchical multi-scale representations using transformers for medical image segmentation. In WACV, pages 6202–6212, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.894, + 0.579 + ], + "angle": 0, + "content": "[19] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. J. Digital Imaging, 32:582-596, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.58, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[20] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet \\(3+\\): A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.647, + 0.894, + 0.713 + ], + "angle": 0, + "content": "[21] Ruobing Huang, Mingrong Lin, Haoran Dou, Zehui Lin, Qilong Ying, Xiaohong Jia, Wenwen Xu, Zihan Mei, Xin Yang, Yijie Dong, et al. Boundary-rendering network for breast lesion segmentation in ultrasound images. Med. Image Anal., 80:102478, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.715, + 0.894, + 0.768 + ], + "angle": 0, + "content": "[22] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods, 18(2):203-211, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.769, + 0.894, + 0.809 + ], + "angle": 0, + "content": "[23] Tianqin Li, Zijie Li, Harold Rockwell, Amir Farimani, and Tai Sing Lee. Prototype memory and attention mechanisms for few shot image generation. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.81, + 0.894, + 0.889 + ], + "angle": 0, + "content": "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11480-11491, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "23017" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[25] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.173, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[26] Rodrigo Quian Quiroga, Itzhak Fried, and Christof Koch. Brain cells for grandmother. Sci. Am., 308(2):30-35, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.215, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[27] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In WACV, pages 6222-6231, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.257, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[28] Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. In WACV, pages 11769-11779, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.312, + 0.483, + 0.365 + ], + "angle": 0, + "content": "[29] KKDSamesh, G Kiran Kumar, K Swapna, Debabrata Datta, and S Suman Rajest. A review of medical image segmentation algorithms. EAI Endorsed Transactions on Pervasive Health & Technology, 7(27), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.367, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.409, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[31] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In ICML, pages 1842-1850. PMLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.464, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[32] Tracey J Shors. From stem cells to grandmother cells: how neurogenesis relates to learning and memory. Cell Stem Cell, 3(3):253-258, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.506, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[33] Shiming Tang, Tai Sing Lee, Ming Li, Yimeng Zhang, Yue Xu, Fang Liu, Benjamin Teo, and Hongfei Jiang. Complex pattern selectivity in macaque primary visual cortex revealed by large-scale two-photon imaging. Curr. Biol., 28(1):38-48, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.574, + 0.483, + 0.64 + ], + "angle": 0, + "content": "[34] Shiming Tang, Yimeng Zhang, Zhihao Li, Ming Li, Fang Liu, Hongfei Jiang, and Tai Sing Lee. Large-scale two-photon imaging revealed super-sparse population codes in the v1 superficial layer of awake monkeys. ELIFE, 7:e33370, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.643, + 0.483, + 0.696 + ], + "angle": 0, + "content": "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, pages 459-479. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.698, + 0.483, + 0.751 + ], + "angle": 0, + "content": "[36] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In MICCAI, pages 36-46. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.753, + 0.483, + 0.806 + ], + "angle": 0, + "content": "[37] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.808, + 0.483, + 0.861 + ], + "angle": 0, + "content": "[38] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390-2394. IEEE, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.864, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[39] Zhehao Wang, Xian Lin, Nannan Wu, Li Yu, Kwang-Ting Cheng, and Zengqiang Yan. Dtmformer: Dynamic token" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.894, + 0.131 + ], + "angle": 0, + "content": "merging for boosting transformer-based medical image segmentation. In AAAI, pages 5814-5822, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.133, + 0.894, + 0.172 + ], + "angle": 0, + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.175, + 0.894, + 0.226 + ], + "angle": 0, + "content": "[41] Hanguang Xiao, Li Li, Qiyuan Liu, Xiuhong Zhu, and Qihang Zhang. Transformers in medical image segmentation: A review. Biomed. Signal Process. Control, 84:104791, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.229, + 0.894, + 0.282 + ], + "angle": 0, + "content": "[42] Lianghui Xu, Liejun Wang, Yongming Li, and Anyu Du. Big model and small model: Remote modeling and local information extraction module for medical image segmentation. Appl. Soft Comput., 136:110128, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.284, + 0.894, + 0.337 + ], + "angle": 0, + "content": "[43] Cheng Xue, Lei Zhu, Huazhu Fu, Xiaowei Hu, Xiaomeng Li, Hai Zhang, and Pheng-Ann Heng. Global guidance network for breast lesion segmentation in ultrasound images. Med. Image Anal., 70:101989, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.34, + 0.894, + 0.377 + ], + "angle": 0, + "content": "[44] Tianyu Yang and Antoni B Chan. Learning dynamic memory networks for object tracking. In ECCV, pages 152-167, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.381, + 0.894, + 0.434 + ], + "angle": 0, + "content": "[45] Ziyun Yang and Sina Farsiu. Directional connectivity-based segmentation of medical images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11535, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.437, + 0.894, + 0.489 + ], + "angle": 0, + "content": "[46] Chang Yao, Menghan Hu, Qingli Li, Guangtao Zhai, and Xiao-Ping Zhang. Transclaw u-net: claw u-net with transformers for medical image segmentation. In ICASSP, pages 280-284. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.491, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[47] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In MICCAI, pages 14-24. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.533, + 0.894, + 0.599 + ], + "angle": 0, + "content": "[48] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Proc. Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis Support., pages 3-11. Springer, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.601, + 0.894, + 0.681 + ], + "angle": 0, + "content": "[49] Wenhui Zhu, Xiwen Chen, Peijie Qiu, Mohammad Farazi, Aristeidis Sotiras, Abolfazl Razi, and Yalin Wang. Selffregunet: Self-regularized unet for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 601–611. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.683, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[50] Yanjie Zhu, Jing Cheng, Zhuo-Xu Cui, Qingyong Zhu, Leslie Ying, and Dong Liang. Physics-driven deep learning methods for fast quantitative magnetic resonance imaging: Performance improvements through integration with deep neural networks. IEEE Signal Process Mag., 40(2):116-128, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.894, + 0.761 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "23018" + } + ] +] \ No newline at end of file diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_origin.pdf b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0853546bd00e81ea9982ac08da3947948863094e --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/7124b243-3435-4191-a8a2-f64fe05a8d96_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d2e4e4b2fe8fe6841d3c54a1954de07df6fc6a8759b7de14f715f2a94eb6ddc +size 2530425 diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/full.md b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2304e9a662bf1d051ff0dbb90600e65887b30b3f --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/full.md @@ -0,0 +1,244 @@ +# Similarity Memory Prior is All You Need for Medical Image Segmentation + +Hao Tang $^{1}$ , Zhiqing Guo $^{1,2,*}$ , Liejun Wang $^{1,3,*}$ , Chao Liu $^{1}$ + +$^{1}$ School of Computer Science and Technology, Xinjiang University + +$^{2}$ Xinjiang Multimodal Intelligent Processing and Information Security Engineering Technology Research Center + +$^{3}$ Silk Road Multilingual Cognitive Computing International Cooperation Joint Laboratory, Xinjiang University + +{107552201330@stu.，guozhiqing@，wljxju@，107556522208@stu.}xju.edu.cn + +# Abstract + +In recent years, it has been found that "grandmother cells" in the primary visual cortex (VI) of macaques can directly recognize visual input with complex shapes. This inspires us to examine the value of these cells in promoting the research of medical image segmentation. In this paper, we design a Similarity Memory Prior Network (Sim-MPNet) for medical image segmentation. Specifically, we propose a Dynamic Memory Weights-Loss Attention (DMW-LA), which matches and remembers the category features of specific lesions or organs in medical images through the similarity memory prior in the prototype memory bank, thus helping the network to learn subtle texture changes between categories. DMW-LA also dynamically updates the similarity memory prior in reverse through Weight-Loss Dynamic (W-LD) update strategy, effectively assisting the network directly extract category features. In addition, we propose the Double-Similarity Global Internal Enhancement Module (DS-GIM) to deeply explore the internal differences in the feature distribution of input data through cosine similarity and euclidean distance. Extensive experiments on four public datasets show that Sim-MPNet has better segmentation performance than other state-of-the-art methods. Our code is available on https://github.com/vpsg-research/Sim-MPNet. + +# 1. Introduction + +Medical image segmentation, as a basic and key technology, is increasingly highlighting its importance in disease diagnosis, treatment planning, and curative effect evaluation [9, 29, 50]. With the rapid development of artificial intelligence and computer vision, constructing automatic and accurate medical image segmentation methods has become a + +![](images/4fd2a829bc5f7179d5eea0fa0163b9b8640b415d410288568140dec6c8169f0d.jpg) + +![](images/6f763ff7868dddc722125d4e0d6e3d5b37540d6f69bff7f87b8195fe39d58f90.jpg) +(a) primate visual system +(b) similarity memory prior directly memorizes target features +Figure 1. Motivation. (a) Schematic diagram of the primate visual system. (b) We use similarity memory priors stored in the prototype memory bank to imitate "grandmother cells" in V1, and utilize these features to directly match various organs in medical images. + +research hotspot [19, 41]. These methods aim to accurately identify and segment the anatomical structures or pathological features of interest from medical images, to provide clinicians with more intuitive and quantitative medical information [21, 43]. + +In existing medical image segmentation works, researchers attempt to use Convolutional Neural Network (CNN) or Vision Transformer (ViT) to enable networks to generate images with precise labels [3, 6, 8, 30, 48]. However, although the local receptive field design of CNNs conform to the translational invariance of images, its inductive bias is essentially a passive response to local textures (edges and corners), rather than actively modeling semantic objects (lesions and organs). ViTs require massive data to learn se + +![](images/28fe4afcde9d2c5ece0cf7713f07e0b9c4f6058cd21f94a8b7dab42baae99501.jpg) +Figure 2. Compared with the existing works. CNN passively responds to local texture through convolution kernel. ViT only focuses on high-frequency textures through self-attention. DMW-LA uses similarity memory priors to directly identify and extract category features of organs. + +mantic associations of position encoding, while medical image annotation is scarce, leading to model degradation into "pseudo-global modeling" that focuses on high-frequency texture rather than anatomical structure. In addition, CNNs and ViTs belong to data-driven "implicit learning" and lack prior knowledge to quickly identify key features. Recently, researchers have used two-photon calcium imaging technology to discover the presence of curvature detecting cells in the V1 [33, 34]. These curvature detecting cells can be regarded as a type of "grandmother cells" with specific activation conditions, as shown in Figure 1(a). The "grandmother cells" have strong recognition abilities and only respond to specific input stimuli [14, 26, 32]. If the design of medical segmentation networks is not limited to traditional calculation mechanisms and directly focuses on lesions or organs (category features), it may be a promising design paradigm. + +Inspired by this, we try to construct the prototype memory bank with different similarity memory priors to simulate the biological mechanism of "grandmother cells", as shown in Figure 1(b). Specifically, similarity memory priors are matched with organ features in medical images through similarity, and a multi-dimensional mapping relationship based on similarity is constructed to extract rich category features in medical images. Meanwhile, this mapping relationship is strengthened by updating the prototype memory bank. In this process, the prototype memory bank is equivalent to a dynamic prior knowledge repository. With the continuous input of new organ features, the similarity memory priors will be continuously optimized to match organ category features more accurately. + +In this paper, we design the Dynamic Memory Weights-Loss Attention (DMW-LA) to directly extract rich category features from medical images. DMW-LA accurately matches the similarity memory priors with the category information in medical images through similarity. This process enables the network to directly capture the abstract category information in medical images, thus significantly enhancing the network's segmentation ability. In addition, to ensure that the similarity memory priors in the prototype memory bank can accurately identify lesions or organs, + +we design a novel Weight-Loss Dynamic (W-LD) updating strategy that dynamically updates the prototype memory bank through feature weights and training losses to maintain the effectiveness of similarity memory priors. To deeply explore the hidden patterns and structures in features, we design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand and distinguish subtle differences between different lesions or organs. Finally, we distill the above insights, and construct a Similarity Memory Prior Network (Sim-MPNet) with dual encoder structure by introducing MaxViT [35] to strengthen the learning of global context. In summary, the main contributions of this paper can be summarized as follows: + +- We design a novel Similarity Memory Prior Network (Sim-MPNet), which can fully integrate and utilize category features and global context in medical images, thus improving the accuracy in segmentation tasks. +- We design Dynamic Memory Weights-Loss Attention (DMW-LA). DMW-LA utilizes prototype memory bank to recognize and extract category information, and dynamically adjusts the semantic cluster features through Weight-Loss Dynamic (W-LD) update strategy. This method is a novel feature extraction paradigm that provides a new perspective for the design of medical image segmentation networks. +- We design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand the content of medical images. DS-GIM utilizes cosine similarity and euclidean distance to finely distinguish and capture subtle differences between lesions or organs, thus enhancing the network representation ability. + +# 2. Related Work + +Medical Image Segmentation. In medical image segmentation, a common method is to segment the image with a full convolutional neural network (FCNN) [11, 15]. Especially after the landmark network U-Net [30] came out, researchers successively launched a series of similar U-shaped Convolutional Neural Network (CNN), such as U + +Net++ [48], U-Net3+ [20] and nnU-Net [22]. However, CNNs tend to explore the correlation between local elements, which results in poor performance in segmenting large lesions or organs. Therefore, some researchers use Vision Transformer (ViT) [12] for segmentation, such as MEDT [36], DAE-former [2] and DTMFormer [39]. The self-attention mechanism in ViT focuses on information interaction from a global perspective when processing image information, which lacks exploration of the relationships between local elements. Therefore, some scholars combine CNNs with ViTs to improve the local and global modeling ability of networks, such as TransUNet [8], MixTrans [38], UCTransNet [37] and Transfuse [47]. However, this strategy of integrating local details and global context may be difficult to capture useful category features directly, thus limiting the segmentation performance, as shown in Figure 2. Different from the traditional computing paradigm, our method directly extracts and remembers the representative category information from images, which realizes the accurate segmentation of the target contour. + +Prototype Memory Mechanism. The memory bank can capture the cross-time features in network training and has been proven effective in other fields. Santoro et al. [31] used a memory bank to store new training data for few-shot learning, while Yang and Chan [44] employed it to memorize long-term object information for video tracking. Additionally, Caron et al. [7] utilized memory queues to accumulate negative samples for unsupervised learning, and Li et al. [23] proposed Memory Concept Attention (MOCA) to improve the quality of image generation. However, medical images contain various lesions and organs structures, and these structures exhibit significant differences in different individuals and different disease states. For example, the shape, size, position and surrounding tissues of tumors in medical images are ever-changing. The existing prototype memory mechanism lacks sufficient flexibility to update the representation and memory of different structural features in a timely manner. In our work, a novel dynamic memory prior mechanism is designed, which is called DMW-LA. DMW-LA utilizes prototype memory bank to effectively identify the category features of lesions or organs, and dynamically updates the memory bank through the W-LD update strategy. In addition, DS-GIM is also designed to enhance the difference of feature distribution between different lesions or organs. + +# 3. Methodology + +# 3.1. Overview + +Figure 3 shows the overall structure of the Sim-MPNet. Among them, the encoder 1 consists of four stages DMW- LA. We alternately stack DMW-LA and DS-GIM in the third stage. The encoder 2 consists of a four-stage visual + +![](images/669c66d6f905afa256a36325d59af6aabb8392b39821b470d3fcfa68ca370a8d.jpg) +Figure 3. Overview of Sim-MPNet. The network consists of two encoders, one decoder and one skip connection, and it efficiently models intra-class and global dependencies of input features through the dual encoder structure. + +backbone MaxViT [35]. Each encoder uses overlapping convolution operations to achieve down-sampling. The input image will be re-fused with the stage features of each encoder through bilinear interpolation to reduce information loss caused by resolution reduction. After the encoder completes feature extraction, the network fuses the corresponding stage features of the two encoders, as shown in Equation (1). + +$$ +E _ {f _ {i}} = E _ {f _ {i}} ^ {1} + E _ {f _ {i}} ^ {2} \tag {1} +$$ + +Where $i = \{1,2,3,4\}$ , indicating the encoder stage. $E_{f_i}^1$ and $E_{f_i}^2$ represent the features of encoder 1 and encoder 2 in stage $i$ respectively. Then, the $E_{f_1}, E_{f_2}$ and $E_{f_3}$ is passed into the skip connection, $E_{f_4}$ as the input of the decoder. In skip connection, we connect channel attention (CA) and spatial attention (SA) [40] in series to fuse the corresponding features of the encoder and decoder, as shown in Equation (2). + +$$ +O _ {i} = S A \left(C A \left(E _ {f 4 - i} + D _ {f i}\right)\right) \tag {2} +$$ + +Where $i = \{1,2,3\}$ , $D_{f_i}$ represents the output of decoder in stage $i$ , and $O_i$ represents the output of skip connection. Then, $O_i + D_{f_i}$ is used as the input of the next stage. The decoder uses DMW-LA to extract deep semantic information and bilinear interpolation to restore the image to the same resolution as the input image to achieve pixel-level segmentation. Next, we will introduce each module in detail. + +![](images/1cdacf815e0318754cb4359f0e39b7c582bf8d3a6a16c6c4f4ac56aac1d89ec3.jpg) +(a) DMW-LA + +![](images/bf92095c259c9add8798ba5b7826bcdca2306d90843341251fc1bc72f8f63a07.jpg) +(b) W-LD +Figure 4. Overview of DMW-LA and W-LD update strategy. (a) DMW-LA consists of intra-cluster interaction and global interaction. (b) The W-LD update strategy adopts weight and loss dual guidance to update the prototype memory bank. + +# 3.2. Dynamic Memory Weight-Loss Attention + +In this paper, we design Dynamic Memory Weight-Loss Attention (DMW-LA) to extract key category features from medical images, as shown in Figure 4(a). The intra-cluster interaction is responsible for obtaining internal dependencies of category features, while global interaction is responsible for initializing prototype memory bank and modeling global dependencies. Assuming input image $X_{1} \in \mathbb{R}^{H \times W \times C}$ , where $H$ , $W$ , and $C$ refer to the height, width, and channel dimensions of the image, respectively. First, $X_{1}$ is linearly transformed with weights $W^{q}$ , $W^{k}$ and $W^{v}$ to generate Query, Key, and Value for global interaction. At the beginning of network training, we perform K-means on the result $X^{\prime}$ of global interaction to fill the memory bank. Here, we call the generated semantic clusters as similarity memory prior, and the central point of each cluster is called the similarity core prior. Therefore, the similarity memory prior can be represented as the $\{\{E_{11}, E_{12}, \dots, E_{1c}\}, \dots, \{E_{k1}, E_{k2}, \dots, E_{kc}\}\}$ , and the similarity core prior can be represented as the $\{\tilde{E}_1, \tilde{E}_2, \tilde{E}_3, \dots, \tilde{E}_k\}$ . After the initialization of the memory bank is completed, we can supplement prior knowledge for network training. Here, we first calculate the cosine similarity between the Query and the similarity core prior to match the corresponding similarity memory prior. Assuming there is $\{X^{i} \in \mathbb{R}^{N \times C}, N \leqslant H \times W\}$ in the Query that matches the similarity memory prior $\{E_i = E_{i1}, E_{i2}, \dots, E_{ic} \mid E_i^T \in \mathbb{R}^{C \times M}\}$ , then $E_i$ will perform self-attention with $X^{i}$ , as shown in Equation (3). + +$$ +A _ {i} = S M \left(X ^ {i} \cdot E _ {i} ^ {T}\right) \cdot E _ {i} \tag {3} +$$ + +Where $i = \{1,2,3,\ldots\}$ , indicating the number of segmented categories, and $SM$ represents the softmax activation function. This step simulates the process of the "grandmother cell" recognizing specific visual stimuli, so the $E_{i}$ plays both the role of Key and Value. Then, we concatenate $A_{i}$ to obtain semantic category features $A$ . Finally, we aggregate semantic category features and global features, and then map to the original feature space through the convolution operation $\psi(\cdot)$ , as shown in Equation (4). + +$$ +X _ {2} = \psi \left(A + X ^ {\prime}\right) \tag {4} +$$ + +After the initialization of the memory bank is completed, it still needs to continuously update it to ensure that it can adapt to complex medical images. At present, the common method is to use a random update strategy, that is, randomly insert a portion of the input features into the corresponding semantic clusters in the memory. This method is not suitable for complex and ever-changing medical images and may introduce interference information. Therefore, we propose a Weight-Loss Dynamic (W-LD) update strategy to dynamically update the memory bank, as shown in Figure 4(b). Specifically, assume that $X^{i} \in \mathbb{R}^{N \times C}$ in the output $X_{2} \in \mathbb{R}^{H \times W \times C}$ of the DMW-LA matches the similarity memory prior $E_{i} \in \mathbb{R}^{M \times C}$ . Then use the Softmax to calculate the weights of $X^{i}$ and $E_{i}$ separately, and sort the columns of $X^{i}$ and $E_{i}$ respectively according to the weight size. Note that the columns with large weight in $X^{i}$ contain more semantic information, while the columns with small weight in $E_{i}$ contain less semantic information. Therefore, we replace the $K$ columns in $E_{i}$ with the smallest weight with the $K$ columns in $X^{i}$ with the largest weight. In addition, we evaluate the quality of the current prototype mem + +![](images/e3dfc5627fc4e2104e7a7c0f6f748a6f6960e3c779a8d7b8e02ae78a3a40ad75.jpg) +Figure 5. Overview of DS-GIM. DS-GIM uses cosine similarity and euclidean distance to comprehensively evaluate and highlight the differences between spatial element structures. + +ory bank by calculating the difference between the current epoch loss $(Loss^{\Delta})$ and the previous epoch loss $(Loss^{\nabla})$ . Specifically, when the value of $Loss^{\nabla} - Loss^{\Delta}$ is greater than 0, it shows that the prototype memory bank has a positive impact on network training, leading to the reduction of the current epoch loss. Therefore, the $K$ should be reduced to preserve these valuable features. Conversely, when the value of $Loss^{\nabla} - Loss^{\Delta}$ is less than 0, the $K$ should be increased to expand the range of updates and eliminate useless features as much as possible. In this way, the prototype memory bank can accumulate high-value semantic information to the maximum extent in each training epoch. In addition, a linear function is designed to link $K$ and $Loss^{\nabla} - Loss^{\Delta}$ , as shown in Equation (5). + +$$ +K = (- \alpha \cdot (L o s s ^ {\nabla} - L o s s ^ {\Delta}) + \beta) \cdot \theta \tag {5} +$$ + +In the above formula, $\alpha$ and $\beta$ are decay factors, which are set to 0.5, and $\theta = M$ . In addition, we set $\frac{M}{4} \leq K \leq \frac{3M}{4}$ to avoid $K$ falling into extreme situations and ensure stable training, see appendix A. In this way, we can not only minimize the loss of information but also ensure the semantic clarity and representation ability of the similarity memory prior. + +# 3.3. Double-Similarity Global Internal Enhancement Module + +The structure of the Double Similarity Global Internal Enhancement Module (DS-GIM) is shown in Figure 5. Assuming the input size is $X \in \mathbb{R}^{H \times W \times C}$ , we first divide $X$ into $\left\{X_{1}, X_{2}, \dots, X_{\frac{H \times W}{p^{2}}} \mid X_{i} \in \mathbb{R}^{P \times P \times C}\right\}$ with the windows of $P \times P$ , and calculate the internal cosine similarity of these windows. Weight different decay coefficients for each window based on similarity (from large to small). The purpose is to use the decay coefficient to increase the difference between windows. The formula for calculating the decay coefficient is as follows. + +$$ +\gamma = \exp \left(- \left(0. 2 5 - 2 ^ {- 2. 5 - \frac {5 \cdot n}{l}}\right)\right) \tag {6} +$$ + +In the above formula, $l$ represents the total number of layers in the current stage, $n = \{0,1,2,\dots ,l - 1\}$ . Next, concatenate these windows to form a new feature representation $X^{\prime}$ . Map $X^{\prime}$ to a new feature space for feature calculation to obtain $X_{I}\in \mathbb{R}^{HW\times HW}$ . Then average the columns of $X_{I}$ to get $X_{II}\in \mathbb{R}^{HW\times 1}$ , and then copy $X_{II}$ to get a new matrix $S\in \mathbb{R}^{HW\times HW}$ . At this time, the elements in $S$ can be regarded as a set of spatial location elements, and we can calculate the euclidean distance between elements using the formula $Euc = |S - S^T|$ to measure their proximity in the feature space. Based on size, we can divide the internal elements of $Euc$ into $Ecu_{I}$ and $Ecu_{II}$ . Next, these two parts are calculated by the Softmax function and areadamard products with $X_{I}$ respectively. Finally, splice these two parts together to form a new matrix and multiply it with $X^{\prime}$ , as shown in Equation (7). + +$$ +O _ {D} = X ^ {\prime} \cdot \left(S M (E c u _ {I}) * X _ {I} + S M (E c u _ {I I}) * X _ {I}\right) \tag {7} +$$ + +In the above equation, $*$ represents hadamard product, and $SM$ represents softmax activation function. In this way, distance information between spatial elements will be embedded into feature encoding to enhance the model's understanding of the internal structure of input data. + +# 4. Experiments + +# 4.1. Datasets + +We conduct extensive experiments on four publicly available datasets: Automated Cardiac Diagnosis Challenge (ACDC, 4 classes) [4], Segmentation of Multiple Myeloma Plasma Cells in Microscopic Images (SegPC-2021, 3 classes) [13, 16, 17], International Skin Imaging Collaboration 2018 Challenge (ISIC-2018, 2 classes) [10], and MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge (Synapse, 9 classes). (1) ACDC dataset consists of Magnetic Resonance Imaging (MRI) of the left ventricle (LV), right ventricle (RV) and myocardium (Myo). The division of training set and testing set is consistent with TransUNet++ [42]. (2) SegPC-2021 dataset consists of myeloma plasma cell images. The segmentation target is the cytoplasm and nucleus of the cells. We divide the training set and the testing set by 4:1 ratio. (3) ISIC-2018 dataset consists of dermoscopy images. We also divide the training set and the testing set by 4:1 ratio. (4) Synapse dataset consists of axial abdominal clinical CT images. Each image contains part or all of the eight abdominal organs (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach). We divide the training set and the testing set according to the TransUNet [8]. + +# 4.2. Implementation Details and Evaluation Metrics + +In all experiments, the hyperparameters of DS-GIM (decay coefficient) and DMW-LA (linear function) remained un + +
TypeMethodsACDCSegPC-2021ISIC-2018
DSC↑HD95↓RVMyoLVDSC↑HD95↓CytoplasmNucleusDSC↑HD95↓
CNNU-Net [30]90.141.4889.6186.5894.2380.5435.8980.7780.3089.494.66
Att-UNet [25]90.481.2489.1587.2995.0180.4036.0479.5681.3789.585.11
U-Net++ [48]89.541.2587.9786.6294.0580.0733.1879.4080.1289.194.88
EMCAD [28]91.431.1390.6588.7894.8680.3834.8479.4981.2790.034.33
MADGNet [24]91.131.1790.0188.3095.0780.0233.2179.5580.4889.774.30
SelfReg-UNet [49]91.431.1988.9289.4995.8877.6035.6577.4477.7588.995.74
DconnNet [45]89.961.8089.1686.1294.5879.8433.0779.0080.6890.014.10
ViTSwinUNet [6]89.492.0388.0186.0394.4578.2235.2477.3379.1187.325.47
TransDeepLab [1]89.222.0787.4885.7494.4478.8133.0277.3880.2490.114.07
DAE-former [2]89.781.9189.9184.3895.0478.5234.8677.1079.9389.854.22
SelfReg-SwinUNet [49]91.491.9789.4989.2795.7076.4032.4574.7078.0989.164.90
CNN&ViTTransUNet [8]89.371.4889.2085.2393.6776.6933.9578.6980.6889.924.34
MixTrans [38]90.982.2889.5288.6094.8280.7232.6279.7881.6686.467.34
TransClawUNet [46]91.291.1091.3187.9394.6479.7734.4579.0480.5089.624.81
HiFormer [18]90.122.1591.0684.5494.7779.5432.0877.7381.3589.794.58
TransCASCADE [27]91.481.1291.3788.3294.7680.3632.8779.1881.5389.764.42
PAG-TransYnet [5]91.331.1590.8088.2894.9279.4133.8878.6780.1689.654.53
Memory MechanismSim-MPNet (Ours)92.181.0892.0689.1995.2983.1229.6682.0084.2490.853.55
+ +Table 1. Performance on ACDC, SegPC-2021 and ISIC-2018. We report the DSC and HD95 scores of different networks on the ACDC, SegPC-2021 and ISIC-2018. We adopt publicly available SOTA methods for implementation, and all results are obtained under the same dataset partitioning conditions. DSC are reported for individual organs or lesions. $\uparrow (\downarrow)$ denotes the higher (lower) the better. + +
TypeMethodsP(M)F(G)DSC↑HD95↓AortaGBKLKRLiverPCSPSMTi
CNNU-Net [30]31.0450.1476.8539.7089.0769.7277.7768.6093.4353.9886.6775.58729
Att-UNet [25]34.8834.8877.7736.0289.5568.8877.9871.1193.5758.0487.3075.75761
U-Net++ [48]36.63107.0178.9830.6587.4565.8681.7574.2793.8564.1687.8276.651139
EMCAD [28]26.764.2983.6315.6888.1868.8788.0884.1095.2668.5192.1783.92752
MADGNet [24]32.167.7180.9827.4688.0165.1282.8179.0994.9267.2390.1580.501152
SelfReg-UNet [49]17.2630.7180.3426.9888.7471.7885.3280.7193.8062.2184.7875.39625
DconnNet [45]36.404.8081.5619.0187.4868.5683.7081.0494.6465.2490.6481.15849
ViTSwinUNet [6]27.175.9279.1321.5585.4766.5383.2879.6194.2956.5890.6676.60625
TransDeepLab [1]28.6117.0880.1621.2586.0469.1684.0879.8893.5361.1989.0078.40830
DAE-Former [2]48.0726.0782.4317.4688.9672.3086.0880.8894.9865.1291.9479.191187
SelfReg-SwinUNet [49]27.155.9280.5420.1086.0769.6585.1282.5894.1861.0987.4278.221014
CNN&ViTTransUNet [8]105.3224.6677.4831.6987.2363.1381.8777.0294.0855.8685.0875.62794
MixTrans [38]79.0744.7578.5926.5987.9264.9981.4777.2993.0659.4687.7576.81958
TransClawUNet [46]113.0238.0878.0926.3885.8761.3884.8379.3694.2857.6587.7473.551253
HiFormer [18]29.5211.8880.6919.1487.0368.6184.2378.3794.0760.7790.4482.03931
TransCASCIDE [27]123.4756.9382.6817.3486.6368.4887.6684.5694.4365.3390.7983.521472
PAG-TransYnet [5]144.2233.6483.4315.8289.6768.8986.7484.8895.8768.7592.0180.66940
Memory MechanismSim-MPNet (Ours)91.3043.7484.3414.8587.0673.4988.6585.3495.2869.2591.3684.30873
+ +Table 2. Performance on Synapse. P and F represent Parameters and FLOPs respectively. FLOPs of all the methods are reported for ${224} \times {224}$ inputs,except Sim-MPNet (dual encoder: ${224} \times {224},{256} \times {256}$ ). ${T}_{i}$ is the average inference efficiency of a single image in milliseconds. DSC are reported for individual organs. GB, KL, KR, PC, SP and SM represent gallbladder, left kidney, right kidney, pancreas, spleen and stomach respectively. + +![](images/75b6e15b9ee26c2ace8a24f9d3adda6328bb58cd2015f0a717e805d2d90d4aa4.jpg) +Figure 6. Multi-organ and cell segmentation. Comparison of visualization results between Sim-MPNet and other SOTA methods on Synapse and SegPC-2021. The different segmentation targets are represented by different colors. + +changed. For different datasets, just set $\mathrm{k}$ in $\mathrm{K}$ -means as the number of categories in the dataset. The input resolutions of two encoders are set to $224 \times 224$ and $256 \times 256$ respectively. In addition, we train the model using an AdamW optimizer with a learning rate of 0.0001, and a weight decay of 1e-4. The model is trained on a single RTX 3090 GPU with a batch size 20 using Dice $(L_{dice})$ and Cross-entropy $(L_{ce})$ losses for 300 epochs. The specific proportion of these two losses is $Loss = 0.7 \cdot L_{dice} + 0.3 \cdot L_{ce}$ . Finally, all experiments are performed in the Python 3.10.0 and Torch 2.3.0 environment. Only use random rotation and flipping as data augmentation. We choose the Dice score (DSC) and $95\%$ Hausdorff distance (HD95) as evaluation metrics. The HD95 represents the $95^{th}$ percentile distance between the boundaries of the model prediction graph and real labels. + +# 4.3. Comparison with SOTA Methods + +ACDC. ACDC datasets test the model's ability to explore local details, and ViT-based methods are often difficult to achieve better segmentation results. Table 1 shows that the DSC of SwinUNet, TransDeepLab and DAE-Former does not exceed $90\%$ . Although the hybrid CNN-ViT methods partially compensate for the shortcomings of ViT, its highest DSC is only $91.48\%$ . In contrast, Sim-MPNet can directly capture the category features of organs in ACDC by dynamic memory prior mechanism, thus achieving $92.18\%$ DSC and acquiring better segmentation results. + +SegPC-2021. Table 1 shows that the Sim-MPNet outperforms existing methods and achieves a new SOTA on SegPC-2021. Because most of the myeloma plasma cells + +in SegPC-2021 are closely connected, this requires the network to have a strong ability to extract local details. Therefore, we can observe that the CNN-based methods and the hybrid CNN-ViT methods are better than the ViT-based methods in segmentation effect. However, compared with the optimal network MixTrans in the above methods, Sim-MPNet improved DSC by $2.40\%$ and HD95 by $2.42\mathrm{mm}$ . This proves that Sim-MPNet can effectively handle complex medical image. + +ISIC-2018. Table 1 shows that the Sim-MPNet also achieved SOTA results on the ISIC-2018 dataset. Since there are many large-scale lesions in the ISIC-2018, it verifies the global modeling ability of the network. Thus, the ViT-based TransDeepLab method has achieved the best segmentation results among the existing methods. However, compared with TransDeepLab method, Sim-MPNet improves DSC by $0.74\%$ and HD95 by $0.52\mathrm{mm}$ . This proves once again that Sim-MPNet has the advantage of capturing category dependencies. + +Synapse. To further verify the performance of Sim-MPNet model in different complex scenarios, we introduce more diverse Synapse datasets for evaluation. Table 2 shows the comparison results on the Synapse. Firstly, CNN-based methods have significant advantages in small organ segmentation due to their excellent local modeling ability. For example, EMCAD achieved DSC of $88.08\%$ and $84.10\%$ on the left kidney (Kidney (L)) and right kidney (Kidney (R)), respectively. In contrast, Sim-MPNet performed better on these two organs, with DSC of $0.57\%$ and $1.24\%$ higher than EMCAD, respectively. However, on the Spleen + +
Encoder 1Encoder 2ACDCSegPC-2021ISIC-2018Synapse
DMW-LA (Ours)—91.5482.6589.5783.46
—MaxViT91.0681.9289.2383.27
MaxViTMaxViT91.3382.0790.0483.59
DMW-LA (Ours)MaxViT92.1883.1290.8584.34
+ +with a complex contour, Sim-MPNet cannot fully calculate the correlation between local elements, and thus it cannot achieve the optimal segmentation result. Secondly, ViT-based methods have advantages in large-scale organ segmentation due to their global context exploration ability. DAE-Former achieved DSC of $72.30\%$ and $94.98\%$ on Gallbladder and Liver, respectively. Sim-MPNet has a DSC of $1.19\%$ and $0.3\%$ higher than DAE-Former in these two organs, respectively. Finally, the hybrid CNN-ViT methods combine the advantages of CNN and ViT, enabling the PAG-TransYnet network to explore both local and global features simultaneously, achieving optimal segmentation results in Aorta with the smallest area and Liver with the largest area. Although Sim-MPNet is slightly inferior to PAG-TransYnet in segmentation of the Aorta and Liver, its unique dynamic memory prior mechanism enables it to achieve the best average DSC and HD95 scores on the Synapse. This indicates that Sim-MPNet has excellent robustness in various scenarios. In addition, Sim-MPNet does not have significant advantages in terms of Params and FLOPs, but our method extracts medical image features from a new perspective and achieves excellent segmentation results. This shows the application potential of our method in the field of medical image segmentation and opens up a new path for future research. + +Visual Comparison and Analysis. Figure 6 shows the visualizations of Sim-MPNet and other networks on Synapse and SegPC-2021. Firstly, the segmentation results of Sim-MPNet are very close to the ground truth in Synapse. Especially when dealing with organs like the Pancreas, which have complex contours and rich details, Sim-MPNet shows better segmentation performance compared to other networks. Secondly, Sim-MPNet still obtains the best segmentation results for images containing multiple cells in SegPC-2021. Although false positives and false negatives may also occur, the incidence of these errors is significantly reduced compared to other networks. In addition, the semantic clustering receptive fields of DMW-LA on Synapse, ACDC and SegPC-2021 see appendix B. + +Table 3. Ablation study of DMW-LA on four datasets. We use different encoders to verify their effectiveness separately. The input resolution of each encoder is the same as that of the comparative experiment, which is $224 \times 224$ (Encoder 1) and $256 \times 256$ (Encoder 2) respectively. + +
DatasetACDCSegPC-2021ISIC-2018Synapse
w/o W-LD91.5382.6390.2283.75
w/o DS-GIM91.6782.5890.4783.64
w/o W-LD&DS-GIM91.0282.2390.0783.39
Sim-MPNet92.1883.1290.8584.34
+ +Table 4. Ablation study of DS-GIM and W-LD update strategy on four datasets. In w/o W-LD, we use a random update strategy to replace the W-LD update strategy. The figures in the table are DSC. + +# 4.4. Ablation Study + +Effectiveness of DMW-LA. In this paper, Sim-MPNet adopts dual encoder structure, where Encoder 1 is composed of DMW-LA and Encoder 2 is composed of MaxViT. To evaluate the effectiveness of DMW-LA, we conducted experiments on each encoder. As shown in Table 3, the obvious performance difference indicates that DMW-LA can help the network obtain rich category features, thereby improving segmentation performance. + +Effectiveness of DS-GIM and W-LD Update Strategy. We conducted ablation experiments on DS-GIM and W-LD update strategy. It can be seen from Table 4 that compared with the random update strategy, our W-LD update strategy can minimize redundant information and retain useful features. In addition, the remarkable performance improvement in the experiment further confirms the effectiveness of DS-GIM in capturing the internal differences of features. + +# 5. Conclusion + +In this paper, we break the limitations of traditional methods and construct a novel medical image segmentation network, namely Sim-MPNet. Sim-MPNet directly extracts and locates the category features of lesions or organs. We develop two core modules: DMW-LA and DS-GIM. DMW-LA obtains category information through prototype memory bank and W-LD update strategy, while DS-GIM enhances the intrinsic differences between elements through a dual similarity mechanism. The comparative experiments show that Sim-MPNet outperforms the previous SOTA methods in segmenting multiple organs, cells, and lesions. We believe that Sim-MPNet brings new perspectives and inspirations to the design of medical image segmentation networks in future. + +# Acknowledgements + +This work was supported in part by the Tianshan Talent Training Program under Grant 2022TSYCLJ0036, in part by the National Natural Science Foundation of China under Grant 62472368, Grant 62302427 and Grant 62462060. + +# References + +[1] Reza Azad, Moein Heidari, Moein Shariatnia, Ehsan Khodapanah Aghdam, Sanaz Karimijafarbigloo, Ehsan Adeli, and Dorit Merhof. Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation. In Int. Workshop PRedict. Intell. Med., pages 91-102. Springer, 2022. 6 +[2] Reza Azad, René Arimond, Ehsan Khodapanah Aghdam, Amirhossein Kazerouni, and Dorit Merhof. Dae-former: Dual attention-guided efficient transformer for medical image segmentation. In In Int. Workshop PRedict. Intell. Med., pages 83-95. Springer, 2023. 3, 6 +[3] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1 +[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans. Med. Imaging, 37(11):2514-2525, 2018. 5 +[5] Fares Bougourzi, Fadi Dornaika, Abdelmalik Taleb-Ahmed, and Vinh Truong Hoang. Rethinking attention gated with hybrid dual pyramid transformer-cnn for generalized segmentation in medical imaging. In International Conference on Pattern Recognition, pages 243-258. Springer, 2024. 6 +[6] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, pages 205-218. Springer, 2022. 1, 6 +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Proc. Adv. Neural Inf. Process. Syst., 33:9912-9924, 2020. 3 +[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 3, 5, 6 +[9] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE transactions on medical imaging, 37(11):2453-2462, 2018. 1 +[10] Noel Codella, Veronica Rotemberg, Philipp Tschandi, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 5 +[11] Shuang Cong and Yang Zhou. A review of convolutional neural network architectures and their optimizations. Artif. Intell. Rev., 56(3):1905-1969, 2023. 2 + +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[13] Shiv Gehlot, Anubha Gupta, and Ritu Gupta. Ednfc-net: Convolutional neural network with nested feature concatenation for nuclei-instance segmentation. In ICASSP, pages 1389-1393. IEEE, 2020. 5 +[14] Charles G Gross. Genealogy of the "grandmother cell". Neuroscientist., 8(5):512-518, 2002. 2 +[15] Jieuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al. Recent advances in convolutional neural networks. Pattern Recognit., 77:354-377, 2018. 2 +[16] Anubha Gupta, Pramit Mallick, Ojaswa Sharma, Ritu Gupta, and Rahul Duggal. Pcseg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma. PloS one, 13(12):e0207908, 2018. 5 +[17] Anubha Gupta, Rahul Duggal, Shiv Gehlot, Ritu Gupta, Anvit Mangal, Lalit Kumar, Nisarg Thakkar, and Devprakash Satpathy. Gcti-sn: Geometry-inspired chemical and tissue invariant stain normalization of microscopic medical images. Med. Image Anal., 65:101788, 2020. 5 +[18] Moein Heidari, Amirhossein Kazerouni, Milad Soltany, Reza Azad, Ehsan Khodapanah Aghdam, Julien Cohen-Adad, and Dorit Merhof. Hformer: Hierarchical multi-scale representations using transformers for medical image segmentation. In WACV, pages 6202–6212, 2023. 6 +[19] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. J. Digital Imaging, 32:582-596, 2019. 1 +[20] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet $3+$ : A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 3 +[21] Ruobing Huang, Mingrong Lin, Haoran Dou, Zehui Lin, Qilong Ying, Xiaohong Jia, Wenwen Xu, Zihan Mei, Xin Yang, Yijie Dong, et al. Boundary-rendering network for breast lesion segmentation in ultrasound images. Med. Image Anal., 80:102478, 2022. 1 +[22] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods, 18(2):203-211, 2021. 3 +[23] Tianqin Li, Zijie Li, Harold Rockwell, Amir Farimani, and Tai Sing Lee. Prototype memory and attention mechanisms for few shot image generation. In ICLR, 2022. 3 +[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11480-11491, 2024. 6 + +[25] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 6 +[26] Rodrigo Quian Quiroga, Itzhak Fried, and Christof Koch. Brain cells for grandmother. Sci. Am., 308(2):30-35, 2013. 2 +[27] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In WACV, pages 6222-6231, 2023. 6 +[28] Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. In WACV, pages 11769-11779, 2024. 6 +[29] KKDSamesh, G Kiran Kumar, K Swapna, Debabrata Datta, and S Suman Rajest. A review of medical image segmentation algorithms. EAI Endorsed Transactions on Pervasive Health & Technology, 7(27), 2021. 1 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 1, 2, 6 +[31] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In ICML, pages 1842-1850. PMLR, 2016. 3 +[32] Tracey J Shors. From stem cells to grandmother cells: how neurogenesis relates to learning and memory. Cell Stem Cell, 3(3):253-258, 2008. 2 +[33] Shiming Tang, Tai Sing Lee, Ming Li, Yimeng Zhang, Yue Xu, Fang Liu, Benjamin Teo, and Hongfei Jiang. Complex pattern selectivity in macaque primary visual cortex revealed by large-scale two-photon imaging. Curr. Biol., 28(1):38-48, 2018. 2 +[34] Shiming Tang, Yimeng Zhang, Zhihao Li, Ming Li, Fang Liu, Hongfei Jiang, and Tai Sing Lee. Large-scale two-photon imaging revealed super-sparse population codes in the v1 superficial layer of awake monkeys. ELIFE, 7:e33370, 2018. 2 +[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, pages 459-479. Springer, 2022. 2, 3 +[36] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In MICCAI, pages 36-46. Springer, 2021. 3 +[37] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 3 +[38] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390-2394. IEEE, 2022. 3, 6 +[39] Zhehao Wang, Xian Lin, Nannan Wu, Li Yu, Kwang-Ting Cheng, and Zengqiang Yan. Dtmformer: Dynamic token + +merging for boosting transformer-based medical image segmentation. In AAAI, pages 5814-5822, 2024. 3 +[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 3 +[41] Hanguang Xiao, Li Li, Qiyuan Liu, Xiuhong Zhu, and Qihang Zhang. Transformers in medical image segmentation: A review. Biomed. Signal Process. Control, 84:104791, 2023. 1 +[42] Lianghui Xu, Liejun Wang, Yongming Li, and Anyu Du. Big model and small model: Remote modeling and local information extraction module for medical image segmentation. Appl. Soft Comput., 136:110128, 2023. 5 +[43] Cheng Xue, Lei Zhu, Huazhu Fu, Xiaowei Hu, Xiaomeng Li, Hai Zhang, and Pheng-Ann Heng. Global guidance network for breast lesion segmentation in ultrasound images. Med. Image Anal., 70:101989, 2021. 1 +[44] Tianyu Yang and Antoni B Chan. Learning dynamic memory networks for object tracking. In ECCV, pages 152-167, 2018. 3 +[45] Ziyun Yang and Sina Farsiu. Directional connectivity-based segmentation of medical images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11535, 2023. 6 +[46] Chang Yao, Menghan Hu, Qingli Li, Guangtao Zhai, and Xiao-Ping Zhang. Transclaw u-net: claw u-net with transformers for medical image segmentation. In ICASSP, pages 280-284. IEEE, 2022. 6 +[47] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In MICCAI, pages 14-24. Springer, 2021. 3 +[48] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Proc. Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis Support., pages 3-11. Springer, 2018. 1, 3, 6 +[49] Wenhui Zhu, Xiwen Chen, Peijie Qiu, Mohammad Farazi, Aristeidis Sotiras, Abolfazl Razi, and Yalin Wang. Selffregunet: Self-regularized unet for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 601–611. Springer, 2024. 6 +[50] Yanjie Zhu, Jing Cheng, Zhuo-Xu Cui, Qingyong Zhu, Leslie Ying, and Dong Liang. Physics-driven deep learning methods for fast quantitative magnetic resonance imaging: Performance improvements through integration with deep neural networks. IEEE Signal Process Mag., 40(2):116-128, 2023. 1 \ No newline at end of file diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/images.zip b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36ab98cd9d5b6cfe87f19b5049e0c26ef9bbc142 --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5664bba3e9998504ae57532d5432c5fc39074d46a7f7097267c9fec2acc0fcc2 +size 954435 diff --git a/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/layout.json b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05d58079148087ecd9cd66ab0336036cd49a4b7e --- /dev/null +++ b/2025/Similarity Memory Prior is All You Need for Medical Image Segmentation/layout.json @@ -0,0 +1,7895 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 111, + 529, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 111, + 529, + 129 + ], + "spans": [ + { + "bbox": [ + 82, + 111, + 529, + 129 + ], + "type": "text", + "content": "Similarity Memory Prior is All You Need for Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "spans": [ + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "text", + "content": "Hao Tang" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "text", + "content": ", Zhiqing Guo" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "text", + "content": ", Liejun Wang" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "text", + "content": ", Chao Liu" + }, + { + "bbox": [ + 171, + 149, + 436, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 163, + 462, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 163, + 462, + 178 + ], + "spans": [ + { + "bbox": [ + 149, + 163, + 462, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 163, + 462, + 178 + ], + "type": "text", + "content": "School of Computer Science and Technology, Xinjiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 178, + 469, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 178, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 140, + 178, + 469, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 178, + 469, + 205 + ], + "type": "text", + "content": "Xinjiang Multimodal Intelligent Processing and Information Security Engineering Technology Research Center" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 205, + 475, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 205, + 475, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 205, + 475, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 205, + 475, + 232 + ], + "type": "text", + "content": "Silk Road Multilingual Cognitive Computing International Cooperation Joint Laboratory, Xinjiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 234, + 487, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 234, + 487, + 245 + ], + "spans": [ + { + "bbox": [ + 120, + 234, + 487, + 245 + ], + "type": "text", + "content": "{107552201330@stu.，guozhiqing@，wljxju@，107556522208@stu.}xju.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 156, + 272, + 202, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 272, + 202, + 284 + ], + "spans": [ + { + "bbox": [ + 156, + 272, + 202, + 284 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 297, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 574 + ], + "type": "text", + "content": "In recent years, it has been found that \"grandmother cells\" in the primary visual cortex (VI) of macaques can directly recognize visual input with complex shapes. This inspires us to examine the value of these cells in promoting the research of medical image segmentation. In this paper, we design a Similarity Memory Prior Network (Sim-MPNet) for medical image segmentation. Specifically, we propose a Dynamic Memory Weights-Loss Attention (DMW-LA), which matches and remembers the category features of specific lesions or organs in medical images through the similarity memory prior in the prototype memory bank, thus helping the network to learn subtle texture changes between categories. DMW-LA also dynamically updates the similarity memory prior in reverse through Weight-Loss Dynamic (W-LD) update strategy, effectively assisting the network directly extract category features. In addition, we propose the Double-Similarity Global Internal Enhancement Module (DS-GIM) to deeply explore the internal differences in the feature distribution of input data through cosine similarity and euclidean distance. Extensive experiments on four public datasets show that Sim-MPNet has better segmentation performance than other state-of-the-art methods. Our code is available on https://github.com/vpsg-research/Sim-MPNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 597, + 141, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 597, + 141, + 609 + ], + "spans": [ + { + "bbox": [ + 63, + 597, + 141, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 616, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 616, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 61, + 616, + 296, + 687 + ], + "type": "text", + "content": "Medical image segmentation, as a basic and key technology, is increasingly highlighting its importance in disease diagnosis, treatment planning, and curative effect evaluation [9, 29, 50]. With the rapid development of artificial intelligence and computer vision, constructing automatic and accurate medical image segmentation methods has become a" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 272, + 423, + 413 + ], + "blocks": [ + { + "bbox": [ + 309, + 272, + 423, + 413 + ], + "lines": [ + { + "bbox": [ + 309, + 272, + 423, + 413 + ], + "spans": [ + { + "bbox": [ + 309, + 272, + 423, + 413 + ], + "type": "image", + "image_path": "4fd2a829bc5f7179d5eea0fa0163b9b8640b415d410288568140dec6c8169f0d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 426, + 271, + 547, + 430 + ], + "blocks": [ + { + "bbox": [ + 324, + 435, + 412, + 445 + ], + "lines": [ + { + "bbox": [ + 324, + 435, + 412, + 445 + ], + "spans": [ + { + "bbox": [ + 324, + 435, + 412, + 445 + ], + "type": "text", + "content": "(a) primate visual system" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 271, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 426, + 271, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 426, + 271, + 547, + 430 + ], + "type": "image", + "image_path": "6f763ff7868dddc722125d4e0d6e3d5b37540d6f69bff7f87b8195fe39d58f90.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 434, + 552, + 455 + ], + "lines": [ + { + "bbox": [ + 429, + 434, + 552, + 455 + ], + "spans": [ + { + "bbox": [ + 429, + 434, + 552, + 455 + ], + "type": "text", + "content": "(b) similarity memory prior directly memorizes target features" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 464, + 548, + 518 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 548, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 548, + 518 + ], + "type": "text", + "content": "Figure 1. Motivation. (a) Schematic diagram of the primate visual system. (b) We use similarity memory priors stored in the prototype memory bank to imitate \"grandmother cells\" in V1, and utilize these features to directly match various organs in medical images." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 540, + 548, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 540, + 548, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 540, + 548, + 598 + ], + "type": "text", + "content": "research hotspot [19, 41]. These methods aim to accurately identify and segment the anatomical structures or pathological features of interest from medical images, to provide clinicians with more intuitive and quantitative medical information [21, 43]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "type": "text", + "content": "In existing medical image segmentation works, researchers attempt to use Convolutional Neural Network (CNN) or Vision Transformer (ViT) to enable networks to generate images with precise labels [3, 6, 8, 30, 48]. However, although the local receptive field design of CNNs conform to the translational invariance of images, its inductive bias is essentially a passive response to local textures (edges and corners), rather than actively modeling semantic objects (lesions and organs). ViTs require massive data to learn se" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 73, + 694, + 148, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 694, + 148, + 704 + ], + "spans": [ + { + "bbox": [ + 73, + 694, + 148, + 704 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "type": "text", + "content": "23009" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 79, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 64, + 79, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 64, + 79, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 64, + 79, + 547, + 171 + ], + "type": "image", + "image_path": "28fe4afcde9d2c5ece0cf7713f07e0b9c4f6058cd21f94a8b7dab42baae99501.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 177, + 548, + 210 + ], + "lines": [ + { + "bbox": [ + 61, + 177, + 548, + 210 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 548, + 210 + ], + "type": "text", + "content": "Figure 2. Compared with the existing works. CNN passively responds to local texture through convolution kernel. ViT only focuses on high-frequency textures through self-attention. DMW-LA uses similarity memory priors to directly identify and extract category features of organs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 231, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 231, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 231, + 296, + 417 + ], + "type": "text", + "content": "mantic associations of position encoding, while medical image annotation is scarce, leading to model degradation into \"pseudo-global modeling\" that focuses on high-frequency texture rather than anatomical structure. In addition, CNNs and ViTs belong to data-driven \"implicit learning\" and lack prior knowledge to quickly identify key features. Recently, researchers have used two-photon calcium imaging technology to discover the presence of curvature detecting cells in the V1 [33, 34]. These curvature detecting cells can be regarded as a type of \"grandmother cells\" with specific activation conditions, as shown in Figure 1(a). The \"grandmother cells\" have strong recognition abilities and only respond to specific input stimuli [14, 26, 32]. If the design of medical segmentation networks is not limited to traditional calculation mechanisms and directly focuses on lesions or organs (category features), it may be a promising design paradigm." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 422, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 422, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 61, + 422, + 296, + 583 + ], + "type": "text", + "content": "Inspired by this, we try to construct the prototype memory bank with different similarity memory priors to simulate the biological mechanism of \"grandmother cells\", as shown in Figure 1(b). Specifically, similarity memory priors are matched with organ features in medical images through similarity, and a multi-dimensional mapping relationship based on similarity is constructed to extract rich category features in medical images. Meanwhile, this mapping relationship is strengthened by updating the prototype memory bank. In this process, the prototype memory bank is equivalent to a dynamic prior knowledge repository. With the continuous input of new organ features, the similarity memory priors will be continuously optimized to match organ category features more accurately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 588, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 296, + 705 + ], + "type": "text", + "content": "In this paper, we design the Dynamic Memory Weights-Loss Attention (DMW-LA) to directly extract rich category features from medical images. DMW-LA accurately matches the similarity memory priors with the category information in medical images through similarity. This process enables the network to directly capture the abstract category information in medical images, thus significantly enhancing the network's segmentation ability. In addition, to ensure that the similarity memory priors in the prototype memory bank can accurately identify lesions or organs," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 231, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 547, + 392 + ], + "type": "text", + "content": "we design a novel Weight-Loss Dynamic (W-LD) updating strategy that dynamically updates the prototype memory bank through feature weights and training losses to maintain the effectiveness of similarity memory priors. To deeply explore the hidden patterns and structures in features, we design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand and distinguish subtle differences between different lesions or organs. Finally, we distill the above insights, and construct a Similarity Memory Prior Network (Sim-MPNet) with dual encoder structure by introducing MaxViT [35] to strengthen the learning of global context. In summary, the main contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 394, + 547, + 603 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 394, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 547, + 440 + ], + "type": "text", + "content": "- We design a novel Similarity Memory Prior Network (Sim-MPNet), which can fully integrate and utilize category features and global context in medical images, thus improving the accuracy in segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 441, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 547, + 533 + ], + "type": "text", + "content": "- We design Dynamic Memory Weights-Loss Attention (DMW-LA). DMW-LA utilizes prototype memory bank to recognize and extract category information, and dynamically adjusts the semantic cluster features through Weight-Loss Dynamic (W-LD) update strategy. This method is a novel feature extraction paradigm that provides a new perspective for the design of medical image segmentation networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 547, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 547, + 603 + ], + "type": "text", + "content": "- We design a Double-Similarity Global Internal Enhancement Module (DS-GIM) to help the network better understand the content of medical images. DS-GIM utilizes cosine similarity and euclidean distance to finely distinguish and capture subtle differences between lesions or organs, thus enhancing the network representation ability." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 615, + 399, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 399, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 399, + 627 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 547, + 704 + ], + "type": "text", + "content": "Medical Image Segmentation. In medical image segmentation, a common method is to segment the image with a full convolutional neural network (FCNN) [11, 15]. Especially after the landmark network U-Net [30] came out, researchers successively launched a series of similar U-shaped Convolutional Neural Network (CNN), such as U" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23010" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 296, + 312 + ], + "type": "text", + "content": "Net++ [48], U-Net3+ [20] and nnU-Net [22]. However, CNNs tend to explore the correlation between local elements, which results in poor performance in segmenting large lesions or organs. Therefore, some researchers use Vision Transformer (ViT) [12] for segmentation, such as MEDT [36], DAE-former [2] and DTMFormer [39]. The self-attention mechanism in ViT focuses on information interaction from a global perspective when processing image information, which lacks exploration of the relationships between local elements. Therefore, some scholars combine CNNs with ViTs to improve the local and global modeling ability of networks, such as TransUNet [8], MixTrans [38], UCTransNet [37] and Transfuse [47]. However, this strategy of integrating local details and global context may be difficult to capture useful category features directly, thus limiting the segmentation performance, as shown in Figure 2. Different from the traditional computing paradigm, our method directly extracts and remembers the representative category information from images, which realizes the accurate segmentation of the target contour." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 314, + 298, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 314, + 298, + 605 + ], + "spans": [ + { + "bbox": [ + 61, + 314, + 298, + 605 + ], + "type": "text", + "content": "Prototype Memory Mechanism. The memory bank can capture the cross-time features in network training and has been proven effective in other fields. Santoro et al. [31] used a memory bank to store new training data for few-shot learning, while Yang and Chan [44] employed it to memorize long-term object information for video tracking. Additionally, Caron et al. [7] utilized memory queues to accumulate negative samples for unsupervised learning, and Li et al. [23] proposed Memory Concept Attention (MOCA) to improve the quality of image generation. However, medical images contain various lesions and organs structures, and these structures exhibit significant differences in different individuals and different disease states. For example, the shape, size, position and surrounding tissues of tumors in medical images are ever-changing. The existing prototype memory mechanism lacks sufficient flexibility to update the representation and memory of different structural features in a timely manner. In our work, a novel dynamic memory prior mechanism is designed, which is called DMW-LA. DMW-LA utilizes prototype memory bank to effectively identify the category features of lesions or organs, and dynamically updates the memory bank through the W-LD update strategy. In addition, DS-GIM is also designed to enhance the difference of feature distribution between different lesions or organs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 619, + 143, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 619, + 143, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 619, + 143, + 633 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 62, + 639, + 128, + 651 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 296, + 706 + ], + "type": "text", + "content": "Figure 3 shows the overall structure of the Sim-MPNet. Among them, the encoder 1 consists of four stages DMW- LA. We alternately stack DMW-LA and DS-GIM in the third stage. The encoder 2 consists of a four-stage visual" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 79, + 551, + 322 + ], + "blocks": [ + { + "bbox": [ + 310, + 79, + 551, + 322 + ], + "lines": [ + { + "bbox": [ + 310, + 79, + 551, + 322 + ], + "spans": [ + { + "bbox": [ + 310, + 79, + 551, + 322 + ], + "type": "image", + "image_path": "669c66d6f905afa256a36325d59af6aabb8392b39821b470d3fcfa68ca370a8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 330, + 548, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 330, + 548, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 548, + 373 + ], + "type": "text", + "content": "Figure 3. Overview of Sim-MPNet. The network consists of two encoders, one decoder and one skip connection, and it efficiently models intra-class and global dependencies of input features through the dual encoder structure." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 394, + 548, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 548, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 548, + 487 + ], + "type": "text", + "content": "backbone MaxViT [35]. Each encoder uses overlapping convolution operations to achieve down-sampling. The input image will be re-fused with the stage features of each encoder through bilinear interpolation to reduce information loss caused by resolution reduction. After the encoder completes feature extraction, the network fuses the corresponding stage features of the two encoders, as shown in Equation (1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 394, + 487, + 547, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 487, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 394, + 487, + 547, + 503 + ], + "type": "interline_equation", + "content": "E _ {f _ {i}} = E _ {f _ {i}} ^ {1} + E _ {f _ {i}} ^ {2} \\tag {1}", + "image_path": "cff9c2255af308449f950c2a5214027f1da6983f0309aef91cb608eb92296d68.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "i = \\{1,2,3,4\\}" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": ", indicating the encoder stage. " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "E_{f_i}^1" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "E_{f_i}^2" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " represent the features of encoder 1 and encoder 2 in stage " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " respectively. Then, the " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "E_{f_1}, E_{f_2}" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "E_{f_3}" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " is passed into the skip connection, " + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "inline_equation", + "content": "E_{f_4}" + }, + { + "bbox": [ + 313, + 508, + 548, + 601 + ], + "type": "text", + "content": " as the input of the decoder. In skip connection, we connect channel attention (CA) and spatial attention (SA) [40] in series to fuse the corresponding features of the encoder and decoder, as shown in Equation (2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 604, + 547, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 604, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 369, + 604, + 547, + 617 + ], + "type": "interline_equation", + "content": "O _ {i} = S A \\left(C A \\left(E _ {f 4 - i} + D _ {f i}\\right)\\right) \\tag {2}", + "image_path": "dc41a40ca7865eeebee3f63bc5e5d6fc25adf3c457a217936cff1611b92d3346.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "inline_equation", + "content": "i = \\{1,2,3\\}" + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "inline_equation", + "content": "D_{f_i}" + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": " represents the output of decoder in stage " + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": " represents the output of skip connection. Then, " + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "inline_equation", + "content": "O_i + D_{f_i}" + }, + { + "bbox": [ + 313, + 623, + 548, + 704 + ], + "type": "text", + "content": " is used as the input of the next stage. The decoder uses DMW-LA to extract deep semantic information and bilinear interpolation to restore the image to the same resolution as the input image to achieve pixel-level segmentation. Next, we will introduce each module in detail." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 744 + ], + "type": "text", + "content": "23011" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 79, + 371, + 266 + ], + "blocks": [ + { + "bbox": [ + 66, + 79, + 371, + 266 + ], + "lines": [ + { + "bbox": [ + 66, + 79, + 371, + 266 + ], + "spans": [ + { + "bbox": [ + 66, + 79, + 371, + 266 + ], + "type": "image", + "image_path": "1cdacf815e0318754cb4359f0e39b7c582bf8d3a6a16c6c4f4ac56aac1d89ec3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 268, + 243, + 278 + ], + "lines": [ + { + "bbox": [ + 181, + 268, + 243, + 278 + ], + "spans": [ + { + "bbox": [ + 181, + 268, + 243, + 278 + ], + "type": "text", + "content": "(a) DMW-LA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 375, + 79, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 375, + 79, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 375, + 79, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 375, + 79, + 545, + 266 + ], + "type": "image", + "image_path": "bf92095c259c9add8798ba5b7826bcdca2306d90843341251fc1bc72f8f63a07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 267, + 478, + 277 + ], + "lines": [ + { + "bbox": [ + 429, + 267, + 478, + 277 + ], + "spans": [ + { + "bbox": [ + 429, + 267, + 478, + 277 + ], + "type": "text", + "content": "(b) W-LD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 286, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 61, + 286, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 61, + 286, + 547, + 309 + ], + "type": "text", + "content": "Figure 4. Overview of DMW-LA and W-LD update strategy. (a) DMW-LA consists of intra-cluster interaction and global interaction. (b) The W-LD update strategy adopts weight and loss dual guidance to update the prototype memory bank." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 329, + 272, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 272, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 272, + 341 + ], + "type": "text", + "content": "3.2. Dynamic Memory Weight-Loss Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": "In this paper, we design Dynamic Memory Weight-Loss Attention (DMW-LA) to extract key category features from medical images, as shown in Figure 4(a). The intra-cluster interaction is responsible for obtaining internal dependencies of category features, while global interaction is responsible for initializing prototype memory bank and modeling global dependencies. Assuming input image " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "X_{1} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " refer to the height, width, and channel dimensions of the image, respectively. First, " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " is linearly transformed with weights " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "W^{q}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "W^{k}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "W^{v}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " to generate Query, Key, and Value for global interaction. At the beginning of network training, we perform K-means on the result " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " of global interaction to fill the memory bank. Here, we call the generated semantic clusters as similarity memory prior, and the central point of each cluster is called the similarity core prior. Therefore, the similarity memory prior can be represented as the " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\{\\{E_{11}, E_{12}, \\dots, E_{1c}\\}, \\dots, \\{E_{k1}, E_{k2}, \\dots, E_{kc}\\}\\}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", and the similarity core prior can be represented as the " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\{\\tilde{E}_1, \\tilde{E}_2, \\tilde{E}_3, \\dots, \\tilde{E}_k\\}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ". After the initialization of the memory bank is completed, we can supplement prior knowledge for network training. Here, we first calculate the cosine similarity between the Query and the similarity core prior to match the corresponding similarity memory prior. Assuming there is " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\{X^{i} \\in \\mathbb{R}^{N \\times C}, N \\leqslant H \\times W\\}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " in the Query that matches the similarity memory prior " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\{E_i = E_{i1}, E_{i2}, \\dots, E_{ic} \\mid E_i^T \\in \\mathbb{R}^{C \\times M}\\}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "E_i" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": " will perform self-attention with " + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 61, + 347, + 296, + 677 + ], + "type": "text", + "content": ", as shown in Equation (3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 691, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 691, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 128, + 691, + 295, + 706 + ], + "type": "interline_equation", + "content": "A _ {i} = S M \\left(X ^ {i} \\cdot E _ {i} ^ {T}\\right) \\cdot E _ {i} \\tag {3}", + "image_path": "aed0b84b7a3d1ae8b2b9ebfe0d3d988448a6d923da0ec7588f2f6a82f3b2519c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "i = \\{1,2,3,\\ldots\\}" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": ", indicating the number of segmented categories, and " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "SM" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": " represents the softmax activation function. This step simulates the process of the \"grandmother cell\" recognizing specific visual stimuli, so the " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": " plays both the role of Key and Value. Then, we concatenate " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": " to obtain semantic category features " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": ". Finally, we aggregate semantic category features and global features, and then map to the original feature space through the convolution operation " + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "inline_equation", + "content": "\\psi(\\cdot)" + }, + { + "bbox": [ + 313, + 329, + 547, + 434 + ], + "type": "text", + "content": ", as shown in Equation (4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 394, + 441, + 547, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 441, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 394, + 441, + 547, + 454 + ], + "type": "interline_equation", + "content": "X _ {2} = \\psi \\left(A + X ^ {\\prime}\\right) \\tag {4}", + "image_path": "f2775c3493cbe5f4fb62d1cf1cf408da38cec16519abcdd71aa756316a5f6925.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": "After the initialization of the memory bank is completed, it still needs to continuously update it to ensure that it can adapt to complex medical images. At present, the common method is to use a random update strategy, that is, randomly insert a portion of the input features into the corresponding semantic clusters in the memory. This method is not suitable for complex and ever-changing medical images and may introduce interference information. Therefore, we propose a Weight-Loss Dynamic (W-LD) update strategy to dynamically update the memory bank, as shown in Figure 4(b). Specifically, assume that " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X^{i} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " in the output " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X_{2} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " of the DMW-LA matches the similarity memory prior " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "E_{i} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": ". Then use the Softmax to calculate the weights of " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " separately, and sort the columns of " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " respectively according to the weight size. Note that the columns with large weight in " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " contain more semantic information, while the columns with small weight in " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " contain less semantic information. Therefore, we replace the " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " columns in " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " with the smallest weight with the " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " columns in " + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "inline_equation", + "content": "X^{i}" + }, + { + "bbox": [ + 313, + 460, + 547, + 705 + ], + "type": "text", + "content": " with the largest weight. In addition, we evaluate the quality of the current prototype mem" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23012" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 79, + 298, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 79, + 298, + 186 + ], + "lines": [ + { + "bbox": [ + 62, + 79, + 298, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 79, + 298, + 186 + ], + "type": "image", + "image_path": "e3dfc5627fc4e2104e7a7c0f6f748a6f6960e3c779a8d7b8e02ae78a3a40ad75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 194, + 296, + 227 + ], + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 227 + ], + "type": "text", + "content": "Figure 5. Overview of DS-GIM. DS-GIM uses cosine similarity and euclidean distance to comprehensively evaluate and highlight the differences between spatial element structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": "ory bank by calculating the difference between the current epoch loss " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "(Loss^{\\Delta})" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " and the previous epoch loss " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "(Loss^{\\nabla})" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": ". Specifically, when the value of " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "Loss^{\\nabla} - Loss^{\\Delta}" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " is greater than 0, it shows that the prototype memory bank has a positive impact on network training, leading to the reduction of the current epoch loss. Therefore, the " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " should be reduced to preserve these valuable features. Conversely, when the value of " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "Loss^{\\nabla} - Loss^{\\Delta}" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " is less than 0, the " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " should be increased to expand the range of updates and eliminate useless features as much as possible. In this way, the prototype memory bank can accumulate high-value semantic information to the maximum extent in each training epoch. In addition, a linear function is designed to link " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "inline_equation", + "content": "Loss^{\\nabla} - Loss^{\\Delta}" + }, + { + "bbox": [ + 61, + 247, + 296, + 411 + ], + "type": "text", + "content": ", as shown in Equation (5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 421, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 421, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 100, + 421, + 295, + 435 + ], + "type": "interline_equation", + "content": "K = (- \\alpha \\cdot (L o s s ^ {\\nabla} - L o s s ^ {\\Delta}) + \\beta) \\cdot \\theta \\tag {5}", + "image_path": "5191e7e9b033c9d4181f4c48125be90275babadfa081ab8f524dedca86335631.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": "In the above formula, " + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": " are decay factors, which are set to 0.5, and " + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\theta = M" + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": ". In addition, we set " + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\frac{M}{4} \\leq K \\leq \\frac{3M}{4}" + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": " to avoid " + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 61, + 439, + 296, + 521 + ], + "type": "text", + "content": " falling into extreme situations and ensure stable training, see appendix A. In this way, we can not only minimize the loss of information but also ensure the semantic clarity and representation ability of the similarity memory prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 528, + 296, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 528, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 528, + 296, + 552 + ], + "type": "text", + "content": "3.3. Double-Similarity Global Internal Enhancement Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "content": "The structure of the Double Similarity Global Internal Enhancement Module (DS-GIM) is shown in Figure 5. Assuming the input size is " + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "content": ", we first divide " + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\left\\{X_{1}, X_{2}, \\dots, X_{\\frac{H \\times W}{p^{2}}} \\mid X_{i} \\in \\mathbb{R}^{P \\times P \\times C}\\right\\}" + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "content": " with the windows of " + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 61, + 556, + 296, + 679 + ], + "type": "text", + "content": ", and calculate the internal cosine similarity of these windows. Weight different decay coefficients for each window based on similarity (from large to small). The purpose is to use the decay coefficient to increase the difference between windows. The formula for calculating the decay coefficient is as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 687, + 295, + 708 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 687, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 108, + 687, + 295, + 708 + ], + "type": "interline_equation", + "content": "\\gamma = \\exp \\left(- \\left(0. 2 5 - 2 ^ {- 2. 5 - \\frac {5 \\cdot n}{l}}\\right)\\right) \\tag {6}", + "image_path": "d6f4535bcb7b09a8e5edd501f4921aec4f851b66724c7512afda035596233be3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": "In the above formula, " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " represents the total number of layers in the current stage, " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "n = \\{0,1,2,\\dots ,l - 1\\}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ". Next, concatenate these windows to form a new feature representation " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ". Map " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " to a new feature space for feature calculation to obtain " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X_{I}\\in \\mathbb{R}^{HW\\times HW}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ". Then average the columns of " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X_{I}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X_{II}\\in \\mathbb{R}^{HW\\times 1}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ", and then copy " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X_{II}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " to get a new matrix " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "S\\in \\mathbb{R}^{HW\\times HW}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ". At this time, the elements in " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " can be regarded as a set of spatial location elements, and we can calculate the euclidean distance between elements using the formula " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "Euc = |S - S^T|" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " to measure their proximity in the feature space. Based on size, we can divide the internal elements of " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "Euc" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "Ecu_{I}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "Ecu_{II}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ". Next, these two parts are calculated by the Softmax function and areadamard products with " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X_{I}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": " respectively. Finally, splice these two parts together to form a new matrix and multiply it with " + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 313, + 81, + 547, + 267 + ], + "type": "text", + "content": ", as shown in Equation (7)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 271, + 547, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 271, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 271, + 547, + 285 + ], + "type": "interline_equation", + "content": "O _ {D} = X ^ {\\prime} \\cdot \\left(S M (E c u _ {I}) * X _ {I} + S M (E c u _ {I I}) * X _ {I}\\right) \\tag {7}", + "image_path": "c80829730cb73fc55ef092de0d96f053b11fbebbff459c88b00fcda8b1733a60.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "text", + "content": "In the above equation, " + }, + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "text", + "content": " represents hadamard product, and " + }, + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "inline_equation", + "content": "SM" + }, + { + "bbox": [ + 313, + 289, + 547, + 346 + ], + "type": "text", + "content": " represents softmax activation function. In this way, distance information between spatial elements will be embedded into feature encoding to enhance the model's understanding of the internal structure of input data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 356, + 392, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 392, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 392, + 369 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 375, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 375, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 375, + 385 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 392, + 547, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 547, + 647 + ], + "type": "text", + "content": "We conduct extensive experiments on four publicly available datasets: Automated Cardiac Diagnosis Challenge (ACDC, 4 classes) [4], Segmentation of Multiple Myeloma Plasma Cells in Microscopic Images (SegPC-2021, 3 classes) [13, 16, 17], International Skin Imaging Collaboration 2018 Challenge (ISIC-2018, 2 classes) [10], and MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge (Synapse, 9 classes). (1) ACDC dataset consists of Magnetic Resonance Imaging (MRI) of the left ventricle (LV), right ventricle (RV) and myocardium (Myo). The division of training set and testing set is consistent with TransUNet++ [42]. (2) SegPC-2021 dataset consists of myeloma plasma cell images. The segmentation target is the cytoplasm and nucleus of the cells. We divide the training set and the testing set by 4:1 ratio. (3) ISIC-2018 dataset consists of dermoscopy images. We also divide the training set and the testing set by 4:1 ratio. (4) Synapse dataset consists of axial abdominal clinical CT images. Each image contains part or all of the eight abdominal organs (aorta, gallbladder, left kidney, right kidney, liver, pancreas, spleen, stomach). We divide the training set and the testing set according to the TransUNet [8]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 652, + 547, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 547, + 676 + ], + "type": "text", + "content": "4.2. Implementation Details and Evaluation Metrics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 681, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 681, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 681, + 547, + 704 + ], + "type": "text", + "content": "In all experiments, the hyperparameters of DS-GIM (decay coefficient) and DMW-LA (linear function) remained un" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23013" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 79, + 547, + 338 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 547, + 338 + ], + "type": "table", + "html": "
TypeMethodsACDCSegPC-2021ISIC-2018
DSC↑HD95↓RVMyoLVDSC↑HD95↓CytoplasmNucleusDSC↑HD95↓
CNNU-Net [30]90.141.4889.6186.5894.2380.5435.8980.7780.3089.494.66
Att-UNet [25]90.481.2489.1587.2995.0180.4036.0479.5681.3789.585.11
U-Net++ [48]89.541.2587.9786.6294.0580.0733.1879.4080.1289.194.88
EMCAD [28]91.431.1390.6588.7894.8680.3834.8479.4981.2790.034.33
MADGNet [24]91.131.1790.0188.3095.0780.0233.2179.5580.4889.774.30
SelfReg-UNet [49]91.431.1988.9289.4995.8877.6035.6577.4477.7588.995.74
DconnNet [45]89.961.8089.1686.1294.5879.8433.0779.0080.6890.014.10
ViTSwinUNet [6]89.492.0388.0186.0394.4578.2235.2477.3379.1187.325.47
TransDeepLab [1]89.222.0787.4885.7494.4478.8133.0277.3880.2490.114.07
DAE-former [2]89.781.9189.9184.3895.0478.5234.8677.1079.9389.854.22
SelfReg-SwinUNet [49]91.491.9789.4989.2795.7076.4032.4574.7078.0989.164.90
CNN&ViTTransUNet [8]89.371.4889.2085.2393.6776.6933.9578.6980.6889.924.34
MixTrans [38]90.982.2889.5288.6094.8280.7232.6279.7881.6686.467.34
TransClawUNet [46]91.291.1091.3187.9394.6479.7734.4579.0480.5089.624.81
HiFormer [18]90.122.1591.0684.5494.7779.5432.0877.7381.3589.794.58
TransCASCADE [27]91.481.1291.3788.3294.7680.3632.8779.1881.5389.764.42
PAG-TransYnet [5]91.331.1590.8088.2894.9279.4133.8878.6780.1689.654.53
Memory MechanismSim-MPNet (Ours)92.181.0892.0689.1995.2983.1229.6682.0084.2490.853.55
", + "image_path": "70b516c0227c791bea67404ad793781148bcf533a7c755bbd3d40404018635bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 389, + 548, + 632 + ], + "blocks": [ + { + "bbox": [ + 61, + 346, + 548, + 380 + ], + "lines": [ + { + "bbox": [ + 61, + 346, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 346, + 548, + 380 + ], + "type": "text", + "content": "Table 1. Performance on ACDC, SegPC-2021 and ISIC-2018. We report the DSC and HD95 scores of different networks on the ACDC, SegPC-2021 and ISIC-2018. We adopt publicly available SOTA methods for implementation, and all results are obtained under the same dataset partitioning conditions. DSC are reported for individual organs or lesions. " + }, + { + "bbox": [ + 61, + 346, + 548, + 380 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 61, + 346, + 548, + 380 + ], + "type": "text", + "content": " denotes the higher (lower) the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 389, + 548, + 632 + ], + "lines": [ + { + "bbox": [ + 63, + 389, + 548, + 632 + ], + "spans": [ + { + "bbox": [ + 63, + 389, + 548, + 632 + ], + "type": "table", + "html": "
TypeMethodsP(M)F(G)DSC↑HD95↓AortaGBKLKRLiverPCSPSMTi
CNNU-Net [30]31.0450.1476.8539.7089.0769.7277.7768.6093.4353.9886.6775.58729
Att-UNet [25]34.8834.8877.7736.0289.5568.8877.9871.1193.5758.0487.3075.75761
U-Net++ [48]36.63107.0178.9830.6587.4565.8681.7574.2793.8564.1687.8276.651139
EMCAD [28]26.764.2983.6315.6888.1868.8788.0884.1095.2668.5192.1783.92752
MADGNet [24]32.167.7180.9827.4688.0165.1282.8179.0994.9267.2390.1580.501152
SelfReg-UNet [49]17.2630.7180.3426.9888.7471.7885.3280.7193.8062.2184.7875.39625
DconnNet [45]36.404.8081.5619.0187.4868.5683.7081.0494.6465.2490.6481.15849
ViTSwinUNet [6]27.175.9279.1321.5585.4766.5383.2879.6194.2956.5890.6676.60625
TransDeepLab [1]28.6117.0880.1621.2586.0469.1684.0879.8893.5361.1989.0078.40830
DAE-Former [2]48.0726.0782.4317.4688.9672.3086.0880.8894.9865.1291.9479.191187
SelfReg-SwinUNet [49]27.155.9280.5420.1086.0769.6585.1282.5894.1861.0987.4278.221014
CNN&ViTTransUNet [8]105.3224.6677.4831.6987.2363.1381.8777.0294.0855.8685.0875.62794
MixTrans [38]79.0744.7578.5926.5987.9264.9981.4777.2993.0659.4687.7576.81958
TransClawUNet [46]113.0238.0878.0926.3885.8761.3884.8379.3694.2857.6587.7473.551253
HiFormer [18]29.5211.8880.6919.1487.0368.6184.2378.3794.0760.7790.4482.03931
TransCASCIDE [27]123.4756.9382.6817.3486.6368.4887.6684.5694.4365.3390.7983.521472
PAG-TransYnet [5]144.2233.6483.4315.8289.6768.8986.7484.8895.8768.7592.0180.66940
Memory MechanismSim-MPNet (Ours)91.3043.7484.3414.8587.0673.4988.6585.3495.2869.2591.3684.30873
", + "image_path": "d8caade8590db4c9c095689a94be16d28d9b73f19b37261696cc77815ed02379.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "lines": [ + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "spans": [ + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "text", + "content": "Table 2. Performance on Synapse. P and F represent Parameters and FLOPs respectively. FLOPs of all the methods are reported for " + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "inline_equation", + "content": "{224} \\times {224}" + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "text", + "content": " inputs,except Sim-MPNet (dual encoder: " + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "inline_equation", + "content": "{224} \\times {224},{256} \\times {256}" + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "text", + "content": " ). " + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "inline_equation", + "content": "{T}_{i}" + }, + { + "bbox": [ + 61, + 639, + 548, + 684 + ], + "type": "text", + "content": " is the average inference efficiency of a single image in milliseconds. DSC are reported for individual organs. GB, KL, KR, PC, SP and SM represent gallbladder, left kidney, right kidney, pancreas, spleen and stomach respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 744 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 744 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 744 + ], + "type": "text", + "content": "23014" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 79, + 545, + 296 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 545, + 296 + ], + "type": "image", + "image_path": "75b6e15b9ee26c2ace8a24f9d3adda6328bb58cd2015f0a717e805d2d90d4aa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 304, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 61, + 304, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 548, + 327 + ], + "type": "text", + "content": "Figure 6. Multi-organ and cell segmentation. Comparison of visualization results between Sim-MPNet and other SOTA methods on Synapse and SegPC-2021. The different segmentation targets are represented by different colors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": "changed. For different datasets, just set " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{k}" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": "-means as the number of categories in the dataset. The input resolutions of two encoders are set to " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " respectively. In addition, we train the model using an AdamW optimizer with a learning rate of 0.0001, and a weight decay of 1e-4. The model is trained on a single RTX 3090 GPU with a batch size 20 using Dice " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "(L_{dice})" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " and Cross-entropy " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "(L_{ce})" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " losses for 300 epochs. The specific proportion of these two losses is " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "Loss = 0.7 \\cdot L_{dice} + 0.3 \\cdot L_{ce}" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": ". Finally, all experiments are performed in the Python 3.10.0 and Torch 2.3.0 environment. Only use random rotation and flipping as data augmentation. We choose the Dice score (DSC) and " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " Hausdorff distance (HD95) as evaluation metrics. The HD95 represents the " + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "inline_equation", + "content": "95^{th}" + }, + { + "bbox": [ + 61, + 346, + 296, + 521 + ], + "type": "text", + "content": " percentile distance between the boundaries of the model prediction graph and real labels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 534, + 237, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 237, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 237, + 547 + ], + "type": "text", + "content": "4.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "text", + "content": "ACDC. ACDC datasets test the model's ability to explore local details, and ViT-based methods are often difficult to achieve better segmentation results. Table 1 shows that the DSC of SwinUNet, TransDeepLab and DAE-Former does not exceed " + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "text", + "content": ". Although the hybrid CNN-ViT methods partially compensate for the shortcomings of ViT, its highest DSC is only " + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "inline_equation", + "content": "91.48\\%" + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "text", + "content": ". In contrast, Sim-MPNet can directly capture the category features of organs in ACDC by dynamic memory prior mechanism, thus achieving " + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "inline_equation", + "content": "92.18\\%" + }, + { + "bbox": [ + 61, + 553, + 295, + 669 + ], + "type": "text", + "content": " DSC and acquiring better segmentation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 670, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 670, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 670, + 296, + 704 + ], + "type": "text", + "content": "SegPC-2021. Table 1 shows that the Sim-MPNet outperforms existing methods and achieves a new SOTA on SegPC-2021. Because most of the myeloma plasma cells" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "text", + "content": "in SegPC-2021 are closely connected, this requires the network to have a strong ability to extract local details. Therefore, we can observe that the CNN-based methods and the hybrid CNN-ViT methods are better than the ViT-based methods in segmentation effect. However, compared with the optimal network MixTrans in the above methods, Sim-MPNet improved DSC by " + }, + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "inline_equation", + "content": "2.40\\%" + }, + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "text", + "content": " and HD95 by " + }, + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "inline_equation", + "content": "2.42\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 346, + 547, + 450 + ], + "type": "text", + "content": ". This proves that Sim-MPNet can effectively handle complex medical image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "text", + "content": "ISIC-2018. Table 1 shows that the Sim-MPNet also achieved SOTA results on the ISIC-2018 dataset. Since there are many large-scale lesions in the ISIC-2018, it verifies the global modeling ability of the network. Thus, the ViT-based TransDeepLab method has achieved the best segmentation results among the existing methods. However, compared with TransDeepLab method, Sim-MPNet improves DSC by " + }, + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "inline_equation", + "content": "0.74\\%" + }, + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "text", + "content": " and HD95 by " + }, + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "inline_equation", + "content": "0.52\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 451, + 548, + 567 + ], + "type": "text", + "content": ". This proves once again that Sim-MPNet has the advantage of capturing category dependencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "content": "Synapse. To further verify the performance of Sim-MPNet model in different complex scenarios, we introduce more diverse Synapse datasets for evaluation. Table 2 shows the comparison results on the Synapse. Firstly, CNN-based methods have significant advantages in small organ segmentation due to their excellent local modeling ability. For example, EMCAD achieved DSC of " + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "inline_equation", + "content": "88.08\\%" + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "inline_equation", + "content": "84.10\\%" + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "content": " on the left kidney (Kidney (L)) and right kidney (Kidney (R)), respectively. In contrast, Sim-MPNet performed better on these two organs, with DSC of " + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "inline_equation", + "content": "0.57\\%" + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "inline_equation", + "content": "1.24\\%" + }, + { + "bbox": [ + 313, + 567, + 548, + 696 + ], + "type": "text", + "content": " higher than EMCAD, respectively. However, on the Spleen" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23015" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 80, + 295, + 153 + ], + "blocks": [ + { + "bbox": [ + 63, + 80, + 295, + 153 + ], + "lines": [ + { + "bbox": [ + 63, + 80, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 295, + 153 + ], + "type": "table", + "html": "
Encoder 1Encoder 2ACDCSegPC-2021ISIC-2018Synapse
DMW-LA (Ours)—91.5482.6589.5783.46
—MaxViT91.0681.9289.2383.27
MaxViTMaxViT91.3382.0790.0483.59
DMW-LA (Ours)MaxViT92.1883.1290.8584.34
", + "image_path": "4cc662e60cdb7ccd2ed545b1fedfcb074d76f195504652185611730f216576ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "content": "with a complex contour, Sim-MPNet cannot fully calculate the correlation between local elements, and thus it cannot achieve the optimal segmentation result. Secondly, ViT-based methods have advantages in large-scale organ segmentation due to their global context exploration ability. DAE-Former achieved DSC of " + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "inline_equation", + "content": "72.30\\%" + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "inline_equation", + "content": "94.98\\%" + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "content": " on Gallbladder and Liver, respectively. Sim-MPNet has a DSC of " + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 61, + 251, + 295, + 540 + ], + "type": "text", + "content": " higher than DAE-Former in these two organs, respectively. Finally, the hybrid CNN-ViT methods combine the advantages of CNN and ViT, enabling the PAG-TransYnet network to explore both local and global features simultaneously, achieving optimal segmentation results in Aorta with the smallest area and Liver with the largest area. Although Sim-MPNet is slightly inferior to PAG-TransYnet in segmentation of the Aorta and Liver, its unique dynamic memory prior mechanism enables it to achieve the best average DSC and HD95 scores on the Synapse. This indicates that Sim-MPNet has excellent robustness in various scenarios. In addition, Sim-MPNet does not have significant advantages in terms of Params and FLOPs, but our method extracts medical image features from a new perspective and achieves excellent segmentation results. This shows the application potential of our method in the field of medical image segmentation and opens up a new path for future research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 542, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 542, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 542, + 295, + 705 + ], + "type": "text", + "content": "Visual Comparison and Analysis. Figure 6 shows the visualizations of Sim-MPNet and other networks on Synapse and SegPC-2021. Firstly, the segmentation results of Sim-MPNet are very close to the ground truth in Synapse. Especially when dealing with organs like the Pancreas, which have complex contours and rich details, Sim-MPNet shows better segmentation performance compared to other networks. Secondly, Sim-MPNet still obtains the best segmentation results for images containing multiple cells in SegPC-2021. Although false positives and false negatives may also occur, the incidence of these errors is significantly reduced compared to other networks. In addition, the semantic clustering receptive fields of DMW-LA on Synapse, ACDC and SegPC-2021 see appendix B." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 314, + 80, + 550, + 159 + ], + "blocks": [ + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "type": "text", + "content": "Table 3. Ablation study of DMW-LA on four datasets. We use different encoders to verify their effectiveness separately. The input resolution of each encoder is the same as that of the comparative experiment, which is " + }, + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "type": "text", + "content": " (Encoder 1) and " + }, + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 61, + 161, + 295, + 216 + ], + "type": "text", + "content": " (Encoder 2) respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 80, + 550, + 159 + ], + "lines": [ + { + "bbox": [ + 314, + 80, + 550, + 159 + ], + "spans": [ + { + "bbox": [ + 314, + 80, + 550, + 159 + ], + "type": "table", + "html": "
DatasetACDCSegPC-2021ISIC-2018Synapse
w/o W-LD91.5382.6390.2283.75
w/o DS-GIM91.6782.5890.4783.64
w/o W-LD&DS-GIM91.0282.2390.0783.39
Sim-MPNet92.1883.1290.8584.34
", + "image_path": "dd31a3f0d7a3a9ab3af7831cd05c43b8f9aba10463d83a33c278da596b6c7c8e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 167, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 547, + 210 + ], + "type": "text", + "content": "Table 4. Ablation study of DS-GIM and W-LD update strategy on four datasets. In w/o W-LD, we use a random update strategy to replace the W-LD update strategy. The figures in the table are DSC." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 231, + 405, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 405, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 405, + 243 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 247, + 547, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 547, + 341 + ], + "type": "text", + "content": "Effectiveness of DMW-LA. In this paper, Sim-MPNet adopts dual encoder structure, where Encoder 1 is composed of DMW-LA and Encoder 2 is composed of MaxViT. To evaluate the effectiveness of DMW-LA, we conducted experiments on each encoder. As shown in Table 3, the obvious performance difference indicates that DMW-LA can help the network obtain rich category features, thereby improving segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 341, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 547, + 434 + ], + "type": "text", + "content": "Effectiveness of DS-GIM and W-LD Update Strategy. We conducted ablation experiments on DS-GIM and W-LD update strategy. It can be seen from Table 4 that compared with the random update strategy, our W-LD update strategy can minimize redundant information and retain useful features. In addition, the remarkable performance improvement in the experiment further confirms the effectiveness of DS-GIM in capturing the internal differences of features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 444, + 384, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 384, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 384, + 456 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 547, + 626 + ], + "type": "text", + "content": "In this paper, we break the limitations of traditional methods and construct a novel medical image segmentation network, namely Sim-MPNet. Sim-MPNet directly extracts and locates the category features of lesions or organs. We develop two core modules: DMW-LA and DS-GIM. DMW-LA obtains category information through prototype memory bank and W-LD update strategy, while DS-GIM enhances the intrinsic differences between elements through a dual similarity mechanism. The comparative experiments show that Sim-MPNet outperforms the previous SOTA methods in segmenting multiple organs, cells, and lesions. We believe that Sim-MPNet brings new perspectives and inspirations to the design of medical image segmentation networks in future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 637, + 414, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 414, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 414, + 651 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 656, + 547, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 547, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 547, + 703 + ], + "type": "text", + "content": "This work was supported in part by the Tianshan Talent Training Program under Grant 2022TSYCLJ0036, in part by the National Natural Science Foundation of China under Grant 62472368, Grant 62302427 and Grant 62462060." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23016" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 99, + 295, + 704 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 99, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 295, + 163 + ], + "type": "text", + "content": "[1] Reza Azad, Moein Heidari, Moein Shariatnia, Ehsan Khodapanah Aghdam, Sanaz Karimijafarbigloo, Ehsan Adeli, and Dorit Merhof. Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation. In Int. Workshop PRedict. Intell. Med., pages 91-102. Springer, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 164, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 164, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 164, + 295, + 217 + ], + "type": "text", + "content": "[2] Reza Azad, René Arimond, Ehsan Khodapanah Aghdam, Amirhossein Kazerouni, and Dorit Merhof. Dae-former: Dual attention-guided efficient transformer for medical image segmentation. In In Int. Workshop PRedict. Intell. Med., pages 83-95. Springer, 2023. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 219, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 219, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 295, + 281 + ], + "type": "text", + "content": "[3] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 283, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 283, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 67, + 283, + 295, + 357 + ], + "type": "text", + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Trans. Med. Imaging, 37(11):2514-2525, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 358, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 358, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 358, + 295, + 412 + ], + "type": "text", + "content": "[5] Fares Bougourzi, Fadi Dornaika, Abdelmalik Taleb-Ahmed, and Vinh Truong Hoang. Rethinking attention gated with hybrid dual pyramid transformer-cnn for generalized segmentation in medical imaging. In International Conference on Pattern Recognition, pages 243-258. Springer, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "type": "text", + "content": "[6] Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, pages 205-218. Springer, 2022. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 456, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 456, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 67, + 456, + 295, + 498 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Proc. Adv. Neural Inf. Process. Syst., 33:9912-9924, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 500, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 500, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 67, + 500, + 295, + 552 + ], + "type": "text", + "content": "[8] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 554, + 295, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 295, + 596 + ], + "type": "text", + "content": "[9] Liang Chen, Paul Bentley, Kensaku Mori, Kazunari Misawa, Michitaka Fujiwara, and Daniel Rueckert. Drinet for medical image segmentation. IEEE transactions on medical imaging, 37(11):2453-2462, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 597, + 295, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 597, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 63, + 597, + 295, + 671 + ], + "type": "text", + "content": "[10] Noel Codella, Veronica Rotemberg, Philipp Tschandi, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 673, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 673, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 673, + 295, + 704 + ], + "type": "text", + "content": "[11] Shuang Cong and Yang Zhou. A review of convolutional neural network architectures and their optimizations. Artif. Intell. Rev., 56(3):1905-1969, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 315, + 82, + 547, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 82, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 547, + 145 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 147, + 547, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 147, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 315, + 147, + 547, + 188 + ], + "type": "text", + "content": "[13] Shiv Gehlot, Anubha Gupta, and Ritu Gupta. Ednfc-net: Convolutional neural network with nested feature concatenation for nuclei-instance segmentation. In ICASSP, pages 1389-1393. IEEE, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 190, + 547, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 190, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 315, + 190, + 547, + 210 + ], + "type": "text", + "content": "[14] Charles G Gross. Genealogy of the \"grandmother cell\". Neuroscientist., 8(5):512-518, 2002. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 212, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 212, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 547, + 262 + ], + "type": "text", + "content": "[15] Jieuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al. Recent advances in convolutional neural networks. Pattern Recognit., 77:354-377, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 547, + 308 + ], + "type": "text", + "content": "[16] Anubha Gupta, Pramit Mallick, Ojaswa Sharma, Ritu Gupta, and Rahul Duggal. Pcseg: Color model driven probabilistic multiphase level set based tool for plasma cell segmentation in multiple myeloma. PloS one, 13(12):e0207908, 2018. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 308, + 547, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 547, + 361 + ], + "type": "text", + "content": "[17] Anubha Gupta, Rahul Duggal, Shiv Gehlot, Ritu Gupta, Anvit Mangal, Lalit Kumar, Nisarg Thakkar, and Devprakash Satpathy. Gcti-sn: Geometry-inspired chemical and tissue invariant stain normalization of microscopic medical images. Med. Image Anal., 65:101788, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 361, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 547, + 415 + ], + "type": "text", + "content": "[18] Moein Heidari, Amirhossein Kazerouni, Milad Soltany, Reza Azad, Ehsan Khodapanah Aghdam, Julien Cohen-Adad, and Dorit Merhof. Hformer: Hierarchical multi-scale representations using transformers for medical image segmentation. In WACV, pages 6202–6212, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 415, + 547, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 547, + 458 + ], + "type": "text", + "content": "[19] Mohammad Hesam Hesamian, Wenjing Jia, Xiangjian He, and Paul Kennedy. Deep learning techniques for medical image segmentation: achievements and challenges. J. Digital Imaging, 32:582-596, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 459, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 459, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 459, + 547, + 510 + ], + "type": "text", + "content": "[20] Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu. Unet " + }, + { + "bbox": [ + 315, + 459, + 547, + 510 + ], + "type": "inline_equation", + "content": "3+" + }, + { + "bbox": [ + 315, + 459, + 547, + 510 + ], + "type": "text", + "content": ": A full-scale connected unet for medical image segmentation. In ICASSP, pages 1055–1059. IEEE, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 512, + 547, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 512, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 315, + 512, + 547, + 564 + ], + "type": "text", + "content": "[21] Ruobing Huang, Mingrong Lin, Haoran Dou, Zehui Lin, Qilong Ying, Xiaohong Jia, Wenwen Xu, Zihan Mei, Xin Yang, Yijie Dong, et al. Boundary-rendering network for breast lesion segmentation in ultrasound images. Med. Image Anal., 80:102478, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 566, + 547, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 566, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 315, + 566, + 547, + 608 + ], + "type": "text", + "content": "[22] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nat. Methods, 18(2):203-211, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 609, + 547, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 609, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 315, + 609, + 547, + 640 + ], + "type": "text", + "content": "[23] Tianqin Li, Zijie Li, Harold Rockwell, Amir Farimani, and Tai Sing Lee. Prototype memory and attention mechanisms for few shot image generation. In ICLR, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 641, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 641, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 641, + 547, + 704 + ], + "type": "text", + "content": "[24] Ju-Hyeon Nam, Nur Suriza Syazwany, Su Jung Kim, and Sang-Chul Lee. Modality-agnostic domain generalizable medical image segmentation by multi-frequency in multiscale attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11480-11491, 2024. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23017" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 296, + 135 + ], + "type": "text", + "content": "[25] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 137, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 137, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 295, + 167 + ], + "type": "text", + "content": "[26] Rodrigo Quian Quiroga, Itzhak Fried, and Christof Koch. Brain cells for grandmother. Sci. Am., 308(2):30-35, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 170, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 170, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 63, + 170, + 295, + 201 + ], + "type": "text", + "content": "[27] Md Mostafijur Rahman and Radu Marculescu. Medical image segmentation via cascaded attention decoding. In WACV, pages 6222-6231, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 203, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 203, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 63, + 203, + 295, + 245 + ], + "type": "text", + "content": "[28] Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. In WACV, pages 11769-11779, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 247, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 247, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 63, + 247, + 295, + 289 + ], + "type": "text", + "content": "[29] KKDSamesh, G Kiran Kumar, K Swapna, Debabrata Datta, and S Suman Rajest. A review of medical image segmentation algorithms. EAI Endorsed Transactions on Pervasive Health & Technology, 7(27), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 290, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 290, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 63, + 290, + 295, + 322 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234-241. Springer, 2015. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 323, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 323, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 63, + 323, + 295, + 365 + ], + "type": "text", + "content": "[31] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap. Meta-learning with memory-augmented neural networks. In ICML, pages 1842-1850. PMLR, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 367, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 367, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 63, + 367, + 295, + 398 + ], + "type": "text", + "content": "[32] Tracey J Shors. From stem cells to grandmother cells: how neurogenesis relates to learning and memory. Cell Stem Cell, 3(3):253-258, 2008. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 400, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 400, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 63, + 400, + 295, + 452 + ], + "type": "text", + "content": "[33] Shiming Tang, Tai Sing Lee, Ming Li, Yimeng Zhang, Yue Xu, Fang Liu, Benjamin Teo, and Hongfei Jiang. Complex pattern selectivity in macaque primary visual cortex revealed by large-scale two-photon imaging. Curr. Biol., 28(1):38-48, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 454, + 295, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 454, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 63, + 454, + 295, + 506 + ], + "type": "text", + "content": "[34] Shiming Tang, Yimeng Zhang, Zhihao Li, Ming Li, Fang Liu, Hongfei Jiang, and Tai Sing Lee. Large-scale two-photon imaging revealed super-sparse population codes in the v1 superficial layer of awake monkeys. ELIFE, 7:e33370, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 509, + 295, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 509, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 63, + 509, + 295, + 551 + ], + "type": "text", + "content": "[35] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, pages 459-479. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 552, + 295, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 552, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 63, + 552, + 295, + 594 + ], + "type": "text", + "content": "[36] Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Hacihaliloglu, and Vishal M Patel. Medical transformer: Gated axial-attention for medical image segmentation. In MICCAI, pages 36-46. Springer, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 596, + 295, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 596, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 63, + 596, + 295, + 638 + ], + "type": "text", + "content": "[37] Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. In AAAI, pages 2441-2449, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 639, + 295, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 639, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 63, + 639, + 295, + 681 + ], + "type": "text", + "content": "[38] Hongyi Wang, Shiao Xie, Lanfen Lin, Yutaro Iwamoto, Xian-Hua Han, Yen-Wei Chen, and Ruofeng Tong. Mixed transformer u-net for medical image segmentation. In ICASSP, pages 2390-2394. IEEE, 2022. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "type": "text", + "content": "[39] Zhehao Wang, Xian Lin, Nannan Wu, Li Yu, Kwang-Ting Cheng, and Zengqiang Yan. Dtmformer: Dynamic token" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 602 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 103 + ], + "type": "text", + "content": "merging for boosting transformer-based medical image segmentation. In AAAI, pages 5814-5822, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 105, + 547, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 105, + 547, + 136 + ], + "spans": [ + { + "bbox": [ + 315, + 105, + 547, + 136 + ], + "type": "text", + "content": "[40] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, pages 3–19, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 138, + 547, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 138, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 138, + 547, + 178 + ], + "type": "text", + "content": "[41] Hanguang Xiao, Li Li, Qiyuan Liu, Xiuhong Zhu, and Qihang Zhang. Transformers in medical image segmentation: A review. Biomed. Signal Process. Control, 84:104791, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 181, + 547, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 181, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 315, + 181, + 547, + 223 + ], + "type": "text", + "content": "[42] Lianghui Xu, Liejun Wang, Yongming Li, and Anyu Du. Big model and small model: Remote modeling and local information extraction module for medical image segmentation. Appl. Soft Comput., 136:110128, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 224, + 547, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 224, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 547, + 266 + ], + "type": "text", + "content": "[43] Cheng Xue, Lei Zhu, Huazhu Fu, Xiaowei Hu, Xiaomeng Li, Hai Zhang, and Pheng-Ann Heng. Global guidance network for breast lesion segmentation in ultrasound images. Med. Image Anal., 70:101989, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 269, + 547, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 547, + 298 + ], + "type": "text", + "content": "[44] Tianyu Yang and Antoni B Chan. Learning dynamic memory networks for object tracking. In ECCV, pages 152-167, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 301, + 547, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 301, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 315, + 301, + 547, + 343 + ], + "type": "text", + "content": "[45] Ziyun Yang and Sina Farsiu. Directional connectivity-based segmentation of medical images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11525-11535, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 346, + 547, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 346, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 315, + 346, + 547, + 387 + ], + "type": "text", + "content": "[46] Chang Yao, Menghan Hu, Qingli Li, Guangtao Zhai, and Xiao-Ping Zhang. Transclaw u-net: claw u-net with transformers for medical image segmentation. In ICASSP, pages 280-284. IEEE, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 388, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 388, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 315, + 388, + 547, + 420 + ], + "type": "text", + "content": "[47] Yundong Zhang, Huiye Liu, and Qiang Hu. Transfuse: Fusing transformers and cnns for medical image segmentation. In MICCAI, pages 14-24. Springer, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 422, + 547, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 547, + 474 + ], + "type": "text", + "content": "[48] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Proc. Deep Learn. Med. Image Anal. Multimodal Learn. Clin. Decis Support., pages 3-11. Springer, 2018. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 475, + 547, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 547, + 539 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 547, + 539 + ], + "type": "text", + "content": "[49] Wenhui Zhu, Xiwen Chen, Peijie Qiu, Mohammad Farazi, Aristeidis Sotiras, Abolfazl Razi, and Yalin Wang. Selffregunet: Self-regularized unet for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 601–611. Springer, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 540, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 540, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 540, + 547, + 602 + ], + "type": "text", + "content": "[50] Yanjie Zhu, Jing Cheng, Zhuo-Xu Cui, Qingyong Zhu, Leslie Ying, and Dong Liang. Physics-driven deep learning methods for fast quantitative magnetic resonance imaging: Performance improvements through integration with deep neural networks. IEEE Signal Process Mag., 40(2):116-128, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 743 + ], + "type": "text", + "content": "23018" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_content_list.json b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdeaf4c9dc865d984b9fdd94df2940d7de8950c2 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_content_list.json @@ -0,0 +1,1321 @@ +[ + { + "type": "text", + "text": "SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 178, + 128, + 818, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianfu Cheng $^{1,2*}$ , Wei Zhang $^{1*}$ , Shiwei Zhang $^{2*}$ , Jian Yang $^{1*†}$ , Xiangyuan Guan $^{1}$ , Xianjie Wu $^{1}$ , Xiang Li $^{1†}$ , Ge Zhang $^{3}$ , Jiaheng Liu $^{3}$ , Yuying Mai $^{3}$ , Yutao Zeng $^{1}$ , Zhoufutu Wen $^{3}$ , Ke Jin $^{1,2}$ , Baorui Wang $^{1}$ , Weixiao Zhou $^{1}$ , Yunhong Lu $^{3}$ , Hangyuan Ji $^{1}$ , Tongliang Li $^{1}$ , Wenhao Huang $^{3}$ , Zhoujun Li $^{1,4}$ , $^{1}$ Beihang University; $^{2}$ Baidu Inc., China; $^{3}$ M-A-P; $^{4}$ Shenzhen Intelligent Strong Technology Co., Ltd. {buaacxf, lizj} @ buaa.edu.cn, zhangshiwei05@baidu.com", + "bbox": [ + 125, + 202, + 870, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 377, + 326, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The increasing application of multi-modal large language models (MLLMs) across various sectors has spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by 7 key features: it is based on bilingual, it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 scenario domains. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases.", + "bbox": [ + 89, + 409, + 483, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 789, + 220, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A significant challenge in large language models (LLMs) is ensuring that LLMs [1, 25] generate factually accurate and evidence-based responses. Current state-of-the-art (SOTA)", + "bbox": [ + 89, + 815, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f61626ebd2b36af0ed3c9cb4cea56ac0ba065e59bd05249dec00a580c3babc3.jpg", + "image_caption": [ + "Figure 1. An example from our proposed SimpleVQA." + ], + "image_footnote": [], + "bbox": [ + 547, + 361, + 874, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LLMs often produce outputs that are misleading or unsupported by evidence phenomenon known as \"hallucinations\" [3, 29, 43]. The generation of incorrect or unsubstantiated information remains a major barrier to the broader adoption and reliability of general-purpose AI technologies.", + "bbox": [ + 511, + 642, + 906, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OpenAI proposes SimpleQA [30] to measure factuality simple and reliable with nearly 4K concise and fact-seeking questions. Further, Chinese SimpleQA [8] comprised of 3K Chinese questions spanning 6 major topics is proposed to target the Chinese language. However, the SimpleQA and Chinese SimpleQA benchmarks mainly evaluate the model capabilities in text modality, ignoring wider real-world scenarios (e.g. vision modality). For the vision modality, the research progress of the multi-modal large language models (MLLMs) is still hindered by the \"hallucinations\" introduced by the given images. Therefore, The community of MLLMs has an urgent need for how to measure the simple", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contributions.", + "bbox": [ + 114, + 875, + 241, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Authors.", + "bbox": [ + 114, + 887, + 256, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4637", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and reliable factuality introduced by the image.", + "bbox": [ + 89, + 90, + 403, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this limitation, we develop the SimpleVQA benchmark as shown in Figure 1, where we define the factual question-answering capability of the visual language model. For the proposed factual Visual Question Answering (VQA), we collect 2,025 high-quality question-answer pairs covering 9 different scenario domains across 9 different application tasks. As a factual benchmark for a short answer, SimpleVQA has the following advantages: (1) English and Chinese: SimpleVQA provides general knowledge visual Q&A in both English and Chinese backgrounds, and comprehensively assesses the fact-generating capacity of MLLMs in Chinese and English communities. (2) Multi-task division: We divide the SimpleVQA assessment set into 9 different forms of VQA tasks according to the collected questions and different needs of pictures. (3) Diversified scenarios: SimpleVQA divided the VQA tasks into 9 uniquely representative scenario domains based on the category of the image and the source of the collected images. (4) High quality: We implement a comprehensive and rigorous quality control process to ensure the quality of questions and the accuracy of answers at SimpleVQA. (5) Challenge: simpleVQA focuses on factual questions that mainstream MLLMs cannot answer accurately, and cannot trace the cause of errors through the model itself. (6) Static answers: Following SimpleQA's factual definition, all the standard answers provided in our benchmark don't change over time. (7) Easy to evaluate: SimpleVQA's short answers make it possible to use existing LLMs to run a judge program (called LLM-as-a-Judge [15, 34]) to quickly determine right or wrong and get an overall accuracy rate.", + "bbox": [ + 89, + 109, + 485, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We systematically evaluate 18 MLLMs on SimpleVQA and create a dynamic leaderboard to show results. Further, a series of probing experiments are performed to explore the effect of the key factors for SimpleVQA. We classify the capabilities possessed by MLLMs for factual questions into two aspects, visual understanding and internalized knowledge capabilities: (1) visual understanding refers to the ability of the model to identify the subject of the question being asked in the question; and (2) internalized knowledge capabilities test whether the model has already mastered the relevant knowledge of the subject of the question being asked, and thus is able to answer the relevant question correctly after identifying that subject. Based on this definition, we added an abductive reasoning experiment to the basic assessment to help determine whether the badcase came from a lack of visual understanding ability or a lack of internalized knowledge ability by generating and labeling atomic questions (each atomic question corresponds to an atomic fact) for each VQA example.", + "bbox": [ + 89, + 565, + 483, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remarkable findings from SimpleVQA are summarized as: (1) The factual accuracy of most evaluation models in the field of visual question-answering is insufficient.", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) The training data of MLLMs contains knowledge errors and they are overconfident in what they generate. (3) Image content understanding is still a major challenge for MLLMs to achieve improved capabilities. (4) The ability of MLLMs to internalize massive world knowledge still needs to be improved, and overcoming illusions remains a great challenge for large language models. (5) Further aligning visual core information with the knowledge stored in LLMs is also key to improving the factuality of MLLMs.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. SimpleVQA", + "text_level": 1, + "bbox": [ + 511, + 241, + 640, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 266, + 624, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The SimpleVQA benchmark consists of 2,025 samples spanning 9 core tasks and 9 primary domains, with each question-image pair categorized into relevant subcategories, enabling a comprehensive evaluation of MLLMs across diverse knowledge areas. The dataset 9 tasks, including covers Logic & Science (LS), Object Identification Recognition (OIR), Time & Event (TE), Person & Emotion (PE), Location & Building (LB), Text Processing (TP), Quantity & Position Relationship (QPR), Art & Culture (AC), and Object Attributes Recognition (OAR). To ensure broad topic coverage, SimpleVQA is structured around 9 key domains: Literature, education & sports (LES), Euro-American History & Culture (EHC), Contemporary Society (CS), Engineering, Technology & Application (ETA), Film, Television & Media (FTM), Natural Science (NS), Art (AR), Chinese History & Culture (CHC), and Life (LI).", + "bbox": [ + 511, + 287, + 906, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Table 1, SimpleVQA differs from existing MLLM benchmarks by focusing on factual knowledge boundaries instead of general vision-language understanding. Politically sensitive and ideological content is excluded to maintain neutrality and avoid controversy. Designed for efficiency, the dataset features concise questions and standardized answers, reducing complexity in model evaluation. All samples follow a short-answer Q&A format, enabling simple and objective assessment through direct answer matching. These refinements ensure SimpleVQA serves as a robust benchmark for evaluating MLLMs' factual reasoning abilities.", + "bbox": [ + 511, + 531, + 908, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Dataset Criteria", + "text_level": 1, + "bbox": [ + 511, + 722, + 673, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SimpleVQA maintains the following rigorous standards and guidelines for question, image, and answer selection, ensuring objectivity, temporal stability, and verifiability.", + "bbox": [ + 511, + 744, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Question Guidelines Clear and Unique Answers: Questions must have a single, undisputed answer. They should precisely define scope (e.g., \"Which city?\" instead of \"Which location?\") and specify time references (e.g., \"Which year?\" rather than \"When?\"). Evidence-Based: Each question must be supported by verifiable", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4638", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/fd9421bb8aecc531862e5809f037e64e73eac3fdc98dfebd6cce61ba4ff8640f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkMultimodalData SizeLanguageData SourceDomainFactualityReasoningMetric
MMbench [23]Image&Text2,438Chi.&Eng.Real WorldKnowledge××MCQ Eval
CCBench [23]Image&Text510ChineseKnowledgeKnowledge××MCQ Eval
MME [13]Image&Text1300EnglishReal WorldGeneral××TFQ Eval
MM-Vet [39]Image&Text200EnglishHumanGeneral××LLM-as-a-Judge
Dynamath [45]Image&Text5000EnglishExamsMath××Accuracy
MMMU [40]Image&Text11.5kEnglishExamsGeneral×✓Accuracy
MMMU-Pro [41]Image&Text3460EnglishHuman&GPTGeneral×✓Accuracy
ChineseFactEval [35]Text Only125ChineseHumanKnowledge✓×LLM-as-a-Judge
AGI-Eval [44]Text Only8062Chi.&Eng.ExamsKnowledge××Accuracy
C-Eval [11]Text Only13,948ChineseExamsKnowledge××Accuracy
SimpleQA [31]Text Only4,326EnglishHumanKnowledge✓×LLM-as-a-Judge
Chinese SimpleQA [9]Text Only3,000ChineseHuman&GPTKnowledge✓×LLM-as-a-Judge
SimpleVQA (Ours)Image&Text2,025Chi.&Eng.Human&GPTKnowledge✓✓LLM-as-a-Judge
", + "bbox": [ + 116, + 88, + 885, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons between our SimpleVQA and other benchmarks, where \"TFQ\" means True or False questions, \"MCQ\" means multi-choice questions, \"Chi.& Eng.\" means Chinese and English.", + "bbox": [ + 89, + 287, + 906, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4921ea471c671e02e65159677490f3b4bfd18e4521c7feb450659072bb6262a.jpg", + "image_caption": [ + "Figure 2. An overview of the data construction process of SimpleVQA." + ], + "image_footnote": [], + "bbox": [ + 114, + 329, + 885, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sources. Manually annotated questions include reference links, while automatically generated ones undergo independent validation by two AI trainers. *Challenging for MLLMs*: Questions are tested on GPT-4o, GPT-4o-mini, doubao-vision-pro, and ERNIE-VL. Only those that at least one model answers incorrectly are retained; others are revised. Answerable by August 2024: All questions must be answerable based on knowledge available before September 1, 2024, ensuring a fair evaluation across models with similar knowledge cutoffs.", + "bbox": [ + 88, + 563, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Guidelines No Direct Textual Clues: Images must not contain text revealing the answer. Authenticity: Only real, unaltered images are allowed to prevent factual distortion. Supports Question Reasoning: Each image must provide sufficient context for answering. Manually labeled samples undergo multi-annotator verification. Fixed Before August 2024: Image content must be valid and confirmable before August 2024.", + "bbox": [ + 88, + 731, + 482, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Answer Guidelines Temporal Stability: Answers must remain unchanged and unaffected by new information.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Time-sensitive topics (e.g., sports, media) should specify a timeframe rather than a general answer that may change. Sufficiently Challenging: Answers are tested against four high-precision MLLMs. If all models respond correctly, the question is revised to increase difficulty. Fully Objective and Evaluable: Answers must be precise, verifiable, and free from subjective interpretation. Unambiguous: Each answer must have a single, clear meaning to prevent misinterpretation.", + "bbox": [ + 511, + 563, + 906, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Data Collection and Processing", + "text_level": 1, + "bbox": [ + 511, + 710, + 790, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows an overview of the data construction process of SimpleVQA, which adheres to a meticulously designed five-step process, ensuring the highest standards of data quality and reliability.", + "bbox": [ + 511, + 733, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Seed Example Collection. SimpleVQA's seed examples are sourced from two primary channels. First, we filter images and Q&A pairs from publicly available VQA datasets that align with factual knowledge criteria. We select MMVet (English), MME (English), Dynamath (English), MMbench_CN (Chinese), and CCBench (Chinese) due to their recent construction (post-2024) and their rele", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4639", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/26e4c00b037160baf5e3e726920f54967a517ffee0bd512f20c6f99ed3259ba7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StatisticsNumberStatisticsNumber
Data2025Domain Categories9
- Chinsne(CN)1012- Literature, Education & Sports(LES)13.48%
- English(EN)1013- Euro-American History & Culture(EHC)12.89%
Task Categories9- Contemporary Society(CS)11.51%
- Logic & Science(LS)5.04%- Engineering, Technology & Application(ETA)7.95%
- Object Identification Recognition(OIR)14.07%- Film, Television & Media(FTM)10.62%
- Time & Event(TE)9.98%- Natural Science (NS)12.64%
- Person & Emotion(PE)13.58%- Art (AR)7.65%
- Location & Building(LB)21.53%- Chinese History & Culture(CHC)9.68%
- Text Processing(TP)10.07%- Life (LI)13.58%
- Quantity & Position Relationship(QPR)10.12%Word length of Questions
- Art & Culture(AC)9.09%- Maximum question length314
- Object Attributes Recognition(OAR)6.52%- Minimum problem length5
", + "bbox": [ + 197, + 88, + 802, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Statistics of SimpleVQA", + "bbox": [ + 395, + 270, + 599, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vance to real-world applications. Second, we collect images and relevant factual knowledge from search engines (e.g., Google, Baidu, Wikipedia), with expert annotators generating corresponding questions and answers. These data focus on entities and events across multiple domains, ensuring answers are objective, fact-based, and centered on entity recognition or attribute extraction.", + "bbox": [ + 89, + 300, + 480, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Data Enhancement and QA Pair Generation. Once sufficient seed examples are gathered, we employ GPT-4o [12] to refine the data and generate Q&A pairs for factual categories. For multiple-choice questions (MCQs) from sources like MMbench-CN and CCBench, to ensure answer uniqueness, we use LLMs to rephrase the original question and introduce qualifiers that precisely align with the correct response. For MME, we extract the answer entity and rewrite the question based on its attributes, ensuring a one-to-one correspondence. Datasets like MMVet, Dynamath, and CCBench, which contain discrepancies from factual Q&A formats (e.g., incorrect answer options, image descriptions, or MCQ distractors), are processed using GPT-4o to align the content with factual reasoning. These refinements produce the initial version of SimpleVQA.", + "bbox": [ + 89, + 409, + 482, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: LLM-Based Quality Verification. The refined dataset undergoes verification using GPT-4 to assess adherence to quality standards, ensuring answer stability, uniqueness, and question difficulty. Following LLM screening, two professional annotators conduct rigorous quality checks and refine samples as needed.", + "bbox": [ + 89, + 638, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 4: Difficulty Screening. To maximize the dataset's utility in model evaluation, we filter out overly simple Q&A pairs. We assess responses from four mainstream MLLMs (GPT-4o, GPT-4o-mini, Doubao-vision-pro, and ERNIE-VL). Any question correctly answered by all four models is deemed too simple and excluded from the dataset, thereby maintaining a challenging benchmark.", + "bbox": [ + 89, + 731, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 5: Extracting Atomic Facts. To analyze visual comprehension and language alignment in MLLMs more precisely, we generate atomic questions from each SimpleVQA entry. An atomic fact represents the most funda", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mental, indivisible attribute or characteristic of an object. For instance, given the question \"In what year was the person in the image born?\", the corresponding atomic question is \"Who is the person in the image?\". MLLMs generate candidate answers, which are then reviewed and refined by professional annotators to ensure accuracy.", + "bbox": [ + 511, + 300, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Human Annotation & Quality Control", + "text_level": 1, + "bbox": [ + 511, + 402, + 844, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure dataset quality, we implement a rigorous manual validation process following automated data collection. Most of the collaborators participated in the necessary data annotation, and we also selected three domain experts from the collaborators. Each question is independently reviewed by two expert annotators to verify factual accuracy. If either annotator finds a question unsuitable, it is discarded. Annotations fact-check answers using authoritative sources such as Wikipedia and Baidu Encyclopedia, providing at least two supporting URLs. If their answers differ, a third expert conducts a final review to ensure consistency and correctness. Only Q&A pairs that fully align with both human evaluations and LLM-generated responses are retained.", + "bbox": [ + 511, + 426, + 906, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A difficulty assessment further refines the dataset. We begin with 8,360 Q&A pairs, filtering out $22\\%$ of image-based samples that lack challenge or fail to meet predefined criteria. 1,108 pairs are removed through multi-model testing to ensure that questions pose a meaningful challenge to MLLMs. To maintain category balance, we carefully select 200 high-difficulty mathematical Q&As from 5,000 Dynamath samples, avoiding an overrepresentation of simpler factual questions. Through multiple validation rounds, we retain 2,025 high-precision Q&A pairs, accounting for $24\\%$ of the original dataset. This process ensures factual integrity, topic diversity, and appropriate difficulty levels, making SimpleVQA a robust benchmark for evaluating MLLMs' reasoning and knowledge boundaries.", + "bbox": [ + 511, + 623, + 908, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.5. Dataset Statistics", + "text_level": 1, + "bbox": [ + 511, + 847, + 681, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Table 1 and 2, our SimpleVQA benchmark consists of 2,025 samples across 9 major tasks, 9 major do", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4640", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mains, and 244 image types. Examples of each category can be found in supplementary material. This design facilitates a comprehensive assessment of MLLMs across different domains. Regarding the distribution of domain topics and image types in SimpleVQA, 9 main domains are defined and subcategories are assigned based on each topic. Table 2 displays statistics related to SimpleVQA, including the amount of data, categorization and percentage, and the text length of the questions. In Table 1, we also compare SimpleVQA with several mainstream MLLMs' evaluation benchmarks, which suggests that SimpleVQA is the first MLLMs' benchmark that focuses on the evaluation of knowledge boundaries in factual categories. We excluded ideological and politically relevant data from the dataset to prevent social controversies and negative impacts. In addition, we implemented several optimizations to improve the efficiency of the evaluation. The dataset features concise questions and standardized answers, minimizing the input and output markers required for GPT assessment. In addition, all examples are in short-answer question-and-answer (QA) format, and they can be assessed by simple matching.", + "bbox": [ + 89, + 90, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 89, + 431, + 222, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 460, + 279, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement a standardized protocol across all experiments, maintaining a consistent prompt format throughout. All temperature and sampling parameters are configured according to each LLM's official recommendations or default settings. For evaluation and benchmark construction, we employ GPT-4o as the primary model due to its robust performance in multimodal tasks.", + "bbox": [ + 89, + 484, + 483, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Baseline Models", + "text_level": 1, + "bbox": [ + 89, + 609, + 251, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate 18 representative MLLMs in total, comprising 8 closed-source and 10 open-source models, which provide a diverse evaluation of model capabilities across different architectures and training paradigms. The closed-source models include GPT-4o and 4o-mini, Doubao-pro-128k and 32k, Gemini-2.0-flash, Claude-3.5-Sonnet, Qwen-Max, ERNIE-VL. The open-source models cover a wide range of frameworks, including InternVL2.5 and In", + "bbox": [ + 89, + 636, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ternVL2, Qwen2.5-VL and Qwen2-VL, Janus-pro-7B.", + "bbox": [ + 511, + 90, + 877, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 116, + 697, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SimpleVQA benchmark employs a set of evaluation metrics to assess the accuracy, reliability, and consistency of the model's predictions. These metrics include: (1) Correct (CO) evaluates whether the predicted answer matches the reference answer exactly without any contradiction. (2) Not Attempted (NA) identifies cases where the model does not attempt to answer, ensuring no contradictions are present. (3) Incorrect (IN) flags instances where the predicted answer contradicts the reference answer, even if resolved. (4) Correct Given Attempted (CGA) measures the proportion of correct answers among those attempted by the model, reflecting its performance when engaged. (5) F-score computes the harmonic mean between \"Correct\" and \"Correct Given Attempted\", providing a balanced evaluation that combines accuracy and attempt success.", + "bbox": [ + 511, + 138, + 906, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Main Results", + "text_level": 1, + "bbox": [ + 511, + 376, + 651, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on Overall and Different Tasks Table 3 presents the performance of MLLMs on SimpleVQA, highlighting their F-scores across different tasks. Among the closed-source models, Gemini-2.0-Flash demonstrates superior performance, particularly in tasks such as LS, OIR, TE, PE, LB, TP, and QPR, achieving state-of-the-art (SOTA) scores across these task categories. Claude-3.5-Sonnet, GPT-4o, and ERNIE also exhibit strong performance. Open-source models, including InternVL2.5-78BMPO and Qwen2.5-VL-72B-Instruct, show competitive results, though slightly below the leading closed-source models. Notably, Qwen2.5-VL-72B-Instruct outperforms most mainstream closed-source MLLMs in several aspects.", + "bbox": [ + 511, + 398, + 908, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on Cross-lingual and Different Domains Table 4 lists the performance of various closed-source and open-source MLLMs on SimpleVQA, highlighting their Q&A performance across languages and across different scenarios. Among the closed-source models, Gemini-2.0-Flash and ERNIE-VL show strong Chinese quizzing ability; Gemini-2.0-Flash and Claude-3.5-Sonnet show strong English quizzing ability; especially on the scale evaluation with scenarios as the categorization, the three models mentioned above achieve almost all the SOTA. opensource models InternVL2.5-78B-MPO and Qwen2.5-VL", + "bbox": [ + 511, + 614, + 908, + 782 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://platform.openai.com/docs/models#gpt-4o; https://platform.openai.com/docs/models#gpt-4o-mini \nhttps://www.volcengine.com/product/ark \nhttps://deepmind.google/technologies/gemini/flash/ \nhttps://www.anthropic.com/news/claude-3-5-sonnet \nhttps://huggingface.co/spaces/Qwen/Qwen-Max \nhttps://yiyan.baidu.com/", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/OpenGVLab/InternVL2_5-78B-MPO; https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B \nhttps://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct; https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct; https://huggingface.co/Qwen/Qwen2-VL-72B-Instruct \nhttps://huggingface.co/deepseek-ai/Janus-Pro-7B", + "bbox": [ + 513, + 791, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4641", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fc4c03ba13ec9a9aafa871809bbf52083a05527e351d2240aa570539f4599adb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsOverall resultsF-score on 9 task categories
CONAINCGAF-scoreLSOIRTEPELBTPQPRACOAR
Closed-source Multi-modal Large Language Models
GPT-4o47.27.845.051.249.158.753.448.136.444.357.561.835.161.0
GPT-4o-mini35.510.054.539.437.330.044.438.529.636.242.145.925.840.6
Doubao-vision-pro-128k39.720.839.550.144.348.353.553.439.635.038.546.935.962.9
Doubao-vision-pro-32k25.423.651.432.728.324.237.918.031.621.339.232.918.636.6
Gemini-2.0-flash52.86.041.256.154.463.760.954.355.044.361.565.733.964.4
Claude-3.5-Sonnet48.59.841.853.750.957.153.952.129.747.258.462.746.760.1
Qwen-Max25.415.359.330.027.515.133.713.530.127.736.034.412.136.4
ERNIE-VL46.59.544.051.448.849.055.948.340.740.754.459.733.670.8
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO45.45.948.648.346.857.454.650.326.434.457.065.832.872.3
InternVL2.5-78B41.57.750.845.043.249.549.445.028.431.049.663.732.065.2
InternVL2-Llama3-76B35.78.455.938.937.234.743.535.626.228.444.353.529.054.1
InternVL2.5-38B-MPO42.95.451.845.344.051.752.545.722.134.052.067.832.860.4
InternVL2.5-26B-MPO39.97.652.543.241.543.444.547.727.131.745.958.629.068.0
InternVL2.5-8B-MPO33.67.558.936.334.945.137.430.919.126.044.352.226.456.2
Qwen2.5-VL-72B49.45.445.252.250.857.750.851.038.651.657.865.829.862.8
Qwen2-VL-72B-Instruct44.710.345.049.847.148.851.946.337.938.555.363.035.959.7
Qwen2.5-VL-7B-Instruct43.25.151.745.644.338.652.737.632.741.052.653.639.256.7
Janus-pro-7B31.310.658.135.033.027.043.426.523.728.224.050.436.242.3
", + "bbox": [ + 94, + 88, + 903, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. We report the scores across different tasks, including \"Logic & Science (LS)\", \"Object Identification Recognition (OIR)\", \"Time & Event (TE)\", \"Person & Emotion (PE)\", \"Location & Building (LB)\", \"Text Processing (TP)\", \"Quantity & Position Relationship (QPR)\", \"Art & Culture (AC)\", and \"Object Attributes Recognition (OAR)\".", + "bbox": [ + 89, + 440, + 906, + 497 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbe3542cfd831a11ea3b5c93cd42dbd2f0a72191801e204326f5fc95813e44eb.jpg", + "image_caption": [ + "Figure 3. F-score results for eight different models across nine task categories." + ], + "image_footnote": [], + "bbox": [ + 94, + 510, + 259, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08b983dde6cf1fe599a6e7cf36c0e18842c62f5acdeba12412b677e5284c2578.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 308, + 510, + 472, + 720 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/41d280c1a3c62f683c79d835c9eb9c5b20ccdcbc506524a75c71c09fe6512203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 510, + 687, + 719 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6bd0d01f88c97d128d233b2ff65e32ee10a39f0f89a9c27841b4655a0646f20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 510, + 901, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "72B-Instruct, respectively, achieved the highest performance, albeit slightly lower than the top closed-source models. It is worthwhile to note that Qwen2.5-VL-72B-Instruct achieved the highest score for the open source model on the English review and was reviewed for strong performance on several domains.", + "bbox": [ + 89, + 758, + 485, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Confidence Analysis The study of confidence analysis evaluates the consistency between an MLLM's expressed", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "confidence and the accuracy of its answers by requiring models to report confidence levels with their responses. As illustrated in Figure 4, two key patterns emerge: first, models demonstrate higher confidence when more reference materials are available, irrespective of architectural differences; second, while higher confidence generally correlates with greater accuracy, we observe numerous instances of high-confidence errors. This finding suggests that MLLMs exhibit overconfidence in their knowledge representation", + "bbox": [ + 511, + 758, + 908, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4642", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88b4a2daeb4e68c99f5692b25712cc60f28982b28c62d60e1982122135358ab4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsChinese partial resultsEnglish partial resultsF-score on 9 domains categories
COCGAF-scoreCOCGAF-scoreLESEHCCSETAFTMNSARCHCLI
Closed-source Multi-modal Large Language Models
GPT-4o48.751.650.145.750.748.147.037.558.262.050.561.730.329.658.8
GPT-4o-mini33.036.034.438.243.140.536.025.444.441.541.140.622.926.652.3
Doubao-vision-pro-128k51.156.453.628.541.833.951.823.750.152.249.044.332.644.948.7
Doubao-vision-pro-32k29.037.832.815.626.219.634.913.433.038.229.323.415.229.137.4
Gemini-2.0-flash54.857.956.350.754.352.554.135.459.967.570.964.629.236.364.6
Claude-3.5-Sonnet48.050.649.349.257.052.849.142.752.756.548.161.138.734.466.5
Qwen-Max25.229.427.125.630.627.928.121.035.434.831.620.014.211.944.1
ERNIE-VL54.055.354.640.744.942.755.932.960.054.642.850.727.845.459.4
", + "bbox": [ + 94, + 88, + 906, + 242 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b716aae26cc2cfaf17cd1eae6b2a8e0861f449b3326649d673b95eb15f6c067b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. SimpleVQA is structured around nine key domains: \"Literature, education & sports (LES)\", \"Euro-American History & Culture (EHC)\", \"Contemporary Society (CS)\", \"Engineering, Technology & Application (ETA)\", \"Film\", \"Television & Media (FTM)\", \"Natural Science (NS)\", \"Art (AR)\", \"Chinese History & Culture (CHC)\", and \"Life (LI)\"." + ], + "table_body": "
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO51.754.753.139.241.840.554.222.560.253.635.556.727.537.963.6
InternVL2.5-78B46.749.247.936.340.638.348.318.655.256.033.053.925.633.757.5
InternVL2-Llama3-76B34.237.435.737.140.438.738.319.044.044.435.445.620.421.457.1
InternVL2.5-38B-MPO48.049.548.837.740.939.252.321.054.259.529.351.527.931.661.6
InternVL2.5-26B-MPO45.547.346.334.438.836.446.920.654.547.632.451.620.534.454.3
InternVL2.5-8B-MPO35.838.537.131.434.032.736.014.043.743.226.446.420.322.553.5
Qwen2.5-VL-72B-Instruct48.050.449.250.754.052.349.445.264.362.453.658.320.930.361.8
Qwen2-VL-72B-Instruct46.148.647.343.351.246.945.230.454.057.251.653.529.429.765.2
Qwen2.5-VL-7B-Instruct41.944.042.944.547.145.842.942.854.949.840.446.425.930.356.9
Janus-pro-7B29.532.130.733.238.135.530.521.540.240.826.237.126.815.353.2
", + "bbox": [ + 94, + 243, + 906, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and comprehension capabilities.", + "bbox": [ + 89, + 491, + 305, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2864bd4d699dae78fb9c82d6705346f1ef588c5936c2b1f4d8847151af76d7c5.jpg", + "image_caption": [ + "Figure 4. Calibration of LLMs based on their stated confidence. The x-axis represents the confidence level of the LLMs, and the y-axis represents the accuracy." + ], + "image_footnote": [], + "bbox": [ + 101, + 523, + 473, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Different LLMs To validate the benchmark's reliance on visual information, we conduct experiments with eight mainstream LLMs using text-only inputs (without images). These models include well-known LLMs", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/01e87b0d754b457bfc9e558c45c262d5b86fd20f5ae85b67857ba33f176d31af.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. The CO, NA, IN, and CGA results for 8 LLMs across simpleVQA without image input." + ], + "table_body": "
ModelsCOINNACGA
o1-preview3.65%22.12%74.22%14.16%
o1-mini4.1%19.51%76.4%17.36%
DeepSeek-R110.08%53.6%36.32%15.82%
Qwen-Max7.6%70.77%21.63%9.69%
GPT-4o5.23%31.6%63.16%14.2%
GPT-4o-mini6.47%47.65%45.88%11.95%
Claude-3.5-Sonnet22.02%7.21%90.77%21.88%
Gemini-2.0-flash7.06%61.33%31.6%10.32%
", + "bbox": [ + 527, + 488, + 893, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "such as o1-Preview, o1-mini, and DeepSeek-R1. Table 5 presents the experimental results. We specifically design visual questions that theoretically require image inputs to be answered correctly. Despite the absence of visual information, LLMs achieve minimal accuracy on these tasks, with DeepSeek-R1 demonstrating notably higher performance in inferring correct answers from textual context alone.", + "bbox": [ + 511, + 667, + 906, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 786, + 681, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on SimpleVQA, we conduct a further evaluation of MLLMs, revealing significant factual accuracy issues. We", + "bbox": [ + 511, + 811, + 906, + 844 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "https://platform.openai.com/docs/models#ol \nhttps://platform.openai.com/docs/models#ol-mini \nhttps://huggingface.co/deepseek-ai/DeepSeek-R1", + "bbox": [ + 511, + 851, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4643", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "also perform a deep analysis of these factual problems by examining both the visual understanding capabilities and text generation mechanisms of MLLMs, which provides insights for future model optimization.", + "bbox": [ + 93, + 90, + 480, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we identify the three most robust MLLMs through systematic evaluation. For each VQA task where an MLLM provides an incorrect response, we use a carefully designed prompt to decompose the question into an atomic problem focused on basic content recognition. Each atomic problem corresponds to a fundamental atomic fact that, when identified, converts the original question into a purely factual text-based query. When an MLLM fails to answer even these simplified atomic queries correctly, we attribute this failure to insufficient visual understanding of the image. Next, we distinguish between original questions that are already atomic in nature and those that are not. We systematically identify cases where the derived atomic questions differ from the original questions and compile them into a test set called the Complex Fact Question (CFQ) set. This dataset is specifically designed to evaluate whether model performance improves when atomic facts are explicitly provided. In a follow-up experiment, we augment each CFQ query by incorporating the answer to its corresponding atomic question as an explicit hint. We then reassess model performance with these augmented queries. When models still provide incorrect answers despite having access to the relevant atomic facts, we attribute such failures to deficiencies in background knowledge rather than visual understanding.", + "bbox": [ + 93, + 152, + 480, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents the results of the CFQ experiments. From our complete dataset, we select 570 challenging CFQ examples to evaluate both visual understanding and knowledge internalization capabilities. In addition, another 90 questions are selected from which none MLLMs can answer correctly, to test the performance of mainstream LLMs in deep thinking. We test pure language models (LLMs) including o1-preview, o1-mini, and DeepSeek-R1, as well as multimodal models (MLLMs) such as GPT-4o, Qwen2.5-VL-72B-Instruct, and InternVL2.5-78B-MPO. Our findings reveal several key insights. First, LLMs, despite their reasoning capabilities, fail to effectively utilize atomic facts without accompanying images. Second, MLLMs demonstrate a significant disparity between their visual understanding abilities and knowledge internalization capabilities. While MLLMs show marginally better knowledge retention in relation to visual content, substantial room for improvement remains. Finally, the performance patterns observed when MLLMs answer atomic questions suggest potential for enhancing visual understanding through targeted training approaches.", + "bbox": [ + 93, + 513, + 480, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Works", + "text_level": 1, + "bbox": [ + 93, + 844, + 238, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal Benchmarks. Recent vision-language benchmarks have been developed to assess models' capabilities", + "bbox": [ + 93, + 869, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b3657a43b8760ecc62f17a14498967c78096207b972cf6e7d5383375153f853a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsOriginAtomicAtomic-GivenFix-Badcase
o1-preview--62.74%13.33% (90)
o1-mini--51.49%10.00% (90)
DeepSeek-R1--55.01%17.78% (90)
Qwen-Max--54.83%5.56% (90)
GPT-4o56.24%56.94%61.69%35.60% (250)
Qwen2.5-VL-72B-Instruct51.67%59.58%64.15%36.25% (276)
InternVL2.5-78B-MPO55.36%55.71%69.95%43.53% (255)
", + "bbox": [ + 517, + 88, + 903, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. CFQ experiment in SimpleVQA. \"Origin\": original Q&A accuracy; \"Atomic\": atomic Q&A accuracy; \"Atomic-Given\": original Q&A with atomic facts; \"Fix-Badcase\": corrected responses with atomic facts. Sample sizes are in parentheses.", + "bbox": [ + 517, + 200, + 903, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in integrating visual and textual information across various tasks [14, 32, 33, 42], including OCR [4, 6], spatial awareness [17], multimodal representation learning and retrieval [5, 19, 20], and reasoning skills [24, 26]. For example, MMBench [22] employs multiple-choice tasks in both Chinese and English, covering a wide range of domains. MMMU [40] focuses on complex vision-language tasks, particularly those requiring advanced multimodal reasoning. MMStar [2] utilizes multi-task evaluations to test models' ability to fuse different modalities.", + "bbox": [ + 517, + 279, + 903, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Factuality Benchmarks. Factuality refers a model's capacity to produce content consistent with factual knowledge, encompassing commonsense, general world knowledge, and domain-specific information. The capability can be evaluated through comparison against authoritative references, including Wikipedia and academic literature. Recently, Various benchmarks have been developed to evaluate factuality in LLMs [11, 16, 21, 27, 28, 36-38, 44]. For example, MMLU [10] assesses multitask accuracy across 57 diverse tasks. HaluEval [18] explores the propensity of LLMs to produce hallucinations or false information. SimpleQA [31] and Chinese SimpleQA [7] have been proposed to measure the short-form factuality in LLMs.", + "bbox": [ + 517, + 435, + 903, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 517, + 659, + 630, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents SimpleVQA, the first bilingual visual question-answering benchmark for evaluating fact-based reasoning in MLLMs. SimpleVQA features seven key attributes: Chinese-English bilingualism, multitask/multi-scene adaptability, high-quality content, challenging queries, static design, and evaluation efficiency. Our comprehensive evaluation of 18 MLLMs and 8 LLMs demonstrates its effectiveness in assessing fact-based reasoning capabilities. We also introduce the calibration methodology to evaluate MLLMs' visual comprehension and visual-linguistic alignment through atomic question decomposition of factual queries. SimpleVQA establishes a robust benchmark for assessing factuality and advancing the development of reliable MLLMs.", + "bbox": [ + 517, + 689, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4644", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex & Critical Software Environment (Grant No. SKLCCSE-2024ZX-18).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] AI@Meta. Llama 3 model card. 2024. 1", + "[2] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.8", + "[3] Qinyuan Cheng, Tianxiang Sun, Wenwei Zhang, Siyin Wang, Xiangyang Liu, Mozhi Zhang, Junliang He, Mianqiu Huang, Zhangyue Yin, Kai Chen, et al. Evaluating hallucinations in Chinese large language models. arXiv preprint arXiv:2310.03368, 2023. 1", + "[4] XianFu Cheng, YanQing Yao, Liying Zhang, Ao Liu, and Zhoujun Li. An improved stochastic gradient descent algorithm based on rényi differential privacy. International Journal of Intelligent Systems, 37(12):10694-10714, 2022. 8", + "[5] Xianfu Cheng, Hang Zhang, Jian Yang, Xiang Li, Weixiao Zhou, Kui Wu, Fei Liu, Wei Zhang, Tao Sun, Tongliang Li, et al. Xformparser: A simple and effective multimodal multilingual semi-structured form parser. arXiv preprint arXiv:2405.17336, 2024.8", + "[6] Xianfu Cheng, Weixiao Zhou, Xiang Li, Jian Yang, Hang Zhang, Tao Sun, Wei Zhang, Yuying Mai, Tongliang Li, Xiaoming Chen, et al. Sviptr: Fast and efficient scene text recognition with vision permutable extractor. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 365-373, 2024. 8", + "[7] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, Zhuoran Lin, Xuepeng Liu, Dekai Sun, Shirong Lin, Zhicheng Zheng, Xiaoyong Zhu, Wenbo Su, and Bo Zheng. Chinese simpleqa: A Chinese factuality evaluation for large language models, 2024. 8", + "[8] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 1", + "[9] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 3", + "[10] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021. 8" + ], + "bbox": [ + 94, + 232, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. arXiv preprint arXiv:2305.08322, 2023. 3, 8", + "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4", + "[13] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3", + "[14] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 8", + "[15] Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv: 2411.16594, 2024. 2", + "[16] Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. Cmmlu: Measuring massive multitask language understanding in Chinese, 2023. 8", + "[17] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 8", + "[18] Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. 8", + "[19] Xiang Li, Xianfu Cheng, Dezhuang Miao, Xiaoming Zhang, and Zhoujun Li. Tf-mamba: Text-enhanced fusion mamba with missing modalities for robust multimodal sentiment analysis. arXiv preprint arXiv:2505.14329, 2025. 8", + "[20] Xiang Li, Hajjun Zhang, Zhiqiang Dong, Xianfu Cheng, Yun Liu, and Xiaoming Zhang. Learning fine-grained representation with token-level alignment for multimodal sentiment analysis. Expert Systems with Applications, 269:126274, 2025. 8", + "[21] Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland, 2022. Association for Computational Linguistics. 8", + "[22] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3", + "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 8", + "[25] OpenAI. Gpt-4 technical report. PREPRINT, 2023. 1", + "[26] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 8", + "[27] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 8", + "[28] Yingshui Tan, Boren Zheng, Baihui Zheng, Kerui Cao, Huiyun Jing, Jincheng Wei, Jiaheng Liu, Yancheng He, Wenbo Su, Xiangyong Zhu, and Bo Zheng. Chinese safetyqa: A safety short-form factuality benchmark for large language models, 2024. 8", + "[29] SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024. 1", + "[30] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. URL https://apisemantic scholar.org/CorpusID, 273877483. 1", + "[31] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. 3, 8", + "[32] Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kajing Ma, Chenghao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, et al. Scimmir: Benchmarking scientific multi-modal information retrieval. arXiv preprint arXiv:2401.13478, 2024. 8", + "[33] Siwei Wu, Kang Zhu, Yu Bai, Yiming Liang, Yizhi Li, Haoning Wu, Jiaheng Liu, Ruibo Liu, Xingwei Qu, Xuxin Cheng, et al. Mmra: A benchmark for multi-granularity multi-image relational association. arXiv preprint arXiv:2407.17379, 2024.8", + "[34] Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et al. Tablebench: A comprehensive and complex benchmark for table question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 25497-25506, 2025. 2", + "[35] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 3", + "[36] Jian Yang, Jiaxi Yang, Ke Jin, Yibo Miao, Lei Zhang, Liquan Yang, Zeyu Cui, Yichang Zhang, Binyuan Hui, and Junyang Lin. Evaluating and aligning codellms on human preference. arXiv preprint arXiv:2412.05210, 2024. 8", + "[37] Jian Yang, Jiajun Zhang, Jiaxi Yang, Ke Jin, Lei Zhang, Qiyao Peng, Ken Deng, Yibo Miao, Tianyu Liu, Zeyu Cui, et al. Excrepobench: Multi-level executable code completion evaluation. arXiv preprint arXiv:2412.11990, 2024.", + "[38] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 8", + "[39] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3", + "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3, 8", + "[41] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3", + "[42] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, Wenhu Chen, and Jie Fu. Cmmm: A chinese massive multidiscipline multimodal understanding benchmark. ArXiv, abs/2401.11944, 2024. 8", + "[43] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1", + "[44] Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models, 2023. 3, 8", + "[45] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 823 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4646", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_model.json b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..461af0dacf3253a7f898a852219ce25973719ab4 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_model.json @@ -0,0 +1,1804 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.13, + 0.82, + 0.176 + ], + "angle": 0, + "content": "SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.203, + 0.871, + 0.327 + ], + "angle": 0, + "content": "Xianfu Cheng\\(^{1,2*}\\), Wei Zhang\\(^{1*}\\), Shiwei Zhang\\(^{2*}\\), Jian Yang\\(^{1*†}\\), Xiangyuan Guan\\(^{1}\\), Xianjie Wu\\(^{1}\\), Xiang Li\\(^{1†}\\), Ge Zhang\\(^{3}\\), Jiaheng Liu\\(^{3}\\), Yuying Mai\\(^{3}\\), Yutao Zeng\\(^{1}\\), Zhoufutu Wen\\(^{3}\\), Ke Jin\\(^{1,2}\\), Baorui Wang\\(^{1}\\), Weixiao Zhou\\(^{1}\\), Yunhong Lu\\(^{3}\\), Hangyuan Ji\\(^{1}\\), Tongliang Li\\(^{1}\\), Wenhao Huang\\(^{3}\\), Zhoujun Li\\(^{1,4}\\), \\(^{1}\\)Beihang University; \\(^{2}\\)Baidu Inc., China; \\(^{3}\\)M-A-P; \\(^{4}\\)Shenzhen Intelligent Strong Technology Co., Ltd. {buaacxf, lizj} @ buaa.edu.cn, zhangshiwei05@baidu.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.378, + 0.327, + 0.394 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.485, + 0.759 + ], + "angle": 0, + "content": "The increasing application of multi-modal large language models (MLLMs) across various sectors has spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by 7 key features: it is based on bilingual, it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 scenario domains. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.79, + 0.222, + 0.806 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.484, + 0.862 + ], + "angle": 0, + "content": "A significant challenge in large language models (LLMs) is ensuring that LLMs [1, 25] generate factually accurate and evidence-based responses. Current state-of-the-art (SOTA)" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.362, + 0.875, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.615, + 0.874, + 0.63 + ], + "angle": 0, + "content": "Figure 1. An example from our proposed SimpleVQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.72 + ], + "angle": 0, + "content": "LLMs often produce outputs that are misleading or unsupported by evidence phenomenon known as \"hallucinations\" [3, 29, 43]. The generation of incorrect or unsubstantiated information remains a major barrier to the broader adoption and reliability of general-purpose AI technologies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "OpenAI proposes SimpleQA [30] to measure factuality simple and reliable with nearly 4K concise and fact-seeking questions. Further, Chinese SimpleQA [8] comprised of 3K Chinese questions spanning 6 major topics is proposed to target the Chinese language. However, the SimpleQA and Chinese SimpleQA benchmarks mainly evaluate the model capabilities in text modality, ignoring wider real-world scenarios (e.g. vision modality). For the vision modality, the research progress of the multi-modal large language models (MLLMs) is still hindered by the \"hallucinations\" introduced by the given images. Therefore, The community of MLLMs has an urgent need for how to measure the simple" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.875, + 0.242, + 0.888 + ], + "angle": 0, + "content": "* Equal Contributions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.257, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Authors." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.875, + 0.257, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4637" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.405, + 0.107 + ], + "angle": 0, + "content": "and reliable factuality introduced by the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.486, + 0.563 + ], + "angle": 0, + "content": "To address this limitation, we develop the SimpleVQA benchmark as shown in Figure 1, where we define the factual question-answering capability of the visual language model. For the proposed factual Visual Question Answering (VQA), we collect 2,025 high-quality question-answer pairs covering 9 different scenario domains across 9 different application tasks. As a factual benchmark for a short answer, SimpleVQA has the following advantages: (1) English and Chinese: SimpleVQA provides general knowledge visual Q&A in both English and Chinese backgrounds, and comprehensively assesses the fact-generating capacity of MLLMs in Chinese and English communities. (2) Multi-task division: We divide the SimpleVQA assessment set into 9 different forms of VQA tasks according to the collected questions and different needs of pictures. (3) Diversified scenarios: SimpleVQA divided the VQA tasks into 9 uniquely representative scenario domains based on the category of the image and the source of the collected images. (4) High quality: We implement a comprehensive and rigorous quality control process to ensure the quality of questions and the accuracy of answers at SimpleVQA. (5) Challenge: simpleVQA focuses on factual questions that mainstream MLLMs cannot answer accurately, and cannot trace the cause of errors through the model itself. (6) Static answers: Following SimpleQA's factual definition, all the standard answers provided in our benchmark don't change over time. (7) Easy to evaluate: SimpleVQA's short answers make it possible to use existing LLMs to run a judge program (called LLM-as-a-Judge [15, 34]) to quickly determine right or wrong and get an overall accuracy rate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.485, + 0.853 + ], + "angle": 0, + "content": "We systematically evaluate 18 MLLMs on SimpleVQA and create a dynamic leaderboard to show results. Further, a series of probing experiments are performed to explore the effect of the key factors for SimpleVQA. We classify the capabilities possessed by MLLMs for factual questions into two aspects, visual understanding and internalized knowledge capabilities: (1) visual understanding refers to the ability of the model to identify the subject of the question being asked in the question; and (2) internalized knowledge capabilities test whether the model has already mastered the relevant knowledge of the subject of the question being asked, and thus is able to answer the relevant question correctly after identifying that subject. Based on this definition, we added an abductive reasoning experiment to the basic assessment to help determine whether the badcase came from a lack of visual understanding ability or a lack of internalized knowledge ability by generating and labeling atomic questions (each atomic question corresponds to an atomic fact) for each VQA example." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The remarkable findings from SimpleVQA are summarized as: (1) The factual accuracy of most evaluation models in the field of visual question-answering is insufficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.228 + ], + "angle": 0, + "content": "(2) The training data of MLLMs contains knowledge errors and they are overconfident in what they generate. (3) Image content understanding is still a major challenge for MLLMs to achieve improved capabilities. (4) The ability of MLLMs to internalize massive world knowledge still needs to be improved, and overcoming illusions remains a great challenge for large language models. (5) Further aligning visual core information with the knowledge stored in LLMs is also key to improving the factuality of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.242, + 0.642, + 0.26 + ], + "angle": 0, + "content": "2. SimpleVQA" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.625, + 0.282 + ], + "angle": 0, + "content": "2.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.908, + 0.53 + ], + "angle": 0, + "content": "The SimpleVQA benchmark consists of 2,025 samples spanning 9 core tasks and 9 primary domains, with each question-image pair categorized into relevant subcategories, enabling a comprehensive evaluation of MLLMs across diverse knowledge areas. The dataset 9 tasks, including covers Logic & Science (LS), Object Identification Recognition (OIR), Time & Event (TE), Person & Emotion (PE), Location & Building (LB), Text Processing (TP), Quantity & Position Relationship (QPR), Art & Culture (AC), and Object Attributes Recognition (OAR). To ensure broad topic coverage, SimpleVQA is structured around 9 key domains: Literature, education & sports (LES), Euro-American History & Culture (EHC), Contemporary Society (CS), Engineering, Technology & Application (ETA), Film, Television & Media (FTM), Natural Science (NS), Art (AR), Chinese History & Culture (CHC), and Life (LI)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.713 + ], + "angle": 0, + "content": "As shown in Table 1, SimpleVQA differs from existing MLLM benchmarks by focusing on factual knowledge boundaries instead of general vision-language understanding. Politically sensitive and ideological content is excluded to maintain neutrality and avoid controversy. Designed for efficiency, the dataset features concise questions and standardized answers, reducing complexity in model evaluation. All samples follow a short-answer Q&A format, enabling simple and objective assessment through direct answer matching. These refinements ensure SimpleVQA serves as a robust benchmark for evaluating MLLMs' factual reasoning abilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.674, + 0.737 + ], + "angle": 0, + "content": "2.2. Dataset Criteria" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.908, + 0.793 + ], + "angle": 0, + "content": "SimpleVQA maintains the following rigorous standards and guidelines for question, image, and answer selection, ensuring objectivity, temporal stability, and verifiability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Question Guidelines Clear and Unique Answers: Questions must have a single, undisputed answer. They should precisely define scope (e.g., \"Which city?\" instead of \"Which location?\") and specify time references (e.g., \"Which year?\" rather than \"When?\"). Evidence-Based: Each question must be supported by verifiable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4638" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.089, + 0.887, + 0.278 + ], + "angle": 0, + "content": "
BenchmarkMultimodalData SizeLanguageData SourceDomainFactualityReasoningMetric
MMbench [23]Image&Text2,438Chi.&Eng.Real WorldKnowledge××MCQ Eval
CCBench [23]Image&Text510ChineseKnowledgeKnowledge××MCQ Eval
MME [13]Image&Text1300EnglishReal WorldGeneral××TFQ Eval
MM-Vet [39]Image&Text200EnglishHumanGeneral××LLM-as-a-Judge
Dynamath [45]Image&Text5000EnglishExamsMath××Accuracy
MMMU [40]Image&Text11.5kEnglishExamsGeneral×✓Accuracy
MMMU-Pro [41]Image&Text3460EnglishHuman&GPTGeneral×✓Accuracy
ChineseFactEval [35]Text Only125ChineseHumanKnowledge✓×LLM-as-a-Judge
AGI-Eval [44]Text Only8062Chi.&Eng.ExamsKnowledge××Accuracy
C-Eval [11]Text Only13,948ChineseExamsKnowledge××Accuracy
SimpleQA [31]Text Only4,326EnglishHumanKnowledge✓×LLM-as-a-Judge
Chinese SimpleQA [9]Text Only3,000ChineseHuman&GPTKnowledge✓×LLM-as-a-Judge
SimpleVQA (Ours)Image&Text2,025Chi.&Eng.Human&GPTKnowledge✓✓LLM-as-a-Judge
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.288, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Table 1. Comparisons between our SimpleVQA and other benchmarks, where \"TFQ\" means True or False questions, \"MCQ\" means multi-choice questions, \"Chi.& Eng.\" means Chinese and English." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.33, + 0.887, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.536, + 0.712, + 0.55 + ], + "angle": 0, + "content": "Figure 2. An overview of the data construction process of SimpleVQA." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.564, + 0.483, + 0.716 + ], + "angle": 0, + "content": "sources. Manually annotated questions include reference links, while automatically generated ones undergo independent validation by two AI trainers. *Challenging for MLLMs*: Questions are tested on GPT-4o, GPT-4o-mini, doubao-vision-pro, and ERNIE-VL. Only those that at least one model answers incorrectly are retained; others are revised. Answerable by August 2024: All questions must be answerable based on knowledge available before September 1, 2024, ensuring a fair evaluation across models with similar knowledge cutoffs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.732, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Visual Guidelines No Direct Textual Clues: Images must not contain text revealing the answer. Authenticity: Only real, unaltered images are allowed to prevent factual distortion. Supports Question Reasoning: Each image must provide sufficient context for answering. Manually labeled samples undergo multi-annotator verification. Fixed Before August 2024: Image content must be valid and confirmable before August 2024." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Answer Guidelines Temporal Stability: Answers must remain unchanged and unaffected by new information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Time-sensitive topics (e.g., sports, media) should specify a timeframe rather than a general answer that may change. Sufficiently Challenging: Answers are tested against four high-precision MLLMs. If all models respond correctly, the question is revised to increase difficulty. Fully Objective and Evaluable: Answers must be precise, verifiable, and free from subjective interpretation. Unambiguous: Each answer must have a single, clear meaning to prevent misinterpretation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.791, + 0.729 + ], + "angle": 0, + "content": "2.3. Data Collection and Processing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Figure 2 shows an overview of the data construction process of SimpleVQA, which adheres to a meticulously designed five-step process, ensuring the highest standards of data quality and reliability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Step 1: Seed Example Collection. SimpleVQA's seed examples are sourced from two primary channels. First, we filter images and Q&A pairs from publicly available VQA datasets that align with factual knowledge criteria. We select MMVet (English), MME (English), Dynamath (English), MMbench_CN (Chinese), and CCBench (Chinese) due to their recent construction (post-2024) and their rele" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4639" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.198, + 0.089, + 0.803, + 0.263 + ], + "angle": 0, + "content": "
StatisticsNumberStatisticsNumber
Data2025Domain Categories9
- Chinsne(CN)1012- Literature, Education & Sports(LES)13.48%
- English(EN)1013- Euro-American History & Culture(EHC)12.89%
Task Categories9- Contemporary Society(CS)11.51%
- Logic & Science(LS)5.04%- Engineering, Technology & Application(ETA)7.95%
- Object Identification Recognition(OIR)14.07%- Film, Television & Media(FTM)10.62%
- Time & Event(TE)9.98%- Natural Science (NS)12.64%
- Person & Emotion(PE)13.58%- Art (AR)7.65%
- Location & Building(LB)21.53%- Chinese History & Culture(CHC)9.68%
- Text Processing(TP)10.07%- Life (LI)13.58%
- Quantity & Position Relationship(QPR)10.12%Word length of Questions
- Art & Culture(AC)9.09%- Maximum question length314
- Object Attributes Recognition(OAR)6.52%- Minimum problem length5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.397, + 0.271, + 0.6, + 0.286 + ], + "angle": 0, + "content": "Table 2. Statistics of SimpleVQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.482, + 0.406 + ], + "angle": 0, + "content": "vance to real-world applications. Second, we collect images and relevant factual knowledge from search engines (e.g., Google, Baidu, Wikipedia), with expert annotators generating corresponding questions and answers. These data focus on entities and events across multiple domains, ensuring answers are objective, fact-based, and centered on entity recognition or attribute extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Step 2: Data Enhancement and QA Pair Generation. Once sufficient seed examples are gathered, we employ GPT-4o [12] to refine the data and generate Q&A pairs for factual categories. For multiple-choice questions (MCQs) from sources like MMbench-CN and CCBench, to ensure answer uniqueness, we use LLMs to rephrase the original question and introduce qualifiers that precisely align with the correct response. For MME, we extract the answer entity and rewrite the question based on its attributes, ensuring a one-to-one correspondence. Datasets like MMVet, Dynamath, and CCBench, which contain discrepancies from factual Q&A formats (e.g., incorrect answer options, image descriptions, or MCQ distractors), are processed using GPT-4o to align the content with factual reasoning. These refinements produce the initial version of SimpleVQA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Step 3: LLM-Based Quality Verification. The refined dataset undergoes verification using GPT-4 to assess adherence to quality standards, ensuring answer stability, uniqueness, and question difficulty. Following LLM screening, two professional annotators conduct rigorous quality checks and refine samples as needed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Step 4: Difficulty Screening. To maximize the dataset's utility in model evaluation, we filter out overly simple Q&A pairs. We assess responses from four mainstream MLLMs (GPT-4o, GPT-4o-mini, Doubao-vision-pro, and ERNIE-VL). Any question correctly answered by all four models is deemed too simple and excluded from the dataset, thereby maintaining a challenging benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Step 5: Extracting Atomic Facts. To analyze visual comprehension and language alignment in MLLMs more precisely, we generate atomic questions from each SimpleVQA entry. An atomic fact represents the most funda" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.905, + 0.392 + ], + "angle": 0, + "content": "mental, indivisible attribute or characteristic of an object. For instance, given the question \"In what year was the person in the image born?\", the corresponding atomic question is \"Who is the person in the image?\". MLLMs generate candidate answers, which are then reviewed and refined by professional annotators to ensure accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.845, + 0.42 + ], + "angle": 0, + "content": "2.4. Human Annotation & Quality Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.623 + ], + "angle": 0, + "content": "To ensure dataset quality, we implement a rigorous manual validation process following automated data collection. Most of the collaborators participated in the necessary data annotation, and we also selected three domain experts from the collaborators. Each question is independently reviewed by two expert annotators to verify factual accuracy. If either annotator finds a question unsuitable, it is discarded. Annotations fact-check answers using authoritative sources such as Wikipedia and Baidu Encyclopedia, providing at least two supporting URLs. If their answers differ, a third expert conducts a final review to ensure consistency and correctness. Only Q&A pairs that fully align with both human evaluations and LLM-generated responses are retained." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.837 + ], + "angle": 0, + "content": "A difficulty assessment further refines the dataset. We begin with 8,360 Q&A pairs, filtering out \\(22\\%\\) of image-based samples that lack challenge or fail to meet predefined criteria. 1,108 pairs are removed through multi-model testing to ensure that questions pose a meaningful challenge to MLLMs. To maintain category balance, we carefully select 200 high-difficulty mathematical Q&As from 5,000 Dynamath samples, avoiding an overrepresentation of simpler factual questions. Through multiple validation rounds, we retain 2,025 high-precision Q&A pairs, accounting for \\(24\\%\\) of the original dataset. This process ensures factual integrity, topic diversity, and appropriate difficulty levels, making SimpleVQA a robust benchmark for evaluating MLLMs' reasoning and knowledge boundaries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.682, + 0.862 + ], + "angle": 0, + "content": "2.5. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1 and 2, our SimpleVQA benchmark consists of 2,025 samples across 9 major tasks, 9 major do" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4640" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.41 + ], + "angle": 0, + "content": "mains, and 244 image types. Examples of each category can be found in supplementary material. This design facilitates a comprehensive assessment of MLLMs across different domains. Regarding the distribution of domain topics and image types in SimpleVQA, 9 main domains are defined and subcategories are assigned based on each topic. Table 2 displays statistics related to SimpleVQA, including the amount of data, categorization and percentage, and the text length of the questions. In Table 1, we also compare SimpleVQA with several mainstream MLLMs' evaluation benchmarks, which suggests that SimpleVQA is the first MLLMs' benchmark that focuses on the evaluation of knowledge boundaries in factual categories. We excluded ideological and politically relevant data from the dataset to prevent social controversies and negative impacts. In addition, we implemented several optimizations to improve the efficiency of the evaluation. The dataset features concise questions and standardized answers, minimizing the input and output markers required for GPT assessment. In addition, all examples are in short-answer question-and-answer (QA) format, and they can be assessed by simple matching." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.223, + 0.451 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.281, + 0.478 + ], + "angle": 0, + "content": "3.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.592 + ], + "angle": 0, + "content": "We implement a standardized protocol across all experiments, maintaining a consistent prompt format throughout. All temperature and sampling parameters are configured according to each LLM's official recommendations or default settings. For evaluation and benchmark construction, we employ GPT-4o as the primary model due to its robust performance in multimodal tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.253, + 0.627 + ], + "angle": 0, + "content": "3.2. Baseline Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.759 + ], + "angle": 0, + "content": "We evaluate 18 representative MLLMs in total, comprising 8 closed-source and 10 open-source models, which provide a diverse evaluation of model capabilities across different architectures and training paradigms. The closed-source models include GPT-4o and 4o-mini, Doubao-pro-128k and 32k, Gemini-2.0-flash, Claude-3.5-Sonnet, Qwen-Max, ERNIE-VL. The open-source models cover a wide range of frameworks, including InternVL2.5 and In" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.878, + 0.108 + ], + "angle": 0, + "content": "ternVL2, Qwen2.5-VL and Qwen2-VL, Janus-pro-7B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.117, + 0.698, + 0.133 + ], + "angle": 0, + "content": "3.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.908, + 0.368 + ], + "angle": 0, + "content": "The SimpleVQA benchmark employs a set of evaluation metrics to assess the accuracy, reliability, and consistency of the model's predictions. These metrics include: (1) Correct (CO) evaluates whether the predicted answer matches the reference answer exactly without any contradiction. (2) Not Attempted (NA) identifies cases where the model does not attempt to answer, ensuring no contradictions are present. (3) Incorrect (IN) flags instances where the predicted answer contradicts the reference answer, even if resolved. (4) Correct Given Attempted (CGA) measures the proportion of correct answers among those attempted by the model, reflecting its performance when engaged. (5) F-score computes the harmonic mean between \"Correct\" and \"Correct Given Attempted\", providing a balanced evaluation that combines accuracy and attempt success." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.377, + 0.652, + 0.391 + ], + "angle": 0, + "content": "3.4. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.598 + ], + "angle": 0, + "content": "Results on Overall and Different Tasks Table 3 presents the performance of MLLMs on SimpleVQA, highlighting their F-scores across different tasks. Among the closed-source models, Gemini-2.0-Flash demonstrates superior performance, particularly in tasks such as LS, OIR, TE, PE, LB, TP, and QPR, achieving state-of-the-art (SOTA) scores across these task categories. Claude-3.5-Sonnet, GPT-4o, and ERNIE also exhibit strong performance. Open-source models, including InternVL2.5-78BMPO and Qwen2.5-VL-72B-Instruct, show competitive results, though slightly below the leading closed-source models. Notably, Qwen2.5-VL-72B-Instruct outperforms most mainstream closed-source MLLMs in several aspects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.909, + 0.783 + ], + "angle": 0, + "content": "Results on Cross-lingual and Different Domains Table 4 lists the performance of various closed-source and open-source MLLMs on SimpleVQA, highlighting their Q&A performance across languages and across different scenarios. Among the closed-source models, Gemini-2.0-Flash and ERNIE-VL show strong Chinese quizzing ability; Gemini-2.0-Flash and Claude-3.5-Sonnet show strong English quizzing ability; especially on the scale evaluation with scenarios as the categorization, the three models mentioned above achieve almost all the SOTA. opensource models InternVL2.5-78B-MPO and Qwen2.5-VL" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "https://platform.openai.com/docs/models#gpt-4o; https://platform.openai.com/docs/models#gpt-4o-mini \nhttps://www.volcengine.com/product/ark \nhttps://deepmind.google/technologies/gemini/flash/ \nhttps://www.anthropic.com/news/claude-3-5-sonnet \nhttps://huggingface.co/spaces/Qwen/Qwen-Max \nhttps://yiyan.baidu.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.514, + 0.792, + 0.907, + 0.9 + ], + "angle": 0, + "content": "https://huggingface.co/OpenGVLab/InternVL2_5-78B-MPO; https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B \nhttps://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct; https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct; https://huggingface.co/Qwen/Qwen2-VL-72B-Instruct \nhttps://huggingface.co/deepseek-ai/Janus-Pro-7B" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4641" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.43 + ], + "angle": 0, + "content": "
ModelsOverall resultsF-score on 9 task categories
CONAINCGAF-scoreLSOIRTEPELBTPQPRACOAR
Closed-source Multi-modal Large Language Models
GPT-4o47.27.845.051.249.158.753.448.136.444.357.561.835.161.0
GPT-4o-mini35.510.054.539.437.330.044.438.529.636.242.145.925.840.6
Doubao-vision-pro-128k39.720.839.550.144.348.353.553.439.635.038.546.935.962.9
Doubao-vision-pro-32k25.423.651.432.728.324.237.918.031.621.339.232.918.636.6
Gemini-2.0-flash52.86.041.256.154.463.760.954.355.044.361.565.733.964.4
Claude-3.5-Sonnet48.59.841.853.750.957.153.952.129.747.258.462.746.760.1
Qwen-Max25.415.359.330.027.515.133.713.530.127.736.034.412.136.4
ERNIE-VL46.59.544.051.448.849.055.948.340.740.754.459.733.670.8
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO45.45.948.648.346.857.454.650.326.434.457.065.832.872.3
InternVL2.5-78B41.57.750.845.043.249.549.445.028.431.049.663.732.065.2
InternVL2-Llama3-76B35.78.455.938.937.234.743.535.626.228.444.353.529.054.1
InternVL2.5-38B-MPO42.95.451.845.344.051.752.545.722.134.052.067.832.860.4
InternVL2.5-26B-MPO39.97.652.543.241.543.444.547.727.131.745.958.629.068.0
InternVL2.5-8B-MPO33.67.558.936.334.945.137.430.919.126.044.352.226.456.2
Qwen2.5-VL-72B49.45.445.252.250.857.750.851.038.651.657.865.829.862.8
Qwen2-VL-72B-Instruct44.710.345.049.847.148.851.946.337.938.555.363.035.959.7
Qwen2.5-VL-7B-Instruct43.25.151.745.644.338.652.737.632.741.052.653.639.256.7
Janus-pro-7B31.310.658.135.033.027.043.426.523.728.224.050.436.242.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.441, + 0.907, + 0.498 + ], + "angle": 0, + "content": "Table 3. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. We report the scores across different tasks, including \"Logic & Science (LS)\", \"Object Identification Recognition (OIR)\", \"Time & Event (TE)\", \"Person & Emotion (PE)\", \"Location & Building (LB)\", \"Text Processing (TP)\", \"Quantity & Position Relationship (QPR)\", \"Art & Culture (AC)\", and \"Object Attributes Recognition (OAR)\"." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.511, + 0.26, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.511, + 0.473, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.511, + 0.688, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.511, + 0.902, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.732, + 0.731, + 0.746 + ], + "angle": 0, + "content": "Figure 3. F-score results for eight different models across nine task categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.486, + 0.851 + ], + "angle": 0, + "content": "72B-Instruct, respectively, achieved the highest performance, albeit slightly lower than the top closed-source models. It is worthwhile to note that Qwen2.5-VL-72B-Instruct achieved the highest score for the open source model on the English review and was reviewed for strong performance on several domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Confidence Analysis The study of confidence analysis evaluates the consistency between an MLLM's expressed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.909, + 0.898 + ], + "angle": 0, + "content": "confidence and the accuracy of its answers by requiring models to report confidence levels with their responses. As illustrated in Figure 4, two key patterns emerge: first, models demonstrate higher confidence when more reference materials are available, irrespective of architectural differences; second, while higher confidence generally correlates with greater accuracy, we observe numerous instances of high-confidence errors. This finding suggests that MLLMs exhibit overconfidence in their knowledge representation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4642" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.243 + ], + "angle": 0, + "content": "
ModelsChinese partial resultsEnglish partial resultsF-score on 9 domains categories
COCGAF-scoreCOCGAF-scoreLESEHCCSETAFTMNSARCHCLI
Closed-source Multi-modal Large Language Models
GPT-4o48.751.650.145.750.748.147.037.558.262.050.561.730.329.658.8
GPT-4o-mini33.036.034.438.243.140.536.025.444.441.541.140.622.926.652.3
Doubao-vision-pro-128k51.156.453.628.541.833.951.823.750.152.249.044.332.644.948.7
Doubao-vision-pro-32k29.037.832.815.626.219.634.913.433.038.229.323.415.229.137.4
Gemini-2.0-flash54.857.956.350.754.352.554.135.459.967.570.964.629.236.364.6
Claude-3.5-Sonnet48.050.649.349.257.052.849.142.752.756.548.161.138.734.466.5
Qwen-Max25.229.427.125.630.627.928.121.035.434.831.620.014.211.944.1
ERNIE-VL54.055.354.640.744.942.755.932.960.054.642.850.727.845.459.4
" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.244, + 0.907, + 0.398 + ], + "angle": 0, + "content": "
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO51.754.753.139.241.840.554.222.560.253.635.556.727.537.963.6
InternVL2.5-78B46.749.247.936.340.638.348.318.655.256.033.053.925.633.757.5
InternVL2-Llama3-76B34.237.435.737.140.438.738.319.044.044.435.445.620.421.457.1
InternVL2.5-38B-MPO48.049.548.837.740.939.252.321.054.259.529.351.527.931.661.6
InternVL2.5-26B-MPO45.547.346.334.438.836.446.920.654.547.632.451.620.534.454.3
InternVL2.5-8B-MPO35.838.537.131.434.032.736.014.043.743.226.446.420.322.553.5
Qwen2.5-VL-72B-Instruct48.050.449.250.754.052.349.445.264.362.453.658.320.930.361.8
Qwen2-VL-72B-Instruct46.148.647.343.351.246.945.230.454.057.251.653.529.429.765.2
Qwen2.5-VL-7B-Instruct41.944.042.944.547.145.842.942.854.949.840.446.425.930.356.9
Janus-pro-7B29.532.130.733.238.135.530.521.540.240.826.237.126.815.353.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.409, + 0.907, + 0.466 + ], + "angle": 0, + "content": "Table 4. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. SimpleVQA is structured around nine key domains: \"Literature, education & sports (LES)\", \"Euro-American History & Culture (EHC)\", \"Contemporary Society (CS)\", \"Engineering, Technology & Application (ETA)\", \"Film\", \"Television & Media (FTM)\", \"Natural Science (NS)\", \"Art (AR)\", \"Chinese History & Culture (CHC)\", and \"Life (LI)\"." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.492, + 0.306, + 0.507 + ], + "angle": 0, + "content": "and comprehension capabilities." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.525, + 0.475, + 0.762 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Figure 4. Calibration of LLMs based on their stated confidence. The x-axis represents the confidence level of the LLMs, and the y-axis represents the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Results on Different LLMs To validate the benchmark's reliance on visual information, we conduct experiments with eight mainstream LLMs using text-only inputs (without images). These models include well-known LLMs" + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.489, + 0.895, + 0.617 + ], + "angle": 0, + "content": "
ModelsCOINNACGA
o1-preview3.65%22.12%74.22%14.16%
o1-mini4.1%19.51%76.4%17.36%
DeepSeek-R110.08%53.6%36.32%15.82%
Qwen-Max7.6%70.77%21.63%9.69%
GPT-4o5.23%31.6%63.16%14.2%
GPT-4o-mini6.47%47.65%45.88%11.95%
Claude-3.5-Sonnet22.02%7.21%90.77%21.88%
Gemini-2.0-flash7.06%61.33%31.6%10.32%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.626, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Table 5. The CO, NA, IN, and CGA results for 8 LLMs across simpleVQA without image input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.775 + ], + "angle": 0, + "content": "such as o1-Preview, o1-mini, and DeepSeek-R1. Table 5 presents the experimental results. We specifically design visual questions that theoretically require image inputs to be answered correctly. Despite the absence of visual information, LLMs achieve minimal accuracy on these tasks, with DeepSeek-R1 demonstrating notably higher performance in inferring correct answers from textual context alone." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.682, + 0.804 + ], + "angle": 0, + "content": "4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.813, + 0.907, + 0.845 + ], + "angle": 0, + "content": "Based on SimpleVQA, we conduct a further evaluation of MLLMs, revealing significant factual accuracy issues. We" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.852, + 0.906, + 0.901 + ], + "angle": 0, + "content": "https://platform.openai.com/docs/models#ol \nhttps://platform.openai.com/docs/models#ol-mini \nhttps://huggingface.co/deepseek-ai/DeepSeek-R1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4643" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.481, + 0.151 + ], + "angle": 0, + "content": "also perform a deep analysis of these factual problems by examining both the visual understanding capabilities and text generation mechanisms of MLLMs, which provides insights for future model optimization." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.153, + 0.482, + 0.513 + ], + "angle": 0, + "content": "First, we identify the three most robust MLLMs through systematic evaluation. For each VQA task where an MLLM provides an incorrect response, we use a carefully designed prompt to decompose the question into an atomic problem focused on basic content recognition. Each atomic problem corresponds to a fundamental atomic fact that, when identified, converts the original question into a purely factual text-based query. When an MLLM fails to answer even these simplified atomic queries correctly, we attribute this failure to insufficient visual understanding of the image. Next, we distinguish between original questions that are already atomic in nature and those that are not. We systematically identify cases where the derived atomic questions differ from the original questions and compile them into a test set called the Complex Fact Question (CFQ) set. This dataset is specifically designed to evaluate whether model performance improves when atomic facts are explicitly provided. In a follow-up experiment, we augment each CFQ query by incorporating the answer to its corresponding atomic question as an explicit hint. We then reassess model performance with these augmented queries. When models still provide incorrect answers despite having access to the relevant atomic facts, we attribute such failures to deficiencies in background knowledge rather than visual understanding." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.515, + 0.482, + 0.831 + ], + "angle": 0, + "content": "Table 6 presents the results of the CFQ experiments. From our complete dataset, we select 570 challenging CFQ examples to evaluate both visual understanding and knowledge internalization capabilities. In addition, another 90 questions are selected from which none MLLMs can answer correctly, to test the performance of mainstream LLMs in deep thinking. We test pure language models (LLMs) including o1-preview, o1-mini, and DeepSeek-R1, as well as multimodal models (MLLMs) such as GPT-4o, Qwen2.5-VL-72B-Instruct, and InternVL2.5-78B-MPO. Our findings reveal several key insights. First, LLMs, despite their reasoning capabilities, fail to effectively utilize atomic facts without accompanying images. Second, MLLMs demonstrate a significant disparity between their visual understanding abilities and knowledge internalization capabilities. While MLLMs show marginally better knowledge retention in relation to visual content, substantial room for improvement remains. Finally, the performance patterns observed when MLLMs answer atomic questions suggest potential for enhancing visual understanding through targeted training approaches." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.845, + 0.239, + 0.861 + ], + "angle": 0, + "content": "5. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.871, + 0.481, + 0.9 + ], + "angle": 0, + "content": "Multimodal Benchmarks. Recent vision-language benchmarks have been developed to assess models' capabilities" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.904, + 0.189 + ], + "angle": 0, + "content": "
ModelsOriginAtomicAtomic-GivenFix-Badcase
o1-preview--62.74%13.33% (90)
o1-mini--51.49%10.00% (90)
DeepSeek-R1--55.01%17.78% (90)
Qwen-Max--54.83%5.56% (90)
GPT-4o56.24%56.94%61.69%35.60% (250)
Qwen2.5-VL-72B-Instruct51.67%59.58%64.15%36.25% (276)
InternVL2.5-78B-MPO55.36%55.71%69.95%43.53% (255)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.201, + 0.904, + 0.255 + ], + "angle": 0, + "content": "Table 6. CFQ experiment in SimpleVQA. \"Origin\": original Q&A accuracy; \"Atomic\": atomic Q&A accuracy; \"Atomic-Given\": original Q&A with atomic facts; \"Fix-Badcase\": corrected responses with atomic facts. Sample sizes are in parentheses." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.28, + 0.904, + 0.43 + ], + "angle": 0, + "content": "in integrating visual and textual information across various tasks [14, 32, 33, 42], including OCR [4, 6], spatial awareness [17], multimodal representation learning and retrieval [5, 19, 20], and reasoning skills [24, 26]. For example, MMBench [22] employs multiple-choice tasks in both Chinese and English, covering a wide range of domains. MMMU [40] focuses on complex vision-language tasks, particularly those requiring advanced multimodal reasoning. MMStar [2] utilizes multi-task evaluations to test models' ability to fuse different modalities." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.436, + 0.904, + 0.63 + ], + "angle": 0, + "content": "Factuality Benchmarks. Factuality refers a model's capacity to produce content consistent with factual knowledge, encompassing commonsense, general world knowledge, and domain-specific information. The capability can be evaluated through comparison against authoritative references, including Wikipedia and academic literature. Recently, Various benchmarks have been developed to evaluate factuality in LLMs [11, 16, 21, 27, 28, 36-38, 44]. For example, MMLU [10] assesses multitask accuracy across 57 diverse tasks. HaluEval [18] explores the propensity of LLMs to produce hallucinations or false information. SimpleQA [31] and Chinese SimpleQA [7] have been proposed to measure the short-form factuality in LLMs." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.66, + 0.631, + 0.675 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.69, + 0.904, + 0.9 + ], + "angle": 0, + "content": "This paper presents SimpleVQA, the first bilingual visual question-answering benchmark for evaluating fact-based reasoning in MLLMs. SimpleVQA features seven key attributes: Chinese-English bilingualism, multitask/multi-scene adaptability, high-quality content, challenging queries, static design, and evaluation efficiency. Our comprehensive evaluation of 18 MLLMs and 8 LLMs demonstrates its effectiveness in assessing fact-based reasoning capabilities. We also introduce the calibration methodology to evaluate MLLMs' visual comprehension and visual-linguistic alignment through atomic question decomposition of factual queries. SimpleVQA establishes a robust benchmark for assessing factuality and advancing the development of reliable MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4644" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex & Critical Software Environment (Grant No. SKLCCSE-2024ZX-18)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.369, + 0.246 + ], + "angle": 0, + "content": "[1] AI@Meta. Llama 3 model card. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[2] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[3] Qinyuan Cheng, Tianxiang Sun, Wenwei Zhang, Siyin Wang, Xiangyang Liu, Mozhi Zhang, Junliang He, Mianqiu Huang, Zhangyue Yin, Kai Chen, et al. Evaluating hallucinations in Chinese large language models. arXiv preprint arXiv:2310.03368, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[4] XianFu Cheng, YanQing Yao, Liying Zhang, Ao Liu, and Zhoujun Li. An improved stochastic gradient descent algorithm based on rényi differential privacy. International Journal of Intelligent Systems, 37(12):10694-10714, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.449, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[5] Xianfu Cheng, Hang Zhang, Jian Yang, Xiang Li, Weixiao Zhou, Kui Wu, Fei Liu, Wei Zhang, Tao Sun, Tongliang Li, et al. Xformparser: A simple and effective multimodal multilingual semi-structured form parser. arXiv preprint arXiv:2405.17336, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[6] Xianfu Cheng, Weixiao Zhou, Xiang Li, Jian Yang, Hang Zhang, Tao Sun, Wei Zhang, Yuying Mai, Tongliang Li, Xiaoming Chen, et al. Sviptr: Fast and efficient scene text recognition with vision permutable extractor. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 365-373, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[7] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, Zhuoran Lin, Xuepeng Liu, Dekai Sun, Shirong Lin, Zhicheng Zheng, Xiaoyong Zhu, Wenbo Su, and Bo Zheng. Chinese simpleqa: A Chinese factuality evaluation for large language models, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[9] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.233, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[11] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. arXiv preprint arXiv:2305.08322, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.304, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[15] Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv: 2411.16594, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.904, + 0.509 + ], + "angle": 0, + "content": "[16] Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. Cmmlu: Measuring massive multitask language understanding in Chinese, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[17] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[18] Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[19] Xiang Li, Xianfu Cheng, Dezhuang Miao, Xiaoming Zhang, and Zhoujun Li. Tf-mamba: Text-enhanced fusion mamba with missing modalities for robust multimodal sentiment analysis. arXiv preprint arXiv:2505.14329, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.68, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[20] Xiang Li, Hajjun Zhang, Zhiqiang Dong, Xianfu Cheng, Yun Liu, and Xiaoming Zhang. Learning fine-grained representation with token-level alignment for multimodal sentiment analysis. Expert Systems with Applications, 269:126274, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[21] Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland, 2022. Association for Computational Linguistics. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[22] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4645" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.443, + 0.246 + ], + "angle": 0, + "content": "[25] OpenAI. Gpt-4 technical report. PREPRINT, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[26] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[27] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[28] Yingshui Tan, Boren Zheng, Baihui Zheng, Kerui Cao, Huiyun Jing, Jincheng Wei, Jiaheng Liu, Yancheng He, Wenbo Su, Xiangyong Zhu, and Bo Zheng. Chinese safetyqa: A safety short-form factuality benchmark for large language models, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[29] SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[30] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. URL https://apisemantic scholar.org/CorpusID, 273877483. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[31] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[32] Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kajing Ma, Chenghao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, et al. Scimmir: Benchmarking scientific multi-modal information retrieval. arXiv preprint arXiv:2401.13478, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[33] Siwei Wu, Kang Zhu, Yu Bai, Yiming Liang, Yizhi Li, Haoning Wu, Jiaheng Liu, Ruibo Liu, Xingwei Qu, Xuxin Cheng, et al. Mmra: A benchmark for multi-granularity multi-image relational association. arXiv preprint arXiv:2407.17379, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et al. Tablebench: A comprehensive and complex benchmark for table question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 25497-25506, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[36] Jian Yang, Jiaxi Yang, Ke Jin, Yibo Miao, Lei Zhang, Liquan Yang, Zeyu Cui, Yichang Zhang, Binyuan Hui, and Junyang Lin. Evaluating and aligning codellms on human preference. arXiv preprint arXiv:2412.05210, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[37] Jian Yang, Jiajun Zhang, Jiaxi Yang, Ke Jin, Lei Zhang, Qiyao Peng, Ken Deng, Yibo Miao, Tianyu Liu, Zeyu Cui, et al. Excrepobench: Multi-level executable code completion evaluation. arXiv preprint arXiv:2412.11990, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[38] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[39] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[41] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[42] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, Wenhu Chen, and Jie Fu. Cmmm: A chinese massive multidiscipline multimodal understanding benchmark. ArXiv, abs/2401.11944, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[43] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[44] Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.756, + 0.905, + 0.824 + ], + "angle": 0, + "content": "[45] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4646" + } + ] +] \ No newline at end of file diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_origin.pdf b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..aa00e774ee53f94cffb9de4b0fa637ce30485e71 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/d1c51684-4f7b-42eb-97c5-436f06ea7c3e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20e719be145eb177ccc11371fbfd92e8dba6c260d8477dca83b284b13b856580 +size 2920296 diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/full.md b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a488d05dec78018ad2a15227aaa39823586cc42 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/full.md @@ -0,0 +1,233 @@ +# SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models + +Xianfu Cheng $^{1,2*}$ , Wei Zhang $^{1*}$ , Shiwei Zhang $^{2*}$ , Jian Yang $^{1*†}$ , Xiangyuan Guan $^{1}$ , Xianjie Wu $^{1}$ , Xiang Li $^{1†}$ , Ge Zhang $^{3}$ , Jiaheng Liu $^{3}$ , Yuying Mai $^{3}$ , Yutao Zeng $^{1}$ , Zhoufutu Wen $^{3}$ , Ke Jin $^{1,2}$ , Baorui Wang $^{1}$ , Weixiao Zhou $^{1}$ , Yunhong Lu $^{3}$ , Hangyuan Ji $^{1}$ , Tongliang Li $^{1}$ , Wenhao Huang $^{3}$ , Zhoujun Li $^{1,4}$ , $^{1}$ Beihang University; $^{2}$ Baidu Inc., China; $^{3}$ M-A-P; $^{4}$ Shenzhen Intelligent Strong Technology Co., Ltd. {buaacxf, lizj} @ buaa.edu.cn, zhangshiwei05@baidu.com + +# Abstract + +The increasing application of multi-modal large language models (MLLMs) across various sectors has spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by 7 key features: it is based on bilingual, it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 scenario domains. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases. + +# 1. Introduction + +A significant challenge in large language models (LLMs) is ensuring that LLMs [1, 25] generate factually accurate and evidence-based responses. Current state-of-the-art (SOTA) + +![](images/1f61626ebd2b36af0ed3c9cb4cea56ac0ba065e59bd05249dec00a580c3babc3.jpg) +Figure 1. An example from our proposed SimpleVQA. + +LLMs often produce outputs that are misleading or unsupported by evidence phenomenon known as "hallucinations" [3, 29, 43]. The generation of incorrect or unsubstantiated information remains a major barrier to the broader adoption and reliability of general-purpose AI technologies. + +OpenAI proposes SimpleQA [30] to measure factuality simple and reliable with nearly 4K concise and fact-seeking questions. Further, Chinese SimpleQA [8] comprised of 3K Chinese questions spanning 6 major topics is proposed to target the Chinese language. However, the SimpleQA and Chinese SimpleQA benchmarks mainly evaluate the model capabilities in text modality, ignoring wider real-world scenarios (e.g. vision modality). For the vision modality, the research progress of the multi-modal large language models (MLLMs) is still hindered by the "hallucinations" introduced by the given images. Therefore, The community of MLLMs has an urgent need for how to measure the simple + +and reliable factuality introduced by the image. + +To address this limitation, we develop the SimpleVQA benchmark as shown in Figure 1, where we define the factual question-answering capability of the visual language model. For the proposed factual Visual Question Answering (VQA), we collect 2,025 high-quality question-answer pairs covering 9 different scenario domains across 9 different application tasks. As a factual benchmark for a short answer, SimpleVQA has the following advantages: (1) English and Chinese: SimpleVQA provides general knowledge visual Q&A in both English and Chinese backgrounds, and comprehensively assesses the fact-generating capacity of MLLMs in Chinese and English communities. (2) Multi-task division: We divide the SimpleVQA assessment set into 9 different forms of VQA tasks according to the collected questions and different needs of pictures. (3) Diversified scenarios: SimpleVQA divided the VQA tasks into 9 uniquely representative scenario domains based on the category of the image and the source of the collected images. (4) High quality: We implement a comprehensive and rigorous quality control process to ensure the quality of questions and the accuracy of answers at SimpleVQA. (5) Challenge: simpleVQA focuses on factual questions that mainstream MLLMs cannot answer accurately, and cannot trace the cause of errors through the model itself. (6) Static answers: Following SimpleQA's factual definition, all the standard answers provided in our benchmark don't change over time. (7) Easy to evaluate: SimpleVQA's short answers make it possible to use existing LLMs to run a judge program (called LLM-as-a-Judge [15, 34]) to quickly determine right or wrong and get an overall accuracy rate. + +We systematically evaluate 18 MLLMs on SimpleVQA and create a dynamic leaderboard to show results. Further, a series of probing experiments are performed to explore the effect of the key factors for SimpleVQA. We classify the capabilities possessed by MLLMs for factual questions into two aspects, visual understanding and internalized knowledge capabilities: (1) visual understanding refers to the ability of the model to identify the subject of the question being asked in the question; and (2) internalized knowledge capabilities test whether the model has already mastered the relevant knowledge of the subject of the question being asked, and thus is able to answer the relevant question correctly after identifying that subject. Based on this definition, we added an abductive reasoning experiment to the basic assessment to help determine whether the badcase came from a lack of visual understanding ability or a lack of internalized knowledge ability by generating and labeling atomic questions (each atomic question corresponds to an atomic fact) for each VQA example. + +The remarkable findings from SimpleVQA are summarized as: (1) The factual accuracy of most evaluation models in the field of visual question-answering is insufficient. + +(2) The training data of MLLMs contains knowledge errors and they are overconfident in what they generate. (3) Image content understanding is still a major challenge for MLLMs to achieve improved capabilities. (4) The ability of MLLMs to internalize massive world knowledge still needs to be improved, and overcoming illusions remains a great challenge for large language models. (5) Further aligning visual core information with the knowledge stored in LLMs is also key to improving the factuality of MLLMs. + +# 2. SimpleVQA + +# 2.1. Overview + +The SimpleVQA benchmark consists of 2,025 samples spanning 9 core tasks and 9 primary domains, with each question-image pair categorized into relevant subcategories, enabling a comprehensive evaluation of MLLMs across diverse knowledge areas. The dataset 9 tasks, including covers Logic & Science (LS), Object Identification Recognition (OIR), Time & Event (TE), Person & Emotion (PE), Location & Building (LB), Text Processing (TP), Quantity & Position Relationship (QPR), Art & Culture (AC), and Object Attributes Recognition (OAR). To ensure broad topic coverage, SimpleVQA is structured around 9 key domains: Literature, education & sports (LES), Euro-American History & Culture (EHC), Contemporary Society (CS), Engineering, Technology & Application (ETA), Film, Television & Media (FTM), Natural Science (NS), Art (AR), Chinese History & Culture (CHC), and Life (LI). + +As shown in Table 1, SimpleVQA differs from existing MLLM benchmarks by focusing on factual knowledge boundaries instead of general vision-language understanding. Politically sensitive and ideological content is excluded to maintain neutrality and avoid controversy. Designed for efficiency, the dataset features concise questions and standardized answers, reducing complexity in model evaluation. All samples follow a short-answer Q&A format, enabling simple and objective assessment through direct answer matching. These refinements ensure SimpleVQA serves as a robust benchmark for evaluating MLLMs' factual reasoning abilities. + +# 2.2. Dataset Criteria + +SimpleVQA maintains the following rigorous standards and guidelines for question, image, and answer selection, ensuring objectivity, temporal stability, and verifiability. + +Question Guidelines Clear and Unique Answers: Questions must have a single, undisputed answer. They should precisely define scope (e.g., "Which city?" instead of "Which location?") and specify time references (e.g., "Which year?" rather than "When?"). Evidence-Based: Each question must be supported by verifiable + +
BenchmarkMultimodalData SizeLanguageData SourceDomainFactualityReasoningMetric
MMbench [23]Image&Text2,438Chi.&Eng.Real WorldKnowledge××MCQ Eval
CCBench [23]Image&Text510ChineseKnowledgeKnowledge××MCQ Eval
MME [13]Image&Text1300EnglishReal WorldGeneral××TFQ Eval
MM-Vet [39]Image&Text200EnglishHumanGeneral××LLM-as-a-Judge
Dynamath [45]Image&Text5000EnglishExamsMath××Accuracy
MMMU [40]Image&Text11.5kEnglishExamsGeneral×✓Accuracy
MMMU-Pro [41]Image&Text3460EnglishHuman&GPTGeneral×✓Accuracy
ChineseFactEval [35]Text Only125ChineseHumanKnowledge✓×LLM-as-a-Judge
AGI-Eval [44]Text Only8062Chi.&Eng.ExamsKnowledge××Accuracy
C-Eval [11]Text Only13,948ChineseExamsKnowledge××Accuracy
SimpleQA [31]Text Only4,326EnglishHumanKnowledge✓×LLM-as-a-Judge
Chinese SimpleQA [9]Text Only3,000ChineseHuman&GPTKnowledge✓×LLM-as-a-Judge
SimpleVQA (Ours)Image&Text2,025Chi.&Eng.Human&GPTKnowledge✓✓LLM-as-a-Judge
+ +Table 1. Comparisons between our SimpleVQA and other benchmarks, where "TFQ" means True or False questions, "MCQ" means multi-choice questions, "Chi.& Eng." means Chinese and English. + +![](images/f4921ea471c671e02e65159677490f3b4bfd18e4521c7feb450659072bb6262a.jpg) +Figure 2. An overview of the data construction process of SimpleVQA. + +sources. Manually annotated questions include reference links, while automatically generated ones undergo independent validation by two AI trainers. *Challenging for MLLMs*: Questions are tested on GPT-4o, GPT-4o-mini, doubao-vision-pro, and ERNIE-VL. Only those that at least one model answers incorrectly are retained; others are revised. Answerable by August 2024: All questions must be answerable based on knowledge available before September 1, 2024, ensuring a fair evaluation across models with similar knowledge cutoffs. + +Visual Guidelines No Direct Textual Clues: Images must not contain text revealing the answer. Authenticity: Only real, unaltered images are allowed to prevent factual distortion. Supports Question Reasoning: Each image must provide sufficient context for answering. Manually labeled samples undergo multi-annotator verification. Fixed Before August 2024: Image content must be valid and confirmable before August 2024. + +Answer Guidelines Temporal Stability: Answers must remain unchanged and unaffected by new information. + +Time-sensitive topics (e.g., sports, media) should specify a timeframe rather than a general answer that may change. Sufficiently Challenging: Answers are tested against four high-precision MLLMs. If all models respond correctly, the question is revised to increase difficulty. Fully Objective and Evaluable: Answers must be precise, verifiable, and free from subjective interpretation. Unambiguous: Each answer must have a single, clear meaning to prevent misinterpretation. + +# 2.3. Data Collection and Processing + +Figure 2 shows an overview of the data construction process of SimpleVQA, which adheres to a meticulously designed five-step process, ensuring the highest standards of data quality and reliability. + +Step 1: Seed Example Collection. SimpleVQA's seed examples are sourced from two primary channels. First, we filter images and Q&A pairs from publicly available VQA datasets that align with factual knowledge criteria. We select MMVet (English), MME (English), Dynamath (English), MMbench_CN (Chinese), and CCBench (Chinese) due to their recent construction (post-2024) and their rele + +
StatisticsNumberStatisticsNumber
Data2025Domain Categories9
- Chinsne(CN)1012- Literature, Education & Sports(LES)13.48%
- English(EN)1013- Euro-American History & Culture(EHC)12.89%
Task Categories9- Contemporary Society(CS)11.51%
- Logic & Science(LS)5.04%- Engineering, Technology & Application(ETA)7.95%
- Object Identification Recognition(OIR)14.07%- Film, Television & Media(FTM)10.62%
- Time & Event(TE)9.98%- Natural Science (NS)12.64%
- Person & Emotion(PE)13.58%- Art (AR)7.65%
- Location & Building(LB)21.53%- Chinese History & Culture(CHC)9.68%
- Text Processing(TP)10.07%- Life (LI)13.58%
- Quantity & Position Relationship(QPR)10.12%Word length of Questions
- Art & Culture(AC)9.09%- Maximum question length314
- Object Attributes Recognition(OAR)6.52%- Minimum problem length5
+ +Table 2. Statistics of SimpleVQA + +vance to real-world applications. Second, we collect images and relevant factual knowledge from search engines (e.g., Google, Baidu, Wikipedia), with expert annotators generating corresponding questions and answers. These data focus on entities and events across multiple domains, ensuring answers are objective, fact-based, and centered on entity recognition or attribute extraction. + +Step 2: Data Enhancement and QA Pair Generation. Once sufficient seed examples are gathered, we employ GPT-4o [12] to refine the data and generate Q&A pairs for factual categories. For multiple-choice questions (MCQs) from sources like MMbench-CN and CCBench, to ensure answer uniqueness, we use LLMs to rephrase the original question and introduce qualifiers that precisely align with the correct response. For MME, we extract the answer entity and rewrite the question based on its attributes, ensuring a one-to-one correspondence. Datasets like MMVet, Dynamath, and CCBench, which contain discrepancies from factual Q&A formats (e.g., incorrect answer options, image descriptions, or MCQ distractors), are processed using GPT-4o to align the content with factual reasoning. These refinements produce the initial version of SimpleVQA. + +Step 3: LLM-Based Quality Verification. The refined dataset undergoes verification using GPT-4 to assess adherence to quality standards, ensuring answer stability, uniqueness, and question difficulty. Following LLM screening, two professional annotators conduct rigorous quality checks and refine samples as needed. + +Step 4: Difficulty Screening. To maximize the dataset's utility in model evaluation, we filter out overly simple Q&A pairs. We assess responses from four mainstream MLLMs (GPT-4o, GPT-4o-mini, Doubao-vision-pro, and ERNIE-VL). Any question correctly answered by all four models is deemed too simple and excluded from the dataset, thereby maintaining a challenging benchmark. + +Step 5: Extracting Atomic Facts. To analyze visual comprehension and language alignment in MLLMs more precisely, we generate atomic questions from each SimpleVQA entry. An atomic fact represents the most funda + +mental, indivisible attribute or characteristic of an object. For instance, given the question "In what year was the person in the image born?", the corresponding atomic question is "Who is the person in the image?". MLLMs generate candidate answers, which are then reviewed and refined by professional annotators to ensure accuracy. + +# 2.4. Human Annotation & Quality Control + +To ensure dataset quality, we implement a rigorous manual validation process following automated data collection. Most of the collaborators participated in the necessary data annotation, and we also selected three domain experts from the collaborators. Each question is independently reviewed by two expert annotators to verify factual accuracy. If either annotator finds a question unsuitable, it is discarded. Annotations fact-check answers using authoritative sources such as Wikipedia and Baidu Encyclopedia, providing at least two supporting URLs. If their answers differ, a third expert conducts a final review to ensure consistency and correctness. Only Q&A pairs that fully align with both human evaluations and LLM-generated responses are retained. + +A difficulty assessment further refines the dataset. We begin with 8,360 Q&A pairs, filtering out $22\%$ of image-based samples that lack challenge or fail to meet predefined criteria. 1,108 pairs are removed through multi-model testing to ensure that questions pose a meaningful challenge to MLLMs. To maintain category balance, we carefully select 200 high-difficulty mathematical Q&As from 5,000 Dynamath samples, avoiding an overrepresentation of simpler factual questions. Through multiple validation rounds, we retain 2,025 high-precision Q&A pairs, accounting for $24\%$ of the original dataset. This process ensures factual integrity, topic diversity, and appropriate difficulty levels, making SimpleVQA a robust benchmark for evaluating MLLMs' reasoning and knowledge boundaries. + +# 2.5. Dataset Statistics + +As shown in Table 1 and 2, our SimpleVQA benchmark consists of 2,025 samples across 9 major tasks, 9 major do + +mains, and 244 image types. Examples of each category can be found in supplementary material. This design facilitates a comprehensive assessment of MLLMs across different domains. Regarding the distribution of domain topics and image types in SimpleVQA, 9 main domains are defined and subcategories are assigned based on each topic. Table 2 displays statistics related to SimpleVQA, including the amount of data, categorization and percentage, and the text length of the questions. In Table 1, we also compare SimpleVQA with several mainstream MLLMs' evaluation benchmarks, which suggests that SimpleVQA is the first MLLMs' benchmark that focuses on the evaluation of knowledge boundaries in factual categories. We excluded ideological and politically relevant data from the dataset to prevent social controversies and negative impacts. In addition, we implemented several optimizations to improve the efficiency of the evaluation. The dataset features concise questions and standardized answers, minimizing the input and output markers required for GPT assessment. In addition, all examples are in short-answer question-and-answer (QA) format, and they can be assessed by simple matching. + +# 3. Experiments + +# 3.1. Experimental Setup + +We implement a standardized protocol across all experiments, maintaining a consistent prompt format throughout. All temperature and sampling parameters are configured according to each LLM's official recommendations or default settings. For evaluation and benchmark construction, we employ GPT-4o as the primary model due to its robust performance in multimodal tasks. + +# 3.2. Baseline Models + +We evaluate 18 representative MLLMs in total, comprising 8 closed-source and 10 open-source models, which provide a diverse evaluation of model capabilities across different architectures and training paradigms. The closed-source models include GPT-4o and 4o-mini, Doubao-pro-128k and 32k, Gemini-2.0-flash, Claude-3.5-Sonnet, Qwen-Max, ERNIE-VL. The open-source models cover a wide range of frameworks, including InternVL2.5 and In + +ternVL2, Qwen2.5-VL and Qwen2-VL, Janus-pro-7B. + +# 3.3. Evaluation Metrics + +The SimpleVQA benchmark employs a set of evaluation metrics to assess the accuracy, reliability, and consistency of the model's predictions. These metrics include: (1) Correct (CO) evaluates whether the predicted answer matches the reference answer exactly without any contradiction. (2) Not Attempted (NA) identifies cases where the model does not attempt to answer, ensuring no contradictions are present. (3) Incorrect (IN) flags instances where the predicted answer contradicts the reference answer, even if resolved. (4) Correct Given Attempted (CGA) measures the proportion of correct answers among those attempted by the model, reflecting its performance when engaged. (5) F-score computes the harmonic mean between "Correct" and "Correct Given Attempted", providing a balanced evaluation that combines accuracy and attempt success. + +# 3.4. Main Results + +Results on Overall and Different Tasks Table 3 presents the performance of MLLMs on SimpleVQA, highlighting their F-scores across different tasks. Among the closed-source models, Gemini-2.0-Flash demonstrates superior performance, particularly in tasks such as LS, OIR, TE, PE, LB, TP, and QPR, achieving state-of-the-art (SOTA) scores across these task categories. Claude-3.5-Sonnet, GPT-4o, and ERNIE also exhibit strong performance. Open-source models, including InternVL2.5-78BMPO and Qwen2.5-VL-72B-Instruct, show competitive results, though slightly below the leading closed-source models. Notably, Qwen2.5-VL-72B-Instruct outperforms most mainstream closed-source MLLMs in several aspects. + +Results on Cross-lingual and Different Domains Table 4 lists the performance of various closed-source and open-source MLLMs on SimpleVQA, highlighting their Q&A performance across languages and across different scenarios. Among the closed-source models, Gemini-2.0-Flash and ERNIE-VL show strong Chinese quizzing ability; Gemini-2.0-Flash and Claude-3.5-Sonnet show strong English quizzing ability; especially on the scale evaluation with scenarios as the categorization, the three models mentioned above achieve almost all the SOTA. opensource models InternVL2.5-78B-MPO and Qwen2.5-VL + +
ModelsOverall resultsF-score on 9 task categories
CONAINCGAF-scoreLSOIRTEPELBTPQPRACOAR
Closed-source Multi-modal Large Language Models
GPT-4o47.27.845.051.249.158.753.448.136.444.357.561.835.161.0
GPT-4o-mini35.510.054.539.437.330.044.438.529.636.242.145.925.840.6
Doubao-vision-pro-128k39.720.839.550.144.348.353.553.439.635.038.546.935.962.9
Doubao-vision-pro-32k25.423.651.432.728.324.237.918.031.621.339.232.918.636.6
Gemini-2.0-flash52.86.041.256.154.463.760.954.355.044.361.565.733.964.4
Claude-3.5-Sonnet48.59.841.853.750.957.153.952.129.747.258.462.746.760.1
Qwen-Max25.415.359.330.027.515.133.713.530.127.736.034.412.136.4
ERNIE-VL46.59.544.051.448.849.055.948.340.740.754.459.733.670.8
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO45.45.948.648.346.857.454.650.326.434.457.065.832.872.3
InternVL2.5-78B41.57.750.845.043.249.549.445.028.431.049.663.732.065.2
InternVL2-Llama3-76B35.78.455.938.937.234.743.535.626.228.444.353.529.054.1
InternVL2.5-38B-MPO42.95.451.845.344.051.752.545.722.134.052.067.832.860.4
InternVL2.5-26B-MPO39.97.652.543.241.543.444.547.727.131.745.958.629.068.0
InternVL2.5-8B-MPO33.67.558.936.334.945.137.430.919.126.044.352.226.456.2
Qwen2.5-VL-72B49.45.445.252.250.857.750.851.038.651.657.865.829.862.8
Qwen2-VL-72B-Instruct44.710.345.049.847.148.851.946.337.938.555.363.035.959.7
Qwen2.5-VL-7B-Instruct43.25.151.745.644.338.652.737.632.741.052.653.639.256.7
Janus-pro-7B31.310.658.135.033.027.043.426.523.728.224.050.436.242.3
+ +Table 3. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote "Correct", "Not attempted", "Incorrect", and "Correct given attempted", respectively. We report the scores across different tasks, including "Logic & Science (LS)", "Object Identification Recognition (OIR)", "Time & Event (TE)", "Person & Emotion (PE)", "Location & Building (LB)", "Text Processing (TP)", "Quantity & Position Relationship (QPR)", "Art & Culture (AC)", and "Object Attributes Recognition (OAR)". + +![](images/cbe3542cfd831a11ea3b5c93cd42dbd2f0a72191801e204326f5fc95813e44eb.jpg) +Figure 3. F-score results for eight different models across nine task categories. + +![](images/08b983dde6cf1fe599a6e7cf36c0e18842c62f5acdeba12412b677e5284c2578.jpg) + +![](images/41d280c1a3c62f683c79d835c9eb9c5b20ccdcbc506524a75c71c09fe6512203.jpg) + +![](images/f6bd0d01f88c97d128d233b2ff65e32ee10a39f0f89a9c27841b4655a0646f20.jpg) + +72B-Instruct, respectively, achieved the highest performance, albeit slightly lower than the top closed-source models. It is worthwhile to note that Qwen2.5-VL-72B-Instruct achieved the highest score for the open source model on the English review and was reviewed for strong performance on several domains. + +Confidence Analysis The study of confidence analysis evaluates the consistency between an MLLM's expressed + +confidence and the accuracy of its answers by requiring models to report confidence levels with their responses. As illustrated in Figure 4, two key patterns emerge: first, models demonstrate higher confidence when more reference materials are available, irrespective of architectural differences; second, while higher confidence generally correlates with greater accuracy, we observe numerous instances of high-confidence errors. This finding suggests that MLLMs exhibit overconfidence in their knowledge representation + +
ModelsChinese partial resultsEnglish partial resultsF-score on 9 domains categories
COCGAF-scoreCOCGAF-scoreLESEHCCSETAFTMNSARCHCLI
Closed-source Multi-modal Large Language Models
GPT-4o48.751.650.145.750.748.147.037.558.262.050.561.730.329.658.8
GPT-4o-mini33.036.034.438.243.140.536.025.444.441.541.140.622.926.652.3
Doubao-vision-pro-128k51.156.453.628.541.833.951.823.750.152.249.044.332.644.948.7
Doubao-vision-pro-32k29.037.832.815.626.219.634.913.433.038.229.323.415.229.137.4
Gemini-2.0-flash54.857.956.350.754.352.554.135.459.967.570.964.629.236.364.6
Claude-3.5-Sonnet48.050.649.349.257.052.849.142.752.756.548.161.138.734.466.5
Qwen-Max25.229.427.125.630.627.928.121.035.434.831.620.014.211.944.1
ERNIE-VL54.055.354.640.744.942.755.932.960.054.642.850.727.845.459.4
+ +
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO51.754.753.139.241.840.554.222.560.253.635.556.727.537.963.6
InternVL2.5-78B46.749.247.936.340.638.348.318.655.256.033.053.925.633.757.5
InternVL2-Llama3-76B34.237.435.737.140.438.738.319.044.044.435.445.620.421.457.1
InternVL2.5-38B-MPO48.049.548.837.740.939.252.321.054.259.529.351.527.931.661.6
InternVL2.5-26B-MPO45.547.346.334.438.836.446.920.654.547.632.451.620.534.454.3
InternVL2.5-8B-MPO35.838.537.131.434.032.736.014.043.743.226.446.420.322.553.5
Qwen2.5-VL-72B-Instruct48.050.449.250.754.052.349.445.264.362.453.658.320.930.361.8
Qwen2-VL-72B-Instruct46.148.647.343.351.246.945.230.454.057.251.653.529.429.765.2
Qwen2.5-VL-7B-Instruct41.944.042.944.547.145.842.942.854.949.840.446.425.930.356.9
Janus-pro-7B29.532.130.733.238.135.530.521.540.240.826.237.126.815.353.2
+ +Table 4. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote "Correct", "Not attempted", "Incorrect", and "Correct given attempted", respectively. SimpleVQA is structured around nine key domains: "Literature, education & sports (LES)", "Euro-American History & Culture (EHC)", "Contemporary Society (CS)", "Engineering, Technology & Application (ETA)", "Film", "Television & Media (FTM)", "Natural Science (NS)", "Art (AR)", "Chinese History & Culture (CHC)", and "Life (LI)". + +and comprehension capabilities. + +![](images/2864bd4d699dae78fb9c82d6705346f1ef588c5936c2b1f4d8847151af76d7c5.jpg) +Figure 4. Calibration of LLMs based on their stated confidence. The x-axis represents the confidence level of the LLMs, and the y-axis represents the accuracy. + +Results on Different LLMs To validate the benchmark's reliance on visual information, we conduct experiments with eight mainstream LLMs using text-only inputs (without images). These models include well-known LLMs + +
ModelsCOINNACGA
o1-preview3.65%22.12%74.22%14.16%
o1-mini4.1%19.51%76.4%17.36%
DeepSeek-R110.08%53.6%36.32%15.82%
Qwen-Max7.6%70.77%21.63%9.69%
GPT-4o5.23%31.6%63.16%14.2%
GPT-4o-mini6.47%47.65%45.88%11.95%
Claude-3.5-Sonnet22.02%7.21%90.77%21.88%
Gemini-2.0-flash7.06%61.33%31.6%10.32%
+ +Table 5. The CO, NA, IN, and CGA results for 8 LLMs across simpleVQA without image input. + +such as o1-Preview, o1-mini, and DeepSeek-R1. Table 5 presents the experimental results. We specifically design visual questions that theoretically require image inputs to be answered correctly. Despite the absence of visual information, LLMs achieve minimal accuracy on these tasks, with DeepSeek-R1 demonstrating notably higher performance in inferring correct answers from textual context alone. + +# 4. Further Analysis + +Based on SimpleVQA, we conduct a further evaluation of MLLMs, revealing significant factual accuracy issues. We + +also perform a deep analysis of these factual problems by examining both the visual understanding capabilities and text generation mechanisms of MLLMs, which provides insights for future model optimization. + +First, we identify the three most robust MLLMs through systematic evaluation. For each VQA task where an MLLM provides an incorrect response, we use a carefully designed prompt to decompose the question into an atomic problem focused on basic content recognition. Each atomic problem corresponds to a fundamental atomic fact that, when identified, converts the original question into a purely factual text-based query. When an MLLM fails to answer even these simplified atomic queries correctly, we attribute this failure to insufficient visual understanding of the image. Next, we distinguish between original questions that are already atomic in nature and those that are not. We systematically identify cases where the derived atomic questions differ from the original questions and compile them into a test set called the Complex Fact Question (CFQ) set. This dataset is specifically designed to evaluate whether model performance improves when atomic facts are explicitly provided. In a follow-up experiment, we augment each CFQ query by incorporating the answer to its corresponding atomic question as an explicit hint. We then reassess model performance with these augmented queries. When models still provide incorrect answers despite having access to the relevant atomic facts, we attribute such failures to deficiencies in background knowledge rather than visual understanding. + +Table 6 presents the results of the CFQ experiments. From our complete dataset, we select 570 challenging CFQ examples to evaluate both visual understanding and knowledge internalization capabilities. In addition, another 90 questions are selected from which none MLLMs can answer correctly, to test the performance of mainstream LLMs in deep thinking. We test pure language models (LLMs) including o1-preview, o1-mini, and DeepSeek-R1, as well as multimodal models (MLLMs) such as GPT-4o, Qwen2.5-VL-72B-Instruct, and InternVL2.5-78B-MPO. Our findings reveal several key insights. First, LLMs, despite their reasoning capabilities, fail to effectively utilize atomic facts without accompanying images. Second, MLLMs demonstrate a significant disparity between their visual understanding abilities and knowledge internalization capabilities. While MLLMs show marginally better knowledge retention in relation to visual content, substantial room for improvement remains. Finally, the performance patterns observed when MLLMs answer atomic questions suggest potential for enhancing visual understanding through targeted training approaches. + +# 5. Related Works + +Multimodal Benchmarks. Recent vision-language benchmarks have been developed to assess models' capabilities + +
ModelsOriginAtomicAtomic-GivenFix-Badcase
o1-preview--62.74%13.33% (90)
o1-mini--51.49%10.00% (90)
DeepSeek-R1--55.01%17.78% (90)
Qwen-Max--54.83%5.56% (90)
GPT-4o56.24%56.94%61.69%35.60% (250)
Qwen2.5-VL-72B-Instruct51.67%59.58%64.15%36.25% (276)
InternVL2.5-78B-MPO55.36%55.71%69.95%43.53% (255)
+ +Table 6. CFQ experiment in SimpleVQA. "Origin": original Q&A accuracy; "Atomic": atomic Q&A accuracy; "Atomic-Given": original Q&A with atomic facts; "Fix-Badcase": corrected responses with atomic facts. Sample sizes are in parentheses. + +in integrating visual and textual information across various tasks [14, 32, 33, 42], including OCR [4, 6], spatial awareness [17], multimodal representation learning and retrieval [5, 19, 20], and reasoning skills [24, 26]. For example, MMBench [22] employs multiple-choice tasks in both Chinese and English, covering a wide range of domains. MMMU [40] focuses on complex vision-language tasks, particularly those requiring advanced multimodal reasoning. MMStar [2] utilizes multi-task evaluations to test models' ability to fuse different modalities. + +Factuality Benchmarks. Factuality refers a model's capacity to produce content consistent with factual knowledge, encompassing commonsense, general world knowledge, and domain-specific information. The capability can be evaluated through comparison against authoritative references, including Wikipedia and academic literature. Recently, Various benchmarks have been developed to evaluate factuality in LLMs [11, 16, 21, 27, 28, 36-38, 44]. For example, MMLU [10] assesses multitask accuracy across 57 diverse tasks. HaluEval [18] explores the propensity of LLMs to produce hallucinations or false information. SimpleQA [31] and Chinese SimpleQA [7] have been proposed to measure the short-form factuality in LLMs. + +# 6. Conclusion + +This paper presents SimpleVQA, the first bilingual visual question-answering benchmark for evaluating fact-based reasoning in MLLMs. SimpleVQA features seven key attributes: Chinese-English bilingualism, multitask/multi-scene adaptability, high-quality content, challenging queries, static design, and evaluation efficiency. Our comprehensive evaluation of 18 MLLMs and 8 LLMs demonstrates its effectiveness in assessing fact-based reasoning capabilities. We also introduce the calibration methodology to evaluate MLLMs' visual comprehension and visual-linguistic alignment through atomic question decomposition of factual queries. SimpleVQA establishes a robust benchmark for assessing factuality and advancing the development of reliable MLLMs. + +# 7. Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex & Critical Software Environment (Grant No. SKLCCSE-2024ZX-18). + +# References + +[1] AI@Meta. Llama 3 model card. 2024. 1 +[2] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.8 +[3] Qinyuan Cheng, Tianxiang Sun, Wenwei Zhang, Siyin Wang, Xiangyang Liu, Mozhi Zhang, Junliang He, Mianqiu Huang, Zhangyue Yin, Kai Chen, et al. Evaluating hallucinations in Chinese large language models. arXiv preprint arXiv:2310.03368, 2023. 1 +[4] XianFu Cheng, YanQing Yao, Liying Zhang, Ao Liu, and Zhoujun Li. An improved stochastic gradient descent algorithm based on rényi differential privacy. International Journal of Intelligent Systems, 37(12):10694-10714, 2022. 8 +[5] Xianfu Cheng, Hang Zhang, Jian Yang, Xiang Li, Weixiao Zhou, Kui Wu, Fei Liu, Wei Zhang, Tao Sun, Tongliang Li, et al. Xformparser: A simple and effective multimodal multilingual semi-structured form parser. arXiv preprint arXiv:2405.17336, 2024.8 +[6] Xianfu Cheng, Weixiao Zhou, Xiang Li, Jian Yang, Hang Zhang, Tao Sun, Wei Zhang, Yuying Mai, Tongliang Li, Xiaoming Chen, et al. Sviptr: Fast and efficient scene text recognition with vision permutable extractor. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 365-373, 2024. 8 +[7] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, Zhuoran Lin, Xuepeng Liu, Dekai Sun, Shirong Lin, Zhicheng Zheng, Xiaoyong Zhu, Wenbo Su, and Bo Zheng. Chinese simpleqa: A Chinese factuality evaluation for large language models, 2024. 8 +[8] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 1 +[9] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 3 +[10] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021. 8 + +[11] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. arXiv preprint arXiv:2305.08322, 2023. 3, 8 +[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4 +[13] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3 +[14] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 8 +[15] Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv: 2411.16594, 2024. 2 +[16] Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. Cmmlu: Measuring massive multitask language understanding in Chinese, 2023. 8 +[17] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 8 +[18] Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. 8 +[19] Xiang Li, Xianfu Cheng, Dezhuang Miao, Xiaoming Zhang, and Zhoujun Li. Tf-mamba: Text-enhanced fusion mamba with missing modalities for robust multimodal sentiment analysis. arXiv preprint arXiv:2505.14329, 2025. 8 +[20] Xiang Li, Hajjun Zhang, Zhiqiang Dong, Xianfu Cheng, Yun Liu, and Xiaoming Zhang. Learning fine-grained representation with token-level alignment for multimodal sentiment analysis. Expert Systems with Applications, 269:126274, 2025. 8 +[21] Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland, 2022. Association for Computational Linguistics. 8 +[22] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8 + +[23] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3 +[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 8 +[25] OpenAI. Gpt-4 technical report. PREPRINT, 2023. 1 +[26] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 8 +[27] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 8 +[28] Yingshui Tan, Boren Zheng, Baihui Zheng, Kerui Cao, Huiyun Jing, Jincheng Wei, Jiaheng Liu, Yancheng He, Wenbo Su, Xiangyong Zhu, and Bo Zheng. Chinese safetyqa: A safety short-form factuality benchmark for large language models, 2024. 8 +[29] SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024. 1 +[30] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. URL https://apisemantic scholar.org/CorpusID, 273877483. 1 +[31] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. 3, 8 +[32] Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kajing Ma, Chenghao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, et al. Scimmir: Benchmarking scientific multi-modal information retrieval. arXiv preprint arXiv:2401.13478, 2024. 8 +[33] Siwei Wu, Kang Zhu, Yu Bai, Yiming Liang, Yizhi Li, Haoning Wu, Jiaheng Liu, Ruibo Liu, Xingwei Qu, Xuxin Cheng, et al. Mmra: A benchmark for multi-granularity multi-image relational association. arXiv preprint arXiv:2407.17379, 2024.8 +[34] Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et al. Tablebench: A comprehensive and complex benchmark for table question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 25497-25506, 2025. 2 +[35] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, + +et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 3 +[36] Jian Yang, Jiaxi Yang, Ke Jin, Yibo Miao, Lei Zhang, Liquan Yang, Zeyu Cui, Yichang Zhang, Binyuan Hui, and Junyang Lin. Evaluating and aligning codellms on human preference. arXiv preprint arXiv:2412.05210, 2024. 8 +[37] Jian Yang, Jiajun Zhang, Jiaxi Yang, Ke Jin, Lei Zhang, Qiyao Peng, Ken Deng, Yibo Miao, Tianyu Liu, Zeyu Cui, et al. Excrepobench: Multi-level executable code completion evaluation. arXiv preprint arXiv:2412.11990, 2024. +[38] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 8 +[39] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3 +[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3, 8 +[41] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3 +[42] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, Wenhu Chen, and Jie Fu. Cmmm: A chinese massive multidiscipline multimodal understanding benchmark. ArXiv, abs/2401.11944, 2024. 8 +[43] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1 +[44] Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models, 2023. 3, 8 +[45] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 3 \ No newline at end of file diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/images.zip b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af5d98d583836c00328d0f4a7979f1d4c8cbc889 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c1b7ebb3dbc7ad7cac21b6e222b64630e84284545f4bdc545ecf75afc741460 +size 864226 diff --git a/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/layout.json b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..807472947458392dade8a395a20f921060825497 --- /dev/null +++ b/2025/SimpleVQA_ Multimodal Factuality Evaluation for Multimodal Large Language Models/layout.json @@ -0,0 +1,6048 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 102, + 501, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 102, + 501, + 139 + ], + "spans": [ + { + "bbox": [ + 109, + 102, + 501, + 139 + ], + "type": "text", + "content": "SimpleVQA: Multimodal Factuality Evaluation for Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": "Xianfu Cheng" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Wei Zhang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Shiwei Zhang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Xiangyuan Guan" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Xianjie Wu" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Xiang Li" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Ge Zhang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Jiaheng Liu" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Yuying Mai" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Yutao Zeng" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Zhoufutu Wen" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Ke Jin" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Baorui Wang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Weixiao Zhou" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Yunhong Lu" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Hangyuan Ji" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Tongliang Li" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Wenhao Huang" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", Zhoujun Li" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": "Beihang University; " + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": "Baidu Inc., China; " + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": "M-A-P; " + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 77, + 160, + 533, + 258 + ], + "type": "text", + "content": "Shenzhen Intelligent Strong Technology Co., Ltd. {buaacxf, lizj} @ buaa.edu.cn, zhangshiwei05@baidu.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "spans": [ + { + "bbox": [ + 152, + 299, + 200, + 312 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 324, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 601 + ], + "type": "text", + "content": "The increasing application of multi-modal large language models (MLLMs) across various sectors has spotlighted the essence of their output reliability and accuracy, particularly their ability to produce content grounded in factual information (e.g. common and domain-specific knowledge). In this work, we introduce SimpleVQA, the first comprehensive multi-modal benchmark to evaluate the factuality ability of MLLMs to answer natural language short questions. SimpleVQA is characterized by 7 key features: it is based on bilingual, it covers multiple tasks and multiple scenarios, ensures high quality and challenging queries, maintains static and timeless reference answers, and is straightforward to evaluate. Our approach involves categorizing visual question-answering items into 9 different tasks around objective events or common knowledge and situating these within 9 scenario domains. Rigorous quality control processes are implemented to guarantee high-quality, concise, and clear answers, facilitating evaluation with minimal variance via an LLM-as-a-judge scoring system. Using SimpleVQA, we perform a comprehensive assessment of leading 18 MLLMs and 8 text-only LLMs, delving into their image comprehension and text generation abilities by identifying and analyzing error cases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 625, + 135, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 135, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 135, + 638 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 682 + ], + "type": "text", + "content": "A significant challenge in large language models (LLMs) is ensuring that LLMs [1, 25] generate factually accurate and evidence-based responses. Current state-of-the-art (SOTA)" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 335, + 286, + 535, + 477 + ], + "blocks": [ + { + "bbox": [ + 335, + 286, + 535, + 477 + ], + "lines": [ + { + "bbox": [ + 335, + 286, + 535, + 477 + ], + "spans": [ + { + "bbox": [ + 335, + 286, + 535, + 477 + ], + "type": "image", + "image_path": "1f61626ebd2b36af0ed3c9cb4cea56ac0ba065e59bd05249dec00a580c3babc3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 487, + 534, + 498 + ], + "lines": [ + { + "bbox": [ + 334, + 487, + 534, + 498 + ], + "spans": [ + { + "bbox": [ + 334, + 487, + 534, + 498 + ], + "type": "text", + "content": "Figure 1. An example from our proposed SimpleVQA." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 570 + ], + "type": "text", + "content": "LLMs often produce outputs that are misleading or unsupported by evidence phenomenon known as \"hallucinations\" [3, 29, 43]. The generation of incorrect or unsubstantiated information remains a major barrier to the broader adoption and reliability of general-purpose AI technologies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "OpenAI proposes SimpleQA [30] to measure factuality simple and reliable with nearly 4K concise and fact-seeking questions. Further, Chinese SimpleQA [8] comprised of 3K Chinese questions spanning 6 major topics is proposed to target the Chinese language. However, the SimpleQA and Chinese SimpleQA benchmarks mainly evaluate the model capabilities in text modality, ignoring wider real-world scenarios (e.g. vision modality). For the vision modality, the research progress of the multi-modal large language models (MLLMs) is still hindered by the \"hallucinations\" introduced by the given images. Therefore, The community of MLLMs has an urgent need for how to measure the simple" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 693, + 148, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 148, + 703 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 148, + 703 + ], + "type": "text", + "content": "* Equal Contributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 70, + 703, + 157, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 157, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 157, + 712 + ], + "type": "text", + "content": "† Corresponding Authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4637" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 247, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 247, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 247, + 84 + ], + "type": "text", + "content": "and reliable factuality introduced by the image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 445 + ], + "type": "text", + "content": "To address this limitation, we develop the SimpleVQA benchmark as shown in Figure 1, where we define the factual question-answering capability of the visual language model. For the proposed factual Visual Question Answering (VQA), we collect 2,025 high-quality question-answer pairs covering 9 different scenario domains across 9 different application tasks. As a factual benchmark for a short answer, SimpleVQA has the following advantages: (1) English and Chinese: SimpleVQA provides general knowledge visual Q&A in both English and Chinese backgrounds, and comprehensively assesses the fact-generating capacity of MLLMs in Chinese and English communities. (2) Multi-task division: We divide the SimpleVQA assessment set into 9 different forms of VQA tasks according to the collected questions and different needs of pictures. (3) Diversified scenarios: SimpleVQA divided the VQA tasks into 9 uniquely representative scenario domains based on the category of the image and the source of the collected images. (4) High quality: We implement a comprehensive and rigorous quality control process to ensure the quality of questions and the accuracy of answers at SimpleVQA. (5) Challenge: simpleVQA focuses on factual questions that mainstream MLLMs cannot answer accurately, and cannot trace the cause of errors through the model itself. (6) Static answers: Following SimpleQA's factual definition, all the standard answers provided in our benchmark don't change over time. (7) Easy to evaluate: SimpleVQA's short answers make it possible to use existing LLMs to run a judge program (called LLM-as-a-Judge [15, 34]) to quickly determine right or wrong and get an overall accuracy rate." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 448, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 675 + ], + "type": "text", + "content": "We systematically evaluate 18 MLLMs on SimpleVQA and create a dynamic leaderboard to show results. Further, a series of probing experiments are performed to explore the effect of the key factors for SimpleVQA. We classify the capabilities possessed by MLLMs for factual questions into two aspects, visual understanding and internalized knowledge capabilities: (1) visual understanding refers to the ability of the model to identify the subject of the question being asked in the question; and (2) internalized knowledge capabilities test whether the model has already mastered the relevant knowledge of the subject of the question being asked, and thus is able to answer the relevant question correctly after identifying that subject. Based on this definition, we added an abductive reasoning experiment to the basic assessment to help determine whether the badcase came from a lack of visual understanding ability or a lack of internalized knowledge ability by generating and labeling atomic questions (each atomic question corresponds to an atomic fact) for each VQA example." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "The remarkable findings from SimpleVQA are summarized as: (1) The factual accuracy of most evaluation models in the field of visual question-answering is insufficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "(2) The training data of MLLMs contains knowledge errors and they are overconfident in what they generate. (3) Image content understanding is still a major challenge for MLLMs to achieve improved capabilities. (4) The ability of MLLMs to internalize massive world knowledge still needs to be improved, and overcoming illusions remains a great challenge for large language models. (5) Further aligning visual core information with the knowledge stored in LLMs is also key to improving the factuality of MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 191, + 392, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 392, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 392, + 205 + ], + "type": "text", + "content": "2. SimpleVQA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 211, + 382, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 382, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 382, + 223 + ], + "type": "text", + "content": "2.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "type": "text", + "content": "The SimpleVQA benchmark consists of 2,025 samples spanning 9 core tasks and 9 primary domains, with each question-image pair categorized into relevant subcategories, enabling a comprehensive evaluation of MLLMs across diverse knowledge areas. The dataset 9 tasks, including covers Logic & Science (LS), Object Identification Recognition (OIR), Time & Event (TE), Person & Emotion (PE), Location & Building (LB), Text Processing (TP), Quantity & Position Relationship (QPR), Art & Culture (AC), and Object Attributes Recognition (OAR). To ensure broad topic coverage, SimpleVQA is structured around 9 key domains: Literature, education & sports (LES), Euro-American History & Culture (EHC), Contemporary Society (CS), Engineering, Technology & Application (ETA), Film, Television & Media (FTM), Natural Science (NS), Art (AR), Chinese History & Culture (CHC), and Life (LI)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 421, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 564 + ], + "type": "text", + "content": "As shown in Table 1, SimpleVQA differs from existing MLLM benchmarks by focusing on factual knowledge boundaries instead of general vision-language understanding. Politically sensitive and ideological content is excluded to maintain neutrality and avoid controversy. Designed for efficiency, the dataset features concise questions and standardized answers, reducing complexity in model evaluation. All samples follow a short-answer Q&A format, enabling simple and objective assessment through direct answer matching. These refinements ensure SimpleVQA serves as a robust benchmark for evaluating MLLMs' factual reasoning abilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 572, + 412, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 412, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 412, + 583 + ], + "type": "text", + "content": "2.2. Dataset Criteria" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 590, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 628 + ], + "type": "text", + "content": "SimpleVQA maintains the following rigorous standards and guidelines for question, image, and answer selection, ensuring objectivity, temporal stability, and verifiability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Question Guidelines Clear and Unique Answers: Questions must have a single, undisputed answer. They should precisely define scope (e.g., \"Which city?\" instead of \"Which location?\") and specify time references (e.g., \"Which year?\" rather than \"When?\"). Evidence-Based: Each question must be supported by verifiable" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4638" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 542, + 220 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 542, + 220 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 542, + 220 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 542, + 220 + ], + "type": "table", + "html": "
BenchmarkMultimodalData SizeLanguageData SourceDomainFactualityReasoningMetric
MMbench [23]Image&Text2,438Chi.&Eng.Real WorldKnowledge××MCQ Eval
CCBench [23]Image&Text510ChineseKnowledgeKnowledge××MCQ Eval
MME [13]Image&Text1300EnglishReal WorldGeneral××TFQ Eval
MM-Vet [39]Image&Text200EnglishHumanGeneral××LLM-as-a-Judge
Dynamath [45]Image&Text5000EnglishExamsMath××Accuracy
MMMU [40]Image&Text11.5kEnglishExamsGeneral×✓Accuracy
MMMU-Pro [41]Image&Text3460EnglishHuman&GPTGeneral×✓Accuracy
ChineseFactEval [35]Text Only125ChineseHumanKnowledge✓×LLM-as-a-Judge
AGI-Eval [44]Text Only8062Chi.&Eng.ExamsKnowledge××Accuracy
C-Eval [11]Text Only13,948ChineseExamsKnowledge××Accuracy
SimpleQA [31]Text Only4,326EnglishHumanKnowledge✓×LLM-as-a-Judge
Chinese SimpleQA [9]Text Only3,000ChineseHuman&GPTKnowledge✓×LLM-as-a-Judge
SimpleVQA (Ours)Image&Text2,025Chi.&Eng.Human&GPTKnowledge✓✓LLM-as-a-Judge
", + "image_path": "fd9421bb8aecc531862e5809f037e64e73eac3fdc98dfebd6cce61ba4ff8640f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "type": "text", + "content": "Table 1. Comparisons between our SimpleVQA and other benchmarks, where \"TFQ\" means True or False questions, \"MCQ\" means multi-choice questions, \"Chi.& Eng.\" means Chinese and English." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 261, + 542, + 415 + ], + "blocks": [ + { + "bbox": [ + 70, + 261, + 542, + 415 + ], + "lines": [ + { + "bbox": [ + 70, + 261, + 542, + 415 + ], + "spans": [ + { + "bbox": [ + 70, + 261, + 542, + 415 + ], + "type": "image", + "image_path": "f4921ea471c671e02e65159677490f3b4bfd18e4521c7feb450659072bb6262a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 424, + 435, + 435 + ], + "lines": [ + { + "bbox": [ + 175, + 424, + 435, + 435 + ], + "spans": [ + { + "bbox": [ + 175, + 424, + 435, + 435 + ], + "type": "text", + "content": "Figure 2. An overview of the data construction process of SimpleVQA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 446, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 446, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 446, + 295, + 567 + ], + "type": "text", + "content": "sources. Manually annotated questions include reference links, while automatically generated ones undergo independent validation by two AI trainers. *Challenging for MLLMs*: Questions are tested on GPT-4o, GPT-4o-mini, doubao-vision-pro, and ERNIE-VL. Only those that at least one model answers incorrectly are retained; others are revised. Answerable by August 2024: All questions must be answerable based on knowledge available before September 1, 2024, ensuring a fair evaluation across models with similar knowledge cutoffs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 579, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 579, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 579, + 295, + 675 + ], + "type": "text", + "content": "Visual Guidelines No Direct Textual Clues: Images must not contain text revealing the answer. Authenticity: Only real, unaltered images are allowed to prevent factual distortion. Supports Question Reasoning: Each image must provide sufficient context for answering. Manually labeled samples undergo multi-annotator verification. Fixed Before August 2024: Image content must be valid and confirmable before August 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Answer Guidelines Temporal Stability: Answers must remain unchanged and unaffected by new information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 446, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 555 + ], + "type": "text", + "content": "Time-sensitive topics (e.g., sports, media) should specify a timeframe rather than a general answer that may change. Sufficiently Challenging: Answers are tested against four high-precision MLLMs. If all models respond correctly, the question is revised to increase difficulty. Fully Objective and Evaluable: Answers must be precise, verifiable, and free from subjective interpretation. Unambiguous: Each answer must have a single, clear meaning to prevent misinterpretation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 563, + 484, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 484, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 484, + 577 + ], + "type": "text", + "content": "2.3. Data Collection and Processing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 581, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 629 + ], + "type": "text", + "content": "Figure 2 shows an overview of the data construction process of SimpleVQA, which adheres to a meticulously designed five-step process, ensuring the highest standards of data quality and reliability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Step 1: Seed Example Collection. SimpleVQA's seed examples are sourced from two primary channels. First, we filter images and Q&A pairs from publicly available VQA datasets that align with factual knowledge criteria. We select MMVet (English), MME (English), Dynamath (English), MMbench_CN (Chinese), and CCBench (Chinese) due to their recent construction (post-2024) and their rele" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4639" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 121, + 70, + 491, + 208 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 491, + 208 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 491, + 208 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 491, + 208 + ], + "type": "table", + "html": "
StatisticsNumberStatisticsNumber
Data2025Domain Categories9
- Chinsne(CN)1012- Literature, Education & Sports(LES)13.48%
- English(EN)1013- Euro-American History & Culture(EHC)12.89%
Task Categories9- Contemporary Society(CS)11.51%
- Logic & Science(LS)5.04%- Engineering, Technology & Application(ETA)7.95%
- Object Identification Recognition(OIR)14.07%- Film, Television & Media(FTM)10.62%
- Time & Event(TE)9.98%- Natural Science (NS)12.64%
- Person & Emotion(PE)13.58%- Art (AR)7.65%
- Location & Building(LB)21.53%- Chinese History & Culture(CHC)9.68%
- Text Processing(TP)10.07%- Life (LI)13.58%
- Quantity & Position Relationship(QPR)10.12%Word length of Questions
- Art & Culture(AC)9.09%- Maximum question length314
- Object Attributes Recognition(OAR)6.52%- Minimum problem length5
", + "image_path": "26e4c00b037160baf5e3e726920f54967a517ffee0bd512f20c6f99ed3259ba7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 242, + 214, + 367, + 226 + ], + "lines": [ + { + "bbox": [ + 242, + 214, + 367, + 226 + ], + "spans": [ + { + "bbox": [ + 242, + 214, + 367, + 226 + ], + "type": "text", + "content": "Table 2. Statistics of SimpleVQA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 238, + 294, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 294, + 321 + ], + "type": "text", + "content": "vance to real-world applications. Second, we collect images and relevant factual knowledge from search engines (e.g., Google, Baidu, Wikipedia), with expert annotators generating corresponding questions and answers. These data focus on entities and events across multiple domains, ensuring answers are objective, fact-based, and centered on entity recognition or attribute extraction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 503 + ], + "type": "text", + "content": "Step 2: Data Enhancement and QA Pair Generation. Once sufficient seed examples are gathered, we employ GPT-4o [12] to refine the data and generate Q&A pairs for factual categories. For multiple-choice questions (MCQs) from sources like MMbench-CN and CCBench, to ensure answer uniqueness, we use LLMs to rephrase the original question and introduce qualifiers that precisely align with the correct response. For MME, we extract the answer entity and rewrite the question based on its attributes, ensuring a one-to-one correspondence. Datasets like MMVet, Dynamath, and CCBench, which contain discrepancies from factual Q&A formats (e.g., incorrect answer options, image descriptions, or MCQ distractors), are processed using GPT-4o to align the content with factual reasoning. These refinements produce the initial version of SimpleVQA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 578 + ], + "type": "text", + "content": "Step 3: LLM-Based Quality Verification. The refined dataset undergoes verification using GPT-4 to assess adherence to quality standards, ensuring answer stability, uniqueness, and question difficulty. Following LLM screening, two professional annotators conduct rigorous quality checks and refine samples as needed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 579, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 664 + ], + "type": "text", + "content": "Step 4: Difficulty Screening. To maximize the dataset's utility in model evaluation, we filter out overly simple Q&A pairs. We assess responses from four mainstream MLLMs (GPT-4o, GPT-4o-mini, Doubao-vision-pro, and ERNIE-VL). Any question correctly answered by all four models is deemed too simple and excluded from the dataset, thereby maintaining a challenging benchmark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Step 5: Extracting Atomic Facts. To analyze visual comprehension and language alignment in MLLMs more precisely, we generate atomic questions from each SimpleVQA entry. An atomic fact represents the most funda" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 238, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 553, + 310 + ], + "type": "text", + "content": "mental, indivisible attribute or characteristic of an object. For instance, given the question \"In what year was the person in the image born?\", the corresponding atomic question is \"Who is the person in the image?\". MLLMs generate candidate answers, which are then reviewed and refined by professional annotators to ensure accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 319, + 517, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 517, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 517, + 332 + ], + "type": "text", + "content": "2.4. Human Annotation & Quality Control" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 338, + 555, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 493 + ], + "type": "text", + "content": "To ensure dataset quality, we implement a rigorous manual validation process following automated data collection. Most of the collaborators participated in the necessary data annotation, and we also selected three domain experts from the collaborators. Each question is independently reviewed by two expert annotators to verify factual accuracy. If either annotator finds a question unsuitable, it is discarded. Annotations fact-check answers using authoritative sources such as Wikipedia and Baidu Encyclopedia, providing at least two supporting URLs. If their answers differ, a third expert conducts a final review to ensure consistency and correctness. Only Q&A pairs that fully align with both human evaluations and LLM-generated responses are retained." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "text", + "content": "A difficulty assessment further refines the dataset. We begin with 8,360 Q&A pairs, filtering out " + }, + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "text", + "content": " of image-based samples that lack challenge or fail to meet predefined criteria. 1,108 pairs are removed through multi-model testing to ensure that questions pose a meaningful challenge to MLLMs. To maintain category balance, we carefully select 200 high-difficulty mathematical Q&As from 5,000 Dynamath samples, avoiding an overrepresentation of simpler factual questions. Through multiple validation rounds, we retain 2,025 high-precision Q&A pairs, accounting for " + }, + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 494, + 556, + 662 + ], + "type": "text", + "content": " of the original dataset. This process ensures factual integrity, topic diversity, and appropriate difficulty levels, making SimpleVQA a robust benchmark for evaluating MLLMs' reasoning and knowledge boundaries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 671, + 417, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 417, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 417, + 682 + ], + "type": "text", + "content": "2.5. Dataset Statistics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "As shown in Table 1 and 2, our SimpleVQA benchmark consists of 2,025 samples across 9 major tasks, 9 major do" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": "mains, and 244 image types. Examples of each category can be found in supplementary material. This design facilitates a comprehensive assessment of MLLMs across different domains. Regarding the distribution of domain topics and image types in SimpleVQA, 9 main domains are defined and subcategories are assigned based on each topic. Table 2 displays statistics related to SimpleVQA, including the amount of data, categorization and percentage, and the text length of the questions. In Table 1, we also compare SimpleVQA with several mainstream MLLMs' evaluation benchmarks, which suggests that SimpleVQA is the first MLLMs' benchmark that focuses on the evaluation of knowledge boundaries in factual categories. We excluded ideological and politically relevant data from the dataset to prevent social controversies and negative impacts. In addition, we implemented several optimizations to improve the efficiency of the evaluation. The dataset features concise questions and standardized answers, minimizing the input and output markers required for GPT assessment. In addition, all examples are in short-answer question-and-answer (QA) format, and they can be assessed by simple matching." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 136, + 357 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 171, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 171, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 171, + 378 + ], + "type": "text", + "content": "3.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "content": "We implement a standardized protocol across all experiments, maintaining a consistent prompt format throughout. All temperature and sampling parameters are configured according to each LLM's official recommendations or default settings. For evaluation and benchmark construction, we employ GPT-4o as the primary model due to its robust performance in multimodal tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 154, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 154, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 154, + 496 + ], + "type": "text", + "content": "3.2. Baseline Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 601 + ], + "type": "text", + "content": "We evaluate 18 representative MLLMs in total, comprising 8 closed-source and 10 open-source models, which provide a diverse evaluation of model capabilities across different architectures and training paradigms. The closed-source models include GPT-4o and 4o-mini, Doubao-pro-128k and 32k, Gemini-2.0-flash, Claude-3.5-Sonnet, Qwen-Max, ERNIE-VL. The open-source models cover a wide range of frameworks, including InternVL2.5 and In" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": "ternVL2, Qwen2.5-VL and Qwen2-VL, Janus-pro-7B." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 92, + 427, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 427, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 427, + 105 + ], + "type": "text", + "content": "3.3. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 110, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 291 + ], + "type": "text", + "content": "The SimpleVQA benchmark employs a set of evaluation metrics to assess the accuracy, reliability, and consistency of the model's predictions. These metrics include: (1) Correct (CO) evaluates whether the predicted answer matches the reference answer exactly without any contradiction. (2) Not Attempted (NA) identifies cases where the model does not attempt to answer, ensuring no contradictions are present. (3) Incorrect (IN) flags instances where the predicted answer contradicts the reference answer, even if resolved. (4) Correct Given Attempted (CGA) measures the proportion of correct answers among those attempted by the model, reflecting its performance when engaged. (5) F-score computes the harmonic mean between \"Correct\" and \"Correct Given Attempted\", providing a balanced evaluation that combines accuracy and attempt success." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 298, + 399, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 399, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 399, + 309 + ], + "type": "text", + "content": "3.4. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 556, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 473 + ], + "type": "text", + "content": "Results on Overall and Different Tasks Table 3 presents the performance of MLLMs on SimpleVQA, highlighting their F-scores across different tasks. Among the closed-source models, Gemini-2.0-Flash demonstrates superior performance, particularly in tasks such as LS, OIR, TE, PE, LB, TP, and QPR, achieving state-of-the-art (SOTA) scores across these task categories. Claude-3.5-Sonnet, GPT-4o, and ERNIE also exhibit strong performance. Open-source models, including InternVL2.5-78BMPO and Qwen2.5-VL-72B-Instruct, show competitive results, though slightly below the leading closed-source models. Notably, Qwen2.5-VL-72B-Instruct outperforms most mainstream closed-source MLLMs in several aspects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 487, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 620 + ], + "type": "text", + "content": "Results on Cross-lingual and Different Domains Table 4 lists the performance of various closed-source and open-source MLLMs on SimpleVQA, highlighting their Q&A performance across languages and across different scenarios. Among the closed-source models, Gemini-2.0-Flash and ERNIE-VL show strong Chinese quizzing ability; Gemini-2.0-Flash and Claude-3.5-Sonnet show strong English quizzing ability; especially on the scale evaluation with scenarios as the categorization, the three models mentioned above achieve almost all the SOTA. opensource models InternVL2.5-78B-MPO and Qwen2.5-VL" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "https://platform.openai.com/docs/models#gpt-4o; https://platform.openai.com/docs/models#gpt-4o-mini \nhttps://www.volcengine.com/product/ark \nhttps://deepmind.google/technologies/gemini/flash/ \nhttps://www.anthropic.com/news/claude-3-5-sonnet \nhttps://huggingface.co/spaces/Qwen/Qwen-Max \nhttps://yiyan.baidu.com/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 627, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 627, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 627, + 555, + 712 + ], + "type": "text", + "content": "https://huggingface.co/OpenGVLab/InternVL2_5-78B-MPO; https://huggingface.co/OpenGVLab/InternVL2-Llama3-76B \nhttps://huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct; https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct; https://huggingface.co/Qwen/Qwen2-VL-72B-Instruct \nhttps://huggingface.co/deepseek-ai/Janus-Pro-7B" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4641" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 340 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 340 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 340 + ], + "type": "table", + "html": "
ModelsOverall resultsF-score on 9 task categories
CONAINCGAF-scoreLSOIRTEPELBTPQPRACOAR
Closed-source Multi-modal Large Language Models
GPT-4o47.27.845.051.249.158.753.448.136.444.357.561.835.161.0
GPT-4o-mini35.510.054.539.437.330.044.438.529.636.242.145.925.840.6
Doubao-vision-pro-128k39.720.839.550.144.348.353.553.439.635.038.546.935.962.9
Doubao-vision-pro-32k25.423.651.432.728.324.237.918.031.621.339.232.918.636.6
Gemini-2.0-flash52.86.041.256.154.463.760.954.355.044.361.565.733.964.4
Claude-3.5-Sonnet48.59.841.853.750.957.153.952.129.747.258.462.746.760.1
Qwen-Max25.415.359.330.027.515.133.713.530.127.736.034.412.136.4
ERNIE-VL46.59.544.051.448.849.055.948.340.740.754.459.733.670.8
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO45.45.948.648.346.857.454.650.326.434.457.065.832.872.3
InternVL2.5-78B41.57.750.845.043.249.549.445.028.431.049.663.732.065.2
InternVL2-Llama3-76B35.78.455.938.937.234.743.535.626.228.444.353.529.054.1
InternVL2.5-38B-MPO42.95.451.845.344.051.752.545.722.134.052.067.832.860.4
InternVL2.5-26B-MPO39.97.652.543.241.543.444.547.727.131.745.958.629.068.0
InternVL2.5-8B-MPO33.67.558.936.334.945.137.430.919.126.044.352.226.456.2
Qwen2.5-VL-72B49.45.445.252.250.857.750.851.038.651.657.865.829.862.8
Qwen2-VL-72B-Instruct44.710.345.049.847.148.851.946.337.938.555.363.035.959.7
Qwen2.5-VL-7B-Instruct43.25.151.745.644.338.652.737.632.741.052.653.639.256.7
Janus-pro-7B31.310.658.135.033.027.043.426.523.728.224.050.436.242.3
", + "image_path": "fc4c03ba13ec9a9aafa871809bbf52083a05527e351d2240aa570539f4599adb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 349, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 555, + 394 + ], + "type": "text", + "content": "Table 3. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. We report the scores across different tasks, including \"Logic & Science (LS)\", \"Object Identification Recognition (OIR)\", \"Time & Event (TE)\", \"Person & Emotion (PE)\", \"Location & Building (LB)\", \"Text Processing (TP)\", \"Quantity & Position Relationship (QPR)\", \"Art & Culture (AC)\", and \"Object Attributes Recognition (OAR)\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 404, + 159, + 571 + ], + "blocks": [ + { + "bbox": [ + 58, + 404, + 159, + 571 + ], + "lines": [ + { + "bbox": [ + 58, + 404, + 159, + 571 + ], + "spans": [ + { + "bbox": [ + 58, + 404, + 159, + 571 + ], + "type": "image", + "image_path": "cbe3542cfd831a11ea3b5c93cd42dbd2f0a72191801e204326f5fc95813e44eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 579, + 447, + 590 + ], + "lines": [ + { + "bbox": [ + 162, + 579, + 447, + 590 + ], + "spans": [ + { + "bbox": [ + 162, + 579, + 447, + 590 + ], + "type": "text", + "content": "Figure 3. F-score results for eight different models across nine task categories." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 189, + 404, + 289, + 571 + ], + "blocks": [ + { + "bbox": [ + 189, + 404, + 289, + 571 + ], + "lines": [ + { + "bbox": [ + 189, + 404, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 189, + 404, + 289, + 571 + ], + "type": "image", + "image_path": "08b983dde6cf1fe599a6e7cf36c0e18842c62f5acdeba12412b677e5284c2578.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 320, + 404, + 421, + 570 + ], + "blocks": [ + { + "bbox": [ + 320, + 404, + 421, + 570 + ], + "lines": [ + { + "bbox": [ + 320, + 404, + 421, + 570 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 421, + 570 + ], + "type": "image", + "image_path": "41d280c1a3c62f683c79d835c9eb9c5b20ccdcbc506524a75c71c09fe6512203.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 452, + 404, + 552, + 570 + ], + "blocks": [ + { + "bbox": [ + 452, + 404, + 552, + 570 + ], + "lines": [ + { + "bbox": [ + 452, + 404, + 552, + 570 + ], + "spans": [ + { + "bbox": [ + 452, + 404, + 552, + 570 + ], + "type": "image", + "image_path": "f6bd0d01f88c97d128d233b2ff65e32ee10a39f0f89a9c27841b4655a0646f20.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 297, + 673 + ], + "type": "text", + "content": "72B-Instruct, respectively, achieved the highest performance, albeit slightly lower than the top closed-source models. It is worthwhile to note that Qwen2.5-VL-72B-Instruct achieved the highest score for the open source model on the English review and was reviewed for strong performance on several domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Confidence Analysis The study of confidence analysis evaluates the consistency between an MLLM's expressed" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 601, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 711 + ], + "type": "text", + "content": "confidence and the accuracy of its answers by requiring models to report confidence levels with their responses. As illustrated in Figure 4, two key patterns emerge: first, models demonstrate higher confidence when more reference materials are available, irrespective of architectural differences; second, while higher confidence generally correlates with greater accuracy, we observe numerous instances of high-confidence errors. This finding suggests that MLLMs exhibit overconfidence in their knowledge representation" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4642" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 192 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 192 + ], + "type": "table", + "html": "
ModelsChinese partial resultsEnglish partial resultsF-score on 9 domains categories
COCGAF-scoreCOCGAF-scoreLESEHCCSETAFTMNSARCHCLI
Closed-source Multi-modal Large Language Models
GPT-4o48.751.650.145.750.748.147.037.558.262.050.561.730.329.658.8
GPT-4o-mini33.036.034.438.243.140.536.025.444.441.541.140.622.926.652.3
Doubao-vision-pro-128k51.156.453.628.541.833.951.823.750.152.249.044.332.644.948.7
Doubao-vision-pro-32k29.037.832.815.626.219.634.913.433.038.229.323.415.229.137.4
Gemini-2.0-flash54.857.956.350.754.352.554.135.459.967.570.964.629.236.364.6
Claude-3.5-Sonnet48.050.649.349.257.052.849.142.752.756.548.161.138.734.466.5
Qwen-Max25.229.427.125.630.627.928.121.035.434.831.620.014.211.944.1
ERNIE-VL54.055.354.640.744.942.755.932.960.054.642.850.727.845.459.4
", + "image_path": "88b4a2daeb4e68c99f5692b25712cc60f28982b28c62d60e1982122135358ab4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 193, + 555, + 315 + ], + "blocks": [ + { + "bbox": [ + 58, + 193, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 58, + 193, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 58, + 193, + 555, + 315 + ], + "type": "table", + "html": "
Open-source Multi-modal Large Language Models
InternVL2.5-78B-MPO51.754.753.139.241.840.554.222.560.253.635.556.727.537.963.6
InternVL2.5-78B46.749.247.936.340.638.348.318.655.256.033.053.925.633.757.5
InternVL2-Llama3-76B34.237.435.737.140.438.738.319.044.044.435.445.620.421.457.1
InternVL2.5-38B-MPO48.049.548.837.740.939.252.321.054.259.529.351.527.931.661.6
InternVL2.5-26B-MPO45.547.346.334.438.836.446.920.654.547.632.451.620.534.454.3
InternVL2.5-8B-MPO35.838.537.131.434.032.736.014.043.743.226.446.420.322.553.5
Qwen2.5-VL-72B-Instruct48.050.449.250.754.052.349.445.264.362.453.658.320.930.361.8
Qwen2-VL-72B-Instruct46.148.647.343.351.246.945.230.454.057.251.653.529.429.765.2
Qwen2.5-VL-7B-Instruct41.944.042.944.547.145.842.942.854.949.840.446.425.930.356.9
Janus-pro-7B29.532.130.733.238.135.530.521.540.240.826.237.126.815.353.2
", + "image_path": "b716aae26cc2cfaf17cd1eae6b2a8e0861f449b3326649d673b95eb15f6c067b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 323, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 369 + ], + "type": "text", + "content": "Table 4. Results of different models on SimpleVQA. For metrics, CO, NA, IN, and CGA denote \"Correct\", \"Not attempted\", \"Incorrect\", and \"Correct given attempted\", respectively. SimpleVQA is structured around nine key domains: \"Literature, education & sports (LES)\", \"Euro-American History & Culture (EHC)\", \"Contemporary Society (CS)\", \"Engineering, Technology & Application (ETA)\", \"Film\", \"Television & Media (FTM)\", \"Natural Science (NS)\", \"Art (AR)\", \"Chinese History & Culture (CHC)\", and \"Life (LI)\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 389, + 187, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 187, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 187, + 401 + ], + "type": "text", + "content": "and comprehension capabilities." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 415, + 290, + 603 + ], + "blocks": [ + { + "bbox": [ + 62, + 415, + 290, + 603 + ], + "lines": [ + { + "bbox": [ + 62, + 415, + 290, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 290, + 603 + ], + "type": "image", + "image_path": "2864bd4d699dae78fb9c82d6705346f1ef588c5936c2b1f4d8847151af76d7c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "type": "text", + "content": "Figure 4. Calibration of LLMs based on their stated confidence. The x-axis represents the confidence level of the LLMs, and the y-axis represents the accuracy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Results on Different LLMs To validate the benchmark's reliance on visual information, we conduct experiments with eight mainstream LLMs using text-only inputs (without images). These models include well-known LLMs" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 387, + 547, + 488 + ], + "blocks": [ + { + "bbox": [ + 323, + 387, + 547, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 387, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 387, + 547, + 488 + ], + "type": "table", + "html": "
ModelsCOINNACGA
o1-preview3.65%22.12%74.22%14.16%
o1-mini4.1%19.51%76.4%17.36%
DeepSeek-R110.08%53.6%36.32%15.82%
Qwen-Max7.6%70.77%21.63%9.69%
GPT-4o5.23%31.6%63.16%14.2%
GPT-4o-mini6.47%47.65%45.88%11.95%
Claude-3.5-Sonnet22.02%7.21%90.77%21.88%
Gemini-2.0-flash7.06%61.33%31.6%10.32%
", + "image_path": "01e87b0d754b457bfc9e558c45c262d5b86fd20f5ae85b67857ba33f176d31af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 519 + ], + "type": "text", + "content": "Table 5. The CO, NA, IN, and CGA results for 8 LLMs across simpleVQA without image input." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 613 + ], + "type": "text", + "content": "such as o1-Preview, o1-mini, and DeepSeek-R1. Table 5 presents the experimental results. We specifically design visual questions that theoretically require image inputs to be answered correctly. Despite the absence of visual information, LLMs achieve minimal accuracy on these tasks, with DeepSeek-R1 demonstrating notably higher performance in inferring correct answers from textual context alone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 623, + 417, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 417, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 417, + 636 + ], + "type": "text", + "content": "4. Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 643, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 669 + ], + "type": "text", + "content": "Based on SimpleVQA, we conduct a further evaluation of MLLMs, revealing significant factual accuracy issues. We" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 674, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 554, + 713 + ], + "type": "text", + "content": "https://platform.openai.com/docs/models#ol \nhttps://platform.openai.com/docs/models#ol-mini \nhttps://huggingface.co/deepseek-ai/DeepSeek-R1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4643" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 119 + ], + "type": "text", + "content": "also perform a deep analysis of these factual problems by examining both the visual understanding capabilities and text generation mechanisms of MLLMs, which provides insights for future model optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 121, + 294, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 121, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 57, + 121, + 294, + 406 + ], + "type": "text", + "content": "First, we identify the three most robust MLLMs through systematic evaluation. For each VQA task where an MLLM provides an incorrect response, we use a carefully designed prompt to decompose the question into an atomic problem focused on basic content recognition. Each atomic problem corresponds to a fundamental atomic fact that, when identified, converts the original question into a purely factual text-based query. When an MLLM fails to answer even these simplified atomic queries correctly, we attribute this failure to insufficient visual understanding of the image. Next, we distinguish between original questions that are already atomic in nature and those that are not. We systematically identify cases where the derived atomic questions differ from the original questions and compile them into a test set called the Complex Fact Question (CFQ) set. This dataset is specifically designed to evaluate whether model performance improves when atomic facts are explicitly provided. In a follow-up experiment, we augment each CFQ query by incorporating the answer to its corresponding atomic question as an explicit hint. We then reassess model performance with these augmented queries. When models still provide incorrect answers despite having access to the relevant atomic facts, we attribute such failures to deficiencies in background knowledge rather than visual understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 407, + 294, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 407, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 407, + 294, + 658 + ], + "type": "text", + "content": "Table 6 presents the results of the CFQ experiments. From our complete dataset, we select 570 challenging CFQ examples to evaluate both visual understanding and knowledge internalization capabilities. In addition, another 90 questions are selected from which none MLLMs can answer correctly, to test the performance of mainstream LLMs in deep thinking. We test pure language models (LLMs) including o1-preview, o1-mini, and DeepSeek-R1, as well as multimodal models (MLLMs) such as GPT-4o, Qwen2.5-VL-72B-Instruct, and InternVL2.5-78B-MPO. Our findings reveal several key insights. First, LLMs, despite their reasoning capabilities, fail to effectively utilize atomic facts without accompanying images. Second, MLLMs demonstrate a significant disparity between their visual understanding abilities and knowledge internalization capabilities. While MLLMs show marginally better knowledge retention in relation to visual content, substantial room for improvement remains. Finally, the performance patterns observed when MLLMs answer atomic questions suggest potential for enhancing visual understanding through targeted training approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 669, + 146, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 146, + 681 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 146, + 681 + ], + "type": "text", + "content": "5. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 689, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 689, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 689, + 294, + 712 + ], + "type": "text", + "content": "Multimodal Benchmarks. Recent vision-language benchmarks have been developed to assess models' capabilities" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 149 + ], + "type": "table", + "html": "
ModelsOriginAtomicAtomic-GivenFix-Badcase
o1-preview--62.74%13.33% (90)
o1-mini--51.49%10.00% (90)
DeepSeek-R1--55.01%17.78% (90)
Qwen-Max--54.83%5.56% (90)
GPT-4o56.24%56.94%61.69%35.60% (250)
Qwen2.5-VL-72B-Instruct51.67%59.58%64.15%36.25% (276)
InternVL2.5-78B-MPO55.36%55.71%69.95%43.53% (255)
", + "image_path": "b3657a43b8760ecc62f17a14498967c78096207b972cf6e7d5383375153f853a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 159, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 317, + 159, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 553, + 201 + ], + "type": "text", + "content": "Table 6. CFQ experiment in SimpleVQA. \"Origin\": original Q&A accuracy; \"Atomic\": atomic Q&A accuracy; \"Atomic-Given\": original Q&A with atomic facts; \"Fix-Badcase\": corrected responses with atomic facts. Sample sizes are in parentheses." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 317, + 221, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 221, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 553, + 340 + ], + "type": "text", + "content": "in integrating visual and textual information across various tasks [14, 32, 33, 42], including OCR [4, 6], spatial awareness [17], multimodal representation learning and retrieval [5, 19, 20], and reasoning skills [24, 26]. For example, MMBench [22] employs multiple-choice tasks in both Chinese and English, covering a wide range of domains. MMMU [40] focuses on complex vision-language tasks, particularly those requiring advanced multimodal reasoning. MMStar [2] utilizes multi-task evaluations to test models' ability to fuse different modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 345, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 553, + 498 + ], + "type": "text", + "content": "Factuality Benchmarks. Factuality refers a model's capacity to produce content consistent with factual knowledge, encompassing commonsense, general world knowledge, and domain-specific information. The capability can be evaluated through comparison against authoritative references, including Wikipedia and academic literature. Recently, Various benchmarks have been developed to evaluate factuality in LLMs [11, 16, 21, 27, 28, 36-38, 44]. For example, MMLU [10] assesses multitask accuracy across 57 diverse tasks. HaluEval [18] explores the propensity of LLMs to produce hallucinations or false information. SimpleQA [31] and Chinese SimpleQA [7] have been proposed to measure the short-form factuality in LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 522, + 386, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 386, + 534 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 386, + 534 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 546, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 712 + ], + "type": "text", + "content": "This paper presents SimpleVQA, the first bilingual visual question-answering benchmark for evaluating fact-based reasoning in MLLMs. SimpleVQA features seven key attributes: Chinese-English bilingualism, multitask/multi-scene adaptability, high-quality content, challenging queries, static design, and evaluation efficiency. Our comprehensive evaluation of 18 MLLMs and 8 LLMs demonstrates its effectiveness in assessing fact-based reasoning capabilities. We also introduce the calibration methodology to evaluate MLLMs' visual comprehension and visual-linguistic alignment through atomic question decomposition of factual queries. SimpleVQA establishes a robust benchmark for assessing factuality and advancing the development of reliable MLLMs." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "4644" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 62276017, 62406033, U1636211, 61672081), and the State Key Laboratory of Complex & Critical Software Environment (Grant No. SKLCCSE-2024ZX-18)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 184, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 225, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 225, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 225, + 194 + ], + "type": "text", + "content": "[1] AI@Meta. Llama 3 model card. 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 196, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 295, + 251 + ], + "type": "text", + "content": "[2] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 253, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 295, + 308 + ], + "type": "text", + "content": "[3] Qinyuan Cheng, Tianxiang Sun, Wenwei Zhang, Siyin Wang, Xiangyang Liu, Mozhi Zhang, Junliang He, Mianqiu Huang, Zhangyue Yin, Kai Chen, et al. Evaluating hallucinations in Chinese large language models. arXiv preprint arXiv:2310.03368, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 309, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 295, + 353 + ], + "type": "text", + "content": "[4] XianFu Cheng, YanQing Yao, Liying Zhang, Ao Liu, and Zhoujun Li. An improved stochastic gradient descent algorithm based on rényi differential privacy. International Journal of Intelligent Systems, 37(12):10694-10714, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 355, + 295, + 408 + ], + "type": "text", + "content": "[5] Xianfu Cheng, Hang Zhang, Jian Yang, Xiang Li, Weixiao Zhou, Kui Wu, Fei Liu, Wei Zhang, Tao Sun, Tongliang Li, et al. Xformparser: A simple and effective multimodal multilingual semi-structured form parser. arXiv preprint arXiv:2405.17336, 2024.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 477 + ], + "type": "text", + "content": "[6] Xianfu Cheng, Weixiao Zhou, Xiang Li, Jian Yang, Hang Zhang, Tao Sun, Wei Zhang, Yuying Mai, Tongliang Li, Xiaoming Chen, et al. Sviptr: Fast and efficient scene text recognition with vision permutable extractor. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 365-373, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "type": "text", + "content": "[7] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, Zhuoran Lin, Xuepeng Liu, Dekai Sun, Shirong Lin, Zhicheng Zheng, Xiaoyong Zhu, Wenbo Su, and Bo Zheng. Chinese simpleqa: A Chinese factuality evaluation for large language models, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 601 + ], + "type": "text", + "content": "[8] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 656 + ], + "type": "text", + "content": "[9] Yancheng He, Shilong Li, Jiaheng Liu, Yingshui Tan, Weixun Wang, Hui Huang, Xingyuan Bu, Hangyu Guo, Chengwei Hu, Boren Zheng, et al. Chinese simpleqa: A chinese factuality evaluation for large language models. arXiv preprint arXiv:2411.07140, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "text", + "content": "[10] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[11] Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Jiayi Lei, Yao Fu, Maosong Sun, and Junxian He. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. arXiv preprint arXiv:2305.08322, 2023. 3, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[12] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "text", + "content": "[13] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 240, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 553, + 293 + ], + "type": "text", + "content": "[14] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Naturalbench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "text", + "content": "[15] Dawei Li, Bohan Jiang, Liangjie Huang, Alimohammad Beigi, Chengshuai Zhao, Zhen Tan, Amrita Bhattacharjee, Yuxuan Jiang, Canyu Chen, Tianhao Wu, Kai Shu, Lu Cheng, and Huan Liu. From generation to judgment: Opportunities and challenges of llm-as-a-judge. arXiv preprint arXiv: 2411.16594, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 403 + ], + "type": "text", + "content": "[16] Haonan Li, Yixuan Zhang, Fajri Koto, Yifei Yang, Hai Zhao, Yeyun Gong, Nan Duan, and Timothy Baldwin. Cmmlu: Measuring massive multitask language understanding in Chinese, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "type": "text", + "content": "[17] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 492 + ], + "type": "text", + "content": "[18] Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large-scale hallucination evaluation benchmark for large language models, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 536 + ], + "type": "text", + "content": "[19] Xiang Li, Xianfu Cheng, Dezhuang Miao, Xiaoming Zhang, and Zhoujun Li. Tf-mamba: Text-enhanced fusion mamba with missing modalities for robust multimodal sentiment analysis. arXiv preprint arXiv:2505.14329, 2025. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 538, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 538, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 538, + 553, + 591 + ], + "type": "text", + "content": "[20] Xiang Li, Hajjun Zhang, Zhiqiang Dong, Xianfu Cheng, Yun Liu, and Xiaoming Zhang. Learning fine-grained representation with token-level alignment for multimodal sentiment analysis. Expert Systems with Applications, 269:126274, 2025. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "text", + "content": "[21] Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA: Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland, 2022. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[22] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4645" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[23] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 271, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 271, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 271, + 194 + ], + "type": "text", + "content": "[25] OpenAI. Gpt-4 technical report. PREPRINT, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 250 + ], + "type": "text", + "content": "[26] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146–162. Springer, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "type": "text", + "content": "[27] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on Machine Learning Research, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[28] Yingshui Tan, Boren Zheng, Baihui Zheng, Kerui Cao, Huiyun Jing, Jincheng Wei, Jiaheng Liu, Yancheng He, Wenbo Su, Xiangyong Zhu, and Bo Zheng. Chinese safetyqa: A safety short-form factuality benchmark for large language models, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[29] SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava Das. A comprehensive survey of hallucination mitigation techniques in large language models. arXiv preprint arXiv:2401.01313, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[30] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. URL https://apisemantic scholar.org/CorpusID, 273877483. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[31] Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. Measuring short-form factuality in large language models. 2024. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[32] Siwei Wu, Yizhi Li, Kang Zhu, Ge Zhang, Yiming Liang, Kajing Ma, Chenghao Xiao, Haoran Zhang, Bohao Yang, Wenhu Chen, et al. Scimmir: Benchmarking scientific multi-modal information retrieval. arXiv preprint arXiv:2401.13478, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[33] Siwei Wu, Kang Zhu, Yu Bai, Yiming Liang, Yizhi Li, Haoning Wu, Jiaheng Liu, Ruibo Liu, Xingwei Qu, Xuxin Cheng, et al. Mmra: A benchmark for multi-granularity multi-image relational association. arXiv preprint arXiv:2407.17379, 2024.8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 689 + ], + "type": "text", + "content": "[34] Xianjie Wu, Jian Yang, Linzheng Chai, Ge Zhang, Jiaheng Liu, Xeron Du, Di Liang, Daixin Shu, Xianfu Cheng, Tianzhen Sun, et al. Tablebench: A comprehensive and complex benchmark for table question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 25497-25506, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 652 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[36] Jian Yang, Jiaxi Yang, Ke Jin, Yibo Miao, Lei Zhang, Liquan Yang, Zeyu Cui, Yichang Zhang, Binyuan Hui, and Junyang Lin. Evaluating and aligning codellms on human preference. arXiv preprint arXiv:2412.05210, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[37] Jian Yang, Jiajun Zhang, Jiaxi Yang, Ke Jin, Lei Zhang, Qiyao Peng, Ken Deng, Yibo Miao, Tianyu Liu, Zeyu Cui, et al. Excrepobench: Multi-level executable code completion evaluation. arXiv preprint arXiv:2412.11990, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[38] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[39] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 352 + ], + "type": "text", + "content": "[40] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 407 + ], + "type": "text", + "content": "[41] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, et al. Mmmu-pro: A more robust multidiscipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 496 + ], + "type": "text", + "content": "[42] Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, Chenghua Lin, Wenhao Huang, Wenhu Chen, and Jie Fu. Cmmm: A chinese massive multidiscipline multimodal understanding benchmark. ArXiv, abs/2401.11944, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 552 + ], + "type": "text", + "content": "[43] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[44] Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. Agieval: A human-centric benchmark for evaluating foundation models, 2023. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 553, + 652 + ], + "type": "text", + "content": "[45] Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836, 2024. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4646" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_content_list.json b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..151b533a94bfdbf7793316bddda9e260a06af69c --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_content_list.json @@ -0,0 +1,1419 @@ +[ + { + "type": "text", + "text": "Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation", + "text_level": 1, + "bbox": [ + 138, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fengchen He, Dayang Zhao, Hao Xu, Tingwei Quan, Shaoqun Zeng Huazhong University of Science and Technology", + "bbox": [ + 227, + 180, + 767, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{linyark, dayangzhao, xuhao_2003, quantingwei, sqzeng}@hust.edu.cn", + "bbox": [ + 204, + 219, + 787, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many studies utilize dual-pixel (DP) sensor phase information for various applications, such as depth estimation and deblurring. However, since DP image features are entirely determined by the camera hardware, DP-depth paired datasets are very scarce, especially when performing depth estimation on customized cameras. To overcome this, studies simulate DP images using ideal optical models. However, these simulations often violate real optical propagation laws, leading to poor generalization to real DP data. To address this, we investigate the domain gap between simulated and real DP data, and propose solutions using the Simulating DP Images from Ray Tracing (Sdirt) scheme. Sdirt generates realistic DP images via ray tracing and integrates them into the depth estimation training pipeline. Experimental results show that models trained with Sdirt-simulated images generalize better to real DP data. The code and collected datasets will be available at https://github.com/LinYark/Sdirt.", + "bbox": [ + 88, + 301, + 485, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 602, + 220, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The DP sensor [15, 29] is designed to split each pixel into left and right sub-pixels, utilizing microlenses and subpixels to achieve phase splitting. This allows the capture of a pair of images in one shot, known as DP images, as shown in Fig. 1(a). DP images are not only used for autofocus [7, 11, 31], but also improve performance in tasks such as deblurring [4, 17, 41] and depth estimation [9, 13, 23]. In the field of depth-from-dual-pixel (DfDP) estimation, the scarcity of DP-depth paired datasets stems from hardware limitations. To address this, many studies [2, 16, 18, 23, 26, 37] have focused on simulating DP images using RGBD datasets, which serve as more flexible and accessible data sources.", + "bbox": [ + 89, + 628, + 482, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key to simulating DP images from RGBD datasets lies in accurately modeling the DP point spread function (PSF). In recent years, numerous model-based and calibration-based DP PSF simulators have been proposed. Calibration-based simulators [16, 18, 37] require substantial", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4421f6660aaa09a6026144f2cea6a0a00bc2cbc8836700851d369ae1e0f08e8f.jpg", + "image_caption": [ + "Figure 1. (a) Imaging process of a DP camera. The slight shifts between the left and right DP images caused by phase differences are illustrated with white dashed lines. (b) Example comparison between real and CoC-simulated DP PSFs, showing a significant difference between them." + ], + "image_footnote": [], + "bbox": [ + 522, + 271, + 906, + 505 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time calibrating real cameras to match DP PSFs, but suffer from issues such as interpolation errors due to discrete calibration points and the neglect of lens and DP sensor parameters, limiting model transferability. Model-based simulators [2, 23, 26] employ ideal optical or aberration analysis models to directly compute DP PSFs, bypassing calibration. However, these models often perform poorly on real DP images because they ignore lens aberrations and sensor phase splitting characteristics. As shown in Fig. 1(b), the ideal thin-lens model's simulated circle-of-confusion (CoC) DP PSF exhibits a significant domain gap with the real PSF, adversely affecting depth estimation.", + "bbox": [ + 511, + 609, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge the domain gap between real DP images and DP images simulated by model-based simulators, we propose Sdirt. Sdirt consists of a ray-traced DP PSF simulator and a pixel-wise DP image rendering module, which can accurately generate DP images with aberration and phase information. Specifically, given a fixed focus lens and DP sensor system with known parameters, we use ray trac", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26106", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [35, 39, 40] to calculate the spatially varying DP PSFs and train a multilayer perceptron (MLP) to predict them. Subsequently, we convolve per-pixel DP PSFs with the all-in-focus (AiF) image to simulate the DP images captured by a real camera. After training a DfDP model with DP images simulated by Sdirt, the model can accurately estimate depth using optical aberration and phase information, enhancing its ability to generalize to real DP images.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to validate the effectiveness of Sdirt. The experimental results demonstrate that, compared to other model-based simulators, the DP PSFs and DP images generated by Sdirt better resemble real data, and the DfDP model trained on Sdirt achieves superior generalization. In summary, our contributions are fourfold:", + "bbox": [ + 89, + 212, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a ray-traced DP PSF simulator that computes spatially varying DP PSFs, addressing the domain gap between simulated and real DP PSFs caused by lens aberrations and sensor phase splitting.", + "- We propose a pixel-wise DP image rendering module that uses an MLP to predict the DP PSF for each pixel, narrowing the gap between simulated and real DP images.", + "- Depth estimation results show that the DfDP model trained on Sdirt generalizes better to real DP images.", + "- We collected DP119, a real DP-depth paired test set with an open lens structure and fixed focus, featuring diverse real-world scenes." + ], + "bbox": [ + 89, + 304, + 480, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 500, + 232, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. DP Datasets", + "text_level": 1, + "bbox": [ + 89, + 526, + 220, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most professional and mobile cameras have DP sensors, but only Canon cameras [1, 25, 26] and Google Pixel phones [9, 34, 43] offer DP data. In the field of depth estimation, supervised tasks rely on DP-depth paired datasets [1, 9, 26, 43]. However, changes in the lens, sensor, focal length, or aperture settings alter the PSFs, making it necessary to reacquire DP-depth data. Consequently, many researchers focus on simulating DP-depth datasets using RGBD data to reduce data collection costs and improve flexibility.", + "bbox": [ + 89, + 547, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. DP PSF simulators", + "text_level": 1, + "bbox": [ + 89, + 696, + 272, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Simulating DP images from RGBD datasets relies on DP PSFs, with recent progress in calibration-based [16, 18, 37] and model-based [2, 23, 26] DP PSF simulators. In calibration-based simulators, Xin et al. [37] acquired full-space DP PSFs by sampling discrete points in space and interpolating, while Li et al. [18] acquired them through size scaling using the CoC model. Furthermore, Li et al. [16] generated DP PSFs using a U-Net, which requires a large amount of DP-depth paired data. Calibration-based simulators are time-consuming and prone to interpolation and scaling errors, while model-based simulators circumvent these issues. Pan et al. [23] simplified the optical system to an", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ideal thin lens and used a symmetrically divided rectangular aperture to calculate DP PSFs. Punnappurath et al. [26] modeled the DP PSFs using a depletion kernel that is phase symmetric, meaning the left and right PSFs of the same object point are flip symmetric. Abuolaim et al. [2] used aberration analysis models to calculate DP PSFs, but the resulting patterns still remained phase symmetric. These studies oversimplify optical propagation by neglecting aberrations and phase splitting, leading to unrealistic domain gaps.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Applications driven by DP data", + "text_level": 1, + "bbox": [ + 511, + 234, + 790, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, DP data-driven applications such as depth estimation [9, 10, 13, 14, 23, 24, 43], deblurring [3, 4, 17, 27, 38, 41, 42], and refocusing [5, 7, 11, 32] have undergone rapid development. Additionally, Kang et al. [13] utilized DP data for facial scanning, similar to depth estimation tasks. Wadhwa et al. [34] employed DP data for synthetic shallow depth-of-field imaging, a popular recent application in mobile photography. Shi et al. [28] integrated DP hardware with structured diffractive optical elements for complex image encoding and high-precision multimodal reconstruction. In addition, DP data can be used for disparity estimation [22, 36], rain removal [18], and reflection removal [25]. It is foreseeable that DP data-driven applications will continue to emerge in various new fields.", + "bbox": [ + 511, + 257, + 906, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method of Sdirt", + "text_level": 1, + "bbox": [ + 511, + 481, + 671, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To provide the required input for training the DfDP network, Sdirt generates realistic DP images through ray tracing. As shown in Fig. 2(c), the module achieves pixel-wise rendering of DP images by convolving the DP PSF corresponding to each point in the depth map with the AiF RGB image. The DP PSF for each point is inferred by a pretrained MLP network. As illustrated in Fig. 2(b), the offline training process of the MLP network uses ray-traced DP PSFs (Fig. 2(a)) as ground truth (GT).", + "bbox": [ + 511, + 507, + 905, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Ray-traced DP PSF simulator", + "text_level": 1, + "bbox": [ + 511, + 652, + 779, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 2(a), lens ray tracing can accurately obtain the landing point on the sensor surface [8, 35, 39, 40]. We construct a ray set $A$ containing $n$ rays, each originating from an object point $p$ . The aperture stop's image in object space is treated as the entrance pupil, from which $n$ points are densely sampled. Each ray in $A$ is initially directed from $p$ to its corresponding entrance pupil sample. As rays in $A$ traverse each lens surface, their positions and directions are updated by refraction according to Snell's law and the lens parameters. After multiple refractions, the rays reach the sensor plane. We denote the sensor-plane landing points as $O$ and the corresponding ray directions as $D$ .", + "bbox": [ + 511, + 674, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DP sensor ray tracing is limited because camera manufacturers do not disclose the structural parameters of the microlens and sub-pixel components within the DP pixel.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93ab78c423775cc9cdd65314e09bd144dcac0bf9c2a6709b6aeed6c296cad114.jpg", + "image_caption": [ + "(a)", + "Figure 2. Simulating Dual-Pixel Images from Ray Tracing pipeline. (a) Ray-traced DP PSF simulator. Calculates spatially varying DP PSFs for lens and DP sensor through ray tracing. (b) DP PSF prediction network. Trains an MLP network to predict DP PSFs, using the ray-traced DP PSFs as GT. (c) Pixel-wise DP image rendering module. The network predicts the DP PSFs for all points in the depth map (red pass). Then, each DP PSF is convolved with the AiF RGB image to render the simulated DP image (blue pass)." + ], + "image_footnote": [], + "bbox": [ + 127, + 92, + 906, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a result, we simplify the DP pixel structure (Fig. 3(a)) based on past research [15, 29]. We model the microlens as a thin lens with radius $r$ and focal length $f$ , and define $h$ as distance between the sub-pixel and the microlens, $w$ as sub-pixel width, and $ps$ as DP pixel size.", + "bbox": [ + 89, + 510, + 482, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The DP PSF can be obtained by separately calculating the cumulative integration of the rays for each sub-pixel. We analyze which sub-pixel the $k$ -th ray $A_{k}$ in the ray set $A$ ultimately enters. Through lens ray tracing, we obtain the landing point $O_{k}(x_{k},y_{k},z_{k})$ and direction $D_{k}(x_{k}^{\\prime},y_{k}^{\\prime},z_{k}^{\\prime})$ of the $k$ -th ray $A_{k}$ on the sensor surface. Based on the landing point $O_{k}$ , we can easily calculate that $A_{k}$ lies within the DP pixel $(i,j)$ , which has surface center coordinates $(x_{i},y_{i})$ . For $O_{k}$ , the situation differs depending on whether it lies within or outside the microlens:", + "bbox": [ + 89, + 592, + 482, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When $O_{k}$ lies within the microlens, as shown in Fig. 3(b), $A_{k}$ is refracted by the microlens and then enters a sub-pixel. According to geometric optics theory, $A_{k}$ and the principal ray incident on the thin lens at the same angle both converge on the focal plane of the thin lens. We only need to focus on a few boundary lines $x_{L1}, x_{M1}, x_{R1}$ to determine which sub-pixel $A_{k}$ ultimately enters. If $x_{k}$ is in the interval $[x_{L1}, x_{M1}]$ , it enters the left sub-pixel; if in the interval $[x_{M1}, x_{R1}]$ , it enters the right sub-pixel; otherwise, it is considered a missing ray. The calculation methods for the", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "boundary lines are as follows:", + "bbox": [ + 511, + 510, + 712, + 525 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {L 1} = x _ {i} + w - (f * \\tan \\theta - w) * h / (f - h)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 532, + 864, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {M 1} = x _ {i} - (f * \\tan \\theta) * h / (f - h) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 551, + 903, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {R 1} = x _ {i} - w - (f * \\tan \\theta + w) * h / (f - h)\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 570, + 864, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\tan \\theta$ is the tangent of the angle at which $A_{k}$ lands on the sensor surface, equivalent to $x_{k}^{\\prime} / z_{k}^{\\prime}$ . We provide the derivation hints for Eq. (1) in Fig. 4(a).", + "bbox": [ + 511, + 593, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When $O_{k}$ lies outside the microlens, as shown in Fig. 3(c), $A_{k}$ enters the sub-pixel directly without refraction. Similarly, checking boundary lines $x_{L2}, x_{M2}, x_{R2}$ suffices to determine which sub-pixel $A_{k}$ ultimately enters:", + "bbox": [ + 511, + 638, + 903, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {L 2} = x _ {i} + w - h * \\tan \\theta\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 705, + 802, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {M 2} = x _ {i} - h * \\tan \\theta \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 724, + 903, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {R 2} = x _ {i} - w - h * \\tan \\theta\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 744, + 800, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The left PSF $(PSF_{L})$ can be calculated by integrating the set of rays $A$ emitted from point $p$ over all left subpixels:", + "bbox": [ + 511, + 766, + 906, + 810 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP S F _ {L} (i, j) = \\sum_ {k = 1} ^ {n} A _ {k} * \\delta_ {L, k} (i, j) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 810, + 903, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_{L,k}(i,j)$ represents the energy distribution of ray $A_{k}$ in the left sub-pixel at DP pixel $(i,j)$ . The ray energy distribution is assumed to be a unit impulse, which", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26108", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ae6b7be3ee796de1178f432f1afcb6e403ff915397ef5d0386f7b604cf2c2d5c.jpg", + "image_caption": [ + "Figure 3. (a) DP pixel structure layout. The left part is a right-side view, and the right part is a front perspective view. (b) On hitting the microlens, the $k$ -th ray lands on left/right sub-pixel in red/blue interval, else missed. (c) Without hitting the microlens, the $k$ -th ray lands on left/right sub-pixel in red/blue interval, else missed." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 444, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "equals zero unless it eventually enters the left sub-pixel of DP pixel $(i,j)$ . By following the same calculation steps as for $PSF_{L}$ , $PSF_{R}$ can also be computed.", + "bbox": [ + 89, + 553, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pixel-wise DP image rendering module", + "text_level": 1, + "bbox": [ + 89, + 607, + 424, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2(b), we reduce the computational cost of ray-traced DP PSFs by training an MLP network to predict them, inspired by [33, 39]. The network input is the normalized coordinate $p$ within the camera's valid imaging region (Fig. 4(b)), which is a frustum defined by the field of view, sensor size, and the preset minimum and maximum depths. We normalize the $(x,y)$ coordinates to $[-1,1]$ . After fixing the focus distance, we set the DP PSF kernel size $(ks)$ large enough to ensure that the DP PSF for any point within the valid imaging region can be fully displayed. The network consists of 5 hidden layers, each containing 512 neurons, and an output layer with $2 * ks^2$ neurons. We adopt the L2 loss to supervise predictions $P\\hat{S} F_L$ and $P\\hat{S} F_R$ :", + "bbox": [ + 89, + 628, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = L _ {2} (P \\hat {S} F _ {L}, P S F _ {L}) + L _ {2} (P \\hat {S} F _ {R}, P S F _ {R}) \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 830, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we use ray-traced DP PSFs as GT and apply max normalization to alleviate learning difficulties caused by large radii PSFs. During inference, we apply", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0310068b5819416d0a4c7d18005ee98a792dfcec67246f15baf43eb512f6c8e.jpg", + "image_caption": [ + "Figure 4. (a) Derivation hints for Eq. (1). (b) The valid imaging region is a frustum, and we normalize the $(x,y)$ coordinates to $[-1,1]$ . (c) The DfDP model takes DP image as input to predict the depth map. (d) During cost volume generation in [6], we stack original disparity (green arrows) and add reverse disparity (blue arrows), with $d_{max}$ as max displacement." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 898, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sum normalization to the predicted DP PSFs, approximating the uniform intensity distribution in cameras with vignetting compensation.", + "bbox": [ + 511, + 450, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Because the DP PSF is shift-variant, we convolve each pixel's DP PSF with the AiF image to simulate the DP image captured by a camera. The RGBD dataset [21, 30] provides the paired RGB image $(I_{RGB})$ and depth map $(I_D)$ . As shown in Fig. 2(c), we treat each pixel in $I_D$ as an object point and use the trained MLP to predict its DP PSF. Using $I_{RGB}$ as the AiF image, we apply pixel-wise local convolution [39] to render the DP image containing aberration and phase information.", + "bbox": [ + 511, + 497, + 905, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. DfDP Model Based on Sdtirt", + "text_level": 1, + "bbox": [ + 511, + 647, + 769, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4(c), the input to the DfDP model is provided in real time by the pixel-wise DP image rendering module of Sdirt. At each training iteration, Sdirt feeds the DfDP model with simulated DP images of shape $(B,6,H,W)$ , where $B$ is the batch size, the number of channels is 6, and $H, W$ are the DP image height and width. The model outputs the predicted depth maps $(\\hat{I}_D)$ of shape $(B,1,H,W)$ . We apply the L1 loss to supervise $\\hat{I}_D$ :", + "bbox": [ + 511, + 674, + 906, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = L _ {1} \\left(\\hat {I _ {D}}, I _ {D}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 800, + 903, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We select [6] as the DfDP model and make reasonable adjustments to its cost volume step to accommodate the DP data. In binocular images, disparity from points at different depths always has the same direction, thus the cost volume generation process in [6] is also unidirectional, as shown by", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26109", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b234044e6e568ce38fd4dd4a13c7ec7e60bac25d74f5e9d72763dc3e89c643ad.jpg", + "image_caption": [ + "Figure 5. Qualitative results of simulated DP PSFs. Evaluate the real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) F/1.8 DP PSFs at two depths $(0.5\\mathrm{m}$ and $1.5\\mathrm{m})$ and three different positions. As the object point $p$ moves further from the optical axis, the real $PSF_{L}$ and $PSF_{R}$ become more phase asymmetric, and aberrations increase. Existing simulators neglect aberrations and DP phase splitting, causing a large gap between simulated and real DP PSFs. Only our ray-traced simulator predicts realistic results at all depths and positions." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the green arrows in Fig. 4(d). However, in DP images, disparity from object points before and after the focus distance is opposite in direction. Therefore, we add blue arrows to extend the cost volume generation, in order to enhance the DfDP model's ability to capture reverse disparity.", + "bbox": [ + 89, + 506, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. S skirt Representation Result", + "text_level": 1, + "bbox": [ + 89, + 595, + 346, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 621, + 305, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sdirt employs a Canon RF50mm lens and an R6 Mark II camera body as the real camera. The simulated DP sensor has dimensions of $24\\mathrm{mm} \\times 36\\mathrm{mm}$ with a resolution of $512 \\times 768$ . We set the focus distance of both the simulated and real cameras to $1\\mathrm{m}$ , the valid imaging range to $0.5 - 20\\mathrm{m}$ , the F-number to $F/4$ , and the ray set $A$ to contain 4096 rays. During MLP training, $ks$ is set to 21 to fully display the DP PSF of any point within the valid imaging region. We train the MLP for 100,000 iterations, selecting 128 random points within the valid imaging region during each iteration. We conducted our experiments on a 12700K CPU and a 3090 GPU.", + "bbox": [ + 88, + 643, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [39], we assume chromatic aberration is well corrected compared to other aberrations and defocus effects, allowing us to use a $550\\mathrm{nm}$ wavelength to reduce ray tracing costs. For details on obtaining the structural parameters of the DP pixel, please refer to the Supplementary.", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparison methods", + "text_level": 1, + "bbox": [ + 511, + 506, + 714, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For comparison, we select all existing model-based DP PSF simulators: L2R [2], Modeling [26], DDDNet [23], and an added CoC simulator. Although [24] is a more recent work, its simulator fully reuses DDDNet [23], so we do not include it for comparison. Calibration-based simulators and constraint-training methods are data-driven and require real DP-depth paired data, making them fundamentally different from model-based simulators, so no comparison is made. We directly used the source code of these DP PSF simulators [2, 26], and re-implemented the simulator [23] which only provided executable files. To ensure fairness, we set the focus distance, F-number, and sensor size in each simulator to be consistent with ours.", + "bbox": [ + 511, + 529, + 906, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Evaluation on simulated DP PSFs", + "text_level": 1, + "bbox": [ + 511, + 739, + 808, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To facilitate observation of aberration and phase information, we temporarily set the F-numbers of both the simulated and real cameras to F/1.8. We capture a OLED screen displaying a bright pixel to obtain the real DP PSFs.", + "bbox": [ + 511, + 762, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we evaluate the DP PSF $(PSF_{L}$ on the left, $PSF_{R}$ on the right) of object points $p$ at various depths and positions within the valid imaging region. Observing the real DP PSFs, it can be found that defocus (with focus distance of $1\\mathrm{m}$ ) causes a significant phase difference", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26110", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f04a9898785919d648bab907418b11cf4a868e14099cf0478c7670c820c9399.jpg", + "image_caption": [ + "Figure 6. Qualitative results of simulated DP images. Evaluate the similarity between simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) and real F/4 defocused left DP images at two depths (0.5 m and 2 m). Compared to real F/4 defocused images, images simulated by other methods exhibit varying sizes of patterns, incomplete shapes, and texture shifts in different directions before and after the focus distance (1 m). Our method produces the most realistic simulated images." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "between the real $PSF_{L}$ and $PSF_{R}$ , and the phase difference at depths of $0.5\\mathrm{m}$ and $1.5\\mathrm{m}$ is opposite. When $p$ is on the optical axis, all other simulators fail to provide DP PSF results close to the real ones. Farther from the optical axis, the real $PSF_{L}$ and $PSF_{R}$ become phase asymmetric, with smaller phase differences and more off-axis aberrations. The differences between DP PSFs from other simulators and the real ones become more apparent. In contrast, our ray-traced simulator predicts accurate results at all depths and positions.", + "bbox": [ + 88, + 515, + 480, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the quantitative results in Tab. 1, using OpenCV's normalized squared difference (NSD) and normalized cross-correlation (NCC) to quantify the error and similarity between the real and simulated DP PSFs. We sampled 50 DP PSFs at $0.5\\mathrm{m}$ and $1.5\\mathrm{m}$ within the valid imaging region. Since the real DP PSF has flip symmetry (not phase symmetry) about both the $x$ and $y$ axes, we performed uniform sampling in the first quadrant. Our method achieves the highest similarity (0.915 NCC) and lowest error (0.133 NSD), demonstrating its effectiveness in bridging the domain gap between simulated and real DP PSFs introduced by lens aberrations and sensor phase splitting.", + "bbox": [ + 88, + 670, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In all subsequent experiments, the F-number is adjusted back to F/4 due to the large blur kernel size caused by defocus at F/1.8, which resulted in GPU memory shortages.", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac33c1fec695d79f579046e46e4fdbd64e05f1f1e805c654c0ae642d633ab291.jpg", + "table_caption": [ + "Table 1. Quantitative results of simulated DP PSFs. The NCC and NSD metrics between simulated and real DP PSFs at 50 points were evaluated. Our method achieved the highest similarity and lowest error." + ], + "table_footnote": [], + "table_body": "
MethodOursCoCL2R [2]Modeling [26]DDDNet [23]
NCC↑0.9150.6720.6380.7070.589
NSD↓0.1330.4480.5230.4230.625
", + "bbox": [ + 516, + 566, + 903, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Evaluation on simulated DP images", + "text_level": 1, + "bbox": [ + 511, + 636, + 821, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When comparing simulated and real DP images, to assess the degradation of simulated DP images caused by defocus effects at different depths, we captured richly textured planar scenes from $0.5\\mathrm{m}$ to infinity, perpendicular to the optical axis. Shallower depths were sampled more densely to better capture defocus variations. This resulted in 56 scenes with varying depths, including 11 infinite-depth scenes. For each scene, real DP images were captured at F/4 and F/20. The F/20 image served as the AiF RGB image, and its capture-time depth was assigned to the entire depth map. The F/20 RGB-depth pairs were then input into DP simulators to generate the simulated F/4 defocused DP images.", + "bbox": [ + 509, + 657, + 906, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 6, we compare the qualitative results of the defocused left DP images at depths of $0.5\\mathrm{m}$ and $2\\mathrm{m}$ . Due to errors in simulating the DP PSF, other methods produce simulated images with textures shifted to the right at", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26111", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7d1b9ebb9ac384c1bf71c1d83fb225112b33e28f9a0ea1d26274a2509e6ad1d.jpg", + "image_caption": [ + "Figure 7. Quantitative results of simulated DP images. Evaluate the similarity of 56 real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) planar scene F/4 DP images at different depths using PSNR $(\\uparrow)$ and SSIM $(\\uparrow)$ metrics. As the depth of simulated images deviates further from the focus distance $(1\\mathrm{m})$ , the less realistic the other methods become, whereas our scheme maintains the highest accuracy across all depths." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 285, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9dc4a618d98703cedecee41e77b587735f70ce1dfee490cddf9c457d0fdc2e3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 480, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$0.5\\mathrm{m}$ depth and to the left at $2\\mathrm{m}$ depth, compared to the real left DP image. Moreover, other methods show variations in image texture sizes and incomplete shapes. However, the DP images simulated by our Sdirt method show minimal deviation in texture details.", + "bbox": [ + 89, + 311, + 483, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 7, which presents the quantitative results, we assess the similarity between simulated and real images using PSNR and SSIM. As the simulated image's depth moves further away from the focus distance (1 m), the neglect of aberrations and phase splitting in other methods leads to a more significant decrease in their metrics. In contrast, our Sdirt method achieves the best results at all depths, with average PSNR/SSIM values of 37.1982/0.9845, indicating that our simulated DP images are highly realistic.", + "bbox": [ + 89, + 388, + 483, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. DfDP Estimation Results", + "text_level": 1, + "bbox": [ + 89, + 536, + 323, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 561, + 305, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To train the DfDP model, we simulate F/4 DP images using Sdirt on RGBD datasets [21, 30]. The NYU Depth dataset [30], preprocessed by [12], contains 50,688 indoor scenes. We adopt the AdamW optimizer [19] and CosineAnnealing scheduler [20] with an initial learning rate of $1 \\times 10^{-4}$ . The model is trained with a batch size of 4 for 50 epochs. In each epoch, 2,000 randomly selected DP-depth pairs are used for training. After training, the model is directly evaluated on a real-world test set without any fin-tuning to assess its generalization ability.", + "bbox": [ + 89, + 583, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. A new real-world test set DP119", + "text_level": 1, + "bbox": [ + 89, + 742, + 372, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike previous simulation methods, S skirt addresses data scarcity by simulating hardware, requiring consistent camera and setup during training and inference. However, as shown in Tab. 2, no public dataset satisfies all of the following conditions: 1. Real DP-depth paired data for simulating DP images; 2. A known lens structure for ray tracing; 3. A fixed focus distance, since changing it alters the DP PSF of the same object point. This makes evaluation on public datasets with unknown hardware impractical. We there-", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fore collected the DP119 test set, which contains diverse test scenes.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ca049ba0eb72b1ff94fc5d7b0c9426ec0aedd4c52be084bb3e6927c2e38e7d4.jpg", + "table_caption": [ + "Table 2. Summary of existing DP datasets." + ], + "table_footnote": [], + "table_body": "
DPDD [1]L2R [2]DPNet [9]DP5K [16]DDDNet [23]Modeling [26]DP119 Ours
Real captured✓×✓✓×✓✓
Paired depth×✓✓✓✓×✓
Lens structure✓-×✓-×✓
Fixed focus dist.✓-××-×✓
", + "bbox": [ + 517, + 146, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DP119 consists of 45 planar scenes, 44 box scenes, and 30 casual scenes, totaling 119 scenes. The casual scenes represent common scenarios, suitable for evaluating the robustness of simulation models. The planar and box scenes are richly textured, helping the model utilize aberration and phase cues for depth estimation. Textureless areas, even when defocused, lack aberrations or phase differences, making depth estimation difficult and evaluation unreliable. Thus, richly textured planar and box scenes are ideal for assessing domain gaps between simulated and real DP images. Real DP images are captured by the Canon RF50mm lens with a R6 Mark II at F/4 and $1\\mathrm{m}$ focus. GT depth maps for planar scenes are created by the depths at the capture time, while those for box and casual scenes are obtained from LiDAR scans. For more information about the DP119 test set, please refer to the Supplementary.", + "bbox": [ + 511, + 234, + 906, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Evaluation", + "text_level": 1, + "bbox": [ + 513, + 486, + 633, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present qualitative and quantitative results for the DfDP model trained by CoC, L2R [2], Modeling [26], DDDNet [23], and our Sdirt. For a fair comparison, we deploy their DfDP models using their DP PSF simulator source code [2, 26] and re-implemented code [23], and have all models share the same depth estimation network architecture, initialize from a common CoC pre-trained checkpoint, and follow the same implementation details.", + "bbox": [ + 511, + 508, + 905, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8 shows depth estimation results for four casual scenes, with a color bar in meters for visualization. It can be seen that although other methods provide some relative positional information in the center region, they fail to correctly estimate relative positional relationships in the edge regions, and both the center and edge regions have large absolute positional errors. This is because the DP PSFs predicted by these methods exhibit phase symmetry and shift-invariance, whereas real DP PSFs exhibit phase asymmetry, aberrations, and shift-variance. Their methods' mismatch with real DP PSFs leads to significant depth estimation errors, especially near the edges. The results of L2R [2] and Modeling [26] are similar because the peak positions of their DP PSFs are quite close. In contrast, our method provides accurate relative positional information across the entire image area with minimal absolute positional errors, as our simulated DP PSFs are highly consistent with the real ones in both the center and edge regions.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26112", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d184cd67a6648a800ad3c990640da5270722ca35992508579011ff2c887b3397.jpg", + "image_caption": [ + "Figure 8. Qualitative results of absolute depth estimation on DP119. Evaluate DfDP models with CoC, L2R [2], Modeling [26], DDDNet [23] and our Sdirt at four casual scenes. Each result image includes a color bar showing depth in meters. Their depth estimation results show partial accuracy in relative positional relationships but large absolute positional errors. Our depth estimation results, however, demonstrate high accuracy in both relative and absolute positions with minimal errors. (Best viewed in color and enlarge on screen.)" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 329 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac4fa10e03dac76224fe0b30ee03638acf4a7743d8893a1df4aedb160f4e8ee4.jpg", + "table_caption": [ + "Table 3. Quantitative results of absolute depth estimation on DP119. The DfDP model trained with our Sdirt achieves superior performance across most scenarios and evaluation metrics." + ], + "table_footnote": [], + "table_body": "
SceneMethodMAE↓MSE↓Abs.r↓Sq.r↓Acc-1↑Acc-2↑
PlanarOurs0.08450.01090.08710.00950.98490.9997
CoC0.20850.10010.18010.06590.66700.8990
L2R [2]0.24180.12710.21120.08410.63190.8536
Modeling [26]0.22840.11420.20040.07660.64960.8725
DDDNet [23]0.25830.14850.21910.09580.56480.8089
BoxOurs0.11970.03390.09060.02310.94740.9812
CoC0.33750.18040.24420.11160.44120.8277
L2R [2]0.38660.22840.28030.14120.36510.7156
Modeling [26]0.36550.20550.26600.12780.39070.7758
DDDNet [23]0.41770.26760.29750.16360.34560.6274
CasualOurs0.27020.22940.46320.72410.82360.9314
CoC0.79251.85790.54610.68210.33180.6103
L2R [2]0.81701.74870.55970.67190.27600.5315
Modeling [26]0.79341.72560.55100.66550.29780.5732
DDDNet [23]0.89312.06240.57520.71350.24810.4685
", + "bbox": [ + 93, + 445, + 480, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate all models on the DP119 test set using the following metrics to assess depth estimation performance: mean absolute error (MAE), mean squared error (MSE), absolute relative error (Abs.r.), squared relative error (Sq.r.), accuracy with $\\delta < 1.25$ (Acc-1) and accuracy with $\\delta < 1.25^2$ (Acc-2). Since the planar and box scenes are richly textured and do not have interference from textureless areas, the simulation models can rely on abundant aberration and phase difference cues for depth estimation. As shown in Tab. 3, in the simplest planar scene, our model achieves the best results across all metrics (0.9849 Acc-1). In the box scenes, all models experience a performance drop compared to the planar scenes, but we still achieve the optimal metrics (0.9474 Acc-1). These results demonstrate that our model is highly realistic and effectively bridges the domain gap between simulated and real DP images. Furthermore, we test the robustness of all models on the casual scenes. Due to", + "bbox": [ + 88, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the textureless areas in the casual scenes, the performance of all models significantly degrades. However, our model still achieves the best metrics (0.8236 Acc-1), far surpassing the second-best model (0.3318 Acc-1). This further demonstrates that our model is the most realistic and has the best generalization performance. For more sample results from the DP119 test set, please refer to the Supplementary.", + "bbox": [ + 511, + 404, + 906, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 544, + 761, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Simulated DP images help address the scarcity of DP-depth paired data but face a domain gap between simulated and real DP data. In this work, we propose a novel simulation framework called Sdirt to bridge this gap. Specifically, we calculate the DP PSF for points in object space using ray tracing, and employ a network to predict them. Then, we render DP images based on the predicted DP PSFs. Experimental results show that the proposed Sdirt can simulate more realistic DP data. Moreover, depth estimation models trained based on Sdirt generalize better to real DP images.", + "bbox": [ + 509, + 575, + 906, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that Sdirt is not limited to DfDP tasks, it can provide extra depth information for any task with known optical imaging system parameters, promising significant applications in scenarios such as smartphones, automobiles, and microscopes in the future. However, Sdirt is only applicable to cameras equipped with a fixed-focus lens (the structure must be open) and a DP sensor (DP images must be available). Currently, only Canon (5D4, R series) meets these requirements. To further expand the application of Sdirt, more camera and smartphone manufacturers need to make these data accessible.", + "bbox": [ + 509, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26113", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (Grant No. 32471146) and the project N20240194. The authors thank Echosom, Miya, and Xinge for valuable discussions and assistance.", + "bbox": [ + 89, + 114, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 111-126. Springer, 2020. 2, 7", + "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 2, 5, 6, 7, 8", + "[3] Abdullah Abuolaim, Radu Timofte, and Michael S Brown. Ntire 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 578-587, 2021. 2", + "[4] Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2", + "[5] Hadi Alzayer, Abdullah Abuolaim, Leung Chun Chan, Yang Yang, Ying Chen Lou, Jia-Bin Huang, and Abhishek Kar. Dc2: Dual-camera defocus control by learning to refocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21488-21497, 2023. 2", + "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 4", + "[7] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1, 2", + "[8] Geoffroi Côté, Fahim Mannan, Simon Thibault, Jean-François Lalonde, and Felix Heide. The differentiable lens: Compound lens search over glass surfaces and materials for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20803-20812, 2023. 2", + "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7628-7637, 2019. 1, 2, 7", + "[10] Bhargav Ghanekar, Salman Siddique Khan, Pranav Sharma, Shreyas Singh, Vivek Boominathan, Kaushik Mitra, and" + ], + "bbox": [ + 93, + 213, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ashok Veeraraghavan. Passive snapshot coded aperture dual-pixel rgb-d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25348-25357, 2024. 2", + "[11] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 1, 2", + "[12] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 7", + "[13] Minjun Kang, Jaesung Choe, Hyowon Ha, Hae-Gon Jeon, Sunghoon Im, In So Kweon, and Kuk-Jin Yoon. Facial depth and normal estimation using single dual-pixel camera. In European Conference on Computer Vision, pages 181-200. Springer, 2022. 1, 2", + "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 2", + "[15] Masahiro Kobayashi, Michiko Johnson, Yoichi Wada, Hiromasa Tsuboi, Hideaki Takada, Kenji Togo, Takafumi Kishi, Hidekazu Takahashi, Takeshi Ichikawa, and Shunsuke Inoue. A low noise and high sensitivity image sensor with imaging and phase-difference detection of all pixels. ITE Transactions on Media Technology and Applications, 4(2):123-128, 2016. 1, 3", + "[16] Feiran Li, Heng Guo, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Learning to synthesize photorealistic dual-pixel images from rgbd frames. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2023. 1, 2, 7", + "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 1, 2", + "[18] Yizhou Li, Yusuke Monno, and Masatoshi Okutomi. Dual-pixel raindrop removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2", + "[19] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[20] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7", + "[21] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 4, 7", + "[22] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. In 2024 International Conference on 3D Vision (3DV), pages 675-684. IEEE, 2024. 2" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26114", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 5, 6, 7, 8", + "[24] Liyuan Pan, Richard Hartley, Liu Liu, Zhiwei Xu, Shah Chowdhury, Yan Yang, Hongguang Zhang, Hongdong Li, and Miaomiao Liu. Weakly-supervised depth estimation and image deblurring via dual-pixel sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5", + "[25] Abhijith Punnappurath and Michael S Brown. Reflection removal using a dual-pixel sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1556-1565, 2019. 2", + "[26] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 5, 6, 7, 8", + "[27] Lingyan Ruan, Martin Balint, Mojtaba Bemana, Krzysztof Wolski, Hans-Peter Seidel, Karol Myszkowski, and Bin Chen. Self-supervised video defocus deblurring with atlas learning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[28] Zheng Shi, Ilya Chugunov, Mario Bijelic, Geoffroi Côté, Jiwoon Yeom, Qiang Fu, Hadi Amata, Wolfgang Heidrich, and Felix Heide. Split-aperture 2-in-1 computational cameras. ACM Transactions on Graphics (TOG), 43(4):1-19, 2024. 2", + "[29] Eun Sub Shim, Kyungho Lee, Junghyung Pyo, Wooseok Choi, Jungbin Yun, Taesub Jung, Kyungduck Lee, Seyoung Kim, Chanhee Lee, Seungki Baek, et al. All-directional dual pixel auto focus technology in cmos image sensors. In 2021 Symposium on VLSI Technology, pages 1-2. IEEE, 2021. 1, 3", + "[30] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 7", + "[31] Przemysław Śliwiński and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 1", + "[32] Kunal Swami. Adjust your focus: Defocus deblurring from dual-pixel images using explicit multi-scale cross-correlation. In International Conference on Computer Vision and Image Processing, pages 318-330. Springer, 2023. 2", + "[33] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 4", + "[34] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy." + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2", + "[35] Congli Wang, Ni Chen, and Wolfgang Heidrich. do: A differentiable engine for deep lens design of computational imaging systems. IEEE Transactions on Computational Imaging, 8:905-916, 2022. 2", + "[36] Zhuofeng Wu, Doehyung Lee, Zihua Liu, Kazunori Yoshizaki, Yusuke Monno, and Masatoshi Okutomi. Disparity estimation using a quad-pixel sensor. arXiv preprint arXiv:2409.00665, 2024. 2", + "[37] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2", + "[38] Hao Yang, Liyuan Pan, Yan Yang, Richard Hartley, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24078-24087, 2024. 2", + "[39] Xinge Yang, Qiang Fu, Mohamed Elhoseiny, and Wolfgang Heidrich. Aberration-aware depth-from-focus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 2, 4, 5", + "[40] Xinge Yang, Qiang Fu, and Wolfgang Heidrich. Curriculum learning for ab initio deep learned refractive optics. Nature communications, 15(1):6572, 2024. 2", + "[41] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13263-13272, 2023. 1, 2", + "[42] Dafeng Zhang and Xiaobing Wang. Dynamic multi-scale network for dual-pixel images defocus deblurring with transformer. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 2", + "[43] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du 2 net: Learning depth estimation from dual-cameras and dual-pixels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 582-598. Springer, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 699 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26115", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_model.json b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23a57eb814228b39fa08c9c59f356bda81cb7a7f --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.181, + 0.768, + 0.218 + ], + "angle": 0, + "content": "Fengchen He, Dayang Zhao, Hao Xu, Tingwei Quan, Shaoqun Zeng Huazhong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.22, + 0.788, + 0.234 + ], + "angle": 0, + "content": "{linyark, dayangzhao, xuhao_2003, quantingwei, sqzeng}@hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.573 + ], + "angle": 0, + "content": "Many studies utilize dual-pixel (DP) sensor phase information for various applications, such as depth estimation and deblurring. However, since DP image features are entirely determined by the camera hardware, DP-depth paired datasets are very scarce, especially when performing depth estimation on customized cameras. To overcome this, studies simulate DP images using ideal optical models. However, these simulations often violate real optical propagation laws, leading to poor generalization to real DP data. To address this, we investigate the domain gap between simulated and real DP data, and propose solutions using the Simulating DP Images from Ray Tracing (Sdirt) scheme. Sdirt generates realistic DP images via ray tracing and integrates them into the depth estimation training pipeline. Experimental results show that models trained with Sdirt-simulated images generalize better to real DP data. The code and collected datasets will be available at https://github.com/LinYark/Sdirt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.222, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.824 + ], + "angle": 0, + "content": "The DP sensor [15, 29] is designed to split each pixel into left and right sub-pixels, utilizing microlenses and subpixels to achieve phase splitting. This allows the capture of a pair of images in one shot, known as DP images, as shown in Fig. 1(a). DP images are not only used for autofocus [7, 11, 31], but also improve performance in tasks such as deblurring [4, 17, 41] and depth estimation [9, 13, 23]. In the field of depth-from-dual-pixel (DfDP) estimation, the scarcity of DP-depth paired datasets stems from hardware limitations. To address this, many studies [2, 16, 18, 23, 26, 37] have focused on simulating DP images using RGBD datasets, which serve as more flexible and accessible data sources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The key to simulating DP images from RGBD datasets lies in accurately modeling the DP point spread function (PSF). In recent years, numerous model-based and calibration-based DP PSF simulators have been proposed. Calibration-based simulators [16, 18, 37] require substantial" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.272, + 0.907, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.508, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 1. (a) Imaging process of a DP camera. The slight shifts between the left and right DP images caused by phase differences are illustrated with white dashed lines. (b) Example comparison between real and CoC-simulated DP PSFs, showing a significant difference between them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.793 + ], + "angle": 0, + "content": "time calibrating real cameras to match DP PSFs, but suffer from issues such as interpolation errors due to discrete calibration points and the neglect of lens and DP sensor parameters, limiting model transferability. Model-based simulators [2, 23, 26] employ ideal optical or aberration analysis models to directly compute DP PSFs, bypassing calibration. However, these models often perform poorly on real DP images because they ignore lens aberrations and sensor phase splitting characteristics. As shown in Fig. 1(b), the ideal thin-lens model's simulated circle-of-confusion (CoC) DP PSF exhibits a significant domain gap with the real PSF, adversely affecting depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To bridge the domain gap between real DP images and DP images simulated by model-based simulators, we propose Sdirt. Sdirt consists of a ray-traced DP PSF simulator and a pixel-wise DP image rendering module, which can accurately generate DP images with aberration and phase information. Specifically, given a fixed focus lens and DP sensor system with known parameters, we use ray trac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26106" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ing [35, 39, 40] to calculate the spatially varying DP PSFs and train a multilayer perceptron (MLP) to predict them. Subsequently, we convolve per-pixel DP PSFs with the all-in-focus (AiF) image to simulate the DP images captured by a real camera. After training a DfDP model with DP images simulated by Sdirt, the model can accurately estimate depth using optical aberration and phase information, enhancing its ability to generalize to real DP images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.304 + ], + "angle": 0, + "content": "We conduct extensive experiments to validate the effectiveness of Sdirt. The experimental results demonstrate that, compared to other model-based simulators, the DP PSFs and DP images generated by Sdirt better resemble real data, and the DfDP model trained on Sdirt achieves superior generalization. In summary, our contributions are fourfold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.305, + 0.482, + 0.365 + ], + "angle": 0, + "content": "- We propose a ray-traced DP PSF simulator that computes spatially varying DP PSFs, addressing the domain gap between simulated and real DP PSFs caused by lens aberrations and sensor phase splitting." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.366, + 0.482, + 0.41 + ], + "angle": 0, + "content": "- We propose a pixel-wise DP image rendering module that uses an MLP to predict the DP PSF for each pixel, narrowing the gap between simulated and real DP images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.41, + 0.482, + 0.44 + ], + "angle": 0, + "content": "- Depth estimation results show that the DfDP model trained on Sdirt generalizes better to real DP images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.44, + 0.482, + 0.484 + ], + "angle": 0, + "content": "- We collected DP119, a real DP-depth paired test set with an open lens structure and fixed focus, featuring diverse real-world scenes." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.305, + 0.482, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.233, + 0.517 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.221, + 0.541 + ], + "angle": 0, + "content": "2.1. DP Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Most professional and mobile cameras have DP sensors, but only Canon cameras [1, 25, 26] and Google Pixel phones [9, 34, 43] offer DP data. In the field of depth estimation, supervised tasks rely on DP-depth paired datasets [1, 9, 26, 43]. However, changes in the lens, sensor, focal length, or aperture settings alter the PSFs, making it necessary to reacquire DP-depth data. Consequently, many researchers focus on simulating DP-depth datasets using RGBD data to reduce data collection costs and improve flexibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.273, + 0.712 + ], + "angle": 0, + "content": "2.2. DP PSF simulators" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Simulating DP images from RGBD datasets relies on DP PSFs, with recent progress in calibration-based [16, 18, 37] and model-based [2, 23, 26] DP PSF simulators. In calibration-based simulators, Xin et al. [37] acquired full-space DP PSFs by sampling discrete points in space and interpolating, while Li et al. [18] acquired them through size scaling using the CoC model. Furthermore, Li et al. [16] generated DP PSFs using a U-Net, which requires a large amount of DP-depth paired data. Calibration-based simulators are time-consuming and prone to interpolation and scaling errors, while model-based simulators circumvent these issues. Pan et al. [23] simplified the optical system to an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "ideal thin lens and used a symmetrically divided rectangular aperture to calculate DP PSFs. Punnappurath et al. [26] modeled the DP PSFs using a depletion kernel that is phase symmetric, meaning the left and right PSFs of the same object point are flip symmetric. Abuolaim et al. [2] used aberration analysis models to calculate DP PSFs, but the resulting patterns still remained phase symmetric. These studies oversimplify optical propagation by neglecting aberrations and phase splitting, leading to unrealistic domain gaps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.792, + 0.252 + ], + "angle": 0, + "content": "2.3. Applications driven by DP data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.47 + ], + "angle": 0, + "content": "In recent years, DP data-driven applications such as depth estimation [9, 10, 13, 14, 23, 24, 43], deblurring [3, 4, 17, 27, 38, 41, 42], and refocusing [5, 7, 11, 32] have undergone rapid development. Additionally, Kang et al. [13] utilized DP data for facial scanning, similar to depth estimation tasks. Wadhwa et al. [34] employed DP data for synthetic shallow depth-of-field imaging, a popular recent application in mobile photography. Shi et al. [28] integrated DP hardware with structured diffractive optical elements for complex image encoding and high-precision multimodal reconstruction. In addition, DP data can be used for disparity estimation [22, 36], rain removal [18], and reflection removal [25]. It is foreseeable that DP data-driven applications will continue to emerge in various new fields." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.673, + 0.497 + ], + "angle": 0, + "content": "3. Method of Sdirt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.644 + ], + "angle": 0, + "content": "To provide the required input for training the DfDP network, Sdirt generates realistic DP images through ray tracing. As shown in Fig. 2(c), the module achieves pixel-wise rendering of DP images by convolving the DP PSF corresponding to each point in the depth map with the AiF RGB image. The DP PSF for each point is inferred by a pretrained MLP network. As illustrated in Fig. 2(b), the offline training process of the MLP network uses ray-traced DP PSFs (Fig. 2(a)) as ground truth (GT)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.78, + 0.668 + ], + "angle": 0, + "content": "3.1. Ray-traced DP PSF simulator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.855 + ], + "angle": 0, + "content": "As shown in Fig. 2(a), lens ray tracing can accurately obtain the landing point on the sensor surface [8, 35, 39, 40]. We construct a ray set \\( A \\) containing \\( n \\) rays, each originating from an object point \\( p \\). The aperture stop's image in object space is treated as the entrance pupil, from which \\( n \\) points are densely sampled. Each ray in \\( A \\) is initially directed from \\( p \\) to its corresponding entrance pupil sample. As rays in \\( A \\) traverse each lens surface, their positions and directions are updated by refraction according to Snell's law and the lens parameters. After multiple refractions, the rays reach the sensor plane. We denote the sensor-plane landing points as \\( O \\) and the corresponding ray directions as \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "DP sensor ray tracing is limited because camera manufacturers do not disclose the structural parameters of the microlens and sub-pixel components within the DP pixel." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26107" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.097, + 0.129, + 0.115 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.093, + 0.907, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.436, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Figure 2. Simulating Dual-Pixel Images from Ray Tracing pipeline. (a) Ray-traced DP PSF simulator. Calculates spatially varying DP PSFs for lens and DP sensor through ray tracing. (b) DP PSF prediction network. Trains an MLP network to predict DP PSFs, using the ray-traced DP PSFs as GT. (c) Pixel-wise DP image rendering module. The network predicts the DP PSFs for all points in the depth map (red pass). Then, each DP PSF is convolved with the AiF RGB image to render the simulated DP image (blue pass)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.587 + ], + "angle": 0, + "content": "As a result, we simplify the DP pixel structure (Fig. 3(a)) based on past research [15, 29]. We model the microlens as a thin lens with radius \\( r \\) and focal length \\( f \\), and define \\( h \\) as distance between the sub-pixel and the microlens, \\( w \\) as sub-pixel width, and \\( ps \\) as DP pixel size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.744 + ], + "angle": 0, + "content": "The DP PSF can be obtained by separately calculating the cumulative integration of the rays for each sub-pixel. We analyze which sub-pixel the \\(k\\)-th ray \\(A_{k}\\) in the ray set \\(A\\) ultimately enters. Through lens ray tracing, we obtain the landing point \\(O_{k}(x_{k},y_{k},z_{k})\\) and direction \\(D_{k}(x_{k}^{\\prime},y_{k}^{\\prime},z_{k}^{\\prime})\\) of the \\(k\\)-th ray \\(A_{k}\\) on the sensor surface. Based on the landing point \\(O_{k}\\), we can easily calculate that \\(A_{k}\\) lies within the DP pixel \\((i,j)\\), which has surface center coordinates \\((x_{i},y_{i})\\). For \\(O_{k}\\), the situation differs depending on whether it lies within or outside the microlens:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "When \\(O_{k}\\) lies within the microlens, as shown in Fig. 3(b), \\(A_{k}\\) is refracted by the microlens and then enters a sub-pixel. According to geometric optics theory, \\(A_{k}\\) and the principal ray incident on the thin lens at the same angle both converge on the focal plane of the thin lens. We only need to focus on a few boundary lines \\(x_{L1}, x_{M1}, x_{R1}\\) to determine which sub-pixel \\(A_{k}\\) ultimately enters. If \\(x_{k}\\) is in the interval \\([x_{L1}, x_{M1}]\\), it enters the left sub-pixel; if in the interval \\([x_{M1}, x_{R1}]\\), it enters the right sub-pixel; otherwise, it is considered a missing ray. The calculation methods for the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.511, + 0.714, + 0.526 + ], + "angle": 0, + "content": "boundary lines are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.533, + 0.866, + 0.55 + ], + "angle": 0, + "content": "\\[\nx _ {L 1} = x _ {i} + w - (f * \\tan \\theta - w) * h / (f - h)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.552, + 0.905, + 0.568 + ], + "angle": 0, + "content": "\\[\nx _ {M 1} = x _ {i} - (f * \\tan \\theta) * h / (f - h) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.571, + 0.866, + 0.587 + ], + "angle": 0, + "content": "\\[\nx _ {R 1} = x _ {i} - w - (f * \\tan \\theta + w) * h / (f - h)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.905, + 0.639 + ], + "angle": 0, + "content": "Where \\(\\tan \\theta\\) is the tangent of the angle at which \\(A_{k}\\) lands on the sensor surface, equivalent to \\(x_{k}^{\\prime} / z_{k}^{\\prime}\\). We provide the derivation hints for Eq. (1) in Fig. 4(a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.905, + 0.7 + ], + "angle": 0, + "content": "When \\(O_{k}\\) lies outside the microlens, as shown in Fig. 3(c), \\(A_{k}\\) enters the sub-pixel directly without refraction. Similarly, checking boundary lines \\(x_{L2}, x_{M2}, x_{R2}\\) suffices to determine which sub-pixel \\(A_{k}\\) ultimately enters:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.707, + 0.803, + 0.722 + ], + "angle": 0, + "content": "\\[\nx _ {L 2} = x _ {i} + w - h * \\tan \\theta\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.726, + 0.905, + 0.741 + ], + "angle": 0, + "content": "\\[\nx _ {M 2} = x _ {i} - h * \\tan \\theta \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.745, + 0.802, + 0.759 + ], + "angle": 0, + "content": "\\[\nx _ {R 2} = x _ {i} - w - h * \\tan \\theta\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.811 + ], + "angle": 0, + "content": "The left PSF \\((PSF_{L})\\) can be calculated by integrating the set of rays \\(A\\) emitted from point \\(p\\) over all left subpixels:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.811, + 0.905, + 0.85 + ], + "angle": 0, + "content": "\\[\nP S F _ {L} (i, j) = \\sum_ {k = 1} ^ {n} A _ {k} * \\delta_ {L, k} (i, j) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\delta_{L,k}(i,j)\\) represents the energy distribution of ray \\(A_{k}\\) in the left sub-pixel at DP pixel \\((i,j)\\). The ray energy distribution is assumed to be a unit impulse, which" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.445, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.541 + ], + "angle": 0, + "content": "Figure 3. (a) DP pixel structure layout. The left part is a right-side view, and the right part is a front perspective view. (b) On hitting the microlens, the \\( k \\)-th ray lands on left/right sub-pixel in red/blue interval, else missed. (c) Without hitting the microlens, the \\( k \\)-th ray lands on left/right sub-pixel in red/blue interval, else missed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.601 + ], + "angle": 0, + "content": "equals zero unless it eventually enters the left sub-pixel of DP pixel \\((i,j)\\). By following the same calculation steps as for \\(PSF_{L}\\), \\(PSF_{R}\\) can also be computed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.425, + 0.625 + ], + "angle": 0, + "content": "3.2. Pixel-wise DP image rendering module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.827 + ], + "angle": 0, + "content": "As shown in Fig. 2(b), we reduce the computational cost of ray-traced DP PSFs by training an MLP network to predict them, inspired by [33, 39]. The network input is the normalized coordinate \\( p \\) within the camera's valid imaging region (Fig. 4(b)), which is a frustum defined by the field of view, sensor size, and the preset minimum and maximum depths. We normalize the \\( (x,y) \\) coordinates to \\( [-1,1] \\). After fixing the focus distance, we set the DP PSF kernel size \\( (ks) \\) large enough to ensure that the DP PSF for any point within the valid imaging region can be fully displayed. The network consists of 5 hidden layers, each containing 512 neurons, and an output layer with \\( 2 * ks^2 \\) neurons. We adopt the L2 loss to supervise predictions \\( P\\hat{S} F_L \\) and \\( P\\hat{S} F_R \\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.832, + 0.483, + 0.851 + ], + "angle": 0, + "content": "\\[\nL o s s = L _ {2} (P \\hat {S} F _ {L}, P S F _ {L}) + L _ {2} (P \\hat {S} F _ {R}, P S F _ {R}) \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "During training, we use ray-traced DP PSFs as GT and apply max normalization to alleviate learning difficulties caused by large radii PSFs. During inference, we apply" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.9, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.351, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 4. (a) Derivation hints for Eq. (1). (b) The valid imaging region is a frustum, and we normalize the \\((x,y)\\) coordinates to \\([-1,1]\\). (c) The DfDP model takes DP image as input to predict the depth map. (d) During cost volume generation in [6], we stack original disparity (green arrows) and add reverse disparity (blue arrows), with \\(d_{max}\\) as max displacement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.905, + 0.496 + ], + "angle": 0, + "content": "sum normalization to the predicted DP PSFs, approximating the uniform intensity distribution in cameras with vignetting compensation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.633 + ], + "angle": 0, + "content": "Because the DP PSF is shift-variant, we convolve each pixel's DP PSF with the AiF image to simulate the DP image captured by a camera. The RGBD dataset [21, 30] provides the paired RGB image \\((I_{RGB})\\) and depth map \\((I_D)\\). As shown in Fig. 2(c), we treat each pixel in \\(I_D\\) as an object point and use the trained MLP to predict its DP PSF. Using \\(I_{RGB}\\) as the AiF image, we apply pixel-wise local convolution [39] to render the DP image containing aberration and phase information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.771, + 0.665 + ], + "angle": 0, + "content": "4. DfDP Model Based on Sdtirt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.797 + ], + "angle": 0, + "content": "As shown in Fig. 4(c), the input to the DfDP model is provided in real time by the pixel-wise DP image rendering module of Sdirt. At each training iteration, Sdirt feeds the DfDP model with simulated DP images of shape \\((B,6,H,W)\\), where \\(B\\) is the batch size, the number of channels is 6, and \\(H, W\\) are the DP image height and width. The model outputs the predicted depth maps \\((\\hat{I}_D)\\) of shape \\((B,1,H,W)\\). We apply the L1 loss to supervise \\(\\hat{I}_D\\):" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.801, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\nL o s s = L _ {1} \\left(\\hat {I _ {D}}, I _ {D}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We select [6] as the DfDP model and make reasonable adjustments to its cost volume step to accommodate the DP data. In binocular images, disparity from points at different depths always has the same direction, thus the cost volume generation process in [6] is also unidirectional, as shown by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.907, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.907, + 0.491 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of simulated DP PSFs. Evaluate the real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) F/1.8 DP PSFs at two depths \\((0.5\\mathrm{m}\\) and \\(1.5\\mathrm{m})\\) and three different positions. As the object point \\(p\\) moves further from the optical axis, the real \\(PSF_{L}\\) and \\(PSF_{R}\\) become more phase asymmetric, and aberrations increase. Existing simulators neglect aberrations and DP phase splitting, causing a large gap between simulated and real DP PSFs. Only our ray-traced simulator predicts realistic results at all depths and positions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.584 + ], + "angle": 0, + "content": "the green arrows in Fig. 4(d). However, in DP images, disparity from object points before and after the focus distance is opposite in direction. Therefore, we add blue arrows to extend the cost volume generation, in order to enhance the DfDP model's ability to capture reverse disparity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.597, + 0.348, + 0.614 + ], + "angle": 0, + "content": "5. S skirt Representation Result" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.622, + 0.306, + 0.639 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Sdirt employs a Canon RF50mm lens and an R6 Mark II camera body as the real camera. The simulated DP sensor has dimensions of \\(24\\mathrm{mm} \\times 36\\mathrm{mm}\\) with a resolution of \\(512 \\times 768\\). We set the focus distance of both the simulated and real cameras to \\(1\\mathrm{m}\\), the valid imaging range to \\(0.5 - 20\\mathrm{m}\\), the F-number to \\(F/4\\), and the ray set \\(A\\) to contain 4096 rays. During MLP training, \\(ks\\) is set to 21 to fully display the DP PSF of any point within the valid imaging region. We train the MLP for 100,000 iterations, selecting 128 random points within the valid imaging region during each iteration. We conducted our experiments on a 12700K CPU and a 3090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Following [39], we assume chromatic aberration is well corrected compared to other aberrations and defocus effects, allowing us to use a \\(550\\mathrm{nm}\\) wavelength to reduce ray tracing costs. For details on obtaining the structural parameters of the DP pixel, please refer to the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.715, + 0.523 + ], + "angle": 0, + "content": "5.2. Comparison methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.727 + ], + "angle": 0, + "content": "For comparison, we select all existing model-based DP PSF simulators: L2R [2], Modeling [26], DDDNet [23], and an added CoC simulator. Although [24] is a more recent work, its simulator fully reuses DDDNet [23], so we do not include it for comparison. Calibration-based simulators and constraint-training methods are data-driven and require real DP-depth paired data, making them fundamentally different from model-based simulators, so no comparison is made. We directly used the source code of these DP PSF simulators [2, 26], and re-implemented the simulator [23] which only provided executable files. To ensure fairness, we set the focus distance, F-number, and sensor size in each simulator to be consistent with ours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.809, + 0.755 + ], + "angle": 0, + "content": "5.3. Evaluation on simulated DP PSFs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.824 + ], + "angle": 0, + "content": "To facilitate observation of aberration and phase information, we temporarily set the F-numbers of both the simulated and real cameras to F/1.8. We capture a OLED screen displaying a bright pixel to obtain the real DP PSFs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 5, we evaluate the DP PSF \\((PSF_{L}\\) on the left, \\(PSF_{R}\\) on the right) of object points \\(p\\) at various depths and positions within the valid imaging region. Observing the real DP PSFs, it can be found that defocus (with focus distance of \\(1\\mathrm{m}\\)) causes a significant phase difference" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.907, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.445, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of simulated DP images. Evaluate the similarity between simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) and real F/4 defocused left DP images at two depths (0.5 m and 2 m). Compared to real F/4 defocused images, images simulated by other methods exhibit varying sizes of patterns, incomplete shapes, and texture shifts in different directions before and after the focus distance (1 m). Our method produces the most realistic simulated images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.482, + 0.668 + ], + "angle": 0, + "content": "between the real \\(PSF_{L}\\) and \\(PSF_{R}\\), and the phase difference at depths of \\(0.5\\mathrm{m}\\) and \\(1.5\\mathrm{m}\\) is opposite. When \\(p\\) is on the optical axis, all other simulators fail to provide DP PSF results close to the real ones. Farther from the optical axis, the real \\(PSF_{L}\\) and \\(PSF_{R}\\) become phase asymmetric, with smaller phase differences and more off-axis aberrations. The differences between DP PSFs from other simulators and the real ones become more apparent. In contrast, our ray-traced simulator predicts accurate results at all depths and positions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.853 + ], + "angle": 0, + "content": "We present the quantitative results in Tab. 1, using OpenCV's normalized squared difference (NSD) and normalized cross-correlation (NCC) to quantify the error and similarity between the real and simulated DP PSFs. We sampled 50 DP PSFs at \\(0.5\\mathrm{m}\\) and \\(1.5\\mathrm{m}\\) within the valid imaging region. Since the real DP PSF has flip symmetry (not phase symmetry) about both the \\(x\\) and \\(y\\) axes, we performed uniform sampling in the first quadrant. Our method achieves the highest similarity (0.915 NCC) and lowest error (0.133 NSD), demonstrating its effectiveness in bridging the domain gap between simulated and real DP PSFs introduced by lens aberrations and sensor phase splitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In all subsequent experiments, the F-number is adjusted back to F/4 due to the large blur kernel size caused by defocus at F/1.8, which resulted in GPU memory shortages." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.511, + 0.907, + 0.565 + ], + "angle": 0, + "content": "Table 1. Quantitative results of simulated DP PSFs. The NCC and NSD metrics between simulated and real DP PSFs at 50 points were evaluated. Our method achieved the highest similarity and lowest error." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.567, + 0.904, + 0.619 + ], + "angle": 0, + "content": "
MethodOursCoCL2R [2]Modeling [26]DDDNet [23]
NCC↑0.9150.6720.6380.7070.589
NSD↓0.1330.4480.5230.4230.625
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.822, + 0.654 + ], + "angle": 0, + "content": "5.4. Evaluation on simulated DP images" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.84 + ], + "angle": 0, + "content": "When comparing simulated and real DP images, to assess the degradation of simulated DP images caused by defocus effects at different depths, we captured richly textured planar scenes from \\(0.5\\mathrm{m}\\) to infinity, perpendicular to the optical axis. Shallower depths were sampled more densely to better capture defocus variations. This resulted in 56 scenes with varying depths, including 11 infinite-depth scenes. For each scene, real DP images were captured at F/4 and F/20. The F/20 image served as the AiF RGB image, and its capture-time depth was assigned to the entire depth map. The F/20 RGB-depth pairs were then input into DP simulators to generate the simulated F/4 defocused DP images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 6, we compare the qualitative results of the defocused left DP images at depths of \\(0.5\\mathrm{m}\\) and \\(2\\mathrm{m}\\). Due to errors in simulating the DP PSF, other methods produce simulated images with textures shifted to the right at" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.287, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.091, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.302 + ], + "angle": 0, + "content": "Figure 7. Quantitative results of simulated DP images. Evaluate the similarity of 56 real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) planar scene F/4 DP images at different depths using PSNR \\((\\uparrow)\\) and SSIM \\((\\uparrow)\\) metrics. As the depth of simulated images deviates further from the focus distance \\((1\\mathrm{m})\\), the less realistic the other methods become, whereas our scheme maintains the highest accuracy across all depths." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.387 + ], + "angle": 0, + "content": "\\(0.5\\mathrm{m}\\) depth and to the left at \\(2\\mathrm{m}\\) depth, compared to the real left DP image. Moreover, other methods show variations in image texture sizes and incomplete shapes. However, the DP images simulated by our Sdirt method show minimal deviation in texture details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.525 + ], + "angle": 0, + "content": "As shown in Fig. 7, which presents the quantitative results, we assess the similarity between simulated and real images using PSNR and SSIM. As the simulated image's depth moves further away from the focus distance (1 m), the neglect of aberrations and phase splitting in other methods leads to a more significant decrease in their metrics. In contrast, our Sdirt method achieves the best results at all depths, with average PSNR/SSIM values of 37.1982/0.9845, indicating that our simulated DP images are highly realistic." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.325, + 0.552 + ], + "angle": 0, + "content": "6. DfDP Estimation Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.307, + 0.577 + ], + "angle": 0, + "content": "6.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.735 + ], + "angle": 0, + "content": "To train the DfDP model, we simulate F/4 DP images using Sdirt on RGBD datasets [21, 30]. The NYU Depth dataset [30], preprocessed by [12], contains 50,688 indoor scenes. We adopt the AdamW optimizer [19] and CosineAnnealing scheduler [20] with an initial learning rate of \\(1 \\times 10^{-4}\\). The model is trained with a batch size of 4 for 50 epochs. In each epoch, 2,000 randomly selected DP-depth pairs are used for training. After training, the model is directly evaluated on a real-world test set without any fin-tuning to assess its generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.373, + 0.758 + ], + "angle": 0, + "content": "6.2. A new real-world test set DP119" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Unlike previous simulation methods, S skirt addresses data scarcity by simulating hardware, requiring consistent camera and setup during training and inference. However, as shown in Tab. 2, no public dataset satisfies all of the following conditions: 1. Real DP-depth paired data for simulating DP images; 2. A known lens structure for ray tracing; 3. A fixed focus distance, since changing it alters the DP PSF of the same object point. This makes evaluation on public datasets with unknown hardware impractical. We there-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "fore collected the DP119 test set, which contains diverse test scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.582, + 0.133, + 0.837, + 0.146 + ], + "angle": 0, + "content": "Table 2. Summary of existing DP datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.147, + 0.905, + 0.228 + ], + "angle": 0, + "content": "
DPDD [1]L2R [2]DPNet [9]DP5K [16]DDDNet [23]Modeling [26]DP119 Ours
Real captured✓×✓✓×✓✓
Paired depth×✓✓✓✓×✓
Lens structure✓-×✓-×✓
Fixed focus dist.✓-××-×✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.907, + 0.478 + ], + "angle": 0, + "content": "DP119 consists of 45 planar scenes, 44 box scenes, and 30 casual scenes, totaling 119 scenes. The casual scenes represent common scenarios, suitable for evaluating the robustness of simulation models. The planar and box scenes are richly textured, helping the model utilize aberration and phase cues for depth estimation. Textureless areas, even when defocused, lack aberrations or phase differences, making depth estimation difficult and evaluation unreliable. Thus, richly textured planar and box scenes are ideal for assessing domain gaps between simulated and real DP images. Real DP images are captured by the Canon RF50mm lens with a R6 Mark II at F/4 and \\(1\\mathrm{m}\\) focus. GT depth maps for planar scenes are created by the depths at the capture time, while those for box and casual scenes are obtained from LiDAR scans. For more information about the DP119 test set, please refer to the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.634, + 0.501 + ], + "angle": 0, + "content": "6.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.629 + ], + "angle": 0, + "content": "We present qualitative and quantitative results for the DfDP model trained by CoC, L2R [2], Modeling [26], DDDNet [23], and our Sdirt. For a fair comparison, we deploy their DfDP models using their DP PSF simulator source code [2, 26] and re-implemented code [23], and have all models share the same depth estimation network architecture, initialize from a common CoC pre-trained checkpoint, and follow the same implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Figure 8 shows depth estimation results for four casual scenes, with a color bar in meters for visualization. It can be seen that although other methods provide some relative positional information in the center region, they fail to correctly estimate relative positional relationships in the edge regions, and both the center and edge regions have large absolute positional errors. This is because the DP PSFs predicted by these methods exhibit phase symmetry and shift-invariance, whereas real DP PSFs exhibit phase asymmetry, aberrations, and shift-variance. Their methods' mismatch with real DP PSFs leads to significant depth estimation errors, especially near the edges. The results of L2R [2] and Modeling [26] are similar because the peak positions of their DP PSFs are quite close. In contrast, our method provides accurate relative positional information across the entire image area with minimal absolute positional errors, as our simulated DP PSFs are highly consistent with the real ones in both the center and edge regions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26112" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.904, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 8. Qualitative results of absolute depth estimation on DP119. Evaluate DfDP models with CoC, L2R [2], Modeling [26], DDDNet [23] and our Sdirt at four casual scenes. Each result image includes a color bar showing depth in meters. Their depth estimation results show partial accuracy in relative positional relationships but large absolute positional errors. Our depth estimation results, however, demonstrate high accuracy in both relative and absolute positions with minimal errors. (Best viewed in color and enlarge on screen.)" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Table 3. Quantitative results of absolute depth estimation on DP119. The DfDP model trained with our Sdirt achieves superior performance across most scenarios and evaluation metrics." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.446, + 0.481, + 0.622 + ], + "angle": 0, + "content": "
SceneMethodMAE↓MSE↓Abs.r↓Sq.r↓Acc-1↑Acc-2↑
PlanarOurs0.08450.01090.08710.00950.98490.9997
CoC0.20850.10010.18010.06590.66700.8990
L2R [2]0.24180.12710.21120.08410.63190.8536
Modeling [26]0.22840.11420.20040.07660.64960.8725
DDDNet [23]0.25830.14850.21910.09580.56480.8089
BoxOurs0.11970.03390.09060.02310.94740.9812
CoC0.33750.18040.24420.11160.44120.8277
L2R [2]0.38660.22840.28030.14120.36510.7156
Modeling [26]0.36550.20550.26600.12780.39070.7758
DDDNet [23]0.41770.26760.29750.16360.34560.6274
CasualOurs0.27020.22940.46320.72410.82360.9314
CoC0.79251.85790.54610.68210.33180.6103
L2R [2]0.81701.74870.55970.67190.27600.5315
Modeling [26]0.79341.72560.55100.66550.29780.5732
DDDNet [23]0.89312.06240.57520.71350.24810.4685
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate all models on the DP119 test set using the following metrics to assess depth estimation performance: mean absolute error (MAE), mean squared error (MSE), absolute relative error (Abs.r.), squared relative error (Sq.r.), accuracy with \\(\\delta < 1.25\\) (Acc-1) and accuracy with \\(\\delta < 1.25^2\\) (Acc-2). Since the planar and box scenes are richly textured and do not have interference from textureless areas, the simulation models can rely on abundant aberration and phase difference cues for depth estimation. As shown in Tab. 3, in the simplest planar scene, our model achieves the best results across all metrics (0.9849 Acc-1). In the box scenes, all models experience a performance drop compared to the planar scenes, but we still achieve the optimal metrics (0.9474 Acc-1). These results demonstrate that our model is highly realistic and effectively bridges the domain gap between simulated and real DP images. Furthermore, we test the robustness of all models on the casual scenes. Due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.513 + ], + "angle": 0, + "content": "the textureless areas in the casual scenes, the performance of all models significantly degrades. However, our model still achieves the best metrics (0.8236 Acc-1), far surpassing the second-best model (0.3318 Acc-1). This further demonstrates that our model is the most realistic and has the best generalization performance. For more sample results from the DP119 test set, please refer to the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.762, + 0.56 + ], + "angle": 0, + "content": "7. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.577, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Simulated DP images help address the scarcity of DP-depth paired data but face a domain gap between simulated and real DP data. In this work, we propose a novel simulation framework called Sdirt to bridge this gap. Specifically, we calculate the DP PSF for points in object space using ray tracing, and employ a network to predict them. Then, we render DP images based on the predicted DP PSFs. Experimental results show that the proposed Sdirt can simulate more realistic DP data. Moreover, depth estimation models trained based on Sdirt generalize better to real DP images." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We believe that Sdirt is not limited to DfDP tasks, it can provide extra depth information for any task with known optical imaging system parameters, promising significant applications in scenarios such as smartphones, automobiles, and microscopes in the future. However, Sdirt is only applicable to cameras equipped with a fixed-focus lens (the structure must be open) and a DP sensor (DP images must be available). Currently, only Canon (5D4, R series) meets these requirements. To further expand the application of Sdirt, more camera and smartphone manufacturers need to make these data accessible." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26113" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.177 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 32471146) and the project N20240194. The authors thank Echosom, Miya, and Xinge for valuable discussions and assistance." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 111-126. Springer, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.284, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.485, + 0.424 + ], + "angle": 0, + "content": "[3] Abdullah Abuolaim, Radu Timofte, and Michael S Brown. Ntire 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 578-587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.485, + 0.495 + ], + "angle": 0, + "content": "[4] Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Hadi Alzayer, Abdullah Abuolaim, Leung Chun Chan, Yang Yang, Ying Chen Lou, Jia-Bin Huang, and Abhishek Kar. Dc2: Dual-camera defocus control by learning to refocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21488-21497, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.485, + 0.648 + ], + "angle": 0, + "content": "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[7] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[8] Geoffroi Côté, Fahim Mannan, Simon Thibault, Jean-François Lalonde, and Felix Heide. The differentiable lens: Compound lens search over glass surfaces and materials for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20803-20812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7628-7637, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Bhargav Ghanekar, Salman Siddique Khan, Pranav Sharma, Shreyas Singh, Vivek Boominathan, Kaushik Mitra, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Ashok Veeraraghavan. Passive snapshot coded aperture dual-pixel rgb-d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25348-25357, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.207 + ], + "angle": 0, + "content": "[11] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.907, + 0.278 + ], + "angle": 0, + "content": "[12] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.906, + 0.35 + ], + "angle": 0, + "content": "[13] Minjun Kang, Jaesung Choe, Hyowon Ha, Hae-Gon Jeon, Sunghoon Im, In So Kweon, and Kuk-Jin Yoon. Facial depth and normal estimation using single dual-pixel camera. In European Conference on Computer Vision, pages 181-200. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.906, + 0.422 + ], + "angle": 0, + "content": "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.906, + 0.522 + ], + "angle": 0, + "content": "[15] Masahiro Kobayashi, Michiko Johnson, Yoichi Wada, Hiromasa Tsuboi, Hideaki Takada, Kenji Togo, Takafumi Kishi, Hidekazu Takahashi, Takeshi Ichikawa, and Shunsuke Inoue. A low noise and high sensitivity image sensor with imaging and phase-difference detection of all pixels. ITE Transactions on Media Technology and Applications, 4(2):123-128, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[16] Feiran Li, Heng Guo, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Learning to synthesize photorealistic dual-pixel images from rgbd frames. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.906, + 0.639 + ], + "angle": 0, + "content": "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.906, + 0.683 + ], + "angle": 0, + "content": "[18] Yizhou Li, Yusuke Monno, and Masatoshi Okutomi. Dual-pixel raindrop removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.685, + 0.906, + 0.713 + ], + "angle": 0, + "content": "[19] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[21] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. In 2024 International Conference on 3D Vision (3DV), pages 675-684. IEEE, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26114" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[23] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[24] Liyuan Pan, Richard Hartley, Liu Liu, Zhiwei Xu, Shah Chowdhury, Yan Yang, Hongguang Zhang, Hongdong Li, and Miaomiao Liu. Weakly-supervised depth estimation and image deblurring via dual-pixel sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[25] Abhijith Punnappurath and Michael S Brown. Reflection removal using a dual-pixel sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1556-1565, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[26] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[27] Lingyan Ruan, Martin Balint, Mojtaba Bemana, Krzysztof Wolski, Hans-Peter Seidel, Karol Myszkowski, and Bin Chen. Self-supervised video defocus deblurring with atlas learning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[28] Zheng Shi, Ilya Chugunov, Mario Bijelic, Geoffroi Côté, Jiwoon Yeom, Qiang Fu, Hadi Amata, Wolfgang Heidrich, and Felix Heide. Split-aperture 2-in-1 computational cameras. ACM Transactions on Graphics (TOG), 43(4):1-19, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[29] Eun Sub Shim, Kyungho Lee, Junghyung Pyo, Wooseok Choi, Jungbin Yun, Taesub Jung, Kyungduck Lee, Seyoung Kim, Chanhee Lee, Seungki Baek, et al. All-directional dual pixel auto focus technology in cmos image sensors. In 2021 Symposium on VLSI Technology, pages 1-2. IEEE, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[30] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[31] Przemysław Śliwiński and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[32] Kunal Swami. Adjust your focus: Defocus deblurring from dual-pixel images using explicit multi-scale cross-correlation. In International Conference on Computer Vision and Image Processing, pages 318-330. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[33] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[34] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[35] Congli Wang, Ni Chen, and Wolfgang Heidrich. do: A differentiable engine for deep lens design of computational imaging systems. IEEE Transactions on Computational Imaging, 8:905-916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[36] Zhuofeng Wu, Doehyung Lee, Zihua Liu, Kazunori Yoshizaki, Yusuke Monno, and Masatoshi Okutomi. Disparity estimation using a quad-pixel sensor. arXiv preprint arXiv:2409.00665, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[37] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[38] Hao Yang, Liyuan Pan, Yan Yang, Richard Hartley, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24078-24087, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[39] Xinge Yang, Qiang Fu, Mohamed Elhoseiny, and Wolfgang Heidrich. Aberration-aware depth-from-focus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[40] Xinge Yang, Qiang Fu, and Wolfgang Heidrich. Curriculum learning for ab initio deep learned refractive optics. Nature communications, 15(1):6572, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[41] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13263-13272, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[42] Dafeng Zhang and Xiaobing Wang. Dynamic multi-scale network for dual-pixel images defocus deblurring with transformer. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[43] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du 2 net: Learning depth estimation from dual-cameras and dual-pixels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 582-598. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26115" + } + ] +] \ No newline at end of file diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_origin.pdf b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..baf9c2856c4ec6c68a909ff85eb88fa903dae630 --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/daa8d996-9820-46cb-8540-26b2ea341407_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53bb9f2314f669c4524e9f1f81631a31ea3de210762ad7db692ca58a754cc063 +size 4923839 diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/full.md b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f4014f8211331537e61fe5f5bd702c1e9fe81425 --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/full.md @@ -0,0 +1,284 @@ +# Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation + +Fengchen He, Dayang Zhao, Hao Xu, Tingwei Quan, Shaoqun Zeng Huazhong University of Science and Technology + +{linyark, dayangzhao, xuhao_2003, quantingwei, sqzeng}@hust.edu.cn + +# Abstract + +Many studies utilize dual-pixel (DP) sensor phase information for various applications, such as depth estimation and deblurring. However, since DP image features are entirely determined by the camera hardware, DP-depth paired datasets are very scarce, especially when performing depth estimation on customized cameras. To overcome this, studies simulate DP images using ideal optical models. However, these simulations often violate real optical propagation laws, leading to poor generalization to real DP data. To address this, we investigate the domain gap between simulated and real DP data, and propose solutions using the Simulating DP Images from Ray Tracing (Sdirt) scheme. Sdirt generates realistic DP images via ray tracing and integrates them into the depth estimation training pipeline. Experimental results show that models trained with Sdirt-simulated images generalize better to real DP data. The code and collected datasets will be available at https://github.com/LinYark/Sdirt. + +# 1. Introduction + +The DP sensor [15, 29] is designed to split each pixel into left and right sub-pixels, utilizing microlenses and subpixels to achieve phase splitting. This allows the capture of a pair of images in one shot, known as DP images, as shown in Fig. 1(a). DP images are not only used for autofocus [7, 11, 31], but also improve performance in tasks such as deblurring [4, 17, 41] and depth estimation [9, 13, 23]. In the field of depth-from-dual-pixel (DfDP) estimation, the scarcity of DP-depth paired datasets stems from hardware limitations. To address this, many studies [2, 16, 18, 23, 26, 37] have focused on simulating DP images using RGBD datasets, which serve as more flexible and accessible data sources. + +The key to simulating DP images from RGBD datasets lies in accurately modeling the DP point spread function (PSF). In recent years, numerous model-based and calibration-based DP PSF simulators have been proposed. Calibration-based simulators [16, 18, 37] require substantial + +![](images/4421f6660aaa09a6026144f2cea6a0a00bc2cbc8836700851d369ae1e0f08e8f.jpg) +Figure 1. (a) Imaging process of a DP camera. The slight shifts between the left and right DP images caused by phase differences are illustrated with white dashed lines. (b) Example comparison between real and CoC-simulated DP PSFs, showing a significant difference between them. + +time calibrating real cameras to match DP PSFs, but suffer from issues such as interpolation errors due to discrete calibration points and the neglect of lens and DP sensor parameters, limiting model transferability. Model-based simulators [2, 23, 26] employ ideal optical or aberration analysis models to directly compute DP PSFs, bypassing calibration. However, these models often perform poorly on real DP images because they ignore lens aberrations and sensor phase splitting characteristics. As shown in Fig. 1(b), the ideal thin-lens model's simulated circle-of-confusion (CoC) DP PSF exhibits a significant domain gap with the real PSF, adversely affecting depth estimation. + +To bridge the domain gap between real DP images and DP images simulated by model-based simulators, we propose Sdirt. Sdirt consists of a ray-traced DP PSF simulator and a pixel-wise DP image rendering module, which can accurately generate DP images with aberration and phase information. Specifically, given a fixed focus lens and DP sensor system with known parameters, we use ray trac + +ing [35, 39, 40] to calculate the spatially varying DP PSFs and train a multilayer perceptron (MLP) to predict them. Subsequently, we convolve per-pixel DP PSFs with the all-in-focus (AiF) image to simulate the DP images captured by a real camera. After training a DfDP model with DP images simulated by Sdirt, the model can accurately estimate depth using optical aberration and phase information, enhancing its ability to generalize to real DP images. + +We conduct extensive experiments to validate the effectiveness of Sdirt. The experimental results demonstrate that, compared to other model-based simulators, the DP PSFs and DP images generated by Sdirt better resemble real data, and the DfDP model trained on Sdirt achieves superior generalization. In summary, our contributions are fourfold: + +- We propose a ray-traced DP PSF simulator that computes spatially varying DP PSFs, addressing the domain gap between simulated and real DP PSFs caused by lens aberrations and sensor phase splitting. +- We propose a pixel-wise DP image rendering module that uses an MLP to predict the DP PSF for each pixel, narrowing the gap between simulated and real DP images. +- Depth estimation results show that the DfDP model trained on Sdirt generalizes better to real DP images. +- We collected DP119, a real DP-depth paired test set with an open lens structure and fixed focus, featuring diverse real-world scenes. + +# 2. Related Work + +# 2.1. DP Datasets + +Most professional and mobile cameras have DP sensors, but only Canon cameras [1, 25, 26] and Google Pixel phones [9, 34, 43] offer DP data. In the field of depth estimation, supervised tasks rely on DP-depth paired datasets [1, 9, 26, 43]. However, changes in the lens, sensor, focal length, or aperture settings alter the PSFs, making it necessary to reacquire DP-depth data. Consequently, many researchers focus on simulating DP-depth datasets using RGBD data to reduce data collection costs and improve flexibility. + +# 2.2. DP PSF simulators + +Simulating DP images from RGBD datasets relies on DP PSFs, with recent progress in calibration-based [16, 18, 37] and model-based [2, 23, 26] DP PSF simulators. In calibration-based simulators, Xin et al. [37] acquired full-space DP PSFs by sampling discrete points in space and interpolating, while Li et al. [18] acquired them through size scaling using the CoC model. Furthermore, Li et al. [16] generated DP PSFs using a U-Net, which requires a large amount of DP-depth paired data. Calibration-based simulators are time-consuming and prone to interpolation and scaling errors, while model-based simulators circumvent these issues. Pan et al. [23] simplified the optical system to an + +ideal thin lens and used a symmetrically divided rectangular aperture to calculate DP PSFs. Punnappurath et al. [26] modeled the DP PSFs using a depletion kernel that is phase symmetric, meaning the left and right PSFs of the same object point are flip symmetric. Abuolaim et al. [2] used aberration analysis models to calculate DP PSFs, but the resulting patterns still remained phase symmetric. These studies oversimplify optical propagation by neglecting aberrations and phase splitting, leading to unrealistic domain gaps. + +# 2.3. Applications driven by DP data + +In recent years, DP data-driven applications such as depth estimation [9, 10, 13, 14, 23, 24, 43], deblurring [3, 4, 17, 27, 38, 41, 42], and refocusing [5, 7, 11, 32] have undergone rapid development. Additionally, Kang et al. [13] utilized DP data for facial scanning, similar to depth estimation tasks. Wadhwa et al. [34] employed DP data for synthetic shallow depth-of-field imaging, a popular recent application in mobile photography. Shi et al. [28] integrated DP hardware with structured diffractive optical elements for complex image encoding and high-precision multimodal reconstruction. In addition, DP data can be used for disparity estimation [22, 36], rain removal [18], and reflection removal [25]. It is foreseeable that DP data-driven applications will continue to emerge in various new fields. + +# 3. Method of Sdirt + +To provide the required input for training the DfDP network, Sdirt generates realistic DP images through ray tracing. As shown in Fig. 2(c), the module achieves pixel-wise rendering of DP images by convolving the DP PSF corresponding to each point in the depth map with the AiF RGB image. The DP PSF for each point is inferred by a pretrained MLP network. As illustrated in Fig. 2(b), the offline training process of the MLP network uses ray-traced DP PSFs (Fig. 2(a)) as ground truth (GT). + +# 3.1. Ray-traced DP PSF simulator + +As shown in Fig. 2(a), lens ray tracing can accurately obtain the landing point on the sensor surface [8, 35, 39, 40]. We construct a ray set $A$ containing $n$ rays, each originating from an object point $p$ . The aperture stop's image in object space is treated as the entrance pupil, from which $n$ points are densely sampled. Each ray in $A$ is initially directed from $p$ to its corresponding entrance pupil sample. As rays in $A$ traverse each lens surface, their positions and directions are updated by refraction according to Snell's law and the lens parameters. After multiple refractions, the rays reach the sensor plane. We denote the sensor-plane landing points as $O$ and the corresponding ray directions as $D$ . + +DP sensor ray tracing is limited because camera manufacturers do not disclose the structural parameters of the microlens and sub-pixel components within the DP pixel. + +![](images/93ab78c423775cc9cdd65314e09bd144dcac0bf9c2a6709b6aeed6c296cad114.jpg) +(a) +Figure 2. Simulating Dual-Pixel Images from Ray Tracing pipeline. (a) Ray-traced DP PSF simulator. Calculates spatially varying DP PSFs for lens and DP sensor through ray tracing. (b) DP PSF prediction network. Trains an MLP network to predict DP PSFs, using the ray-traced DP PSFs as GT. (c) Pixel-wise DP image rendering module. The network predicts the DP PSFs for all points in the depth map (red pass). Then, each DP PSF is convolved with the AiF RGB image to render the simulated DP image (blue pass). + +As a result, we simplify the DP pixel structure (Fig. 3(a)) based on past research [15, 29]. We model the microlens as a thin lens with radius $r$ and focal length $f$ , and define $h$ as distance between the sub-pixel and the microlens, $w$ as sub-pixel width, and $ps$ as DP pixel size. + +The DP PSF can be obtained by separately calculating the cumulative integration of the rays for each sub-pixel. We analyze which sub-pixel the $k$ -th ray $A_{k}$ in the ray set $A$ ultimately enters. Through lens ray tracing, we obtain the landing point $O_{k}(x_{k},y_{k},z_{k})$ and direction $D_{k}(x_{k}^{\prime},y_{k}^{\prime},z_{k}^{\prime})$ of the $k$ -th ray $A_{k}$ on the sensor surface. Based on the landing point $O_{k}$ , we can easily calculate that $A_{k}$ lies within the DP pixel $(i,j)$ , which has surface center coordinates $(x_{i},y_{i})$ . For $O_{k}$ , the situation differs depending on whether it lies within or outside the microlens: + +When $O_{k}$ lies within the microlens, as shown in Fig. 3(b), $A_{k}$ is refracted by the microlens and then enters a sub-pixel. According to geometric optics theory, $A_{k}$ and the principal ray incident on the thin lens at the same angle both converge on the focal plane of the thin lens. We only need to focus on a few boundary lines $x_{L1}, x_{M1}, x_{R1}$ to determine which sub-pixel $A_{k}$ ultimately enters. If $x_{k}$ is in the interval $[x_{L1}, x_{M1}]$ , it enters the left sub-pixel; if in the interval $[x_{M1}, x_{R1}]$ , it enters the right sub-pixel; otherwise, it is considered a missing ray. The calculation methods for the + +boundary lines are as follows: + +$$ +x _ {L 1} = x _ {i} + w - (f * \tan \theta - w) * h / (f - h) +$$ + +$$ +x _ {M 1} = x _ {i} - (f * \tan \theta) * h / (f - h) \tag {1} +$$ + +$$ +x _ {R 1} = x _ {i} - w - (f * \tan \theta + w) * h / (f - h) +$$ + +Where $\tan \theta$ is the tangent of the angle at which $A_{k}$ lands on the sensor surface, equivalent to $x_{k}^{\prime} / z_{k}^{\prime}$ . We provide the derivation hints for Eq. (1) in Fig. 4(a). + +When $O_{k}$ lies outside the microlens, as shown in Fig. 3(c), $A_{k}$ enters the sub-pixel directly without refraction. Similarly, checking boundary lines $x_{L2}, x_{M2}, x_{R2}$ suffices to determine which sub-pixel $A_{k}$ ultimately enters: + +$$ +x _ {L 2} = x _ {i} + w - h * \tan \theta +$$ + +$$ +x _ {M 2} = x _ {i} - h * \tan \theta \tag {2} +$$ + +$$ +x _ {R 2} = x _ {i} - w - h * \tan \theta +$$ + +The left PSF $(PSF_{L})$ can be calculated by integrating the set of rays $A$ emitted from point $p$ over all left subpixels: + +$$ +P S F _ {L} (i, j) = \sum_ {k = 1} ^ {n} A _ {k} * \delta_ {L, k} (i, j) \tag {3} +$$ + +where $\delta_{L,k}(i,j)$ represents the energy distribution of ray $A_{k}$ in the left sub-pixel at DP pixel $(i,j)$ . The ray energy distribution is assumed to be a unit impulse, which + +![](images/ae6b7be3ee796de1178f432f1afcb6e403ff915397ef5d0386f7b604cf2c2d5c.jpg) +Figure 3. (a) DP pixel structure layout. The left part is a right-side view, and the right part is a front perspective view. (b) On hitting the microlens, the $k$ -th ray lands on left/right sub-pixel in red/blue interval, else missed. (c) Without hitting the microlens, the $k$ -th ray lands on left/right sub-pixel in red/blue interval, else missed. + +equals zero unless it eventually enters the left sub-pixel of DP pixel $(i,j)$ . By following the same calculation steps as for $PSF_{L}$ , $PSF_{R}$ can also be computed. + +# 3.2. Pixel-wise DP image rendering module + +As shown in Fig. 2(b), we reduce the computational cost of ray-traced DP PSFs by training an MLP network to predict them, inspired by [33, 39]. The network input is the normalized coordinate $p$ within the camera's valid imaging region (Fig. 4(b)), which is a frustum defined by the field of view, sensor size, and the preset minimum and maximum depths. We normalize the $(x,y)$ coordinates to $[-1,1]$ . After fixing the focus distance, we set the DP PSF kernel size $(ks)$ large enough to ensure that the DP PSF for any point within the valid imaging region can be fully displayed. The network consists of 5 hidden layers, each containing 512 neurons, and an output layer with $2 * ks^2$ neurons. We adopt the L2 loss to supervise predictions $P\hat{S} F_L$ and $P\hat{S} F_R$ : + +$$ +L o s s = L _ {2} (P \hat {S} F _ {L}, P S F _ {L}) + L _ {2} (P \hat {S} F _ {R}, P S F _ {R}) \quad (4) +$$ + +During training, we use ray-traced DP PSFs as GT and apply max normalization to alleviate learning difficulties caused by large radii PSFs. During inference, we apply + +![](images/c0310068b5819416d0a4c7d18005ee98a792dfcec67246f15baf43eb512f6c8e.jpg) +Figure 4. (a) Derivation hints for Eq. (1). (b) The valid imaging region is a frustum, and we normalize the $(x,y)$ coordinates to $[-1,1]$ . (c) The DfDP model takes DP image as input to predict the depth map. (d) During cost volume generation in [6], we stack original disparity (green arrows) and add reverse disparity (blue arrows), with $d_{max}$ as max displacement. + +sum normalization to the predicted DP PSFs, approximating the uniform intensity distribution in cameras with vignetting compensation. + +Because the DP PSF is shift-variant, we convolve each pixel's DP PSF with the AiF image to simulate the DP image captured by a camera. The RGBD dataset [21, 30] provides the paired RGB image $(I_{RGB})$ and depth map $(I_D)$ . As shown in Fig. 2(c), we treat each pixel in $I_D$ as an object point and use the trained MLP to predict its DP PSF. Using $I_{RGB}$ as the AiF image, we apply pixel-wise local convolution [39] to render the DP image containing aberration and phase information. + +# 4. DfDP Model Based on Sdtirt + +As shown in Fig. 4(c), the input to the DfDP model is provided in real time by the pixel-wise DP image rendering module of Sdirt. At each training iteration, Sdirt feeds the DfDP model with simulated DP images of shape $(B,6,H,W)$ , where $B$ is the batch size, the number of channels is 6, and $H, W$ are the DP image height and width. The model outputs the predicted depth maps $(\hat{I}_D)$ of shape $(B,1,H,W)$ . We apply the L1 loss to supervise $\hat{I}_D$ : + +$$ +L o s s = L _ {1} \left(\hat {I _ {D}}, I _ {D}\right) \tag {5} +$$ + +We select [6] as the DfDP model and make reasonable adjustments to its cost volume step to accommodate the DP data. In binocular images, disparity from points at different depths always has the same direction, thus the cost volume generation process in [6] is also unidirectional, as shown by + +![](images/b234044e6e568ce38fd4dd4a13c7ec7e60bac25d74f5e9d72763dc3e89c643ad.jpg) +Figure 5. Qualitative results of simulated DP PSFs. Evaluate the real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) F/1.8 DP PSFs at two depths $(0.5\mathrm{m}$ and $1.5\mathrm{m})$ and three different positions. As the object point $p$ moves further from the optical axis, the real $PSF_{L}$ and $PSF_{R}$ become more phase asymmetric, and aberrations increase. Existing simulators neglect aberrations and DP phase splitting, causing a large gap between simulated and real DP PSFs. Only our ray-traced simulator predicts realistic results at all depths and positions. + +the green arrows in Fig. 4(d). However, in DP images, disparity from object points before and after the focus distance is opposite in direction. Therefore, we add blue arrows to extend the cost volume generation, in order to enhance the DfDP model's ability to capture reverse disparity. + +# 5. S skirt Representation Result + +# 5.1. Implementation details + +Sdirt employs a Canon RF50mm lens and an R6 Mark II camera body as the real camera. The simulated DP sensor has dimensions of $24\mathrm{mm} \times 36\mathrm{mm}$ with a resolution of $512 \times 768$ . We set the focus distance of both the simulated and real cameras to $1\mathrm{m}$ , the valid imaging range to $0.5 - 20\mathrm{m}$ , the F-number to $F/4$ , and the ray set $A$ to contain 4096 rays. During MLP training, $ks$ is set to 21 to fully display the DP PSF of any point within the valid imaging region. We train the MLP for 100,000 iterations, selecting 128 random points within the valid imaging region during each iteration. We conducted our experiments on a 12700K CPU and a 3090 GPU. + +Following [39], we assume chromatic aberration is well corrected compared to other aberrations and defocus effects, allowing us to use a $550\mathrm{nm}$ wavelength to reduce ray tracing costs. For details on obtaining the structural parameters of the DP pixel, please refer to the Supplementary. + +# 5.2. Comparison methods + +For comparison, we select all existing model-based DP PSF simulators: L2R [2], Modeling [26], DDDNet [23], and an added CoC simulator. Although [24] is a more recent work, its simulator fully reuses DDDNet [23], so we do not include it for comparison. Calibration-based simulators and constraint-training methods are data-driven and require real DP-depth paired data, making them fundamentally different from model-based simulators, so no comparison is made. We directly used the source code of these DP PSF simulators [2, 26], and re-implemented the simulator [23] which only provided executable files. To ensure fairness, we set the focus distance, F-number, and sensor size in each simulator to be consistent with ours. + +# 5.3. Evaluation on simulated DP PSFs + +To facilitate observation of aberration and phase information, we temporarily set the F-numbers of both the simulated and real cameras to F/1.8. We capture a OLED screen displaying a bright pixel to obtain the real DP PSFs. + +As shown in Fig. 5, we evaluate the DP PSF $(PSF_{L}$ on the left, $PSF_{R}$ on the right) of object points $p$ at various depths and positions within the valid imaging region. Observing the real DP PSFs, it can be found that defocus (with focus distance of $1\mathrm{m}$ ) causes a significant phase difference + +![](images/6f04a9898785919d648bab907418b11cf4a868e14099cf0478c7670c820c9399.jpg) +Figure 6. Qualitative results of simulated DP images. Evaluate the similarity between simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) and real F/4 defocused left DP images at two depths (0.5 m and 2 m). Compared to real F/4 defocused images, images simulated by other methods exhibit varying sizes of patterns, incomplete shapes, and texture shifts in different directions before and after the focus distance (1 m). Our method produces the most realistic simulated images. + +between the real $PSF_{L}$ and $PSF_{R}$ , and the phase difference at depths of $0.5\mathrm{m}$ and $1.5\mathrm{m}$ is opposite. When $p$ is on the optical axis, all other simulators fail to provide DP PSF results close to the real ones. Farther from the optical axis, the real $PSF_{L}$ and $PSF_{R}$ become phase asymmetric, with smaller phase differences and more off-axis aberrations. The differences between DP PSFs from other simulators and the real ones become more apparent. In contrast, our ray-traced simulator predicts accurate results at all depths and positions. + +We present the quantitative results in Tab. 1, using OpenCV's normalized squared difference (NSD) and normalized cross-correlation (NCC) to quantify the error and similarity between the real and simulated DP PSFs. We sampled 50 DP PSFs at $0.5\mathrm{m}$ and $1.5\mathrm{m}$ within the valid imaging region. Since the real DP PSF has flip symmetry (not phase symmetry) about both the $x$ and $y$ axes, we performed uniform sampling in the first quadrant. Our method achieves the highest similarity (0.915 NCC) and lowest error (0.133 NSD), demonstrating its effectiveness in bridging the domain gap between simulated and real DP PSFs introduced by lens aberrations and sensor phase splitting. + +In all subsequent experiments, the F-number is adjusted back to F/4 due to the large blur kernel size caused by defocus at F/1.8, which resulted in GPU memory shortages. + +Table 1. Quantitative results of simulated DP PSFs. The NCC and NSD metrics between simulated and real DP PSFs at 50 points were evaluated. Our method achieved the highest similarity and lowest error. + +
MethodOursCoCL2R [2]Modeling [26]DDDNet [23]
NCC↑0.9150.6720.6380.7070.589
NSD↓0.1330.4480.5230.4230.625
+ +# 5.4. Evaluation on simulated DP images + +When comparing simulated and real DP images, to assess the degradation of simulated DP images caused by defocus effects at different depths, we captured richly textured planar scenes from $0.5\mathrm{m}$ to infinity, perpendicular to the optical axis. Shallower depths were sampled more densely to better capture defocus variations. This resulted in 56 scenes with varying depths, including 11 infinite-depth scenes. For each scene, real DP images were captured at F/4 and F/20. The F/20 image served as the AiF RGB image, and its capture-time depth was assigned to the entire depth map. The F/20 RGB-depth pairs were then input into DP simulators to generate the simulated F/4 defocused DP images. + +As shown in Fig. 6, we compare the qualitative results of the defocused left DP images at depths of $0.5\mathrm{m}$ and $2\mathrm{m}$ . Due to errors in simulating the DP PSF, other methods produce simulated images with textures shifted to the right at + +![](images/e7d1b9ebb9ac384c1bf71c1d83fb225112b33e28f9a0ea1d26274a2509e6ad1d.jpg) +Figure 7. Quantitative results of simulated DP images. Evaluate the similarity of 56 real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) planar scene F/4 DP images at different depths using PSNR $(\uparrow)$ and SSIM $(\uparrow)$ metrics. As the depth of simulated images deviates further from the focus distance $(1\mathrm{m})$ , the less realistic the other methods become, whereas our scheme maintains the highest accuracy across all depths. + +![](images/9dc4a618d98703cedecee41e77b587735f70ce1dfee490cddf9c457d0fdc2e3d.jpg) + +$0.5\mathrm{m}$ depth and to the left at $2\mathrm{m}$ depth, compared to the real left DP image. Moreover, other methods show variations in image texture sizes and incomplete shapes. However, the DP images simulated by our Sdirt method show minimal deviation in texture details. + +As shown in Fig. 7, which presents the quantitative results, we assess the similarity between simulated and real images using PSNR and SSIM. As the simulated image's depth moves further away from the focus distance (1 m), the neglect of aberrations and phase splitting in other methods leads to a more significant decrease in their metrics. In contrast, our Sdirt method achieves the best results at all depths, with average PSNR/SSIM values of 37.1982/0.9845, indicating that our simulated DP images are highly realistic. + +# 6. DfDP Estimation Results + +# 6.1. Implementation details + +To train the DfDP model, we simulate F/4 DP images using Sdirt on RGBD datasets [21, 30]. The NYU Depth dataset [30], preprocessed by [12], contains 50,688 indoor scenes. We adopt the AdamW optimizer [19] and CosineAnnealing scheduler [20] with an initial learning rate of $1 \times 10^{-4}$ . The model is trained with a batch size of 4 for 50 epochs. In each epoch, 2,000 randomly selected DP-depth pairs are used for training. After training, the model is directly evaluated on a real-world test set without any fin-tuning to assess its generalization ability. + +# 6.2. A new real-world test set DP119 + +Unlike previous simulation methods, S skirt addresses data scarcity by simulating hardware, requiring consistent camera and setup during training and inference. However, as shown in Tab. 2, no public dataset satisfies all of the following conditions: 1. Real DP-depth paired data for simulating DP images; 2. A known lens structure for ray tracing; 3. A fixed focus distance, since changing it alters the DP PSF of the same object point. This makes evaluation on public datasets with unknown hardware impractical. We there- + +fore collected the DP119 test set, which contains diverse test scenes. + +Table 2. Summary of existing DP datasets. + +
DPDD [1]L2R [2]DPNet [9]DP5K [16]DDDNet [23]Modeling [26]DP119 Ours
Real captured✓×✓✓×✓✓
Paired depth×✓✓✓✓×✓
Lens structure✓-×✓-×✓
Fixed focus dist.✓-××-×✓
+ +DP119 consists of 45 planar scenes, 44 box scenes, and 30 casual scenes, totaling 119 scenes. The casual scenes represent common scenarios, suitable for evaluating the robustness of simulation models. The planar and box scenes are richly textured, helping the model utilize aberration and phase cues for depth estimation. Textureless areas, even when defocused, lack aberrations or phase differences, making depth estimation difficult and evaluation unreliable. Thus, richly textured planar and box scenes are ideal for assessing domain gaps between simulated and real DP images. Real DP images are captured by the Canon RF50mm lens with a R6 Mark II at F/4 and $1\mathrm{m}$ focus. GT depth maps for planar scenes are created by the depths at the capture time, while those for box and casual scenes are obtained from LiDAR scans. For more information about the DP119 test set, please refer to the Supplementary. + +# 6.3. Evaluation + +We present qualitative and quantitative results for the DfDP model trained by CoC, L2R [2], Modeling [26], DDDNet [23], and our Sdirt. For a fair comparison, we deploy their DfDP models using their DP PSF simulator source code [2, 26] and re-implemented code [23], and have all models share the same depth estimation network architecture, initialize from a common CoC pre-trained checkpoint, and follow the same implementation details. + +Figure 8 shows depth estimation results for four casual scenes, with a color bar in meters for visualization. It can be seen that although other methods provide some relative positional information in the center region, they fail to correctly estimate relative positional relationships in the edge regions, and both the center and edge regions have large absolute positional errors. This is because the DP PSFs predicted by these methods exhibit phase symmetry and shift-invariance, whereas real DP PSFs exhibit phase asymmetry, aberrations, and shift-variance. Their methods' mismatch with real DP PSFs leads to significant depth estimation errors, especially near the edges. The results of L2R [2] and Modeling [26] are similar because the peak positions of their DP PSFs are quite close. In contrast, our method provides accurate relative positional information across the entire image area with minimal absolute positional errors, as our simulated DP PSFs are highly consistent with the real ones in both the center and edge regions. + +![](images/d184cd67a6648a800ad3c990640da5270722ca35992508579011ff2c887b3397.jpg) +Figure 8. Qualitative results of absolute depth estimation on DP119. Evaluate DfDP models with CoC, L2R [2], Modeling [26], DDDNet [23] and our Sdirt at four casual scenes. Each result image includes a color bar showing depth in meters. Their depth estimation results show partial accuracy in relative positional relationships but large absolute positional errors. Our depth estimation results, however, demonstrate high accuracy in both relative and absolute positions with minimal errors. (Best viewed in color and enlarge on screen.) + +Table 3. Quantitative results of absolute depth estimation on DP119. The DfDP model trained with our Sdirt achieves superior performance across most scenarios and evaluation metrics. + +
SceneMethodMAE↓MSE↓Abs.r↓Sq.r↓Acc-1↑Acc-2↑
PlanarOurs0.08450.01090.08710.00950.98490.9997
CoC0.20850.10010.18010.06590.66700.8990
L2R [2]0.24180.12710.21120.08410.63190.8536
Modeling [26]0.22840.11420.20040.07660.64960.8725
DDDNet [23]0.25830.14850.21910.09580.56480.8089
BoxOurs0.11970.03390.09060.02310.94740.9812
CoC0.33750.18040.24420.11160.44120.8277
L2R [2]0.38660.22840.28030.14120.36510.7156
Modeling [26]0.36550.20550.26600.12780.39070.7758
DDDNet [23]0.41770.26760.29750.16360.34560.6274
CasualOurs0.27020.22940.46320.72410.82360.9314
CoC0.79251.85790.54610.68210.33180.6103
L2R [2]0.81701.74870.55970.67190.27600.5315
Modeling [26]0.79341.72560.55100.66550.29780.5732
DDDNet [23]0.89312.06240.57520.71350.24810.4685
+ +We evaluate all models on the DP119 test set using the following metrics to assess depth estimation performance: mean absolute error (MAE), mean squared error (MSE), absolute relative error (Abs.r.), squared relative error (Sq.r.), accuracy with $\delta < 1.25$ (Acc-1) and accuracy with $\delta < 1.25^2$ (Acc-2). Since the planar and box scenes are richly textured and do not have interference from textureless areas, the simulation models can rely on abundant aberration and phase difference cues for depth estimation. As shown in Tab. 3, in the simplest planar scene, our model achieves the best results across all metrics (0.9849 Acc-1). In the box scenes, all models experience a performance drop compared to the planar scenes, but we still achieve the optimal metrics (0.9474 Acc-1). These results demonstrate that our model is highly realistic and effectively bridges the domain gap between simulated and real DP images. Furthermore, we test the robustness of all models on the casual scenes. Due to + +the textureless areas in the casual scenes, the performance of all models significantly degrades. However, our model still achieves the best metrics (0.8236 Acc-1), far surpassing the second-best model (0.3318 Acc-1). This further demonstrates that our model is the most realistic and has the best generalization performance. For more sample results from the DP119 test set, please refer to the Supplementary. + +# 7. Conclusion and Discussion + +Simulated DP images help address the scarcity of DP-depth paired data but face a domain gap between simulated and real DP data. In this work, we propose a novel simulation framework called Sdirt to bridge this gap. Specifically, we calculate the DP PSF for points in object space using ray tracing, and employ a network to predict them. Then, we render DP images based on the predicted DP PSFs. Experimental results show that the proposed Sdirt can simulate more realistic DP data. Moreover, depth estimation models trained based on Sdirt generalize better to real DP images. + +We believe that Sdirt is not limited to DfDP tasks, it can provide extra depth information for any task with known optical imaging system parameters, promising significant applications in scenarios such as smartphones, automobiles, and microscopes in the future. However, Sdirt is only applicable to cameras equipped with a fixed-focus lens (the structure must be open) and a DP sensor (DP images must be available). Currently, only Canon (5D4, R series) meets these requirements. To further expand the application of Sdirt, more camera and smartphone manufacturers need to make these data accessible. + +# Acknowledgments + +This work was supported by National Natural Science Foundation of China (Grant No. 32471146) and the project N20240194. The authors thank Echosom, Miya, and Xinge for valuable discussions and assistance. + +# References + +[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 111-126. Springer, 2020. 2, 7 +[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 2, 5, 6, 7, 8 +[3] Abdullah Abuolaim, Radu Timofte, and Michael S Brown. Ntire 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 578-587, 2021. 2 +[4] Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2 +[5] Hadi Alzayer, Abdullah Abuolaim, Leung Chun Chan, Yang Yang, Ying Chen Lou, Jia-Bin Huang, and Abhishek Kar. Dc2: Dual-camera defocus control by learning to refocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21488-21497, 2023. 2 +[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 4 +[7] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1, 2 +[8] Geoffroi Côté, Fahim Mannan, Simon Thibault, Jean-François Lalonde, and Felix Heide. The differentiable lens: Compound lens search over glass surfaces and materials for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20803-20812, 2023. 2 +[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7628-7637, 2019. 1, 2, 7 +[10] Bhargav Ghanekar, Salman Siddique Khan, Pranav Sharma, Shreyas Singh, Vivek Boominathan, Kaushik Mitra, and + +Ashok Veeraraghavan. Passive snapshot coded aperture dual-pixel rgb-d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25348-25357, 2024. 2 +[11] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 1, 2 +[12] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 7 +[13] Minjun Kang, Jaesung Choe, Hyowon Ha, Hae-Gon Jeon, Sunghoon Im, In So Kweon, and Kuk-Jin Yoon. Facial depth and normal estimation using single dual-pixel camera. In European Conference on Computer Vision, pages 181-200. Springer, 2022. 1, 2 +[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 2 +[15] Masahiro Kobayashi, Michiko Johnson, Yoichi Wada, Hiromasa Tsuboi, Hideaki Takada, Kenji Togo, Takafumi Kishi, Hidekazu Takahashi, Takeshi Ichikawa, and Shunsuke Inoue. A low noise and high sensitivity image sensor with imaging and phase-difference detection of all pixels. ITE Transactions on Media Technology and Applications, 4(2):123-128, 2016. 1, 3 +[16] Feiran Li, Heng Guo, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Learning to synthesize photorealistic dual-pixel images from rgbd frames. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2023. 1, 2, 7 +[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 1, 2 +[18] Yizhou Li, Yusuke Monno, and Masatoshi Okutomi. Dual-pixel raindrop removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[19] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[20] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7 +[21] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 4, 7 +[22] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. In 2024 International Conference on 3D Vision (3DV), pages 675-684. IEEE, 2024. 2 + +[23] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 5, 6, 7, 8 +[24] Liyuan Pan, Richard Hartley, Liu Liu, Zhiwei Xu, Shah Chowdhury, Yan Yang, Hongguang Zhang, Hongdong Li, and Miaomiao Liu. Weakly-supervised depth estimation and image deblurring via dual-pixel sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5 +[25] Abhijith Punnappurath and Michael S Brown. Reflection removal using a dual-pixel sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1556-1565, 2019. 2 +[26] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 5, 6, 7, 8 +[27] Lingyan Ruan, Martin Balint, Mojtaba Bemana, Krzysztof Wolski, Hans-Peter Seidel, Karol Myszkowski, and Bin Chen. Self-supervised video defocus deblurring with atlas learning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[28] Zheng Shi, Ilya Chugunov, Mario Bijelic, Geoffroi Côté, Jiwoon Yeom, Qiang Fu, Hadi Amata, Wolfgang Heidrich, and Felix Heide. Split-aperture 2-in-1 computational cameras. ACM Transactions on Graphics (TOG), 43(4):1-19, 2024. 2 +[29] Eun Sub Shim, Kyungho Lee, Junghyung Pyo, Wooseok Choi, Jungbin Yun, Taesub Jung, Kyungduck Lee, Seyoung Kim, Chanhee Lee, Seungki Baek, et al. All-directional dual pixel auto focus technology in cmos image sensors. In 2021 Symposium on VLSI Technology, pages 1-2. IEEE, 2021. 1, 3 +[30] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 7 +[31] Przemysław Śliwiński and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 1 +[32] Kunal Swami. Adjust your focus: Defocus deblurring from dual-pixel images using explicit multi-scale cross-correlation. In International Conference on Computer Vision and Image Processing, pages 318-330. Springer, 2023. 2 +[33] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 4 +[34] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. + +Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2 +[35] Congli Wang, Ni Chen, and Wolfgang Heidrich. do: A differentiable engine for deep lens design of computational imaging systems. IEEE Transactions on Computational Imaging, 8:905-916, 2022. 2 +[36] Zhuofeng Wu, Doehyung Lee, Zihua Liu, Kazunori Yoshizaki, Yusuke Monno, and Masatoshi Okutomi. Disparity estimation using a quad-pixel sensor. arXiv preprint arXiv:2409.00665, 2024. 2 +[37] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2 +[38] Hao Yang, Liyuan Pan, Yan Yang, Richard Hartley, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24078-24087, 2024. 2 +[39] Xinge Yang, Qiang Fu, Mohamed Elhoseiny, and Wolfgang Heidrich. Aberration-aware depth-from-focus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 2, 4, 5 +[40] Xinge Yang, Qiang Fu, and Wolfgang Heidrich. Curriculum learning for ab initio deep learned refractive optics. Nature communications, 15(1):6572, 2024. 2 +[41] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13263-13272, 2023. 1, 2 +[42] Dafeng Zhang and Xiaobing Wang. Dynamic multi-scale network for dual-pixel images defocus deblurring with transformer. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 2 +[43] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du 2 net: Learning depth estimation from dual-cameras and dual-pixels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 582-598. Springer, 2020. 2 \ No newline at end of file diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/images.zip b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9693de0c1085ca4f6d07fad9f36f350613ac987a --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90ae5311ee1f422703f2267ae592834382d83223f4c47aff66ef5e0f40172386 +size 834174 diff --git a/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/layout.json b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dca37dee3268a3e49d8f82a95031aa542fd4a1c9 --- /dev/null +++ b/2025/Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation/layout.json @@ -0,0 +1,8248 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "text", + "content": "Simulating Dual-Pixel Images From Ray Tracing For Depth Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 143, + 470, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 143, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 139, + 143, + 470, + 172 + ], + "type": "text", + "content": "Fengchen He, Dayang Zhao, Hao Xu, Tingwei Quan, Shaoqun Zeng Huazhong University of Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 174, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 174, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 174, + 482, + 185 + ], + "type": "text", + "content": "{linyark, dayangzhao, xuhao_2003, quantingwei, sqzeng}@hust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 453 + ], + "type": "text", + "content": "Many studies utilize dual-pixel (DP) sensor phase information for various applications, such as depth estimation and deblurring. However, since DP image features are entirely determined by the camera hardware, DP-depth paired datasets are very scarce, especially when performing depth estimation on customized cameras. To overcome this, studies simulate DP images using ideal optical models. However, these simulations often violate real optical propagation laws, leading to poor generalization to real DP data. To address this, we investigate the domain gap between simulated and real DP data, and propose solutions using the Simulating DP Images from Ray Tracing (Sdirt) scheme. Sdirt generates realistic DP images via ray tracing and integrates them into the depth estimation training pipeline. Experimental results show that models trained with Sdirt-simulated images generalize better to real DP data. The code and collected datasets will be available at https://github.com/LinYark/Sdirt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 135, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "content": "The DP sensor [15, 29] is designed to split each pixel into left and right sub-pixels, utilizing microlenses and subpixels to achieve phase splitting. This allows the capture of a pair of images in one shot, known as DP images, as shown in Fig. 1(a). DP images are not only used for autofocus [7, 11, 31], but also improve performance in tasks such as deblurring [4, 17, 41] and depth estimation [9, 13, 23]. In the field of depth-from-dual-pixel (DfDP) estimation, the scarcity of DP-depth paired datasets stems from hardware limitations. To address this, many studies [2, 16, 18, 23, 26, 37] have focused on simulating DP images using RGBD datasets, which serve as more flexible and accessible data sources." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "The key to simulating DP images from RGBD datasets lies in accurately modeling the DP point spread function (PSF). In recent years, numerous model-based and calibration-based DP PSF simulators have been proposed. Calibration-based simulators [16, 18, 37] require substantial" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 215, + 555, + 400 + ], + "blocks": [ + { + "bbox": [ + 320, + 215, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 320, + 215, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 320, + 215, + 555, + 400 + ], + "type": "image", + "image_path": "4421f6660aaa09a6026144f2cea6a0a00bc2cbc8836700851d369ae1e0f08e8f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. (a) Imaging process of a DP camera. The slight shifts between the left and right DP images caused by phase differences are illustrated with white dashed lines. (b) Example comparison between real and CoC-simulated DP PSFs, showing a significant difference between them." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 628 + ], + "type": "text", + "content": "time calibrating real cameras to match DP PSFs, but suffer from issues such as interpolation errors due to discrete calibration points and the neglect of lens and DP sensor parameters, limiting model transferability. Model-based simulators [2, 23, 26] employ ideal optical or aberration analysis models to directly compute DP PSFs, bypassing calibration. However, these models often perform poorly on real DP images because they ignore lens aberrations and sensor phase splitting characteristics. As shown in Fig. 1(b), the ideal thin-lens model's simulated circle-of-confusion (CoC) DP PSF exhibits a significant domain gap with the real PSF, adversely affecting depth estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "To bridge the domain gap between real DP images and DP images simulated by model-based simulators, we propose Sdirt. Sdirt consists of a ray-traced DP PSF simulator and a pixel-wise DP image rendering module, which can accurately generate DP images with aberration and phase information. Specifically, given a fixed focus lens and DP sensor system with known parameters, we use ray trac" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26106" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ing [35, 39, 40] to calculate the spatially varying DP PSFs and train a multilayer perceptron (MLP) to predict them. Subsequently, we convolve per-pixel DP PSFs with the all-in-focus (AiF) image to simulate the DP images captured by a real camera. After training a DfDP model with DP images simulated by Sdirt, the model can accurately estimate depth using optical aberration and phase information, enhancing its ability to generalize to real DP images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 240 + ], + "type": "text", + "content": "We conduct extensive experiments to validate the effectiveness of Sdirt. The experimental results demonstrate that, compared to other model-based simulators, the DP PSFs and DP images generated by Sdirt better resemble real data, and the DfDP model trained on Sdirt achieves superior generalization. In summary, our contributions are fourfold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 294, + 383 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 241, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 294, + 289 + ], + "type": "text", + "content": "- We propose a ray-traced DP PSF simulator that computes spatially varying DP PSFs, addressing the domain gap between simulated and real DP PSFs caused by lens aberrations and sensor phase splitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 289, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 294, + 324 + ], + "type": "text", + "content": "- We propose a pixel-wise DP image rendering module that uses an MLP to predict the DP PSF for each pixel, narrowing the gap between simulated and real DP images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 324, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 294, + 348 + ], + "type": "text", + "content": "- Depth estimation results show that the DfDP model trained on Sdirt generalizes better to real DP images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 348, + 294, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 383 + ], + "type": "text", + "content": "- We collected DP119, a real DP-depth paired test set with an open lens structure and fixed focus, featuring diverse real-world scenes." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 396, + 142, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 142, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 142, + 409 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 417, + 135, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 135, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 135, + 428 + ], + "type": "text", + "content": "2.1. DP Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 434, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 543 + ], + "type": "text", + "content": "Most professional and mobile cameras have DP sensors, but only Canon cameras [1, 25, 26] and Google Pixel phones [9, 34, 43] offer DP data. In the field of depth estimation, supervised tasks rely on DP-depth paired datasets [1, 9, 26, 43]. However, changes in the lens, sensor, focal length, or aperture settings alter the PSFs, making it necessary to reacquire DP-depth data. Consequently, many researchers focus on simulating DP-depth datasets using RGBD data to reduce data collection costs and improve flexibility." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 167, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 167, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 167, + 563 + ], + "type": "text", + "content": "2.2. DP PSF simulators" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Simulating DP images from RGBD datasets relies on DP PSFs, with recent progress in calibration-based [16, 18, 37] and model-based [2, 23, 26] DP PSF simulators. In calibration-based simulators, Xin et al. [37] acquired full-space DP PSFs by sampling discrete points in space and interpolating, while Li et al. [18] acquired them through size scaling using the CoC model. Furthermore, Li et al. [16] generated DP PSFs using a U-Net, which requires a large amount of DP-depth paired data. Calibration-based simulators are time-consuming and prone to interpolation and scaling errors, while model-based simulators circumvent these issues. Pan et al. [23] simplified the optical system to an" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "ideal thin lens and used a symmetrically divided rectangular aperture to calculate DP PSFs. Punnappurath et al. [26] modeled the DP PSFs using a depletion kernel that is phase symmetric, meaning the left and right PSFs of the same object point are flip symmetric. Abuolaim et al. [2] used aberration analysis models to calculate DP PSFs, but the resulting patterns still remained phase symmetric. These studies oversimplify optical propagation by neglecting aberrations and phase splitting, leading to unrealistic domain gaps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 186, + 484, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 484, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 484, + 199 + ], + "type": "text", + "content": "2.3. Applications driven by DP data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 204, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 372 + ], + "type": "text", + "content": "In recent years, DP data-driven applications such as depth estimation [9, 10, 13, 14, 23, 24, 43], deblurring [3, 4, 17, 27, 38, 41, 42], and refocusing [5, 7, 11, 32] have undergone rapid development. Additionally, Kang et al. [13] utilized DP data for facial scanning, similar to depth estimation tasks. Wadhwa et al. [34] employed DP data for synthetic shallow depth-of-field imaging, a popular recent application in mobile photography. Shi et al. [28] integrated DP hardware with structured diffractive optical elements for complex image encoding and high-precision multimodal reconstruction. In addition, DP data can be used for disparity estimation [22, 36], rain removal [18], and reflection removal [25]. It is foreseeable that DP data-driven applications will continue to emerge in various new fields." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 381, + 411, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 411, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 411, + 393 + ], + "type": "text", + "content": "3. Method of Sdirt" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 510 + ], + "type": "text", + "content": "To provide the required input for training the DfDP network, Sdirt generates realistic DP images through ray tracing. As shown in Fig. 2(c), the module achieves pixel-wise rendering of DP images by convolving the DP PSF corresponding to each point in the depth map with the AiF RGB image. The DP PSF for each point is inferred by a pretrained MLP network. As illustrated in Fig. 2(b), the offline training process of the MLP network uses ray-traced DP PSFs (Fig. 2(a)) as ground truth (GT)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 517, + 477, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 477, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 477, + 529 + ], + "type": "text", + "content": "3.1. Ray-traced DP PSF simulator" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": "As shown in Fig. 2(a), lens ray tracing can accurately obtain the landing point on the sensor surface [8, 35, 39, 40]. We construct a ray set " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " rays, each originating from an object point " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": ". The aperture stop's image in object space is treated as the entrance pupil, from which " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " points are densely sampled. Each ray in " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " is initially directed from " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " to its corresponding entrance pupil sample. As rays in " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " traverse each lens surface, their positions and directions are updated by refraction according to Snell's law and the lens parameters. After multiple refractions, the rays reach the sensor plane. We denote the sensor-plane landing points as " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": " and the corresponding ray directions as " + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 554, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "DP sensor ray tracing is limited because camera manufacturers do not disclose the structural parameters of the microlens and sub-pixel components within the DP pixel." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26107" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 73, + 555, + 339 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 78, + 91 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 78, + 91 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 78, + 91 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 73, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 78, + 73, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 78, + 73, + 555, + 339 + ], + "type": "image", + "image_path": "93ab78c423775cc9cdd65314e09bd144dcac0bf9c2a6709b6aeed6c296cad114.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 345, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 345, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 555, + 390 + ], + "type": "text", + "content": "Figure 2. Simulating Dual-Pixel Images from Ray Tracing pipeline. (a) Ray-traced DP PSF simulator. Calculates spatially varying DP PSFs for lens and DP sensor through ray tracing. (b) DP PSF prediction network. Trains an MLP network to predict DP PSFs, using the ray-traced DP PSFs as GT. (c) Pixel-wise DP image rendering module. The network predicts the DP PSFs for all points in the depth map (red pass). Then, each DP PSF is convolved with the AiF RGB image to render the simulated DP image (blue pass)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": "As a result, we simplify the DP pixel structure (Fig. 3(a)) based on past research [15, 29]. We model the microlens as a thin lens with radius " + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": " and focal length " + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": ", and define " + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": " as distance between the sub-pixel and the microlens, " + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": " as sub-pixel width, and " + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "inline_equation", + "content": "ps" + }, + { + "bbox": [ + 55, + 404, + 295, + 464 + ], + "type": "text", + "content": " as DP pixel size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": "The DP PSF can be obtained by separately calculating the cumulative integration of the rays for each sub-pixel. We analyze which sub-pixel the " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": "-th ray " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " in the ray set " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " ultimately enters. Through lens ray tracing, we obtain the landing point " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "O_{k}(x_{k},y_{k},z_{k})" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " and direction " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "D_{k}(x_{k}^{\\prime},y_{k}^{\\prime},z_{k}^{\\prime})" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": "-th ray " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " on the sensor surface. Based on the landing point " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "O_{k}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": ", we can easily calculate that " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " lies within the DP pixel " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": ", which has surface center coordinates " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "O_{k}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": ", the situation differs depending on whether it lies within or outside the microlens:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "O_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " lies within the microlens, as shown in Fig. 3(b), " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is refracted by the microlens and then enters a sub-pixel. According to geometric optics theory, " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and the principal ray incident on the thin lens at the same angle both converge on the focal plane of the thin lens. We only need to focus on a few boundary lines " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{L1}, x_{M1}, x_{R1}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " to determine which sub-pixel " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " ultimately enters. If " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " is in the interval " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "[x_{L1}, x_{M1}]" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", it enters the left sub-pixel; if in the interval " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "[x_{M1}, x_{R1}]" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", it enters the right sub-pixel; otherwise, it is considered a missing ray. The calculation methods for the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 404, + 436, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 436, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 436, + 416 + ], + "type": "text", + "content": "boundary lines are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 340, + 422, + 529, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 422, + 529, + 435 + ], + "spans": [ + { + "bbox": [ + 340, + 422, + 529, + 435 + ], + "type": "interline_equation", + "content": "x _ {L 1} = x _ {i} + w - (f * \\tan \\theta - w) * h / (f - h)", + "image_path": "e0bf34f86501ed46b53e3d3b227fc45de166e054b15b55e6f1feb8eb5c76d1aa.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 437, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 437, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 340, + 437, + 553, + 449 + ], + "type": "interline_equation", + "content": "x _ {M 1} = x _ {i} - (f * \\tan \\theta) * h / (f - h) \\tag {1}", + "image_path": "cfde52f962348fd1bed66f1a41aea0cf0aa14eac930e9bc8273b415e62d04fee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 452, + 529, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 452, + 529, + 464 + ], + "spans": [ + { + "bbox": [ + 341, + 452, + 529, + 464 + ], + "type": "interline_equation", + "content": "x _ {R 1} = x _ {i} - w - (f * \\tan \\theta + w) * h / (f - h)", + "image_path": "540ea8da0c81c6ae05413980a2cd6dd7b96a8e0a9ff593047be67d552f6bee3b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "inline_equation", + "content": "\\tan \\theta" + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "content": " is the tangent of the angle at which " + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "content": " lands on the sensor surface, equivalent to " + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "inline_equation", + "content": "x_{k}^{\\prime} / z_{k}^{\\prime}" + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "content": ". We provide the derivation hints for Eq. (1) in Fig. 4(a)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "O_{k}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " lies outside the microlens, as shown in Fig. 3(c), " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " enters the sub-pixel directly without refraction. Similarly, checking boundary lines " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "x_{L2}, x_{M2}, x_{R2}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " suffices to determine which sub-pixel " + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 313, + 506, + 553, + 554 + ], + "type": "text", + "content": " ultimately enters:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 559, + 491, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 559, + 491, + 571 + ], + "spans": [ + { + "bbox": [ + 380, + 559, + 491, + 571 + ], + "type": "interline_equation", + "content": "x _ {L 2} = x _ {i} + w - h * \\tan \\theta", + "image_path": "f78c7dc9327073a8e5a436f87208c7522d5cc0290bfab5e3490aa54d9f5ba42e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 574, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 574, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 378, + 574, + 553, + 586 + ], + "type": "interline_equation", + "content": "x _ {M 2} = x _ {i} - h * \\tan \\theta \\tag {2}", + "image_path": "a9e831243f1658748a043ac6b46fa3abe42a7bf4603740db63c4321aaabb7665.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 590, + 490, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 590, + 490, + 601 + ], + "spans": [ + { + "bbox": [ + 380, + 590, + 490, + 601 + ], + "type": "interline_equation", + "content": "x _ {R 2} = x _ {i} - w - h * \\tan \\theta", + "image_path": "07a132be67041737c3667c00c087b2abd870798f5eeae25db6d60abeb20091b9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "text", + "content": "The left PSF " + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "inline_equation", + "content": "(PSF_{L})" + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "text", + "content": " can be calculated by integrating the set of rays " + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "text", + "content": " emitted from point " + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 607, + 555, + 642 + ], + "type": "text", + "content": " over all left subpixels:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 642, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 642, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 364, + 642, + 553, + 673 + ], + "type": "interline_equation", + "content": "P S F _ {L} (i, j) = \\sum_ {k = 1} ^ {n} A _ {k} * \\delta_ {L, k} (i, j) \\tag {3}", + "image_path": "84c1fa8d0d9a23cf2e44e57497f79f8343a29670bc0579a2888a80ac3cfba74b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{L,k}(i,j)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " represents the energy distribution of ray " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " in the left sub-pixel at DP pixel " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ". The ray energy distribution is assumed to be a unit impulse, which" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26108" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 272, + 371 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 272, + 371 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 272, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 272, + 371 + ], + "type": "image", + "image_path": "ae6b7be3ee796de1178f432f1afcb6e403ff915397ef5d0386f7b604cf2c2d5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "type": "text", + "content": "Figure 3. (a) DP pixel structure layout. The left part is a right-side view, and the right part is a front perspective view. (b) On hitting the microlens, the " + }, + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "type": "text", + "content": "-th ray lands on left/right sub-pixel in red/blue interval, else missed. (c) Without hitting the microlens, the " + }, + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 372, + 295, + 428 + ], + "type": "text", + "content": "-th ray lands on left/right sub-pixel in red/blue interval, else missed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": "equals zero unless it eventually enters the left sub-pixel of DP pixel " + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": ". By following the same calculation steps as for " + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "inline_equation", + "content": "PSF_{L}" + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "inline_equation", + "content": "PSF_{R}" + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": " can also be computed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 260, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 260, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 260, + 495 + ], + "type": "text", + "content": "3.2. Pixel-wise DP image rendering module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": "As shown in Fig. 2(b), we reduce the computational cost of ray-traced DP PSFs by training an MLP network to predict them, inspired by [33, 39]. The network input is the normalized coordinate " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": " within the camera's valid imaging region (Fig. 4(b)), which is a frustum defined by the field of view, sensor size, and the preset minimum and maximum depths. We normalize the " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": " coordinates to " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": ". After fixing the focus distance, we set the DP PSF kernel size " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "(ks)" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": " large enough to ensure that the DP PSF for any point within the valid imaging region can be fully displayed. The network consists of 5 hidden layers, each containing 512 neurons, and an output layer with " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "2 * ks^2" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": " neurons. We adopt the L2 loss to supervise predictions " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "P\\hat{S} F_L" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "inline_equation", + "content": "P\\hat{S} F_R" + }, + { + "bbox": [ + 55, + 498, + 296, + 654 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 658, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 658, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 658, + 295, + 673 + ], + "type": "interline_equation", + "content": "L o s s = L _ {2} (P \\hat {S} F _ {L}, P S F _ {L}) + L _ {2} (P \\hat {S} F _ {R}, P S F _ {R}) \\quad (4)", + "image_path": "3b4925b988dc66b2ea3d17e4b8ae91ab21fb0fa1702ba740e5dd605a85dbcc65.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "During training, we use ray-traced DP PSFs as GT and apply max normalization to alleviate learning difficulties caused by large radii PSFs. During inference, we apply" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 550, + 277 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 550, + 277 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 550, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 550, + 277 + ], + "type": "image", + "image_path": "c0310068b5819416d0a4c7d18005ee98a792dfcec67246f15baf43eb512f6c8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "text", + "content": "Figure 4. (a) Derivation hints for Eq. (1). (b) The valid imaging region is a frustum, and we normalize the " + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "text", + "content": " coordinates to " + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "text", + "content": ". (c) The DfDP model takes DP image as input to predict the depth map. (d) During cost volume generation in [6], we stack original disparity (green arrows) and add reverse disparity (blue arrows), with " + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "inline_equation", + "content": "d_{max}" + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "text", + "content": " as max displacement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 357, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 553, + 392 + ], + "type": "text", + "content": "sum normalization to the predicted DP PSFs, approximating the uniform intensity distribution in cameras with vignetting compensation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": "Because the DP PSF is shift-variant, we convolve each pixel's DP PSF with the AiF image to simulate the DP image captured by a camera. The RGBD dataset [21, 30] provides the paired RGB image " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "(I_{RGB})" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " and depth map " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "(I_D)" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": ". As shown in Fig. 2(c), we treat each pixel in " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "I_D" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " as an object point and use the trained MLP to predict its DP PSF. Using " + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "inline_equation", + "content": "I_{RGB}" + }, + { + "bbox": [ + 313, + 394, + 554, + 501 + ], + "type": "text", + "content": " as the AiF image, we apply pixel-wise local convolution [39] to render the DP image containing aberration and phase information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 471, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 471, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 471, + 526 + ], + "type": "text", + "content": "4. DfDP Model Based on Sdtirt" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": "As shown in Fig. 4(c), the input to the DfDP model is provided in real time by the pixel-wise DP image rendering module of Sdirt. At each training iteration, Sdirt feeds the DfDP model with simulated DP images of shape " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "(B,6,H,W)" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": " is the batch size, the number of channels is 6, and " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": " are the DP image height and width. The model outputs the predicted depth maps " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\hat{I}_D)" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "(B,1,H,W)" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": ". We apply the L1 loss to supervise " + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\hat{I}_D" + }, + { + "bbox": [ + 313, + 534, + 555, + 631 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 634, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 634, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 392, + 634, + 553, + 649 + ], + "type": "interline_equation", + "content": "L o s s = L _ {1} \\left(\\hat {I _ {D}}, I _ {D}\\right) \\tag {5}", + "image_path": "ee3bdc9bc22d99f137c0fe887c298fb098cfdd3b356574cd31f2012d032fcb72.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "We select [6] as the DfDP model and make reasonable adjustments to its cost volume step to accommodate the DP data. In binocular images, disparity from points at different depths always has the same direction, thus the cost volume generation process in [6] is also unidirectional, as shown by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26109" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 333 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 333 + ], + "type": "image", + "image_path": "b234044e6e568ce38fd4dd4a13c7ec7e60bac25d74f5e9d72763dc3e89c643ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": "Figure 5. Qualitative results of simulated DP PSFs. Evaluate the real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) F/1.8 DP PSFs at two depths " + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "inline_equation", + "content": "(0.5\\mathrm{m}" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m})" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": " and three different positions. As the object point " + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": " moves further from the optical axis, the real " + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "inline_equation", + "content": "PSF_{L}" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "inline_equation", + "content": "PSF_{R}" + }, + { + "bbox": [ + 55, + 334, + 555, + 388 + ], + "type": "text", + "content": " become more phase asymmetric, and aberrations increase. Existing simulators neglect aberrations and DP phase splitting, causing a large gap between simulated and real DP PSFs. Only our ray-traced simulator predicts realistic results at all depths and positions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 401, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 462 + ], + "type": "text", + "content": "the green arrows in Fig. 4(d). However, in DP images, disparity from object points before and after the focus distance is opposite in direction. Therefore, we add blue arrows to extend the cost volume generation, in order to enhance the DfDP model's ability to capture reverse disparity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 212, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 212, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 212, + 486 + ], + "type": "text", + "content": "5. S skirt Representation Result" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 187, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 187, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 187, + 506 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": "Sdirt employs a Canon RF50mm lens and an R6 Mark II camera body as the real camera. The simulated DP sensor has dimensions of " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "24\\mathrm{mm} \\times 36\\mathrm{mm}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " with a resolution of " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "512 \\times 768" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ". We set the focus distance of both the simulated and real cameras to " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ", the valid imaging range to " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "0.5 - 20\\mathrm{m}" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ", the F-number to " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "F/4" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": ", and the ray set " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " to contain 4096 rays. During MLP training, " + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "inline_equation", + "content": "ks" + }, + { + "bbox": [ + 54, + 510, + 295, + 652 + ], + "type": "text", + "content": " is set to 21 to fully display the DP PSF of any point within the valid imaging region. We train the MLP for 100,000 iterations, selecting 128 random points within the valid imaging region during each iteration. We conducted our experiments on a 12700K CPU and a 3090 GPU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Following [39], we assume chromatic aberration is well corrected compared to other aberrations and defocus effects, allowing us to use a " + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "inline_equation", + "content": "550\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": " wavelength to reduce ray tracing costs. For details on obtaining the structural parameters of the DP pixel, please refer to the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 401, + 437, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 437, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 437, + 414 + ], + "type": "text", + "content": "5.2. Comparison methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 575 + ], + "type": "text", + "content": "For comparison, we select all existing model-based DP PSF simulators: L2R [2], Modeling [26], DDDNet [23], and an added CoC simulator. Although [24] is a more recent work, its simulator fully reuses DDDNet [23], so we do not include it for comparison. Calibration-based simulators and constraint-training methods are data-driven and require real DP-depth paired data, making them fundamentally different from model-based simulators, so no comparison is made. We directly used the source code of these DP PSF simulators [2, 26], and re-implemented the simulator [23] which only provided executable files. To ensure fairness, we set the focus distance, F-number, and sensor size in each simulator to be consistent with ours." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 586, + 495, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 495, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 495, + 597 + ], + "type": "text", + "content": "5.3. Evaluation on simulated DP PSFs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 652 + ], + "type": "text", + "content": "To facilitate observation of aberration and phase information, we temporarily set the F-numbers of both the simulated and real cameras to F/1.8. We capture a OLED screen displaying a bright pixel to obtain the real DP PSFs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "As shown in Fig. 5, we evaluate the DP PSF " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "(PSF_{L}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " on the left, " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "PSF_{R}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " on the right) of object points " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " at various depths and positions within the valid imaging region. Observing the real DP PSFs, it can be found that defocus (with focus distance of " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": ") causes a significant phase difference" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26110" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 351 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 351 + ], + "type": "image", + "image_path": "6f04a9898785919d648bab907418b11cf4a868e14099cf0478c7670c820c9399.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 352, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 352, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 555, + 396 + ], + "type": "text", + "content": "Figure 6. Qualitative results of simulated DP images. Evaluate the similarity between simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) and real F/4 defocused left DP images at two depths (0.5 m and 2 m). Compared to real F/4 defocused images, images simulated by other methods exhibit varying sizes of patterns, incomplete shapes, and texture shifts in different directions before and after the focus distance (1 m). Our method produces the most realistic simulated images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": "between the real " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "PSF_{L}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "PSF_{R}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": ", and the phase difference at depths of " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " is opposite. When " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " is on the optical axis, all other simulators fail to provide DP PSF results close to the real ones. Farther from the optical axis, the real " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "PSF_{L}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "inline_equation", + "content": "PSF_{R}" + }, + { + "bbox": [ + 54, + 408, + 294, + 529 + ], + "type": "text", + "content": " become phase asymmetric, with smaller phase differences and more off-axis aberrations. The differences between DP PSFs from other simulators and the real ones become more apparent. In contrast, our ray-traced simulator predicts accurate results at all depths and positions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "content": "We present the quantitative results in Tab. 1, using OpenCV's normalized squared difference (NSD) and normalized cross-correlation (NCC) to quantify the error and similarity between the real and simulated DP PSFs. We sampled 50 DP PSFs at " + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "content": " within the valid imaging region. Since the real DP PSF has flip symmetry (not phase symmetry) about both the " + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 531, + 295, + 675 + ], + "type": "text", + "content": " axes, we performed uniform sampling in the first quadrant. Our method achieves the highest similarity (0.915 NCC) and lowest error (0.133 NSD), demonstrating its effectiveness in bridging the domain gap between simulated and real DP PSFs introduced by lens aberrations and sensor phase splitting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "In all subsequent experiments, the F-number is adjusted back to F/4 due to the large blur kernel size caused by defocus at F/1.8, which resulted in GPU memory shortages." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 449, + 553, + 490 + ], + "blocks": [ + { + "bbox": [ + 313, + 404, + 555, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 447 + ], + "type": "text", + "content": "Table 1. Quantitative results of simulated DP PSFs. The NCC and NSD metrics between simulated and real DP PSFs at 50 points were evaluated. Our method achieved the highest similarity and lowest error." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 449, + 553, + 490 + ], + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 490 + ], + "type": "table", + "html": "
MethodOursCoCL2R [2]Modeling [26]DDDNet [23]
NCC↑0.9150.6720.6380.7070.589
NSD↓0.1330.4480.5230.4230.625
", + "image_path": "ac33c1fec695d79f579046e46e4fdbd64e05f1f1e805c654c0ae642d633ab291.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 504, + 503, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 503, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 503, + 517 + ], + "type": "text", + "content": "5.4. Evaluation on simulated DP images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 521, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 665 + ], + "type": "text", + "content": "When comparing simulated and real DP images, to assess the degradation of simulated DP images caused by defocus effects at different depths, we captured richly textured planar scenes from " + }, + { + "bbox": [ + 312, + 521, + 555, + 665 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 312, + 521, + 555, + 665 + ], + "type": "text", + "content": " to infinity, perpendicular to the optical axis. Shallower depths were sampled more densely to better capture defocus variations. This resulted in 56 scenes with varying depths, including 11 infinite-depth scenes. For each scene, real DP images were captured at F/4 and F/20. The F/20 image served as the AiF RGB image, and its capture-time depth was assigned to the entire depth map. The F/20 RGB-depth pairs were then input into DP simulators to generate the simulated F/4 defocused DP images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "As shown in Fig. 6, we compare the qualitative results of the defocused left DP images at depths of " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": ". Due to errors in simulating the DP PSF, other methods produce simulated images with textures shifted to the right at" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26111" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 175, + 162 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 175, + 162 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 175, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 175, + 162 + ], + "type": "image", + "image_path": "e7d1b9ebb9ac384c1bf71c1d83fb225112b33e28f9a0ea1d26274a2509e6ad1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "text", + "content": "Figure 7. Quantitative results of simulated DP images. Evaluate the similarity of 56 real and simulated (ours, CoC, L2R [2], Modeling [26], and DDDNet [23]) planar scene F/4 DP images at different depths using PSNR " + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "text", + "content": " and SSIM " + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "text", + "content": " metrics. As the depth of simulated images deviates further from the focus distance " + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "inline_equation", + "content": "(1\\mathrm{m})" + }, + { + "bbox": [ + 55, + 163, + 295, + 239 + ], + "type": "text", + "content": ", the less realistic the other methods become, whereas our scheme maintains the highest accuracy across all depths." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 294, + 162 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 294, + 162 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 294, + 162 + ], + "type": "image", + "image_path": "9dc4a618d98703cedecee41e77b587735f70ce1dfee490cddf9c457d0fdc2e3d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "type": "text", + "content": " depth and to the left at " + }, + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "type": "inline_equation", + "content": "2\\mathrm{m}" + }, + { + "bbox": [ + 55, + 247, + 296, + 306 + ], + "type": "text", + "content": " depth, compared to the real left DP image. Moreover, other methods show variations in image texture sizes and incomplete shapes. However, the DP images simulated by our Sdirt method show minimal deviation in texture details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 415 + ], + "type": "text", + "content": "As shown in Fig. 7, which presents the quantitative results, we assess the similarity between simulated and real images using PSNR and SSIM. As the simulated image's depth moves further away from the focus distance (1 m), the neglect of aberrations and phase splitting in other methods leads to a more significant decrease in their metrics. In contrast, our Sdirt method achieves the best results at all depths, with average PSNR/SSIM values of 37.1982/0.9845, indicating that our simulated DP images are highly realistic." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "type": "text", + "content": "6. DfDP Estimation Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 187, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 187, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 187, + 456 + ], + "type": "text", + "content": "6.1. Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "text", + "content": "To train the DfDP model, we simulate F/4 DP images using Sdirt on RGBD datasets [21, 30]. The NYU Depth dataset [30], preprocessed by [12], contains 50,688 indoor scenes. We adopt the AdamW optimizer [19] and CosineAnnealing scheduler [20] with an initial learning rate of " + }, + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 462, + 295, + 582 + ], + "type": "text", + "content": ". The model is trained with a batch size of 4 for 50 epochs. In each epoch, 2,000 randomly selected DP-depth pairs are used for training. After training, the model is directly evaluated on a real-world test set without any fin-tuning to assess its generalization ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 228, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 228, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 228, + 600 + ], + "type": "text", + "content": "6.2. A new real-world test set DP119" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Unlike previous simulation methods, S skirt addresses data scarcity by simulating hardware, requiring consistent camera and setup during training and inference. However, as shown in Tab. 2, no public dataset satisfies all of the following conditions: 1. Real DP-depth paired data for simulating DP images; 2. A known lens structure for ray tracing; 3. A fixed focus distance, since changing it alters the DP PSF of the same object point. This makes evaluation on public datasets with unknown hardware impractical. We there-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "fore collected the DP119 test set, which contains diverse test scenes." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 116, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 356, + 105, + 512, + 115 + ], + "lines": [ + { + "bbox": [ + 356, + 105, + 512, + 115 + ], + "spans": [ + { + "bbox": [ + 356, + 105, + 512, + 115 + ], + "type": "text", + "content": "Table 2. Summary of existing DP datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 116, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 317, + 116, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 553, + 180 + ], + "type": "table", + "html": "
DPDD [1]L2R [2]DPNet [9]DP5K [16]DDDNet [23]Modeling [26]DP119 Ours
Real captured✓×✓✓×✓✓
Paired depth×✓✓✓✓×✓
Lens structure✓-×✓-×✓
Fixed focus dist.✓-××-×✓
", + "image_path": "6ca049ba0eb72b1ff94fc5d7b0c9426ec0aedd4c52be084bb3e6927c2e38e7d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 186, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 378 + ], + "type": "text", + "content": "DP119 consists of 45 planar scenes, 44 box scenes, and 30 casual scenes, totaling 119 scenes. The casual scenes represent common scenarios, suitable for evaluating the robustness of simulation models. The planar and box scenes are richly textured, helping the model utilize aberration and phase cues for depth estimation. Textureless areas, even when defocused, lack aberrations or phase differences, making depth estimation difficult and evaluation unreliable. Thus, richly textured planar and box scenes are ideal for assessing domain gaps between simulated and real DP images. Real DP images are captured by the Canon RF50mm lens with a R6 Mark II at F/4 and " + }, + { + "bbox": [ + 313, + 186, + 555, + 378 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 186, + 555, + 378 + ], + "type": "text", + "content": " focus. GT depth maps for planar scenes are created by the depths at the capture time, while those for box and casual scenes are obtained from LiDAR scans. For more information about the DP119 test set, please refer to the Supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 385, + 388, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 388, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 388, + 396 + ], + "type": "text", + "content": "6.3. Evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 498 + ], + "type": "text", + "content": "We present qualitative and quantitative results for the DfDP model trained by CoC, L2R [2], Modeling [26], DDDNet [23], and our Sdirt. For a fair comparison, we deploy their DfDP models using their DP PSF simulator source code [2, 26] and re-implemented code [23], and have all models share the same depth estimation network architecture, initialize from a common CoC pre-trained checkpoint, and follow the same implementation details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Figure 8 shows depth estimation results for four casual scenes, with a color bar in meters for visualization. It can be seen that although other methods provide some relative positional information in the center region, they fail to correctly estimate relative positional relationships in the edge regions, and both the center and edge regions have large absolute positional errors. This is because the DP PSFs predicted by these methods exhibit phase symmetry and shift-invariance, whereas real DP PSFs exhibit phase asymmetry, aberrations, and shift-variance. Their methods' mismatch with real DP PSFs leads to significant depth estimation errors, especially near the edges. The results of L2R [2] and Modeling [26] are similar because the peak positions of their DP PSFs are quite close. In contrast, our method provides accurate relative positional information across the entire image area with minimal absolute positional errors, as our simulated DP PSFs are highly consistent with the real ones in both the center and edge regions." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26112" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 261 + ], + "type": "image", + "image_path": "d184cd67a6648a800ad3c990640da5270722ca35992508579011ff2c887b3397.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "type": "text", + "content": "Figure 8. Qualitative results of absolute depth estimation on DP119. Evaluate DfDP models with CoC, L2R [2], Modeling [26], DDDNet [23] and our Sdirt at four casual scenes. Each result image includes a color bar showing depth in meters. Their depth estimation results show partial accuracy in relative positional relationships but large absolute positional errors. Our depth estimation results, however, demonstrate high accuracy in both relative and absolute positions with minimal errors. (Best viewed in color and enlarge on screen.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 353, + 294, + 492 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "text", + "content": "Table 3. Quantitative results of absolute depth estimation on DP119. The DfDP model trained with our Sdirt achieves superior performance across most scenarios and evaluation metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 353, + 294, + 492 + ], + "lines": [ + { + "bbox": [ + 57, + 353, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 57, + 353, + 294, + 492 + ], + "type": "table", + "html": "
SceneMethodMAE↓MSE↓Abs.r↓Sq.r↓Acc-1↑Acc-2↑
PlanarOurs0.08450.01090.08710.00950.98490.9997
CoC0.20850.10010.18010.06590.66700.8990
L2R [2]0.24180.12710.21120.08410.63190.8536
Modeling [26]0.22840.11420.20040.07660.64960.8725
DDDNet [23]0.25830.14850.21910.09580.56480.8089
BoxOurs0.11970.03390.09060.02310.94740.9812
CoC0.33750.18040.24420.11160.44120.8277
L2R [2]0.38660.22840.28030.14120.36510.7156
Modeling [26]0.36550.20550.26600.12780.39070.7758
DDDNet [23]0.41770.26760.29750.16360.34560.6274
CasualOurs0.27020.22940.46320.72410.82360.9314
CoC0.79251.85790.54610.68210.33180.6103
L2R [2]0.81701.74870.55970.67190.27600.5315
Modeling [26]0.79341.72560.55100.66550.29780.5732
DDDNet [23]0.89312.06240.57520.71350.24810.4685
", + "image_path": "ac4fa10e03dac76224fe0b30ee03638acf4a7743d8893a1df4aedb160f4e8ee4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": "We evaluate all models on the DP119 test set using the following metrics to assess depth estimation performance: mean absolute error (MAE), mean squared error (MSE), absolute relative error (Abs.r.), squared relative error (Sq.r.), accuracy with " + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": " (Acc-1) and accuracy with " + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta < 1.25^2" + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": " (Acc-2). Since the planar and box scenes are richly textured and do not have interference from textureless areas, the simulation models can rely on abundant aberration and phase difference cues for depth estimation. As shown in Tab. 3, in the simplest planar scene, our model achieves the best results across all metrics (0.9849 Acc-1). In the box scenes, all models experience a performance drop compared to the planar scenes, but we still achieve the optimal metrics (0.9474 Acc-1). These results demonstrate that our model is highly realistic and effectively bridges the domain gap between simulated and real DP images. Furthermore, we test the robustness of all models on the casual scenes. Due to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 320, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 406 + ], + "type": "text", + "content": "the textureless areas in the casual scenes, the performance of all models significantly degrades. However, our model still achieves the best metrics (0.8236 Acc-1), far surpassing the second-best model (0.3318 Acc-1). This further demonstrates that our model is the most realistic and has the best generalization performance. For more sample results from the DP119 test set, please refer to the Supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 466, + 443 + ], + "type": "text", + "content": "7. Conclusion and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 456, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 555, + 578 + ], + "type": "text", + "content": "Simulated DP images help address the scarcity of DP-depth paired data but face a domain gap between simulated and real DP data. In this work, we propose a novel simulation framework called Sdirt to bridge this gap. Specifically, we calculate the DP PSF for points in object space using ray tracing, and employ a network to predict them. Then, we render DP images based on the predicted DP PSFs. Experimental results show that the proposed Sdirt can simulate more realistic DP data. Moreover, depth estimation models trained based on Sdirt generalize better to real DP images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "type": "text", + "content": "We believe that Sdirt is not limited to DfDP tasks, it can provide extra depth information for any task with known optical imaging system parameters, promising significant applications in scenarios such as smartphones, automobiles, and microscopes in the future. However, Sdirt is only applicable to cameras equipped with a fixed-focus lens (the structure must be open) and a DP sensor (DP images must be available). Currently, only Canon (5D4, R series) meets these requirements. To further expand the application of Sdirt, more camera and smartphone manufacturers need to make these data accessible." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26113" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 140 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 32471146) and the project N20240194. The authors thank Echosom, Miya, and Xinge for valuable discussions and assistance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 224 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X 16, pages 111-126. Springer, 2020. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 280 + ], + "type": "text", + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "text", + "content": "[3] Abdullah Abuolaim, Radu Timofte, and Michael S Brown. Ntire 2021 challenge for defocus deblurring using dual-pixel images: Methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 578-587, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 335, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 296, + 392 + ], + "type": "text", + "content": "[4] Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "type": "text", + "content": "[5] Hadi Alzayer, Abdullah Abuolaim, Leung Chun Chan, Yang Yang, Ying Chen Lou, Jia-Bin Huang, and Abhishek Kar. Dc2: Dual-camera defocus control by learning to refocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21488-21497, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 513 + ], + "type": "text", + "content": "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "text", + "content": "[7] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 635 + ], + "type": "text", + "content": "[8] Geoffroi Côté, Fahim Mannan, Simon Thibault, Jean-François Lalonde, and Felix Heide. The differentiable lens: Compound lens search over glass surfaces and materials for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20803-20812, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "text", + "content": "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7628-7637, 2019. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 296, + 714 + ], + "type": "text", + "content": "[10] Bhargav Ghanekar, Salman Siddique Khan, Pranav Sharma, Shreyas Singh, Vivek Boominathan, Kaushik Mitra, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "Ashok Veeraraghavan. Passive snapshot coded aperture dual-pixel rgb-d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25348-25357, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 163 + ], + "type": "text", + "content": "[11] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 165, + 555, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 555, + 220 + ], + "type": "text", + "content": "[12] Junjie Hu, Mete Ozay, Yan Zhang, and Takayuki Okatani. Revisiting single image depth estimation: Toward higher resolution maps with accurate object boundaries. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), 2019. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 222, + 554, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 554, + 277 + ], + "type": "text", + "content": "[13] Minjun Kang, Jaesung Choe, Hyowon Ha, Hae-Gon Jeon, Sunghoon Im, In So Kweon, and Kuk-Jin Yoon. Facial depth and normal estimation using single dual-pixel camera. In European Conference on Computer Vision, pages 181-200. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 278, + 554, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 554, + 334 + ], + "type": "text", + "content": "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 335, + 554, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 554, + 413 + ], + "type": "text", + "content": "[15] Masahiro Kobayashi, Michiko Johnson, Yoichi Wada, Hiromasa Tsuboi, Hideaki Takada, Kenji Togo, Takafumi Kishi, Hidekazu Takahashi, Takeshi Ichikawa, and Shunsuke Inoue. A low noise and high sensitivity image sensor with imaging and phase-difference detection of all pixels. ITE Transactions on Media Technology and Applications, 4(2):123-128, 2016. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 470 + ], + "type": "text", + "content": "[16] Feiran Li, Heng Guo, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Learning to synthesize photorealistic dual-pixel images from rgbd frames. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-11. IEEE, 2023. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 472, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 506 + ], + "type": "text", + "content": "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 554, + 540 + ], + "type": "text", + "content": "[18] Yizhou Li, Yusuke Monno, and Masatoshi Okutomi. Dual-pixel raindrop removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 542, + 554, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 542, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 315, + 542, + 554, + 564 + ], + "type": "text", + "content": "[19] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 566, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 554, + 599 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 667 + ], + "type": "text", + "content": "[21] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 713 + ], + "type": "text", + "content": "[22] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. In 2024 International Conference on 3D Vision (3DV), pages 675-684. IEEE, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26114" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[23] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "text", + "content": "[24] Liyuan Pan, Richard Hartley, Liu Liu, Zhiwei Xu, Shah Chowdhury, Yan Yang, Hongguang Zhang, Hongdong Li, and Miaomiao Liu. Weakly-supervised depth estimation and image deblurring via dual-pixel sensors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "text", + "content": "[25] Abhijith Punnappurath and Michael S Brown. Reflection removal using a dual-pixel sensor. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1556-1565, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 297 + ], + "type": "text", + "content": "[26] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 353 + ], + "type": "text", + "content": "[27] Lingyan Ruan, Martin Balint, Mojtaba Bemana, Krzysztof Wolski, Hans-Peter Seidel, Karol Myszkowski, and Bin Chen. Self-supervised video defocus deblurring with atlas learning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "type": "text", + "content": "[28] Zheng Shi, Ilya Chugunov, Mario Bijelic, Geoffroi Côté, Jiwoon Yeom, Qiang Fu, Hadi Amata, Wolfgang Heidrich, and Felix Heide. Split-aperture 2-in-1 computational cameras. ACM Transactions on Graphics (TOG), 43(4):1-19, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 464 + ], + "type": "text", + "content": "[29] Eun Sub Shim, Kyungho Lee, Junghyung Pyo, Wooseok Choi, Jungbin Yun, Taesub Jung, Kyungduck Lee, Seyoung Kim, Chanhee Lee, Seungki Baek, et al. All-directional dual pixel auto focus technology in cmos image sensors. In 2021 Symposium on VLSI Technology, pages 1-2. IEEE, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 533 + ], + "type": "text", + "content": "[30] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 567 + ], + "type": "text", + "content": "[31] Przemysław Śliwiński and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 612 + ], + "type": "text", + "content": "[32] Kunal Swami. Adjust your focus: Defocus deblurring from dual-pixel images using explicit multi-scale cross-correlation. In International Conference on Computer Vision and Image Processing, pages 318-330. Springer, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 679 + ], + "type": "text", + "content": "[33] Ethan Tseng, Ali Mosleh, Fahim Mannan, Karl St-Arnaud, Avinash Sharma, Yifan Peng, Alexander Braun, Derek Nowrouzezahrai, Jean-Francois Lalonde, and Felix Heide. Differentiable compound optics and processing pipeline optimization for end-to-end camera design. ACM Transactions on Graphics (TOG), 40(2):1-19, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "text", + "content": "[34] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 554 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 139 + ], + "type": "text", + "content": "[35] Congli Wang, Ni Chen, and Wolfgang Heidrich. do: A differentiable engine for deep lens design of computational imaging systems. IEEE Transactions on Computational Imaging, 8:905-916, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[36] Zhuofeng Wu, Doehyung Lee, Zihua Liu, Kazunori Yoshizaki, Yusuke Monno, and Masatoshi Okutomi. Disparity estimation using a quad-pixel sensor. arXiv preprint arXiv:2409.00665, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 252 + ], + "type": "text", + "content": "[37] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "text", + "content": "[38] Hao Yang, Liyuan Pan, Yan Yang, Richard Hartley, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24078-24087, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[39] Xinge Yang, Qiang Fu, Mohamed Elhoseiny, and Wolfgang Heidrich. Aberration-aware depth-from-focus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 2, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[40] Xinge Yang, Qiang Fu, and Wolfgang Heidrich. Curriculum learning for ab initio deep learned refractive optics. Nature communications, 15(1):6572, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[41] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13263-13272, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[42] Dafeng Zhang and Xiaobing Wang. Dynamic multi-scale network for dual-pixel images defocus deblurring with transformer. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 488, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 554 + ], + "type": "text", + "content": "[43] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du 2 net: Learning depth estimation from dual-cameras and dual-pixels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 582-598. Springer, 2020. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26115" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_content_list.json b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44525e8f1c445ad29a25f8b98ea39370d65dac6e --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_content_list.json @@ -0,0 +1,2005 @@ +[ + { + "type": "text", + "text": "Simultaneous Motion And Noise Estimation with Event Cameras", + "text_level": 1, + "bbox": [ + 169, + 130, + 826, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shintaro Shiba $^{1,2,3}$ , Yoshimitsu Aoki $^{1}$ and Guillermo Gallego $^{3,4}$", + "bbox": [ + 250, + 179, + 754, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Keio University, 2 Woven by Toyota, Inc., Japan. 3 Technische Universität Berlin, 4 Einstein Center Digital Future, Robotics Institute Germany, and Science of Intelligence Excellence Cluster, Germany.", + "bbox": [ + 99, + 198, + 898, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras are emerging vision sensors whose noise is challenging to characterize. Existing denoising methods for event cameras are often designed in isolation and thus consider other tasks, such as motion estimation, separately (i.e., sequentially after denoising). However, motion is an intrinsic part of event data, since scene edges cannot be sensed without motion. We propose, to the best of our knowledge, the first method that simultaneously estimates motion in its various forms (e.g., ego-motion, optical flow) and noise. The method is flexible, as it allows replacing the one-step motion estimation of the widely-used Contrast Maximization framework with any other motion estimator, such as deep neural networks. The experiments show that the proposed method achieves state-of-the-art results on the E-MLB denoising benchmark and competitive results on the DND21 benchmark, while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances event-data denoising theory and expands practical denoising use-cases via open-source code. Project page: https://github.com/tub-rip/ESMD", + "bbox": [ + 88, + 301, + 485, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 633, + 220, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras are emerging vision sensors that overcome shortcomings of conventional cameras (e.g., motion blur, limited dynamic range, data and power efficiency), yet they suffer from a considerable amount of noise due to their novelty and operation in low-power (transistor subthreshold) conditions [19, 36]. Since event cameras are suitable for many computer vision tasks, especially motion-related tasks, it is paramount to classify event data that is related to motion (i.e., signal) and that is not (i.e., noise). However, denoising event data is a challenging problem because noise properties have yet to be fully characterized [22, 24] and it is infeasible to define ground-truth (GT) noise labels in real-world data recordings. Previous works either leveraged simulation or prepared real event data and aggressively suppressed noise occurrence.", + "bbox": [ + 89, + 657, + 482, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, such artificial modifications on the real event", + "bbox": [ + 109, + 885, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60f0d0e4b9b4750bd32f26bb9e4daa66eef54a87131f6d80f3004d9d4875aded.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed method. We use only raw events as input, and estimate signal events, noise events, and motion (e.g., dense optical flow and ego-motion estimation)." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 903, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data could change event signal/noise characteristics, possibly deteriorating the information that the original event had for downstream tasks (e.g., motion estimation). Our work is inspired by the observation that motion is an intrinsic part of event data since scene edges cannot be sensed without motion. Rather than solving denoising and motion estimation tasks separately, our idea is to leverage motion information to improve denoising (as \"vice-versa\" is well known). In this work (Fig. 1), we rethink and investigate the connection between denoising and motion estimation, by asking ourselves whether they can be combined in a more integrated way to benefit the results of both tasks simultaneously. We formulate the simultaneous estimation problem by leveraging the idea of motion compensation in the Contrast Maximization (CMax) framework, defining novel metrics about the contribution to the contrast score for each event. In summary, we present several distinctive contributions:", + "bbox": [ + 511, + 547, + 906, + 805 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel method to jointly estimate noise and motion only from event data (Sec. 3).", + "2. We demonstrate that our joint estimation method $(i)$ achieves state-of-the-art results in existing denoising benchmarks (Sec. 4.2), and $(ii)$ improves the robustness of motion estimation (CMax framework) (Sec. 4.3)." + ], + "bbox": [ + 511, + 810, + 903, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6959", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3. We show the flexibility of our method on $(i)$ further applications, such as intensity reconstruction, validating its image quality (Secs. 4.4 and 4.5), and $(ii)$ in combination with learning-based motion estimators (Sec. 4.6).", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hope our work opens a new line of research for event cameras, by considering the problems of motion estimation and denoising together to take advantage of their coaction.", + "bbox": [ + 89, + 151, + 482, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 210, + 232, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising is a fundamental problem in event cameras since there are a lot of leakage events [24], and event-processing algorithms may suffer in scenarios with high noise [3]. Removing event noise has a double effect: it reduces the load of subsequent processing stages (as the load is proportional to the number of events) and improves their output (as the input data becomes of higher quality). In this work, we focus on noise due to the leakage inside the camera hardware, such as background activity (BA) events [36]. Events due to flickering or active lights are therefore not our priority.", + "bbox": [ + 89, + 236, + 482, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classical denoising works focus on the spatio-temporal correlation of events, as BA events have fewer neighboring events in space-time coordinates than signal events. This approach includes widely-used BA Filter (BAF) [9], and other types of spatio-temporal filters [26, 31, 37, 56]. Some works utilize additional sensor inputs (e.g., frames) to estimate noise probabilities [4, 12, 30, 55, 58], since edges are informative of the signal events produced.", + "bbox": [ + 89, + 387, + 482, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, learning-based approaches have gained more attention, such as multilayer perceptrons [24], convolutional neural networks [1, 4, 12], graph neural networks [2], and transformers [29]. They provide the probability of noise (i.e., signal-noise classification). Nonetheless, the challenge of learning-based approaches is the need for ground truth (GT) noise labels for training. GT labels have been obtained through simulation (e.g., [13, 24]) as well as real-world recordings, however, real-world recordings aggressively remove noisy events using existing filters (e.g., [24]) or use unnatural illumination settings (e.g., [29]), by leveraging the differences of noise occurrence due to different scene brightness [24, 35]. While these approaches are effective for benchmarking, the challenge of obtaining GT labels for practical real-world datasets remains unsolved.", + "bbox": [ + 89, + 508, + 482, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work builds upon a series of successful methods for motion estimation and their applications. Event-based motion estimation is an important topic since event cameras naturally respond to motion in the scenes [19]. In particular, Contrast Maximization (CMax) [17, 50] is a state-of-the-art framework for motion estimation, which has various applications, such as rotational motion estimation [15, 23, 32, 42, 45], optical flow estimation [27, 28, 44, 48, 64], and motion segmentation [51, 57, 60]. We investigate the connection and potential benefits of applying the CMax framework to denoising during the estimation process (i.e., simul", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "taneously), demonstrating motion estimation improvements as well as competitive denoising accuracy.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A prior work, \"ProgressiveMotionSeg\" [8], has extended the CMax-based segmentation method [51] with extra weights that accounted for how much each event contributed to the estimated motion clusters. However, [8] mainly focuses on the task of motion segmentation using low degrees-of-freedom (DOF) warps (e.g., 2-DOF feature flow) and does not discuss denoising efficacy or evaluate on standard denoising benchmarks. In contrast, our method supports various types of motion estimation (both low-DOF and high-DOF warps, such as optical flow), and considers that noise shall not be used to estimate motion, as from first principles we deem noise and motion to be uncorrelated.", + "bbox": [ + 511, + 121, + 906, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 314, + 648, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event cameras consist of independent pixels that generate asynchronous \"events\" when the logarithmic brightness at the pixel increases or decreases by a predefined contrast sensitivity. Each event $e_k \\doteq (\\mathbf{x}_k, t_k, p_k)$ contains the pixel coordinates $\\mathbf{x}_k$ , the timestamp $t_k$ , and polarity $p_k \\in \\{+1, -1\\}$ of the brightness change. Events occur asynchronously and sparsely on the pixel grid, resulting in a variable data rate based on the scene texture and dynamics.", + "bbox": [ + 511, + 340, + 905, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Contrast Maximization", + "text_level": 1, + "bbox": [ + 511, + 469, + 728, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The CMax framework [17] assumes events $\\mathcal{E} \\doteq \\{e_k\\}_{k=1}^{N_e}$ are caused by moving edges, and it transforms their coordinates according to a motion model $\\mathbf{W}$ , producing a set of warped events $\\mathcal{E}'_{t_{\\mathrm{ref}}} \\doteq \\{e_k'\\}_{k=1}^{N_e}$ at a reference time $t_{\\mathrm{ref}}$ :", + "bbox": [ + 511, + 489, + 905, + 554 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k} \\dot {=} \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\mapsto e _ {k} ^ {\\prime} \\dot {=} \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 561, + 903, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The warp $\\mathbf{x}_k^{\\prime} = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta})$ transports each event from $t_k$ to $t_{\\mathrm{ref}}$ along the motion curve that passes through it. Then, they are aggregated on an image of warped events (IWE):", + "bbox": [ + 511, + 582, + 903, + 628 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k} ^ {\\prime}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 633, + 903, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where each pixel $\\mathbf{x}$ sums the number of warped events $\\mathbf{x}_k^\\prime$ that fall within it. The Dirac delta is approximated by a Gaussian, $\\delta (\\mathbf{x} - \\pmb {\\mu})\\approx \\mathcal{N}(\\mathbf{x};\\pmb {\\mu},\\epsilon^2\\mathrm{Id})$ with $\\epsilon = 1$ pixel.", + "bbox": [ + 511, + 657, + 905, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, an objective function, such as the contrast of the IWE (2) (image variance [18] $\\mathrm{Var}\\big(I(\\mathbf{x};\\pmb {\\theta})\\big)\\doteq \\frac{1}{|\\Omega|}\\int_{\\Omega}(I(\\mathbf{x};\\pmb {\\theta}) - \\mu_I)^2 d\\mathbf{x})$ , measures the goodness of fit between the events and the candidate motion curves (warp).", + "bbox": [ + 511, + 703, + 905, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, an optimization algorithm iterates the above steps to find the motion parameters $\\theta^{*}$ that maximize the alignment of events caused by the same scene edge. Event alignment is measured by the strength of the edges of the IWE, which is related to image contrast [18].", + "bbox": [ + 511, + 765, + 905, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion Models. As exemplary warps $\\mathbf{W}$ we focus on two popular problem settings: rotational motion (3-DOF angular velocity estimation [23, 25, 42, 45]) and dense optical flow (pixel-wise velocity estimation [43, 48, 50, 64],", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6960", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe741d6d8a61f060504e939b6f26702e5ebdcd616adc04c751cd1d18fca0f992.jpg", + "image_caption": [ + "Figure 2. Qualitative result of denoising and flow estimation on E-MLB and DND21. The driving sequence has $5\\mathrm{-Hz}$ artificial noise. (b) is an alternative visualization of (a) with the same grayscale coding as (d). Flow is visualized on all events and flow color is given in Fig. 1." + ], + "image_footnote": [], + "bbox": [ + 114, + 97, + 885, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resulting in $2N_{p}$ -DOFs, where $N_{p}$ is the number of pixels). For the angular velocity estimation during a small time interval, the warp is parametrized by the angular velocity $\\theta \\equiv \\omega = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}$ , and $\\mathbf{x}^{h}(t)\\sim \\mathbb{R}(\\boldsymbol {\\omega}t)\\mathbf{x}^{h}(0)$ , using calibrated homogeneous coordinates $\\mathbf{x}^{h}$ and exponential coordinates $\\mathbb{R}(\\phi)\\doteq \\exp (\\phi^{\\wedge})$ [6, 16]. The warp for dense optical flow is $\\mathbf{x}_k^\\prime = \\mathbf{x}_k + (t_k - t_{\\mathrm{ref}})\\mathbf{v}(\\mathbf{x}_k)$ , where $\\theta = \\{\\mathbf{v}(\\mathbf{x})\\}_{\\mathbf{x}\\in \\Omega}$ is a flow field on the image plane $\\Omega$ .", + "bbox": [ + 89, + 470, + 482, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Denoising Based on Local Contrast", + "text_level": 1, + "bbox": [ + 89, + 599, + 397, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As Fig. 1 shows, our method is iterative. In each iteration, motion estimation is refined and events are classified into signal $\\mathcal{E}_s$ or noise $\\mathcal{E}_n$ . Without labels or additional information, events rely on each other to realize this classification. To this end, we compute a score $c_k$ for each event, rank the events according to their scores, and threshold at some value $T$ . The threshold is computed by assuming the amount or ratio of noisy events is known $\\eta$ , i.e., $T \\equiv T(\\eta)$ , which allows us to have high-level control over the denoising process. Hence we classify events into signal (the $\\tau \\doteq (1 - \\eta)$ percentage of events with highest scores) and noise:", + "bbox": [ + 89, + 621, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {s} \\equiv \\mathcal {E} _ {\\text {s i g n a l}} \\doteq \\left\\{e _ {k} \\in \\mathcal {E} \\mid c _ {k} > T (\\eta) \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 795, + 483, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {n} \\equiv \\mathcal {E} _ {\\text {n o i s e}} \\dot {=} \\mathcal {E} \\backslash \\mathcal {E} _ {\\text {s i g n a l}} \\quad (\\text {c o m p l e m e n t} \\mathcal {E} _ {\\text {s i g n a l}}).\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 815, + 441, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We studied different choices for the per-event score $c_k$ based on the agreement among warped events $\\mathcal{E}' = \\mathcal{E}_s' \\cup \\mathcal{E}_n'$ , through their IWEs, and concluded on two, as follows. Letting $I_k^s \\equiv I_k^s(\\mathbf{x}_k')$ , $I_k^n \\equiv I_k^n(\\mathbf{x}_k')$ and $I_k \\equiv I(\\mathbf{x}_k') = I_k^s + I_k^n$", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be the IWE values of previously classified signal events, noise events and all events, respectively, the two choices are: (i) $^{1}c_{k} \\doteq I_{k}$ , and (ii) the ratio of signal events $^{2}c_{k} \\doteq I_{k}^{s} / I_{k}$ (probabilistic viewpoint). In fact, both are equivalent: rule $^{1}c_{k} > T_{1}$ is the same as rule $^{2}c_{k} > T_{2}$ if $T_{2} = 1 - I_{k}^{n} / T_{1}$ and the event noise $I_{k}^{n}$ is known. We use $^{1}c_{k}$ , the number of warped events (also called local contrast [18, 51]), because it is the simplest to interpret.", + "bbox": [ + 511, + 470, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The intuition is that \"signal\" events (e.g., due to motion) warped according to the true motion lie all at motion-compensated edges, and therefore produce a sharp IWE. Noise events, produced all over the image plane, are not expected to contribute to the IWE's sharpness. Hence, we may classify events into signal or noise according to how much they contribute to edge strength. The IWE (2) provides one such measure of edge strength: the higher $I(\\mathbf{x})$ , the more events are warped to that pixel location $\\mathbf{x}$ (i.e., more events support the same scene edge), thus producing a sharper IWE. Hence, we define how much an event $e_k$ contributes to edge strength by means of the IWE (2), $^1 c_k$ .", + "bbox": [ + 511, + 593, + 908, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Invariance. Note that (3) is invariant to monotonically increasing functions of $c_k$ , as such functions preserve the order. The same $\\mathcal{E}_{\\mathrm{signal}}$ is obtained if one ranks and thresholds (at the corresponding level) values $\\{c_k\\}$ , $\\{\\ln(c_k)\\}$ , $\\{e^{c_k}\\}$ , etc. They are equivalent under the ranking rule (3).", + "bbox": [ + 511, + 777, + 908, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness to Various Edge Strengths. The Gaussian kernel in the IWE controls the sensitivity to edge strength. Since the proposed method is based on both signal and noise", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6961", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IWEs, increasing the size of the Gaussian kernel in the IWE can emphasize contour edges over isolated points. This preserves signal events better in regions with low IWE intensities, for example at pixels with large depth values.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Alternating Optimization", + "text_level": 1, + "bbox": [ + 89, + 160, + 321, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classifying events into signal and noise requires knowledge of the true motion, and estimation of the true motion (e.g., using CMax) requires knowledge of the signal events because, by definition, noise events carry no information about motion. Hence, it is a circular dependency problem.", + "bbox": [ + 89, + 181, + 482, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We approach its solution iteratively, as shown in Fig. 1, by combining motion estimation (Sec. 3.1) and S/N classification (Sec. 3.2). Assuming some initialization, the current signal events $\\mathcal{E}_{\\mathrm{signal}}$ are used to estimate the motion via CMax (1 step is enough); then all events are warped to compute the IWE and the per-event scores $c_{k}$ , from which the sets (3) are recomputed, in preparation for the next iteration. In practice, we initialize with a random split of $\\mathcal{E}$ into $\\mathcal{E}_{\\mathrm{signal}}$ and $\\mathcal{E}_{\\mathrm{noise}}$ , update $\\mathcal{E}_{\\mathrm{signal}}$ at every warp during the optimization iteration until the process converges (as flagged by the convergence of the motion parameters in CMax).", + "bbox": [ + 88, + 257, + 482, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Computational Complexity. The computational complexity of one iteration of the proposed method is $O(N_{p} + N_{e}\\log N_{e})$ , which is slightly larger than the original CMax (i.e., $O(N_{p} + N_{e})$ ) due to the search for the highest-ranked events during the optimization.", + "bbox": [ + 89, + 424, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 512, + 222, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the proposed method on four datasets using various metrics (in Sec. 4.1). In the following sections using common benchmarks, we assess denoising (Sec. 4.2), joint motion estimation (Sec. 4.3), and the efficacy on intensity reconstruction (Sec. 4.4). We also conduct sensitivity analyses (Sec. 4.5) and an ablation (Sec. 4.6).", + "bbox": [ + 89, + 537, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets, Metrics, and Hyper-parameters", + "text_level": 1, + "bbox": [ + 89, + 637, + 444, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. E-MLB [10] is a large-scale, de-facto dataset to benchmark denoising. It consists of 100 sequences with 4 different brightness levels using neutral-density (ND) filters on a DAVIS346 camera $(346\\times 260\\mathrm{px})$ during day and night.", + "bbox": [ + 89, + 659, + 482, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Historically, the two sequences (hotel and driving) of the DND21 dataset [24] have been widely used for denoising evaluation [10, 24, 29]. Both sequences are recorded with a DAVIS346 camera [53]. They are regarded as \"signal\" data, as the sequences have been aggressively filtered offline. Combined with other pure BA noise sequences, the dataset has event-wise noise annotations that are useful for evaluation. We use the above sequences with different noise rates, 1, 3, 5, 7, $10\\mathrm{Hz}$ per pixel, following prior work [29].", + "bbox": [ + 89, + 719, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ECD dataset [40] is a standard dataset for various tasks including camera ego-motion estimation [15, 41, 47, 49, 61, 62]. Using a DAVIS240C camera $(240\\times 180\\mathrm{px}$ [7]),", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/73514bff06d4d47271fff8cbe639d757948a6aebf7931d4c5eb2bd9be0f85a54.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
E-MLB (Day)E-MLB (Night)DND21
ND1ND4ND16ND64ND1ND4ND16ND64
Model-basedRaw0.8210.8240.8150.7860.8900.8240.7860.7680.869
BAF [9]0.8610.8690.8760.8900.9460.9730.9920.9420.920
TS [34]0.8770.8870.8700.8371.0330.9440.8860.7970.985
KNoise [31]0.8460.8370.8300.8070.9540.9560.8710.8170.887
EvFlow [54]0.8480.8780.8680.8330.9690.9830.8890.7971.006
IETS [5]0.7720.7850.7770.7530.9500.8230.8040.7110.900
Ynoise [14]0.8660.8630.8570.8211.0090.9430.8750.7920.966
GEF [11]1.0510.9380.9350.9271.0270.9550.9460.9350.932
DWF [24]0.8780.8760.8660.8650.9230.9620.9880.9320.905
Ours0.9380.9580.9860.9501.0370.9610.9450.9320.992
LearningEDnCNN [4]0.8870.9080.9030.9121.0011.0241.0791.0860.977
EventZoom [12]0.9960.9880.9960.9701.0551.0071.0100.9881.059
MLPF [24]0.8510.8550.8460.8400.9260.9280.9100.9060.944
EDformer [29]0.9520.9550.9560.9421.0481.0191.0761.0991.041
", + "bbox": [ + 519, + 89, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d3e91ef88b172889809ec4784cccb1f29e26bb9aae2d3d9bb25ed57a7837fd46.jpg", + "table_caption": [ + "Table 1. Mean ESR (MESR↑) results among denoising methods on the event denoising datasets E-MLB [10] and DND21 [24]. In each category, the best is in bold and the second best is underlined. Methods are chronologically sorted within each category." + ], + "table_footnote": [], + "table_body": "
1Hz5Hz10Hz
hoteldrivinghoteldrivinghoteldriving
Model-basedBAF [9]0.9540.8480.8920.7930.8370.748
TS [34]0.9720.9310.9610.9270.9620.920
KNoise [31]0.6770.6300.6700.6240.6410.614
Ynoise [14]0.9690.9410.9230.9090.8990.880
DWF [24]0.9270.7410.8620.6900.7960.656
Ours1.0140.8820.9630.8550.9610.836
LearningEDnCNN [4]0.9570.8870.9370.8750.9010.874
MLPF [24]0.9700.8890.9700.8850.9630.876
EDformer [29]0.9930.9540.9850.9420.9700.926
", + "bbox": [ + 517, + 359, + 903, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. The AUC $\\uparrow$ of ROC on the two DND21 sequence at different noise rates. We follow [29] for baseline methods selection.", + "bbox": [ + 511, + 521, + 903, + 549 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ef37170daa390a676e23b6f9a7959ad5ea887116686e84b9fcbcda148b57c65.jpg", + "image_caption": [ + "Figure 3. ROC curves on the DND 21 dataset sequences. Dash lines represent learning-based methods (see also Tab. 2)." + ], + "image_footnote": [], + "bbox": [ + 516, + 566, + 906, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each sequence provides events, frames, calibration information, IMU data, and ground truth (GT) camera poses (at $200\\mathrm{Hz}$ ). We use boxesRotation and dynamicRotation sequences to evaluate rotational motion estimation accuracy.", + "bbox": [ + 511, + 747, + 903, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we also use sequences from MVSEC [63] and zurich_city_12a sequence from the DSEC dataset [20], which are typically used for optical flow estimation. We use MVSEC sequences for quantitative evaluation and the DSEC sequence for qualitative evaluation, since it is captured during the night and hence includes lots of noise.", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6962", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b28b00a8fa4d581eb8dab8adb212212295062b5100fe9e4ded79d089fd0d483d.jpg", + "image_caption": [ + "Figure 4. Comparison of denoising methods on the two DND21 sequences at two different noise levels." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 911, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. The metrics used for denoising assessment are: receiver operating characteristic (ROC) curve, area under the curve (AUC) and Mean Event Structural Ratio (MESR). Given signal and noise GT labels, the ROC shows the True-Positive Rate and False-Positive Rate over several S/N ratios. The AUC of the ROC becomes closer to 1, the better the denoiser works. The MESR is a denoising metric that does not require GT labels [10]. We follow previous work to fix the number of signal events to 30000 [10, 29], as well as $M = 20000$ [10, Eq.10], to discount the difference in the number of signal events from various denoising methods. The higher, the better. Furthermore, we use the RMS error for angular velocity estimation and the EPE for optical flow estimation, following previous works [23, 25, 42, 50], as well as the FWL metric [52], which is the relative variance (sharpness) of the IWE with respect to that of the identity warp. Further details about the experimental settings (e.g., hyper-parameters) are in the supplementary.", + "bbox": [ + 89, + 332, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Denoising Results", + "text_level": 1, + "bbox": [ + 89, + 618, + 264, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate denoising accuracy using AUC on DND21 data, and using MESR on E-MLB and DND21 data, using the optical flow estimation.", + "bbox": [ + 89, + 642, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoising Performance on E-MLB. Table 1 shows the benchmark results on the E-MLB dataset using the MESR metric. We categorize prior work into model-based methods and learning-based methods for convenience. Compared with other model-based techniques, our method ranks first or second, showcasing the efficacy of the proposed method on denoising. Even compared with learning-based ones, our method achieves competitive scores (e.g., second in E-MLB (Day) ND4-ND64 conditions, etc). Notice that, however, learning-based approaches utilize additional information (GT labels) during training, while our method does not. Nonetheless, the results demonstrate the efficacy of our approach, which contributes to the theoretical foundation of event-denoising methodologies. We discuss further details", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/55bf99e617c6a07d453f0a17812cdd866fd511e704255aafbd8eb929fd94ebf6.jpg", + "image_caption": [ + "Figure 5. Simultaneous estimation of rotational motion and denoising on the dynamicRotation sequence from ECD dataset [40]. Our method produces sharp IWEs $(2^{\\mathrm{nd}}$ row) as well as reasonable S/N classification (rows 4 and 5), keeping $90\\%$ of data $(\\tau = 0.9)$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 337, + 898, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the evaluation metrics (ESR) in Sec. 5.", + "bbox": [ + 511, + 805, + 790, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results are shown in Fig. 2. The denoised outputs (\"Signal\") are reasonable, as dark (underexposed) regions of the scene become cleaner, and the IWEs using the estimated flow provide sharp edges of the original scenes. We compare with state-of-the-art methods in detail next.", + "bbox": [ + 511, + 824, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6963", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5374923d01492c439392167a7cae19ee330bb9f7501ad14b7f5b00700343d021.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bda697145e928ca91ea383e47c245c88b6e5feb26162fe183be61387addcd6d1.jpg", + "image_caption": [ + "Figure 6. Results on the ego-motion estimation application (top: boxes, bottom: dynamic) [40]. Our method relaxes the dependency on the initial value for CMax-based rotational motion estimation (indicated as \"Gain\")." + ], + "image_footnote": [], + "bbox": [ + 94, + 196, + 478, + 324 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a5dccc26f531ba451c3ff0b93e157d4d97978c9840c08e7b97846f7fd07289d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
dynamic_rotboxes_rot
RMS ↓FWL ↑RMS ↓FWL ↑
5 Hz NoiseCMax [17]20.0011.259124.6411.129
- w/ Init.8.2751.27420.6591.214
Downsampling (Best)8.8081.273124.8691.108
- w/ Init.8.2261.27420.6191.214
Ours (Best)8.5221.27397.5851.141
- w/ Init.8.1701.27420.6041.214
BAF19.6751.260125.0281.127
- w/ Init.8.2531.27420.5501.214
", + "bbox": [ + 114, + 393, + 460, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Angular velocity estimation on ECD dataset [40].", + "bbox": [ + 112, + 547, + 459, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Denoising Accuracy on DND21. Table 2 shows the result of AUC for the ROC result on the DND21 dataset. Our method achieves consistently high AUC compared with existing model-based methods. Note that the DND21 dataset aggressively removes events to generate \"pure-signal\" sequences [24, Sec.5]. Due to this step, the labeled signal events may be fewer than the actual ones corresponding to the true motion. Hence, the DND21 sequences could result in a slight degradation in the False Positive score for motion-based (unsupervised) denoising methods like ours, which may result in lower AUC values. Nonetheless, our method provides competitive AUC scores among other state-of-the-art. The ROC curves are shown in Fig. 3.", + "bbox": [ + 88, + 580, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons among several denoising methods is given in Fig. 4. From the results, our method achieves a good trade-off between under-denoising (preserving not only scene-edge details but some noise) and over-denoising (removing not only most of the noise but some edge details). This can be confirmed, for example, in the upper-right edges in the driving example, and the person's arm in the hotel example. Essentially, all denoising methods have", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "some parameters to control the amount of denoising; our proposed method has the signal ratio, which is intuitive; other optimization methods have some parameters about spatio-temporal kernel size, and learning-based approaches use the probability threshold for the network output. While we fix it for each dataset and condition for fair comparison (we do not tune it for each sequence in the results), this is an essential challenge for denoising algorithms in general, regardless of their underlying character (learning-based or model-based). Hence, we believe it is important for research to discuss denoising effectiveness in combination with downstream tasks, as shown in Secs. 4.3 and 4.4.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Joint Motion Estimation Results", + "text_level": 1, + "bbox": [ + 511, + 286, + 797, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ego-Motion on ECD. One of the merits of the proposed joint estimation method is that it can improve motion estimation results. Here, the results of ego-motion (e.g., angular velocity) estimation are displayed in Fig. 5. We used the classical ECD dataset and injected BA noise from the DND21 recordings with different noise rates. As no prior work simultaneously estimates motion and noise, we compare against sequential filters: random sampling (of the input data with a uniform distribution) and BAF [9] that is the most widely-used denoising method to date. As shown in the IWEs, our method produces the sharpest edges (cf. (b) and (c)) and the most reasonable signal-noise classification among others (cf. (a), (b), and (c)). Note that the original ECD recordings have noise events besides the ones that are injected using DND21 recordings, which makes the calculation of ROC (similar to Sec. 3.2) intricate. Nonetheless, we report ROC results in the supplementary.", + "bbox": [ + 511, + 308, + 906, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative evaluation on angular velocity estimation is reported in Fig. 6, as well as the summary of the supplementary detailed table in Tab. 3. As shown in the first row (\"CMax\"), the original CMax degrades due to the noise occurrences. Our method achieves the best results in terms of RMS and FWL, compared with the random sampling method as well as the original CMax (no denoising). Also, using better initialization (the result of the previous slice for sequential estimation) always improves the results (\"w/ Init.\"): the estimation accuracy relies on both the initialization and the denoising strategies. Hence, the result demonstrates that the proposed method relaxes the dependency on the initialization, thus improving the robustness of motion estimation.", + "bbox": [ + 511, + 566, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optical Flow on MVSEC. The results on more complex motion estimation (e.g., optical flow) are shown in Tab. 4, using the MVSEC dataset [63], with BA noise from the DND21 dataset. Due to the $5\\mathrm{Hz}$ BA noise, the baseline method MultiCM [48] degrades. Our denoising method improves the flow accuracy, showcasing the efficacy of simultaneous noise estimation. A challenge of evaluating denoising in these real-world datasets (e.g., ECD, MVSEC) is the", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6964", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3ddb1665febad69ad99d9fc03cf3a5caae3bd7ca7061ef9354807a4dc85e551.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 85, + 261, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4eb34d949be54c3126c44fb3fa44e6ad6c0c9063c5f338a0d95b341f65e5735b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 157, + 261, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/896daaa895844ecc161cd48140cee724820ebb36597a41a26cd15b2c0477a431.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 88, + 377, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b46025279b1f2cdf8dd2c8e62598906c34f547e2139fa6f6a5c04f2e32a155f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 157, + 377, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb5991c30f64a0eb898abf68d72f10da2a6042ac5a0f4153f94a82b0369a999.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 88, + 493, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33f90d53462ceac23f2e016430c7ff3cdbb66e2fcd9afb497c3d947a2b3a3d8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 157, + 493, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3cc8de1acf621db3440d613e46a69aaccc28e8ae735bd87d415f36719468c8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 88, + 611, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fffe399b3e99b647238adf5e4a0ad4c337c66921c44a793c5163144a31b49d57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 157, + 611, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75d5e2ca95ccb2ab0981ef0a32fa07a7d8ace6b28a2d3ec52ecd4fd3f5ce7cea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 88, + 727, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/00b7f39e62ba0bf310a00dfa3bd7ebd01e5679f5dbcabd611dded2c5c041f6e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 157, + 727, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d26dcecf03b28eef0bc8b251694df919ea84d99e6aba98c37901a83773a33939.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 88, + 844, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7885488cd32a03083a66e1cf95ba1f716d358933996b86e9f408580358a6ea25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 157, + 844, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f771427a458bdd8604627c5fd8b1be9dd97cc27df2cfd4a22a91226f998f7604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 227, + 261, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80c6d0b20a4fef589ed86c26a8418ca09c638715bdb74daa1f90679db93e9aeb.jpg", + "image_caption": [ + "(a) Input events" + ], + "image_footnote": [], + "bbox": [ + 145, + 295, + 261, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0ab63322f5ea2373496e5ae024f33d8b9df7338b0099bf0d020e9da1a61f1d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 227, + 377, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b519167a76ca773bf64bc36bfde8c06b952a4df25a3c5c20b785ed0e99c02d9.jpg", + "image_caption": [ + "(b) Estim. flow" + ], + "image_footnote": [], + "bbox": [ + 261, + 295, + 379, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/611881d6462c94893c3e2b305d25417b5e6648cb8e347cec23773c9b44cab91a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 227, + 495, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31a3b6c530eb861e92ed02ea78450de367460d20d0df378eeb3b11bc8a727acb.jpg", + "image_caption": [ + "(c) Signal events" + ], + "image_footnote": [], + "bbox": [ + 379, + 295, + 493, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5459b703e831b4ac8d0a30e1024fd5c641d7bcad17c61c87ece00e3338ab5ba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 227, + 611, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6d44af299743e42c52aae9aca358a335deb0ba6b93ac90fe73178bc4e9b6626.jpg", + "image_caption": [ + "(d) Reconstr. from (c) signal" + ], + "image_footnote": [], + "bbox": [ + 495, + 295, + 611, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1dff928675a01ad4107b3541d35445e31584b0eb2c086492dfd98924a6849a35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 227, + 727, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8b10ecc5da0cd0d858c27c1dbabd4bb3ce6f2189078c3ba480d41719134f08c.jpg", + "image_caption": [ + "(e) Reconstr. from (a) input" + ], + "image_footnote": [], + "bbox": [ + 612, + 295, + 727, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7840d8e7465f76d67efbd4d4a9522c4dec2a557eda0bf09bc091ea9f5ffc36e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 227, + 844, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb877060ab676d280305608a8d7686e52c3100edb8370d9ff8c7baf50b5b7052.jpg", + "image_caption": [ + "(f) Reference frame" + ], + "image_footnote": [], + "bbox": [ + 728, + 295, + 844, + 364 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db3b0a3fc3e197fea688f070fae6e0a02194d75df0ae7749174af2b56c840840.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
indoor1indoor2
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.67632.0982.74633.474
Ours2.51729.4752.69533.047
indoor3outdoor1
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.73233.2191.96018.160
Ours2.65432.2411.86517.554
", + "bbox": [ + 151, + 436, + 421, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Optical flow estimation results on MVSEC [63] ( $dt = 4$ frames, i.e., $89 \\mathrm{~ms}$ ) with BA noise at $5 \\mathrm{~Hz}$ .", + "bbox": [ + 89, + 571, + 482, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "existing noise in the original sequences. Hence, we only evaluate motion estimation metrics, not ROC or AUC.", + "bbox": [ + 89, + 619, + 482, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Application to Intensity Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 662, + 429, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In real-world use cases, denoising works as preprocessing steps for other tasks (e.g., deblur, 3D reconstruction, motion estimation, etc.), hence it is important for denoising methods to demonstrate applicability. Although the proposed approach intrinsically improves motion estimation (Sec. 4.3), we further evaluate the denoising outcome via intensity reconstruction, a well-known task in event-based vision, using the E-MLB dataset [10]. Figure 7 collects the results. We use two state-of-the-art approaches for reconstruction: E2VID [46], which does not rely on motion, and EVILIP [59], which utilizes optical flow as prior.", + "bbox": [ + 89, + 686, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method provides flow (Fig. 7, column (b)) and signal events (sharp IWEs in column (c)). Comparing the reconstruction with our event denoising method (column (d)) and", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2942f1d5329e90790c245bb3b0ffa43e4a385cb733d3fe774bf00e468932af3c.jpg", + "image_caption": [ + "Figure 7. Image intensity reconstruction on E-MLB data [10]. The first two rows use the reconstruction method EVILIP [59] and the last two rows use E2VID [46]. Reconstruction on denoised events produces higher quality images (i.e., fewer artifacts) than on raw event data." + ], + "image_footnote": [], + "bbox": [ + 522, + 446, + 772, + 517 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d848c24e69a6b0310d66bc0ab07dafcecc1b0f870f12b08f4bdcf9ab58ac0da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 517, + 772, + 587 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e3f08c4be859530b52f646523d3008f77d225d5566f79a73cfaf72886384212.jpg", + "image_caption": [ + "(a) Signal", + "(b) Noise" + ], + "image_footnote": [], + "bbox": [ + 522, + 587, + 772, + 657 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e47a2e47b6178fc312284ab664e59473cb3a6bc1f269840816557a453b40025c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 448, + 893, + 518 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a93237ea407609e8ed6854f97b1b8b2e1dd8533b98211bd71b2ed4b3888904a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 518, + 893, + 589 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1090b5d162cf6e76d1326c8bf1468a8da31b27f27afbdba7f2e28b53f23105e7.jpg", + "image_caption": [ + "(c) Reconstruction", + "Figure 8. Sensitivity analysis: effect of different target ratio parameters $\\tau$ on image reconstruction using EVILIP [59]." + ], + "image_footnote": [], + "bbox": [ + 774, + 589, + 893, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the reconstruction from the raw event data (column (e)), the images after event denoising have fewer artifacts than those without denoising, for both reconstruction methods tested. We observe that E2VID is more robust against noise than EVILIP, possibly due to noise events in its training (simulation) data. Due to the ND filters, the quality of the reference frames (last column) is limited, and hence, we discuss the results qualitatively. Nevertheless comparing the reconstructed images to the reference frames, evidences the HDR advantages of event cameras over frame-based ones.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6965", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8114e778fee861c5ec5b676583eb98ef667d0c2040669fad9a5bf24702675f5c.jpg", + "image_caption": [ + "(a) Ref. events" + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 222, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dad65955ddb772fddc961fde6e2fae628a52c0fa9bd7e7ea4a8e82ad65d9c5b4.jpg", + "image_caption": [ + "(b) $k = 1$ px", + "Figure 9. Results of different Gaussian kernel sizes $k$ for sensitivity on different depth scenes." + ], + "image_footnote": [], + "bbox": [ + 225, + 88, + 344, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/382280d018ace21f9cf8d0e6647a8011ff7758662371f1eda87bbce7c5c783b0.jpg", + "image_caption": [ + "(c) $k = 3$ px" + ], + "image_footnote": [], + "bbox": [ + 346, + 88, + 468, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Sensitivity", + "text_level": 1, + "bbox": [ + 89, + 220, + 207, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Noise Ratio. As Sec. 4.2 already showed the sensitivity for denoising accuracy, let us further discuss the sensitivity using image reconstruction. Figure 8 presents the result of the Bicycle sequence with ND64 filter [10]. While the GT image is highly underexposed and not informative (as those on the last column of Fig. 7), the reconstructed images clearly reflect the choice of the signal target ratio $\\tau$ . As more events are removed, $(i)$ the reconstructed image becomes more homogeneous but preserves strong edges, and $(ii)$ the removed events (2nd col.) contain more scene edges (i.e., signal). We quantitatively analyze the former in the supplementary using non-reference image quality metrics.", + "bbox": [ + 89, + 246, + 483, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Kernel Size. Figure 9 shows the sensitivity analysis for different kernel sizes. By increasing the size of the Gaussian kernel in the IWE, our method can emphasize contour edges over isolated points, thus better preserving signal events in regions with low IWE intensities (center of the image, corresponding to far away scene points).", + "bbox": [ + 89, + 429, + 483, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation of Motion Estimator", + "text_level": 1, + "bbox": [ + 89, + 535, + 354, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed approach can be combined with other estimation methods, such as deep-neural-network (DNN) flow estimators. To this end, we show an example of E-RAFT [21]. The model-based method alone [48] is known to output a degraded flow in night sequences with lots of noise, such as zurich_city_12a [20]. As shown in Fig. 10, our method can identify and discard noisy events. E-RAFT provides sharper and more natural edges in IWEs after denoising compared to MultiCM [48] (thanks to its more accurate flow). ESR scores are similar: 2.39 for E-RAFT and 2.45 for MultiCM. While we focus on the inference of DNN in the proposed pipeline, future research could look into accommodating training within the proposed joint estimation framework.", + "bbox": [ + 89, + 559, + 482, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Runtime", + "text_level": 1, + "bbox": [ + 89, + 770, + 194, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using 30k events from the ECD dataset $(240 \\times 180\\mathrm{px})$ on a Mac M1 CPU (2020), the denoising operations (scoring and ranking) take less than $0.1\\mathrm{s}$ extra computation in total, out of $2\\mathrm{s}$ of the entire processing time (i.e., $5\\%$ computation increase with respect to the base method). Nonetheless, accommodating learning-based approaches in our framework could be key to the trade-off between accuracy and speed.", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b6abb16bd161e709d7bd0ac6292b7f59ef6f681827178c5c494a392c0ce2ae2.jpg", + "image_caption": [ + "Raw Event" + ], + "image_footnote": [], + "bbox": [ + 535, + 97, + 656, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06d32f0c215e5bc8e90658d248159d2e5d05798bdac2631a711e2e7c202f3ec9.jpg", + "image_caption": [ + "IWE (No Warp)" + ], + "image_footnote": [], + "bbox": [ + 656, + 98, + 774, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a4bba47789dd768ef59eda58d7f2d069a6a397fa374df9858dbbdfab1a7a1eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 179, + 656, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55b6d476952e74d25445e291eb3240e7f889093496c1f4a41c9b74846768e6f6.jpg", + "image_caption": [ + "(a) Signal Events", + "Figure 10. Effect of adopting two different motion estimators (E-RAFT or MultiCM) on the DSEC night scene (zurich_12a) [20]. Despite small IWE differences (e.g., in the car's hood), the proposed framework supports several motion estimators." + ], + "image_footnote": [], + "bbox": [ + 522, + 247, + 656, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adc7580ce076347908892112af1c604d5112ed7fc73d24fbecbc84f0ffa1d0b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 180, + 774, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99e035bdcbd5ae0e10cd26b64bd30e41b443b1529dc6dbfb309a5ec5b8de37e4.jpg", + "image_caption": [ + "(b) IWE" + ], + "image_footnote": [], + "bbox": [ + 658, + 250, + 774, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53fe5b6e7247e26af7ed36683d7ad1ac96975aca3bca3e49aff6dee3f50caba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 180, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dd08aff9896dbb3b06d5cb095a14ee56deb45039f14ace0ac1ffb0f718fe5b9.jpg", + "image_caption": [ + "(c) Flow Estimation" + ], + "image_footnote": [], + "bbox": [ + 779, + 251, + 893, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 513, + 409, + 635, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method expects as parameter a target noise level, but the intrinsic (\"true\") noise level depends on the camera, scene, etc. and is therefore unknown. More research in this direction is needed for automatic parameter tuning. Also, notice that there might be a potential limitation of the denoising metric ESR, since it depends on the motion parameter (see Suppl. Mat. of [29]). Although our denoising efficacy is confirmed from multiple perspectives, we believe that advancing denoising metrics for real-world data (where it gets interesting due to the absence of GT labels) is a central direction for future work in event-based vision.", + "bbox": [ + 511, + 435, + 906, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, this work proposes that: (i) Joint estimation of noise and motion is a novel problem setting compared to conventional (isolated) denoising, and is sensitive given the event data's inherent motion dependency. (ii) The proposed joint estimation framework can accommodate various motion estimators and has shown its efficacy in improving both denoising and motion estimation accuracies. (iii) It is important to evaluate denoising results on both, denoising benchmarks and downstream tasks (e.g., motion estimation, intensity reconstruction), since current denoising benchmarks have shortcomings in terms of metrics, data, and baselines. In particular, our method achieves notable denoising results (state-of-the-art on the E-MLB benchmark and competitive on DND21) while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances both event-denoising theory and practical denoising use-cases.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6966", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Dr. B. Jiang for useful discussions. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135.", + "bbox": [ + 89, + 114, + 485, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Saeed Afshar, Andrew Peter Nicholson, Andre van Schaik, and Gregory Cohen. Event-based object detection and tracking for space situational awareness. IEEE Sensors Journal, 20(24):15117-15132, 2020.", + "[2] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Trans. Neural Netw. Learn. Syst., 35(3):4110-4124, 2022.", + "[3] Sami Arja, Alexandre Marcireau, Richard L. Balthazor, Matthew G. McHarg, Saeed Afshar, and Gregory Cohen. Density invariant contrast maximization for neuromorphic earth observations. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 3984-3994, 2023.", + "[4] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1701-1710, 2020.", + "[5] R Wes Baldwin, Mohammed Almatrafi, Jason R Kaufman, Vijayan Asari, and Keigo Hirakawa. Inceptive event timesurfaces for object classification using neuromorphic cameras. In Image Analysis and Recognition: 16th Int. Conf., ICIAR 2019, Waterloo, ON, Canada, August 27-29, 2019, Proceedings, Part II 16, pages 395-403. Springer, 2019.", + "[6] T. D. Barfoot. State Estimation for Robotics - A Matrix Lie Group Approach. Cambridge University Press, 2015.", + "[7] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240x180 130dB $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE J. Solid-State Circuits, 49(10):2333-2341, 2014.", + "[8] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In AAAI Conf. Artificial Intell., pages 303-311, 2022.", + "[9] Tobi Delbruck. Frame-free dynamic digital vision. In Proc. Int. Symp. Secure-Life Electron., pages 21-26, 2008.", + "[10] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-MLB: Multilevel benchmark for event-based camera denoising. IEEE Trans. Multimedia, 26:65–76, 2023.", + "[11] Peiqi Duan, Zihao Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2021.", + "[12] Peiqi Duan, Zihao W. Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve" + ], + "bbox": [ + 93, + 227, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neuromorphic events. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12824-12833, 2021.", + "[13] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022.", + "[14] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020.", + "[15] Guillermo Gallego and Davide Scaramuzza. Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett., 2(2):632-639, 2017.", + "[16] Guillermo Gallego and Anthony Yezzi. A compact formula for the derivative of a 3-D rotation in exponential coordinates. J. Math. Imaging Vis., 51(3):378-384, 2014.", + "[17] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3867-3876, 2018.", + "[18] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vision. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12272-12281, 2019.", + "[19] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154-180, 2022.", + "[20] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett., 6(3):4947-4954, 2021.", + "[21] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vision (3DV), pages 197–206, 2021.", + "[22] Rui Graca and Tobi Delbruck. Towards a physically realistic computationally efficient dvs pixel model. In Int. Image Sensor Workshop (IISW), 2025.", + "[23] Cheng Gu, Erik Learned-Miller, Daniel Sheldon, Guillermo Gallego, and Pia Bideau. The spatio-temporal Poisson point process: A simple model for the alignment of event camera data. In Int. Conf. Comput. Vis. (ICCV), pages 13495-13504, 2021.", + "[24] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Trans. Pattern Anal. Mach. Intell., 45(1):785-795, 2023.", + "[25] Shuang Guo and Guillermo Gallego. CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot., 40: 2442-2461, 2024.", + "[26] Shasha Guo, Ziyang Kang, Lei Wang, Shiming Li, and Weixia Xu. Hashheat: An O(C) complexity hashing-based" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6967", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "filter for dynamic vision sensor. In 25th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 452-457, 2020.", + "[27] Shuang Guo, Friedhelm Hamann, and Guillermo Gallego. Unsupervised joint learning of optical flow and intensity with event cameras. In Int. Conf. Comput. Vis. (ICCV), 2025.", + "[28] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In Eur. Conf. Comput. Vis. (ECCV), pages 18-37, 2024.", + "[29] Bin Jiang, bo Xiong, Bohan Qu, M. Slaman Asif, you Zhou, and Zhan Ma. EDformer: Transformer-based event denoising across varied noise levels. In Eur. Conf. Comput. Vis. (ECCV), pages 1-12, 2024.", + "[30] Pritam P. Karmokar, Quan H. Nguyen, and William J. Beksi. Secrets of edge-informed contrast maximization for event-based vision. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), 2025.", + "[31] Alireza Khodamoradi and Ryan Kastner. O(N)-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Trans. Emerg. Topics Comput., PP(99):1-1, 2018.", + "[32] Haram Kim and H. Jin Kim. Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett., 6(3):6016-6023, 2021.", + "[33] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), pages 8601-8608, 2021.", + "[34] Xavier Lagorce, Garrick Orchard, Francesco Gallupi, Bertram E. Shi, and Ryad Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(7):1346-1359, 2017.", + "[35] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128x128 120dB 30mW asynchronous vision sensor that responds to relative intensity change. In IEEE Int. Solid-State Circuits Conf. (ISSCC), pages 2060-2069, 2006.", + "[36] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ dB 15 μs latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits, 43(2):566-576, 2008.", + "[37] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In IEEE Int. Symp. Circuits Syst. (ISCAS), pages 722-725, 2015.", + "[38] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012.", + "[39] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2012.", + "[40] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research, 36(2):142-149, 2017." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot., 34(6):1425-1440, 2018.", + "[42] Urbano Miguel Nunes and Yiannis Demiris. Robust event-based vision model estimation by dispersion minimisation. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9561-9573, 2022.", + "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3445–3454, 2021.", + "[44] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE de Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Int. Conf. Comput. Vis. (ICCV), pages 9661-9671, 2023.", + "[45] Xin Peng, Ling Gao, Yifu Wang, and Laurent Kneip. Globally-optimal contrast maximisation for event cameras. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3479-3495, 2022.", + "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Trans. Pattern Anal. Mach. Intell., 43 (6):1964-1980, 2021.", + "[47] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high speed scenarios. IEEE Robot. Autom. Lett., 3(2): 994-1001, 2018.", + "[48] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis. (ECCV), pages 628-645, 2022.", + "[49] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):1-20, 2022.", + "[50] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):7742-7759, 2024.", + "[51] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Int. Conf. Comput. Vis. (ICCV), pages 7243-7252, 2019.", + "[52] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In Eur. Conf. Comput. Vis. (ECCV), pages 534-549, 2020.", + "[53] Gemma Taverni, Diederik Paul Moeys, Chenghan Li, Celso Cavaco, Vasyl Motsnyi, David San Segundo Bello, and Tobi Delbruck. Front and back illuminated Dynamic and Active Pixel Vision Sensors comparison. IEEE Trans. Circuits Syst. II (TCSII), 65(5):677-681, 2018.", + "[54] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors." + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6968", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6358-6367, 2019.", + "[55] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1606–1616, 2020.", + "[56] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Trans. Multimedia, 23:1148-1159, 2020.", + "[57] Ryo Yamaki, Shintaro Shiba, Gallego Guillermo, and Yoshimitsu Aoki. Iterative event-based motion segmentation by variational contrast maximization. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 4918-4927, 2025.", + "[58] Pei Zhang, Haosen Liu, Zhou Ge, Chutian Wang, and Edmund Y Lam. Neuromorphic imaging with joint image deblurring and event denoising. IEEE Trans. Image Process., 2024.", + "[59] Zelin Zhang, Anthony Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8372-8389, 2023.", + "[60] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Trans. Neural Netw. Learn. Syst., pages 1-13, 2021.", + "[61] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Trans. Robot., 37(5):1433-1450, 2021.", + "[62] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5816-5824, 2017.", + "[63] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. Autom. Lett., 3(3):2032-2039, 2018.", + "[64] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 989-997, 2019." + ], + "bbox": [ + 91, + 90, + 482, + 713 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6969", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_model.json b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0925c2c7819ec2e03b81f4375bc19a29155504dc --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_model.json @@ -0,0 +1,2829 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.828, + 0.151 + ], + "angle": 0, + "content": "Simultaneous Motion And Noise Estimation with Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.18, + 0.756, + 0.199 + ], + "angle": 0, + "content": "Shintaro Shiba\\(^{1,2,3}\\), Yoshimitsu Aoki\\(^{1}\\) and Guillermo Gallego\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.199, + 0.9, + 0.235 + ], + "angle": 0, + "content": "1 Keio University, 2 Woven by Toyota, Inc., Japan. 3 Technische Universität Berlin, 4 Einstein Center Digital Future, Robotics Institute Germany, and Science of Intelligence Excellence Cluster, Germany." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.604 + ], + "angle": 0, + "content": "Event cameras are emerging vision sensors whose noise is challenging to characterize. Existing denoising methods for event cameras are often designed in isolation and thus consider other tasks, such as motion estimation, separately (i.e., sequentially after denoising). However, motion is an intrinsic part of event data, since scene edges cannot be sensed without motion. We propose, to the best of our knowledge, the first method that simultaneously estimates motion in its various forms (e.g., ego-motion, optical flow) and noise. The method is flexible, as it allows replacing the one-step motion estimation of the widely-used Contrast Maximization framework with any other motion estimator, such as deep neural networks. The experiments show that the proposed method achieves state-of-the-art results on the E-MLB denoising benchmark and competitive results on the DND21 benchmark, while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances event-data denoising theory and expands practical denoising use-cases via open-source code. Project page: https://github.com/tub-rip/ESMD" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.222, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Event cameras are emerging vision sensors that overcome shortcomings of conventional cameras (e.g., motion blur, limited dynamic range, data and power efficiency), yet they suffer from a considerable amount of noise due to their novelty and operation in low-power (transistor subthreshold) conditions [19, 36]. Since event cameras are suitable for many computer vision tasks, especially motion-related tasks, it is paramount to classify event data that is related to motion (i.e., signal) and that is not (i.e., noise). However, denoising event data is a challenging problem because noise properties have yet to be fully characterized [22, 24] and it is infeasible to define ground-truth (GT) noise labels in real-world data recordings. Previous works either leveraged simulation or prepared real event data and aggressively suppressed noise occurrence." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.9 + ], + "angle": 0, + "content": "However, such artificial modifications on the real event" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.27, + 0.905, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed method. We use only raw events as input, and estimate signal events, noise events, and motion (e.g., dense optical flow and ego-motion estimation)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.806 + ], + "angle": 0, + "content": "data could change event signal/noise characteristics, possibly deteriorating the information that the original event had for downstream tasks (e.g., motion estimation). Our work is inspired by the observation that motion is an intrinsic part of event data since scene edges cannot be sensed without motion. Rather than solving denoising and motion estimation tasks separately, our idea is to leverage motion information to improve denoising (as \"vice-versa\" is well known). In this work (Fig. 1), we rethink and investigate the connection between denoising and motion estimation, by asking ourselves whether they can be combined in a more integrated way to benefit the results of both tasks simultaneously. We formulate the simultaneous estimation problem by leveraging the idea of motion compensation in the Contrast Maximization (CMax) framework, defining novel metrics about the contribution to the contrast score for each event. In summary, we present several distinctive contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "1. We propose a novel method to jointly estimate noise and motion only from event data (Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.9 + ], + "angle": 0, + "content": "2. We demonstrate that our joint estimation method \\((i)\\) achieves state-of-the-art results in existing denoising benchmarks (Sec. 4.2), and \\((ii)\\) improves the robustness of motion estimation (CMax framework) (Sec. 4.3)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "3. We show the flexibility of our method on \\((i)\\) further applications, such as intensity reconstruction, validating its image quality (Secs. 4.4 and 4.5), and \\((ii)\\) in combination with learning-based motion estimators (Sec. 4.6)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.198 + ], + "angle": 0, + "content": "We hope our work opens a new line of research for event cameras, by considering the problems of motion estimation and denoising together to take advantage of their coaction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.233, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.388 + ], + "angle": 0, + "content": "Denoising is a fundamental problem in event cameras since there are a lot of leakage events [24], and event-processing algorithms may suffer in scenarios with high noise [3]. Removing event noise has a double effect: it reduces the load of subsequent processing stages (as the load is proportional to the number of events) and improves their output (as the input data becomes of higher quality). In this work, we focus on noise due to the leakage inside the camera hardware, such as background activity (BA) events [36]. Events due to flickering or active lights are therefore not our priority." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.508 + ], + "angle": 0, + "content": "Classical denoising works focus on the spatio-temporal correlation of events, as BA events have fewer neighboring events in space-time coordinates than signal events. This approach includes widely-used BA Filter (BAF) [9], and other types of spatio-temporal filters [26, 31, 37, 56]. Some works utilize additional sensor inputs (e.g., frames) to estimate noise probabilities [4, 12, 30, 55, 58], since edges are informative of the signal events produced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Recently, learning-based approaches have gained more attention, such as multilayer perceptrons [24], convolutional neural networks [1, 4, 12], graph neural networks [2], and transformers [29]. They provide the probability of noise (i.e., signal-noise classification). Nonetheless, the challenge of learning-based approaches is the need for ground truth (GT) noise labels for training. GT labels have been obtained through simulation (e.g., [13, 24]) as well as real-world recordings, however, real-world recordings aggressively remove noisy events using existing filters (e.g., [24]) or use unnatural illumination settings (e.g., [29]), by leveraging the differences of noise occurrence due to different scene brightness [24, 35]. While these approaches are effective for benchmarking, the challenge of obtaining GT labels for practical real-world datasets remains unsolved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our work builds upon a series of successful methods for motion estimation and their applications. Event-based motion estimation is an important topic since event cameras naturally respond to motion in the scenes [19]. In particular, Contrast Maximization (CMax) [17, 50] is a state-of-the-art framework for motion estimation, which has various applications, such as rotational motion estimation [15, 23, 32, 42, 45], optical flow estimation [27, 28, 44, 48, 64], and motion segmentation [51, 57, 60]. We investigate the connection and potential benefits of applying the CMax framework to denoising during the estimation process (i.e., simul" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "taneously), demonstrating motion estimation improvements as well as competitive denoising accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.304 + ], + "angle": 0, + "content": "A prior work, \"ProgressiveMotionSeg\" [8], has extended the CMax-based segmentation method [51] with extra weights that accounted for how much each event contributed to the estimated motion clusters. However, [8] mainly focuses on the task of motion segmentation using low degrees-of-freedom (DOF) warps (e.g., 2-DOF feature flow) and does not discuss denoising efficacy or evaluate on standard denoising benchmarks. In contrast, our method supports various types of motion estimation (both low-DOF and high-DOF warps, such as optical flow), and considers that noise shall not be used to estimate motion, as from first principles we deem noise and motion to be uncorrelated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.649, + 0.332 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.906, + 0.462 + ], + "angle": 0, + "content": "Event cameras consist of independent pixels that generate asynchronous \"events\" when the logarithmic brightness at the pixel increases or decreases by a predefined contrast sensitivity. Each event \\( e_k \\doteq (\\mathbf{x}_k, t_k, p_k) \\) contains the pixel coordinates \\( \\mathbf{x}_k \\), the timestamp \\( t_k \\), and polarity \\( p_k \\in \\{+1, -1\\} \\) of the brightness change. Events occur asynchronously and sparsely on the pixel grid, resulting in a variable data rate based on the scene texture and dynamics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.73, + 0.484 + ], + "angle": 0, + "content": "3.1. Contrast Maximization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.555 + ], + "angle": 0, + "content": "The CMax framework [17] assumes events \\(\\mathcal{E} \\doteq \\{e_k\\}_{k=1}^{N_e}\\) are caused by moving edges, and it transforms their coordinates according to a motion model \\(\\mathbf{W}\\), producing a set of warped events \\(\\mathcal{E}'_{t_{\\mathrm{ref}}} \\doteq \\{e_k'\\}_{k=1}^{N_e}\\) at a reference time \\(t_{\\mathrm{ref}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.562, + 0.905, + 0.579 + ], + "angle": 0, + "content": "\\[\ne _ {k} \\dot {=} \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\mapsto e _ {k} ^ {\\prime} \\dot {=} \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.629 + ], + "angle": 0, + "content": "The warp \\(\\mathbf{x}_k^{\\prime} = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta})\\) transports each event from \\(t_k\\) to \\(t_{\\mathrm{ref}}\\) along the motion curve that passes through it. Then, they are aggregated on an image of warped events (IWE):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.635, + 0.905, + 0.656 + ], + "angle": 0, + "content": "\\[\nI (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k} ^ {\\prime}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.703 + ], + "angle": 0, + "content": "where each pixel \\(\\mathbf{x}\\) sums the number of warped events \\(\\mathbf{x}_k^\\prime\\) that fall within it. The Dirac delta is approximated by a Gaussian, \\(\\delta (\\mathbf{x} - \\pmb {\\mu})\\approx \\mathcal{N}(\\mathbf{x};\\pmb {\\mu},\\epsilon^2\\mathrm{Id})\\) with \\(\\epsilon = 1\\) pixel." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.704, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Next, an objective function, such as the contrast of the IWE (2) (image variance [18] \\(\\mathrm{Var}\\big(I(\\mathbf{x};\\pmb {\\theta})\\big)\\doteq \\frac{1}{|\\Omega|}\\int_{\\Omega}(I(\\mathbf{x};\\pmb {\\theta}) - \\mu_I)^2 d\\mathbf{x})\\) , measures the goodness of fit between the events and the candidate motion curves (warp)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Finally, an optimization algorithm iterates the above steps to find the motion parameters \\(\\theta^{*}\\) that maximize the alignment of events caused by the same scene edge. Event alignment is measured by the strength of the edges of the IWE, which is related to image contrast [18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Motion Models. As exemplary warps \\(\\mathbf{W}\\) we focus on two popular problem settings: rotational motion (3-DOF angular velocity estimation [23, 25, 42, 45]) and dense optical flow (pixel-wise velocity estimation [43, 48, 50, 64]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.098, + 0.887, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.416, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Qualitative result of denoising and flow estimation on E-MLB and DND21. The driving sequence has \\(5\\mathrm{-Hz}\\) artificial noise. (b) is an alternative visualization of (a) with the same grayscale coding as (d). Flow is visualized on all events and flow color is given in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.593 + ], + "angle": 0, + "content": "resulting in \\(2N_{p}\\)-DOFs, where \\(N_{p}\\) is the number of pixels). For the angular velocity estimation during a small time interval, the warp is parametrized by the angular velocity \\(\\theta \\equiv \\omega = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}\\), and \\(\\mathbf{x}^{h}(t)\\sim \\mathbb{R}(\\boldsymbol {\\omega}t)\\mathbf{x}^{h}(0)\\), using calibrated homogeneous coordinates \\(\\mathbf{x}^{h}\\) and exponential coordinates \\(\\mathbb{R}(\\phi)\\doteq \\exp (\\phi^{\\wedge})\\) [6, 16]. The warp for dense optical flow is \\(\\mathbf{x}_k^\\prime = \\mathbf{x}_k + (t_k - t_{\\mathrm{ref}})\\mathbf{v}(\\mathbf{x}_k)\\), where \\(\\theta = \\{\\mathbf{v}(\\mathbf{x})\\}_{\\mathbf{x}\\in \\Omega}\\) is a flow field on the image plane \\(\\Omega\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.6, + 0.398, + 0.617 + ], + "angle": 0, + "content": "3.2. Denoising Based on Local Contrast" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.789 + ], + "angle": 0, + "content": "As Fig. 1 shows, our method is iterative. In each iteration, motion estimation is refined and events are classified into signal \\(\\mathcal{E}_s\\) or noise \\(\\mathcal{E}_n\\). Without labels or additional information, events rely on each other to realize this classification. To this end, we compute a score \\(c_k\\) for each event, rank the events according to their scores, and threshold at some value \\(T\\). The threshold is computed by assuming the amount or ratio of noisy events is known \\(\\eta\\), i.e., \\(T \\equiv T(\\eta)\\), which allows us to have high-level control over the denoising process. Hence we classify events into signal (the \\(\\tau \\doteq (1 - \\eta)\\) percentage of events with highest scores) and noise:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.796, + 0.484, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {s} \\equiv \\mathcal {E} _ {\\text {s i g n a l}} \\doteq \\left\\{e _ {k} \\in \\mathcal {E} \\mid c _ {k} > T (\\eta) \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.816, + 0.442, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {n} \\equiv \\mathcal {E} _ {\\text {n o i s e}} \\dot {=} \\mathcal {E} \\backslash \\mathcal {E} _ {\\text {s i g n a l}} \\quad (\\text {c o m p l e m e n t} \\mathcal {E} _ {\\text {s i g n a l}}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.904 + ], + "angle": 0, + "content": "We studied different choices for the per-event score \\( c_k \\) based on the agreement among warped events \\( \\mathcal{E}' = \\mathcal{E}_s' \\cup \\mathcal{E}_n' \\), through their IWEs, and concluded on two, as follows. Letting \\( I_k^s \\equiv I_k^s(\\mathbf{x}_k') \\), \\( I_k^n \\equiv I_k^n(\\mathbf{x}_k') \\) and \\( I_k \\equiv I(\\mathbf{x}_k') = I_k^s + I_k^n \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.592 + ], + "angle": 0, + "content": "be the IWE values of previously classified signal events, noise events and all events, respectively, the two choices are: (i) \\( ^{1}c_{k} \\doteq I_{k} \\), and (ii) the ratio of signal events \\( ^{2}c_{k} \\doteq I_{k}^{s} / I_{k} \\) (probabilistic viewpoint). In fact, both are equivalent: rule \\( ^{1}c_{k} > T_{1} \\) is the same as rule \\( ^{2}c_{k} > T_{2} \\) if \\( T_{2} = 1 - I_{k}^{n} / T_{1} \\) and the event noise \\( I_{k}^{n} \\) is known. We use \\( ^{1}c_{k} \\), the number of warped events (also called local contrast [18, 51]), because it is the simplest to interpret." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.776 + ], + "angle": 0, + "content": "The intuition is that \"signal\" events (e.g., due to motion) warped according to the true motion lie all at motion-compensated edges, and therefore produce a sharp IWE. Noise events, produced all over the image plane, are not expected to contribute to the IWE's sharpness. Hence, we may classify events into signal or noise according to how much they contribute to edge strength. The IWE (2) provides one such measure of edge strength: the higher \\(I(\\mathbf{x})\\), the more events are warped to that pixel location \\(\\mathbf{x}\\) (i.e., more events support the same scene edge), thus producing a sharper IWE. Hence, we define how much an event \\(e_k\\) contributes to edge strength by means of the IWE (2), \\(^1 c_k\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Invariance. Note that (3) is invariant to monotonically increasing functions of \\( c_k \\), as such functions preserve the order. The same \\( \\mathcal{E}_{\\mathrm{signal}} \\) is obtained if one ranks and thresholds (at the corresponding level) values \\( \\{c_k\\} \\), \\( \\{\\ln(c_k)\\} \\), \\( \\{e^{c_k}\\} \\), etc. They are equivalent under the ranking rule (3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Robustness to Various Edge Strengths. The Gaussian kernel in the IWE controls the sensitivity to edge strength. Since the proposed method is based on both signal and noise" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "6961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "IWEs, increasing the size of the Gaussian kernel in the IWE can emphasize contour edges over isolated points. This preserves signal events better in regions with low IWE intensities, for example at pixels with large depth values." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.323, + 0.178 + ], + "angle": 0, + "content": "3.3. Alternating Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Classifying events into signal and noise requires knowledge of the true motion, and estimation of the true motion (e.g., using CMax) requires knowledge of the signal events because, by definition, noise events carry no information about motion. Hence, it is a circular dependency problem." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.258, + 0.483, + 0.424 + ], + "angle": 0, + "content": "We approach its solution iteratively, as shown in Fig. 1, by combining motion estimation (Sec. 3.1) and S/N classification (Sec. 3.2). Assuming some initialization, the current signal events \\(\\mathcal{E}_{\\mathrm{signal}}\\) are used to estimate the motion via CMax (1 step is enough); then all events are warped to compute the IWE and the per-event scores \\(c_{k}\\), from which the sets (3) are recomputed, in preparation for the next iteration. In practice, we initialize with a random split of \\(\\mathcal{E}\\) into \\(\\mathcal{E}_{\\mathrm{signal}}\\) and \\(\\mathcal{E}_{\\mathrm{noise}}\\), update \\(\\mathcal{E}_{\\mathrm{signal}}\\) at every warp during the optimization iteration until the process converges (as flagged by the convergence of the motion parameters in CMax)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Computational Complexity. The computational complexity of one iteration of the proposed method is \\( O(N_{p} + N_{e}\\log N_{e}) \\), which is slightly larger than the original CMax (i.e., \\( O(N_{p} + N_{e}) \\)) due to the search for the highest-ranked events during the optimization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.513, + 0.223, + 0.53 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.63 + ], + "angle": 0, + "content": "We evaluate the proposed method on four datasets using various metrics (in Sec. 4.1). In the following sections using common benchmarks, we assess denoising (Sec. 4.2), joint motion estimation (Sec. 4.3), and the efficacy on intensity reconstruction (Sec. 4.4). We also conduct sensitivity analyses (Sec. 4.5) and an ablation (Sec. 4.6)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.638, + 0.445, + 0.655 + ], + "angle": 0, + "content": "4.1. Datasets, Metrics, and Hyper-parameters" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Datasets. E-MLB [10] is a large-scale, de-facto dataset to benchmark denoising. It consists of 100 sequences with 4 different brightness levels using neutral-density (ND) filters on a DAVIS346 camera \\((346\\times 260\\mathrm{px})\\) during day and night." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Historically, the two sequences (hotel and driving) of the DND21 dataset [24] have been widely used for denoising evaluation [10, 24, 29]. Both sequences are recorded with a DAVIS346 camera [53]. They are regarded as \"signal\" data, as the sequences have been aggressively filtered offline. Combined with other pure BA noise sequences, the dataset has event-wise noise annotations that are useful for evaluation. We use the above sequences with different noise rates, 1, 3, 5, 7, \\(10\\mathrm{Hz}\\) per pixel, following prior work [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The ECD dataset [40] is a standard dataset for various tasks including camera ego-motion estimation [15, 41, 47, 49, 61, 62]. Using a DAVIS240C camera \\((240\\times 180\\mathrm{px}\\) [7])," + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.905, + 0.278 + ], + "angle": 0, + "content": "
E-MLB (Day)E-MLB (Night)DND21
ND1ND4ND16ND64ND1ND4ND16ND64
Model-basedRaw0.8210.8240.8150.7860.8900.8240.7860.7680.869
BAF [9]0.8610.8690.8760.8900.9460.9730.9920.9420.920
TS [34]0.8770.8870.8700.8371.0330.9440.8860.7970.985
KNoise [31]0.8460.8370.8300.8070.9540.9560.8710.8170.887
EvFlow [54]0.8480.8780.8680.8330.9690.9830.8890.7971.006
IETS [5]0.7720.7850.7770.7530.9500.8230.8040.7110.900
Ynoise [14]0.8660.8630.8570.8211.0090.9430.8750.7920.966
GEF [11]1.0510.9380.9350.9271.0270.9550.9460.9350.932
DWF [24]0.8780.8760.8660.8650.9230.9620.9880.9320.905
Ours0.9380.9580.9860.9501.0370.9610.9450.9320.992
LearningEDnCNN [4]0.8870.9080.9030.9121.0011.0241.0791.0860.977
EventZoom [12]0.9960.9880.9960.9701.0551.0071.0100.9881.059
MLPF [24]0.8510.8550.8460.8400.9260.9280.9100.9060.944
EDformer [29]0.9520.9550.9560.9421.0481.0191.0761.0991.041
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.288, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Table 1. Mean ESR (MESR↑) results among denoising methods on the event denoising datasets E-MLB [10] and DND21 [24]. In each category, the best is in bold and the second best is underlined. Methods are chronologically sorted within each category." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.361, + 0.905, + 0.512 + ], + "angle": 0, + "content": "
1Hz5Hz10Hz
hoteldrivinghoteldrivinghoteldriving
Model-basedBAF [9]0.9540.8480.8920.7930.8370.748
TS [34]0.9720.9310.9610.9270.9620.920
KNoise [31]0.6770.6300.6700.6240.6410.614
Ynoise [14]0.9690.9410.9230.9090.8990.880
DWF [24]0.9270.7410.8620.6900.7960.656
Ours1.0140.8820.9630.8550.9610.836
LearningEDnCNN [4]0.9570.8870.9370.8750.9010.874
MLPF [24]0.9700.8890.9700.8850.9630.876
EDformer [29]0.9930.9540.9850.9420.9700.926
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.522, + 0.905, + 0.55 + ], + "angle": 0, + "content": "Table 2. The AUC\\(\\uparrow\\) of ROC on the two DND21 sequence at different noise rates. We follow [29] for baseline methods selection." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.702, + 0.907, + 0.73 + ], + "angle": 0, + "content": "Figure 3. ROC curves on the DND 21 dataset sequences. Dash lines represent learning-based methods (see also Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.905, + 0.809 + ], + "angle": 0, + "content": "each sequence provides events, frames, calibration information, IMU data, and ground truth (GT) camera poses (at \\(200\\mathrm{Hz}\\)). We use boxesRotation and dynamicRotation sequences to evaluate rotational motion estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Finally, we also use sequences from MVSEC [63] and zurich_city_12a sequence from the DSEC dataset [20], which are typically used for optical flow estimation. We use MVSEC sequences for quantitative evaluation and the DSEC sequence for qualitative evaluation, since it is captured during the night and hence includes lots of noise." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6962" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.912, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.297, + 0.807, + 0.312 + ], + "angle": 0, + "content": "Figure 4. Comparison of denoising methods on the two DND21 sequences at two different noise levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.485, + 0.606 + ], + "angle": 0, + "content": "Metrics. The metrics used for denoising assessment are: receiver operating characteristic (ROC) curve, area under the curve (AUC) and Mean Event Structural Ratio (MESR). Given signal and noise GT labels, the ROC shows the True-Positive Rate and False-Positive Rate over several S/N ratios. The AUC of the ROC becomes closer to 1, the better the denoiser works. The MESR is a denoising metric that does not require GT labels [10]. We follow previous work to fix the number of signal events to 30000 [10, 29], as well as \\( M = 20000 \\) [10, Eq.10], to discount the difference in the number of signal events from various denoising methods. The higher, the better. Furthermore, we use the RMS error for angular velocity estimation and the EPE for optical flow estimation, following previous works [23, 25, 42, 50], as well as the FWL metric [52], which is the relative variance (sharpness) of the IWE with respect to that of the identity warp. Further details about the experimental settings (e.g., hyper-parameters) are in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.265, + 0.635 + ], + "angle": 0, + "content": "4.2. Denoising Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.687 + ], + "angle": 0, + "content": "We evaluate denoising accuracy using AUC on DND21 data, and using MESR on E-MLB and DND21 data, using the optical flow estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Denoising Performance on E-MLB. Table 1 shows the benchmark results on the E-MLB dataset using the MESR metric. We categorize prior work into model-based methods and learning-based methods for convenience. Compared with other model-based techniques, our method ranks first or second, showcasing the efficacy of the proposed method on denoising. Even compared with learning-based ones, our method achieves competitive scores (e.g., second in E-MLB (Day) ND4-ND64 conditions, etc). Notice that, however, learning-based approaches utilize additional information (GT labels) during training, while our method does not. Nonetheless, the results demonstrate the efficacy of our approach, which contributes to the theoretical foundation of event-denoising methodologies. We discuss further details" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.338, + 0.9, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.726, + 0.907, + 0.782 + ], + "angle": 0, + "content": "Figure 5. Simultaneous estimation of rotational motion and denoising on the dynamicRotation sequence from ECD dataset [40]. Our method produces sharp IWEs \\((2^{\\mathrm{nd}}\\) row) as well as reasonable S/N classification (rows 4 and 5), keeping \\(90\\%\\) of data \\((\\tau = 0.9)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.806, + 0.791, + 0.821 + ], + "angle": 0, + "content": "of the evaluation metrics (ESR) in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Qualitative results are shown in Fig. 2. The denoised outputs (\"Signal\") are reasonable, as dark (underexposed) regions of the scene become cleaner, and the IWEs using the estimated flow provide sharp edges of the original scenes. We compare with state-of-the-art methods in detail next." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.197, + 0.48, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.389 + ], + "angle": 0, + "content": "Figure 6. Results on the ego-motion estimation application (top: boxes, bottom: dynamic) [40]. Our method relaxes the dependency on the initial value for CMax-based rotational motion estimation (indicated as \"Gain\")." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.395, + 0.462, + 0.538 + ], + "angle": 0, + "content": "
dynamic_rotboxes_rot
RMS ↓FWL ↑RMS ↓FWL ↑
5 Hz NoiseCMax [17]20.0011.259124.6411.129
- w/ Init.8.2751.27420.6591.214
Downsampling (Best)8.8081.273124.8691.108
- w/ Init.8.2261.27420.6191.214
Ours (Best)8.5221.27397.5851.141
- w/ Init.8.1701.27420.6041.214
BAF19.6751.260125.0281.127
- w/ Init.8.2531.27420.5501.214
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.548, + 0.46, + 0.563 + ], + "angle": 0, + "content": "Table 3. Angular velocity estimation on ECD dataset [40]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Denoising Accuracy on DND21. Table 2 shows the result of AUC for the ROC result on the DND21 dataset. Our method achieves consistently high AUC compared with existing model-based methods. Note that the DND21 dataset aggressively removes events to generate \"pure-signal\" sequences [24, Sec.5]. Due to this step, the labeled signal events may be fewer than the actual ones corresponding to the true motion. Hence, the DND21 sequences could result in a slight degradation in the False Positive score for motion-based (unsupervised) denoising methods like ours, which may result in lower AUC values. Nonetheless, our method provides competitive AUC scores among other state-of-the-art. The ROC curves are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative comparisons among several denoising methods is given in Fig. 4. From the results, our method achieves a good trade-off between under-denoising (preserving not only scene-edge details but some noise) and over-denoising (removing not only most of the noise but some edge details). This can be confirmed, for example, in the upper-right edges in the driving example, and the person's arm in the hotel example. Essentially, all denoising methods have" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "some parameters to control the amount of denoising; our proposed method has the signal ratio, which is intuitive; other optimization methods have some parameters about spatio-temporal kernel size, and learning-based approaches use the probability threshold for the network output. While we fix it for each dataset and condition for fair comparison (we do not tune it for each sequence in the results), this is an essential challenge for denoising algorithms in general, regardless of their underlying character (learning-based or model-based). Hence, we believe it is important for research to discuss denoising effectiveness in combination with downstream tasks, as shown in Secs. 4.3 and 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.799, + 0.302 + ], + "angle": 0, + "content": "4.3. Joint Motion Estimation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Ego-Motion on ECD. One of the merits of the proposed joint estimation method is that it can improve motion estimation results. Here, the results of ego-motion (e.g., angular velocity) estimation are displayed in Fig. 5. We used the classical ECD dataset and injected BA noise from the DND21 recordings with different noise rates. As no prior work simultaneously estimates motion and noise, we compare against sequential filters: random sampling (of the input data with a uniform distribution) and BAF [9] that is the most widely-used denoising method to date. As shown in the IWEs, our method produces the sharpest edges (cf. (b) and (c)) and the most reasonable signal-noise classification among others (cf. (a), (b), and (c)). Note that the original ECD recordings have noise events besides the ones that are injected using DND21 recordings, which makes the calculation of ROC (similar to Sec. 3.2) intricate. Nonetheless, we report ROC results in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.779 + ], + "angle": 0, + "content": "The quantitative evaluation on angular velocity estimation is reported in Fig. 6, as well as the summary of the supplementary detailed table in Tab. 3. As shown in the first row (\"CMax\"), the original CMax degrades due to the noise occurrences. Our method achieves the best results in terms of RMS and FWL, compared with the random sampling method as well as the original CMax (no denoising). Also, using better initialization (the result of the previous slice for sequential estimation) always improves the results (\"w/ Init.\"): the estimation accuracy relies on both the initialization and the denoising strategies. Hence, the result demonstrates that the proposed method relaxes the dependency on the initialization, thus improving the robustness of motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Optical Flow on MVSEC. The results on more complex motion estimation (e.g., optical flow) are shown in Tab. 4, using the MVSEC dataset [63], with BA noise from the DND21 dataset. Due to the \\(5\\mathrm{Hz}\\) BA noise, the baseline method MultiCM [48] degrades. Our denoising method improves the flow accuracy, showcasing the efficacy of simultaneous noise estimation. A challenge of evaluating denoising in these real-world datasets (e.g., ECD, MVSEC) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.146, + 0.087, + 0.263, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.158, + 0.262, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.089, + 0.379, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.158, + 0.379, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.089, + 0.495, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.158, + 0.495, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.089, + 0.612, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.158, + 0.612, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.089, + 0.728, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.158, + 0.728, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.089, + 0.845, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.158, + 0.845, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.228, + 0.263, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.146, + 0.296, + 0.262, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.367, + 0.246, + 0.38 + ], + "angle": 0, + "content": "(a) Input events" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.228, + 0.379, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.296, + 0.38, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.367, + 0.361, + 0.379 + ], + "angle": 0, + "content": "(b) Estim. flow" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.228, + 0.496, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.296, + 0.495, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.367, + 0.482, + 0.38 + ], + "angle": 0, + "content": "(c) Signal events" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.228, + 0.612, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.296, + 0.612, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.367, + 0.594, + 0.392 + ], + "angle": 0, + "content": "(d) Reconstr. from (c) signal" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.228, + 0.728, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.296, + 0.728, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.367, + 0.708, + 0.392 + ], + "angle": 0, + "content": "(e) Reconstr. from (a) input" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.228, + 0.845, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.296, + 0.845, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.367, + 0.84, + 0.378 + ], + "angle": 0, + "content": "(f) Reference frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.398, + 0.907, + 0.427 + ], + "angle": 0, + "content": "Figure 7. Image intensity reconstruction on E-MLB data [10]. The first two rows use the reconstruction method EVILIP [59] and the last two rows use E2VID [46]. Reconstruction on denoised events produces higher quality images (i.e., fewer artifacts) than on raw event data." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.438, + 0.422, + 0.563 + ], + "angle": 0, + "content": "
indoor1indoor2
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.67632.0982.74633.474
Ours2.51729.4752.69533.047
indoor3outdoor1
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.73233.2191.96018.160
Ours2.65432.2411.86517.554
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Table 4. Optical flow estimation results on MVSEC [63] (\\( dt = 4 \\) frames, i.e., \\( 89 \\mathrm{~ms} \\)) with BA noise at \\( 5 \\mathrm{~Hz} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.651 + ], + "angle": 0, + "content": "existing noise in the original sequences. Hence, we only evaluate motion estimation metrics, not ROC or AUC." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.664, + 0.431, + 0.681 + ], + "angle": 0, + "content": "4.4. Application to Intensity Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.854 + ], + "angle": 0, + "content": "In real-world use cases, denoising works as preprocessing steps for other tasks (e.g., deblur, 3D reconstruction, motion estimation, etc.), hence it is important for denoising methods to demonstrate applicability. Although the proposed approach intrinsically improves motion estimation (Sec. 4.3), we further evaluate the denoising outcome via intensity reconstruction, a well-known task in event-based vision, using the E-MLB dataset [10]. Figure 7 collects the results. We use two state-of-the-art approaches for reconstruction: E2VID [46], which does not rely on motion, and EVILIP [59], which utilizes optical flow as prior." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method provides flow (Fig. 7, column (b)) and signal events (sharp IWEs in column (c)). Comparing the reconstruction with our event denoising method (column (d)) and" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.447, + 0.773, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.518, + 0.773, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.588, + 0.773, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.66, + 0.621, + 0.671 + ], + "angle": 0, + "content": "(a) Signal" + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.66, + 0.739, + 0.67 + ], + "angle": 0, + "content": "(b) Noise" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.449, + 0.895, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.519, + 0.895, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.59, + 0.895, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.66, + 0.879, + 0.67 + ], + "angle": 0, + "content": "(c) Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.679, + 0.907, + 0.708 + ], + "angle": 0, + "content": "Figure 8. Sensitivity analysis: effect of different target ratio parameters \\(\\tau\\) on image reconstruction using EVILIP [59]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "the reconstruction from the raw event data (column (e)), the images after event denoising have fewer artifacts than those without denoising, for both reconstruction methods tested. We observe that E2VID is more robust against noise than EVILIP, possibly due to noise events in its training (simulation) data. Due to the ND filters, the quality of the reference frames (last column) is limited, and hence, we discuss the results qualitatively. Nevertheless comparing the reconstructed images to the reference frames, evidences the HDR advantages of event cameras over frame-based ones." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.223, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.16, + 0.199, + 0.17 + ], + "angle": 0, + "content": "(a) Ref. events" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.089, + 0.346, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.161, + 0.318, + 0.17 + ], + "angle": 0, + "content": "(b) \\(k = 1\\) px" + }, + { + "type": "image", + "bbox": [ + 0.348, + 0.089, + 0.47, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.161, + 0.44, + 0.17 + ], + "angle": 0, + "content": "(c) \\(k = 3\\) px" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Figure 9. Results of different Gaussian kernel sizes \\( k \\) for sensitivity on different depth scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.222, + 0.208, + 0.239 + ], + "angle": 0, + "content": "4.5. Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Effect of Noise Ratio. As Sec. 4.2 already showed the sensitivity for denoising accuracy, let us further discuss the sensitivity using image reconstruction. Figure 8 presents the result of the Bicycle sequence with ND64 filter [10]. While the GT image is highly underexposed and not informative (as those on the last column of Fig. 7), the reconstructed images clearly reflect the choice of the signal target ratio \\(\\tau\\). As more events are removed, \\((i)\\) the reconstructed image becomes more homogeneous but preserves strong edges, and \\((ii)\\) the removed events (2nd col.) contain more scene edges (i.e., signal). We quantitatively analyze the former in the supplementary using non-reference image quality metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Effect of the Kernel Size. Figure 9 shows the sensitivity analysis for different kernel sizes. By increasing the size of the Gaussian kernel in the IWE, our method can emphasize contour edges over isolated points, thus better preserving signal events in regions with low IWE intensities (center of the image, corresponding to far away scene points)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.536, + 0.356, + 0.55 + ], + "angle": 0, + "content": "4.6. Ablation of Motion Estimator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.756 + ], + "angle": 0, + "content": "The proposed approach can be combined with other estimation methods, such as deep-neural-network (DNN) flow estimators. To this end, we show an example of E-RAFT [21]. The model-based method alone [48] is known to output a degraded flow in night sequences with lots of noise, such as zurich_city_12a [20]. As shown in Fig. 10, our method can identify and discard noisy events. E-RAFT provides sharper and more natural edges in IWEs after denoising compared to MultiCM [48] (thanks to its more accurate flow). ESR scores are similar: 2.39 for E-RAFT and 2.45 for MultiCM. While we focus on the inference of DNN in the proposed pipeline, future research could look into accommodating training within the proposed joint estimation framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.195, + 0.786 + ], + "angle": 0, + "content": "4.7. Runtime" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Using 30k events from the ECD dataset \\((240 \\times 180\\mathrm{px})\\) on a Mac M1 CPU (2020), the denoising operations (scoring and ranking) take less than \\(0.1\\mathrm{s}\\) extra computation in total, out of \\(2\\mathrm{s}\\) of the entire processing time (i.e., \\(5\\%\\) computation increase with respect to the base method). Nonetheless, accommodating learning-based approaches in our framework could be key to the trade-off between accuracy and speed." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.098, + 0.657, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.17, + 0.624, + 0.179 + ], + "angle": 0, + "content": "Raw Event" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.099, + 0.776, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.17, + 0.755, + 0.179 + ], + "angle": 0, + "content": "IWE (No Warp)" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.18, + 0.658, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.248, + 0.658, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.323, + 0.638, + 0.333 + ], + "angle": 0, + "content": "(a) Signal Events" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.181, + 0.776, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.251, + 0.776, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.322, + 0.736, + 0.332 + ], + "angle": 0, + "content": "(b) IWE" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.181, + 0.895, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.252, + 0.895, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.323, + 0.882, + 0.332 + ], + "angle": 0, + "content": "(c) Flow Estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.341, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 10. Effect of adopting two different motion estimators (E-RAFT or MultiCM) on the DSEC night scene (zurich_12a) [20]. Despite small IWE differences (e.g., in the car's hood), the proposed framework supports several motion estimators." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.636, + 0.426 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Our method expects as parameter a target noise level, but the intrinsic (\"true\") noise level depends on the camera, scene, etc. and is therefore unknown. More research in this direction is needed for automatic parameter tuning. Also, notice that there might be a potential limitation of the denoising metric ESR, since it depends on the motion parameter (see Suppl. Mat. of [29]). Although our denoising efficacy is confirmed from multiple perspectives, we believe that advancing denoising metrics for real-world data (where it gets interesting due to the absence of GT labels) is a central direction for future work in event-based vision." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.634, + 0.634 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In summary, this work proposes that: (i) Joint estimation of noise and motion is a novel problem setting compared to conventional (isolated) denoising, and is sensitive given the event data's inherent motion dependency. (ii) The proposed joint estimation framework can accommodate various motion estimators and has shown its efficacy in improving both denoising and motion estimation accuracies. (iii) It is important to evaluate denoising results on both, denoising benchmarks and downstream tasks (e.g., motion estimation, intensity reconstruction), since current denoising benchmarks have shortcomings in terms of metrics, data, and baselines. In particular, our method achieves notable denoising results (state-of-the-art on the E-MLB benchmark and competitive on DND21) while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances both event-denoising theory and practical denoising use-cases." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6966" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.192 + ], + "angle": 0, + "content": "We would like to thank Dr. B. Jiang for useful discussions. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.482, + 0.284 + ], + "angle": 0, + "content": "[1] Saeed Afshar, Andrew Peter Nicholson, Andre van Schaik, and Gregory Cohen. Event-based object detection and tracking for space situational awareness. IEEE Sensors Journal, 20(24):15117-15132, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Trans. Neural Netw. Learn. Syst., 35(3):4110-4124, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.355, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[3] Sami Arja, Alexandre Marcireau, Richard L. Balthazor, Matthew G. McHarg, Saeed Afshar, and Gregory Cohen. Density invariant contrast maximization for neuromorphic earth observations. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 3984-3994, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.425, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[4] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1701-1710, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[5] R Wes Baldwin, Mohammed Almatrafi, Jason R Kaufman, Vijayan Asari, and Keigo Hirakawa. Inceptive event timesurfaces for object classification using neuromorphic cameras. In Image Analysis and Recognition: 16th Int. Conf., ICIAR 2019, Waterloo, ON, Canada, August 27-29, 2019, Proceedings, Part II 16, pages 395-403. Springer, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[6] T. D. Barfoot. State Estimation for Robotics - A Matrix Lie Group Approach. Cambridge University Press, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[7] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240x180 130dB \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE J. Solid-State Circuits, 49(10):2333-2341, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[8] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In AAAI Conf. Artificial Intell., pages 303-311, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[9] Tobi Delbruck. Frame-free dynamic digital vision. In Proc. Int. Symp. Secure-Life Electron., pages 21-26, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[10] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-MLB: Multilevel benchmark for event-based camera denoising. IEEE Trans. Multimedia, 26:65–76, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[11] Peiqi Duan, Zihao Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Peiqi Duan, Zihao W. Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "neuromorphic events. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12824-12833, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.192 + ], + "angle": 0, + "content": "[13] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[14] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.906, + 0.293 + ], + "angle": 0, + "content": "[15] Guillermo Gallego and Davide Scaramuzza. Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett., 2(2):632-639, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.906, + 0.338 + ], + "angle": 0, + "content": "[16] Guillermo Gallego and Anthony Yezzi. A compact formula for the derivative of a 3-D rotation in exponential coordinates. J. Math. Imaging Vis., 51(3):378-384, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.906, + 0.409 + ], + "angle": 0, + "content": "[17] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3867-3876, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[18] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vision. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12272-12281, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[19] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154-180, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.906, + 0.595 + ], + "angle": 0, + "content": "[20] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett., 6(3):4947-4954, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.906, + 0.653 + ], + "angle": 0, + "content": "[21] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vision (3DV), pages 197–206, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.906, + 0.698 + ], + "angle": 0, + "content": "[22] Rui Graca and Tobi Delbruck. Towards a physically realistic computationally efficient dvs pixel model. In Int. Image Sensor Workshop (IISW), 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.768 + ], + "angle": 0, + "content": "[23] Cheng Gu, Erik Learned-Miller, Daniel Sheldon, Guillermo Gallego, and Pia Bideau. The spatio-temporal Poisson point process: A simple model for the alignment of event camera data. In Int. Conf. Comput. Vis. (ICCV), pages 13495-13504, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[24] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Trans. Pattern Anal. Mach. Intell., 45(1):785-795, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[25] Shuang Guo and Guillermo Gallego. CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot., 40: 2442-2461, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[26] Shasha Guo, Ziyang Kang, Lei Wang, Shiming Li, and Weixia Xu. Hashheat: An O(C) complexity hashing-based" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6967" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "filter for dynamic vision sensor. In 25th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 452-457, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[27] Shuang Guo, Friedhelm Hamann, and Guillermo Gallego. Unsupervised joint learning of optical flow and intensity with event cameras. In Int. Conf. Comput. Vis. (ICCV), 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[28] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In Eur. Conf. Comput. Vis. (ECCV), pages 18-37, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[29] Bin Jiang, bo Xiong, Bohan Qu, M. Slaman Asif, you Zhou, and Zhan Ma. EDformer: Transformer-based event denoising across varied noise levels. In Eur. Conf. Comput. Vis. (ECCV), pages 1-12, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[30] Pritam P. Karmokar, Quan H. Nguyen, and William J. Beksi. Secrets of edge-informed contrast maximization for event-based vision. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[31] Alireza Khodamoradi and Ryan Kastner. O(N)-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Trans. Emerg. Topics Comput., PP(99):1-1, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[32] Haram Kim and H. Jin Kim. Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett., 6(3):6016-6023, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[33] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), pages 8601-8608, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[34] Xavier Lagorce, Garrick Orchard, Francesco Gallupi, Bertram E. Shi, and Ryad Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(7):1346-1359, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[35] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128x128 120dB 30mW asynchronous vision sensor that responds to relative intensity change. In IEEE Int. Solid-State Circuits Conf. (ISSCC), pages 2060-2069, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[36] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) dB 15 μs latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits, 43(2):566-576, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[37] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In IEEE Int. Symp. Circuits Syst. (ISCAS), pages 722-725, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[38] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[39] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research, 36(2):142-149, 2017." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[41] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot., 34(6):1425-1440, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[42] Urbano Miguel Nunes and Yiannis Demiris. Robust event-based vision model estimation by dispersion minimisation. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9561-9573, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3445–3454, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.274, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[44] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE de Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Int. Conf. Comput. Vis. (ICCV), pages 9661-9671, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.343, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[45] Xin Peng, Ling Gao, Yifu Wang, and Laurent Kneip. Globally-optimal contrast maximisation for event cameras. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3479-3495, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.399, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Trans. Pattern Anal. Mach. Intell., 43 (6):1964-1980, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.455, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[47] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high speed scenarios. IEEE Robot. Autom. Lett., 3(2): 994-1001, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[48] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis. (ECCV), pages 628-645, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.567, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[49] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):1-20, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[50] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):7742-7759, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[51] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Int. Conf. Comput. Vis. (ICCV), pages 7243-7252, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[52] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In Eur. Conf. Comput. Vis. (ECCV), pages 534-549, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[53] Gemma Taverni, Diederik Paul Moeys, Chenghan Li, Celso Cavaco, Vasyl Motsnyi, David San Segundo Bello, and Tobi Delbruck. Front and back illuminated Dynamic and Active Pixel Vision Sensors comparison. IEEE Trans. Circuits Syst. II (TCSII), 65(5):677-681, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6358-6367, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[55] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1606–1616, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[56] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Trans. Multimedia, 23:1148-1159, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[57] Ryo Yamaki, Shintaro Shiba, Gallego Guillermo, and Yoshimitsu Aoki. Iterative event-based motion segmentation by variational contrast maximization. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 4918-4927, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[58] Pei Zhang, Haosen Liu, Zhou Ge, Chutian Wang, and Edmund Y Lam. Neuromorphic imaging with joint image deblurring and event denoising. IEEE Trans. Image Process., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[59] Zelin Zhang, Anthony Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8372-8389, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[60] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Trans. Neural Netw. Learn. Syst., pages 1-13, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[61] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Trans. Robot., 37(5):1433-1450, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[62] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5816-5824, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[63] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. Autom. Lett., 3(3):2032-2039, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[64] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 989-997, 2019." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6969" + } + ] +] \ No newline at end of file diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_origin.pdf b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6864263f022e6020f1487914a8ff42934a4eb221 --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/6f9dcd49-a085-404d-b287-141b9737e694_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a62e80aecb3edfeb3441ced79dc00a90deddcf09a985bdefd7b340a4a6db279d +size 4912315 diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/full.md b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..948dcede117b3d2a222373ab3b43bd9075488841 --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/full.md @@ -0,0 +1,389 @@ +# Simultaneous Motion And Noise Estimation with Event Cameras + +Shintaro Shiba $^{1,2,3}$ , Yoshimitsu Aoki $^{1}$ and Guillermo Gallego $^{3,4}$ + +1 Keio University, 2 Woven by Toyota, Inc., Japan. 3 Technische Universität Berlin, 4 Einstein Center Digital Future, Robotics Institute Germany, and Science of Intelligence Excellence Cluster, Germany. + +# Abstract + +Event cameras are emerging vision sensors whose noise is challenging to characterize. Existing denoising methods for event cameras are often designed in isolation and thus consider other tasks, such as motion estimation, separately (i.e., sequentially after denoising). However, motion is an intrinsic part of event data, since scene edges cannot be sensed without motion. We propose, to the best of our knowledge, the first method that simultaneously estimates motion in its various forms (e.g., ego-motion, optical flow) and noise. The method is flexible, as it allows replacing the one-step motion estimation of the widely-used Contrast Maximization framework with any other motion estimator, such as deep neural networks. The experiments show that the proposed method achieves state-of-the-art results on the E-MLB denoising benchmark and competitive results on the DND21 benchmark, while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances event-data denoising theory and expands practical denoising use-cases via open-source code. Project page: https://github.com/tub-rip/ESMD + +# 1. Introduction + +Event cameras are emerging vision sensors that overcome shortcomings of conventional cameras (e.g., motion blur, limited dynamic range, data and power efficiency), yet they suffer from a considerable amount of noise due to their novelty and operation in low-power (transistor subthreshold) conditions [19, 36]. Since event cameras are suitable for many computer vision tasks, especially motion-related tasks, it is paramount to classify event data that is related to motion (i.e., signal) and that is not (i.e., noise). However, denoising event data is a challenging problem because noise properties have yet to be fully characterized [22, 24] and it is infeasible to define ground-truth (GT) noise labels in real-world data recordings. Previous works either leveraged simulation or prepared real event data and aggressively suppressed noise occurrence. + +However, such artificial modifications on the real event + +![](images/60f0d0e4b9b4750bd32f26bb9e4daa66eef54a87131f6d80f3004d9d4875aded.jpg) +Figure 1. Overview of the proposed method. We use only raw events as input, and estimate signal events, noise events, and motion (e.g., dense optical flow and ego-motion estimation). + +data could change event signal/noise characteristics, possibly deteriorating the information that the original event had for downstream tasks (e.g., motion estimation). Our work is inspired by the observation that motion is an intrinsic part of event data since scene edges cannot be sensed without motion. Rather than solving denoising and motion estimation tasks separately, our idea is to leverage motion information to improve denoising (as "vice-versa" is well known). In this work (Fig. 1), we rethink and investigate the connection between denoising and motion estimation, by asking ourselves whether they can be combined in a more integrated way to benefit the results of both tasks simultaneously. We formulate the simultaneous estimation problem by leveraging the idea of motion compensation in the Contrast Maximization (CMax) framework, defining novel metrics about the contribution to the contrast score for each event. In summary, we present several distinctive contributions: + +1. We propose a novel method to jointly estimate noise and motion only from event data (Sec. 3). +2. We demonstrate that our joint estimation method $(i)$ achieves state-of-the-art results in existing denoising benchmarks (Sec. 4.2), and $(ii)$ improves the robustness of motion estimation (CMax framework) (Sec. 4.3). + +3. We show the flexibility of our method on $(i)$ further applications, such as intensity reconstruction, validating its image quality (Secs. 4.4 and 4.5), and $(ii)$ in combination with learning-based motion estimators (Sec. 4.6). + +We hope our work opens a new line of research for event cameras, by considering the problems of motion estimation and denoising together to take advantage of their coaction. + +# 2. Related Work + +Denoising is a fundamental problem in event cameras since there are a lot of leakage events [24], and event-processing algorithms may suffer in scenarios with high noise [3]. Removing event noise has a double effect: it reduces the load of subsequent processing stages (as the load is proportional to the number of events) and improves their output (as the input data becomes of higher quality). In this work, we focus on noise due to the leakage inside the camera hardware, such as background activity (BA) events [36]. Events due to flickering or active lights are therefore not our priority. + +Classical denoising works focus on the spatio-temporal correlation of events, as BA events have fewer neighboring events in space-time coordinates than signal events. This approach includes widely-used BA Filter (BAF) [9], and other types of spatio-temporal filters [26, 31, 37, 56]. Some works utilize additional sensor inputs (e.g., frames) to estimate noise probabilities [4, 12, 30, 55, 58], since edges are informative of the signal events produced. + +Recently, learning-based approaches have gained more attention, such as multilayer perceptrons [24], convolutional neural networks [1, 4, 12], graph neural networks [2], and transformers [29]. They provide the probability of noise (i.e., signal-noise classification). Nonetheless, the challenge of learning-based approaches is the need for ground truth (GT) noise labels for training. GT labels have been obtained through simulation (e.g., [13, 24]) as well as real-world recordings, however, real-world recordings aggressively remove noisy events using existing filters (e.g., [24]) or use unnatural illumination settings (e.g., [29]), by leveraging the differences of noise occurrence due to different scene brightness [24, 35]. While these approaches are effective for benchmarking, the challenge of obtaining GT labels for practical real-world datasets remains unsolved. + +Our work builds upon a series of successful methods for motion estimation and their applications. Event-based motion estimation is an important topic since event cameras naturally respond to motion in the scenes [19]. In particular, Contrast Maximization (CMax) [17, 50] is a state-of-the-art framework for motion estimation, which has various applications, such as rotational motion estimation [15, 23, 32, 42, 45], optical flow estimation [27, 28, 44, 48, 64], and motion segmentation [51, 57, 60]. We investigate the connection and potential benefits of applying the CMax framework to denoising during the estimation process (i.e., simul + +taneously), demonstrating motion estimation improvements as well as competitive denoising accuracy. + +A prior work, "ProgressiveMotionSeg" [8], has extended the CMax-based segmentation method [51] with extra weights that accounted for how much each event contributed to the estimated motion clusters. However, [8] mainly focuses on the task of motion segmentation using low degrees-of-freedom (DOF) warps (e.g., 2-DOF feature flow) and does not discuss denoising efficacy or evaluate on standard denoising benchmarks. In contrast, our method supports various types of motion estimation (both low-DOF and high-DOF warps, such as optical flow), and considers that noise shall not be used to estimate motion, as from first principles we deem noise and motion to be uncorrelated. + +# 3. Methodology + +Event cameras consist of independent pixels that generate asynchronous "events" when the logarithmic brightness at the pixel increases or decreases by a predefined contrast sensitivity. Each event $e_k \doteq (\mathbf{x}_k, t_k, p_k)$ contains the pixel coordinates $\mathbf{x}_k$ , the timestamp $t_k$ , and polarity $p_k \in \{+1, -1\}$ of the brightness change. Events occur asynchronously and sparsely on the pixel grid, resulting in a variable data rate based on the scene texture and dynamics. + +# 3.1. Contrast Maximization + +The CMax framework [17] assumes events $\mathcal{E} \doteq \{e_k\}_{k=1}^{N_e}$ are caused by moving edges, and it transforms their coordinates according to a motion model $\mathbf{W}$ , producing a set of warped events $\mathcal{E}'_{t_{\mathrm{ref}}} \doteq \{e_k'\}_{k=1}^{N_e}$ at a reference time $t_{\mathrm{ref}}$ : + +$$ +e _ {k} \dot {=} \left(\mathbf {x} _ {k}, t _ {k}, p _ {k}\right) \mapsto e _ {k} ^ {\prime} \dot {=} \left(\mathbf {x} _ {k} ^ {\prime}, t _ {\text {r e f}}, p _ {k}\right). \tag {1} +$$ + +The warp $\mathbf{x}_k^{\prime} = \mathbf{W}(\mathbf{x}_k, t_k; \boldsymbol{\theta})$ transports each event from $t_k$ to $t_{\mathrm{ref}}$ along the motion curve that passes through it. Then, they are aggregated on an image of warped events (IWE): + +$$ +I (\mathbf {x}; \mathcal {E} _ {t _ {\text {r e f}}} ^ {\prime}, \boldsymbol {\theta}) \doteq \sum_ {k = 1} ^ {N _ {e}} \delta \left(\mathbf {x} - \mathbf {x} _ {k} ^ {\prime}\right), \tag {2} +$$ + +where each pixel $\mathbf{x}$ sums the number of warped events $\mathbf{x}_k^\prime$ that fall within it. The Dirac delta is approximated by a Gaussian, $\delta (\mathbf{x} - \pmb {\mu})\approx \mathcal{N}(\mathbf{x};\pmb {\mu},\epsilon^2\mathrm{Id})$ with $\epsilon = 1$ pixel. + +Next, an objective function, such as the contrast of the IWE (2) (image variance [18] $\mathrm{Var}\big(I(\mathbf{x};\pmb {\theta})\big)\doteq \frac{1}{|\Omega|}\int_{\Omega}(I(\mathbf{x};\pmb {\theta}) - \mu_I)^2 d\mathbf{x})$ , measures the goodness of fit between the events and the candidate motion curves (warp). + +Finally, an optimization algorithm iterates the above steps to find the motion parameters $\theta^{*}$ that maximize the alignment of events caused by the same scene edge. Event alignment is measured by the strength of the edges of the IWE, which is related to image contrast [18]. + +Motion Models. As exemplary warps $\mathbf{W}$ we focus on two popular problem settings: rotational motion (3-DOF angular velocity estimation [23, 25, 42, 45]) and dense optical flow (pixel-wise velocity estimation [43, 48, 50, 64], + +![](images/fe741d6d8a61f060504e939b6f26702e5ebdcd616adc04c751cd1d18fca0f992.jpg) +Figure 2. Qualitative result of denoising and flow estimation on E-MLB and DND21. The driving sequence has $5\mathrm{-Hz}$ artificial noise. (b) is an alternative visualization of (a) with the same grayscale coding as (d). Flow is visualized on all events and flow color is given in Fig. 1. + +resulting in $2N_{p}$ -DOFs, where $N_{p}$ is the number of pixels). For the angular velocity estimation during a small time interval, the warp is parametrized by the angular velocity $\theta \equiv \omega = (\omega_{x},\omega_{y},\omega_{z})^{\top}$ , and $\mathbf{x}^{h}(t)\sim \mathbb{R}(\boldsymbol {\omega}t)\mathbf{x}^{h}(0)$ , using calibrated homogeneous coordinates $\mathbf{x}^{h}$ and exponential coordinates $\mathbb{R}(\phi)\doteq \exp (\phi^{\wedge})$ [6, 16]. The warp for dense optical flow is $\mathbf{x}_k^\prime = \mathbf{x}_k + (t_k - t_{\mathrm{ref}})\mathbf{v}(\mathbf{x}_k)$ , where $\theta = \{\mathbf{v}(\mathbf{x})\}_{\mathbf{x}\in \Omega}$ is a flow field on the image plane $\Omega$ . + +# 3.2. Denoising Based on Local Contrast + +As Fig. 1 shows, our method is iterative. In each iteration, motion estimation is refined and events are classified into signal $\mathcal{E}_s$ or noise $\mathcal{E}_n$ . Without labels or additional information, events rely on each other to realize this classification. To this end, we compute a score $c_k$ for each event, rank the events according to their scores, and threshold at some value $T$ . The threshold is computed by assuming the amount or ratio of noisy events is known $\eta$ , i.e., $T \equiv T(\eta)$ , which allows us to have high-level control over the denoising process. Hence we classify events into signal (the $\tau \doteq (1 - \eta)$ percentage of events with highest scores) and noise: + +$$ +\mathcal {E} _ {s} \equiv \mathcal {E} _ {\text {s i g n a l}} \doteq \left\{e _ {k} \in \mathcal {E} \mid c _ {k} > T (\eta) \right\}, \tag {3} +$$ + +$$ +\mathcal {E} _ {n} \equiv \mathcal {E} _ {\text {n o i s e}} \dot {=} \mathcal {E} \backslash \mathcal {E} _ {\text {s i g n a l}} \quad (\text {c o m p l e m e n t} \mathcal {E} _ {\text {s i g n a l}}). +$$ + +We studied different choices for the per-event score $c_k$ based on the agreement among warped events $\mathcal{E}' = \mathcal{E}_s' \cup \mathcal{E}_n'$ , through their IWEs, and concluded on two, as follows. Letting $I_k^s \equiv I_k^s(\mathbf{x}_k')$ , $I_k^n \equiv I_k^n(\mathbf{x}_k')$ and $I_k \equiv I(\mathbf{x}_k') = I_k^s + I_k^n$ + +be the IWE values of previously classified signal events, noise events and all events, respectively, the two choices are: (i) $^{1}c_{k} \doteq I_{k}$ , and (ii) the ratio of signal events $^{2}c_{k} \doteq I_{k}^{s} / I_{k}$ (probabilistic viewpoint). In fact, both are equivalent: rule $^{1}c_{k} > T_{1}$ is the same as rule $^{2}c_{k} > T_{2}$ if $T_{2} = 1 - I_{k}^{n} / T_{1}$ and the event noise $I_{k}^{n}$ is known. We use $^{1}c_{k}$ , the number of warped events (also called local contrast [18, 51]), because it is the simplest to interpret. + +The intuition is that "signal" events (e.g., due to motion) warped according to the true motion lie all at motion-compensated edges, and therefore produce a sharp IWE. Noise events, produced all over the image plane, are not expected to contribute to the IWE's sharpness. Hence, we may classify events into signal or noise according to how much they contribute to edge strength. The IWE (2) provides one such measure of edge strength: the higher $I(\mathbf{x})$ , the more events are warped to that pixel location $\mathbf{x}$ (i.e., more events support the same scene edge), thus producing a sharper IWE. Hence, we define how much an event $e_k$ contributes to edge strength by means of the IWE (2), $^1 c_k$ . + +Invariance. Note that (3) is invariant to monotonically increasing functions of $c_k$ , as such functions preserve the order. The same $\mathcal{E}_{\mathrm{signal}}$ is obtained if one ranks and thresholds (at the corresponding level) values $\{c_k\}$ , $\{\ln(c_k)\}$ , $\{e^{c_k}\}$ , etc. They are equivalent under the ranking rule (3). + +Robustness to Various Edge Strengths. The Gaussian kernel in the IWE controls the sensitivity to edge strength. Since the proposed method is based on both signal and noise + +IWEs, increasing the size of the Gaussian kernel in the IWE can emphasize contour edges over isolated points. This preserves signal events better in regions with low IWE intensities, for example at pixels with large depth values. + +# 3.3. Alternating Optimization + +Classifying events into signal and noise requires knowledge of the true motion, and estimation of the true motion (e.g., using CMax) requires knowledge of the signal events because, by definition, noise events carry no information about motion. Hence, it is a circular dependency problem. + +We approach its solution iteratively, as shown in Fig. 1, by combining motion estimation (Sec. 3.1) and S/N classification (Sec. 3.2). Assuming some initialization, the current signal events $\mathcal{E}_{\mathrm{signal}}$ are used to estimate the motion via CMax (1 step is enough); then all events are warped to compute the IWE and the per-event scores $c_{k}$ , from which the sets (3) are recomputed, in preparation for the next iteration. In practice, we initialize with a random split of $\mathcal{E}$ into $\mathcal{E}_{\mathrm{signal}}$ and $\mathcal{E}_{\mathrm{noise}}$ , update $\mathcal{E}_{\mathrm{signal}}$ at every warp during the optimization iteration until the process converges (as flagged by the convergence of the motion parameters in CMax). + +Computational Complexity. The computational complexity of one iteration of the proposed method is $O(N_{p} + N_{e}\log N_{e})$ , which is slightly larger than the original CMax (i.e., $O(N_{p} + N_{e})$ ) due to the search for the highest-ranked events during the optimization. + +# 4. Experiments + +We evaluate the proposed method on four datasets using various metrics (in Sec. 4.1). In the following sections using common benchmarks, we assess denoising (Sec. 4.2), joint motion estimation (Sec. 4.3), and the efficacy on intensity reconstruction (Sec. 4.4). We also conduct sensitivity analyses (Sec. 4.5) and an ablation (Sec. 4.6). + +# 4.1. Datasets, Metrics, and Hyper-parameters + +Datasets. E-MLB [10] is a large-scale, de-facto dataset to benchmark denoising. It consists of 100 sequences with 4 different brightness levels using neutral-density (ND) filters on a DAVIS346 camera $(346\times 260\mathrm{px})$ during day and night. + +Historically, the two sequences (hotel and driving) of the DND21 dataset [24] have been widely used for denoising evaluation [10, 24, 29]. Both sequences are recorded with a DAVIS346 camera [53]. They are regarded as "signal" data, as the sequences have been aggressively filtered offline. Combined with other pure BA noise sequences, the dataset has event-wise noise annotations that are useful for evaluation. We use the above sequences with different noise rates, 1, 3, 5, 7, $10\mathrm{Hz}$ per pixel, following prior work [29]. + +The ECD dataset [40] is a standard dataset for various tasks including camera ego-motion estimation [15, 41, 47, 49, 61, 62]. Using a DAVIS240C camera $(240\times 180\mathrm{px}$ [7]), + +
E-MLB (Day)E-MLB (Night)DND21
ND1ND4ND16ND64ND1ND4ND16ND64
Model-basedRaw0.8210.8240.8150.7860.8900.8240.7860.7680.869
BAF [9]0.8610.8690.8760.8900.9460.9730.9920.9420.920
TS [34]0.8770.8870.8700.8371.0330.9440.8860.7970.985
KNoise [31]0.8460.8370.8300.8070.9540.9560.8710.8170.887
EvFlow [54]0.8480.8780.8680.8330.9690.9830.8890.7971.006
IETS [5]0.7720.7850.7770.7530.9500.8230.8040.7110.900
Ynoise [14]0.8660.8630.8570.8211.0090.9430.8750.7920.966
GEF [11]1.0510.9380.9350.9271.0270.9550.9460.9350.932
DWF [24]0.8780.8760.8660.8650.9230.9620.9880.9320.905
Ours0.9380.9580.9860.9501.0370.9610.9450.9320.992
LearningEDnCNN [4]0.8870.9080.9030.9121.0011.0241.0791.0860.977
EventZoom [12]0.9960.9880.9960.9701.0551.0071.0100.9881.059
MLPF [24]0.8510.8550.8460.8400.9260.9280.9100.9060.944
EDformer [29]0.9520.9550.9560.9421.0481.0191.0761.0991.041
+ +Table 1. Mean ESR (MESR↑) results among denoising methods on the event denoising datasets E-MLB [10] and DND21 [24]. In each category, the best is in bold and the second best is underlined. Methods are chronologically sorted within each category. + +
1Hz5Hz10Hz
hoteldrivinghoteldrivinghoteldriving
Model-basedBAF [9]0.9540.8480.8920.7930.8370.748
TS [34]0.9720.9310.9610.9270.9620.920
KNoise [31]0.6770.6300.6700.6240.6410.614
Ynoise [14]0.9690.9410.9230.9090.8990.880
DWF [24]0.9270.7410.8620.6900.7960.656
Ours1.0140.8820.9630.8550.9610.836
LearningEDnCNN [4]0.9570.8870.9370.8750.9010.874
MLPF [24]0.9700.8890.9700.8850.9630.876
EDformer [29]0.9930.9540.9850.9420.9700.926
+ +Table 2. The AUC $\uparrow$ of ROC on the two DND21 sequence at different noise rates. We follow [29] for baseline methods selection. + +![](images/9ef37170daa390a676e23b6f9a7959ad5ea887116686e84b9fcbcda148b57c65.jpg) +Figure 3. ROC curves on the DND 21 dataset sequences. Dash lines represent learning-based methods (see also Tab. 2). + +each sequence provides events, frames, calibration information, IMU data, and ground truth (GT) camera poses (at $200\mathrm{Hz}$ ). We use boxesRotation and dynamicRotation sequences to evaluate rotational motion estimation accuracy. + +Finally, we also use sequences from MVSEC [63] and zurich_city_12a sequence from the DSEC dataset [20], which are typically used for optical flow estimation. We use MVSEC sequences for quantitative evaluation and the DSEC sequence for qualitative evaluation, since it is captured during the night and hence includes lots of noise. + +![](images/b28b00a8fa4d581eb8dab8adb212212295062b5100fe9e4ded79d089fd0d483d.jpg) +Figure 4. Comparison of denoising methods on the two DND21 sequences at two different noise levels. + +Metrics. The metrics used for denoising assessment are: receiver operating characteristic (ROC) curve, area under the curve (AUC) and Mean Event Structural Ratio (MESR). Given signal and noise GT labels, the ROC shows the True-Positive Rate and False-Positive Rate over several S/N ratios. The AUC of the ROC becomes closer to 1, the better the denoiser works. The MESR is a denoising metric that does not require GT labels [10]. We follow previous work to fix the number of signal events to 30000 [10, 29], as well as $M = 20000$ [10, Eq.10], to discount the difference in the number of signal events from various denoising methods. The higher, the better. Furthermore, we use the RMS error for angular velocity estimation and the EPE for optical flow estimation, following previous works [23, 25, 42, 50], as well as the FWL metric [52], which is the relative variance (sharpness) of the IWE with respect to that of the identity warp. Further details about the experimental settings (e.g., hyper-parameters) are in the supplementary. + +# 4.2. Denoising Results + +We evaluate denoising accuracy using AUC on DND21 data, and using MESR on E-MLB and DND21 data, using the optical flow estimation. + +Denoising Performance on E-MLB. Table 1 shows the benchmark results on the E-MLB dataset using the MESR metric. We categorize prior work into model-based methods and learning-based methods for convenience. Compared with other model-based techniques, our method ranks first or second, showcasing the efficacy of the proposed method on denoising. Even compared with learning-based ones, our method achieves competitive scores (e.g., second in E-MLB (Day) ND4-ND64 conditions, etc). Notice that, however, learning-based approaches utilize additional information (GT labels) during training, while our method does not. Nonetheless, the results demonstrate the efficacy of our approach, which contributes to the theoretical foundation of event-denoising methodologies. We discuss further details + +![](images/55bf99e617c6a07d453f0a17812cdd866fd511e704255aafbd8eb929fd94ebf6.jpg) +Figure 5. Simultaneous estimation of rotational motion and denoising on the dynamicRotation sequence from ECD dataset [40]. Our method produces sharp IWEs $(2^{\mathrm{nd}}$ row) as well as reasonable S/N classification (rows 4 and 5), keeping $90\%$ of data $(\tau = 0.9)$ . + +of the evaluation metrics (ESR) in Sec. 5. + +Qualitative results are shown in Fig. 2. The denoised outputs ("Signal") are reasonable, as dark (underexposed) regions of the scene become cleaner, and the IWEs using the estimated flow provide sharp edges of the original scenes. We compare with state-of-the-art methods in detail next. + +![](images/5374923d01492c439392167a7cae19ee330bb9f7501ad14b7f5b00700343d021.jpg) + +![](images/bda697145e928ca91ea383e47c245c88b6e5feb26162fe183be61387addcd6d1.jpg) +Figure 6. Results on the ego-motion estimation application (top: boxes, bottom: dynamic) [40]. Our method relaxes the dependency on the initial value for CMax-based rotational motion estimation (indicated as "Gain"). + +
dynamic_rotboxes_rot
RMS ↓FWL ↑RMS ↓FWL ↑
5 Hz NoiseCMax [17]20.0011.259124.6411.129
- w/ Init.8.2751.27420.6591.214
Downsampling (Best)8.8081.273124.8691.108
- w/ Init.8.2261.27420.6191.214
Ours (Best)8.5221.27397.5851.141
- w/ Init.8.1701.27420.6041.214
BAF19.6751.260125.0281.127
- w/ Init.8.2531.27420.5501.214
+ +Table 3. Angular velocity estimation on ECD dataset [40]. + +Denoising Accuracy on DND21. Table 2 shows the result of AUC for the ROC result on the DND21 dataset. Our method achieves consistently high AUC compared with existing model-based methods. Note that the DND21 dataset aggressively removes events to generate "pure-signal" sequences [24, Sec.5]. Due to this step, the labeled signal events may be fewer than the actual ones corresponding to the true motion. Hence, the DND21 sequences could result in a slight degradation in the False Positive score for motion-based (unsupervised) denoising methods like ours, which may result in lower AUC values. Nonetheless, our method provides competitive AUC scores among other state-of-the-art. The ROC curves are shown in Fig. 3. + +Qualitative comparisons among several denoising methods is given in Fig. 4. From the results, our method achieves a good trade-off between under-denoising (preserving not only scene-edge details but some noise) and over-denoising (removing not only most of the noise but some edge details). This can be confirmed, for example, in the upper-right edges in the driving example, and the person's arm in the hotel example. Essentially, all denoising methods have + +some parameters to control the amount of denoising; our proposed method has the signal ratio, which is intuitive; other optimization methods have some parameters about spatio-temporal kernel size, and learning-based approaches use the probability threshold for the network output. While we fix it for each dataset and condition for fair comparison (we do not tune it for each sequence in the results), this is an essential challenge for denoising algorithms in general, regardless of their underlying character (learning-based or model-based). Hence, we believe it is important for research to discuss denoising effectiveness in combination with downstream tasks, as shown in Secs. 4.3 and 4.4. + +# 4.3. Joint Motion Estimation Results + +Ego-Motion on ECD. One of the merits of the proposed joint estimation method is that it can improve motion estimation results. Here, the results of ego-motion (e.g., angular velocity) estimation are displayed in Fig. 5. We used the classical ECD dataset and injected BA noise from the DND21 recordings with different noise rates. As no prior work simultaneously estimates motion and noise, we compare against sequential filters: random sampling (of the input data with a uniform distribution) and BAF [9] that is the most widely-used denoising method to date. As shown in the IWEs, our method produces the sharpest edges (cf. (b) and (c)) and the most reasonable signal-noise classification among others (cf. (a), (b), and (c)). Note that the original ECD recordings have noise events besides the ones that are injected using DND21 recordings, which makes the calculation of ROC (similar to Sec. 3.2) intricate. Nonetheless, we report ROC results in the supplementary. + +The quantitative evaluation on angular velocity estimation is reported in Fig. 6, as well as the summary of the supplementary detailed table in Tab. 3. As shown in the first row ("CMax"), the original CMax degrades due to the noise occurrences. Our method achieves the best results in terms of RMS and FWL, compared with the random sampling method as well as the original CMax (no denoising). Also, using better initialization (the result of the previous slice for sequential estimation) always improves the results ("w/ Init."): the estimation accuracy relies on both the initialization and the denoising strategies. Hence, the result demonstrates that the proposed method relaxes the dependency on the initialization, thus improving the robustness of motion estimation. + +Optical Flow on MVSEC. The results on more complex motion estimation (e.g., optical flow) are shown in Tab. 4, using the MVSEC dataset [63], with BA noise from the DND21 dataset. Due to the $5\mathrm{Hz}$ BA noise, the baseline method MultiCM [48] degrades. Our denoising method improves the flow accuracy, showcasing the efficacy of simultaneous noise estimation. A challenge of evaluating denoising in these real-world datasets (e.g., ECD, MVSEC) is the + +![](images/c3ddb1665febad69ad99d9fc03cf3a5caae3bd7ca7061ef9354807a4dc85e551.jpg) + +![](images/4eb34d949be54c3126c44fb3fa44e6ad6c0c9063c5f338a0d95b341f65e5735b.jpg) + +![](images/896daaa895844ecc161cd48140cee724820ebb36597a41a26cd15b2c0477a431.jpg) + +![](images/b46025279b1f2cdf8dd2c8e62598906c34f547e2139fa6f6a5c04f2e32a155f4.jpg) + +![](images/ddb5991c30f64a0eb898abf68d72f10da2a6042ac5a0f4153f94a82b0369a999.jpg) + +![](images/33f90d53462ceac23f2e016430c7ff3cdbb66e2fcd9afb497c3d947a2b3a3d8f.jpg) + +![](images/e3cc8de1acf621db3440d613e46a69aaccc28e8ae735bd87d415f36719468c8c.jpg) + +![](images/fffe399b3e99b647238adf5e4a0ad4c337c66921c44a793c5163144a31b49d57.jpg) + +![](images/75d5e2ca95ccb2ab0981ef0a32fa07a7d8ace6b28a2d3ec52ecd4fd3f5ce7cea.jpg) + +![](images/00b7f39e62ba0bf310a00dfa3bd7ebd01e5679f5dbcabd611dded2c5c041f6e3.jpg) + +![](images/d26dcecf03b28eef0bc8b251694df919ea84d99e6aba98c37901a83773a33939.jpg) + +![](images/7885488cd32a03083a66e1cf95ba1f716d358933996b86e9f408580358a6ea25.jpg) + +![](images/f771427a458bdd8604627c5fd8b1be9dd97cc27df2cfd4a22a91226f998f7604.jpg) + +![](images/80c6d0b20a4fef589ed86c26a8418ca09c638715bdb74daa1f90679db93e9aeb.jpg) +(a) Input events + +![](images/e0ab63322f5ea2373496e5ae024f33d8b9df7338b0099bf0d020e9da1a61f1d3.jpg) + +![](images/0b519167a76ca773bf64bc36bfde8c06b952a4df25a3c5c20b785ed0e99c02d9.jpg) +(b) Estim. flow + +![](images/611881d6462c94893c3e2b305d25417b5e6648cb8e347cec23773c9b44cab91a.jpg) + +![](images/31a3b6c530eb861e92ed02ea78450de367460d20d0df378eeb3b11bc8a727acb.jpg) +(c) Signal events + +![](images/5459b703e831b4ac8d0a30e1024fd5c641d7bcad17c61c87ece00e3338ab5ba3.jpg) + +![](images/c6d44af299743e42c52aae9aca358a335deb0ba6b93ac90fe73178bc4e9b6626.jpg) +(d) Reconstr. from (c) signal + +![](images/1dff928675a01ad4107b3541d35445e31584b0eb2c086492dfd98924a6849a35.jpg) + +![](images/b8b10ecc5da0cd0d858c27c1dbabd4bb3ce6f2189078c3ba480d41719134f08c.jpg) +(e) Reconstr. from (a) input + +![](images/7840d8e7465f76d67efbd4d4a9522c4dec2a557eda0bf09bc091ea9f5ffc36e0.jpg) + +![](images/fb877060ab676d280305608a8d7686e52c3100edb8370d9ff8c7baf50b5b7052.jpg) +(f) Reference frame + +
indoor1indoor2
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.67632.0982.74633.474
Ours2.51729.4752.69533.047
indoor3outdoor1
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.73233.2191.96018.160
Ours2.65432.2411.86517.554
+ +Table 4. Optical flow estimation results on MVSEC [63] ( $dt = 4$ frames, i.e., $89 \mathrm{~ms}$ ) with BA noise at $5 \mathrm{~Hz}$ . + +existing noise in the original sequences. Hence, we only evaluate motion estimation metrics, not ROC or AUC. + +# 4.4. Application to Intensity Reconstruction + +In real-world use cases, denoising works as preprocessing steps for other tasks (e.g., deblur, 3D reconstruction, motion estimation, etc.), hence it is important for denoising methods to demonstrate applicability. Although the proposed approach intrinsically improves motion estimation (Sec. 4.3), we further evaluate the denoising outcome via intensity reconstruction, a well-known task in event-based vision, using the E-MLB dataset [10]. Figure 7 collects the results. We use two state-of-the-art approaches for reconstruction: E2VID [46], which does not rely on motion, and EVILIP [59], which utilizes optical flow as prior. + +Our method provides flow (Fig. 7, column (b)) and signal events (sharp IWEs in column (c)). Comparing the reconstruction with our event denoising method (column (d)) and + +![](images/2942f1d5329e90790c245bb3b0ffa43e4a385cb733d3fe774bf00e468932af3c.jpg) +Figure 7. Image intensity reconstruction on E-MLB data [10]. The first two rows use the reconstruction method EVILIP [59] and the last two rows use E2VID [46]. Reconstruction on denoised events produces higher quality images (i.e., fewer artifacts) than on raw event data. + +![](images/4d848c24e69a6b0310d66bc0ab07dafcecc1b0f870f12b08f4bdcf9ab58ac0da.jpg) + +![](images/8e3f08c4be859530b52f646523d3008f77d225d5566f79a73cfaf72886384212.jpg) +(a) Signal +(b) Noise + +![](images/e47a2e47b6178fc312284ab664e59473cb3a6bc1f269840816557a453b40025c.jpg) + +![](images/a93237ea407609e8ed6854f97b1b8b2e1dd8533b98211bd71b2ed4b3888904a4.jpg) + +![](images/1090b5d162cf6e76d1326c8bf1468a8da31b27f27afbdba7f2e28b53f23105e7.jpg) +(c) Reconstruction +Figure 8. Sensitivity analysis: effect of different target ratio parameters $\tau$ on image reconstruction using EVILIP [59]. + +the reconstruction from the raw event data (column (e)), the images after event denoising have fewer artifacts than those without denoising, for both reconstruction methods tested. We observe that E2VID is more robust against noise than EVILIP, possibly due to noise events in its training (simulation) data. Due to the ND filters, the quality of the reference frames (last column) is limited, and hence, we discuss the results qualitatively. Nevertheless comparing the reconstructed images to the reference frames, evidences the HDR advantages of event cameras over frame-based ones. + +![](images/8114e778fee861c5ec5b676583eb98ef667d0c2040669fad9a5bf24702675f5c.jpg) +(a) Ref. events + +![](images/dad65955ddb772fddc961fde6e2fae628a52c0fa9bd7e7ea4a8e82ad65d9c5b4.jpg) +(b) $k = 1$ px +Figure 9. Results of different Gaussian kernel sizes $k$ for sensitivity on different depth scenes. + +![](images/382280d018ace21f9cf8d0e6647a8011ff7758662371f1eda87bbce7c5c783b0.jpg) +(c) $k = 3$ px + +# 4.5. Sensitivity + +Effect of Noise Ratio. As Sec. 4.2 already showed the sensitivity for denoising accuracy, let us further discuss the sensitivity using image reconstruction. Figure 8 presents the result of the Bicycle sequence with ND64 filter [10]. While the GT image is highly underexposed and not informative (as those on the last column of Fig. 7), the reconstructed images clearly reflect the choice of the signal target ratio $\tau$ . As more events are removed, $(i)$ the reconstructed image becomes more homogeneous but preserves strong edges, and $(ii)$ the removed events (2nd col.) contain more scene edges (i.e., signal). We quantitatively analyze the former in the supplementary using non-reference image quality metrics. + +Effect of the Kernel Size. Figure 9 shows the sensitivity analysis for different kernel sizes. By increasing the size of the Gaussian kernel in the IWE, our method can emphasize contour edges over isolated points, thus better preserving signal events in regions with low IWE intensities (center of the image, corresponding to far away scene points). + +# 4.6. Ablation of Motion Estimator + +The proposed approach can be combined with other estimation methods, such as deep-neural-network (DNN) flow estimators. To this end, we show an example of E-RAFT [21]. The model-based method alone [48] is known to output a degraded flow in night sequences with lots of noise, such as zurich_city_12a [20]. As shown in Fig. 10, our method can identify and discard noisy events. E-RAFT provides sharper and more natural edges in IWEs after denoising compared to MultiCM [48] (thanks to its more accurate flow). ESR scores are similar: 2.39 for E-RAFT and 2.45 for MultiCM. While we focus on the inference of DNN in the proposed pipeline, future research could look into accommodating training within the proposed joint estimation framework. + +# 4.7. Runtime + +Using 30k events from the ECD dataset $(240 \times 180\mathrm{px})$ on a Mac M1 CPU (2020), the denoising operations (scoring and ranking) take less than $0.1\mathrm{s}$ extra computation in total, out of $2\mathrm{s}$ of the entire processing time (i.e., $5\%$ computation increase with respect to the base method). Nonetheless, accommodating learning-based approaches in our framework could be key to the trade-off between accuracy and speed. + +![](images/5b6abb16bd161e709d7bd0ac6292b7f59ef6f681827178c5c494a392c0ce2ae2.jpg) +Raw Event + +![](images/06d32f0c215e5bc8e90658d248159d2e5d05798bdac2631a711e2e7c202f3ec9.jpg) +IWE (No Warp) + +![](images/2a4bba47789dd768ef59eda58d7f2d069a6a397fa374df9858dbbdfab1a7a1eb.jpg) + +![](images/55b6d476952e74d25445e291eb3240e7f889093496c1f4a41c9b74846768e6f6.jpg) +(a) Signal Events +Figure 10. Effect of adopting two different motion estimators (E-RAFT or MultiCM) on the DSEC night scene (zurich_12a) [20]. Despite small IWE differences (e.g., in the car's hood), the proposed framework supports several motion estimators. + +![](images/adc7580ce076347908892112af1c604d5112ed7fc73d24fbecbc84f0ffa1d0b8.jpg) + +![](images/99e035bdcbd5ae0e10cd26b64bd30e41b443b1529dc6dbfb309a5ec5b8de37e4.jpg) +(b) IWE + +![](images/53fe5b6e7247e26af7ed36683d7ad1ac96975aca3bca3e49aff6dee3f50caba0.jpg) + +![](images/4dd08aff9896dbb3b06d5cb095a14ee56deb45039f14ace0ac1ffb0f718fe5b9.jpg) +(c) Flow Estimation + +# 5. Limitations + +Our method expects as parameter a target noise level, but the intrinsic ("true") noise level depends on the camera, scene, etc. and is therefore unknown. More research in this direction is needed for automatic parameter tuning. Also, notice that there might be a potential limitation of the denoising metric ESR, since it depends on the motion parameter (see Suppl. Mat. of [29]). Although our denoising efficacy is confirmed from multiple perspectives, we believe that advancing denoising metrics for real-world data (where it gets interesting due to the absence of GT labels) is a central direction for future work in event-based vision. + +# 6. Conclusion + +In summary, this work proposes that: (i) Joint estimation of noise and motion is a novel problem setting compared to conventional (isolated) denoising, and is sensitive given the event data's inherent motion dependency. (ii) The proposed joint estimation framework can accommodate various motion estimators and has shown its efficacy in improving both denoising and motion estimation accuracies. (iii) It is important to evaluate denoising results on both, denoising benchmarks and downstream tasks (e.g., motion estimation, intensity reconstruction), since current denoising benchmarks have shortcomings in terms of metrics, data, and baselines. In particular, our method achieves notable denoising results (state-of-the-art on the E-MLB benchmark and competitive on DND21) while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances both event-denoising theory and practical denoising use-cases. + +# Acknowledgments + +We would like to thank Dr. B. Jiang for useful discussions. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 "Science of Intelligence" - project number 390523135. + +# References + +[1] Saeed Afshar, Andrew Peter Nicholson, Andre van Schaik, and Gregory Cohen. Event-based object detection and tracking for space situational awareness. IEEE Sensors Journal, 20(24):15117-15132, 2020. +[2] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Trans. Neural Netw. Learn. Syst., 35(3):4110-4124, 2022. +[3] Sami Arja, Alexandre Marcireau, Richard L. Balthazor, Matthew G. McHarg, Saeed Afshar, and Gregory Cohen. Density invariant contrast maximization for neuromorphic earth observations. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 3984-3994, 2023. +[4] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1701-1710, 2020. +[5] R Wes Baldwin, Mohammed Almatrafi, Jason R Kaufman, Vijayan Asari, and Keigo Hirakawa. Inceptive event timesurfaces for object classification using neuromorphic cameras. In Image Analysis and Recognition: 16th Int. Conf., ICIAR 2019, Waterloo, ON, Canada, August 27-29, 2019, Proceedings, Part II 16, pages 395-403. Springer, 2019. +[6] T. D. Barfoot. State Estimation for Robotics - A Matrix Lie Group Approach. Cambridge University Press, 2015. +[7] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240x180 130dB $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE J. Solid-State Circuits, 49(10):2333-2341, 2014. +[8] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In AAAI Conf. Artificial Intell., pages 303-311, 2022. +[9] Tobi Delbruck. Frame-free dynamic digital vision. In Proc. Int. Symp. Secure-Life Electron., pages 21-26, 2008. +[10] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-MLB: Multilevel benchmark for event-based camera denoising. IEEE Trans. Multimedia, 26:65–76, 2023. +[11] Peiqi Duan, Zihao Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2021. +[12] Peiqi Duan, Zihao W. Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve + +neuromorphic events. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12824-12833, 2021. +[13] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. +[14] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020. +[15] Guillermo Gallego and Davide Scaramuzza. Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett., 2(2):632-639, 2017. +[16] Guillermo Gallego and Anthony Yezzi. A compact formula for the derivative of a 3-D rotation in exponential coordinates. J. Math. Imaging Vis., 51(3):378-384, 2014. +[17] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3867-3876, 2018. +[18] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vision. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12272-12281, 2019. +[19] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154-180, 2022. +[20] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett., 6(3):4947-4954, 2021. +[21] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vision (3DV), pages 197–206, 2021. +[22] Rui Graca and Tobi Delbruck. Towards a physically realistic computationally efficient dvs pixel model. In Int. Image Sensor Workshop (IISW), 2025. +[23] Cheng Gu, Erik Learned-Miller, Daniel Sheldon, Guillermo Gallego, and Pia Bideau. The spatio-temporal Poisson point process: A simple model for the alignment of event camera data. In Int. Conf. Comput. Vis. (ICCV), pages 13495-13504, 2021. +[24] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Trans. Pattern Anal. Mach. Intell., 45(1):785-795, 2023. +[25] Shuang Guo and Guillermo Gallego. CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot., 40: 2442-2461, 2024. +[26] Shasha Guo, Ziyang Kang, Lei Wang, Shiming Li, and Weixia Xu. Hashheat: An O(C) complexity hashing-based + +filter for dynamic vision sensor. In 25th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 452-457, 2020. +[27] Shuang Guo, Friedhelm Hamann, and Guillermo Gallego. Unsupervised joint learning of optical flow and intensity with event cameras. In Int. Conf. Comput. Vis. (ICCV), 2025. +[28] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In Eur. Conf. Comput. Vis. (ECCV), pages 18-37, 2024. +[29] Bin Jiang, bo Xiong, Bohan Qu, M. Slaman Asif, you Zhou, and Zhan Ma. EDformer: Transformer-based event denoising across varied noise levels. In Eur. Conf. Comput. Vis. (ECCV), pages 1-12, 2024. +[30] Pritam P. Karmokar, Quan H. Nguyen, and William J. Beksi. Secrets of edge-informed contrast maximization for event-based vision. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), 2025. +[31] Alireza Khodamoradi and Ryan Kastner. O(N)-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Trans. Emerg. Topics Comput., PP(99):1-1, 2018. +[32] Haram Kim and H. Jin Kim. Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett., 6(3):6016-6023, 2021. +[33] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), pages 8601-8608, 2021. +[34] Xavier Lagorce, Garrick Orchard, Francesco Gallupi, Bertram E. Shi, and Ryad Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(7):1346-1359, 2017. +[35] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128x128 120dB 30mW asynchronous vision sensor that responds to relative intensity change. In IEEE Int. Solid-State Circuits Conf. (ISSCC), pages 2060-2069, 2006. +[36] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ dB 15 μs latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits, 43(2):566-576, 2008. +[37] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In IEEE Int. Symp. Circuits Syst. (ISCAS), pages 722-725, 2015. +[38] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012. +[39] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2012. +[40] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research, 36(2):142-149, 2017. + +[41] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot., 34(6):1425-1440, 2018. +[42] Urbano Miguel Nunes and Yiannis Demiris. Robust event-based vision model estimation by dispersion minimisation. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9561-9573, 2022. +[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3445–3454, 2021. +[44] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE de Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Int. Conf. Comput. Vis. (ICCV), pages 9661-9671, 2023. +[45] Xin Peng, Ling Gao, Yifu Wang, and Laurent Kneip. Globally-optimal contrast maximisation for event cameras. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3479-3495, 2022. +[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Trans. Pattern Anal. Mach. Intell., 43 (6):1964-1980, 2021. +[47] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high speed scenarios. IEEE Robot. Autom. Lett., 3(2): 994-1001, 2018. +[48] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis. (ECCV), pages 628-645, 2022. +[49] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):1-20, 2022. +[50] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):7742-7759, 2024. +[51] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Int. Conf. Comput. Vis. (ICCV), pages 7243-7252, 2019. +[52] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In Eur. Conf. Comput. Vis. (ECCV), pages 534-549, 2020. +[53] Gemma Taverni, Diederik Paul Moeys, Chenghan Li, Celso Cavaco, Vasyl Motsnyi, David San Segundo Bello, and Tobi Delbruck. Front and back illuminated Dynamic and Active Pixel Vision Sensors comparison. IEEE Trans. Circuits Syst. II (TCSII), 65(5):677-681, 2018. +[54] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. + +In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6358-6367, 2019. +[55] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1606–1616, 2020. +[56] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Trans. Multimedia, 23:1148-1159, 2020. +[57] Ryo Yamaki, Shintaro Shiba, Gallego Guillermo, and Yoshimitsu Aoki. Iterative event-based motion segmentation by variational contrast maximization. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 4918-4927, 2025. +[58] Pei Zhang, Haosen Liu, Zhou Ge, Chutian Wang, and Edmund Y Lam. Neuromorphic imaging with joint image deblurring and event denoising. IEEE Trans. Image Process., 2024. +[59] Zelin Zhang, Anthony Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8372-8389, 2023. +[60] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Trans. Neural Netw. Learn. Syst., pages 1-13, 2021. +[61] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Trans. Robot., 37(5):1433-1450, 2021. +[62] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5816-5824, 2017. +[63] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. Autom. Lett., 3(3):2032-2039, 2018. +[64] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 989-997, 2019. \ No newline at end of file diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/images.zip b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e762e92ccf3abf25ee5ac48ae95343e01381cdfc --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6980bc54ca374ef0ef3888576f49df8f34cc0566c4056fc70292ad7ee946fb65 +size 1176048 diff --git a/2025/Simultaneous Motion And Noise Estimation with Event Cameras/layout.json b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a3c51cfe6dc861d7dbcbf4b48de448f843da53ac --- /dev/null +++ b/2025/Simultaneous Motion And Noise Estimation with Event Cameras/layout.json @@ -0,0 +1,10980 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 506, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 506, + 119 + ], + "type": "text", + "content": "Simultaneous Motion And Noise Estimation with Event Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "spans": [ + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "text", + "content": "Shintaro Shiba" + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "text", + "content": ", Yoshimitsu Aoki" + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "text", + "content": " and Guillermo Gallego" + }, + { + "bbox": [ + 153, + 142, + 462, + 157 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 157, + 550, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 550, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 550, + 186 + ], + "type": "text", + "content": "1 Keio University, 2 Woven by Toyota, Inc., Japan. 3 Technische Universität Berlin, 4 Einstein Center Digital Future, Robotics Institute Germany, and Science of Intelligence Excellence Cluster, Germany." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 478 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 478 + ], + "type": "text", + "content": "Event cameras are emerging vision sensors whose noise is challenging to characterize. Existing denoising methods for event cameras are often designed in isolation and thus consider other tasks, such as motion estimation, separately (i.e., sequentially after denoising). However, motion is an intrinsic part of event data, since scene edges cannot be sensed without motion. We propose, to the best of our knowledge, the first method that simultaneously estimates motion in its various forms (e.g., ego-motion, optical flow) and noise. The method is flexible, as it allows replacing the one-step motion estimation of the widely-used Contrast Maximization framework with any other motion estimator, such as deep neural networks. The experiments show that the proposed method achieves state-of-the-art results on the E-MLB denoising benchmark and competitive results on the DND21 benchmark, while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances event-data denoising theory and expands practical denoising use-cases via open-source code. Project page: https://github.com/tub-rip/ESMD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 135, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 135, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 135, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 700 + ], + "type": "text", + "content": "Event cameras are emerging vision sensors that overcome shortcomings of conventional cameras (e.g., motion blur, limited dynamic range, data and power efficiency), yet they suffer from a considerable amount of noise due to their novelty and operation in low-power (transistor subthreshold) conditions [19, 36]. Since event cameras are suitable for many computer vision tasks, especially motion-related tasks, it is paramount to classify event data that is related to motion (i.e., signal) and that is not (i.e., noise). However, denoising event data is a challenging problem because noise properties have yet to be fully characterized [22, 24] and it is infeasible to define ground-truth (GT) noise labels in real-world data recordings. Previous works either leveraged simulation or prepared real event data and aggressively suppressed noise occurrence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 712 + ], + "type": "text", + "content": "However, such artificial modifications on the real event" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 376 + ], + "type": "image", + "image_path": "60f0d0e4b9b4750bd32f26bb9e4daa66eef54a87131f6d80f3004d9d4875aded.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 378, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 411 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed method. We use only raw events as input, and estimate signal events, noise events, and motion (e.g., dense optical flow and ego-motion estimation)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 638 + ], + "type": "text", + "content": "data could change event signal/noise characteristics, possibly deteriorating the information that the original event had for downstream tasks (e.g., motion estimation). Our work is inspired by the observation that motion is an intrinsic part of event data since scene edges cannot be sensed without motion. Rather than solving denoising and motion estimation tasks separately, our idea is to leverage motion information to improve denoising (as \"vice-versa\" is well known). In this work (Fig. 1), we rethink and investigate the connection between denoising and motion estimation, by asking ourselves whether they can be combined in a more integrated way to benefit the results of both tasks simultaneously. We formulate the simultaneous estimation problem by leveraging the idea of motion compensation in the Contrast Maximization (CMax) framework, defining novel metrics about the contribution to the contrast score for each event. In summary, we present several distinctive contributions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 642, + 553, + 666 + ], + "type": "text", + "content": "1. We propose a novel method to jointly estimate noise and motion only from event data (Sec. 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": "2. We demonstrate that our joint estimation method " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " achieves state-of-the-art results in existing denoising benchmarks (Sec. 4.2), and " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " improves the robustness of motion estimation (CMax framework) (Sec. 4.3)." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6959" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "3. We show the flexibility of our method on " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " further applications, such as intensity reconstruction, validating its image quality (Secs. 4.4 and 4.5), and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " in combination with learning-based motion estimators (Sec. 4.6)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 156 + ], + "type": "text", + "content": "We hope our work opens a new line of research for event cameras, by considering the problems of motion estimation and denoising together to take advantage of their coaction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 142, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 142, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 142, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 187, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 307 + ], + "type": "text", + "content": "Denoising is a fundamental problem in event cameras since there are a lot of leakage events [24], and event-processing algorithms may suffer in scenarios with high noise [3]. Removing event noise has a double effect: it reduces the load of subsequent processing stages (as the load is proportional to the number of events) and improves their output (as the input data becomes of higher quality). In this work, we focus on noise due to the leakage inside the camera hardware, such as background activity (BA) events [36]. Events due to flickering or active lights are therefore not our priority." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 307, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 402 + ], + "type": "text", + "content": "Classical denoising works focus on the spatio-temporal correlation of events, as BA events have fewer neighboring events in space-time coordinates than signal events. This approach includes widely-used BA Filter (BAF) [9], and other types of spatio-temporal filters [26, 31, 37, 56]. Some works utilize additional sensor inputs (e.g., frames) to estimate noise probabilities [4, 12, 30, 55, 58], since edges are informative of the signal events produced." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 582 + ], + "type": "text", + "content": "Recently, learning-based approaches have gained more attention, such as multilayer perceptrons [24], convolutional neural networks [1, 4, 12], graph neural networks [2], and transformers [29]. They provide the probability of noise (i.e., signal-noise classification). Nonetheless, the challenge of learning-based approaches is the need for ground truth (GT) noise labels for training. GT labels have been obtained through simulation (e.g., [13, 24]) as well as real-world recordings, however, real-world recordings aggressively remove noisy events using existing filters (e.g., [24]) or use unnatural illumination settings (e.g., [29]), by leveraging the differences of noise occurrence due to different scene brightness [24, 35]. While these approaches are effective for benchmarking, the challenge of obtaining GT labels for practical real-world datasets remains unsolved." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Our work builds upon a series of successful methods for motion estimation and their applications. Event-based motion estimation is an important topic since event cameras naturally respond to motion in the scenes [19]. In particular, Contrast Maximization (CMax) [17, 50] is a state-of-the-art framework for motion estimation, which has various applications, such as rotational motion estimation [15, 23, 32, 42, 45], optical flow estimation [27, 28, 44, 48, 64], and motion segmentation [51, 57, 60]. We investigate the connection and potential benefits of applying the CMax framework to denoising during the estimation process (i.e., simul" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "taneously), demonstrating motion estimation improvements as well as competitive denoising accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 240 + ], + "type": "text", + "content": "A prior work, \"ProgressiveMotionSeg\" [8], has extended the CMax-based segmentation method [51] with extra weights that accounted for how much each event contributed to the estimated motion clusters. However, [8] mainly focuses on the task of motion segmentation using low degrees-of-freedom (DOF) warps (e.g., 2-DOF feature flow) and does not discuss denoising efficacy or evaluate on standard denoising benchmarks. In contrast, our method supports various types of motion estimation (both low-DOF and high-DOF warps, such as optical flow), and considers that noise shall not be used to estimate motion, as from first principles we deem noise and motion to be uncorrelated." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 249, + 397, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 397, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 397, + 262 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "content": "Event cameras consist of independent pixels that generate asynchronous \"events\" when the logarithmic brightness at the pixel increases or decreases by a predefined contrast sensitivity. Each event " + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "inline_equation", + "content": "e_k \\doteq (\\mathbf{x}_k, t_k, p_k)" + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "content": " contains the pixel coordinates " + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "content": ", the timestamp " + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "content": ", and polarity " + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "inline_equation", + "content": "p_k \\in \\{+1, -1\\}" + }, + { + "bbox": [ + 313, + 270, + 554, + 365 + ], + "type": "text", + "content": " of the brightness change. Events occur asynchronously and sparsely on the pixel grid, resulting in a variable data rate based on the scene texture and dynamics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 372, + 446, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 446, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 446, + 383 + ], + "type": "text", + "content": "3.1. Contrast Maximization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "content": "The CMax framework [17] assumes events " + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{E} \\doteq \\{e_k\\}_{k=1}^{N_e}" + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "content": " are caused by moving edges, and it transforms their coordinates according to a motion model " + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "content": ", producing a set of warped events " + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{E}'_{t_{\\mathrm{ref}}} \\doteq \\{e_k'\\}_{k=1}^{N_e}" + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "content": " at a reference time " + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 388, + 554, + 439 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 445, + 553, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 445, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 348, + 445, + 553, + 458 + ], + "type": "interline_equation", + "content": "e _ {k} \\dot {=} \\left(\\mathbf {x} _ {k}, t _ {k}, p _ {k}\\right) \\mapsto e _ {k} ^ {\\prime} \\dot {=} \\left(\\mathbf {x} _ {k} ^ {\\prime}, t _ {\\text {r e f}}, p _ {k}\\right). \\tag {1}", + "image_path": "03df1e27aa3a24d972fbb0050a0e19cbddee8bdba64c87d3a938888fde80e98f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "text", + "content": "The warp " + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k^{\\prime} = \\mathbf{W}(\\mathbf{x}_k, t_k; \\boldsymbol{\\theta})" + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "text", + "content": " transports each event from " + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 461, + 553, + 498 + ], + "type": "text", + "content": " along the motion curve that passes through it. Then, they are aggregated on an image of warped events (IWE):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 502, + 553, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 502, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 365, + 502, + 553, + 519 + ], + "type": "interline_equation", + "content": "I (\\mathbf {x}; \\mathcal {E} _ {t _ {\\text {r e f}}} ^ {\\prime}, \\boldsymbol {\\theta}) \\doteq \\sum_ {k = 1} ^ {N _ {e}} \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k} ^ {\\prime}\\right), \\tag {2}", + "image_path": "d2697b6cceb15aac52f74e278cf6160a3f72bbe8ab74fa9fcc20bcf329f2f3ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": "where each pixel " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " sums the number of warped events " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k^\\prime" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " that fall within it. The Dirac delta is approximated by a Gaussian, " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\delta (\\mathbf{x} - \\pmb {\\mu})\\approx \\mathcal{N}(\\mathbf{x};\\pmb {\\mu},\\epsilon^2\\mathrm{Id})" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\epsilon = 1" + }, + { + "bbox": [ + 313, + 521, + 554, + 556 + ], + "type": "text", + "content": " pixel." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "content": "Next, an objective function, such as the contrast of the IWE (2) (image variance [18] " + }, + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}\\big(I(\\mathbf{x};\\pmb {\\theta})\\big)\\doteq \\frac{1}{|\\Omega|}\\int_{\\Omega}(I(\\mathbf{x};\\pmb {\\theta}) - \\mu_I)^2 d\\mathbf{x})" + }, + { + "bbox": [ + 313, + 557, + 554, + 605 + ], + "type": "text", + "content": " , measures the goodness of fit between the events and the candidate motion curves (warp)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "text", + "content": "Finally, an optimization algorithm iterates the above steps to find the motion parameters " + }, + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 313, + 606, + 554, + 665 + ], + "type": "text", + "content": " that maximize the alignment of events caused by the same scene edge. Event alignment is measured by the strength of the edges of the IWE, which is related to image contrast [18]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Motion Models. As exemplary warps " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " we focus on two popular problem settings: rotational motion (3-DOF angular velocity estimation [23, 25, 42, 45]) and dense optical flow (pixel-wise velocity estimation [43, 48, 50, 64]," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6960" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 77, + 542, + 321 + ], + "blocks": [ + { + "bbox": [ + 70, + 77, + 542, + 321 + ], + "lines": [ + { + "bbox": [ + 70, + 77, + 542, + 321 + ], + "spans": [ + { + "bbox": [ + 70, + 77, + 542, + 321 + ], + "type": "image", + "image_path": "fe741d6d8a61f060504e939b6f26702e5ebdcd616adc04c751cd1d18fca0f992.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 329, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 555, + 353 + ], + "type": "text", + "content": "Figure 2. Qualitative result of denoising and flow estimation on E-MLB and DND21. The driving sequence has " + }, + { + "bbox": [ + 55, + 329, + 555, + 353 + ], + "type": "inline_equation", + "content": "5\\mathrm{-Hz}" + }, + { + "bbox": [ + 55, + 329, + 555, + 353 + ], + "type": "text", + "content": " artificial noise. (b) is an alternative visualization of (a) with the same grayscale coding as (d). Flow is visualized on all events and flow color is given in Fig. 1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": "resulting in " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "2N_{p}" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": "-DOFs, where " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": " is the number of pixels). For the angular velocity estimation during a small time interval, the warp is parametrized by the angular velocity " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\theta \\equiv \\omega = (\\omega_{x},\\omega_{y},\\omega_{z})^{\\top}" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{h}(t)\\sim \\mathbb{R}(\\boldsymbol {\\omega}t)\\mathbf{x}^{h}(0)" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": ", using calibrated homogeneous coordinates " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{h}" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": " and exponential coordinates " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\mathbb{R}(\\phi)\\doteq \\exp (\\phi^{\\wedge})" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": " [6, 16]. The warp for dense optical flow is " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k^\\prime = \\mathbf{x}_k + (t_k - t_{\\mathrm{ref}})\\mathbf{v}(\\mathbf{x}_k)" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\mathbf{v}(\\mathbf{x})\\}_{\\mathbf{x}\\in \\Omega}" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": " is a flow field on the image plane " + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 373, + 295, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 243, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 243, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 243, + 488 + ], + "type": "text", + "content": "3.2. Denoising Based on Local Contrast" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": "As Fig. 1 shows, our method is iterative. In each iteration, motion estimation is refined and events are classified into signal " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_s" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " or noise " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_n" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": ". Without labels or additional information, events rely on each other to realize this classification. To this end, we compute a score " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " for each event, rank the events according to their scores, and threshold at some value " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": ". The threshold is computed by assuming the amount or ratio of noisy events is known " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "T \\equiv T(\\eta)" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": ", which allows us to have high-level control over the denoising process. Hence we classify events into signal (the " + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\tau \\doteq (1 - \\eta)" + }, + { + "bbox": [ + 55, + 492, + 296, + 624 + ], + "type": "text", + "content": " percentage of events with highest scores) and noise:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 630, + 296, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 630, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 68, + 630, + 296, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {s} \\equiv \\mathcal {E} _ {\\text {s i g n a l}} \\doteq \\left\\{e _ {k} \\in \\mathcal {E} \\mid c _ {k} > T (\\eta) \\right\\}, \\tag {3}", + "image_path": "322f2017a2e7675561604a06337d03c0e8c327d1d6527402327b10a61bc0eed4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 646, + 270, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 646, + 270, + 659 + ], + "spans": [ + { + "bbox": [ + 70, + 646, + 270, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {n} \\equiv \\mathcal {E} _ {\\text {n o i s e}} \\dot {=} \\mathcal {E} \\backslash \\mathcal {E} _ {\\text {s i g n a l}} \\quad (\\text {c o m p l e m e n t} \\mathcal {E} _ {\\text {s i g n a l}}).", + "image_path": "80f6da77865236e4cd66928b87b0dc3a64397fb607f2ece990bc8f80c443ff13.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "We studied different choices for the per-event score " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " based on the agreement among warped events " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}' = \\mathcal{E}_s' \\cup \\mathcal{E}_n'" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": ", through their IWEs, and concluded on two, as follows. Letting " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_k^s \\equiv I_k^s(\\mathbf{x}_k')" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_k^n \\equiv I_k^n(\\mathbf{x}_k')" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "I_k \\equiv I(\\mathbf{x}_k') = I_k^s + I_k^n" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": "be the IWE values of previously classified signal events, noise events and all events, respectively, the two choices are: (i) " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "^{1}c_{k} \\doteq I_{k}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": ", and (ii) the ratio of signal events " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "^{2}c_{k} \\doteq I_{k}^{s} / I_{k}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": " (probabilistic viewpoint). In fact, both are equivalent: rule " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "^{1}c_{k} > T_{1}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": " is the same as rule " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "^{2}c_{k} > T_{2}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "T_{2} = 1 - I_{k}^{n} / T_{1}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": " and the event noise " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "I_{k}^{n}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": " is known. We use " + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "inline_equation", + "content": "^{1}c_{k}" + }, + { + "bbox": [ + 313, + 373, + 555, + 468 + ], + "type": "text", + "content": ", the number of warped events (also called local contrast [18, 51]), because it is the simplest to interpret." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "content": "The intuition is that \"signal\" events (e.g., due to motion) warped according to the true motion lie all at motion-compensated edges, and therefore produce a sharp IWE. Noise events, produced all over the image plane, are not expected to contribute to the IWE's sharpness. Hence, we may classify events into signal or noise according to how much they contribute to edge strength. The IWE (2) provides one such measure of edge strength: the higher " + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "inline_equation", + "content": "I(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "content": ", the more events are warped to that pixel location " + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "content": " (i.e., more events support the same scene edge), thus producing a sharper IWE. Hence, we define how much an event " + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "content": " contributes to edge strength by means of the IWE (2), " + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "inline_equation", + "content": "^1 c_k" + }, + { + "bbox": [ + 313, + 470, + 556, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": "Invariance. Note that (3) is invariant to monotonically increasing functions of " + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": ", as such functions preserve the order. The same " + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{signal}}" + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": " is obtained if one ranks and thresholds (at the corresponding level) values " + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\{c_k\\}" + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\{\\ln(c_k)\\}" + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\{e^{c_k}\\}" + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": ", etc. They are equivalent under the ranking rule (3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Robustness to Various Edge Strengths. The Gaussian kernel in the IWE controls the sensitivity to edge strength. Since the proposed method is based on both signal and noise" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6961" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "IWEs, increasing the size of the Gaussian kernel in the IWE can emphasize contour edges over isolated points. This preserves signal events better in regions with low IWE intensities, for example at pixels with large depth values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 197, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 197, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 197, + 140 + ], + "type": "text", + "content": "3.3. Alternating Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 204 + ], + "type": "text", + "content": "Classifying events into signal and noise requires knowledge of the true motion, and estimation of the true motion (e.g., using CMax) requires knowledge of the signal events because, by definition, noise events carry no information about motion. Hence, it is a circular dependency problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": "We approach its solution iteratively, as shown in Fig. 1, by combining motion estimation (Sec. 3.1) and S/N classification (Sec. 3.2). Assuming some initialization, the current signal events " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{signal}}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": " are used to estimate the motion via CMax (1 step is enough); then all events are warped to compute the IWE and the per-event scores " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": ", from which the sets (3) are recomputed, in preparation for the next iteration. In practice, we initialize with a random split of " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{signal}}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{noise}}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": ", update " + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{signal}}" + }, + { + "bbox": [ + 54, + 204, + 295, + 335 + ], + "type": "text", + "content": " at every warp during the optimization iteration until the process converges (as flagged by the convergence of the motion parameters in CMax)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "text", + "content": "Computational Complexity. The computational complexity of one iteration of the proposed method is " + }, + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "inline_equation", + "content": "O(N_{p} + N_{e}\\log N_{e})" + }, + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "text", + "content": ", which is slightly larger than the original CMax (i.e., " + }, + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "inline_equation", + "content": "O(N_{p} + N_{e})" + }, + { + "bbox": [ + 55, + 336, + 296, + 396 + ], + "type": "text", + "content": ") due to the search for the highest-ranked events during the optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 406, + 136, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 136, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 136, + 419 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "type": "text", + "content": "We evaluate the proposed method on four datasets using various metrics (in Sec. 4.1). In the following sections using common benchmarks, we assess denoising (Sec. 4.2), joint motion estimation (Sec. 4.3), and the efficacy on intensity reconstruction (Sec. 4.4). We also conduct sensitivity analyses (Sec. 4.5) and an ablation (Sec. 4.6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 272, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 272, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 272, + 518 + ], + "type": "text", + "content": "4.1. Datasets, Metrics, and Hyper-parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 570 + ], + "type": "text", + "content": "Datasets. E-MLB [10] is a large-scale, de-facto dataset to benchmark denoising. It consists of 100 sequences with 4 different brightness levels using neutral-density (ND) filters on a DAVIS346 camera " + }, + { + "bbox": [ + 55, + 522, + 295, + 570 + ], + "type": "inline_equation", + "content": "(346\\times 260\\mathrm{px})" + }, + { + "bbox": [ + 55, + 522, + 295, + 570 + ], + "type": "text", + "content": " during day and night." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": "Historically, the two sequences (hotel and driving) of the DND21 dataset [24] have been widely used for denoising evaluation [10, 24, 29]. Both sequences are recorded with a DAVIS346 camera [53]. They are regarded as \"signal\" data, as the sequences have been aggressively filtered offline. Combined with other pure BA noise sequences, the dataset has event-wise noise annotations that are useful for evaluation. We use the above sequences with different noise rates, 1, 3, 5, 7, " + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 570, + 295, + 677 + ], + "type": "text", + "content": " per pixel, following prior work [29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The ECD dataset [40] is a standard dataset for various tasks including camera ego-motion estimation [15, 41, 47, 49, 61, 62]. Using a DAVIS240C camera " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "(240\\times 180\\mathrm{px}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " [7])," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 220 + ], + "type": "table", + "html": "
E-MLB (Day)E-MLB (Night)DND21
ND1ND4ND16ND64ND1ND4ND16ND64
Model-basedRaw0.8210.8240.8150.7860.8900.8240.7860.7680.869
BAF [9]0.8610.8690.8760.8900.9460.9730.9920.9420.920
TS [34]0.8770.8870.8700.8371.0330.9440.8860.7970.985
KNoise [31]0.8460.8370.8300.8070.9540.9560.8710.8170.887
EvFlow [54]0.8480.8780.8680.8330.9690.9830.8890.7971.006
IETS [5]0.7720.7850.7770.7530.9500.8230.8040.7110.900
Ynoise [14]0.8660.8630.8570.8211.0090.9430.8750.7920.966
GEF [11]1.0510.9380.9350.9271.0270.9550.9460.9350.932
DWF [24]0.8780.8760.8660.8650.9230.9620.9880.9320.905
Ours0.9380.9580.9860.9501.0370.9610.9450.9320.992
LearningEDnCNN [4]0.8870.9080.9030.9121.0011.0241.0791.0860.977
EventZoom [12]0.9960.9880.9960.9701.0551.0071.0100.9881.059
MLPF [24]0.8510.8550.8460.8400.9260.9280.9100.9060.944
EDformer [29]0.9520.9550.9560.9421.0481.0191.0761.0991.041
", + "image_path": "73514bff06d4d47271fff8cbe639d757948a6aebf7931d4c5eb2bd9be0f85a54.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 285, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 313, + 228, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 273 + ], + "type": "text", + "content": "Table 1. Mean ESR (MESR↑) results among denoising methods on the event denoising datasets E-MLB [10] and DND21 [24]. In each category, the best is in bold and the second best is underlined. Methods are chronologically sorted within each category." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 285, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 317, + 285, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 285, + 553, + 405 + ], + "type": "table", + "html": "
1Hz5Hz10Hz
hoteldrivinghoteldrivinghoteldriving
Model-basedBAF [9]0.9540.8480.8920.7930.8370.748
TS [34]0.9720.9310.9610.9270.9620.920
KNoise [31]0.6770.6300.6700.6240.6410.614
Ynoise [14]0.9690.9410.9230.9090.8990.880
DWF [24]0.9270.7410.8620.6900.7960.656
Ours1.0140.8820.9630.8550.9610.836
LearningEDnCNN [4]0.9570.8870.9370.8750.9010.874
MLPF [24]0.9700.8890.9700.8850.9630.876
EDformer [29]0.9930.9540.9850.9420.9700.926
", + "image_path": "d3e91ef88b172889809ec4784cccb1f29e26bb9aae2d3d9bb25ed57a7837fd46.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 413, + 553, + 435 + ], + "lines": [ + { + "bbox": [ + 313, + 413, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 553, + 435 + ], + "type": "text", + "content": "Table 2. The AUC" + }, + { + "bbox": [ + 313, + 413, + 553, + 435 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 413, + 553, + 435 + ], + "type": "text", + "content": " of ROC on the two DND21 sequence at different noise rates. We follow [29] for baseline methods selection." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 449, + 555, + 547 + ], + "blocks": [ + { + "bbox": [ + 316, + 449, + 555, + 547 + ], + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 547 + ], + "type": "image", + "image_path": "9ef37170daa390a676e23b6f9a7959ad5ea887116686e84b9fcbcda148b57c65.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 578 + ], + "type": "text", + "content": "Figure 3. ROC curves on the DND 21 dataset sequences. Dash lines represent learning-based methods (see also Tab. 2)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 592, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 553, + 640 + ], + "type": "text", + "content": "each sequence provides events, frames, calibration information, IMU data, and ground truth (GT) camera poses (at " + }, + { + "bbox": [ + 313, + 592, + 553, + 640 + ], + "type": "inline_equation", + "content": "200\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 592, + 553, + 640 + ], + "type": "text", + "content": "). We use boxesRotation and dynamicRotation sequences to evaluate rotational motion estimation accuracy." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Finally, we also use sequences from MVSEC [63] and zurich_city_12a sequence from the DSEC dataset [20], which are typically used for optical flow estimation. We use MVSEC sequences for quantitative evaluation and the DSEC sequence for qualitative evaluation, since it is captured during the night and hence includes lots of noise." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6962" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 558, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 558, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 558, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 558, + 228 + ], + "type": "image", + "image_path": "b28b00a8fa4d581eb8dab8adb212212295062b5100fe9e4ded79d089fd0d483d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 235, + 493, + 247 + ], + "lines": [ + { + "bbox": [ + 116, + 235, + 493, + 247 + ], + "spans": [ + { + "bbox": [ + 116, + 235, + 493, + 247 + ], + "type": "text", + "content": "Figure 4. Comparison of denoising methods on the two DND21 sequences at two different noise levels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 263, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 479 + ], + "type": "text", + "content": "Metrics. The metrics used for denoising assessment are: receiver operating characteristic (ROC) curve, area under the curve (AUC) and Mean Event Structural Ratio (MESR). Given signal and noise GT labels, the ROC shows the True-Positive Rate and False-Positive Rate over several S/N ratios. The AUC of the ROC becomes closer to 1, the better the denoiser works. The MESR is a denoising metric that does not require GT labels [10]. We follow previous work to fix the number of signal events to 30000 [10, 29], as well as " + }, + { + "bbox": [ + 55, + 263, + 296, + 479 + ], + "type": "inline_equation", + "content": "M = 20000" + }, + { + "bbox": [ + 55, + 263, + 296, + 479 + ], + "type": "text", + "content": " [10, Eq.10], to discount the difference in the number of signal events from various denoising methods. The higher, the better. Furthermore, we use the RMS error for angular velocity estimation and the EPE for optical flow estimation, following previous works [23, 25, 42, 50], as well as the FWL metric [52], which is the relative variance (sharpness) of the IWE with respect to that of the identity warp. Further details about the experimental settings (e.g., hyper-parameters) are in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 490, + 162, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 162, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 162, + 502 + ], + "type": "text", + "content": "4.2. Denoising Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 544 + ], + "type": "text", + "content": "We evaluate denoising accuracy using AUC on DND21 data, and using MESR on E-MLB and DND21 data, using the optical flow estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Denoising Performance on E-MLB. Table 1 shows the benchmark results on the E-MLB dataset using the MESR metric. We categorize prior work into model-based methods and learning-based methods for convenience. Compared with other model-based techniques, our method ranks first or second, showcasing the efficacy of the proposed method on denoising. Even compared with learning-based ones, our method achieves competitive scores (e.g., second in E-MLB (Day) ND4-ND64 conditions, etc). Notice that, however, learning-based approaches utilize additional information (GT labels) during training, while our method does not. Nonetheless, the results demonstrate the efficacy of our approach, which contributes to the theoretical foundation of event-denoising methodologies. We discuss further details" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 267, + 550, + 566 + ], + "blocks": [ + { + "bbox": [ + 319, + 267, + 550, + 566 + ], + "lines": [ + { + "bbox": [ + 319, + 267, + 550, + 566 + ], + "spans": [ + { + "bbox": [ + 319, + 267, + 550, + 566 + ], + "type": "image", + "image_path": "55bf99e617c6a07d453f0a17812cdd866fd511e704255aafbd8eb929fd94ebf6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "text", + "content": "Figure 5. Simultaneous estimation of rotational motion and denoising on the dynamicRotation sequence from ECD dataset [40]. Our method produces sharp IWEs " + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "inline_equation", + "content": "(2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "text", + "content": " row) as well as reasonable S/N classification (rows 4 and 5), keeping " + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "text", + "content": " of data " + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "inline_equation", + "content": "(\\tau = 0.9)" + }, + { + "bbox": [ + 313, + 574, + 555, + 619 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 638, + 484, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 484, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 484, + 650 + ], + "type": "text", + "content": "of the evaluation metrics (ESR) in Sec. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": "Qualitative results are shown in Fig. 2. The denoised outputs (\"Signal\") are reasonable, as dark (underexposed) regions of the scene become cleaner, and the IWEs using the estimated flow provide sharp edges of the original scenes. We compare with state-of-the-art methods in detail next." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6963" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 153 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 153 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 153 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 153 + ], + "type": "image", + "image_path": "5374923d01492c439392167a7cae19ee330bb9f7501ad14b7f5b00700343d021.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 156, + 293, + 257 + ], + "blocks": [ + { + "bbox": [ + 58, + 156, + 293, + 257 + ], + "lines": [ + { + "bbox": [ + 58, + 156, + 293, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 156, + 293, + 257 + ], + "type": "image", + "image_path": "bda697145e928ca91ea383e47c245c88b6e5feb26162fe183be61387addcd6d1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 263, + 295, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 308 + ], + "type": "text", + "content": "Figure 6. Results on the ego-motion estimation application (top: boxes, bottom: dynamic) [40]. Our method relaxes the dependency on the initial value for CMax-based rotational motion estimation (indicated as \"Gain\")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 312, + 282, + 426 + ], + "blocks": [ + { + "bbox": [ + 70, + 312, + 282, + 426 + ], + "lines": [ + { + "bbox": [ + 70, + 312, + 282, + 426 + ], + "spans": [ + { + "bbox": [ + 70, + 312, + 282, + 426 + ], + "type": "table", + "html": "
dynamic_rotboxes_rot
RMS ↓FWL ↑RMS ↓FWL ↑
5 Hz NoiseCMax [17]20.0011.259124.6411.129
- w/ Init.8.2751.27420.6591.214
Downsampling (Best)8.8081.273124.8691.108
- w/ Init.8.2261.27420.6191.214
Ours (Best)8.5221.27397.5851.141
- w/ Init.8.1701.27420.6041.214
BAF19.6751.260125.0281.127
- w/ Init.8.2531.27420.5501.214
", + "image_path": "4a5dccc26f531ba451c3ff0b93e157d4d97978c9840c08e7b97846f7fd07289d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 434, + 281, + 445 + ], + "lines": [ + { + "bbox": [ + 69, + 434, + 281, + 445 + ], + "spans": [ + { + "bbox": [ + 69, + 434, + 281, + 445 + ], + "type": "text", + "content": "Table 3. Angular velocity estimation on ECD dataset [40]." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "type": "text", + "content": "Denoising Accuracy on DND21. Table 2 shows the result of AUC for the ROC result on the DND21 dataset. Our method achieves consistently high AUC compared with existing model-based methods. Note that the DND21 dataset aggressively removes events to generate \"pure-signal\" sequences [24, Sec.5]. Due to this step, the labeled signal events may be fewer than the actual ones corresponding to the true motion. Hence, the DND21 sequences could result in a slight degradation in the False Positive score for motion-based (unsupervised) denoising methods like ours, which may result in lower AUC values. Nonetheless, our method provides competitive AUC scores among other state-of-the-art. The ROC curves are shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Qualitative comparisons among several denoising methods is given in Fig. 4. From the results, our method achieves a good trade-off between under-denoising (preserving not only scene-edge details but some noise) and over-denoising (removing not only most of the noise but some edge details). This can be confirmed, for example, in the upper-right edges in the driving example, and the person's arm in the hotel example. Essentially, all denoising methods have" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "some parameters to control the amount of denoising; our proposed method has the signal ratio, which is intuitive; other optimization methods have some parameters about spatio-temporal kernel size, and learning-based approaches use the probability threshold for the network output. While we fix it for each dataset and condition for fair comparison (we do not tune it for each sequence in the results), this is an essential challenge for denoising algorithms in general, regardless of their underlying character (learning-based or model-based). Hence, we believe it is important for research to discuss denoising effectiveness in combination with downstream tasks, as shown in Secs. 4.3 and 4.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 227, + 488, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 488, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 488, + 239 + ], + "type": "text", + "content": "4.3. Joint Motion Estimation Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 244, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 449 + ], + "type": "text", + "content": "Ego-Motion on ECD. One of the merits of the proposed joint estimation method is that it can improve motion estimation results. Here, the results of ego-motion (e.g., angular velocity) estimation are displayed in Fig. 5. We used the classical ECD dataset and injected BA noise from the DND21 recordings with different noise rates. As no prior work simultaneously estimates motion and noise, we compare against sequential filters: random sampling (of the input data with a uniform distribution) and BAF [9] that is the most widely-used denoising method to date. As shown in the IWEs, our method produces the sharpest edges (cf. (b) and (c)) and the most reasonable signal-noise classification among others (cf. (a), (b), and (c)). Note that the original ECD recordings have noise events besides the ones that are injected using DND21 recordings, which makes the calculation of ROC (similar to Sec. 3.2) intricate. Nonetheless, we report ROC results in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "content": "The quantitative evaluation on angular velocity estimation is reported in Fig. 6, as well as the summary of the supplementary detailed table in Tab. 3. As shown in the first row (\"CMax\"), the original CMax degrades due to the noise occurrences. Our method achieves the best results in terms of RMS and FWL, compared with the random sampling method as well as the original CMax (no denoising). Also, using better initialization (the result of the previous slice for sequential estimation) always improves the results (\"w/ Init.\"): the estimation accuracy relies on both the initialization and the denoising strategies. Hence, the result demonstrates that the proposed method relaxes the dependency on the initialization, thus improving the robustness of motion estimation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Optical Flow on MVSEC. The results on more complex motion estimation (e.g., optical flow) are shown in Tab. 4, using the MVSEC dataset [63], with BA noise from the DND21 dataset. Due to the " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " BA noise, the baseline method MultiCM [48] degrades. Our denoising method improves the flow accuracy, showcasing the efficacy of simultaneous noise estimation. A challenge of evaluating denoising in these real-world datasets (e.g., ECD, MVSEC) is the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6964" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 68, + 160, + 125 + ], + "blocks": [ + { + "bbox": [ + 89, + 68, + 160, + 125 + ], + "lines": [ + { + "bbox": [ + 89, + 68, + 160, + 125 + ], + "spans": [ + { + "bbox": [ + 89, + 68, + 160, + 125 + ], + "type": "image", + "image_path": "c3ddb1665febad69ad99d9fc03cf3a5caae3bd7ca7061ef9354807a4dc85e551.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 89, + 125, + 160, + 178 + ], + "blocks": [ + { + "bbox": [ + 89, + 125, + 160, + 178 + ], + "lines": [ + { + "bbox": [ + 89, + 125, + 160, + 178 + ], + "spans": [ + { + "bbox": [ + 89, + 125, + 160, + 178 + ], + "type": "image", + "image_path": "4eb34d949be54c3126c44fb3fa44e6ad6c0c9063c5f338a0d95b341f65e5735b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 160, + 70, + 231, + 125 + ], + "blocks": [ + { + "bbox": [ + 160, + 70, + 231, + 125 + ], + "lines": [ + { + "bbox": [ + 160, + 70, + 231, + 125 + ], + "spans": [ + { + "bbox": [ + 160, + 70, + 231, + 125 + ], + "type": "image", + "image_path": "896daaa895844ecc161cd48140cee724820ebb36597a41a26cd15b2c0477a431.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 125, + 231, + 178 + ], + "blocks": [ + { + "bbox": [ + 160, + 125, + 231, + 178 + ], + "lines": [ + { + "bbox": [ + 160, + 125, + 231, + 178 + ], + "spans": [ + { + "bbox": [ + 160, + 125, + 231, + 178 + ], + "type": "image", + "image_path": "b46025279b1f2cdf8dd2c8e62598906c34f547e2139fa6f6a5c04f2e32a155f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 231, + 70, + 302, + 125 + ], + "blocks": [ + { + "bbox": [ + 231, + 70, + 302, + 125 + ], + "lines": [ + { + "bbox": [ + 231, + 70, + 302, + 125 + ], + "spans": [ + { + "bbox": [ + 231, + 70, + 302, + 125 + ], + "type": "image", + "image_path": "ddb5991c30f64a0eb898abf68d72f10da2a6042ac5a0f4153f94a82b0369a999.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 232, + 125, + 302, + 178 + ], + "blocks": [ + { + "bbox": [ + 232, + 125, + 302, + 178 + ], + "lines": [ + { + "bbox": [ + 232, + 125, + 302, + 178 + ], + "spans": [ + { + "bbox": [ + 232, + 125, + 302, + 178 + ], + "type": "image", + "image_path": "33f90d53462ceac23f2e016430c7ff3cdbb66e2fcd9afb497c3d947a2b3a3d8f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 303, + 70, + 374, + 125 + ], + "blocks": [ + { + "bbox": [ + 303, + 70, + 374, + 125 + ], + "lines": [ + { + "bbox": [ + 303, + 70, + 374, + 125 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 374, + 125 + ], + "type": "image", + "image_path": "e3cc8de1acf621db3440d613e46a69aaccc28e8ae735bd87d415f36719468c8c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 303, + 125, + 374, + 178 + ], + "blocks": [ + { + "bbox": [ + 303, + 125, + 374, + 178 + ], + "lines": [ + { + "bbox": [ + 303, + 125, + 374, + 178 + ], + "spans": [ + { + "bbox": [ + 303, + 125, + 374, + 178 + ], + "type": "image", + "image_path": "fffe399b3e99b647238adf5e4a0ad4c337c66921c44a793c5163144a31b49d57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 375, + 70, + 445, + 125 + ], + "blocks": [ + { + "bbox": [ + 375, + 70, + 445, + 125 + ], + "lines": [ + { + "bbox": [ + 375, + 70, + 445, + 125 + ], + "spans": [ + { + "bbox": [ + 375, + 70, + 445, + 125 + ], + "type": "image", + "image_path": "75d5e2ca95ccb2ab0981ef0a32fa07a7d8ace6b28a2d3ec52ecd4fd3f5ce7cea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 375, + 125, + 445, + 178 + ], + "blocks": [ + { + "bbox": [ + 375, + 125, + 445, + 178 + ], + "lines": [ + { + "bbox": [ + 375, + 125, + 445, + 178 + ], + "spans": [ + { + "bbox": [ + 375, + 125, + 445, + 178 + ], + "type": "image", + "image_path": "00b7f39e62ba0bf310a00dfa3bd7ebd01e5679f5dbcabd611dded2c5c041f6e3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 446, + 70, + 517, + 125 + ], + "blocks": [ + { + "bbox": [ + 446, + 70, + 517, + 125 + ], + "lines": [ + { + "bbox": [ + 446, + 70, + 517, + 125 + ], + "spans": [ + { + "bbox": [ + 446, + 70, + 517, + 125 + ], + "type": "image", + "image_path": "d26dcecf03b28eef0bc8b251694df919ea84d99e6aba98c37901a83773a33939.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 446, + 125, + 517, + 178 + ], + "blocks": [ + { + "bbox": [ + 446, + 125, + 517, + 178 + ], + "lines": [ + { + "bbox": [ + 446, + 125, + 517, + 178 + ], + "spans": [ + { + "bbox": [ + 446, + 125, + 517, + 178 + ], + "type": "image", + "image_path": "7885488cd32a03083a66e1cf95ba1f716d358933996b86e9f408580358a6ea25.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 89, + 180, + 160, + 234 + ], + "blocks": [ + { + "bbox": [ + 89, + 180, + 160, + 234 + ], + "lines": [ + { + "bbox": [ + 89, + 180, + 160, + 234 + ], + "spans": [ + { + "bbox": [ + 89, + 180, + 160, + 234 + ], + "type": "image", + "image_path": "f771427a458bdd8604627c5fd8b1be9dd97cc27df2cfd4a22a91226f998f7604.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 89, + 234, + 160, + 289 + ], + "blocks": [ + { + "bbox": [ + 89, + 234, + 160, + 289 + ], + "lines": [ + { + "bbox": [ + 89, + 234, + 160, + 289 + ], + "spans": [ + { + "bbox": [ + 89, + 234, + 160, + 289 + ], + "type": "image", + "image_path": "80c6d0b20a4fef589ed86c26a8418ca09c638715bdb74daa1f90679db93e9aeb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 290, + 150, + 300 + ], + "lines": [ + { + "bbox": [ + 99, + 290, + 150, + 300 + ], + "spans": [ + { + "bbox": [ + 99, + 290, + 150, + 300 + ], + "type": "text", + "content": "(a) Input events" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 160, + 180, + 231, + 234 + ], + "blocks": [ + { + "bbox": [ + 160, + 180, + 231, + 234 + ], + "lines": [ + { + "bbox": [ + 160, + 180, + 231, + 234 + ], + "spans": [ + { + "bbox": [ + 160, + 180, + 231, + 234 + ], + "type": "image", + "image_path": "e0ab63322f5ea2373496e5ae024f33d8b9df7338b0099bf0d020e9da1a61f1d3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 160, + 234, + 232, + 289 + ], + "blocks": [ + { + "bbox": [ + 160, + 234, + 232, + 289 + ], + "lines": [ + { + "bbox": [ + 160, + 234, + 232, + 289 + ], + "spans": [ + { + "bbox": [ + 160, + 234, + 232, + 289 + ], + "type": "image", + "image_path": "0b519167a76ca773bf64bc36bfde8c06b952a4df25a3c5c20b785ed0e99c02d9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 290, + 220, + 300 + ], + "lines": [ + { + "bbox": [ + 171, + 290, + 220, + 300 + ], + "spans": [ + { + "bbox": [ + 171, + 290, + 220, + 300 + ], + "type": "text", + "content": "(b) Estim. flow" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 232, + 180, + 303, + 234 + ], + "blocks": [ + { + "bbox": [ + 232, + 180, + 303, + 234 + ], + "lines": [ + { + "bbox": [ + 232, + 180, + 303, + 234 + ], + "spans": [ + { + "bbox": [ + 232, + 180, + 303, + 234 + ], + "type": "image", + "image_path": "611881d6462c94893c3e2b305d25417b5e6648cb8e347cec23773c9b44cab91a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 232, + 234, + 302, + 289 + ], + "blocks": [ + { + "bbox": [ + 232, + 234, + 302, + 289 + ], + "lines": [ + { + "bbox": [ + 232, + 234, + 302, + 289 + ], + "spans": [ + { + "bbox": [ + 232, + 234, + 302, + 289 + ], + "type": "image", + "image_path": "31a3b6c530eb861e92ed02ea78450de367460d20d0df378eeb3b11bc8a727acb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 290, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 239, + 290, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 239, + 290, + 294, + 300 + ], + "type": "text", + "content": "(c) Signal events" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 303, + 180, + 374, + 234 + ], + "blocks": [ + { + "bbox": [ + 303, + 180, + 374, + 234 + ], + "lines": [ + { + "bbox": [ + 303, + 180, + 374, + 234 + ], + "spans": [ + { + "bbox": [ + 303, + 180, + 374, + 234 + ], + "type": "image", + "image_path": "5459b703e831b4ac8d0a30e1024fd5c641d7bcad17c61c87ece00e3338ab5ba3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 303, + 234, + 374, + 289 + ], + "blocks": [ + { + "bbox": [ + 303, + 234, + 374, + 289 + ], + "lines": [ + { + "bbox": [ + 303, + 234, + 374, + 289 + ], + "spans": [ + { + "bbox": [ + 303, + 234, + 374, + 289 + ], + "type": "image", + "image_path": "c6d44af299743e42c52aae9aca358a335deb0ba6b93ac90fe73178bc4e9b6626.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 290, + 363, + 310 + ], + "lines": [ + { + "bbox": [ + 315, + 290, + 363, + 310 + ], + "spans": [ + { + "bbox": [ + 315, + 290, + 363, + 310 + ], + "type": "text", + "content": "(d) Reconstr. from (c) signal" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 375, + 180, + 445, + 234 + ], + "blocks": [ + { + "bbox": [ + 375, + 180, + 445, + 234 + ], + "lines": [ + { + "bbox": [ + 375, + 180, + 445, + 234 + ], + "spans": [ + { + "bbox": [ + 375, + 180, + 445, + 234 + ], + "type": "image", + "image_path": "1dff928675a01ad4107b3541d35445e31584b0eb2c086492dfd98924a6849a35.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 375, + 234, + 445, + 289 + ], + "blocks": [ + { + "bbox": [ + 375, + 234, + 445, + 289 + ], + "lines": [ + { + "bbox": [ + 375, + 234, + 445, + 289 + ], + "spans": [ + { + "bbox": [ + 375, + 234, + 445, + 289 + ], + "type": "image", + "image_path": "b8b10ecc5da0cd0d858c27c1dbabd4bb3ce6f2189078c3ba480d41719134f08c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 290, + 433, + 310 + ], + "lines": [ + { + "bbox": [ + 387, + 290, + 433, + 310 + ], + "spans": [ + { + "bbox": [ + 387, + 290, + 433, + 310 + ], + "type": "text", + "content": "(e) Reconstr. from (a) input" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 446, + 180, + 517, + 234 + ], + "blocks": [ + { + "bbox": [ + 446, + 180, + 517, + 234 + ], + "lines": [ + { + "bbox": [ + 446, + 180, + 517, + 234 + ], + "spans": [ + { + "bbox": [ + 446, + 180, + 517, + 234 + ], + "type": "image", + "image_path": "7840d8e7465f76d67efbd4d4a9522c4dec2a557eda0bf09bc091ea9f5ffc36e0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 446, + 234, + 517, + 289 + ], + "blocks": [ + { + "bbox": [ + 446, + 234, + 517, + 289 + ], + "lines": [ + { + "bbox": [ + 446, + 234, + 517, + 289 + ], + "spans": [ + { + "bbox": [ + 446, + 234, + 517, + 289 + ], + "type": "image", + "image_path": "fb877060ab676d280305608a8d7686e52c3100edb8370d9ff8c7baf50b5b7052.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 290, + 514, + 299 + ], + "lines": [ + { + "bbox": [ + 449, + 290, + 514, + 299 + ], + "spans": [ + { + "bbox": [ + 449, + 290, + 514, + 299 + ], + "type": "text", + "content": "(f) Reference frame" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 93, + 346, + 258, + 445 + ], + "blocks": [ + { + "bbox": [ + 93, + 346, + 258, + 445 + ], + "lines": [ + { + "bbox": [ + 93, + 346, + 258, + 445 + ], + "spans": [ + { + "bbox": [ + 93, + 346, + 258, + 445 + ], + "type": "table", + "html": "
indoor1indoor2
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.67632.0982.74633.474
Ours2.51729.4752.69533.047
indoor3outdoor1
EPE ↓%Out ↓EPE ↓%Out ↓
MultiCM [48]2.73233.2191.96018.160
Ours2.65432.2411.86517.554
", + "image_path": "db3b0a3fc3e197fea688f070fae6e0a02194d75df0ae7749174af2b56c840840.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "text", + "content": "Table 4. Optical flow estimation results on MVSEC [63] (" + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "inline_equation", + "content": "dt = 4" + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "text", + "content": " frames, i.e., " + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "inline_equation", + "content": "89 \\mathrm{~ms}" + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "text", + "content": ") with BA noise at " + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "inline_equation", + "content": "5 \\mathrm{~Hz}" + }, + { + "bbox": [ + 55, + 453, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 515 + ], + "type": "text", + "content": "existing noise in the original sequences. Hence, we only evaluate motion estimation metrics, not ROC or AUC." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 525, + 263, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 263, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 263, + 539 + ], + "type": "text", + "content": "4.4. Application to Intensity Reconstruction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 676 + ], + "type": "text", + "content": "In real-world use cases, denoising works as preprocessing steps for other tasks (e.g., deblur, 3D reconstruction, motion estimation, etc.), hence it is important for denoising methods to demonstrate applicability. Although the proposed approach intrinsically improves motion estimation (Sec. 4.3), we further evaluate the denoising outcome via intensity reconstruction, a well-known task in event-based vision, using the E-MLB dataset [10]. Figure 7 collects the results. We use two state-of-the-art approaches for reconstruction: E2VID [46], which does not rely on motion, and EVILIP [59], which utilizes optical flow as prior." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Our method provides flow (Fig. 7, column (b)) and signal events (sharp IWEs in column (c)). Comparing the reconstruction with our event denoising method (column (d)) and" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 320, + 354, + 473, + 410 + ], + "blocks": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 555, + 338 + ], + "type": "text", + "content": "Figure 7. Image intensity reconstruction on E-MLB data [10]. The first two rows use the reconstruction method EVILIP [59] and the last two rows use E2VID [46]. Reconstruction on denoised events produces higher quality images (i.e., fewer artifacts) than on raw event data." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 354, + 473, + 410 + ], + "lines": [ + { + "bbox": [ + 320, + 354, + 473, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 354, + 473, + 410 + ], + "type": "image", + "image_path": "2942f1d5329e90790c245bb3b0ffa43e4a385cb733d3fe774bf00e468932af3c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 320, + 410, + 473, + 465 + ], + "blocks": [ + { + "bbox": [ + 320, + 410, + 473, + 465 + ], + "lines": [ + { + "bbox": [ + 320, + 410, + 473, + 465 + ], + "spans": [ + { + "bbox": [ + 320, + 410, + 473, + 465 + ], + "type": "image", + "image_path": "4d848c24e69a6b0310d66bc0ab07dafcecc1b0f870f12b08f4bdcf9ab58ac0da.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 320, + 465, + 473, + 521 + ], + "blocks": [ + { + "bbox": [ + 320, + 465, + 473, + 521 + ], + "lines": [ + { + "bbox": [ + 320, + 465, + 473, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 465, + 473, + 521 + ], + "type": "image", + "image_path": "8e3f08c4be859530b52f646523d3008f77d225d5566f79a73cfaf72886384212.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 522, + 380, + 531 + ], + "lines": [ + { + "bbox": [ + 351, + 522, + 380, + 531 + ], + "spans": [ + { + "bbox": [ + 351, + 522, + 380, + 531 + ], + "type": "text", + "content": "(a) Signal" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 424, + 522, + 452, + 530 + ], + "lines": [ + { + "bbox": [ + 424, + 522, + 452, + 530 + ], + "spans": [ + { + "bbox": [ + 424, + 522, + 452, + 530 + ], + "type": "text", + "content": "(b) Noise" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 474, + 355, + 547, + 411 + ], + "blocks": [ + { + "bbox": [ + 474, + 355, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 474, + 355, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 474, + 355, + 547, + 411 + ], + "type": "image", + "image_path": "e47a2e47b6178fc312284ab664e59473cb3a6bc1f269840816557a453b40025c.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 474, + 411, + 547, + 467 + ], + "blocks": [ + { + "bbox": [ + 474, + 411, + 547, + 467 + ], + "lines": [ + { + "bbox": [ + 474, + 411, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 474, + 411, + 547, + 467 + ], + "type": "image", + "image_path": "a93237ea407609e8ed6854f97b1b8b2e1dd8533b98211bd71b2ed4b3888904a4.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 474, + 467, + 547, + 521 + ], + "blocks": [ + { + "bbox": [ + 474, + 467, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 474, + 467, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 474, + 467, + 547, + 521 + ], + "type": "image", + "image_path": "1090b5d162cf6e76d1326c8bf1468a8da31b27f27afbdba7f2e28b53f23105e7.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 522, + 537, + 530 + ], + "lines": [ + { + "bbox": [ + 484, + 522, + 537, + 530 + ], + "spans": [ + { + "bbox": [ + 484, + 522, + 537, + 530 + ], + "type": "text", + "content": "(c) Reconstruction" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 537, + 555, + 560 + ], + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 560 + ], + "type": "text", + "content": "Figure 8. Sensitivity analysis: effect of different target ratio parameters " + }, + { + "bbox": [ + 313, + 537, + 555, + 560 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 537, + 555, + 560 + ], + "type": "text", + "content": " on image reconstruction using EVILIP [59]." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "the reconstruction from the raw event data (column (e)), the images after event denoising have fewer artifacts than those without denoising, for both reconstruction methods tested. We observe that E2VID is more robust against noise than EVILIP, possibly due to noise events in its training (simulation) data. Due to the ND filters, the quality of the reference frames (last column) is limited, and hence, we discuss the results qualitatively. Nevertheless comparing the reconstructed images to the reference frames, evidences the HDR advantages of event cameras over frame-based ones." + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6965" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 136, + 125 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 136, + 125 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 136, + 125 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 136, + 125 + ], + "type": "image", + "image_path": "8114e778fee861c5ec5b676583eb98ef667d0c2040669fad9a5bf24702675f5c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 126, + 121, + 134 + ], + "lines": [ + { + "bbox": [ + 78, + 126, + 121, + 134 + ], + "spans": [ + { + "bbox": [ + 78, + 126, + 121, + 134 + ], + "type": "text", + "content": "(a) Ref. events" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 70, + 211, + 125 + ], + "blocks": [ + { + "bbox": [ + 138, + 70, + 211, + 125 + ], + "lines": [ + { + "bbox": [ + 138, + 70, + 211, + 125 + ], + "spans": [ + { + "bbox": [ + 138, + 70, + 211, + 125 + ], + "type": "image", + "image_path": "dad65955ddb772fddc961fde6e2fae628a52c0fa9bd7e7ea4a8e82ad65d9c5b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 127, + 194, + 134 + ], + "lines": [ + { + "bbox": [ + 156, + 127, + 194, + 134 + ], + "spans": [ + { + "bbox": [ + 156, + 127, + 194, + 134 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 156, + 127, + 194, + 134 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 156, + 127, + 194, + 134 + ], + "type": "text", + "content": " px" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "text", + "content": "Figure 9. Results of different Gaussian kernel sizes " + }, + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 138, + 295, + 160 + ], + "type": "text", + "content": " for sensitivity on different depth scenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 287, + 125 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 287, + 125 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 287, + 125 + ], + "type": "image", + "image_path": "382280d018ace21f9cf8d0e6647a8011ff7758662371f1eda87bbce7c5c783b0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 127, + 269, + 134 + ], + "lines": [ + { + "bbox": [ + 231, + 127, + 269, + 134 + ], + "spans": [ + { + "bbox": [ + 231, + 127, + 269, + 134 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 231, + 127, + 269, + 134 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 231, + 127, + 269, + 134 + ], + "type": "text", + "content": " px" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 175, + 127, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 127, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 127, + 189 + ], + "type": "text", + "content": "4.5. Sensitivity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "text", + "content": "Effect of Noise Ratio. As Sec. 4.2 already showed the sensitivity for denoising accuracy, let us further discuss the sensitivity using image reconstruction. Figure 8 presents the result of the Bicycle sequence with ND64 filter [10]. While the GT image is highly underexposed and not informative (as those on the last column of Fig. 7), the reconstructed images clearly reflect the choice of the signal target ratio " + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "text", + "content": ". As more events are removed, " + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "text", + "content": " the reconstructed image becomes more homogeneous but preserves strong edges, and " + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 55, + 195, + 296, + 339 + ], + "type": "text", + "content": " the removed events (2nd col.) contain more scene edges (i.e., signal). We quantitatively analyze the former in the supplementary using non-reference image quality metrics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "type": "text", + "content": "Effect of the Kernel Size. Figure 9 shows the sensitivity analysis for different kernel sizes. By increasing the size of the Gaussian kernel in the IWE, our method can emphasize contour edges over isolated points, thus better preserving signal events in regions with low IWE intensities (center of the image, corresponding to far away scene points)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 424, + 217, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 217, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 217, + 435 + ], + "type": "text", + "content": "4.6. Ablation of Motion Estimator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 598 + ], + "type": "text", + "content": "The proposed approach can be combined with other estimation methods, such as deep-neural-network (DNN) flow estimators. To this end, we show an example of E-RAFT [21]. The model-based method alone [48] is known to output a degraded flow in night sequences with lots of noise, such as zurich_city_12a [20]. As shown in Fig. 10, our method can identify and discard noisy events. E-RAFT provides sharper and more natural edges in IWEs after denoising compared to MultiCM [48] (thanks to its more accurate flow). ESR scores are similar: 2.39 for E-RAFT and 2.45 for MultiCM. While we focus on the inference of DNN in the proposed pipeline, future research could look into accommodating training within the proposed joint estimation framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 610, + 119, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 119, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 119, + 622 + ], + "type": "text", + "content": "4.7. Runtime" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Using 30k events from the ECD dataset " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "(240 \\times 180\\mathrm{px})" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " on a Mac M1 CPU (2020), the denoising operations (scoring and ranking) take less than " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.1\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " extra computation in total, out of " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "2\\mathrm{s}" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " of the entire processing time (i.e., " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " computation increase with respect to the base method). Nonetheless, accommodating learning-based approaches in our framework could be key to the trade-off between accuracy and speed." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 328, + 77, + 402, + 133 + ], + "blocks": [ + { + "bbox": [ + 328, + 77, + 402, + 133 + ], + "lines": [ + { + "bbox": [ + 328, + 77, + 402, + 133 + ], + "spans": [ + { + "bbox": [ + 328, + 77, + 402, + 133 + ], + "type": "image", + "image_path": "5b6abb16bd161e709d7bd0ac6292b7f59ef6f681827178c5c494a392c0ce2ae2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 134, + 381, + 141 + ], + "lines": [ + { + "bbox": [ + 350, + 134, + 381, + 141 + ], + "spans": [ + { + "bbox": [ + 350, + 134, + 381, + 141 + ], + "type": "text", + "content": "Raw Event" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 402, + 78, + 474, + 133 + ], + "blocks": [ + { + "bbox": [ + 402, + 78, + 474, + 133 + ], + "lines": [ + { + "bbox": [ + 402, + 78, + 474, + 133 + ], + "spans": [ + { + "bbox": [ + 402, + 78, + 474, + 133 + ], + "type": "image", + "image_path": "06d32f0c215e5bc8e90658d248159d2e5d05798bdac2631a711e2e7c202f3ec9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 134, + 462, + 141 + ], + "lines": [ + { + "bbox": [ + 415, + 134, + 462, + 141 + ], + "spans": [ + { + "bbox": [ + 415, + 134, + 462, + 141 + ], + "type": "text", + "content": "IWE (No Warp)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 320, + 142, + 402, + 196 + ], + "blocks": [ + { + "bbox": [ + 320, + 142, + 402, + 196 + ], + "lines": [ + { + "bbox": [ + 320, + 142, + 402, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 142, + 402, + 196 + ], + "type": "image", + "image_path": "2a4bba47789dd768ef59eda58d7f2d069a6a397fa374df9858dbbdfab1a7a1eb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 320, + 196, + 402, + 254 + ], + "blocks": [ + { + "bbox": [ + 320, + 196, + 402, + 254 + ], + "lines": [ + { + "bbox": [ + 320, + 196, + 402, + 254 + ], + "spans": [ + { + "bbox": [ + 320, + 196, + 402, + 254 + ], + "type": "image", + "image_path": "55b6d476952e74d25445e291eb3240e7f889093496c1f4a41c9b74846768e6f6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 255, + 390, + 263 + ], + "lines": [ + { + "bbox": [ + 342, + 255, + 390, + 263 + ], + "spans": [ + { + "bbox": [ + 342, + 255, + 390, + 263 + ], + "type": "text", + "content": "(a) Signal Events" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 270, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 314 + ], + "type": "text", + "content": "Figure 10. Effect of adopting two different motion estimators (E-RAFT or MultiCM) on the DSEC night scene (zurich_12a) [20]. Despite small IWE differences (e.g., in the car's hood), the proposed framework supports several motion estimators." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 403, + 143, + 474, + 198 + ], + "blocks": [ + { + "bbox": [ + 403, + 143, + 474, + 198 + ], + "lines": [ + { + "bbox": [ + 403, + 143, + 474, + 198 + ], + "spans": [ + { + "bbox": [ + 403, + 143, + 474, + 198 + ], + "type": "image", + "image_path": "adc7580ce076347908892112af1c604d5112ed7fc73d24fbecbc84f0ffa1d0b8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 403, + 198, + 474, + 254 + ], + "blocks": [ + { + "bbox": [ + 403, + 198, + 474, + 254 + ], + "lines": [ + { + "bbox": [ + 403, + 198, + 474, + 254 + ], + "spans": [ + { + "bbox": [ + 403, + 198, + 474, + 254 + ], + "type": "image", + "image_path": "99e035bdcbd5ae0e10cd26b64bd30e41b443b1529dc6dbfb309a5ec5b8de37e4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 255, + 450, + 262 + ], + "lines": [ + { + "bbox": [ + 427, + 255, + 450, + 262 + ], + "spans": [ + { + "bbox": [ + 427, + 255, + 450, + 262 + ], + "type": "text", + "content": "(b) IWE" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 476, + 143, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 476, + 143, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 476, + 143, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 476, + 143, + 547, + 198 + ], + "type": "image", + "image_path": "53fe5b6e7247e26af7ed36683d7ad1ac96975aca3bca3e49aff6dee3f50caba0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 477, + 199, + 547, + 254 + ], + "blocks": [ + { + "bbox": [ + 477, + 199, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 477, + 199, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 477, + 199, + 547, + 254 + ], + "type": "image", + "image_path": "4dd08aff9896dbb3b06d5cb095a14ee56deb45039f14ace0ac1ffb0f718fe5b9.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 255, + 539, + 262 + ], + "lines": [ + { + "bbox": [ + 482, + 255, + 539, + 262 + ], + "spans": [ + { + "bbox": [ + 482, + 255, + 539, + 262 + ], + "type": "text", + "content": "(c) Flow Estimation" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 324, + 389, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 389, + 337 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 389, + 337 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 345, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 476 + ], + "type": "text", + "content": "Our method expects as parameter a target noise level, but the intrinsic (\"true\") noise level depends on the camera, scene, etc. and is therefore unknown. More research in this direction is needed for automatic parameter tuning. Also, notice that there might be a potential limitation of the denoising metric ESR, since it depends on the motion parameter (see Suppl. Mat. of [29]). Although our denoising efficacy is confirmed from multiple perspectives, we believe that advancing denoising metrics for real-world data (where it gets interesting due to the absence of GT labels) is a central direction for future work in event-based vision." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 388, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In summary, this work proposes that: (i) Joint estimation of noise and motion is a novel problem setting compared to conventional (isolated) denoising, and is sensitive given the event data's inherent motion dependency. (ii) The proposed joint estimation framework can accommodate various motion estimators and has shown its efficacy in improving both denoising and motion estimation accuracies. (iii) It is important to evaluate denoising results on both, denoising benchmarks and downstream tasks (e.g., motion estimation, intensity reconstruction), since current denoising benchmarks have shortcomings in terms of metrics, data, and baselines. In particular, our method achieves notable denoising results (state-of-the-art on the E-MLB benchmark and competitive on DND21) while demonstrating effectiveness across motion estimation and intensity reconstruction tasks. Our approach advances both event-denoising theory and practical denoising use-cases." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6966" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "content": "We would like to thank Dr. B. Jiang for useful discussions. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy - EXC 2002/1 \"Science of Intelligence\" - project number 390523135." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 180, + 294, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 294, + 224 + ], + "type": "text", + "content": "[1] Saeed Afshar, Andrew Peter Nicholson, Andre van Schaik, and Gregory Cohen. Event-based object detection and tracking for space situational awareness. IEEE Sensors Journal, 20(24):15117-15132, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 280 + ], + "type": "text", + "content": "[2] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Trans. Neural Netw. Learn. Syst., 35(3):4110-4124, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 296, + 335 + ], + "type": "text", + "content": "[3] Sami Arja, Alexandre Marcireau, Richard L. Balthazor, Matthew G. McHarg, Saeed Afshar, and Gregory Cohen. Density invariant contrast maximization for neuromorphic earth observations. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 3984-3994, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 336, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 336, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 336, + 296, + 392 + ], + "type": "text", + "content": "[4] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1701-1710, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 458 + ], + "type": "text", + "content": "[5] R Wes Baldwin, Mohammed Almatrafi, Jason R Kaufman, Vijayan Asari, and Keigo Hirakawa. Inceptive event timesurfaces for object classification using neuromorphic cameras. In Image Analysis and Recognition: 16th Int. Conf., ICIAR 2019, Waterloo, ON, Canada, August 27-29, 2019, Proceedings, Part II 16, pages 395-403. Springer, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 479 + ], + "type": "text", + "content": "[6] T. D. Barfoot. State Estimation for Robotics - A Matrix Lie Group Approach. Cambridge University Press, 2015." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": "[7] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A 240x180 130dB " + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE J. Solid-State Circuits, 49(10):2333-2341, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 568 + ], + "type": "text", + "content": "[8] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In AAAI Conf. Artificial Intell., pages 303-311, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 569, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 591 + ], + "type": "text", + "content": "[9] Tobi Delbruck. Frame-free dynamic digital vision. In Proc. Int. Symp. Secure-Life Electron., pages 21-26, 2008." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "text", + "content": "[10] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-MLB: Multilevel benchmark for event-based camera denoising. IEEE Trans. Multimedia, 26:65–76, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "text", + "content": "[11] Peiqi Duan, Zihao Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Trans. Pattern Anal. Mach. Intell., pages 1-1, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[12] Peiqi Duan, Zihao W. Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. EventZoom: Learning to denoise and super resolve" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "neuromorphic events. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12824-12833, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "type": "text", + "content": "[13] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 153, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 554, + 196 + ], + "type": "text", + "content": "[14] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 554, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 232 + ], + "type": "text", + "content": "[15] Guillermo Gallego and Davide Scaramuzza. Accurate angular velocity estimation with an event camera. IEEE Robot. Autom. Lett., 2(2):632-639, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 234, + 554, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 554, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 554, + 267 + ], + "type": "text", + "content": "[16] Guillermo Gallego and Anthony Yezzi. A compact formula for the derivative of a 3-D rotation in exponential coordinates. J. Math. Imaging Vis., 51(3):378-384, 2014." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 269, + 554, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 554, + 323 + ], + "type": "text", + "content": "[17] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3867-3876, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 554, + 369 + ], + "type": "text", + "content": "[18] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus is all you need: Loss functions for event-based vision. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 12272-12281, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 371, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 426 + ], + "type": "text", + "content": "[19] Guillermo Gallego, Tobi Delbruck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew Davison, Jörg Conradt, Kostas Daniilidis, and Davide Scaramuzza. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154-180, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 428, + 554, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 554, + 471 + ], + "type": "text", + "content": "[20] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robot. Autom. Lett., 6(3):4947-4954, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 473, + 554, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 554, + 517 + ], + "type": "text", + "content": "[21] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In Int. Conf. 3D Vision (3DV), pages 197–206, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "type": "text", + "content": "[22] Rui Graca and Tobi Delbruck. Towards a physically realistic computationally efficient dvs pixel model. In Int. Image Sensor Workshop (IISW), 2025." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 554, + 554, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 608 + ], + "type": "text", + "content": "[23] Cheng Gu, Erik Learned-Miller, Daniel Sheldon, Guillermo Gallego, and Pia Bideau. The spatio-temporal Poisson point process: A simple model for the alignment of event camera data. In Int. Conf. Comput. Vis. (ICCV), pages 13495-13504, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 554, + 643 + ], + "type": "text", + "content": "[24] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Trans. Pattern Anal. Mach. Intell., 45(1):785-795, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "text", + "content": "[25] Shuang Guo and Guillermo Gallego. CMax-SLAM: Event-based rotational-motion bundle adjustment and SLAM system using contrast maximization. IEEE Trans. Robot., 40: 2442-2461, 2024." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[26] Shasha Guo, Ziyang Kang, Lei Wang, Shiming Li, and Weixia Xu. Hashheat: An O(C) complexity hashing-based" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6967" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 105 + ], + "type": "text", + "content": "filter for dynamic vision sensor. In 25th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 452-457, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[27] Shuang Guo, Friedhelm Hamann, and Guillermo Gallego. Unsupervised joint learning of optical flow and intensity with event cameras. In Int. Conf. Comput. Vis. (ICCV), 2025." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 193 + ], + "type": "text", + "content": "[28] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In Eur. Conf. Comput. Vis. (ECCV), pages 18-37, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "text", + "content": "[29] Bin Jiang, bo Xiong, Bohan Qu, M. Slaman Asif, you Zhou, and Zhan Ma. EDformer: Transformer-based event denoising across varied noise levels. In Eur. Conf. Comput. Vis. (ECCV), pages 1-12, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 281 + ], + "type": "text", + "content": "[30] Pritam P. Karmokar, Quan H. Nguyen, and William J. Beksi. Secrets of edge-informed contrast maximization for event-based vision. In IEEE Winter Conf. Appl. Comput. Vis. (WACV), 2025." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 325 + ], + "type": "text", + "content": "[31] Alireza Khodamoradi and Ryan Kastner. O(N)-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Trans. Emerg. Topics Comput., PP(99):1-1, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[32] Haram Kim and H. Jin Kim. Real-time rotational motion estimation with contrast maximization over globally aligned events. IEEE Robot. Autom. Lett., 6(3):6016-6023, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[33] Simon Klenk, Jason Chui, Nikolaus Demmel, and Daniel Cremers. TUM-VIE: The TUM stereo visual-inertial event dataset. In IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS), pages 8601-8608, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[34] Xavier Lagorce, Garrick Orchard, Francesco Gallupi, Bertram E. Shi, and Ryad Benosman. HOTS: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(7):1346-1359, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[35] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A 128x128 120dB 30mW asynchronous vision sensor that responds to relative intensity change. In IEEE Int. Solid-State Circuits Conf. (ISSCC), pages 2060-2069, 2006." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "text", + "content": "[36] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "text", + "content": " dB 15 μs latency asynchronous temporal contrast vision sensor. IEEE J. Solid-State Circuits, 43(2):566-576, 2008." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 536, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 581 + ], + "type": "text", + "content": "[37] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In IEEE Int. Symp. Circuits Syst. (ISCAS), pages 722-725, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[38] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Trans. Image Process., 21(12):4695-4708, 2012." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "text", + "content": "[39] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Process. Lett., 20(3):209-212, 2012." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and SLAM. Int. J. Robot. Research, 36(2):142-149, 2017." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "text", + "content": "[41] Elias Mueggler, Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. Continuous-time visual-inertial odometry for event cameras. IEEE Trans. Robot., 34(6):1425-1440, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 160 + ], + "type": "text", + "content": "[42] Urbano Miguel Nunes and Yiannis Demiris. Robust event-based vision model estimation by dispersion minimisation. IEEE Trans. Pattern Anal. Mach. Intell., 44(12):9561-9573, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 315, + 162, + 553, + 215 + ], + "type": "text", + "content": "[43] Federico Paredes-Valles and Guido C. H. E. de Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3445–3454, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 553, + 270 + ], + "type": "text", + "content": "[44] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE de Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Int. Conf. Comput. Vis. (ICCV), pages 9661-9671, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 271, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 271, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 553, + 314 + ], + "type": "text", + "content": "[45] Xin Peng, Ling Gao, Yifu Wang, and Laurent Kneip. Globally-optimal contrast maximisation for event cameras. IEEE Trans. Pattern Anal. Mach. Intell., 44(7):3479-3495, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 553, + 358 + ], + "type": "text", + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE Trans. Pattern Anal. Mach. Intell., 43 (6):1964-1980, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 360, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 360, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 315, + 360, + 553, + 414 + ], + "type": "text", + "content": "[47] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high speed scenarios. IEEE Robot. Autom. Lett., 3(2): 994-1001, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 553, + 448 + ], + "type": "text", + "content": "[48] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow. In Eur. Conf. Comput. Vis. (ECCV), pages 628-645, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 553, + 480 + ], + "type": "text", + "content": "[49] Shintaro Shiba, Yoshimitsu Aoki, and Guillermo Gallego. Event collapse in contrast maximization frameworks. Sensors, 22(14):1-20, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 315, + 482, + 553, + 525 + ], + "type": "text", + "content": "[50] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth, and ego-motion by contrast maximization. IEEE Trans. Pattern Anal. Mach. Intell., 46(12):7742-7759, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 315, + 526, + 553, + 569 + ], + "type": "text", + "content": "[51] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Int. Conf. Comput. Vis. (ICCV), pages 7243-7252, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 571, + 553, + 624 + ], + "type": "text", + "content": "[52] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In Eur. Conf. Comput. Vis. (ECCV), pages 534-549, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 553, + 679 + ], + "type": "text", + "content": "[53] Gemma Taverni, Diederik Paul Moeys, Chenghan Li, Celso Cavaco, Vasyl Motsnyi, David San Segundo Bello, and Tobi Delbruck. Front and back illuminated Dynamic and Active Pixel Vision Sensors comparison. IEEE Trans. Circuits Syst. II (TCSII), 65(5):677-681, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 681, + 553, + 713 + ], + "type": "text", + "content": "[54] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6968" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 565 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 6358-6367, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[55] Zihao W. Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1606–1616, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[56] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Trans. Multimedia, 23:1148-1159, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[57] Ryo Yamaki, Shintaro Shiba, Gallego Guillermo, and Yoshimitsu Aoki. Iterative event-based motion segmentation by variational contrast maximization. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPRW), pages 4918-4927, 2025." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 295 + ], + "type": "text", + "content": "[58] Pei Zhang, Haosen Liu, Zhou Ge, Chutian Wang, and Edmund Y Lam. Neuromorphic imaging with joint image deblurring and event denoising. IEEE Trans. Image Process., 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[59] Zelin Zhang, Anthony Yezzi, and Guillermo Gallego. Formulating event-based image reconstruction as a linear inverse problem with deep regularization using optical flow. IEEE Trans. Pattern Anal. Mach. Intell., 45(7):8372-8389, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "text", + "content": "[60] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Trans. Neural Netw. Learn. Syst., pages 1-13, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 430 + ], + "type": "text", + "content": "[61] Yi Zhou, Guillermo Gallego, and Shaojie Shen. Event-based stereo visual odometry. IEEE Trans. Robot., 37(5):1433-1450, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 464 + ], + "type": "text", + "content": "[62] Alex Zihao Zhu, Nikolay Atanasov, and Kostas Daniilidis. Event-based visual inertial odometry. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 5816-5824, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "text", + "content": "[63] Alex Zihao Zhu, Dinesh Thakur, Tolga Ozaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multivehicle stereo event camera dataset: An event camera dataset for 3D perception. IEEE Robot. Autom. Lett., 3(3):2032-2039, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 522, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 565 + ], + "type": "text", + "content": "[64] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 989-997, 2019." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6969" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_content_list.json b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4810598bb3498901786a324125b0ee660cbd677a --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_content_list.json @@ -0,0 +1,2660 @@ +[ + { + "type": "text", + "text": "Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras", + "text_level": 1, + "bbox": [ + 145, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Petr Hruby", + "bbox": [ + 271, + 181, + 361, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ETH Zürich", + "bbox": [ + 267, + 199, + 366, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rämistrasse 101, 8006 Zürich", + "bbox": [ + 197, + 215, + 437, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "petr.h Ruby@inf.ethz.ch", + "bbox": [ + 217, + 237, + 415, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marc Pollefeys", + "bbox": [ + 581, + 181, + 704, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ETH Zürich / Microsoft Spatial AI Lab", + "bbox": [ + 486, + 199, + 799, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rämistrasse 101, 8006 Zürich", + "bbox": [ + 524, + 217, + 763, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "marc.pollefeys@inf.ethz.ch", + "bbox": [ + 524, + 237, + 756, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel approach for estimating the relative pose between rolling shutter cameras using the intersections of line projections with a single scanline per image. This allows pose estimation without explicitly modeling camera motion. Alternatively, scanlines can be selected within a single image, enabling single-view relative pose estimation for scanlines of rolling shutter cameras. Our approach is designed as a foundational building block for rolling shutter structure-from-motion (SfM), where no motion model is required, and each scanline's pose can be computed independently. We classify minimal solvers for this problem in both generic and specialized settings, including cases with parallel lines and known gravity direction, assuming known intrinsic and no lens distortion. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors, by leveraging connections between this problem and the estimation of 2D structure from 1D cameras. Experiments on rolling shutter images from the Fastec dataset demonstrate the feasibility of our approach for initializing rolling shutter SfM, highlighting its potential for further development. The code will be made publicly available.", + "bbox": [ + 89, + 316, + 483, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 679, + 220, + 695 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Relative pose estimation is a fundamental problem in computer vision with applications in Structure from Motion [61], SLAM, multi-view stereo, and visual odometry. This paper focuses on estimating relative pose between rolling shutter (RS) cameras.", + "bbox": [ + 89, + 703, + 482, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rolling shutter cameras capture images row by row. If the camera is static during capture, the resulting image is equivalent to that of a pinhole camera. However, if the camera moves, the rolling shutter effect distorts the image, making standard pinhole-based methods [48] unreliable.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common approach to handle rolling shutter distortion is modeling camera motion with parametric models such as SLERP [64], Cayley transformation [25], linearized", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9167d9030ae607acdad8367af0c7b8ed46e8ed334a531b47fc5c5102eaeb54d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 287, + 647, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00ca0360a4f25ff5d06d72003db2394c31994a8bd8dcfb333683b0b05e46b345.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 289, + 763, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/423bd1bad9211b1cdc4ac82bfcc6f9d7fb39a98f41734eb1d188f4b5c6f96bd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 289, + 880, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ffcf8b46cdcc7b0f1ea112eaec5af7df9692ec2d7fce9ea88d201e2e19c7626.jpg", + "image_caption": [ + "Figure 1. We estimate a relative pose from the intersections of the line projections with a single scanline per image (green). This allows us to estimate relative pose between rolling shutter cameras without modelling the camera motion." + ], + "image_footnote": [], + "bbox": [ + 593, + 362, + 821, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rotation [4], affine motion [67], pushbroom [46], accelerated motion [19], or order-one motion [24]. This has been successfully applied to absolute pose estimation [2-6, 9, 32, 33, 36, 44, 45, 59], bundle adjustment [27, 41, 42, 47, 60, 62, 71, 74], SfM [26, 31, 58], and motion rectification from single cameras [35, 51, 52, 56, 72, 77], camera rigs [1, 7, 20, 63], and videos [14, 19, 21, 23, 30, 53, 54, 57, 67, 68, 75].", + "bbox": [ + 511, + 546, + 906, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While most methods use point-based features, some employ lines for absolute pose estimation [3], bundle adjustment [74], and motion rectification [35, 51, 52, 56].", + "bbox": [ + 511, + 670, + 905, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The problem of estimating relative pose for rolling shutter cameras has been explored in prior work. Special essential matrices for RS cameras under different motion models were introduced in [15], along with linear non-minimal solvers. The problem was also solved for pure translation models with constant velocity and acceleration [76]. IMU-based approaches reduced the required number of point correspondences from 11 to 5 [38, 39]. Other methods first estimate pose using a global shutter solver before refining with an RS model [37]. Additional works estimate relative pose for RS camera rigs [70] and for planar scenes [34]. Recently, [24] classified RS relative pose minimal problems", + "bbox": [ + 509, + 719, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7143", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "under the order-one motion model. Despite this, the general RS relative pose problem remains unsolved.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One drawback of parametric motion models is their reliance on an assumed motion that may not match real-world trajectories. An alternative is estimating pose independently for each scanline, similar to line-camera calibration [28, 40, 66, 69, 73]. This approach has been used for tracking a single camera [10, 11] and a camera rig [16] with respect to a predefined marker.", + "bbox": [ + 89, + 121, + 483, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We explore the feasibility of estimating relative pose between scanlines of RS cameras, without explicit motion modelling. The goal is to use this method as a building block to initialize rolling-shutter SfM where no motion model is required and each scanline can have its pose independently computed. To the best of our knowledge, this is the first work that enables this. After the initialization, scanlines from remaining images may be registered by solving the generalized relative pose problem [65], similarly to [10, 11, 16], or, in the case of a planar scene, using [9]. Our contributions include:", + "bbox": [ + 89, + 227, + 483, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Formulation of the problem of relative pose estimation from line projections in a single scanline per image, assuming known intrinsics and no lens distortion.", + "- Enumeration of minimal problems for general and specialized settings with parallel lines and gravity priors.", + "- Demonstration that the problems with parallel lines and with vertical lines and gravity prior correspond to estimating relative pose of uncalibrated cameras [55], respectively calibrated [8] 1D camera pose.", + "- Development of minimal solvers for the following cases:" + ], + "bbox": [ + 89, + 393, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 3 cameras, 7 parallel lines (with projective ambiguity).", + "- 3 cameras, gravity prior, 7 parallel lines.", + "- 3 cameras, gravity prior, 5 vertical lines.", + "- 4 cameras, gravity prior, 4 vertical lines." + ], + "bbox": [ + 104, + 545, + 480, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Evaluation of solvers on synthetic and real RS datasets (Fastec, [43]), taking scanlines either from different images or from a single image.", + "bbox": [ + 89, + 604, + 482, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Notation", + "text_level": 1, + "bbox": [ + 89, + 657, + 194, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use bold letters for matrices and vectors, and normal font for scalars. Indexing starts from one, with indices in brackets; e.g., the first element of $\\mathbf{L}_d$ is $\\mathbf{L}_{d(1)}$ . Matrix indexing follows Matlab style, where $\\mathbf{A}_{(:,1:2)}$ selects the first two rows of $\\mathbf{A}$ . The scanlines are denoted with a scalar $y \\in \\mathbb{R}$ giving their y-coordinate. The set of all 3D lines is denoted as $Gr(1,3)$ .", + "bbox": [ + 89, + 680, + 483, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 800, + 295, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section describes the geometry of a line projected into a rolling-shutter camera and derives an algebraic constraint useful for estimating camera poses.", + "bbox": [ + 89, + 824, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consider $n$ rolling-shutter cameras capturing the same 3D scene. Due to motion, the cameras capture different", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0802ace47166ecf5d8b9c5b091b33e6bf11be274f82906f00bdc096d658c563.jpg", + "image_caption": [ + "Figure 2. Illustration of the line projection. Line $\\mathbf{L}$ projects onto scanline $\\mathbf{y}$ at point $\\mathbf{p}$ . Then, the ray $\\mathbf{r}$ obtained by backprojecting $\\mathbf{p}$ intersects $\\mathbf{L}$ , yielding the constraint (3)." + ], + "image_footnote": [], + "bbox": [ + 576, + 88, + 841, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rows at different poses. Let $\\mathbf{R}_i(y) \\in SO(3)$ and $\\mathbf{C}_i(y) \\in \\mathbb{R}^3$ denote the orientation and center of camera $i$ at row $y$ .", + "bbox": [ + 511, + 266, + 903, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to estimate the relative pose between the cameras without explicitly modeling their motion functions. To achieve this, we select a single row $y_{i}$ per camera and measure intersections between these selected rows and the projections of $m$ straight lines $\\mathbf{L}_j \\in Gr(1,3)$ . Then, we use this information to recover the camera poses $\\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i)$ . Figure 1 illustrates the setup.", + "bbox": [ + 511, + 297, + 903, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $\\mathbf{p}_{i,j} \\in \\mathbb{P}^2$ , with $p_{i,j} = [x_{i,j} y_i 1]^T$ , denote the projection of line $\\mathbf{L}_j$ onto row $y_i$ of camera $i$ . Then, the back-projected ray $\\mathbf{r}_{i,j}$ from $\\mathbf{p}_{i,j}$ intersects line $\\mathbf{L}_j$ , giving:", + "bbox": [ + 511, + 402, + 903, + 449 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {i, j} = \\mathbf {C} _ {i} (y _ {i}) + \\lambda_ {i, j} \\mathbf {R} _ {i} (y _ {i}) ^ {T} \\mathbf {p} _ {i, j}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 458, + 903, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This situation is depicted in Fig. 2. Let us represent $\\mathbf{L}_j$ with a pair $(\\mathbf{L}_{0,j},\\mathbf{L}_{d,j})$ , where $\\mathbf{L}_{0,j}$ is a point on the line and $\\mathbf{L}_{d,j}$ is its direction. The ray $\\mathbf{r}_{i,j}$ intersects $\\mathbf{L}_j$ if and only if:", + "bbox": [ + 511, + 486, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\exists \\alpha , \\beta \\in \\mathbb {R}: \\quad \\mathbf {C} _ {i} \\left(y _ {i}\\right) + \\alpha \\mathbf {R} _ {i} \\left(y _ {i}\\right) ^ {T} \\mathbf {p} _ {i, j} = \\mathbf {L} _ {0, j} + \\beta \\mathbf {L} _ {d, j}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 541, + 903, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This is equivalent to requiring that vectors $\\mathbf{R}_i(y_i)^T\\mathbf{p}_{i,j}$ , $\\mathbf{L}_{d,j}$ , and $\\mathbf{L}_{0,j} - \\mathbf{C}_i(y_i)$ are linearly dependent. Since three 3D vectors are linearly dependent if and only if their scalar triple product vanishes, we obtain:", + "bbox": [ + 511, + 573, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {L} _ {d, j} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 643, + 903, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Each projection provides a single constraint on the unknowns (lines and camera poses). A 3D line has four degrees of freedom, requiring four scanlines for triangulation and at least five scanlines for constraining relative pose.", + "bbox": [ + 511, + 672, + 906, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Assuming parallel lines", + "text_level": 1, + "bbox": [ + 511, + 742, + 730, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let us consider the scenario where all lines are parallel, which frequently occurs in real-world settings. Here, all lines share a common direction $\\mathbf{L}_d \\in \\mathbb{R}^3$ . Since Eq. (3) is invariant to scene rotation, we align $\\mathbf{L}_d$ with the y-axis, $\\mathbf{e}_2$ . This allows representing each line $\\mathbf{L}_j$ with a single point $\\mathbf{L}_{0,j}$ in the xz-plane, with its y-coordinate equal to zero.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With a shared direction, each line has only two individual degrees of freedom (DoF). Once $\\mathbf{L}_d$ is known, a line can be triangulated from just two cameras. This reduces the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7144", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "minimum number of cameras needed for pose estimation to three. To simplify Eq. (3), we substitute $\\mathbf{e}_2$ for $\\mathbf{L}_{d,j}$ :", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} (y _ {i}) \\left(\\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (3)} \\\\ 0 \\\\ - \\mathbf {L} _ {0, j (1)} \\end{array} \\right] - \\left[ \\begin{array}{c} \\mathbf {C} _ {i} (y _ {i}) _ {(3)} \\\\ 0 \\\\ - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} \\end{array} \\right]\\right) = 0. \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 128, + 482, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define rotation $\\mathbf{R}_{0,i} \\in SO(3)$ around the x-axis, which maps row $y_{i}$ to $y = 0$ , and set $\\mathbf{p}_{i,j}' = \\mathbf{R}_{0,i}\\mathbf{p}_{i,j}$ . Then, there holds $\\mathbf{p}_{i,j}' = [x_{i,j}'01]^T$ . Defining", + "bbox": [ + 89, + 183, + 483, + 233 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {i} ^ {\\prime} = \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 316, + 241, + 482, + 266 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {i} ^ {\\prime} = - \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}) [ \\mathbf {C} _ {i} (y _ {i}) _ {(3)} 0 - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} ] ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 261, + 441, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and substituting $\\mathbf{R}_i^{\\prime},\\mathbf{t}_i^{\\prime}$ , and $\\mathbf{p}_{i,j}^{\\prime}$ into Eq. (4), we obtain:", + "bbox": [ + 89, + 285, + 460, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l l} x _ {i, j} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} \\left[ \\begin{array}{c c c} - \\mathbf {R} _ {i (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i (1, 1)} ^ {\\prime} & \\mathbf {t} _ {i (0)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i (3, 1)} ^ {\\prime} & \\mathbf {t} _ {i (2)} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (1)} \\\\ \\mathbf {L} _ {0, j (3)} \\\\ 1 \\end{array} \\right] = 0. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 309, + 482, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can write Equation (6) in a vector form as", + "bbox": [ + 89, + 371, + 393, + 385 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {i, j} ^ {T} \\mathbf {A} _ {i} \\mathbf {L} _ {h, j} = 0, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 392, + 482, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_i\\in \\mathbb{R}^{2,3}$ is a matrix encoding the camera pose, $\\mathbf{u}_{i,j}\\in \\mathbb{P}^1$ is a vector encoding the observation, and $\\mathbf{L}_{h,j}\\in$ $\\mathbb{P}^2$ is a homogeneous representation of point $\\mathbf{L}_{0,j}$ . Alternatively, we define $\\mathbf{u}_{i,j}^{\\prime} = [1 - x_{i,j}^{\\prime}]^{T}$ , and rewrite (7) as", + "bbox": [ + 89, + 417, + 483, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\mathbf {A} _ {i} \\mathbf {L} _ {j, h}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 492, + 482, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since constraint (7) is homogeneous, matrix $\\mathbf{A}_i$ has 5 DoF. The elements $\\mathbf{R}_i(y_i)$ , $\\mathbf{C}_i(y_i)_{(1)}$ , $\\mathbf{C}_i(y_i)_{(3)}$ of the pose, which appear in $\\mathbf{A}_i$ , also have 5 DoF. Using the Gröbner basis method, we have verified that for every matrix $\\mathbf{A} \\in \\mathbb{R}^{2,3}$ , there are 8 ways to decompose it into pose parameters $R \\in SO(3)$ , $C_{(1)} \\in \\mathbb{R}$ , $C_{(3)} \\in \\mathbb{R}$ , and a scale factor $\\sigma \\in \\mathbb{R}$ . This means that $\\mathbf{A}_i$ has no internal constraints, and every $2 \\times 3$ matrix can represent $\\mathbf{A}_i$ for some pose.", + "bbox": [ + 89, + 513, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, the relative pose problem can be transformed into finding matrices $\\mathbf{A}_i\\in \\mathbb{R}^{2,3}$ , vectors $\\mathbf{L}_{h,j}\\in \\mathbb{P}^2$ , and scalars $\\lambda_{i,j}\\in \\mathbb{R}$ such that (8) holds for every $i\\in \\{1,\\dots,n\\}$ , $j\\in \\{1,\\dots,m\\}$ . This is equivalent to the problem of estimating 2D structure from images taken by $n$ uncalibrated 1D cameras, which has been studied in [55].", + "bbox": [ + 89, + 633, + 483, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1.1. Ambiguities in the case with parallel lines", + "text_level": 1, + "bbox": [ + 89, + 731, + 426, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The assumption that all lines are parallel introduces ambiguities into the relative pose estimation. First, since all line directions $\\mathbf{L}_{d,j}$ are equal, the values of constraints (3) remain unchanged if we shift $\\mathbf{C}_i(y_i)$ to $\\mathbf{C}_i(y_i) + \\gamma \\mathbf{L}_{j,d}$ for any $\\gamma \\in \\mathbb{R}$ . Therefore, it is only possible to estimate the camera centers up to translation along $\\mathbf{L}_d$ .", + "bbox": [ + 89, + 750, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, if (8) holds for some $\\mathbf{A}_i\\in \\mathbb{R}^{2,3}$ , $\\mathbf{L}_{h,j}\\in \\mathbb{P}^2$ , then (8) also holds for $\\mathbf{A}_i\\cdot \\mathbf{H}^{-1}$ , $\\mathbf{H}\\cdot \\mathbf{L}_{j,h}$ for every invertible $\\mathbf{H}\\in \\mathbb{R}^{3,3}$ . Therefore, the cameras and the structure can only be estimated up to a projective transformation $\\mathbf{H}$ .", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show later in the paper that we can resolve the projective ambiguity by introducing priors on the cameras. To resolve the ambiguity in the camera centers, we can, for instance, observe one line orthogonal to $\\mathbf{L}_d$ and use it to estimate the camera centers.", + "bbox": [ + 511, + 90, + 905, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Enumerating minimal problems", + "text_level": 1, + "bbox": [ + 511, + 179, + 807, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we classify minimal problems for estimating the relative pose from projections of $n$ lines onto $m$ scanlines, along with their degrees. We consider five settings:", + "bbox": [ + 511, + 205, + 905, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A. Fully generic setting", + "text_level": 1, + "bbox": [ + 516, + 252, + 687, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "B. Parallel lines", + "text_level": 1, + "bbox": [ + 517, + 268, + 635, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "C. Gravity prior for cameras + Generic lines", + "text_level": 1, + "bbox": [ + 516, + 282, + 834, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "D. Gravity prior for cameras + Parallel lines", + "text_level": 1, + "bbox": [ + 516, + 297, + 833, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "E. Gravity prior for cameras + Vertical lines", + "text_level": 1, + "bbox": [ + 517, + 313, + 834, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In cases A and C, we estimate the pose up to global rotation, translation, and scale. In cases D and E, we estimate the pose up to these transformations and an additional shift of centers $\\mathbf{C}_i(y_i)$ along direction $\\mathbf{L}_d$ . In case B, we estimate the pose up to these transformations, the shift of centers along $\\mathbf{L}_d$ , and the projective transformation $\\mathbf{H}$ .", + "bbox": [ + 511, + 330, + 905, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each setting, we determine the number of cameras $m$ and the number of lines $n$ s.t. the problem (3) is minimal. We assume complete visibility, meaning each line is projected into each camera. We follow concepts from [17]: Minimal problem [17, Def. 1]: A parametric polynomial system $f(\\theta ,\\mathbf{x}) = 0$ with a finite nonzero number of solutions $\\mathbf{x}^*$ for a generic set of parameters $\\theta$ .", + "bbox": [ + 511, + 421, + 905, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Balanced problem [17, Def. 2]: A parametric polynomial system $f(\\theta, \\mathbf{x}) = 0$ where the dimensionality of the space of variables $\\mathbf{x}$ equals the number of independent equations. Condition for minimal problems [17, Step. 1, Corr. 2]: A problem $f(\\theta, \\mathbf{x}) = 0$ is minimal if and only if it is balanced and its Jacobian $\\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}}$ is invertible for generic $\\theta, \\mathbf{x}^*$ .", + "bbox": [ + 511, + 526, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This summary provides the foundation for enumerating minimal problems. For detailed theory, refer to [17]. In our case, the parameter vector $\\theta$ consists of projections $\\mathbf{p}_{i,j} \\in \\mathbb{P}^2$ , and the variable vector $\\mathbf{x}$ encodes camera poses $(\\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i))$ and lines $(\\mathbf{L}_{0,j}, \\mathbf{L}_{d,j})$ . The polynomial system $f(\\mathbf{x}, \\theta)$ consists of equations in the form of (3). Following [17], we used the following steps to list all minimal problems and determine their degrees:", + "bbox": [ + 511, + 619, + 905, + 739 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Enumerate all balanced problems for each setting A-E.", + "- For each balanced problem, generate random camera poses and lines $\\mathbf{x}^*$ , reproject the lines into the cameras, and obtain projections $\\theta$ such that $f(\\theta, \\mathbf{x}^*) = 0$ .", + "- Compute the Jacobian $\\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}}$ . Label the problem as minimal if $\\det \\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}} \\neq 0$ .", + "- Use monodromy computation from [13] to determine the degree of minimal problems." + ], + "bbox": [ + 511, + 742, + 903, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to algebraic geometry, finding a single instance where the Jacobian is full-rank suffices to claim generic", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7145", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c712fc3ee14f2088beeca32433b54caad13316f790b2deec4d3e0fa702da7cf8.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. List of balanced problems for settings A-E. S. denotes the setting, $m$ and $n$ are the number of cameras and lines, and Deg is the degree. Column Min. contains Y if the problem is minimal. $2^{*}$ indicates two solutions in terms of $2 \\times 3$ matrices, and + denotes computations that did not terminate. See Sec. 3 for details." + ], + "table_body": "
S.mnMin.DegS.mnMin.Deg
A523Y389k+D37Y48
A217Y40k+D45Y232
B37Y2*D64Y1224
B46Y2*E35Y16
C515Y1.8M+E44Y32
C155Y532k+
", + "bbox": [ + 91, + 89, + 483, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "minimality. In the following subsections, we enumerate minimal problems for each setting A-E.", + "bbox": [ + 89, + 303, + 483, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A. Fully Generic Setting Each camera has 6 degrees of freedom (DoF), while each line has 4 DoF. Due to gauge freedom, we fix 7 DoF (3 for rotation, 3 for translation, and 1 for scale). There are $m \\cdot n$ constraints in the form of (3), fixing $m \\cdot n$ DoF. The balanced constraint is therefore:", + "bbox": [ + 89, + 351, + 483, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n6 m + 4 n - 7 = m n. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 435, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define a function $n_A(m)$ , mapping the number of cameras $m$ to the number of lines $n$ , s.t. (9) holds for $m$ , $n_A(m)$ :", + "bbox": [ + 89, + 459, + 483, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {A} (m) = \\frac {6 m - 7}{m - 4}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 500, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As stated in Sec. 2, the minimal number of cameras is $m \\geq 5$ . The function $n_A(m)$ is strictly decreasing for $m > 4$ , and since $\\lim_{m \\to \\infty} n_A(m) = 6$ , no balanced problem has fewer than 7 lines. Evaluating $n_A(m)$ for integer values of $m$ from 5 onward, and searching for integer values of $n_A(m)$ , we identify the following balanced cases:", + "bbox": [ + 89, + 532, + 483, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bullet m = 5, n = 2 3; \\quad m = 2 1, n = 7\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 625, + 334, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. Parallel Lines As discussed in Sec. 2, the pose is estimated up to a shift along $\\mathbf{L}_d$ and a projective transformation $\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}$ . The pose estimation corresponds to estimating matrices $\\mathbf{A}_i \\in \\mathbb{R}^{2 \\times 3}$ and points $\\mathbf{L}_{h,j} \\in \\mathbb{P}^2$ such that (7) holds for all $i, j$ . Here, each camera has 5 DoF, each line has 2 DoF, and we fix 8 DoF due to projective ambiguity. The balanced constraint is:", + "bbox": [ + 89, + 657, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n5 m + 2 n - 8 = m n. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 772, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining $n_B(m)$ as:", + "bbox": [ + 89, + 796, + 227, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {B} (m) = \\frac {5 m - 8}{m - 2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 818, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and noting that $m \\geq 3$ , we find $\\lim_{m \\to \\infty} n_B(m) = 5$ , implying no balanced problem has fewer than 6 lines. Evaluating $n_B(m)$ for integer $m$ , we obtain:", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bullet m = 3, n = 7; \\quad m = 4, n = 6\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 90, + 740, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Gravity Prior for Cameras + Generic Lines With a known vertical direction, each camera has 4 DoF (1 for rotation, 3 for translation), and we fix 5 DoF due to gauge freedom. Lines remain generic, contributing 4 DoF each. The balanced constraint is:", + "bbox": [ + 511, + 123, + 906, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n4 m + 4 n - 5 = m n. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 214, + 906, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The corresponding function is:", + "bbox": [ + 513, + 234, + 720, + 250 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {C} (m) = \\frac {4 m - 5}{m - 4}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 253, + 905, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find $\\lim_{m\\to \\infty}n_C(m) = 4$ so no balanced problem has fewer than 5 lines. Checking integer values of $m$ , we get:", + "bbox": [ + 511, + 290, + 905, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bullet m = 5, n = 1 5; \\quad m = 1 5, n = 5 \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 321, + 754, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Gravity Prior for Cameras + Parallel Lines Here, camera rotation has 1 DoF, and the camera center has 2 DoF (up to translation along the line). We fix 4 DoF due to gauge freedom. Lines share 2 DoF for direction $\\mathbf{L}_d$ , with each line having 2 additional DoF. The balanced constraint is:", + "bbox": [ + 511, + 352, + 905, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n3 m + 2 + 2 n - 4 = m n. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 436, + 903, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We get $n_D(m)$ as:", + "bbox": [ + 513, + 459, + 640, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {D} (m) = \\frac {3 m - 2}{m - 2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 479, + 903, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $m \\geq 3$ , we find $\\lim_{m \\to \\infty} n_D(m) = 3$ , meaning no balanced problem has fewer than 4 lines. We have found:", + "bbox": [ + 511, + 515, + 905, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bullet m = 3, n = 7; \\quad m = 4, n = 5; \\quad m = 6, n = 4\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 547, + 854, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "E. Gravity Prior for Cameras + Vertical Lines Similar to the previous case, each camera has 3 DoF, and we fix 4 DoF due to gauge freedom. Since the direction $\\mathbf{L}_d$ is fixed, each line has 2 DoF. The balanced constraint is:", + "bbox": [ + 511, + 579, + 905, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n3 m + 2 n - 4 = m n. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 655, + 903, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define $n_E(m)$ as:", + "bbox": [ + 513, + 675, + 658, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nn _ {E} (m) = \\frac {3 m - 4}{m - 2}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 695, + 903, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For $m \\geq 3$ , we find $\\lim_{m \\to \\infty} n_E(m) = 3$ , meaning no balanced problem has fewer than 4 lines. We found problems:", + "bbox": [ + 511, + 731, + 905, + 761 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bullet m = 3, n = 5; \\quad m = 4, n = 4\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 763, + 740, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All balanced problems for these settings are summarized in Table 1, along with their minimality status and degree. The table shows that problems with parallel/vertical lines (B, D, E) are minimal and computationally feasible. Problems with generic lines (A, C) are minimal but have degrees too high for practical RANSAC solutions. For problems $(A,5,23)$ and $(C,5,15)$ , monodromy computations did not terminate within a month.", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7146", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Minimal solvers", + "text_level": 1, + "bbox": [ + 89, + 89, + 250, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce solvers for the minimal problems $(B,3,7)$ , $(D,3,7)$ , $(E,3,5)$ , $(E,4,4)$ from Section 3. We use the notation and concepts from Sec. 2.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Parallel lines with generic cameras", + "text_level": 1, + "bbox": [ + 89, + 167, + 393, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we describe the solution to the problem $(B,3,7)$ . Since the lines are parallel, our goal is to find matrices $\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3\\in \\mathbb{R}^{2\\times 3}$ and vectors $\\mathbf{L}_{h,j}$ for $j\\in \\{1,\\dots ,7\\}$ , such that equation (7) holds for every camera $i$ and line $j$ . Due to projective ambiguity, we aim to recover the camera poses up to a common projective transformation $\\mathbf{H}\\in \\mathbb{R}^{3\\times 3}$ . Each camera triplet $\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3$ can be expressed as:", + "bbox": [ + 89, + 190, + 483, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {1} = \\mathbf {A} _ {c, 1} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {2} = \\mathbf {A} _ {c, 2} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {3} = \\mathbf {A} _ {c, 3} \\cdot \\mathbf {H}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 306, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for some $\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}$ , with $\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}$ taking the form:", + "bbox": [ + 89, + 330, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {c, 1} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {1} & \\alpha_ {1} & \\alpha_ {1} \\end{array} \\right], \\mathbf {A} _ {c, 2} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {2} & \\alpha_ {3} & \\alpha_ {4} \\end{array} \\right], \\mathbf {A} _ {c, 3} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {5} & \\alpha_ {6} & \\alpha_ {7} \\end{array} \\right]. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 368, + 498, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof follows analogously to Proposition 2 in [29]. These canonical-form matrices have 7 DoF.", + "bbox": [ + 89, + 401, + 480, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To estimate $\\mathbf{A}_{c,1},\\mathbf{A}_{c,2},\\mathbf{A}_{c,3}$ , we leverage the fact that, as shown in [55], a trifocal tensor exists between three camera matrices $\\mathbf{A}_i\\in \\mathbb{R}^{2\\times 3}$ . To compute this tensor, we stack the constraints from (8) into the matrix equation:", + "bbox": [ + 89, + 431, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c c} \\mathbf {A} _ {1} & \\mathbf {u} _ {1, j} ^ {\\prime} & 0 & 0 \\\\ \\mathbf {A} _ {2} & 0 & \\mathbf {u} _ {2, j} ^ {\\prime} & 0 \\\\ \\mathbf {A} _ {3} & 0 & 0 & \\mathbf {u} _ {3, j} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {h, j} \\\\ - \\lambda_ {1, j} \\\\ - \\lambda_ {2, j} \\\\ - \\lambda_ {3, j} \\end{array} \\right] = 0. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 500, + 480, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For this equation to hold, the determinant of the coefficient matrix must vanish, leading to the constraint:", + "bbox": [ + 89, + 570, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {a = 1} ^ {2} \\sum_ {b = 1} ^ {2} \\sum_ {c = 1} ^ {2} \\mathbf {u} _ {1, j (a)} ^ {\\prime} \\mathbf {u} _ {2, j (b)} ^ {\\prime} \\mathbf {u} _ {3, j (c)} ^ {\\prime} \\mathbf {T} _ {(a, b, c)} = 0, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 608, + 480, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}$ is the trifocal tensor with elements:", + "bbox": [ + 89, + 657, + 459, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {(a, b, c)} = (- 1) ^ {a + b + c} \\det \\left(\\left[ \\mathbf {A} _ {1 (a,:)} ^ {T} \\mathbf {A} _ {2 (b,:)} ^ {T} \\mathbf {A} _ {3 (c,:)} ^ {T} \\right]\\right). \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 678, + 480, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in [55], $\\mathbf{T}$ has no internal constraints and it can be estimated using 7 line correspondences.", + "bbox": [ + 89, + 717, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $\\mathbf{T}$ is homogeneous, it has 7 DoF, matching the DoF of the canonical camera matrices. We verified that for every tensor $\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}$ , there are exactly two corresponding triplets of matrices $\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}$ . The proof and the decomposition algorithm for recovering these matrices can be found in the SM. Once $\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}$ are found, we can triangulate the lines up to the projective transformation $\\mathbf{H}$ .", + "bbox": [ + 89, + 747, + 482, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The solver for the $(B,3,7)$ problem, which estimates the trifocal tensor $\\mathbf{T}$ , operates as follows:", + "bbox": [ + 89, + 853, + 482, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Input: Projections $\\mathbf{p}_{i,j}$ of 7 lines in 3 cameras.", + "bbox": [ + 89, + 885, + 423, + 901 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. Compute matrices $\\mathbf{R}_{0,i}$ and vectors $\\mathbf{u}_{i,j}$ and $\\mathbf{u}_{i,j}^{\\prime}$ .", + "3. Construct the linear system (22).", + "4. Solve (22) using singular value decomposition (SVD) to estimate the tensor $\\mathbf{T}$ ." + ], + "bbox": [ + 511, + 90, + 903, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On average, the solver runs in $6.92\\mu s$", + "bbox": [ + 511, + 152, + 769, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Vertical lines and known gravity", + "text_level": 1, + "bbox": [ + 511, + 176, + 797, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we analyze the solutions to the problems $(E,3,5)$ and $(E,4,4)$ . We assume that all lines are vertical and the gravity directions $\\mathbf{v}_i\\in \\mathbb{P}^2$ for the cameras are known.", + "bbox": [ + 511, + 198, + 903, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to Sec. 4.1, we set $\\mathbf{L}_{d,j} = \\mathbf{e}_2$ . Therefore, Equation (6) still holds. Furthermore the vanishing points $\\mathbf{v}_i$ for $i \\in \\{1,\\dots,m\\}$ associated with the vertical lines are known, and there holds $\\mathbf{v}_i = \\mathbf{R}_i(y_i)\\mathbf{e}_2$ . We can decompose $\\mathbf{R}_i(y_i) = \\mathbf{R}_{v,i}\\mathbf{R}_{y,i}$ , where $\\mathbf{R}_{y,i}$ is a rotation around y-axis, and $\\mathbf{R}_{v,i}$ is a rotation around an axis in the xz-plane, such that $\\mathbf{R}_{v,i}\\mathbf{e}_2 = \\mathbf{v}_i$ . Since $\\mathbf{R}_{v,i}$ is calculated directly from $\\mathbf{v}_i$ , we assume it to be known.", + "bbox": [ + 511, + 244, + 905, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this case, $\\mathbf{R}_i^{\\prime}$ (5) takes the form $\\mathbf{R}_i^{\\prime} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}$ . While both $\\mathbf{R}_{0,i}$ and $\\mathbf{R}_{v,i}$ have their axes in the xz-plane, $\\mathbf{R}_{0,i}\\mathbf{R}_{v,i}$ does generically not share this property. However, we can decompose $\\mathbf{R}_{0,i}\\mathbf{R}_{v,i} = \\mathbf{R}_{a,i}\\mathbf{R}_{b,i}$ , s.t. $\\mathbf{R}_{a,i}$ has its axis in the xz-plane and $\\mathbf{R}_{b,i}$ rotates around the y-axis.", + "bbox": [ + 511, + 364, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, we decompose $\\mathbf{R}_i^{\\prime}$ as $\\mathbf{R}_{a,i} \\cdot \\mathbf{R}_i^{\\prime \\prime}$ , where $\\mathbf{R}_{a,i}$ is a known rotation around axis in the xz-plane, and $\\mathbf{R}_i^{\\prime \\prime} = \\mathbf{R}_{b,i}$ . $\\mathbf{R}_{y,i}$ is an unknown rotation around y-axis.", + "bbox": [ + 511, + 439, + 903, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We represent $\\mathbf{R}_{a,i},\\mathbf{R}_i^{\\prime \\prime}$ with quaternions $\\mathbf{q}_i = [\\mathbf{q}_{i,w}\\mathbf{q}_{i,x}0\\mathbf{q}_{i,z}]^T$ $\\mathbf{q}_i^{\\prime \\prime} = [\\mathbf{q}_{i,w}^{\\prime \\prime}0\\mathbf{q}_{i,y}^{\\prime \\prime}0]^T$ , write the elements of $\\mathbf{R}'$ in the form", + "bbox": [ + 511, + 484, + 903, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {R} _ {i (1, 1)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) - 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (1, 3)} ^ {\\prime} = 2 \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} + 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right), \\\\ \\mathbf {R} _ {i (3, 1)} ^ {\\prime} = 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) - 2 (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (3, 3)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, x} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) + 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\tag {24} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 534, + 903, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and decompose the left $2 \\times 2$ submatrix of $\\mathbf{A}_i$ (7) as", + "bbox": [ + 511, + 642, + 861, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l l} - \\mathbf {R} _ {i, (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (1, 1)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i, (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (3, 1)} ^ {\\prime} \\end{array} \\right] = \\mathbf {A} _ {i, a} \\cdot \\mathbf {A} _ {i, b}, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 664, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 513, + 710, + 547, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i, a} = \\left[ \\begin{array}{l l} - 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} & (1 - 2 \\mathbf {q} _ {i, z} ^ {2}) \\\\ - (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) & 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\end{array} \\right], \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 729, + 903, + 765 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i, b} = \\left[ \\begin{array}{c c} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) & 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ - 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) \\end{array} \\right]. \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 773, + 903, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{A}_{i,a}$ can be directly obtained from the known gravity direction and the current scanline $y_{i}$ , and $\\mathbf{A}_{i,b}$ is an unknown $2D$ rotation matrix. We define $\\mathbf{A}_i^{\\prime} = \\mathbf{A}_{i,a}^{-1}\\mathbf{A}_i$ , and $\\mathbf{u}_{i,j}^{\\prime \\prime} = \\mathbf{A}_{i,a}^{\\prime T}\\mathbf{u}_{i,j}$ . Then, there holds", + "bbox": [ + 511, + 811, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {i, j} ^ {\\prime \\prime T} \\mathbf {A} _ {i} ^ {\\prime} \\mathbf {L} _ {j, h} = 0. \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 883, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7147", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the left $2 \\times 2$ submatrix of $\\mathbf{A}_i'$ is $\\mathbf{A}_{i,b}$ , $\\mathbf{A}_i'$ has a form of a calibrated $2 \\times 3$ camera matrix, and the problem with known gravity direction and vertical lines is equivalent to estimating a 2D structure from images taken by $n$ calibrated cameras [8]. Unlike its uncalibrated counterpart, the calibrated problem can be solved without projective ambiguity. Solvers for the cases $m = 3, n = 5$ and $m = 4, n = 4$ have been presented in [8]. Notably, these numbers exactly correspond to the minimal problems discussed in Sec. 3.", + "bbox": [ + 89, + 90, + 483, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The solver for $m = 3, n = 5$ estimates a trifocal tensor of the form (23) from 5 point correspondences and 2 linear internal constraints, and solves a quadratic function to decompose the tensor into camera matrices. The solver for $m = 4, n = 4$ estimates a dual quadrifocal tensor from 4 point correspondences and 11 linear internal constraints, and also solves a quadratic function to decompose the tensor into camera matrices. In both cases, there are two solutions in terms of matrices $\\mathbf{A}_i'$ . Typically, the correct solution is selected by counting the points lying in front of all cameras.", + "bbox": [ + 89, + 227, + 483, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After finding $\\mathbf{A}_i^{\\prime}$ with [8], we construct the camera poses $\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)$ . For this, we first obtain $\\mathbf{R}_i^{\\prime \\prime}$ . Because (28) is homogeneous, there are two possible matrices $\\mathbf{R}_i^{\\prime \\prime}$ :", + "bbox": [ + 89, + 378, + 483, + 422 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {R} _ {i} ^ {\\prime \\prime} \\in \\left\\{\\left[ \\begin{array}{c c c} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right], \\right. \\tag {29} \\\\ \\left. \\left[ \\begin{array}{c c c} - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right] \\right\\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 431, + 483, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The existence of two choices of $\\mathbf{R}_i^{\\prime \\prime}$ per $\\mathbf{A}_i^\\prime$ accounts for the presence of 16 solutions in the 3 view case and 32 solutions in the 4 view case. Finally, we construct $\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)$ :", + "bbox": [ + 89, + 540, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {0, i} ^ {T} \\mathbf {R} _ {a, i} \\mathbf {R} _ {i} ^ {\\prime \\prime}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {i (1: 2, 3)}. \\tag {30} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 594, + 480, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note, that $\\mathbf{C}_i(y_i)_{(2)}$ is always zero. This is in accordance with the Section 2, which explains that the camera centers can only be estimated up to the shift along the direction $\\mathbf{L}_d$ .", + "bbox": [ + 89, + 643, + 482, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, we perform the following steps to solve problems $(E,3,5)$ and $(E,4,4)$ :", + "bbox": [ + 89, + 689, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Input: Projections $\\mathbf{p}_{i,j}$ , vertical directions $\\mathbf{v}_i$ .", + "2. Calculate matrices $\\mathbf{R}_{0,i},\\mathbf{R}_{v,i},\\mathbf{R}_{a,i}$ , and vectors $\\mathbf{u}_{i,j}$ (7).", + "3. Compose matrices $\\mathbf{A}_{i,a}$ (26) and vectors $\\mathbf{u}_{i,j}^{\\prime \\prime}(28)$ .", + "4. Use the 3 view or the 4 view solver from [8] to find 2 sets of matrices $\\mathbf{A}_i^{\\prime}$ , $i\\in \\{1,\\dots,m\\}$ from $\\mathbf{u}_{i,j}^{\\prime \\prime}$ .", + "5. Construct $\\mathbf{R}_i^{\\prime \\prime}$ (29), $\\mathbf{R}_i(y_i)$ , and $\\mathbf{C}_i(y_i)$ (30).", + "On average, the 3 view solver takes $9.16\\mu s$ , and the 4 view solver takes $70.84\\mu s$ ." + ], + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Parallel lines and known gravity", + "text_level": 1, + "bbox": [ + 89, + 848, + 375, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the solution to the problem $(D,3,7)$ . In this case, all lines share a common direction $\\mathbf{L}_d$ , which is", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generically not aligned with the vertical direction $\\mathbf{e}_2$ . Let $\\mathbf{R}_d \\in SO(3)$ be a rotation around an axis in the xz-plane, s.t. $\\mathbf{L}_d = \\mathbf{R}_d\\mathbf{e}_2$ . Substituting this into (3), we get", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {R} _ {d} \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0, \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 143, + 903, + 162 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\mathbf {R} _ {d} ^ {T} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {32}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 181, + 903, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We define $\\mathbf{L}_{0,j}^{\\prime} = \\mathbf{R}_d^T\\mathbf{L}_{0,j}$ , $\\mathbf{C}_i^\\prime = \\mathbf{R}_d^T\\mathbf{C}_i(y_i)$ and substitute these into (32):", + "bbox": [ + 511, + 204, + 903, + 234 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} ^ {\\prime} - \\mathbf {C} _ {i} ^ {\\prime}\\right) = 0. \\tag {33}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 243, + 903, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following Sec. 4.2, we find rotation matrices $\\mathbf{R}_{0,i}$ and $\\mathbf{R}_{v,i}$ , decompose $\\mathbf{R}_i(y_i)$ as $\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}$ , and find vectors $\\mathbf{u}_{i,j}^{\\prime}$ according to (8). Then, we define $\\mathbf{R}_{D,i} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\mathbf{R}_d$ , and $\\mathbf{t}_{D,i} = -\\mathbf{R}_{D,i}^{-1}[\\mathbf{e}_2]_{\\times}\\mathbf{C}_i^{\\prime}$ , and rewrite the constraint as", + "bbox": [ + 511, + 268, + 905, + 330 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\left[ \\begin{array}{l l l} - \\mathbf {R} _ {D, i (1, 3)} & \\mathbf {R} _ {D, i (1, 1)} & \\mathbf {t} _ {D, i (1)} \\\\ - \\mathbf {R} _ {D, i (3, 3)} & \\mathbf {R} _ {D, i (3, 1)} & \\mathbf {t} _ {D, i (3)} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {L} _ {0, j (1)} ^ {\\prime} \\\\ \\mathbf {L} _ {0, j (3)} ^ {\\prime} \\\\ 1 \\end{array} \\right], \\tag {34}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 339, + 903, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which can be written in matrix form as", + "bbox": [ + 513, + 402, + 772, + 417 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} = \\mathbf {A} _ {D, i} \\mathbf {L} _ {h, j} ^ {\\prime}. \\tag {35}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 426, + 903, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This equation has the same form as (8), ensuring the existence of a trifocal tensor $\\mathbf{T} \\in \\mathbb{R}^{2,2,2}$ satisfying (22). We estimate this tensor linearly using 7 correspondences, as in Sec. 4.1.", + "bbox": [ + 511, + 450, + 905, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let us now discuss how to decompose $\\mathbf{T}$ to estimate the relative pose. First, we fix the coordinate frame by setting $\\mathbf{R}_1(y_1) = I$ , $\\mathbf{C}_1(y_1) = 0$ . Both the tensor $\\mathbf{T}$ and the pose have 7 DoF. Using Gröbner basis, we have discovered that for every tensor $\\mathbf{T} \\in \\mathbb{R}^{2,2,2}$ , there are 48 camera poses $\\mathbf{R}_{y,2}, \\mathbf{t}_{D,2}, \\mathbf{R}_{y,3}, \\mathbf{t}_{D,3}, \\mathbf{R}_d$ , such that the matrices $\\mathbf{A}_{D,i}$ constructed via (34) have the tensor $\\mathbf{T}$ .", + "bbox": [ + 511, + 511, + 905, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the high number of variables and equation complexity, we could not generate an elimination template for a Gröbner basis solver. Therefore, we solve this problem with homotopy continuation [18], with an average runtime $19089\\mu s$ .", + "bbox": [ + 511, + 617, + 905, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we compose $\\mathbf{A}_{i,D}$ according to (34), and the pose $\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)$ as", + "bbox": [ + 511, + 691, + 903, + 723 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {v, i} \\mathbf {R} _ {y, i}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = R _ {d} \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {D, i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {D, i (1: 2, 3)} \\tag {36} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 729, + 903, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To summarize, we perform the following steps to solve the problem $(D,3,7)$ :", + "bbox": [ + 511, + 776, + 905, + 806 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Input: Projections $\\mathbf{p}_{i,j}$ , vertical directions $\\mathbf{v}_i$ .", + "2. Calculate matrices $\\mathbf{R}_{0,i},\\mathbf{R}_{v,i}$ and vectors $\\mathbf{u}_{i,j}$ (7).", + "3. Construct linear system (22) and solve it to get $\\mathbf{T}$ .", + "4. Use homotopy continuation to decompose $\\mathbf{T}$ into $\\mathbf{R}_{y,2}$ , $\\mathbf{t}_{D,2}$ , $\\mathbf{R}_{y,3}$ , $\\mathbf{t}_{3,D}$ , $\\mathbf{R}_d$ .", + "5. Construct $\\mathbf{R}_i(y_i)$ and $\\mathbf{C}_i(y_i)$ using (36)." + ], + "bbox": [ + 513, + 806, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7148", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/564df0bbbfe616916c75a04a33e36c49e60c771b06e0ddae69bb51c83b93c721.jpg", + "image_caption": [ + "Figure 3. Histogram of tensor, rotation, and translation errors of the proposed solvers, over $10^{5}$ noiseless samples. In radians." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Degenerate scenarios", + "text_level": 1, + "bbox": [ + 89, + 265, + 290, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we analyze the degenerate cases for the solvers introduced in Sections 4.1, 4.2, and 4.3. Specifically, we consider scenarios where some lines are coplanar or where camera centers are collinear or coincident.", + "bbox": [ + 89, + 287, + 483, + 347 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Solvers $(E,3,5)$ , $(E,4,4)$ are degenerate if all $m$ lines are coplanar or if any two camera centers coincide.", + "- Solvers $(B,3,7)$ , $(D,3,7)$ are degenerate if at least 5 lines are coplanar or if any two camera centers coincide.", + "- There is no degeneracy if the camera centers are colinear." + ], + "bbox": [ + 89, + 347, + 483, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 438, + 222, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the solvers $(B,3,7)$ , $(D,3,7)$ , $(E,3,5)$ , $(E,4,4)$ on both synthetic and real data.", + "bbox": [ + 89, + 463, + 482, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Synthetic experiments", + "text_level": 1, + "bbox": [ + 89, + 503, + 297, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Numerical stability. We sample scanlines $y_{i}$ from $\\mathcal{U}_{[-1,1]}$ and generate random camera rotations $\\mathbf{R}_{i,GT}$ and centers $\\mathbf{C}_{i,GT}$ . Since we use one scanline per image, we only sample one pose per image. We get the vertical directions as $v_{i} = \\mathbf{R}_{i,GT}\\mathbf{e}_{2}$ . To generate line $\\mathbf{L}_j$ , we sample scalars $\\mu_{j,x},\\mu_{j,z}$ from normalized Gaussian distribution and set $\\mathbf{L}_{0,j} = [\\mu_{j,x}0\\mu_{j,z} + 5]^T$ . For solver $(D,3,7)$ , we also generate scalars $\\nu_{x},\\nu_{z}$ , and set $L_{j,d} = [\\nu_x1\\nu_z]^T$ . In other cases, we use $L_{j,d} = \\mathbf{e}_2$ . We generate $\\mathbf{p}_{i,j}$ by projecting $\\mathbf{L}_j$ into the camera with pose $(\\mathbf{R}_{i,GT},\\mathbf{C}_{i,GT})$ and intersecting the projected line with scanline $y_{i}$ .", + "bbox": [ + 89, + 526, + 483, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Before measuring the error, we transform both the ground truth pose, and the estimated pose $(\\mathbf{R}_{i,\\mathrm{est}},\\mathbf{C}_{i,\\mathrm{est}})$ into a canonical coordinate frame with $\\mathbf{C}_1 = 0$ $\\mathbf{R_{y,1}} = \\mathbf{I},\\mathbf{L_d} = \\mathbf{e_1}$ . Since the y-coordinates of all camera centers are unobservable, we set them to zero. Then, we calculate the rotation error $err_{\\mathbf{R}} = \\max_{i = 1}^{m}(\\angle \\mathbf{R}_{i,\\mathrm{est}}^{T}\\mathbf{R}_{i,GT})$ and the translation error $err_{\\mathbf{C}} = \\max_{i = 1}^{m}\\angle (\\mathbf{C}_{i,GT},\\mathbf{C}_{i,\\mathrm{est}})$ .", + "bbox": [ + 89, + 693, + 482, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we construct tensors $\\mathbf{T}_{GT}$ from GT poses and estimate $\\mathbf{T}_{est}$ using the solvers. For trifocal problems, $\\mathbf{T}_{GT},\\mathbf{T}_{est}$ are trifocal tensors (23), while for $(E,4,4)$ , they are dual quadrifocal tensors. We measure tensor error errT as the Frobenius norm of $\\mathbf{T}_{GT} - \\mathbf{T}_{est}$ . For each solver, we generate $10^{5}$ synthetic noise-free scenes and measure errR, errC, and errT. Results in Fig. 3 show solver stability.", + "bbox": [ + 89, + 800, + 483, + 905 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d808fa8edf1daa28750b34ebcb215b0c48d1687308749456a6175f163299a052.jpg", + "image_caption": [ + "Figure 4. Rotation, translation, and tensor errors averaged over 10000 successful runs as a function of the point and the vertical noise, measured from the real solution closest to the ground truth. If a solver returns only complex solutions, we report a failure. The failure rates are reported in the last row. Plots in the same row share the y-axis and plots in the same column share the x-axis. The fixed parameters are in the titles." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tests with noise. To evaluate the robustness to noise, we set focal length $f = 1000px$ , and add Gaussian noise with standard deviation $\\frac{\\sigma_p}{f}$ to each projection $\\mathbf{p}_{i,j}$ . We also perturb vertical directions $\\mathbf{v}_i$ by applying a random 3D rotation with angle $\\sigma_v$ . Fig. 4 shows $err_{\\mathbf{R}}$ , $err_{\\mathbf{C}}$ , and $err_{\\mathbf{T}}$ as the function of $\\sigma_p$ , $\\sigma_v$ , averaged over $10^4$ runs. The results confirm that, since the line detections tend to be relatively precise, our solvers exhibit reasonable behaviour under realistic noise levels.", + "bbox": [ + 511, + 465, + 906, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Real-world Experiments", + "text_level": 1, + "bbox": [ + 511, + 614, + 738, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our solvers using the Fastec dataset [43], which consists of 19 sequences captured from a moving car in Tübingen. Each sequence contains 34 rolling shutter (RS) images and two global shutter (GS) images for every RS image: one with the pose of the first scanline and one with the pose of the middle scanline. The images have a resolution of $640 \\times 480$ px and a focal length $768$ px.", + "bbox": [ + 511, + 636, + 905, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since GT poses are unavailable, we use COLMAP [61] on the GS images to obtain pseudo-GT poses $(\\mathbf{R}_{i,F},\\mathbf{C}_{i,F})$ for the first scanline of image $i$ and $(\\mathbf{R}_{i,M},\\mathbf{C}_{i,M})$ for the middle scanline. The pseudo-GT pose for scanline $y_{i}$ is interpolated as: $\\mathbf{R}_{i,GT}(y_i) = \\mathbf{R}_{i,M\\to y_i}\\mathbf{R}_{i,M},\\mathbf{C}_{i,GT}(y_i) =$ $\\mathbf{C}_{i,M} + \\frac{h - 2y_i}{h} (\\mathbf{C}_{i,F} - \\mathbf{C}_{i,M})$ , where $\\mathbf{R}_{i,M\\to y_i}$ is a rotation matrix with the axis of $\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}$ , and the angle $\\frac{h - 2y_i}{h}$ times the angle of $\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}$ . We evaluate the solvers in multi-view and single-view settings.", + "bbox": [ + 511, + 743, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-view. We estimate poses for $m$ consecutive im", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7149", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d91acf15aa45a920189060ebc2d175ea2b850f070510cf43cbe62e0a24f8ab4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
solver≤ 5°≤ 10°≤ 20°≤ 30°
MV(E,3,5)131015
(E,4,4)581114
(D,3,7)6101517
SV(E,3,5)33510
(E,4,4)24713
(D,3,7)4479
", + "bbox": [ + 109, + 88, + 460, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ages, selecting one scanline per image. We use DeepLSD [49] and Gluestick [50] to detect and match the lines. Although the projections of lines by RS cameras are generically curves, the line-based methods work well in the case of Fastec dataset. To select scanlines $y_{1}, \\ldots, y_{m}$ , we exhaustively search for the scanlines that maximize the number of matched lines intersecting them. Then, we find $p_{i,j}$ as the intersections of the detected lines and the scanlines. Following [12], we approximate the vertical direction by $[0 1 0]^T$ . This setting works well for many datasets, including Fastec, and does not require gravity direction measurement.", + "bbox": [ + 88, + 268, + 482, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-view. In this case, select $m$ scanlines in the same image and estimate the pose between them. We detect lines with DeepLSD [49], and select scanlines $y_{1},\\ldots ,y_{m}$ with spacing at least 25px with the highest number of shared lines. We obtain $p_{i,j}$ by intersecting lines with scanlines.", + "bbox": [ + 89, + 435, + 483, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the solvers in RANSAC [22]. We label a correspondence as inlier, if the triangulated line $\\mathbf{L}_j$ lies in front of all cameras and its reprojection error $\\varepsilon$ is below $1px$ . In the case of $(B,3,7)$ , we decompose the tensor into canonical cameras (20) before triangulation. Details about triangulation are in the SM. The score for a model is the sum over all lines of $(1px - \\varepsilon)^2$ if $\\varepsilon < 1px$ , and zero otherwise. We run RANSAC for 1000 iterations, and select the model with the highest score. We do not use local optimization.", + "bbox": [ + 89, + 511, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are summarized in Tables 2, 3 summarize the results for the solvers $(E,3,5)$ , $(E,4,4)$ , and $(D,3,7)$ , and Table 4 shows the tensor error for the solver $(B,3,7)$ . Although the total number of camera poses with an error below $10^{\\circ}$ and $20^{\\circ}$ remains relatively low, in Structure-from-Motion (SfM) applications, it is sufficient to estimate a single correct relative pose per sequence to initialize the reconstruction. Table 2 confirms that this is achievable for the vast majority of sequences. A more detailed evaluation and further experiments are provided in the SM.", + "bbox": [ + 89, + 648, + 483, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 89, + 814, + 362, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach shows promising results for rolling shutter relative pose estimation, several limitations remain. A key challenge is the lack of robust curve detectors and matchers, as rolling shutter distortion bends straight lines.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cbf416a8dfc6489a5b2e2f1685fb30b5ac7c36e0a48fb9ba4484e4d8a96b36d5.jpg", + "table_caption": [ + "Table 2. Number of sequences (out of 19) from the Fastec dataset [43], with at least one estimated pose with error below $5/10/20/30^{\\circ}$ . MV: Multi-view, SV: Single-view." + ], + "table_footnote": [], + "table_body": "
solver≤ 10°≤ 20°≤ 30°
MV(E,3,5)1.0 %6.2 %21.9 %
(E,4,4)1.9 %7.9 %22.4 %
(D,3,7)5.6 %13.8 %21.7 %
SV(E,3,5)0.5 %2.5 %5.3 %
(E,4,4)1.2 %4.1 %10.0 %
(D,3,7)0.5 %2.2 %6.6 %
", + "bbox": [ + 553, + 88, + 864, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/198797a15b67b996ca061593b94779fe2faed69db70f1daa5f1b1b4e24fbc67c.jpg", + "table_caption": [ + "Table 3. Percentage of camera poses from Fastec [43] with pose error below $10/20/30^{\\circ}$ . MV: Multi-view, SV: Single-view." + ], + "table_footnote": [], + "table_body": "
settingMedian errorMinimum error
MV0.850.44
SV0.890.62
", + "bbox": [ + 558, + 241, + 857, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Median and Minimum tensor error of the $(B,3,7)$ solver, averaged over all sequences from Fastec [43]. A more detailed evaluation is in the SM.", + "bbox": [ + 511, + 291, + 905, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, our method relies on chirality constraints to resolve ambiguity, as noted in [8]. The method does not use any motion model, which makes it broadly applicable and model-agnostic. However, this universality comes at the cost of stability: we expect methods that incorporate motion models to be more stable. Our solvers are also limited to specific scene types and assume a known gravity direction. However, this reliance could be mitigated using scene knowledge, such as geometric relations between lines, in combination with solver $(B,3,7)$ . Finally, our method estimates pose only up to a shift along the line direction. Future work could integrate additional geometric constraints, improve feature detection, and integrate the solvers to a rolling-shutter SfM pipeline.", + "bbox": [ + 511, + 359, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 587, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new approach for estimating the relative pose of rolling shutter cameras using projections of lines onto multiple scanlines. This enables estimating relative pose of rolling shutter cameras without explicitly modelling camera motion. The proposed solvers provide an important building block for rolling-shutter SfM where each scanline has its pose computed independently, without relying on a global motion model.", + "bbox": [ + 511, + 613, + 905, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We identify minimal problems for different variations of the single-scanline relative pose problem, including cases with parallel lines, vertical lines, and known gravity direction. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors. To evaluate our approach, we conduct experiments on both synthetic data and real-world rolling shutter images from the Fastec dataset, considering scanlines from either the same image or different images. These experiments demonstrate that our solvers successfully estimate at least one correct pose in the majority of scenes, confirming their feasibility for SfM initialization.", + "bbox": [ + 511, + 733, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7150", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omar Ait-Aider and François Berry. Structure and kinematics triangulation with a rolling shutter stereo rig. In IEEE 12th International Conference on Computer Vision, ICCV 2009, Kyoto, Japan, September 27 - October 4, 2009, 2009.", + "[2] Omar Ait-Aider, Nicolas Andreff, Jean-Marc Lavest, and Philippe Martinet. Simultaneous object pose and velocity computation using a single view from a rolling shutter camera. In Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part II, 2006. 1", + "[3] Omar Ait-Aider, Adrien Bartoli, and Nicolas Andreff. Kinematics from lines in a single rolling shutter image. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA, 2007. 1", + "[4] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. R6P - rolling shutter absolute pose problem. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, 2015. 1", + "[5] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. Rolling shutter absolute pose problem with known vertical direction. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016.", + "[6] Cenek Albl, Zuzana Kukelova, Viktor Larsson, and Tomás Pajdla. Rolling shutter camera absolute pose. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1", + "[7] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomás Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 1", + "[8] Kalle Åström and Magnus Oskarsson. Solutions and ambiguities of the structure and motion problem for 1d retinal vision. J. Math. Imaging Vis., 2000, 2, 6, 8", + "[9] Fang Bai, Agniva Sengupta, and Adrien Bartoli. Scanline homographies for rolling-shutter plane absolute pose. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 2022. 1, 2", + "[10] Akash Bapat, Enrique Dunn, and Jan-Michael Frahm. Towards kilo-hertz 6-dof visual tracking using an egocentric cluster of rolling shutter cameras. IEEE Trans. Vis. Comput. Graph., 2016. 2", + "[11] Akash Bapat, True Price, and Jan-Michael Frahm. Rolling shutter and radial distortion are features for high frame rate multi-camera tracking. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 2", + "[12] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. StereoGlue: Robust estimation with single-point solvers. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 8" + ], + "bbox": [ + 93, + 143, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Paul Breiding and Sascha Timme. Homotopycontinuation.jl: A package for homotopy continuation in julia. In Mathematical Software - ICMS 2018 - 6th International Conference, South Bend, IN, USA, July 24-27, 2018, Proceedings. Springer, 2018. 3", + "[14] Mingdeng Cao, Zhihang Zhong, Jiahao Wang, Yinqiang Zheng, and Yujiu Yang. Learning adaptive warping for real-world rolling shutter correction. CoRR, 2022. 1", + "[15] Yuchao Dai, Hongdong Li, and Laurent Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1", + "[16] Juan Carlos Dibene, Yazmín Maldonado, Leonardo Trujillo, and Enrique Dunn. Prepare for ludicrous speed: Marker-based instantaneous binocular rolling shutter localization. IEEE Trans. Vis. Comput. Graph., 2022. 2", + "[17] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. PLMP - point-line minimal problems in complete multiview visibility. IEEE Trans. Pattern Anal. Mach. Intell., 2024, 3", + "[18] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 6", + "[19] Bin Fan, Yuchao Dai, and Ke Wang. Rolling-shutter-stereo-aware motion estimation and image correction. Comput. Vis. Image Underst., 2021. 1", + "[20] Bin Fan, Yuchao Dai, Zhiyuan Zhang, and Ke Wang. Differential sfm and image correction for a rolling shutter stereog. Image Vis. Comput., 2022. 1", + "[21] Bin Fan, Yuchao Dai, and Hongdong Li. Rolling shutter inversion: Bring rolling shutter images to high framerate global shutter video. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1", + "[22] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 1981. 8", + "[23] Per-Erik Forssén and Erik Ringaby. Rectifying rolling shutter video from hand-held devices. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, 2010. 1", + "[24] Marvin Anas Hahn, Kathlen Kohn, Orlando Marigliano, and Tomás Pajdla. Order-one rolling shutter cameras. CoRR, 2024. 1", + "[25] Michiel Hazewinkel. Encyclopaedia of mathematics. 2002. 1", + "[26] Johan Hedborg, Erik Ringaby, Per-Erik Forssén, and Michael Felsberg. Structure and motion estimation from rolling shutter video. In IEEE International Conference on Computer Vision Workshops, ICCV 2011 Workshops, Barcelona, Spain, November 6-13, 2011, 2011. 1" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7151", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Johan Hedborg, Per-Erik Forssén, Michael Felsberg, and Erik Ringaby. Rolling shutter bundle adjustment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012, 2012. 1", + "[28] Radu Horaud, Roger Mohr, and Boguslaw Lorecki. On single-scanline camera calibration. IEEE Trans. Robotics Autom., 1993. 2", + "[29] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 5", + "[30] Chao Jia and Brian L. Evans. Probabilistic 3-d motion estimation for rolling shutter video rectification from visual and inertial measurements. In 14th IEEE International Workshop on Multimedia Signal Processing, MMSP 2012, Banff, AB, Canada, September 17-19, 2012, 2012. 1", + "[31] Georg Klein and David William Murray. Parallel tracking and mapping on a camera phone. In Science & Technology Proceedings, 8th IEEE International Symposium on Mixed and Augmented Reality 2009, ISMAR 2009, Orlando, Florida, USA, October 19-22, 2009, 2009. 1", + "[32] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, and Tomás Pajdla. Linear solution to the minimal absolute pose rolling shutter problem. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III, 2018. 1", + "[33] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, Konrad Schindler, and Tomás Pajdla. Minimal rolling shutter absolute pose with unknown focal length and radial distortion. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V, 2020. 1", + "[34] Yizhen Lao and Omar Ait-Aider. Rolling shutter homography and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2780-2793, 2021. 1", + "[35] Yizhen Lao, Omar Ait-Aider, and Helder Araujo. Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognit. Lett., 2018. 1", + "[36] Yizhen Lao, Omar Ait-Aider, and Adrien Bartoli. Rolling shutter pose and ego-motion estimation using shape-from-template. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part II, 2018. 1", + "[37] Chang-Ryeol Lee and Kuk-Jin Yoon. Monocular visual odometry with a rolling shutter camera. CoRR, 2017. 1", + "[38] Chang-Ryeol Lee and Kuk-Jin Yoon. Inertial-aided rolling shutter relative pose estimation. CoRR, 2017. 1", + "[39] Chang-Ryeol Lee, Ju Hong Yoon, Min-Gyu Park, and Kuk-Jin Yoon. Gyroscope-aided relative pose estimation for rolling shutter cameras. CoRR, 2019. 1", + "[40] Dongdong Li, Gongjian Wen, Bing Wei Hui, Shaohua Qiu, and Wenfei Wang. Cross-ratio invariant based line scan camera geometric calibration with static linear data. Optics and Lasers in Engineering, 2014. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Moyang Li, Peng Wang, Lingzhe Zhao, Bangyan Liao, and Peidong Liu. Usb-nerf: Unrolling shutter bundle adjusted neural radiance fields. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024. 1", + "[42] Bangyan Liao, Delin Qu, Yifei Xue, Huiqing Zhang, and Yizhen Lao. Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 1", + "[43] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, 2, 7, 8", + "[44] Ludovic Magerand, Adrien Bartoli, Omar Ait-Aider, and Daniel Pizarro. Global optimization of object pose and motion from a single rolling shutter image with automatic 2d-3d matching. In Computer Vision - ECCV 2012 - 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I, 2012. 1", + "[45] Marci Meingast, Christopher Geyer, and Shankar Sastry. Geometric models of rolling-shutter cameras. CoRR, abs/cs/0503076, 2005. 1", + "[46] Branislav Micusik and Georgios Evangelidis. Renormalization for initialization of rolling shutter visual-inertial odometry. Int. J. Comput. Vis., 2021. 1", + "[47] Thanh-Tin Nguyen and Maxime Lhuillier. Adding synchronization and rolling shutter in multi-camera bundle adjustment. In Proceedings of the British Machine Vision Conference 2016, BMVC 2016, York, UK, September 19-22, 2016, 2016. 1", + "[48] David Nister. An efficient solution to the five-point relative pose problem. IEEE Trans. Pattern Anal. Mach. Intell., 2004. 1", + "[49] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 8", + "[50] Rémi* Pautrat, Iago* Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. GlueStick: Robust image matching by sticking points and lines together. In International Conference on Computer Vision (ICCV), 2023. 8", + "[51] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018, Lake Tahoe, NV, USA, March 12-15, 2018, 2018. 1", + "[52] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in Manhattan world. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1", + "[53] Delin Qu, Yizhen Lao, Zhigang Wang, Dong Wang, Bin Zhao, and Xuelong Li. Towards nonlinear-motion-aware and occlusion-robust rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + ], + "bbox": [ + 514, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7152", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1", + "[55] Long Quan and Takeo Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Trans. Pattern Anal. Mach. Intell., 1997. 2, 3, 5", + "[56] Vijay Rengarajan, A. N. Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1", + "[57] Erik Ringaby and Per-Erik Forssén. Efficient video rectification and stabilisation for cell-phones. Int. J. Comput. Vis., 2012. 1", + "[58] Olivier Saurer, Kevin Koser, Jean-Yves Bouguet, and Marc Pollefeys. Rolling shutter stereo. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 1", + "[59] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. A minimal solution to the rolling shutter pose estimation problem. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2015, Hamburg, Germany, September 28 - October 2, 2015, 2015. 1", + "[60] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. Sparse to dense 3d reconstruction from rolling shutter images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1", + "[61] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 7", + "[62] David Schubert, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. Direct sparse odometry with rolling shutter. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VIII, 2018. 1", + "[63] Wei Shang, Dongwei Ren, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised learning to bring dual reversed rolling shutter images alive. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1", + "[64] Ken Shoemake. Animating rotation with quaternion curves. In SIGGRAPH, 1985. 1", + "[65] Henrik Stewénius, David Nister, Magnus Oskarsson, and Kalle Åström. Solutions to minimal generalized relative pose problems. Workshop on omnidirectional vision, 2005. 2", + "[66] Daobilige Su, Asher Bender, and Salah Sukkarieh. Improved cross-ratio invariant-based intrinsic calibration of A hyperspectral line-scan camera. Sensors, 2018. 2", + "[67] Yufen Sun, Gang Liu, and Yue Sun. An affine motion model for removing rolling shutter distortions. IEEE Signal Process. Lett., 2016. 1", + "[68] Subeesh Vasu, Mahesh Mohan M. R., and A. N. Rajagopalan. Occlusion-aware rolling shutter rectification of 3d scenes. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Carlos Andres Luna Vázquez, Manuel Mazo, José Luis Lázaro, and Juan F. Vázquez. Calibration of line-scan cameras. IEEE Trans. Instrum. Meas., 2010. 2", + "[70] Ke Wang, Bin Fan, and Yuchao Dai. Relative pose estimation for stereo rolling shutter cameras. In IEEE International Conference on Image Processing, ICIP 2020, Abu Dhabi, United Arab Emirates, October 25-28, 2020, 2020. 1", + "[71] Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. Urs-nerf: Unordered rolling shutter bundle adjustment for neural radiance fields. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV, 2024. 1", + "[72] Weilong Yan, Robby T. Tan, Bing Zeng, and Shuaicheng Liu. Deep homography mixture for single image rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1", + "[73] Ming Yao, Zuyun Zhao, and Bugao Xu. Geometric calibration of line-scan camera using a planar pattern. J. Electronic Imaging, 2014. 2", + "[74] Yongcong Zhang, Bangyan Liao, Yifei Xue, Chen Lu, Peidong Liu, and Yizhen Lao. RSL-BA: rolling shutter line bundle adjustment. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LIX, 2024. 1", + "[75] Bingbing Zhuang and Quoc-Huy Tran. Image stitching and rectification for hand-held cameras. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, 2020. 1", + "[76] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1", + "[77] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1" + ], + "bbox": [ + 516, + 92, + 903, + 627 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7153", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_model.json b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2d24a706687cdac3957e77c0bdf2c7cdca9e1f5e --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_model.json @@ -0,0 +1,3687 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.182, + 0.362, + 0.199 + ], + "angle": 0, + "content": "Petr Hruby" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.2, + 0.367, + 0.215 + ], + "angle": 0, + "content": "ETH Zürich" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.438, + 0.233 + ], + "angle": 0, + "content": "Rämistrasse 101, 8006 Zürich" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.238, + 0.416, + 0.251 + ], + "angle": 0, + "content": "petr.h Ruby@inf.ethz.ch" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.182, + 0.705, + 0.199 + ], + "angle": 0, + "content": "Marc Pollefeys" + }, + { + "type": "text", + "bbox": [ + 0.488, + 0.2, + 0.8, + 0.217 + ], + "angle": 0, + "content": "ETH Zürich / Microsoft Spatial AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.218, + 0.764, + 0.233 + ], + "angle": 0, + "content": "Rämistrasse 101, 8006 Zürich" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.238, + 0.758, + 0.251 + ], + "angle": 0, + "content": "marc.pollefeys@inf.ethz.ch" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.651 + ], + "angle": 0, + "content": "We propose a novel approach for estimating the relative pose between rolling shutter cameras using the intersections of line projections with a single scanline per image. This allows pose estimation without explicitly modeling camera motion. Alternatively, scanlines can be selected within a single image, enabling single-view relative pose estimation for scanlines of rolling shutter cameras. Our approach is designed as a foundational building block for rolling shutter structure-from-motion (SfM), where no motion model is required, and each scanline's pose can be computed independently. We classify minimal solvers for this problem in both generic and specialized settings, including cases with parallel lines and known gravity direction, assuming known intrinsic and no lens distortion. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors, by leveraging connections between this problem and the estimation of 2D structure from 1D cameras. Experiments on rolling shutter images from the Fastec dataset demonstrate the feasibility of our approach for initializing rolling shutter SfM, highlighting its potential for further development. The code will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.222, + 0.696 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Relative pose estimation is a fundamental problem in computer vision with applications in Structure from Motion [61], SLAM, multi-view stereo, and visual odometry. This paper focuses on estimating relative pose between rolling shutter (RS) cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Rolling shutter cameras capture images row by row. If the camera is static during capture, the resulting image is equivalent to that of a pinhole camera. However, if the camera moves, the rolling shutter effect distorts the image, making standard pinhole-based methods [48] unreliable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A common approach to handle rolling shutter distortion is modeling camera motion with parametric models such as SLERP [64], Cayley transformation [25], linearized" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.289, + 0.648, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.29, + 0.764, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.29, + 0.881, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.363, + 0.822, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.458, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Figure 1. We estimate a relative pose from the intersections of the line projections with a single scanline per image (green). This allows us to estimate relative pose between rolling shutter cameras without modelling the camera motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.669 + ], + "angle": 0, + "content": "rotation [4], affine motion [67], pushbroom [46], accelerated motion [19], or order-one motion [24]. This has been successfully applied to absolute pose estimation [2-6, 9, 32, 33, 36, 44, 45, 59], bundle adjustment [27, 41, 42, 47, 60, 62, 71, 74], SfM [26, 31, 58], and motion rectification from single cameras [35, 51, 52, 56, 72, 77], camera rigs [1, 7, 20, 63], and videos [14, 19, 21, 23, 30, 53, 54, 57, 67, 68, 75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.717 + ], + "angle": 0, + "content": "While most methods use point-based features, some employ lines for absolute pose estimation [3], bundle adjustment [74], and motion rectification [35, 51, 52, 56]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The problem of estimating relative pose for rolling shutter cameras has been explored in prior work. Special essential matrices for RS cameras under different motion models were introduced in [15], along with linear non-minimal solvers. The problem was also solved for pure translation models with constant velocity and acceleration [76]. IMU-based approaches reduced the required number of point correspondences from 11 to 5 [38, 39]. Other methods first estimate pose using a global shutter solver before refining with an RS model [37]. Additional works estimate relative pose for RS camera rigs [70] and for planar scenes [34]. Recently, [24] classified RS relative pose minimal problems" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7143" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "under the order-one motion model. Despite this, the general RS relative pose problem remains unsolved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.227 + ], + "angle": 0, + "content": "One drawback of parametric motion models is their reliance on an assumed motion that may not match real-world trajectories. An alternative is estimating pose independently for each scanline, similar to line-camera calibration [28, 40, 66, 69, 73]. This approach has been used for tracking a single camera [10, 11] and a camera rig [16] with respect to a predefined marker." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.392 + ], + "angle": 0, + "content": "We explore the feasibility of estimating relative pose between scanlines of RS cameras, without explicit motion modelling. The goal is to use this method as a building block to initialize rolling-shutter SfM where no motion model is required and each scanline can have its pose independently computed. To the best of our knowledge, this is the first work that enables this. After the initialization, scanlines from remaining images may be registered by solving the generalized relative pose problem [65], similarly to [10, 11, 16], or, in the case of a planar scene, using [9]. Our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.394, + 0.483, + 0.438 + ], + "angle": 0, + "content": "- Formulation of the problem of relative pose estimation from line projections in a single scanline per image, assuming known intrinsics and no lens distortion." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.439, + 0.483, + 0.469 + ], + "angle": 0, + "content": "- Enumeration of minimal problems for general and specialized settings with parallel lines and gravity priors." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.47, + 0.484, + 0.529 + ], + "angle": 0, + "content": "- Demonstration that the problems with parallel lines and with vertical lines and gravity prior correspond to estimating relative pose of uncalibrated cameras [55], respectively calibrated [8] 1D camera pose." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.483, + 0.544 + ], + "angle": 0, + "content": "- Development of minimal solvers for the following cases:" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.394, + 0.484, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.546, + 0.481, + 0.559 + ], + "angle": 0, + "content": "- 3 cameras, 7 parallel lines (with projective ambiguity)." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.561, + 0.388, + 0.574 + ], + "angle": 0, + "content": "- 3 cameras, gravity prior, 7 parallel lines." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.576, + 0.388, + 0.589 + ], + "angle": 0, + "content": "- 3 cameras, gravity prior, 5 vertical lines." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.591, + 0.388, + 0.604 + ], + "angle": 0, + "content": "- 4 cameras, gravity prior, 4 vertical lines." + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.546, + 0.481, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.605, + 0.483, + 0.651 + ], + "angle": 0, + "content": "- Evaluation of solvers on synthetic and real RS datasets (Fastec, [43]), taking scanlines either from different images or from a single image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.195, + 0.673 + ], + "angle": 0, + "content": "1.1. Notation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.787 + ], + "angle": 0, + "content": "We use bold letters for matrices and vectors, and normal font for scalars. Indexing starts from one, with indices in brackets; e.g., the first element of \\(\\mathbf{L}_d\\) is \\(\\mathbf{L}_{d(1)}\\). Matrix indexing follows Matlab style, where \\(\\mathbf{A}_{(:,1:2)}\\) selects the first two rows of \\(\\mathbf{A}\\). The scanlines are denoted with a scalar \\(y \\in \\mathbb{R}\\) giving their y-coordinate. The set of all 3D lines is denoted as \\(Gr(1,3)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.297, + 0.815 + ], + "angle": 0, + "content": "2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.871 + ], + "angle": 0, + "content": "This section describes the geometry of a line projected into a rolling-shutter camera and derives an algebraic constraint useful for estimating camera poses." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Consider \\( n \\) rolling-shutter cameras capturing the same 3D scene. Due to motion, the cameras capture different" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.089, + 0.843, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.199, + 0.907, + 0.242 + ], + "angle": 0, + "content": "Figure 2. Illustration of the line projection. Line \\(\\mathbf{L}\\) projects onto scanline \\(\\mathbf{y}\\) at point \\(\\mathbf{p}\\). Then, the ray \\(\\mathbf{r}\\) obtained by backprojecting \\(\\mathbf{p}\\) intersects \\(\\mathbf{L}\\), yielding the constraint (3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.298 + ], + "angle": 0, + "content": "rows at different poses. Let \\(\\mathbf{R}_i(y) \\in SO(3)\\) and \\(\\mathbf{C}_i(y) \\in \\mathbb{R}^3\\) denote the orientation and center of camera \\(i\\) at row \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.905, + 0.404 + ], + "angle": 0, + "content": "We aim to estimate the relative pose between the cameras without explicitly modeling their motion functions. To achieve this, we select a single row \\( y_{i} \\) per camera and measure intersections between these selected rows and the projections of \\( m \\) straight lines \\( \\mathbf{L}_j \\in Gr(1,3) \\). Then, we use this information to recover the camera poses \\( \\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i) \\). Figure 1 illustrates the setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.905, + 0.45 + ], + "angle": 0, + "content": "Let \\(\\mathbf{p}_{i,j} \\in \\mathbb{P}^2\\), with \\(p_{i,j} = [x_{i,j} y_i 1]^T\\), denote the projection of line \\(\\mathbf{L}_j\\) onto row \\(y_i\\) of camera \\(i\\). Then, the back-projected ray \\(\\mathbf{r}_{i,j}\\) from \\(\\mathbf{p}_{i,j}\\) intersects line \\(\\mathbf{L}_j\\), giving:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.459, + 0.905, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {i, j} = \\mathbf {C} _ {i} (y _ {i}) + \\lambda_ {i, j} \\mathbf {R} _ {i} (y _ {i}) ^ {T} \\mathbf {p} _ {i, j}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.487, + 0.905, + 0.534 + ], + "angle": 0, + "content": "This situation is depicted in Fig. 2. Let us represent \\(\\mathbf{L}_j\\) with a pair \\((\\mathbf{L}_{0,j},\\mathbf{L}_{d,j})\\), where \\(\\mathbf{L}_{0,j}\\) is a point on the line and \\(\\mathbf{L}_{d,j}\\) is its direction. The ray \\(\\mathbf{r}_{i,j}\\) intersects \\(\\mathbf{L}_j\\) if and only if:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.542, + 0.905, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\exists \\alpha , \\beta \\in \\mathbb {R}: \\quad \\mathbf {C} _ {i} \\left(y _ {i}\\right) + \\alpha \\mathbf {R} _ {i} \\left(y _ {i}\\right) ^ {T} \\mathbf {p} _ {i, j} = \\mathbf {L} _ {0, j} + \\beta \\mathbf {L} _ {d, j}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.905, + 0.635 + ], + "angle": 0, + "content": "This is equivalent to requiring that vectors \\(\\mathbf{R}_i(y_i)^T\\mathbf{p}_{i,j}\\), \\(\\mathbf{L}_{d,j}\\), and \\(\\mathbf{L}_{0,j} - \\mathbf{C}_i(y_i)\\) are linearly dependent. Since three 3D vectors are linearly dependent if and only if their scalar triple product vanishes, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.645, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {L} _ {d, j} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Each projection provides a single constraint on the unknowns (lines and camera poses). A 3D line has four degrees of freedom, requiring four scanlines for triangulation and at least five scanlines for constraining relative pose." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.731, + 0.759 + ], + "angle": 0, + "content": "2.1. Assuming parallel lines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Let us consider the scenario where all lines are parallel, which frequently occurs in real-world settings. Here, all lines share a common direction \\(\\mathbf{L}_d \\in \\mathbb{R}^3\\). Since Eq. (3) is invariant to scene rotation, we align \\(\\mathbf{L}_d\\) with the y-axis, \\(\\mathbf{e}_2\\). This allows representing each line \\(\\mathbf{L}_j\\) with a single point \\(\\mathbf{L}_{0,j}\\) in the xz-plane, with its y-coordinate equal to zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "With a shared direction, each line has only two individual degrees of freedom (DoF). Once \\(\\mathbf{L}_d\\) is known, a line can be triangulated from just two cameras. This reduces the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7144" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "minimum number of cameras needed for pose estimation to three. To simplify Eq. (3), we substitute \\(\\mathbf{e}_2\\) for \\(\\mathbf{L}_{d,j}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.129, + 0.483, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} (y _ {i}) \\left(\\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (3)} \\\\ 0 \\\\ - \\mathbf {L} _ {0, j (1)} \\end{array} \\right] - \\left[ \\begin{array}{c} \\mathbf {C} _ {i} (y _ {i}) _ {(3)} \\\\ 0 \\\\ - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} \\end{array} \\right]\\right) = 0. \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.484, + 0.234 + ], + "angle": 0, + "content": "We define rotation \\(\\mathbf{R}_{0,i} \\in SO(3)\\) around the x-axis, which maps row \\(y_{i}\\) to \\(y = 0\\), and set \\(\\mathbf{p}_{i,j}' = \\mathbf{R}_{0,i}\\mathbf{p}_{i,j}\\). Then, there holds \\(\\mathbf{p}_{i,j}' = [x_{i,j}'01]^T\\). Defining" + }, + { + "type": "equation", + "bbox": [ + 0.318, + 0.242, + 0.483, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {i} ^ {\\prime} = \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.262, + 0.442, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {i} ^ {\\prime} = - \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}) [ \\mathbf {C} _ {i} (y _ {i}) _ {(3)} 0 - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} ] ^ {T},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.462, + 0.304 + ], + "angle": 0, + "content": "and substituting \\(\\mathbf{R}_i^{\\prime},\\mathbf{t}_i^{\\prime}\\) , and \\(\\mathbf{p}_{i,j}^{\\prime}\\) into Eq. (4), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.31, + 0.483, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l l} x _ {i, j} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} \\left[ \\begin{array}{c c c} - \\mathbf {R} _ {i (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i (1, 1)} ^ {\\prime} & \\mathbf {t} _ {i (0)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i (3, 1)} ^ {\\prime} & \\mathbf {t} _ {i (2)} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (1)} \\\\ \\mathbf {L} _ {0, j (3)} \\\\ 1 \\end{array} \\right] = 0. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.372, + 0.395, + 0.386 + ], + "angle": 0, + "content": "We can write Equation (6) in a vector form as" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.393, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {i, j} ^ {T} \\mathbf {A} _ {i} \\mathbf {L} _ {h, j} = 0, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.482 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}_i\\in \\mathbb{R}^{2,3}\\) is a matrix encoding the camera pose, \\(\\mathbf{u}_{i,j}\\in \\mathbb{P}^1\\) is a vector encoding the observation, and \\(\\mathbf{L}_{h,j}\\in\\) \\(\\mathbb{P}^2\\) is a homogeneous representation of point \\(\\mathbf{L}_{0,j}\\). Alternatively, we define \\(\\mathbf{u}_{i,j}^{\\prime} = [1 - x_{i,j}^{\\prime}]^{T}\\), and rewrite (7) as" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.493, + 0.483, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\mathbf {A} _ {i} \\mathbf {L} _ {j, h}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.635 + ], + "angle": 0, + "content": "Since constraint (7) is homogeneous, matrix \\(\\mathbf{A}_i\\) has 5 DoF. The elements \\(\\mathbf{R}_i(y_i)\\), \\(\\mathbf{C}_i(y_i)_{(1)}\\), \\(\\mathbf{C}_i(y_i)_{(3)}\\) of the pose, which appear in \\(\\mathbf{A}_i\\), also have 5 DoF. Using the Gröbner basis method, we have verified that for every matrix \\(\\mathbf{A} \\in \\mathbb{R}^{2,3}\\), there are 8 ways to decompose it into pose parameters \\(R \\in SO(3)\\), \\(C_{(1)} \\in \\mathbb{R}\\), \\(C_{(3)} \\in \\mathbb{R}\\), and a scale factor \\(\\sigma \\in \\mathbb{R}\\). This means that \\(\\mathbf{A}_i\\) has no internal constraints, and every \\(2 \\times 3\\) matrix can represent \\(\\mathbf{A}_i\\) for some pose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.726 + ], + "angle": 0, + "content": "Therefore, the relative pose problem can be transformed into finding matrices \\(\\mathbf{A}_i\\in \\mathbb{R}^{2,3}\\), vectors \\(\\mathbf{L}_{h,j}\\in \\mathbb{P}^2\\), and scalars \\(\\lambda_{i,j}\\in \\mathbb{R}\\) such that (8) holds for every \\(i\\in \\{1,\\dots,n\\}\\), \\(j\\in \\{1,\\dots,m\\}\\). This is equivalent to the problem of estimating 2D structure from images taken by \\(n\\) uncalibrated 1D cameras, which has been studied in [55]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.427, + 0.747 + ], + "angle": 0, + "content": "2.1.1. Ambiguities in the case with parallel lines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The assumption that all lines are parallel introduces ambiguities into the relative pose estimation. First, since all line directions \\(\\mathbf{L}_{d,j}\\) are equal, the values of constraints (3) remain unchanged if we shift \\(\\mathbf{C}_i(y_i)\\) to \\(\\mathbf{C}_i(y_i) + \\gamma \\mathbf{L}_{j,d}\\) for any \\(\\gamma \\in \\mathbb{R}\\). Therefore, it is only possible to estimate the camera centers up to translation along \\(\\mathbf{L}_d\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Furthermore, if (8) holds for some \\(\\mathbf{A}_i\\in \\mathbb{R}^{2,3}\\), \\(\\mathbf{L}_{h,j}\\in \\mathbb{P}^2\\), then (8) also holds for \\(\\mathbf{A}_i\\cdot \\mathbf{H}^{-1}\\), \\(\\mathbf{H}\\cdot \\mathbf{L}_{j,h}\\) for every invertible \\(\\mathbf{H}\\in \\mathbb{R}^{3,3}\\). Therefore, the cameras and the structure can only be estimated up to a projective transformation \\(\\mathbf{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.167 + ], + "angle": 0, + "content": "We show later in the paper that we can resolve the projective ambiguity by introducing priors on the cameras. To resolve the ambiguity in the camera centers, we can, for instance, observe one line orthogonal to \\(\\mathbf{L}_d\\) and use it to estimate the camera centers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.808, + 0.198 + ], + "angle": 0, + "content": "3. Enumerating minimal problems" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.906, + 0.251 + ], + "angle": 0, + "content": "In this section, we classify minimal problems for estimating the relative pose from projections of \\( n \\) lines onto \\( m \\) scanlines, along with their degrees. We consider five settings:" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.253, + 0.689, + 0.268 + ], + "angle": 0, + "content": "A. Fully generic setting" + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.269, + 0.636, + 0.281 + ], + "angle": 0, + "content": "B. Parallel lines" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.283, + 0.835, + 0.298 + ], + "angle": 0, + "content": "C. Gravity prior for cameras + Generic lines" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.299, + 0.834, + 0.313 + ], + "angle": 0, + "content": "D. Gravity prior for cameras + Parallel lines" + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.314, + 0.835, + 0.328 + ], + "angle": 0, + "content": "E. Gravity prior for cameras + Vertical lines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.906, + 0.421 + ], + "angle": 0, + "content": "In cases A and C, we estimate the pose up to global rotation, translation, and scale. In cases D and E, we estimate the pose up to these transformations and an additional shift of centers \\(\\mathbf{C}_i(y_i)\\) along direction \\(\\mathbf{L}_d\\). In case B, we estimate the pose up to these transformations, the shift of centers along \\(\\mathbf{L}_d\\), and the projective transformation \\(\\mathbf{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.527 + ], + "angle": 0, + "content": "For each setting, we determine the number of cameras \\( m \\) and the number of lines \\( n \\) s.t. the problem (3) is minimal. We assume complete visibility, meaning each line is projected into each camera. We follow concepts from [17]: Minimal problem [17, Def. 1]: A parametric polynomial system \\( f(\\theta ,\\mathbf{x}) = 0 \\) with a finite nonzero number of solutions \\( \\mathbf{x}^* \\) for a generic set of parameters \\( \\theta \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Balanced problem [17, Def. 2]: A parametric polynomial system \\( f(\\theta, \\mathbf{x}) = 0 \\) where the dimensionality of the space of variables \\( \\mathbf{x} \\) equals the number of independent equations. Condition for minimal problems [17, Step. 1, Corr. 2]: A problem \\( f(\\theta, \\mathbf{x}) = 0 \\) is minimal if and only if it is balanced and its Jacobian \\( \\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}} \\) is invertible for generic \\( \\theta, \\mathbf{x}^* \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.74 + ], + "angle": 0, + "content": "This summary provides the foundation for enumerating minimal problems. For detailed theory, refer to [17]. In our case, the parameter vector \\(\\theta\\) consists of projections \\(\\mathbf{p}_{i,j} \\in \\mathbb{P}^2\\), and the variable vector \\(\\mathbf{x}\\) encodes camera poses \\((\\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i))\\) and lines \\((\\mathbf{L}_{0,j}, \\mathbf{L}_{d,j})\\). The polynomial system \\(f(\\mathbf{x}, \\theta)\\) consists of equations in the form of (3). Following [17], we used the following steps to list all minimal problems and determine their degrees:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.743, + 0.892, + 0.757 + ], + "angle": 0, + "content": "- Enumerate all balanced problems for each setting A-E." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.758, + 0.905, + 0.802 + ], + "angle": 0, + "content": "- For each balanced problem, generate random camera poses and lines \\(\\mathbf{x}^*\\), reproject the lines into the cameras, and obtain projections \\(\\theta\\) such that \\(f(\\theta, \\mathbf{x}^*) = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.802, + 0.905, + 0.839 + ], + "angle": 0, + "content": "- Compute the Jacobian \\(\\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}}\\). Label the problem as minimal if \\(\\det \\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}} \\neq 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.839, + 0.905, + 0.868 + ], + "angle": 0, + "content": "- Use monodromy computation from [13] to determine the degree of minimal problems." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.743, + 0.905, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "According to algebraic geometry, finding a single instance where the Jacobian is full-rank suffices to claim generic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7145" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.485, + 0.208 + ], + "angle": 0, + "content": "
S.mnMin.DegS.mnMin.Deg
A523Y389k+D37Y48
A217Y40k+D45Y232
B37Y2*D64Y1224
B46Y2*E35Y16
C515Y1.8M+E44Y32
C155Y532k+
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.21, + 0.484, + 0.281 + ], + "angle": 0, + "content": "Table 1. List of balanced problems for settings A-E. S. denotes the setting, \\( m \\) and \\( n \\) are the number of cameras and lines, and Deg is the degree. Column Min. contains Y if the problem is minimal. \\( 2^{*} \\) indicates two solutions in terms of \\( 2 \\times 3 \\) matrices, and + denotes computations that did not terminate. See Sec. 3 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.336 + ], + "angle": 0, + "content": "minimality. In the following subsections, we enumerate minimal problems for each setting A-E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.428 + ], + "angle": 0, + "content": "A. Fully Generic Setting Each camera has 6 degrees of freedom (DoF), while each line has 4 DoF. Due to gauge freedom, we fix 7 DoF (3 for rotation, 3 for translation, and 1 for scale). There are \\( m \\cdot n \\) constraints in the form of (3), fixing \\( m \\cdot n \\) DoF. The balanced constraint is therefore:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.436, + 0.483, + 0.452 + ], + "angle": 0, + "content": "\\[\n6 m + 4 n - 7 = m n. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.492 + ], + "angle": 0, + "content": "We define a function \\( n_A(m) \\), mapping the number of cameras \\( m \\) to the number of lines \\( n \\), s.t. (9) holds for \\( m \\), \\( n_A(m) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.5, + 0.483, + 0.531 + ], + "angle": 0, + "content": "\\[\nn _ {A} (m) = \\frac {6 m - 7}{m - 4}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.624 + ], + "angle": 0, + "content": "As stated in Sec. 2, the minimal number of cameras is \\( m \\geq 5 \\). The function \\( n_A(m) \\) is strictly decreasing for \\( m > 4 \\), and since \\( \\lim_{m \\to \\infty} n_A(m) = 6 \\), no balanced problem has fewer than 7 lines. Evaluating \\( n_A(m) \\) for integer values of \\( m \\) from 5 onward, and searching for integer values of \\( n_A(m) \\), we identify the following balanced cases:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.626, + 0.336, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\bullet m = 5, n = 2 3; \\quad m = 2 1, n = 7\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.763 + ], + "angle": 0, + "content": "B. Parallel Lines As discussed in Sec. 2, the pose is estimated up to a shift along \\(\\mathbf{L}_d\\) and a projective transformation \\(\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}\\). The pose estimation corresponds to estimating matrices \\(\\mathbf{A}_i \\in \\mathbb{R}^{2 \\times 3}\\) and points \\(\\mathbf{L}_{h,j} \\in \\mathbb{P}^2\\) such that (7) holds for all \\(i, j\\). Here, each camera has 5 DoF, each line has 2 DoF, and we fix 8 DoF due to projective ambiguity. The balanced constraint is:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.773, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\n5 m + 2 n - 8 = m n. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.797, + 0.228, + 0.812 + ], + "angle": 0, + "content": "Defining \\(n_B(m)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.819, + 0.483, + 0.85 + ], + "angle": 0, + "content": "\\[\nn _ {B} (m) = \\frac {5 m - 8}{m - 2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "and noting that \\(m \\geq 3\\), we find \\(\\lim_{m \\to \\infty} n_B(m) = 5\\), implying no balanced problem has fewer than 6 lines. Evaluating \\(n_B(m)\\) for integer \\(m\\), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.092, + 0.741, + 0.106 + ], + "angle": 0, + "content": "\\[\n\\bullet m = 3, n = 7; \\quad m = 4, n = 6\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.907, + 0.199 + ], + "angle": 0, + "content": "C. Gravity Prior for Cameras + Generic Lines With a known vertical direction, each camera has 4 DoF (1 for rotation, 3 for translation), and we fix 5 DoF due to gauge freedom. Lines remain generic, contributing 4 DoF each. The balanced constraint is:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.215, + 0.907, + 0.231 + ], + "angle": 0, + "content": "\\[\n4 m + 4 n - 5 = m n. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.235, + 0.722, + 0.25 + ], + "angle": 0, + "content": "The corresponding function is:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.255, + 0.906, + 0.285 + ], + "angle": 0, + "content": "\\[\nn _ {C} (m) = \\frac {4 m - 5}{m - 4}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.291, + 0.906, + 0.321 + ], + "angle": 0, + "content": "We find \\(\\lim_{m\\to \\infty}n_C(m) = 4\\) so no balanced problem has fewer than 5 lines. Checking integer values of \\(m\\), we get:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.322, + 0.756, + 0.335 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bullet m = 5, n = 1 5; \\quad m = 1 5, n = 5 \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.43 + ], + "angle": 0, + "content": "D. Gravity Prior for Cameras + Parallel Lines Here, camera rotation has 1 DoF, and the camera center has 2 DoF (up to translation along the line). We fix 4 DoF due to gauge freedom. Lines share 2 DoF for direction \\(\\mathbf{L}_d\\), with each line having 2 additional DoF. The balanced constraint is:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.437, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\[\n3 m + 2 + 2 n - 4 = m n. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.46, + 0.641, + 0.475 + ], + "angle": 0, + "content": "We get \\( n_D(m) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.48, + 0.905, + 0.51 + ], + "angle": 0, + "content": "\\[\nn _ {D} (m) = \\frac {3 m - 2}{m - 2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.546 + ], + "angle": 0, + "content": "For \\(m \\geq 3\\), we find \\(\\lim_{m \\to \\infty} n_D(m) = 3\\), meaning no balanced problem has fewer than 4 lines. We have found:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.548, + 0.855, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\bullet m = 3, n = 7; \\quad m = 4, n = 5; \\quad m = 6, n = 4\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.64 + ], + "angle": 0, + "content": "E. Gravity Prior for Cameras + Vertical Lines Similar to the previous case, each camera has 3 DoF, and we fix 4 DoF due to gauge freedom. Since the direction \\(\\mathbf{L}_d\\) is fixed, each line has 2 DoF. The balanced constraint is:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.656, + 0.905, + 0.671 + ], + "angle": 0, + "content": "\\[\n3 m + 2 n - 4 = m n. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.676, + 0.66, + 0.691 + ], + "angle": 0, + "content": "We define \\(n_E(m)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.696, + 0.905, + 0.727 + ], + "angle": 0, + "content": "\\[\nn _ {E} (m) = \\frac {3 m - 4}{m - 2}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.762 + ], + "angle": 0, + "content": "For \\(m \\geq 3\\), we find \\(\\lim_{m \\to \\infty} n_E(m) = 3\\), meaning no balanced problem has fewer than 4 lines. We found problems:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.764, + 0.741, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\bullet m = 3, n = 5; \\quad m = 4, n = 4\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "All balanced problems for these settings are summarized in Table 1, along with their minimality status and degree. The table shows that problems with parallel/vertical lines (B, D, E) are minimal and computationally feasible. Problems with generic lines (A, C) are minimal but have degrees too high for practical RANSAC solutions. For problems \\((A,5,23)\\) and \\((C,5,15)\\), monodromy computations did not terminate within a month." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7146" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.251, + 0.106 + ], + "angle": 0, + "content": "4. Minimal solvers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "In this section, we introduce solvers for the minimal problems \\((B,3,7)\\), \\((D,3,7)\\), \\((E,3,5)\\), \\((E,4,4)\\) from Section 3. We use the notation and concepts from Sec. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.169, + 0.394, + 0.185 + ], + "angle": 0, + "content": "4.1. Parallel lines with generic cameras" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.484, + 0.298 + ], + "angle": 0, + "content": "Here, we describe the solution to the problem \\((B,3,7)\\). Since the lines are parallel, our goal is to find matrices \\(\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3\\in \\mathbb{R}^{2\\times 3}\\) and vectors \\(\\mathbf{L}_{h,j}\\) for \\(j\\in \\{1,\\dots ,7\\}\\), such that equation (7) holds for every camera \\(i\\) and line \\(j\\). Due to projective ambiguity, we aim to recover the camera poses up to a common projective transformation \\(\\mathbf{H}\\in \\mathbb{R}^{3\\times 3}\\). Each camera triplet \\(\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.307, + 0.484, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {1} = \\mathbf {A} _ {c, 1} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {2} = \\mathbf {A} _ {c, 2} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {3} = \\mathbf {A} _ {c, 3} \\cdot \\mathbf {H}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.331, + 0.483, + 0.347 + ], + "angle": 0, + "content": "for some \\(\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}\\), with \\(\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}\\) taking the form:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.369, + 0.499, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {c, 1} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {1} & \\alpha_ {1} & \\alpha_ {1} \\end{array} \\right], \\mathbf {A} _ {c, 2} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {2} & \\alpha_ {3} & \\alpha_ {4} \\end{array} \\right], \\mathbf {A} _ {c, 3} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {5} & \\alpha_ {6} & \\alpha_ {7} \\end{array} \\right]. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.482, + 0.431 + ], + "angle": 0, + "content": "The proof follows analogously to Proposition 2 in [29]. These canonical-form matrices have 7 DoF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.492 + ], + "angle": 0, + "content": "To estimate \\(\\mathbf{A}_{c,1},\\mathbf{A}_{c,2},\\mathbf{A}_{c,3}\\), we leverage the fact that, as shown in [55], a trifocal tensor exists between three camera matrices \\(\\mathbf{A}_i\\in \\mathbb{R}^{2\\times 3}\\). To compute this tensor, we stack the constraints from (8) into the matrix equation:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.5, + 0.482, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c c} \\mathbf {A} _ {1} & \\mathbf {u} _ {1, j} ^ {\\prime} & 0 & 0 \\\\ \\mathbf {A} _ {2} & 0 & \\mathbf {u} _ {2, j} ^ {\\prime} & 0 \\\\ \\mathbf {A} _ {3} & 0 & 0 & \\mathbf {u} _ {3, j} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {h, j} \\\\ - \\lambda_ {1, j} \\\\ - \\lambda_ {2, j} \\\\ - \\lambda_ {3, j} \\end{array} \\right] = 0. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.601 + ], + "angle": 0, + "content": "For this equation to hold, the determinant of the coefficient matrix must vanish, leading to the constraint:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.609, + 0.482, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\sum_ {a = 1} ^ {2} \\sum_ {b = 1} ^ {2} \\sum_ {c = 1} ^ {2} \\mathbf {u} _ {1, j (a)} ^ {\\prime} \\mathbf {u} _ {2, j (b)} ^ {\\prime} \\mathbf {u} _ {3, j (c)} ^ {\\prime} \\mathbf {T} _ {(a, b, c)} = 0, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.659, + 0.46, + 0.675 + ], + "angle": 0, + "content": "where \\(\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}\\) is the trifocal tensor with elements:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.679, + 0.482, + 0.717 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {(a, b, c)} = (- 1) ^ {a + b + c} \\det \\left(\\left[ \\mathbf {A} _ {1 (a,:)} ^ {T} \\mathbf {A} _ {2 (b,:)} ^ {T} \\mathbf {A} _ {3 (c,:)} ^ {T} \\right]\\right). \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.748 + ], + "angle": 0, + "content": "As shown in [55], \\(\\mathbf{T}\\) has no internal constraints and it can be estimated using 7 line correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Since \\(\\mathbf{T}\\) is homogeneous, it has 7 DoF, matching the DoF of the canonical camera matrices. We verified that for every tensor \\(\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}\\), there are exactly two corresponding triplets of matrices \\(\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}\\). The proof and the decomposition algorithm for recovering these matrices can be found in the SM. Once \\(\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}\\) are found, we can triangulate the lines up to the projective transformation \\(\\mathbf{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.854, + 0.483, + 0.883 + ], + "angle": 0, + "content": "The solver for the \\((B,3,7)\\) problem, which estimates the trifocal tensor \\(\\mathbf{T}\\), operates as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.424, + 0.902 + ], + "angle": 0, + "content": "1. Input: Projections \\(\\mathbf{p}_{i,j}\\) of 7 lines in 3 cameras." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.862, + 0.109 + ], + "angle": 0, + "content": "2. Compute matrices \\(\\mathbf{R}_{0,i}\\) and vectors \\(\\mathbf{u}_{i,j}\\) and \\(\\mathbf{u}_{i,j}^{\\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.753, + 0.122 + ], + "angle": 0, + "content": "3. Construct the linear system (22)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.15 + ], + "angle": 0, + "content": "4. Solve (22) using singular value decomposition (SVD) to estimate the tensor \\(\\mathbf{T}\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.154, + 0.771, + 0.169 + ], + "angle": 0, + "content": "On average, the solver runs in \\(6.92\\mu s\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.177, + 0.799, + 0.193 + ], + "angle": 0, + "content": "4.2. Vertical lines and known gravity" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.244 + ], + "angle": 0, + "content": "Here, we analyze the solutions to the problems \\((E,3,5)\\) and \\((E,4,4)\\). We assume that all lines are vertical and the gravity directions \\(\\mathbf{v}_i\\in \\mathbb{P}^2\\) for the cameras are known." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.906, + 0.364 + ], + "angle": 0, + "content": "Similar to Sec. 4.1, we set \\(\\mathbf{L}_{d,j} = \\mathbf{e}_2\\). Therefore, Equation (6) still holds. Furthermore the vanishing points \\(\\mathbf{v}_i\\) for \\(i \\in \\{1,\\dots,m\\}\\) associated with the vertical lines are known, and there holds \\(\\mathbf{v}_i = \\mathbf{R}_i(y_i)\\mathbf{e}_2\\). We can decompose \\(\\mathbf{R}_i(y_i) = \\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\), where \\(\\mathbf{R}_{y,i}\\) is a rotation around y-axis, and \\(\\mathbf{R}_{v,i}\\) is a rotation around an axis in the xz-plane, such that \\(\\mathbf{R}_{v,i}\\mathbf{e}_2 = \\mathbf{v}_i\\). Since \\(\\mathbf{R}_{v,i}\\) is calculated directly from \\(\\mathbf{v}_i\\), we assume it to be known." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.905, + 0.44 + ], + "angle": 0, + "content": "In this case, \\(\\mathbf{R}_i^{\\prime}\\) (5) takes the form \\(\\mathbf{R}_i^{\\prime} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\). While both \\(\\mathbf{R}_{0,i}\\) and \\(\\mathbf{R}_{v,i}\\) have their axes in the xz-plane, \\(\\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\) does generically not share this property. However, we can decompose \\(\\mathbf{R}_{0,i}\\mathbf{R}_{v,i} = \\mathbf{R}_{a,i}\\mathbf{R}_{b,i}\\), s.t. \\(\\mathbf{R}_{a,i}\\) has its axis in the xz-plane and \\(\\mathbf{R}_{b,i}\\) rotates around the y-axis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.486 + ], + "angle": 0, + "content": "Now, we decompose \\(\\mathbf{R}_i^{\\prime}\\) as \\(\\mathbf{R}_{a,i} \\cdot \\mathbf{R}_i^{\\prime \\prime}\\), where \\(\\mathbf{R}_{a,i}\\) is a known rotation around axis in the xz-plane, and \\(\\mathbf{R}_i^{\\prime \\prime} = \\mathbf{R}_{b,i}\\). \\(\\mathbf{R}_{y,i}\\) is an unknown rotation around y-axis." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.486, + 0.905, + 0.53 + ], + "angle": 0, + "content": "We represent \\(\\mathbf{R}_{a,i},\\mathbf{R}_i^{\\prime \\prime}\\) with quaternions \\(\\mathbf{q}_i = [\\mathbf{q}_{i,w}\\mathbf{q}_{i,x}0\\mathbf{q}_{i,z}]^T\\) \\(\\mathbf{q}_i^{\\prime \\prime} = [\\mathbf{q}_{i,w}^{\\prime \\prime}0\\mathbf{q}_{i,y}^{\\prime \\prime}0]^T\\) , write the elements of \\(\\mathbf{R}'\\) in the form" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.535, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {R} _ {i (1, 1)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) - 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (1, 3)} ^ {\\prime} = 2 \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} + 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right), \\\\ \\mathbf {R} _ {i (3, 1)} ^ {\\prime} = 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) - 2 (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (3, 3)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, x} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) + 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\tag {24} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.862, + 0.658 + ], + "angle": 0, + "content": "and decompose the left \\(2 \\times 2\\) submatrix of \\(\\mathbf{A}_i\\) (7) as" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.665, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l l} - \\mathbf {R} _ {i, (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (1, 1)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i, (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (3, 1)} ^ {\\prime} \\end{array} \\right] = \\mathbf {A} _ {i, a} \\cdot \\mathbf {A} _ {i, b}, \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.712, + 0.548, + 0.725 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.731, + 0.905, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i, a} = \\left[ \\begin{array}{l l} - 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} & (1 - 2 \\mathbf {q} _ {i, z} ^ {2}) \\\\ - (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) & 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\end{array} \\right], \\tag {26}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.775, + 0.905, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i, b} = \\left[ \\begin{array}{c c} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) & 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ - 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) \\end{array} \\right]. \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.813, + 0.905, + 0.877 + ], + "angle": 0, + "content": "\\(\\mathbf{A}_{i,a}\\) can be directly obtained from the known gravity direction and the current scanline \\(y_{i}\\), and \\(\\mathbf{A}_{i,b}\\) is an unknown \\(2D\\) rotation matrix. We define \\(\\mathbf{A}_i^{\\prime} = \\mathbf{A}_{i,a}^{-1}\\mathbf{A}_i\\), and \\(\\mathbf{u}_{i,j}^{\\prime \\prime} = \\mathbf{A}_{i,a}^{\\prime T}\\mathbf{u}_{i,j}\\). Then, there holds" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.884, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {i, j} ^ {\\prime \\prime T} \\mathbf {A} _ {i} ^ {\\prime} \\mathbf {L} _ {j, h} = 0. \\tag {28}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7147" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.227 + ], + "angle": 0, + "content": "Since the left \\(2 \\times 2\\) submatrix of \\(\\mathbf{A}_i'\\) is \\(\\mathbf{A}_{i,b}\\), \\(\\mathbf{A}_i'\\) has a form of a calibrated \\(2 \\times 3\\) camera matrix, and the problem with known gravity direction and vertical lines is equivalent to estimating a 2D structure from images taken by \\(n\\) calibrated cameras [8]. Unlike its uncalibrated counterpart, the calibrated problem can be solved without projective ambiguity. Solvers for the cases \\(m = 3, n = 5\\) and \\(m = 4, n = 4\\) have been presented in [8]. Notably, these numbers exactly correspond to the minimal problems discussed in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.379 + ], + "angle": 0, + "content": "The solver for \\( m = 3, n = 5 \\) estimates a trifocal tensor of the form (23) from 5 point correspondences and 2 linear internal constraints, and solves a quadratic function to decompose the tensor into camera matrices. The solver for \\( m = 4, n = 4 \\) estimates a dual quadrifocal tensor from 4 point correspondences and 11 linear internal constraints, and also solves a quadratic function to decompose the tensor into camera matrices. In both cases, there are two solutions in terms of matrices \\( \\mathbf{A}_i' \\). Typically, the correct solution is selected by counting the points lying in front of all cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.424 + ], + "angle": 0, + "content": "After finding \\(\\mathbf{A}_i^{\\prime}\\) with [8], we construct the camera poses \\(\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)\\). For this, we first obtain \\(\\mathbf{R}_i^{\\prime \\prime}\\). Because (28) is homogeneous, there are two possible matrices \\(\\mathbf{R}_i^{\\prime \\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.432, + 0.484, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {R} _ {i} ^ {\\prime \\prime} \\in \\left\\{\\left[ \\begin{array}{c c c} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right], \\right. \\tag {29} \\\\ \\left. \\left[ \\begin{array}{c c c} - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right] \\right\\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.484, + 0.587 + ], + "angle": 0, + "content": "The existence of two choices of \\(\\mathbf{R}_i^{\\prime \\prime}\\) per \\(\\mathbf{A}_i^\\prime\\) accounts for the presence of 16 solutions in the 3 view case and 32 solutions in the 4 view case. Finally, we construct \\(\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {0, i} ^ {T} \\mathbf {R} _ {a, i} \\mathbf {R} _ {i} ^ {\\prime \\prime}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {i (1: 2, 3)}. \\tag {30} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Note, that \\(\\mathbf{C}_i(y_i)_{(2)}\\) is always zero. This is in accordance with the Section 2, which explains that the camera centers can only be estimated up to the shift along the direction \\(\\mathbf{L}_d\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.719 + ], + "angle": 0, + "content": "To summarize, we perform the following steps to solve problems \\((E,3,5)\\) and \\((E,4,4)\\):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.415, + 0.735 + ], + "angle": 0, + "content": "1. Input: Projections \\(\\mathbf{p}_{i,j}\\), vertical directions \\(\\mathbf{v}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.735, + 0.481, + 0.751 + ], + "angle": 0, + "content": "2. Calculate matrices \\(\\mathbf{R}_{0,i},\\mathbf{R}_{v,i},\\mathbf{R}_{a,i}\\) , and vectors \\(\\mathbf{u}_{i,j}\\) (7)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.442, + 0.766 + ], + "angle": 0, + "content": "3. Compose matrices \\(\\mathbf{A}_{i,a}\\) (26) and vectors \\(\\mathbf{u}_{i,j}^{\\prime \\prime}(28)\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "4. Use the 3 view or the 4 view solver from [8] to find 2 sets of matrices \\(\\mathbf{A}_i^{\\prime}\\), \\(i\\in \\{1,\\dots,m\\}\\) from \\(\\mathbf{u}_{i,j}^{\\prime \\prime}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.408, + 0.811 + ], + "angle": 0, + "content": "5. Construct \\(\\mathbf{R}_i^{\\prime \\prime}\\) (29), \\(\\mathbf{R}_i(y_i)\\), and \\(\\mathbf{C}_i(y_i)\\) (30)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.84 + ], + "angle": 0, + "content": "On average, the 3 view solver takes \\(9.16\\mu s\\), and the 4 view solver takes \\(70.84\\mu s\\)." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.376, + 0.866 + ], + "angle": 0, + "content": "4.3. Parallel lines and known gravity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We present the solution to the problem \\((D,3,7)\\). In this case, all lines share a common direction \\(\\mathbf{L}_d\\), which is" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "generically not aligned with the vertical direction \\(\\mathbf{e}_2\\). Let \\(\\mathbf{R}_d \\in SO(3)\\) be a rotation around an axis in the xz-plane, s.t. \\(\\mathbf{L}_d = \\mathbf{R}_d\\mathbf{e}_2\\). Substituting this into (3), we get" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.144, + 0.905, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {R} _ {d} \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0, \\tag {31}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.183, + 0.905, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\mathbf {R} _ {d} ^ {T} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {32}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.205, + 0.905, + 0.236 + ], + "angle": 0, + "content": "We define \\(\\mathbf{L}_{0,j}^{\\prime} = \\mathbf{R}_d^T\\mathbf{L}_{0,j}\\), \\(\\mathbf{C}_i^\\prime = \\mathbf{R}_d^T\\mathbf{C}_i(y_i)\\) and substitute these into (32):" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.244, + 0.905, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} ^ {\\prime} - \\mathbf {C} _ {i} ^ {\\prime}\\right) = 0. \\tag {33}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.906, + 0.332 + ], + "angle": 0, + "content": "Following Sec. 4.2, we find rotation matrices \\(\\mathbf{R}_{0,i}\\) and \\(\\mathbf{R}_{v,i}\\), decompose \\(\\mathbf{R}_i(y_i)\\) as \\(\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\), and find vectors \\(\\mathbf{u}_{i,j}^{\\prime}\\) according to (8). Then, we define \\(\\mathbf{R}_{D,i} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\mathbf{R}_d\\), and \\(\\mathbf{t}_{D,i} = -\\mathbf{R}_{D,i}^{-1}[\\mathbf{e}_2]_{\\times}\\mathbf{C}_i^{\\prime}\\), and rewrite the constraint as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\left[ \\begin{array}{l l l} - \\mathbf {R} _ {D, i (1, 3)} & \\mathbf {R} _ {D, i (1, 1)} & \\mathbf {t} _ {D, i (1)} \\\\ - \\mathbf {R} _ {D, i (3, 3)} & \\mathbf {R} _ {D, i (3, 1)} & \\mathbf {t} _ {D, i (3)} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {L} _ {0, j (1)} ^ {\\prime} \\\\ \\mathbf {L} _ {0, j (3)} ^ {\\prime} \\\\ 1 \\end{array} \\right], \\tag {34}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.404, + 0.774, + 0.418 + ], + "angle": 0, + "content": "which can be written in matrix form as" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.427, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i, j} \\mathbf {u} _ {i, j} = \\mathbf {A} _ {D, i} \\mathbf {L} _ {h, j} ^ {\\prime}. \\tag {35}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.906, + 0.51 + ], + "angle": 0, + "content": "This equation has the same form as (8), ensuring the existence of a trifocal tensor \\(\\mathbf{T} \\in \\mathbb{R}^{2,2,2}\\) satisfying (22). We estimate this tensor linearly using 7 correspondences, as in Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.512, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Let us now discuss how to decompose \\(\\mathbf{T}\\) to estimate the relative pose. First, we fix the coordinate frame by setting \\(\\mathbf{R}_1(y_1) = I\\), \\(\\mathbf{C}_1(y_1) = 0\\). Both the tensor \\(\\mathbf{T}\\) and the pose have 7 DoF. Using Gröbner basis, we have discovered that for every tensor \\(\\mathbf{T} \\in \\mathbb{R}^{2,2,2}\\), there are 48 camera poses \\(\\mathbf{R}_{y,2}, \\mathbf{t}_{D,2}, \\mathbf{R}_{y,3}, \\mathbf{t}_{D,3}, \\mathbf{R}_d\\), such that the matrices \\(\\mathbf{A}_{D,i}\\) constructed via (34) have the tensor \\(\\mathbf{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.618, + 0.906, + 0.692 + ], + "angle": 0, + "content": "Due to the high number of variables and equation complexity, we could not generate an elimination template for a Gröbner basis solver. Therefore, we solve this problem with homotopy continuation [18], with an average runtime \\( 19089\\mu s \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.693, + 0.905, + 0.724 + ], + "angle": 0, + "content": "Finally, we compose \\(\\mathbf{A}_{i,D}\\) according to (34), and the pose \\(\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.731, + 0.905, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {v, i} \\mathbf {R} _ {y, i}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = R _ {d} \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {D, i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {D, i (1: 2, 3)} \\tag {36} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.906, + 0.807 + ], + "angle": 0, + "content": "To summarize, we perform the following steps to solve the problem \\((D,3,7)\\):" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.808, + 0.839, + 0.823 + ], + "angle": 0, + "content": "1. Input: Projections \\(\\mathbf{p}_{i,j}\\), vertical directions \\(\\mathbf{v}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.823, + 0.866, + 0.838 + ], + "angle": 0, + "content": "2. Calculate matrices \\(\\mathbf{R}_{0,i},\\mathbf{R}_{v,i}\\) and vectors \\(\\mathbf{u}_{i,j}\\) (7)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.838, + 0.865, + 0.853 + ], + "angle": 0, + "content": "3. Construct linear system (22) and solve it to get \\(\\mathbf{T}\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.853, + 0.905, + 0.883 + ], + "angle": 0, + "content": "4. Use homotopy continuation to decompose \\(\\mathbf{T}\\) into \\(\\mathbf{R}_{y,2}\\), \\(\\mathbf{t}_{D,2}\\), \\(\\mathbf{R}_{y,3}\\), \\(\\mathbf{t}_{3,D}\\), \\(\\mathbf{R}_d\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.884, + 0.803, + 0.899 + ], + "angle": 0, + "content": "5. Construct \\(\\mathbf{R}_i(y_i)\\) and \\(\\mathbf{C}_i(y_i)\\) using (36)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.808, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.482, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.211, + 0.485, + 0.24 + ], + "angle": 0, + "content": "Figure 3. Histogram of tensor, rotation, and translation errors of the proposed solvers, over \\(10^{5}\\) noiseless samples. In radians." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.266, + 0.292, + 0.282 + ], + "angle": 0, + "content": "4.4. Degenerate scenarios" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Here, we analyze the degenerate cases for the solvers introduced in Sections 4.1, 4.2, and 4.3. Specifically, we consider scenarios where some lines are coplanar or where camera centers are collinear or coincident." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.348, + 0.483, + 0.378 + ], + "angle": 0, + "content": "- Solvers \\((E,3,5)\\), \\((E,4,4)\\) are degenerate if all \\(m\\) lines are coplanar or if any two camera centers coincide." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.378, + 0.484, + 0.409 + ], + "angle": 0, + "content": "- Solvers \\((B,3,7)\\), \\((D,3,7)\\) are degenerate if at least 5 lines are coplanar or if any two camera centers coincide." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.41, + 0.483, + 0.424 + ], + "angle": 0, + "content": "- There is no degeneracy if the camera centers are colinear." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.348, + 0.484, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.223, + 0.456 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.496 + ], + "angle": 0, + "content": "In this section, we evaluate the solvers \\((B,3,7)\\), \\((D,3,7)\\), \\((E,3,5)\\), \\((E,4,4)\\) on both synthetic and real data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.299, + 0.522 + ], + "angle": 0, + "content": "5.1. Synthetic experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.693 + ], + "angle": 0, + "content": "Numerical stability. We sample scanlines \\( y_{i} \\) from \\( \\mathcal{U}_{[-1,1]} \\) and generate random camera rotations \\( \\mathbf{R}_{i,GT} \\) and centers \\( \\mathbf{C}_{i,GT} \\). Since we use one scanline per image, we only sample one pose per image. We get the vertical directions as \\( v_{i} = \\mathbf{R}_{i,GT}\\mathbf{e}_{2} \\). To generate line \\( \\mathbf{L}_j \\), we sample scalars \\( \\mu_{j,x},\\mu_{j,z} \\) from normalized Gaussian distribution and set \\( \\mathbf{L}_{0,j} = [\\mu_{j,x}0\\mu_{j,z} + 5]^T \\). For solver \\( (D,3,7) \\), we also generate scalars \\( \\nu_{x},\\nu_{z} \\), and set \\( L_{j,d} = [\\nu_x1\\nu_z]^T \\). In other cases, we use \\( L_{j,d} = \\mathbf{e}_2 \\). We generate \\( \\mathbf{p}_{i,j} \\) by projecting \\( \\mathbf{L}_j \\) into the camera with pose \\( (\\mathbf{R}_{i,GT},\\mathbf{C}_{i,GT}) \\) and intersecting the projected line with scanline \\( y_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.483, + 0.8 + ], + "angle": 0, + "content": "Before measuring the error, we transform both the ground truth pose, and the estimated pose \\((\\mathbf{R}_{i,\\mathrm{est}},\\mathbf{C}_{i,\\mathrm{est}})\\) into a canonical coordinate frame with \\(\\mathbf{C}_1 = 0\\) \\(\\mathbf{R_{y,1}} = \\mathbf{I},\\mathbf{L_d} = \\mathbf{e_1}\\). Since the y-coordinates of all camera centers are unobservable, we set them to zero. Then, we calculate the rotation error \\(err_{\\mathbf{R}} = \\max_{i = 1}^{m}(\\angle \\mathbf{R}_{i,\\mathrm{est}}^{T}\\mathbf{R}_{i,GT})\\) and the translation error \\(err_{\\mathbf{C}} = \\max_{i = 1}^{m}\\angle (\\mathbf{C}_{i,GT},\\mathbf{C}_{i,\\mathrm{est}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.484, + 0.906 + ], + "angle": 0, + "content": "Additionally, we construct tensors \\(\\mathbf{T}_{GT}\\) from GT poses and estimate \\(\\mathbf{T}_{est}\\) using the solvers. For trifocal problems, \\(\\mathbf{T}_{GT},\\mathbf{T}_{est}\\) are trifocal tensors (23), while for \\((E,4,4)\\), they are dual quadrifocal tensors. We measure tensor error errT as the Frobenius norm of \\(\\mathbf{T}_{GT} - \\mathbf{T}_{est}\\). For each solver, we generate \\(10^{5}\\) synthetic noise-free scenes and measure errR, errC, and errT. Results in Fig. 3 show solver stability." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Figure 4. Rotation, translation, and tensor errors averaged over 10000 successful runs as a function of the point and the vertical noise, measured from the real solution closest to the ground truth. If a solver returns only complex solutions, we report a failure. The failure rates are reported in the last row. Plots in the same row share the y-axis and plots in the same column share the x-axis. The fixed parameters are in the titles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Tests with noise. To evaluate the robustness to noise, we set focal length \\( f = 1000px \\), and add Gaussian noise with standard deviation \\( \\frac{\\sigma_p}{f} \\) to each projection \\( \\mathbf{p}_{i,j} \\). We also perturb vertical directions \\( \\mathbf{v}_i \\) by applying a random 3D rotation with angle \\( \\sigma_v \\). Fig. 4 shows \\( err_{\\mathbf{R}} \\), \\( err_{\\mathbf{C}} \\), and \\( err_{\\mathbf{T}} \\) as the function of \\( \\sigma_p \\), \\( \\sigma_v \\), averaged over \\( 10^4 \\) runs. The results confirm that, since the line detections tend to be relatively precise, our solvers exhibit reasonable behaviour under realistic noise levels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.74, + 0.631 + ], + "angle": 0, + "content": "5.2. Real-world Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.744 + ], + "angle": 0, + "content": "We evaluate our solvers using the Fastec dataset [43], which consists of 19 sequences captured from a moving car in Tübingen. Each sequence contains 34 rolling shutter (RS) images and two global shutter (GS) images for every RS image: one with the pose of the first scanline and one with the pose of the middle scanline. The images have a resolution of \\(640 \\times 480\\) px and a focal length \\(768\\) px." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Since GT poses are unavailable, we use COLMAP [61] on the GS images to obtain pseudo-GT poses \\((\\mathbf{R}_{i,F},\\mathbf{C}_{i,F})\\) for the first scanline of image \\(i\\) and \\((\\mathbf{R}_{i,M},\\mathbf{C}_{i,M})\\) for the middle scanline. The pseudo-GT pose for scanline \\(y_{i}\\) is interpolated as: \\(\\mathbf{R}_{i,GT}(y_i) = \\mathbf{R}_{i,M\\to y_i}\\mathbf{R}_{i,M},\\mathbf{C}_{i,GT}(y_i) =\\) \\(\\mathbf{C}_{i,M} + \\frac{h - 2y_i}{h} (\\mathbf{C}_{i,F} - \\mathbf{C}_{i,M})\\) , where \\(\\mathbf{R}_{i,M\\to y_i}\\) is a rotation matrix with the axis of \\(\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}\\) , and the angle \\(\\frac{h - 2y_i}{h}\\) times the angle of \\(\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}\\) . We evaluate the solvers in multi-view and single-view settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multi-view. We estimate poses for \\(m\\) consecutive im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7149" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.462, + 0.199 + ], + "angle": 0, + "content": "
solver≤ 5°≤ 10°≤ 20°≤ 30°
MV(E,3,5)131015
(E,4,4)581114
(D,3,7)6101517
SV(E,3,5)33510
(E,4,4)24713
(D,3,7)4479
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 2. Number of sequences (out of 19) from the Fastec dataset [43], with at least one estimated pose with error below \\(5/10/20/30^{\\circ}\\). MV: Multi-view, SV: Single-view." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.269, + 0.483, + 0.435 + ], + "angle": 0, + "content": "ages, selecting one scanline per image. We use DeepLSD [49] and Gluestick [50] to detect and match the lines. Although the projections of lines by RS cameras are generically curves, the line-based methods work well in the case of Fastec dataset. To select scanlines \\( y_{1}, \\ldots, y_{m} \\), we exhaustively search for the scanlines that maximize the number of matched lines intersecting them. Then, we find \\( p_{i,j} \\) as the intersections of the detected lines and the scanlines. Following [12], we approximate the vertical direction by \\( [0 1 0]^T \\). This setting works well for many datasets, including Fastec, and does not require gravity direction measurement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.512 + ], + "angle": 0, + "content": "Single-view. In this case, select \\( m \\) scanlines in the same image and estimate the pose between them. We detect lines with DeepLSD [49], and select scanlines \\( y_{1},\\ldots ,y_{m} \\) with spacing at least 25px with the highest number of shared lines. We obtain \\( p_{i,j} \\) by intersecting lines with scanlines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.648 + ], + "angle": 0, + "content": "We evaluate the solvers in RANSAC [22]. We label a correspondence as inlier, if the triangulated line \\(\\mathbf{L}_j\\) lies in front of all cameras and its reprojection error \\(\\varepsilon\\) is below \\(1px\\). In the case of \\((B,3,7)\\), we decompose the tensor into canonical cameras (20) before triangulation. Details about triangulation are in the SM. The score for a model is the sum over all lines of \\((1px - \\varepsilon)^2\\) if \\(\\varepsilon < 1px\\), and zero otherwise. We run RANSAC for 1000 iterations, and select the model with the highest score. We do not use local optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.8 + ], + "angle": 0, + "content": "The results are summarized in Tables 2, 3 summarize the results for the solvers \\((E,3,5)\\), \\((E,4,4)\\), and \\((D,3,7)\\), and Table 4 shows the tensor error for the solver \\((B,3,7)\\). Although the total number of camera poses with an error below \\(10^{\\circ}\\) and \\(20^{\\circ}\\) remains relatively low, in Structure-from-Motion (SfM) applications, it is sufficient to estimate a single correct relative pose per sequence to initialize the reconstruction. Table 2 confirms that this is achievable for the vast majority of sequences. A more detailed evaluation and further experiments are provided in the SM." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.815, + 0.364, + 0.831 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "While our approach shows promising results for rolling shutter relative pose estimation, several limitations remain. A key challenge is the lack of robust curve detectors and matchers, as rolling shutter distortion bends straight lines." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.089, + 0.865, + 0.199 + ], + "angle": 0, + "content": "
solver≤ 10°≤ 20°≤ 30°
MV(E,3,5)1.0 %6.2 %21.9 %
(E,4,4)1.9 %7.9 %22.4 %
(D,3,7)5.6 %13.8 %21.7 %
SV(E,3,5)0.5 %2.5 %5.3 %
(E,4,4)1.2 %4.1 %10.0 %
(D,3,7)0.5 %2.2 %6.6 %
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Table 3. Percentage of camera poses from Fastec [43] with pose error below \\(10/20/30^{\\circ}\\). MV: Multi-view, SV: Single-view." + }, + { + "type": "table", + "bbox": [ + 0.56, + 0.242, + 0.858, + 0.291 + ], + "angle": 0, + "content": "
settingMedian errorMinimum error
MV0.850.44
SV0.890.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.292, + 0.906, + 0.334 + ], + "angle": 0, + "content": "Table 4. Median and Minimum tensor error of the \\((B,3,7)\\) solver, averaged over all sequences from Fastec [43]. A more detailed evaluation is in the SM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Additionally, our method relies on chirality constraints to resolve ambiguity, as noted in [8]. The method does not use any motion model, which makes it broadly applicable and model-agnostic. However, this universality comes at the cost of stability: we expect methods that incorporate motion models to be more stable. Our solvers are also limited to specific scene types and assume a known gravity direction. However, this reliance could be mitigated using scene knowledge, such as geometric relations between lines, in combination with solver \\((B,3,7)\\). Finally, our method estimates pose only up to a shift along the line direction. Future work could integrate additional geometric constraints, improve feature detection, and integrate the solvers to a rolling-shutter SfM pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.634, + 0.604 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.734 + ], + "angle": 0, + "content": "In this paper, we propose a new approach for estimating the relative pose of rolling shutter cameras using projections of lines onto multiple scanlines. This enables estimating relative pose of rolling shutter cameras without explicitly modelling camera motion. The proposed solvers provide an important building block for rolling-shutter SfM where each scanline has its pose computed independently, without relying on a global motion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We identify minimal problems for different variations of the single-scanline relative pose problem, including cases with parallel lines, vertical lines, and known gravity direction. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors. To evaluate our approach, we conduct experiments on both synthetic data and real-world rolling shutter images from the Fastec dataset, considering scanlines from either the same image or different images. These experiments demonstrate that our solvers successfully estimate at least one correct pose in the majority of scenes, confirming their feasibility for SfM initialization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7150" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[1] Omar Ait-Aider and François Berry. Structure and kinematics triangulation with a rolling shutter stereo rig. In IEEE 12th International Conference on Computer Vision, ICCV 2009, Kyoto, Japan, September 27 - October 4, 2009, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[2] Omar Ait-Aider, Nicolas Andreff, Jean-Marc Lavest, and Philippe Martinet. Simultaneous object pose and velocity computation using a single view from a rolling shutter camera. In Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part II, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[3] Omar Ait-Aider, Adrien Bartoli, and Nicolas Andreff. Kinematics from lines in a single rolling shutter image. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[4] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. R6P - rolling shutter absolute pose problem. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[5] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. Rolling shutter absolute pose problem with known vertical direction. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[6] Cenek Albl, Zuzana Kukelova, Viktor Larsson, and Tomás Pajdla. Rolling shutter camera absolute pose. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[7] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomás Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[8] Kalle Åström and Magnus Oskarsson. Solutions and ambiguities of the structure and motion problem for 1d retinal vision. J. Math. Imaging Vis., 2000, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[9] Fang Bai, Agniva Sengupta, and Adrien Bartoli. Scanline homographies for rolling-shutter plane absolute pose. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[10] Akash Bapat, Enrique Dunn, and Jan-Michael Frahm. Towards kilo-hertz 6-dof visual tracking using an egocentric cluster of rolling shutter cameras. IEEE Trans. Vis. Comput. Graph., 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[11] Akash Bapat, True Price, and Jan-Michael Frahm. Rolling shutter and radial distortion are features for high frame rate multi-camera tracking. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. StereoGlue: Robust estimation with single-point solvers. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Paul Breiding and Sascha Timme. Homotopycontinuation.jl: A package for homotopy continuation in julia. In Mathematical Software - ICMS 2018 - 6th International Conference, South Bend, IN, USA, July 24-27, 2018, Proceedings. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Mingdeng Cao, Zhihang Zhong, Jiahao Wang, Yinqiang Zheng, and Yujiu Yang. Learning adaptive warping for real-world rolling shutter correction. CoRR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Yuchao Dai, Hongdong Li, and Laurent Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.904, + 0.332 + ], + "angle": 0, + "content": "[16] Juan Carlos Dibene, Yazmín Maldonado, Leonardo Trujillo, and Enrique Dunn. Prepare for ludicrous speed: Marker-based instantaneous binocular rolling shutter localization. IEEE Trans. Vis. Comput. Graph., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[17] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. PLMP - point-line minimal problems in complete multiview visibility. IEEE Trans. Pattern Anal. Mach. Intell., 2024, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.391, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[18] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.904, + 0.531 + ], + "angle": 0, + "content": "[19] Bin Fan, Yuchao Dai, and Ke Wang. Rolling-shutter-stereo-aware motion estimation and image correction. Comput. Vis. Image Underst., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.904, + 0.573 + ], + "angle": 0, + "content": "[20] Bin Fan, Yuchao Dai, Zhiyuan Zhang, and Ke Wang. Differential sfm and image correction for a rolling shutter stereog. Image Vis. Comput., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.904, + 0.63 + ], + "angle": 0, + "content": "[21] Bin Fan, Yuchao Dai, and Hongdong Li. Rolling shutter inversion: Bring rolling shutter images to high framerate global shutter video. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[22] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 1981. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[23] Per-Erik Forssén and Erik Ringaby. Rectifying rolling shutter video from hand-held devices. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[24] Marvin Anas Hahn, Kathlen Kohn, Orlando Marigliano, and Tomás Pajdla. Order-one rolling shutter cameras. CoRR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.904, + 0.829 + ], + "angle": 0, + "content": "[25] Michiel Hazewinkel. Encyclopaedia of mathematics. 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Johan Hedborg, Erik Ringaby, Per-Erik Forssén, and Michael Felsberg. Structure and motion estimation from rolling shutter video. In IEEE International Conference on Computer Vision Workshops, ICCV 2011 Workshops, Barcelona, Spain, November 6-13, 2011, 2011. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Johan Hedborg, Per-Erik Forssén, Michael Felsberg, and Erik Ringaby. Rolling shutter bundle adjustment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[28] Radu Horaud, Roger Mohr, and Boguslaw Lorecki. On single-scanline camera calibration. IEEE Trans. Robotics Autom., 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[29] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[30] Chao Jia and Brian L. Evans. Probabilistic 3-d motion estimation for rolling shutter video rectification from visual and inertial measurements. In 14th IEEE International Workshop on Multimedia Signal Processing, MMSP 2012, Banff, AB, Canada, September 17-19, 2012, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[31] Georg Klein and David William Murray. Parallel tracking and mapping on a camera phone. In Science & Technology Proceedings, 8th IEEE International Symposium on Mixed and Augmented Reality 2009, ISMAR 2009, Orlando, Florida, USA, October 19-22, 2009, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[32] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, and Tomás Pajdla. Linear solution to the minimal absolute pose rolling shutter problem. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[33] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, Konrad Schindler, and Tomás Pajdla. Minimal rolling shutter absolute pose with unknown focal length and radial distortion. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[34] Yizhen Lao and Omar Ait-Aider. Rolling shutter homography and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2780-2793, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[35] Yizhen Lao, Omar Ait-Aider, and Helder Araujo. Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognit. Lett., 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[36] Yizhen Lao, Omar Ait-Aider, and Adrien Bartoli. Rolling shutter pose and ego-motion estimation using shape-from-template. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part II, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[37] Chang-Ryeol Lee and Kuk-Jin Yoon. Monocular visual odometry with a rolling shutter camera. CoRR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[38] Chang-Ryeol Lee and Kuk-Jin Yoon. Inertial-aided rolling shutter relative pose estimation. CoRR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[39] Chang-Ryeol Lee, Ju Hong Yoon, Min-Gyu Park, and Kuk-Jin Yoon. Gyroscope-aided relative pose estimation for rolling shutter cameras. CoRR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Dongdong Li, Gongjian Wen, Bing Wei Hui, Shaohua Qiu, and Wenfei Wang. Cross-ratio invariant based line scan camera geometric calibration with static linear data. Optics and Lasers in Engineering, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Moyang Li, Peng Wang, Lingzhe Zhao, Bangyan Liao, and Peidong Liu. Usb-nerf: Unrolling shutter bundle adjusted neural radiance fields. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[42] Bangyan Liao, Delin Qu, Yifei Xue, Huiqing Zhang, and Yizhen Lao. Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[43] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.295, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[44] Ludovic Magerand, Adrien Bartoli, Omar Ait-Aider, and Daniel Pizarro. Global optimization of object pose and motion from a single rolling shutter image with automatic 2d-3d matching. In Computer Vision - ECCV 2012 - 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[45] Marci Meingast, Christopher Geyer, and Shankar Sastry. Geometric models of rolling-shutter cameras. CoRR, abs/cs/0503076, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[46] Branislav Micusik and Georgios Evangelidis. Renormalization for initialization of rolling shutter visual-inertial odometry. Int. J. Comput. Vis., 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[47] Thanh-Tin Nguyen and Maxime Lhuillier. Adding synchronization and rolling shutter in multi-camera bundle adjustment. In Proceedings of the British Machine Vision Conference 2016, BMVC 2016, York, UK, September 19-22, 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[48] David Nister. An efficient solution to the five-point relative pose problem. IEEE Trans. Pattern Anal. Mach. Intell., 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[49] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[50] Rémi* Pautrat, Iago* Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. GlueStick: Robust image matching by sticking points and lines together. In International Conference on Computer Vision (ICCV), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.907, + 0.77 + ], + "angle": 0, + "content": "[51] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018, Lake Tahoe, NV, USA, March 12-15, 2018, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[52] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in Manhattan world. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[53] Delin Qu, Yizhen Lao, Zhigang Wang, Dong Wang, Bin Zhao, and Xuelong Li. Towards nonlinear-motion-aware and occlusion-robust rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7152" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[54] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[55] Long Quan and Takeo Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Trans. Pattern Anal. Mach. Intell., 1997. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[56] Vijay Rengarajan, A. N. Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[57] Erik Ringaby and Per-Erik Forssén. Efficient video rectification and stabilisation for cell-phones. Int. J. Comput. Vis., 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[58] Olivier Saurer, Kevin Koser, Jean-Yves Bouguet, and Marc Pollefeys. Rolling shutter stereo. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[59] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. A minimal solution to the rolling shutter pose estimation problem. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2015, Hamburg, Germany, September 28 - October 2, 2015, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[60] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. Sparse to dense 3d reconstruction from rolling shutter images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[61] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[62] David Schubert, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. Direct sparse odometry with rolling shutter. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VIII, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[63] Wei Shang, Dongwei Ren, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised learning to bring dual reversed rolling shutter images alive. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[64] Ken Shoemake. Animating rotation with quaternion curves. In SIGGRAPH, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[65] Henrik Stewénius, David Nister, Magnus Oskarsson, and Kalle Åström. Solutions to minimal generalized relative pose problems. Workshop on omnidirectional vision, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[66] Daobilige Su, Asher Bender, and Salah Sukkarieh. Improved cross-ratio invariant-based intrinsic calibration of A hyperspectral line-scan camera. Sensors, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[67] Yufen Sun, Gang Liu, and Yue Sun. An affine motion model for removing rolling shutter distortions. IEEE Signal Process. Lett., 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[68] Subeesh Vasu, Mahesh Mohan M. R., and A. N. Rajagopalan. Occlusion-aware rolling shutter rectification of 3d scenes. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[69] Carlos Andres Luna Vázquez, Manuel Mazo, José Luis Lázaro, and Juan F. Vázquez. Calibration of line-scan cameras. IEEE Trans. Instrum. Meas., 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Ke Wang, Bin Fan, and Yuchao Dai. Relative pose estimation for stereo rolling shutter cameras. In IEEE International Conference on Image Processing, ICIP 2020, Abu Dhabi, United Arab Emirates, October 25-28, 2020, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[71] Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. Urs-nerf: Unordered rolling shutter bundle adjustment for neural radiance fields. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[72] Weilong Yan, Robby T. Tan, Bing Zeng, and Shuaicheng Liu. Deep homography mixture for single image rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[73] Ming Yao, Zuyun Zhao, and Bugao Xu. Geometric calibration of line-scan camera using a planar pattern. J. Electronic Imaging, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[74] Yongcong Zhang, Bangyan Liao, Yifei Xue, Chen Lu, Peidong Liu, and Yizhen Lao. RSL-BA: rolling shutter line bundle adjustment. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LIX, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[75] Bingbing Zhuang and Quoc-Huy Tran. Image stitching and rectification for hand-held cameras. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[76] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[77] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7153" + } + ] +] \ No newline at end of file diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_origin.pdf b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..86cf7d447f61bcff3e96ad5d2bfdfa58b1dde8d4 --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/1be14823-e24f-4aa3-8a5b-8d0994a9e44f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6636bd9f23f5f204c12d598ad311fe6895ca410f949b36a0c85b0e1d39208ae +size 1744999 diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/full.md b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a42a396dd58c24af33cf25df638b6bdba5d720f --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/full.md @@ -0,0 +1,600 @@ +# Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras + +Petr Hruby + +ETH Zürich + +Rämistrasse 101, 8006 Zürich + +petr.h Ruby@inf.ethz.ch + +Marc Pollefeys + +ETH Zürich / Microsoft Spatial AI Lab + +Rämistrasse 101, 8006 Zürich + +marc.pollefeys@inf.ethz.ch + +# Abstract + +We propose a novel approach for estimating the relative pose between rolling shutter cameras using the intersections of line projections with a single scanline per image. This allows pose estimation without explicitly modeling camera motion. Alternatively, scanlines can be selected within a single image, enabling single-view relative pose estimation for scanlines of rolling shutter cameras. Our approach is designed as a foundational building block for rolling shutter structure-from-motion (SfM), where no motion model is required, and each scanline's pose can be computed independently. We classify minimal solvers for this problem in both generic and specialized settings, including cases with parallel lines and known gravity direction, assuming known intrinsic and no lens distortion. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors, by leveraging connections between this problem and the estimation of 2D structure from 1D cameras. Experiments on rolling shutter images from the Fastec dataset demonstrate the feasibility of our approach for initializing rolling shutter SfM, highlighting its potential for further development. The code will be made publicly available. + +# 1. Introduction + +Relative pose estimation is a fundamental problem in computer vision with applications in Structure from Motion [61], SLAM, multi-view stereo, and visual odometry. This paper focuses on estimating relative pose between rolling shutter (RS) cameras. + +Rolling shutter cameras capture images row by row. If the camera is static during capture, the resulting image is equivalent to that of a pinhole camera. However, if the camera moves, the rolling shutter effect distorts the image, making standard pinhole-based methods [48] unreliable. + +A common approach to handle rolling shutter distortion is modeling camera motion with parametric models such as SLERP [64], Cayley transformation [25], linearized + +![](images/9167d9030ae607acdad8367af0c7b8ed46e8ed334a531b47fc5c5102eaeb54d1.jpg) + +![](images/00ca0360a4f25ff5d06d72003db2394c31994a8bd8dcfb333683b0b05e46b345.jpg) + +![](images/423bd1bad9211b1cdc4ac82bfcc6f9d7fb39a98f41734eb1d188f4b5c6f96bd5.jpg) + +![](images/5ffcf8b46cdcc7b0f1ea112eaec5af7df9692ec2d7fce9ea88d201e2e19c7626.jpg) +Figure 1. We estimate a relative pose from the intersections of the line projections with a single scanline per image (green). This allows us to estimate relative pose between rolling shutter cameras without modelling the camera motion. + +rotation [4], affine motion [67], pushbroom [46], accelerated motion [19], or order-one motion [24]. This has been successfully applied to absolute pose estimation [2-6, 9, 32, 33, 36, 44, 45, 59], bundle adjustment [27, 41, 42, 47, 60, 62, 71, 74], SfM [26, 31, 58], and motion rectification from single cameras [35, 51, 52, 56, 72, 77], camera rigs [1, 7, 20, 63], and videos [14, 19, 21, 23, 30, 53, 54, 57, 67, 68, 75]. + +While most methods use point-based features, some employ lines for absolute pose estimation [3], bundle adjustment [74], and motion rectification [35, 51, 52, 56]. + +The problem of estimating relative pose for rolling shutter cameras has been explored in prior work. Special essential matrices for RS cameras under different motion models were introduced in [15], along with linear non-minimal solvers. The problem was also solved for pure translation models with constant velocity and acceleration [76]. IMU-based approaches reduced the required number of point correspondences from 11 to 5 [38, 39]. Other methods first estimate pose using a global shutter solver before refining with an RS model [37]. Additional works estimate relative pose for RS camera rigs [70] and for planar scenes [34]. Recently, [24] classified RS relative pose minimal problems + +under the order-one motion model. Despite this, the general RS relative pose problem remains unsolved. + +One drawback of parametric motion models is their reliance on an assumed motion that may not match real-world trajectories. An alternative is estimating pose independently for each scanline, similar to line-camera calibration [28, 40, 66, 69, 73]. This approach has been used for tracking a single camera [10, 11] and a camera rig [16] with respect to a predefined marker. + +We explore the feasibility of estimating relative pose between scanlines of RS cameras, without explicit motion modelling. The goal is to use this method as a building block to initialize rolling-shutter SfM where no motion model is required and each scanline can have its pose independently computed. To the best of our knowledge, this is the first work that enables this. After the initialization, scanlines from remaining images may be registered by solving the generalized relative pose problem [65], similarly to [10, 11, 16], or, in the case of a planar scene, using [9]. Our contributions include: + +- Formulation of the problem of relative pose estimation from line projections in a single scanline per image, assuming known intrinsics and no lens distortion. +- Enumeration of minimal problems for general and specialized settings with parallel lines and gravity priors. +- Demonstration that the problems with parallel lines and with vertical lines and gravity prior correspond to estimating relative pose of uncalibrated cameras [55], respectively calibrated [8] 1D camera pose. +- Development of minimal solvers for the following cases: + +- 3 cameras, 7 parallel lines (with projective ambiguity). +- 3 cameras, gravity prior, 7 parallel lines. +- 3 cameras, gravity prior, 5 vertical lines. +- 4 cameras, gravity prior, 4 vertical lines. + +- Evaluation of solvers on synthetic and real RS datasets (Fastec, [43]), taking scanlines either from different images or from a single image. + +# 1.1. Notation + +We use bold letters for matrices and vectors, and normal font for scalars. Indexing starts from one, with indices in brackets; e.g., the first element of $\mathbf{L}_d$ is $\mathbf{L}_{d(1)}$ . Matrix indexing follows Matlab style, where $\mathbf{A}_{(:,1:2)}$ selects the first two rows of $\mathbf{A}$ . The scanlines are denoted with a scalar $y \in \mathbb{R}$ giving their y-coordinate. The set of all 3D lines is denoted as $Gr(1,3)$ . + +# 2. Problem Formulation + +This section describes the geometry of a line projected into a rolling-shutter camera and derives an algebraic constraint useful for estimating camera poses. + +Consider $n$ rolling-shutter cameras capturing the same 3D scene. Due to motion, the cameras capture different + +![](images/f0802ace47166ecf5d8b9c5b091b33e6bf11be274f82906f00bdc096d658c563.jpg) +Figure 2. Illustration of the line projection. Line $\mathbf{L}$ projects onto scanline $\mathbf{y}$ at point $\mathbf{p}$ . Then, the ray $\mathbf{r}$ obtained by backprojecting $\mathbf{p}$ intersects $\mathbf{L}$ , yielding the constraint (3). + +rows at different poses. Let $\mathbf{R}_i(y) \in SO(3)$ and $\mathbf{C}_i(y) \in \mathbb{R}^3$ denote the orientation and center of camera $i$ at row $y$ . + +We aim to estimate the relative pose between the cameras without explicitly modeling their motion functions. To achieve this, we select a single row $y_{i}$ per camera and measure intersections between these selected rows and the projections of $m$ straight lines $\mathbf{L}_j \in Gr(1,3)$ . Then, we use this information to recover the camera poses $\mathbf{R}_i(y_i), \mathbf{C}_i(y_i)$ . Figure 1 illustrates the setup. + +Let $\mathbf{p}_{i,j} \in \mathbb{P}^2$ , with $p_{i,j} = [x_{i,j} y_i 1]^T$ , denote the projection of line $\mathbf{L}_j$ onto row $y_i$ of camera $i$ . Then, the back-projected ray $\mathbf{r}_{i,j}$ from $\mathbf{p}_{i,j}$ intersects line $\mathbf{L}_j$ , giving: + +$$ +\mathbf {r} _ {i, j} = \mathbf {C} _ {i} (y _ {i}) + \lambda_ {i, j} \mathbf {R} _ {i} (y _ {i}) ^ {T} \mathbf {p} _ {i, j}, \quad \lambda_ {i, j} \in \mathbb {R}. \tag {1} +$$ + +This situation is depicted in Fig. 2. Let us represent $\mathbf{L}_j$ with a pair $(\mathbf{L}_{0,j},\mathbf{L}_{d,j})$ , where $\mathbf{L}_{0,j}$ is a point on the line and $\mathbf{L}_{d,j}$ is its direction. The ray $\mathbf{r}_{i,j}$ intersects $\mathbf{L}_j$ if and only if: + +$$ +\exists \alpha , \beta \in \mathbb {R}: \quad \mathbf {C} _ {i} \left(y _ {i}\right) + \alpha \mathbf {R} _ {i} \left(y _ {i}\right) ^ {T} \mathbf {p} _ {i, j} = \mathbf {L} _ {0, j} + \beta \mathbf {L} _ {d, j}. \tag {2} +$$ + +This is equivalent to requiring that vectors $\mathbf{R}_i(y_i)^T\mathbf{p}_{i,j}$ , $\mathbf{L}_{d,j}$ , and $\mathbf{L}_{0,j} - \mathbf{C}_i(y_i)$ are linearly dependent. Since three 3D vectors are linearly dependent if and only if their scalar triple product vanishes, we obtain: + +$$ +\mathbf {p} _ {i, j} ^ {T} \mathbf {R} _ {i} \left(y _ {i}\right) \left[ \mathbf {L} _ {d, j} \right] _ {\times} \left(\mathbf {L} _ {0, j} - \mathbf {C} _ {i} \left(y _ {i}\right)\right) = 0. \tag {3} +$$ + +Each projection provides a single constraint on the unknowns (lines and camera poses). A 3D line has four degrees of freedom, requiring four scanlines for triangulation and at least five scanlines for constraining relative pose. + +# 2.1. Assuming parallel lines + +Let us consider the scenario where all lines are parallel, which frequently occurs in real-world settings. Here, all lines share a common direction $\mathbf{L}_d \in \mathbb{R}^3$ . Since Eq. (3) is invariant to scene rotation, we align $\mathbf{L}_d$ with the y-axis, $\mathbf{e}_2$ . This allows representing each line $\mathbf{L}_j$ with a single point $\mathbf{L}_{0,j}$ in the xz-plane, with its y-coordinate equal to zero. + +With a shared direction, each line has only two individual degrees of freedom (DoF). Once $\mathbf{L}_d$ is known, a line can be triangulated from just two cameras. This reduces the + +minimum number of cameras needed for pose estimation to three. To simplify Eq. (3), we substitute $\mathbf{e}_2$ for $\mathbf{L}_{d,j}$ : + +$$ +\mathbf {p} _ {i, j} ^ {T} \mathbf {R} _ {i} (y _ {i}) \left(\left[ \begin{array}{c} \mathbf {L} _ {0, j (3)} \\ 0 \\ - \mathbf {L} _ {0, j (1)} \end{array} \right] - \left[ \begin{array}{c} \mathbf {C} _ {i} (y _ {i}) _ {(3)} \\ 0 \\ - \mathbf {C} _ {i} (y _ {i}) _ {(1)} \end{array} \right]\right) = 0. \quad (4) +$$ + +We define rotation $\mathbf{R}_{0,i} \in SO(3)$ around the x-axis, which maps row $y_{i}$ to $y = 0$ , and set $\mathbf{p}_{i,j}' = \mathbf{R}_{0,i}\mathbf{p}_{i,j}$ . Then, there holds $\mathbf{p}_{i,j}' = [x_{i,j}'01]^T$ . Defining + +$$ +\mathbf {R} _ {i} ^ {\prime} = \mathbf {R} _ {0, i} \mathbf {R} _ {i} (y _ {i}), \tag {5} +$$ + +$$ +\mathbf {t} _ {i} ^ {\prime} = - \mathbf {R} _ {0, i} \mathbf {R} _ {i} (y _ {i}) [ \mathbf {C} _ {i} (y _ {i}) _ {(3)} 0 - \mathbf {C} _ {i} (y _ {i}) _ {(1)} ] ^ {T}, +$$ + +and substituting $\mathbf{R}_i^{\prime},\mathbf{t}_i^{\prime}$ , and $\mathbf{p}_{i,j}^{\prime}$ into Eq. (4), we obtain: + +$$ +\left[ \begin{array}{l l} x _ {i, j} ^ {\prime} & 1 \end{array} \right] ^ {T} \left[ \begin{array}{c c c} - \mathbf {R} _ {i (1, 3)} ^ {\prime} & \mathbf {R} _ {i (1, 1)} ^ {\prime} & \mathbf {t} _ {i (0)} ^ {\prime} \\ - \mathbf {R} _ {i (3, 3)} ^ {\prime} & \mathbf {R} _ {i (3, 1)} ^ {\prime} & \mathbf {t} _ {i (2)} ^ {\prime} \end{array} \right] \left[ \begin{array}{c} \mathbf {L} _ {0, j (1)} \\ \mathbf {L} _ {0, j (3)} \\ 1 \end{array} \right] = 0. \tag {6} +$$ + +We can write Equation (6) in a vector form as + +$$ +\mathbf {u} _ {i, j} ^ {T} \mathbf {A} _ {i} \mathbf {L} _ {h, j} = 0, \tag {7} +$$ + +where $\mathbf{A}_i\in \mathbb{R}^{2,3}$ is a matrix encoding the camera pose, $\mathbf{u}_{i,j}\in \mathbb{P}^1$ is a vector encoding the observation, and $\mathbf{L}_{h,j}\in$ $\mathbb{P}^2$ is a homogeneous representation of point $\mathbf{L}_{0,j}$ . Alternatively, we define $\mathbf{u}_{i,j}^{\prime} = [1 - x_{i,j}^{\prime}]^{T}$ , and rewrite (7) as + +$$ +\lambda_ {i, j} \mathbf {u} _ {i, j} ^ {\prime} = \mathbf {A} _ {i} \mathbf {L} _ {j, h}, \quad \lambda_ {i, j} \in \mathbb {R}. \tag {8} +$$ + +Since constraint (7) is homogeneous, matrix $\mathbf{A}_i$ has 5 DoF. The elements $\mathbf{R}_i(y_i)$ , $\mathbf{C}_i(y_i)_{(1)}$ , $\mathbf{C}_i(y_i)_{(3)}$ of the pose, which appear in $\mathbf{A}_i$ , also have 5 DoF. Using the Gröbner basis method, we have verified that for every matrix $\mathbf{A} \in \mathbb{R}^{2,3}$ , there are 8 ways to decompose it into pose parameters $R \in SO(3)$ , $C_{(1)} \in \mathbb{R}$ , $C_{(3)} \in \mathbb{R}$ , and a scale factor $\sigma \in \mathbb{R}$ . This means that $\mathbf{A}_i$ has no internal constraints, and every $2 \times 3$ matrix can represent $\mathbf{A}_i$ for some pose. + +Therefore, the relative pose problem can be transformed into finding matrices $\mathbf{A}_i\in \mathbb{R}^{2,3}$ , vectors $\mathbf{L}_{h,j}\in \mathbb{P}^2$ , and scalars $\lambda_{i,j}\in \mathbb{R}$ such that (8) holds for every $i\in \{1,\dots,n\}$ , $j\in \{1,\dots,m\}$ . This is equivalent to the problem of estimating 2D structure from images taken by $n$ uncalibrated 1D cameras, which has been studied in [55]. + +# 2.1.1. Ambiguities in the case with parallel lines + +The assumption that all lines are parallel introduces ambiguities into the relative pose estimation. First, since all line directions $\mathbf{L}_{d,j}$ are equal, the values of constraints (3) remain unchanged if we shift $\mathbf{C}_i(y_i)$ to $\mathbf{C}_i(y_i) + \gamma \mathbf{L}_{j,d}$ for any $\gamma \in \mathbb{R}$ . Therefore, it is only possible to estimate the camera centers up to translation along $\mathbf{L}_d$ . + +Furthermore, if (8) holds for some $\mathbf{A}_i\in \mathbb{R}^{2,3}$ , $\mathbf{L}_{h,j}\in \mathbb{P}^2$ , then (8) also holds for $\mathbf{A}_i\cdot \mathbf{H}^{-1}$ , $\mathbf{H}\cdot \mathbf{L}_{j,h}$ for every invertible $\mathbf{H}\in \mathbb{R}^{3,3}$ . Therefore, the cameras and the structure can only be estimated up to a projective transformation $\mathbf{H}$ . + +We show later in the paper that we can resolve the projective ambiguity by introducing priors on the cameras. To resolve the ambiguity in the camera centers, we can, for instance, observe one line orthogonal to $\mathbf{L}_d$ and use it to estimate the camera centers. + +# 3. Enumerating minimal problems + +In this section, we classify minimal problems for estimating the relative pose from projections of $n$ lines onto $m$ scanlines, along with their degrees. We consider five settings: + +# A. Fully generic setting + +# B. Parallel lines + +# C. Gravity prior for cameras + Generic lines + +# D. Gravity prior for cameras + Parallel lines + +# E. Gravity prior for cameras + Vertical lines + +In cases A and C, we estimate the pose up to global rotation, translation, and scale. In cases D and E, we estimate the pose up to these transformations and an additional shift of centers $\mathbf{C}_i(y_i)$ along direction $\mathbf{L}_d$ . In case B, we estimate the pose up to these transformations, the shift of centers along $\mathbf{L}_d$ , and the projective transformation $\mathbf{H}$ . + +For each setting, we determine the number of cameras $m$ and the number of lines $n$ s.t. the problem (3) is minimal. We assume complete visibility, meaning each line is projected into each camera. We follow concepts from [17]: Minimal problem [17, Def. 1]: A parametric polynomial system $f(\theta ,\mathbf{x}) = 0$ with a finite nonzero number of solutions $\mathbf{x}^*$ for a generic set of parameters $\theta$ . + +Balanced problem [17, Def. 2]: A parametric polynomial system $f(\theta, \mathbf{x}) = 0$ where the dimensionality of the space of variables $\mathbf{x}$ equals the number of independent equations. Condition for minimal problems [17, Step. 1, Corr. 2]: A problem $f(\theta, \mathbf{x}) = 0$ is minimal if and only if it is balanced and its Jacobian $\frac{\partial f(\theta, \mathbf{x}^*)}{\partial \mathbf{x}}$ is invertible for generic $\theta, \mathbf{x}^*$ . + +This summary provides the foundation for enumerating minimal problems. For detailed theory, refer to [17]. In our case, the parameter vector $\theta$ consists of projections $\mathbf{p}_{i,j} \in \mathbb{P}^2$ , and the variable vector $\mathbf{x}$ encodes camera poses $(\mathbf{R}_i(y_i), \mathbf{C}_i(y_i))$ and lines $(\mathbf{L}_{0,j}, \mathbf{L}_{d,j})$ . The polynomial system $f(\mathbf{x}, \theta)$ consists of equations in the form of (3). Following [17], we used the following steps to list all minimal problems and determine their degrees: + +- Enumerate all balanced problems for each setting A-E. +- For each balanced problem, generate random camera poses and lines $\mathbf{x}^*$ , reproject the lines into the cameras, and obtain projections $\theta$ such that $f(\theta, \mathbf{x}^*) = 0$ . +- Compute the Jacobian $\frac{\partial f(\theta, \mathbf{x}^*)}{\partial \mathbf{x}}$ . Label the problem as minimal if $\det \frac{\partial f(\theta, \mathbf{x}^*)}{\partial \mathbf{x}} \neq 0$ . +- Use monodromy computation from [13] to determine the degree of minimal problems. + +According to algebraic geometry, finding a single instance where the Jacobian is full-rank suffices to claim generic + +
S.mnMin.DegS.mnMin.Deg
A523Y389k+D37Y48
A217Y40k+D45Y232
B37Y2*D64Y1224
B46Y2*E35Y16
C515Y1.8M+E44Y32
C155Y532k+
+ +Table 1. List of balanced problems for settings A-E. S. denotes the setting, $m$ and $n$ are the number of cameras and lines, and Deg is the degree. Column Min. contains Y if the problem is minimal. $2^{*}$ indicates two solutions in terms of $2 \times 3$ matrices, and + denotes computations that did not terminate. See Sec. 3 for details. + +minimality. In the following subsections, we enumerate minimal problems for each setting A-E. + +A. Fully Generic Setting Each camera has 6 degrees of freedom (DoF), while each line has 4 DoF. Due to gauge freedom, we fix 7 DoF (3 for rotation, 3 for translation, and 1 for scale). There are $m \cdot n$ constraints in the form of (3), fixing $m \cdot n$ DoF. The balanced constraint is therefore: + +$$ +6 m + 4 n - 7 = m n. \tag {9} +$$ + +We define a function $n_A(m)$ , mapping the number of cameras $m$ to the number of lines $n$ , s.t. (9) holds for $m$ , $n_A(m)$ : + +$$ +n _ {A} (m) = \frac {6 m - 7}{m - 4}. \tag {10} +$$ + +As stated in Sec. 2, the minimal number of cameras is $m \geq 5$ . The function $n_A(m)$ is strictly decreasing for $m > 4$ , and since $\lim_{m \to \infty} n_A(m) = 6$ , no balanced problem has fewer than 7 lines. Evaluating $n_A(m)$ for integer values of $m$ from 5 onward, and searching for integer values of $n_A(m)$ , we identify the following balanced cases: + +$$ +\bullet m = 5, n = 2 3; \quad m = 2 1, n = 7 +$$ + +B. Parallel Lines As discussed in Sec. 2, the pose is estimated up to a shift along $\mathbf{L}_d$ and a projective transformation $\mathbf{H} \in \mathbb{R}^{3 \times 3}$ . The pose estimation corresponds to estimating matrices $\mathbf{A}_i \in \mathbb{R}^{2 \times 3}$ and points $\mathbf{L}_{h,j} \in \mathbb{P}^2$ such that (7) holds for all $i, j$ . Here, each camera has 5 DoF, each line has 2 DoF, and we fix 8 DoF due to projective ambiguity. The balanced constraint is: + +$$ +5 m + 2 n - 8 = m n. \tag {11} +$$ + +Defining $n_B(m)$ as: + +$$ +n _ {B} (m) = \frac {5 m - 8}{m - 2}, \tag {12} +$$ + +and noting that $m \geq 3$ , we find $\lim_{m \to \infty} n_B(m) = 5$ , implying no balanced problem has fewer than 6 lines. Evaluating $n_B(m)$ for integer $m$ , we obtain: + +$$ +\bullet m = 3, n = 7; \quad m = 4, n = 6 +$$ + +C. Gravity Prior for Cameras + Generic Lines With a known vertical direction, each camera has 4 DoF (1 for rotation, 3 for translation), and we fix 5 DoF due to gauge freedom. Lines remain generic, contributing 4 DoF each. The balanced constraint is: + +$$ +4 m + 4 n - 5 = m n. \tag {13} +$$ + +The corresponding function is: + +$$ +n _ {C} (m) = \frac {4 m - 5}{m - 4}. \tag {14} +$$ + +We find $\lim_{m\to \infty}n_C(m) = 4$ so no balanced problem has fewer than 5 lines. Checking integer values of $m$ , we get: + +$$ +\begin{array}{l} \bullet m = 5, n = 1 5; \quad m = 1 5, n = 5 \end{array} +$$ + +D. Gravity Prior for Cameras + Parallel Lines Here, camera rotation has 1 DoF, and the camera center has 2 DoF (up to translation along the line). We fix 4 DoF due to gauge freedom. Lines share 2 DoF for direction $\mathbf{L}_d$ , with each line having 2 additional DoF. The balanced constraint is: + +$$ +3 m + 2 + 2 n - 4 = m n. \tag {15} +$$ + +We get $n_D(m)$ as: + +$$ +n _ {D} (m) = \frac {3 m - 2}{m - 2}. \tag {16} +$$ + +For $m \geq 3$ , we find $\lim_{m \to \infty} n_D(m) = 3$ , meaning no balanced problem has fewer than 4 lines. We have found: + +$$ +\bullet m = 3, n = 7; \quad m = 4, n = 5; \quad m = 6, n = 4 +$$ + +E. Gravity Prior for Cameras + Vertical Lines Similar to the previous case, each camera has 3 DoF, and we fix 4 DoF due to gauge freedom. Since the direction $\mathbf{L}_d$ is fixed, each line has 2 DoF. The balanced constraint is: + +$$ +3 m + 2 n - 4 = m n. \tag {17} +$$ + +We define $n_E(m)$ as: + +$$ +n _ {E} (m) = \frac {3 m - 4}{m - 2}. \tag {18} +$$ + +For $m \geq 3$ , we find $\lim_{m \to \infty} n_E(m) = 3$ , meaning no balanced problem has fewer than 4 lines. We found problems: + +$$ +\bullet m = 3, n = 5; \quad m = 4, n = 4 +$$ + +All balanced problems for these settings are summarized in Table 1, along with their minimality status and degree. The table shows that problems with parallel/vertical lines (B, D, E) are minimal and computationally feasible. Problems with generic lines (A, C) are minimal but have degrees too high for practical RANSAC solutions. For problems $(A,5,23)$ and $(C,5,15)$ , monodromy computations did not terminate within a month. + +# 4. Minimal solvers + +In this section, we introduce solvers for the minimal problems $(B,3,7)$ , $(D,3,7)$ , $(E,3,5)$ , $(E,4,4)$ from Section 3. We use the notation and concepts from Sec. 2. + +# 4.1. Parallel lines with generic cameras + +Here, we describe the solution to the problem $(B,3,7)$ . Since the lines are parallel, our goal is to find matrices $\mathbf{A}_1,\mathbf{A}_2,\mathbf{A}_3\in \mathbb{R}^{2\times 3}$ and vectors $\mathbf{L}_{h,j}$ for $j\in \{1,\dots ,7\}$ , such that equation (7) holds for every camera $i$ and line $j$ . Due to projective ambiguity, we aim to recover the camera poses up to a common projective transformation $\mathbf{H}\in \mathbb{R}^{3\times 3}$ . Each camera triplet $\mathbf{A}_1,\mathbf{A}_2,\mathbf{A}_3$ can be expressed as: + +$$ +\mathbf {A} _ {1} = \mathbf {A} _ {c, 1} \cdot \mathbf {H}, \quad \mathbf {A} _ {2} = \mathbf {A} _ {c, 2} \cdot \mathbf {H}, \quad \mathbf {A} _ {3} = \mathbf {A} _ {c, 3} \cdot \mathbf {H}, \tag {19} +$$ + +for some $\mathbf{H} \in \mathbb{R}^{3 \times 3}$ , with $\mathbf{A}_{c,1}, \mathbf{A}_{c,2}, \mathbf{A}_{c,3}$ taking the form: + +$$ +\mathbf {A} _ {c, 1} = \left[ \begin{array}{l l l} 1 & 0 & 0 \\ \alpha_ {1} & \alpha_ {1} & \alpha_ {1} \end{array} \right], \mathbf {A} _ {c, 2} = \left[ \begin{array}{l l l} 1 & 0 & 0 \\ \alpha_ {2} & \alpha_ {3} & \alpha_ {4} \end{array} \right], \mathbf {A} _ {c, 3} = \left[ \begin{array}{l l l} 1 & 0 & 0 \\ \alpha_ {5} & \alpha_ {6} & \alpha_ {7} \end{array} \right]. \tag {20} +$$ + +The proof follows analogously to Proposition 2 in [29]. These canonical-form matrices have 7 DoF. + +To estimate $\mathbf{A}_{c,1},\mathbf{A}_{c,2},\mathbf{A}_{c,3}$ , we leverage the fact that, as shown in [55], a trifocal tensor exists between three camera matrices $\mathbf{A}_i\in \mathbb{R}^{2\times 3}$ . To compute this tensor, we stack the constraints from (8) into the matrix equation: + +$$ +\left[ \begin{array}{c c c c} \mathbf {A} _ {1} & \mathbf {u} _ {1, j} ^ {\prime} & 0 & 0 \\ \mathbf {A} _ {2} & 0 & \mathbf {u} _ {2, j} ^ {\prime} & 0 \\ \mathbf {A} _ {3} & 0 & 0 & \mathbf {u} _ {3, j} ^ {\prime} \end{array} \right] \left[ \begin{array}{c} \mathbf {L} _ {h, j} \\ - \lambda_ {1, j} \\ - \lambda_ {2, j} \\ - \lambda_ {3, j} \end{array} \right] = 0. \tag {21} +$$ + +For this equation to hold, the determinant of the coefficient matrix must vanish, leading to the constraint: + +$$ +\sum_ {a = 1} ^ {2} \sum_ {b = 1} ^ {2} \sum_ {c = 1} ^ {2} \mathbf {u} _ {1, j (a)} ^ {\prime} \mathbf {u} _ {2, j (b)} ^ {\prime} \mathbf {u} _ {3, j (c)} ^ {\prime} \mathbf {T} _ {(a, b, c)} = 0, \tag {22} +$$ + +where $\mathbf{T} \in \mathbb{R}^{2 \times 2 \times 2}$ is the trifocal tensor with elements: + +$$ +\mathbf {T} _ {(a, b, c)} = (- 1) ^ {a + b + c} \det \left(\left[ \mathbf {A} _ {1 (a,:)} ^ {T} \mathbf {A} _ {2 (b,:)} ^ {T} \mathbf {A} _ {3 (c,:)} ^ {T} \right]\right). \tag {23} +$$ + +As shown in [55], $\mathbf{T}$ has no internal constraints and it can be estimated using 7 line correspondences. + +Since $\mathbf{T}$ is homogeneous, it has 7 DoF, matching the DoF of the canonical camera matrices. We verified that for every tensor $\mathbf{T} \in \mathbb{R}^{2 \times 2 \times 2}$ , there are exactly two corresponding triplets of matrices $\mathbf{A}_{c,1}, \mathbf{A}_{c,2}, \mathbf{A}_{c,3}$ . The proof and the decomposition algorithm for recovering these matrices can be found in the SM. Once $\mathbf{A}_{c,1}, \mathbf{A}_{c,2}, \mathbf{A}_{c,3}$ are found, we can triangulate the lines up to the projective transformation $\mathbf{H}$ . + +The solver for the $(B,3,7)$ problem, which estimates the trifocal tensor $\mathbf{T}$ , operates as follows: + +1. Input: Projections $\mathbf{p}_{i,j}$ of 7 lines in 3 cameras. + +2. Compute matrices $\mathbf{R}_{0,i}$ and vectors $\mathbf{u}_{i,j}$ and $\mathbf{u}_{i,j}^{\prime}$ . +3. Construct the linear system (22). +4. Solve (22) using singular value decomposition (SVD) to estimate the tensor $\mathbf{T}$ . + +On average, the solver runs in $6.92\mu s$ + +# 4.2. Vertical lines and known gravity + +Here, we analyze the solutions to the problems $(E,3,5)$ and $(E,4,4)$ . We assume that all lines are vertical and the gravity directions $\mathbf{v}_i\in \mathbb{P}^2$ for the cameras are known. + +Similar to Sec. 4.1, we set $\mathbf{L}_{d,j} = \mathbf{e}_2$ . Therefore, Equation (6) still holds. Furthermore the vanishing points $\mathbf{v}_i$ for $i \in \{1,\dots,m\}$ associated with the vertical lines are known, and there holds $\mathbf{v}_i = \mathbf{R}_i(y_i)\mathbf{e}_2$ . We can decompose $\mathbf{R}_i(y_i) = \mathbf{R}_{v,i}\mathbf{R}_{y,i}$ , where $\mathbf{R}_{y,i}$ is a rotation around y-axis, and $\mathbf{R}_{v,i}$ is a rotation around an axis in the xz-plane, such that $\mathbf{R}_{v,i}\mathbf{e}_2 = \mathbf{v}_i$ . Since $\mathbf{R}_{v,i}$ is calculated directly from $\mathbf{v}_i$ , we assume it to be known. + +In this case, $\mathbf{R}_i^{\prime}$ (5) takes the form $\mathbf{R}_i^{\prime} = \mathbf{R}_{0,i}\mathbf{R}_{v,i}\mathbf{R}_{y,i}$ . While both $\mathbf{R}_{0,i}$ and $\mathbf{R}_{v,i}$ have their axes in the xz-plane, $\mathbf{R}_{0,i}\mathbf{R}_{v,i}$ does generically not share this property. However, we can decompose $\mathbf{R}_{0,i}\mathbf{R}_{v,i} = \mathbf{R}_{a,i}\mathbf{R}_{b,i}$ , s.t. $\mathbf{R}_{a,i}$ has its axis in the xz-plane and $\mathbf{R}_{b,i}$ rotates around the y-axis. + +Now, we decompose $\mathbf{R}_i^{\prime}$ as $\mathbf{R}_{a,i} \cdot \mathbf{R}_i^{\prime \prime}$ , where $\mathbf{R}_{a,i}$ is a known rotation around axis in the xz-plane, and $\mathbf{R}_i^{\prime \prime} = \mathbf{R}_{b,i}$ . $\mathbf{R}_{y,i}$ is an unknown rotation around y-axis. + +We represent $\mathbf{R}_{a,i},\mathbf{R}_i^{\prime \prime}$ with quaternions $\mathbf{q}_i = [\mathbf{q}_{i,w}\mathbf{q}_{i,x}0\mathbf{q}_{i,z}]^T$ $\mathbf{q}_i^{\prime \prime} = [\mathbf{q}_{i,w}^{\prime \prime}0\mathbf{q}_{i,y}^{\prime \prime}0]^T$ , write the elements of $\mathbf{R}'$ in the form + +$$ +\begin{array}{l} \mathbf {R} _ {i (1, 1)} ^ {\prime} = \left(1 - 2 \mathbf {q} _ {i, z} ^ {2}\right) \left(1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}\right) - 4 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime}, \\ \mathbf {R} _ {i (1, 3)} ^ {\prime} = 2 \left(1 - 2 \mathbf {q} _ {i, z} ^ {2}\right) \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} + 2 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} \left(1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}\right), \\ \mathbf {R} _ {i (3, 1)} ^ {\prime} = 2 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} (1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}) - 2 (1 - 2 \mathbf {q} _ {i, x} ^ {2}) \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime}, \\ \mathbf {R} _ {i (3, 3)} ^ {\prime} = \left(1 - 2 \mathbf {q} _ {i, x} ^ {2}\right) \left(1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}\right) + 4 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime}, \tag {24} \\ \end{array} +$$ + +and decompose the left $2 \times 2$ submatrix of $\mathbf{A}_i$ (7) as + +$$ +\left[ \begin{array}{l l} - \mathbf {R} _ {i, (1, 3)} ^ {\prime} & \mathbf {R} _ {i, (1, 1)} ^ {\prime} \\ - \mathbf {R} _ {i, (3, 3)} ^ {\prime} & \mathbf {R} _ {i, (3, 1)} ^ {\prime} \end{array} \right] = \mathbf {A} _ {i, a} \cdot \mathbf {A} _ {i, b}, \tag {25} +$$ + +with + +$$ +\mathbf {A} _ {i, a} = \left[ \begin{array}{l l} - 2 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} & (1 - 2 \mathbf {q} _ {i, z} ^ {2}) \\ - (1 - 2 \mathbf {q} _ {i, x} ^ {2}) & 2 \mathbf {q} _ {i, x} \mathbf {q} _ {i, z} \end{array} \right], \tag {26} +$$ + +$$ +\mathbf {A} _ {i, b} = \left[ \begin{array}{c c} \left(1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}\right) & 2 q _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} \\ - 2 q _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} & \left(1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}\right) \end{array} \right]. \tag {27} +$$ + +$\mathbf{A}_{i,a}$ can be directly obtained from the known gravity direction and the current scanline $y_{i}$ , and $\mathbf{A}_{i,b}$ is an unknown $2D$ rotation matrix. We define $\mathbf{A}_i^{\prime} = \mathbf{A}_{i,a}^{-1}\mathbf{A}_i$ , and $\mathbf{u}_{i,j}^{\prime \prime} = \mathbf{A}_{i,a}^{\prime T}\mathbf{u}_{i,j}$ . Then, there holds + +$$ +\mathbf {u} _ {i, j} ^ {\prime \prime T} \mathbf {A} _ {i} ^ {\prime} \mathbf {L} _ {j, h} = 0. \tag {28} +$$ + +Since the left $2 \times 2$ submatrix of $\mathbf{A}_i'$ is $\mathbf{A}_{i,b}$ , $\mathbf{A}_i'$ has a form of a calibrated $2 \times 3$ camera matrix, and the problem with known gravity direction and vertical lines is equivalent to estimating a 2D structure from images taken by $n$ calibrated cameras [8]. Unlike its uncalibrated counterpart, the calibrated problem can be solved without projective ambiguity. Solvers for the cases $m = 3, n = 5$ and $m = 4, n = 4$ have been presented in [8]. Notably, these numbers exactly correspond to the minimal problems discussed in Sec. 3. + +The solver for $m = 3, n = 5$ estimates a trifocal tensor of the form (23) from 5 point correspondences and 2 linear internal constraints, and solves a quadratic function to decompose the tensor into camera matrices. The solver for $m = 4, n = 4$ estimates a dual quadrifocal tensor from 4 point correspondences and 11 linear internal constraints, and also solves a quadratic function to decompose the tensor into camera matrices. In both cases, there are two solutions in terms of matrices $\mathbf{A}_i'$ . Typically, the correct solution is selected by counting the points lying in front of all cameras. + +After finding $\mathbf{A}_i^{\prime}$ with [8], we construct the camera poses $\mathbf{R}_i(y_i),\mathbf{C}_i(y_i)$ . For this, we first obtain $\mathbf{R}_i^{\prime \prime}$ . Because (28) is homogeneous, there are two possible matrices $\mathbf{R}_i^{\prime \prime}$ : + +$$ +\begin{array}{l} \mathbf {R} _ {i} ^ {\prime \prime} \in \left\{\left[ \begin{array}{c c c} (1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}) & 0 & 2 \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} \\ 0 & 1 & 0 \\ - 2 \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} & 0 & (1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}) \end{array} \right], \right. \tag {29} \\ \left. \left[ \begin{array}{c c c} - (1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}) & 0 & - 2 \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} \\ 0 & 1 & 0 \\ 2 \mathbf {q} _ {i, w} ^ {\prime \prime} \mathbf {q} _ {i, y} ^ {\prime \prime} & 0 & - (1 - 2 \mathbf {q} _ {i, y} ^ {\prime \prime 2}) \end{array} \right] \right\}. \\ \end{array} +$$ + +The existence of two choices of $\mathbf{R}_i^{\prime \prime}$ per $\mathbf{A}_i^\prime$ accounts for the presence of 16 solutions in the 3 view case and 32 solutions in the 4 view case. Finally, we construct $\mathbf{R}_i(y_i),\mathbf{C}_i(y_i)$ : + +$$ +\begin{array}{l} \mathbf {R} _ {i} (y _ {i}) = \mathbf {R} _ {0, i} ^ {T} \mathbf {R} _ {a, i} \mathbf {R} _ {i} ^ {\prime \prime}, \\ \mathbf {C} _ {i} \left(y _ {i}\right) = \left[ \begin{array}{l l} \mathbf {e} _ {1} & \mathbf {e} _ {3} \end{array} \right] \mathbf {A} _ {i (1: 2, 1: 2)} ^ {- 1} \mathbf {A} _ {i (1: 2, 3)}. \tag {30} \\ \end{array} +$$ + +Note, that $\mathbf{C}_i(y_i)_{(2)}$ is always zero. This is in accordance with the Section 2, which explains that the camera centers can only be estimated up to the shift along the direction $\mathbf{L}_d$ . + +To summarize, we perform the following steps to solve problems $(E,3,5)$ and $(E,4,4)$ : + +1. Input: Projections $\mathbf{p}_{i,j}$ , vertical directions $\mathbf{v}_i$ . +2. Calculate matrices $\mathbf{R}_{0,i},\mathbf{R}_{v,i},\mathbf{R}_{a,i}$ , and vectors $\mathbf{u}_{i,j}$ (7). +3. Compose matrices $\mathbf{A}_{i,a}$ (26) and vectors $\mathbf{u}_{i,j}^{\prime \prime}(28)$ . +4. Use the 3 view or the 4 view solver from [8] to find 2 sets of matrices $\mathbf{A}_i^{\prime}$ , $i\in \{1,\dots,m\}$ from $\mathbf{u}_{i,j}^{\prime \prime}$ . +5. Construct $\mathbf{R}_i^{\prime \prime}$ (29), $\mathbf{R}_i(y_i)$ , and $\mathbf{C}_i(y_i)$ (30). +On average, the 3 view solver takes $9.16\mu s$ , and the 4 view solver takes $70.84\mu s$ . + +# 4.3. Parallel lines and known gravity + +We present the solution to the problem $(D,3,7)$ . In this case, all lines share a common direction $\mathbf{L}_d$ , which is + +generically not aligned with the vertical direction $\mathbf{e}_2$ . Let $\mathbf{R}_d \in SO(3)$ be a rotation around an axis in the xz-plane, s.t. $\mathbf{L}_d = \mathbf{R}_d\mathbf{e}_2$ . Substituting this into (3), we get + +$$ +\mathbf {p} _ {i, j} ^ {T} \mathbf {R} _ {i} \left(y _ {i}\right) \left[ \mathbf {R} _ {d} \mathbf {e} _ {2} \right] _ {\times} \left(\mathbf {L} _ {0, j} - \mathbf {C} _ {i} \left(y _ {i}\right)\right) = 0, \tag {31} +$$ + +$$ +\mathbf {p} _ {i, j} ^ {T} \mathbf {R} _ {i} \left(y _ {i}\right) \mathbf {R} _ {d} \left[ \mathbf {e} _ {2} \right] _ {\times} \mathbf {R} _ {d} ^ {T} \left(\mathbf {L} _ {0, j} - \mathbf {C} _ {i} \left(y _ {i}\right)\right) = 0. \tag {32} +$$ + +We define $\mathbf{L}_{0,j}^{\prime} = \mathbf{R}_d^T\mathbf{L}_{0,j}$ , $\mathbf{C}_i^\prime = \mathbf{R}_d^T\mathbf{C}_i(y_i)$ and substitute these into (32): + +$$ +\mathbf {p} _ {i, j} ^ {T} \mathbf {R} _ {i} \left(y _ {i}\right) \mathbf {R} _ {d} \left[ \mathbf {e} _ {2} \right] _ {\times} \left(\mathbf {L} _ {0, j} ^ {\prime} - \mathbf {C} _ {i} ^ {\prime}\right) = 0. \tag {33} +$$ + +Following Sec. 4.2, we find rotation matrices $\mathbf{R}_{0,i}$ and $\mathbf{R}_{v,i}$ , decompose $\mathbf{R}_i(y_i)$ as $\mathbf{R}_{v,i}\mathbf{R}_{y,i}$ , and find vectors $\mathbf{u}_{i,j}^{\prime}$ according to (8). Then, we define $\mathbf{R}_{D,i} = \mathbf{R}_{0,i}\mathbf{R}_{v,i}\mathbf{R}_{y,i}\mathbf{R}_d$ , and $\mathbf{t}_{D,i} = -\mathbf{R}_{D,i}^{-1}[\mathbf{e}_2]_{\times}\mathbf{C}_i^{\prime}$ , and rewrite the constraint as + +$$ +\lambda_ {i, j} \mathbf {u} _ {i, j} ^ {\prime} = \left[ \begin{array}{l l l} - \mathbf {R} _ {D, i (1, 3)} & \mathbf {R} _ {D, i (1, 1)} & \mathbf {t} _ {D, i (1)} \\ - \mathbf {R} _ {D, i (3, 3)} & \mathbf {R} _ {D, i (3, 1)} & \mathbf {t} _ {D, i (3)} \end{array} \right] \left[ \begin{array}{l} \mathbf {L} _ {0, j (1)} ^ {\prime} \\ \mathbf {L} _ {0, j (3)} ^ {\prime} \\ 1 \end{array} \right], \tag {34} +$$ + +which can be written in matrix form as + +$$ +\lambda_ {i, j} \mathbf {u} _ {i, j} = \mathbf {A} _ {D, i} \mathbf {L} _ {h, j} ^ {\prime}. \tag {35} +$$ + +This equation has the same form as (8), ensuring the existence of a trifocal tensor $\mathbf{T} \in \mathbb{R}^{2,2,2}$ satisfying (22). We estimate this tensor linearly using 7 correspondences, as in Sec. 4.1. + +Let us now discuss how to decompose $\mathbf{T}$ to estimate the relative pose. First, we fix the coordinate frame by setting $\mathbf{R}_1(y_1) = I$ , $\mathbf{C}_1(y_1) = 0$ . Both the tensor $\mathbf{T}$ and the pose have 7 DoF. Using Gröbner basis, we have discovered that for every tensor $\mathbf{T} \in \mathbb{R}^{2,2,2}$ , there are 48 camera poses $\mathbf{R}_{y,2}, \mathbf{t}_{D,2}, \mathbf{R}_{y,3}, \mathbf{t}_{D,3}, \mathbf{R}_d$ , such that the matrices $\mathbf{A}_{D,i}$ constructed via (34) have the tensor $\mathbf{T}$ . + +Due to the high number of variables and equation complexity, we could not generate an elimination template for a Gröbner basis solver. Therefore, we solve this problem with homotopy continuation [18], with an average runtime $19089\mu s$ . + +Finally, we compose $\mathbf{A}_{i,D}$ according to (34), and the pose $\mathbf{R}_i(y_i),\mathbf{C}_i(y_i)$ as + +$$ +\begin{array}{l} \mathbf {R} _ {i} (y _ {i}) = \mathbf {R} _ {v, i} \mathbf {R} _ {y, i}, \\ \mathbf {C} _ {i} \left(y _ {i}\right) = R _ {d} \left[ \begin{array}{l l} \mathbf {e} _ {1} & \mathbf {e} _ {3} \end{array} \right] \mathbf {A} _ {D, i (1: 2, 1: 2)} ^ {- 1} \mathbf {A} _ {D, i (1: 2, 3)} \tag {36} \\ \end{array} +$$ + +To summarize, we perform the following steps to solve the problem $(D,3,7)$ : + +1. Input: Projections $\mathbf{p}_{i,j}$ , vertical directions $\mathbf{v}_i$ . +2. Calculate matrices $\mathbf{R}_{0,i},\mathbf{R}_{v,i}$ and vectors $\mathbf{u}_{i,j}$ (7). +3. Construct linear system (22) and solve it to get $\mathbf{T}$ . +4. Use homotopy continuation to decompose $\mathbf{T}$ into $\mathbf{R}_{y,2}$ , $\mathbf{t}_{D,2}$ , $\mathbf{R}_{y,3}$ , $\mathbf{t}_{3,D}$ , $\mathbf{R}_d$ . +5. Construct $\mathbf{R}_i(y_i)$ and $\mathbf{C}_i(y_i)$ using (36). + +![](images/564df0bbbfe616916c75a04a33e36c49e60c771b06e0ddae69bb51c83b93c721.jpg) +Figure 3. Histogram of tensor, rotation, and translation errors of the proposed solvers, over $10^{5}$ noiseless samples. In radians. + +# 4.4. Degenerate scenarios + +Here, we analyze the degenerate cases for the solvers introduced in Sections 4.1, 4.2, and 4.3. Specifically, we consider scenarios where some lines are coplanar or where camera centers are collinear or coincident. + +- Solvers $(E,3,5)$ , $(E,4,4)$ are degenerate if all $m$ lines are coplanar or if any two camera centers coincide. +- Solvers $(B,3,7)$ , $(D,3,7)$ are degenerate if at least 5 lines are coplanar or if any two camera centers coincide. +- There is no degeneracy if the camera centers are colinear. + +# 5. Experiments + +In this section, we evaluate the solvers $(B,3,7)$ , $(D,3,7)$ , $(E,3,5)$ , $(E,4,4)$ on both synthetic and real data. + +# 5.1. Synthetic experiments + +Numerical stability. We sample scanlines $y_{i}$ from $\mathcal{U}_{[-1,1]}$ and generate random camera rotations $\mathbf{R}_{i,GT}$ and centers $\mathbf{C}_{i,GT}$ . Since we use one scanline per image, we only sample one pose per image. We get the vertical directions as $v_{i} = \mathbf{R}_{i,GT}\mathbf{e}_{2}$ . To generate line $\mathbf{L}_j$ , we sample scalars $\mu_{j,x},\mu_{j,z}$ from normalized Gaussian distribution and set $\mathbf{L}_{0,j} = [\mu_{j,x}0\mu_{j,z} + 5]^T$ . For solver $(D,3,7)$ , we also generate scalars $\nu_{x},\nu_{z}$ , and set $L_{j,d} = [\nu_x1\nu_z]^T$ . In other cases, we use $L_{j,d} = \mathbf{e}_2$ . We generate $\mathbf{p}_{i,j}$ by projecting $\mathbf{L}_j$ into the camera with pose $(\mathbf{R}_{i,GT},\mathbf{C}_{i,GT})$ and intersecting the projected line with scanline $y_{i}$ . + +Before measuring the error, we transform both the ground truth pose, and the estimated pose $(\mathbf{R}_{i,\mathrm{est}},\mathbf{C}_{i,\mathrm{est}})$ into a canonical coordinate frame with $\mathbf{C}_1 = 0$ $\mathbf{R_{y,1}} = \mathbf{I},\mathbf{L_d} = \mathbf{e_1}$ . Since the y-coordinates of all camera centers are unobservable, we set them to zero. Then, we calculate the rotation error $err_{\mathbf{R}} = \max_{i = 1}^{m}(\angle \mathbf{R}_{i,\mathrm{est}}^{T}\mathbf{R}_{i,GT})$ and the translation error $err_{\mathbf{C}} = \max_{i = 1}^{m}\angle (\mathbf{C}_{i,GT},\mathbf{C}_{i,\mathrm{est}})$ . + +Additionally, we construct tensors $\mathbf{T}_{GT}$ from GT poses and estimate $\mathbf{T}_{est}$ using the solvers. For trifocal problems, $\mathbf{T}_{GT},\mathbf{T}_{est}$ are trifocal tensors (23), while for $(E,4,4)$ , they are dual quadrifocal tensors. We measure tensor error errT as the Frobenius norm of $\mathbf{T}_{GT} - \mathbf{T}_{est}$ . For each solver, we generate $10^{5}$ synthetic noise-free scenes and measure errR, errC, and errT. Results in Fig. 3 show solver stability. + +![](images/d808fa8edf1daa28750b34ebcb215b0c48d1687308749456a6175f163299a052.jpg) +Figure 4. Rotation, translation, and tensor errors averaged over 10000 successful runs as a function of the point and the vertical noise, measured from the real solution closest to the ground truth. If a solver returns only complex solutions, we report a failure. The failure rates are reported in the last row. Plots in the same row share the y-axis and plots in the same column share the x-axis. The fixed parameters are in the titles. + +Tests with noise. To evaluate the robustness to noise, we set focal length $f = 1000px$ , and add Gaussian noise with standard deviation $\frac{\sigma_p}{f}$ to each projection $\mathbf{p}_{i,j}$ . We also perturb vertical directions $\mathbf{v}_i$ by applying a random 3D rotation with angle $\sigma_v$ . Fig. 4 shows $err_{\mathbf{R}}$ , $err_{\mathbf{C}}$ , and $err_{\mathbf{T}}$ as the function of $\sigma_p$ , $\sigma_v$ , averaged over $10^4$ runs. The results confirm that, since the line detections tend to be relatively precise, our solvers exhibit reasonable behaviour under realistic noise levels. + +# 5.2. Real-world Experiments + +We evaluate our solvers using the Fastec dataset [43], which consists of 19 sequences captured from a moving car in Tübingen. Each sequence contains 34 rolling shutter (RS) images and two global shutter (GS) images for every RS image: one with the pose of the first scanline and one with the pose of the middle scanline. The images have a resolution of $640 \times 480$ px and a focal length $768$ px. + +Since GT poses are unavailable, we use COLMAP [61] on the GS images to obtain pseudo-GT poses $(\mathbf{R}_{i,F},\mathbf{C}_{i,F})$ for the first scanline of image $i$ and $(\mathbf{R}_{i,M},\mathbf{C}_{i,M})$ for the middle scanline. The pseudo-GT pose for scanline $y_{i}$ is interpolated as: $\mathbf{R}_{i,GT}(y_i) = \mathbf{R}_{i,M\to y_i}\mathbf{R}_{i,M},\mathbf{C}_{i,GT}(y_i) =$ $\mathbf{C}_{i,M} + \frac{h - 2y_i}{h} (\mathbf{C}_{i,F} - \mathbf{C}_{i,M})$ , where $\mathbf{R}_{i,M\to y_i}$ is a rotation matrix with the axis of $\mathbf{R}_{i,F}\mathbf{R}_{i,M}^{T}$ , and the angle $\frac{h - 2y_i}{h}$ times the angle of $\mathbf{R}_{i,F}\mathbf{R}_{i,M}^{T}$ . We evaluate the solvers in multi-view and single-view settings. + +Multi-view. We estimate poses for $m$ consecutive im + +
solver≤ 5°≤ 10°≤ 20°≤ 30°
MV(E,3,5)131015
(E,4,4)581114
(D,3,7)6101517
SV(E,3,5)33510
(E,4,4)24713
(D,3,7)4479
+ +ages, selecting one scanline per image. We use DeepLSD [49] and Gluestick [50] to detect and match the lines. Although the projections of lines by RS cameras are generically curves, the line-based methods work well in the case of Fastec dataset. To select scanlines $y_{1}, \ldots, y_{m}$ , we exhaustively search for the scanlines that maximize the number of matched lines intersecting them. Then, we find $p_{i,j}$ as the intersections of the detected lines and the scanlines. Following [12], we approximate the vertical direction by $[0 1 0]^T$ . This setting works well for many datasets, including Fastec, and does not require gravity direction measurement. + +Single-view. In this case, select $m$ scanlines in the same image and estimate the pose between them. We detect lines with DeepLSD [49], and select scanlines $y_{1},\ldots ,y_{m}$ with spacing at least 25px with the highest number of shared lines. We obtain $p_{i,j}$ by intersecting lines with scanlines. + +We evaluate the solvers in RANSAC [22]. We label a correspondence as inlier, if the triangulated line $\mathbf{L}_j$ lies in front of all cameras and its reprojection error $\varepsilon$ is below $1px$ . In the case of $(B,3,7)$ , we decompose the tensor into canonical cameras (20) before triangulation. Details about triangulation are in the SM. The score for a model is the sum over all lines of $(1px - \varepsilon)^2$ if $\varepsilon < 1px$ , and zero otherwise. We run RANSAC for 1000 iterations, and select the model with the highest score. We do not use local optimization. + +The results are summarized in Tables 2, 3 summarize the results for the solvers $(E,3,5)$ , $(E,4,4)$ , and $(D,3,7)$ , and Table 4 shows the tensor error for the solver $(B,3,7)$ . Although the total number of camera poses with an error below $10^{\circ}$ and $20^{\circ}$ remains relatively low, in Structure-from-Motion (SfM) applications, it is sufficient to estimate a single correct relative pose per sequence to initialize the reconstruction. Table 2 confirms that this is achievable for the vast majority of sequences. A more detailed evaluation and further experiments are provided in the SM. + +# 6. Limitations and Future Work + +While our approach shows promising results for rolling shutter relative pose estimation, several limitations remain. A key challenge is the lack of robust curve detectors and matchers, as rolling shutter distortion bends straight lines. + +Table 2. Number of sequences (out of 19) from the Fastec dataset [43], with at least one estimated pose with error below $5/10/20/30^{\circ}$ . MV: Multi-view, SV: Single-view. + +
solver≤ 10°≤ 20°≤ 30°
MV(E,3,5)1.0 %6.2 %21.9 %
(E,4,4)1.9 %7.9 %22.4 %
(D,3,7)5.6 %13.8 %21.7 %
SV(E,3,5)0.5 %2.5 %5.3 %
(E,4,4)1.2 %4.1 %10.0 %
(D,3,7)0.5 %2.2 %6.6 %
+ +Table 3. Percentage of camera poses from Fastec [43] with pose error below $10/20/30^{\circ}$ . MV: Multi-view, SV: Single-view. + +
settingMedian errorMinimum error
MV0.850.44
SV0.890.62
+ +Table 4. Median and Minimum tensor error of the $(B,3,7)$ solver, averaged over all sequences from Fastec [43]. A more detailed evaluation is in the SM. + +Additionally, our method relies on chirality constraints to resolve ambiguity, as noted in [8]. The method does not use any motion model, which makes it broadly applicable and model-agnostic. However, this universality comes at the cost of stability: we expect methods that incorporate motion models to be more stable. Our solvers are also limited to specific scene types and assume a known gravity direction. However, this reliance could be mitigated using scene knowledge, such as geometric relations between lines, in combination with solver $(B,3,7)$ . Finally, our method estimates pose only up to a shift along the line direction. Future work could integrate additional geometric constraints, improve feature detection, and integrate the solvers to a rolling-shutter SfM pipeline. + +# 7. Conclusion + +In this paper, we propose a new approach for estimating the relative pose of rolling shutter cameras using projections of lines onto multiple scanlines. This enables estimating relative pose of rolling shutter cameras without explicitly modelling camera motion. The proposed solvers provide an important building block for rolling-shutter SfM where each scanline has its pose computed independently, without relying on a global motion model. + +We identify minimal problems for different variations of the single-scanline relative pose problem, including cases with parallel lines, vertical lines, and known gravity direction. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors. To evaluate our approach, we conduct experiments on both synthetic data and real-world rolling shutter images from the Fastec dataset, considering scanlines from either the same image or different images. These experiments demonstrate that our solvers successfully estimate at least one correct pose in the majority of scenes, confirming their feasibility for SfM initialization. + +# References + +[1] Omar Ait-Aider and François Berry. Structure and kinematics triangulation with a rolling shutter stereo rig. In IEEE 12th International Conference on Computer Vision, ICCV 2009, Kyoto, Japan, September 27 - October 4, 2009, 2009. +[2] Omar Ait-Aider, Nicolas Andreff, Jean-Marc Lavest, and Philippe Martinet. Simultaneous object pose and velocity computation using a single view from a rolling shutter camera. In Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part II, 2006. 1 +[3] Omar Ait-Aider, Adrien Bartoli, and Nicolas Andreff. Kinematics from lines in a single rolling shutter image. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA, 2007. 1 +[4] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. R6P - rolling shutter absolute pose problem. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, 2015. 1 +[5] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. Rolling shutter absolute pose problem with known vertical direction. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. +[6] Cenek Albl, Zuzana Kukelova, Viktor Larsson, and Tomás Pajdla. Rolling shutter camera absolute pose. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1 +[7] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomás Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 1 +[8] Kalle Åström and Magnus Oskarsson. Solutions and ambiguities of the structure and motion problem for 1d retinal vision. J. Math. Imaging Vis., 2000, 2, 6, 8 +[9] Fang Bai, Agniva Sengupta, and Adrien Bartoli. Scanline homographies for rolling-shutter plane absolute pose. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 2022. 1, 2 +[10] Akash Bapat, Enrique Dunn, and Jan-Michael Frahm. Towards kilo-hertz 6-dof visual tracking using an egocentric cluster of rolling shutter cameras. IEEE Trans. Vis. Comput. Graph., 2016. 2 +[11] Akash Bapat, True Price, and Jan-Michael Frahm. Rolling shutter and radial distortion are features for high frame rate multi-camera tracking. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 2 +[12] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. StereoGlue: Robust estimation with single-point solvers. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 8 + +[13] Paul Breiding and Sascha Timme. Homotopycontinuation.jl: A package for homotopy continuation in julia. In Mathematical Software - ICMS 2018 - 6th International Conference, South Bend, IN, USA, July 24-27, 2018, Proceedings. Springer, 2018. 3 +[14] Mingdeng Cao, Zhihang Zhong, Jiahao Wang, Yinqiang Zheng, and Yujiu Yang. Learning adaptive warping for real-world rolling shutter correction. CoRR, 2022. 1 +[15] Yuchao Dai, Hongdong Li, and Laurent Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1 +[16] Juan Carlos Dibene, Yazmín Maldonado, Leonardo Trujillo, and Enrique Dunn. Prepare for ludicrous speed: Marker-based instantaneous binocular rolling shutter localization. IEEE Trans. Vis. Comput. Graph., 2022. 2 +[17] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. PLMP - point-line minimal problems in complete multiview visibility. IEEE Trans. Pattern Anal. Mach. Intell., 2024, 3 +[18] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 6 +[19] Bin Fan, Yuchao Dai, and Ke Wang. Rolling-shutter-stereo-aware motion estimation and image correction. Comput. Vis. Image Underst., 2021. 1 +[20] Bin Fan, Yuchao Dai, Zhiyuan Zhang, and Ke Wang. Differential sfm and image correction for a rolling shutter stereog. Image Vis. Comput., 2022. 1 +[21] Bin Fan, Yuchao Dai, and Hongdong Li. Rolling shutter inversion: Bring rolling shutter images to high framerate global shutter video. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1 +[22] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 1981. 8 +[23] Per-Erik Forssén and Erik Ringaby. Rectifying rolling shutter video from hand-held devices. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, 2010. 1 +[24] Marvin Anas Hahn, Kathlen Kohn, Orlando Marigliano, and Tomás Pajdla. Order-one rolling shutter cameras. CoRR, 2024. 1 +[25] Michiel Hazewinkel. Encyclopaedia of mathematics. 2002. 1 +[26] Johan Hedborg, Erik Ringaby, Per-Erik Forssén, and Michael Felsberg. Structure and motion estimation from rolling shutter video. In IEEE International Conference on Computer Vision Workshops, ICCV 2011 Workshops, Barcelona, Spain, November 6-13, 2011, 2011. 1 + +[27] Johan Hedborg, Per-Erik Forssén, Michael Felsberg, and Erik Ringaby. Rolling shutter bundle adjustment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012, 2012. 1 +[28] Radu Horaud, Roger Mohr, and Boguslaw Lorecki. On single-scanline camera calibration. IEEE Trans. Robotics Autom., 1993. 2 +[29] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 5 +[30] Chao Jia and Brian L. Evans. Probabilistic 3-d motion estimation for rolling shutter video rectification from visual and inertial measurements. In 14th IEEE International Workshop on Multimedia Signal Processing, MMSP 2012, Banff, AB, Canada, September 17-19, 2012, 2012. 1 +[31] Georg Klein and David William Murray. Parallel tracking and mapping on a camera phone. In Science & Technology Proceedings, 8th IEEE International Symposium on Mixed and Augmented Reality 2009, ISMAR 2009, Orlando, Florida, USA, October 19-22, 2009, 2009. 1 +[32] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, and Tomás Pajdla. Linear solution to the minimal absolute pose rolling shutter problem. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III, 2018. 1 +[33] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, Konrad Schindler, and Tomás Pajdla. Minimal rolling shutter absolute pose with unknown focal length and radial distortion. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V, 2020. 1 +[34] Yizhen Lao and Omar Ait-Aider. Rolling shutter homography and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2780-2793, 2021. 1 +[35] Yizhen Lao, Omar Ait-Aider, and Helder Araujo. Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognit. Lett., 2018. 1 +[36] Yizhen Lao, Omar Ait-Aider, and Adrien Bartoli. Rolling shutter pose and ego-motion estimation using shape-from-template. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part II, 2018. 1 +[37] Chang-Ryeol Lee and Kuk-Jin Yoon. Monocular visual odometry with a rolling shutter camera. CoRR, 2017. 1 +[38] Chang-Ryeol Lee and Kuk-Jin Yoon. Inertial-aided rolling shutter relative pose estimation. CoRR, 2017. 1 +[39] Chang-Ryeol Lee, Ju Hong Yoon, Min-Gyu Park, and Kuk-Jin Yoon. Gyroscope-aided relative pose estimation for rolling shutter cameras. CoRR, 2019. 1 +[40] Dongdong Li, Gongjian Wen, Bing Wei Hui, Shaohua Qiu, and Wenfei Wang. Cross-ratio invariant based line scan camera geometric calibration with static linear data. Optics and Lasers in Engineering, 2014. 2 + +[41] Moyang Li, Peng Wang, Lingzhe Zhao, Bangyan Liao, and Peidong Liu. Usb-nerf: Unrolling shutter bundle adjusted neural radiance fields. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024. 1 +[42] Bangyan Liao, Delin Qu, Yifei Xue, Huiqing Zhang, and Yizhen Lao. Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 1 +[43] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, 2, 7, 8 +[44] Ludovic Magerand, Adrien Bartoli, Omar Ait-Aider, and Daniel Pizarro. Global optimization of object pose and motion from a single rolling shutter image with automatic 2d-3d matching. In Computer Vision - ECCV 2012 - 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I, 2012. 1 +[45] Marci Meingast, Christopher Geyer, and Shankar Sastry. Geometric models of rolling-shutter cameras. CoRR, abs/cs/0503076, 2005. 1 +[46] Branislav Micusik and Georgios Evangelidis. Renormalization for initialization of rolling shutter visual-inertial odometry. Int. J. Comput. Vis., 2021. 1 +[47] Thanh-Tin Nguyen and Maxime Lhuillier. Adding synchronization and rolling shutter in multi-camera bundle adjustment. In Proceedings of the British Machine Vision Conference 2016, BMVC 2016, York, UK, September 19-22, 2016, 2016. 1 +[48] David Nister. An efficient solution to the five-point relative pose problem. IEEE Trans. Pattern Anal. Mach. Intell., 2004. 1 +[49] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 8 +[50] Rémi* Pautrat, Iago* Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. GlueStick: Robust image matching by sticking points and lines together. In International Conference on Computer Vision (ICCV), 2023. 8 +[51] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018, Lake Tahoe, NV, USA, March 12-15, 2018, 2018. 1 +[52] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in Manhattan world. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1 +[53] Delin Qu, Yizhen Lao, Zhigang Wang, Dong Wang, Bin Zhao, and Xuelong Li. Towards nonlinear-motion-aware and occlusion-robust rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1 + +[54] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1 +[55] Long Quan and Takeo Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Trans. Pattern Anal. Mach. Intell., 1997. 2, 3, 5 +[56] Vijay Rengarajan, A. N. Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1 +[57] Erik Ringaby and Per-Erik Forssén. Efficient video rectification and stabilisation for cell-phones. Int. J. Comput. Vis., 2012. 1 +[58] Olivier Saurer, Kevin Koser, Jean-Yves Bouguet, and Marc Pollefeys. Rolling shutter stereo. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 1 +[59] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. A minimal solution to the rolling shutter pose estimation problem. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2015, Hamburg, Germany, September 28 - October 2, 2015, 2015. 1 +[60] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. Sparse to dense 3d reconstruction from rolling shutter images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1 +[61] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 7 +[62] David Schubert, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. Direct sparse odometry with rolling shutter. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VIII, 2018. 1 +[63] Wei Shang, Dongwei Ren, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised learning to bring dual reversed rolling shutter images alive. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1 +[64] Ken Shoemake. Animating rotation with quaternion curves. In SIGGRAPH, 1985. 1 +[65] Henrik Stewénius, David Nister, Magnus Oskarsson, and Kalle Åström. Solutions to minimal generalized relative pose problems. Workshop on omnidirectional vision, 2005. 2 +[66] Daobilige Su, Asher Bender, and Salah Sukkarieh. Improved cross-ratio invariant-based intrinsic calibration of A hyperspectral line-scan camera. Sensors, 2018. 2 +[67] Yufen Sun, Gang Liu, and Yue Sun. An affine motion model for removing rolling shutter distortions. IEEE Signal Process. Lett., 2016. 1 +[68] Subeesh Vasu, Mahesh Mohan M. R., and A. N. Rajagopalan. Occlusion-aware rolling shutter rectification of 3d scenes. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 1 + +[69] Carlos Andres Luna Vázquez, Manuel Mazo, José Luis Lázaro, and Juan F. Vázquez. Calibration of line-scan cameras. IEEE Trans. Instrum. Meas., 2010. 2 +[70] Ke Wang, Bin Fan, and Yuchao Dai. Relative pose estimation for stereo rolling shutter cameras. In IEEE International Conference on Image Processing, ICIP 2020, Abu Dhabi, United Arab Emirates, October 25-28, 2020, 2020. 1 +[71] Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. Urs-nerf: Unordered rolling shutter bundle adjustment for neural radiance fields. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV, 2024. 1 +[72] Weilong Yan, Robby T. Tan, Bing Zeng, and Shuaicheng Liu. Deep homography mixture for single image rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1 +[73] Ming Yao, Zuyun Zhao, and Bugao Xu. Geometric calibration of line-scan camera using a planar pattern. J. Electronic Imaging, 2014. 2 +[74] Yongcong Zhang, Bangyan Liao, Yifei Xue, Chen Lu, Peidong Liu, and Yizhen Lao. RSL-BA: rolling shutter line bundle adjustment. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LIX, 2024. 1 +[75] Bingbing Zhuang and Quoc-Huy Tran. Image stitching and rectification for hand-held cameras. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, 2020. 1 +[76] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1 +[77] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1 \ No newline at end of file diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/images.zip b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7daa8cf9cb317d3043eb456e9aba49c0381e4dd9 --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bd13cab2549d48e42cc7039da5b34257bb7a5037517697304621e07798f9c69 +size 464746 diff --git a/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/layout.json b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25538acba6c2d54da85dc5c4c7331ebecef7890e --- /dev/null +++ b/2025/Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras/layout.json @@ -0,0 +1,18687 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "type": "text", + "content": "Single-Scanline Relative Pose Estimation for Rolling Shutter Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 144, + 221, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 144, + 221, + 157 + ], + "spans": [ + { + "bbox": [ + 166, + 144, + 221, + 157 + ], + "type": "text", + "content": "Petr Hruby" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 158, + 224, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 158, + 224, + 170 + ], + "spans": [ + { + "bbox": [ + 164, + 158, + 224, + 170 + ], + "type": "text", + "content": "ETH Zürich" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 171, + 268, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 268, + 184 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 268, + 184 + ], + "type": "text", + "content": "Rämistrasse 101, 8006 Zürich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 188, + 254, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 188, + 254, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 188, + 254, + 198 + ], + "type": "text", + "content": "petr.h Ruby@inf.ethz.ch" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 356, + 144, + 431, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 144, + 431, + 157 + ], + "spans": [ + { + "bbox": [ + 356, + 144, + 431, + 157 + ], + "type": "text", + "content": "Marc Pollefeys" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 298, + 158, + 489, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 158, + 489, + 171 + ], + "spans": [ + { + "bbox": [ + 298, + 158, + 489, + 171 + ], + "type": "text", + "content": "ETH Zürich / Microsoft Spatial AI Lab" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 172, + 467, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 172, + 467, + 184 + ], + "spans": [ + { + "bbox": [ + 321, + 172, + 467, + 184 + ], + "type": "text", + "content": "Rämistrasse 101, 8006 Zürich" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 188, + 463, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 188, + 463, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 188, + 463, + 198 + ], + "type": "text", + "content": "marc.pollefeys@inf.ethz.ch" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 251, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 515 + ], + "type": "text", + "content": "We propose a novel approach for estimating the relative pose between rolling shutter cameras using the intersections of line projections with a single scanline per image. This allows pose estimation without explicitly modeling camera motion. Alternatively, scanlines can be selected within a single image, enabling single-view relative pose estimation for scanlines of rolling shutter cameras. Our approach is designed as a foundational building block for rolling shutter structure-from-motion (SfM), where no motion model is required, and each scanline's pose can be computed independently. We classify minimal solvers for this problem in both generic and specialized settings, including cases with parallel lines and known gravity direction, assuming known intrinsic and no lens distortion. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors, by leveraging connections between this problem and the estimation of 2D structure from 1D cameras. Experiments on rolling shutter images from the Fastec dataset demonstrate the feasibility of our approach for initializing rolling shutter SfM, highlighting its potential for further development. The code will be made publicly available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 135, + 551 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 557, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 617 + ], + "type": "text", + "content": "Relative pose estimation is a fundamental problem in computer vision with applications in Structure from Motion [61], SLAM, multi-view stereo, and visual odometry. This paper focuses on estimating relative pose between rolling shutter (RS) cameras." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "Rolling shutter cameras capture images row by row. If the camera is static during capture, the resulting image is equivalent to that of a pinhole camera. However, if the camera moves, the rolling shutter effect distorts the image, making standard pinhole-based methods [48] unreliable." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "A common approach to handle rolling shutter distortion is modeling camera motion with parametric models such as SLERP [64], Cayley transformation [25], linearized" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 331, + 228, + 396, + 278 + ], + "blocks": [ + { + "bbox": [ + 331, + 228, + 396, + 278 + ], + "lines": [ + { + "bbox": [ + 331, + 228, + 396, + 278 + ], + "spans": [ + { + "bbox": [ + 331, + 228, + 396, + 278 + ], + "type": "image", + "image_path": "9167d9030ae607acdad8367af0c7b8ed46e8ed334a531b47fc5c5102eaeb54d1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 402, + 229, + 467, + 278 + ], + "blocks": [ + { + "bbox": [ + 402, + 229, + 467, + 278 + ], + "lines": [ + { + "bbox": [ + 402, + 229, + 467, + 278 + ], + "spans": [ + { + "bbox": [ + 402, + 229, + 467, + 278 + ], + "type": "image", + "image_path": "00ca0360a4f25ff5d06d72003db2394c31994a8bd8dcfb333683b0b05e46b345.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 474, + 229, + 539, + 278 + ], + "blocks": [ + { + "bbox": [ + 474, + 229, + 539, + 278 + ], + "lines": [ + { + "bbox": [ + 474, + 229, + 539, + 278 + ], + "spans": [ + { + "bbox": [ + 474, + 229, + 539, + 278 + ], + "type": "image", + "image_path": "423bd1bad9211b1cdc4ac82bfcc6f9d7fb39a98f41734eb1d188f4b5c6f96bd5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 363, + 287, + 503, + 356 + ], + "blocks": [ + { + "bbox": [ + 363, + 287, + 503, + 356 + ], + "lines": [ + { + "bbox": [ + 363, + 287, + 503, + 356 + ], + "spans": [ + { + "bbox": [ + 363, + 287, + 503, + 356 + ], + "type": "image", + "image_path": "5ffcf8b46cdcc7b0f1ea112eaec5af7df9692ec2d7fce9ea88d201e2e19c7626.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. We estimate a relative pose from the intersections of the line projections with a single scanline per image (green). This allows us to estimate relative pose between rolling shutter cameras without modelling the camera motion." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 433, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 529 + ], + "type": "text", + "content": "rotation [4], affine motion [67], pushbroom [46], accelerated motion [19], or order-one motion [24]. This has been successfully applied to absolute pose estimation [2-6, 9, 32, 33, 36, 44, 45, 59], bundle adjustment [27, 41, 42, 47, 60, 62, 71, 74], SfM [26, 31, 58], and motion rectification from single cameras [35, 51, 52, 56, 72, 77], camera rigs [1, 7, 20, 63], and videos [14, 19, 21, 23, 30, 53, 54, 57, 67, 68, 75]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "type": "text", + "content": "While most methods use point-based features, some employ lines for absolute pose estimation [3], bundle adjustment [74], and motion rectification [35, 51, 52, 56]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 714 + ], + "type": "text", + "content": "The problem of estimating relative pose for rolling shutter cameras has been explored in prior work. Special essential matrices for RS cameras under different motion models were introduced in [15], along with linear non-minimal solvers. The problem was also solved for pure translation models with constant velocity and acceleration [76]. IMU-based approaches reduced the required number of point correspondences from 11 to 5 [38, 39]. Other methods first estimate pose using a global shutter solver before refining with an RS model [37]. Additional works estimate relative pose for RS camera rigs [70] and for planar scenes [34]. Recently, [24] classified RS relative pose minimal problems" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7143" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "under the order-one motion model. Despite this, the general RS relative pose problem remains unsolved." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 179 + ], + "type": "text", + "content": "One drawback of parametric motion models is their reliance on an assumed motion that may not match real-world trajectories. An alternative is estimating pose independently for each scanline, similar to line-camera calibration [28, 40, 66, 69, 73]. This approach has been used for tracking a single camera [10, 11] and a camera rig [16] with respect to a predefined marker." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 310 + ], + "type": "text", + "content": "We explore the feasibility of estimating relative pose between scanlines of RS cameras, without explicit motion modelling. The goal is to use this method as a building block to initialize rolling-shutter SfM where no motion model is required and each scanline can have its pose independently computed. To the best of our knowledge, this is the first work that enables this. After the initialization, scanlines from remaining images may be registered by solving the generalized relative pose problem [65], similarly to [10, 11, 16], or, in the case of a planar scene, using [9]. Our contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 430 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "type": "text", + "content": "- Formulation of the problem of relative pose estimation from line projections in a single scanline per image, assuming known intrinsics and no lens distortion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 371 + ], + "type": "text", + "content": "- Enumeration of minimal problems for general and specialized settings with parallel lines and gravity priors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 418 + ], + "type": "text", + "content": "- Demonstration that the problems with parallel lines and with vertical lines and gravity prior correspond to estimating relative pose of uncalibrated cameras [55], respectively calibrated [8] 1D camera pose." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 419, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 430 + ], + "type": "text", + "content": "- Development of minimal solvers for the following cases:" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 64, + 432, + 294, + 478 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 64, + 432, + 294, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 432, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 64, + 432, + 294, + 442 + ], + "type": "text", + "content": "- 3 cameras, 7 parallel lines (with projective ambiguity)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 444, + 237, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 444, + 237, + 454 + ], + "spans": [ + { + "bbox": [ + 64, + 444, + 237, + 454 + ], + "type": "text", + "content": "- 3 cameras, gravity prior, 7 parallel lines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 456, + 237, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 456, + 237, + 466 + ], + "spans": [ + { + "bbox": [ + 64, + 456, + 237, + 466 + ], + "type": "text", + "content": "- 3 cameras, gravity prior, 5 vertical lines." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 468, + 237, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 468, + 237, + 478 + ], + "spans": [ + { + "bbox": [ + 64, + 468, + 237, + 478 + ], + "type": "text", + "content": "- 4 cameras, gravity prior, 4 vertical lines." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 479, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 515 + ], + "type": "text", + "content": "- Evaluation of solvers on synthetic and real RS datasets (Fastec, [43]), taking scanlines either from different images or from a single image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 521, + 119, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 119, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 119, + 533 + ], + "type": "text", + "content": "1.1. Notation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": "We use bold letters for matrices and vectors, and normal font for scalars. Indexing starts from one, with indices in brackets; e.g., the first element of " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d(1)}" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": ". Matrix indexing follows Matlab style, where " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{(:,1:2)}" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": " selects the first two rows of " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": ". The scanlines are denoted with a scalar " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": " giving their y-coordinate. The set of all 3D lines is denoted as " + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "inline_equation", + "content": "Gr(1,3)" + }, + { + "bbox": [ + 55, + 539, + 296, + 623 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 634, + 181, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 181, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 181, + 645 + ], + "type": "text", + "content": "2. Problem Formulation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 689 + ], + "type": "text", + "content": "This section describes the geometry of a line projected into a rolling-shutter camera and derives an algebraic constraint useful for estimating camera poses." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Consider " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " rolling-shutter cameras capturing the same 3D scene. Due to motion, the cameras capture different" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 353, + 70, + 515, + 150 + ], + "blocks": [ + { + "bbox": [ + 353, + 70, + 515, + 150 + ], + "lines": [ + { + "bbox": [ + 353, + 70, + 515, + 150 + ], + "spans": [ + { + "bbox": [ + 353, + 70, + 515, + 150 + ], + "type": "image", + "image_path": "f0802ace47166ecf5d8b9c5b091b33e6bf11be274f82906f00bdc096d658c563.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": "Figure 2. Illustration of the line projection. Line " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": " projects onto scanline " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": ". Then, the ray " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": " obtained by backprojecting " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": " intersects " + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 313, + 157, + 555, + 191 + ], + "type": "text", + "content": ", yielding the constraint (3)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "content": "rows at different poses. Let " + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y) \\in SO(3)" + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "content": " denote the orientation and center of camera " + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "content": " at row " + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 211, + 553, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "content": "We aim to estimate the relative pose between the cameras without explicitly modeling their motion functions. To achieve this, we select a single row " + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "content": " per camera and measure intersections between these selected rows and the projections of " + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "content": " straight lines " + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j \\in Gr(1,3)" + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "content": ". Then, we use this information to recover the camera poses " + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 313, + 236, + 553, + 319 + ], + "type": "text", + "content": ". Figure 1 illustrates the setup." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j} \\in \\mathbb{P}^2" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "p_{i,j} = [x_{i,j} y_i 1]^T" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": ", denote the projection of line " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": " onto row " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": " of camera " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": ". Then, the back-projected ray " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i,j}" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": " intersects line " + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 313, + 319, + 553, + 356 + ], + "type": "text", + "content": ", giving:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 339, + 363, + 553, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 363, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 339, + 363, + 553, + 378 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {i, j} = \\mathbf {C} _ {i} (y _ {i}) + \\lambda_ {i, j} \\mathbf {R} _ {i} (y _ {i}) ^ {T} \\mathbf {p} _ {i, j}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {1}", + "image_path": "c6ed30b246c8d7c1af6c8928877ac6a52c9d644d7f4b8e1b963b0567384a770f.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": "This situation is depicted in Fig. 2. Let us represent " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": " with a pair " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "(\\mathbf{L}_{0,j},\\mathbf{L}_{d,j})" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j}" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": " is a point on the line and " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d,j}" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": " is its direction. The ray " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{i,j}" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": " intersects " + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 313, + 385, + 553, + 422 + ], + "type": "text", + "content": " if and only if:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 429, + 553, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 429, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 318, + 429, + 553, + 454 + ], + "type": "interline_equation", + "content": "\\exists \\alpha , \\beta \\in \\mathbb {R}: \\quad \\mathbf {C} _ {i} \\left(y _ {i}\\right) + \\alpha \\mathbf {R} _ {i} \\left(y _ {i}\\right) ^ {T} \\mathbf {p} _ {i, j} = \\mathbf {L} _ {0, j} + \\beta \\mathbf {L} _ {d, j}. \\tag {2}", + "image_path": "cfdcb92dbf893086063886c202c41e93161508c12a759b4289ee4645bc2155be.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "text", + "content": "This is equivalent to requiring that vectors " + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i)^T\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d,j}" + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j} - \\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 313, + 454, + 553, + 502 + ], + "type": "text", + "content": " are linearly dependent. Since three 3D vectors are linearly dependent if and only if their scalar triple product vanishes, we obtain:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 352, + 510, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 510, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 352, + 510, + 553, + 526 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {L} _ {d, j} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {3}", + "image_path": "c19527f2d571ac9d027346764725e7485ec5809555ee91cd4a99a305b0dd231a.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 582 + ], + "type": "text", + "content": "Each projection provides a single constraint on the unknowns (lines and camera poses). A 3D line has four degrees of freedom, requiring four scanlines for triangulation and at least five scanlines for constraining relative pose." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 447, + 601 + ], + "type": "text", + "content": "2.1. Assuming parallel lines" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "Let us consider the scenario where all lines are parallel, which frequently occurs in real-world settings. Here, all lines share a common direction " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": ". Since Eq. (3) is invariant to scene rotation, we align " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " with the y-axis, " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_2" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": ". This allows representing each line " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " with a single point " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j}" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " in the xz-plane, with its y-coordinate equal to zero." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "With a shared direction, each line has only two individual degrees of freedom (DoF). Once " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " is known, a line can be triangulated from just two cameras. This reduces the" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7144" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "minimum number of cameras needed for pose estimation to three. To simplify Eq. (3), we substitute " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_2" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d,j}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 102, + 295, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 295, + 141 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 295, + 141 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} (y _ {i}) \\left(\\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (3)} \\\\ 0 \\\\ - \\mathbf {L} _ {0, j (1)} \\end{array} \\right] - \\left[ \\begin{array}{c} \\mathbf {C} _ {i} (y _ {i}) _ {(3)} \\\\ 0 \\\\ - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} \\end{array} \\right]\\right) = 0. \\quad (4)", + "image_path": "129852eefb829e061f3ebe2d41c879aa141548dbe7ae73504e84d5a9bc1ab344.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": "We define rotation " + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i} \\in SO(3)" + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": " around the x-axis, which maps row " + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}' = \\mathbf{R}_{0,i}\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": ". Then, there holds " + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}' = [x_{i,j}'01]^T" + }, + { + "bbox": [ + 55, + 145, + 296, + 185 + ], + "type": "text", + "content": ". Defining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 194, + 191, + 295, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 191, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 194, + 191, + 295, + 211 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {i} ^ {\\prime} = \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}), \\tag {5}", + "image_path": "01efdc978c9555ac30bd900a50e41b963afb0af565bf548b5e630ec816f644b8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 207, + 270, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 207, + 270, + 221 + ], + "spans": [ + { + "bbox": [ + 80, + 207, + 270, + 221 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {i} ^ {\\prime} = - \\mathbf {R} _ {0, i} \\mathbf {R} _ {i} (y _ {i}) [ \\mathbf {C} _ {i} (y _ {i}) _ {(3)} 0 - \\mathbf {C} _ {i} (y _ {i}) _ {(1)} ] ^ {T},", + "image_path": "ea1195c3fa97f50a0f6699af4dfe4eb74b8f699b7fe0e75f9280d61021ae3415.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "text", + "content": "and substituting " + }, + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime},\\mathbf{t}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}^{\\prime}" + }, + { + "bbox": [ + 55, + 226, + 282, + 240 + ], + "type": "text", + "content": " into Eq. (4), we obtain:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 245, + 295, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 295, + 293 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l l} x _ {i, j} ^ {\\prime} & 1 \\end{array} \\right] ^ {T} \\left[ \\begin{array}{c c c} - \\mathbf {R} _ {i (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i (1, 1)} ^ {\\prime} & \\mathbf {t} _ {i (0)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i (3, 1)} ^ {\\prime} & \\mathbf {t} _ {i (2)} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {0, j (1)} \\\\ \\mathbf {L} _ {0, j (3)} \\\\ 1 \\end{array} \\right] = 0. \\tag {6}", + "image_path": "86033ce8c1a15166a04aadf38b4480f2be8ada682670b569dd3c7e4691811ad4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 294, + 241, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 241, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 241, + 305 + ], + "type": "text", + "content": "We can write Equation (6) in a vector form as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 311, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 311, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 140, + 311, + 295, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {i, j} ^ {T} \\mathbf {A} _ {i} \\mathbf {L} _ {h, j} = 0, \\tag {7}", + "image_path": "b35ae6ede69e7203a63c75eae429aa52fe09794aedb40c6b7836b4fd1614eda9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{2,3}" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": " is a matrix encoding the camera pose, " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}\\in \\mathbb{P}^1" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": " is a vector encoding the observation, and " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{h,j}\\in" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbb{P}^2" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": " is a homogeneous representation of point " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j}" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": ". Alternatively, we define " + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime} = [1 - x_{i,j}^{\\prime}]^{T}" + }, + { + "bbox": [ + 55, + 331, + 296, + 381 + ], + "type": "text", + "content": ", and rewrite (7) as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 110, + 390, + 295, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 390, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 110, + 390, + 295, + 404 + ], + "type": "interline_equation", + "content": "\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\mathbf {A} _ {i} \\mathbf {L} _ {j, h}, \\quad \\lambda_ {i, j} \\in \\mathbb {R}. \\tag {8}", + "image_path": "50116fb69920da6f593830c94c188a605f732cfd62ad4b8a50b7d9ce8d238c24.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": "Since constraint (7) is homogeneous, matrix " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": " has 5 DoF. The elements " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i)" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)_{(1)}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)_{(3)}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": " of the pose, which appear in " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", also have 5 DoF. Using the Gröbner basis method, we have verified that for every matrix " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{2,3}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", there are 8 ways to decompose it into pose parameters " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "R \\in SO(3)" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "C_{(1)} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "C_{(3)} \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ", and a scale factor " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": ". This means that " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": " has no internal constraints, and every " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "2 \\times 3" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": " matrix can represent " + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 502 + ], + "type": "text", + "content": " for some pose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": "Therefore, the relative pose problem can be transformed into finding matrices " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{2,3}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": ", vectors " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{h,j}\\in \\mathbb{P}^2" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": ", and scalars " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": " such that (8) holds for every " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,n\\}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "j\\in \\{1,\\dots,m\\}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": ". This is equivalent to the problem of estimating 2D structure from images taken by " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": " uncalibrated 1D cameras, which has been studied in [55]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 579, + 261, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 261, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 261, + 591 + ], + "type": "text", + "content": "2.1.1. Ambiguities in the case with parallel lines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": "The assumption that all lines are parallel introduces ambiguities into the relative pose estimation. First, since all line directions " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d,j}" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": " are equal, the values of constraints (3) remain unchanged if we shift " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i) + \\gamma \\mathbf{L}_{j,d}" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": ". Therefore, it is only possible to estimate the camera centers up to translation along " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Furthermore, if (8) holds for some " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{2,3}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{h,j}\\in \\mathbb{P}^2" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", then (8) also holds for " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\cdot \\mathbf{H}^{-1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\cdot \\mathbf{L}_{j,h}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " for every invertible " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{3,3}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Therefore, the cameras and the structure can only be estimated up to a projective transformation " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": "We show later in the paper that we can resolve the projective ambiguity by introducing priors on the cameras. To resolve the ambiguity in the camera centers, we can, for instance, observe one line orthogonal to " + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 72, + 554, + 132 + ], + "type": "text", + "content": " and use it to estimate the camera centers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 142, + 494, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 494, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 494, + 156 + ], + "type": "text", + "content": "3. Enumerating minimal problems" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "content": "In this section, we classify minimal problems for estimating the relative pose from projections of " + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "content": " lines onto " + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "content": " scanlines, along with their degrees. We consider five settings:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 200, + 421, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 421, + 212 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 421, + 212 + ], + "type": "text", + "content": "A. Fully generic setting" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 213, + 389, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 213, + 389, + 222 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 389, + 222 + ], + "type": "text", + "content": "B. Parallel lines" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 224, + 511, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 511, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 511, + 236 + ], + "type": "text", + "content": "C. Gravity prior for cameras + Generic lines" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 236, + 510, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 510, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 510, + 247 + ], + "type": "text", + "content": "D. Gravity prior for cameras + Parallel lines" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 248, + 511, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 248, + 511, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 248, + 511, + 259 + ], + "type": "text", + "content": "E. Gravity prior for cameras + Vertical lines" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "content": "In cases A and C, we estimate the pose up to global rotation, translation, and scale. In cases D and E, we estimate the pose up to these transformations and an additional shift of centers " + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "content": " along direction " + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "content": ". In case B, we estimate the pose up to these transformations, the shift of centers along " + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "content": ", and the projective transformation " + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 262, + 554, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": "For each setting, we determine the number of cameras " + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": " and the number of lines " + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": " s.t. the problem (3) is minimal. We assume complete visibility, meaning each line is projected into each camera. We follow concepts from [17]: Minimal problem [17, Def. 1]: A parametric polynomial system " + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "inline_equation", + "content": "f(\\theta ,\\mathbf{x}) = 0" + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": " with a finite nonzero number of solutions " + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^*" + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": " for a generic set of parameters " + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 334, + 554, + 417 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": "Balanced problem [17, Def. 2]: A parametric polynomial system " + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "inline_equation", + "content": "f(\\theta, \\mathbf{x}) = 0" + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": " where the dimensionality of the space of variables " + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": " equals the number of independent equations. Condition for minimal problems [17, Step. 1, Corr. 2]: A problem " + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "inline_equation", + "content": "f(\\theta, \\mathbf{x}) = 0" + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": " is minimal if and only if it is balanced and its Jacobian " + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}}" + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": " is invertible for generic " + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta, \\mathbf{x}^*" + }, + { + "bbox": [ + 313, + 417, + 554, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": "This summary provides the foundation for enumerating minimal problems. For detailed theory, refer to [17]. In our case, the parameter vector " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": " consists of projections " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j} \\in \\mathbb{P}^2" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": ", and the variable vector " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": " encodes camera poses " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_i(y_i), \\mathbf{C}_i(y_i))" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": " and lines " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "(\\mathbf{L}_{0,j}, \\mathbf{L}_{d,j})" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": ". The polynomial system " + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}, \\theta)" + }, + { + "bbox": [ + 313, + 491, + 554, + 586 + ], + "type": "text", + "content": " consists of equations in the form of (3). Following [17], we used the following steps to list all minimal problems and determine their degrees:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 588, + 553, + 687 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 313, + 588, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 545, + 599 + ], + "type": "text", + "content": "- Enumerate all balanced problems for each setting A-E." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "text", + "content": "- For each balanced problem, generate random camera poses and lines " + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^*" + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "text", + "content": ", reproject the lines into the cameras, and obtain projections " + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "inline_equation", + "content": "f(\\theta, \\mathbf{x}^*) = 0" + }, + { + "bbox": [ + 313, + 600, + 553, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "text", + "content": "- Compute the Jacobian " + }, + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}}" + }, + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "text", + "content": ". Label the problem as minimal if " + }, + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\det \\frac{\\partial f(\\theta, \\mathbf{x}^*)}{\\partial \\mathbf{x}} \\neq 0" + }, + { + "bbox": [ + 313, + 635, + 553, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 664, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 553, + 687 + ], + "type": "text", + "content": "- Use monodromy computation from [13] to determine the degree of minimal problems." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "According to algebraic geometry, finding a single instance where the Jacobian is full-rank suffices to claim generic" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7145" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 296, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 296, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 296, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 296, + 164 + ], + "type": "table", + "html": "
S.mnMin.DegS.mnMin.Deg
A523Y389k+D37Y48
A217Y40k+D45Y232
B37Y2*D64Y1224
B46Y2*E35Y16
C515Y1.8M+E44Y32
C155Y532k+
", + "image_path": "c712fc3ee14f2088beeca32433b54caad13316f790b2deec4d3e0fa702da7cf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "text", + "content": "Table 1. List of balanced problems for settings A-E. S. denotes the setting, " + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "text", + "content": " are the number of cameras and lines, and Deg is the degree. Column Min. contains Y if the problem is minimal. " + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "inline_equation", + "content": "2^{*}" + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "text", + "content": " indicates two solutions in terms of " + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "inline_equation", + "content": "2 \\times 3" + }, + { + "bbox": [ + 55, + 166, + 296, + 222 + ], + "type": "text", + "content": " matrices, and + denotes computations that did not terminate. See Sec. 3 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 266 + ], + "type": "text", + "content": "minimality. In the following subsections, we enumerate minimal problems for each setting A-E." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": "A. Fully Generic Setting Each camera has 6 degrees of freedom (DoF), while each line has 4 DoF. Due to gauge freedom, we fix 7 DoF (3 for rotation, 3 for translation, and 1 for scale). There are " + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "inline_equation", + "content": "m \\cdot n" + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": " constraints in the form of (3), fixing " + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "inline_equation", + "content": "m \\cdot n" + }, + { + "bbox": [ + 55, + 278, + 296, + 338 + ], + "type": "text", + "content": " DoF. The balanced constraint is therefore:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 345, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 345, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 131, + 345, + 295, + 357 + ], + "type": "interline_equation", + "content": "6 m + 4 n - 7 = m n. \\tag {9}", + "image_path": "1db17ee7f165a5914c00bf96f9ad85d2fcfcac7a29e6f9f7b32f66be9a904723.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": "We define a function " + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "inline_equation", + "content": "n_A(m)" + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": ", mapping the number of cameras " + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": " to the number of lines " + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": ", s.t. (9) holds for " + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "inline_equation", + "content": "n_A(m)" + }, + { + "bbox": [ + 55, + 364, + 296, + 389 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 396, + 295, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 396, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 134, + 396, + 295, + 420 + ], + "type": "interline_equation", + "content": "n _ {A} (m) = \\frac {6 m - 7}{m - 4}. \\tag {10}", + "image_path": "e0c606db72311729836f1039028f464172eabe42c2170820f7a5e3a55d4c42c7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": "As stated in Sec. 2, the minimal number of cameras is " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "m \\geq 5" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": ". The function " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "n_A(m)" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": " is strictly decreasing for " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "m > 4" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": ", and since " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\lim_{m \\to \\infty} n_A(m) = 6" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": ", no balanced problem has fewer than 7 lines. Evaluating " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "n_A(m)" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": " for integer values of " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": " from 5 onward, and searching for integer values of " + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "inline_equation", + "content": "n_A(m)" + }, + { + "bbox": [ + 55, + 422, + 296, + 494 + ], + "type": "text", + "content": ", we identify the following balanced cases:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 495, + 205, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 205, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 205, + 507 + ], + "type": "interline_equation", + "content": "\\bullet m = 5, n = 2 3; \\quad m = 2 1, n = 7", + "image_path": "475c7ab2965e681659c4ad51b57dff2f7daf9e8823358ed08ef2d043f31cd17e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": "B. Parallel Lines As discussed in Sec. 2, the pose is estimated up to a shift along " + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": " and a projective transformation " + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": ". The pose estimation corresponds to estimating matrices " + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": " and points " + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{h,j} \\in \\mathbb{P}^2" + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": " such that (7) holds for all " + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 55, + 521, + 296, + 604 + ], + "type": "text", + "content": ". Here, each camera has 5 DoF, each line has 2 DoF, and we fix 8 DoF due to projective ambiguity. The balanced constraint is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 612, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 612, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 131, + 612, + 295, + 624 + ], + "type": "interline_equation", + "content": "5 m + 2 n - 8 = m n. \\tag {11}", + "image_path": "6177eec659b4077194415deb017a024a2e71e496453ba29889169e696253920a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 631, + 139, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 139, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 139, + 643 + ], + "type": "text", + "content": "Defining " + }, + { + "bbox": [ + 55, + 631, + 139, + 643 + ], + "type": "inline_equation", + "content": "n_B(m)" + }, + { + "bbox": [ + 55, + 631, + 139, + 643 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 134, + 648, + 295, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 648, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 134, + 648, + 295, + 673 + ], + "type": "interline_equation", + "content": "n _ {B} (m) = \\frac {5 m - 8}{m - 2}, \\tag {12}", + "image_path": "f840f6eb42066d3276681f2356a7752aefe5b10cf8c25032377dcf9b37c77d3e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "and noting that " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we find " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lim_{m \\to \\infty} n_B(m) = 5" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", implying no balanced problem has fewer than 6 lines. Evaluating " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "n_B(m)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " for integer " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we obtain:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 453, + 83 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 453, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 453, + 83 + ], + "type": "interline_equation", + "content": "\\bullet m = 3, n = 7; \\quad m = 4, n = 6", + "image_path": "b732bdd985a7c8fd0e2aad1db22c3727767e18c5a8ed37c5c712c701d5f08e45.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 98, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 157 + ], + "type": "text", + "content": "C. Gravity Prior for Cameras + Generic Lines With a known vertical direction, each camera has 4 DoF (1 for rotation, 3 for translation), and we fix 5 DoF due to gauge freedom. Lines remain generic, contributing 4 DoF each. The balanced constraint is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 170, + 555, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 170, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 390, + 170, + 555, + 182 + ], + "type": "interline_equation", + "content": "4 m + 4 n - 5 = m n. \\tag {13}", + "image_path": "8256b7affaca034bb515028dcf940402bea1e3b41456d5fec6c6fb4a455b34a6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 186, + 441, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 186, + 441, + 198 + ], + "spans": [ + { + "bbox": [ + 314, + 186, + 441, + 198 + ], + "type": "text", + "content": "The corresponding function is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 392, + 201, + 554, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 201, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 392, + 201, + 554, + 225 + ], + "type": "interline_equation", + "content": "n _ {C} (m) = \\frac {4 m - 5}{m - 4}. \\tag {14}", + "image_path": "11f891ad606c3e4e0acace581fa2ca7e6abeb75d8480579579d1ec54e86aa3bc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "text", + "content": "We find " + }, + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "inline_equation", + "content": "\\lim_{m\\to \\infty}n_C(m) = 4" + }, + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "text", + "content": " so no balanced problem has fewer than 5 lines. Checking integer values of " + }, + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 230, + 554, + 254 + ], + "type": "text", + "content": ", we get:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 255, + 462, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 255, + 462, + 265 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 462, + 265 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bullet m = 5, n = 1 5; \\quad m = 1 5, n = 5 \\end{array}", + "image_path": "32d76e773e79f1cc4ae0625b8e7f70ea30a06163ccd355bb1df8bee8329e1ed3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 279, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 340 + ], + "type": "text", + "content": "D. Gravity Prior for Cameras + Parallel Lines Here, camera rotation has 1 DoF, and the camera center has 2 DoF (up to translation along the line). We fix 4 DoF due to gauge freedom. Lines share 2 DoF for direction " + }, + { + "bbox": [ + 313, + 279, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 279, + 554, + 340 + ], + "type": "text", + "content": ", with each line having 2 additional DoF. The balanced constraint is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 381, + 346, + 553, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 346, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 381, + 346, + 553, + 358 + ], + "type": "interline_equation", + "content": "3 m + 2 + 2 n - 4 = m n. \\tag {15}", + "image_path": "f6defc9c42e7016b65eeb0296df1377c9ec84d0590be6e8723708872b6cfe424.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 364, + 392, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 364, + 392, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 364, + 392, + 376 + ], + "type": "text", + "content": "We get " + }, + { + "bbox": [ + 314, + 364, + 392, + 376 + ], + "type": "inline_equation", + "content": "n_D(m)" + }, + { + "bbox": [ + 314, + 364, + 392, + 376 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 392, + 380, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 380, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 392, + 380, + 553, + 403 + ], + "type": "interline_equation", + "content": "n _ {D} (m) = \\frac {3 m - 2}{m - 2}. \\tag {16}", + "image_path": "63da66a4b13c43e18b77e9c2ed5b3b012073c4f084c19e221bc5742cc0072364.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "text", + "content": ", we find " + }, + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "inline_equation", + "content": "\\lim_{m \\to \\infty} n_D(m) = 3" + }, + { + "bbox": [ + 313, + 408, + 554, + 432 + ], + "type": "text", + "content": ", meaning no balanced problem has fewer than 4 lines. We have found:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 434, + 523, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 434, + 523, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 434, + 523, + 445 + ], + "type": "interline_equation", + "content": "\\bullet m = 3, n = 7; \\quad m = 4, n = 5; \\quad m = 6, n = 4", + "image_path": "ec76ba006530819f82402ef44040582050b622962a155a37e6866a06f8b5c77a.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 459, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 506 + ], + "type": "text", + "content": "E. Gravity Prior for Cameras + Vertical Lines Similar to the previous case, each camera has 3 DoF, and we fix 4 DoF due to gauge freedom. Since the direction " + }, + { + "bbox": [ + 313, + 459, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 313, + 459, + 554, + 506 + ], + "type": "text", + "content": " is fixed, each line has 2 DoF. The balanced constraint is:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 389, + 519, + 553, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 519, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 389, + 519, + 553, + 531 + ], + "type": "interline_equation", + "content": "3 m + 2 n - 4 = m n. \\tag {17}", + "image_path": "e03bd3eef8eb16dda19f2b513a627217cd4f9d2544356609de6b129b4879d236.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 535, + 403, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 403, + 547 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 403, + 547 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 314, + 535, + 403, + 547 + ], + "type": "inline_equation", + "content": "n_E(m)" + }, + { + "bbox": [ + 314, + 535, + 403, + 547 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 392, + 551, + 553, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 551, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 392, + 551, + 553, + 575 + ], + "type": "interline_equation", + "content": "n _ {E} (m) = \\frac {3 m - 4}{m - 2}. \\tag {18}", + "image_path": "b16f6c6b5c1e6dcc402db9b0c28d0183dcc46e9693dc2141e95e2982463c2180.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "inline_equation", + "content": "m \\geq 3" + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": ", we find " + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\lim_{m \\to \\infty} n_E(m) = 3" + }, + { + "bbox": [ + 313, + 579, + 554, + 603 + ], + "type": "text", + "content": ", meaning no balanced problem has fewer than 4 lines. We found problems:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 605, + 453, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 605, + 453, + 616 + ], + "spans": [ + { + "bbox": [ + 314, + 605, + 453, + 616 + ], + "type": "interline_equation", + "content": "\\bullet m = 3, n = 5; \\quad m = 4, n = 4", + "image_path": "21111b0ae107e5cd70a2fd084f0214ef6bc0eaa957e177e01f2c3823ec3e58cd.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "All balanced problems for these settings are summarized in Table 1, along with their minimality status and degree. The table shows that problems with parallel/vertical lines (B, D, E) are minimal and computationally feasible. Problems with generic lines (A, C) are minimal but have degrees too high for practical RANSAC solutions. For problems " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "(A,5,23)" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "(C,5,15)" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", monodromy computations did not terminate within a month." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7146" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 153, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 153, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 153, + 83 + ], + "type": "text", + "content": "4. Minimal solvers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "In this section, we introduce solvers for the minimal problems " + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": " from Section 3. We use the notation and concepts from Sec. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 241, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 241, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 241, + 146 + ], + "type": "text", + "content": "4.1. Parallel lines with generic cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": "Here, we describe the solution to the problem " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": ". Since the lines are parallel, our goal is to find matrices " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3\\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": " and vectors " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{h,j}" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "j\\in \\{1,\\dots ,7\\}" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": ", such that equation (7) holds for every camera " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": " and line " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": ". Due to projective ambiguity, we aim to recover the camera poses up to a common projective transformation " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": ". Each camera triplet " + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1,\\mathbf{A}_2,\\mathbf{A}_3" + }, + { + "bbox": [ + 55, + 151, + 296, + 236 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 255 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {1} = \\mathbf {A} _ {c, 1} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {2} = \\mathbf {A} _ {c, 2} \\cdot \\mathbf {H}, \\quad \\mathbf {A} _ {3} = \\mathbf {A} _ {c, 3} \\cdot \\mathbf {H}, \\tag {19}", + "image_path": "7d86ef61a019788478d15f4f1b2a752fafd7e64bd735dd3e10937f3c3276bcdb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "text", + "content": "for some " + }, + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}" + }, + { + "bbox": [ + 55, + 262, + 295, + 274 + ], + "type": "text", + "content": " taking the form:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 292, + 305, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 305, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 305, + 316 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {c, 1} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {1} & \\alpha_ {1} & \\alpha_ {1} \\end{array} \\right], \\mathbf {A} _ {c, 2} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {2} & \\alpha_ {3} & \\alpha_ {4} \\end{array} \\right], \\mathbf {A} _ {c, 3} = \\left[ \\begin{array}{l l l} 1 & 0 & 0 \\\\ \\alpha_ {5} & \\alpha_ {6} & \\alpha_ {7} \\end{array} \\right]. \\tag {20}", + "image_path": "c30e7f40849dc335ef15daadf4239e9281cc92aac92bdd3f4c120dac69171abe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 318, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 294, + 341 + ], + "type": "text", + "content": "The proof follows analogously to Proposition 2 in [29]. These canonical-form matrices have 7 DoF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "text", + "content": "To estimate " + }, + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{c,1},\\mathbf{A}_{c,2},\\mathbf{A}_{c,3}" + }, + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "text", + "content": ", we leverage the fact that, as shown in [55], a trifocal tensor exists between three camera matrices " + }, + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i\\in \\mathbb{R}^{2\\times 3}" + }, + { + "bbox": [ + 55, + 342, + 296, + 389 + ], + "type": "text", + "content": ". To compute this tensor, we stack the constraints from (8) into the matrix equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 396, + 294, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 396, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 91, + 396, + 294, + 447 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c c} \\mathbf {A} _ {1} & \\mathbf {u} _ {1, j} ^ {\\prime} & 0 & 0 \\\\ \\mathbf {A} _ {2} & 0 & \\mathbf {u} _ {2, j} ^ {\\prime} & 0 \\\\ \\mathbf {A} _ {3} & 0 & 0 & \\mathbf {u} _ {3, j} ^ {\\prime} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {L} _ {h, j} \\\\ - \\lambda_ {1, j} \\\\ - \\lambda_ {2, j} \\\\ - \\lambda_ {3, j} \\end{array} \\right] = 0. \\tag {21}", + "image_path": "257c8b06739bc418771b7a26864fd170569a73925f213344abe9e6a827e4cd2e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 452, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 475 + ], + "type": "text", + "content": "For this equation to hold, the determinant of the coefficient matrix must vanish, leading to the constraint:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 482, + 294, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 482, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 74, + 482, + 294, + 515 + ], + "type": "interline_equation", + "content": "\\sum_ {a = 1} ^ {2} \\sum_ {b = 1} ^ {2} \\sum_ {c = 1} ^ {2} \\mathbf {u} _ {1, j (a)} ^ {\\prime} \\mathbf {u} _ {2, j (b)} ^ {\\prime} \\mathbf {u} _ {3, j (c)} ^ {\\prime} \\mathbf {T} _ {(a, b, c)} = 0, \\tag {22}", + "image_path": "779b825a85170abdf2c609e51554ec0669ed78be325282b5ab9f6c0320f0c5f5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 521, + 281, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 281, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 281, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 281, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}" + }, + { + "bbox": [ + 55, + 521, + 281, + 534 + ], + "type": "text", + "content": " is the trifocal tensor with elements:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 537, + 294, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 537, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 63, + 537, + 294, + 567 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {(a, b, c)} = (- 1) ^ {a + b + c} \\det \\left(\\left[ \\mathbf {A} _ {1 (a,:)} ^ {T} \\mathbf {A} _ {2 (b,:)} ^ {T} \\mathbf {A} _ {3 (c,:)} ^ {T} \\right]\\right). \\tag {23}", + "image_path": "891e01afc3f8b8d11ab35a99269e9d0cf77060e6977b2f9f2107e236eb944fe8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "content": "As shown in [55], " + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 568, + 295, + 592 + ], + "type": "text", + "content": " has no internal constraints and it can be estimated using 7 line correspondences." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": " is homogeneous, it has 7 DoF, matching the DoF of the canonical camera matrices. We verified that for every tensor " + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{2 \\times 2 \\times 2}" + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": ", there are exactly two corresponding triplets of matrices " + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}" + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": ". The proof and the decomposition algorithm for recovering these matrices can be found in the SM. Once " + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{c,1}, \\mathbf{A}_{c,2}, \\mathbf{A}_{c,3}" + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": " are found, we can triangulate the lines up to the projective transformation " + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 592, + 295, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "text", + "content": "The solver for the " + }, + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "text", + "content": " problem, which estimates the trifocal tensor " + }, + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 676, + 295, + 699 + ], + "type": "text", + "content": ", operates as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 701, + 259, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 259, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 259, + 714 + ], + "type": "text", + "content": "1. Input: Projections " + }, + { + "bbox": [ + 55, + 701, + 259, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 55, + 701, + 259, + 714 + ], + "type": "text", + "content": " of 7 lines in 3 cameras." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 72, + 553, + 118 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "text", + "content": "2. Compute matrices " + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i}" + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "text", + "content": " and vectors " + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}" + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 527, + 86 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 86, + 460, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 460, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 460, + 96 + ], + "type": "text", + "content": "3. Construct the linear system (22)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 96, + 553, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 118 + ], + "type": "text", + "content": "4. Solve (22) using singular value decomposition (SVD) to estimate the tensor " + }, + { + "bbox": [ + 313, + 96, + 553, + 118 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 96, + 553, + 118 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 121, + 471, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 471, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 471, + 133 + ], + "type": "text", + "content": "On average, the solver runs in " + }, + { + "bbox": [ + 313, + 121, + 471, + 133 + ], + "type": "inline_equation", + "content": "6.92\\mu s" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 140, + 488, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 488, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 488, + 152 + ], + "type": "text", + "content": "4.2. Vertical lines and known gravity" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "text", + "content": "Here, we analyze the solutions to the problems " + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "text", + "content": ". We assume that all lines are vertical and the gravity directions " + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i\\in \\mathbb{P}^2" + }, + { + "bbox": [ + 313, + 157, + 553, + 193 + ], + "type": "text", + "content": " for the cameras are known." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": "Similar to Sec. 4.1, we set " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{d,j} = \\mathbf{e}_2" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": ". Therefore, Equation (6) still holds. Furthermore the vanishing points " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "i \\in \\{1,\\dots,m\\}" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": " associated with the vertical lines are known, and there holds " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = \\mathbf{R}_i(y_i)\\mathbf{e}_2" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": ". We can decompose " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i) = \\mathbf{R}_{v,i}\\mathbf{R}_{y,i}" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{y,i}" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": " is a rotation around y-axis, and " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": " is a rotation around an axis in the xz-plane, such that " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}\\mathbf{e}_2 = \\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": " is calculated directly from " + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 194, + 554, + 288 + ], + "type": "text", + "content": ", we assume it to be known." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": "In this case, " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " (5) takes the form " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": ". While both " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " have their axes in the xz-plane, " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i}\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " does generically not share this property. However, we can decompose " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i}\\mathbf{R}_{v,i} = \\mathbf{R}_{a,i}\\mathbf{R}_{b,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": ", s.t. " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{a,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " has its axis in the xz-plane and " + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{b,i}" + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": " rotates around the y-axis." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": "Now, we decompose " + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime}" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{a,i} \\cdot \\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{a,i}" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": " is a known rotation around axis in the xz-plane, and " + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime \\prime} = \\mathbf{R}_{b,i}" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{y,i}" + }, + { + "bbox": [ + 313, + 348, + 553, + 384 + ], + "type": "text", + "content": " is an unknown rotation around y-axis." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "text", + "content": "We represent " + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{a,i},\\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "text", + "content": " with quaternions " + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i = [\\mathbf{q}_{i,w}\\mathbf{q}_{i,x}0\\mathbf{q}_{i,z}]^T" + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^{\\prime \\prime} = [\\mathbf{q}_{i,w}^{\\prime \\prime}0\\mathbf{q}_{i,y}^{\\prime \\prime}0]^T" + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "text", + "content": " , write the elements of " + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{R}'" + }, + { + "bbox": [ + 313, + 384, + 553, + 419 + ], + "type": "text", + "content": " in the form" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 423, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 423, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 321, + 423, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {R} _ {i (1, 1)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) - 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (1, 3)} ^ {\\prime} = 2 \\left(1 - 2 \\mathbf {q} _ {i, z} ^ {2}\\right) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} + 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right), \\\\ \\mathbf {R} _ {i (3, 1)} ^ {\\prime} = 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) - 2 (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\\\ \\mathbf {R} _ {i (3, 3)} ^ {\\prime} = \\left(1 - 2 \\mathbf {q} _ {i, x} ^ {2}\\right) \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) + 4 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime}, \\tag {24} \\\\ \\end{array}", + "image_path": "c81161fa420d8fa94c9ec5cd49a51528f0483eb15bb197117f5817a20ecea7d3.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "text", + "content": "and decompose the left " + }, + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "text", + "content": " submatrix of " + }, + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 509, + 527, + 521 + ], + "type": "text", + "content": " (7) as" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 358, + 526, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 526, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 358, + 526, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l l} - \\mathbf {R} _ {i, (1, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (1, 1)} ^ {\\prime} \\\\ - \\mathbf {R} _ {i, (3, 3)} ^ {\\prime} & \\mathbf {R} _ {i, (3, 1)} ^ {\\prime} \\end{array} \\right] = \\mathbf {A} _ {i, a} \\cdot \\mathbf {A} _ {i, b}, \\tag {25}", + "image_path": "c32fd26ccb446ca0116bf22ff3f9fcd6a33b98031e740e2534758e5687500160.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 563, + 335, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 335, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 335, + 574 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 354, + 578, + 553, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 578, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 354, + 578, + 553, + 606 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i, a} = \\left[ \\begin{array}{l l} - 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} & (1 - 2 \\mathbf {q} _ {i, z} ^ {2}) \\\\ - (1 - 2 \\mathbf {q} _ {i, x} ^ {2}) & 2 \\mathbf {q} _ {i, x} \\mathbf {q} _ {i, z} \\end{array} \\right], \\tag {26}", + "image_path": "61d750edd539710b517e0599e94775602b6bde589e5718986c2bef3175d3ab8b.jpg" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 359, + 613, + 553, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 613, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 359, + 613, + 553, + 641 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i, b} = \\left[ \\begin{array}{c c} \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) & 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ - 2 q _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & \\left(1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}\\right) \\end{array} \\right]. \\tag {27}", + "image_path": "a39526a55b1575df1adbdc0daf7ebbddad128c4384087242849475010a58f6c8.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,a}" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": " can be directly obtained from the known gravity direction and the current scanline " + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,b}" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": " is an unknown " + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": " rotation matrix. We define " + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^{\\prime} = \\mathbf{A}_{i,a}^{-1}\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime \\prime} = \\mathbf{A}_{i,a}^{\\prime T}\\mathbf{u}_{i,j}" + }, + { + "bbox": [ + 313, + 643, + 553, + 694 + ], + "type": "text", + "content": ". Then, there holds" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 399, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 399, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {i, j} ^ {\\prime \\prime T} \\mathbf {A} _ {i} ^ {\\prime} \\mathbf {L} _ {j, h} = 0. \\tag {28}", + "image_path": "ad08dc83e5802448d810f8113c7ca1047a1bde94096f9c4601acf17b9b3206be.jpg" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "7147" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": "Since the left " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " submatrix of " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i'" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,b}" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i'" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " has a form of a calibrated " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "2 \\times 3" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " camera matrix, and the problem with known gravity direction and vertical lines is equivalent to estimating a 2D structure from images taken by " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " calibrated cameras [8]. Unlike its uncalibrated counterpart, the calibrated problem can be solved without projective ambiguity. Solvers for the cases " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "m = 3, n = 5" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "m = 4, n = 4" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " have been presented in [8]. Notably, these numbers exactly correspond to the minimal problems discussed in Sec. 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "text", + "content": "The solver for " + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "inline_equation", + "content": "m = 3, n = 5" + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "text", + "content": " estimates a trifocal tensor of the form (23) from 5 point correspondences and 2 linear internal constraints, and solves a quadratic function to decompose the tensor into camera matrices. The solver for " + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "inline_equation", + "content": "m = 4, n = 4" + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "text", + "content": " estimates a dual quadrifocal tensor from 4 point correspondences and 11 linear internal constraints, and also solves a quadratic function to decompose the tensor into camera matrices. In both cases, there are two solutions in terms of matrices " + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i'" + }, + { + "bbox": [ + 55, + 180, + 296, + 300 + ], + "type": "text", + "content": ". Typically, the correct solution is selected by counting the points lying in front of all cameras." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "content": "After finding " + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "content": " with [8], we construct the camera poses " + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "content": ". For this, we first obtain " + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "content": ". Because (28) is homogeneous, there are two possible matrices " + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 300, + 296, + 335 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 342, + 296, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 342, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 79, + 342, + 296, + 422 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {R} _ {i} ^ {\\prime \\prime} \\in \\left\\{\\left[ \\begin{array}{c c c} (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right], \\right. \\tag {29} \\\\ \\left. \\left[ \\begin{array}{c c c} - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) & 0 & - 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} \\\\ 0 & 1 & 0 \\\\ 2 \\mathbf {q} _ {i, w} ^ {\\prime \\prime} \\mathbf {q} _ {i, y} ^ {\\prime \\prime} & 0 & - (1 - 2 \\mathbf {q} _ {i, y} ^ {\\prime \\prime 2}) \\end{array} \\right] \\right\\}. \\\\ \\end{array}", + "image_path": "614a6dd39fd1d7bd75284f8c29e79497ae65f26be6120af961ff3ce87dcdc7f8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": "The existence of two choices of " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": " per " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^\\prime" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": " accounts for the presence of 16 solutions in the 3 view case and 32 solutions in the 4 view case. Finally, we construct " + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 55, + 428, + 296, + 464 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 471, + 294, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 95, + 471, + 294, + 503 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {0, i} ^ {T} \\mathbf {R} _ {a, i} \\mathbf {R} _ {i} ^ {\\prime \\prime}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {i (1: 2, 3)}. \\tag {30} \\\\ \\end{array}", + "image_path": "b0be7e99fe556c583797056d9d7e0f5b772f51dea4c6cf1b2659c029ff5fe482.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "content": "Note, that " + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)_{(2)}" + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "content": " is always zero. This is in accordance with the Section 2, which explains that the camera centers can only be estimated up to the shift along the direction " + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 55, + 510, + 295, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": "To summarize, we perform the following steps to solve problems " + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "text", + "content": "1. Input: Projections " + }, + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "text", + "content": ", vertical directions " + }, + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 570, + 253, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "text", + "content": "2. Calculate matrices " + }, + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i},\\mathbf{R}_{v,i},\\mathbf{R}_{a,i}" + }, + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "text", + "content": " , and vectors " + }, + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}" + }, + { + "bbox": [ + 55, + 582, + 294, + 594 + ], + "type": "text", + "content": " (7)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "text", + "content": "3. Compose matrices " + }, + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,a}" + }, + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "text", + "content": " (26) and vectors " + }, + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime \\prime}(28)" + }, + { + "bbox": [ + 55, + 594, + 270, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "4. Use the 3 view or the 4 view solver from [8] to find 2 sets of matrices " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots,m\\}" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "text", + "content": "5. Construct " + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i^{\\prime \\prime}" + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "text", + "content": " (29), " + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i)" + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 55, + 629, + 249, + 642 + ], + "type": "text", + "content": " (30)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "text", + "content": "On average, the 3 view solver takes " + }, + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "inline_equation", + "content": "9.16\\mu s" + }, + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "text", + "content": ", and the 4 view solver takes " + }, + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "inline_equation", + "content": "70.84\\mu s" + }, + { + "bbox": [ + 55, + 642, + 294, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 672, + 230, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 230, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 230, + 685 + ], + "type": "text", + "content": "4.3. Parallel lines and known gravity" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We present the solution to the problem " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". In this case, all lines share a common direction " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", which is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "generically not aligned with the vertical direction " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_2" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_d \\in SO(3)" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " be a rotation around an axis in the xz-plane, s.t. " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_d = \\mathbf{R}_d\\mathbf{e}_2" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": ". Substituting this into (3), we get" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 349, + 114, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 114, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 349, + 114, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\left[ \\mathbf {R} _ {d} \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0, \\tag {31}", + "image_path": "04cf6f67fa8665e63a2da638f1afae8ba51ebf0f181cbdd6c1b75e7a0dad8b9e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 334, + 144, + 553, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 144, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 334, + 144, + 553, + 159 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\mathbf {R} _ {d} ^ {T} \\left(\\mathbf {L} _ {0, j} - \\mathbf {C} _ {i} \\left(y _ {i}\\right)\\right) = 0. \\tag {32}", + "image_path": "2f96e6dd3690ac45866feb2fd9ed6dd8b6ef655e21edd5c5705de64f14bf3bbe.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j}^{\\prime} = \\mathbf{R}_d^T\\mathbf{L}_{0,j}" + }, + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i^\\prime = \\mathbf{R}_d^T\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 313, + 162, + 553, + 186 + ], + "type": "text", + "content": " and substitute these into (32):" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 358, + 193, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 193, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 358, + 193, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {i, j} ^ {T} \\mathbf {R} _ {i} \\left(y _ {i}\\right) \\mathbf {R} _ {d} \\left[ \\mathbf {e} _ {2} \\right] _ {\\times} \\left(\\mathbf {L} _ {0, j} ^ {\\prime} - \\mathbf {C} _ {i} ^ {\\prime}\\right) = 0. \\tag {33}", + "image_path": "2b5459f87b953ab3395ee45e20d4cc1a22c4d8afb165d947094d987c2de51d76.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": "Following Sec. 4.2, we find rotation matrices " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": ", decompose " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i)" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": ", and find vectors " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}^{\\prime}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": " according to (8). Then, we define " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{D,i} = \\mathbf{R}_{0,i}\\mathbf{R}_{v,i}\\mathbf{R}_{y,i}\\mathbf{R}_d" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{D,i} = -\\mathbf{R}_{D,i}^{-1}[\\mathbf{e}_2]_{\\times}\\mathbf{C}_i^{\\prime}" + }, + { + "bbox": [ + 313, + 213, + 554, + 262 + ], + "type": "text", + "content": ", and rewrite the constraint as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 269, + 553, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 319 + ], + "type": "interline_equation", + "content": "\\lambda_ {i, j} \\mathbf {u} _ {i, j} ^ {\\prime} = \\left[ \\begin{array}{l l l} - \\mathbf {R} _ {D, i (1, 3)} & \\mathbf {R} _ {D, i (1, 1)} & \\mathbf {t} _ {D, i (1)} \\\\ - \\mathbf {R} _ {D, i (3, 3)} & \\mathbf {R} _ {D, i (3, 1)} & \\mathbf {t} _ {D, i (3)} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {L} _ {0, j (1)} ^ {\\prime} \\\\ \\mathbf {L} _ {0, j (3)} ^ {\\prime} \\\\ 1 \\end{array} \\right], \\tag {34}", + "image_path": "9d3c76933a68ea8bc132a13d4176d589d16669b7e19dce459bb4bc18f1a69100.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 319, + 473, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 473, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 473, + 331 + ], + "type": "text", + "content": "which can be written in matrix form as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 389, + 338, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 338, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 389, + 338, + 553, + 352 + ], + "type": "interline_equation", + "content": "\\lambda_ {i, j} \\mathbf {u} _ {i, j} = \\mathbf {A} _ {D, i} \\mathbf {L} _ {h, j} ^ {\\prime}. \\tag {35}", + "image_path": "a33d9ac54e5e2ccda4d2a2f5b5ad2dbb54dcbaa2e3797db30d4d45905e66a8ea.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 357, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 403 + ], + "type": "text", + "content": "This equation has the same form as (8), ensuring the existence of a trifocal tensor " + }, + { + "bbox": [ + 313, + 357, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{2,2,2}" + }, + { + "bbox": [ + 313, + 357, + 554, + 403 + ], + "type": "text", + "content": " satisfying (22). We estimate this tensor linearly using 7 correspondences, as in Sec. 4.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": "Let us now discuss how to decompose " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": " to estimate the relative pose. First, we fix the coordinate frame by setting " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_1(y_1) = I" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1(y_1) = 0" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": ". Both the tensor " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": " and the pose have 7 DoF. Using Gröbner basis, we have discovered that for every tensor " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{2,2,2}" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": ", there are 48 camera poses " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{y,2}, \\mathbf{t}_{D,2}, \\mathbf{R}_{y,3}, \\mathbf{t}_{D,3}, \\mathbf{R}_d" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": ", such that the matrices " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{D,i}" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": " constructed via (34) have the tensor " + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 405, + 554, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 489, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 554, + 548 + ], + "type": "text", + "content": "Due to the high number of variables and equation complexity, we could not generate an elimination template for a Gröbner basis solver. Therefore, we solve this problem with homotopy continuation [18], with an average runtime " + }, + { + "bbox": [ + 313, + 489, + 554, + 548 + ], + "type": "inline_equation", + "content": "19089\\mu s" + }, + { + "bbox": [ + 313, + 489, + 554, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "text", + "content": "Finally, we compose " + }, + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{i,D}" + }, + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "text", + "content": " according to (34), and the pose " + }, + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i),\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 313, + 548, + 553, + 573 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 331, + 578, + 553, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 331, + 578, + 553, + 610 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {R} _ {i} (y _ {i}) = \\mathbf {R} _ {v, i} \\mathbf {R} _ {y, i}, \\\\ \\mathbf {C} _ {i} \\left(y _ {i}\\right) = R _ {d} \\left[ \\begin{array}{l l} \\mathbf {e} _ {1} & \\mathbf {e} _ {3} \\end{array} \\right] \\mathbf {A} _ {D, i (1: 2, 1: 2)} ^ {- 1} \\mathbf {A} _ {D, i (1: 2, 3)} \\tag {36} \\\\ \\end{array}", + "image_path": "3c6888f35823b2909f28d5958e857d8d6f19614dbdc89490bdea8b992eec0af9.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "text", + "content": "To summarize, we perform the following steps to solve the problem " + }, + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 313, + 615, + 554, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 639, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "text", + "content": "1. Input: Projections " + }, + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "text", + "content": ", vertical directions " + }, + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 314, + 639, + 513, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "text", + "content": "2. Calculate matrices " + }, + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{0,i},\\mathbf{R}_{v,i}" + }, + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "text", + "content": " and vectors " + }, + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{i,j}" + }, + { + "bbox": [ + 314, + 651, + 529, + 663 + ], + "type": "text", + "content": " (7)." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 663, + 529, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 663, + 529, + 675 + ], + "spans": [ + { + "bbox": [ + 314, + 663, + 529, + 675 + ], + "type": "text", + "content": "3. Construct linear system (22) and solve it to get " + }, + { + "bbox": [ + 314, + 663, + 529, + 675 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 314, + 663, + 529, + 675 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "spans": [ + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": "4. Use homotopy continuation to decompose " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{y,2}" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{D,2}" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{y,3}" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{3,D}" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_d" + }, + { + "bbox": [ + 314, + 675, + 553, + 699 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "text", + "content": "5. Construct " + }, + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i(y_i)" + }, + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_i(y_i)" + }, + { + "bbox": [ + 314, + 700, + 491, + 712 + ], + "type": "text", + "content": " using (36)." + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7148" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 166 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 166 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 166 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 166 + ], + "type": "image", + "image_path": "564df0bbbfe616916c75a04a33e36c49e60c771b06e0ddae69bb51c83b93c721.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "type": "text", + "content": "Figure 3. Histogram of tensor, rotation, and translation errors of the proposed solvers, over " + }, + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 55, + 167, + 296, + 190 + ], + "type": "text", + "content": " noiseless samples. In radians." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 210, + 178, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 178, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 178, + 223 + ], + "type": "text", + "content": "4.4. Degenerate scenarios" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "text", + "content": "Here, we analyze the degenerate cases for the solvers introduced in Sections 4.1, 4.2, and 4.3. Specifically, we consider scenarios where some lines are coplanar or where camera centers are collinear or coincident." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 275, + 296, + 335 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "text", + "content": "- Solvers " + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "text", + "content": " are degenerate if all " + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 275, + 295, + 299 + ], + "type": "text", + "content": " lines are coplanar or if any two camera centers coincide." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "text", + "content": "- Solvers " + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 299, + 296, + 323 + ], + "type": "text", + "content": " are degenerate if at least 5 lines are coplanar or if any two camera centers coincide." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 324, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 295, + 335 + ], + "type": "text", + "content": "- There is no degeneracy if the camera centers are colinear." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 347, + 136, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 136, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 136, + 361 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": "In this section, we evaluate the solvers " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 367, + 295, + 392 + ], + "type": "text", + "content": " on both synthetic and real data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 399, + 182, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 182, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 182, + 413 + ], + "type": "text", + "content": "5.1. Synthetic experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": "Numerical stability. We sample scanlines " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{[-1,1]}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " and generate random camera rotations " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i,GT}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " and centers " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{i,GT}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ". Since we use one scanline per image, we only sample one pose per image. We get the vertical directions as " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "v_{i} = \\mathbf{R}_{i,GT}\\mathbf{e}_{2}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ". To generate line " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ", we sample scalars " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mu_{j,x},\\mu_{j,z}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " from normalized Gaussian distribution and set " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{0,j} = [\\mu_{j,x}0\\mu_{j,z} + 5]^T" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ". For solver " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ", we also generate scalars " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\nu_{x},\\nu_{z}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ", and set " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "L_{j,d} = [\\nu_x1\\nu_z]^T" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ". In other cases, we use " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "L_{j,d} = \\mathbf{e}_2" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": ". We generate " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " into the camera with pose " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{i,GT},\\mathbf{C}_{i,GT})" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": " and intersecting the projected line with scanline " + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 417, + 296, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": "Before measuring the error, we transform both the ground truth pose, and the estimated pose " + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{i,\\mathrm{est}},\\mathbf{C}_{i,\\mathrm{est}})" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": " into a canonical coordinate frame with " + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1 = 0" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{R_{y,1}} = \\mathbf{I},\\mathbf{L_d} = \\mathbf{e_1}" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": ". Since the y-coordinates of all camera centers are unobservable, we set them to zero. Then, we calculate the rotation error " + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "err_{\\mathbf{R}} = \\max_{i = 1}^{m}(\\angle \\mathbf{R}_{i,\\mathrm{est}}^{T}\\mathbf{R}_{i,GT})" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": " and the translation error " + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "inline_equation", + "content": "err_{\\mathbf{C}} = \\max_{i = 1}^{m}\\angle (\\mathbf{C}_{i,GT},\\mathbf{C}_{i,\\mathrm{est}})" + }, + { + "bbox": [ + 55, + 549, + 295, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": "Additionally, we construct tensors " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{GT}" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": " from GT poses and estimate " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{est}" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": " using the solvers. For trifocal problems, " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{GT},\\mathbf{T}_{est}" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": " are trifocal tensors (23), while for " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": ", they are dual quadrifocal tensors. We measure tensor error errT as the Frobenius norm of " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{GT} - \\mathbf{T}_{est}" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": ". For each solver, we generate " + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 55, + 634, + 296, + 717 + ], + "type": "text", + "content": " synthetic noise-free scenes and measure errR, errC, and errT. Results in Fig. 3 show solver stability." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 270 + ], + "type": "image", + "image_path": "d808fa8edf1daa28750b34ebcb215b0c48d1687308749456a6175f163299a052.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 270, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 348 + ], + "type": "text", + "content": "Figure 4. Rotation, translation, and tensor errors averaged over 10000 successful runs as a function of the point and the vertical noise, measured from the real solution closest to the ground truth. If a solver returns only complex solutions, we report a failure. The failure rates are reported in the last row. Plots in the same row share the y-axis and plots in the same column share the x-axis. The fixed parameters are in the titles." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": "Tests with noise. To evaluate the robustness to noise, we set focal length " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "f = 1000px" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ", and add Gaussian noise with standard deviation " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\frac{\\sigma_p}{f}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " to each projection " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ". We also perturb vertical directions " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " by applying a random 3D rotation with angle " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_v" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ". Fig. 4 shows " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "err_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "err_{\\mathbf{C}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "err_{\\mathbf{T}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " as the function of " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_p" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\sigma_v" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ", averaged over " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "10^4" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " runs. The results confirm that, since the line detections tend to be relatively precise, our solvers exhibit reasonable behaviour under realistic noise levels." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 487, + 452, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 452, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 452, + 499 + ], + "type": "text", + "content": "5.2. Real-world Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "text", + "content": "We evaluate our solvers using the Fastec dataset [43], which consists of 19 sequences captured from a moving car in Tübingen. Each sequence contains 34 rolling shutter (RS) images and two global shutter (GS) images for every RS image: one with the pose of the first scanline and one with the pose of the middle scanline. The images have a resolution of " + }, + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "text", + "content": " px and a focal length " + }, + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "inline_equation", + "content": "768" + }, + { + "bbox": [ + 313, + 504, + 554, + 589 + ], + "type": "text", + "content": " px." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": "Since GT poses are unavailable, we use COLMAP [61] on the GS images to obtain pseudo-GT poses " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{i,F},\\mathbf{C}_{i,F})" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " for the first scanline of image " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{i,M},\\mathbf{C}_{i,M})" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " for the middle scanline. The pseudo-GT pose for scanline " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " is interpolated as: " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i,GT}(y_i) = \\mathbf{R}_{i,M\\to y_i}\\mathbf{R}_{i,M},\\mathbf{C}_{i,GT}(y_i) =" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{i,M} + \\frac{h - 2y_i}{h} (\\mathbf{C}_{i,F} - \\mathbf{C}_{i,M})" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i,M\\to y_i}" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " is a rotation matrix with the axis of " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " , and the angle " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\frac{h - 2y_i}{h}" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " times the angle of " + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{i,F}\\mathbf{R}_{i,M}^{T}" + }, + { + "bbox": [ + 313, + 589, + 554, + 700 + ], + "type": "text", + "content": " . We evaluate the solvers in multi-view and single-view settings." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Multi-view. We estimate poses for " + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": " consecutive im" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7149" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 282, + 157 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 282, + 157 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 282, + 157 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 282, + 157 + ], + "type": "table", + "html": "
solver≤ 5°≤ 10°≤ 20°≤ 30°
MV(E,3,5)131015
(E,4,4)581114
(D,3,7)6101517
SV(E,3,5)33510
(E,4,4)24713
(D,3,7)4479
", + "image_path": "6d91acf15aa45a920189060ebc2d175ea2b850f070510cf43cbe62e0a24f8ab4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": "ages, selecting one scanline per image. We use DeepLSD [49] and Gluestick [50] to detect and match the lines. Although the projections of lines by RS cameras are generically curves, the line-based methods work well in the case of Fastec dataset. To select scanlines " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "y_{1}, \\ldots, y_{m}" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": ", we exhaustively search for the scanlines that maximize the number of matched lines intersecting them. Then, we find " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": " as the intersections of the detected lines and the scanlines. Following [12], we approximate the vertical direction by " + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "inline_equation", + "content": "[0 1 0]^T" + }, + { + "bbox": [ + 54, + 213, + 295, + 344 + ], + "type": "text", + "content": ". This setting works well for many datasets, including Fastec, and does not require gravity direction measurement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "text", + "content": "Single-view. In this case, select " + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "text", + "content": " scanlines in the same image and estimate the pose between them. We detect lines with DeepLSD [49], and select scanlines " + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "inline_equation", + "content": "y_{1},\\ldots ,y_{m}" + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "text", + "content": " with spacing at least 25px with the highest number of shared lines. We obtain " + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "inline_equation", + "content": "p_{i,j}" + }, + { + "bbox": [ + 55, + 345, + 296, + 405 + ], + "type": "text", + "content": " by intersecting lines with scanlines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": "We evaluate the solvers in RANSAC [22]. We label a correspondence as inlier, if the triangulated line " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_j" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": " lies in front of all cameras and its reprojection error " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": " is below " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "1px" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": ". In the case of " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": ", we decompose the tensor into canonical cameras (20) before triangulation. Details about triangulation are in the SM. The score for a model is the sum over all lines of " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "(1px - \\varepsilon)^2" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\varepsilon < 1px" + }, + { + "bbox": [ + 55, + 405, + 296, + 513 + ], + "type": "text", + "content": ", and zero otherwise. We run RANSAC for 1000 iterations, and select the model with the highest score. We do not use local optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": "The results are summarized in Tables 2, 3 summarize the results for the solvers " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "(E,3,5)" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "(E,4,4)" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "(D,3,7)" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": ", and Table 4 shows the tensor error for the solver " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": ". Although the total number of camera poses with an error below " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 55, + 514, + 296, + 633 + ], + "type": "text", + "content": " remains relatively low, in Structure-from-Motion (SfM) applications, it is sufficient to estimate a single correct relative pose per sequence to initialize the reconstruction. Table 2 confirms that this is achievable for the vast majority of sequences. A more detailed evaluation and further experiments are provided in the SM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 645, + 222, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 222, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 222, + 658 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "While our approach shows promising results for rolling shutter relative pose estimation, several limitations remain. A key challenge is the lack of robust curve detectors and matchers, as rolling shutter distortion bends straight lines." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 339, + 70, + 529, + 157 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "type": "text", + "content": "Table 2. Number of sequences (out of 19) from the Fastec dataset [43], with at least one estimated pose with error below " + }, + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "type": "inline_equation", + "content": "5/10/20/30^{\\circ}" + }, + { + "bbox": [ + 55, + 158, + 295, + 191 + ], + "type": "text", + "content": ". MV: Multi-view, SV: Single-view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 339, + 70, + 529, + 157 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 529, + 157 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 529, + 157 + ], + "type": "table", + "html": "
solver≤ 10°≤ 20°≤ 30°
MV(E,3,5)1.0 %6.2 %21.9 %
(E,4,4)1.9 %7.9 %22.4 %
(D,3,7)5.6 %13.8 %21.7 %
SV(E,3,5)0.5 %2.5 %5.3 %
(E,4,4)1.2 %4.1 %10.0 %
(D,3,7)0.5 %2.2 %6.6 %
", + "image_path": "cbf416a8dfc6489a5b2e2f1685fb30b5ac7c36e0a48fb9ba4484e4d8a96b36d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 342, + 191, + 525, + 230 + ], + "blocks": [ + { + "bbox": [ + 313, + 157, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 180 + ], + "type": "text", + "content": "Table 3. Percentage of camera poses from Fastec [43] with pose error below " + }, + { + "bbox": [ + 313, + 157, + 553, + 180 + ], + "type": "inline_equation", + "content": "10/20/30^{\\circ}" + }, + { + "bbox": [ + 313, + 157, + 553, + 180 + ], + "type": "text", + "content": ". MV: Multi-view, SV: Single-view." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 342, + 191, + 525, + 230 + ], + "lines": [ + { + "bbox": [ + 342, + 191, + 525, + 230 + ], + "spans": [ + { + "bbox": [ + 342, + 191, + 525, + 230 + ], + "type": "table", + "html": "
settingMedian errorMinimum error
MV0.850.44
SV0.890.62
", + "image_path": "198797a15b67b996ca061593b94779fe2faed69db70f1daa5f1b1b4e24fbc67c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 231, + 554, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 554, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 554, + 264 + ], + "type": "text", + "content": "Table 4. Median and Minimum tensor error of the " + }, + { + "bbox": [ + 313, + 231, + 554, + 264 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 313, + 231, + 554, + 264 + ], + "type": "text", + "content": " solver, averaged over all sequences from Fastec [43]. A more detailed evaluation is in the SM." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 285, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 453 + ], + "type": "text", + "content": "Additionally, our method relies on chirality constraints to resolve ambiguity, as noted in [8]. The method does not use any motion model, which makes it broadly applicable and model-agnostic. However, this universality comes at the cost of stability: we expect methods that incorporate motion models to be more stable. Our solvers are also limited to specific scene types and assume a known gravity direction. However, this reliance could be mitigated using scene knowledge, such as geometric relations between lines, in combination with solver " + }, + { + "bbox": [ + 313, + 285, + 555, + 453 + ], + "type": "inline_equation", + "content": "(B,3,7)" + }, + { + "bbox": [ + 313, + 285, + 555, + 453 + ], + "type": "text", + "content": ". Finally, our method estimates pose only up to a shift along the line direction. Future work could integrate additional geometric constraints, improve feature detection, and integrate the solvers to a rolling-shutter SfM pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": "In this paper, we propose a new approach for estimating the relative pose of rolling shutter cameras using projections of lines onto multiple scanlines. This enables estimating relative pose of rolling shutter cameras without explicitly modelling camera motion. The proposed solvers provide an important building block for rolling-shutter SfM where each scanline has its pose computed independently, without relying on a global motion model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 713 + ], + "type": "text", + "content": "We identify minimal problems for different variations of the single-scanline relative pose problem, including cases with parallel lines, vertical lines, and known gravity direction. Furthermore, we develop minimal solvers for the parallel-lines scenario, both with and without gravity priors. To evaluate our approach, we conduct experiments on both synthetic data and real-world rolling shutter images from the Fastec dataset, considering scanlines from either the same image or different images. These experiments demonstrate that our solvers successfully estimate at least one correct pose in the majority of scenes, confirming their feasibility for SfM initialization." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7150" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 296, + 167 + ], + "type": "text", + "content": "[1] Omar Ait-Aider and François Berry. Structure and kinematics triangulation with a rolling shutter stereo rig. In IEEE 12th International Conference on Computer Vision, ICCV 2009, Kyoto, Japan, September 27 - October 4, 2009, 2009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 235 + ], + "type": "text", + "content": "[2] Omar Ait-Aider, Nicolas Andreff, Jean-Marc Lavest, and Philippe Martinet. Simultaneous object pose and velocity computation using a single view from a rolling shutter camera. In Computer Vision - ECCV 2006, 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part II, 2006. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "type": "text", + "content": "[3] Omar Ait-Aider, Adrien Bartoli, and Nicolas Andreff. Kinematics from lines in a single rolling shutter image. In 2007 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2007), 18-23 June 2007, Minneapolis, Minnesota, USA, 2007. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 292, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 334 + ], + "type": "text", + "content": "[4] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. R6P - rolling shutter absolute pose problem. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "type": "text", + "content": "[5] Cenek Albl, Zuzana Kukelova, and Tomás Pajdla. Rolling shutter absolute pose problem with known vertical direction. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 423 + ], + "type": "text", + "content": "[6] Cenek Albl, Zuzana Kukelova, Viktor Larsson, and Tomás Pajdla. Rolling shutter camera absolute pose. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 426, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 479 + ], + "type": "text", + "content": "[7] Cenek Albl, Zuzana Kukelova, Viktor Larsson, Michal Polic, Tomás Pajdla, and Konrad Schindler. From two rolling shutters to one global shutter. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "text", + "content": "[8] Kalle Åström and Magnus Oskarsson. Solutions and ambiguities of the structure and motion problem for 1d retinal vision. J. Math. Imaging Vis., 2000, 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 567 + ], + "type": "text", + "content": "[9] Fang Bai, Agniva Sengupta, and Adrien Bartoli. Scanline homographies for rolling-shutter plane absolute pose. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 613 + ], + "type": "text", + "content": "[10] Akash Bapat, Enrique Dunn, and Jan-Michael Frahm. Towards kilo-hertz 6-dof visual tracking using an egocentric cluster of rolling shutter cameras. IEEE Trans. Vis. Comput. Graph., 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "type": "text", + "content": "[11] Akash Bapat, True Price, and Jan-Michael Frahm. Rolling shutter and radial distortion are features for high frame rate multi-camera tracking. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[12] Daniel Barath, Dmytro Mishkin, Luca Cavalli, Paul-Edouard Sarlin, Petr Hruby, and Marc Pollefeys. StereoGlue: Robust estimation with single-point solvers. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Paul Breiding and Sascha Timme. Homotopycontinuation.jl: A package for homotopy continuation in julia. In Mathematical Software - ICMS 2018 - 6th International Conference, South Bend, IN, USA, July 24-27, 2018, Proceedings. Springer, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "text", + "content": "[14] Mingdeng Cao, Zhihang Zhong, Jiahao Wang, Yinqiang Zheng, and Yujiu Yang. Learning adaptive warping for real-world rolling shutter correction. CoRR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[15] Yuchao Dai, Hongdong Li, and Laurent Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[16] Juan Carlos Dibene, Yazmín Maldonado, Leonardo Trujillo, and Enrique Dunn. Prepare for ludicrous speed: Marker-based instantaneous binocular rolling shutter localization. IEEE Trans. Vis. Comput. Graph., 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "text", + "content": "[17] Timothy Duff, Kathlen Kohn, Anton Leykin, and Tomas Pajdla. PLMP - point-line minimal problems in complete multiview visibility. IEEE Trans. Pattern Anal. Mach. Intell., 2024, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 309, + 553, + 385 + ], + "type": "text", + "content": "[18] Ricardo Fabbri, Timothy Duff, Hongyi Fan, Margaret H. Regan, David da Costa de Pinho, Elias P. Tsigaridas, Charles W. Wampler, Jonathan D. Hauenstein, Peter J. Giblin, Benjamin B. Kimia, Anton Leykin, and Tomás Pajdla. TRPLP - trifocal relative pose from lines at points. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[19] Bin Fan, Yuchao Dai, and Ke Wang. Rolling-shutter-stereo-aware motion estimation and image correction. Comput. Vis. Image Underst., 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[20] Bin Fan, Yuchao Dai, Zhiyuan Zhang, and Ke Wang. Differential sfm and image correction for a rolling shutter stereog. Image Vis. Comput., 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[21] Bin Fan, Yuchao Dai, and Hongdong Li. Rolling shutter inversion: Bring rolling shutter images to high framerate global shutter video. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "text", + "content": "[22] Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 1981. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "text", + "content": "[23] Per-Erik Forssén and Erik Ringaby. Rectifying rolling shutter video from hand-held devices. In The Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, San Francisco, CA, USA, 13-18 June 2010, 2010. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[24] Marvin Anas Hahn, Kathlen Kohn, Orlando Marigliano, and Tomás Pajdla. Order-one rolling shutter cameras. CoRR, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "type": "text", + "content": "[25] Michiel Hazewinkel. Encyclopaedia of mathematics. 2002. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Johan Hedborg, Erik Ringaby, Per-Erik Forssén, and Michael Felsberg. Structure and motion estimation from rolling shutter video. In IEEE International Conference on Computer Vision Workshops, ICCV 2011 Workshops, Barcelona, Spain, November 6-13, 2011, 2011. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7151" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Johan Hedborg, Per-Erik Forssén, Michael Felsberg, and Erik Ringaby. Rolling shutter bundle adjustment. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, June 16-21, 2012, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[28] Radu Horaud, Roger Mohr, and Boguslaw Lorecki. On single-scanline camera calibration. IEEE Trans. Robotics Autom., 1993. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 215 + ], + "type": "text", + "content": "[29] Petr Hruby, Viktor Korotynskiy, Timothy Duff, Luke Oeding, Marc Pollefeys, Tomás Pajdla, and Viktor Larsson. Four-view geometry with unknown radial distortion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[30] Chao Jia and Brian L. Evans. Probabilistic 3-d motion estimation for rolling shutter video rectification from visual and inertial measurements. In 14th IEEE International Workshop on Multimedia Signal Processing, MMSP 2012, Banff, AB, Canada, September 17-19, 2012, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[31] Georg Klein and David William Murray. Parallel tracking and mapping on a camera phone. In Science & Technology Proceedings, 8th IEEE International Symposium on Mixed and Augmented Reality 2009, ISMAR 2009, Orlando, Florida, USA, October 19-22, 2009, 2009. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[32] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, and Tomás Pajdla. Linear solution to the minimal absolute pose rolling shutter problem. In Computer Vision - ACCV 2018 - 14th Asian Conference on Computer Vision, Perth, Australia, December 2-6, 2018, Revised Selected Papers, Part III, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 458 + ], + "type": "text", + "content": "[33] Zuzana Kukelova, Cenek Albl, Akihiro Sugimoto, Konrad Schindler, and Tomás Pajdla. Minimal rolling shutter absolute pose with unknown focal length and radial distortion. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[34] Yizhen Lao and Omar Ait-Aider. Rolling shutter homography and its applications. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(8):2780-2793, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 536 + ], + "type": "text", + "content": "[35] Yizhen Lao, Omar Ait-Aider, and Helder Araujo. Robustified structure from motion with rolling-shutter camera using straightness constraint. Pattern Recognit. Lett., 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[36] Yizhen Lao, Omar Ait-Aider, and Adrien Bartoli. Rolling shutter pose and ego-motion estimation using shape-from-template. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part II, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "type": "text", + "content": "[37] Chang-Ryeol Lee and Kuk-Jin Yoon. Monocular visual odometry with a rolling shutter camera. CoRR, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 635 + ], + "type": "text", + "content": "[38] Chang-Ryeol Lee and Kuk-Jin Yoon. Inertial-aided rolling shutter relative pose estimation. CoRR, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[39] Chang-Ryeol Lee, Ju Hong Yoon, Min-Gyu Park, and Kuk-Jin Yoon. Gyroscope-aided relative pose estimation for rolling shutter cameras. CoRR, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Dongdong Li, Gongjian Wen, Bing Wei Hui, Shaohua Qiu, and Wenfei Wang. Cross-ratio invariant based line scan camera geometric calibration with static linear data. Optics and Lasers in Engineering, 2014. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 128 + ], + "type": "text", + "content": "[41] Moyang Li, Peng Wang, Lingzhe Zhao, Bangyan Liao, and Peidong Liu. Usb-nerf: Unrolling shutter bundle adjusted neural radiance fields. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "text", + "content": "[42] Bangyan Liao, Delin Qu, Yifei Xue, Huiqing Zhang, and Yizhen Lao. Revisiting rolling shutter bundle adjustment: Toward accurate and fast solution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[43] Peidong Liu, Zhaopeng Cui, Viktor Larsson, and Marc Pollefeys. Deep shutter unrolling network. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020, 2, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 233, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 233, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 233, + 553, + 299 + ], + "type": "text", + "content": "[44] Ludovic Magerand, Adrien Bartoli, Omar Ait-Aider, and Daniel Pizarro. Global optimization of object pose and motion from a single rolling shutter image with automatic 2d-3d matching. In Computer Vision - ECCV 2012 - 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part I, 2012. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "text", + "content": "[45] Marci Meingast, Christopher Geyer, and Shankar Sastry. Geometric models of rolling-shutter cameras. CoRR, abs/cs/0503076, 2005. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "type": "text", + "content": "[46] Branislav Micusik and Georgios Evangelidis. Renormalization for initialization of rolling shutter visual-inertial odometry. Int. J. Comput. Vis., 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "text", + "content": "[47] Thanh-Tin Nguyen and Maxime Lhuillier. Adding synchronization and rolling shutter in multi-camera bundle adjustment. In Proceedings of the British Machine Vision Conference 2016, BMVC 2016, York, UK, September 19-22, 2016, 2016. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "text", + "content": "[48] David Nister. An efficient solution to the five-point relative pose problem. IEEE Trans. Pattern Anal. Mach. Intell., 2004. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 507 + ], + "type": "text", + "content": "[49] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R. Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Computer Vision and Pattern Recognition (CVPR), 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "type": "text", + "content": "[50] Rémi* Pautrat, Iago* Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. GlueStick: Robust image matching by sticking points and lines together. In International Conference on Computer Vision (ICCV), 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 555, + 609 + ], + "type": "text", + "content": "[51] Pulak Purkait and Christopher Zach. Minimal solvers for monocular rolling shutter compensation under ackermann motion. In 2018 IEEE Winter Conference on Applications of Computer Vision, WACV 2018, Lake Tahoe, NV, USA, March 12-15, 2018, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[52] Pulak Purkait, Christopher Zach, and Ales Leonardis. Rolling shutter correction in Manhattan world. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[53] Delin Qu, Yizhen Lao, Zhigang Wang, Dong Wang, Bin Zhao, and Xuelong Li. Towards nonlinear-motion-aware and occlusion-robust rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7152" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[54] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Trans. Pattern Anal. Mach. Intell., 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[55] Long Quan and Takeo Kanade. Affine structure from line correspondences with uncalibrated affine cameras. IEEE Trans. Pattern Anal. Mach. Intell., 1997. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "text", + "content": "[56] Vijay Rengarajan, A. N. Rajagopalan, and Rangarajan Aravind. From bows to arrows: Rolling shutter rectification of urban scenes. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "text", + "content": "[57] Erik Ringaby and Per-Erik Forssén. Efficient video rectification and stabilisation for cell-phones. Int. J. Comput. Vis., 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[58] Olivier Saurer, Kevin Koser, Jean-Yves Bouguet, and Marc Pollefeys. Rolling shutter stereo. In IEEE International Conference on Computer Vision, ICCV 2013, Sydney, Australia, December 1-8, 2013, 2013. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[59] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. A minimal solution to the rolling shutter pose estimation problem. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2015, Hamburg, Germany, September 28 - October 2, 2015, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "text", + "content": "[60] Olivier Saurer, Marc Pollefeys, and Gim Hee Lee. Sparse to dense 3d reconstruction from rolling shutter images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[61] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[62] David Schubert, Nikolaus Demmel, Vladyslav Usenko, Jörg Stuckler, and Daniel Cremers. Direct sparse odometry with rolling shutter. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VIII, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "text", + "content": "[63] Wei Shang, Dongwei Ren, Chaoyu Feng, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised learning to bring dual reversed rolling shutter images alive. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 555 + ], + "type": "text", + "content": "[64] Ken Shoemake. Animating rotation with quaternion curves. In SIGGRAPH, 1985. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[65] Henrik Stewénius, David Nister, Magnus Oskarsson, and Kalle Åström. Solutions to minimal generalized relative pose problems. Workshop on omnidirectional vision, 2005. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 623 + ], + "type": "text", + "content": "[66] Daobilige Su, Asher Bender, and Salah Sukkarieh. Improved cross-ratio invariant-based intrinsic calibration of A hyperspectral line-scan camera. Sensors, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[67] Yufen Sun, Gang Liu, and Yue Sun. An affine motion model for removing rolling shutter distortions. IEEE Signal Process. Lett., 2016. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[68] Subeesh Vasu, Mahesh Mohan M. R., and A. N. Rajagopalan. Occlusion-aware rolling shutter rectification of 3d scenes. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, 2018. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 497 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[69] Carlos Andres Luna Vázquez, Manuel Mazo, José Luis Lázaro, and Juan F. Vázquez. Calibration of line-scan cameras. IEEE Trans. Instrum. Meas., 2010. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[70] Ke Wang, Bin Fan, and Yuchao Dai. Relative pose estimation for stereo rolling shutter cameras. In IEEE International Conference on Image Processing, ICIP 2020, Abu Dhabi, United Arab Emirates, October 25-28, 2020, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[71] Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. Urs-nerf: Unordered rolling shutter bundle adjustment for neural radiance fields. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXXV, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[72] Weilong Yan, Robby T. Tan, Bing Zeng, and Shuaicheng Liu. Deep homography mixture for single image rolling shutter correction. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "text", + "content": "[73] Ming Yao, Zuyun Zhao, and Bugao Xu. Geometric calibration of line-scan camera using a planar pattern. J. Electronic Imaging, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[74] Yongcong Zhang, Bangyan Liao, Yifei Xue, Chen Lu, Peidong Liu, and Yizhen Lao. RSL-BA: rolling shutter line bundle adjustment. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LIX, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[75] Bingbing Zhuang and Quoc-Huy Tran. Image stitching and rectification for hand-held cameras. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[76] Bingbing Zhuang, Loong-Fah Cheong, and Gim Hee Lee. Rolling-shutter-aware differential sfm and image rectification. In IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 497 + ], + "type": "text", + "content": "[77] Bingbing Zhuang, Quoc-Huy Tran, Pan Ji, Loong-Fah Cheong, and Manmohan Chandraker. Learning structure-and-motion-aware rolling shutter correction. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, 2019. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7153" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_content_list.json b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a20c822ba2193df1f7bf51ed1f600bd3d33b4e66 --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Uzay Gokay $^{1}$ Federico Spurio $^{1,3}$ Dominik R. Bach $^{1,2}$ Juergen Gall $^{1,3}$ $^{1}$ University of Bonn, Germany $^{2}$ University College London, UK \n $^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence, Germany {u.goekay, d.bach}@uni-bonn.de {fspurio, gall}@iai.uni-bonn.de", + "bbox": [ + 196, + 202, + 800, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods. Code is available at github.com/bachlab/SMQ.", + "bbox": [ + 88, + 340, + 483, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 691, + 220, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in motion capture technology and pose estimation have enabled the precise recording of human motion, represented as time series of joint positions, orientations, or other representations of skeleton sequences. Recently, there has been an increased interest in temporally segmenting long skeleton sequences into action segments [14, 19, 21, 32, 53-55, 60]. These approaches, however, are predominantly supervised, requiring extensive labeled data for training. In contrast, prior unsupervised skeleton-based approaches [16, 30, 34, 35, 45, 52, 67] focus mainly on action recognition, assuming that the sequences are very short and contain only a single action per sequence.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ac97c4981373e8d051e7dd1a45bfd017e3e8c6df02acf97b09241af91c08df4.jpg", + "image_caption": [ + "Figure 1. Our unsupervised skeleton-based temporal action segmentation framework, Skeleton Motion Quantization (SMQ), takes a set of untrimmed skeleton sequences as input (top) and discovers actions consistent across all sequences (bottom). This is achieved by jointly segmenting and clustering all sequences. The letters A, B, C, and D correspond to the identified actions." + ], + "image_footnote": [], + "bbox": [ + 514, + 306, + 908, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This highly constrained setting limits their applicability to real-world scenarios, where multiple actions are performed within a single sequence.", + "bbox": [ + 511, + 590, + 906, + 637 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the context of RGB videos, unsupervised approaches for temporal action segmentation have been proposed [26, 27, 29, 36, 50, 56, 61]. These methods discover actions in a set of long untrimmed video sequences by identifying actions across sequences and multiple actions within each sequence. While these approaches achieved promising results on video data, they have not been evaluated on skeleton sequences. Our investigations reveal that such methods either struggle to recognize re-occurring actions due to post-processing with Viterbi decoding or produce overly small segments, as visualized in qualitative results.", + "bbox": [ + 511, + 640, + 908, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we thus propose a novel approach for unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Our model uses a dilated temporal convolution based sequence-to-sequence autoencoder, which preserves the skeleton structure by ensuring that the information of the different joints remains disen", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tangled in the embedding space. We will show that this joint-based disentanglement is essential for improving performance. Subsequently, the latent representations generated by the temporal encoder are then patchified into nonoverlapping sub-sequences of motion, which are quantized by a skeleton motion quantization module. Specifically, each latent motion patch is assigned to the closest cluster prototype, and these prototypes are learned during training with the aim of capturing the most representative skeleton motion words of each action in the dataset. We term the approach thus Skeleton Motion Quantization (SMQ).", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To assess the effectiveness of SMQ, we evaluated it on three skeleton-based datasets, namely HuGaDB [9], LARa [44], and BABEL [46]. We demonstrate that it outperforms state-of-the-art approaches for unsupervised temporal action segmentation by a large margin, which highlights the importance of developing methods that exploit the inherent structure of skeleton sequences. We also show that it outperforms existing skeleton-based representation learning approaches on the task of unsupervised skeleton-based temporal action segmentation. In summary, our contributions are the following:", + "bbox": [ + 89, + 257, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We thoroughly evaluate existing unsupervised temporal action segmentation approaches for skeleton data and provide protocols for three datasets.", + "2. We introduce SMQ, a novel unsupervised framework for skeleton-based temporal action segmentation. It learns representative skeleton motion words in a joint-disentangled embedding space.", + "3. We demonstrate that SMQ outperforms existing unsupervised action segmentation methods, underscoring the importance of utilizing the structure of skeleton data." + ], + "bbox": [ + 89, + 422, + 482, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 588, + 232, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Motion Segmentation. Human motion segmentation aims to break down continuous motion into distinct, meaningful segments, often at a very fine granularity [2, 23, 48, 69], such as phases of a gait cycle. In contrast, temporal action segmentation targets broader, higher-level actions over long, untrimmed sequences [10]. One of the earliest approaches for human motion segmentation was introduced by Barbic et al. [2], using a probabilistic PCA-based method to detect significant distribution changes in human poses. Other early methods include hierarchical clustering and graph-based techniques to align sequences and segment motion based on structural similarity [23, 58, 69, 70]. Recent methods for human motion segmentation use heuristics such as extended Douglas-Peucker keypoint detection [42], zero-velocity crossing with adaptive clustering [59], and curvature analysis on kinematic data [1] to isolate motion primitives within individual sequences. While these approaches focus on fine-grained motion patterns, there are recent methods [4, 11, 47] that aim to", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "identify complex, high-level actions within individual video sequences. However, they are still limited to segmenting a single sequence at a time. Our approach, in contrast, identifies globally learned actions across multiple sequences and subjects.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Action Segmentation. Temporal action segmentation involves segmenting long untrimmed videos, often containing multiple actions per video, into contiguous segments, each labeled with a specific action [3, 12, 20, 24, 28, 38, 41]. Unlike traditional action recognition, which aims to classify short videos containing a single action, temporal action segmentation addresses the sequential nature of actions. To avoid an expensive annotation process, unsupervised methods [26, 27, 29, 36, 49, 50, 56, 61] have been proposed for this task. Specifically, Mallow [49] is a two-stage iterative approach, combining discriminative learning from visual features with generative modeling of temporal structure using a Generalized Mallows Model. CTE [26] aims to learn continuous temporal embeddings with the supervision of the relative time-stamps and then clusters the learned features in a two step fashion. TOT [27] proposed a joint representation learning and online clustering framework which applies temporal optimal transport on the features from the encoder and cluster prototypes to obtain pseudo action labels for training. UFSA [56] combines the encoder from ASFormer [65] and the decoder from UVAST [3], allowing the use of segment-level cues. Lastly, ASOT [61] utilizes temporally consistent unbalanced optimal transport, relaxing the balanced assignment assumption in TOT. Our results demonstrate that despite progress in unsupervised temporal action segmentation, existing methods do not perform well on skeleton data, as they do not exploit the inherent structure of skeleton data.", + "bbox": [ + 511, + 170, + 903, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skeleton-based Temporal Action Segmentation. Specifically for skeleton data, a variety of fully supervised temporal action segmentation methods have been proposed. [14, 54, 60, 63] combine GCNs [22] and TCNs [12, 28, 31] to capture both the spatial and temporal aspects of these sequences. Additionally, some skeleton-based temporal action segmentation approaches further investigate the separation [33] and decoupling [32] of spatial and temporal information. Various attention mechanisms [39, 54] have been proposed to capture spatio-temporal dependencies while other approaches emphasize distinct frameworks, including action synthesis [60] and language priors [21]. However, all of these methods rely on full supervision, requiring extensive labeling. In this work, we propose a skeleton-based temporal action segmentation approach that automatically learns holistic, globally consistent actions across the entire dataset in an unsupervised manner.", + "bbox": [ + 511, + 595, + 903, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Representation Learning for Skeleton Data. Self-supervised representation learning [5-8, 13, 15, 17, 18] aims to learn feature representations with", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "out human annotations by solving pretext tasks that exploit the structure of unlabeled data. Recent methods have explored this paradigm for skeleton-based action recognition [16, 30, 34, 35, 52, 62, 67, 68], focusing on short, trimmed sequences containing a single action. In particular, CrosSCLR [30], AimCLR [16], and ActCLR [35] adopt contrastive learning to extract discriminative representations. However, these methods solely learn representations without predicting action labels during training, operating on short, single-action sequences. They thus require annotations for downstream evaluation and are not applicable to long, untrimmed sequences.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hBehaveMAE [51] uses a hierarchical framework with a masked autoencoder to learn interpretable latent representations of actions at varying levels of granularity. LAC [64] utilizes pre-trained visual encoders with the task of learning composable actions from synthesized data. The model is then fine-tuned for different tasks including action segmentation to evaluate the learned representation. Unlike self-supervised approaches that learn action-related representations without assigning actions during training, and therefore require annotated data for downstream fine-tuning, our method discovers and segments actions directly during training in a fully unsupervised and label-free manner.", + "bbox": [ + 89, + 272, + 480, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Unsupervised Skeleton-based Temporal Action Segmentation", + "text_level": 1, + "bbox": [ + 89, + 464, + 480, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we address the task of unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Given input skeleton sequences encoding joint positions, orientations, or related features, the objective of skeleton-based action segmentation is to map each frame to one of the predefined action categories. In the unsupervised case, the actions are unknown and need to be discovered from the data. Formally, the input skeleton sequences are defined as $\\mathbf{X} \\in \\mathbb{R}^{N \\times C \\times T \\times V}$ , where $N$ is the number of sequences or batch size, $C$ is the number of feature channels representing a single joint, $T$ is the sequence length, and $V$ is the number of joints. The goal is to discover $K$ semantically meaningful clusters and to generate an output label sequence $\\mathbf{Y} \\in \\mathbb{N}^{N \\times T}$ , where each element $\\mathbf{Y}_{nt} \\in \\{1,2,\\dots,K\\}$ corresponds to the cluster assigned to the $t$ -th frame of the $n$ -th sequence. The main challenge is to discover actions across sequences and subjects, since it requires dealing with a large variability in how actions can be executed by different individuals.", + "bbox": [ + 89, + 508, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the task has been addressed for video data, it has not been studied for skeleton data. In our experiments, we show that state-of-the-art unsupervised approaches for video-based temporal action segmentation do not perform well for skeleton data since they do not exploit the nature of skeleton data. We therefore propose an approach for unsupervised skeleton-based temporal action segmentation.", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The approach is illustrated in Fig. 2. Each joint of a sequence will be mapped to an embedding space by an encoder, which will be described in Section 3.1. The encoder learns an embedding for each joint separately, which ensures that the information of the different joints remains disentangled in the embedding space and that the representation is not dominated by a subset of joints. Another important aspect of the approach is the conversion of a frame-wise representation into a patch-based representation, which is then quantized. We term this approach skeleton motion quantization (SMQ) and describe it in Section 3.2. Finally, the decoder reconstructs the original skeleton sequence from the quantized patches as described in Section 3.3. While the end-to-end training is described in Section 3.4, the following sections provide a detailed description of each component.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.Encoder", + "text_level": 1, + "bbox": [ + 511, + 344, + 614, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our approach, we process input skeleton sequences to capture joint-specific temporal dynamics effectively. To learn latent subspaces for each joint individually, we reshape and permute $\\mathbf{X}$ into $\\mathbf{X}' \\in \\mathbb{R}^{(N \\cdot V) \\times C \\times T}$ . This transformation treats each joint's temporal sequence $\\mathbf{X}_{nv}' \\in \\mathbb{R}^{C \\times T}$ as an independent sample, allowing the encoder to capture fine-grained motion patterns unique to each joint. Specifically, the encoder is a variant of a Temporal Convolutional Network (TCN) [12, 31], which consists of a series of dilated residual layers that progressively refine the temporal features of $\\mathbf{X}'$ . Each stage begins with a $1 \\times 1$ convolution to project the input features into a higher-dimensional space, followed by several residual blocks with dilated convolutions to capture long-range temporal dependencies. The dilation factor in these residual layers increases exponentially, allowing the network to model temporal patterns at multiple scales. Finally, a $1 \\times 1$ convolutional layer maps the features to the latent dimension $D$ , yielding latent embeddings $\\mathbf{Z}_{nv} \\in \\mathbb{R}^{D \\times T}$ .", + "bbox": [ + 511, + 367, + 903, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Skeleton Motion Words", + "text_level": 1, + "bbox": [ + 511, + 665, + 732, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since skeleton sequences represent joint positions, orientations, or higher-order representations as time series, we reformulate the unsupervised skeleton-based temporal action segmentation task as an online time series clustering problem [57]. In this formulation, a skeleton sequence is mapped to a sequence of skeleton motion words which are learnable discrete units representing prototypical short-term temporal skeleton motion patterns. This allows us to capture the dynamic and temporal dependencies inherent in human motion data, aligning well with the nature of sequential joint movements across time.", + "bbox": [ + 511, + 688, + 903, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, the skeletons are first mapped into an embedding space using a temporal encoder, while keeping the information of the different joints disentangled. To", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8433902983024d70186dcbf6a1943c935008ce5ce320f8c351c6aa154d7eaf16.jpg", + "image_caption": [ + "Figure 2. Overview of SMQ. Given a set of skeleton sequences, our approach discovers actions that are semantically consistent across all sequences. The discovered actions correspond to clusters of learned motion words in the codebook. The model first encodes the skeleton sequences into a joint-based disentangled embedding space. Each embedded sequence is then divided into short non-overlapping temporal patches and a codebook with $K$ motion words is learned for the entire dataset via a patch-based quantization process, which assigns each patch to its nearest motion word using Euclidean distance. This assignment directly provides the segmentation of each sequence based on the motion word indices. In order to learn meaningful representations, the decoder reconstructs the input sequences from the quantized patches, where each patch is replaced by its corresponding motion word." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 869, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model skeleton motion, the embeddings are divided into non-overlapping patches to capture local temporal structure over fixed intervals. During training, skeleton motion words are learned, and each patch is assigned to its closest motion word based on the learned representation. Therefore, it effectively captures both spatial and temporal patterns in skeleton sequences, allowing for more robust and coherent action segmentation.", + "bbox": [ + 88, + 505, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Temporal Patching", + "text_level": 1, + "bbox": [ + 89, + 637, + 272, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to achieve skeleton motion quantization, we start by patching the latent representations from the encoder into fixed-length short time series. Specifically, after reorganizing the per-joint embeddings $\\mathbf{Z}_{nv}$ and concatenating joint features of the embedded skeleton to form $\\mathbf{Z}_{\\text{concat}} \\in \\mathbb{R}^{N \\times T \\times (V \\cdot D)}$ , we ensure that the sequence length $T$ is divisible by the patch size $P$ by applying zero-padding as necessary. We then partition the sequence into $M = \\frac{T}{P}$ non-overlapping temporal patches, yielding a patched tensor $\\mathbf{Z}_p \\in \\mathbb{R}^{N \\times M \\times P \\times (V \\cdot D)}$ .", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This structured patching allows us to measure similarity based on sequences of embedding vectors rather than individual vectors, accommodating greater temporal variability. This enables the learning of representative skeleton motion words that encapsulate temporal dynamics, rather than relying on single-frame representations.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Quantization", + "text_level": 1, + "bbox": [ + 511, + 505, + 651, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the temporal patching $\\mathbf{Z}_p$ , the network learns a patch-based codebook $\\mathcal{C} = \\{\\mathbf{c}_k\\}_{k=1}^K$ , where each codebook element $\\mathbf{c}_k \\in \\mathbb{R}^{P \\times (V \\cdot D)}$ . As before, $P$ denotes the patch size, $V$ the number of joints, and $D$ the latent dimension. Each element $\\mathbf{c}_k$ in the codebook defines what we refer to as a skeleton motion word, which serves as a discrete representation of skeleton motion over a temporal patch. The codebook $\\mathcal{C}$ is initialized randomly using Kaiming uniform initialization.", + "bbox": [ + 511, + 523, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During quantization, each patch in the batch $\\mathbf{p}_i\\in \\mathbf{Z}_p$ is assigned to its nearest motion word in the codebook based on the Euclidean distance. Each patch is then replaced with its closest motion word $\\mathbf{c}_{k_i}$ , where $\\mathbf{c}_{k_i}$ denotes the codebook element assigned to $\\mathbf{p}_i$ , yielding the quantized patch $\\mathbf{q}_i\\in \\mathbb{R}^{P\\times (V\\cdot D)}$ , i.e.,", + "bbox": [ + 511, + 659, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {i} = \\mathbf {c} _ {k _ {i}} \\quad \\text {w i t h} \\quad k _ {i} = \\underset {k} {\\arg \\min } \\| \\mathbf {p} _ {i} - \\mathbf {c} _ {k} \\| _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 758, + 906, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This produces the quantized tensor $\\mathbf{Q}_p\\in \\mathbb{R}^{N\\times M\\times P\\times (V\\cdot D)}$ . We then depatchify the quantized tensor $\\mathbf{Q}_p$ to obtain $\\mathbf{Q}\\in \\mathbb{R}^{N\\times T\\times (V\\cdot D)}$ . From the assignment of a patch $\\mathbf{p}_i$ to the motion word $\\mathbf{c}_{k_i}$ , we directly derive the segmentation $\\mathbf{Y}\\in \\mathbb{N}^{N\\times T}$ of all sequences in the batch, where each element $\\mathbf{Y}_{nt}\\in \\{1,2,\\dots ,K\\}$ denotes the index of the assigned motion word at the $t$ -th frame of the $n$ -th sequence.", + "bbox": [ + 511, + 792, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12104", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Codebook Update", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the quantization process, where patches are assigned to their nearest motion words, we update the motion words by averaging the assigned patches:", + "bbox": [ + 89, + 109, + 483, + 157 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {k} = \\frac {1}{| \\mathcal {P} _ {k} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 166, + 482, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_k = \\{\\mathbf{p}_i\\mid k_i = k\\}$ denotes the set of patches assigned to the $k$ -th motion word. To enhance the stability of codebook updates within the SMQ framework, we use Exponential Moving Average (EMA) updates [57]:", + "bbox": [ + 89, + 215, + 483, + 277 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {k} ^ {(t)} \\leftarrow \\alpha \\mathbf {c} _ {k} ^ {(t - 1)} + (1 - \\alpha) \\tilde {\\mathbf {c}} _ {k} ^ {(t)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 286, + 482, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {c}} _ {k} ^ {(t)} = \\frac {1}{| \\mathcal {P} _ {k} ^ {(t)} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k} ^ {(t)}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 311, + 482, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_k^{(t)} = \\{\\mathbf{p}_i\\mid k_i = k\\}$ is the set of patches assigned to the $k$ -th motion word at iteration $t$ , and $\\alpha = 0.5$ is the EMA decay factor that controls the smoothness of the updates.", + "bbox": [ + 89, + 366, + 483, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.Decoder", + "text_level": 1, + "bbox": [ + 89, + 422, + 191, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reconstruct the input skeleton sequences from the quantized latent representations, we design a decoder that mirrors the encoder's architecture. After quantization, the depatchified tensor $\\mathbf{Q}$ is reshaped into $\\mathbf{Q}' \\in \\mathbb{R}^{(N \\cdot V) \\times D \\times T}$ to obtain the joint independence. This reshaping allows the decoder to process each joint's temporal sequence individually, similar to the encoder's approach. The reconstructed skeleton, denoted as $\\hat{\\mathbf{X}} \\in \\mathbb{R}^{N \\times C \\times T \\times V}$ , is then obtained after reshaping.", + "bbox": [ + 89, + 446, + 483, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 89, + 590, + 233, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The network is trained end-to-end using two loss functions. The first term is the reconstruction loss that compares the reconstructed sequences $\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times C\\times T\\times V}$ with the original sequences $\\mathbf{X}\\in \\mathbb{R}^{N\\times C\\times T\\times V}$ . The loss ensures that the encoder and codebook preserve as much information from the input sequences as possible. While the mean squared error (MSE) between $\\mathbf{X}$ and $\\hat{\\mathbf{X}}$ is an obvious choice, we propose to compute inter-joint distance MSE loss, defined over the distance from each joint to all other joints within the same frame:", + "bbox": [ + 89, + 614, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e c}} = \\frac {1}{N \\cdot T \\cdot V ^ {2}} \\sum_ {n = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {v = 1} ^ {V} \\sum_ {w = 1} ^ {V} \\left(d X _ {n t v w} - d \\hat {X} _ {n t v w}\\right) ^ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 773, + 480, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} \\quad \\begin{array}{r l} d X _ {n t v w} & = \\left\\| \\mathbf {X} _ {n t v} - \\mathbf {X} _ {n t w} \\right\\| _ {2}, \\\\ d \\hat {X} _ {n t v w} & = \\left\\| \\hat {\\mathbf {X}} _ {n t v} - \\hat {\\mathbf {X}} _ {n t w} \\right\\| _ {2}. \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 816, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss has the advantage that it is inherently translation and rotation-invariant and only measures differences in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "poses. In our experiments, we show that the proposed loss performs better than a standard MSE loss.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the learned embeddings are separated into distinct skeleton motion words, we employ the commitment loss [57]. This loss encourages the encoder to produce latent embeddings that are close to their assigned motion words. The commitment loss is defined as", + "bbox": [ + 511, + 121, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o m m i t}} = \\sum_ {\\mathbf {p} _ {i} \\in \\mathbf {Z} _ {p}} \\| \\operatorname {s g} \\left[ \\mathbf {c} _ {k _ {i}} \\right] - \\mathbf {p} _ {i} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 203, + 903, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{c}_{k_i} \\in \\{\\mathbf{c}_1, \\dots, \\mathbf{c}_K\\}$ is the motion word that the latent patch embedding $\\mathbf{p}_i$ has been assigned to, and $\\mathrm{sg}[\\cdot]$ denotes the stop-gradient operator, which stops the backpropagation of the gradient.", + "bbox": [ + 511, + 246, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss function combines both the reconstruction loss and the commitment loss", + "bbox": [ + 511, + 306, + 905, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = \\lambda L _ {\\text {r e c}} + L _ {\\text {c o m m i t}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 347, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda = 0.001$ is the hyperparameter that balances the loss terms.", + "bbox": [ + 511, + 373, + 905, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 416, + 645, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 440, + 614, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated our method on three skeleton-based action datasets: HuGaDB [9], LARa [44], and BABEL [46]. Additional results on the PKU-MMD v2 dataset [37] are reported in the suppl. material.", + "bbox": [ + 511, + 462, + 905, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HuGaDB consists of 364 trials capturing 10 lower limb activities, including walking, running, and sitting. Data were collected from 18 participants using six IMUs (3-axis accelerometer and 3-axis gyroscope) placed on the thighs, shins, and feet, recorded at $60\\mathrm{Hz}$ .", + "bbox": [ + 511, + 523, + 905, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LARa comprises 439 trials across 8 warehouse-related action categories. Data were collected from 16 participants using marker-based MoCap (200 fps), recording the 3-axis position and orientation of 22 full body joints. The dataset includes 960 minutes of recordings. LARa is downsampled by a factor of 4, reducing the frame rate to 50 fps [14]. Additionally, the skeleton is centered from the root joint for translation invariance.", + "bbox": [ + 511, + 598, + 905, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The BABEL dataset contains 43 hours of 3D human motion sequences from AMASS [43], annotated with over 63,000 frame-level labels across more than 250 action categories. Following the protocol of [66], we extract 25 full-body joint skeletons from SMPL [40] parameters, construct three subsets, each focusing on four action categories, and downsample all sequences to 30 fps, after centering them on the root joint. In addition to the original protocol, we exclude sequences containing more than $50\\%$ background actions to ensure that the models learn relevant actions. The evaluation metrics and implementation details are described in the suppl. material.", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5be54b1e991989d447837a7e77ae6476d857c65317947da025f12b85525ca548.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHuGaDBLARa
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
SupervisedTCN [28]88.3---56.861.5---20.0
ST-GCN [63]88.7---67.767.9---25.8
MS-TCN [12]86.8---89.965.8---39.6
MS-GCN [14]90.4---93.065.6---43.6
UnsupervisedCTE [26]33.84.70.60.60.523.316.88.15.22.3
CTE + Viterbi [26]39.221.713.29.57.523.017.76.83.71.6
TOT [27]33.83.10.70.50.421.47.85.32.71.1
TOT + Viterbi [27]33.820.815.610.57.532.617.711.67.43.2
ASOT [61]33.917.44.53.83.022.923.417.812.15.7
SMQ (ours)42.036.138.531.524.337.439.434.728.416.4
", + "bbox": [ + 130, + 88, + 867, + 272 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e3be8f6bfcbea0b6ad9e553110e09ec60455d95f63be70c7fbe0a2f5f293afd.jpg", + "table_caption": [ + "Table 1. Comparison to supervised and unsupervised temporal action segmentation methods on the HuGaDB and LARa datasets." + ], + "table_footnote": [], + "table_body": "
MethodBABEL Subset-1BABEL Subset-2BABEL Subset-3
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
CTE [26]34.828.625.017.59.540.330.617.812.27.431.413.18.25.83.6
CTE + Viterbi [26]30.936.223.215.27.342.430.724.319.512.831.230.920.715.28.4
TOT [27]31.818.714.27.64.435.412.813.78.64.331.57.14.92.91.7
TOT + Viterbi [27]29.129.331.520.89.935.336.835.930.019.834.033.831.326.817.9
ASOT [61]35.343.142.334.124.543.137.740.333.423.438.027.127.421.614.3
SMQ (ours)36.638.540.932.822.349.137.843.837.427.440.638.638.029.319.3
", + "bbox": [ + 93, + 299, + 906, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison to unsupervised temporal action segmentation methods on the BABEL dataset.", + "bbox": [ + 200, + 420, + 794, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to State of the Art", + "text_level": 1, + "bbox": [ + 89, + 460, + 364, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our proposed method, SMQ, with state-of-the-art unsupervised temporal action segmentation methods on the HuGaDB, LARa and BABEL datasets, as shown in Tables 1 and 2. Across all evaluation metrics, SMQ consistently outperforms unsupervised temporal action segmentation approaches.", + "bbox": [ + 89, + 484, + 482, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the HuGaDB dataset in Table 1, SMQ achieves a MoF of 42.0 and Edit Score of 36.1, surpassing the best performing approach CTE [26] by 2.8 and 14.4, respectively. For the F1@50 metric, our method has the score of 24.3, which is by 16.8 higher than CTE [26] and TOT [27]. On the LARa dataset, SMQ achieves a MoF of 37.4, Edit Score of 39.4, and F1@50 of 16.4, outperforming the best approach ASOT [61] by a large margin as well. For completeness, we also include results from supervised approaches, which reported only MoF and F1@50. While supervised approaches achieve much higher scores, unsupervised approaches do not use any annotations and solve thus a much harder task.", + "bbox": [ + 89, + 578, + 482, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the BABEL dataset in Table 2, SMQ achieves a higher MoF than ASOT [61] for Subset-1, but ASOT performs slightly better for the other metrics. Whereas on Subset-2 and Subset-3, SMQ outperforms all unsupervised temporal action segmentation methods.", + "bbox": [ + 89, + 762, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 shows a qualitative comparison between the methods. Both CTE and TOT generate too many small segments. They thus employ a post-processing technique utilizing Viterbi decoding [10, 25], to refine initial predictions", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and achieve temporally consistent action segments. The Viterbi decoding [25] computes the most probable sequence of actions by maximizing the joint likelihood under a first-order Markov model, using transition probabilities that penalize frequent label changes. While this improves segmental metrics by reducing fragmentation, it also imposes a strong prior that each action appears only once per sequence, which is not the case in the datasets.", + "bbox": [ + 511, + 460, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ASOT can identify re-occurring actions and does not suffer from small segments. The approach, however, has the tendency to predict rather large segments. The small segments in the HuGaDB and LARa datasets are not discovered. Since the segments are longer in the BABEL subsets, ASOT performs better on this dataset. While SMQ does not generate perfect results, for instance it mixes standing up and sitting down in the example from the HuGaDB dataset, the results are better aligned with the ground truth. It is able to detect re-occurring actions, and it does not have the tendency to generate too many small segments, as CTE and TOT, or overly large segments as ASOT. Overall, the results show that SMQ is more effective in handling diverse segmentation scenarios. However, as can be seen in Fig. 3, identifying precise action transitions remains a challenge. Since actions are assigned at the patch level, transitions can only occur at patch boundaries. Improving the detection of action boundaries needs to be investigated in future work.", + "bbox": [ + 511, + 582, + 905, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare SMQ to self-supervised representation learning approaches for skeleton data [16, 35, 52]. These approaches are typically first trained in a self-supervised", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12106", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a144090599241001b4c78ad6d0f7fa0da08f0ba96db58e94ea859f4c335e1ccd.jpg", + "image_caption": [ + "(a) HuGaDB" + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 495, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/204e6da2416d2db82df422d8ae97130dc0311abb2dccca41870d299805429278.jpg", + "image_caption": [ + "(b) LARa" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 893, + 212 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6c8fca9f9c9a2cea07211823538c2df5409a24d1c4a5729dfa8a1ab5e89375a.jpg", + "image_caption": [ + "(c) BABEL Subset-1" + ], + "image_footnote": [], + "bbox": [ + 102, + 229, + 491, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecadcd828e68923d725761c7363ed820a2dddaa62a18a6b481803b89863b803d.jpg", + "image_caption": [ + "(d) BABEL Subset-3", + "Figure 3. Qualitative results for unsupervised action segmentation algorithms." + ], + "image_footnote": [], + "bbox": [ + 504, + 229, + 893, + 349 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6b1d3a53f21ee94d8d170d97a0aee279f86b7927c6a3822acee9749290a6223b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Predict&Cluster [52]39.331.533.224.218.318.828.519.212.95.432.232.132.224.314.2
AimCLR [16]34.730.630.923.711.322.028.022.115.67.740.329.732.525.715.5
ActCLR [35]28.221.610.46.24.123.330.623.716.77.832.633.232.824.413.4
SMQ (ours)42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
", + "bbox": [ + 94, + 391, + 906, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison to self-supervised skeleton-based representation learning methods.", + "bbox": [ + 235, + 482, + 758, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "way and then fine-tuned for different downstream tasks like skeleton-based action recognition using annotated data. In order to apply these methods to unsupervised temporal action segmentation, we divided the input skeleton sequences into non-overlapping patches, consistent with the latent patching in SMQ. These patches were fed through the self-supervised representation learning approaches to learn patch-level action representations. We then performed k-means clustering over the patch-level representations. The results in Table 3 show that SMQ consistently outperforms existing self-supervised representation learning approaches on all the datasets. It needs to be noted that these approaches are designed for datasets with only one action per sequence, whereas unsupervised temporal action segmentation requires processing long, untrimmed skeleton sequences containing multiple actions per sequence. Compared to SMQ, they therefore learn less distinctive representations for this task as it is visualized in the suppl. material.", + "bbox": [ + 88, + 508, + 482, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 797, + 200, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the impact of different design choices in our model.", + "bbox": [ + 89, + 821, + 482, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Disentangled Embedding. As illustrated in Fig. 2, the encoder processes each joint independently, which ensures that the information of the different joints", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "remains disentangled in the embedding space. To evaluate the effectiveness of the disentangled embedding space, we conducted an ablation study. In Table 4, we compare our approach to a setup where we concatenate the joints before learning an embedding. In the latter case, the joints get entangled in the embedding and the structural information of the skeleton gets lost. The results demonstrate that the disentangled embedding enhances our model's performance, indicating that learning joint representations separately allows for more accurate action segmentation by enabling each joint to reside in distinct subspaces.", + "bbox": [ + 511, + 508, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Patch Size. We apply temporal patching to the latent skeleton embeddings to obtain skeleton motion words. Specifically, we use a fixed patch size to create non-overlapping patches of the representations. To investigate the effect of patch size on the performance of our model, we conducted an ablation study with varying patch sizes, as presented in Table 5. Without patching, i.e., patch size is 1, we obtain a frame-wise representation. The results indicate that SMQ consistently outperforms the frameswise representation across all evaluated metrics, with significant improvements in segmental metrics. These findings emphasize that frame-wise quantization leads to over-segmentation, as it fails to capture the inherent temporal consistency of actions. In contrast, SMQ captures the tem", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12107", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/876c9c501d6892b7722ad8cda9cfff6625e0ba1e03791696f7e1e5d4625a229a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EmbeddingHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Entangled38.934.235.627.918.926.635.929.522.812.340.132.134.927.116.9
Disentangled42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
", + "bbox": [ + 94, + 88, + 906, + 150 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e55b70a41b42c0f27dcd999ea2351ec373fa98bf0314a488a1742c3cdab9a792.jpg", + "table_caption": [ + "Table 4. Impact of disentangled joint embedding on the HuGaDB, LARa, and BABEL datasets." + ], + "table_footnote": [], + "table_body": "
Patch SizeMoFEditF1@{10, 25, 50}
1 (Frame-wise)33.925.621.214.98.6
25 (0.5 sec)35.235.732.425.014.9
50 (1 sec)37.439.434.728.416.4
75 (1.5 sec)37.239.134.728.417.1
100 (2 sec)30.836.133.027.015.4
150 (3 sec)31.434.032.626.514.7
", + "bbox": [ + 119, + 174, + 452, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cad07acd7eee282143bba74fc2fc249381bc7ddf18fd0d46556e560e5a2af90d.jpg", + "table_caption": [ + "Table 5. Impact of motion patch size on the LARa dataset." + ], + "table_footnote": [], + "table_body": "
LcommitReconstruction LossMoFEditF1@{10,25,50}
✓X29.929.926.817.78.8
XInter-Joint Dist. MSE31.034.630.624.414.4
✓MSE34.639.434.527.415.7
✓Root Distance MSE28.034.930.223.814.1
✓Inter-Joint Dist. MSE37.439.434.728.416.4
", + "bbox": [ + 99, + 325, + 482, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "poral variability of actions more effectively by grouping contiguous latent embeddings into patches. This results in more semantically meaningful action clusters that yield temporally consistent segmentations. On the LARa dataset, a patch size of 50 performs best, which corresponds to one second of motion, as the data is downsampled to 50 fps. We keep this one-second patching consistent across all datasets. Impact of Loss Variants. To evaluate the effectiveness of different loss components, we conduct an ablation study examining various configurations of the commitment loss and reconstruction loss. We first evaluate three variants of the reconstruction loss in the last three rows of Table 6. The proposed inter-joint distance MSE loss performs better than the commonly used MSE loss since it is invariant to translation and rotation. The root distance MSE loss is obtained by computing the MSE over the distances of each joint to the root joint within the same frame. While this loss is also translation and rotation invariant, it performs even worse than the inter-joint variant. Furthermore, we evaluate the impact of the commitment loss and reconstruction loss by omitting one of them as presented in the first two rows of the Table 6. Removing either of the loss terms substantially decreases the performance.", + "bbox": [ + 89, + 460, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Reconstruction Loss Weight. In our ablation study, we varied the reconstruction loss weight, $\\lambda$ (7), to understand its effect on model performance. By adjusting $\\lambda$ , we explore the balance between the reconstruction and commitment loss in Table 7. Our model is not very sensitive to this parameter and performs well for $\\lambda = 0.001$ .", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f5f638b9adbcfacf0cc2240a17479036cd9f00e829e4c94975d69bed85cf1c21.jpg", + "table_caption": [ + "Table 6. Impact of loss function variants on the LARa dataset." + ], + "table_footnote": [], + "table_body": "
λMoFEditF1@{10, 25, 50}
0.135.938.433.426.915.6
0.0136.238.934.427.815.9
0.00137.439.434.728.416.4
0.000134.139.835.427.916.1
", + "bbox": [ + 571, + 176, + 846, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b7d80ad1093b3f76ebb9371898bbb73ff8b697e594edd1b9dad6da1e3ab41533.jpg", + "table_caption": [ + "Table 7. Impact of $\\lambda$ on the LARa dataset." + ], + "table_footnote": [], + "table_body": "
αMoFEditF1@{10, 25, 50}
0.3534.839.435.027.615.5
0.5037.439.434.728.416.4
0.6535.538.533.927.215.8
0.8034.237.733.427.115.5
", + "bbox": [ + 578, + 300, + 841, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Impact of EMA decay rate $\\left( \\alpha \\right)$ on the LARa dataset.", + "bbox": [ + 527, + 397, + 888, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Decay Factor. To examine the role of the EMA decay factor $\\alpha$ (3), we perform an ablation study by evaluating multiple $\\alpha$ values in Table 8. The decay factor governs the balance between incorporating new patch information and maintaining historical motion word estimates, thereby influencing the smoothness and responsiveness of the updates. Our model performs best when $\\alpha$ is set to 0.5. Additional ablation studies are provided in the suppl. material.", + "bbox": [ + 511, + 438, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 584, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced SMQ, a method for unsupervised skeleton-based temporal action segmentation. It uses an autoencoder that preserves the skeleton structure by ensuring that the information of the different joints remains disentangled in the embedding space. The skeleton motion quantization module then patches the latent representations into short time series of fixed length and quantizes them into skeleton motion words. We performed an extensive evaluation and compared the proposed approach to state-of-the-art unsupervised temporal action segmentation approaches, as well as approaches for self-supervised skeleton-based representation learning. SMQ outperforms them by a large margin. Furthermore, we demonstrated the importance of learning the representation over patches instead of frames and the importance of an embedding space where the information of the different joints remains disentangled. As part of future work, it would be interesting to investigate a hierarchical version of SMQ and improve the detection of action boundaries.", + "bbox": [ + 511, + 613, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work has been supported by the ERC Consolidator Grants FORHUE (101044724) and ActionContraThreat (816564), the Ministry of Culture and Science of the State of North Rhine-Westphalia (iBehave network, Programme \"Netzwerke 2021\"), and the TRA Modelling (University of Bonn) as part of the Excellence Strategy of the federal and state governments.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Robert T Arn, Pradyumna Narayana, Tegan Emerson, Bruce A Draper, Michael Kirby, and Chris Peterson. Motion segmentation via generalized curvatures. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41 (12):2919-2932, 2018. 2", + "[2] Jernej Barbič, Alla Safonova, Jia-Yu Pan, Christos Faloutsos, Jessica K Hodgins, and Nancy S Pollard. Segmenting motion capture data into distinct behaviors. In Proceedings of Graphics Interface (GI), pages 185-194, 2004. 2", + "[3] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-68. Springer, 2022. 2", + "[4] Elena Belén Bueno-Benito, Biel Tura Vecino, and Mariella Dimiccoli. Leveraging triplet loss for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4921-4929, 2023. 2", + "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 2", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020.", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021.", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2", + "[9] Roman Chereshnev and Attila Kertesz-Farkas. Hugadb: Human gait database for activity recognition from wearable inertial sensor networks. In Analysis of Images, Social Networks and Texts (AIST), pages 131-141. Springer, 2018. 2, 5", + "[10] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques." + ], + "bbox": [ + 93, + 261, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 6", + "[11] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3323-3332, 2022. 2", + "[12] Yazan Abu Farha and Juergen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3575-3584, 2019. 2, 3, 6", + "[13] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2", + "[14] Benjamin Filtjens, Bart Vanrumste, and Peter Slaets. Skeleton-based action segmentation with multi-stage spatial-temporal graph convolutional neural networks. IEEE Transactions on Emerging Topics in Computing, 12(1):202-212, 2022. 1, 2, 5, 6", + "[15] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 33:21271-21284, 2020. 2", + "[16] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 762-770, 2022. 1, 3, 6, 7", + "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 2", + "[19] Babak Hosseini, Romain Montagne, and Barbara Hammer. Deep-aligned convolutional neural network for skeleton-based action recognition and segmentation. Data Science and Engineering (DSE), 5(2):126-139, 2020. 1", + "[20] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14024-14034, 2020. 2", + "[21] Haoyu Ji, Bowen Chen, Xinglong Xu, Weihong Ren, Zhiyong Wang, and Honghai Liu. Language-assisted skeleton action understanding for skeleton-based temporal action segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12109", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 2", + "[23] Björn Krüger, Anna Vögele, Tobias Willig, Angela Yao, Reinhard Klein, and Andreas Weber. Efficient unsupervised temporal segmentation of motion data. IEEE Transactions on Multimedia (T-MM), 19(4):797-812, 2016. 2", + "[24] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2", + "[25] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(4):765-779, 2020. 6", + "[26] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Juergen Gall. Unsupervised learning of action classes with continuous temporal embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12066-12074, 2019. 1, 2, 6", + "[27] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 1, 2, 6", + "[28] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 156-165, 2017. 2, 6", + "[29] Jun Li and Sinisa Todorovic. Action shuffle alternating learning for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12628-12636, 2021. 1, 2", + "[30] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4741-4750, 2021. 1, 3", + "[31] Shijie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45 (6):6647-6658, 2023. 2, 3", + "[32] Yunheng Li, Zhongyu Li, Shanghua Gao, Qilong Wang, Qibin Hou, and Ming-Ming Cheng. A decoupled spatiotemporal framework for skeleton-based action segmentation. arXiv preprint arXiv:2312.05830, 2023. 1, 2", + "[33] Yun-Heng Li, Kai-Yuan Liu, Sheng-Lan Liu, Lin Feng, and Hong Qiao. Involving distinguished temporal graph convolutional networks for skeleton-based temporal action segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):647-660, 2023. 2", + "[34] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "recognition. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2490-2498, 2020. 1, 3", + "[35] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Action-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2363-2372, 2023. 1, 3, 6, 7", + "[36] Wei Lin, Anna Kukleva, Horst Possegger, Hilde Kuehne, and Horst Bischof. Taec: Unsupervised action segmentation with temporal-aware embedding and clustering. In Computer Vision Winter Workshop, 2023. 1, 2", + "[37] Chunhui Liu, Yueyu Hu, Yanghao Li, Sijie Song, and Jiaying Liu. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv preprint arXiv:1703.07475, 2017. 5", + "[38] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10139-10149, 2023. 2", + "[39] Kaiyuan Liu, Yunheng Li, Yuanfeng Xu, Shuai Liu, and Shenglan Liu. Spatial focus attention for fine-grained skeleton-based action tasks. IEEE Signal Processing Letters, 29:1883-1887, 2022. 2", + "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 5", + "[41] Zijia Lu and Ehsan Elhamifar. FACT: Frame-action cross-attention temporal modeling for efficient supervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[42] Hao Ma, Zaiyue Yang, and Haoyang Liu. Fine-grained unsupervised temporal action segmentation and distributed representation for skeleton-based human motion analysis. IEEE Transactions on Cybernetics (T-CYB), 52(12):13411-13424, 2021. 2", + "[43] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5442-5451, 2019. 5", + "[44] Friedrich Niemann, Christopher Reining, Fernando Moya Rueda, Nilah Ravi Nair, Janine Anika Steffens, Gernot A Fink, and Michael Ten Hompel. Lara: Creating a dataset for human activity recognition in logistics using semantic attributes. Sensors, 20(15):4083, 2020. 2, 5", + "[45] Giancarlo Paoletti, Jacopo Cavazza, Cigdem Bryan, and Alessio Del Bue. Unsupervised human action recognition with skeletal graph laplacian and self-supervised viewpoints invariance. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 1", + "[46] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12110", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2, 5", + "[47] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11225-11234, 2021. 2", + "[48] Jan Sedmidubsky, Petra Budikova, Vlastislav Dohnal, and Pavel Zezula. Motion words: a text-like representation of 3d skeleton sequences. In Proceedings of the European Conference on Information Retrieval (ECIR), pages 527-541. Springer, 2020. 2", + "[49] Fadime Sener and Angela Yao. Unsupervised learning and segmentation of complex activities from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8368-8376, 2018. 2", + "[50] Federico Spurio, Emad Bahrami, Gianpiero Francesca, and Juergen Gall. Hierarchical vector quantization for unsupervised action segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6996-7005, 2025. 1, 2", + "[51] Lucas Stoffl, Andy Bonnetto, Stephane d'Ascoli, and Alexander Mathis. Elucidating the hierarchical nature of behavior with masked autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 106-125, 2024. 3", + "[52] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9631-9640, 2020. 1, 3, 6, 7", + "[53] Chenwei Tan, Tao Sun, Talas Fu, Yuhan Wang, Minjie Xu, and Shenglan Liu. Hierarchical spatial-temporal network for skeleton-based temporal action segmentation. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 28-39. Springer, 2023. 1", + "[54] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Stga-net: Spatial-temporal graph attention network for skeleton-based temporal action segmentation. In Proceedings of the IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 218-223. IEEE, 2023. 2", + "[55] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Spatial-temporal graph transformer network for skeleton-based temporal action segmentation. Multimedia Tools and Applications (MTAP), 83(15):44273-44297, 2024. 1", + "[56] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6426-6436, 2024. 1, 2", + "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3, 5", + "[58] Anna Vögele, Björn Krüger, and Reinhard Klein. Efficient unsupervised temporal segmentation of human motion. In" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), pages 167-176, 2014. 2", + "[59] Qifei Wang, Gregorij Kurillo, Ferda Ofli, and Ruzena Bajcsy. Unsupervised temporal segmentation of repetitive human actions based on kinematic modeling and frequency analysis. In Proceedings of the IEEE International Conference on 3D Vision (3DV), pages 562-570. IEEE, 2015. 2", + "[60] Leiyang Xu, Qiang Wang, Xiaotian Lin, and Lin Yuan. An efficient framework for few-shot skeleton-based temporal action segmentation. Proceedings of the European Conference on Computer Vision (ECCV), 232:103707, 2023. 1, 2", + "[61] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14618-14627, 2024. 1, 2, 6", + "[62] Ziwei Xu, Xudong Shen, Yongkang Wong, and Mohan S. Kankanhalli. Unsupervised motion representation learning with capsule autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 34:3205-3217, 2021. 3", + "[63] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2, 6", + "[64] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13679-13690, 2023. 3", + "[65] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASFormer: Transformer for action segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 2", + "[66] Qing Yu and Kent Fujiwara. Frame-level label refinement for skeleton-based weakly-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3322-3330, 2023. 5", + "[67] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 36-51. Springer, 2022. 1, 3", + "[68] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 3", + "[69] Feng Zhou, Fernando De la Torre, and Jessica K. Hodgins. Aligned cluster analysis for temporal segmentation of human motion. In Proceedings of the IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-7, 2008. 2", + "[70] Feng Zhou, Fernando De la Torre, and Jessica K Hodgins. Hierarchical aligned cluster analysis for temporal clustering of human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(3):582-596, 2012. 2" + ], + "bbox": [ + 516, + 92, + 903, + 896 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12111", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_model.json b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ef9879048b8d2161a301d136d5ab6c8fdfe156b --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.203, + 0.802, + 0.277 + ], + "angle": 0, + "content": "Uzay Gokay\\(^{1}\\) Federico Spurio\\(^{1,3}\\) Dominik R. Bach\\(^{1,2}\\) Juergen Gall\\(^{1,3}\\) \n\\(^{1}\\)University of Bonn, Germany \\(^{2}\\)University College London, UK \n\\(^{3}\\)Lamarr Institute for Machine Learning and Artificial Intelligence, Germany {u.goekay, d.bach}@uni-bonn.de {fspurio, gall}@iai.uni-bonn.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.485, + 0.66 + ], + "angle": 0, + "content": "Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods. Code is available at github.com/bachlab/SMQ." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.222, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Advancements in motion capture technology and pose estimation have enabled the precise recording of human motion, represented as time series of joint positions, orientations, or other representations of skeleton sequences. Recently, there has been an increased interest in temporally segmenting long skeleton sequences into action segments [14, 19, 21, 32, 53-55, 60]. These approaches, however, are predominantly supervised, requiring extensive labeled data for training. In contrast, prior unsupervised skeleton-based approaches [16, 30, 34, 35, 45, 52, 67] focus mainly on action recognition, assuming that the sequences are very short and contain only a single action per sequence." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.308, + 0.909, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.573 + ], + "angle": 0, + "content": "Figure 1. Our unsupervised skeleton-based temporal action segmentation framework, Skeleton Motion Quantization (SMQ), takes a set of untrimmed skeleton sequences as input (top) and discovers actions consistent across all sequences (bottom). This is achieved by jointly segmenting and clustering all sequences. The letters A, B, C, and D correspond to the identified actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.638 + ], + "angle": 0, + "content": "This highly constrained setting limits their applicability to real-world scenarios, where multiple actions are performed within a single sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.909, + 0.807 + ], + "angle": 0, + "content": "In the context of RGB videos, unsupervised approaches for temporal action segmentation have been proposed [26, 27, 29, 36, 50, 56, 61]. These methods discover actions in a set of long untrimmed video sequences by identifying actions across sequences and multiple actions within each sequence. While these approaches achieved promising results on video data, they have not been evaluated on skeleton sequences. Our investigations reveal that such methods either struggle to recognize re-occurring actions due to post-processing with Viterbi decoding or produce overly small segments, as visualized in qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we thus propose a novel approach for unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Our model uses a dilated temporal convolution based sequence-to-sequence autoencoder, which preserves the skeleton structure by ensuring that the information of the different joints remains disen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "tangled in the embedding space. We will show that this joint-based disentanglement is essential for improving performance. Subsequently, the latent representations generated by the temporal encoder are then patchified into nonoverlapping sub-sequences of motion, which are quantized by a skeleton motion quantization module. Specifically, each latent motion patch is assigned to the closest cluster prototype, and these prototypes are learned during training with the aim of capturing the most representative skeleton motion words of each action in the dataset. We term the approach thus Skeleton Motion Quantization (SMQ)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.424 + ], + "angle": 0, + "content": "To assess the effectiveness of SMQ, we evaluated it on three skeleton-based datasets, namely HuGaDB [9], LARa [44], and BABEL [46]. We demonstrate that it outperforms state-of-the-art approaches for unsupervised temporal action segmentation by a large margin, which highlights the importance of developing methods that exploit the inherent structure of skeleton sequences. We also show that it outperforms existing skeleton-based representation learning approaches on the task of unsupervised skeleton-based temporal action segmentation. In summary, our contributions are the following:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.424, + 0.483, + 0.468 + ], + "angle": 0, + "content": "1. We thoroughly evaluate existing unsupervised temporal action segmentation approaches for skeleton data and provide protocols for three datasets." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.469, + 0.482, + 0.529 + ], + "angle": 0, + "content": "2. We introduce SMQ, a novel unsupervised framework for skeleton-based temporal action segmentation. It learns representative skeleton motion words in a joint-disentangled embedding space." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.482, + 0.575 + ], + "angle": 0, + "content": "3. We demonstrate that SMQ outperforms existing unsupervised action segmentation methods, underscoring the importance of utilizing the structure of skeleton data." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.424, + 0.483, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.233, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Human Motion Segmentation. Human motion segmentation aims to break down continuous motion into distinct, meaningful segments, often at a very fine granularity [2, 23, 48, 69], such as phases of a gait cycle. In contrast, temporal action segmentation targets broader, higher-level actions over long, untrimmed sequences [10]. One of the earliest approaches for human motion segmentation was introduced by Barbic et al. [2], using a probabilistic PCA-based method to detect significant distribution changes in human poses. Other early methods include hierarchical clustering and graph-based techniques to align sequences and segment motion based on structural similarity [23, 58, 69, 70]. Recent methods for human motion segmentation use heuristics such as extended Douglas-Peucker keypoint detection [42], zero-velocity crossing with adaptive clustering [59], and curvature analysis on kinematic data [1] to isolate motion primitives within individual sequences. While these approaches focus on fine-grained motion patterns, there are recent methods [4, 11, 47] that aim to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "identify complex, high-level actions within individual video sequences. However, they are still limited to segmenting a single sequence at a time. Our approach, in contrast, identifies globally learned actions across multiple sequences and subjects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.171, + 0.905, + 0.593 + ], + "angle": 0, + "content": "Temporal Action Segmentation. Temporal action segmentation involves segmenting long untrimmed videos, often containing multiple actions per video, into contiguous segments, each labeled with a specific action [3, 12, 20, 24, 28, 38, 41]. Unlike traditional action recognition, which aims to classify short videos containing a single action, temporal action segmentation addresses the sequential nature of actions. To avoid an expensive annotation process, unsupervised methods [26, 27, 29, 36, 49, 50, 56, 61] have been proposed for this task. Specifically, Mallow [49] is a two-stage iterative approach, combining discriminative learning from visual features with generative modeling of temporal structure using a Generalized Mallows Model. CTE [26] aims to learn continuous temporal embeddings with the supervision of the relative time-stamps and then clusters the learned features in a two step fashion. TOT [27] proposed a joint representation learning and online clustering framework which applies temporal optimal transport on the features from the encoder and cluster prototypes to obtain pseudo action labels for training. UFSA [56] combines the encoder from ASFormer [65] and the decoder from UVAST [3], allowing the use of segment-level cues. Lastly, ASOT [61] utilizes temporally consistent unbalanced optimal transport, relaxing the balanced assignment assumption in TOT. Our results demonstrate that despite progress in unsupervised temporal action segmentation, existing methods do not perform well on skeleton data, as they do not exploit the inherent structure of skeleton data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.597, + 0.905, + 0.852 + ], + "angle": 0, + "content": "Skeleton-based Temporal Action Segmentation. Specifically for skeleton data, a variety of fully supervised temporal action segmentation methods have been proposed. [14, 54, 60, 63] combine GCNs [22] and TCNs [12, 28, 31] to capture both the spatial and temporal aspects of these sequences. Additionally, some skeleton-based temporal action segmentation approaches further investigate the separation [33] and decoupling [32] of spatial and temporal information. Various attention mechanisms [39, 54] have been proposed to capture spatio-temporal dependencies while other approaches emphasize distinct frameworks, including action synthesis [60] and language priors [21]. However, all of these methods rely on full supervision, requiring extensive labeling. In this work, we propose a skeleton-based temporal action segmentation approach that automatically learns holistic, globally consistent actions across the entire dataset in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Self-Supervised Representation Learning for Skeleton Data. Self-supervised representation learning [5-8, 13, 15, 17, 18] aims to learn feature representations with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "out human annotations by solving pretext tasks that exploit the structure of unlabeled data. Recent methods have explored this paradigm for skeleton-based action recognition [16, 30, 34, 35, 52, 62, 67, 68], focusing on short, trimmed sequences containing a single action. In particular, CrosSCLR [30], AimCLR [16], and ActCLR [35] adopt contrastive learning to extract discriminative representations. However, these methods solely learn representations without predicting action labels during training, operating on short, single-action sequences. They thus require annotations for downstream evaluation and are not applicable to long, untrimmed sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.453 + ], + "angle": 0, + "content": "hBehaveMAE [51] uses a hierarchical framework with a masked autoencoder to learn interpretable latent representations of actions at varying levels of granularity. LAC [64] utilizes pre-trained visual encoders with the task of learning composable actions from synthesized data. The model is then fine-tuned for different tasks including action segmentation to evaluate the learned representation. Unlike self-supervised approaches that learn action-related representations without assigning actions during training, and therefore require annotated data for downstream fine-tuning, our method discovers and segments actions directly during training in a fully unsupervised and label-free manner." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.482, + 0.501 + ], + "angle": 0, + "content": "3. Unsupervised Skeleton-based Temporal Action Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.794 + ], + "angle": 0, + "content": "In this work, we address the task of unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Given input skeleton sequences encoding joint positions, orientations, or related features, the objective of skeleton-based action segmentation is to map each frame to one of the predefined action categories. In the unsupervised case, the actions are unknown and need to be discovered from the data. Formally, the input skeleton sequences are defined as \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C \\times T \\times V}\\), where \\(N\\) is the number of sequences or batch size, \\(C\\) is the number of feature channels representing a single joint, \\(T\\) is the sequence length, and \\(V\\) is the number of joints. The goal is to discover \\(K\\) semantically meaningful clusters and to generate an output label sequence \\(\\mathbf{Y} \\in \\mathbb{N}^{N \\times T}\\), where each element \\(\\mathbf{Y}_{nt} \\in \\{1,2,\\dots,K\\}\\) corresponds to the cluster assigned to the \\(t\\)-th frame of the \\(n\\)-th sequence. The main challenge is to discover actions across sequences and subjects, since it requires dealing with a large variability in how actions can be executed by different individuals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "While the task has been addressed for video data, it has not been studied for skeleton data. In our experiments, we show that state-of-the-art unsupervised approaches for video-based temporal action segmentation do not perform well for skeleton data since they do not exploit the nature of skeleton data. We therefore propose an approach for unsupervised skeleton-based temporal action segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "The approach is illustrated in Fig. 2. Each joint of a sequence will be mapped to an embedding space by an encoder, which will be described in Section 3.1. The encoder learns an embedding for each joint separately, which ensures that the information of the different joints remains disentangled in the embedding space and that the representation is not dominated by a subset of joints. Another important aspect of the approach is the conversion of a frame-wise representation into a patch-based representation, which is then quantized. We term this approach skeleton motion quantization (SMQ) and describe it in Section 3.2. Finally, the decoder reconstructs the original skeleton sequence from the quantized patches as described in Section 3.3. While the end-to-end training is described in Section 3.4, the following sections provide a detailed description of each component." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.615, + 0.359 + ], + "angle": 0, + "content": "3.1.Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.655 + ], + "angle": 0, + "content": "In our approach, we process input skeleton sequences to capture joint-specific temporal dynamics effectively. To learn latent subspaces for each joint individually, we reshape and permute \\(\\mathbf{X}\\) into \\(\\mathbf{X}' \\in \\mathbb{R}^{(N \\cdot V) \\times C \\times T}\\). This transformation treats each joint's temporal sequence \\(\\mathbf{X}_{nv}' \\in \\mathbb{R}^{C \\times T}\\) as an independent sample, allowing the encoder to capture fine-grained motion patterns unique to each joint. Specifically, the encoder is a variant of a Temporal Convolutional Network (TCN) [12, 31], which consists of a series of dilated residual layers that progressively refine the temporal features of \\(\\mathbf{X}'\\). Each stage begins with a \\(1 \\times 1\\) convolution to project the input features into a higher-dimensional space, followed by several residual blocks with dilated convolutions to capture long-range temporal dependencies. The dilation factor in these residual layers increases exponentially, allowing the network to model temporal patterns at multiple scales. Finally, a \\(1 \\times 1\\) convolutional layer maps the features to the latent dimension \\(D\\), yielding latent embeddings \\(\\mathbf{Z}_{nv} \\in \\mathbb{R}^{D \\times T}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.733, + 0.681 + ], + "angle": 0, + "content": "3.2. Skeleton Motion Words" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.853 + ], + "angle": 0, + "content": "Since skeleton sequences represent joint positions, orientations, or higher-order representations as time series, we reformulate the unsupervised skeleton-based temporal action segmentation task as an online time series clustering problem [57]. In this formulation, a skeleton sequence is mapped to a sequence of skeleton motion words which are learnable discrete units representing prototypical short-term temporal skeleton motion patterns. This allows us to capture the dynamic and temporal dependencies inherent in human motion data, aligning well with the nature of sequential joint movements across time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 2, the skeletons are first mapped into an embedding space using a temporal encoder, while keeping the information of the different joints disentangled. To" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.87, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.395, + 0.908, + 0.493 + ], + "angle": 0, + "content": "Figure 2. Overview of SMQ. Given a set of skeleton sequences, our approach discovers actions that are semantically consistent across all sequences. The discovered actions correspond to clusters of learned motion words in the codebook. The model first encodes the skeleton sequences into a joint-based disentangled embedding space. Each embedded sequence is then divided into short non-overlapping temporal patches and a codebook with \\( K \\) motion words is learned for the entire dataset via a patch-based quantization process, which assigns each patch to its nearest motion word using Euclidean distance. This assignment directly provides the segmentation of each sequence based on the motion word indices. In order to learn meaningful representations, the decoder reconstructs the input sequences from the quantized patches, where each patch is replaced by its corresponding motion word." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.484, + 0.626 + ], + "angle": 0, + "content": "model skeleton motion, the embeddings are divided into non-overlapping patches to capture local temporal structure over fixed intervals. During training, skeleton motion words are learned, and each patch is assigned to its closest motion word based on the learned representation. Therefore, it effectively captures both spatial and temporal patterns in skeleton sequences, allowing for more robust and coherent action segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.273, + 0.654 + ], + "angle": 0, + "content": "3.2.1. Temporal Patching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In order to achieve skeleton motion quantization, we start by patching the latent representations from the encoder into fixed-length short time series. Specifically, after reorganizing the per-joint embeddings \\(\\mathbf{Z}_{nv}\\) and concatenating joint features of the embedded skeleton to form \\(\\mathbf{Z}_{\\text{concat}} \\in \\mathbb{R}^{N \\times T \\times (V \\cdot D)}\\), we ensure that the sequence length \\(T\\) is divisible by the patch size \\(P\\) by applying zero-padding as necessary. We then partition the sequence into \\(M = \\frac{T}{P}\\) non-overlapping temporal patches, yielding a patched tensor \\(\\mathbf{Z}_p \\in \\mathbb{R}^{N \\times M \\times P \\times (V \\cdot D)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This structured patching allows us to measure similarity based on sequences of embedding vectors rather than individual vectors, accommodating greater temporal variability. This enables the learning of representative skeleton motion words that encapsulate temporal dynamics, rather than relying on single-frame representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.652, + 0.52 + ], + "angle": 0, + "content": "3.2.2. Quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Based on the temporal patching \\(\\mathbf{Z}_p\\), the network learns a patch-based codebook \\(\\mathcal{C} = \\{\\mathbf{c}_k\\}_{k=1}^K\\), where each codebook element \\(\\mathbf{c}_k \\in \\mathbb{R}^{P \\times (V \\cdot D)}\\). As before, \\(P\\) denotes the patch size, \\(V\\) the number of joints, and \\(D\\) the latent dimension. Each element \\(\\mathbf{c}_k\\) in the codebook defines what we refer to as a skeleton motion word, which serves as a discrete representation of skeleton motion over a temporal patch. The codebook \\(\\mathcal{C}\\) is initialized randomly using Kaiming uniform initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.751 + ], + "angle": 0, + "content": "During quantization, each patch in the batch \\(\\mathbf{p}_i\\in \\mathbf{Z}_p\\) is assigned to its nearest motion word in the codebook based on the Euclidean distance. Each patch is then replaced with its closest motion word \\(\\mathbf{c}_{k_i}\\), where \\(\\mathbf{c}_{k_i}\\) denotes the codebook element assigned to \\(\\mathbf{p}_i\\), yielding the quantized patch \\(\\mathbf{q}_i\\in \\mathbb{R}^{P\\times (V\\cdot D)}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.76, + 0.907, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {i} = \\mathbf {c} _ {k _ {i}} \\quad \\text {w i t h} \\quad k _ {i} = \\underset {k} {\\arg \\min } \\| \\mathbf {p} _ {i} - \\mathbf {c} _ {k} \\| _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.794, + 0.909, + 0.902 + ], + "angle": 0, + "content": "This produces the quantized tensor \\(\\mathbf{Q}_p\\in \\mathbb{R}^{N\\times M\\times P\\times (V\\cdot D)}\\). We then depatchify the quantized tensor \\(\\mathbf{Q}_p\\) to obtain \\(\\mathbf{Q}\\in \\mathbb{R}^{N\\times T\\times (V\\cdot D)}\\). From the assignment of a patch \\(\\mathbf{p}_i\\) to the motion word \\(\\mathbf{c}_{k_i}\\), we directly derive the segmentation \\(\\mathbf{Y}\\in \\mathbb{N}^{N\\times T}\\) of all sequences in the batch, where each element \\(\\mathbf{Y}_{nt}\\in \\{1,2,\\dots ,K\\}\\) denotes the index of the assigned motion word at the \\(t\\)-th frame of the \\(n\\)-th sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12104" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.264, + 0.106 + ], + "angle": 0, + "content": "3.2.3. Codebook Update" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.111, + 0.485, + 0.158 + ], + "angle": 0, + "content": "Following the quantization process, where patches are assigned to their nearest motion words, we update the motion words by averaging the assigned patches:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.167, + 0.483, + 0.205 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {k} = \\frac {1}{| \\mathcal {P} _ {k} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.485, + 0.279 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_k = \\{\\mathbf{p}_i\\mid k_i = k\\}\\) denotes the set of patches assigned to the \\(k\\) -th motion word. To enhance the stability of codebook updates within the SMQ framework, we use Exponential Moving Average (EMA) updates [57]:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.287, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {k} ^ {(t)} \\leftarrow \\alpha \\mathbf {c} _ {k} ^ {(t - 1)} + (1 - \\alpha) \\tilde {\\mathbf {c}} _ {k} ^ {(t)}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.312, + 0.483, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {c}} _ {k} ^ {(t)} = \\frac {1}{| \\mathcal {P} _ {k} ^ {(t)} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k} ^ {(t)}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.485, + 0.415 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_k^{(t)} = \\{\\mathbf{p}_i\\mid k_i = k\\}\\) is the set of patches assigned to the \\(k\\) -th motion word at iteration \\(t\\) , and \\(\\alpha = 0.5\\) is the EMA decay factor that controls the smoothness of the updates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.192, + 0.437 + ], + "angle": 0, + "content": "3.3.Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.584 + ], + "angle": 0, + "content": "To reconstruct the input skeleton sequences from the quantized latent representations, we design a decoder that mirrors the encoder's architecture. After quantization, the depatchified tensor \\(\\mathbf{Q}\\) is reshaped into \\(\\mathbf{Q}' \\in \\mathbb{R}^{(N \\cdot V) \\times D \\times T}\\) to obtain the joint independence. This reshaping allows the decoder to process each joint's temporal sequence individually, similar to the encoder's approach. The reconstructed skeleton, denoted as \\(\\hat{\\mathbf{X}} \\in \\mathbb{R}^{N \\times C \\times T \\times V}\\), is then obtained after reshaping." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.235, + 0.608 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.765 + ], + "angle": 0, + "content": "The network is trained end-to-end using two loss functions. The first term is the reconstruction loss that compares the reconstructed sequences \\(\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times C\\times T\\times V}\\) with the original sequences \\(\\mathbf{X}\\in \\mathbb{R}^{N\\times C\\times T\\times V}\\). The loss ensures that the encoder and codebook preserve as much information from the input sequences as possible. While the mean squared error (MSE) between \\(\\mathbf{X}\\) and \\(\\hat{\\mathbf{X}}\\) is an obvious choice, we propose to compute inter-joint distance MSE loss, defined over the distance from each joint to all other joints within the same frame:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.775, + 0.481, + 0.816 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e c}} = \\frac {1}{N \\cdot T \\cdot V ^ {2}} \\sum_ {n = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {v = 1} ^ {V} \\sum_ {w = 1} ^ {V} \\left(d X _ {n t v w} - d \\hat {X} _ {n t v w}\\right) ^ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.818, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} \\quad \\begin{array}{r l} d X _ {n t v w} & = \\left\\| \\mathbf {X} _ {n t v} - \\mathbf {X} _ {n t w} \\right\\| _ {2}, \\\\ d \\hat {X} _ {n t v w} & = \\left\\| \\hat {\\mathbf {X}} _ {n t v} - \\hat {\\mathbf {X}} _ {n t w} \\right\\| _ {2}. \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The loss has the advantage that it is inherently translation and rotation-invariant and only measures differences in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "poses. In our experiments, we show that the proposed loss performs better than a standard MSE loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.197 + ], + "angle": 0, + "content": "To ensure that the learned embeddings are separated into distinct skeleton motion words, we employ the commitment loss [57]. This loss encourages the encoder to produce latent embeddings that are close to their assigned motion words. The commitment loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.204, + 0.905, + 0.238 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o m m i t}} = \\sum_ {\\mathbf {p} _ {i} \\in \\mathbf {Z} _ {p}} \\| \\operatorname {s g} \\left[ \\mathbf {c} _ {k _ {i}} \\right] - \\mathbf {p} _ {i} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.906, + 0.307 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}_{k_i} \\in \\{\\mathbf{c}_1, \\dots, \\mathbf{c}_K\\}\\) is the motion word that the latent patch embedding \\(\\mathbf{p}_i\\) has been assigned to, and \\(\\mathrm{sg}[\\cdot]\\) denotes the stop-gradient operator, which stops the backpropagation of the gradient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.906, + 0.337 + ], + "angle": 0, + "content": "The total loss function combines both the reconstruction loss and the commitment loss" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.348, + 0.905, + 0.365 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = \\lambda L _ {\\text {r e c}} + L _ {\\text {c o m m i t}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.403 + ], + "angle": 0, + "content": "where \\(\\lambda = 0.001\\) is the hyperparameter that balances the loss terms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.417, + 0.646, + 0.433 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.441, + 0.616, + 0.456 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.523 + ], + "angle": 0, + "content": "We evaluated our method on three skeleton-based action datasets: HuGaDB [9], LARa [44], and BABEL [46]. Additional results on the PKU-MMD v2 dataset [37] are reported in the suppl. material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.599 + ], + "angle": 0, + "content": "HuGaDB consists of 364 trials capturing 10 lower limb activities, including walking, running, and sitting. Data were collected from 18 participants using six IMUs (3-axis accelerometer and 3-axis gyroscope) placed on the thighs, shins, and feet, recorded at \\(60\\mathrm{Hz}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.718 + ], + "angle": 0, + "content": "LARa comprises 439 trials across 8 warehouse-related action categories. Data were collected from 16 participants using marker-based MoCap (200 fps), recording the 3-axis position and orientation of 22 full body joints. The dataset includes 960 minutes of recordings. LARa is downsampled by a factor of 4, reducing the frame rate to 50 fps [14]. Additionally, the skeleton is centered from the root joint for translation invariance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The BABEL dataset contains 43 hours of 3D human motion sequences from AMASS [43], annotated with over 63,000 frame-level labels across more than 250 action categories. Following the protocol of [66], we extract 25 full-body joint skeletons from SMPL [40] parameters, construct three subsets, each focusing on four action categories, and downsample all sequences to 30 fps, after centering them on the root joint. In addition to the original protocol, we exclude sequences containing more than \\(50\\%\\) background actions to ensure that the models learn relevant actions. The evaluation metrics and implementation details are described in the suppl. material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12105" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.868, + 0.273 + ], + "angle": 0, + "content": "
MethodHuGaDBLARa
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
SupervisedTCN [28]88.3---56.861.5---20.0
ST-GCN [63]88.7---67.767.9---25.8
MS-TCN [12]86.8---89.965.8---39.6
MS-GCN [14]90.4---93.065.6---43.6
UnsupervisedCTE [26]33.84.70.60.60.523.316.88.15.22.3
CTE + Viterbi [26]39.221.713.29.57.523.017.76.83.71.6
TOT [27]33.83.10.70.50.421.47.85.32.71.1
TOT + Viterbi [27]33.820.815.610.57.532.617.711.67.43.2
ASOT [61]33.917.44.53.83.022.923.417.812.15.7
SMQ (ours)42.036.138.531.524.337.439.434.728.416.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.279, + 0.878, + 0.293 + ], + "angle": 0, + "content": "Table 1. Comparison to supervised and unsupervised temporal action segmentation methods on the HuGaDB and LARa datasets." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.3, + 0.907, + 0.416 + ], + "angle": 0, + "content": "
MethodBABEL Subset-1BABEL Subset-2BABEL Subset-3
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
CTE [26]34.828.625.017.59.540.330.617.812.27.431.413.18.25.83.6
CTE + Viterbi [26]30.936.223.215.27.342.430.724.319.512.831.230.920.715.28.4
TOT [27]31.818.714.27.64.435.412.813.78.64.331.57.14.92.91.7
TOT + Viterbi [27]29.129.331.520.89.935.336.835.930.019.834.033.831.326.817.9
ASOT [61]35.343.142.334.124.543.137.740.333.423.438.027.127.421.614.3
SMQ (ours)36.638.540.932.822.349.137.843.837.427.440.638.638.029.319.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.421, + 0.795, + 0.435 + ], + "angle": 0, + "content": "Table 2. Comparison to unsupervised temporal action segmentation methods on the BABEL dataset." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.461, + 0.365, + 0.477 + ], + "angle": 0, + "content": "4.2. Comparison to State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.576 + ], + "angle": 0, + "content": "We compare our proposed method, SMQ, with state-of-the-art unsupervised temporal action segmentation methods on the HuGaDB, LARa and BABEL datasets, as shown in Tables 1 and 2. Across all evaluation metrics, SMQ consistently outperforms unsupervised temporal action segmentation approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.76 + ], + "angle": 0, + "content": "On the HuGaDB dataset in Table 1, SMQ achieves a MoF of 42.0 and Edit Score of 36.1, surpassing the best performing approach CTE [26] by 2.8 and 14.4, respectively. For the F1@50 metric, our method has the score of 24.3, which is by 16.8 higher than CTE [26] and TOT [27]. On the LARa dataset, SMQ achieves a MoF of 37.4, Edit Score of 39.4, and F1@50 of 16.4, outperforming the best approach ASOT [61] by a large margin as well. For completeness, we also include results from supervised approaches, which reported only MoF and F1@50. While supervised approaches achieve much higher scores, unsupervised approaches do not use any annotations and solve thus a much harder task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.839 + ], + "angle": 0, + "content": "On the BABEL dataset in Table 2, SMQ achieves a higher MoF than ASOT [61] for Subset-1, but ASOT performs slightly better for the other metrics. Whereas on Subset-2 and Subset-3, SMQ outperforms all unsupervised temporal action segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fig. 3 shows a qualitative comparison between the methods. Both CTE and TOT generate too many small segments. They thus employ a post-processing technique utilizing Viterbi decoding [10, 25], to refine initial predictions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.581 + ], + "angle": 0, + "content": "and achieve temporally consistent action segments. The Viterbi decoding [25] computes the most probable sequence of actions by maximizing the joint likelihood under a first-order Markov model, using transition probabilities that penalize frequent label changes. While this improves segmental metrics by reducing fragmentation, it also imposes a strong prior that each action appears only once per sequence, which is not the case in the datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.854 + ], + "angle": 0, + "content": "ASOT can identify re-occurring actions and does not suffer from small segments. The approach, however, has the tendency to predict rather large segments. The small segments in the HuGaDB and LARa datasets are not discovered. Since the segments are longer in the BABEL subsets, ASOT performs better on this dataset. While SMQ does not generate perfect results, for instance it mixes standing up and sitting down in the example from the HuGaDB dataset, the results are better aligned with the ground truth. It is able to detect re-occurring actions, and it does not have the tendency to generate too many small segments, as CTE and TOT, or overly large segments as ASOT. Overall, the results show that SMQ is more effective in handling diverse segmentation scenarios. However, as can be seen in Fig. 3, identifying precise action transitions remains a challenge. Since actions are assigned at the patch level, transitions can only occur at patch boundaries. Improving the detection of action boundaries needs to be investigated in future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We also compare SMQ to self-supervised representation learning approaches for skeleton data [16, 35, 52]. These approaches are typically first trained in a self-supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.496, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.215, + 0.334, + 0.227 + ], + "angle": 0, + "content": "(a) HuGaDB" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.895, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.215, + 0.726, + 0.227 + ], + "angle": 0, + "content": "(b) LARa" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.23, + 0.493, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.354, + 0.352, + 0.365 + ], + "angle": 0, + "content": "(c) BABEL Subset-1" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.23, + 0.895, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.354, + 0.753, + 0.365 + ], + "angle": 0, + "content": "(d) BABEL Subset-3" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.371, + 0.729, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Qualitative results for unsupervised action segmentation algorithms." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.392, + 0.907, + 0.479 + ], + "angle": 0, + "content": "
MethodHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Predict&Cluster [52]39.331.533.224.218.318.828.519.212.95.432.232.132.224.314.2
AimCLR [16]34.730.630.923.711.322.028.022.115.67.740.329.732.525.715.5
ActCLR [35]28.221.610.46.24.123.330.623.716.77.832.633.232.824.413.4
SMQ (ours)42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.483, + 0.759, + 0.497 + ], + "angle": 0, + "content": "Table 3. Comparison to self-supervised skeleton-based representation learning methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.483, + 0.782 + ], + "angle": 0, + "content": "way and then fine-tuned for different downstream tasks like skeleton-based action recognition using annotated data. In order to apply these methods to unsupervised temporal action segmentation, we divided the input skeleton sequences into non-overlapping patches, consistent with the latent patching in SMQ. These patches were fed through the self-supervised representation learning approaches to learn patch-level action representations. We then performed k-means clustering over the patch-level representations. The results in Table 3 show that SMQ consistently outperforms existing self-supervised representation learning approaches on all the datasets. It needs to be noted that these approaches are designed for datasets with only one action per sequence, whereas unsupervised temporal action segmentation requires processing long, untrimmed skeleton sequences containing multiple actions per sequence. Compared to SMQ, they therefore learn less distinctive representations for this task as it is visualized in the suppl. material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.202, + 0.813 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.483, + 0.853 + ], + "angle": 0, + "content": "In this section, we analyze the impact of different design choices in our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Impact of Disentangled Embedding. As illustrated in Fig. 2, the encoder processes each joint independently, which ensures that the information of the different joints" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.678 + ], + "angle": 0, + "content": "remains disentangled in the embedding space. To evaluate the effectiveness of the disentangled embedding space, we conducted an ablation study. In Table 4, we compare our approach to a setup where we concatenate the joints before learning an embedding. In the latter case, the joints get entangled in the embedding and the structural information of the skeleton gets lost. The results demonstrate that the disentangled embedding enhances our model's performance, indicating that learning joint representations separately allows for more accurate action segmentation by enabling each joint to reside in distinct subspaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Impact of Patch Size. We apply temporal patching to the latent skeleton embeddings to obtain skeleton motion words. Specifically, we use a fixed patch size to create non-overlapping patches of the representations. To investigate the effect of patch size on the performance of our model, we conducted an ablation study with varying patch sizes, as presented in Table 5. Without patching, i.e., patch size is 1, we obtain a frame-wise representation. The results indicate that SMQ consistently outperforms the frameswise representation across all evaluated metrics, with significant improvements in segmental metrics. These findings emphasize that frame-wise quantization leads to over-segmentation, as it fails to capture the inherent temporal consistency of actions. In contrast, SMQ captures the tem" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12107" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.151 + ], + "angle": 0, + "content": "
EmbeddingHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Entangled38.934.235.627.918.926.635.929.522.812.340.132.134.927.116.9
Disentangled42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.155, + 0.782, + 0.171 + ], + "angle": 0, + "content": "Table 4. Impact of disentangled joint embedding on the HuGaDB, LARa, and BABEL datasets." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.175, + 0.454, + 0.298 + ], + "angle": 0, + "content": "
Patch SizeMoFEditF1@{10, 25, 50}
1 (Frame-wise)33.925.621.214.98.6
25 (0.5 sec)35.235.732.425.014.9
50 (1 sec)37.439.434.728.416.4
75 (1.5 sec)37.239.134.728.417.1
100 (2 sec)30.836.133.027.015.4
150 (3 sec)31.434.032.626.514.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.303, + 0.46, + 0.317 + ], + "angle": 0, + "content": "Table 5. Impact of motion patch size on the LARa dataset." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.326, + 0.483, + 0.422 + ], + "angle": 0, + "content": "
LcommitReconstruction LossMoFEditF1@{10,25,50}
✓X29.929.926.817.78.8
XInter-Joint Dist. MSE31.034.630.624.414.4
✓MSE34.639.434.527.415.7
✓Root Distance MSE28.034.930.223.814.1
✓Inter-Joint Dist. MSE37.439.434.728.416.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.426, + 0.472, + 0.44 + ], + "angle": 0, + "content": "Table 6. Impact of loss function variants on the LARa dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.809 + ], + "angle": 0, + "content": "poral variability of actions more effectively by grouping contiguous latent embeddings into patches. This results in more semantically meaningful action clusters that yield temporally consistent segmentations. On the LARa dataset, a patch size of 50 performs best, which corresponds to one second of motion, as the data is downsampled to 50 fps. We keep this one-second patching consistent across all datasets. Impact of Loss Variants. To evaluate the effectiveness of different loss components, we conduct an ablation study examining various configurations of the commitment loss and reconstruction loss. We first evaluate three variants of the reconstruction loss in the last three rows of Table 6. The proposed inter-joint distance MSE loss performs better than the commonly used MSE loss since it is invariant to translation and rotation. The root distance MSE loss is obtained by computing the MSE over the distances of each joint to the root joint within the same frame. While this loss is also translation and rotation invariant, it performs even worse than the inter-joint variant. Furthermore, we evaluate the impact of the commitment loss and reconstruction loss by omitting one of them as presented in the first two rows of the Table 6. Removing either of the loss terms substantially decreases the performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Impact of Reconstruction Loss Weight. In our ablation study, we varied the reconstruction loss weight, \\(\\lambda\\) (7), to understand its effect on model performance. By adjusting \\(\\lambda\\), we explore the balance between the reconstruction and commitment loss in Table 7. Our model is not very sensitive to this parameter and performs well for \\(\\lambda = 0.001\\)." + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.177, + 0.848, + 0.269 + ], + "angle": 0, + "content": "
λMoFEditF1@{10, 25, 50}
0.135.938.433.426.915.6
0.0136.238.934.427.815.9
0.00137.439.434.728.416.4
0.000134.139.835.427.916.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.273, + 0.836, + 0.287 + ], + "angle": 0, + "content": "Table 7. Impact of \\( \\lambda \\) on the LARa dataset." + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.301, + 0.842, + 0.393 + ], + "angle": 0, + "content": "
αMoFEditF1@{10, 25, 50}
0.3534.839.435.027.615.5
0.5037.439.434.728.416.4
0.6535.538.533.927.215.8
0.8034.237.733.427.115.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.398, + 0.89, + 0.412 + ], + "angle": 0, + "content": "Table 8. Impact of EMA decay rate \\( \\left( \\alpha \\right) \\) on the LARa dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Impact of Decay Factor. To examine the role of the EMA decay factor \\(\\alpha\\) (3), we perform an ablation study by evaluating multiple \\(\\alpha\\) values in Table 8. The decay factor governs the balance between incorporating new patch information and maintaining historical motion word estimates, thereby influencing the smoothness and responsiveness of the updates. Our model performs best when \\(\\alpha\\) is set to 0.5. Additional ablation studies are provided in the suppl. material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.634, + 0.601 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduced SMQ, a method for unsupervised skeleton-based temporal action segmentation. It uses an autoencoder that preserves the skeleton structure by ensuring that the information of the different joints remains disentangled in the embedding space. The skeleton motion quantization module then patches the latent representations into short time series of fixed length and quantizes them into skeleton motion words. We performed an extensive evaluation and compared the proposed approach to state-of-the-art unsupervised temporal action segmentation approaches, as well as approaches for self-supervised skeleton-based representation learning. SMQ outperforms them by a large margin. Furthermore, we demonstrated the importance of learning the representation over patches instead of frames and the importance of an embedding space where the information of the different joints remains disentangled. As part of future work, it would be interesting to investigate a hierarchical version of SMQ and improve the detection of action boundaries." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12108" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work has been supported by the ERC Consolidator Grants FORHUE (101044724) and ActionContraThreat (816564), the Ministry of Culture and Science of the State of North Rhine-Westphalia (iBehave network, Programme \"Netzwerke 2021\"), and the TRA Modelling (University of Bonn) as part of the Excellence Strategy of the federal and state governments." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[1] Robert T Arn, Pradyumna Narayana, Tegan Emerson, Bruce A Draper, Michael Kirby, and Chris Peterson. Motion segmentation via generalized curvatures. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41 (12):2919-2932, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[2] Jernej Barbič, Alla Safonova, Jia-Yu Pan, Christos Faloutsos, Jessica K Hodgins, and Nancy S Pollard. Segmenting motion capture data into distinct behaviors. In Proceedings of Graphics Interface (GI), pages 185-194, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[3] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-68. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[4] Elena Belén Bueno-Benito, Biel Tura Vecino, and Mariella Dimiccoli. Leveraging triplet loss for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4921-4929, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[9] Roman Chereshnev and Attila Kertesz-Farkas. Hugadb: Human gait database for activity recognition from wearable inertial sensor networks. In Analysis of Images, Social Networks and Texts (AIST), pages 131-141. Springer, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3323-3332, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[12] Yazan Abu Farha and Juergen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3575-3584, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.904, + 0.318 + ], + "angle": 0, + "content": "[13] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[14] Benjamin Filtjens, Bart Vanrumste, and Peter Slaets. Skeleton-based action segmentation with multi-stage spatial-temporal graph convolutional neural networks. IEEE Transactions on Emerging Topics in Computing, 12(1):202-212, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[16] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 762-770, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[19] Babak Hosseini, Romain Montagne, and Barbara Hammer. Deep-aligned convolutional neural network for skeleton-based action recognition and segmentation. Data Science and Engineering (DSE), 5(2):126-139, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[20] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14024-14034, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Haoyu Ji, Bowen Chen, Xinglong Xu, Weihong Ren, Zhiyong Wang, and Honghai Liu. Language-assisted skeleton action understanding for skeleton-based temporal action segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12109" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[22] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[23] Björn Krüger, Anna Vögele, Tobias Willig, Angela Yao, Reinhard Klein, and Andreas Weber. Efficient unsupervised temporal segmentation of motion data. IEEE Transactions on Multimedia (T-MM), 19(4):797-812, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[24] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[25] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(4):765-779, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[26] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Juergen Gall. Unsupervised learning of action classes with continuous temporal embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12066-12074, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[27] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[28] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 156-165, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[29] Jun Li and Sinisa Todorovic. Action shuffle alternating learning for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12628-12636, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[30] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4741-4750, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[31] Shijie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45 (6):6647-6658, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[32] Yunheng Li, Zhongyu Li, Shanghua Gao, Qilong Wang, Qibin Hou, and Ming-Ming Cheng. A decoupled spatiotemporal framework for skeleton-based action segmentation. arXiv preprint arXiv:2312.05830, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[33] Yun-Heng Li, Kai-Yuan Liu, Sheng-Lan Liu, Lin Feng, and Hong Qiao. Involving distinguished temporal graph convolutional networks for skeleton-based temporal action segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):647-660, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[34] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.134 + ], + "angle": 0, + "content": "recognition. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2490-2498, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[35] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Action-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2363-2372, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.906, + 0.263 + ], + "angle": 0, + "content": "[36] Wei Lin, Anna Kukleva, Horst Possegger, Hilde Kuehne, and Horst Bischof. Taec: Unsupervised action segmentation with temporal-aware embedding and clustering. In Computer Vision Winter Workshop, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[37] Chunhui Liu, Yueyu Hu, Yanghao Li, Sijie Song, and Jiaying Liu. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv preprint arXiv:1703.07475, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[38] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10139-10149, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[39] Kaiyuan Liu, Yunheng Li, Yuanfeng Xu, Shuai Liu, and Shenglan Liu. Spatial focus attention for fine-grained skeleton-based action tasks. IEEE Signal Processing Letters, 29:1883-1887, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[41] Zijia Lu and Ehsan Elhamifar. FACT: Frame-action cross-attention temporal modeling for efficient supervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[42] Hao Ma, Zaiyue Yang, and Haoyang Liu. Fine-grained unsupervised temporal action segmentation and distributed representation for skeleton-based human motion analysis. IEEE Transactions on Cybernetics (T-CYB), 52(12):13411-13424, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[43] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5442-5451, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[44] Friedrich Niemann, Christopher Reining, Fernando Moya Rueda, Nilah Ravi Nair, Janine Anika Steffens, Gernot A Fink, and Michael Ten Hompel. Lara: Creating a dataset for human activity recognition in logistics using semantic attributes. Sensors, 20(15):4083, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[45] Giancarlo Paoletti, Jacopo Cavazza, Cigdem Bryan, and Alessio Del Bue. Unsupervised human action recognition with skeletal graph laplacian and self-supervised viewpoints invariance. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[46] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[47] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11225-11234, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[48] Jan Sedmidubsky, Petra Budikova, Vlastislav Dohnal, and Pavel Zezula. Motion words: a text-like representation of 3d skeleton sequences. In Proceedings of the European Conference on Information Retrieval (ECIR), pages 527-541. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[49] Fadime Sener and Angela Yao. Unsupervised learning and segmentation of complex activities from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8368-8376, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[50] Federico Spurio, Emad Bahrami, Gianpiero Francesca, and Juergen Gall. Hierarchical vector quantization for unsupervised action segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6996-7005, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[51] Lucas Stoffl, Andy Bonnetto, Stephane d'Ascoli, and Alexander Mathis. Elucidating the hierarchical nature of behavior with masked autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 106-125, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[52] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9631-9640, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[53] Chenwei Tan, Tao Sun, Talas Fu, Yuhan Wang, Minjie Xu, and Shenglan Liu. Hierarchical spatial-temporal network for skeleton-based temporal action segmentation. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 28-39. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[54] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Stga-net: Spatial-temporal graph attention network for skeleton-based temporal action segmentation. In Proceedings of the IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 218-223. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[55] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Spatial-temporal graph transformer network for skeleton-based temporal action segmentation. Multimedia Tools and Applications (MTAP), 83(15):44273-44297, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[56] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6426-6436, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[58] Anna Vögele, Björn Krüger, and Reinhard Klein. Efficient unsupervised temporal segmentation of human motion. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), pages 167-176, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[59] Qifei Wang, Gregorij Kurillo, Ferda Ofli, and Ruzena Bajcsy. Unsupervised temporal segmentation of repetitive human actions based on kinematic modeling and frequency analysis. In Proceedings of the IEEE International Conference on 3D Vision (3DV), pages 562-570. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[60] Leiyang Xu, Qiang Wang, Xiaotian Lin, and Lin Yuan. An efficient framework for few-shot skeleton-based temporal action segmentation. Proceedings of the European Conference on Computer Vision (ECCV), 232:103707, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[61] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14618-14627, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[62] Ziwei Xu, Xudong Shen, Yongkang Wong, and Mohan S. Kankanhalli. Unsupervised motion representation learning with capsule autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 34:3205-3217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[63] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[64] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13679-13690, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[65] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASFormer: Transformer for action segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[66] Qing Yu and Kent Fujiwara. Frame-level label refinement for skeleton-based weakly-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3322-3330, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[67] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 36-51. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[68] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.905, + 0.84 + ], + "angle": 0, + "content": "[69] Feng Zhou, Fernando De la Torre, and Jessica K. Hodgins. Aligned cluster analysis for temporal segmentation of human motion. In Proceedings of the IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-7, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.842, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[70] Feng Zhou, Fernando De la Torre, and Jessica K Hodgins. Hierarchical aligned cluster analysis for temporal clustering of human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(3):582-596, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12111" + } + ] +] \ No newline at end of file diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_origin.pdf b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8f596811803d37a6a3c9b1d68a7d79c6021a2e6f --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/d38bb523-583d-4f38-b51f-72d4fb904420_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84485245245ed936e937ef70aba8dd0d4879320655171b1dba81f76588b20d0f +size 794872 diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/full.md b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eca078254aca148eece4753f8ec658cad505cc30 --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/full.md @@ -0,0 +1,323 @@ +# Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation + +Uzay Gokay $^{1}$ Federico Spurio $^{1,3}$ Dominik R. Bach $^{1,2}$ Juergen Gall $^{1,3}$ $^{1}$ University of Bonn, Germany $^{2}$ University College London, UK + $^{3}$ Lamarr Institute for Machine Learning and Artificial Intelligence, Germany {u.goekay, d.bach}@uni-bonn.de {fspurio, gall}@iai.uni-bonn.de + +# Abstract + +Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods. Code is available at github.com/bachlab/SMQ. + +# 1. Introduction + +Advancements in motion capture technology and pose estimation have enabled the precise recording of human motion, represented as time series of joint positions, orientations, or other representations of skeleton sequences. Recently, there has been an increased interest in temporally segmenting long skeleton sequences into action segments [14, 19, 21, 32, 53-55, 60]. These approaches, however, are predominantly supervised, requiring extensive labeled data for training. In contrast, prior unsupervised skeleton-based approaches [16, 30, 34, 35, 45, 52, 67] focus mainly on action recognition, assuming that the sequences are very short and contain only a single action per sequence. + +![](images/6ac97c4981373e8d051e7dd1a45bfd017e3e8c6df02acf97b09241af91c08df4.jpg) +Figure 1. Our unsupervised skeleton-based temporal action segmentation framework, Skeleton Motion Quantization (SMQ), takes a set of untrimmed skeleton sequences as input (top) and discovers actions consistent across all sequences (bottom). This is achieved by jointly segmenting and clustering all sequences. The letters A, B, C, and D correspond to the identified actions. + +This highly constrained setting limits their applicability to real-world scenarios, where multiple actions are performed within a single sequence. + +In the context of RGB videos, unsupervised approaches for temporal action segmentation have been proposed [26, 27, 29, 36, 50, 56, 61]. These methods discover actions in a set of long untrimmed video sequences by identifying actions across sequences and multiple actions within each sequence. While these approaches achieved promising results on video data, they have not been evaluated on skeleton sequences. Our investigations reveal that such methods either struggle to recognize re-occurring actions due to post-processing with Viterbi decoding or produce overly small segments, as visualized in qualitative results. + +In this paper, we thus propose a novel approach for unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Our model uses a dilated temporal convolution based sequence-to-sequence autoencoder, which preserves the skeleton structure by ensuring that the information of the different joints remains disen + +tangled in the embedding space. We will show that this joint-based disentanglement is essential for improving performance. Subsequently, the latent representations generated by the temporal encoder are then patchified into nonoverlapping sub-sequences of motion, which are quantized by a skeleton motion quantization module. Specifically, each latent motion patch is assigned to the closest cluster prototype, and these prototypes are learned during training with the aim of capturing the most representative skeleton motion words of each action in the dataset. We term the approach thus Skeleton Motion Quantization (SMQ). + +To assess the effectiveness of SMQ, we evaluated it on three skeleton-based datasets, namely HuGaDB [9], LARa [44], and BABEL [46]. We demonstrate that it outperforms state-of-the-art approaches for unsupervised temporal action segmentation by a large margin, which highlights the importance of developing methods that exploit the inherent structure of skeleton sequences. We also show that it outperforms existing skeleton-based representation learning approaches on the task of unsupervised skeleton-based temporal action segmentation. In summary, our contributions are the following: + +1. We thoroughly evaluate existing unsupervised temporal action segmentation approaches for skeleton data and provide protocols for three datasets. +2. We introduce SMQ, a novel unsupervised framework for skeleton-based temporal action segmentation. It learns representative skeleton motion words in a joint-disentangled embedding space. +3. We demonstrate that SMQ outperforms existing unsupervised action segmentation methods, underscoring the importance of utilizing the structure of skeleton data. + +# 2. Related Work + +Human Motion Segmentation. Human motion segmentation aims to break down continuous motion into distinct, meaningful segments, often at a very fine granularity [2, 23, 48, 69], such as phases of a gait cycle. In contrast, temporal action segmentation targets broader, higher-level actions over long, untrimmed sequences [10]. One of the earliest approaches for human motion segmentation was introduced by Barbic et al. [2], using a probabilistic PCA-based method to detect significant distribution changes in human poses. Other early methods include hierarchical clustering and graph-based techniques to align sequences and segment motion based on structural similarity [23, 58, 69, 70]. Recent methods for human motion segmentation use heuristics such as extended Douglas-Peucker keypoint detection [42], zero-velocity crossing with adaptive clustering [59], and curvature analysis on kinematic data [1] to isolate motion primitives within individual sequences. While these approaches focus on fine-grained motion patterns, there are recent methods [4, 11, 47] that aim to + +identify complex, high-level actions within individual video sequences. However, they are still limited to segmenting a single sequence at a time. Our approach, in contrast, identifies globally learned actions across multiple sequences and subjects. + +Temporal Action Segmentation. Temporal action segmentation involves segmenting long untrimmed videos, often containing multiple actions per video, into contiguous segments, each labeled with a specific action [3, 12, 20, 24, 28, 38, 41]. Unlike traditional action recognition, which aims to classify short videos containing a single action, temporal action segmentation addresses the sequential nature of actions. To avoid an expensive annotation process, unsupervised methods [26, 27, 29, 36, 49, 50, 56, 61] have been proposed for this task. Specifically, Mallow [49] is a two-stage iterative approach, combining discriminative learning from visual features with generative modeling of temporal structure using a Generalized Mallows Model. CTE [26] aims to learn continuous temporal embeddings with the supervision of the relative time-stamps and then clusters the learned features in a two step fashion. TOT [27] proposed a joint representation learning and online clustering framework which applies temporal optimal transport on the features from the encoder and cluster prototypes to obtain pseudo action labels for training. UFSA [56] combines the encoder from ASFormer [65] and the decoder from UVAST [3], allowing the use of segment-level cues. Lastly, ASOT [61] utilizes temporally consistent unbalanced optimal transport, relaxing the balanced assignment assumption in TOT. Our results demonstrate that despite progress in unsupervised temporal action segmentation, existing methods do not perform well on skeleton data, as they do not exploit the inherent structure of skeleton data. + +Skeleton-based Temporal Action Segmentation. Specifically for skeleton data, a variety of fully supervised temporal action segmentation methods have been proposed. [14, 54, 60, 63] combine GCNs [22] and TCNs [12, 28, 31] to capture both the spatial and temporal aspects of these sequences. Additionally, some skeleton-based temporal action segmentation approaches further investigate the separation [33] and decoupling [32] of spatial and temporal information. Various attention mechanisms [39, 54] have been proposed to capture spatio-temporal dependencies while other approaches emphasize distinct frameworks, including action synthesis [60] and language priors [21]. However, all of these methods rely on full supervision, requiring extensive labeling. In this work, we propose a skeleton-based temporal action segmentation approach that automatically learns holistic, globally consistent actions across the entire dataset in an unsupervised manner. + +Self-Supervised Representation Learning for Skeleton Data. Self-supervised representation learning [5-8, 13, 15, 17, 18] aims to learn feature representations with + +out human annotations by solving pretext tasks that exploit the structure of unlabeled data. Recent methods have explored this paradigm for skeleton-based action recognition [16, 30, 34, 35, 52, 62, 67, 68], focusing on short, trimmed sequences containing a single action. In particular, CrosSCLR [30], AimCLR [16], and ActCLR [35] adopt contrastive learning to extract discriminative representations. However, these methods solely learn representations without predicting action labels during training, operating on short, single-action sequences. They thus require annotations for downstream evaluation and are not applicable to long, untrimmed sequences. + +hBehaveMAE [51] uses a hierarchical framework with a masked autoencoder to learn interpretable latent representations of actions at varying levels of granularity. LAC [64] utilizes pre-trained visual encoders with the task of learning composable actions from synthesized data. The model is then fine-tuned for different tasks including action segmentation to evaluate the learned representation. Unlike self-supervised approaches that learn action-related representations without assigning actions during training, and therefore require annotated data for downstream fine-tuning, our method discovers and segments actions directly during training in a fully unsupervised and label-free manner. + +# 3. Unsupervised Skeleton-based Temporal Action Segmentation + +In this work, we address the task of unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Given input skeleton sequences encoding joint positions, orientations, or related features, the objective of skeleton-based action segmentation is to map each frame to one of the predefined action categories. In the unsupervised case, the actions are unknown and need to be discovered from the data. Formally, the input skeleton sequences are defined as $\mathbf{X} \in \mathbb{R}^{N \times C \times T \times V}$ , where $N$ is the number of sequences or batch size, $C$ is the number of feature channels representing a single joint, $T$ is the sequence length, and $V$ is the number of joints. The goal is to discover $K$ semantically meaningful clusters and to generate an output label sequence $\mathbf{Y} \in \mathbb{N}^{N \times T}$ , where each element $\mathbf{Y}_{nt} \in \{1,2,\dots,K\}$ corresponds to the cluster assigned to the $t$ -th frame of the $n$ -th sequence. The main challenge is to discover actions across sequences and subjects, since it requires dealing with a large variability in how actions can be executed by different individuals. + +While the task has been addressed for video data, it has not been studied for skeleton data. In our experiments, we show that state-of-the-art unsupervised approaches for video-based temporal action segmentation do not perform well for skeleton data since they do not exploit the nature of skeleton data. We therefore propose an approach for unsupervised skeleton-based temporal action segmentation. + +The approach is illustrated in Fig. 2. Each joint of a sequence will be mapped to an embedding space by an encoder, which will be described in Section 3.1. The encoder learns an embedding for each joint separately, which ensures that the information of the different joints remains disentangled in the embedding space and that the representation is not dominated by a subset of joints. Another important aspect of the approach is the conversion of a frame-wise representation into a patch-based representation, which is then quantized. We term this approach skeleton motion quantization (SMQ) and describe it in Section 3.2. Finally, the decoder reconstructs the original skeleton sequence from the quantized patches as described in Section 3.3. While the end-to-end training is described in Section 3.4, the following sections provide a detailed description of each component. + +# 3.1.Encoder + +In our approach, we process input skeleton sequences to capture joint-specific temporal dynamics effectively. To learn latent subspaces for each joint individually, we reshape and permute $\mathbf{X}$ into $\mathbf{X}' \in \mathbb{R}^{(N \cdot V) \times C \times T}$ . This transformation treats each joint's temporal sequence $\mathbf{X}_{nv}' \in \mathbb{R}^{C \times T}$ as an independent sample, allowing the encoder to capture fine-grained motion patterns unique to each joint. Specifically, the encoder is a variant of a Temporal Convolutional Network (TCN) [12, 31], which consists of a series of dilated residual layers that progressively refine the temporal features of $\mathbf{X}'$ . Each stage begins with a $1 \times 1$ convolution to project the input features into a higher-dimensional space, followed by several residual blocks with dilated convolutions to capture long-range temporal dependencies. The dilation factor in these residual layers increases exponentially, allowing the network to model temporal patterns at multiple scales. Finally, a $1 \times 1$ convolutional layer maps the features to the latent dimension $D$ , yielding latent embeddings $\mathbf{Z}_{nv} \in \mathbb{R}^{D \times T}$ . + +# 3.2. Skeleton Motion Words + +Since skeleton sequences represent joint positions, orientations, or higher-order representations as time series, we reformulate the unsupervised skeleton-based temporal action segmentation task as an online time series clustering problem [57]. In this formulation, a skeleton sequence is mapped to a sequence of skeleton motion words which are learnable discrete units representing prototypical short-term temporal skeleton motion patterns. This allows us to capture the dynamic and temporal dependencies inherent in human motion data, aligning well with the nature of sequential joint movements across time. + +As shown in Fig. 2, the skeletons are first mapped into an embedding space using a temporal encoder, while keeping the information of the different joints disentangled. To + +![](images/8433902983024d70186dcbf6a1943c935008ce5ce320f8c351c6aa154d7eaf16.jpg) +Figure 2. Overview of SMQ. Given a set of skeleton sequences, our approach discovers actions that are semantically consistent across all sequences. The discovered actions correspond to clusters of learned motion words in the codebook. The model first encodes the skeleton sequences into a joint-based disentangled embedding space. Each embedded sequence is then divided into short non-overlapping temporal patches and a codebook with $K$ motion words is learned for the entire dataset via a patch-based quantization process, which assigns each patch to its nearest motion word using Euclidean distance. This assignment directly provides the segmentation of each sequence based on the motion word indices. In order to learn meaningful representations, the decoder reconstructs the input sequences from the quantized patches, where each patch is replaced by its corresponding motion word. + +model skeleton motion, the embeddings are divided into non-overlapping patches to capture local temporal structure over fixed intervals. During training, skeleton motion words are learned, and each patch is assigned to its closest motion word based on the learned representation. Therefore, it effectively captures both spatial and temporal patterns in skeleton sequences, allowing for more robust and coherent action segmentation. + +# 3.2.1. Temporal Patching + +In order to achieve skeleton motion quantization, we start by patching the latent representations from the encoder into fixed-length short time series. Specifically, after reorganizing the per-joint embeddings $\mathbf{Z}_{nv}$ and concatenating joint features of the embedded skeleton to form $\mathbf{Z}_{\text{concat}} \in \mathbb{R}^{N \times T \times (V \cdot D)}$ , we ensure that the sequence length $T$ is divisible by the patch size $P$ by applying zero-padding as necessary. We then partition the sequence into $M = \frac{T}{P}$ non-overlapping temporal patches, yielding a patched tensor $\mathbf{Z}_p \in \mathbb{R}^{N \times M \times P \times (V \cdot D)}$ . + +This structured patching allows us to measure similarity based on sequences of embedding vectors rather than individual vectors, accommodating greater temporal variability. This enables the learning of representative skeleton motion words that encapsulate temporal dynamics, rather than relying on single-frame representations. + +# 3.2.2. Quantization + +Based on the temporal patching $\mathbf{Z}_p$ , the network learns a patch-based codebook $\mathcal{C} = \{\mathbf{c}_k\}_{k=1}^K$ , where each codebook element $\mathbf{c}_k \in \mathbb{R}^{P \times (V \cdot D)}$ . As before, $P$ denotes the patch size, $V$ the number of joints, and $D$ the latent dimension. Each element $\mathbf{c}_k$ in the codebook defines what we refer to as a skeleton motion word, which serves as a discrete representation of skeleton motion over a temporal patch. The codebook $\mathcal{C}$ is initialized randomly using Kaiming uniform initialization. + +During quantization, each patch in the batch $\mathbf{p}_i\in \mathbf{Z}_p$ is assigned to its nearest motion word in the codebook based on the Euclidean distance. Each patch is then replaced with its closest motion word $\mathbf{c}_{k_i}$ , where $\mathbf{c}_{k_i}$ denotes the codebook element assigned to $\mathbf{p}_i$ , yielding the quantized patch $\mathbf{q}_i\in \mathbb{R}^{P\times (V\cdot D)}$ , i.e., + +$$ +\mathbf {q} _ {i} = \mathbf {c} _ {k _ {i}} \quad \text {w i t h} \quad k _ {i} = \underset {k} {\arg \min } \| \mathbf {p} _ {i} - \mathbf {c} _ {k} \| _ {2}. \tag {1} +$$ + +This produces the quantized tensor $\mathbf{Q}_p\in \mathbb{R}^{N\times M\times P\times (V\cdot D)}$ . We then depatchify the quantized tensor $\mathbf{Q}_p$ to obtain $\mathbf{Q}\in \mathbb{R}^{N\times T\times (V\cdot D)}$ . From the assignment of a patch $\mathbf{p}_i$ to the motion word $\mathbf{c}_{k_i}$ , we directly derive the segmentation $\mathbf{Y}\in \mathbb{N}^{N\times T}$ of all sequences in the batch, where each element $\mathbf{Y}_{nt}\in \{1,2,\dots ,K\}$ denotes the index of the assigned motion word at the $t$ -th frame of the $n$ -th sequence. + +# 3.2.3. Codebook Update + +Following the quantization process, where patches are assigned to their nearest motion words, we update the motion words by averaging the assigned patches: + +$$ +\mathbf {c} _ {k} = \frac {1}{| \mathcal {P} _ {k} |} \sum_ {\mathbf {p} _ {i} \in \mathcal {P} _ {k}} \mathbf {p} _ {i}, \quad \forall k \in \{1, \dots , K \}, \tag {2} +$$ + +where $\mathcal{P}_k = \{\mathbf{p}_i\mid k_i = k\}$ denotes the set of patches assigned to the $k$ -th motion word. To enhance the stability of codebook updates within the SMQ framework, we use Exponential Moving Average (EMA) updates [57]: + +$$ +\mathbf {c} _ {k} ^ {(t)} \leftarrow \alpha \mathbf {c} _ {k} ^ {(t - 1)} + (1 - \alpha) \tilde {\mathbf {c}} _ {k} ^ {(t)}, \tag {3} +$$ + +$$ +\tilde {\mathbf {c}} _ {k} ^ {(t)} = \frac {1}{| \mathcal {P} _ {k} ^ {(t)} |} \sum_ {\mathbf {p} _ {i} \in \mathcal {P} _ {k} ^ {(t)}} \mathbf {p} _ {i}, \quad \forall k \in \{1, \dots , K \}, \tag {4} +$$ + +where $\mathcal{P}_k^{(t)} = \{\mathbf{p}_i\mid k_i = k\}$ is the set of patches assigned to the $k$ -th motion word at iteration $t$ , and $\alpha = 0.5$ is the EMA decay factor that controls the smoothness of the updates. + +# 3.3.Decoder + +To reconstruct the input skeleton sequences from the quantized latent representations, we design a decoder that mirrors the encoder's architecture. After quantization, the depatchified tensor $\mathbf{Q}$ is reshaped into $\mathbf{Q}' \in \mathbb{R}^{(N \cdot V) \times D \times T}$ to obtain the joint independence. This reshaping allows the decoder to process each joint's temporal sequence individually, similar to the encoder's approach. The reconstructed skeleton, denoted as $\hat{\mathbf{X}} \in \mathbb{R}^{N \times C \times T \times V}$ , is then obtained after reshaping. + +# 3.4. Training Loss + +The network is trained end-to-end using two loss functions. The first term is the reconstruction loss that compares the reconstructed sequences $\hat{\mathbf{X}}\in \mathbb{R}^{N\times C\times T\times V}$ with the original sequences $\mathbf{X}\in \mathbb{R}^{N\times C\times T\times V}$ . The loss ensures that the encoder and codebook preserve as much information from the input sequences as possible. While the mean squared error (MSE) between $\mathbf{X}$ and $\hat{\mathbf{X}}$ is an obvious choice, we propose to compute inter-joint distance MSE loss, defined over the distance from each joint to all other joints within the same frame: + +$$ +L _ {\text {r e c}} = \frac {1}{N \cdot T \cdot V ^ {2}} \sum_ {n = 1} ^ {N} \sum_ {t = 1} ^ {T} \sum_ {v = 1} ^ {V} \sum_ {w = 1} ^ {V} \left(d X _ {n t v w} - d \hat {X} _ {n t v w}\right) ^ {2} +$$ + +$$ +\text {w i t h} \quad \begin{array}{r l} d X _ {n t v w} & = \left\| \mathbf {X} _ {n t v} - \mathbf {X} _ {n t w} \right\| _ {2}, \\ d \hat {X} _ {n t v w} & = \left\| \hat {\mathbf {X}} _ {n t v} - \hat {\mathbf {X}} _ {n t w} \right\| _ {2}. \end{array} \tag {5} +$$ + +The loss has the advantage that it is inherently translation and rotation-invariant and only measures differences in + +poses. In our experiments, we show that the proposed loss performs better than a standard MSE loss. + +To ensure that the learned embeddings are separated into distinct skeleton motion words, we employ the commitment loss [57]. This loss encourages the encoder to produce latent embeddings that are close to their assigned motion words. The commitment loss is defined as + +$$ +L _ {\text {c o m m i t}} = \sum_ {\mathbf {p} _ {i} \in \mathbf {Z} _ {p}} \| \operatorname {s g} \left[ \mathbf {c} _ {k _ {i}} \right] - \mathbf {p} _ {i} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\mathbf{c}_{k_i} \in \{\mathbf{c}_1, \dots, \mathbf{c}_K\}$ is the motion word that the latent patch embedding $\mathbf{p}_i$ has been assigned to, and $\mathrm{sg}[\cdot]$ denotes the stop-gradient operator, which stops the backpropagation of the gradient. + +The total loss function combines both the reconstruction loss and the commitment loss + +$$ +L _ {\text {t o t a l}} = \lambda L _ {\text {r e c}} + L _ {\text {c o m m i t}}, \tag {7} +$$ + +where $\lambda = 0.001$ is the hyperparameter that balances the loss terms. + +# 4. Experiments + +# 4.1. Datasets + +We evaluated our method on three skeleton-based action datasets: HuGaDB [9], LARa [44], and BABEL [46]. Additional results on the PKU-MMD v2 dataset [37] are reported in the suppl. material. + +HuGaDB consists of 364 trials capturing 10 lower limb activities, including walking, running, and sitting. Data were collected from 18 participants using six IMUs (3-axis accelerometer and 3-axis gyroscope) placed on the thighs, shins, and feet, recorded at $60\mathrm{Hz}$ . + +LARa comprises 439 trials across 8 warehouse-related action categories. Data were collected from 16 participants using marker-based MoCap (200 fps), recording the 3-axis position and orientation of 22 full body joints. The dataset includes 960 minutes of recordings. LARa is downsampled by a factor of 4, reducing the frame rate to 50 fps [14]. Additionally, the skeleton is centered from the root joint for translation invariance. + +The BABEL dataset contains 43 hours of 3D human motion sequences from AMASS [43], annotated with over 63,000 frame-level labels across more than 250 action categories. Following the protocol of [66], we extract 25 full-body joint skeletons from SMPL [40] parameters, construct three subsets, each focusing on four action categories, and downsample all sequences to 30 fps, after centering them on the root joint. In addition to the original protocol, we exclude sequences containing more than $50\%$ background actions to ensure that the models learn relevant actions. The evaluation metrics and implementation details are described in the suppl. material. + +
MethodHuGaDBLARa
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
SupervisedTCN [28]88.3---56.861.5---20.0
ST-GCN [63]88.7---67.767.9---25.8
MS-TCN [12]86.8---89.965.8---39.6
MS-GCN [14]90.4---93.065.6---43.6
UnsupervisedCTE [26]33.84.70.60.60.523.316.88.15.22.3
CTE + Viterbi [26]39.221.713.29.57.523.017.76.83.71.6
TOT [27]33.83.10.70.50.421.47.85.32.71.1
TOT + Viterbi [27]33.820.815.610.57.532.617.711.67.43.2
ASOT [61]33.917.44.53.83.022.923.417.812.15.7
SMQ (ours)42.036.138.531.524.337.439.434.728.416.4
+ +Table 1. Comparison to supervised and unsupervised temporal action segmentation methods on the HuGaDB and LARa datasets. + +
MethodBABEL Subset-1BABEL Subset-2BABEL Subset-3
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
CTE [26]34.828.625.017.59.540.330.617.812.27.431.413.18.25.83.6
CTE + Viterbi [26]30.936.223.215.27.342.430.724.319.512.831.230.920.715.28.4
TOT [27]31.818.714.27.64.435.412.813.78.64.331.57.14.92.91.7
TOT + Viterbi [27]29.129.331.520.89.935.336.835.930.019.834.033.831.326.817.9
ASOT [61]35.343.142.334.124.543.137.740.333.423.438.027.127.421.614.3
SMQ (ours)36.638.540.932.822.349.137.843.837.427.440.638.638.029.319.3
+ +Table 2. Comparison to unsupervised temporal action segmentation methods on the BABEL dataset. + +# 4.2. Comparison to State of the Art + +We compare our proposed method, SMQ, with state-of-the-art unsupervised temporal action segmentation methods on the HuGaDB, LARa and BABEL datasets, as shown in Tables 1 and 2. Across all evaluation metrics, SMQ consistently outperforms unsupervised temporal action segmentation approaches. + +On the HuGaDB dataset in Table 1, SMQ achieves a MoF of 42.0 and Edit Score of 36.1, surpassing the best performing approach CTE [26] by 2.8 and 14.4, respectively. For the F1@50 metric, our method has the score of 24.3, which is by 16.8 higher than CTE [26] and TOT [27]. On the LARa dataset, SMQ achieves a MoF of 37.4, Edit Score of 39.4, and F1@50 of 16.4, outperforming the best approach ASOT [61] by a large margin as well. For completeness, we also include results from supervised approaches, which reported only MoF and F1@50. While supervised approaches achieve much higher scores, unsupervised approaches do not use any annotations and solve thus a much harder task. + +On the BABEL dataset in Table 2, SMQ achieves a higher MoF than ASOT [61] for Subset-1, but ASOT performs slightly better for the other metrics. Whereas on Subset-2 and Subset-3, SMQ outperforms all unsupervised temporal action segmentation methods. + +Fig. 3 shows a qualitative comparison between the methods. Both CTE and TOT generate too many small segments. They thus employ a post-processing technique utilizing Viterbi decoding [10, 25], to refine initial predictions + +and achieve temporally consistent action segments. The Viterbi decoding [25] computes the most probable sequence of actions by maximizing the joint likelihood under a first-order Markov model, using transition probabilities that penalize frequent label changes. While this improves segmental metrics by reducing fragmentation, it also imposes a strong prior that each action appears only once per sequence, which is not the case in the datasets. + +ASOT can identify re-occurring actions and does not suffer from small segments. The approach, however, has the tendency to predict rather large segments. The small segments in the HuGaDB and LARa datasets are not discovered. Since the segments are longer in the BABEL subsets, ASOT performs better on this dataset. While SMQ does not generate perfect results, for instance it mixes standing up and sitting down in the example from the HuGaDB dataset, the results are better aligned with the ground truth. It is able to detect re-occurring actions, and it does not have the tendency to generate too many small segments, as CTE and TOT, or overly large segments as ASOT. Overall, the results show that SMQ is more effective in handling diverse segmentation scenarios. However, as can be seen in Fig. 3, identifying precise action transitions remains a challenge. Since actions are assigned at the patch level, transitions can only occur at patch boundaries. Improving the detection of action boundaries needs to be investigated in future work. + +We also compare SMQ to self-supervised representation learning approaches for skeleton data [16, 35, 52]. These approaches are typically first trained in a self-supervised + +![](images/a144090599241001b4c78ad6d0f7fa0da08f0ba96db58e94ea859f4c335e1ccd.jpg) +(a) HuGaDB + +![](images/204e6da2416d2db82df422d8ae97130dc0311abb2dccca41870d299805429278.jpg) +(b) LARa + +![](images/b6c8fca9f9c9a2cea07211823538c2df5409a24d1c4a5729dfa8a1ab5e89375a.jpg) +(c) BABEL Subset-1 + +![](images/ecadcd828e68923d725761c7363ed820a2dddaa62a18a6b481803b89863b803d.jpg) +(d) BABEL Subset-3 +Figure 3. Qualitative results for unsupervised action segmentation algorithms. + +
MethodHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Predict&Cluster [52]39.331.533.224.218.318.828.519.212.95.432.232.132.224.314.2
AimCLR [16]34.730.630.923.711.322.028.022.115.67.740.329.732.525.715.5
ActCLR [35]28.221.610.46.24.123.330.623.716.77.832.633.232.824.413.4
SMQ (ours)42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
+ +Table 3. Comparison to self-supervised skeleton-based representation learning methods. + +way and then fine-tuned for different downstream tasks like skeleton-based action recognition using annotated data. In order to apply these methods to unsupervised temporal action segmentation, we divided the input skeleton sequences into non-overlapping patches, consistent with the latent patching in SMQ. These patches were fed through the self-supervised representation learning approaches to learn patch-level action representations. We then performed k-means clustering over the patch-level representations. The results in Table 3 show that SMQ consistently outperforms existing self-supervised representation learning approaches on all the datasets. It needs to be noted that these approaches are designed for datasets with only one action per sequence, whereas unsupervised temporal action segmentation requires processing long, untrimmed skeleton sequences containing multiple actions per sequence. Compared to SMQ, they therefore learn less distinctive representations for this task as it is visualized in the suppl. material. + +# 4.3. Ablations + +In this section, we analyze the impact of different design choices in our model. + +Impact of Disentangled Embedding. As illustrated in Fig. 2, the encoder processes each joint independently, which ensures that the information of the different joints + +remains disentangled in the embedding space. To evaluate the effectiveness of the disentangled embedding space, we conducted an ablation study. In Table 4, we compare our approach to a setup where we concatenate the joints before learning an embedding. In the latter case, the joints get entangled in the embedding and the structural information of the skeleton gets lost. The results demonstrate that the disentangled embedding enhances our model's performance, indicating that learning joint representations separately allows for more accurate action segmentation by enabling each joint to reside in distinct subspaces. + +Impact of Patch Size. We apply temporal patching to the latent skeleton embeddings to obtain skeleton motion words. Specifically, we use a fixed patch size to create non-overlapping patches of the representations. To investigate the effect of patch size on the performance of our model, we conducted an ablation study with varying patch sizes, as presented in Table 5. Without patching, i.e., patch size is 1, we obtain a frame-wise representation. The results indicate that SMQ consistently outperforms the frameswise representation across all evaluated metrics, with significant improvements in segmental metrics. These findings emphasize that frame-wise quantization leads to over-segmentation, as it fails to capture the inherent temporal consistency of actions. In contrast, SMQ captures the tem + +
EmbeddingHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Entangled38.934.235.627.918.926.635.929.522.812.340.132.134.927.116.9
Disentangled42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
+ +Table 4. Impact of disentangled joint embedding on the HuGaDB, LARa, and BABEL datasets. + +
Patch SizeMoFEditF1@{10, 25, 50}
1 (Frame-wise)33.925.621.214.98.6
25 (0.5 sec)35.235.732.425.014.9
50 (1 sec)37.439.434.728.416.4
75 (1.5 sec)37.239.134.728.417.1
100 (2 sec)30.836.133.027.015.4
150 (3 sec)31.434.032.626.514.7
+ +Table 5. Impact of motion patch size on the LARa dataset. + +
LcommitReconstruction LossMoFEditF1@{10,25,50}
✓X29.929.926.817.78.8
XInter-Joint Dist. MSE31.034.630.624.414.4
✓MSE34.639.434.527.415.7
✓Root Distance MSE28.034.930.223.814.1
✓Inter-Joint Dist. MSE37.439.434.728.416.4
+ +poral variability of actions more effectively by grouping contiguous latent embeddings into patches. This results in more semantically meaningful action clusters that yield temporally consistent segmentations. On the LARa dataset, a patch size of 50 performs best, which corresponds to one second of motion, as the data is downsampled to 50 fps. We keep this one-second patching consistent across all datasets. Impact of Loss Variants. To evaluate the effectiveness of different loss components, we conduct an ablation study examining various configurations of the commitment loss and reconstruction loss. We first evaluate three variants of the reconstruction loss in the last three rows of Table 6. The proposed inter-joint distance MSE loss performs better than the commonly used MSE loss since it is invariant to translation and rotation. The root distance MSE loss is obtained by computing the MSE over the distances of each joint to the root joint within the same frame. While this loss is also translation and rotation invariant, it performs even worse than the inter-joint variant. Furthermore, we evaluate the impact of the commitment loss and reconstruction loss by omitting one of them as presented in the first two rows of the Table 6. Removing either of the loss terms substantially decreases the performance. + +Impact of Reconstruction Loss Weight. In our ablation study, we varied the reconstruction loss weight, $\lambda$ (7), to understand its effect on model performance. By adjusting $\lambda$ , we explore the balance between the reconstruction and commitment loss in Table 7. Our model is not very sensitive to this parameter and performs well for $\lambda = 0.001$ . + +Table 6. Impact of loss function variants on the LARa dataset. + +
λMoFEditF1@{10, 25, 50}
0.135.938.433.426.915.6
0.0136.238.934.427.815.9
0.00137.439.434.728.416.4
0.000134.139.835.427.916.1
+ +Table 7. Impact of $\lambda$ on the LARa dataset. + +
αMoFEditF1@{10, 25, 50}
0.3534.839.435.027.615.5
0.5037.439.434.728.416.4
0.6535.538.533.927.215.8
0.8034.237.733.427.115.5
+ +Table 8. Impact of EMA decay rate $\left( \alpha \right)$ on the LARa dataset. + +Impact of Decay Factor. To examine the role of the EMA decay factor $\alpha$ (3), we perform an ablation study by evaluating multiple $\alpha$ values in Table 8. The decay factor governs the balance between incorporating new patch information and maintaining historical motion word estimates, thereby influencing the smoothness and responsiveness of the updates. Our model performs best when $\alpha$ is set to 0.5. Additional ablation studies are provided in the suppl. material. + +# 5. Conclusion + +In this paper, we introduced SMQ, a method for unsupervised skeleton-based temporal action segmentation. It uses an autoencoder that preserves the skeleton structure by ensuring that the information of the different joints remains disentangled in the embedding space. The skeleton motion quantization module then patches the latent representations into short time series of fixed length and quantizes them into skeleton motion words. We performed an extensive evaluation and compared the proposed approach to state-of-the-art unsupervised temporal action segmentation approaches, as well as approaches for self-supervised skeleton-based representation learning. SMQ outperforms them by a large margin. Furthermore, we demonstrated the importance of learning the representation over patches instead of frames and the importance of an embedding space where the information of the different joints remains disentangled. As part of future work, it would be interesting to investigate a hierarchical version of SMQ and improve the detection of action boundaries. + +# Acknowledgments + +This work has been supported by the ERC Consolidator Grants FORHUE (101044724) and ActionContraThreat (816564), the Ministry of Culture and Science of the State of North Rhine-Westphalia (iBehave network, Programme "Netzwerke 2021"), and the TRA Modelling (University of Bonn) as part of the Excellence Strategy of the federal and state governments. + +# References + +[1] Robert T Arn, Pradyumna Narayana, Tegan Emerson, Bruce A Draper, Michael Kirby, and Chris Peterson. Motion segmentation via generalized curvatures. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41 (12):2919-2932, 2018. 2 +[2] Jernej Barbič, Alla Safonova, Jia-Yu Pan, Christos Faloutsos, Jessica K Hodgins, and Nancy S Pollard. Segmenting motion capture data into distinct behaviors. In Proceedings of Graphics Interface (GI), pages 185-194, 2004. 2 +[3] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-68. Springer, 2022. 2 +[4] Elena Belén Bueno-Benito, Biel Tura Vecino, and Mariella Dimiccoli. Leveraging triplet loss for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4921-4929, 2023. 2 +[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 2 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021. +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2 +[9] Roman Chereshnev and Attila Kertesz-Farkas. Hugadb: Human gait database for activity recognition from wearable inertial sensor networks. In Analysis of Images, Social Networks and Texts (AIST), pages 131-141. Springer, 2018. 2, 5 +[10] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques. + +IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 6 +[11] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3323-3332, 2022. 2 +[12] Yazan Abu Farha and Juergen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3575-3584, 2019. 2, 3, 6 +[13] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2 +[14] Benjamin Filtjens, Bart Vanrumste, and Peter Slaets. Skeleton-based action segmentation with multi-stage spatial-temporal graph convolutional neural networks. IEEE Transactions on Emerging Topics in Computing, 12(1):202-212, 2022. 1, 2, 5, 6 +[15] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 33:21271-21284, 2020. 2 +[16] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 762-770, 2022. 1, 3, 6, 7 +[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 2 +[19] Babak Hosseini, Romain Montagne, and Barbara Hammer. Deep-aligned convolutional neural network for skeleton-based action recognition and segmentation. Data Science and Engineering (DSE), 5(2):126-139, 2020. 1 +[20] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14024-14034, 2020. 2 +[21] Haoyu Ji, Bowen Chen, Xinglong Xu, Weihong Ren, Zhiyong Wang, and Honghai Liu. Language-assisted skeleton action understanding for skeleton-based temporal action segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2 + +[22] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 2 +[23] Björn Krüger, Anna Vögele, Tobias Willig, Angela Yao, Reinhard Klein, and Andreas Weber. Efficient unsupervised temporal segmentation of motion data. IEEE Transactions on Multimedia (T-MM), 19(4):797-812, 2016. 2 +[24] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2 +[25] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(4):765-779, 2020. 6 +[26] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Juergen Gall. Unsupervised learning of action classes with continuous temporal embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12066-12074, 2019. 1, 2, 6 +[27] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 1, 2, 6 +[28] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 156-165, 2017. 2, 6 +[29] Jun Li and Sinisa Todorovic. Action shuffle alternating learning for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12628-12636, 2021. 1, 2 +[30] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4741-4750, 2021. 1, 3 +[31] Shijie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45 (6):6647-6658, 2023. 2, 3 +[32] Yunheng Li, Zhongyu Li, Shanghua Gao, Qilong Wang, Qibin Hou, and Ming-Ming Cheng. A decoupled spatiotemporal framework for skeleton-based action segmentation. arXiv preprint arXiv:2312.05830, 2023. 1, 2 +[33] Yun-Heng Li, Kai-Yuan Liu, Sheng-Lan Liu, Lin Feng, and Hong Qiao. Involving distinguished temporal graph convolutional networks for skeleton-based temporal action segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):647-660, 2023. 2 +[34] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action + +recognition. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2490-2498, 2020. 1, 3 +[35] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Action-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2363-2372, 2023. 1, 3, 6, 7 +[36] Wei Lin, Anna Kukleva, Horst Possegger, Hilde Kuehne, and Horst Bischof. Taec: Unsupervised action segmentation with temporal-aware embedding and clustering. In Computer Vision Winter Workshop, 2023. 1, 2 +[37] Chunhui Liu, Yueyu Hu, Yanghao Li, Sijie Song, and Jiaying Liu. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv preprint arXiv:1703.07475, 2017. 5 +[38] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10139-10149, 2023. 2 +[39] Kaiyuan Liu, Yunheng Li, Yuanfeng Xu, Shuai Liu, and Shenglan Liu. Spatial focus attention for fine-grained skeleton-based action tasks. IEEE Signal Processing Letters, 29:1883-1887, 2022. 2 +[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 5 +[41] Zijia Lu and Ehsan Elhamifar. FACT: Frame-action cross-attention temporal modeling for efficient supervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[42] Hao Ma, Zaiyue Yang, and Haoyang Liu. Fine-grained unsupervised temporal action segmentation and distributed representation for skeleton-based human motion analysis. IEEE Transactions on Cybernetics (T-CYB), 52(12):13411-13424, 2021. 2 +[43] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5442-5451, 2019. 5 +[44] Friedrich Niemann, Christopher Reining, Fernando Moya Rueda, Nilah Ravi Nair, Janine Anika Steffens, Gernot A Fink, and Michael Ten Hompel. Lara: Creating a dataset for human activity recognition in logistics using semantic attributes. Sensors, 20(15):4083, 2020. 2, 5 +[45] Giancarlo Paoletti, Jacopo Cavazza, Cigdem Bryan, and Alessio Del Bue. Unsupervised human action recognition with skeletal graph laplacian and self-supervised viewpoints invariance. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 1 +[46] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english + +labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2, 5 +[47] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11225-11234, 2021. 2 +[48] Jan Sedmidubsky, Petra Budikova, Vlastislav Dohnal, and Pavel Zezula. Motion words: a text-like representation of 3d skeleton sequences. In Proceedings of the European Conference on Information Retrieval (ECIR), pages 527-541. Springer, 2020. 2 +[49] Fadime Sener and Angela Yao. Unsupervised learning and segmentation of complex activities from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8368-8376, 2018. 2 +[50] Federico Spurio, Emad Bahrami, Gianpiero Francesca, and Juergen Gall. Hierarchical vector quantization for unsupervised action segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6996-7005, 2025. 1, 2 +[51] Lucas Stoffl, Andy Bonnetto, Stephane d'Ascoli, and Alexander Mathis. Elucidating the hierarchical nature of behavior with masked autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 106-125, 2024. 3 +[52] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9631-9640, 2020. 1, 3, 6, 7 +[53] Chenwei Tan, Tao Sun, Talas Fu, Yuhan Wang, Minjie Xu, and Shenglan Liu. Hierarchical spatial-temporal network for skeleton-based temporal action segmentation. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 28-39. Springer, 2023. 1 +[54] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Stga-net: Spatial-temporal graph attention network for skeleton-based temporal action segmentation. In Proceedings of the IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 218-223. IEEE, 2023. 2 +[55] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Spatial-temporal graph transformer network for skeleton-based temporal action segmentation. Multimedia Tools and Applications (MTAP), 83(15):44273-44297, 2024. 1 +[56] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6426-6436, 2024. 1, 2 +[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3, 5 +[58] Anna Vögele, Björn Krüger, and Reinhard Klein. Efficient unsupervised temporal segmentation of human motion. In + +Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), pages 167-176, 2014. 2 +[59] Qifei Wang, Gregorij Kurillo, Ferda Ofli, and Ruzena Bajcsy. Unsupervised temporal segmentation of repetitive human actions based on kinematic modeling and frequency analysis. In Proceedings of the IEEE International Conference on 3D Vision (3DV), pages 562-570. IEEE, 2015. 2 +[60] Leiyang Xu, Qiang Wang, Xiaotian Lin, and Lin Yuan. An efficient framework for few-shot skeleton-based temporal action segmentation. Proceedings of the European Conference on Computer Vision (ECCV), 232:103707, 2023. 1, 2 +[61] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14618-14627, 2024. 1, 2, 6 +[62] Ziwei Xu, Xudong Shen, Yongkang Wong, and Mohan S. Kankanhalli. Unsupervised motion representation learning with capsule autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 34:3205-3217, 2021. 3 +[63] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2, 6 +[64] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13679-13690, 2023. 3 +[65] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASFormer: Transformer for action segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 2 +[66] Qing Yu and Kent Fujiwara. Frame-level label refinement for skeleton-based weakly-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3322-3330, 2023. 5 +[67] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 36-51. Springer, 2022. 1, 3 +[68] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 3 +[69] Feng Zhou, Fernando De la Torre, and Jessica K. Hodgins. Aligned cluster analysis for temporal segmentation of human motion. In Proceedings of the IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-7, 2008. 2 +[70] Feng Zhou, Fernando De la Torre, and Jessica K Hodgins. Hierarchical aligned cluster analysis for temporal clustering of human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(3):582-596, 2012. 2 \ No newline at end of file diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/images.zip b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b84386cfb3e28f486bb87aa0161e5fde3f152737 --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba7ed1bac759f7ea91a102c572d40c68505b7c55ac27a6307cabf416a3306bcf +size 556107 diff --git a/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/layout.json b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3297e53afa807b609c5a31810c74a63cac37851f --- /dev/null +++ b/2025/Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation/layout.json @@ -0,0 +1,8784 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "spans": [ + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": "Uzay Gokay" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": " Federico Spurio" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": " Dominik R. Bach" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": " Juergen Gall" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": "University of Bonn, Germany " + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": "University College London, UK \n" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 120, + 160, + 490, + 219 + ], + "type": "text", + "content": "Lamarr Institute for Machine Learning and Artificial Intelligence, Germany {u.goekay, d.bach}@uni-bonn.de {fspurio, gall}@iai.uni-bonn.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 522 + ], + "type": "text", + "content": "Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods. Code is available at github.com/bachlab/SMQ." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Advancements in motion capture technology and pose estimation have enabled the precise recording of human motion, represented as time series of joint positions, orientations, or other representations of skeleton sequences. Recently, there has been an increased interest in temporally segmenting long skeleton sequences into action segments [14, 19, 21, 32, 53-55, 60]. These approaches, however, are predominantly supervised, requiring extensive labeled data for training. In contrast, prior unsupervised skeleton-based approaches [16, 30, 34, 35, 45, 52, 67] focus mainly on action recognition, assuming that the sequences are very short and contain only a single action per sequence." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 556, + 378 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 556, + 378 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 556, + 378 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 556, + 378 + ], + "type": "image", + "image_path": "6ac97c4981373e8d051e7dd1a45bfd017e3e8c6df02acf97b09241af91c08df4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 387, + 556, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 453 + ], + "type": "text", + "content": "Figure 1. Our unsupervised skeleton-based temporal action segmentation framework, Skeleton Motion Quantization (SMQ), takes a set of untrimmed skeleton sequences as input (top) and discovers actions consistent across all sequences (bottom). This is achieved by jointly segmenting and clustering all sequences. The letters A, B, C, and D correspond to the identified actions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 505 + ], + "type": "text", + "content": "This highly constrained setting limits their applicability to real-world scenarios, where multiple actions are performed within a single sequence." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "text", + "content": "In the context of RGB videos, unsupervised approaches for temporal action segmentation have been proposed [26, 27, 29, 36, 50, 56, 61]. These methods discover actions in a set of long untrimmed video sequences by identifying actions across sequences and multiple actions within each sequence. While these approaches achieved promising results on video data, they have not been evaluated on skeleton sequences. Our investigations reveal that such methods either struggle to recognize re-occurring actions due to post-processing with Viterbi decoding or produce overly small segments, as visualized in qualitative results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we thus propose a novel approach for unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Our model uses a dilated temporal convolution based sequence-to-sequence autoencoder, which preserves the skeleton structure by ensuring that the information of the different joints remains disen" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12101" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "tangled in the embedding space. We will show that this joint-based disentanglement is essential for improving performance. Subsequently, the latent representations generated by the temporal encoder are then patchified into nonoverlapping sub-sequences of motion, which are quantized by a skeleton motion quantization module. Specifically, each latent motion patch is assigned to the closest cluster prototype, and these prototypes are learned during training with the aim of capturing the most representative skeleton motion words of each action in the dataset. We term the approach thus Skeleton Motion Quantization (SMQ)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 335 + ], + "type": "text", + "content": "To assess the effectiveness of SMQ, we evaluated it on three skeleton-based datasets, namely HuGaDB [9], LARa [44], and BABEL [46]. We demonstrate that it outperforms state-of-the-art approaches for unsupervised temporal action segmentation by a large margin, which highlights the importance of developing methods that exploit the inherent structure of skeleton sequences. We also show that it outperforms existing skeleton-based representation learning approaches on the task of unsupervised skeleton-based temporal action segmentation. In summary, our contributions are the following:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 295, + 455 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 335, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 295, + 370 + ], + "type": "text", + "content": "1. We thoroughly evaluate existing unsupervised temporal action segmentation approaches for skeleton data and provide protocols for three datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 294, + 418 + ], + "type": "text", + "content": "2. We introduce SMQ, a novel unsupervised framework for skeleton-based temporal action segmentation. It learns representative skeleton motion words in a joint-disentangled embedding space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 294, + 455 + ], + "type": "text", + "content": "3. We demonstrate that SMQ outperforms existing unsupervised action segmentation methods, underscoring the importance of utilizing the structure of skeleton data." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Human Motion Segmentation. Human motion segmentation aims to break down continuous motion into distinct, meaningful segments, often at a very fine granularity [2, 23, 48, 69], such as phases of a gait cycle. In contrast, temporal action segmentation targets broader, higher-level actions over long, untrimmed sequences [10]. One of the earliest approaches for human motion segmentation was introduced by Barbic et al. [2], using a probabilistic PCA-based method to detect significant distribution changes in human poses. Other early methods include hierarchical clustering and graph-based techniques to align sequences and segment motion based on structural similarity [23, 58, 69, 70]. Recent methods for human motion segmentation use heuristics such as extended Douglas-Peucker keypoint detection [42], zero-velocity crossing with adaptive clustering [59], and curvature analysis on kinematic data [1] to isolate motion primitives within individual sequences. While these approaches focus on fine-grained motion patterns, there are recent methods [4, 11, 47] that aim to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "identify complex, high-level actions within individual video sequences. However, they are still limited to segmenting a single sequence at a time. Our approach, in contrast, identifies globally learned actions across multiple sequences and subjects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 135, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 469 + ], + "type": "text", + "content": "Temporal Action Segmentation. Temporal action segmentation involves segmenting long untrimmed videos, often containing multiple actions per video, into contiguous segments, each labeled with a specific action [3, 12, 20, 24, 28, 38, 41]. Unlike traditional action recognition, which aims to classify short videos containing a single action, temporal action segmentation addresses the sequential nature of actions. To avoid an expensive annotation process, unsupervised methods [26, 27, 29, 36, 49, 50, 56, 61] have been proposed for this task. Specifically, Mallow [49] is a two-stage iterative approach, combining discriminative learning from visual features with generative modeling of temporal structure using a Generalized Mallows Model. CTE [26] aims to learn continuous temporal embeddings with the supervision of the relative time-stamps and then clusters the learned features in a two step fashion. TOT [27] proposed a joint representation learning and online clustering framework which applies temporal optimal transport on the features from the encoder and cluster prototypes to obtain pseudo action labels for training. UFSA [56] combines the encoder from ASFormer [65] and the decoder from UVAST [3], allowing the use of segment-level cues. Lastly, ASOT [61] utilizes temporally consistent unbalanced optimal transport, relaxing the balanced assignment assumption in TOT. Our results demonstrate that despite progress in unsupervised temporal action segmentation, existing methods do not perform well on skeleton data, as they do not exploit the inherent structure of skeleton data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 553, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 674 + ], + "type": "text", + "content": "Skeleton-based Temporal Action Segmentation. Specifically for skeleton data, a variety of fully supervised temporal action segmentation methods have been proposed. [14, 54, 60, 63] combine GCNs [22] and TCNs [12, 28, 31] to capture both the spatial and temporal aspects of these sequences. Additionally, some skeleton-based temporal action segmentation approaches further investigate the separation [33] and decoupling [32] of spatial and temporal information. Various attention mechanisms [39, 54] have been proposed to capture spatio-temporal dependencies while other approaches emphasize distinct frameworks, including action synthesis [60] and language priors [21]. However, all of these methods rely on full supervision, requiring extensive labeling. In this work, we propose a skeleton-based temporal action segmentation approach that automatically learns holistic, globally consistent actions across the entire dataset in an unsupervised manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Self-Supervised Representation Learning for Skeleton Data. Self-supervised representation learning [5-8, 13, 15, 17, 18] aims to learn feature representations with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12102" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "out human annotations by solving pretext tasks that exploit the structure of unlabeled data. Recent methods have explored this paradigm for skeleton-based action recognition [16, 30, 34, 35, 52, 62, 67, 68], focusing on short, trimmed sequences containing a single action. In particular, CrosSCLR [30], AimCLR [16], and ActCLR [35] adopt contrastive learning to extract discriminative representations. However, these methods solely learn representations without predicting action labels during training, operating on short, single-action sequences. They thus require annotations for downstream evaluation and are not applicable to long, untrimmed sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 358 + ], + "type": "text", + "content": "hBehaveMAE [51] uses a hierarchical framework with a masked autoencoder to learn interpretable latent representations of actions at varying levels of granularity. LAC [64] utilizes pre-trained visual encoders with the task of learning composable actions from synthesized data. The model is then fine-tuned for different tasks including action segmentation to evaluate the learned representation. Unlike self-supervised approaches that learn action-related representations without assigning actions during training, and therefore require annotated data for downstream fine-tuning, our method discovers and segments actions directly during training in a fully unsupervised and label-free manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 368, + 294, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 294, + 396 + ], + "type": "text", + "content": "3. Unsupervised Skeleton-based Temporal Action Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": "In this work, we address the task of unsupervised skeleton-based temporal action segmentation, which is illustrated in Fig. 1. Given input skeleton sequences encoding joint positions, orientations, or related features, the objective of skeleton-based action segmentation is to map each frame to one of the predefined action categories. In the unsupervised case, the actions are unknown and need to be discovered from the data. Formally, the input skeleton sequences are defined as " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C \\times T \\times V}" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " is the number of sequences or batch size, " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " is the number of feature channels representing a single joint, " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " is the sequence length, and " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " is the number of joints. The goal is to discover " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " semantically meaningful clusters and to generate an output label sequence " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathbb{N}^{N \\times T}" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": ", where each element " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{nt} \\in \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": " corresponds to the cluster assigned to the " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": "-th frame of the " + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 403, + 295, + 628 + ], + "type": "text", + "content": "-th sequence. The main challenge is to discover actions across sequences and subjects, since it requires dealing with a large variability in how actions can be executed by different individuals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "While the task has been addressed for video data, it has not been studied for skeleton data. In our experiments, we show that state-of-the-art unsupervised approaches for video-based temporal action segmentation do not perform well for skeleton data since they do not exploit the nature of skeleton data. We therefore propose an approach for unsupervised skeleton-based temporal action segmentation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "The approach is illustrated in Fig. 2. Each joint of a sequence will be mapped to an embedding space by an encoder, which will be described in Section 3.1. The encoder learns an embedding for each joint separately, which ensures that the information of the different joints remains disentangled in the embedding space and that the representation is not dominated by a subset of joints. Another important aspect of the approach is the conversion of a frame-wise representation into a patch-based representation, which is then quantized. We term this approach skeleton motion quantization (SMQ) and describe it in Section 3.2. Finally, the decoder reconstructs the original skeleton sequence from the quantized patches as described in Section 3.3. While the end-to-end training is described in Section 3.4, the following sections provide a detailed description of each component." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 273, + 376, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 376, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 376, + 284 + ], + "type": "text", + "content": "3.1.Encoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": "In our approach, we process input skeleton sequences to capture joint-specific temporal dynamics effectively. To learn latent subspaces for each joint individually, we reshape and permute " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}' \\in \\mathbb{R}^{(N \\cdot V) \\times C \\times T}" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": ". This transformation treats each joint's temporal sequence " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{nv}' \\in \\mathbb{R}^{C \\times T}" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": " as an independent sample, allowing the encoder to capture fine-grained motion patterns unique to each joint. Specifically, the encoder is a variant of a Temporal Convolutional Network (TCN) [12, 31], which consists of a series of dilated residual layers that progressively refine the temporal features of " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}'" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": ". Each stage begins with a " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": " convolution to project the input features into a higher-dimensional space, followed by several residual blocks with dilated convolutions to capture long-range temporal dependencies. The dilation factor in these residual layers increases exponentially, allowing the network to model temporal patterns at multiple scales. Finally, a " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": " convolutional layer maps the features to the latent dimension " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": ", yielding latent embeddings " + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{nv} \\in \\mathbb{R}^{D \\times T}" + }, + { + "bbox": [ + 313, + 291, + 553, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 527, + 448, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 448, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 448, + 539 + ], + "type": "text", + "content": "3.2. Skeleton Motion Words" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 675 + ], + "type": "text", + "content": "Since skeleton sequences represent joint positions, orientations, or higher-order representations as time series, we reformulate the unsupervised skeleton-based temporal action segmentation task as an online time series clustering problem [57]. In this formulation, a skeleton sequence is mapped to a sequence of skeleton motion words which are learnable discrete units representing prototypical short-term temporal skeleton motion patterns. This allows us to capture the dynamic and temporal dependencies inherent in human motion data, aligning well with the nature of sequential joint movements across time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "As shown in Fig. 2, the skeletons are first mapped into an embedding space using a temporal encoder, while keeping the information of the different joints disentangled. To" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12103" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 532, + 305 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 532, + 305 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 532, + 305 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 532, + 305 + ], + "type": "image", + "image_path": "8433902983024d70186dcbf6a1943c935008ce5ce320f8c351c6aa154d7eaf16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 312, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 54, + 312, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 555, + 390 + ], + "type": "text", + "content": "Figure 2. Overview of SMQ. Given a set of skeleton sequences, our approach discovers actions that are semantically consistent across all sequences. The discovered actions correspond to clusters of learned motion words in the codebook. The model first encodes the skeleton sequences into a joint-based disentangled embedding space. Each embedded sequence is then divided into short non-overlapping temporal patches and a codebook with " + }, + { + "bbox": [ + 54, + 312, + 555, + 390 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 312, + 555, + 390 + ], + "type": "text", + "content": " motion words is learned for the entire dataset via a patch-based quantization process, which assigns each patch to its nearest motion word using Euclidean distance. This assignment directly provides the segmentation of each sequence based on the motion word indices. In order to learn meaningful representations, the decoder reconstructs the input sequences from the quantized patches, where each patch is replaced by its corresponding motion word." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 400, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 296, + 495 + ], + "type": "text", + "content": "model skeleton motion, the embeddings are divided into non-overlapping patches to capture local temporal structure over fixed intervals. During training, skeleton motion words are learned, and each patch is assigned to its closest motion word based on the learned representation. Therefore, it effectively captures both spatial and temporal patterns in skeleton sequences, allowing for more robust and coherent action segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 505, + 167, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 167, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 167, + 517 + ], + "type": "text", + "content": "3.2.1. Temporal Patching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "In order to achieve skeleton motion quantization, we start by patching the latent representations from the encoder into fixed-length short time series. Specifically, after reorganizing the per-joint embeddings " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{nv}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " and concatenating joint features of the embedded skeleton to form " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\text{concat}} \\in \\mathbb{R}^{N \\times T \\times (V \\cdot D)}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": ", we ensure that the sequence length " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " is divisible by the patch size " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " by applying zero-padding as necessary. We then partition the sequence into " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "M = \\frac{T}{P}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " non-overlapping temporal patches, yielding a patched tensor " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_p \\in \\mathbb{R}^{N \\times M \\times P \\times (V \\cdot D)}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "This structured patching allows us to measure similarity based on sequences of embedding vectors rather than individual vectors, accommodating greater temporal variability. This enables the learning of representative skeleton motion words that encapsulate temporal dynamics, rather than relying on single-frame representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 400, + 399, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 399, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 399, + 411 + ], + "type": "text", + "content": "3.2.2. Quantization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": "Based on the temporal patching " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_p" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": ", the network learns a patch-based codebook " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathbf{c}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": ", where each codebook element " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k \\in \\mathbb{R}^{P \\times (V \\cdot D)}" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": ". As before, " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": " denotes the patch size, " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": " the number of joints, and " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": " the latent dimension. Each element " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_k" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": " in the codebook defines what we refer to as a skeleton motion word, which serves as a discrete representation of skeleton motion over a temporal patch. The codebook " + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 415, + 555, + 521 + ], + "type": "text", + "content": " is initialized randomly using Kaiming uniform initialization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": "During quantization, each patch in the batch " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i\\in \\mathbf{Z}_p" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": " is assigned to its nearest motion word in the codebook based on the Euclidean distance. Each patch is then replaced with its closest motion word " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{k_i}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{k_i}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": " denotes the codebook element assigned to " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": ", yielding the quantized patch " + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i\\in \\mathbb{R}^{P\\times (V\\cdot D)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 594 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 601, + 555, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 601, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 341, + 601, + 555, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {i} = \\mathbf {c} _ {k _ {i}} \\quad \\text {w i t h} \\quad k _ {i} = \\underset {k} {\\arg \\min } \\| \\mathbf {p} _ {i} - \\mathbf {c} _ {k} \\| _ {2}. \\tag {1}", + "image_path": "9986d37679b2b704a7c10504011a6032d9fee5cd8b4f18b4ba1ccd1ccbee32c7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": "This produces the quantized tensor " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p\\in \\mathbb{R}^{N\\times M\\times P\\times (V\\cdot D)}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": ". We then depatchify the quantized tensor " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_p" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\in \\mathbb{R}^{N\\times T\\times (V\\cdot D)}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": ". From the assignment of a patch " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " to the motion word " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{k_i}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": ", we directly derive the segmentation " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\in \\mathbb{N}^{N\\times T}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " of all sequences in the batch, where each element " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{nt}\\in \\{1,2,\\dots ,K\\}" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": " denotes the index of the assigned motion word at the " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": "-th frame of the " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": "-th sequence." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12104" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 83 + ], + "type": "text", + "content": "3.2.3. Codebook Update" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 125 + ], + "type": "text", + "content": "Following the quantization process, where patches are assigned to their nearest motion words, we update the motion words by averaging the assigned patches:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 132, + 295, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 132, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 89, + 132, + 295, + 162 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {k} = \\frac {1}{| \\mathcal {P} _ {k} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {2}", + "image_path": "ed4b1b838c7419281e8bfa64185095e02e603860c4b886ebb8d0255784645237.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k = \\{\\mathbf{p}_i\\mid k_i = k\\}" + }, + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "text", + "content": " denotes the set of patches assigned to the " + }, + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 171, + 296, + 220 + ], + "type": "text", + "content": " -th motion word. To enhance the stability of codebook updates within the SMQ framework, we use Exponential Moving Average (EMA) updates [57]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 227, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 227, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 82, + 227, + 295, + 245 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {k} ^ {(t)} \\leftarrow \\alpha \\mathbf {c} _ {k} ^ {(t - 1)} + (1 - \\alpha) \\tilde {\\mathbf {c}} _ {k} ^ {(t)}, \\tag {3}", + "image_path": "01e4c2920c836a656674af0d05359ec18356cf5ef728094c9fee9baa0129a7e9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 247, + 295, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 247, + 295, + 279 + ], + "spans": [ + { + "bbox": [ + 83, + 247, + 295, + 279 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {c}} _ {k} ^ {(t)} = \\frac {1}{| \\mathcal {P} _ {k} ^ {(t)} |} \\sum_ {\\mathbf {p} _ {i} \\in \\mathcal {P} _ {k} ^ {(t)}} \\mathbf {p} _ {i}, \\quad \\forall k \\in \\{1, \\dots , K \\}, \\tag {4}", + "image_path": "42fd8e467c9e490d080bec2a678e87b0e8aa0cdfb5e347bd33f183c5b8cd7a96.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_k^{(t)} = \\{\\mathbf{p}_i\\mid k_i = k\\}" + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "content": " is the set of patches assigned to the " + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "content": " -th motion word at iteration " + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 55, + 290, + 296, + 328 + ], + "type": "text", + "content": " is the EMA decay factor that controls the smoothness of the updates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 335, + 117, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 117, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 117, + 346 + ], + "type": "text", + "content": "3.3.Decoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "content": "To reconstruct the input skeleton sequences from the quantized latent representations, we design a decoder that mirrors the encoder's architecture. After quantization, the depatchified tensor " + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "content": " is reshaped into " + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}' \\in \\mathbb{R}^{(N \\cdot V) \\times D \\times T}" + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "content": " to obtain the joint independence. This reshaping allows the decoder to process each joint's temporal sequence individually, similar to the encoder's approach. The reconstructed skeleton, denoted as " + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\in \\mathbb{R}^{N \\times C \\times T \\times V}" + }, + { + "bbox": [ + 55, + 354, + 296, + 462 + ], + "type": "text", + "content": ", is then obtained after reshaping." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 468, + 143, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 143, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 143, + 481 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "content": "The network is trained end-to-end using two loss functions. The first term is the reconstruction loss that compares the reconstructed sequences " + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times C\\times T\\times V}" + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "content": " with the original sequences " + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{N\\times C\\times T\\times V}" + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "content": ". The loss ensures that the encoder and codebook preserve as much information from the input sequences as possible. While the mean squared error (MSE) between " + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 487, + 296, + 605 + ], + "type": "text", + "content": " is an obvious choice, we propose to compute inter-joint distance MSE loss, defined over the distance from each joint to all other joints within the same frame:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 613, + 294, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 613, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 58, + 613, + 294, + 646 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e c}} = \\frac {1}{N \\cdot T \\cdot V ^ {2}} \\sum_ {n = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\sum_ {v = 1} ^ {V} \\sum_ {w = 1} ^ {V} \\left(d X _ {n t v w} - d \\hat {X} _ {n t v w}\\right) ^ {2}", + "image_path": "c57378570a3bee4d5ff5e29181b1ff7e638b1e6885cff9e26214911a033c5663.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 84, + 647, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 647, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 84, + 647, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\text {w i t h} \\quad \\begin{array}{r l} d X _ {n t v w} & = \\left\\| \\mathbf {X} _ {n t v} - \\mathbf {X} _ {n t w} \\right\\| _ {2}, \\\\ d \\hat {X} _ {n t v w} & = \\left\\| \\hat {\\mathbf {X}} _ {n t v} - \\hat {\\mathbf {X}} _ {n t w} \\right\\| _ {2}. \\end{array} \\tag {5}", + "image_path": "f981b3a2526f12fe792d4a682c6af9611eff42d8210218415302653ae493be6a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The loss has the advantage that it is inherently translation and rotation-invariant and only measures differences in" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "poses. In our experiments, we show that the proposed loss performs better than a standard MSE loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 96, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 156 + ], + "type": "text", + "content": "To ensure that the learned embeddings are separated into distinct skeleton motion words, we employ the commitment loss [57]. This loss encourages the encoder to produce latent embeddings that are close to their assigned motion words. The commitment loss is defined as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 161, + 553, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 161, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 364, + 161, + 553, + 188 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o m m i t}} = \\sum_ {\\mathbf {p} _ {i} \\in \\mathbf {Z} _ {p}} \\| \\operatorname {s g} \\left[ \\mathbf {c} _ {k _ {i}} \\right] - \\mathbf {p} _ {i} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "a9a1d97efc298db59260c0d912c236341b5629c3a9d65076943cc3a9d13eb490.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{k_i} \\in \\{\\mathbf{c}_1, \\dots, \\mathbf{c}_K\\}" + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "content": " is the motion word that the latent patch embedding " + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "content": " has been assigned to, and " + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot]" + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "content": " denotes the stop-gradient operator, which stops the backpropagation of the gradient." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 243, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 266 + ], + "type": "text", + "content": "The total loss function combines both the reconstruction loss and the commitment loss" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 383, + 275, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 275, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 383, + 275, + 553, + 289 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = \\lambda L _ {\\text {r e c}} + L _ {\\text {c o m m i t}}, \\tag {7}", + "image_path": "18ee9d11bead5ee8527b99c84b5c865136ade74619e87ed3f7da698829fb1db2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 296, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 296, + 554, + 319 + ], + "type": "inline_equation", + "content": "\\lambda = 0.001" + }, + { + "bbox": [ + 313, + 296, + 554, + 319 + ], + "type": "text", + "content": " is the hyperparameter that balances the loss terms." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 395, + 342 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 349, + 376, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 376, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 376, + 361 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 366, + 554, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 414 + ], + "type": "text", + "content": "We evaluated our method on three skeleton-based action datasets: HuGaDB [9], LARa [44], and BABEL [46]. Additional results on the PKU-MMD v2 dataset [37] are reported in the suppl. material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 415, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 474 + ], + "type": "text", + "content": "HuGaDB consists of 364 trials capturing 10 lower limb activities, including walking, running, and sitting. Data were collected from 18 participants using six IMUs (3-axis accelerometer and 3-axis gyroscope) placed on the thighs, shins, and feet, recorded at " + }, + { + "bbox": [ + 313, + 415, + 554, + 474 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 415, + 554, + 474 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 474, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 568 + ], + "type": "text", + "content": "LARa comprises 439 trials across 8 warehouse-related action categories. Data were collected from 16 participants using marker-based MoCap (200 fps), recording the 3-axis position and orientation of 22 full body joints. The dataset includes 960 minutes of recordings. LARa is downsampled by a factor of 4, reducing the frame rate to 50 fps [14]. Additionally, the skeleton is centered from the root joint for translation invariance." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "The BABEL dataset contains 43 hours of 3D human motion sequences from AMASS [43], annotated with over 63,000 frame-level labels across more than 250 action categories. Following the protocol of [66], we extract 25 full-body joint skeletons from SMPL [40] parameters, construct three subsets, each focusing on four action categories, and downsample all sequences to 30 fps, after centering them on the root joint. In addition to the original protocol, we exclude sequences containing more than " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " background actions to ensure that the models learn relevant actions. The evaluation metrics and implementation details are described in the suppl. material." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12105" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 531, + 216 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 531, + 216 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 531, + 216 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 531, + 216 + ], + "type": "table", + "html": "
MethodHuGaDBLARa
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
SupervisedTCN [28]88.3---56.861.5---20.0
ST-GCN [63]88.7---67.767.9---25.8
MS-TCN [12]86.8---89.965.8---39.6
MS-GCN [14]90.4---93.065.6---43.6
UnsupervisedCTE [26]33.84.70.60.60.523.316.88.15.22.3
CTE + Viterbi [26]39.221.713.29.57.523.017.76.83.71.6
TOT [27]33.83.10.70.50.421.47.85.32.71.1
TOT + Viterbi [27]33.820.815.610.57.532.617.711.67.43.2
ASOT [61]33.917.44.53.83.022.923.417.812.15.7
SMQ (ours)42.036.138.531.524.337.439.434.728.416.4
", + "image_path": "5be54b1e991989d447837a7e77ae6476d857c65317947da025f12b85525ca548.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 237, + 555, + 329 + ], + "blocks": [ + { + "bbox": [ + 71, + 220, + 537, + 232 + ], + "lines": [ + { + "bbox": [ + 71, + 220, + 537, + 232 + ], + "spans": [ + { + "bbox": [ + 71, + 220, + 537, + 232 + ], + "type": "text", + "content": "Table 1. Comparison to supervised and unsupervised temporal action segmentation methods on the HuGaDB and LARa datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 237, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 57, + 237, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 57, + 237, + 555, + 329 + ], + "type": "table", + "html": "
MethodBABEL Subset-1BABEL Subset-2BABEL Subset-3
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
CTE [26]34.828.625.017.59.540.330.617.812.27.431.413.18.25.83.6
CTE + Viterbi [26]30.936.223.215.27.342.430.724.319.512.831.230.920.715.28.4
TOT [27]31.818.714.27.64.435.412.813.78.64.331.57.14.92.91.7
TOT + Viterbi [27]29.129.331.520.89.935.336.835.930.019.834.033.831.326.817.9
ASOT [61]35.343.142.334.124.543.137.740.333.423.438.027.127.421.614.3
SMQ (ours)36.638.540.932.822.349.137.843.837.427.440.638.638.029.319.3
", + "image_path": "4e3be8f6bfcbea0b6ad9e553110e09ec60455d95f63be70c7fbe0a2f5f293afd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 333, + 486, + 344 + ], + "lines": [ + { + "bbox": [ + 123, + 333, + 486, + 344 + ], + "spans": [ + { + "bbox": [ + 123, + 333, + 486, + 344 + ], + "type": "text", + "content": "Table 2. Comparison to unsupervised temporal action segmentation methods on the BABEL dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 365, + 223, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 223, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 223, + 377 + ], + "type": "text", + "content": "4.2. Comparison to State of the Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 456 + ], + "type": "text", + "content": "We compare our proposed method, SMQ, with state-of-the-art unsupervised temporal action segmentation methods on the HuGaDB, LARa and BABEL datasets, as shown in Tables 1 and 2. Across all evaluation metrics, SMQ consistently outperforms unsupervised temporal action segmentation approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 458, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 601 + ], + "type": "text", + "content": "On the HuGaDB dataset in Table 1, SMQ achieves a MoF of 42.0 and Edit Score of 36.1, surpassing the best performing approach CTE [26] by 2.8 and 14.4, respectively. For the F1@50 metric, our method has the score of 24.3, which is by 16.8 higher than CTE [26] and TOT [27]. On the LARa dataset, SMQ achieves a MoF of 37.4, Edit Score of 39.4, and F1@50 of 16.4, outperforming the best approach ASOT [61] by a large margin as well. For completeness, we also include results from supervised approaches, which reported only MoF and F1@50. While supervised approaches achieve much higher scores, unsupervised approaches do not use any annotations and solve thus a much harder task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 664 + ], + "type": "text", + "content": "On the BABEL dataset in Table 2, SMQ achieves a higher MoF than ASOT [61] for Subset-1, but ASOT performs slightly better for the other metrics. Whereas on Subset-2 and Subset-3, SMQ outperforms all unsupervised temporal action segmentation methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Fig. 3 shows a qualitative comparison between the methods. Both CTE and TOT generate too many small segments. They thus employ a post-processing technique utilizing Viterbi decoding [10, 25], to refine initial predictions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 460 + ], + "type": "text", + "content": "and achieve temporally consistent action segments. The Viterbi decoding [25] computes the most probable sequence of actions by maximizing the joint likelihood under a first-order Markov model, using transition probabilities that penalize frequent label changes. While this improves segmental metrics by reducing fragmentation, it also imposes a strong prior that each action appears only once per sequence, which is not the case in the datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 676 + ], + "type": "text", + "content": "ASOT can identify re-occurring actions and does not suffer from small segments. The approach, however, has the tendency to predict rather large segments. The small segments in the HuGaDB and LARa datasets are not discovered. Since the segments are longer in the BABEL subsets, ASOT performs better on this dataset. While SMQ does not generate perfect results, for instance it mixes standing up and sitting down in the example from the HuGaDB dataset, the results are better aligned with the ground truth. It is able to detect re-occurring actions, and it does not have the tendency to generate too many small segments, as CTE and TOT, or overly large segments as ASOT. Overall, the results show that SMQ is more effective in handling diverse segmentation scenarios. However, as can be seen in Fig. 3, identifying precise action transitions remains a challenge. Since actions are assigned at the patch level, transitions can only occur at patch boundaries. Improving the detection of action boundaries needs to be investigated in future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We also compare SMQ to self-supervised representation learning approaches for skeleton data [16, 35, 52]. These approaches are typically first trained in a self-supervised" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12106" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 303, + 167 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 303, + 167 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 303, + 167 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 303, + 167 + ], + "type": "image", + "image_path": "a144090599241001b4c78ad6d0f7fa0da08f0ba96db58e94ea859f4c335e1ccd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 170, + 204, + 179 + ], + "lines": [ + { + "bbox": [ + 161, + 170, + 204, + 179 + ], + "spans": [ + { + "bbox": [ + 161, + 170, + 204, + 179 + ], + "type": "text", + "content": "(a) HuGaDB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 547, + 168 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 547, + 168 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 547, + 168 + ], + "type": "image", + "image_path": "204e6da2416d2db82df422d8ae97130dc0311abb2dccca41870d299805429278.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 170, + 444, + 179 + ], + "lines": [ + { + "bbox": [ + 411, + 170, + 444, + 179 + ], + "spans": [ + { + "bbox": [ + 411, + 170, + 444, + 179 + ], + "type": "text", + "content": "(b) LARa" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 182, + 301, + 277 + ], + "blocks": [ + { + "bbox": [ + 63, + 182, + 301, + 277 + ], + "lines": [ + { + "bbox": [ + 63, + 182, + 301, + 277 + ], + "spans": [ + { + "bbox": [ + 63, + 182, + 301, + 277 + ], + "type": "image", + "image_path": "b6c8fca9f9c9a2cea07211823538c2df5409a24d1c4a5729dfa8a1ab5e89375a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 280, + 215, + 289 + ], + "lines": [ + { + "bbox": [ + 147, + 280, + 215, + 289 + ], + "spans": [ + { + "bbox": [ + 147, + 280, + 215, + 289 + ], + "type": "text", + "content": "(c) BABEL Subset-1" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 309, + 182, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 309, + 182, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 309, + 182, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 309, + 182, + 547, + 277 + ], + "type": "image", + "image_path": "ecadcd828e68923d725761c7363ed820a2dddaa62a18a6b481803b89863b803d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 280, + 460, + 289 + ], + "lines": [ + { + "bbox": [ + 392, + 280, + 460, + 289 + ], + "spans": [ + { + "bbox": [ + 392, + 280, + 460, + 289 + ], + "type": "text", + "content": "(d) BABEL Subset-3" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 293, + 446, + 304 + ], + "lines": [ + { + "bbox": [ + 162, + 293, + 446, + 304 + ], + "spans": [ + { + "bbox": [ + 162, + 293, + 446, + 304 + ], + "type": "text", + "content": "Figure 3. Qualitative results for unsupervised action segmentation algorithms." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 310, + 555, + 379 + ], + "blocks": [ + { + "bbox": [ + 58, + 310, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 58, + 310, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 58, + 310, + 555, + 379 + ], + "type": "table", + "html": "
MethodHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Predict&Cluster [52]39.331.533.224.218.318.828.519.212.95.432.232.132.224.314.2
AimCLR [16]34.730.630.923.711.322.028.022.115.67.740.329.732.525.715.5
ActCLR [35]28.221.610.46.24.123.330.623.716.77.832.633.232.824.413.4
SMQ (ours)42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
", + "image_path": "6b1d3a53f21ee94d8d170d97a0aee279f86b7927c6a3822acee9749290a6223b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 144, + 382, + 464, + 393 + ], + "lines": [ + { + "bbox": [ + 144, + 382, + 464, + 393 + ], + "spans": [ + { + "bbox": [ + 144, + 382, + 464, + 393 + ], + "type": "text", + "content": "Table 3. Comparison to self-supervised skeleton-based representation learning methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 403, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 619 + ], + "type": "text", + "content": "way and then fine-tuned for different downstream tasks like skeleton-based action recognition using annotated data. In order to apply these methods to unsupervised temporal action segmentation, we divided the input skeleton sequences into non-overlapping patches, consistent with the latent patching in SMQ. These patches were fed through the self-supervised representation learning approaches to learn patch-level action representations. We then performed k-means clustering over the patch-level representations. The results in Table 3 show that SMQ consistently outperforms existing self-supervised representation learning approaches on all the datasets. It needs to be noted that these approaches are designed for datasets with only one action per sequence, whereas unsupervised temporal action segmentation requires processing long, untrimmed skeleton sequences containing multiple actions per sequence. Compared to SMQ, they therefore learn less distinctive representations for this task as it is visualized in the suppl. material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 632, + 123, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 123, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 123, + 643 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "content": "In this section, we analyze the impact of different design choices in our model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Impact of Disentangled Embedding. As illustrated in Fig. 2, the encoder processes each joint independently, which ensures that the information of the different joints" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 536 + ], + "type": "text", + "content": "remains disentangled in the embedding space. To evaluate the effectiveness of the disentangled embedding space, we conducted an ablation study. In Table 4, we compare our approach to a setup where we concatenate the joints before learning an embedding. In the latter case, the joints get entangled in the embedding and the structural information of the skeleton gets lost. The results demonstrate that the disentangled embedding enhances our model's performance, indicating that learning joint representations separately allows for more accurate action segmentation by enabling each joint to reside in distinct subspaces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Impact of Patch Size. We apply temporal patching to the latent skeleton embeddings to obtain skeleton motion words. Specifically, we use a fixed patch size to create non-overlapping patches of the representations. To investigate the effect of patch size on the performance of our model, we conducted an ablation study with varying patch sizes, as presented in Table 5. Without patching, i.e., patch size is 1, we obtain a frame-wise representation. The results indicate that SMQ consistently outperforms the frameswise representation across all evaluated metrics, with significant improvements in segmental metrics. These findings emphasize that frame-wise quantization leads to over-segmentation, as it fails to capture the inherent temporal consistency of actions. In contrast, SMQ captures the tem" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12107" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 119 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 119 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 119 + ], + "type": "table", + "html": "
EmbeddingHuGaDBLARaBABEL - Avg
MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}MoFEditF1@{10, 25, 50}
Entangled38.934.235.627.918.926.635.929.522.812.340.132.134.927.116.9
Disentangled42.036.138.531.524.337.439.434.728.416.442.138.340.933.223.0
", + "image_path": "876c9c501d6892b7722ad8cda9cfff6625e0ba1e03791696f7e1e5d4625a229a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 138, + 277, + 236 + ], + "blocks": [ + { + "bbox": [ + 130, + 122, + 478, + 135 + ], + "lines": [ + { + "bbox": [ + 130, + 122, + 478, + 135 + ], + "spans": [ + { + "bbox": [ + 130, + 122, + 478, + 135 + ], + "type": "text", + "content": "Table 4. Impact of disentangled joint embedding on the HuGaDB, LARa, and BABEL datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 138, + 277, + 236 + ], + "lines": [ + { + "bbox": [ + 73, + 138, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 73, + 138, + 277, + 236 + ], + "type": "table", + "html": "
Patch SizeMoFEditF1@{10, 25, 50}
1 (Frame-wise)33.925.621.214.98.6
25 (0.5 sec)35.235.732.425.014.9
50 (1 sec)37.439.434.728.416.4
75 (1.5 sec)37.239.134.728.417.1
100 (2 sec)30.836.133.027.015.4
150 (3 sec)31.434.032.626.514.7
", + "image_path": "e55b70a41b42c0f27dcd999ea2351ec373fa98bf0314a488a1742c3cdab9a792.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 258, + 295, + 334 + ], + "blocks": [ + { + "bbox": [ + 69, + 239, + 281, + 251 + ], + "lines": [ + { + "bbox": [ + 69, + 239, + 281, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 239, + 281, + 251 + ], + "type": "text", + "content": "Table 5. Impact of motion patch size on the LARa dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 258, + 295, + 334 + ], + "lines": [ + { + "bbox": [ + 61, + 258, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 295, + 334 + ], + "type": "table", + "html": "
LcommitReconstruction LossMoFEditF1@{10,25,50}
✓X29.929.926.817.78.8
XInter-Joint Dist. MSE31.034.630.624.414.4
✓MSE34.639.434.527.415.7
✓Root Distance MSE28.034.930.223.814.1
✓Inter-Joint Dist. MSE37.439.434.728.416.4
", + "image_path": "cad07acd7eee282143bba74fc2fc249381bc7ddf18fd0d46556e560e5a2af90d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 640 + ], + "type": "text", + "content": "poral variability of actions more effectively by grouping contiguous latent embeddings into patches. This results in more semantically meaningful action clusters that yield temporally consistent segmentations. On the LARa dataset, a patch size of 50 performs best, which corresponds to one second of motion, as the data is downsampled to 50 fps. We keep this one-second patching consistent across all datasets. Impact of Loss Variants. To evaluate the effectiveness of different loss components, we conduct an ablation study examining various configurations of the commitment loss and reconstruction loss. We first evaluate three variants of the reconstruction loss in the last three rows of Table 6. The proposed inter-joint distance MSE loss performs better than the commonly used MSE loss since it is invariant to translation and rotation. The root distance MSE loss is obtained by computing the MSE over the distances of each joint to the root joint within the same frame. While this loss is also translation and rotation invariant, it performs even worse than the inter-joint variant. Furthermore, we evaluate the impact of the commitment loss and reconstruction loss by omitting one of them as presented in the first two rows of the Table 6. Removing either of the loss terms substantially decreases the performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Impact of Reconstruction Loss Weight. In our ablation study, we varied the reconstruction loss weight, " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " (7), to understand its effect on model performance. By adjusting " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", we explore the balance between the reconstruction and commitment loss in Table 7. Our model is not very sensitive to this parameter and performs well for " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.001" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 350, + 140, + 518, + 213 + ], + "blocks": [ + { + "bbox": [ + 62, + 337, + 288, + 348 + ], + "lines": [ + { + "bbox": [ + 62, + 337, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 288, + 348 + ], + "type": "text", + "content": "Table 6. Impact of loss function variants on the LARa dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 140, + 518, + 213 + ], + "lines": [ + { + "bbox": [ + 350, + 140, + 518, + 213 + ], + "spans": [ + { + "bbox": [ + 350, + 140, + 518, + 213 + ], + "type": "table", + "html": "
λMoFEditF1@{10, 25, 50}
0.135.938.433.426.915.6
0.0136.238.934.427.815.9
0.00137.439.434.728.416.4
0.000134.139.835.427.916.1
", + "image_path": "f5f638b9adbcfacf0cc2240a17479036cd9f00e829e4c94975d69bed85cf1c21.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 354, + 238, + 515, + 311 + ], + "blocks": [ + { + "bbox": [ + 356, + 216, + 511, + 227 + ], + "lines": [ + { + "bbox": [ + 356, + 216, + 511, + 227 + ], + "spans": [ + { + "bbox": [ + 356, + 216, + 511, + 227 + ], + "type": "text", + "content": "Table 7. Impact of " + }, + { + "bbox": [ + 356, + 216, + 511, + 227 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 356, + 216, + 511, + 227 + ], + "type": "text", + "content": " on the LARa dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 238, + 515, + 311 + ], + "lines": [ + { + "bbox": [ + 354, + 238, + 515, + 311 + ], + "spans": [ + { + "bbox": [ + 354, + 238, + 515, + 311 + ], + "type": "table", + "html": "
αMoFEditF1@{10, 25, 50}
0.3534.839.435.027.615.5
0.5037.439.434.728.416.4
0.6535.538.533.927.215.8
0.8034.237.733.427.115.5
", + "image_path": "b7d80ad1093b3f76ebb9371898bbb73ff8b697e594edd1b9dad6da1e3ab41533.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 315, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 323, + 315, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 323, + 315, + 544, + 326 + ], + "type": "text", + "content": "Table 8. Impact of EMA decay rate " + }, + { + "bbox": [ + 323, + 315, + 544, + 326 + ], + "type": "inline_equation", + "content": "\\left( \\alpha \\right)" + }, + { + "bbox": [ + 323, + 315, + 544, + 326 + ], + "type": "text", + "content": " on the LARa dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "text", + "content": "Impact of Decay Factor. To examine the role of the EMA decay factor " + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "text", + "content": " (3), we perform an ablation study by evaluating multiple " + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "text", + "content": " values in Table 8. The decay factor governs the balance between incorporating new patch information and maintaining historical motion word estimates, thereby influencing the smoothness and responsiveness of the updates. Our model performs best when " + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 347, + 555, + 445 + ], + "type": "text", + "content": " is set to 0.5. Additional ablation studies are provided in the suppl. material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we introduced SMQ, a method for unsupervised skeleton-based temporal action segmentation. It uses an autoencoder that preserves the skeleton structure by ensuring that the information of the different joints remains disentangled in the embedding space. The skeleton motion quantization module then patches the latent representations into short time series of fixed length and quantizes them into skeleton motion words. We performed an extensive evaluation and compared the proposed approach to state-of-the-art unsupervised temporal action segmentation approaches, as well as approaches for self-supervised skeleton-based representation learning. SMQ outperforms them by a large margin. Furthermore, we demonstrated the importance of learning the representation over patches instead of frames and the importance of an embedding space where the information of the different joints remains disentangled. As part of future work, it would be interesting to investigate a hierarchical version of SMQ and improve the detection of action boundaries." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work has been supported by the ERC Consolidator Grants FORHUE (101044724) and ActionContraThreat (816564), the Ministry of Culture and Science of the State of North Rhine-Westphalia (iBehave network, Programme \"Netzwerke 2021\"), and the TRA Modelling (University of Bonn) as part of the Excellence Strategy of the federal and state governments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 295, + 262 + ], + "type": "text", + "content": "[1] Robert T Arn, Pradyumna Narayana, Tegan Emerson, Bruce A Draper, Michael Kirby, and Chris Peterson. Motion segmentation via generalized curvatures. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 41 (12):2919-2932, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 295, + 308 + ], + "type": "text", + "content": "[2] Jernej Barbič, Alla Safonova, Jia-Yu Pan, Christos Faloutsos, Jessica K Hodgins, and Nancy S Pollard. Segmenting motion capture data into distinct behaviors. In Proceedings of Graphics Interface (GI), pages 185-194, 2004. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 294, + 373 + ], + "type": "text", + "content": "[3] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 52-68. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "text", + "content": "[4] Elena Belén Bueno-Benito, Biel Tura Vecino, and Mariella Dimiccoli. Leveraging triplet loss for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4921-4929, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 476 + ], + "type": "text", + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), pages 132-149, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 531 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 294, + 588 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9650-9660, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 634 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 688 + ], + "type": "text", + "content": "[9] Roman Chereshnev and Attila Kertesz-Farkas. Hugadb: Human gait database for activity recognition from wearable inertial sensor networks. In Analysis of Images, Social Networks and Texts (AIST), pages 131-141. Springer, 2018. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[10] Guodong Ding, Fadime Sener, and Angela Yao. Temporal action segmentation: An analysis of modern techniques." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "text", + "content": "[11] Zexing Du, Xue Wang, Guoqing Zhou, and Qing Wang. Fast and unsupervised action boundary detection for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3323-3332, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "text", + "content": "[12] Yazan Abu Farha and Juergen Gall. Ms-tn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3575-3584, 2019. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "text", + "content": "[13] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "text", + "content": "[14] Benjamin Filtjens, Bart Vanrumste, and Peter Slaets. Skeleton-based action segmentation with multi-stage spatial-temporal graph convolutional neural networks. IEEE Transactions on Emerging Topics in Computing, 12(1):202-212, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 376 + ], + "type": "text", + "content": "[15] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 442 + ], + "type": "text", + "content": "[16] Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 762-770, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[17] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 601 + ], + "type": "text", + "content": "[19] Babak Hosseini, Romain Montagne, and Barbara Hammer. Deep-aligned convolutional neural network for skeleton-based action recognition and segmentation. Data Science and Engineering (DSE), 5(2):126-139, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 602, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 655 + ], + "type": "text", + "content": "[20] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14024-14034, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[21] Haoyu Ji, Bowen Chen, Xinglong Xu, Weihong Ren, Zhiyong Wang, and Honghai Liu. Language-assisted skeleton action understanding for skeleton-based temporal action segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12109" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 106 + ], + "type": "text", + "content": "[22] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 150 + ], + "type": "text", + "content": "[23] Björn Krüger, Anna Vögele, Tobias Willig, Angela Yao, Reinhard Klein, and Andreas Weber. Efficient unsupervised temporal segmentation of motion data. IEEE Transactions on Multimedia (T-MM), 19(4):797-812, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[24] Hilde Kuehne, Juergen Gall, and Thomas Serre. An end-to-end generative framework for video segmentation and recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1-8. IEEE, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 250 + ], + "type": "text", + "content": "[25] Hilde Kuehne, Alexander Richard, and Juergen Gall. A hybrid rnn-hmm approach for weakly supervised temporal action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(4):765-779, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 305 + ], + "type": "text", + "content": "[26] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Juergen Gall. Unsupervised learning of action classes with continuous temporal embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12066-12074, 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 370 + ], + "type": "text", + "content": "[27] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 426 + ], + "type": "text", + "content": "[28] Colin Lea, Michael D Flynn, Rene Vidal, Austin Reiter, and Gregory D Hager. Temporal convolutional networks for action segmentation and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 156-165, 2017. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 471 + ], + "type": "text", + "content": "[29] Jun Li and Sinisa Todorovic. Action shuffle alternating learning for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12628-12636, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 536 + ], + "type": "text", + "content": "[30] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4741-4750, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[31] Shijie Li, Yazan Abu Farha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45 (6):6647-6658, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "text", + "content": "[32] Yunheng Li, Zhongyu Li, Shanghua Gao, Qilong Wang, Qibin Hou, and Ming-Ming Cheng. A decoupled spatiotemporal framework for skeleton-based action segmentation. arXiv preprint arXiv:2312.05830, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 691 + ], + "type": "text", + "content": "[33] Yun-Heng Li, Kai-Yuan Liu, Sheng-Lan Liu, Lin Feng, and Hong Qiao. Involving distinguished temporal graph convolutional networks for skeleton-based temporal action segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):647-660, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[34] Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "text", + "content": "recognition. In Proceedings of the ACM International Conference on Multimedia (ACM MM), pages 2490-2498, 2020. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[35] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Action-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2363-2372, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 554, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 208 + ], + "type": "text", + "content": "[36] Wei Lin, Anna Kukleva, Horst Possegger, Hilde Kuehne, and Horst Bischof. Taec: Unsupervised action segmentation with temporal-aware embedding and clustering. In Computer Vision Winter Workshop, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "text", + "content": "[37] Chunhui Liu, Yueyu Hu, Yanghao Li, Sijie Song, and Jiaying Liu. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. arXiv preprint arXiv:1703.07475, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "text", + "content": "[38] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10139-10149, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[39] Kaiyuan Liu, Yunheng Li, Yuanfeng Xu, Shuai Liu, and Shenglan Liu. Spatial focus attention for fine-grained skeleton-based action tasks. IEEE Signal Processing Letters, 29:1883-1887, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 399 + ], + "type": "text", + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "text", + "content": "[41] Zijia Lu and Ehsan Elhamifar. FACT: Frame-action cross-attention temporal modeling for efficient supervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 510 + ], + "type": "text", + "content": "[42] Hao Ma, Zaiyue Yang, and Haoyang Liu. Fine-grained unsupervised temporal action segmentation and distributed representation for skeleton-based human motion analysis. IEEE Transactions on Cybernetics (T-CYB), 52(12):13411-13424, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 567 + ], + "type": "text", + "content": "[43] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5442-5451, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 623 + ], + "type": "text", + "content": "[44] Friedrich Niemann, Christopher Reining, Fernando Moya Rueda, Nilah Ravi Nair, Janine Anika Steffens, Gernot A Fink, and Michael Ten Hompel. Lara: Creating a dataset for human activity recognition in logistics using semantic attributes. Sensors, 20(15):4083, 2020. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "text", + "content": "[45] Giancarlo Paoletti, Jacopo Cavazza, Cigdem Bryan, and Alessio Del Bue. Unsupervised human action recognition with skeletal graph laplacian and self-supervised viewpoints invariance. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[46] Abhinanda R. Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J. Black. BABEL: Bodies, action and behavior with english" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12110" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "labels. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 722-731, 2021. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[47] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11225-11234, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[48] Jan Sedmidubsky, Petra Budikova, Vlastislav Dohnal, and Pavel Zezula. Motion words: a text-like representation of 3d skeleton sequences. In Proceedings of the European Conference on Information Retrieval (ECIR), pages 527-541. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "text", + "content": "[49] Fadime Sener and Angela Yao. Unsupervised learning and segmentation of complex activities from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8368-8376, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "text", + "content": "[50] Federico Spurio, Emad Bahrami, Gianpiero Francesca, and Juergen Gall. Hierarchical vector quantization for unsupervised action segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6996-7005, 2025. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[51] Lucas Stoffl, Andy Bonnetto, Stephane d'Ascoli, and Alexander Mathis. Elucidating the hierarchical nature of behavior with masked autoencoders. In Proceedings of the European Conference on Computer Vision (ECCV), pages 106-125, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[52] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9631-9640, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[53] Chenwei Tan, Tao Sun, Talas Fu, Yuhan Wang, Minjie Xu, and Shenglan Liu. Hierarchical spatial-temporal network for skeleton-based temporal action segmentation. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 28-39. Springer, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "text", + "content": "[54] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Stga-net: Spatial-temporal graph attention network for skeleton-based temporal action segmentation. In Proceedings of the IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 218-223. IEEE, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 590 + ], + "type": "text", + "content": "[55] Xiaoyan Tian, Ye Jin, Zhao Zhang, Peng Liu, and Xianglong Tang. Spatial-temporal graph transformer network for skeleton-based temporal action segmentation. Multimedia Tools and Applications (MTAP), 83(15):44273-44297, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 658 + ], + "type": "text", + "content": "[56] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6426-6436, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems (NeurIPS), 30, 2017. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[58] Anna Vögele, Björn Krüger, and Reinhard Klein. Efficient unsupervised temporal segmentation of human motion. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Proceedings of the ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), pages 167-176, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[59] Qifei Wang, Gregorij Kurillo, Ferda Ofli, and Ruzena Bajcsy. Unsupervised temporal segmentation of repetitive human actions based on kinematic modeling and frequency analysis. In Proceedings of the IEEE International Conference on 3D Vision (3DV), pages 562-570. IEEE, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[60] Leiyang Xu, Qiang Wang, Xiaotian Lin, and Lin Yuan. An efficient framework for few-shot skeleton-based temporal action segmentation. Proceedings of the European Conference on Computer Vision (ECCV), 232:103707, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 261 + ], + "type": "text", + "content": "[61] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14618-14627, 2024. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[62] Ziwei Xu, Xudong Shen, Yongkang Wong, and Mohan S. Kankanhalli. Unsupervised motion representation learning with capsule autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 34:3205-3217, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[63] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 418 + ], + "type": "text", + "content": "[64] Di Yang, Yaohui Wang, Antitza Dantcheva, Quan Kong, Lorenzo Garattoni, Gianpiero Francesca, and Francois Bremond. Lac-latent action composition for skeleton-based action segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13679-13690, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "text", + "content": "[65] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASFormer: Transformer for action segmentation. In Proceedings of the British Machine Vision Conference (BMVC), 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 497 + ], + "type": "text", + "content": "[66] Qing Yu and Kent Fujiwara. Frame-level label refinement for skeleton-based weakly-supervised action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3322-3330, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[67] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In Proceedings of the European Conference on Computer Vision (ECCV), pages 36-51. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "text", + "content": "[68] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "text", + "content": "[69] Feng Zhou, Fernando De la Torre, and Jessica K. Hodgins. Aligned cluster analysis for temporal segmentation of human motion. In Proceedings of the IEEE International Conference on Automatic Face and Gesture Recognition (FG), pages 1-7, 2008. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 666, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 553, + 710 + ], + "type": "text", + "content": "[70] Feng Zhou, Fernando De la Torre, and Jessica K Hodgins. Hierarchical aligned cluster analysis for temporal clustering of human motion. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 35(3):582-596, 2012. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12111" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_content_list.json b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb4fc69fc6157b234d435b9d4f82e962b6dd96c9 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_content_list.json @@ -0,0 +1,1372 @@ +[ + { + "type": "text", + "text": "SketchSplat: 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting", + "text_level": 1, + "bbox": [ + 217, + 130, + 781, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haiyang Ying, Matthias Zwicker \nUniversity of Maryland, College Park", + "bbox": [ + 320, + 203, + 643, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://oceanying.github.io/SketchSplat", + "bbox": [ + 320, + 241, + 666, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edges are one of the most basic parametric primitives to describe structural information in 3D. In this paper, we study parametric 3D edge reconstruction from calibrated multi-view images. Previous methods usually reconstruct a 3D edge point set from multi-view 2D edge images, and then fit 3D edges to the point set. However, noise in the point set may cause gaps among fitted edges, and the recovered edges may not align with input multi-view images since the edge fitting depends only on the reconstructed 3D point set. To mitigate these problems, we propose SketchSplat, a method to reconstruct accurate, complete, and compact 3D edges via differentiable multi-view sketch splatting. We represent 3D edges as sketches, which are parametric lines and curves defined by attributes including control points, scales, and opacity. During reconstruction, we iteratively sample Gaussian points from a set of sketches and rasterize the Gaussians onto 2D edge images. Then the gradient of the image loss can be back-propagated to optimize the sketch attributes. Our method bridges 2D edge images and 3D edges in a differentiable manner, which ensures that 3D edges align well with 2D images and leads to accurate and complete results. We also propose a series of adaptive topological operations to reduce redundant edges and apply them along with the sketch optimization, yielding a more compact reconstruction. Finally, we contribute an accurate 2D edge detector that improves the performance of both ours and existing methods. Experiments show that our method achieves state-of-the-art accuracy, completeness, and compactness on a benchmark CAD dataset.", + "bbox": [ + 89, + 325, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Edges serve as one of the most fundamental elements of visual perception, playing a crucial role in understanding the geometry and structure of the 3D world. Unlike point-based representations, continuous edges offer a more compact and perceptually aligned way to describe 3D structures. As a", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d6f51cef5ca6b312b3eaeaeb38899e5d7d4e6cd889cff2f3e06149849a622cdf.jpg", + "image_caption": [ + "Figure 1. We propose SketchSplat, a multi-view 3D edge reconstruction method based on differentiable sketch splatting. Previous methods [7, 19] first recover edge points and then fit them to 3D edges, which is sensitive to noise and often leads to incomplete results. Instead, we sample Gaussian points from 3D sketches and rasterize them to compute a rendering loss, enabling direct edge parameter optimization under image supervision. Additionally, we design adaptive topology control methods to refine edge structures during training and contribute an improved 2D edge detection method for more accurate reconstruction." + ], + "image_footnote": [], + "bbox": [ + 517, + 287, + 906, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cornerstone of computer vision and graphics, 3D edge reconstruction underpins a wide range of applications, including CAD modeling [10, 20, 56], SLAM [13, 24, 41, 50], and autonomous driving [8, 35, 36]. Despite its significance, achieving high-quality 3D edge reconstructions remains a challenging task due to the inherent difficulty of ensuring", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25649", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the accuracy and completeness of reconstructed edges and maintaining the connectivity of the reconstructed structure.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional methods [5, 12, 23, 48] usually detect 2D line segments in each view. Then the 2D lines are lifted to 3D via line matching and triangulation across views. While these methods achieve high accuracy and efficiency, multiview consistent detection and robust line matching are difficult to achieve, leading to incomplete results. In addition, these methods are usually not capable of handling lines and curves simultaneously.", + "bbox": [ + 89, + 123, + 480, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, differentiable rendering has emerged as a promising technique to reconstruct 3D contents for novel view synthesis [3, 27, 28, 55], surface reconstruction [11, 22, 25, 47], and animation [29, 34, 49]. In the field of edge reconstruction, several works introduced the idea of differentiable rendering to recover edge information [7, 19, 54]. Instead of detecting parametric 2D edge segments explicitly, they leverage pixel-wise 2D edge intensity images [33, 44] as a supervision signal and optimize a point-based 3D edge representation via differentiable multiview rendering, resulting in an extracted 3D edge point set. Parametric 3D edges are then obtained by fitting lines and curves to the extracted 3D points. Since the edge-fitting step relies purely on the recovered 3D edge points but not the input 2D edge images, the fitted 3D edges may not align well with the 2D edge images. In addition, noise in the 3D edge points usually leads to discontinuous and fragmented edges.", + "bbox": [ + 91, + 247, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose SketchSplat, a differentiable edge reconstruction framework that directly optimizes parametric 3D edges by fitting them to the input 2D edge images using differentiable rendering. Our experimental results demonstrate that this leads to more accurate, complete, and compact reconstructions, more faithfully reflecting the original connectivity.", + "bbox": [ + 89, + 506, + 480, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, we propose to model 3D edges as sketches, which are parametric lines and curves with control points, scales, and opacity. These sketches can be initialized from existing methods [7, 19, 54], and can be sampled as a set of 3D Gaussian points for differentiable rendering. To optimize the sketches, we sample Gaussian points from the sketches and rasterize them in each view to calculate the rendering loss between the rendered and the ground-truth edge image. Then the error can be back-propagated to update the sketch attributes. Since we optimize our sketch parameters directly using the 2D edge images, the resulting edges will align well with the 2D edges in the input images. To further improve the compactness of the reconstructed edges, we leverage a series of topological operations to connect, merge, and filter sketches during optimization. Finally, we observe that one common limitation of existing methods [7, 19, 54] lies in the inaccuracy of 2D edge images. To mitigate this, we propose an accurate and robust edge detection method, which is validated to improve both", + "bbox": [ + 91, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our approach and existing methods by a large margin.", + "bbox": [ + 511, + 90, + 867, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments demonstrate that our method achieves state-of-the-art performance on the ABC-NEF dataset [54], surpassing previous approaches by over $10\\%$ in F-score. In summary, key innovations of our method include:", + "bbox": [ + 511, + 108, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Differentiable Parametric Optimization: We contribute a pipeline that directly optimizes parametric edge representations via differentiable multi-view rasterization, leading to SOTA accuracy and completeness.", + "- Adaptive Topological Control: We propose a series of topological operations to further enhance edge continuity and reduce redundant edges during training, yielding more compact reconstruction results.", + "- Accurate 2D Edge Detection: We propose a more accurate 2D edge detection method, which shows large improvements for both our approach and existing methods." + ], + "bbox": [ + 511, + 171, + 903, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 353, + 653, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Line Matching and Triangulation. Conventional approaches reconstruct 3D lines [2, 4-6, 12, 26] or curves [15, 16, 38] by enforcing multi-view geometric constraints. Given detected 2D edge segments and their descriptors, they match the line descriptors across multiple views and perform triangulation to lift 2D edges into 3D. However, these methods face challenges in consistently detecting edges and robustly matching lines, especially when occlusions cause endpoint inconsistencies across views. Despite notable advancements by involving structural [39, 48] or learning-based priors [23, 30-32], line detection and matching remain critical bottlenecks, limiting overall performance. While most of the methods can only handle a single type of edges, in this paper we aim at recovering both lines and curves.", + "bbox": [ + 511, + 378, + 906, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differentiable Rendering Methods. The advancement of differentiable rendering opens up a new way for 3D reconstruction [22, 27, 47]. Inspired by NeRF [27], NEF [54] proposes to optimize an implicit neural radiance field with a multi-view rendering loss. After training, 3D edge points are extracted from the radiance field and curves are fitted to these points. NEAT [51] builds on VolSDF [52] and uses a signed distance field to represent both object surfaces and endpoints of 3D lines. The surface-based modeling approach restricts this method to objects with rich texture. EMAP [19] adopts an unsigned distance function (UDF), where sampled points can be displaced towards the zero-level set of the UDF to be more accurate. While methods based on neural field are time consuming to optimize, EdgeGS [7] represents 3D edge fields with explicit Gaussian points and leverages Gaussian splatting [17] to render and optimize Gaussian points, leading to more efficient edge reconstruction. A common issue of these methods", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25650", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is that the edge fitting process relies only on the extracted 3D edge points, hence noise in the reconstructed 3D points may lead to inaccurate and incomplete edges. Instead, we represent 3D edges as sketches defined parametrically by control points, scales, and opacity, and directly optimize these parameters via differentiable rendering. By leveraging sketch sampling and rasterization, our differentiable rendering aligns 3D edges with 2D edge images, resulting in more accurate and complete reconstructions.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Vector Graphics Optimization. DiffVG [21] proposes a differentiable rasterizer that can optimize 2D vector graphics with 2D image supervision in a differentiable manner. However, the differentiation process relies on distance measurements between the rasterized curves and pixels, which needs to be custom tailored for each type of curve. 3Doodle [9] extends the idea of DiffVG [21] to reconstruct 3D curve-based object abstractions using multiple input views. They optimize a set of Bezier curves by minimizing the perceptual loss between the rasterized image and multi-view CLIP feature maps [37, 45]. Since CLIP features only contain abstract information, 3Doodle is not suitable for recovering accurate edges of objects like CAD models. Moreover, they approximate the projection of Bezier curves using an orthographic camera model. Instead, we get inspiration from 3DGS [17] and propose to optimize 3D sketches via sampling and rasterization of Gaussian points sampled from sketch parameters, which can be easily extended to different types of curves and achieves accurate 3D edge reconstruction.", + "bbox": [ + 91, + 247, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 564, + 428, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage 3D Gaussian splatting (3DGS) [17] for initialization and differentiable sketch rendering. 3DGS proposes to model 3D scenes with a collection of anisotropic 3D Gaussian points. Each point is defined as a Gaussian distribution $G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)}$ , where $\\mu \\in R^3$ is the center position or mean, and $\\Sigma \\in R^{3\\times 3}$ is the covariance matrix. Each Gaussian has an opacity $o \\in R$ and a view-dependent color value $c(d) = SH(d) \\in R^3$ modeled with spherical harmonics. Following Zwicker et al. [57], the Gaussians can be rendered to 2D images using an efficient volume rendering approach, where the color for each camera ray $r$ is", + "bbox": [ + 89, + 590, + 482, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc (r) = \\sum_ {i \\in N} c _ {i} (d (r)) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\alpha_ {i} = o _ {i} g _ {i} (r), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 785, + 482, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g_{i}(r)$ is the line integral of the 3D Gaussian $G_{i}$ along ray $r$ , and the indices $i$ are ordered in front to back manner. By calculating the gradient of a loss $\\mathcal{L}$ between the rendered and ground truth image, the parameters of all Gaussians can", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be optimized using gradient descent. A typical loss function is a combination of an $L_{1}$ and a perceptual loss such as", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {D - S S I M}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 133, + 906, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given input images, the initial Gaussians are usually obtained from SfM [40] methods. During training, several operations are adopted to adaptively control the density and distribution of Gaussians. For example, Gaussians with low opacity will be culled, and Gaussians with high gradient will be duplicated or split to densify specific regions.", + "bbox": [ + 511, + 162, + 906, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In SketchSplat, we leverage 3DGS [17] as an intermediary to bridge 2D edge images and 3D parametric edges, enabling the optimization of 3D edges through differentiable image rendering.", + "bbox": [ + 511, + 253, + 905, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 329, + 604, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given multi-view posed RGB images as input, our target is to recover the parametric edges for the 3D objects. Fig. 2 provides an overview of our approach. We firstly compute an edge image for each RGB image using our proposed 2D edge detector. Then we represent edges as 3D parametric sketches, including both lines $l \\in \\mathbb{R}^{2 \\times 3}$ and third-order Bézier curves with four 3D control points $c \\in \\mathbb{R}^{4 \\times 3}$ . We initialize a set of sketches from existing edge reconstruction methods, and treat their control points, along with persketch scales and opacity, as estimizable parameters. During optimization, we use a sample-based method to convert sketches to Gaussian points, which can be rasterized onto 2D images to calculate the image loss. We back-propagate the gradient of the loss to optimize the sketch parameters. In addition, we introduce a set of topological operations to adaptively refine the sketches during training, yielding more accurate and compact edge reconstruction results.", + "bbox": [ + 511, + 354, + 906, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Edge Image Extraction", + "text_level": 1, + "bbox": [ + 511, + 621, + 725, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previously, some methods [23, 51] extract edges with line detectors like LSD [46] and SOLD2 [30]. While these methods produce accurate parametric lines in 2D, they may fail to handle curve regions and result in incomplete results. Recent methods [7, 19, 54] use neural network-based edge detectors [33, 44] to detect edge intensity maps as supervision signal. However, since these detectors are not trained specifically on CAD-like objects, they may miss some significant edges in 2D or wrongly recover shading textures as real edges. Finally, we observed that these detectors introduce bias by consistently shifting detected edges slightly from certain real edges, which leads to multi-view inconsistency in 3D edge reconstruction (see supplementary). We address these issues by proposing a new edge detection method, which leverages geometric cues (e.g., per-pixel depth and normal estimation of 2D RGB images) to produce more accurate edge detection results.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25651", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f08285c5181b9f452d9955574aa0a5bf4bd45b2af4b31cc23ad11f38d28ddc57.jpg", + "image_caption": [ + "(A) Sketch initialization", + "(B) Differentiable Multi-view Sketch Optimization", + "Figure 2. Overview of SketchSplat pipeline: (a) We extract edge images using a novel approach (Sec. 4.1). (b) We then obtain initial edge points using EdgeGS [7] and (c) fit parametric sketches defined as lines and third-order Bézier curves. The core step of our approach is to optimize the sketches by (d) sampling and rasterizing them in a differentiable manner and back-propagating gradients of the image loss to (e) sketch parameters. In addition, during training we apply (f) a set of topological operations to improve compactness and connectivity of the reconstructed sketches. (g) After training, we take the optimized geometric parameters of sketches as the reconstructed 3D edges." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each RGB image, we combine information from three parts: foreground mask $A$ , depth map $D$ , and normal map $N$ . The foreground mask can be extracted using the image alpha-channel (if available), a thresholding operation or possibly using semantic segmentation, providing object boundary information. We obtain a depth image for each view by reconstructing the scene with 2DGS [14] and rendering depth maps from each view. We use a normal estimator [53] to generate normal maps $N$ for each view. Then we calculate gradient maps $g(D)$ and $g(N)$ with the Sobel operator [42], threshold them with $t_d$ and $t_n$ , and combine them via the OR operation, followed by a Gaussian filter $G_f(\\cdot)$ to smooth the boundaries of the edges. In summary, we call our edge detector 2DGS-SN, defined as", + "bbox": [ + 88, + 438, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE = G _ {f} * (A | (g (D) > t _ {d}) | (g (N) > t _ {n})). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 660, + 482, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Sec. 5 we show that this simple approach improves existing methods [7] on CAD-like objects by a large margin.", + "bbox": [ + 89, + 688, + 482, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Sketch Splatting", + "text_level": 1, + "bbox": [ + 89, + 727, + 253, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Representation. We represent edges as 3D parametric sketches, including both lines $l \\in \\mathbb{R}^{2 \\times 3}$ and third-order Bézier curves with control points $c \\in \\mathbb{R}^{4 \\times 3}$ . This is the same as the output format of EdgeGS[7]. While EdgeGS optimizes 3D point positions first and then fits 3D lines and curves to the 3D points, we optimize the parameters of lines and curves directly via multi-view differentiable rasterization. In addition to the estimizable geometry parameters $(l$ or $c)$ , we equip each sketch with an estimizable opacity $o \\in \\mathbb{R}$ and a local scale $s \\in \\mathbb{R}^3$ to model the sketch intensity", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and thickness, respectively. These attributes will support Gaussian sampling and rasterization in the next steps.", + "bbox": [ + 511, + 438, + 906, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sketch Initialization. Unlike point-based methods [7, 19], it is non-trivial to initialize sketches directly from multi-view edge images. Instead, inspired by 3DGS [17], we first initialize sketches with existing point-based methods like EdgeGS [7]. By default, we select EdgeGS because it provides a sufficient number of short 3D edges to cover the objects, and achieves the highest completeness among previous methods, despite some fragmentation and noise.", + "bbox": [ + 511, + 487, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable Sketch Optimization. Our key idea is to bridge the gap between 2D edge images and 3D sketches via differentiable rendering and gradient descent optimization, which significantly improves sketch reconstruction over initialization method EdgeGS [7]. Inspired by 3DGS [17] and DiffVG [21], we choose 3D Gaussians as an intermediate representation to achieve differentiable sketch rendering.", + "bbox": [ + 511, + 628, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each gradient descent step, we first sample points from each sketch with a constant step size (5mm for all experiments), resulting in a point set $P$ . Each point will be assigned the opacity and scale value of the corresponding sketch. To model the orientation, the main direction of each point is sampled as the local tangent direction of the sketch. Then we follow 3DGS[17] to rasterize the Gaussian points onto a chosen 2D view to generate a rendered edge image $E^{*}$ . We calculate the $L_{1}$ loss between $E^{*}$ and the ground truth edge image $E_{gt}$ , and back-propagate the gradient to update the sketch parameters $S_{i} = \\{l_{i} \\text{ or } c_{i}, o_{i}, s_{i}\\}$ . We fol", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25652", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "low EdgeGS [7] to sample an equal number of foreground and background pixels for $L_{1}$ loss calculation.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Adaptive Topology Control", + "text_level": 1, + "bbox": [ + 89, + 128, + 336, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A good edge reconstruction, which could be used for CAD modeling for example, should not only accurately match the 2D edge images, but also provide a compact, well-connected set of edges. Neighboring edges should be connected correctly, overlapping sketches should be merged, and the number of sketches should be as small as possible while ensuring accurate matching of 2D edge images. For this purpose we deploy topology control operations adaptively during optimization. While some existing methods [7] do not consider topology optimization, others [19, 54] try to connect and merge neighboring sketches in the post-processing stage. However, this may be sensitive to noise in the extracted 3D point set and cannot ensure that the resulting edges are still consistent with the input multi-view 2D edge images. Instead, since our method directly optimizes sketch parameters using iterative gradient descent, we can execute topological operations along with the multi-view optimization steps, as shown in Fig. 2(f).", + "bbox": [ + 88, + 151, + 482, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "End-point Merging. In our representation, each sketch has two end-points. If there exist two end-points, one from sketch $S_{i}$ and one from sketch $S_{j}$ , that are sufficiently close (i.e. the point distance is smaller than threshold $th_{connect} = 10mm$ ), we will connect sketch $S_{i}$ and $S_{j}$ by merging those two end-points into one point (see Fig. 2(f-i)). However, this means the number of estimable points could differ from the number of control points $|\\{l\\}| * 2 + |\\{c\\}| * 4$ . To accommodate this, we maintain an estimable point set $P$ , and the control points of each line $l$ and curve $c$ are represented by indexing point set $P$ . This representation also supports our additional topological operations.", + "bbox": [ + 89, + 441, + 483, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overlapping Sketch Merging. As shown in Fig. 2(f-ii), if one sketch is almost entirely covered by another, the smaller one will be merged into the larger one. Specifically, for sketch $S_{i}$ and $S_{j}$ , we sample two point sets $P_{i}$ and $P_{j}$ and calculate the distance between each pair of points. Then for $P_{i}$ , we calculate the ratio $r_{i\\rightarrow j}$ of points that have close neighboring points in $P_{j}$ , where the neighbor threshold is $th_{neighbor} = 10mm$ . If $r_{i\\rightarrow j} > th_{overlap} = 80\\%$ , sketch $S_{i}$ will be merged in sketch $S_{j}$ . If both $r_{i\\rightarrow j}$ and $r_{j\\rightarrow i}$ surpass $th_{overlap}$ , we will choose the sketch with larger overlapping ratio to be merged.", + "bbox": [ + 89, + 641, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Co-linear Sketch Merging. If two lines are co-linear and nearly connected, they should be merged into a longer line (see Fig. 2(f-iii)). The co-linear condition is: two lines should have similar direction ( $th_{dir} = 5^{\\circ}$ ), and the maximum of the projective distance/offset should be small", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "enough ( $th_{offset} = 10mm$ ). In addition, to prevent gaps between two lines, one line will be projected onto the other one to calculate if there is a gap and if the gap is small enough ( $th_{connect}$ ). To merge the lines, we take the two end-points with the furthest distance after projection and delete the other two endpoints from point set $P$ . Additionally, to accelerate this procedure, we use the axis-aligned bounding box of each sketch to quickly filter out all the sketch pairs that cannot meet the merging criterion.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sketch Filtering. After training, we adopt a multi-view filter method to filter sketches that are not visible. We borrow the refinement idea from EMAP [19]. 3D points can be projected onto 2D edge images to see if they are aligned with the 2D edges by checking the edge intensity of the projected pixel position. A point will be labeled as invisible if it is invisible in more than $90\\%$ of the views. For each sketch, we will evaluate the visibility for all of the points sampled from the sketch. If more than $th_{vis} = 50\\%$ of the sampled points are invisible points, the sketch will be removed. When rasterizing sketches, some sketches that should not appear due to occlusion still appear in the rendered image, which sometimes leads to incorrect reconstruction. This filter scheme will help filter out such outlier sketches.", + "bbox": [ + 511, + 253, + 908, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 478, + 732, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unless otherwise specified, we initialize with EdgeGS [7], which supplies a sufficient number of 3D edges to capture object structure. For each scene, we optimize the sketches for 1000 epochs with Adam. We update parameters only once for each epoch, where the loss is accumulated across all the training views. All of the scenes are scaled to fit within a $1\\mathrm{m}^3$ bounding box, and topology operation hyperparameters are set accordingly. Distance thresholds (e.g., $th_{connect}$ ) are set based on the desired precision and we set all of them as $1\\%$ of the scene size $(1m^3)$ . The percentage and angular thresholds (e.g., $th_{overlap}$ and $th_{dir}$ ) work across different scenes without adjustment. All our experiments are implemented on a single RTX A5000 NVIDIA GPU. In average, each scene in ABC-NEF [54] dataset costs about 10 minutes for training, where both initialization and sketch optimization takes about 5 minutes, respectively.", + "bbox": [ + 511, + 501, + 908, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 760, + 648, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 786, + 705, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with other multi-view edge reconstruction methods following the evaluation setup of EdgeGS [7]. The results show that our method achieves the state-of-the-art performance in both accuracy and completeness on the ABC-NEF dataset [18, 54]. Our method also produces highly compact results with fast inference speed.", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25653", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c761d3eac7ac4270f20e443de41fa3d14ed3655c0ad6578d241e8e5f2b3f68d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDetectorModalA↓C↓R5↑R10↑R20↑P5↑P10↑P20↑F5↑F10↑F20↑
LIMAP[23]LSDLine9.918.736.282.387.943.087.693.939.084.390.4
SOLD2Line5.929.664.276.679.688.196.497.972.984.086.7
NEF[54]PiDiNeT†Curve11.916.911.462.091.315.768.596.313.064.093.3
PiDiNeTCurve15.116.511.753.393.912.361.395.812.351.888.7
DexiNedCurve21.915.711.348.393.711.558.991.710.842.176.8
EMAP[19]PiDiNeTEdge9.215.630.275.789.535.679.195.432.477.092.2
DexiNedEdge8.88.956.488.994.862.989.995.759.188.994.9
2DGS-SNEdge8.87.963.590.996.270.491.094.966.390.495.1
EdgeGS[7]PiDiNeTEdge11.710.317.173.983.126.087.292.520.679.386.7
DexiNedEdge9.68.442.491.795.849.194.896.345.293.795.7
2DGS-SNEdge7.47.275.791.096.086.995.996.480.392.995.8
OursPiDiNeTEdge10.411.025.188.694.530.989.695.427.588.594.5
DexiNedEdge8.58.647.593.596.154.196.397.150.394.496.3
2DGS-SNEdge6.85.890.895.797.292.996.396.691.395.496.5
", + "bbox": [ + 119, + 88, + 880, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation on ABC-NEF dataset [54]. Our method achieves the state-of-the-art performance on most metrics. A: Accuracy (mm), C: Completeness (mm), Rx: Recall (threshold $x\\mathrm{\\;{mm}}$ ),Px: Precision (mm), Fx: F-score (mm).", + "bbox": [ + 89, + 311, + 906, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c8d4e47194c27dc9c43e037339f3201485ad20aa9ba071b39da67dcfcf37012.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on the ABC-NEF dataset [54]. Compared to the baselines, our method achieves superior accuracy and completeness, along with a favorable balance between completeness and compactness." + ], + "image_footnote": [], + "bbox": [ + 94, + 356, + 906, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three datasets: ABC-NEF [18, 54] DTU [1], and Replica [43]. The ABC-NEF dataset [18, 54] comprises 115 CAD models, each containing 50 RGB images of the target object and a parametric edge file providing ground truth 3D edges. Following EMAP [19] and EdgeGS [7], we use 82 CAD models from ABC-NEF, excluding the ones with cylinders and spheres since smooth convex surfaces will produce silhouette edges that are not consistent in multi-views. For 2D edge image generation, we use PiDiNet [44], DexiNed [33], and our proposed 2DGS-SN as 2D edge detectors. The", + "bbox": [ + 88, + 727, + 485, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DTU dataset [1] contains multi-view images of real-world scenes. We follow EdgeGS [7] and evaluate on 6 DTU scenes. We use edge images and the pseudo ground truth edges provided by EMAP [19]. The Replica dataset [43] consists of multiple indoor scenes. We follow EdgeGS [7] to conduct qualitative evaluation on Replica [43].", + "bbox": [ + 511, + 727, + 908, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with five baseline methods, including the state-of-the-art 3D line mapping method (LIMAP [23]), and several methods based on volumetric differentiable rendering (NEF [54], NEAT [51],", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25654", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acb3124d7edac71cac255d2bffe7da517135cd4a7ea61b907a43a6402e83c6ca.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on the DTU and Replica datasets [1, 43]. Enum is the edge number and edges are visualized with varied colors. Compared to the baselines, our method achieves higher compactness and retains faithful completeness." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 439, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db3d13ef0cdde8892a7c652c6e24f491c908fb5c181d467651da1c94be2866ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 88, + 903, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EMAP [19], EdgeGS [7]). Following EMAP and EdgeGS, we do not evaluate NEAT on the ABC-NEF dataset since NEAT often fails to train for texture-less objects.", + "bbox": [ + 89, + 286, + 483, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. Following the metrics used in baseline methods [7, 19], we evaluate Accuracy (A) and Completeness (C) in millimeters, and Recall (R), Precision (P), and F-score (F) as percentages at three thresholds (5mm, 10mm, 20mm). See supplementary material for more details.", + "bbox": [ + 89, + 352, + 483, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 439, + 184, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on ABC-NEF [54]. The qualitative and quantitative evaluation results on ABC-NEF dataset are shown in Fig. 3 and Tab. 1, respectively. Overall, the proposed method achieves a new state-of-the-art across most of the metrics. To the best of our knowledge, SketchSplat is the first method to achieve both accuracy (A) and completeness (C) below $7.0\\mathrm{mm}$ , while also exceeding $90\\%$ in recall (R5), precision (P5), and F-score (F5).", + "bbox": [ + 89, + 460, + 482, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 shows LIMAP [23] achieves the highest accuracy due to its robust line mapping method. However, its completeness and recall are low, as LIMAP struggles with curve reconstruction and consistent edge detection results (Fig. 3(b)). NEF [54] produces more complete results in Fig. 3(c), but performs poorly on metrics as its fitting process is sensitive to noise in the extracted 3D points. EMAP [19] outperforms NEF with its unsigned distance field modeling and point shifting mechanism. However, Fig. 3(d) shows that EMAP still generates duplicated edges and misses some edges. With PidiNeT and DexiNed as 2D detector, EdgeGS [7] surpasses EMAP in recall and precision at $10mm$ and $20mm$ thresholds, but the trend reverses for measurements under $5mm$ . This is due to the thickness of 2D edges introduces bias in smaller distance evaluations. When using our detector 2DGS-SN, EdgeGS performs better than EMAP to a large margin under $5mm$ metrics. As shown in Fig. 3, our SketchSplat achieves the highest precision and completeness compared to all the baselines.", + "bbox": [ + 89, + 583, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab.1, the proposed 2DGS-SN significantly boosts the performance of EMAP, EdgeGS, and Sketch-", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a08edc33a78ddf092bf4501ada7bb5fe9c1918854d417c52ff5ccfe20e439eea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodE-numTime (h)
NEF [54]22.91:26
EMAP [19]45.82:30
EdgeGS [7]140.80:05
ours44.30:10
", + "bbox": [ + 588, + 282, + 834, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of the number of reconstructed edges (E-num) and training time on ABC-NEF dataset.", + "bbox": [ + 511, + 376, + 905, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Splat, addressing the bias from PiDiNet[44] and Dex-iNed [33]. Notably, across all edge detectors, our method consistently achieves higher Recall, Accuracy, and F-score than EdgeGS [7], demonstrating effectiveness of Sketch-Splat in aligning 3D edges with 2D edge images.", + "bbox": [ + 511, + 431, + 906, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on DTU [1] and Replica [43]. Fig. 4 presents qualitative comparisons on the DTU and Replica datasets. Compared to the baselines, our method consistently produces accurate and complete results using significantly fewer edges, demonstrating a strong balance between completeness and compactness. Although the ground-truth edges in the DTU scenes are not sufficiently accurate for reliable quantitative evaluation [7], our method still achieves metrics comparable to state-of-the-art methods EMAP and EdgeGS (see supplementary material for more details).", + "bbox": [ + 511, + 527, + 908, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speed Analysis. The initialization with EdgeGS costs about 5 minutes. For optimization part, thanks to the efficiency of rasterization-based optimization, our method achieves comparable speed (about 5 minute per scene) to EdgeGS while being significantly faster than other neural network-based approaches [19, 54], as shown in Tab. 2.", + "bbox": [ + 511, + 699, + 908, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compactness. EdgeGS achieves higher recall than NEF and EMAP due to its strict edge fitting process, which produces numerous short edges, covering more edge regions as the number of edges increases. Instead, our method incorporates adaptive topology control to reduce redundant edges, as shown in Tab. 2, Fig. 4, and Fig. 5(a,b,d). This", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25655", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/680192c8ba1f43ce645fd6dbf40c92af000d6b8df06cc92819349e85a2c039ee.jpg", + "image_caption": [ + "(a) Initialization (EdgeGS)" + ], + "image_footnote": [], + "bbox": [ + 94, + 97, + 251, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4a7d33a00db868f194705e7ddcf51748eeaa53a735c5d08707d89630f5afe73.jpg", + "image_caption": [ + "(b) Ours (w/o Topology Control)" + ], + "image_footnote": [], + "bbox": [ + 256, + 98, + 415, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e4bd38b7d8a0e1d516eb3ee29ff5294fd52ae5bad1940b4e7ed31906c8c342b.jpg", + "image_caption": [ + "(c) Ours (w/o 2DGS-SN)" + ], + "image_footnote": [], + "bbox": [ + 418, + 98, + 576, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b472d9b45f4c50771158f5ae1d104fbceb38c9cc778fca995a8c6d181ac0574b.jpg", + "image_caption": [ + "(d) Ours", + "Figure 5. Ablation study on the key components of SketchSplat. Removing key components leads to noisier or redundant reconstructions." + ], + "image_footnote": [], + "bbox": [ + 581, + 98, + 738, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5dbe9e8cf8a17c27180f39223ca9140be3d9da8b7873a077494187c19545a86f.jpg", + "image_caption": [ + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 746, + 98, + 903, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2b5c2be289eb38030d72764ac3150f777342d96a1b13e9d18e620360f401afa1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodA↓C↓R5↑P5↑F5↑E-num
Ours6.85.890.892.991.344.28
w/o 2DGS-SN8.58.647.554.150.333.52
w/o Merge6.65.791.892.791.7140.72
w/o Filter6.85.890.992.991.344.37
", + "bbox": [ + 96, + 323, + 478, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/980bf2acaa6efc5a04361030ee56a67c267381d7c3ff4470612dcd0d41180183.jpg", + "table_caption": [ + "Table 3. Ablation on the proposed components using ABC-NEF. w/o 2DGS-SN denotes SketchSplat trained with edge images detected by DexiNed, while the others are trained with 2DGS-SN." + ], + "table_footnote": [], + "table_body": "
MethodA↓C↓R5↑P5↑F5↑E-num
EdgeGS7.47.275.786.980.3140.76
Ours (EdgeGS)6.85.890.892.991.344.28
EMAP8.87.963.570.466.348.71
Ours (EMAP)6.96.288.890.889.331.70
", + "bbox": [ + 94, + 465, + 480, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation on initialization methods using ABC-NEF. Ours (EdgeGS) and Ours (EMAP) denote our method initialized with EdgeGS and EMAP, respectively. 2DGS-SN is chosen as 2D edge detector.", + "bbox": [ + 89, + 551, + 482, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mechanism helps SketchSplat achieve higher compactness, while maintaining the highest accuracy and completeness.", + "bbox": [ + 89, + 636, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation", + "text_level": 1, + "bbox": [ + 89, + 680, + 194, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key Components. We use the ABC-NEF dataset to conduct an ablation study on the key contributions of our method, with results presented in Fig. 5 and Tab. 3. Fig. 5(b,d) shows that without topology control, noisy edge segments appear around the ground truth edges, with some extending beyond the object boundary. Incorporating topology control, particularly merging operations, effectively reduces the required edge count while maintaining comparable performance on other metrics (see Tab. 3). In addition, compared to edges reconstructed using DexiNed 2D edge images (Fig. 5(c)), our proposed 2DGS-SN better aligns edges with the ground truth (Fig. 5(d)), improving all metrics by effectively reducing bias in 2D edge images.", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization Methods. In our main experiments, we use EdgeGS for initialization. However, we demonstrate SketchSplat can also achieve strong results when initialized with alternative methods like EMAP. As shown in Tab. 4, our approach consistently outperforms the initialization methods by a large margin and achieves comparable results across different initializations.", + "bbox": [ + 511, + 325, + 906, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 450, + 633, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SketchSplat, a method for parametric 3D edge reconstruction from calibrated multi-view images. Unlike previous approaches that rely on 3D edge point reconstruction and edge fitting, SketchSplat directly optimizes edges through multi-view sketch splatting, ensuring accurate alignment with input 2D images. A series of topological operations are proposed to adaptively refine the edge structure during training, which reduces redundancy while maintaining high accuracy and completeness. Additionally, an improved 2D edge detection method is adopted to achieve higher reconstruction quality. Experiments show that SketchSplat achieves the state-of-the-art accuracy and completeness on the CAD dataset while maintaining compact results and efficient training speed.", + "bbox": [ + 511, + 479, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research is based in part upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODN, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.", + "bbox": [ + 511, + 719, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25656", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 6, 7", + "[2] Caroline Baillard, Cordelia Schmid, Andrew Zisserman, and Andrew Fitzgibbon. Automatic line matching and 3d reconstruction of buildings from multiple views. In ISPRS Conference on Automatic Extraction of GIS Objects from Digital Imagery, pages 69-80, 1999. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 2", + "[5] Andrea Bignoli, Andrea Romanoni, Matteo Matteucci, and Politecnico di Milano. Multi-view stereo 3d edge reconstruction. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 867-875. IEEE, 2018. 2", + "[6] Manmohan Chandraker, Jongwoo Lim, and David Kriegman. Moving in stereo: Efficient structure and motion using lines. In 2009 IEEE 12th International Conference on Computer Vision, pages 1741-1748. IEEE, 2009. 2", + "[7] Kunal Chelani, Assia Benbihi, Torsten Sattler, and Fredrik Kahl. Edgegaussians-3d edge mapping via gaussian splatting. arXiv preprint arXiv:2409.12886, 2024. 1, 2, 3, 4, 5, 6, 7", + "[8] Wentao Cheng, Sheng Yang, Maomin Zhou, Ziyuan Liu, Yiming Chen, and Mingyang Li. Road mapping and localization using sparse semantic visual features. IEEE robotics and automation letters, 6(4):8118-8125, 2021. 1", + "[9] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3", + "[10] Hyelim Choi, Minji Lee, Jiseock Kang, and Dongjun Lee. Online 3d edge reconstruction of wiry structures from monocular image sequences. IEEE Robotics and Automation Letters, 8(11):7479-7486, 2023. 1", + "[11] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2", + "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Improving sparse 3d models for man-made environments using line-based 3d reconstruction. In 2014 2nd International Conference on 3D Vision, pages 535-542. IEEE, 2014. 2", + "[13] Petr Hruby, Shaohui Liu, Rémi Pautrat, Marc Pollefeys, and Daniel Barath. Handbook on leveraging lines for two-view relative pose estimation. In 2024 International Conference on 3D Vision (3DV), pages 376-386. IEEE, 2024. 1", + "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "curate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 4", + "[15] Kahl and August. Multiview reconstruction of space curves. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 1017-1024. IEEE, 2003. 2", + "[16] Jeremy Yermiyahou Kaminski, Michael Fryers, Amnon Shashua, and Mina Teicher. Multiple view geometry of non-planar algebraic curves. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 181-186. IEEE, 2001. 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4", + "[18] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 5, 6", + "[19] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21219-21229, 2024. 1, 2, 3, 4, 5, 6, 7", + "[20] Pu Li, Jianwei Guo, Huibin Li, Bedrich Benes, and Dong-Ming Yan. Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2024. 1", + "[21] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 3, 4", + "[22] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2", + "[23] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 2, 3, 6, 7", + "[24] Yuncai Liu, Thomas S Huang, and Olivier D Faugeras. Determination of camera location from 2-d to 3-d line and point correspondences. IEEE Transactions on pattern analysis and machine intelligence, 12(1):28-37, 1990. 1", + "[25] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2", + "[26] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision, 124:65-79, 2017. 2", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2", + "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2", + "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 2, 3", + "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023.", + "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 2", + "[33] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 2, 3, 6, 7", + "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 2", + "[35] Zhijian Qiao, Zehuan Yu, Huan Yin, and Shaojie Shen. Online monocular lane mapping using catmull-rom spline. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7179–7186. IEEE, 2023. 1", + "[36] Tong Qin, Yuxin Zheng, Tongqing Chen, Yilun Chen, and Qing Su. A light-weight semantic map for visual localization towards autonomous driving. In 2021 IEEE international conference on robotics and automation (ICRA), pages 11248-11254. IEEE, 2021. 1", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3", + "[38] Luc Robert and Olivier D Faugeras. Curve-based stereo: Figural continuity and curvature. In Proceedings. 1991 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 57-58. IEEE Computer Society, 1991. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853. IEEE, 2006. 2", + "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3", + "[41] Paul Smith, Ian Reid, and Andrew J Davison. Real-time monocular slam with straight lines. In BMVC, pages 17-26, 2006. 1", + "[42] Irwin Sobel, Gary Feldman, et al. A 3x3 isotropic gradient operator for image processing. a talk at the Stanford Artificial Project in, 1968:271-272, 1968. 4", + "[43] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7", + "[44] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 2, 3, 6, 7", + "[45] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3", + "[46] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4): 722-732, 2008. 3", + "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2", + "[48] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 2", + "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2", + "[50] Xiaolong Wu, Patricio A Vela, and Cedric Pradalier. Robust monocular edge visual odometry through coarse-to-fine data association. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4923-4929. IEEE, 2020. 1", + "[51] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. Neat: Distilling 3d wireframes from neural attraction fields. In Proceedings of" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25658", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19968-19977, 2024. 2, 3, 6", + "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2", + "[53] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 2024. 4", + "[54] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 2, 3, 5, 6, 7", + "[55] Haiyang Ying, Baowei Jiang, Jinzhi Zhang, Di Xu, Tao Yu, Qionghai Dai, and Lu Fang. Parf: Primitive-aware radiance fusion for indoor scene novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17706-17716, 2023. 2", + "[56] Chao Zhang, Romain Pinquié, Arnaud Polette, Gregorio Carasi, Henri De Charnace, and Jean-Philippe Pernot. Automatic 3d cad models reconstruction from 2d orthographic drawings. Computers & Graphics, 114:179–189, 2023. 1", + "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + ], + "bbox": [ + 91, + 90, + 482, + 474 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25659", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_model.json b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b189dd3487fe6eacacb8cdd1cb8b65ccefcf7e03 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_model.json @@ -0,0 +1,2081 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.782, + 0.177 + ], + "angle": 0, + "content": "SketchSplat: 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.204, + 0.644, + 0.24 + ], + "angle": 0, + "content": "Haiyang Ying, Matthias Zwicker \nUniversity of Maryland, College Park" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.242, + 0.668, + 0.257 + ], + "angle": 0, + "content": "https://oceanying.github.io/SketchSplat" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.486, + 0.766 + ], + "angle": 0, + "content": "Edges are one of the most basic parametric primitives to describe structural information in 3D. In this paper, we study parametric 3D edge reconstruction from calibrated multi-view images. Previous methods usually reconstruct a 3D edge point set from multi-view 2D edge images, and then fit 3D edges to the point set. However, noise in the point set may cause gaps among fitted edges, and the recovered edges may not align with input multi-view images since the edge fitting depends only on the reconstructed 3D point set. To mitigate these problems, we propose SketchSplat, a method to reconstruct accurate, complete, and compact 3D edges via differentiable multi-view sketch splatting. We represent 3D edges as sketches, which are parametric lines and curves defined by attributes including control points, scales, and opacity. During reconstruction, we iteratively sample Gaussian points from a set of sketches and rasterize the Gaussians onto 2D edge images. Then the gradient of the image loss can be back-propagated to optimize the sketch attributes. Our method bridges 2D edge images and 3D edges in a differentiable manner, which ensures that 3D edges align well with 2D images and leads to accurate and complete results. We also propose a series of adaptive topological operations to reduce redundant edges and apply them along with the sketch optimization, yielding a more compact reconstruction. Finally, we contribute an accurate 2D edge detector that improves the performance of both ours and existing methods. Experiments show that our method achieves state-of-the-art accuracy, completeness, and compactness on a benchmark CAD dataset." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.798, + 0.222, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Edges serve as one of the most fundamental elements of visual perception, playing a crucial role in understanding the geometry and structure of the 3D world. Unlike point-based representations, continuous edges offer a more compact and perceptually aligned way to describe 3D structures. As a" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.288, + 0.907, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Figure 1. We propose SketchSplat, a multi-view 3D edge reconstruction method based on differentiable sketch splatting. Previous methods [7, 19] first recover edge points and then fit them to 3D edges, which is sensitive to noise and often leads to incomplete results. Instead, we sample Gaussian points from 3D sketches and rasterize them to compute a rendering loss, enabling direct edge parameter optimization under image supervision. Additionally, we design adaptive topology control methods to refine edge structures during training and contribute an improved 2D edge detection method for more accurate reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "cornerstone of computer vision and graphics, 3D edge reconstruction underpins a wide range of applications, including CAD modeling [10, 20, 56], SLAM [13, 24, 41, 50], and autonomous driving [8, 35, 36]. Despite its significance, achieving high-quality 3D edge reconstructions remains a challenging task due to the inherent difficulty of ensuring" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "the accuracy and completeness of reconstructed edges and maintaining the connectivity of the reconstructed structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.482, + 0.244 + ], + "angle": 0, + "content": "Traditional methods [5, 12, 23, 48] usually detect 2D line segments in each view. Then the 2D lines are lifted to 3D via line matching and triangulation across views. While these methods achieve high accuracy and efficiency, multiview consistent detection and robust line matching are difficult to achieve, leading to incomplete results. In addition, these methods are usually not capable of handling lines and curves simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.483, + 0.504 + ], + "angle": 0, + "content": "In recent years, differentiable rendering has emerged as a promising technique to reconstruct 3D contents for novel view synthesis [3, 27, 28, 55], surface reconstruction [11, 22, 25, 47], and animation [29, 34, 49]. In the field of edge reconstruction, several works introduced the idea of differentiable rendering to recover edge information [7, 19, 54]. Instead of detecting parametric 2D edge segments explicitly, they leverage pixel-wise 2D edge intensity images [33, 44] as a supervision signal and optimize a point-based 3D edge representation via differentiable multiview rendering, resulting in an extracted 3D edge point set. Parametric 3D edges are then obtained by fitting lines and curves to the extracted 3D points. Since the edge-fitting step relies purely on the recovered 3D edge points but not the input 2D edge images, the fitted 3D edges may not align well with the 2D edge images. In addition, noise in the 3D edge points usually leads to discontinuous and fragmented edges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.482, + 0.612 + ], + "angle": 0, + "content": "To address these limitations, we propose SketchSplat, a differentiable edge reconstruction framework that directly optimizes parametric 3D edges by fitting them to the input 2D edge images using differentiable rendering. Our experimental results demonstrate that this leads to more accurate, complete, and compact reconstructions, more faithfully reflecting the original connectivity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Firstly, we propose to model 3D edges as sketches, which are parametric lines and curves with control points, scales, and opacity. These sketches can be initialized from existing methods [7, 19, 54], and can be sampled as a set of 3D Gaussian points for differentiable rendering. To optimize the sketches, we sample Gaussian points from the sketches and rasterize them in each view to calculate the rendering loss between the rendered and the ground-truth edge image. Then the error can be back-propagated to update the sketch attributes. Since we optimize our sketch parameters directly using the 2D edge images, the resulting edges will align well with the 2D edges in the input images. To further improve the compactness of the reconstructed edges, we leverage a series of topological operations to connect, merge, and filter sketches during optimization. Finally, we observe that one common limitation of existing methods [7, 19, 54] lies in the inaccuracy of 2D edge images. To mitigate this, we propose an accurate and robust edge detection method, which is validated to improve both" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.868, + 0.107 + ], + "angle": 0, + "content": "our approach and existing methods by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.109, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Experiments demonstrate that our method achieves state-of-the-art performance on the ABC-NEF dataset [54], surpassing previous approaches by over \\(10\\%\\) in F-score. In summary, key innovations of our method include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.172, + 0.905, + 0.233 + ], + "angle": 0, + "content": "- Differentiable Parametric Optimization: We contribute a pipeline that directly optimizes parametric edge representations via differentiable multi-view rasterization, leading to SOTA accuracy and completeness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.233, + 0.905, + 0.292 + ], + "angle": 0, + "content": "- Adaptive Topological Control: We propose a series of topological operations to further enhance edge continuity and reduce redundant edges during training, yielding more compact reconstruction results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.905, + 0.338 + ], + "angle": 0, + "content": "- Accurate 2D Edge Detection: We propose a more accurate 2D edge detection method, which shows large improvements for both our approach and existing methods." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.172, + 0.905, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.354, + 0.655, + 0.37 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Line Matching and Triangulation. Conventional approaches reconstruct 3D lines [2, 4-6, 12, 26] or curves [15, 16, 38] by enforcing multi-view geometric constraints. Given detected 2D edge segments and their descriptors, they match the line descriptors across multiple views and perform triangulation to lift 2D edges into 3D. However, these methods face challenges in consistently detecting edges and robustly matching lines, especially when occlusions cause endpoint inconsistencies across views. Despite notable advancements by involving structural [39, 48] or learning-based priors [23, 30-32], line detection and matching remain critical bottlenecks, limiting overall performance. While most of the methods can only handle a single type of edges, in this paper we aim at recovering both lines and curves." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Differentiable Rendering Methods. The advancement of differentiable rendering opens up a new way for 3D reconstruction [22, 27, 47]. Inspired by NeRF [27], NEF [54] proposes to optimize an implicit neural radiance field with a multi-view rendering loss. After training, 3D edge points are extracted from the radiance field and curves are fitted to these points. NEAT [51] builds on VolSDF [52] and uses a signed distance field to represent both object surfaces and endpoints of 3D lines. The surface-based modeling approach restricts this method to objects with rich texture. EMAP [19] adopts an unsigned distance function (UDF), where sampled points can be displaced towards the zero-level set of the UDF to be more accurate. While methods based on neural field are time consuming to optimize, EdgeGS [7] represents 3D edge fields with explicit Gaussian points and leverages Gaussian splatting [17] to render and optimize Gaussian points, leading to more efficient edge reconstruction. A common issue of these methods" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25650" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "is that the edge fitting process relies only on the extracted 3D edge points, hence noise in the reconstructed 3D points may lead to inaccurate and incomplete edges. Instead, we represent 3D edges as sketches defined parametrically by control points, scales, and opacity, and directly optimize these parameters via differentiable rendering. By leveraging sketch sampling and rasterization, our differentiable rendering aligns 3D edges with 2D edge images, resulting in more accurate and complete reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.55 + ], + "angle": 0, + "content": "Differentiable Vector Graphics Optimization. DiffVG [21] proposes a differentiable rasterizer that can optimize 2D vector graphics with 2D image supervision in a differentiable manner. However, the differentiation process relies on distance measurements between the rasterized curves and pixels, which needs to be custom tailored for each type of curve. 3Doodle [9] extends the idea of DiffVG [21] to reconstruct 3D curve-based object abstractions using multiple input views. They optimize a set of Bezier curves by minimizing the perceptual loss between the rasterized image and multi-view CLIP feature maps [37, 45]. Since CLIP features only contain abstract information, 3Doodle is not suitable for recovering accurate edges of objects like CAD models. Moreover, they approximate the projection of Bezier curves using an orthographic camera model. Instead, we get inspiration from 3DGS [17] and propose to optimize 3D sketches via sampling and rasterization of Gaussian points sampled from sketch parameters, which can be easily extended to different types of curves and achieves accurate 3D edge reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.429, + 0.583 + ], + "angle": 0, + "content": "3. Preliminaries: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.483, + 0.775 + ], + "angle": 0, + "content": "We leverage 3D Gaussian splatting (3DGS) [17] for initialization and differentiable sketch rendering. 3DGS proposes to model 3D scenes with a collection of anisotropic 3D Gaussian points. Each point is defined as a Gaussian distribution \\( G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)} \\), where \\( \\mu \\in R^3 \\) is the center position or mean, and \\( \\Sigma \\in R^{3\\times 3} \\) is the covariance matrix. Each Gaussian has an opacity \\( o \\in R \\) and a view-dependent color value \\( c(d) = SH(d) \\in R^3 \\) modeled with spherical harmonics. Following Zwicker et al. [57], the Gaussians can be rendered to 2D images using an efficient volume rendering approach, where the color for each camera ray \\( r \\) is" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.786, + 0.483, + 0.829 + ], + "angle": 0, + "content": "\\[\nc (r) = \\sum_ {i \\in N} c _ {i} (d (r)) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\alpha_ {i} = o _ {i} g _ {i} (r), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\( g_{i}(r) \\) is the line integral of the 3D Gaussian \\( G_{i} \\) along ray \\( r \\), and the indices \\( i \\) are ordered in front to back manner. By calculating the gradient of a loss \\( \\mathcal{L} \\) between the rendered and ground truth image, the parameters of all Gaussians can" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "be optimized using gradient descent. A typical loss function is a combination of an \\( L_{1} \\) and a perceptual loss such as" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.135, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {D - S S I M}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.164, + 0.907, + 0.254 + ], + "angle": 0, + "content": "Given input images, the initial Gaussians are usually obtained from SfM [40] methods. During training, several operations are adopted to adaptively control the density and distribution of Gaussians. For example, Gaussians with low opacity will be culled, and Gaussians with high gradient will be duplicated or split to densify specific regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.906, + 0.315 + ], + "angle": 0, + "content": "In SketchSplat, we leverage 3DGS [17] as an intermediary to bridge 2D edge images and 3D parametric edges, enabling the optimization of 3D edges through differentiable image rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.33, + 0.605, + 0.345 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Given multi-view posed RGB images as input, our target is to recover the parametric edges for the 3D objects. Fig. 2 provides an overview of our approach. We firstly compute an edge image for each RGB image using our proposed 2D edge detector. Then we represent edges as 3D parametric sketches, including both lines \\( l \\in \\mathbb{R}^{2 \\times 3} \\) and third-order Bézier curves with four 3D control points \\( c \\in \\mathbb{R}^{4 \\times 3} \\). We initialize a set of sketches from existing edge reconstruction methods, and treat their control points, along with persketch scales and opacity, as estimizable parameters. During optimization, we use a sample-based method to convert sketches to Gaussian points, which can be rasterized onto 2D images to calculate the image loss. We back-propagate the gradient of the loss to optimize the sketch parameters. In addition, we introduce a set of topological operations to adaptively refine the sketches during training, yielding more accurate and compact edge reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.727, + 0.637 + ], + "angle": 0, + "content": "4.1. Edge Image Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Previously, some methods [23, 51] extract edges with line detectors like LSD [46] and SOLD2 [30]. While these methods produce accurate parametric lines in 2D, they may fail to handle curve regions and result in incomplete results. Recent methods [7, 19, 54] use neural network-based edge detectors [33, 44] to detect edge intensity maps as supervision signal. However, since these detectors are not trained specifically on CAD-like objects, they may miss some significant edges in 2D or wrongly recover shading textures as real edges. Finally, we observed that these detectors introduce bias by consistently shifting detected edges slightly from certain real edges, which leads to multi-view inconsistency in 3D edge reconstruction (see supplementary). We address these issues by proposing a new edge detection method, which leverages geometric cues (e.g., per-pixel depth and normal estimation of 2D RGB images) to produce more accurate edge detection results." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.315, + 0.321, + 0.328 + ], + "angle": 0, + "content": "(A) Sketch initialization" + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.315, + 0.807, + 0.329 + ], + "angle": 0, + "content": "(B) Differentiable Multi-view Sketch Optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Overview of SketchSplat pipeline: (a) We extract edge images using a novel approach (Sec. 4.1). (b) We then obtain initial edge points using EdgeGS [7] and (c) fit parametric sketches defined as lines and third-order Bézier curves. The core step of our approach is to optimize the sketches by (d) sampling and rasterizing them in a differentiable manner and back-propagating gradients of the image loss to (e) sketch parameters. In addition, during training we apply (f) a set of topological operations to improve compactness and connectivity of the reconstructed sketches. (g) After training, we take the optimized geometric parameters of sketches as the reconstructed 3D edges." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.484, + 0.651 + ], + "angle": 0, + "content": "For each RGB image, we combine information from three parts: foreground mask \\(A\\), depth map \\(D\\), and normal map \\(N\\). The foreground mask can be extracted using the image alpha-channel (if available), a thresholding operation or possibly using semantic segmentation, providing object boundary information. We obtain a depth image for each view by reconstructing the scene with 2DGS [14] and rendering depth maps from each view. We use a normal estimator [53] to generate normal maps \\(N\\) for each view. Then we calculate gradient maps \\(g(D)\\) and \\(g(N)\\) with the Sobel operator [42], threshold them with \\(t_d\\) and \\(t_n\\), and combine them via the OR operation, followed by a Gaussian filter \\(G_f(\\cdot)\\) to smooth the boundaries of the edges. In summary, we call our edge detector 2DGS-SN, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.661, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\nE = G _ {f} * (A | (g (D) > t _ {d}) | (g (N) > t _ {n})). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.72 + ], + "angle": 0, + "content": "In Sec. 5 we show that this simple approach improves existing methods [7] on CAD-like objects by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.254, + 0.745 + ], + "angle": 0, + "content": "4.2. Sketch Splatting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Representation. We represent edges as 3D parametric sketches, including both lines \\( l \\in \\mathbb{R}^{2 \\times 3} \\) and third-order Bézier curves with control points \\( c \\in \\mathbb{R}^{4 \\times 3} \\). This is the same as the output format of EdgeGS[7]. While EdgeGS optimizes 3D point positions first and then fits 3D lines and curves to the 3D points, we optimize the parameters of lines and curves directly via multi-view differentiable rasterization. In addition to the estimizable geometry parameters \\( (l \\) or \\( c) \\), we equip each sketch with an estimizable opacity \\( o \\in \\mathbb{R} \\) and a local scale \\( s \\in \\mathbb{R}^3 \\) to model the sketch intensity" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.47 + ], + "angle": 0, + "content": "and thickness, respectively. These attributes will support Gaussian sampling and rasterization in the next steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Sketch Initialization. Unlike point-based methods [7, 19], it is non-trivial to initialize sketches directly from multi-view edge images. Instead, inspired by 3DGS [17], we first initialize sketches with existing point-based methods like EdgeGS [7]. By default, we select EdgeGS because it provides a sufficient number of short 3D edges to cover the objects, and achieves the highest completeness among previous methods, despite some fragmentation and noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Differentiable Sketch Optimization. Our key idea is to bridge the gap between 2D edge images and 3D sketches via differentiable rendering and gradient descent optimization, which significantly improves sketch reconstruction over initialization method EdgeGS [7]. Inspired by 3DGS [17] and DiffVG [21], we choose 3D Gaussians as an intermediate representation to achieve differentiable sketch rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "For each gradient descent step, we first sample points from each sketch with a constant step size (5mm for all experiments), resulting in a point set \\( P \\). Each point will be assigned the opacity and scale value of the corresponding sketch. To model the orientation, the main direction of each point is sampled as the local tangent direction of the sketch. Then we follow 3DGS[17] to rasterize the Gaussian points onto a chosen 2D view to generate a rendered edge image \\( E^{*} \\). We calculate the \\( L_{1} \\) loss between \\( E^{*} \\) and the ground truth edge image \\( E_{gt} \\), and back-propagate the gradient to update the sketch parameters \\( S_{i} = \\{l_{i} \\text{ or } c_{i}, o_{i}, s_{i}\\} \\). We fol" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "low EdgeGS [7] to sample an equal number of foreground and background pixels for \\( L_{1} \\) loss calculation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.13, + 0.338, + 0.146 + ], + "angle": 0, + "content": "4.3. Adaptive Topology Control" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.152, + 0.483, + 0.425 + ], + "angle": 0, + "content": "A good edge reconstruction, which could be used for CAD modeling for example, should not only accurately match the 2D edge images, but also provide a compact, well-connected set of edges. Neighboring edges should be connected correctly, overlapping sketches should be merged, and the number of sketches should be as small as possible while ensuring accurate matching of 2D edge images. For this purpose we deploy topology control operations adaptively during optimization. While some existing methods [7] do not consider topology optimization, others [19, 54] try to connect and merge neighboring sketches in the post-processing stage. However, this may be sensitive to noise in the extracted 3D point set and cannot ensure that the resulting edges are still consistent with the input multi-view 2D edge images. Instead, since our method directly optimizes sketch parameters using iterative gradient descent, we can execute topological operations along with the multi-view optimization steps, as shown in Fig. 2(f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.624 + ], + "angle": 0, + "content": "End-point Merging. In our representation, each sketch has two end-points. If there exist two end-points, one from sketch \\( S_{i} \\) and one from sketch \\( S_{j} \\), that are sufficiently close (i.e. the point distance is smaller than threshold \\( th_{connect} = 10mm \\)), we will connect sketch \\( S_{i} \\) and \\( S_{j} \\) by merging those two end-points into one point (see Fig. 2(f-i)). However, this means the number of estimable points could differ from the number of control points \\( |\\{l\\}| * 2 + |\\{c\\}| * 4 \\). To accommodate this, we maintain an estimable point set \\( P \\), and the control points of each line \\( l \\) and curve \\( c \\) are represented by indexing point set \\( P \\). This representation also supports our additional topological operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Overlapping Sketch Merging. As shown in Fig. 2(f-ii), if one sketch is almost entirely covered by another, the smaller one will be merged into the larger one. Specifically, for sketch \\( S_{i} \\) and \\( S_{j} \\), we sample two point sets \\( P_{i} \\) and \\( P_{j} \\) and calculate the distance between each pair of points. Then for \\( P_{i} \\), we calculate the ratio \\( r_{i\\rightarrow j} \\) of points that have close neighboring points in \\( P_{j} \\), where the neighbor threshold is \\( th_{neighbor} = 10mm \\). If \\( r_{i\\rightarrow j} > th_{overlap} = 80\\% \\), sketch \\( S_{i} \\) will be merged in sketch \\( S_{j} \\). If both \\( r_{i\\rightarrow j} \\) and \\( r_{j\\rightarrow i} \\) surpass \\( th_{overlap} \\), we will choose the sketch with larger overlapping ratio to be merged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Co-linear Sketch Merging. If two lines are co-linear and nearly connected, they should be merged into a longer line (see Fig. 2(f-iii)). The co-linear condition is: two lines should have similar direction (\\(th_{dir} = 5^{\\circ}\\)), and the maximum of the projective distance/offset should be small" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "enough (\\( th_{offset} = 10mm \\)). In addition, to prevent gaps between two lines, one line will be projected onto the other one to calculate if there is a gap and if the gap is small enough (\\( th_{connect} \\)). To merge the lines, we take the two end-points with the furthest distance after projection and delete the other two endpoints from point set \\( P \\). Additionally, to accelerate this procedure, we use the axis-aligned bounding box of each sketch to quickly filter out all the sketch pairs that cannot meet the merging criterion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.909, + 0.466 + ], + "angle": 0, + "content": "Sketch Filtering. After training, we adopt a multi-view filter method to filter sketches that are not visible. We borrow the refinement idea from EMAP [19]. 3D points can be projected onto 2D edge images to see if they are aligned with the 2D edges by checking the edge intensity of the projected pixel position. A point will be labeled as invisible if it is invisible in more than \\(90\\%\\) of the views. For each sketch, we will evaluate the visibility for all of the points sampled from the sketch. If more than \\(th_{vis} = 50\\%\\) of the sampled points are invisible points, the sketch will be removed. When rasterizing sketches, some sketches that should not appear due to occlusion still appear in the rendered image, which sometimes leads to incorrect reconstruction. This filter scheme will help filter out such outlier sketches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.733, + 0.495 + ], + "angle": 0, + "content": "4.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Unless otherwise specified, we initialize with EdgeGS [7], which supplies a sufficient number of 3D edges to capture object structure. For each scene, we optimize the sketches for 1000 epochs with Adam. We update parameters only once for each epoch, where the loss is accumulated across all the training views. All of the scenes are scaled to fit within a \\(1\\mathrm{m}^3\\) bounding box, and topology operation hyperparameters are set accordingly. Distance thresholds (e.g., \\(th_{connect}\\)) are set based on the desired precision and we set all of them as \\(1\\%\\) of the scene size \\((1m^3)\\). The percentage and angular thresholds (e.g., \\(th_{overlap}\\) and \\(th_{dir}\\)) work across different scenes without adjustment. All our experiments are implemented on a single RTX A5000 NVIDIA GPU. In average, each scene in ABC-NEF [54] dataset costs about 10 minutes for training, where both initialization and sketch optimization takes about 5 minutes, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.649, + 0.778 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.707, + 0.804 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We compare our method with other multi-view edge reconstruction methods following the evaluation setup of EdgeGS [7]. The results show that our method achieves the state-of-the-art performance in both accuracy and completeness on the ABC-NEF dataset [18, 54]. Our method also produces highly compact results with fast inference speed." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25653" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.882, + 0.303 + ], + "angle": 0, + "content": "
MethodDetectorModalA↓C↓R5↑R10↑R20↑P5↑P10↑P20↑F5↑F10↑F20↑
LIMAP[23]LSDLine9.918.736.282.387.943.087.693.939.084.390.4
SOLD2Line5.929.664.276.679.688.196.497.972.984.086.7
NEF[54]PiDiNeT†Curve11.916.911.462.091.315.768.596.313.064.093.3
PiDiNeTCurve15.116.511.753.393.912.361.395.812.351.888.7
DexiNedCurve21.915.711.348.393.711.558.991.710.842.176.8
EMAP[19]PiDiNeTEdge9.215.630.275.789.535.679.195.432.477.092.2
DexiNedEdge8.88.956.488.994.862.989.995.759.188.994.9
2DGS-SNEdge8.87.963.590.996.270.491.094.966.390.495.1
EdgeGS[7]PiDiNeTEdge11.710.317.173.983.126.087.292.520.679.386.7
DexiNedEdge9.68.442.491.795.849.194.896.345.293.795.7
2DGS-SNEdge7.47.275.791.096.086.995.996.480.392.995.8
OursPiDiNeTEdge10.411.025.188.694.530.989.695.427.588.594.5
DexiNedEdge8.58.647.593.596.154.196.397.150.394.496.3
2DGS-SNEdge6.85.890.895.797.292.996.396.691.395.496.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.313, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on ABC-NEF dataset [54]. Our method achieves the state-of-the-art performance on most metrics. A: Accuracy (mm), C: Completeness (mm), Rx: Recall (threshold \\( x\\mathrm{\\;{mm}} \\) ),Px: Precision (mm), Fx: F-score (mm)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.357, + 0.907, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.673, + 0.908, + 0.703 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on the ABC-NEF dataset [54]. Compared to the baselines, our method achieves superior accuracy and completeness, along with a favorable balance between completeness and compactness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.728, + 0.486, + 0.894 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three datasets: ABC-NEF [18, 54] DTU [1], and Replica [43]. The ABC-NEF dataset [18, 54] comprises 115 CAD models, each containing 50 RGB images of the target object and a parametric edge file providing ground truth 3D edges. Following EMAP [19] and EdgeGS [7], we use 82 CAD models from ABC-NEF, excluding the ones with cylinders and spheres since smooth convex surfaces will produce silhouette edges that are not consistent in multi-views. For 2D edge image generation, we use PiDiNet [44], DexiNed [33], and our proposed 2DGS-SN as 2D edge detectors. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.909, + 0.82 + ], + "angle": 0, + "content": "DTU dataset [1] contains multi-view images of real-world scenes. We follow EdgeGS [7] and evaluate on 6 DTU scenes. We use edge images and the pseudo ground truth edges provided by EMAP [19]. The Replica dataset [43] consists of multiple indoor scenes. We follow EdgeGS [7] to conduct qualitative evaluation on Replica [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare our method with five baseline methods, including the state-of-the-art 3D line mapping method (LIMAP [23]), and several methods based on volumetric differentiable rendering (NEF [54], NEAT [51]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25654" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.44, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.089, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.232, + 0.908, + 0.262 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on the DTU and Replica datasets [1, 43]. Enum is the edge number and edges are visualized with varied colors. Compared to the baselines, our method achieves higher compactness and retains faithful completeness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.334 + ], + "angle": 0, + "content": "EMAP [19], EdgeGS [7]). Following EMAP and EdgeGS, we do not evaluate NEAT on the ABC-NEF dataset since NEAT often fails to train for texture-less objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.43 + ], + "angle": 0, + "content": "Metrics. Following the metrics used in baseline methods [7, 19], we evaluate Accuracy (A) and Completeness (C) in millimeters, and Recall (R), Precision (P), and F-score (F) as percentages at three thresholds (5mm, 10mm, 20mm). See supplementary material for more details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.185, + 0.455 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.583 + ], + "angle": 0, + "content": "Evaluation on ABC-NEF [54]. The qualitative and quantitative evaluation results on ABC-NEF dataset are shown in Fig. 3 and Tab. 1, respectively. Overall, the proposed method achieves a new state-of-the-art across most of the metrics. To the best of our knowledge, SketchSplat is the first method to achieve both accuracy (A) and completeness (C) below \\(7.0\\mathrm{mm}\\), while also exceeding \\(90\\%\\) in recall (R5), precision (P5), and F-score (F5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Tab. 1 shows LIMAP [23] achieves the highest accuracy due to its robust line mapping method. However, its completeness and recall are low, as LIMAP struggles with curve reconstruction and consistent edge detection results (Fig. 3(b)). NEF [54] produces more complete results in Fig. 3(c), but performs poorly on metrics as its fitting process is sensitive to noise in the extracted 3D points. EMAP [19] outperforms NEF with its unsigned distance field modeling and point shifting mechanism. However, Fig. 3(d) shows that EMAP still generates duplicated edges and misses some edges. With PidiNeT and DexiNed as 2D detector, EdgeGS [7] surpasses EMAP in recall and precision at \\(10mm\\) and \\(20mm\\) thresholds, but the trend reverses for measurements under \\(5mm\\). This is due to the thickness of 2D edges introduces bias in smaller distance evaluations. When using our detector 2DGS-SN, EdgeGS performs better than EMAP to a large margin under \\(5mm\\) metrics. As shown in Fig. 3, our SketchSplat achieves the highest precision and completeness compared to all the baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab.1, the proposed 2DGS-SN significantly boosts the performance of EMAP, EdgeGS, and Sketch-" + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.284, + 0.835, + 0.367 + ], + "angle": 0, + "content": "
MethodE-numTime (h)
NEF [54]22.91:26
EMAP [19]45.82:30
EdgeGS [7]140.80:05
ours44.30:10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.377, + 0.906, + 0.406 + ], + "angle": 0, + "content": "Table 2. Comparison of the number of reconstructed edges (E-num) and training time on ABC-NEF dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Splat, addressing the bias from PiDiNet[44] and Dex-iNed [33]. Notably, across all edge detectors, our method consistently achieves higher Recall, Accuracy, and F-score than EdgeGS [7], demonstrating effectiveness of Sketch-Splat in aligning 3D edges with 2D edge images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.909, + 0.68 + ], + "angle": 0, + "content": "Evaluation on DTU [1] and Replica [43]. Fig. 4 presents qualitative comparisons on the DTU and Replica datasets. Compared to the baselines, our method consistently produces accurate and complete results using significantly fewer edges, demonstrating a strong balance between completeness and compactness. Although the ground-truth edges in the DTU scenes are not sufficiently accurate for reliable quantitative evaluation [7], our method still achieves metrics comparable to state-of-the-art methods EMAP and EdgeGS (see supplementary material for more details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.791 + ], + "angle": 0, + "content": "Speed Analysis. The initialization with EdgeGS costs about 5 minutes. For optimization part, thanks to the efficiency of rasterization-based optimization, our method achieves comparable speed (about 5 minute per scene) to EdgeGS while being significantly faster than other neural network-based approaches [19, 54], as shown in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Compactness. EdgeGS achieves higher recall than NEF and EMAP due to its strict edge fitting process, which produces numerous short edges, covering more edge regions as the number of edges increases. Instead, our method incorporates adaptive topology control to reduce redundant edges, as shown in Tab. 2, Fig. 4, and Fig. 5(a,b,d). This" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "25655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.098, + 0.252, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.26, + 0.233, + 0.272 + ], + "angle": 0, + "content": "(a) Initialization (EdgeGS)" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.099, + 0.416, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.26, + 0.407, + 0.272 + ], + "angle": 0, + "content": "(b) Ours (w/o Topology Control)" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.099, + 0.577, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.26, + 0.553, + 0.271 + ], + "angle": 0, + "content": "(c) Ours (w/o 2DGS-SN)" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.099, + 0.74, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.26, + 0.68, + 0.271 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.099, + 0.904, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.26, + 0.84, + 0.271 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.286, + 0.905, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Ablation study on the key components of SketchSplat. Removing key components leads to noisier or redundant reconstructions." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.324, + 0.48, + 0.399 + ], + "angle": 0, + "content": "
MethodA↓C↓R5↑P5↑F5↑E-num
Ours6.85.890.892.991.344.28
w/o 2DGS-SN8.58.647.554.150.333.52
w/o Merge6.65.791.892.791.7140.72
w/o Filter6.85.890.992.991.344.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.409, + 0.483, + 0.451 + ], + "angle": 0, + "content": "Table 3. Ablation on the proposed components using ABC-NEF. w/o 2DGS-SN denotes SketchSplat trained with edge images detected by DexiNed, while the others are trained with 2DGS-SN." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.467, + 0.481, + 0.542 + ], + "angle": 0, + "content": "
MethodA↓C↓R5↑P5↑F5↑E-num
EdgeGS7.47.275.786.980.3140.76
Ours (EdgeGS)6.85.890.892.991.344.28
EMAP8.87.963.570.466.348.71
Ours (EMAP)6.96.288.890.889.331.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.608 + ], + "angle": 0, + "content": "Table 4. Ablation on initialization methods using ABC-NEF. Ours (EdgeGS) and Ours (EMAP) denote our method initialized with EdgeGS and EMAP, respectively. 2DGS-SN is chosen as 2D edge detector." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.637, + 0.483, + 0.669 + ], + "angle": 0, + "content": "mechanism helps SketchSplat achieve higher compactness, while maintaining the highest accuracy and completeness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.196, + 0.696 + ], + "angle": 0, + "content": "5.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Key Components. We use the ABC-NEF dataset to conduct an ablation study on the key contributions of our method, with results presented in Fig. 5 and Tab. 3. Fig. 5(b,d) shows that without topology control, noisy edge segments appear around the ground truth edges, with some extending beyond the object boundary. Incorporating topology control, particularly merging operations, effectively reduces the required edge count while maintaining comparable performance on other metrics (see Tab. 3). In addition, compared to edges reconstructed using DexiNed 2D edge images (Fig. 5(c)), our proposed 2DGS-SN better aligns edges with the ground truth (Fig. 5(d)), improving all metrics by effectively reducing bias in 2D edge images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Initialization Methods. In our main experiments, we use EdgeGS for initialization. However, we demonstrate SketchSplat can also achieve strong results when initialized with alternative methods like EMAP. As shown in Tab. 4, our approach consistently outperforms the initialization methods by a large margin and achieves comparable results across different initializations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.452, + 0.634, + 0.468 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.692 + ], + "angle": 0, + "content": "We present SketchSplat, a method for parametric 3D edge reconstruction from calibrated multi-view images. Unlike previous approaches that rely on 3D edge point reconstruction and edge fitting, SketchSplat directly optimizes edges through multi-view sketch splatting, ensuring accurate alignment with input 2D images. A series of topological operations are proposed to adaptively refine the edge structure during training, which reduces redundancy while maintaining high accuracy and completeness. Additionally, an improved 2D edge detection method is adopted to achieve higher reconstruction quality. Experiments show that SketchSplat achieves the state-of-the-art accuracy and completeness on the CAD dataset while maintaining compact results and efficient training speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements This research is based in part upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODN, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25656" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Caroline Baillard, Cordelia Schmid, Andrew Zisserman, and Andrew Fitzgibbon. Automatic line matching and 3d reconstruction of buildings from multiple views. In ISPRS Conference on Automatic Extraction of GIS Objects from Digital Imagery, pages 69-80, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[5] Andrea Bignoli, Andrea Romanoni, Matteo Matteucci, and Politecnico di Milano. Multi-view stereo 3d edge reconstruction. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 867-875. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.479 + ], + "angle": 0, + "content": "[6] Manmohan Chandraker, Jongwoo Lim, and David Kriegman. Moving in stereo: Efficient structure and motion using lines. In 2009 IEEE 12th International Conference on Computer Vision, pages 1741-1748. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[7] Kunal Chelani, Assia Benbihi, Torsten Sattler, and Fredrik Kahl. Edgegaussians-3d edge mapping via gaussian splatting. arXiv preprint arXiv:2409.12886, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.591 + ], + "angle": 0, + "content": "[8] Wentao Cheng, Sheng Yang, Maomin Zhou, Ziyuan Liu, Yiming Chen, and Mingyang Li. Road mapping and localization using sparse semantic visual features. IEEE robotics and automation letters, 6(4):8118-8125, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[9] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[10] Hyelim Choi, Minji Lee, Jiseock Kang, and Dongjun Lee. Online 3d edge reconstruction of wiry structures from monocular image sequences. IEEE Robotics and Automation Letters, 8(11):7479-7486, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[11] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Improving sparse 3d models for man-made environments using line-based 3d reconstruction. In 2014 2nd International Conference on 3D Vision, pages 535-542. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Petr Hruby, Shaohui Liu, Rémi Pautrat, Marc Pollefeys, and Daniel Barath. Handbook on leveraging lines for two-view relative pose estimation. In 2024 International Conference on 3D Vision (3DV), pages 376-386. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "curate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[15] Kahl and August. Multiview reconstruction of space curves. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 1017-1024. IEEE, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] Jeremy Yermiyahou Kaminski, Michael Fryers, Amnon Shashua, and Mina Teicher. Multiple view geometry of non-planar algebraic curves. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 181-186. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21219-21229, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[20] Pu Li, Jianwei Guo, Huibin Li, Bedrich Benes, and Dong-Ming Yan. Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[21] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[22] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[23] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[24] Yuncai Liu, Thomas S Huang, and Olivier D Faugeras. Determination of camera location from 2-d to 3-d line and point correspondences. IEEE Transactions on pattern analysis and machine intelligence, 12(1):28-37, 1990. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[25] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision, 124:65-79, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25657" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[33] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[35] Zhijian Qiao, Zehuan Yu, Huan Yin, and Shaojie Shen. Online monocular lane mapping using catmull-rom spline. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7179–7186. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[36] Tong Qin, Yuxin Zheng, Tongqing Chen, Yilun Chen, and Qing Su. A light-weight semantic map for visual localization towards autonomous driving. In 2021 IEEE international conference on robotics and automation (ICRA), pages 11248-11254. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Luc Robert and Olivier D Faugeras. Curve-based stereo: Figural continuity and curvature. In Proceedings. 1991 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 57-58. IEEE Computer Society, 1991. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[39] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Paul Smith, Ian Reid, and Andrew J Davison. Real-time monocular slam with straight lines. In BMVC, pages 17-26, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[42] Irwin Sobel, Gary Feldman, et al. A 3x3 isotropic gradient operator for image processing. a talk at the Stanford Artificial Project in, 1968:271-272, 1968. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[43] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[44] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[45] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[46] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4): 722-732, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[48] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[50] Xiaolong Wu, Patricio A Vela, and Cedric Pradalier. Robust monocular edge visual odometry through coarse-to-fine data association. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4923-4929. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[51] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. Neat: Distilling 3d wireframes from neural attraction fields. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25658" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19968-19977, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[53] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[54] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[55] Haiyang Ying, Baowei Jiang, Jinzhi Zhang, Di Xu, Tao Yu, Qionghai Dai, and Lu Fang. Parf: Primitive-aware radiance fusion for indoor scene novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17706-17716, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[56] Chao Zhang, Romain Pinquié, Arnaud Polette, Gregorio Carasi, Henri De Charnace, and Jean-Philippe Pernot. Automatic 3d cad models reconstruction from 2d orthographic drawings. Computers & Graphics, 114:179–189, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25659" + } + ] +] \ No newline at end of file diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_origin.pdf b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..df1aaee67b4fb092c360ad2f3f128aa11a14e168 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/03d6f604-4658-4429-92bd-44a5220bf2b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2974caf2c134016d12b1db24e24f72f13dbe16a375c114f02b0ff313a2aafd28 +size 6264831 diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/full.md b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f024d20fb6bd92ff0f8264df140b70e3c26a53b4 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/full.md @@ -0,0 +1,273 @@ +# SketchSplat: 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting + +Haiyang Ying, Matthias Zwicker +University of Maryland, College Park + +https://oceanying.github.io/SketchSplat + +# Abstract + +Edges are one of the most basic parametric primitives to describe structural information in 3D. In this paper, we study parametric 3D edge reconstruction from calibrated multi-view images. Previous methods usually reconstruct a 3D edge point set from multi-view 2D edge images, and then fit 3D edges to the point set. However, noise in the point set may cause gaps among fitted edges, and the recovered edges may not align with input multi-view images since the edge fitting depends only on the reconstructed 3D point set. To mitigate these problems, we propose SketchSplat, a method to reconstruct accurate, complete, and compact 3D edges via differentiable multi-view sketch splatting. We represent 3D edges as sketches, which are parametric lines and curves defined by attributes including control points, scales, and opacity. During reconstruction, we iteratively sample Gaussian points from a set of sketches and rasterize the Gaussians onto 2D edge images. Then the gradient of the image loss can be back-propagated to optimize the sketch attributes. Our method bridges 2D edge images and 3D edges in a differentiable manner, which ensures that 3D edges align well with 2D images and leads to accurate and complete results. We also propose a series of adaptive topological operations to reduce redundant edges and apply them along with the sketch optimization, yielding a more compact reconstruction. Finally, we contribute an accurate 2D edge detector that improves the performance of both ours and existing methods. Experiments show that our method achieves state-of-the-art accuracy, completeness, and compactness on a benchmark CAD dataset. + +# 1. Introduction + +Edges serve as one of the most fundamental elements of visual perception, playing a crucial role in understanding the geometry and structure of the 3D world. Unlike point-based representations, continuous edges offer a more compact and perceptually aligned way to describe 3D structures. As a + +![](images/d6f51cef5ca6b312b3eaeaeb38899e5d7d4e6cd889cff2f3e06149849a622cdf.jpg) +Figure 1. We propose SketchSplat, a multi-view 3D edge reconstruction method based on differentiable sketch splatting. Previous methods [7, 19] first recover edge points and then fit them to 3D edges, which is sensitive to noise and often leads to incomplete results. Instead, we sample Gaussian points from 3D sketches and rasterize them to compute a rendering loss, enabling direct edge parameter optimization under image supervision. Additionally, we design adaptive topology control methods to refine edge structures during training and contribute an improved 2D edge detection method for more accurate reconstruction. + +cornerstone of computer vision and graphics, 3D edge reconstruction underpins a wide range of applications, including CAD modeling [10, 20, 56], SLAM [13, 24, 41, 50], and autonomous driving [8, 35, 36]. Despite its significance, achieving high-quality 3D edge reconstructions remains a challenging task due to the inherent difficulty of ensuring + +the accuracy and completeness of reconstructed edges and maintaining the connectivity of the reconstructed structure. + +Traditional methods [5, 12, 23, 48] usually detect 2D line segments in each view. Then the 2D lines are lifted to 3D via line matching and triangulation across views. While these methods achieve high accuracy and efficiency, multiview consistent detection and robust line matching are difficult to achieve, leading to incomplete results. In addition, these methods are usually not capable of handling lines and curves simultaneously. + +In recent years, differentiable rendering has emerged as a promising technique to reconstruct 3D contents for novel view synthesis [3, 27, 28, 55], surface reconstruction [11, 22, 25, 47], and animation [29, 34, 49]. In the field of edge reconstruction, several works introduced the idea of differentiable rendering to recover edge information [7, 19, 54]. Instead of detecting parametric 2D edge segments explicitly, they leverage pixel-wise 2D edge intensity images [33, 44] as a supervision signal and optimize a point-based 3D edge representation via differentiable multiview rendering, resulting in an extracted 3D edge point set. Parametric 3D edges are then obtained by fitting lines and curves to the extracted 3D points. Since the edge-fitting step relies purely on the recovered 3D edge points but not the input 2D edge images, the fitted 3D edges may not align well with the 2D edge images. In addition, noise in the 3D edge points usually leads to discontinuous and fragmented edges. + +To address these limitations, we propose SketchSplat, a differentiable edge reconstruction framework that directly optimizes parametric 3D edges by fitting them to the input 2D edge images using differentiable rendering. Our experimental results demonstrate that this leads to more accurate, complete, and compact reconstructions, more faithfully reflecting the original connectivity. + +Firstly, we propose to model 3D edges as sketches, which are parametric lines and curves with control points, scales, and opacity. These sketches can be initialized from existing methods [7, 19, 54], and can be sampled as a set of 3D Gaussian points for differentiable rendering. To optimize the sketches, we sample Gaussian points from the sketches and rasterize them in each view to calculate the rendering loss between the rendered and the ground-truth edge image. Then the error can be back-propagated to update the sketch attributes. Since we optimize our sketch parameters directly using the 2D edge images, the resulting edges will align well with the 2D edges in the input images. To further improve the compactness of the reconstructed edges, we leverage a series of topological operations to connect, merge, and filter sketches during optimization. Finally, we observe that one common limitation of existing methods [7, 19, 54] lies in the inaccuracy of 2D edge images. To mitigate this, we propose an accurate and robust edge detection method, which is validated to improve both + +our approach and existing methods by a large margin. + +Experiments demonstrate that our method achieves state-of-the-art performance on the ABC-NEF dataset [54], surpassing previous approaches by over $10\%$ in F-score. In summary, key innovations of our method include: + +- Differentiable Parametric Optimization: We contribute a pipeline that directly optimizes parametric edge representations via differentiable multi-view rasterization, leading to SOTA accuracy and completeness. +- Adaptive Topological Control: We propose a series of topological operations to further enhance edge continuity and reduce redundant edges during training, yielding more compact reconstruction results. +- Accurate 2D Edge Detection: We propose a more accurate 2D edge detection method, which shows large improvements for both our approach and existing methods. + +# 2. Related Work + +Line Matching and Triangulation. Conventional approaches reconstruct 3D lines [2, 4-6, 12, 26] or curves [15, 16, 38] by enforcing multi-view geometric constraints. Given detected 2D edge segments and their descriptors, they match the line descriptors across multiple views and perform triangulation to lift 2D edges into 3D. However, these methods face challenges in consistently detecting edges and robustly matching lines, especially when occlusions cause endpoint inconsistencies across views. Despite notable advancements by involving structural [39, 48] or learning-based priors [23, 30-32], line detection and matching remain critical bottlenecks, limiting overall performance. While most of the methods can only handle a single type of edges, in this paper we aim at recovering both lines and curves. + +Differentiable Rendering Methods. The advancement of differentiable rendering opens up a new way for 3D reconstruction [22, 27, 47]. Inspired by NeRF [27], NEF [54] proposes to optimize an implicit neural radiance field with a multi-view rendering loss. After training, 3D edge points are extracted from the radiance field and curves are fitted to these points. NEAT [51] builds on VolSDF [52] and uses a signed distance field to represent both object surfaces and endpoints of 3D lines. The surface-based modeling approach restricts this method to objects with rich texture. EMAP [19] adopts an unsigned distance function (UDF), where sampled points can be displaced towards the zero-level set of the UDF to be more accurate. While methods based on neural field are time consuming to optimize, EdgeGS [7] represents 3D edge fields with explicit Gaussian points and leverages Gaussian splatting [17] to render and optimize Gaussian points, leading to more efficient edge reconstruction. A common issue of these methods + +is that the edge fitting process relies only on the extracted 3D edge points, hence noise in the reconstructed 3D points may lead to inaccurate and incomplete edges. Instead, we represent 3D edges as sketches defined parametrically by control points, scales, and opacity, and directly optimize these parameters via differentiable rendering. By leveraging sketch sampling and rasterization, our differentiable rendering aligns 3D edges with 2D edge images, resulting in more accurate and complete reconstructions. + +Differentiable Vector Graphics Optimization. DiffVG [21] proposes a differentiable rasterizer that can optimize 2D vector graphics with 2D image supervision in a differentiable manner. However, the differentiation process relies on distance measurements between the rasterized curves and pixels, which needs to be custom tailored for each type of curve. 3Doodle [9] extends the idea of DiffVG [21] to reconstruct 3D curve-based object abstractions using multiple input views. They optimize a set of Bezier curves by minimizing the perceptual loss between the rasterized image and multi-view CLIP feature maps [37, 45]. Since CLIP features only contain abstract information, 3Doodle is not suitable for recovering accurate edges of objects like CAD models. Moreover, they approximate the projection of Bezier curves using an orthographic camera model. Instead, we get inspiration from 3DGS [17] and propose to optimize 3D sketches via sampling and rasterization of Gaussian points sampled from sketch parameters, which can be easily extended to different types of curves and achieves accurate 3D edge reconstruction. + +# 3. Preliminaries: 3D Gaussian Splitting + +We leverage 3D Gaussian splatting (3DGS) [17] for initialization and differentiable sketch rendering. 3DGS proposes to model 3D scenes with a collection of anisotropic 3D Gaussian points. Each point is defined as a Gaussian distribution $G(x) = e^{-\frac{1}{2}(x - \mu)^T\Sigma^{-1}(x - \mu)}$ , where $\mu \in R^3$ is the center position or mean, and $\Sigma \in R^{3\times 3}$ is the covariance matrix. Each Gaussian has an opacity $o \in R$ and a view-dependent color value $c(d) = SH(d) \in R^3$ modeled with spherical harmonics. Following Zwicker et al. [57], the Gaussians can be rendered to 2D images using an efficient volume rendering approach, where the color for each camera ray $r$ is + +$$ +c (r) = \sum_ {i \in N} c _ {i} (d (r)) \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \alpha_ {i} = o _ {i} g _ {i} (r), \tag {1} +$$ + +where $g_{i}(r)$ is the line integral of the 3D Gaussian $G_{i}$ along ray $r$ , and the indices $i$ are ordered in front to back manner. By calculating the gradient of a loss $\mathcal{L}$ between the rendered and ground truth image, the parameters of all Gaussians can + +be optimized using gradient descent. A typical loss function is a combination of an $L_{1}$ and a perceptual loss such as + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {1} + (1 - \lambda) \mathcal {L} _ {D - S S I M}. \tag {2} +$$ + +Given input images, the initial Gaussians are usually obtained from SfM [40] methods. During training, several operations are adopted to adaptively control the density and distribution of Gaussians. For example, Gaussians with low opacity will be culled, and Gaussians with high gradient will be duplicated or split to densify specific regions. + +In SketchSplat, we leverage 3DGS [17] as an intermediary to bridge 2D edge images and 3D parametric edges, enabling the optimization of 3D edges through differentiable image rendering. + +# 4. Method + +Given multi-view posed RGB images as input, our target is to recover the parametric edges for the 3D objects. Fig. 2 provides an overview of our approach. We firstly compute an edge image for each RGB image using our proposed 2D edge detector. Then we represent edges as 3D parametric sketches, including both lines $l \in \mathbb{R}^{2 \times 3}$ and third-order Bézier curves with four 3D control points $c \in \mathbb{R}^{4 \times 3}$ . We initialize a set of sketches from existing edge reconstruction methods, and treat their control points, along with persketch scales and opacity, as estimizable parameters. During optimization, we use a sample-based method to convert sketches to Gaussian points, which can be rasterized onto 2D images to calculate the image loss. We back-propagate the gradient of the loss to optimize the sketch parameters. In addition, we introduce a set of topological operations to adaptively refine the sketches during training, yielding more accurate and compact edge reconstruction results. + +# 4.1. Edge Image Extraction + +Previously, some methods [23, 51] extract edges with line detectors like LSD [46] and SOLD2 [30]. While these methods produce accurate parametric lines in 2D, they may fail to handle curve regions and result in incomplete results. Recent methods [7, 19, 54] use neural network-based edge detectors [33, 44] to detect edge intensity maps as supervision signal. However, since these detectors are not trained specifically on CAD-like objects, they may miss some significant edges in 2D or wrongly recover shading textures as real edges. Finally, we observed that these detectors introduce bias by consistently shifting detected edges slightly from certain real edges, which leads to multi-view inconsistency in 3D edge reconstruction (see supplementary). We address these issues by proposing a new edge detection method, which leverages geometric cues (e.g., per-pixel depth and normal estimation of 2D RGB images) to produce more accurate edge detection results. + +![](images/f08285c5181b9f452d9955574aa0a5bf4bd45b2af4b31cc23ad11f38d28ddc57.jpg) +(A) Sketch initialization +(B) Differentiable Multi-view Sketch Optimization +Figure 2. Overview of SketchSplat pipeline: (a) We extract edge images using a novel approach (Sec. 4.1). (b) We then obtain initial edge points using EdgeGS [7] and (c) fit parametric sketches defined as lines and third-order Bézier curves. The core step of our approach is to optimize the sketches by (d) sampling and rasterizing them in a differentiable manner and back-propagating gradients of the image loss to (e) sketch parameters. In addition, during training we apply (f) a set of topological operations to improve compactness and connectivity of the reconstructed sketches. (g) After training, we take the optimized geometric parameters of sketches as the reconstructed 3D edges. + +For each RGB image, we combine information from three parts: foreground mask $A$ , depth map $D$ , and normal map $N$ . The foreground mask can be extracted using the image alpha-channel (if available), a thresholding operation or possibly using semantic segmentation, providing object boundary information. We obtain a depth image for each view by reconstructing the scene with 2DGS [14] and rendering depth maps from each view. We use a normal estimator [53] to generate normal maps $N$ for each view. Then we calculate gradient maps $g(D)$ and $g(N)$ with the Sobel operator [42], threshold them with $t_d$ and $t_n$ , and combine them via the OR operation, followed by a Gaussian filter $G_f(\cdot)$ to smooth the boundaries of the edges. In summary, we call our edge detector 2DGS-SN, defined as + +$$ +E = G _ {f} * (A | (g (D) > t _ {d}) | (g (N) > t _ {n})). \tag {3} +$$ + +In Sec. 5 we show that this simple approach improves existing methods [7] on CAD-like objects by a large margin. + +# 4.2. Sketch Splatting + +Representation. We represent edges as 3D parametric sketches, including both lines $l \in \mathbb{R}^{2 \times 3}$ and third-order Bézier curves with control points $c \in \mathbb{R}^{4 \times 3}$ . This is the same as the output format of EdgeGS[7]. While EdgeGS optimizes 3D point positions first and then fits 3D lines and curves to the 3D points, we optimize the parameters of lines and curves directly via multi-view differentiable rasterization. In addition to the estimizable geometry parameters $(l$ or $c)$ , we equip each sketch with an estimizable opacity $o \in \mathbb{R}$ and a local scale $s \in \mathbb{R}^3$ to model the sketch intensity + +and thickness, respectively. These attributes will support Gaussian sampling and rasterization in the next steps. + +Sketch Initialization. Unlike point-based methods [7, 19], it is non-trivial to initialize sketches directly from multi-view edge images. Instead, inspired by 3DGS [17], we first initialize sketches with existing point-based methods like EdgeGS [7]. By default, we select EdgeGS because it provides a sufficient number of short 3D edges to cover the objects, and achieves the highest completeness among previous methods, despite some fragmentation and noise. + +Differentiable Sketch Optimization. Our key idea is to bridge the gap between 2D edge images and 3D sketches via differentiable rendering and gradient descent optimization, which significantly improves sketch reconstruction over initialization method EdgeGS [7]. Inspired by 3DGS [17] and DiffVG [21], we choose 3D Gaussians as an intermediate representation to achieve differentiable sketch rendering. + +For each gradient descent step, we first sample points from each sketch with a constant step size (5mm for all experiments), resulting in a point set $P$ . Each point will be assigned the opacity and scale value of the corresponding sketch. To model the orientation, the main direction of each point is sampled as the local tangent direction of the sketch. Then we follow 3DGS[17] to rasterize the Gaussian points onto a chosen 2D view to generate a rendered edge image $E^{*}$ . We calculate the $L_{1}$ loss between $E^{*}$ and the ground truth edge image $E_{gt}$ , and back-propagate the gradient to update the sketch parameters $S_{i} = \{l_{i} \text{ or } c_{i}, o_{i}, s_{i}\}$ . We fol + +low EdgeGS [7] to sample an equal number of foreground and background pixels for $L_{1}$ loss calculation. + +# 4.3. Adaptive Topology Control + +A good edge reconstruction, which could be used for CAD modeling for example, should not only accurately match the 2D edge images, but also provide a compact, well-connected set of edges. Neighboring edges should be connected correctly, overlapping sketches should be merged, and the number of sketches should be as small as possible while ensuring accurate matching of 2D edge images. For this purpose we deploy topology control operations adaptively during optimization. While some existing methods [7] do not consider topology optimization, others [19, 54] try to connect and merge neighboring sketches in the post-processing stage. However, this may be sensitive to noise in the extracted 3D point set and cannot ensure that the resulting edges are still consistent with the input multi-view 2D edge images. Instead, since our method directly optimizes sketch parameters using iterative gradient descent, we can execute topological operations along with the multi-view optimization steps, as shown in Fig. 2(f). + +End-point Merging. In our representation, each sketch has two end-points. If there exist two end-points, one from sketch $S_{i}$ and one from sketch $S_{j}$ , that are sufficiently close (i.e. the point distance is smaller than threshold $th_{connect} = 10mm$ ), we will connect sketch $S_{i}$ and $S_{j}$ by merging those two end-points into one point (see Fig. 2(f-i)). However, this means the number of estimable points could differ from the number of control points $|\{l\}| * 2 + |\{c\}| * 4$ . To accommodate this, we maintain an estimable point set $P$ , and the control points of each line $l$ and curve $c$ are represented by indexing point set $P$ . This representation also supports our additional topological operations. + +Overlapping Sketch Merging. As shown in Fig. 2(f-ii), if one sketch is almost entirely covered by another, the smaller one will be merged into the larger one. Specifically, for sketch $S_{i}$ and $S_{j}$ , we sample two point sets $P_{i}$ and $P_{j}$ and calculate the distance between each pair of points. Then for $P_{i}$ , we calculate the ratio $r_{i\rightarrow j}$ of points that have close neighboring points in $P_{j}$ , where the neighbor threshold is $th_{neighbor} = 10mm$ . If $r_{i\rightarrow j} > th_{overlap} = 80\%$ , sketch $S_{i}$ will be merged in sketch $S_{j}$ . If both $r_{i\rightarrow j}$ and $r_{j\rightarrow i}$ surpass $th_{overlap}$ , we will choose the sketch with larger overlapping ratio to be merged. + +Co-linear Sketch Merging. If two lines are co-linear and nearly connected, they should be merged into a longer line (see Fig. 2(f-iii)). The co-linear condition is: two lines should have similar direction ( $th_{dir} = 5^{\circ}$ ), and the maximum of the projective distance/offset should be small + +enough ( $th_{offset} = 10mm$ ). In addition, to prevent gaps between two lines, one line will be projected onto the other one to calculate if there is a gap and if the gap is small enough ( $th_{connect}$ ). To merge the lines, we take the two end-points with the furthest distance after projection and delete the other two endpoints from point set $P$ . Additionally, to accelerate this procedure, we use the axis-aligned bounding box of each sketch to quickly filter out all the sketch pairs that cannot meet the merging criterion. + +Sketch Filtering. After training, we adopt a multi-view filter method to filter sketches that are not visible. We borrow the refinement idea from EMAP [19]. 3D points can be projected onto 2D edge images to see if they are aligned with the 2D edges by checking the edge intensity of the projected pixel position. A point will be labeled as invisible if it is invisible in more than $90\%$ of the views. For each sketch, we will evaluate the visibility for all of the points sampled from the sketch. If more than $th_{vis} = 50\%$ of the sampled points are invisible points, the sketch will be removed. When rasterizing sketches, some sketches that should not appear due to occlusion still appear in the rendered image, which sometimes leads to incorrect reconstruction. This filter scheme will help filter out such outlier sketches. + +# 4.4. Implementation Details + +Unless otherwise specified, we initialize with EdgeGS [7], which supplies a sufficient number of 3D edges to capture object structure. For each scene, we optimize the sketches for 1000 epochs with Adam. We update parameters only once for each epoch, where the loss is accumulated across all the training views. All of the scenes are scaled to fit within a $1\mathrm{m}^3$ bounding box, and topology operation hyperparameters are set accordingly. Distance thresholds (e.g., $th_{connect}$ ) are set based on the desired precision and we set all of them as $1\%$ of the scene size $(1m^3)$ . The percentage and angular thresholds (e.g., $th_{overlap}$ and $th_{dir}$ ) work across different scenes without adjustment. All our experiments are implemented on a single RTX A5000 NVIDIA GPU. In average, each scene in ABC-NEF [54] dataset costs about 10 minutes for training, where both initialization and sketch optimization takes about 5 minutes, respectively. + +# 5. Experiments + +# 5.1. Experimental Setup + +We compare our method with other multi-view edge reconstruction methods following the evaluation setup of EdgeGS [7]. The results show that our method achieves the state-of-the-art performance in both accuracy and completeness on the ABC-NEF dataset [18, 54]. Our method also produces highly compact results with fast inference speed. + +
MethodDetectorModalA↓C↓R5↑R10↑R20↑P5↑P10↑P20↑F5↑F10↑F20↑
LIMAP[23]LSDLine9.918.736.282.387.943.087.693.939.084.390.4
SOLD2Line5.929.664.276.679.688.196.497.972.984.086.7
NEF[54]PiDiNeT†Curve11.916.911.462.091.315.768.596.313.064.093.3
PiDiNeTCurve15.116.511.753.393.912.361.395.812.351.888.7
DexiNedCurve21.915.711.348.393.711.558.991.710.842.176.8
EMAP[19]PiDiNeTEdge9.215.630.275.789.535.679.195.432.477.092.2
DexiNedEdge8.88.956.488.994.862.989.995.759.188.994.9
2DGS-SNEdge8.87.963.590.996.270.491.094.966.390.495.1
EdgeGS[7]PiDiNeTEdge11.710.317.173.983.126.087.292.520.679.386.7
DexiNedEdge9.68.442.491.795.849.194.896.345.293.795.7
2DGS-SNEdge7.47.275.791.096.086.995.996.480.392.995.8
OursPiDiNeTEdge10.411.025.188.694.530.989.695.427.588.594.5
DexiNedEdge8.58.647.593.596.154.196.397.150.394.496.3
2DGS-SNEdge6.85.890.895.797.292.996.396.691.395.496.5
+ +Table 1. Quantitative evaluation on ABC-NEF dataset [54]. Our method achieves the state-of-the-art performance on most metrics. A: Accuracy (mm), C: Completeness (mm), Rx: Recall (threshold $x\mathrm{\;{mm}}$ ),Px: Precision (mm), Fx: F-score (mm). + +![](images/8c8d4e47194c27dc9c43e037339f3201485ad20aa9ba071b39da67dcfcf37012.jpg) +Figure 3. Qualitative comparison on the ABC-NEF dataset [54]. Compared to the baselines, our method achieves superior accuracy and completeness, along with a favorable balance between completeness and compactness. + +Datasets. We evaluate our method on three datasets: ABC-NEF [18, 54] DTU [1], and Replica [43]. The ABC-NEF dataset [18, 54] comprises 115 CAD models, each containing 50 RGB images of the target object and a parametric edge file providing ground truth 3D edges. Following EMAP [19] and EdgeGS [7], we use 82 CAD models from ABC-NEF, excluding the ones with cylinders and spheres since smooth convex surfaces will produce silhouette edges that are not consistent in multi-views. For 2D edge image generation, we use PiDiNet [44], DexiNed [33], and our proposed 2DGS-SN as 2D edge detectors. The + +DTU dataset [1] contains multi-view images of real-world scenes. We follow EdgeGS [7] and evaluate on 6 DTU scenes. We use edge images and the pseudo ground truth edges provided by EMAP [19]. The Replica dataset [43] consists of multiple indoor scenes. We follow EdgeGS [7] to conduct qualitative evaluation on Replica [43]. + +Baselines. We compare our method with five baseline methods, including the state-of-the-art 3D line mapping method (LIMAP [23]), and several methods based on volumetric differentiable rendering (NEF [54], NEAT [51], + +![](images/acb3124d7edac71cac255d2bffe7da517135cd4a7ea61b907a43a6402e83c6ca.jpg) +Figure 4. Qualitative comparison on the DTU and Replica datasets [1, 43]. Enum is the edge number and edges are visualized with varied colors. Compared to the baselines, our method achieves higher compactness and retains faithful completeness. + +![](images/db3d13ef0cdde8892a7c652c6e24f491c908fb5c181d467651da1c94be2866ca.jpg) + +EMAP [19], EdgeGS [7]). Following EMAP and EdgeGS, we do not evaluate NEAT on the ABC-NEF dataset since NEAT often fails to train for texture-less objects. + +Metrics. Following the metrics used in baseline methods [7, 19], we evaluate Accuracy (A) and Completeness (C) in millimeters, and Recall (R), Precision (P), and F-score (F) as percentages at three thresholds (5mm, 10mm, 20mm). See supplementary material for more details. + +# 5.2. Results + +Evaluation on ABC-NEF [54]. The qualitative and quantitative evaluation results on ABC-NEF dataset are shown in Fig. 3 and Tab. 1, respectively. Overall, the proposed method achieves a new state-of-the-art across most of the metrics. To the best of our knowledge, SketchSplat is the first method to achieve both accuracy (A) and completeness (C) below $7.0\mathrm{mm}$ , while also exceeding $90\%$ in recall (R5), precision (P5), and F-score (F5). + +Tab. 1 shows LIMAP [23] achieves the highest accuracy due to its robust line mapping method. However, its completeness and recall are low, as LIMAP struggles with curve reconstruction and consistent edge detection results (Fig. 3(b)). NEF [54] produces more complete results in Fig. 3(c), but performs poorly on metrics as its fitting process is sensitive to noise in the extracted 3D points. EMAP [19] outperforms NEF with its unsigned distance field modeling and point shifting mechanism. However, Fig. 3(d) shows that EMAP still generates duplicated edges and misses some edges. With PidiNeT and DexiNed as 2D detector, EdgeGS [7] surpasses EMAP in recall and precision at $10mm$ and $20mm$ thresholds, but the trend reverses for measurements under $5mm$ . This is due to the thickness of 2D edges introduces bias in smaller distance evaluations. When using our detector 2DGS-SN, EdgeGS performs better than EMAP to a large margin under $5mm$ metrics. As shown in Fig. 3, our SketchSplat achieves the highest precision and completeness compared to all the baselines. + +As shown in Tab.1, the proposed 2DGS-SN significantly boosts the performance of EMAP, EdgeGS, and Sketch- + +
MethodE-numTime (h)
NEF [54]22.91:26
EMAP [19]45.82:30
EdgeGS [7]140.80:05
ours44.30:10
+ +Table 2. Comparison of the number of reconstructed edges (E-num) and training time on ABC-NEF dataset. + +Splat, addressing the bias from PiDiNet[44] and Dex-iNed [33]. Notably, across all edge detectors, our method consistently achieves higher Recall, Accuracy, and F-score than EdgeGS [7], demonstrating effectiveness of Sketch-Splat in aligning 3D edges with 2D edge images. + +Evaluation on DTU [1] and Replica [43]. Fig. 4 presents qualitative comparisons on the DTU and Replica datasets. Compared to the baselines, our method consistently produces accurate and complete results using significantly fewer edges, demonstrating a strong balance between completeness and compactness. Although the ground-truth edges in the DTU scenes are not sufficiently accurate for reliable quantitative evaluation [7], our method still achieves metrics comparable to state-of-the-art methods EMAP and EdgeGS (see supplementary material for more details). + +Speed Analysis. The initialization with EdgeGS costs about 5 minutes. For optimization part, thanks to the efficiency of rasterization-based optimization, our method achieves comparable speed (about 5 minute per scene) to EdgeGS while being significantly faster than other neural network-based approaches [19, 54], as shown in Tab. 2. + +Compactness. EdgeGS achieves higher recall than NEF and EMAP due to its strict edge fitting process, which produces numerous short edges, covering more edge regions as the number of edges increases. Instead, our method incorporates adaptive topology control to reduce redundant edges, as shown in Tab. 2, Fig. 4, and Fig. 5(a,b,d). This + +![](images/680192c8ba1f43ce645fd6dbf40c92af000d6b8df06cc92819349e85a2c039ee.jpg) +(a) Initialization (EdgeGS) + +![](images/c4a7d33a00db868f194705e7ddcf51748eeaa53a735c5d08707d89630f5afe73.jpg) +(b) Ours (w/o Topology Control) + +![](images/9e4bd38b7d8a0e1d516eb3ee29ff5294fd52ae5bad1940b4e7ed31906c8c342b.jpg) +(c) Ours (w/o 2DGS-SN) + +![](images/b472d9b45f4c50771158f5ae1d104fbceb38c9cc778fca995a8c6d181ac0574b.jpg) +(d) Ours +Figure 5. Ablation study on the key components of SketchSplat. Removing key components leads to noisier or redundant reconstructions. + +![](images/5dbe9e8cf8a17c27180f39223ca9140be3d9da8b7873a077494187c19545a86f.jpg) +(e) GT + +
MethodA↓C↓R5↑P5↑F5↑E-num
Ours6.85.890.892.991.344.28
w/o 2DGS-SN8.58.647.554.150.333.52
w/o Merge6.65.791.892.791.7140.72
w/o Filter6.85.890.992.991.344.37
+ +Table 3. Ablation on the proposed components using ABC-NEF. w/o 2DGS-SN denotes SketchSplat trained with edge images detected by DexiNed, while the others are trained with 2DGS-SN. + +
MethodA↓C↓R5↑P5↑F5↑E-num
EdgeGS7.47.275.786.980.3140.76
Ours (EdgeGS)6.85.890.892.991.344.28
EMAP8.87.963.570.466.348.71
Ours (EMAP)6.96.288.890.889.331.70
+ +Table 4. Ablation on initialization methods using ABC-NEF. Ours (EdgeGS) and Ours (EMAP) denote our method initialized with EdgeGS and EMAP, respectively. 2DGS-SN is chosen as 2D edge detector. + +mechanism helps SketchSplat achieve higher compactness, while maintaining the highest accuracy and completeness. + +# 5.3. Ablation + +Key Components. We use the ABC-NEF dataset to conduct an ablation study on the key contributions of our method, with results presented in Fig. 5 and Tab. 3. Fig. 5(b,d) shows that without topology control, noisy edge segments appear around the ground truth edges, with some extending beyond the object boundary. Incorporating topology control, particularly merging operations, effectively reduces the required edge count while maintaining comparable performance on other metrics (see Tab. 3). In addition, compared to edges reconstructed using DexiNed 2D edge images (Fig. 5(c)), our proposed 2DGS-SN better aligns edges with the ground truth (Fig. 5(d)), improving all metrics by effectively reducing bias in 2D edge images. + +Initialization Methods. In our main experiments, we use EdgeGS for initialization. However, we demonstrate SketchSplat can also achieve strong results when initialized with alternative methods like EMAP. As shown in Tab. 4, our approach consistently outperforms the initialization methods by a large margin and achieves comparable results across different initializations. + +# 6. Conclusion + +We present SketchSplat, a method for parametric 3D edge reconstruction from calibrated multi-view images. Unlike previous approaches that rely on 3D edge point reconstruction and edge fitting, SketchSplat directly optimizes edges through multi-view sketch splatting, ensuring accurate alignment with input 2D images. A series of topological operations are proposed to adaptively refine the edge structure during training, which reduces redundancy while maintaining high accuracy and completeness. Additionally, an improved 2D edge detection method is adopted to achieve higher reconstruction quality. Experiments show that SketchSplat achieves the state-of-the-art accuracy and completeness on the CAD dataset while maintaining compact results and efficient training speed. + +Acknowledgements This research is based in part upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODN, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 6, 7 +[2] Caroline Baillard, Cordelia Schmid, Andrew Zisserman, and Andrew Fitzgibbon. Automatic line matching and 3d reconstruction of buildings from multiple views. In ISPRS Conference on Automatic Extraction of GIS Objects from Digital Imagery, pages 69-80, 1999. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 2 +[5] Andrea Bignoli, Andrea Romanoni, Matteo Matteucci, and Politecnico di Milano. Multi-view stereo 3d edge reconstruction. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 867-875. IEEE, 2018. 2 +[6] Manmohan Chandraker, Jongwoo Lim, and David Kriegman. Moving in stereo: Efficient structure and motion using lines. In 2009 IEEE 12th International Conference on Computer Vision, pages 1741-1748. IEEE, 2009. 2 +[7] Kunal Chelani, Assia Benbihi, Torsten Sattler, and Fredrik Kahl. Edgegaussians-3d edge mapping via gaussian splatting. arXiv preprint arXiv:2409.12886, 2024. 1, 2, 3, 4, 5, 6, 7 +[8] Wentao Cheng, Sheng Yang, Maomin Zhou, Ziyuan Liu, Yiming Chen, and Mingyang Li. Road mapping and localization using sparse semantic visual features. IEEE robotics and automation letters, 6(4):8118-8125, 2021. 1 +[9] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3 +[10] Hyelim Choi, Minji Lee, Jiseock Kang, and Dongjun Lee. Online 3d edge reconstruction of wiry structures from monocular image sequences. IEEE Robotics and Automation Letters, 8(11):7479-7486, 2023. 1 +[11] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2 +[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Improving sparse 3d models for man-made environments using line-based 3d reconstruction. In 2014 2nd International Conference on 3D Vision, pages 535-542. IEEE, 2014. 2 +[13] Petr Hruby, Shaohui Liu, Rémi Pautrat, Marc Pollefeys, and Daniel Barath. Handbook on leveraging lines for two-view relative pose estimation. In 2024 International Conference on 3D Vision (3DV), pages 376-386. IEEE, 2024. 1 +[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac + +curate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 4 +[15] Kahl and August. Multiview reconstruction of space curves. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 1017-1024. IEEE, 2003. 2 +[16] Jeremy Yermiyahou Kaminski, Michael Fryers, Amnon Shashua, and Mina Teicher. Multiple view geometry of non-planar algebraic curves. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 181-186. IEEE, 2001. 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4 +[18] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 5, 6 +[19] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21219-21229, 2024. 1, 2, 3, 4, 5, 6, 7 +[20] Pu Li, Jianwei Guo, Huibin Li, Bedrich Benes, and Dong-Ming Yan. Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2024. 1 +[21] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 3, 4 +[22] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2 +[23] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 2, 3, 6, 7 +[24] Yuncai Liu, Thomas S Huang, and Olivier D Faugeras. Determination of camera location from 2-d to 3-d line and point correspondences. IEEE Transactions on pattern analysis and machine intelligence, 12(1):28-37, 1990. 1 +[25] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2 +[26] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision, 124:65-79, 2017. 2 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2 +[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 2, 3 +[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023. +[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 2 +[33] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 2, 3, 6, 7 +[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 2 +[35] Zhijian Qiao, Zehuan Yu, Huan Yin, and Shaojie Shen. Online monocular lane mapping using catmull-rom spline. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7179–7186. IEEE, 2023. 1 +[36] Tong Qin, Yuxin Zheng, Tongqing Chen, Yilun Chen, and Qing Su. A light-weight semantic map for visual localization towards autonomous driving. In 2021 IEEE international conference on robotics and automation (ICRA), pages 11248-11254. IEEE, 2021. 1 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3 +[38] Luc Robert and Olivier D Faugeras. Curve-based stereo: Figural continuity and curvature. In Proceedings. 1991 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 57-58. IEEE Computer Society, 1991. 2 + +[39] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853. IEEE, 2006. 2 +[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3 +[41] Paul Smith, Ian Reid, and Andrew J Davison. Real-time monocular slam with straight lines. In BMVC, pages 17-26, 2006. 1 +[42] Irwin Sobel, Gary Feldman, et al. A 3x3 isotropic gradient operator for image processing. a talk at the Stanford Artificial Project in, 1968:271-272, 1968. 4 +[43] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7 +[44] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 2, 3, 6, 7 +[45] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3 +[46] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4): 722-732, 2008. 3 +[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2 +[48] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 2 +[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2 +[50] Xiaolong Wu, Patricio A Vela, and Cedric Pradalier. Robust monocular edge visual odometry through coarse-to-fine data association. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4923-4929. IEEE, 2020. 1 +[51] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. Neat: Distilling 3d wireframes from neural attraction fields. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19968-19977, 2024. 2, 3, 6 +[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2 +[53] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 2024. 4 +[54] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 2, 3, 5, 6, 7 +[55] Haiyang Ying, Baowei Jiang, Jinzhi Zhang, Di Xu, Tao Yu, Qionghai Dai, and Lu Fang. Parf: Primitive-aware radiance fusion for indoor scene novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17706-17716, 2023. 2 +[56] Chao Zhang, Romain Pinquié, Arnaud Polette, Gregorio Carasi, Henri De Charnace, and Jean-Philippe Pernot. Automatic 3d cad models reconstruction from 2d orthographic drawings. Computers & Graphics, 114:179–189, 2023. 1 +[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3 \ No newline at end of file diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/images.zip b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8205ff3eb0a612c6bde92c924c5dd9b9dbb13f19 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9e76e4bf07741114845a84661973a37b138c404ef9914eb46758ddb18b2a592 +size 597706 diff --git a/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/layout.json b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5baf2e03ac473df82a5a8c6cb3cdca6cfbdca7f3 --- /dev/null +++ b/2025/SketchSplat_ 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting/layout.json @@ -0,0 +1,7959 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 478, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 478, + 140 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 478, + 140 + ], + "type": "text", + "content": "SketchSplat: 3D Edge Reconstruction via Differentiable Multi-view Sketch Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 161, + 394, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 161, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 196, + 161, + 394, + 190 + ], + "type": "text", + "content": "Haiyang Ying, Matthias Zwicker \nUniversity of Maryland, College Park" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 191, + 408, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 191, + 408, + 203 + ], + "spans": [ + { + "bbox": [ + 196, + 191, + 408, + 203 + ], + "type": "text", + "content": "https://oceanying.github.io/SketchSplat" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 606 + ], + "type": "text", + "content": "Edges are one of the most basic parametric primitives to describe structural information in 3D. In this paper, we study parametric 3D edge reconstruction from calibrated multi-view images. Previous methods usually reconstruct a 3D edge point set from multi-view 2D edge images, and then fit 3D edges to the point set. However, noise in the point set may cause gaps among fitted edges, and the recovered edges may not align with input multi-view images since the edge fitting depends only on the reconstructed 3D point set. To mitigate these problems, we propose SketchSplat, a method to reconstruct accurate, complete, and compact 3D edges via differentiable multi-view sketch splatting. We represent 3D edges as sketches, which are parametric lines and curves defined by attributes including control points, scales, and opacity. During reconstruction, we iteratively sample Gaussian points from a set of sketches and rasterize the Gaussians onto 2D edge images. Then the gradient of the image loss can be back-propagated to optimize the sketch attributes. Our method bridges 2D edge images and 3D edges in a differentiable manner, which ensures that 3D edges align well with 2D images and leads to accurate and complete results. We also propose a series of adaptive topological operations to reduce redundant edges and apply them along with the sketch optimization, yielding a more compact reconstruction. Finally, we contribute an accurate 2D edge detector that improves the performance of both ours and existing methods. Experiments show that our method achieves state-of-the-art accuracy, completeness, and compactness on a benchmark CAD dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Edges serve as one of the most fundamental elements of visual perception, playing a crucial role in understanding the geometry and structure of the 3D world. Unlike point-based representations, continuous edges offer a more compact and perceptually aligned way to describe 3D structures. As a" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 555, + 497 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 497 + ], + "type": "image", + "image_path": "d6f51cef5ca6b312b3eaeaeb38899e5d7d4e6cd889cff2f3e06149849a622cdf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 506, + 555, + 616 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 616 + ], + "type": "text", + "content": "Figure 1. We propose SketchSplat, a multi-view 3D edge reconstruction method based on differentiable sketch splatting. Previous methods [7, 19] first recover edge points and then fit them to 3D edges, which is sensitive to noise and often leads to incomplete results. Instead, we sample Gaussian points from 3D sketches and rasterize them to compute a rendering loss, enabling direct edge parameter optimization under image supervision. Additionally, we design adaptive topology control methods to refine edge structures during training and contribute an improved 2D edge detection method for more accurate reconstruction." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "cornerstone of computer vision and graphics, 3D edge reconstruction underpins a wide range of applications, including CAD modeling [10, 20, 56], SLAM [13, 24, 41, 50], and autonomous driving [8, 35, 36]. Despite its significance, achieving high-quality 3D edge reconstructions remains a challenging task due to the inherent difficulty of ensuring" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25649" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "the accuracy and completeness of reconstructed edges and maintaining the connectivity of the reconstructed structure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 294, + 193 + ], + "type": "text", + "content": "Traditional methods [5, 12, 23, 48] usually detect 2D line segments in each view. Then the 2D lines are lifted to 3D via line matching and triangulation across views. While these methods achieve high accuracy and efficiency, multiview consistent detection and robust line matching are difficult to achieve, leading to incomplete results. In addition, these methods are usually not capable of handling lines and curves simultaneously." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 295, + 399 + ], + "type": "text", + "content": "In recent years, differentiable rendering has emerged as a promising technique to reconstruct 3D contents for novel view synthesis [3, 27, 28, 55], surface reconstruction [11, 22, 25, 47], and animation [29, 34, 49]. In the field of edge reconstruction, several works introduced the idea of differentiable rendering to recover edge information [7, 19, 54]. Instead of detecting parametric 2D edge segments explicitly, they leverage pixel-wise 2D edge intensity images [33, 44] as a supervision signal and optimize a point-based 3D edge representation via differentiable multiview rendering, resulting in an extracted 3D edge point set. Parametric 3D edges are then obtained by fitting lines and curves to the extracted 3D points. Since the edge-fitting step relies purely on the recovered 3D edge points but not the input 2D edge images, the fitted 3D edges may not align well with the 2D edge images. In addition, noise in the 3D edge points usually leads to discontinuous and fragmented edges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 294, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 294, + 484 + ], + "type": "text", + "content": "To address these limitations, we propose SketchSplat, a differentiable edge reconstruction framework that directly optimizes parametric 3D edges by fitting them to the input 2D edge images using differentiable rendering. Our experimental results demonstrate that this leads to more accurate, complete, and compact reconstructions, more faithfully reflecting the original connectivity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": "Firstly, we propose to model 3D edges as sketches, which are parametric lines and curves with control points, scales, and opacity. These sketches can be initialized from existing methods [7, 19, 54], and can be sampled as a set of 3D Gaussian points for differentiable rendering. To optimize the sketches, we sample Gaussian points from the sketches and rasterize them in each view to calculate the rendering loss between the rendered and the ground-truth edge image. Then the error can be back-propagated to update the sketch attributes. Since we optimize our sketch parameters directly using the 2D edge images, the resulting edges will align well with the 2D edges in the input images. To further improve the compactness of the reconstructed edges, we leverage a series of topological operations to connect, merge, and filter sketches during optimization. Finally, we observe that one common limitation of existing methods [7, 19, 54] lies in the inaccuracy of 2D edge images. To mitigate this, we propose an accurate and robust edge detection method, which is validated to improve both" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "text", + "content": "our approach and existing methods by a large margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 86, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 553, + 133 + ], + "type": "text", + "content": "Experiments demonstrate that our method achieves state-of-the-art performance on the ABC-NEF dataset [54], surpassing previous approaches by over " + }, + { + "bbox": [ + 313, + 86, + 553, + 133 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 86, + 553, + 133 + ], + "type": "text", + "content": " in F-score. In summary, key innovations of our method include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 136, + 553, + 267 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 136, + 553, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 136, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 314, + 136, + 553, + 184 + ], + "type": "text", + "content": "- Differentiable Parametric Optimization: We contribute a pipeline that directly optimizes parametric edge representations via differentiable multi-view rasterization, leading to SOTA accuracy and completeness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 184, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 231 + ], + "type": "text", + "content": "- Adaptive Topological Control: We propose a series of topological operations to further enhance edge continuity and reduce redundant edges during training, yielding more compact reconstruction results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 267 + ], + "type": "text", + "content": "- Accurate 2D Edge Detection: We propose a more accurate 2D edge detection method, which shows large improvements for both our approach and existing methods." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 280, + 400, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 280, + 400, + 293 + ], + "spans": [ + { + "bbox": [ + 314, + 280, + 400, + 293 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 479 + ], + "type": "text", + "content": "Line Matching and Triangulation. Conventional approaches reconstruct 3D lines [2, 4-6, 12, 26] or curves [15, 16, 38] by enforcing multi-view geometric constraints. Given detected 2D edge segments and their descriptors, they match the line descriptors across multiple views and perform triangulation to lift 2D edges into 3D. However, these methods face challenges in consistently detecting edges and robustly matching lines, especially when occlusions cause endpoint inconsistencies across views. Despite notable advancements by involving structural [39, 48] or learning-based priors [23, 30-32], line detection and matching remain critical bottlenecks, limiting overall performance. While most of the methods can only handle a single type of edges, in this paper we aim at recovering both lines and curves." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Differentiable Rendering Methods. The advancement of differentiable rendering opens up a new way for 3D reconstruction [22, 27, 47]. Inspired by NeRF [27], NEF [54] proposes to optimize an implicit neural radiance field with a multi-view rendering loss. After training, 3D edge points are extracted from the radiance field and curves are fitted to these points. NEAT [51] builds on VolSDF [52] and uses a signed distance field to represent both object surfaces and endpoints of 3D lines. The surface-based modeling approach restricts this method to objects with rich texture. EMAP [19] adopts an unsigned distance function (UDF), where sampled points can be displaced towards the zero-level set of the UDF to be more accurate. While methods based on neural field are time consuming to optimize, EdgeGS [7] represents 3D edge fields with explicit Gaussian points and leverages Gaussian splatting [17] to render and optimize Gaussian points, leading to more efficient edge reconstruction. A common issue of these methods" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25650" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "is that the edge fitting process relies only on the extracted 3D edge points, hence noise in the reconstructed 3D points may lead to inaccurate and incomplete edges. Instead, we represent 3D edges as sketches defined parametrically by control points, scales, and opacity, and directly optimize these parameters via differentiable rendering. By leveraging sketch sampling and rasterization, our differentiable rendering aligns 3D edges with 2D edge images, resulting in more accurate and complete reconstructions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 196, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 435 + ], + "type": "text", + "content": "Differentiable Vector Graphics Optimization. DiffVG [21] proposes a differentiable rasterizer that can optimize 2D vector graphics with 2D image supervision in a differentiable manner. However, the differentiation process relies on distance measurements between the rasterized curves and pixels, which needs to be custom tailored for each type of curve. 3Doodle [9] extends the idea of DiffVG [21] to reconstruct 3D curve-based object abstractions using multiple input views. They optimize a set of Bezier curves by minimizing the perceptual loss between the rasterized image and multi-view CLIP feature maps [37, 45]. Since CLIP features only contain abstract information, 3Doodle is not suitable for recovering accurate edges of objects like CAD models. Moreover, they approximate the projection of Bezier curves using an orthographic camera model. Instead, we get inspiration from 3DGS [17] and propose to optimize 3D sketches via sampling and rasterization of Gaussian points sampled from sketch parameters, which can be easily extended to different types of curves and achieves accurate 3D edge reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 447, + 262, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 262, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 262, + 461 + ], + "type": "text", + "content": "3. Preliminaries: 3D Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": "We leverage 3D Gaussian splatting (3DGS) [17] for initialization and differentiable sketch rendering. 3DGS proposes to model 3D scenes with a collection of anisotropic 3D Gaussian points. Each point is defined as a Gaussian distribution " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "G(x) = e^{-\\frac{1}{2}(x - \\mu)^T\\Sigma^{-1}(x - \\mu)}" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mu \\in R^3" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " is the center position or mean, and " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\Sigma \\in R^{3\\times 3}" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " is the covariance matrix. Each Gaussian has an opacity " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "o \\in R" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " and a view-dependent color value " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "c(d) = SH(d) \\in R^3" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " modeled with spherical harmonics. Following Zwicker et al. [57], the Gaussians can be rendered to 2D images using an efficient volume rendering approach, where the color for each camera ray " + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 468, + 295, + 613 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 622, + 295, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 622, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 67, + 622, + 295, + 656 + ], + "type": "interline_equation", + "content": "c (r) = \\sum_ {i \\in N} c _ {i} (d (r)) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\alpha_ {i} = o _ {i} g _ {i} (r), \\tag {1}", + "image_path": "1fd0314cbbced2b6e12f9182dba5d905d20a0798e6a89a60ab6c79baecb26727.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "g_{i}(r)" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " is the line integral of the 3D Gaussian " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " along ray " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": ", and the indices " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " are ordered in front to back manner. By calculating the gradient of a loss " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " between the rendered and ground truth image, the parameters of all Gaussians can" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "be optimized using gradient descent. A typical loss function is a combination of an " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " and a perceptual loss such as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 369, + 106, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 106, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 369, + 106, + 555, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {1} + (1 - \\lambda) \\mathcal {L} _ {D - S S I M}. \\tag {2}", + "image_path": "5b48d8f50426fb8c67d111e0962bcab77454801eb0679fe1a9fd2638d22d48c4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 129, + 555, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 201 + ], + "type": "text", + "content": "Given input images, the initial Gaussians are usually obtained from SfM [40] methods. During training, several operations are adopted to adaptively control the density and distribution of Gaussians. For example, Gaussians with low opacity will be culled, and Gaussians with high gradient will be duplicated or split to densify specific regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 554, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 554, + 249 + ], + "type": "text", + "content": "In SketchSplat, we leverage 3DGS [17] as an intermediary to bridge 2D edge images and 3D parametric edges, enabling the optimization of 3D edges through differentiable image rendering." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 261, + 370, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 370, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 370, + 273 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "text", + "content": "Given multi-view posed RGB images as input, our target is to recover the parametric edges for the 3D objects. Fig. 2 provides an overview of our approach. We firstly compute an edge image for each RGB image using our proposed 2D edge detector. Then we represent edges as 3D parametric sketches, including both lines " + }, + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "inline_equation", + "content": "l \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "text", + "content": " and third-order Bézier curves with four 3D control points " + }, + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{4 \\times 3}" + }, + { + "bbox": [ + 313, + 281, + 555, + 484 + ], + "type": "text", + "content": ". We initialize a set of sketches from existing edge reconstruction methods, and treat their control points, along with persketch scales and opacity, as estimizable parameters. During optimization, we use a sample-based method to convert sketches to Gaussian points, which can be rasterized onto 2D images to calculate the image loss. We back-propagate the gradient of the loss to optimize the sketch parameters. In addition, we introduce a set of topological operations to adaptively refine the sketches during training, yielding more accurate and compact edge reconstruction results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 492, + 444, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 444, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 444, + 504 + ], + "type": "text", + "content": "4.1. Edge Image Extraction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Previously, some methods [23, 51] extract edges with line detectors like LSD [46] and SOLD2 [30]. While these methods produce accurate parametric lines in 2D, they may fail to handle curve regions and result in incomplete results. Recent methods [7, 19, 54] use neural network-based edge detectors [33, 44] to detect edge intensity maps as supervision signal. However, since these detectors are not trained specifically on CAD-like objects, they may miss some significant edges in 2D or wrongly recover shading textures as real edges. Finally, we observed that these detectors introduce bias by consistently shifting detected edges slightly from certain real edges, which leads to multi-view inconsistency in 3D edge reconstruction (see supplementary). We address these issues by proposing a new edge detection method, which leverages geometric cues (e.g., per-pixel depth and normal estimation of 2D RGB images) to produce more accurate edge detection results." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25651" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 245 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 245 + ], + "type": "image", + "image_path": "f08285c5181b9f452d9955574aa0a5bf4bd45b2af4b31cc23ad11f38d28ddc57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 249, + 196, + 259 + ], + "lines": [ + { + "bbox": [ + 102, + 249, + 196, + 259 + ], + "spans": [ + { + "bbox": [ + 102, + 249, + 196, + 259 + ], + "type": "text", + "content": "(A) Sketch initialization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 249, + 493, + 260 + ], + "lines": [ + { + "bbox": [ + 296, + 249, + 493, + 260 + ], + "spans": [ + { + "bbox": [ + 296, + 249, + 493, + 260 + ], + "type": "text", + "content": "(B) Differentiable Multi-view Sketch Optimization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 327 + ], + "type": "text", + "content": "Figure 2. Overview of SketchSplat pipeline: (a) We extract edge images using a novel approach (Sec. 4.1). (b) We then obtain initial edge points using EdgeGS [7] and (c) fit parametric sketches defined as lines and third-order Bézier curves. The core step of our approach is to optimize the sketches by (d) sampling and rasterizing them in a differentiable manner and back-propagating gradients of the image loss to (e) sketch parameters. In addition, during training we apply (f) a set of topological operations to improve compactness and connectivity of the reconstructed sketches. (g) After training, we take the optimized geometric parameters of sketches as the reconstructed 3D edges." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": "For each RGB image, we combine information from three parts: foreground mask " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": ", depth map " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": ", and normal map " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": ". The foreground mask can be extracted using the image alpha-channel (if available), a thresholding operation or possibly using semantic segmentation, providing object boundary information. We obtain a depth image for each view by reconstructing the scene with 2DGS [14] and rendering depth maps from each view. We use a normal estimator [53] to generate normal maps " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": " for each view. Then we calculate gradient maps " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "g(D)" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "g(N)" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": " with the Sobel operator [42], threshold them with " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "t_d" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": ", and combine them via the OR operation, followed by a Gaussian filter " + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "inline_equation", + "content": "G_f(\\cdot)" + }, + { + "bbox": [ + 54, + 347, + 296, + 515 + ], + "type": "text", + "content": " to smooth the boundaries of the edges. In summary, we call our edge detector 2DGS-SN, defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 523, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 523, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 88, + 523, + 295, + 537 + ], + "type": "interline_equation", + "content": "E = G _ {f} * (A | (g (D) > t _ {d}) | (g (N) > t _ {n})). \\tag {3}", + "image_path": "32952a3e74815fad415a5c3fd48108a8f80787ad7bf99f7508ba424dd1e99dd8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 570 + ], + "type": "text", + "content": "In Sec. 5 we show that this simple approach improves existing methods [7] on CAD-like objects by a large margin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 155, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 155, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 155, + 590 + ], + "type": "text", + "content": "4.2. Sketch Splatting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Representation. We represent edges as 3D parametric sketches, including both lines " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "l \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " and third-order Bézier curves with control points " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{4 \\times 3}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". This is the same as the output format of EdgeGS[7]. While EdgeGS optimizes 3D point positions first and then fits 3D lines and curves to the 3D points, we optimize the parameters of lines and curves directly via multi-view differentiable rasterization. In addition to the estimizable geometry parameters " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "(l" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "c)" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", we equip each sketch with an estimizable opacity " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " and a local scale " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " to model the sketch intensity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 372 + ], + "type": "text", + "content": "and thickness, respectively. These attributes will support Gaussian sampling and rasterization in the next steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 386, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 483 + ], + "type": "text", + "content": "Sketch Initialization. Unlike point-based methods [7, 19], it is non-trivial to initialize sketches directly from multi-view edge images. Instead, inspired by 3DGS [17], we first initialize sketches with existing point-based methods like EdgeGS [7]. By default, we select EdgeGS because it provides a sufficient number of short 3D edges to cover the objects, and achieves the highest completeness among previous methods, despite some fragmentation and noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": "Differentiable Sketch Optimization. Our key idea is to bridge the gap between 2D edge images and 3D sketches via differentiable rendering and gradient descent optimization, which significantly improves sketch reconstruction over initialization method EdgeGS [7]. Inspired by 3DGS [17] and DiffVG [21], we choose 3D Gaussians as an intermediate representation to achieve differentiable sketch rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "For each gradient descent step, we first sample points from each sketch with a constant step size (5mm for all experiments), resulting in a point set " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". Each point will be assigned the opacity and scale value of the corresponding sketch. To model the orientation, the main direction of each point is sampled as the local tangent direction of the sketch. Then we follow 3DGS[17] to rasterize the Gaussian points onto a chosen 2D view to generate a rendered edge image " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "E^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". We calculate the " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " loss between " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "E^{*}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and the ground truth edge image " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "E_{gt}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ", and back-propagate the gradient to update the sketch parameters " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "S_{i} = \\{l_{i} \\text{ or } c_{i}, o_{i}, s_{i}\\}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". We fol" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25652" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "low EdgeGS [7] to sample an equal number of foreground and background pixels for " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " loss calculation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 206, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 206, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 206, + 115 + ], + "type": "text", + "content": "4.3. Adaptive Topology Control" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 120, + 295, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 120, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 120, + 295, + 336 + ], + "type": "text", + "content": "A good edge reconstruction, which could be used for CAD modeling for example, should not only accurately match the 2D edge images, but also provide a compact, well-connected set of edges. Neighboring edges should be connected correctly, overlapping sketches should be merged, and the number of sketches should be as small as possible while ensuring accurate matching of 2D edge images. For this purpose we deploy topology control operations adaptively during optimization. While some existing methods [7] do not consider topology optimization, others [19, 54] try to connect and merge neighboring sketches in the post-processing stage. However, this may be sensitive to noise in the extracted 3D point set and cannot ensure that the resulting edges are still consistent with the input multi-view 2D edge images. Instead, since our method directly optimizes sketch parameters using iterative gradient descent, we can execute topological operations along with the multi-view optimization steps, as shown in Fig. 2(f)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": "End-point Merging. In our representation, each sketch has two end-points. If there exist two end-points, one from sketch " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": " and one from sketch " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": ", that are sufficiently close (i.e. the point distance is smaller than threshold " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "th_{connect} = 10mm" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": "), we will connect sketch " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": " by merging those two end-points into one point (see Fig. 2(f-i)). However, this means the number of estimable points could differ from the number of control points " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "|\\{l\\}| * 2 + |\\{c\\}| * 4" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": ". To accommodate this, we maintain an estimable point set " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": ", and the control points of each line " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": " and curve " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": " are represented by indexing point set " + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 350, + 296, + 494 + ], + "type": "text", + "content": ". This representation also supports our additional topological operations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": "Overlapping Sketch Merging. As shown in Fig. 2(f-ii), if one sketch is almost entirely covered by another, the smaller one will be merged into the larger one. Specifically, for sketch " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ", we sample two point sets " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "P_{j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " and calculate the distance between each pair of points. Then for " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ", we calculate the ratio " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "r_{i\\rightarrow j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " of points that have close neighboring points in " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "P_{j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ", where the neighbor threshold is " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "th_{neighbor} = 10mm" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "r_{i\\rightarrow j} > th_{overlap} = 80\\%" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ", sketch " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " will be merged in sketch " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ". If both " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "r_{i\\rightarrow j}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "r_{j\\rightarrow i}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": " surpass " + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "inline_equation", + "content": "th_{overlap}" + }, + { + "bbox": [ + 55, + 508, + 295, + 640 + ], + "type": "text", + "content": ", we will choose the sketch with larger overlapping ratio to be merged." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Co-linear Sketch Merging. If two lines are co-linear and nearly connected, they should be merged into a longer line (see Fig. 2(f-iii)). The co-linear condition is: two lines should have similar direction (" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "th_{dir} = 5^{\\circ}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "), and the maximum of the projective distance/offset should be small" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "enough (" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "th_{offset} = 10mm" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "). In addition, to prevent gaps between two lines, one line will be projected onto the other one to calculate if there is a gap and if the gap is small enough (" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "th_{connect}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "). To merge the lines, we take the two end-points with the furthest distance after projection and delete the other two endpoints from point set " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ". Additionally, to accelerate this procedure, we use the axis-aligned bounding box of each sketch to quickly filter out all the sketch pairs that cannot meet the merging criterion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "text", + "content": "Sketch Filtering. After training, we adopt a multi-view filter method to filter sketches that are not visible. We borrow the refinement idea from EMAP [19]. 3D points can be projected onto 2D edge images to see if they are aligned with the 2D edges by checking the edge intensity of the projected pixel position. A point will be labeled as invisible if it is invisible in more than " + }, + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "text", + "content": " of the views. For each sketch, we will evaluate the visibility for all of the points sampled from the sketch. If more than " + }, + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "inline_equation", + "content": "th_{vis} = 50\\%" + }, + { + "bbox": [ + 313, + 201, + 556, + 369 + ], + "type": "text", + "content": " of the sampled points are invisible points, the sketch will be removed. When rasterizing sketches, some sketches that should not appear due to occlusion still appear in the rendered image, which sometimes leads to incorrect reconstruction. This filter scheme will help filter out such outlier sketches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 379, + 448, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 448, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 448, + 392 + ], + "type": "text", + "content": "4.4. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": "Unless otherwise specified, we initialize with EdgeGS [7], which supplies a sufficient number of 3D edges to capture object structure. For each scene, we optimize the sketches for 1000 epochs with Adam. We update parameters only once for each epoch, where the loss is accumulated across all the training views. All of the scenes are scaled to fit within a " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}^3" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": " bounding box, and topology operation hyperparameters are set accordingly. Distance thresholds (e.g., " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "th_{connect}" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": ") are set based on the desired precision and we set all of them as " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": " of the scene size " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "(1m^3)" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": ". The percentage and angular thresholds (e.g., " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "th_{overlap}" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "inline_equation", + "content": "th_{dir}" + }, + { + "bbox": [ + 313, + 397, + 556, + 589 + ], + "type": "text", + "content": ") work across different scenes without adjustment. All our experiments are implemented on a single RTX A5000 NVIDIA GPU. In average, each scene in ABC-NEF [54] dataset costs about 10 minutes for training, where both initialization and sketch optimization takes about 5 minutes, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 602, + 397, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 397, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 397, + 616 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 623, + 432, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 432, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 432, + 636 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "We compare our method with other multi-view edge reconstruction methods following the evaluation setup of EdgeGS [7]. The results show that our method achieves the state-of-the-art performance in both accuracy and completeness on the ABC-NEF dataset [18, 54]. Our method also produces highly compact results with fast inference speed." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25653" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 539, + 239 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 539, + 239 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 539, + 239 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 539, + 239 + ], + "type": "table", + "html": "
MethodDetectorModalA↓C↓R5↑R10↑R20↑P5↑P10↑P20↑F5↑F10↑F20↑
LIMAP[23]LSDLine9.918.736.282.387.943.087.693.939.084.390.4
SOLD2Line5.929.664.276.679.688.196.497.972.984.086.7
NEF[54]PiDiNeT†Curve11.916.911.462.091.315.768.596.313.064.093.3
PiDiNeTCurve15.116.511.753.393.912.361.395.812.351.888.7
DexiNedCurve21.915.711.348.393.711.558.991.710.842.176.8
EMAP[19]PiDiNeTEdge9.215.630.275.789.535.679.195.432.477.092.2
DexiNedEdge8.88.956.488.994.862.989.995.759.188.994.9
2DGS-SNEdge8.87.963.590.996.270.491.094.966.390.495.1
EdgeGS[7]PiDiNeTEdge11.710.317.173.983.126.087.292.520.679.386.7
DexiNedEdge9.68.442.491.795.849.194.896.345.293.795.7
2DGS-SNEdge7.47.275.791.096.086.995.996.480.392.995.8
OursPiDiNeTEdge10.411.025.188.694.530.989.695.427.588.594.5
DexiNedEdge8.58.647.593.596.154.196.397.150.394.496.3
2DGS-SNEdge6.85.890.895.797.292.996.396.691.395.496.5
", + "image_path": "c761d3eac7ac4270f20e443de41fa3d14ed3655c0ad6578d241e8e5f2b3f68d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on ABC-NEF dataset [54]. Our method achieves the state-of-the-art performance on most metrics. A: Accuracy (mm), C: Completeness (mm), Rx: Recall (threshold " + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "inline_equation", + "content": "x\\mathrm{\\;{mm}}" + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": " ),Px: Precision (mm), Fx: F-score (mm)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 282, + 555, + 524 + ], + "blocks": [ + { + "bbox": [ + 58, + 282, + 555, + 524 + ], + "lines": [ + { + "bbox": [ + 58, + 282, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 58, + 282, + 555, + 524 + ], + "type": "image", + "image_path": "8c8d4e47194c27dc9c43e037339f3201485ad20aa9ba071b39da67dcfcf37012.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 555, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 555, + 556 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on the ABC-NEF dataset [54]. Compared to the baselines, our method achieves superior accuracy and completeness, along with a favorable balance between completeness and compactness." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 576, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 576, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 54, + 576, + 297, + 708 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three datasets: ABC-NEF [18, 54] DTU [1], and Replica [43]. The ABC-NEF dataset [18, 54] comprises 115 CAD models, each containing 50 RGB images of the target object and a parametric edge file providing ground truth 3D edges. Following EMAP [19] and EdgeGS [7], we use 82 CAD models from ABC-NEF, excluding the ones with cylinders and spheres since smooth convex surfaces will produce silhouette edges that are not consistent in multi-views. For 2D edge image generation, we use PiDiNet [44], DexiNed [33], and our proposed 2DGS-SN as 2D edge detectors. The" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 576, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 556, + 649 + ], + "type": "text", + "content": "DTU dataset [1] contains multi-view images of real-world scenes. We follow EdgeGS [7] and evaluate on 6 DTU scenes. We use edge images and the pseudo ground truth edges provided by EMAP [19]. The Replica dataset [43] consists of multiple indoor scenes. We follow EdgeGS [7] to conduct qualitative evaluation on Replica [43]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Baselines. We compare our method with five baseline methods, including the state-of-the-art 3D line mapping method (LIMAP [23]), and several methods based on volumetric differentiable rendering (NEF [54], NEAT [51]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25654" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 269, + 175 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 269, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 269, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 269, + 175 + ], + "type": "image", + "image_path": "acb3124d7edac71cac255d2bffe7da517135cd4a7ea61b907a43a6402e83c6ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 54, + 183, + 555, + 207 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on the DTU and Replica datasets [1, 43]. Enum is the edge number and edges are visualized with varied colors. Compared to the baselines, our method achieves higher compactness and retains faithful completeness." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 70, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 276, + 70, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 276, + 70, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 276, + 70, + 553, + 175 + ], + "type": "image", + "image_path": "db3d13ef0cdde8892a7c652c6e24f491c908fb5c181d467651da1c94be2866ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 264 + ], + "type": "text", + "content": "EMAP [19], EdgeGS [7]). Following EMAP and EdgeGS, we do not evaluate NEAT on the ABC-NEF dataset since NEAT often fails to train for texture-less objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 340 + ], + "type": "text", + "content": "Metrics. Following the metrics used in baseline methods [7, 19], we evaluate Accuracy (A) and Completeness (C) in millimeters, and Recall (R), Precision (P), and F-score (F) as percentages at three thresholds (5mm, 10mm, 20mm). See supplementary material for more details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 348, + 113, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 113, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 113, + 360 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "text", + "content": "Evaluation on ABC-NEF [54]. The qualitative and quantitative evaluation results on ABC-NEF dataset are shown in Fig. 3 and Tab. 1, respectively. Overall, the proposed method achieves a new state-of-the-art across most of the metrics. To the best of our knowledge, SketchSplat is the first method to achieve both accuracy (A) and completeness (C) below " + }, + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "inline_equation", + "content": "7.0\\mathrm{mm}" + }, + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "text", + "content": ", while also exceeding " + }, + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 365, + 295, + 461 + ], + "type": "text", + "content": " in recall (R5), precision (P5), and F-score (F5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "content": "Tab. 1 shows LIMAP [23] achieves the highest accuracy due to its robust line mapping method. However, its completeness and recall are low, as LIMAP struggles with curve reconstruction and consistent edge detection results (Fig. 3(b)). NEF [54] produces more complete results in Fig. 3(c), but performs poorly on metrics as its fitting process is sensitive to noise in the extracted 3D points. EMAP [19] outperforms NEF with its unsigned distance field modeling and point shifting mechanism. However, Fig. 3(d) shows that EMAP still generates duplicated edges and misses some edges. With PidiNeT and DexiNed as 2D detector, EdgeGS [7] surpasses EMAP in recall and precision at " + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "inline_equation", + "content": "10mm" + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "inline_equation", + "content": "20mm" + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "content": " thresholds, but the trend reverses for measurements under " + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "inline_equation", + "content": "5mm" + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "content": ". This is due to the thickness of 2D edges introduces bias in smaller distance evaluations. When using our detector 2DGS-SN, EdgeGS performs better than EMAP to a large margin under " + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "inline_equation", + "content": "5mm" + }, + { + "bbox": [ + 55, + 462, + 295, + 689 + ], + "type": "text", + "content": " metrics. As shown in Fig. 3, our SketchSplat achieves the highest precision and completeness compared to all the baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab.1, the proposed 2DGS-SN significantly boosts the performance of EMAP, EdgeGS, and Sketch-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 360, + 224, + 511, + 290 + ], + "blocks": [ + { + "bbox": [ + 360, + 224, + 511, + 290 + ], + "lines": [ + { + "bbox": [ + 360, + 224, + 511, + 290 + ], + "spans": [ + { + "bbox": [ + 360, + 224, + 511, + 290 + ], + "type": "table", + "html": "
MethodE-numTime (h)
NEF [54]22.91:26
EMAP [19]45.82:30
EdgeGS [7]140.80:05
ours44.30:10
", + "image_path": "a08edc33a78ddf092bf4501ada7bb5fe9c1918854d417c52ff5ccfe20e439eea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 298, + 554, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 321 + ], + "type": "text", + "content": "Table 2. Comparison of the number of reconstructed edges (E-num) and training time on ABC-NEF dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "Splat, addressing the bias from PiDiNet[44] and Dex-iNed [33]. Notably, across all edge detectors, our method consistently achieves higher Recall, Accuracy, and F-score than EdgeGS [7], demonstrating effectiveness of Sketch-Splat in aligning 3D edges with 2D edge images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 538 + ], + "type": "text", + "content": "Evaluation on DTU [1] and Replica [43]. Fig. 4 presents qualitative comparisons on the DTU and Replica datasets. Compared to the baselines, our method consistently produces accurate and complete results using significantly fewer edges, demonstrating a strong balance between completeness and compactness. Although the ground-truth edges in the DTU scenes are not sufficiently accurate for reliable quantitative evaluation [7], our method still achieves metrics comparable to state-of-the-art methods EMAP and EdgeGS (see supplementary material for more details)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 554, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 626 + ], + "type": "text", + "content": "Speed Analysis. The initialization with EdgeGS costs about 5 minutes. For optimization part, thanks to the efficiency of rasterization-based optimization, our method achieves comparable speed (about 5 minute per scene) to EdgeGS while being significantly faster than other neural network-based approaches [19, 54], as shown in Tab. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Compactness. EdgeGS achieves higher recall than NEF and EMAP due to its strict edge fitting process, which produces numerous short edges, covering more edge regions as the number of edges increases. Instead, our method incorporates adaptive topology control to reduce redundant edges, as shown in Tab. 2, Fig. 4, and Fig. 5(a,b,d). This" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "25655" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 77, + 154, + 204 + ], + "blocks": [ + { + "bbox": [ + 58, + 77, + 154, + 204 + ], + "lines": [ + { + "bbox": [ + 58, + 77, + 154, + 204 + ], + "spans": [ + { + "bbox": [ + 58, + 77, + 154, + 204 + ], + "type": "image", + "image_path": "680192c8ba1f43ce645fd6dbf40c92af000d6b8df06cc92819349e85a2c039ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 205, + 142, + 215 + ], + "lines": [ + { + "bbox": [ + 70, + 205, + 142, + 215 + ], + "spans": [ + { + "bbox": [ + 70, + 205, + 142, + 215 + ], + "type": "text", + "content": "(a) Initialization (EdgeGS)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 78, + 254, + 205 + ], + "blocks": [ + { + "bbox": [ + 157, + 78, + 254, + 205 + ], + "lines": [ + { + "bbox": [ + 157, + 78, + 254, + 205 + ], + "spans": [ + { + "bbox": [ + 157, + 78, + 254, + 205 + ], + "type": "image", + "image_path": "c4a7d33a00db868f194705e7ddcf51748eeaa53a735c5d08707d89630f5afe73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 205, + 249, + 215 + ], + "lines": [ + { + "bbox": [ + 162, + 205, + 249, + 215 + ], + "spans": [ + { + "bbox": [ + 162, + 205, + 249, + 215 + ], + "type": "text", + "content": "(b) Ours (w/o Topology Control)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 256, + 78, + 353, + 204 + ], + "blocks": [ + { + "bbox": [ + 256, + 78, + 353, + 204 + ], + "lines": [ + { + "bbox": [ + 256, + 78, + 353, + 204 + ], + "spans": [ + { + "bbox": [ + 256, + 78, + 353, + 204 + ], + "type": "image", + "image_path": "9e4bd38b7d8a0e1d516eb3ee29ff5294fd52ae5bad1940b4e7ed31906c8c342b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 205, + 338, + 214 + ], + "lines": [ + { + "bbox": [ + 271, + 205, + 338, + 214 + ], + "spans": [ + { + "bbox": [ + 271, + 205, + 338, + 214 + ], + "type": "text", + "content": "(c) Ours (w/o 2DGS-SN)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 356, + 78, + 452, + 205 + ], + "blocks": [ + { + "bbox": [ + 356, + 78, + 452, + 205 + ], + "lines": [ + { + "bbox": [ + 356, + 78, + 452, + 205 + ], + "spans": [ + { + "bbox": [ + 356, + 78, + 452, + 205 + ], + "type": "image", + "image_path": "b472d9b45f4c50771158f5ae1d104fbceb38c9cc778fca995a8c6d181ac0574b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 205, + 416, + 214 + ], + "lines": [ + { + "bbox": [ + 392, + 205, + 416, + 214 + ], + "spans": [ + { + "bbox": [ + 392, + 205, + 416, + 214 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 226, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 553, + 239 + ], + "type": "text", + "content": "Figure 5. Ablation study on the key components of SketchSplat. Removing key components leads to noisier or redundant reconstructions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 457, + 78, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 457, + 78, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 457, + 78, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 457, + 78, + 553, + 205 + ], + "type": "image", + "image_path": "5dbe9e8cf8a17c27180f39223ca9140be3d9da8b7873a077494187c19545a86f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 205, + 514, + 214 + ], + "lines": [ + { + "bbox": [ + 494, + 205, + 514, + 214 + ], + "spans": [ + { + "bbox": [ + 494, + 205, + 514, + 214 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 59, + 256, + 293, + 316 + ], + "blocks": [ + { + "bbox": [ + 59, + 256, + 293, + 316 + ], + "lines": [ + { + "bbox": [ + 59, + 256, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 59, + 256, + 293, + 316 + ], + "type": "table", + "html": "
MethodA↓C↓R5↑P5↑F5↑E-num
Ours6.85.890.892.991.344.28
w/o 2DGS-SN8.58.647.554.150.333.52
w/o Merge6.65.791.892.791.7140.72
w/o Filter6.85.890.992.991.344.37
", + "image_path": "2b5c2be289eb38030d72764ac3150f777342d96a1b13e9d18e620360f401afa1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 58, + 369, + 294, + 429 + ], + "blocks": [ + { + "bbox": [ + 55, + 323, + 295, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 357 + ], + "type": "text", + "content": "Table 3. Ablation on the proposed components using ABC-NEF. w/o 2DGS-SN denotes SketchSplat trained with edge images detected by DexiNed, while the others are trained with 2DGS-SN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 369, + 294, + 429 + ], + "lines": [ + { + "bbox": [ + 58, + 369, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 294, + 429 + ], + "type": "table", + "html": "
MethodA↓C↓R5↑P5↑F5↑E-num
EdgeGS7.47.275.786.980.3140.76
Ours (EdgeGS)6.85.890.892.991.344.28
EMAP8.87.963.570.466.348.71
Ours (EMAP)6.96.288.890.889.331.70
", + "image_path": "980bf2acaa6efc5a04361030ee56a67c267381d7c3ff4470612dcd0d41180183.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 437, + 295, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 481 + ], + "type": "text", + "content": "Table 4. Ablation on initialization methods using ABC-NEF. Ours (EdgeGS) and Ours (EMAP) denote our method initialized with EdgeGS and EMAP, respectively. 2DGS-SN is chosen as 2D edge detector." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 529 + ], + "type": "text", + "content": "mechanism helps SketchSplat achieve higher compactness, while maintaining the highest accuracy and completeness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 539, + 119, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 119, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 119, + 551 + ], + "type": "text", + "content": "5.3. Ablation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "Key Components. We use the ABC-NEF dataset to conduct an ablation study on the key contributions of our method, with results presented in Fig. 5 and Tab. 3. Fig. 5(b,d) shows that without topology control, noisy edge segments appear around the ground truth edges, with some extending beyond the object boundary. Incorporating topology control, particularly merging operations, effectively reduces the required edge count while maintaining comparable performance on other metrics (see Tab. 3). In addition, compared to edges reconstructed using DexiNed 2D edge images (Fig. 5(c)), our proposed 2DGS-SN better aligns edges with the ground truth (Fig. 5(d)), improving all metrics by effectively reducing bias in 2D edge images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 258, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 342 + ], + "type": "text", + "content": "Initialization Methods. In our main experiments, we use EdgeGS for initialization. However, we demonstrate SketchSplat can also achieve strong results when initialized with alternative methods like EMAP. As shown in Tab. 4, our approach consistently outperforms the initialization methods by a large margin and achieves comparable results across different initializations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 357, + 388, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 388, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 388, + 370 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 380, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 548 + ], + "type": "text", + "content": "We present SketchSplat, a method for parametric 3D edge reconstruction from calibrated multi-view images. Unlike previous approaches that rely on 3D edge point reconstruction and edge fitting, SketchSplat directly optimizes edges through multi-view sketch splatting, ensuring accurate alignment with input 2D images. A series of topological operations are proposed to adaptively refine the edge structure during training, which reduces redundancy while maintaining high accuracy and completeness. Additionally, an improved 2D edge detection method is adopted to achieve higher reconstruction quality. Experiments show that SketchSplat achieves the state-of-the-art accuracy and completeness on the CAD dataset while maintaining compact results and efficient training speed." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "content": "Acknowledgements This research is based in part upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via IARPA R&D Contract No. 140D0423C0076. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODN, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25656" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Caroline Baillard, Cordelia Schmid, Andrew Zisserman, and Andrew Fitzgibbon. Automatic line matching and 3d reconstruction of buildings from multiple views. In ISPRS Conference on Automatic Extraction of GIS Objects from Digital Imagery, pages 69-80, 1999. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[4] Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment. Computer vision and image understanding, 100(3): 416-441, 2005. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 335 + ], + "type": "text", + "content": "[5] Andrea Bignoli, Andrea Romanoni, Matteo Matteucci, and Politecnico di Milano. Multi-view stereo 3d edge reconstruction. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 867-875. IEEE, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 379 + ], + "type": "text", + "content": "[6] Manmohan Chandraker, Jongwoo Lim, and David Kriegman. Moving in stereo: Efficient structure and motion using lines. In 2009 IEEE 12th International Conference on Computer Vision, pages 1741-1748. IEEE, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "text", + "content": "[7] Kunal Chelani, Assia Benbihi, Torsten Sattler, and Fredrik Kahl. Edgegaussians-3d edge mapping via gaussian splatting. arXiv preprint arXiv:2409.12886, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 468 + ], + "type": "text", + "content": "[8] Wentao Cheng, Sheng Yang, Maomin Zhou, Ziyuan Liu, Yiming Chen, and Mingyang Li. Road mapping and localization using sparse semantic visual features. IEEE robotics and automation letters, 6(4):8118-8125, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "text", + "content": "[9] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 556 + ], + "type": "text", + "content": "[10] Hyelim Choi, Minji Lee, Jiseock Kang, and Dongjun Lee. Online 3d edge reconstruction of wiry structures from monocular image sequences. IEEE Robotics and Automation Letters, 8(11):7479-7486, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 601 + ], + "type": "text", + "content": "[11] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[12] Manuel Hofer, Michael Maurer, and Horst Bischof. Improving sparse 3d models for man-made environments using line-based 3d reconstruction. In 2014 2nd International Conference on 3D Vision, pages 535-542. IEEE, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[13] Petr Hruby, Shaohui Liu, Rémi Pautrat, Marc Pollefeys, and Daniel Barath. Handbook on leveraging lines for two-view relative pose estimation. In 2024 International Conference on 3D Vision (3DV), pages 376-386. IEEE, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[14] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "curate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[15] Kahl and August. Multiview reconstruction of space curves. In Proceedings Ninth IEEE International Conference on Computer Vision, pages 1017-1024. IEEE, 2003. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[16] Jeremy Yermiyahou Kaminski, Michael Fryers, Amnon Shashua, and Mina Teicher. Multiple view geometry of non-planar algebraic curves. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 181-186. IEEE, 2001. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 297 + ], + "type": "text", + "content": "[18] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9601-9611, 2019. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "text", + "content": "[19] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Rémi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3d neural edge reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21219-21229, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "text", + "content": "[20] Pu Li, Jianwei Guo, Huibin Li, Bedrich Benes, and Dong-Ming Yan. Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4671-4680, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[21] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 455, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 510 + ], + "type": "text", + "content": "[22] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "text", + "content": "[23] Shaohui Liu, Yifan Yu, Rémi Pautrat, Marc Pollefeys, and Viktor Larsson. 3d line mapping revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21445-21455, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 555, + 600 + ], + "type": "text", + "content": "[24] Yuncai Liu, Thomas S Huang, and Olivier D Faugeras. Determination of camera location from 2-d to 3-d line and point correspondences. IEEE Transactions on pattern analysis and machine intelligence, 12(1):28-37, 1990. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[25] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[26] Branislav Micusik and Horst Wildenauer. Structure from motion with line segments under relaxed endpoint constraints. International Journal of Computer Vision, 124:65-79, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25657" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[28] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[29] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 261 + ], + "type": "text", + "content": "[30] Rémi Pautrat, Juan-Ting Lin, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Sold2: Self-supervised occlusion-aware line description and detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11368-11378, 2021. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "text", + "content": "[31] Rémi Pautrat, Daniel Barath, Viktor Larsson, Martin R Oswald, and Marc Pollefeys. Deeplsd: Line segment detection and refinement with deep image gradients. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17327-17336, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "type": "text", + "content": "[32] Rémi Pautrat, Iago Suárez, Yifan Yu, Marc Pollefeys, and Viktor Larsson. Gluestick: Robust image matching by sticking points and lines together. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9706-9716, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[33] Xavier Soria Poma, Edgar Riba, and Angel Sappa. Dense extreme inception network: Towards a robust cnn model for edge detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1923-1932, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 536 + ], + "type": "text", + "content": "[35] Zhijian Qiao, Zehuan Yu, Huan Yin, and Shaojie Shen. Online monocular lane mapping using catmull-rom spline. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7179–7186. IEEE, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[36] Tong Qin, Yuxin Zheng, Tongqing Chen, Yilun Chen, and Qing Su. A light-weight semantic map for visual localization towards autonomous driving. In 2021 IEEE international conference on robotics and automation (ICRA), pages 11248-11254. IEEE, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 657 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Luc Robert and Olivier D Faugeras. Curve-based stereo: Figural continuity and curvature. In Proceedings. 1991 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 57-58. IEEE Computer Society, 1991. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[39] Grant Schindler, Panchapagesan Krishnamurthy, and Frank Dellaert. Line-based structure from motion for urban environments. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06), pages 846-853. IEEE, 2006. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[40] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "text", + "content": "[41] Paul Smith, Ian Reid, and Andrew J Davison. Real-time monocular slam with straight lines. In BMVC, pages 17-26, 2006. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "text", + "content": "[42] Irwin Sobel, Gary Feldman, et al. A 3x3 isotropic gradient operator for image processing. a talk at the Stanford Artificial Project in, 1968:271-272, 1968. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "text", + "content": "[43] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "type": "text", + "content": "[44] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5117-5127, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 410 + ], + "type": "text", + "content": "[45] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[46] Rafael Grompone Von Gioi, Jeremie Jakubowicz, Jean-Michel Morel, and Gregory Randall. Lsd: A fast line segment detector with a false detection control. IEEE transactions on pattern analysis and machine intelligence, 32(4): 722-732, 2008. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[48] Dong Wei, Yi Wan, Yongjun Zhang, Xinyi Liu, Bin Zhang, and Xiqi Wang. Elsr: Efficient line segment reconstruction with planes and points guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15807-15815, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "text", + "content": "[49] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[50] Xiaolong Wu, Patricio A Vela, and Cedric Pradalier. Robust monocular edge visual odometry through coarse-to-fine data association. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4923-4929. IEEE, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 679, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 714 + ], + "type": "text", + "content": "[51] Nan Xue, Bin Tan, Yuxi Xiao, Liang Dong, Gui-Song Xia, Tianfu Wu, and Yujun Shen. Neat: Distilling 3d wireframes from neural attraction fields. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25658" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 376 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19968-19977, 2024. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[52] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 184 + ], + "type": "text", + "content": "[53] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. Stablenormal: Reducing diffusion variance for stable and sharp normal. ACM Transactions on Graphics (TOG), 2024. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "type": "text", + "content": "[54] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. Nef: Neural edge fields for 3d parametric curve reconstruction from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8486-8495, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "text", + "content": "[55] Haiyang Ying, Baowei Jiang, Jinzhi Zhang, Di Xu, Tao Yu, Qionghai Dai, and Lu Fang. Parf: Primitive-aware radiance fusion for indoor scene novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17706-17716, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[56] Chao Zhang, Romain Pinquié, Arnaud Polette, Gregorio Carasi, Henri De Charnace, and Jean-Philippe Pernot. Automatic 3d cad models reconstruction from 2d orthographic drawings. Computers & Graphics, 114:179–189, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 376 + ], + "type": "text", + "content": "[57] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25659" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_content_list.json b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d8b833cd915e08d7221daf51914166d57733d8f3 --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_content_list.json @@ -0,0 +1,1946 @@ +[ + { + "type": "text", + "text": "Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weili Zeng $^{1*}$ , Ziyuan Huang $^{2}$ , Kaixiang Ji $^{2}$ , Yichao Yan $^{1\\dagger}$", + "bbox": [ + 267, + 209, + 727, + 228 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Ant Group", + "bbox": [ + 109, + 228, + 885, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zw1666,yanyichao}@sjtu.edu.cn, ziyuan.huang@u.nus.edu, kaixiang.jkx@antgroup.com", + "bbox": [ + 137, + 250, + 854, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 297, + 326, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that SkipVision reduces training time by up to $35\\%$ , inference FLOPs by $75\\%$ , and latency by $45\\%$ , while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency. Our project page can be found here.", + "bbox": [ + 88, + 330, + 485, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based large-scale models have significantly advanced progress in Artificial Intelligence (AI) [10, 86, 101]. Their remarkable capabilities have driven the emergence of modern Multimodal Large Language Models (MLLMs) [2, 3, 69, 88], which demonstrate vision-language competencies comparable to human performance. Empirical evidence suggests that scaling laws continue to", + "bbox": [ + 89, + 753, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/854eda7a7f731484f8da78f37fae6a357bb004a72bcf968b8248987b0bd9c0f4.jpg", + "image_caption": [ + "Figure 1. Performance-efficiency trade-off curve. Each circle denotes a model configuration, where our models utilize the Skip-Vision framework, with CoS [40], LLaVA-HD [67] and LLaVA [69] serving as baselines. Circle sizes reflect the inference FLOPs ratio. Skip-Vision demonstrate superior performance, scaling effectively with increased FLOPs and data and achieving higher inference efficiency when compared to baselines and other efficient MLLM methods. All methods utilize LLaMA3 8B as the foundational large language model." + ], + "image_footnote": [], + "bbox": [ + 522, + 297, + 901, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "be effective, primarily across three dimensions: visual scaling [40, 67, 112], data scaling [63, 64], and model scaling [56, 104]. However, these scaling methods present substantial efficiency issues, significantly increasing the training and inference burden. In a typical LLaVA-1.5 framework [69], where the model receives 576 visual tokens and 40 text tokens, inference using Vicuna-13B [25] LLM backbone demands 200 trillion FLOPS [42]. These motivate our pursuit of an MLLM architecture that optimally balances efficiency and performance.", + "bbox": [ + 511, + 717, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in Multimodal Large Language Models (MLLMs), such as Chain-of-Sight (CoS)[40] and LLaVA-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Weili Zeng interned at Ant Group.", + "bbox": [ + 112, + 875, + 442, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 112, + 887, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21384", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/64ce9bea74e44b023b4a2e1bda2858c8d7e02926213e779c6c78d6ea8aefb695.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Llama layerComplexityParameters
FFN10.5LNC25.64B(70%)
Attention2LN2C + 1.5NC21.34B(17%)
Embedding-0.5B(6%)
", + "bbox": [ + 109, + 88, + 467, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Computational complexity and parameter statistics for Llama3 8B. L denotes the number of blocks, N represents the number of input tokens, and C is the feature dimensionality", + "bbox": [ + 89, + 162, + 482, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next[68], have improved performance by increasing visual tokens. However, during the Supervised Fine-Tuning (SFT) stage, the computational cost rises dramatically, leading to diminishing returns and making the scaling unsustainable. As shown in Figure 1, CoS requires a $165\\%$ increase in training resources and a $400\\%$ increase in inference computation for a mere $1.7\\%$ performance improvement, while LLaVA-Next faces similar inefficiencies. This challenge emphasizes the need to reduce visual token computation. Approaches like token compression or pruning [15, 22, 39, 60, 116, 119], while efficient, risk losing critical visual information, revealing a trade-off between efficiency and performance. This work aims to optimize the model's computational efficiency from an architectural perspective, proposing a scalable solution that balances efficiency with visual fidelity.", + "bbox": [ + 88, + 228, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We systematically examine the parameter count and FLOPs distribution across different components of the LLM. As shown in Table 1, using LLaMA-3 [29] as an example, the parameter and FLOPS requirements for the FFN are significantly higher than the attention module. Furthermore, as illustrated in Figure 2, when the number of visual tokens is large, the majority of computations are concentrated on FFN of these tokens. In particular, we computed the difference in the magnitudes of the features before and after the FFN transformation for different tokens. As illustrated in the figure 3, we observe that the updates introduced by the FFN to the visual tokens are significantly smaller than those that affect the text tokens. This is similar to the observation in [123]. This observation suggests a potential approach: could we bypass FFN computations for a substantial portion of visual tokens? Furthermore, in the autoregressive training framework of MLLMs, only the final visual token utilizes its last hidden state for next-token prediction, while the other visual tokens primarily serve to convey causal information. During model inference, inspired by [7], we observe that most visual information is progressively integrated into the final token through the early causal layers. This raises the possibility of reducing reliance on the full visual scaling KV cache throughout inference, potentially accelerating the inference process.", + "bbox": [ + 91, + 474, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose Skip-Vision, a novel and flexible architecture designed for more efficient scaling (see Figure 1 for a comparison). Visual tokens de", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8be379eb26551a301c26d0ca932d87f74a22843ebb31fc8e68dbdd11bc04f8eb.jpg", + "image_caption": [ + "Figure 2. Flops ratio of visual tokens in FFN. When visual tokens greatly outnumber text tokens, computation is predominantly consumed by the FFN of visual tokens. Reducing their passage through the FFN can thus markedly decrease computational costs." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 906, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rived from smaller window/scale sizes constitute approximately $80\\%$ of the total visual tokens and account for most of the computational overhead. To mitigate this, as shown in Figure 4, we allow these tokens to skip the Feed-Forward Network (FFN) computation, significantly reducing overall computational costs. To address redundancy within skipped vision tokens, we can apply a token merging strategy for further simplification. During inference, skipped tokens are removed from the KV cache after the pre-fill phase, accelerating subsequent computation. Empirical results demonstrate that Skip-Vision reduces training time by up to $35\\%$ , FLOPs of inference by $75\\%$ and inference latency by $45\\%$ , while maintaining or even improving performance. These efficiency gains provide a scalable solution for large-scale multimodal learning, enabling better data utilization and enhanced model scalability.", + "bbox": [ + 511, + 381, + 906, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, Skip-Vision is designed to be seamlessly integrated into the standard SFT pipeline of MLLMs without introducing additional retraining or decoupled modules. It directly modifies the transformer's computation flow, offering a practical and theoretically grounded acceleration solution for MLLM training and inference jointly. Our contributions are as follows:", + "bbox": [ + 511, + 625, + 905, + 731 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel and efficient framework, using token merge and skip FFN strategy during training to reduce redundant computation for visual tokens.", + "- In inference, our framework employs a skip KV-cache mechanism that removes skip-FFN visual tokens from the KV-cache, enhancing efficiency.", + "- We present a theoretical analysis of the performance error in Skip-Vision and establish an error bound, which will be empirically validated through experiments.", + "- Experiments show our model's superior efficiency, effective data scaling, and performance on par with state-of" + ], + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90b52cde8ad17b80416bc12f56fd3c53c4709de6a0bbfeb3e77d9f6d0defaf22.jpg", + "image_caption": [ + "(a) FFN updates for CoS" + ], + "image_footnote": [], + "bbox": [ + 132, + 93, + 370, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3043d0a625953e2bf208b09527d18729cad7c27bec26f8071729072df4e4959f.jpg", + "image_caption": [ + "(b) FFN updates for LLaVA-HD", + "Figure 3. FFN imapct for different MLLMs. We evaluate FFN impact by computing the feature modulus ratio before $(\\| h_{\\mathrm{attn}}\\|_2)$ and after FFN $(\\| \\mathrm{FFN}(h_{\\mathrm{attn}})\\|_2)$ . The FFN introduces significantly smaller updates to visual tokens compared to text tokens." + ], + "image_footnote": [], + "bbox": [ + 377, + 90, + 620, + 227 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a1f9d01c36c74b7a0bef88cd90a6a14d09d9558f6e27fcb01b5e6b2402bd5f80.jpg", + "image_caption": [ + "(c) FFN updates for LLaV" + ], + "image_footnote": [], + "bbox": [ + 625, + 90, + 870, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the-art models of similar scale.", + "bbox": [ + 104, + 304, + 310, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 351, + 227, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Vison-Language models", + "text_level": 1, + "bbox": [ + 89, + 382, + 310, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rapid progress in large language models (LLMs) [25, 29, 114] has laid a solid foundation for the emergence of large-scale vision-language models (VLMs) [3, 6, 19, 58]. Flamingo [3] was a pioneering effort, integrating pretrained image encoders with LLMs to handle multimodal data, thereby allowing comprehensive understanding and reasoning across modalities. Following this, the major models developed by leading corporations, such as GPT-4v [2] and Gemini-1.5-Pro [88], have led the progress in this domain. Using proprietary data sets and undisclosed training methodologies, these models have elevated multimodal intelligence to unprecedented levels.", + "bbox": [ + 89, + 411, + 483, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the same time, the open-source community has worked diligently to stay abreast of these advancements. The LLava series models [56, 67, 69, 112] exemplify a key strategy in this domain by mapping visual features directly into the input embedding space of the language model, integrating them as input tokens. However, this approach generates a substantial volume of visual tokens, frequently exceeding the number of text tokens, creating efficiency challenges. InternVL-1.5 [24] adopts dynamic high-resolution techniques, segmenting large images into smaller fragments for processing and scaling them accordingly. MiniCPMV [117] utilizes a perceiver resampler structure with a single layer of cross-attention to compress visual tokens. Models such as Vary [107], SPHINX [64], Cambrian-1 [100], and BRAVE [47] leverage multi-visual encoder architectures to strengthen their visual processing capabilities. Furthermore, a crucial factor contributing to the effectiveness of modern VLMs is instruction fine-tuning [28, 67, 134], enabling these models to function as conversational agents and interact with users through natural, human-like dialogue.", + "bbox": [ + 89, + 599, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Efficient multimodal large language models", + "text_level": 1, + "bbox": [ + 511, + 304, + 882, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the efficiency of vision-language models, one of the primary and widely employed approaches in contemporary research involves reducing the size of the backbone models [27, 124, 133?]. Simultaneously, given that visual tokens contribute significantly to computational demands, research focused on compressing or refining these tokens has garnered increasing attention. Such methods are often coupled with the design of vision-language projectors. For instance, techniques like the Perceiver Resampler [3, 6, 41] and Q-Former [58] utilize transformer-based mechanisms to consolidate visual tokens into more compact query sets. In this way, the transformer outputs corresponding to the positions of the learnable latent queries serve as the aggregated representation of the visual features [42]. Honeybee [14] propose two visual projectors, namely C-Abstractor and D-Abstractor. LLaVA-PruMerge [91] and MADTP [12] introduce adaptive methods for reducing visual tokens. In addition, while the large language model (LLM) field has employed various techniques for token reduction to accelerate inference and compress KV cache [35, 131], research in this area remains limited within the MLLM domain.", + "bbox": [ + 511, + 325, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 657, + 651, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Decoder-only LLM", + "text_level": 1, + "bbox": [ + 511, + 683, + 697, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each Transformer layer in a decoder-only LLM primarily consists of a self-attention layer and a feed-forward neural network. To facilitate autoregressive inference, the use of a KV-cache has been developed to accelerate the decoding process.", + "bbox": [ + 511, + 705, + 905, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Attention Mechanism: The self-attention mechanism enables each token in a sequence to selectively attend to other tokens, capturing contextual dependencies. Each token generates a query $(Q)$ , key $(K)$ , and value $(V)$ vector. Attention scores are calculated by the scaled dot-product:", + "bbox": [ + 511, + 782, + 905, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 869, + 903, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21386", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/df2c8b74ad126f2c273114dbae5e81ef25bd7e6e89b6c1b8ffa8b57f2ec80f5a.jpg", + "image_caption": [ + "Figure 4. The framework of Skip-Vision. a) While visual scaling enriches visual information, it also increases computational overhead. Skip-Vision uses a skip-FFN strategy during training to reduce redundant computation for visual tokens. The numerous skipped tokens will be limited to the attention layer and bypass FFN. b) At the beginning of inference, Skip-Vision will remove skip-FFN visual tokens from the initial KV-cache, enhancing efficiency. c) During inference, skip attention leverages the skip KV-cache to accelerate generation." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{k}$ is the dimension of $K$ . This formulation allows each token to incorporate information from others based on relevance, making self-attention critical in capturing relationships across tokens.", + "bbox": [ + 88, + 470, + 482, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feed-Forward Neural Network: The feed-forward neural network enhances token representations following the self-attention layer. The operation can be expressed as:", + "bbox": [ + 89, + 531, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F F N} (x) = \\text {A c t i v a t i o n} \\left(x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 587, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is the input, $W_{1}$ and $W_{2}$ are weight matrices, and $b_{1}$ and $b_{2}$ are biases. This architecture allows the model to capture complex patterns and improve the expressiveness of token representations.", + "bbox": [ + 89, + 614, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KV-cache: The key-value cache (KV-cache) stores key $(K)$ and value $(V)$ representations from previous self-attention computations, enabling efficient reuse during inference. This approach avoids redundant calculations for earlier tokens, improving inference speed.", + "bbox": [ + 89, + 675, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention score is computed using Attention $(Q, K_{\\text{cache}}^t, V_{\\text{cache}}^t)$ , where $Q$ is the query for the current token. At each time step $t$ , the cache is updated as:", + "bbox": [ + 89, + 751, + 483, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\text {c a c h e}} ^ {(t)} = \\left[ K _ {\\text {c a c h e}} ^ {(t - 1)}; K _ {t} \\right], \\quad V _ {\\text {c a c h e}} ^ {(t)} = \\left[ V _ {\\text {c a c h e}} ^ {(t - 1)}; V _ {t} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 820, + 482, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K_{t}$ and $V_{t}$ are the current token's key and value. This mechanism enhances computational efficiency and accelerates token generation.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 468, + 604, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Skipped token selection and token merge", + "text_level": 1, + "bbox": [ + 511, + 494, + 862, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before applying the skip FFN strategy, we first identify the skipped and retained visual tokens. The retained tokens should preserve most essential visual information, while the skipped tokens serve as complementary details, ensuring efficiency without significant loss of information.", + "bbox": [ + 511, + 518, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fortunately, in MLLM architectures, we can naturally distinguish retained and skipped tokens based on global and local context tokens. Global context tokens are retained, while local context tokens are skipped. In architectures like LLaVA-HD, global tokens originate from the resized full image, whereas local tokens come from smaller image patches. In CoS, global tokens correspond to larger window sizes, while local tokens are derived from smaller windows. Since LLaVA lacks a clear distinction between global and local tokens, inspired by [115], we select the top- $n$ tokens most similar to the CLS token as retained tokens, categorizing the rest as skipped tokens.", + "bbox": [ + 511, + 597, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further reduce the computational burden, it is necessary to refine the visual tokens to eliminate excessive redundancy. As shown in Figure 5, we conducted a similarity analysis of the visual tokens in CoS at different levels, calculating their similarity density. We found that the smaller the granularity of the visual tokens, the higher their average similarity, while their token count is significantly greater than that of other granularities. Consequently, inspired by", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ToMe [9], we employ a token merge method to reduce redundancy in skipped tokens, as detailed below:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Compute cosine similarities: calculate the cosine similarity between each pair of tokens in the set, obtaining an average similarity score for each token.", + "- Select top $k$ distinctive Tokens: sort tokens by average similarity in ascending order and retain the top $k$ tokens with the lowest average similarity, representing the most distinctive information.", + "- Merge remaining tokens: for the remaining $n - k$ tokens, merge each one with the most similar token among the top $k$ , thus consolidating information efficiently." + ], + "bbox": [ + 89, + 121, + 482, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Training speed-up: skip FFN", + "text_level": 1, + "bbox": [ + 89, + 279, + 349, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Figure 4a, we propose the FFN skip strategy: The retained visual tokens, which are few in number, proceed conventionally through all decoder layers of the LLM, while the numerous skipped visual tokens will be limited to the self-attention layers of each transformer block. For a given layer $l$ , the output of retained tokens $h_{\\mathrm{retained}}^{(l)}$ is the sum of the self-attention output $h_{\\mathrm{attn}}^{(l)}$ and the FFN-transformed features $\\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)})$ :", + "bbox": [ + 89, + 301, + 483, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\text {r e t a i n e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)} + \\operatorname {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 434, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of the skipped tokens is:", + "bbox": [ + 89, + 460, + 330, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\text {s k i p p e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 481, + 480, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, bypassing the FFN layers presents a nontrivial challenge. In the autoregressive training process, the final token of the vision sequence is used to predict the initial token of the text sequence, necessitating a specialized design for this terminal vision token. To facilitate this transition effectively, it must pass through the FFN layers, which are enriched with stored linguistic information [32, 83].", + "bbox": [ + 89, + 508, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, as shown in Figure 4, we introduce adaptive summary tokens to efficiently consolidate visual information. This token is generated by the Adaptive Summary Layer before the token sequence is input into the LLM. The layer comprises a simple linear transformation, and its computation is defined as follows:", + "bbox": [ + 89, + 628, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {s} = \\left(\\operatorname {s o f t m a x} \\left(X \\cdot W ^ {T}\\right)\\right) ^ {T} \\cdot X, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 724, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X$ denotes the retained or skipped visual tokens, $x^{s} \\in \\mathbb{R}^{1 \\times C}$ is the summary token, and $W \\in \\mathbb{R}^{1 \\times C}$ is the weight matrix. Leveraging this layer, we concatenate a summary token on both ends of the skipped visual tokens before their input to the LLM. The first summary token aggregates essential information from retained tokens, mitigating potential loss from extended token sequences. The second summary token integrates prior sequence information and plays a key role in predicting the next text token.", + "bbox": [ + 89, + 750, + 483, + 885 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0c96dbe6580415cba281479de909b3324509d8ed8de123e3a955dd1de121f6a7.jpg", + "image_caption": [ + "Figure 5. Similarity density statistics. We conducted a similarity density analysis in CoS, examining the visual tokens derived from various window scale. The results indicate that as the window scale decreases, the similarity between visual tokens becomes more pronounced, reflecting higher token redundancy." + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 885, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Inference speed-up: skip KV-cache", + "text_level": 1, + "bbox": [ + 511, + 358, + 818, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training, our Skip-Vision framework incorporates skip KV-cache method to further improve efficiency:", + "bbox": [ + 511, + 380, + 906, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK _ {\\text {s k i p - c a c h e}} ^ {(0)} = K _ {\\text {c a c h e}} ^ {(0)} \\backslash K _ {\\text {s k i p}}, \\quad V _ {\\text {s k i p - c a c h e}} ^ {(0)} = V _ {\\text {c a c h e}} ^ {(0)} \\backslash V _ {\\text {s k i p}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 422, + 903, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equation 7 indicates that visual tokens skipped by FFN are removed from the KV cache after the pre-filling stage. As depicted in Figures 4 b) and c), given the autoregressive nature of LLMs, earlier token information is progressively transferred to later tokens. Skip FFN further enhances this integration by suppressing the attention weights of skipped tokens, forcing the model to focus on key tokens and improving information aggregation efficiency. Empirical results show that, due to the inclusion of summary tokens, this omission does not compromise model performance, underscoring a distinctive advantage of our framework for more efficient inference.", + "bbox": [ + 511, + 458, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Theoretical analysis of Skip-Vision", + "text_level": 1, + "bbox": [ + 511, + 654, + 828, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the core of this analysis lies the layer error incurred when bypassing the FFN layer. For a given layer $l$ , the original output $h_{\\mathrm{original}}^{(l)}$ is: $h_{\\mathrm{original}}^{(l)} = h_{\\mathrm{attn}}^{(l)} + \\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)})$ . When skipping the FFN, the output becomes: $h_{\\mathrm{skip}}^{(l)} = h_{\\mathrm{attn}}^{(l)}$ . The per-layer skipping error is:", + "bbox": [ + 511, + 680, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {(l)} = \\left\\| h _ {\\text {o r i g i n a l}} ^ {(l)} - h _ {\\text {s k i p}} ^ {(l)} \\right\\| _ {2} = \\left\\| \\mathrm {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right) \\right\\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 775, + 903, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For redundant tokens, such as those in homogeneous image regions, this error is negligible ( $\\epsilon^{(l)} \\approx 0$ ) due to minimal feature transformations by the FFN.", + "bbox": [ + 511, + 809, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, errors propagate through subsequent layers, amplified by the recursive nature of transformer architectures. Leveraging Lipschitz continuity assumptions for self", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21388", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "attention and FFN operations $(L_{attn}^{(l + 1)}$ and $L_{FFN}^{(l + 1)}$ ), the cumulative error at layer $l + 1$ is bounded by:", + "bbox": [ + 89, + 88, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {(l + 1)} \\leq \\left(L _ {\\mathrm {a t t n}} ^ {(l + 1)} + L _ {\\mathrm {F F N}} ^ {(l + 1)}\\right) \\cdot \\epsilon^ {(l)} + \\epsilon_ {\\mathrm {s k i p}} ^ {(l + 1)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 131, + 482, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $(\\epsilon_{\\mathrm{skip}}^{(l + 1)})$ represents new errors from skipping deeper layers $(l + 1)$ .", + "bbox": [ + 89, + 162, + 482, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Over $L$ layers, the total error telescopes to:", + "bbox": [ + 109, + 198, + 393, + 213 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\text {t o t a l}} \\leq \\sum_ {l = 1} ^ {L} \\epsilon_ {\\text {s k i p}} ^ {(l)} \\cdot \\prod_ {i = 1} ^ {L - l} \\left(L _ {\\text {a t t n}} ^ {(i + 1)} + L _ {\\text {F F N}} ^ {(i + 1)}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 223, + 482, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 5.1 Bounded Lipschitz. If the spectral norms of $W_{1}, W_{2}, W_{K}, W_{Q}$ , and $W_{V}$ are bounded by 1, then:", + "bbox": [ + 89, + 275, + 485, + 304 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $L(attn) \\leq 1 / \\sqrt{d_k}$", + "$L(FFN)\\leq 1$" + ], + "bbox": [ + 89, + 305, + 223, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Follow Theorem 5.1, we assume Lipschitz constants $L_{\\mathrm{attn}} + L_{\\mathrm{FFN}} \\leq \\gamma$ and skipping errors $\\epsilon_{\\mathrm{skip}}^{(l)} \\leq \\epsilon$ , the total error is scaled to:", + "bbox": [ + 89, + 345, + 483, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\text {t o t a l}} \\leq \\epsilon \\cdot \\frac {\\gamma^ {L} - 1}{\\gamma - 1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 391, + 482, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Equation 11 establishes that the skip error is bounded when $\\gamma < 1$ , provided the model is trained with modern regularization techniques. This ensures that the MLLM remains less sensitive to the effects of skipping.", + "bbox": [ + 89, + 426, + 483, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This error also impacts the KL divergence between the original and skipped outputs, bounded by:", + "bbox": [ + 89, + 488, + 483, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {K L}} \\left(p _ {\\text {s k i p}} \\| p _ {\\text {o r i g i n a l}}\\right) \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\epsilon_ {\\text {t o t a l}} ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 527, + 482, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma^2$ is the variance of the logits. Further integrating feature similarity errors $(\\epsilon_{\\mathrm{sim}} = O(\\sqrt{1 - \\theta}))$ from low-attention tokens, the final bound becomes:", + "bbox": [ + 89, + 566, + 483, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {K L}} \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\left(\\epsilon_ {\\text {t o t a l}} + \\epsilon_ {\\text {s i m}}\\right) ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 621, + 482, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Practically, this analysis motivates a layer-wise skipping strategy, alongside token selection and merge based on feature similarity $(\\theta)$ . More analysis and the proof of Theorem 5.1 are provided in the Appendix.", + "bbox": [ + 89, + 659, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 733, + 220, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 757, + 223, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively evaluate our model, we conducted experiments on eight different benchmarks. These include MME [30], which assesses the perceptual and cognitive capabilities of multimodal language models; TextVQA [95], focused on fine-grained visual question answering; MM-Bench [72] and MMStar [21] for general diagnostic capabilities; MMMU [125], designed to test STEM-related reasoning; MathVista [76], which evaluates mathematical problem", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "solving skills; OCRBench [70], for optical character recognition tasks; and MMVet [121], used for subjective assessment. Taken together, these benchmarks offer a comprehensive framework for assessing the efficacy of the model in various multimodal challenges.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Efficiency evaluation", + "text_level": 1, + "bbox": [ + 511, + 181, + 709, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We systematically evaluate both training time (hours), inference FLOP computation and inference latency (ms) on single A100 GPU. For training efficiency, we report actual GPU hours measured on consistent hardware configurations. For inference FLOPs, we adopt the methodology from FastV [22], specifically tracking the computational demands of processing visual tokens. In the LLama architecture, we calculate FLOPs for the causal attention and feed forward network modules as $4N C^{2} + 2N^{2}C + 3N C M$ where $N$ is the token count, $C$ denotes the hidden state dimension, and $M$ is the intermediate FFN dimension. SkipVision framework adjusts token counts dynamically between the attention and FFN modules, allowing us to refine FLOP calculations as $4N_{1}C^{2} + 2N_{1}^{2}C + 3N_{2}CM$ , where $N_{1}$ and $N_{2}$ represent the visual token counts processed in the causal attention and FFN layers, respectively.", + "bbox": [ + 511, + 204, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.3. Data setup", + "text_level": 1, + "bbox": [ + 511, + 460, + 633, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During pretraining, we adopt the experimental settings of LLaVA and CoS, utilizing LLaVA-558K [69] and 65 million image-text pairs for respective comparisons. For the Supervised Fine-Tuning phase, we use LLaVA-665K [69] for all comparative experiments. By leveraging the computational efficiency of Skip-Vision framework, the resources saved can be redirected toward scaling the dataset to further enhance model performance. To this end, we extend our dataset to SV-1M, comprising 1 million samples, demonstrating the substantial performance improvements achievable through data expansion. To further assess our approach against state-of-the-art models of similar scale, we extend the dataset to SV-9M, underscoring the scalability and potential of our method. Detailed dataset statistics are provided in the Appendix.", + "bbox": [ + 511, + 484, + 908, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.4. Training setup", + "text_level": 1, + "bbox": [ + 511, + 726, + 661, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For LLaVA and LLaVA-HD setting, we use input resolutions of 336 and a dynamic resolution ranging from 336 to 672, respectively, with CLIP ViT-L/336px [86] as the vision encoder and LLaMA3 8B Instruct [29] as the backbone (additional comparisons under LLaVA-1.5-7B training setup are in the Appendix 10.1). For CoS setting, pretraining is conducted at 224 resolution, training only the multi-scale visual resampler with CLIP ViT-L/224px [86]. The model processes 80 visual tokens, including 16 global and 64 local tokens, and scales to 336 and 1296 tokens during SFT.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21389", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/22513fff52ba8b8c07051b6ab66729d33d0c1c4ba4cd4d8bf2e53e60173c024d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MMETextvqaMMBMMVetMMMUMMathVOCRBMMStarAvgHoursFLOPs Latency
LLaVA setting:
LLaVA [69]153558.672.232.839.621.332.736.542.010.2100%125
LLaVA-FastV [22]145356.17132.437.320.628.836.340.49.630%103
MQT-LLaVA [39]148753.66928.237.816.828.136.538.67.544%106
LLaVA-Tokenpacker [60]153856.272.332.139.019.329.835.040.58.824%97
SV-LLaVA151956.572.532.438.821.230.736.541.37.925%90
(Nr=100,Ns=156)-1%-3.6%+0.4%-1.2%-2%-0.5%-3.6%-6.1%-1.7%-22.5%-75%-28%
SV-LLaVA153057.372.735.639.221.432.040.042.69.060%103
(Nr=256,Ns=320)-0.3%-2.2%+0.7%+8.5%-1%+0.5%-2.1%+9.6%+1.4%-11.8%-40%-17.6%
LLaVA-HD setting:
LLaVA-HD [67]15336471.835.339.32037.440.544.019100%450
SV-LLaVA-HD153461.273.135.939.721.234.139.343.512.250%250
(Nr=576,Ns=576)+0%-4.3%+1.8%+1.7%+1.5%-6%-8.8%-2.9%-1.1%-35.8%-50%-44.4%
CoS setting:
CoS [40]158564.477.139.439.221.539.241.246.015.3100%160
SV-CoS156363.776.541.740.321.23741.946.09.825%90
(Nr=272,Ns=256)-1.4%-1.1%-0.8%+5.8%+2.8%-1.4%-5.6%+1.7%-0%-40%-75%-43.8%
SV-CoS (SV-1M)156967.177.143.340.722.050.744.449.316.325%90
", + "bbox": [ + 94, + 87, + 903, + 381 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05e78d0a2944b11792ca413f593403bf0bc551a4a8e80bbbb6ae7b26692c76cc.jpg", + "table_caption": [ + "Table 2. Performance and efficiency evaluation. We evaluate Skip-Vision on LLaVA, LLaVA-HD, and CoS and compare it with state-of-the-art efficiency optimization models under the LLaVA setting. $N_r$ and $N_s$ denote the number of retained and skipped tokens, respectively." + ], + "table_footnote": [], + "table_body": "
LLMMMETextvqaMMBMMVetMMMUMathVOCRBMMStar
Open-source models in 8B tier:
mPLUG-Owl3 [118]Qwen2-69.077.640.1-65.0-40.1
Cambrian [100]LLaMA3154771.775.9-42.749.062.4-
LLaVA [69]LLaMA3153558.672.232.839.621.332.736.5
Mini-Gemini-HD [61]LLaMA3160670.272.7-37.337.047.7-
LLaVA-Next [56]LLaMa3160464.672.1-41.736.349.0-
Ovis [77]LLaMA3--77.4-44.740.8-49.5
MiniCPM-V2.5 [117]LLaMA3-76.677.6-45.854.372.5-
SV-CoS (SV-9M)LLaMA3155069.178.451.74361.369.452.3
", + "bbox": [ + 106, + 422, + 890, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. MLLM evaluation. We present the performance of SV-CoS on SV-9M, comparing it against the current SOTA models of a similar scale. The highest scores are highlighted in bold, while the second-highest scores are indicated with underlines.", + "bbox": [ + 89, + 588, + 906, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In SFT, the resolution increases to 448, and all model parameters are unfrozen for full optimization.", + "bbox": [ + 89, + 628, + 482, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In all experiments, we used a learning rate of $1e^{-3}$ for pre-training, which is reduced to $1e^{-5}$ during SFT. Training follows a cosine decay schedule after a $3\\%$ warm-up phase. For skipped token selection, unless otherwise specified, the LLaVA setting includes 156 skipped tokens and 100 retained tokens, the LLaVA-HD setting has 576 skipped and 576 retained tokens, and the CoS setting consists of 256 skipped and 272 retained tokens. All skipped tokens undergo a token merging process to enhance efficiency.", + "bbox": [ + 89, + 659, + 482, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Main experimental results", + "text_level": 1, + "bbox": [ + 89, + 803, + 328, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance and efficiency. We first evaluate the effectiveness of Skip-Vision across three different frameworks: CoS, LLaVA, and LLaVA-HD, demonstrating its ability to maximize efficiency while minimizing performance degradation. As shown in Figure 1 and Table 6, Skip-Vision re", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duces training time by $22.5\\%$ , inference FLOPs by $75\\%$ , and inference latency by $28\\%$ for LLaVA. For LLaVA-HD, it achieves a $35.8\\%$ reduction in training time, $50\\%$ in inference FLOPs, and $45\\%$ in inference latency. In CoS, it reduces training time by $35.8\\%$ , inference FLOPs by $75\\%$ , and inference latency by $44\\%$ .", + "bbox": [ + 511, + 628, + 906, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling. We evaluated the scalability of Skip-Vision, leveraging CoS's visual scaling to increase visual tokens from 336 to 1296 while maintaining efficiency. Expanding the fine-tuning dataset from LLaVA-665k to SV-1M further improved performance by $8\\%$ over $C o S_{1296}$ with similar resources. To benchmark against state-of-the-art models, we scaled to SV-9M, with training completed in 72 hours on 16 NVIDIA A100 GPUs (Table 3).", + "bbox": [ + 511, + 720, + 908, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.6. Ablation study and analysis", + "text_level": 1, + "bbox": [ + 511, + 848, + 764, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of each component within the Skip-Vision framework, we conducted ablation and com", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21390", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/86470e3723517fd4c5b4f14c13cb728d4bdcefa1ba6b2fc4928f1eea925e1cc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MMETextvqaMMBMMVetMMMUMathVOCRBMMStarAvgLatency
SV-LLaVA (w/o SK)151956.772.531.438.821.230.936.541.1103
SV-LLaVA151956.572.532.438.821.230.736.541.390
+0%-0.4%+0%+3.2%+0%+0%-0.6%+0%+0.5%-9.7%
SV-LLaVA-HD (w/o SK)153361.673.035.839.921.234.539.743.7310
SV-LLaVA-HD153461.273.135.939.721.234.139.343.5250
+0%-0.7%+0.1%+0.3%-0.5%+0%-1.2%-1%-0.5%-19.4%
SV-LLaVA-CoS (w/o SK)156263.976.540.240.221.237.241.945.9126
SV-LLaVA-CoS156363.776.541.740.321.23741.946.090
+0%-0.3%+0%+3.7%+0.2%+0%-0.5%+0%+0.2%-28.6%
", + "bbox": [ + 91, + 87, + 906, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20bd027391091a7b44c86e87db03dc371646ae53d0deea8bdf5c02fd6b73d14e.jpg", + "image_caption": [ + "Table 4. Ablation on skip KV-cache. We evaluate SK (using skip KV-cache during inference) of Skip-Vision on different settings.", + "(a) Ablation on LLaVA" + ], + "image_footnote": [], + "bbox": [ + 133, + 270, + 359, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/892870d1a207c1e5fed03929e6ea26e4f6a68289766f5d21b24157546f59af90.jpg", + "image_caption": [ + "(b) Ablation on CoS", + "" + ], + "image_footnote": [], + "bbox": [ + 359, + 270, + 614, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4ea6c94363d42f2866abbfcffe002a70241e63763992acd3b944562eeeb51eb.jpg", + "image_caption": [ + "(c) Ablation on LLaVA-HD" + ], + "image_footnote": [], + "bbox": [ + 627, + 270, + 862, + 396 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9baf878fc339ad47abe3903f68fb1225de82816dfc51f51f9eaa5315a5efca38.jpg", + "image_caption": [ + "Figure 6. Ablation study. We performed an ablation study on each component of the Skip-Vision. SF (skip FFN), TM (token merge)." + ], + "image_footnote": [], + "bbox": [ + 96, + 446, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28caf95df1a05c1466a032b57792f37bb6d1c1d30afce73eb7aad7557cb9b2e2.jpg", + "image_caption": [ + "Figure 7. Visualization of attention map in MMVet.", + "Figure 8. Visualization of attention map in TextVQA." + ], + "image_footnote": [], + "bbox": [ + 96, + 601, + 483, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "parative experiments on token merge, skip FFN, and skip KV-cache.", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 6, both the skip FFN strategy and token merging enhance computational efficiency across different model settings while maintaining performance. Notably, skip FFN performs better in the CoS setting than in LLaVA and LLaVA-HD. We attribute this to CoS allowing the vision backbone to be fine-tuned during the SFT stage, enabling key information to be more effectively captured by", + "bbox": [ + 88, + 794, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "summary tokens, thereby reducing the errors introduced by skip FFN. As shown in Table 4, skip KV-cache slightly affects fine-grained tasks such as TextVQA and OCRBench but improves performance in long-form response tasks like MMVet.", + "bbox": [ + 511, + 446, + 905, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further analyze this, we visualized attention maps across layers for visual tokens in SV-CoS during the MMVet and TextVQA tasks. In MMVet (Figure 7), the attention of generated tokens primarily focuses on retained and summary tokens, with skip KV-cache filtering out unnecessary noise. In TextVQA (Figure 8), which requires fine-grained understanding, attention is spread across both retained and skipped tokens, leading to some performance loss with skip KV-cache. However, causal attention effectively aggregates information from skipped tokens into the final summary tokens, allowing us to retain only the summary tokens and retained visual tokens in the KV-cache.", + "bbox": [ + 511, + 523, + 906, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Skip-Vision, an efficient multimodal large language model architecture that boosts computational efficiency through a skip-FFN strategy, reducing redundant computations over visual tokens during training. During inference, Skip-Vision uses a skip KV-cache to accelerate processing by omitting non-essential visual tokens from the KV cache. Extensive experiments show our model outperforms current methods in efficiency and scales effectively with data, delivering competitive performance among leading models of similar scale.", + "bbox": [ + 509, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by NSFC (62201342), and Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102). This work was supported by Ant Group Research Intern Program.", + "bbox": [ + 89, + 113, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 5", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 3", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 23716-23736, 2022. 1, 3", + "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 5", + "[5] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 5", + "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3", + "[7] Samyadeep Basu, Martin Grayson, Cecily Morrison, Besmira Nushi, Soheil Feizi, and Daniela Massiceti. Understanding information storage and transfer in multi-modal large language models. arXiv preprint arXiv:2406.04236, 2024. 2", + "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5", + "[9] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022.5", + "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[11] Shihao Cai, Keqin Bao, Hangyu Guo, Jizhi Zhang, Jun Song, and Bo Zheng. Geogpt4v: Towards geometric multi-" + ], + "bbox": [ + 99, + 223, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "modal large language models with geometric image generation. arXiv preprint arXiv:2406.11503, 2024. 5", + "[12] Jianjian Cao, Peng Ye, Shengze Li, Chong Yu, Yansong Tang, Jiwen Lu, and Tao Chen. Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15710-15719, 2024. 3", + "[13] Jimmy Carter. Textocr-gpt4v. https://huggingface.co/datasets/jimmycarter/textocr-gpt4v,2024.5", + "[14] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 3", + "[15] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 2", + "[16] Chungkwong Chan. Stroke extraction for offline handwritten mathematical expression recognition. IEEE Access, 8: 61565-61575, 2020. 5", + "[17] Shuaichen Chang, David Palzer, Jialin Li, Eric Fosler-Lussier, and Ningchuan Xiao. Mapqa: A dataset for question answering on choropleth maps. arXiv preprint arXiv:2211.08545, 2022. 5", + "[18] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024. 5", + "[19] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3", + "[20] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5", + "[21] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.6", + "[22] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 6, 7", + "[23] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5", + "[24] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21392", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 3", + "[25] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/, 2023.1.3", + "[26] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaiqiao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1571–1576. IEEE, 2019. 5", + "[27] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 3", + "[28] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3", + "[29] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6", + "[30] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2024. 6", + "[31] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023.5", + "[32] Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767, 2023.5", + "[33] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2315-2324, 2016. 5", + "[34] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939–948, 2019. 5", + "[35] Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Zero-shot extreme" + ], + "bbox": [ + 99, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "length generalization for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3991-4008, 2024. 3", + "[36] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In 2018 24th international conference on pattern recognition (ICPR), pages 7-12. IEEE, 2018. 5", + "[37] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.5", + "[38] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024.5", + "[39] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2, 7", + "[40] Ziyuan Huang, Kaixiang Ji, Biao Gong, Zhiwu Qing, Qinglong Zhang, Kecheng Zheng, Jian Wang, Jingdong Chen, and Ming Yang. Accelerating pre-training of multimodal llms via chain-of-sight. arXiv preprint arXiv:2407.15819, 2024. 1, 7", + "[41] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3", + "[42] Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024. 1, 3", + "[43] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 5", + "[44] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 5", + "[45] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figurea: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 5", + "[46] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 5" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21393", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 3", + "[48] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5", + "[49] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5", + "[50] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern recognition, pages 4999-5007, 2017. 5", + "[51] Sunjun Kweon, Yeonsu Kwon, Seonhee Cho, Yohan Jo, and Edward Choi. Open-wikitable: Dataset for open domain question answering with complex reasoning over table. arXiv preprint arXiv:2305.07288, 2023. 5", + "[52] Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1):1-10, 2018. 5", + "[53] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 5", + "[54] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: mitigating multimodal hallucination through self-feedback guided revision. arXiv preprint arXiv:2311.07362, 2023. 5", + "[55] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 5", + "[56] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild. https://llava-v1.github.io/blog/2024-05-10-llava-last-stronger-llms/, 2024.1,3,7", + "[57] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5", + "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3", + "[59] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 5", + "[60] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 7", + "[61] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 7", + "[62] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 5", + "[63] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 1", + "[64] Dongyang Liu, Renrui Zhang, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, et al. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models. arXiv preprint arXiv:2402.05935, 2024. 1, 3", + "[65] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 5", + "[66] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023.5", + "[67] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 7", + "[68] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.2", + "[69] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7, 5" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21394", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023.6", + "[71] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 5", + "[72] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6", + "[73] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Icdar 2023 competition on hierarchical text detection and recognition. arXiv preprint arXiv:2305.09750, 2023. 5", + "[74] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021.5", + "[75] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5", + "[76] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6", + "[77] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 7", + "[78] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5", + "[79] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5", + "[80] Minesh Mathew, Lluis Gomez, Dimosthenis Karatzas, and CV Jawahar. Asking questions on handwritten document collections. International Journal on Document Analysis and Recognition (IJDAR), 24(3):235-249, 2021. 5", + "[81] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5", + "[82] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa." + ], + "bbox": [ + 99, + 92, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5", + "[83] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022. 5", + "[84] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 5", + "[85] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5", + "[86] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6", + "[87] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia, 2018. Association for Computational Linguistics. 5", + "[88] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 3", + "[89] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 5", + "[90] Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar. Kvqa: Knowledge-aware visual question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8876-8884, 2019. 5", + "[91] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3, 2", + "[92] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 5", + "[93] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024.5", + "[94] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. 2020. 5" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21395", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6, 5", + "[96] Dominik Stammbach and Elliott Ash. Docscan: Unsupervised text classification via learning from neighbors. arXiv preprint arXiv:2105.04024, 2021. 5", + "[97] Hamed Rahimi Sujet AI, Allaa Boutaleb. Sujet-finance-qavision-100k: A large-scale dataset for financial document vqa, 2024. 5", + "[98] S Svetlichnaya. Deepform: Understand structured documents at scale. 2020. 5", + "[99] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 5", + "[100] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3, 7", + "[101] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[102] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016.5", + "[103] Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[104] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1", + "[105] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 5", + "[106] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Conghui He, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation, 2024. 5", + "[107] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jian-Yuan Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. ArXiv, abs/2312.06109, 2023. 3", + "[108] Yuxin Wen, Qingqing Cao, Qichen Fu, Sachin Mehta, and Mahyar Najibi. Efficient vision-language models by" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "summarizing visual tokens into compact registers. arXiv preprint arXiv:2410.14072, 2024. 2", + "[109] C. Wendler. Renderedtext. https://huggingface.co/datasets/wendlerc/RenderedText, 2023.5", + "[110] Jianshu Chen Yunkai Zhang Hong Wang Shiyang Li Xiyou Zhou Wenhu Chen, Hongmin Wang and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 2020. 5", + "[111] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2", + "[112] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 1, 3", + "[113] Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts. arXiv preprint arXiv:2312.15915, 2023. 5", + "[114] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 3", + "[115] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 4, 2", + "[116] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2", + "[117] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 7", + "[118] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 7", + "[119] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 2", + "[120] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthyness. arXiv preprint arXiv:2405.17220, 2024. 5", + "[121] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "21396", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.6", + "[122] Xinyan Velocity Yu, Sewon Min, Luke Zettlemoyer, and Hannaneh Hajishirzi. Crepe: Open-domain question answering with false presuppositions. arXiv preprint arXiv:2211.17257, 2022. 5", + "[123] Qianhao Yuan, Qingyu Zhang, Yanjiang Liu, Jiawei Chen, Yaojie Lu, Hongyu Lin, Jia Zheng, Xianpei Han, and Le Sun. Shortv: Efficient multimodal large language models by freezing visual tokens in ineffective layers. arXiv preprint arXiv:2504.00502, 2025. 2", + "[124] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 3", + "[125] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6", + "[126] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5317-5327, 2019. 5", + "[127] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 2", + "[128] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 5", + "[129] Ying Zhang, Lionel Gueguen, Ilya Zharkov, Peter Zhang, Keith Seifert, and Ben Kadlec. Uber-text: A large-scale dataset for optical character recognition from street-level imagery. In SUNw: Scene Understanding Workshop-CVPR, page 5, 2017. 5", + "[130] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2", + "[131] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3", + "[132] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[133] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 3", + "[134] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[135] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. 5" + ], + "bbox": [ + 516, + 90, + 905, + 273 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "21397", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_model.json b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_model.json new file mode 100644 index 0000000000000000000000000000000000000000..14b4a9a3d7d2e1e1f107f5ef6424b42821f5b29e --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_model.json @@ -0,0 +1,3550 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.9, + 0.178 + ], + "angle": 0, + "content": "Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.21, + 0.728, + 0.229 + ], + "angle": 0, + "content": "Weili Zeng\\(^{1*}\\), Ziyuan Huang\\(^{2}\\), Kaixiang Ji\\(^{2}\\), Yichao Yan\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.229, + 0.886, + 0.249 + ], + "angle": 0, + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.25, + 0.855, + 0.266 + ], + "angle": 0, + "content": "{zw1666,yanyichao}@sjtu.edu.cn, ziyuan.huang@u.nus.edu, kaixiang.jkx@antgroup.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.299, + 0.327, + 0.316 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.486, + 0.697 + ], + "angle": 0, + "content": "Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that SkipVision reduces training time by up to \\(35\\%\\), inference FLOPs by \\(75\\%\\), and latency by \\(45\\%\\), while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency. Our project page can be found here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.485, + 0.862 + ], + "angle": 0, + "content": "Transformer-based large-scale models have significantly advanced progress in Artificial Intelligence (AI) [10, 86, 101]. Their remarkable capabilities have driven the emergence of modern Multimodal Large Language Models (MLLMs) [2, 3, 69, 88], which demonstrate vision-language competencies comparable to human performance. Empirical evidence suggests that scaling laws continue to" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.299, + 0.903, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Figure 1. Performance-efficiency trade-off curve. Each circle denotes a model configuration, where our models utilize the Skip-Vision framework, with CoS [40], LLaVA-HD [67] and LLaVA [69] serving as baselines. Circle sizes reflect the inference FLOPs ratio. Skip-Vision demonstrate superior performance, scaling effectively with increased FLOPs and data and achieving higher inference efficiency when compared to baselines and other efficient MLLM methods. All methods utilize LLaMA3 8B as the foundational large language model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.869 + ], + "angle": 0, + "content": "be effective, primarily across three dimensions: visual scaling [40, 67, 112], data scaling [63, 64], and model scaling [56, 104]. However, these scaling methods present substantial efficiency issues, significantly increasing the training and inference burden. In a typical LLaVA-1.5 framework [69], where the model receives 576 visual tokens and 40 text tokens, inference using Vicuna-13B [25] LLM backbone demands 200 trillion FLOPS [42]. These motivate our pursuit of an MLLM architecture that optimally balances efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recent advances in Multimodal Large Language Models (MLLMs), such as Chain-of-Sight (CoS)[40] and LLaVA-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.444, + 0.888 + ], + "angle": 0, + "content": "*This work was done when Weili Zeng interned at Ant Group." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.24, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.444, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21384" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.468, + 0.155 + ], + "angle": 0, + "content": "
Llama layerComplexityParameters
FFN10.5LNC25.64B(70%)
Attention2LN2C + 1.5NC21.34B(17%)
Embedding-0.5B(6%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.164, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 1. Computational complexity and parameter statistics for Llama3 8B. L denotes the number of blocks, N represents the number of input tokens, and C is the feature dimensionality" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.229, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Next[68], have improved performance by increasing visual tokens. However, during the Supervised Fine-Tuning (SFT) stage, the computational cost rises dramatically, leading to diminishing returns and making the scaling unsustainable. As shown in Figure 1, CoS requires a \\(165\\%\\) increase in training resources and a \\(400\\%\\) increase in inference computation for a mere \\(1.7\\%\\) performance improvement, while LLaVA-Next faces similar inefficiencies. This challenge emphasizes the need to reduce visual token computation. Approaches like token compression or pruning [15, 22, 39, 60, 116, 119], while efficient, risk losing critical visual information, revealing a trade-off between efficiency and performance. This work aims to optimize the model's computational efficiency from an architectural perspective, proposing a scalable solution that balances efficiency with visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.852 + ], + "angle": 0, + "content": "We systematically examine the parameter count and FLOPs distribution across different components of the LLM. As shown in Table 1, using LLaMA-3 [29] as an example, the parameter and FLOPS requirements for the FFN are significantly higher than the attention module. Furthermore, as illustrated in Figure 2, when the number of visual tokens is large, the majority of computations are concentrated on FFN of these tokens. In particular, we computed the difference in the magnitudes of the features before and after the FFN transformation for different tokens. As illustrated in the figure 3, we observe that the updates introduced by the FFN to the visual tokens are significantly smaller than those that affect the text tokens. This is similar to the observation in [123]. This observation suggests a potential approach: could we bypass FFN computations for a substantial portion of visual tokens? Furthermore, in the autoregressive training framework of MLLMs, only the final visual token utilizes its last hidden state for next-token prediction, while the other visual tokens primarily serve to convey causal information. During model inference, inspired by [7], we observe that most visual information is progressively integrated into the final token through the early causal layers. This raises the possibility of reducing reliance on the full visual scaling KV cache throughout inference, potentially accelerating the inference process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Building on these insights, we propose Skip-Vision, a novel and flexible architecture designed for more efficient scaling (see Figure 1 for a comparison). Visual tokens de" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.09, + 0.907, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.305, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 2. Flops ratio of visual tokens in FFN. When visual tokens greatly outnumber text tokens, computation is predominantly consumed by the FFN of visual tokens. Reducing their passage through the FFN can thus markedly decrease computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.625 + ], + "angle": 0, + "content": "rived from smaller window/scale sizes constitute approximately \\(80\\%\\) of the total visual tokens and account for most of the computational overhead. To mitigate this, as shown in Figure 4, we allow these tokens to skip the Feed-Forward Network (FFN) computation, significantly reducing overall computational costs. To address redundancy within skipped vision tokens, we can apply a token merging strategy for further simplification. During inference, skipped tokens are removed from the KV cache after the pre-fill phase, accelerating subsequent computation. Empirical results demonstrate that Skip-Vision reduces training time by up to \\(35\\%\\), FLOPs of inference by \\(75\\%\\) and inference latency by \\(45\\%\\), while maintaining or even improving performance. These efficiency gains provide a scalable solution for large-scale multimodal learning, enabling better data utilization and enhanced model scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.732 + ], + "angle": 0, + "content": "In summary, Skip-Vision is designed to be seamlessly integrated into the standard SFT pipeline of MLLMs without introducing additional retraining or decoupled modules. It directly modifies the transformer's computation flow, offering a practical and theoretically grounded acceleration solution for MLLM training and inference jointly. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.779 + ], + "angle": 0, + "content": "- We introduce a novel and efficient framework, using token merge and skip FFN strategy during training to reduce redundant computation for visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- In inference, our framework employs a skip KV-cache mechanism that removes skip-FFN visual tokens from the KV-cache, enhancing efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- We present a theoretical analysis of the performance error in Skip-Vision and establish an error bound, which will be empirically validated through experiments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Experiments show our model's superior efficiency, effective data scaling, and performance on par with state-of" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.094, + 0.371, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.236, + 0.319, + 0.248 + ], + "angle": 0, + "content": "(a) FFN updates for CoS" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.092, + 0.621, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.236, + 0.585, + 0.248 + ], + "angle": 0, + "content": "(b) FFN updates for LLaVA-HD" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.092, + 0.871, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.236, + 0.82, + 0.248 + ], + "angle": 0, + "content": "(c) FFN updates for LLaV" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.29 + ], + "angle": 0, + "content": "Figure 3. FFN imapct for different MLLMs. We evaluate FFN impact by computing the feature modulus ratio before \\((\\| h_{\\mathrm{attn}}\\|_2)\\) and after FFN \\((\\| \\mathrm{FFN}(h_{\\mathrm{attn}})\\|_2)\\). The FFN introduces significantly smaller updates to visual tokens compared to text tokens." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.305, + 0.312, + 0.319 + ], + "angle": 0, + "content": "the-art models of similar scale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.352, + 0.228, + 0.367 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.312, + 0.399 + ], + "angle": 0, + "content": "2.1. Vison-Language models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Rapid progress in large language models (LLMs) [25, 29, 114] has laid a solid foundation for the emergence of large-scale vision-language models (VLMs) [3, 6, 19, 58]. Flamingo [3] was a pioneering effort, integrating pretrained image encoders with LLMs to handle multimodal data, thereby allowing comprehensive understanding and reasoning across modalities. Following this, the major models developed by leading corporations, such as GPT-4v [2] and Gemini-1.5-Pro [88], have led the progress in this domain. Using proprietary data sets and undisclosed training methodologies, these models have elevated multimodal intelligence to unprecedented levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.901 + ], + "angle": 0, + "content": "At the same time, the open-source community has worked diligently to stay abreast of these advancements. The LLava series models [56, 67, 69, 112] exemplify a key strategy in this domain by mapping visual features directly into the input embedding space of the language model, integrating them as input tokens. However, this approach generates a substantial volume of visual tokens, frequently exceeding the number of text tokens, creating efficiency challenges. InternVL-1.5 [24] adopts dynamic high-resolution techniques, segmenting large images into smaller fragments for processing and scaling them accordingly. MiniCPMV [117] utilizes a perceiver resampler structure with a single layer of cross-attention to compress visual tokens. Models such as Vary [107], SPHINX [64], Cambrian-1 [100], and BRAVE [47] leverage multi-visual encoder architectures to strengthen their visual processing capabilities. Furthermore, a crucial factor contributing to the effectiveness of modern VLMs is instruction fine-tuning [28, 67, 134], enabling these models to function as conversational agents and interact with users through natural, human-like dialogue." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.883, + 0.32 + ], + "angle": 0, + "content": "2.2. Efficient multimodal large language models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.644 + ], + "angle": 0, + "content": "To enhance the efficiency of vision-language models, one of the primary and widely employed approaches in contemporary research involves reducing the size of the backbone models [27, 124, 133?]. Simultaneously, given that visual tokens contribute significantly to computational demands, research focused on compressing or refining these tokens has garnered increasing attention. Such methods are often coupled with the design of vision-language projectors. For instance, techniques like the Perceiver Resampler [3, 6, 41] and Q-Former [58] utilize transformer-based mechanisms to consolidate visual tokens into more compact query sets. In this way, the transformer outputs corresponding to the positions of the learnable latent queries serve as the aggregated representation of the visual features [42]. Honeybee [14] propose two visual projectors, namely C-Abstractor and D-Abstractor. LLaVA-PruMerge [91] and MADTP [12] introduce adaptive methods for reducing visual tokens. In addition, while the large language model (LLM) field has employed various techniques for token reduction to accelerate inference and compress KV cache [35, 131], research in this area remains limited within the MLLM domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.652, + 0.674 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.684, + 0.698, + 0.7 + ], + "angle": 0, + "content": "3.1. Decoder-only LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.782 + ], + "angle": 0, + "content": "Each Transformer layer in a decoder-only LLM primarily consists of a self-attention layer and a feed-forward neural network. To facilitate autoregressive inference, the use of a KV-cache has been developed to accelerate the decoding process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.906, + 0.859 + ], + "angle": 0, + "content": "Self-Attention Mechanism: The self-attention mechanism enables each token in a sequence to selectively attend to other tokens, capturing contextual dependencies. Each token generates a query \\((Q)\\), key \\((K)\\), and value \\((V)\\) vector. Attention scores are calculated by the scaled dot-product:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.87, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21386" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.399, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 4. The framework of Skip-Vision. a) While visual scaling enriches visual information, it also increases computational overhead. Skip-Vision uses a skip-FFN strategy during training to reduce redundant computation for visual tokens. The numerous skipped tokens will be limited to the attention layer and bypass FFN. b) At the beginning of inference, Skip-Vision will remove skip-FFN visual tokens from the initial KV-cache, enhancing efficiency. c) During inference, skip attention leverages the skip KV-cache to accelerate generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.531 + ], + "angle": 0, + "content": "where \\( d_{k} \\) is the dimension of \\( K \\). This formulation allows each token to incorporate information from others based on relevance, making self-attention critical in capturing relationships across tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Feed-Forward Neural Network: The feed-forward neural network enhances token representations following the self-attention layer. The operation can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.588, + 0.483, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F F N} (x) = \\text {A c t i v a t i o n} \\left(x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\( x \\) is the input, \\( W_{1} \\) and \\( W_{2} \\) are weight matrices, and \\( b_{1} \\) and \\( b_{2} \\) are biases. This architecture allows the model to capture complex patterns and improve the expressiveness of token representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.484, + 0.751 + ], + "angle": 0, + "content": "KV-cache: The key-value cache (KV-cache) stores key \\((K)\\) and value \\((V)\\) representations from previous self-attention computations, enabling efficient reuse during inference. This approach avoids redundant calculations for earlier tokens, improving inference speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.811 + ], + "angle": 0, + "content": "The attention score is computed using Attention \\((Q, K_{\\text{cache}}^t, V_{\\text{cache}}^t)\\), where \\(Q\\) is the query for the current token. At each time step \\(t\\), the cache is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.821, + 0.483, + 0.848 + ], + "angle": 0, + "content": "\\[\nK _ {\\text {c a c h e}} ^ {(t)} = \\left[ K _ {\\text {c a c h e}} ^ {(t - 1)}; K _ {t} \\right], \\quad V _ {\\text {c a c h e}} ^ {(t)} = \\left[ V _ {\\text {c a c h e}} ^ {(t - 1)}; V _ {t} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( K_{t} \\) and \\( V_{t} \\) are the current token's key and value. This mechanism enhances computational efficiency and accelerates token generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.469, + 0.605, + 0.485 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.864, + 0.513 + ], + "angle": 0, + "content": "4.1. Skipped token selection and token merge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Before applying the skip FFN strategy, we first identify the skipped and retained visual tokens. The retained tokens should preserve most essential visual information, while the skipped tokens serve as complementary details, ensuring efficiency without significant loss of information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Fortunately, in MLLM architectures, we can naturally distinguish retained and skipped tokens based on global and local context tokens. Global context tokens are retained, while local context tokens are skipped. In architectures like LLaVA-HD, global tokens originate from the resized full image, whereas local tokens come from smaller image patches. In CoS, global tokens correspond to larger window sizes, while local tokens are derived from smaller windows. Since LLaVA lacks a clear distinction between global and local tokens, inspired by [115], we select the top-\\(n\\) tokens most similar to the CLS token as retained tokens, categorizing the rest as skipped tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further reduce the computational burden, it is necessary to refine the visual tokens to eliminate excessive redundancy. As shown in Figure 5, we conducted a similarity analysis of the visual tokens in CoS at different levels, calculating their similarity density. We found that the smaller the granularity of the visual tokens, the higher their average similarity, while their token count is significantly greater than that of other granularities. Consequently, inspired by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "ToMe [9], we employ a token merge method to reduce redundancy in skipped tokens, as detailed below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.167 + ], + "angle": 0, + "content": "- Compute cosine similarities: calculate the cosine similarity between each pair of tokens in the set, obtaining an average similarity score for each token." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.483, + 0.226 + ], + "angle": 0, + "content": "- Select top \\( k \\) distinctive Tokens: sort tokens by average similarity in ascending order and retain the top \\( k \\) tokens with the lowest average similarity, representing the most distinctive information." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.228, + 0.483, + 0.273 + ], + "angle": 0, + "content": "- Merge remaining tokens: for the remaining \\( n - k \\) tokens, merge each one with the most similar token among the top \\( k \\), thus consolidating information efficiently." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.28, + 0.351, + 0.296 + ], + "angle": 0, + "content": "4.2. Training speed-up: skip FFN" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.484, + 0.429 + ], + "angle": 0, + "content": "As depicted in Figure 4a, we propose the FFN skip strategy: The retained visual tokens, which are few in number, proceed conventionally through all decoder layers of the LLM, while the numerous skipped visual tokens will be limited to the self-attention layers of each transformer block. For a given layer \\( l \\), the output of retained tokens \\( h_{\\mathrm{retained}}^{(l)} \\) is the sum of the self-attention output \\( h_{\\mathrm{attn}}^{(l)} \\) and the FFN-transformed features \\( \\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.435, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\nh _ {\\text {r e t a i n e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)} + \\operatorname {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.461, + 0.331, + 0.476 + ], + "angle": 0, + "content": "The output of the skipped tokens is:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.482, + 0.482, + 0.505 + ], + "angle": 0, + "content": "\\[\nh _ {\\text {s k i p p e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.629 + ], + "angle": 0, + "content": "However, bypassing the FFN layers presents a nontrivial challenge. In the autoregressive training process, the final token of the vision sequence is used to predict the initial token of the text sequence, necessitating a specialized design for this terminal vision token. To facilitate this transition effectively, it must pass through the FFN layers, which are enriched with stored linguistic information [32, 83]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.719 + ], + "angle": 0, + "content": "To address this, as shown in Figure 4, we introduce adaptive summary tokens to efficiently consolidate visual information. This token is generated by the Adaptive Summary Layer before the token sequence is input into the LLM. The layer comprises a simple linear transformation, and its computation is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.725, + 0.483, + 0.745 + ], + "angle": 0, + "content": "\\[\nx ^ {s} = \\left(\\operatorname {s o f t m a x} \\left(X \\cdot W ^ {T}\\right)\\right) ^ {T} \\cdot X, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.886 + ], + "angle": 0, + "content": "where \\( X \\) denotes the retained or skipped visual tokens, \\( x^{s} \\in \\mathbb{R}^{1 \\times C} \\) is the summary token, and \\( W \\in \\mathbb{R}^{1 \\times C} \\) is the weight matrix. Leveraging this layer, we concatenate a summary token on both ends of the skipped visual tokens before their input to the LLM. The first summary token aggregates essential information from retained tokens, mitigating potential loss from extended token sequences. The second summary token integrates prior sequence information and plays a key role in predicting the next text token." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.09, + 0.887, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Figure 5. Similarity density statistics. We conducted a similarity density analysis in CoS, examining the visual tokens derived from various window scale. The results indicate that as the window scale decreases, the similarity between visual tokens becomes more pronounced, reflecting higher token redundancy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.82, + 0.375 + ], + "angle": 0, + "content": "4.3. Inference speed-up: skip KV-cache" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.412 + ], + "angle": 0, + "content": "After training, our Skip-Vision framework incorporates skip KV-cache method to further improve efficiency:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.423, + 0.905, + 0.458 + ], + "angle": 0, + "content": "\\[\nK _ {\\text {s k i p - c a c h e}} ^ {(0)} = K _ {\\text {c a c h e}} ^ {(0)} \\backslash K _ {\\text {s k i p}}, \\quad V _ {\\text {s k i p - c a c h e}} ^ {(0)} = V _ {\\text {c a c h e}} ^ {(0)} \\backslash V _ {\\text {s k i p}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Equation 7 indicates that visual tokens skipped by FFN are removed from the KV cache after the pre-filling stage. As depicted in Figures 4 b) and c), given the autoregressive nature of LLMs, earlier token information is progressively transferred to later tokens. Skip FFN further enhances this integration by suppressing the attention weights of skipped tokens, forcing the model to focus on key tokens and improving information aggregation efficiency. Empirical results show that, due to the inclusion of summary tokens, this omission does not compromise model performance, underscoring a distinctive advantage of our framework for more efficient inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.829, + 0.673 + ], + "angle": 0, + "content": "5. Theoretical analysis of Skip-Vision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.765 + ], + "angle": 0, + "content": "At the core of this analysis lies the layer error incurred when bypassing the FFN layer. For a given layer \\( l \\), the original output \\( h_{\\mathrm{original}}^{(l)} \\) is: \\( h_{\\mathrm{original}}^{(l)} = h_{\\mathrm{attn}}^{(l)} + \\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)}) \\). When skipping the FFN, the output becomes: \\( h_{\\mathrm{skip}}^{(l)} = h_{\\mathrm{attn}}^{(l)} \\). The per-layer skipping error is:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.776, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {(l)} = \\left\\| h _ {\\text {o r i g i n a l}} ^ {(l)} - h _ {\\text {s k i p}} ^ {(l)} \\right\\| _ {2} = \\left\\| \\mathrm {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right) \\right\\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.855 + ], + "angle": 0, + "content": "For redundant tokens, such as those in homogeneous image regions, this error is negligible (\\(\\epsilon^{(l)} \\approx 0\\)) due to minimal feature transformations by the FFN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "However, errors propagate through subsequent layers, amplified by the recursive nature of transformer architectures. Leveraging Lipschitz continuity assumptions for self" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.123 + ], + "angle": 0, + "content": "attention and FFN operations \\((L_{attn}^{(l + 1)}\\) and \\(L_{FFN}^{(l + 1)}\\)), the cumulative error at layer \\(l + 1\\) is bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.132, + 0.483, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {(l + 1)} \\leq \\left(L _ {\\mathrm {a t t n}} ^ {(l + 1)} + L _ {\\mathrm {F F N}} ^ {(l + 1)}\\right) \\cdot \\epsilon^ {(l)} + \\epsilon_ {\\mathrm {s k i p}} ^ {(l + 1)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.164, + 0.483, + 0.198 + ], + "angle": 0, + "content": "where \\((\\epsilon_{\\mathrm{skip}}^{(l + 1)})\\) represents new errors from skipping deeper layers \\((l + 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.199, + 0.395, + 0.214 + ], + "angle": 0, + "content": "Over \\(L\\) layers, the total error telescopes to:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.224, + 0.483, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\text {t o t a l}} \\leq \\sum_ {l = 1} ^ {L} \\epsilon_ {\\text {s k i p}} ^ {(l)} \\cdot \\prod_ {i = 1} ^ {L - l} \\left(L _ {\\text {a t t n}} ^ {(i + 1)} + L _ {\\text {F F N}} ^ {(i + 1)}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.486, + 0.305 + ], + "angle": 0, + "content": "Theorem 5.1 Bounded Lipschitz. If the spectral norms of \\( W_{1}, W_{2}, W_{K}, W_{Q} \\), and \\( W_{V} \\) are bounded by 1, then:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.306, + 0.225, + 0.321 + ], + "angle": 0, + "content": "- \\( L(attn) \\leq 1 / \\sqrt{d_k} \\)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.322, + 0.192, + 0.335 + ], + "angle": 0, + "content": "\\(L(FFN)\\leq 1\\)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.306, + 0.225, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Follow Theorem 5.1, we assume Lipschitz constants \\( L_{\\mathrm{attn}} + L_{\\mathrm{FFN}} \\leq \\gamma \\) and skipping errors \\( \\epsilon_{\\mathrm{skip}}^{(l)} \\leq \\epsilon \\), the total error is scaled to:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.392, + 0.483, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\text {t o t a l}} \\leq \\epsilon \\cdot \\frac {\\gamma^ {L} - 1}{\\gamma - 1}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.489 + ], + "angle": 0, + "content": "Equation 11 establishes that the skip error is bounded when \\(\\gamma < 1\\), provided the model is trained with modern regularization techniques. This ensures that the MLLM remains less sensitive to the effects of skipping." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.52 + ], + "angle": 0, + "content": "This error also impacts the KL divergence between the original and skipped outputs, bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.528, + 0.483, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {K L}} \\left(p _ {\\text {s k i p}} \\| p _ {\\text {o r i g i n a l}}\\right) \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\epsilon_ {\\text {t o t a l}} ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.613 + ], + "angle": 0, + "content": "where \\(\\sigma^2\\) is the variance of the logits. Further integrating feature similarity errors \\((\\epsilon_{\\mathrm{sim}} = O(\\sqrt{1 - \\theta}))\\) from low-attention tokens, the final bound becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.622, + 0.483, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {K L}} \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\left(\\epsilon_ {\\text {t o t a l}} + \\epsilon_ {\\text {s i m}}\\right) ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.721 + ], + "angle": 0, + "content": "Practically, this analysis motivates a layer-wise skipping strategy, alongside token selection and merge based on feature similarity \\((\\theta)\\). More analysis and the proof of Theorem 5.1 are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.222, + 0.75 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.225, + 0.772 + ], + "angle": 0, + "content": "6.1. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To comprehensively evaluate our model, we conducted experiments on eight different benchmarks. These include MME [30], which assesses the perceptual and cognitive capabilities of multimodal language models; TextVQA [95], focused on fine-grained visual question answering; MM-Bench [72] and MMStar [21] for general diagnostic capabilities; MMMU [125], designed to test STEM-related reasoning; MathVista [76], which evaluates mathematical problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "solving skills; OCRBench [70], for optical character recognition tasks; and MMVet [121], used for subjective assessment. Taken together, these benchmarks offer a comprehensive framework for assessing the efficacy of the model in various multimodal challenges." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.182, + 0.71, + 0.198 + ], + "angle": 0, + "content": "6.2. Efficiency evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.448 + ], + "angle": 0, + "content": "We systematically evaluate both training time (hours), inference FLOP computation and inference latency (ms) on single A100 GPU. For training efficiency, we report actual GPU hours measured on consistent hardware configurations. For inference FLOPs, we adopt the methodology from FastV [22], specifically tracking the computational demands of processing visual tokens. In the LLama architecture, we calculate FLOPs for the causal attention and feed forward network modules as \\(4N C^{2} + 2N^{2}C + 3N C M\\) where \\(N\\) is the token count, \\(C\\) denotes the hidden state dimension, and \\(M\\) is the intermediate FFN dimension. SkipVision framework adjusts token counts dynamically between the attention and FFN modules, allowing us to refine FLOP calculations as \\(4N_{1}C^{2} + 2N_{1}^{2}C + 3N_{2}CM\\), where \\(N_{1}\\) and \\(N_{2}\\) represent the visual token counts processed in the causal attention and FFN layers, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.462, + 0.634, + 0.478 + ], + "angle": 0, + "content": "6.3. Data setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.713 + ], + "angle": 0, + "content": "During pretraining, we adopt the experimental settings of LLaVA and CoS, utilizing LLaVA-558K [69] and 65 million image-text pairs for respective comparisons. For the Supervised Fine-Tuning phase, we use LLaVA-665K [69] for all comparative experiments. By leveraging the computational efficiency of Skip-Vision framework, the resources saved can be redirected toward scaling the dataset to further enhance model performance. To this end, we extend our dataset to SV-1M, comprising 1 million samples, demonstrating the substantial performance improvements achievable through data expansion. To further assess our approach against state-of-the-art models of similar scale, we extend the dataset to SV-9M, underscoring the scalability and potential of our method. Detailed dataset statistics are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.663, + 0.743 + ], + "angle": 0, + "content": "6.4. Training setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For LLaVA and LLaVA-HD setting, we use input resolutions of 336 and a dynamic resolution ranging from 336 to 672, respectively, with CLIP ViT-L/336px [86] as the vision encoder and LLaMA3 8B Instruct [29] as the backbone (additional comparisons under LLaVA-1.5-7B training setup are in the Appendix 10.1). For CoS setting, pretraining is conducted at 224 resolution, training only the multi-scale visual resampler with CLIP ViT-L/224px [86]. The model processes 80 visual tokens, including 16 global and 64 local tokens, and scales to 336 and 1296 tokens during SFT." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21389" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.382 + ], + "angle": 0, + "content": "
MMETextvqaMMBMMVetMMMUMMathVOCRBMMStarAvgHoursFLOPs Latency
LLaVA setting:
LLaVA [69]153558.672.232.839.621.332.736.542.010.2100%125
LLaVA-FastV [22]145356.17132.437.320.628.836.340.49.630%103
MQT-LLaVA [39]148753.66928.237.816.828.136.538.67.544%106
LLaVA-Tokenpacker [60]153856.272.332.139.019.329.835.040.58.824%97
SV-LLaVA151956.572.532.438.821.230.736.541.37.925%90
(Nr=100,Ns=156)-1%-3.6%+0.4%-1.2%-2%-0.5%-3.6%-6.1%-1.7%-22.5%-75%-28%
SV-LLaVA153057.372.735.639.221.432.040.042.69.060%103
(Nr=256,Ns=320)-0.3%-2.2%+0.7%+8.5%-1%+0.5%-2.1%+9.6%+1.4%-11.8%-40%-17.6%
LLaVA-HD setting:
LLaVA-HD [67]15336471.835.339.32037.440.544.019100%450
SV-LLaVA-HD153461.273.135.939.721.234.139.343.512.250%250
(Nr=576,Ns=576)+0%-4.3%+1.8%+1.7%+1.5%-6%-8.8%-2.9%-1.1%-35.8%-50%-44.4%
CoS setting:
CoS [40]158564.477.139.439.221.539.241.246.015.3100%160
SV-CoS156363.776.541.740.321.23741.946.09.825%90
(Nr=272,Ns=256)-1.4%-1.1%-0.8%+5.8%+2.8%-1.4%-5.6%+1.7%-0%-40%-75%-43.8%
SV-CoS (SV-1M)156967.177.143.340.722.050.744.449.316.325%90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.392, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Table 2. Performance and efficiency evaluation. We evaluate Skip-Vision on LLaVA, LLaVA-HD, and CoS and compare it with state-of-the-art efficiency optimization models under the LLaVA setting. \\( N_r \\) and \\( N_s \\) denote the number of retained and skipped tokens, respectively." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.423, + 0.892, + 0.58 + ], + "angle": 0, + "content": "
LLMMMETextvqaMMBMMVetMMMUMathVOCRBMMStar
Open-source models in 8B tier:
mPLUG-Owl3 [118]Qwen2-69.077.640.1-65.0-40.1
Cambrian [100]LLaMA3154771.775.9-42.749.062.4-
LLaVA [69]LLaMA3153558.672.232.839.621.332.736.5
Mini-Gemini-HD [61]LLaMA3160670.272.7-37.337.047.7-
LLaVA-Next [56]LLaMa3160464.672.1-41.736.349.0-
Ovis [77]LLaMA3--77.4-44.740.8-49.5
MiniCPM-V2.5 [117]LLaMA3-76.677.6-45.854.372.5-
SV-CoS (SV-9M)LLaMA3155069.178.451.74361.369.452.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.589, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Table 3. MLLM evaluation. We present the performance of SV-CoS on SV-9M, comparing it against the current SOTA models of a similar scale. The highest scores are highlighted in bold, while the second-highest scores are indicated with underlines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.66 + ], + "angle": 0, + "content": "In SFT, the resolution increases to 448, and all model parameters are unfrozen for full optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.797 + ], + "angle": 0, + "content": "In all experiments, we used a learning rate of \\(1e^{-3}\\) for pre-training, which is reduced to \\(1e^{-5}\\) during SFT. Training follows a cosine decay schedule after a \\(3\\%\\) warm-up phase. For skipped token selection, unless otherwise specified, the LLaVA setting includes 156 skipped tokens and 100 retained tokens, the LLaVA-HD setting has 576 skipped and 576 retained tokens, and the CoS setting consists of 256 skipped and 272 retained tokens. All skipped tokens undergo a token merging process to enhance efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.33, + 0.82 + ], + "angle": 0, + "content": "6.5. Main experimental results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Performance and efficiency. We first evaluate the effectiveness of Skip-Vision across three different frameworks: CoS, LLaVA, and LLaVA-HD, demonstrating its ability to maximize efficiency while minimizing performance degradation. As shown in Figure 1 and Table 6, Skip-Vision re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.72 + ], + "angle": 0, + "content": "duces training time by \\(22.5\\%\\), inference FLOPs by \\(75\\%\\), and inference latency by \\(28\\%\\) for LLaVA. For LLaVA-HD, it achieves a \\(35.8\\%\\) reduction in training time, \\(50\\%\\) in inference FLOPs, and \\(45\\%\\) in inference latency. In CoS, it reduces training time by \\(35.8\\%\\), inference FLOPs by \\(75\\%\\), and inference latency by \\(44\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.909, + 0.842 + ], + "angle": 0, + "content": "Scaling. We evaluated the scalability of Skip-Vision, leveraging CoS's visual scaling to increase visual tokens from 336 to 1296 while maintaining efficiency. Expanding the fine-tuning dataset from LLaVA-665k to SV-1M further improved performance by \\(8\\%\\) over \\(C o S_{1296}\\) with similar resources. To benchmark against state-of-the-art models, we scaled to SV-9M, with training completed in 72 hours on 16 NVIDIA A100 GPUs (Table 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.766, + 0.866 + ], + "angle": 0, + "content": "6.6. Ablation study and analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To validate the effectiveness of each component within the Skip-Vision framework, we conducted ablation and com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21390" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.246 + ], + "angle": 0, + "content": "
MMETextvqaMMBMMVetMMMUMathVOCRBMMStarAvgLatency
SV-LLaVA (w/o SK)151956.772.531.438.821.230.936.541.1103
SV-LLaVA151956.572.532.438.821.230.736.541.390
+0%-0.4%+0%+3.2%+0%+0%-0.6%+0%+0.5%-9.7%
SV-LLaVA-HD (w/o SK)153361.673.035.839.921.234.539.743.7310
SV-LLaVA-HD153461.273.135.939.721.234.139.343.5250
+0%-0.7%+0.1%+0.3%-0.5%+0%-1.2%-1%-0.5%-19.4%
SV-LLaVA-CoS (w/o SK)156263.976.540.240.221.237.241.945.9126
SV-LLaVA-CoS156363.776.541.740.321.23741.946.090
+0%-0.3%+0%+3.7%+0.2%+0%-0.5%+0%+0.2%-28.6%
" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.256, + 0.885, + 0.27 + ], + "angle": 0, + "content": "Table 4. Ablation on skip KV-cache. We evaluate SK (using skip KV-cache during inference) of Skip-Vision on different settings." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.271, + 0.36, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.403, + 0.314, + 0.415 + ], + "angle": 0, + "content": "(a) Ablation on LLaVA" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.271, + 0.616, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.445, + 0.403, + 0.553, + 0.415 + ], + "angle": 0, + "content": "(b) Ablation on CoS" + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.308, + 0.628, + 0.345 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.271, + 0.864, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.403, + 0.82, + 0.415 + ], + "angle": 0, + "content": "(c) Ablation on LLaVA-HD" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.427, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 6. Ablation study. We performed an ablation study on each component of the Skip-Vision. SF (skip FFN), TM (token merge)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.447, + 0.485, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.586, + 0.449, + 0.6 + ], + "angle": 0, + "content": "Figure 7. Visualization of attention map in MMVet." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.602, + 0.484, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.741, + 0.455, + 0.756 + ], + "angle": 0, + "content": "Figure 8. Visualization of attention map in TextVQA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.793 + ], + "angle": 0, + "content": "parative experiments on token merge, skip FFN, and skip KV-cache." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 6, both the skip FFN strategy and token merging enhance computational efficiency across different model settings while maintaining performance. Notably, skip FFN performs better in the CoS setting than in LLaVA and LLaVA-HD. We attribute this to CoS allowing the vision backbone to be fine-tuned during the SFT stage, enabling key information to be more effectively captured by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.52 + ], + "angle": 0, + "content": "summary tokens, thereby reducing the errors introduced by skip FFN. As shown in Table 4, skip KV-cache slightly affects fine-grained tasks such as TextVQA and OCRBench but improves performance in long-form response tasks like MMVet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.705 + ], + "angle": 0, + "content": "To further analyze this, we visualized attention maps across layers for visual tokens in SV-CoS during the MMVet and TextVQA tasks. In MMVet (Figure 7), the attention of generated tokens primarily focuses on retained and summary tokens, with skip KV-cache filtering out unnecessary noise. In TextVQA (Figure 8), which requires fine-grained understanding, attention is spread across both retained and skipped tokens, leading to some performance loss with skip KV-cache. However, causal attention effectively aggregates information from skipped tokens into the final summary tokens, allowing us to retain only the summary tokens and retained visual tokens in the KV-cache." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose Skip-Vision, an efficient multimodal large language model architecture that boosts computational efficiency through a skip-FFN strategy, reducing redundant computations over visual tokens during training. During inference, Skip-Vision uses a skip KV-cache to accelerate processing by omitting non-essential visual tokens from the KV cache. Extensive experiments show our model outperforms current methods in efficiency and scales effectively with data, delivering competitive performance among leading models of similar scale." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21391" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.173 + ], + "angle": 0, + "content": "This work was supported in part by NSFC (62201342), and Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102). This work was supported by Ant Group Research Intern Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.224, + 0.484, + 0.278 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.281, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.351, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 23716-23736, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.435, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.506, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[5] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.577, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.647, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[7] Samyadeep Basu, Martin Grayson, Cecily Morrison, Besmira Nushi, Soheil Feizi, and Daniela Massiceti. Understanding information storage and transfer in multi-modal large language models. arXiv preprint arXiv:2406.04236, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.717, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.788, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[9] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.845, + 0.481, + 0.872 + ], + "angle": 0, + "content": "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Shihao Cai, Keqin Bao, Hangyu Guo, Jizhi Zhang, Jun Song, and Bo Zheng. Geogpt4v: Towards geometric multi-" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.224, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "modal large language models with geometric image generation. arXiv preprint arXiv:2406.11503, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[12] Jianjian Cao, Peng Ye, Shengze Li, Chong Yu, Yansong Tang, Jiwen Lu, and Tao Chen. Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15710-15719, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Jimmy Carter. Textocr-gpt4v. https://huggingface.co/datasets/jimmycarter/textocr-gpt4v,2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[14] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[16] Chungkwong Chan. Stroke extraction for offline handwritten mathematical expression recognition. IEEE Access, 8: 61565-61575, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[17] Shuaichen Chang, David Palzer, Jialin Li, Eric Fosler-Lussier, and Ningchuan Xiao. Mapqa: A dataset for question answering on choropleth maps. arXiv preprint arXiv:2211.08545, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[18] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[19] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[21] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[22] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[25] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/, 2023.1.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.218, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[26] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaiqiao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1571–1576. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[27] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[28] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[29] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[30] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[31] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[32] Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[33] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2315-2324, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[34] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939–948, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Zero-shot extreme" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "length generalization for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3991-4008, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[36] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In 2018 24th international conference on pattern recognition (ICPR), pages 7-12. IEEE, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.246, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[37] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.316, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[38] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.385, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[39] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[40] Ziyuan Huang, Kaixiang Ji, Biao Gong, Zhiwu Qing, Qinglong Zhang, Kecheng Zheng, Jian Wang, Jingdong Chen, and Ming Yang. Accelerating pre-training of multimodal llms via chain-of-sight. arXiv preprint arXiv:2407.15819, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[41] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.58, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[42] Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.637, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[43] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[44] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[45] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figurea: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21393" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[47] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[48] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[49] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[50] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern recognition, pages 4999-5007, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[51] Sunjun Kweon, Yeonsu Kwon, Seonhee Cho, Yohan Jo, and Edward Choi. Open-wikitable: Dataset for open domain question answering with complex reasoning over table. arXiv preprint arXiv:2305.07288, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[52] Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1):1-10, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[53] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[54] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: mitigating multimodal hallucination through self-feedback guided revision. arXiv preprint arXiv:2311.07362, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[55] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[56] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild. https://llava-v1.github.io/blog/2024-05-10-llava-last-stronger-llms/, 2024.1,3,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[57] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[59] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[60] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.307, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[61] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.377, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[62] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[63] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[64] Dongyang Liu, Renrui Zhang, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, et al. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models. arXiv preprint arXiv:2402.05935, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[65] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[66] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[67] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[68] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[69] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7, 5" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21394" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[70] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[71] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[72] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[73] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Icdar 2023 competition on hierarchical text detection and recognition. arXiv preprint arXiv:2305.09750, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[74] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[75] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[76] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[77] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[78] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[79] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[80] Minesh Mathew, Lluis Gomez, Dimosthenis Karatzas, and CV Jawahar. Asking questions on handwritten document collections. International Journal on Document Analysis and Recognition (IJDAR), 24(3):235-249, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[81] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[82] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[83] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.166, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[84] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.223, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[85] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.266, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[86] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.351, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[87] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia, 2018. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.436, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[88] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.521, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[89] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[90] Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar. Kvqa: Knowledge-aware visual question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8876-8884, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[91] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.719, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[92] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[93] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[94] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21395" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[95] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[96] Dominik Stammbach and Elliott Ash. Docscan: Unsupervised text classification via learning from neighbors. arXiv preprint arXiv:2105.04024, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[97] Hamed Rahimi Sujet AI, Allaa Boutaleb. Sujet-finance-qavision-100k: A large-scale dataset for financial document vqa, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[98] S Svetlichnaya. Deepform: Understand structured documents at scale. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[99] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[100] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.401, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[101] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[102] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.526, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[103] Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.581, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[104] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.651, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[105] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[106] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Conghui He, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[107] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jian-Yuan Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. ArXiv, abs/2312.06109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[108] Yuxin Wen, Qingqing Cao, Qichen Fu, Sachin Mehta, and Mahyar Najibi. Efficient vision-language models by" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "summarizing visual tokens into compact registers. arXiv preprint arXiv:2410.14072, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.903, + 0.149 + ], + "angle": 0, + "content": "[109] C. Wendler. Renderedtext. https://huggingface.co/datasets/wendlerc/RenderedText, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[110] Jianshu Chen Yunkai Zhang Hong Wang Shiyang Li Xiyou Zhou Wenhu Chen, Hongmin Wang and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[111] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[112] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[113] Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts. arXiv preprint arXiv:2312.15915, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[114] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[115] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[116] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[117] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[118] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[119] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[120] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthyness. arXiv preprint arXiv:2405.17220, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[121] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21396" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[122] Xinyan Velocity Yu, Sewon Min, Luke Zettlemoyer, and Hannaneh Hajishirzi. Crepe: Open-domain question answering with false presuppositions. arXiv preprint arXiv:2211.17257, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[123] Qianhao Yuan, Qingyu Zhang, Yanjiang Liu, Jiawei Chen, Yaojie Lu, Hongyu Lin, Jia Zheng, Xianpei Han, and Le Sun. Shortv: Efficient multimodal large language models by freezing visual tokens in ineffective layers. arXiv preprint arXiv:2504.00502, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[124] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[125] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[126] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5317-5327, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[127] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[128] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[129] Ying Zhang, Lionel Gueguen, Ilya Zharkov, Peter Zhang, Keith Seifert, and Ben Kadlec. Uber-text: A large-scale dataset for optical character recognition from street-level imagery. In SUNw: Scene Understanding Workshop-CVPR, page 5, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[130] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[131] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[132] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[133] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[134] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[135] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21397" + } + ] +] \ No newline at end of file diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_origin.pdf b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9373f0f8832c6c10f6f50cfbcde3c6f041523dc4 --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/1d1ce44b-d4d0-40e2-b544-e98ad6e37776_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5d4aa05041a1c5edfd1132d4793ec0a42aee3798df9ad376c77aa43c0d6d172 +size 9473710 diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/full.md b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ed2a225c167518121bbc972e6a4c64684768e3d --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/full.md @@ -0,0 +1,444 @@ +# Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping + +Weili Zeng $^{1*}$ , Ziyuan Huang $^{2}$ , Kaixiang Ji $^{2}$ , Yichao Yan $^{1\dagger}$ + +1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Ant Group + +{zw1666,yanyichao}@sjtu.edu.cn, ziyuan.huang@u.nus.edu, kaixiang.jkx@antgroup.com + +# Abstract + +Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that SkipVision reduces training time by up to $35\%$ , inference FLOPs by $75\%$ , and latency by $45\%$ , while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency. Our project page can be found here. + +# 1. Introduction + +Transformer-based large-scale models have significantly advanced progress in Artificial Intelligence (AI) [10, 86, 101]. Their remarkable capabilities have driven the emergence of modern Multimodal Large Language Models (MLLMs) [2, 3, 69, 88], which demonstrate vision-language competencies comparable to human performance. Empirical evidence suggests that scaling laws continue to + +![](images/854eda7a7f731484f8da78f37fae6a357bb004a72bcf968b8248987b0bd9c0f4.jpg) +Figure 1. Performance-efficiency trade-off curve. Each circle denotes a model configuration, where our models utilize the Skip-Vision framework, with CoS [40], LLaVA-HD [67] and LLaVA [69] serving as baselines. Circle sizes reflect the inference FLOPs ratio. Skip-Vision demonstrate superior performance, scaling effectively with increased FLOPs and data and achieving higher inference efficiency when compared to baselines and other efficient MLLM methods. All methods utilize LLaMA3 8B as the foundational large language model. + +be effective, primarily across three dimensions: visual scaling [40, 67, 112], data scaling [63, 64], and model scaling [56, 104]. However, these scaling methods present substantial efficiency issues, significantly increasing the training and inference burden. In a typical LLaVA-1.5 framework [69], where the model receives 576 visual tokens and 40 text tokens, inference using Vicuna-13B [25] LLM backbone demands 200 trillion FLOPS [42]. These motivate our pursuit of an MLLM architecture that optimally balances efficiency and performance. + +Recent advances in Multimodal Large Language Models (MLLMs), such as Chain-of-Sight (CoS)[40] and LLaVA- + +
Llama layerComplexityParameters
FFN10.5LNC25.64B(70%)
Attention2LN2C + 1.5NC21.34B(17%)
Embedding-0.5B(6%)
+ +Table 1. Computational complexity and parameter statistics for Llama3 8B. L denotes the number of blocks, N represents the number of input tokens, and C is the feature dimensionality + +Next[68], have improved performance by increasing visual tokens. However, during the Supervised Fine-Tuning (SFT) stage, the computational cost rises dramatically, leading to diminishing returns and making the scaling unsustainable. As shown in Figure 1, CoS requires a $165\%$ increase in training resources and a $400\%$ increase in inference computation for a mere $1.7\%$ performance improvement, while LLaVA-Next faces similar inefficiencies. This challenge emphasizes the need to reduce visual token computation. Approaches like token compression or pruning [15, 22, 39, 60, 116, 119], while efficient, risk losing critical visual information, revealing a trade-off between efficiency and performance. This work aims to optimize the model's computational efficiency from an architectural perspective, proposing a scalable solution that balances efficiency with visual fidelity. + +We systematically examine the parameter count and FLOPs distribution across different components of the LLM. As shown in Table 1, using LLaMA-3 [29] as an example, the parameter and FLOPS requirements for the FFN are significantly higher than the attention module. Furthermore, as illustrated in Figure 2, when the number of visual tokens is large, the majority of computations are concentrated on FFN of these tokens. In particular, we computed the difference in the magnitudes of the features before and after the FFN transformation for different tokens. As illustrated in the figure 3, we observe that the updates introduced by the FFN to the visual tokens are significantly smaller than those that affect the text tokens. This is similar to the observation in [123]. This observation suggests a potential approach: could we bypass FFN computations for a substantial portion of visual tokens? Furthermore, in the autoregressive training framework of MLLMs, only the final visual token utilizes its last hidden state for next-token prediction, while the other visual tokens primarily serve to convey causal information. During model inference, inspired by [7], we observe that most visual information is progressively integrated into the final token through the early causal layers. This raises the possibility of reducing reliance on the full visual scaling KV cache throughout inference, potentially accelerating the inference process. + +Building on these insights, we propose Skip-Vision, a novel and flexible architecture designed for more efficient scaling (see Figure 1 for a comparison). Visual tokens de + +![](images/8be379eb26551a301c26d0ca932d87f74a22843ebb31fc8e68dbdd11bc04f8eb.jpg) +Figure 2. Flops ratio of visual tokens in FFN. When visual tokens greatly outnumber text tokens, computation is predominantly consumed by the FFN of visual tokens. Reducing their passage through the FFN can thus markedly decrease computational costs. + +rived from smaller window/scale sizes constitute approximately $80\%$ of the total visual tokens and account for most of the computational overhead. To mitigate this, as shown in Figure 4, we allow these tokens to skip the Feed-Forward Network (FFN) computation, significantly reducing overall computational costs. To address redundancy within skipped vision tokens, we can apply a token merging strategy for further simplification. During inference, skipped tokens are removed from the KV cache after the pre-fill phase, accelerating subsequent computation. Empirical results demonstrate that Skip-Vision reduces training time by up to $35\%$ , FLOPs of inference by $75\%$ and inference latency by $45\%$ , while maintaining or even improving performance. These efficiency gains provide a scalable solution for large-scale multimodal learning, enabling better data utilization and enhanced model scalability. + +In summary, Skip-Vision is designed to be seamlessly integrated into the standard SFT pipeline of MLLMs without introducing additional retraining or decoupled modules. It directly modifies the transformer's computation flow, offering a practical and theoretically grounded acceleration solution for MLLM training and inference jointly. Our contributions are as follows: + +- We introduce a novel and efficient framework, using token merge and skip FFN strategy during training to reduce redundant computation for visual tokens. +- In inference, our framework employs a skip KV-cache mechanism that removes skip-FFN visual tokens from the KV-cache, enhancing efficiency. +- We present a theoretical analysis of the performance error in Skip-Vision and establish an error bound, which will be empirically validated through experiments. +- Experiments show our model's superior efficiency, effective data scaling, and performance on par with state-of + +![](images/90b52cde8ad17b80416bc12f56fd3c53c4709de6a0bbfeb3e77d9f6d0defaf22.jpg) +(a) FFN updates for CoS + +![](images/3043d0a625953e2bf208b09527d18729cad7c27bec26f8071729072df4e4959f.jpg) +(b) FFN updates for LLaVA-HD +Figure 3. FFN imapct for different MLLMs. We evaluate FFN impact by computing the feature modulus ratio before $(\| h_{\mathrm{attn}}\|_2)$ and after FFN $(\| \mathrm{FFN}(h_{\mathrm{attn}})\|_2)$ . The FFN introduces significantly smaller updates to visual tokens compared to text tokens. + +![](images/a1f9d01c36c74b7a0bef88cd90a6a14d09d9558f6e27fcb01b5e6b2402bd5f80.jpg) +(c) FFN updates for LLaV + +the-art models of similar scale. + +# 2. Related work + +# 2.1. Vison-Language models + +Rapid progress in large language models (LLMs) [25, 29, 114] has laid a solid foundation for the emergence of large-scale vision-language models (VLMs) [3, 6, 19, 58]. Flamingo [3] was a pioneering effort, integrating pretrained image encoders with LLMs to handle multimodal data, thereby allowing comprehensive understanding and reasoning across modalities. Following this, the major models developed by leading corporations, such as GPT-4v [2] and Gemini-1.5-Pro [88], have led the progress in this domain. Using proprietary data sets and undisclosed training methodologies, these models have elevated multimodal intelligence to unprecedented levels. + +At the same time, the open-source community has worked diligently to stay abreast of these advancements. The LLava series models [56, 67, 69, 112] exemplify a key strategy in this domain by mapping visual features directly into the input embedding space of the language model, integrating them as input tokens. However, this approach generates a substantial volume of visual tokens, frequently exceeding the number of text tokens, creating efficiency challenges. InternVL-1.5 [24] adopts dynamic high-resolution techniques, segmenting large images into smaller fragments for processing and scaling them accordingly. MiniCPMV [117] utilizes a perceiver resampler structure with a single layer of cross-attention to compress visual tokens. Models such as Vary [107], SPHINX [64], Cambrian-1 [100], and BRAVE [47] leverage multi-visual encoder architectures to strengthen their visual processing capabilities. Furthermore, a crucial factor contributing to the effectiveness of modern VLMs is instruction fine-tuning [28, 67, 134], enabling these models to function as conversational agents and interact with users through natural, human-like dialogue. + +# 2.2. Efficient multimodal large language models + +To enhance the efficiency of vision-language models, one of the primary and widely employed approaches in contemporary research involves reducing the size of the backbone models [27, 124, 133?]. Simultaneously, given that visual tokens contribute significantly to computational demands, research focused on compressing or refining these tokens has garnered increasing attention. Such methods are often coupled with the design of vision-language projectors. For instance, techniques like the Perceiver Resampler [3, 6, 41] and Q-Former [58] utilize transformer-based mechanisms to consolidate visual tokens into more compact query sets. In this way, the transformer outputs corresponding to the positions of the learnable latent queries serve as the aggregated representation of the visual features [42]. Honeybee [14] propose two visual projectors, namely C-Abstractor and D-Abstractor. LLaVA-PruMerge [91] and MADTP [12] introduce adaptive methods for reducing visual tokens. In addition, while the large language model (LLM) field has employed various techniques for token reduction to accelerate inference and compress KV cache [35, 131], research in this area remains limited within the MLLM domain. + +# 3. Preliminaries + +# 3.1. Decoder-only LLM + +Each Transformer layer in a decoder-only LLM primarily consists of a self-attention layer and a feed-forward neural network. To facilitate autoregressive inference, the use of a KV-cache has been developed to accelerate the decoding process. + +Self-Attention Mechanism: The self-attention mechanism enables each token in a sequence to selectively attend to other tokens, capturing contextual dependencies. Each token generates a query $(Q)$ , key $(K)$ , and value $(V)$ vector. Attention scores are calculated by the scaled dot-product: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V, \tag {1} +$$ + +![](images/df2c8b74ad126f2c273114dbae5e81ef25bd7e6e89b6c1b8ffa8b57f2ec80f5a.jpg) +Figure 4. The framework of Skip-Vision. a) While visual scaling enriches visual information, it also increases computational overhead. Skip-Vision uses a skip-FFN strategy during training to reduce redundant computation for visual tokens. The numerous skipped tokens will be limited to the attention layer and bypass FFN. b) At the beginning of inference, Skip-Vision will remove skip-FFN visual tokens from the initial KV-cache, enhancing efficiency. c) During inference, skip attention leverages the skip KV-cache to accelerate generation. + +where $d_{k}$ is the dimension of $K$ . This formulation allows each token to incorporate information from others based on relevance, making self-attention critical in capturing relationships across tokens. + +Feed-Forward Neural Network: The feed-forward neural network enhances token representations following the self-attention layer. The operation can be expressed as: + +$$ +\operatorname {F F N} (x) = \text {A c t i v a t i o n} \left(x W _ {1} + b _ {1}\right) W _ {2} + b _ {2}, \tag {2} +$$ + +where $x$ is the input, $W_{1}$ and $W_{2}$ are weight matrices, and $b_{1}$ and $b_{2}$ are biases. This architecture allows the model to capture complex patterns and improve the expressiveness of token representations. + +KV-cache: The key-value cache (KV-cache) stores key $(K)$ and value $(V)$ representations from previous self-attention computations, enabling efficient reuse during inference. This approach avoids redundant calculations for earlier tokens, improving inference speed. + +The attention score is computed using Attention $(Q, K_{\text{cache}}^t, V_{\text{cache}}^t)$ , where $Q$ is the query for the current token. At each time step $t$ , the cache is updated as: + +$$ +K _ {\text {c a c h e}} ^ {(t)} = \left[ K _ {\text {c a c h e}} ^ {(t - 1)}; K _ {t} \right], \quad V _ {\text {c a c h e}} ^ {(t)} = \left[ V _ {\text {c a c h e}} ^ {(t - 1)}; V _ {t} \right], \tag {3} +$$ + +where $K_{t}$ and $V_{t}$ are the current token's key and value. This mechanism enhances computational efficiency and accelerates token generation. + +# 4. Method + +# 4.1. Skipped token selection and token merge + +Before applying the skip FFN strategy, we first identify the skipped and retained visual tokens. The retained tokens should preserve most essential visual information, while the skipped tokens serve as complementary details, ensuring efficiency without significant loss of information. + +Fortunately, in MLLM architectures, we can naturally distinguish retained and skipped tokens based on global and local context tokens. Global context tokens are retained, while local context tokens are skipped. In architectures like LLaVA-HD, global tokens originate from the resized full image, whereas local tokens come from smaller image patches. In CoS, global tokens correspond to larger window sizes, while local tokens are derived from smaller windows. Since LLaVA lacks a clear distinction between global and local tokens, inspired by [115], we select the top- $n$ tokens most similar to the CLS token as retained tokens, categorizing the rest as skipped tokens. + +To further reduce the computational burden, it is necessary to refine the visual tokens to eliminate excessive redundancy. As shown in Figure 5, we conducted a similarity analysis of the visual tokens in CoS at different levels, calculating their similarity density. We found that the smaller the granularity of the visual tokens, the higher their average similarity, while their token count is significantly greater than that of other granularities. Consequently, inspired by + +ToMe [9], we employ a token merge method to reduce redundancy in skipped tokens, as detailed below: + +- Compute cosine similarities: calculate the cosine similarity between each pair of tokens in the set, obtaining an average similarity score for each token. +- Select top $k$ distinctive Tokens: sort tokens by average similarity in ascending order and retain the top $k$ tokens with the lowest average similarity, representing the most distinctive information. +- Merge remaining tokens: for the remaining $n - k$ tokens, merge each one with the most similar token among the top $k$ , thus consolidating information efficiently. + +# 4.2. Training speed-up: skip FFN + +As depicted in Figure 4a, we propose the FFN skip strategy: The retained visual tokens, which are few in number, proceed conventionally through all decoder layers of the LLM, while the numerous skipped visual tokens will be limited to the self-attention layers of each transformer block. For a given layer $l$ , the output of retained tokens $h_{\mathrm{retained}}^{(l)}$ is the sum of the self-attention output $h_{\mathrm{attn}}^{(l)}$ and the FFN-transformed features $\mathrm{FFN}^{(l)}(h_{\mathrm{attn}}^{(l)})$ : + +$$ +h _ {\text {r e t a i n e d}} ^ {(l)} = h _ {\text {a t t n}} ^ {(l)} + \operatorname {F F N} ^ {(l)} \left(h _ {\text {a t t n}} ^ {(l)}\right). \tag {4} +$$ + +The output of the skipped tokens is: + +$$ +h _ {\text {s k i p p e d}} ^ {(l)} = h _ {\text {a t t n}} ^ {(l)}. \tag {5} +$$ + +However, bypassing the FFN layers presents a nontrivial challenge. In the autoregressive training process, the final token of the vision sequence is used to predict the initial token of the text sequence, necessitating a specialized design for this terminal vision token. To facilitate this transition effectively, it must pass through the FFN layers, which are enriched with stored linguistic information [32, 83]. + +To address this, as shown in Figure 4, we introduce adaptive summary tokens to efficiently consolidate visual information. This token is generated by the Adaptive Summary Layer before the token sequence is input into the LLM. The layer comprises a simple linear transformation, and its computation is defined as follows: + +$$ +x ^ {s} = \left(\operatorname {s o f t m a x} \left(X \cdot W ^ {T}\right)\right) ^ {T} \cdot X, \tag {6} +$$ + +where $X$ denotes the retained or skipped visual tokens, $x^{s} \in \mathbb{R}^{1 \times C}$ is the summary token, and $W \in \mathbb{R}^{1 \times C}$ is the weight matrix. Leveraging this layer, we concatenate a summary token on both ends of the skipped visual tokens before their input to the LLM. The first summary token aggregates essential information from retained tokens, mitigating potential loss from extended token sequences. The second summary token integrates prior sequence information and plays a key role in predicting the next text token. + +![](images/0c96dbe6580415cba281479de909b3324509d8ed8de123e3a955dd1de121f6a7.jpg) +Figure 5. Similarity density statistics. We conducted a similarity density analysis in CoS, examining the visual tokens derived from various window scale. The results indicate that as the window scale decreases, the similarity between visual tokens becomes more pronounced, reflecting higher token redundancy. + +# 4.3. Inference speed-up: skip KV-cache + +After training, our Skip-Vision framework incorporates skip KV-cache method to further improve efficiency: + +$$ +K _ {\text {s k i p - c a c h e}} ^ {(0)} = K _ {\text {c a c h e}} ^ {(0)} \backslash K _ {\text {s k i p}}, \quad V _ {\text {s k i p - c a c h e}} ^ {(0)} = V _ {\text {c a c h e}} ^ {(0)} \backslash V _ {\text {s k i p}}, \tag {7} +$$ + +Equation 7 indicates that visual tokens skipped by FFN are removed from the KV cache after the pre-filling stage. As depicted in Figures 4 b) and c), given the autoregressive nature of LLMs, earlier token information is progressively transferred to later tokens. Skip FFN further enhances this integration by suppressing the attention weights of skipped tokens, forcing the model to focus on key tokens and improving information aggregation efficiency. Empirical results show that, due to the inclusion of summary tokens, this omission does not compromise model performance, underscoring a distinctive advantage of our framework for more efficient inference. + +# 5. Theoretical analysis of Skip-Vision + +At the core of this analysis lies the layer error incurred when bypassing the FFN layer. For a given layer $l$ , the original output $h_{\mathrm{original}}^{(l)}$ is: $h_{\mathrm{original}}^{(l)} = h_{\mathrm{attn}}^{(l)} + \mathrm{FFN}^{(l)}(h_{\mathrm{attn}}^{(l)})$ . When skipping the FFN, the output becomes: $h_{\mathrm{skip}}^{(l)} = h_{\mathrm{attn}}^{(l)}$ . The per-layer skipping error is: + +$$ +\epsilon^ {(l)} = \left\| h _ {\text {o r i g i n a l}} ^ {(l)} - h _ {\text {s k i p}} ^ {(l)} \right\| _ {2} = \left\| \mathrm {F F N} ^ {(l)} \left(h _ {\text {a t t n}} ^ {(l)}\right) \right\| _ {2}. \tag {8} +$$ + +For redundant tokens, such as those in homogeneous image regions, this error is negligible ( $\epsilon^{(l)} \approx 0$ ) due to minimal feature transformations by the FFN. + +However, errors propagate through subsequent layers, amplified by the recursive nature of transformer architectures. Leveraging Lipschitz continuity assumptions for self + +attention and FFN operations $(L_{attn}^{(l + 1)}$ and $L_{FFN}^{(l + 1)}$ ), the cumulative error at layer $l + 1$ is bounded by: + +$$ +\epsilon^ {(l + 1)} \leq \left(L _ {\mathrm {a t t n}} ^ {(l + 1)} + L _ {\mathrm {F F N}} ^ {(l + 1)}\right) \cdot \epsilon^ {(l)} + \epsilon_ {\mathrm {s k i p}} ^ {(l + 1)}, \tag {9} +$$ + +where $(\epsilon_{\mathrm{skip}}^{(l + 1)})$ represents new errors from skipping deeper layers $(l + 1)$ . + +Over $L$ layers, the total error telescopes to: + +$$ +\epsilon_ {\text {t o t a l}} \leq \sum_ {l = 1} ^ {L} \epsilon_ {\text {s k i p}} ^ {(l)} \cdot \prod_ {i = 1} ^ {L - l} \left(L _ {\text {a t t n}} ^ {(i + 1)} + L _ {\text {F F N}} ^ {(i + 1)}\right). \tag {10} +$$ + +Theorem 5.1 Bounded Lipschitz. If the spectral norms of $W_{1}, W_{2}, W_{K}, W_{Q}$ , and $W_{V}$ are bounded by 1, then: + +- $L(attn) \leq 1 / \sqrt{d_k}$ +$L(FFN)\leq 1$ + +Follow Theorem 5.1, we assume Lipschitz constants $L_{\mathrm{attn}} + L_{\mathrm{FFN}} \leq \gamma$ and skipping errors $\epsilon_{\mathrm{skip}}^{(l)} \leq \epsilon$ , the total error is scaled to: + +$$ +\epsilon_ {\text {t o t a l}} \leq \epsilon \cdot \frac {\gamma^ {L} - 1}{\gamma - 1}, \tag {11} +$$ + +Equation 11 establishes that the skip error is bounded when $\gamma < 1$ , provided the model is trained with modern regularization techniques. This ensures that the MLLM remains less sensitive to the effects of skipping. + +This error also impacts the KL divergence between the original and skipped outputs, bounded by: + +$$ +\mathcal {D} _ {\mathrm {K L}} \left(p _ {\text {s k i p}} \| p _ {\text {o r i g i n a l}}\right) \leq \frac {1}{2 \sigma^ {2}} \cdot \epsilon_ {\text {t o t a l}} ^ {2}, \tag {12} +$$ + +where $\sigma^2$ is the variance of the logits. Further integrating feature similarity errors $(\epsilon_{\mathrm{sim}} = O(\sqrt{1 - \theta}))$ from low-attention tokens, the final bound becomes: + +$$ +\mathcal {D} _ {\mathrm {K L}} \leq \frac {1}{2 \sigma^ {2}} \cdot \left(\epsilon_ {\text {t o t a l}} + \epsilon_ {\text {s i m}}\right) ^ {2}. \tag {13} +$$ + +Practically, this analysis motivates a layer-wise skipping strategy, alongside token selection and merge based on feature similarity $(\theta)$ . More analysis and the proof of Theorem 5.1 are provided in the Appendix. + +# 6. Experiments + +# 6.1. Benchmarks + +To comprehensively evaluate our model, we conducted experiments on eight different benchmarks. These include MME [30], which assesses the perceptual and cognitive capabilities of multimodal language models; TextVQA [95], focused on fine-grained visual question answering; MM-Bench [72] and MMStar [21] for general diagnostic capabilities; MMMU [125], designed to test STEM-related reasoning; MathVista [76], which evaluates mathematical problem + +solving skills; OCRBench [70], for optical character recognition tasks; and MMVet [121], used for subjective assessment. Taken together, these benchmarks offer a comprehensive framework for assessing the efficacy of the model in various multimodal challenges. + +# 6.2. Efficiency evaluation + +We systematically evaluate both training time (hours), inference FLOP computation and inference latency (ms) on single A100 GPU. For training efficiency, we report actual GPU hours measured on consistent hardware configurations. For inference FLOPs, we adopt the methodology from FastV [22], specifically tracking the computational demands of processing visual tokens. In the LLama architecture, we calculate FLOPs for the causal attention and feed forward network modules as $4N C^{2} + 2N^{2}C + 3N C M$ where $N$ is the token count, $C$ denotes the hidden state dimension, and $M$ is the intermediate FFN dimension. SkipVision framework adjusts token counts dynamically between the attention and FFN modules, allowing us to refine FLOP calculations as $4N_{1}C^{2} + 2N_{1}^{2}C + 3N_{2}CM$ , where $N_{1}$ and $N_{2}$ represent the visual token counts processed in the causal attention and FFN layers, respectively. + +# 6.3. Data setup + +During pretraining, we adopt the experimental settings of LLaVA and CoS, utilizing LLaVA-558K [69] and 65 million image-text pairs for respective comparisons. For the Supervised Fine-Tuning phase, we use LLaVA-665K [69] for all comparative experiments. By leveraging the computational efficiency of Skip-Vision framework, the resources saved can be redirected toward scaling the dataset to further enhance model performance. To this end, we extend our dataset to SV-1M, comprising 1 million samples, demonstrating the substantial performance improvements achievable through data expansion. To further assess our approach against state-of-the-art models of similar scale, we extend the dataset to SV-9M, underscoring the scalability and potential of our method. Detailed dataset statistics are provided in the Appendix. + +# 6.4. Training setup + +For LLaVA and LLaVA-HD setting, we use input resolutions of 336 and a dynamic resolution ranging from 336 to 672, respectively, with CLIP ViT-L/336px [86] as the vision encoder and LLaMA3 8B Instruct [29] as the backbone (additional comparisons under LLaVA-1.5-7B training setup are in the Appendix 10.1). For CoS setting, pretraining is conducted at 224 resolution, training only the multi-scale visual resampler with CLIP ViT-L/224px [86]. The model processes 80 visual tokens, including 16 global and 64 local tokens, and scales to 336 and 1296 tokens during SFT. + +
MMETextvqaMMBMMVetMMMUMMathVOCRBMMStarAvgHoursFLOPs Latency
LLaVA setting:
LLaVA [69]153558.672.232.839.621.332.736.542.010.2100%125
LLaVA-FastV [22]145356.17132.437.320.628.836.340.49.630%103
MQT-LLaVA [39]148753.66928.237.816.828.136.538.67.544%106
LLaVA-Tokenpacker [60]153856.272.332.139.019.329.835.040.58.824%97
SV-LLaVA151956.572.532.438.821.230.736.541.37.925%90
(Nr=100,Ns=156)-1%-3.6%+0.4%-1.2%-2%-0.5%-3.6%-6.1%-1.7%-22.5%-75%-28%
SV-LLaVA153057.372.735.639.221.432.040.042.69.060%103
(Nr=256,Ns=320)-0.3%-2.2%+0.7%+8.5%-1%+0.5%-2.1%+9.6%+1.4%-11.8%-40%-17.6%
LLaVA-HD setting:
LLaVA-HD [67]15336471.835.339.32037.440.544.019100%450
SV-LLaVA-HD153461.273.135.939.721.234.139.343.512.250%250
(Nr=576,Ns=576)+0%-4.3%+1.8%+1.7%+1.5%-6%-8.8%-2.9%-1.1%-35.8%-50%-44.4%
CoS setting:
CoS [40]158564.477.139.439.221.539.241.246.015.3100%160
SV-CoS156363.776.541.740.321.23741.946.09.825%90
(Nr=272,Ns=256)-1.4%-1.1%-0.8%+5.8%+2.8%-1.4%-5.6%+1.7%-0%-40%-75%-43.8%
SV-CoS (SV-1M)156967.177.143.340.722.050.744.449.316.325%90
+ +Table 2. Performance and efficiency evaluation. We evaluate Skip-Vision on LLaVA, LLaVA-HD, and CoS and compare it with state-of-the-art efficiency optimization models under the LLaVA setting. $N_r$ and $N_s$ denote the number of retained and skipped tokens, respectively. + +
LLMMMETextvqaMMBMMVetMMMUMathVOCRBMMStar
Open-source models in 8B tier:
mPLUG-Owl3 [118]Qwen2-69.077.640.1-65.0-40.1
Cambrian [100]LLaMA3154771.775.9-42.749.062.4-
LLaVA [69]LLaMA3153558.672.232.839.621.332.736.5
Mini-Gemini-HD [61]LLaMA3160670.272.7-37.337.047.7-
LLaVA-Next [56]LLaMa3160464.672.1-41.736.349.0-
Ovis [77]LLaMA3--77.4-44.740.8-49.5
MiniCPM-V2.5 [117]LLaMA3-76.677.6-45.854.372.5-
SV-CoS (SV-9M)LLaMA3155069.178.451.74361.369.452.3
+ +Table 3. MLLM evaluation. We present the performance of SV-CoS on SV-9M, comparing it against the current SOTA models of a similar scale. The highest scores are highlighted in bold, while the second-highest scores are indicated with underlines. + +In SFT, the resolution increases to 448, and all model parameters are unfrozen for full optimization. + +In all experiments, we used a learning rate of $1e^{-3}$ for pre-training, which is reduced to $1e^{-5}$ during SFT. Training follows a cosine decay schedule after a $3\%$ warm-up phase. For skipped token selection, unless otherwise specified, the LLaVA setting includes 156 skipped tokens and 100 retained tokens, the LLaVA-HD setting has 576 skipped and 576 retained tokens, and the CoS setting consists of 256 skipped and 272 retained tokens. All skipped tokens undergo a token merging process to enhance efficiency. + +# 6.5. Main experimental results + +Performance and efficiency. We first evaluate the effectiveness of Skip-Vision across three different frameworks: CoS, LLaVA, and LLaVA-HD, demonstrating its ability to maximize efficiency while minimizing performance degradation. As shown in Figure 1 and Table 6, Skip-Vision re + +duces training time by $22.5\%$ , inference FLOPs by $75\%$ , and inference latency by $28\%$ for LLaVA. For LLaVA-HD, it achieves a $35.8\%$ reduction in training time, $50\%$ in inference FLOPs, and $45\%$ in inference latency. In CoS, it reduces training time by $35.8\%$ , inference FLOPs by $75\%$ , and inference latency by $44\%$ . + +Scaling. We evaluated the scalability of Skip-Vision, leveraging CoS's visual scaling to increase visual tokens from 336 to 1296 while maintaining efficiency. Expanding the fine-tuning dataset from LLaVA-665k to SV-1M further improved performance by $8\%$ over $C o S_{1296}$ with similar resources. To benchmark against state-of-the-art models, we scaled to SV-9M, with training completed in 72 hours on 16 NVIDIA A100 GPUs (Table 3). + +# 6.6. Ablation study and analysis + +To validate the effectiveness of each component within the Skip-Vision framework, we conducted ablation and com + +
MMETextvqaMMBMMVetMMMUMathVOCRBMMStarAvgLatency
SV-LLaVA (w/o SK)151956.772.531.438.821.230.936.541.1103
SV-LLaVA151956.572.532.438.821.230.736.541.390
+0%-0.4%+0%+3.2%+0%+0%-0.6%+0%+0.5%-9.7%
SV-LLaVA-HD (w/o SK)153361.673.035.839.921.234.539.743.7310
SV-LLaVA-HD153461.273.135.939.721.234.139.343.5250
+0%-0.7%+0.1%+0.3%-0.5%+0%-1.2%-1%-0.5%-19.4%
SV-LLaVA-CoS (w/o SK)156263.976.540.240.221.237.241.945.9126
SV-LLaVA-CoS156363.776.541.740.321.23741.946.090
+0%-0.3%+0%+3.7%+0.2%+0%-0.5%+0%+0.2%-28.6%
+ +![](images/20bd027391091a7b44c86e87db03dc371646ae53d0deea8bdf5c02fd6b73d14e.jpg) +Table 4. Ablation on skip KV-cache. We evaluate SK (using skip KV-cache during inference) of Skip-Vision on different settings. +(a) Ablation on LLaVA + +![](images/892870d1a207c1e5fed03929e6ea26e4f6a68289766f5d21b24157546f59af90.jpg) +(b) Ablation on CoS + +![](images/b4ea6c94363d42f2866abbfcffe002a70241e63763992acd3b944562eeeb51eb.jpg) +(c) Ablation on LLaVA-HD + +![](images/9baf878fc339ad47abe3903f68fb1225de82816dfc51f51f9eaa5315a5efca38.jpg) +Figure 6. Ablation study. We performed an ablation study on each component of the Skip-Vision. SF (skip FFN), TM (token merge). + +![](images/28caf95df1a05c1466a032b57792f37bb6d1c1d30afce73eb7aad7557cb9b2e2.jpg) +Figure 7. Visualization of attention map in MMVet. +Figure 8. Visualization of attention map in TextVQA. + +parative experiments on token merge, skip FFN, and skip KV-cache. + +As shown in Figure 6, both the skip FFN strategy and token merging enhance computational efficiency across different model settings while maintaining performance. Notably, skip FFN performs better in the CoS setting than in LLaVA and LLaVA-HD. We attribute this to CoS allowing the vision backbone to be fine-tuned during the SFT stage, enabling key information to be more effectively captured by + +summary tokens, thereby reducing the errors introduced by skip FFN. As shown in Table 4, skip KV-cache slightly affects fine-grained tasks such as TextVQA and OCRBench but improves performance in long-form response tasks like MMVet. + +To further analyze this, we visualized attention maps across layers for visual tokens in SV-CoS during the MMVet and TextVQA tasks. In MMVet (Figure 7), the attention of generated tokens primarily focuses on retained and summary tokens, with skip KV-cache filtering out unnecessary noise. In TextVQA (Figure 8), which requires fine-grained understanding, attention is spread across both retained and skipped tokens, leading to some performance loss with skip KV-cache. However, causal attention effectively aggregates information from skipped tokens into the final summary tokens, allowing us to retain only the summary tokens and retained visual tokens in the KV-cache. + +# 7. Conclusion + +We propose Skip-Vision, an efficient multimodal large language model architecture that boosts computational efficiency through a skip-FFN strategy, reducing redundant computations over visual tokens during training. During inference, Skip-Vision uses a skip KV-cache to accelerate processing by omitting non-essential visual tokens from the KV cache. Extensive experiments show our model outperforms current methods in efficiency and scales effectively with data, delivering competitive performance among leading models of similar scale. + +# 8. Acknowledgements + +This work was supported in part by NSFC (62201342), and Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102). This work was supported by Ant Group Research Intern Program. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 5 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 3 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 23716-23736, 2022. 1, 3 +[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 5 +[5] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 5 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3 +[7] Samyadeep Basu, Martin Grayson, Cecily Morrison, Besmira Nushi, Soheil Feizi, and Daniela Massiceti. Understanding information storage and transfer in multi-modal large language models. arXiv preprint arXiv:2406.04236, 2024. 2 +[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5 +[9] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022.5 +[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[11] Shihao Cai, Keqin Bao, Hangyu Guo, Jizhi Zhang, Jun Song, and Bo Zheng. Geogpt4v: Towards geometric multi- + +modal large language models with geometric image generation. arXiv preprint arXiv:2406.11503, 2024. 5 +[12] Jianjian Cao, Peng Ye, Shengze Li, Chong Yu, Yansong Tang, Jiwen Lu, and Tao Chen. Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15710-15719, 2024. 3 +[13] Jimmy Carter. Textocr-gpt4v. https://huggingface.co/datasets/jimmycarter/textocr-gpt4v,2024.5 +[14] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 3 +[15] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 2 +[16] Chungkwong Chan. Stroke extraction for offline handwritten mathematical expression recognition. IEEE Access, 8: 61565-61575, 2020. 5 +[17] Shuaichen Chang, David Palzer, Jialin Li, Eric Fosler-Lussier, and Ningchuan Xiao. Mapqa: A dataset for question answering on choropleth maps. arXiv preprint arXiv:2211.08545, 2022. 5 +[18] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024. 5 +[19] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3 +[20] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5 +[21] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.6 +[22] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 6, 7 +[23] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5 +[24] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng + +Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 3 +[25] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/, 2023.1.3 +[26] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaiqiao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1571–1576. IEEE, 2019. 5 +[27] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 3 +[28] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3 +[29] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6 +[30] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2024. 6 +[31] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023.5 +[32] Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767, 2023.5 +[33] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2315-2324, 2016. 5 +[34] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939–948, 2019. 5 +[35] Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Zero-shot extreme + +length generalization for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3991-4008, 2024. 3 +[36] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In 2018 24th international conference on pattern recognition (ICPR), pages 7-12. IEEE, 2018. 5 +[37] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.5 +[38] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024.5 +[39] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2, 7 +[40] Ziyuan Huang, Kaixiang Ji, Biao Gong, Zhiwu Qing, Qinglong Zhang, Kecheng Zheng, Jian Wang, Jingdong Chen, and Ming Yang. Accelerating pre-training of multimodal llms via chain-of-sight. arXiv preprint arXiv:2407.15819, 2024. 1, 7 +[41] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3 +[42] Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024. 1, 3 +[43] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 5 +[44] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 5 +[45] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figurea: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 5 +[46] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 5 + +[47] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 3 +[48] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5 +[49] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5 +[50] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern recognition, pages 4999-5007, 2017. 5 +[51] Sunjun Kweon, Yeonsu Kwon, Seonhee Cho, Yohan Jo, and Edward Choi. Open-wikitable: Dataset for open domain question answering with complex reasoning over table. arXiv preprint arXiv:2305.07288, 2023. 5 +[52] Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1):1-10, 2018. 5 +[53] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 5 +[54] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: mitigating multimodal hallucination through self-feedback guided revision. arXiv preprint arXiv:2311.07362, 2023. 5 +[55] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 5 +[56] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild. https://llava-v1.github.io/blog/2024-05-10-llava-last-stronger-llms/, 2024.1,3,7 +[57] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5 +[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[59] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 5 +[60] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 7 +[61] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 7 +[62] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 5 +[63] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 1 +[64] Dongyang Liu, Renrui Zhang, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, et al. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models. arXiv preprint arXiv:2402.05935, 2024. 1, 3 +[65] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 5 +[66] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023.5 +[67] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 7 +[68] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.2 +[69] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7, 5 + +[70] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023.6 +[71] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 5 +[72] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6 +[73] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Icdar 2023 competition on hierarchical text detection and recognition. arXiv preprint arXiv:2305.09750, 2023. 5 +[74] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021.5 +[75] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5 +[76] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6 +[77] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 7 +[78] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5 +[79] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5 +[80] Minesh Mathew, Lluis Gomez, Dimosthenis Karatzas, and CV Jawahar. Asking questions on handwritten document collections. International Journal on Document Analysis and Recognition (IJDAR), 24(3):235-249, 2021. 5 +[81] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5 +[82] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. + +In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5 +[83] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022. 5 +[84] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 5 +[85] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5 +[86] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6 +[87] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia, 2018. Association for Computational Linguistics. 5 +[88] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 3 +[89] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 5 +[90] Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar. Kvqa: Knowledge-aware visual question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8876-8884, 2019. 5 +[91] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3, 2 +[92] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 5 +[93] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024.5 +[94] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. 2020. 5 + +[95] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6, 5 +[96] Dominik Stammbach and Elliott Ash. Docscan: Unsupervised text classification via learning from neighbors. arXiv preprint arXiv:2105.04024, 2021. 5 +[97] Hamed Rahimi Sujet AI, Allaa Boutaleb. Sujet-finance-qavision-100k: A large-scale dataset for financial document vqa, 2024. 5 +[98] S Svetlichnaya. Deepform: Understand structured documents at scale. 2020. 5 +[99] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 5 +[100] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3, 7 +[101] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[102] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016.5 +[103] Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in Neural Information Processing Systems, 31, 2018. 2 +[104] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[105] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 5 +[106] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Conghui He, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation, 2024. 5 +[107] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jian-Yuan Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. ArXiv, abs/2312.06109, 2023. 3 +[108] Yuxin Wen, Qingqing Cao, Qichen Fu, Sachin Mehta, and Mahyar Najibi. Efficient vision-language models by + +summarizing visual tokens into compact registers. arXiv preprint arXiv:2410.14072, 2024. 2 +[109] C. Wendler. Renderedtext. https://huggingface.co/datasets/wendlerc/RenderedText, 2023.5 +[110] Jianshu Chen Yunkai Zhang Hong Wang Shiyang Li Xiyou Zhou Wenhu Chen, Hongmin Wang and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 2020. 5 +[111] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2 +[112] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 1, 3 +[113] Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts. arXiv preprint arXiv:2312.15915, 2023. 5 +[114] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 3 +[115] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 4, 2 +[116] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2 +[117] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 7 +[118] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 7 +[119] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 2 +[120] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthyness. arXiv preprint arXiv:2405.17220, 2024. 5 +[121] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. + +Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.6 +[122] Xinyan Velocity Yu, Sewon Min, Luke Zettlemoyer, and Hannaneh Hajishirzi. Crepe: Open-domain question answering with false presuppositions. arXiv preprint arXiv:2211.17257, 2022. 5 +[123] Qianhao Yuan, Qingyu Zhang, Yanjiang Liu, Jiawei Chen, Yaojie Lu, Hongyu Lin, Jia Zheng, Xianpei Han, and Le Sun. Shortv: Efficient multimodal large language models by freezing visual tokens in ineffective layers. arXiv preprint arXiv:2504.00502, 2025. 2 +[124] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 3 +[125] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6 +[126] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5317-5327, 2019. 5 +[127] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 2 +[128] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 5 +[129] Ying Zhang, Lionel Gueguen, Ilya Zharkov, Peter Zhang, Keith Seifert, and Ben Kadlec. Uber-text: A large-scale dataset for optical character recognition from street-level imagery. In SUNw: Scene Understanding Workshop-CVPR, page 5, 2017. 5 +[130] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2 +[131] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3 +[132] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. 5 + +[133] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 3 +[134] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[135] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. 5 \ No newline at end of file diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/images.zip b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6b336c97645c4f9b16a40e3390b0fa47abc3a43f --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a009905d52eaa073caec59d96b9756caf0a86b6f60b62fab0eed16ecab5e432 +size 792026 diff --git a/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/layout.json b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..834fa1d9c720af47d06d52d91bf9e9ffce3ac7bf --- /dev/null +++ b/2025/Skip-Vision_ Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping/layout.json @@ -0,0 +1,12371 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "type": "text", + "content": "Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "spans": [ + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "text", + "content": "Weili Zeng" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "text", + "content": ", Ziyuan Huang" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "text", + "content": ", Kaixiang Ji" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "text", + "content": ", Yichao Yan" + }, + { + "bbox": [ + 164, + 166, + 445, + 181 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 181, + 542, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 181, + 542, + 197 + ], + "spans": [ + { + "bbox": [ + 67, + 181, + 542, + 197 + ], + "type": "text", + "content": "1 MoE Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2 Ant Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 198, + 523, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 198, + 523, + 210 + ], + "spans": [ + { + "bbox": [ + 84, + 198, + 523, + 210 + ], + "type": "text", + "content": "{zw1666,yanyichao}@sjtu.edu.cn, ziyuan.huang@u.nus.edu, kaixiang.jkx@antgroup.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 236, + 200, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 236, + 200, + 250 + ], + "spans": [ + { + "bbox": [ + 151, + 236, + 200, + 250 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "text", + "content": "Transformer-based models have driven significant advancements in Multimodal Large Language Models (MLLMs), yet their computational costs surge drastically when scaling resolution, training data, and model parameters. A key bottleneck stems from the proliferation of visual tokens required for fine-grained image understanding. We propose Skip-Vision, a unified framework addressing both training and inference inefficiencies in vision-language models. On top of conventional token compression approaches, our method introduces two complementary acceleration strategies. For training acceleration, we observe that Feed-Forward Network (FFN) computations on visual tokens induce marginal feature updates. This motivates our Skip-FFN strategy, which bypasses FFN layers for redundant visual tokens. For inference acceleration, we design a selective KV-cache removal mechanism that prunes the skipped key-value pairs during decoding while preserving model performance. Experimental results demonstrate that SkipVision reduces training time by up to " + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "text", + "content": ", inference FLOPs by " + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "text", + "content": ", and latency by " + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 54, + 262, + 297, + 552 + ], + "type": "text", + "content": ", while achieving comparable or superior performance to existing methods. Our work provides a practical solution for scaling high-performance MLLMs with enhanced efficiency. Our project page can be found here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 682 + ], + "type": "text", + "content": "Transformer-based large-scale models have significantly advanced progress in Artificial Intelligence (AI) [10, 86, 101]. Their remarkable capabilities have driven the emergence of modern Multimodal Large Language Models (MLLMs) [2, 3, 69, 88], which demonstrate vision-language competencies comparable to human performance. Empirical evidence suggests that scaling laws continue to" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 236, + 552, + 441 + ], + "blocks": [ + { + "bbox": [ + 320, + 236, + 552, + 441 + ], + "lines": [ + { + "bbox": [ + 320, + 236, + 552, + 441 + ], + "spans": [ + { + "bbox": [ + 320, + 236, + 552, + 441 + ], + "type": "image", + "image_path": "854eda7a7f731484f8da78f37fae6a357bb004a72bcf968b8248987b0bd9c0f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 455, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 555 + ], + "type": "text", + "content": "Figure 1. Performance-efficiency trade-off curve. Each circle denotes a model configuration, where our models utilize the Skip-Vision framework, with CoS [40], LLaVA-HD [67] and LLaVA [69] serving as baselines. Circle sizes reflect the inference FLOPs ratio. Skip-Vision demonstrate superior performance, scaling effectively with increased FLOPs and data and achieving higher inference efficiency when compared to baselines and other efficient MLLM methods. All methods utilize LLaMA3 8B as the foundational large language model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "content": "be effective, primarily across three dimensions: visual scaling [40, 67, 112], data scaling [63, 64], and model scaling [56, 104]. However, these scaling methods present substantial efficiency issues, significantly increasing the training and inference burden. In a typical LLaVA-1.5 framework [69], where the model receives 576 visual tokens and 40 text tokens, inference using Vicuna-13B [25] LLM backbone demands 200 trillion FLOPS [42]. These motivate our pursuit of an MLLM architecture that optimally balances efficiency and performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Recent advances in Multimodal Large Language Models (MLLMs), such as Chain-of-Sight (CoS)[40] and LLaVA-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 271, + 703 + ], + "type": "text", + "content": "*This work was done when Weili Zeng interned at Ant Group." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 703, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 146, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21384" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 286, + 122 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 286, + 122 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 286, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 286, + 122 + ], + "type": "table", + "html": "
Llama layerComplexityParameters
FFN10.5LNC25.64B(70%)
Attention2LN2C + 1.5NC21.34B(17%)
Embedding-0.5B(6%)
", + "image_path": "64ce9bea74e44b023b4a2e1bda2858c8d7e02926213e779c6c78d6ea8aefb695.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 163 + ], + "type": "text", + "content": "Table 1. Computational complexity and parameter statistics for Llama3 8B. L denotes the number of blocks, N represents the number of input tokens, and C is the feature dimensionality" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "text", + "content": "Next[68], have improved performance by increasing visual tokens. However, during the Supervised Fine-Tuning (SFT) stage, the computational cost rises dramatically, leading to diminishing returns and making the scaling unsustainable. As shown in Figure 1, CoS requires a " + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "inline_equation", + "content": "165\\%" + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "text", + "content": " increase in training resources and a " + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "inline_equation", + "content": "400\\%" + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "text", + "content": " increase in inference computation for a mere " + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 54, + 181, + 295, + 373 + ], + "type": "text", + "content": " performance improvement, while LLaVA-Next faces similar inefficiencies. This challenge emphasizes the need to reduce visual token computation. Approaches like token compression or pruning [15, 22, 39, 60, 116, 119], while efficient, risk losing critical visual information, revealing a trade-off between efficiency and performance. This work aims to optimize the model's computational efficiency from an architectural perspective, proposing a scalable solution that balances efficiency with visual fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 376, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 674 + ], + "type": "text", + "content": "We systematically examine the parameter count and FLOPs distribution across different components of the LLM. As shown in Table 1, using LLaMA-3 [29] as an example, the parameter and FLOPS requirements for the FFN are significantly higher than the attention module. Furthermore, as illustrated in Figure 2, when the number of visual tokens is large, the majority of computations are concentrated on FFN of these tokens. In particular, we computed the difference in the magnitudes of the features before and after the FFN transformation for different tokens. As illustrated in the figure 3, we observe that the updates introduced by the FFN to the visual tokens are significantly smaller than those that affect the text tokens. This is similar to the observation in [123]. This observation suggests a potential approach: could we bypass FFN computations for a substantial portion of visual tokens? Furthermore, in the autoregressive training framework of MLLMs, only the final visual token utilizes its last hidden state for next-token prediction, while the other visual tokens primarily serve to convey causal information. During model inference, inspired by [7], we observe that most visual information is progressively integrated into the final token through the early causal layers. This raises the possibility of reducing reliance on the full visual scaling KV cache throughout inference, potentially accelerating the inference process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Building on these insights, we propose Skip-Vision, a novel and flexible architecture designed for more efficient scaling (see Figure 1 for a comparison). Visual tokens de" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 555, + 233 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 555, + 233 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 555, + 233 + ], + "type": "image", + "image_path": "8be379eb26551a301c26d0ca932d87f74a22843ebb31fc8e68dbdd11bc04f8eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 241, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 286 + ], + "type": "text", + "content": "Figure 2. Flops ratio of visual tokens in FFN. When visual tokens greatly outnumber text tokens, computation is predominantly consumed by the FFN of visual tokens. Reducing their passage through the FFN can thus markedly decrease computational costs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "content": "rived from smaller window/scale sizes constitute approximately " + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "content": " of the total visual tokens and account for most of the computational overhead. To mitigate this, as shown in Figure 4, we allow these tokens to skip the Feed-Forward Network (FFN) computation, significantly reducing overall computational costs. To address redundancy within skipped vision tokens, we can apply a token merging strategy for further simplification. During inference, skipped tokens are removed from the KV cache after the pre-fill phase, accelerating subsequent computation. Empirical results demonstrate that Skip-Vision reduces training time by up to " + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "content": ", FLOPs of inference by " + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "content": " and inference latency by " + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 302, + 555, + 495 + ], + "type": "text", + "content": ", while maintaining or even improving performance. These efficiency gains provide a scalable solution for large-scale multimodal learning, enabling better data utilization and enhanced model scalability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": "In summary, Skip-Vision is designed to be seamlessly integrated into the standard SFT pipeline of MLLMs without introducing additional retraining or decoupled modules. It directly modifies the transformer's computation flow, offering a practical and theoretically grounded acceleration solution for MLLM training and inference jointly. Our contributions are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 582, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 616 + ], + "type": "text", + "content": "- We introduce a novel and efficient framework, using token merge and skip FFN strategy during training to reduce redundant computation for visual tokens." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 653 + ], + "type": "text", + "content": "- In inference, our framework employs a skip KV-cache mechanism that removes skip-FFN visual tokens from the KV-cache, enhancing efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "- We present a theoretical analysis of the performance error in Skip-Vision and establish an error bound, which will be empirically validated through experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Experiments show our model's superior efficiency, effective data scaling, and performance on par with state-of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21385" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 74, + 227, + 180 + ], + "blocks": [ + { + "bbox": [ + 81, + 74, + 227, + 180 + ], + "lines": [ + { + "bbox": [ + 81, + 74, + 227, + 180 + ], + "spans": [ + { + "bbox": [ + 81, + 74, + 227, + 180 + ], + "type": "image", + "image_path": "90b52cde8ad17b80416bc12f56fd3c53c4709de6a0bbfeb3e77d9f6d0defaf22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 186, + 195, + 196 + ], + "lines": [ + { + "bbox": [ + 113, + 186, + 195, + 196 + ], + "spans": [ + { + "bbox": [ + 113, + 186, + 195, + 196 + ], + "type": "text", + "content": "(a) FFN updates for CoS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 72, + 380, + 180 + ], + "blocks": [ + { + "bbox": [ + 231, + 72, + 380, + 180 + ], + "lines": [ + { + "bbox": [ + 231, + 72, + 380, + 180 + ], + "spans": [ + { + "bbox": [ + 231, + 72, + 380, + 180 + ], + "type": "image", + "image_path": "3043d0a625953e2bf208b09527d18729cad7c27bec26f8071729072df4e4959f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 186, + 358, + 196 + ], + "lines": [ + { + "bbox": [ + 252, + 186, + 358, + 196 + ], + "spans": [ + { + "bbox": [ + 252, + 186, + 358, + 196 + ], + "type": "text", + "content": "(b) FFN updates for LLaVA-HD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "text", + "content": "Figure 3. FFN imapct for different MLLMs. We evaluate FFN impact by computing the feature modulus ratio before " + }, + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "inline_equation", + "content": "(\\| h_{\\mathrm{attn}}\\|_2)" + }, + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "text", + "content": " and after FFN " + }, + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "inline_equation", + "content": "(\\| \\mathrm{FFN}(h_{\\mathrm{attn}})\\|_2)" + }, + { + "bbox": [ + 55, + 206, + 555, + 229 + ], + "type": "text", + "content": ". The FFN introduces significantly smaller updates to visual tokens compared to text tokens." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 72, + 533, + 180 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 533, + 180 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 533, + 180 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 533, + 180 + ], + "type": "image", + "image_path": "a1f9d01c36c74b7a0bef88cd90a6a14d09d9558f6e27fcb01b5e6b2402bd5f80.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 186, + 501, + 196 + ], + "lines": [ + { + "bbox": [ + 410, + 186, + 501, + 196 + ], + "spans": [ + { + "bbox": [ + 410, + 186, + 501, + 196 + ], + "type": "text", + "content": "(c) FFN updates for LLaV" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 241, + 190, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 241, + 190, + 252 + ], + "spans": [ + { + "bbox": [ + 64, + 241, + 190, + 252 + ], + "type": "text", + "content": "the-art models of similar scale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 278, + 139, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 139, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 139, + 290 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 303, + 190, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 190, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 190, + 316 + ], + "type": "text", + "content": "2.1. Vison-Language models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 326, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 469 + ], + "type": "text", + "content": "Rapid progress in large language models (LLMs) [25, 29, 114] has laid a solid foundation for the emergence of large-scale vision-language models (VLMs) [3, 6, 19, 58]. Flamingo [3] was a pioneering effort, integrating pretrained image encoders with LLMs to handle multimodal data, thereby allowing comprehensive understanding and reasoning across modalities. Following this, the major models developed by leading corporations, such as GPT-4v [2] and Gemini-1.5-Pro [88], have led the progress in this domain. Using proprietary data sets and undisclosed training methodologies, these models have elevated multimodal intelligence to unprecedented levels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 475, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 713 + ], + "type": "text", + "content": "At the same time, the open-source community has worked diligently to stay abreast of these advancements. The LLava series models [56, 67, 69, 112] exemplify a key strategy in this domain by mapping visual features directly into the input embedding space of the language model, integrating them as input tokens. However, this approach generates a substantial volume of visual tokens, frequently exceeding the number of text tokens, creating efficiency challenges. InternVL-1.5 [24] adopts dynamic high-resolution techniques, segmenting large images into smaller fragments for processing and scaling them accordingly. MiniCPMV [117] utilizes a perceiver resampler structure with a single layer of cross-attention to compress visual tokens. Models such as Vary [107], SPHINX [64], Cambrian-1 [100], and BRAVE [47] leverage multi-visual encoder architectures to strengthen their visual processing capabilities. Furthermore, a crucial factor contributing to the effectiveness of modern VLMs is instruction fine-tuning [28, 67, 134], enabling these models to function as conversational agents and interact with users through natural, human-like dialogue." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 241, + 540, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 540, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 540, + 253 + ], + "type": "text", + "content": "2.2. Efficient multimodal large language models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 510 + ], + "type": "text", + "content": "To enhance the efficiency of vision-language models, one of the primary and widely employed approaches in contemporary research involves reducing the size of the backbone models [27, 124, 133?]. Simultaneously, given that visual tokens contribute significantly to computational demands, research focused on compressing or refining these tokens has garnered increasing attention. Such methods are often coupled with the design of vision-language projectors. For instance, techniques like the Perceiver Resampler [3, 6, 41] and Q-Former [58] utilize transformer-based mechanisms to consolidate visual tokens into more compact query sets. In this way, the transformer outputs corresponding to the positions of the learnable latent queries serve as the aggregated representation of the visual features [42]. Honeybee [14] propose two visual projectors, namely C-Abstractor and D-Abstractor. LLaVA-PruMerge [91] and MADTP [12] introduce adaptive methods for reducing visual tokens. In addition, while the large language model (LLM) field has employed various techniques for token reduction to accelerate inference and compress KV cache [35, 131], research in this area remains limited within the MLLM domain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 399, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 399, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 399, + 533 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 541, + 427, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 427, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 427, + 554 + ], + "type": "text", + "content": "3.1. Decoder-only LLM" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 559, + 554, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 619 + ], + "type": "text", + "content": "Each Transformer layer in a decoder-only LLM primarily consists of a self-attention layer and a feed-forward neural network. To facilitate autoregressive inference, the use of a KV-cache has been developed to accelerate the decoding process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "text", + "content": "Self-Attention Mechanism: The self-attention mechanism enables each token in a sequence to selectively attend to other tokens, capturing contextual dependencies. Each token generates a query " + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 313, + 620, + 554, + 680 + ], + "type": "text", + "content": " vector. Attention scores are calculated by the scaled dot-product:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 689, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 689, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 340, + 689, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}", + "image_path": "3519f45f2a79342adb6bcc96ea7a108cff273ab1c31a460e06a9edb84c681ffb.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21386" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 304 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 304 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 304 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 304 + ], + "type": "image", + "image_path": "df2c8b74ad126f2c273114dbae5e81ef25bd7e6e89b6c1b8ffa8b57f2ec80f5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "type": "text", + "content": "Figure 4. The framework of Skip-Vision. a) While visual scaling enriches visual information, it also increases computational overhead. Skip-Vision uses a skip-FFN strategy during training to reduce redundant computation for visual tokens. The numerous skipped tokens will be limited to the attention layer and bypass FFN. b) At the beginning of inference, Skip-Vision will remove skip-FFN visual tokens from the initial KV-cache, enhancing efficiency. c) During inference, skip attention leverages the skip KV-cache to accelerate generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 373, + 295, + 420 + ], + "type": "text", + "content": ". This formulation allows each token to incorporate information from others based on relevance, making self-attention critical in capturing relationships across tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 456 + ], + "type": "text", + "content": "Feed-Forward Neural Network: The feed-forward neural network enhances token representations following the self-attention layer. The operation can be expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 465, + 295, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 465, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 82, + 465, + 295, + 479 + ], + "type": "interline_equation", + "content": "\\operatorname {F F N} (x) = \\text {A c t i v a t i o n} \\left(x W _ {1} + b _ {1}\\right) W _ {2} + b _ {2}, \\tag {2}", + "image_path": "dd3169c318bede2a0a2c03eb6ec2a6612e9d97b2cddc2262c64f667db9bb7227.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": " is the input, " + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": " are weight matrices, and " + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "inline_equation", + "content": "b_{1}" + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "inline_equation", + "content": "b_{2}" + }, + { + "bbox": [ + 55, + 487, + 296, + 534 + ], + "type": "text", + "content": " are biases. This architecture allows the model to capture complex patterns and improve the expressiveness of token representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "text", + "content": "KV-cache: The key-value cache (KV-cache) stores key " + }, + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 535, + 296, + 594 + ], + "type": "text", + "content": " representations from previous self-attention computations, enabling efficient reuse during inference. This approach avoids redundant calculations for earlier tokens, improving inference speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "text", + "content": "The attention score is computed using Attention " + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "inline_equation", + "content": "(Q, K_{\\text{cache}}^t, V_{\\text{cache}}^t)" + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "text", + "content": " is the query for the current token. At each time step " + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 595, + 296, + 642 + ], + "type": "text", + "content": ", the cache is updated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 650, + 295, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 650, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 66, + 650, + 295, + 671 + ], + "type": "interline_equation", + "content": "K _ {\\text {c a c h e}} ^ {(t)} = \\left[ K _ {\\text {c a c h e}} ^ {(t - 1)}; K _ {t} \\right], \\quad V _ {\\text {c a c h e}} ^ {(t)} = \\left[ V _ {\\text {c a c h e}} ^ {(t - 1)}; V _ {t} \\right], \\tag {3}", + "image_path": "ce2867d6923ab001d2ed5918fa9e66daeb2fb10f7e55b70d94f2a77cda22028d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " are the current token's key and value. This mechanism enhances computational efficiency and accelerates token generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 371, + 370, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 370, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 370, + 384 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 528, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 528, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 528, + 406 + ], + "type": "text", + "content": "4.1. Skipped token selection and token merge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 471 + ], + "type": "text", + "content": "Before applying the skip FFN strategy, we first identify the skipped and retained visual tokens. The retained tokens should preserve most essential visual information, while the skipped tokens serve as complementary details, ensuring efficiency without significant loss of information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "text", + "content": "Fortunately, in MLLM architectures, we can naturally distinguish retained and skipped tokens based on global and local context tokens. Global context tokens are retained, while local context tokens are skipped. In architectures like LLaVA-HD, global tokens originate from the resized full image, whereas local tokens come from smaller image patches. In CoS, global tokens correspond to larger window sizes, while local tokens are derived from smaller windows. Since LLaVA lacks a clear distinction between global and local tokens, inspired by [115], we select the top-" + }, + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "text", + "content": " tokens most similar to the CLS token as retained tokens, categorizing the rest as skipped tokens." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "To further reduce the computational burden, it is necessary to refine the visual tokens to eliminate excessive redundancy. As shown in Figure 5, we conducted a similarity analysis of the visual tokens in CoS at different levels, calculating their similarity density. We found that the smaller the granularity of the visual tokens, the higher their average similarity, while their token count is significantly greater than that of other granularities. Consequently, inspired by" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21387" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "ToMe [9], we employ a token merge method to reduce redundancy in skipped tokens, as detailed below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 216 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "type": "text", + "content": "- Compute cosine similarities: calculate the cosine similarity between each pair of tokens in the set, obtaining an average similarity score for each token." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "content": "- Select top " + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "content": " distinctive Tokens: sort tokens by average similarity in ascending order and retain the top " + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 133, + 295, + 178 + ], + "type": "text", + "content": " tokens with the lowest average similarity, representing the most distinctive information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "text", + "content": "- Merge remaining tokens: for the remaining " + }, + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "inline_equation", + "content": "n - k" + }, + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "text", + "content": " tokens, merge each one with the most similar token among the top " + }, + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 180, + 295, + 216 + ], + "type": "text", + "content": ", thus consolidating information efficiently." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 221, + 214, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 214, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 214, + 234 + ], + "type": "text", + "content": "4.2. Training speed-up: skip FFN" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "content": "As depicted in Figure 4a, we propose the FFN skip strategy: The retained visual tokens, which are few in number, proceed conventionally through all decoder layers of the LLM, while the numerous skipped visual tokens will be limited to the self-attention layers of each transformer block. For a given layer " + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "content": ", the output of retained tokens " + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{retained}}^{(l)}" + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "content": " is the sum of the self-attention output " + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{attn}}^{(l)}" + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "content": " and the FFN-transformed features " + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)})" + }, + { + "bbox": [ + 55, + 239, + 296, + 339 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 344, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 344, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 110, + 344, + 295, + 360 + ], + "type": "interline_equation", + "content": "h _ {\\text {r e t a i n e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)} + \\operatorname {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right). \\tag {4}", + "image_path": "d28780d4e71f2fb5d676a27fd8d69b320a21a3cd5bad0bd0e238da7c1f56bb98.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 365, + 202, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 202, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 202, + 376 + ], + "type": "text", + "content": "The output of the skipped tokens is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 381, + 294, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 381, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 143, + 381, + 294, + 399 + ], + "type": "interline_equation", + "content": "h _ {\\text {s k i p p e d}} ^ {(l)} = h _ {\\text {a t t n}} ^ {(l)}. \\tag {5}", + "image_path": "1208e1845741dcfa4b0763841bedc2f58b34a91bc97e16f9da592be212862b99.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 403, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 498 + ], + "type": "text", + "content": "However, bypassing the FFN layers presents a nontrivial challenge. In the autoregressive training process, the final token of the vision sequence is used to predict the initial token of the text sequence, necessitating a specialized design for this terminal vision token. To facilitate this transition effectively, it must pass through the FFN layers, which are enriched with stored linguistic information [32, 83]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 569 + ], + "type": "text", + "content": "To address this, as shown in Figure 4, we introduce adaptive summary tokens to efficiently consolidate visual information. This token is generated by the Adaptive Summary Layer before the token sequence is input into the LLM. The layer comprises a simple linear transformation, and its computation is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 574, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 574, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 107, + 574, + 295, + 590 + ], + "type": "interline_equation", + "content": "x ^ {s} = \\left(\\operatorname {s o f t m a x} \\left(X \\cdot W ^ {T}\\right)\\right) ^ {T} \\cdot X, \\tag {6}", + "image_path": "a24e058dcf930847c4e69b00e4105bb86388faa98c863f6a9acabd6a9ba9b843.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "text", + "content": " denotes the retained or skipped visual tokens, " + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "inline_equation", + "content": "x^{s} \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "text", + "content": " is the summary token, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{1 \\times C}" + }, + { + "bbox": [ + 55, + 594, + 296, + 701 + ], + "type": "text", + "content": " is the weight matrix. Leveraging this layer, we concatenate a summary token on both ends of the skipped visual tokens before their input to the LLM. The first summary token aggregates essential information from retained tokens, mitigating potential loss from extended token sequences. The second summary token integrates prior sequence information and plays a key role in predicting the next text token." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 542, + 209 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 542, + 209 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 542, + 209 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 542, + 209 + ], + "type": "image", + "image_path": "0c96dbe6580415cba281479de909b3324509d8ed8de123e3a955dd1de121f6a7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 274 + ], + "type": "text", + "content": "Figure 5. Similarity density statistics. We conducted a similarity density analysis in CoS, examining the visual tokens derived from various window scale. The results indicate that as the window scale decreases, the similarity between visual tokens becomes more pronounced, reflecting higher token redundancy." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 284, + 501, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 501, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 501, + 297 + ], + "type": "text", + "content": "4.3. Inference speed-up: skip KV-cache" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 301, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 326 + ], + "type": "text", + "content": "After training, our Skip-Vision framework incorporates skip KV-cache method to further improve efficiency:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 335, + 553, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 553, + 362 + ], + "type": "interline_equation", + "content": "K _ {\\text {s k i p - c a c h e}} ^ {(0)} = K _ {\\text {c a c h e}} ^ {(0)} \\backslash K _ {\\text {s k i p}}, \\quad V _ {\\text {s k i p - c a c h e}} ^ {(0)} = V _ {\\text {c a c h e}} ^ {(0)} \\backslash V _ {\\text {s k i p}}, \\tag {7}", + "image_path": "efacf266d9a3db8266855265577d57e38fe5de5ebe71db72517656b02c9e9814.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": "Equation 7 indicates that visual tokens skipped by FFN are removed from the KV cache after the pre-filling stage. As depicted in Figures 4 b) and c), given the autoregressive nature of LLMs, earlier token information is progressively transferred to later tokens. Skip FFN further enhances this integration by suppressing the attention weights of skipped tokens, forcing the model to focus on key tokens and improving information aggregation efficiency. Empirical results show that, due to the inclusion of summary tokens, this omission does not compromise model performance, underscoring a distinctive advantage of our framework for more efficient inference." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 518, + 507, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 507, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 507, + 533 + ], + "type": "text", + "content": "5. Theoretical analysis of Skip-Vision" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "content": "At the core of this analysis lies the layer error incurred when bypassing the FFN layer. For a given layer " + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "content": ", the original output " + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{original}}^{(l)}" + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "content": " is: " + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{original}}^{(l)} = h_{\\mathrm{attn}}^{(l)} + \\mathrm{FFN}^{(l)}(h_{\\mathrm{attn}}^{(l)})" + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "content": ". When skipping the FFN, the output becomes: " + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{skip}}^{(l)} = h_{\\mathrm{attn}}^{(l)}" + }, + { + "bbox": [ + 313, + 539, + 554, + 605 + ], + "type": "text", + "content": ". The per-layer skipping error is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 341, + 614, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 614, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 341, + 614, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\epsilon^ {(l)} = \\left\\| h _ {\\text {o r i g i n a l}} ^ {(l)} - h _ {\\text {s k i p}} ^ {(l)} \\right\\| _ {2} = \\left\\| \\mathrm {F F N} ^ {(l)} \\left(h _ {\\text {a t t n}} ^ {(l)}\\right) \\right\\| _ {2}. \\tag {8}", + "image_path": "7ac872a31c00875f19097a53f2595a3f7e17773900ded6eaee0201774828cecd.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "content": "For redundant tokens, such as those in homogeneous image regions, this error is negligible (" + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon^{(l)} \\approx 0" + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "content": ") due to minimal feature transformations by the FFN." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "However, errors propagate through subsequent layers, amplified by the recursive nature of transformer architectures. Leveraging Lipschitz continuity assumptions for self" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21388" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "text", + "content": "attention and FFN operations " + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "inline_equation", + "content": "(L_{attn}^{(l + 1)}" + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "inline_equation", + "content": "L_{FFN}^{(l + 1)}" + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "text", + "content": "), the cumulative error at layer " + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 55, + 70, + 297, + 97 + ], + "type": "text", + "content": " is bounded by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 104, + 295, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 104, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 89, + 104, + 295, + 124 + ], + "type": "interline_equation", + "content": "\\epsilon^ {(l + 1)} \\leq \\left(L _ {\\mathrm {a t t n}} ^ {(l + 1)} + L _ {\\mathrm {F F N}} ^ {(l + 1)}\\right) \\cdot \\epsilon^ {(l)} + \\epsilon_ {\\mathrm {s k i p}} ^ {(l + 1)}, \\tag {9}", + "image_path": "75d02bc2daebeedd75cddab6c04adb3f2f321d1d99784ec76259b6c0a4389989.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\mathrm{skip}}^{(l + 1)})" + }, + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "text", + "content": " represents new errors from skipping deeper layers " + }, + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "inline_equation", + "content": "(l + 1)" + }, + { + "bbox": [ + 55, + 129, + 295, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 157, + 241, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 157, + 241, + 169 + ], + "spans": [ + { + "bbox": [ + 67, + 157, + 241, + 169 + ], + "type": "text", + "content": "Over " + }, + { + "bbox": [ + 67, + 157, + 241, + 169 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 67, + 157, + 241, + 169 + ], + "type": "text", + "content": " layers, the total error telescopes to:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 177, + 295, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 177, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 93, + 177, + 295, + 209 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\text {t o t a l}} \\leq \\sum_ {l = 1} ^ {L} \\epsilon_ {\\text {s k i p}} ^ {(l)} \\cdot \\prod_ {i = 1} ^ {L - l} \\left(L _ {\\text {a t t n}} ^ {(i + 1)} + L _ {\\text {F F N}} ^ {(i + 1)}\\right). \\tag {10}", + "image_path": "aeef97ee51c5ae295faf6b820bc20eade7a5538b8951bec818ef35d09dbe4fe7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "text", + "content": "Theorem 5.1 Bounded Lipschitz. If the spectral norms of " + }, + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "inline_equation", + "content": "W_{1}, W_{2}, W_{K}, W_{Q}" + }, + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 55, + 218, + 297, + 241 + ], + "type": "text", + "content": " are bounded by 1, then:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 242, + 137, + 265 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 242, + 137, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 137, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 137, + 254 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 55, + 242, + 137, + 254 + ], + "type": "inline_equation", + "content": "L(attn) \\leq 1 / \\sqrt{d_k}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 255, + 117, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 117, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 117, + 265 + ], + "type": "inline_equation", + "content": "L(FFN)\\leq 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": "Follow Theorem 5.1, we assume Lipschitz constants " + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{attn}} + L_{\\mathrm{FFN}} \\leq \\gamma" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": " and skipping errors " + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{skip}}^{(l)} \\leq \\epsilon" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": ", the total error is scaled to:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 135, + 310, + 295, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 310, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 135, + 310, + 295, + 335 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\text {t o t a l}} \\leq \\epsilon \\cdot \\frac {\\gamma^ {L} - 1}{\\gamma - 1}, \\tag {11}", + "image_path": "f54da8e1afca00884d3f8cd0fb5b26871147586bd137f7b33738ef5d44a55323.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 338, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 387 + ], + "type": "text", + "content": "Equation 11 establishes that the skip error is bounded when " + }, + { + "bbox": [ + 55, + 338, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\gamma < 1" + }, + { + "bbox": [ + 55, + 338, + 296, + 387 + ], + "type": "text", + "content": ", provided the model is trained with modern regularization techniques. This ensures that the MLLM remains less sensitive to the effects of skipping." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 411 + ], + "type": "text", + "content": "This error also impacts the KL divergence between the original and skipped outputs, bounded by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 418, + 295, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 418, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 104, + 418, + 295, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {K L}} \\left(p _ {\\text {s k i p}} \\| p _ {\\text {o r i g i n a l}}\\right) \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\epsilon_ {\\text {t o t a l}} ^ {2}, \\tag {12}", + "image_path": "ed5555423b429893f254e52e0e1d5273baa938d738b0e5b3285be641128a8431.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": " is the variance of the logits. Further integrating feature similarity errors " + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\mathrm{sim}} = O(\\sqrt{1 - \\theta}))" + }, + { + "bbox": [ + 55, + 449, + 296, + 485 + ], + "type": "text", + "content": " from low-attention tokens, the final bound becomes:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 492, + 295, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 492, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 115, + 492, + 295, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {K L}} \\leq \\frac {1}{2 \\sigma^ {2}} \\cdot \\left(\\epsilon_ {\\text {t o t a l}} + \\epsilon_ {\\text {s i m}}\\right) ^ {2}. \\tag {13}", + "image_path": "9e4d99d94c9f05298e85ffc8d1c95556298b84db574d5b0a3d1da60efaa5f7f0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "text", + "content": "Practically, this analysis motivates a layer-wise skipping strategy, alongside token selection and merge based on feature similarity " + }, + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "inline_equation", + "content": "(\\theta)" + }, + { + "bbox": [ + 55, + 522, + 296, + 571 + ], + "type": "text", + "content": ". More analysis and the proof of Theorem 5.1 are provided in the Appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 600, + 137, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 137, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 137, + 611 + ], + "type": "text", + "content": "6.1. Benchmarks" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "To comprehensively evaluate our model, we conducted experiments on eight different benchmarks. These include MME [30], which assesses the perceptual and cognitive capabilities of multimodal language models; TextVQA [95], focused on fine-grained visual question answering; MM-Bench [72] and MMStar [21] for general diagnostic capabilities; MMMU [125], designed to test STEM-related reasoning; MathVista [76], which evaluates mathematical problem" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "solving skills; OCRBench [70], for optical character recognition tasks; and MMVet [121], used for subjective assessment. Taken together, these benchmarks offer a comprehensive framework for assessing the efficacy of the model in various multimodal challenges." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 144, + 434, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 434, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 434, + 156 + ], + "type": "text", + "content": "6.2. Efficiency evaluation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": "We systematically evaluate both training time (hours), inference FLOP computation and inference latency (ms) on single A100 GPU. For training efficiency, we report actual GPU hours measured on consistent hardware configurations. For inference FLOPs, we adopt the methodology from FastV [22], specifically tracking the computational demands of processing visual tokens. In the LLama architecture, we calculate FLOPs for the causal attention and feed forward network modules as " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "4N C^{2} + 2N^{2}C + 3N C M" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " is the token count, " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " denotes the hidden state dimension, and " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " is the intermediate FFN dimension. SkipVision framework adjusts token counts dynamically between the attention and FFN modules, allowing us to refine FLOP calculations as " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "4N_{1}C^{2} + 2N_{1}^{2}C + 3N_{2}CM" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 313, + 162, + 555, + 354 + ], + "type": "text", + "content": " represent the visual token counts processed in the causal attention and FFN layers, respectively." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 365, + 388, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 388, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 388, + 378 + ], + "type": "text", + "content": "6.3. Data setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "type": "text", + "content": "During pretraining, we adopt the experimental settings of LLaVA and CoS, utilizing LLaVA-558K [69] and 65 million image-text pairs for respective comparisons. For the Supervised Fine-Tuning phase, we use LLaVA-665K [69] for all comparative experiments. By leveraging the computational efficiency of Skip-Vision framework, the resources saved can be redirected toward scaling the dataset to further enhance model performance. To this end, we extend our dataset to SV-1M, comprising 1 million samples, demonstrating the substantial performance improvements achievable through data expansion. To further assess our approach against state-of-the-art models of similar scale, we extend the dataset to SV-9M, underscoring the scalability and potential of our method. Detailed dataset statistics are provided in the Appendix." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 575, + 405, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 405, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 405, + 588 + ], + "type": "text", + "content": "6.4. Training setup" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "For LLaVA and LLaVA-HD setting, we use input resolutions of 336 and a dynamic resolution ranging from 336 to 672, respectively, with CLIP ViT-L/336px [86] as the vision encoder and LLaMA3 8B Instruct [29] as the backbone (additional comparisons under LLaVA-1.5-7B training setup are in the Appendix 10.1). For CoS setting, pretraining is conducted at 224 resolution, training only the multi-scale visual resampler with CLIP ViT-L/224px [86]. The model processes 80 visual tokens, including 16 global and 64 local tokens, and scales to 336 and 1296 tokens during SFT." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21389" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 302 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 302 + ], + "type": "table", + "html": "
MMETextvqaMMBMMVetMMMUMMathVOCRBMMStarAvgHoursFLOPs Latency
LLaVA setting:
LLaVA [69]153558.672.232.839.621.332.736.542.010.2100%125
LLaVA-FastV [22]145356.17132.437.320.628.836.340.49.630%103
MQT-LLaVA [39]148753.66928.237.816.828.136.538.67.544%106
LLaVA-Tokenpacker [60]153856.272.332.139.019.329.835.040.58.824%97
SV-LLaVA151956.572.532.438.821.230.736.541.37.925%90
(Nr=100,Ns=156)-1%-3.6%+0.4%-1.2%-2%-0.5%-3.6%-6.1%-1.7%-22.5%-75%-28%
SV-LLaVA153057.372.735.639.221.432.040.042.69.060%103
(Nr=256,Ns=320)-0.3%-2.2%+0.7%+8.5%-1%+0.5%-2.1%+9.6%+1.4%-11.8%-40%-17.6%
LLaVA-HD setting:
LLaVA-HD [67]15336471.835.339.32037.440.544.019100%450
SV-LLaVA-HD153461.273.135.939.721.234.139.343.512.250%250
(Nr=576,Ns=576)+0%-4.3%+1.8%+1.7%+1.5%-6%-8.8%-2.9%-1.1%-35.8%-50%-44.4%
CoS setting:
CoS [40]158564.477.139.439.221.539.241.246.015.3100%160
SV-CoS156363.776.541.740.321.23741.946.09.825%90
(Nr=272,Ns=256)-1.4%-1.1%-0.8%+5.8%+2.8%-1.4%-5.6%+1.7%-0%-40%-75%-43.8%
SV-CoS (SV-1M)156967.177.143.340.722.050.744.449.316.325%90
", + "image_path": "22513fff52ba8b8c07051b6ab66729d33d0c1c4ba4cd4d8bf2e53e60173c024d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 335, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "text", + "content": "Table 2. Performance and efficiency evaluation. We evaluate Skip-Vision on LLaVA, LLaVA-HD, and CoS and compare it with state-of-the-art efficiency optimization models under the LLaVA setting. " + }, + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "inline_equation", + "content": "N_s" + }, + { + "bbox": [ + 55, + 310, + 555, + 333 + ], + "type": "text", + "content": " denote the number of retained and skipped tokens, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 335, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 65, + 335, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 65, + 335, + 545, + 459 + ], + "type": "table", + "html": "
LLMMMETextvqaMMBMMVetMMMUMathVOCRBMMStar
Open-source models in 8B tier:
mPLUG-Owl3 [118]Qwen2-69.077.640.1-65.0-40.1
Cambrian [100]LLaMA3154771.775.9-42.749.062.4-
LLaVA [69]LLaMA3153558.672.232.839.621.332.736.5
Mini-Gemini-HD [61]LLaMA3160670.272.7-37.337.047.7-
LLaVA-Next [56]LLaMa3160464.672.1-41.736.349.0-
Ovis [77]LLaMA3--77.4-44.740.8-49.5
MiniCPM-V2.5 [117]LLaMA3-76.677.6-45.854.372.5-
SV-CoS (SV-9M)LLaMA3155069.178.451.74361.369.452.3
", + "image_path": "05e78d0a2944b11792ca413f593403bf0bc551a4a8e80bbbb6ae7b26692c76cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 555, + 489 + ], + "type": "text", + "content": "Table 3. MLLM evaluation. We present the performance of SV-CoS on SV-9M, comparing it against the current SOTA models of a similar scale. The highest scores are highlighted in bold, while the second-highest scores are indicated with underlines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 498, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 522 + ], + "type": "text", + "content": "In SFT, the resolution increases to 448, and all model parameters are unfrozen for full optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": "In all experiments, we used a learning rate of " + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": " for pre-training, which is reduced to " + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": " during SFT. Training follows a cosine decay schedule after a " + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 631 + ], + "type": "text", + "content": " warm-up phase. For skipped token selection, unless otherwise specified, the LLaVA setting includes 156 skipped tokens and 100 retained tokens, the LLaVA-HD setting has 576 skipped and 576 retained tokens, and the CoS setting consists of 256 skipped and 272 retained tokens. All skipped tokens undergo a token merging process to enhance efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 201, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 201, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 201, + 649 + ], + "type": "text", + "content": "6.5. Main experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Performance and efficiency. We first evaluate the effectiveness of Skip-Vision across three different frameworks: CoS, LLaVA, and LLaVA-HD, demonstrating its ability to maximize efficiency while minimizing performance degradation. As shown in Figure 1 and Table 6, Skip-Vision re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": "duces training time by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": ", inference FLOPs by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": ", and inference latency by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": " for LLaVA. For LLaVA-HD, it achieves a " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": " reduction in training time, " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": " in inference FLOPs, and " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": " in inference latency. In CoS, it reduces training time by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": ", inference FLOPs by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": ", and inference latency by " + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "text", + "content": "Scaling. We evaluated the scalability of Skip-Vision, leveraging CoS's visual scaling to increase visual tokens from 336 to 1296 while maintaining efficiency. Expanding the fine-tuning dataset from LLaVA-665k to SV-1M further improved performance by " + }, + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "inline_equation", + "content": "C o S_{1296}" + }, + { + "bbox": [ + 313, + 571, + 556, + 666 + ], + "type": "text", + "content": " with similar resources. To benchmark against state-of-the-art models, we scaled to SV-9M, with training completed in 72 hours on 16 NVIDIA A100 GPUs (Table 3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 672, + 468, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 468, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 468, + 685 + ], + "type": "text", + "content": "6.6. Ablation study and analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "To validate the effectiveness of each component within the Skip-Vision framework, we conducted ablation and com" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21390" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 194 + ], + "type": "table", + "html": "
MMETextvqaMMBMMVetMMMUMathVOCRBMMStarAvgLatency
SV-LLaVA (w/o SK)151956.772.531.438.821.230.936.541.1103
SV-LLaVA151956.572.532.438.821.230.736.541.390
+0%-0.4%+0%+3.2%+0%+0%-0.6%+0%+0.5%-9.7%
SV-LLaVA-HD (w/o SK)153361.673.035.839.921.234.539.743.7310
SV-LLaVA-HD153461.273.135.939.721.234.139.343.5250
+0%-0.7%+0.1%+0.3%-0.5%+0%-1.2%-1%-0.5%-19.4%
SV-LLaVA-CoS (w/o SK)156263.976.540.240.221.237.241.945.9126
SV-LLaVA-CoS156363.776.541.740.321.23741.946.090
+0%-0.3%+0%+3.7%+0.2%+0%-0.5%+0%+0.2%-28.6%
", + "image_path": "86470e3723517fd4c5b4f14c13cb728d4bdcefa1ba6b2fc4928f1eea925e1cc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 214, + 220, + 314 + ], + "blocks": [ + { + "bbox": [ + 67, + 202, + 541, + 213 + ], + "lines": [ + { + "bbox": [ + 67, + 202, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 67, + 202, + 541, + 213 + ], + "type": "text", + "content": "Table 4. Ablation on skip KV-cache. We evaluate SK (using skip KV-cache during inference) of Skip-Vision on different settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 214, + 220, + 314 + ], + "lines": [ + { + "bbox": [ + 82, + 214, + 220, + 314 + ], + "spans": [ + { + "bbox": [ + 82, + 214, + 220, + 314 + ], + "type": "image", + "image_path": "20bd027391091a7b44c86e87db03dc371646ae53d0deea8bdf5c02fd6b73d14e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 319, + 192, + 328 + ], + "lines": [ + { + "bbox": [ + 116, + 319, + 192, + 328 + ], + "spans": [ + { + "bbox": [ + 116, + 319, + 192, + 328 + ], + "type": "text", + "content": "(a) Ablation on LLaVA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 220, + 214, + 376, + 314 + ], + "blocks": [ + { + "bbox": [ + 220, + 214, + 376, + 314 + ], + "lines": [ + { + "bbox": [ + 220, + 214, + 376, + 314 + ], + "spans": [ + { + "bbox": [ + 220, + 214, + 376, + 314 + ], + "type": "image", + "image_path": "892870d1a207c1e5fed03929e6ea26e4f6a68289766f5d21b24157546f59af90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 319, + 338, + 328 + ], + "lines": [ + { + "bbox": [ + 272, + 319, + 338, + 328 + ], + "spans": [ + { + "bbox": [ + 272, + 319, + 338, + 328 + ], + "type": "text", + "content": "(b) Ablation on CoS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 243, + 384, + 273 + ], + "lines": [ + { + "bbox": [ + 376, + 243, + 384, + 273 + ], + "spans": [ + { + "bbox": [ + 376, + 243, + 384, + 273 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 384, + 214, + 528, + 314 + ], + "blocks": [ + { + "bbox": [ + 384, + 214, + 528, + 314 + ], + "lines": [ + { + "bbox": [ + 384, + 214, + 528, + 314 + ], + "spans": [ + { + "bbox": [ + 384, + 214, + 528, + 314 + ], + "type": "image", + "image_path": "b4ea6c94363d42f2866abbfcffe002a70241e63763992acd3b944562eeeb51eb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 319, + 501, + 328 + ], + "lines": [ + { + "bbox": [ + 410, + 319, + 501, + 328 + ], + "spans": [ + { + "bbox": [ + 410, + 319, + 501, + 328 + ], + "type": "text", + "content": "(c) Ablation on LLaVA-HD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 354, + 296, + 453 + ], + "blocks": [ + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 546, + 350 + ], + "type": "text", + "content": "Figure 6. Ablation study. We performed an ablation study on each component of the Skip-Vision. SF (skip FFN), TM (token merge)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 354, + 296, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 354, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 296, + 453 + ], + "type": "image", + "image_path": "9baf878fc339ad47abe3903f68fb1225de82816dfc51f51f9eaa5315a5efca38.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 476, + 296, + 577 + ], + "blocks": [ + { + "bbox": [ + 76, + 464, + 274, + 475 + ], + "lines": [ + { + "bbox": [ + 76, + 464, + 274, + 475 + ], + "spans": [ + { + "bbox": [ + 76, + 464, + 274, + 475 + ], + "type": "text", + "content": "Figure 7. Visualization of attention map in MMVet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 476, + 296, + 577 + ], + "lines": [ + { + "bbox": [ + 59, + 476, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 59, + 476, + 296, + 577 + ], + "type": "image", + "image_path": "28caf95df1a05c1466a032b57792f37bb6d1c1d30afce73eb7aad7557cb9b2e2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 586, + 278, + 598 + ], + "lines": [ + { + "bbox": [ + 72, + 586, + 278, + 598 + ], + "spans": [ + { + "bbox": [ + 72, + 586, + 278, + 598 + ], + "type": "text", + "content": "Figure 8. Visualization of attention map in TextVQA." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "parative experiments on token merge, skip FFN, and skip KV-cache." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 714 + ], + "type": "text", + "content": "As shown in Figure 6, both the skip FFN strategy and token merging enhance computational efficiency across different model settings while maintaining performance. Notably, skip FFN performs better in the CoS setting than in LLaVA and LLaVA-HD. We attribute this to CoS allowing the vision backbone to be fine-tuned during the SFT stage, enabling key information to be more effectively captured by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 354, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 411 + ], + "type": "text", + "content": "summary tokens, thereby reducing the errors introduced by skip FFN. As shown in Table 4, skip KV-cache slightly affects fine-grained tasks such as TextVQA and OCRBench but improves performance in long-form response tasks like MMVet." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 558 + ], + "type": "text", + "content": "To further analyze this, we visualized attention maps across layers for visual tokens in SV-CoS during the MMVet and TextVQA tasks. In MMVet (Figure 7), the attention of generated tokens primarily focuses on retained and summary tokens, with skip KV-cache filtering out unnecessary noise. In TextVQA (Figure 8), which requires fine-grained understanding, attention is spread across both retained and skipped tokens, leading to some performance loss with skip KV-cache. However, causal attention effectively aggregates information from skipped tokens into the final summary tokens, allowing us to retain only the summary tokens and retained visual tokens in the KV-cache." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "type": "text", + "content": "We propose Skip-Vision, an efficient multimodal large language model architecture that boosts computational efficiency through a skip-FFN strategy, reducing redundant computations over visual tokens during training. During inference, Skip-Vision uses a skip KV-cache to accelerate processing by omitting non-essential visual tokens from the KV cache. Extensive experiments show our model outperforms current methods in efficiency and scales effectively with data, delivering competitive performance among leading models of similar scale." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21391" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "content": "This work was supported in part by NSFC (62201342), and Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102). This work was supported by Ant Group Research Intern Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 177, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 177, + 296, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 177, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 66, + 177, + 296, + 220 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In Proceedings of the AAAI conference on artificial intelligence, pages 8076-8084, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 222, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 222, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 66, + 222, + 296, + 276 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 277, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 295, + 342 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35: 23716-23736, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 344, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 344, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 66, + 344, + 295, + 399 + ], + "type": "text", + "content": "[4] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 400, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 400, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 66, + 400, + 295, + 455 + ], + "type": "text", + "content": "[5] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 66, + 456, + 295, + 510 + ], + "type": "text", + "content": "[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 512, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 512, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 66, + 512, + 295, + 566 + ], + "type": "text", + "content": "[7] Samyadeep Basu, Martin Grayson, Cecily Morrison, Besmira Nushi, Soheil Feizi, and Daniela Massiceti. Understanding information storage and transfer in multi-modal large language models. arXiv preprint arXiv:2406.04236, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 567, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 567, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 66, + 567, + 295, + 622 + ], + "type": "text", + "content": "[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 66, + 624, + 295, + 666 + ], + "type": "text", + "content": "[9] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022.5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 669, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 669, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 61, + 669, + 294, + 690 + ], + "type": "text", + "content": "[10] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 692, + 294, + 713 + ], + "type": "text", + "content": "[11] Shihao Cai, Keqin Bao, Hangyu Guo, Jizhi Zhang, Jun Song, and Bo Zheng. Geogpt4v: Towards geometric multi-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "modal large language models with geometric image generation. arXiv preprint arXiv:2406.11503, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 162 + ], + "type": "text", + "content": "[12] Jianjian Cao, Peng Ye, Shengze Li, Chong Yu, Yansong Tang, Jiwen Lu, and Tao Chen. Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15710-15719, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 196 + ], + "type": "text", + "content": "[13] Jimmy Carter. Textocr-gpt4v. https://huggingface.co/datasets/jimmycarter/textocr-gpt4v,2024.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 251 + ], + "type": "text", + "content": "[14] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 253, + 553, + 308 + ], + "type": "text", + "content": "[15] Wenhao Chai, Enxin Song, Yilun Du, Chenlin Meng, Vashisht Madhavan, Omer Bar-Tal, Jenq-Neng Hwang, Saining Xie, and Christopher D Manning. Auroracap: Efficient, performant video detailed captioning and a new benchmark. arXiv preprint arXiv:2410.03051, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 553, + 342 + ], + "type": "text", + "content": "[16] Chungkwong Chan. Stroke extraction for offline handwritten mathematical expression recognition. IEEE Access, 8: 61565-61575, 2020. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 320, + 343, + 553, + 387 + ], + "type": "text", + "content": "[17] Shuaichen Chang, David Palzer, Jialin Li, Eric Fosler-Lussier, and Ningchuan Xiao. Mapqa: A dataset for question answering on choropleth maps. arXiv preprint arXiv:2211.08545, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 442 + ], + "type": "text", + "content": "[18] Guiming Hardy Chen, Shunian Chen, Ruifei Zhang, Junying Chen, Xiangbo Wu, Zhiyi Zhang, Zhihong Chen, Jianquan Li, Xiang Wan, and Benyou Wang. Allava: Harnessing gpt4v-synthesized data for a lite vision-language model, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 553, + 487 + ], + "type": "text", + "content": "[19] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 489, + 553, + 533 + ], + "type": "text", + "content": "[20] Lin Chen, Jisong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 588 + ], + "type": "text", + "content": "[21] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024.6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 553, + 644 + ], + "type": "text", + "content": "[22] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 553, + 690 + ], + "type": "text", + "content": "[23] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21392" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "text", + "content": "Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 172 + ], + "type": "text", + "content": "[25] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 61, + 106, + 295, + 172 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 61, + 106, + 295, + 172 + ], + "type": "text", + "content": " chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna/, 2023.1.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 172, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 294, + 239 + ], + "type": "text", + "content": "[26] Chee Kheng Chng, Yuliang Liu, Yipeng Sun, Chun Chet Ng, Canjie Luo, Zihan Ni, ChuanMing Fang, Shuaiqiao Zhang, Junyu Han, Errui Ding, et al. Icdar2019 robust reading challenge on arbitrary-shaped text-rrc-art. In 2019 International Conference on Document Analysis and Recognition (ICDAR), pages 1571–1576. IEEE, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[27] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevlm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 358 + ], + "type": "text", + "content": "[28] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "text", + "content": "[29] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "type": "text", + "content": "[30] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "text", + "content": "[31] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 580 + ], + "type": "text", + "content": "[32] Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations in auto-regressive language models. arXiv preprint arXiv:2304.14767, 2023.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 624 + ], + "type": "text", + "content": "[33] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2315-2324, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 625, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 295, + 691 + ], + "type": "text", + "content": "[34] Danna Gurari, Qing Li, Chi Lin, Yinan Zhao, Anhong Guo, Abigale Stangl, and Jeffrey P Bigham. Vizwiz-priv: A dataset for recognizing the presence and purpose of private visual information in images taken by blind people. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 939–948, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "text", + "content": "[35] Chi Han, Qifan Wang, Hao Peng, Wenhan Xiong, Yu Chen, Heng Ji, and Sinong Wang. Lm-infinite: Zero-shot extreme" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 128 + ], + "type": "text", + "content": "length generalization for large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3991-4008, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "type": "text", + "content": "[36] Mengchao He, Yuliang Liu, Zhibo Yang, Sheng Zhang, Canjie Luo, Feiyu Gao, Qi Zheng, Yongpan Wang, Xin Zhang, and Lianwen Jin. Icpr2018 contest on robust reading for multi-type web images. In 2018 24th international conference on pattern recognition (ICPR), pages 7-12. IEEE, 2018. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 194, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 194, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 194, + 553, + 249 + ], + "type": "text", + "content": "[37] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 553, + 304 + ], + "type": "text", + "content": "[38] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 304, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 304, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 320, + 304, + 553, + 350 + ], + "type": "text", + "content": "[39] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 350, + 553, + 403 + ], + "type": "text", + "content": "[40] Ziyuan Huang, Kaixiang Ji, Biao Gong, Zhiwu Qing, Qinglong Zhang, Kecheng Zheng, Jian Wang, Jingdong Chen, and Ming Yang. Accelerating pre-training of multimodal llms via chain-of-sight. arXiv preprint arXiv:2407.15819, 2024. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 553, + 458 + ], + "type": "text", + "content": "[41] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In International conference on machine learning, pages 4651-4664. PMLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 459, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 459, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 320, + 459, + 553, + 504 + ], + "type": "text", + "content": "[42] Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 504, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 504, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 504, + 553, + 569 + ], + "type": "text", + "content": "[43] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "type": "text", + "content": "[44] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 669 + ], + "type": "text", + "content": "[45] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figurea: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[46] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21393" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[47] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. arXiv preprint arXiv:2404.07204, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 194 + ], + "type": "text", + "content": "[48] Dimosthenis Karatzas, Lluis Gomez-Bigorda, Anguelos Nicolaou, Suman Ghosh, Andrew Bagdanov, Masakazu Iwamura, Jiri Matas, Lukas Neumann, Vijay Ramaseshan Chandrasekhar, Shijian Lu, et al. Icdar 2015 competition on robust reading. In 2015 13th international conference on document analysis and recognition (ICDAR), pages 1156-1160. IEEE, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 197, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 294, + 262 + ], + "type": "text", + "content": "[49] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "type": "text", + "content": "[50] Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern recognition, pages 4999-5007, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 374 + ], + "type": "text", + "content": "[51] Sunjun Kweon, Yeonsu Kwon, Seonhee Cho, Yohan Jo, and Edward Choi. Open-wikitable: Dataset for open domain question answering with complex reasoning over table. arXiv preprint arXiv:2305.07288, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 419 + ], + "type": "text", + "content": "[52] Jason J Lau, Soumya Gayen, Asma Ben Abacha, and Dina Demner-Fushman. A dataset of clinically generated visual questions and answers about radiology images. Scientific data, 5(1):1-10, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 454 + ], + "type": "text", + "content": "[53] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 498 + ], + "type": "text", + "content": "[54] Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Minjoon Seo. Volcano: mitigating multimodal hallucination through self-feedback guided revision. arXiv preprint arXiv:2311.07362, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 577 + ], + "type": "text", + "content": "[55] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 579, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 294, + 644 + ], + "type": "text", + "content": "[56] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llava-last: Stronger llms supercharge multimodal capabilities in the wild. https://llava-v1.github.io/blog/2024-05-10-llava-last-stronger-llms/, 2024.1,3,7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[57] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "text", + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 108, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 194 + ], + "type": "text", + "content": "[59] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 197, + 553, + 239 + ], + "type": "text", + "content": "[60] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 243, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 553, + 296 + ], + "type": "text", + "content": "[61] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 298, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 298, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 321, + 298, + 553, + 363 + ], + "type": "text", + "content": "[62] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 419 + ], + "type": "text", + "content": "[63] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 475 + ], + "type": "text", + "content": "[64] Dongyang Liu, Renrui Zhang, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, et al. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models. arXiv preprint arXiv:2402.05935, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "text", + "content": "[65] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 566 + ], + "type": "text", + "content": "[66] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "type": "text", + "content": "[67] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "text", + "content": "[68] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge. https://llava-v1.github.io/blog/2024-01-30-llava-last/, 2024.2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[69] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7, 5" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21394" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[70] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 162 + ], + "type": "text", + "content": "[71] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 294, + 228 + ], + "type": "text", + "content": "[72] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multimodal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 294, + 274 + ], + "type": "text", + "content": "[73] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Icdar 2023 competition on hierarchical text detection and recognition. arXiv preprint arXiv:2305.09750, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 294, + 330 + ], + "type": "text", + "content": "[74] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 332, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 396 + ], + "type": "text", + "content": "[75] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "type": "text", + "content": "[76] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 498 + ], + "type": "text", + "content": "[77] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv:2405.20797, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 554 + ], + "type": "text", + "content": "[78] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 599 + ], + "type": "text", + "content": "[79] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 644 + ], + "type": "text", + "content": "[80] Minesh Mathew, Lluis Gomez, Dimosthenis Karatzas, and CV Jawahar. Asking questions on handwritten document collections. International Journal on Document Analysis and Recognition (IJDAR), 24(3):235-249, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[81] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[82] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 97, + 553, + 129 + ], + "type": "text", + "content": "[83] Kevin Meng, Arnab Sen Sharma, Alex Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. arXiv preprint arXiv:2210.07229, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 131, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 131, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 321, + 131, + 553, + 174 + ], + "type": "text", + "content": "[84] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 176, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 176, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 321, + 176, + 553, + 208 + ], + "type": "text", + "content": "[85] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 210, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 210, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 321, + 210, + 553, + 275 + ], + "type": "text", + "content": "[86] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 277, + 553, + 342 + ], + "type": "text", + "content": "[87] Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don't know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784-789, Melbourne, Australia, 2018. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 345, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 345, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 321, + 345, + 553, + 410 + ], + "type": "text", + "content": "[88] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 412, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 412, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 412, + 553, + 476 + ], + "type": "text", + "content": "[89] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 479, + 553, + 521 + ], + "type": "text", + "content": "[90] Sanket Shah, Anand Mishra, Naganand Yadati, and Partha Pratim Talukdar. Kvqa: Knowledge-aware visual question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8876-8884, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 553, + 566 + ], + "type": "text", + "content": "[91] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 3, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 569, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 569, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 569, + 553, + 621 + ], + "type": "text", + "content": "[92] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "text", + "content": "[93] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[94] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. 2020. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21395" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[95] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 294, + 161 + ], + "type": "text", + "content": "[96] Dominik Stammbach and Elliott Ash. Docscan: Unsupervised text classification via learning from neighbors. arXiv preprint arXiv:2105.04024, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 194 + ], + "type": "text", + "content": "[97] Hamed Rahimi Sujet AI, Allaa Boutaleb. Sujet-finance-qavision-100k: A large-scale dataset for financial document vqa, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 195, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 295, + 216 + ], + "type": "text", + "content": "[98] S Svetlichnaya. Deepform: Understand structured documents at scale. 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 259 + ], + "type": "text", + "content": "[99] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 295, + 316 + ], + "type": "text", + "content": "[100] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 317, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 317, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 57, + 317, + 295, + 370 + ], + "type": "text", + "content": "[101] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 294, + 415 + ], + "type": "text", + "content": "[102] Andreas Veit, Tomas Matera, Lukas Neumann, Jiri Matas, and Serge Belongie. Coco-text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016.5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 416, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 416, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 57, + 416, + 294, + 458 + ], + "type": "text", + "content": "[103] Aladin Virmaux and Kevin Scaman. Lipschitz regularity of deep neural networks: analysis and efficient estimation. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 294, + 514 + ], + "type": "text", + "content": "[104] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 515, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 515, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 57, + 515, + 294, + 569 + ], + "type": "text", + "content": "[105] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 571, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 294, + 633 + ], + "type": "text", + "content": "[106] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Conghui He, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "text", + "content": "[107] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jian-Yuan Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. ArXiv, abs/2312.06109, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[108] Yuxin Wen, Qingqing Cao, Qichen Fu, Sachin Mehta, and Mahyar Najibi. Efficient vision-language models by" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "summarizing visual tokens into compact registers. arXiv preprint arXiv:2410.14072, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 552, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 552, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 552, + 118 + ], + "type": "text", + "content": "[109] C. Wendler. Renderedtext. https://huggingface.co/datasets/wendlerc/RenderedText, 2023.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 185 + ], + "type": "text", + "content": "[110] Jianshu Chen Yunkai Zhang Hong Wang Shiyang Li Xiyou Zhou Wenhu Chen, Hongmin Wang and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "text", + "content": "[111] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 307 + ], + "type": "text", + "content": "[112] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[113] Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts. arXiv preprint arXiv:2312.15915, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[114] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[115] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 4, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[116] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[117] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[118] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl3: Towards long image-sequence understanding in multi-modal large language models. arXiv preprint arXiv:2408.04840, 2024. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[119] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[120] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. Rlaif-v: Aligning mllms through open-source ai feedback for super gpt-4v trustworthyness. arXiv preprint arXiv:2405.17220, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[121] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21396" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "text", + "content": "Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[122] Xinyan Velocity Yu, Sewon Min, Luke Zettlemoyer, and Hannaneh Hajishirzi. Crepe: Open-domain question answering with false presuppositions. arXiv preprint arXiv:2211.17257, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[123] Qianhao Yuan, Qingyu Zhang, Yanjiang Liu, Jiawei Chen, Yaojie Lu, Hongyu Lin, Jia Zheng, Xianpei Han, and Le Sun. Shortv: Efficient multimodal large language models by freezing visual tokens in ineffective layers. arXiv preprint arXiv:2504.00502, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[124] Zhengqing Yuan, Zhaoxu Li, Weiran Huang, Yanfang Ye, and Lichao Sun. Tinygpt-v: Efficient multimodal large language model via small backbones. arXiv preprint arXiv:2312.16862, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 330 + ], + "type": "text", + "content": "[125] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multidiscipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[126] Chi Zhang, Feng Gao, Baoxiong Jia, Yixin Zhu, and Song-Chun Zhu. Raven: A dataset for relational and analogical visual reasoning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5317-5327, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 443 + ], + "type": "text", + "content": "[127] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[128] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Peng Gao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? arXiv preprint arXiv:2403.14624, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[129] Ying Zhang, Lionel Gueguen, Ilya Zharkov, Peter Zhang, Keith Seifert, and Ben Kadlec. Uber-text: A large-scale dataset for optical character recognition from street-level imagery. In SUNw: Scene Understanding Workshop-CVPR, page 5, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[130] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[131] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[132] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 217 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[133] Baichuan Zhou, Ying Hu, Xi Weng, Junlong Jia, Jie Luo, Xien Liu, Ji Wu, and Lei Huang. Tinyllava: A framework of small-scale large multimodal models. arXiv preprint arXiv:2402.14289, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[134] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 217 + ], + "type": "text", + "content": "[135] Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. Tat-qa: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21397" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_content_list.json b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71db3de0695bdc82b1d8b1270fcc708c4e6fb43a --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingying Zhang $^{1}$ Lixiang Ru $^{1}$ Kang Wu $^{2}$ Lei Yu $^{1}$ Lei Liang $^{1}$ Yansheng Li $^{2}$ Jingdong Chen $^{1}$ $^{1}$ Ant Group $^{2}$ Wuhan University qichu.zyy@antgroup.com", + "bbox": [ + 101, + 179, + 913, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The multi-modal remote sensing foundation model (MM-RSFM) has significantly advanced various Earth observation tasks, such as urban planning, environmental monitoring, and natural disaster management. However, most existing approaches generally require the training of separate backbone networks for each data modality, leading to redundancy and inefficient parameter utilization. Moreover, prevalent pre-training methods typically apply self-supervised learning (SSL) techniques from natural images without adequately accommodating the characteristics of remote sensing (RS) images, such as the complicated semantic distribution within a single RS image. In this work, we present SkySense V2, a unified MM-RSFM that employs a single transformer backbone to handle multiple modalities. This backbone is pre-trained with a novel SSL strategy tailored to the distinct traits of RS data. In particular, SkySense V2 incorporates an innovative adaptive patch merging module and learnable modality prompt tokens to address challenges related to varying resolutions and limited feature diversity across modalities. In additional, we incorporate the mixture of experts (MoE) module to further enhance the performance of the foundation model. SkySense V2 demonstrates impressive generalization abilities through an extensive evaluation involving 16 datasets over 7 tasks, outperforming SkySense by an average of 1.8 points.", + "bbox": [ + 89, + 300, + 485, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in multi-modal remote sensing foundation models (MM-RSFM) have significantly enhanced the utilization of remote sensing (RS) data across numerous Earth observation (EO) applications[5, 10, 36, 65], such as environmental monitoring, agriculture management, natural disaster response, land cover (or use) mapping etc. Recently, Guo et al. introduced SkySense [20], the largest MM-RSFM to date. SkySense has demonstrated impressive generalization capabilities through extensive evaluations across many different tasks. A key factor in its success is the simultaneous training of geographically aligned multi-modal", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a82f94c635106848da56b5529fc72525d5d6555cc4f85d6d7f6647edec4eda9.jpg", + "image_caption": [ + "Figure 1. SkySense V2 has achieved superior performance on 16 datasets over 7 distinct tasks compared with its predecessor Sky-Sense and supports a board range of tasks." + ], + "image_footnote": [], + "bbox": [ + 537, + 268, + 880, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and multi-temporal RS data. To accommodate the varying resolutions of different modalities, SkySense employs the Swin Transformer [31] for processing high-resolution (HR) optical images, while using the Vision Transformer (ViT) [16] for medium-resolution multi-spectral (MS) data and synthetic aperture radar (SAR) data. However, this separate backbone design results in redundancy and inefficient use of parameters. On the other hand, SkySense's pre-training approach is primarily based on DINOv2 [42], renowned for its efficiency in self-supervised learning (SSL). Although it has demonstrated strong performance with RS images, there remains considerable room for improvement by adequately accommodating the involved semantic distribution within a single RS image.", + "bbox": [ + 511, + 564, + 906, + 775 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce SkySense V2, an advanced MM-RSFM pre-trained through a novel SSL approach tailored to the distinctive traits of RS data. To be specific, SkySense V2 leverages a unified transformer backbone capable of processing data across diverse modalities and resolutions. It is pre-trained using a new SSL strategy that utilizes query-based attention to gather the similar semantic features distributed in different regions of RS images.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9136", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Designing a unified backbone for geographically aligned multi-modal data presents two primary challenges. The first is processing feature resolutions across data with varying ground sample distance (GSD) during simultaneous pretraining of multiple modalities. To address this challenge, we introduce an innovative Adaptive Patch Merging (APM) module, which is integrated after each stage of the unified backbone. This module determines whether to reduce feature resolution according to the specific feature resolution requirements of each modality. For instance, when dealing with HR optical images, the APM module activates resolution reduction at each stage. In contrast, for medium-resolution data, such as MS and SAR data, the feature resolution is preserved throughout the stages. This design enables the backbone to efficiently process data of different resolutions while adapting output feature resolutions according to actual requirements, which is crucial for the joint training and fusion of multi-modal features.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second challenge is that fully sharing parameters across the different modalities can reduce the feature diversity. To address this, inspired by the work of [14] and [26], we introduce learnable modality prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pretrained model can capture the unique characteristics of each modality better. Compared to SkySense, our unified design significantly enhances the efficiency of parameter utilization. To be specific, SkySense uses three separate backbones, Swin-H for HR optical data, ViT-L for MS data, and ViT-L for SAR data, collectively totaling 1.26 billion parameters. In contrast, SkySense V2's unified transformer backbone is capable of simultaneously processing all three modalities while utilizing only 665 million parameters. This efficient use of parameters not only optimizes the architecture but also allows for potential model scaling to further boost performance. Therefore, we integrate the mixture of experts module [24], a technique commonly used in various large language models [17, 45, 68] to further enhance the performance of the foundation model.", + "bbox": [ + 89, + 367, + 483, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively adapt SSL techniques developed for natural images to the pre-training of RSFM, it is essential to consider the distinct differences in data characteristics. Natural images typically feature a single, clear subject, such as a cat, dog, or person. In contrast, RS images comprise multiple subjects distributed across various regions of the image. For instance, as illustrated in Figure 2, one section of an RS image may show a building, while other areas might show a forest, pond, or land. Current SSL methods perform contrastive learning directly between different views using various augmented crops, which is effective for natural images due to their singular focus. However, when applied to RS images, this approach can result in semantic inaccuracies, as different views may capture different subjects. For", + "bbox": [ + 89, + 689, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/01538eb678699ee9904ed6057dca4de016e8849455a461b9bd592ff9d08c68b1.jpg", + "image_caption": [ + "(a) Views of natural image", + "Figure 2. Comparison of different augmented views from natural and RS images in SSL. Natural image views often concentrate on a singular subject, such as a \"dog\". RS image views encompass a variety of subjects, including building, forest, pond, and land." + ], + "image_footnote": [], + "bbox": [ + 524, + 87, + 705, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7e60dacfbae18e7330c45c0796f26ae6463f9e18b86279806a71a496a7bf06c5.jpg", + "image_caption": [ + "(b) Views of RS image" + ], + "image_footnote": [], + "bbox": [ + 714, + 89, + 893, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "example, in Figure 2b, view 1 of the RS image primarily contains building, whereas view 2 shows forest and pond.", + "bbox": [ + 511, + 323, + 903, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this challenge, we propose a novel approach called Query-based Semantic Aggregation Contrastive Learning (QSACL). This method employs multiple learnable queries to perform cross-attention with features from different views, generating semantically aggregated features. We then apply contrastive learning to these aggregated features pairs derived from the same query. As demonstrated in Figure 9 from our ablation study, utilizing different queries can aggregate consistent semantic features across multiple views, which enhances the accuracy of contrastive learning.", + "bbox": [ + 511, + 354, + 906, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated SkySense V2 on a diverse set of 16 datasets [8, 11-13, 15, 19, 27, 48, 51, 53, 59, 61], covering a range of task types, modalities, and spatial scales. As illustrated in Figure 1, SkySense V2 demonstrates substantial improvements in performance over its predecessor, SkySense, achieving state-of-the-art (SOTA) results across various modalities of EO tasks. The experimental results across all test scenarios highlight its competitive advantage over existing RSFMs in a wide array of EO interpretation tasks.", + "bbox": [ + 511, + 521, + 908, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 671, + 655, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Remote Sensing Foundation Model", + "text_level": 1, + "bbox": [ + 511, + 696, + 818, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RSFMs are characterized by their ability to leverage vast amounts of data through self-supervised learning (SSL) technologies, enabling them to learn robust feature representations without the need for extensive annotations. The predominant RSFMs utilize Contrastive Learning (CL) or Masked Image Modeling (MIM). Typical works in CL include RS-BYOL [25], GASSL [64], DINO-MC [58], SeCo [38], and CACo [37]. Other research efforts focus on enhancing MIM framework, such as RingMo [50], S2MAE [29], MA3E [30], SatMAE [13], and SatMAE++ [41]. Additionally, studies like CMID [40] and GFM [39] explore the intersection of CL and MIM through a self-distillation", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9137", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach. Recently, Guo et al. introduced a comprehensive MM-RSFM known as SkySense [20], which features a factorized multi-modal spatiotemporal encoder. This architecture facilitates independent spatial feature extraction and multi-modal temporal fusion. As the largest MM-RSFM to date, SkySense exhibits exceptional generalization capabilities across a wide range of RS datasets. However, it does face a challenge: the inefficient utilization of parameters due to the separate backbone designs for different modalities. In this work, we propose SkySense V2 to address this issue by employing a unified transformer backbone design.", + "bbox": [ + 89, + 90, + 483, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Unified Framework for Multi-modal Learning", + "text_level": 1, + "bbox": [ + 89, + 267, + 482, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal learning aims to train models that effectively process and relate information from various modalities. Substantial advancements have been made in this field, culminating in numerous studies focused on creating unified networks for the integration and processing of diverse modalities. The main work includes VLMO [46], MetaTransformer [66], Uni-Perceiver [67], and UniTR [21], etc. There are several methods that utilize a unified backbone design for multi-modal RS data. For example, OFA-Net [63] pre-trains a single transformer backbone on a curated multi-modal dataset using MIM. Han et al. proposed a RSFM called msGFM[22], which utilizes a shared transformer encoder while employing different decoders for each modality. In contrast to our approach, both OFA-Net and msGFM simply share all of parameters of backbone and learn feature representations for multi-modal data separately, which results in a lack of geographical alignment and fusion between features from different modalities. AnySat [2] employs a joint embedding predictive architecture for a multi-modal model, training a single model on heterogeneous data in a SSL manner. However, the backbone design and SSL pre-training methods of our SkySense V2 differ significantly from AnySat.", + "bbox": [ + 89, + 290, + 483, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SkySense V2", + "text_level": 1, + "bbox": [ + 89, + 652, + 223, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 676, + 277, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Unified Transformer Backbone", + "text_level": 1, + "bbox": [ + 89, + 699, + 354, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 3, our unified transformer backbone is a hierarchical encoder structure with four stages. In the first two stages, we employ Swin Transformer V2 Blocks (SwinV2B) [32] to incorporate essential visual priors, such as locality and translation invariance. The window-based self-attention mechanisms in SwinV2B also help reduce computational complexity compared to global self-attention, particularly given the high spatial resolution of features in these initial stages. In the last two stages, we utilize vanilla Transformer Blocks (TB) [16] for two main reasons: First, the spatial resolution of features in these stages is relatively low, making the computational", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5da551ae3edf10f05bd3a5f8d8ade3e925a500f1bab788fb4533d966a80b63ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 94, + 913, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/81f83f84530a2ece1a0a9856e9aeae73bb9c474b601085994f6c8743748e02e1.jpg", + "image_caption": [ + "Figure 3. Overview of the unified transformer backbone model in SkySense V2: The entire backbone shares all parameters across different modalities, with the exception of distinct tokenizers." + ], + "image_footnote": [], + "bbox": [ + 580, + 213, + 870, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "costs of global self-attention more manageable. Second, our ablation study demonstrates that global self-attention can be complementary to window-based self-attention, enabling the model to achieve stronger representations.", + "bbox": [ + 511, + 402, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a group of multi-modal inputs consisting of a high-resolution optical image $x_{HR}$ , multi-spectral data $x_{MS}$ , and synthetic aperture radar data $x_{SAR}$ where each 'pixel' in the different modalities is naturally aligned by geo-location, we employ three distinct tokenizers to process these inputs into tokens. In each tokenizer, the input data is first divided into non-overlapping $4 \\times 4$ patches. Subsequently, a linear embedding layer is applied to this raw, patched data to project it into patch tokens. Subsequently, four stages of either SwinV2B or TB, with shared parameters across modalities, are applied to these patch tokens.", + "bbox": [ + 511, + 463, + 908, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the varying GSD of satellite sensors across different modalities, input data corresponding to the same region at specific geo-locations during joint aligned training exhibit differing spatial resolutions. To harmonize these spatial resolutions for tokens from various modalities, we propose Adaptive Patch Merging (APM) and utilize it to selectively reduce spatial resolution at each stage, except for Stage 1. Specifically, for the tokens of optical image data with high spatial resolution, APM concatenates the features from groups of $2 \\times 2$ neighboring patches and applies a linear layer to the $4c$ -dimensional concatenated features. This process reduces the number of tokens by a factor of $2 \\times 2 = 4$ downsampling of spatial resolution, while the output dimension is set to $2c$ . In contrast, for the tokens of MS and SAR data with lower spatial resolution, APM maintains the resolution by applying a linear projection with averaging the weights across the input dimensions. By integrating the APM module, our unified backbone can efficiently process", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9138", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6fad7f3c67a3374040e087878f6577842cdc793c1fc760e1ddea740e1950af2c.jpg", + "image_caption": [ + "Figure 4. The illustration of modality-specific prompt tokens added to the unified backbone." + ], + "image_footnote": [], + "bbox": [ + 99, + 97, + 491, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multi-modal data with varying resolutions while maintaining the spatial alignment of features across modalities", + "bbox": [ + 89, + 318, + 482, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Modality-specific Prompt Tokens", + "text_level": 1, + "bbox": [ + 89, + 356, + 366, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our unified transformer backbone utilizes fully shared parameters across different modalities. To enhance feature diversity, we introduce learnable modality-specific prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pre-trained model can better capture the unique characteristics of each modality. As illustrated in Figure 4, we incorporate $N$ learnable prompt tokens for each modality in the last two stages. We denote the input tokens of each stage as $E_{i}^{j}\\in \\mathbb{R}^{h_{j}w_{j}\\times c_{j}}$ , $i\\in \\{HR,MS,SAR\\}$ , $j\\in \\{3,4\\}$ , where $h_j$ and $w_{j}$ are the height and width of the spatial resolution at stage $j$ , while $c_{j}$ refers to the dimension of the tokens. For each modality $i$ , we insert the modality-specific learnable prompt tokens $P_{i}^{j}\\in \\mathbb{R}^{N\\times c_{j}}$ into the input of the first block in $j$ -th stage $\\mathcal{F}_j$ . At the last block of each stage, these prompt tokens are discarded. The entire process can be formulated as:", + "bbox": [ + 89, + 375, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ E _ {d r o p}, E _ {i} ^ {4} \\right] = \\mathcal {F} _ {3} \\left(\\left[ P _ {i} ^ {3}, E _ {i} ^ {3} \\right]\\right), \\tag {1} \\\\ [ E _ {d r o p}, E _ {i} ^ {o u t} ] = \\mathcal {F} _ {4} ([ P _ {i} ^ {4}, E _ {i} ^ {4} ]). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 640, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $E_{drop}$ denotes the discarded tokens, and $E_{i}^{out}$ represents the final output tokens of backbone. Our design of modality-specific prompts enhances feature diversity while maintaining full parameter sharing by incorporating only a small number of modality-specific parameters.", + "bbox": [ + 89, + 691, + 483, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3. Scaling Up Model Capacity", + "text_level": 1, + "bbox": [ + 89, + 775, + 328, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our unified backbone design across various modalities significantly enhances parameter utilization efficiency. To further boost the model's capabilities, we integrate a Mixture of Experts (MoE) [24] into the backbone network. We choose MoE instead of simply increasing the backbone's width and depth because MoE enables models to be pretrained with substantially less computation by leveraging", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sparse feed-forward layers (i.e., experts) for individual tokens. Following common approaches [33, 60], we insert MoE modules into the last $L$ transformer blocks, replacing the original feed-forward network (FFN) layers. Each MoE module consists of $M$ experts (denoted as $\\mathcal{E}_i(\\cdot), i = 1,2,\\dots,M$ ), which share the same structure as the FFN and operate as independent networks. For the gating network in MoE, we employ a learnable linear layer followed by a Softmax function, $\\mathcal{G}(x) = \\text{Softmax}(Wx)$ , where W is the gating parameter. Finally, the output of the MoE is computed as a linear combination of the outputs from the selected experts, weighted by the corresponding gate values. This process is formulated as follows:", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM O E (x) = \\sum_ {i \\in \\mathcal {T}} \\mathcal {G} _ {i} (x) \\cdot \\mathcal {E} _ {i} (x), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 294, + 906, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{T}$ represent the set of the top- $k$ indices. In our SkySense V2 backbone, we set $L = 6,M = 8$ , and $k = 1$", + "bbox": [ + 511, + 333, + 906, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Pre-training Pipeline", + "text_level": 1, + "bbox": [ + 511, + 371, + 774, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 5 provides an overview of our pre-training procedure. The pre-training framework of SkySense V2 primarily adopts the teacher-student architecture from SkySense [20], where the teacher network's parameters are updated using an exponential moving average (EMA) [42] of the student network's parameters. To train SkySense V2, we employ the Multi-Granularity Contrastive Learning (MGCL) loss $\\mathcal{L}_{MGCL}$ as proposed in SkySense. For each modality, MGCL utilizes contrastive loss [42] to align the representations of teacher and student networks at pixel, object and image-level granularity, respectively. Additionally, we incorporate the unsupervised Geo-Context Prototype Learning (GCPL), as validated in SkySense, to enhance the learning of complementary regional context clues, aiding downstream tasks. To further enhance dense interpretation capabilities, we introduce an auxiliary supervision loss $\\mathcal{L}_{ITA}$ to employ dense Image-Text Alignment (ITA) according to OpenStreetMap (OSM) labels1. Detailed implementations of MGCL, GCPL, and ITA are included in the Appendix A. In following paragraph, we will focus on proposed query-based semantic aggregation contrastive learning (QSACL).", + "bbox": [ + 511, + 393, + 906, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query-based Semantic Aggregation Contrastive Learning. QSACL utilizes learnable queries to aggregate similar semantics across different regions of images, enabling more accurate contrastive learning. Given features $g_{1}, g_{2}$ from two global views and $l_{1}, l_{2}, \\ldots, l_{n}$ from $n$ local views (where $g$ and $l$ denotes the fused features $F_{\\mathrm{fus}}$ from global and local view), we employ a transformer decoder layer to perform cross-attention between $m$ learnable queries $q_{1}, q_{2}, \\ldots, q_{m}$ and the feature sets $g$ and $l$ . This process generates sets of globally and locally aggregated features set", + "bbox": [ + 511, + 728, + 908, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://www.openstreetmap.org/", + "bbox": [ + 531, + 886, + 710, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9139", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70a45e8d4136055737c6aa275dd9546d43ce537f250f09d2f596ef1fea283769.jpg", + "image_caption": [ + "SkySense V2 Pre-training", + "Figure 5. Overview of the SkySense V2 pre-training pipeline. The SkySense V2 employs data augmentations on multi-modal inputs, which are then fed into both the student and teacher networks. The pre-training process incorporates Multi-Granularity Contrastive Learning, Dense Image-Text Alignment, and Query-based Semantic Aggregation Contrastive Learning to effectively train the network. Once pretraining is complete, the parameters from the teacher branch are utilized for downstream applications." + ], + "image_footnote": [], + "bbox": [ + 114, + 112, + 861, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$z_{i}^{g}$ and $z_{i}^{l}$ for each $i = 1,2,\\ldots m$ . As is shown in Figure 6, we use superscript $\\prime$ denotes features from the teacher network, the QSACL loss $\\mathcal{L}_{QSACL}$ is computed by averaging the loss $\\mathcal{L}_{CL}$ of aggregation features between local and global views for each query:", + "bbox": [ + 89, + 483, + 483, + 560 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {Q S A C L} = \\frac {1}{2 m} \\sum_ {i = 1} ^ {m} \\left(\\mathcal {L} _ {C L} \\left(z _ {i} ^ {g}, z _ {i} ^ {l ^ {\\prime}}\\right) + \\mathcal {L} _ {C L} \\left(z _ {i} ^ {l}, z _ {i} ^ {g ^ {\\prime}}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 571, + 482, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C L} (x, x ^ {\\prime}) = - \\mathcal {H} (x) \\log \\left(\\mathcal {H} ^ {\\prime} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 627, + 482, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{H}$ and $\\mathcal{H}'$ denotes the learning head and corresponding EMA part as defined in[6]. The overall training objective is the weighted sum of $\\mathcal{L}_{MGCL}$ , $\\mathcal{L}_{ITA}$ , and $\\mathcal{L}_{QSACL}$ , i.e.,", + "bbox": [ + 89, + 651, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M G C L} + \\lambda_ {2} \\mathcal {L} _ {I T A} + \\lambda_ {3} \\mathcal {L} _ {Q S A C L}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 724, + 482, + 742 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2692c01802c74e2d82df136a5360141bd0c5b23806fdc2f686651197675ef1d3.jpg", + "image_caption": [ + "Figure 6. Computation pipelines for QSACL with multi-crop augmentation (2 global and $n$ local views)." + ], + "image_footnote": [], + "bbox": [ + 94, + 747, + 480, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 482, + 645, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following SkySense [20], we conducted experiments on 16 datasets across various modalities and tasks to thoroughly evaluate SkySense V2. The applications of pre-trained Sky-Sense V2 to diverse tasks are detailed in Appendix B. Each module is engineered for combined or individual use, providing the flexibility to be frozen or fine-tuned as needed.", + "bbox": [ + 511, + 508, + 906, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training Implementation", + "text_level": 1, + "bbox": [ + 511, + 612, + 769, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. We use the same training dataset of SkySense [20] to train SkySense V2. This extensive dataset comprises approximately 21 million multi-modal RS imagery sets. Each set comprises three components: a high-resolution optical RGB image of spatial dimension $2048 \\times 2048$ , a sequence of Sentinel-2 (S2) multi-spectral (MS) images with 10 spectral bands and spatial dimensions of $64 \\times 64$ (average sequence length of 65 images), and a sequence of Sentinel-1 (S1) synthetic aperture radar (SAR) images with 2 bands, also having spatial dimensions of $64 \\times 64$ (average sequence length of 10 images). In each training iteration, we randomly sample 10 S2 images and 5 S1 images of different time series to accommodate the GPU memory constraints.", + "bbox": [ + 511, + 635, + 908, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setting of Unified Backbone. We initially set the token dimension to $C = 352$ . In the first two stages, the win-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9140", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dow size for the Swin Transformer V2 Blocks (SwinV2B) is configured to 8. Each head has a query dimension of 32, and each MLP layer consists of two full-connected layers with dimensions expanding by a factor of 4 consistently across all blocks within the unified backbone. Prompt tokens are incorporated during the last two stages, with the number of prompt tokens being the same for each modality. We allocate 4 prompt tokens for stage-3 and stage-4.", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-training Settings. SkySense V2 is trained using a batch size of 1024, distributed across 128 H20 GPUs. The model undergoes a total of 600k iterations, utilizing the AdamW optimizer [35]. The learning rate is initially set to $2 \\times 10^{-4}$ and decays to $1 \\times 10^{-6}$ following a cosine annealing schedule [34]. Additional implementation details can be found in the Appendix C.", + "bbox": [ + 89, + 229, + 480, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Single-Modal Tasks", + "text_level": 1, + "bbox": [ + 89, + 344, + 277, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate SkySense V2 on four common single-modal RS interpretation tasks: scene classification, object detection, semantic segmentation, and change detection. The experiments are conducted via fine-tuning a pre-trained Sky-Sense V2 backbone along with a task-specific head. Further implementation details can be found in the Appendix D.", + "bbox": [ + 89, + 366, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Classification. In Table 1, we report the performance of scene classification on four widely used datasets: AID [61] and RESISC-45 [11], both featuring static RGB optical images; BEN-S2 [48], which includes static MS images; and fMoW-S2 [13], which contains temporal MS images. For each dataset, the performance is evaluated using different training ratios (TR) following the methodology [13, 38]. In our experiments, we employ a simple linear classifier as the classification head. For multi-label datasets, we use the mean average precision (mAP) as the evaluation metric, whereas for single-label datasets, we report the overall accuracy (OA). As illustrated in Table 1, SkySense V2 consistently outperforms previous RSFMs across various datasets and training configurations, particularly in low TR settings. This demonstrates that SkySense V2 exhibits a stronger representation capability compared to other RSFMs, including its predecessor, SkySense [20].", + "bbox": [ + 89, + 474, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semantic Segmentation. In Table 2a, we report the segmentation results from four representative semantic segmentation datasets: iSAID [59] and Potsdam [47] for high-resolution optical image segmentation, and Dyna.-Pla [53] and Dyna.-S2 [53] for multi-spectral and multi-temporal image segmentation. Following previous work, we use the mean F1-score (mF1) as the evaluation metric for the Potsdam dataset. While for the other datasets, we report the mean intersection over union (mIoU). All segmentation experiments employ the UperNet [62] as the decoder head.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/17eb365830aac4a3e0f2bb189c38788dc0cdb092d95e31fa05fb7dc32411a3aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [3]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [38]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [13]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo [50]96.90/98.3494.25/95.67--
RVSA [55]97.03/98.5093.93/95.69--
DINO-MC [58]--84.20/88.7560.16/83.49
TOV [52]95.16/97.0990.97/93.79--
SSL4EO [57]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [40]96.11/97.7994.05/95.53--
CACo [37]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMA [18]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [39]95.47/97.0992.73/94.6486.30/--
Scale-MAE [43]96.44/97.5892.63/95.04--
MA3E [30]-/99.04-/96.23--
SatMAE++ [41]--/97.4885.11/-63.23/-
SkySense [20]97.68/98.6094.85/96.3288.67/92.0964.38/87.27
SkySense V298.34/99.0596.42/97.2489.13/93.7866.65/89.32
", + "bbox": [ + 526, + 88, + 893, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Scene classification results. - means the task is not supported or the value is unavailable in the paper.", + "bbox": [ + 511, + 354, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Table 2a, our SkySense V2 achieve higher performance than recent RSFMs across all four datasets. Specifically, compared to the previous SOTA method SkySense [20], SkySense V2 achieves $1.5\\%$ higher performance on average.", + "bbox": [ + 511, + 401, + 903, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Horizontal & Oriented Object Detection. We utilize the DIOR dataset [27] to evaluate SkySense V2's performance in horizontal object detection, while utilizing the DIOR-R [12] and FAIR1M [51] datasets for oriented object detection. These datasets only consist of RGB RS imagery. Consistent with previous research [20], we adopt Faster R-CNN [44] and Oriented R-CNN [28] as the basic detectors for horizontal and oriented object detection, respectively. As reported in Table 2b, our SkySense V2 outperforms previous RSFMs by a notable margin. Specifically, SkySense V2 surpasses the previous SOTA, SkySense [20], by an average of $1.1\\%$ mAP.", + "bbox": [ + 511, + 500, + 905, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Change Detection. For change detection, we conducted experiments on LEVIR-CD [8], OSCD [15] and Dyna.-S2 [53] datasets. Since LEVIR-CD and OSCD datasets only contain binary change detection results, we use the basic change detector in [9] and report the F1 score as the evaluation metric. While Dyna.-S2 focuses on semantic change detection, we use UperNet [62] as the segmentation network and calculate the semantic change detection score (SCS) on validation and test set following [20]. As presented in Table 2c, SkySense V2 consistently achieves higher performance than the previous SOTA method SkySense on all 3 datasets, and surpasses SkySense by $2.7\\%$ on average. Particularly, on the OSCD dataset, our SkySense V2 remark", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9141", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1880090de436d3b750c653c2a577482c7b986d28f2fde08b5361e3f956322c8c.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Semantic segmentation results." + ], + "table_body": "
ModelsPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [3]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [38]ICCV'21-57.2089.0329.4/39.8
SatMAE [13]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo [50]TGRS'22-67.2091.27-
RVSA [55]TGRS'2234.3/44.464.49--
BFM [7]Arxiv'23--92.12-
TOV [52]JSTARS'2332.1/37.866.2492.03-
SSL4EO [57]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [40]TGRS'2336.4/43.566.2191.86-
CACo [37]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS [56]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [39]ICCV'2336.7/45.666.6291.85-
Scale-MAE [43]ICCV'2334.0/41.765.7791.54-
MA3E [30]ECCV'24-64.0691.50-
SkySense [20]CVPR'2439.7/46.570.9193.9933.1/46.2
SkySense V2-41.2/47.671.8795.8635.2/47.5
", + "bbox": [ + 91, + 88, + 424, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef03af63540b704b25950ecd390cf28eb3cf952f8dea7a80fe691affc1c06897.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Object detection results." + ], + "table_body": "
ModelsHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [3]67.4065.6548.15
SatMAE [13]70.8965.6646.55
RingMo [50]75.90-46.21
RVSA [55]73.2271.0547.04
BFM [7]-73.62-
TOV [52]70.1666.3349.62
SSL4EO [57]64.8261.2349.37
CMID [40]75.1166.3750.58
CACo [37]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [39]72.8467.6749.69
Scale-MAE [43]73.8166.4748.31
MA3E [30]-71.82-
SkySense [20]78.7374.2754.57
SkySense V279.5075.2955.96
", + "bbox": [ + 429, + 90, + 661, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67ebc81055753547a3938edff7634ad9966ba026c5573527472c6e921596264d.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Change detection results." + ], + "table_body": "
ModelsLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [3]78.1946.2613.6/16.7
SeCo [38]90.1447.6713.9/16.0
SatMAE [13]87.6552.7614.8/16.2
RingMo [50]91.86--
RVSA [55]90.86--
SpectralGPT [23]-54.29-
MATTER [1]-59.37-
DINO-MC [58]-52.7014.5/15.6
SSL4EO [57]89.0535.0812.3/17.5
CMID [40]91.72--
CACo [37]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [39]91.7359.82-
Scale-MAE [43]92.07--
SkySense [20]92.5860.0615.4/18.0
SkySense V294.8365.2916.0/20.7
", + "bbox": [ + 666, + 92, + 905, + 303 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f2fd6a57db011827a029766f6867a719b02c55d3557d07176a5baa3f29a80b1.jpg", + "table_caption": [ + "Table 2. Results of semantic segmentation, object detection and change detection." + ], + "table_footnote": [ + "Table 3. Fine-tuning results on multi-modal tasks." + ], + "table_body": "
Task & DatasetData SourceSkySense[20]SkySense V2
(a) Multi-modal Seg: Dyna.-MM(i) Planet.46.547.6
(ii) S246.247.5
(iii) Planet. + S247.348.7
(iv) Planet. + S2 + S147.748.9
(b) Multi-modal Seg: PASTIS-MM(i) S273.575.0
(ii) S2-Ts84.685.5
(iii) S2-Ts + S1-Ts84.885.6
(iv) S2-Ts + GEP85.886.7
(c) Multi-modal Cls: BEN-MM(i) S186.286.5
(ii) S2 + S192.293.8
", + "bbox": [ + 101, + 358, + 470, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ably outperform SkySense by $5.2\\%$ on F1-score.", + "bbox": [ + 89, + 542, + 410, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Multi-modal Tasks", + "text_level": 1, + "bbox": [ + 89, + 568, + 272, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate SkySense V2 on two representative multi-modal tasks, i.e., Multi-modal Segmentation and Multi-modal Scene Classification, to demonstrate its generalization capability across data from various modalities.", + "bbox": [ + 89, + 590, + 482, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modal Segmentation. In Table 3 (a) and (b), we present the performance results of Dyna.-MM [53] and PASTIS-MM [19], respectively. Dyna.-MM contains high-resolution optical imagery from PlanetFusion (Planet.), multi-spectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). For the experiments conducted on Dyna.-MM, we employ a basic UperNet [62] as the decoder head and utilize mIoU as the evaluation metric. As depicted in Table 3 (a), SkySense V2 consistently outperforms the SkySense across various modalities, demonstrating a stronger representation ability. Furthermore, integrating multiple modalities yields higher performance compared to using a single modality. This results confirms that the proposed SkySense V2 can effectively extract representations from diverse modalities using a unified backbone.", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PASTIS-MM is a crop mapping dataset that contains high-resolution optical imagery from Google Earth Pro (GEP), multi-temporal multi-spectral imagery from Sentinel-2 (S2-Ts), and multi-spectral SAR imagery from Sentinel-1 (S1-Ts). We employ a straightforward FCN head to decode the semantic segmentation results and report the overall accuracy as the evaluation metric in Table 3 (b). The results demonstrate that SkySense V2 outperforms the previous SOTA method SkySense by an average of $1.0\\%$ . Additionally, it is worth noting that S2-Ts can remarkably improve the segmentation accuracy compared to S2, underscoring the importance of temporal information for crop mapping.", + "bbox": [ + 511, + 359, + 906, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-modal Scene Classification. We conduct multimodal scene classification experiments on the BEN-MM dataset [49]. As is shown in Table 3 (c), the proposed SkySense V2 surpasses SkySense, demonstrating a stronger representation ability. Furthermore, by integrating S1 and S2 imagery, SkySense V2 achieves higher improvements, highlighting its enhanced ability to extract superior representations from different modalities compared to SkySense.", + "bbox": [ + 511, + 565, + 908, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation and Discussion", + "text_level": 1, + "bbox": [ + 511, + 696, + 735, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study of Components. To understand the contributions of each component, we conduct a k-nearest neighbors (k-NN) evaluation on the AID and RESISC-45 datasets. Figure 7 illustrates the performance of different component combinations across various pre-training iterations. Notably, the unified backbone design significantly accelerates representation learning. This is primarily due to shared parameters across different modalities, which allow gradients to aggregate, thereby speeding up convergence. Moreover, the unified design allows the backbone to be trained with data from different modalities, thereby enhancing the model's generalization capabilities (see Ap", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9142", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/157b50fa09f6a9bb3c383d2cbfb764cf3ce022dc61b0ebacd6e457ac628d5358.jpg", + "image_caption": [ + "(a) AID dataset" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 285, + 193 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ddab7a34b9c460ff759e00ca77fa0ac2f30b72d847516100b464b78fbb3ac768.jpg", + "image_caption": [ + "(b) RESISC-45 dataset" + ], + "image_footnote": [], + "bbox": [ + 289, + 89, + 478, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pendix F.3 for more details). The results also indicate that global self-attention complements window-based self-attention, facilitating improved representations. Additionally, modality-specific prompt tokens and MoE scaling enhance representation learning by increasing feature diversity and model capacity, respectively. QSACL enhances model performance by improving the semantic precision of features in contrastive learning.", + "bbox": [ + 89, + 358, + 483, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8afe768d3b37554c5227266fccc75147b6e8a6aa0fea4fdc8844b8fd3bf8c00.jpg", + "image_caption": [ + "(a) w/o modality-specific prompt", + "Figure 8. t-SNE visualization of the output feature representation from the final block of the unified transformer backbone. HROI, SARI, and MSI represent the features from high-resolution optical imagery, synthetic aperture radar imagery, and multi-spectral imagery, respectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 491, + 289, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab0b69daa3fd05c6b98ef39b824a4d8b9019942209e76506731358ad8be19be5.jpg", + "image_caption": [ + "(b) w/ modality-specific prompt" + ], + "image_footnote": [], + "bbox": [ + 292, + 491, + 480, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How Do Modality-Specific Prompt Tokens Facilitate Pre-training? To investigate the impact of modality-specific prompt (MsP) tokens on the multi-modal pretraining process within a unified backbone, we employed t-SNE [54] for visualizing the features corresponding to different modalities. These features are extracted from the last block of the unified transformer backbone, and the comparative results are presented in Figure 8. Without the use of MsP, the features of high-resolution optical imagery (HROI) and synthetic aperture radar imagery (SARI) appear closely clustered, leading to overlapping feature distri", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "butions. However, with the inclusion of MsP, the features of HROI and SARI become distinctly separable. These visualization outcomes clearly demonstrate that MsP effectively enhances feature diversity and imparts modality-specific characteristics.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/721ae3dd381463ee86702aa22004dd666abd304c68e69858a1ed9f9c2f219d30.jpg", + "image_caption": [ + "Figure 7. Ablation study using k-NN classification was conducted on the AID and RESISC-45 datasets. In this context, the abbreviations have the following meanings: GA refers to replacing window-based self-attention with global self-attention in the last two stages; MsP involves the addition of modality-specific prompt tokens; MoE indicates scaling up the model through a mixture of experts approach; and QSACL means adding additional query-based semantic aggregation contrastive learning." + ], + "image_footnote": [], + "bbox": [ + 516, + 183, + 709, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ec7f5d7fa36c20e60ba37f477b4a6224da6008f0e3b2066617163a2ea2177e87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 184, + 903, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e10518bb815b3b681427b5d6f84a2776cfe4be3d5ca83fb3d896f24a99b5925.jpg", + "image_caption": [ + "Figure 9. Visualization of cross-attention weights for features patch corresponding to different queries in QSACL. Queries can effectively aggregate features with specific semantics." + ], + "image_footnote": [], + "bbox": [ + 517, + 237, + 709, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46b0d1342509ec15d7fe3af9840b4562ba3b30bf5cdb50ec90b034c69a2b0e3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 237, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "What Types of Features Does Query-based Semantic Aggregation Contrastive Learning Capture? To provide a clearer understanding of our proposed QSACL, we visualize the attention weights assigned to different features during their interaction with certain query, as illustrated in Figure 9. During pre-training, we use two global augmentation crops and six local augmentation crops. Here, we present two distinct queries, each focusing on different semantic features of image patches. For example, one query focuses on features with building semantics (left part of the first row in Figure 9), while another focuses on features with land semantics (right part of the first row in Figure 9). The attention weights of different queries aggregate features from each patch into distinct feature representations. Subsequently, we perform contrastive learning between aggregated features from the same query across teacher and student models. This approach significantly enhances semantic accuracy compared to traditional contrastive learning applied to entire image features.", + "bbox": [ + 511, + 371, + 906, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce SkySense V2, a MM-RSFM that utilizes a unified backbone to accommodate various modalities. This unified approach enhances parameter utilization efficiency and improves the model's generalization ability. SkySense V2 is pre-trained with an innovative QSACL strategy, specifically designed to leverage the unique characteristics of RS images. As a result, SkySense V2 significantly boosts performance while maintaining the multi-modal flexibility benefits of its predecessor, SkySense. Looking ahead, our future work will focus on integrating the language modality and incorporating a geographical knowledge graph to develop a more powerful and versatile MM-RSFM.", + "bbox": [ + 511, + 703, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9143", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 7", + "[2] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. AnySat: An Earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3", + "[3] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 6, 7", + "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 6, 7", + "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5", + "[7] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 7", + "[8] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6", + "[9] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 6", + "[10] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1", + "[11] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 6", + "[12] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 6", + "[13] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[15] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3", + "[17] William Fedus, Barret Zoph, and Noam M. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ArXiv, abs/2101.03961, 2021. 2", + "[18] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 6", + "[19] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7", + "[20] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. Skysense: A multimodal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27672-27683, 2024. 1, 3, 4, 5, 6, 7", + "[21] Wang Haiyang, Tang Hao, Shi Shaoshuai, Li Aoxue, Li Zhenguo, Schiele Bernt, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. 3", + "[22] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27852-27862, 2024. 3", + "[23] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Pedram Ghamisi, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, (8):5227-5244, 2024. 7", + "[24] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3:79–87, 1991. 2, 4", + "[25] Pallavi Jain, Bianca Schoen-Phelan, and Robert J. Ross. Self-supervised learning for invariant representations from" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9144", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multi-spectral and sar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 7797-7808, 2022. 2", + "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2", + "[27] Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6", + "[28] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reppoints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6", + "[29] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24088-24097, 2024. 2", + "[30] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images. In European Conference on Computer Vision, pages 260-278. Springer, 2024. 2, 6, 7", + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1", + "[32] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 3", + "[33] Zhili Liu, Kai Chen, Jianhua Han, Lanqing Hong, Hang Xu, Zhenguo Li, and James Tin-Yau Kwok. Task-customized masked autoencoder via mixture of cluster-conditional experts. ArXiv, abs/2402.05382, 2024. 4", + "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning, 2016. 6", + "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. CoRR, abs/1711.05101, 2017. 6", + "[36] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1", + "[37] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7", + "[38] Oscar Manas, Alexandre Lacoste, Xavier Giro-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 6, 7", + "[40] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 6, 7", + "[41] Mubashir Noman, Muzammal Naseer, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, and Fahad Shahbaz Khan. Rethinking transformers pre-training for multispectral satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27811-27819, 2024. 2, 6", + "[42] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 4", + "[43] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 6, 7", + "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6", + "[45] Sheng Shen, Le Hou, Yan-Quan Zhou, Nan Du, S. Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In International Conference on Learning Representations, 2023. 2", + "[46] Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Y Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[47] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6", + "[48] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Felipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 6" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9145", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7", + "[50] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7", + "[51] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fairlm: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6", + "[52] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 6, 7", + "[53] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7", + "[54] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8", + "[55] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 6, 7", + "[56] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 7", + "[57] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 6, 7", + "[58] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6, 7", + "[59] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6" + ], + "bbox": [ + 91, + 90, + 480, + 897 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Lemeng Wu, Mengchen Liu, Yinpeng Chen, Dongdong Chen, Xiyang Dai, and Lu Yuan. Residual mixture of experts. ArXiv, abs/2204.09636, 2022. 4", + "[61] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 6", + "[62] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6, 7", + "[63] Zhitong Xiong, Yi Wang, Fahong Zhang, and Xiao Xiang Zhu. One for all: Toward unified foundation models for earth vision. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 2734-2738, 2024. 3", + "[64] Longqi Yang, Liangliang Zhang, and Wenjing Yang. Graph adversarial self-supervised learning. In Advances in Neural Information Processing Systems, 2021. 2", + "[65] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1", + "[66] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 3", + "[67] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uniperceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16783-16794, 2021. 3", + "[68] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam M. Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 612 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9146", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_model.json b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d6b6f98ef61af032ec5a2655c1f0c48ae75bd6c --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.18, + 0.914, + 0.235 + ], + "angle": 0, + "content": "Yingying Zhang\\(^{1}\\) Lixiang Ru\\(^{1}\\) Kang Wu\\(^{2}\\) Lei Yu\\(^{1}\\) Lei Liang\\(^{1}\\) Yansheng Li\\(^{2}\\) Jingdong Chen\\(^{1}\\) \n\\(^{1}\\)Ant Group \\(^{2}\\) Wuhan University qichu.zyy@antgroup.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.694 + ], + "angle": 0, + "content": "The multi-modal remote sensing foundation model (MM-RSFM) has significantly advanced various Earth observation tasks, such as urban planning, environmental monitoring, and natural disaster management. However, most existing approaches generally require the training of separate backbone networks for each data modality, leading to redundancy and inefficient parameter utilization. Moreover, prevalent pre-training methods typically apply self-supervised learning (SSL) techniques from natural images without adequately accommodating the characteristics of remote sensing (RS) images, such as the complicated semantic distribution within a single RS image. In this work, we present SkySense V2, a unified MM-RSFM that employs a single transformer backbone to handle multiple modalities. This backbone is pre-trained with a novel SSL strategy tailored to the distinct traits of RS data. In particular, SkySense V2 incorporates an innovative adaptive patch merging module and learnable modality prompt tokens to address challenges related to varying resolutions and limited feature diversity across modalities. In additional, we incorporate the mixture of experts (MoE) module to further enhance the performance of the foundation model. SkySense V2 demonstrates impressive generalization abilities through an extensive evaluation involving 16 datasets over 7 tasks, outperforming SkySense by an average of 1.8 points." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Advancements in multi-modal remote sensing foundation models (MM-RSFM) have significantly enhanced the utilization of remote sensing (RS) data across numerous Earth observation (EO) applications[5, 10, 36, 65], such as environmental monitoring, agriculture management, natural disaster response, land cover (or use) mapping etc. Recently, Guo et al. introduced SkySense [20], the largest MM-RSFM to date. SkySense has demonstrated impressive generalization capabilities through extensive evaluations across many different tasks. A key factor in its success is the simultaneous training of geographically aligned multi-modal" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.269, + 0.882, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.507, + 0.908, + 0.55 + ], + "angle": 0, + "content": "Figure 1. SkySense V2 has achieved superior performance on 16 datasets over 7 distinct tasks compared with its predecessor Sky-Sense and supports a board range of tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.776 + ], + "angle": 0, + "content": "and multi-temporal RS data. To accommodate the varying resolutions of different modalities, SkySense employs the Swin Transformer [31] for processing high-resolution (HR) optical images, while using the Vision Transformer (ViT) [16] for medium-resolution multi-spectral (MS) data and synthetic aperture radar (SAR) data. However, this separate backbone design results in redundancy and inefficient use of parameters. On the other hand, SkySense's pre-training approach is primarily based on DINOv2 [42], renowned for its efficiency in self-supervised learning (SSL). Although it has demonstrated strong performance with RS images, there remains considerable room for improvement by adequately accommodating the involved semantic distribution within a single RS image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce SkySense V2, an advanced MM-RSFM pre-trained through a novel SSL approach tailored to the distinctive traits of RS data. To be specific, SkySense V2 leverages a unified transformer backbone capable of processing data across diverse modalities and resolutions. It is pre-trained using a new SSL strategy that utilizes query-based attention to gather the similar semantic features distributed in different regions of RS images." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.364 + ], + "angle": 0, + "content": "Designing a unified backbone for geographically aligned multi-modal data presents two primary challenges. The first is processing feature resolutions across data with varying ground sample distance (GSD) during simultaneous pretraining of multiple modalities. To address this challenge, we introduce an innovative Adaptive Patch Merging (APM) module, which is integrated after each stage of the unified backbone. This module determines whether to reduce feature resolution according to the specific feature resolution requirements of each modality. For instance, when dealing with HR optical images, the APM module activates resolution reduction at each stage. In contrast, for medium-resolution data, such as MS and SAR data, the feature resolution is preserved throughout the stages. This design enables the backbone to efficiently process data of different resolutions while adapting output feature resolutions according to actual requirements, which is crucial for the joint training and fusion of multi-modal features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.685 + ], + "angle": 0, + "content": "The second challenge is that fully sharing parameters across the different modalities can reduce the feature diversity. To address this, inspired by the work of [14] and [26], we introduce learnable modality prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pretrained model can capture the unique characteristics of each modality better. Compared to SkySense, our unified design significantly enhances the efficiency of parameter utilization. To be specific, SkySense uses three separate backbones, Swin-H for HR optical data, ViT-L for MS data, and ViT-L for SAR data, collectively totaling 1.26 billion parameters. In contrast, SkySense V2's unified transformer backbone is capable of simultaneously processing all three modalities while utilizing only 665 million parameters. This efficient use of parameters not only optimizes the architecture but also allows for potential model scaling to further boost performance. Therefore, we integrate the mixture of experts module [24], a technique commonly used in various large language models [17, 45, 68] to further enhance the performance of the foundation model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To effectively adapt SSL techniques developed for natural images to the pre-training of RSFM, it is essential to consider the distinct differences in data characteristics. Natural images typically feature a single, clear subject, such as a cat, dog, or person. In contrast, RS images comprise multiple subjects distributed across various regions of the image. For instance, as illustrated in Figure 2, one section of an RS image may show a building, while other areas might show a forest, pond, or land. Current SSL methods perform contrastive learning directly between different views using various augmented crops, which is effective for natural images due to their singular focus. However, when applied to RS images, this approach can result in semantic inaccuracies, as different views may capture different subjects. For" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.088, + 0.707, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.229, + 0.686, + 0.242 + ], + "angle": 0, + "content": "(a) Views of natural image" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.09, + 0.895, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.228, + 0.868, + 0.242 + ], + "angle": 0, + "content": "(b) Views of RS image" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.254, + 0.908, + 0.31 + ], + "angle": 0, + "content": "Figure 2. Comparison of different augmented views from natural and RS images in SSL. Natural image views often concentrate on a singular subject, such as a \"dog\". RS image views encompass a variety of subjects, including building, forest, pond, and land." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.354 + ], + "angle": 0, + "content": "example, in Figure 2b, view 1 of the RS image primarily contains building, whereas view 2 shows forest and pond." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.521 + ], + "angle": 0, + "content": "To tackle this challenge, we propose a novel approach called Query-based Semantic Aggregation Contrastive Learning (QSACL). This method employs multiple learnable queries to perform cross-attention with features from different views, generating semantically aggregated features. We then apply contrastive learning to these aggregated features pairs derived from the same query. As demonstrated in Figure 9 from our ablation study, utilizing different queries can aggregate consistent semantic features across multiple views, which enhances the accuracy of contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.658 + ], + "angle": 0, + "content": "We evaluated SkySense V2 on a diverse set of 16 datasets [8, 11-13, 15, 19, 27, 48, 51, 53, 59, 61], covering a range of task types, modalities, and spatial scales. As illustrated in Figure 1, SkySense V2 demonstrates substantial improvements in performance over its predecessor, SkySense, achieving state-of-the-art (SOTA) results across various modalities of EO tasks. The experimental results across all test scenarios highlight its competitive advantage over existing RSFMs in a wide array of EO interpretation tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.672, + 0.656, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.82, + 0.714 + ], + "angle": 0, + "content": "2.1. Remote Sensing Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "RSFMs are characterized by their ability to leverage vast amounts of data through self-supervised learning (SSL) technologies, enabling them to learn robust feature representations without the need for extensive annotations. The predominant RSFMs utilize Contrastive Learning (CL) or Masked Image Modeling (MIM). Typical works in CL include RS-BYOL [25], GASSL [64], DINO-MC [58], SeCo [38], and CACo [37]. Other research efforts focus on enhancing MIM framework, such as RingMo [50], S2MAE [29], MA3E [30], SatMAE [13], and SatMAE++ [41]. Additionally, studies like CMID [40] and GFM [39] explore the intersection of CL and MIM through a self-distillation" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.259 + ], + "angle": 0, + "content": "approach. Recently, Guo et al. introduced a comprehensive MM-RSFM known as SkySense [20], which features a factorized multi-modal spatiotemporal encoder. This architecture facilitates independent spatial feature extraction and multi-modal temporal fusion. As the largest MM-RSFM to date, SkySense exhibits exceptional generalization capabilities across a wide range of RS datasets. However, it does face a challenge: the inefficient utilization of parameters due to the separate backbone designs for different modalities. In this work, we propose SkySense V2 to address this issue by employing a unified transformer backbone design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.268, + 0.483, + 0.285 + ], + "angle": 0, + "content": "2.2. Unified Framework for Multi-modal Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.485, + 0.641 + ], + "angle": 0, + "content": "Multi-modal learning aims to train models that effectively process and relate information from various modalities. Substantial advancements have been made in this field, culminating in numerous studies focused on creating unified networks for the integration and processing of diverse modalities. The main work includes VLMO [46], MetaTransformer [66], Uni-Perceiver [67], and UniTR [21], etc. There are several methods that utilize a unified backbone design for multi-modal RS data. For example, OFA-Net [63] pre-trains a single transformer backbone on a curated multi-modal dataset using MIM. Han et al. proposed a RSFM called msGFM[22], which utilizes a shared transformer encoder while employing different decoders for each modality. In contrast to our approach, both OFA-Net and msGFM simply share all of parameters of backbone and learn feature representations for multi-modal data separately, which results in a lack of geographical alignment and fusion between features from different modalities. AnySat [2] employs a joint embedding predictive architecture for a multi-modal model, training a single model on heterogeneous data in a SSL manner. However, the backbone design and SSL pre-training methods of our SkySense V2 differ significantly from AnySat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.224, + 0.67 + ], + "angle": 0, + "content": "3. SkySense V2" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.279, + 0.693 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.355, + 0.715 + ], + "angle": 0, + "content": "3.1.1. Unified Transformer Backbone" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Figure 3, our unified transformer backbone is a hierarchical encoder structure with four stages. In the first two stages, we employ Swin Transformer V2 Blocks (SwinV2B) [32] to incorporate essential visual priors, such as locality and translation invariance. The window-based self-attention mechanisms in SwinV2B also help reduce computational complexity compared to global self-attention, particularly given the high spatial resolution of features in these initial stages. In the last two stages, we utilize vanilla Transformer Blocks (TB) [16] for two main reasons: First, the spatial resolution of features in these stages is relatively low, making the computational" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.095, + 0.915, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.214, + 0.872, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.345, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 3. Overview of the unified transformer backbone model in SkySense V2: The entire backbone shares all parameters across different modalities, with the exception of distinct tokenizers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.463 + ], + "angle": 0, + "content": "costs of global self-attention more manageable. Second, our ablation study demonstrates that global self-attention can be complementary to window-based self-attention, enabling the model to achieve stronger representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.629 + ], + "angle": 0, + "content": "Given a group of multi-modal inputs consisting of a high-resolution optical image \\( x_{HR} \\), multi-spectral data \\( x_{MS} \\), and synthetic aperture radar data \\( x_{SAR} \\) where each 'pixel' in the different modalities is naturally aligned by geo-location, we employ three distinct tokenizers to process these inputs into tokens. In each tokenizer, the input data is first divided into non-overlapping \\( 4 \\times 4 \\) patches. Subsequently, a linear embedding layer is applied to this raw, patched data to project it into patch tokens. Subsequently, four stages of either SwinV2B or TB, with shared parameters across modalities, are applied to these patch tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Due to the varying GSD of satellite sensors across different modalities, input data corresponding to the same region at specific geo-locations during joint aligned training exhibit differing spatial resolutions. To harmonize these spatial resolutions for tokens from various modalities, we propose Adaptive Patch Merging (APM) and utilize it to selectively reduce spatial resolution at each stage, except for Stage 1. Specifically, for the tokens of optical image data with high spatial resolution, APM concatenates the features from groups of \\(2 \\times 2\\) neighboring patches and applies a linear layer to the \\(4c\\)-dimensional concatenated features. This process reduces the number of tokens by a factor of \\(2 \\times 2 = 4\\) downsampling of spatial resolution, while the output dimension is set to \\(2c\\). In contrast, for the tokens of MS and SAR data with lower spatial resolution, APM maintains the resolution by applying a linear projection with averaging the weights across the input dimensions. By integrating the APM module, our unified backbone can efficiently process" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.098, + 0.493, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 4. The illustration of modality-specific prompt tokens added to the unified backbone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.351 + ], + "angle": 0, + "content": "multi-modal data with varying resolutions while maintaining the spatial alignment of features across modalities" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.367, + 0.373 + ], + "angle": 0, + "content": "3.1.2. Modality-specific Prompt Tokens" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Our unified transformer backbone utilizes fully shared parameters across different modalities. To enhance feature diversity, we introduce learnable modality-specific prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pre-trained model can better capture the unique characteristics of each modality. As illustrated in Figure 4, we incorporate \\(N\\) learnable prompt tokens for each modality in the last two stages. We denote the input tokens of each stage as \\(E_{i}^{j}\\in \\mathbb{R}^{h_{j}w_{j}\\times c_{j}}\\), \\(i\\in \\{HR,MS,SAR\\}\\), \\(j\\in \\{3,4\\}\\), where \\(h_j\\) and \\(w_{j}\\) are the height and width of the spatial resolution at stage \\(j\\), while \\(c_{j}\\) refers to the dimension of the tokens. For each modality \\(i\\), we insert the modality-specific learnable prompt tokens \\(P_{i}^{j}\\in \\mathbb{R}^{N\\times c_{j}}\\) into the input of the first block in \\(j\\)-th stage \\(\\mathcal{F}_j\\). At the last block of each stage, these prompt tokens are discarded. The entire process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.641, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ E _ {d r o p}, E _ {i} ^ {4} \\right] = \\mathcal {F} _ {3} \\left(\\left[ P _ {i} ^ {3}, E _ {i} ^ {3} \\right]\\right), \\tag {1} \\\\ [ E _ {d r o p}, E _ {i} ^ {o u t} ] = \\mathcal {F} _ {4} ([ P _ {i} ^ {4}, E _ {i} ^ {4} ]). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Here, \\( E_{drop} \\) denotes the discarded tokens, and \\( E_{i}^{out} \\) represents the final output tokens of backbone. Our design of modality-specific prompts enhances feature diversity while maintaining full parameter sharing by incorporating only a small number of modality-specific parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.329, + 0.792 + ], + "angle": 0, + "content": "3.1.3. Scaling Up Model Capacity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our unified backbone design across various modalities significantly enhances parameter utilization efficiency. To further boost the model's capabilities, we integrate a Mixture of Experts (MoE) [24] into the backbone network. We choose MoE instead of simply increasing the backbone's width and depth because MoE enables models to be pretrained with substantially less computation by leveraging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "sparse feed-forward layers (i.e., experts) for individual tokens. Following common approaches [33, 60], we insert MoE modules into the last \\(L\\) transformer blocks, replacing the original feed-forward network (FFN) layers. Each MoE module consists of \\(M\\) experts (denoted as \\(\\mathcal{E}_i(\\cdot), i = 1,2,\\dots,M\\)), which share the same structure as the FFN and operate as independent networks. For the gating network in MoE, we employ a learnable linear layer followed by a Softmax function, \\(\\mathcal{G}(x) = \\text{Softmax}(Wx)\\), where W is the gating parameter. Finally, the output of the MoE is computed as a linear combination of the outputs from the selected experts, weighted by the corresponding gate values. This process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.295, + 0.907, + 0.328 + ], + "angle": 0, + "content": "\\[\nM O E (x) = \\sum_ {i \\in \\mathcal {T}} \\mathcal {G} _ {i} (x) \\cdot \\mathcal {E} _ {i} (x), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.907, + 0.364 + ], + "angle": 0, + "content": "\\(\\mathcal{T}\\) represent the set of the top- \\(k\\) indices. In our SkySense V2 backbone, we set \\(L = 6,M = 8\\) , and \\(k = 1\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.776, + 0.389 + ], + "angle": 0, + "content": "3.2. Overall Pre-training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.712 + ], + "angle": 0, + "content": "Figure 5 provides an overview of our pre-training procedure. The pre-training framework of SkySense V2 primarily adopts the teacher-student architecture from SkySense [20], where the teacher network's parameters are updated using an exponential moving average (EMA) [42] of the student network's parameters. To train SkySense V2, we employ the Multi-Granularity Contrastive Learning (MGCL) loss \\(\\mathcal{L}_{MGCL}\\) as proposed in SkySense. For each modality, MGCL utilizes contrastive loss [42] to align the representations of teacher and student networks at pixel, object and image-level granularity, respectively. Additionally, we incorporate the unsupervised Geo-Context Prototype Learning (GCPL), as validated in SkySense, to enhance the learning of complementary regional context clues, aiding downstream tasks. To further enhance dense interpretation capabilities, we introduce an auxiliary supervision loss \\(\\mathcal{L}_{ITA}\\) to employ dense Image-Text Alignment (ITA) according to OpenStreetMap (OSM) labels1. Detailed implementations of MGCL, GCPL, and ITA are included in the Appendix A. In following paragraph, we will focus on proposed query-based semantic aggregation contrastive learning (QSACL)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.909, + 0.88 + ], + "angle": 0, + "content": "Query-based Semantic Aggregation Contrastive Learning. QSACL utilizes learnable queries to aggregate similar semantics across different regions of images, enabling more accurate contrastive learning. Given features \\( g_{1}, g_{2} \\) from two global views and \\( l_{1}, l_{2}, \\ldots, l_{n} \\) from \\( n \\) local views (where \\( g \\) and \\( l \\) denotes the fused features \\( F_{\\mathrm{fus}} \\) from global and local view), we employ a transformer decoder layer to perform cross-attention between \\( m \\) learnable queries \\( q_{1}, q_{2}, \\ldots, q_{m} \\) and the feature sets \\( g \\) and \\( l \\). This process generates sets of globally and locally aggregated features set" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.712, + 0.901 + ], + "angle": 0, + "content": "1https://www.openstreetmap.org/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9139" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.099, + 0.593, + 0.116 + ], + "angle": 0, + "content": "SkySense V2 Pre-training" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.113, + 0.862, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.402, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 5. Overview of the SkySense V2 pre-training pipeline. The SkySense V2 employs data augmentations on multi-modal inputs, which are then fed into both the student and teacher networks. The pre-training process incorporates Multi-Granularity Contrastive Learning, Dense Image-Text Alignment, and Query-based Semantic Aggregation Contrastive Learning to effectively train the network. Once pretraining is complete, the parameters from the teacher branch are utilized for downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.561 + ], + "angle": 0, + "content": "\\(z_{i}^{g}\\) and \\(z_{i}^{l}\\) for each \\(i = 1,2,\\ldots m\\). As is shown in Figure 6, we use superscript \\(\\prime\\) denotes features from the teacher network, the QSACL loss \\(\\mathcal{L}_{QSACL}\\) is computed by averaging the loss \\(\\mathcal{L}_{CL}\\) of aggregation features between local and global views for each query:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.572, + 0.483, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {Q S A C L} = \\frac {1}{2 m} \\sum_ {i = 1} ^ {m} \\left(\\mathcal {L} _ {C L} \\left(z _ {i} ^ {g}, z _ {i} ^ {l ^ {\\prime}}\\right) + \\mathcal {L} _ {C L} \\left(z _ {i} ^ {l}, z _ {i} ^ {g ^ {\\prime}}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.628, + 0.483, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C L} (x, x ^ {\\prime}) = - \\mathcal {H} (x) \\log \\left(\\mathcal {H} ^ {\\prime} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{H}\\) and \\(\\mathcal{H}'\\) denotes the learning head and corresponding EMA part as defined in[6]. The overall training objective is the weighted sum of \\(\\mathcal{L}_{MGCL}\\), \\(\\mathcal{L}_{ITA}\\), and \\(\\mathcal{L}_{QSACL}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.725, + 0.483, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M G C L} + \\lambda_ {2} \\mathcal {L} _ {I T A} + \\lambda_ {3} \\mathcal {L} _ {Q S A C L}. \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.748, + 0.481, + 0.858 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.869, + 0.483, + 0.898 + ], + "angle": 0, + "content": "Figure 6. Computation pipelines for QSACL with multi-crop augmentation (2 global and \\( n \\) local views)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.646, + 0.501 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Following SkySense [20], we conducted experiments on 16 datasets across various modalities and tasks to thoroughly evaluate SkySense V2. The applications of pre-trained Sky-Sense V2 to diverse tasks are detailed in Appendix B. Each module is engineered for combined or individual use, providing the flexibility to be frozen or fine-tuned as needed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.771, + 0.63 + ], + "angle": 0, + "content": "4.1. Pre-training Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.848 + ], + "angle": 0, + "content": "Training Datasets. We use the same training dataset of SkySense [20] to train SkySense V2. This extensive dataset comprises approximately 21 million multi-modal RS imagery sets. Each set comprises three components: a high-resolution optical RGB image of spatial dimension \\(2048 \\times 2048\\), a sequence of Sentinel-2 (S2) multi-spectral (MS) images with 10 spectral bands and spatial dimensions of \\(64 \\times 64\\) (average sequence length of 65 images), and a sequence of Sentinel-1 (S1) synthetic aperture radar (SAR) images with 2 bands, also having spatial dimensions of \\(64 \\times 64\\) (average sequence length of 10 images). In each training iteration, we randomly sample 10 S2 images and 5 S1 images of different time series to accommodate the GPU memory constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Setting of Unified Backbone. We initially set the token dimension to \\( C = 352 \\). In the first two stages, the win-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9140" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.213 + ], + "angle": 0, + "content": "dow size for the Swin Transformer V2 Blocks (SwinV2B) is configured to 8. Each head has a query dimension of 32, and each MLP layer consists of two full-connected layers with dimensions expanding by a factor of 4 consistently across all blocks within the unified backbone. Prompt tokens are incorporated during the last two stages, with the number of prompt tokens being the same for each modality. We allocate 4 prompt tokens for stage-3 and stage-4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Pre-training Settings. SkySense V2 is trained using a batch size of 1024, distributed across 128 H20 GPUs. The model undergoes a total of 600k iterations, utilizing the AdamW optimizer [35]. The learning rate is initially set to \\(2 \\times 10^{-4}\\) and decays to \\(1 \\times 10^{-6}\\) following a cosine annealing schedule [34]. Additional implementation details can be found in the Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.279, + 0.36 + ], + "angle": 0, + "content": "4.2. Single-Modal Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.457 + ], + "angle": 0, + "content": "We evaluate SkySense V2 on four common single-modal RS interpretation tasks: scene classification, object detection, semantic segmentation, and change detection. The experiments are conducted via fine-tuning a pre-trained Sky-Sense V2 backbone along with a task-specific head. Further implementation details can be found in the Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Scene Classification. In Table 1, we report the performance of scene classification on four widely used datasets: AID [61] and RESISC-45 [11], both featuring static RGB optical images; BEN-S2 [48], which includes static MS images; and fMoW-S2 [13], which contains temporal MS images. For each dataset, the performance is evaluated using different training ratios (TR) following the methodology [13, 38]. In our experiments, we employ a simple linear classifier as the classification head. For multi-label datasets, we use the mean average precision (mAP) as the evaluation metric, whereas for single-label datasets, we report the overall accuracy (OA). As illustrated in Table 1, SkySense V2 consistently outperforms previous RSFMs across various datasets and training configurations, particularly in low TR settings. This demonstrates that SkySense V2 exhibits a stronger representation capability compared to other RSFMs, including its predecessor, SkySense [20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Semantic Segmentation. In Table 2a, we report the segmentation results from four representative semantic segmentation datasets: iSAID [59] and Potsdam [47] for high-resolution optical image segmentation, and Dyna.-Pla [53] and Dyna.-S2 [53] for multi-spectral and multi-temporal image segmentation. Following previous work, we use the mean F1-score (mF1) as the evaluation metric for the Potsdam dataset. While for the other datasets, we report the mean intersection over union (mIoU). All segmentation experiments employ the UperNet [62] as the decoder head." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.345 + ], + "angle": 0, + "content": "
ModelsSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [3]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [38]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [13]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo [50]96.90/98.3494.25/95.67--
RVSA [55]97.03/98.5093.93/95.69--
DINO-MC [58]--84.20/88.7560.16/83.49
TOV [52]95.16/97.0990.97/93.79--
SSL4EO [57]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [40]96.11/97.7994.05/95.53--
CACo [37]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMA [18]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [39]95.47/97.0992.73/94.6486.30/--
Scale-MAE [43]96.44/97.5892.63/95.04--
MA3E [30]-/99.04-/96.23--
SatMAE++ [41]--/97.4885.11/-63.23/-
SkySense [20]97.68/98.6094.85/96.3288.67/92.0964.38/87.27
SkySense V298.34/99.0596.42/97.2489.13/93.7866.65/89.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.356, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Table 1. Scene classification results. - means the task is not supported or the value is unavailable in the paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.905, + 0.477 + ], + "angle": 0, + "content": "As illustrated in Table 2a, our SkySense V2 achieve higher performance than recent RSFMs across all four datasets. Specifically, compared to the previous SOTA method SkySense [20], SkySense V2 achieves \\(1.5\\%\\) higher performance on average." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.68 + ], + "angle": 0, + "content": "Horizontal & Oriented Object Detection. We utilize the DIOR dataset [27] to evaluate SkySense V2's performance in horizontal object detection, while utilizing the DIOR-R [12] and FAIR1M [51] datasets for oriented object detection. These datasets only consist of RGB RS imagery. Consistent with previous research [20], we adopt Faster R-CNN [44] and Oriented R-CNN [28] as the basic detectors for horizontal and oriented object detection, respectively. As reported in Table 2b, our SkySense V2 outperforms previous RSFMs by a notable margin. Specifically, SkySense V2 surpasses the previous SOTA, SkySense [20], by an average of \\(1.1\\%\\) mAP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Change Detection. For change detection, we conducted experiments on LEVIR-CD [8], OSCD [15] and Dyna.-S2 [53] datasets. Since LEVIR-CD and OSCD datasets only contain binary change detection results, we use the basic change detector in [9] and report the F1 score as the evaluation metric. While Dyna.-S2 focuses on semantic change detection, we use UperNet [62] as the segmentation network and calculate the semantic change detection score (SCS) on validation and test set following [20]. As presented in Table 2c, SkySense V2 consistently achieves higher performance than the previous SOTA method SkySense on all 3 datasets, and surpasses SkySense by \\(2.7\\%\\) on average. Particularly, on the OSCD dataset, our SkySense V2 remark" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9141" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.426, + 0.308 + ], + "angle": 0, + "content": "
ModelsPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [3]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [38]ICCV'21-57.2089.0329.4/39.8
SatMAE [13]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo [50]TGRS'22-67.2091.27-
RVSA [55]TGRS'2234.3/44.464.49--
BFM [7]Arxiv'23--92.12-
TOV [52]JSTARS'2332.1/37.866.2492.03-
SSL4EO [57]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [40]TGRS'2336.4/43.566.2191.86-
CACo [37]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS [56]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [39]ICCV'2336.7/45.666.6291.85-
Scale-MAE [43]ICCV'2334.0/41.765.7791.54-
MA3E [30]ECCV'24-64.0691.50-
SkySense [20]CVPR'2439.7/46.570.9193.9933.1/46.2
SkySense V2-41.2/47.671.8795.8635.2/47.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.165, + 0.311, + 0.347, + 0.323 + ], + "angle": 0, + "content": "(a) Semantic segmentation results." + }, + { + "type": "table", + "bbox": [ + 0.43, + 0.091, + 0.663, + 0.307 + ], + "angle": 0, + "content": "
ModelsHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [3]67.4065.6548.15
SatMAE [13]70.8965.6646.55
RingMo [50]75.90-46.21
RVSA [55]73.2271.0547.04
BFM [7]-73.62-
TOV [52]70.1666.3349.62
SSL4EO [57]64.8261.2349.37
CMID [40]75.1166.3750.58
CACo [37]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [39]72.8467.6749.69
Scale-MAE [43]73.8166.4748.31
MA3E [30]-71.82-
SkySense [20]78.7374.2754.57
SkySense V279.5075.2955.96
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.47, + 0.309, + 0.618, + 0.321 + ], + "angle": 0, + "content": "(b) Object detection results." + }, + { + "type": "table", + "bbox": [ + 0.668, + 0.093, + 0.906, + 0.304 + ], + "angle": 0, + "content": "
ModelsLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [3]78.1946.2613.6/16.7
SeCo [38]90.1447.6713.9/16.0
SatMAE [13]87.6552.7614.8/16.2
RingMo [50]91.86--
RVSA [55]90.86--
SpectralGPT [23]-54.29-
MATTER [1]-59.37-
DINO-MC [58]-52.7014.5/15.6
SSL4EO [57]89.0535.0812.3/17.5
CMID [40]91.72--
CACo [37]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [39]91.7359.82-
Scale-MAE [43]92.07--
SkySense [20]92.5860.0615.4/18.0
SkySense V294.8365.2916.0/20.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.708, + 0.306, + 0.859, + 0.318 + ], + "angle": 0, + "content": "(c) Change detection results." + }, + { + "type": "table_caption", + "bbox": [ + 0.255, + 0.335, + 0.741, + 0.35 + ], + "angle": 0, + "content": "Table 2. Results of semantic segmentation, object detection and change detection." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.359, + 0.472, + 0.499 + ], + "angle": 0, + "content": "
Task & DatasetData SourceSkySense[20]SkySense V2
(a) Multi-modal Seg: Dyna.-MM(i) Planet.46.547.6
(ii) S246.247.5
(iii) Planet. + S247.348.7
(iv) Planet. + S2 + S147.748.9
(b) Multi-modal Seg: PASTIS-MM(i) S273.575.0
(ii) S2-Ts84.685.5
(iii) S2-Ts + S1-Ts84.885.6
(iv) S2-Ts + GEP85.886.7
(c) Multi-modal Cls: BEN-MM(i) S186.286.5
(ii) S2 + S192.293.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.138, + 0.51, + 0.434, + 0.523 + ], + "angle": 0, + "content": "Table 3. Fine-tuning results on multi-modal tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.543, + 0.411, + 0.558 + ], + "angle": 0, + "content": "ably outperform SkySense by \\(5.2\\%\\) on F1-score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.569, + 0.273, + 0.584 + ], + "angle": 0, + "content": "4.3. Multi-modal Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.653 + ], + "angle": 0, + "content": "In this section, we evaluate SkySense V2 on two representative multi-modal tasks, i.e., Multi-modal Segmentation and Multi-modal Scene Classification, to demonstrate its generalization capability across data from various modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-modal Segmentation. In Table 3 (a) and (b), we present the performance results of Dyna.-MM [53] and PASTIS-MM [19], respectively. Dyna.-MM contains high-resolution optical imagery from PlanetFusion (Planet.), multi-spectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). For the experiments conducted on Dyna.-MM, we employ a basic UperNet [62] as the decoder head and utilize mIoU as the evaluation metric. As depicted in Table 3 (a), SkySense V2 consistently outperforms the SkySense across various modalities, demonstrating a stronger representation ability. Furthermore, integrating multiple modalities yields higher performance compared to using a single modality. This results confirms that the proposed SkySense V2 can effectively extract representations from diverse modalities using a unified backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.558 + ], + "angle": 0, + "content": "PASTIS-MM is a crop mapping dataset that contains high-resolution optical imagery from Google Earth Pro (GEP), multi-temporal multi-spectral imagery from Sentinel-2 (S2-Ts), and multi-spectral SAR imagery from Sentinel-1 (S1-Ts). We employ a straightforward FCN head to decode the semantic segmentation results and report the overall accuracy as the evaluation metric in Table 3 (b). The results demonstrate that SkySense V2 outperforms the previous SOTA method SkySense by an average of \\(1.0\\%\\). Additionally, it is worth noting that S2-Ts can remarkably improve the segmentation accuracy compared to S2, underscoring the importance of temporal information for crop mapping." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.687 + ], + "angle": 0, + "content": "Multi-modal Scene Classification. We conduct multimodal scene classification experiments on the BEN-MM dataset [49]. As is shown in Table 3 (c), the proposed SkySense V2 surpasses SkySense, demonstrating a stronger representation ability. Furthermore, by integrating S1 and S2 imagery, SkySense V2 achieves higher improvements, highlighting its enhanced ability to extract superior representations from different modalities compared to SkySense." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.736, + 0.713 + ], + "angle": 0, + "content": "4.4. Ablation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation Study of Components. To understand the contributions of each component, we conduct a k-nearest neighbors (k-NN) evaluation on the AID and RESISC-45 datasets. Figure 7 illustrates the performance of different component combinations across various pre-training iterations. Notably, the unified backbone design significantly accelerates representation learning. This is primarily due to shared parameters across different modalities, which allow gradients to aggregate, thereby speeding up convergence. Moreover, the unified design allows the backbone to be trained with data from different modalities, thereby enhancing the model's generalization capabilities (see Ap" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.287, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.198, + 0.233, + 0.21 + ], + "angle": 0, + "content": "(a) AID dataset" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.09, + 0.48, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.197, + 0.446, + 0.209 + ], + "angle": 0, + "content": "(b) RESISC-45 dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.334 + ], + "angle": 0, + "content": "Figure 7. Ablation study using k-NN classification was conducted on the AID and RESISC-45 datasets. In this context, the abbreviations have the following meanings: GA refers to replacing window-based self-attention with global self-attention in the last two stages; MsP involves the addition of modality-specific prompt tokens; MoE indicates scaling up the model through a mixture of experts approach; and QSACL means adding additional query-based semantic aggregation contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.481 + ], + "angle": 0, + "content": "pendix F.3 for more details). The results also indicate that global self-attention complements window-based self-attention, facilitating improved representations. Additionally, modality-specific prompt tokens and MoE scaling enhance representation learning by increasing feature diversity and model capacity, respectively. QSACL enhances model performance by improving the semantic precision of features in contrastive learning." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.492, + 0.29, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.607, + 0.279, + 0.62 + ], + "angle": 0, + "content": "(a) w/o modality-specific prompt" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.492, + 0.481, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.607, + 0.469, + 0.62 + ], + "angle": 0, + "content": "(b) w/ modality-specific prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Figure 8. t-SNE visualization of the output feature representation from the final block of the unified transformer backbone. HROI, SARI, and MSI represent the features from high-resolution optical imagery, synthetic aperture radar imagery, and multi-spectral imagery, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "How Do Modality-Specific Prompt Tokens Facilitate Pre-training? To investigate the impact of modality-specific prompt (MsP) tokens on the multi-modal pretraining process within a unified backbone, we employed t-SNE [54] for visualizing the features corresponding to different modalities. These features are extracted from the last block of the unified transformer backbone, and the comparative results are presented in Figure 8. Without the use of MsP, the features of high-resolution optical imagery (HROI) and synthetic aperture radar imagery (SARI) appear closely clustered, leading to overlapping feature distri" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "butions. However, with the inclusion of MsP, the features of HROI and SARI become distinctly separable. These visualization outcomes clearly demonstrate that MsP effectively enhances feature diversity and imparts modality-specific characteristics." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.184, + 0.71, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.185, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.238, + 0.71, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.238, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.3, + 0.906, + 0.342 + ], + "angle": 0, + "content": "Figure 9. Visualization of cross-attention weights for features patch corresponding to different queries in QSACL. Queries can effectively aggregate features with specific semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.66 + ], + "angle": 0, + "content": "What Types of Features Does Query-based Semantic Aggregation Contrastive Learning Capture? To provide a clearer understanding of our proposed QSACL, we visualize the attention weights assigned to different features during their interaction with certain query, as illustrated in Figure 9. During pre-training, we use two global augmentation crops and six local augmentation crops. Here, we present two distinct queries, each focusing on different semantic features of image patches. For example, one query focuses on features with building semantics (left part of the first row in Figure 9), while another focuses on features with land semantics (right part of the first row in Figure 9). The attention weights of different queries aggregate features from each patch into distinct feature representations. Subsequently, we perform contrastive learning between aggregated features from the same query across teacher and student models. This approach significantly enhances semantic accuracy compared to traditional contrastive learning applied to entire image features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduce SkySense V2, a MM-RSFM that utilizes a unified backbone to accommodate various modalities. This unified approach enhances parameter utilization efficiency and improves the model's generalization ability. SkySense V2 is pre-trained with an innovative QSACL strategy, specifically designed to leverage the unique characteristics of RS images. As a result, SkySense V2 significantly boosts performance while maintaining the multi-modal flexibility benefits of its predecessor, SkySense. Looking ahead, our future work will focus on integrating the language modality and incorporating a geographical knowledge graph to develop a more powerful and versatile MM-RSFM." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9143" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. AnySat: An Earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[7] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[8] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[9] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[10] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[11] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[12] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[13] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[17] William Fedus, Barret Zoph, and Noam M. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ArXiv, abs/2101.03961, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.418 + ], + "angle": 0, + "content": "[18] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[19] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.491, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[20] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. Skysense: A multimodal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27672-27683, 2024. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[21] Wang Haiyang, Tang Hao, Shi Shaoshuai, Li Aoxue, Li Zhenguo, Schiele Bernt, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[22] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27852-27862, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[23] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Pedram Ghamisi, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, (8):5227-5244, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3:79–87, 1991. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Pallavi Jain, Bianca Schoen-Phelan, and Robert J. Ross. Self-supervised learning for invariant representations from" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9144" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "multi-spectral and sar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 7797-7808, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reppoints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[29] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24088-24097, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[30] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images. In European Conference on Computer Vision, pages 260-278. Springer, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[32] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[33] Zhili Liu, Kai Chen, Jianhua Han, Lanqing Hong, Hang Xu, Zhenguo Li, and James Tin-Yau Kwok. Task-customized masked autoencoder via mixture of cluster-conditional experts. ArXiv, abs/2402.05382, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. CoRR, abs/1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[36] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Oscar Manas, Alexandre Lacoste, Xavier Giro-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[39] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[40] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[41] Mubashir Noman, Muzammal Naseer, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, and Fahad Shahbaz Khan. Rethinking transformers pre-training for multispectral satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27811-27819, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[42] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[43] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[45] Sheng Shen, Le Hou, Yan-Quan Zhou, Nan Du, S. Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[46] Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Y Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[47] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[48] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Felipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9145" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[49] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[50] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[51] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fairlm: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[52] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[53] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[54] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[55] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[56] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[57] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[58] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.898 + ], + "angle": 0, + "content": "[59] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[60] Lemeng Wu, Mengchen Liu, Yinpeng Chen, Dongdong Chen, Xiyang Dai, and Lu Yuan. Residual mixture of experts. ArXiv, abs/2204.09636, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[61] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[62] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[63] Zhitong Xiong, Yi Wang, Fahong Zhang, and Xiao Xiang Zhu. One for all: Toward unified foundation models for earth vision. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 2734-2738, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[64] Longqi Yang, Liangliang Zhang, and Wenjing Yang. Graph adversarial self-supervised learning. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[65] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[66] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[67] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uniperceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16783-16794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[68] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam M. Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9146" + } + ] +] \ No newline at end of file diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_origin.pdf b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55cfc28d4f77cd8890e40b90649abc6babbc43e3 --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/5fb33625-ede3-48c2-9ec0-3741fd1fba6c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:824992d5a659acc65f610aec1410042e398b598d3731478c6ee33f97a65e691b +size 1472418 diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/full.md b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bd4b9f30e19466e3ee64d96670746116846b68f4 --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/full.md @@ -0,0 +1,302 @@ +# SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing + +Yingying Zhang $^{1}$ Lixiang Ru $^{1}$ Kang Wu $^{2}$ Lei Yu $^{1}$ Lei Liang $^{1}$ Yansheng Li $^{2}$ Jingdong Chen $^{1}$ $^{1}$ Ant Group $^{2}$ Wuhan University qichu.zyy@antgroup.com + +# Abstract + +The multi-modal remote sensing foundation model (MM-RSFM) has significantly advanced various Earth observation tasks, such as urban planning, environmental monitoring, and natural disaster management. However, most existing approaches generally require the training of separate backbone networks for each data modality, leading to redundancy and inefficient parameter utilization. Moreover, prevalent pre-training methods typically apply self-supervised learning (SSL) techniques from natural images without adequately accommodating the characteristics of remote sensing (RS) images, such as the complicated semantic distribution within a single RS image. In this work, we present SkySense V2, a unified MM-RSFM that employs a single transformer backbone to handle multiple modalities. This backbone is pre-trained with a novel SSL strategy tailored to the distinct traits of RS data. In particular, SkySense V2 incorporates an innovative adaptive patch merging module and learnable modality prompt tokens to address challenges related to varying resolutions and limited feature diversity across modalities. In additional, we incorporate the mixture of experts (MoE) module to further enhance the performance of the foundation model. SkySense V2 demonstrates impressive generalization abilities through an extensive evaluation involving 16 datasets over 7 tasks, outperforming SkySense by an average of 1.8 points. + +# 1. Introduction + +Advancements in multi-modal remote sensing foundation models (MM-RSFM) have significantly enhanced the utilization of remote sensing (RS) data across numerous Earth observation (EO) applications[5, 10, 36, 65], such as environmental monitoring, agriculture management, natural disaster response, land cover (or use) mapping etc. Recently, Guo et al. introduced SkySense [20], the largest MM-RSFM to date. SkySense has demonstrated impressive generalization capabilities through extensive evaluations across many different tasks. A key factor in its success is the simultaneous training of geographically aligned multi-modal + +![](images/8a82f94c635106848da56b5529fc72525d5d6555cc4f85d6d7f6647edec4eda9.jpg) +Figure 1. SkySense V2 has achieved superior performance on 16 datasets over 7 distinct tasks compared with its predecessor Sky-Sense and supports a board range of tasks. + +and multi-temporal RS data. To accommodate the varying resolutions of different modalities, SkySense employs the Swin Transformer [31] for processing high-resolution (HR) optical images, while using the Vision Transformer (ViT) [16] for medium-resolution multi-spectral (MS) data and synthetic aperture radar (SAR) data. However, this separate backbone design results in redundancy and inefficient use of parameters. On the other hand, SkySense's pre-training approach is primarily based on DINOv2 [42], renowned for its efficiency in self-supervised learning (SSL). Although it has demonstrated strong performance with RS images, there remains considerable room for improvement by adequately accommodating the involved semantic distribution within a single RS image. + +In this paper, we introduce SkySense V2, an advanced MM-RSFM pre-trained through a novel SSL approach tailored to the distinctive traits of RS data. To be specific, SkySense V2 leverages a unified transformer backbone capable of processing data across diverse modalities and resolutions. It is pre-trained using a new SSL strategy that utilizes query-based attention to gather the similar semantic features distributed in different regions of RS images. + +Designing a unified backbone for geographically aligned multi-modal data presents two primary challenges. The first is processing feature resolutions across data with varying ground sample distance (GSD) during simultaneous pretraining of multiple modalities. To address this challenge, we introduce an innovative Adaptive Patch Merging (APM) module, which is integrated after each stage of the unified backbone. This module determines whether to reduce feature resolution according to the specific feature resolution requirements of each modality. For instance, when dealing with HR optical images, the APM module activates resolution reduction at each stage. In contrast, for medium-resolution data, such as MS and SAR data, the feature resolution is preserved throughout the stages. This design enables the backbone to efficiently process data of different resolutions while adapting output feature resolutions according to actual requirements, which is crucial for the joint training and fusion of multi-modal features. + +The second challenge is that fully sharing parameters across the different modalities can reduce the feature diversity. To address this, inspired by the work of [14] and [26], we introduce learnable modality prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pretrained model can capture the unique characteristics of each modality better. Compared to SkySense, our unified design significantly enhances the efficiency of parameter utilization. To be specific, SkySense uses three separate backbones, Swin-H for HR optical data, ViT-L for MS data, and ViT-L for SAR data, collectively totaling 1.26 billion parameters. In contrast, SkySense V2's unified transformer backbone is capable of simultaneously processing all three modalities while utilizing only 665 million parameters. This efficient use of parameters not only optimizes the architecture but also allows for potential model scaling to further boost performance. Therefore, we integrate the mixture of experts module [24], a technique commonly used in various large language models [17, 45, 68] to further enhance the performance of the foundation model. + +To effectively adapt SSL techniques developed for natural images to the pre-training of RSFM, it is essential to consider the distinct differences in data characteristics. Natural images typically feature a single, clear subject, such as a cat, dog, or person. In contrast, RS images comprise multiple subjects distributed across various regions of the image. For instance, as illustrated in Figure 2, one section of an RS image may show a building, while other areas might show a forest, pond, or land. Current SSL methods perform contrastive learning directly between different views using various augmented crops, which is effective for natural images due to their singular focus. However, when applied to RS images, this approach can result in semantic inaccuracies, as different views may capture different subjects. For + +![](images/01538eb678699ee9904ed6057dca4de016e8849455a461b9bd592ff9d08c68b1.jpg) +(a) Views of natural image +Figure 2. Comparison of different augmented views from natural and RS images in SSL. Natural image views often concentrate on a singular subject, such as a "dog". RS image views encompass a variety of subjects, including building, forest, pond, and land. + +![](images/7e60dacfbae18e7330c45c0796f26ae6463f9e18b86279806a71a496a7bf06c5.jpg) +(b) Views of RS image + +example, in Figure 2b, view 1 of the RS image primarily contains building, whereas view 2 shows forest and pond. + +To tackle this challenge, we propose a novel approach called Query-based Semantic Aggregation Contrastive Learning (QSACL). This method employs multiple learnable queries to perform cross-attention with features from different views, generating semantically aggregated features. We then apply contrastive learning to these aggregated features pairs derived from the same query. As demonstrated in Figure 9 from our ablation study, utilizing different queries can aggregate consistent semantic features across multiple views, which enhances the accuracy of contrastive learning. + +We evaluated SkySense V2 on a diverse set of 16 datasets [8, 11-13, 15, 19, 27, 48, 51, 53, 59, 61], covering a range of task types, modalities, and spatial scales. As illustrated in Figure 1, SkySense V2 demonstrates substantial improvements in performance over its predecessor, SkySense, achieving state-of-the-art (SOTA) results across various modalities of EO tasks. The experimental results across all test scenarios highlight its competitive advantage over existing RSFMs in a wide array of EO interpretation tasks. + +# 2. Related Work + +# 2.1. Remote Sensing Foundation Model + +RSFMs are characterized by their ability to leverage vast amounts of data through self-supervised learning (SSL) technologies, enabling them to learn robust feature representations without the need for extensive annotations. The predominant RSFMs utilize Contrastive Learning (CL) or Masked Image Modeling (MIM). Typical works in CL include RS-BYOL [25], GASSL [64], DINO-MC [58], SeCo [38], and CACo [37]. Other research efforts focus on enhancing MIM framework, such as RingMo [50], S2MAE [29], MA3E [30], SatMAE [13], and SatMAE++ [41]. Additionally, studies like CMID [40] and GFM [39] explore the intersection of CL and MIM through a self-distillation + +approach. Recently, Guo et al. introduced a comprehensive MM-RSFM known as SkySense [20], which features a factorized multi-modal spatiotemporal encoder. This architecture facilitates independent spatial feature extraction and multi-modal temporal fusion. As the largest MM-RSFM to date, SkySense exhibits exceptional generalization capabilities across a wide range of RS datasets. However, it does face a challenge: the inefficient utilization of parameters due to the separate backbone designs for different modalities. In this work, we propose SkySense V2 to address this issue by employing a unified transformer backbone design. + +# 2.2. Unified Framework for Multi-modal Learning + +Multi-modal learning aims to train models that effectively process and relate information from various modalities. Substantial advancements have been made in this field, culminating in numerous studies focused on creating unified networks for the integration and processing of diverse modalities. The main work includes VLMO [46], MetaTransformer [66], Uni-Perceiver [67], and UniTR [21], etc. There are several methods that utilize a unified backbone design for multi-modal RS data. For example, OFA-Net [63] pre-trains a single transformer backbone on a curated multi-modal dataset using MIM. Han et al. proposed a RSFM called msGFM[22], which utilizes a shared transformer encoder while employing different decoders for each modality. In contrast to our approach, both OFA-Net and msGFM simply share all of parameters of backbone and learn feature representations for multi-modal data separately, which results in a lack of geographical alignment and fusion between features from different modalities. AnySat [2] employs a joint embedding predictive architecture for a multi-modal model, training a single model on heterogeneous data in a SSL manner. However, the backbone design and SSL pre-training methods of our SkySense V2 differ significantly from AnySat. + +# 3. SkySense V2 + +# 3.1. Model Architecture + +# 3.1.1. Unified Transformer Backbone + +As illustrated in Figure 3, our unified transformer backbone is a hierarchical encoder structure with four stages. In the first two stages, we employ Swin Transformer V2 Blocks (SwinV2B) [32] to incorporate essential visual priors, such as locality and translation invariance. The window-based self-attention mechanisms in SwinV2B also help reduce computational complexity compared to global self-attention, particularly given the high spatial resolution of features in these initial stages. In the last two stages, we utilize vanilla Transformer Blocks (TB) [16] for two main reasons: First, the spatial resolution of features in these stages is relatively low, making the computational + +![](images/5da551ae3edf10f05bd3a5f8d8ade3e925a500f1bab788fb4533d966a80b63ac.jpg) + +![](images/81f83f84530a2ece1a0a9856e9aeae73bb9c474b601085994f6c8743748e02e1.jpg) +Figure 3. Overview of the unified transformer backbone model in SkySense V2: The entire backbone shares all parameters across different modalities, with the exception of distinct tokenizers. + +costs of global self-attention more manageable. Second, our ablation study demonstrates that global self-attention can be complementary to window-based self-attention, enabling the model to achieve stronger representations. + +Given a group of multi-modal inputs consisting of a high-resolution optical image $x_{HR}$ , multi-spectral data $x_{MS}$ , and synthetic aperture radar data $x_{SAR}$ where each 'pixel' in the different modalities is naturally aligned by geo-location, we employ three distinct tokenizers to process these inputs into tokens. In each tokenizer, the input data is first divided into non-overlapping $4 \times 4$ patches. Subsequently, a linear embedding layer is applied to this raw, patched data to project it into patch tokens. Subsequently, four stages of either SwinV2B or TB, with shared parameters across modalities, are applied to these patch tokens. + +Due to the varying GSD of satellite sensors across different modalities, input data corresponding to the same region at specific geo-locations during joint aligned training exhibit differing spatial resolutions. To harmonize these spatial resolutions for tokens from various modalities, we propose Adaptive Patch Merging (APM) and utilize it to selectively reduce spatial resolution at each stage, except for Stage 1. Specifically, for the tokens of optical image data with high spatial resolution, APM concatenates the features from groups of $2 \times 2$ neighboring patches and applies a linear layer to the $4c$ -dimensional concatenated features. This process reduces the number of tokens by a factor of $2 \times 2 = 4$ downsampling of spatial resolution, while the output dimension is set to $2c$ . In contrast, for the tokens of MS and SAR data with lower spatial resolution, APM maintains the resolution by applying a linear projection with averaging the weights across the input dimensions. By integrating the APM module, our unified backbone can efficiently process + +![](images/6fad7f3c67a3374040e087878f6577842cdc793c1fc760e1ddea740e1950af2c.jpg) +Figure 4. The illustration of modality-specific prompt tokens added to the unified backbone. + +multi-modal data with varying resolutions while maintaining the spatial alignment of features across modalities + +# 3.1.2. Modality-specific Prompt Tokens + +Our unified transformer backbone utilizes fully shared parameters across different modalities. To enhance feature diversity, we introduce learnable modality-specific prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pre-trained model can better capture the unique characteristics of each modality. As illustrated in Figure 4, we incorporate $N$ learnable prompt tokens for each modality in the last two stages. We denote the input tokens of each stage as $E_{i}^{j}\in \mathbb{R}^{h_{j}w_{j}\times c_{j}}$ , $i\in \{HR,MS,SAR\}$ , $j\in \{3,4\}$ , where $h_j$ and $w_{j}$ are the height and width of the spatial resolution at stage $j$ , while $c_{j}$ refers to the dimension of the tokens. For each modality $i$ , we insert the modality-specific learnable prompt tokens $P_{i}^{j}\in \mathbb{R}^{N\times c_{j}}$ into the input of the first block in $j$ -th stage $\mathcal{F}_j$ . At the last block of each stage, these prompt tokens are discarded. The entire process can be formulated as: + +$$ +\begin{array}{l} \left[ E _ {d r o p}, E _ {i} ^ {4} \right] = \mathcal {F} _ {3} \left(\left[ P _ {i} ^ {3}, E _ {i} ^ {3} \right]\right), \tag {1} \\ [ E _ {d r o p}, E _ {i} ^ {o u t} ] = \mathcal {F} _ {4} ([ P _ {i} ^ {4}, E _ {i} ^ {4} ]). \\ \end{array} +$$ + +Here, $E_{drop}$ denotes the discarded tokens, and $E_{i}^{out}$ represents the final output tokens of backbone. Our design of modality-specific prompts enhances feature diversity while maintaining full parameter sharing by incorporating only a small number of modality-specific parameters. + +# 3.1.3. Scaling Up Model Capacity + +Our unified backbone design across various modalities significantly enhances parameter utilization efficiency. To further boost the model's capabilities, we integrate a Mixture of Experts (MoE) [24] into the backbone network. We choose MoE instead of simply increasing the backbone's width and depth because MoE enables models to be pretrained with substantially less computation by leveraging + +sparse feed-forward layers (i.e., experts) for individual tokens. Following common approaches [33, 60], we insert MoE modules into the last $L$ transformer blocks, replacing the original feed-forward network (FFN) layers. Each MoE module consists of $M$ experts (denoted as $\mathcal{E}_i(\cdot), i = 1,2,\dots,M$ ), which share the same structure as the FFN and operate as independent networks. For the gating network in MoE, we employ a learnable linear layer followed by a Softmax function, $\mathcal{G}(x) = \text{Softmax}(Wx)$ , where W is the gating parameter. Finally, the output of the MoE is computed as a linear combination of the outputs from the selected experts, weighted by the corresponding gate values. This process is formulated as follows: + +$$ +M O E (x) = \sum_ {i \in \mathcal {T}} \mathcal {G} _ {i} (x) \cdot \mathcal {E} _ {i} (x), \tag {2} +$$ + +$\mathcal{T}$ represent the set of the top- $k$ indices. In our SkySense V2 backbone, we set $L = 6,M = 8$ , and $k = 1$ + +# 3.2. Overall Pre-training Pipeline + +Figure 5 provides an overview of our pre-training procedure. The pre-training framework of SkySense V2 primarily adopts the teacher-student architecture from SkySense [20], where the teacher network's parameters are updated using an exponential moving average (EMA) [42] of the student network's parameters. To train SkySense V2, we employ the Multi-Granularity Contrastive Learning (MGCL) loss $\mathcal{L}_{MGCL}$ as proposed in SkySense. For each modality, MGCL utilizes contrastive loss [42] to align the representations of teacher and student networks at pixel, object and image-level granularity, respectively. Additionally, we incorporate the unsupervised Geo-Context Prototype Learning (GCPL), as validated in SkySense, to enhance the learning of complementary regional context clues, aiding downstream tasks. To further enhance dense interpretation capabilities, we introduce an auxiliary supervision loss $\mathcal{L}_{ITA}$ to employ dense Image-Text Alignment (ITA) according to OpenStreetMap (OSM) labels1. Detailed implementations of MGCL, GCPL, and ITA are included in the Appendix A. In following paragraph, we will focus on proposed query-based semantic aggregation contrastive learning (QSACL). + +Query-based Semantic Aggregation Contrastive Learning. QSACL utilizes learnable queries to aggregate similar semantics across different regions of images, enabling more accurate contrastive learning. Given features $g_{1}, g_{2}$ from two global views and $l_{1}, l_{2}, \ldots, l_{n}$ from $n$ local views (where $g$ and $l$ denotes the fused features $F_{\mathrm{fus}}$ from global and local view), we employ a transformer decoder layer to perform cross-attention between $m$ learnable queries $q_{1}, q_{2}, \ldots, q_{m}$ and the feature sets $g$ and $l$ . This process generates sets of globally and locally aggregated features set + +![](images/70a45e8d4136055737c6aa275dd9546d43ce537f250f09d2f596ef1fea283769.jpg) +SkySense V2 Pre-training +Figure 5. Overview of the SkySense V2 pre-training pipeline. The SkySense V2 employs data augmentations on multi-modal inputs, which are then fed into both the student and teacher networks. The pre-training process incorporates Multi-Granularity Contrastive Learning, Dense Image-Text Alignment, and Query-based Semantic Aggregation Contrastive Learning to effectively train the network. Once pretraining is complete, the parameters from the teacher branch are utilized for downstream applications. + +$z_{i}^{g}$ and $z_{i}^{l}$ for each $i = 1,2,\ldots m$ . As is shown in Figure 6, we use superscript $\prime$ denotes features from the teacher network, the QSACL loss $\mathcal{L}_{QSACL}$ is computed by averaging the loss $\mathcal{L}_{CL}$ of aggregation features between local and global views for each query: + +$$ +\mathcal {L} _ {Q S A C L} = \frac {1}{2 m} \sum_ {i = 1} ^ {m} \left(\mathcal {L} _ {C L} \left(z _ {i} ^ {g}, z _ {i} ^ {l ^ {\prime}}\right) + \mathcal {L} _ {C L} \left(z _ {i} ^ {l}, z _ {i} ^ {g ^ {\prime}}\right)\right), \tag {3} +$$ + +$$ +\mathcal {L} _ {C L} (x, x ^ {\prime}) = - \mathcal {H} (x) \log \left(\mathcal {H} ^ {\prime} \left(x ^ {\prime}\right)\right). \tag {4} +$$ + +Here, $\mathcal{H}$ and $\mathcal{H}'$ denotes the learning head and corresponding EMA part as defined in[6]. The overall training objective is the weighted sum of $\mathcal{L}_{MGCL}$ , $\mathcal{L}_{ITA}$ , and $\mathcal{L}_{QSACL}$ , i.e., + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {M G C L} + \lambda_ {2} \mathcal {L} _ {I T A} + \lambda_ {3} \mathcal {L} _ {Q S A C L}. \tag {5} +$$ + +![](images/2692c01802c74e2d82df136a5360141bd0c5b23806fdc2f686651197675ef1d3.jpg) +Figure 6. Computation pipelines for QSACL with multi-crop augmentation (2 global and $n$ local views). + +# 4. Experiments + +Following SkySense [20], we conducted experiments on 16 datasets across various modalities and tasks to thoroughly evaluate SkySense V2. The applications of pre-trained Sky-Sense V2 to diverse tasks are detailed in Appendix B. Each module is engineered for combined or individual use, providing the flexibility to be frozen or fine-tuned as needed. + +# 4.1. Pre-training Implementation + +Training Datasets. We use the same training dataset of SkySense [20] to train SkySense V2. This extensive dataset comprises approximately 21 million multi-modal RS imagery sets. Each set comprises three components: a high-resolution optical RGB image of spatial dimension $2048 \times 2048$ , a sequence of Sentinel-2 (S2) multi-spectral (MS) images with 10 spectral bands and spatial dimensions of $64 \times 64$ (average sequence length of 65 images), and a sequence of Sentinel-1 (S1) synthetic aperture radar (SAR) images with 2 bands, also having spatial dimensions of $64 \times 64$ (average sequence length of 10 images). In each training iteration, we randomly sample 10 S2 images and 5 S1 images of different time series to accommodate the GPU memory constraints. + +Setting of Unified Backbone. We initially set the token dimension to $C = 352$ . In the first two stages, the win- + +dow size for the Swin Transformer V2 Blocks (SwinV2B) is configured to 8. Each head has a query dimension of 32, and each MLP layer consists of two full-connected layers with dimensions expanding by a factor of 4 consistently across all blocks within the unified backbone. Prompt tokens are incorporated during the last two stages, with the number of prompt tokens being the same for each modality. We allocate 4 prompt tokens for stage-3 and stage-4. + +Pre-training Settings. SkySense V2 is trained using a batch size of 1024, distributed across 128 H20 GPUs. The model undergoes a total of 600k iterations, utilizing the AdamW optimizer [35]. The learning rate is initially set to $2 \times 10^{-4}$ and decays to $1 \times 10^{-6}$ following a cosine annealing schedule [34]. Additional implementation details can be found in the Appendix C. + +# 4.2. Single-Modal Tasks + +We evaluate SkySense V2 on four common single-modal RS interpretation tasks: scene classification, object detection, semantic segmentation, and change detection. The experiments are conducted via fine-tuning a pre-trained Sky-Sense V2 backbone along with a task-specific head. Further implementation details can be found in the Appendix D. + +Scene Classification. In Table 1, we report the performance of scene classification on four widely used datasets: AID [61] and RESISC-45 [11], both featuring static RGB optical images; BEN-S2 [48], which includes static MS images; and fMoW-S2 [13], which contains temporal MS images. For each dataset, the performance is evaluated using different training ratios (TR) following the methodology [13, 38]. In our experiments, we employ a simple linear classifier as the classification head. For multi-label datasets, we use the mean average precision (mAP) as the evaluation metric, whereas for single-label datasets, we report the overall accuracy (OA). As illustrated in Table 1, SkySense V2 consistently outperforms previous RSFMs across various datasets and training configurations, particularly in low TR settings. This demonstrates that SkySense V2 exhibits a stronger representation capability compared to other RSFMs, including its predecessor, SkySense [20]. + +Semantic Segmentation. In Table 2a, we report the segmentation results from four representative semantic segmentation datasets: iSAID [59] and Potsdam [47] for high-resolution optical image segmentation, and Dyna.-Pla [53] and Dyna.-S2 [53] for multi-spectral and multi-temporal image segmentation. Following previous work, we use the mean F1-score (mF1) as the evaluation metric for the Potsdam dataset. While for the other datasets, we report the mean intersection over union (mIoU). All segmentation experiments employ the UperNet [62] as the decoder head. + +
ModelsSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [3]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [38]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [13]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo [50]96.90/98.3494.25/95.67--
RVSA [55]97.03/98.5093.93/95.69--
DINO-MC [58]--84.20/88.7560.16/83.49
TOV [52]95.16/97.0990.97/93.79--
SSL4EO [57]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [40]96.11/97.7994.05/95.53--
CACo [37]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMA [18]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [39]95.47/97.0992.73/94.6486.30/--
Scale-MAE [43]96.44/97.5892.63/95.04--
MA3E [30]-/99.04-/96.23--
SatMAE++ [41]--/97.4885.11/-63.23/-
SkySense [20]97.68/98.6094.85/96.3288.67/92.0964.38/87.27
SkySense V298.34/99.0596.42/97.2489.13/93.7866.65/89.32
+ +Table 1. Scene classification results. - means the task is not supported or the value is unavailable in the paper. + +As illustrated in Table 2a, our SkySense V2 achieve higher performance than recent RSFMs across all four datasets. Specifically, compared to the previous SOTA method SkySense [20], SkySense V2 achieves $1.5\%$ higher performance on average. + +Horizontal & Oriented Object Detection. We utilize the DIOR dataset [27] to evaluate SkySense V2's performance in horizontal object detection, while utilizing the DIOR-R [12] and FAIR1M [51] datasets for oriented object detection. These datasets only consist of RGB RS imagery. Consistent with previous research [20], we adopt Faster R-CNN [44] and Oriented R-CNN [28] as the basic detectors for horizontal and oriented object detection, respectively. As reported in Table 2b, our SkySense V2 outperforms previous RSFMs by a notable margin. Specifically, SkySense V2 surpasses the previous SOTA, SkySense [20], by an average of $1.1\%$ mAP. + +Change Detection. For change detection, we conducted experiments on LEVIR-CD [8], OSCD [15] and Dyna.-S2 [53] datasets. Since LEVIR-CD and OSCD datasets only contain binary change detection results, we use the basic change detector in [9] and report the F1 score as the evaluation metric. While Dyna.-S2 focuses on semantic change detection, we use UperNet [62] as the segmentation network and calculate the semantic change detection score (SCS) on validation and test set following [20]. As presented in Table 2c, SkySense V2 consistently achieves higher performance than the previous SOTA method SkySense on all 3 datasets, and surpasses SkySense by $2.7\%$ on average. Particularly, on the OSCD dataset, our SkySense V2 remark + +
ModelsPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [3]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [38]ICCV'21-57.2089.0329.4/39.8
SatMAE [13]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo [50]TGRS'22-67.2091.27-
RVSA [55]TGRS'2234.3/44.464.49--
BFM [7]Arxiv'23--92.12-
TOV [52]JSTARS'2332.1/37.866.2492.03-
SSL4EO [57]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [40]TGRS'2336.4/43.566.2191.86-
CACo [37]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS [56]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [39]ICCV'2336.7/45.666.6291.85-
Scale-MAE [43]ICCV'2334.0/41.765.7791.54-
MA3E [30]ECCV'24-64.0691.50-
SkySense [20]CVPR'2439.7/46.570.9193.9933.1/46.2
SkySense V2-41.2/47.671.8795.8635.2/47.5
+ +(a) Semantic segmentation results. + +
ModelsHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [3]67.4065.6548.15
SatMAE [13]70.8965.6646.55
RingMo [50]75.90-46.21
RVSA [55]73.2271.0547.04
BFM [7]-73.62-
TOV [52]70.1666.3349.62
SSL4EO [57]64.8261.2349.37
CMID [40]75.1166.3750.58
CACo [37]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [39]72.8467.6749.69
Scale-MAE [43]73.8166.4748.31
MA3E [30]-71.82-
SkySense [20]78.7374.2754.57
SkySense V279.5075.2955.96
+ +(b) Object detection results. + +
ModelsLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [3]78.1946.2613.6/16.7
SeCo [38]90.1447.6713.9/16.0
SatMAE [13]87.6552.7614.8/16.2
RingMo [50]91.86--
RVSA [55]90.86--
SpectralGPT [23]-54.29-
MATTER [1]-59.37-
DINO-MC [58]-52.7014.5/15.6
SSL4EO [57]89.0535.0812.3/17.5
CMID [40]91.72--
CACo [37]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [39]91.7359.82-
Scale-MAE [43]92.07--
SkySense [20]92.5860.0615.4/18.0
SkySense V294.8365.2916.0/20.7
+ +(c) Change detection results. + +Table 2. Results of semantic segmentation, object detection and change detection. + +
Task & DatasetData SourceSkySense[20]SkySense V2
(a) Multi-modal Seg: Dyna.-MM(i) Planet.46.547.6
(ii) S246.247.5
(iii) Planet. + S247.348.7
(iv) Planet. + S2 + S147.748.9
(b) Multi-modal Seg: PASTIS-MM(i) S273.575.0
(ii) S2-Ts84.685.5
(iii) S2-Ts + S1-Ts84.885.6
(iv) S2-Ts + GEP85.886.7
(c) Multi-modal Cls: BEN-MM(i) S186.286.5
(ii) S2 + S192.293.8
+ +Table 3. Fine-tuning results on multi-modal tasks. + +ably outperform SkySense by $5.2\%$ on F1-score. + +# 4.3. Multi-modal Tasks + +In this section, we evaluate SkySense V2 on two representative multi-modal tasks, i.e., Multi-modal Segmentation and Multi-modal Scene Classification, to demonstrate its generalization capability across data from various modalities. + +Multi-modal Segmentation. In Table 3 (a) and (b), we present the performance results of Dyna.-MM [53] and PASTIS-MM [19], respectively. Dyna.-MM contains high-resolution optical imagery from PlanetFusion (Planet.), multi-spectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). For the experiments conducted on Dyna.-MM, we employ a basic UperNet [62] as the decoder head and utilize mIoU as the evaluation metric. As depicted in Table 3 (a), SkySense V2 consistently outperforms the SkySense across various modalities, demonstrating a stronger representation ability. Furthermore, integrating multiple modalities yields higher performance compared to using a single modality. This results confirms that the proposed SkySense V2 can effectively extract representations from diverse modalities using a unified backbone. + +PASTIS-MM is a crop mapping dataset that contains high-resolution optical imagery from Google Earth Pro (GEP), multi-temporal multi-spectral imagery from Sentinel-2 (S2-Ts), and multi-spectral SAR imagery from Sentinel-1 (S1-Ts). We employ a straightforward FCN head to decode the semantic segmentation results and report the overall accuracy as the evaluation metric in Table 3 (b). The results demonstrate that SkySense V2 outperforms the previous SOTA method SkySense by an average of $1.0\%$ . Additionally, it is worth noting that S2-Ts can remarkably improve the segmentation accuracy compared to S2, underscoring the importance of temporal information for crop mapping. + +Multi-modal Scene Classification. We conduct multimodal scene classification experiments on the BEN-MM dataset [49]. As is shown in Table 3 (c), the proposed SkySense V2 surpasses SkySense, demonstrating a stronger representation ability. Furthermore, by integrating S1 and S2 imagery, SkySense V2 achieves higher improvements, highlighting its enhanced ability to extract superior representations from different modalities compared to SkySense. + +# 4.4. Ablation and Discussion + +Ablation Study of Components. To understand the contributions of each component, we conduct a k-nearest neighbors (k-NN) evaluation on the AID and RESISC-45 datasets. Figure 7 illustrates the performance of different component combinations across various pre-training iterations. Notably, the unified backbone design significantly accelerates representation learning. This is primarily due to shared parameters across different modalities, which allow gradients to aggregate, thereby speeding up convergence. Moreover, the unified design allows the backbone to be trained with data from different modalities, thereby enhancing the model's generalization capabilities (see Ap + +![](images/157b50fa09f6a9bb3c383d2cbfb764cf3ce022dc61b0ebacd6e457ac628d5358.jpg) +(a) AID dataset + +![](images/ddab7a34b9c460ff759e00ca77fa0ac2f30b72d847516100b464b78fbb3ac768.jpg) +(b) RESISC-45 dataset + +pendix F.3 for more details). The results also indicate that global self-attention complements window-based self-attention, facilitating improved representations. Additionally, modality-specific prompt tokens and MoE scaling enhance representation learning by increasing feature diversity and model capacity, respectively. QSACL enhances model performance by improving the semantic precision of features in contrastive learning. + +![](images/a8afe768d3b37554c5227266fccc75147b6e8a6aa0fea4fdc8844b8fd3bf8c00.jpg) +(a) w/o modality-specific prompt +Figure 8. t-SNE visualization of the output feature representation from the final block of the unified transformer backbone. HROI, SARI, and MSI represent the features from high-resolution optical imagery, synthetic aperture radar imagery, and multi-spectral imagery, respectively. + +![](images/ab0b69daa3fd05c6b98ef39b824a4d8b9019942209e76506731358ad8be19be5.jpg) +(b) w/ modality-specific prompt + +How Do Modality-Specific Prompt Tokens Facilitate Pre-training? To investigate the impact of modality-specific prompt (MsP) tokens on the multi-modal pretraining process within a unified backbone, we employed t-SNE [54] for visualizing the features corresponding to different modalities. These features are extracted from the last block of the unified transformer backbone, and the comparative results are presented in Figure 8. Without the use of MsP, the features of high-resolution optical imagery (HROI) and synthetic aperture radar imagery (SARI) appear closely clustered, leading to overlapping feature distri + +butions. However, with the inclusion of MsP, the features of HROI and SARI become distinctly separable. These visualization outcomes clearly demonstrate that MsP effectively enhances feature diversity and imparts modality-specific characteristics. + +![](images/721ae3dd381463ee86702aa22004dd666abd304c68e69858a1ed9f9c2f219d30.jpg) +Figure 7. Ablation study using k-NN classification was conducted on the AID and RESISC-45 datasets. In this context, the abbreviations have the following meanings: GA refers to replacing window-based self-attention with global self-attention in the last two stages; MsP involves the addition of modality-specific prompt tokens; MoE indicates scaling up the model through a mixture of experts approach; and QSACL means adding additional query-based semantic aggregation contrastive learning. + +![](images/ec7f5d7fa36c20e60ba37f477b4a6224da6008f0e3b2066617163a2ea2177e87.jpg) + +![](images/4e10518bb815b3b681427b5d6f84a2776cfe4be3d5ca83fb3d896f24a99b5925.jpg) +Figure 9. Visualization of cross-attention weights for features patch corresponding to different queries in QSACL. Queries can effectively aggregate features with specific semantics. + +![](images/46b0d1342509ec15d7fe3af9840b4562ba3b30bf5cdb50ec90b034c69a2b0e3d.jpg) + +What Types of Features Does Query-based Semantic Aggregation Contrastive Learning Capture? To provide a clearer understanding of our proposed QSACL, we visualize the attention weights assigned to different features during their interaction with certain query, as illustrated in Figure 9. During pre-training, we use two global augmentation crops and six local augmentation crops. Here, we present two distinct queries, each focusing on different semantic features of image patches. For example, one query focuses on features with building semantics (left part of the first row in Figure 9), while another focuses on features with land semantics (right part of the first row in Figure 9). The attention weights of different queries aggregate features from each patch into distinct feature representations. Subsequently, we perform contrastive learning between aggregated features from the same query across teacher and student models. This approach significantly enhances semantic accuracy compared to traditional contrastive learning applied to entire image features. + +# 5. Conclusion + +In this paper, we introduce SkySense V2, a MM-RSFM that utilizes a unified backbone to accommodate various modalities. This unified approach enhances parameter utilization efficiency and improves the model's generalization ability. SkySense V2 is pre-trained with an innovative QSACL strategy, specifically designed to leverage the unique characteristics of RS images. As a result, SkySense V2 significantly boosts performance while maintaining the multi-modal flexibility benefits of its predecessor, SkySense. Looking ahead, our future work will focus on integrating the language modality and incorporating a geographical knowledge graph to develop a more powerful and versatile MM-RSFM. + +# References + +[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 7 +[2] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. AnySat: An Earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3 +[3] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 6, 7 +[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 6, 7 +[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5 +[7] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 7 +[8] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6 +[9] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 6 +[10] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1 +[11] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 6 +[12] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 6 +[13] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 6, 7 + +[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2 +[15] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3 +[17] William Fedus, Barret Zoph, and Noam M. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ArXiv, abs/2101.03961, 2021. 2 +[18] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 6 +[19] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7 +[20] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. Skysense: A multimodal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27672-27683, 2024. 1, 3, 4, 5, 6, 7 +[21] Wang Haiyang, Tang Hao, Shi Shaoshuai, Li Aoxue, Li Zhenguo, Schiele Bernt, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. 3 +[22] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27852-27862, 2024. 3 +[23] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Pedram Ghamisi, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, (8):5227-5244, 2024. 7 +[24] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3:79–87, 1991. 2, 4 +[25] Pallavi Jain, Bianca Schoen-Phelan, and Robert J. Ross. Self-supervised learning for invariant representations from + +multi-spectral and sar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 7797-7808, 2022. 2 +[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2 +[27] Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6 +[28] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reppoints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6 +[29] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24088-24097, 2024. 2 +[30] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images. In European Conference on Computer Vision, pages 260-278. Springer, 2024. 2, 6, 7 +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1 +[32] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 3 +[33] Zhili Liu, Kai Chen, Jianhua Han, Lanqing Hong, Hang Xu, Zhenguo Li, and James Tin-Yau Kwok. Task-customized masked autoencoder via mixture of cluster-conditional experts. ArXiv, abs/2402.05382, 2024. 4 +[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning, 2016. 6 +[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. CoRR, abs/1711.05101, 2017. 6 +[36] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1 +[37] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7 +[38] Oscar Manas, Alexandre Lacoste, Xavier Giro-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7 + +[39] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 6, 7 +[40] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 6, 7 +[41] Mubashir Noman, Muzammal Naseer, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, and Fahad Shahbaz Khan. Rethinking transformers pre-training for multispectral satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27811-27819, 2024. 2, 6 +[42] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 4 +[43] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 6, 7 +[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6 +[45] Sheng Shen, Le Hou, Yan-Quan Zhou, Nan Du, S. Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In International Conference on Learning Representations, 2023. 2 +[46] Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Y Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3 +[47] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6 +[48] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Felipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 6 + +[49] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7 +[50] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7 +[51] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fairlm: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6 +[52] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 6, 7 +[53] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7 +[54] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8 +[55] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 6, 7 +[56] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 7 +[57] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 6, 7 +[58] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6, 7 +[59] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6 + +[60] Lemeng Wu, Mengchen Liu, Yinpeng Chen, Dongdong Chen, Xiyang Dai, and Lu Yuan. Residual mixture of experts. ArXiv, abs/2204.09636, 2022. 4 +[61] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 6 +[62] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6, 7 +[63] Zhitong Xiong, Yi Wang, Fahong Zhang, and Xiao Xiang Zhu. One for all: Toward unified foundation models for earth vision. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 2734-2738, 2024. 3 +[64] Longqi Yang, Liangliang Zhang, and Wenjing Yang. Graph adversarial self-supervised learning. In Advances in Neural Information Processing Systems, 2021. 2 +[65] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1 +[66] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 3 +[67] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uniperceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16783-16794, 2021. 3 +[68] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam M. Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. 2022. 2 \ No newline at end of file diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/images.zip b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..59c96fc0e705fa33d2e7c0a697db7c917aadaa20 --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1dc0eb7e607f57a70224953b87b5990dadddd9a1970c1df8820ccdb37b2f81f +size 701322 diff --git a/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/layout.json b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d31786ed0929595a445a02d2f448bf1d527af8e --- /dev/null +++ b/2025/SkySense V2_ A Unified Foundation Model for Multi-modal Remote Sensing/layout.json @@ -0,0 +1,8513 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "SkySense V2: A Unified Foundation Model for Multi-modal Remote Sensing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": "Yingying Zhang" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Lixiang Ru" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Kang Wu" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Lei Yu" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Lei Liang" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Yansheng Li" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Jingdong Chen" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 559, + 186 + ], + "type": "text", + "content": " Wuhan University qichu.zyy@antgroup.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 549 + ], + "type": "text", + "content": "The multi-modal remote sensing foundation model (MM-RSFM) has significantly advanced various Earth observation tasks, such as urban planning, environmental monitoring, and natural disaster management. However, most existing approaches generally require the training of separate backbone networks for each data modality, leading to redundancy and inefficient parameter utilization. Moreover, prevalent pre-training methods typically apply self-supervised learning (SSL) techniques from natural images without adequately accommodating the characteristics of remote sensing (RS) images, such as the complicated semantic distribution within a single RS image. In this work, we present SkySense V2, a unified MM-RSFM that employs a single transformer backbone to handle multiple modalities. This backbone is pre-trained with a novel SSL strategy tailored to the distinct traits of RS data. In particular, SkySense V2 incorporates an innovative adaptive patch merging module and learnable modality prompt tokens to address challenges related to varying resolutions and limited feature diversity across modalities. In additional, we incorporate the mixture of experts (MoE) module to further enhance the performance of the foundation model. SkySense V2 demonstrates impressive generalization abilities through an extensive evaluation involving 16 datasets over 7 tasks, outperforming SkySense by an average of 1.8 points." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Advancements in multi-modal remote sensing foundation models (MM-RSFM) have significantly enhanced the utilization of remote sensing (RS) data across numerous Earth observation (EO) applications[5, 10, 36, 65], such as environmental monitoring, agriculture management, natural disaster response, land cover (or use) mapping etc. Recently, Guo et al. introduced SkySense [20], the largest MM-RSFM to date. SkySense has demonstrated impressive generalization capabilities through extensive evaluations across many different tasks. A key factor in its success is the simultaneous training of geographically aligned multi-modal" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 213, + 539, + 392 + ], + "blocks": [ + { + "bbox": [ + 329, + 213, + 539, + 392 + ], + "lines": [ + { + "bbox": [ + 329, + 213, + 539, + 392 + ], + "spans": [ + { + "bbox": [ + 329, + 213, + 539, + 392 + ], + "type": "image", + "image_path": "8a82f94c635106848da56b5529fc72525d5d6555cc4f85d6d7f6647edec4eda9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 401, + 555, + 435 + ], + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 435 + ], + "type": "text", + "content": "Figure 1. SkySense V2 has achieved superior performance on 16 datasets over 7 distinct tasks compared with its predecessor Sky-Sense and supports a board range of tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 614 + ], + "type": "text", + "content": "and multi-temporal RS data. To accommodate the varying resolutions of different modalities, SkySense employs the Swin Transformer [31] for processing high-resolution (HR) optical images, while using the Vision Transformer (ViT) [16] for medium-resolution multi-spectral (MS) data and synthetic aperture radar (SAR) data. However, this separate backbone design results in redundancy and inefficient use of parameters. On the other hand, SkySense's pre-training approach is primarily based on DINOv2 [42], renowned for its efficiency in self-supervised learning (SSL). Although it has demonstrated strong performance with RS images, there remains considerable room for improvement by adequately accommodating the involved semantic distribution within a single RS image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we introduce SkySense V2, an advanced MM-RSFM pre-trained through a novel SSL approach tailored to the distinctive traits of RS data. To be specific, SkySense V2 leverages a unified transformer backbone capable of processing data across diverse modalities and resolutions. It is pre-trained using a new SSL strategy that utilizes query-based attention to gather the similar semantic features distributed in different regions of RS images." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9136" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "Designing a unified backbone for geographically aligned multi-modal data presents two primary challenges. The first is processing feature resolutions across data with varying ground sample distance (GSD) during simultaneous pretraining of multiple modalities. To address this challenge, we introduce an innovative Adaptive Patch Merging (APM) module, which is integrated after each stage of the unified backbone. This module determines whether to reduce feature resolution according to the specific feature resolution requirements of each modality. For instance, when dealing with HR optical images, the APM module activates resolution reduction at each stage. In contrast, for medium-resolution data, such as MS and SAR data, the feature resolution is preserved throughout the stages. This design enables the backbone to efficiently process data of different resolutions while adapting output feature resolutions according to actual requirements, which is crucial for the joint training and fusion of multi-modal features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 542 + ], + "type": "text", + "content": "The second challenge is that fully sharing parameters across the different modalities can reduce the feature diversity. To address this, inspired by the work of [14] and [26], we introduce learnable modality prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pretrained model can capture the unique characteristics of each modality better. Compared to SkySense, our unified design significantly enhances the efficiency of parameter utilization. To be specific, SkySense uses three separate backbones, Swin-H for HR optical data, ViT-L for MS data, and ViT-L for SAR data, collectively totaling 1.26 billion parameters. In contrast, SkySense V2's unified transformer backbone is capable of simultaneously processing all three modalities while utilizing only 665 million parameters. This efficient use of parameters not only optimizes the architecture but also allows for potential model scaling to further boost performance. Therefore, we integrate the mixture of experts module [24], a technique commonly used in various large language models [17, 45, 68] to further enhance the performance of the foundation model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": "To effectively adapt SSL techniques developed for natural images to the pre-training of RSFM, it is essential to consider the distinct differences in data characteristics. Natural images typically feature a single, clear subject, such as a cat, dog, or person. In contrast, RS images comprise multiple subjects distributed across various regions of the image. For instance, as illustrated in Figure 2, one section of an RS image may show a building, while other areas might show a forest, pond, or land. Current SSL methods perform contrastive learning directly between different views using various augmented crops, which is effective for natural images due to their singular focus. However, when applied to RS images, this approach can result in semantic inaccuracies, as different views may capture different subjects. For" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 321, + 69, + 432, + 178 + ], + "blocks": [ + { + "bbox": [ + 321, + 69, + 432, + 178 + ], + "lines": [ + { + "bbox": [ + 321, + 69, + 432, + 178 + ], + "spans": [ + { + "bbox": [ + 321, + 69, + 432, + 178 + ], + "type": "image", + "image_path": "01538eb678699ee9904ed6057dca4de016e8849455a461b9bd592ff9d08c68b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 181, + 419, + 191 + ], + "lines": [ + { + "bbox": [ + 332, + 181, + 419, + 191 + ], + "spans": [ + { + "bbox": [ + 332, + 181, + 419, + 191 + ], + "type": "text", + "content": "(a) Views of natural image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 245 + ], + "type": "text", + "content": "Figure 2. Comparison of different augmented views from natural and RS images in SSL. Natural image views often concentrate on a singular subject, such as a \"dog\". RS image views encompass a variety of subjects, including building, forest, pond, and land." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 437, + 71, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 437, + 71, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 437, + 71, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 437, + 71, + 547, + 178 + ], + "type": "image", + "image_path": "7e60dacfbae18e7330c45c0796f26ae6463f9e18b86279806a71a496a7bf06c5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 180, + 531, + 191 + ], + "lines": [ + { + "bbox": [ + 456, + 180, + 531, + 191 + ], + "spans": [ + { + "bbox": [ + 456, + 180, + 531, + 191 + ], + "type": "text", + "content": "(b) Views of RS image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 280 + ], + "type": "text", + "content": "example, in Figure 2b, view 1 of the RS image primarily contains building, whereas view 2 shows forest and pond." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 412 + ], + "type": "text", + "content": "To tackle this challenge, we propose a novel approach called Query-based Semantic Aggregation Contrastive Learning (QSACL). This method employs multiple learnable queries to perform cross-attention with features from different views, generating semantically aggregated features. We then apply contrastive learning to these aggregated features pairs derived from the same query. As demonstrated in Figure 9 from our ablation study, utilizing different queries can aggregate consistent semantic features across multiple views, which enhances the accuracy of contrastive learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 521 + ], + "type": "text", + "content": "We evaluated SkySense V2 on a diverse set of 16 datasets [8, 11-13, 15, 19, 27, 48, 51, 53, 59, 61], covering a range of task types, modalities, and spatial scales. As illustrated in Figure 1, SkySense V2 demonstrates substantial improvements in performance over its predecessor, SkySense, achieving state-of-the-art (SOTA) results across various modalities of EO tasks. The experimental results across all test scenarios highlight its competitive advantage over existing RSFMs in a wide array of EO interpretation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 532, + 401, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 532, + 401, + 544 + ], + "spans": [ + { + "bbox": [ + 314, + 532, + 401, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 501, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 501, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 501, + 565 + ], + "type": "text", + "content": "2.1. Remote Sensing Foundation Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "RSFMs are characterized by their ability to leverage vast amounts of data through self-supervised learning (SSL) technologies, enabling them to learn robust feature representations without the need for extensive annotations. The predominant RSFMs utilize Contrastive Learning (CL) or Masked Image Modeling (MIM). Typical works in CL include RS-BYOL [25], GASSL [64], DINO-MC [58], SeCo [38], and CACo [37]. Other research efforts focus on enhancing MIM framework, such as RingMo [50], S2MAE [29], MA3E [30], SatMAE [13], and SatMAE++ [41]. Additionally, studies like CMID [40] and GFM [39] explore the intersection of CL and MIM through a self-distillation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9137" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": "approach. Recently, Guo et al. introduced a comprehensive MM-RSFM known as SkySense [20], which features a factorized multi-modal spatiotemporal encoder. This architecture facilitates independent spatial feature extraction and multi-modal temporal fusion. As the largest MM-RSFM to date, SkySense exhibits exceptional generalization capabilities across a wide range of RS datasets. However, it does face a challenge: the inefficient utilization of parameters due to the separate backbone designs for different modalities. In this work, we propose SkySense V2 to address this issue by employing a unified transformer backbone design." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 295, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 225 + ], + "type": "text", + "content": "2.2. Unified Framework for Multi-modal Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 507 + ], + "type": "text", + "content": "Multi-modal learning aims to train models that effectively process and relate information from various modalities. Substantial advancements have been made in this field, culminating in numerous studies focused on creating unified networks for the integration and processing of diverse modalities. The main work includes VLMO [46], MetaTransformer [66], Uni-Perceiver [67], and UniTR [21], etc. There are several methods that utilize a unified backbone design for multi-modal RS data. For example, OFA-Net [63] pre-trains a single transformer backbone on a curated multi-modal dataset using MIM. Han et al. proposed a RSFM called msGFM[22], which utilizes a shared transformer encoder while employing different decoders for each modality. In contrast to our approach, both OFA-Net and msGFM simply share all of parameters of backbone and learn feature representations for multi-modal data separately, which results in a lack of geographical alignment and fusion between features from different modalities. AnySat [2] employs a joint embedding predictive architecture for a multi-modal model, training a single model on heterogeneous data in a SSL manner. However, the backbone design and SSL pre-training methods of our SkySense V2 differ significantly from AnySat." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 137, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 137, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 137, + 530 + ], + "type": "text", + "content": "3. SkySense V2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 170, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 170, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 170, + 548 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 217, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 217, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 217, + 566 + ], + "type": "text", + "content": "3.1.1. Unified Transformer Backbone" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "As illustrated in Figure 3, our unified transformer backbone is a hierarchical encoder structure with four stages. In the first two stages, we employ Swin Transformer V2 Blocks (SwinV2B) [32] to incorporate essential visual priors, such as locality and translation invariance. The window-based self-attention mechanisms in SwinV2B also help reduce computational complexity compared to global self-attention, particularly given the high spatial resolution of features in these initial stages. In the last two stages, we utilize vanilla Transformer Blocks (TB) [16] for two main reasons: First, the spatial resolution of features in these stages is relatively low, making the computational" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 75, + 559, + 168 + ], + "blocks": [ + { + "bbox": [ + 318, + 75, + 559, + 168 + ], + "lines": [ + { + "bbox": [ + 318, + 75, + 559, + 168 + ], + "spans": [ + { + "bbox": [ + 318, + 75, + 559, + 168 + ], + "type": "image", + "image_path": "5da551ae3edf10f05bd3a5f8d8ade3e925a500f1bab788fb4533d966a80b63ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 355, + 169, + 533, + 258 + ], + "blocks": [ + { + "bbox": [ + 355, + 169, + 533, + 258 + ], + "lines": [ + { + "bbox": [ + 355, + 169, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 355, + 169, + 533, + 258 + ], + "type": "image", + "image_path": "81f83f84530a2ece1a0a9856e9aeae73bb9c474b601085994f6c8743748e02e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 273, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 307 + ], + "type": "text", + "content": "Figure 3. Overview of the unified transformer backbone model in SkySense V2: The entire backbone shares all parameters across different modalities, with the exception of distinct tokenizers." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 366 + ], + "type": "text", + "content": "costs of global self-attention more manageable. Second, our ablation study demonstrates that global self-attention can be complementary to window-based self-attention, enabling the model to achieve stronger representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": "Given a group of multi-modal inputs consisting of a high-resolution optical image " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "x_{HR}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": ", multi-spectral data " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "x_{MS}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": ", and synthetic aperture radar data " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "x_{SAR}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " where each 'pixel' in the different modalities is naturally aligned by geo-location, we employ three distinct tokenizers to process these inputs into tokens. In each tokenizer, the input data is first divided into non-overlapping " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " patches. Subsequently, a linear embedding layer is applied to this raw, patched data to project it into patch tokens. Subsequently, four stages of either SwinV2B or TB, with shared parameters across modalities, are applied to these patch tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Due to the varying GSD of satellite sensors across different modalities, input data corresponding to the same region at specific geo-locations during joint aligned training exhibit differing spatial resolutions. To harmonize these spatial resolutions for tokens from various modalities, we propose Adaptive Patch Merging (APM) and utilize it to selectively reduce spatial resolution at each stage, except for Stage 1. Specifically, for the tokens of optical image data with high spatial resolution, APM concatenates the features from groups of " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " neighboring patches and applies a linear layer to the " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "4c" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "-dimensional concatenated features. This process reduces the number of tokens by a factor of " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "2 \\times 2 = 4" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " downsampling of spatial resolution, while the output dimension is set to " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "2c" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": ". In contrast, for the tokens of MS and SAR data with lower spatial resolution, APM maintains the resolution by applying a linear projection with averaging the weights across the input dimensions. By integrating the APM module, our unified backbone can efficiently process" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9138" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 301, + 205 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 301, + 205 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 301, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 301, + 205 + ], + "type": "image", + "image_path": "6fad7f3c67a3374040e087878f6577842cdc793c1fc760e1ddea740e1950af2c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 240 + ], + "type": "text", + "content": "Figure 4. The illustration of modality-specific prompt tokens added to the unified backbone." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 277 + ], + "type": "text", + "content": "multi-modal data with varying resolutions while maintaining the spatial alignment of features across modalities" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 224, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 224, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 224, + 295 + ], + "type": "text", + "content": "3.1.2. Modality-specific Prompt Tokens" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": "Our unified transformer backbone utilizes fully shared parameters across different modalities. To enhance feature diversity, we introduce learnable modality-specific prompt tokens for each modality. By interacting with these individual modal prompt tokens through an attention mechanism, the pre-trained model can better capture the unique characteristics of each modality. As illustrated in Figure 4, we incorporate " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": " learnable prompt tokens for each modality in the last two stages. We denote the input tokens of each stage as " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "E_{i}^{j}\\in \\mathbb{R}^{h_{j}w_{j}\\times c_{j}}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "i\\in \\{HR,MS,SAR\\}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "j\\in \\{3,4\\}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": " are the height and width of the spatial resolution at stage " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": " refers to the dimension of the tokens. For each modality " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ", we insert the modality-specific learnable prompt tokens " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "P_{i}^{j}\\in \\mathbb{R}^{N\\times c_{j}}" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": " into the input of the first block in " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": "-th stage " + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_j" + }, + { + "bbox": [ + 55, + 297, + 296, + 501 + ], + "type": "text", + "content": ". At the last block of each stage, these prompt tokens are discarded. The entire process can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 507, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 507, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 111, + 507, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ E _ {d r o p}, E _ {i} ^ {4} \\right] = \\mathcal {F} _ {3} \\left(\\left[ P _ {i} ^ {3}, E _ {i} ^ {3} \\right]\\right), \\tag {1} \\\\ [ E _ {d r o p}, E _ {i} ^ {o u t} ] = \\mathcal {F} _ {4} ([ P _ {i} ^ {4}, E _ {i} ^ {4} ]). \\\\ \\end{array}", + "image_path": "e1cafc0a3306eb51af24b8f78bffac228066613a4f692f3a6dc3e462427aa7c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "E_{drop}" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": " denotes the discarded tokens, and " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "E_{i}^{out}" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": " represents the final output tokens of backbone. Our design of modality-specific prompts enhances feature diversity while maintaining full parameter sharing by incorporating only a small number of modality-specific parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 614, + 201, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 201, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 201, + 627 + ], + "type": "text", + "content": "3.1.3. Scaling Up Model Capacity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Our unified backbone design across various modalities significantly enhances parameter utilization efficiency. To further boost the model's capabilities, we integrate a Mixture of Experts (MoE) [24] into the backbone network. We choose MoE instead of simply increasing the backbone's width and depth because MoE enables models to be pretrained with substantially less computation by leveraging" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "sparse feed-forward layers (i.e., experts) for individual tokens. Following common approaches [33, 60], we insert MoE modules into the last " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " transformer blocks, replacing the original feed-forward network (FFN) layers. Each MoE module consists of " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " experts (denoted as " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i(\\cdot), i = 1,2,\\dots,M" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "), which share the same structure as the FFN and operate as independent networks. For the gating network in MoE, we employ a learnable linear layer followed by a Softmax function, " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(x) = \\text{Softmax}(Wx)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ", where W is the gating parameter. Finally, the output of the MoE is computed as a linear combination of the outputs from the selected experts, weighted by the corresponding gate values. This process is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 233, + 555, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 233, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 365, + 233, + 555, + 259 + ], + "type": "interline_equation", + "content": "M O E (x) = \\sum_ {i \\in \\mathcal {T}} \\mathcal {G} _ {i} (x) \\cdot \\mathcal {E} _ {i} (x), \\tag {2}", + "image_path": "f8fe03ba677ce1866d8ab2890ac2b87793d733fbc03ae884664ff82492fba7f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "text", + "content": " represent the set of the top- " + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "text", + "content": " indices. In our SkySense V2 backbone, we set " + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "inline_equation", + "content": "L = 6,M = 8" + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 264, + 555, + 288 + ], + "type": "inline_equation", + "content": "k = 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 294, + 474, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 474, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 474, + 308 + ], + "type": "text", + "content": "3.2. Overall Pre-training Pipeline" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "text", + "content": "Figure 5 provides an overview of our pre-training procedure. The pre-training framework of SkySense V2 primarily adopts the teacher-student architecture from SkySense [20], where the teacher network's parameters are updated using an exponential moving average (EMA) [42] of the student network's parameters. To train SkySense V2, we employ the Multi-Granularity Contrastive Learning (MGCL) loss " + }, + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MGCL}" + }, + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "text", + "content": " as proposed in SkySense. For each modality, MGCL utilizes contrastive loss [42] to align the representations of teacher and student networks at pixel, object and image-level granularity, respectively. Additionally, we incorporate the unsupervised Geo-Context Prototype Learning (GCPL), as validated in SkySense, to enhance the learning of complementary regional context clues, aiding downstream tasks. To further enhance dense interpretation capabilities, we introduce an auxiliary supervision loss " + }, + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITA}" + }, + { + "bbox": [ + 313, + 312, + 555, + 563 + ], + "type": "text", + "content": " to employ dense Image-Text Alignment (ITA) according to OpenStreetMap (OSM) labels1. Detailed implementations of MGCL, GCPL, and ITA are included in the Appendix A. In following paragraph, we will focus on proposed query-based semantic aggregation contrastive learning (QSACL)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": "Query-based Semantic Aggregation Contrastive Learning. QSACL utilizes learnable queries to aggregate similar semantics across different regions of images, enabling more accurate contrastive learning. Given features " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "g_{1}, g_{2}" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " from two global views and " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "l_{1}, l_{2}, \\ldots, l_{n}" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " local views (where " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " denotes the fused features " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{fus}}" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " from global and local view), we employ a transformer decoder layer to perform cross-attention between " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " learnable queries " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "q_{1}, q_{2}, \\ldots, q_{m}" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " and the feature sets " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 577, + 556, + 696 + ], + "type": "text", + "content": ". This process generates sets of globally and locally aggregated features set" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 435, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 435, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 435, + 713 + ], + "type": "text", + "content": "1https://www.openstreetmap.org/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9139" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 89, + 527, + 295 + ], + "blocks": [ + { + "bbox": [ + 239, + 78, + 362, + 91 + ], + "lines": [ + { + "bbox": [ + 239, + 78, + 362, + 91 + ], + "spans": [ + { + "bbox": [ + 239, + 78, + 362, + 91 + ], + "type": "text", + "content": "SkySense V2 Pre-training" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 89, + 527, + 295 + ], + "lines": [ + { + "bbox": [ + 70, + 89, + 527, + 295 + ], + "spans": [ + { + "bbox": [ + 70, + 89, + 527, + 295 + ], + "type": "image", + "image_path": "70a45e8d4136055737c6aa275dd9546d43ce537f250f09d2f596ef1fea283769.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 318, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 54, + 318, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 555, + 364 + ], + "type": "text", + "content": "Figure 5. Overview of the SkySense V2 pre-training pipeline. The SkySense V2 employs data augmentations on multi-modal inputs, which are then fed into both the student and teacher networks. The pre-training process incorporates Multi-Granularity Contrastive Learning, Dense Image-Text Alignment, and Query-based Semantic Aggregation Contrastive Learning to effectively train the network. Once pretraining is complete, the parameters from the teacher branch are utilized for downstream applications." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "z_{i}^{g}" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "z_{i}^{l}" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "i = 1,2,\\ldots m" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": ". As is shown in Figure 6, we use superscript " + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\prime" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": " denotes features from the teacher network, the QSACL loss " + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{QSACL}" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": " is computed by averaging the loss " + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CL}" + }, + { + "bbox": [ + 55, + 383, + 296, + 444 + ], + "type": "text", + "content": " of aggregation features between local and global views for each query:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 453, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 453, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 64, + 453, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {Q S A C L} = \\frac {1}{2 m} \\sum_ {i = 1} ^ {m} \\left(\\mathcal {L} _ {C L} \\left(z _ {i} ^ {g}, z _ {i} ^ {l ^ {\\prime}}\\right) + \\mathcal {L} _ {C L} \\left(z _ {i} ^ {l}, z _ {i} ^ {g ^ {\\prime}}\\right)\\right), \\tag {3}", + "image_path": "824a385026f860a815077b11965a34a2ffbf9467a59551642064e15d83530949.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 497, + 295, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 497, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 103, + 497, + 295, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C L} (x, x ^ {\\prime}) = - \\mathcal {H} (x) \\log \\left(\\mathcal {H} ^ {\\prime} \\left(x ^ {\\prime}\\right)\\right). \\tag {4}", + "image_path": "8275beef2a49e5c0b95e6a24c1837153fe7b66fd9735bfb14376760828da27ea.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{H}'" + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": " denotes the learning head and corresponding EMA part as defined in[6]. The overall training objective is the weighted sum of " + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{MGCL}" + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ITA}" + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{QSACL}" + }, + { + "bbox": [ + 55, + 516, + 296, + 564 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 574, + 295, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 574, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 88, + 574, + 295, + 588 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {M G C L} + \\lambda_ {2} \\mathcal {L} _ {I T A} + \\lambda_ {3} \\mathcal {L} _ {Q S A C L}. \\tag {5}", + "image_path": "43d0217289621d7cb70c98ed568f472005797ebea7a1c027a88874ccd0aa8118.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 592, + 294, + 679 + ], + "blocks": [ + { + "bbox": [ + 58, + 592, + 294, + 679 + ], + "lines": [ + { + "bbox": [ + 58, + 592, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 58, + 592, + 294, + 679 + ], + "type": "image", + "image_path": "2692c01802c74e2d82df136a5360141bd0c5b23806fdc2f686651197675ef1d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "type": "text", + "content": "Figure 6. Computation pipelines for QSACL with multi-crop augmentation (2 global and " + }, + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 688, + 295, + 711 + ], + "type": "text", + "content": " local views)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 382, + 395, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 395, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 395, + 396 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 475 + ], + "type": "text", + "content": "Following SkySense [20], we conducted experiments on 16 datasets across various modalities and tasks to thoroughly evaluate SkySense V2. The applications of pre-trained Sky-Sense V2 to diverse tasks are detailed in Appendix B. Each module is engineered for combined or individual use, providing the flexibility to be frozen or fine-tuned as needed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 471, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 471, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 471, + 498 + ], + "type": "text", + "content": "4.1. Pre-training Implementation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "text", + "content": "Training Datasets. We use the same training dataset of SkySense [20] to train SkySense V2. This extensive dataset comprises approximately 21 million multi-modal RS imagery sets. Each set comprises three components: a high-resolution optical RGB image of spatial dimension " + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "inline_equation", + "content": "2048 \\times 2048" + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "text", + "content": ", a sequence of Sentinel-2 (S2) multi-spectral (MS) images with 10 spectral bands and spatial dimensions of " + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "text", + "content": " (average sequence length of 65 images), and a sequence of Sentinel-1 (S1) synthetic aperture radar (SAR) images with 2 bands, also having spatial dimensions of " + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 503, + 556, + 671 + ], + "type": "text", + "content": " (average sequence length of 10 images). In each training iteration, we randomly sample 10 S2 images and 5 S1 images of different time series to accommodate the GPU memory constraints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Setting of Unified Backbone. We initially set the token dimension to " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "C = 352" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". In the first two stages, the win-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9140" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "dow size for the Swin Transformer V2 Blocks (SwinV2B) is configured to 8. Each head has a query dimension of 32, and each MLP layer consists of two full-connected layers with dimensions expanding by a factor of 4 consistently across all blocks within the unified backbone. Prompt tokens are incorporated during the last two stages, with the number of prompt tokens being the same for each modality. We allocate 4 prompt tokens for stage-3 and stage-4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "text", + "content": "Pre-training Settings. SkySense V2 is trained using a batch size of 1024, distributed across 128 H20 GPUs. The model undergoes a total of 600k iterations, utilizing the AdamW optimizer [35]. The learning rate is initially set to " + }, + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "text", + "content": " and decays to " + }, + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 55, + 182, + 294, + 266 + ], + "type": "text", + "content": " following a cosine annealing schedule [34]. Additional implementation details can be found in the Appendix C." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 273, + 170, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 170, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 170, + 285 + ], + "type": "text", + "content": "4.2. Single-Modal Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 361 + ], + "type": "text", + "content": "We evaluate SkySense V2 on four common single-modal RS interpretation tasks: scene classification, object detection, semantic segmentation, and change detection. The experiments are conducted via fine-tuning a pre-trained Sky-Sense V2 backbone along with a task-specific head. Further implementation details can be found in the Appendix D." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 579 + ], + "type": "text", + "content": "Scene Classification. In Table 1, we report the performance of scene classification on four widely used datasets: AID [61] and RESISC-45 [11], both featuring static RGB optical images; BEN-S2 [48], which includes static MS images; and fMoW-S2 [13], which contains temporal MS images. For each dataset, the performance is evaluated using different training ratios (TR) following the methodology [13, 38]. In our experiments, we employ a simple linear classifier as the classification head. For multi-label datasets, we use the mean average precision (mAP) as the evaluation metric, whereas for single-label datasets, we report the overall accuracy (OA). As illustrated in Table 1, SkySense V2 consistently outperforms previous RSFMs across various datasets and training configurations, particularly in low TR settings. This demonstrates that SkySense V2 exhibits a stronger representation capability compared to other RSFMs, including its predecessor, SkySense [20]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Semantic Segmentation. In Table 2a, we report the segmentation results from four representative semantic segmentation datasets: iSAID [59] and Potsdam [47] for high-resolution optical image segmentation, and Dyna.-Pla [53] and Dyna.-S2 [53] for multi-spectral and multi-temporal image segmentation. Following previous work, we use the mean F1-score (mF1) as the evaluation metric for the Potsdam dataset. While for the other datasets, we report the mean intersection over union (mIoU). All segmentation experiments employ the UperNet [62] as the decoder head." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 547, + 273 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 273 + ], + "type": "table", + "html": "
ModelsSingle-labelMulti-labelTemporal
AID (TR=20%/50%)RESISC-45 (TR=10%/20%)BEN-S2 (TR=10%/100%)fMoW-S2 (TR=100%)
OAOAmAPTop-1/5 Acc
GASSL [3]93.55/95.9290.86/93.0679.24/87.4050.69/77.99
SeCo [38]93.47/95.9989.64/92.9182.62/87.8151.65/77.40
SatMAE [13]95.02/96.9491.72/94.1086.18/89.5063.84/-
RingMo [50]96.90/98.3494.25/95.67--
RVSA [55]97.03/98.5093.93/95.69--
DINO-MC [58]--84.20/88.7560.16/83.49
TOV [52]95.16/97.0990.97/93.79--
SSL4EO [57]91.06/94.7487.60/91.2787.10/91.8051.70/76.77
CMID [40]96.11/97.7994.05/95.53--
CACo [37]90.88/95.0588.28/91.9481.30/87.0050.72/76.31
CROMA [18]--88.29/-63.59/-
SatLas [4]94.96/97.3892.16/94.7082.80/88.3757.95/79.00
GFM [39]95.47/97.0992.73/94.6486.30/--
Scale-MAE [43]96.44/97.5892.63/95.04--
MA3E [30]-/99.04-/96.23--
SatMAE++ [41]--/97.4885.11/-63.23/-
SkySense [20]97.68/98.6094.85/96.3288.67/92.0964.38/87.27
SkySense V298.34/99.0596.42/97.2489.13/93.7866.65/89.32
", + "image_path": "17eb365830aac4a3e0f2bb189c38788dc0cdb092d95e31fa05fb7dc32411a3aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 304 + ], + "type": "text", + "content": "Table 1. Scene classification results. - means the task is not supported or the value is unavailable in the paper." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 318, + 553, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 377 + ], + "type": "text", + "content": "As illustrated in Table 2a, our SkySense V2 achieve higher performance than recent RSFMs across all four datasets. Specifically, compared to the previous SOTA method SkySense [20], SkySense V2 achieves " + }, + { + "bbox": [ + 313, + 318, + 553, + 377 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 318, + 553, + 377 + ], + "type": "text", + "content": " higher performance on average." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 538 + ], + "type": "text", + "content": "Horizontal & Oriented Object Detection. We utilize the DIOR dataset [27] to evaluate SkySense V2's performance in horizontal object detection, while utilizing the DIOR-R [12] and FAIR1M [51] datasets for oriented object detection. These datasets only consist of RGB RS imagery. Consistent with previous research [20], we adopt Faster R-CNN [44] and Oriented R-CNN [28] as the basic detectors for horizontal and oriented object detection, respectively. As reported in Table 2b, our SkySense V2 outperforms previous RSFMs by a notable margin. Specifically, SkySense V2 surpasses the previous SOTA, SkySense [20], by an average of " + }, + { + "bbox": [ + 313, + 396, + 554, + 538 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 313, + 396, + 554, + 538 + ], + "type": "text", + "content": " mAP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Change Detection. For change detection, we conducted experiments on LEVIR-CD [8], OSCD [15] and Dyna.-S2 [53] datasets. Since LEVIR-CD and OSCD datasets only contain binary change detection results, we use the basic change detector in [9] and report the F1 score as the evaluation metric. While Dyna.-S2 focuses on semantic change detection, we use UperNet [62] as the segmentation network and calculate the semantic change detection score (SCS) on validation and test set following [20]. As presented in Table 2c, SkySense V2 consistently achieves higher performance than the previous SOTA method SkySense on all 3 datasets, and surpasses SkySense by " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " on average. Particularly, on the OSCD dataset, our SkySense V2 remark" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9141" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 260, + 243 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 260, + 243 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 260, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 260, + 243 + ], + "type": "table", + "html": "
ModelsPublicationDyna.-Pla.iSAIDPotsdamDyna.-S2
mIoUmIoUmF1mIoU
GASSL [3]ICCV'2134.0/40.865.9591.2728.1/41.0
SeCo [38]ICCV'21-57.2089.0329.4/39.8
SatMAE [13]NIPS'2232.8/39.962.9790.6330.1/38.7
RingMo [50]TGRS'22-67.2091.27-
RVSA [55]TGRS'2234.3/44.464.49--
BFM [7]Arxiv'23--92.12-
TOV [52]JSTARS'2332.1/37.866.2492.03-
SSL4EO [57]GRSM'2335.3/42.164.0191.5431.8/42.7
CMID [40]TGRS'2336.4/43.566.2191.86-
CACo [37]CVPR'2335.4/42.764.3291.3530.2/42.5
SAMRS [56]NIPS'23-66.2691.43-
SatLas [4]ICCV'2337.4/40.768.7191.2831.9/43.5
GFM [39]ICCV'2336.7/45.666.6291.85-
Scale-MAE [43]ICCV'2334.0/41.765.7791.54-
MA3E [30]ECCV'24-64.0691.50-
SkySense [20]CVPR'2439.7/46.570.9193.9933.1/46.2
SkySense V2-41.2/47.671.8795.8635.2/47.5
", + "image_path": "1880090de436d3b750c653c2a577482c7b986d28f2fde08b5361e3f956322c8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 100, + 246, + 212, + 255 + ], + "lines": [ + { + "bbox": [ + 100, + 246, + 212, + 255 + ], + "spans": [ + { + "bbox": [ + 100, + 246, + 212, + 255 + ], + "type": "text", + "content": "(a) Semantic segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 263, + 72, + 405, + 243 + ], + "blocks": [ + { + "bbox": [ + 263, + 72, + 405, + 243 + ], + "lines": [ + { + "bbox": [ + 263, + 72, + 405, + 243 + ], + "spans": [ + { + "bbox": [ + 263, + 72, + 405, + 243 + ], + "type": "table", + "html": "
ModelsHorizontalOriented
DIORDIOR-RFAIR1M
mAP50mAPmAP
GASSL [3]67.4065.6548.15
SatMAE [13]70.8965.6646.55
RingMo [50]75.90-46.21
RVSA [55]73.2271.0547.04
BFM [7]-73.62-
TOV [52]70.1666.3349.62
SSL4EO [57]64.8261.2349.37
CMID [40]75.1166.3750.58
CACo [37]66.9164.1047.83
SatLas [4]74.1067.5946.19
GFM [39]72.8467.6749.69
Scale-MAE [43]73.8166.4748.31
MA3E [30]-71.82-
SkySense [20]78.7374.2754.57
SkySense V279.5075.2955.96
", + "image_path": "ef03af63540b704b25950ecd390cf28eb3cf952f8dea7a80fe691affc1c06897.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 287, + 244, + 378, + 254 + ], + "lines": [ + { + "bbox": [ + 287, + 244, + 378, + 254 + ], + "spans": [ + { + "bbox": [ + 287, + 244, + 378, + 254 + ], + "type": "text", + "content": "(b) Object detection results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 408, + 73, + 554, + 240 + ], + "blocks": [ + { + "bbox": [ + 408, + 73, + 554, + 240 + ], + "lines": [ + { + "bbox": [ + 408, + 73, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 408, + 73, + 554, + 240 + ], + "type": "table", + "html": "
ModelsLEVIR-CDOSCDDyna.-S2
F1F1SCS
GASSL [3]78.1946.2613.6/16.7
SeCo [38]90.1447.6713.9/16.0
SatMAE [13]87.6552.7614.8/16.2
RingMo [50]91.86--
RVSA [55]90.86--
SpectralGPT [23]-54.29-
MATTER [1]-59.37-
DINO-MC [58]-52.7014.5/15.6
SSL4EO [57]89.0535.0812.3/17.5
CMID [40]91.72--
CACo [37]81.0452.1115.3/15.8
SatLas [4]90.62-13.3/17.8
GFM [39]91.7359.82-
Scale-MAE [43]92.07--
SkySense [20]92.5860.0615.4/18.0
SkySense V294.8365.2916.0/20.7
", + "image_path": "67ebc81055753547a3938edff7634ad9966ba026c5573527472c6e921596264d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 433, + 242, + 525, + 251 + ], + "lines": [ + { + "bbox": [ + 433, + 242, + 525, + 251 + ], + "spans": [ + { + "bbox": [ + 433, + 242, + 525, + 251 + ], + "type": "text", + "content": "(c) Change detection results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 62, + 284, + 288, + 395 + ], + "blocks": [ + { + "bbox": [ + 156, + 265, + 453, + 277 + ], + "lines": [ + { + "bbox": [ + 156, + 265, + 453, + 277 + ], + "spans": [ + { + "bbox": [ + 156, + 265, + 453, + 277 + ], + "type": "text", + "content": "Table 2. Results of semantic segmentation, object detection and change detection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 284, + 288, + 395 + ], + "lines": [ + { + "bbox": [ + 62, + 284, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 288, + 395 + ], + "type": "table", + "html": "
Task & DatasetData SourceSkySense[20]SkySense V2
(a) Multi-modal Seg: Dyna.-MM(i) Planet.46.547.6
(ii) S246.247.5
(iii) Planet. + S247.348.7
(iv) Planet. + S2 + S147.748.9
(b) Multi-modal Seg: PASTIS-MM(i) S273.575.0
(ii) S2-Ts84.685.5
(iii) S2-Ts + S1-Ts84.885.6
(iv) S2-Ts + GEP85.886.7
(c) Multi-modal Cls: BEN-MM(i) S186.286.5
(ii) S2 + S192.293.8
", + "image_path": "6f2fd6a57db011827a029766f6867a719b02c55d3557d07176a5baa3f29a80b1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 84, + 403, + 265, + 414 + ], + "lines": [ + { + "bbox": [ + 84, + 403, + 265, + 414 + ], + "spans": [ + { + "bbox": [ + 84, + 403, + 265, + 414 + ], + "type": "text", + "content": "Table 3. Fine-tuning results on multi-modal tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 251, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 251, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 251, + 441 + ], + "type": "text", + "content": "ably outperform SkySense by " + }, + { + "bbox": [ + 55, + 430, + 251, + 441 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 55, + 430, + 251, + 441 + ], + "type": "text", + "content": " on F1-score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 450, + 167, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 167, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 167, + 462 + ], + "type": "text", + "content": "4.3. Multi-modal Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 468, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 517 + ], + "type": "text", + "content": "In this section, we evaluate SkySense V2 on two representative multi-modal tasks, i.e., Multi-modal Segmentation and Multi-modal Scene Classification, to demonstrate its generalization capability across data from various modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Multi-modal Segmentation. In Table 3 (a) and (b), we present the performance results of Dyna.-MM [53] and PASTIS-MM [19], respectively. Dyna.-MM contains high-resolution optical imagery from PlanetFusion (Planet.), multi-spectral imagery from Sentinel-2 (S2), and SAR imagery from Sentinel-1 (S1). For the experiments conducted on Dyna.-MM, we employ a basic UperNet [62] as the decoder head and utilize mIoU as the evaluation metric. As depicted in Table 3 (a), SkySense V2 consistently outperforms the SkySense across various modalities, demonstrating a stronger representation ability. Furthermore, integrating multiple modalities yields higher performance compared to using a single modality. This results confirms that the proposed SkySense V2 can effectively extract representations from diverse modalities using a unified backbone." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 441 + ], + "type": "text", + "content": "PASTIS-MM is a crop mapping dataset that contains high-resolution optical imagery from Google Earth Pro (GEP), multi-temporal multi-spectral imagery from Sentinel-2 (S2-Ts), and multi-spectral SAR imagery from Sentinel-1 (S1-Ts). We employ a straightforward FCN head to decode the semantic segmentation results and report the overall accuracy as the evaluation metric in Table 3 (b). The results demonstrate that SkySense V2 outperforms the previous SOTA method SkySense by an average of " + }, + { + "bbox": [ + 313, + 285, + 555, + 441 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 313, + 285, + 555, + 441 + ], + "type": "text", + "content": ". Additionally, it is worth noting that S2-Ts can remarkably improve the segmentation accuracy compared to S2, underscoring the importance of temporal information for crop mapping." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 544 + ], + "type": "text", + "content": "Multi-modal Scene Classification. We conduct multimodal scene classification experiments on the BEN-MM dataset [49]. As is shown in Table 3 (c), the proposed SkySense V2 surpasses SkySense, demonstrating a stronger representation ability. Furthermore, by integrating S1 and S2 imagery, SkySense V2 achieves higher improvements, highlighting its enhanced ability to extract superior representations from different modalities compared to SkySense." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 450, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 450, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 450, + 564 + ], + "type": "text", + "content": "4.4. Ablation and Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Ablation Study of Components. To understand the contributions of each component, we conduct a k-nearest neighbors (k-NN) evaluation on the AID and RESISC-45 datasets. Figure 7 illustrates the performance of different component combinations across various pre-training iterations. Notably, the unified backbone design significantly accelerates representation learning. This is primarily due to shared parameters across different modalities, which allow gradients to aggregate, thereby speeding up convergence. Moreover, the unified design allows the backbone to be trained with data from different modalities, thereby enhancing the model's generalization capabilities (see Ap" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9142" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 175, + 153 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 175, + 153 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 175, + 153 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 175, + 153 + ], + "type": "image", + "image_path": "157b50fa09f6a9bb3c383d2cbfb764cf3ce022dc61b0ebacd6e457ac628d5358.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 156, + 142, + 166 + ], + "lines": [ + { + "bbox": [ + 91, + 156, + 142, + 166 + ], + "spans": [ + { + "bbox": [ + 91, + 156, + 142, + 166 + ], + "type": "text", + "content": "(a) AID dataset" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 71, + 293, + 153 + ], + "blocks": [ + { + "bbox": [ + 177, + 71, + 293, + 153 + ], + "lines": [ + { + "bbox": [ + 177, + 71, + 293, + 153 + ], + "spans": [ + { + "bbox": [ + 177, + 71, + 293, + 153 + ], + "type": "image", + "image_path": "ddab7a34b9c460ff759e00ca77fa0ac2f30b72d847516100b464b78fbb3ac768.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 156, + 272, + 165 + ], + "lines": [ + { + "bbox": [ + 198, + 156, + 272, + 165 + ], + "spans": [ + { + "bbox": [ + 198, + 156, + 272, + 165 + ], + "type": "text", + "content": "(b) RESISC-45 dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 380 + ], + "type": "text", + "content": "pendix F.3 for more details). The results also indicate that global self-attention complements window-based self-attention, facilitating improved representations. Additionally, modality-specific prompt tokens and MoE scaling enhance representation learning by increasing feature diversity and model capacity, respectively. QSACL enhances model performance by improving the semantic precision of features in contrastive learning." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 389, + 177, + 478 + ], + "blocks": [ + { + "bbox": [ + 58, + 389, + 177, + 478 + ], + "lines": [ + { + "bbox": [ + 58, + 389, + 177, + 478 + ], + "spans": [ + { + "bbox": [ + 58, + 389, + 177, + 478 + ], + "type": "image", + "image_path": "a8afe768d3b37554c5227266fccc75147b6e8a6aa0fea4fdc8844b8fd3bf8c00.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 480, + 170, + 491 + ], + "lines": [ + { + "bbox": [ + 63, + 480, + 170, + 491 + ], + "spans": [ + { + "bbox": [ + 63, + 480, + 170, + 491 + ], + "type": "text", + "content": "(a) w/o modality-specific prompt" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 499, + 296, + 555 + ], + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 555 + ], + "type": "text", + "content": "Figure 8. t-SNE visualization of the output feature representation from the final block of the unified transformer backbone. HROI, SARI, and MSI represent the features from high-resolution optical imagery, synthetic aperture radar imagery, and multi-spectral imagery, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 389, + 294, + 478 + ], + "blocks": [ + { + "bbox": [ + 179, + 389, + 294, + 478 + ], + "lines": [ + { + "bbox": [ + 179, + 389, + 294, + 478 + ], + "spans": [ + { + "bbox": [ + 179, + 389, + 294, + 478 + ], + "type": "image", + "image_path": "ab0b69daa3fd05c6b98ef39b824a4d8b9019942209e76506731358ad8be19be5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 480, + 287, + 491 + ], + "lines": [ + { + "bbox": [ + 183, + 480, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 183, + 480, + 287, + 491 + ], + "type": "text", + "content": "(b) w/ modality-specific prompt" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "How Do Modality-Specific Prompt Tokens Facilitate Pre-training? To investigate the impact of modality-specific prompt (MsP) tokens on the multi-modal pretraining process within a unified backbone, we employed t-SNE [54] for visualizing the features corresponding to different modalities. These features are extracted from the last block of the unified transformer backbone, and the comparative results are presented in Figure 8. Without the use of MsP, the features of high-resolution optical imagery (HROI) and synthetic aperture radar imagery (SARI) appear closely clustered, leading to overlapping feature distri" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "butions. However, with the inclusion of MsP, the features of HROI and SARI become distinctly separable. These visualization outcomes clearly demonstrate that MsP effectively enhances feature diversity and imparts modality-specific characteristics." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 145, + 434, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 264 + ], + "type": "text", + "content": "Figure 7. Ablation study using k-NN classification was conducted on the AID and RESISC-45 datasets. In this context, the abbreviations have the following meanings: GA refers to replacing window-based self-attention with global self-attention in the last two stages; MsP involves the addition of modality-specific prompt tokens; MoE indicates scaling up the model through a mixture of experts approach; and QSACL means adding additional query-based semantic aggregation contrastive learning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 145, + 434, + 185 + ], + "lines": [ + { + "bbox": [ + 316, + 145, + 434, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 145, + 434, + 185 + ], + "type": "image", + "image_path": "721ae3dd381463ee86702aa22004dd666abd304c68e69858a1ed9f9c2f219d30.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 146, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 435, + 146, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 435, + 146, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 435, + 146, + 553, + 185 + ], + "type": "image", + "image_path": "ec7f5d7fa36c20e60ba37f477b4a6224da6008f0e3b2066617163a2ea2177e87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 188, + 434, + 228 + ], + "blocks": [ + { + "bbox": [ + 317, + 188, + 434, + 228 + ], + "lines": [ + { + "bbox": [ + 317, + 188, + 434, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 434, + 228 + ], + "type": "image", + "image_path": "4e10518bb815b3b681427b5d6f84a2776cfe4be3d5ca83fb3d896f24a99b5925.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 237, + 554, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 554, + 270 + ], + "type": "text", + "content": "Figure 9. Visualization of cross-attention weights for features patch corresponding to different queries in QSACL. Queries can effectively aggregate features with specific semantics." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 188, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 435, + 188, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 435, + 188, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 435, + 188, + 553, + 228 + ], + "type": "image", + "image_path": "46b0d1342509ec15d7fe3af9840b4562ba3b30bf5cdb50ec90b034c69a2b0e3d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 294, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 522 + ], + "type": "text", + "content": "What Types of Features Does Query-based Semantic Aggregation Contrastive Learning Capture? To provide a clearer understanding of our proposed QSACL, we visualize the attention weights assigned to different features during their interaction with certain query, as illustrated in Figure 9. During pre-training, we use two global augmentation crops and six local augmentation crops. Here, we present two distinct queries, each focusing on different semantic features of image patches. For example, one query focuses on features with building semantics (left part of the first row in Figure 9), while another focuses on features with land semantics (right part of the first row in Figure 9). The attention weights of different queries aggregate features from each patch into distinct feature representations. Subsequently, we perform contrastive learning between aggregated features from the same query across teacher and student models. This approach significantly enhances semantic accuracy compared to traditional contrastive learning applied to entire image features." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we introduce SkySense V2, a MM-RSFM that utilizes a unified backbone to accommodate various modalities. This unified approach enhances parameter utilization efficiency and improves the model's generalization ability. SkySense V2 is pre-trained with an innovative QSACL strategy, specifically designed to leverage the unique characteristics of RS images. As a result, SkySense V2 significantly boosts performance while maintaining the multi-modal flexibility benefits of its predecessor, SkySense. Looking ahead, our future work will focus on integrating the language modality and incorporating a geographical knowledge graph to develop a more powerful and versatile MM-RSFM." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9143" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Peri Akiva, Matthew Purri, and Matthew Leotta. Self-supervised material and texture representation learning for remote sensing tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8203-8215, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. AnySat: An Earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "text", + "content": "[3] Kumar Ayush, Burak Uzkent, Chenlin Meng, Kumar Tanmay, Marshall Burke, David Lobell, and Stefano Ermon. Geography-aware self-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10190, 2021. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "text", + "content": "[4] Favyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. Satlaspretrain: A large-scale dataset for remote sensing image understanding. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16772-16782, 2023. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "type": "text", + "content": "[5] Yinxia Cao, Xin Huang, and Qihao Weng. A multiscale weakly supervised learning method with adaptive online noise correction for high-resolution change detection of built-up areas. Remote Sensing of Environment, 297:113779, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 417 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "text", + "content": "[7] Keumgang Cha, Junghoon Seo, and Taekyung Lee. A billion-scale foundation model for remote sensing images. arXiv preprint arXiv:2304.05215, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 453, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 486 + ], + "type": "text", + "content": "[8] Hao Chen and Zhenwei Shi. A spatial-temporal attention-based method and a new dataset for remote sensing image change detection. Remote Sensing, 12(10):1662, 2020. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "text", + "content": "[9] Hao Chen, Zipeng Qi, and Zhenwei Shi. Remote sensing image change detection with transformers. IEEE Transactions on Geoscience and Remote Sensing, 60:1-14, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 555 + ], + "type": "text", + "content": "[10] Gong Cheng and Junwei Han. A survey on object detection in optical remote sensing images. ISPRS journal of photogrammetry and remote sensing, 117:11-28, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 600 + ], + "type": "text", + "content": "[11] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[12] Gong Cheng, Jiabao Wang, Ke Li, Xingxing Xie, Chunbo Lang, Yanqing Yao, and Junwei Han. Anchor-free oriented proposal generator for object detection. IEEE Transactions on Geoscience and Remote Sensing, 60:1-11, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "text", + "content": "[13] Yezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke, David Lobell, and Stefano Ermon. Satmae: Pre-training transformers for temporal and multi-spectral satellite imagery. Advances in Neural Information Processing Systems, 35:197-211, 2022. 2, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 183 + ], + "type": "text", + "content": "[15] Rodrigo Caye Daudt, Bertr Le Saux, Alexandre Boulch, and Yann Gousseau. Urban change detection for multispectral earth observation using convolutional neural networks. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium, pages 2115-2118. IEEE, 2018. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 250 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "type": "text", + "content": "[17] William Fedus, Barret Zoph, and Noam M. Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. ArXiv, abs/2101.03961, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 331 + ], + "type": "text", + "content": "[18] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders. Advances in Neural Information Processing Systems, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 333, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 386 + ], + "type": "text", + "content": "[19] Vivien Sainte Fare Garnot, Loic Landrieu, and Nesrine Chehata. Multi-modal temporal attention models for crop mapping from satellite time series. ISPRS Journal of Photogrammetry and Remote Sensing, 187:294-305, 2022. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 486 + ], + "type": "text", + "content": "[20] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun Zhang, and Yansheng Li. Skysense: A multimodal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27672-27683, 2024. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "text", + "content": "[21] Wang Haiyang, Tang Hao, Shi Shaoshuai, Li Aoxue, Li Zhenguo, Schiele Bernt, and Liwei Wang. Unitr: A unified and efficient multi-modal transformer for bird's-eye-view representation. In ICCV, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "text", + "content": "[22] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27852-27862, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 655 + ], + "type": "text", + "content": "[23] Danfeng Hong, Bing Zhang, Xuyang Li, Yuxuan Li, Chenyu Li, Jing Yao, Pedram Ghamisi, Naoto Yokoya, Hao Li, Xiuping Jia, Antonio Plaza, Paolo Gamba, Jon Atli Benediktsson, and Jocelyn Chanussot. Spectralgpt: Spectral remote sensing foundation model. IEEE Transactions on Pattern Analysis and Machine Intelligence, (8):5227-5244, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[24] Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive mixtures of local experts. *Neural Computation*, 3:79–87, 1991. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Pallavi Jain, Bianca Schoen-Phelan, and Robert J. Ross. Self-supervised learning for invariant representations from" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9144" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "multi-spectral and sar images. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 15: 7797-7808, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[27] Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "text", + "content": "[28] Wentong Li, Yijie Chen, Kaixuan Hu, and Jianke Zhu. Oriented reppoints for aerial object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1829-1838, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[29] Xuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2mae: A spatial-spectral pretraining foundation model for spectral remote sensing data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24088-24097, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 338 + ], + "type": "text", + "content": "[30] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images. In European Conference on Computer Vision, pages 260-278. Springer, 2024. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 459 + ], + "type": "text", + "content": "[32] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "text", + "content": "[33] Zhili Liu, Kai Chen, Jianhua Han, Lanqing Hong, Hang Xu, Zhenguo Li, and James Tin-Yau Kwok. Task-customized masked autoencoder via mixture of cluster-conditional experts. ArXiv, abs/2402.05382, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 525 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv: Learning, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 548 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. CoRR, abs/1711.05101, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[36] ZhiYong Lv, HaiTao Huang, Xinghua Li, MingHua Zhao, Jon Atli Benediktsson, WeiWei Sun, and Nicola Falco. Land cover change detection with heterogeneous remote sensing images: Review, progress, and perspective. Proceedings of the IEEE, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 658 + ], + "type": "text", + "content": "[37] Utkarsh Mall, Bharath Hariharan, and Kavita Bala. Change-aware sampling and contrastive learning for satellite images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5261-5270, 2023. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Oscar Manas, Alexandre Lacoste, Xavier Giro-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[39] Matías Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Towards geospatial foundation models via continual pretraining. Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16806-16816, 2023. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[40] Dilxat Muhtar, Xueliang Zhang, Pengfeng Xiao, Zhenshi Li, and Feng Gu. Cmid: A unified self-supervised learning framework for remote sensing image understanding. IEEE Transactions on Geoscience and Remote Sensing, 2023. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 252 + ], + "type": "text", + "content": "[41] Mubashir Noman, Muzammal Naseer, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, and Fahad Shahbaz Khan. Rethinking transformers pre-training for multispectral satellite imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27811-27819, 2024. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 308 + ], + "type": "text", + "content": "[42] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 387 + ], + "type": "text", + "content": "[43] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[44] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 521 + ], + "type": "text", + "content": "[45] Sheng Shen, Le Hou, Yan-Quan Zhou, Nan Du, S. Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 611 + ], + "type": "text", + "content": "[46] Sheng Shen, Le Hou, Yanqi Zhou, Nan Du, Shayne Long-pre, Jason Wei, Hyung Won Chung, Barret Zoph, William Fedus, Xinyun Chen, Tu Vu, Yuexin Wu, Wuyang Chen, Albert Webson, Yunxuan Li, Vincent Y Zhao, Hongkun Yu, Kurt Keutzer, Trevor Darrell, and Denny Zhou. Mixture-of-experts meets instruction tuning: A winning combination for large language models. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 611, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 554, + 646 + ], + "type": "text", + "content": "[47] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[48] Gencer Sumbul, Jian Kang, Tristan Kreuziger, Felipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, and Begüm Demir. Bigearthnet: A large-scale benchmark archive for remote sensing image understanding. In IEEE International Geoscience and Remote Sensing Symposium, pages 5901-5904, 2019. 2, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9145" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 711 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 148 + ], + "type": "text", + "content": "[49] Gencer Sumbul, Arne de Wall, Tristan Kreuziger, Filipe Marcelino, Hugo Costa, Pedro Benevides, Mario Caetano, Begüm Demir, and Volkerl Mark. BigEarthNet-MM: A large-scale, multimodal, multilabel benchmark archive for remote sensing image classification and retrieval. IEEE Geoscience and Remote Sensing Magazine, 9(3):174-180, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[50] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xue Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 2022. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 270 + ], + "type": "text", + "content": "[51] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fairlm: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[52] Chao Tao, Ji Qi, Guo Zhang, Qing Zhu, Weipeng Lu, and Haifeng Li. Tov: The original vision model for optical remote sensing image understanding via self-supervised learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 403 + ], + "type": "text", + "content": "[53] Aysim Token, Lukas Kondmann, Mark Weber, Marvin Eisenberger, Andrés Camero, Jingliang Hu, Ariadna Pregel Hoderlein, Caglar Şenaras, Timothy Davis, Daniel Cremers, et al. Dynamicearthnet: Daily multi-spectral satellite dataset for semantic change segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21158-21167, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 404, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 435 + ], + "type": "text", + "content": "[54] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[55] Di Wang, Qiming Zhang, Yufei Xu, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Advancing plain vision transformer toward remote sensing foundation model. IEEE Transactions on Geoscience and Remote Sensing, 61:1-15, 2022. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 535 + ], + "type": "text", + "content": "[56] Di Wang, Jing Zhang, Bo Du, Dacheng Tao, and Liangpei Zhang. Scaling-up remote sensing segmentation dataset with segment anything model. Advances in Neural Information Processing Systems, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[57] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eo-s12: A large-scale multi-modal, multi-temporal dataset for self-supervised learning in earth observation. IEEE Geoscience and Remote Sensing Magazine, 11(3):98-106, 2023. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "text", + "content": "[58] Xinye Wanyan, Sachith Seneviratne, Shuchang Shen, and Michael Kirley. Dino-mc: Self-supervised contrastive learning for remote sensing imagery with multi-sized local crops. arXiv preprint arXiv:2303.06670, 2023. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 637, + 294, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 711 + ], + "type": "text", + "content": "[59] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 2, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 485 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[60] Lemeng Wu, Mengchen Liu, Yinpeng Chen, Dongdong Chen, Xiyang Dai, and Lu Yuan. Residual mixture of experts. ArXiv, abs/2204.09636, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[61] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[62] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[63] Zhitong Xiong, Yi Wang, Fahong Zhang, and Xiao Xiang Zhu. One for all: Toward unified foundation models for earth vision. IGARSS 2024 - 2024 IEEE International Geoscience and Remote Sensing Symposium, pages 2734-2738, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[64] Longqi Yang, Liangliang Zhang, and Wenjing Yang. Graph adversarial self-supervised learning. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[65] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[66] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "text", + "content": "[67] Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Xiaogang Wang, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uniperceiver: Pre-training unified architecture for generic perception for zero-shot and few-shot tasks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16783-16794, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 485 + ], + "type": "text", + "content": "[68] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam M. Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. 2022. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9146" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_content_list.json b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b36edb4426c52de9bd425c003e2e842b0bab70e8 --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_content_list.json @@ -0,0 +1,1530 @@ +[ + { + "type": "text", + "text": "Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 109, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheyun Qin $^{1}$ , Deng Yu $^{2}$ , Chuanchen Luo $^{2}$ , Zhumin Chen $^{2*}$", + "bbox": [ + 238, + 186, + 761, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Shandong University $^{2}$ School of Artificial Intelligence, Shandong University", + "bbox": [ + 125, + 210, + 870, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 260, + 326, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, researchers have explored the task of open-vocabulary video instance segmentation, which aims to identify, track, and segment any instance within an open set of categories. The core challenge of Open-Vocabulary VIS lies in solving the cross-domain alignment problem, including spatial-temporal and text-visual domain alignments. Existing methods have made progress but still face shortcomings in addressing these alignments, especially due to data heterogeneity. Inspired by metric learning, we propose an innovative Sliced Wasserstein Bridging Learning Framework. This framework utilizes the Sliced Wasserstein distance as the core tool for metric learning, effectively bridging the four domains involved in the task. Our innovations are threefold: (1) Domain Alignment: By mapping features from different domains into a unified metric space, our method maintains temporal consistency and learns intrinsic consistent features between modalities, improving the fusion of text and visual information. (2) Weighting Mechanism: We introduce an importance weighting mechanism to enhance the discriminative ability of our method when dealing with imbalanced or significantly different data. (3) High Efficiency: Our method inherits the computational efficiency of the Sliced Wasserstein distance, allowing for online processing of large-scale video data while maintaining segmentation accuracy. Through extensive experimental evaluations, we have validated the robustness of our concept and the effectiveness of our framework.", + "bbox": [ + 91, + 292, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, video instance segmentation methods [23, 27, 28] have demonstrated significant improvements and breakthroughs in various application fields. However, most existing methods rely on the Closed-Set Assumption [4, 22], where the learning and recognition scope of the model are strictly limited to known categories or labels during training and testing. This assumption ignores the possibility of unknown categories in the real world, limiting the model's ability to effectively detect unknown categories.", + "bbox": [ + 89, + 741, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this limitation, researchers have begun to explore the task of open-vocabulary video instance segmentation [15, 26] (Open-Vocabulary VIS), which aims to identify, track, and segment any instance within an open set of categories. These Open-Vocabulary VIS methods can not only handle traditional predefined categories but also flexibly adapt to new user-specified categories, achieving the advanced function of on-demand segmentation.", + "bbox": [ + 511, + 261, + 903, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The core of Open-Vocabulary VIS methods lies in solving the cross-domain alignment problem, including spatial-temporal domain alignment and text-visual domain alignment. Existing methods have different focuses. In terms of spatial-temporal domain alignment, current methods extract class-agnostic instance features based on the foundational model Mask2Former-VIS. Furthermore, OV2Seg [26], OpenVIS [15], and CLIP-VIS [32] employ memory-based approaches to perform cross-frame instance matching. In terms of text-visual domain alignment, existing methods mainly rely on the zero-shot performance exhibited by visual-language models, i.e., CLIP [24] pretrained on largescale image-text pairs. Specifically, BriVIS [8] and OV-Former [11] introduce additional CLIP image encoders on top of visual feature extractors. The difference lies in that BriVIS integrates them based on contrastive learning, while OV-Former maps text and visual features into the same space vis the attention mechanism.", + "bbox": [ + 511, + 383, + 903, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the significant progress made by existing Open-Vocabulary VIS methods, there are still many shortcomings. The cross-domain alignment problem remains inadequately addressed, and the inherent data heterogeneity results in significant differences in feature distributions across different domains. In light of this, we cannot help but trace back to the development history of traditional machine learning methods, seeking a tool that can simultaneously solve spatial-temporal domain alignment and text-visual domain alignment. We find that the metric learning method in the field of machine learning provides an effective approach to addressing the alignment problem [12, 17]. By learning a suitable metric space, this paradigm makes similar samples closer while dissimilar samples relatively farther apart, providing a powerful tool for solving the alignment problem.", + "bbox": [ + 511, + 657, + 903, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guided by the idea of metric learning, we propose an", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Zhumin Chen.", + "bbox": [ + 109, + 886, + 315, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12470", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "innovative Sliced Wasserstein Bridging Learning Framework, i.e., SWbridge. This framework utilizes the Sliced Wasserstein distance [3] as the core tool for metric learning. This framework cleverly utilizes the Sliced Wasserstein distance as the core tool for metric learning. Specifically, we minimize the Sliced Wasserstein distance between samples based on random paths through sampling, achieving feature embedding alignment. We bridge four domains based on a strategy to address the challenges of cross-domain alignment (spatial-temporal domain and text-visual domain).", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our innovations are mainly concentrated in three aspects: (1) Domain Alignment. The Sliced Wasserstein distance has the advantage of accurately capturing differences in data distributions in high-dimensional spaces. Our method uses the Sliced Wasserstein distance as a bridge to map features from different domains into a unified metric space. This not only maintains temporal consistency by comparing instance embeddings between adjacent frames, achieving temporal alignment and effectively mitigating challenges such as progressive occlusion, but also learns intrinsic consistent features between modalities, narrowing the semantic gap and improving the fusion of text and visual information through modality alignment. (2) Weighting Mechanism: We introduce an importance weighting mechanism to further enhance its discriminative ability when dealing with imbalanced data or data with significant feature differences. In open-vocabulary video instance segmentation tasks, there may be significant feature differences between different instances, and some instances may be more critical or informative. Through the importance weighting mechanism, our method can dynamically adjust the contribution of different instances in the metric space, allowing key instances to have greater weight in distance calculations. (3) High Efficiency: Our SWbridge inherits the computational efficiency advantage of the Sliced Wasserstein distance, making it highly efficient when processing large-scale video data. Without introducing additional network parameters, our method can handle video in an online manner while maintaining segmentation accuracy.", + "bbox": [ + 91, + 244, + 483, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Combining the aforementioned perspectives and innovations, we propose an open-vocabulary video instance segmentation method capable of bridging multiple domains. This method has achieved remarkable results on multiple datasets related to open-vocabulary and video tasks.", + "bbox": [ + 89, + 683, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 775, + 241, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Open-Vocabulary Detection and Segmentation", + "text_level": 1, + "bbox": [ + 89, + 801, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open Vocabulary Detection (OVD) [9, 25] and Open Vocabulary Segmentation (OVS) [13, 18] are cutting-edge tasks in the field of computer vision, enabling models to be trained on images containing unannotated novel objects, thereby breaking the closed-set constraint. This break-", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "through is primarily attributed to the application of weak supervision signals, namely the utilization of image-text pairs (such as image-caption pairs and image-level labels) or large pre-trained Visual-Language Models (VLMs), such as CLIP [24]. Supported by weak supervision signals, methods for OVD and OVS can be categorized into four main types [31]. The first type is region-aware training, which learns object feature representations by exploring the intrinsic links between image regions and text descriptions without relying on the VLMs' image encoder (VLMsIE) or direct object annotations. The second type is the pseudo-labeling method, which enhances model generalization by generating pseudo-labels from preliminary predictions on unannotated images using both image-text pairs and the VLMs-IE. The other two types are knowledge distillation and transfer learning. Knowledge distillation transfers knowledge from pre-trained VLMs to new models or tasks via a distillation mechanism, while transfer learning applies knowledge learned by pre-trained models on specific tasks to new tasks or domains. Both rely on the VLMsIE and seldom involve direct training on image-text pairs.", + "bbox": [ + 511, + 90, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, zero-shot learning [29] addresses the closed-set constraint in scene-aware tasks and differs from open vocabulary methods. It prohibits access to weak supervision signals during training, but can be converted into Open-Vocabulary tasks upon gaining access, enhancing the model's generalization and recognition of novel objects.", + "bbox": [ + 511, + 409, + 906, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Open-Vocabulary Video Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 513, + 903, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary video instance segmentation, as an emerging vision task, aims to simultaneously classify, track, and segment arbitrary objects within open categories in videos, attracting considerable attention from researchers in recent years. OV2Seg [26] has taken the lead in this field by not only constructing the Large Vocabulary Video Instance Segmentation (LV-VIS) dataset but also proposing the first end-to-end benchmark for open-vocabulary video instance segmentation. OV2Seg method leverages the Mask2Former [6] framework to extract class-agnostic masks and query embeddings, utilizes the CLIP text encoder for precise mask classification, and achieves cross-frame instance tracking through a long-term matching strategy. Subsequent research has mostly followed this basic approach and expanded upon it. OpenVIS [15] introduces a two-stage framework, OVIS, which adopts a neighboring matching strategy for instance tracking, effectively simplifying the processing pipeline. BriVIS [8] models instances as Brownian bridges and closely aligns bridge-level instance representations with category texts through contrastive learning, also employing a neighboring matching strategy to enhance the accuracy of instance recognition. OVFormer [11] introduces a novel unified embedding alignment module that effectively addresses the domain gap between instance", + "bbox": [ + 511, + 537, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "queries and text embeddings, demonstrating good practicality with its semi-online processing approach. However, the aforementioned methods primarily focus on bridging the domain discrepancy between text and vision, with insufficient attention to spatiotemporal consistency among cross-frame instances. CLIP-VIS [32] improves upon this by introducing a time-topK enhanced matching strategy, which strengthens temporal modeling capabilities between frames and improves matching accuracy.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We believe that the core of the open-vocabulary video instance segmentation task lies in simultaneously addressing the alignment of the spatial-temporal domain and the text-visual domain. To this end, based on the concept of metric learning, we employ the Sliced Wasserstein distance to bridge various domains. This approach not only effectively narrows the domain gap between modalities but also ensures spatio-temporal consistency of instances across frames, providing a novel solution for the open-vocabulary video instance segmentation task.", + "bbox": [ + 89, + 227, + 482, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 383, + 316, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of Visual Language Models (VLMs) [5] is their deep training on large-scale image-text paired data, enabling the fusion and understanding of visual and textual information, and endowing them with powerful zero-shot object recognition capabilities [19]. CLIP [24], a prominent VLM, uses an image encoder to extract visual features and a text encoder to generate text embeddings, jointly constructing a cross-modal representation space. However, applying VLMs to open-vocabulary video instance segmentation is challenging. VLMs, primarily trained on images, struggle with understanding dynamic video scenes rich in spatiotemporal information, and lack the capability to maintain object consistency across frames. Additionally, the domain shift between text and vision poses a significant challenge, potentially leading to biases in mapping textual information to the visual space. To overcome these limitations, we propose a progressive approach that addresses the spatiotemporal consistency issue and the domain shift problem, enabling VLMs to better adapt to video instance segmentation tasks.", + "bbox": [ + 89, + 406, + 483, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 705, + 227, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Open-Vocabulary VIS Formulation", + "text_level": 1, + "bbox": [ + 89, + 729, + 397, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a test video $\\mathcal{D}_{test}$ with $T$ frames, the objective in Open-Vocabulary VIS is to accurately predict all $N$ instances belonging to the categories in $\\mathcal{C} = \\mathcal{C}_{base} \\cup \\mathcal{C}_{novel}$ by the trained model $f_{\\theta}$ . $\\mathcal{C}_{base}$ is the set of base (training) categories, and $\\mathcal{C}_{novel}$ is the set of novel categories that are not seen during training but may appear in the test videos. The prediction can be formulated as:", + "bbox": [ + 89, + 751, + 482, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\left\\{m _ {1}, m _ {2}, \\dots , m _ {T} \\right\\}, c \\right\\} _ {n} ^ {N} = f _ {\\theta} \\left(\\mathcal {D} _ {t e s t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 862, + 480, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{m_t\\}_{t = 1}^T$ is the segmentation masks, and $c\\in \\mathcal{C}$", + "bbox": [ + 89, + 883, + 449, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sliced Wasserstein Distance", + "text_level": 1, + "bbox": [ + 513, + 90, + 763, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One-dimensional Wasserstein Distance. For one-dimensional probability measures $\\mu$ and $\\nu$ in $\\mathcal{P}_p(\\mathbb{R})$ , the $p$ -Wasserstein distance is defined as:", + "bbox": [ + 511, + 112, + 903, + 157 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {W} _ {p} ^ {p} (\\mu , \\nu) = \\int_ {0} ^ {1} | F _ {\\mu} ^ {- 1} (z) - F _ {\\nu} ^ {- 1} (z) | ^ {p} d z, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 165, + 903, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{\\mu}$ and $F_{\\nu}$ are the cumulative distribution functions (CDFs) of $\\mu$ and $\\nu$ . This formulation provides a closed form for computing the Wasserstein distance in one-dimensional spaces, making it well-suited for projected measures.", + "bbox": [ + 511, + 208, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sliced Wasserstein Distance. To generalize the Wasserstein distance to higher-dimensional measures, the Sliced Wasserstein Distance (SWD) projects the measures $\\mu$ and $\\nu$ in $\\mathcal{P}_p(\\mathbb{R}^d)$ onto one-dimensional subspaces, and then averages the one-dimensional Wasserstein distances from these projections. For $\\mu, \\nu \\in \\mathcal{P}_p(\\mathbb{R}^d)$ , the SWD is defined as:", + "bbox": [ + 511, + 268, + 905, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S W} _ {p} ^ {p} (\\mu , \\nu) = \\mathbb {E} _ {\\theta \\sim \\mathcal {U} \\left(\\mathbb {S} ^ {d - 1}\\right)} \\left[ \\mathrm {W} _ {p} ^ {p} \\left(\\theta \\sharp \\mu , \\theta \\sharp \\nu\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 369, + 903, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta \\sharp \\mu$ and $\\theta \\sharp \\nu$ represent the push-forward measures of $\\mu$ and $\\nu$ along direction $\\theta \\in \\mathbb{S}^{d - 1}$ , the unit sphere in $\\mathbb{R}^d$ . This projection $f(x) = \\theta^{\\top}x$ maps points from $\\mathbb{R}^d$ to $\\mathbb{R}$ , enabling the computation of Wasserstein distances in a one-dimensional space.", + "bbox": [ + 511, + 396, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the expectation in Eq. 3 is computationally expensive, SWD is typically approximated by averaging over $L$ independent directions, $\\theta_{1},\\ldots ,\\theta_{L}$ , sampled from $\\mathcal{U}(\\mathbb{S}^{d - 1})$ :", + "bbox": [ + 511, + 472, + 903, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbf {S W}} _ {p} ^ {p} (\\mu , \\nu ; L) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {W} _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 527, + 903, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where each $\\theta_{l}\\sharp \\mu$ and $\\theta_{l}\\sharp \\nu$ are projected representations of $\\mu$ and $\\nu$ along the direction $\\theta_{l}$ . The number of projections $L$ controls the accuracy of the Monte Carlo approximation.", + "bbox": [ + 511, + 575, + 905, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 633, + 604, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, we propose the SWbridge framework. Firstly, in §4.1, we integrate a category-agnostic feature extractor. Then, in §4.2, we introduce a spatio-temporal bridging module that robustly establishes instance associations between adjacent frames, ensuring the consistency and coherence of temporal information. Finally, in §4.3, we design a modal bridging module that performs fine-grained semantic mapping across the text and visual domains to enhance cross-domain understanding and effectively address the challenges posed by domain shifts.", + "bbox": [ + 511, + 657, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Class-agnostic Feature Extractor", + "text_level": 1, + "bbox": [ + 511, + 818, + 802, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video $\\mathcal{D}$ consisting of $T$ frames, $\\{V_t \\in \\mathbb{R}^{3 \\times H \\times W}\\}_{t=1}^T$ , where $H$ and $W$ denote the height and width of each frame, our model follows the conventional paradigm [11, 26, 32] by adopting the Mask2Former", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12472", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a47e114bbf7d6746aa848776aae236520e165cbbf227cd6371392875237e0c7.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed SWBridge framework (left) and the Sliced Wasserstein distance calculation process (right)." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 524, + 345 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4d44f98b684a3c9e393ec78b641d1b754f9f2e02dacbedee2631a771fb76644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 88, + 681, + 343 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b29e81eceee9d52640a0f0f236c5595caef8d913c35c350151ff7d0954387c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 88, + 883, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "architecture as a class-agnostic feature extractor. The Mask2Former comprises three core components: an image encoder $f_{\\mathrm{CLIP}}$ , a pixel decoder $f_{\\mathrm{decode}}$ , and a Transformer decoder $f_{\\mathrm{Transformer}}$ . Specifically, the frozen CLIP model $f_{\\mathrm{CLIP}}$ processes each frame $V_{t}$ independently to extract high-level visual features $F_{t}$ . These features are then input to the pixel decoder $f_{\\mathrm{decode}}$ , which upsamples them to produce a multiscale feature map $D_{t}$ preserving spatial details. Furthermore, a set of learnable queries $Q$ is initialized and fed into the Transformer decoder $f_{\\mathrm{Transformer}}$ along with the high-resolution feature map $D_{t}$ . The Transformer decoder processes these inputs to generate the query embeddings $E_{t}$ .", + "bbox": [ + 88, + 397, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Spatio-Temporal Bridge", + "text_level": 1, + "bbox": [ + 89, + 590, + 313, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After extracting category-agnostic features, the next challenge is to construct a spatio-temporal bridge to establish the correspondence between instances in consecutive frames $t$ and $t + 1$ . Given the dynamic nature of video content, objects may encounter occlusion, movement, and appearance changes. Therefore, it is particularly important to ensure the accuracy of instance tracking over time. To achieve this, we employ the Sliced Wasserstein metric.", + "bbox": [ + 89, + 614, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, for the query embeddings of frames $t$ and $t + 1$ , we perform random sampling to construct random paths $Z_{i,j} = \\mathbf{q}_{t,i} - \\mathbf{q}_{t + 1,j}$ , where $\\mathbf{q}_{t,i} \\in E_t$ and $\\mathbf{q}_{t + 1,i} \\in E_{t + 1}$ are outputs of the Transformer decoder $f_{\\mathrm{Transformer}}$ . These paths are then normalized to obtain the unit direction $\\theta_{i,j} = \\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2}$ , which serves as the projection direction.", + "bbox": [ + 89, + 737, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we sample a new projection direction $\\hat{\\theta}_{i,j}$ from $\\sigma_{\\kappa}(\\theta_{i,j};\\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2})$ . This introduces randomness, potentially yielding different projection directions each time, which aids in exploring multiple relative positional relation", + "bbox": [ + 89, + 835, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ships between feature points. This, in turn, may enhance the model's robustness and generalization capability.", + "bbox": [ + 511, + 397, + 903, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the projection direction $\\hat{\\theta}_{i,j}$ in hand, we proceed to calculate the one-dimensional Wasserstein distance. This involves projecting $\\mathcal{M}_t$ and $\\mathcal{T}$ along the $\\theta$ direction to obtain one-dimensional distributions $\\theta \\sharp \\mu$ and $\\theta \\sharp \\nu$ . These distributions are then sorted, and the inverse functions of their cumulative distribution functions (CDFs) are computed. Finally, the one-dimensional Wasserstein distance $W_{p}^{p}(\\theta \\sharp \\mu ,\\theta \\sharp \\nu)$ is calculated using the formula:", + "bbox": [ + 511, + 428, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} ^ {p} (F, G) = \\int_ {0} ^ {1} | F ^ {- 1} (t) - G ^ {- 1} (t) | ^ {p} d t, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 556, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F$ and $G$ are the CDFs of $\\theta \\sharp \\mu$ and $\\theta \\sharp \\nu$ , respectively, and $p$ is typically set to 2.", + "bbox": [ + 511, + 595, + 903, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step involves importance weighting and Monte Carlo estimation. We repeat the aforementioned process $L$ times to obtain $L$ projection directions $\\{\\theta_1, \\dots, \\theta_L\\}$ and the corresponding set of one-dimensional Wasserstein distances. Each distance is then weighted with $w$ , where $w$ is determined by an increasing function $f(\\cdot)$ . Increasing function is $f(x) = e^x$ , places higher importance on larger differences, which may highlight instances with more substantial temporal changes. The weighted average of these distances serves as the approximation of SW, calculated using the formula:", + "bbox": [ + 511, + 626, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS W = \\frac {1}{L} \\sum_ {l = 1} ^ {L} w \\cdot \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mathcal {M} _ {t} , \\theta_ {l} \\sharp \\mathcal {T}\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mathcal {M} _ {t} , \\theta_ {j} \\sharp \\mathcal {T}\\right)\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 797, + 903, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the weight $w$ for each projection direction is:", + "bbox": [ + 511, + 844, + 854, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw = \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mu , \\theta_ {j} \\sharp \\nu\\right)\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 864, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq.6 yields a weighted Sliced Wasserstein distance that reflects the relationship between adjacent frames along random projection directions, considering both the geometric positions of data points and their differences through importance weighting.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To establish instance correspondences, we compute the SW values between all pairs of instances across consecutive frames $t$ and $t + 1$ , denoted as $\\mathrm{SW}_p^p (E_t,E_{t + 1})$ . These values quantify the difference between instances along the projection direction, with smaller SW values indicating stronger spatio-temporal consistency. By selecting the instance pairs with the smallest SW values, we identify those with the highest temporal alignment, ensuring robust and meaningful associations for instance tracking. Formally, the instance matching objective is:", + "bbox": [ + 89, + 167, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T i m e}} = \\arg \\min \\mathrm {S W} \\left(E _ {t}, E _ {t + 1}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 329, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the selected pairs correspond to instances with the strongest spatio-temporal consistency, facilitating accurate tracking and correlation across frames. Our method thus emphasizes directions and features that capture significant inter-frame changes, effectively prioritizing instances that reflect spatial and appearance consistency. The adaptive weighting in Eq.6 further enhances tracking by highlighting subtle variations, ensuring robust temporal consistency.", + "bbox": [ + 89, + 356, + 482, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Text-Visual Bridge", + "text_level": 1, + "bbox": [ + 89, + 486, + 272, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After bridging the temporal consistency between the spatial and temporal domains through the spatio-temporal bridge, we are still confronted with the challenge of addressing the modal domain shift between the text and visual domains. Video content, with its dynamic and diverse nature, introduces additional complexity as textual descriptions and visual appearances can vary significantly. To ensure the accurate alignment between text and visual features, which is crucial in this context, we continue to leverage the Sliced Wasserstein metric. This metric has proven effective in our previous spatio-temporal bridging efforts, leading us to adopt it for modal alignment in our current method.", + "bbox": [ + 89, + 508, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first combine the temporal feature $D_{t}$ with the query embeddings $E_{t}$ through multiplication to obtain the visual embeddings $M_{t}$ . Subsequently, we apply a mask pooling operation to derive a class-agnostic mask $\\mathcal{M}_t\\in \\mathbb{R}^{N\\times D}$ based on $M_{t}$ and the frame features $F_{t}$ . Lastly, we obtain the class text embeddings $\\mathcal{H}\\in \\mathbb{R}^{L\\times D}$ generated by the CLIP text encoder, where $L$ represents the number of categories.", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that our cross-modal alignment method bears similarity to the Sliced Wasserstein metric procedure detailed in §4.2. We begin by defining the projection path of the cross-modal module as $Z_{i,j}^{c} = \\pmb{m}_{t,i} - \\pmb{h}_{t,j}$ . To obtain the projection direction, we normalize this path, resulting in $\\theta_{i,j}$ . Given the relatively small sizes of $\\mathcal{M}_t$ and $\\mathcal{H}$ ,", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "this module departs from the random sampling approach and instead adopts a full-path full-projection strategy. Upon successfully obtaining the projection direction $\\theta$ , we proceed to calculate the one-dimensional Wasserstein distance. We then apply importance weighting and incorporate the Monte Carlo estimation method in our subsequent calculations. Following the logic of Eq. 6 and Eq. 7, we ultimately compute the weighted average as an approximation of the Sliced Wasserstein distance.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, similar to Eq. 8, we introduce an additional loss function $\\mathcal{L}_{Modal}$ to constrain the modality alignment process. This loss function is defined as:", + "bbox": [ + 511, + 227, + 903, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {M o d a l}} = \\arg \\min \\operatorname {S W} \\left(\\mathcal {M} _ {t}, \\mathcal {H}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 287, + 903, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Loss Function", + "text_level": 1, + "bbox": [ + 513, + 311, + 658, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mine similarity relations in cross-text video pairs, we draw on the successful experience of cross-modal learning, as demonstrated in [8]. To this end, we introduce a cross-text similarity alignment loss based on contrastive learning. The loss function is formulated as follows:", + "bbox": [ + 511, + 333, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V T}} = - \\frac {1}{N L} \\sum_ {i, j} \\log \\frac {\\exp \\left(d _ {i j} / \\tau\\right)}{\\sum_ {k} \\exp \\left(d _ {i k} / \\tau\\right) + \\sum_ {l} \\exp \\left(d _ {j l} / \\tau\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 417, + 903, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{ij} = \\mathrm{sim}(\\pmb {m}_i,\\pmb {h}_j),\\mathrm{sim}(\\cdot ,\\cdot)$ represents the similarity metric function, and $\\tau$ is the temperature hyperparameter.", + "bbox": [ + 511, + 468, + 903, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Total Loss $\\mathcal{L}_{\\mathrm{Total}}$ is a weighted combination of multiple loss components, which can be written as:", + "bbox": [ + 511, + 500, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T o t a l}} = \\alpha_ {T} \\mathcal {L} _ {\\text {T i m e}} + \\alpha_ {M} \\mathcal {L} _ {\\text {M o d a l}} + \\alpha_ {V} \\mathcal {L} _ {\\mathrm {V T}} + \\alpha_ {S} \\mathcal {L} _ {\\text {S e g}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 541, + 903, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{Seg}}$ represents the binary cross-entropy and dice losses for mask prediction, following the approach in [26]. The coefficients $\\alpha_{T}, \\alpha_{M}, \\alpha_{V}$ , and $\\alpha_{S}$ are the corresponding weighting factors for each loss component.", + "bbox": [ + 511, + 568, + 903, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computational complexity. In our Sliced Wasserstein metric process, the core steps involve matrix operations and sorting procedures. Initially, we generate a set of initial projection vectors by randomly selecting samples and computing the differences among these samples. The complexity of this selection operation is $O(L)$ , while the complexity of the difference calculation is $O(L \\cdot D)$ , where $L$ represents the number of projection vectors and $D$ is the dimensionality of the samples. Subsequently, we proceed to the normalization and resampling phase, which also maintains a complexity of $O(L \\cdot D)$ . When calculating the one-dimensional Wasserstein distance, we need to perform matrix multiplication, a step with a complexity of $O(N \\cdot D \\cdot L)$ , where $N$ is the number of samples. Additionally, we need to sort the projected data, with a sorting operation complexity of $O(N \\log N)$ . In summary, the overall complexity of our method can be approximated as $O(N \\cdot D \\cdot L + N \\log N)$ . In practical applications, as the number of samples $N$ and the number of", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12474", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/85e66470a14b8dde44e58982d0305e40405cc70360334052215d11f2473584b3.jpg", + "table_caption": [ + "Table 1. Statistics of the dataset. Cat. and Ins. represent the number of categories and instances respectively." + ], + "table_footnote": [], + "table_body": "
DatasetVideosCat.Ins.TrainVal.Test
LV-VIS4,8281,196-3,083837908
YT-VIS 192,883404,8832,238302343
YT-VIS 213,859408,1712,985421453
OVIS901255,223607140154
BURST2,91448216,0005009931,421
", + "bbox": [ + 91, + 128, + 486, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "projection vectors $L$ increase, the computational cost of matrix multiplication will become dominant. Although sorting and other operations with linear complexity also exist, their contribution to the overall complexity is relatively small.", + "bbox": [ + 89, + 262, + 482, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 337, + 222, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 362, + 377, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initially train our method on a combined set of common and frequent categories derived from the LVIS dataset [16]. Subsequently, we assess the performance of our method on both the validation and test sets of the LV-VIS dataset [26], as well as the validation set of several datasets, namely OVIS [21], YT-VIS 2019/2021 [27], and BURST [1]. The statistics of the evaluation dataset are detailed in Table 1.", + "bbox": [ + 89, + 383, + 482, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LVIS [16] dataset comprises 1,203 object categories, which significantly exceeds the number of categories in COCO. Following ViLD[14], we select 866 frequent and common categories as the training categories and reserve the remaining 337 rare categories as novel categories.", + "bbox": [ + 89, + 489, + 482, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following [26], we present the mean Average Precision $(AP)$ for all categories in aggregate. Additionally, we report on $AP_{n}$ for the novel category.", + "bbox": [ + 89, + 566, + 482, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 621, + 307, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our implementation, the CLIP image encoder and text encoder are kept frozen, whereas the pixel decoder, transformer decoder, and mask prediction head undergo training. This strategy effectively harnesses the pre-trained capabilities of CLIP for mask classification and query matching, utilizing query embeddings, thereby obviating the need for category and identity annotations. The CLIP model employed in our framework adopts the ViTB/32 architecture for both the text and image encoders. Throughout the training process, the parameters of these encoders remain frozen. For optimization purposes, we utilize the AdamW optimizer, with the base learning rate set to $1 \\times 10^{-4}$ . $\\tau = 0.05$ , $\\alpha_{T}$ , $\\alpha_{M}$ , $\\alpha_{V}$ , and $\\alpha_{S}$ are all set to 1 based on grid parameter tuning experience. Following a step-wise learning rate schedule, this rate is subsequently reduced by a factor of 10 at $90\\%$ and $95\\%$ of the total training steps. To assess the robustness and generality of our proposed method, we con", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "duct experiments using two distinct backbones: ResNet-50 and SwinB. Our experimental setup is analogous to that of OV2Seg [26]. All experiments are conducted on a high-performance machine equipped with eight NVIDIA A800 GPUs, each possessing 80GB of memory. To facilitate a fair comparison, we also present the inference speed of our method when executed on a single NVIDIA A100 GPU.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 209, + 651, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results presented in Table 2 provide a comprehensive comparison of our proposed SWBridge method with several state-of-the-art approaches across multiple video instance segmentation datasets. A significant finding reveals that SWBridge exhibited superior performance compared to existing methods across 19 out of 22 evaluation metrics in various settings. Furthermore, it attained suboptimal results in the remaining three metrics, thereby underscoring its efficacy in addressing the inherent complexity associated with Open-Vocabulary VIS. This result not only verifies the effectiveness of SWBridge in its design, but also shows that it can adapt well to the characteristics of different datasets, such as changes in video length, number of instances, scene complexity, etc.", + "bbox": [ + 511, + 233, + 903, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LV-VIS. When trained on the diverse and extensive LVIS dataset, which includes a large number of frequent and common categories, SWBridge exhibits robust generalization capabilities. This is particularly evident in its performance on the LV-VIS dataset, where it achieves the highest $AP$ for both the validation and test sets, including the novel categories $(AP_{n})$ . The superior performance on LV-VIS underscores SWBridge's ability to effectively recognize and segment objects from both base and novel categories, even in the presence of a significant number of unseen classes during training. Although powerful backbone networks such as ConvNeXt-B can enhance performance, we find that the Swin-B backbone network, when combined with the SWBridge, yields even better results. This indicates that SWBridge has a synergistic effect with the backbone network, enabling more efficient capture of spatio-temporal features in videos, and the performance improvement exceeds that achieved by mere network upgrades.", + "bbox": [ + 511, + 446, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "YT-VIS. SWBridge demonstrates superior performance on the YT-VIS2019 and YT-VIS2021 datasets, further attesting to its strong capability in handling video instance segmentation tasks with a moderate number of categories. Remarkably, this method achieves competitive performance levels without the need for fine-tuning on specific video datasets. This characteristic indicates that SWBridge can effectively leverage image-based training data and successfully generalize to video data, thereby significantly reducing the reliance on extensive video-specific annotations. In comparison with state-of-the-art methods, although OV-Former achieves excellent results by incorporating an ad", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12475", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c84e31c1e966cc2992ac9f77e51ac7f3734967614cb2f0640cfb4c1fa4195926.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art methods. All the methods are trained on image dataset LVIS and evaluated on the video instance segmentation datasets directly, i.e., they are not fine-tuned using the training set of each dataset." + ], + "table_footnote": [], + "table_body": "
MethodBackboneLV-VIS val.LV-VIS testYT-VIS2019YT-VIS2021OVISBURST
APAPnAPAPnAPAPnAPAPnAPAPAPAP
DetPro[10]-SORT[2]R506.43.55.82.1--------
Detric[30]-SORT[2]R506.53.45.72.114.63.512.73.16.71.92.5
DetPro[10]-OWTB[20]R507.94.27.02.9-------
Detric[30]-OWTB[20]R507.74.27.02.817.94.516.75.89.02.71.8
Detric[30]-XMem[7]R508.85.47.73.6-------
OV2Seg[26]R5014.211.911.48.927.211.123.67.311.23.72.4
OVFormer[11]R50+ViT-B----34.816.529.815.715.1--
CLIP-VIS [32]R5019.524.214.615.932.223.830.117.914.15.27.7
SWBridge (Our)R5020.724.915.416.233.424.030.918.415.25.58.2
Detric[30]-SORT[2]SwinB12.86.69.44.723.87.921.69.811.72.51.0
Detric[30]-OWTB[20]SwinB14.58.511.86.130.09.727.111.413.63.92.4
Detric[30]-XMem[7]SwinB16.310.613.17.7-------
OV2Seg[26]SwinB21.116.316.411.537.621.333.918.217.54.93.0
OVFormer[11]SwinB+ViT-B----44.321.537.618.321.3--
CLIP-VIS [32]ConvNeXt-B32.240.225.330.642.127.537.922.018.58.312.7
SWBridge (Our)SwinB33.040.526.031.043.028.038.522.519.78.312.8
", + "bbox": [ + 106, + 128, + 893, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ditional CLIP image encoder, SWBridge still outperforms CLIP-VIS in terms of the $AP$ metric under similar model architectures. More notably, in the $AP_{n}$ metric, which pertains to the segmentation of new class instances, our bridge strategy comprehensively surpasses the performance gains brought about by the CLIP encoder.", + "bbox": [ + 88, + 474, + 480, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OVIS. Since the categories of the OVIS dataset highly overlap with the LVIS dataset used for training, we only reports the AP (average precision) performance metric. SWBridge even outperformed OVFormer with the extra CLIP on the R50 backbone and was second only to OVFoemer on the SwinB backbone. SWBridge's ability to maintain high accuracy under these conditions demonstrates its robustness in complex visual environments, where the visibility of objects changes significantly over time.", + "bbox": [ + 88, + 566, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BURST. The BURST dataset, with its diverse and uncommon categories, poses a unique challenge due to the wide range of object types. SWBridge's competitive performance on this dataset, especially in terms of $AP_{n}$ , suggests that it can effectively manage the intricacies of uncommon object categories, enhancing its applicability to real-world scenarios where encountering unseen objects is common.", + "bbox": [ + 88, + 702, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 816, + 243, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our ablation experiments continue to use the LVIS dataset for training with ResNet50 as the Backbone and are evaluated on the validation sets of the LV-VIS, OVIS, and BURST datasets, aiming to measure the performance of the", + "bbox": [ + 88, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ecd91e3f6ebb42c77f25f0323ea2e446e497c2c651fcfc6cdd6f122b2f1e72bf.jpg", + "table_caption": [ + "Table 3. Comparison of ablation results of bridge strategy." + ], + "table_footnote": [], + "table_body": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Baseline8.510.25.32.62.3
+ Spatio-Temporal Bridge15.018.010.23.15.4
+ Text-Visual Bridge19.923.314.25.37.9
+ LVT in Eq. 1020.724.915.25.58.2
", + "bbox": [ + 514, + 498, + 923, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proposed SWBridge more comprehensively and rigorously. Effectiveness of Bridge Strategy. We first established a baseline variant that removed the Spatio-Temporal Bridge bridge and Text-Visual Bridge modules, while discarding the additional contrast loss $\\mathcal{L}_{VT}$ and retaining only the basic $\\mathcal{L}_{Seg}$ loss function. In order to achieve instance association, we adopted the memory module proposed in OV2Seg to replace the original spatiotemporal bridge function.", + "bbox": [ + 511, + 638, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The experimental results presented in Table 3 confirm the significant effectiveness of each component in the SWBridge method. Specifically, upon integrating the spatio-temporal bridge module into the baseline variant, we observed notable improvements in both $AP$ and $AP_{n}$ metrics across all datasets. This outcome indicates that the spatio-temporal bridge module efficiently captures spatiotemporal correlations between video frames, thereby aiding the model in more accurately understanding and tracking", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12476", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/993e0530489659d6d13e78cd27585c76de1c3919412b7a21e41355b4fc957793.jpg", + "table_caption": [ + "Table 4. Impact of ablation results of Importance Weighting." + ], + "table_footnote": [], + "table_body": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
w/o Weight15.317.012.34.55.3
w/o Weight in §4.118.522.313.74.96.8
w/o Weight in §4.319.623.514.55.27.5
SWBridge(Our)20.724.915.25.58.2
", + "bbox": [ + 104, + 114, + 467, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4e5d229e4a30ec6b871b4e9f07e8f604f4e938c1d8d5bc1911854923e67b1d61.jpg", + "table_caption": [ + "Table 5. Impact of ablation results of Importance Weighting." + ], + "table_footnote": [], + "table_body": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Huberized18.822.413.84.76.4
Cosine18.622.213.94.86.5
Wasserstein19.523.314.45.17.0
Sliced Wasserstein20.724.915.25.58.2
", + "bbox": [ + 102, + 263, + 467, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the dynamic changes of instances within videos. Furthermore, the model's performance was significantly enhanced once again after adding the text-visual bridge module on top of the spatio-temporal bridge module. By introducing textual information, the text-visual bridge provides the model with rich semantic cues, greatly enhancing its accuracy in recognizing and understanding instances. Additionally, $\\mathcal{L}_{VT}$ synergizes with other bridge strategy modules, collectively driving a comprehensive improvement in the model's performance and generalization.", + "bbox": [ + 89, + 398, + 482, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of SW Weighting Mechanism. Table 4 shows that without a weighting mechanism in the spatio-temporal or text-visual bridges, the model struggles to utilize key information from instances with marked feature differences, limiting performance. However, incorporating the weighting mechanism into both bridges notably improves model performance. This indicates the mechanism helps the model focus on crucial instances with significant feature differences, assigning them greater weights during distance calculation. This adjustment enhances the model's accuracy and efficiency in processing complex data.", + "bbox": [ + 89, + 550, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of Distance Measurement. In Tab. 5, we compare distance metrics for SWBridge. Huberized distance is sensitive to large differences, Cosine similarity ignores feature magnitudes, and Wasserstein distance misses local details. Our SW distance comprehensively considers high-dimensional feature differences with importance weighting, making it robust for complex domain shifts.", + "bbox": [ + 89, + 718, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity analysis of $L$ . In the bridge strategy, we ablated the number of sampling iterations to explore its relationship with model performance (Fig.2). Results show a nonlinear relationship: increasing iterations significantly improves $AP$ and $AP_{n}$ across all datasets. However, excessive", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/27355c8d0edde4c5a96964f2667359330784395772273dd47b5c5044581ddc2a.jpg", + "image_caption": [ + "Figure 2. Ablation experiment on the number of random projection sampling $L$ ." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 897, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "iterations introduce noise and computational burden, potentially causing overfitting or reducing training efficiency.", + "bbox": [ + 511, + 339, + 903, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 375, + 633, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an innovative Sliced Wasserstein Bridging Learning Framework (SWbridge), aimed at tackling the challenging task of Open-Vocabulary Video Instance Segmentation. Inspired by metric learning, our framework utilizes the Sliced Wasserstein distance as a core tool to effectively bridge the spatial-temporal and text-visual domains involved in the task. Through domain alignment, we successfully map features from different domains into a unified metric space, preserving temporal consistency and learning intrinsic consistent features across modalities. This significantly improves the fusion of text and visual information, addressing one of the core challenges in Open-Vocabulary VIS. Looking ahead, we plan to explore better optimization methods for the Sliced Wasserstein distance to further enhance the performance of our framework.", + "bbox": [ + 511, + 398, + 906, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in part by the China Postdoctoral Science Foundation (No. 2024M761805), the Postdoctoral Fellowship Program of CPSF (No. GZB20250390), the Qingdao Postdoctoral Innovation Project (No. QDBSH20240201015), the Major Science and Technology Innovation Projects in Shandong Province (No. 2024CXGC010108).", + "bbox": [ + 511, + 627, + 905, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 747, + 609, + 762 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 6", + "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 7", + "[3] Nicolas Bonneel, Julien Rabin, Gabriel Peyre, and Hanspeter Pfister. Sliced and radon wasserstein barycenters of mea" + ], + "bbox": [ + 521, + 772, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12477", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sures. Journal of Mathematical Imaging and Vision, 51:22-45, 2015. 2", + "[4] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In ECCV, 2020. 1", + "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3", + "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2", + "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7", + "[8] Zesen Cheng, Kehan Li, Hao Li, Peng Jin, Chang Liu, Xiawu Zheng, Rongrong Ji, and Jie Chen. Instance brownian bridge as texts for open-vocabulary video instance segmentation. arXiv preprint arXiv:2401.09732, 2024. 1, 2, 5", + "[9] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 2", + "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 7", + "[11] Hao Fang, Peng Wu, Yawei Li, Xinxin Zhang, and Xiankai Lu. Unified embedding alignment for open-vocabulary video instance segmentation. In ECCV, 2024. 1, 2, 3, 7", + "[12] Damien Garreau, Rémi Lajugie, Sylvain Arlot, and Francis Bach. Metric learning for temporal sequence alignment. NeurIPS, 2014. 1", + "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2", + "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.6", + "[15] Pinxue Guo, Tony Huang, Peiyang He, Xuefeng Liu, Tianjun Xiao, Zhaoyu Chen, and Wenqiang Zhang. Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835, 2023. 1, 2", + "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 6", + "[17] Fusheng Hao, Fengxiang He, Jun Cheng, Lei Wang, Jianzhong Cao, and Dacheng Tao. Collect and select: Semantic alignment metric learning for few-shot learning. In ICCV, 2019. 1", + "[18] Dat Huynh, Jason Kuen, Zhe Lin, Jieuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In CVPR, 2022. 2", + "[19] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In CVPR, 2023. 3" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 7", + "[21] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, 130(8): 2022-2039, 2022. 6", + "[22] Zheyun Qin, Xiankai Lu, Dongfang Liu, Xiushan Nie, Yi long Yin, Jianbing Shen, and Alexander C Loui. Reformulating graph kernels for self-supervised space-time correspondence learning. IEEE Transactions on Image Processing, 32: 6543-6557, 2023. 1", + "[23] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1192-1208, 2023. 1", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3", + "[25] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 2", + "[26] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 2, 3, 5, 6, 7", + "[27] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 6", + "[28] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In ICCV, 2023. 1", + "[29] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In CVPR, 2021. 2", + "[30] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7", + "[31] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[32] Wenqi Zhu, Jiale Cao, Jin Xie, Shuangming Yang, and Yanwei Pang. Clip-vis: Adapting clip for open-vocabulary video instance segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 757 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12478", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_model.json b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c766430f60957e4bcb6d5ca0f2c4a9c82e0d0d12 --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_model.json @@ -0,0 +1,1879 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.888, + 0.153 + ], + "angle": 0, + "content": "Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.188, + 0.762, + 0.209 + ], + "angle": 0, + "content": "Zheyun Qin\\(^{1}\\), Deng Yu\\(^{2}\\), Chuanchen Luo\\(^{2}\\), Zhumin Chen\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.211, + 0.871, + 0.226 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Shandong University \\(^{2}\\)School of Artificial Intelligence, Shandong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.261, + 0.327, + 0.277 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.293, + 0.485, + 0.701 + ], + "angle": 0, + "content": "In recent years, researchers have explored the task of open-vocabulary video instance segmentation, which aims to identify, track, and segment any instance within an open set of categories. The core challenge of Open-Vocabulary VIS lies in solving the cross-domain alignment problem, including spatial-temporal and text-visual domain alignments. Existing methods have made progress but still face shortcomings in addressing these alignments, especially due to data heterogeneity. Inspired by metric learning, we propose an innovative Sliced Wasserstein Bridging Learning Framework. This framework utilizes the Sliced Wasserstein distance as the core tool for metric learning, effectively bridging the four domains involved in the task. Our innovations are threefold: (1) Domain Alignment: By mapping features from different domains into a unified metric space, our method maintains temporal consistency and learns intrinsic consistent features between modalities, improving the fusion of text and visual information. (2) Weighting Mechanism: We introduce an importance weighting mechanism to enhance the discriminative ability of our method when dealing with imbalanced or significantly different data. (3) High Efficiency: Our method inherits the computational efficiency of the Sliced Wasserstein distance, allowing for online processing of large-scale video data while maintaining segmentation accuracy. Through extensive experimental evaluations, we have validated the robustness of our concept and the effectiveness of our framework." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.878 + ], + "angle": 0, + "content": "In recent years, video instance segmentation methods [23, 27, 28] have demonstrated significant improvements and breakthroughs in various application fields. However, most existing methods rely on the Closed-Set Assumption [4, 22], where the learning and recognition scope of the model are strictly limited to known categories or labels during training and testing. This assumption ignores the possibility of unknown categories in the real world, limiting the model's ability to effectively detect unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.905, + 0.383 + ], + "angle": 0, + "content": "To overcome this limitation, researchers have begun to explore the task of open-vocabulary video instance segmentation [15, 26] (Open-Vocabulary VIS), which aims to identify, track, and segment any instance within an open set of categories. These Open-Vocabulary VIS methods can not only handle traditional predefined categories but also flexibly adapt to new user-specified categories, achieving the advanced function of on-demand segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.656 + ], + "angle": 0, + "content": "The core of Open-Vocabulary VIS methods lies in solving the cross-domain alignment problem, including spatial-temporal domain alignment and text-visual domain alignment. Existing methods have different focuses. In terms of spatial-temporal domain alignment, current methods extract class-agnostic instance features based on the foundational model Mask2Former-VIS. Furthermore, OV2Seg [26], OpenVIS [15], and CLIP-VIS [32] employ memory-based approaches to perform cross-frame instance matching. In terms of text-visual domain alignment, existing methods mainly rely on the zero-shot performance exhibited by visual-language models, i.e., CLIP [24] pretrained on largescale image-text pairs. Specifically, BriVIS [8] and OV-Former [11] introduce additional CLIP image encoders on top of visual feature extractors. The difference lies in that BriVIS integrates them based on contrastive learning, while OV-Former maps text and visual features into the same space vis the attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Despite the significant progress made by existing Open-Vocabulary VIS methods, there are still many shortcomings. The cross-domain alignment problem remains inadequately addressed, and the inherent data heterogeneity results in significant differences in feature distributions across different domains. In light of this, we cannot help but trace back to the development history of traditional machine learning methods, seeking a tool that can simultaneously solve spatial-temporal domain alignment and text-visual domain alignment. We find that the metric learning method in the field of machine learning provides an effective approach to addressing the alignment problem [12, 17]. By learning a suitable metric space, this paradigm makes similar samples closer while dissimilar samples relatively farther apart, providing a powerful tool for solving the alignment problem." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Guided by the idea of metric learning, we propose an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.316, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Zhumin Chen." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12470" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "innovative Sliced Wasserstein Bridging Learning Framework, i.e., SWbridge. This framework utilizes the Sliced Wasserstein distance [3] as the core tool for metric learning. This framework cleverly utilizes the Sliced Wasserstein distance as the core tool for metric learning. Specifically, we minimize the Sliced Wasserstein distance between samples based on random paths through sampling, achieving feature embedding alignment. We bridge four domains based on a strategy to address the challenges of cross-domain alignment (spatial-temporal domain and text-visual domain)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.682 + ], + "angle": 0, + "content": "Our innovations are mainly concentrated in three aspects: (1) Domain Alignment. The Sliced Wasserstein distance has the advantage of accurately capturing differences in data distributions in high-dimensional spaces. Our method uses the Sliced Wasserstein distance as a bridge to map features from different domains into a unified metric space. This not only maintains temporal consistency by comparing instance embeddings between adjacent frames, achieving temporal alignment and effectively mitigating challenges such as progressive occlusion, but also learns intrinsic consistent features between modalities, narrowing the semantic gap and improving the fusion of text and visual information through modality alignment. (2) Weighting Mechanism: We introduce an importance weighting mechanism to further enhance its discriminative ability when dealing with imbalanced data or data with significant feature differences. In open-vocabulary video instance segmentation tasks, there may be significant feature differences between different instances, and some instances may be more critical or informative. Through the importance weighting mechanism, our method can dynamically adjust the contribution of different instances in the metric space, allowing key instances to have greater weight in distance calculations. (3) High Efficiency: Our SWbridge inherits the computational efficiency advantage of the Sliced Wasserstein distance, making it highly efficient when processing large-scale video data. Without introducing additional network parameters, our method can handle video in an online manner while maintaining segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Combining the aforementioned perspectives and innovations, we propose an open-vocabulary video instance segmentation method capable of bridging multiple domains. This method has achieved remarkable results on multiple datasets related to open-vocabulary and video tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.242, + 0.792 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.483, + 0.818 + ], + "angle": 0, + "content": "2.1. Open-Vocabulary Detection and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Open Vocabulary Detection (OVD) [9, 25] and Open Vocabulary Segmentation (OVS) [13, 18] are cutting-edge tasks in the field of computer vision, enabling models to be trained on images containing unannotated novel objects, thereby breaking the closed-set constraint. This break-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.409 + ], + "angle": 0, + "content": "through is primarily attributed to the application of weak supervision signals, namely the utilization of image-text pairs (such as image-caption pairs and image-level labels) or large pre-trained Visual-Language Models (VLMs), such as CLIP [24]. Supported by weak supervision signals, methods for OVD and OVS can be categorized into four main types [31]. The first type is region-aware training, which learns object feature representations by exploring the intrinsic links between image regions and text descriptions without relying on the VLMs' image encoder (VLMsIE) or direct object annotations. The second type is the pseudo-labeling method, which enhances model generalization by generating pseudo-labels from preliminary predictions on unannotated images using both image-text pairs and the VLMs-IE. The other two types are knowledge distillation and transfer learning. Knowledge distillation transfers knowledge from pre-trained VLMs to new models or tasks via a distillation mechanism, while transfer learning applies knowledge learned by pre-trained models on specific tasks to new tasks or domains. Both rely on the VLMsIE and seldom involve direct training on image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Furthermore, zero-shot learning [29] addresses the closed-set constraint in scene-aware tasks and differs from open vocabulary methods. It prohibits access to weak supervision signals during training, but can be converted into Open-Vocabulary tasks upon gaining access, enhancing the model's generalization and recognition of novel objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.905, + 0.532 + ], + "angle": 0, + "content": "2.2. Open-Vocabulary Video Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Open-vocabulary video instance segmentation, as an emerging vision task, aims to simultaneously classify, track, and segment arbitrary objects within open categories in videos, attracting considerable attention from researchers in recent years. OV2Seg [26] has taken the lead in this field by not only constructing the Large Vocabulary Video Instance Segmentation (LV-VIS) dataset but also proposing the first end-to-end benchmark for open-vocabulary video instance segmentation. OV2Seg method leverages the Mask2Former [6] framework to extract class-agnostic masks and query embeddings, utilizes the CLIP text encoder for precise mask classification, and achieves cross-frame instance tracking through a long-term matching strategy. Subsequent research has mostly followed this basic approach and expanded upon it. OpenVIS [15] introduces a two-stage framework, OVIS, which adopts a neighboring matching strategy for instance tracking, effectively simplifying the processing pipeline. BriVIS [8] models instances as Brownian bridges and closely aligns bridge-level instance representations with category texts through contrastive learning, also employing a neighboring matching strategy to enhance the accuracy of instance recognition. OVFormer [11] introduces a novel unified embedding alignment module that effectively addresses the domain gap between instance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "queries and text embeddings, demonstrating good practicality with its semi-online processing approach. However, the aforementioned methods primarily focus on bridging the domain discrepancy between text and vision, with insufficient attention to spatiotemporal consistency among cross-frame instances. CLIP-VIS [32] improves upon this by introducing a time-topK enhanced matching strategy, which strengthens temporal modeling capabilities between frames and improves matching accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.379 + ], + "angle": 0, + "content": "We believe that the core of the open-vocabulary video instance segmentation task lies in simultaneously addressing the alignment of the spatial-temporal domain and the text-visual domain. To this end, based on the concept of metric learning, we employ the Sliced Wasserstein distance to bridge various domains. This approach not only effectively narrows the domain gap between modalities but also ensures spatio-temporal consistency of instances across frames, providing a novel solution for the open-vocabulary video instance segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.385, + 0.318, + 0.401 + ], + "angle": 0, + "content": "2.3. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.695 + ], + "angle": 0, + "content": "The core of Visual Language Models (VLMs) [5] is their deep training on large-scale image-text paired data, enabling the fusion and understanding of visual and textual information, and endowing them with powerful zero-shot object recognition capabilities [19]. CLIP [24], a prominent VLM, uses an image encoder to extract visual features and a text encoder to generate text embeddings, jointly constructing a cross-modal representation space. However, applying VLMs to open-vocabulary video instance segmentation is challenging. VLMs, primarily trained on images, struggle with understanding dynamic video scenes rich in spatiotemporal information, and lack the capability to maintain object consistency across frames. Additionally, the domain shift between text and vision poses a significant challenge, potentially leading to biases in mapping textual information to the visual space. To overcome these limitations, we propose a progressive approach that addresses the spatiotemporal consistency issue and the domain shift problem, enabling VLMs to better adapt to video instance segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.228, + 0.721 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.398, + 0.746 + ], + "angle": 0, + "content": "3.1. Open-Vocabulary VIS Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.483, + 0.857 + ], + "angle": 0, + "content": "Given a test video \\( \\mathcal{D}_{test} \\) with \\( T \\) frames, the objective in Open-Vocabulary VIS is to accurately predict all \\( N \\) instances belonging to the categories in \\( \\mathcal{C} = \\mathcal{C}_{base} \\cup \\mathcal{C}_{novel} \\) by the trained model \\( f_{\\theta} \\). \\( \\mathcal{C}_{base} \\) is the set of base (training) categories, and \\( \\mathcal{C}_{novel} \\) is the set of novel categories that are not seen during training but may appear in the test videos. The prediction can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.863, + 0.482, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\left\\{m _ {1}, m _ {2}, \\dots , m _ {T} \\right\\}, c \\right\\} _ {n} ^ {N} = f _ {\\theta} \\left(\\mathcal {D} _ {t e s t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.45, + 0.902 + ], + "angle": 0, + "content": "where \\(\\{m_t\\}_{t = 1}^T\\) is the segmentation masks, and \\(c\\in \\mathcal{C}\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.764, + 0.106 + ], + "angle": 0, + "content": "3.2. Sliced Wasserstein Distance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.158 + ], + "angle": 0, + "content": "One-dimensional Wasserstein Distance. For one-dimensional probability measures \\(\\mu\\) and \\(\\nu\\) in \\(\\mathcal{P}_p(\\mathbb{R})\\), the \\(p\\)-Wasserstein distance is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.166, + 0.905, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathrm {W} _ {p} ^ {p} (\\mu , \\nu) = \\int_ {0} ^ {1} | F _ {\\mu} ^ {- 1} (z) - F _ {\\nu} ^ {- 1} (z) | ^ {p} d z, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.209, + 0.905, + 0.27 + ], + "angle": 0, + "content": "where \\(F_{\\mu}\\) and \\(F_{\\nu}\\) are the cumulative distribution functions (CDFs) of \\(\\mu\\) and \\(\\nu\\). This formulation provides a closed form for computing the Wasserstein distance in one-dimensional spaces, making it well-suited for projected measures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.27, + 0.906, + 0.362 + ], + "angle": 0, + "content": "Sliced Wasserstein Distance. To generalize the Wasserstein distance to higher-dimensional measures, the Sliced Wasserstein Distance (SWD) projects the measures \\(\\mu\\) and \\(\\nu\\) in \\(\\mathcal{P}_p(\\mathbb{R}^d)\\) onto one-dimensional subspaces, and then averages the one-dimensional Wasserstein distances from these projections. For \\(\\mu, \\nu \\in \\mathcal{P}_p(\\mathbb{R}^d)\\), the SWD is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.37, + 0.905, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S W} _ {p} ^ {p} (\\mu , \\nu) = \\mathbb {E} _ {\\theta \\sim \\mathcal {U} \\left(\\mathbb {S} ^ {d - 1}\\right)} \\left[ \\mathrm {W} _ {p} ^ {p} \\left(\\theta \\sharp \\mu , \\theta \\sharp \\nu\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.906, + 0.472 + ], + "angle": 0, + "content": "where \\(\\theta \\sharp \\mu\\) and \\(\\theta \\sharp \\nu\\) represent the push-forward measures of \\(\\mu\\) and \\(\\nu\\) along direction \\(\\theta \\in \\mathbb{S}^{d - 1}\\), the unit sphere in \\(\\mathbb{R}^d\\). This projection \\(f(x) = \\theta^{\\top}x\\) maps points from \\(\\mathbb{R}^d\\) to \\(\\mathbb{R}\\), enabling the computation of Wasserstein distances in a one-dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.905, + 0.519 + ], + "angle": 0, + "content": "Since the expectation in Eq. 3 is computationally expensive, SWD is typically approximated by averaging over \\(L\\) independent directions, \\(\\theta_{1},\\ldots ,\\theta_{L}\\), sampled from \\(\\mathcal{U}(\\mathbb{S}^{d - 1})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.529, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbf {S W}} _ {p} ^ {p} (\\mu , \\nu ; L) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {W} _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.576, + 0.906, + 0.621 + ], + "angle": 0, + "content": "where each \\(\\theta_{l}\\sharp \\mu\\) and \\(\\theta_{l}\\sharp \\nu\\) are projected representations of \\(\\mu\\) and \\(\\nu\\) along the direction \\(\\theta_{l}\\). The number of projections \\(L\\) controls the accuracy of the Monte Carlo approximation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.605, + 0.648 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.81 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, we propose the SWbridge framework. Firstly, in §4.1, we integrate a category-agnostic feature extractor. Then, in §4.2, we introduce a spatio-temporal bridging module that robustly establishes instance associations between adjacent frames, ensuring the consistency and coherence of temporal information. Finally, in §4.3, we design a modal bridging module that performs fine-grained semantic mapping across the text and visual domains to enhance cross-domain understanding and effectively address the challenges posed by domain shifts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.803, + 0.835 + ], + "angle": 0, + "content": "4.1. Class-agnostic Feature Extractor" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Given a video \\(\\mathcal{D}\\) consisting of \\(T\\) frames, \\(\\{V_t \\in \\mathbb{R}^{3 \\times H \\times W}\\}_{t=1}^T\\), where \\(H\\) and \\(W\\) denote the height and width of each frame, our model follows the conventional paradigm [11, 26, 32] by adopting the Mask2Former" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.526, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.089, + 0.683, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.089, + 0.885, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.357, + 0.895, + 0.373 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed SWBridge framework (left) and the Sliced Wasserstein distance calculation process (right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.484, + 0.581 + ], + "angle": 0, + "content": "architecture as a class-agnostic feature extractor. The Mask2Former comprises three core components: an image encoder \\( f_{\\mathrm{CLIP}} \\), a pixel decoder \\( f_{\\mathrm{decode}} \\), and a Transformer decoder \\( f_{\\mathrm{Transformer}} \\). Specifically, the frozen CLIP model \\( f_{\\mathrm{CLIP}} \\) processes each frame \\( V_{t} \\) independently to extract high-level visual features \\( F_{t} \\). These features are then input to the pixel decoder \\( f_{\\mathrm{decode}} \\), which upsamples them to produce a multiscale feature map \\( D_{t} \\) preserving spatial details. Furthermore, a set of learnable queries \\( Q \\) is initialized and fed into the Transformer decoder \\( f_{\\mathrm{Transformer}} \\) along with the high-resolution feature map \\( D_{t} \\). The Transformer decoder processes these inputs to generate the query embeddings \\( E_{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.592, + 0.315, + 0.609 + ], + "angle": 0, + "content": "4.2. Spatio-Temporal Bridge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.484, + 0.736 + ], + "angle": 0, + "content": "After extracting category-agnostic features, the next challenge is to construct a spatio-temporal bridge to establish the correspondence between instances in consecutive frames \\( t \\) and \\( t + 1 \\). Given the dynamic nature of video content, objects may encounter occlusion, movement, and appearance changes. Therefore, it is particularly important to ensure the accuracy of instance tracking over time. To achieve this, we employ the Sliced Wasserstein metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Initially, for the query embeddings of frames \\(t\\) and \\(t + 1\\), we perform random sampling to construct random paths \\(Z_{i,j} = \\mathbf{q}_{t,i} - \\mathbf{q}_{t + 1,j}\\), where \\(\\mathbf{q}_{t,i} \\in E_t\\) and \\(\\mathbf{q}_{t + 1,i} \\in E_{t + 1}\\) are outputs of the Transformer decoder \\(f_{\\mathrm{Transformer}}\\). These paths are then normalized to obtain the unit direction \\(\\theta_{i,j} = \\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2}\\), which serves as the projection direction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Subsequently, we sample a new projection direction \\(\\hat{\\theta}_{i,j}\\) from \\(\\sigma_{\\kappa}(\\theta_{i,j};\\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2})\\). This introduces randomness, potentially yielding different projection directions each time, which aids in exploring multiple relative positional relation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.429 + ], + "angle": 0, + "content": "ships between feature points. This, in turn, may enhance the model's robustness and generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.552 + ], + "angle": 0, + "content": "With the projection direction \\(\\hat{\\theta}_{i,j}\\) in hand, we proceed to calculate the one-dimensional Wasserstein distance. This involves projecting \\(\\mathcal{M}_t\\) and \\(\\mathcal{T}\\) along the \\(\\theta\\) direction to obtain one-dimensional distributions \\(\\theta \\sharp \\mu\\) and \\(\\theta \\sharp \\nu\\). These distributions are then sorted, and the inverse functions of their cumulative distribution functions (CDFs) are computed. Finally, the one-dimensional Wasserstein distance \\(W_{p}^{p}(\\theta \\sharp \\mu ,\\theta \\sharp \\nu)\\) is calculated using the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.557, + 0.905, + 0.592 + ], + "angle": 0, + "content": "\\[\nW _ {p} ^ {p} (F, G) = \\int_ {0} ^ {1} | F ^ {- 1} (t) - G ^ {- 1} (t) | ^ {p} d t, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.904, + 0.626 + ], + "angle": 0, + "content": "where \\(F\\) and \\(G\\) are the CDFs of \\(\\theta \\sharp \\mu\\) and \\(\\theta \\sharp \\nu\\), respectively, and \\(p\\) is typically set to 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.793 + ], + "angle": 0, + "content": "The next step involves importance weighting and Monte Carlo estimation. We repeat the aforementioned process \\( L \\) times to obtain \\( L \\) projection directions \\( \\{\\theta_1, \\dots, \\theta_L\\} \\) and the corresponding set of one-dimensional Wasserstein distances. Each distance is then weighted with \\( w \\), where \\( w \\) is determined by an increasing function \\( f(\\cdot) \\). Increasing function is \\( f(x) = e^x \\), places higher importance on larger differences, which may highlight instances with more substantial temporal changes. The weighted average of these distances serves as the approximation of SW, calculated using the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.798, + 0.905, + 0.842 + ], + "angle": 0, + "content": "\\[\nS W = \\frac {1}{L} \\sum_ {l = 1} ^ {L} w \\cdot \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mathcal {M} _ {t} , \\theta_ {l} \\sharp \\mathcal {T}\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mathcal {M} _ {t} , \\theta_ {j} \\sharp \\mathcal {T}\\right)\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.845, + 0.855, + 0.861 + ], + "angle": 0, + "content": "where the weight \\( w \\) for each projection direction is:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.865, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nw = \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mu , \\theta_ {j} \\sharp \\nu\\right)\\right)}. \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12473" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "Eq.6 yields a weighted Sliced Wasserstein distance that reflects the relationship between adjacent frames along random projection directions, considering both the geometric positions of data points and their differences through importance weighting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.318 + ], + "angle": 0, + "content": "To establish instance correspondences, we compute the SW values between all pairs of instances across consecutive frames \\( t \\) and \\( t + 1 \\), denoted as \\( \\mathrm{SW}_p^p (E_t,E_{t + 1}) \\). These values quantify the difference between instances along the projection direction, with smaller SW values indicating stronger spatio-temporal consistency. By selecting the instance pairs with the smallest SW values, we identify those with the highest temporal alignment, ensuring robust and meaningful associations for instance tracking. Formally, the instance matching objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.33, + 0.483, + 0.347 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T i m e}} = \\arg \\min \\mathrm {S W} \\left(E _ {t}, E _ {t + 1}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.478 + ], + "angle": 0, + "content": "where the selected pairs correspond to instances with the strongest spatio-temporal consistency, facilitating accurate tracking and correlation across frames. Our method thus emphasizes directions and features that capture significant inter-frame changes, effectively prioritizing instances that reflect spatial and appearance consistency. The adaptive weighting in Eq.6 further enhances tracking by highlighting subtle variations, ensuring robust temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.273, + 0.503 + ], + "angle": 0, + "content": "4.3. Text-Visual Bridge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.689 + ], + "angle": 0, + "content": "After bridging the temporal consistency between the spatial and temporal domains through the spatio-temporal bridge, we are still confronted with the challenge of addressing the modal domain shift between the text and visual domains. Video content, with its dynamic and diverse nature, introduces additional complexity as textual descriptions and visual appearances can vary significantly. To ensure the accurate alignment between text and visual features, which is crucial in this context, we continue to leverage the Sliced Wasserstein metric. This metric has proven effective in our previous spatio-temporal bridging efforts, leading us to adopt it for modal alignment in our current method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Specifically, we first combine the temporal feature \\( D_{t} \\) with the query embeddings \\( E_{t} \\) through multiplication to obtain the visual embeddings \\( M_{t} \\). Subsequently, we apply a mask pooling operation to derive a class-agnostic mask \\( \\mathcal{M}_t\\in \\mathbb{R}^{N\\times D} \\) based on \\( M_{t} \\) and the frame features \\( F_{t} \\). Lastly, we obtain the class text embeddings \\( \\mathcal{H}\\in \\mathbb{R}^{L\\times D} \\) generated by the CLIP text encoder, where \\( L \\) represents the number of categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "It is worth noting that our cross-modal alignment method bears similarity to the Sliced Wasserstein metric procedure detailed in §4.2. We begin by defining the projection path of the cross-modal module as \\( Z_{i,j}^{c} = \\pmb{m}_{t,i} - \\pmb{h}_{t,j} \\). To obtain the projection direction, we normalize this path, resulting in \\( \\theta_{i,j} \\). Given the relatively small sizes of \\( \\mathcal{M}_t \\) and \\( \\mathcal{H} \\)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "this module departs from the random sampling approach and instead adopts a full-path full-projection strategy. Upon successfully obtaining the projection direction \\(\\theta\\), we proceed to calculate the one-dimensional Wasserstein distance. We then apply importance weighting and incorporate the Monte Carlo estimation method in our subsequent calculations. Following the logic of Eq. 6 and Eq. 7, we ultimately compute the weighted average as an approximation of the Sliced Wasserstein distance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Additionally, similar to Eq. 8, we introduce an additional loss function \\(\\mathcal{L}_{Modal}\\) to constrain the modality alignment process. This loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.288, + 0.905, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {M o d a l}} = \\arg \\min \\operatorname {S W} \\left(\\mathcal {M} _ {t}, \\mathcal {H}\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.312, + 0.66, + 0.327 + ], + "angle": 0, + "content": "4.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.409 + ], + "angle": 0, + "content": "To mine similarity relations in cross-text video pairs, we draw on the successful experience of cross-modal learning, as demonstrated in [8]. To this end, we introduce a cross-text similarity alignment loss based on contrastive learning. The loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.418, + 0.905, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V T}} = - \\frac {1}{N L} \\sum_ {i, j} \\log \\frac {\\exp \\left(d _ {i j} / \\tau\\right)}{\\sum_ {k} \\exp \\left(d _ {i k} / \\tau\\right) + \\sum_ {l} \\exp \\left(d _ {j l} / \\tau\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.905, + 0.499 + ], + "angle": 0, + "content": "where \\(d_{ij} = \\mathrm{sim}(\\pmb {m}_i,\\pmb {h}_j),\\mathrm{sim}(\\cdot ,\\cdot)\\) represents the similarity metric function, and \\(\\tau\\) is the temperature hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.905, + 0.529 + ], + "angle": 0, + "content": "The Total Loss \\(\\mathcal{L}_{\\mathrm{Total}}\\) is a weighted combination of multiple loss components, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.542, + 0.905, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T o t a l}} = \\alpha_ {T} \\mathcal {L} _ {\\text {T i m e}} + \\alpha_ {M} \\mathcal {L} _ {\\text {M o d a l}} + \\alpha_ {V} \\mathcal {L} _ {\\mathrm {V T}} + \\alpha_ {S} \\mathcal {L} _ {\\text {S e g}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{Seg}}\\) represents the binary cross-entropy and dice losses for mask prediction, following the approach in [26]. The coefficients \\(\\alpha_{T}, \\alpha_{M}, \\alpha_{V}\\), and \\(\\alpha_{S}\\) are the corresponding weighting factors for each loss component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Computational complexity. In our Sliced Wasserstein metric process, the core steps involve matrix operations and sorting procedures. Initially, we generate a set of initial projection vectors by randomly selecting samples and computing the differences among these samples. The complexity of this selection operation is \\( O(L) \\), while the complexity of the difference calculation is \\( O(L \\cdot D) \\), where \\( L \\) represents the number of projection vectors and \\( D \\) is the dimensionality of the samples. Subsequently, we proceed to the normalization and resampling phase, which also maintains a complexity of \\( O(L \\cdot D) \\). When calculating the one-dimensional Wasserstein distance, we need to perform matrix multiplication, a step with a complexity of \\( O(N \\cdot D \\cdot L) \\), where \\( N \\) is the number of samples. Additionally, we need to sort the projected data, with a sorting operation complexity of \\( O(N \\log N) \\). In summary, the overall complexity of our method can be approximated as \\( O(N \\cdot D \\cdot L + N \\log N) \\). In practical applications, as the number of samples \\( N \\) and the number of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12474" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 1. Statistics of the dataset. Cat. and Ins. represent the number of categories and instances respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.129, + 0.487, + 0.238 + ], + "angle": 0, + "content": "
DatasetVideosCat.Ins.TrainVal.Test
LV-VIS4,8281,196-3,083837908
YT-VIS 192,883404,8832,238302343
YT-VIS 213,859408,1712,985421453
OVIS901255,223607140154
BURST2,91448216,0005009931,421
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.326 + ], + "angle": 0, + "content": "projection vectors \\( L \\) increase, the computational cost of matrix multiplication will become dominant. Although sorting and other operations with linear complexity also exist, their contribution to the overall complexity is relatively small." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.338, + 0.223, + 0.357 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.363, + 0.378, + 0.379 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.489 + ], + "angle": 0, + "content": "We initially train our method on a combined set of common and frequent categories derived from the LVIS dataset [16]. Subsequently, we assess the performance of our method on both the validation and test sets of the LV-VIS dataset [26], as well as the validation set of several datasets, namely OVIS [21], YT-VIS 2019/2021 [27], and BURST [1]. The statistics of the evaluation dataset are detailed in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.566 + ], + "angle": 0, + "content": "LVIS [16] dataset comprises 1,203 object categories, which significantly exceeds the number of categories in COCO. Following ViLD[14], we select 866 frequent and common categories as the training categories and reserve the remaining 337 rare categories as novel categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Evaluation Metrics. Following [26], we present the mean Average Precision \\((AP)\\) for all categories in aggregate. Additionally, we report on \\(AP_{n}\\) for the novel category." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.308, + 0.639 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our implementation, the CLIP image encoder and text encoder are kept frozen, whereas the pixel decoder, transformer decoder, and mask prediction head undergo training. This strategy effectively harnesses the pre-trained capabilities of CLIP for mask classification and query matching, utilizing query embeddings, thereby obviating the need for category and identity annotations. The CLIP model employed in our framework adopts the ViTB/32 architecture for both the text and image encoders. Throughout the training process, the parameters of these encoders remain frozen. For optimization purposes, we utilize the AdamW optimizer, with the base learning rate set to \\(1 \\times 10^{-4}\\). \\(\\tau = 0.05\\), \\(\\alpha_{T}\\), \\(\\alpha_{M}\\), \\(\\alpha_{V}\\), and \\(\\alpha_{S}\\) are all set to 1 based on grid parameter tuning experience. Following a step-wise learning rate schedule, this rate is subsequently reduced by a factor of 10 at \\(90\\%\\) and \\(95\\%\\) of the total training steps. To assess the robustness and generality of our proposed method, we con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "duct experiments using two distinct backbones: ResNet-50 and SwinB. Our experimental setup is analogous to that of OV2Seg [26]. All experiments are conducted on a high-performance machine equipped with eight NVIDIA A800 GPUs, each possessing 80GB of memory. To facilitate a fair comparison, we also present the inference speed of our method when executed on a single NVIDIA A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.21, + 0.653, + 0.226 + ], + "angle": 0, + "content": "5.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.446 + ], + "angle": 0, + "content": "The experimental results presented in Table 2 provide a comprehensive comparison of our proposed SWBridge method with several state-of-the-art approaches across multiple video instance segmentation datasets. A significant finding reveals that SWBridge exhibited superior performance compared to existing methods across 19 out of 22 evaluation metrics in various settings. Furthermore, it attained suboptimal results in the remaining three metrics, thereby underscoring its efficacy in addressing the inherent complexity associated with Open-Vocabulary VIS. This result not only verifies the effectiveness of SWBridge in its design, but also shows that it can adapt well to the characteristics of different datasets, such as changes in video length, number of instances, scene complexity, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.719 + ], + "angle": 0, + "content": "LV-VIS. When trained on the diverse and extensive LVIS dataset, which includes a large number of frequent and common categories, SWBridge exhibits robust generalization capabilities. This is particularly evident in its performance on the LV-VIS dataset, where it achieves the highest \\(AP\\) for both the validation and test sets, including the novel categories \\((AP_{n})\\). The superior performance on LV-VIS underscores SWBridge's ability to effectively recognize and segment objects from both base and novel categories, even in the presence of a significant number of unseen classes during training. Although powerful backbone networks such as ConvNeXt-B can enhance performance, we find that the Swin-B backbone network, when combined with the SWBridge, yields even better results. This indicates that SWBridge has a synergistic effect with the backbone network, enabling more efficient capture of spatio-temporal features in videos, and the performance improvement exceeds that achieved by mere network upgrades." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "YT-VIS. SWBridge demonstrates superior performance on the YT-VIS2019 and YT-VIS2021 datasets, further attesting to its strong capability in handling video instance segmentation tasks with a moderate number of categories. Remarkably, this method achieves competitive performance levels without the need for fine-tuning on specific video datasets. This characteristic indicates that SWBridge can effectively leverage image-based training data and successfully generalize to video data, thereby significantly reducing the reliance on extensive video-specific annotations. In comparison with state-of-the-art methods, although OV-Former achieves excellent results by incorporating an ad" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12475" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods. All the methods are trained on image dataset LVIS and evaluated on the video instance segmentation datasets directly, i.e., they are not fine-tuned using the training set of each dataset." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.129, + 0.894, + 0.451 + ], + "angle": 0, + "content": "
MethodBackboneLV-VIS val.LV-VIS testYT-VIS2019YT-VIS2021OVISBURST
APAPnAPAPnAPAPnAPAPnAPAPAPAP
DetPro[10]-SORT[2]R506.43.55.82.1--------
Detric[30]-SORT[2]R506.53.45.72.114.63.512.73.16.71.92.5
DetPro[10]-OWTB[20]R507.94.27.02.9-------
Detric[30]-OWTB[20]R507.74.27.02.817.94.516.75.89.02.71.8
Detric[30]-XMem[7]R508.85.47.73.6-------
OV2Seg[26]R5014.211.911.48.927.211.123.67.311.23.72.4
OVFormer[11]R50+ViT-B----34.816.529.815.715.1--
CLIP-VIS [32]R5019.524.214.615.932.223.830.117.914.15.27.7
SWBridge (Our)R5020.724.915.416.233.424.030.918.415.25.58.2
Detric[30]-SORT[2]SwinB12.86.69.44.723.87.921.69.811.72.51.0
Detric[30]-OWTB[20]SwinB14.58.511.86.130.09.727.111.413.63.92.4
Detric[30]-XMem[7]SwinB16.310.613.17.7-------
OV2Seg[26]SwinB21.116.316.411.537.621.333.918.217.54.93.0
OVFormer[11]SwinB+ViT-B----44.321.537.618.321.3--
CLIP-VIS [32]ConvNeXt-B32.240.225.330.642.127.537.922.018.58.312.7
SWBridge (Our)SwinB33.040.526.031.043.028.038.522.519.78.312.8
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.482, + 0.566 + ], + "angle": 0, + "content": "ditional CLIP image encoder, SWBridge still outperforms CLIP-VIS in terms of the \\(AP\\) metric under similar model architectures. More notably, in the \\(AP_{n}\\) metric, which pertains to the segmentation of new class instances, our bridge strategy comprehensively surpasses the performance gains brought about by the CLIP encoder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.483, + 0.702 + ], + "angle": 0, + "content": "OVIS. Since the categories of the OVIS dataset highly overlap with the LVIS dataset used for training, we only reports the AP (average precision) performance metric. SWBridge even outperformed OVFormer with the extra CLIP on the R50 backbone and was second only to OVFoemer on the SwinB backbone. SWBridge's ability to maintain high accuracy under these conditions demonstrates its robustness in complex visual environments, where the visibility of objects changes significantly over time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.703, + 0.483, + 0.809 + ], + "angle": 0, + "content": "BURST. The BURST dataset, with its diverse and uncommon categories, poses a unique challenge due to the wide range of object types. SWBridge's competitive performance on this dataset, especially in terms of \\(AP_{n}\\), suggests that it can effectively manage the intricacies of uncommon object categories, enhancing its applicability to real-world scenarios where encountering unseen objects is common." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.245, + 0.835 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our ablation experiments continue to use the LVIS dataset for training with ResNet50 as the Backbone and are evaluated on the validation sets of the LV-VIS, OVIS, and BURST datasets, aiming to measure the performance of the" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.474, + 0.894, + 0.489 + ], + "angle": 0, + "content": "Table 3. Comparison of ablation results of bridge strategy." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.499, + 0.924, + 0.609 + ], + "angle": 0, + "content": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Baseline8.510.25.32.62.3
+ Spatio-Temporal Bridge15.018.010.23.15.4
+ Text-Visual Bridge19.923.314.25.37.9
+ LVT in Eq. 1020.724.915.25.58.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.763 + ], + "angle": 0, + "content": "proposed SWBridge more comprehensively and rigorously. Effectiveness of Bridge Strategy. We first established a baseline variant that removed the Spatio-Temporal Bridge bridge and Text-Visual Bridge modules, while discarding the additional contrast loss \\(\\mathcal{L}_{VT}\\) and retaining only the basic \\(\\mathcal{L}_{Seg}\\) loss function. In order to achieve instance association, we adopted the memory module proposed in OV2Seg to replace the original spatiotemporal bridge function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The experimental results presented in Table 3 confirm the significant effectiveness of each component in the SWBridge method. Specifically, upon integrating the spatio-temporal bridge module into the baseline variant, we observed notable improvements in both \\(AP\\) and \\(AP_{n}\\) metrics across all datasets. This outcome indicates that the spatio-temporal bridge module efficiently captures spatiotemporal correlations between video frames, thereby aiding the model in more accurately understanding and tracking" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "12476" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.09, + 0.477, + 0.104 + ], + "angle": 0, + "content": "Table 4. Impact of ablation results of Importance Weighting." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.115, + 0.468, + 0.223 + ], + "angle": 0, + "content": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
w/o Weight15.317.012.34.55.3
w/o Weight in §4.118.522.313.74.96.8
w/o Weight in §4.319.623.514.55.27.5
SWBridge(Our)20.724.915.25.58.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.238, + 0.477, + 0.252 + ], + "angle": 0, + "content": "Table 5. Impact of ablation results of Importance Weighting." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.264, + 0.468, + 0.371 + ], + "angle": 0, + "content": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Huberized18.822.413.84.76.4
Cosine18.622.213.94.86.5
Wasserstein19.523.314.45.17.0
Sliced Wasserstein20.724.915.25.58.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.549 + ], + "angle": 0, + "content": "the dynamic changes of instances within videos. Furthermore, the model's performance was significantly enhanced once again after adding the text-visual bridge module on top of the spatio-temporal bridge module. By introducing textual information, the text-visual bridge provides the model with rich semantic cues, greatly enhancing its accuracy in recognizing and understanding instances. Additionally, \\(\\mathcal{L}_{VT}\\) synergizes with other bridge strategy modules, collectively driving a comprehensive improvement in the model's performance and generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Impact of SW Weighting Mechanism. Table 4 shows that without a weighting mechanism in the spatio-temporal or text-visual bridges, the model struggles to utilize key information from instances with marked feature differences, limiting performance. However, incorporating the weighting mechanism into both bridges notably improves model performance. This indicates the mechanism helps the model focus on crucial instances with significant feature differences, assigning them greater weights during distance calculation. This adjustment enhances the model's accuracy and efficiency in processing complex data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Comparison of Distance Measurement. In Tab. 5, we compare distance metrics for SWBridge. Huberized distance is sensitive to large differences, Cosine similarity ignores feature magnitudes, and Wasserstein distance misses local details. Our SW distance comprehensively considers high-dimensional feature differences with importance weighting, making it robust for complex domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Sensitivity analysis of \\( L \\). In the bridge strategy, we ablated the number of sampling iterations to explore its relationship with model performance (Fig.2). Results show a nonlinear relationship: increasing iterations significantly improves \\( AP \\) and \\( AP_{n} \\) across all datasets. However, excessive" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.898, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.284, + 0.905, + 0.312 + ], + "angle": 0, + "content": "Figure 2. Ablation experiment on the number of random projection sampling \\( L \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.371 + ], + "angle": 0, + "content": "iterations introduce noise and computational burden, potentially causing overfitting or reducing training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.375, + 0.634, + 0.39 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.626 + ], + "angle": 0, + "content": "In this paper, we propose an innovative Sliced Wasserstein Bridging Learning Framework (SWbridge), aimed at tackling the challenging task of Open-Vocabulary Video Instance Segmentation. Inspired by metric learning, our framework utilizes the Sliced Wasserstein distance as a core tool to effectively bridge the spatial-temporal and text-visual domains involved in the task. Through domain alignment, we successfully map features from different domains into a unified metric space, preserving temporal consistency and learning intrinsic consistent features across modalities. This significantly improves the fusion of text and visual information, addressing one of the core challenges in Open-Vocabulary VIS. Looking ahead, we plan to explore better optimization methods for the Sliced Wasserstein distance to further enhance the performance of our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in part by the China Postdoctoral Science Foundation (No. 2024M761805), the Postdoctoral Fellowship Program of CPSF (No. GZB20250390), the Qingdao Postdoctoral Innovation Project (No. QDBSH20240201015), the Major Science and Technology Innovation Projects in Shandong Province (No. 2024CXGC010108)." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.748, + 0.611, + 0.763 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.773, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[1] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.83, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[3] Nicolas Bonneel, Julien Rabin, Gabriel Peyre, and Hanspeter Pfister. Sliced and radon wasserstein barycenters of mea" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.773, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12477" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "sures. Journal of Mathematical Imaging and Vision, 51:22-45, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[4] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[8] Zesen Cheng, Kehan Li, Hao Li, Peng Jin, Chang Liu, Xiawu Zheng, Rongrong Ji, and Jie Chen. Instance brownian bridge as texts for open-vocabulary video instance segmentation. arXiv preprint arXiv:2401.09732, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[9] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.427, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.482, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[11] Hao Fang, Peng Wu, Yawei Li, Xinxin Zhang, and Xiankai Lu. Unified embedding alignment for open-vocabulary video instance segmentation. In ECCV, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.525, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[12] Damien Garreau, Rémi Lajugie, Sylvain Arlot, and Francis Bach. Metric learning for temporal sequence alignment. NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.567, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.608, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[15] Pinxue Guo, Tony Huang, Peiyang He, Xuefeng Liu, Tianjun Xiao, Zhaoyu Chen, and Wenqiang Zhang. Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[17] Fusheng Hao, Fengxiang He, Jun Cheng, Lei Wang, Jianzhong Cao, and Dacheng Tao. Collect and select: Semantic alignment metric learning for few-shot learning. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[18] Dat Huynh, Jason Kuen, Zhe Lin, Jieuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[19] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In CVPR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[20] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[21] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, 130(8): 2022-2039, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[22] Zheyun Qin, Xiankai Lu, Dongfang Liu, Xiushan Nie, Yi long Yin, Jianbing Shen, and Alexander C Loui. Reformulating graph kernels for self-supervised space-time correspondence learning. IEEE Transactions on Image Processing, 32: 6543-6557, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[23] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1192-1208, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[25] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[26] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[27] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[28] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[29] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[30] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[31] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[32] Wenqi Zhu, Jiale Cao, Jin Xie, Shuangming Yang, and Yanwei Pang. Clip-vis: Adapting clip for open-vocabulary video instance segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12478" + } + ] +] \ No newline at end of file diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_origin.pdf b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4398342ab1087d6757e9ed8b7d931d0731bd9daa --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/320731ee-724e-4243-b22e-a9cb1f48c443_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71a333b89b6f6f8bff98d79e91b4e1fc51246678df7a455700f1d21d462b5435 +size 2304292 diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/full.md b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7675f2a1785b1f94b5d3282166e04a536c427ff3 --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/full.md @@ -0,0 +1,297 @@ +# Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation + +Zheyun Qin $^{1}$ , Deng Yu $^{2}$ , Chuanchen Luo $^{2}$ , Zhumin Chen $^{2*}$ + +$^{1}$ School of Computer Science and Technology, Shandong University $^{2}$ School of Artificial Intelligence, Shandong University + +# Abstract + +In recent years, researchers have explored the task of open-vocabulary video instance segmentation, which aims to identify, track, and segment any instance within an open set of categories. The core challenge of Open-Vocabulary VIS lies in solving the cross-domain alignment problem, including spatial-temporal and text-visual domain alignments. Existing methods have made progress but still face shortcomings in addressing these alignments, especially due to data heterogeneity. Inspired by metric learning, we propose an innovative Sliced Wasserstein Bridging Learning Framework. This framework utilizes the Sliced Wasserstein distance as the core tool for metric learning, effectively bridging the four domains involved in the task. Our innovations are threefold: (1) Domain Alignment: By mapping features from different domains into a unified metric space, our method maintains temporal consistency and learns intrinsic consistent features between modalities, improving the fusion of text and visual information. (2) Weighting Mechanism: We introduce an importance weighting mechanism to enhance the discriminative ability of our method when dealing with imbalanced or significantly different data. (3) High Efficiency: Our method inherits the computational efficiency of the Sliced Wasserstein distance, allowing for online processing of large-scale video data while maintaining segmentation accuracy. Through extensive experimental evaluations, we have validated the robustness of our concept and the effectiveness of our framework. + +# 1. Introduction + +In recent years, video instance segmentation methods [23, 27, 28] have demonstrated significant improvements and breakthroughs in various application fields. However, most existing methods rely on the Closed-Set Assumption [4, 22], where the learning and recognition scope of the model are strictly limited to known categories or labels during training and testing. This assumption ignores the possibility of unknown categories in the real world, limiting the model's ability to effectively detect unknown categories. + +To overcome this limitation, researchers have begun to explore the task of open-vocabulary video instance segmentation [15, 26] (Open-Vocabulary VIS), which aims to identify, track, and segment any instance within an open set of categories. These Open-Vocabulary VIS methods can not only handle traditional predefined categories but also flexibly adapt to new user-specified categories, achieving the advanced function of on-demand segmentation. + +The core of Open-Vocabulary VIS methods lies in solving the cross-domain alignment problem, including spatial-temporal domain alignment and text-visual domain alignment. Existing methods have different focuses. In terms of spatial-temporal domain alignment, current methods extract class-agnostic instance features based on the foundational model Mask2Former-VIS. Furthermore, OV2Seg [26], OpenVIS [15], and CLIP-VIS [32] employ memory-based approaches to perform cross-frame instance matching. In terms of text-visual domain alignment, existing methods mainly rely on the zero-shot performance exhibited by visual-language models, i.e., CLIP [24] pretrained on largescale image-text pairs. Specifically, BriVIS [8] and OV-Former [11] introduce additional CLIP image encoders on top of visual feature extractors. The difference lies in that BriVIS integrates them based on contrastive learning, while OV-Former maps text and visual features into the same space vis the attention mechanism. + +Despite the significant progress made by existing Open-Vocabulary VIS methods, there are still many shortcomings. The cross-domain alignment problem remains inadequately addressed, and the inherent data heterogeneity results in significant differences in feature distributions across different domains. In light of this, we cannot help but trace back to the development history of traditional machine learning methods, seeking a tool that can simultaneously solve spatial-temporal domain alignment and text-visual domain alignment. We find that the metric learning method in the field of machine learning provides an effective approach to addressing the alignment problem [12, 17]. By learning a suitable metric space, this paradigm makes similar samples closer while dissimilar samples relatively farther apart, providing a powerful tool for solving the alignment problem. + +Guided by the idea of metric learning, we propose an + +innovative Sliced Wasserstein Bridging Learning Framework, i.e., SWbridge. This framework utilizes the Sliced Wasserstein distance [3] as the core tool for metric learning. This framework cleverly utilizes the Sliced Wasserstein distance as the core tool for metric learning. Specifically, we minimize the Sliced Wasserstein distance between samples based on random paths through sampling, achieving feature embedding alignment. We bridge four domains based on a strategy to address the challenges of cross-domain alignment (spatial-temporal domain and text-visual domain). + +Our innovations are mainly concentrated in three aspects: (1) Domain Alignment. The Sliced Wasserstein distance has the advantage of accurately capturing differences in data distributions in high-dimensional spaces. Our method uses the Sliced Wasserstein distance as a bridge to map features from different domains into a unified metric space. This not only maintains temporal consistency by comparing instance embeddings between adjacent frames, achieving temporal alignment and effectively mitigating challenges such as progressive occlusion, but also learns intrinsic consistent features between modalities, narrowing the semantic gap and improving the fusion of text and visual information through modality alignment. (2) Weighting Mechanism: We introduce an importance weighting mechanism to further enhance its discriminative ability when dealing with imbalanced data or data with significant feature differences. In open-vocabulary video instance segmentation tasks, there may be significant feature differences between different instances, and some instances may be more critical or informative. Through the importance weighting mechanism, our method can dynamically adjust the contribution of different instances in the metric space, allowing key instances to have greater weight in distance calculations. (3) High Efficiency: Our SWbridge inherits the computational efficiency advantage of the Sliced Wasserstein distance, making it highly efficient when processing large-scale video data. Without introducing additional network parameters, our method can handle video in an online manner while maintaining segmentation accuracy. + +Combining the aforementioned perspectives and innovations, we propose an open-vocabulary video instance segmentation method capable of bridging multiple domains. This method has achieved remarkable results on multiple datasets related to open-vocabulary and video tasks. + +# 2. Related Works + +# 2.1. Open-Vocabulary Detection and Segmentation + +Open Vocabulary Detection (OVD) [9, 25] and Open Vocabulary Segmentation (OVS) [13, 18] are cutting-edge tasks in the field of computer vision, enabling models to be trained on images containing unannotated novel objects, thereby breaking the closed-set constraint. This break- + +through is primarily attributed to the application of weak supervision signals, namely the utilization of image-text pairs (such as image-caption pairs and image-level labels) or large pre-trained Visual-Language Models (VLMs), such as CLIP [24]. Supported by weak supervision signals, methods for OVD and OVS can be categorized into four main types [31]. The first type is region-aware training, which learns object feature representations by exploring the intrinsic links between image regions and text descriptions without relying on the VLMs' image encoder (VLMsIE) or direct object annotations. The second type is the pseudo-labeling method, which enhances model generalization by generating pseudo-labels from preliminary predictions on unannotated images using both image-text pairs and the VLMs-IE. The other two types are knowledge distillation and transfer learning. Knowledge distillation transfers knowledge from pre-trained VLMs to new models or tasks via a distillation mechanism, while transfer learning applies knowledge learned by pre-trained models on specific tasks to new tasks or domains. Both rely on the VLMsIE and seldom involve direct training on image-text pairs. + +Furthermore, zero-shot learning [29] addresses the closed-set constraint in scene-aware tasks and differs from open vocabulary methods. It prohibits access to weak supervision signals during training, but can be converted into Open-Vocabulary tasks upon gaining access, enhancing the model's generalization and recognition of novel objects. + +# 2.2. Open-Vocabulary Video Instance Segmentation + +Open-vocabulary video instance segmentation, as an emerging vision task, aims to simultaneously classify, track, and segment arbitrary objects within open categories in videos, attracting considerable attention from researchers in recent years. OV2Seg [26] has taken the lead in this field by not only constructing the Large Vocabulary Video Instance Segmentation (LV-VIS) dataset but also proposing the first end-to-end benchmark for open-vocabulary video instance segmentation. OV2Seg method leverages the Mask2Former [6] framework to extract class-agnostic masks and query embeddings, utilizes the CLIP text encoder for precise mask classification, and achieves cross-frame instance tracking through a long-term matching strategy. Subsequent research has mostly followed this basic approach and expanded upon it. OpenVIS [15] introduces a two-stage framework, OVIS, which adopts a neighboring matching strategy for instance tracking, effectively simplifying the processing pipeline. BriVIS [8] models instances as Brownian bridges and closely aligns bridge-level instance representations with category texts through contrastive learning, also employing a neighboring matching strategy to enhance the accuracy of instance recognition. OVFormer [11] introduces a novel unified embedding alignment module that effectively addresses the domain gap between instance + +queries and text embeddings, demonstrating good practicality with its semi-online processing approach. However, the aforementioned methods primarily focus on bridging the domain discrepancy between text and vision, with insufficient attention to spatiotemporal consistency among cross-frame instances. CLIP-VIS [32] improves upon this by introducing a time-topK enhanced matching strategy, which strengthens temporal modeling capabilities between frames and improves matching accuracy. + +We believe that the core of the open-vocabulary video instance segmentation task lies in simultaneously addressing the alignment of the spatial-temporal domain and the text-visual domain. To this end, based on the concept of metric learning, we employ the Sliced Wasserstein distance to bridge various domains. This approach not only effectively narrows the domain gap between modalities but also ensures spatio-temporal consistency of instances across frames, providing a novel solution for the open-vocabulary video instance segmentation task. + +# 2.3. Vision-Language Models + +The core of Visual Language Models (VLMs) [5] is their deep training on large-scale image-text paired data, enabling the fusion and understanding of visual and textual information, and endowing them with powerful zero-shot object recognition capabilities [19]. CLIP [24], a prominent VLM, uses an image encoder to extract visual features and a text encoder to generate text embeddings, jointly constructing a cross-modal representation space. However, applying VLMs to open-vocabulary video instance segmentation is challenging. VLMs, primarily trained on images, struggle with understanding dynamic video scenes rich in spatiotemporal information, and lack the capability to maintain object consistency across frames. Additionally, the domain shift between text and vision poses a significant challenge, potentially leading to biases in mapping textual information to the visual space. To overcome these limitations, we propose a progressive approach that addresses the spatiotemporal consistency issue and the domain shift problem, enabling VLMs to better adapt to video instance segmentation tasks. + +# 3. Preliminaries + +# 3.1. Open-Vocabulary VIS Formulation + +Given a test video $\mathcal{D}_{test}$ with $T$ frames, the objective in Open-Vocabulary VIS is to accurately predict all $N$ instances belonging to the categories in $\mathcal{C} = \mathcal{C}_{base} \cup \mathcal{C}_{novel}$ by the trained model $f_{\theta}$ . $\mathcal{C}_{base}$ is the set of base (training) categories, and $\mathcal{C}_{novel}$ is the set of novel categories that are not seen during training but may appear in the test videos. The prediction can be formulated as: + +$$ +\left\{\left\{m _ {1}, m _ {2}, \dots , m _ {T} \right\}, c \right\} _ {n} ^ {N} = f _ {\theta} \left(\mathcal {D} _ {t e s t}\right), \tag {1} +$$ + +where $\{m_t\}_{t = 1}^T$ is the segmentation masks, and $c\in \mathcal{C}$ + +# 3.2. Sliced Wasserstein Distance + +One-dimensional Wasserstein Distance. For one-dimensional probability measures $\mu$ and $\nu$ in $\mathcal{P}_p(\mathbb{R})$ , the $p$ -Wasserstein distance is defined as: + +$$ +\mathrm {W} _ {p} ^ {p} (\mu , \nu) = \int_ {0} ^ {1} | F _ {\mu} ^ {- 1} (z) - F _ {\nu} ^ {- 1} (z) | ^ {p} d z, \tag {2} +$$ + +where $F_{\mu}$ and $F_{\nu}$ are the cumulative distribution functions (CDFs) of $\mu$ and $\nu$ . This formulation provides a closed form for computing the Wasserstein distance in one-dimensional spaces, making it well-suited for projected measures. + +Sliced Wasserstein Distance. To generalize the Wasserstein distance to higher-dimensional measures, the Sliced Wasserstein Distance (SWD) projects the measures $\mu$ and $\nu$ in $\mathcal{P}_p(\mathbb{R}^d)$ onto one-dimensional subspaces, and then averages the one-dimensional Wasserstein distances from these projections. For $\mu, \nu \in \mathcal{P}_p(\mathbb{R}^d)$ , the SWD is defined as: + +$$ +\mathrm {S W} _ {p} ^ {p} (\mu , \nu) = \mathbb {E} _ {\theta \sim \mathcal {U} \left(\mathbb {S} ^ {d - 1}\right)} \left[ \mathrm {W} _ {p} ^ {p} \left(\theta \sharp \mu , \theta \sharp \nu\right) \right], \tag {3} +$$ + +where $\theta \sharp \mu$ and $\theta \sharp \nu$ represent the push-forward measures of $\mu$ and $\nu$ along direction $\theta \in \mathbb{S}^{d - 1}$ , the unit sphere in $\mathbb{R}^d$ . This projection $f(x) = \theta^{\top}x$ maps points from $\mathbb{R}^d$ to $\mathbb{R}$ , enabling the computation of Wasserstein distances in a one-dimensional space. + +Since the expectation in Eq. 3 is computationally expensive, SWD is typically approximated by averaging over $L$ independent directions, $\theta_{1},\ldots ,\theta_{L}$ , sampled from $\mathcal{U}(\mathbb{S}^{d - 1})$ : + +$$ +\widehat {\mathbf {S W}} _ {p} ^ {p} (\mu , \nu ; L) = \frac {1}{L} \sum_ {l = 1} ^ {L} \mathbf {W} _ {p} ^ {p} \left(\theta_ {l} \sharp \mu , \theta_ {l} \sharp \nu\right), \tag {4} +$$ + +where each $\theta_{l}\sharp \mu$ and $\theta_{l}\sharp \nu$ are projected representations of $\mu$ and $\nu$ along the direction $\theta_{l}$ . The number of projections $L$ controls the accuracy of the Monte Carlo approximation. + +# 4. Method + +As illustrated in Fig. 1, we propose the SWbridge framework. Firstly, in §4.1, we integrate a category-agnostic feature extractor. Then, in §4.2, we introduce a spatio-temporal bridging module that robustly establishes instance associations between adjacent frames, ensuring the consistency and coherence of temporal information. Finally, in §4.3, we design a modal bridging module that performs fine-grained semantic mapping across the text and visual domains to enhance cross-domain understanding and effectively address the challenges posed by domain shifts. + +# 4.1. Class-agnostic Feature Extractor + +Given a video $\mathcal{D}$ consisting of $T$ frames, $\{V_t \in \mathbb{R}^{3 \times H \times W}\}_{t=1}^T$ , where $H$ and $W$ denote the height and width of each frame, our model follows the conventional paradigm [11, 26, 32] by adopting the Mask2Former + +![](images/8a47e114bbf7d6746aa848776aae236520e165cbbf227cd6371392875237e0c7.jpg) +Figure 1. Overview of the proposed SWBridge framework (left) and the Sliced Wasserstein distance calculation process (right). + +![](images/e4d44f98b684a3c9e393ec78b641d1b754f9f2e02dacbedee2631a771fb76644.jpg) + +![](images/b29e81eceee9d52640a0f0f236c5595caef8d913c35c350151ff7d0954387c84.jpg) + +architecture as a class-agnostic feature extractor. The Mask2Former comprises three core components: an image encoder $f_{\mathrm{CLIP}}$ , a pixel decoder $f_{\mathrm{decode}}$ , and a Transformer decoder $f_{\mathrm{Transformer}}$ . Specifically, the frozen CLIP model $f_{\mathrm{CLIP}}$ processes each frame $V_{t}$ independently to extract high-level visual features $F_{t}$ . These features are then input to the pixel decoder $f_{\mathrm{decode}}$ , which upsamples them to produce a multiscale feature map $D_{t}$ preserving spatial details. Furthermore, a set of learnable queries $Q$ is initialized and fed into the Transformer decoder $f_{\mathrm{Transformer}}$ along with the high-resolution feature map $D_{t}$ . The Transformer decoder processes these inputs to generate the query embeddings $E_{t}$ . + +# 4.2. Spatio-Temporal Bridge + +After extracting category-agnostic features, the next challenge is to construct a spatio-temporal bridge to establish the correspondence between instances in consecutive frames $t$ and $t + 1$ . Given the dynamic nature of video content, objects may encounter occlusion, movement, and appearance changes. Therefore, it is particularly important to ensure the accuracy of instance tracking over time. To achieve this, we employ the Sliced Wasserstein metric. + +Initially, for the query embeddings of frames $t$ and $t + 1$ , we perform random sampling to construct random paths $Z_{i,j} = \mathbf{q}_{t,i} - \mathbf{q}_{t + 1,j}$ , where $\mathbf{q}_{t,i} \in E_t$ and $\mathbf{q}_{t + 1,i} \in E_{t + 1}$ are outputs of the Transformer decoder $f_{\mathrm{Transformer}}$ . These paths are then normalized to obtain the unit direction $\theta_{i,j} = \frac{Z_{i,j}}{\|Z_{i,j}\|_2}$ , which serves as the projection direction. + +Subsequently, we sample a new projection direction $\hat{\theta}_{i,j}$ from $\sigma_{\kappa}(\theta_{i,j};\frac{Z_{i,j}}{\|Z_{i,j}\|_2})$ . This introduces randomness, potentially yielding different projection directions each time, which aids in exploring multiple relative positional relation + +ships between feature points. This, in turn, may enhance the model's robustness and generalization capability. + +With the projection direction $\hat{\theta}_{i,j}$ in hand, we proceed to calculate the one-dimensional Wasserstein distance. This involves projecting $\mathcal{M}_t$ and $\mathcal{T}$ along the $\theta$ direction to obtain one-dimensional distributions $\theta \sharp \mu$ and $\theta \sharp \nu$ . These distributions are then sorted, and the inverse functions of their cumulative distribution functions (CDFs) are computed. Finally, the one-dimensional Wasserstein distance $W_{p}^{p}(\theta \sharp \mu ,\theta \sharp \nu)$ is calculated using the formula: + +$$ +W _ {p} ^ {p} (F, G) = \int_ {0} ^ {1} | F ^ {- 1} (t) - G ^ {- 1} (t) | ^ {p} d t, \tag {5} +$$ + +where $F$ and $G$ are the CDFs of $\theta \sharp \mu$ and $\theta \sharp \nu$ , respectively, and $p$ is typically set to 2. + +The next step involves importance weighting and Monte Carlo estimation. We repeat the aforementioned process $L$ times to obtain $L$ projection directions $\{\theta_1, \dots, \theta_L\}$ and the corresponding set of one-dimensional Wasserstein distances. Each distance is then weighted with $w$ , where $w$ is determined by an increasing function $f(\cdot)$ . Increasing function is $f(x) = e^x$ , places higher importance on larger differences, which may highlight instances with more substantial temporal changes. The weighted average of these distances serves as the approximation of SW, calculated using the formula: + +$$ +S W = \frac {1}{L} \sum_ {l = 1} ^ {L} w \cdot \frac {f \left(W _ {p} ^ {p} \left(\theta_ {l} \sharp \mathcal {M} _ {t} , \theta_ {l} \sharp \mathcal {T}\right)\right)}{\sum_ {j = 1} ^ {L} f \left(W _ {p} ^ {p} \left(\theta_ {j} \sharp \mathcal {M} _ {t} , \theta_ {j} \sharp \mathcal {T}\right)\right)}, \tag {6} +$$ + +where the weight $w$ for each projection direction is: + +$$ +w = \frac {f \left(W _ {p} ^ {p} \left(\theta_ {l} \sharp \mu , \theta_ {l} \sharp \nu\right)\right)}{\sum_ {j = 1} ^ {L} f \left(W _ {p} ^ {p} \left(\theta_ {j} \sharp \mu , \theta_ {j} \sharp \nu\right)\right)}. \tag {7} +$$ + +Eq.6 yields a weighted Sliced Wasserstein distance that reflects the relationship between adjacent frames along random projection directions, considering both the geometric positions of data points and their differences through importance weighting. + +To establish instance correspondences, we compute the SW values between all pairs of instances across consecutive frames $t$ and $t + 1$ , denoted as $\mathrm{SW}_p^p (E_t,E_{t + 1})$ . These values quantify the difference between instances along the projection direction, with smaller SW values indicating stronger spatio-temporal consistency. By selecting the instance pairs with the smallest SW values, we identify those with the highest temporal alignment, ensuring robust and meaningful associations for instance tracking. Formally, the instance matching objective is: + +$$ +\mathcal {L} _ {\text {T i m e}} = \arg \min \mathrm {S W} \left(E _ {t}, E _ {t + 1}\right), \tag {8} +$$ + +where the selected pairs correspond to instances with the strongest spatio-temporal consistency, facilitating accurate tracking and correlation across frames. Our method thus emphasizes directions and features that capture significant inter-frame changes, effectively prioritizing instances that reflect spatial and appearance consistency. The adaptive weighting in Eq.6 further enhances tracking by highlighting subtle variations, ensuring robust temporal consistency. + +# 4.3. Text-Visual Bridge + +After bridging the temporal consistency between the spatial and temporal domains through the spatio-temporal bridge, we are still confronted with the challenge of addressing the modal domain shift between the text and visual domains. Video content, with its dynamic and diverse nature, introduces additional complexity as textual descriptions and visual appearances can vary significantly. To ensure the accurate alignment between text and visual features, which is crucial in this context, we continue to leverage the Sliced Wasserstein metric. This metric has proven effective in our previous spatio-temporal bridging efforts, leading us to adopt it for modal alignment in our current method. + +Specifically, we first combine the temporal feature $D_{t}$ with the query embeddings $E_{t}$ through multiplication to obtain the visual embeddings $M_{t}$ . Subsequently, we apply a mask pooling operation to derive a class-agnostic mask $\mathcal{M}_t\in \mathbb{R}^{N\times D}$ based on $M_{t}$ and the frame features $F_{t}$ . Lastly, we obtain the class text embeddings $\mathcal{H}\in \mathbb{R}^{L\times D}$ generated by the CLIP text encoder, where $L$ represents the number of categories. + +It is worth noting that our cross-modal alignment method bears similarity to the Sliced Wasserstein metric procedure detailed in §4.2. We begin by defining the projection path of the cross-modal module as $Z_{i,j}^{c} = \pmb{m}_{t,i} - \pmb{h}_{t,j}$ . To obtain the projection direction, we normalize this path, resulting in $\theta_{i,j}$ . Given the relatively small sizes of $\mathcal{M}_t$ and $\mathcal{H}$ , + +this module departs from the random sampling approach and instead adopts a full-path full-projection strategy. Upon successfully obtaining the projection direction $\theta$ , we proceed to calculate the one-dimensional Wasserstein distance. We then apply importance weighting and incorporate the Monte Carlo estimation method in our subsequent calculations. Following the logic of Eq. 6 and Eq. 7, we ultimately compute the weighted average as an approximation of the Sliced Wasserstein distance. + +Additionally, similar to Eq. 8, we introduce an additional loss function $\mathcal{L}_{Modal}$ to constrain the modality alignment process. This loss function is defined as: + +$$ +\mathcal {L} _ {\text {M o d a l}} = \arg \min \operatorname {S W} \left(\mathcal {M} _ {t}, \mathcal {H}\right). \tag {9} +$$ + +# 4.4. Loss Function + +To mine similarity relations in cross-text video pairs, we draw on the successful experience of cross-modal learning, as demonstrated in [8]. To this end, we introduce a cross-text similarity alignment loss based on contrastive learning. The loss function is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {V T}} = - \frac {1}{N L} \sum_ {i, j} \log \frac {\exp \left(d _ {i j} / \tau\right)}{\sum_ {k} \exp \left(d _ {i k} / \tau\right) + \sum_ {l} \exp \left(d _ {j l} / \tau\right)}, \tag {10} +$$ + +where $d_{ij} = \mathrm{sim}(\pmb {m}_i,\pmb {h}_j),\mathrm{sim}(\cdot ,\cdot)$ represents the similarity metric function, and $\tau$ is the temperature hyperparameter. + +The Total Loss $\mathcal{L}_{\mathrm{Total}}$ is a weighted combination of multiple loss components, which can be written as: + +$$ +\mathcal {L} _ {\text {T o t a l}} = \alpha_ {T} \mathcal {L} _ {\text {T i m e}} + \alpha_ {M} \mathcal {L} _ {\text {M o d a l}} + \alpha_ {V} \mathcal {L} _ {\mathrm {V T}} + \alpha_ {S} \mathcal {L} _ {\text {S e g}}, \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{Seg}}$ represents the binary cross-entropy and dice losses for mask prediction, following the approach in [26]. The coefficients $\alpha_{T}, \alpha_{M}, \alpha_{V}$ , and $\alpha_{S}$ are the corresponding weighting factors for each loss component. + +Computational complexity. In our Sliced Wasserstein metric process, the core steps involve matrix operations and sorting procedures. Initially, we generate a set of initial projection vectors by randomly selecting samples and computing the differences among these samples. The complexity of this selection operation is $O(L)$ , while the complexity of the difference calculation is $O(L \cdot D)$ , where $L$ represents the number of projection vectors and $D$ is the dimensionality of the samples. Subsequently, we proceed to the normalization and resampling phase, which also maintains a complexity of $O(L \cdot D)$ . When calculating the one-dimensional Wasserstein distance, we need to perform matrix multiplication, a step with a complexity of $O(N \cdot D \cdot L)$ , where $N$ is the number of samples. Additionally, we need to sort the projected data, with a sorting operation complexity of $O(N \log N)$ . In summary, the overall complexity of our method can be approximated as $O(N \cdot D \cdot L + N \log N)$ . In practical applications, as the number of samples $N$ and the number of + +Table 1. Statistics of the dataset. Cat. and Ins. represent the number of categories and instances respectively. + +
DatasetVideosCat.Ins.TrainVal.Test
LV-VIS4,8281,196-3,083837908
YT-VIS 192,883404,8832,238302343
YT-VIS 213,859408,1712,985421453
OVIS901255,223607140154
BURST2,91448216,0005009931,421
+ +projection vectors $L$ increase, the computational cost of matrix multiplication will become dominant. Although sorting and other operations with linear complexity also exist, their contribution to the overall complexity is relatively small. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Metrics + +We initially train our method on a combined set of common and frequent categories derived from the LVIS dataset [16]. Subsequently, we assess the performance of our method on both the validation and test sets of the LV-VIS dataset [26], as well as the validation set of several datasets, namely OVIS [21], YT-VIS 2019/2021 [27], and BURST [1]. The statistics of the evaluation dataset are detailed in Table 1. + +LVIS [16] dataset comprises 1,203 object categories, which significantly exceeds the number of categories in COCO. Following ViLD[14], we select 866 frequent and common categories as the training categories and reserve the remaining 337 rare categories as novel categories. + +Evaluation Metrics. Following [26], we present the mean Average Precision $(AP)$ for all categories in aggregate. Additionally, we report on $AP_{n}$ for the novel category. + +# 5.2. Implementation Details + +In our implementation, the CLIP image encoder and text encoder are kept frozen, whereas the pixel decoder, transformer decoder, and mask prediction head undergo training. This strategy effectively harnesses the pre-trained capabilities of CLIP for mask classification and query matching, utilizing query embeddings, thereby obviating the need for category and identity annotations. The CLIP model employed in our framework adopts the ViTB/32 architecture for both the text and image encoders. Throughout the training process, the parameters of these encoders remain frozen. For optimization purposes, we utilize the AdamW optimizer, with the base learning rate set to $1 \times 10^{-4}$ . $\tau = 0.05$ , $\alpha_{T}$ , $\alpha_{M}$ , $\alpha_{V}$ , and $\alpha_{S}$ are all set to 1 based on grid parameter tuning experience. Following a step-wise learning rate schedule, this rate is subsequently reduced by a factor of 10 at $90\%$ and $95\%$ of the total training steps. To assess the robustness and generality of our proposed method, we con + +duct experiments using two distinct backbones: ResNet-50 and SwinB. Our experimental setup is analogous to that of OV2Seg [26]. All experiments are conducted on a high-performance machine equipped with eight NVIDIA A800 GPUs, each possessing 80GB of memory. To facilitate a fair comparison, we also present the inference speed of our method when executed on a single NVIDIA A100 GPU. + +# 5.3. Main Results + +The experimental results presented in Table 2 provide a comprehensive comparison of our proposed SWBridge method with several state-of-the-art approaches across multiple video instance segmentation datasets. A significant finding reveals that SWBridge exhibited superior performance compared to existing methods across 19 out of 22 evaluation metrics in various settings. Furthermore, it attained suboptimal results in the remaining three metrics, thereby underscoring its efficacy in addressing the inherent complexity associated with Open-Vocabulary VIS. This result not only verifies the effectiveness of SWBridge in its design, but also shows that it can adapt well to the characteristics of different datasets, such as changes in video length, number of instances, scene complexity, etc. + +LV-VIS. When trained on the diverse and extensive LVIS dataset, which includes a large number of frequent and common categories, SWBridge exhibits robust generalization capabilities. This is particularly evident in its performance on the LV-VIS dataset, where it achieves the highest $AP$ for both the validation and test sets, including the novel categories $(AP_{n})$ . The superior performance on LV-VIS underscores SWBridge's ability to effectively recognize and segment objects from both base and novel categories, even in the presence of a significant number of unseen classes during training. Although powerful backbone networks such as ConvNeXt-B can enhance performance, we find that the Swin-B backbone network, when combined with the SWBridge, yields even better results. This indicates that SWBridge has a synergistic effect with the backbone network, enabling more efficient capture of spatio-temporal features in videos, and the performance improvement exceeds that achieved by mere network upgrades. + +YT-VIS. SWBridge demonstrates superior performance on the YT-VIS2019 and YT-VIS2021 datasets, further attesting to its strong capability in handling video instance segmentation tasks with a moderate number of categories. Remarkably, this method achieves competitive performance levels without the need for fine-tuning on specific video datasets. This characteristic indicates that SWBridge can effectively leverage image-based training data and successfully generalize to video data, thereby significantly reducing the reliance on extensive video-specific annotations. In comparison with state-of-the-art methods, although OV-Former achieves excellent results by incorporating an ad + +Table 2. Comparison with state-of-the-art methods. All the methods are trained on image dataset LVIS and evaluated on the video instance segmentation datasets directly, i.e., they are not fine-tuned using the training set of each dataset. + +
MethodBackboneLV-VIS val.LV-VIS testYT-VIS2019YT-VIS2021OVISBURST
APAPnAPAPnAPAPnAPAPnAPAPAPAP
DetPro[10]-SORT[2]R506.43.55.82.1--------
Detric[30]-SORT[2]R506.53.45.72.114.63.512.73.16.71.92.5
DetPro[10]-OWTB[20]R507.94.27.02.9-------
Detric[30]-OWTB[20]R507.74.27.02.817.94.516.75.89.02.71.8
Detric[30]-XMem[7]R508.85.47.73.6-------
OV2Seg[26]R5014.211.911.48.927.211.123.67.311.23.72.4
OVFormer[11]R50+ViT-B----34.816.529.815.715.1--
CLIP-VIS [32]R5019.524.214.615.932.223.830.117.914.15.27.7
SWBridge (Our)R5020.724.915.416.233.424.030.918.415.25.58.2
Detric[30]-SORT[2]SwinB12.86.69.44.723.87.921.69.811.72.51.0
Detric[30]-OWTB[20]SwinB14.58.511.86.130.09.727.111.413.63.92.4
Detric[30]-XMem[7]SwinB16.310.613.17.7-------
OV2Seg[26]SwinB21.116.316.411.537.621.333.918.217.54.93.0
OVFormer[11]SwinB+ViT-B----44.321.537.618.321.3--
CLIP-VIS [32]ConvNeXt-B32.240.225.330.642.127.537.922.018.58.312.7
SWBridge (Our)SwinB33.040.526.031.043.028.038.522.519.78.312.8
+ +ditional CLIP image encoder, SWBridge still outperforms CLIP-VIS in terms of the $AP$ metric under similar model architectures. More notably, in the $AP_{n}$ metric, which pertains to the segmentation of new class instances, our bridge strategy comprehensively surpasses the performance gains brought about by the CLIP encoder. + +OVIS. Since the categories of the OVIS dataset highly overlap with the LVIS dataset used for training, we only reports the AP (average precision) performance metric. SWBridge even outperformed OVFormer with the extra CLIP on the R50 backbone and was second only to OVFoemer on the SwinB backbone. SWBridge's ability to maintain high accuracy under these conditions demonstrates its robustness in complex visual environments, where the visibility of objects changes significantly over time. + +BURST. The BURST dataset, with its diverse and uncommon categories, poses a unique challenge due to the wide range of object types. SWBridge's competitive performance on this dataset, especially in terms of $AP_{n}$ , suggests that it can effectively manage the intricacies of uncommon object categories, enhancing its applicability to real-world scenarios where encountering unseen objects is common. + +# 5.4. Ablation Study + +Our ablation experiments continue to use the LVIS dataset for training with ResNet50 as the Backbone and are evaluated on the validation sets of the LV-VIS, OVIS, and BURST datasets, aiming to measure the performance of the + +Table 3. Comparison of ablation results of bridge strategy. + +
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Baseline8.510.25.32.62.3
+ Spatio-Temporal Bridge15.018.010.23.15.4
+ Text-Visual Bridge19.923.314.25.37.9
+ LVT in Eq. 1020.724.915.25.58.2
+ +proposed SWBridge more comprehensively and rigorously. Effectiveness of Bridge Strategy. We first established a baseline variant that removed the Spatio-Temporal Bridge bridge and Text-Visual Bridge modules, while discarding the additional contrast loss $\mathcal{L}_{VT}$ and retaining only the basic $\mathcal{L}_{Seg}$ loss function. In order to achieve instance association, we adopted the memory module proposed in OV2Seg to replace the original spatiotemporal bridge function. + +The experimental results presented in Table 3 confirm the significant effectiveness of each component in the SWBridge method. Specifically, upon integrating the spatio-temporal bridge module into the baseline variant, we observed notable improvements in both $AP$ and $AP_{n}$ metrics across all datasets. This outcome indicates that the spatio-temporal bridge module efficiently captures spatiotemporal correlations between video frames, thereby aiding the model in more accurately understanding and tracking + +Table 4. Impact of ablation results of Importance Weighting. + +
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
w/o Weight15.317.012.34.55.3
w/o Weight in §4.118.522.313.74.96.8
w/o Weight in §4.319.623.514.55.27.5
SWBridge(Our)20.724.915.25.58.2
+ +Table 5. Impact of ablation results of Importance Weighting. + +
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Huberized18.822.413.84.76.4
Cosine18.622.213.94.86.5
Wasserstein19.523.314.45.17.0
Sliced Wasserstein20.724.915.25.58.2
+ +the dynamic changes of instances within videos. Furthermore, the model's performance was significantly enhanced once again after adding the text-visual bridge module on top of the spatio-temporal bridge module. By introducing textual information, the text-visual bridge provides the model with rich semantic cues, greatly enhancing its accuracy in recognizing and understanding instances. Additionally, $\mathcal{L}_{VT}$ synergizes with other bridge strategy modules, collectively driving a comprehensive improvement in the model's performance and generalization. + +Impact of SW Weighting Mechanism. Table 4 shows that without a weighting mechanism in the spatio-temporal or text-visual bridges, the model struggles to utilize key information from instances with marked feature differences, limiting performance. However, incorporating the weighting mechanism into both bridges notably improves model performance. This indicates the mechanism helps the model focus on crucial instances with significant feature differences, assigning them greater weights during distance calculation. This adjustment enhances the model's accuracy and efficiency in processing complex data. + +Comparison of Distance Measurement. In Tab. 5, we compare distance metrics for SWBridge. Huberized distance is sensitive to large differences, Cosine similarity ignores feature magnitudes, and Wasserstein distance misses local details. Our SW distance comprehensively considers high-dimensional feature differences with importance weighting, making it robust for complex domain shifts. + +Sensitivity analysis of $L$ . In the bridge strategy, we ablated the number of sampling iterations to explore its relationship with model performance (Fig.2). Results show a nonlinear relationship: increasing iterations significantly improves $AP$ and $AP_{n}$ across all datasets. However, excessive + +![](images/27355c8d0edde4c5a96964f2667359330784395772273dd47b5c5044581ddc2a.jpg) +Figure 2. Ablation experiment on the number of random projection sampling $L$ . + +iterations introduce noise and computational burden, potentially causing overfitting or reducing training efficiency. + +# 6. Conclusion + +In this paper, we propose an innovative Sliced Wasserstein Bridging Learning Framework (SWbridge), aimed at tackling the challenging task of Open-Vocabulary Video Instance Segmentation. Inspired by metric learning, our framework utilizes the Sliced Wasserstein distance as a core tool to effectively bridge the spatial-temporal and text-visual domains involved in the task. Through domain alignment, we successfully map features from different domains into a unified metric space, preserving temporal consistency and learning intrinsic consistent features across modalities. This significantly improves the fusion of text and visual information, addressing one of the core challenges in Open-Vocabulary VIS. Looking ahead, we plan to explore better optimization methods for the Sliced Wasserstein distance to further enhance the performance of our framework. + +Acknowledgements This work was supported in part by the China Postdoctoral Science Foundation (No. 2024M761805), the Postdoctoral Fellowship Program of CPSF (No. GZB20250390), the Qingdao Postdoctoral Innovation Project (No. QDBSH20240201015), the Major Science and Technology Innovation Projects in Shandong Province (No. 2024CXGC010108). + +# References + +[1] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 6 +[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 7 +[3] Nicolas Bonneel, Julien Rabin, Gabriel Peyre, and Hanspeter Pfister. Sliced and radon wasserstein barycenters of mea + +sures. Journal of Mathematical Imaging and Vision, 51:22-45, 2015. 2 +[4] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In ECCV, 2020. 1 +[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3 +[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2 +[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7 +[8] Zesen Cheng, Kehan Li, Hao Li, Peng Jin, Chang Liu, Xiawu Zheng, Rongrong Ji, and Jie Chen. Instance brownian bridge as texts for open-vocabulary video instance segmentation. arXiv preprint arXiv:2401.09732, 2024. 1, 2, 5 +[9] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 2 +[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 7 +[11] Hao Fang, Peng Wu, Yawei Li, Xinxin Zhang, and Xiankai Lu. Unified embedding alignment for open-vocabulary video instance segmentation. In ECCV, 2024. 1, 2, 3, 7 +[12] Damien Garreau, Rémi Lajugie, Sylvain Arlot, and Francis Bach. Metric learning for temporal sequence alignment. NeurIPS, 2014. 1 +[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2 +[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.6 +[15] Pinxue Guo, Tony Huang, Peiyang He, Xuefeng Liu, Tianjun Xiao, Zhaoyu Chen, and Wenqiang Zhang. Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835, 2023. 1, 2 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 6 +[17] Fusheng Hao, Fengxiang He, Jun Cheng, Lei Wang, Jianzhong Cao, and Dacheng Tao. Collect and select: Semantic alignment metric learning for few-shot learning. In ICCV, 2019. 1 +[18] Dat Huynh, Jason Kuen, Zhe Lin, Jieuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In CVPR, 2022. 2 +[19] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In CVPR, 2023. 3 + +[20] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 7 +[21] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, 130(8): 2022-2039, 2022. 6 +[22] Zheyun Qin, Xiankai Lu, Dongfang Liu, Xiushan Nie, Yi long Yin, Jianbing Shen, and Alexander C Loui. Reformulating graph kernels for self-supervised space-time correspondence learning. IEEE Transactions on Image Processing, 32: 6543-6557, 2023. 1 +[23] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1192-1208, 2023. 1 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3 +[25] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 2 +[26] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 2, 3, 5, 6, 7 +[27] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 6 +[28] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In ICCV, 2023. 1 +[29] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In CVPR, 2021. 2 +[30] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7 +[31] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[32] Wenqi Zhu, Jiale Cao, Jin Xie, Shuangming Yang, and Yanwei Pang. Clip-vis: Adapting clip for open-vocabulary video instance segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 3, 7 \ No newline at end of file diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/images.zip b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..554979ec6ef9787e78c506c3818865d0e1aa83e4 --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92029e5649ba0ca85535ebf1a642eecc3dd7ba46d91cf94c1da31ba4c07c383a +size 445424 diff --git a/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/layout.json b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c21bcc7a47945d2440b6d167917cf47c18bc25f --- /dev/null +++ b/2025/Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation/layout.json @@ -0,0 +1,8739 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "type": "text", + "content": "Sliced Wasserstein Bridge for Open-Vocabulary Video Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "spans": [ + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "text", + "content": "Zheyun Qin" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "text", + "content": ", Deng Yu" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "text", + "content": ", Chuanchen Luo" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "text", + "content": ", Zhumin Chen" + }, + { + "bbox": [ + 146, + 148, + 466, + 165 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "spans": [ + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "type": "text", + "content": "School of Computer Science and Technology, Shandong University " + }, + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 167, + 533, + 178 + ], + "type": "text", + "content": "School of Artificial Intelligence, Shandong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 206, + 200, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 206, + 200, + 219 + ], + "spans": [ + { + "bbox": [ + 151, + 206, + 200, + 219 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 232, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 555 + ], + "type": "text", + "content": "In recent years, researchers have explored the task of open-vocabulary video instance segmentation, which aims to identify, track, and segment any instance within an open set of categories. The core challenge of Open-Vocabulary VIS lies in solving the cross-domain alignment problem, including spatial-temporal and text-visual domain alignments. Existing methods have made progress but still face shortcomings in addressing these alignments, especially due to data heterogeneity. Inspired by metric learning, we propose an innovative Sliced Wasserstein Bridging Learning Framework. This framework utilizes the Sliced Wasserstein distance as the core tool for metric learning, effectively bridging the four domains involved in the task. Our innovations are threefold: (1) Domain Alignment: By mapping features from different domains into a unified metric space, our method maintains temporal consistency and learns intrinsic consistent features between modalities, improving the fusion of text and visual information. (2) Weighting Mechanism: We introduce an importance weighting mechanism to enhance the discriminative ability of our method when dealing with imbalanced or significantly different data. (3) High Efficiency: Our method inherits the computational efficiency of the Sliced Wasserstein distance, allowing for online processing of large-scale video data while maintaining segmentation accuracy. Through extensive experimental evaluations, we have validated the robustness of our concept and the effectiveness of our framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 695 + ], + "type": "text", + "content": "In recent years, video instance segmentation methods [23, 27, 28] have demonstrated significant improvements and breakthroughs in various application fields. However, most existing methods rely on the Closed-Set Assumption [4, 22], where the learning and recognition scope of the model are strictly limited to known categories or labels during training and testing. This assumption ignores the possibility of unknown categories in the real world, limiting the model's ability to effectively detect unknown categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 207, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 303 + ], + "type": "text", + "content": "To overcome this limitation, researchers have begun to explore the task of open-vocabulary video instance segmentation [15, 26] (Open-Vocabulary VIS), which aims to identify, track, and segment any instance within an open set of categories. These Open-Vocabulary VIS methods can not only handle traditional predefined categories but also flexibly adapt to new user-specified categories, achieving the advanced function of on-demand segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 553, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 519 + ], + "type": "text", + "content": "The core of Open-Vocabulary VIS methods lies in solving the cross-domain alignment problem, including spatial-temporal domain alignment and text-visual domain alignment. Existing methods have different focuses. In terms of spatial-temporal domain alignment, current methods extract class-agnostic instance features based on the foundational model Mask2Former-VIS. Furthermore, OV2Seg [26], OpenVIS [15], and CLIP-VIS [32] employ memory-based approaches to perform cross-frame instance matching. In terms of text-visual domain alignment, existing methods mainly rely on the zero-shot performance exhibited by visual-language models, i.e., CLIP [24] pretrained on largescale image-text pairs. Specifically, BriVIS [8] and OV-Former [11] introduce additional CLIP image encoders on top of visual feature extractors. The difference lies in that BriVIS integrates them based on contrastive learning, while OV-Former maps text and visual features into the same space vis the attention mechanism." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 700 + ], + "type": "text", + "content": "Despite the significant progress made by existing Open-Vocabulary VIS methods, there are still many shortcomings. The cross-domain alignment problem remains inadequately addressed, and the inherent data heterogeneity results in significant differences in feature distributions across different domains. In light of this, we cannot help but trace back to the development history of traditional machine learning methods, seeking a tool that can simultaneously solve spatial-temporal domain alignment and text-visual domain alignment. We find that the metric learning method in the field of machine learning provides an effective approach to addressing the alignment problem [12, 17]. By learning a suitable metric space, this paradigm makes similar samples closer while dissimilar samples relatively farther apart, providing a powerful tool for solving the alignment problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Guided by the idea of metric learning, we propose an" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 193, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 193, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 193, + 713 + ], + "type": "text", + "content": "*Corresponding author: Zhumin Chen." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12470" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "innovative Sliced Wasserstein Bridging Learning Framework, i.e., SWbridge. This framework utilizes the Sliced Wasserstein distance [3] as the core tool for metric learning. This framework cleverly utilizes the Sliced Wasserstein distance as the core tool for metric learning. Specifically, we minimize the Sliced Wasserstein distance between samples based on random paths through sampling, achieving feature embedding alignment. We bridge four domains based on a strategy to address the challenges of cross-domain alignment (spatial-temporal domain and text-visual domain)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 540 + ], + "type": "text", + "content": "Our innovations are mainly concentrated in three aspects: (1) Domain Alignment. The Sliced Wasserstein distance has the advantage of accurately capturing differences in data distributions in high-dimensional spaces. Our method uses the Sliced Wasserstein distance as a bridge to map features from different domains into a unified metric space. This not only maintains temporal consistency by comparing instance embeddings between adjacent frames, achieving temporal alignment and effectively mitigating challenges such as progressive occlusion, but also learns intrinsic consistent features between modalities, narrowing the semantic gap and improving the fusion of text and visual information through modality alignment. (2) Weighting Mechanism: We introduce an importance weighting mechanism to further enhance its discriminative ability when dealing with imbalanced data or data with significant feature differences. In open-vocabulary video instance segmentation tasks, there may be significant feature differences between different instances, and some instances may be more critical or informative. Through the importance weighting mechanism, our method can dynamically adjust the contribution of different instances in the metric space, allowing key instances to have greater weight in distance calculations. (3) High Efficiency: Our SWbridge inherits the computational efficiency advantage of the Sliced Wasserstein distance, making it highly efficient when processing large-scale video data. Without introducing additional network parameters, our method can handle video in an online manner while maintaining segmentation accuracy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "content": "Combining the aforementioned perspectives and innovations, we propose an open-vocabulary video instance segmentation method capable of bridging multiple domains. This method has achieved remarkable results on multiple datasets related to open-vocabulary and video tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 614, + 148, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 148, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 148, + 627 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 635, + 295, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 647 + ], + "type": "text", + "content": "2.1. Open-Vocabulary Detection and Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Open Vocabulary Detection (OVD) [9, 25] and Open Vocabulary Segmentation (OVS) [13, 18] are cutting-edge tasks in the field of computer vision, enabling models to be trained on images containing unannotated novel objects, thereby breaking the closed-set constraint. This break-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "type": "text", + "content": "through is primarily attributed to the application of weak supervision signals, namely the utilization of image-text pairs (such as image-caption pairs and image-level labels) or large pre-trained Visual-Language Models (VLMs), such as CLIP [24]. Supported by weak supervision signals, methods for OVD and OVS can be categorized into four main types [31]. The first type is region-aware training, which learns object feature representations by exploring the intrinsic links between image regions and text descriptions without relying on the VLMs' image encoder (VLMsIE) or direct object annotations. The second type is the pseudo-labeling method, which enhances model generalization by generating pseudo-labels from preliminary predictions on unannotated images using both image-text pairs and the VLMs-IE. The other two types are knowledge distillation and transfer learning. Knowledge distillation transfers knowledge from pre-trained VLMs to new models or tasks via a distillation mechanism, while transfer learning applies knowledge learned by pre-trained models on specific tasks to new tasks or domains. Both rely on the VLMsIE and seldom involve direct training on image-text pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 397 + ], + "type": "text", + "content": "Furthermore, zero-shot learning [29] addresses the closed-set constraint in scene-aware tasks and differs from open vocabulary methods. It prohibits access to weak supervision signals during training, but can be converted into Open-Vocabulary tasks upon gaining access, enhancing the model's generalization and recognition of novel objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 553, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 421 + ], + "type": "text", + "content": "2.2. Open-Vocabulary Video Instance Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 715 + ], + "type": "text", + "content": "Open-vocabulary video instance segmentation, as an emerging vision task, aims to simultaneously classify, track, and segment arbitrary objects within open categories in videos, attracting considerable attention from researchers in recent years. OV2Seg [26] has taken the lead in this field by not only constructing the Large Vocabulary Video Instance Segmentation (LV-VIS) dataset but also proposing the first end-to-end benchmark for open-vocabulary video instance segmentation. OV2Seg method leverages the Mask2Former [6] framework to extract class-agnostic masks and query embeddings, utilizes the CLIP text encoder for precise mask classification, and achieves cross-frame instance tracking through a long-term matching strategy. Subsequent research has mostly followed this basic approach and expanded upon it. OpenVIS [15] introduces a two-stage framework, OVIS, which adopts a neighboring matching strategy for instance tracking, effectively simplifying the processing pipeline. BriVIS [8] models instances as Brownian bridges and closely aligns bridge-level instance representations with category texts through contrastive learning, also employing a neighboring matching strategy to enhance the accuracy of instance recognition. OVFormer [11] introduces a novel unified embedding alignment module that effectively addresses the domain gap between instance" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12471" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "queries and text embeddings, demonstrating good practicality with its semi-online processing approach. However, the aforementioned methods primarily focus on bridging the domain discrepancy between text and vision, with insufficient attention to spatiotemporal consistency among cross-frame instances. CLIP-VIS [32] improves upon this by introducing a time-topK enhanced matching strategy, which strengthens temporal modeling capabilities between frames and improves matching accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 300 + ], + "type": "text", + "content": "We believe that the core of the open-vocabulary video instance segmentation task lies in simultaneously addressing the alignment of the spatial-temporal domain and the text-visual domain. To this end, based on the concept of metric learning, we employ the Sliced Wasserstein distance to bridge various domains. This approach not only effectively narrows the domain gap between modalities but also ensures spatio-temporal consistency of instances across frames, providing a novel solution for the open-vocabulary video instance segmentation task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 194, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 194, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 194, + 317 + ], + "type": "text", + "content": "2.3. Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 550 + ], + "type": "text", + "content": "The core of Visual Language Models (VLMs) [5] is their deep training on large-scale image-text paired data, enabling the fusion and understanding of visual and textual information, and endowing them with powerful zero-shot object recognition capabilities [19]. CLIP [24], a prominent VLM, uses an image encoder to extract visual features and a text encoder to generate text embeddings, jointly constructing a cross-modal representation space. However, applying VLMs to open-vocabulary video instance segmentation is challenging. VLMs, primarily trained on images, struggle with understanding dynamic video scenes rich in spatiotemporal information, and lack the capability to maintain object consistency across frames. Additionally, the domain shift between text and vision poses a significant challenge, potentially leading to biases in mapping textual information to the visual space. To overcome these limitations, we propose a progressive approach that addresses the spatiotemporal consistency issue and the domain shift problem, enabling VLMs to better adapt to video instance segmentation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 139, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 139, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 139, + 571 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 243, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 243, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 243, + 590 + ], + "type": "text", + "content": "3.1. Open-Vocabulary VIS Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": "Given a test video " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{test}" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " frames, the objective in Open-Vocabulary VIS is to accurately predict all " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " instances belonging to the categories in " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\mathcal{C}_{base} \\cup \\mathcal{C}_{novel}" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " by the trained model " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{base}" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " is the set of base (training) categories, and " + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{novel}" + }, + { + "bbox": [ + 55, + 595, + 295, + 678 + ], + "type": "text", + "content": " is the set of novel categories that are not seen during training but may appear in the test videos. The prediction can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 683, + 294, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 683, + 294, + 697 + ], + "spans": [ + { + "bbox": [ + 94, + 683, + 294, + 697 + ], + "type": "interline_equation", + "content": "\\left\\{\\left\\{m _ {1}, m _ {2}, \\dots , m _ {T} \\right\\}, c \\right\\} _ {n} ^ {N} = f _ {\\theta} \\left(\\mathcal {D} _ {t e s t}\\right), \\tag {1}", + "image_path": "d462ce904ee4bda0147609c9913ab320747935710b407cdf0df94dbeaa3aa46b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "type": "inline_equation", + "content": "\\{m_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "type": "text", + "content": " is the segmentation masks, and " + }, + { + "bbox": [ + 55, + 700, + 275, + 714 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 467, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 467, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 467, + 83 + ], + "type": "text", + "content": "3.2. Sliced Wasserstein Distance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": "One-dimensional Wasserstein Distance. For one-dimensional probability measures " + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_p(\\mathbb{R})" + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": "-Wasserstein distance is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 131, + 553, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 131, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 348, + 131, + 553, + 159 + ], + "type": "interline_equation", + "content": "\\mathrm {W} _ {p} ^ {p} (\\mu , \\nu) = \\int_ {0} ^ {1} | F _ {\\mu} ^ {- 1} (z) - F _ {\\nu} ^ {- 1} (z) | ^ {p} d z, \\tag {2}", + "image_path": "f62d409ac436f84964c9875193ae9c178bd32d3b31a74048d616146ca9f09800.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "inline_equation", + "content": "F_{\\mu}" + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "inline_equation", + "content": "F_{\\nu}" + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": " are the cumulative distribution functions (CDFs) of " + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 165, + 553, + 213 + ], + "type": "text", + "content": ". This formulation provides a closed form for computing the Wasserstein distance in one-dimensional spaces, making it well-suited for projected measures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "content": "Sliced Wasserstein Distance. To generalize the Wasserstein distance to higher-dimensional measures, the Sliced Wasserstein Distance (SWD) projects the measures " + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_p(\\mathbb{R}^d)" + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "content": " onto one-dimensional subspaces, and then averages the one-dimensional Wasserstein distances from these projections. For " + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "inline_equation", + "content": "\\mu, \\nu \\in \\mathcal{P}_p(\\mathbb{R}^d)" + }, + { + "bbox": [ + 313, + 213, + 554, + 286 + ], + "type": "text", + "content": ", the SWD is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 293, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 293, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 345, + 293, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\mathrm {S W} _ {p} ^ {p} (\\mu , \\nu) = \\mathbb {E} _ {\\theta \\sim \\mathcal {U} \\left(\\mathbb {S} ^ {d - 1}\\right)} \\left[ \\mathrm {W} _ {p} ^ {p} \\left(\\theta \\sharp \\mu , \\theta \\sharp \\nu\\right) \\right], \\tag {3}", + "image_path": "8f1c47199d33c5ef035fc78bbedf81cf7f35646224a755ef6ccfe27504f76829.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\mu" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\nu" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " represent the push-forward measures of " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " along direction " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{S}^{d - 1}" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": ", the unit sphere in " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": ". This projection " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "f(x) = \\theta^{\\top}x" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " maps points from " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 313, + 314, + 554, + 373 + ], + "type": "text", + "content": ", enabling the computation of Wasserstein distances in a one-dimensional space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "text", + "content": "Since the expectation in Eq. 3 is computationally expensive, SWD is typically approximated by averaging over " + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "text", + "content": " independent directions, " + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\theta_{1},\\ldots ,\\theta_{L}" + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "text", + "content": ", sampled from " + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\mathbb{S}^{d - 1})" + }, + { + "bbox": [ + 313, + 374, + 553, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 418, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 418, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 352, + 418, + 553, + 452 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbf {S W}} _ {p} ^ {p} (\\mu , \\nu ; L) = \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\mathbf {W} _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right), \\tag {4}", + "image_path": "02de60be971583d004ec309226070ca955948dc70da99457aff87d8bdd331e4b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{l}\\sharp \\mu" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{l}\\sharp \\nu" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": " are projected representations of " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": " along the direction " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": ". The number of projections " + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 456, + 554, + 491 + ], + "type": "text", + "content": " controls the accuracy of the Monte Carlo approximation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 502, + 370, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 370, + 513 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 370, + 513 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 521, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 641 + ], + "type": "text", + "content": "As illustrated in Fig. 1, we propose the SWbridge framework. Firstly, in §4.1, we integrate a category-agnostic feature extractor. Then, in §4.2, we introduce a spatio-temporal bridging module that robustly establishes instance associations between adjacent frames, ensuring the consistency and coherence of temporal information. Finally, in §4.3, we design a modal bridging module that performs fine-grained semantic mapping across the text and visual domains to enhance cross-domain understanding and effectively address the challenges posed by domain shifts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 648, + 491, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 491, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 491, + 661 + ], + "type": "text", + "content": "4.1. Class-agnostic Feature Extractor" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Given a video " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " frames, " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\{V_t \\in \\mathbb{R}^{3 \\times H \\times W}\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " denote the height and width of each frame, our model follows the conventional paradigm [11, 26, 32] by adopting the Mask2Former" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12472" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 321, + 274 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 321, + 274 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 321, + 274 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 321, + 274 + ], + "type": "image", + "image_path": "8a47e114bbf7d6746aa848776aae236520e165cbbf227cd6371392875237e0c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 282, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 61, + 282, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 547, + 295 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed SWBridge framework (left) and the Sliced Wasserstein distance calculation process (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 321, + 70, + 417, + 272 + ], + "blocks": [ + { + "bbox": [ + 321, + 70, + 417, + 272 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 417, + 272 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 417, + 272 + ], + "type": "image", + "image_path": "e4d44f98b684a3c9e393ec78b641d1b754f9f2e02dacbedee2631a771fb76644.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 417, + 70, + 541, + 274 + ], + "blocks": [ + { + "bbox": [ + 417, + 70, + 541, + 274 + ], + "lines": [ + { + "bbox": [ + 417, + 70, + 541, + 274 + ], + "spans": [ + { + "bbox": [ + 417, + 70, + 541, + 274 + ], + "type": "image", + "image_path": "b29e81eceee9d52640a0f0f236c5595caef8d913c35c350151ff7d0954387c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": "architecture as a class-agnostic feature extractor. The Mask2Former comprises three core components: an image encoder " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ", a pixel decoder " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{decode}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ", and a Transformer decoder " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Transformer}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ". Specifically, the frozen CLIP model " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": " processes each frame " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": " independently to extract high-level visual features " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ". These features are then input to the pixel decoder " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{decode}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ", which upsamples them to produce a multiscale feature map " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": " preserving spatial details. Furthermore, a set of learnable queries " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": " is initialized and fed into the Transformer decoder " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Transformer}}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": " along with the high-resolution feature map " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": ". The Transformer decoder processes these inputs to generate the query embeddings " + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 54, + 315, + 296, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 192, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 192, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 192, + 482 + ], + "type": "text", + "content": "4.2. Spatio-Temporal Bridge" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "content": "After extracting category-agnostic features, the next challenge is to construct a spatio-temporal bridge to establish the correspondence between instances in consecutive frames " + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "content": ". Given the dynamic nature of video content, objects may encounter occlusion, movement, and appearance changes. Therefore, it is particularly important to ensure the accuracy of instance tracking over time. To achieve this, we employ the Sliced Wasserstein metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": "Initially, for the query embeddings of frames " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": ", we perform random sampling to construct random paths " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "Z_{i,j} = \\mathbf{q}_{t,i} - \\mathbf{q}_{t + 1,j}" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{t,i} \\in E_t" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{t + 1,i} \\in E_{t + 1}" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": " are outputs of the Transformer decoder " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{Transformer}}" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": ". These paths are then normalized to obtain the unit direction " + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\theta_{i,j} = \\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2}" + }, + { + "bbox": [ + 55, + 584, + 296, + 661 + ], + "type": "text", + "content": ", which serves as the projection direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "text", + "content": "Subsequently, we sample a new projection direction " + }, + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{i,j}" + }, + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{\\kappa}(\\theta_{i,j};\\frac{Z_{i,j}}{\\|Z_{i,j}\\|_2})" + }, + { + "bbox": [ + 55, + 662, + 296, + 714 + ], + "type": "text", + "content": ". This introduces randomness, potentially yielding different projection directions each time, which aids in exploring multiple relative positional relation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 339 + ], + "type": "text", + "content": "ships between feature points. This, in turn, may enhance the model's robustness and generalization capability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": "With the projection direction " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}_{i,j}" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " in hand, we proceed to calculate the one-dimensional Wasserstein distance. This involves projecting " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " direction to obtain one-dimensional distributions " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\mu" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\nu" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": ". These distributions are then sorted, and the inverse functions of their cumulative distribution functions (CDFs) are computed. Finally, the one-dimensional Wasserstein distance " + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "inline_equation", + "content": "W_{p}^{p}(\\theta \\sharp \\mu ,\\theta \\sharp \\nu)" + }, + { + "bbox": [ + 313, + 339, + 554, + 437 + ], + "type": "text", + "content": " is calculated using the formula:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 441, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 441, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 349, + 441, + 553, + 468 + ], + "type": "interline_equation", + "content": "W _ {p} ^ {p} (F, G) = \\int_ {0} ^ {1} | F ^ {- 1} (t) - G ^ {- 1} (t) | ^ {p} d t, \\tag {5}", + "image_path": "5e8b6f7e7793d115278c88cddb420d972192f5a5f5c764bbdb3e232da2ac0622.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": " are the CDFs of " + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\mu" + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "inline_equation", + "content": "\\theta \\sharp \\nu" + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 472, + 553, + 495 + ], + "type": "text", + "content": " is typically set to 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": "The next step involves importance weighting and Monte Carlo estimation. We repeat the aforementioned process " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": " times to obtain " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": " projection directions " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\{\\theta_1, \\dots, \\theta_L\\}" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": " and the corresponding set of one-dimensional Wasserstein distances. Each distance is then weighted with " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": " is determined by an increasing function " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": ". Increasing function is " + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "inline_equation", + "content": "f(x) = e^x" + }, + { + "bbox": [ + 313, + 496, + 554, + 628 + ], + "type": "text", + "content": ", places higher importance on larger differences, which may highlight instances with more substantial temporal changes. The weighted average of these distances serves as the approximation of SW, calculated using the formula:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 632, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 632, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 329, + 632, + 553, + 666 + ], + "type": "interline_equation", + "content": "S W = \\frac {1}{L} \\sum_ {l = 1} ^ {L} w \\cdot \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mathcal {M} _ {t} , \\theta_ {l} \\sharp \\mathcal {T}\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mathcal {M} _ {t} , \\theta_ {j} \\sharp \\mathcal {T}\\right)\\right)}, \\tag {6}", + "image_path": "1573a015a8f651734b4e3a8ae4838e237739f4160b67109fb41cb4beb181a70f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 669, + 523, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 523, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 523, + 681 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 313, + 669, + 523, + 681 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 669, + 523, + 681 + ], + "type": "text", + "content": " for each projection direction is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 685, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 685, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 366, + 685, + 553, + 717 + ], + "type": "interline_equation", + "content": "w = \\frac {f \\left(W _ {p} ^ {p} \\left(\\theta_ {l} \\sharp \\mu , \\theta_ {l} \\sharp \\nu\\right)\\right)}{\\sum_ {j = 1} ^ {L} f \\left(W _ {p} ^ {p} \\left(\\theta_ {j} \\sharp \\mu , \\theta_ {j} \\sharp \\nu\\right)\\right)}. \\tag {7}", + "image_path": "667206747d14f2d9afff5a1cdccbf3a77974499e1d9a40f34600a75de1dccd10.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12473" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "Eq.6 yields a weighted Sliced Wasserstein distance that reflects the relationship between adjacent frames along random projection directions, considering both the geometric positions of data points and their differences through importance weighting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": "To establish instance correspondences, we compute the SW values between all pairs of instances across consecutive frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{SW}_p^p (E_t,E_{t + 1})" + }, + { + "bbox": [ + 55, + 133, + 295, + 251 + ], + "type": "text", + "content": ". These values quantify the difference between instances along the projection direction, with smaller SW values indicating stronger spatio-temporal consistency. By selecting the instance pairs with the smallest SW values, we identify those with the highest temporal alignment, ensuring robust and meaningful associations for instance tracking. Formally, the instance matching objective is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 261, + 295, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 261, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 261, + 295, + 274 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T i m e}} = \\arg \\min \\mathrm {S W} \\left(E _ {t}, E _ {t + 1}\\right), \\tag {8}", + "image_path": "e079c122908618b08e810cba187b1b8041ca48d12625d360be1466f7bf6c001b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 378 + ], + "type": "text", + "content": "where the selected pairs correspond to instances with the strongest spatio-temporal consistency, facilitating accurate tracking and correlation across frames. Our method thus emphasizes directions and features that capture significant inter-frame changes, effectively prioritizing instances that reflect spatial and appearance consistency. The adaptive weighting in Eq.6 further enhances tracking by highlighting subtle variations, ensuring robust temporal consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 167, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 167, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 167, + 398 + ], + "type": "text", + "content": "4.3. Text-Visual Bridge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 545 + ], + "type": "text", + "content": "After bridging the temporal consistency between the spatial and temporal domains through the spatio-temporal bridge, we are still confronted with the challenge of addressing the modal domain shift between the text and visual domains. Video content, with its dynamic and diverse nature, introduces additional complexity as textual descriptions and visual appearances can vary significantly. To ensure the accurate alignment between text and visual features, which is crucial in this context, we continue to leverage the Sliced Wasserstein metric. This metric has proven effective in our previous spatio-temporal bridging efforts, leading us to adopt it for modal alignment in our current method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "Specifically, we first combine the temporal feature " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " with the query embeddings " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "E_{t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " through multiplication to obtain the visual embeddings " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ". Subsequently, we apply a mask pooling operation to derive a class-agnostic mask " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " and the frame features " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ". Lastly, we obtain the class text embeddings " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{H}\\in \\mathbb{R}^{L\\times D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " generated by the CLIP text encoder, where " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " represents the number of categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "It is worth noting that our cross-modal alignment method bears similarity to the Sliced Wasserstein metric procedure detailed in §4.2. We begin by defining the projection path of the cross-modal module as " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "Z_{i,j}^{c} = \\pmb{m}_{t,i} - \\pmb{h}_{t,j}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". To obtain the projection direction, we normalize this path, resulting in " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i,j}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". Given the relatively small sizes of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_t" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "this module departs from the random sampling approach and instead adopts a full-path full-projection strategy. Upon successfully obtaining the projection direction " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", we proceed to calculate the one-dimensional Wasserstein distance. We then apply importance weighting and incorporate the Monte Carlo estimation method in our subsequent calculations. Following the logic of Eq. 6 and Eq. 7, we ultimately compute the weighted average as an approximation of the Sliced Wasserstein distance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": "Additionally, similar to Eq. 8, we introduce an additional loss function " + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Modal}" + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": " to constrain the modality alignment process. This loss function is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 228, + 553, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 228, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 365, + 228, + 553, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {M o d a l}} = \\arg \\min \\operatorname {S W} \\left(\\mathcal {M} _ {t}, \\mathcal {H}\\right). \\tag {9}", + "image_path": "da006c5e18444480ed1613300310197fcaa0969f68a6c0bc09fae30605b7c4fd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 247, + 403, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 247, + 403, + 258 + ], + "spans": [ + { + "bbox": [ + 314, + 247, + 403, + 258 + ], + "type": "text", + "content": "4.4. Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "type": "text", + "content": "To mine similarity relations in cross-text video pairs, we draw on the successful experience of cross-modal learning, as demonstrated in [8]. To this end, we introduce a cross-text similarity alignment loss based on contrastive learning. The loss function is formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 331, + 553, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 331, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 553, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V T}} = - \\frac {1}{N L} \\sum_ {i, j} \\log \\frac {\\exp \\left(d _ {i j} / \\tau\\right)}{\\sum_ {k} \\exp \\left(d _ {i k} / \\tau\\right) + \\sum_ {l} \\exp \\left(d _ {j l} / \\tau\\right)}, \\tag {10}", + "image_path": "49b0428db6960d2e3761a0a329a4bcbd52a40b36078ae4d250763f55b506e6a1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "inline_equation", + "content": "d_{ij} = \\mathrm{sim}(\\pmb {m}_i,\\pmb {h}_j),\\mathrm{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "text", + "content": " represents the similarity metric function, and " + }, + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 371, + 553, + 395 + ], + "type": "text", + "content": " is the temperature hyperparameter." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "content": "The Total Loss " + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Total}}" + }, + { + "bbox": [ + 313, + 396, + 553, + 418 + ], + "type": "text", + "content": " is a weighted combination of multiple loss components, which can be written as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 429, + 553, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 429, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 319, + 429, + 553, + 442 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T o t a l}} = \\alpha_ {T} \\mathcal {L} _ {\\text {T i m e}} + \\alpha_ {M} \\mathcal {L} _ {\\text {M o d a l}} + \\alpha_ {V} \\mathcal {L} _ {\\mathrm {V T}} + \\alpha_ {S} \\mathcal {L} _ {\\text {S e g}}, \\tag {11}", + "image_path": "53abda3794974283001d6d1e033142a3329084b7573e8d37fca8c4ab0c81cd99.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Seg}}" + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "text", + "content": " represents the binary cross-entropy and dice losses for mask prediction, following the approach in [26]. The coefficients " + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_{T}, \\alpha_{M}, \\alpha_{V}" + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "inline_equation", + "content": "\\alpha_{S}" + }, + { + "bbox": [ + 313, + 450, + 553, + 498 + ], + "type": "text", + "content": " are the corresponding weighting factors for each loss component." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Computational complexity. In our Sliced Wasserstein metric process, the core steps involve matrix operations and sorting procedures. Initially, we generate a set of initial projection vectors by randomly selecting samples and computing the differences among these samples. The complexity of this selection operation is " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(L)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", while the complexity of the difference calculation is " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(L \\cdot D)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " represents the number of projection vectors and " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " is the dimensionality of the samples. Subsequently, we proceed to the normalization and resampling phase, which also maintains a complexity of " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(L \\cdot D)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". When calculating the one-dimensional Wasserstein distance, we need to perform matrix multiplication, a step with a complexity of " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(N \\cdot D \\cdot L)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " is the number of samples. Additionally, we need to sort the projected data, with a sorting operation complexity of " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(N \\log N)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". In summary, the overall complexity of our method can be approximated as " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "O(N \\cdot D \\cdot L + N \\log N)" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". In practical applications, as the number of samples " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " and the number of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12474" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 298, + 188 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 94 + ], + "type": "text", + "content": "Table 1. Statistics of the dataset. Cat. and Ins. represent the number of categories and instances respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 298, + 188 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 298, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 298, + 188 + ], + "type": "table", + "html": "
DatasetVideosCat.Ins.TrainVal.Test
LV-VIS4,8281,196-3,083837908
YT-VIS 192,883404,8832,238302343
YT-VIS 213,859408,1712,985421453
OVIS901255,223607140154
BURST2,91448216,0005009931,421
", + "image_path": "85e66470a14b8dde44e58982d0305e40405cc70360334052215d11f2473584b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": "projection vectors " + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": " increase, the computational cost of matrix multiplication will become dominant. Although sorting and other operations with linear complexity also exist, their contribution to the overall complexity is relatively small." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 136, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 136, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 136, + 282 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 231, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 231, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 231, + 300 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 304, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 387 + ], + "type": "text", + "content": "We initially train our method on a combined set of common and frequent categories derived from the LVIS dataset [16]. Subsequently, we assess the performance of our method on both the validation and test sets of the LV-VIS dataset [26], as well as the validation set of several datasets, namely OVIS [21], YT-VIS 2019/2021 [27], and BURST [1]. The statistics of the evaluation dataset are detailed in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 448 + ], + "type": "text", + "content": "LVIS [16] dataset comprises 1,203 object categories, which significantly exceeds the number of categories in COCO. Following ViLD[14], we select 866 frequent and common categories as the training categories and reserve the remaining 337 rare categories as novel categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": "Evaluation Metrics. Following [26], we present the mean Average Precision " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "(AP)" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " for all categories in aggregate. Additionally, we report on " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "AP_{n}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " for the novel category." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 492, + 188, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 188, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 188, + 506 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "In our implementation, the CLIP image encoder and text encoder are kept frozen, whereas the pixel decoder, transformer decoder, and mask prediction head undergo training. This strategy effectively harnesses the pre-trained capabilities of CLIP for mask classification and query matching, utilizing query embeddings, thereby obviating the need for category and identity annotations. The CLIP model employed in our framework adopts the ViTB/32 architecture for both the text and image encoders. Throughout the training process, the parameters of these encoders remain frozen. For optimization purposes, we utilize the AdamW optimizer, with the base learning rate set to " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{T}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{M}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{V}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{S}" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " are all set to 1 based on grid parameter tuning experience. Following a step-wise learning rate schedule, this rate is subsequently reduced by a factor of 10 at " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " of the total training steps. To assess the robustness and generality of our proposed method, we con" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "duct experiments using two distinct backbones: ResNet-50 and SwinB. Our experimental setup is analogous to that of OV2Seg [26]. All experiments are conducted on a high-performance machine equipped with eight NVIDIA A800 GPUs, each possessing 80GB of memory. To facilitate a fair comparison, we also present the inference speed of our method when executed on a single NVIDIA A100 GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 166, + 399, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 399, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 399, + 178 + ], + "type": "text", + "content": "5.3. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 185, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 353 + ], + "type": "text", + "content": "The experimental results presented in Table 2 provide a comprehensive comparison of our proposed SWBridge method with several state-of-the-art approaches across multiple video instance segmentation datasets. A significant finding reveals that SWBridge exhibited superior performance compared to existing methods across 19 out of 22 evaluation metrics in various settings. Furthermore, it attained suboptimal results in the remaining three metrics, thereby underscoring its efficacy in addressing the inherent complexity associated with Open-Vocabulary VIS. This result not only verifies the effectiveness of SWBridge in its design, but also shows that it can adapt well to the characteristics of different datasets, such as changes in video length, number of instances, scene complexity, etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "text", + "content": "LV-VIS. When trained on the diverse and extensive LVIS dataset, which includes a large number of frequent and common categories, SWBridge exhibits robust generalization capabilities. This is particularly evident in its performance on the LV-VIS dataset, where it achieves the highest " + }, + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "text", + "content": " for both the validation and test sets, including the novel categories " + }, + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "inline_equation", + "content": "(AP_{n})" + }, + { + "bbox": [ + 313, + 354, + 554, + 569 + ], + "type": "text", + "content": ". The superior performance on LV-VIS underscores SWBridge's ability to effectively recognize and segment objects from both base and novel categories, even in the presence of a significant number of unseen classes during training. Although powerful backbone networks such as ConvNeXt-B can enhance performance, we find that the Swin-B backbone network, when combined with the SWBridge, yields even better results. This indicates that SWBridge has a synergistic effect with the backbone network, enabling more efficient capture of spatio-temporal features in videos, and the performance improvement exceeds that achieved by mere network upgrades." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "YT-VIS. SWBridge demonstrates superior performance on the YT-VIS2019 and YT-VIS2021 datasets, further attesting to its strong capability in handling video instance segmentation tasks with a moderate number of categories. Remarkably, this method achieves competitive performance levels without the need for fine-tuning on specific video datasets. This characteristic indicates that SWBridge can effectively leverage image-based training data and successfully generalize to video data, thereby significantly reducing the reliance on extensive video-specific annotations. In comparison with state-of-the-art methods, although OV-Former achieves excellent results by incorporating an ad" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12475" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 102, + 547, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods. All the methods are trained on image dataset LVIS and evaluated on the video instance segmentation datasets directly, i.e., they are not fine-tuned using the training set of each dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 102, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 357 + ], + "type": "table", + "html": "
MethodBackboneLV-VIS val.LV-VIS testYT-VIS2019YT-VIS2021OVISBURST
APAPnAPAPnAPAPnAPAPnAPAPAPAP
DetPro[10]-SORT[2]R506.43.55.82.1--------
Detric[30]-SORT[2]R506.53.45.72.114.63.512.73.16.71.92.5
DetPro[10]-OWTB[20]R507.94.27.02.9-------
Detric[30]-OWTB[20]R507.74.27.02.817.94.516.75.89.02.71.8
Detric[30]-XMem[7]R508.85.47.73.6-------
OV2Seg[26]R5014.211.911.48.927.211.123.67.311.23.72.4
OVFormer[11]R50+ViT-B----34.816.529.815.715.1--
CLIP-VIS [32]R5019.524.214.615.932.223.830.117.914.15.27.7
SWBridge (Our)R5020.724.915.416.233.424.030.918.415.25.58.2
Detric[30]-SORT[2]SwinB12.86.69.44.723.87.921.69.811.72.51.0
Detric[30]-OWTB[20]SwinB14.58.511.86.130.09.727.111.413.63.92.4
Detric[30]-XMem[7]SwinB16.310.613.17.7-------
OV2Seg[26]SwinB21.116.316.411.537.621.333.918.217.54.93.0
OVFormer[11]SwinB+ViT-B----44.321.537.618.321.3--
CLIP-VIS [32]ConvNeXt-B32.240.225.330.642.127.537.922.018.58.312.7
SWBridge (Our)SwinB33.040.526.031.043.028.038.522.519.78.312.8
", + "image_path": "c84e31c1e966cc2992ac9f77e51ac7f3734967614cb2f0640cfb4c1fa4195926.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "text", + "content": "ditional CLIP image encoder, SWBridge still outperforms CLIP-VIS in terms of the " + }, + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "text", + "content": " metric under similar model architectures. More notably, in the " + }, + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "inline_equation", + "content": "AP_{n}" + }, + { + "bbox": [ + 54, + 376, + 294, + 448 + ], + "type": "text", + "content": " metric, which pertains to the segmentation of new class instances, our bridge strategy comprehensively surpasses the performance gains brought about by the CLIP encoder." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 555 + ], + "type": "text", + "content": "OVIS. Since the categories of the OVIS dataset highly overlap with the LVIS dataset used for training, we only reports the AP (average precision) performance metric. SWBridge even outperformed OVFormer with the extra CLIP on the R50 backbone and was second only to OVFoemer on the SwinB backbone. SWBridge's ability to maintain high accuracy under these conditions demonstrates its robustness in complex visual environments, where the visibility of objects changes significantly over time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 556, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 556, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 556, + 295, + 640 + ], + "type": "text", + "content": "BURST. The BURST dataset, with its diverse and uncommon categories, poses a unique challenge due to the wide range of object types. SWBridge's competitive performance on this dataset, especially in terms of " + }, + { + "bbox": [ + 54, + 556, + 295, + 640 + ], + "type": "inline_equation", + "content": "AP_{n}" + }, + { + "bbox": [ + 54, + 556, + 295, + 640 + ], + "type": "text", + "content": ", suggests that it can effectively manage the intricacies of uncommon object categories, enhancing its applicability to real-world scenarios where encountering unseen objects is common." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 149, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 149, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 149, + 661 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our ablation experiments continue to use the LVIS dataset for training with ResNet50 as the Backbone and are evaluated on the validation sets of the LV-VIS, OVIS, and BURST datasets, aiming to measure the performance of the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 395, + 565, + 482 + ], + "blocks": [ + { + "bbox": [ + 321, + 375, + 547, + 387 + ], + "lines": [ + { + "bbox": [ + 321, + 375, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 321, + 375, + 547, + 387 + ], + "type": "text", + "content": "Table 3. Comparison of ablation results of bridge strategy." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 395, + 565, + 482 + ], + "lines": [ + { + "bbox": [ + 315, + 395, + 565, + 482 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 565, + 482 + ], + "type": "table", + "html": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Baseline8.510.25.32.62.3
+ Spatio-Temporal Bridge15.018.010.23.15.4
+ Text-Visual Bridge19.923.314.25.37.9
+ LVT in Eq. 1020.724.915.25.58.2
", + "image_path": "ecd91e3f6ebb42c77f25f0323ea2e446e497c2c651fcfc6cdd6f122b2f1e72bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "text", + "content": "proposed SWBridge more comprehensively and rigorously. Effectiveness of Bridge Strategy. We first established a baseline variant that removed the Spatio-Temporal Bridge bridge and Text-Visual Bridge modules, while discarding the additional contrast loss " + }, + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VT}" + }, + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "text", + "content": " and retaining only the basic " + }, + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Seg}" + }, + { + "bbox": [ + 313, + 506, + 555, + 604 + ], + "type": "text", + "content": " loss function. In order to achieve instance association, we adopted the memory module proposed in OV2Seg to replace the original spatiotemporal bridge function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "The experimental results presented in Table 3 confirm the significant effectiveness of each component in the SWBridge method. Specifically, upon integrating the spatio-temporal bridge module into the baseline variant, we observed notable improvements in both " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "AP_{n}" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " metrics across all datasets. This outcome indicates that the spatio-temporal bridge module efficiently captures spatiotemporal correlations between video frames, thereby aiding the model in more accurately understanding and tracking" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12476" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 91, + 286, + 176 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 291, + 82 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 291, + 82 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 291, + 82 + ], + "type": "text", + "content": "Table 4. Impact of ablation results of Importance Weighting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 91, + 286, + 176 + ], + "lines": [ + { + "bbox": [ + 64, + 91, + 286, + 176 + ], + "spans": [ + { + "bbox": [ + 64, + 91, + 286, + 176 + ], + "type": "table", + "html": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
w/o Weight15.317.012.34.55.3
w/o Weight in §4.118.522.313.74.96.8
w/o Weight in §4.319.623.514.55.27.5
SWBridge(Our)20.724.915.25.58.2
", + "image_path": "993e0530489659d6d13e78cd27585c76de1c3919412b7a21e41355b4fc957793.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 209, + 286, + 293 + ], + "blocks": [ + { + "bbox": [ + 59, + 188, + 291, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 188, + 291, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 188, + 291, + 199 + ], + "type": "text", + "content": "Table 5. Impact of ablation results of Importance Weighting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 209, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 63, + 209, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 63, + 209, + 286, + 293 + ], + "type": "table", + "html": "
VariantLV-VISOVISBURST
APAPnAPAPAPAPn
Huberized18.822.413.84.76.4
Cosine18.622.213.94.86.5
Wasserstein19.523.314.45.17.0
Sliced Wasserstein20.724.915.25.58.2
", + "image_path": "4e5d229e4a30ec6b871b4e9f07e8f604f4e938c1d8d5bc1911854923e67b1d61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 434 + ], + "type": "text", + "content": "the dynamic changes of instances within videos. Furthermore, the model's performance was significantly enhanced once again after adding the text-visual bridge module on top of the spatio-temporal bridge module. By introducing textual information, the text-visual bridge provides the model with rich semantic cues, greatly enhancing its accuracy in recognizing and understanding instances. Additionally, " + }, + { + "bbox": [ + 55, + 316, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VT}" + }, + { + "bbox": [ + 55, + 316, + 295, + 434 + ], + "type": "text", + "content": " synergizes with other bridge strategy modules, collectively driving a comprehensive improvement in the model's performance and generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 567 + ], + "type": "text", + "content": "Impact of SW Weighting Mechanism. Table 4 shows that without a weighting mechanism in the spatio-temporal or text-visual bridges, the model struggles to utilize key information from instances with marked feature differences, limiting performance. However, incorporating the weighting mechanism into both bridges notably improves model performance. This indicates the mechanism helps the model focus on crucial instances with significant feature differences, assigning them greater weights during distance calculation. This adjustment enhances the model's accuracy and efficiency in processing complex data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 652 + ], + "type": "text", + "content": "Comparison of Distance Measurement. In Tab. 5, we compare distance metrics for SWBridge. Huberized distance is sensitive to large differences, Cosine similarity ignores feature magnitudes, and Wasserstein distance misses local details. Our SW distance comprehensively considers high-dimensional feature differences with importance weighting, making it robust for complex domain shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Sensitivity analysis of " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": ". In the bridge strategy, we ablated the number of sampling iterations to explore its relationship with model performance (Fig.2). Results show a nonlinear relationship: increasing iterations significantly improves " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "AP_{n}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " across all datasets. However, excessive" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 549, + 217 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 549, + 217 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 549, + 217 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 549, + 217 + ], + "type": "image", + "image_path": "27355c8d0edde4c5a96964f2667359330784395772273dd47b5c5044581ddc2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 224, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 247 + ], + "type": "text", + "content": "Figure 2. Ablation experiment on the number of random projection sampling " + }, + { + "bbox": [ + 313, + 224, + 553, + 247 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 224, + 553, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 293 + ], + "type": "text", + "content": "iterations introduce noise and computational burden, potentially causing overfitting or reducing training efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 388, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 388, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 388, + 308 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 495 + ], + "type": "text", + "content": "In this paper, we propose an innovative Sliced Wasserstein Bridging Learning Framework (SWbridge), aimed at tackling the challenging task of Open-Vocabulary Video Instance Segmentation. Inspired by metric learning, our framework utilizes the Sliced Wasserstein distance as a core tool to effectively bridge the spatial-temporal and text-visual domains involved in the task. Through domain alignment, we successfully map features from different domains into a unified metric space, preserving temporal consistency and learning intrinsic consistent features across modalities. This significantly improves the fusion of text and visual information, addressing one of the core challenges in Open-Vocabulary VIS. Looking ahead, we plan to explore better optimization methods for the Sliced Wasserstein distance to further enhance the performance of our framework." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 497, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 579 + ], + "type": "text", + "content": "Acknowledgements This work was supported in part by the China Postdoctoral Science Foundation (No. 2024M761805), the Postdoctoral Fellowship Program of CPSF (No. GZB20250390), the Qingdao Postdoctoral Innovation Project (No. QDBSH20240201015), the Major Science and Technology Innovation Projects in Shandong Province (No. 2024CXGC010108)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 592, + 373, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 592, + 373, + 604 + ], + "spans": [ + { + "bbox": [ + 315, + 592, + 373, + 604 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 612, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 319, + 612, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 612, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 319, + 612, + 553, + 655 + ], + "type": "text", + "content": "[1] Ali Athar, Jonathon Luiten, Paul Voigtlaender, Tarasha Khurana, Achal Dave, Bastian Leibe, and Deva Ramanan. Burst: A benchmark for unifying object recognition, segmentation and tracking in video. In WACV, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 319, + 657, + 553, + 689 + ], + "type": "text", + "content": "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 319, + 691, + 553, + 712 + ], + "type": "text", + "content": "[3] Nicolas Bonneel, Julien Rabin, Gabriel Peyre, and Hanspeter Pfister. Sliced and radon wasserstein barycenters of mea" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12477" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "sures. Journal of Mathematical Imaging and Vision, 51:22-45, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 62, + 95, + 295, + 128 + ], + "type": "text", + "content": "[4] Silvia Bucci, Mohammad Reza Loghmani, and Tatiana Tommasi. On the effectiveness of image rotation for open set domain adaptation. In ECCV, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 128, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 295, + 171 + ], + "type": "text", + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 295, + 216 + ], + "type": "text", + "content": "[6] Bowen Cheng, Anwesa Choudhuri, Ishan Misra, Alexander Kirillov, Rohit Girdhar, and Alexander G Schwing. Mask2former for video instance segmentation. arXiv preprint arXiv:2112.10764, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "text", + "content": "[7] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 250, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 294 + ], + "type": "text", + "content": "[8] Zesen Cheng, Kehan Li, Hao Li, Peng Jin, Chang Liu, Xiawu Zheng, Rongrong Ji, and Jie Chen. Instance brownian bridge as texts for open-vocabulary video instance segmentation. arXiv preprint arXiv:2401.09732, 2024. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 294, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 295, + 337 + ], + "type": "text", + "content": "[9] Penghui Du, Yu Wang, Yifan Sun, Luting Wang, Yue Liao, Gang Zhang, Errui Ding, Yan Wang, Jingdong Wang, and Si Liu. Lami-detr: Open-vocabulary detection with language model instruction. In ECCV, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 338, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 338, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 295, + 380 + ], + "type": "text", + "content": "[10] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 381, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 381, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 381, + 295, + 415 + ], + "type": "text", + "content": "[11] Hao Fang, Peng Wu, Yawei Li, Xinxin Zhang, and Xiankai Lu. Unified embedding alignment for open-vocabulary video instance segmentation. In ECCV, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 415, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 415, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 295, + 448 + ], + "type": "text", + "content": "[12] Damien Garreau, Rémi Lajugie, Sylvain Arlot, and Francis Bach. Metric learning for temporal sequence alignment. NeurIPS, 2014. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 449, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 295, + 480 + ], + "type": "text", + "content": "[13] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 57, + 481, + 295, + 525 + ], + "type": "text", + "content": "[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 568 + ], + "type": "text", + "content": "[15] Pinxue Guo, Tony Huang, Peiyang He, Xuefeng Liu, Tianjun Xiao, Zhaoyu Chen, and Wenqiang Zhang. Openvis: Open-vocabulary video instance segmentation. arXiv preprint arXiv:2305.16835, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[17] Fusheng Hao, Fengxiang He, Jun Cheng, Lei Wang, Jianzhong Cao, and Dacheng Tao. Collect and select: Semantic alignment metric learning for few-shot learning. In ICCV, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 680 + ], + "type": "text", + "content": "[18] Dat Huynh, Jason Kuen, Zhe Lin, Jieuxiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[19] Siyuan Li, Tobias Fischer, Lei Ke, Henghui Ding, Martin Danelljan, and Fisher Yu. Ovtrack: Open-vocabulary multiple object tracking. In CVPR, 2023. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 600 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[20] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[21] Jiyang Qi, Yan Gao, Yao Hu, Xinggang Wang, Xiaoyu Liu, Xiang Bai, Serge Belongie, Alan Yuille, Philip HS Torr, and Song Bai. Occluded video instance segmentation: A benchmark. International Journal of Computer Vision, 130(8): 2022-2039, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "text", + "content": "[22] Zheyun Qin, Xiankai Lu, Dongfang Liu, Xiushan Nie, Yi long Yin, Jianbing Shen, and Alexander C Loui. Reformulating graph kernels for self-supervised space-time correspondence learning. IEEE Transactions on Image Processing, 32: 6543-6557, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 218, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 272 + ], + "type": "text", + "content": "[23] Zheyun Qin, Xiankai Lu, Xiushan Nie, Dongfang Liu, Yi long Yin, and Wenguan Wang. Coarse-to-fine video instance segmentation with factorized conditional appearance flows. IEEE/CAA Journal of Automatica Sinica, 10(5):1192-1208, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 329, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 553, + 352 + ], + "type": "text", + "content": "[25] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 396 + ], + "type": "text", + "content": "[26] Haochen Wang, Cilin Yan, Shuai Wang, Xiaolong Jiang, Xu Tang, Yao Hu, Weidi Xie, and Efstratios Gavves. Towards open-vocabulary video instance segmentation. In ICCV, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 397, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 419 + ], + "type": "text", + "content": "[27] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 453 + ], + "type": "text", + "content": "[28] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. Dvis: Decoupled video instance segmentation framework. In ICCV, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 454, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 477 + ], + "type": "text", + "content": "[29] Ye Zheng, Jiahong Wu, Yongqiang Qin, Faen Zhang, and Li Cui. Zero-shot instance segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 510 + ], + "type": "text", + "content": "[30] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "text", + "content": "[31] Chaoyang Zhu and Long Chen. A survey on open-vocabulary detection and segmentation: Past, present, and future. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "text", + "content": "[32] Wenqi Zhu, Jiale Cao, Jin Xie, Shuangming Yang, and Yanwei Pang. Clip-vis: Adapting clip for open-vocabulary video instance segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 3, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12478" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_content_list.json b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2af59b02be75471c40aabbbd1993b159648d7c08 --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "SliderSpace: Decomposing the Visual Capabilities of Diffusion Models", + "text_level": 1, + "bbox": [ + 143, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Gandikota $^{1*}$ Zongze Wu $^{2}$ Richard Zhang $^{2}$ David Bau $^{1}$ Eli Shechtman $^{2}$ Nick Kolkin $^{2}$ Northeastern University $^{2}$ Adobe Research", + "bbox": [ + 99, + 178, + 903, + 220 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/920cee0ec944fef56241de91be0d5d59992603853c34d38ac3f394f2340b0e7b.jpg", + "image_caption": [ + "* The labels are generated by Claude-3.5 post-discovery of SliderSpace" + ], + "image_footnote": [], + "bbox": [ + 102, + 257, + 658, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3dac2e37bb3af9403ab5a3cfda07d1896900b86c7b4039104f553eb968494a75.jpg", + "image_caption": [ + "Figure 1. Given a prompt, SliderSpace identifies the principal directions of the visual capabilities of a diffusion model by decomposing the image distribution over the prompt. By manipulating these directions as sliders, users can control and combine them to explore the creative possibilities of the model. We visualize the top directions discovered by SliderSpace within Flux Schnell [4] for the concept \"Toy\"." + ], + "image_footnote": [], + "bbox": [ + 673, + 258, + 900, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 532, + 326, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at sliderspace.baulab.info", + "bbox": [ + 89, + 565, + 483, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 532, + 643, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models are capable of generating remarkable visual variations from a single prompt through different random initializations. However, this vast creative potential remains largely opaque to users—while we can generate diverse images, we lack understanding of the underlying structure of these variations. This presents a fundamental challenge: how can we discover and expose the latent visual capabilities encoded within these models?", + "bbox": [ + 509, + 558, + 906, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The challenge touches on a key limitation in how we interact with diffusion models today. Current control methods require users to explicitly specify their desired edits in advance through prompts [16], reference images [8, 24, 29, 41, 42, 52], or attribute vectors [20, 21, 30, 32, 46, 50]. That contrasts sharply with natural human creative workflows, where artists dynamically explore creative ideas and jointly refine them toward meaningful artistic outcomes [23]. The need for pre-specified controls creates a barrier between users and the full creative potential of these models.", + "bbox": [ + 509, + 680, + 908, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interestingly, earlier generative models like GANs [6, 17, 28] naturally developed more interpretable internal structures. Their compact latent spaces often exhibited", + "bbox": [ + 511, + 830, + 906, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence to gandikota.ro@northeastern.edu", + "bbox": [ + 537, + 887, + 877, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15994", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emergent disentanglement [19, 35, 45, 47], enabling continuous and compositional control over generated images. Users could explore these spaces to discover interesting variations that would be difficult to describe in words [47], then combine them to achieve their creative goals [18].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models have largely superseded GANs in conditional image synthesis [11], achieving greater diversity through much higher-dimensional latents. And yet an understanding of the underlying structure of these larger latent spaces has remained elusive. In this work, we ask a fundamental question: Can we automatically discover the visual structure within a diffusion model's knowledge of a concept? Rather than requiring user-specified controls, we aim to decompose the model's internal representations into expressive directions that users can explore and combine.", + "bbox": [ + 88, + 167, + 480, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these needs, we present SliderSpace, a framework that brings systematic explorability to diffusion models. Given just a text prompt, SliderSpace discovers a canonical set of meaningful, diverse, and controllable directions within the model's knowledge of that concept. Each direction is implemented as a low-rank adapter [24] that can be scaled and composed with others, allowing users to explore and smoothly combine different aspects of variation, as shown in Figure 1.", + "bbox": [ + 89, + 320, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We ground SliderSpace discovery in three key requirements for meaningful decomposition of a diffusion model's visual manifold:", + "bbox": [ + 89, + 458, + 482, + 502 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Unsupervised Discovery: The decomposition process should emerge from the intrinsic structure of the model's learned representation, rather than being guided by predefined attributes. This ensures we capture the true topology of the model's knowledge space rather than projecting our assumptions onto it.", + "2. Semantic Orthogonality: Each discovered control must represent a distinct semantic direction. This is enforced in a semantic feature space, like CLIP, where every slider has an orthogonal effect in embeddings. This prevents discovering multiple controls that create similar semantic effects [15, 33, 47], making the system more efficient.", + "3. Distribution Consistency: Directions must induce consistent transformations across both random seeds and prompt variations." + ], + "bbox": [ + 89, + 505, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These requirements naturally lead to our proposed framework, which we formalize in Section 4. As we show in our experiments, SliderSpace is architecture-agnostic, working with both conventional U-Net based models like Stable Diffusion [34, 37, 40, 43, 51] and recent transformer-based architectures like Flux [4].", + "bbox": [ + 89, + 733, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the expressiveness of SliderSpace through three applications: First, we show how SliderSpace can decompose high-level concepts into diverse and expressive components, revealing the natural axes of variation in the model's understanding. Second, we explore artistic", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "style variation, where SliderSpace discovers directions that match or exceed the diversity of manually curated artist lists while being judged more useful by human evaluators. Finally, we show how SliderSpace can help reverse the mode collapse commonly observed in distilled diffusion models, restoring diversity while maintaining generation speed.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond providing practical creative control, SliderSpace opens new avenues for understanding and utilizing the latent capabilities of diffusion models. By mapping these models' visual potential into intuitive, composable directions, we take a step toward making their creative possibilities more accessible and interpretable to users.", + "bbox": [ + 511, + 181, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 286, + 661, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent text-to-image diffusion models have demonstrated remarkable capabilities in generating diverse visual concepts [40]. While newer foundation models enhance text-image alignment through LLM-generated captions [13, 26], the fundamental challenge remains: text-conditioned generation is inherently under-determined, with multiple distinct outputs potentially satisfying the same prompt, making precise control challenging.", + "bbox": [ + 511, + 311, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior work has explored various approaches to enhance generation control. One direction introduces additional conditioning modalities: spatial signals via adapters (ControlNet [52]), attention-based regional control [8], and image-based conditioning for identity preservation [30, 46, 50] or style transfer via attention manipulation [21]. Another line of research focuses on refining existing images through disentangled text-driven editing [5, 7, 10, 32, 48, 49], where specific attributes are modified while preserving others, often using spatial masks. However, these text-driven approaches inherit the same under-determination challenges as the base models.", + "bbox": [ + 511, + 431, + 905, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Notably, disentangled control has been more naturally achieved in GANs [17], particularly through StyleGAN's low-dimensional latent space [28], which exhibits emergent disentanglement even at scale [27]. This has enabled powerful latent space and image editing capabilities [1, 2, 47]. While diffusion models offer superior generation quality and diversity, they lack two key advantages of GANs' latent space: continuous, compositional edits and emergent disentanglement. This limitation prevents users from making serendipitous discoveries about visual variations captured in the training data, instead constraining them to variations they can explicitly describe through prompts.", + "bbox": [ + 511, + 614, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works have explored different approaches to discovering interpretable directions in diffusion models. The weights2weights method [12] learns a manifold of personalized model weights by fine-tuning individual LoRA adapters for each identity and applying PCA to discover a weight space that enables editing. While effective, this requires training separate models per instance. NoiseCLR [9]", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15995", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learns text embeddings through contrastive learning on a data distribution, but the discovered directions can be arbitrary and lack semantic interpretability. Similarly, Liu et al. [31] propose unsupervised decomposition of images into compositional concepts by learning multiple embeddings, but their approach often yields redundant or non-semantic directions. In contrast, our work directly learns a small set of semantically grounded sliders that decompose the distribution into interpretable and composable directions, enabling infinite creative variations through systematic exploration of the visual manifold.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Sliders [16] addressed continuous control by leveraging LoRA adaptors [24] to learn user-defined attributes. Our work complements this by tackling emergent disentanglement through self-supervised decomposition of the model's inherent variations into composable control dimensions, enabling systematic exploration of the model's creative capabilities.", + "bbox": [ + 89, + 257, + 482, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 377, + 218, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 401, + 313, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State-of-the-art text-to-image diffusion models [4, 34, 38, 39] often belong to the class of latent diffusion. Unlike traditional diffusion models that operate in pixel space, latent diffusion models work in a compressed latent space, offering significant computational advantages [40]. The diffusion modelling can be formalized as follows:", + "bbox": [ + 89, + 424, + 482, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathbf{x}_0$ be an initial image and $\\mathbf{x}_T$ be pure Gaussian noise. The forward diffusion process gradually adds noise to the image. The generative process aims to reverse this diffusion, starting from $x_{T}$ and progressively denoising to reconstruct $\\mathbf{x}_0$ . At a timestep $t$ the model takes $\\mathbf{x}_t$ as input and predicts a noise $\\epsilon_t$ such that the next step $\\mathbf{x}_{t - 1}$ :", + "bbox": [ + 89, + 515, + 483, + 606 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}}{\\sqrt {\\alpha_ {t}}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 617, + 480, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can estimate the final image $\\tilde{\\mathbf{x}}_{0,t}$ by taking the same direction $\\epsilon_t$ for remaining diffusion steps. This can be achieved by recursively applying the denoising from Equation 1 with the same direction. We label this \"Final Image Extrapolation\":", + "bbox": [ + 89, + 662, + 482, + 736 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {0, t} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 750, + 480, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This enables us to visualize the final image $\\tilde{\\mathbf{x}}_{0,t}$ that the diffusion model is planning of at each timestep $t$ without actually running denoising forward passes through all timesteps.", + "bbox": [ + 89, + 792, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LoRA: Low Rank Adaptors", + "text_level": 1, + "bbox": [ + 89, + 848, + 341, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-rank adaptor (LoRA) [24] are a class of light-weight adaptors that are attachable to the weights of the model.", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pre-trained model layer with weights $W_0 \\in \\mathbb{R}^{d \\times k}$ , LoRA decomposes the weight update $\\Delta W$ as:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta W = B A, \\quad B \\in \\mathbb {R} ^ {d \\times r}, A \\in \\mathbb {R} ^ {r \\times k}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 138, + 903, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r \\ll \\min(d, k)$ is a small rank that constrains the update to a low-dimensional subspace. This decomposition allows for efficient parameter updates and has shown success in various downstream tasks. However, the application of LoRA to unsupervised discovery of semantic directions in diffusion model weight space remains unexplored.", + "bbox": [ + 511, + 169, + 906, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 277, + 604, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present SliderSpace, a framework for decomposing a diffusion model's visual capabilities into semantically orthogonal control dimensions (Fig. 1). Given a pre-trained text-to-image diffusion model $\\theta$ and a prompt $c$ , our goal is to discover $n$ independent directions that capture the principal modes of variation in the model's learned distribution.", + "bbox": [ + 511, + 304, + 905, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 409, + 718, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{M}_{\\theta}(c)$ denote the manifold of possible images that model $\\theta$ can generate for prompt $c$ . We aim to identify a set of controllable directions $\\{\\mathcal{T}_i\\}_{i=1}^n$ that: (1) span the major modes of variation in $\\mathcal{M}_{\\theta}(c)$ , (2) maintain semantic consistency across different initializations, and (3) are mutually orthogonal in semantic space.", + "bbox": [ + 511, + 431, + 905, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on recent advances in model adaptation [16], we formulate each control dimension as a LoRA adaptor [24], updating $\\mathcal{T}_i$ , where $i \\in \\{1, \\dots, n\\}$ . These lightweight adapters introduce targeted modifications to the model's cross-attention layers, enabling precise control over specific generative attributes.", + "bbox": [ + 511, + 523, + 905, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. SliderSpace: Unsupervised Visual Discovery", + "text_level": 1, + "bbox": [ + 511, + 628, + 888, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SliderSpace discovery process consists of three key steps as depicted in Figure 2:", + "bbox": [ + 511, + 652, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distribution Sampling First, we generate a diverse set of samples $\\{x_{j}\\}_{j = 1}^{m}$ from $\\mathcal{M}_{\\theta}(c)$ by varying the random seed (for stability, $m\\approx 5000$ ). For each sample, we extract the estimated final image $\\tilde{\\mathbf{x}}_{0,t}$ at each timestep $t$ using Eq. 2.", + "bbox": [ + 511, + 708, + 906, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic Decomposition We map each sample to a semantic embedding space, like CLIP, $\\phi (\\tilde{\\mathbf{x}}_{0,t})$ and compute the principal components $V = \\{v_{i}\\}_{i = 1}^{n}$ of the resulting distribution. These components represent orthogonal directions of maximal variation in semantic space:", + "bbox": [ + 511, + 794, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV = \\operatorname {P C A} \\left(\\left\\{\\phi \\left(\\tilde {x} _ {0, t}\\right) \\right\\} _ {j, t}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 883, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15996", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/44f0c7fc1eed88d6179270d33846652e8a5f5ecd3507b6edd6b63162736fa5e7.jpg", + "image_caption": [ + "Figure 2. Given a prompt, SliderSpace first generates images and extracts the CLIP features. We then compute the spectral decomposition of the CLIP features and align the each slider with extracted principle components. Each slider is therefore trained to represent a unique semantic direction that are relevant in the diffusion model's knowledge of the prompt." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Slider Training For each principal direction $v_{i}$ , we train a corresponding adapter $\\mathcal{T}_i$ to induce transformations that align with $v_{i}$ in semantic CLIP space. The training objective for each slider is:", + "bbox": [ + 88, + 378, + 486, + 440 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s l i d e r s p a c e}} = \\sum_ {i = 1} ^ {n} 1 - \\cos \\left(\\Delta \\phi_ {i}, v _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 455, + 482, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta \\phi_{i} = \\phi (\\tilde{\\mathbf{x}}_{0,t}^{i}) - \\phi (\\tilde{\\mathbf{x}}_{0,t})$ represents the transformation induced by slider $i$ in embedding space and $\\tilde{x}_{0,t}^{i}$ is the estimated final image with the slider $\\mathcal{T}_i$ attached. This ensures that each slider's effect is semantically aligned with its corresponding principal direction while maintaining orthogonality with other sliders, building on established work showing that embedding differences encode semantic relationships [15, 33].", + "bbox": [ + 88, + 516, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our formulation satisfies the three key requirements outlined in Section 1. First, unsupervised discovery is achieved by deriving directions directly from the model's intrinsic variation through PCA in a semantic embedding space, without imposing predefined attributes or external supervision. Second, distribution consistency is enforced by our training objective $\\mathcal{L}_{\\mathrm{sliderspace}}$ , which ensures each slider's transformation maintains consistent direction in embedding space across different seeds and timesteps. Finally, semantic orthogonality is guaranteed through the PCA-based initialization of directions as each principal components are mutually orthogonal, ensuring each slider captures a distinct mode of variation. Together, these components enable the discovery of interpretable and reliable control dimensions that effectively decompose the model's learned distribution. In majority of this paper, we use CLIP [36] as our primary semantic encoder.", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Interpretability & Control", + "text_level": 1, + "bbox": [ + 511, + 377, + 754, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SliderSpace provides a dual contribution: it serves as both a framework for discovering expressive dimensions of control and as a mechanism for decomposing the model's learned concept space into semantically meaningful components. Each adapter functions as a \"slider\" controlling a specific attribute of the generated image, enabling fine-grained manipulation of the output while maintaining semantics.", + "bbox": [ + 511, + 400, + 906, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The resulting set of adapters provides valuable insights into the model's conceptual understanding, revealing nuanced semantic relationships that may not be immediately apparent from the text prompt alone.", + "bbox": [ + 511, + 507, + 905, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, the low-rank structure of our adapters ensures computational efficiency and minimal memory overhead, making them particularly suitable for real-time interactive applications. This enables users to explore the concept space dynamically by modulating the influence of each adapter, facilitating intuitive creative control over the image generation process while maintaining the underlying semantic integrity of the original prompt.", + "bbox": [ + 511, + 569, + 906, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 707, + 645, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct our main experiments to evaluate SliderSpace using SDXL-DMD [51], a 4-step distilled diffusion model. Our implementation requires less than 24GB VRAM and can discover 64 semantic directions in under 2 hrs on a single A100 GPU. When concepts exhibit severe mode collapse, the discovery process can be enhanced by generating data using undistilled base models or LLM-expanded prompts for increased sample diversity. We demonstrate SliderSpace's generalization to SDXL [34], SDXL-Turbo [43], and transformer-based FLUX Schnell [4] in appendix. Our analysis focuses on three key applications:", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15997", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cfc3a3fc362612dc20df1e7364ec5bb07150c7ef2d43349f054c0a37e817dde8.jpg", + "image_caption": [ + "Figure 3. SliderSpace decomposes the visual variation of diffusion model's knowledge corresponding to a concept. These directions can be perceived as interpretable directions of the model's hierarchical knowledge. We show the decomposed slider direction for a concept using SliderSpace and the corresponding labels generated by Claude 3.5 Sonnet." + ], + "image_footnote": [], + "bbox": [ + 91, + 84, + 903, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "concept decomposition, art styles exploration, and diversity enhancement in distilled models.", + "bbox": [ + 89, + 407, + 480, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Concept Decomposition", + "text_level": 1, + "bbox": [ + 89, + 452, + 312, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first demonstrate how SliderSpace can serve as an exploratory tool by decomposing high-level concepts into semantic directions that align with the diffusion model's internal representations. Summarizing and exposing the dominant variations a model is capable of for a particular prompt.", + "bbox": [ + 89, + 476, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a concept prompt (e.g., \"picture of a monster\"), we discover SliderSpace directions. Figure 3 shows discovered sliders for \"Monster\", and \"car\" concepts. We label the sliders using Claude 3.5 Sonnet [3] by showing multiple image pairs showing the effect of each slider and prompting to identify the semantic transformation being applied. Through these discovered directions, we demonstrate the manipulation of individual attributes. As these directions are intended to capture the model's visual possibilities for a prompt, we naturally ask the question, \"How much variation is enabled by these directions?\" To quantitatively evaluate the diversity enabled by SliderSpace against base model, we compute DreamSim [14] distance to measure inter-image variation across 2500 generated samples per concept (Figure 4). For SliderSpace-augmented generation, we generate the images by randomly activating a sparse subset of 3 sliders (out of 32 discovered directions) for each generation. Our analysis reveals that SliderSpace-generated samples exhibit significantly higher inter-image diversity compared to the baseline model outputs. We measure the CLIP-Score [22] between the input prompts and the generated images to measure text-alignment with the prompt. We find that they have similar CLIP Scores to the images gen", + "bbox": [ + 89, + 553, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c5c8d0d75537cb3083682bf3ff2b9a4efc6e6b08683bc2c617abcc402acbc72f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method vs.User Study (Win Rate %)
“Diverse”“Useful”“Creative”
SDXL-DMD72.466.068.1
LLM + SDXL-DMD62.562.562.5
SDXL65.361.259.2
", + "bbox": [ + 514, + 406, + 906, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Users perceive SliderSpace generated images to be more diverse, useful, and interesting. We show win-rate percentages of SliderSpace samples against baselines.", + "bbox": [ + 511, + 488, + 906, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "erated by the base model. This suggests that SliderSpace effectively expands the achievable variation in the model's knowledge, while maintaining semantic consistency. To validate our findings through human perception, we conducted pairwise comparisons of image grids generated by SliderSpace versus baseline methods. As shown in Table 1, users consistently preferred SliderSpace outputs across diversity, utility, and creative potential.", + "bbox": [ + 511, + 559, + 906, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Art Styles Exploration", + "text_level": 1, + "bbox": [ + 511, + 694, + 723, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also evaluate to what degree SliderSpace can expose \"all\" of the art styles that the diffusion model has learned from its training data. As a proxy for \"all\", we use a diverse set of art styles, manually discovered and documented by ParrotZone [25]. We explore the visual artistic space by decomposing the prompt \"artwork in the style of a famous artist\" into 64 directions. This process enables us to discover the SliderSpace of art and create a comprehensive dictionary of discoverable art styles in a diffusion model.", + "bbox": [ + 511, + 717, + 906, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compare how this measures up against supervised methods like Concept Slider [16], we train 64 manually curated concept sliders using LLM generated training", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15998", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76d2d6e63c6f42822c0d0e4e0d28f217dc37b7fb50991f881a8a975aa8bd0e0b.jpg", + "image_caption": [ + "Figure 4. (left) SliderSpace generates diverse variations of a concept, as measured by DreamSim [14] distance across generated samples (higher is better). (right) The method maintains similar text-to-image alignment, as measured by CLIP Scores [22] (lower is better)." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 450, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1632b0ce9a286263773b7ce8a092ed7bfd9ef2bfd506bafbdf3ccfb7c9f7df8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 88, + 893, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompts. We compute FID scores against the generated samples conditioned on actual artist names from ParrotZone [25] dataset. The dataset contains 4388 artists mimicked by SDXL, are discovered through exhaustive manual search. We also establish two other baselines. First, we use GPT/Claude to generate 4388 different artistic style names (e.g., \"cubism\", \"Warhol\"). Second, we use a generic prompt like \"[subject] in the style of a famous artist\". Using our discovered SliderSpace, we randomly sample a sparse set of 3 sliders to generate an equivalent number of images. Specifically, for each baseline, we match the provided samples of [25] by creating two images per style for \"building landscape\" and \"character portraits\" each (total of four images per style) to ensure same semantic structure across the datasets and that FID measures the artistic spread. Figure 5 shows random, non-cherry picked images from all the methods and their corresponding FID scores. We find that SliderSpace generates a distribution significantly more closely matches the manually curated artist names than the baselines, including supervised methods like Concept Sliders. Figure 6 shows qualitative comparison of art styles discovered by SliderSpace and styles discovered manually [25].", + "bbox": [ + 88, + 277, + 485, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the practical utility of our \"artistic\" SliderSpace, we conduct a user study examining both usefulness and diversity of generated samples (Table 2). We conduct comparative study with 2 grids of 9 images (total of 1000 pairs), one with slider generated images and other with baselines. Users find the SliderSpace-generated images are far more diverse and useful than the baseline methods. We also find that users would prefer using SliderSpace images over images prompted with the real artist names [25], while finding these both to be equally diverse. We provide user-study details and qualitative examples in Appendix.", + "bbox": [ + 89, + 625, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Diversity Enhancement", + "text_level": 1, + "bbox": [ + 89, + 801, + 307, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we explore SliderSpace's ability to addressing mode collapse in distilled models. Instead of discovering SliderSpace for narrow distributions, we train it on a larger spread of 8000 randomly selected COCO-30k prompts and enhance them with LLM. We use DMD model for gener", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c0653ba4f60d849fe0066aedf533e042ce056e7fa1b909c4f7ec42917cbafb3a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUser study (win rate %)
“Diverse”“Useful”
vs. Real Artist Prompts [25]5463
vs. LLM Prompts7367
vs. Generic Prompts8788
", + "bbox": [ + 527, + 275, + 893, + 347 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/745b04bd8702771bfe81143ae10479a54c7a85a946a896f07bd8e029a9aaf6bf.jpg", + "table_caption": [ + "Table 2. Pairwise comparison user study (win rates in %) reveals that SliderSpace-extracted artistic styles achieve comparable diversity to manually curated artist [25], while being significantly preferred for creative utilization." + ], + "table_footnote": [], + "table_body": "
MethodFID-30k (↓)CLIP (↑)
Real-30.14
SDXL11.7229.41
SDXL-DMD15.5228.92
DMD-SliderSpace12.1229.13
", + "bbox": [ + 545, + 430, + 874, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. SliderSpace trained on larger range of knowledge can improve the diversity of the distilled models (FID) while having a good text-image alignment (CLIP).", + "bbox": [ + 511, + 518, + 908, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ating training dataset since we wish to undo mode collapse of DMD models. We discover a generic SliderSpace of 64 sliders that captures the generally applicable visual variations within SDXL-DMD. Figure 7 showcases qualitative examples of two distributions for the prompts \"Car driving through a forest\" and \"Picture of a person\". We demonstrate how generating images by randomly sampling from these SliderSpace directions effectively increases output diversity. This demonstrates that the mode collapse in distilled model can be reversed by discovering and exploring the visual structure. We show more examples in appendix.", + "bbox": [ + 511, + 590, + 906, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, we evaluate this improvement by measuring FID and CLIP scores on all COCO-30k prompts before and after applying SliderSpace to SDXL-DMD. We find that DMD-SliderSpace improves FID from the distilled version, almost matching the FID of undistilled SDXL.", + "bbox": [ + 511, + 758, + 908, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Slider Transferability", + "text_level": 1, + "bbox": [ + 511, + 845, + 718, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using FaceNet [44] as the semantic encoder, we train SliderSpace on the concept \"person\" and discover interpretable", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15999", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d80110c9f12f3f9954d4b3a4ec9f41c4444d3671cf8287be8d0006894e6e0897.jpg", + "image_caption": [ + "Real Artists", + "(a)", + "Figure 5. SliderSpace demonstrates broader artistic style coverage, as evidenced by the lower FID scores compared to both supervised Concept Sliders. Comparison of artistic style diversity using FID scores against reference distribution (a) derived from the complete artist dataset [25]. We compare against outputs from (b) generic art prompts (b), (c) LLM-generated art prompts, and (d) Concept Sliders." + ], + "image_footnote": [], + "bbox": [ + 93, + 102, + 251, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1a168bb77bc4386aed65278bcc123c9fbf4647d0930c6770081aa810aa80c8c.jpg", + "image_caption": [ + "Generic Prompts (FID: 76.08)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 256, + 101, + 415, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/047d5a705de86afb97ece042b464f4fbd570cb44cb28e556600f9b06a417a557.jpg", + "image_caption": [ + "LLM Prompts (FID: 38.84)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 419, + 101, + 578, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/978c80621cd18437b6128771157752b3c98c675a3bc3485b4d0f4bb979c6d1e3.jpg", + "image_caption": [ + "Concept Sliders (FID: 32.86)", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 583, + 101, + 741, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88c7521209cf47f21218615b7bc43de90d267b42c4a618ed474d825c41f98143.jpg", + "image_caption": [ + "SliderSpace (FID: 19.12)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 746, + 101, + 903, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/498ef8e9a906b5bd57e11f713f202600ca486ca0643c891bfb42945ec28d06e1.jpg", + "image_caption": [ + "Elenore Abbott", + "Wayne Barlowe" + ], + "image_footnote": [], + "bbox": [ + 93, + 337, + 287, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ad2537634c15198edf91fe3acc1d43bd4f7c7c0a2df506033b8e3428e16f351.jpg", + "image_caption": [ + "Figure 6. Comparison of artistic styles discovered in SDXL-DMD: (Left) Representative samples from artist-specific prompts manually curated by the Parrotzone community [25] through extensive exploration. (Right) Automatically discovered artistic directions using SliderSpace, which captures diverse and semantically meaningful variations without requiring explicit artist references." + ], + "image_footnote": [], + "bbox": [ + 94, + 449, + 287, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4304f4149c1af9f75e2daa39d59232faeb0ecaa09490372d2460d93d38b201ab.jpg", + "image_caption": [ + "Manually Discovered Styles by ParrotZone", + "Miki Asai" + ], + "image_footnote": [], + "bbox": [ + 295, + 335, + 488, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95e0fc2f3b6c5cd7e1a4c1128b35b2674f2294f20f2c003ce470933f73d2f7d7.jpg", + "image_caption": [ + "Edward Gorey" + ], + "image_footnote": [], + "bbox": [ + 295, + 449, + 488, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc4a9b33d232f45b5091c94c841b9bda05bcf6f252491808d13d222046d9fffe.jpg", + "image_caption": [ + "Automatically Discovered Styles by SliderSpace", + "Direction 1" + ], + "image_footnote": [], + "bbox": [ + 506, + 335, + 702, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bbce1fb0c35f38f87b891054cfb07c24799191f60e9f8bfda8b403a8ad9787a.jpg", + "image_caption": [ + "Direction 25" + ], + "image_footnote": [], + "bbox": [ + 506, + 448, + 702, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a81b70628fccbdbca3d7f116f7373cc9b6683b1cc913c08939e01b3b5c480d85.jpg", + "image_caption": [ + "Direction 23", + "Direction 31" + ], + "image_footnote": [], + "bbox": [ + 709, + 335, + 903, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e1ba86a094f362ebb0f7069db2c523607f1554b9b1fa9726b41003df294a020.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 449, + 903, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "directions controlling attributes like appearance and age (Figure 8). These directions not only transfer effectively to related concepts like \"police\" and \"athlete\" but also generalize surprisingly well to out-of-domain concepts like \"dog\", suggesting that SliderSpace captures fundamental visual transformations in the model's knowledge space.", + "bbox": [ + 89, + 628, + 482, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additional experiments analyzing hyperparameter choices (App. B.2), alternative semantic embeddings (App. B.1), and ablation studies (App. E) are in Appendix.", + "bbox": [ + 89, + 720, + 483, + 766 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 89, + 782, + 212, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method can work with several semantic embeddings - each encoder might introduce different inherent biases present in its training data. While these embeddings enable semantic consistency, they may not capture certain culturally-specific or nuanced concepts. This highlights the need for application specific semantic embedding choice.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The current discovery process requires $\\approx 65$ mins on A100. We note that our method trains 8 times faster than Concept Sliders for the same number of sliders. To further optimize, it opens avenues for future research into training time optimizations. For art style discovery, the discovered directions are not one-to-one matched with the real artists, this can be limiting based on users' application. Further work can address discovery that nudges the directions to be aligned with real artists.", + "bbox": [ + 511, + 628, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 782, + 633, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SliderSpace is a simple framework that automatically decomposes diffusion models' capabilities into semantically meaningful and controllable directions. By leveraging spectral decomposition in semantic space combined with low-rank adaptation, our method enables systematic exploration of a model's latent creative space without requiring manual", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16000", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb754555c6b0f0f77b2abd71066bc942065255aa3a1317903cdabee3daeec71b.jpg", + "image_caption": [ + "Figure 7. SliderSpace can also decompose general broad visual variation of diffusion model's and can be used to overcome model collapse in distilled models. We randomly sample a sparse set of sliders and generate the sample showing higher variation than base distilled model." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fba636fa30ffefdd9cf8b55832fc124f5f59891838834dc2629b516e7cf6cb2.jpg", + "image_caption": [ + "Figure 8. SliderSpace directions for the \"person\" concept successfully generalize to related \"police\" and \"athlete\" concepts. They also transfer to out-of-domain concepts like \"dog\"" + ], + "image_footnote": [], + "bbox": [ + 114, + 343, + 885, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "attribute specification. Through extensive experiments, we demonstrated SliderSpace's effectiveness across three key applications. First, our concept decomposition revealed interpretable variations within the model's knowledge representation, enabling fine-grained control while maintaining semantic consistency. Second, our exploration of artistic capabilities showed that SliderSpace can discover directions matching or exceeding the diversity of manually curated artist lists, while being rated more useful by human evaluators. Finally, we demonstrate how SliderSpace can help", + "bbox": [ + 89, + 748, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "address mode collapse in distilled diffusion models, restoring diversity while preserving computational efficiency.", + "bbox": [ + 511, + 748, + 906, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ability of SliderSpace to uncover interpretable directions suggests that diffusion models may develop structured internal representations of visual concepts during training, without explicit supervision. By mapping these models' vast creative potential into intuitive, composable directions, our work takes a step toward making their capabilities more transparent and accessible.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16001", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was done by RG at Adobe during their internship. RG and DB are supported by Open Philanthropy and NSF grant #2403304.", + "bbox": [ + 89, + 114, + 482, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 186, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432-4441, 2019. 2", + "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2", + "[3] Anthropic. Introducing claude 3.5 sonnet, 2024. 5, 2", + "[4] BlackForestLabs. Announcing state-of-the-art flux.1 dev and Schnell models, 2024. 1, 2, 3, 4, 17", + "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2", + "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv e-prints, pages arXiv-1809, 2018. 1", + "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[8] Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompting for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024. 1, 2", + "[9] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24209-24218, 2024. 2, 3", + "[10] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 2", + "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[12] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 2", + "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + ], + "bbox": [ + 93, + 200, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Forty-first International Conference on Machine Learning, 2024. 2", + "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6", + "[15] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4", + "[16] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In European Conference on Computer Vision, pages 172-188. Springer, 2024. 1, 3, 5", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2", + "[18] Imke Grabe, Miguel González-Duque, Sebastian Risi, and Jichen Zhu. Towards a framework for human-ai interaction patterns in co-creative gan applications. In Joint Proceedings of the ACM IUI Workshops, 2022. 2", + "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2", + "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1", + "[21] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2", + "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6", + "[23] Oliver Hoffmann. On modeling human-computer cocreativity. In Knowledge, Information and Creativity Support Systems: Selected Papers from KICSS'2014-9th International Conference, held in Limassol, Cyprus, on November 6-8, 2014, pages 37-48. Springer, 2016. 1", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2, 3", + "[25] Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies, 2022. 5, 6, 7", + "[26] et al James Betker. Improving image generation with better captions. OpenAI Reports, 2023. 2", + "[27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16002", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2", + "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2", + "[29] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3", + "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2", + "[31] Nan Liu, Yilun Du, Shuang Li, Joshua B Tenenbaum, and Antonio Torralba. Unsupervised compositional concepts discovery with text-to-image generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085-2095, 2023. 3", + "[32] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2", + "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 2, 4", + "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 4, 8", + "[35] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[37] Robin Rombach. Stable diffusion 2.0 release, 2022. 2", + "[38] Robin Rombach and Patrick Esser. Stable diffusion v1-4 model card, 2022. 3", + "[39] Robin Rombach and Patrick Esser. Stable diffusion v2 model card, 2022. 3", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and BjA¶rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3", + "[42] Simo Ryu. Cloneofsimo/lora: Using low-rank adaptation to quickly fine-tune diffusion models.s. GitHub, 2023. 1", + "[43] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 2, 4, 16", + "[44] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6", + "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2", + "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 1, 2", + "[47] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 2", + "[48] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024. 2", + "[49] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2", + "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2", + "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15", + "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 756 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16003", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_model.json b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_model.json new file mode 100644 index 0000000000000000000000000000000000000000..633337853df47d9fd86a63c5747b2407b33c7e3b --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.853, + 0.154 + ], + "angle": 0, + "content": "SliderSpace: Decomposing the Visual Capabilities of Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.905, + 0.222 + ], + "angle": 0, + "content": "Rohit Gandikota\\(^{1*}\\) Zongze Wu\\(^{2}\\) Richard Zhang\\(^{2}\\) David Bau\\(^{1}\\) Eli Shechtman\\(^{2}\\) Nick Kolkin\\(^{2}\\) Northeastern University\\(^{2}\\) Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.258, + 0.659, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.259, + 0.901, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.451, + 0.653, + 0.461 + ], + "angle": 0, + "content": "* The labels are generated by Claude-3.5 post-discovery of SliderSpace" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.476, + 0.908, + 0.52 + ], + "angle": 0, + "content": "Figure 1. Given a prompt, SliderSpace identifies the principal directions of the visual capabilities of a diffusion model by decomposing the image distribution over the prompt. By manipulating these directions as sliders, users can control and combine them to explore the creative possibilities of the model. We visualize the top directions discovered by SliderSpace within Flux Schnell [4] for the concept \"Toy\"." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.534, + 0.327, + 0.549 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.485, + 0.884 + ], + "angle": 0, + "content": "We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at sliderspace.baulab.info" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.534, + 0.645, + 0.549 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.559, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Text-to-image diffusion models are capable of generating remarkable visual variations from a single prompt through different random initializations. However, this vast creative potential remains largely opaque to users—while we can generate diverse images, we lack understanding of the underlying structure of these variations. This presents a fundamental challenge: how can we discover and expose the latent visual capabilities encoded within these models?" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.681, + 0.909, + 0.831 + ], + "angle": 0, + "content": "The challenge touches on a key limitation in how we interact with diffusion models today. Current control methods require users to explicitly specify their desired edits in advance through prompts [16], reference images [8, 24, 29, 41, 42, 52], or attribute vectors [20, 21, 30, 32, 46, 50]. That contrasts sharply with natural human creative workflows, where artists dynamically explore creative ideas and jointly refine them toward meaningful artistic outcomes [23]. The need for pre-specified controls creates a barrier between users and the full creative potential of these models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.832, + 0.908, + 0.879 + ], + "angle": 0, + "content": "Interestingly, earlier generative models like GANs [6, 17, 28] naturally developed more interpretable internal structures. Their compact latent spaces often exhibited" + }, + { + "type": "page_footnote", + "bbox": [ + 0.538, + 0.888, + 0.878, + 0.9 + ], + "angle": 0, + "content": "*Correspondence to gandikota.ro@northeastern.edu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15994" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "emergent disentanglement [19, 35, 45, 47], enabling continuous and compositional control over generated images. Users could explore these spaces to discover interesting variations that would be difficult to describe in words [47], then combine them to achieve their creative goals [18]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.169, + 0.482, + 0.319 + ], + "angle": 0, + "content": "Diffusion models have largely superseded GANs in conditional image synthesis [11], achieving greater diversity through much higher-dimensional latents. And yet an understanding of the underlying structure of these larger latent spaces has remained elusive. In this work, we ask a fundamental question: Can we automatically discover the visual structure within a diffusion model's knowledge of a concept? Rather than requiring user-specified controls, we aim to decompose the model's internal representations into expressive directions that users can explore and combine." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.457 + ], + "angle": 0, + "content": "To address these needs, we present SliderSpace, a framework that brings systematic explorability to diffusion models. Given just a text prompt, SliderSpace discovers a canonical set of meaningful, diverse, and controllable directions within the model's knowledge of that concept. Each direction is implemented as a low-rank adapter [24] that can be scaled and composed with others, allowing users to explore and smoothly combine different aspects of variation, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.503 + ], + "angle": 0, + "content": "We ground SliderSpace discovery in three key requirements for meaningful decomposition of a diffusion model's visual manifold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.506, + 0.483, + 0.596 + ], + "angle": 0, + "content": "1. Unsupervised Discovery: The decomposition process should emerge from the intrinsic structure of the model's learned representation, rather than being guided by predefined attributes. This ensures we capture the true topology of the model's knowledge space rather than projecting our assumptions onto it." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.597, + 0.483, + 0.687 + ], + "angle": 0, + "content": "2. Semantic Orthogonality: Each discovered control must represent a distinct semantic direction. This is enforced in a semantic feature space, like CLIP, where every slider has an orthogonal effect in embeddings. This prevents discovering multiple controls that create similar semantic effects [15, 33, 47], making the system more efficient." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.687, + 0.483, + 0.732 + ], + "angle": 0, + "content": "3. Distribution Consistency: Directions must induce consistent transformations across both random seeds and prompt variations." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.506, + 0.483, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.824 + ], + "angle": 0, + "content": "These requirements naturally lead to our proposed framework, which we formalize in Section 4. As we show in our experiments, SliderSpace is architecture-agnostic, working with both conventional U-Net based models like Stable Diffusion [34, 37, 40, 43, 51] and recent transformer-based architectures like Flux [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We demonstrate the expressiveness of SliderSpace through three applications: First, we show how SliderSpace can decompose high-level concepts into diverse and expressive components, revealing the natural axes of variation in the model's understanding. Second, we explore artistic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "style variation, where SliderSpace discovers directions that match or exceed the diversity of manually curated artist lists while being judged more useful by human evaluators. Finally, we show how SliderSpace can help reverse the mode collapse commonly observed in distilled diffusion models, restoring diversity while maintaining generation speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Beyond providing practical creative control, SliderSpace opens new avenues for understanding and utilizing the latent capabilities of diffusion models. By mapping these models' visual potential into intuitive, composable directions, we take a step toward making their creative possibilities more accessible and interpretable to users." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.287, + 0.663, + 0.303 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.433 + ], + "angle": 0, + "content": "Recent text-to-image diffusion models have demonstrated remarkable capabilities in generating diverse visual concepts [40]. While newer foundation models enhance text-image alignment through LLM-generated captions [13, 26], the fundamental challenge remains: text-conditioned generation is inherently under-determined, with multiple distinct outputs potentially satisfying the same prompt, making precise control challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.613 + ], + "angle": 0, + "content": "Prior work has explored various approaches to enhance generation control. One direction introduces additional conditioning modalities: spatial signals via adapters (ControlNet [52]), attention-based regional control [8], and image-based conditioning for identity preservation [30, 46, 50] or style transfer via attention manipulation [21]. Another line of research focuses on refining existing images through disentangled text-driven editing [5, 7, 10, 32, 48, 49], where specific attributes are modified while preserving others, often using spatial masks. However, these text-driven approaches inherit the same under-determination challenges as the base models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Notably, disentangled control has been more naturally achieved in GANs [17], particularly through StyleGAN's low-dimensional latent space [28], which exhibits emergent disentanglement even at scale [27]. This has enabled powerful latent space and image editing capabilities [1, 2, 47]. While diffusion models offer superior generation quality and diversity, they lack two key advantages of GANs' latent space: continuous, compositional edits and emergent disentanglement. This limitation prevents users from making serendipitous discoveries about visual variations captured in the training data, instead constraining them to variations they can explicitly describe through prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent works have explored different approaches to discovering interpretable directions in diffusion models. The weights2weights method [12] learns a manifold of personalized model weights by fine-tuning individual LoRA adapters for each identity and applying PCA to discover a weight space that enables editing. While effective, this requires training separate models per instance. NoiseCLR [9]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15995" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "learns text embeddings through contrastive learning on a data distribution, but the discovered directions can be arbitrary and lack semantic interpretability. Similarly, Liu et al. [31] propose unsupervised decomposition of images into compositional concepts by learning multiple embeddings, but their approach often yields redundant or non-semantic directions. In contrast, our work directly learns a small set of semantically grounded sliders that decompose the distribution into interpretable and composable directions, enabling infinite creative variations through systematic exploration of the visual manifold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Concept Sliders [16] addressed continuous control by leveraging LoRA adaptors [24] to learn user-defined attributes. Our work complements this by tackling emergent disentanglement through self-supervised decomposition of the model's inherent variations into composable control dimensions, enabling systematic exploration of the model's creative capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.378, + 0.22, + 0.394 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.402, + 0.314, + 0.417 + ], + "angle": 0, + "content": "3.1. Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.515 + ], + "angle": 0, + "content": "State-of-the-art text-to-image diffusion models [4, 34, 38, 39] often belong to the class of latent diffusion. Unlike traditional diffusion models that operate in pixel space, latent diffusion models work in a compressed latent space, offering significant computational advantages [40]. The diffusion modelling can be formalized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Let \\(\\mathbf{x}_0\\) be an initial image and \\(\\mathbf{x}_T\\) be pure Gaussian noise. The forward diffusion process gradually adds noise to the image. The generative process aims to reverse this diffusion, starting from \\(x_{T}\\) and progressively denoising to reconstruct \\(\\mathbf{x}_0\\). At a timestep \\(t\\) the model takes \\(\\mathbf{x}_t\\) as input and predicts a noise \\(\\epsilon_t\\) such that the next step \\(\\mathbf{x}_{t - 1}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.618, + 0.482, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}}{\\sqrt {\\alpha_ {t}}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.737 + ], + "angle": 0, + "content": "We can estimate the final image \\(\\tilde{\\mathbf{x}}_{0,t}\\) by taking the same direction \\(\\epsilon_t\\) for remaining diffusion steps. This can be achieved by recursively applying the denoising from Equation 1 with the same direction. We label this \"Final Image Extrapolation\":" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.75, + 0.482, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {0, t} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.84 + ], + "angle": 0, + "content": "This enables us to visualize the final image \\(\\tilde{\\mathbf{x}}_{0,t}\\) that the diffusion model is planning of at each timestep \\(t\\) without actually running denoising forward passes through all timesteps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.342, + 0.865 + ], + "angle": 0, + "content": "3.2. LoRA: Low Rank Adaptors" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Low-rank adaptor (LoRA) [24] are a class of light-weight adaptors that are attachable to the weights of the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Given a pre-trained model layer with weights \\(W_0 \\in \\mathbb{R}^{d \\times k}\\), LoRA decomposes the weight update \\(\\Delta W\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.14, + 0.905, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\Delta W = B A, \\quad B \\in \\mathbb {R} ^ {d \\times r}, A \\in \\mathbb {R} ^ {r \\times k}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.262 + ], + "angle": 0, + "content": "where \\( r \\ll \\min(d, k) \\) is a small rank that constrains the update to a low-dimensional subspace. This decomposition allows for efficient parameter updates and has shown success in various downstream tasks. However, the application of LoRA to unsupervised discovery of semantic directions in diffusion model weight space remains unexplored." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.279, + 0.605, + 0.294 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.397 + ], + "angle": 0, + "content": "We present SliderSpace, a framework for decomposing a diffusion model's visual capabilities into semantically orthogonal control dimensions (Fig. 1). Given a pre-trained text-to-image diffusion model \\(\\theta\\) and a prompt \\(c\\), our goal is to discover \\(n\\) independent directions that capture the principal modes of variation in the model's learned distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.719, + 0.424 + ], + "angle": 0, + "content": "4.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.524 + ], + "angle": 0, + "content": "Let \\(\\mathcal{M}_{\\theta}(c)\\) denote the manifold of possible images that model \\(\\theta\\) can generate for prompt \\(c\\). We aim to identify a set of controllable directions \\(\\{\\mathcal{T}_i\\}_{i=1}^n\\) that: (1) span the major modes of variation in \\(\\mathcal{M}_{\\theta}(c)\\), (2) maintain semantic consistency across different initializations, and (3) are mutually orthogonal in semantic space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.906, + 0.616 + ], + "angle": 0, + "content": "Building on recent advances in model adaptation [16], we formulate each control dimension as a LoRA adaptor [24], updating \\(\\mathcal{T}_i\\), where \\(i \\in \\{1, \\dots, n\\}\\). These lightweight adapters introduce targeted modifications to the model's cross-attention layers, enabling precise control over specific generative attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.89, + 0.646 + ], + "angle": 0, + "content": "4.2. SliderSpace: Unsupervised Visual Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.906, + 0.683 + ], + "angle": 0, + "content": "SliderSpace discovery process consists of three key steps as depicted in Figure 2:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.709, + 0.907, + 0.77 + ], + "angle": 0, + "content": "Distribution Sampling First, we generate a diverse set of samples \\(\\{x_{j}\\}_{j = 1}^{m}\\) from \\(\\mathcal{M}_{\\theta}(c)\\) by varying the random seed (for stability, \\(m\\approx 5000\\)). For each sample, we extract the estimated final image \\(\\tilde{\\mathbf{x}}_{0,t}\\) at each timestep \\(t\\) using Eq. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.871 + ], + "angle": 0, + "content": "Semantic Decomposition We map each sample to a semantic embedding space, like CLIP, \\(\\phi (\\tilde{\\mathbf{x}}_{0,t})\\) and compute the principal components \\(V = \\{v_{i}\\}_{i = 1}^{n}\\) of the resulting distribution. These components represent orthogonal directions of maximal variation in semantic space:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nV = \\operatorname {P C A} \\left(\\left\\{\\phi \\left(\\tilde {x} _ {0, t}\\right) \\right\\} _ {j, t}\\right) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15996" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.309, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Given a prompt, SliderSpace first generates images and extracts the CLIP features. We then compute the spectral decomposition of the CLIP features and align the each slider with extracted principle components. Each slider is therefore trained to represent a unique semantic direction that are relevant in the diffusion model's knowledge of the prompt." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.379, + 0.487, + 0.441 + ], + "angle": 0, + "content": "Slider Training For each principal direction \\( v_{i} \\), we train a corresponding adapter \\( \\mathcal{T}_i \\) to induce transformations that align with \\( v_{i} \\) in semantic CLIP space. The training objective for each slider is:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.457, + 0.483, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s l i d e r s p a c e}} = \\sum_ {i = 1} ^ {n} 1 - \\cos \\left(\\Delta \\phi_ {i}, v _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.483, + 0.64 + ], + "angle": 0, + "content": "where \\(\\Delta \\phi_{i} = \\phi (\\tilde{\\mathbf{x}}_{0,t}^{i}) - \\phi (\\tilde{\\mathbf{x}}_{0,t})\\) represents the transformation induced by slider \\(i\\) in embedding space and \\(\\tilde{x}_{0,t}^{i}\\) is the estimated final image with the slider \\(\\mathcal{T}_i\\) attached. This ensures that each slider's effect is semantically aligned with its corresponding principal direction while maintaining orthogonality with other sliders, building on established work showing that embedding differences encode semantic relationships [15, 33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our formulation satisfies the three key requirements outlined in Section 1. First, unsupervised discovery is achieved by deriving directions directly from the model's intrinsic variation through PCA in a semantic embedding space, without imposing predefined attributes or external supervision. Second, distribution consistency is enforced by our training objective \\(\\mathcal{L}_{\\mathrm{sliderspace}}\\), which ensures each slider's transformation maintains consistent direction in embedding space across different seeds and timesteps. Finally, semantic orthogonality is guaranteed through the PCA-based initialization of directions as each principal components are mutually orthogonal, ensuring each slider captures a distinct mode of variation. Together, these components enable the discovery of interpretable and reliable control dimensions that effectively decompose the model's learned distribution. In majority of this paper, we use CLIP [36] as our primary semantic encoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.755, + 0.394 + ], + "angle": 0, + "content": "4.3. Interpretability & Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.507 + ], + "angle": 0, + "content": "SliderSpace provides a dual contribution: it serves as both a framework for discovering expressive dimensions of control and as a mechanism for decomposing the model's learned concept space into semantically meaningful components. Each adapter functions as a \"slider\" controlling a specific attribute of the generated image, enabling fine-grained manipulation of the output while maintaining semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.569 + ], + "angle": 0, + "content": "The resulting set of adapters provides valuable insights into the model's conceptual understanding, revealing nuanced semantic relationships that may not be immediately apparent from the text prompt alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.691 + ], + "angle": 0, + "content": "Moreover, the low-rank structure of our adapters ensures computational efficiency and minimal memory overhead, making them particularly suitable for real-time interactive applications. This enables users to explore the concept space dynamically by modulating the influence of each adapter, facilitating intuitive creative control over the image generation process while maintaining the underlying semantic integrity of the original prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.646, + 0.725 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We conduct our main experiments to evaluate SliderSpace using SDXL-DMD [51], a 4-step distilled diffusion model. Our implementation requires less than 24GB VRAM and can discover 64 semantic directions in under 2 hrs on a single A100 GPU. When concepts exhibit severe mode collapse, the discovery process can be enhanced by generating data using undistilled base models or LLM-expanded prompts for increased sample diversity. We demonstrate SliderSpace's generalization to SDXL [34], SDXL-Turbo [43], and transformer-based FLUX Schnell [4] in appendix. Our analysis focuses on three key applications:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15997" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.085, + 0.905, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 3. SliderSpace decomposes the visual variation of diffusion model's knowledge corresponding to a concept. These directions can be perceived as interpretable directions of the model's hierarchical knowledge. We show the decomposed slider direction for a concept using SliderSpace and the corresponding labels generated by Claude 3.5 Sonnet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.482, + 0.439 + ], + "angle": 0, + "content": "concept decomposition, art styles exploration, and diversity enhancement in distilled models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.453, + 0.313, + 0.47 + ], + "angle": 0, + "content": "5.1. Concept Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.553 + ], + "angle": 0, + "content": "We first demonstrate how SliderSpace can serve as an exploratory tool by decomposing high-level concepts into semantic directions that align with the diffusion model's internal representations. Summarizing and exposing the dominant variations a model is capable of for a particular prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Given a concept prompt (e.g., \"picture of a monster\"), we discover SliderSpace directions. Figure 3 shows discovered sliders for \"Monster\", and \"car\" concepts. We label the sliders using Claude 3.5 Sonnet [3] by showing multiple image pairs showing the effect of each slider and prompting to identify the semantic transformation being applied. Through these discovered directions, we demonstrate the manipulation of individual attributes. As these directions are intended to capture the model's visual possibilities for a prompt, we naturally ask the question, \"How much variation is enabled by these directions?\" To quantitatively evaluate the diversity enabled by SliderSpace against base model, we compute DreamSim [14] distance to measure inter-image variation across 2500 generated samples per concept (Figure 4). For SliderSpace-augmented generation, we generate the images by randomly activating a sparse subset of 3 sliders (out of 32 discovered directions) for each generation. Our analysis reveals that SliderSpace-generated samples exhibit significantly higher inter-image diversity compared to the baseline model outputs. We measure the CLIP-Score [22] between the input prompts and the generated images to measure text-alignment with the prompt. We find that they have similar CLIP Scores to the images gen" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.407, + 0.907, + 0.477 + ], + "angle": 0, + "content": "
Method vs.User Study (Win Rate %)
“Diverse”“Useful”“Creative”
SDXL-DMD72.466.068.1
LLM + SDXL-DMD62.562.562.5
SDXL65.361.259.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.489, + 0.908, + 0.531 + ], + "angle": 0, + "content": "Table 1. Users perceive SliderSpace generated images to be more diverse, useful, and interesting. We show win-rate percentages of SliderSpace samples against baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.682 + ], + "angle": 0, + "content": "erated by the base model. This suggests that SliderSpace effectively expands the achievable variation in the model's knowledge, while maintaining semantic consistency. To validate our findings through human perception, we conducted pairwise comparisons of image grids generated by SliderSpace versus baseline methods. As shown in Table 1, users consistently preferred SliderSpace outputs across diversity, utility, and creative potential." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.724, + 0.712 + ], + "angle": 0, + "content": "5.2. Art Styles Exploration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.854 + ], + "angle": 0, + "content": "We also evaluate to what degree SliderSpace can expose \"all\" of the art styles that the diffusion model has learned from its training data. As a proxy for \"all\", we use a diverse set of art styles, manually discovered and documented by ParrotZone [25]. We explore the visual artistic space by decomposing the prompt \"artwork in the style of a famous artist\" into 64 directions. This process enables us to discover the SliderSpace of art and create a comprehensive dictionary of discoverable art styles in a diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To compare how this measures up against supervised methods like Concept Slider [16], we train 64 manually curated concept sliders using LLM generated training" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15998" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.452, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.089, + 0.895, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.908, + 0.253 + ], + "angle": 0, + "content": "Figure 4. (left) SliderSpace generates diverse variations of a concept, as measured by DreamSim [14] distance across generated samples (higher is better). (right) The method maintains similar text-to-image alignment, as measured by CLIP Scores [22] (lower is better)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.486, + 0.611 + ], + "angle": 0, + "content": "prompts. We compute FID scores against the generated samples conditioned on actual artist names from ParrotZone [25] dataset. The dataset contains 4388 artists mimicked by SDXL, are discovered through exhaustive manual search. We also establish two other baselines. First, we use GPT/Claude to generate 4388 different artistic style names (e.g., \"cubism\", \"Warhol\"). Second, we use a generic prompt like \"[subject] in the style of a famous artist\". Using our discovered SliderSpace, we randomly sample a sparse set of 3 sliders to generate an equivalent number of images. Specifically, for each baseline, we match the provided samples of [25] by creating two images per style for \"building landscape\" and \"character portraits\" each (total of four images per style) to ensure same semantic structure across the datasets and that FID measures the artistic spread. Figure 5 shows random, non-cherry picked images from all the methods and their corresponding FID scores. We find that SliderSpace generates a distribution significantly more closely matches the manually curated artist names than the baselines, including supervised methods like Concept Sliders. Figure 6 shows qualitative comparison of art styles discovered by SliderSpace and styles discovered manually [25]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.793 + ], + "angle": 0, + "content": "To validate the practical utility of our \"artistic\" SliderSpace, we conduct a user study examining both usefulness and diversity of generated samples (Table 2). We conduct comparative study with 2 grids of 9 images (total of 1000 pairs), one with slider generated images and other with baselines. Users find the SliderSpace-generated images are far more diverse and useful than the baseline methods. We also find that users would prefer using SliderSpace images over images prompted with the real artist names [25], while finding these both to be equally diverse. We provide user-study details and qualitative examples in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.308, + 0.819 + ], + "angle": 0, + "content": "5.3. Diversity Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finally, we explore SliderSpace's ability to addressing mode collapse in distilled models. Instead of discovering SliderSpace for narrow distributions, we train it on a larger spread of 8000 randomly selected COCO-30k prompts and enhance them with LLM. We use DMD model for gener" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.276, + 0.895, + 0.348 + ], + "angle": 0, + "content": "
MethodUser study (win rate %)
“Diverse”“Useful”
vs. Real Artist Prompts [25]5463
vs. LLM Prompts7367
vs. Generic Prompts8788
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.358, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Table 2. Pairwise comparison user study (win rates in %) reveals that SliderSpace-extracted artistic styles achieve comparable diversity to manually curated artist [25], while being significantly preferred for creative utilization." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.431, + 0.875, + 0.509 + ], + "angle": 0, + "content": "
MethodFID-30k (↓)CLIP (↑)
Real-30.14
SDXL11.7229.41
SDXL-DMD15.5228.92
DMD-SliderSpace12.1229.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.52, + 0.909, + 0.563 + ], + "angle": 0, + "content": "Table 3. SliderSpace trained on larger range of knowledge can improve the diversity of the distilled models (FID) while having a good text-image alignment (CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.757 + ], + "angle": 0, + "content": "ating training dataset since we wish to undo mode collapse of DMD models. We discover a generic SliderSpace of 64 sliders that captures the generally applicable visual variations within SDXL-DMD. Figure 7 showcases qualitative examples of two distributions for the prompts \"Car driving through a forest\" and \"Picture of a person\". We demonstrate how generating images by randomly sampling from these SliderSpace directions effectively increases output diversity. This demonstrates that the mode collapse in distilled model can be reversed by discovering and exploring the visual structure. We show more examples in appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.909, + 0.835 + ], + "angle": 0, + "content": "In Table 3, we evaluate this improvement by measuring FID and CLIP scores on all COCO-30k prompts before and after applying SliderSpace to SDXL-DMD. We find that DMD-SliderSpace improves FID from the distilled version, almost matching the FID of undistilled SDXL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.719, + 0.864 + ], + "angle": 0, + "content": "5.4. Slider Transferability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Using FaceNet [44] as the semantic encoder, we train SliderSpace on the concept \"person\" and discover interpretable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15999" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.089, + 0.203, + 0.099 + ], + "angle": 0, + "content": "Real Artists" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.103, + 0.252, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.226, + 0.182, + 0.238 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.089, + 0.41, + 0.1 + ], + "angle": 0, + "content": "Generic Prompts (FID: 76.08)" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.102, + 0.416, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.226, + 0.345, + 0.238 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.089, + 0.566, + 0.1 + ], + "angle": 0, + "content": "LLM Prompts (FID: 38.84)" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.102, + 0.579, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.226, + 0.508, + 0.238 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.089, + 0.733, + 0.1 + ], + "angle": 0, + "content": "Concept Sliders (FID: 32.86)" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.102, + 0.742, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.226, + 0.672, + 0.238 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.089, + 0.886, + 0.1 + ], + "angle": 0, + "content": "SliderSpace (FID: 19.12)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.102, + 0.905, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.226, + 0.834, + 0.238 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 5. SliderSpace demonstrates broader artistic style coverage, as evidenced by the lower FID scores compared to both supervised Concept Sliders. Comparison of artistic style diversity using FID scores against reference distribution (a) derived from the complete artist dataset [25]. We compare against outputs from (b) generic art prompts (b), (c) LLM-generated art prompts, and (d) Concept Sliders." + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.304, + 0.425, + 0.319 + ], + "angle": 0, + "content": "Manually Discovered Styles by ParrotZone" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.325, + 0.238, + 0.337 + ], + "angle": 0, + "content": "Elenore Abbott" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.338, + 0.288, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.438, + 0.24, + 0.449 + ], + "angle": 0, + "content": "Wayne Barlowe" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.45, + 0.288, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.325, + 0.424, + 0.337 + ], + "angle": 0, + "content": "Miki Asai" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.337, + 0.49, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.438, + 0.438, + 0.45 + ], + "angle": 0, + "content": "Edward Gorey" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.45, + 0.49, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.304, + 0.856, + 0.319 + ], + "angle": 0, + "content": "Automatically Discovered Styles by SliderSpace" + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.325, + 0.64, + 0.337 + ], + "angle": 0, + "content": "Direction 1" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.337, + 0.703, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.438, + 0.645, + 0.449 + ], + "angle": 0, + "content": "Direction 25" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.449, + 0.703, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.325, + 0.847, + 0.337 + ], + "angle": 0, + "content": "Direction 23" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.337, + 0.905, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.438, + 0.846, + 0.449 + ], + "angle": 0, + "content": "Direction 31" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.45, + 0.905, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 6. Comparison of artistic styles discovered in SDXL-DMD: (Left) Representative samples from artist-specific prompts manually curated by the Parrotzone community [25] through extensive exploration. (Right) Automatically discovered artistic directions using SliderSpace, which captures diverse and semantically meaningful variations without requiring explicit artist references." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.72 + ], + "angle": 0, + "content": "directions controlling attributes like appearance and age (Figure 8). These directions not only transfer effectively to related concepts like \"police\" and \"athlete\" but also generalize surprisingly well to out-of-domain concepts like \"dog\", suggesting that SliderSpace captures fundamental visual transformations in the model's knowledge space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.484, + 0.767 + ], + "angle": 0, + "content": "Additional experiments analyzing hyperparameter choices (App. B.2), alternative semantic embeddings (App. B.1), and ablation studies (App. E) are in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.784, + 0.214, + 0.799 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our method can work with several semantic embeddings - each encoder might introduce different inherent biases present in its training data. While these embeddings enable semantic consistency, they may not capture certain culturally-specific or nuanced concepts. This highlights the need for application specific semantic embedding choice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.765 + ], + "angle": 0, + "content": "The current discovery process requires \\(\\approx 65\\) mins on A100. We note that our method trains 8 times faster than Concept Sliders for the same number of sliders. To further optimize, it opens avenues for future research into training time optimizations. For art style discovery, the discovered directions are not one-to-one matched with the real artists, this can be limiting based on users' application. Further work can address discovery that nudges the directions to be aligned with real artists." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.634, + 0.799 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "SliderSpace is a simple framework that automatically decomposes diffusion models' capabilities into semantically meaningful and controllable directions. By leveraging spectral decomposition in semantic space combined with low-rank adaptation, our method enables systematic exploration of a model's latent creative space without requiring manual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16000" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.086, + 0.885, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 7. SliderSpace can also decompose general broad visual variation of diffusion model's and can be used to overcome model collapse in distilled models. We randomly sample a sparse set of sliders and generate the sample showing higher variation than base distilled model." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.344, + 0.887, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.694, + 0.908, + 0.723 + ], + "angle": 0, + "content": "Figure 8. SliderSpace directions for the \"person\" concept successfully generalize to related \"police\" and \"athlete\" concepts. They also transfer to out-of-domain concepts like \"dog\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.486, + 0.901 + ], + "angle": 0, + "content": "attribute specification. Through extensive experiments, we demonstrated SliderSpace's effectiveness across three key applications. First, our concept decomposition revealed interpretable variations within the model's knowledge representation, enabling fine-grained control while maintaining semantic consistency. Second, our exploration of artistic capabilities showed that SliderSpace can discover directions matching or exceeding the diversity of manually curated artist lists, while being rated more useful by human evaluators. Finally, we demonstrate how SliderSpace can help" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.749, + 0.907, + 0.78 + ], + "angle": 0, + "content": "address mode collapse in distilled diffusion models, restoring diversity while preserving computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The ability of SliderSpace to uncover interpretable directions suggests that diffusion models may develop structured internal representations of visual concepts during training, without explicit supervision. By mapping these models' vast creative potential into intuitive, composable directions, our work takes a step toward making their capabilities more transparent and accessible." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16001" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.162 + ], + "angle": 0, + "content": "This work was done by RG at Adobe during their internship. RG and DB are supported by Open Philanthropy and NSF grant #2403304." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.187, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432-4441, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.441, + 0.344 + ], + "angle": 0, + "content": "[3] Anthropic. Introducing claude 3.5 sonnet, 2024. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[4] BlackForestLabs. Announcing state-of-the-art flux.1 dev and Schnell models, 2024. 1, 2, 3, 4, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv e-prints, pages arXiv-1809, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[8] Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompting for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.485, + 0.7 + ], + "angle": 0, + "content": "[9] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24209-24218, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[10] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "In Forty-first International Conference on Machine Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In European Conference on Computer Vision, pages 172-188. Springer, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[18] Imke Grabe, Miguel González-Duque, Sebastian Risi, and Jichen Zhu. Towards a framework for human-ai interaction patterns in co-creative gan applications. In Joint Proceedings of the ACM IUI Workshops, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[21] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[23] Oliver Hoffmann. On modeling human-computer cocreativity. In Knowledge, Information and Creativity Support Systems: Selected Papers from KICSS'2014-9th International Conference, held in Limassol, Cyprus, on November 6-8, 2014, pages 37-48. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[26] et al James Betker. Improving image generation with better captions. OpenAI Reports, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16002" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[29] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[31] Nan Liu, Yilun Du, Shuang Li, Joshua B Tenenbaum, and Antonio Torralba. Unsupervised compositional concepts discovery with text-to-image generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085-2095, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[32] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[35] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.447, + 0.729 + ], + "angle": 0, + "content": "[37] Robin Rombach. Stable diffusion 2.0 release, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[38] Robin Rombach and Patrick Esser. Stable diffusion v1-4 model card, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[39] Robin Rombach and Patrick Esser. Stable diffusion v2 model card, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and BjA¶rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[42] Simo Ryu. Cloneofsimo/lora: Using low-rank adaptation to quickly fine-tune diffusion models.s. GitHub, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[43] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 2, 4, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[44] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[47] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[48] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[49] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16003" + } + ] +] \ No newline at end of file diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_origin.pdf b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee11f5bdb918491bdc2d8179ec3c2062346d1b38 --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/50d20dec-7e48-4238-88fb-f013284e5706_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f579002cbadf445653e7c82045cb511759f57d084219ee509103e3464c4604cd +size 4004988 diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/full.md b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f3d2adb330ce701a9bf08bf58b5fa720d765aa2 --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/full.md @@ -0,0 +1,325 @@ +# SliderSpace: Decomposing the Visual Capabilities of Diffusion Models + +Rohit Gandikota $^{1*}$ Zongze Wu $^{2}$ Richard Zhang $^{2}$ David Bau $^{1}$ Eli Shechtman $^{2}$ Nick Kolkin $^{2}$ Northeastern University $^{2}$ Adobe Research + +![](images/920cee0ec944fef56241de91be0d5d59992603853c34d38ac3f394f2340b0e7b.jpg) +* The labels are generated by Claude-3.5 post-discovery of SliderSpace + +![](images/3dac2e37bb3af9403ab5a3cfda07d1896900b86c7b4039104f553eb968494a75.jpg) +Figure 1. Given a prompt, SliderSpace identifies the principal directions of the visual capabilities of a diffusion model by decomposing the image distribution over the prompt. By manipulating these directions as sliders, users can control and combine them to explore the creative possibilities of the model. We visualize the top directions discovered by SliderSpace within Flux Schnell [4] for the concept "Toy". + +# Abstract + +We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at sliderspace.baulab.info + +# 1. Introduction + +Text-to-image diffusion models are capable of generating remarkable visual variations from a single prompt through different random initializations. However, this vast creative potential remains largely opaque to users—while we can generate diverse images, we lack understanding of the underlying structure of these variations. This presents a fundamental challenge: how can we discover and expose the latent visual capabilities encoded within these models? + +The challenge touches on a key limitation in how we interact with diffusion models today. Current control methods require users to explicitly specify their desired edits in advance through prompts [16], reference images [8, 24, 29, 41, 42, 52], or attribute vectors [20, 21, 30, 32, 46, 50]. That contrasts sharply with natural human creative workflows, where artists dynamically explore creative ideas and jointly refine them toward meaningful artistic outcomes [23]. The need for pre-specified controls creates a barrier between users and the full creative potential of these models. + +Interestingly, earlier generative models like GANs [6, 17, 28] naturally developed more interpretable internal structures. Their compact latent spaces often exhibited + +emergent disentanglement [19, 35, 45, 47], enabling continuous and compositional control over generated images. Users could explore these spaces to discover interesting variations that would be difficult to describe in words [47], then combine them to achieve their creative goals [18]. + +Diffusion models have largely superseded GANs in conditional image synthesis [11], achieving greater diversity through much higher-dimensional latents. And yet an understanding of the underlying structure of these larger latent spaces has remained elusive. In this work, we ask a fundamental question: Can we automatically discover the visual structure within a diffusion model's knowledge of a concept? Rather than requiring user-specified controls, we aim to decompose the model's internal representations into expressive directions that users can explore and combine. + +To address these needs, we present SliderSpace, a framework that brings systematic explorability to diffusion models. Given just a text prompt, SliderSpace discovers a canonical set of meaningful, diverse, and controllable directions within the model's knowledge of that concept. Each direction is implemented as a low-rank adapter [24] that can be scaled and composed with others, allowing users to explore and smoothly combine different aspects of variation, as shown in Figure 1. + +We ground SliderSpace discovery in three key requirements for meaningful decomposition of a diffusion model's visual manifold: + +1. Unsupervised Discovery: The decomposition process should emerge from the intrinsic structure of the model's learned representation, rather than being guided by predefined attributes. This ensures we capture the true topology of the model's knowledge space rather than projecting our assumptions onto it. +2. Semantic Orthogonality: Each discovered control must represent a distinct semantic direction. This is enforced in a semantic feature space, like CLIP, where every slider has an orthogonal effect in embeddings. This prevents discovering multiple controls that create similar semantic effects [15, 33, 47], making the system more efficient. +3. Distribution Consistency: Directions must induce consistent transformations across both random seeds and prompt variations. + +These requirements naturally lead to our proposed framework, which we formalize in Section 4. As we show in our experiments, SliderSpace is architecture-agnostic, working with both conventional U-Net based models like Stable Diffusion [34, 37, 40, 43, 51] and recent transformer-based architectures like Flux [4]. + +We demonstrate the expressiveness of SliderSpace through three applications: First, we show how SliderSpace can decompose high-level concepts into diverse and expressive components, revealing the natural axes of variation in the model's understanding. Second, we explore artistic + +style variation, where SliderSpace discovers directions that match or exceed the diversity of manually curated artist lists while being judged more useful by human evaluators. Finally, we show how SliderSpace can help reverse the mode collapse commonly observed in distilled diffusion models, restoring diversity while maintaining generation speed. + +Beyond providing practical creative control, SliderSpace opens new avenues for understanding and utilizing the latent capabilities of diffusion models. By mapping these models' visual potential into intuitive, composable directions, we take a step toward making their creative possibilities more accessible and interpretable to users. + +# 2. Related Works + +Recent text-to-image diffusion models have demonstrated remarkable capabilities in generating diverse visual concepts [40]. While newer foundation models enhance text-image alignment through LLM-generated captions [13, 26], the fundamental challenge remains: text-conditioned generation is inherently under-determined, with multiple distinct outputs potentially satisfying the same prompt, making precise control challenging. + +Prior work has explored various approaches to enhance generation control. One direction introduces additional conditioning modalities: spatial signals via adapters (ControlNet [52]), attention-based regional control [8], and image-based conditioning for identity preservation [30, 46, 50] or style transfer via attention manipulation [21]. Another line of research focuses on refining existing images through disentangled text-driven editing [5, 7, 10, 32, 48, 49], where specific attributes are modified while preserving others, often using spatial masks. However, these text-driven approaches inherit the same under-determination challenges as the base models. + +Notably, disentangled control has been more naturally achieved in GANs [17], particularly through StyleGAN's low-dimensional latent space [28], which exhibits emergent disentanglement even at scale [27]. This has enabled powerful latent space and image editing capabilities [1, 2, 47]. While diffusion models offer superior generation quality and diversity, they lack two key advantages of GANs' latent space: continuous, compositional edits and emergent disentanglement. This limitation prevents users from making serendipitous discoveries about visual variations captured in the training data, instead constraining them to variations they can explicitly describe through prompts. + +Recent works have explored different approaches to discovering interpretable directions in diffusion models. The weights2weights method [12] learns a manifold of personalized model weights by fine-tuning individual LoRA adapters for each identity and applying PCA to discover a weight space that enables editing. While effective, this requires training separate models per instance. NoiseCLR [9] + +learns text embeddings through contrastive learning on a data distribution, but the discovered directions can be arbitrary and lack semantic interpretability. Similarly, Liu et al. [31] propose unsupervised decomposition of images into compositional concepts by learning multiple embeddings, but their approach often yields redundant or non-semantic directions. In contrast, our work directly learns a small set of semantically grounded sliders that decompose the distribution into interpretable and composable directions, enabling infinite creative variations through systematic exploration of the visual manifold. + +Concept Sliders [16] addressed continuous control by leveraging LoRA adaptors [24] to learn user-defined attributes. Our work complements this by tackling emergent disentanglement through self-supervised decomposition of the model's inherent variations into composable control dimensions, enabling systematic exploration of the model's creative capabilities. + +# 3. Background + +# 3.1. Latent Diffusion Models + +State-of-the-art text-to-image diffusion models [4, 34, 38, 39] often belong to the class of latent diffusion. Unlike traditional diffusion models that operate in pixel space, latent diffusion models work in a compressed latent space, offering significant computational advantages [40]. The diffusion modelling can be formalized as follows: + +Let $\mathbf{x}_0$ be an initial image and $\mathbf{x}_T$ be pure Gaussian noise. The forward diffusion process gradually adds noise to the image. The generative process aims to reverse this diffusion, starting from $x_{T}$ and progressively denoising to reconstruct $\mathbf{x}_0$ . At a timestep $t$ the model takes $\mathbf{x}_t$ as input and predicts a noise $\epsilon_t$ such that the next step $\mathbf{x}_{t - 1}$ : + +$$ +\mathbf {x} _ {t - 1} \leftarrow \frac {\mathbf {x} _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {t}}{\sqrt {\alpha_ {t}}} \tag {1} +$$ + +We can estimate the final image $\tilde{\mathbf{x}}_{0,t}$ by taking the same direction $\epsilon_t$ for remaining diffusion steps. This can be achieved by recursively applying the denoising from Equation 1 with the same direction. We label this "Final Image Extrapolation": + +$$ +\tilde {\mathbf {x}} _ {0, t} \leftarrow \frac {\mathbf {x} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}}{\sqrt {\bar {\alpha} _ {t}}}. \tag {2} +$$ + +This enables us to visualize the final image $\tilde{\mathbf{x}}_{0,t}$ that the diffusion model is planning of at each timestep $t$ without actually running denoising forward passes through all timesteps. + +# 3.2. LoRA: Low Rank Adaptors + +Low-rank adaptor (LoRA) [24] are a class of light-weight adaptors that are attachable to the weights of the model. + +Given a pre-trained model layer with weights $W_0 \in \mathbb{R}^{d \times k}$ , LoRA decomposes the weight update $\Delta W$ as: + +$$ +\Delta W = B A, \quad B \in \mathbb {R} ^ {d \times r}, A \in \mathbb {R} ^ {r \times k}, \tag {3} +$$ + +where $r \ll \min(d, k)$ is a small rank that constrains the update to a low-dimensional subspace. This decomposition allows for efficient parameter updates and has shown success in various downstream tasks. However, the application of LoRA to unsupervised discovery of semantic directions in diffusion model weight space remains unexplored. + +# 4. Method + +We present SliderSpace, a framework for decomposing a diffusion model's visual capabilities into semantically orthogonal control dimensions (Fig. 1). Given a pre-trained text-to-image diffusion model $\theta$ and a prompt $c$ , our goal is to discover $n$ independent directions that capture the principal modes of variation in the model's learned distribution. + +# 4.1. Problem Formulation + +Let $\mathcal{M}_{\theta}(c)$ denote the manifold of possible images that model $\theta$ can generate for prompt $c$ . We aim to identify a set of controllable directions $\{\mathcal{T}_i\}_{i=1}^n$ that: (1) span the major modes of variation in $\mathcal{M}_{\theta}(c)$ , (2) maintain semantic consistency across different initializations, and (3) are mutually orthogonal in semantic space. + +Building on recent advances in model adaptation [16], we formulate each control dimension as a LoRA adaptor [24], updating $\mathcal{T}_i$ , where $i \in \{1, \dots, n\}$ . These lightweight adapters introduce targeted modifications to the model's cross-attention layers, enabling precise control over specific generative attributes. + +# 4.2. SliderSpace: Unsupervised Visual Discovery + +SliderSpace discovery process consists of three key steps as depicted in Figure 2: + +Distribution Sampling First, we generate a diverse set of samples $\{x_{j}\}_{j = 1}^{m}$ from $\mathcal{M}_{\theta}(c)$ by varying the random seed (for stability, $m\approx 5000$ ). For each sample, we extract the estimated final image $\tilde{\mathbf{x}}_{0,t}$ at each timestep $t$ using Eq. 2. + +Semantic Decomposition We map each sample to a semantic embedding space, like CLIP, $\phi (\tilde{\mathbf{x}}_{0,t})$ and compute the principal components $V = \{v_{i}\}_{i = 1}^{n}$ of the resulting distribution. These components represent orthogonal directions of maximal variation in semantic space: + +$$ +V = \operatorname {P C A} \left(\left\{\phi \left(\tilde {x} _ {0, t}\right) \right\} _ {j, t}\right) \tag {4} +$$ + +![](images/44f0c7fc1eed88d6179270d33846652e8a5f5ecd3507b6edd6b63162736fa5e7.jpg) +Figure 2. Given a prompt, SliderSpace first generates images and extracts the CLIP features. We then compute the spectral decomposition of the CLIP features and align the each slider with extracted principle components. Each slider is therefore trained to represent a unique semantic direction that are relevant in the diffusion model's knowledge of the prompt. + +Slider Training For each principal direction $v_{i}$ , we train a corresponding adapter $\mathcal{T}_i$ to induce transformations that align with $v_{i}$ in semantic CLIP space. The training objective for each slider is: + +$$ +\mathcal {L} _ {\text {s l i d e r s p a c e}} = \sum_ {i = 1} ^ {n} 1 - \cos \left(\Delta \phi_ {i}, v _ {i}\right), \tag {5} +$$ + +where $\Delta \phi_{i} = \phi (\tilde{\mathbf{x}}_{0,t}^{i}) - \phi (\tilde{\mathbf{x}}_{0,t})$ represents the transformation induced by slider $i$ in embedding space and $\tilde{x}_{0,t}^{i}$ is the estimated final image with the slider $\mathcal{T}_i$ attached. This ensures that each slider's effect is semantically aligned with its corresponding principal direction while maintaining orthogonality with other sliders, building on established work showing that embedding differences encode semantic relationships [15, 33]. + +Our formulation satisfies the three key requirements outlined in Section 1. First, unsupervised discovery is achieved by deriving directions directly from the model's intrinsic variation through PCA in a semantic embedding space, without imposing predefined attributes or external supervision. Second, distribution consistency is enforced by our training objective $\mathcal{L}_{\mathrm{sliderspace}}$ , which ensures each slider's transformation maintains consistent direction in embedding space across different seeds and timesteps. Finally, semantic orthogonality is guaranteed through the PCA-based initialization of directions as each principal components are mutually orthogonal, ensuring each slider captures a distinct mode of variation. Together, these components enable the discovery of interpretable and reliable control dimensions that effectively decompose the model's learned distribution. In majority of this paper, we use CLIP [36] as our primary semantic encoder. + +# 4.3. Interpretability & Control + +SliderSpace provides a dual contribution: it serves as both a framework for discovering expressive dimensions of control and as a mechanism for decomposing the model's learned concept space into semantically meaningful components. Each adapter functions as a "slider" controlling a specific attribute of the generated image, enabling fine-grained manipulation of the output while maintaining semantics. + +The resulting set of adapters provides valuable insights into the model's conceptual understanding, revealing nuanced semantic relationships that may not be immediately apparent from the text prompt alone. + +Moreover, the low-rank structure of our adapters ensures computational efficiency and minimal memory overhead, making them particularly suitable for real-time interactive applications. This enables users to explore the concept space dynamically by modulating the influence of each adapter, facilitating intuitive creative control over the image generation process while maintaining the underlying semantic integrity of the original prompt. + +# 5. Experiments + +We conduct our main experiments to evaluate SliderSpace using SDXL-DMD [51], a 4-step distilled diffusion model. Our implementation requires less than 24GB VRAM and can discover 64 semantic directions in under 2 hrs on a single A100 GPU. When concepts exhibit severe mode collapse, the discovery process can be enhanced by generating data using undistilled base models or LLM-expanded prompts for increased sample diversity. We demonstrate SliderSpace's generalization to SDXL [34], SDXL-Turbo [43], and transformer-based FLUX Schnell [4] in appendix. Our analysis focuses on three key applications: + +![](images/cfc3a3fc362612dc20df1e7364ec5bb07150c7ef2d43349f054c0a37e817dde8.jpg) +Figure 3. SliderSpace decomposes the visual variation of diffusion model's knowledge corresponding to a concept. These directions can be perceived as interpretable directions of the model's hierarchical knowledge. We show the decomposed slider direction for a concept using SliderSpace and the corresponding labels generated by Claude 3.5 Sonnet. + +concept decomposition, art styles exploration, and diversity enhancement in distilled models. + +# 5.1. Concept Decomposition + +We first demonstrate how SliderSpace can serve as an exploratory tool by decomposing high-level concepts into semantic directions that align with the diffusion model's internal representations. Summarizing and exposing the dominant variations a model is capable of for a particular prompt. + +Given a concept prompt (e.g., "picture of a monster"), we discover SliderSpace directions. Figure 3 shows discovered sliders for "Monster", and "car" concepts. We label the sliders using Claude 3.5 Sonnet [3] by showing multiple image pairs showing the effect of each slider and prompting to identify the semantic transformation being applied. Through these discovered directions, we demonstrate the manipulation of individual attributes. As these directions are intended to capture the model's visual possibilities for a prompt, we naturally ask the question, "How much variation is enabled by these directions?" To quantitatively evaluate the diversity enabled by SliderSpace against base model, we compute DreamSim [14] distance to measure inter-image variation across 2500 generated samples per concept (Figure 4). For SliderSpace-augmented generation, we generate the images by randomly activating a sparse subset of 3 sliders (out of 32 discovered directions) for each generation. Our analysis reveals that SliderSpace-generated samples exhibit significantly higher inter-image diversity compared to the baseline model outputs. We measure the CLIP-Score [22] between the input prompts and the generated images to measure text-alignment with the prompt. We find that they have similar CLIP Scores to the images gen + +
Method vs.User Study (Win Rate %)
“Diverse”“Useful”“Creative”
SDXL-DMD72.466.068.1
LLM + SDXL-DMD62.562.562.5
SDXL65.361.259.2
+ +Table 1. Users perceive SliderSpace generated images to be more diverse, useful, and interesting. We show win-rate percentages of SliderSpace samples against baselines. + +erated by the base model. This suggests that SliderSpace effectively expands the achievable variation in the model's knowledge, while maintaining semantic consistency. To validate our findings through human perception, we conducted pairwise comparisons of image grids generated by SliderSpace versus baseline methods. As shown in Table 1, users consistently preferred SliderSpace outputs across diversity, utility, and creative potential. + +# 5.2. Art Styles Exploration + +We also evaluate to what degree SliderSpace can expose "all" of the art styles that the diffusion model has learned from its training data. As a proxy for "all", we use a diverse set of art styles, manually discovered and documented by ParrotZone [25]. We explore the visual artistic space by decomposing the prompt "artwork in the style of a famous artist" into 64 directions. This process enables us to discover the SliderSpace of art and create a comprehensive dictionary of discoverable art styles in a diffusion model. + +To compare how this measures up against supervised methods like Concept Slider [16], we train 64 manually curated concept sliders using LLM generated training + +![](images/76d2d6e63c6f42822c0d0e4e0d28f217dc37b7fb50991f881a8a975aa8bd0e0b.jpg) +Figure 4. (left) SliderSpace generates diverse variations of a concept, as measured by DreamSim [14] distance across generated samples (higher is better). (right) The method maintains similar text-to-image alignment, as measured by CLIP Scores [22] (lower is better). + +![](images/d1632b0ce9a286263773b7ce8a092ed7bfd9ef2bfd506bafbdf3ccfb7c9f7df8.jpg) + +prompts. We compute FID scores against the generated samples conditioned on actual artist names from ParrotZone [25] dataset. The dataset contains 4388 artists mimicked by SDXL, are discovered through exhaustive manual search. We also establish two other baselines. First, we use GPT/Claude to generate 4388 different artistic style names (e.g., "cubism", "Warhol"). Second, we use a generic prompt like "[subject] in the style of a famous artist". Using our discovered SliderSpace, we randomly sample a sparse set of 3 sliders to generate an equivalent number of images. Specifically, for each baseline, we match the provided samples of [25] by creating two images per style for "building landscape" and "character portraits" each (total of four images per style) to ensure same semantic structure across the datasets and that FID measures the artistic spread. Figure 5 shows random, non-cherry picked images from all the methods and their corresponding FID scores. We find that SliderSpace generates a distribution significantly more closely matches the manually curated artist names than the baselines, including supervised methods like Concept Sliders. Figure 6 shows qualitative comparison of art styles discovered by SliderSpace and styles discovered manually [25]. + +To validate the practical utility of our "artistic" SliderSpace, we conduct a user study examining both usefulness and diversity of generated samples (Table 2). We conduct comparative study with 2 grids of 9 images (total of 1000 pairs), one with slider generated images and other with baselines. Users find the SliderSpace-generated images are far more diverse and useful than the baseline methods. We also find that users would prefer using SliderSpace images over images prompted with the real artist names [25], while finding these both to be equally diverse. We provide user-study details and qualitative examples in Appendix. + +# 5.3. Diversity Enhancement + +Finally, we explore SliderSpace's ability to addressing mode collapse in distilled models. Instead of discovering SliderSpace for narrow distributions, we train it on a larger spread of 8000 randomly selected COCO-30k prompts and enhance them with LLM. We use DMD model for gener + +
MethodUser study (win rate %)
“Diverse”“Useful”
vs. Real Artist Prompts [25]5463
vs. LLM Prompts7367
vs. Generic Prompts8788
+ +Table 2. Pairwise comparison user study (win rates in %) reveals that SliderSpace-extracted artistic styles achieve comparable diversity to manually curated artist [25], while being significantly preferred for creative utilization. + +
MethodFID-30k (↓)CLIP (↑)
Real-30.14
SDXL11.7229.41
SDXL-DMD15.5228.92
DMD-SliderSpace12.1229.13
+ +Table 3. SliderSpace trained on larger range of knowledge can improve the diversity of the distilled models (FID) while having a good text-image alignment (CLIP). + +ating training dataset since we wish to undo mode collapse of DMD models. We discover a generic SliderSpace of 64 sliders that captures the generally applicable visual variations within SDXL-DMD. Figure 7 showcases qualitative examples of two distributions for the prompts "Car driving through a forest" and "Picture of a person". We demonstrate how generating images by randomly sampling from these SliderSpace directions effectively increases output diversity. This demonstrates that the mode collapse in distilled model can be reversed by discovering and exploring the visual structure. We show more examples in appendix. + +In Table 3, we evaluate this improvement by measuring FID and CLIP scores on all COCO-30k prompts before and after applying SliderSpace to SDXL-DMD. We find that DMD-SliderSpace improves FID from the distilled version, almost matching the FID of undistilled SDXL. + +# 5.4. Slider Transferability + +Using FaceNet [44] as the semantic encoder, we train SliderSpace on the concept "person" and discover interpretable + +![](images/d80110c9f12f3f9954d4b3a4ec9f41c4444d3671cf8287be8d0006894e6e0897.jpg) +Real Artists +(a) +Figure 5. SliderSpace demonstrates broader artistic style coverage, as evidenced by the lower FID scores compared to both supervised Concept Sliders. Comparison of artistic style diversity using FID scores against reference distribution (a) derived from the complete artist dataset [25]. We compare against outputs from (b) generic art prompts (b), (c) LLM-generated art prompts, and (d) Concept Sliders. + +![](images/d1a168bb77bc4386aed65278bcc123c9fbf4647d0930c6770081aa810aa80c8c.jpg) +Generic Prompts (FID: 76.08) +(b) + +![](images/047d5a705de86afb97ece042b464f4fbd570cb44cb28e556600f9b06a417a557.jpg) +LLM Prompts (FID: 38.84) +(c) + +![](images/978c80621cd18437b6128771157752b3c98c675a3bc3485b4d0f4bb979c6d1e3.jpg) +Concept Sliders (FID: 32.86) +(d) + +![](images/88c7521209cf47f21218615b7bc43de90d267b42c4a618ed474d825c41f98143.jpg) +SliderSpace (FID: 19.12) +(e) + +![](images/498ef8e9a906b5bd57e11f713f202600ca486ca0643c891bfb42945ec28d06e1.jpg) +Elenore Abbott +Wayne Barlowe + +![](images/9ad2537634c15198edf91fe3acc1d43bd4f7c7c0a2df506033b8e3428e16f351.jpg) +Figure 6. Comparison of artistic styles discovered in SDXL-DMD: (Left) Representative samples from artist-specific prompts manually curated by the Parrotzone community [25] through extensive exploration. (Right) Automatically discovered artistic directions using SliderSpace, which captures diverse and semantically meaningful variations without requiring explicit artist references. + +![](images/4304f4149c1af9f75e2daa39d59232faeb0ecaa09490372d2460d93d38b201ab.jpg) +Manually Discovered Styles by ParrotZone +Miki Asai + +![](images/95e0fc2f3b6c5cd7e1a4c1128b35b2674f2294f20f2c003ce470933f73d2f7d7.jpg) +Edward Gorey + +![](images/dc4a9b33d232f45b5091c94c841b9bda05bcf6f252491808d13d222046d9fffe.jpg) +Automatically Discovered Styles by SliderSpace +Direction 1 + +![](images/3bbce1fb0c35f38f87b891054cfb07c24799191f60e9f8bfda8b403a8ad9787a.jpg) +Direction 25 + +![](images/a81b70628fccbdbca3d7f116f7373cc9b6683b1cc913c08939e01b3b5c480d85.jpg) +Direction 23 +Direction 31 + +![](images/4e1ba86a094f362ebb0f7069db2c523607f1554b9b1fa9726b41003df294a020.jpg) + +directions controlling attributes like appearance and age (Figure 8). These directions not only transfer effectively to related concepts like "police" and "athlete" but also generalize surprisingly well to out-of-domain concepts like "dog", suggesting that SliderSpace captures fundamental visual transformations in the model's knowledge space. + +Additional experiments analyzing hyperparameter choices (App. B.2), alternative semantic embeddings (App. B.1), and ablation studies (App. E) are in Appendix. + +# 6. Limitations + +Our method can work with several semantic embeddings - each encoder might introduce different inherent biases present in its training data. While these embeddings enable semantic consistency, they may not capture certain culturally-specific or nuanced concepts. This highlights the need for application specific semantic embedding choice. + +The current discovery process requires $\approx 65$ mins on A100. We note that our method trains 8 times faster than Concept Sliders for the same number of sliders. To further optimize, it opens avenues for future research into training time optimizations. For art style discovery, the discovered directions are not one-to-one matched with the real artists, this can be limiting based on users' application. Further work can address discovery that nudges the directions to be aligned with real artists. + +# 7. Conclusion + +SliderSpace is a simple framework that automatically decomposes diffusion models' capabilities into semantically meaningful and controllable directions. By leveraging spectral decomposition in semantic space combined with low-rank adaptation, our method enables systematic exploration of a model's latent creative space without requiring manual + +![](images/eb754555c6b0f0f77b2abd71066bc942065255aa3a1317903cdabee3daeec71b.jpg) +Figure 7. SliderSpace can also decompose general broad visual variation of diffusion model's and can be used to overcome model collapse in distilled models. We randomly sample a sparse set of sliders and generate the sample showing higher variation than base distilled model. + +![](images/5fba636fa30ffefdd9cf8b55832fc124f5f59891838834dc2629b516e7cf6cb2.jpg) +Figure 8. SliderSpace directions for the "person" concept successfully generalize to related "police" and "athlete" concepts. They also transfer to out-of-domain concepts like "dog" + +attribute specification. Through extensive experiments, we demonstrated SliderSpace's effectiveness across three key applications. First, our concept decomposition revealed interpretable variations within the model's knowledge representation, enabling fine-grained control while maintaining semantic consistency. Second, our exploration of artistic capabilities showed that SliderSpace can discover directions matching or exceeding the diversity of manually curated artist lists, while being rated more useful by human evaluators. Finally, we demonstrate how SliderSpace can help + +address mode collapse in distilled diffusion models, restoring diversity while preserving computational efficiency. + +The ability of SliderSpace to uncover interpretable directions suggests that diffusion models may develop structured internal representations of visual concepts during training, without explicit supervision. By mapping these models' vast creative potential into intuitive, composable directions, our work takes a step toward making their capabilities more transparent and accessible. + +# Acknowledgment + +This work was done by RG at Adobe during their internship. RG and DB are supported by Open Philanthropy and NSF grant #2403304. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432-4441, 2019. 2 +[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2 +[3] Anthropic. Introducing claude 3.5 sonnet, 2024. 5, 2 +[4] BlackForestLabs. Announcing state-of-the-art flux.1 dev and Schnell models, 2024. 1, 2, 3, 4, 17 +[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2 +[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv e-prints, pages arXiv-1809, 2018. 1 +[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[8] Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompting for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024. 1, 2 +[9] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24209-24218, 2024. 2, 3 +[10] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 2 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[12] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 2 +[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. + +In Forty-first International Conference on Machine Learning, 2024. 2 +[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[15] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4 +[16] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In European Conference on Computer Vision, pages 172-188. Springer, 2024. 1, 3, 5 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2 +[18] Imke Grabe, Miguel González-Duque, Sebastian Risi, and Jichen Zhu. Towards a framework for human-ai interaction patterns in co-creative gan applications. In Joint Proceedings of the ACM IUI Workshops, 2022. 2 +[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2 +[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1 +[21] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2 +[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6 +[23] Oliver Hoffmann. On modeling human-computer cocreativity. In Knowledge, Information and Creativity Support Systems: Selected Papers from KICSS'2014-9th International Conference, held in Limassol, Cyprus, on November 6-8, 2014, pages 37-48. Springer, 2016. 1 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2, 3 +[25] Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies, 2022. 5, 6, 7 +[26] et al James Betker. Improving image generation with better captions. OpenAI Reports, 2023. 2 +[27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2 +[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2 +[29] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3 +[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2 +[31] Nan Liu, Yilun Du, Shuang Li, Joshua B Tenenbaum, and Antonio Torralba. Unsupervised compositional concepts discovery with text-to-image generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085-2095, 2023. 3 +[32] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2 +[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 2, 4 +[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 4, 8 +[35] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[37] Robin Rombach. Stable diffusion 2.0 release, 2022. 2 +[38] Robin Rombach and Patrick Esser. Stable diffusion v1-4 model card, 2022. 3 +[39] Robin Rombach and Patrick Esser. Stable diffusion v2 model card, 2022. 3 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and BjA¶rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven + +generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3 +[42] Simo Ryu. Cloneofsimo/lora: Using low-rank adaptation to quickly fine-tune diffusion models.s. GitHub, 2023. 1 +[43] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 2, 4, 16 +[44] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6 +[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2 +[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 1, 2 +[47] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 2 +[48] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024. 2 +[49] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2 +[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2 +[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15 +[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2 \ No newline at end of file diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/images.zip b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6403dca575a2b782e55c562cafee404a4aa847ef --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78868fe438e2b82c1ffe99eecef386a39e3f2a24310ed451ff301e399ab8b99e +size 1020812 diff --git a/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/layout.json b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c056a6be568f839b820cbb738a98c675538f4635 --- /dev/null +++ b/2025/SliderSpace_ Decomposing the Visual Capabilities of Diffusion Models/layout.json @@ -0,0 +1,8438 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 522, + 121 + ], + "type": "text", + "content": "SliderSpace: Decomposing the Visual Capabilities of Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": "Rohit Gandikota" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Zongze Wu" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Richard Zhang" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " David Bau" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Eli Shechtman" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Nick Kolkin" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Northeastern University" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 553, + 175 + ], + "type": "text", + "content": " Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 204, + 403, + 357 + ], + "blocks": [ + { + "bbox": [ + 63, + 204, + 403, + 357 + ], + "lines": [ + { + "bbox": [ + 63, + 204, + 403, + 357 + ], + "spans": [ + { + "bbox": [ + 63, + 204, + 403, + 357 + ], + "type": "image", + "image_path": "920cee0ec944fef56241de91be0d5d59992603853c34d38ac3f394f2340b0e7b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 357, + 399, + 365 + ], + "lines": [ + { + "bbox": [ + 222, + 357, + 399, + 365 + ], + "spans": [ + { + "bbox": [ + 222, + 357, + 399, + 365 + ], + "type": "text", + "content": "* The labels are generated by Claude-3.5 post-discovery of SliderSpace" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 412, + 205, + 551, + 356 + ], + "blocks": [ + { + "bbox": [ + 412, + 205, + 551, + 356 + ], + "lines": [ + { + "bbox": [ + 412, + 205, + 551, + 356 + ], + "spans": [ + { + "bbox": [ + 412, + 205, + 551, + 356 + ], + "type": "image", + "image_path": "3dac2e37bb3af9403ab5a3cfda07d1896900b86c7b4039104f553eb968494a75.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 376, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 555, + 411 + ], + "type": "text", + "content": "Figure 1. Given a prompt, SliderSpace identifies the principal directions of the visual capabilities of a diffusion model by decomposing the image distribution over the prompt. By manipulating these directions as sliders, users can control and combine them to explore the creative possibilities of the model. We visualize the top directions discovered by SliderSpace within Flux Schnell [4] for the concept \"Toy\"." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 422, + 200, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 422, + 200, + 434 + ], + "spans": [ + { + "bbox": [ + 152, + 422, + 200, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 448, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 700 + ], + "type": "text", + "content": "We present SliderSpace, a framework for automatically decomposing the visual capabilities of diffusion models into controllable and human-understandable directions. Unlike existing control methods that require a user to specify attributes for each edit direction individually, SliderSpace discovers multiple interpretable and diverse directions simultaneously from a single text prompt. Each direction is trained as a low-rank adaptor, enabling compositional control and the discovery of surprising possibilities in the model's latent space. Through extensive experiments on state-of-the-art diffusion models, we demonstrate SliderSpace's effectiveness across three applications: concept decomposition, artistic style exploration, and diversity enhancement. Our quantitative evaluation shows that SliderSpace-discovered directions decompose the visual structure of model's knowledge effectively, offering insights into the latent capabilities encoded within diffusion models. User studies further validate that our method produces more diverse and useful variations compared to baselines. Our code, data and trained weights are available at sliderspace.baulab.info" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 442, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 442, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 312, + 442, + 555, + 537 + ], + "type": "text", + "content": "Text-to-image diffusion models are capable of generating remarkable visual variations from a single prompt through different random initializations. However, this vast creative potential remains largely opaque to users—while we can generate diverse images, we lack understanding of the underlying structure of these variations. This presents a fundamental challenge: how can we discover and expose the latent visual capabilities encoded within these models?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 539, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 539, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 539, + 556, + 658 + ], + "type": "text", + "content": "The challenge touches on a key limitation in how we interact with diffusion models today. Current control methods require users to explicitly specify their desired edits in advance through prompts [16], reference images [8, 24, 29, 41, 42, 52], or attribute vectors [20, 21, 30, 32, 46, 50]. That contrasts sharply with natural human creative workflows, where artists dynamically explore creative ideas and jointly refine them toward meaningful artistic outcomes [23]. The need for pre-specified controls creates a barrier between users and the full creative potential of these models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 555, + 696 + ], + "type": "text", + "content": "Interestingly, earlier generative models like GANs [6, 17, 28] naturally developed more interpretable internal structures. Their compact latent spaces often exhibited" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 329, + 703, + 537, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 703, + 537, + 712 + ], + "spans": [ + { + "bbox": [ + 329, + 703, + 537, + 712 + ], + "type": "text", + "content": "*Correspondence to gandikota.ro@northeastern.edu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15994" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "emergent disentanglement [19, 35, 45, 47], enabling continuous and compositional control over generated images. Users could explore these spaces to discover interesting variations that would be difficult to describe in words [47], then combine them to achieve their creative goals [18]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 294, + 252 + ], + "type": "text", + "content": "Diffusion models have largely superseded GANs in conditional image synthesis [11], achieving greater diversity through much higher-dimensional latents. And yet an understanding of the underlying structure of these larger latent spaces has remained elusive. In this work, we ask a fundamental question: Can we automatically discover the visual structure within a diffusion model's knowledge of a concept? Rather than requiring user-specified controls, we aim to decompose the model's internal representations into expressive directions that users can explore and combine." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 361 + ], + "type": "text", + "content": "To address these needs, we present SliderSpace, a framework that brings systematic explorability to diffusion models. Given just a text prompt, SliderSpace discovers a canonical set of meaningful, diverse, and controllable directions within the model's knowledge of that concept. Each direction is implemented as a low-rank adapter [24] that can be scaled and composed with others, allowing users to explore and smoothly combine different aspects of variation, as shown in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 398 + ], + "type": "text", + "content": "We ground SliderSpace discovery in three key requirements for meaningful decomposition of a diffusion model's visual manifold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 295, + 579 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 400, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 472 + ], + "type": "text", + "content": "1. Unsupervised Discovery: The decomposition process should emerge from the intrinsic structure of the model's learned representation, rather than being guided by predefined attributes. This ensures we capture the true topology of the model's knowledge space rather than projecting our assumptions onto it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "2. Semantic Orthogonality: Each discovered control must represent a distinct semantic direction. This is enforced in a semantic feature space, like CLIP, where every slider has an orthogonal effect in embeddings. This prevents discovering multiple controls that create similar semantic effects [15, 33, 47], making the system more efficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": "3. Distribution Consistency: Directions must induce consistent transformations across both random seeds and prompt variations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 652 + ], + "type": "text", + "content": "These requirements naturally lead to our proposed framework, which we formalize in Section 4. As we show in our experiments, SliderSpace is architecture-agnostic, working with both conventional U-Net based models like Stable Diffusion [34, 37, 40, 43, 51] and recent transformer-based architectures like Flux [4]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "We demonstrate the expressiveness of SliderSpace through three applications: First, we show how SliderSpace can decompose high-level concepts into diverse and expressive components, revealing the natural axes of variation in the model's understanding. Second, we explore artistic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "style variation, where SliderSpace discovers directions that match or exceed the diversity of manually curated artist lists while being judged more useful by human evaluators. Finally, we show how SliderSpace can help reverse the mode collapse commonly observed in distilled diffusion models, restoring diversity while maintaining generation speed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 216 + ], + "type": "text", + "content": "Beyond providing practical creative control, SliderSpace opens new avenues for understanding and utilizing the latent capabilities of diffusion models. By mapping these models' visual potential into intuitive, composable directions, we take a step toward making their creative possibilities more accessible and interpretable to users." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 227, + 405, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 227, + 405, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 227, + 405, + 239 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 247, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 342 + ], + "type": "text", + "content": "Recent text-to-image diffusion models have demonstrated remarkable capabilities in generating diverse visual concepts [40]. While newer foundation models enhance text-image alignment through LLM-generated captions [13, 26], the fundamental challenge remains: text-conditioned generation is inherently under-determined, with multiple distinct outputs potentially satisfying the same prompt, making precise control challenging." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "content": "Prior work has explored various approaches to enhance generation control. One direction introduces additional conditioning modalities: spatial signals via adapters (ControlNet [52]), attention-based regional control [8], and image-based conditioning for identity preservation [30, 46, 50] or style transfer via attention manipulation [21]. Another line of research focuses on refining existing images through disentangled text-driven editing [5, 7, 10, 32, 48, 49], where specific attributes are modified while preserving others, often using spatial masks. However, these text-driven approaches inherit the same under-determination challenges as the base models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 629 + ], + "type": "text", + "content": "Notably, disentangled control has been more naturally achieved in GANs [17], particularly through StyleGAN's low-dimensional latent space [28], which exhibits emergent disentanglement even at scale [27]. This has enabled powerful latent space and image editing capabilities [1, 2, 47]. While diffusion models offer superior generation quality and diversity, they lack two key advantages of GANs' latent space: continuous, compositional edits and emergent disentanglement. This limitation prevents users from making serendipitous discoveries about visual variations captured in the training data, instead constraining them to variations they can explicitly describe through prompts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Recent works have explored different approaches to discovering interpretable directions in diffusion models. The weights2weights method [12] learns a manifold of personalized model weights by fine-tuning individual LoRA adapters for each identity and applying PCA to discover a weight space that enables editing. While effective, this requires training separate models per instance. NoiseCLR [9]" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15995" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "learns text embeddings through contrastive learning on a data distribution, but the discovered directions can be arbitrary and lack semantic interpretability. Similarly, Liu et al. [31] propose unsupervised decomposition of images into compositional concepts by learning multiple embeddings, but their approach often yields redundant or non-semantic directions. In contrast, our work directly learns a small set of semantically grounded sliders that decompose the distribution into interpretable and composable directions, enabling infinite creative variations through systematic exploration of the visual manifold." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 288 + ], + "type": "text", + "content": "Concept Sliders [16] addressed continuous control by leveraging LoRA adaptors [24] to learn user-defined attributes. Our work complements this by tackling emergent disentanglement through self-supervised decomposition of the model's inherent variations into composable control dimensions, enabling systematic exploration of the model's creative capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 134, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 134, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 134, + 312 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 192, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 192, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 192, + 330 + ], + "type": "text", + "content": "3.1. Latent Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 407 + ], + "type": "text", + "content": "State-of-the-art text-to-image diffusion models [4, 34, 38, 39] often belong to the class of latent diffusion. Unlike traditional diffusion models that operate in pixel space, latent diffusion models work in a compressed latent space, offering significant computational advantages [40]. The diffusion modelling can be formalized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " be an initial image and " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " be pure Gaussian noise. The forward diffusion process gradually adds noise to the image. The generative process aims to reverse this diffusion, starting from " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " and progressively denoising to reconstruct " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": ". At a timestep " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " the model takes " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " as input and predicts a noise " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": " such that the next step " + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t - 1}" + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 489, + 294, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 489, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 122, + 489, + 294, + 517 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {t}}{\\sqrt {\\alpha_ {t}}} \\tag {1}", + "image_path": "afbf5f8186017bc9e5dfd33c14768a25ac73fb7f0dfaedad9b591abf07f7e773.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "text", + "content": "We can estimate the final image " + }, + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{0,t}" + }, + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "text", + "content": " by taking the same direction " + }, + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 55, + 525, + 295, + 583 + ], + "type": "text", + "content": " for remaining diffusion steps. This can be achieved by recursively applying the denoising from Equation 1 with the same direction. We label this \"Final Image Extrapolation\":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 594, + 294, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 594, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 123, + 594, + 294, + 620 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {0, t} \\leftarrow \\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {2}", + "image_path": "29c22073f9de81cb0ff767de1663bec2140d7c78242acf148c1aae34bca2a8dd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "content": "This enables us to visualize the final image " + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{0,t}" + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "content": " that the diffusion model is planning of at each timestep " + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "content": " without actually running denoising forward passes through all timesteps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 209, + 685 + ], + "type": "text", + "content": "3.2. LoRA: Low Rank Adaptors" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Low-rank adaptor (LoRA) [24] are a class of light-weight adaptors that are attachable to the weights of the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "Given a pre-trained model layer with weights " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", LoRA decomposes the weight update " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 110, + 553, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 110, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 356, + 110, + 553, + 125 + ], + "type": "interline_equation", + "content": "\\Delta W = B A, \\quad B \\in \\mathbb {R} ^ {d \\times r}, A \\in \\mathbb {R} ^ {r \\times k}, \\tag {3}", + "image_path": "91b4e4fd8952df467ac24be511f4538959480dcd0289b92f33b791c1a35d5882.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 134, + 555, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 134, + 555, + 207 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d, k)" + }, + { + "bbox": [ + 313, + 134, + 555, + 207 + ], + "type": "text", + "content": " is a small rank that constrains the update to a low-dimensional subspace. This decomposition allows for efficient parameter updates and has shown success in various downstream tasks. However, the application of LoRA to unsupervised discovery of semantic directions in diffusion model weight space remains unexplored." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 220, + 370, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 220, + 370, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 220, + 370, + 232 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "text", + "content": "We present SliderSpace, a framework for decomposing a diffusion model's visual capabilities into semantically orthogonal control dimensions (Fig. 1). Given a pre-trained text-to-image diffusion model " + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "text", + "content": " and a prompt " + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "text", + "content": ", our goal is to discover " + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 241, + 554, + 314 + ], + "type": "text", + "content": " independent directions that capture the principal modes of variation in the model's learned distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 324, + 440, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 440, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 440, + 335 + ], + "type": "text", + "content": "4.1. Problem Formulation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}(c)" + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": " denote the manifold of possible images that model " + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": " can generate for prompt " + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": ". We aim to identify a set of controllable directions " + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{T}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": " that: (1) span the major modes of variation in " + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}(c)" + }, + { + "bbox": [ + 313, + 342, + 554, + 415 + ], + "type": "text", + "content": ", (2) maintain semantic consistency across different initializations, and (3) are mutually orthogonal in semantic space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "content": "Building on recent advances in model adaptation [16], we formulate each control dimension as a LoRA adaptor [24], updating " + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, n\\}" + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "content": ". These lightweight adapters introduce targeted modifications to the model's cross-attention layers, enabling precise control over specific generative attributes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 544, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 544, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 544, + 511 + ], + "type": "text", + "content": "4.2. SliderSpace: Unsupervised Visual Discovery" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 540 + ], + "type": "text", + "content": "SliderSpace discovery process consists of three key steps as depicted in Figure 2:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": "Distribution Sampling First, we generate a diverse set of samples " + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\{x_{j}\\}_{j = 1}^{m}" + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}(c)" + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": " by varying the random seed (for stability, " + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "inline_equation", + "content": "m\\approx 5000" + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": "). For each sample, we extract the estimated final image " + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{0,t}" + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 561, + 555, + 609 + ], + "type": "text", + "content": " using Eq. 2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": "Semantic Decomposition We map each sample to a semantic embedding space, like CLIP, " + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\phi (\\tilde{\\mathbf{x}}_{0,t})" + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": " and compute the principal components " + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "inline_equation", + "content": "V = \\{v_{i}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 313, + 629, + 554, + 689 + ], + "type": "text", + "content": " of the resulting distribution. These components represent orthogonal directions of maximal variation in semantic space:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "V = \\operatorname {P C A} \\left(\\left\\{\\phi \\left(\\tilde {x} _ {0, t}\\right) \\right\\} _ {j, t}\\right) \\tag {4}", + "image_path": "e93e2ac67eb290bd833069f0e44eb813e8749305d137d6ff535099d792a7daa1.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15996" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 239 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 239 + ], + "type": "image", + "image_path": "44f0c7fc1eed88d6179270d33846652e8a5f5ecd3507b6edd6b63162736fa5e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Given a prompt, SliderSpace first generates images and extracts the CLIP features. We then compute the spectral decomposition of the CLIP features and align the each slider with extracted principle components. Each slider is therefore trained to represent a unique semantic direction that are relevant in the diffusion model's knowledge of the prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "text", + "content": "Slider Training For each principal direction " + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "text", + "content": ", we train a corresponding adapter " + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "text", + "content": " to induce transformations that align with " + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 54, + 300, + 298, + 349 + ], + "type": "text", + "content": " in semantic CLIP space. The training objective for each slider is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 361, + 295, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 361, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 295, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s l i d e r s p a c e}} = \\sum_ {i = 1} ^ {n} 1 - \\cos \\left(\\Delta \\phi_ {i}, v _ {i}\\right), \\tag {5}", + "image_path": "e4617108ca783669c346097b37792fe13d74f91f4692a8bc59fd97f8ab2e2e05.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\Delta \\phi_{i} = \\phi (\\tilde{\\mathbf{x}}_{0,t}^{i}) - \\phi (\\tilde{\\mathbf{x}}_{0,t})" + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "content": " represents the transformation induced by slider " + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "content": " in embedding space and " + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{0,t}^{i}" + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "content": " is the estimated final image with the slider " + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 54, + 409, + 295, + 506 + ], + "type": "text", + "content": " attached. This ensures that each slider's effect is semantically aligned with its corresponding principal direction while maintaining orthogonality with other sliders, building on established work showing that embedding differences encode semantic relationships [15, 33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Our formulation satisfies the three key requirements outlined in Section 1. First, unsupervised discovery is achieved by deriving directions directly from the model's intrinsic variation through PCA in a semantic embedding space, without imposing predefined attributes or external supervision. Second, distribution consistency is enforced by our training objective " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sliderspace}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ", which ensures each slider's transformation maintains consistent direction in embedding space across different seeds and timesteps. Finally, semantic orthogonality is guaranteed through the PCA-based initialization of directions as each principal components are mutually orthogonal, ensuring each slider captures a distinct mode of variation. Together, these components enable the discovery of interpretable and reliable control dimensions that effectively decompose the model's learned distribution. In majority of this paper, we use CLIP [36] as our primary semantic encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 299, + 462, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 462, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 462, + 312 + ], + "type": "text", + "content": "4.3. Interpretability & Control" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 401 + ], + "type": "text", + "content": "SliderSpace provides a dual contribution: it serves as both a framework for discovering expressive dimensions of control and as a mechanism for decomposing the model's learned concept space into semantically meaningful components. Each adapter functions as a \"slider\" controlling a specific attribute of the generated image, enabling fine-grained manipulation of the output while maintaining semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 450 + ], + "type": "text", + "content": "The resulting set of adapters provides valuable insights into the model's conceptual understanding, revealing nuanced semantic relationships that may not be immediately apparent from the text prompt alone." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 451, + 555, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 547 + ], + "type": "text", + "content": "Moreover, the low-rank structure of our adapters ensures computational efficiency and minimal memory overhead, making them particularly suitable for real-time interactive applications. This enables users to explore the concept space dynamically by modulating the influence of each adapter, facilitating intuitive creative control over the image generation process while maintaining the underlying semantic integrity of the original prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 560, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 395, + 574 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We conduct our main experiments to evaluate SliderSpace using SDXL-DMD [51], a 4-step distilled diffusion model. Our implementation requires less than 24GB VRAM and can discover 64 semantic directions in under 2 hrs on a single A100 GPU. When concepts exhibit severe mode collapse, the discovery process can be enhanced by generating data using undistilled base models or LLM-expanded prompts for increased sample diversity. We demonstrate SliderSpace's generalization to SDXL [34], SDXL-Turbo [43], and transformer-based FLUX Schnell [4] in appendix. Our analysis focuses on three key applications:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15997" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 67, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 56, + 67, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 56, + 67, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 67, + 553, + 261 + ], + "type": "image", + "image_path": "cfc3a3fc362612dc20df1e7364ec5bb07150c7ef2d43349f054c0a37e817dde8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 304 + ], + "type": "text", + "content": "Figure 3. SliderSpace decomposes the visual variation of diffusion model's knowledge corresponding to a concept. These directions can be perceived as interpretable directions of the model's hierarchical knowledge. We show the decomposed slider direction for a concept using SliderSpace and the corresponding labels generated by Claude 3.5 Sonnet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 294, + 347 + ], + "type": "text", + "content": "concept decomposition, art styles exploration, and diversity enhancement in distilled models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 191, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 191, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 191, + 372 + ], + "type": "text", + "content": "5.1. Concept Decomposition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 437 + ], + "type": "text", + "content": "We first demonstrate how SliderSpace can serve as an exploratory tool by decomposing high-level concepts into semantic directions that align with the diffusion model's internal representations. Summarizing and exposing the dominant variations a model is capable of for a particular prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 715 + ], + "type": "text", + "content": "Given a concept prompt (e.g., \"picture of a monster\"), we discover SliderSpace directions. Figure 3 shows discovered sliders for \"Monster\", and \"car\" concepts. We label the sliders using Claude 3.5 Sonnet [3] by showing multiple image pairs showing the effect of each slider and prompting to identify the semantic transformation being applied. Through these discovered directions, we demonstrate the manipulation of individual attributes. As these directions are intended to capture the model's visual possibilities for a prompt, we naturally ask the question, \"How much variation is enabled by these directions?\" To quantitatively evaluate the diversity enabled by SliderSpace against base model, we compute DreamSim [14] distance to measure inter-image variation across 2500 generated samples per concept (Figure 4). For SliderSpace-augmented generation, we generate the images by randomly activating a sparse subset of 3 sliders (out of 32 discovered directions) for each generation. Our analysis reveals that SliderSpace-generated samples exhibit significantly higher inter-image diversity compared to the baseline model outputs. We measure the CLIP-Score [22] between the input prompts and the generated images to measure text-alignment with the prompt. We find that they have similar CLIP Scores to the images gen" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 322, + 555, + 377 + ], + "blocks": [ + { + "bbox": [ + 315, + 322, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 315, + 322, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 555, + 377 + ], + "type": "table", + "html": "
Method vs.User Study (Win Rate %)
“Diverse”“Useful”“Creative”
SDXL-DMD72.466.068.1
LLM + SDXL-DMD62.562.562.5
SDXL65.361.259.2
", + "image_path": "c5c8d0d75537cb3083682bf3ff2b9a4efc6e6b08683bc2c617abcc402acbc72f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 420 + ], + "type": "text", + "content": "Table 1. Users perceive SliderSpace generated images to be more diverse, useful, and interesting. We show win-rate percentages of SliderSpace samples against baselines." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 443, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 540 + ], + "type": "text", + "content": "erated by the base model. This suggests that SliderSpace effectively expands the achievable variation in the model's knowledge, while maintaining semantic consistency. To validate our findings through human perception, we conducted pairwise comparisons of image grids generated by SliderSpace versus baseline methods. As shown in Table 1, users consistently preferred SliderSpace outputs across diversity, utility, and creative potential." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 550, + 443, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 443, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 443, + 563 + ], + "type": "text", + "content": "5.2. Art Styles Exploration" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 676 + ], + "type": "text", + "content": "We also evaluate to what degree SliderSpace can expose \"all\" of the art styles that the diffusion model has learned from its training data. As a proxy for \"all\", we use a diverse set of art styles, manually discovered and documented by ParrotZone [25]. We explore the visual artistic space by decomposing the prompt \"artwork in the style of a famous artist\" into 64 directions. This process enables us to discover the SliderSpace of art and create a comprehensive dictionary of discoverable art styles in a diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To compare how this measures up against supervised methods like Concept Slider [16], we train 64 manually curated concept sliders using LLM generated training" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15998" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 276, + 163 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 276, + 163 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 276, + 163 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 276, + 163 + ], + "type": "image", + "image_path": "76d2d6e63c6f42822c0d0e4e0d28f217dc37b7fb50991f881a8a975aa8bd0e0b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 555, + 200 + ], + "type": "text", + "content": "Figure 4. (left) SliderSpace generates diverse variations of a concept, as measured by DreamSim [14] distance across generated samples (higher is better). (right) The method maintains similar text-to-image alignment, as measured by CLIP Scores [22] (lower is better)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 286, + 70, + 547, + 164 + ], + "blocks": [ + { + "bbox": [ + 286, + 70, + 547, + 164 + ], + "lines": [ + { + "bbox": [ + 286, + 70, + 547, + 164 + ], + "spans": [ + { + "bbox": [ + 286, + 70, + 547, + 164 + ], + "type": "image", + "image_path": "d1632b0ce9a286263773b7ce8a092ed7bfd9ef2bfd506bafbdf3ccfb7c9f7df8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 220, + 297, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 297, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 297, + 483 + ], + "type": "text", + "content": "prompts. We compute FID scores against the generated samples conditioned on actual artist names from ParrotZone [25] dataset. The dataset contains 4388 artists mimicked by SDXL, are discovered through exhaustive manual search. We also establish two other baselines. First, we use GPT/Claude to generate 4388 different artistic style names (e.g., \"cubism\", \"Warhol\"). Second, we use a generic prompt like \"[subject] in the style of a famous artist\". Using our discovered SliderSpace, we randomly sample a sparse set of 3 sliders to generate an equivalent number of images. Specifically, for each baseline, we match the provided samples of [25] by creating two images per style for \"building landscape\" and \"character portraits\" each (total of four images per style) to ensure same semantic structure across the datasets and that FID measures the artistic spread. Figure 5 shows random, non-cherry picked images from all the methods and their corresponding FID scores. We find that SliderSpace generates a distribution significantly more closely matches the manually curated artist names than the baselines, including supervised methods like Concept Sliders. Figure 6 shows qualitative comparison of art styles discovered by SliderSpace and styles discovered manually [25]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 628 + ], + "type": "text", + "content": "To validate the practical utility of our \"artistic\" SliderSpace, we conduct a user study examining both usefulness and diversity of generated samples (Table 2). We conduct comparative study with 2 grids of 9 images (total of 1000 pairs), one with slider generated images and other with baselines. Users find the SliderSpace-generated images are far more diverse and useful than the baseline methods. We also find that users would prefer using SliderSpace images over images prompted with the real artist names [25], while finding these both to be equally diverse. We provide user-study details and qualitative examples in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 188, + 648 + ], + "type": "text", + "content": "5.3. Diversity Enhancement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Finally, we explore SliderSpace's ability to addressing mode collapse in distilled models. Instead of discovering SliderSpace for narrow distributions, we train it on a larger spread of 8000 randomly selected COCO-30k prompts and enhance them with LLM. We use DMD model for gener" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 218, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 323, + 218, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 323, + 218, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 323, + 218, + 547, + 275 + ], + "type": "table", + "html": "
MethodUser study (win rate %)
“Diverse”“Useful”
vs. Real Artist Prompts [25]5463
vs. LLM Prompts7367
vs. Generic Prompts8788
", + "image_path": "c0653ba4f60d849fe0066aedf533e042ce056e7fa1b909c4f7ec42917cbafb3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 341, + 535, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 328 + ], + "type": "text", + "content": "Table 2. Pairwise comparison user study (win rates in %) reveals that SliderSpace-extracted artistic styles achieve comparable diversity to manually curated artist [25], while being significantly preferred for creative utilization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 341, + 535, + 403 + ], + "lines": [ + { + "bbox": [ + 334, + 341, + 535, + 403 + ], + "spans": [ + { + "bbox": [ + 334, + 341, + 535, + 403 + ], + "type": "table", + "html": "
MethodFID-30k (↓)CLIP (↑)
Real-30.14
SDXL11.7229.41
SDXL-DMD15.5228.92
DMD-SliderSpace12.1229.13
", + "image_path": "745b04bd8702771bfe81143ae10479a54c7a85a946a896f07bd8e029a9aaf6bf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 556, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 445 + ], + "type": "text", + "content": "Table 3. SliderSpace trained on larger range of knowledge can improve the diversity of the distilled models (FID) while having a good text-image alignment (CLIP)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 468, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 599 + ], + "type": "text", + "content": "ating training dataset since we wish to undo mode collapse of DMD models. We discover a generic SliderSpace of 64 sliders that captures the generally applicable visual variations within SDXL-DMD. Figure 7 showcases qualitative examples of two distributions for the prompts \"Car driving through a forest\" and \"Picture of a person\". We demonstrate how generating images by randomly sampling from these SliderSpace directions effectively increases output diversity. This demonstrates that the mode collapse in distilled model can be reversed by discovering and exploring the visual structure. We show more examples in appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 601, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 661 + ], + "type": "text", + "content": "In Table 3, we evaluate this improvement by measuring FID and CLIP scores on all COCO-30k prompts before and after applying SliderSpace to SDXL-DMD. We find that DMD-SliderSpace improves FID from the distilled version, almost matching the FID of undistilled SDXL." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 670, + 440, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 440, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 440, + 684 + ], + "type": "text", + "content": "5.4. Slider Transferability" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Using FaceNet [44] as the semantic encoder, we train SliderSpace on the concept \"person\" and discover interpretable" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15999" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 81, + 154, + 176 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 124, + 78 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 124, + 78 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 124, + 78 + ], + "type": "text", + "content": "Real Artists" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 81, + 154, + 176 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 154, + 176 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 154, + 176 + ], + "type": "image", + "image_path": "d80110c9f12f3f9954d4b3a4ec9f41c4444d3671cf8287be8d0006894e6e0897.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 178, + 111, + 188 + ], + "lines": [ + { + "bbox": [ + 100, + 178, + 111, + 188 + ], + "spans": [ + { + "bbox": [ + 100, + 178, + 111, + 188 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. SliderSpace demonstrates broader artistic style coverage, as evidenced by the lower FID scores compared to both supervised Concept Sliders. Comparison of artistic style diversity using FID scores against reference distribution (a) derived from the complete artist dataset [25]. We compare against outputs from (b) generic art prompts (b), (c) LLM-generated art prompts, and (d) Concept Sliders." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 157, + 80, + 254, + 176 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 250, + 79 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 250, + 79 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 250, + 79 + ], + "type": "text", + "content": "Generic Prompts (FID: 76.08)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 157, + 80, + 254, + 176 + ], + "lines": [ + { + "bbox": [ + 157, + 80, + 254, + 176 + ], + "spans": [ + { + "bbox": [ + 157, + 80, + 254, + 176 + ], + "type": "image", + "image_path": "d1a168bb77bc4386aed65278bcc123c9fbf4647d0930c6770081aa810aa80c8c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 178, + 211, + 188 + ], + "lines": [ + { + "bbox": [ + 200, + 178, + 211, + 188 + ], + "spans": [ + { + "bbox": [ + 200, + 178, + 211, + 188 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 257, + 80, + 354, + 176 + ], + "blocks": [ + { + "bbox": [ + 264, + 70, + 346, + 79 + ], + "lines": [ + { + "bbox": [ + 264, + 70, + 346, + 79 + ], + "spans": [ + { + "bbox": [ + 264, + 70, + 346, + 79 + ], + "type": "text", + "content": "LLM Prompts (FID: 38.84)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 80, + 354, + 176 + ], + "lines": [ + { + "bbox": [ + 257, + 80, + 354, + 176 + ], + "spans": [ + { + "bbox": [ + 257, + 80, + 354, + 176 + ], + "type": "image", + "image_path": "047d5a705de86afb97ece042b464f4fbd570cb44cb28e556600f9b06a417a557.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 178, + 310, + 188 + ], + "lines": [ + { + "bbox": [ + 300, + 178, + 310, + 188 + ], + "spans": [ + { + "bbox": [ + 300, + 178, + 310, + 188 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 357, + 80, + 454, + 176 + ], + "blocks": [ + { + "bbox": [ + 362, + 70, + 448, + 79 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 448, + 79 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 448, + 79 + ], + "type": "text", + "content": "Concept Sliders (FID: 32.86)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 80, + 454, + 176 + ], + "lines": [ + { + "bbox": [ + 357, + 80, + 454, + 176 + ], + "spans": [ + { + "bbox": [ + 357, + 80, + 454, + 176 + ], + "type": "image", + "image_path": "978c80621cd18437b6128771157752b3c98c675a3bc3485b4d0f4bb979c6d1e3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 178, + 411, + 188 + ], + "lines": [ + { + "bbox": [ + 400, + 178, + 411, + 188 + ], + "spans": [ + { + "bbox": [ + 400, + 178, + 411, + 188 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 457, + 80, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 542, + 79 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 542, + 79 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 542, + 79 + ], + "type": "text", + "content": "SliderSpace (FID: 19.12)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 80, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 457, + 80, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 457, + 80, + 553, + 176 + ], + "type": "image", + "image_path": "88c7521209cf47f21218615b7bc43de90d267b42c4a618ed474d825c41f98143.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 178, + 510, + 188 + ], + "lines": [ + { + "bbox": [ + 500, + 178, + 510, + 188 + ], + "spans": [ + { + "bbox": [ + 500, + 178, + 510, + 188 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 57, + 267, + 176, + 345 + ], + "blocks": [ + { + "bbox": [ + 88, + 257, + 145, + 266 + ], + "lines": [ + { + "bbox": [ + 88, + 257, + 145, + 266 + ], + "spans": [ + { + "bbox": [ + 88, + 257, + 145, + 266 + ], + "type": "text", + "content": "Elenore Abbott" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 267, + 176, + 345 + ], + "lines": [ + { + "bbox": [ + 57, + 267, + 176, + 345 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 176, + 345 + ], + "type": "image", + "image_path": "498ef8e9a906b5bd57e11f713f202600ca486ca0643c891bfb42945ec28d06e1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 346, + 146, + 355 + ], + "lines": [ + { + "bbox": [ + 86, + 346, + 146, + 355 + ], + "spans": [ + { + "bbox": [ + 86, + 346, + 146, + 355 + ], + "type": "text", + "content": "Wayne Barlowe" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 58, + 356, + 176, + 435 + ], + "blocks": [ + { + "bbox": [ + 58, + 356, + 176, + 435 + ], + "lines": [ + { + "bbox": [ + 58, + 356, + 176, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 356, + 176, + 435 + ], + "type": "image", + "image_path": "9ad2537634c15198edf91fe3acc1d43bd4f7c7c0a2df506033b8e3428e16f351.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 555, + 477 + ], + "type": "text", + "content": "Figure 6. Comparison of artistic styles discovered in SDXL-DMD: (Left) Representative samples from artist-specific prompts manually curated by the Parrotzone community [25] through extensive exploration. (Right) Automatically discovered artistic directions using SliderSpace, which captures diverse and semantically meaningful variations without requiring explicit artist references." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 181, + 266, + 299, + 345 + ], + "blocks": [ + { + "bbox": [ + 96, + 240, + 260, + 252 + ], + "lines": [ + { + "bbox": [ + 96, + 240, + 260, + 252 + ], + "spans": [ + { + "bbox": [ + 96, + 240, + 260, + 252 + ], + "type": "text", + "content": "Manually Discovered Styles by ParrotZone" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 257, + 259, + 266 + ], + "lines": [ + { + "bbox": [ + 221, + 257, + 259, + 266 + ], + "spans": [ + { + "bbox": [ + 221, + 257, + 259, + 266 + ], + "type": "text", + "content": "Miki Asai" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 266, + 299, + 345 + ], + "lines": [ + { + "bbox": [ + 181, + 266, + 299, + 345 + ], + "spans": [ + { + "bbox": [ + 181, + 266, + 299, + 345 + ], + "type": "image", + "image_path": "4304f4149c1af9f75e2daa39d59232faeb0ecaa09490372d2460d93d38b201ab.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 181, + 356, + 299, + 435 + ], + "blocks": [ + { + "bbox": [ + 212, + 346, + 268, + 356 + ], + "lines": [ + { + "bbox": [ + 212, + 346, + 268, + 356 + ], + "spans": [ + { + "bbox": [ + 212, + 346, + 268, + 356 + ], + "type": "text", + "content": "Edward Gorey" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 356, + 299, + 435 + ], + "lines": [ + { + "bbox": [ + 181, + 356, + 299, + 435 + ], + "spans": [ + { + "bbox": [ + 181, + 356, + 299, + 435 + ], + "type": "image", + "image_path": "95e0fc2f3b6c5cd7e1a4c1128b35b2674f2294f20f2c003ce470933f73d2f7d7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 310, + 266, + 430, + 345 + ], + "blocks": [ + { + "bbox": [ + 340, + 240, + 523, + 252 + ], + "lines": [ + { + "bbox": [ + 340, + 240, + 523, + 252 + ], + "spans": [ + { + "bbox": [ + 340, + 240, + 523, + 252 + ], + "type": "text", + "content": "Automatically Discovered Styles by SliderSpace" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 257, + 391, + 266 + ], + "lines": [ + { + "bbox": [ + 349, + 257, + 391, + 266 + ], + "spans": [ + { + "bbox": [ + 349, + 257, + 391, + 266 + ], + "type": "text", + "content": "Direction 1" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 266, + 430, + 345 + ], + "lines": [ + { + "bbox": [ + 310, + 266, + 430, + 345 + ], + "spans": [ + { + "bbox": [ + 310, + 266, + 430, + 345 + ], + "type": "image", + "image_path": "dc4a9b33d232f45b5091c94c841b9bda05bcf6f252491808d13d222046d9fffe.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 310, + 355, + 430, + 435 + ], + "blocks": [ + { + "bbox": [ + 347, + 346, + 394, + 355 + ], + "lines": [ + { + "bbox": [ + 347, + 346, + 394, + 355 + ], + "spans": [ + { + "bbox": [ + 347, + 346, + 394, + 355 + ], + "type": "text", + "content": "Direction 25" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 355, + 430, + 435 + ], + "lines": [ + { + "bbox": [ + 310, + 355, + 430, + 435 + ], + "spans": [ + { + "bbox": [ + 310, + 355, + 430, + 435 + ], + "type": "image", + "image_path": "3bbce1fb0c35f38f87b891054cfb07c24799191f60e9f8bfda8b403a8ad9787a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 434, + 266, + 553, + 345 + ], + "blocks": [ + { + "bbox": [ + 470, + 257, + 518, + 266 + ], + "lines": [ + { + "bbox": [ + 470, + 257, + 518, + 266 + ], + "spans": [ + { + "bbox": [ + 470, + 257, + 518, + 266 + ], + "type": "text", + "content": "Direction 23" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 266, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 434, + 266, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 434, + 266, + 553, + 345 + ], + "type": "image", + "image_path": "a81b70628fccbdbca3d7f116f7373cc9b6683b1cc913c08939e01b3b5c480d85.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 346, + 517, + 355 + ], + "lines": [ + { + "bbox": [ + 470, + 346, + 517, + 355 + ], + "spans": [ + { + "bbox": [ + 470, + 346, + 517, + 355 + ], + "type": "text", + "content": "Direction 31" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 434, + 356, + 553, + 435 + ], + "blocks": [ + { + "bbox": [ + 434, + 356, + 553, + 435 + ], + "lines": [ + { + "bbox": [ + 434, + 356, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 434, + 356, + 553, + 435 + ], + "type": "image", + "image_path": "4e1ba86a094f362ebb0f7069db2c523607f1554b9b1fa9726b41003df294a020.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 570 + ], + "type": "text", + "content": "directions controlling attributes like appearance and age (Figure 8). These directions not only transfer effectively to related concepts like \"police\" and \"athlete\" but also generalize surprisingly well to out-of-domain concepts like \"dog\", suggesting that SliderSpace captures fundamental visual transformations in the model's knowledge space." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 607 + ], + "type": "text", + "content": "Additional experiments analyzing hyperparameter choices (App. B.2), alternative semantic embeddings (App. B.1), and ablation studies (App. E) are in Appendix." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 620, + 130, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 130, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 130, + 632 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Our method can work with several semantic embeddings - each encoder might introduce different inherent biases present in its training data. While these embeddings enable semantic consistency, they may not capture certain culturally-specific or nuanced concepts. This highlights the need for application specific semantic embedding choice." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": "The current discovery process requires " + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\approx 65" + }, + { + "bbox": [ + 313, + 498, + 555, + 605 + ], + "type": "text", + "content": " mins on A100. We note that our method trains 8 times faster than Concept Sliders for the same number of sliders. To further optimize, it opens avenues for future research into training time optimizations. For art style discovery, the discovered directions are not one-to-one matched with the real artists, this can be limiting based on users' application. Further work can address discovery that nudges the directions to be aligned with real artists." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 388, + 632 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "SliderSpace is a simple framework that automatically decomposes diffusion models' capabilities into semantically meaningful and controllable directions. By leveraging spectral decomposition in semantic space combined with low-rank adaptation, our method enables systematic exploration of a model's latent creative space without requiring manual" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16000" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 233 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 233 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 233 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 233 + ], + "type": "image", + "image_path": "eb754555c6b0f0f77b2abd71066bc942065255aa3a1317903cdabee3daeec71b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "type": "text", + "content": "Figure 7. SliderSpace can also decompose general broad visual variation of diffusion model's and can be used to overcome model collapse in distilled models. We randomly sample a sparse set of sliders and generate the sample showing higher variation than base distilled model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 272, + 542, + 541 + ], + "blocks": [ + { + "bbox": [ + 70, + 272, + 542, + 541 + ], + "lines": [ + { + "bbox": [ + 70, + 272, + 542, + 541 + ], + "spans": [ + { + "bbox": [ + 70, + 272, + 542, + 541 + ], + "type": "image", + "image_path": "5fba636fa30ffefdd9cf8b55832fc124f5f59891838834dc2629b516e7cf6cb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 549, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 549, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 555, + 572 + ], + "type": "text", + "content": "Figure 8. SliderSpace directions for the \"person\" concept successfully generalize to related \"police\" and \"athlete\" concepts. They also transfer to out-of-domain concepts like \"dog\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 593, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 297, + 713 + ], + "type": "text", + "content": "attribute specification. Through extensive experiments, we demonstrated SliderSpace's effectiveness across three key applications. First, our concept decomposition revealed interpretable variations within the model's knowledge representation, enabling fine-grained control while maintaining semantic consistency. Second, our exploration of artistic capabilities showed that SliderSpace can discover directions matching or exceeding the diversity of manually curated artist lists, while being rated more useful by human evaluators. Finally, we demonstrate how SliderSpace can help" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 593, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 617 + ], + "type": "text", + "content": "address mode collapse in distilled diffusion models, restoring diversity while preserving computational efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "The ability of SliderSpace to uncover interpretable directions suggests that diffusion models may develop structured internal representations of visual concepts during training, without explicit supervision. By mapping these models' vast creative potential into intuitive, composable directions, our work takes a step toward making their capabilities more transparent and accessible." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16001" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 128 + ], + "type": "text", + "content": "This work was done by RG at Adobe during their internship. RG and DB are supported by Open Philanthropy and NSF grant #2403304." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 114, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 204 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF international conference on computer vision, pages 4432-4441, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "text", + "content": "[2] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Styleflow: Attribute-conditioned exploration of stylegan-generated images using conditional continuous normalizing flows. ACM Transactions on Graphics (ToG), 40(3):1-21, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 269, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 269, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 269, + 272 + ], + "type": "text", + "content": "[3] Anthropic. Introducing claude 3.5 sonnet, 2024. 5, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 295 + ], + "type": "text", + "content": "[4] BlackForestLabs. Announcing state-of-the-art flux.1 dev and Schnell models, 2024. 1, 2, 3, 4, 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 361 + ], + "type": "text", + "content": "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 364, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 397 + ], + "type": "text", + "content": "[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv e-prints, pages arXiv-1809, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "text", + "content": "[7] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "type": "text", + "content": "[8] Anthony Chen, Jianjin Xu, Wenzhao Zheng, Gaole Dai, Yida Wang, Renrui Zhang, Haofan Wang, and Shanghang Zhang. Training-free regional prompting for diffusion transformers. arXiv preprint arXiv:2411.02395, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "type": "text", + "content": "[9] Yusuf Dalva and Pinar Yanardag. Noiseclar: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24209-24218, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "type": "text", + "content": "[10] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "text", + "content": "[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[12] Amir Dravid, Yossi Gandelsman, Kuan-Chieh Wang, Rameen Abdal, Gordon Wetzstein, Alexei A Efros, and Kfir Aberman. Interpreting the weight space of customized diffusion models. arXiv preprint arXiv:2406.09413, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "In Forty-first International Conference on Machine Learning, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[14] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[15] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[16] Rohit Gandikota, Joanna Materzyńska, Tingrui Zhou, Antonio Torralba, and David Bau. Concept sliders: Lora adaptors for precise control in diffusion models. In European Conference on Computer Vision, pages 172-188. Springer, 2024. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[18] Imke Grabe, Miguel González-Duque, Sebastian Risi, and Jichen Zhu. Towards a framework for human-ai interaction patterns in co-creative gan applications. In Joint Proceedings of the ACM IUI Workshops, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "text", + "content": "[19] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in neural information processing systems, 33:9841-9850, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 430 + ], + "type": "text", + "content": "[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 486 + ], + "type": "text", + "content": "[21] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "type": "text", + "content": "[22] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 588 + ], + "type": "text", + "content": "[23] Oliver Hoffmann. On modeling human-computer cocreativity. In Knowledge, Information and Creativity Support Systems: Selected Papers from KICSS'2014-9th International Conference, held in Limassol, Cyprus, on November 6-8, 2014, pages 37-48. Springer, 2016. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 656 + ], + "type": "text", + "content": "[25] Surea I, Proxima Centauri B, Erratica, and Stephen Young. Image synthesis style studies, 2022. 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[26] et al James Betker. Improving image generation with better captions. OpenAI Reports, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "type": "text", + "content": "[27] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16002" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[28] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 196 + ], + "type": "text", + "content": "[29] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "text", + "content": "[30] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 308 + ], + "type": "text", + "content": "[31] Nan Liu, Yilun Du, Shuang Li, Joshua B Tenenbaum, and Antonio Torralba. Unsupervised compositional concepts discovery with text-to-image generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2085-2095, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "text", + "content": "[32] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "text", + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2085–2094, 2021. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "text", + "content": "[34] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 4, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 498 + ], + "type": "text", + "content": "[35] Alec Radford. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 564 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 273, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 273, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 273, + 577 + ], + "type": "text", + "content": "[37] Robin Rombach. Stable diffusion 2.0 release, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 599 + ], + "type": "text", + "content": "[38] Robin Rombach and Patrick Esser. Stable diffusion v1-4 model card, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 622 + ], + "type": "text", + "content": "[39] Robin Rombach and Patrick Esser. Stable diffusion v2 model card, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and BjA¶rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[41] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 599 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[42] Simo Ryu. Cloneofsimo/lora: Using low-rank adaptation to quickly fine-tune diffusion models.s. GitHub, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[43] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 2, 4, 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[44] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[45] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE transactions on pattern analysis and machine intelligence, 44(4):2004-2018, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[46] Jing Shi, Wei Xiong, Zhe Lin, and Hyun Joon Jung. Instantbooth: Personalized text-to-image generation without test-time finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8543-8552, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[47] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863-12872, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[48] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. Turboedit: Instant text-based image editing. ECCV, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[49] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[50] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 553 + ], + "type": "text", + "content": "[51] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 2, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 599 + ], + "type": "text", + "content": "[52] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16003" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_content_list.json b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06928221941b2535b1b528a433c911782d70824e --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_content_list.json @@ -0,0 +1,1382 @@ +[ + { + "type": "text", + "text": "SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ahmed Nassar Matteo Omenetti Maksym Lysak Nikolaos Livathinos Christoph Auer Lucas Morin \nRafael Teixeira de Lima Yusik Kim A. Said Gurbuz Michele Dolfi Peter W. J. Staar \n{ahn, mly, nli, cau, dol, taa}@zurich.ibm.com, {Matteo.Omenetti1}@ibm.com \nIBM Research", + "bbox": [ + 99, + 203, + 919, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms — significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition. Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model weights and datasets are available at: https://huggingface.co/ds4sd", + "bbox": [ + 88, + 342, + 485, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For decades, converting complex digital documents into a structured, machine-processable format has been a significant technical challenge. This challenge primarily stems from the substantial variability in document layouts and styles, as well as the inherently opaque nature of the widely used PDF format, which is optimized for printing rather than semantic parsing. Intricate layout styles and visually challenging elements such as forms, tables, and complex charts can significantly impact the reading order and gen", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eral understanding of documents. These problems have driven extensive research and development across multiple domains of computer science. On one hand, sophisticated ensemble systems emerged, which decompose the conversion problem into several sub-tasks (e.g. OCR, layout analysis, table structure recognition, classification) and tackle each sub-task independently. Although such systems can achieve high-quality results for many document types while maintaining relatively low computational demands, they are often difficult to tune and generalize.", + "bbox": [ + 511, + 310, + 906, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, in the recent past, great interest has developed around large foundational multimodal models that can solve the whole conversion task in one shot, while simultaneously offering flexible querying and parametrization through prompts. This approach has been made possible by the advent of multimodal pre-training of large vision-language models (LVLMs), which opened up a vast array of opportunities for leveraging diverse data sources, including PDF documents. However, the literature on this topic highlights a significant gap in the availability of high-quality and open-access datasets suitable for training robust multimodal models for the task of document understanding. Furthermore, relying on LVLMs may introduce common issues associated with such models, including hallucinations and the use of significant computational resources, making them impractical from both a quality and cost perspective.", + "bbox": [ + 511, + 460, + 908, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we outline how we close the gaps left by publicly available datasets and establish a training approach to achieve end-to-end, full-featured document conversion through a vision-language model, which we name SmolDoccling. Our main contributions can be summarized as follows:", + "bbox": [ + 511, + 702, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "- SmolDocling: An ultra-compact VLM for end-to-end document conversion, based on Hugging Face's most recent SmolVLM-256M [53], which is between 5 and 10 times smaller in parameters than comparable VLMs tuned on document understanding tasks. Hence, we reduce the computational complexity to roughly the same order of magnitude as typical representatives of ensemble meth", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21972", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0211634d0b93a26372b178c2ce62ccc68ddba7148f16e50aafb340745dbdf4fb.jpg", + "image_caption": [ + "Figure 1. SmolDocling/SmolVLM architecture. SmolDocling converts images of document pages to DocTags sequences. First, input images are encoded using a vision encoder and reshaped via projection and pooling. Then, the projected embeddings are concatenated with the text embeddings of the user prompt, possibly with interleaving. Finally, the sequence is used by an LLM to autoregressively predict the DocTags sequence." + ], + "image_footnote": [], + "bbox": [ + 169, + 85, + 831, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods. Uniquely, SmolDocling learns and produces a unified document representation that captures content, structure and spatial location of all elements in a page.", + "bbox": [ + 102, + 373, + 482, + 417 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We augment existing document pre-training datasets with additional feature annotations, and create new datasets for tasks with insufficient coverage in open multi-modal training datasets, including code listings, equations, and chart understanding. Importantly, we construct full-page ground truth data that combines key annotation features such as layout, table structure, code listings, charts, and formulas in their natural context, and derive instruction datasets from all training datasets used.", + "- We introduce DocTags, a document markup format optimized to efficiently represent the full content and layout characteristics of a document, inspired by [52].", + "- We evaluate the achieved prediction quality of SmolDoccling on several tasks, including comparisons to popular community models." + ], + "bbox": [ + 89, + 419, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 657, + 232, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Vision-Language Models (LVLMs)", + "text_level": 1, + "bbox": [ + 89, + 681, + 442, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advent of large language models (LLMs) [2, 4, 8, 25, 31, 61, 67, 73, 80] characterized by their powerful contextual understanding and adaptability, several approaches have emerged to extend them into large vision-language models. Some proprietary, closed-source LLMs have been adapted into multimodal large language models (MLLMs), such as GPT-4o [60], Gemini [71], and Claude 3.5 [5], demonstrating exceptional capabilities across various modalities, including vision. Open-source methods are actively advancing to match the capabilities of proprietary MLLMs. BLIP-2 [44] is one of the earliest works combining a vision encoder with a frozen LLM (OPT [91] or FlanT5 [14]) using a lightweight transformer (Q-former", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[90]). Building upon BLIP-2, MiniGPT-4 [94] integrated a frozen visual encoder with a frozen LLM (Vicuna [76]) using a Q-Former network and a single projection layer. LLaVA [48, 49] employs a similar architecture but uses a minimal adapter layer. LLaVA-OneVision [42] and LLaVA-NeXT-Interleave [43] build upon LLaVA to support multi-image, higher resolution, and video understanding. For handling high resolutions and text-image compositions and comprehension, InternLM-XComposer introduces a family of models [19, 20, 88, 89] designed specifically for these tasks. Qwen-VL [8] introduces a position-aware adapter to address efficiency issues caused by the vision encoder generating long image feature sequences. Qwen2.5-VL [9] employs windowed attention with 2D rotary positional embeddings to process native-resolution inputs efficiently and integrates an vision-language merger for dynamic feature compression.", + "bbox": [ + 511, + 372, + 906, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we build upon the SmolVLM architecture approach [53]. Precisely, we adopt SmolVLM-256M, a compact variant of base SmolVLM-2.2B model, inspired by the architecture of Idefics3 [38] which employs the shape-optimized SigLIP encoder as its visual backbone. It applies an aggressive pixel shuffle strategy to compress visual features, reduces the number of image hidden states and introduces new special tokens to separate sub-images and improve tokenization efficiency.", + "bbox": [ + 511, + 628, + 908, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. State of the Art in Document Understanding", + "text_level": 1, + "bbox": [ + 511, + 772, + 890, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several powerful document understanding solutions have emerged on the market in the past decade as commercial cloud offerings on hyperscalers [3, 6, 24, 59], frontier models such as GPT-4o [60] or Claude [5], or as open-source libraries such as Docling [50], Grobid [1], Marker [62], MinerU [77] or Unstructured [75]. Typical tasks that document understanding encompasses are document classifi", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21973", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cation [29, 84], OCR [45], layout analysis [65, 85], table recognition [52, 69, 92], key-value extraction [29, 30, 34], chart understanding [13, 26, 32, 33, 46, 47, 54, 55, 57, 82, 83], equation recognition [17, 23, 63, 68, 86], and more.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ensemble systems, such as the aforementioned open-source libraries [1, 50, 62, 77], implement a pipeline in source code, which conditionally applies specialized, single-task models on a given input document, composing the prediction results into a meaningful document representation. Typically, each task entails human-crafted handling of pre-processing and post-processing logic (i.e. setting confidence thresholds for layout detections, matching layout elements to text cells, conditionally applying models).", + "bbox": [ + 88, + 154, + 480, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, multi-task models aim at providing a single model capable of handling various document-understanding related tasks simultaneously, benefiting from shared context and shared representations learned across these different tasks. OCR-reliant methods, such as LayoutLM [29, 84, 85] and UDOP [70], use text extracted from an external OCR engine, along with image and text bounding box locations, as input. In contrast, OCR-free methods such as Donut [34], Dessurt [16], DocParser [18], Pix2Struct [40] are transformer-based models trained end-to-end to take an image as input and directly output text. Large Vision Language Models (LVLMs) such as LLaVA [48], LLaVA-OneVision [42], UReader [87], Kosmos-2 [64], and Qwen-VL [8] all leverage various vision encoders, projection adapters, and LLMs. While they may use different image patching techniques and instruction-tuning datasets to handle a range of vision tasks, they commonly evaluate their document understanding capabilities on datasets like DocVQA [56] and mPLUG-DocOwl 1.5 datasets [27], primarily focusing on question answering and reasoning.", + "bbox": [ + 91, + 292, + 483, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work focuses primarily on document understanding tasks related to conversion and structural recognition, aiming to represent documents with high accuracy and completeness. The closest works to our approach are Nougat [12], DocOwl 2 [28], GOT [79], and Qwen2.5-VL [9]. DocOwl 2 and GOT are dedicated to precise conversion and recognition of document structures. On one hand, DocOwl 2 employs a dynamic shape-adaptive cropping module to process high-resolution images, feeding them into a ViT vision encoder. A dedicated module compresses the processed visual features before feeding them into a LLaMa-based LLM. On the other hand, GOT focuses on converting diverse elements—such as text, formulas, molecular diagrams, tables, charts, sheet music, and geometric shapes—into structured formats. It achieves a lightweight architecture by utilizing a SAM vision encoder [35], a linear projection layer, and a Qwen-0.5B LLM. Building on these approaches, our work not only extends the range of tasks but also broadens the types of documents handled, moving beyond a strictly scientific focus", + "bbox": [ + 91, + 599, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to support a more diverse array of document formats. In parallel, Qwen2.5-VL introduces the Omni-Parsing strategy, that integrates multiple document elements into a unified HTML-based representation. This format encodes layout bounding box information and reading order for various document elements such as paragraphs and charts.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.SmolDocling", + "text_level": 1, + "bbox": [ + 511, + 198, + 645, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on Hugging Face's SmolVLM [53], SmolDocling converts document page images into sequences representing their content and structure. It produces a unified page representation (DocTags markup) that is optimized for the LLM backbone and interpretable by Docling [50]. This new DocTags format covers the reproduction of captions, charts, forms, code, equations, tables, footnotes, lists, page footers and headers, section headings, and text. SmolDocling both recognizes the type and location of these elements as well as performs OCR within them. Our approach enables fullpage conversion with reading-order and hierarchical linking between elements (e.g., linking captions with figures and tables), providing bounding boxes for each component, supporting isolated predictions on individual elements (e.g., cropped tables) and grouping for lists and nested lists.", + "bbox": [ + 511, + 224, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 464, + 700, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model architecture is illustrated in Figure 1. SmolVLM-256M relies on SigLIP base patch-16/512 (93M) as visual backbone and compared to the 2.2B version of the same model, its training data was rebalanced to emphasize document understanding (41%) and image captioning (14%), combining The Cauldron [39], Docmatix [38] datasets with the addition of MathWriting [23]. It uses lightweight variant of the SmolLM-2 family (135M) [2] as the language backbone and employs a radical pixel shuffle method that compresses each 512x512 image patch into 64 visual tokens. Last but not least, tokenization efficiency was also improved in SmolVLM-256M by increasing the pixel-to-token ratio to 4096 pixels per token and introducing special tokens for sub-image separators.", + "bbox": [ + 511, + 487, + 908, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. DocTags", + "text_level": 1, + "bbox": [ + 511, + 710, + 616, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train VLMs for standardized document conversion, we introduce the DocTags markup format, illustrated in Figure 2. Inspired by OTSL [52], DocTags define a structured vocabulary of unambiguous tags and rules that explicitly separate textual content from document structure. This disentanglement improves performances of Imago-to-Sequence models by minimizing confusions. In contrast, direct conversion to standard formats such as HTML or Markdown is often lossy and ambiguous, lacking explicit structural markup, failing to preserve page layout context, and increasing total tokens count. Pieces", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21974", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/697ddb11402df582bbb872662903ea9f82bca7a09917fcb97dc904996e087583.jpg", + "image_caption": [ + "Figure 2. DocTags format describes key features of document elements: type of an element (text, picture, table, code, etc.), location on the page, and content. Nested tags convey additional information: pictures and tables may nest captions, table structure is represented in OTSL tags, lists nest list items, code and pictures carry a classification. All DocTags output is taken from SmolDocling predictions, artificial line-breaks and dots (...) were inserted for better readability. A, B, C - snippets of various pages from [52], D - from [81]" + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 854, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of text content that represent basic elements in DocTags are enclosed in XML-style tags. Tags define document block types such as: text, caption, footnote, formula, title, list_item, page Footer, page_header, picture, section_header, otsl, code, document_index, etc. Each element can nest additional location tags encoding its position on the page as a bounding box, represented in DocTags as $<\\mathrm{loc\\_x1}><\\mathrm{loc\\_y1}><\\mathrm{loc\\_x2}><\\mathrm{loc\\_y2}>$ .", + "bbox": [ + 88, + 536, + 480, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Special blocks, such as tables and images, nest additional descriptors for captions, tabular structure, or image classes. To encode table structures, we fully include the OTSL vocabulary into DocTags. OTSL has the advantage of being minimalist and supporting cell spans and header information. Its tokens are interleaved with text, allowing simultaneous encoding of both table structure and content. In order to promote robust visual-semantic alignment in our document understanding pipeline, we maintain a uniform DocTags representation for cropped page elements (e.g., tables, code, equations) that is identical to their full-page counterparts. This consistency ensures that the model can leverage similar visual cues whether an element is presented in isolation or in the context of an entire page, thereby strengthening the learning signal for both computer vision and downstream machine learning tasks.", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a complete list of tags and detailed description of the format, consult the supplementary material.", + "bbox": [ + 511, + 536, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Model Training", + "text_level": 1, + "bbox": [ + 511, + 578, + 669, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a curriculum learning approach to progressively align our model for document conversion, ensuring faster convergence. As an initial step, we incorporate DocTags as tokens in our tokenizer. To align the LLM part we freeze the vision encoder and train only the remaining network to adapt it to the new output format which it hasn't seen before. To ensure comprehensive coverage of all DocTags, we maintain a balanced mix of tasks and data types during training. Next, we unfreeze the vision encoder and train the model on our pretraining datasets (see Sec. 4.1), along with all task-specific conversion datasets, including tables, code, equations, and charts. Finally, we fine-tune using all available datasets.", + "bbox": [ + 511, + 601, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Data", + "text_level": 1, + "bbox": [ + 511, + 814, + 578, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For document understanding tasks, each page image represents the visual layout of textual elements—such as paragraphs, formulas, tables, code, or figures with embedded textual content. Publicly available datasets that unify all", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21975", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "these characteristics remain limited, and existing corpora often suffer from small sample sizes or fragmented formats. In this section, we explain how we mitigate these issues in both pre-training and task-specific training scenarios.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training datasets", + "text_level": 1, + "bbox": [ + 89, + 162, + 287, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the domain of document understanding, currently available open-access document pre-training datasets include OCR-IDL [10], IIT-CDIP [41], CCpdf [74], and WordScape [78]. OCR-IDL and IIT-CDIP, industry-sourced from the 1990s, provide only page images and OCR text. CCpdf and WordScape, derived from Common Crawl, focus on raw PDFs and Word documents, respectively. While WordScape enriches documents with layout, table structure, and topic information via XML markup, its visual variability is limited. Notably, none of these datasets offer rich annotations for all page elements.", + "bbox": [ + 89, + 185, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DocLayNet-PT. Given the lack of public multimodal document pre-training data with sufficient annotation features, we constructed a new document pre-training dataset named DocLayNet-PT. DocLayNet-PT is a 1.4M pages dataset extracted from the DocFM dataset [25] of unique PDF documents, sourced from CommonCrawl, Wikipedia, and business-related documents. We focused on this dataset to look for documents including visually distinct content such as equations, tables, code and charts with colorful layouts. This dataset was augmented with weak annotations through a series of processing steps. First, we applied PDF parsing to all pages to obtain the text layout including positional information using Docling [7, 50]. Next, we performed an enrichment step to obtain an enhanced, machine-friendly representation of each page. Specifically, we provide annotations for layout elements, table structure, language, topic, and figure classification on each page in DocTag format.", + "bbox": [ + 89, + 352, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Docmatix. To preserve SmolVLM's original DocVQA capabilities, we adapted the same weak annotation strategy used for DocLayNet-PT and applied it to the 1.3M documents inside Docmatix dataset [38]. In addition to the 9.5 million DocVQA questions, we introduced another instruction requiring the full conversion of multi-page documents to DocTags.", + "bbox": [ + 89, + 625, + 482, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Task-specific Datasets", + "text_level": 1, + "bbox": [ + 89, + 742, + 294, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layout. To optimize prediction quality for document layout and table structure, we sampled 76K pages from DocLayNet-PT. Each page underwent human annotation and rigorous quality review, resulting in a high-quality ground-truth dataset. Of these, 60K pages were utilized for fine-tuning, constituting what we refer to as DocLayNet v2. Additionally, we extracted 63K pages from WordScape, specifically filtering for pages containing both text and tables, and utilized their inherently structure-preserving for", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mat as a reliable source of ground truth. To further enhance the model's adaptability to diverse layouts, colors, and fonts, we synthetically generated 250K pages using content from Wikipedia, creating the dataset known as SynthDocNet. This synthetic dataset enables the controlled generation of various document types with precise annotations. All datasets share a unified annotation format and class definitions, ensuring consistency and comprehensive document representation.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tables. For the table structure recognition task we trained our model on public datasets such as PubTables-1M [69], FinTabNet [92], WikiTableSet [51] and tabular information extracted from WordScape documents. We transformed table structure information from the original ground-truth to OTSL format [52]. Then, we interleaved each cell tag with ground-truth text. This yields a condensed sequence which represents the structure of the table (including merged cells, and header declaration) along with its textual content.", + "bbox": [ + 511, + 229, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Charts. For chart reconstruction, several datasets are publicly available [32, 33, 54, 82] that consist of both synthetic and web-scraped charts. However, these datasets are limited by either the quantity of real-world data or the visual diversity of synthetic examples. This has prompted recent studies [13, 26, 47, 55, 57, 82, 83] to crawl or render additional datasets, which unfortunately are rarely shared publicly. To address these gaps, we generated our own chart dataset comprising four types of charts: line, pie, bar, and stacked bar. Using data from 90,000 tables sourced from the FinTabNet dataset [92], we generated a comprehensive dataset containing 5,000 line charts, 380,000 pie charts, 380,000 bar charts, and 77,000 stacked bar charts. Each data point was rendered using three different visualization libraries (details provided in the Supplementary Material), yielding a total of 2.5 million visually diverse charts.", + "bbox": [ + 511, + 383, + 903, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Code. Technical books and scientific documents frequently include code snippets, making it essential for document-processing models to accurately interpret code structure, particularly indentation, which carries semantic significance in certain programming languages. Several publicly available datasets provide extensive collections of code snippets with permissive licenses in text format [36, 66], typically aggregated from sources like GitHub. However, a common limitation is that these datasets contain only text-based code snippets, lacking any image-based representation. To address this gap, we leveraged LaTeX and Pygments [15] to generate visually diverse renderings of code (details provided in the Supplementary Material). Our dataset includes 9.3M million rendered code snippets generated at 120 dpi, covering 56 programming languages.", + "bbox": [ + 511, + 626, + 903, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equations. Accurately parsing mathematical formulas is crucial for document-processing models, as formulas often contain nuanced structures and semantics criti-", + "bbox": [ + 511, + 856, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cal for scientific understanding. To effectively train our model for formula transcription, we combined publicly available datasets with a custom-rendered dataset. The public datasets [11, 17, 22, 23] collectively comprise approximately $730\\mathrm{k}$ unique formulas. In addition, we extracted 4.7 million formulas from arXiv by applying a regular expression to the source LaTeX code. The extracted formulas underwent a rigorous normalization procedure to ensure the model was trained only on correct and standardized LaTeX code. Subsequently, these normalized equations were visually rendered using LaTeX (details provided in the Supplementary Material). Overall, our dataset includes 5.5 million unique formulas, each rendered at a resolution of 120 dpi.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24b504bdc5514a0fcbd114496f123dcb2d2d3672d2c3a85065d66a5889121100.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset NameSizeType
TheCauldron [39]2.63M
DocLayNet-PT-Instruct1.62MMulti Task
DocMatix w/ Doc Conversion* [38]952K
SynthChartNet*2.5M
PlotQA [58]225K
FigureQA [32]100KChart
Unichart [55]600K
ChartQA [54]28K
DocLayNet-PT1.40M
Doc4MLlaVa - RVLCDIP742K
SynthDocNet375KConversion
DocLayNet [65]81K
DocLayNet v276K
WordScape-PT63K
PubMed1M [69]1M
WordScape Tables [78]207KTable
WikiTableSet (EN) [51]204K
FinTabNet [92]90K
SynthFormulaNet*5.5MFormula
SynthCodeNet*9.3MCode
", + "bbox": [ + 94, + 297, + 483, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Datasets. Overview of the datasets used for the training and evaluation of SmolDocling. Datasets we contribute as part of this publication are denoted with $(^{*})$ .", + "bbox": [ + 89, + 648, + 482, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Document Instruction Tuning datasets", + "text_level": 1, + "bbox": [ + 89, + 712, + 421, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To reinforce the recognition of different page elements and introduce document-related features and no-code pipelines, we leveraged rule-based techniques and the Granite-3.1-2b-instruct [25] LLM. Using samples from DocLayNet-PT pages, we generated one instruction by randomly sampling layout elements from a page. These instructions included tasks such as \"Perform OCR atbbox,\" \"Identify page element type atbbox,\" and \"Extract all section headers from the page.\" In addition, we train with Cauldron [39] to avoid catastrophic forgetting with the amount of conversation datasets introduced.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we assess the performance of SmolDocling on document understanding tasks, and compare the results to previously published models and baselines. Capturing the qualities and defects of SmolDocling and other models in comparable numbers is a non-trivial effort for reasons we outline further below. Therefore, we report results on standard evaluation tasks and metrics independently (see section 5.2), but also provide a qualitative assessment of output characteristics (see section 5.3).", + "bbox": [ + 511, + 114, + 906, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 258, + 730, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our training was conducted using 64 NVIDIA A100 80GB GPUs, with each epoch requiring 28 hours. The model was trained for a total of 4 epochs using the AdamW optimizer with a learning rate of $2 \\times 10^{-4}$ and $2 \\times 10^{-6}$ for the vision encoder once unfrozen. We also employed gradient clipping to 1.0 for stability and a warmup ratio of 0.03. For inference, using VLLM [37] on an A100 GPU, SmolDocling achieves a page conversion time of 0.35 seconds while occupying only 0.489 GB of VRAM. The model supports a maximum sequence length of 8,192 tokens and is trained to process up to three pages at a time.", + "bbox": [ + 511, + 279, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 452, + 707, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable valid cross-model comparisons with SmolDooling, we addressed multiple dataset and harmonization challenges:", + "bbox": [ + 511, + 474, + 905, + 518 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Public evaluation datasets lack comprehensive multi-task annotations with diverse layouts. We augmented DocLayNet [65] with text content and reading-order to enable evaluation of both layout analysis and text recognition. Table structure, code listing, formula and chart recognition are evaluated on specific datasets.", + "- Different models each follow their own conventions for output markup (e.g., Markdown, HTML, DocTags) and produce annotations which do not fully align in semantics. For example, SmolDocling is trained to produce 14 distinct layout-class tags via DocTags, while Qwen2.5-VL combines standard HTML tags and a few explicit class attributes with only partially compatible definitions. GOT and Nougat output Markdown, which can express paragraphs, lists and tables without location annotation.", + "- Input resolution of page images can affect model performance, but is rarely specified in literature. We standardized at 144 DPI (dots per inch), which reproduces enough detail to be perfectly legible by humans while keeping resource demand for inference reasonable." + ], + "bbox": [ + 513, + 521, + 905, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text Recognition", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/65d7ee92eb96bc0581d7c1c382210464f757d0e25f4e5171b77d64587183a5f7.jpg", + "table_caption": [], + "table_footnote": [ + "Values for Dockling provide a reference for the performance of an ensemble system." + ], + "table_body": "
MethodSizeEdit Distance ↓F1-score ↑Precision ↑Recall ↑BLEU ↑METEOR ↑
Full-page \nDocling† [7] v2.25-0.400.910.930.920.720.80
Full-page \nQwen2.5 VL [9]7B0.560.720.800.700.460.57
GOT [79]580M0.610.690.710.730.480.59
Nougat (base) [12]350M0.620.660.720.670.440.54
SmolDocling (Ours)256M0.480.800.890.790.580.67
Code listings \nSmolDocling (Ours)256M0.110.920.940.910.870.89
Equations \nQwen2.5 VL [9]7B0.220.890.910.870.680.77
GOT [79]580M0.110.950.950.960.850.91
Nougat (base) [12]350M0.620.600.600.530.330.41
SmolDocling (Ours)256M0.110.950.960.950.830.89
", + "bbox": [ + 96, + 88, + 906, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "non-textual elements. For different tasks, we conduct evaluations on diverse content types: full-page documents (from DocLayNet test set) in Markdown format, code snippets (from our SynthCodeNet dataset) in plaintext, and equations (from Im2Latex-230k [21]) in LaTeX format. For Qwen2.5-VL and SmolDocling, we convert HTML or DocTags output to Markdown while preserving formatting where possible.", + "bbox": [ + 88, + 438, + 480, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables are omitted from the full-page output since they do not compare well through Markdown, due to loss of row and column spans. Code snippets must be reproduced with intact line breaks and indentation for accurate scoring. Formulas are evaluated separately in LaTeX format, which is natively produced by all the models tested in this work. The LaTeX outputs from all models, including the ground truth, undergo the same normalization procedure used for generating the SynthFormulaNet dataset. This ensures consistency and fairness in model comparisons.", + "bbox": [ + 88, + 561, + 482, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SmolDocling significantly outperforms GOT, Nougat and also Qwen2.5-VL, which is 27 times larger, on every metric in full-page transcription (see Table 2. For the code parsing task, we report results exclusively for SmolDocling, as we introduce this task for the first time, and the other models were not explicitly trained for it. SmolDocling also demonstrates superior performance compared to Nougat and Qwen2.5-VL in the formula recognition task, closely matching GOT's results.", + "bbox": [ + 88, + 715, + 482, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Layout. We evaluate element localization accuracy using the DocLayNet test set (Table 3), measuring the ability to reconstruct document representations that support con", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tent grounding and visual feature extraction. SmolDocling is compared against Qwen2.5-VL-7b [9], currently the only other VLM addressing this task. For a fair comparison, we map each model's label outputs and the ground-truth to six compatible classes: Text, Section Heading, List Item, Table, Picture, and Formula. Overall, SmolDocling outperforms Qwen2.5-VL-7b by a significant margin, however both models score far below a human baseline. This can be partially explained through inherent difficulty of the dataset, as indicated by relatively low human agreement scores. The scores are primarily impacted by moderate recall of element bounding boxes and different degrees of label confusion in both SmolDocling and Qwen2.5-VL. A majority of samples show accurate element bounding boxes as seen in the supplementary material.", + "bbox": [ + 511, + 438, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/997a757850cb44a3c33b7e9ce5ea90adf3a8379a0bb14b679b0c87f684bdf624.jpg", + "table_caption": [ + "Table 2. Structured document text recognition. We evaluate OCR performance and document formatting on DocLayNet, focusing on textual elements, excluding tables. Text accuracy is measured through multiple metrics (Edit Distance, F1-Score, Precision, Recall, BLEU and METEOR) as proposed in [12]." + ], + "table_footnote": [], + "table_body": "
Docling [7] v2.25SmolDoclingQwen2.5-VLHuman
Formula0.7370.2670.0590.84
Table0.6670.2660.2620.79
List Item0.5280.2960.0900.87
Section Header0.4710.2890.1290.83
Text0.4440.2040.1800.85
Picture0.3280.0660.0780.70
Overall0.5290.2310.1330.82
", + "bbox": [ + 516, + 676, + 903, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Layout analysis comparison. Evaluation of layout analysis task on DocLayNet with a subset of 6 class labels in mAP[0.5:0.95] metric. Additional reference points are provided for an estimate of human performance [65].", + "bbox": [ + 511, + 782, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table Structure Recognition. We evaluate the accuracy of recovering table structures, i.e. shaping columns and rows with the correct text content and spans, through", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21978", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the TEDS metric [93], once including table text content and once on structure only with omitted text (see Table 4). For isolated assessment, we provide cropped table images from FinTabNet [92] and PubTables-1M [69] test sets, which have poor image quality (72 dpi, compression artifacts). SmolDocling performs competitively against significantly larger models, despite challenges with text transcription from the low-resolution image crops, a limitation likely attributable to insufficient training at such resolutions. Scores for structure only TEDS evaluation are therefore much stronger.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34aab7261b746c162248f2e8f3cb2bd13bb3d4873c83d2b26332d549b74c27e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModel SizeFinTabNetPubTables-1M
SmolVLM [53]2.2B0.180.32
Granite Vision [72]2.98B0.540.7
TableFormer† [52]52M0.89 f0.84 (0.86)
Qwen2.5-VL 7B7B0.66- (0.786)-
SmolDocling (Ours)256M0.72 (0.81)0.83 (0.88)
SmolDocling (Ours) @150DPI256M0.76 (0.82)-
", + "bbox": [ + 93, + 268, + 482, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Chart Extraction. In this task, we evaluate the ability to convert cropped sections of charts into datapoint tables on the ChartQA [54] dataset. All table output is transformed into HTML and compared against the ground-truth HTML using TEDS score, as reported in Table 5. We attribute SmolDocling's performance on charts to the variability in ground-truth styles and the differing interpretations of representations across datasets, which we observe leads to inconsistencies in the model's predictions. Despite its significantly smaller size, SmolDocling achieves a competitive TEDs score.", + "bbox": [ + 89, + 460, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, SmolDocling achieves leading performance across a wide range of document conversion tasks and outperforms significantly larger models in text recognition, layout analysis and structure extraction. Furthermore, the model's high performance in downstream tasks, such as code listing and formula recognition, demonstrates its versatility and real-world applicability. Notably, we establish benchmark results for the code listing recognition, setting a standard for future evaluations.", + "bbox": [ + 89, + 628, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 772, + 272, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several characteristics of SmolDocling output are not evident from task-based metrics alone. In this section, we highlight different output characteristics observed in practice. More illustrated samples can be found in the supplementary material.", + "bbox": [ + 89, + 794, + 480, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As presented in Figure 2, SmolDocling outputs a sequence of DocTags to encode content, structure and layout.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8cc44181d4dd5e4f8ac6d8310691ff9de091c61de341de58fedfc42e4b9499a5.jpg", + "table_caption": [ + "Table 4. Table structure and cell content reconstruction. Numbers are given in TEDS metric including text content, with structure-only score in brackets where available. We produced results for SmolDocling and TableFormer, other results are reflected here as published. †TableFormer extracts the content through external OCR." + ], + "table_footnote": [], + "table_body": "
ModelSizeTEDs↑
Phi-3.5-vision4B0.40
Granite Vision [72]3B0.95
SmolVLM [53]2.2B0.02
Molmo-E1B0.54
SmolDocling (Ours)256M0.83
", + "bbox": [ + 566, + 88, + 854, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Chart analysis. We compare SmolDocling to other small-sized VLMs on the chart-to-table task using the TEDS metric on the ChartQA dataset. The results in this table are based on the work reported in [25].", + "bbox": [ + 511, + 194, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A few notable traits beyond straight content reproduction include: 1) Labeling of page headers and footers, which allows ignoring these repetitive elements in downstream applications. 2) Linking of information, such as captions to tables or pictures, or list elements to parent list through nesting of tags. 3) Preserved line breaks and indentation for code snippets, while line wraps in paragraphs are correctly removed.", + "bbox": [ + 511, + 284, + 903, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Among typical defects found in the output of SmolDocling are missing tags (e.g. absent location tags), malformed structure (such as missing closing tags), and endless repetition of tokens from some point in a page. Both of these phenomena are commonly seen in auto-regressive models and may prevent parsing or rendering the output correctly (see supplementary material). Despite these limitations, SmolDocling demonstrates unique robustness advantages compared to ensemble systems like Docling, since conversion output is inferred in a single pass. In the latter, errors may accumulate throughout the ensemble model pipeline. For example, a mislocated table detection will lead to missing or distorted output of table structure and content, since this depends on correct table localization. In contrast, SmolDocling can reproduce the tables correctly, even with incorrectly determined location.", + "bbox": [ + 511, + 407, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 675, + 633, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce SmolDocling, an efficient and compact VLM optimized for document conversion with rich output representation. We also present a suite of new datasets with a unified format for document conversion, including the novel task of code listing transcription. We identify page element localization as a critical area requiring further refinement, where targeted techniques will significantly enhance performance in future iterations. Our results conclusively demonstrate that smaller models with unified, optimized output formats such as DocTags can effectively compete with significantly larger models, establishing a clear pathway for resource-efficient multi-task document understanding models.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21979", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: We thank our collaborators at Hugging Face, Andrés Marafioti and Miquel Farré, who kindly and rapidly developed SmolVLM-256M-Instruct in response to our collaboration, providing the foundational model upon which this work is built.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Grobid. https://github.com/kermitt2/grobid, 2008-2025. 2, 3", + "[2] Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martin Blázquez, Guilherme Penedo, Lewis Tunstall, Andres Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werra, and Thomas Wolf. SmolLM2: When Smol Goes Big - Data-Centric Training of a Small Language Model, 2025. 2, 3", + "[3] Amazon Web Services. Amazon Texttract. https://aws.amazon.com/texttract/, 2024. Accessed: 2024-11-11. 2", + "[4] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2", + "[5] Anthropic. Claude-3.5. https://www.anthropic.com/news/claude-3-5-sonnet, 2024. Accessed: 2024-02-11. 2", + "[6] Christoph Auer, Michele Dolfi, André Carvalho, Cesar BerroSpi Ramis, and Peter WJ Staar. Delivering Document Conversion as a Cloud Service with High Throughput and Responsiveness. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD), pages 363-373. IEEE, 2022. 2", + "[7] Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Nikolaos Livathinos, Panos Vagenas, Cesar Berrospi Ramis, Matteo Omenetti, Fabian Lindlbauer, Kasper Dinkla, et al. Docling Technical Report. arXiv preprint arXiv:2408.09869, 2024. 5, 7", + "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3", + "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL Technical Report, 2025. 2, 3, 7", + "[10] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. In European Conference on Computer Vision, pages 241-252. Springer, 2022. 5", + "[11] Lukas Blecher. LaTeX-OCR, 2024. Accessed: 2024-11-09. 6" + ], + "bbox": [ + 93, + 205, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. 3, 6, 7", + "[13] Jinyue Chen, Lingyu Kong, Haoran Wei, Chenglong Liu, Zheng Ge, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. OneChart: Purify the Chart Structural Extraction via One Auxiliary Token. arXiv preprint arXiv:2404.09987, 2024. 3, 5", + "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2", + "[15] Pygments Contributors. Pygments: Python syntax highlighter, 2024. Accessed: 2024-11-06. 5", + "[16] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 3", + "[17] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning, pages 980-989. PMLR, 2017. 3, 6", + "[18] Mohamed Dhouib, Ghassen Bettaieb, and Aymen Shabou. Docparser: End-to-endOCR-free information extraction from visually rich documents. In Document Analysis and Recognition - ICDAR 2023, pages 155-172, Cham, 2023. Springer Nature Switzerland. 3", + "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model. arXiv preprint arXiv:2401.16420, 2024. 2", + "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Zhe Chen, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Kai Chen, Conghui He, Xingcheng Zhang, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD. arXiv preprint arXiv:2404.06512, 2024. 2", + "[21] Gregory Eritsyan. Im2LaTeX 230K. https://www.kaggle.com/datasets/gregoryeritsyan/im2latex-230k, 2023. Accessed: 2025-03-06. 7", + "[22] Hugging Face. latex-formulas Dataset, 2024. Accessed: 2024-11-09. 6", + "[23] Philippe Gervais, Asya Fadeeva, and Andrii Maksai. Math-writing: A dataset for handwritten mathematical expression recognition. URL https://arxiv.org/abs/2404.10690, 2024.3, 6" + ], + "bbox": [ + 514, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21980", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Google Cloud. Document AI. https://cloud.google.com/document-ai, 2024. Accessed: 2024-11-11. 2", + "[25] IBM Granite Team. Granite 3.0 Language Models, 2024. 2, 5, 6, 8", + "[26] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 3, 5", + "[27] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3", + "[28] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 3", + "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 3", + "[30] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), pages 1–6. IEEE, 2019. 3", + "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 2", + "[32] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 3, 5, 6", + "[33] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 3, 5", + "[34] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 3", + "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3", + "[36] Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Munoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The Stack: 3 TB of permissively licensed source code. Preprint, 2022. 5" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6", + "[38] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 2, 3, 5, 6", + "[39] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models?, 2024. 3, 6", + "[40] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 3", + "[41] David Lewis, Gady Agam, Shlomo Argamon, Ophir Frieder, David Grossman, and Jefferson Heard. Building a test collection for complex document information processing. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 665-666, 2006. 5", + "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3", + "[43] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2", + "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[45] Han Lin, Peng Yang, and Fanlong Zhang. Review of scene text detection and recognition. Archives of computational methods in engineering, 27(2):433-454, 2020. 3", + "[46] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 3", + "[47] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 3, 5", + "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2023. 2, 3", + "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21981", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Nikolaos Livathinos, Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Panos Vagenas, Cesar Berropi Ramis, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Shubham Gupta, Rafael Teixeira de Lima, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, and Peter W. J. Staar. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion, 2025. 2, 3, 5", + "[51] Nam Ly., Atsuhiro Takasu., Phuc Nguyen., and Hideaki Takeda. Rethinking Image-Based Table Recognition Using Weakly Supervised Methods. pages 872-880, 2023. 5, 6", + "[52] Maksym Lysak, Ahmed Nassar, Nikolaos Livathinos, Christoph Auer, and Peter Staar. Optimized Table Tokenization for Table Structure Recognition. In Document Analysis and Recognition - ICDAR 2023, pages 37-50, Cham, 2023. Springer Nature Switzerland. 2, 3, 4, 5, 8", + "[53] Andres Marafioti, Orr Zohar, Miquel Farre, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Alal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. SmolVLM: Redefining small and efficient multimodal models. 2025. 1, 2, 3, 8", + "[54] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3, 5, 6, 8", + "[55] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 3, 5, 6", + "[56] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 3", + "[57] Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. Chartassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. arXiv preprint arXiv:2401.02384, 2024. 3, 5", + "[58] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. PlotQA: Reasoning over Scientific Plots. In The IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 6", + "[59] Microsoft Azure. AI Document Intelligence. https://azure.microsoft.com/en-us/products/ai-services.ai-document-intelligence, 2024. Accessed: 2024-11-11. 2", + "[60] OpenAI. Hello GPT-4O. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-02-09, 2024-02-11, 2024-02-12. 2", + "[61] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 2", + "[62] Vik Paruchuri. Marker: Convert PDF to Markdown Quickly with High Accuracy. https://github.com/VikParuchuri/marker, 2024. 2, 3", + "[63] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2024. Accessed: 2024-11-11. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3", + "[65] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 3743–3751, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6, 7", + "[66] Ruchir Puri, David Kung, Geert Janssen, Wei Zhang, Giacomo Domeniconi, Vladimir Zolotov, Julian Dolby, Jie Chen, Mihir Choudhury, Lindsey Decker, Veronika Thost, Luca Buratti, Saurabh Pujar, Shyam Ramji, Ulrich Finkler, Susan Malaika, and Frederick Reiss. CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Tasks, 2021. 5", + "[67] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2", + "[68] Felix M Schmitt-Koopmann, Elaine M Huang, Hans-Peter Hutter, Thilo Stadelmann, and Alireza Darvishy. MathNet: A Data-Centric Approach for Printed Mathematical Expression Recognition. IEEE Access, 2024. 3", + "[69] Brandon Smock, Rohith Pesala, and Robin Abraham. PubTables-1M: Towards comprehensive table extraction from unstructured documents. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4624–4632, 2022. 3, 5, 6, 8", + "[70] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 3", + "[71] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2", + "[72] Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kondic, Nimrod Shabtay, Pengyuan Li, Roei Herzig, Shafiq Abedin, Shaked Perek, Sivan Harary, Udi Barzelay, Adi Raz Goldfarb, Aude Oliva, Ben Wieles, Bishwaranjan Bhattacharjee, Brandon Huang, Christoph Auer, Dan GUTFreund, David Beymer, David Wood, Hilde Kuehne, Jacob Hansen, Joseph Shtok, Ken Wong, Luis Angel Bathen, Mayank Mishra, Maksym Lysak, Michele Dolfi, Mikhail Yurochkin, Nikolaos Livathinos, Nimrod Harel, Ophir Azulai, Oshri Naparstek, Rafael Teixeira de Lima, Rameswar Panda, Sivan Doveh, Shubham Gupta, Subhro Das, Syed Zawad, Yusik Kim, Zexue He, Alexander Brooks, Gabe Goodhart, Anita Govindjee, Derek Leist, Ibrahim Ibrahim, Aya Soffer, David Cox," + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21982", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kate Soule, Luis Lastras, Nirmit Desai, Shila Ofek-koifman, Sriram Raghavan, Tanveer Syeda-Mahmood, Peter Staar, Tal Drory, and Rogerio Feris. Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence, 2025. 8", + "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[74] Michal Turski, Tomasz Stanisławek, Karol Kaczmarek, Paweł Dyda, and Filip Graliński. CCpdf: Building a High Quality Corpus for Visually Rich Documents from Web Crawl Data. In Document Analysis and Recognition - IC-DAR 2023, pages 348-365, Cham, 2023. Springer Nature Switzerland. 5", + "[75] Unstructured.io Team. Unstructured.io: Open-Source Pre-Processing Tools for Unstructured Data. https://unstructured.io, 2024. Accessed: 2024-11-19. 2", + "[76] Vicuna. Vicuna: An open-source chatbot impressing GPT-4 with $90\\%$ ChatGPT quality. https://vicuna.lmsys.org/, 2023.2", + "[77] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, Bo Zhang, Liquun Wei, Zhihao Sui, Wei Li, Botian Shi, Yu Qiao, Dahua Lin, and Conghui He. MinerU: An Open-Source Solution for Precise Document Content Extraction, 2024. 2, 3", + "[78] Maurice Weber, Carlo Siebenschuh, Rory Butler, Anton Alexandrov, Valdemar Thanner, Georgios Tsolakis, Haris Jabbar, Ian Foster, Bo Li, Rick Stevens, et al. WordScape: a Pipeline to extract multilingual, visually rich Documents with Layout Annotations from Web Crawl Data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6", + "[79] Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model. arXiv preprint arXiv:2409.01704, 2024. 3, 7", + "[80] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2", + "[81] Wes McKinney and the Pandas Development Team. Pandas: powerful Python data analysis toolkit, Release 1.4.4. https://pandas.pydata.org/pandas-docs/ version/1.4/pandas.pdf, 2022.4", + "[82] Renqiu Xia, Bo Zhang, Haoyang Peng, Hancheng Ye, Xiangchao Yan, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. StructChart: Perception, Structuring, Reasoning for Visual Chart Understanding. arXiv preprint arXiv:2309.11268, 2023. 3, 5", + "[83] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024. 3, 5" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 3", + "[85] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020.3", + "[86] Zuoyu Yan, Xiaode Zhang, Liangcai Gao, Ke Yuan, and Zhi Tang. ConvMath: a convolutional sequence network for mathematical expression recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4566-4572. IEEE, 2021. 3", + "[87] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3", + "[88] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition. arXiv preprint arXiv:2309.15112, 2023. 2", + "[89] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer-2.5: A Versatile Large Vision Language Model Supporting Long-Contextual Input and Output. arXiv preprint arXiv:2407.03320, 2024. 2", + "[90] Qiming Zhang, Jing Zhang, Yufei Xu, and Dacheng Tao. Vision Transformer with Quadrangle Attention. arXiv preprint arXiv:2303.15105, 2023. 2", + "[91] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pretrained transformer language models. URL https://arxiv.org/abs/2205.01068, 3:19-0, 2023. 2", + "[92] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 697-706, 2021. 3, 5, 6, 8", + "[93] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. arXiv preprint arXiv:1911.10683, 2019. 8", + "[94] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21983", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_model.json b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..17246f711fe43c11abbab50195704cd252de6a4c --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_model.json @@ -0,0 +1,2402 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.175 + ], + "angle": 0, + "content": "SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.204, + 0.921, + 0.274 + ], + "angle": 0, + "content": "Ahmed Nassar Matteo Omenetti Maksym Lysak Nikolaos Livathinos Christoph Auer Lucas Morin \nRafael Teixeira de Lima Yusik Kim A. Said Gurbuz Michele Dolfi Peter W. J. Staar \n{ahn, mly, nli, cau, dol, taa}@zurich.ibm.com, {Matteo.Omenetti1}@ibm.com \nIBM Research" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.486, + 0.707 + ], + "angle": 0, + "content": "We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms — significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition. Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model weights and datasets are available at: https://huggingface.co/ds4sd" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.903 + ], + "angle": 0, + "content": "For decades, converting complex digital documents into a structured, machine-processable format has been a significant technical challenge. This challenge primarily stems from the substantial variability in document layouts and styles, as well as the inherently opaque nature of the widely used PDF format, which is optimized for printing rather than semantic parsing. Intricate layout styles and visually challenging elements such as forms, tables, and complex charts can significantly impact the reading order and gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.908, + 0.462 + ], + "angle": 0, + "content": "eral understanding of documents. These problems have driven extensive research and development across multiple domains of computer science. On one hand, sophisticated ensemble systems emerged, which decompose the conversion problem into several sub-tasks (e.g. OCR, layout analysis, table structure recognition, classification) and tackle each sub-task independently. Although such systems can achieve high-quality results for many document types while maintaining relatively low computational demands, they are often difficult to tune and generalize." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.703 + ], + "angle": 0, + "content": "On the other hand, in the recent past, great interest has developed around large foundational multimodal models that can solve the whole conversion task in one shot, while simultaneously offering flexible querying and parametrization through prompts. This approach has been made possible by the advent of multimodal pre-training of large vision-language models (LVLMs), which opened up a vast array of opportunities for leveraging diverse data sources, including PDF documents. However, the literature on this topic highlights a significant gap in the availability of high-quality and open-access datasets suitable for training robust multimodal models for the task of document understanding. Furthermore, relying on LVLMs may introduce common issues associated with such models, including hallucinations and the use of significant computational resources, making them impractical from both a quality and cost perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.793 + ], + "angle": 0, + "content": "In this work, we outline how we close the gaps left by publicly available datasets and establish a training approach to achieve end-to-end, full-featured document conversion through a vision-language model, which we name SmolDoccling. Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "- SmolDocling: An ultra-compact VLM for end-to-end document conversion, based on Hugging Face's most recent SmolVLM-256M [53], which is between 5 and 10 times smaller in parameters than comparable VLMs tuned on document understanding tasks. Hence, we reduce the computational complexity to roughly the same order of magnitude as typical representatives of ensemble meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21972" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.087, + 0.833, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 1. SmolDocling/SmolVLM architecture. SmolDocling converts images of document pages to DocTags sequences. First, input images are encoded using a vision encoder and reshaped via projection and pooling. Then, the projected embeddings are concatenated with the text embeddings of the user prompt, possibly with interleaving. Finally, the sequence is used by an LLM to autoregressively predict the DocTags sequence." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.374, + 0.483, + 0.419 + ], + "angle": 0, + "content": "ods. Uniquely, SmolDocling learns and produces a unified document representation that captures content, structure and spatial location of all elements in a page." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.42, + 0.483, + 0.554 + ], + "angle": 0, + "content": "- We augment existing document pre-training datasets with additional feature annotations, and create new datasets for tasks with insufficient coverage in open multi-modal training datasets, including code listings, equations, and chart understanding. Importantly, we construct full-page ground truth data that combines key annotation features such as layout, table structure, code listings, charts, and formulas in their natural context, and derive instruction datasets from all training datasets used." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.555, + 0.483, + 0.6 + ], + "angle": 0, + "content": "- We introduce DocTags, a document markup format optimized to efficiently represent the full content and layout characteristics of a document, inspired by [52]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.6, + 0.483, + 0.645 + ], + "angle": 0, + "content": "- We evaluate the achieved prediction quality of SmolDoccling on several tasks, including comparisons to popular community models." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.42, + 0.483, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.233, + 0.673 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.444, + 0.699 + ], + "angle": 0, + "content": "2.1. Large Vision-Language Models (LVLMs)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With the advent of large language models (LLMs) [2, 4, 8, 25, 31, 61, 67, 73, 80] characterized by their powerful contextual understanding and adaptability, several approaches have emerged to extend them into large vision-language models. Some proprietary, closed-source LLMs have been adapted into multimodal large language models (MLLMs), such as GPT-4o [60], Gemini [71], and Claude 3.5 [5], demonstrating exceptional capabilities across various modalities, including vision. Open-source methods are actively advancing to match the capabilities of proprietary MLLMs. BLIP-2 [44] is one of the earliest works combining a vision encoder with a frozen LLM (OPT [91] or FlanT5 [14]) using a lightweight transformer (Q-former" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.629 + ], + "angle": 0, + "content": "[90]). Building upon BLIP-2, MiniGPT-4 [94] integrated a frozen visual encoder with a frozen LLM (Vicuna [76]) using a Q-Former network and a single projection layer. LLaVA [48, 49] employs a similar architecture but uses a minimal adapter layer. LLaVA-OneVision [42] and LLaVA-NeXT-Interleave [43] build upon LLaVA to support multi-image, higher resolution, and video understanding. For handling high resolutions and text-image compositions and comprehension, InternLM-XComposer introduces a family of models [19, 20, 88, 89] designed specifically for these tasks. Qwen-VL [8] introduces a position-aware adapter to address efficiency issues caused by the vision encoder generating long image feature sequences. Qwen2.5-VL [9] employs windowed attention with 2D rotary positional embeddings to process native-resolution inputs efficiently and integrates an vision-language merger for dynamic feature compression." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.768 + ], + "angle": 0, + "content": "In this work, we build upon the SmolVLM architecture approach [53]. Precisely, we adopt SmolVLM-256M, a compact variant of base SmolVLM-2.2B model, inspired by the architecture of Idefics3 [38] which employs the shape-optimized SigLIP encoder as its visual backbone. It applies an aggressive pixel shuffle strategy to compress visual features, reduces the number of image hidden states and introduces new special tokens to separate sub-images and improve tokenization efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.892, + 0.79 + ], + "angle": 0, + "content": "2.2. State of the Art in Document Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Several powerful document understanding solutions have emerged on the market in the past decade as commercial cloud offerings on hyperscalers [3, 6, 24, 59], frontier models such as GPT-4o [60] or Claude [5], or as open-source libraries such as Docling [50], Grobid [1], Marker [62], MinerU [77] or Unstructured [75]. Typical tasks that document understanding encompasses are document classifi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21973" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "cation [29, 84], OCR [45], layout analysis [65, 85], table recognition [52, 69, 92], key-value extraction [29, 30, 34], chart understanding [13, 26, 32, 33, 46, 47, 54, 55, 57, 82, 83], equation recognition [17, 23, 63, 68, 86], and more." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.155, + 0.482, + 0.292 + ], + "angle": 0, + "content": "Ensemble systems, such as the aforementioned open-source libraries [1, 50, 62, 77], implement a pipeline in source code, which conditionally applies specialized, single-task models on a given input document, composing the prediction results into a meaningful document representation. Typically, each task entails human-crafted handling of pre-processing and post-processing logic (i.e. setting confidence thresholds for layout detections, matching layout elements to text cells, conditionally applying models)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.485, + 0.597 + ], + "angle": 0, + "content": "Conversely, multi-task models aim at providing a single model capable of handling various document-understanding related tasks simultaneously, benefiting from shared context and shared representations learned across these different tasks. OCR-reliant methods, such as LayoutLM [29, 84, 85] and UDOP [70], use text extracted from an external OCR engine, along with image and text bounding box locations, as input. In contrast, OCR-free methods such as Donut [34], Dessurt [16], DocParser [18], Pix2Struct [40] are transformer-based models trained end-to-end to take an image as input and directly output text. Large Vision Language Models (LVLMs) such as LLaVA [48], LLaVA-OneVision [42], UReader [87], Kosmos-2 [64], and Qwen-VL [8] all leverage various vision encoders, projection adapters, and LLMs. While they may use different image patching techniques and instruction-tuning datasets to handle a range of vision tasks, they commonly evaluate their document understanding capabilities on datasets like DocVQA [56] and mPLUG-DocOwl 1.5 datasets [27], primarily focusing on question answering and reasoning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.6, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our work focuses primarily on document understanding tasks related to conversion and structural recognition, aiming to represent documents with high accuracy and completeness. The closest works to our approach are Nougat [12], DocOwl 2 [28], GOT [79], and Qwen2.5-VL [9]. DocOwl 2 and GOT are dedicated to precise conversion and recognition of document structures. On one hand, DocOwl 2 employs a dynamic shape-adaptive cropping module to process high-resolution images, feeding them into a ViT vision encoder. A dedicated module compresses the processed visual features before feeding them into a LLaMa-based LLM. On the other hand, GOT focuses on converting diverse elements—such as text, formulas, molecular diagrams, tables, charts, sheet music, and geometric shapes—into structured formats. It achieves a lightweight architecture by utilizing a SAM vision encoder [35], a linear projection layer, and a Qwen-0.5B LLM. Building on these approaches, our work not only extends the range of tasks but also broadens the types of documents handled, moving beyond a strictly scientific focus" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "to support a more diverse array of document formats. In parallel, Qwen2.5-VL introduces the Omni-Parsing strategy, that integrates multiple document elements into a unified HTML-based representation. This format encodes layout bounding box information and reading order for various document elements such as paragraphs and charts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.199, + 0.646, + 0.217 + ], + "angle": 0, + "content": "3.SmolDocling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Building on Hugging Face's SmolVLM [53], SmolDocling converts document page images into sequences representing their content and structure. It produces a unified page representation (DocTags markup) that is optimized for the LLM backbone and interpretable by Docling [50]. This new DocTags format covers the reproduction of captions, charts, forms, code, equations, tables, footnotes, lists, page footers and headers, section headings, and text. SmolDocling both recognizes the type and location of these elements as well as performs OCR within them. Our approach enables fullpage conversion with reading-order and hierarchical linking between elements (e.g., linking captions with figures and tables), providing bounding boxes for each component, supporting isolated predictions on individual elements (e.g., cropped tables) and grouping for lists and nested lists." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.702, + 0.48 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.7 + ], + "angle": 0, + "content": "The model architecture is illustrated in Figure 1. SmolVLM-256M relies on SigLIP base patch-16/512 (93M) as visual backbone and compared to the 2.2B version of the same model, its training data was rebalanced to emphasize document understanding (41%) and image captioning (14%), combining The Cauldron [39], Docmatix [38] datasets with the addition of MathWriting [23]. It uses lightweight variant of the SmolLM-2 family (135M) [2] as the language backbone and employs a radical pixel shuffle method that compresses each 512x512 image patch into 64 visual tokens. Last but not least, tokenization efficiency was also improved in SmolVLM-256M by increasing the pixel-to-token ratio to 4096 pixels per token and introducing special tokens for sub-image separators." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.617, + 0.728 + ], + "angle": 0, + "content": "3.2. DocTags" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To train VLMs for standardized document conversion, we introduce the DocTags markup format, illustrated in Figure 2. Inspired by OTSL [52], DocTags define a structured vocabulary of unambiguous tags and rules that explicitly separate textual content from document structure. This disentanglement improves performances of Imago-to-Sequence models by minimizing confusions. In contrast, direct conversion to standard formats such as HTML or Markdown is often lossy and ambiguous, lacking explicit structural markup, failing to preserve page layout context, and increasing total tokens count. Pieces" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.089, + 0.856, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.454, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Figure 2. DocTags format describes key features of document elements: type of an element (text, picture, table, code, etc.), location on the page, and content. Nested tags convey additional information: pictures and tables may nest captions, table structure is represented in OTSL tags, lists nest list items, code and pictures carry a classification. All DocTags output is taken from SmolDocling predictions, artificial line-breaks and dots (...) were inserted for better readability. A, B, C - snippets of various pages from [52], D - from [81]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.482, + 0.658 + ], + "angle": 0, + "content": "of text content that represent basic elements in DocTags are enclosed in XML-style tags. Tags define document block types such as: text, caption, footnote, formula, title, list_item, page Footer, page_header, picture, section_header, otsl, code, document_index, etc. Each element can nest additional location tags encoding its position on the page as a bounding box, represented in DocTags as \\(<\\mathrm{loc\\_x1}><\\mathrm{loc\\_y1}><\\mathrm{loc\\_x2}><\\mathrm{loc\\_y2}>\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Special blocks, such as tables and images, nest additional descriptors for captions, tabular structure, or image classes. To encode table structures, we fully include the OTSL vocabulary into DocTags. OTSL has the advantage of being minimalist and supporting cell spans and header information. Its tokens are interleaved with text, allowing simultaneous encoding of both table structure and content. In order to promote robust visual-semantic alignment in our document understanding pipeline, we maintain a uniform DocTags representation for cropped page elements (e.g., tables, code, equations) that is identical to their full-page counterparts. This consistency ensures that the model can leverage similar visual cues whether an element is presented in isolation or in the context of an entire page, thereby strengthening the learning signal for both computer vision and downstream machine learning tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.537, + 0.907, + 0.568 + ], + "angle": 0, + "content": "For a complete list of tags and detailed description of the format, consult the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.579, + 0.671, + 0.595 + ], + "angle": 0, + "content": "3.3. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.798 + ], + "angle": 0, + "content": "We employ a curriculum learning approach to progressively align our model for document conversion, ensuring faster convergence. As an initial step, we incorporate DocTags as tokens in our tokenizer. To align the LLM part we freeze the vision encoder and train only the remaining network to adapt it to the new output format which it hasn't seen before. To ensure comprehensive coverage of all DocTags, we maintain a balanced mix of tasks and data types during training. Next, we unfreeze the vision encoder and train the model on our pretraining datasets (see Sec. 4.1), along with all task-specific conversion datasets, including tables, code, equations, and charts. Finally, we fine-tune using all available datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.58, + 0.83 + ], + "angle": 0, + "content": "4. Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For document understanding tasks, each page image represents the visual layout of textual elements—such as paragraphs, formulas, tables, code, or figures with embedded textual content. Publicly available datasets that unify all" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21975" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "these characteristics remain limited, and existing corpora often suffer from small sample sizes or fragmented formats. In this section, we explain how we mitigate these issues in both pre-training and task-specific training scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.164, + 0.289, + 0.179 + ], + "angle": 0, + "content": "4.1. Pre-training datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.351 + ], + "angle": 0, + "content": "In the domain of document understanding, currently available open-access document pre-training datasets include OCR-IDL [10], IIT-CDIP [41], CCpdf [74], and WordScape [78]. OCR-IDL and IIT-CDIP, industry-sourced from the 1990s, provide only page images and OCR text. CCpdf and WordScape, derived from Common Crawl, focus on raw PDFs and Word documents, respectively. While WordScape enriches documents with layout, table structure, and topic information via XML markup, its visual variability is limited. Notably, none of these datasets offer rich annotations for all page elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.625 + ], + "angle": 0, + "content": "DocLayNet-PT. Given the lack of public multimodal document pre-training data with sufficient annotation features, we constructed a new document pre-training dataset named DocLayNet-PT. DocLayNet-PT is a 1.4M pages dataset extracted from the DocFM dataset [25] of unique PDF documents, sourced from CommonCrawl, Wikipedia, and business-related documents. We focused on this dataset to look for documents including visually distinct content such as equations, tables, code and charts with colorful layouts. This dataset was augmented with weak annotations through a series of processing steps. First, we applied PDF parsing to all pages to obtain the text layout including positional information using Docling [7, 50]. Next, we performed an enrichment step to obtain an enhanced, machine-friendly representation of each page. Specifically, we provide annotations for layout elements, table structure, language, topic, and figure classification on each page in DocTag format." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Docmatix. To preserve SmolVLM's original DocVQA capabilities, we adapted the same weak annotation strategy used for DocLayNet-PT and applied it to the 1.3M documents inside Docmatix dataset [38]. In addition to the 9.5 million DocVQA questions, we introduced another instruction requiring the full conversion of multi-page documents to DocTags." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.295, + 0.758 + ], + "angle": 0, + "content": "4.2. Task-specific Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Layout. To optimize prediction quality for document layout and table structure, we sampled 76K pages from DocLayNet-PT. Each page underwent human annotation and rigorous quality review, resulting in a high-quality ground-truth dataset. Of these, 60K pages were utilized for fine-tuning, constituting what we refer to as DocLayNet v2. Additionally, we extracted 63K pages from WordScape, specifically filtering for pages containing both text and tables, and utilized their inherently structure-preserving for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "mat as a reliable source of ground truth. To further enhance the model's adaptability to diverse layouts, colors, and fonts, we synthetically generated 250K pages using content from Wikipedia, creating the dataset known as SynthDocNet. This synthetic dataset enables the controlled generation of various document types with precise annotations. All datasets share a unified annotation format and class definitions, ensuring consistency and comprehensive document representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.905, + 0.38 + ], + "angle": 0, + "content": "Tables. For the table structure recognition task we trained our model on public datasets such as PubTables-1M [69], FinTabNet [92], WikiTableSet [51] and tabular information extracted from WordScape documents. We transformed table structure information from the original ground-truth to OTSL format [52]. Then, we interleaved each cell tag with ground-truth text. This yields a condensed sequence which represents the structure of the table (including merged cells, and header declaration) along with its textual content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.625 + ], + "angle": 0, + "content": "Charts. For chart reconstruction, several datasets are publicly available [32, 33, 54, 82] that consist of both synthetic and web-scraped charts. However, these datasets are limited by either the quantity of real-world data or the visual diversity of synthetic examples. This has prompted recent studies [13, 26, 47, 55, 57, 82, 83] to crawl or render additional datasets, which unfortunately are rarely shared publicly. To address these gaps, we generated our own chart dataset comprising four types of charts: line, pie, bar, and stacked bar. Using data from 90,000 tables sourced from the FinTabNet dataset [92], we generated a comprehensive dataset containing 5,000 line charts, 380,000 pie charts, 380,000 bar charts, and 77,000 stacked bar charts. Each data point was rendered using three different visualization libraries (details provided in the Supplementary Material), yielding a total of 2.5 million visually diverse charts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.853 + ], + "angle": 0, + "content": "Code. Technical books and scientific documents frequently include code snippets, making it essential for document-processing models to accurately interpret code structure, particularly indentation, which carries semantic significance in certain programming languages. Several publicly available datasets provide extensive collections of code snippets with permissive licenses in text format [36, 66], typically aggregated from sources like GitHub. However, a common limitation is that these datasets contain only text-based code snippets, lacking any image-based representation. To address this gap, we leveraged LaTeX and Pygments [15] to generate visually diverse renderings of code (details provided in the Supplementary Material). Our dataset includes 9.3M million rendered code snippets generated at 120 dpi, covering 56 programming languages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Equations. Accurately parsing mathematical formulas is crucial for document-processing models, as formulas often contain nuanced structures and semantics criti-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "cal for scientific understanding. To effectively train our model for formula transcription, we combined publicly available datasets with a custom-rendered dataset. The public datasets [11, 17, 22, 23] collectively comprise approximately \\(730\\mathrm{k}\\) unique formulas. In addition, we extracted 4.7 million formulas from arXiv by applying a regular expression to the source LaTeX code. The extracted formulas underwent a rigorous normalization procedure to ensure the model was trained only on correct and standardized LaTeX code. Subsequently, these normalized equations were visually rendered using LaTeX (details provided in the Supplementary Material). Overall, our dataset includes 5.5 million unique formulas, each rendered at a resolution of 120 dpi." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.298, + 0.485, + 0.639 + ], + "angle": 0, + "content": "
Dataset NameSizeType
TheCauldron [39]2.63M
DocLayNet-PT-Instruct1.62MMulti Task
DocMatix w/ Doc Conversion* [38]952K
SynthChartNet*2.5M
PlotQA [58]225K
FigureQA [32]100KChart
Unichart [55]600K
ChartQA [54]28K
DocLayNet-PT1.40M
Doc4MLlaVa - RVLCDIP742K
SynthDocNet375KConversion
DocLayNet [65]81K
DocLayNet v276K
WordScape-PT63K
PubMed1M [69]1M
WordScape Tables [78]207KTable
WikiTableSet (EN) [51]204K
FinTabNet [92]90K
SynthFormulaNet*5.5MFormula
SynthCodeNet*9.3MCode
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Table 1. Datasets. Overview of the datasets used for the training and evaluation of SmolDocling. Datasets we contribute as part of this publication are denoted with \\((^{*})\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.422, + 0.73 + ], + "angle": 0, + "content": "4.3. Document Instruction Tuning datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To reinforce the recognition of different page elements and introduce document-related features and no-code pipelines, we leveraged rule-based techniques and the Granite-3.1-2b-instruct [25] LLM. Using samples from DocLayNet-PT pages, we generated one instruction by randomly sampling layout elements from a page. These instructions included tasks such as \"Perform OCR atbbox,\" \"Identify page element type atbbox,\" and \"Extract all section headers from the page.\" In addition, we train with Cauldron [39] to avoid catastrophic forgetting with the amount of conversation datasets introduced." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.252 + ], + "angle": 0, + "content": "In this section, we assess the performance of SmolDocling on document understanding tasks, and compare the results to previously published models and baselines. Capturing the qualities and defects of SmolDocling and other models in comparable numbers is a non-trivial effort for reasons we outline further below. Therefore, we report results on standard evaluation tasks and metrics independently (see section 5.2), but also provide a qualitative assessment of output characteristics (see section 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.259, + 0.731, + 0.274 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Our training was conducted using 64 NVIDIA A100 80GB GPUs, with each epoch requiring 28 hours. The model was trained for a total of 4 epochs using the AdamW optimizer with a learning rate of \\(2 \\times 10^{-4}\\) and \\(2 \\times 10^{-6}\\) for the vision encoder once unfrozen. We also employed gradient clipping to 1.0 for stability and a warmup ratio of 0.03. For inference, using VLLM [37] on an A100 GPU, SmolDocling achieves a page conversion time of 0.35 seconds while occupying only 0.489 GB of VRAM. The model supports a maximum sequence length of 8,192 tokens and is trained to process up to three pages at a time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.708, + 0.469 + ], + "angle": 0, + "content": "5.2.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.52 + ], + "angle": 0, + "content": "To enable valid cross-model comparisons with SmolDooling, we addressed multiple dataset and harmonization challenges:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.522, + 0.906, + 0.612 + ], + "angle": 0, + "content": "- Public evaluation datasets lack comprehensive multi-task annotations with diverse layouts. We augmented DocLayNet [65] with text content and reading-order to enable evaluation of both layout analysis and text recognition. Table structure, code listing, formula and chart recognition are evaluated on specific datasets." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.613, + 0.906, + 0.748 + ], + "angle": 0, + "content": "- Different models each follow their own conventions for output markup (e.g., Markdown, HTML, DocTags) and produce annotations which do not fully align in semantics. For example, SmolDocling is trained to produce 14 distinct layout-class tags via DocTags, while Qwen2.5-VL combines standard HTML tags and a few explicit class attributes with only partially compatible definitions. GOT and Nougat output Markdown, which can express paragraphs, lists and tables without location annotation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.749, + 0.905, + 0.823 + ], + "angle": 0, + "content": "- Input resolution of page images can affect model performance, but is rarely specified in literature. We standardized at 144 DPI (dots per inch), which reproduces enough detail to be perfectly legible by humans while keeping resource demand for inference reasonable." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.522, + 0.906, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Text Recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21977" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.346 + ], + "angle": 0, + "content": "
MethodSizeEdit Distance ↓F1-score ↑Precision ↑Recall ↑BLEU ↑METEOR ↑
Full-page \nDocling† [7] v2.25-0.400.910.930.920.720.80
Full-page \nQwen2.5 VL [9]7B0.560.720.800.700.460.57
GOT [79]580M0.610.690.710.730.480.59
Nougat (base) [12]350M0.620.660.720.670.440.54
SmolDocling (Ours)256M0.480.800.890.790.580.67
Code listings \nSmolDocling (Ours)256M0.110.920.940.910.870.89
Equations \nQwen2.5 VL [9]7B0.220.890.910.870.680.77
GOT [79]580M0.110.950.950.960.850.91
Nougat (base) [12]350M0.620.600.600.530.330.41
SmolDocling (Ours)256M0.110.950.960.950.830.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.357, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Table 2. Structured document text recognition. We evaluate OCR performance and document formatting on DocLayNet, focusing on textual elements, excluding tables. Text accuracy is measured through multiple metrics (Edit Distance, F1-Score, Precision, Recall, BLEU and METEOR) as proposed in [12]." + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.4, + 0.592, + 0.413 + ], + "angle": 0, + "content": "Values for Dockling provide a reference for the performance of an ensemble system." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.482, + 0.56 + ], + "angle": 0, + "content": "non-textual elements. For different tasks, we conduct evaluations on diverse content types: full-page documents (from DocLayNet test set) in Markdown format, code snippets (from our SynthCodeNet dataset) in plaintext, and equations (from Im2Latex-230k [21]) in LaTeX format. For Qwen2.5-VL and SmolDocling, we convert HTML or DocTags output to Markdown while preserving formatting where possible." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.563, + 0.483, + 0.715 + ], + "angle": 0, + "content": "Tables are omitted from the full-page output since they do not compare well through Markdown, due to loss of row and column spans. Code snippets must be reproduced with intact line breaks and indentation for accurate scoring. Formulas are evaluated separately in LaTeX format, which is natively produced by all the models tested in this work. The LaTeX outputs from all models, including the ground truth, undergo the same normalization procedure used for generating the SynthFormulaNet dataset. This ensures consistency and fairness in model comparisons." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.716, + 0.483, + 0.853 + ], + "angle": 0, + "content": "SmolDocling significantly outperforms GOT, Nougat and also Qwen2.5-VL, which is 27 times larger, on every metric in full-page transcription (see Table 2. For the code parsing task, we report results exclusively for SmolDocling, as we introduce this task for the first time, and the other models were not explicitly trained for it. SmolDocling also demonstrates superior performance compared to Nougat and Qwen2.5-VL in the formula recognition task, closely matching GOT's results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Layout. We evaluate element localization accuracy using the DocLayNet test set (Table 3), measuring the ability to reconstruct document representations that support con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.668 + ], + "angle": 0, + "content": "tent grounding and visual feature extraction. SmolDocling is compared against Qwen2.5-VL-7b [9], currently the only other VLM addressing this task. For a fair comparison, we map each model's label outputs and the ground-truth to six compatible classes: Text, Section Heading, List Item, Table, Picture, and Formula. Overall, SmolDocling outperforms Qwen2.5-VL-7b by a significant margin, however both models score far below a human baseline. This can be partially explained through inherent difficulty of the dataset, as indicated by relatively low human agreement scores. The scores are primarily impacted by moderate recall of element bounding boxes and different degrees of label confusion in both SmolDocling and Qwen2.5-VL. A majority of samples show accurate element bounding boxes as seen in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.773 + ], + "angle": 0, + "content": "
Docling [7] v2.25SmolDoclingQwen2.5-VLHuman
Formula0.7370.2670.0590.84
Table0.6670.2660.2620.79
List Item0.5280.2960.0900.87
Section Header0.4710.2890.1290.83
Text0.4440.2040.1800.85
Picture0.3280.0660.0780.70
Overall0.5290.2310.1330.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.784, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Table 3. Layout analysis comparison. Evaluation of layout analysis task on DocLayNet with a subset of 6 class labels in mAP[0.5:0.95] metric. Additional reference points are provided for an estimate of human performance [65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Table Structure Recognition. We evaluate the accuracy of recovering table structures, i.e. shaping columns and rows with the correct text content and spans, through" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21978" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "the TEDS metric [93], once including table text content and once on structure only with omitted text (see Table 4). For isolated assessment, we provide cropped table images from FinTabNet [92] and PubTables-1M [69] test sets, which have poor image quality (72 dpi, compression artifacts). SmolDocling performs competitively against significantly larger models, despite challenges with text transcription from the low-resolution image crops, a limitation likely attributable to insufficient training at such resolutions. Scores for structure only TEDS evaluation are therefore much stronger." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.269, + 0.483, + 0.352 + ], + "angle": 0, + "content": "
MethodModel SizeFinTabNetPubTables-1M
SmolVLM [53]2.2B0.180.32
Granite Vision [72]2.98B0.540.7
TableFormer† [52]52M0.89 f0.84 (0.86)
Qwen2.5-VL 7B7B0.66- (0.786)-
SmolDocling (Ours)256M0.72 (0.81)0.83 (0.88)
SmolDocling (Ours) @150DPI256M0.76 (0.82)-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Table 4. Table structure and cell content reconstruction. Numbers are given in TEDS metric including text content, with structure-only score in brackets where available. We produced results for SmolDocling and TableFormer, other results are reflected here as published. †TableFormer extracts the content through external OCR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Chart Extraction. In this task, we evaluate the ability to convert cropped sections of charts into datapoint tables on the ChartQA [54] dataset. All table output is transformed into HTML and compared against the ground-truth HTML using TEDS score, as reported in Table 5. We attribute SmolDocling's performance on charts to the variability in ground-truth styles and the differing interpretations of representations across datasets, which we observe leads to inconsistencies in the model's predictions. Despite its significantly smaller size, SmolDocling achieves a competitive TEDs score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Overall, SmolDocling achieves leading performance across a wide range of document conversion tasks and outperforms significantly larger models in text recognition, layout analysis and structure extraction. Furthermore, the model's high performance in downstream tasks, such as code listing and formula recognition, demonstrates its versatility and real-world applicability. Notably, we establish benchmark results for the code listing recognition, setting a standard for future evaluations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.274, + 0.789 + ], + "angle": 0, + "content": "5.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.87 + ], + "angle": 0, + "content": "Several characteristics of SmolDocling output are not evident from task-based metrics alone. In this section, we highlight different output characteristics observed in practice. More illustrated samples can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As presented in Figure 2, SmolDocling outputs a sequence of DocTags to encode content, structure and layout." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.089, + 0.855, + 0.184 + ], + "angle": 0, + "content": "
ModelSizeTEDs↑
Phi-3.5-vision4B0.40
Granite Vision [72]3B0.95
SmolVLM [53]2.2B0.02
Molmo-E1B0.54
SmolDocling (Ours)256M0.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Table 5. Chart analysis. We compare SmolDocling to other small-sized VLMs on the chart-to-table task using the TEDS metric on the ChartQA dataset. The results in this table are based on the work reported in [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.905, + 0.405 + ], + "angle": 0, + "content": "A few notable traits beyond straight content reproduction include: 1) Labeling of page headers and footers, which allows ignoring these repetitive elements in downstream applications. 2) Linking of information, such as captions to tables or pictures, or list elements to parent list through nesting of tags. 3) Preserved line breaks and indentation for code snippets, while line wraps in paragraphs are correctly removed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Among typical defects found in the output of SmolDocling are missing tags (e.g. absent location tags), malformed structure (such as missing closing tags), and endless repetition of tokens from some point in a page. Both of these phenomena are commonly seen in auto-regressive models and may prevent parsing or rendering the output correctly (see supplementary material). Despite these limitations, SmolDocling demonstrates unique robustness advantages compared to ensemble systems like Docling, since conversion output is inferred in a single pass. In the latter, errors may accumulate throughout the ensemble model pipeline. For example, a mislocated table detection will lead to missing or distorted output of table structure and content, since this depends on correct table localization. In contrast, SmolDocling can reproduce the tables correctly, even with incorrectly determined location." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.676, + 0.634, + 0.691 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce SmolDocling, an efficient and compact VLM optimized for document conversion with rich output representation. We also present a suite of new datasets with a unified format for document conversion, including the novel task of code listing transcription. We identify page element localization as a critical area requiring further refinement, where targeted techniques will significantly enhance performance in future iterations. Our results conclusively demonstrate that smaller models with unified, optimized output formats such as DocTags can effectively compete with significantly larger models, establishing a clear pathway for resource-efficient multi-task document understanding models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgements: We thank our collaborators at Hugging Face, Andrés Marafioti and Miquel Farré, who kindly and rapidly developed SmolVLM-256M-Instruct in response to our collaboration, providing the foundational model upon which this work is built." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[1] Grobid. https://github.com/kermitt2/grobid, 2008-2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[2] Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martin Blázquez, Guilherme Penedo, Lewis Tunstall, Andres Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werra, and Thomas Wolf. SmolLM2: When Smol Goes Big - Data-Centric Training of a Small Language Model, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[3] Amazon Web Services. Amazon Texttract. https://aws.amazon.com/texttract/, 2024. Accessed: 2024-11-11. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[4] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[5] Anthropic. Claude-3.5. https://www.anthropic.com/news/claude-3-5-sonnet, 2024. Accessed: 2024-02-11. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[6] Christoph Auer, Michele Dolfi, André Carvalho, Cesar BerroSpi Ramis, and Peter WJ Staar. Delivering Document Conversion as a Cloud Service with High Throughput and Responsiveness. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD), pages 363-373. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[7] Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Nikolaos Livathinos, Panos Vagenas, Cesar Berrospi Ramis, Matteo Omenetti, Fabian Lindlbauer, Kasper Dinkla, et al. Docling Technical Report. arXiv preprint arXiv:2408.09869, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL Technical Report, 2025. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[10] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. In European Conference on Computer Vision, pages 241-252. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[11] Lukas Blecher. LaTeX-OCR, 2024. Accessed: 2024-11-09. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[13] Jinyue Chen, Lingyu Kong, Haoran Wei, Chenglong Liu, Zheng Ge, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. OneChart: Purify the Chart Structural Extraction via One Auxiliary Token. arXiv preprint arXiv:2404.09987, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[15] Pygments Contributors. Pygments: Python syntax highlighter, 2024. Accessed: 2024-11-06. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[16] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[17] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning, pages 980-989. PMLR, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.45, + 0.906, + 0.518 + ], + "angle": 0, + "content": "[18] Mohamed Dhouib, Ghassen Bettaieb, and Aymen Shabou. Docparser: End-to-endOCR-free information extraction from visually rich documents. In Document Analysis and Recognition - ICDAR 2023, pages 155-172, Cham, 2023. Springer Nature Switzerland. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model. arXiv preprint arXiv:2401.16420, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Zhe Chen, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Kai Chen, Conghui He, Xingcheng Zhang, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD. arXiv preprint arXiv:2404.06512, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[21] Gregory Eritsyan. Im2LaTeX 230K. https://www.kaggle.com/datasets/gregoryeritsyan/im2latex-230k, 2023. Accessed: 2025-03-06. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[22] Hugging Face. latex-formulas Dataset, 2024. Accessed: 2024-11-09. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Philippe Gervais, Asya Fadeeva, and Andrii Maksai. Math-writing: A dataset for handwritten mathematical expression recognition. URL https://arxiv.org/abs/2404.10690, 2024.3, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21980" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[24] Google Cloud. Document AI. https://cloud.google.com/document-ai, 2024. Accessed: 2024-11-11. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[25] IBM Granite Team. Granite 3.0 Language Models, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[26] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[27] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[28] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), pages 1–6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[32] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[33] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[34] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[36] Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Munoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The Stack: 3 TB of permissively licensed source code. Preprint, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[37] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[39] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models?, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.364 + ], + "angle": 0, + "content": "[40] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[41] David Lewis, Gady Agam, Shlomo Argamon, Ophir Frieder, David Grossman, and Jefferson Heard. Building a test collection for complex document information processing. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 665-666, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.452, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.51, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[43] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.64, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[45] Han Lin, Peng Yang, and Fanlong Zhang. Review of scene text detection and recognition. Archives of computational methods in engineering, 27(2):433-454, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.685, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[46] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[47] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.829, + 0.905, + 0.869 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21981" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[50] Nikolaos Livathinos, Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Panos Vagenas, Cesar Berropi Ramis, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Shubham Gupta, Rafael Teixeira de Lima, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, and Peter W. J. Staar. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion, 2025. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[51] Nam Ly., Atsuhiro Takasu., Phuc Nguyen., and Hideaki Takeda. Rethinking Image-Based Table Recognition Using Weakly Supervised Methods. pages 872-880, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[52] Maksym Lysak, Ahmed Nassar, Nikolaos Livathinos, Christoph Auer, and Peter Staar. Optimized Table Tokenization for Table Structure Recognition. In Document Analysis and Recognition - ICDAR 2023, pages 37-50, Cham, 2023. Springer Nature Switzerland. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[53] Andres Marafioti, Orr Zohar, Miquel Farre, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Alal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. SmolVLM: Redefining small and efficient multimodal models. 2025. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[54] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[55] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[56] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[57] Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. Chartassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. arXiv preprint arXiv:2401.02384, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[58] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. PlotQA: Reasoning over Scientific Plots. In The IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[59] Microsoft Azure. AI Document Intelligence. https://azure.microsoft.com/en-us/products/ai-services.ai-document-intelligence, 2024. Accessed: 2024-11-11. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[60] OpenAI. Hello GPT-4O. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-02-09, 2024-02-11, 2024-02-12. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[61] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[62] Vik Paruchuri. Marker: Convert PDF to Markdown Quickly with High Accuracy. https://github.com/VikParuchuri/marker, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[63] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2024. Accessed: 2024-11-11. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[65] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 3743–3751, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[66] Ruchir Puri, David Kung, Geert Janssen, Wei Zhang, Giacomo Domeniconi, Vladimir Zolotov, Julian Dolby, Jie Chen, Mihir Choudhury, Lindsey Decker, Veronika Thost, Luca Buratti, Saurabh Pujar, Shyam Ramji, Ulrich Finkler, Susan Malaika, and Frederick Reiss. CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Tasks, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[67] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[68] Felix M Schmitt-Koopmann, Elaine M Huang, Hans-Peter Hutter, Thilo Stadelmann, and Alireza Darvishy. MathNet: A Data-Centric Approach for Printed Mathematical Expression Recognition. IEEE Access, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[69] Brandon Smock, Rohith Pesala, and Robin Abraham. PubTables-1M: Towards comprehensive table extraction from unstructured documents. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4624–4632, 2022. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[70] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[71] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[72] Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kondic, Nimrod Shabtay, Pengyuan Li, Roei Herzig, Shafiq Abedin, Shaked Perek, Sivan Harary, Udi Barzelay, Adi Raz Goldfarb, Aude Oliva, Ben Wieles, Bishwaranjan Bhattacharjee, Brandon Huang, Christoph Auer, Dan GUTFreund, David Beymer, David Wood, Hilde Kuehne, Jacob Hansen, Joseph Shtok, Ken Wong, Luis Angel Bathen, Mayank Mishra, Maksym Lysak, Michele Dolfi, Mikhail Yurochkin, Nikolaos Livathinos, Nimrod Harel, Ophir Azulai, Oshri Naparstek, Rafael Teixeira de Lima, Rameswar Panda, Sivan Doveh, Shubham Gupta, Subhro Das, Syed Zawad, Yusik Kim, Zexue He, Alexander Brooks, Gabe Goodhart, Anita Govindjee, Derek Leist, Ibrahim Ibrahim, Aya Soffer, David Cox," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21982" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "Kate Soule, Luis Lastras, Nirmit Desai, Shila Ofek-koifman, Sriram Raghavan, Tanveer Syeda-Mahmood, Peter Staar, Tal Drory, and Rogerio Feris. Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[74] Michal Turski, Tomasz Stanisławek, Karol Kaczmarek, Paweł Dyda, and Filip Graliński. CCpdf: Building a High Quality Corpus for Visually Rich Documents from Web Crawl Data. In Document Analysis and Recognition - IC-DAR 2023, pages 348-365, Cham, 2023. Springer Nature Switzerland. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[75] Unstructured.io Team. Unstructured.io: Open-Source Pre-Processing Tools for Unstructured Data. https://unstructured.io, 2024. Accessed: 2024-11-19. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[76] Vicuna. Vicuna: An open-source chatbot impressing GPT-4 with \\(90\\%\\) ChatGPT quality. https://vicuna.lmsys.org/, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[77] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, Bo Zhang, Liquun Wei, Zhihao Sui, Wei Li, Botian Shi, Yu Qiao, Dahua Lin, and Conghui He. MinerU: An Open-Source Solution for Precise Document Content Extraction, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[78] Maurice Weber, Carlo Siebenschuh, Rory Butler, Anton Alexandrov, Valdemar Thanner, Georgios Tsolakis, Haris Jabbar, Ian Foster, Bo Li, Rick Stevens, et al. WordScape: a Pipeline to extract multilingual, visually rich Documents with Layout Annotations from Web Crawl Data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[79] Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model. arXiv preprint arXiv:2409.01704, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[80] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[81] Wes McKinney and the Pandas Development Team. Pandas: powerful Python data analysis toolkit, Release 1.4.4. https://pandas.pydata.org/pandas-docs/ version/1.4/pandas.pdf, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[82] Renqiu Xia, Bo Zhang, Haoyang Peng, Hancheng Ye, Xiangchao Yan, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. StructChart: Perception, Structuring, Reasoning for Visual Chart Understanding. arXiv preprint arXiv:2309.11268, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[83] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[84] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[85] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[86] Zuoyu Yan, Xiaode Zhang, Liangcai Gao, Ke Yuan, and Zhi Tang. ConvMath: a convolutional sequence network for mathematical expression recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4566-4572. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[87] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[88] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition. arXiv preprint arXiv:2309.15112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[89] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer-2.5: A Versatile Large Vision Language Model Supporting Long-Contextual Input and Output. arXiv preprint arXiv:2407.03320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.648 + ], + "angle": 0, + "content": "[90] Qiming Zhang, Jing Zhang, Yufei Xu, and Dacheng Tao. Vision Transformer with Quadrangle Attention. arXiv preprint arXiv:2303.15105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[91] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pretrained transformer language models. URL https://arxiv.org/abs/2205.01068, 3:19-0, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[92] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 697-706, 2021. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[93] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. arXiv preprint arXiv:1911.10683, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[94] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21983" + } + ] +] \ No newline at end of file diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_origin.pdf b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2852ad6a9a6fc329931460af0a3bfbcb5fba2d48 --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/52fdd715-223a-408c-b30d-b6b134926ef6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f491c151e698e1aa501423dd4413d4c89ac5cecd62fee1a6c675ffd9b54475d +size 1103465 diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/full.md b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b765b4743e660b9bb4467b35e42403a6ca85e28 --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/full.md @@ -0,0 +1,297 @@ +# SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion + +Ahmed Nassar Matteo Omenetti Maksym Lysak Nikolaos Livathinos Christoph Auer Lucas Morin +Rafael Teixeira de Lima Yusik Kim A. Said Gurbuz Michele Dolfi Peter W. J. Staar +{ahn, mly, nli, cau, dol, taa}@zurich.ibm.com, {Matteo.Omenetti1}@ibm.com +IBM Research + +# Abstract + +We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms — significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition. Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model weights and datasets are available at: https://huggingface.co/ds4sd + +# 1. Introduction + +For decades, converting complex digital documents into a structured, machine-processable format has been a significant technical challenge. This challenge primarily stems from the substantial variability in document layouts and styles, as well as the inherently opaque nature of the widely used PDF format, which is optimized for printing rather than semantic parsing. Intricate layout styles and visually challenging elements such as forms, tables, and complex charts can significantly impact the reading order and gen + +eral understanding of documents. These problems have driven extensive research and development across multiple domains of computer science. On one hand, sophisticated ensemble systems emerged, which decompose the conversion problem into several sub-tasks (e.g. OCR, layout analysis, table structure recognition, classification) and tackle each sub-task independently. Although such systems can achieve high-quality results for many document types while maintaining relatively low computational demands, they are often difficult to tune and generalize. + +On the other hand, in the recent past, great interest has developed around large foundational multimodal models that can solve the whole conversion task in one shot, while simultaneously offering flexible querying and parametrization through prompts. This approach has been made possible by the advent of multimodal pre-training of large vision-language models (LVLMs), which opened up a vast array of opportunities for leveraging diverse data sources, including PDF documents. However, the literature on this topic highlights a significant gap in the availability of high-quality and open-access datasets suitable for training robust multimodal models for the task of document understanding. Furthermore, relying on LVLMs may introduce common issues associated with such models, including hallucinations and the use of significant computational resources, making them impractical from both a quality and cost perspective. + +In this work, we outline how we close the gaps left by publicly available datasets and establish a training approach to achieve end-to-end, full-featured document conversion through a vision-language model, which we name SmolDoccling. Our main contributions can be summarized as follows: + +- SmolDocling: An ultra-compact VLM for end-to-end document conversion, based on Hugging Face's most recent SmolVLM-256M [53], which is between 5 and 10 times smaller in parameters than comparable VLMs tuned on document understanding tasks. Hence, we reduce the computational complexity to roughly the same order of magnitude as typical representatives of ensemble meth + +![](images/0211634d0b93a26372b178c2ce62ccc68ddba7148f16e50aafb340745dbdf4fb.jpg) +Figure 1. SmolDocling/SmolVLM architecture. SmolDocling converts images of document pages to DocTags sequences. First, input images are encoded using a vision encoder and reshaped via projection and pooling. Then, the projected embeddings are concatenated with the text embeddings of the user prompt, possibly with interleaving. Finally, the sequence is used by an LLM to autoregressively predict the DocTags sequence. + +ods. Uniquely, SmolDocling learns and produces a unified document representation that captures content, structure and spatial location of all elements in a page. + +- We augment existing document pre-training datasets with additional feature annotations, and create new datasets for tasks with insufficient coverage in open multi-modal training datasets, including code listings, equations, and chart understanding. Importantly, we construct full-page ground truth data that combines key annotation features such as layout, table structure, code listings, charts, and formulas in their natural context, and derive instruction datasets from all training datasets used. +- We introduce DocTags, a document markup format optimized to efficiently represent the full content and layout characteristics of a document, inspired by [52]. +- We evaluate the achieved prediction quality of SmolDoccling on several tasks, including comparisons to popular community models. + +# 2. Related Work + +# 2.1. Large Vision-Language Models (LVLMs) + +With the advent of large language models (LLMs) [2, 4, 8, 25, 31, 61, 67, 73, 80] characterized by their powerful contextual understanding and adaptability, several approaches have emerged to extend them into large vision-language models. Some proprietary, closed-source LLMs have been adapted into multimodal large language models (MLLMs), such as GPT-4o [60], Gemini [71], and Claude 3.5 [5], demonstrating exceptional capabilities across various modalities, including vision. Open-source methods are actively advancing to match the capabilities of proprietary MLLMs. BLIP-2 [44] is one of the earliest works combining a vision encoder with a frozen LLM (OPT [91] or FlanT5 [14]) using a lightweight transformer (Q-former + +[90]). Building upon BLIP-2, MiniGPT-4 [94] integrated a frozen visual encoder with a frozen LLM (Vicuna [76]) using a Q-Former network and a single projection layer. LLaVA [48, 49] employs a similar architecture but uses a minimal adapter layer. LLaVA-OneVision [42] and LLaVA-NeXT-Interleave [43] build upon LLaVA to support multi-image, higher resolution, and video understanding. For handling high resolutions and text-image compositions and comprehension, InternLM-XComposer introduces a family of models [19, 20, 88, 89] designed specifically for these tasks. Qwen-VL [8] introduces a position-aware adapter to address efficiency issues caused by the vision encoder generating long image feature sequences. Qwen2.5-VL [9] employs windowed attention with 2D rotary positional embeddings to process native-resolution inputs efficiently and integrates an vision-language merger for dynamic feature compression. + +In this work, we build upon the SmolVLM architecture approach [53]. Precisely, we adopt SmolVLM-256M, a compact variant of base SmolVLM-2.2B model, inspired by the architecture of Idefics3 [38] which employs the shape-optimized SigLIP encoder as its visual backbone. It applies an aggressive pixel shuffle strategy to compress visual features, reduces the number of image hidden states and introduces new special tokens to separate sub-images and improve tokenization efficiency. + +# 2.2. State of the Art in Document Understanding + +Several powerful document understanding solutions have emerged on the market in the past decade as commercial cloud offerings on hyperscalers [3, 6, 24, 59], frontier models such as GPT-4o [60] or Claude [5], or as open-source libraries such as Docling [50], Grobid [1], Marker [62], MinerU [77] or Unstructured [75]. Typical tasks that document understanding encompasses are document classifi + +cation [29, 84], OCR [45], layout analysis [65, 85], table recognition [52, 69, 92], key-value extraction [29, 30, 34], chart understanding [13, 26, 32, 33, 46, 47, 54, 55, 57, 82, 83], equation recognition [17, 23, 63, 68, 86], and more. + +Ensemble systems, such as the aforementioned open-source libraries [1, 50, 62, 77], implement a pipeline in source code, which conditionally applies specialized, single-task models on a given input document, composing the prediction results into a meaningful document representation. Typically, each task entails human-crafted handling of pre-processing and post-processing logic (i.e. setting confidence thresholds for layout detections, matching layout elements to text cells, conditionally applying models). + +Conversely, multi-task models aim at providing a single model capable of handling various document-understanding related tasks simultaneously, benefiting from shared context and shared representations learned across these different tasks. OCR-reliant methods, such as LayoutLM [29, 84, 85] and UDOP [70], use text extracted from an external OCR engine, along with image and text bounding box locations, as input. In contrast, OCR-free methods such as Donut [34], Dessurt [16], DocParser [18], Pix2Struct [40] are transformer-based models trained end-to-end to take an image as input and directly output text. Large Vision Language Models (LVLMs) such as LLaVA [48], LLaVA-OneVision [42], UReader [87], Kosmos-2 [64], and Qwen-VL [8] all leverage various vision encoders, projection adapters, and LLMs. While they may use different image patching techniques and instruction-tuning datasets to handle a range of vision tasks, they commonly evaluate their document understanding capabilities on datasets like DocVQA [56] and mPLUG-DocOwl 1.5 datasets [27], primarily focusing on question answering and reasoning. + +Our work focuses primarily on document understanding tasks related to conversion and structural recognition, aiming to represent documents with high accuracy and completeness. The closest works to our approach are Nougat [12], DocOwl 2 [28], GOT [79], and Qwen2.5-VL [9]. DocOwl 2 and GOT are dedicated to precise conversion and recognition of document structures. On one hand, DocOwl 2 employs a dynamic shape-adaptive cropping module to process high-resolution images, feeding them into a ViT vision encoder. A dedicated module compresses the processed visual features before feeding them into a LLaMa-based LLM. On the other hand, GOT focuses on converting diverse elements—such as text, formulas, molecular diagrams, tables, charts, sheet music, and geometric shapes—into structured formats. It achieves a lightweight architecture by utilizing a SAM vision encoder [35], a linear projection layer, and a Qwen-0.5B LLM. Building on these approaches, our work not only extends the range of tasks but also broadens the types of documents handled, moving beyond a strictly scientific focus + +to support a more diverse array of document formats. In parallel, Qwen2.5-VL introduces the Omni-Parsing strategy, that integrates multiple document elements into a unified HTML-based representation. This format encodes layout bounding box information and reading order for various document elements such as paragraphs and charts. + +# 3.SmolDocling + +Building on Hugging Face's SmolVLM [53], SmolDocling converts document page images into sequences representing their content and structure. It produces a unified page representation (DocTags markup) that is optimized for the LLM backbone and interpretable by Docling [50]. This new DocTags format covers the reproduction of captions, charts, forms, code, equations, tables, footnotes, lists, page footers and headers, section headings, and text. SmolDocling both recognizes the type and location of these elements as well as performs OCR within them. Our approach enables fullpage conversion with reading-order and hierarchical linking between elements (e.g., linking captions with figures and tables), providing bounding boxes for each component, supporting isolated predictions on individual elements (e.g., cropped tables) and grouping for lists and nested lists. + +# 3.1. Model Architecture + +The model architecture is illustrated in Figure 1. SmolVLM-256M relies on SigLIP base patch-16/512 (93M) as visual backbone and compared to the 2.2B version of the same model, its training data was rebalanced to emphasize document understanding (41%) and image captioning (14%), combining The Cauldron [39], Docmatix [38] datasets with the addition of MathWriting [23]. It uses lightweight variant of the SmolLM-2 family (135M) [2] as the language backbone and employs a radical pixel shuffle method that compresses each 512x512 image patch into 64 visual tokens. Last but not least, tokenization efficiency was also improved in SmolVLM-256M by increasing the pixel-to-token ratio to 4096 pixels per token and introducing special tokens for sub-image separators. + +# 3.2. DocTags + +To train VLMs for standardized document conversion, we introduce the DocTags markup format, illustrated in Figure 2. Inspired by OTSL [52], DocTags define a structured vocabulary of unambiguous tags and rules that explicitly separate textual content from document structure. This disentanglement improves performances of Imago-to-Sequence models by minimizing confusions. In contrast, direct conversion to standard formats such as HTML or Markdown is often lossy and ambiguous, lacking explicit structural markup, failing to preserve page layout context, and increasing total tokens count. Pieces + +![](images/697ddb11402df582bbb872662903ea9f82bca7a09917fcb97dc904996e087583.jpg) +Figure 2. DocTags format describes key features of document elements: type of an element (text, picture, table, code, etc.), location on the page, and content. Nested tags convey additional information: pictures and tables may nest captions, table structure is represented in OTSL tags, lists nest list items, code and pictures carry a classification. All DocTags output is taken from SmolDocling predictions, artificial line-breaks and dots (...) were inserted for better readability. A, B, C - snippets of various pages from [52], D - from [81] + +of text content that represent basic elements in DocTags are enclosed in XML-style tags. Tags define document block types such as: text, caption, footnote, formula, title, list_item, page Footer, page_header, picture, section_header, otsl, code, document_index, etc. Each element can nest additional location tags encoding its position on the page as a bounding box, represented in DocTags as $<\mathrm{loc\_x1}><\mathrm{loc\_y1}><\mathrm{loc\_x2}><\mathrm{loc\_y2}>$ . + +Special blocks, such as tables and images, nest additional descriptors for captions, tabular structure, or image classes. To encode table structures, we fully include the OTSL vocabulary into DocTags. OTSL has the advantage of being minimalist and supporting cell spans and header information. Its tokens are interleaved with text, allowing simultaneous encoding of both table structure and content. In order to promote robust visual-semantic alignment in our document understanding pipeline, we maintain a uniform DocTags representation for cropped page elements (e.g., tables, code, equations) that is identical to their full-page counterparts. This consistency ensures that the model can leverage similar visual cues whether an element is presented in isolation or in the context of an entire page, thereby strengthening the learning signal for both computer vision and downstream machine learning tasks. + +For a complete list of tags and detailed description of the format, consult the supplementary material. + +# 3.3. Model Training + +We employ a curriculum learning approach to progressively align our model for document conversion, ensuring faster convergence. As an initial step, we incorporate DocTags as tokens in our tokenizer. To align the LLM part we freeze the vision encoder and train only the remaining network to adapt it to the new output format which it hasn't seen before. To ensure comprehensive coverage of all DocTags, we maintain a balanced mix of tasks and data types during training. Next, we unfreeze the vision encoder and train the model on our pretraining datasets (see Sec. 4.1), along with all task-specific conversion datasets, including tables, code, equations, and charts. Finally, we fine-tune using all available datasets. + +# 4. Data + +For document understanding tasks, each page image represents the visual layout of textual elements—such as paragraphs, formulas, tables, code, or figures with embedded textual content. Publicly available datasets that unify all + +these characteristics remain limited, and existing corpora often suffer from small sample sizes or fragmented formats. In this section, we explain how we mitigate these issues in both pre-training and task-specific training scenarios. + +# 4.1. Pre-training datasets + +In the domain of document understanding, currently available open-access document pre-training datasets include OCR-IDL [10], IIT-CDIP [41], CCpdf [74], and WordScape [78]. OCR-IDL and IIT-CDIP, industry-sourced from the 1990s, provide only page images and OCR text. CCpdf and WordScape, derived from Common Crawl, focus on raw PDFs and Word documents, respectively. While WordScape enriches documents with layout, table structure, and topic information via XML markup, its visual variability is limited. Notably, none of these datasets offer rich annotations for all page elements. + +DocLayNet-PT. Given the lack of public multimodal document pre-training data with sufficient annotation features, we constructed a new document pre-training dataset named DocLayNet-PT. DocLayNet-PT is a 1.4M pages dataset extracted from the DocFM dataset [25] of unique PDF documents, sourced from CommonCrawl, Wikipedia, and business-related documents. We focused on this dataset to look for documents including visually distinct content such as equations, tables, code and charts with colorful layouts. This dataset was augmented with weak annotations through a series of processing steps. First, we applied PDF parsing to all pages to obtain the text layout including positional information using Docling [7, 50]. Next, we performed an enrichment step to obtain an enhanced, machine-friendly representation of each page. Specifically, we provide annotations for layout elements, table structure, language, topic, and figure classification on each page in DocTag format. + +Docmatix. To preserve SmolVLM's original DocVQA capabilities, we adapted the same weak annotation strategy used for DocLayNet-PT and applied it to the 1.3M documents inside Docmatix dataset [38]. In addition to the 9.5 million DocVQA questions, we introduced another instruction requiring the full conversion of multi-page documents to DocTags. + +# 4.2. Task-specific Datasets + +Layout. To optimize prediction quality for document layout and table structure, we sampled 76K pages from DocLayNet-PT. Each page underwent human annotation and rigorous quality review, resulting in a high-quality ground-truth dataset. Of these, 60K pages were utilized for fine-tuning, constituting what we refer to as DocLayNet v2. Additionally, we extracted 63K pages from WordScape, specifically filtering for pages containing both text and tables, and utilized their inherently structure-preserving for + +mat as a reliable source of ground truth. To further enhance the model's adaptability to diverse layouts, colors, and fonts, we synthetically generated 250K pages using content from Wikipedia, creating the dataset known as SynthDocNet. This synthetic dataset enables the controlled generation of various document types with precise annotations. All datasets share a unified annotation format and class definitions, ensuring consistency and comprehensive document representation. + +Tables. For the table structure recognition task we trained our model on public datasets such as PubTables-1M [69], FinTabNet [92], WikiTableSet [51] and tabular information extracted from WordScape documents. We transformed table structure information from the original ground-truth to OTSL format [52]. Then, we interleaved each cell tag with ground-truth text. This yields a condensed sequence which represents the structure of the table (including merged cells, and header declaration) along with its textual content. + +Charts. For chart reconstruction, several datasets are publicly available [32, 33, 54, 82] that consist of both synthetic and web-scraped charts. However, these datasets are limited by either the quantity of real-world data or the visual diversity of synthetic examples. This has prompted recent studies [13, 26, 47, 55, 57, 82, 83] to crawl or render additional datasets, which unfortunately are rarely shared publicly. To address these gaps, we generated our own chart dataset comprising four types of charts: line, pie, bar, and stacked bar. Using data from 90,000 tables sourced from the FinTabNet dataset [92], we generated a comprehensive dataset containing 5,000 line charts, 380,000 pie charts, 380,000 bar charts, and 77,000 stacked bar charts. Each data point was rendered using three different visualization libraries (details provided in the Supplementary Material), yielding a total of 2.5 million visually diverse charts. + +Code. Technical books and scientific documents frequently include code snippets, making it essential for document-processing models to accurately interpret code structure, particularly indentation, which carries semantic significance in certain programming languages. Several publicly available datasets provide extensive collections of code snippets with permissive licenses in text format [36, 66], typically aggregated from sources like GitHub. However, a common limitation is that these datasets contain only text-based code snippets, lacking any image-based representation. To address this gap, we leveraged LaTeX and Pygments [15] to generate visually diverse renderings of code (details provided in the Supplementary Material). Our dataset includes 9.3M million rendered code snippets generated at 120 dpi, covering 56 programming languages. + +Equations. Accurately parsing mathematical formulas is crucial for document-processing models, as formulas often contain nuanced structures and semantics criti- + +cal for scientific understanding. To effectively train our model for formula transcription, we combined publicly available datasets with a custom-rendered dataset. The public datasets [11, 17, 22, 23] collectively comprise approximately $730\mathrm{k}$ unique formulas. In addition, we extracted 4.7 million formulas from arXiv by applying a regular expression to the source LaTeX code. The extracted formulas underwent a rigorous normalization procedure to ensure the model was trained only on correct and standardized LaTeX code. Subsequently, these normalized equations were visually rendered using LaTeX (details provided in the Supplementary Material). Overall, our dataset includes 5.5 million unique formulas, each rendered at a resolution of 120 dpi. + +
Dataset NameSizeType
TheCauldron [39]2.63M
DocLayNet-PT-Instruct1.62MMulti Task
DocMatix w/ Doc Conversion* [38]952K
SynthChartNet*2.5M
PlotQA [58]225K
FigureQA [32]100KChart
Unichart [55]600K
ChartQA [54]28K
DocLayNet-PT1.40M
Doc4MLlaVa - RVLCDIP742K
SynthDocNet375KConversion
DocLayNet [65]81K
DocLayNet v276K
WordScape-PT63K
PubMed1M [69]1M
WordScape Tables [78]207KTable
WikiTableSet (EN) [51]204K
FinTabNet [92]90K
SynthFormulaNet*5.5MFormula
SynthCodeNet*9.3MCode
+ +Table 1. Datasets. Overview of the datasets used for the training and evaluation of SmolDocling. Datasets we contribute as part of this publication are denoted with $(^{*})$ . + +# 4.3. Document Instruction Tuning datasets + +To reinforce the recognition of different page elements and introduce document-related features and no-code pipelines, we leveraged rule-based techniques and the Granite-3.1-2b-instruct [25] LLM. Using samples from DocLayNet-PT pages, we generated one instruction by randomly sampling layout elements from a page. These instructions included tasks such as "Perform OCR atbbox," "Identify page element type atbbox," and "Extract all section headers from the page." In addition, we train with Cauldron [39] to avoid catastrophic forgetting with the amount of conversation datasets introduced. + +# 5. Experiments + +In this section, we assess the performance of SmolDocling on document understanding tasks, and compare the results to previously published models and baselines. Capturing the qualities and defects of SmolDocling and other models in comparable numbers is a non-trivial effort for reasons we outline further below. Therefore, we report results on standard evaluation tasks and metrics independently (see section 5.2), but also provide a qualitative assessment of output characteristics (see section 5.3). + +# 5.1. Implementation Details + +Our training was conducted using 64 NVIDIA A100 80GB GPUs, with each epoch requiring 28 hours. The model was trained for a total of 4 epochs using the AdamW optimizer with a learning rate of $2 \times 10^{-4}$ and $2 \times 10^{-6}$ for the vision encoder once unfrozen. We also employed gradient clipping to 1.0 for stability and a warmup ratio of 0.03. For inference, using VLLM [37] on an A100 GPU, SmolDocling achieves a page conversion time of 0.35 seconds while occupying only 0.489 GB of VRAM. The model supports a maximum sequence length of 8,192 tokens and is trained to process up to three pages at a time. + +# 5.2.Quantitative Results + +To enable valid cross-model comparisons with SmolDooling, we addressed multiple dataset and harmonization challenges: + +- Public evaluation datasets lack comprehensive multi-task annotations with diverse layouts. We augmented DocLayNet [65] with text content and reading-order to enable evaluation of both layout analysis and text recognition. Table structure, code listing, formula and chart recognition are evaluated on specific datasets. +- Different models each follow their own conventions for output markup (e.g., Markdown, HTML, DocTags) and produce annotations which do not fully align in semantics. For example, SmolDocling is trained to produce 14 distinct layout-class tags via DocTags, while Qwen2.5-VL combines standard HTML tags and a few explicit class attributes with only partially compatible definitions. GOT and Nougat output Markdown, which can express paragraphs, lists and tables without location annotation. +- Input resolution of page images can affect model performance, but is rarely specified in literature. We standardized at 144 DPI (dots per inch), which reproduces enough detail to be perfectly legible by humans while keeping resource demand for inference reasonable. + +Text Recognition + +
MethodSizeEdit Distance ↓F1-score ↑Precision ↑Recall ↑BLEU ↑METEOR ↑
Full-page +Docling† [7] v2.25-0.400.910.930.920.720.80
Full-page +Qwen2.5 VL [9]7B0.560.720.800.700.460.57
GOT [79]580M0.610.690.710.730.480.59
Nougat (base) [12]350M0.620.660.720.670.440.54
SmolDocling (Ours)256M0.480.800.890.790.580.67
Code listings +SmolDocling (Ours)256M0.110.920.940.910.870.89
Equations +Qwen2.5 VL [9]7B0.220.890.910.870.680.77
GOT [79]580M0.110.950.950.960.850.91
Nougat (base) [12]350M0.620.600.600.530.330.41
SmolDocling (Ours)256M0.110.950.960.950.830.89
+ +Values for Dockling provide a reference for the performance of an ensemble system. + +non-textual elements. For different tasks, we conduct evaluations on diverse content types: full-page documents (from DocLayNet test set) in Markdown format, code snippets (from our SynthCodeNet dataset) in plaintext, and equations (from Im2Latex-230k [21]) in LaTeX format. For Qwen2.5-VL and SmolDocling, we convert HTML or DocTags output to Markdown while preserving formatting where possible. + +Tables are omitted from the full-page output since they do not compare well through Markdown, due to loss of row and column spans. Code snippets must be reproduced with intact line breaks and indentation for accurate scoring. Formulas are evaluated separately in LaTeX format, which is natively produced by all the models tested in this work. The LaTeX outputs from all models, including the ground truth, undergo the same normalization procedure used for generating the SynthFormulaNet dataset. This ensures consistency and fairness in model comparisons. + +SmolDocling significantly outperforms GOT, Nougat and also Qwen2.5-VL, which is 27 times larger, on every metric in full-page transcription (see Table 2. For the code parsing task, we report results exclusively for SmolDocling, as we introduce this task for the first time, and the other models were not explicitly trained for it. SmolDocling also demonstrates superior performance compared to Nougat and Qwen2.5-VL in the formula recognition task, closely matching GOT's results. + +Layout. We evaluate element localization accuracy using the DocLayNet test set (Table 3), measuring the ability to reconstruct document representations that support con + +tent grounding and visual feature extraction. SmolDocling is compared against Qwen2.5-VL-7b [9], currently the only other VLM addressing this task. For a fair comparison, we map each model's label outputs and the ground-truth to six compatible classes: Text, Section Heading, List Item, Table, Picture, and Formula. Overall, SmolDocling outperforms Qwen2.5-VL-7b by a significant margin, however both models score far below a human baseline. This can be partially explained through inherent difficulty of the dataset, as indicated by relatively low human agreement scores. The scores are primarily impacted by moderate recall of element bounding boxes and different degrees of label confusion in both SmolDocling and Qwen2.5-VL. A majority of samples show accurate element bounding boxes as seen in the supplementary material. + +Table 2. Structured document text recognition. We evaluate OCR performance and document formatting on DocLayNet, focusing on textual elements, excluding tables. Text accuracy is measured through multiple metrics (Edit Distance, F1-Score, Precision, Recall, BLEU and METEOR) as proposed in [12]. + +
Docling [7] v2.25SmolDoclingQwen2.5-VLHuman
Formula0.7370.2670.0590.84
Table0.6670.2660.2620.79
List Item0.5280.2960.0900.87
Section Header0.4710.2890.1290.83
Text0.4440.2040.1800.85
Picture0.3280.0660.0780.70
Overall0.5290.2310.1330.82
+ +Table 3. Layout analysis comparison. Evaluation of layout analysis task on DocLayNet with a subset of 6 class labels in mAP[0.5:0.95] metric. Additional reference points are provided for an estimate of human performance [65]. + +Table Structure Recognition. We evaluate the accuracy of recovering table structures, i.e. shaping columns and rows with the correct text content and spans, through + +the TEDS metric [93], once including table text content and once on structure only with omitted text (see Table 4). For isolated assessment, we provide cropped table images from FinTabNet [92] and PubTables-1M [69] test sets, which have poor image quality (72 dpi, compression artifacts). SmolDocling performs competitively against significantly larger models, despite challenges with text transcription from the low-resolution image crops, a limitation likely attributable to insufficient training at such resolutions. Scores for structure only TEDS evaluation are therefore much stronger. + +
MethodModel SizeFinTabNetPubTables-1M
SmolVLM [53]2.2B0.180.32
Granite Vision [72]2.98B0.540.7
TableFormer† [52]52M0.89 f0.84 (0.86)
Qwen2.5-VL 7B7B0.66- (0.786)-
SmolDocling (Ours)256M0.72 (0.81)0.83 (0.88)
SmolDocling (Ours) @150DPI256M0.76 (0.82)-
+ +Chart Extraction. In this task, we evaluate the ability to convert cropped sections of charts into datapoint tables on the ChartQA [54] dataset. All table output is transformed into HTML and compared against the ground-truth HTML using TEDS score, as reported in Table 5. We attribute SmolDocling's performance on charts to the variability in ground-truth styles and the differing interpretations of representations across datasets, which we observe leads to inconsistencies in the model's predictions. Despite its significantly smaller size, SmolDocling achieves a competitive TEDs score. + +Overall, SmolDocling achieves leading performance across a wide range of document conversion tasks and outperforms significantly larger models in text recognition, layout analysis and structure extraction. Furthermore, the model's high performance in downstream tasks, such as code listing and formula recognition, demonstrates its versatility and real-world applicability. Notably, we establish benchmark results for the code listing recognition, setting a standard for future evaluations. + +# 5.3. Qualitative Results + +Several characteristics of SmolDocling output are not evident from task-based metrics alone. In this section, we highlight different output characteristics observed in practice. More illustrated samples can be found in the supplementary material. + +As presented in Figure 2, SmolDocling outputs a sequence of DocTags to encode content, structure and layout. + +Table 4. Table structure and cell content reconstruction. Numbers are given in TEDS metric including text content, with structure-only score in brackets where available. We produced results for SmolDocling and TableFormer, other results are reflected here as published. †TableFormer extracts the content through external OCR. + +
ModelSizeTEDs↑
Phi-3.5-vision4B0.40
Granite Vision [72]3B0.95
SmolVLM [53]2.2B0.02
Molmo-E1B0.54
SmolDocling (Ours)256M0.83
+ +Table 5. Chart analysis. We compare SmolDocling to other small-sized VLMs on the chart-to-table task using the TEDS metric on the ChartQA dataset. The results in this table are based on the work reported in [25]. + +A few notable traits beyond straight content reproduction include: 1) Labeling of page headers and footers, which allows ignoring these repetitive elements in downstream applications. 2) Linking of information, such as captions to tables or pictures, or list elements to parent list through nesting of tags. 3) Preserved line breaks and indentation for code snippets, while line wraps in paragraphs are correctly removed. + +Among typical defects found in the output of SmolDocling are missing tags (e.g. absent location tags), malformed structure (such as missing closing tags), and endless repetition of tokens from some point in a page. Both of these phenomena are commonly seen in auto-regressive models and may prevent parsing or rendering the output correctly (see supplementary material). Despite these limitations, SmolDocling demonstrates unique robustness advantages compared to ensemble systems like Docling, since conversion output is inferred in a single pass. In the latter, errors may accumulate throughout the ensemble model pipeline. For example, a mislocated table detection will lead to missing or distorted output of table structure and content, since this depends on correct table localization. In contrast, SmolDocling can reproduce the tables correctly, even with incorrectly determined location. + +# 6. Conclusion + +In this work, we introduce SmolDocling, an efficient and compact VLM optimized for document conversion with rich output representation. We also present a suite of new datasets with a unified format for document conversion, including the novel task of code listing transcription. We identify page element localization as a critical area requiring further refinement, where targeted techniques will significantly enhance performance in future iterations. Our results conclusively demonstrate that smaller models with unified, optimized output formats such as DocTags can effectively compete with significantly larger models, establishing a clear pathway for resource-efficient multi-task document understanding models. + +Acknowledgements: We thank our collaborators at Hugging Face, Andrés Marafioti and Miquel Farré, who kindly and rapidly developed SmolVLM-256M-Instruct in response to our collaboration, providing the foundational model upon which this work is built. + +# References + +[1] Grobid. https://github.com/kermitt2/grobid, 2008-2025. 2, 3 +[2] Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martin Blázquez, Guilherme Penedo, Lewis Tunstall, Andres Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werra, and Thomas Wolf. SmolLM2: When Smol Goes Big - Data-Centric Training of a Small Language Model, 2025. 2, 3 +[3] Amazon Web Services. Amazon Texttract. https://aws.amazon.com/texttract/, 2024. Accessed: 2024-11-11. 2 +[4] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2 +[5] Anthropic. Claude-3.5. https://www.anthropic.com/news/claude-3-5-sonnet, 2024. Accessed: 2024-02-11. 2 +[6] Christoph Auer, Michele Dolfi, André Carvalho, Cesar BerroSpi Ramis, and Peter WJ Staar. Delivering Document Conversion as a Cloud Service with High Throughput and Responsiveness. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD), pages 363-373. IEEE, 2022. 2 +[7] Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Nikolaos Livathinos, Panos Vagenas, Cesar Berrospi Ramis, Matteo Omenetti, Fabian Lindlbauer, Kasper Dinkla, et al. Docling Technical Report. arXiv preprint arXiv:2408.09869, 2024. 5, 7 +[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3 +[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL Technical Report, 2025. 2, 3, 7 +[10] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. In European Conference on Computer Vision, pages 241-252. Springer, 2022. 5 +[11] Lukas Blecher. LaTeX-OCR, 2024. Accessed: 2024-11-09. 6 + +[12] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. 3, 6, 7 +[13] Jinyue Chen, Lingyu Kong, Haoran Wei, Chenglong Liu, Zheng Ge, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. OneChart: Purify the Chart Structural Extraction via One Auxiliary Token. arXiv preprint arXiv:2404.09987, 2024. 3, 5 +[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2 +[15] Pygments Contributors. Pygments: Python syntax highlighter, 2024. Accessed: 2024-11-06. 5 +[16] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 3 +[17] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning, pages 980-989. PMLR, 2017. 3, 6 +[18] Mohamed Dhouib, Ghassen Bettaieb, and Aymen Shabou. Docparser: End-to-endOCR-free information extraction from visually rich documents. In Document Analysis and Recognition - ICDAR 2023, pages 155-172, Cham, 2023. Springer Nature Switzerland. 3 +[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model. arXiv preprint arXiv:2401.16420, 2024. 2 +[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Zhe Chen, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Kai Chen, Conghui He, Xingcheng Zhang, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD. arXiv preprint arXiv:2404.06512, 2024. 2 +[21] Gregory Eritsyan. Im2LaTeX 230K. https://www.kaggle.com/datasets/gregoryeritsyan/im2latex-230k, 2023. Accessed: 2025-03-06. 7 +[22] Hugging Face. latex-formulas Dataset, 2024. Accessed: 2024-11-09. 6 +[23] Philippe Gervais, Asya Fadeeva, and Andrii Maksai. Math-writing: A dataset for handwritten mathematical expression recognition. URL https://arxiv.org/abs/2404.10690, 2024.3, 6 + +[24] Google Cloud. Document AI. https://cloud.google.com/document-ai, 2024. Accessed: 2024-11-11. 2 +[25] IBM Granite Team. Granite 3.0 Language Models, 2024. 2, 5, 6, 8 +[26] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 3, 5 +[27] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3 +[28] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 3 +[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 3 +[30] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), pages 1–6. IEEE, 2019. 3 +[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 2 +[32] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 3, 5, 6 +[33] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 3, 5 +[34] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 3 +[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3 +[36] Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Munoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The Stack: 3 TB of permissively licensed source code. Preprint, 2022. 5 + +[37] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6 +[38] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 2, 3, 5, 6 +[39] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models?, 2024. 3, 6 +[40] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 3 +[41] David Lewis, Gady Agam, Shlomo Argamon, Ophir Frieder, David Grossman, and Jefferson Heard. Building a test collection for complex document information processing. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 665-666, 2006. 5 +[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3 +[43] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2 +[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[45] Han Lin, Peng Yang, and Fanlong Zhang. Review of scene text detection and recognition. Archives of computational methods in engineering, 27(2):433-454, 2020. 3 +[46] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 3 +[47] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 3, 5 +[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2023. 2, 3 +[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2023. 2 + +[50] Nikolaos Livathinos, Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Panos Vagenas, Cesar Berropi Ramis, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Shubham Gupta, Rafael Teixeira de Lima, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, and Peter W. J. Staar. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion, 2025. 2, 3, 5 +[51] Nam Ly., Atsuhiro Takasu., Phuc Nguyen., and Hideaki Takeda. Rethinking Image-Based Table Recognition Using Weakly Supervised Methods. pages 872-880, 2023. 5, 6 +[52] Maksym Lysak, Ahmed Nassar, Nikolaos Livathinos, Christoph Auer, and Peter Staar. Optimized Table Tokenization for Table Structure Recognition. In Document Analysis and Recognition - ICDAR 2023, pages 37-50, Cham, 2023. Springer Nature Switzerland. 2, 3, 4, 5, 8 +[53] Andres Marafioti, Orr Zohar, Miquel Farre, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Alal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. SmolVLM: Redefining small and efficient multimodal models. 2025. 1, 2, 3, 8 +[54] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3, 5, 6, 8 +[55] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 3, 5, 6 +[56] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 3 +[57] Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. Chartassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. arXiv preprint arXiv:2401.02384, 2024. 3, 5 +[58] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. PlotQA: Reasoning over Scientific Plots. In The IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 6 +[59] Microsoft Azure. AI Document Intelligence. https://azure.microsoft.com/en-us/products/ai-services.ai-document-intelligence, 2024. Accessed: 2024-11-11. 2 +[60] OpenAI. Hello GPT-4O. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-02-09, 2024-02-11, 2024-02-12. 2 +[61] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 2 +[62] Vik Paruchuri. Marker: Convert PDF to Markdown Quickly with High Accuracy. https://github.com/VikParuchuri/marker, 2024. 2, 3 +[63] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2024. Accessed: 2024-11-11. 3 + +[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3 +[65] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 3743–3751, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6, 7 +[66] Ruchir Puri, David Kung, Geert Janssen, Wei Zhang, Giacomo Domeniconi, Vladimir Zolotov, Julian Dolby, Jie Chen, Mihir Choudhury, Lindsey Decker, Veronika Thost, Luca Buratti, Saurabh Pujar, Shyam Ramji, Ulrich Finkler, Susan Malaika, and Frederick Reiss. CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Tasks, 2021. 5 +[67] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2 +[68] Felix M Schmitt-Koopmann, Elaine M Huang, Hans-Peter Hutter, Thilo Stadelmann, and Alireza Darvishy. MathNet: A Data-Centric Approach for Printed Mathematical Expression Recognition. IEEE Access, 2024. 3 +[69] Brandon Smock, Rohith Pesala, and Robin Abraham. PubTables-1M: Towards comprehensive table extraction from unstructured documents. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4624–4632, 2022. 3, 5, 6, 8 +[70] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 3 +[71] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 +[72] Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kondic, Nimrod Shabtay, Pengyuan Li, Roei Herzig, Shafiq Abedin, Shaked Perek, Sivan Harary, Udi Barzelay, Adi Raz Goldfarb, Aude Oliva, Ben Wieles, Bishwaranjan Bhattacharjee, Brandon Huang, Christoph Auer, Dan GUTFreund, David Beymer, David Wood, Hilde Kuehne, Jacob Hansen, Joseph Shtok, Ken Wong, Luis Angel Bathen, Mayank Mishra, Maksym Lysak, Michele Dolfi, Mikhail Yurochkin, Nikolaos Livathinos, Nimrod Harel, Ophir Azulai, Oshri Naparstek, Rafael Teixeira de Lima, Rameswar Panda, Sivan Doveh, Shubham Gupta, Subhro Das, Syed Zawad, Yusik Kim, Zexue He, Alexander Brooks, Gabe Goodhart, Anita Govindjee, Derek Leist, Ibrahim Ibrahim, Aya Soffer, David Cox, + +Kate Soule, Luis Lastras, Nirmit Desai, Shila Ofek-koifman, Sriram Raghavan, Tanveer Syeda-Mahmood, Peter Staar, Tal Drory, and Rogerio Feris. Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence, 2025. 8 +[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[74] Michal Turski, Tomasz Stanisławek, Karol Kaczmarek, Paweł Dyda, and Filip Graliński. CCpdf: Building a High Quality Corpus for Visually Rich Documents from Web Crawl Data. In Document Analysis and Recognition - IC-DAR 2023, pages 348-365, Cham, 2023. Springer Nature Switzerland. 5 +[75] Unstructured.io Team. Unstructured.io: Open-Source Pre-Processing Tools for Unstructured Data. https://unstructured.io, 2024. Accessed: 2024-11-19. 2 +[76] Vicuna. Vicuna: An open-source chatbot impressing GPT-4 with $90\%$ ChatGPT quality. https://vicuna.lmsys.org/, 2023.2 +[77] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, Bo Zhang, Liquun Wei, Zhihao Sui, Wei Li, Botian Shi, Yu Qiao, Dahua Lin, and Conghui He. MinerU: An Open-Source Solution for Precise Document Content Extraction, 2024. 2, 3 +[78] Maurice Weber, Carlo Siebenschuh, Rory Butler, Anton Alexandrov, Valdemar Thanner, Georgios Tsolakis, Haris Jabbar, Ian Foster, Bo Li, Rick Stevens, et al. WordScape: a Pipeline to extract multilingual, visually rich Documents with Layout Annotations from Web Crawl Data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6 +[79] Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model. arXiv preprint arXiv:2409.01704, 2024. 3, 7 +[80] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2 +[81] Wes McKinney and the Pandas Development Team. Pandas: powerful Python data analysis toolkit, Release 1.4.4. https://pandas.pydata.org/pandas-docs/ version/1.4/pandas.pdf, 2022.4 +[82] Renqiu Xia, Bo Zhang, Haoyang Peng, Hancheng Ye, Xiangchao Yan, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. StructChart: Perception, Structuring, Reasoning for Visual Chart Understanding. arXiv preprint arXiv:2309.11268, 2023. 3, 5 +[83] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024. 3, 5 + +[84] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 3 +[85] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020.3 +[86] Zuoyu Yan, Xiaode Zhang, Liangcai Gao, Ke Yuan, and Zhi Tang. ConvMath: a convolutional sequence network for mathematical expression recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4566-4572. IEEE, 2021. 3 +[87] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3 +[88] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition. arXiv preprint arXiv:2309.15112, 2023. 2 +[89] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer-2.5: A Versatile Large Vision Language Model Supporting Long-Contextual Input and Output. arXiv preprint arXiv:2407.03320, 2024. 2 +[90] Qiming Zhang, Jing Zhang, Yufei Xu, and Dacheng Tao. Vision Transformer with Quadrangle Attention. arXiv preprint arXiv:2303.15105, 2023. 2 +[91] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pretrained transformer language models. URL https://arxiv.org/abs/2205.01068, 3:19-0, 2023. 2 +[92] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 697-706, 2021. 3, 5, 6, 8 +[93] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. arXiv preprint arXiv:1911.10683, 2019. 8 +[94] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 \ No newline at end of file diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/images.zip b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1dc356307ad368a46078c4224a51816dcdcdd615 --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ab11e53c7577bb3be465dcf08b657d1c763fa7ea4c514bc08ee612983604f9e +size 431323 diff --git a/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/layout.json b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a203d9097b4dbfc8ee35219ba6f36f668ac95165 --- /dev/null +++ b/2025/SmolDocling_ An ultra-compact vision-language model for end-to-end multi-modal document conversion/layout.json @@ -0,0 +1,7285 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 138 + ], + "type": "text", + "content": "SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 161, + 563, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 563, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 563, + 217 + ], + "type": "text", + "content": "Ahmed Nassar Matteo Omenetti Maksym Lysak Nikolaos Livathinos Christoph Auer Lucas Morin \nRafael Teixeira de Lima Yusik Kim A. Said Gurbuz Michele Dolfi Peter W. J. Staar \n{ahn, mly, nli, cau, dol, taa}@zurich.ibm.com, {Matteo.Omenetti1}@ibm.com \nIBM Research" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 271, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 297, + 559 + ], + "type": "text", + "content": "We introduce SmolDocling, an ultra-compact vision-language model targeting end-to-end document conversion. Our model comprehensively processes entire pages by generating DocTags, a new universal markup format that captures all page elements in their full context with location. Unlike existing approaches that rely on large foundational models, or ensemble solutions that rely on handcrafted pipelines of multiple specialized models, SmolDocling offers an end-to-end conversion for accurately capturing content, structure and spatial location of document elements in a 256M parameters vision-language model. SmolDocling exhibits robust performance in correctly reproducing document features such as code listings, tables, equations, charts, lists, and more across a diverse range of document types including business documents, academic papers, technical reports, patents, and forms — significantly extending beyond the commonly observed focus on scientific papers. Additionally, we contribute novel publicly sourced datasets for charts, tables, equations, and code recognition. Experimental results demonstrate that SmolDocling competes with other Vision Language Models that are up to 27 times larger in size, while reducing computational requirements substantially. The model weights and datasets are available at: https://huggingface.co/ds4sd" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "For decades, converting complex digital documents into a structured, machine-processable format has been a significant technical challenge. This challenge primarily stems from the substantial variability in document layouts and styles, as well as the inherently opaque nature of the widely used PDF format, which is optimized for printing rather than semantic parsing. Intricate layout styles and visually challenging elements such as forms, tables, and complex charts can significantly impact the reading order and gen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 365 + ], + "type": "text", + "content": "eral understanding of documents. These problems have driven extensive research and development across multiple domains of computer science. On one hand, sophisticated ensemble systems emerged, which decompose the conversion problem into several sub-tasks (e.g. OCR, layout analysis, table structure recognition, classification) and tackle each sub-task independently. Although such systems can achieve high-quality results for many document types while maintaining relatively low computational demands, they are often difficult to tune and generalize." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 556 + ], + "type": "text", + "content": "On the other hand, in the recent past, great interest has developed around large foundational multimodal models that can solve the whole conversion task in one shot, while simultaneously offering flexible querying and parametrization through prompts. This approach has been made possible by the advent of multimodal pre-training of large vision-language models (LVLMs), which opened up a vast array of opportunities for leveraging diverse data sources, including PDF documents. However, the literature on this topic highlights a significant gap in the availability of high-quality and open-access datasets suitable for training robust multimodal models for the task of document understanding. Furthermore, relying on LVLMs may introduce common issues associated with such models, including hallucinations and the use of significant computational resources, making them impractical from both a quality and cost perspective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "type": "text", + "content": "In this work, we outline how we close the gaps left by publicly available datasets and establish a training approach to achieve end-to-end, full-featured document conversion through a vision-language model, which we name SmolDoccling. Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "- SmolDocling: An ultra-compact VLM for end-to-end document conversion, based on Hugging Face's most recent SmolVLM-256M [53], which is between 5 and 10 times smaller in parameters than comparable VLMs tuned on document understanding tasks. Hence, we reduce the computational complexity to roughly the same order of magnitude as typical representatives of ensemble meth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21972" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 68, + 509, + 210 + ], + "blocks": [ + { + "bbox": [ + 104, + 68, + 509, + 210 + ], + "lines": [ + { + "bbox": [ + 104, + 68, + 509, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 68, + 509, + 210 + ], + "type": "image", + "image_path": "0211634d0b93a26372b178c2ce62ccc68ddba7148f16e50aafb340745dbdf4fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 275 + ], + "type": "text", + "content": "Figure 1. SmolDocling/SmolVLM architecture. SmolDocling converts images of document pages to DocTags sequences. First, input images are encoded using a vision encoder and reshaped via projection and pooling. Then, the projected embeddings are concatenated with the text embeddings of the user prompt, possibly with interleaving. Finally, the sequence is used by an LLM to autoregressively predict the DocTags sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 296, + 295, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 296, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 63, + 296, + 295, + 331 + ], + "type": "text", + "content": "ods. Uniquely, SmolDocling learns and produces a unified document representation that captures content, structure and spatial location of all elements in a page." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 332, + 295, + 510 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 332, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 295, + 438 + ], + "type": "text", + "content": "- We augment existing document pre-training datasets with additional feature annotations, and create new datasets for tasks with insufficient coverage in open multi-modal training datasets, including code listings, equations, and chart understanding. Importantly, we construct full-page ground truth data that combines key annotation features such as layout, table structure, code listings, charts, and formulas in their natural context, and derive instruction datasets from all training datasets used." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 439, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 475 + ], + "type": "text", + "content": "- We introduce DocTags, a document markup format optimized to efficiently represent the full content and layout characteristics of a document, inspired by [52]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": "- We evaluate the achieved prediction quality of SmolDoccling on several tasks, including comparisons to popular community models." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 142, + 533 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 271, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 271, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 271, + 553 + ], + "type": "text", + "content": "2.1. Large Vision-Language Models (LVLMs)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "With the advent of large language models (LLMs) [2, 4, 8, 25, 31, 61, 67, 73, 80] characterized by their powerful contextual understanding and adaptability, several approaches have emerged to extend them into large vision-language models. Some proprietary, closed-source LLMs have been adapted into multimodal large language models (MLLMs), such as GPT-4o [60], Gemini [71], and Claude 3.5 [5], demonstrating exceptional capabilities across various modalities, including vision. Open-source methods are actively advancing to match the capabilities of proprietary MLLMs. BLIP-2 [44] is one of the earliest works combining a vision encoder with a frozen LLM (OPT [91] or FlanT5 [14]) using a lightweight transformer (Q-former" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 498 + ], + "type": "text", + "content": "[90]). Building upon BLIP-2, MiniGPT-4 [94] integrated a frozen visual encoder with a frozen LLM (Vicuna [76]) using a Q-Former network and a single projection layer. LLaVA [48, 49] employs a similar architecture but uses a minimal adapter layer. LLaVA-OneVision [42] and LLaVA-NeXT-Interleave [43] build upon LLaVA to support multi-image, higher resolution, and video understanding. For handling high resolutions and text-image compositions and comprehension, InternLM-XComposer introduces a family of models [19, 20, 88, 89] designed specifically for these tasks. Qwen-VL [8] introduces a position-aware adapter to address efficiency issues caused by the vision encoder generating long image feature sequences. Qwen2.5-VL [9] employs windowed attention with 2D rotary positional embeddings to process native-resolution inputs efficiently and integrates an vision-language merger for dynamic feature compression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 608 + ], + "type": "text", + "content": "In this work, we build upon the SmolVLM architecture approach [53]. Precisely, we adopt SmolVLM-256M, a compact variant of base SmolVLM-2.2B model, inspired by the architecture of Idefics3 [38] which employs the shape-optimized SigLIP encoder as its visual backbone. It applies an aggressive pixel shuffle strategy to compress visual features, reduces the number of image hidden states and introduces new special tokens to separate sub-images and improve tokenization efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 545, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 545, + 625 + ], + "type": "text", + "content": "2.2. State of the Art in Document Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Several powerful document understanding solutions have emerged on the market in the past decade as commercial cloud offerings on hyperscalers [3, 6, 24, 59], frontier models such as GPT-4o [60] or Claude [5], or as open-source libraries such as Docling [50], Grobid [1], Marker [62], MinerU [77] or Unstructured [75]. Typical tasks that document understanding encompasses are document classifi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21973" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "cation [29, 84], OCR [45], layout analysis [65, 85], table recognition [52, 69, 92], key-value extraction [29, 30, 34], chart understanding [13, 26, 32, 33, 46, 47, 54, 55, 57, 82, 83], equation recognition [17, 23, 63, 68, 86], and more." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 122, + 294, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 122, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 54, + 122, + 294, + 231 + ], + "type": "text", + "content": "Ensemble systems, such as the aforementioned open-source libraries [1, 50, 62, 77], implement a pipeline in source code, which conditionally applies specialized, single-task models on a given input document, composing the prediction results into a meaningful document representation. Typically, each task entails human-crafted handling of pre-processing and post-processing logic (i.e. setting confidence thresholds for layout detections, matching layout elements to text cells, conditionally applying models)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 472 + ], + "type": "text", + "content": "Conversely, multi-task models aim at providing a single model capable of handling various document-understanding related tasks simultaneously, benefiting from shared context and shared representations learned across these different tasks. OCR-reliant methods, such as LayoutLM [29, 84, 85] and UDOP [70], use text extracted from an external OCR engine, along with image and text bounding box locations, as input. In contrast, OCR-free methods such as Donut [34], Dessurt [16], DocParser [18], Pix2Struct [40] are transformer-based models trained end-to-end to take an image as input and directly output text. Large Vision Language Models (LVLMs) such as LLaVA [48], LLaVA-OneVision [42], UReader [87], Kosmos-2 [64], and Qwen-VL [8] all leverage various vision encoders, projection adapters, and LLMs. While they may use different image patching techniques and instruction-tuning datasets to handle a range of vision tasks, they commonly evaluate their document understanding capabilities on datasets like DocVQA [56] and mPLUG-DocOwl 1.5 datasets [27], primarily focusing on question answering and reasoning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "type": "text", + "content": "Our work focuses primarily on document understanding tasks related to conversion and structural recognition, aiming to represent documents with high accuracy and completeness. The closest works to our approach are Nougat [12], DocOwl 2 [28], GOT [79], and Qwen2.5-VL [9]. DocOwl 2 and GOT are dedicated to precise conversion and recognition of document structures. On one hand, DocOwl 2 employs a dynamic shape-adaptive cropping module to process high-resolution images, feeding them into a ViT vision encoder. A dedicated module compresses the processed visual features before feeding them into a LLaMa-based LLM. On the other hand, GOT focuses on converting diverse elements—such as text, formulas, molecular diagrams, tables, charts, sheet music, and geometric shapes—into structured formats. It achieves a lightweight architecture by utilizing a SAM vision encoder [35], a linear projection layer, and a Qwen-0.5B LLM. Building on these approaches, our work not only extends the range of tasks but also broadens the types of documents handled, moving beyond a strictly scientific focus" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "to support a more diverse array of document formats. In parallel, Qwen2.5-VL introduces the Omni-Parsing strategy, that integrates multiple document elements into a unified HTML-based representation. This format encodes layout bounding box information and reading order for various document elements such as paragraphs and charts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 395, + 171 + ], + "type": "text", + "content": "3.SmolDocling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 178, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 357 + ], + "type": "text", + "content": "Building on Hugging Face's SmolVLM [53], SmolDocling converts document page images into sequences representing their content and structure. It produces a unified page representation (DocTags markup) that is optimized for the LLM backbone and interpretable by Docling [50]. This new DocTags format covers the reproduction of captions, charts, forms, code, equations, tables, footnotes, lists, page footers and headers, section headings, and text. SmolDocling both recognizes the type and location of these elements as well as performs OCR within them. Our approach enables fullpage conversion with reading-order and hierarchical linking between elements (e.g., linking captions with figures and tables), providing bounding boxes for each component, supporting isolated predictions on individual elements (e.g., cropped tables) and grouping for lists and nested lists." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 368, + 429, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 429, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 429, + 380 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 554 + ], + "type": "text", + "content": "The model architecture is illustrated in Figure 1. SmolVLM-256M relies on SigLIP base patch-16/512 (93M) as visual backbone and compared to the 2.2B version of the same model, its training data was rebalanced to emphasize document understanding (41%) and image captioning (14%), combining The Cauldron [39], Docmatix [38] datasets with the addition of MathWriting [23]. It uses lightweight variant of the SmolLM-2 family (135M) [2] as the language backbone and employs a radical pixel shuffle method that compresses each 512x512 image patch into 64 visual tokens. Last but not least, tokenization efficiency was also improved in SmolVLM-256M by increasing the pixel-to-token ratio to 4096 pixels per token and introducing special tokens for sub-image separators." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 377, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 377, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 377, + 576 + ], + "type": "text", + "content": "3.2. DocTags" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "To train VLMs for standardized document conversion, we introduce the DocTags markup format, illustrated in Figure 2. Inspired by OTSL [52], DocTags define a structured vocabulary of unambiguous tags and rules that explicitly separate textual content from document structure. This disentanglement improves performances of Imago-to-Sequence models by minimizing confusions. In contrast, direct conversion to standard formats such as HTML or Markdown is often lossy and ambiguous, lacking explicit structural markup, failing to preserve page layout context, and increasing total tokens count. Pieces" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21974" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 523, + 352 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 523, + 352 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 523, + 352 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 523, + 352 + ], + "type": "image", + "image_path": "697ddb11402df582bbb872662903ea9f82bca7a09917fcb97dc904996e087583.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 359, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 54, + 359, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 555, + 407 + ], + "type": "text", + "content": "Figure 2. DocTags format describes key features of document elements: type of an element (text, picture, table, code, etc.), location on the page, and content. Nested tags convey additional information: pictures and tables may nest captions, table structure is represented in OTSL tags, lists nest list items, code and pictures carry a classification. All DocTags output is taken from SmolDocling predictions, artificial line-breaks and dots (...) were inserted for better readability. A, B, C - snippets of various pages from [52], D - from [81]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 425, + 294, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 294, + 521 + ], + "type": "text", + "content": "of text content that represent basic elements in DocTags are enclosed in XML-style tags. Tags define document block types such as: text, caption, footnote, formula, title, list_item, page Footer, page_header, picture, section_header, otsl, code, document_index, etc. Each element can nest additional location tags encoding its position on the page as a bounding box, represented in DocTags as " + }, + { + "bbox": [ + 54, + 425, + 294, + 521 + ], + "type": "inline_equation", + "content": "<\\mathrm{loc\\_x1}><\\mathrm{loc\\_y1}><\\mathrm{loc\\_x2}><\\mathrm{loc\\_y2}>" + }, + { + "bbox": [ + 54, + 425, + 294, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "Special blocks, such as tables and images, nest additional descriptors for captions, tabular structure, or image classes. To encode table structures, we fully include the OTSL vocabulary into DocTags. OTSL has the advantage of being minimalist and supporting cell spans and header information. Its tokens are interleaved with text, allowing simultaneous encoding of both table structure and content. In order to promote robust visual-semantic alignment in our document understanding pipeline, we maintain a uniform DocTags representation for cropped page elements (e.g., tables, code, equations) that is identical to their full-page counterparts. This consistency ensures that the model can leverage similar visual cues whether an element is presented in isolation or in the context of an entire page, thereby strengthening the learning signal for both computer vision and downstream machine learning tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 449 + ], + "type": "text", + "content": "For a complete list of tags and detailed description of the format, consult the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 458, + 410, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 410, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 410, + 471 + ], + "type": "text", + "content": "3.3. Model Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 476, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 632 + ], + "type": "text", + "content": "We employ a curriculum learning approach to progressively align our model for document conversion, ensuring faster convergence. As an initial step, we incorporate DocTags as tokens in our tokenizer. To align the LLM part we freeze the vision encoder and train only the remaining network to adapt it to the new output format which it hasn't seen before. To ensure comprehensive coverage of all DocTags, we maintain a balanced mix of tasks and data types during training. Next, we unfreeze the vision encoder and train the model on our pretraining datasets (see Sec. 4.1), along with all task-specific conversion datasets, including tables, code, equations, and charts. Finally, we fine-tune using all available datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 645, + 354, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 354, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 354, + 657 + ], + "type": "text", + "content": "4. Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "For document understanding tasks, each page image represents the visual layout of textual elements—such as paragraphs, formulas, tables, code, or figures with embedded textual content. Publicly available datasets that unify all" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21975" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "these characteristics remain limited, and existing corpora often suffer from small sample sizes or fragmented formats. In this section, we explain how we mitigate these issues in both pre-training and task-specific training scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 176, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 176, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 176, + 141 + ], + "type": "text", + "content": "4.1. Pre-training datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 277 + ], + "type": "text", + "content": "In the domain of document understanding, currently available open-access document pre-training datasets include OCR-IDL [10], IIT-CDIP [41], CCpdf [74], and WordScape [78]. OCR-IDL and IIT-CDIP, industry-sourced from the 1990s, provide only page images and OCR text. CCpdf and WordScape, derived from Common Crawl, focus on raw PDFs and Word documents, respectively. While WordScape enriches documents with layout, table structure, and topic information via XML markup, its visual variability is limited. Notably, none of these datasets offer rich annotations for all page elements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 495 + ], + "type": "text", + "content": "DocLayNet-PT. Given the lack of public multimodal document pre-training data with sufficient annotation features, we constructed a new document pre-training dataset named DocLayNet-PT. DocLayNet-PT is a 1.4M pages dataset extracted from the DocFM dataset [25] of unique PDF documents, sourced from CommonCrawl, Wikipedia, and business-related documents. We focused on this dataset to look for documents including visually distinct content such as equations, tables, code and charts with colorful layouts. This dataset was augmented with weak annotations through a series of processing steps. First, we applied PDF parsing to all pages to obtain the text layout including positional information using Docling [7, 50]. Next, we performed an enrichment step to obtain an enhanced, machine-friendly representation of each page. Specifically, we provide annotations for layout elements, table structure, language, topic, and figure classification on each page in DocTag format." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 579 + ], + "type": "text", + "content": "Docmatix. To preserve SmolVLM's original DocVQA capabilities, we adapted the same weak annotation strategy used for DocLayNet-PT and applied it to the 1.3M documents inside Docmatix dataset [38]. In addition to the 9.5 million DocVQA questions, we introduced another instruction requiring the full conversion of multi-page documents to DocTags." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 180, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 180, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 180, + 600 + ], + "type": "text", + "content": "4.2. Task-specific Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Layout. To optimize prediction quality for document layout and table structure, we sampled 76K pages from DocLayNet-PT. Each page underwent human annotation and rigorous quality review, resulting in a high-quality ground-truth dataset. Of these, 60K pages were utilized for fine-tuning, constituting what we refer to as DocLayNet v2. Additionally, we extracted 63K pages from WordScape, specifically filtering for pages containing both text and tables, and utilized their inherently structure-preserving for" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "mat as a reliable source of ground truth. To further enhance the model's adaptability to diverse layouts, colors, and fonts, we synthetically generated 250K pages using content from Wikipedia, creating the dataset known as SynthDocNet. This synthetic dataset enables the controlled generation of various document types with precise annotations. All datasets share a unified annotation format and class definitions, ensuring consistency and comprehensive document representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 300 + ], + "type": "text", + "content": "Tables. For the table structure recognition task we trained our model on public datasets such as PubTables-1M [69], FinTabNet [92], WikiTableSet [51] and tabular information extracted from WordScape documents. We transformed table structure information from the original ground-truth to OTSL format [52]. Then, we interleaved each cell tag with ground-truth text. This yields a condensed sequence which represents the structure of the table (including merged cells, and header declaration) along with its textual content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 495 + ], + "type": "text", + "content": "Charts. For chart reconstruction, several datasets are publicly available [32, 33, 54, 82] that consist of both synthetic and web-scraped charts. However, these datasets are limited by either the quantity of real-world data or the visual diversity of synthetic examples. This has prompted recent studies [13, 26, 47, 55, 57, 82, 83] to crawl or render additional datasets, which unfortunately are rarely shared publicly. To address these gaps, we generated our own chart dataset comprising four types of charts: line, pie, bar, and stacked bar. Using data from 90,000 tables sourced from the FinTabNet dataset [92], we generated a comprehensive dataset containing 5,000 line charts, 380,000 pie charts, 380,000 bar charts, and 77,000 stacked bar charts. Each data point was rendered using three different visualization libraries (details provided in the Supplementary Material), yielding a total of 2.5 million visually diverse charts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 675 + ], + "type": "text", + "content": "Code. Technical books and scientific documents frequently include code snippets, making it essential for document-processing models to accurately interpret code structure, particularly indentation, which carries semantic significance in certain programming languages. Several publicly available datasets provide extensive collections of code snippets with permissive licenses in text format [36, 66], typically aggregated from sources like GitHub. However, a common limitation is that these datasets contain only text-based code snippets, lacking any image-based representation. To address this gap, we leveraged LaTeX and Pygments [15] to generate visually diverse renderings of code (details provided in the Supplementary Material). Our dataset includes 9.3M million rendered code snippets generated at 120 dpi, covering 56 programming languages." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "content": "Equations. Accurately parsing mathematical formulas is crucial for document-processing models, as formulas often contain nuanced structures and semantics criti-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21976" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "cal for scientific understanding. To effectively train our model for formula transcription, we combined publicly available datasets with a custom-rendered dataset. The public datasets [11, 17, 22, 23] collectively comprise approximately " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "730\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " unique formulas. In addition, we extracted 4.7 million formulas from arXiv by applying a regular expression to the source LaTeX code. The extracted formulas underwent a rigorous normalization procedure to ensure the model was trained only on correct and standardized LaTeX code. Subsequently, these normalized equations were visually rendered using LaTeX (details provided in the Supplementary Material). Overall, our dataset includes 5.5 million unique formulas, each rendered at a resolution of 120 dpi." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 236, + 296, + 506 + ], + "blocks": [ + { + "bbox": [ + 58, + 236, + 296, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 236, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 296, + 506 + ], + "type": "table", + "html": "
Dataset NameSizeType
TheCauldron [39]2.63M
DocLayNet-PT-Instruct1.62MMulti Task
DocMatix w/ Doc Conversion* [38]952K
SynthChartNet*2.5M
PlotQA [58]225K
FigureQA [32]100KChart
Unichart [55]600K
ChartQA [54]28K
DocLayNet-PT1.40M
Doc4MLlaVa - RVLCDIP742K
SynthDocNet375KConversion
DocLayNet [65]81K
DocLayNet v276K
WordScape-PT63K
PubMed1M [69]1M
WordScape Tables [78]207KTable
WikiTableSet (EN) [51]204K
FinTabNet [92]90K
SynthFormulaNet*5.5MFormula
SynthCodeNet*9.3MCode
", + "image_path": "24b504bdc5514a0fcbd114496f123dcb2d2d3672d2c3a85065d66a5889121100.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "text", + "content": "Table 1. Datasets. Overview of the datasets used for the training and evaluation of SmolDocling. Datasets we contribute as part of this publication are denoted with " + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 55, + 514, + 295, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 564, + 258, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 258, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 258, + 578 + ], + "type": "text", + "content": "4.3. Document Instruction Tuning datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "To reinforce the recognition of different page elements and introduce document-related features and no-code pipelines, we leveraged rule-based techniques and the Granite-3.1-2b-instruct [25] LLM. Using samples from DocLayNet-PT pages, we generated one instruction by randomly sampling layout elements from a page. These instructions included tasks such as \"Perform OCR atbbox,\" \"Identify page element type atbbox,\" and \"Extract all section headers from the page.\" In addition, we train with Cauldron [39] to avoid catastrophic forgetting with the amount of conversation datasets introduced." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 199 + ], + "type": "text", + "content": "In this section, we assess the performance of SmolDocling on document understanding tasks, and compare the results to previously published models and baselines. Capturing the qualities and defects of SmolDocling and other models in comparable numbers is a non-trivial effort for reasons we outline further below. Therefore, we report results on standard evaluation tasks and metrics independently (see section 5.2), but also provide a qualitative assessment of output characteristics (see section 5.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 447, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 447, + 217 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "content": "Our training was conducted using 64 NVIDIA A100 80GB GPUs, with each epoch requiring 28 hours. The model was trained for a total of 4 epochs using the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 221, + 555, + 354 + ], + "type": "text", + "content": " for the vision encoder once unfrozen. We also employed gradient clipping to 1.0 for stability and a warmup ratio of 0.03. For inference, using VLLM [37] on an A100 GPU, SmolDocling achieves a page conversion time of 0.35 seconds while occupying only 0.489 GB of VRAM. The model supports a maximum sequence length of 8,192 tokens and is trained to process up to three pages at a time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 433, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 433, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 433, + 371 + ], + "type": "text", + "content": "5.2.Quantitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 411 + ], + "type": "text", + "content": "To enable valid cross-model comparisons with SmolDooling, we addressed multiple dataset and harmonization challenges:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 413, + 554, + 651 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 413, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 554, + 484 + ], + "type": "text", + "content": "- Public evaluation datasets lack comprehensive multi-task annotations with diverse layouts. We augmented DocLayNet [65] with text content and reading-order to enable evaluation of both layout analysis and text recognition. Table structure, code listing, formula and chart recognition are evaluated on specific datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 485, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 554, + 592 + ], + "type": "text", + "content": "- Different models each follow their own conventions for output markup (e.g., Markdown, HTML, DocTags) and produce annotations which do not fully align in semantics. For example, SmolDocling is trained to produce 14 distinct layout-class tags via DocTags, while Qwen2.5-VL combines standard HTML tags and a few explicit class attributes with only partially compatible definitions. GOT and Nougat output Markdown, which can express paragraphs, lists and tables without location annotation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 593, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 593, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 593, + 553, + 651 + ], + "type": "text", + "content": "- Input resolution of page images can affect model performance, but is rarely specified in literature. We standardized at 144 DPI (dots per inch), which reproduces enough detail to be perfectly legible by humans while keeping resource demand for inference reasonable." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Text Recognition" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21977" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 274 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 274 + ], + "type": "table", + "html": "
MethodSizeEdit Distance ↓F1-score ↑Precision ↑Recall ↑BLEU ↑METEOR ↑
Full-page \nDocling† [7] v2.25-0.400.910.930.920.720.80
Full-page \nQwen2.5 VL [9]7B0.560.720.800.700.460.57
GOT [79]580M0.610.690.710.730.480.59
Nougat (base) [12]350M0.620.660.720.670.440.54
SmolDocling (Ours)256M0.480.800.890.790.580.67
Code listings \nSmolDocling (Ours)256M0.110.920.940.910.870.89
Equations \nQwen2.5 VL [9]7B0.220.890.910.870.680.77
GOT [79]580M0.110.950.950.960.850.91
Nougat (base) [12]350M0.620.600.600.530.330.41
SmolDocling (Ours)256M0.110.950.960.950.830.89
", + "image_path": "65d7ee92eb96bc0581d7c1c382210464f757d0e25f4e5171b77d64587183a5f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 316, + 362, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 362, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 362, + 327 + ], + "type": "text", + "content": "Values for Dockling provide a reference for the performance of an ensemble system." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": "non-textual elements. For different tasks, we conduct evaluations on diverse content types: full-page documents (from DocLayNet test set) in Markdown format, code snippets (from our SynthCodeNet dataset) in plaintext, and equations (from Im2Latex-230k [21]) in LaTeX format. For Qwen2.5-VL and SmolDocling, we convert HTML or DocTags output to Markdown while preserving formatting where possible." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 445, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 566 + ], + "type": "text", + "content": "Tables are omitted from the full-page output since they do not compare well through Markdown, due to loss of row and column spans. Code snippets must be reproduced with intact line breaks and indentation for accurate scoring. Formulas are evaluated separately in LaTeX format, which is natively produced by all the models tested in this work. The LaTeX outputs from all models, including the ground truth, undergo the same normalization procedure used for generating the SynthFormulaNet dataset. This ensures consistency and fairness in model comparisons." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 567, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 567, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 295, + 675 + ], + "type": "text", + "content": "SmolDocling significantly outperforms GOT, Nougat and also Qwen2.5-VL, which is 27 times larger, on every metric in full-page transcription (see Table 2. For the code parsing task, we report results exclusively for SmolDocling, as we introduce this task for the first time, and the other models were not explicitly trained for it. SmolDocling also demonstrates superior performance compared to Nougat and Qwen2.5-VL in the formula recognition task, closely matching GOT's results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Layout. We evaluate element localization accuracy using the DocLayNet test set (Table 3), measuring the ability to reconstruct document representations that support con" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 529 + ], + "type": "text", + "content": "tent grounding and visual feature extraction. SmolDocling is compared against Qwen2.5-VL-7b [9], currently the only other VLM addressing this task. For a fair comparison, we map each model's label outputs and the ground-truth to six compatible classes: Text, Section Heading, List Item, Table, Picture, and Formula. Overall, SmolDocling outperforms Qwen2.5-VL-7b by a significant margin, however both models score far below a human baseline. This can be partially explained through inherent difficulty of the dataset, as indicated by relatively low human agreement scores. The scores are primarily impacted by moderate recall of element bounding boxes and different degrees of label confusion in both SmolDocling and Qwen2.5-VL. A majority of samples show accurate element bounding boxes as seen in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 536, + 553, + 612 + ], + "blocks": [ + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "text", + "content": "Table 2. Structured document text recognition. We evaluate OCR performance and document formatting on DocLayNet, focusing on textual elements, excluding tables. Text accuracy is measured through multiple metrics (Edit Distance, F1-Score, Precision, Recall, BLEU and METEOR) as proposed in [12]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 536, + 553, + 612 + ], + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 612 + ], + "type": "table", + "html": "
Docling [7] v2.25SmolDoclingQwen2.5-VLHuman
Formula0.7370.2670.0590.84
Table0.6670.2660.2620.79
List Item0.5280.2960.0900.87
Section Header0.4710.2890.1290.83
Text0.4440.2040.1800.85
Picture0.3280.0660.0780.70
Overall0.5290.2310.1330.82
", + "image_path": "997a757850cb44a3c33b7e9ce5ea90adf3a8379a0bb14b679b0c87f684bdf624.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 620, + 555, + 665 + ], + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 665 + ], + "type": "text", + "content": "Table 3. Layout analysis comparison. Evaluation of layout analysis task on DocLayNet with a subset of 6 class labels in mAP[0.5:0.95] metric. Additional reference points are provided for an estimate of human performance [65]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Table Structure Recognition. We evaluate the accuracy of recovering table structures, i.e. shaping columns and rows with the correct text content and spans, through" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21978" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "the TEDS metric [93], once including table text content and once on structure only with omitted text (see Table 4). For isolated assessment, we provide cropped table images from FinTabNet [92] and PubTables-1M [69] test sets, which have poor image quality (72 dpi, compression artifacts). SmolDocling performs competitively against significantly larger models, despite challenges with text transcription from the low-resolution image crops, a limitation likely attributable to insufficient training at such resolutions. Scores for structure only TEDS evaluation are therefore much stronger." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 213, + 295, + 278 + ], + "blocks": [ + { + "bbox": [ + 57, + 213, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 57, + 213, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 57, + 213, + 295, + 278 + ], + "type": "table", + "html": "
MethodModel SizeFinTabNetPubTables-1M
SmolVLM [53]2.2B0.180.32
Granite Vision [72]2.98B0.540.7
TableFormer† [52]52M0.89 f0.84 (0.86)
Qwen2.5-VL 7B7B0.66- (0.786)-
SmolDocling (Ours)256M0.72 (0.81)0.83 (0.88)
SmolDocling (Ours) @150DPI256M0.76 (0.82)-
", + "image_path": "34aab7261b746c162248f2e8f3cb2bd13bb3d4873c83d2b26332d549b74c27e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "content": "Chart Extraction. In this task, we evaluate the ability to convert cropped sections of charts into datapoint tables on the ChartQA [54] dataset. All table output is transformed into HTML and compared against the ground-truth HTML using TEDS score, as reported in Table 5. We attribute SmolDocling's performance on charts to the variability in ground-truth styles and the differing interpretations of representations across datasets, which we observe leads to inconsistencies in the model's predictions. Despite its significantly smaller size, SmolDocling achieves a competitive TEDs score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 605 + ], + "type": "text", + "content": "Overall, SmolDocling achieves leading performance across a wide range of document conversion tasks and outperforms significantly larger models in text recognition, layout analysis and structure extraction. Furthermore, the model's high performance in downstream tasks, such as code listing and formula recognition, demonstrates its versatility and real-world applicability. Notably, we establish benchmark results for the code listing recognition, setting a standard for future evaluations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 167, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 167, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 167, + 624 + ], + "type": "text", + "content": "5.3. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 689 + ], + "type": "text", + "content": "Several characteristics of SmolDocling output are not evident from task-based metrics alone. In this section, we highlight different output characteristics observed in practice. More illustrated samples can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "As presented in Figure 2, SmolDocling outputs a sequence of DocTags to encode content, structure and layout." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 70, + 523, + 145 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 295, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 353 + ], + "type": "text", + "content": "Table 4. Table structure and cell content reconstruction. Numbers are given in TEDS metric including text content, with structure-only score in brackets where available. We produced results for SmolDocling and TableFormer, other results are reflected here as published. †TableFormer extracts the content through external OCR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 70, + 523, + 145 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 523, + 145 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 523, + 145 + ], + "type": "table", + "html": "
ModelSizeTEDs↑
Phi-3.5-vision4B0.40
Granite Vision [72]3B0.95
SmolVLM [53]2.2B0.02
Molmo-E1B0.54
SmolDocling (Ours)256M0.83
", + "image_path": "8cc44181d4dd5e4f8ac6d8310691ff9de091c61de341de58fedfc42e4b9499a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 154, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 553, + 198 + ], + "type": "text", + "content": "Table 5. Chart analysis. We compare SmolDocling to other small-sized VLMs on the chart-to-table task using the TEDS metric on the ChartQA dataset. The results in this table are based on the work reported in [25]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 225, + 553, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 553, + 320 + ], + "type": "text", + "content": "A few notable traits beyond straight content reproduction include: 1) Labeling of page headers and footers, which allows ignoring these repetitive elements in downstream applications. 2) Linking of information, such as captions to tables or pictures, or list elements to parent list through nesting of tags. 3) Preserved line breaks and indentation for code snippets, while line wraps in paragraphs are correctly removed." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 515 + ], + "type": "text", + "content": "Among typical defects found in the output of SmolDocling are missing tags (e.g. absent location tags), malformed structure (such as missing closing tags), and endless repetition of tokens from some point in a page. Both of these phenomena are commonly seen in auto-regressive models and may prevent parsing or rendering the output correctly (see supplementary material). Despite these limitations, SmolDocling demonstrates unique robustness advantages compared to ensemble systems like Docling, since conversion output is inferred in a single pass. In the latter, errors may accumulate throughout the ensemble model pipeline. For example, a mislocated table detection will lead to missing or distorted output of table structure and content, since this depends on correct table localization. In contrast, SmolDocling can reproduce the tables correctly, even with incorrectly determined location." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 535, + 388, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 388, + 547 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 388, + 547 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce SmolDocling, an efficient and compact VLM optimized for document conversion with rich output representation. We also present a suite of new datasets with a unified format for document conversion, including the novel task of code listing transcription. We identify page element localization as a critical area requiring further refinement, where targeted techniques will significantly enhance performance in future iterations. Our results conclusively demonstrate that smaller models with unified, optimized output formats such as DocTags can effectively compete with significantly larger models, establishing a clear pathway for resource-efficient multi-task document understanding models." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21979" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgements: We thank our collaborators at Hugging Face, Andrés Marafioti and Miquel Farré, who kindly and rapidly developed SmolVLM-256M-Instruct in response to our collaboration, providing the foundational model upon which this work is built." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 163, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 185 + ], + "type": "text", + "content": "[1] Grobid. https://github.com/kermitt2/grobid, 2008-2025. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 186, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 285 + ], + "type": "text", + "content": "[2] Loubna Ben Allal, Anton Lozhkov, Elie Bakouch, Gabriel Martin Blázquez, Guilherme Penedo, Lewis Tunstall, Andres Marafioti, Hynek Kydlíček, Agustín Piqueres Lajarín, Vaibhav Srivastav, Joshua Lochner, Caleb Fahlgren, Xuan-Son Nguyen, Clémentine Fourrier, Ben Burtenshaw, Hugo Larcher, Haojun Zhao, Cyril Zakka, Mathieu Morlon, Colin Raffel, Leandro von Werra, and Thomas Wolf. SmolLM2: When Smol Goes Big - Data-Centric Training of a Small Language Model, 2025. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "text", + "content": "[3] Amazon Web Services. Amazon Texttract. https://aws.amazon.com/texttract/, 2024. Accessed: 2024-11-11. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 365 + ], + "type": "text", + "content": "[4] Rohan Anil, Andrew M Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 366, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 398 + ], + "type": "text", + "content": "[5] Anthropic. Claude-3.5. https://www.anthropic.com/news/claude-3-5-sonnet, 2024. Accessed: 2024-02-11. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 400, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 295, + 465 + ], + "type": "text", + "content": "[6] Christoph Auer, Michele Dolfi, André Carvalho, Cesar BerroSpi Ramis, and Peter WJ Staar. Delivering Document Conversion as a Cloud Service with High Throughput and Responsiveness. In 2022 IEEE 15th International Conference on Cloud Computing (CLOUD), pages 363-373. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 521 + ], + "type": "text", + "content": "[7] Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Nikolaos Livathinos, Panos Vagenas, Cesar Berrospi Ramis, Matteo Omenetti, Fabian Lindlbauer, Kasper Dinkla, et al. Docling Technical Report. arXiv preprint arXiv:2408.09869, 2024. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 567 + ], + "type": "text", + "content": "[8] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 295, + 644 + ], + "type": "text", + "content": "[9] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL Technical Report, 2025. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[10] Ali Furkan Biten, Ruben Tito, Lluis Gomez, Ernest Valveny, and Dimosthenis Karatzas. Ocr-idl: Ocr annotations for industry document library dataset. In European Conference on Computer Vision, pages 241-252. Springer, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[11] Lukas Blecher. LaTeX-OCR, 2024. Accessed: 2024-11-09. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Lukas Blecher, Guillem Cucurull, Thomas Scialom, and Robert Stojnic. Nougat: Neural optical understanding for academic documents. arXiv preprint arXiv:2308.13418, 2023. 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[13] Jinyue Chen, Lingyu Kong, Haoran Wei, Chenglong Liu, Zheng Ge, Liang Zhao, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. OneChart: Purify the Chart Structural Extraction via One Auxiliary Token. arXiv preprint arXiv:2404.09987, 2024. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 229 + ], + "type": "text", + "content": "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 253 + ], + "type": "text", + "content": "[15] Pygments Contributors. Pygments: Python syntax highlighter, 2024. Accessed: 2024-11-06. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "text", + "content": "[16] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In European Conference on Computer Vision, pages 280-296. Springer, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[17] Yuntian Deng, Anssi Kanervisto, Jeffrey Ling, and Alexander M Rush. Image-to-markup generation with coarse-to-fine attention. In International Conference on Machine Learning, pages 980-989. PMLR, 2017. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 356, + 554, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 356, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 554, + 410 + ], + "type": "text", + "content": "[18] Mohamed Dhouib, Ghassen Bettaieb, and Aymen Shabou. Docparser: End-to-endOCR-free information extraction from visually rich documents. In Document Analysis and Recognition - ICDAR 2023, pages 155-172, Cham, 2023. Springer Nature Switzerland. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 412, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 554, + 509 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2: Mastering Free-form Text-Image Composition and Comprehension in Vision-Language Large Model. arXiv preprint arXiv:2401.16420, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 554, + 609 + ], + "type": "text", + "content": "[20] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Zhe Chen, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Kai Chen, Conghui He, Xingcheng Zhang, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer2-4KHD: A Pioneering Large Vision-Language Model Handling Resolutions from 336 Pixels to 4K HD. arXiv preprint arXiv:2404.06512, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 644 + ], + "type": "text", + "content": "[21] Gregory Eritsyan. Im2LaTeX 230K. https://www.kaggle.com/datasets/gregoryeritsyan/im2latex-230k, 2023. Accessed: 2025-03-06. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 667 + ], + "type": "text", + "content": "[22] Hugging Face. latex-formulas Dataset, 2024. Accessed: 2024-11-09. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[23] Philippe Gervais, Asya Fadeeva, and Andrii Maksai. Math-writing: A dataset for handwritten mathematical expression recognition. URL https://arxiv.org/abs/2404.10690, 2024.3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21980" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[24] Google Cloud. Document AI. https://cloud.google.com/document-ai, 2024. Accessed: 2024-11-11. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "text", + "content": "[25] IBM Granite Team. Granite 3.0 Language Models, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[26] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[27] Anwen Hu, Haiyang Xu, Jiabo Ye, Ming Yan, Liang Zhang, Bo Zhang, Chen Li, Ji Zhang, Qin Jin, Fei Huang, et al. mplug-docowl 1.5: Unified structure learning forOCR-free document understanding. arXiv preprint arXiv:2403.12895, 2024.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[28] Anwen Hu, Haiyang Xu, Liang Zhang, Jiabo Ye, Ming Yan, Ji Zhang, Qin Jin, Fei Huang, and Jingren Zhou. mplug-docowl2: High-resolution compressing for ocr-free multi-page document understanding. arXiv preprint arXiv:2409.03420, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[29] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4083-4091, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[30] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents. In 2019 International Conference on Document Analysis and Recognition Workshops (ICDARW), pages 1–6. IEEE, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "text", + "content": "[31] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[32] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[33] Shankar Kantharaj, Rixie Tiffany Ko Leong, Xiang Lin, Ahmed Masry, Megh Thakkar, Enamul Hoque, and Shafiq Joty. Chart-to-text: A large-scale benchmark for chart summarization. arXiv preprint arXiv:2203.06486, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[34] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7(15):2, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 647 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[36] Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Munoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra, and Harm de Vries. The Stack: 3 TB of permissively licensed source code. Preprint, 2022. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[37] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[38] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions., 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 219 + ], + "type": "text", + "content": "[39] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models?, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 222, + 553, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 288 + ], + "type": "text", + "content": "[40] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 289, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 356 + ], + "type": "text", + "content": "[41] David Lewis, Gady Agam, Shlomo Argamon, Ophir Frieder, David Grossman, and Jefferson Heard. Building a test collection for complex document information processing. In Proceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval, pages 665-666, 2006. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 357, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 357, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 357, + 553, + 402 + ], + "type": "text", + "content": "[42] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 403, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 553, + 448 + ], + "type": "text", + "content": "[43] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 504 + ], + "type": "text", + "content": "[44] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 553, + 540 + ], + "type": "text", + "content": "[45] Han Lin, Peng Yang, and Fanlong Zhang. Review of scene text detection and recognition. Archives of computational methods in engineering, 27(2):433-454, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 542, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 542, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 553, + 597 + ], + "type": "text", + "content": "[46] Fangyu Liu, Julian Martin Eisenschlos, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Wenhu Chen, Nigel Collier, and Yasemin Altun. Deplot: One-shot visual language reasoning by plot-to-table translation. arXiv preprint arXiv:2212.10505, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "text", + "content": "[47] Fangyu Liu, Francesco Piccinno, Syrine Krichene, Chenxi Pang, Kenton Lee, Mandar Joshi, Yasemin Altun, Nigel Collier, and Julian Martin Eisenschlos. Matcha: Enhancing visual language pretraining with math reasoning and chart derendering. arXiv preprint arXiv:2212.09662, 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 656, + 553, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 656, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 316, + 656, + 553, + 688 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21981" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "text", + "content": "[50] Nikolaos Livathinos, Christoph Auer, Maksym Lysak, Ahmed Nassar, Michele Dolfi, Panos Vagenas, Cesar Berropi Ramis, Matteo Omenetti, Kasper Dinkla, Yusik Kim, Shubham Gupta, Rafael Teixeira de Lima, Valery Weber, Lucas Morin, Ingmar Meijer, Viktor Kuropiatnyk, and Peter W. J. Staar. Docling: An Efficient Open-Source Toolkit for AI-driven Document Conversion, 2025. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[51] Nam Ly., Atsuhiro Takasu., Phuc Nguyen., and Hideaki Takeda. Rethinking Image-Based Table Recognition Using Weakly Supervised Methods. pages 872-880, 2023. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 238 + ], + "type": "text", + "content": "[52] Maksym Lysak, Ahmed Nassar, Nikolaos Livathinos, Christoph Auer, and Peter Staar. Optimized Table Tokenization for Table Structure Recognition. In Document Analysis and Recognition - ICDAR 2023, pages 37-50, Cham, 2023. Springer Nature Switzerland. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 314 + ], + "type": "text", + "content": "[53] Andres Marafioti, Orr Zohar, Miquel Farre, Merve Noyan, Elie Bakouch, Pedro Cuenca, Cyril Zakka, Loubna Ben Alal, Anton Lozhkov, Nouamane Tazi, Vaibhav Srivastav, Joshua Lochner, Hugo Larcher, Mathieu Morlon, Lewis Tunstall, Leandro von Werra, and Thomas Wolf. SmolVLM: Redefining small and efficient multimodal models. 2025. 1, 2, 3, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[54] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3, 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[55] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 447 + ], + "type": "text", + "content": "[56] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[57] Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. Chartassistant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning. arXiv preprint arXiv:2401.02384, 2024. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[58] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. PlotQA: Reasoning over Scientific Plots. In The IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "text", + "content": "[59] Microsoft Azure. AI Document Intelligence. https://azure.microsoft.com/en-us/products/ai-services.ai-document-intelligence, 2024. Accessed: 2024-11-11. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "text", + "content": "[60] OpenAI. Hello GPT-4O. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-02-09, 2024-02-11, 2024-02-12. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 646 + ], + "type": "text", + "content": "[61] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[62] Vik Paruchuri. Marker: Convert PDF to Markdown Quickly with High Accuracy. https://github.com/VikParuchuri/marker, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[63] Vik Paruchuri. Texify. https://github.com/VikParuchuri/texify, 2024. Accessed: 2024-11-11. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[64] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "text", + "content": "[65] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A Large Human-Annotated Dataset for Document-Layout Segmentation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, page 3743–3751, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "type": "text", + "content": "[66] Ruchir Puri, David Kung, Geert Janssen, Wei Zhang, Giacomo Domeniconi, Vladimir Zolotov, Julian Dolby, Jie Chen, Mihir Choudhury, Lindsey Decker, Veronika Thost, Luca Buratti, Saurabh Pujar, Shyam Ramji, Ulrich Finkler, Susan Malaika, and Frederick Reiss. CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Tasks, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[67] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[68] Felix M Schmitt-Koopmann, Elaine M Huang, Hans-Peter Hutter, Thilo Stadelmann, and Alireza Darvishy. MathNet: A Data-Centric Approach for Printed Mathematical Expression Recognition. IEEE Access, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "text", + "content": "[69] Brandon Smock, Rohith Pesala, and Robin Abraham. PubTables-1M: Towards comprehensive table extraction from unstructured documents. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4624–4632, 2022. 3, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 491 + ], + "type": "text", + "content": "[70] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19254-19264, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 547 + ], + "type": "text", + "content": "[71] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 712 + ], + "type": "text", + "content": "[72] Granite Vision Team, Leonid Karlinsky, Assaf Arbelle, Abraham Daniels, Ahmed Nassar, Amit Alfassi, Bo Wu, Eli Schwartz, Dhiraj Joshi, Jovana Kondic, Nimrod Shabtay, Pengyuan Li, Roei Herzig, Shafiq Abedin, Shaked Perek, Sivan Harary, Udi Barzelay, Adi Raz Goldfarb, Aude Oliva, Ben Wieles, Bishwaranjan Bhattacharjee, Brandon Huang, Christoph Auer, Dan GUTFreund, David Beymer, David Wood, Hilde Kuehne, Jacob Hansen, Joseph Shtok, Ken Wong, Luis Angel Bathen, Mayank Mishra, Maksym Lysak, Michele Dolfi, Mikhail Yurochkin, Nikolaos Livathinos, Nimrod Harel, Ophir Azulai, Oshri Naparstek, Rafael Teixeira de Lima, Rameswar Panda, Sivan Doveh, Shubham Gupta, Subhro Das, Syed Zawad, Yusik Kim, Zexue He, Alexander Brooks, Gabe Goodhart, Anita Govindjee, Derek Leist, Ibrahim Ibrahim, Aya Soffer, David Cox," + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21982" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 126 + ], + "type": "text", + "content": "Kate Soule, Luis Lastras, Nirmit Desai, Shila Ofek-koifman, Sriram Raghavan, Tanveer Syeda-Mahmood, Peter Staar, Tal Drory, and Rogerio Feris. Granite Vision: a lightweight, open-source multimodal model for enterprise Intelligence, 2025. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[73] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "type": "text", + "content": "[74] Michal Turski, Tomasz Stanisławek, Karol Kaczmarek, Paweł Dyda, and Filip Graliński. CCpdf: Building a High Quality Corpus for Visually Rich Documents from Web Crawl Data. In Document Analysis and Recognition - IC-DAR 2023, pages 348-365, Cham, 2023. Springer Nature Switzerland. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "text", + "content": "[75] Unstructured.io Team. Unstructured.io: Open-Source Pre-Processing Tools for Unstructured Data. https://unstructured.io, 2024. Accessed: 2024-11-19. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "text", + "content": "[76] Vicuna. Vicuna: An open-source chatbot impressing GPT-4 with " + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "text", + "content": " ChatGPT quality. https://vicuna.lmsys.org/, 2023.2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 381 + ], + "type": "text", + "content": "[77] Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, Fukai Shang, Bo Zhang, Liquun Wei, Zhihao Sui, Wei Li, Botian Shi, Yu Qiao, Dahua Lin, and Conghui He. MinerU: An Open-Source Solution for Precise Document Content Extraction, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 449 + ], + "type": "text", + "content": "[78] Maurice Weber, Carlo Siebenschuh, Rory Butler, Anton Alexandrov, Valdemar Thanner, Georgios Tsolakis, Haris Jabbar, Ian Foster, Bo Li, Rick Stevens, et al. WordScape: a Pipeline to extract multilingual, visually rich Documents with Layout Annotations from Web Crawl Data. Advances in Neural Information Processing Systems, 36, 2024. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[79] Haoran Wei, Chenglong Liu, Jinyue Chen, Jia Wang, Lingyu Kong, Yanming Xu, Zheng Ge, Liang Zhao, Jianjian Sun, Yuang Peng, et al. General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model. arXiv preprint arXiv:2409.01704, 2024. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "text", + "content": "[80] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[81] Wes McKinney and the Pandas Development Team. Pandas: powerful Python data analysis toolkit, Release 1.4.4. https://pandas.pydata.org/pandas-docs/ version/1.4/pandas.pdf, 2022.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "text", + "content": "[82] Renqiu Xia, Bo Zhang, Haoyang Peng, Hancheng Ye, Xiangchao Yan, Peng Ye, Botian Shi, Junchi Yan, and Yu Qiao. StructChart: Perception, Structuring, Reasoning for Visual Chart Understanding. arXiv preprint arXiv:2309.11268, 2023. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[83] Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Min Dou, Botian Shi, Junchi Yan, et al. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning. arXiv preprint arXiv:2402.12185, 2024. 3, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[84] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[85] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, et al. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. arXiv preprint arXiv:2012.14740, 2020.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 237 + ], + "type": "text", + "content": "[86] Zuoyu Yan, Xiaode Zhang, Liangcai Gao, Ke Yuan, and Zhi Tang. ConvMath: a convolutional sequence network for mathematical expression recognition. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 4566-4572. IEEE, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[87] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 294, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 381 + ], + "type": "text", + "content": "[88] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Shuangrui Ding, Songyang Zhang, Haodong Duan, Wenwei Zhang, Hang Yan, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition. arXiv preprint arXiv:2309.15112, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 383, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 480 + ], + "type": "text", + "content": "[89] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. InternLM-XComposer-2.5: A Versatile Large Vision Language Model Supporting Long-Contextual Input and Output. arXiv preprint arXiv:2407.03320, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 481, + 553, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 513 + ], + "type": "text", + "content": "[90] Qiming Zhang, Jing Zhang, Yufei Xu, and Dacheng Tao. Vision Transformer with Quadrangle Attention. arXiv preprint arXiv:2303.15105, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[91] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pretrained transformer language models. URL https://arxiv.org/abs/2205.01068, 3:19-0, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 570, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 634 + ], + "type": "text", + "content": "[92] Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 697-706, 2021. 3, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[93] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. arXiv preprint arXiv:1911.10683, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[94] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21983" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_content_list.json b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bdd60a3dfec903c21e9fecbb8023b1fb6add883c --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_content_list.json @@ -0,0 +1,1956 @@ +[ + { + "type": "text", + "text": "Snakes and Ladders: Two Steps Up for VideoMamba", + "text_level": 1, + "bbox": [ + 228, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hui Lu \nUtrecht University \nh.lu1@uu.nl", + "bbox": [ + 218, + 181, + 369, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Albert A. Salah \nUtrecht University \na.a.salah@uu.nl", + "bbox": [ + 423, + 181, + 573, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ronald Poppe \nUtrecht University \nr.w.poppe@uu.nl", + "bbox": [ + 627, + 181, + 777, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding requires the extraction of rich spatiotemporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by $1.6 - 3.0\\%$ and $1.1 - 1.9\\%$ top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models.", + "bbox": [ + 89, + 300, + 485, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 647, + 220, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video understanding is challenging and requires models that can extract rich spatio-temporal representations from video inputs. Transformers are powerful neural networks capable of effectively capturing temporal and spatial information from videos [19, 29, 50]. Current state-of-the-art video understanding models are transformers [41, 51]. At the core of transformers is self-attention [49], the self-alignment between tokens in an input obtained by estimating the relative importance of a token with respect to all other tokens. The long-range token dependency accounts for much of the success of transformer models [5, 49].", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Calculating self-attention is computationally costly, which eventually limits the application of powerful transformer models in practical settings [17]. Recently, alternative models with lower-cost operators have been proposed", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for national language processing (NLP), including S4 [13], RWKV [38], and RetNet [46]. Among these, Mamba [11] shows the best performance on long-range and causal tasks such as language understanding [31] and content-based reasoning [37].", + "bbox": [ + 511, + 268, + 903, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by the favorable computational cost, researchers have recently extended Mamba from the NLP domain to the computer vision domain. The core adaptation involved splitting the input image into multiple regions and embedding these as continuous tokens [61]. For video understanding, the recently proposed VideoMamba [22] extracts key frames from videos as the continuous input sequence. However, compared to previous transformer-based methods, VideoMamba's performance on video benchmarks is markedly lower. For example, VideoMamba achieves $82.4\\%$ top-1 on Kinetics-400, compared to $85.2\\%$ for VideoMAE [47], indicating room for improvement.", + "bbox": [ + 511, + 348, + 906, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we first analyze differences in the feature extraction capabilities of transformers and Mamba. We identify two limitations of Mamba: historical decay and element contradiction. We then extend VideoMamba [22] to mitigate these limitations. The proposed VideoMambaPro (VMP) addresses historical decay through masked backward computation in the bi-directional Mamba process, allowing the network to better handle historical tokens. We introduce residual connections to Mamba's matrix elements to tackle element contradiction. VideoMambaPro consistently improves the performance of VideoMamba on video understanding tasks, positioning it as a strong, efficient competitor to transformers. Our contributions are:", + "bbox": [ + 511, + 532, + 905, + 728 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We derive a formal representation of Mamba from the perspective of self-attention and identify two limitations of Mamba in the video analysis domain.", + "- We propose VideoMambaPro, effectively addressing Mamba's limitations for video understanding.", + "- We report strong video action recognition performance compared to recent Vision Mamba-based models, and surpass the original VideoMamba by clear margins." + ], + "bbox": [ + 511, + 731, + 903, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We first discuss related work. Then, we provide a theoretical analysis and introduce VideoMambaPro. Experiments are presented in Section 5. We conclude in Section 6.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24234", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 230, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers. One core aspect of transformers is self-attention [49] to achieve long-range interactions by measuring the similarity between tokens. Self-attention was introduced in the computer vision domain for tasks such as image recognition [26, 45] and object detection [7, 58]. Subsequent works (e.g., [9, 21, 47, 53] extended vision transformers to the video domain, to achieve superior performance. However, the mechanism of self-attention introduces significant computational overhead because of the similarity between all pairs of tokens needs to be calculated.", + "bbox": [ + 89, + 114, + 483, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternative models. Recent work has introduced alternative models with reduced computational complexity, while maintaining the advantages of self-attention [30, 40, 59]. SOFT [30] uses Gaussian kernel functions to replace the dot-product similarity, which enables a full self-attention matrix to be approximated with a low-rank matrix decomposition. Combiner [40] employs a structured factorization to approximate full self-attention, realizing low computation and memory complexity.", + "bbox": [ + 89, + 266, + 483, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Peng et al. [38] propose the Receptance Weighted Key Value (RWKV) architecture that combines self-attention training with an efficient recurrent neural network (RNN) inference using a linear attention mechanism. Through parallel computation, a lower, constant-level computational and memory complexity is achieved. RetNet [46] includes another variant of self-attention, by dividing the input into multiple chunks. Within each chunk, the self-attention mechanism can be computed in parallel, while information is transmitted between chunks based on an RNN.", + "bbox": [ + 89, + 402, + 483, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State-space models. The S4 model completely abandons self-attention and, instead, builds upon a state space model [13]. Instead of calculating a similarity matrix by performing matrix multiplications for pairs of tokens, it enables the network to directly learn a global high-order polynomial projection operator (HiPPO) matrix to handle relations between tokens. Additionally, for the simultaneous input of multiple tokens, S4 proposes a convolutional processing approach, enabling parallel training and thereby accelerating the training process.", + "bbox": [ + 89, + 553, + 483, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on S4, Mamba [11] proposes a selection mechanism where, for each input token, a unique HiPPO matrix [12] is generated. This allows the model to selectively process input tokens, enabling it to focus on or ignore specific inputs. Due to Mamba's strong representation ability in NLP, and linear-time complexity, it has garnered attention as a promising alternative to transformers. In the computer vision domain, researchers have proposed Vision Mamba [61] and VMamba [27] for tasks such as image classification and object detection.", + "bbox": [ + 89, + 704, + 483, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the video domain, several Mamba variants have been proposed [22, 23, 34]. Their performance is somewhat lower than expected, with limited understanding of the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "causes. We argue that a systematic, mathematical analysis of Mamba from the perspective of self-attention could reveal shortcomings of Mamba's inner workings. Better understanding of these limitations allow us to develop improvements, and to close the accuracy performance gap with transformers, while enjoying Mamba's efficiency.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 511, + 194, + 709, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We revisit Mamba from the perspective of self-attention. Then, we analyze its limitations for video understanding. To address these, we propose VideoMambaPro in Section 4.", + "bbox": [ + 511, + 219, + 905, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Mamba from the perspective of self-attention", + "text_level": 1, + "bbox": [ + 511, + 273, + 895, + 289 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-attention. Given an input sequence $\\mathbf{X} \\coloneqq [x_1, \\dots, x_N] \\in \\mathbb{R}^{N \\times D_x}$ of $N$ feature vectors of depth $D_x$ , self-attention [49, 60] computes the output sequence $\\mathbf{Y}$ from $\\mathbf{X}$ following two steps:", + "bbox": [ + 511, + 295, + 905, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step 1: Similarity matrix computation. The input sequence $\\mathbf{X}$ is linearly projected onto the three different subspaces query $\\mathbf{Q} \\in \\mathbb{R}^{N \\times D}$ , key $\\mathbf{K} \\in \\mathbb{R}^{N \\times D}$ , and value $\\mathbf{V} \\in \\mathbb{R}^{N \\times D_V}$ :", + "bbox": [ + 511, + 354, + 905, + 415 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\boldsymbol {X} \\mathbf {W} _ {Q} ^ {\\top}, \\mathbf {K} = \\boldsymbol {X} \\mathbf {W} _ {K} ^ {\\top}, \\mathbf {V} = \\boldsymbol {X} \\mathbf {W} _ {V} ^ {\\top}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 428, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with $\\mathbf{W}_Q, \\mathbf{W}_K \\in \\mathbb{R}^{D \\times D_x}$ , and $\\mathbf{W}_V \\in \\mathbb{R}^{D_v \\times D_x}$ the corresponding weight matrices. Specifically, $\\mathbf{Q} := [q_1, \\dots, q_N]^\\top$ , $\\mathbf{K} := [k_1, \\dots, k_N]^\\top$ , and $\\mathbf{V} := [v_1, \\dots, v_N]^\\top$ with vectors $q_i, k_i, v_i$ for $i = 1, \\dots, N$ the query, key, and value vectors, respectively, for input $i$ . Based on $\\mathbf{Q}$ and $\\mathbf{K}$ , similarity matrix $\\mathbf{S} \\in \\mathbb{R}^{N \\times N}$ contains the correlations between all query and key vectors, with a softmax function applied to each row of $\\mathbf{S}$ :", + "bbox": [ + 511, + 452, + 905, + 577 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top} / \\sqrt {D}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 584, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Each component $s_{ij} (i,j = 1,\\dots ,N)$ represents the similarity score between $\\pmb{q}_i$ and $\\pmb{k}_j$ .", + "bbox": [ + 511, + 612, + 903, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step 2: Output computation. Output sequence $\\mathbf{Y} := [y_1, \\dots, y_N]^\\top$ is then calculated based on $\\mathbf{S}$ as:", + "bbox": [ + 513, + 642, + 903, + 674 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {S V}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 683, + 903, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It follows that each output vector $\\pmb{y}_i$ ( $i = 1, \\dots, N$ ) can be written in vector form as:", + "bbox": [ + 511, + 708, + 905, + 737 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {i} = \\sum_ {j = 1} ^ {N} s _ {i j} \\boldsymbol {v} _ {j}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 744, + 903, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Any output vector $\\mathbf{y}_i$ is a linear combination of vectors $\\mathbf{v}_j (j = 1, \\dots, N)$ , with similarity score $s_{ij}$ serving as coefficient. The larger the similarity score, the greater the influence of $\\mathbf{v}_j$ on output $\\mathbf{y}_i$ [43].", + "bbox": [ + 511, + 795, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mamba. State Space Models (SSMs) serve as the foundation of Mamba [11]. They are based on continuous systems that map 1D functions or sequences, $x(t) \\in \\mathbb{R}^L \\to$", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24235", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$y(t)\\in \\mathbb{R}^{L}$ to output sequences $y(t)$ through a hidden state $h(t)\\in \\mathbb{R}^N$ . Formally, SSM implements the mapping as:", + "bbox": [ + 89, + 90, + 485, + 125 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 143, + 482, + 159 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny (t) = \\mathbf {C} h (t) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 162, + 482, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ is the evolution matrix of the system, and $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ , $\\mathbf{C} \\in \\mathbb{R}^{N \\times 1}$ are the projection matrices. Often, the input is discrete rather than a continuous function $x(t)$ . Therefore, Mamba performs discretization, effectively creating a discrete version of the continuous system. A timescale parameter $\\Delta$ is used to transform the continuous parameters $\\mathbf{A}$ and $\\mathbf{B}$ into their discrete counterparts $\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}$ , and the transformation typically employs the zero-order hold method [57]. This process is expressed as:", + "bbox": [ + 89, + 185, + 483, + 323 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 328, + 482, + 345 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} \\left(\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}\\right) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 348, + 482, + 366 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 369, + 482, + 386 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = \\mathbf {C h} _ {t}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 390, + 480, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that parameters $\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}$ in the original SSM are independent of the input data $x(t)$ and cannot be tailored to specific input data, Mamba employs a Selective Scan Mechanism as its core operator. More precisely, three functions $S_B(x),S_C(x),S_\\Delta (x)$ are introduced to associate parameters $\\overline{\\mathbf{B}}$ , C, and $\\pmb{\\Delta}$ in Eqs. 7-10 to the input data $x$ Based on $S_{\\Delta}(x)$ $\\overline{\\mathbf{A}}$ can also be associated with the input data $x$ . For example, given the input $x_{1}$ , functions $S_{\\Delta}(x)$ will produce the corresponding $\\overline{\\mathbf{A}}_1$ based on Eq. 7, and functions $S_B(x)$ and $S_{\\Delta}(x)$ will produce the corresponding $\\overline{\\mathbf{B}}_1$ based on Eq. 8. $\\overline{\\mathbf{C}}_1$ is obtained based on function $S_{C}(x)$ . Following Eqs. 9 and 10, we analyze the process to obtain output sequence $\\mathbf{Y}$ when given an input sequence $\\mathbf{X}:= [x_1,\\dots ,x_N]\\in \\mathbb{R}^{N\\times D_x}$ of $N$ feature vectors. Each vector's hidden state is denoted as:", + "bbox": [ + 89, + 412, + 483, + 641 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {1} = \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 662, + 482, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h _ {2} = \\overline {{\\mathbf {A}}} _ {2} h _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ = \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2}, \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 683, + 480, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h _ {3} = \\overline {{\\mathbf {A}}} _ {3} h _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3} \\\\ = \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {I} + \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3}, \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 724, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": ".", + "bbox": [ + 124, + 768, + 145, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h _ {N} = \\overline {{\\mathbf {A}}} _ {N} h _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N} \\\\ = \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {B}}} _ {N - 1} \\boldsymbol {x} _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N}. \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 784, + 480, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eqs. 11-14 can be written in matrix form:", + "bbox": [ + 532, + 90, + 812, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {H} = \\left[ h _ {1}, h _ {2}, h _ {3}, \\dots , h _ {N} \\right] ^ {\\top} \\\\ = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {1} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right]. \\tag {15} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 116, + 903, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For output sequence $\\mathbf{Y} := [\\pmb{y}_1, \\dots, \\pmb{y}_N]^\\top$ , each vector $\\pmb{y}_i$ ( $i = 1, \\dots, N$ ) can be expressed as:", + "bbox": [ + 513, + 228, + 903, + 260 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {N} = \\overline {{\\mathbf {C}}} _ {N} h _ {N}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 270, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and in matrix form as:", + "bbox": [ + 513, + 299, + 663, + 313 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {C}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ 0 & \\overline {{\\mathbf {C}}} _ {2} & 0 & \\dots & 0 \\\\ 0 & 0 & \\overline {{\\mathbf {C}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ 0 & 0 & 0 & \\dots & \\overline {{\\mathbf {C}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\\\ \\vdots \\\\ h _ {N} \\end{array} \\right] = \\overline {{\\mathbf {C}}} h. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 321, + 903, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By substituting Eq. 15 into Eq. 17, we obtain:", + "bbox": [ + 532, + 421, + 836, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\overline {{\\mathbf {C}}} \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {I} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right], \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 446, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which can be expressed as:", + "bbox": [ + 513, + 532, + 694, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\overline {{\\mathbf {C}}} (\\mathbf {M} \\mathbf {X}), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 558, + 903, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ represents the second term on the right-hand side of Eq. 18. Recall from Eq. 3 that the result $\\mathbf{Y}$ obtained by self-attention processing can be expressed as:", + "bbox": [ + 511, + 587, + 903, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = \\mathbf {S V} = (\\mathbf {S X}) \\mathbf {W} _ {V} ^ {\\top} \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 643, + 903, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the perspective of self-attention, by comparing Eqs. 19 and 20, the essence of Mamba is to generate a matrix $\\mathbf{M}$ similar to similarity matrix $\\mathbf{S}$ , such that the result of $\\mathbf{M}\\mathbf{X}$ is based on the correlation between vectors of $\\mathbf{X}$ . Although the final result of $\\mathbf{M}\\mathbf{X}$ is left multiplied by a mapping matrix $\\mathbf{C}$ , while the result of $\\mathbf{S}\\mathbf{X}$ is right multiplied by a mapping matrix $\\mathbf{W}_V^\\top$ , the geometric meaning of the two are the same.", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limitations of Mamba in video understanding", + "text_level": 1, + "bbox": [ + 511, + 801, + 903, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From the perspective of self-attention, the concept of Mamba is similar: both use similarity matrices. We now analyze the differences between the similarity matrices of Mamba and self-attention, and discuss the limitations of Mamba in the context of the video understanding task.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The original SSM [13] employs $h'(t) = Ah(t) + Bx(t)$ , with $h(t)$ the hidden state from previous time step $t - 1$ , and $h'(t)$ the updated current hidden state, replacing $h(t)$ . Considering this approach may lead to ambiguity, we have adopted the updated description.", + "bbox": [ + 89, + 849, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24236", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitation 1: Historical decay. Matrix M in Eq. 19 corresponds to the second right-hand term in Eq. 18, which is a lower triangular matrix of the form:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\left[ \\begin{array}{c c c c c} m _ {1 1} & 0 & 0 & \\dots & 0 \\\\ m _ {2 1} & m _ {2 2} & 0 & \\dots & 0 \\\\ m _ {3 1} & m _ {3 2} & m _ {3 3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {N 1} & m _ {N 2} & m _ {N 3} & \\dots & m _ {N N} \\end{array} \\right]. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 143, + 483, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By comparing $\\mathbf{M}$ with matrix $\\mathbf{S}$ in self-attention, we find that outputs in Mamba favor more recent information, because the more weights are zero, the earlier the token is observed. For example, for input $[x_1, x_2, x_3]$ , the output $\\mathbf{M}x_1$ in Mamba is $m_{11}x_1$ while the output $\\mathbf{S}x_1$ is $s_{11}x_1 + s_{12}x_2 + s_{13}x_3$ in self-attention. This indicates that, in Mamba, the influence of earlier observed tokens on the final result is greatly diminished. We refer to this limitation as historical decay.", + "bbox": [ + 89, + 239, + 483, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the NLP domain, more recent dialogue information often has more impact on the final judgment, so this effect is acceptable. However, in the computer vision domain, the order of the tokens has less meaning. Previous works such as Vision Mamba [61] and VMamba [27] tried to solve this issue by processing the token sequence in both forward and backward directions. This produces better results, but significant deficiencies still exist. We will explain this theoretically below.", + "bbox": [ + 89, + 375, + 483, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When processing bi-directionally, the results generated from input forward tokens $\\left[x_1,\\dots ,x_N\\right]$ , denoted as $\\mathbf{M}_f\\mathbf{X}$ and the results generated from input backward tokens $\\left[x_{N},\\dots ,x_{1}\\right]$ , denoted as $\\mathbf{M}_b\\mathbf{X}$ , are linearly combined to generate the final result $\\mathbf{M}_{bi}\\mathbf{X}$ with $\\mathbf{M}_{bi}$ a dense matrix. As a result, the influence of historical information on the result is increased, consequently leading to better results.", + "bbox": [ + 89, + 511, + 483, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For example, for the input tokens $[x_1, x_2, x_3]$ , $\\mathbf{M}_f\\mathbf{X}$ and $\\mathbf{M}_b\\mathbf{X}$ can be expressed as:", + "bbox": [ + 89, + 616, + 482, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {f} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} f _ {1 1} & 0 & 0 \\\\ f _ {2 1} & f _ {2 2} & 0 \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {1 f} \\\\ h _ {2 f} \\\\ h _ {3 f} \\end{array} \\right], \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 654, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {b} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} b _ {3 3} & 0 & 0 \\\\ b _ {2 3} & b _ {2 2} & 0 \\\\ b _ {1 3} & b _ {1 2} & b _ {1 1} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {3} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {1} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {3 b} \\\\ h _ {2 b} \\\\ h _ {1 b} \\end{array} \\right], \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 723, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{ij}$ represents the similarity score during the forward process, and $b_{ij}$ is the similarity score in the backward direction. After bi-directional computation, with the outputs linearly combined, the results are expressed as:", + "bbox": [ + 89, + 780, + 483, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {1} = h _ {1 f} + h _ {1 b} = f _ {1 1} \\boldsymbol {x} _ {1} + b _ {1 3} \\boldsymbol {x} _ {3} + b _ {1 2} \\boldsymbol {x} _ {2} + b _ {1 1} \\boldsymbol {x} _ {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 849, + 442, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {2} = h _ {2 f} + h _ {2 b} = f _ {2 1} \\boldsymbol {x} _ {1} + f _ {2 2} \\boldsymbol {x} _ {2} + b _ {2 3} \\boldsymbol {x} _ {3} + b _ {2 2} \\boldsymbol {x} _ {2} \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 868, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {3} = h _ {3 f} + h _ {3 b} = f _ {3 1} \\boldsymbol {x} _ {1} + f _ {3 2} \\boldsymbol {x} _ {2} + f _ {3 3} \\boldsymbol {x} _ {3} + b _ {3 3} \\boldsymbol {x} _ {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 887, + 444, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can write Eq. 24 in matrix form:", + "bbox": [ + 532, + 90, + 774, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\end{array} \\right] = \\left[ \\begin{array}{c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] \\tag {25} \\\\ = \\mathbf {M} _ {b i} \\left[ \\begin{array}{c} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 116, + 903, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The bi-directional computation transforms matrix $\\mathbf{M}$ from a lower triangular matrix to a dense matrix $\\mathbf{M}_{bi}$ , thereby capturing more historical information and effectively avoiding historical decay. When extending to the case of $N$ input tokens $[\\pmb{x}_1,\\dots ,\\pmb{x}_N]$ , $\\mathbf{M}_{bi}$ can be written as:", + "bbox": [ + 511, + 224, + 906, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} + b _ {N N} \\end{array} \\right]. \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 311, + 905, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diagonal elements of $\\mathbf{M}_{bi}$ contain duplicates of the similarity between a token and itself. For example, $f_{33}$ and $b_{33}$ each represent the similarity between token $x_{3}$ and itself. Consequently, the similarity is effectively doubled which weakens the association with other tokens. One possible approach is to adjust $\\mathbf{M}_f$ and $\\mathbf{M}_b$ using a weight coefficient $z$ through a linear combination. However, this method is not very effective. For example, to prevent the similarity of diagonal elements from being doubled, the $z$ value is set to 0.5. However, after multiplying other elements by 0.5, their weight values are significantly reduced, which means that the association of each token with other tokens is weakened.", + "bbox": [ + 511, + 391, + 906, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitation 2: Element contradiction. By analyzing non-zero elements $m_{ij}$ in $\\mathbf{M}$ of Eq. 18, we derive:", + "bbox": [ + 511, + 587, + 905, + 618 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 628, + 903, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After multiple iterations, the above equation results in implicit consideration of the correlation between previous tokens and token $j$ when computing the correlation between token $i$ and token $j$ . As a result, $m_{ij}$ exhibits stronger contextual dependencies compared to the elements $s_{ij}$ in the matrix S. This might explain why Mamba achieves better performance than transformers in the field of NLP. While this is advantageous in the NLP domain, for the computer vision domain, input tokens often lack semantic connections. The consideration of the influence of other tokens on each element can lead to significant drawbacks.", + "bbox": [ + 511, + 657, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interleaved token structures are common when processing images. Tokens that \"belong together\" might not be subsequently processed. For example, in an image classification task, input tokens $\\left[\\pmb{x}_1,\\pmb{x}_2,\\pmb{x}_3\\right]$ might represent image regions [dog, other, dog]. Ideally, $m_{31}$ should be high and", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24237", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba5a91a49994f4e3e836d8a9169fa896086e5500ab54a7d793597e9edf1ea360.jpg", + "image_caption": [ + "(a) VideoMambaPro" + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 580, + 316 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f59be29a79303a6ae787aed361d4bf21fdb4bb8b47d06148eb1469d74093c07.jpg", + "image_caption": [ + "(b) Bi-Directional Mamba Block", + "Figure 1. (a) Framework of VideoMambaPro with $K$ bi-directional Mamba blocks. (b) In each bi-directional Mamba block, we employ forward residual SSM and masked backward Residual SSM." + ], + "image_footnote": [], + "bbox": [ + 606, + 89, + 864, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$m_{21}$ low. Following Eq. 27, $m_{31} = \\overline{\\mathbf{A}}_3 m_{21}$ , so $\\overline{\\mathbf{A}}_3$ needs to increase. However, this causes $m_{32} = \\overline{\\mathbf{A}}_3 m_{22}$ to increase because $m_{22}$ is also high. But, theoretically, $m_{32}$ should be low. This causes an element contradiction. Especially for video understanding, such contradictions are common because most video regions contain background and other irrelevant information, making relevant tokens sparse. Consequently, the performance of Mamba applied to video analysis tasks is underwhelming [22, 24, 56].", + "bbox": [ + 88, + 396, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. VideoMambaPro", + "text_level": 1, + "bbox": [ + 89, + 545, + 261, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use VideoMamba [22] as backbone, and propose two adaptations to address historical decay and element contradiction. The resulting architecture is termed VideoMambaPro (VMP). With minor adjustments, our adaptations can also be applied to related Mamba models.", + "bbox": [ + 89, + 570, + 483, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address historical decay, we keep the result of $\\mathbf{M}_f\\mathbf{X}$ but we use masked computation in the backward process. Specifically, we assign a mask to the diagonal elements of $\\mathbf{M}_b$ , setting their values to 0, and then proceed with the calculations in Eqs 21-25. We thus eliminate the duplicate similarity on the diagonal, without affecting other elements. The final $\\mathbf{M}_{bi}$ is expressed as:", + "bbox": [ + 89, + 646, + 483, + 751 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} \\end{array} \\right]. \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 760, + 482, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve element contradiction, we propose residual SSM. This solution is inspired by residual connections, to distribute the requirement for $\\overline{\\mathbf{A}}_i$ in $m_{ij}$ over multiple $\\overline{\\mathbf{A}}_i$ .", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This helps to avoid contradictions caused by interleaved sequence structures. For example, for our previous example input sequence $[x_1, x_2, x_3]$ , which represents regions [dog, other, dog], we let $m_{31} = \\overline{\\mathbf{A}}_3 m_{21} + \\overline{\\mathbf{A}}_3$ . This way, the requirement for a single $\\overline{\\mathbf{A}}_3$ can be split into two parts, thus avoiding contradictions. This can be expressed as:", + "bbox": [ + 511, + 396, + 905, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} + \\overline {{\\mathbf {A}}} _ {i} \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 494, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement these solutions into VideoMamba [22], to form VideoMambaPro (see Figure 1). Given input video $\\pmb{X}^v \\in \\mathbb{R}^{3 \\times T \\times H \\times W}$ , we first use a 3D convolution with a $1 \\times 16 \\times 16$ kernel to convert $\\pmb{X}^v$ into $L$ non-overlapping patch-wise tokens $\\pmb{X}^p \\in \\mathbb{R}^{L \\times C}$ with $L = t \\times h \\times w$ ( $t = T, h = \\frac{H}{16}, w = \\frac{W}{16}$ ). Because SSM is sensitive to token positions, and in line with [22], we include learnable spatial and temporal position embeddings $\\pmb{p}_s \\in \\mathbb{R}^{(hw + 1) \\times C}$ and $\\pmb{p}_t \\in \\mathbb{R}^{t \\times C}$ . Input tokens $\\pmb{X}$ are expressed as:", + "bbox": [ + 511, + 520, + 906, + 657 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} = \\left[ \\boldsymbol {X} _ {\\text {c l s}}, \\boldsymbol {X} \\right] + \\boldsymbol {p} _ {s} + \\boldsymbol {p} _ {t}, \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 666, + 903, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{cls}$ is a learnable classification token positioned at the start of the sequence. Input tokens $X$ pass through $K$ Mamba blocks, and the final layer's [CLS] token is used for classification, after normalization and linear projection.", + "bbox": [ + 511, + 690, + 905, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 762, + 648, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 787, + 700, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate VideoMambaPro on five video benchmarks: (a) Kinetics-400 (K400, [3]) comprises $\\sim 240\\mathrm{K}$ training and $\\sim 20\\mathrm{K}$ validation videos, each with an average duration of 10 seconds and categorized into 400 classes. (b) Something-Something V2 (SSv2, [10]) includes $\\sim 160\\mathrm{K}$ training and $\\sim 20\\mathrm{K}$ validation videos with", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24238", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an average duration of 4 seconds, and 174 motion-centric classes. (c) UCF-101 [44] is a relatively small dataset, consisting of $\\sim 9.5\\mathrm{K}$ training and $\\sim 3.5\\mathrm{K}$ validation videos. (d) HMDB51 [18] is also a compact video dataset, containing $\\sim 3.5\\mathrm{K}$ training and $\\sim 1.5\\mathrm{K}$ validation videos. (e) AVA [14] is a dataset for spatio-temporal localization of human actions with $\\sim 211\\mathrm{k}$ and $\\sim 57\\mathrm{k}$ validation video segments.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. In line with VideoMamba, we introduce three models with increasing embedding dimension and number of bi-directional Mamba blocks $K$ : Tiny, Small, and Middle (details in supplementary material). To compare with VideoMamba, we pre-train VideoMambaPro on ImageNet-1K (IN-1K). On K400, we also pre-train with IN-1K, fine-tune on the training set and report on the validation set. For K400, we also report on the larger $336^2$ input size. During pre-training, we follow DeiT [48] by applying a center crop to obtain the $224^2$ sized images. We apply random cropping, random horizontal flipping, label-smoothing regularization, mix-up, and random erasing as data augmentations. We use AdamW [28] with a momentum of 0.9, a batch size of 1024, and a weight decay of 0.05. We employ a cosine learning rate schedule during training, $1 \\times 10^{-3}$ initial learning rate over 300 epochs. The fine-tuning settings follow VideoMAE [47]. We resize frames to $224^2$ , and use AdamW with a momentum of 0.9 and a batch size of 512. The evaluation process is the same as VideoMamba, and more details are in the supplementary materials. These materials also include results on IN-1K image classification.", + "bbox": [ + 91, + 199, + 486, + 517 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d614d751fe34406f1a0340a6651b03a8f5f86d3e0f74235b004d6380b03d710c.jpg", + "table_caption": [], + "table_footnote": [ + "$^{1}$ IN-21K+WTS", + "2 CLIP-400M+WebVid+HowTo+K710+SSv2+AVA2.2+more.", + "3 LAION-300M+KMash+WebVid+InternVid+LLaVA+more." + ], + "table_body": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]32 × 22425×137M350G80.294.4
MViTv2-S [25]16 × 22425×135M320G81.094.6
Uniformer-S [20]IN-1K16 × 22424×121M168G80.894.7
Uniformer-B [20]IN-1K16 × 22424×150M388G82.095.1
Uniformer-B [20]IN-1K32 × 22424×350M3.1T83.095.4
STAM [42]IN-21K64 × 22421×1121M1.0T79.2-
TimeSformer-L [2]IN-21K96 × 22421×3121M7.1T80.794.7
ViViT-L [1]IN-21K16 × 22424×3311M47.9T81.394.7
Mformer-HR [36]IN-21K16 × 336210×3311M28.8T81.195.2
VideoMAE-H [47]IN-21K16 × 22425×3633M17.9T86.697.1
X-CLIP-L/14 [32]CLIP-400M16 × 33624×3453M37.0T87.7-
MTV-H [55]60M132 × 22424×31120M44.5T89.198.2
InternVideo-1B [51]412M264 × 224216×41300M86.2T91.198.9
InternVideo-2B [52]414M316 × 224216×41000M-91.6-
InternVideo-2B-6B [52]414M316 × 224216×45903M-92.1-
VideoMamba-TiIN-1K32 × 22424×37M0.4T78.893.9
VideoMamba-TiIN-1K64 × 38424×37M2.4T80.394.8
VideoMamba-SIN-1K32 × 22424×326M1.6T81.595.2
VideoMamba-SIN-1K64 × 38424×326M4.7T82.795.6
VideoMamba-MIN-1K32 × 22424×374M4.8T82.495.7
VideoMamba-MIN-1K64 × 38424×374M28.4T83.396.1
VideoMambaPro-TiIN-1K32 × 22424×37M0.4T81.695.9
VideoMambaPro-TiIN-1K64 × 38424×37M2.2T83.396.1
VideoMambaPro-SIN-1K32 × 22424×325M1.6T83.396.0
VideoMambaPro-SIN-1K64 × 38424×325M4.4T84.596.6
VideoMambaPro-MIN-1K32 × 22424×372M4.7T84.096.4
VideoMambaPro-MIN-1K64 × 38424×372M27.0T85.096.7
", + "bbox": [ + 93, + 532, + 480, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with state-of-the-art", + "text_level": 1, + "bbox": [ + 513, + 90, + 803, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "K400. Results appear in Table 1. Compared to VideoMamba, VideoMambaPro has slightly fewer parameters and FLOPs. This is primarily because VideoMamba employs an additional projection layer to generate the weight coefficient $z$ to adjust $\\mathbf{A}_f$ and $\\mathbf{A}_b$ . See the supplementary materials for an architecture comparison. VideoMambaPro outperforms VideoMamba across model and input sizes. With $224^2$ inputs and pre-trained only on IN-1K, the best-performing VideoMambaPro-M achieves a top-1 accuracy of $84.0\\%$ , $1.6\\%$ higher than VideoMamba-M. Further comparisons appear in Section 5.4. Increasing the input size to $336^2$ leads to a performance improvement of $1.0 - 1.7\\%$ .", + "bbox": [ + 511, + 118, + 906, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VideoMambaPro scores lower than the recent InternVideo2-1B [52] by $7.6\\%$ , but was only pre-trained on IN-1K and has significantly fewer parameters (1000M vs 72M) and inference only takes $\\sim 5.5\\%$ of the FLOPs.", + "bbox": [ + 511, + 306, + 906, + 367 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9873961af44110943ecde5138752f5bda859091ec957061dc4b165199e54e5ed.jpg", + "table_caption": [ + "Table 1. Performance on K400. Top part of the table are Transformer models, bottom part are Mamba models. We report crops (temporal $\\times$ spatial) and FLOPs for inference. —: not reported." + ], + "table_footnote": [], + "table_body": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]K40016 × 22421×337M213G64.789.2
MViTv1-B [6]K40032 × 22421×337M510G67.190.8
MViTv2-S [25]K40016 × 22421×335M195G68.291.4
MViTv2-B [25]K40032 × 22421×351M675G70.592.7
Uniformer-S [20]IN-1K+K40016 × 22421×321M126G67.791.4
Uniformer-B [20]IN-1K+K40016 × 22421×350M291G70.492.8
TimeSformer-L [2]IN-21K16 × 22421×3121M5.1T62.5-
ViViT-L [1]IN-21K+K40016 × 22424×3311M47.9T65.489.8
Mformer-HR [36]IN-21K+K40016 × 33621×3311M3.6T68.191.2
MaskFeat-L [53]IN-21K64 × 31224×3218M8.5T75.095.0
VideoMAE-L [47]IN-21K32 × 22421×3305M4.3T75.495.2
TubeViT-L [39]IN-1K32 × 22424×3311M9.5T76.195.2
InternVideo-1B [51]See Table 164 × 224216×41300M86.2T77.295.9
InternVideo2-1B [52]See Table 164 × 224216×41000M-77.1-
InternVideo2-6B [52]See Table 164 × 224216×45903M-77.4-
VideoMamba-TiIN-1K16 × 22422×37M102G66.089.6
VideoMamba-SIN-1K16 × 22422×326M408G67.690.9
VideoMamba-MIN-1K16 × 22424×374M2.4T68.391.4
VideoMambaPro-TiIN-1K16 × 22422×37M96G67.991.2
VideoMambaPro-SIN-1K16 × 22422×325M382G68.891.4
VideoMambaPro-MIN-1K16 × 22424×372M2.2T69.491.6
", + "bbox": [ + 517, + 392, + 903, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on SSv2. —: not reported. Top part of the table are Transformer models, bottom part are Mamba models.", + "bbox": [ + 511, + 609, + 906, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SSv2. Results appear in Table 2. VideoMambaPro outperforms VideoMamba by $1.1 - 1.9\\%$ . It also outperforms several popular transformer models. Although InternVideo-1B [51] and InternVideo2-6B [52] outperform our VideoMambaPro-M by $7.8\\%$ and $8.0\\%$ , respectively, they require 18.0-82 times more parameters and at least 39 times more FLOPs. Again, we expect that the performance for VideoMambaPro will increase with more pre-training.", + "bbox": [ + 511, + 667, + 908, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "UCF-101/HMDB51/AVA V2.2. From Table 3, it shows that VideoMambaPro-M is competitive, and outperforms VideoMamba by $3.4\\%$ and $1.8\\%$ on UCF-101 and HMDB51, respectively. VideoMambaPro-M achieves 31.9 mAP on AVA V2.2, which is $10.7\\%$ lower than VideoMAE V2 [50] but with an order of magnitude fewer parameters and FLOPs and pre-trained only on IN-1K (see Table 4).", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24239", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d285d62d221af0537f4a82fa9470e2657666abc389021f9da8f1bdf87498721b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsUCF-101HMDB51
VideoMoCo [33]15M78.749.2
CoCLR [16]9M81.452.1
MemDPC [15]32M86.154.5
Vi²CLR [4]9M89.155.7
VideoMAE [47]87M91.362.6
GDT [35]33M95.272.8
VideoMAE V2 [50]1050M99.688.1
VideoMamba-M74M88.260.8
VideoMambaPro-M72M91.663.2
", + "bbox": [ + 135, + 88, + 441, + 224 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2e62904c263e90f6a88724f7ebf48cc52d0d8c876beadd24dd8b2cdb471cff7.jpg", + "table_caption": [ + "Table 3. Results on UCF-101 and HMDB51." + ], + "table_footnote": [], + "table_body": "
MethodFLOPsParammAP
SlowFast R101 [8]138G53M23.8
VideoMAE-B [47]180G87M26.7
MViTv2-B [25]225G51M30.5
ObjectTransformer [54]243G86M31.0
MViTv2-L [25]2828G213M34.4
ST-MAE-H [9]1193G632M36.2
VideoMAE V2 [50]4220G1050M42.6
VideoMamba-M [19]202G74M30.1
VideoMambaPro-M183G72M31.9
", + "bbox": [ + 140, + 263, + 433, + 404 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24e1d191be1bc618ae7d3a6dcddcea07b75523d035868ceef99a506010b1da90.jpg", + "table_caption": [ + "Table 4. Results on AVA V2.2." + ], + "table_footnote": [ + "Table 5. Ablation study on K400, with and without masked backward computation and elemental residual connections." + ], + "table_body": "
ModelsInputTop-1Top-5
VideoMamba-M (baseline)32 × 224282.495.7
VideoMambaPro-M (w/o residual)32 × 224283.6 (+1.2)96.0 (+0.3)
VideoMambaPro-M (w/o masking)32 × 224283.0 (+1.0)95.8 (+0.1)
VideoMambaPro-M32 × 224284.0 (+1.6)96.4 (+0.7)
", + "bbox": [ + 94, + 441, + 480, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 575, + 240, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Influence of masked backward computation and elemental residual connection. We have identified two limitations that exist in VideoMamba: historical decay and element contradiction. We introduced masked backward computation and elemental residual connections to address these respective issue. Here, we analyze the impact of each solution. We use the same settings as before, with VideoMambaPro-M and pre-training on IN-1K. We summarize the performance of VideoMambaPro-M on K400 in Table 5. Both solutions contribute to an improved score, and their effect is partly complementary. This indicates that the two limitations exist simultaneously in VideoMamba.", + "bbox": [ + 89, + 598, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evidence of limitations of Mamba. We have theoretically identified historical decay and element contradiction as fundamental problems in Mamba, and propose the masked backward computation and elemental residual connection to solve the limitations. Previous experiments on image and video action recognition demonstrated the effectiveness of our approach, and we further provide concrete evidence that the increased performance stems from", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "addressing these two issues.", + "bbox": [ + 511, + 90, + 700, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We performed two tests on ImageNet-1K image classification task to build the connection between the issues in practice and our approach. First, to investigate whether the residual connection alleviates element contradiction, we randomly replaced a percentage of the patches by randomly selected patches from other classes, which will increase element contradiction because more irrelevant information is present. The table 6 shows that the relative performance of VideoMambaPro over VideoMamba increases to $9.2\\%$ when $20\\%$ of the patches is replaced. For higher percentages, the gap is reduced along with the overall score. This demonstrates that irrelevant (or even misleading) information can better be dealt with using our approach.", + "bbox": [ + 511, + 106, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa5c63bcd7319ecf61a9690a929f5a35edc2cfcf10d2d3250c1add5ec6540b93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Replacement ratio
051020305080
VideoMamba-Ti76.973.268.059.452.947.628.9
VideoMambaPro-Ti78.978.076.168.655.948.229.1
", + "bbox": [ + 517, + 316, + 901, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Second, to verify that masking addresses the issue of historical decay, we adopted a progressive masking approach, gradually decreasing the mask values of the diagonal elements of $\\mathbf{M}_b$ from 1 to 0, and then retrained the model on ImageNet-1K. The results below show that VideoMambaPro's accuracy gradually increases when more masking is applied, evidencing the merits of the approach.", + "bbox": [ + 511, + 426, + 905, + 532 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23987e356e9aa53af9fe33547a3b254732d1f64eec05fd8a1b04b9e966d7bb74.jpg", + "table_caption": [ + "Table 6. Top-1 accuracy of VideoMambaPro-Ti and VideoMamba-Ti with various replacement ratios (in %) on ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
Mask value
10.70.50
VideoMambaPro-Ti76.977.077.277.8
", + "bbox": [ + 576, + 545, + 846, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Top-1 accuracy across mask values on ImageNet-1K.", + "bbox": [ + 522, + 604, + 890, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Comparison with VideoMamba on K400", + "text_level": 1, + "bbox": [ + 511, + 636, + 861, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We more thoroughly compare the differences between VideoMamba and VideoMambaPro by investigating the relative performance per class. We then present a statistical comparison between the results of both backbones.", + "bbox": [ + 511, + 657, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Class analysis. We compare VideoMambaPro-M with $224^{2}$ image size pre-trained on IN-1K to a VideoMamba-M baseline with the same settings. We show the relative performance for all classes of K400 in Figure 2. For over $95\\%$ of the classes, VideoMambaPro shows improvement. Although there is a lower performance for certain classes, the decrease is typically limited.", + "bbox": [ + 511, + 718, + 903, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The majority of the classes sees a $\\sim 1.8\\%$ improvement, which is substantial. For a small number of classes, VideoMambaPro performs $>2\\%$ better than VideoMamba. Only a fraction of the classes is negatively affected by the solutions introduced in VideoMambaPro.", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24240", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28cbe6524cbac5873fbb947b49b348fef62046a69a53c8f71d72fff8b7321620.jpg", + "image_caption": [ + "Figure 2. Relative accuracy per class on Kinetics-400 by comparing VideoMambaPro-M to VideoMamba-M." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Statistical comparison. In order to understand whether the improvements of VideoMambaPro over VideoMamba are statistically significant, we compare the K400 results of the respective Middle models, both pre-trained on IN-1K and size $224 \\times 224$ . Other settings are also the same. For each test sample, we check whether it is correctly classified by either model. The results appear in Table 8.", + "bbox": [ + 89, + 335, + 483, + 443 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d0cc73a449a637818985e725ae1a5b2f9d98f128cda0929830777960ea7ac849.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VideoMambaPro-M
TrueFalseTotal
VideoMamba-MTrue14,30246914,771
False1,8333,3025,135
Total16,1353,77119,906
", + "bbox": [ + 119, + 454, + 455, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Contingency table for K400 test items for VideoMamba-M and VideoMambaPro-M.", + "bbox": [ + 89, + 560, + 482, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used the McNemar test, a non-parametric test with a single degree of freedom. It checks whether the number of items that are incorrectly classified by VideoMambaPro $(n_{10})$ but not VideoMamba is substantially lower than the number of items misclassified by VideoMamba but not VideoMambaPro $(n_{01})$ . The test is calculated as $\\chi^2 = \\frac{(n_{01} - n_{10})^2}{(n_{01} + n_{10})}$ . The resulting value of 808.2 corresponds to a significance level of $p < 0.001$ . We can thus conclude that VideoMambaPro-M is statistically significantly better than VideoMamba-M. Because we relied on the aggregated performance reported in papers for other methods, we cannot report statistical comparisons here.", + "bbox": [ + 89, + 606, + 483, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Computation cost analysis", + "text_level": 1, + "bbox": [ + 89, + 801, + 328, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, we compare the performance of VideoMambaPro with various model sizes with other approaches on K400. We map the top-1 to the number of parameters and FLOPs in Figures 3 and 4, respectively. VideoMambaPro performs favorably in terms of the accuracy-compute trade", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71a21b66d61b9f0c71b7cc6c12157524b8c07ce69f225eb685dcf1319631d141.jpg", + "image_caption": [ + "Figure 3. Top-1 accuracy versus number of parameters of VideoMambaPro and other models on Kinetics-400." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 898, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef37a2df50cef53f9627a9b71dade869f27255428064217d2bd48f7c18889190.jpg", + "image_caption": [ + "Figure 4. Top-1 accuracy versus number of FLOPs of VideoMambaPro and other models on Kinetics-400." + ], + "image_footnote": [], + "bbox": [ + 517, + 316, + 906, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "off. Importantly, VideoMambaPro was trained on much less data than other models, which might provide opportunities for further accuracy gains without additional compute and memory requirements.", + "bbox": [ + 511, + 573, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From a mathematical comparison with self-attention, we have identified two limitations in how Mamba processes token sequences. We argue that these limitations constrain Mamba's potential, especially in video understanding tasks. To address the two limitations, we have introduced VideoMambaPro (VMP). It takes VideoMamba and introduces the masked backward State Space Model (SSM), and adds residual connections in both forward and backward SSM. In experiments on Kinetics-400, Something-Something V2, HMDB51, UCF-101, and AVA V2.2, VideoMambaPro consistently demonstrates improved performance over the vanilla VideoMamba. We expect that extensive pre-training further elevates the performance of Mamba models for video tasks, making it an increasingly attractive, efficient alternative to large transformer models.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24241", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 6", + "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, pages 813-824, 2021. 6", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 5", + "[4] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2CLR: Video and image for visual contrastive learning of representation. In CVPR, pages 1502-1512, 2021. 7", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[6] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 6", + "[7] Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. NeurIPS, 34:26183-26197, 2021. 2", + "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 7", + "[9] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 35:35946-35958, 2022. 2, 7", + "[10] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"Something Something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 5", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2", + "[12] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 33:1474-1487, 2020. 2", + "[13] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2, 3", + "[14] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. AVA: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, pages 6047-6056, 2018. 6" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, pages 312-329, 2020. 7", + "[16] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 7", + "[17] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of selfattention. In International Conference on Algorithmic Learning Theory, pages 597-619. PMLR, 2023. 1", + "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. HMDB: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 6", + "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. UniformerV2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 1, 7", + "[20] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 6", + "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2", + "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 1, 2, 5", + "[23] Shufan Li, Harkanwar Singh, and Aditya Grover. Mamba-nd: Selective state space modeling for multi-dimensional data. In European Conference on Computer Vision, pages 75–92. Springer, 2024. 2", + "[24] Wenrui Li, Xiaopeng Hong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 5", + "[25] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MViTv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 6, 7", + "[26] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. SimpleClick: Interactive image segmentation with simple vision transformers. In ICCV, pages 22290-22300, 2023. 2", + "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[29] Hui Lu, Hu Jian, Ronald Poppe, and Albert Ali Salah. Enhancing video transformers for action understanding with vlm-aided training. arXiv preprint arXiv:2403.16128, 2024.1" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24242", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. NeurIPS, 34:21297-21309, 2021. 2", + "[31] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 1", + "[32] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18, 2022. 6", + "[33] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. VideoMoCo: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 7", + "[34] Jinyoung Park, Hee-Seon Kim, Kangwook Ko, Minbeom Kim, and Changick Kim. Videomamba: Spatio-temporal selective state space model. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2", + "[35] Mandela Patrick, Yuki Markus Asano, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multimodal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 7", + "[36] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 6", + "[37] Badri Narayana Patro and Vijay Srinivas Agneeswaran. Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges. arXiv preprint arXiv:2404.16112, 2024. 1", + "[38] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, et al. RWKV: Reinventing RNNs for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 1, 2", + "[39] AJ Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking video ViTs: Sparse video tubes for joint image and video learning. In CVPR, pages 2214-2224, 2023. 6", + "[40] Hongyu Ren, Hanjun Dai, Zihang Dai, Mengjiao Yang, Jure Leskovec, Dale Schuurmans, and Bo Dai. Combiner: Full attention transformer with sparse computation cost. NeurIPS, 34:22470-22482, 2021. 2", + "[41] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 1", + "[42] Gilad Sharir, Asaf Noy, and Lihi Zelnik-Manor. An image is worth 16x16 words, what is a video worth? arXiv preprint arXiv:2103.13915, 2021. 6", + "[43] Yuanyuan Shen, Edmund M-K Lai, and Mahsa Mohagheh. Effects of similarity score functions in attention mechanisms on the performance of neural question answering systems. Neural Processing Letters, 54(3):2283-2302, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.6", + "[45] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2", + "[46] Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621, 2023. 1, 2", + "[47] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 1, 2, 6, 7", + "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 6", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 2", + "[50] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 1, 6, 7", + "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6", + "[52] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. InternVideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6", + "[53] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 6", + "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 7", + "[55] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 6", + "[56] Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianxiang Chen, and Zi Ye. A survey on Visual Mamba. arXiv preprint arXiv:2404.15956, 2024. 5", + "[57] Zheng Zhang and Kil To Chong. Comparison between first-order hold with zero-order hold in discretization of input-delay nonlinear systems. In 2007 International Conference on Control, Automation and Systems, pages 2892-2896, 2007. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24243", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. ViT-YOLO: Transformer-basedYOLO for object detection. In ICCV, pages 2799-2808, 2021. 2", + "[59] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. NeurIPS, 34:17723-17736, 2021. 2", + "[60] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. BiFormer: Vision transformer with bilevel routing attention. In CVPR, pages 10323-10333, 2023. 2", + "[61] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2, 4" + ], + "bbox": [ + 91, + 92, + 480, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24244", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_model.json b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6042bbfa19e64ac032fa783e92ba6d4cc31a4be1 --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.768, + 0.153 + ], + "angle": 0, + "content": "Snakes and Ladders: Two Steps Up for VideoMamba" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.182, + 0.37, + 0.234 + ], + "angle": 0, + "content": "Hui Lu \nUtrecht University \nh.lu1@uu.nl" + }, + { + "type": "text", + "bbox": [ + 0.424, + 0.182, + 0.574, + 0.234 + ], + "angle": 0, + "content": "Albert A. Salah \nUtrecht University \na.a.salah@uu.nl" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.182, + 0.779, + 0.234 + ], + "angle": 0, + "content": "Ronald Poppe \nUtrecht University \nr.w.poppe@uu.nl" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.604 + ], + "angle": 0, + "content": "Video understanding requires the extraction of rich spatiotemporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by \\(1.6 - 3.0\\%\\) and \\(1.1 - 1.9\\%\\) top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.648, + 0.222, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Video understanding is challenging and requires models that can extract rich spatio-temporal representations from video inputs. Transformers are powerful neural networks capable of effectively capturing temporal and spatial information from videos [19, 29, 50]. Current state-of-the-art video understanding models are transformers [41, 51]. At the core of transformers is self-attention [49], the self-alignment between tokens in an input obtained by estimating the relative importance of a token with respect to all other tokens. The long-range token dependency accounts for much of the success of transformer models [5, 49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Calculating self-attention is computationally costly, which eventually limits the application of powerful transformer models in practical settings [17]. Recently, alternative models with lower-cost operators have been proposed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.905, + 0.346 + ], + "angle": 0, + "content": "for national language processing (NLP), including S4 [13], RWKV [38], and RetNet [46]. Among these, Mamba [11] shows the best performance on long-range and causal tasks such as language understanding [31] and content-based reasoning [37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Motivated by the favorable computational cost, researchers have recently extended Mamba from the NLP domain to the computer vision domain. The core adaptation involved splitting the input image into multiple regions and embedding these as continuous tokens [61]. For video understanding, the recently proposed VideoMamba [22] extracts key frames from videos as the continuous input sequence. However, compared to previous transformer-based methods, VideoMamba's performance on video benchmarks is markedly lower. For example, VideoMamba achieves \\(82.4\\%\\) top-1 on Kinetics-400, compared to \\(85.2\\%\\) for VideoMAE [47], indicating room for improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.729 + ], + "angle": 0, + "content": "In this paper, we first analyze differences in the feature extraction capabilities of transformers and Mamba. We identify two limitations of Mamba: historical decay and element contradiction. We then extend VideoMamba [22] to mitigate these limitations. The proposed VideoMambaPro (VMP) addresses historical decay through masked backward computation in the bi-directional Mamba process, allowing the network to better handle historical tokens. We introduce residual connections to Mamba's matrix elements to tackle element contradiction. VideoMambaPro consistently improves the performance of VideoMamba on video understanding tasks, positioning it as a strong, efficient competitor to transformers. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.732, + 0.905, + 0.775 + ], + "angle": 0, + "content": "- We derive a formal representation of Mamba from the perspective of self-attention and identify two limitations of Mamba in the video analysis domain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.904, + 0.806 + ], + "angle": 0, + "content": "- We propose VideoMambaPro, effectively addressing Mamba's limitations for video understanding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.808, + 0.904, + 0.851 + ], + "angle": 0, + "content": "- We report strong video action recognition performance compared to recent Vision Mamba-based models, and surpass the original VideoMamba by clear margins." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.732, + 0.905, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We first discuss related work. Then, we provide a theoretical analysis and introduce VideoMambaPro. Experiments are presented in Section 5. We conclude in Section 6." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24234" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.232, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.266 + ], + "angle": 0, + "content": "Transformers. One core aspect of transformers is self-attention [49] to achieve long-range interactions by measuring the similarity between tokens. Self-attention was introduced in the computer vision domain for tasks such as image recognition [26, 45] and object detection [7, 58]. Subsequent works (e.g., [9, 21, 47, 53] extended vision transformers to the video domain, to achieve superior performance. However, the mechanism of self-attention introduces significant computational overhead because of the similarity between all pairs of tokens needs to be calculated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.403 + ], + "angle": 0, + "content": "Alternative models. Recent work has introduced alternative models with reduced computational complexity, while maintaining the advantages of self-attention [30, 40, 59]. SOFT [30] uses Gaussian kernel functions to replace the dot-product similarity, which enables a full self-attention matrix to be approximated with a low-rank matrix decomposition. Combiner [40] employs a structured factorization to approximate full self-attention, realizing low computation and memory complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.553 + ], + "angle": 0, + "content": "Peng et al. [38] propose the Receptance Weighted Key Value (RWKV) architecture that combines self-attention training with an efficient recurrent neural network (RNN) inference using a linear attention mechanism. Through parallel computation, a lower, constant-level computational and memory complexity is achieved. RetNet [46] includes another variant of self-attention, by dividing the input into multiple chunks. Within each chunk, the self-attention mechanism can be computed in parallel, while information is transmitted between chunks based on an RNN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.705 + ], + "angle": 0, + "content": "State-space models. The S4 model completely abandons self-attention and, instead, builds upon a state space model [13]. Instead of calculating a similarity matrix by performing matrix multiplications for pairs of tokens, it enables the network to directly learn a global high-order polynomial projection operator (HiPPO) matrix to handle relations between tokens. Additionally, for the simultaneous input of multiple tokens, S4 proposes a convolutional processing approach, enabling parallel training and thereby accelerating the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Based on S4, Mamba [11] proposes a selection mechanism where, for each input token, a unique HiPPO matrix [12] is generated. This allows the model to selectively process input tokens, enabling it to focus on or ignore specific inputs. Due to Mamba's strong representation ability in NLP, and linear-time complexity, it has garnered attention as a promising alternative to transformers. In the computer vision domain, researchers have proposed Vision Mamba [61] and VMamba [27] for tasks such as image classification and object detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the video domain, several Mamba variants have been proposed [22, 23, 34]. Their performance is somewhat lower than expected, with limited understanding of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "causes. We argue that a systematic, mathematical analysis of Mamba from the perspective of self-attention could reveal shortcomings of Mamba's inner workings. Better understanding of these limitations allow us to develop improvements, and to close the accuracy performance gap with transformers, while enjoying Mamba's efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.71, + 0.212 + ], + "angle": 0, + "content": "3. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.906, + 0.266 + ], + "angle": 0, + "content": "We revisit Mamba from the perspective of self-attention. Then, we analyze its limitations for video understanding. To address these, we propose VideoMambaPro in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.896, + 0.29 + ], + "angle": 0, + "content": "3.1. Mamba from the perspective of self-attention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.906, + 0.356 + ], + "angle": 0, + "content": "Self-attention. Given an input sequence \\( \\mathbf{X} \\coloneqq [x_1, \\dots, x_N] \\in \\mathbb{R}^{N \\times D_x} \\) of \\( N \\) feature vectors of depth \\( D_x \\), self-attention [49, 60] computes the output sequence \\( \\mathbf{Y} \\) from \\( \\mathbf{X} \\) following two steps:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.906, + 0.416 + ], + "angle": 0, + "content": "Step 1: Similarity matrix computation. The input sequence \\(\\mathbf{X}\\) is linearly projected onto the three different subspaces query \\(\\mathbf{Q} \\in \\mathbb{R}^{N \\times D}\\), key \\(\\mathbf{K} \\in \\mathbb{R}^{N \\times D}\\), and value \\(\\mathbf{V} \\in \\mathbb{R}^{N \\times D_V}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.429, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\boldsymbol {X} \\mathbf {W} _ {Q} ^ {\\top}, \\mathbf {K} = \\boldsymbol {X} \\mathbf {W} _ {K} ^ {\\top}, \\mathbf {V} = \\boldsymbol {X} \\mathbf {W} _ {V} ^ {\\top}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.453, + 0.906, + 0.578 + ], + "angle": 0, + "content": "with \\(\\mathbf{W}_Q, \\mathbf{W}_K \\in \\mathbb{R}^{D \\times D_x}\\), and \\(\\mathbf{W}_V \\in \\mathbb{R}^{D_v \\times D_x}\\) the corresponding weight matrices. Specifically, \\(\\mathbf{Q} := [q_1, \\dots, q_N]^\\top\\), \\(\\mathbf{K} := [k_1, \\dots, k_N]^\\top\\), and \\(\\mathbf{V} := [v_1, \\dots, v_N]^\\top\\) with vectors \\(q_i, k_i, v_i\\) for \\(i = 1, \\dots, N\\) the query, key, and value vectors, respectively, for input \\(i\\). Based on \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\), similarity matrix \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times N}\\) contains the correlations between all query and key vectors, with a softmax function applied to each row of \\(\\mathbf{S}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.585, + 0.905, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top} / \\sqrt {D}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.904, + 0.643 + ], + "angle": 0, + "content": "Each component \\( s_{ij} (i,j = 1,\\dots ,N) \\) represents the similarity score between \\( \\pmb{q}_i \\) and \\( \\pmb{k}_j \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.643, + 0.905, + 0.675 + ], + "angle": 0, + "content": "Step 2: Output computation. Output sequence \\(\\mathbf{Y} := [y_1, \\dots, y_N]^\\top\\) is then calculated based on \\(\\mathbf{S}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.684, + 0.905, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {S V}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.709, + 0.906, + 0.738 + ], + "angle": 0, + "content": "It follows that each output vector \\(\\pmb{y}_i\\) (\\(i = 1, \\dots, N\\)) can be written in vector form as:" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.745, + 0.905, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {i} = \\sum_ {j = 1} ^ {N} s _ {i j} \\boldsymbol {v} _ {j}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Any output vector \\( \\mathbf{y}_i \\) is a linear combination of vectors \\( \\mathbf{v}_j (j = 1, \\dots, N) \\), with similarity score \\( s_{ij} \\) serving as coefficient. The larger the similarity score, the greater the influence of \\( \\mathbf{v}_j \\) on output \\( \\mathbf{y}_i \\) [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Mamba. State Space Models (SSMs) serve as the foundation of Mamba [11]. They are based on continuous systems that map 1D functions or sequences, \\( x(t) \\in \\mathbb{R}^L \\to \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24235" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.125 + ], + "angle": 0, + "content": "\\(y(t)\\in \\mathbb{R}^{L}\\) to output sequences \\(y(t)\\) through a hidden state \\(h(t)\\in \\mathbb{R}^N\\) . Formally, SSM implements the mapping as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.144, + 0.483, + 0.16 + ], + "angle": 0, + "content": "\\[\nh (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.163, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\ny (t) = \\mathbf {C} h (t) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.485, + 0.324 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\) is the evolution matrix of the system, and \\(\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}\\), \\(\\mathbf{C} \\in \\mathbb{R}^{N \\times 1}\\) are the projection matrices. Often, the input is discrete rather than a continuous function \\(x(t)\\). Therefore, Mamba performs discretization, effectively creating a discrete version of the continuous system. A timescale parameter \\(\\Delta\\) is used to transform the continuous parameters \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) into their discrete counterparts \\(\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}\\), and the transformation typically employs the zero-order hold method [57]. This process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.329, + 0.483, + 0.346 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.349, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} \\left(\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}\\right) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.37, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.391, + 0.482, + 0.407 + ], + "angle": 0, + "content": "\\[\ny _ {t} = \\mathbf {C h} _ {t}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.485, + 0.642 + ], + "angle": 0, + "content": "Considering that parameters \\(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}\\) in the original SSM are independent of the input data \\(x(t)\\) and cannot be tailored to specific input data, Mamba employs a Selective Scan Mechanism as its core operator. More precisely, three functions \\(S_B(x),S_C(x),S_\\Delta (x)\\) are introduced to associate parameters \\(\\overline{\\mathbf{B}}\\) , C, and \\(\\pmb{\\Delta}\\) in Eqs. 7-10 to the input data \\(x\\) Based on \\(S_{\\Delta}(x)\\) \\(\\overline{\\mathbf{A}}\\) can also be associated with the input data \\(x\\) . For example, given the input \\(x_{1}\\) , functions \\(S_{\\Delta}(x)\\) will produce the corresponding \\(\\overline{\\mathbf{A}}_1\\) based on Eq. 7, and functions \\(S_B(x)\\) and \\(S_{\\Delta}(x)\\) will produce the corresponding \\(\\overline{\\mathbf{B}}_1\\) based on Eq. 8. \\(\\overline{\\mathbf{C}}_1\\) is obtained based on function \\(S_{C}(x)\\) . Following Eqs. 9 and 10, we analyze the process to obtain output sequence \\(\\mathbf{Y}\\) when given an input sequence \\(\\mathbf{X}:= [x_1,\\dots ,x_N]\\in \\mathbb{R}^{N\\times D_x}\\) of \\(N\\) feature vectors. Each vector's hidden state is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.663, + 0.483, + 0.68 + ], + "angle": 0, + "content": "\\[\nh _ {1} = \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.684, + 0.482, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h _ {2} = \\overline {{\\mathbf {A}}} _ {2} h _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ = \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2}, \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.725, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h _ {3} = \\overline {{\\mathbf {A}}} _ {3} h _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3} \\\\ = \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {I} + \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3}, \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.769, + 0.146, + 0.777 + ], + "angle": 0, + "content": "." + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.785, + 0.482, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h _ {N} = \\overline {{\\mathbf {A}}} _ {N} h _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N} \\\\ = \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {B}}} _ {N - 1} \\boldsymbol {x} _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N}. \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.485, + 0.902 + ], + "angle": 0, + "content": "1The original SSM [13] employs \\( h'(t) = Ah(t) + Bx(t) \\), with \\( h(t) \\) the hidden state from previous time step \\( t - 1 \\), and \\( h'(t) \\) the updated current hidden state, replacing \\( h(t) \\). Considering this approach may lead to ambiguity, we have adopted the updated description." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.813, + 0.107 + ], + "angle": 0, + "content": "Eqs. 11-14 can be written in matrix form:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.117, + 0.905, + 0.229 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {H} = \\left[ h _ {1}, h _ {2}, h _ {3}, \\dots , h _ {N} \\right] ^ {\\top} \\\\ = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {1} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right]. \\tag {15} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.229, + 0.905, + 0.261 + ], + "angle": 0, + "content": "For output sequence \\(\\mathbf{Y} := [\\pmb{y}_1, \\dots, \\pmb{y}_N]^\\top\\), each vector \\(\\pmb{y}_i\\) (\\(i = 1, \\dots, N\\)) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.271, + 0.905, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {N} = \\overline {{\\mathbf {C}}} _ {N} h _ {N}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.3, + 0.664, + 0.314 + ], + "angle": 0, + "content": "and in matrix form as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.323, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {C}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ 0 & \\overline {{\\mathbf {C}}} _ {2} & 0 & \\dots & 0 \\\\ 0 & 0 & \\overline {{\\mathbf {C}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ 0 & 0 & 0 & \\dots & \\overline {{\\mathbf {C}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\\\ \\vdots \\\\ h _ {N} \\end{array} \\right] = \\overline {{\\mathbf {C}}} h. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.422, + 0.838, + 0.437 + ], + "angle": 0, + "content": "By substituting Eq. 15 into Eq. 17, we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.447, + 0.905, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\overline {{\\mathbf {C}}} \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {I} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right], \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.534, + 0.695, + 0.548 + ], + "angle": 0, + "content": "which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.559, + 0.905, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\overline {{\\mathbf {C}}} (\\mathbf {M} \\mathbf {X}), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.905, + 0.635 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) represents the second term on the right-hand side of Eq. 18. Recall from Eq. 3 that the result \\(\\mathbf{Y}\\) obtained by self-attention processing can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.644, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = \\mathbf {S V} = (\\mathbf {S X}) \\mathbf {W} _ {V} ^ {\\top} \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.793 + ], + "angle": 0, + "content": "From the perspective of self-attention, by comparing Eqs. 19 and 20, the essence of Mamba is to generate a matrix \\(\\mathbf{M}\\) similar to similarity matrix \\(\\mathbf{S}\\), such that the result of \\(\\mathbf{M}\\mathbf{X}\\) is based on the correlation between vectors of \\(\\mathbf{X}\\). Although the final result of \\(\\mathbf{M}\\mathbf{X}\\) is left multiplied by a mapping matrix \\(\\mathbf{C}\\), while the result of \\(\\mathbf{S}\\mathbf{X}\\) is right multiplied by a mapping matrix \\(\\mathbf{W}_V^\\top\\), the geometric meaning of the two are the same." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.905, + 0.82 + ], + "angle": 0, + "content": "3.2. Limitations of Mamba in video understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "From the perspective of self-attention, the concept of Mamba is similar: both use similarity matrices. We now analyze the differences between the similarity matrices of Mamba and self-attention, and discuss the limitations of Mamba in the context of the video understanding task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24236" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "Limitation 1: Historical decay. Matrix M in Eq. 19 corresponds to the second right-hand term in Eq. 18, which is a lower triangular matrix of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.145, + 0.484, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\left[ \\begin{array}{c c c c c} m _ {1 1} & 0 & 0 & \\dots & 0 \\\\ m _ {2 1} & m _ {2 2} & 0 & \\dots & 0 \\\\ m _ {3 1} & m _ {3 2} & m _ {3 3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {N 1} & m _ {N 2} & m _ {N 3} & \\dots & m _ {N N} \\end{array} \\right]. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.375 + ], + "angle": 0, + "content": "By comparing \\(\\mathbf{M}\\) with matrix \\(\\mathbf{S}\\) in self-attention, we find that outputs in Mamba favor more recent information, because the more weights are zero, the earlier the token is observed. For example, for input \\([x_1, x_2, x_3]\\), the output \\(\\mathbf{M}x_1\\) in Mamba is \\(m_{11}x_1\\) while the output \\(\\mathbf{S}x_1\\) is \\(s_{11}x_1 + s_{12}x_2 + s_{13}x_3\\) in self-attention. This indicates that, in Mamba, the influence of earlier observed tokens on the final result is greatly diminished. We refer to this limitation as historical decay." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.512 + ], + "angle": 0, + "content": "In the NLP domain, more recent dialogue information often has more impact on the final judgment, so this effect is acceptable. However, in the computer vision domain, the order of the tokens has less meaning. Previous works such as Vision Mamba [61] and VMamba [27] tried to solve this issue by processing the token sequence in both forward and backward directions. This produces better results, but significant deficiencies still exist. We will explain this theoretically below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.617 + ], + "angle": 0, + "content": "When processing bi-directionally, the results generated from input forward tokens \\(\\left[x_1,\\dots ,x_N\\right]\\), denoted as \\(\\mathbf{M}_f\\mathbf{X}\\) and the results generated from input backward tokens \\(\\left[x_{N},\\dots ,x_{1}\\right]\\), denoted as \\(\\mathbf{M}_b\\mathbf{X}\\), are linearly combined to generate the final result \\(\\mathbf{M}_{bi}\\mathbf{X}\\) with \\(\\mathbf{M}_{bi}\\) a dense matrix. As a result, the influence of historical information on the result is increased, consequently leading to better results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.483, + 0.648 + ], + "angle": 0, + "content": "For example, for the input tokens \\([x_1, x_2, x_3]\\), \\(\\mathbf{M}_f\\mathbf{X}\\) and \\(\\mathbf{M}_b\\mathbf{X}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.655, + 0.484, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {f} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} f _ {1 1} & 0 & 0 \\\\ f _ {2 1} & f _ {2 2} & 0 \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {1 f} \\\\ h _ {2 f} \\\\ h _ {3 f} \\end{array} \\right], \\tag {22}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.724, + 0.484, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {b} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} b _ {3 3} & 0 & 0 \\\\ b _ {2 3} & b _ {2 2} & 0 \\\\ b _ {1 3} & b _ {1 2} & b _ {1 1} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {3} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {1} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {3 b} \\\\ h _ {2 b} \\\\ h _ {1 b} \\end{array} \\right], \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.843 + ], + "angle": 0, + "content": "where \\( f_{ij} \\) represents the similarity score during the forward process, and \\( b_{ij} \\) is the similarity score in the backward direction. After bi-directional computation, with the outputs linearly combined, the results are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.851, + 0.443, + 0.868 + ], + "angle": 0, + "content": "\\[\nh _ {1} = h _ {1 f} + h _ {1 b} = f _ {1 1} \\boldsymbol {x} _ {1} + b _ {1 3} \\boldsymbol {x} _ {3} + b _ {1 2} \\boldsymbol {x} _ {2} + b _ {1 1} \\boldsymbol {x} _ {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.869, + 0.483, + 0.885 + ], + "angle": 0, + "content": "\\[\nh _ {2} = h _ {2 f} + h _ {2 b} = f _ {2 1} \\boldsymbol {x} _ {1} + f _ {2 2} \\boldsymbol {x} _ {2} + b _ {2 3} \\boldsymbol {x} _ {3} + b _ {2 2} \\boldsymbol {x} _ {2} \\tag {24}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.888, + 0.445, + 0.904 + ], + "angle": 0, + "content": "\\[\nh _ {3} = h _ {3 f} + h _ {3 b} = f _ {3 1} \\boldsymbol {x} _ {1} + f _ {3 2} \\boldsymbol {x} _ {2} + f _ {3 3} \\boldsymbol {x} _ {3} + b _ {3 3} \\boldsymbol {x} _ {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.775, + 0.107 + ], + "angle": 0, + "content": "We can write Eq. 24 in matrix form:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.117, + 0.905, + 0.216 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\end{array} \\right] = \\left[ \\begin{array}{c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] \\tag {25} \\\\ = \\mathbf {M} _ {b i} \\left[ \\begin{array}{c} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.907, + 0.303 + ], + "angle": 0, + "content": "The bi-directional computation transforms matrix \\(\\mathbf{M}\\) from a lower triangular matrix to a dense matrix \\(\\mathbf{M}_{bi}\\), thereby capturing more historical information and effectively avoiding historical decay. When extending to the case of \\(N\\) input tokens \\([\\pmb{x}_1,\\dots ,\\pmb{x}_N]\\), \\(\\mathbf{M}_{bi}\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.312, + 0.906, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} + b _ {N N} \\end{array} \\right]. \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.587 + ], + "angle": 0, + "content": "The diagonal elements of \\(\\mathbf{M}_{bi}\\) contain duplicates of the similarity between a token and itself. For example, \\(f_{33}\\) and \\(b_{33}\\) each represent the similarity between token \\(x_{3}\\) and itself. Consequently, the similarity is effectively doubled which weakens the association with other tokens. One possible approach is to adjust \\(\\mathbf{M}_f\\) and \\(\\mathbf{M}_b\\) using a weight coefficient \\(z\\) through a linear combination. However, this method is not very effective. For example, to prevent the similarity of diagonal elements from being doubled, the \\(z\\) value is set to 0.5. However, after multiplying other elements by 0.5, their weight values are significantly reduced, which means that the association of each token with other tokens is weakened." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.588, + 0.906, + 0.619 + ], + "angle": 0, + "content": "Limitation 2: Element contradiction. By analyzing non-zero elements \\( m_{ij} \\) in \\( \\mathbf{M} \\) of Eq. 18, we derive:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.63, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\nm _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.825 + ], + "angle": 0, + "content": "After multiple iterations, the above equation results in implicit consideration of the correlation between previous tokens and token \\( j \\) when computing the correlation between token \\( i \\) and token \\( j \\). As a result, \\( m_{ij} \\) exhibits stronger contextual dependencies compared to the elements \\( s_{ij} \\) in the matrix S. This might explain why Mamba achieves better performance than transformers in the field of NLP. While this is advantageous in the NLP domain, for the computer vision domain, input tokens often lack semantic connections. The consideration of the influence of other tokens on each element can lead to significant drawbacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Interleaved token structures are common when processing images. Tokens that \"belong together\" might not be subsequently processed. For example, in an image classification task, input tokens \\(\\left[\\pmb{x}_1,\\pmb{x}_2,\\pmb{x}_3\\right]\\) might represent image regions [dog, other, dog]. Ideally, \\(m_{31}\\) should be high and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24237" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.581, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.318, + 0.419, + 0.332 + ], + "angle": 0, + "content": "(a) VideoMambaPro" + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.09, + 0.865, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.318, + 0.826, + 0.332 + ], + "angle": 0, + "content": "(b) Bi-Directional Mamba Block" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 1. (a) Framework of VideoMambaPro with \\( K \\) bi-directional Mamba blocks. (b) In each bi-directional Mamba block, we employ forward residual SSM and masked backward Residual SSM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.484, + 0.534 + ], + "angle": 0, + "content": "\\(m_{21}\\) low. Following Eq. 27, \\(m_{31} = \\overline{\\mathbf{A}}_3 m_{21}\\), so \\(\\overline{\\mathbf{A}}_3\\) needs to increase. However, this causes \\(m_{32} = \\overline{\\mathbf{A}}_3 m_{22}\\) to increase because \\(m_{22}\\) is also high. But, theoretically, \\(m_{32}\\) should be low. This causes an element contradiction. Especially for video understanding, such contradictions are common because most video regions contain background and other irrelevant information, making relevant tokens sparse. Consequently, the performance of Mamba applied to video analysis tasks is underwhelming [22, 24, 56]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.262, + 0.563 + ], + "angle": 0, + "content": "4. VideoMambaPro" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.646 + ], + "angle": 0, + "content": "We use VideoMamba [22] as backbone, and propose two adaptations to address historical decay and element contradiction. The resulting architecture is termed VideoMambaPro (VMP). With minor adjustments, our adaptations can also be applied to related Mamba models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.752 + ], + "angle": 0, + "content": "To address historical decay, we keep the result of \\(\\mathbf{M}_f\\mathbf{X}\\) but we use masked computation in the backward process. Specifically, we assign a mask to the diagonal elements of \\(\\mathbf{M}_b\\), setting their values to 0, and then proceed with the calculations in Eqs 21-25. We thus eliminate the duplicate similarity on the diagonal, without affecting other elements. The final \\(\\mathbf{M}_{bi}\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.761, + 0.483, + 0.849 + ], + "angle": 0, + "content": "\\[\nM _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} \\end{array} \\right]. \\tag {28}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To solve element contradiction, we propose residual SSM. This solution is inspired by residual connections, to distribute the requirement for \\(\\overline{\\mathbf{A}}_i\\) in \\(m_{ij}\\) over multiple \\(\\overline{\\mathbf{A}}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.489 + ], + "angle": 0, + "content": "This helps to avoid contradictions caused by interleaved sequence structures. For example, for our previous example input sequence \\([x_1, x_2, x_3]\\), which represents regions [dog, other, dog], we let \\(m_{31} = \\overline{\\mathbf{A}}_3 m_{21} + \\overline{\\mathbf{A}}_3\\). This way, the requirement for a single \\(\\overline{\\mathbf{A}}_3\\) can be split into two parts, thus avoiding contradictions. This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.496, + 0.906, + 0.515 + ], + "angle": 0, + "content": "\\[\nm _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} + \\overline {{\\mathbf {A}}} _ {i} \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.907, + 0.658 + ], + "angle": 0, + "content": "We implement these solutions into VideoMamba [22], to form VideoMambaPro (see Figure 1). Given input video \\( \\pmb{X}^v \\in \\mathbb{R}^{3 \\times T \\times H \\times W} \\), we first use a 3D convolution with a \\( 1 \\times 16 \\times 16 \\) kernel to convert \\( \\pmb{X}^v \\) into \\( L \\) non-overlapping patch-wise tokens \\( \\pmb{X}^p \\in \\mathbb{R}^{L \\times C} \\) with \\( L = t \\times h \\times w \\) (\\( t = T, h = \\frac{H}{16}, w = \\frac{W}{16} \\)). Because SSM is sensitive to token positions, and in line with [22], we include learnable spatial and temporal position embeddings \\( \\pmb{p}_s \\in \\mathbb{R}^{(hw + 1) \\times C} \\) and \\( \\pmb{p}_t \\in \\mathbb{R}^{t \\times C} \\). Input tokens \\( \\pmb{X} \\) are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.667, + 0.905, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} = \\left[ \\boldsymbol {X} _ {\\text {c l s}}, \\boldsymbol {X} \\right] + \\boldsymbol {p} _ {s} + \\boldsymbol {p} _ {t}, \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.906, + 0.752 + ], + "angle": 0, + "content": "where \\( X_{cls} \\) is a learnable classification token positioned at the start of the sequence. Input tokens \\( X \\) pass through \\( K \\) Mamba blocks, and the final layer's [CLS] token is used for classification, after normalization and linear projection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.649, + 0.781 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.702, + 0.805 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate VideoMambaPro on five video benchmarks: (a) Kinetics-400 (K400, [3]) comprises \\(\\sim 240\\mathrm{K}\\) training and \\(\\sim 20\\mathrm{K}\\) validation videos, each with an average duration of 10 seconds and categorized into 400 classes. (b) Something-Something V2 (SSv2, [10]) includes \\(\\sim 160\\mathrm{K}\\) training and \\(\\sim 20\\mathrm{K}\\) validation videos with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24238" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "an average duration of 4 seconds, and 174 motion-centric classes. (c) UCF-101 [44] is a relatively small dataset, consisting of \\(\\sim 9.5\\mathrm{K}\\) training and \\(\\sim 3.5\\mathrm{K}\\) validation videos. (d) HMDB51 [18] is also a compact video dataset, containing \\(\\sim 3.5\\mathrm{K}\\) training and \\(\\sim 1.5\\mathrm{K}\\) validation videos. (e) AVA [14] is a dataset for spatio-temporal localization of human actions with \\(\\sim 211\\mathrm{k}\\) and \\(\\sim 57\\mathrm{k}\\) validation video segments." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.487, + 0.518 + ], + "angle": 0, + "content": "Implementation. In line with VideoMamba, we introduce three models with increasing embedding dimension and number of bi-directional Mamba blocks \\( K \\): Tiny, Small, and Middle (details in supplementary material). To compare with VideoMamba, we pre-train VideoMambaPro on ImageNet-1K (IN-1K). On K400, we also pre-train with IN-1K, fine-tune on the training set and report on the validation set. For K400, we also report on the larger \\( 336^2 \\) input size. During pre-training, we follow DeiT [48] by applying a center crop to obtain the \\( 224^2 \\) sized images. We apply random cropping, random horizontal flipping, label-smoothing regularization, mix-up, and random erasing as data augmentations. We use AdamW [28] with a momentum of 0.9, a batch size of 1024, and a weight decay of 0.05. We employ a cosine learning rate schedule during training, \\( 1 \\times 10^{-3} \\) initial learning rate over 300 epochs. The fine-tuning settings follow VideoMAE [47]. We resize frames to \\( 224^2 \\), and use AdamW with a momentum of 0.9 and a batch size of 512. The evaluation process is the same as VideoMamba, and more details are in the supplementary materials. These materials also include results on IN-1K image classification." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.534, + 0.482, + 0.803 + ], + "angle": 0, + "content": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]32 × 22425×137M350G80.294.4
MViTv2-S [25]16 × 22425×135M320G81.094.6
Uniformer-S [20]IN-1K16 × 22424×121M168G80.894.7
Uniformer-B [20]IN-1K16 × 22424×150M388G82.095.1
Uniformer-B [20]IN-1K32 × 22424×350M3.1T83.095.4
STAM [42]IN-21K64 × 22421×1121M1.0T79.2-
TimeSformer-L [2]IN-21K96 × 22421×3121M7.1T80.794.7
ViViT-L [1]IN-21K16 × 22424×3311M47.9T81.394.7
Mformer-HR [36]IN-21K16 × 336210×3311M28.8T81.195.2
VideoMAE-H [47]IN-21K16 × 22425×3633M17.9T86.697.1
X-CLIP-L/14 [32]CLIP-400M16 × 33624×3453M37.0T87.7-
MTV-H [55]60M132 × 22424×31120M44.5T89.198.2
InternVideo-1B [51]412M264 × 224216×41300M86.2T91.198.9
InternVideo-2B [52]414M316 × 224216×41000M-91.6-
InternVideo-2B-6B [52]414M316 × 224216×45903M-92.1-
VideoMamba-TiIN-1K32 × 22424×37M0.4T78.893.9
VideoMamba-TiIN-1K64 × 38424×37M2.4T80.394.8
VideoMamba-SIN-1K32 × 22424×326M1.6T81.595.2
VideoMamba-SIN-1K64 × 38424×326M4.7T82.795.6
VideoMamba-MIN-1K32 × 22424×374M4.8T82.495.7
VideoMamba-MIN-1K64 × 38424×374M28.4T83.396.1
VideoMambaPro-TiIN-1K32 × 22424×37M0.4T81.695.9
VideoMambaPro-TiIN-1K64 × 38424×37M2.2T83.396.1
VideoMambaPro-SIN-1K32 × 22424×325M1.6T83.396.0
VideoMambaPro-SIN-1K64 × 38424×325M4.4T84.596.6
VideoMambaPro-MIN-1K32 × 22424×372M4.7T84.096.4
VideoMambaPro-MIN-1K64 × 38424×372M27.0T85.096.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.814, + 0.482, + 0.854 + ], + "angle": 0, + "content": "Table 1. Performance on K400. Top part of the table are Transformer models, bottom part are Mamba models. We report crops (temporal \\(\\times\\) spatial) and FLOPs for inference. —: not reported." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.856, + 0.192, + 0.868 + ], + "angle": 0, + "content": "\\(^{1}\\) IN-21K+WTS" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.869, + 0.461, + 0.882 + ], + "angle": 0, + "content": "2 CLIP-400M+WebVid+HowTo+K710+SSv2+AVA2.2+more." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.883, + 0.458, + 0.896 + ], + "angle": 0, + "content": "3 LAION-300M+KMash+WebVid+InternVid+LLaVA+more." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.856, + 0.461, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.805, + 0.108 + ], + "angle": 0, + "content": "5.2. Comparison with state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.119, + 0.907, + 0.302 + ], + "angle": 0, + "content": "K400. Results appear in Table 1. Compared to VideoMamba, VideoMambaPro has slightly fewer parameters and FLOPs. This is primarily because VideoMamba employs an additional projection layer to generate the weight coefficient \\( z \\) to adjust \\( \\mathbf{A}_f \\) and \\( \\mathbf{A}_b \\). See the supplementary materials for an architecture comparison. VideoMambaPro outperforms VideoMamba across model and input sizes. With \\( 224^2 \\) inputs and pre-trained only on IN-1K, the best-performing VideoMambaPro-M achieves a top-1 accuracy of \\( 84.0\\% \\), \\( 1.6\\% \\) higher than VideoMamba-M. Further comparisons appear in Section 5.4. Increasing the input size to \\( 336^2 \\) leads to a performance improvement of \\( 1.0 - 1.7\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.907, + 0.368 + ], + "angle": 0, + "content": "VideoMambaPro scores lower than the recent InternVideo2-1B [52] by \\(7.6\\%\\), but was only pre-trained on IN-1K and has significantly fewer parameters (1000M vs 72M) and inference only takes \\(\\sim 5.5\\%\\) of the FLOPs." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.393, + 0.905, + 0.6 + ], + "angle": 0, + "content": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]K40016 × 22421×337M213G64.789.2
MViTv1-B [6]K40032 × 22421×337M510G67.190.8
MViTv2-S [25]K40016 × 22421×335M195G68.291.4
MViTv2-B [25]K40032 × 22421×351M675G70.592.7
Uniformer-S [20]IN-1K+K40016 × 22421×321M126G67.791.4
Uniformer-B [20]IN-1K+K40016 × 22421×350M291G70.492.8
TimeSformer-L [2]IN-21K16 × 22421×3121M5.1T62.5-
ViViT-L [1]IN-21K+K40016 × 22424×3311M47.9T65.489.8
Mformer-HR [36]IN-21K+K40016 × 33621×3311M3.6T68.191.2
MaskFeat-L [53]IN-21K64 × 31224×3218M8.5T75.095.0
VideoMAE-L [47]IN-21K32 × 22421×3305M4.3T75.495.2
TubeViT-L [39]IN-1K32 × 22424×3311M9.5T76.195.2
InternVideo-1B [51]See Table 164 × 224216×41300M86.2T77.295.9
InternVideo2-1B [52]See Table 164 × 224216×41000M-77.1-
InternVideo2-6B [52]See Table 164 × 224216×45903M-77.4-
VideoMamba-TiIN-1K16 × 22422×37M102G66.089.6
VideoMamba-SIN-1K16 × 22422×326M408G67.690.9
VideoMamba-MIN-1K16 × 22424×374M2.4T68.391.4
VideoMambaPro-TiIN-1K16 × 22422×37M96G67.991.2
VideoMambaPro-SIN-1K16 × 22422×325M382G68.891.4
VideoMambaPro-MIN-1K16 × 22424×372M2.2T69.491.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.61, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Table 2. Performance on SSv2. —: not reported. Top part of the table are Transformer models, bottom part are Mamba models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.909, + 0.79 + ], + "angle": 0, + "content": "SSv2. Results appear in Table 2. VideoMambaPro outperforms VideoMamba by \\(1.1 - 1.9\\%\\). It also outperforms several popular transformer models. Although InternVideo-1B [51] and InternVideo2-6B [52] outperform our VideoMambaPro-M by \\(7.8\\%\\) and \\(8.0\\%\\), respectively, they require 18.0-82 times more parameters and at least 39 times more FLOPs. Again, we expect that the performance for VideoMambaPro will increase with more pre-training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "UCF-101/HMDB51/AVA V2.2. From Table 3, it shows that VideoMambaPro-M is competitive, and outperforms VideoMamba by \\(3.4\\%\\) and \\(1.8\\%\\) on UCF-101 and HMDB51, respectively. VideoMambaPro-M achieves 31.9 mAP on AVA V2.2, which is \\(10.7\\%\\) lower than VideoMAE V2 [50] but with an order of magnitude fewer parameters and FLOPs and pre-trained only on IN-1K (see Table 4)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24239" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.442, + 0.226 + ], + "angle": 0, + "content": "
MethodParamsUCF-101HMDB51
VideoMoCo [33]15M78.749.2
CoCLR [16]9M81.452.1
MemDPC [15]32M86.154.5
Vi²CLR [4]9M89.155.7
VideoMAE [47]87M91.362.6
GDT [35]33M95.272.8
VideoMAE V2 [50]1050M99.688.1
VideoMamba-M74M88.260.8
VideoMambaPro-M72M91.663.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.237, + 0.42, + 0.25 + ], + "angle": 0, + "content": "Table 3. Results on UCF-101 and HMDB51." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.265, + 0.434, + 0.405 + ], + "angle": 0, + "content": "
MethodFLOPsParammAP
SlowFast R101 [8]138G53M23.8
VideoMAE-B [47]180G87M26.7
MViTv2-B [25]225G51M30.5
ObjectTransformer [54]243G86M31.0
MViTv2-L [25]2828G213M34.4
ST-MAE-H [9]1193G632M36.2
VideoMAE V2 [50]4220G1050M42.6
VideoMamba-M [19]202G74M30.1
VideoMambaPro-M183G72M31.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.415, + 0.38, + 0.428 + ], + "angle": 0, + "content": "Table 4. Results on AVA V2.2." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.443, + 0.481, + 0.513 + ], + "angle": 0, + "content": "
ModelsInputTop-1Top-5
VideoMamba-M (baseline)32 × 224282.495.7
VideoMambaPro-M (w/o residual)32 × 224283.6 (+1.2)96.0 (+0.3)
VideoMambaPro-M (w/o masking)32 × 224283.0 (+1.0)95.8 (+0.1)
VideoMambaPro-M32 × 224284.0 (+1.6)96.4 (+0.7)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.522, + 0.483, + 0.551 + ], + "angle": 0, + "content": "Table 5. Ablation study on K400, with and without masked backward computation and elemental residual connections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.241, + 0.594 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Influence of masked backward computation and elemental residual connection. We have identified two limitations that exist in VideoMamba: historical decay and element contradiction. We introduced masked backward computation and elemental residual connections to address these respective issue. Here, we analyze the impact of each solution. We use the same settings as before, with VideoMambaPro-M and pre-training on IN-1K. We summarize the performance of VideoMambaPro-M on K400 in Table 5. Both solutions contribute to an improved score, and their effect is partly complementary. This indicates that the two limitations exist simultaneously in VideoMamba." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evidence of limitations of Mamba. We have theoretically identified historical decay and element contradiction as fundamental problems in Mamba, and propose the masked backward computation and elemental residual connection to solve the limitations. Previous experiments on image and video action recognition demonstrated the effectiveness of our approach, and we further provide concrete evidence that the increased performance stems from" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.702, + 0.106 + ], + "angle": 0, + "content": "addressing these two issues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.304 + ], + "angle": 0, + "content": "We performed two tests on ImageNet-1K image classification task to build the connection between the issues in practice and our approach. First, to investigate whether the residual connection alleviates element contradiction, we randomly replaced a percentage of the patches by randomly selected patches from other classes, which will increase element contradiction because more irrelevant information is present. The table 6 shows that the relative performance of VideoMambaPro over VideoMamba increases to \\(9.2\\%\\) when \\(20\\%\\) of the patches is replaced. For higher percentages, the gap is reduced along with the overall score. This demonstrates that irrelevant (or even misleading) information can better be dealt with using our approach." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.317, + 0.903, + 0.382 + ], + "angle": 0, + "content": "
Replacement ratio
051020305080
VideoMamba-Ti76.973.268.059.452.947.628.9
VideoMambaPro-Ti78.978.076.168.655.948.229.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.392, + 0.905, + 0.422 + ], + "angle": 0, + "content": "Table 6. Top-1 accuracy of VideoMambaPro-Ti and VideoMamba-Ti with various replacement ratios (in %) on ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Second, to verify that masking addresses the issue of historical decay, we adopted a progressive masking approach, gradually decreasing the mask values of the diagonal elements of \\(\\mathbf{M}_b\\) from 1 to 0, and then retrained the model on ImageNet-1K. The results below show that VideoMambaPro's accuracy gradually increases when more masking is applied, evidencing the merits of the approach." + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.546, + 0.847, + 0.596 + ], + "angle": 0, + "content": "
Mask value
10.70.50
VideoMambaPro-Ti76.977.077.277.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.605, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Table 7. Top-1 accuracy across mask values on ImageNet-1K." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.862, + 0.653 + ], + "angle": 0, + "content": "5.4. Comparison with VideoMamba on K400" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.718 + ], + "angle": 0, + "content": "We more thoroughly compare the differences between VideoMamba and VideoMambaPro by investigating the relative performance per class. We then present a statistical comparison between the results of both backbones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Class analysis. We compare VideoMambaPro-M with \\(224^{2}\\) image size pre-trained on IN-1K to a VideoMamba-M baseline with the same settings. We show the relative performance for all classes of K400 in Figure 2. For over \\(95\\%\\) of the classes, VideoMambaPro shows improvement. Although there is a lower performance for certain classes, the decrease is typically limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The majority of the classes sees a \\(\\sim 1.8\\%\\) improvement, which is substantial. For a small number of classes, VideoMambaPro performs \\(>2\\%\\) better than VideoMamba. Only a fraction of the classes is negatively affected by the solutions introduced in VideoMambaPro." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.309 + ], + "angle": 0, + "content": "Figure 2. Relative accuracy per class on Kinetics-400 by comparing VideoMambaPro-M to VideoMamba-M." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.444 + ], + "angle": 0, + "content": "Statistical comparison. In order to understand whether the improvements of VideoMambaPro over VideoMamba are statistically significant, we compare the K400 results of the respective Middle models, both pre-trained on IN-1K and size \\(224 \\times 224\\). Other settings are also the same. For each test sample, we check whether it is correctly classified by either model. The results appear in Table 8." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.455, + 0.456, + 0.551 + ], + "angle": 0, + "content": "
VideoMambaPro-M
TrueFalseTotal
VideoMamba-MTrue14,30246914,771
False1,8333,3025,135
Total16,1353,77119,906
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Table 8. Contingency table for K400 test items for VideoMamba-M and VideoMambaPro-M." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.792 + ], + "angle": 0, + "content": "We used the McNemar test, a non-parametric test with a single degree of freedom. It checks whether the number of items that are incorrectly classified by VideoMambaPro \\((n_{10})\\) but not VideoMamba is substantially lower than the number of items misclassified by VideoMamba but not VideoMambaPro \\((n_{01})\\). The test is calculated as \\(\\chi^2 = \\frac{(n_{01} - n_{10})^2}{(n_{01} + n_{10})}\\). The resulting value of 808.2 corresponds to a significance level of \\(p < 0.001\\). We can thus conclude that VideoMambaPro-M is statistically significantly better than VideoMamba-M. Because we relied on the aggregated performance reported in papers for other methods, we cannot report statistical comparisons here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.33, + 0.82 + ], + "angle": 0, + "content": "5.5. Computation cost analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Finally, we compare the performance of VideoMambaPro with various model sizes with other approaches on K400. We map the top-1 to the number of parameters and FLOPs in Figures 3 and 4, respectively. VideoMambaPro performs favorably in terms of the accuracy-compute trade" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.9, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.301 + ], + "angle": 0, + "content": "Figure 3. Top-1 accuracy versus number of parameters of VideoMambaPro and other models on Kinetics-400." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.318, + 0.908, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.519, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Figure 4. Top-1 accuracy versus number of FLOPs of VideoMambaPro and other models on Kinetics-400." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.635 + ], + "angle": 0, + "content": "off. Importantly, VideoMambaPro was trained on much less data than other models, which might provide opportunities for further accuracy gains without additional compute and memory requirements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "From a mathematical comparison with self-attention, we have identified two limitations in how Mamba processes token sequences. We argue that these limitations constrain Mamba's potential, especially in video understanding tasks. To address the two limitations, we have introduced VideoMambaPro (VMP). It takes VideoMamba and introduces the masked backward State Space Model (SSM), and adds residual connections in both forward and backward SSM. In experiments on Kinetics-400, Something-Something V2, HMDB51, UCF-101, and AVA V2.2, VideoMambaPro consistently demonstrates improved performance over the vanilla VideoMamba. We expect that extensive pre-training further elevates the performance of Mamba models for video tasks, making it an increasingly attractive, efficient alternative to large transformer models." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24241" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, pages 813-824, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.243 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[4] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2CLR: Video and image for visual contrastive learning of representation. In CVPR, pages 1502-1512, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[6] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[7] Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. NeurIPS, 34:26183-26197, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 35:35946-35958, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[10] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"Something Something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 33:1474-1487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[13] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. AVA: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, pages 6047-6056, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, pages 312-329, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[16] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[17] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of selfattention. In International Conference on Algorithmic Learning Theory, pages 597-619. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. HMDB: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. UniformerV2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[20] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[23] Shufan Li, Harkanwar Singh, and Aditya Grover. Mamba-nd: Selective state space modeling for multi-dimensional data. In European Conference on Computer Vision, pages 75–92. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[24] Wenrui Li, Xiaopeng Hong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[25] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MViTv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. SimpleClick: Interactive image segmentation with simple vision transformers. In ICCV, pages 22290-22300, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[29] Hui Lu, Hu Jian, Ronald Poppe, and Albert Ali Salah. Enhancing video transformers for action understanding with vlm-aided training. arXiv preprint arXiv:2403.16128, 2024.1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24242" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. NeurIPS, 34:21297-21309, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[31] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[32] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[33] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. VideoMoCo: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[34] Jinyoung Park, Hee-Seon Kim, Kangwook Ko, Minbeom Kim, and Changick Kim. Videomamba: Spatio-temporal selective state space model. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[35] Mandela Patrick, Yuki Markus Asano, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multimodal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[36] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[37] Badri Narayana Patro and Vijay Srinivas Agneeswaran. Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges. arXiv preprint arXiv:2404.16112, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[38] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, et al. RWKV: Reinventing RNNs for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[39] AJ Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking video ViTs: Sparse video tubes for joint image and video learning. In CVPR, pages 2214-2224, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[40] Hongyu Ren, Hanjun Dai, Zihang Dai, Mengjiao Yang, Jure Leskovec, Dale Schuurmans, and Bo Dai. Combiner: Full attention transformer with sparse computation cost. NeurIPS, 34:22470-22482, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[41] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Gilad Sharir, Asaf Noy, and Lihi Zelnik-Manor. An image is worth 16x16 words, what is a video worth? arXiv preprint arXiv:2103.13915, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Yuanyuan Shen, Edmund M-K Lai, and Mahsa Mohagheh. Effects of similarity score functions in attention mechanisms on the performance of neural question answering systems. Neural Processing Letters, 54(3):2283-2302, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[45] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[46] Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[47] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[50] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[52] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. InternVideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[53] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[55] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[56] Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianxiang Chen, and Zi Ye. A survey on Visual Mamba. arXiv preprint arXiv:2404.15956, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[57] Zheng Zhang and Kil To Chong. Comparison between first-order hold with zero-order hold in discretization of input-delay nonlinear systems. In 2007 International Conference on Control, Automation and Systems, pages 2892-2896, 2007. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24243" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[58] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. ViT-YOLO: Transformer-basedYOLO for object detection. In ICCV, pages 2799-2808, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[59] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. NeurIPS, 34:17723-17736, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[60] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. BiFormer: Vision transformer with bilevel routing attention. In CVPR, pages 10323-10333, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[61] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24244" + } + ] +] \ No newline at end of file diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_origin.pdf b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e89abd2bb64ba054f64190f1d787037231fb25c9 --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/b11c3116-c63e-4aac-8755-cf4eabafe782_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4214c44f84795ccac4fb69d4b8f4a83891931cd70722a406417787869a3953b0 +size 993872 diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/full.md b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/full.md new file mode 100644 index 0000000000000000000000000000000000000000..29b0121b2807ccfebe02ba976fff83aa8731c42e --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/full.md @@ -0,0 +1,441 @@ +# Snakes and Ladders: Two Steps Up for VideoMamba + +Hui Lu +Utrecht University +h.lu1@uu.nl + +Albert A. Salah +Utrecht University +a.a.salah@uu.nl + +Ronald Poppe +Utrecht University +r.w.poppe@uu.nl + +# Abstract + +Video understanding requires the extraction of rich spatiotemporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by $1.6 - 3.0\%$ and $1.1 - 1.9\%$ top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models. + +# 1. Introduction + +Video understanding is challenging and requires models that can extract rich spatio-temporal representations from video inputs. Transformers are powerful neural networks capable of effectively capturing temporal and spatial information from videos [19, 29, 50]. Current state-of-the-art video understanding models are transformers [41, 51]. At the core of transformers is self-attention [49], the self-alignment between tokens in an input obtained by estimating the relative importance of a token with respect to all other tokens. The long-range token dependency accounts for much of the success of transformer models [5, 49]. + +Calculating self-attention is computationally costly, which eventually limits the application of powerful transformer models in practical settings [17]. Recently, alternative models with lower-cost operators have been proposed + +for national language processing (NLP), including S4 [13], RWKV [38], and RetNet [46]. Among these, Mamba [11] shows the best performance on long-range and causal tasks such as language understanding [31] and content-based reasoning [37]. + +Motivated by the favorable computational cost, researchers have recently extended Mamba from the NLP domain to the computer vision domain. The core adaptation involved splitting the input image into multiple regions and embedding these as continuous tokens [61]. For video understanding, the recently proposed VideoMamba [22] extracts key frames from videos as the continuous input sequence. However, compared to previous transformer-based methods, VideoMamba's performance on video benchmarks is markedly lower. For example, VideoMamba achieves $82.4\%$ top-1 on Kinetics-400, compared to $85.2\%$ for VideoMAE [47], indicating room for improvement. + +In this paper, we first analyze differences in the feature extraction capabilities of transformers and Mamba. We identify two limitations of Mamba: historical decay and element contradiction. We then extend VideoMamba [22] to mitigate these limitations. The proposed VideoMambaPro (VMP) addresses historical decay through masked backward computation in the bi-directional Mamba process, allowing the network to better handle historical tokens. We introduce residual connections to Mamba's matrix elements to tackle element contradiction. VideoMambaPro consistently improves the performance of VideoMamba on video understanding tasks, positioning it as a strong, efficient competitor to transformers. Our contributions are: + +- We derive a formal representation of Mamba from the perspective of self-attention and identify two limitations of Mamba in the video analysis domain. +- We propose VideoMambaPro, effectively addressing Mamba's limitations for video understanding. +- We report strong video action recognition performance compared to recent Vision Mamba-based models, and surpass the original VideoMamba by clear margins. + +We first discuss related work. Then, we provide a theoretical analysis and introduce VideoMambaPro. Experiments are presented in Section 5. We conclude in Section 6. + +# 2. Related Work + +Transformers. One core aspect of transformers is self-attention [49] to achieve long-range interactions by measuring the similarity between tokens. Self-attention was introduced in the computer vision domain for tasks such as image recognition [26, 45] and object detection [7, 58]. Subsequent works (e.g., [9, 21, 47, 53] extended vision transformers to the video domain, to achieve superior performance. However, the mechanism of self-attention introduces significant computational overhead because of the similarity between all pairs of tokens needs to be calculated. + +Alternative models. Recent work has introduced alternative models with reduced computational complexity, while maintaining the advantages of self-attention [30, 40, 59]. SOFT [30] uses Gaussian kernel functions to replace the dot-product similarity, which enables a full self-attention matrix to be approximated with a low-rank matrix decomposition. Combiner [40] employs a structured factorization to approximate full self-attention, realizing low computation and memory complexity. + +Peng et al. [38] propose the Receptance Weighted Key Value (RWKV) architecture that combines self-attention training with an efficient recurrent neural network (RNN) inference using a linear attention mechanism. Through parallel computation, a lower, constant-level computational and memory complexity is achieved. RetNet [46] includes another variant of self-attention, by dividing the input into multiple chunks. Within each chunk, the self-attention mechanism can be computed in parallel, while information is transmitted between chunks based on an RNN. + +State-space models. The S4 model completely abandons self-attention and, instead, builds upon a state space model [13]. Instead of calculating a similarity matrix by performing matrix multiplications for pairs of tokens, it enables the network to directly learn a global high-order polynomial projection operator (HiPPO) matrix to handle relations between tokens. Additionally, for the simultaneous input of multiple tokens, S4 proposes a convolutional processing approach, enabling parallel training and thereby accelerating the training process. + +Based on S4, Mamba [11] proposes a selection mechanism where, for each input token, a unique HiPPO matrix [12] is generated. This allows the model to selectively process input tokens, enabling it to focus on or ignore specific inputs. Due to Mamba's strong representation ability in NLP, and linear-time complexity, it has garnered attention as a promising alternative to transformers. In the computer vision domain, researchers have proposed Vision Mamba [61] and VMamba [27] for tasks such as image classification and object detection. + +In the video domain, several Mamba variants have been proposed [22, 23, 34]. Their performance is somewhat lower than expected, with limited understanding of the + +causes. We argue that a systematic, mathematical analysis of Mamba from the perspective of self-attention could reveal shortcomings of Mamba's inner workings. Better understanding of these limitations allow us to develop improvements, and to close the accuracy performance gap with transformers, while enjoying Mamba's efficiency. + +# 3. Theoretical Analysis + +We revisit Mamba from the perspective of self-attention. Then, we analyze its limitations for video understanding. To address these, we propose VideoMambaPro in Section 4. + +# 3.1. Mamba from the perspective of self-attention + +Self-attention. Given an input sequence $\mathbf{X} \coloneqq [x_1, \dots, x_N] \in \mathbb{R}^{N \times D_x}$ of $N$ feature vectors of depth $D_x$ , self-attention [49, 60] computes the output sequence $\mathbf{Y}$ from $\mathbf{X}$ following two steps: + +Step 1: Similarity matrix computation. The input sequence $\mathbf{X}$ is linearly projected onto the three different subspaces query $\mathbf{Q} \in \mathbb{R}^{N \times D}$ , key $\mathbf{K} \in \mathbb{R}^{N \times D}$ , and value $\mathbf{V} \in \mathbb{R}^{N \times D_V}$ : + +$$ +\mathbf {Q} = \boldsymbol {X} \mathbf {W} _ {Q} ^ {\top}, \mathbf {K} = \boldsymbol {X} \mathbf {W} _ {K} ^ {\top}, \mathbf {V} = \boldsymbol {X} \mathbf {W} _ {V} ^ {\top}. \tag {1} +$$ + +with $\mathbf{W}_Q, \mathbf{W}_K \in \mathbb{R}^{D \times D_x}$ , and $\mathbf{W}_V \in \mathbb{R}^{D_v \times D_x}$ the corresponding weight matrices. Specifically, $\mathbf{Q} := [q_1, \dots, q_N]^\top$ , $\mathbf{K} := [k_1, \dots, k_N]^\top$ , and $\mathbf{V} := [v_1, \dots, v_N]^\top$ with vectors $q_i, k_i, v_i$ for $i = 1, \dots, N$ the query, key, and value vectors, respectively, for input $i$ . Based on $\mathbf{Q}$ and $\mathbf{K}$ , similarity matrix $\mathbf{S} \in \mathbb{R}^{N \times N}$ contains the correlations between all query and key vectors, with a softmax function applied to each row of $\mathbf{S}$ : + +$$ +\mathbf {S} = \operatorname {s o f t m a x} \left(\mathbf {Q} \mathbf {K} ^ {\top} / \sqrt {D}\right). \tag {2} +$$ + +Each component $s_{ij} (i,j = 1,\dots ,N)$ represents the similarity score between $\pmb{q}_i$ and $\pmb{k}_j$ . + +Step 2: Output computation. Output sequence $\mathbf{Y} := [y_1, \dots, y_N]^\top$ is then calculated based on $\mathbf{S}$ as: + +$$ +\mathbf {Y} = \mathbf {S V}. \tag {3} +$$ + +It follows that each output vector $\pmb{y}_i$ ( $i = 1, \dots, N$ ) can be written in vector form as: + +$$ +\boldsymbol {y} _ {i} = \sum_ {j = 1} ^ {N} s _ {i j} \boldsymbol {v} _ {j}. \tag {4} +$$ + +Any output vector $\mathbf{y}_i$ is a linear combination of vectors $\mathbf{v}_j (j = 1, \dots, N)$ , with similarity score $s_{ij}$ serving as coefficient. The larger the similarity score, the greater the influence of $\mathbf{v}_j$ on output $\mathbf{y}_i$ [43]. + +Mamba. State Space Models (SSMs) serve as the foundation of Mamba [11]. They are based on continuous systems that map 1D functions or sequences, $x(t) \in \mathbb{R}^L \to$ + +$y(t)\in \mathbb{R}^{L}$ to output sequences $y(t)$ through a hidden state $h(t)\in \mathbb{R}^N$ . Formally, SSM implements the mapping as: + +$$ +h (t) = \mathbf {A} h (t - 1) + \mathbf {B} x (t) \tag {5} +$$ + +$$ +y (t) = \mathbf {C} h (t) \tag {6} +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ is the evolution matrix of the system, and $\mathbf{B} \in \mathbb{R}^{N \times 1}$ , $\mathbf{C} \in \mathbb{R}^{N \times 1}$ are the projection matrices. Often, the input is discrete rather than a continuous function $x(t)$ . Therefore, Mamba performs discretization, effectively creating a discrete version of the continuous system. A timescale parameter $\Delta$ is used to transform the continuous parameters $\mathbf{A}$ and $\mathbf{B}$ into their discrete counterparts $\overline{\mathbf{A}}, \overline{\mathbf{B}}$ , and the transformation typically employs the zero-order hold method [57]. This process is expressed as: + +$$ +\overline {{\mathbf {A}}} = \exp (\Delta \mathbf {A}), \tag {7} +$$ + +$$ +\overline {{\mathbf {B}}} = (\boldsymbol {\Delta} \mathbf {A}) ^ {- 1} \left(\exp (\boldsymbol {\Delta} \mathbf {A}) - \mathbf {I}\right) \cdot \boldsymbol {\Delta} \mathbf {B}, \tag {8} +$$ + +$$ +h _ {t} = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \tag {9} +$$ + +$$ +y _ {t} = \mathbf {C h} _ {t}. \tag {10} +$$ + +Considering that parameters $\overline{\mathbf{A}},\overline{\mathbf{B}},\mathbf{C}$ in the original SSM are independent of the input data $x(t)$ and cannot be tailored to specific input data, Mamba employs a Selective Scan Mechanism as its core operator. More precisely, three functions $S_B(x),S_C(x),S_\Delta (x)$ are introduced to associate parameters $\overline{\mathbf{B}}$ , C, and $\pmb{\Delta}$ in Eqs. 7-10 to the input data $x$ Based on $S_{\Delta}(x)$ $\overline{\mathbf{A}}$ can also be associated with the input data $x$ . For example, given the input $x_{1}$ , functions $S_{\Delta}(x)$ will produce the corresponding $\overline{\mathbf{A}}_1$ based on Eq. 7, and functions $S_B(x)$ and $S_{\Delta}(x)$ will produce the corresponding $\overline{\mathbf{B}}_1$ based on Eq. 8. $\overline{\mathbf{C}}_1$ is obtained based on function $S_{C}(x)$ . Following Eqs. 9 and 10, we analyze the process to obtain output sequence $\mathbf{Y}$ when given an input sequence $\mathbf{X}:= [x_1,\dots ,x_N]\in \mathbb{R}^{N\times D_x}$ of $N$ feature vectors. Each vector's hidden state is denoted as: + +$$ +h _ {1} = \overline {{\mathbf {B}}} _ {1} \boldsymbol {x} _ {1}, \tag {11} +$$ + +$$ +\begin{array}{l} h _ {2} = \overline {{\mathbf {A}}} _ {2} h _ {1} + \overline {{\mathbf {B}}} _ {2} \boldsymbol {x} _ {2} \\ = \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} \boldsymbol {x} _ {1} + \overline {{\mathbf {B}}} _ {2} \boldsymbol {x} _ {2}, \tag {12} \\ \end{array} +$$ + +$$ +\begin{array}{l} h _ {3} = \overline {{\mathbf {A}}} _ {3} h _ {2} + \overline {{\mathbf {B}}} _ {3} \boldsymbol {x} _ {3} \\ = \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} \boldsymbol {x} _ {I} + \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {B}}} _ {2} \boldsymbol {x} _ {2} + \overline {{\mathbf {B}}} _ {3} \boldsymbol {x} _ {3}, \tag {13} \\ \end{array} +$$ + +. + +$$ +\begin{array}{l} h _ {N} = \overline {{\mathbf {A}}} _ {N} h _ {N - 1} + \overline {{\mathbf {B}}} _ {N} \boldsymbol {x} _ {N} \\ = \overline {{\mathbf {A}}} _ {N} \overline {{\mathbf {A}}} _ {N - 1} \dots \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} \boldsymbol {x} _ {1} + \overline {{\mathbf {A}}} _ {N} \overline {{\mathbf {A}}} _ {N - 1} \dots \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {B}}} _ {2} \boldsymbol {x} _ {2} \\ + \overline {{\mathbf {A}}} _ {N} \overline {{\mathbf {B}}} _ {N - 1} \boldsymbol {x} _ {N - 1} + \overline {{\mathbf {B}}} _ {N} \boldsymbol {x} _ {N}. \tag {14} \\ \end{array} +$$ + +Eqs. 11-14 can be written in matrix form: + +$$ +\begin{array}{l} \mathbf {H} = \left[ h _ {1}, h _ {2}, h _ {3}, \dots , h _ {N} \right] ^ {\top} \\ = \left[ \begin{array}{c c c c c} \overline {{\mathbf {B}}} _ {1} & 0 & 0 & \dots & 0 \\ \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} & \overline {{\mathbf {B}}} _ {2} & 0 & \dots & 0 \\ \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} & \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {B}}} _ {2} & \overline {{\mathbf {B}}} _ {3} & \dots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ (\prod_ {\mathrm {j} = \mathrm {N}} ^ {2} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {1} & (\prod_ {\mathrm {j} = \mathrm {N}} ^ {3} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {2} & (\prod_ {\mathrm {j} = \mathrm {N}} ^ {4} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {3} & \dots & \overline {{\mathbf {B}}} _ {N} \end{array} \right] \left[ \begin{array}{l} x _ {1} \\ x _ {2} \\ x _ {3} \\ \vdots \\ x _ {N} \end{array} \right]. \tag {15} \\ \end{array} +$$ + +For output sequence $\mathbf{Y} := [\pmb{y}_1, \dots, \pmb{y}_N]^\top$ , each vector $\pmb{y}_i$ ( $i = 1, \dots, N$ ) can be expressed as: + +$$ +\boldsymbol {y} _ {N} = \overline {{\mathbf {C}}} _ {N} h _ {N}, \tag {16} +$$ + +and in matrix form as: + +$$ +\mathbf {Y} = \left[ \begin{array}{c c c c c} \overline {{\mathbf {C}}} _ {1} & 0 & 0 & \dots & 0 \\ 0 & \overline {{\mathbf {C}}} _ {2} & 0 & \dots & 0 \\ 0 & 0 & \overline {{\mathbf {C}}} _ {3} & \dots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ 0 & 0 & 0 & \dots & \overline {{\mathbf {C}}} _ {N} \end{array} \right] \left[ \begin{array}{l} h _ {1} \\ h _ {2} \\ h _ {3} \\ \vdots \\ h _ {N} \end{array} \right] = \overline {{\mathbf {C}}} h. \tag {17} +$$ + +By substituting Eq. 15 into Eq. 17, we obtain: + +$$ +\mathbf {Y} = \overline {{\mathbf {C}}} \left[ \begin{array}{c c c c c} \overline {{\mathbf {B}}} _ {1} & 0 & 0 & \dots & 0 \\ \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} & \overline {{\mathbf {B}}} _ {2} & 0 & \dots & 0 \\ \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {A}}} _ {2} \overline {{\mathbf {B}}} _ {1} & \overline {{\mathbf {A}}} _ {3} \overline {{\mathbf {B}}} _ {2} & \overline {{\mathbf {B}}} _ {3} & \dots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ (\prod_ {\mathrm {j} = \mathrm {N}} ^ {2} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {1} & (\prod_ {\mathrm {j} = \mathrm {N}} ^ {3} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {2} & (\prod_ {\mathrm {j} = \mathrm {N}} ^ {4} \overline {{\mathbf {A}}} _ {\mathrm {j}}) \overline {{\mathbf {B}}} _ {3} & \dots & \overline {{\mathbf {B}}} _ {N} \end{array} \right] \left[ \begin{array}{l} x _ {I} \\ x _ {2} \\ x _ {3} \\ \vdots \\ x _ {N} \end{array} \right], \tag {18} +$$ + +which can be expressed as: + +$$ +\mathbf {Y} = \overline {{\mathbf {C}}} (\mathbf {M} \mathbf {X}), \tag {19} +$$ + +where $\mathbf{M}$ represents the second term on the right-hand side of Eq. 18. Recall from Eq. 3 that the result $\mathbf{Y}$ obtained by self-attention processing can be expressed as: + +$$ +\mathbf {Y} = \mathbf {S V} = (\mathbf {S X}) \mathbf {W} _ {V} ^ {\top} \tag {20} +$$ + +From the perspective of self-attention, by comparing Eqs. 19 and 20, the essence of Mamba is to generate a matrix $\mathbf{M}$ similar to similarity matrix $\mathbf{S}$ , such that the result of $\mathbf{M}\mathbf{X}$ is based on the correlation between vectors of $\mathbf{X}$ . Although the final result of $\mathbf{M}\mathbf{X}$ is left multiplied by a mapping matrix $\mathbf{C}$ , while the result of $\mathbf{S}\mathbf{X}$ is right multiplied by a mapping matrix $\mathbf{W}_V^\top$ , the geometric meaning of the two are the same. + +# 3.2. Limitations of Mamba in video understanding + +From the perspective of self-attention, the concept of Mamba is similar: both use similarity matrices. We now analyze the differences between the similarity matrices of Mamba and self-attention, and discuss the limitations of Mamba in the context of the video understanding task. + +Limitation 1: Historical decay. Matrix M in Eq. 19 corresponds to the second right-hand term in Eq. 18, which is a lower triangular matrix of the form: + +$$ +\mathbf {M} = \left[ \begin{array}{c c c c c} m _ {1 1} & 0 & 0 & \dots & 0 \\ m _ {2 1} & m _ {2 2} & 0 & \dots & 0 \\ m _ {3 1} & m _ {3 2} & m _ {3 3} & \dots & 0 \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ m _ {N 1} & m _ {N 2} & m _ {N 3} & \dots & m _ {N N} \end{array} \right]. \tag {21} +$$ + +By comparing $\mathbf{M}$ with matrix $\mathbf{S}$ in self-attention, we find that outputs in Mamba favor more recent information, because the more weights are zero, the earlier the token is observed. For example, for input $[x_1, x_2, x_3]$ , the output $\mathbf{M}x_1$ in Mamba is $m_{11}x_1$ while the output $\mathbf{S}x_1$ is $s_{11}x_1 + s_{12}x_2 + s_{13}x_3$ in self-attention. This indicates that, in Mamba, the influence of earlier observed tokens on the final result is greatly diminished. We refer to this limitation as historical decay. + +In the NLP domain, more recent dialogue information often has more impact on the final judgment, so this effect is acceptable. However, in the computer vision domain, the order of the tokens has less meaning. Previous works such as Vision Mamba [61] and VMamba [27] tried to solve this issue by processing the token sequence in both forward and backward directions. This produces better results, but significant deficiencies still exist. We will explain this theoretically below. + +When processing bi-directionally, the results generated from input forward tokens $\left[x_1,\dots ,x_N\right]$ , denoted as $\mathbf{M}_f\mathbf{X}$ and the results generated from input backward tokens $\left[x_{N},\dots ,x_{1}\right]$ , denoted as $\mathbf{M}_b\mathbf{X}$ , are linearly combined to generate the final result $\mathbf{M}_{bi}\mathbf{X}$ with $\mathbf{M}_{bi}$ a dense matrix. As a result, the influence of historical information on the result is increased, consequently leading to better results. + +For example, for the input tokens $[x_1, x_2, x_3]$ , $\mathbf{M}_f\mathbf{X}$ and $\mathbf{M}_b\mathbf{X}$ can be expressed as: + +$$ +\mathbf {M} _ {f} \boldsymbol {X} = \left[ \begin{array}{c c c} f _ {1 1} & 0 & 0 \\ f _ {2 1} & f _ {2 2} & 0 \\ f _ {3 1} & f _ {3 2} & f _ {3 3} \end{array} \right] \left[ \begin{array}{l} \boldsymbol {x} _ {1} \\ \boldsymbol {x} _ {2} \\ \boldsymbol {x} _ {3} \end{array} \right] = \left[ \begin{array}{l} h _ {1 f} \\ h _ {2 f} \\ h _ {3 f} \end{array} \right], \tag {22} +$$ + +$$ +\mathbf {M} _ {b} \boldsymbol {X} = \left[ \begin{array}{c c c} b _ {3 3} & 0 & 0 \\ b _ {2 3} & b _ {2 2} & 0 \\ b _ {1 3} & b _ {1 2} & b _ {1 1} \end{array} \right] \left[ \begin{array}{l} \boldsymbol {x} _ {3} \\ \boldsymbol {x} _ {2} \\ \boldsymbol {x} _ {1} \end{array} \right] = \left[ \begin{array}{l} h _ {3 b} \\ h _ {2 b} \\ h _ {1 b} \end{array} \right], \tag {23} +$$ + +where $f_{ij}$ represents the similarity score during the forward process, and $b_{ij}$ is the similarity score in the backward direction. After bi-directional computation, with the outputs linearly combined, the results are expressed as: + +$$ +h _ {1} = h _ {1 f} + h _ {1 b} = f _ {1 1} \boldsymbol {x} _ {1} + b _ {1 3} \boldsymbol {x} _ {3} + b _ {1 2} \boldsymbol {x} _ {2} + b _ {1 1} \boldsymbol {x} _ {1} +$$ + +$$ +h _ {2} = h _ {2 f} + h _ {2 b} = f _ {2 1} \boldsymbol {x} _ {1} + f _ {2 2} \boldsymbol {x} _ {2} + b _ {2 3} \boldsymbol {x} _ {3} + b _ {2 2} \boldsymbol {x} _ {2} \tag {24} +$$ + +$$ +h _ {3} = h _ {3 f} + h _ {3 b} = f _ {3 1} \boldsymbol {x} _ {1} + f _ {3 2} \boldsymbol {x} _ {2} + f _ {3 3} \boldsymbol {x} _ {3} + b _ {3 3} \boldsymbol {x} _ {3} +$$ + +We can write Eq. 24 in matrix form: + +$$ +\begin{array}{l} \left[ \begin{array}{l} h _ {1} \\ h _ {2} \\ h _ {3} \end{array} \right] = \left[ \begin{array}{c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} \\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} \\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} \end{array} \right] \left[ \begin{array}{l} \boldsymbol {x} _ {1} \\ \boldsymbol {x} _ {2} \\ \boldsymbol {x} _ {3} \end{array} \right] \tag {25} \\ = \mathbf {M} _ {b i} \left[ \begin{array}{c} \boldsymbol {x} _ {1} \\ \boldsymbol {x} _ {2} \\ \boldsymbol {x} _ {3} \end{array} \right]. \\ \end{array} +$$ + +The bi-directional computation transforms matrix $\mathbf{M}$ from a lower triangular matrix to a dense matrix $\mathbf{M}_{bi}$ , thereby capturing more historical information and effectively avoiding historical decay. When extending to the case of $N$ input tokens $[\pmb{x}_1,\dots ,\pmb{x}_N]$ , $\mathbf{M}_{bi}$ can be written as: + +$$ +\mathbf {M} _ {b i} = \left[ \begin{array}{c c c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} & \dots & b _ {1 N} \\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} & \dots & b _ {2 N} \\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} & \dots & b _ {3 N} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \dots & f _ {N N} + b _ {N N} \end{array} \right]. \tag {26} +$$ + +The diagonal elements of $\mathbf{M}_{bi}$ contain duplicates of the similarity between a token and itself. For example, $f_{33}$ and $b_{33}$ each represent the similarity between token $x_{3}$ and itself. Consequently, the similarity is effectively doubled which weakens the association with other tokens. One possible approach is to adjust $\mathbf{M}_f$ and $\mathbf{M}_b$ using a weight coefficient $z$ through a linear combination. However, this method is not very effective. For example, to prevent the similarity of diagonal elements from being doubled, the $z$ value is set to 0.5. However, after multiplying other elements by 0.5, their weight values are significantly reduced, which means that the association of each token with other tokens is weakened. + +Limitation 2: Element contradiction. By analyzing non-zero elements $m_{ij}$ in $\mathbf{M}$ of Eq. 18, we derive: + +$$ +m _ {i j} = \overline {{\mathbf {A}}} _ {i} m _ {i - 1, j} \tag {27} +$$ + +After multiple iterations, the above equation results in implicit consideration of the correlation between previous tokens and token $j$ when computing the correlation between token $i$ and token $j$ . As a result, $m_{ij}$ exhibits stronger contextual dependencies compared to the elements $s_{ij}$ in the matrix S. This might explain why Mamba achieves better performance than transformers in the field of NLP. While this is advantageous in the NLP domain, for the computer vision domain, input tokens often lack semantic connections. The consideration of the influence of other tokens on each element can lead to significant drawbacks. + +Interleaved token structures are common when processing images. Tokens that "belong together" might not be subsequently processed. For example, in an image classification task, input tokens $\left[\pmb{x}_1,\pmb{x}_2,\pmb{x}_3\right]$ might represent image regions [dog, other, dog]. Ideally, $m_{31}$ should be high and + +![](images/ba5a91a49994f4e3e836d8a9169fa896086e5500ab54a7d793597e9edf1ea360.jpg) +(a) VideoMambaPro + +![](images/3f59be29a79303a6ae787aed361d4bf21fdb4bb8b47d06148eb1469d74093c07.jpg) +(b) Bi-Directional Mamba Block +Figure 1. (a) Framework of VideoMambaPro with $K$ bi-directional Mamba blocks. (b) In each bi-directional Mamba block, we employ forward residual SSM and masked backward Residual SSM. + +$m_{21}$ low. Following Eq. 27, $m_{31} = \overline{\mathbf{A}}_3 m_{21}$ , so $\overline{\mathbf{A}}_3$ needs to increase. However, this causes $m_{32} = \overline{\mathbf{A}}_3 m_{22}$ to increase because $m_{22}$ is also high. But, theoretically, $m_{32}$ should be low. This causes an element contradiction. Especially for video understanding, such contradictions are common because most video regions contain background and other irrelevant information, making relevant tokens sparse. Consequently, the performance of Mamba applied to video analysis tasks is underwhelming [22, 24, 56]. + +# 4. VideoMambaPro + +We use VideoMamba [22] as backbone, and propose two adaptations to address historical decay and element contradiction. The resulting architecture is termed VideoMambaPro (VMP). With minor adjustments, our adaptations can also be applied to related Mamba models. + +To address historical decay, we keep the result of $\mathbf{M}_f\mathbf{X}$ but we use masked computation in the backward process. Specifically, we assign a mask to the diagonal elements of $\mathbf{M}_b$ , setting their values to 0, and then proceed with the calculations in Eqs 21-25. We thus eliminate the duplicate similarity on the diagonal, without affecting other elements. The final $\mathbf{M}_{bi}$ is expressed as: + +$$ +M _ {b i} = \left[ \begin{array}{c c c c c} f _ {1 1} & b _ {1 2} & b _ {1 3} & \dots & b _ {1 N} \\ f _ {2 1} & f _ {2 2} & b _ {2 3} & \dots & b _ {2 N} \\ f _ {3 1} & f _ {3 2} & f _ {3 3} & \dots & b _ {3 N} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \dots & f _ {N N} \end{array} \right]. \tag {28} +$$ + +To solve element contradiction, we propose residual SSM. This solution is inspired by residual connections, to distribute the requirement for $\overline{\mathbf{A}}_i$ in $m_{ij}$ over multiple $\overline{\mathbf{A}}_i$ . + +This helps to avoid contradictions caused by interleaved sequence structures. For example, for our previous example input sequence $[x_1, x_2, x_3]$ , which represents regions [dog, other, dog], we let $m_{31} = \overline{\mathbf{A}}_3 m_{21} + \overline{\mathbf{A}}_3$ . This way, the requirement for a single $\overline{\mathbf{A}}_3$ can be split into two parts, thus avoiding contradictions. This can be expressed as: + +$$ +m _ {i j} = \overline {{\mathbf {A}}} _ {i} m _ {i - 1, j} + \overline {{\mathbf {A}}} _ {i} \tag {29} +$$ + +We implement these solutions into VideoMamba [22], to form VideoMambaPro (see Figure 1). Given input video $\pmb{X}^v \in \mathbb{R}^{3 \times T \times H \times W}$ , we first use a 3D convolution with a $1 \times 16 \times 16$ kernel to convert $\pmb{X}^v$ into $L$ non-overlapping patch-wise tokens $\pmb{X}^p \in \mathbb{R}^{L \times C}$ with $L = t \times h \times w$ ( $t = T, h = \frac{H}{16}, w = \frac{W}{16}$ ). Because SSM is sensitive to token positions, and in line with [22], we include learnable spatial and temporal position embeddings $\pmb{p}_s \in \mathbb{R}^{(hw + 1) \times C}$ and $\pmb{p}_t \in \mathbb{R}^{t \times C}$ . Input tokens $\pmb{X}$ are expressed as: + +$$ +\boldsymbol {X} = \left[ \boldsymbol {X} _ {\text {c l s}}, \boldsymbol {X} \right] + \boldsymbol {p} _ {s} + \boldsymbol {p} _ {t}, \tag {30} +$$ + +where $X_{cls}$ is a learnable classification token positioned at the start of the sequence. Input tokens $X$ pass through $K$ Mamba blocks, and the final layer's [CLS] token is used for classification, after normalization and linear projection. + +# 5. Experiments + +# 5.1. Experimental setup + +Datasets. We evaluate VideoMambaPro on five video benchmarks: (a) Kinetics-400 (K400, [3]) comprises $\sim 240\mathrm{K}$ training and $\sim 20\mathrm{K}$ validation videos, each with an average duration of 10 seconds and categorized into 400 classes. (b) Something-Something V2 (SSv2, [10]) includes $\sim 160\mathrm{K}$ training and $\sim 20\mathrm{K}$ validation videos with + +an average duration of 4 seconds, and 174 motion-centric classes. (c) UCF-101 [44] is a relatively small dataset, consisting of $\sim 9.5\mathrm{K}$ training and $\sim 3.5\mathrm{K}$ validation videos. (d) HMDB51 [18] is also a compact video dataset, containing $\sim 3.5\mathrm{K}$ training and $\sim 1.5\mathrm{K}$ validation videos. (e) AVA [14] is a dataset for spatio-temporal localization of human actions with $\sim 211\mathrm{k}$ and $\sim 57\mathrm{k}$ validation video segments. + +Implementation. In line with VideoMamba, we introduce three models with increasing embedding dimension and number of bi-directional Mamba blocks $K$ : Tiny, Small, and Middle (details in supplementary material). To compare with VideoMamba, we pre-train VideoMambaPro on ImageNet-1K (IN-1K). On K400, we also pre-train with IN-1K, fine-tune on the training set and report on the validation set. For K400, we also report on the larger $336^2$ input size. During pre-training, we follow DeiT [48] by applying a center crop to obtain the $224^2$ sized images. We apply random cropping, random horizontal flipping, label-smoothing regularization, mix-up, and random erasing as data augmentations. We use AdamW [28] with a momentum of 0.9, a batch size of 1024, and a weight decay of 0.05. We employ a cosine learning rate schedule during training, $1 \times 10^{-3}$ initial learning rate over 300 epochs. The fine-tuning settings follow VideoMAE [47]. We resize frames to $224^2$ , and use AdamW with a momentum of 0.9 and a batch size of 512. The evaluation process is the same as VideoMamba, and more details are in the supplementary materials. These materials also include results on IN-1K image classification. + +
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]32 × 22425×137M350G80.294.4
MViTv2-S [25]16 × 22425×135M320G81.094.6
Uniformer-S [20]IN-1K16 × 22424×121M168G80.894.7
Uniformer-B [20]IN-1K16 × 22424×150M388G82.095.1
Uniformer-B [20]IN-1K32 × 22424×350M3.1T83.095.4
STAM [42]IN-21K64 × 22421×1121M1.0T79.2-
TimeSformer-L [2]IN-21K96 × 22421×3121M7.1T80.794.7
ViViT-L [1]IN-21K16 × 22424×3311M47.9T81.394.7
Mformer-HR [36]IN-21K16 × 336210×3311M28.8T81.195.2
VideoMAE-H [47]IN-21K16 × 22425×3633M17.9T86.697.1
X-CLIP-L/14 [32]CLIP-400M16 × 33624×3453M37.0T87.7-
MTV-H [55]60M132 × 22424×31120M44.5T89.198.2
InternVideo-1B [51]412M264 × 224216×41300M86.2T91.198.9
InternVideo-2B [52]414M316 × 224216×41000M-91.6-
InternVideo-2B-6B [52]414M316 × 224216×45903M-92.1-
VideoMamba-TiIN-1K32 × 22424×37M0.4T78.893.9
VideoMamba-TiIN-1K64 × 38424×37M2.4T80.394.8
VideoMamba-SIN-1K32 × 22424×326M1.6T81.595.2
VideoMamba-SIN-1K64 × 38424×326M4.7T82.795.6
VideoMamba-MIN-1K32 × 22424×374M4.8T82.495.7
VideoMamba-MIN-1K64 × 38424×374M28.4T83.396.1
VideoMambaPro-TiIN-1K32 × 22424×37M0.4T81.695.9
VideoMambaPro-TiIN-1K64 × 38424×37M2.2T83.396.1
VideoMambaPro-SIN-1K32 × 22424×325M1.6T83.396.0
VideoMambaPro-SIN-1K64 × 38424×325M4.4T84.596.6
VideoMambaPro-MIN-1K32 × 22424×372M4.7T84.096.4
VideoMambaPro-MIN-1K64 × 38424×372M27.0T85.096.7
+ +$^{1}$ IN-21K+WTS +2 CLIP-400M+WebVid+HowTo+K710+SSv2+AVA2.2+more. +3 LAION-300M+KMash+WebVid+InternVid+LLaVA+more. + +# 5.2. Comparison with state-of-the-art + +K400. Results appear in Table 1. Compared to VideoMamba, VideoMambaPro has slightly fewer parameters and FLOPs. This is primarily because VideoMamba employs an additional projection layer to generate the weight coefficient $z$ to adjust $\mathbf{A}_f$ and $\mathbf{A}_b$ . See the supplementary materials for an architecture comparison. VideoMambaPro outperforms VideoMamba across model and input sizes. With $224^2$ inputs and pre-trained only on IN-1K, the best-performing VideoMambaPro-M achieves a top-1 accuracy of $84.0\%$ , $1.6\%$ higher than VideoMamba-M. Further comparisons appear in Section 5.4. Increasing the input size to $336^2$ leads to a performance improvement of $1.0 - 1.7\%$ . + +VideoMambaPro scores lower than the recent InternVideo2-1B [52] by $7.6\%$ , but was only pre-trained on IN-1K and has significantly fewer parameters (1000M vs 72M) and inference only takes $\sim 5.5\%$ of the FLOPs. + +Table 1. Performance on K400. Top part of the table are Transformer models, bottom part are Mamba models. We report crops (temporal $\times$ spatial) and FLOPs for inference. —: not reported. + +
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]K40016 × 22421×337M213G64.789.2
MViTv1-B [6]K40032 × 22421×337M510G67.190.8
MViTv2-S [25]K40016 × 22421×335M195G68.291.4
MViTv2-B [25]K40032 × 22421×351M675G70.592.7
Uniformer-S [20]IN-1K+K40016 × 22421×321M126G67.791.4
Uniformer-B [20]IN-1K+K40016 × 22421×350M291G70.492.8
TimeSformer-L [2]IN-21K16 × 22421×3121M5.1T62.5-
ViViT-L [1]IN-21K+K40016 × 22424×3311M47.9T65.489.8
Mformer-HR [36]IN-21K+K40016 × 33621×3311M3.6T68.191.2
MaskFeat-L [53]IN-21K64 × 31224×3218M8.5T75.095.0
VideoMAE-L [47]IN-21K32 × 22421×3305M4.3T75.495.2
TubeViT-L [39]IN-1K32 × 22424×3311M9.5T76.195.2
InternVideo-1B [51]See Table 164 × 224216×41300M86.2T77.295.9
InternVideo2-1B [52]See Table 164 × 224216×41000M-77.1-
InternVideo2-6B [52]See Table 164 × 224216×45903M-77.4-
VideoMamba-TiIN-1K16 × 22422×37M102G66.089.6
VideoMamba-SIN-1K16 × 22422×326M408G67.690.9
VideoMamba-MIN-1K16 × 22424×374M2.4T68.391.4
VideoMambaPro-TiIN-1K16 × 22422×37M96G67.991.2
VideoMambaPro-SIN-1K16 × 22422×325M382G68.891.4
VideoMambaPro-MIN-1K16 × 22424×372M2.2T69.491.6
+ +Table 2. Performance on SSv2. —: not reported. Top part of the table are Transformer models, bottom part are Mamba models. + +SSv2. Results appear in Table 2. VideoMambaPro outperforms VideoMamba by $1.1 - 1.9\%$ . It also outperforms several popular transformer models. Although InternVideo-1B [51] and InternVideo2-6B [52] outperform our VideoMambaPro-M by $7.8\%$ and $8.0\%$ , respectively, they require 18.0-82 times more parameters and at least 39 times more FLOPs. Again, we expect that the performance for VideoMambaPro will increase with more pre-training. + +UCF-101/HMDB51/AVA V2.2. From Table 3, it shows that VideoMambaPro-M is competitive, and outperforms VideoMamba by $3.4\%$ and $1.8\%$ on UCF-101 and HMDB51, respectively. VideoMambaPro-M achieves 31.9 mAP on AVA V2.2, which is $10.7\%$ lower than VideoMAE V2 [50] but with an order of magnitude fewer parameters and FLOPs and pre-trained only on IN-1K (see Table 4). + +
MethodParamsUCF-101HMDB51
VideoMoCo [33]15M78.749.2
CoCLR [16]9M81.452.1
MemDPC [15]32M86.154.5
Vi²CLR [4]9M89.155.7
VideoMAE [47]87M91.362.6
GDT [35]33M95.272.8
VideoMAE V2 [50]1050M99.688.1
VideoMamba-M74M88.260.8
VideoMambaPro-M72M91.663.2
+ +Table 3. Results on UCF-101 and HMDB51. + +
MethodFLOPsParammAP
SlowFast R101 [8]138G53M23.8
VideoMAE-B [47]180G87M26.7
MViTv2-B [25]225G51M30.5
ObjectTransformer [54]243G86M31.0
MViTv2-L [25]2828G213M34.4
ST-MAE-H [9]1193G632M36.2
VideoMAE V2 [50]4220G1050M42.6
VideoMamba-M [19]202G74M30.1
VideoMambaPro-M183G72M31.9
+ +Table 4. Results on AVA V2.2. + +
ModelsInputTop-1Top-5
VideoMamba-M (baseline)32 × 224282.495.7
VideoMambaPro-M (w/o residual)32 × 224283.6 (+1.2)96.0 (+0.3)
VideoMambaPro-M (w/o masking)32 × 224283.0 (+1.0)95.8 (+0.1)
VideoMambaPro-M32 × 224284.0 (+1.6)96.4 (+0.7)
+ +Table 5. Ablation study on K400, with and without masked backward computation and elemental residual connections. + +# 5.3. Ablation study + +Influence of masked backward computation and elemental residual connection. We have identified two limitations that exist in VideoMamba: historical decay and element contradiction. We introduced masked backward computation and elemental residual connections to address these respective issue. Here, we analyze the impact of each solution. We use the same settings as before, with VideoMambaPro-M and pre-training on IN-1K. We summarize the performance of VideoMambaPro-M on K400 in Table 5. Both solutions contribute to an improved score, and their effect is partly complementary. This indicates that the two limitations exist simultaneously in VideoMamba. + +Evidence of limitations of Mamba. We have theoretically identified historical decay and element contradiction as fundamental problems in Mamba, and propose the masked backward computation and elemental residual connection to solve the limitations. Previous experiments on image and video action recognition demonstrated the effectiveness of our approach, and we further provide concrete evidence that the increased performance stems from + +addressing these two issues. + +We performed two tests on ImageNet-1K image classification task to build the connection between the issues in practice and our approach. First, to investigate whether the residual connection alleviates element contradiction, we randomly replaced a percentage of the patches by randomly selected patches from other classes, which will increase element contradiction because more irrelevant information is present. The table 6 shows that the relative performance of VideoMambaPro over VideoMamba increases to $9.2\%$ when $20\%$ of the patches is replaced. For higher percentages, the gap is reduced along with the overall score. This demonstrates that irrelevant (or even misleading) information can better be dealt with using our approach. + +
Replacement ratio
051020305080
VideoMamba-Ti76.973.268.059.452.947.628.9
VideoMambaPro-Ti78.978.076.168.655.948.229.1
+ +Second, to verify that masking addresses the issue of historical decay, we adopted a progressive masking approach, gradually decreasing the mask values of the diagonal elements of $\mathbf{M}_b$ from 1 to 0, and then retrained the model on ImageNet-1K. The results below show that VideoMambaPro's accuracy gradually increases when more masking is applied, evidencing the merits of the approach. + +Table 6. Top-1 accuracy of VideoMambaPro-Ti and VideoMamba-Ti with various replacement ratios (in %) on ImageNet-1K. + +
Mask value
10.70.50
VideoMambaPro-Ti76.977.077.277.8
+ +Table 7. Top-1 accuracy across mask values on ImageNet-1K. + +# 5.4. Comparison with VideoMamba on K400 + +We more thoroughly compare the differences between VideoMamba and VideoMambaPro by investigating the relative performance per class. We then present a statistical comparison between the results of both backbones. + +Class analysis. We compare VideoMambaPro-M with $224^{2}$ image size pre-trained on IN-1K to a VideoMamba-M baseline with the same settings. We show the relative performance for all classes of K400 in Figure 2. For over $95\%$ of the classes, VideoMambaPro shows improvement. Although there is a lower performance for certain classes, the decrease is typically limited. + +The majority of the classes sees a $\sim 1.8\%$ improvement, which is substantial. For a small number of classes, VideoMambaPro performs $>2\%$ better than VideoMamba. Only a fraction of the classes is negatively affected by the solutions introduced in VideoMambaPro. + +![](images/28cbe6524cbac5873fbb947b49b348fef62046a69a53c8f71d72fff8b7321620.jpg) +Figure 2. Relative accuracy per class on Kinetics-400 by comparing VideoMambaPro-M to VideoMamba-M. + +Statistical comparison. In order to understand whether the improvements of VideoMambaPro over VideoMamba are statistically significant, we compare the K400 results of the respective Middle models, both pre-trained on IN-1K and size $224 \times 224$ . Other settings are also the same. For each test sample, we check whether it is correctly classified by either model. The results appear in Table 8. + +
VideoMambaPro-M
TrueFalseTotal
VideoMamba-MTrue14,30246914,771
False1,8333,3025,135
Total16,1353,77119,906
+ +Table 8. Contingency table for K400 test items for VideoMamba-M and VideoMambaPro-M. + +We used the McNemar test, a non-parametric test with a single degree of freedom. It checks whether the number of items that are incorrectly classified by VideoMambaPro $(n_{10})$ but not VideoMamba is substantially lower than the number of items misclassified by VideoMamba but not VideoMambaPro $(n_{01})$ . The test is calculated as $\chi^2 = \frac{(n_{01} - n_{10})^2}{(n_{01} + n_{10})}$ . The resulting value of 808.2 corresponds to a significance level of $p < 0.001$ . We can thus conclude that VideoMambaPro-M is statistically significantly better than VideoMamba-M. Because we relied on the aggregated performance reported in papers for other methods, we cannot report statistical comparisons here. + +# 5.5. Computation cost analysis + +Finally, we compare the performance of VideoMambaPro with various model sizes with other approaches on K400. We map the top-1 to the number of parameters and FLOPs in Figures 3 and 4, respectively. VideoMambaPro performs favorably in terms of the accuracy-compute trade + +![](images/71a21b66d61b9f0c71b7cc6c12157524b8c07ce69f225eb685dcf1319631d141.jpg) +Figure 3. Top-1 accuracy versus number of parameters of VideoMambaPro and other models on Kinetics-400. + +![](images/ef37a2df50cef53f9627a9b71dade869f27255428064217d2bd48f7c18889190.jpg) +Figure 4. Top-1 accuracy versus number of FLOPs of VideoMambaPro and other models on Kinetics-400. + +off. Importantly, VideoMambaPro was trained on much less data than other models, which might provide opportunities for further accuracy gains without additional compute and memory requirements. + +# 6. Conclusion + +From a mathematical comparison with self-attention, we have identified two limitations in how Mamba processes token sequences. We argue that these limitations constrain Mamba's potential, especially in video understanding tasks. To address the two limitations, we have introduced VideoMambaPro (VMP). It takes VideoMamba and introduces the masked backward State Space Model (SSM), and adds residual connections in both forward and backward SSM. In experiments on Kinetics-400, Something-Something V2, HMDB51, UCF-101, and AVA V2.2, VideoMambaPro consistently demonstrates improved performance over the vanilla VideoMamba. We expect that extensive pre-training further elevates the performance of Mamba models for video tasks, making it an increasingly attractive, efficient alternative to large transformer models. + +# References + +[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 6 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, pages 813-824, 2021. 6 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 5 +[4] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2CLR: Video and image for visual contrastive learning of representation. In CVPR, pages 1502-1512, 2021. 7 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[6] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 6 +[7] Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. NeurIPS, 34:26183-26197, 2021. 2 +[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 7 +[9] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 35:35946-35958, 2022. 2, 7 +[10] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The "Something Something" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 5 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2 +[12] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 33:1474-1487, 2020. 2 +[13] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2, 3 +[14] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. AVA: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, pages 6047-6056, 2018. 6 + +[15] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, pages 312-329, 2020. 7 +[16] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 7 +[17] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of selfattention. In International Conference on Algorithmic Learning Theory, pages 597-619. PMLR, 2023. 1 +[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. HMDB: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 6 +[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. UniformerV2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 1, 7 +[20] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 6 +[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2 +[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 1, 2, 5 +[23] Shufan Li, Harkanwar Singh, and Aditya Grover. Mamba-nd: Selective state space modeling for multi-dimensional data. In European Conference on Computer Vision, pages 75–92. Springer, 2024. 2 +[24] Wenrui Li, Xiaopeng Hong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 5 +[25] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MViTv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 6, 7 +[26] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. SimpleClick: Interactive image segmentation with simple vision transformers. In ICCV, pages 22290-22300, 2023. 2 +[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[29] Hui Lu, Hu Jian, Ronald Poppe, and Albert Ali Salah. Enhancing video transformers for action understanding with vlm-aided training. arXiv preprint arXiv:2403.16128, 2024.1 + +[30] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. NeurIPS, 34:21297-21309, 2021. 2 +[31] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 1 +[32] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18, 2022. 6 +[33] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. VideoMoCo: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 7 +[34] Jinyoung Park, Hee-Seon Kim, Kangwook Ko, Minbeom Kim, and Changick Kim. Videomamba: Spatio-temporal selective state space model. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2 +[35] Mandela Patrick, Yuki Markus Asano, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multimodal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 7 +[36] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 6 +[37] Badri Narayana Patro and Vijay Srinivas Agneeswaran. Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges. arXiv preprint arXiv:2404.16112, 2024. 1 +[38] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, et al. RWKV: Reinventing RNNs for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 1, 2 +[39] AJ Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking video ViTs: Sparse video tubes for joint image and video learning. In CVPR, pages 2214-2224, 2023. 6 +[40] Hongyu Ren, Hanjun Dai, Zihang Dai, Mengjiao Yang, Jure Leskovec, Dale Schuurmans, and Bo Dai. Combiner: Full attention transformer with sparse computation cost. NeurIPS, 34:22470-22482, 2021. 2 +[41] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 1 +[42] Gilad Sharir, Asaf Noy, and Lihi Zelnik-Manor. An image is worth 16x16 words, what is a video worth? arXiv preprint arXiv:2103.13915, 2021. 6 +[43] Yuanyuan Shen, Edmund M-K Lai, and Mahsa Mohagheh. Effects of similarity score functions in attention mechanisms on the performance of neural question answering systems. Neural Processing Letters, 54(3):2283-2302, 2022. 2 + +[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.6 +[45] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2 +[46] Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621, 2023. 1, 2 +[47] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 1, 2, 6, 7 +[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 6 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 2 +[50] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 1, 6, 7 +[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6 +[52] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. InternVideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6 +[53] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 6 +[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 7 +[55] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 6 +[56] Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianxiang Chen, and Zi Ye. A survey on Visual Mamba. arXiv preprint arXiv:2404.15956, 2024. 5 +[57] Zheng Zhang and Kil To Chong. Comparison between first-order hold with zero-order hold in discretization of input-delay nonlinear systems. In 2007 International Conference on Control, Automation and Systems, pages 2892-2896, 2007. 3 + +[58] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. ViT-YOLO: Transformer-basedYOLO for object detection. In ICCV, pages 2799-2808, 2021. 2 +[59] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. NeurIPS, 34:17723-17736, 2021. 2 +[60] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. BiFormer: Vision transformer with bilevel routing attention. In CVPR, pages 10323-10333, 2023. 2 +[61] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2, 4 \ No newline at end of file diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/images.zip b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b000435574e1ed692f73bebd55df99513b11415c --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2542579a1db18d020519374f11341c9dddc754e8ac6a006e765c4213e1dcfd6 +size 601313 diff --git a/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/layout.json b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..360b3428c0c8363594234e6c9bcbc4351cad5e3a --- /dev/null +++ b/2025/Snakes and Ladders_ Two Steps Up for VideoMamba/layout.json @@ -0,0 +1,12066 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 470, + 121 + ], + "type": "text", + "content": "Snakes and Ladders: Two Steps Up for VideoMamba" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 144, + 226, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 144, + 226, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 144, + 226, + 185 + ], + "type": "text", + "content": "Hui Lu \nUtrecht University \nh.lu1@uu.nl" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 259, + 144, + 351, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 259, + 144, + 351, + 185 + ], + "spans": [ + { + "bbox": [ + 259, + 144, + 351, + 185 + ], + "type": "text", + "content": "Albert A. Salah \nUtrecht University \na.a.salah@uu.nl" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 384, + 144, + 476, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 144, + 476, + 185 + ], + "spans": [ + { + "bbox": [ + 384, + 144, + 476, + 185 + ], + "type": "text", + "content": "Ronald Poppe \nUtrecht University \nr.w.poppe@uu.nl" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "content": "Video understanding requires the extraction of rich spatiotemporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by " + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "inline_equation", + "content": "1.6 - 3.0\\%" + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "inline_equation", + "content": "1.1 - 1.9\\%" + }, + { + "bbox": [ + 55, + 238, + 297, + 478 + ], + "type": "text", + "content": " top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 513, + 135, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 135, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 135, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "Video understanding is challenging and requires models that can extract rich spatio-temporal representations from video inputs. Transformers are powerful neural networks capable of effectively capturing temporal and spatial information from videos [19, 29, 50]. Current state-of-the-art video understanding models are transformers [41, 51]. At the core of transformers is self-attention [49], the self-alignment between tokens in an input obtained by estimating the relative importance of a token with respect to all other tokens. The long-range token dependency accounts for much of the success of transformer models [5, 49]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Calculating self-attention is computationally costly, which eventually limits the application of powerful transformer models in practical settings [17]. Recently, alternative models with lower-cost operators have been proposed" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 274 + ], + "type": "text", + "content": "for national language processing (NLP), including S4 [13], RWKV [38], and RetNet [46]. Among these, Mamba [11] shows the best performance on long-range and causal tasks such as language understanding [31] and content-based reasoning [37]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "content": "Motivated by the favorable computational cost, researchers have recently extended Mamba from the NLP domain to the computer vision domain. The core adaptation involved splitting the input image into multiple regions and embedding these as continuous tokens [61]. For video understanding, the recently proposed VideoMamba [22] extracts key frames from videos as the continuous input sequence. However, compared to previous transformer-based methods, VideoMamba's performance on video benchmarks is markedly lower. For example, VideoMamba achieves " + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "inline_equation", + "content": "82.4\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "content": " top-1 on Kinetics-400, compared to " + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 313, + 276, + 555, + 419 + ], + "type": "text", + "content": " for VideoMAE [47], indicating room for improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 422, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 577 + ], + "type": "text", + "content": "In this paper, we first analyze differences in the feature extraction capabilities of transformers and Mamba. We identify two limitations of Mamba: historical decay and element contradiction. We then extend VideoMamba [22] to mitigate these limitations. The proposed VideoMambaPro (VMP) addresses historical decay through masked backward computation in the bi-directional Mamba process, allowing the network to better handle historical tokens. We introduce residual connections to Mamba's matrix elements to tackle element contradiction. VideoMambaPro consistently improves the performance of VideoMamba on video understanding tasks, positioning it as a strong, efficient competitor to transformers. Our contributions are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 553, + 673 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 579, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 553, + 613 + ], + "type": "text", + "content": "- We derive a formal representation of Mamba from the perspective of self-attention and identify two limitations of Mamba in the video analysis domain." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 615, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 553, + 638 + ], + "type": "text", + "content": "- We propose VideoMambaPro, effectively addressing Mamba's limitations for video understanding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 639, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 673 + ], + "type": "text", + "content": "- We report strong video action recognition performance compared to recent Vision Mamba-based models, and surpass the original VideoMamba by clear margins." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We first discuss related work. Then, we provide a theoretical analysis and introduce VideoMambaPro. Experiments are presented in Section 5. We conclude in Section 6." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24234" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "content": "Transformers. One core aspect of transformers is self-attention [49] to achieve long-range interactions by measuring the similarity between tokens. Self-attention was introduced in the computer vision domain for tasks such as image recognition [26, 45] and object detection [7, 58]. Subsequent works (e.g., [9, 21, 47, 53] extended vision transformers to the video domain, to achieve superior performance. However, the mechanism of self-attention introduces significant computational overhead because of the similarity between all pairs of tokens needs to be calculated." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 319 + ], + "type": "text", + "content": "Alternative models. Recent work has introduced alternative models with reduced computational complexity, while maintaining the advantages of self-attention [30, 40, 59]. SOFT [30] uses Gaussian kernel functions to replace the dot-product similarity, which enables a full self-attention matrix to be approximated with a low-rank matrix decomposition. Combiner [40] employs a structured factorization to approximate full self-attention, realizing low computation and memory complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 437 + ], + "type": "text", + "content": "Peng et al. [38] propose the Receptance Weighted Key Value (RWKV) architecture that combines self-attention training with an efficient recurrent neural network (RNN) inference using a linear attention mechanism. Through parallel computation, a lower, constant-level computational and memory complexity is achieved. RetNet [46] includes another variant of self-attention, by dividing the input into multiple chunks. Within each chunk, the self-attention mechanism can be computed in parallel, while information is transmitted between chunks based on an RNN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 558 + ], + "type": "text", + "content": "State-space models. The S4 model completely abandons self-attention and, instead, builds upon a state space model [13]. Instead of calculating a similarity matrix by performing matrix multiplications for pairs of tokens, it enables the network to directly learn a global high-order polynomial projection operator (HiPPO) matrix to handle relations between tokens. Additionally, for the simultaneous input of multiple tokens, S4 proposes a convolutional processing approach, enabling parallel training and thereby accelerating the training process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 677 + ], + "type": "text", + "content": "Based on S4, Mamba [11] proposes a selection mechanism where, for each input token, a unique HiPPO matrix [12] is generated. This allows the model to selectively process input tokens, enabling it to focus on or ignore specific inputs. Due to Mamba's strong representation ability in NLP, and linear-time complexity, it has garnered attention as a promising alternative to transformers. In the computer vision domain, researchers have proposed Vision Mamba [61] and VMamba [27] for tasks such as image classification and object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In the video domain, several Mamba variants have been proposed [22, 23, 34]. Their performance is somewhat lower than expected, with limited understanding of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "causes. We argue that a systematic, mathematical analysis of Mamba from the perspective of self-attention could reveal shortcomings of Mamba's inner workings. Better understanding of these limitations allow us to develop improvements, and to close the accuracy performance gap with transformers, while enjoying Mamba's efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 154, + 434, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 434, + 167 + ], + "type": "text", + "content": "3. Theoretical Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 210 + ], + "type": "text", + "content": "We revisit Mamba from the perspective of self-attention. Then, we analyze its limitations for video understanding. To address these, we propose VideoMambaPro in Section 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 548, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 548, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 548, + 229 + ], + "type": "text", + "content": "3.1. Mamba from the perspective of self-attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": "Self-attention. Given an input sequence " + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\coloneqq [x_1, \\dots, x_N] \\in \\mathbb{R}^{N \\times D_x}" + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": " feature vectors of depth " + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "inline_equation", + "content": "D_x" + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": ", self-attention [49, 60] computes the output sequence " + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 234, + 554, + 281 + ], + "type": "text", + "content": " following two steps:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "content": "Step 1: Similarity matrix computation. The input sequence " + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "content": " is linearly projected onto the three different subspaces query " + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{N \\times D_V}" + }, + { + "bbox": [ + 313, + 281, + 554, + 329 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 339, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 339, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 349, + 339, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\boldsymbol {X} \\mathbf {W} _ {Q} ^ {\\top}, \\mathbf {K} = \\boldsymbol {X} \\mathbf {W} _ {K} ^ {\\top}, \\mathbf {V} = \\boldsymbol {X} \\mathbf {W} _ {V} ^ {\\top}. \\tag {1}", + "image_path": "cd347c782579c2168f89094928a846551cfa1024789c460fc7733fd48d34c23a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q, \\mathbf{W}_K \\in \\mathbb{R}^{D \\times D_x}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V \\in \\mathbb{R}^{D_v \\times D_x}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " the corresponding weight matrices. Specifically, " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} := [q_1, \\dots, q_N]^\\top" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{K} := [k_1, \\dots, k_N]^\\top" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{V} := [v_1, \\dots, v_N]^\\top" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " with vectors " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "q_i, k_i, v_i" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " the query, key, and value vectors, respectively, for input " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ", similarity matrix " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": " contains the correlations between all query and key vectors, with a softmax function applied to each row of " + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 358, + 554, + 457 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 463, + 553, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 463, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 377, + 463, + 553, + 478 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top} / \\sqrt {D}\\right). \\tag {2}", + "image_path": "202a01ac82c882c8d92dffb8a3f7f6f7b7e36f289f7e436f5e00ac3a7f399646.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": "Each component " + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "inline_equation", + "content": "s_{ij} (i,j = 1,\\dots ,N)" + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": " represents the similarity score between " + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{q}_i" + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\pmb{k}_j" + }, + { + "bbox": [ + 313, + 485, + 553, + 509 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "text", + "content": "Step 2: Output computation. Output sequence " + }, + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} := [y_1, \\dots, y_N]^\\top" + }, + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "text", + "content": " is then calculated based on " + }, + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 314, + 509, + 553, + 534 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 412, + 541, + 553, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 541, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 412, + 541, + 553, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {S V}. \\tag {3}", + "image_path": "a25527e11ea1b031a0d6bd0aacdc29aec277319a76be6a9257ce14193abdbdb0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "text", + "content": "It follows that each output vector " + }, + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i" + }, + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 313, + 561, + 554, + 584 + ], + "type": "text", + "content": ") can be written in vector form as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 401, + 590, + 553, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 401, + 590, + 553, + 623 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {i} = \\sum_ {j = 1} ^ {N} s _ {i j} \\boldsymbol {v} _ {j}. \\tag {4}", + "image_path": "6b435e4ea5c3bf3bd14a5a0ad3f7bfb3d5fb3896b8f85bec710fcc84082bcac1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": "Any output vector " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " is a linear combination of vectors " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j (j = 1, \\dots, N)" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": ", with similarity score " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " serving as coefficient. The larger the similarity score, the greater the influence of " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " on output " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " [43]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Mamba. State Space Models (SSMs) serve as the foundation of Mamba [11]. They are based on continuous systems that map 1D functions or sequences, " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}^L \\to" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24235" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "inline_equation", + "content": "y(t)\\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "text", + "content": " to output sequences " + }, + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "inline_equation", + "content": "y(t)" + }, + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "text", + "content": " through a hidden state " + }, + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "inline_equation", + "content": "h(t)\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 72, + 297, + 99 + ], + "type": "text", + "content": " . Formally, SSM implements the mapping as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 114, + 295, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 114, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 118, + 114, + 295, + 126 + ], + "type": "interline_equation", + "content": "h (t) = \\mathbf {A} h (t - 1) + \\mathbf {B} x (t) \\tag {5}", + "image_path": "3d864d7f7f5b03abbc9092b9b87b5d47ab26fd46f736ad0f240a1b8b71b3176c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 119, + 129, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 129, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 119, + 129, + 295, + 142 + ], + "type": "interline_equation", + "content": "y (t) = \\mathbf {C} h (t) \\tag {6}", + "image_path": "0d9eb97a32498ab3cea66a9f9f35fc16a91e3fafe824104a10f202461e57e44c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": " is the evolution matrix of the system, and " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": " are the projection matrices. Often, the input is discrete rather than a continuous function " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": ". Therefore, Mamba performs discretization, effectively creating a discrete version of the continuous system. A timescale parameter " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": " is used to transform the continuous parameters " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": " into their discrete counterparts " + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}, \\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 55, + 147, + 296, + 256 + ], + "type": "text", + "content": ", and the transformation typically employs the zero-order hold method [57]. This process is expressed as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 260, + 295, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 260, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 97, + 260, + 295, + 274 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\tag {7}", + "image_path": "bb62b8b85074361a7812258a827aaa3299fb6eeaadafa112f66e394f90fe4c8c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 276, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 276, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 99, + 276, + 295, + 290 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {B}}} = (\\boldsymbol {\\Delta} \\mathbf {A}) ^ {- 1} \\left(\\exp (\\boldsymbol {\\Delta} \\mathbf {A}) - \\mathbf {I}\\right) \\cdot \\boldsymbol {\\Delta} \\mathbf {B}, \\tag {8}", + "image_path": "6942a162511090b8b21dfc21a123aa9663e1215f24e22103262ec12b3af14fde.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 293, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 293, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 99, + 293, + 295, + 306 + ], + "type": "interline_equation", + "content": "h _ {t} = \\overline {{\\mathbf {A}}} h _ {t - 1} + \\overline {{\\mathbf {B}}} x _ {t}, \\tag {9}", + "image_path": "34f18bb0fadeefa5d7e04fbd466703de3ccb267189a94217a1d931b21d5b8c46.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 309, + 294, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 309, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 99, + 309, + 294, + 322 + ], + "type": "interline_equation", + "content": "y _ {t} = \\mathbf {C h} _ {t}. \\tag {10}", + "image_path": "c509d0e8969911af0898a2e49c271c8cc3ffde2df45731e1b71f2f93b8a67650.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": "Considering that parameters " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " in the original SSM are independent of the input data " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " and cannot be tailored to specific input data, Mamba employs a Selective Scan Mechanism as its core operator. More precisely, three functions " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_B(x),S_C(x),S_\\Delta (x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " are introduced to associate parameters " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " , C, and " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " in Eqs. 7-10 to the input data " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " Based on " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_{\\Delta}(x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " can also be associated with the input data " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " . For example, given the input " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " , functions " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_{\\Delta}(x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " will produce the corresponding " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_1" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " based on Eq. 7, and functions " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_B(x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_{\\Delta}(x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " will produce the corresponding " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}_1" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " based on Eq. 8. " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{C}}_1" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " is obtained based on function " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "S_{C}(x)" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " . Following Eqs. 9 and 10, we analyze the process to obtain output sequence " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " when given an input sequence " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{X}:= [x_1,\\dots ,x_N]\\in \\mathbb{R}^{N\\times D_x}" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 327, + 296, + 508 + ], + "type": "text", + "content": " feature vectors. Each vector's hidden state is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 525, + 295, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 525, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 63, + 525, + 295, + 538 + ], + "type": "interline_equation", + "content": "h _ {1} = \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1}, \\tag {11}", + "image_path": "cdbfb7031da463a8a59cfed8b9fd3d679905878c7ed295944b68dc4987465e83.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 541, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 541, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 63, + 541, + 294, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h _ {2} = \\overline {{\\mathbf {A}}} _ {2} h _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ = \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2}, \\tag {12} \\\\ \\end{array}", + "image_path": "de0d2325d33431e2f9bd3569a021d3c5c37ff6bf222d0b05bedc25876c9e0d68.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 574, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 574, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 63, + 574, + 295, + 604 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h _ {3} = \\overline {{\\mathbf {A}}} _ {3} h _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3} \\\\ = \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {I} + \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} + \\overline {{\\mathbf {B}}} _ {3} \\boldsymbol {x} _ {3}, \\tag {13} \\\\ \\end{array}", + "image_path": "d4d8a7079fa817dc5ddaa95cef55a05ce4da8efa3fcbb5a905d2ccb5b0e9c1a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 76, + 609, + 89, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 609, + 89, + 615 + ], + "spans": [ + { + "bbox": [ + 76, + 609, + 89, + 615 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 60, + 621, + 294, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 621, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 60, + 621, + 294, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h _ {N} = \\overline {{\\mathbf {A}}} _ {N} h _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N} \\\\ = \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} \\boldsymbol {x} _ {1} + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {A}}} _ {N - 1} \\dots \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} \\boldsymbol {x} _ {2} \\\\ + \\overline {{\\mathbf {A}}} _ {N} \\overline {{\\mathbf {B}}} _ {N - 1} \\boldsymbol {x} _ {N - 1} + \\overline {{\\mathbf {B}}} _ {N} \\boldsymbol {x} _ {N}. \\tag {14} \\\\ \\end{array}", + "image_path": "42d3f0eb8c14b23443a90c89874e1772a4b3115585684684b2724a09b273c6d9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 72, + 497, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 497, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 497, + 84 + ], + "type": "text", + "content": "Eqs. 11-14 can be written in matrix form:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 92, + 553, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 92, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 326, + 92, + 553, + 181 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {H} = \\left[ h _ {1}, h _ {2}, h _ {3}, \\dots , h _ {N} \\right] ^ {\\top} \\\\ = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {1} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right]. \\tag {15} \\\\ \\end{array}", + "image_path": "a0cd2b51b1af10bd4b68da5e7b7a01acb2b8927eb79d9ddc675a4a12c0058d70.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "text", + "content": "For output sequence " + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} := [\\pmb{y}_1, \\dots, \\pmb{y}_N]^\\top" + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "text", + "content": ", each vector " + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i" + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 314, + 181, + 553, + 206 + ], + "type": "text", + "content": ") can be expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 403, + 214, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 214, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 403, + 214, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {N} = \\overline {{\\mathbf {C}}} _ {N} h _ {N}, \\tag {16}", + "image_path": "64860d762dd35e3dbde3f936efa51e3e75e4d5282312faae1e270b1146eceba8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 237, + 406, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 406, + 248 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 406, + 248 + ], + "type": "text", + "content": "and in matrix form as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 255, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 255, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 323, + 255, + 553, + 326 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {C}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ 0 & \\overline {{\\mathbf {C}}} _ {2} & 0 & \\dots & 0 \\\\ 0 & 0 & \\overline {{\\mathbf {C}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ 0 & 0 & 0 & \\dots & \\overline {{\\mathbf {C}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\\\ \\vdots \\\\ h _ {N} \\end{array} \\right] = \\overline {{\\mathbf {C}}} h. \\tag {17}", + "image_path": "14a81355c30415f3b1817b12e3740b0a585532c424fa87e6066dded0118d5888.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 334, + 512, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 334, + 512, + 346 + ], + "spans": [ + { + "bbox": [ + 326, + 334, + 512, + 346 + ], + "type": "text", + "content": "By substituting Eq. 15 into Eq. 17, we obtain:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 327, + 354, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 354, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 327, + 354, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\overline {{\\mathbf {C}}} \\left[ \\begin{array}{c c c c c} \\overline {{\\mathbf {B}}} _ {1} & 0 & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {B}}} _ {2} & 0 & \\dots & 0 \\\\ \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {A}}} _ {2} \\overline {{\\mathbf {B}}} _ {1} & \\overline {{\\mathbf {A}}} _ {3} \\overline {{\\mathbf {B}}} _ {2} & \\overline {{\\mathbf {B}}} _ {3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {2} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {1} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {3} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {2} & (\\prod_ {\\mathrm {j} = \\mathrm {N}} ^ {4} \\overline {{\\mathbf {A}}} _ {\\mathrm {j}}) \\overline {{\\mathbf {B}}} _ {3} & \\dots & \\overline {{\\mathbf {B}}} _ {N} \\end{array} \\right] \\left[ \\begin{array}{l} x _ {I} \\\\ x _ {2} \\\\ x _ {3} \\\\ \\vdots \\\\ x _ {N} \\end{array} \\right], \\tag {18}", + "image_path": "242bbaacef4e348b7c4721df4a277ccec75b02ed2d3767a874ce54e907805dd6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 422, + 425, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 425, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 425, + 434 + ], + "type": "text", + "content": "which can be expressed as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 401, + 442, + 553, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 442, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 401, + 442, + 553, + 456 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\overline {{\\mathbf {C}}} (\\mathbf {M} \\mathbf {X}), \\tag {19}", + "image_path": "db4bc46aa1adc7ceaea2e8ed70787a8c3eb4c69819afac828b9c746c6d99ef4a.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "text", + "content": " represents the second term on the right-hand side of Eq. 18. Recall from Eq. 3 that the result " + }, + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 465, + 553, + 502 + ], + "type": "text", + "content": " obtained by self-attention processing can be expressed as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 385, + 510, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 510, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 385, + 510, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = \\mathbf {S V} = (\\mathbf {S X}) \\mathbf {W} _ {V} ^ {\\top} \\tag {20}", + "image_path": "7db7323d88306fd71a3c9e22f6c5e81e65f522bc64c3675ece57471784e96711.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "From the perspective of self-attention, by comparing Eqs. 19 and 20, the essence of Mamba is to generate a matrix " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " similar to similarity matrix " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ", such that the result of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " is based on the correlation between vectors of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ". Although the final result of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " is left multiplied by a mapping matrix " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ", while the result of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{S}\\mathbf{X}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " is right multiplied by a mapping matrix " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_V^\\top" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ", the geometric meaning of the two are the same." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 635, + 553, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 649 + ], + "type": "text", + "content": "3.2. Limitations of Mamba in video understanding" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "From the perspective of self-attention, the concept of Mamba is similar: both use similarity matrices. We now analyze the differences between the similarity matrices of Mamba and self-attention, and discuss the limitations of Mamba in the context of the video understanding task." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": "1The original SSM [13] employs " + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "inline_equation", + "content": "h'(t) = Ah(t) + Bx(t)" + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": " the hidden state from previous time step " + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "inline_equation", + "content": "h'(t)" + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": " the updated current hidden state, replacing " + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 55, + 673, + 296, + 714 + ], + "type": "text", + "content": ". Considering this approach may lead to ambiguity, we have adopted the updated description." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24236" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "Limitation 1: Historical decay. Matrix M in Eq. 19 corresponds to the second right-hand term in Eq. 18, which is a lower triangular matrix of the form:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 114, + 296, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 114, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 75, + 114, + 296, + 184 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\left[ \\begin{array}{c c c c c} m _ {1 1} & 0 & 0 & \\dots & 0 \\\\ m _ {2 1} & m _ {2 2} & 0 & \\dots & 0 \\\\ m _ {3 1} & m _ {3 2} & m _ {3 3} & \\dots & 0 \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ m _ {N 1} & m _ {N 2} & m _ {N 3} & \\dots & m _ {N N} \\end{array} \\right]. \\tag {21}", + "image_path": "e28aa5693f7bffd045e99fbe2098a180859b460532f7554cc73b2b7b088c1fa6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": "By comparing " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " with matrix " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " in self-attention, we find that outputs in Mamba favor more recent information, because the more weights are zero, the earlier the token is observed. For example, for input " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "[x_1, x_2, x_3]" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{M}x_1" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " in Mamba is " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "m_{11}x_1" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " while the output " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{S}x_1" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "inline_equation", + "content": "s_{11}x_1 + s_{12}x_2 + s_{13}x_3" + }, + { + "bbox": [ + 55, + 190, + 296, + 297 + ], + "type": "text", + "content": " in self-attention. This indicates that, in Mamba, the influence of earlier observed tokens on the final result is greatly diminished. We refer to this limitation as historical decay." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 405 + ], + "type": "text", + "content": "In the NLP domain, more recent dialogue information often has more impact on the final judgment, so this effect is acceptable. However, in the computer vision domain, the order of the tokens has less meaning. Previous works such as Vision Mamba [61] and VMamba [27] tried to solve this issue by processing the token sequence in both forward and backward directions. This produces better results, but significant deficiencies still exist. We will explain this theoretically below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": "When processing bi-directionally, the results generated from input forward tokens " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\left[x_1,\\dots ,x_N\\right]" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_f\\mathbf{X}" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": " and the results generated from input backward tokens " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\left[x_{N},\\dots ,x_{1}\\right]" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b\\mathbf{X}" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": ", are linearly combined to generate the final result " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}\\mathbf{X}" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}" + }, + { + "bbox": [ + 55, + 405, + 296, + 488 + ], + "type": "text", + "content": " a dense matrix. As a result, the influence of historical information on the result is increased, consequently leading to better results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "text", + "content": "For example, for the input tokens " + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "inline_equation", + "content": "[x_1, x_2, x_3]" + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_f\\mathbf{X}" + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b\\mathbf{X}" + }, + { + "bbox": [ + 55, + 488, + 295, + 513 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 518, + 296, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 518, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 73, + 518, + 296, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {f} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} f _ {1 1} & 0 & 0 \\\\ f _ {2 1} & f _ {2 2} & 0 \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {1 f} \\\\ h _ {2 f} \\\\ h _ {3 f} \\end{array} \\right], \\tag {22}", + "image_path": "8f3caf142554f6619008f710aec44807005788e207d3b1b9f8360688ea1e2bed.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 573, + 296, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 573, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 75, + 573, + 296, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {b} \\boldsymbol {X} = \\left[ \\begin{array}{c c c} b _ {3 3} & 0 & 0 \\\\ b _ {2 3} & b _ {2 2} & 0 \\\\ b _ {1 3} & b _ {1 2} & b _ {1 1} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {3} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {1} \\end{array} \\right] = \\left[ \\begin{array}{l} h _ {3 b} \\\\ h _ {2 b} \\\\ h _ {1 b} \\end{array} \\right], \\tag {23}", + "image_path": "194cc27a8115cfa434fa1fac6aeb75ec13b3e96a1d7b29d211ca4393986abe24.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "inline_equation", + "content": "f_{ij}" + }, + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "text", + "content": " represents the similarity score during the forward process, and " + }, + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "inline_equation", + "content": "b_{ij}" + }, + { + "bbox": [ + 55, + 618, + 296, + 667 + ], + "type": "text", + "content": " is the similarity score in the backward direction. After bi-directional computation, with the outputs linearly combined, the results are expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 673, + 271, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 673, + 271, + 687 + ], + "spans": [ + { + "bbox": [ + 61, + 673, + 271, + 687 + ], + "type": "interline_equation", + "content": "h _ {1} = h _ {1 f} + h _ {1 b} = f _ {1 1} \\boldsymbol {x} _ {1} + b _ {1 3} \\boldsymbol {x} _ {3} + b _ {1 2} \\boldsymbol {x} _ {2} + b _ {1 1} \\boldsymbol {x} _ {1}", + "image_path": "bddd838ae0c1d965b2ef9ad60bbca2a387fd73b4f22c11781ee7c716a24e0249.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 688, + 295, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 688, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 688, + 295, + 700 + ], + "type": "interline_equation", + "content": "h _ {2} = h _ {2 f} + h _ {2 b} = f _ {2 1} \\boldsymbol {x} _ {1} + f _ {2 2} \\boldsymbol {x} _ {2} + b _ {2 3} \\boldsymbol {x} _ {3} + b _ {2 2} \\boldsymbol {x} _ {2} \\tag {24}", + "image_path": "6363dc802865b6a7bada7adb48cef4df9efaeb5863ddf496082b9d7af3721667.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 703, + 272, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 272, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 272, + 715 + ], + "type": "interline_equation", + "content": "h _ {3} = h _ {3 f} + h _ {3 b} = f _ {3 1} \\boldsymbol {x} _ {1} + f _ {3 2} \\boldsymbol {x} _ {2} + f _ {3 3} \\boldsymbol {x} _ {3} + b _ {3 3} \\boldsymbol {x} _ {3}", + "image_path": "8c5ecab4491276fce1d96985e64c195f71f4fda42d0d517b5303f4e178f9e16b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 72, + 474, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 474, + 84 + ], + "type": "text", + "content": "We can write Eq. 24 in matrix form:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 92, + 553, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 92, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 323, + 92, + 553, + 171 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\begin{array}{l} h _ {1} \\\\ h _ {2} \\\\ h _ {3} \\end{array} \\right] = \\left[ \\begin{array}{c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} \\end{array} \\right] \\left[ \\begin{array}{l} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right] \\tag {25} \\\\ = \\mathbf {M} _ {b i} \\left[ \\begin{array}{c} \\boldsymbol {x} _ {1} \\\\ \\boldsymbol {x} _ {2} \\\\ \\boldsymbol {x} _ {3} \\end{array} \\right]. \\\\ \\end{array}", + "image_path": "331857e940e84c6b287a083ecea26addd04903680ed1da2658c61fc2cfba6c80.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": "The bi-directional computation transforms matrix " + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": " from a lower triangular matrix to a dense matrix " + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}" + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": ", thereby capturing more historical information and effectively avoiding historical decay. When extending to the case of " + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": " input tokens " + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "inline_equation", + "content": "[\\pmb{x}_1,\\dots ,\\pmb{x}_N]" + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}" + }, + { + "bbox": [ + 313, + 178, + 555, + 239 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 247, + 554, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 247, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 320, + 247, + 554, + 301 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} + b _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} + b _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} + b _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} + b _ {N N} \\end{array} \\right]. \\tag {26}", + "image_path": "80c26424b5a08f613f1fa32ecf12d37865638efeb7f8daf393a84f351179d090.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": "The diagonal elements of " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " contain duplicates of the similarity between a token and itself. For example, " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "f_{33}" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "b_{33}" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " each represent the similarity between token " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "x_{3}" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " and itself. Consequently, the similarity is effectively doubled which weakens the association with other tokens. One possible approach is to adjust " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_f" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " using a weight coefficient " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " through a linear combination. However, this method is not very effective. For example, to prevent the similarity of diagonal elements from being doubled, the " + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 310, + 555, + 464 + ], + "type": "text", + "content": " value is set to 0.5. However, after multiplying other elements by 0.5, their weight values are significantly reduced, which means that the association of each token with other tokens is weakened." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "text", + "content": "Limitation 2: Element contradiction. By analyzing non-zero elements " + }, + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 465, + 554, + 490 + ], + "type": "text", + "content": " of Eq. 18, we derive:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 398, + 498, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 498, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 398, + 498, + 553, + 513 + ], + "type": "interline_equation", + "content": "m _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} \\tag {27}", + "image_path": "071f1e01ca964403481c72be7613a88443182dcf0545d80854edae7c28f352af.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": "After multiple iterations, the above equation results in implicit consideration of the correlation between previous tokens and token " + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": " when computing the correlation between token " + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": " and token " + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": ". As a result, " + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": " exhibits stronger contextual dependencies compared to the elements " + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 313, + 521, + 555, + 653 + ], + "type": "text", + "content": " in the matrix S. This might explain why Mamba achieves better performance than transformers in the field of NLP. While this is advantageous in the NLP domain, for the computer vision domain, input tokens often lack semantic connections. The consideration of the influence of other tokens on each element can lead to significant drawbacks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Interleaved token structures are common when processing images. Tokens that \"belong together\" might not be subsequently processed. For example, in an image classification task, input tokens " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\left[\\pmb{x}_1,\\pmb{x}_2,\\pmb{x}_3\\right]" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " might represent image regions [dog, other, dog]. Ideally, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "m_{31}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " should be high and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24237" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 355, + 251 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 355, + 251 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 355, + 251 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 355, + 251 + ], + "type": "image", + "image_path": "ba5a91a49994f4e3e836d8a9169fa896086e5500ab54a7d793597e9edf1ea360.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 251, + 256, + 262 + ], + "lines": [ + { + "bbox": [ + 181, + 251, + 256, + 262 + ], + "spans": [ + { + "bbox": [ + 181, + 251, + 256, + 262 + ], + "type": "text", + "content": "(a) VideoMambaPro" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 371, + 71, + 529, + 248 + ], + "blocks": [ + { + "bbox": [ + 371, + 71, + 529, + 248 + ], + "lines": [ + { + "bbox": [ + 371, + 71, + 529, + 248 + ], + "spans": [ + { + "bbox": [ + 371, + 71, + 529, + 248 + ], + "type": "image", + "image_path": "3f59be29a79303a6ae787aed361d4bf21fdb4bb8b47d06148eb1469d74093c07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 251, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 394, + 251, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 394, + 251, + 505, + 262 + ], + "type": "text", + "content": "(b) Bi-Directional Mamba Block" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "type": "text", + "content": "Figure 1. (a) Framework of VideoMambaPro with " + }, + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "type": "text", + "content": " bi-directional Mamba blocks. (b) In each bi-directional Mamba block, we employ forward residual SSM and masked backward Residual SSM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "m_{21}" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": " low. Following Eq. 27, " + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "m_{31} = \\overline{\\mathbf{A}}_3 m_{21}" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": ", so " + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_3" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": " needs to increase. However, this causes " + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "m_{32} = \\overline{\\mathbf{A}}_3 m_{22}" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": " to increase because " + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "m_{22}" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": " is also high. But, theoretically, " + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "inline_equation", + "content": "m_{32}" + }, + { + "bbox": [ + 54, + 314, + 296, + 422 + ], + "type": "text", + "content": " should be low. This causes an element contradiction. Especially for video understanding, such contradictions are common because most video regions contain background and other irrelevant information, making relevant tokens sparse. Consequently, the performance of Mamba applied to video analysis tasks is underwhelming [22, 24, 56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 432, + 160, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 160, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 160, + 445 + ], + "type": "text", + "content": "4. VideoMambaPro" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 452, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 511 + ], + "type": "text", + "content": "We use VideoMamba [22] as backbone, and propose two adaptations to address historical decay and element contradiction. The resulting architecture is termed VideoMambaPro (VMP). With minor adjustments, our adaptations can also be applied to related Mamba models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "content": "To address historical decay, we keep the result of " + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_f\\mathbf{X}" + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "content": " but we use masked computation in the backward process. Specifically, we assign a mask to the diagonal elements of " + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "content": ", setting their values to 0, and then proceed with the calculations in Eqs 21-25. We thus eliminate the duplicate similarity on the diagonal, without affecting other elements. The final " + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{bi}" + }, + { + "bbox": [ + 55, + 512, + 296, + 595 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 602, + 295, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 602, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 89, + 602, + 295, + 672 + ], + "type": "interline_equation", + "content": "M _ {b i} = \\left[ \\begin{array}{c c c c c} f _ {1 1} & b _ {1 2} & b _ {1 3} & \\dots & b _ {1 N} \\\\ f _ {2 1} & f _ {2 2} & b _ {2 3} & \\dots & b _ {2 N} \\\\ f _ {3 1} & f _ {3 2} & f _ {3 3} & \\dots & b _ {3 N} \\\\ \\vdots & \\vdots & \\vdots & \\ddots & \\vdots \\\\ f _ {N 1} & f _ {N 2} & f _ {N 3} & \\dots & f _ {N N} \\end{array} \\right]. \\tag {28}", + "image_path": "bbe223ba609cf7d942576aaabfab44dbba835ad94c50b7095ca7237789e7a869.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "To solve element contradiction, we propose residual SSM. This solution is inspired by residual connections, to distribute the requirement for " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " over multiple " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "text", + "content": "This helps to avoid contradictions caused by interleaved sequence structures. For example, for our previous example input sequence " + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "inline_equation", + "content": "[x_1, x_2, x_3]" + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "text", + "content": ", which represents regions [dog, other, dog], we let " + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "inline_equation", + "content": "m_{31} = \\overline{\\mathbf{A}}_3 m_{21} + \\overline{\\mathbf{A}}_3" + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "text", + "content": ". This way, the requirement for a single " + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}_3" + }, + { + "bbox": [ + 313, + 314, + 554, + 387 + ], + "type": "text", + "content": " can be split into two parts, thus avoiding contradictions. This can be expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 384, + 392, + 554, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 392, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 384, + 392, + 554, + 407 + ], + "type": "interline_equation", + "content": "m _ {i j} = \\overline {{\\mathbf {A}}} _ {i} m _ {i - 1, j} + \\overline {{\\mathbf {A}}} _ {i} \\tag {29}", + "image_path": "6b2bb57b76ccedac9de095abb64e169f094c016dd252b010f35875aea4e08028.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": "We implement these solutions into VideoMamba [22], to form VideoMambaPro (see Figure 1). Given input video " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{X}^v \\in \\mathbb{R}^{3 \\times T \\times H \\times W}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": ", we first use a 3D convolution with a " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "1 \\times 16 \\times 16" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " kernel to convert " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{X}^v" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " non-overlapping patch-wise tokens " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{X}^p \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "L = t \\times h \\times w" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "t = T, h = \\frac{H}{16}, w = \\frac{W}{16}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": "). Because SSM is sensitive to token positions, and in line with [22], we include learnable spatial and temporal position embeddings " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{p}_s \\in \\mathbb{R}^{(hw + 1) \\times C}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t \\in \\mathbb{R}^{t \\times C}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": ". Input tokens " + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 313, + 412, + 555, + 521 + ], + "type": "text", + "content": " are expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 528, + 553, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 528, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 377, + 528, + 553, + 541 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} = \\left[ \\boldsymbol {X} _ {\\text {c l s}}, \\boldsymbol {X} \\right] + \\boldsymbol {p} _ {s} + \\boldsymbol {p} _ {t}, \\tag {30}", + "image_path": "8ff7842789be2bb6d1a77e9df33cbc41cb4452ceff56da5cb2275bd48260d271.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "inline_equation", + "content": "X_{cls}" + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "text", + "content": " is a learnable classification token positioned at the start of the sequence. Input tokens " + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "text", + "content": " pass through " + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 547, + 554, + 595 + ], + "type": "text", + "content": " Mamba blocks, and the final layer's [CLS] token is used for classification, after normalization and linear projection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 397, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 397, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 397, + 618 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate VideoMambaPro on five video benchmarks: (a) Kinetics-400 (K400, [3]) comprises " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sim 240\\mathrm{K}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " training and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{K}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " validation videos, each with an average duration of 10 seconds and categorized into 400 classes. (b) Something-Something V2 (SSv2, [10]) includes " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sim 160\\mathrm{K}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " training and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sim 20\\mathrm{K}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " validation videos with" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24238" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "an average duration of 4 seconds, and 174 motion-centric classes. (c) UCF-101 [44] is a relatively small dataset, consisting of " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 9.5\\mathrm{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " training and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 3.5\\mathrm{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " validation videos. (d) HMDB51 [18] is also a compact video dataset, containing " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 3.5\\mathrm{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " training and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 1.5\\mathrm{K}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " validation videos. (e) AVA [14] is a dataset for spatio-temporal localization of human actions with " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 211\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 57\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " validation video segments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": "Implementation. In line with VideoMamba, we introduce three models with increasing embedding dimension and number of bi-directional Mamba blocks " + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": ": Tiny, Small, and Middle (details in supplementary material). To compare with VideoMamba, we pre-train VideoMambaPro on ImageNet-1K (IN-1K). On K400, we also pre-train with IN-1K, fine-tune on the training set and report on the validation set. For K400, we also report on the larger " + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "inline_equation", + "content": "336^2" + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": " input size. During pre-training, we follow DeiT [48] by applying a center crop to obtain the " + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": " sized images. We apply random cropping, random horizontal flipping, label-smoothing regularization, mix-up, and random erasing as data augmentations. We use AdamW [28] with a momentum of 0.9, a batch size of 1024, and a weight decay of 0.05. We employ a cosine learning rate schedule during training, " + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": " initial learning rate over 300 epochs. The fine-tuning settings follow VideoMAE [47]. We resize frames to " + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 56, + 158, + 298, + 410 + ], + "type": "text", + "content": ", and use AdamW with a momentum of 0.9 and a batch size of 512. The evaluation process is the same as VideoMamba, and more details are in the supplementary materials. These materials also include results on IN-1K image classification." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 422, + 294, + 635 + ], + "blocks": [ + { + "bbox": [ + 57, + 422, + 294, + 635 + ], + "lines": [ + { + "bbox": [ + 57, + 422, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 294, + 635 + ], + "type": "table", + "html": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]32 × 22425×137M350G80.294.4
MViTv2-S [25]16 × 22425×135M320G81.094.6
Uniformer-S [20]IN-1K16 × 22424×121M168G80.894.7
Uniformer-B [20]IN-1K16 × 22424×150M388G82.095.1
Uniformer-B [20]IN-1K32 × 22424×350M3.1T83.095.4
STAM [42]IN-21K64 × 22421×1121M1.0T79.2-
TimeSformer-L [2]IN-21K96 × 22421×3121M7.1T80.794.7
ViViT-L [1]IN-21K16 × 22424×3311M47.9T81.394.7
Mformer-HR [36]IN-21K16 × 336210×3311M28.8T81.195.2
VideoMAE-H [47]IN-21K16 × 22425×3633M17.9T86.697.1
X-CLIP-L/14 [32]CLIP-400M16 × 33624×3453M37.0T87.7-
MTV-H [55]60M132 × 22424×31120M44.5T89.198.2
InternVideo-1B [51]412M264 × 224216×41300M86.2T91.198.9
InternVideo-2B [52]414M316 × 224216×41000M-91.6-
InternVideo-2B-6B [52]414M316 × 224216×45903M-92.1-
VideoMamba-TiIN-1K32 × 22424×37M0.4T78.893.9
VideoMamba-TiIN-1K64 × 38424×37M2.4T80.394.8
VideoMamba-SIN-1K32 × 22424×326M1.6T81.595.2
VideoMamba-SIN-1K64 × 38424×326M4.7T82.795.6
VideoMamba-MIN-1K32 × 22424×374M4.8T82.495.7
VideoMamba-MIN-1K64 × 38424×374M28.4T83.396.1
VideoMambaPro-TiIN-1K32 × 22424×37M0.4T81.695.9
VideoMambaPro-TiIN-1K64 × 38424×37M2.2T83.396.1
VideoMambaPro-SIN-1K32 × 22424×325M1.6T83.396.0
VideoMambaPro-SIN-1K64 × 38424×325M4.4T84.596.6
VideoMambaPro-MIN-1K32 × 22424×372M4.7T84.096.4
VideoMambaPro-MIN-1K64 × 38424×372M27.0T85.096.7
", + "image_path": "d614d751fe34406f1a0340a6651b03a8f5f86d3e0f74235b004d6380b03d710c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 677, + 117, + 687 + ], + "lines": [ + { + "bbox": [ + 56, + 677, + 117, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 117, + 687 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 56, + 677, + 117, + 687 + ], + "type": "text", + "content": " IN-21K+WTS" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 688, + 282, + 698 + ], + "lines": [ + { + "bbox": [ + 56, + 688, + 282, + 698 + ], + "spans": [ + { + "bbox": [ + 56, + 688, + 282, + 698 + ], + "type": "text", + "content": "2 CLIP-400M+WebVid+HowTo+K710+SSv2+AVA2.2+more." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 699, + 280, + 709 + ], + "lines": [ + { + "bbox": [ + 56, + 699, + 280, + 709 + ], + "spans": [ + { + "bbox": [ + 56, + 699, + 280, + 709 + ], + "type": "text", + "content": "3 LAION-300M+KMash+WebVid+InternVid+LLaVA+more." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 72, + 492, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 492, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 492, + 85 + ], + "type": "text", + "content": "5.2. Comparison with state-of-the-art" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": "K400. Results appear in Table 1. Compared to VideoMamba, VideoMambaPro has slightly fewer parameters and FLOPs. This is primarily because VideoMamba employs an additional projection layer to generate the weight coefficient " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": " to adjust " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_f" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_b" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": ". See the supplementary materials for an architecture comparison. VideoMambaPro outperforms VideoMamba across model and input sizes. With " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": " inputs and pre-trained only on IN-1K, the best-performing VideoMambaPro-M achieves a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "84.0\\%" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": " higher than VideoMamba-M. Further comparisons appear in Section 5.4. Increasing the input size to " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "336^2" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": " leads to a performance improvement of " + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "inline_equation", + "content": "1.0 - 1.7\\%" + }, + { + "bbox": [ + 313, + 94, + 555, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "content": "VideoMambaPro scores lower than the recent InternVideo2-1B [52] by " + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "content": ", but was only pre-trained on IN-1K and has significantly fewer parameters (1000M vs 72M) and inference only takes " + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\sim 5.5\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "content": " of the FLOPs." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 311, + 553, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 644, + 294, + 676 + ], + "lines": [ + { + "bbox": [ + 55, + 644, + 294, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 294, + 676 + ], + "type": "text", + "content": "Table 1. Performance on K400. Top part of the table are Transformer models, bottom part are Mamba models. We report crops (temporal " + }, + { + "bbox": [ + 55, + 644, + 294, + 676 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 644, + 294, + 676 + ], + "type": "text", + "content": " spatial) and FLOPs for inference. —: not reported." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 311, + 553, + 475 + ], + "lines": [ + { + "bbox": [ + 317, + 311, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 311, + 553, + 475 + ], + "type": "table", + "html": "
MethodPre-trainInputCropsParamFLOPTop1Top5
MViTv1-B [6]K40016 × 22421×337M213G64.789.2
MViTv1-B [6]K40032 × 22421×337M510G67.190.8
MViTv2-S [25]K40016 × 22421×335M195G68.291.4
MViTv2-B [25]K40032 × 22421×351M675G70.592.7
Uniformer-S [20]IN-1K+K40016 × 22421×321M126G67.791.4
Uniformer-B [20]IN-1K+K40016 × 22421×350M291G70.492.8
TimeSformer-L [2]IN-21K16 × 22421×3121M5.1T62.5-
ViViT-L [1]IN-21K+K40016 × 22424×3311M47.9T65.489.8
Mformer-HR [36]IN-21K+K40016 × 33621×3311M3.6T68.191.2
MaskFeat-L [53]IN-21K64 × 31224×3218M8.5T75.095.0
VideoMAE-L [47]IN-21K32 × 22421×3305M4.3T75.495.2
TubeViT-L [39]IN-1K32 × 22424×3311M9.5T76.195.2
InternVideo-1B [51]See Table 164 × 224216×41300M86.2T77.295.9
InternVideo2-1B [52]See Table 164 × 224216×41000M-77.1-
InternVideo2-6B [52]See Table 164 × 224216×45903M-77.4-
VideoMamba-TiIN-1K16 × 22422×37M102G66.089.6
VideoMamba-SIN-1K16 × 22422×326M408G67.690.9
VideoMamba-MIN-1K16 × 22424×374M2.4T68.391.4
VideoMambaPro-TiIN-1K16 × 22422×37M96G67.991.2
VideoMambaPro-SIN-1K16 × 22422×325M382G68.891.4
VideoMambaPro-MIN-1K16 × 22424×372M2.2T69.491.6
", + "image_path": "9873961af44110943ecde5138752f5bda859091ec957061dc4b165199e54e5ed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "Table 2. Performance on SSv2. —: not reported. Top part of the table are Transformer models, bottom part are Mamba models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "content": "SSv2. Results appear in Table 2. VideoMambaPro outperforms VideoMamba by " + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "inline_equation", + "content": "1.1 - 1.9\\%" + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "content": ". It also outperforms several popular transformer models. Although InternVideo-1B [51] and InternVideo2-6B [52] outperform our VideoMambaPro-M by " + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "inline_equation", + "content": "8.0\\%" + }, + { + "bbox": [ + 313, + 529, + 556, + 625 + ], + "type": "text", + "content": ", respectively, they require 18.0-82 times more parameters and at least 39 times more FLOPs. Again, we expect that the performance for VideoMambaPro will increase with more pre-training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "UCF-101/HMDB51/AVA V2.2. From Table 3, it shows that VideoMambaPro-M is competitive, and outperforms VideoMamba by " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " on UCF-101 and HMDB51, respectively. VideoMambaPro-M achieves 31.9 mAP on AVA V2.2, which is " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " lower than VideoMAE V2 [50] but with an order of magnitude fewer parameters and FLOPs and pre-trained only on IN-1K (see Table 4)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24239" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 270, + 178 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 270, + 178 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 270, + 178 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 270, + 178 + ], + "type": "table", + "html": "
MethodParamsUCF-101HMDB51
VideoMoCo [33]15M78.749.2
CoCLR [16]9M81.452.1
MemDPC [15]32M86.154.5
Vi²CLR [4]9M89.155.7
VideoMAE [47]87M91.362.6
GDT [35]33M95.272.8
VideoMAE V2 [50]1050M99.688.1
VideoMamba-M74M88.260.8
VideoMambaPro-M72M91.663.2
", + "image_path": "d285d62d221af0537f4a82fa9470e2657666abc389021f9da8f1bdf87498721b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 209, + 265, + 320 + ], + "blocks": [ + { + "bbox": [ + 93, + 187, + 257, + 198 + ], + "lines": [ + { + "bbox": [ + 93, + 187, + 257, + 198 + ], + "spans": [ + { + "bbox": [ + 93, + 187, + 257, + 198 + ], + "type": "text", + "content": "Table 3. Results on UCF-101 and HMDB51." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 209, + 265, + 320 + ], + "lines": [ + { + "bbox": [ + 86, + 209, + 265, + 320 + ], + "spans": [ + { + "bbox": [ + 86, + 209, + 265, + 320 + ], + "type": "table", + "html": "
MethodFLOPsParammAP
SlowFast R101 [8]138G53M23.8
VideoMAE-B [47]180G87M26.7
MViTv2-B [25]225G51M30.5
ObjectTransformer [54]243G86M31.0
MViTv2-L [25]2828G213M34.4
ST-MAE-H [9]1193G632M36.2
VideoMAE V2 [50]4220G1050M42.6
VideoMamba-M [19]202G74M30.1
VideoMambaPro-M183G72M31.9
", + "image_path": "f2e62904c263e90f6a88724f7ebf48cc52d0d8c876beadd24dd8b2cdb471cff7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 350, + 294, + 406 + ], + "blocks": [ + { + "bbox": [ + 118, + 328, + 232, + 338 + ], + "lines": [ + { + "bbox": [ + 118, + 328, + 232, + 338 + ], + "spans": [ + { + "bbox": [ + 118, + 328, + 232, + 338 + ], + "type": "text", + "content": "Table 4. Results on AVA V2.2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 350, + 294, + 406 + ], + "lines": [ + { + "bbox": [ + 58, + 350, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 58, + 350, + 294, + 406 + ], + "type": "table", + "html": "
ModelsInputTop-1Top-5
VideoMamba-M (baseline)32 × 224282.495.7
VideoMambaPro-M (w/o residual)32 × 224283.6 (+1.2)96.0 (+0.3)
VideoMambaPro-M (w/o masking)32 × 224283.0 (+1.0)95.8 (+0.1)
VideoMambaPro-M32 × 224284.0 (+1.6)96.4 (+0.7)
", + "image_path": "24e1d191be1bc618ae7d3a6dcddcea07b75523d035868ceef99a506010b1da90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "type": "text", + "content": "Table 5. Ablation study on K400, with and without masked backward computation and elemental residual connections." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 147, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 147, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 147, + 470 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 617 + ], + "type": "text", + "content": "Influence of masked backward computation and elemental residual connection. We have identified two limitations that exist in VideoMamba: historical decay and element contradiction. We introduced masked backward computation and elemental residual connections to address these respective issue. Here, we analyze the impact of each solution. We use the same settings as before, with VideoMambaPro-M and pre-training on IN-1K. We summarize the performance of VideoMambaPro-M on K400 in Table 5. Both solutions contribute to an improved score, and their effect is partly complementary. This indicates that the two limitations exist simultaneously in VideoMamba." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Evidence of limitations of Mamba. We have theoretically identified historical decay and element contradiction as fundamental problems in Mamba, and propose the masked backward computation and elemental residual connection to solve the limitations. Previous experiments on image and video action recognition demonstrated the effectiveness of our approach, and we further provide concrete evidence that the increased performance stems from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 429, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 429, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 429, + 83 + ], + "type": "text", + "content": "addressing these two issues." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "text", + "content": "We performed two tests on ImageNet-1K image classification task to build the connection between the issues in practice and our approach. First, to investigate whether the residual connection alleviates element contradiction, we randomly replaced a percentage of the patches by randomly selected patches from other classes, which will increase element contradiction because more irrelevant information is present. The table 6 shows that the relative performance of VideoMambaPro over VideoMamba increases to " + }, + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 84, + 553, + 240 + ], + "type": "text", + "content": " of the patches is replaced. For higher percentages, the gap is reduced along with the overall score. This demonstrates that irrelevant (or even misleading) information can better be dealt with using our approach." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 251, + 552, + 302 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 552, + 302 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 552, + 302 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 552, + 302 + ], + "type": "table", + "html": "
Replacement ratio
051020305080
VideoMamba-Ti76.973.268.059.452.947.628.9
VideoMambaPro-Ti78.978.076.168.655.948.229.1
", + "image_path": "aa5c63bcd7319ecf61a9690a929f5a35edc2cfcf10d2d3250c1add5ec6540b93.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "content": "Second, to verify that masking addresses the issue of historical decay, we adopted a progressive masking approach, gradually decreasing the mask values of the diagonal elements of " + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 313, + 338, + 554, + 422 + ], + "type": "text", + "content": " from 1 to 0, and then retrained the model on ImageNet-1K. The results below show that VideoMambaPro's accuracy gradually increases when more masking is applied, evidencing the merits of the approach." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 353, + 432, + 518, + 472 + ], + "blocks": [ + { + "bbox": [ + 313, + 310, + 553, + 334 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 553, + 334 + ], + "type": "text", + "content": "Table 6. Top-1 accuracy of VideoMambaPro-Ti and VideoMamba-Ti with various replacement ratios (in %) on ImageNet-1K." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 432, + 518, + 472 + ], + "lines": [ + { + "bbox": [ + 353, + 432, + 518, + 472 + ], + "spans": [ + { + "bbox": [ + 353, + 432, + 518, + 472 + ], + "type": "table", + "html": "
Mask value
10.70.50
VideoMambaPro-Ti76.977.077.277.8
", + "image_path": "23987e356e9aa53af9fe33547a3b254732d1f64eec05fd8a1b04b9e966d7bb74.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 479, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 320, + 479, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 545, + 491 + ], + "type": "text", + "content": "Table 7. Top-1 accuracy across mask values on ImageNet-1K." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 504, + 527, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 527, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 527, + 517 + ], + "type": "text", + "content": "5.4. Comparison with VideoMamba on K400" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 568 + ], + "type": "text", + "content": "We more thoroughly compare the differences between VideoMamba and VideoMambaPro by investigating the relative performance per class. We then present a statistical comparison between the results of both backbones." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "text", + "content": "Class analysis. We compare VideoMambaPro-M with " + }, + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "text", + "content": " image size pre-trained on IN-1K to a VideoMamba-M baseline with the same settings. We show the relative performance for all classes of K400 in Figure 2. For over " + }, + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 569, + 553, + 653 + ], + "type": "text", + "content": " of the classes, VideoMambaPro shows improvement. Although there is a lower performance for certain classes, the decrease is typically limited." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "The majority of the classes sees a " + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\sim 1.8\\%" + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": " improvement, which is substantial. For a small number of classes, VideoMambaPro performs " + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "inline_equation", + "content": ">2\\%" + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": " better than VideoMamba. Only a fraction of the classes is negatively affected by the solutions introduced in VideoMambaPro." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24240" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 214 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 214 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 214 + ], + "type": "image", + "image_path": "28cbe6524cbac5873fbb947b49b348fef62046a69a53c8f71d72fff8b7321620.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 295, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 244 + ], + "type": "text", + "content": "Figure 2. Relative accuracy per class on Kinetics-400 by comparing VideoMambaPro-M to VideoMamba-M." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 351 + ], + "type": "text", + "content": "Statistical comparison. In order to understand whether the improvements of VideoMambaPro over VideoMamba are statistically significant, we compare the K400 results of the respective Middle models, both pre-trained on IN-1K and size " + }, + { + "bbox": [ + 55, + 266, + 296, + 351 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 266, + 296, + 351 + ], + "type": "text", + "content": ". Other settings are also the same. For each test sample, we check whether it is correctly classified by either model. The results appear in Table 8." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 360, + 279, + 436 + ], + "blocks": [ + { + "bbox": [ + 73, + 360, + 279, + 436 + ], + "lines": [ + { + "bbox": [ + 73, + 360, + 279, + 436 + ], + "spans": [ + { + "bbox": [ + 73, + 360, + 279, + 436 + ], + "type": "table", + "html": "
VideoMambaPro-M
TrueFalseTotal
VideoMamba-MTrue14,30246914,771
False1,8333,3025,135
Total16,1353,77119,906
", + "image_path": "d0cc73a449a637818985e725ae1a5b2f9d98f128cda0929830777960ea7ac849.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 444, + 295, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 466 + ], + "type": "text", + "content": "Table 8. Contingency table for K400 test items for VideoMamba-M and VideoMambaPro-M." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "content": "We used the McNemar test, a non-parametric test with a single degree of freedom. It checks whether the number of items that are incorrectly classified by VideoMambaPro " + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "inline_equation", + "content": "(n_{10})" + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "content": " but not VideoMamba is substantially lower than the number of items misclassified by VideoMamba but not VideoMambaPro " + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "inline_equation", + "content": "(n_{01})" + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "content": ". The test is calculated as " + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\chi^2 = \\frac{(n_{01} - n_{10})^2}{(n_{01} + n_{10})}" + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "content": ". The resulting value of 808.2 corresponds to a significance level of " + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "inline_equation", + "content": "p < 0.001" + }, + { + "bbox": [ + 55, + 480, + 296, + 627 + ], + "type": "text", + "content": ". We can thus conclude that VideoMambaPro-M is statistically significantly better than VideoMamba-M. Because we relied on the aggregated performance reported in papers for other methods, we cannot report statistical comparisons here." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 201, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 201, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 201, + 649 + ], + "type": "text", + "content": "5.5. Computation cost analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Finally, we compare the performance of VideoMambaPro with various model sizes with other approaches on K400. We map the top-1 to the number of parameters and FLOPs in Figures 3 and 4, respectively. VideoMambaPro performs favorably in terms of the accuracy-compute trade" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 550, + 207 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 550, + 207 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 550, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 550, + 207 + ], + "type": "image", + "image_path": "71a21b66d61b9f0c71b7cc6c12157524b8c07ce69f225eb685dcf1319631d141.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 238 + ], + "type": "text", + "content": "Figure 3. Top-1 accuracy versus number of parameters of VideoMambaPro and other models on Kinetics-400." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 251, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 317, + 251, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 555, + 401 + ], + "type": "image", + "image_path": "ef37a2df50cef53f9627a9b71dade869f27255428064217d2bd48f7c18889190.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 433 + ], + "type": "text", + "content": "Figure 4. Top-1 accuracy versus number of FLOPs of VideoMambaPro and other models on Kinetics-400." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 502 + ], + "type": "text", + "content": "off. Importantly, VideoMambaPro was trained on much less data than other models, which might provide opportunities for further accuracy gains without additional compute and memory requirements." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "From a mathematical comparison with self-attention, we have identified two limitations in how Mamba processes token sequences. We argue that these limitations constrain Mamba's potential, especially in video understanding tasks. To address the two limitations, we have introduced VideoMambaPro (VMP). It takes VideoMamba and introduces the masked backward State Space Model (SSM), and adds residual connections in both forward and backward SSM. In experiments on Kinetics-400, Something-Something V2, HMDB51, UCF-101, and AVA V2.2, VideoMambaPro consistently demonstrates improved performance over the vanilla VideoMamba. We expect that extensive pre-training further elevates the performance of Mamba models for video tasks, making it an increasingly attractive, efficient alternative to large transformer models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24241" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. Vivit: A video vision transformer. In ICCV, pages 6836-6846, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, pages 813-824, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 192 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the Kinetics dataset. In CVPR, pages 6299-6308, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[4] Ali Diba, Vivek Sharma, Reza Safdari, Dariush Lotfi, Saqib Sarfraz, Rainer Stiefelhagen, and Luc Van Gool. Vi2CLR: Video and image for visual contrastive learning of representation. In CVPR, pages 1502-1512, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 304 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 361 + ], + "type": "text", + "content": "[6] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 417 + ], + "type": "text", + "content": "[7] Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object detection. NeurIPS, 34:26183-26197, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 294, + 452 + ], + "type": "text", + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "type": "text", + "content": "[9] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. NeurIPS, 35:35946-35958, 2022. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 553 + ], + "type": "text", + "content": "[10] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The \"Something Something\" video database for learning and evaluating visual common sense. In ICCV, pages 5842-5850, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 555, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 587 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "text", + "content": "[12] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. NeurIPS, 33:1474-1487, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "text", + "content": "[13] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Chunhui Gu, Chen Sun, David A Ross, Carl Vondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. AVA: A video dataset of spatio-temporally localized atomic visual actions. In CVPR, pages 6047-6056, 2018. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Tengda Han, Weidi Xie, and Andrew Zisserman. Memory-augmented dense predictive coding for video representation learning. In ECCV, pages 312-329, 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[16] Tengda Han, Weidi Xie, and Andrew Zisserman. Self-supervised co-training for video representation learning. Advances in neural information processing systems, 33:5679-5690, 2020. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[17] Feyza Duman Keles, Pruthuvi Mahesakya Wijewardena, and Chinmay Hegde. On the computational complexity of selfattention. In International Conference on Algorithmic Learning Theory, pages 597-619. PMLR, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[18] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. HMDB: a large video database for human motion recognition. In ICCV, pages 2556-2563, 2011. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[19] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. UniformerV2: Spatiotemporal learning by arming image vits with video uniformer. arXiv preprint arXiv:2211.09552, 2022. 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[20] Kunchang Li, Yali Wang, Gao Peng, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatial-temporal representation learning. In ICLR, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[21] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In ICCV, pages 19948-19960, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. VideoMamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[23] Shufan Li, Harkanwar Singh, and Aditya Grover. Mamba-nd: Selective state space modeling for multi-dimensional data. In European Conference on Computer Vision, pages 75–92. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[24] Wenrui Li, Xiaopeng Hong, and Xiaopeng Fan. Spikemba: Multi-modal spiking saliency mamba for temporal video grounding. arXiv preprint arXiv:2404.01174, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[25] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MViTv2: Improved multiscale vision transformers for classification and detection. In CVPR, pages 4804-4814, 2022. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "text", + "content": "[26] Qin Liu, Zhenlin Xu, Gedas Bertasius, and Marc Niethammer. SimpleClick: Interactive image segmentation with simple vision transformers. In ICCV, pages 22290-22300, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[27] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. VMamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 647, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 553, + 669 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[29] Hui Lu, Hu Jian, Ronald Poppe, and Albert Ali Salah. Enhancing video transformers for action understanding with vlm-aided training. arXiv preprint arXiv:2403.16128, 2024.1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24242" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[30] Jiachen Lu, Jinghan Yao, Junge Zhang, Xiatian Zhu, Hang Xu, Weiguo Gao, Chunjing Xu, Tao Xiang, and Li Zhang. Soft: Softmax-free transformer with linear complexity. NeurIPS, 34:21297-21309, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[31] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[32] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In ECCV, pages 1-18, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 240 + ], + "type": "text", + "content": "[33] Tian Pan, Yibing Song, Tianyu Yang, Wenhao Jiang, and Wei Liu. VideoMoCo: Contrastive video representation learning with temporally adversarial examples. In CVPR, pages 11205-11214, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[34] Jinyoung Park, Hee-Seon Kim, Kangwook Ko, Minbeom Kim, and Changick Kim. Videomamba: Spatio-temporal selective state space model. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 331 + ], + "type": "text", + "content": "[35] Mandela Patrick, Yuki Markus Asano, Ruth Fong, João F. Henriques, Geoffrey Zweig, and Andrea Vedaldi. Multimodal self-supervision from generalized data transformations. arXiv preprint arXiv:2003.04298, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[36] Mandela Patrick, Dylan Campbell, Yuki Asano, Ishan Misra, Florian Metze, Christoph Feichtenhofer, Andrea Vedaldi, and Joao F Henriques. Keeping your eye on the ball: Trajectory attention in video transformers. Advances in neural information processing systems, 34:12493-12506, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[37] Badri Narayana Patro and Vijay Srinivas Agneeswaran. Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges. arXiv preprint arXiv:2404.16112, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[38] Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, et al. RWKV: Reinventing RNNs for the transformer era. arXiv preprint arXiv:2305.13048, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[39] AJ Piergiovanni, Weicheng Kuo, and Anelia Angelova. Rethinking video ViTs: Sparse video tubes for joint image and video learning. In CVPR, pages 2214-2224, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[40] Hongyu Ren, Hanjun Dai, Zihang Dai, Mengjiao Yang, Jure Leskovec, Dale Schuurmans, and Bo Dai. Combiner: Full attention transformer with sparse computation cost. NeurIPS, 34:22470-22482, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[41] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[42] Gilad Sharir, Asaf Noy, and Lihi Zelnik-Manor. An image is worth 16x16 words, what is a video worth? arXiv preprint arXiv:2103.13915, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[43] Yuanyuan Shen, Edmund M-K Lai, and Mahsa Mohagheh. Effects of similarity score functions in attention mechanisms on the performance of neural question answering systems. Neural Processing Letters, 54(3):2283-2302, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[44] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "text", + "content": "[45] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In ICCV, pages 7262-7272, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[46] Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, and Furu Wei. Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[47] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[48] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[50] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. VideoMAE V2: Scaling video masked autoencoders with dual masking. In CVPR, pages 14549-14560, 2023. 1, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 431 + ], + "type": "text", + "content": "[51] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. InternVideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[52] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. InternVideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 533 + ], + "type": "text", + "content": "[53] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, pages 14668-14678, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 622 + ], + "type": "text", + "content": "[55] Shen Yan, Xuehan Xiong, Anurag Arnab, Zhichao Lu, Mi Zhang, Chen Sun, and Cordelia Schmid. Multiview transformers for video recognition. In CVPR, pages 3333-3343, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[56] Hanwei Zhang, Ying Zhu, Dan Wang, Lijun Zhang, Tianxiang Chen, and Zi Ye. A survey on Visual Mamba. arXiv preprint arXiv:2404.15956, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[57] Zheng Zhang and Kil To Chong. Comparison between first-order hold with zero-order hold in discretization of input-delay nonlinear systems. In 2007 International Conference on Control, Automation and Systems, pages 2892-2896, 2007. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24243" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[58] Zixiao Zhang, Xiaogiang Lu, Guojin Cao, Yuting Yang, Licheng Jiao, and Fang Liu. ViT-YOLO: Transformer-basedYOLO for object detection. In ICCV, pages 2799-2808, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[59] Chen Zhu, Wei Ping, Chaowei Xiao, Mohammad Shoeybi, Tom Goldstein, Anima Anandkumar, and Bryan Catanzaro. Long-short transformer: Efficient transformers for language and vision. NeurIPS, 34:17723-17736, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 205 + ], + "type": "text", + "content": "[60] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. BiFormer: Vision transformer with bilevel routing attention. In CVPR, pages 10323-10333, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[61] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision Mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2, 4" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24244" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_content_list.json b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d9dc868549941268c6ae4bc369848adb50012a6 --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_content_list.json @@ -0,0 +1,1844 @@ +[ + { + "type": "text", + "text": "Social Debiasing for Fair Multi-modal LLMs", + "text_level": 1, + "bbox": [ + 269, + 130, + 725, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Harry Cheng $^{1}$ , Yangyang Guo $^{2}$ , Qingpei Guo $^{3,4,*}$ , Ming Yang $^{4}$ , Tian Gan $^{1}$ , Weili Guan $^{5}$ , Liqiang Nie $^{5}$ $^{1}$ Shandong University, $^{2}$ National University of Singapore, $^{3}$ University of Warwick, $^{4}$ Ant Group, $^{5}$ Harbin Institute of Technology (Shenzhen)", + "bbox": [ + 101, + 179, + 895, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xaCheng1996@gmail.com, guoyang.eric@gmail.com, qingpei.gqp@antgroup.com, m.yang@antgroup.com, gantian@sdu.edu.cn, honeyguan@gmail.com, nieliqiang@gmail.com", + "bbox": [ + 130, + 237, + 859, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks.", + "bbox": [ + 88, + 335, + 485, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 684, + 220, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-modal Large Language Models (MLLMs) have revolutionized the field of general-purpose vision-language understanding. Representative models, such as LLaVA [38], Qwen-VL [3], and Bunny [25], exhibit remarkable zero-shot performance and can be easily fine-tuned for diverse downstream applications.", + "bbox": [ + 89, + 710, + 482, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite MLLMs' widespread use, it is imperative to recognize that these models can exhibit severe social biases with respect to attributes such as race and gender [18, 65]. Figure 1a illustrates one example that a biased MLLM is", + "bbox": [ + 89, + 801, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65d3a9cf4c9d75841914648b70a3c4a0ec34c40ef09447d883cfe1466deba406.jpg", + "image_caption": [ + "Figure 1. Social bias examples. Left: Gender-biased prediction on Nurse from LLaVA-7B1. Right: Maximum Skews across different human attributes on the FairFace dataset from Qwen-VL-7B and LLaVA-7B. As a metric for measuring model bias, a larger Skew implies a higher degree of social bias. Our CSD method outperforms existing approaches by a significant margin." + ], + "image_footnote": [], + "bbox": [ + 535, + 301, + 880, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "much more likely to associate the role of nurse with female rather than male, which may be stereotypical. Such biased predictions often occur unconsciously within MLLMs, making them difficult to detect and avoid by explicit rules. A key reason for this issue lies in the composition of the MLLM training data, which can contain content related to violence or racism [6]. These inappropriate samples may collectively demonstrate some stereotypes, resulting in subtle yet uncomfortable responses [8, 63].", + "bbox": [ + 511, + 506, + 905, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing studies on mitigating the social bias problem in MLLMs remain largely under-explored. A naive approach is to collect attribute-balanced vision-language counterfactual datasets [24, 34], which can be used to directly finetune biased models to predict fairer distributions. However, as demonstrated in Table 1, existing large-scale datasets are limited by their focus on a single social concept, such as occupation [27], without attending to multifaceted social stereotypes [62]. This significantly hampers the model in learning more comprehensive representations. Therefore, we are motivated to construct a more extensive and diverse counterfactual dataset. From a methodological perspective, directly fine-tuning on a counterfactual dataset assigns equal importance to instances receiving different social biases, resulting in sub-optimal debiasing performance (as demonstrated in Section 6.2). Analogy to a common solution in chemistry—neutralizing acidic water requires", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 107, + 875, + 235, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We manually visualize the bounding boxes from LLaVA's text output.", + "bbox": [ + 109, + 887, + 480, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1740", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/fd60fc8597ff847691196f8a2cf8e7cce02d63ac56848780835ed19f3304e899.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetVenueType#ImagesSocial Concepts
CoCo-Counterfactuals [34]NeurIPS'23General34KNo annotations
FairFace [33]WACV'21Social108KNo annotations
VisoGender [24]NeurIPS'23Social0.6KOccupation
PATA [55]CVPR'23Social5KOccupation
SocialCounterfactuals [27]CVPR'24Social171KOccupation
MM-Bias [31]EACL'23Social3K14 minorities
CMSC (Ours)-Social60K18 balanced concepts
", + "bbox": [ + 101, + 88, + 470, + 185 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of CMSC with mainstream datasets used for model debiasing. Existing datasets are limited either by scale or by the number of social concepts they cover. Type - The category of biases the dataset focuses on.", + "bbox": [ + 89, + 188, + 482, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adding an alkaline substance rather than plain water—we are prompted to speculate: Can we leverage the opposite of the suffered social bias to build a fairer model?", + "bbox": [ + 89, + 253, + 483, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these two issues, we first construct a large-scale, high-quality Counterfactual dataset with Multiple Social Concepts (CMSC). As demonstrated in Table 1, our CMSC dataset narrows the gap in data scale and concept richness. We conduct extensive experiments on four different MLLM architectures [3, 25, 38] using several existing counterfactual datasets as well as our CMSC. The results demonstrate that MLLMs fine-tuned on CMSC exhibit significantly lower social bias compared to those fine-tuned on single-concept datasets. For instance, LLaVA-7B [38] fine-tuned on CMSC achieves a $64\\%$ debiasing effect, which significantly outperforms the $30\\%$ bias reduction achieved when fine-tuned on the SocialCounterfactuals dataset $[27]^2$ . This indicates that fine-tuning the models with a rich set of social concept is beneficial in MLLM debiasing.", + "bbox": [ + 91, + 300, + 483, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to implement the idea of 'debiasing with the opposite of the social biases', we propose a Counter-Stereotype Debiasing strategy (CSD) to effectively reduce social biases in MLLMs. Our CSD is equipped with two techniques: 1) we design a novel data sampling method based on the bias level, 2) we rescale the previously used autoregressive loss function to a new Social Fairness Loss (SFLoss). In this manner, previously under-represented cases, e.g., male nurses, will receive more attention, serving as new 'alkaline' ones to counteract the bias of MLLMs that prefer female nurses. The experimental results demonstrate that our CSD method is an effective debiasing strategy. As illustrated in Figure 1b, our method reduces the bias by over $50\\%$ for the Qwen-VL-7B model compared to prior training strategies, especially naive fine-tuning. Additionally, CSD does not compromise the original model performance on general multi-modal benchmarks.", + "bbox": [ + 89, + 527, + 482, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are three-fold:", + "bbox": [ + 109, + 784, + 408, + 797 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a high-quality counterfactual dataset that includes 18 social concepts, which demonstrated to be superior to existing ones on MLLM social debiasing.", + "- We propose a novel approach that applying counter-stereotype debiasing to mitigate social biases. To the best" + ], + "bbox": [ + 91, + 800, + 482, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of our knowledge, this is the first research effort dedicated to addressing the social bias problem in autoregressive MLLMs.", + "bbox": [ + 524, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We apply our debiasing strategy to four multi-modal LLMs architectures to show the generalization capability and effectiveness of our method.", + "bbox": [ + 513, + 136, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 196, + 653, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-modal Large Language Model", + "text_level": 1, + "bbox": [ + 511, + 223, + 830, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rapid development of LLMs [9, 51, 57, 60, 64], increasing efforts have been dedicated to extending the powerful reasoning capabilities of LLMs to multi-modal applications [1, 10, 14, 28, 32, 35, 40, 42, 58, 70, 71]. Specifically, MLLMs utilize LLMs as the foundational base, aligning features from other modalities with text embeddings to enable LLMs to perceive multi-modal inputs. For instance, BLIP-2 [36] utilizes a Q-Former to align vision encoders with LLMs. LLaVA [38, 39] presents to directly map vision features to the word embedding space of Vicuna [11]. Qwen-VL [3] employs a single-layer cross-attention module as the vision-language adapter, and introduces multitask pre-training to improve the model performance. These approaches significantly enhance multi-modal understanding capabilities. However, some studies highlight the existence of social biases in pre-trained LLMs [7, 8, 18]. These biases often manifest as harmful spurious correlations related to human attributes such as gender and race, which significantly undermine the fairness. For instance, LLMs may 'efficiently' filter resumes based on race rather than the candidates' qualifications [61]. Nevertheless, research on social biases in MLLMs remains largely unexplored [68].", + "bbox": [ + 511, + 244, + 906, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Social Bias Reduction", + "text_level": 1, + "bbox": [ + 513, + 590, + 718, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We roughly categorize the bias mitigation strategies into two groups: data-based and objective-based.", + "bbox": [ + 511, + 612, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data-based debiasing typically refers to data augmentation techniques. For instance, Chuang et al. [12] employ mixup [69] to construct interpolated samples among groups with different distributions. Ramaswamy et al. [52] utilize perturbed GAN-generated images [21] in latent space to augment the original dataset. Moreover, some studies collect fair datasets across human attributes for continual finetuning [5, 24, 31, 34, 54, 72, 75]. In particular, FairFace [33] contains $108\\mathrm{K}$ images that are balanced for the race attribute. Socialcounterfactuals [27] collects $171\\mathrm{K}$ imagetext pairs to probe biases across race, gender, and physical characteristics. However, these collected counterfactual datasets often focus on only one common concept - occupation, which hinders effective debiasing [67]. Though a few datasets involve several concepts, they nevertheless, are largely limited by their data scale (e.g., 3K [31]), making it less feasible to fine-tune MLLMs.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2In this paper, we interchangeably use debiasing and bias reduction.", + "bbox": [ + 107, + 886, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1741", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37ca1d70d0659a45d23a2014baf3a63c47507e136d54b8357619328d3c7eb611.jpg", + "image_caption": [ + "Figure 2. Image Generation Pipeline. Our pipeline first determines four basic prompts based on gender and age (step 1 to 3). Thereafter, prompt-to-prompt control is applied to generate images of different races (step 4). Finally, we filter out low-quality images (step 5)." + ], + "image_footnote": [], + "bbox": [ + 173, + 89, + 823, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective-based debiasing modifies the model training process to achieve better fairness. Early studies are proposed to reduce social bias within uni-modal models [2, 17, 19, 44, 45, 66]. For example, Bolukbasi et al. [7] optimizes word embeddings to remove gender stereotypes. In contrast, multi-modal debiasing methods often focus on contrastive learning-based vision language models, e.g., CLIP [50]. Representative approaches are to rectify the modality similarity matrix in VLMs [15] via learning additive adapters [55, 73], eliminating biased directions [13], and using adversarial samples [4]. However, among these methods, uni-modal approaches often require re-training the full model. This operation, though plausible for previous small-scale models, is less practical for existing MLLMs with billions of parameters [53]. On the other hand, mainstream multi-modal techniques cannot be utilized for these models due to the divergent training objectives between MLLMs and previous CLIP-style models. In particular, MLLMs are mostly trained in an autoregressive way rather than through contrastive learning.", + "bbox": [ + 91, + 243, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CMSC Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 561, + 356, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As demonstrated in Table 1, existing counterfactual datasets are either restricted by their small scale [24] or a narrow coverage of concepts [33]. To bridge this gap between limited existing datasets and diverse real-world stereotypes, we introduce the CMSC dataset, encompassing 60k high-quality images across eighteen social concepts.", + "bbox": [ + 89, + 587, + 482, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Social Attributes and Concepts", + "text_level": 1, + "bbox": [ + 89, + 689, + 366, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A counterfactual dataset for social bias reduction and evaluation typically contains two key aspects: Social Attribute (SA) and Social Concept (SC). The former, i.e., SA, is defined as characteristics shared by a group of people [33]. Specifically, we investigate three types of attributes: i) genders: {Male, Female}; ii) races: {White, Black, Indian, Southeast Asian, East Asian, Middle Eastern, Latino}; and iii) ages: {Young, Old}. These attributes are intrinsic to individuals3. For each image $\\mathbf{I}_i$ in CMSC, three SA labels are provided with respect to these three types of SA, and are", + "bbox": [ + 89, + 712, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "combined as a SA set $\\mathcal{A}_i$ .", + "bbox": [ + 511, + 244, + 684, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As for the SC, we define it as the societal label attributed to an individual. Drawing from the sociological research [16, 46, 48, 59, 74], we employ 18 SCs in CMSC. In particular, these SCs are categorized into three groups: personality, responsibility, and education. Specifically, personality relates to concepts pertaining to an individual's character. We use five concepts: {compassionate, belligerent, authority, pleasant, unpleasant}. Responsibility is the roles or duties that individuals are expected to fulfill in society or family. We identify six concepts: {tool user, weapon user, career, family, chef working, earning money}. Education pertains to the level of education a person has received. We include seven concepts: {middle school, high school, university, good student, bad student, science, arts}. Each image in CMSC is annotated with one SC label $c_i \\in \\mathcal{C}$ , where $\\mathcal{C}$ is the union of the concepts from all above groups.", + "bbox": [ + 511, + 260, + 906, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, each instance $\\mathcal{P}_i$ in our dataset consists of an image $\\mathbf{I}_i$ , a set of SA labels $\\mathcal{A}_i$ , and a SC label $c_i$ . To the best of our knowledge, CMSC is the first large-scale counterfactual dataset that includes a variety of social concepts.", + "bbox": [ + 511, + 502, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image Generation Pipeline", + "text_level": 1, + "bbox": [ + 511, + 575, + 756, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ Stable Diffusion XL (SDXL) [49] to generate instances in CMSC. For each SC, we carefully design prompts that provide a detailed explanation of the concept. These prompts not only help eliminate potential ambiguities associated with the SC but also enhance the quality of the generated images [29]. For instance, for the SC 'pleasant,' we use the prompt: 'A photo of a pleasant [race] [gender] person at [age], who has a friendly smile, a relaxed posture, and ...' To ensure diverse and representative image generation, these prompts are randomly modified by replacing words with synonyms during image generation. Subsequently, the intersectional generation strategy [27] is applied. As illustrated in Figure 2, we first determine gender and age (Step 2) and generate four base images (Step 3). Thereafter, we apply Prompt-to-Prompt (P2P) control [26] to produce visually similar images that vary only in terms of race (Step 4). For each SC, we create 28 prompts across attribute combinations. Each prompt helps generate 100 images. To ensure image quality, we apply a rigorous image filtering mechanism (step 5). First, we use CLIP-Score [50]", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3All SAs are perceived, made by human annotators or models. We acknowledge that the SAs are not representative of all people [27, 33].", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1742", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to filter out images that are semantically misaligned with the prompts. Subsequently, a panel of three experts conducts a voting process to remove low-quality and NSFW content. If the generated images perform poorly on a specific SA, the entire set of images (including those of other SAs) will be discarded. Moreover, we observe that the model can generate high-quality images for certain SCs or SAs, while struggling to do so for others. As a result, we filter out excessively generated ones for balanced distribution. Our filtering process is very strict. For instance, the testing set initially contains over 50,000 images, but only 12,019 images are finally used in CMSC. A similar process is applied to the training set, which comprises 48,134 images. In total, our dataset contains 60K images.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. CMSC Dataset Evaluation", + "text_level": 1, + "bbox": [ + 89, + 316, + 336, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 89, + 342, + 284, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Protocols. We employed four different MLLM architectures, i.e., LLaVA-7B [38], LLaVA-13B [38], Qwen-VL-7B [3], and Bunny-8B [25], to evaluate our CMSC dataset. For each MLLM, we used the same template to collect their responses: 'Based on the race, gender, and age of the face , do you think the person is {SCs}? Respond with one of them.' Here, is the image from the dataset, and {SCs} represents the list of used SCs. For instance, when testing personality, {SCs} includes 'pleasant,' 'unpleasant,' etc. After collecting these predictions, we used them to compute MaxSkew@C and MinSkew@C (which will be elaborated in the following section) to measure bias.", + "bbox": [ + 89, + 364, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Metrics. We employ Skew-based [20] metrics to quantify the extent of social biases. Specifically, for a given dataset $\\mathcal{D}$ , we define the subset containing instances with a specific SC label $c'$ as $\\mathcal{D}_{c'}$ ,", + "bbox": [ + 89, + 561, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D}, c _ {i} = c ^ {\\prime} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 635, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From $\\mathcal{D}_{c'}$ , we further extract the subset consisting of instances with a specific SA label $a'$ as $\\mathcal{D}_{a'|c'}$ ,", + "bbox": [ + 89, + 665, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D} _ {c ^ {\\prime}}, a ^ {\\prime} \\in \\mathcal {A} _ {i} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 708, + 482, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then utilize the aforementioned template to guide MLLMs in predicting the SC label corresponding to each $\\mathcal{P}_i$ . We denote the predicted SC label as $\\hat{c}_i$ . By applying this process to the entire dataset $\\mathcal{D}$ , we can construct a new predicted set $\\hat{\\mathcal{D}} = \\{\\hat{\\mathcal{P}}_i\\}_{i=1}^N$ , where $\\hat{\\mathcal{P}}_i = \\{\\mathbf{I}_i, \\mathcal{A}_i, \\hat{c}_i\\}$ , and $N = |\\mathcal{D}|$ . In $\\hat{\\mathcal{D}}$ , subsets $\\hat{\\mathcal{D}}_{c'}$ and $\\hat{\\mathcal{D}}_{a'|c'}$ can be derived with Equation (1) and Equation (2), respectively. The Skew for SC $c'$ and SA $a'$ could be formulated as", + "bbox": [ + 89, + 738, + 483, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S k e w} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\log \\left(\\frac {\\hat {\\gamma} _ {a ^ {\\prime}} \\mid c ^ {\\prime}}{\\gamma_ {a ^ {\\prime}} \\mid c ^ {\\prime}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 869, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f68790acb97a2aa051ef1aec18b1d1c6c30527132566eb54a8f59f72122c616d.jpg", + "image_caption": [ + "Figure 3. Performance of MLLMs when evaluated on FairFace. In each model, the bars labeled 'FT w SCounter' and 'FT w CMSC' represent the model fine-tuned on SocialCounterfactuals and our CMSC, respectively." + ], + "image_footnote": [], + "bbox": [ + 544, + 89, + 874, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 339, + 558, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {\\gamma} _ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\hat {\\mathcal {D}} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\hat {\\mathcal {D}} _ {c ^ {\\prime}} \\right|}, \\\\ \\gamma_ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\mathcal {D} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\mathcal {D} _ {c ^ {\\prime}} \\right|}. \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 349, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $\\mathrm{Skew}_{a'|c'} > 0$ , the MLLM tends to predict instances with attribute $a'$ as concept $c'$ . For instance, in Figure 1a, $\\mathrm{Skew}_{\\text{Female|Nurse}} > 0$ . In contrast, when $\\mathrm{Skew}_{a'|c'} < 0$ , the MLLM is inclined to not predict these instances containing SA $a'$ as SC $c'$ , e.g., predicting male as nurse. A fair MLLM should have Skew close to 0 across all concepts and attributes, i.e., $\\sum_{a,c} |\\mathrm{Skew}_{a|c}| \\to 0$ .", + "bbox": [ + 511, + 405, + 905, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although Skew can effectively measure MLLMs' bias towards a particular SA-SC pair, a counterfactual dataset often contains hundreds of such combinations, e.g., our CMSC includes 198 SA-SC pairs. This confounds the comprehensive assessment of MLLM bias. Therefore, we propose two variants of Skew to enable a more thorough analysis. Specifically, we first identify the maximum and minimum Skew values for each SC across all SAs [13],", + "bbox": [ + 511, + 512, + 905, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\operatorname {M a x S k e w} _ {c ^ {\\prime}} = \\operatorname {M a x} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}, \\\\ \\operatorname {M i n S k e w} _ {c ^ {\\prime}} = \\operatorname {M i n} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}. \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 643, + 903, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For all $c' \\in \\mathcal{C}$ , we separately calculate the average of $\\mathrm{MaxSkew}_{c'}$ and $\\mathrm{MinSkew}_{c'}$ , resulting in two aggregated Skew values. We refer to these two values as MaxSkew@C (MaxS@C) and MinSkew@C (MinS@C), which represent the overall bias level of the MLLM. Both of these metrics indicate better fairness as they approach zero.", + "bbox": [ + 511, + 688, + 905, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We compared our dataset with the previous state-of-the-art counterfactual dataset. Specifically, we separately fine-tuned MLLMs on CMSC and SocialCounterfactual [27] to evaluate their effectiveness in mitigating bias. The SocialCounterfactual dataset focuses on the SC of occupation and contains 171K high-quality synthetic images. Moreover, we employed the FairFace dataset [33], which contains 108K images with a balanced distribution of SAs,", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1743", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/90700cb2f6e6b41caa8f0d988170e5d3322848f98c8560bccd1c2502495be328.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FT w/ SCPer.Res.Edu.
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA×-0.94862.4950-0.76622.2188-0.85693.9821
+ FT Per.-0.84691.6675-0.59152.2025-0.60943.1794
+ FT Res.-0.91521.7409-0.49211.4397-0.81582.7705
+ FT Edu.-0.97561.7842-0.68721.5719-0.52572.4362
Qwen.VL×-2.46882.5772-2.53012.2663-2.82422.4059
+ FT Per.-0.73861.2509-1.90770.3470-2.55242.3506
+ FT Res.-0.73841.9044-1.32050.1202-2.85452.3872
+ FT Edu.-0.84311.6817-1.87590.2236-2.47062.3502
", + "bbox": [ + 107, + 88, + 464, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Performance comparison of LLaVA-7B and Qwen-VL-7B when fine-tuned and tested on SCs in CMSC. Per.: Personality, Res.: Responsibility, Edu.: Education, $\\times$ : Original performance.", + "bbox": [ + 89, + 198, + 482, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as an additional testing set to assess the cross-dataset debiasing performance of models fine-tuned on CMSC and SocialCounterfactual [55].", + "bbox": [ + 89, + 257, + 482, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison on Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 316, + 341, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We reported MinSkew@C and MaxSkew@C of several MLLMs in Figure 3. The models, i.e., LLaVA [38], Qwen-VL [3], and Bunny [25], are fine-tuned on the SocialCounterfactuals and CMSC datasets individually and evaluated on the FairFace dataset, respectively. For clarity, we took the absolute values of both metrics. From Figure 3, we observed that models fine-tuned on CMSC exhibit superior debiasing effects. For instance, LLaVA-13B achieves a MinSkew@C of -1.3132, a significant advantage over that from the model fine-tuned on SocialCounterfactuals. This implies that the broader range of social concepts in CMSC enables the model to learn fairer distributions.", + "bbox": [ + 89, + 340, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Fine-tuning with Specific SC Group", + "text_level": 1, + "bbox": [ + 89, + 536, + 401, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Section 3.1, we split the SCs in CMSC into three groups: personality, responsibility, and education. We applied LLaVA and Qwen-VL to fine-tune separately on these subsets. The results are reported in Table 2. It can be observed that in intra-subset evaluations, i.e., fine-tuned and tested on the same subset, the model can generally achieve lower biases. For instance, when fine-tuning and evaluating Qwen-VL on the personality subset, it achieves a MaxS@C of 1.2509, showing an absolute difference of 0.65 and 0.43 compared to performance when fine-tuned on the responsibility and education subsets, respectively.", + "bbox": [ + 89, + 560, + 483, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Comparison on Image Distribution", + "text_level": 1, + "bbox": [ + 89, + 739, + 395, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We calculated the Fréchet Inception Distance (FID) scores for the synthetic counterfactual datasets SocialCounterfactuals and our CMSC. Specifically, we randomly sampled 1,000 synthetic images from each dataset and then computed the distributional differences with the same set of 1,000 real images. A lower FID signifies better image quality. With this process, the two datasets received FID scores of 27.17 and 24.35, respectively. This indicates that the images in our dataset bear a closer resemblance to reality.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Counter-stereotype Debiasing", + "text_level": 1, + "bbox": [ + 513, + 89, + 787, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 114, + 653, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before introducing our proposed method, we first revisit the training objectives of MLLMs. Subsequently, we introduce the concept of $\\mathrm{Skew}(\\mathcal{P}_i)$ to measure stereotypes associated with specific instances, upon which we base the reformulation of the autoregressive fine-tuning paradigm of MLLMs.", + "bbox": [ + 511, + 136, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Traning Objective of MLLMs", + "text_level": 1, + "bbox": [ + 511, + 219, + 767, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Current mainstream MLLMs employ pre-trained LLMs [11, 60] as the output interface [3, 25, 38, 43]. Under this context, the base LLM is often trained in an autoregressive way,", + "bbox": [ + 511, + 238, + 905, + 284 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} \\left(\\mathbf {x} _ {1: T}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {< t}; \\theta\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 292, + 905, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is a text sequence with $T$ tokens, and $\\theta$ denotes the model parameters. During pre-training, the LLM predicts the $t$ -th token $\\mathbf{x}_t$ based on all preceding ones, i.e., $\\mathbf{x}_{< t}$ . After that, another instruction tuning stage [47] is often followed to enable LLMs to better understand user intentions,", + "bbox": [ + 511, + 344, + 905, + 420 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}; \\theta\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 429, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{\\mathrm{ins}}$ and $\\mathbf{y}$ are textual instructions and responses, respectively. MLLMs extend the inputs of LLMs with the supplement of image features $\\mathbf{x}_{\\mathrm{img}}$ that are extracted using pre-trained vision encoders such as ViT [30, 50],", + "bbox": [ + 511, + 479, + 905, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 551, + 905, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_{\\mathrm{img}}$ is aligned with text features via a connector, e.g., a trainable projection matrix. To optimize the pipeline, previous cross-entropy loss from LLMs is directly inherited,", + "bbox": [ + 511, + 602, + 905, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {y}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta) = - \\sum_ {t = 1} ^ {T} \\log \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 657, + 905, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.2. Stereotype Measurement", + "text_level": 1, + "bbox": [ + 513, + 715, + 733, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that $\\mathrm{Skew}_{a|c}$ quantifies the degree of social biases in an MLLM across the entire dataset. In this section, we are more interested in the social bias degree for each specific instance $\\mathcal{P}_i$ . We then define $\\mathrm{Skew}(\\mathcal{P}_i)$ as a selected $\\mathrm{Skew}_{a|c}$ with the maximum absolute value across all SAs for instance $\\mathcal{P}_i$ . For example, suppose that $\\mathcal{P}_i$ contains three SA labels: {'White,' Female,' 'Young'}, along with one SC label: 'Nurse.' $\\mathrm{Skew}(\\mathcal{P}_i)$ is the one with the highest absolute value from $\\{\\mathrm{Skew}_{White|Nurse}, \\mathrm{Skew}_{Female|Nurse}, \\mathrm{Skew}_{Young|Nurse}\\}$ . It describes the SA that receives the most severe bias, thereby indicating the level of bias in $\\mathcal{P}_i$ .", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1744", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad79bf07f31990abf72184d663161dfa8b9fbb59190a858fbacfa135387412c4.jpg", + "image_caption": [ + "Figure 4. Comparison between fine-tuning (upper part) and our CSD (lower part). CSD adjusts the loss and datasets based on $\\mathrm{Skew}(\\mathcal{P}_i)$ , guiding the model to focus on overlooked instances." + ], + "image_footnote": [ + ": Parameter updates based on a specific loss function." + ], + "bbox": [ + 122, + 88, + 447, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. CSD Approach", + "text_level": 1, + "bbox": [ + 89, + 292, + 243, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, vanilla fine-tuning approach involves sampling images from a balanced dataset and updating MLLMs through the original autoregressive objective. We argue that this method, which treats all instances equally, is ineffective in addressing the social bias problem in MLLMs. Therefore, we propose an CSD method from the view of counter-stereotype. Specifically, our CSD is composed of two components: i) Dataset Resampling, which enhances the data sampling process to include more underrepresented instances. ii) Loss Rescaling, where we adjust the loss function to place larger emphasis on instances that are overlooked in terms of social attributes.", + "bbox": [ + 89, + 314, + 483, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Resampling. The debiasing approach, e.g., directly fine-tuning, utilizes counterfactual datasets that are balanced across all SAs. We believe that such datasets make it difficult for MLLMs to recognize which SAs are subjected to greater bias and which are not4. To address this, we resample the dataset to increase the frequency of instances that MLLMs tend to ignore and reduce the frequency of emphasized ones, thereby 'neutralizing' social biases.", + "bbox": [ + 88, + 497, + 483, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\operatorname{Skew}(\\mathcal{P}_i)$ serves as an indicator for dataset resampling. As demonstrated in Algorithm 1, for each instance $\\mathcal{P}_i$ , when $\\operatorname{Skew}(\\mathcal{P}_i) > 0$ , i.e., $\\mathcal{P}_i$ has received more attention – such as the 'female-nurse' in Figure 1 – we reduce its probability in the resampled dataset $\\mathcal{D}_r$ for the following training epoch,", + "bbox": [ + 89, + 618, + 483, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {r} = \\left\\{ \\begin{array}{l} \\mathcal {D} _ {r} \\cup \\mathcal {P} _ {i}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) > \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\\\ \\mathcal {D} _ {r}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) \\leq \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 705, + 483, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\operatorname{rand}(\\cdot)$ is to randomly draw from 0 to $\\operatorname{Skew}(\\mathcal{P}_i) + \\tau_1$ , and $\\tau_{1}$ is a pre-defined threshold. As such, a larger $\\operatorname{Skew}(\\mathcal{P}_i)$ corresponds to a lower chance of being included by $\\mathcal{D}_r$ . In contrast, for instances with $\\operatorname{Skew}(\\mathcal{P}_i) \\leq 0$ , we believe that increasing their proportion in the new dataset $\\mathcal{D}_r$ is beneficial for the MLLM to learn the features of these overlooked parts, thereby achieving a fairer distribution.", + "bbox": [ + 89, + 757, + 483, + 864 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Dataset Resampling with Skew $(\\mathcal{P}_i)$", + "Input: The original dataset $\\mathcal{D} = \\{\\mathbf{I}_i, \\mathcal{A}_i, c_i\\}$ , the social attribute sets $\\mathcal{A}$ , the social concept set $\\mathcal{C}$", + "Output: The resampled dataset $\\mathcal{D}_r$" + ], + "code_body": "1: $\\mathcal{D}_r\\gets []$ ,AcmSkew $a_{i}|c_{i}\\gets 0$ ; Initialization \n2: for $\\mathcal{P}_i\\in \\mathcal{D}$ do \n3: if Skew $(\\mathcal{P}_i) > 0$ then \n4: if Rand(0,Skew $(\\mathcal{P}_i) + \\tau_1) > \\mathrm{Skew}(\\mathcal{P}_i)$ then \n5: $\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;$ \n6: end if \n7: else $\\triangleright$ Skew $(\\mathcal{P}_i)\\leq 0$ \n8: $\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;$ \n9: AcmSkew $a_{i}|c_{i}\\gets$ AcmSkew $a_{i}|c_{i} + |\\mathrm{Skew}(\\mathcal{P}_i)|;$ \n10: if AcmSkew $a_{i}|c_{i} > \\tau_{2}$ then \n11: $\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;$ Over-resampling \n12: AcmSkew $a_{i}|c_{i}\\gets 0$ \n13: end if \n14: end if \n15: end for \n16: return $\\mathcal{D}_r$", + "bbox": [ + 522, + 157, + 906, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, these instances are directly accepted into $\\mathcal{D}_r$ . Moreover, we design an over-resampling mechanism to further increase the occurrence frequency of these instances. Specifically, for each $\\mathcal{P}_i$ , we employ an accumulative value, AcmSkew, to gradually accumulate the current $\\mathrm{Skew}(\\mathcal{P}_i)$", + "bbox": [ + 511, + 429, + 906, + 506 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} = \\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} + | \\operatorname {S k e w} (\\mathcal {P} _ {i}) |, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 517, + 906, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $c_{i}$ and $a_{i}$ are SC label and SA label corresponding to $\\operatorname{Skew}(\\mathcal{P}_i)$ , respectively. When $\\operatorname{AcmSkew}_{a_i|c_i}$ exceeds a threshold $\\tau_2$ , we add the $\\mathcal{P}_i$ into $\\mathcal{D}_r$ again. The $\\operatorname{AcmSkew}_{a_i|c_i}$ can then be set to 0 for a new round of accumulation. With the above operations, instances with $\\operatorname{Skew}(\\mathcal{P}_i) \\leq 0$ will be resampled multiple times. This resampling process is executed once before each training epoch. For evaluation, we employ the balanced testing set to ensure a fair comparison.", + "bbox": [ + 511, + 547, + 906, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Rescaling. During MLLM fine-tuning, we rescale the autoregressive loss in Equation (9) to a new Social Fairness Loss (SFLoss) for effective debiasing. Specifically, the empirical risk during training can be represented as", + "bbox": [ + 511, + 683, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\mathrm {i n s}} ^ {i}, \\mathbf {x} _ {\\mathrm {i m g}} ^ {i}; \\theta\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 755, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N = |\\mathcal{D}|$ is the dataset size, and $\\mathbf{y}^i,\\mathbf{x}_{\\mathrm{ins}}^i,\\mathbf{x}_{\\mathrm{img}}^i$ are the predicted responses, text instructions, and image features for the $i$ -th instance $\\mathcal{P}_i$ , respectively. However, as we discussed before, treating each instance equally does limited help in addressing the model bias toward overly represented SAs and SCs. Our solution to this issue is inspired by", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4Collecting an imbalanced datasets is impractical, as different MLLMs may exhibit different biases.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1745", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/effe5dcc9f3fc6d169433bf230417c74963797b54ac08ec431a4a2ca50fc38ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model#ParamsSocialCounterfactualsFairFaceCMSCVQAv2MMBenchTextVQA
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLaVA7B-2.05670.3973-2.87920.6457-1.61591.481778.5064.6958.21
LLaVA+POPE-0.51010.4833-1.59330.6056-2.54241.1154---
LLaVA+FT-0.47270.3625-1.01990.4865-0.71420.805878.1163.8858.20
LLaVA+CSD-0.35090.3110-0.86220.3950-0.49330.563378.1664.2058.39
LLaVA13B-2.57300.3799-3.36040.5863-1.67300.535080.067.7061.30
LLaVA+POPE-0.38400.4410-0.95080.4051-2.25421.1454---
LLaVA+FT-0.47480.4051-1.31230.4066-1.51070.460579.1767.2861.02
LLaVA+CSD-0.31130.3718-0.71140.3752-0.81670.419279.7568.3061.42
Qwen-VL7B-0.61170.5966-1.63050.8469-1.51141.096179.3774.1461.39
Qwen-VL+POPE-0.30640.5399-1.31670.9207-2.24381.7575---
Qwen-VL+FT-0.28010.3966-1.39250.6916-0.81661.022779.3774.8260.86
Qwen-VL+CSD-0.24220.2921-0.76720.2607-0.61930.718079.3775.5960.88
Bunny8B-0.42550.6064-1.13750.5349-1.58291.417382.6076.4665.31
Bunny+POPE-0.33700.5899-1.36700.4918-2.80851.7269---
Bunny+FT-0.42020.5851-1.00350.4199-0.82370.902682.4876.3965.20
Bunny+CSD-0.40010.5532-0.90030.2632-0.77450.157582.4276.2865.20
", + "bbox": [ + 187, + 88, + 803, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance comparison. All models are fine-tuned on our CMSC dataset. Among the six datasets, SocialCounterfactuals, FairFace, and CMSC are employed to evaluate social bias; and VQAv2, MMBench, and TextVQA are general benchmarks. Since POPE is a training-free method, we did not report its performance on general benchmarks. #Params: the scale of the base LLM's parameters.", + "bbox": [ + 89, + 301, + 906, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the approaches that have been proven effective in the class imbalance research area [23]. Instead of scaling loss based on the class frequency, we leverage the stereotype quantification metric Skew to rescale the loss value,", + "bbox": [ + 89, + 354, + 482, + 415 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {f a i r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} e ^ {- \\operatorname {S k e w} \\left(\\mathcal {P} _ {i}\\right)} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\text {i n s}} ^ {i}, \\mathbf {x} _ {\\text {i m g}} ^ {i}; \\theta\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 425, + 482, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this scenario, when $\\mathrm{Skew}(\\mathcal{P}_i) > 0$ , i.e., the MLLM tends to predict the input instance as the SC label $c_{i}$ , the fairness term $e^{-\\mathrm{Skew}(\\mathcal{P}_i)}$ is less than 1.0. Consequently, this instance will receive less attention during training. Similarly, when $\\mathrm{Skew}(\\mathcal{P}_i) < 0$ , the fairness term will make the model pay more attention to this overlooked instance. This operation allows the model to dynamically adjust weights during fine-tuning based on the Skew of input instances, enabling the model to learn a fairer distribution.", + "bbox": [ + 89, + 478, + 483, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 627, + 222, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 652, + 189, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. In addition to the three social bias datasets mentioned in Section 4.1, we also employed three recent benchmark datasets to evaluate MLLMs' original zero-shot capabilities. Among them, VQAv2 [22] and TextVQA [56] are benchmarks for general visual question answering and text-oriented visual question answering, respectively. MM-Bench [41] evaluates the model robustness with comprehensive multiple-choice answers.", + "bbox": [ + 89, + 674, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. Given the limited exploration of debiasing strategies for MLLMs, we compared our CSD with the direct fine-tuning (FT), i.e., fine-tuning the model on balanced counterfactual dataset [63]. Moreover, we adapt POPE [37], which addresses hallucinations in MLLMs, as a baseline.", + "bbox": [ + 89, + 795, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Both thresholds $\\tau_{1}$ and $\\tau_{2}$ in Section 5.2 are set to 1.0. We trained the models on $8^{*}\\mathrm{A}800$", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPUs. Each training session takes $\\sim 18$ GPU hours.", + "bbox": [ + 511, + 354, + 854, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 382, + 651, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We reported MaxSkew@C and MinSkew@C of the four utilized MLLMs in Table 3, wherein we have three main observations: i) Compared to the baselines, our CSD demonstrates the best debiasing effectiveness across four different MLLM architectures. In intra-dataset evaluation, i.e., fine-tuning and testing on the CMSC dataset, LLaVA-13B+CSD achieved a MinSkew@C of -0.8167, improving by $46\\%$ compared to LLaVA-13B+FT. In cross-dataset evaluation, our CSD also outperforms the two baselines by a significant margin. For example, Qwen-VL-7B+CSD attains a MaxSkew@C of 0.2607 on the FairFace dataset, showing a reduction of 0.6 in absolute value. ii) A model with a larger size does not necessarily correspond to a lower social bias. This might because the increased scale of parameters makes the model to better learn the biases present in the pre-training dataset. iii) POPE's debiasing effectiveness underperforms FT. One possible reason is that POPE is a training-free method, causing its predictions to remain biased towards the pre-training data distribution.", + "bbox": [ + 511, + 405, + 906, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pertaining to the model results on general multi-modal benchmarks, we observed that our CSD method has a negligible impact, with the effect on all architectures across the three datasets being less than $0.5\\%$ . This indicates that our CSD method enhances the model's social fairness without significantly sacrificing its general capabilities.", + "bbox": [ + 511, + 694, + 906, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 797, + 678, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3.1. Comparison on Ablated CSD", + "text_level": 1, + "bbox": [ + 511, + 820, + 764, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our CSD method consists of two key components: the dataset resampling and the rescaled SFLoss. In Table 4, we reported the performance of them. We observe that both components help alleviate social bias. For example,", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1746", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2204cca243627680d3dbb7afebd09b66859ca697485375153739f2dc9aa1786.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSocialCounterfactualsFairFaceCMSC
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA-7B×-2.05670.3973-2.87920.6457-1.61591.4817
+FT-0.47270.3625-1.01990.4865-0.71420.8058
+Resample-0.44550.3327-0.92380.4838-0.69790.7741
+SFLoss-0.38510.3453-0.89040.4246-0.53850.5838
+CSD-0.35090.3110-0.86220.3950-0.49330.5633
LLAVA-13B×-2.57300.3799-3.36040.5863-1.67300.5350
+FT-0.47480.4051-1.31230.4066-1.51070.4605
+Resample-0.36030.3844-0.81540.3823-1.03640.4292
+SFLoss-0.34700.3905-1.05460.3946-1.03980.4388
+CSD-0.31130.3718-0.71140.3752-0.81670.4192
Qwen-VL-7B×-0.61170.5966-1.63050.8469-1.51141.0961
+FT-0.28010.3966-1.39250.6916-0.81661.0227
+Resample-0.24280.3631-0.80120.3207-0.74140.7813
+SFLoss-0.25800.3004-0.76880.3311-0.63860.8041
+CSD-0.24220.2921-0.76720.2607-0.61930.7180
Bunny-8B×-0.42550.6064-1.13750.5349-1.58291.4173
+FT-0.42020.5851-1.00350.4199-0.82370.9026
+Resample-0.41580.5670-0.98750.3127-0.81410.4008
+SFLoss-0.41710.5534-0.94480.2789-0.79410.4155
+CSD-0.40010.5532-0.90030.2632-0.77450.1575
", + "bbox": [ + 106, + 88, + 467, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The performance of ablated CSD. FT: Directly fin-tuning. Resample: Fine-tuning with resampling. SFLoss: Fin-tuning with SFLoss. The best performance is highlighted in bold.", + "bbox": [ + 89, + 314, + 482, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8a2b166ec860b7df28de920edc7bbe7df780391fecc1e7a94a8f6736bed26f9.jpg", + "image_caption": [ + "Figure 5. MaxSkews of (a) Qwen-VL-7B and (b) Bunny-8B across occupations in SocialCounterfactuals for different races." + ], + "image_footnote": [], + "bbox": [ + 120, + 364, + 284, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbeddbcd22e246b84c115a37077dd9cc486373bf0e2ece775d953f8ae62ba80c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 364, + 447, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLaVA-7B+SFLoss achieves a $70\\%$ MinSkew@C improvement on FairFace. On CMSC, Qwen-VL-7B+Resample achieves a MinSkew@C of -0.6193, reducing the absolute Skew value by more than 0.9. A notable observation is that either of the SFLoss and data resampling shows better debiasing effect compared to the naive FT strategy.", + "bbox": [ + 89, + 497, + 483, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3.2. Comparison on SAs", + "text_level": 1, + "bbox": [ + 89, + 597, + 276, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 5, we illustrate the MaxSkews [13] of Qwen-VL-7B and Bunny-8B across different races in the SocialCounterfactuals dataset. We can observe that both models exhibit significant social bias without any fine-tuning. For instance, Qwen-VL-7B has a Skew value of approximately 0.76 for White, indicating a strong preference for predicting occupations for this race. Our CSD method achieves better debiasing effects across all races for both models.", + "bbox": [ + 89, + 616, + 483, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3.3. Skew Distributions", + "text_level": 1, + "bbox": [ + 89, + 744, + 269, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figure 6, we illustrate the MaxSkew distribution across different SCs in CMSC. It can be observed that LLaVA-13B exhibits larger mean and median MaxSkew values, along with significant outliers in certain attributes. For instance, in the 'personality' category, the Skew value for Old-Belligerent reaches 1.6. After direct fine-tuning, the average degree of social bias decreases. However, the outlier issue remains unresolved. In contrast, applying our CSD method achieves a more optimized Skew distribution", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f5689938d15febc645e230610274cc70d8986847f665f2bc0161b08e7450c2a2.jpg", + "image_caption": [ + "Figure 6. MaxSkew distribution for LLaVA-13B on the CMSC dataset. Each box represents the different $\\mathrm{Skew}_{a|c}$ values. Our CSD method achieves the most optimized Skew distribution." + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 880, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across different concepts. Furthermore, noticeable outlier are eliminated, indicating that CSD provides a comprehensive debiasing effect for MLLMs.", + "bbox": [ + 511, + 335, + 906, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 513, + 393, + 761, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we present to address the notorious social bias problem in MLLMs. Our first contribution is a comprehensive dataset that covers more diverse social concepts than previous datasets. In addition, we advocate an counter-stereotype debiasing approach to perform both dataset resampling and loss rescaling, thereby improving fairness of MLLMs. Extensive experiments demonstrate that our method is promising to alleviate the social bias in MLLMs, with minimal negative impact on their original general multi-modal understanding capabilities.", + "bbox": [ + 511, + 419, + 905, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Social Impact. As MLLMs become increasingly integrated into real-world applications, their societal influence is poised to expand, particularly in sectors requiring sensitive handling of human attributes and social concepts. However, intrinsic social biases within these models can undermine their trustworthiness, raising ethical concerns and potential risks. This paper presents a bias mitigation approach to enhance fairness in MLLMs, fostering trust and accountability by reducing harmful stereotypes. We hope our work inspires further research on bias and fairness in MLLMs, contributing to the development of more equitable models.", + "bbox": [ + 511, + 570, + 905, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 750, + 663, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported in part by National Natural Science Foundation of China (Grant No. 62476071, No. U24A20328); in part by National Natural Science Foundation of China (No. 62176137).", + "bbox": [ + 511, + 773, + 905, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 847, + 609, + 863 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men", + "bbox": [ + 519, + 872, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1747", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 1-20, 2022. 2", + "[2] Mohsan S. Alvi, Andrew Zisserman, and Christoffer Nelläker. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings. In EC-CVW, pages 556-572, 2018. 3", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. CoRR, abs/2308.12966:1-15, 2023. 1, 2, 4, 5", + "[4] Hugo Berg, Siobhan Mackenzie Hall, Yash Bhalgat, Hannah Kirk, Aleksandar Shtedritski, and Max Bain. A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. In AACL/IJCNLP, pages 806-822, 2022. 3", + "[5] Karan Bhanot, Miao Qi, John S. Erickson, Isabelle Guyon, and Kristin P. Bennett. The problem of fairness in synthetic healthcare data. Entropy, 23(9):1165, 2021. 2", + "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. CoRR, abs/2110.01963:1-18, 2021. 1", + "[7] Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Tauman Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, pages 4349-4357, 2016. 2, 3", + "[8] Jannik Brinkmann, Paul Swoboda, and Christian Bartelt. A multidimensional analysis of social biases in vision transformers. In ICCV, pages 4914-4923, 2023. 1, 2", + "[9] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1-40, 2020. 2", + "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly-scaled multilingual language-image model. In ICLR, pages 1-13, 2023. 2", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 2, 5" + ], + "bbox": [ + 93, + 90, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ching-Yao Chuang and Youssef Mroueh. Fair mixup: Fairness via interpolation. In ICLR, pages 1-11, 2021. 2", + "[13] Ching-Yao Chuang, Jampani Varun, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. CoRR, abs/2302.00070:1-13, 2023. 3, 4, 8", + "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Ho. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 1-12, 2023. 2", + "[15] Sepehr Dehdashtian, Lan Wang, and Vishnu Naresh Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In ICLR, pages 1-12, 2024. 3", + "[16] Richard B Felson and Noah Painter-Davis. Another cost of being a young black male: Race, weaponry, and lethal outcomes in assaults. Social Science Research, 41(5):1241-1253, 2012. 3", + "[17] Kathleen C. Fraser and Svetlana Kiritchenko. Examining gender and racial bias in large vision-language models using a novel dataset of parallel images. In EACL, pages 690-713, 2024. 3", + "[18] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1, 2", + "[19] Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. Word embeddings quantify 100 years of gender and ethnic stereotypes. PNAS, 115(16):E3635-E3644, 2018. 3", + "[20] Sahin Cem Geyik, Stuart Ambler, and Krishnamaram Kenthapadi. Fairness-aware ranking in search & recommendation systems with application to linkedin talent search. In KDD, pages 2221-2231, 2019. 4", + "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6325-6334, 2017. 7", + "[23] Yangyang Guo, Liqiang Nie, Harry Cheng, Zhiyong Cheng, Mohan S. Kankanhalli, and Alberto Del Bimbo. On modality bias recognition and reduction. ACM ToMM, 19(3):103:1-103:22, 2023. 7", + "[24] Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu, Grace Sodunke, Aleksandar Shtedritski, and Hannah Rose Kirk. Visogender: A dataset for benchmarking gender bias in image-text pronoun resolution. In NeurIPS, pages 1-15, 2023. 1, 2, 3", + "[25] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. CoRR, abs/2402.11530:1-13, 2024. 1, 2, 4, 5", + "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1748", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "editing with cross-attention control. In ICLR, pages 1-15, 2023. 3", + "[27] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 1, 2, 3, 4", + "[28] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Nils Johan Bertil Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. In NeurIPS, pages 1-22, 2023. 2", + "[29] Ekaterina Iakovleva, Fabio Pizzati, Philip Torr, and Stéphane Lathuilière. Specify and edit: Overcoming ambiguity in text-based image editing. CoRR, abs/2407.20232:1-10, 2024. 3", + "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 5", + "[31] Sepehr Janghorbani and Gerard De Melo. Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision-language models. In EACL, pages 1725-1735, 2023. 2", + "[32] Fangkai Jiao, Zhiyang Teng, Bosheng Ding, Zhengyuan Liu, Nancy Chen, and Shafiq Joty. Exploring self-supervised logic-enhanced training for large language models. In NAACL, pages 926-941, 2024. 2", + "[33] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, pages 1548-1558, 2021. 2, 3, 4", + "[34] Tiep Le, Vasudev Lal, and Phillip Howard. Cocos-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. NeurIPS, pages 1-13, 2024. 1, 2", + "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2", + "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 2", + "[37] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 7", + "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 2, 4, 5", + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, pages 1-13, 2024. 2", + "[40] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? CoRR, abs/2307.06281:1-12, 2023. 7", + "[42] Guangben Lu, Yuzhen Du, Zhimin Sun, Ran Yi, Yifan Qi, Yizhe Tang, Tianyi Wang, Lizhuang Ma, and Fangyuan Zou. Pinco: Position-induced consistent adapter for diffusion transformer in foreground-conditioned inpainting. arXiv preprint arXiv:2412.03812, 2024. 2", + "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding. CoRR, abs/2403.05525:1-29, 2024. 5", + "[44] David Madras, Elliot Creager, Toniann Pitassi, and Richard S. Zemel. Learning adversarially fair and transferable representations. In ICML, pages 3381-3390, 2018. 3", + "[45] Thomas Manzini, Yao Chong Lim, Alan W. Black, and Yulia Tsvetkov. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. In NAACL, pages 615-621, 2019. 3", + "[46] Allison Master. Gender stereotypes influence children's stem motivation. Child Development Perspectives, 15(3):203-210, 2021. 3", + "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, pages 27730-27744, 2022. 5", + "[48] David S Pedulla. The positive consequences of negative stereotypes: Race, sexual orientation, and the job application process. Social Psychology Quarterly, 77(1):75-94, 2014. 3", + "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952: 1-21, 2023. 3", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 5", + "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21:140:1-140:67, 2020. 2", + "[52] Vikram V. Ramaswamy, Sunnie S. Y. Kim, and Olga Russakovsky. Fair attribute classification through latent space de-biasing. In CVPR, pages 9301-9310, 2021. 2", + "[53] Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck, Estelle Aflalo, Shao-Yen Tseng, Vasudev Lal, and Phillip" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1749", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Howard. Debias your large multi-modal model at test-time with non-contrastive visual attribute steering. arXiv preprint arXiv:2411.12590, 2024.3", + "[54] Prasanna Sattigeri, Samuel C. Hoffman, Vijil Chenthamarakshan, and Kush R. Varshney. Fairness GAN: generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3:1-3:9, 2019. 2", + "[55] Ashish Seth, Mayur Hemani, and Chirag Agarwal. Dear: Debiasing vision-language models with additive residuals. In CVPR, pages 6820-6829, 2023. 2, 3, 5", + "[56] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: A dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 7", + "[57] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: pre-training of generic visual-linguistic representations. In ICLR, pages 1–13, 2020. 2", + "[58] Yizhe Tang, Zhimin Sun, Yuzhen Du, Ran Yi, Guangben Lu, Teng Hu, Luying Li, Lizhuang Ma, and Fangyuan Zou. Ata: Adaptive transformation agent for text-guided subject-position variable background inpainting. In CVPR, pages 18335-18345, 2025. 2", + "[59] Özge Tarhan. Children's understanding of the concept of social stereotypes. Child Indicators Research, 15(3):989-1023, 2022. 3", + "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971:1-16, 2023. 2, 5", + "[61] Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. Are emily and greg still more employable than lakisha and jamal? investigating algorithmic hiring bias in the era of chatgpt. CoRR, abs/2310.05135:1-13, 2023. 2", + "[62] W Edgar Vinacke. Stereotypes as social concepts. The journal of social Psychology, 46(2):229-243, 1957. 1", + "[63] Angelina Wang and Olga Russakovsky. Overwriting pretrained bias with finetuning data. In ICCV, pages 3934-3945, 2023. 1, 7", + "[64] Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy Chen. SeaEval for multilingual foundation models: From cross-lingual alignment to cultural reasoning. In NAACL, pages 370–390. Association for Computational Linguistics, 2024. 2", + "[65] Jialu Wang, Yang Liu, and Xin Eric Wang. Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. In EMNLP, pages 1995-2008, 2021. 1", + "[66] Mei Wang and Weihong Deng. Mitigating bias in face recognition using skewness-aware reinforcement learning. In CVPR, pages 9319-9328, 2020. 3", + "[67] Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. Evaluating fairness in large vision-language models across diverse demographic attributes and prompts. CoRR, abs/2406.17974:1-11, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Yisong Xiao, Aishan Liu, QianJia Cheng, Zhenfei Yin, Siyuan Liang, Jiapeng Li, Jing Shao, Xianglong Liu, and Dacheng Tao. Genderbias-vl: Benchmarking gender bias in vision language models via counterfactual probing. CoRR, abs/2407.00600:1-13, 2024. 2", + "[69] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, pages 1-13, 2018. 2", + "[70] Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE TPAMI, 45(12):14144–14160, 2023. 2", + "[71] Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In ICML, pages 59310-59328. PMLR, 2024. 2", + "[72] Jie Zhang, Sibo Wang, Xiangkui Cao, Zheng Yuan, Shiguang Shan, Xilin Chen, and Wen Gao. Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model. CoRR, abs/2406.14194:1-13, 2024. 2", + "[73] Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. NeurIPS, pages 21682-21697, 2022. 3", + "[74] Xiaobin Zhang, Ziqi Cui, Ning He, and Xuqun You. Are boys associated with weapons and girls associated with kitchenware? the extent to which gender stereotypes regarding adults extend to children. Acta psychologica, 230: 103754, 2022. 3", + "[75] Kankan Zhou, Eason Lai, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In AACL/IJCNLP, pages 527-538, 2022. 2" + ], + "bbox": [ + 516, + 90, + 903, + 529 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1750", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_model.json b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dae27580d3eef6a676783c1d5d724514a423e7ab --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_model.json @@ -0,0 +1,2675 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.271, + 0.131, + 0.726, + 0.153 + ], + "angle": 0, + "content": "Social Debiasing for Fair Multi-modal LLMs" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.18, + 0.896, + 0.234 + ], + "angle": 0, + "content": "Harry Cheng\\(^{1}\\), Yangyang Guo\\(^{2}\\), Qingpei Guo\\(^{3,4,*}\\), Ming Yang\\(^{4}\\), Tian Gan\\(^{1}\\), Weili Guan\\(^{5}\\), Liqiang Nie\\(^{5}\\) \n\\(^{1}\\)Shandong University, \\(^{2}\\)National University of Singapore, \\(^{3}\\)University of Warwick, \\(^{4}\\)Ant Group, \\(^{5}\\)Harbin Institute of Technology (Shenzhen)" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.238, + 0.861, + 0.27 + ], + "angle": 0, + "content": "xaCheng1996@gmail.com, guoyang.eric@gmail.com, qingpei.gqp@antgroup.com, m.yang@antgroup.com, gantian@sdu.edu.cn, honeyguan@gmail.com, nieliqiang@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.655 + ], + "angle": 0, + "content": "Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.222, + 0.7 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Multi-modal Large Language Models (MLLMs) have revolutionized the field of general-purpose vision-language understanding. Representative models, such as LLaVA [38], Qwen-VL [3], and Bunny [25], exhibit remarkable zero-shot performance and can be easily fine-tuned for diverse downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Despite MLLMs' widespread use, it is imperative to recognize that these models can exhibit severe social biases with respect to attributes such as race and gender [18, 65]. Figure 1a illustrates one example that a biased MLLM is" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.303, + 0.882, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Figure 1. Social bias examples. Left: Gender-biased prediction on Nurse from LLaVA-7B1. Right: Maximum Skews across different human attributes on the FairFace dataset from Qwen-VL-7B and LLaVA-7B. As a metric for measuring model bias, a larger Skew implies a higher degree of social bias. Our CSD method outperforms existing approaches by a significant margin." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.644 + ], + "angle": 0, + "content": "much more likely to associate the role of nurse with female rather than male, which may be stereotypical. Such biased predictions often occur unconsciously within MLLMs, making them difficult to detect and avoid by explicit rules. A key reason for this issue lies in the composition of the MLLM training data, which can contain content related to violence or racism [6]. These inappropriate samples may collectively demonstrate some stereotypes, resulting in subtle yet uncomfortable responses [8, 63]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Existing studies on mitigating the social bias problem in MLLMs remain largely under-explored. A naive approach is to collect attribute-balanced vision-language counterfactual datasets [24, 34], which can be used to directly finetune biased models to predict fairer distributions. However, as demonstrated in Table 1, existing large-scale datasets are limited by their focus on a single social concept, such as occupation [27], without attending to multifaceted social stereotypes [62]. This significantly hampers the model in learning more comprehensive representations. Therefore, we are motivated to construct a more extensive and diverse counterfactual dataset. From a methodological perspective, directly fine-tuning on a counterfactual dataset assigns equal importance to instances receiving different social biases, resulting in sub-optimal debiasing performance (as demonstrated in Section 6.2). Analogy to a common solution in chemistry—neutralizing acidic water requires" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.482, + 0.9 + ], + "angle": 0, + "content": "1We manually visualize the bounding boxes from LLaVA's text output." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1740" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.472, + 0.186 + ], + "angle": 0, + "content": "
DatasetVenueType#ImagesSocial Concepts
CoCo-Counterfactuals [34]NeurIPS'23General34KNo annotations
FairFace [33]WACV'21Social108KNo annotations
VisoGender [24]NeurIPS'23Social0.6KOccupation
PATA [55]CVPR'23Social5KOccupation
SocialCounterfactuals [27]CVPR'24Social171KOccupation
MM-Bias [31]EACL'23Social3K14 minorities
CMSC (Ours)-Social60K18 balanced concepts
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.189, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Table 1. Comparison of CMSC with mainstream datasets used for model debiasing. Existing datasets are limited either by scale or by the number of social concepts they cover. Type - The category of biases the dataset focuses on." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.3 + ], + "angle": 0, + "content": "adding an alkaline substance rather than plain water—we are prompted to speculate: Can we leverage the opposite of the suffered social bias to build a fairer model?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.301, + 0.484, + 0.526 + ], + "angle": 0, + "content": "To address these two issues, we first construct a large-scale, high-quality Counterfactual dataset with Multiple Social Concepts (CMSC). As demonstrated in Table 1, our CMSC dataset narrows the gap in data scale and concept richness. We conduct extensive experiments on four different MLLM architectures [3, 25, 38] using several existing counterfactual datasets as well as our CMSC. The results demonstrate that MLLMs fine-tuned on CMSC exhibit significantly lower social bias compared to those fine-tuned on single-concept datasets. For instance, LLaVA-7B [38] fine-tuned on CMSC achieves a \\(64\\%\\) debiasing effect, which significantly outperforms the \\(30\\%\\) bias reduction achieved when fine-tuned on the SocialCounterfactuals dataset \\([27]^2\\). This indicates that fine-tuning the models with a rich set of social concept is beneficial in MLLM debiasing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Furthermore, to implement the idea of 'debiasing with the opposite of the social biases', we propose a Counter-Stereotype Debiasing strategy (CSD) to effectively reduce social biases in MLLMs. Our CSD is equipped with two techniques: 1) we design a novel data sampling method based on the bias level, 2) we rescale the previously used autoregressive loss function to a new Social Fairness Loss (SFLoss). In this manner, previously under-represented cases, e.g., male nurses, will receive more attention, serving as new 'alkaline' ones to counteract the bias of MLLMs that prefer female nurses. The experimental results demonstrate that our CSD method is an effective debiasing strategy. As illustrated in Figure 1b, our method reduces the bias by over \\(50\\%\\) for the Qwen-VL-7B model compared to prior training strategies, especially naive fine-tuning. Additionally, CSD does not compromise the original model performance on general multi-modal benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.785, + 0.41, + 0.799 + ], + "angle": 0, + "content": "In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.801, + 0.483, + 0.845 + ], + "angle": 0, + "content": "- We construct a high-quality counterfactual dataset that includes 18 social concepts, which demonstrated to be superior to existing ones on MLLM social debiasing." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.847, + 0.483, + 0.876 + ], + "angle": 0, + "content": "- We propose a novel approach that applying counter-stereotype debiasing to mitigate social biases. To the best" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.801, + 0.483, + 0.876 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "of our knowledge, this is the first research effort dedicated to addressing the social bias problem in autoregressive MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We apply our debiasing strategy to four multi-modal LLMs architectures to show the generalization capability and effectiveness of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.198, + 0.655, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.224, + 0.831, + 0.24 + ], + "angle": 0, + "content": "2.1. Multi-modal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.58 + ], + "angle": 0, + "content": "With the rapid development of LLMs [9, 51, 57, 60, 64], increasing efforts have been dedicated to extending the powerful reasoning capabilities of LLMs to multi-modal applications [1, 10, 14, 28, 32, 35, 40, 42, 58, 70, 71]. Specifically, MLLMs utilize LLMs as the foundational base, aligning features from other modalities with text embeddings to enable LLMs to perceive multi-modal inputs. For instance, BLIP-2 [36] utilizes a Q-Former to align vision encoders with LLMs. LLaVA [38, 39] presents to directly map vision features to the word embedding space of Vicuna [11]. Qwen-VL [3] employs a single-layer cross-attention module as the vision-language adapter, and introduces multitask pre-training to improve the model performance. These approaches significantly enhance multi-modal understanding capabilities. However, some studies highlight the existence of social biases in pre-trained LLMs [7, 8, 18]. These biases often manifest as harmful spurious correlations related to human attributes such as gender and race, which significantly undermine the fairness. For instance, LLMs may 'efficiently' filter resumes based on race rather than the candidates' qualifications [61]. Nevertheless, research on social biases in MLLMs remains largely unexplored [68]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.591, + 0.719, + 0.606 + ], + "angle": 0, + "content": "2.2. Social Bias Reduction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.905, + 0.644 + ], + "angle": 0, + "content": "We roughly categorize the bias mitigation strategies into two groups: data-based and objective-based." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Data-based debiasing typically refers to data augmentation techniques. For instance, Chuang et al. [12] employ mixup [69] to construct interpolated samples among groups with different distributions. Ramaswamy et al. [52] utilize perturbed GAN-generated images [21] in latent space to augment the original dataset. Moreover, some studies collect fair datasets across human attributes for continual finetuning [5, 24, 31, 34, 54, 72, 75]. In particular, FairFace [33] contains \\(108\\mathrm{K}\\) images that are balanced for the race attribute. Socialcounterfactuals [27] collects \\(171\\mathrm{K}\\) imagetext pairs to probe biases across race, gender, and physical characteristics. However, these collected counterfactual datasets often focus on only one common concept - occupation, which hinders effective debiasing [67]. Though a few datasets involve several concepts, they nevertheless, are largely limited by their data scale (e.g., 3K [31]), making it less feasible to fine-tune MLLMs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.468, + 0.901 + ], + "angle": 0, + "content": "2In this paper, we interchangeably use debiasing and bias reduction." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1741" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.09, + 0.824, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.907, + 0.237 + ], + "angle": 0, + "content": "Figure 2. Image Generation Pipeline. Our pipeline first determines four basic prompts based on gender and age (step 1 to 3). Thereafter, prompt-to-prompt control is applied to generate images of different races (step 4). Finally, we filter out low-quality images (step 5)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.244, + 0.485, + 0.548 + ], + "angle": 0, + "content": "Objective-based debiasing modifies the model training process to achieve better fairness. Early studies are proposed to reduce social bias within uni-modal models [2, 17, 19, 44, 45, 66]. For example, Bolukbasi et al. [7] optimizes word embeddings to remove gender stereotypes. In contrast, multi-modal debiasing methods often focus on contrastive learning-based vision language models, e.g., CLIP [50]. Representative approaches are to rectify the modality similarity matrix in VLMs [15] via learning additive adapters [55, 73], eliminating biased directions [13], and using adversarial samples [4]. However, among these methods, uni-modal approaches often require re-training the full model. This operation, though plausible for previous small-scale models, is less practical for existing MLLMs with billions of parameters [53]. On the other hand, mainstream multi-modal techniques cannot be utilized for these models due to the divergent training objectives between MLLMs and previous CLIP-style models. In particular, MLLMs are mostly trained in an autoregressive way rather than through contrastive learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.357, + 0.577 + ], + "angle": 0, + "content": "3. CMSC Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.68 + ], + "angle": 0, + "content": "As demonstrated in Table 1, existing counterfactual datasets are either restricted by their small scale [24] or a narrow coverage of concepts [33]. To bridge this gap between limited existing datasets and diverse real-world stereotypes, we introduce the CMSC dataset, encompassing 60k high-quality images across eighteen social concepts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.367, + 0.707 + ], + "angle": 0, + "content": "3.1. Social Attributes and Concepts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.864 + ], + "angle": 0, + "content": "A counterfactual dataset for social bias reduction and evaluation typically contains two key aspects: Social Attribute (SA) and Social Concept (SC). The former, i.e., SA, is defined as characteristics shared by a group of people [33]. Specifically, we investigate three types of attributes: i) genders: {Male, Female}; ii) races: {White, Black, Indian, Southeast Asian, East Asian, Middle Eastern, Latino}; and iii) ages: {Young, Old}. These attributes are intrinsic to individuals3. For each image \\(\\mathbf{I}_i\\) in CMSC, three SA labels are provided with respect to these three types of SA, and are" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.686, + 0.259 + ], + "angle": 0, + "content": "combined as a SA set \\(\\mathcal{A}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.907, + 0.502 + ], + "angle": 0, + "content": "As for the SC, we define it as the societal label attributed to an individual. Drawing from the sociological research [16, 46, 48, 59, 74], we employ 18 SCs in CMSC. In particular, these SCs are categorized into three groups: personality, responsibility, and education. Specifically, personality relates to concepts pertaining to an individual's character. We use five concepts: {compassionate, belligerent, authority, pleasant, unpleasant}. Responsibility is the roles or duties that individuals are expected to fulfill in society or family. We identify six concepts: {tool user, weapon user, career, family, chef working, earning money}. Education pertains to the level of education a person has received. We include seven concepts: {middle school, high school, university, good student, bad student, science, arts}. Each image in CMSC is annotated with one SC label \\( c_i \\in \\mathcal{C} \\), where \\( \\mathcal{C} \\) is the union of the concepts from all above groups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.563 + ], + "angle": 0, + "content": "In summary, each instance \\(\\mathcal{P}_i\\) in our dataset consists of an image \\(\\mathbf{I}_i\\), a set of SA labels \\(\\mathcal{A}_i\\), and a SC label \\(c_i\\). To the best of our knowledge, CMSC is the first large-scale counterfactual dataset that includes a variety of social concepts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.757, + 0.592 + ], + "angle": 0, + "content": "3.2. Image Generation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We employ Stable Diffusion XL (SDXL) [49] to generate instances in CMSC. For each SC, we carefully design prompts that provide a detailed explanation of the concept. These prompts not only help eliminate potential ambiguities associated with the SC but also enhance the quality of the generated images [29]. For instance, for the SC 'pleasant,' we use the prompt: 'A photo of a pleasant [race] [gender] person at [age], who has a friendly smile, a relaxed posture, and ...' To ensure diverse and representative image generation, these prompts are randomly modified by replacing words with synonyms during image generation. Subsequently, the intersectional generation strategy [27] is applied. As illustrated in Figure 2, we first determine gender and age (Step 2) and generate four base images (Step 3). Thereafter, we apply Prompt-to-Prompt (P2P) control [26] to produce visually similar images that vary only in terms of race (Step 4). For each SC, we create 28 prompts across attribute combinations. Each prompt helps generate 100 images. To ensure image quality, we apply a rigorous image filtering mechanism (step 5). First, we use CLIP-Score [50]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3All SAs are perceived, made by human annotators or models. We acknowledge that the SAs are not representative of all people [27, 33]." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1742" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "to filter out images that are semantically misaligned with the prompts. Subsequently, a panel of three experts conducts a voting process to remove low-quality and NSFW content. If the generated images perform poorly on a specific SA, the entire set of images (including those of other SAs) will be discarded. Moreover, we observe that the model can generate high-quality images for certain SCs or SAs, while struggling to do so for others. As a result, we filter out excessively generated ones for balanced distribution. Our filtering process is very strict. For instance, the testing set initially contains over 50,000 images, but only 12,019 images are finally used in CMSC. A similar process is applied to the training set, which comprises 48,134 images. In total, our dataset contains 60K images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.318, + 0.338, + 0.334 + ], + "angle": 0, + "content": "4. CMSC Dataset Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.343, + 0.285, + 0.36 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Evaluation Protocols. We employed four different MLLM architectures, i.e., LLaVA-7B [38], LLaVA-13B [38], Qwen-VL-7B [3], and Bunny-8B [25], to evaluate our CMSC dataset. For each MLLM, we used the same template to collect their responses: 'Based on the race, gender, and age of the face , do you think the person is {SCs}? Respond with one of them.' Here, is the image from the dataset, and {SCs} represents the list of used SCs. For instance, when testing personality, {SCs} includes 'pleasant,' 'unpleasant,' etc. After collecting these predictions, we used them to compute MaxSkew@C and MinSkew@C (which will be elaborated in the following section) to measure bias." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Metrics. We employ Skew-based [20] metrics to quantify the extent of social biases. Specifically, for a given dataset \\(\\mathcal{D}\\), we define the subset containing instances with a specific SC label \\(c'\\) as \\(\\mathcal{D}_{c'}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.636, + 0.483, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D}, c _ {i} = c ^ {\\prime} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.484, + 0.698 + ], + "angle": 0, + "content": "From \\(\\mathcal{D}_{c'}\\), we further extract the subset consisting of instances with a specific SA label \\(a'\\) as \\(\\mathcal{D}_{a'|c'}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.709, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D} _ {c ^ {\\prime}}, a ^ {\\prime} \\in \\mathcal {A} _ {i} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.484, + 0.86 + ], + "angle": 0, + "content": "We then utilize the aforementioned template to guide MLLMs in predicting the SC label corresponding to each \\(\\mathcal{P}_i\\). We denote the predicted SC label as \\(\\hat{c}_i\\). By applying this process to the entire dataset \\(\\mathcal{D}\\), we can construct a new predicted set \\(\\hat{\\mathcal{D}} = \\{\\hat{\\mathcal{P}}_i\\}_{i=1}^N\\), where \\(\\hat{\\mathcal{P}}_i = \\{\\mathbf{I}_i, \\mathcal{A}_i, \\hat{c}_i\\}\\), and \\(N = |\\mathcal{D}|\\). In \\(\\hat{\\mathcal{D}}\\), subsets \\(\\hat{\\mathcal{D}}_{c'}\\) and \\(\\hat{\\mathcal{D}}_{a'|c'}\\) can be derived with Equation (1) and Equation (2), respectively. The Skew for SC \\(c'\\) and SA \\(a'\\) could be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.871, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S k e w} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\log \\left(\\frac {\\hat {\\gamma} _ {a ^ {\\prime}} \\mid c ^ {\\prime}}{\\gamma_ {a ^ {\\prime}} \\mid c ^ {\\prime}}\\right), \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.09, + 0.875, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.324 + ], + "angle": 0, + "content": "Figure 3. Performance of MLLMs when evaluated on FairFace. In each model, the bars labeled 'FT w SCounter' and 'FT w CMSC' represent the model fine-tuned on SocialCounterfactuals and our CMSC, respectively." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.34, + 0.56, + 0.352 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.351, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {\\gamma} _ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\hat {\\mathcal {D}} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\hat {\\mathcal {D}} _ {c ^ {\\prime}} \\right|}, \\\\ \\gamma_ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\mathcal {D} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\mathcal {D} _ {c ^ {\\prime}} \\right|}. \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.514 + ], + "angle": 0, + "content": "When \\(\\mathrm{Skew}_{a'|c'} > 0\\), the MLLM tends to predict instances with attribute \\(a'\\) as concept \\(c'\\). For instance, in Figure 1a, \\(\\mathrm{Skew}_{\\text{Female|Nurse}} > 0\\). In contrast, when \\(\\mathrm{Skew}_{a'|c'} < 0\\), the MLLM is inclined to not predict these instances containing SA \\(a'\\) as SC \\(c'\\), e.g., predicting male as nurse. A fair MLLM should have Skew close to 0 across all concepts and attributes, i.e., \\(\\sum_{a,c} |\\mathrm{Skew}_{a|c}| \\to 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.906, + 0.633 + ], + "angle": 0, + "content": "Although Skew can effectively measure MLLMs' bias towards a particular SA-SC pair, a counterfactual dataset often contains hundreds of such combinations, e.g., our CMSC includes 198 SA-SC pairs. This confounds the comprehensive assessment of MLLM bias. Therefore, we propose two variants of Skew to enable a more thorough analysis. Specifically, we first identify the maximum and minimum Skew values for each SC across all SAs [13]," + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.644, + 0.905, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\operatorname {M a x S k e w} _ {c ^ {\\prime}} = \\operatorname {M a x} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}, \\\\ \\operatorname {M i n S k e w} _ {c ^ {\\prime}} = \\operatorname {M i n} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}. \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.78 + ], + "angle": 0, + "content": "For all \\(c' \\in \\mathcal{C}\\), we separately calculate the average of \\(\\mathrm{MaxSkew}_{c'}\\) and \\(\\mathrm{MinSkew}_{c'}\\), resulting in two aggregated Skew values. We refer to these two values as MaxSkew@C (MaxS@C) and MinSkew@C (MinS@C), which represent the overall bias level of the MLLM. Both of these metrics indicate better fairness as they approach zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. We compared our dataset with the previous state-of-the-art counterfactual dataset. Specifically, we separately fine-tuned MLLMs on CMSC and SocialCounterfactual [27] to evaluate their effectiveness in mitigating bias. The SocialCounterfactual dataset focuses on the SC of occupation and contains 171K high-quality synthetic images. Moreover, we employed the FairFace dataset [33], which contains 108K images with a balanced distribution of SAs," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1743" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.465, + 0.199 + ], + "angle": 0, + "content": "
FT w/ SCPer.Res.Edu.
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA×-0.94862.4950-0.76622.2188-0.85693.9821
+ FT Per.-0.84691.6675-0.59152.2025-0.60943.1794
+ FT Res.-0.91521.7409-0.49211.4397-0.81582.7705
+ FT Edu.-0.97561.7842-0.68721.5719-0.52572.4362
Qwen.VL×-2.46882.5772-2.53012.2663-2.82422.4059
+ FT Per.-0.73861.2509-1.90770.3470-2.55242.3506
+ FT Res.-0.73841.9044-1.32050.1202-2.85452.3872
+ FT Edu.-0.84311.6817-1.87590.2236-2.47062.3502
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.199, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Table 2. Performance comparison of LLaVA-7B and Qwen-VL-7B when fine-tuned and tested on SCs in CMSC. Per.: Personality, Res.: Responsibility, Edu.: Education, \\(\\times\\): Original performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.303 + ], + "angle": 0, + "content": "as an additional testing set to assess the cross-dataset debiasing performance of models fine-tuned on CMSC and SocialCounterfactual [55]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.318, + 0.342, + 0.334 + ], + "angle": 0, + "content": "4.2. Comparison on Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.522 + ], + "angle": 0, + "content": "We reported MinSkew@C and MaxSkew@C of several MLLMs in Figure 3. The models, i.e., LLaVA [38], Qwen-VL [3], and Bunny [25], are fine-tuned on the SocialCounterfactuals and CMSC datasets individually and evaluated on the FairFace dataset, respectively. For clarity, we took the absolute values of both metrics. From Figure 3, we observed that models fine-tuned on CMSC exhibit superior debiasing effects. For instance, LLaVA-13B achieves a MinSkew@C of -1.3132, a significant advantage over that from the model fine-tuned on SocialCounterfactuals. This implies that the broader range of social concepts in CMSC enables the model to learn fairer distributions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.537, + 0.403, + 0.554 + ], + "angle": 0, + "content": "4.3. Fine-tuning with Specific SC Group" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.728 + ], + "angle": 0, + "content": "As discussed in Section 3.1, we split the SCs in CMSC into three groups: personality, responsibility, and education. We applied LLaVA and Qwen-VL to fine-tune separately on these subsets. The results are reported in Table 2. It can be observed that in intra-subset evaluations, i.e., fine-tuned and tested on the same subset, the model can generally achieve lower biases. For instance, when fine-tuning and evaluating Qwen-VL on the personality subset, it achieves a MaxS@C of 1.2509, showing an absolute difference of 0.65 and 0.43 compared to performance when fine-tuned on the responsibility and education subsets, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.396, + 0.758 + ], + "angle": 0, + "content": "4.4. Comparison on Image Distribution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We calculated the Fréchet Inception Distance (FID) scores for the synthetic counterfactual datasets SocialCounterfactuals and our CMSC. Specifically, we randomly sampled 1,000 synthetic images from each dataset and then computed the distributional differences with the same set of 1,000 real images. A lower FID signifies better image quality. With this process, the two datasets received FID scores of 27.17 and 24.35, respectively. This indicates that the images in our dataset bear a closer resemblance to reality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.788, + 0.108 + ], + "angle": 0, + "content": "5. Counter-stereotype Debiasing" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.655, + 0.13 + ], + "angle": 0, + "content": "5.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.906, + 0.213 + ], + "angle": 0, + "content": "Before introducing our proposed method, we first revisit the training objectives of MLLMs. Subsequently, we introduce the concept of \\(\\mathrm{Skew}(\\mathcal{P}_i)\\) to measure stereotypes associated with specific instances, upon which we base the reformulation of the autoregressive fine-tuning paradigm of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.768, + 0.235 + ], + "angle": 0, + "content": "5.1.1. Traning Objective of MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.239, + 0.906, + 0.285 + ], + "angle": 0, + "content": "Current mainstream MLLMs employ pre-trained LLMs [11, 60] as the output interface [3, 25, 38, 43]. Under this context, the base LLM is often trained in an autoregressive way," + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.294, + 0.906, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} \\left(\\mathbf {x} _ {1: T}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {< t}; \\theta\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.345, + 0.906, + 0.421 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is a text sequence with \\(T\\) tokens, and \\(\\theta\\) denotes the model parameters. During pre-training, the LLM predicts the \\(t\\)-th token \\(\\mathbf{x}_t\\) based on all preceding ones, i.e., \\(\\mathbf{x}_{< t}\\). After that, another instruction tuning stage [47] is often followed to enable LLMs to better understand user intentions," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.43, + 0.906, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}; \\theta\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.481, + 0.906, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{\\mathrm{ins}}\\) and \\(\\mathbf{y}\\) are textual instructions and responses, respectively. MLLMs extend the inputs of LLMs with the supplement of image features \\(\\mathbf{x}_{\\mathrm{img}}\\) that are extracted using pre-trained vision encoders such as ViT [30, 50]," + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.552, + 0.906, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.603, + 0.906, + 0.649 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_{\\mathrm{img}}\\) is aligned with text features via a connector, e.g., a trainable projection matrix. To optimize the pipeline, previous cross-entropy loss from LLMs is directly inherited," + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.658, + 0.906, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {y}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta) = - \\sum_ {t = 1} ^ {T} \\log \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.716, + 0.734, + 0.731 + ], + "angle": 0, + "content": "5.1.2. Stereotype Measurement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Recall that \\(\\mathrm{Skew}_{a|c}\\) quantifies the degree of social biases in an MLLM across the entire dataset. In this section, we are more interested in the social bias degree for each specific instance \\(\\mathcal{P}_i\\). We then define \\(\\mathrm{Skew}(\\mathcal{P}_i)\\) as a selected \\(\\mathrm{Skew}_{a|c}\\) with the maximum absolute value across all SAs for instance \\(\\mathcal{P}_i\\). For example, suppose that \\(\\mathcal{P}_i\\) contains three SA labels: {'White,' Female,' 'Young'}, along with one SC label: 'Nurse.' \\(\\mathrm{Skew}(\\mathcal{P}_i)\\) is the one with the highest absolute value from \\(\\{\\mathrm{Skew}_{White|Nurse}, \\mathrm{Skew}_{Female|Nurse}, \\mathrm{Skew}_{Young|Nurse}\\}\\). It describes the SA that receives the most severe bias, thereby indicating the level of bias in \\(\\mathcal{P}_i\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1744" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.089, + 0.449, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.264 + ], + "angle": 0, + "content": "Figure 4. Comparison between fine-tuning (upper part) and our CSD (lower part). CSD adjusts the loss and datasets based on \\(\\mathrm{Skew}(\\mathcal{P}_i)\\), guiding the model to focus on overlooked instances." + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.265, + 0.424, + 0.278 + ], + "angle": 0, + "content": ": Parameter updates based on a specific loss function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.293, + 0.245, + 0.31 + ], + "angle": 0, + "content": "5.2. CSD Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.496 + ], + "angle": 0, + "content": "As illustrated in Figure 4, vanilla fine-tuning approach involves sampling images from a balanced dataset and updating MLLMs through the original autoregressive objective. We argue that this method, which treats all instances equally, is ineffective in addressing the social bias problem in MLLMs. Therefore, we propose an CSD method from the view of counter-stereotype. Specifically, our CSD is composed of two components: i) Dataset Resampling, which enhances the data sampling process to include more underrepresented instances. ii) Loss Rescaling, where we adjust the loss function to place larger emphasis on instances that are overlooked in terms of social attributes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.498, + 0.484, + 0.619 + ], + "angle": 0, + "content": "Dataset Resampling. The debiasing approach, e.g., directly fine-tuning, utilizes counterfactual datasets that are balanced across all SAs. We believe that such datasets make it difficult for MLLMs to recognize which SAs are subjected to greater bias and which are not4. To address this, we resample the dataset to increase the frequency of instances that MLLMs tend to ignore and reduce the frequency of emphasized ones, thereby 'neutralizing' social biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.695 + ], + "angle": 0, + "content": "\\(\\operatorname{Skew}(\\mathcal{P}_i)\\) serves as an indicator for dataset resampling. As demonstrated in Algorithm 1, for each instance \\(\\mathcal{P}_i\\), when \\(\\operatorname{Skew}(\\mathcal{P}_i) > 0\\), i.e., \\(\\mathcal{P}_i\\) has received more attention – such as the 'female-nurse' in Figure 1 – we reduce its probability in the resampled dataset \\(\\mathcal{D}_r\\) for the following training epoch," + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.706, + 0.484, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {r} = \\left\\{ \\begin{array}{l} \\mathcal {D} _ {r} \\cup \\mathcal {P} _ {i}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) > \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\\\ \\mathcal {D} _ {r}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) \\leq \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.484, + 0.865 + ], + "angle": 0, + "content": "where \\(\\operatorname{rand}(\\cdot)\\) is to randomly draw from 0 to \\(\\operatorname{Skew}(\\mathcal{P}_i) + \\tau_1\\), and \\(\\tau_{1}\\) is a pre-defined threshold. As such, a larger \\(\\operatorname{Skew}(\\mathcal{P}_i)\\) corresponds to a lower chance of being included by \\(\\mathcal{D}_r\\). In contrast, for instances with \\(\\operatorname{Skew}(\\mathcal{P}_i) \\leq 0\\), we believe that increasing their proportion in the new dataset \\(\\mathcal{D}_r\\) is beneficial for the MLLM to learn the features of these overlooked parts, thereby achieving a fairer distribution." + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.092, + 0.842, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1 Dataset Resampling with Skew \\((\\mathcal{P}_i)\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.513, + 0.112, + 0.907, + 0.143 + ], + "angle": 0, + "content": "Input: The original dataset \\(\\mathcal{D} = \\{\\mathbf{I}_i, \\mathcal{A}_i, c_i\\}\\), the social attribute sets \\(\\mathcal{A}\\), the social concept set \\(\\mathcal{C}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.144, + 0.753, + 0.158 + ], + "angle": 0, + "content": "Output: The resampled dataset \\(\\mathcal{D}_r\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.524, + 0.159, + 0.907, + 0.404 + ], + "angle": 0, + "content": "1: \\(\\mathcal{D}_r\\gets []\\) ,AcmSkew \\(a_{i}|c_{i}\\gets 0\\) ; Initialization \n2: for \\(\\mathcal{P}_i\\in \\mathcal{D}\\) do \n3: if Skew \\((\\mathcal{P}_i) > 0\\) then \n4: if Rand(0,Skew \\((\\mathcal{P}_i) + \\tau_1) > \\mathrm{Skew}(\\mathcal{P}_i)\\) then \n5: \\(\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;\\) \n6: end if \n7: else \\(\\triangleright\\) Skew \\((\\mathcal{P}_i)\\leq 0\\) \n8: \\(\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;\\) \n9: AcmSkew \\(a_{i}|c_{i}\\gets\\) AcmSkew \\(a_{i}|c_{i} + |\\mathrm{Skew}(\\mathcal{P}_i)|;\\) \n10: if AcmSkew \\(a_{i}|c_{i} > \\tau_{2}\\) then \n11: \\(\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;\\) Over-resampling \n12: AcmSkew \\(a_{i}|c_{i}\\gets 0\\) \n13: end if \n14: end if \n15: end for \n16: return \\(\\mathcal{D}_r\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.507 + ], + "angle": 0, + "content": "To this end, these instances are directly accepted into \\(\\mathcal{D}_r\\). Moreover, we design an over-resampling mechanism to further increase the occurrence frequency of these instances. Specifically, for each \\(\\mathcal{P}_i\\), we employ an accumulative value, AcmSkew, to gradually accumulate the current \\(\\mathrm{Skew}(\\mathcal{P}_i)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.518, + 0.907, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} = \\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} + | \\operatorname {S k e w} (\\mathcal {P} _ {i}) |, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.683 + ], + "angle": 0, + "content": "where \\( c_{i} \\) and \\( a_{i} \\) are SC label and SA label corresponding to \\( \\operatorname{Skew}(\\mathcal{P}_i) \\), respectively. When \\( \\operatorname{AcmSkew}_{a_i|c_i} \\) exceeds a threshold \\( \\tau_2 \\), we add the \\( \\mathcal{P}_i \\) into \\( \\mathcal{D}_r \\) again. The \\( \\operatorname{AcmSkew}_{a_i|c_i} \\) can then be set to 0 for a new round of accumulation. With the above operations, instances with \\( \\operatorname{Skew}(\\mathcal{P}_i) \\leq 0 \\) will be resampled multiple times. This resampling process is executed once before each training epoch. For evaluation, we employ the balanced testing set to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.745 + ], + "angle": 0, + "content": "Loss Rescaling. During MLLM fine-tuning, we rescale the autoregressive loss in Equation (9) to a new Social Fairness Loss (SFLoss) for effective debiasing. Specifically, the empirical risk during training can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.756, + 0.907, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\mathrm {i n s}} ^ {i}, \\mathbf {x} _ {\\mathrm {i m g}} ^ {i}; \\theta\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(N = |\\mathcal{D}|\\) is the dataset size, and \\(\\mathbf{y}^i,\\mathbf{x}_{\\mathrm{ins}}^i,\\mathbf{x}_{\\mathrm{img}}^i\\) are the predicted responses, text instructions, and image features for the \\(i\\)-th instance \\(\\mathcal{P}_i\\), respectively. However, as we discussed before, treating each instance equally does limited help in addressing the model bias toward overly represented SAs and SCs. Our solution to this issue is inspired by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "4Collecting an imbalanced datasets is impractical, as different MLLMs may exhibit different biases." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1745" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.189, + 0.089, + 0.805, + 0.297 + ], + "angle": 0, + "content": "
Model#ParamsSocialCounterfactualsFairFaceCMSCVQAv2MMBenchTextVQA
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLaVA7B-2.05670.3973-2.87920.6457-1.61591.481778.5064.6958.21
LLaVA+POPE-0.51010.4833-1.59330.6056-2.54241.1154---
LLaVA+FT-0.47270.3625-1.01990.4865-0.71420.805878.1163.8858.20
LLaVA+CSD-0.35090.3110-0.86220.3950-0.49330.563378.1664.2058.39
LLaVA13B-2.57300.3799-3.36040.5863-1.67300.535080.067.7061.30
LLaVA+POPE-0.38400.4410-0.95080.4051-2.25421.1454---
LLaVA+FT-0.47480.4051-1.31230.4066-1.51070.460579.1767.2861.02
LLaVA+CSD-0.31130.3718-0.71140.3752-0.81670.419279.7568.3061.42
Qwen-VL7B-0.61170.5966-1.63050.8469-1.51141.096179.3774.1461.39
Qwen-VL+POPE-0.30640.5399-1.31670.9207-2.24381.7575---
Qwen-VL+FT-0.28010.3966-1.39250.6916-0.81661.022779.3774.8260.86
Qwen-VL+CSD-0.24220.2921-0.76720.2607-0.61930.718079.3775.5960.88
Bunny8B-0.42550.6064-1.13750.5349-1.58291.417382.6076.4665.31
Bunny+POPE-0.33700.5899-1.36700.4918-2.80851.7269---
Bunny+FT-0.42020.5851-1.00350.4199-0.82370.902682.4876.3965.20
Bunny+CSD-0.40010.5532-0.90030.2632-0.77450.157582.4276.2865.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.302, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Table 3. Performance comparison. All models are fine-tuned on our CMSC dataset. Among the six datasets, SocialCounterfactuals, FairFace, and CMSC are employed to evaluate social bias; and VQAv2, MMBench, and TextVQA are general benchmarks. Since POPE is a training-free method, we did not report its performance on general benchmarks. #Params: the scale of the base LLM's parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.416 + ], + "angle": 0, + "content": "the approaches that have been proven effective in the class imbalance research area [23]. Instead of scaling loss based on the class frequency, we leverage the stereotype quantification metric Skew to rescale the loss value," + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.426, + 0.483, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {f a i r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} e ^ {- \\operatorname {S k e w} \\left(\\mathcal {P} _ {i}\\right)} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\text {i n s}} ^ {i}, \\mathbf {x} _ {\\text {i m g}} ^ {i}; \\theta\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.613 + ], + "angle": 0, + "content": "In this scenario, when \\(\\mathrm{Skew}(\\mathcal{P}_i) > 0\\), i.e., the MLLM tends to predict the input instance as the SC label \\(c_{i}\\), the fairness term \\(e^{-\\mathrm{Skew}(\\mathcal{P}_i)}\\) is less than 1.0. Consequently, this instance will receive less attention during training. Similarly, when \\(\\mathrm{Skew}(\\mathcal{P}_i) < 0\\), the fairness term will make the model pay more attention to this overlooked instance. This operation allows the model to dynamically adjust weights during fine-tuning based on the Skew of input instances, enabling the model to learn a fairer distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.223, + 0.645 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.19, + 0.669 + ], + "angle": 0, + "content": "6.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Datasets. In addition to the three social bias datasets mentioned in Section 4.1, we also employed three recent benchmark datasets to evaluate MLLMs' original zero-shot capabilities. Among them, VQAv2 [22] and TextVQA [56] are benchmarks for general visual question answering and text-oriented visual question answering, respectively. MM-Bench [41] evaluates the model robustness with comprehensive multiple-choice answers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Baselines. Given the limited exploration of debiasing strategies for MLLMs, we compared our CSD with the direct fine-tuning (FT), i.e., fine-tuning the model on balanced counterfactual dataset [63]. Moreover, we adapt POPE [37], which addresses hallucinations in MLLMs, as a baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Both thresholds \\(\\tau_{1}\\) and \\(\\tau_{2}\\) in Section 5.2 are set to 1.0. We trained the models on \\(8^{*}\\mathrm{A}800\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.855, + 0.371 + ], + "angle": 0, + "content": "GPUs. Each training session takes \\(\\sim 18\\) GPU hours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.383, + 0.652, + 0.398 + ], + "angle": 0, + "content": "6.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.907, + 0.693 + ], + "angle": 0, + "content": "We reported MaxSkew@C and MinSkew@C of the four utilized MLLMs in Table 3, wherein we have three main observations: i) Compared to the baselines, our CSD demonstrates the best debiasing effectiveness across four different MLLM architectures. In intra-dataset evaluation, i.e., fine-tuning and testing on the CMSC dataset, LLaVA-13B+CSD achieved a MinSkew@C of -0.8167, improving by \\(46\\%\\) compared to LLaVA-13B+FT. In cross-dataset evaluation, our CSD also outperforms the two baselines by a significant margin. For example, Qwen-VL-7B+CSD attains a MaxSkew@C of 0.2607 on the FairFace dataset, showing a reduction of 0.6 in absolute value. ii) A model with a larger size does not necessarily correspond to a lower social bias. This might because the increased scale of parameters makes the model to better learn the biases present in the pre-training dataset. iii) POPE's debiasing effectiveness underperforms FT. One possible reason is that POPE is a training-free method, causing its predictions to remain biased towards the pre-training data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Pertaining to the model results on general multi-modal benchmarks, we observed that our CSD method has a negligible impact, with the effect on all architectures across the three datasets being less than \\(0.5\\%\\). This indicates that our CSD method enhances the model's social fairness without significantly sacrificing its general capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.798, + 0.679, + 0.812 + ], + "angle": 0, + "content": "6.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.765, + 0.835 + ], + "angle": 0, + "content": "6.3.1. Comparison on Ablated CSD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our CSD method consists of two key components: the dataset resampling and the rescaled SFLoss. In Table 4, we reported the performance of them. We observe that both components help alleviate social bias. For example," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1746" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.468, + 0.31 + ], + "angle": 0, + "content": "
ModelSocialCounterfactualsFairFaceCMSC
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA-7B×-2.05670.3973-2.87920.6457-1.61591.4817
+FT-0.47270.3625-1.01990.4865-0.71420.8058
+Resample-0.44550.3327-0.92380.4838-0.69790.7741
+SFLoss-0.38510.3453-0.89040.4246-0.53850.5838
+CSD-0.35090.3110-0.86220.3950-0.49330.5633
LLAVA-13B×-2.57300.3799-3.36040.5863-1.67300.5350
+FT-0.47480.4051-1.31230.4066-1.51070.4605
+Resample-0.36030.3844-0.81540.3823-1.03640.4292
+SFLoss-0.34700.3905-1.05460.3946-1.03980.4388
+CSD-0.31130.3718-0.71140.3752-0.81670.4192
Qwen-VL-7B×-0.61170.5966-1.63050.8469-1.51141.0961
+FT-0.28010.3966-1.39250.6916-0.81661.0227
+Resample-0.24280.3631-0.80120.3207-0.74140.7813
+SFLoss-0.25800.3004-0.76880.3311-0.63860.8041
+CSD-0.24220.2921-0.76720.2607-0.61930.7180
Bunny-8B×-0.42550.6064-1.13750.5349-1.58291.4173
+FT-0.42020.5851-1.00350.4199-0.82370.9026
+Resample-0.41580.5670-0.98750.3127-0.81410.4008
+SFLoss-0.41710.5534-0.94480.2789-0.79410.4155
+CSD-0.40010.5532-0.90030.2632-0.77450.1575
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.315, + 0.483, + 0.358 + ], + "angle": 0, + "content": "Table 4. The performance of ablated CSD. FT: Directly fin-tuning. Resample: Fine-tuning with resampling. SFLoss: Fin-tuning with SFLoss. The best performance is highlighted in bold." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.365, + 0.285, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.365, + 0.449, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.484 + ], + "angle": 0, + "content": "Figure 5. MaxSkews of (a) Qwen-VL-7B and (b) Bunny-8B across occupations in SocialCounterfactuals for different races." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.59 + ], + "angle": 0, + "content": "LLaVA-7B+SFLoss achieves a \\(70\\%\\) MinSkew@C improvement on FairFace. On CMSC, Qwen-VL-7B+Resample achieves a MinSkew@C of -0.6193, reducing the absolute Skew value by more than 0.9. A notable observation is that either of the SFLoss and data resampling shows better debiasing effect compared to the naive FT strategy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.277, + 0.613 + ], + "angle": 0, + "content": "6.3.2. Comparison on SAs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.737 + ], + "angle": 0, + "content": "In Figure 5, we illustrate the MaxSkews [13] of Qwen-VL-7B and Bunny-8B across different races in the SocialCounterfactuals dataset. We can observe that both models exhibit significant social bias without any fine-tuning. For instance, Qwen-VL-7B has a Skew value of approximately 0.76 for White, indicating a strong preference for predicting occupations for this race. Our CSD method achieves better debiasing effects across all races for both models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.27, + 0.76 + ], + "angle": 0, + "content": "6.3.3. Skew Distributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Figure 6, we illustrate the MaxSkew distribution across different SCs in CMSC. It can be observed that LLaVA-13B exhibits larger mean and median MaxSkew values, along with significant outliers in certain attributes. For instance, in the 'personality' category, the Skew value for Old-Belligerent reaches 1.6. After direct fine-tuning, the average degree of social bias decreases. However, the outlier issue remains unresolved. In contrast, applying our CSD method achieves a more optimized Skew distribution" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.09, + 0.882, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Figure 6. MaxSkew distribution for LLaVA-13B on the CMSC dataset. Each box represents the different \\(\\mathrm{Skew}_{a|c}\\) values. Our CSD method achieves the most optimized Skew distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.337, + 0.907, + 0.382 + ], + "angle": 0, + "content": "across different concepts. Furthermore, noticeable outlier are eliminated, indicating that CSD provides a comprehensive debiasing effect for MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.395, + 0.763, + 0.41 + ], + "angle": 0, + "content": "7. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.42, + 0.906, + 0.571 + ], + "angle": 0, + "content": "Conclusion. In this paper, we present to address the notorious social bias problem in MLLMs. Our first contribution is a comprehensive dataset that covers more diverse social concepts than previous datasets. In addition, we advocate an counter-stereotype debiasing approach to perform both dataset resampling and loss rescaling, thereby improving fairness of MLLMs. Extensive experiments demonstrate that our method is promising to alleviate the social bias in MLLMs, with minimal negative impact on their original general multi-modal understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Social Impact. As MLLMs become increasingly integrated into real-world applications, their societal influence is poised to expand, particularly in sectors requiring sensitive handling of human attributes and social concepts. However, intrinsic social biases within these models can undermine their trustworthiness, raising ethical concerns and potential risks. This paper presents a bias mitigation approach to enhance fairness in MLLMs, fostering trust and accountability by reducing harmful stereotypes. We hope our work inspires further research on bias and fairness in MLLMs, contributing to the development of more equitable models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.75, + 0.665, + 0.766 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.775, + 0.906, + 0.835 + ], + "angle": 0, + "content": "This research is supported in part by National Natural Science Foundation of China (Grant No. 62476071, No. U24A20328); in part by National Natural Science Foundation of China (No. 62176137)." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.848, + 0.611, + 0.864 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.521, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1747" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.203 + ], + "angle": 0, + "content": "sch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 1-20, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[2] Mohsan S. Alvi, Andrew Zisserman, and Christoffer Nelläker. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings. In EC-CVW, pages 556-572, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. CoRR, abs/2308.12966:1-15, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[4] Hugo Berg, Siobhan Mackenzie Hall, Yash Bhalgat, Hannah Kirk, Aleksandar Shtedritski, and Max Bain. A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. In AACL/IJCNLP, pages 806-822, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[5] Karan Bhanot, Miao Qi, John S. Erickson, Isabelle Guyon, and Kristin P. Bennett. The problem of fairness in synthetic healthcare data. Entropy, 23(9):1165, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. CoRR, abs/2110.01963:1-18, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[7] Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Tauman Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, pages 4349-4357, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[8] Jannik Brinkmann, Paul Swoboda, and Christian Bartelt. A multidimensional analysis of social biases in vision transformers. In ICCV, pages 4914-4923, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[9] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1-40, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly-scaled multilingual language-image model. In ICLR, pages 1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[12] Ching-Yao Chuang and Youssef Mroueh. Fair mixup: Fairness via interpolation. In ICLR, pages 1-11, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Ching-Yao Chuang, Jampani Varun, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. CoRR, abs/2302.00070:1-13, 2023. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Ho. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.904, + 0.29 + ], + "angle": 0, + "content": "[15] Sepehr Dehdashtian, Lan Wang, and Vishnu Naresh Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In ICLR, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Richard B Felson and Noah Painter-Davis. Another cost of being a young black male: Race, weaponry, and lethal outcomes in assaults. Social Science Research, 41(5):1241-1253, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] Kathleen C. Fraser and Svetlana Kiritchenko. Examining gender and racial bias in large vision-language models using a novel dataset of parallel images. In EACL, pages 690-713, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[19] Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. Word embeddings quantify 100 years of gender and ethnic stereotypes. PNAS, 115(16):E3635-E3644, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.904, + 0.574 + ], + "angle": 0, + "content": "[20] Sahin Cem Geyik, Stuart Ambler, and Krishnamaram Kenthapadi. Fairness-aware ranking in search & recommendation systems with application to linkedin talent search. In KDD, pages 2221-2231, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.904, + 0.63 + ], + "angle": 0, + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.904, + 0.687 + ], + "angle": 0, + "content": "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6325-6334, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[23] Yangyang Guo, Liqiang Nie, Harry Cheng, Zhiyong Cheng, Mohan S. Kankanhalli, and Alberto Del Bimbo. On modality bias recognition and reduction. ACM ToMM, 19(3):103:1-103:22, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu, Grace Sodunke, Aleksandar Shtedritski, and Hannah Rose Kirk. Visogender: A dataset for benchmarking gender bias in image-text pronoun resolution. In NeurIPS, pages 1-15, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[25] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. CoRR, abs/2402.11530:1-13, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1748" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "editing with cross-attention control. In ICLR, pages 1-15, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[27] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[28] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Nils Johan Bertil Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. In NeurIPS, pages 1-22, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Ekaterina Iakovleva, Fabio Pizzati, Philip Torr, and Stéphane Lathuilière. Specify and edit: Overcoming ambiguity in text-based image editing. CoRR, abs/2407.20232:1-10, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[31] Sepehr Janghorbani and Gerard De Melo. Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision-language models. In EACL, pages 1725-1735, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[32] Fangkai Jiao, Zhiyang Teng, Bosheng Ding, Zhengyuan Liu, Nancy Chen, and Shafiq Joty. Exploring self-supervised logic-enhanced training for large language models. In NAACL, pages 926-941, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[33] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, pages 1548-1558, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[34] Tiep Le, Vasudev Lal, and Phillip Howard. Cocos-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. NeurIPS, pages 1-13, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[37] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, pages 1-13, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? CoRR, abs/2307.06281:1-12, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[42] Guangben Lu, Yuzhen Du, Zhimin Sun, Ran Yi, Yifan Qi, Yizhe Tang, Tianyi Wang, Lizhuang Ma, and Fangyuan Zou. Pinco: Position-induced consistent adapter for diffusion transformer in foreground-conditioned inpainting. arXiv preprint arXiv:2412.03812, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding. CoRR, abs/2403.05525:1-29, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[44] David Madras, Elliot Creager, Toniann Pitassi, and Richard S. Zemel. Learning adversarially fair and transferable representations. In ICML, pages 3381-3390, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[45] Thomas Manzini, Yao Chong Lim, Alan W. Black, and Yulia Tsvetkov. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. In NAACL, pages 615-621, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.454 + ], + "angle": 0, + "content": "[46] Allison Master. Gender stereotypes influence children's stem motivation. Child Development Perspectives, 15(3):203-210, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, pages 27730-27744, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[48] David S Pedulla. The positive consequences of negative stereotypes: Race, sexual orientation, and the job application process. Social Psychology Quarterly, 77(1):75-94, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952: 1-21, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21:140:1-140:67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[52] Vikram V. Ramaswamy, Sunnie S. Y. Kim, and Olga Russakovsky. Fair attribute classification through latent space de-biasing. In CVPR, pages 9301-9310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck, Estelle Aflalo, Shao-Yen Tseng, Vasudev Lal, and Phillip" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1749" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.133 + ], + "angle": 0, + "content": "Howard. Debias your large multi-modal model at test-time with non-contrastive visual attribute steering. arXiv preprint arXiv:2411.12590, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[54] Prasanna Sattigeri, Samuel C. Hoffman, Vijil Chenthamarakshan, and Kush R. Varshney. Fairness GAN: generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3:1-3:9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[55] Ashish Seth, Mayur Hemani, and Chirag Agarwal. Dear: Debiasing vision-language models with additive residuals. In CVPR, pages 6820-6829, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[56] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: A dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[57] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: pre-training of generic visual-linguistic representations. In ICLR, pages 1–13, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[58] Yizhe Tang, Zhimin Sun, Yuzhen Du, Ran Yi, Guangben Lu, Teng Hu, Luying Li, Lizhuang Ma, and Fangyuan Zou. Ata: Adaptive transformation agent for text-guided subject-position variable background inpainting. In CVPR, pages 18335-18345, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[59] Özge Tarhan. Children's understanding of the concept of social stereotypes. Child Indicators Research, 15(3):989-1023, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971:1-16, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[61] Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. Are emily and greg still more employable than lakisha and jamal? investigating algorithmic hiring bias in the era of chatgpt. CoRR, abs/2310.05135:1-13, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[62] W Edgar Vinacke. Stereotypes as social concepts. The journal of social Psychology, 46(2):229-243, 1957. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[63] Angelina Wang and Olga Russakovsky. Overwriting pretrained bias with finetuning data. In ICCV, pages 3934-3945, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[64] Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy Chen. SeaEval for multilingual foundation models: From cross-lingual alignment to cultural reasoning. In NAACL, pages 370–390. Association for Computational Linguistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[65] Jialu Wang, Yang Liu, and Xin Eric Wang. Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. In EMNLP, pages 1995-2008, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[66] Mei Wang and Weihong Deng. Mitigating bias in face recognition using skewness-aware reinforcement learning. In CVPR, pages 9319-9328, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[67] Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. Evaluating fairness in large vision-language models across diverse demographic attributes and prompts. CoRR, abs/2406.17974:1-11, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[68] Yisong Xiao, Aishan Liu, QianJia Cheng, Zhenfei Yin, Siyuan Liang, Jiapeng Li, Jing Shao, Xianglong Liu, and Dacheng Tao. Genderbias-vl: Benchmarking gender bias in vision language models via counterfactual probing. CoRR, abs/2407.00600:1-13, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[69] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, pages 1-13, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[70] Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE TPAMI, 45(12):14144–14160, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[71] Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In ICML, pages 59310-59328. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[72] Jie Zhang, Sibo Wang, Xiangkui Cao, Zheng Yuan, Shiguang Shan, Xilin Chen, and Wen Gao. Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model. CoRR, abs/2406.14194:1-13, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[73] Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. NeurIPS, pages 21682-21697, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[74] Xiaobin Zhang, Ziqi Cui, Ning He, and Xuqun You. Are boys associated with weapons and girls associated with kitchenware? the extent to which gender stereotypes regarding adults extend to children. Acta psychologica, 230: 103754, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[75] Kankan Zhou, Eason Lai, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In AACL/IJCNLP, pages 527-538, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1750" + } + ] +] \ No newline at end of file diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_origin.pdf b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d61d8480f2836906bfeaddecfb34f4209899de8 --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/c46b809c-5d37-4c83-a43f-cbf8d15658b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b878fc1e327f7bb907e53104100c3a47828e3422f113e52391fd2d08a6d8f4b +size 4083273 diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/full.md b/2025/Social Debiasing for Fair Multi-modal LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eeea04ce59b23ba958079859599f29bc90b4989a --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/full.md @@ -0,0 +1,388 @@ +# Social Debiasing for Fair Multi-modal LLMs + +Harry Cheng $^{1}$ , Yangyang Guo $^{2}$ , Qingpei Guo $^{3,4,*}$ , Ming Yang $^{4}$ , Tian Gan $^{1}$ , Weili Guan $^{5}$ , Liqiang Nie $^{5}$ $^{1}$ Shandong University, $^{2}$ National University of Singapore, $^{3}$ University of Warwick, $^{4}$ Ant Group, $^{5}$ Harbin Institute of Technology (Shenzhen) + +xaCheng1996@gmail.com, guoyang.eric@gmail.com, qingpei.gqp@antgroup.com, m.yang@antgroup.com, gantian@sdu.edu.cn, honeyguan@gmail.com, nieliqiang@gmail.com + +# Abstract + +Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks. + +# 1. Introduction + +Multi-modal Large Language Models (MLLMs) have revolutionized the field of general-purpose vision-language understanding. Representative models, such as LLaVA [38], Qwen-VL [3], and Bunny [25], exhibit remarkable zero-shot performance and can be easily fine-tuned for diverse downstream applications. + +Despite MLLMs' widespread use, it is imperative to recognize that these models can exhibit severe social biases with respect to attributes such as race and gender [18, 65]. Figure 1a illustrates one example that a biased MLLM is + +![](images/65d3a9cf4c9d75841914648b70a3c4a0ec34c40ef09447d883cfe1466deba406.jpg) +Figure 1. Social bias examples. Left: Gender-biased prediction on Nurse from LLaVA-7B1. Right: Maximum Skews across different human attributes on the FairFace dataset from Qwen-VL-7B and LLaVA-7B. As a metric for measuring model bias, a larger Skew implies a higher degree of social bias. Our CSD method outperforms existing approaches by a significant margin. + +much more likely to associate the role of nurse with female rather than male, which may be stereotypical. Such biased predictions often occur unconsciously within MLLMs, making them difficult to detect and avoid by explicit rules. A key reason for this issue lies in the composition of the MLLM training data, which can contain content related to violence or racism [6]. These inappropriate samples may collectively demonstrate some stereotypes, resulting in subtle yet uncomfortable responses [8, 63]. + +Existing studies on mitigating the social bias problem in MLLMs remain largely under-explored. A naive approach is to collect attribute-balanced vision-language counterfactual datasets [24, 34], which can be used to directly finetune biased models to predict fairer distributions. However, as demonstrated in Table 1, existing large-scale datasets are limited by their focus on a single social concept, such as occupation [27], without attending to multifaceted social stereotypes [62]. This significantly hampers the model in learning more comprehensive representations. Therefore, we are motivated to construct a more extensive and diverse counterfactual dataset. From a methodological perspective, directly fine-tuning on a counterfactual dataset assigns equal importance to instances receiving different social biases, resulting in sub-optimal debiasing performance (as demonstrated in Section 6.2). Analogy to a common solution in chemistry—neutralizing acidic water requires + +
DatasetVenueType#ImagesSocial Concepts
CoCo-Counterfactuals [34]NeurIPS'23General34KNo annotations
FairFace [33]WACV'21Social108KNo annotations
VisoGender [24]NeurIPS'23Social0.6KOccupation
PATA [55]CVPR'23Social5KOccupation
SocialCounterfactuals [27]CVPR'24Social171KOccupation
MM-Bias [31]EACL'23Social3K14 minorities
CMSC (Ours)-Social60K18 balanced concepts
+ +Table 1. Comparison of CMSC with mainstream datasets used for model debiasing. Existing datasets are limited either by scale or by the number of social concepts they cover. Type - The category of biases the dataset focuses on. + +adding an alkaline substance rather than plain water—we are prompted to speculate: Can we leverage the opposite of the suffered social bias to build a fairer model? + +To address these two issues, we first construct a large-scale, high-quality Counterfactual dataset with Multiple Social Concepts (CMSC). As demonstrated in Table 1, our CMSC dataset narrows the gap in data scale and concept richness. We conduct extensive experiments on four different MLLM architectures [3, 25, 38] using several existing counterfactual datasets as well as our CMSC. The results demonstrate that MLLMs fine-tuned on CMSC exhibit significantly lower social bias compared to those fine-tuned on single-concept datasets. For instance, LLaVA-7B [38] fine-tuned on CMSC achieves a $64\%$ debiasing effect, which significantly outperforms the $30\%$ bias reduction achieved when fine-tuned on the SocialCounterfactuals dataset $[27]^2$ . This indicates that fine-tuning the models with a rich set of social concept is beneficial in MLLM debiasing. + +Furthermore, to implement the idea of 'debiasing with the opposite of the social biases', we propose a Counter-Stereotype Debiasing strategy (CSD) to effectively reduce social biases in MLLMs. Our CSD is equipped with two techniques: 1) we design a novel data sampling method based on the bias level, 2) we rescale the previously used autoregressive loss function to a new Social Fairness Loss (SFLoss). In this manner, previously under-represented cases, e.g., male nurses, will receive more attention, serving as new 'alkaline' ones to counteract the bias of MLLMs that prefer female nurses. The experimental results demonstrate that our CSD method is an effective debiasing strategy. As illustrated in Figure 1b, our method reduces the bias by over $50\%$ for the Qwen-VL-7B model compared to prior training strategies, especially naive fine-tuning. Additionally, CSD does not compromise the original model performance on general multi-modal benchmarks. + +In summary, our contributions are three-fold: + +- We construct a high-quality counterfactual dataset that includes 18 social concepts, which demonstrated to be superior to existing ones on MLLM social debiasing. +- We propose a novel approach that applying counter-stereotype debiasing to mitigate social biases. To the best + +of our knowledge, this is the first research effort dedicated to addressing the social bias problem in autoregressive MLLMs. + +- We apply our debiasing strategy to four multi-modal LLMs architectures to show the generalization capability and effectiveness of our method. + +# 2. Related Work + +# 2.1. Multi-modal Large Language Model + +With the rapid development of LLMs [9, 51, 57, 60, 64], increasing efforts have been dedicated to extending the powerful reasoning capabilities of LLMs to multi-modal applications [1, 10, 14, 28, 32, 35, 40, 42, 58, 70, 71]. Specifically, MLLMs utilize LLMs as the foundational base, aligning features from other modalities with text embeddings to enable LLMs to perceive multi-modal inputs. For instance, BLIP-2 [36] utilizes a Q-Former to align vision encoders with LLMs. LLaVA [38, 39] presents to directly map vision features to the word embedding space of Vicuna [11]. Qwen-VL [3] employs a single-layer cross-attention module as the vision-language adapter, and introduces multitask pre-training to improve the model performance. These approaches significantly enhance multi-modal understanding capabilities. However, some studies highlight the existence of social biases in pre-trained LLMs [7, 8, 18]. These biases often manifest as harmful spurious correlations related to human attributes such as gender and race, which significantly undermine the fairness. For instance, LLMs may 'efficiently' filter resumes based on race rather than the candidates' qualifications [61]. Nevertheless, research on social biases in MLLMs remains largely unexplored [68]. + +# 2.2. Social Bias Reduction + +We roughly categorize the bias mitigation strategies into two groups: data-based and objective-based. + +Data-based debiasing typically refers to data augmentation techniques. For instance, Chuang et al. [12] employ mixup [69] to construct interpolated samples among groups with different distributions. Ramaswamy et al. [52] utilize perturbed GAN-generated images [21] in latent space to augment the original dataset. Moreover, some studies collect fair datasets across human attributes for continual finetuning [5, 24, 31, 34, 54, 72, 75]. In particular, FairFace [33] contains $108\mathrm{K}$ images that are balanced for the race attribute. Socialcounterfactuals [27] collects $171\mathrm{K}$ imagetext pairs to probe biases across race, gender, and physical characteristics. However, these collected counterfactual datasets often focus on only one common concept - occupation, which hinders effective debiasing [67]. Though a few datasets involve several concepts, they nevertheless, are largely limited by their data scale (e.g., 3K [31]), making it less feasible to fine-tune MLLMs. + +![](images/37ca1d70d0659a45d23a2014baf3a63c47507e136d54b8357619328d3c7eb611.jpg) +Figure 2. Image Generation Pipeline. Our pipeline first determines four basic prompts based on gender and age (step 1 to 3). Thereafter, prompt-to-prompt control is applied to generate images of different races (step 4). Finally, we filter out low-quality images (step 5). + +Objective-based debiasing modifies the model training process to achieve better fairness. Early studies are proposed to reduce social bias within uni-modal models [2, 17, 19, 44, 45, 66]. For example, Bolukbasi et al. [7] optimizes word embeddings to remove gender stereotypes. In contrast, multi-modal debiasing methods often focus on contrastive learning-based vision language models, e.g., CLIP [50]. Representative approaches are to rectify the modality similarity matrix in VLMs [15] via learning additive adapters [55, 73], eliminating biased directions [13], and using adversarial samples [4]. However, among these methods, uni-modal approaches often require re-training the full model. This operation, though plausible for previous small-scale models, is less practical for existing MLLMs with billions of parameters [53]. On the other hand, mainstream multi-modal techniques cannot be utilized for these models due to the divergent training objectives between MLLMs and previous CLIP-style models. In particular, MLLMs are mostly trained in an autoregressive way rather than through contrastive learning. + +# 3. CMSC Dataset Construction + +As demonstrated in Table 1, existing counterfactual datasets are either restricted by their small scale [24] or a narrow coverage of concepts [33]. To bridge this gap between limited existing datasets and diverse real-world stereotypes, we introduce the CMSC dataset, encompassing 60k high-quality images across eighteen social concepts. + +# 3.1. Social Attributes and Concepts + +A counterfactual dataset for social bias reduction and evaluation typically contains two key aspects: Social Attribute (SA) and Social Concept (SC). The former, i.e., SA, is defined as characteristics shared by a group of people [33]. Specifically, we investigate three types of attributes: i) genders: {Male, Female}; ii) races: {White, Black, Indian, Southeast Asian, East Asian, Middle Eastern, Latino}; and iii) ages: {Young, Old}. These attributes are intrinsic to individuals3. For each image $\mathbf{I}_i$ in CMSC, three SA labels are provided with respect to these three types of SA, and are + +combined as a SA set $\mathcal{A}_i$ . + +As for the SC, we define it as the societal label attributed to an individual. Drawing from the sociological research [16, 46, 48, 59, 74], we employ 18 SCs in CMSC. In particular, these SCs are categorized into three groups: personality, responsibility, and education. Specifically, personality relates to concepts pertaining to an individual's character. We use five concepts: {compassionate, belligerent, authority, pleasant, unpleasant}. Responsibility is the roles or duties that individuals are expected to fulfill in society or family. We identify six concepts: {tool user, weapon user, career, family, chef working, earning money}. Education pertains to the level of education a person has received. We include seven concepts: {middle school, high school, university, good student, bad student, science, arts}. Each image in CMSC is annotated with one SC label $c_i \in \mathcal{C}$ , where $\mathcal{C}$ is the union of the concepts from all above groups. + +In summary, each instance $\mathcal{P}_i$ in our dataset consists of an image $\mathbf{I}_i$ , a set of SA labels $\mathcal{A}_i$ , and a SC label $c_i$ . To the best of our knowledge, CMSC is the first large-scale counterfactual dataset that includes a variety of social concepts. + +# 3.2. Image Generation Pipeline + +We employ Stable Diffusion XL (SDXL) [49] to generate instances in CMSC. For each SC, we carefully design prompts that provide a detailed explanation of the concept. These prompts not only help eliminate potential ambiguities associated with the SC but also enhance the quality of the generated images [29]. For instance, for the SC 'pleasant,' we use the prompt: 'A photo of a pleasant [race] [gender] person at [age], who has a friendly smile, a relaxed posture, and ...' To ensure diverse and representative image generation, these prompts are randomly modified by replacing words with synonyms during image generation. Subsequently, the intersectional generation strategy [27] is applied. As illustrated in Figure 2, we first determine gender and age (Step 2) and generate four base images (Step 3). Thereafter, we apply Prompt-to-Prompt (P2P) control [26] to produce visually similar images that vary only in terms of race (Step 4). For each SC, we create 28 prompts across attribute combinations. Each prompt helps generate 100 images. To ensure image quality, we apply a rigorous image filtering mechanism (step 5). First, we use CLIP-Score [50] + +to filter out images that are semantically misaligned with the prompts. Subsequently, a panel of three experts conducts a voting process to remove low-quality and NSFW content. If the generated images perform poorly on a specific SA, the entire set of images (including those of other SAs) will be discarded. Moreover, we observe that the model can generate high-quality images for certain SCs or SAs, while struggling to do so for others. As a result, we filter out excessively generated ones for balanced distribution. Our filtering process is very strict. For instance, the testing set initially contains over 50,000 images, but only 12,019 images are finally used in CMSC. A similar process is applied to the training set, which comprises 48,134 images. In total, our dataset contains 60K images. + +# 4. CMSC Dataset Evaluation + +# 4.1. Experiment Settings + +Evaluation Protocols. We employed four different MLLM architectures, i.e., LLaVA-7B [38], LLaVA-13B [38], Qwen-VL-7B [3], and Bunny-8B [25], to evaluate our CMSC dataset. For each MLLM, we used the same template to collect their responses: 'Based on the race, gender, and age of the face , do you think the person is {SCs}? Respond with one of them.' Here, is the image from the dataset, and {SCs} represents the list of used SCs. For instance, when testing personality, {SCs} includes 'pleasant,' 'unpleasant,' etc. After collecting these predictions, we used them to compute MaxSkew@C and MinSkew@C (which will be elaborated in the following section) to measure bias. + +Metrics. We employ Skew-based [20] metrics to quantify the extent of social biases. Specifically, for a given dataset $\mathcal{D}$ , we define the subset containing instances with a specific SC label $c'$ as $\mathcal{D}_{c'}$ , + +$$ +\mathcal {D} _ {c ^ {\prime}} = \left\{\mathcal {P} _ {i} \mid \mathcal {P} _ {i} \in \mathcal {D}, c _ {i} = c ^ {\prime} \right\}. \tag {1} +$$ + +From $\mathcal{D}_{c'}$ , we further extract the subset consisting of instances with a specific SA label $a'$ as $\mathcal{D}_{a'|c'}$ , + +$$ +\mathcal {D} _ {a ^ {\prime} \mid c ^ {\prime}} = \left\{\mathcal {P} _ {i} \mid \mathcal {P} _ {i} \in \mathcal {D} _ {c ^ {\prime}}, a ^ {\prime} \in \mathcal {A} _ {i} \right\}. \tag {2} +$$ + +We then utilize the aforementioned template to guide MLLMs in predicting the SC label corresponding to each $\mathcal{P}_i$ . We denote the predicted SC label as $\hat{c}_i$ . By applying this process to the entire dataset $\mathcal{D}$ , we can construct a new predicted set $\hat{\mathcal{D}} = \{\hat{\mathcal{P}}_i\}_{i=1}^N$ , where $\hat{\mathcal{P}}_i = \{\mathbf{I}_i, \mathcal{A}_i, \hat{c}_i\}$ , and $N = |\mathcal{D}|$ . In $\hat{\mathcal{D}}$ , subsets $\hat{\mathcal{D}}_{c'}$ and $\hat{\mathcal{D}}_{a'|c'}$ can be derived with Equation (1) and Equation (2), respectively. The Skew for SC $c'$ and SA $a'$ could be formulated as + +$$ +\operatorname {S k e w} _ {a ^ {\prime} \mid c ^ {\prime}} = \log \left(\frac {\hat {\gamma} _ {a ^ {\prime}} \mid c ^ {\prime}}{\gamma_ {a ^ {\prime}} \mid c ^ {\prime}}\right), \tag {3} +$$ + +![](images/f68790acb97a2aa051ef1aec18b1d1c6c30527132566eb54a8f59f72122c616d.jpg) +Figure 3. Performance of MLLMs when evaluated on FairFace. In each model, the bars labeled 'FT w SCounter' and 'FT w CMSC' represent the model fine-tuned on SocialCounterfactuals and our CMSC, respectively. + +where + +$$ +\left\{ \begin{array}{l} \hat {\gamma} _ {a ^ {\prime} | c ^ {\prime}} = \frac {\left| \hat {\mathcal {D}} _ {a ^ {\prime} | c ^ {\prime}} \right|}{\left| \hat {\mathcal {D}} _ {c ^ {\prime}} \right|}, \\ \gamma_ {a ^ {\prime} | c ^ {\prime}} = \frac {\left| \mathcal {D} _ {a ^ {\prime} | c ^ {\prime}} \right|}{\left| \mathcal {D} _ {c ^ {\prime}} \right|}. \end{array} \right. \tag {4} +$$ + +When $\mathrm{Skew}_{a'|c'} > 0$ , the MLLM tends to predict instances with attribute $a'$ as concept $c'$ . For instance, in Figure 1a, $\mathrm{Skew}_{\text{Female|Nurse}} > 0$ . In contrast, when $\mathrm{Skew}_{a'|c'} < 0$ , the MLLM is inclined to not predict these instances containing SA $a'$ as SC $c'$ , e.g., predicting male as nurse. A fair MLLM should have Skew close to 0 across all concepts and attributes, i.e., $\sum_{a,c} |\mathrm{Skew}_{a|c}| \to 0$ . + +Although Skew can effectively measure MLLMs' bias towards a particular SA-SC pair, a counterfactual dataset often contains hundreds of such combinations, e.g., our CMSC includes 198 SA-SC pairs. This confounds the comprehensive assessment of MLLM bias. Therefore, we propose two variants of Skew to enable a more thorough analysis. Specifically, we first identify the maximum and minimum Skew values for each SC across all SAs [13], + +$$ +\left\{ \begin{array}{l} \operatorname {M a x S k e w} _ {c ^ {\prime}} = \operatorname {M a x} _ {a _ {i} \in \mathcal {A}} \left\{\operatorname {S k e w} _ {a _ {i} \mid c ^ {\prime}} \right\}, \\ \operatorname {M i n S k e w} _ {c ^ {\prime}} = \operatorname {M i n} _ {a _ {i} \in \mathcal {A}} \left\{\operatorname {S k e w} _ {a _ {i} \mid c ^ {\prime}} \right\}. \end{array} \right. \tag {5} +$$ + +For all $c' \in \mathcal{C}$ , we separately calculate the average of $\mathrm{MaxSkew}_{c'}$ and $\mathrm{MinSkew}_{c'}$ , resulting in two aggregated Skew values. We refer to these two values as MaxSkew@C (MaxS@C) and MinSkew@C (MinS@C), which represent the overall bias level of the MLLM. Both of these metrics indicate better fairness as they approach zero. + +Datasets. We compared our dataset with the previous state-of-the-art counterfactual dataset. Specifically, we separately fine-tuned MLLMs on CMSC and SocialCounterfactual [27] to evaluate their effectiveness in mitigating bias. The SocialCounterfactual dataset focuses on the SC of occupation and contains 171K high-quality synthetic images. Moreover, we employed the FairFace dataset [33], which contains 108K images with a balanced distribution of SAs, + +
FT w/ SCPer.Res.Edu.
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA×-0.94862.4950-0.76622.2188-0.85693.9821
+ FT Per.-0.84691.6675-0.59152.2025-0.60943.1794
+ FT Res.-0.91521.7409-0.49211.4397-0.81582.7705
+ FT Edu.-0.97561.7842-0.68721.5719-0.52572.4362
Qwen.VL×-2.46882.5772-2.53012.2663-2.82422.4059
+ FT Per.-0.73861.2509-1.90770.3470-2.55242.3506
+ FT Res.-0.73841.9044-1.32050.1202-2.85452.3872
+ FT Edu.-0.84311.6817-1.87590.2236-2.47062.3502
+ +Table 2. Performance comparison of LLaVA-7B and Qwen-VL-7B when fine-tuned and tested on SCs in CMSC. Per.: Personality, Res.: Responsibility, Edu.: Education, $\times$ : Original performance. + +as an additional testing set to assess the cross-dataset debiasing performance of models fine-tuned on CMSC and SocialCounterfactual [55]. + +# 4.2. Comparison on Fine-tuning + +We reported MinSkew@C and MaxSkew@C of several MLLMs in Figure 3. The models, i.e., LLaVA [38], Qwen-VL [3], and Bunny [25], are fine-tuned on the SocialCounterfactuals and CMSC datasets individually and evaluated on the FairFace dataset, respectively. For clarity, we took the absolute values of both metrics. From Figure 3, we observed that models fine-tuned on CMSC exhibit superior debiasing effects. For instance, LLaVA-13B achieves a MinSkew@C of -1.3132, a significant advantage over that from the model fine-tuned on SocialCounterfactuals. This implies that the broader range of social concepts in CMSC enables the model to learn fairer distributions. + +# 4.3. Fine-tuning with Specific SC Group + +As discussed in Section 3.1, we split the SCs in CMSC into three groups: personality, responsibility, and education. We applied LLaVA and Qwen-VL to fine-tune separately on these subsets. The results are reported in Table 2. It can be observed that in intra-subset evaluations, i.e., fine-tuned and tested on the same subset, the model can generally achieve lower biases. For instance, when fine-tuning and evaluating Qwen-VL on the personality subset, it achieves a MaxS@C of 1.2509, showing an absolute difference of 0.65 and 0.43 compared to performance when fine-tuned on the responsibility and education subsets, respectively. + +# 4.4. Comparison on Image Distribution + +We calculated the Fréchet Inception Distance (FID) scores for the synthetic counterfactual datasets SocialCounterfactuals and our CMSC. Specifically, we randomly sampled 1,000 synthetic images from each dataset and then computed the distributional differences with the same set of 1,000 real images. A lower FID signifies better image quality. With this process, the two datasets received FID scores of 27.17 and 24.35, respectively. This indicates that the images in our dataset bear a closer resemblance to reality. + +# 5. Counter-stereotype Debiasing + +# 5.1. Preliminaries + +Before introducing our proposed method, we first revisit the training objectives of MLLMs. Subsequently, we introduce the concept of $\mathrm{Skew}(\mathcal{P}_i)$ to measure stereotypes associated with specific instances, upon which we base the reformulation of the autoregressive fine-tuning paradigm of MLLMs. + +# 5.1.1. Traning Objective of MLLMs + +Current mainstream MLLMs employ pre-trained LLMs [11, 60] as the output interface [3, 25, 38, 43]. Under this context, the base LLM is often trained in an autoregressive way, + +$$ +\mathrm {P} \left(\mathbf {x} _ {1: T}; \theta\right) = \prod_ {t = 1} ^ {T} \mathrm {P} \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {< t}; \theta\right), \tag {6} +$$ + +where $\mathbf{x}$ is a text sequence with $T$ tokens, and $\theta$ denotes the model parameters. During pre-training, the LLM predicts the $t$ -th token $\mathbf{x}_t$ based on all preceding ones, i.e., $\mathbf{x}_{< t}$ . After that, another instruction tuning stage [47] is often followed to enable LLMs to better understand user intentions, + +$$ +\mathrm {P} \left(\mathbf {y} _ {1: T} \mid \mathbf {x} _ {\text {i n s}}; \theta\right) = \prod_ {t = 1} ^ {T} \mathrm {P} \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {< t}, \mathbf {x} _ {\text {i n s}}; \theta\right), \tag {7} +$$ + +where $\mathbf{x}_{\mathrm{ins}}$ and $\mathbf{y}$ are textual instructions and responses, respectively. MLLMs extend the inputs of LLMs with the supplement of image features $\mathbf{x}_{\mathrm{img}}$ that are extracted using pre-trained vision encoders such as ViT [30, 50], + +$$ +\mathrm {P} \left(\mathbf {y} _ {1: T} \mid \mathbf {x} _ {\text {i n s}}, \mathbf {x} _ {\text {i m g}}; \theta\right) = \prod_ {t = 1} ^ {T} \mathrm {P} \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {< t}, \mathbf {x} _ {\text {i n s}}, \mathbf {x} _ {\text {i m g}}; \theta\right), \tag {8} +$$ + +where $\mathbf{x}_{\mathrm{img}}$ is aligned with text features via a connector, e.g., a trainable projection matrix. To optimize the pipeline, previous cross-entropy loss from LLMs is directly inherited, + +$$ +\mathcal {L} (\mathbf {y}, \mathbf {x} _ {\text {i n s}}, \mathbf {x} _ {\text {i m g}}; \theta) = - \sum_ {t = 1} ^ {T} \log \mathrm {P} \left(\mathbf {y} _ {t} \mid \mathbf {y} _ {< t}, \mathbf {x} _ {\text {i n s}}, \mathbf {x} _ {\text {i m g}}; \theta\right). \tag {9} +$$ + +# 5.1.2. Stereotype Measurement + +Recall that $\mathrm{Skew}_{a|c}$ quantifies the degree of social biases in an MLLM across the entire dataset. In this section, we are more interested in the social bias degree for each specific instance $\mathcal{P}_i$ . We then define $\mathrm{Skew}(\mathcal{P}_i)$ as a selected $\mathrm{Skew}_{a|c}$ with the maximum absolute value across all SAs for instance $\mathcal{P}_i$ . For example, suppose that $\mathcal{P}_i$ contains three SA labels: {'White,' Female,' 'Young'}, along with one SC label: 'Nurse.' $\mathrm{Skew}(\mathcal{P}_i)$ is the one with the highest absolute value from $\{\mathrm{Skew}_{White|Nurse}, \mathrm{Skew}_{Female|Nurse}, \mathrm{Skew}_{Young|Nurse}\}$ . It describes the SA that receives the most severe bias, thereby indicating the level of bias in $\mathcal{P}_i$ . + +Figure 4. Comparison between fine-tuning (upper part) and our CSD (lower part). CSD adjusts the loss and datasets based on $\mathrm{Skew}(\mathcal{P}_i)$ , guiding the model to focus on overlooked instances. +![](images/ad79bf07f31990abf72184d663161dfa8b9fbb59190a858fbacfa135387412c4.jpg) +: Parameter updates based on a specific loss function. + +# 5.2. CSD Approach + +As illustrated in Figure 4, vanilla fine-tuning approach involves sampling images from a balanced dataset and updating MLLMs through the original autoregressive objective. We argue that this method, which treats all instances equally, is ineffective in addressing the social bias problem in MLLMs. Therefore, we propose an CSD method from the view of counter-stereotype. Specifically, our CSD is composed of two components: i) Dataset Resampling, which enhances the data sampling process to include more underrepresented instances. ii) Loss Rescaling, where we adjust the loss function to place larger emphasis on instances that are overlooked in terms of social attributes. + +Dataset Resampling. The debiasing approach, e.g., directly fine-tuning, utilizes counterfactual datasets that are balanced across all SAs. We believe that such datasets make it difficult for MLLMs to recognize which SAs are subjected to greater bias and which are not4. To address this, we resample the dataset to increase the frequency of instances that MLLMs tend to ignore and reduce the frequency of emphasized ones, thereby 'neutralizing' social biases. + +$\operatorname{Skew}(\mathcal{P}_i)$ serves as an indicator for dataset resampling. As demonstrated in Algorithm 1, for each instance $\mathcal{P}_i$ , when $\operatorname{Skew}(\mathcal{P}_i) > 0$ , i.e., $\mathcal{P}_i$ has received more attention – such as the 'female-nurse' in Figure 1 – we reduce its probability in the resampled dataset $\mathcal{D}_r$ for the following training epoch, + +$$ +\mathcal {D} _ {r} = \left\{ \begin{array}{l} \mathcal {D} _ {r} \cup \mathcal {P} _ {i}, \operatorname {r a n d} (0, \operatorname {S k e w} (\mathcal {P} _ {i}) + \tau_ {1}) > \operatorname {S k e w} (\mathcal {P} _ {i}), \\ \mathcal {D} _ {r}, \operatorname {r a n d} (0, \operatorname {S k e w} (\mathcal {P} _ {i}) + \tau_ {1}) \leq \operatorname {S k e w} (\mathcal {P} _ {i}), \end{array} \right. \tag {10} +$$ + +where $\operatorname{rand}(\cdot)$ is to randomly draw from 0 to $\operatorname{Skew}(\mathcal{P}_i) + \tau_1$ , and $\tau_{1}$ is a pre-defined threshold. As such, a larger $\operatorname{Skew}(\mathcal{P}_i)$ corresponds to a lower chance of being included by $\mathcal{D}_r$ . In contrast, for instances with $\operatorname{Skew}(\mathcal{P}_i) \leq 0$ , we believe that increasing their proportion in the new dataset $\mathcal{D}_r$ is beneficial for the MLLM to learn the features of these overlooked parts, thereby achieving a fairer distribution. + +Algorithm 1 Dataset Resampling with Skew $(\mathcal{P}_i)$ +Input: The original dataset $\mathcal{D} = \{\mathbf{I}_i, \mathcal{A}_i, c_i\}$ , the social attribute sets $\mathcal{A}$ , the social concept set $\mathcal{C}$ +Output: The resampled dataset $\mathcal{D}_r$ +1: $\mathcal{D}_r\gets []$ ,AcmSkew $a_{i}|c_{i}\gets 0$ ; Initialization +2: for $\mathcal{P}_i\in \mathcal{D}$ do +3: if Skew $(\mathcal{P}_i) > 0$ then +4: if Rand(0,Skew $(\mathcal{P}_i) + \tau_1) > \mathrm{Skew}(\mathcal{P}_i)$ then +5: $\mathcal{D}_r\gets \mathcal{D}_r\cup \mathcal{P}_i;$ +6: end if +7: else $\triangleright$ Skew $(\mathcal{P}_i)\leq 0$ +8: $\mathcal{D}_r\gets \mathcal{D}_r\cup \mathcal{P}_i;$ +9: AcmSkew $a_{i}|c_{i}\gets$ AcmSkew $a_{i}|c_{i} + |\mathrm{Skew}(\mathcal{P}_i)|;$ +10: if AcmSkew $a_{i}|c_{i} > \tau_{2}$ then +11: $\mathcal{D}_r\gets \mathcal{D}_r\cup \mathcal{P}_i;$ Over-resampling +12: AcmSkew $a_{i}|c_{i}\gets 0$ +13: end if +14: end if +15: end for +16: return $\mathcal{D}_r$ + +To this end, these instances are directly accepted into $\mathcal{D}_r$ . Moreover, we design an over-resampling mechanism to further increase the occurrence frequency of these instances. Specifically, for each $\mathcal{P}_i$ , we employ an accumulative value, AcmSkew, to gradually accumulate the current $\mathrm{Skew}(\mathcal{P}_i)$ + +$$ +\operatorname {A c m S k e w} _ {a _ {i} \mid c _ {i}} = \operatorname {A c m S k e w} _ {a _ {i} \mid c _ {i}} + | \operatorname {S k e w} (\mathcal {P} _ {i}) |, \tag {11} +$$ + +where $c_{i}$ and $a_{i}$ are SC label and SA label corresponding to $\operatorname{Skew}(\mathcal{P}_i)$ , respectively. When $\operatorname{AcmSkew}_{a_i|c_i}$ exceeds a threshold $\tau_2$ , we add the $\mathcal{P}_i$ into $\mathcal{D}_r$ again. The $\operatorname{AcmSkew}_{a_i|c_i}$ can then be set to 0 for a new round of accumulation. With the above operations, instances with $\operatorname{Skew}(\mathcal{P}_i) \leq 0$ will be resampled multiple times. This resampling process is executed once before each training epoch. For evaluation, we employ the balanced testing set to ensure a fair comparison. + +Loss Rescaling. During MLLM fine-tuning, we rescale the autoregressive loss in Equation (9) to a new Social Fairness Loss (SFLoss) for effective debiasing. Specifically, the empirical risk during training can be represented as + +$$ +\mathcal {E} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathcal {L} \left(\mathbf {y} ^ {i}, \mathbf {x} _ {\mathrm {i n s}} ^ {i}, \mathbf {x} _ {\mathrm {i m g}} ^ {i}; \theta\right), \tag {12} +$$ + +where $N = |\mathcal{D}|$ is the dataset size, and $\mathbf{y}^i,\mathbf{x}_{\mathrm{ins}}^i,\mathbf{x}_{\mathrm{img}}^i$ are the predicted responses, text instructions, and image features for the $i$ -th instance $\mathcal{P}_i$ , respectively. However, as we discussed before, treating each instance equally does limited help in addressing the model bias toward overly represented SAs and SCs. Our solution to this issue is inspired by + +
Model#ParamsSocialCounterfactualsFairFaceCMSCVQAv2MMBenchTextVQA
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLaVA7B-2.05670.3973-2.87920.6457-1.61591.481778.5064.6958.21
LLaVA+POPE-0.51010.4833-1.59330.6056-2.54241.1154---
LLaVA+FT-0.47270.3625-1.01990.4865-0.71420.805878.1163.8858.20
LLaVA+CSD-0.35090.3110-0.86220.3950-0.49330.563378.1664.2058.39
LLaVA13B-2.57300.3799-3.36040.5863-1.67300.535080.067.7061.30
LLaVA+POPE-0.38400.4410-0.95080.4051-2.25421.1454---
LLaVA+FT-0.47480.4051-1.31230.4066-1.51070.460579.1767.2861.02
LLaVA+CSD-0.31130.3718-0.71140.3752-0.81670.419279.7568.3061.42
Qwen-VL7B-0.61170.5966-1.63050.8469-1.51141.096179.3774.1461.39
Qwen-VL+POPE-0.30640.5399-1.31670.9207-2.24381.7575---
Qwen-VL+FT-0.28010.3966-1.39250.6916-0.81661.022779.3774.8260.86
Qwen-VL+CSD-0.24220.2921-0.76720.2607-0.61930.718079.3775.5960.88
Bunny8B-0.42550.6064-1.13750.5349-1.58291.417382.6076.4665.31
Bunny+POPE-0.33700.5899-1.36700.4918-2.80851.7269---
Bunny+FT-0.42020.5851-1.00350.4199-0.82370.902682.4876.3965.20
Bunny+CSD-0.40010.5532-0.90030.2632-0.77450.157582.4276.2865.20
+ +Table 3. Performance comparison. All models are fine-tuned on our CMSC dataset. Among the six datasets, SocialCounterfactuals, FairFace, and CMSC are employed to evaluate social bias; and VQAv2, MMBench, and TextVQA are general benchmarks. Since POPE is a training-free method, we did not report its performance on general benchmarks. #Params: the scale of the base LLM's parameters. + +the approaches that have been proven effective in the class imbalance research area [23]. Instead of scaling loss based on the class frequency, we leverage the stereotype quantification metric Skew to rescale the loss value, + +$$ +\mathcal {E} _ {f a i r} = \frac {1}{N} \sum_ {i = 1} ^ {N} e ^ {- \operatorname {S k e w} \left(\mathcal {P} _ {i}\right)} \mathcal {L} \left(\mathbf {y} ^ {i}, \mathbf {x} _ {\text {i n s}} ^ {i}, \mathbf {x} _ {\text {i m g}} ^ {i}; \theta\right). \tag {13} +$$ + +In this scenario, when $\mathrm{Skew}(\mathcal{P}_i) > 0$ , i.e., the MLLM tends to predict the input instance as the SC label $c_{i}$ , the fairness term $e^{-\mathrm{Skew}(\mathcal{P}_i)}$ is less than 1.0. Consequently, this instance will receive less attention during training. Similarly, when $\mathrm{Skew}(\mathcal{P}_i) < 0$ , the fairness term will make the model pay more attention to this overlooked instance. This operation allows the model to dynamically adjust weights during fine-tuning based on the Skew of input instances, enabling the model to learn a fairer distribution. + +# 6. Experiments + +# 6.1. Settings + +Datasets. In addition to the three social bias datasets mentioned in Section 4.1, we also employed three recent benchmark datasets to evaluate MLLMs' original zero-shot capabilities. Among them, VQAv2 [22] and TextVQA [56] are benchmarks for general visual question answering and text-oriented visual question answering, respectively. MM-Bench [41] evaluates the model robustness with comprehensive multiple-choice answers. + +Baselines. Given the limited exploration of debiasing strategies for MLLMs, we compared our CSD with the direct fine-tuning (FT), i.e., fine-tuning the model on balanced counterfactual dataset [63]. Moreover, we adapt POPE [37], which addresses hallucinations in MLLMs, as a baseline. + +Implementation Details. Both thresholds $\tau_{1}$ and $\tau_{2}$ in Section 5.2 are set to 1.0. We trained the models on $8^{*}\mathrm{A}800$ + +GPUs. Each training session takes $\sim 18$ GPU hours. + +# 6.2. Main Results + +We reported MaxSkew@C and MinSkew@C of the four utilized MLLMs in Table 3, wherein we have three main observations: i) Compared to the baselines, our CSD demonstrates the best debiasing effectiveness across four different MLLM architectures. In intra-dataset evaluation, i.e., fine-tuning and testing on the CMSC dataset, LLaVA-13B+CSD achieved a MinSkew@C of -0.8167, improving by $46\%$ compared to LLaVA-13B+FT. In cross-dataset evaluation, our CSD also outperforms the two baselines by a significant margin. For example, Qwen-VL-7B+CSD attains a MaxSkew@C of 0.2607 on the FairFace dataset, showing a reduction of 0.6 in absolute value. ii) A model with a larger size does not necessarily correspond to a lower social bias. This might because the increased scale of parameters makes the model to better learn the biases present in the pre-training dataset. iii) POPE's debiasing effectiveness underperforms FT. One possible reason is that POPE is a training-free method, causing its predictions to remain biased towards the pre-training data distribution. + +Pertaining to the model results on general multi-modal benchmarks, we observed that our CSD method has a negligible impact, with the effect on all architectures across the three datasets being less than $0.5\%$ . This indicates that our CSD method enhances the model's social fairness without significantly sacrificing its general capabilities. + +# 6.3. Ablation Studies + +# 6.3.1. Comparison on Ablated CSD + +Our CSD method consists of two key components: the dataset resampling and the rescaled SFLoss. In Table 4, we reported the performance of them. We observe that both components help alleviate social bias. For example, + +
ModelSocialCounterfactualsFairFaceCMSC
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA-7B×-2.05670.3973-2.87920.6457-1.61591.4817
+FT-0.47270.3625-1.01990.4865-0.71420.8058
+Resample-0.44550.3327-0.92380.4838-0.69790.7741
+SFLoss-0.38510.3453-0.89040.4246-0.53850.5838
+CSD-0.35090.3110-0.86220.3950-0.49330.5633
LLAVA-13B×-2.57300.3799-3.36040.5863-1.67300.5350
+FT-0.47480.4051-1.31230.4066-1.51070.4605
+Resample-0.36030.3844-0.81540.3823-1.03640.4292
+SFLoss-0.34700.3905-1.05460.3946-1.03980.4388
+CSD-0.31130.3718-0.71140.3752-0.81670.4192
Qwen-VL-7B×-0.61170.5966-1.63050.8469-1.51141.0961
+FT-0.28010.3966-1.39250.6916-0.81661.0227
+Resample-0.24280.3631-0.80120.3207-0.74140.7813
+SFLoss-0.25800.3004-0.76880.3311-0.63860.8041
+CSD-0.24220.2921-0.76720.2607-0.61930.7180
Bunny-8B×-0.42550.6064-1.13750.5349-1.58291.4173
+FT-0.42020.5851-1.00350.4199-0.82370.9026
+Resample-0.41580.5670-0.98750.3127-0.81410.4008
+SFLoss-0.41710.5534-0.94480.2789-0.79410.4155
+CSD-0.40010.5532-0.90030.2632-0.77450.1575
+ +Table 4. The performance of ablated CSD. FT: Directly fin-tuning. Resample: Fine-tuning with resampling. SFLoss: Fin-tuning with SFLoss. The best performance is highlighted in bold. + +![](images/c8a2b166ec860b7df28de920edc7bbe7df780391fecc1e7a94a8f6736bed26f9.jpg) +Figure 5. MaxSkews of (a) Qwen-VL-7B and (b) Bunny-8B across occupations in SocialCounterfactuals for different races. + +![](images/fbeddbcd22e246b84c115a37077dd9cc486373bf0e2ece775d953f8ae62ba80c.jpg) + +LLaVA-7B+SFLoss achieves a $70\%$ MinSkew@C improvement on FairFace. On CMSC, Qwen-VL-7B+Resample achieves a MinSkew@C of -0.6193, reducing the absolute Skew value by more than 0.9. A notable observation is that either of the SFLoss and data resampling shows better debiasing effect compared to the naive FT strategy. + +# 6.3.2. Comparison on SAs + +In Figure 5, we illustrate the MaxSkews [13] of Qwen-VL-7B and Bunny-8B across different races in the SocialCounterfactuals dataset. We can observe that both models exhibit significant social bias without any fine-tuning. For instance, Qwen-VL-7B has a Skew value of approximately 0.76 for White, indicating a strong preference for predicting occupations for this race. Our CSD method achieves better debiasing effects across all races for both models. + +# 6.3.3. Skew Distributions + +In Figure 6, we illustrate the MaxSkew distribution across different SCs in CMSC. It can be observed that LLaVA-13B exhibits larger mean and median MaxSkew values, along with significant outliers in certain attributes. For instance, in the 'personality' category, the Skew value for Old-Belligerent reaches 1.6. After direct fine-tuning, the average degree of social bias decreases. However, the outlier issue remains unresolved. In contrast, applying our CSD method achieves a more optimized Skew distribution + +![](images/f5689938d15febc645e230610274cc70d8986847f665f2bc0161b08e7450c2a2.jpg) +Figure 6. MaxSkew distribution for LLaVA-13B on the CMSC dataset. Each box represents the different $\mathrm{Skew}_{a|c}$ values. Our CSD method achieves the most optimized Skew distribution. + +across different concepts. Furthermore, noticeable outlier are eliminated, indicating that CSD provides a comprehensive debiasing effect for MLLMs. + +# 7. Discussion and Conclusion + +Conclusion. In this paper, we present to address the notorious social bias problem in MLLMs. Our first contribution is a comprehensive dataset that covers more diverse social concepts than previous datasets. In addition, we advocate an counter-stereotype debiasing approach to perform both dataset resampling and loss rescaling, thereby improving fairness of MLLMs. Extensive experiments demonstrate that our method is promising to alleviate the social bias in MLLMs, with minimal negative impact on their original general multi-modal understanding capabilities. + +Social Impact. As MLLMs become increasingly integrated into real-world applications, their societal influence is poised to expand, particularly in sectors requiring sensitive handling of human attributes and social concepts. However, intrinsic social biases within these models can undermine their trustworthiness, raising ethical concerns and potential risks. This paper presents a bias mitigation approach to enhance fairness in MLLMs, fostering trust and accountability by reducing harmful stereotypes. We hope our work inspires further research on bias and fairness in MLLMs, contributing to the development of more equitable models. + +# Acknowledgment + +This research is supported in part by National Natural Science Foundation of China (Grant No. 62476071, No. U24A20328); in part by National Natural Science Foundation of China (No. 62176137). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men + +sch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 1-20, 2022. 2 +[2] Mohsan S. Alvi, Andrew Zisserman, and Christoffer Nelläker. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings. In EC-CVW, pages 556-572, 2018. 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. CoRR, abs/2308.12966:1-15, 2023. 1, 2, 4, 5 +[4] Hugo Berg, Siobhan Mackenzie Hall, Yash Bhalgat, Hannah Kirk, Aleksandar Shtedritski, and Max Bain. A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. In AACL/IJCNLP, pages 806-822, 2022. 3 +[5] Karan Bhanot, Miao Qi, John S. Erickson, Isabelle Guyon, and Kristin P. Bennett. The problem of fairness in synthetic healthcare data. Entropy, 23(9):1165, 2021. 2 +[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. CoRR, abs/2110.01963:1-18, 2021. 1 +[7] Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Tauman Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, pages 4349-4357, 2016. 2, 3 +[8] Jannik Brinkmann, Paul Swoboda, and Christian Bartelt. A multidimensional analysis of social biases in vision transformers. In ICCV, pages 4914-4923, 2023. 1, 2 +[9] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1-40, 2020. 2 +[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly-scaled multilingual language-image model. In ICLR, pages 1-13, 2023. 2 +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 2, 5 + +[12] Ching-Yao Chuang and Youssef Mroueh. Fair mixup: Fairness via interpolation. In ICLR, pages 1-11, 2021. 2 +[13] Ching-Yao Chuang, Jampani Varun, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. CoRR, abs/2302.00070:1-13, 2023. 3, 4, 8 +[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Ho. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 1-12, 2023. 2 +[15] Sepehr Dehdashtian, Lan Wang, and Vishnu Naresh Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In ICLR, pages 1-12, 2024. 3 +[16] Richard B Felson and Noah Painter-Davis. Another cost of being a young black male: Race, weaponry, and lethal outcomes in assaults. Social Science Research, 41(5):1241-1253, 2012. 3 +[17] Kathleen C. Fraser and Svetlana Kiritchenko. Examining gender and racial bias in large vision-language models using a novel dataset of parallel images. In EACL, pages 690-713, 2024. 3 +[18] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1, 2 +[19] Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. Word embeddings quantify 100 years of gender and ethnic stereotypes. PNAS, 115(16):E3635-E3644, 2018. 3 +[20] Sahin Cem Geyik, Stuart Ambler, and Krishnamaram Kenthapadi. Fairness-aware ranking in search & recommendation systems with application to linkedin talent search. In KDD, pages 2221-2231, 2019. 4 +[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6325-6334, 2017. 7 +[23] Yangyang Guo, Liqiang Nie, Harry Cheng, Zhiyong Cheng, Mohan S. Kankanhalli, and Alberto Del Bimbo. On modality bias recognition and reduction. ACM ToMM, 19(3):103:1-103:22, 2023. 7 +[24] Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu, Grace Sodunke, Aleksandar Shtedritski, and Hannah Rose Kirk. Visogender: A dataset for benchmarking gender bias in image-text pronoun resolution. In NeurIPS, pages 1-15, 2023. 1, 2, 3 +[25] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. CoRR, abs/2402.11530:1-13, 2024. 1, 2, 4, 5 +[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image + +editing with cross-attention control. In ICLR, pages 1-15, 2023. 3 +[27] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 1, 2, 3, 4 +[28] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Nils Johan Bertil Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. In NeurIPS, pages 1-22, 2023. 2 +[29] Ekaterina Iakovleva, Fabio Pizzati, Philip Torr, and Stéphane Lathuilière. Specify and edit: Overcoming ambiguity in text-based image editing. CoRR, abs/2407.20232:1-10, 2024. 3 +[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 5 +[31] Sepehr Janghorbani and Gerard De Melo. Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision-language models. In EACL, pages 1725-1735, 2023. 2 +[32] Fangkai Jiao, Zhiyang Teng, Bosheng Ding, Zhengyuan Liu, Nancy Chen, and Shafiq Joty. Exploring self-supervised logic-enhanced training for large language models. In NAACL, pages 926-941, 2024. 2 +[33] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, pages 1548-1558, 2021. 2, 3, 4 +[34] Tiep Le, Vasudev Lal, and Phillip Howard. Cocos-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. NeurIPS, pages 1-13, 2024. 1, 2 +[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2 +[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 2 +[37] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 7 +[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 2, 4, 5 +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, pages 1-13, 2024. 2 +[40] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2 + +[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? CoRR, abs/2307.06281:1-12, 2023. 7 +[42] Guangben Lu, Yuzhen Du, Zhimin Sun, Ran Yi, Yifan Qi, Yizhe Tang, Tianyi Wang, Lizhuang Ma, and Fangyuan Zou. Pinco: Position-induced consistent adapter for diffusion transformer in foreground-conditioned inpainting. arXiv preprint arXiv:2412.03812, 2024. 2 +[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding. CoRR, abs/2403.05525:1-29, 2024. 5 +[44] David Madras, Elliot Creager, Toniann Pitassi, and Richard S. Zemel. Learning adversarially fair and transferable representations. In ICML, pages 3381-3390, 2018. 3 +[45] Thomas Manzini, Yao Chong Lim, Alan W. Black, and Yulia Tsvetkov. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. In NAACL, pages 615-621, 2019. 3 +[46] Allison Master. Gender stereotypes influence children's stem motivation. Child Development Perspectives, 15(3):203-210, 2021. 3 +[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, pages 27730-27744, 2022. 5 +[48] David S Pedulla. The positive consequences of negative stereotypes: Race, sexual orientation, and the job application process. Social Psychology Quarterly, 77(1):75-94, 2014. 3 +[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952: 1-21, 2023. 3 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 5 +[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21:140:1-140:67, 2020. 2 +[52] Vikram V. Ramaswamy, Sunnie S. Y. Kim, and Olga Russakovsky. Fair attribute classification through latent space de-biasing. In CVPR, pages 9301-9310, 2021. 2 +[53] Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck, Estelle Aflalo, Shao-Yen Tseng, Vasudev Lal, and Phillip + +Howard. Debias your large multi-modal model at test-time with non-contrastive visual attribute steering. arXiv preprint arXiv:2411.12590, 2024.3 +[54] Prasanna Sattigeri, Samuel C. Hoffman, Vijil Chenthamarakshan, and Kush R. Varshney. Fairness GAN: generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3:1-3:9, 2019. 2 +[55] Ashish Seth, Mayur Hemani, and Chirag Agarwal. Dear: Debiasing vision-language models with additive residuals. In CVPR, pages 6820-6829, 2023. 2, 3, 5 +[56] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: A dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 7 +[57] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: pre-training of generic visual-linguistic representations. In ICLR, pages 1–13, 2020. 2 +[58] Yizhe Tang, Zhimin Sun, Yuzhen Du, Ran Yi, Guangben Lu, Teng Hu, Luying Li, Lizhuang Ma, and Fangyuan Zou. Ata: Adaptive transformation agent for text-guided subject-position variable background inpainting. In CVPR, pages 18335-18345, 2025. 2 +[59] Özge Tarhan. Children's understanding of the concept of social stereotypes. Child Indicators Research, 15(3):989-1023, 2022. 3 +[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971:1-16, 2023. 2, 5 +[61] Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. Are emily and greg still more employable than lakisha and jamal? investigating algorithmic hiring bias in the era of chatgpt. CoRR, abs/2310.05135:1-13, 2023. 2 +[62] W Edgar Vinacke. Stereotypes as social concepts. The journal of social Psychology, 46(2):229-243, 1957. 1 +[63] Angelina Wang and Olga Russakovsky. Overwriting pretrained bias with finetuning data. In ICCV, pages 3934-3945, 2023. 1, 7 +[64] Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy Chen. SeaEval for multilingual foundation models: From cross-lingual alignment to cultural reasoning. In NAACL, pages 370–390. Association for Computational Linguistics, 2024. 2 +[65] Jialu Wang, Yang Liu, and Xin Eric Wang. Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. In EMNLP, pages 1995-2008, 2021. 1 +[66] Mei Wang and Weihong Deng. Mitigating bias in face recognition using skewness-aware reinforcement learning. In CVPR, pages 9319-9328, 2020. 3 +[67] Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. Evaluating fairness in large vision-language models across diverse demographic attributes and prompts. CoRR, abs/2406.17974:1-11, 2024. 2 + +[68] Yisong Xiao, Aishan Liu, QianJia Cheng, Zhenfei Yin, Siyuan Liang, Jiapeng Li, Jing Shao, Xianglong Liu, and Dacheng Tao. Genderbias-vl: Benchmarking gender bias in vision language models via counterfactual probing. CoRR, abs/2407.00600:1-13, 2024. 2 +[69] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, pages 1-13, 2018. 2 +[70] Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE TPAMI, 45(12):14144–14160, 2023. 2 +[71] Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In ICML, pages 59310-59328. PMLR, 2024. 2 +[72] Jie Zhang, Sibo Wang, Xiangkui Cao, Zheng Yuan, Shiguang Shan, Xilin Chen, and Wen Gao. Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model. CoRR, abs/2406.14194:1-13, 2024. 2 +[73] Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. NeurIPS, pages 21682-21697, 2022. 3 +[74] Xiaobin Zhang, Ziqi Cui, Ning He, and Xuqun You. Are boys associated with weapons and girls associated with kitchenware? the extent to which gender stereotypes regarding adults extend to children. Acta psychologica, 230: 103754, 2022. 3 +[75] Kankan Zhou, Eason Lai, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In AACL/IJCNLP, pages 527-538, 2022. 2 \ No newline at end of file diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/images.zip b/2025/Social Debiasing for Fair Multi-modal LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..312b1c1544d71e0216fa4a90904f43d255b41e01 --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2aef17b567962b42228d203b2d4ba4580fa085ff6898441f55318597b21e27b +size 520235 diff --git a/2025/Social Debiasing for Fair Multi-modal LLMs/layout.json b/2025/Social Debiasing for Fair Multi-modal LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..78ba4333ec49376aeb39c662ab9caaf94736a0ef --- /dev/null +++ b/2025/Social Debiasing for Fair Multi-modal LLMs/layout.json @@ -0,0 +1,10736 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 165, + 103, + 444, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 103, + 444, + 121 + ], + "spans": [ + { + "bbox": [ + 165, + 103, + 444, + 121 + ], + "type": "text", + "content": "Social Debiasing for Fair Multi-modal LLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "spans": [ + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "Harry Cheng" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Yangyang Guo" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Qingpei Guo" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{3,4,*}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Ming Yang" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Tian Gan" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Weili Guan" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": ", Liqiang Nie" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "Shandong University, " + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "University of Warwick, " + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "Ant Group, " + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 142, + 548, + 185 + ], + "type": "text", + "content": "Harbin Institute of Technology (Shenzhen)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 188, + 526, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 188, + 526, + 213 + ], + "spans": [ + { + "bbox": [ + 80, + 188, + 526, + 213 + ], + "type": "text", + "content": "xaCheng1996@gmail.com, guoyang.eric@gmail.com, qingpei.gqp@antgroup.com, m.yang@antgroup.com, gantian@sdu.edu.cn, honeyguan@gmail.com, nieliqiang@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 297, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 518 + ], + "type": "text", + "content": "Multi-modal Large Language Models (MLLMs) have dramatically advanced the research field and delivered powerful vision-language understanding capabilities. However, these models often inherit deep-rooted social biases from their training data, leading to uncomfortable responses with respect to attributes such as race and gender. This paper addresses the issue of social biases in MLLMs by i) introducing a comprehensive counterfactual dataset with multiple social concepts (CMSC), which complements existing datasets by providing 18 diverse and balanced social concepts; and ii) proposing a counter-stereotype debiasing (CSD) strategy that mitigates social biases in MLLMs by leveraging the opposites of prevalent stereotypes. CSD incorporates both a novel bias-aware data sampling method and a loss rescaling method, enabling the model to effectively reduce biases. We conduct extensive experiments with four prevalent MLLM architectures. The results demonstrate the advantage of the CMSC dataset and the edge of CSD strategy in reducing social biases compared to existing competing methods, without compromising the overall performance on general multi-modal reasoning benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 542, + 135, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 135, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 135, + 554 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 635 + ], + "type": "text", + "content": "Multi-modal Large Language Models (MLLMs) have revolutionized the field of general-purpose vision-language understanding. Representative models, such as LLaVA [38], Qwen-VL [3], and Bunny [25], exhibit remarkable zero-shot performance and can be easily fine-tuned for diverse downstream applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 684 + ], + "type": "text", + "content": "Despite MLLMs' widespread use, it is imperative to recognize that these models can exhibit severe social biases with respect to attributes such as race and gender [18, 65]. Figure 1a illustrates one example that a biased MLLM is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 239, + 539, + 329 + ], + "blocks": [ + { + "bbox": [ + 328, + 239, + 539, + 329 + ], + "lines": [ + { + "bbox": [ + 328, + 239, + 539, + 329 + ], + "spans": [ + { + "bbox": [ + 328, + 239, + 539, + 329 + ], + "type": "image", + "image_path": "65d3a9cf4c9d75841914648b70a3c4a0ec34c40ef09447d883cfe1466deba406.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 397 + ], + "type": "text", + "content": "Figure 1. Social bias examples. Left: Gender-biased prediction on Nurse from LLaVA-7B1. Right: Maximum Skews across different human attributes on the FairFace dataset from Qwen-VL-7B and LLaVA-7B. As a metric for measuring model bias, a larger Skew implies a higher degree of social bias. Our CSD method outperforms existing approaches by a significant margin." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 510 + ], + "type": "text", + "content": "much more likely to associate the role of nurse with female rather than male, which may be stereotypical. Such biased predictions often occur unconsciously within MLLMs, making them difficult to detect and avoid by explicit rules. A key reason for this issue lies in the composition of the MLLM training data, which can contain content related to violence or racism [6]. These inappropriate samples may collectively demonstrate some stereotypes, resulting in subtle yet uncomfortable responses [8, 63]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Existing studies on mitigating the social bias problem in MLLMs remain largely under-explored. A naive approach is to collect attribute-balanced vision-language counterfactual datasets [24, 34], which can be used to directly finetune biased models to predict fairer distributions. However, as demonstrated in Table 1, existing large-scale datasets are limited by their focus on a single social concept, such as occupation [27], without attending to multifaceted social stereotypes [62]. This significantly hampers the model in learning more comprehensive representations. Therefore, we are motivated to construct a more extensive and diverse counterfactual dataset. From a methodological perspective, directly fine-tuning on a counterfactual dataset assigns equal importance to instances receiving different social biases, resulting in sub-optimal debiasing performance (as demonstrated in Section 6.2). Analogy to a common solution in chemistry—neutralizing acidic water requires" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 144, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 294, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 294, + 712 + ], + "type": "text", + "content": "1We manually visualize the bounding boxes from LLaVA's text output." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1740" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 288, + 147 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 288, + 147 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 288, + 147 + ], + "type": "table", + "html": "
DatasetVenueType#ImagesSocial Concepts
CoCo-Counterfactuals [34]NeurIPS'23General34KNo annotations
FairFace [33]WACV'21Social108KNo annotations
VisoGender [24]NeurIPS'23Social0.6KOccupation
PATA [55]CVPR'23Social5KOccupation
SocialCounterfactuals [27]CVPR'24Social171KOccupation
MM-Bias [31]EACL'23Social3K14 minorities
CMSC (Ours)-Social60K18 balanced concepts
", + "image_path": "fd60fc8597ff847691196f8a2cf8e7cce02d63ac56848780835ed19f3304e899.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 194 + ], + "type": "text", + "content": "Table 1. Comparison of CMSC with mainstream datasets used for model debiasing. Existing datasets are limited either by scale or by the number of social concepts they cover. Type - The category of biases the dataset focuses on." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 201, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 237 + ], + "type": "text", + "content": "adding an alkaline substance rather than plain water—we are prompted to speculate: Can we leverage the opposite of the suffered social bias to build a fairer model?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "text", + "content": "To address these two issues, we first construct a large-scale, high-quality Counterfactual dataset with Multiple Social Concepts (CMSC). As demonstrated in Table 1, our CMSC dataset narrows the gap in data scale and concept richness. We conduct extensive experiments on four different MLLM architectures [3, 25, 38] using several existing counterfactual datasets as well as our CMSC. The results demonstrate that MLLMs fine-tuned on CMSC exhibit significantly lower social bias compared to those fine-tuned on single-concept datasets. For instance, LLaVA-7B [38] fine-tuned on CMSC achieves a " + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "text", + "content": " debiasing effect, which significantly outperforms the " + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "text", + "content": " bias reduction achieved when fine-tuned on the SocialCounterfactuals dataset " + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "inline_equation", + "content": "[27]^2" + }, + { + "bbox": [ + 56, + 238, + 296, + 416 + ], + "type": "text", + "content": ". This indicates that fine-tuning the models with a rich set of social concept is beneficial in MLLM debiasing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "type": "text", + "content": "Furthermore, to implement the idea of 'debiasing with the opposite of the social biases', we propose a Counter-Stereotype Debiasing strategy (CSD) to effectively reduce social biases in MLLMs. Our CSD is equipped with two techniques: 1) we design a novel data sampling method based on the bias level, 2) we rescale the previously used autoregressive loss function to a new Social Fairness Loss (SFLoss). In this manner, previously under-represented cases, e.g., male nurses, will receive more attention, serving as new 'alkaline' ones to counteract the bias of MLLMs that prefer female nurses. The experimental results demonstrate that our CSD method is an effective debiasing strategy. As illustrated in Figure 1b, our method reduces the bias by over " + }, + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 418, + 295, + 620 + ], + "type": "text", + "content": " for the Qwen-VL-7B model compared to prior training strategies, especially naive fine-tuning. Additionally, CSD does not compromise the original model performance on general multi-modal benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 621, + 250, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 621, + 250, + 632 + ], + "spans": [ + { + "bbox": [ + 67, + 621, + 250, + 632 + ], + "type": "text", + "content": "In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 634, + 295, + 693 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 634, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 295, + 669 + ], + "type": "text", + "content": "- We construct a high-quality counterfactual dataset that includes 18 social concepts, which demonstrated to be superior to existing ones on MLLM social debiasing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 670, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 693 + ], + "type": "text", + "content": "- We propose a novel approach that applying counter-stereotype debiasing to mitigate social biases. To the best" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 107 + ], + "type": "text", + "content": "of our knowledge, this is the first research effort dedicated to addressing the social bias problem in autoregressive MLLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We apply our debiasing strategy to four multi-modal LLMs architectures to show the generalization capability and effectiveness of our method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 156, + 400, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 400, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 400, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 508, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 508, + 190 + ], + "type": "text", + "content": "2.1. Multi-modal Large Language Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 459 + ], + "type": "text", + "content": "With the rapid development of LLMs [9, 51, 57, 60, 64], increasing efforts have been dedicated to extending the powerful reasoning capabilities of LLMs to multi-modal applications [1, 10, 14, 28, 32, 35, 40, 42, 58, 70, 71]. Specifically, MLLMs utilize LLMs as the foundational base, aligning features from other modalities with text embeddings to enable LLMs to perceive multi-modal inputs. For instance, BLIP-2 [36] utilizes a Q-Former to align vision encoders with LLMs. LLaVA [38, 39] presents to directly map vision features to the word embedding space of Vicuna [11]. Qwen-VL [3] employs a single-layer cross-attention module as the vision-language adapter, and introduces multitask pre-training to improve the model performance. These approaches significantly enhance multi-modal understanding capabilities. However, some studies highlight the existence of social biases in pre-trained LLMs [7, 8, 18]. These biases often manifest as harmful spurious correlations related to human attributes such as gender and race, which significantly undermine the fairness. For instance, LLMs may 'efficiently' filter resumes based on race rather than the candidates' qualifications [61]. Nevertheless, research on social biases in MLLMs remains largely unexplored [68]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 468, + 440, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 440, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 440, + 479 + ], + "type": "text", + "content": "2.2. Social Bias Reduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 510 + ], + "type": "text", + "content": "We roughly categorize the bias mitigation strategies into two groups: data-based and objective-based." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Data-based debiasing typically refers to data augmentation techniques. For instance, Chuang et al. [12] employ mixup [69] to construct interpolated samples among groups with different distributions. Ramaswamy et al. [52] utilize perturbed GAN-generated images [21] in latent space to augment the original dataset. Moreover, some studies collect fair datasets across human attributes for continual finetuning [5, 24, 31, 34, 54, 72, 75]. In particular, FairFace [33] contains " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "108\\mathrm{K}" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " images that are balanced for the race attribute. Socialcounterfactuals [27] collects " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "171\\mathrm{K}" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " imagetext pairs to probe biases across race, gender, and physical characteristics. However, these collected counterfactual datasets often focus on only one common concept - occupation, which hinders effective debiasing [67]. Though a few datasets involve several concepts, they nevertheless, are largely limited by their data scale (e.g., 3K [31]), making it less feasible to fine-tune MLLMs." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 286, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 286, + 713 + ], + "type": "text", + "content": "2In this paper, we interchangeably use debiasing and bias reduction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1741" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 71, + 504, + 164 + ], + "blocks": [ + { + "bbox": [ + 106, + 71, + 504, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 71, + 504, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 71, + 504, + 164 + ], + "type": "image", + "image_path": "37ca1d70d0659a45d23a2014baf3a63c47507e136d54b8357619328d3c7eb611.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 555, + 187 + ], + "type": "text", + "content": "Figure 2. Image Generation Pipeline. Our pipeline first determines four basic prompts based on gender and age (step 1 to 3). Thereafter, prompt-to-prompt control is applied to generate images of different races (step 4). Finally, we filter out low-quality images (step 5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 296, + 434 + ], + "type": "text", + "content": "Objective-based debiasing modifies the model training process to achieve better fairness. Early studies are proposed to reduce social bias within uni-modal models [2, 17, 19, 44, 45, 66]. For example, Bolukbasi et al. [7] optimizes word embeddings to remove gender stereotypes. In contrast, multi-modal debiasing methods often focus on contrastive learning-based vision language models, e.g., CLIP [50]. Representative approaches are to rectify the modality similarity matrix in VLMs [15] via learning additive adapters [55, 73], eliminating biased directions [13], and using adversarial samples [4]. However, among these methods, uni-modal approaches often require re-training the full model. This operation, though plausible for previous small-scale models, is less practical for existing MLLMs with billions of parameters [53]. On the other hand, mainstream multi-modal techniques cannot be utilized for these models due to the divergent training objectives between MLLMs and previous CLIP-style models. In particular, MLLMs are mostly trained in an autoregressive way rather than through contrastive learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 218, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 218, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 218, + 456 + ], + "type": "text", + "content": "3. CMSC Dataset Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 538 + ], + "type": "text", + "content": "As demonstrated in Table 1, existing counterfactual datasets are either restricted by their small scale [24] or a narrow coverage of concepts [33]. To bridge this gap between limited existing datasets and diverse real-world stereotypes, we introduce the CMSC dataset, encompassing 60k high-quality images across eighteen social concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 224, + 559 + ], + "type": "text", + "content": "3.1. Social Attributes and Concepts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 684 + ], + "type": "text", + "content": "A counterfactual dataset for social bias reduction and evaluation typically contains two key aspects: Social Attribute (SA) and Social Concept (SC). The former, i.e., SA, is defined as characteristics shared by a group of people [33]. Specifically, we investigate three types of attributes: i) genders: {Male, Female}; ii) races: {White, Black, Indian, Southeast Asian, East Asian, Middle Eastern, Latino}; and iii) ages: {Young, Old}. These attributes are intrinsic to individuals3. For each image " + }, + { + "bbox": [ + 55, + 564, + 295, + 684 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 564, + 295, + 684 + ], + "type": "text", + "content": " in CMSC, three SA labels are provided with respect to these three types of SA, and are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 194, + 419, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 419, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 419, + 205 + ], + "type": "text", + "content": "combined as a SA set " + }, + { + "bbox": [ + 313, + 194, + 419, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 313, + 194, + 419, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "text", + "content": "As for the SC, we define it as the societal label attributed to an individual. Drawing from the sociological research [16, 46, 48, 59, 74], we employ 18 SCs in CMSC. In particular, these SCs are categorized into three groups: personality, responsibility, and education. Specifically, personality relates to concepts pertaining to an individual's character. We use five concepts: {compassionate, belligerent, authority, pleasant, unpleasant}. Responsibility is the roles or duties that individuals are expected to fulfill in society or family. We identify six concepts: {tool user, weapon user, career, family, chef working, earning money}. Education pertains to the level of education a person has received. We include seven concepts: {middle school, high school, university, good student, bad student, science, arts}. Each image in CMSC is annotated with one SC label " + }, + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "inline_equation", + "content": "c_i \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 206, + 555, + 397 + ], + "type": "text", + "content": " is the union of the concepts from all above groups." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "content": "In summary, each instance " + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "content": " in our dataset consists of an image " + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "content": ", a set of SA labels " + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_i" + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "content": ", and a SC label " + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 398, + 555, + 445 + ], + "type": "text", + "content": ". To the best of our knowledge, CMSC is the first large-scale counterfactual dataset that includes a variety of social concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 463, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 463, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 463, + 468 + ], + "type": "text", + "content": "3.2. Image Generation Pipeline" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "We employ Stable Diffusion XL (SDXL) [49] to generate instances in CMSC. For each SC, we carefully design prompts that provide a detailed explanation of the concept. These prompts not only help eliminate potential ambiguities associated with the SC but also enhance the quality of the generated images [29]. For instance, for the SC 'pleasant,' we use the prompt: 'A photo of a pleasant [race] [gender] person at [age], who has a friendly smile, a relaxed posture, and ...' To ensure diverse and representative image generation, these prompts are randomly modified by replacing words with synonyms during image generation. Subsequently, the intersectional generation strategy [27] is applied. As illustrated in Figure 2, we first determine gender and age (Step 2) and generate four base images (Step 3). Thereafter, we apply Prompt-to-Prompt (P2P) control [26] to produce visually similar images that vary only in terms of race (Step 4). For each SC, we create 28 prompts across attribute combinations. Each prompt helps generate 100 images. To ensure image quality, we apply a rigorous image filtering mechanism (step 5). First, we use CLIP-Score [50]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "3All SAs are perceived, made by human annotators or models. We acknowledge that the SAs are not representative of all people [27, 33]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1742" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "to filter out images that are semantically misaligned with the prompts. Subsequently, a panel of three experts conducts a voting process to remove low-quality and NSFW content. If the generated images perform poorly on a specific SA, the entire set of images (including those of other SAs) will be discarded. Moreover, we observe that the model can generate high-quality images for certain SCs or SAs, while struggling to do so for others. As a result, we filter out excessively generated ones for balanced distribution. Our filtering process is very strict. For instance, the testing set initially contains over 50,000 images, but only 12,019 images are finally used in CMSC. A similar process is applied to the training set, which comprises 48,134 images. In total, our dataset contains 60K images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 206, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 206, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 206, + 264 + ], + "type": "text", + "content": "4. CMSC Dataset Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 174, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 174, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 174, + 285 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "type": "text", + "content": "Evaluation Protocols. We employed four different MLLM architectures, i.e., LLaVA-7B [38], LLaVA-13B [38], Qwen-VL-7B [3], and Bunny-8B [25], to evaluate our CMSC dataset. For each MLLM, we used the same template to collect their responses: 'Based on the race, gender, and age of the face , do you think the person is {SCs}? Respond with one of them.' Here, is the image from the dataset, and {SCs} represents the list of used SCs. For instance, when testing personality, {SCs} includes 'pleasant,' 'unpleasant,' etc. After collecting these predictions, we used them to compute MaxSkew@C and MinSkew@C (which will be elaborated in the following section) to measure bias." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": "Metrics. We employ Skew-based [20] metrics to quantify the extent of social biases. Specifically, for a given dataset " + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": ", we define the subset containing instances with a specific SC label " + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{c'}" + }, + { + "bbox": [ + 55, + 445, + 296, + 493 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 503, + 295, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 503, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 115, + 503, + 295, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D}, c _ {i} = c ^ {\\prime} \\right\\}. \\tag {1}", + "image_path": "a0f7ffd1a868670f3582dd36dd194d5c97749f5b3e421c0fb659c82de5172ced.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "text", + "content": "From " + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{c'}" + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "text", + "content": ", we further extract the subset consisting of instances with a specific SA label " + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "inline_equation", + "content": "a'" + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{a'|c'}" + }, + { + "bbox": [ + 55, + 527, + 296, + 552 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 561, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 561, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 295, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\left\\{\\mathcal {P} _ {i} \\mid \\mathcal {P} _ {i} \\in \\mathcal {D} _ {c ^ {\\prime}}, a ^ {\\prime} \\in \\mathcal {A} _ {i} \\right\\}. \\tag {2}", + "image_path": "9a408b6d5280a178b1bcc737f27a382e03edf2f640ef2d2c8ec1ed38ab76b08c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": "We then utilize the aforementioned template to guide MLLMs in predicting the SC label corresponding to each " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ". We denote the predicted SC label as " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ". By applying this process to the entire dataset " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ", we can construct a new predicted set " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}} = \\{\\hat{\\mathcal{P}}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_i = \\{\\mathbf{I}_i, \\mathcal{A}_i, \\hat{c}_i\\}" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "N = |\\mathcal{D}|" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": ", subsets " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}_{c'}" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}_{a'|c'}" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": " can be derived with Equation (1) and Equation (2), respectively. The Skew for SC " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": " and SA " + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "inline_equation", + "content": "a'" + }, + { + "bbox": [ + 55, + 585, + 296, + 681 + ], + "type": "text", + "content": " could be formulated as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 689, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 689, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 124, + 689, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\operatorname {S k e w} _ {a ^ {\\prime} \\mid c ^ {\\prime}} = \\log \\left(\\frac {\\hat {\\gamma} _ {a ^ {\\prime}} \\mid c ^ {\\prime}}{\\gamma_ {a ^ {\\prime}} \\mid c ^ {\\prime}}\\right), \\tag {3}", + "image_path": "8a63a55a27c4a42b520c4cb800635097a5e7c4ff3382f6c444223998a24fbc16.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 333, + 71, + 535, + 212 + ], + "blocks": [ + { + "bbox": [ + 333, + 71, + 535, + 212 + ], + "lines": [ + { + "bbox": [ + 333, + 71, + 535, + 212 + ], + "spans": [ + { + "bbox": [ + 333, + 71, + 535, + 212 + ], + "type": "image", + "image_path": "f68790acb97a2aa051ef1aec18b1d1c6c30527132566eb54a8f59f72122c616d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 256 + ], + "type": "text", + "content": "Figure 3. Performance of MLLMs when evaluated on FairFace. In each model, the bars labeled 'FT w SCounter' and 'FT w CMSC' represent the model fine-tuned on SocialCounterfactuals and our CMSC, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 269, + 342, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 269, + 342, + 278 + ], + "spans": [ + { + "bbox": [ + 314, + 269, + 342, + 278 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 391, + 277, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 277, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 391, + 277, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {\\gamma} _ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\hat {\\mathcal {D}} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\hat {\\mathcal {D}} _ {c ^ {\\prime}} \\right|}, \\\\ \\gamma_ {a ^ {\\prime} | c ^ {\\prime}} = \\frac {\\left| \\mathcal {D} _ {a ^ {\\prime} | c ^ {\\prime}} \\right|}{\\left| \\mathcal {D} _ {c ^ {\\prime}} \\right|}. \\end{array} \\right. \\tag {4}", + "image_path": "c33f51fc0173323c62c589b2399fca8377c59374fff833ab5331379520f10ce6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{a'|c'} > 0" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": ", the MLLM tends to predict instances with attribute " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "a'" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": " as concept " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": ". For instance, in Figure 1a, " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{\\text{Female|Nurse}} > 0" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": ". In contrast, when " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{a'|c'} < 0" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": ", the MLLM is inclined to not predict these instances containing SA " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "a'" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": " as SC " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": ", e.g., predicting male as nurse. A fair MLLM should have Skew close to 0 across all concepts and attributes, i.e., " + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\sum_{a,c} |\\mathrm{Skew}_{a|c}| \\to 0" + }, + { + "bbox": [ + 313, + 321, + 554, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 406, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 501 + ], + "type": "text", + "content": "Although Skew can effectively measure MLLMs' bias towards a particular SA-SC pair, a counterfactual dataset often contains hundreds of such combinations, e.g., our CMSC includes 198 SA-SC pairs. This confounds the comprehensive assessment of MLLM bias. Therefore, we propose two variants of Skew to enable a more thorough analysis. Specifically, we first identify the maximum and minimum Skew values for each SC across all SAs [13]," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 510, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 510, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 348, + 510, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\operatorname {M a x S k e w} _ {c ^ {\\prime}} = \\operatorname {M a x} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}, \\\\ \\operatorname {M i n S k e w} _ {c ^ {\\prime}} = \\operatorname {M i n} _ {a _ {i} \\in \\mathcal {A}} \\left\\{\\operatorname {S k e w} _ {a _ {i} \\mid c ^ {\\prime}} \\right\\}. \\end{array} \\right. \\tag {5}", + "image_path": "3d05173d2cbade2ef84c02ba90ec46cc312603e648ec7601e5602c8a567c63dc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": "For all " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "c' \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": ", we separately calculate the average of " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{MaxSkew}_{c'}" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathrm{MinSkew}_{c'}" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": ", resulting in two aggregated Skew values. We refer to these two values as MaxSkew@C (MaxS@C) and MinSkew@C (MinS@C), which represent the overall bias level of the MLLM. Both of these metrics indicate better fairness as they approach zero." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Datasets. We compared our dataset with the previous state-of-the-art counterfactual dataset. Specifically, we separately fine-tuned MLLMs on CMSC and SocialCounterfactual [27] to evaluate their effectiveness in mitigating bias. The SocialCounterfactual dataset focuses on the SC of occupation and contains 171K high-quality synthetic images. Moreover, we employed the FairFace dataset [33], which contains 108K images with a balanced distribution of SAs," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1743" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 284, + 157 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 284, + 157 + ], + "type": "table", + "html": "
FT w/ SCPer.Res.Edu.
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA×-0.94862.4950-0.76622.2188-0.85693.9821
+ FT Per.-0.84691.6675-0.59152.2025-0.60943.1794
+ FT Res.-0.91521.7409-0.49211.4397-0.81582.7705
+ FT Edu.-0.97561.7842-0.68721.5719-0.52572.4362
Qwen.VL×-2.46882.5772-2.53012.2663-2.82422.4059
+ FT Per.-0.73861.2509-1.90770.3470-2.55242.3506
+ FT Res.-0.73841.9044-1.32050.1202-2.85452.3872
+ FT Edu.-0.84311.6817-1.87590.2236-2.47062.3502
", + "image_path": "90700cb2f6e6b41caa8f0d988170e5d3322848f98c8560bccd1c2502495be328.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "type": "text", + "content": "Table 2. Performance comparison of LLaVA-7B and Qwen-VL-7B when fine-tuned and tested on SCs in CMSC. Per.: Personality, Res.: Responsibility, Edu.: Education, " + }, + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 157, + 295, + 191 + ], + "type": "text", + "content": ": Original performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 239 + ], + "type": "text", + "content": "as an additional testing set to assess the cross-dataset debiasing performance of models fine-tuned on CMSC and SocialCounterfactual [55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 209, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 209, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 209, + 264 + ], + "type": "text", + "content": "4.2. Comparison on Fine-tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 413 + ], + "type": "text", + "content": "We reported MinSkew@C and MaxSkew@C of several MLLMs in Figure 3. The models, i.e., LLaVA [38], Qwen-VL [3], and Bunny [25], are fine-tuned on the SocialCounterfactuals and CMSC datasets individually and evaluated on the FairFace dataset, respectively. For clarity, we took the absolute values of both metrics. From Figure 3, we observed that models fine-tuned on CMSC exhibit superior debiasing effects. For instance, LLaVA-13B achieves a MinSkew@C of -1.3132, a significant advantage over that from the model fine-tuned on SocialCounterfactuals. This implies that the broader range of social concepts in CMSC enables the model to learn fairer distributions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 246, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 246, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 246, + 438 + ], + "type": "text", + "content": "4.3. Fine-tuning with Specific SC Group" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 576 + ], + "type": "text", + "content": "As discussed in Section 3.1, we split the SCs in CMSC into three groups: personality, responsibility, and education. We applied LLaVA and Qwen-VL to fine-tune separately on these subsets. The results are reported in Table 2. It can be observed that in intra-subset evaluations, i.e., fine-tuned and tested on the same subset, the model can generally achieve lower biases. For instance, when fine-tuning and evaluating Qwen-VL on the personality subset, it achieves a MaxS@C of 1.2509, showing an absolute difference of 0.65 and 0.43 compared to performance when fine-tuned on the responsibility and education subsets, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 242, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 242, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 242, + 600 + ], + "type": "text", + "content": "4.4. Comparison on Image Distribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "We calculated the Fréchet Inception Distance (FID) scores for the synthetic counterfactual datasets SocialCounterfactuals and our CMSC. Specifically, we randomly sampled 1,000 synthetic images from each dataset and then computed the distributional differences with the same set of 1,000 real images. A lower FID signifies better image quality. With this process, the two datasets received FID scores of 27.17 and 24.35, respectively. This indicates that the images in our dataset bear a closer resemblance to reality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 482, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 482, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 482, + 85 + ], + "type": "text", + "content": "5. Counter-stereotype Debiasing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 400, + 102 + ], + "type": "text", + "content": "5.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "text", + "content": "Before introducing our proposed method, we first revisit the training objectives of MLLMs. Subsequently, we introduce the concept of " + }, + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "text", + "content": " to measure stereotypes associated with specific instances, upon which we base the reformulation of the autoregressive fine-tuning paradigm of MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 174, + 470, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 470, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 470, + 186 + ], + "type": "text", + "content": "5.1.1. Traning Objective of MLLMs" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 189, + 554, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 554, + 225 + ], + "type": "text", + "content": "Current mainstream MLLMs employ pre-trained LLMs [11, 60] as the output interface [3, 25, 38, 43]. Under this context, the base LLM is often trained in an autoregressive way," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 232, + 554, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 232, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 367, + 232, + 554, + 266 + ], + "type": "interline_equation", + "content": "\\mathrm {P} \\left(\\mathbf {x} _ {1: T}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {< t}; \\theta\\right), \\tag {6}", + "image_path": "dbe3989819ea49a2a790c85a7ca25b71094875e7af45500a8dd7fb0bf87bf503.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": " is a text sequence with " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": " tokens, and " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": " denotes the model parameters. During pre-training, the LLM predicts the " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": "-th token " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": " based on all preceding ones, i.e., " + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{< t}" + }, + { + "bbox": [ + 313, + 273, + 554, + 333 + ], + "type": "text", + "content": ". After that, another instruction tuning stage [47] is often followed to enable LLMs to better understand user intentions," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 340, + 554, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 340, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 349, + 340, + 554, + 373 + ], + "type": "interline_equation", + "content": "\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}; \\theta\\right), \\tag {7}", + "image_path": "cfc4583174c65f9aac6aae243a9dd8ffafc255befa7d3d6e17aea38f23da2a5a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{ins}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "text", + "content": " are textual instructions and responses, respectively. MLLMs extend the inputs of LLMs with the supplement of image features " + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 380, + 554, + 430 + ], + "type": "text", + "content": " that are extracted using pre-trained vision encoders such as ViT [30, 50]," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 437, + 554, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 437, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 321, + 437, + 554, + 469 + ], + "type": "interline_equation", + "content": "\\mathrm {P} \\left(\\mathbf {y} _ {1: T} \\mid \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right) = \\prod_ {t = 1} ^ {T} \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right), \\tag {8}", + "image_path": "078f55db32c750c154b787f452cbfbaee936049af1efbd3c9980f0a44c6feac1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 477, + 554, + 514 + ], + "type": "text", + "content": " is aligned with text features via a connector, e.g., a trainable projection matrix. To optimize the pipeline, previous cross-entropy loss from LLMs is directly inherited," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 521, + 554, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 521, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 321, + 521, + 554, + 554 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {y}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta) = - \\sum_ {t = 1} ^ {T} \\log \\mathrm {P} \\left(\\mathbf {y} _ {t} \\mid \\mathbf {y} _ {< t}, \\mathbf {x} _ {\\text {i n s}}, \\mathbf {x} _ {\\text {i m g}}; \\theta\\right). \\tag {9}", + "image_path": "024c0b2353e3717a5e1bf938f8752715a17d49d77ab315b5ce55c1ee2a5b1964.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 567, + 449, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 449, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 449, + 578 + ], + "type": "text", + "content": "5.1.2. Stereotype Measurement" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Recall that " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{a|c}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " quantifies the degree of social biases in an MLLM across the entire dataset. In this section, we are more interested in the social bias degree for each specific instance " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". We then define " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " as a selected " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{a|c}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " with the maximum absolute value across all SAs for instance " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". For example, suppose that " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " contains three SA labels: {'White,' Female,' 'Young'}, along with one SC label: 'Nurse.' " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " is the one with the highest absolute value from " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{Skew}_{White|Nurse}, \\mathrm{Skew}_{Female|Nurse}, \\mathrm{Skew}_{Young|Nurse}\\}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". It describes the SA that receives the most severe bias, thereby indicating the level of bias in " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1744" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 274, + 175 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 274, + 175 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 274, + 175 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 274, + 175 + ], + "type": "image", + "image_path": "ad79bf07f31990abf72184d663161dfa8b9fbb59190a858fbacfa135387412c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 296, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 209 + ], + "type": "text", + "content": "Figure 4. Comparison between fine-tuning (upper part) and our CSD (lower part). CSD adjusts the loss and datasets based on " + }, + { + "bbox": [ + 55, + 175, + 296, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 55, + 175, + 296, + 209 + ], + "type": "text", + "content": ", guiding the model to focus on overlooked instances." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 209, + 259, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 259, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 259, + 220 + ], + "type": "text", + "content": ": Parameter updates based on a specific loss function." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 232, + 149, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 149, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 149, + 245 + ], + "type": "text", + "content": "5.2. CSD Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 249, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 392 + ], + "type": "text", + "content": "As illustrated in Figure 4, vanilla fine-tuning approach involves sampling images from a balanced dataset and updating MLLMs through the original autoregressive objective. We argue that this method, which treats all instances equally, is ineffective in addressing the social bias problem in MLLMs. Therefore, we propose an CSD method from the view of counter-stereotype. Specifically, our CSD is composed of two components: i) Dataset Resampling, which enhances the data sampling process to include more underrepresented instances. ii) Loss Rescaling, where we adjust the loss function to place larger emphasis on instances that are overlooked in terms of social attributes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 394, + 296, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 394, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 296, + 490 + ], + "type": "text", + "content": "Dataset Resampling. The debiasing approach, e.g., directly fine-tuning, utilizes counterfactual datasets that are balanced across all SAs. We believe that such datasets make it difficult for MLLMs to recognize which SAs are subjected to greater bias and which are not4. To address this, we resample the dataset to increase the frequency of instances that MLLMs tend to ignore and reduce the frequency of emphasized ones, thereby 'neutralizing' social biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " serves as an indicator for dataset resampling. As demonstrated in Algorithm 1, for each instance " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": ", when " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i) > 0" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " has received more attention – such as the 'female-nurse' in Figure 1 – we reduce its probability in the resampled dataset " + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 490, + 296, + 550 + ], + "type": "text", + "content": " for the following training epoch," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 559, + 296, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 559, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 59, + 559, + 296, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {r} = \\left\\{ \\begin{array}{l} \\mathcal {D} _ {r} \\cup \\mathcal {P} _ {i}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) > \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\\\ \\mathcal {D} _ {r}, \\operatorname {r a n d} (0, \\operatorname {S k e w} (\\mathcal {P} _ {i}) + \\tau_ {1}) \\leq \\operatorname {S k e w} (\\mathcal {P} _ {i}), \\end{array} \\right. \\tag {10}", + "image_path": "45cea0379cd78861b12eab50637c8aaea40022d40599a5e02714bafd90711df4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\operatorname{rand}(\\cdot)" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": " is to randomly draw from 0 to " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i) + \\tau_1" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": " is a pre-defined threshold. As such, a larger " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": " corresponds to a lower chance of being included by " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": ". In contrast, for instances with " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i) \\leq 0" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": ", we believe that increasing their proportion in the new dataset " + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": " is beneficial for the MLLM to learn the features of these overlooked parts, thereby achieving a fairer distribution." + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 320, + 125, + 555, + 319 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 515, + 85 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 515, + 85 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 515, + 85 + ], + "type": "text", + "content": "Algorithm 1 Dataset Resampling with Skew " + }, + { + "bbox": [ + 315, + 72, + 515, + 85 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_i)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "lines": [ + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "spans": [ + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "text", + "content": "Input: The original dataset " + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathbf{I}_i, \\mathcal{A}_i, c_i\\}" + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "text", + "content": ", the social attribute sets " + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "text", + "content": ", the social concept set " + }, + { + "bbox": [ + 313, + 88, + 555, + 113 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 114, + 460, + 125 + ], + "lines": [ + { + "bbox": [ + 315, + 114, + 460, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 114, + 460, + 125 + ], + "type": "text", + "content": "Output: The resampled dataset " + }, + { + "bbox": [ + 315, + 114, + 460, + 125 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\gets []" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " ,AcmSkew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "a_{i}|c_{i}\\gets 0" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " ; Initialization \n2: for " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i\\in \\mathcal{D}" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " do \n3: if Skew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_i) > 0" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " then \n4: if Rand(0,Skew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_i) + \\tau_1) > \\mathrm{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " then \n5: " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " \n6: end if \n7: else " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " Skew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_i)\\leq 0" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " \n9: AcmSkew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "a_{i}|c_{i}\\gets" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " AcmSkew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "a_{i}|c_{i} + |\\mathrm{Skew}(\\mathcal{P}_i)|;" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " \n10: if AcmSkew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "a_{i}|c_{i} > \\tau_{2}" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " then \n11: " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r\\gets \\mathcal{D}_r\\cup \\mathcal{P}_i;" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " Over-resampling \n12: AcmSkew " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "a_{i}|c_{i}\\gets 0" + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "text", + "content": " \n13: end if \n14: end if \n15: end for \n16: return " + }, + { + "bbox": [ + 320, + 125, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "text", + "content": "To this end, these instances are directly accepted into " + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "text", + "content": ". Moreover, we design an over-resampling mechanism to further increase the occurrence frequency of these instances. Specifically, for each " + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "text", + "content": ", we employ an accumulative value, AcmSkew, to gradually accumulate the current " + }, + { + "bbox": [ + 313, + 340, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 328, + 410, + 555, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 410, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 328, + 410, + 555, + 425 + ], + "type": "interline_equation", + "content": "\\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} = \\operatorname {A c m S k e w} _ {a _ {i} \\mid c _ {i}} + | \\operatorname {S k e w} (\\mathcal {P} _ {i}) |, \\tag {11}", + "image_path": "66a39c790e6dddce9120cc6f68de46f5ed5d81f3f4c7f1bc1cf7a5f2ab9a97f9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " are SC label and SA label corresponding to " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i)" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": ", respectively. When " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\operatorname{AcmSkew}_{a_i|c_i}" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " exceeds a threshold " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\tau_2" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": ", we add the " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " again. The " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\operatorname{AcmSkew}_{a_i|c_i}" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " can then be set to 0 for a new round of accumulation. With the above operations, instances with " + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\operatorname{Skew}(\\mathcal{P}_i) \\leq 0" + }, + { + "bbox": [ + 313, + 434, + 555, + 540 + ], + "type": "text", + "content": " will be resampled multiple times. This resampling process is executed once before each training epoch. For evaluation, we employ the balanced testing set to ensure a fair comparison." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 541, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 590 + ], + "type": "text", + "content": "Loss Rescaling. During MLLM fine-tuning, we rescale the autoregressive loss in Equation (9) to a new Social Fairness Loss (SFLoss) for effective debiasing. Specifically, the empirical risk during training can be represented as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 598, + 555, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 598, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 367, + 598, + 555, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {E} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\mathrm {i n s}} ^ {i}, \\mathbf {x} _ {\\mathrm {i m g}} ^ {i}; \\theta\\right), \\tag {12}", + "image_path": "54737d3ba2eec1f72523bf4d3e6d7730f54376c3d357d684ce06b5644a73bcb2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "N = |\\mathcal{D}|" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " is the dataset size, and " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^i,\\mathbf{x}_{\\mathrm{ins}}^i,\\mathbf{x}_{\\mathrm{img}}^i" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " are the predicted responses, text instructions, and image features for the " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "-th instance " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", respectively. However, as we discussed before, treating each instance equally does limited help in addressing the model bias toward overly represented SAs and SCs. Our solution to this issue is inspired by" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "4Collecting an imbalanced datasets is impractical, as different MLLMs may exhibit different biases." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1745" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 70, + 492, + 235 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 492, + 235 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 492, + 235 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 492, + 235 + ], + "type": "table", + "html": "
Model#ParamsSocialCounterfactualsFairFaceCMSCVQAv2MMBenchTextVQA
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLaVA7B-2.05670.3973-2.87920.6457-1.61591.481778.5064.6958.21
LLaVA+POPE-0.51010.4833-1.59330.6056-2.54241.1154---
LLaVA+FT-0.47270.3625-1.01990.4865-0.71420.805878.1163.8858.20
LLaVA+CSD-0.35090.3110-0.86220.3950-0.49330.563378.1664.2058.39
LLaVA13B-2.57300.3799-3.36040.5863-1.67300.535080.067.7061.30
LLaVA+POPE-0.38400.4410-0.95080.4051-2.25421.1454---
LLaVA+FT-0.47480.4051-1.31230.4066-1.51070.460579.1767.2861.02
LLaVA+CSD-0.31130.3718-0.71140.3752-0.81670.419279.7568.3061.42
Qwen-VL7B-0.61170.5966-1.63050.8469-1.51141.096179.3774.1461.39
Qwen-VL+POPE-0.30640.5399-1.31670.9207-2.24381.7575---
Qwen-VL+FT-0.28010.3966-1.39250.6916-0.81661.022779.3774.8260.86
Qwen-VL+CSD-0.24220.2921-0.76720.2607-0.61930.718079.3775.5960.88
Bunny8B-0.42550.6064-1.13750.5349-1.58291.417382.6076.4665.31
Bunny+POPE-0.33700.5899-1.36700.4918-2.80851.7269---
Bunny+FT-0.42020.5851-1.00350.4199-0.82370.902682.4876.3965.20
Bunny+CSD-0.40010.5532-0.90030.2632-0.77450.157582.4276.2865.20
", + "image_path": "effe5dcc9f3fc6d169433bf230417c74963797b54ac08ec431a4a2ca50fc38ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 274 + ], + "type": "text", + "content": "Table 3. Performance comparison. All models are fine-tuned on our CMSC dataset. Among the six datasets, SocialCounterfactuals, FairFace, and CMSC are employed to evaluate social bias; and VQAv2, MMBench, and TextVQA are general benchmarks. Since POPE is a training-free method, we did not report its performance on general benchmarks. #Params: the scale of the base LLM's parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 281, + 295, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 329 + ], + "type": "text", + "content": "the approaches that have been proven effective in the class imbalance research area [23]. Instead of scaling loss based on the class frequency, we leverage the stereotype quantification metric Skew to rescale the loss value," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 337, + 295, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 337, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 70, + 337, + 295, + 370 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {f a i r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} e ^ {- \\operatorname {S k e w} \\left(\\mathcal {P} _ {i}\\right)} \\mathcal {L} \\left(\\mathbf {y} ^ {i}, \\mathbf {x} _ {\\text {i n s}} ^ {i}, \\mathbf {x} _ {\\text {i m g}} ^ {i}; \\theta\\right). \\tag {13}", + "image_path": "9ea3f90c12c73f9d3d5b19fb5139c6549469b5198a939a565afc89f48d8665b2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "content": "In this scenario, when " + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i) > 0" + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "content": ", i.e., the MLLM tends to predict the input instance as the SC label " + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "content": ", the fairness term " + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "inline_equation", + "content": "e^{-\\mathrm{Skew}(\\mathcal{P}_i)}" + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "content": " is less than 1.0. Consequently, this instance will receive less attention during training. Similarly, when " + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}(\\mathcal{P}_i) < 0" + }, + { + "bbox": [ + 55, + 379, + 296, + 485 + ], + "type": "text", + "content": ", the fairness term will make the model pay more attention to this overlooked instance. This operation allows the model to dynamically adjust weights during fine-tuning based on the Skew of input instances, enabling the model to learn a fairer distribution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 136, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 136, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 136, + 510 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 116, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 116, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 116, + 529 + ], + "type": "text", + "content": "6.1. Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 629 + ], + "type": "text", + "content": "Datasets. In addition to the three social bias datasets mentioned in Section 4.1, we also employed three recent benchmark datasets to evaluate MLLMs' original zero-shot capabilities. Among them, VQAv2 [22] and TextVQA [56] are benchmarks for general visual question answering and text-oriented visual question answering, respectively. MM-Bench [41] evaluates the model robustness with comprehensive multiple-choice answers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "content": "Baselines. Given the limited exploration of debiasing strategies for MLLMs, we compared our CSD with the direct fine-tuning (FT), i.e., fine-tuning the model on balanced counterfactual dataset [63]. Moreover, we adapt POPE [37], which addresses hallucinations in MLLMs, as a baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. Both thresholds " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " in Section 5.2 are set to 1.0. We trained the models on " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "8^{*}\\mathrm{A}800" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 523, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 523, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 523, + 293 + ], + "type": "text", + "content": "GPUs. Each training session takes " + }, + { + "bbox": [ + 313, + 281, + 523, + 293 + ], + "type": "inline_equation", + "content": "\\sim 18" + }, + { + "bbox": [ + 313, + 281, + 523, + 293 + ], + "type": "text", + "content": " GPU hours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 399, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 399, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 399, + 315 + ], + "type": "text", + "content": "6.2. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 321, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 548 + ], + "type": "text", + "content": "We reported MaxSkew@C and MinSkew@C of the four utilized MLLMs in Table 3, wherein we have three main observations: i) Compared to the baselines, our CSD demonstrates the best debiasing effectiveness across four different MLLM architectures. In intra-dataset evaluation, i.e., fine-tuning and testing on the CMSC dataset, LLaVA-13B+CSD achieved a MinSkew@C of -0.8167, improving by " + }, + { + "bbox": [ + 313, + 321, + 555, + 548 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 313, + 321, + 555, + 548 + ], + "type": "text", + "content": " compared to LLaVA-13B+FT. In cross-dataset evaluation, our CSD also outperforms the two baselines by a significant margin. For example, Qwen-VL-7B+CSD attains a MaxSkew@C of 0.2607 on the FairFace dataset, showing a reduction of 0.6 in absolute value. ii) A model with a larger size does not necessarily correspond to a lower social bias. This might because the increased scale of parameters makes the model to better learn the biases present in the pre-training dataset. iii) POPE's debiasing effectiveness underperforms FT. One possible reason is that POPE is a training-free method, causing its predictions to remain biased towards the pre-training data distribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 550, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 622 + ], + "type": "text", + "content": "Pertaining to the model results on general multi-modal benchmarks, we observed that our CSD method has a negligible impact, with the effect on all architectures across the three datasets being less than " + }, + { + "bbox": [ + 313, + 550, + 555, + 622 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 550, + 555, + 622 + ], + "type": "text", + "content": ". This indicates that our CSD method enhances the model's social fairness without significantly sacrificing its general capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 632, + 415, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 415, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 415, + 643 + ], + "type": "text", + "content": "6.3. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 650, + 468, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 468, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 468, + 661 + ], + "type": "text", + "content": "6.3.1. Comparison on Ablated CSD" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Our CSD method consists of two key components: the dataset resampling and the rescaled SFLoss. In Table 4, we reported the performance of them. We observe that both components help alleviate social bias. For example," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1746" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 286, + 245 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 286, + 245 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 286, + 245 + ], + "type": "table", + "html": "
ModelSocialCounterfactualsFairFaceCMSC
MinS@CMaxS@CMinS@CMaxS@CMinS@CMaxS@C
LLAVA-7B×-2.05670.3973-2.87920.6457-1.61591.4817
+FT-0.47270.3625-1.01990.4865-0.71420.8058
+Resample-0.44550.3327-0.92380.4838-0.69790.7741
+SFLoss-0.38510.3453-0.89040.4246-0.53850.5838
+CSD-0.35090.3110-0.86220.3950-0.49330.5633
LLAVA-13B×-2.57300.3799-3.36040.5863-1.67300.5350
+FT-0.47480.4051-1.31230.4066-1.51070.4605
+Resample-0.36030.3844-0.81540.3823-1.03640.4292
+SFLoss-0.34700.3905-1.05460.3946-1.03980.4388
+CSD-0.31130.3718-0.71140.3752-0.81670.4192
Qwen-VL-7B×-0.61170.5966-1.63050.8469-1.51141.0961
+FT-0.28010.3966-1.39250.6916-0.81661.0227
+Resample-0.24280.3631-0.80120.3207-0.74140.7813
+SFLoss-0.25800.3004-0.76880.3311-0.63860.8041
+CSD-0.24220.2921-0.76720.2607-0.61930.7180
Bunny-8B×-0.42550.6064-1.13750.5349-1.58291.4173
+FT-0.42020.5851-1.00350.4199-0.82370.9026
+Resample-0.41580.5670-0.98750.3127-0.81410.4008
+SFLoss-0.41710.5534-0.94480.2789-0.79410.4155
+CSD-0.40010.5532-0.90030.2632-0.77450.1575
", + "image_path": "d2204cca243627680d3dbb7afebd09b66859ca697485375153739f2dc9aa1786.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 295, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 295, + 283 + ], + "type": "text", + "content": "Table 4. The performance of ablated CSD. FT: Directly fin-tuning. Resample: Fine-tuning with resampling. SFLoss: Fin-tuning with SFLoss. The best performance is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 74, + 289, + 174, + 357 + ], + "blocks": [ + { + "bbox": [ + 74, + 289, + 174, + 357 + ], + "lines": [ + { + "bbox": [ + 74, + 289, + 174, + 357 + ], + "spans": [ + { + "bbox": [ + 74, + 289, + 174, + 357 + ], + "type": "image", + "image_path": "c8a2b166ec860b7df28de920edc7bbe7df780391fecc1e7a94a8f6736bed26f9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 295, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 383 + ], + "type": "text", + "content": "Figure 5. MaxSkews of (a) Qwen-VL-7B and (b) Bunny-8B across occupations in SocialCounterfactuals for different races." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 289, + 274, + 357 + ], + "blocks": [ + { + "bbox": [ + 175, + 289, + 274, + 357 + ], + "lines": [ + { + "bbox": [ + 175, + 289, + 274, + 357 + ], + "spans": [ + { + "bbox": [ + 175, + 289, + 274, + 357 + ], + "type": "image", + "image_path": "fbeddbcd22e246b84c115a37077dd9cc486373bf0e2ece775d953f8ae62ba80c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 394, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 467 + ], + "type": "text", + "content": "LLaVA-7B+SFLoss achieves a " + }, + { + "bbox": [ + 55, + 394, + 296, + 467 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 394, + 296, + 467 + ], + "type": "text", + "content": " MinSkew@C improvement on FairFace. On CMSC, Qwen-VL-7B+Resample achieves a MinSkew@C of -0.6193, reducing the absolute Skew value by more than 0.9. A notable observation is that either of the SFLoss and data resampling shows better debiasing effect compared to the naive FT strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 169, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 169, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 169, + 485 + ], + "type": "text", + "content": "6.3.2. Comparison on SAs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 583 + ], + "type": "text", + "content": "In Figure 5, we illustrate the MaxSkews [13] of Qwen-VL-7B and Bunny-8B across different races in the SocialCounterfactuals dataset. We can observe that both models exhibit significant social bias without any fine-tuning. For instance, Qwen-VL-7B has a Skew value of approximately 0.76 for White, indicating a strong preference for predicting occupations for this race. Our CSD method achieves better debiasing effects across all races for both models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 165, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 165, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 165, + 601 + ], + "type": "text", + "content": "6.3.3. Skew Distributions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "In Figure 6, we illustrate the MaxSkew distribution across different SCs in CMSC. It can be observed that LLaVA-13B exhibits larger mean and median MaxSkew values, along with significant outliers in certain attributes. For instance, in the 'personality' category, the Skew value for Old-Belligerent reaches 1.6. After direct fine-tuning, the average degree of social bias decreases. However, the outlier issue remains unresolved. In contrast, applying our CSD method achieves a more optimized Skew distribution" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 539, + 228 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 539, + 228 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 539, + 228 + ], + "type": "image", + "image_path": "f5689938d15febc645e230610274cc70d8986847f665f2bc0161b08e7450c2a2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "text", + "content": "Figure 6. MaxSkew distribution for LLaVA-13B on the CMSC dataset. Each box represents the different " + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{Skew}_{a|c}" + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "text", + "content": " values. Our CSD method achieves the most optimized Skew distribution." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 302 + ], + "type": "text", + "content": "across different concepts. Furthermore, noticeable outlier are eliminated, indicating that CSD provides a comprehensive debiasing effect for MLLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 312, + 466, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 466, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 466, + 324 + ], + "type": "text", + "content": "7. Discussion and Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 332, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 452 + ], + "type": "text", + "content": "Conclusion. In this paper, we present to address the notorious social bias problem in MLLMs. Our first contribution is a comprehensive dataset that covers more diverse social concepts than previous datasets. In addition, we advocate an counter-stereotype debiasing approach to perform both dataset resampling and loss rescaling, thereby improving fairness of MLLMs. Extensive experiments demonstrate that our method is promising to alleviate the social bias in MLLMs, with minimal negative impact on their original general multi-modal understanding capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 452, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 583 + ], + "type": "text", + "content": "Social Impact. As MLLMs become increasingly integrated into real-world applications, their societal influence is poised to expand, particularly in sectors requiring sensitive handling of human attributes and social concepts. However, intrinsic social biases within these models can undermine their trustworthiness, raising ethical concerns and potential risks. This paper presents a bias mitigation approach to enhance fairness in MLLMs, fostering trust and accountability by reducing harmful stereotypes. We hope our work inspires further research on bias and fairness in MLLMs, contributing to the development of more equitable models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 594, + 406, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 406, + 606 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 406, + 606 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 661 + ], + "type": "text", + "content": "This research is supported in part by National Natural Science Foundation of China (Grant No. 62476071, No. U24A20328); in part by National Natural Science Foundation of China (No. 62176137)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 671, + 373, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 671, + 373, + 684 + ], + "spans": [ + { + "bbox": [ + 315, + 671, + 373, + 684 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 691, + 555, + 713 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1747" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 160 + ], + "type": "text", + "content": "sch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L. Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Nikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, pages 1-20, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 161, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 295, + 205 + ], + "type": "text", + "content": "[2] Mohsan S. Alvi, Andrew Zisserman, and Christoffer Nelläker. Turning a blind eye: Explicit removal of biases and variation from deep neural network embeddings. In EC-CVW, pages 556-572, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 205, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 294, + 259 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. CoRR, abs/2308.12966:1-15, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 261, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 314 + ], + "type": "text", + "content": "[4] Hugo Berg, Siobhan Mackenzie Hall, Yash Bhalgat, Hannah Kirk, Aleksandar Shtedritski, and Max Bain. A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning. In AACL/IJCNLP, pages 806-822, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 316, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 348 + ], + "type": "text", + "content": "[5] Karan Bhanot, Miao Qi, John S. Erickson, Isabelle Guyon, and Kristin P. Bennett. The problem of fairness in synthetic healthcare data. Entropy, 23(9):1165, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 350, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 381 + ], + "type": "text", + "content": "[6] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. CoRR, abs/2110.01963:1-18, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 426 + ], + "type": "text", + "content": "[7] Tolga Bolukbasi, Kai-Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Tauman Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, pages 4349-4357, 2016. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 427, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 459 + ], + "type": "text", + "content": "[8] Jannik Brinkmann, Paul Swoboda, and Christian Bartelt. A multidimensional analysis of social biases in vision transformers. In ICCV, pages 4914-4923, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 569 + ], + "type": "text", + "content": "[9] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS, pages 1-40, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 571, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 295, + 656 + ], + "type": "text", + "content": "[10] Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, and Weicheng Kuo. Pali: A jointly-scaled multilingual language-image model. In ICLR, pages 1-13, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 2, 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[12] Ching-Yao Chuang and Youssef Mroueh. Fair mixup: Fairness via interpolation. In ICLR, pages 1-11, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 139 + ], + "type": "text", + "content": "[13] Ching-Yao Chuang, Jampani Varun, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. CoRR, abs/2302.00070:1-13, 2023. 3, 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[14] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Ho. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, pages 1-12, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[15] Sepehr Dehdashtian, Lan Wang, and Vishnu Naresh Boddeti. Fairerclip: Debiasing clip's zero-shot predictions using functions in rkhss. In ICLR, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Richard B Felson and Noah Painter-Davis. Another cost of being a young black male: Race, weaponry, and lethal outcomes in assaults. Social Science Research, 41(5):1241-1253, 2012. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Kathleen C. Fraser and Svetlana Kiritchenko. Examining gender and racial bias in large vision-language models using a novel dataset of parallel images. In EACL, pages 690-713, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "text", + "content": "[18] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1-79, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "text", + "content": "[19] Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. Word embeddings quantify 100 years of gender and ethnic stereotypes. PNAS, 115(16):E3635-E3644, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[20] Sahin Cem Geyik, Stuart Ambler, and Krishnamaram Kenthapadi. Fairness-aware ranking in search & recommendation systems with application to linkedin talent search. In KDD, pages 2221-2231, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[21] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the V in VQA matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6325-6334, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "type": "text", + "content": "[23] Yangyang Guo, Liqiang Nie, Harry Cheng, Zhiyong Cheng, Mohan S. Kankanhalli, and Alberto Del Bimbo. On modality bias recognition and reduction. ACM ToMM, 19(3):103:1-103:22, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 644 + ], + "type": "text", + "content": "[24] Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu, Grace Sodunke, Aleksandar Shtedritski, and Hannah Rose Kirk. Visogender: A dataset for benchmarking gender bias in image-text pronoun resolution. In NeurIPS, pages 1-15, 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Muyang He, Yexin Liu, Boya Wu, Jianhao Yuan, Yueze Wang, Tiejun Huang, and Bo Zhao. Efficient multimodal learning from data-centric perspective. CoRR, abs/2402.11530:1-13, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1748" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "editing with cross-attention control. In ICLR, pages 1-15, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "text", + "content": "[27] Phillip Howard, Avinash Madasu, Tiep Le, Gustavo Lujan Moreno, Anahita Bhiwandiwalla, and Vasudev Lal. Social-counterfactuals: Probing and mitigating intersectional social biases in vision-language models with counterfactual examples. In CVPR, pages 11975-11985, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 227 + ], + "type": "text", + "content": "[28] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Barun Patra, Qiang Liu, Kriti Aggarwal, Zewen Chi, Nils Johan Bertil Bjorck, Vishrav Chaudhary, Subhojit Som, Xia Song, and Furu Wei. Language is not all you need: Aligning perception with language models. In NeurIPS, pages 1-22, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[29] Ekaterina Iakovleva, Fabio Pizzati, Philip Torr, and Stéphane Lathuilière. Specify and edit: Overcoming ambiguity in text-based image editing. CoRR, abs/2407.20232:1-10, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 315 + ], + "type": "text", + "content": "[30] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[31] Sepehr Janghorbani and Gerard De Melo. Multi-modal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision-language models. In EACL, pages 1725-1735, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[32] Fangkai Jiao, Zhiyang Teng, Bosheng Ding, Zhengyuan Liu, Nancy Chen, and Shafiq Joty. Exploring self-supervised logic-enhanced training for large language models. In NAACL, pages 926-941, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 446 + ], + "type": "text", + "content": "[33] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, pages 1548-1558, 2021. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[34] Tiep Le, Vasudev Lal, and Phillip Howard. Cocos-counterfactuals: Automatically constructed counterfactual examples for image-text pairs. NeurIPS, pages 1-13, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[37] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, pages 1-13, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[40] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player? CoRR, abs/2307.06281:1-12, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[42] Guangben Lu, Yuzhen Du, Zhimin Sun, Ran Yi, Yifan Qi, Yizhe Tang, Tianyi Wang, Lizhuang Ma, and Fangyuan Zou. Pinco: Position-induced consistent adapter for diffusion transformer in foreground-conditioned inpainting. arXiv preprint arXiv:2412.03812, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 247 + ], + "type": "text", + "content": "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek-vl: Towards real-world vision-language understanding. CoRR, abs/2403.05525:1-29, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "text", + "content": "[44] David Madras, Elliot Creager, Toniann Pitassi, and Richard S. Zemel. Learning adversarially fair and transferable representations. In ICML, pages 3381-3390, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[45] Thomas Manzini, Yao Chong Lim, Alan W. Black, and Yulia Tsvetkov. Black is to criminal as caucasian is to police: Detecting and removing multiclass bias in word embeddings. In NAACL, pages 615-621, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 359 + ], + "type": "text", + "content": "[46] Allison Master. Gender stereotypes influence children's stem motivation. Child Development Perspectives, 15(3):203-210, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 446 + ], + "type": "text", + "content": "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In NeurIPS, pages 27730-27744, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[48] David S Pedulla. The positive consequences of negative stereotypes: Race, sexual orientation, and the job application process. Social Psychology Quarterly, 77(1):75-94, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "text", + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis. CoRR, abs/2307.01952: 1-21, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 602 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "type": "text", + "content": "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21:140:1-140:67, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "text", + "content": "[52] Vikram V. Ramaswamy, Sunnie S. Y. Kim, and Olga Russakovsky. Fair attribute classification through latent space de-biasing. In CVPR, pages 9301-9310, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[53] Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck, Estelle Aflalo, Shao-Yen Tseng, Vasudev Lal, and Phillip" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1749" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 105 + ], + "type": "text", + "content": "Howard. Debias your large multi-modal model at test-time with non-contrastive visual attribute steering. arXiv preprint arXiv:2411.12590, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[54] Prasanna Sattigeri, Samuel C. Hoffman, Vijil Chenthamarakshan, and Kush R. Varshney. Fairness GAN: generating datasets with fairness properties using a generative adversarial network. IBM Journal of Research and Development, 63 (4/5):3:1-3:9, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "text", + "content": "[55] Ashish Seth, Mayur Hemani, and Chirag Agarwal. Dear: Debiasing vision-language models with additive residuals. In CVPR, pages 6820-6829, 2023. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "text", + "content": "[56] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: A dataset for image captioning with reading comprehension. In ECCV, pages 742-758, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 281 + ], + "type": "text", + "content": "[57] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. VL-BERT: pre-training of generic visual-linguistic representations. In ICLR, pages 1–13, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 282, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 336 + ], + "type": "text", + "content": "[58] Yizhe Tang, Zhimin Sun, Yuzhen Du, Ran Yi, Guangben Lu, Teng Hu, Luying Li, Lizhuang Ma, and Fangyuan Zou. Ata: Adaptive transformation agent for text-guided subject-position variable background inpainting. In CVPR, pages 18335-18345, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 370 + ], + "type": "text", + "content": "[59] Özge Tarhan. Children's understanding of the concept of social stereotypes. Child Indicators Research, 15(3):989-1023, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 436 + ], + "type": "text", + "content": "[60] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971:1-16, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[61] Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. Are emily and greg still more employable than lakisha and jamal? investigating algorithmic hiring bias in the era of chatgpt. CoRR, abs/2310.05135:1-13, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 514 + ], + "type": "text", + "content": "[62] W Edgar Vinacke. Stereotypes as social concepts. The journal of social Psychology, 46(2):229-243, 1957. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 546 + ], + "type": "text", + "content": "[63] Angelina Wang and Olga Russakovsky. Overwriting pretrained bias with finetuning data. In ICCV, pages 3934-3945, 2023. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 547, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 294, + 602 + ], + "type": "text", + "content": "[64] Bin Wang, Zhengyuan Liu, Xin Huang, Fangkai Jiao, Yang Ding, AiTi Aw, and Nancy Chen. SeaEval for multilingual foundation models: From cross-lingual alignment to cultural reasoning. In NAACL, pages 370–390. Association for Computational Linguistics, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 635 + ], + "type": "text", + "content": "[65] Jialu Wang, Yang Liu, and Xin Eric Wang. Are gender-neutral queries really gender-neutral? mitigating gender bias in image search. In EMNLP, pages 1995-2008, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[66] Mei Wang and Weihong Deng. Mitigating bias in face recognition using skewness-aware reinforcement learning. In CVPR, pages 9319-9328, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[67] Xuyang Wu, Yuan Wang, Hsin-Tai Wu, Zhiqiang Tao, and Yi Fang. Evaluating fairness in large vision-language models across diverse demographic attributes and prompts. CoRR, abs/2406.17974:1-11, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 419 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[68] Yisong Xiao, Aishan Liu, QianJia Cheng, Zhenfei Yin, Siyuan Liang, Jiapeng Li, Jing Shao, Xianglong Liu, and Dacheng Tao. Genderbias-vl: Benchmarking gender bias in vision language models via counterfactual probing. CoRR, abs/2407.00600:1-13, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 128, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 162 + ], + "type": "text", + "content": "[69] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, pages 1-13, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[70] Haoyu Zhang, Meng Liu, Yuhong Li, Ming Yan, Zan Gao, Xiaojun Chang, and Liqiang Nie. Attribute-guided collaborative learning for partial person re-identification. IEEE TPAMI, 45(12):14144–14160, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[71] Haoyu Zhang, Meng Liu, Zixin Liu, Xuemeng Song, Yaowei Wang, and Liqiang Nie. Multi-factor adaptive vision selection for egocentric video question answering. In ICML, pages 59310-59328. PMLR, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "text", + "content": "[72] Jie Zhang, Sibo Wang, Xiangkui Cao, Zheng Yuan, Shiguang Shan, Xilin Chen, and Wen Gao. Vlbiasbench: A comprehensive benchmark for evaluating bias in large vision-language model. CoRR, abs/2406.14194:1-13, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[73] Michael Zhang and Christopher Ré. Contrastive adapters for foundation model group robustness. NeurIPS, pages 21682-21697, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[74] Xiaobin Zhang, Ziqi Cui, Ning He, and Xuqun You. Are boys associated with weapons and girls associated with kitchenware? the extent to which gender stereotypes regarding adults extend to children. Acta psychologica, 230: 103754, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 419 + ], + "type": "text", + "content": "[75] Kankan Zhou, Eason Lai, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In AACL/IJCNLP, pages 527-538, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1750" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_content_list.json b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a6e2f0094cdfc1f6b6d0428cbd763e68d1fce2a9 --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "Soft Local Completeness: Rethinking Completeness in XAI", + "text_level": 1, + "bbox": [ + 199, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziv Weiss Haddad* Oren Barkan*", + "bbox": [ + 119, + 180, + 413, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Open University", + "bbox": [ + 209, + 207, + 351, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yehonatan Elisha Noam Koenigstein", + "bbox": [ + 558, + 181, + 874, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tel Aviv University", + "bbox": [ + 650, + 207, + 782, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/xaisloc/sloc", + "bbox": [ + 341, + 232, + 650, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 282, + 326, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Completeness is a widely discussed property in explainability research, requiring that the attributions sum to the model's response to the input. While completeness intuitively suggests that the model's prediction is \"completely explained\" by the attributions, its global formulation alone is insufficient to ensure faithful explanations. We contend that promoting completeness locally within attribution subregions, in a soft manner, can serve as a standalone guiding principle for producing faithful attributions. To this end, we introduce the concept of the completeness gap as a flexible measure of completeness and propose an optimization procedure that minimizes this gap across subregions within the attribution map. Extensive evaluations across various model architectures demonstrate that our method produces state-of-the-art results.", + "bbox": [ + 89, + 316, + 483, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 574, + 220, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Explainable AI (XAI) has emerged as a critical aspect of machine learning, especially for deep learning models, which are often regarded as \"black boxes\" due to their lack of transparency [8, 9, 25, 41, 66]. XAI aims to provide insights into how these models make predictions, offering a way to ensure their reliability and trustworthiness [34]. A common approach to model explainability is through attribution maps, which assign importance scores to individual input features based on their contribution to the model's prediction [22, 27, 65]. These attributions serve as a tool for understanding which aspects of the input were most influential in determining the output.", + "bbox": [ + 89, + 599, + 482, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A widely discussed notion in XAI is completeness (also known as conservation or efficiency) [4, 55, 60, 65]. Completeness is defined as the requirement that the elements in an attribution map should sum to the difference between the model's prediction for a given input and a baseline (e.g., the null representation). This difference is also known as the", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model's response. When completeness is achieved inherently by an attribution method, it is intuitively appealing, as it suggests that the attribution map fully captures the factors contributing to the model's response. However, even a low-quality attribution map can be superficially adjusted to satisfy completeness through post hoc global normalization. Yet, such an approach lacks genuine explanatory power, as the normalization is decoupled from the underlying attribution mechanism.", + "bbox": [ + 511, + 284, + 905, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper recognizes that completeness, as defined in the literature as a global property that an explanation method should satisfy, is too weak to serve as a standalone criterion. Indeed, several well-regarded explanation methods [32, 39, 59] do not inherently satisfy completeness. Nevertheless, these methods have demonstrated considerable effectiveness across diverse objective evaluation metrics, providing faithful explanations and successfully passing various sanity checks [3]. This suggests that completeness is neither a necessary nor sufficient condition for generating high-quality explanations.", + "bbox": [ + 511, + 424, + 905, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we differentiate between global completeness, as originally defined in the literature, which imposes a completeness constraint on the entire attribution map, and local completeness, which requires completeness within individual subregions of the attribution map. By rethinking completeness as a local and flexible guiding measure rather than a strict global constraint, we seek to overcome its limitations while harnessing the appealing motivation behind it for producing meaningful and faithful explanations. To achieve this, we require subregions of the attribution map, referred to as sub-maps, to promote completeness locally by accounting for the model's response to the corresponding subregions of the input image, yet in soft manner. An attribution map that adheres to this guiding principle should assign high (low) importance to image subregions where the model exhibits a strong (weak) response, thereby maintaining faithfulness to the model.", + "bbox": [ + 511, + 594, + 905, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike global completeness, which is a relatively weak requirement that can be easily satisfied, achieving local completeness across an arbitrary set of sub-maps may not", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19794", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "always be feasible. To address this challenge, we introduce the completeness gap - a measure that quantifies the deviation of a sub-map from completeness. Specifically, the completeness gap measures the difference between the sum of elements in each sub-map and the model's response to the corresponding subregions of the input image. Rather than imposing a strict binary constraint, the completeness gap serves as a soft, quantifiable criterion for local completeness. We argue that minimizing this gap over a diverse set of sub-maps leads to faithful explanations.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present Soft LOcal Completeness (SLOC) - a novel explainability method that promotes completeness locally within sub-maps of the attribution map in a soft and flexible manner. SLOC accomplishes this by seeking to minimize the completeness gap locally for each individual sub-map, simultaneously.", + "bbox": [ + 89, + 242, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The motivation for minimizing the completeness gap locally is to emphasize or attenuate sub-maps based on the actual impact of their corresponding input subregions on the model's output. SLOC achieves this through a gradient-based optimization process, where each sub-map is adjusted to achieve local completeness in a soft manner. Therefore, SLOC's goal is not to enforce strict completeness, globally, but to use it as a guiding principle for refining a large set of sub-maps (subregions within the attribution), iteratively and simultaneously.", + "bbox": [ + 89, + 332, + 480, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution is the introduction of the SLOC method, which facilitates a novel optimization procedure to promote completeness in a soft local manner. The effectiveness of SLOC is demonstrated through extensive experiments on various model architectures, where it is shown to produce state-of-the-art results across multiple benchmarks.", + "bbox": [ + 89, + 484, + 480, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 588, + 232, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The XAI literature encompasses a broad range of approaches for attributing model predictions to specific input features across different tasks [2, 4, 6, 7, 16, 24, 29, 40, 54]. Early gradient-based methods produce explanation maps by leveraging gradients directly [5, 60, 62] or through functions that combine class activation maps with or without their gradients [10, 26, 48, 58, 59, 63]. With the rise of transformer architectures, new explainability techniques for transformers have emerged [1, 4, 11, 67]. For example, Transformer Attribution [28] introduces a class-specific Deep Taylor Decomposition, applying relevance propagation for both positive and negative attributions. Building on this, Generic Attention Explainability [27] generalizes Transformer Attribution to explain bi-modal transformers.", + "bbox": [ + 89, + 613, + 480, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Path integration methods form another prominent family of attribution techniques [13, 15, 18, 50, 65]. Integrated Gradients (IG) [65] computes attributions by integrating gradients along an interpolation path from a baseline to the input. A refined variant, Guided Integrated", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradients (GIG) [50], uses an adaptive path that avoids high-gradient regions, thereby reducing irrelevant attributions. Recently, Deep Integrated Explanations (DIX) [14] proposed performing integration over intermediate network representations instead of inputs, producing more faithful attribution maps than IG and GIG.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perturbation-based (occlusion) techniques generate attribution maps by perturbing parts of the input to assess the contributions of specific elements to the prediction [17, 19, 20, 53]. For example, RISE [57], creates perturbations by masking areas in the image through the upsampling of randomly drawn low-resolution binary grids. The class score corresponding to each masked version of the image serves as an importance score for that specific mask. A linear combination of all masks, weighted by their importance, then forms the final attribution map. Another branch of perturbation methods are learning-based perturbation techniques [12, 30, 38, 39]. For example, Meaningful Perturbation (MP) [39] learns to produce perturbations by masking the smallest region possible while significantly altering the original prediction, isolating the minimum necessary content for the prediction. Alternatively, LTX [12] introduces a surrogate 'explainer' model pretrained to mask as much of the input as possible while preserving the original prediction, thereby ensuring the retained features are those most relevant to the model's prediction. Then, LTX finetunes the attribution per specific example, allowing the selection of the best-performing attribution w.r.t. to the metric at hand.", + "bbox": [ + 511, + 183, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our SLOC method falls within the family of learning-based perturbation techniques, yet it diverges from prior works in several key aspects. First, conceptually, SLOC presents a distinct methodology by promoting local completeness rather than attempting to maximize or minimize the model prediction through learned perturbations [12, 30, 39]. Consequently, SLOC does not require gradient backpropagation through the model, and incorporates the model's predictions on the masked inputs into the loss as constant nodes within the computation graph. This design yields a simple, computationally efficient gradient expression, and relies solely on the model's forward pass, hence operating in a true black-box setting. Finally, unlike LTX, SLOC avoids the overhead of learning an 'explainer' function, which would also require a dataset aligned with the training data distribution of the model being explained.", + "bbox": [ + 511, + 530, + 903, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Soft Local Completeness", + "text_level": 1, + "bbox": [ + 511, + 787, + 743, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let $f: \\mathbb{R}^n \\to [0,1]^C$ be a classification model that takes an input image $\\mathbf{x} \\in \\mathbb{R}^n$ and outputs a discrete probability distribution $f(\\mathbf{x}) \\in [0,1]^C$ over $C$ classes. Given a target class $y$ to explain, our goal is to produce an attribution map", + "bbox": [ + 511, + 813, + 903, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ W.l.o.g, we represent images as vectors in $\\mathbb{R}^n$ .", + "bbox": [ + 531, + 886, + 784, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\mathbf{a}_{\\mathbf{x}}^{y} \\in \\mathbb{R}^{n}$ that attributes the contribution of each element $i$ in $\\mathbf{a}_{\\mathbf{x}}^{y}$ to the prediction score $f(\\mathbf{x})[y]$ for the class $y$ .", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Completeness requires that the attributions in $\\mathbf{a}_{\\mathbf{x}}^{y}$ collectively account for the difference between the model's prediction on the input $\\mathbf{x}$ and the baseline representation $\\mathbf{b}$ , which is set to the black image in this work.", + "bbox": [ + 89, + 121, + 482, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model's response to $\\mathbf{x}$ is defined as $r(\\mathbf{x})\\coloneqq f(\\mathbf{x}) - f(\\mathbf{b})$ . Formally, the attribution map $\\mathbf{a}_{\\mathbf{x}}^{y}$ satisfies completeness if $\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i] = r(\\mathbf{x})[y]$ . We note that this property can be superficially imposed on an attribution map through a post-processing step by normalizing each element using the factor $\\frac{r(\\mathbf{x})[y]}{\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i]}$ . However, artificially satisfying completeness through such normalization does not inherently yield a meaningful explanation, as the imposed completeness is an external adjustment rather than an intrinsic characteristic of the explanation method itself. For example, even a flat attribution map (where all elements are equally weighted) could be normalized to satisfy the completeness criterion, yet such a map lacks any explanatory value. Consequently, relying solely on completeness, in its global form, is deemed insufficient for producing meaningful attribution maps.", + "bbox": [ + 89, + 181, + 483, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 438, + 212, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our SLOC approach is motivated by the toy example illustrated in Fig. 1, which depicts an image where the key region influencing the model's prediction for the 'great-grey-owl' class is approximately centered. Figures 1(a)-(d) present four images: the original image (a) and three perturbed versions obtained by masking different regions (b)-(d).", + "bbox": [ + 89, + 459, + 482, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The four different masks applied in Figures 1(a)-(d) are illustrated in Figure 1(e). Each mask selectively perturbs the original image by replacing certain regions with the baseline (in this work, the null pixels). For instance, the mask corresponding to Figure 1(c), shown in the top-right corner of Figure 1(e), applies a perturbation by replacing all pixels within the black rectangle (located at the bottom-right corner) with the baseline.", + "bbox": [ + 89, + 566, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, the model's response for the great-grey-owl class remains consistent across all four images (a-d), with a value of 0.8. Recall that the model response is defined as the difference between the model's class probability for a given image and that for the baseline (the null image).", + "bbox": [ + 89, + 686, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As defined earlier, a sub-map represents a subset of elements in the attribution map that are associated with spe", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(a) Original image: 0.8", + "bbox": [ + 526, + 92, + 638, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(c) Masked Image: 0.8", + "bbox": [ + 650, + 92, + 767, + 103 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(e) Four masks", + "bbox": [ + 774, + 92, + 854, + 103 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b533a8a769d8c12588b6a047ff01364eb955602d2c445a1184a530175715222c.jpg", + "image_caption": [ + "(b) Masked Image: 0.8" + ], + "image_footnote": [], + "bbox": [ + 524, + 107, + 645, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa29e25e40b9a32422368f876b13d79fd6f925d63204c3f28384f530f25bcfde.jpg", + "image_caption": [ + "(f) Attribution Map" + ], + "image_footnote": [], + "bbox": [ + 650, + 107, + 771, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f83fdecf052eb91401b67d7503b9af933f85e31381411d9718162246f9b80f01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 107, + 897, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/86d871f7e3f47b7d9e769583d6be8bffc4198da1c7b6ea0c536cec1aea088375.jpg", + "image_caption": [ + "(d) Masked Image: 0.8" + ], + "image_footnote": [], + "bbox": [ + 524, + 198, + 645, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc026f536af31b0b5bb11d6d9ec64ef1776710d20a0973f005bad06109110bde.jpg", + "image_caption": [ + "Figure 1. Subfigures (a)-(d) present the model's response (0.8) for the class 'great-grey-owl' when applied to the original input (a) and its masked versions (b)-(d). The four masks used to generate (a)-(d) are depicted in (e). A sub-map refers to a subset of elements in the attribution map that correspond to specific subregions in the input. We consider four distinct sub-maps (not shown), each corresponding to the non-masked (visible) subregions in (a)-(d). For instance, the sub-map associated with (c) includes all elements in the attribution map except those corresponding to the black rectangle in the bottom-right corner of (c). This exclusion aligns with the masked region in the input, effectively determining which attributions are included in that sub-map. A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by (c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to (c). By enforcing local completeness across the four sub-maps induced by (a)-(d), the remaining central stripe, which is unmasked in all images (a)-(d), must account for the model's entire response of 0.8. This results in the attribution depicted in (f) that highlights the central stripe containing the owl in red. See 3.1 for more details." + ], + "image_footnote": [], + "bbox": [ + 650, + 198, + 771, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/73a75665b9f038bf3a5911b1cd128e608ad969183787279c05dcab096c01a115.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 198, + 897, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cific subregions in the input. In the context of Fig. 1, we consider four different sub-maps, each corresponding to the non-masked (visible) subregions in Figs.1(a)-(d). For instance, the sub-map corresponding to Fig. 1(c) consists of all elements in the attribution map except those associated with the black rectangle in the bottom-right corner of Fig.1(c). This exclusion aligns with the masked region in the input, effectively defining which attributions are considered within that sub-map. Notably, the sub-map corresponding to the original image (Fig. 1(a)) is equivalent to the entire attribution map, as no regions are masked. This formulation allows us to analyze how different subregions contribute to completeness by evaluating attributions over various sub-maps.", + "bbox": [ + 511, + 594, + 906, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by Fig. 1(c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to Fig. 1(c).", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2The baseline representation should resemble missing information or the neutral representation. While determining the ultimate representation for the baseline is an open question and an active research field in explainability [21, 36], in this work, we opt for the simple choice of a null baseline represented by the black image. This choice is on par with other notable works [57, 65]. Additionally, we conducted experiments using alternative baselines and observed that the resulting trends remained consistent with those obtained using the black image.", + "bbox": [ + 89, + 803, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19796", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key insight is that enforcing local completeness across the four sub-maps induced by Figs.1(a)-(d) allows us to pinpoint the important region of the image w.r.t. the model's prediction. Specifically, the intersection of these four sub-maps forms a central stripe around the owl, illustrated in red in Fig.1(f). By evenly distributing the total attribution sum of 0.8 across the pixels within this intersection (red stripe) and assigning 0 elsewhere (highlighted in dark blue in Fig. 1(f)), we obtain an attribution map that ensures local completeness for all four sub-maps. Since each sub-map contains the intersection while the attribution values outside it are zero, the sum of attributions within each sub-map is exactly the sum of attributions within the intersection: 0.8. This resulting attribution map effectively captures the key region containing the owl.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This simple example can be slightly generalized under similar settings. We make the following assumptions: (1) The model response for a subregion is $R$ if the subregion contains the object of interest and 0 otherwise. (2) Each sub-map either fully contains the object of interest or does not intersect with it at all. (3) One of the sub-maps corresponds to the complete attribution map. (4) All attribution values are non-negative. It can then be shown that if an attribution map satisfies local completeness for all sub-maps, the intersection defined by these sub-maps is the largest subregion that can have non-zero attribution. Simply put, this intersection highlights the important regions in the image responsible for the model's prediction. Moreover, as the number of sub-maps increases, the important regions can potentially become more precisely defined. The formalization and proof follow straightforwardly and are provided in Appendix K.", + "bbox": [ + 89, + 320, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keep in mind, however, that this setting and its assumptions are simplistic and unrealistic, serving only to illustrate the motivation. Therefore, SLOC does not actually rely on these assumptions. In more realistic scenarios, satisfying local completeness for all sub-maps is often infeasible (an example is provided in Fig. 11 in Appendix I). To address this challenge, we introduce the completeness gap, which quantifies the degree to which a sub-map deviates from local completeness. Our approach seeks to minimize the completeness gap across multiple sub-maps simultaneously, hence promoting completeness locally within subregions of the attribution map, in a soft manner. We formulate this as an optimization problem and employ gradient descent to find a solution.", + "bbox": [ + 89, + 580, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, recall that in our simple example, in the absence of additional information, we distributed the attribution uniformly across the central red stripe. When introducing the loss terms for the SLOC optimization, we will also discuss regularization terms. One such term is the Total Variation loss, which encourages spatial smoothness in the attribution. As we demonstrate in Sec. 4, by encouraging this soft", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "form of local completeness, the resulting attribution map offers robust explanatory performance, both visually and according to a variety of evaluation metrics.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SLOC optimization", + "text_level": 1, + "bbox": [ + 511, + 145, + 702, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{m} \\in \\{0,1\\}^n$ be a binary mask. We define the masked input by", + "bbox": [ + 511, + 166, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\mathbf {m}} = \\mathbf {x} \\circ \\mathbf {m} + (1 - \\mathbf {m}) \\circ \\mathbf {b}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 196, + 906, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\circ$ stands for the elementwise product. In words, $\\mathbf{x}^{\\mathbf{m}}$ is a perturbation (masked version) of $\\mathbf{x}$ , according to the mask $\\mathbf{m}$ , where all masked elements of $\\mathbf{x}^{\\mathbf{m}}$ are replaced with their respective elements from the baseline $\\mathbf{b}$ .", + "bbox": [ + 511, + 219, + 905, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we consider binary masks constructed over patches, where each patch is randomly assigned a Bernoulli outcome with probability $p$ . We explore two methodologies for automatically selecting $p$ : one that determines $p$ dynamically per input image, and another that uses a fixed value per model. The reader is referred to Appendix D for the exact details of the masks construction process, including the methodologies used for determining $p$ .", + "bbox": [ + 511, + 279, + 905, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an initial attribution map $\\mathbf{a}_{\\mathbf{x}}^{y}$ and a set of masks $\\mathcal{M} \\subset \\{0,1\\}^n$ , the SLOC loss is defined as follows:", + "bbox": [ + 511, + 400, + 905, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}; \\mathcal {M}\\right) = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} \\underbrace {(r \\left(\\mathbf {x} ^ {\\mathbf {m}}\\right) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {c o m p l e t e n e s s g a p}} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 439, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cdot$ is the dot-product operator, and $|\\cdot|$ denotes the L1 norm. As observed in Eq. 2, $\\mathcal{L}_c$ promotes the minimization of the completeness gap, a soft measure of completeness, locally, for all sub-maps induced by $\\mathcal{M}$ . The rationale behind the local minimization of the completeness gap is that attributions within a sub-map should be strengthened or reduced so that their sum closely matches $r(\\mathbf{x}^{\\mathbf{m}})[y]$ - the model's response to the respective regions exposed in $\\mathbf{x}^{\\mathbf{m}}$ . This behavior is achieved by applying gradient descent on $\\mathcal{L}_c$ with respect to $\\mathbf{a}_{\\mathbf{x}}^{y}$ .", + "bbox": [ + 511, + 496, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better understand the optimization process, one can inspect the gradient:", + "bbox": [ + 511, + 647, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y}} = - \\frac {1}{| \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) \\mathbf {m}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 685, + 903, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity, assume $\\mathcal{M}$ contains a single mask $\\mathbf{m}$ . Then, for each element $\\mathbf{a}_{\\mathbf{x}}^{y}[i]$ that is exposed in $\\mathbf{x}^{\\mathbf{m}}$ (i.e., $\\mathbf{m}[i] = 1$ ), the gradient is the negative of the completeness gap normalized by the number of elements in the sub-map, i.e., $\\frac{\\partial\\mathcal{L}_c}{\\partial\\mathbf{a}_x^y[i]} = -\\frac{1}{|\\mathbf{m}|} (r(\\mathbf{x}^{\\mathbf{m}})[y] - \\mathbf{a}_{\\mathbf{x}}^y\\cdot \\mathbf{m})$ , and for each $i$ where $\\mathbf{m}[i] = 0$ , the gradient vanishes. Accordingly, the gradient descent update ensures an equal increase (decrease) for all elements in the sub-map when the sum of the attributions falls short of (exceeds) the model's response $r(\\mathbf{x}^{\\mathbf{m}})[y]$ . Assuming a learning rate of 1, it follows that after the gradient descent update, the completeness gap vanishes, resulting in", + "bbox": [ + 511, + 732, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19797", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an updated sub-map satisfying completeness. This can be seen mathematically as follows: for a sub-map induced by the mask $\\mathbf{m}$ , the sum of its elements after the gradient update is given by", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] - \\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ]} \\\\ = \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] + \\underbrace {\\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {i n d e p e n d e n t o f} i} \\\\ = | \\mathbf {m} | \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) + \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] \\\\ = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} + \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 161, + 472, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the second transition follows from the fact that the normalized completeness gap is independent of $i$ and is therefore summed $|\\mathbf{m}|$ times, as the summation is taken over every $i$ where $\\mathbf{m}[i] = 1$ , with $\\mathbf{m}$ being a binary tensor. The penultimate transition then arises from expressing the sum of elements in the sub-map as the dot product, $\\mathbf{a}_{\\mathbf{x}}^{y} \\cdot \\mathbf{m}$ . This derivation justifies the normalization of the squared completeness gap by $|\\mathbf{m}|$ in Eq. 2, as it ensures that after the gradient update, the sub-map satisfies local completeness.", + "bbox": [ + 89, + 325, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To give further intuition for using the normalized completeness gap in Eq. 2, consider two different masks: one that exposes a single pixel (element), and another that exposes a large portion of the input. Given that the gradient update of elements in the sub-map is equal for all elements, it makes sense to distribute the completeness gap among all elements. In the case of a mask exposing a single pixel, we have full confidence that this specific element is responsible for the model's response $r(\\mathbf{x}^{\\mathrm{m}})[y]$ . Therefore, it is reasonable to expect that the gradient update for this element will match the size of the completeness gap. However, in the case of a sub-map associated with a mask exposing large portions of the input, we cannot differentiate the accountability of individual elements within the exposed portions for the model's response. Thus, without further information, it is reasonable to apply an update step that evenly distributes the completeness gap across all exposed elements. Hence, the squared completeness gap is divided by $|\\mathbf{m}|$ .", + "bbox": [ + 89, + 478, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "So far, the discussion and analysis have been limited to the case of a single mask, i.e., $|\\mathcal{M}| = 1$ . When considering a set of masks, the update for $\\mathbf{a}_{\\mathbf{x}}^{y}[i]$ is determined by the accumulation of gradients from all the sub-maps it is associated with. Specifically, sub-maps that include the $i$ -th element and whose total sum of elements is below (exceeds) the model's response $r(\\mathbf{x}^{\\mathrm{m}})[y]$ will contribute to the intensification (reduction) of $\\mathbf{a}_{\\mathbf{x}}^{y}[i]$ , with the contribution from each individual mask weighted according to the normalized completeness gap it induces.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4249ef80320f2516bc639b72cde85f1ac389b5da773249ce5d9982fa591efaeb.jpg", + "image_caption": [ + "Masks & Responses Generation Phase", + "Figure 2. SLOC overview: The upper section illustrates the Masks & Responses Generation Phase. In this phase, (1) $\\mathcal{M}$ masks are randomly generated. (2) Perturbations are produced by combining the original image and baseline according to the masks, as defined in Eq. (1) (3) These perturbations are passed as inputs to the model, and the corresponding outputs are obtained. (4) Subtracting the model's output for the baseline from these outputs produces the model responses. (5) These masks and their corresponding responses are then passed to the Optimization Phase. (6) Starting with an initial attribution. (7) Gradient descent is used to iteratively update the attribution. The loss function $\\mathcal{L}$ (Eq. 4) consists of three components: the completeness gap (computed using the masks, responses, and attributions), as well as TV and L1 regularization. See Sec. 3.2 for details." + ], + "image_footnote": [], + "bbox": [ + 516, + 104, + 903, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To promote smooth and focused attribution maps, we combine $\\mathcal{L}_c$ with two types of regularization applied to the attribution map: Total Variation (TV) and L1 regularization that encourage smoothness and sparsity, respectively. The final loss function is then defined as:", + "bbox": [ + 511, + 556, + 906, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) = \\mathcal {L} _ {c} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) + \\lambda_ {1} | \\mathbf {a _ {x} ^ {y}} | + \\lambda_ {2} \\mathrm {T V} (\\mathbf {a _ {x} ^ {y}}), (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 641, + 906, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 513, + 667, + 549, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T V} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}\\right) := \\sum_ {i, j} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i + 1, j ]\\right) ^ {2} + \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j + 1 ]\\right) ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 688, + 928, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{a}_{\\mathbf{x}}^{y}$ in Eq. 5 is reshaped to match the spatial dimensions of the final attribution map. Additionally, $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters controlling the strength of the regularization terms. In our experiments, these hyperparameters are set based on a separate validation set.", + "bbox": [ + 511, + 734, + 906, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization of $\\mathbf{a}_{\\mathbf{x}}^{y}$ is carried out via gradient descent on $\\mathcal{L}$ . The optimization process allows for tracking the performance of the metric during the optimization and selecting the best-performing attribution map based on this metric from the resulting maps [12]. A diagram of the SLOC method is depicted in Fig. 2.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19798", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35fc463f0d7766e6c4af7289daae1f6ac631fc65990a80dc1d83d2907e3e4592.jpg", + "image_caption": [ + "Figure 3. SLOC attribution maps across training steps. Faithful explanations emerge after a few hundred gradient updates, with an appropriate learning rate decay." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practice, we observed that a few hundred gradient updates, combined with an appropriate learning rate decay, were sufficient to produce faithful and meaningful explanations. Therefore, in our experiments, we report SLOC results for $T = 500$ . An example for SLOC's attribution across gradient update steps is shown in Fig. 3. For the exact implementation details and hyperparameter configuration, the reader is referred to Appendix D.", + "bbox": [ + 88, + 248, + 482, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to space limitations, Appendix G provides a detailed discussion of SLOC's computational complexity. Appendix H presents a runtime analysis, comparing SLOC against various methods and demonstrating its efficiency over other state-of-the-art learning-based perturbation methods. Additionally, Appendix L examines additional axioms and explains which of them are satisfied by SLOC.", + "bbox": [ + 88, + 369, + 482, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, it is important to clarify that the SLOC optimization procedure does not guarantee convergence to an attribution map that fully satisfies completeness. Instead, it leverages the completeness gap as a guiding measure to update subregions within the attribution map, thereby promoting local completeness for these subregions in a soft manner. Nevertheless, as demonstrated in Sec. 4, SLOC has proven effective in generating attribution maps that outperform state-of-the-art explanation methods.", + "bbox": [ + 88, + 489, + 482, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 638, + 323, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 662, + 277, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experiments were conducted on an NVIDIA DGX 8xA100 Server, utilizing the PyTorch package. As explanations are multifaceted by nature, no single evaluation metric can fully capture their quality, and no definitive metric exists [21, 23, 35, 45, 47, 51, 68]. Therefore, our evaluation encompasses several distinct protocols designed to assess explanation quality from multiple complementary perspectives, on different datasets, and across four model architectures: ResNet50 (RN) [43], DenseNet201 (DN) [46], ViT-Base (ViT-B) and ViT-Small (ViT-S) [33].", + "bbox": [ + 88, + 685, + 482, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment 1: Faithfulness evaluation This evaluation assesses the faithfulness (also known as correctness or fidelity) of the explanation via a set of perturbation-based", + "bbox": [ + 88, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metrics. These metrics are designed to reveal the actual input elements the model relies on for its predictions. Following prior works [28, 49, 50, 69], we conducted an extensive evaluation using a comprehensive set of perturbation-based explanation metrics to assess the faithfulness of the generated explanations: the Area Under the Curve (AUC) of Positive (POS) and Negative (NEG) perturbation tests [28], AUC of the Insertion (INS) and Deletion (DEL) tests [57], and AUC of the Softmax Information Curve (SIC) and Accuracy Information Curve (AIC) [49]. For POS and DEL, lower values indicate better performance, while for NEG, INS, SIC, and AIC, higher values are preferred. Since NEG and POS, as well as INS and DEL, are complementary metrics, we also report the NEG-POS Difference (NPD) and the INS-DEL Difference (IDD) as single summaries of their respective complementary performances. Faithfulness performance is reported on a sample from the ImageNet [31] ILSVRC 2012 (IN) validation set, which consists of 10,000 images across 1,000 classes. Since most images in the IN dataset contain a single, centrally focused object, we extend our evaluation to multi-object scenes by assessing explanation faithfulness on a sample from the Pascal VOC 2012 [37] (VOC) test set, which includes 1,000 images spanning 20 object classes.", + "bbox": [ + 511, + 90, + 906, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment 2: FunnyBirds evaluation While perturbation-based protocols are a standard approach for assessing explanation faithfulness, they are criticized for introducing domain shifts that undermine the validity of the evaluation [45]. To address this, we additionally report results on the FunnyBirds (FB) dataset which consists of 500 images from 50 classes, following the evaluation protocol from [44]. The FB evaluation employs controlled interventions to estimate ground-truth (GT) importance scores at the level of object parts rather than individual pixels, mitigating issues associated with pixelwise perturbation-based evaluations [68]. Furthermore, FB avoids domain shifts by incorporating semantically meaningful interventions during training. By evaluating explanations at the part level—closer to human visual understanding—FB reduces the gap between automatic XAI evaluation (e.g., faithfulness metrics) and human-centric studies, providing a more interpretable and robust assessment of the explanations. FB evaluates explainability in three aspects: Completeness, Correctness, and Contrastivity - and provides a combined overall score (higher is better).", + "bbox": [ + 511, + 472, + 908, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment 3: Segmentation evaluation To ensure a comprehensive comparison with prior works [27, 28, 48], we conduct an extensive evaluation using segmentation tests on the ImageNet-Seg dataset (IN-Seg) [42], a subset of the ImageNet validation set comprising 4,276 human-annotated", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19799", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ground-truth (GT) segmentation maps across 445 classes. As detailed in [28], these tests generalize the \"Pointing Game\" [45] and serve as a human-grounded evaluation [56] by measuring the alignment between GT segmentation maps and the attributions generated by the explanation method. While higher segmentation accuracy does not necessarily indicate greater explanatory power [57], segmentation tests assess the alignment between explanations and human-annotated GT, hence providing a complementary perspective to other evaluations. We follow the evaluation protocol from [28], assessing segmentation performance using Pixel Accuracy (PA), mean Intersection-over-Union (mIoU), and mean Average Precision (mAP), higher is better for all metrics.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Detailed descriptions of the evaluation protocols and metrics used in Experiments 1, 2, and 3 are provided in Appendices B.1, B.2, and B.3, respectively.", + "bbox": [ + 89, + 301, + 480, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluated methods We evaluate SLOC against a comprehensive suite of 14 explanation methods encompassing gradient-based, path-integration, perturbation, and CAM methods. For CNN models, we include the following methods: Grad-CAM (GC) [59], Grad-CAM++ $(\\mathbf{GC} + +)$ [26], Deep Integrated Explanations (DIX) [14], FullGrad (FG) [63], Ablation-CAM (AC) [32], Layer-CAM (LC) [48], Learning To Explain (LTX) [12], RISE [57], Meaningful Perturbation (MP) [39], Extremal Perturbations (EP) [38], Integrated Gradients (IG) [64], and Guided IG (GIG) [50]. For ViT models, we considered the following methods: Transformer Attribution (T-Attr) [28] and Generic Attention Explainability (GAE) [27], alongside DIX, RISE, LTX, EP, and MP, which are applicable to both CNN and ViT architectures. Hyperparameters for all methods were set according to the recommended settings published by the authors, unless a better configuration was found. A detailed description of all explanation methods is provided in Appendix C. Finally, we evaluated three different versions of our SLOC method on both CNN and ViT models: SLOC, which tunes $p$ dynamically per input, $\\mathrm{SLOC}_{\\mathrm{xp}}$ , where $p$ is calibrated per model (i.e., fixed across all inputs for a given model) based on a validation dataset, and $\\mathrm{SLOC}_{\\mathrm{m}}$ , which is identical to SLOC but additionally monitors the IDD metric on the specific input and selects the explanation with the best performance according to this metric. All versions were run with $T = 500$ . The precise implementation details, including the monitoring procedure, the $p$ tuning procedures, optimization process, and all hyperparameter configurations, are provided in Appendix D and in our GitHub repository.", + "bbox": [ + 91, + 354, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 832, + 184, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tables 1 and 2 present a quantitative comparison of SLOC with other state-of-the-art explanation methods across multiple faithfulness metrics, for DN and ViT-S, respectively.", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b34f2d404e7a71d8bfc772af98fc76d041ea4179f6613b74e8afd3377dbb3026.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)11.6176.549.3965.7964.9356.479.2178.35
SLOC11.5270.979.1260.5359.4551.4177.7977.72
\\( SLOC_{xp} \\)11.7870.209.3459.7558.4350.4177.3376.80
AC17.2467.6813.2757.1850.4443.9177.7875.41
DIX13.3662.9510.3652.4349.5942.0774.6271.47
EP16.1265.6812.9455.049.5542.0677.3874.97
FG19.0644.6615.2637.6225.622.3758.2353.86
GC16.8868.5413.0457.9551.6644.9178.3876.01
GC++17.3667.2913.3456.7549.9343.4178.0475.62
GIG14.8149.9612.2841.9335.1629.6561.1257.6
IG14.1451.9911.244.0837.8532.8861.2658.48
LC17.2867.2713.3156.7149.9943.477.9575.42
LTX16.2471.0912.9559.6954.8446.7478.9276.25
MP18.5453.2414.8743.7934.728.9266.9864.13
RISE18.4262.7514.2652.9944.3338.7376.8274.24
", + "bbox": [ + 517, + 87, + 924, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cfd6aae8bc9b8a0dc1af31383467bd0143a3e6682a136976cb7912e76a99399b.jpg", + "table_caption": [ + "Table 1. Faithfulness results for all combinations of method and metric, using the DN model on the IN dataset." + ], + "table_footnote": [], + "table_body": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)14.8381.8112.2770.3166.9858.0483.7883.34
SLOC15.2577.9712.4766.5162.7254.0483.1982.88
\\( SLOC_{xp} \\)15.7977.8512.8566.1862.0653.3383.0681.97
DIX18.6968.1714.8656.8349.4841.9776.8875.05
EP27.3772.5221.6660.845.1439.1479.4577.56
GAE19.9866.9315.9355.7246.9539.7975.4273.9
LTX20.8468.516.6356.8947.6640.2774.2271.56
MP27.7263.2522.4652.2835.5329.8174.2271.25
RISE29.5169.5223.4757.9340.0234.4679.9377.23
TATTR19.0667.5214.9156.4148.4641.578.0375.69
", + "bbox": [ + 517, + 308, + 929, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Faithfulness results for all combinations of method and metric, using the ViT-S model on the IN dataset.", + "bbox": [ + 511, + 433, + 906, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results for RN and ViT-B are available in Tabs. 7 and 8 in Appendix E, respectively. Overall, the results indicate that all SLOC versions are the best-performing methods, with LTX, DIX, RISE, and EP as runners-up, depending on the combination of architecture and evaluation metric.", + "bbox": [ + 511, + 489, + 906, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among the three versions of SLOC, $\\mathrm{SLOC}_{\\mathrm{m}}$ performs best. This can be attributed to its use of the IDD metric, which measures the difference between INS and DEL performance. The fact that $\\mathrm{SLOC}_{\\mathrm{m}}$ outperforms across the vast majority of metrics suggests that the IDD metric correlates well with other faithfulness metrics. SLOC and $\\mathrm{SLOC}_{\\mathrm{xp}}$ follow, with SLOC performing slightly better. This suggests that tuning $p$ per instance offers a modest advantage in metrics, at the cost of increased inference complexity. When a representative sample of data is available, $p$ can be calibrated per model using this data. Otherwise, $p$ should be adjusted dynamically per input during inference.", + "bbox": [ + 511, + 566, + 908, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, both SLOC and $\\mathrm{SLOC}_{\\mathrm{xp}}$ surpass learning-based approaches such as LTX, MP, and EP, achieving state-of-the-art results without requiring gradient backpropagation through the model (see the discussion on SLOC complexity in Appendix G) or metric monitoring (as in LTX). Figures 6 and 5 present comparative examples of the top-performing methods for DN and ViT-S. Arguably, SLOC produces focused explanation maps highlighting relevant class-discriminative features in the image.", + "bbox": [ + 511, + 748, + 908, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents faithfulness results on the VOC dataset", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/21c8ed707eccc1e9584f1e9e63e06c9b1ef8611b7710b02765014f79d133d6d1.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of attributions produced by different methods, using the DN model w.r.t. the classes (top to bottom): 'sea anemone', 'French horn', 'airliner', 'brown bear'." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 491, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d4d3509e74a0afee2ec14a636adfc58791fef98b47e496f1499d49e77d090a3.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of attributions produced by different methods, using the ViT-S model w.r.t. the classes (top to bottom): 'American aligator', 'spiny lobster', 'black and gold garden spider', 'promontory'." + ], + "image_footnote": [], + "bbox": [ + 93, + 314, + 491, + 479 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/76af45be9a434eb5bc482e68f05f7c68cbd7525042adf56306a5aaf55495384f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)7.3169.25.0748.8361.8943.7675.6379.09
SLOC7.3263.565.0944.5856.2539.574.9677.84
\\( SLOC_{xp} \\)7.6163.235.2243.7255.6538.4873.6977.59
DIX10.1948.666.7431.7338.4724.9964.5168.66
EP14.0655.969.2636.5541.8927.2968.4372.04
GAE11.247.47.4730.9236.2023.4464.2267.95
LTX12.5649.288.232.2636.7124.0656.5563.16
RISE15.6753.2110.2635.3537.5325.0966.9170.5
T-Attr10.3747.566.8131.3537.1924.5465.6269.53
", + "bbox": [ + 94, + 561, + 496, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using ViT-S. The VOC dataset contains multi-object images, allowing multiple classes to appear within the same image. These results demonstrate the superior performance of all SLOC variants not only on single-object images but also on images containing multiple objects from different classes. Results for the DN model are provided in Tab. 9 in Appendix E.", + "bbox": [ + 89, + 733, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tables 4 and 5 present the overall summarized performance scores (higher is better) on the FB benchmark for the RN and ViT-B models, respectively. Detailed fine-grained results are provided in Appendix E. Notably, SLOC", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f85b8bbd97eb43a97563c7a4468fc84af137ba9d3c5b2179165c9c874b96948.jpg", + "table_caption": [ + "Table 3. Faithfulness results for combinations of method and metric, using the ViT-S model on the VOC dataset." + ], + "table_footnote": [], + "table_body": "
MethodSLOCACDIXEPFGGCGC++GIGIGLCRISE
Score0.780.700.720.730.700.720.720.560.630.720.62
", + "bbox": [ + 517, + 87, + 937, + 111 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d945241084ec291498353bfb3bb4d01880017fe4f2d4cc7d885c21da0d9cf475.jpg", + "table_caption": [ + "Table 4. FunnyBirds results for the RN model." + ], + "table_footnote": [], + "table_body": "
MethodSLOCDIXEPRISET-Attr
Score0.880.870.790.770.87
", + "bbox": [ + 570, + 143, + 849, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5ba4108174f1d0242e0c781bd1515ca4918820960b7864d9b3f2b57441925d2.jpg", + "table_caption": [ + "Table 5. FunnyBirds results for the ViT-B model." + ], + "table_footnote": [], + "table_body": "
MethodSLOCACDIXEPGCGC++GIGIGLCLTXRISE
mIoU↑0.560.550.660.520.550.560.510.480.550.560.51
mAP↑0.800.860.840.760.850.850.780.760.850.830.79
PA↑0.770.720.820.720.730.730.740.790.730.510.7
", + "bbox": [ + 517, + 203, + 950, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Segmentation tests results for the RN model.", + "bbox": [ + 549, + 250, + 867, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieves state-of-the-art performance on the FB benchmark as well. The distinct evaluation protocol employed in FB offers a complementary perspective to the faithfulness evaluation conducted on the IN dataset, further reinforcing our confidence in the robustness of SLOC.", + "bbox": [ + 511, + 291, + 906, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents segmentation tests results on the IN-Seg dataset using the RN model. We observe that DIX achieves the best average performance across all metrics, while the runner-up varies depending on the specific metric. Overall, SLOC is found to be competitive, on average, with all runner-up methods. We note that higher segmentation accuracy may reflect a method's ability to facilitate strong object detection rather than to identify the most informative features that explain the model's prediction. In reality, the most explanatory features do not always cover the entire object; instead, they may correspond to a subset of features that are critical to the model's decision. Consequently, higher segmentation accuracy does not necessarily indicate superior explanatory value.", + "bbox": [ + 511, + 367, + 906, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to space limitation, ablation studies, runtime comparison, sanity checks [3], and additional qualitative analyses, are provided in Appendices F, H, J, and I, respectively. These experiments offer further insights into the effectiveness and efficiency of SLOC.", + "bbox": [ + 511, + 578, + 905, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 685, + 633, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduced SLOC, a novel and efficient explainability method. By rethinking completeness as a guiding principle promoted locally rather than as a strict global constraint, SLOC exhibits state-of-the-art performance across various explainability benchmarks. These findings suggest that completeness, when promoted in a soft and local manner, provides a robust foundation for generating high-quality attribution maps that closely align with the model's predictive behavior and human comprehension. Finally, discussions on the limitations of SLOC and potential avenues for future research are provided in Appendix M.", + "bbox": [ + 511, + 710, + 906, + 877 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 2", + "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2", + "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, pages 9505-9515, 2018. 1, 8, 12, 25, 26", + "[4] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International conference on machine learning, pages 435-451. PMLR, 2022. 1, 2", + "[5] Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. In Proceedings of the International Conference on Learning Representations, 2018. 2", + "[6] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10 (7):e0130140, 2015. 2", + "[7] Oren Barkan, Yonatan Fuchs, Avi Caciularu, and Noam Koenigstein. Explainable recommendations via attentive multi-persona collaborative filtering. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 468-473, 2020. 2", + "[8] Oren Barkan, Ori Katz, and Noam Koenigstein. Neural attentive multiview machines. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3357-3361. IEEE, 2020. 1", + "[9] Oren Barkan, Idan Rejwan, Avi Caciularu, and Noam Koenigstein. Bayesian hierarchical words representation learning. arXiv preprint arXiv:2004.07126, 2020. 1", + "[10] Oren Barkan, Omri Armstrong, Amir Hertz, Avi Caciularu, Ori Katz, Itzik Malkiel, and Noam Koenigstein. Gam: Explainable visual similarity and classification via gradient activation maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 68-77, 2021. 2", + "[11] Oren Barkan, Edan Hauon, Avi Caciularu, Ori Katz, Itzik Malkiel, Omri Armstrong, and Noam Koenigstein. Gradsam: Explaining transformers via gradient self-attention maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 2882–2887, 2021. 2", + "[12] Oren Barkan, Yuval Asher, Amit Eshel, Yehonatan Elisha, and Noam Koenigstein. Learning to explain: A model-agnostic framework for explaining black box mod" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "els. In 2023 IEEE International Conference on Data Mining (ICDM), pages 944-949. IEEE, 2023. 2, 5, 7, 15, 23", + "[13] Oren Barkan, Yehonatan Elisha, Yuval Asher, Amit Eshel, and Noam Koenigstein. Visual explanations via iterated integrated attributions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2073-2084, 2023. 2", + "[14] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Deep integrated explanations. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 57-67, 2023. 2, 7, 15", + "[15] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Stochastic integrated explanations for vision models. In 2023 IEEE International Conference on Data Mining (ICDM), pages 938-943. IEEE, 2023. 2", + "[16] Oren Barkan, Tom Shaked, Yonatan Fuchs, and Noam Koenigstein. Modeling users' heterogeneous taste with diversified attentive user profiles. User Modeling and User-Adapted Interaction, pages 1–31, 2023. 2", + "[17] Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, and Noam Koenigstein. A counterfactual framework for learning and evaluating explanations for recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3723-3733, 2024. 2", + "[18] Oren Barkan, Yehonatan Elisha, Yonatan Toib, Jonathan Weill, and Noam Koenigstein. Improving llm attributions with randomized path-integration. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9430-9446, 2024. 2", + "[19] Oren Barkan, Yonatan Toib, Yehonatan Elisha, and Noam Koenigstein. A learning-based approach for explaining language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 98-108, 2024. 2", + "[20] Oren Barkan, Yonatan Toib, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Llm explainability via attributive masking learning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9522-9537, 2024. 2", + "[21] Oren Barkan, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Bee: Metric-adapted explanations via baseline exploration-exploitation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1835-1843, 2025. 3, 6", + "[22] Alexander Binder, Gregoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In International Conference on Artificial Neural Networks, pages 63-71. Springer, 2016. 1", + "[23] Alexander Binder, Leander Weber, Sebastian Lapuschkin, Grégoire Montavon, Klaus-Robert Müller, and Wojciech Samek. Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16143-16152, 2023. 6, 25" + ], + "bbox": [ + 517, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19802", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2", + "[25] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1", + "[26] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847. IEEE, 2018. 2, 7, 12, 14", + "[27] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7, 15", + "[28] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 2, 6, 7, 12, 15", + "[29] Jianbo Chen, Le Song, Martin J. Wainwright, and Michael I. Jordan. L-shapley and c-shapley: Efficient model interpretation for structured data. In International Conference on Learning Representations, 2019. 2", + "[30] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In Advances in Neural Information Processing Systems, pages 6970-6979, 2017. 2, 13", + "[31] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Computer Vision and Pattern Recognition (CVPR), 2009. 6, 25", + "[32] Saurabh Satish Desai and H. G. Ramaswamy. Ablationcam: Visual explanations for deep convolutional network via gradient-free localization. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 972-980, 2020. 1, 7, 15", + "[33] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[34] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1", + "[35] Yehonatan Elisha, Oren Barkan, and Noam Koenigstein. Probabilistic path integration with mixture of baseline distributions. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 570-580, 2024. 6", + "[36] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving performance of deep" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning models with axiomatic attribution priors and expected gradients. Nature machine intelligence, 3(7):620-631, 2021. 3, 30", + "[37] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 6", + "[38] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2950-2958, 2019. 2, 7, 15", + "[39] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3429-3437, 2017. 1, 2, 7, 15, 23", + "[40] Keren Gaiger, Oren Barkan, Shir Tsipory-Samuel, and Noam Koenigstein. Not all memories created equal: Dynamic user representations for collaborative filtering. IEEE Access, 11: 34746-34763, 2023. 2", + "[41] Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR), 51(5):1-42, 2018. 1", + "[42] Matthieu Guillaumin, Daniel Kuttel, and Vittorio Ferrari. Imagenet auto-annotation with segmentation propagation. International Journal of Computer Vision, 110(3):328-348, 2014. 6", + "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[44] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods, 2023. 6, 13, 14, 16", + "[45] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems, 32, 2019. 6, 7", + "[46] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6", + "[47] Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4198-4205, 2020. 6", + "[48] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 7, 15", + "[49] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4948-4957, 2019. 6, 12, 13" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050-5058, 2021. 2, 6, 7, 14", + "[51] Sunnie SY Kim, Nicole Meister, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision, pages 280-298. Springer, 2022. 6", + "[52] D Liu, R Nicoescu, and R Klette. Stereo-based bokeh effects for photography. Machine Vision and Applications, pages 1-13, 2016. 13", + "[53] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pages 4765-4774, 2017. 2", + "[54] Itzik Malkiel, Dvir Ginzburg, Oren Barkan, Avi Caciularu, Jonathan Weill, and Noam Koenigstein. Interpreting bert-based text similarity via activation and saliency maps. In Proceedings of the ACM Web Conference 2022, pages 3259-3268, 2022. 2", + "[55] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern Recognition, 65:211-222, 2017. 1", + "[56] Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, Jörg Schlötterer, Maurice Van Keulen, and Christin Seifert. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai. ACM Computing Surveys, 55(13s): 1-42, 2023. 7", + "[57] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2, 3, 6, 7, 12, 15, 16, 24", + "[58] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 983-991, 2020. 2", + "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 7, 14", + "[60] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3145-3153. JMLR.org, 2017. 1, 2", + "[61] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. 25", + "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2", + "[63] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. In NeurIPS, 2019. 2, 7, 14", + "[64] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, pages 3319-3328, 2017. 7, 14", + "[65] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319-3328. JMLR.org, 2017. 1, 2, 3, 30", + "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1", + "[67] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, 2019. 2", + "[68] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 6", + "[69] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9680-9689, 2020. 6" + ], + "bbox": [ + 516, + 92, + 906, + 502 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19804", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_model.json b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_model.json new file mode 100644 index 0000000000000000000000000000000000000000..02c6c0ca330218c66cd492e585fbd0d89cd5f8e5 --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.2, + 0.131, + 0.798, + 0.153 + ], + "angle": 0, + "content": "Soft Local Completeness: Rethinking Completeness in XAI" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.181, + 0.415, + 0.198 + ], + "angle": 0, + "content": "Ziv Weiss Haddad* Oren Barkan*" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.208, + 0.352, + 0.224 + ], + "angle": 0, + "content": "The Open University" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.182, + 0.875, + 0.199 + ], + "angle": 0, + "content": "Yehonatan Elisha Noam Koenigstein" + }, + { + "type": "text", + "bbox": [ + 0.651, + 0.208, + 0.783, + 0.224 + ], + "angle": 0, + "content": "Tel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.233, + 0.651, + 0.248 + ], + "angle": 0, + "content": "https://github.com/xaisloc/sloc" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.284, + 0.327, + 0.299 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Completeness is a widely discussed property in explainability research, requiring that the attributions sum to the model's response to the input. While completeness intuitively suggests that the model's prediction is \"completely explained\" by the attributions, its global formulation alone is insufficient to ensure faithful explanations. We contend that promoting completeness locally within attribution subregions, in a soft manner, can serve as a standalone guiding principle for producing faithful attributions. To this end, we introduce the concept of the completeness gap as a flexible measure of completeness and propose an optimization procedure that minimizes this gap across subregions within the attribution map. Extensive evaluations across various model architectures demonstrate that our method produces state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.222, + 0.59 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.782 + ], + "angle": 0, + "content": "In recent years, Explainable AI (XAI) has emerged as a critical aspect of machine learning, especially for deep learning models, which are often regarded as \"black boxes\" due to their lack of transparency [8, 9, 25, 41, 66]. XAI aims to provide insights into how these models make predictions, offering a way to ensure their reliability and trustworthiness [34]. A common approach to model explainability is through attribution maps, which assign importance scores to individual input features based on their contribution to the model's prediction [22, 27, 65]. These attributions serve as a tool for understanding which aspects of the input were most influential in determining the output." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.875 + ], + "angle": 0, + "content": "A widely discussed notion in XAI is completeness (also known as conservation or efficiency) [4, 55, 60, 65]. Completeness is defined as the requirement that the elements in an attribution map should sum to the difference between the model's prediction for a given input and a baseline (e.g., the null representation). This difference is also known as the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.906, + 0.42 + ], + "angle": 0, + "content": "model's response. When completeness is achieved inherently by an attribution method, it is intuitively appealing, as it suggests that the attribution map fully captures the factors contributing to the model's response. However, even a low-quality attribution map can be superficially adjusted to satisfy completeness through post hoc global normalization. Yet, such an approach lacks genuine explanatory power, as the normalization is decoupled from the underlying attribution mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.591 + ], + "angle": 0, + "content": "This paper recognizes that completeness, as defined in the literature as a global property that an explanation method should satisfy, is too weak to serve as a standalone criterion. Indeed, several well-regarded explanation methods [32, 39, 59] do not inherently satisfy completeness. Nevertheless, these methods have demonstrated considerable effectiveness across diverse objective evaluation metrics, providing faithful explanations and successfully passing various sanity checks [3]. This suggests that completeness is neither a necessary nor sufficient condition for generating high-quality explanations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.852 + ], + "angle": 0, + "content": "In this work, we differentiate between global completeness, as originally defined in the literature, which imposes a completeness constraint on the entire attribution map, and local completeness, which requires completeness within individual subregions of the attribution map. By rethinking completeness as a local and flexible guiding measure rather than a strict global constraint, we seek to overcome its limitations while harnessing the appealing motivation behind it for producing meaningful and faithful explanations. To achieve this, we require subregions of the attribution map, referred to as sub-maps, to promote completeness locally by accounting for the model's response to the corresponding subregions of the input image, yet in soft manner. An attribution map that adheres to this guiding principle should assign high (low) importance to image subregions where the model exhibits a strong (weak) response, thereby maintaining faithfulness to the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unlike global completeness, which is a relatively weak requirement that can be easily satisfied, achieving local completeness across an arbitrary set of sub-maps may not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "always be feasible. To address this challenge, we introduce the completeness gap - a measure that quantifies the deviation of a sub-map from completeness. Specifically, the completeness gap measures the difference between the sum of elements in each sub-map and the model's response to the corresponding subregions of the input image. Rather than imposing a strict binary constraint, the completeness gap serves as a soft, quantifiable criterion for local completeness. We argue that minimizing this gap over a diverse set of sub-maps leads to faithful explanations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.332 + ], + "angle": 0, + "content": "To this end, we present Soft LOcal Completeness (SLOC) - a novel explainability method that promotes completeness locally within sub-maps of the attribution map in a soft and flexible manner. SLOC accomplishes this by seeking to minimize the completeness gap locally for each individual sub-map, simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.482, + 0.484 + ], + "angle": 0, + "content": "The motivation for minimizing the completeness gap locally is to emphasize or attenuate sub-maps based on the actual impact of their corresponding input subregions on the model's output. SLOC achieves this through a gradient-based optimization process, where each sub-map is adjusted to achieve local completeness in a soft manner. Therefore, SLOC's goal is not to enforce strict completeness, globally, but to use it as a guiding principle for refining a large set of sub-maps (subregions within the attribution), iteratively and simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.482, + 0.576 + ], + "angle": 0, + "content": "Our contribution is the introduction of the SLOC method, which facilitates a novel optimization procedure to promote completeness in a soft local manner. The effectiveness of SLOC is demonstrated through extensive experiments on various model architectures, where it is shown to produce state-of-the-art results across multiple benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.233, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.482, + 0.825 + ], + "angle": 0, + "content": "The XAI literature encompasses a broad range of approaches for attributing model predictions to specific input features across different tasks [2, 4, 6, 7, 16, 24, 29, 40, 54]. Early gradient-based methods produce explanation maps by leveraging gradients directly [5, 60, 62] or through functions that combine class activation maps with or without their gradients [10, 26, 48, 58, 59, 63]. With the rise of transformer architectures, new explainability techniques for transformers have emerged [1, 4, 11, 67]. For example, Transformer Attribution [28] introduces a class-specific Deep Taylor Decomposition, applying relevance propagation for both positive and negative attributions. Building on this, Generic Attention Explainability [27] generalizes Transformer Attribution to explain bi-modal transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Path integration methods form another prominent family of attribution techniques [13, 15, 18, 50, 65]. Integrated Gradients (IG) [65] computes attributions by integrating gradients along an interpolation path from a baseline to the input. A refined variant, Guided Integrated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Gradients (GIG) [50], uses an adaptive path that avoids high-gradient regions, thereby reducing irrelevant attributions. Recently, Deep Integrated Explanations (DIX) [14] proposed performing integration over intermediate network representations instead of inputs, producing more faithful attribution maps than IG and GIG." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.905, + 0.529 + ], + "angle": 0, + "content": "Perturbation-based (occlusion) techniques generate attribution maps by perturbing parts of the input to assess the contributions of specific elements to the prediction [17, 19, 20, 53]. For example, RISE [57], creates perturbations by masking areas in the image through the upsampling of randomly drawn low-resolution binary grids. The class score corresponding to each masked version of the image serves as an importance score for that specific mask. A linear combination of all masks, weighted by their importance, then forms the final attribution map. Another branch of perturbation methods are learning-based perturbation techniques [12, 30, 38, 39]. For example, Meaningful Perturbation (MP) [39] learns to produce perturbations by masking the smallest region possible while significantly altering the original prediction, isolating the minimum necessary content for the prediction. Alternatively, LTX [12] introduces a surrogate 'explainer' model pretrained to mask as much of the input as possible while preserving the original prediction, thereby ensuring the retained features are those most relevant to the model's prediction. Then, LTX finetunes the attribution per specific example, allowing the selection of the best-performing attribution w.r.t. to the metric at hand." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.905, + 0.773 + ], + "angle": 0, + "content": "Our SLOC method falls within the family of learning-based perturbation techniques, yet it diverges from prior works in several key aspects. First, conceptually, SLOC presents a distinct methodology by promoting local completeness rather than attempting to maximize or minimize the model prediction through learned perturbations [12, 30, 39]. Consequently, SLOC does not require gradient backpropagation through the model, and incorporates the model's predictions on the masked inputs into the loss as constant nodes within the computation graph. This design yields a simple, computationally efficient gradient expression, and relies solely on the model's forward pass, hence operating in a true black-box setting. Finally, unlike LTX, SLOC avoids the overhead of learning an 'explainer' function, which would also require a dataset aligned with the training data distribution of the model being explained." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.744, + 0.807 + ], + "angle": 0, + "content": "3. Soft Local Completeness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.814, + 0.905, + 0.876 + ], + "angle": 0, + "content": "Let \\( f: \\mathbb{R}^n \\to [0,1]^C \\) be a classification model that takes an input image \\( \\mathbf{x} \\in \\mathbb{R}^n \\) and outputs a discrete probability distribution \\( f(\\mathbf{x}) \\in [0,1]^C \\) over \\( C \\) classes. Given a target class \\( y \\) to explain, our goal is to produce an attribution map" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.785, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)W.l.o.g, we represent images as vectors in \\(\\mathbb{R}^n\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19795" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "\\(\\mathbf{a}_{\\mathbf{x}}^{y} \\in \\mathbb{R}^{n}\\) that attributes the contribution of each element \\(i\\) in \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\) to the prediction score \\(f(\\mathbf{x})[y]\\) for the class \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.182 + ], + "angle": 0, + "content": "Completeness requires that the attributions in \\( \\mathbf{a}_{\\mathbf{x}}^{y} \\) collectively account for the difference between the model's prediction on the input \\( \\mathbf{x} \\) and the baseline representation \\( \\mathbf{b} \\), which is set to the black image in this work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.429 + ], + "angle": 0, + "content": "The model's response to \\(\\mathbf{x}\\) is defined as \\(r(\\mathbf{x})\\coloneqq f(\\mathbf{x}) - f(\\mathbf{b})\\). Formally, the attribution map \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\) satisfies completeness if \\(\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i] = r(\\mathbf{x})[y]\\). We note that this property can be superficially imposed on an attribution map through a post-processing step by normalizing each element using the factor \\(\\frac{r(\\mathbf{x})[y]}{\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i]}\\). However, artificially satisfying completeness through such normalization does not inherently yield a meaningful explanation, as the imposed completeness is an external adjustment rather than an intrinsic characteristic of the explanation method itself. For example, even a flat attribution map (where all elements are equally weighted) could be normalized to satisfy the completeness criterion, yet such a map lacks any explanatory value. Consequently, relying solely on completeness, in its global form, is deemed insufficient for producing meaningful attribution maps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.214, + 0.453 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Our SLOC approach is motivated by the toy example illustrated in Fig. 1, which depicts an image where the key region influencing the model's prediction for the 'great-grey-owl' class is approximately centered. Figures 1(a)-(d) present four images: the original image (a) and three perturbed versions obtained by masking different regions (b)-(d)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.687 + ], + "angle": 0, + "content": "The four different masks applied in Figures 1(a)-(d) are illustrated in Figure 1(e). Each mask selectively perturbs the original image by replacing certain regions with the baseline (in this work, the null pixels). For instance, the mask corresponding to Figure 1(c), shown in the top-right corner of Figure 1(e), applies a perturbation by replacing all pixels within the black rectangle (located at the bottom-right corner) with the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Notably, the model's response for the great-grey-owl class remains consistent across all four images (a-d), with a value of 0.8. Recall that the model response is defined as the difference between the model's class probability for a given image and that for the baseline (the null image)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.794 + ], + "angle": 0, + "content": "As defined earlier, a sub-map represents a subset of elements in the attribution map that are associated with spe" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.093, + 0.64, + 0.105 + ], + "angle": 0, + "content": "(a) Original image: 0.8" + }, + { + "type": "text", + "bbox": [ + 0.651, + 0.093, + 0.768, + 0.104 + ], + "angle": 0, + "content": "(c) Masked Image: 0.8" + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.093, + 0.856, + 0.104 + ], + "angle": 0, + "content": "(e) Four masks" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.108, + 0.646, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.185, + 0.64, + 0.196 + ], + "angle": 0, + "content": "(b) Masked Image: 0.8" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.108, + 0.772, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.185, + 0.766, + 0.195 + ], + "angle": 0, + "content": "(d) Masked Image: 0.8" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.108, + 0.898, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.185, + 0.866, + 0.196 + ], + "angle": 0, + "content": "(f) Attribution Map" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.199, + 0.647, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.199, + 0.772, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.199, + 0.898, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 1. Subfigures (a)-(d) present the model's response (0.8) for the class 'great-grey-owl' when applied to the original input (a) and its masked versions (b)-(d). The four masks used to generate (a)-(d) are depicted in (e). A sub-map refers to a subset of elements in the attribution map that correspond to specific subregions in the input. We consider four distinct sub-maps (not shown), each corresponding to the non-masked (visible) subregions in (a)-(d). For instance, the sub-map associated with (c) includes all elements in the attribution map except those corresponding to the black rectangle in the bottom-right corner of (c). This exclusion aligns with the masked region in the input, effectively determining which attributions are included in that sub-map. A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by (c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to (c). By enforcing local completeness across the four sub-maps induced by (a)-(d), the remaining central stripe, which is unmasked in all images (a)-(d), must account for the model's entire response of 0.8. This results in the attribution depicted in (f) that highlights the central stripe containing the owl in red. See 3.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.806 + ], + "angle": 0, + "content": "cific subregions in the input. In the context of Fig. 1, we consider four different sub-maps, each corresponding to the non-masked (visible) subregions in Figs.1(a)-(d). For instance, the sub-map corresponding to Fig. 1(c) consists of all elements in the attribution map except those associated with the black rectangle in the bottom-right corner of Fig.1(c). This exclusion aligns with the masked region in the input, effectively defining which attributions are considered within that sub-map. Notably, the sub-map corresponding to the original image (Fig. 1(a)) is equivalent to the entire attribution map, as no regions are masked. This formulation allows us to analyze how different subregions contribute to completeness by evaluating attributions over various sub-maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by Fig. 1(c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to Fig. 1(c)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.804, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2The baseline representation should resemble missing information or the neutral representation. While determining the ultimate representation for the baseline is an open question and an active research field in explainability [21, 36], in this work, we opt for the simple choice of a null baseline represented by the black image. This choice is on par with other notable works [57, 65]. Additionally, we conducted experiments using alternative baselines and observed that the resulting trends remained consistent with those obtained using the black image." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Our key insight is that enforcing local completeness across the four sub-maps induced by Figs.1(a)-(d) allows us to pinpoint the important region of the image w.r.t. the model's prediction. Specifically, the intersection of these four sub-maps forms a central stripe around the owl, illustrated in red in Fig.1(f). By evenly distributing the total attribution sum of 0.8 across the pixels within this intersection (red stripe) and assigning 0 elsewhere (highlighted in dark blue in Fig. 1(f)), we obtain an attribution map that ensures local completeness for all four sub-maps. Since each sub-map contains the intersection while the attribution values outside it are zero, the sum of attributions within each sub-map is exactly the sum of attributions within the intersection: 0.8. This resulting attribution map effectively captures the key region containing the owl." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.577 + ], + "angle": 0, + "content": "This simple example can be slightly generalized under similar settings. We make the following assumptions: (1) The model response for a subregion is \\( R \\) if the subregion contains the object of interest and 0 otherwise. (2) Each sub-map either fully contains the object of interest or does not intersect with it at all. (3) One of the sub-maps corresponds to the complete attribution map. (4) All attribution values are non-negative. It can then be shown that if an attribution map satisfies local completeness for all sub-maps, the intersection defined by these sub-maps is the largest subregion that can have non-zero attribution. Simply put, this intersection highlights the important regions in the image responsible for the model's prediction. Moreover, as the number of sub-maps increases, the important regions can potentially become more precisely defined. The formalization and proof follow straightforwardly and are provided in Appendix K." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Keep in mind, however, that this setting and its assumptions are simplistic and unrealistic, serving only to illustrate the motivation. Therefore, SLOC does not actually rely on these assumptions. In more realistic scenarios, satisfying local completeness for all sub-maps is often infeasible (an example is provided in Fig. 11 in Appendix I). To address this challenge, we introduce the completeness gap, which quantifies the degree to which a sub-map deviates from local completeness. Our approach seeks to minimize the completeness gap across multiple sub-maps simultaneously, hence promoting completeness locally within subregions of the attribution map, in a soft manner. We formulate this as an optimization problem and employ gradient descent to find a solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Finally, recall that in our simple example, in the absence of additional information, we distributed the attribution uniformly across the central red stripe. When introducing the loss terms for the SLOC optimization, we will also discuss regularization terms. One such term is the Total Variation loss, which encourages spatial smoothness in the attribution. As we demonstrate in Sec. 4, by encouraging this soft" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "form of local completeness, the resulting attribution map offers robust explanatory performance, both visually and according to a variety of evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.704, + 0.162 + ], + "angle": 0, + "content": "3.2. SLOC optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.907, + 0.198 + ], + "angle": 0, + "content": "Let \\(\\mathbf{m} \\in \\{0,1\\}^n\\) be a binary mask. We define the masked input by" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.198, + 0.907, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\mathbf {m}} = \\mathbf {x} \\circ \\mathbf {m} + (1 - \\mathbf {m}) \\circ \\mathbf {b}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.906, + 0.28 + ], + "angle": 0, + "content": "where \\( \\circ \\) stands for the elementwise product. In words, \\( \\mathbf{x}^{\\mathbf{m}} \\) is a perturbation (masked version) of \\( \\mathbf{x} \\), according to the mask \\( \\mathbf{m} \\), where all masked elements of \\( \\mathbf{x}^{\\mathbf{m}} \\) are replaced with their respective elements from the baseline \\( \\mathbf{b} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.906, + 0.401 + ], + "angle": 0, + "content": "In this work, we consider binary masks constructed over patches, where each patch is randomly assigned a Bernoulli outcome with probability \\( p \\). We explore two methodologies for automatically selecting \\( p \\): one that determines \\( p \\) dynamically per input image, and another that uses a fixed value per model. The reader is referred to Appendix D for the exact details of the masks construction process, including the methodologies used for determining \\( p \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.906, + 0.432 + ], + "angle": 0, + "content": "Given an initial attribution map \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\) and a set of masks \\(\\mathcal{M} \\subset \\{0,1\\}^n\\), the SLOC loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.44, + 0.905, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}; \\mathcal {M}\\right) = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} \\underbrace {(r \\left(\\mathbf {x} ^ {\\mathbf {m}}\\right) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {c o m p l e t e n e s s g a p}} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.648 + ], + "angle": 0, + "content": "where \\(\\cdot\\) is the dot-product operator, and \\(|\\cdot|\\) denotes the L1 norm. As observed in Eq. 2, \\(\\mathcal{L}_c\\) promotes the minimization of the completeness gap, a soft measure of completeness, locally, for all sub-maps induced by \\(\\mathcal{M}\\). The rationale behind the local minimization of the completeness gap is that attributions within a sub-map should be strengthened or reduced so that their sum closely matches \\(r(\\mathbf{x}^{\\mathbf{m}})[y]\\) - the model's response to the respective regions exposed in \\(\\mathbf{x}^{\\mathbf{m}}\\). This behavior is achieved by applying gradient descent on \\(\\mathcal{L}_c\\) with respect to \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.678 + ], + "angle": 0, + "content": "To better understand the optimization process, one can inspect the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.686, + 0.905, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y}} = - \\frac {1}{| \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) \\mathbf {m}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For simplicity, assume \\(\\mathcal{M}\\) contains a single mask \\(\\mathbf{m}\\). Then, for each element \\(\\mathbf{a}_{\\mathbf{x}}^{y}[i]\\) that is exposed in \\(\\mathbf{x}^{\\mathbf{m}}\\) (i.e., \\(\\mathbf{m}[i] = 1\\)), the gradient is the negative of the completeness gap normalized by the number of elements in the sub-map, i.e., \\(\\frac{\\partial\\mathcal{L}_c}{\\partial\\mathbf{a}_x^y[i]} = -\\frac{1}{|\\mathbf{m}|} (r(\\mathbf{x}^{\\mathbf{m}})[y] - \\mathbf{a}_{\\mathbf{x}}^y\\cdot \\mathbf{m})\\), and for each \\(i\\) where \\(\\mathbf{m}[i] = 0\\), the gradient vanishes. Accordingly, the gradient descent update ensures an equal increase (decrease) for all elements in the sub-map when the sum of the attributions falls short of (exceeds) the model's response \\(r(\\mathbf{x}^{\\mathbf{m}})[y]\\). Assuming a learning rate of 1, it follows that after the gradient descent update, the completeness gap vanishes, resulting in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "an updated sub-map satisfying completeness. This can be seen mathematically as follows: for a sub-map induced by the mask \\(\\mathbf{m}\\), the sum of its elements after the gradient update is given by" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.162, + 0.473, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] - \\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ]} \\\\ = \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] + \\underbrace {\\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {i n d e p e n d e n t o f} i} \\\\ = | \\mathbf {m} | \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) + \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] \\\\ = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} + \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.485, + 0.478 + ], + "angle": 0, + "content": "where the second transition follows from the fact that the normalized completeness gap is independent of \\(i\\) and is therefore summed \\(|\\mathbf{m}|\\) times, as the summation is taken over every \\(i\\) where \\(\\mathbf{m}[i] = 1\\), with \\(\\mathbf{m}\\) being a binary tensor. The penultimate transition then arises from expressing the sum of elements in the sub-map as the dot product, \\(\\mathbf{a}_{\\mathbf{x}}^{y} \\cdot \\mathbf{m}\\). This derivation justifies the normalization of the squared completeness gap by \\(|\\mathbf{m}|\\) in Eq. 2, as it ensures that after the gradient update, the sub-map satisfies local completeness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.75 + ], + "angle": 0, + "content": "To give further intuition for using the normalized completeness gap in Eq. 2, consider two different masks: one that exposes a single pixel (element), and another that exposes a large portion of the input. Given that the gradient update of elements in the sub-map is equal for all elements, it makes sense to distribute the completeness gap among all elements. In the case of a mask exposing a single pixel, we have full confidence that this specific element is responsible for the model's response \\( r(\\mathbf{x}^{\\mathrm{m}})[y] \\). Therefore, it is reasonable to expect that the gradient update for this element will match the size of the completeness gap. However, in the case of a sub-map associated with a mask exposing large portions of the input, we cannot differentiate the accountability of individual elements within the exposed portions for the model's response. Thus, without further information, it is reasonable to apply an update step that evenly distributes the completeness gap across all exposed elements. Hence, the squared completeness gap is divided by \\( |\\mathbf{m}| \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "So far, the discussion and analysis have been limited to the case of a single mask, i.e., \\( |\\mathcal{M}| = 1 \\). When considering a set of masks, the update for \\( \\mathbf{a}_{\\mathbf{x}}^{y}[i] \\) is determined by the accumulation of gradients from all the sub-maps it is associated with. Specifically, sub-maps that include the \\( i \\)-th element and whose total sum of elements is below (exceeds) the model's response \\( r(\\mathbf{x}^{\\mathrm{m}})[y] \\) will contribute to the intensification (reduction) of \\( \\mathbf{a}_{\\mathbf{x}}^{y}[i] \\), with the contribution from each individual mask weighted according to the normalized completeness gap it induces." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.093, + 0.757, + 0.105 + ], + "angle": 0, + "content": "Masks & Responses Generation Phase" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.106, + 0.905, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.34, + 0.908, + 0.534 + ], + "angle": 0, + "content": "Figure 2. SLOC overview: The upper section illustrates the Masks & Responses Generation Phase. In this phase, (1) \\(\\mathcal{M}\\) masks are randomly generated. (2) Perturbations are produced by combining the original image and baseline according to the masks, as defined in Eq. (1) (3) These perturbations are passed as inputs to the model, and the corresponding outputs are obtained. (4) Subtracting the model's output for the baseline from these outputs produces the model responses. (5) These masks and their corresponding responses are then passed to the Optimization Phase. (6) Starting with an initial attribution. (7) Gradient descent is used to iteratively update the attribution. The loss function \\(\\mathcal{L}\\) (Eq. 4) consists of three components: the completeness gap (computed using the masks, responses, and attributions), as well as TV and L1 regularization. See Sec. 3.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.634 + ], + "angle": 0, + "content": "To promote smooth and focused attribution maps, we combine \\(\\mathcal{L}_c\\) with two types of regularization applied to the attribution map: Total Variation (TV) and L1 regularization that encourage smoothness and sparsity, respectively. The final loss function is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.642, + 0.907, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) = \\mathcal {L} _ {c} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) + \\lambda_ {1} | \\mathbf {a _ {x} ^ {y}} | + \\lambda_ {2} \\mathrm {T V} (\\mathbf {a _ {x} ^ {y}}), (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.668, + 0.55, + 0.681 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.689, + 0.929, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T V} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}\\right) := \\sum_ {i, j} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i + 1, j ]\\right) ^ {2} + \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j + 1 ]\\right) ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\) in Eq. 5 is reshaped to match the spatial dimensions of the final attribution map. Additionally, \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters controlling the strength of the regularization terms. In our experiments, these hyperparameters are set based on a separate validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The optimization of \\(\\mathbf{a}_{\\mathbf{x}}^{y}\\) is carried out via gradient descent on \\(\\mathcal{L}\\). The optimization process allows for tracking the performance of the metric during the optimization and selecting the best-performing attribution map based on this metric from the resulting maps [12]. A diagram of the SLOC method is depicted in Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.223 + ], + "angle": 0, + "content": "Figure 3. SLOC attribution maps across training steps. Faithful explanations emerge after a few hundred gradient updates, with an appropriate learning rate decay." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.249, + 0.483, + 0.369 + ], + "angle": 0, + "content": "In practice, we observed that a few hundred gradient updates, combined with an appropriate learning rate decay, were sufficient to produce faithful and meaningful explanations. Therefore, in our experiments, we report SLOC results for \\( T = 500 \\). An example for SLOC's attribution across gradient update steps is shown in Fig. 3. For the exact implementation details and hyperparameter configuration, the reader is referred to Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.37, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Due to space limitations, Appendix G provides a detailed discussion of SLOC's computational complexity. Appendix H presents a runtime analysis, comparing SLOC against various methods and demonstrating its efficiency over other state-of-the-art learning-based perturbation methods. Additionally, Appendix L examines additional axioms and explains which of them are satisfied by SLOC." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Finally, it is important to clarify that the SLOC optimization procedure does not guarantee convergence to an attribution map that fully satisfies completeness. Instead, it leverages the completeness gap as a guiding measure to update subregions within the attribution map, thereby promoting local completeness for these subregions in a soft manner. Nevertheless, as demonstrated in Sec. 4, SLOC has proven effective in generating attribution maps that outperform state-of-the-art explanation methods." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.64, + 0.325, + 0.657 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.664, + 0.279, + 0.681 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.686, + 0.483, + 0.837 + ], + "angle": 0, + "content": "The experiments were conducted on an NVIDIA DGX 8xA100 Server, utilizing the PyTorch package. As explanations are multifaceted by nature, no single evaluation metric can fully capture their quality, and no definitive metric exists [21, 23, 35, 45, 47, 51, 68]. Therefore, our evaluation encompasses several distinct protocols designed to assess explanation quality from multiple complementary perspectives, on different datasets, and across four model architectures: ResNet50 (RN) [43], DenseNet201 (DN) [46], ViT-Base (ViT-B) and ViT-Small (ViT-S) [33]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experiment 1: Faithfulness evaluation This evaluation assesses the faithfulness (also known as correctness or fidelity) of the explanation via a set of perturbation-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.455 + ], + "angle": 0, + "content": "metrics. These metrics are designed to reveal the actual input elements the model relies on for its predictions. Following prior works [28, 49, 50, 69], we conducted an extensive evaluation using a comprehensive set of perturbation-based explanation metrics to assess the faithfulness of the generated explanations: the Area Under the Curve (AUC) of Positive (POS) and Negative (NEG) perturbation tests [28], AUC of the Insertion (INS) and Deletion (DEL) tests [57], and AUC of the Softmax Information Curve (SIC) and Accuracy Information Curve (AIC) [49]. For POS and DEL, lower values indicate better performance, while for NEG, INS, SIC, and AIC, higher values are preferred. Since NEG and POS, as well as INS and DEL, are complementary metrics, we also report the NEG-POS Difference (NPD) and the INS-DEL Difference (IDD) as single summaries of their respective complementary performances. Faithfulness performance is reported on a sample from the ImageNet [31] ILSVRC 2012 (IN) validation set, which consists of 10,000 images across 1,000 classes. Since most images in the IN dataset contain a single, centrally focused object, we extend our evaluation to multi-object scenes by assessing explanation faithfulness on a sample from the Pascal VOC 2012 [37] (VOC) test set, which includes 1,000 images spanning 20 object classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Experiment 2: FunnyBirds evaluation While perturbation-based protocols are a standard approach for assessing explanation faithfulness, they are criticized for introducing domain shifts that undermine the validity of the evaluation [45]. To address this, we additionally report results on the FunnyBirds (FB) dataset which consists of 500 images from 50 classes, following the evaluation protocol from [44]. The FB evaluation employs controlled interventions to estimate ground-truth (GT) importance scores at the level of object parts rather than individual pixels, mitigating issues associated with pixelwise perturbation-based evaluations [68]. Furthermore, FB avoids domain shifts by incorporating semantically meaningful interventions during training. By evaluating explanations at the part level—closer to human visual understanding—FB reduces the gap between automatic XAI evaluation (e.g., faithfulness metrics) and human-centric studies, providing a more interpretable and robust assessment of the explanations. FB evaluates explainability in three aspects: Completeness, Correctness, and Contrastivity - and provides a combined overall score (higher is better)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Experiment 3: Segmentation evaluation To ensure a comprehensive comparison with prior works [27, 28, 48], we conduct an extensive evaluation using segmentation tests on the ImageNet-Seg dataset (IN-Seg) [42], a subset of the ImageNet validation set comprising 4,276 human-annotated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "ground-truth (GT) segmentation maps across 445 classes. As detailed in [28], these tests generalize the \"Pointing Game\" [45] and serve as a human-grounded evaluation [56] by measuring the alignment between GT segmentation maps and the attributions generated by the explanation method. While higher segmentation accuracy does not necessarily indicate greater explanatory power [57], segmentation tests assess the alignment between explanations and human-annotated GT, hence providing a complementary perspective to other evaluations. We follow the evaluation protocol from [28], assessing segmentation performance using Pixel Accuracy (PA), mean Intersection-over-Union (mIoU), and mean Average Precision (mAP), higher is better for all metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Detailed descriptions of the evaluation protocols and metrics used in Experiments 1, 2, and 3 are provided in Appendices B.1, B.2, and B.3, respectively." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.356, + 0.485, + 0.825 + ], + "angle": 0, + "content": "Evaluated methods We evaluate SLOC against a comprehensive suite of 14 explanation methods encompassing gradient-based, path-integration, perturbation, and CAM methods. For CNN models, we include the following methods: Grad-CAM (GC) [59], Grad-CAM++ \\((\\mathbf{GC} + +)\\) [26], Deep Integrated Explanations (DIX) [14], FullGrad (FG) [63], Ablation-CAM (AC) [32], Layer-CAM (LC) [48], Learning To Explain (LTX) [12], RISE [57], Meaningful Perturbation (MP) [39], Extremal Perturbations (EP) [38], Integrated Gradients (IG) [64], and Guided IG (GIG) [50]. For ViT models, we considered the following methods: Transformer Attribution (T-Attr) [28] and Generic Attention Explainability (GAE) [27], alongside DIX, RISE, LTX, EP, and MP, which are applicable to both CNN and ViT architectures. Hyperparameters for all methods were set according to the recommended settings published by the authors, unless a better configuration was found. A detailed description of all explanation methods is provided in Appendix C. Finally, we evaluated three different versions of our SLOC method on both CNN and ViT models: SLOC, which tunes \\(p\\) dynamically per input, \\(\\mathrm{SLOC}_{\\mathrm{xp}}\\), where \\(p\\) is calibrated per model (i.e., fixed across all inputs for a given model) based on a validation dataset, and \\(\\mathrm{SLOC}_{\\mathrm{m}}\\), which is identical to SLOC but additionally monitors the IDD metric on the specific input and selects the explanation with the best performance according to this metric. All versions were run with \\(T = 500\\). The precise implementation details, including the monitoring procedure, the \\(p\\) tuning procedures, optimization process, and all hyperparameter configurations, are provided in Appendix D and in our GitHub repository." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.185, + 0.849 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Tables 1 and 2 present a quantitative comparison of SLOC with other state-of-the-art explanation methods across multiple faithfulness metrics, for DN and ViT-S, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.925, + 0.263 + ], + "angle": 0, + "content": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)11.6176.549.3965.7964.9356.479.2178.35
SLOC11.5270.979.1260.5359.4551.4177.7977.72
\\( SLOC_{xp} \\)11.7870.209.3459.7558.4350.4177.3376.80
AC17.2467.6813.2757.1850.4443.9177.7875.41
DIX13.3662.9510.3652.4349.5942.0774.6271.47
EP16.1265.6812.9455.049.5542.0677.3874.97
FG19.0644.6615.2637.6225.622.3758.2353.86
GC16.8868.5413.0457.9551.6644.9178.3876.01
GC++17.3667.2913.3456.7549.9343.4178.0475.62
GIG14.8149.9612.2841.9335.1629.6561.1257.6
IG14.1451.9911.244.0837.8532.8861.2658.48
LC17.2867.2713.3156.7149.9943.477.9575.42
LTX16.2471.0912.9559.6954.8446.7478.9276.25
MP18.5453.2414.8743.7934.728.9266.9864.13
RISE18.4262.7514.2652.9944.3338.7376.8274.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.265, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Table 1. Faithfulness results for all combinations of method and metric, using the DN model on the IN dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.309, + 0.93, + 0.43 + ], + "angle": 0, + "content": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)14.8381.8112.2770.3166.9858.0483.7883.34
SLOC15.2577.9712.4766.5162.7254.0483.1982.88
\\( SLOC_{xp} \\)15.7977.8512.8566.1862.0653.3383.0681.97
DIX18.6968.1714.8656.8349.4841.9776.8875.05
EP27.3772.5221.6660.845.1439.1479.4577.56
GAE19.9866.9315.9355.7246.9539.7975.4273.9
LTX20.8468.516.6356.8947.6640.2774.2271.56
MP27.7263.2522.4652.2835.5329.8174.2271.25
RISE29.5169.5223.4757.9340.0234.4679.9377.23
TATTR19.0667.5214.9156.4148.4641.578.0375.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.434, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Table 2. Faithfulness results for all combinations of method and metric, using the ViT-S model on the IN dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Results for RN and ViT-B are available in Tabs. 7 and 8 in Appendix E, respectively. Overall, the results indicate that all SLOC versions are the best-performing methods, with LTX, DIX, RISE, and EP as runners-up, depending on the combination of architecture and evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Among the three versions of SLOC, \\( \\mathrm{SLOC}_{\\mathrm{m}} \\) performs best. This can be attributed to its use of the IDD metric, which measures the difference between INS and DEL performance. The fact that \\( \\mathrm{SLOC}_{\\mathrm{m}} \\) outperforms across the vast majority of metrics suggests that the IDD metric correlates well with other faithfulness metrics. SLOC and \\( \\mathrm{SLOC}_{\\mathrm{xp}} \\) follow, with SLOC performing slightly better. This suggests that tuning \\( p \\) per instance offers a modest advantage in metrics, at the cost of increased inference complexity. When a representative sample of data is available, \\( p \\) can be calibrated per model using this data. Otherwise, \\( p \\) should be adjusted dynamically per input during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Notably, both SLOC and \\(\\mathrm{SLOC}_{\\mathrm{xp}}\\) surpass learning-based approaches such as LTX, MP, and EP, achieving state-of-the-art results without requiring gradient backpropagation through the model (see the discussion on SLOC complexity in Appendix G) or metric monitoring (as in LTX). Figures 6 and 5 present comparative examples of the top-performing methods for DN and ViT-S. Arguably, SLOC produces focused explanation maps highlighting relevant class-discriminative features in the image." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Table 3 presents faithfulness results on the VOC dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.493, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.485, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of attributions produced by different methods, using the DN model w.r.t. the classes (top to bottom): 'sea anemone', 'French horn', 'airliner', 'brown bear'." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.315, + 0.493, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of attributions produced by different methods, using the ViT-S model w.r.t. the classes (top to bottom): 'American aligator', 'spiny lobster', 'black and gold garden spider', 'promontory'." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.563, + 0.498, + 0.674 + ], + "angle": 0, + "content": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)7.3169.25.0748.8361.8943.7675.6379.09
SLOC7.3263.565.0944.5856.2539.574.9677.84
\\( SLOC_{xp} \\)7.6163.235.2243.7255.6538.4873.6977.59
DIX10.1948.666.7431.7338.4724.9964.5168.66
EP14.0655.969.2636.5541.8927.2968.4372.04
GAE11.247.47.4730.9236.2023.4464.2267.95
LTX12.5649.288.232.2636.7124.0656.5563.16
RISE15.6753.2110.2635.3537.5325.0966.9170.5
T-Attr10.3747.566.8131.3537.1924.5465.6269.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Table 3. Faithfulness results for combinations of method and metric, using the ViT-S model on the VOC dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.839 + ], + "angle": 0, + "content": "using ViT-S. The VOC dataset contains multi-object images, allowing multiple classes to appear within the same image. These results demonstrate the superior performance of all SLOC variants not only on single-object images but also on images containing multiple objects from different classes. Results for the DN model are provided in Tab. 9 in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Tables 4 and 5 present the overall summarized performance scores (higher is better) on the FB benchmark for the RN and ViT-B models, respectively. Detailed fine-grained results are provided in Appendix E. Notably, SLOC" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.939, + 0.112 + ], + "angle": 0, + "content": "
MethodSLOCACDIXEPFGGCGC++GIGIGLCRISE
Score0.780.700.720.730.700.720.720.560.630.720.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.571, + 0.116, + 0.848, + 0.13 + ], + "angle": 0, + "content": "Table 4. FunnyBirds results for the RN model." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.144, + 0.85, + 0.173 + ], + "angle": 0, + "content": "
MethodSLOCDIXEPRISET-Attr
Score0.880.870.790.770.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.176, + 0.855, + 0.189 + ], + "angle": 0, + "content": "Table 5. FunnyBirds results for the ViT-B model." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.204, + 0.951, + 0.248 + ], + "angle": 0, + "content": "
MethodSLOCACDIXEPGCGC++GIGIGLCLTXRISE
mIoU↑0.560.550.660.520.550.560.510.480.550.560.51
mAP↑0.800.860.840.760.850.850.780.760.850.830.79
PA↑0.770.720.820.720.730.730.740.790.730.510.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.251, + 0.868, + 0.265 + ], + "angle": 0, + "content": "Table 6. Segmentation tests results for the RN model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.367 + ], + "angle": 0, + "content": "achieves state-of-the-art performance on the FB benchmark as well. The distinct evaluation protocol employed in FB offers a complementary perspective to the faithfulness evaluation conducted on the IN dataset, further reinforcing our confidence in the robustness of SLOC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Table 6 presents segmentation tests results on the IN-Seg dataset using the RN model. We observe that DIX achieves the best average performance across all metrics, while the runner-up varies depending on the specific metric. Overall, SLOC is found to be competitive, on average, with all runner-up methods. We note that higher segmentation accuracy may reflect a method's ability to facilitate strong object detection rather than to identify the most informative features that explain the model's prediction. In reality, the most explanatory features do not always cover the entire object; instead, they may correspond to a subset of features that are critical to the model's decision. Consequently, higher segmentation accuracy does not necessarily indicate superior explanatory value." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Due to space limitation, ablation studies, runtime comparison, sanity checks [3], and additional qualitative analyses, are provided in Appendices F, H, J, and I, respectively. These experiments offer further insights into the effectiveness and efficiency of SLOC." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.686, + 0.634, + 0.701 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.878 + ], + "angle": 0, + "content": "This work introduced SLOC, a novel and efficient explainability method. By rethinking completeness as a guiding principle promoted locally rather than as a strict global constraint, SLOC exhibits state-of-the-art performance across various explainability benchmarks. These findings suggest that completeness, when promoted in a soft and local manner, provides a robust foundation for generating high-quality attribution maps that closely align with the model's predictive behavior and human comprehension. Finally, discussions on the limitations of SLOC and potential avenues for future research are provided in Appendix M." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19801" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.155 + ], + "angle": 0, + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, pages 9505-9515, 2018. 1, 8, 12, 25, 26" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International conference on machine learning, pages 435-451. PMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[5] Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. In Proceedings of the International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[6] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10 (7):e0130140, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[7] Oren Barkan, Yonatan Fuchs, Avi Caciularu, and Noam Koenigstein. Explainable recommendations via attentive multi-persona collaborative filtering. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 468-473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[8] Oren Barkan, Ori Katz, and Noam Koenigstein. Neural attentive multiview machines. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3357-3361. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[9] Oren Barkan, Idan Rejwan, Avi Caciularu, and Noam Koenigstein. Bayesian hierarchical words representation learning. arXiv preprint arXiv:2004.07126, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[10] Oren Barkan, Omri Armstrong, Amir Hertz, Avi Caciularu, Ori Katz, Itzik Malkiel, and Noam Koenigstein. Gam: Explainable visual similarity and classification via gradient activation maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 68-77, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Oren Barkan, Edan Hauon, Avi Caciularu, Ori Katz, Itzik Malkiel, Omri Armstrong, and Noam Koenigstein. Gradsam: Explaining transformers via gradient self-attention maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 2882–2887, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Oren Barkan, Yuval Asher, Amit Eshel, Yehonatan Elisha, and Noam Koenigstein. Learning to explain: A model-agnostic framework for explaining black box mod" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "els. In 2023 IEEE International Conference on Data Mining (ICDM), pages 944-949. IEEE, 2023. 2, 5, 7, 15, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[13] Oren Barkan, Yehonatan Elisha, Yuval Asher, Amit Eshel, and Noam Koenigstein. Visual explanations via iterated integrated attributions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2073-2084, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Deep integrated explanations. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 57-67, 2023. 2, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[15] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Stochastic integrated explanations for vision models. In 2023 IEEE International Conference on Data Mining (ICDM), pages 938-943. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[16] Oren Barkan, Tom Shaked, Yonatan Fuchs, and Noam Koenigstein. Modeling users' heterogeneous taste with diversified attentive user profiles. User Modeling and User-Adapted Interaction, pages 1–31, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, and Noam Koenigstein. A counterfactual framework for learning and evaluating explanations for recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3723-3733, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.45, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[18] Oren Barkan, Yehonatan Elisha, Yonatan Toib, Jonathan Weill, and Noam Koenigstein. Improving llm attributions with randomized path-integration. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9430-9446, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[19] Oren Barkan, Yonatan Toib, Yehonatan Elisha, and Noam Koenigstein. A learning-based approach for explaining language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 98-108, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.592, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[20] Oren Barkan, Yonatan Toib, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Llm explainability via attributive masking learning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9522-9537, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.663, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Oren Barkan, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Bee: Metric-adapted explanations via baseline exploration-exploitation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1835-1843, 2025. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[22] Alexander Binder, Gregoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In International Conference on Artificial Neural Networks, pages 63-71. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Alexander Binder, Leander Weber, Sebastian Lapuschkin, Grégoire Montavon, Klaus-Robert Müller, and Wojciech Samek. Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16143-16152, 2023. 6, 25" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[24] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[25] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[26] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847. IEEE, 2018. 2, 7, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[27] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[28] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 2, 6, 7, 12, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[29] Jianbo Chen, Le Song, Martin J. Wainwright, and Michael I. Jordan. L-shapley and c-shapley: Efficient model interpretation for structured data. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[30] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In Advances in Neural Information Processing Systems, pages 6970-6979, 2017. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[31] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Computer Vision and Pattern Recognition (CVPR), 2009. 6, 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[32] Saurabh Satish Desai and H. G. Ramaswamy. Ablationcam: Visual explanations for deep convolutional network via gradient-free localization. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 972-980, 2020. 1, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[33] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[34] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Yehonatan Elisha, Oren Barkan, and Noam Koenigstein. Probabilistic path integration with mixture of baseline distributions. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 570-580, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[36] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving performance of deep" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "learning models with axiomatic attribution priors and expected gradients. Nature machine intelligence, 3(7):620-631, 2021. 3, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[37] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[38] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2950-2958, 2019. 2, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[39] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3429-3437, 2017. 1, 2, 7, 15, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[40] Keren Gaiger, Oren Barkan, Shir Tsipory-Samuel, and Noam Koenigstein. Not all memories created equal: Dynamic user representations for collaborative filtering. IEEE Access, 11: 34746-34763, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[41] Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR), 51(5):1-42, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[42] Matthieu Guillaumin, Daniel Kuttel, and Vittorio Ferrari. Imagenet auto-annotation with segmentation propagation. International Journal of Computer Vision, 110(3):328-348, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[44] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods, 2023. 6, 13, 14, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[45] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems, 32, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[46] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.908, + 0.786 + ], + "angle": 0, + "content": "[47] Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4198-4205, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[48] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[49] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4948-4957, 2019. 6, 12, 13" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19803" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[50] Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050-5058, 2021. 2, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[51] Sunnie SY Kim, Nicole Meister, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision, pages 280-298. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[52] D Liu, R Nicoescu, and R Klette. Stereo-based bokeh effects for photography. Machine Vision and Applications, pages 1-13, 2016. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[53] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pages 4765-4774, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[54] Itzik Malkiel, Dvir Ginzburg, Oren Barkan, Avi Caciularu, Jonathan Weill, and Noam Koenigstein. Interpreting bert-based text similarity via activation and saliency maps. In Proceedings of the ACM Web Conference 2022, pages 3259-3268, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[55] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern Recognition, 65:211-222, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[56] Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, Jörg Schlötterer, Maurice Van Keulen, and Christin Seifert. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai. ACM Computing Surveys, 55(13s): 1-42, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[57] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2, 3, 6, 7, 12, 15, 16, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[58] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 983-991, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[60] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3145-3153. JMLR.org, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[61] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[63] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. In NeurIPS, 2019. 2, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[64] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, pages 3319-3328, 2017. 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[65] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319-3328. JMLR.org, 2017. 1, 2, 3, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[67] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[68] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[69] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9680-9689, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19804" + } + ] +] \ No newline at end of file diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_origin.pdf b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bea0faad310573790a753e0616ec4bab62406185 --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/4e70f0b0-09f1-4a52-a8df-083516f56711_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57339adf63eaace4a79054cbbcabbda7d8b7688dcd97c153609bae699acdf6a9 +size 570660 diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/full.md b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f36055148f316d177f465043c2123b169b822a29 --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/full.md @@ -0,0 +1,330 @@ +# Soft Local Completeness: Rethinking Completeness in XAI + +Ziv Weiss Haddad* Oren Barkan* + +The Open University + +Yehonatan Elisha Noam Koenigstein + +Tel Aviv University + +https://github.com/xaisloc/sloc + +# Abstract + +Completeness is a widely discussed property in explainability research, requiring that the attributions sum to the model's response to the input. While completeness intuitively suggests that the model's prediction is "completely explained" by the attributions, its global formulation alone is insufficient to ensure faithful explanations. We contend that promoting completeness locally within attribution subregions, in a soft manner, can serve as a standalone guiding principle for producing faithful attributions. To this end, we introduce the concept of the completeness gap as a flexible measure of completeness and propose an optimization procedure that minimizes this gap across subregions within the attribution map. Extensive evaluations across various model architectures demonstrate that our method produces state-of-the-art results. + +# 1. Introduction + +In recent years, Explainable AI (XAI) has emerged as a critical aspect of machine learning, especially for deep learning models, which are often regarded as "black boxes" due to their lack of transparency [8, 9, 25, 41, 66]. XAI aims to provide insights into how these models make predictions, offering a way to ensure their reliability and trustworthiness [34]. A common approach to model explainability is through attribution maps, which assign importance scores to individual input features based on their contribution to the model's prediction [22, 27, 65]. These attributions serve as a tool for understanding which aspects of the input were most influential in determining the output. + +A widely discussed notion in XAI is completeness (also known as conservation or efficiency) [4, 55, 60, 65]. Completeness is defined as the requirement that the elements in an attribution map should sum to the difference between the model's prediction for a given input and a baseline (e.g., the null representation). This difference is also known as the + +model's response. When completeness is achieved inherently by an attribution method, it is intuitively appealing, as it suggests that the attribution map fully captures the factors contributing to the model's response. However, even a low-quality attribution map can be superficially adjusted to satisfy completeness through post hoc global normalization. Yet, such an approach lacks genuine explanatory power, as the normalization is decoupled from the underlying attribution mechanism. + +This paper recognizes that completeness, as defined in the literature as a global property that an explanation method should satisfy, is too weak to serve as a standalone criterion. Indeed, several well-regarded explanation methods [32, 39, 59] do not inherently satisfy completeness. Nevertheless, these methods have demonstrated considerable effectiveness across diverse objective evaluation metrics, providing faithful explanations and successfully passing various sanity checks [3]. This suggests that completeness is neither a necessary nor sufficient condition for generating high-quality explanations. + +In this work, we differentiate between global completeness, as originally defined in the literature, which imposes a completeness constraint on the entire attribution map, and local completeness, which requires completeness within individual subregions of the attribution map. By rethinking completeness as a local and flexible guiding measure rather than a strict global constraint, we seek to overcome its limitations while harnessing the appealing motivation behind it for producing meaningful and faithful explanations. To achieve this, we require subregions of the attribution map, referred to as sub-maps, to promote completeness locally by accounting for the model's response to the corresponding subregions of the input image, yet in soft manner. An attribution map that adheres to this guiding principle should assign high (low) importance to image subregions where the model exhibits a strong (weak) response, thereby maintaining faithfulness to the model. + +Unlike global completeness, which is a relatively weak requirement that can be easily satisfied, achieving local completeness across an arbitrary set of sub-maps may not + +always be feasible. To address this challenge, we introduce the completeness gap - a measure that quantifies the deviation of a sub-map from completeness. Specifically, the completeness gap measures the difference between the sum of elements in each sub-map and the model's response to the corresponding subregions of the input image. Rather than imposing a strict binary constraint, the completeness gap serves as a soft, quantifiable criterion for local completeness. We argue that minimizing this gap over a diverse set of sub-maps leads to faithful explanations. + +To this end, we present Soft LOcal Completeness (SLOC) - a novel explainability method that promotes completeness locally within sub-maps of the attribution map in a soft and flexible manner. SLOC accomplishes this by seeking to minimize the completeness gap locally for each individual sub-map, simultaneously. + +The motivation for minimizing the completeness gap locally is to emphasize or attenuate sub-maps based on the actual impact of their corresponding input subregions on the model's output. SLOC achieves this through a gradient-based optimization process, where each sub-map is adjusted to achieve local completeness in a soft manner. Therefore, SLOC's goal is not to enforce strict completeness, globally, but to use it as a guiding principle for refining a large set of sub-maps (subregions within the attribution), iteratively and simultaneously. + +Our contribution is the introduction of the SLOC method, which facilitates a novel optimization procedure to promote completeness in a soft local manner. The effectiveness of SLOC is demonstrated through extensive experiments on various model architectures, where it is shown to produce state-of-the-art results across multiple benchmarks. + +# 2. Related Work + +The XAI literature encompasses a broad range of approaches for attributing model predictions to specific input features across different tasks [2, 4, 6, 7, 16, 24, 29, 40, 54]. Early gradient-based methods produce explanation maps by leveraging gradients directly [5, 60, 62] or through functions that combine class activation maps with or without their gradients [10, 26, 48, 58, 59, 63]. With the rise of transformer architectures, new explainability techniques for transformers have emerged [1, 4, 11, 67]. For example, Transformer Attribution [28] introduces a class-specific Deep Taylor Decomposition, applying relevance propagation for both positive and negative attributions. Building on this, Generic Attention Explainability [27] generalizes Transformer Attribution to explain bi-modal transformers. + +Path integration methods form another prominent family of attribution techniques [13, 15, 18, 50, 65]. Integrated Gradients (IG) [65] computes attributions by integrating gradients along an interpolation path from a baseline to the input. A refined variant, Guided Integrated + +Gradients (GIG) [50], uses an adaptive path that avoids high-gradient regions, thereby reducing irrelevant attributions. Recently, Deep Integrated Explanations (DIX) [14] proposed performing integration over intermediate network representations instead of inputs, producing more faithful attribution maps than IG and GIG. + +Perturbation-based (occlusion) techniques generate attribution maps by perturbing parts of the input to assess the contributions of specific elements to the prediction [17, 19, 20, 53]. For example, RISE [57], creates perturbations by masking areas in the image through the upsampling of randomly drawn low-resolution binary grids. The class score corresponding to each masked version of the image serves as an importance score for that specific mask. A linear combination of all masks, weighted by their importance, then forms the final attribution map. Another branch of perturbation methods are learning-based perturbation techniques [12, 30, 38, 39]. For example, Meaningful Perturbation (MP) [39] learns to produce perturbations by masking the smallest region possible while significantly altering the original prediction, isolating the minimum necessary content for the prediction. Alternatively, LTX [12] introduces a surrogate 'explainer' model pretrained to mask as much of the input as possible while preserving the original prediction, thereby ensuring the retained features are those most relevant to the model's prediction. Then, LTX finetunes the attribution per specific example, allowing the selection of the best-performing attribution w.r.t. to the metric at hand. + +Our SLOC method falls within the family of learning-based perturbation techniques, yet it diverges from prior works in several key aspects. First, conceptually, SLOC presents a distinct methodology by promoting local completeness rather than attempting to maximize or minimize the model prediction through learned perturbations [12, 30, 39]. Consequently, SLOC does not require gradient backpropagation through the model, and incorporates the model's predictions on the masked inputs into the loss as constant nodes within the computation graph. This design yields a simple, computationally efficient gradient expression, and relies solely on the model's forward pass, hence operating in a true black-box setting. Finally, unlike LTX, SLOC avoids the overhead of learning an 'explainer' function, which would also require a dataset aligned with the training data distribution of the model being explained. + +# 3. Soft Local Completeness + +Let $f: \mathbb{R}^n \to [0,1]^C$ be a classification model that takes an input image $\mathbf{x} \in \mathbb{R}^n$ and outputs a discrete probability distribution $f(\mathbf{x}) \in [0,1]^C$ over $C$ classes. Given a target class $y$ to explain, our goal is to produce an attribution map + +$\mathbf{a}_{\mathbf{x}}^{y} \in \mathbb{R}^{n}$ that attributes the contribution of each element $i$ in $\mathbf{a}_{\mathbf{x}}^{y}$ to the prediction score $f(\mathbf{x})[y]$ for the class $y$ . + +Completeness requires that the attributions in $\mathbf{a}_{\mathbf{x}}^{y}$ collectively account for the difference between the model's prediction on the input $\mathbf{x}$ and the baseline representation $\mathbf{b}$ , which is set to the black image in this work. + +The model's response to $\mathbf{x}$ is defined as $r(\mathbf{x})\coloneqq f(\mathbf{x}) - f(\mathbf{b})$ . Formally, the attribution map $\mathbf{a}_{\mathbf{x}}^{y}$ satisfies completeness if $\sum_{i = 1}^{n}\mathbf{a}_{\mathbf{x}}^{y}[i] = r(\mathbf{x})[y]$ . We note that this property can be superficially imposed on an attribution map through a post-processing step by normalizing each element using the factor $\frac{r(\mathbf{x})[y]}{\sum_{i = 1}^{n}\mathbf{a}_{\mathbf{x}}^{y}[i]}$ . However, artificially satisfying completeness through such normalization does not inherently yield a meaningful explanation, as the imposed completeness is an external adjustment rather than an intrinsic characteristic of the explanation method itself. For example, even a flat attribution map (where all elements are equally weighted) could be normalized to satisfy the completeness criterion, yet such a map lacks any explanatory value. Consequently, relying solely on completeness, in its global form, is deemed insufficient for producing meaningful attribution maps. + +# 3.1. Motivation + +Our SLOC approach is motivated by the toy example illustrated in Fig. 1, which depicts an image where the key region influencing the model's prediction for the 'great-grey-owl' class is approximately centered. Figures 1(a)-(d) present four images: the original image (a) and three perturbed versions obtained by masking different regions (b)-(d). + +The four different masks applied in Figures 1(a)-(d) are illustrated in Figure 1(e). Each mask selectively perturbs the original image by replacing certain regions with the baseline (in this work, the null pixels). For instance, the mask corresponding to Figure 1(c), shown in the top-right corner of Figure 1(e), applies a perturbation by replacing all pixels within the black rectangle (located at the bottom-right corner) with the baseline. + +Notably, the model's response for the great-grey-owl class remains consistent across all four images (a-d), with a value of 0.8. Recall that the model response is defined as the difference between the model's class probability for a given image and that for the baseline (the null image). + +As defined earlier, a sub-map represents a subset of elements in the attribution map that are associated with spe + +(a) Original image: 0.8 + +(c) Masked Image: 0.8 + +(e) Four masks + +![](images/b533a8a769d8c12588b6a047ff01364eb955602d2c445a1184a530175715222c.jpg) +(b) Masked Image: 0.8 + +![](images/fa29e25e40b9a32422368f876b13d79fd6f925d63204c3f28384f530f25bcfde.jpg) +(f) Attribution Map + +![](images/f83fdecf052eb91401b67d7503b9af933f85e31381411d9718162246f9b80f01.jpg) + +![](images/86d871f7e3f47b7d9e769583d6be8bffc4198da1c7b6ea0c536cec1aea088375.jpg) +(d) Masked Image: 0.8 + +![](images/dc026f536af31b0b5bb11d6d9ec64ef1776710d20a0973f005bad06109110bde.jpg) +Figure 1. Subfigures (a)-(d) present the model's response (0.8) for the class 'great-grey-owl' when applied to the original input (a) and its masked versions (b)-(d). The four masks used to generate (a)-(d) are depicted in (e). A sub-map refers to a subset of elements in the attribution map that correspond to specific subregions in the input. We consider four distinct sub-maps (not shown), each corresponding to the non-masked (visible) subregions in (a)-(d). For instance, the sub-map associated with (c) includes all elements in the attribution map except those corresponding to the black rectangle in the bottom-right corner of (c). This exclusion aligns with the masked region in the input, effectively determining which attributions are included in that sub-map. A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by (c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to (c). By enforcing local completeness across the four sub-maps induced by (a)-(d), the remaining central stripe, which is unmasked in all images (a)-(d), must account for the model's entire response of 0.8. This results in the attribution depicted in (f) that highlights the central stripe containing the owl in red. See 3.1 for more details. + +![](images/73a75665b9f038bf3a5911b1cd128e608ad969183787279c05dcab096c01a115.jpg) + +cific subregions in the input. In the context of Fig. 1, we consider four different sub-maps, each corresponding to the non-masked (visible) subregions in Figs.1(a)-(d). For instance, the sub-map corresponding to Fig. 1(c) consists of all elements in the attribution map except those associated with the black rectangle in the bottom-right corner of Fig.1(c). This exclusion aligns with the masked region in the input, effectively defining which attributions are considered within that sub-map. Notably, the sub-map corresponding to the original image (Fig. 1(a)) is equivalent to the entire attribution map, as no regions are masked. This formulation allows us to analyze how different subregions contribute to completeness by evaluating attributions over various sub-maps. + +A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by Fig. 1(c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to Fig. 1(c). + +Our key insight is that enforcing local completeness across the four sub-maps induced by Figs.1(a)-(d) allows us to pinpoint the important region of the image w.r.t. the model's prediction. Specifically, the intersection of these four sub-maps forms a central stripe around the owl, illustrated in red in Fig.1(f). By evenly distributing the total attribution sum of 0.8 across the pixels within this intersection (red stripe) and assigning 0 elsewhere (highlighted in dark blue in Fig. 1(f)), we obtain an attribution map that ensures local completeness for all four sub-maps. Since each sub-map contains the intersection while the attribution values outside it are zero, the sum of attributions within each sub-map is exactly the sum of attributions within the intersection: 0.8. This resulting attribution map effectively captures the key region containing the owl. + +This simple example can be slightly generalized under similar settings. We make the following assumptions: (1) The model response for a subregion is $R$ if the subregion contains the object of interest and 0 otherwise. (2) Each sub-map either fully contains the object of interest or does not intersect with it at all. (3) One of the sub-maps corresponds to the complete attribution map. (4) All attribution values are non-negative. It can then be shown that if an attribution map satisfies local completeness for all sub-maps, the intersection defined by these sub-maps is the largest subregion that can have non-zero attribution. Simply put, this intersection highlights the important regions in the image responsible for the model's prediction. Moreover, as the number of sub-maps increases, the important regions can potentially become more precisely defined. The formalization and proof follow straightforwardly and are provided in Appendix K. + +Keep in mind, however, that this setting and its assumptions are simplistic and unrealistic, serving only to illustrate the motivation. Therefore, SLOC does not actually rely on these assumptions. In more realistic scenarios, satisfying local completeness for all sub-maps is often infeasible (an example is provided in Fig. 11 in Appendix I). To address this challenge, we introduce the completeness gap, which quantifies the degree to which a sub-map deviates from local completeness. Our approach seeks to minimize the completeness gap across multiple sub-maps simultaneously, hence promoting completeness locally within subregions of the attribution map, in a soft manner. We formulate this as an optimization problem and employ gradient descent to find a solution. + +Finally, recall that in our simple example, in the absence of additional information, we distributed the attribution uniformly across the central red stripe. When introducing the loss terms for the SLOC optimization, we will also discuss regularization terms. One such term is the Total Variation loss, which encourages spatial smoothness in the attribution. As we demonstrate in Sec. 4, by encouraging this soft + +form of local completeness, the resulting attribution map offers robust explanatory performance, both visually and according to a variety of evaluation metrics. + +# 3.2. SLOC optimization + +Let $\mathbf{m} \in \{0,1\}^n$ be a binary mask. We define the masked input by + +$$ +\mathbf {x} ^ {\mathbf {m}} = \mathbf {x} \circ \mathbf {m} + (1 - \mathbf {m}) \circ \mathbf {b}, \tag {1} +$$ + +where $\circ$ stands for the elementwise product. In words, $\mathbf{x}^{\mathbf{m}}$ is a perturbation (masked version) of $\mathbf{x}$ , according to the mask $\mathbf{m}$ , where all masked elements of $\mathbf{x}^{\mathbf{m}}$ are replaced with their respective elements from the baseline $\mathbf{b}$ . + +In this work, we consider binary masks constructed over patches, where each patch is randomly assigned a Bernoulli outcome with probability $p$ . We explore two methodologies for automatically selecting $p$ : one that determines $p$ dynamically per input image, and another that uses a fixed value per model. The reader is referred to Appendix D for the exact details of the masks construction process, including the methodologies used for determining $p$ . + +Given an initial attribution map $\mathbf{a}_{\mathbf{x}}^{y}$ and a set of masks $\mathcal{M} \subset \{0,1\}^n$ , the SLOC loss is defined as follows: + +$$ +\mathcal {L} _ {c} \left(\mathbf {a} _ {\mathbf {x}} ^ {y}; \mathcal {M}\right) = \frac {1}{2 | \mathcal {M} |} \sum_ {\mathbf {m} \in \mathcal {M}} \frac {1}{| \mathbf {m} |} \underbrace {(r \left(\mathbf {x} ^ {\mathbf {m}}\right) [ y ] - \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m})} _ {\text {c o m p l e t e n e s s g a p}} ^ {2}, \tag {2} +$$ + +where $\cdot$ is the dot-product operator, and $|\cdot|$ denotes the L1 norm. As observed in Eq. 2, $\mathcal{L}_c$ promotes the minimization of the completeness gap, a soft measure of completeness, locally, for all sub-maps induced by $\mathcal{M}$ . The rationale behind the local minimization of the completeness gap is that attributions within a sub-map should be strengthened or reduced so that their sum closely matches $r(\mathbf{x}^{\mathbf{m}})[y]$ - the model's response to the respective regions exposed in $\mathbf{x}^{\mathbf{m}}$ . This behavior is achieved by applying gradient descent on $\mathcal{L}_c$ with respect to $\mathbf{a}_{\mathbf{x}}^{y}$ . + +To better understand the optimization process, one can inspect the gradient: + +$$ +\frac {\partial \mathcal {L} _ {c}}{\partial \mathbf {a} _ {\mathbf {x}} ^ {y}} = - \frac {1}{| \mathcal {M} |} \sum_ {\mathbf {m} \in \mathcal {M}} \frac {1}{| \mathbf {m} |} (r (\mathbf {x} ^ {\mathbf {m}}) [ y ] - \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m}) \mathbf {m}. \tag {3} +$$ + +For simplicity, assume $\mathcal{M}$ contains a single mask $\mathbf{m}$ . Then, for each element $\mathbf{a}_{\mathbf{x}}^{y}[i]$ that is exposed in $\mathbf{x}^{\mathbf{m}}$ (i.e., $\mathbf{m}[i] = 1$ ), the gradient is the negative of the completeness gap normalized by the number of elements in the sub-map, i.e., $\frac{\partial\mathcal{L}_c}{\partial\mathbf{a}_x^y[i]} = -\frac{1}{|\mathbf{m}|} (r(\mathbf{x}^{\mathbf{m}})[y] - \mathbf{a}_{\mathbf{x}}^y\cdot \mathbf{m})$ , and for each $i$ where $\mathbf{m}[i] = 0$ , the gradient vanishes. Accordingly, the gradient descent update ensures an equal increase (decrease) for all elements in the sub-map when the sum of the attributions falls short of (exceeds) the model's response $r(\mathbf{x}^{\mathbf{m}})[y]$ . Assuming a learning rate of 1, it follows that after the gradient descent update, the completeness gap vanishes, resulting in + +an updated sub-map satisfying completeness. This can be seen mathematically as follows: for a sub-map induced by the mask $\mathbf{m}$ , the sum of its elements after the gradient update is given by + +$$ +\begin{array}{l} \sum_ {i: \mathbf {m} [ i ] = 1} \mathbf {a} _ {\mathbf {x}} ^ {y} [ i ] - \frac {\partial \mathcal {L} _ {c}}{\partial \mathbf {a} _ {\mathbf {x}} ^ {y} [ i ]} \\ = \sum_ {i: \mathbf {m} [ i ] = 1} \mathbf {a} _ {\mathbf {x}} ^ {y} [ i ] + \underbrace {\frac {1}{| \mathbf {m} |} (r (\mathbf {x} ^ {\mathbf {m}}) [ y ] - \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m})} _ {\text {i n d e p e n d e n t o f} i} \\ = | \mathbf {m} | \frac {1}{| \mathbf {m} |} (r (\mathbf {x} ^ {\mathbf {m}}) [ y ] - \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m}) + \sum_ {i: \mathbf {m} [ i ] = 1} \mathbf {a} _ {\mathbf {x}} ^ {y} [ i ] \\ = r (\mathbf {x} ^ {\mathbf {m}}) [ y ] - \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m} + \mathbf {a} _ {\mathbf {x}} ^ {y} \cdot \mathbf {m} = r (\mathbf {x} ^ {\mathbf {m}}) [ y ], \\ \end{array} +$$ + +where the second transition follows from the fact that the normalized completeness gap is independent of $i$ and is therefore summed $|\mathbf{m}|$ times, as the summation is taken over every $i$ where $\mathbf{m}[i] = 1$ , with $\mathbf{m}$ being a binary tensor. The penultimate transition then arises from expressing the sum of elements in the sub-map as the dot product, $\mathbf{a}_{\mathbf{x}}^{y} \cdot \mathbf{m}$ . This derivation justifies the normalization of the squared completeness gap by $|\mathbf{m}|$ in Eq. 2, as it ensures that after the gradient update, the sub-map satisfies local completeness. + +To give further intuition for using the normalized completeness gap in Eq. 2, consider two different masks: one that exposes a single pixel (element), and another that exposes a large portion of the input. Given that the gradient update of elements in the sub-map is equal for all elements, it makes sense to distribute the completeness gap among all elements. In the case of a mask exposing a single pixel, we have full confidence that this specific element is responsible for the model's response $r(\mathbf{x}^{\mathrm{m}})[y]$ . Therefore, it is reasonable to expect that the gradient update for this element will match the size of the completeness gap. However, in the case of a sub-map associated with a mask exposing large portions of the input, we cannot differentiate the accountability of individual elements within the exposed portions for the model's response. Thus, without further information, it is reasonable to apply an update step that evenly distributes the completeness gap across all exposed elements. Hence, the squared completeness gap is divided by $|\mathbf{m}|$ . + +So far, the discussion and analysis have been limited to the case of a single mask, i.e., $|\mathcal{M}| = 1$ . When considering a set of masks, the update for $\mathbf{a}_{\mathbf{x}}^{y}[i]$ is determined by the accumulation of gradients from all the sub-maps it is associated with. Specifically, sub-maps that include the $i$ -th element and whose total sum of elements is below (exceeds) the model's response $r(\mathbf{x}^{\mathrm{m}})[y]$ will contribute to the intensification (reduction) of $\mathbf{a}_{\mathbf{x}}^{y}[i]$ , with the contribution from each individual mask weighted according to the normalized completeness gap it induces. + +![](images/4249ef80320f2516bc639b72cde85f1ac389b5da773249ce5d9982fa591efaeb.jpg) +Masks & Responses Generation Phase +Figure 2. SLOC overview: The upper section illustrates the Masks & Responses Generation Phase. In this phase, (1) $\mathcal{M}$ masks are randomly generated. (2) Perturbations are produced by combining the original image and baseline according to the masks, as defined in Eq. (1) (3) These perturbations are passed as inputs to the model, and the corresponding outputs are obtained. (4) Subtracting the model's output for the baseline from these outputs produces the model responses. (5) These masks and their corresponding responses are then passed to the Optimization Phase. (6) Starting with an initial attribution. (7) Gradient descent is used to iteratively update the attribution. The loss function $\mathcal{L}$ (Eq. 4) consists of three components: the completeness gap (computed using the masks, responses, and attributions), as well as TV and L1 regularization. See Sec. 3.2 for details. + +To promote smooth and focused attribution maps, we combine $\mathcal{L}_c$ with two types of regularization applied to the attribution map: Total Variation (TV) and L1 regularization that encourage smoothness and sparsity, respectively. The final loss function is then defined as: + +$$ +\mathcal {L} (\mathbf {a _ {x} ^ {y}}; \mathcal {M}) = \mathcal {L} _ {c} (\mathbf {a _ {x} ^ {y}}; \mathcal {M}) + \lambda_ {1} | \mathbf {a _ {x} ^ {y}} | + \lambda_ {2} \mathrm {T V} (\mathbf {a _ {x} ^ {y}}), (4) +$$ + +with + +$$ +\operatorname {T V} \left(\mathbf {a} _ {\mathbf {x}} ^ {y}\right) := \sum_ {i, j} \left(\mathbf {a} _ {\mathbf {x}} ^ {y} [ i, j ] - \mathbf {a} _ {\mathbf {x}} ^ {y} [ i + 1, j ]\right) ^ {2} + \left(\mathbf {a} _ {\mathbf {x}} ^ {y} [ i, j ] - \mathbf {a} _ {\mathbf {x}} ^ {y} [ i, j + 1 ]\right) ^ {2} \tag {5} +$$ + +where $\mathbf{a}_{\mathbf{x}}^{y}$ in Eq. 5 is reshaped to match the spatial dimensions of the final attribution map. Additionally, $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters controlling the strength of the regularization terms. In our experiments, these hyperparameters are set based on a separate validation set. + +The optimization of $\mathbf{a}_{\mathbf{x}}^{y}$ is carried out via gradient descent on $\mathcal{L}$ . The optimization process allows for tracking the performance of the metric during the optimization and selecting the best-performing attribution map based on this metric from the resulting maps [12]. A diagram of the SLOC method is depicted in Fig. 2. + +![](images/35fc463f0d7766e6c4af7289daae1f6ac631fc65990a80dc1d83d2907e3e4592.jpg) +Figure 3. SLOC attribution maps across training steps. Faithful explanations emerge after a few hundred gradient updates, with an appropriate learning rate decay. + +In practice, we observed that a few hundred gradient updates, combined with an appropriate learning rate decay, were sufficient to produce faithful and meaningful explanations. Therefore, in our experiments, we report SLOC results for $T = 500$ . An example for SLOC's attribution across gradient update steps is shown in Fig. 3. For the exact implementation details and hyperparameter configuration, the reader is referred to Appendix D. + +Due to space limitations, Appendix G provides a detailed discussion of SLOC's computational complexity. Appendix H presents a runtime analysis, comparing SLOC against various methods and demonstrating its efficiency over other state-of-the-art learning-based perturbation methods. Additionally, Appendix L examines additional axioms and explains which of them are satisfied by SLOC. + +Finally, it is important to clarify that the SLOC optimization procedure does not guarantee convergence to an attribution map that fully satisfies completeness. Instead, it leverages the completeness gap as a guiding measure to update subregions within the attribution map, thereby promoting local completeness for these subregions in a soft manner. Nevertheless, as demonstrated in Sec. 4, SLOC has proven effective in generating attribution maps that outperform state-of-the-art explanation methods. + +# 4. Experiments and Results + +# 4.1. Experimental setup + +The experiments were conducted on an NVIDIA DGX 8xA100 Server, utilizing the PyTorch package. As explanations are multifaceted by nature, no single evaluation metric can fully capture their quality, and no definitive metric exists [21, 23, 35, 45, 47, 51, 68]. Therefore, our evaluation encompasses several distinct protocols designed to assess explanation quality from multiple complementary perspectives, on different datasets, and across four model architectures: ResNet50 (RN) [43], DenseNet201 (DN) [46], ViT-Base (ViT-B) and ViT-Small (ViT-S) [33]. + +Experiment 1: Faithfulness evaluation This evaluation assesses the faithfulness (also known as correctness or fidelity) of the explanation via a set of perturbation-based + +metrics. These metrics are designed to reveal the actual input elements the model relies on for its predictions. Following prior works [28, 49, 50, 69], we conducted an extensive evaluation using a comprehensive set of perturbation-based explanation metrics to assess the faithfulness of the generated explanations: the Area Under the Curve (AUC) of Positive (POS) and Negative (NEG) perturbation tests [28], AUC of the Insertion (INS) and Deletion (DEL) tests [57], and AUC of the Softmax Information Curve (SIC) and Accuracy Information Curve (AIC) [49]. For POS and DEL, lower values indicate better performance, while for NEG, INS, SIC, and AIC, higher values are preferred. Since NEG and POS, as well as INS and DEL, are complementary metrics, we also report the NEG-POS Difference (NPD) and the INS-DEL Difference (IDD) as single summaries of their respective complementary performances. Faithfulness performance is reported on a sample from the ImageNet [31] ILSVRC 2012 (IN) validation set, which consists of 10,000 images across 1,000 classes. Since most images in the IN dataset contain a single, centrally focused object, we extend our evaluation to multi-object scenes by assessing explanation faithfulness on a sample from the Pascal VOC 2012 [37] (VOC) test set, which includes 1,000 images spanning 20 object classes. + +Experiment 2: FunnyBirds evaluation While perturbation-based protocols are a standard approach for assessing explanation faithfulness, they are criticized for introducing domain shifts that undermine the validity of the evaluation [45]. To address this, we additionally report results on the FunnyBirds (FB) dataset which consists of 500 images from 50 classes, following the evaluation protocol from [44]. The FB evaluation employs controlled interventions to estimate ground-truth (GT) importance scores at the level of object parts rather than individual pixels, mitigating issues associated with pixelwise perturbation-based evaluations [68]. Furthermore, FB avoids domain shifts by incorporating semantically meaningful interventions during training. By evaluating explanations at the part level—closer to human visual understanding—FB reduces the gap between automatic XAI evaluation (e.g., faithfulness metrics) and human-centric studies, providing a more interpretable and robust assessment of the explanations. FB evaluates explainability in three aspects: Completeness, Correctness, and Contrastivity - and provides a combined overall score (higher is better). + +Experiment 3: Segmentation evaluation To ensure a comprehensive comparison with prior works [27, 28, 48], we conduct an extensive evaluation using segmentation tests on the ImageNet-Seg dataset (IN-Seg) [42], a subset of the ImageNet validation set comprising 4,276 human-annotated + +ground-truth (GT) segmentation maps across 445 classes. As detailed in [28], these tests generalize the "Pointing Game" [45] and serve as a human-grounded evaluation [56] by measuring the alignment between GT segmentation maps and the attributions generated by the explanation method. While higher segmentation accuracy does not necessarily indicate greater explanatory power [57], segmentation tests assess the alignment between explanations and human-annotated GT, hence providing a complementary perspective to other evaluations. We follow the evaluation protocol from [28], assessing segmentation performance using Pixel Accuracy (PA), mean Intersection-over-Union (mIoU), and mean Average Precision (mAP), higher is better for all metrics. + +Detailed descriptions of the evaluation protocols and metrics used in Experiments 1, 2, and 3 are provided in Appendices B.1, B.2, and B.3, respectively. + +Evaluated methods We evaluate SLOC against a comprehensive suite of 14 explanation methods encompassing gradient-based, path-integration, perturbation, and CAM methods. For CNN models, we include the following methods: Grad-CAM (GC) [59], Grad-CAM++ $(\mathbf{GC} + +)$ [26], Deep Integrated Explanations (DIX) [14], FullGrad (FG) [63], Ablation-CAM (AC) [32], Layer-CAM (LC) [48], Learning To Explain (LTX) [12], RISE [57], Meaningful Perturbation (MP) [39], Extremal Perturbations (EP) [38], Integrated Gradients (IG) [64], and Guided IG (GIG) [50]. For ViT models, we considered the following methods: Transformer Attribution (T-Attr) [28] and Generic Attention Explainability (GAE) [27], alongside DIX, RISE, LTX, EP, and MP, which are applicable to both CNN and ViT architectures. Hyperparameters for all methods were set according to the recommended settings published by the authors, unless a better configuration was found. A detailed description of all explanation methods is provided in Appendix C. Finally, we evaluated three different versions of our SLOC method on both CNN and ViT models: SLOC, which tunes $p$ dynamically per input, $\mathrm{SLOC}_{\mathrm{xp}}$ , where $p$ is calibrated per model (i.e., fixed across all inputs for a given model) based on a validation dataset, and $\mathrm{SLOC}_{\mathrm{m}}$ , which is identical to SLOC but additionally monitors the IDD metric on the specific input and selects the explanation with the best performance according to this metric. All versions were run with $T = 500$ . The precise implementation details, including the monitoring procedure, the $p$ tuning procedures, optimization process, and all hyperparameter configurations, are provided in Appendix D and in our GitHub repository. + +# 4.2. Results + +Tables 1 and 2 present a quantitative comparison of SLOC with other state-of-the-art explanation methods across multiple faithfulness metrics, for DN and ViT-S, respectively. + +
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\( SLOC_m \)11.6176.549.3965.7964.9356.479.2178.35
SLOC11.5270.979.1260.5359.4551.4177.7977.72
\( SLOC_{xp} \)11.7870.209.3459.7558.4350.4177.3376.80
AC17.2467.6813.2757.1850.4443.9177.7875.41
DIX13.3662.9510.3652.4349.5942.0774.6271.47
EP16.1265.6812.9455.049.5542.0677.3874.97
FG19.0644.6615.2637.6225.622.3758.2353.86
GC16.8868.5413.0457.9551.6644.9178.3876.01
GC++17.3667.2913.3456.7549.9343.4178.0475.62
GIG14.8149.9612.2841.9335.1629.6561.1257.6
IG14.1451.9911.244.0837.8532.8861.2658.48
LC17.2867.2713.3156.7149.9943.477.9575.42
LTX16.2471.0912.9559.6954.8446.7478.9276.25
MP18.5453.2414.8743.7934.728.9266.9864.13
RISE18.4262.7514.2652.9944.3338.7376.8274.24
+ +Table 1. Faithfulness results for all combinations of method and metric, using the DN model on the IN dataset. + +
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\( SLOC_m \)14.8381.8112.2770.3166.9858.0483.7883.34
SLOC15.2577.9712.4766.5162.7254.0483.1982.88
\( SLOC_{xp} \)15.7977.8512.8566.1862.0653.3383.0681.97
DIX18.6968.1714.8656.8349.4841.9776.8875.05
EP27.3772.5221.6660.845.1439.1479.4577.56
GAE19.9866.9315.9355.7246.9539.7975.4273.9
LTX20.8468.516.6356.8947.6640.2774.2271.56
MP27.7263.2522.4652.2835.5329.8174.2271.25
RISE29.5169.5223.4757.9340.0234.4679.9377.23
TATTR19.0667.5214.9156.4148.4641.578.0375.69
+ +Table 2. Faithfulness results for all combinations of method and metric, using the ViT-S model on the IN dataset. + +Results for RN and ViT-B are available in Tabs. 7 and 8 in Appendix E, respectively. Overall, the results indicate that all SLOC versions are the best-performing methods, with LTX, DIX, RISE, and EP as runners-up, depending on the combination of architecture and evaluation metric. + +Among the three versions of SLOC, $\mathrm{SLOC}_{\mathrm{m}}$ performs best. This can be attributed to its use of the IDD metric, which measures the difference between INS and DEL performance. The fact that $\mathrm{SLOC}_{\mathrm{m}}$ outperforms across the vast majority of metrics suggests that the IDD metric correlates well with other faithfulness metrics. SLOC and $\mathrm{SLOC}_{\mathrm{xp}}$ follow, with SLOC performing slightly better. This suggests that tuning $p$ per instance offers a modest advantage in metrics, at the cost of increased inference complexity. When a representative sample of data is available, $p$ can be calibrated per model using this data. Otherwise, $p$ should be adjusted dynamically per input during inference. + +Notably, both SLOC and $\mathrm{SLOC}_{\mathrm{xp}}$ surpass learning-based approaches such as LTX, MP, and EP, achieving state-of-the-art results without requiring gradient backpropagation through the model (see the discussion on SLOC complexity in Appendix G) or metric monitoring (as in LTX). Figures 6 and 5 present comparative examples of the top-performing methods for DN and ViT-S. Arguably, SLOC produces focused explanation maps highlighting relevant class-discriminative features in the image. + +Table 3 presents faithfulness results on the VOC dataset + +![](images/21c8ed707eccc1e9584f1e9e63e06c9b1ef8611b7710b02765014f79d133d6d1.jpg) +Figure 4. Qualitative comparison of attributions produced by different methods, using the DN model w.r.t. the classes (top to bottom): 'sea anemone', 'French horn', 'airliner', 'brown bear'. + +![](images/1d4d3509e74a0afee2ec14a636adfc58791fef98b47e496f1499d49e77d090a3.jpg) +Figure 5. Qualitative comparison of attributions produced by different methods, using the ViT-S model w.r.t. the classes (top to bottom): 'American aligator', 'spiny lobster', 'black and gold garden spider', 'promontory'. + +
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\( SLOC_m \)7.3169.25.0748.8361.8943.7675.6379.09
SLOC7.3263.565.0944.5856.2539.574.9677.84
\( SLOC_{xp} \)7.6163.235.2243.7255.6538.4873.6977.59
DIX10.1948.666.7431.7338.4724.9964.5168.66
EP14.0655.969.2636.5541.8927.2968.4372.04
GAE11.247.47.4730.9236.2023.4464.2267.95
LTX12.5649.288.232.2636.7124.0656.5563.16
RISE15.6753.2110.2635.3537.5325.0966.9170.5
T-Attr10.3747.566.8131.3537.1924.5465.6269.53
+ +using ViT-S. The VOC dataset contains multi-object images, allowing multiple classes to appear within the same image. These results demonstrate the superior performance of all SLOC variants not only on single-object images but also on images containing multiple objects from different classes. Results for the DN model are provided in Tab. 9 in Appendix E. + +Tables 4 and 5 present the overall summarized performance scores (higher is better) on the FB benchmark for the RN and ViT-B models, respectively. Detailed fine-grained results are provided in Appendix E. Notably, SLOC + +Table 3. Faithfulness results for combinations of method and metric, using the ViT-S model on the VOC dataset. + +
MethodSLOCACDIXEPFGGCGC++GIGIGLCRISE
Score0.780.700.720.730.700.720.720.560.630.720.62
+ +Table 4. FunnyBirds results for the RN model. + +
MethodSLOCDIXEPRISET-Attr
Score0.880.870.790.770.87
+ +Table 5. FunnyBirds results for the ViT-B model. + +
MethodSLOCACDIXEPGCGC++GIGIGLCLTXRISE
mIoU↑0.560.550.660.520.550.560.510.480.550.560.51
mAP↑0.800.860.840.760.850.850.780.760.850.830.79
PA↑0.770.720.820.720.730.730.740.790.730.510.7
+ +Table 6. Segmentation tests results for the RN model. + +achieves state-of-the-art performance on the FB benchmark as well. The distinct evaluation protocol employed in FB offers a complementary perspective to the faithfulness evaluation conducted on the IN dataset, further reinforcing our confidence in the robustness of SLOC. + +Table 6 presents segmentation tests results on the IN-Seg dataset using the RN model. We observe that DIX achieves the best average performance across all metrics, while the runner-up varies depending on the specific metric. Overall, SLOC is found to be competitive, on average, with all runner-up methods. We note that higher segmentation accuracy may reflect a method's ability to facilitate strong object detection rather than to identify the most informative features that explain the model's prediction. In reality, the most explanatory features do not always cover the entire object; instead, they may correspond to a subset of features that are critical to the model's decision. Consequently, higher segmentation accuracy does not necessarily indicate superior explanatory value. + +Due to space limitation, ablation studies, runtime comparison, sanity checks [3], and additional qualitative analyses, are provided in Appendices F, H, J, and I, respectively. These experiments offer further insights into the effectiveness and efficiency of SLOC. + +# 5. Conclusion + +This work introduced SLOC, a novel and efficient explainability method. By rethinking completeness as a guiding principle promoted locally rather than as a strict global constraint, SLOC exhibits state-of-the-art performance across various explainability benchmarks. These findings suggest that completeness, when promoted in a soft and local manner, provides a robust foundation for generating high-quality attribution maps that closely align with the model's predictive behavior and human comprehension. Finally, discussions on the limitations of SLOC and potential avenues for future research are provided in Appendix M. + +# References + +[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 2 +[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2 +[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, pages 9505-9515, 2018. 1, 8, 12, 25, 26 +[4] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International conference on machine learning, pages 435-451. PMLR, 2022. 1, 2 +[5] Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. In Proceedings of the International Conference on Learning Representations, 2018. 2 +[6] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10 (7):e0130140, 2015. 2 +[7] Oren Barkan, Yonatan Fuchs, Avi Caciularu, and Noam Koenigstein. Explainable recommendations via attentive multi-persona collaborative filtering. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 468-473, 2020. 2 +[8] Oren Barkan, Ori Katz, and Noam Koenigstein. Neural attentive multiview machines. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3357-3361. IEEE, 2020. 1 +[9] Oren Barkan, Idan Rejwan, Avi Caciularu, and Noam Koenigstein. Bayesian hierarchical words representation learning. arXiv preprint arXiv:2004.07126, 2020. 1 +[10] Oren Barkan, Omri Armstrong, Amir Hertz, Avi Caciularu, Ori Katz, Itzik Malkiel, and Noam Koenigstein. Gam: Explainable visual similarity and classification via gradient activation maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 68-77, 2021. 2 +[11] Oren Barkan, Edan Hauon, Avi Caciularu, Ori Katz, Itzik Malkiel, Omri Armstrong, and Noam Koenigstein. Gradsam: Explaining transformers via gradient self-attention maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 2882–2887, 2021. 2 +[12] Oren Barkan, Yuval Asher, Amit Eshel, Yehonatan Elisha, and Noam Koenigstein. Learning to explain: A model-agnostic framework for explaining black box mod + +els. In 2023 IEEE International Conference on Data Mining (ICDM), pages 944-949. IEEE, 2023. 2, 5, 7, 15, 23 +[13] Oren Barkan, Yehonatan Elisha, Yuval Asher, Amit Eshel, and Noam Koenigstein. Visual explanations via iterated integrated attributions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2073-2084, 2023. 2 +[14] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Deep integrated explanations. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 57-67, 2023. 2, 7, 15 +[15] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Stochastic integrated explanations for vision models. In 2023 IEEE International Conference on Data Mining (ICDM), pages 938-943. IEEE, 2023. 2 +[16] Oren Barkan, Tom Shaked, Yonatan Fuchs, and Noam Koenigstein. Modeling users' heterogeneous taste with diversified attentive user profiles. User Modeling and User-Adapted Interaction, pages 1–31, 2023. 2 +[17] Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, and Noam Koenigstein. A counterfactual framework for learning and evaluating explanations for recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3723-3733, 2024. 2 +[18] Oren Barkan, Yehonatan Elisha, Yonatan Toib, Jonathan Weill, and Noam Koenigstein. Improving llm attributions with randomized path-integration. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9430-9446, 2024. 2 +[19] Oren Barkan, Yonatan Toib, Yehonatan Elisha, and Noam Koenigstein. A learning-based approach for explaining language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 98-108, 2024. 2 +[20] Oren Barkan, Yonatan Toib, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Llm explainability via attributive masking learning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9522-9537, 2024. 2 +[21] Oren Barkan, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Bee: Metric-adapted explanations via baseline exploration-exploitation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1835-1843, 2025. 3, 6 +[22] Alexander Binder, Gregoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In International Conference on Artificial Neural Networks, pages 63-71. Springer, 2016. 1 +[23] Alexander Binder, Leander Weber, Sebastian Lapuschkin, Grégoire Montavon, Klaus-Robert Müller, and Wojciech Samek. Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16143-16152, 2023. 6, 25 + +[24] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2 +[25] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1 +[26] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847. IEEE, 2018. 2, 7, 12, 14 +[27] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7, 15 +[28] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 2, 6, 7, 12, 15 +[29] Jianbo Chen, Le Song, Martin J. Wainwright, and Michael I. Jordan. L-shapley and c-shapley: Efficient model interpretation for structured data. In International Conference on Learning Representations, 2019. 2 +[30] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In Advances in Neural Information Processing Systems, pages 6970-6979, 2017. 2, 13 +[31] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Computer Vision and Pattern Recognition (CVPR), 2009. 6, 25 +[32] Saurabh Satish Desai and H. G. Ramaswamy. Ablationcam: Visual explanations for deep convolutional network via gradient-free localization. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 972-980, 2020. 1, 7, 15 +[33] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[34] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1 +[35] Yehonatan Elisha, Oren Barkan, and Noam Koenigstein. Probabilistic path integration with mixture of baseline distributions. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 570-580, 2024. 6 +[36] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving performance of deep + +learning models with axiomatic attribution priors and expected gradients. Nature machine intelligence, 3(7):620-631, 2021. 3, 30 +[37] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 6 +[38] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2950-2958, 2019. 2, 7, 15 +[39] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3429-3437, 2017. 1, 2, 7, 15, 23 +[40] Keren Gaiger, Oren Barkan, Shir Tsipory-Samuel, and Noam Koenigstein. Not all memories created equal: Dynamic user representations for collaborative filtering. IEEE Access, 11: 34746-34763, 2023. 2 +[41] Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR), 51(5):1-42, 2018. 1 +[42] Matthieu Guillaumin, Daniel Kuttel, and Vittorio Ferrari. Imagenet auto-annotation with segmentation propagation. International Journal of Computer Vision, 110(3):328-348, 2014. 6 +[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[44] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods, 2023. 6, 13, 14, 16 +[45] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems, 32, 2019. 6, 7 +[46] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6 +[47] Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4198-4205, 2020. 6 +[48] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 7, 15 +[49] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4948-4957, 2019. 6, 12, 13 + +[50] Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050-5058, 2021. 2, 6, 7, 14 +[51] Sunnie SY Kim, Nicole Meister, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision, pages 280-298. Springer, 2022. 6 +[52] D Liu, R Nicoescu, and R Klette. Stereo-based bokeh effects for photography. Machine Vision and Applications, pages 1-13, 2016. 13 +[53] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pages 4765-4774, 2017. 2 +[54] Itzik Malkiel, Dvir Ginzburg, Oren Barkan, Avi Caciularu, Jonathan Weill, and Noam Koenigstein. Interpreting bert-based text similarity via activation and saliency maps. In Proceedings of the ACM Web Conference 2022, pages 3259-3268, 2022. 2 +[55] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern Recognition, 65:211-222, 2017. 1 +[56] Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, Jörg Schlötterer, Maurice Van Keulen, and Christin Seifert. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai. ACM Computing Surveys, 55(13s): 1-42, 2023. 7 +[57] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2, 3, 6, 7, 12, 15, 16, 24 +[58] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 983-991, 2020. 2 +[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 7, 14 +[60] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3145-3153. JMLR.org, 2017. 1, 2 +[61] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. 25 +[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image + +classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2 +[63] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. In NeurIPS, 2019. 2, 7, 14 +[64] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, pages 3319-3328, 2017. 7, 14 +[65] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319-3328. JMLR.org, 2017. 1, 2, 3, 30 +[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1 +[67] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, 2019. 2 +[68] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 6 +[69] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9680-9689, 2020. 6 \ No newline at end of file diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/images.zip b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08259379a49d8f251f0672d947b39ae229851410 --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c8f14c7a8afed664a316eb6c62db840ca9f145f4811a4912194e5ae0c927889 +size 453754 diff --git a/2025/Soft Local Completeness_ Rethinking Completeness in XAI/layout.json b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d504dd863a1753ef6f77d3c8397d9c81114e36bd --- /dev/null +++ b/2025/Soft Local Completeness_ Rethinking Completeness in XAI/layout.json @@ -0,0 +1,9204 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 122, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 122, + 103, + 488, + 121 + ], + "type": "text", + "content": "Soft Local Completeness: Rethinking Completeness in XAI" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 143, + 253, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 143, + 253, + 156 + ], + "spans": [ + { + "bbox": [ + 73, + 143, + 253, + 156 + ], + "type": "text", + "content": "Ziv Weiss Haddad* Oren Barkan*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 164, + 215, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 164, + 215, + 177 + ], + "spans": [ + { + "bbox": [ + 128, + 164, + 215, + 177 + ], + "type": "text", + "content": "The Open University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 342, + 144, + 535, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 144, + 535, + 157 + ], + "spans": [ + { + "bbox": [ + 342, + 144, + 535, + 157 + ], + "type": "text", + "content": "Yehonatan Elisha Noam Koenigstein" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 398, + 164, + 479, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 164, + 479, + 177 + ], + "spans": [ + { + "bbox": [ + 398, + 164, + 479, + 177 + ], + "type": "text", + "content": "Tel Aviv University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 184, + 398, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 184, + 398, + 196 + ], + "spans": [ + { + "bbox": [ + 209, + 184, + 398, + 196 + ], + "type": "text", + "content": "https://github.com/xaisloc/sloc" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 224, + 200, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 224, + 200, + 236 + ], + "spans": [ + { + "bbox": [ + 152, + 224, + 200, + 236 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 251, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 430 + ], + "type": "text", + "content": "Completeness is a widely discussed property in explainability research, requiring that the attributions sum to the model's response to the input. While completeness intuitively suggests that the model's prediction is \"completely explained\" by the attributions, its global formulation alone is insufficient to ensure faithful explanations. We contend that promoting completeness locally within attribution subregions, in a soft manner, can serve as a standalone guiding principle for producing faithful attributions. To this end, we introduce the concept of the completeness gap as a flexible measure of completeness and propose an optimization procedure that minimizes this gap across subregions within the attribution map. Extensive evaluations across various model architectures demonstrate that our method produces state-of-the-art results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 455, + 135, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 135, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 135, + 467 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 619 + ], + "type": "text", + "content": "In recent years, Explainable AI (XAI) has emerged as a critical aspect of machine learning, especially for deep learning models, which are often regarded as \"black boxes\" due to their lack of transparency [8, 9, 25, 41, 66]. XAI aims to provide insights into how these models make predictions, offering a way to ensure their reliability and trustworthiness [34]. A common approach to model explainability is through attribution maps, which assign importance scores to individual input features based on their contribution to the model's prediction [22, 27, 65]. These attributions serve as a tool for understanding which aspects of the input were most influential in determining the output." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "A widely discussed notion in XAI is completeness (also known as conservation or efficiency) [4, 55, 60, 65]. Completeness is defined as the requirement that the elements in an attribution map should sum to the difference between the model's prediction for a given input and a baseline (e.g., the null representation). This difference is also known as the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 225, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 554, + 332 + ], + "type": "text", + "content": "model's response. When completeness is achieved inherently by an attribution method, it is intuitively appealing, as it suggests that the attribution map fully captures the factors contributing to the model's response. However, even a low-quality attribution map can be superficially adjusted to satisfy completeness through post hoc global normalization. Yet, such an approach lacks genuine explanatory power, as the normalization is decoupled from the underlying attribution mechanism." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 468 + ], + "type": "text", + "content": "This paper recognizes that completeness, as defined in the literature as a global property that an explanation method should satisfy, is too weak to serve as a standalone criterion. Indeed, several well-regarded explanation methods [32, 39, 59] do not inherently satisfy completeness. Nevertheless, these methods have demonstrated considerable effectiveness across diverse objective evaluation metrics, providing faithful explanations and successfully passing various sanity checks [3]. This suggests that completeness is neither a necessary nor sufficient condition for generating high-quality explanations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 471, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 674 + ], + "type": "text", + "content": "In this work, we differentiate between global completeness, as originally defined in the literature, which imposes a completeness constraint on the entire attribution map, and local completeness, which requires completeness within individual subregions of the attribution map. By rethinking completeness as a local and flexible guiding measure rather than a strict global constraint, we seek to overcome its limitations while harnessing the appealing motivation behind it for producing meaningful and faithful explanations. To achieve this, we require subregions of the attribution map, referred to as sub-maps, to promote completeness locally by accounting for the model's response to the corresponding subregions of the input image, yet in soft manner. An attribution map that adheres to this guiding principle should assign high (low) importance to image subregions where the model exhibits a strong (weak) response, thereby maintaining faithfulness to the model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Unlike global completeness, which is a relatively weak requirement that can be easily satisfied, achieving local completeness across an arbitrary set of sub-maps may not" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19794" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "always be feasible. To address this challenge, we introduce the completeness gap - a measure that quantifies the deviation of a sub-map from completeness. Specifically, the completeness gap measures the difference between the sum of elements in each sub-map and the model's response to the corresponding subregions of the input image. Rather than imposing a strict binary constraint, the completeness gap serves as a soft, quantifiable criterion for local completeness. We argue that minimizing this gap over a diverse set of sub-maps leads to faithful explanations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 262 + ], + "type": "text", + "content": "To this end, we present Soft LOcal Completeness (SLOC) - a novel explainability method that promotes completeness locally within sub-maps of the attribution map in a soft and flexible manner. SLOC accomplishes this by seeking to minimize the completeness gap locally for each individual sub-map, simultaneously." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 294, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 294, + 383 + ], + "type": "text", + "content": "The motivation for minimizing the completeness gap locally is to emphasize or attenuate sub-maps based on the actual impact of their corresponding input subregions on the model's output. SLOC achieves this through a gradient-based optimization process, where each sub-map is adjusted to achieve local completeness in a soft manner. Therefore, SLOC's goal is not to enforce strict completeness, globally, but to use it as a guiding principle for refining a large set of sub-maps (subregions within the attribution), iteratively and simultaneously." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "type": "text", + "content": "Our contribution is the introduction of the SLOC method, which facilitates a novel optimization procedure to promote completeness in a soft local manner. The effectiveness of SLOC is demonstrated through extensive experiments on various model architectures, where it is shown to produce state-of-the-art results across multiple benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 294, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 294, + 653 + ], + "type": "text", + "content": "The XAI literature encompasses a broad range of approaches for attributing model predictions to specific input features across different tasks [2, 4, 6, 7, 16, 24, 29, 40, 54]. Early gradient-based methods produce explanation maps by leveraging gradients directly [5, 60, 62] or through functions that combine class activation maps with or without their gradients [10, 26, 48, 58, 59, 63]. With the rise of transformer architectures, new explainability techniques for transformers have emerged [1, 4, 11, 67]. For example, Transformer Attribution [28] introduces a class-specific Deep Taylor Decomposition, applying relevance propagation for both positive and negative attributions. Building on this, Generic Attention Explainability [27] generalizes Transformer Attribution to explain bi-modal transformers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "Path integration methods form another prominent family of attribution techniques [13, 15, 18, 50, 65]. Integrated Gradients (IG) [65] computes attributions by integrating gradients along an interpolation path from a baseline to the input. A refined variant, Guided Integrated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "Gradients (GIG) [50], uses an adaptive path that avoids high-gradient regions, thereby reducing irrelevant attributions. Recently, Deep Integrated Explanations (DIX) [14] proposed performing integration over intermediate network representations instead of inputs, producing more faithful attribution maps than IG and GIG." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 145, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 418 + ], + "type": "text", + "content": "Perturbation-based (occlusion) techniques generate attribution maps by perturbing parts of the input to assess the contributions of specific elements to the prediction [17, 19, 20, 53]. For example, RISE [57], creates perturbations by masking areas in the image through the upsampling of randomly drawn low-resolution binary grids. The class score corresponding to each masked version of the image serves as an importance score for that specific mask. A linear combination of all masks, weighted by their importance, then forms the final attribution map. Another branch of perturbation methods are learning-based perturbation techniques [12, 30, 38, 39]. For example, Meaningful Perturbation (MP) [39] learns to produce perturbations by masking the smallest region possible while significantly altering the original prediction, isolating the minimum necessary content for the prediction. Alternatively, LTX [12] introduces a surrogate 'explainer' model pretrained to mask as much of the input as possible while preserving the original prediction, thereby ensuring the retained features are those most relevant to the model's prediction. Then, LTX finetunes the attribution per specific example, allowing the selection of the best-performing attribution w.r.t. to the metric at hand." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 420, + 553, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 612 + ], + "type": "text", + "content": "Our SLOC method falls within the family of learning-based perturbation techniques, yet it diverges from prior works in several key aspects. First, conceptually, SLOC presents a distinct methodology by promoting local completeness rather than attempting to maximize or minimize the model prediction through learned perturbations [12, 30, 39]. Consequently, SLOC does not require gradient backpropagation through the model, and incorporates the model's predictions on the masked inputs into the loss as constant nodes within the computation graph. This design yields a simple, computationally efficient gradient expression, and relies solely on the model's forward pass, hence operating in a true black-box setting. Finally, unlike LTX, SLOC avoids the overhead of learning an 'explainer' function, which would also require a dataset aligned with the training data distribution of the model being explained." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 455, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 455, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 455, + 639 + ], + "type": "text", + "content": "3. Soft Local Completeness" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^n \\to [0,1]^C" + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": " be a classification model that takes an input image " + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": " and outputs a discrete probability distribution " + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}) \\in [0,1]^C" + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": " classes. Given a target class " + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 644, + 553, + 693 + ], + "type": "text", + "content": " to explain, our goal is to produce an attribution map" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "type": "text", + "content": "W.l.o.g, we represent images as vectors in " + }, + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 325, + 702, + 480, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19795" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y} \\in \\mathbb{R}^{n}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " that attributes the contribution of each element " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " to the prediction score " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})[y]" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " for the class " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "content": "Completeness requires that the attributions in " + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "content": " collectively account for the difference between the model's prediction on the input " + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "content": " and the baseline representation " + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 55, + 96, + 295, + 144 + ], + "type": "text", + "content": ", which is set to the black image in this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": "The model's response to " + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "inline_equation", + "content": "r(\\mathbf{x})\\coloneqq f(\\mathbf{x}) - f(\\mathbf{b})" + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": ". Formally, the attribution map " + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": " satisfies completeness if " + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i] = r(\\mathbf{x})[y]" + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": ". We note that this property can be superficially imposed on an attribution map through a post-processing step by normalizing each element using the factor " + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\frac{r(\\mathbf{x})[y]}{\\sum_{i = 1}^{n}\\mathbf{a}_{\\mathbf{x}}^{y}[i]}" + }, + { + "bbox": [ + 55, + 144, + 296, + 339 + ], + "type": "text", + "content": ". However, artificially satisfying completeness through such normalization does not inherently yield a meaningful explanation, as the imposed completeness is an external adjustment rather than an intrinsic characteristic of the explanation method itself. For example, even a flat attribution map (where all elements are equally weighted) could be normalized to satisfy the completeness criterion, yet such a map lacks any explanatory value. Consequently, relying solely on completeness, in its global form, is deemed insufficient for producing meaningful attribution maps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 130, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 130, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 130, + 358 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 448 + ], + "type": "text", + "content": "Our SLOC approach is motivated by the toy example illustrated in Fig. 1, which depicts an image where the key region influencing the model's prediction for the 'great-grey-owl' class is approximately centered. Figures 1(a)-(d) present four images: the original image (a) and three perturbed versions obtained by masking different regions (b)-(d)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 544 + ], + "type": "text", + "content": "The four different masks applied in Figures 1(a)-(d) are illustrated in Figure 1(e). Each mask selectively perturbs the original image by replacing certain regions with the baseline (in this work, the null pixels). For instance, the mask corresponding to Figure 1(c), shown in the top-right corner of Figure 1(e), applies a perturbation by replacing all pixels within the black rectangle (located at the bottom-right corner) with the baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 604 + ], + "type": "text", + "content": "Notably, the model's response for the great-grey-owl class remains consistent across all four images (a-d), with a value of 0.8. Recall that the model response is defined as the difference between the model's class probability for a given image and that for the baseline (the null image)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "As defined earlier, a sub-map represents a subset of elements in the attribution map that are associated with spe" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 73, + 391, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 73, + 391, + 83 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 391, + 83 + ], + "type": "text", + "content": "(a) Original image: 0.8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 398, + 73, + 470, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 73, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 398, + 73, + 470, + 82 + ], + "type": "text", + "content": "(c) Masked Image: 0.8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 474, + 73, + 523, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 73, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 474, + 73, + 523, + 82 + ], + "type": "text", + "content": "(e) Four masks" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 85, + 395, + 144 + ], + "blocks": [ + { + "bbox": [ + 321, + 85, + 395, + 144 + ], + "lines": [ + { + "bbox": [ + 321, + 85, + 395, + 144 + ], + "spans": [ + { + "bbox": [ + 321, + 85, + 395, + 144 + ], + "type": "image", + "image_path": "b533a8a769d8c12588b6a047ff01364eb955602d2c445a1184a530175715222c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 146, + 391, + 155 + ], + "lines": [ + { + "bbox": [ + 322, + 146, + 391, + 155 + ], + "spans": [ + { + "bbox": [ + 322, + 146, + 391, + 155 + ], + "type": "text", + "content": "(b) Masked Image: 0.8" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 398, + 85, + 472, + 144 + ], + "blocks": [ + { + "bbox": [ + 398, + 85, + 472, + 144 + ], + "lines": [ + { + "bbox": [ + 398, + 85, + 472, + 144 + ], + "spans": [ + { + "bbox": [ + 398, + 85, + 472, + 144 + ], + "type": "image", + "image_path": "fa29e25e40b9a32422368f876b13d79fd6f925d63204c3f28384f530f25bcfde.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 146, + 529, + 155 + ], + "lines": [ + { + "bbox": [ + 475, + 146, + 529, + 155 + ], + "spans": [ + { + "bbox": [ + 475, + 146, + 529, + 155 + ], + "type": "text", + "content": "(f) Attribution Map" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 474, + 85, + 549, + 144 + ], + "blocks": [ + { + "bbox": [ + 474, + 85, + 549, + 144 + ], + "lines": [ + { + "bbox": [ + 474, + 85, + 549, + 144 + ], + "spans": [ + { + "bbox": [ + 474, + 85, + 549, + 144 + ], + "type": "image", + "image_path": "f83fdecf052eb91401b67d7503b9af933f85e31381411d9718162246f9b80f01.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 321, + 157, + 395, + 217 + ], + "blocks": [ + { + "bbox": [ + 399, + 146, + 468, + 154 + ], + "lines": [ + { + "bbox": [ + 399, + 146, + 468, + 154 + ], + "spans": [ + { + "bbox": [ + 399, + 146, + 468, + 154 + ], + "type": "text", + "content": "(d) Masked Image: 0.8" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 157, + 395, + 217 + ], + "lines": [ + { + "bbox": [ + 321, + 157, + 395, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 157, + 395, + 217 + ], + "type": "image", + "image_path": "86d871f7e3f47b7d9e769583d6be8bffc4198da1c7b6ea0c536cec1aea088375.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 398, + 157, + 472, + 217 + ], + "blocks": [ + { + "bbox": [ + 398, + 157, + 472, + 217 + ], + "lines": [ + { + "bbox": [ + 398, + 157, + 472, + 217 + ], + "spans": [ + { + "bbox": [ + 398, + 157, + 472, + 217 + ], + "type": "image", + "image_path": "dc026f536af31b0b5bb11d6d9ec64ef1776710d20a0973f005bad06109110bde.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 225, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. Subfigures (a)-(d) present the model's response (0.8) for the class 'great-grey-owl' when applied to the original input (a) and its masked versions (b)-(d). The four masks used to generate (a)-(d) are depicted in (e). A sub-map refers to a subset of elements in the attribution map that correspond to specific subregions in the input. We consider four distinct sub-maps (not shown), each corresponding to the non-masked (visible) subregions in (a)-(d). For instance, the sub-map associated with (c) includes all elements in the attribution map except those corresponding to the black rectangle in the bottom-right corner of (c). This exclusion aligns with the masked region in the input, effectively determining which attributions are included in that sub-map. A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by (c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to (c). By enforcing local completeness across the four sub-maps induced by (a)-(d), the remaining central stripe, which is unmasked in all images (a)-(d), must account for the model's entire response of 0.8. This results in the attribution depicted in (f) that highlights the central stripe containing the owl in red. See 3.1 for more details." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 474, + 157, + 549, + 217 + ], + "blocks": [ + { + "bbox": [ + 474, + 157, + 549, + 217 + ], + "lines": [ + { + "bbox": [ + 474, + 157, + 549, + 217 + ], + "spans": [ + { + "bbox": [ + 474, + 157, + 549, + 217 + ], + "type": "image", + "image_path": "73a75665b9f038bf3a5911b1cd128e608ad969183787279c05dcab096c01a115.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 471, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 638 + ], + "type": "text", + "content": "cific subregions in the input. In the context of Fig. 1, we consider four different sub-maps, each corresponding to the non-masked (visible) subregions in Figs.1(a)-(d). For instance, the sub-map corresponding to Fig. 1(c) consists of all elements in the attribution map except those associated with the black rectangle in the bottom-right corner of Fig.1(c). This exclusion aligns with the masked region in the input, effectively defining which attributions are considered within that sub-map. Notably, the sub-map corresponding to the original image (Fig. 1(a)) is equivalent to the entire attribution map, as no regions are masked. This formulation allows us to analyze how different subregions contribute to completeness by evaluating attributions over various sub-maps." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "A sub-map satisfies local completeness if the sum of attributions within it equals the model's response for the corresponding perturbed image. For example, the sub-map induced by Fig. 1(c) satisfies local completeness if the sum of its elements equals 0.8, which is the model's response to Fig. 1(c)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 636, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 295, + 713 + ], + "type": "text", + "content": "2The baseline representation should resemble missing information or the neutral representation. While determining the ultimate representation for the baseline is an open question and an active research field in explainability [21, 36], in this work, we opt for the simple choice of a null baseline represented by the black image. This choice is on par with other notable works [57, 65]. Additionally, we conducted experiments using alternative baselines and observed that the resulting trends remained consistent with those obtained using the black image." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19796" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "Our key insight is that enforcing local completeness across the four sub-maps induced by Figs.1(a)-(d) allows us to pinpoint the important region of the image w.r.t. the model's prediction. Specifically, the intersection of these four sub-maps forms a central stripe around the owl, illustrated in red in Fig.1(f). By evenly distributing the total attribution sum of 0.8 across the pixels within this intersection (red stripe) and assigning 0 elsewhere (highlighted in dark blue in Fig. 1(f)), we obtain an attribution map that ensures local completeness for all four sub-maps. Since each sub-map contains the intersection while the attribution values outside it are zero, the sum of attributions within each sub-map is exactly the sum of attributions within the intersection: 0.8. This resulting attribution map effectively captures the key region containing the owl." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": "This simple example can be slightly generalized under similar settings. We make the following assumptions: (1) The model response for a subregion is " + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 254, + 296, + 456 + ], + "type": "text", + "content": " if the subregion contains the object of interest and 0 otherwise. (2) Each sub-map either fully contains the object of interest or does not intersect with it at all. (3) One of the sub-maps corresponds to the complete attribution map. (4) All attribution values are non-negative. It can then be shown that if an attribution map satisfies local completeness for all sub-maps, the intersection defined by these sub-maps is the largest subregion that can have non-zero attribution. Simply put, this intersection highlights the important regions in the image responsible for the model's prediction. Moreover, as the number of sub-maps increases, the important regions can potentially become more precisely defined. The formalization and proof follow straightforwardly and are provided in Appendix K." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": "Keep in mind, however, that this setting and its assumptions are simplistic and unrealistic, serving only to illustrate the motivation. Therefore, SLOC does not actually rely on these assumptions. In more realistic scenarios, satisfying local completeness for all sub-maps is often infeasible (an example is provided in Fig. 11 in Appendix I). To address this challenge, we introduce the completeness gap, which quantifies the degree to which a sub-map deviates from local completeness. Our approach seeks to minimize the completeness gap across multiple sub-maps simultaneously, hence promoting completeness locally within subregions of the attribution map, in a soft manner. We formulate this as an optimization problem and employ gradient descent to find a solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Finally, recall that in our simple example, in the absence of additional information, we distributed the attribution uniformly across the central red stripe. When introducing the loss terms for the SLOC optimization, we will also discuss regularization terms. One such term is the Total Variation loss, which encourages spatial smoothness in the attribution. As we demonstrate in Sec. 4, by encouraging this soft" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "form of local completeness, the resulting attribution map offers robust explanatory performance, both visually and according to a variety of evaluation metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 115, + 430, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 430, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 430, + 128 + ], + "type": "text", + "content": "3.2. SLOC optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 132, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 156 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 132, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\{0,1\\}^n" + }, + { + "bbox": [ + 313, + 132, + 555, + 156 + ], + "type": "text", + "content": " be a binary mask. We define the masked input by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 373, + 156, + 555, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 156, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 373, + 156, + 555, + 170 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\mathbf {m}} = \\mathbf {x} \\circ \\mathbf {m} + (1 - \\mathbf {m}) \\circ \\mathbf {b}, \\tag {1}", + "image_path": "b3fec69f2c70b0c5e0dbca0b77a2ad67921f8fe01e774b460c8613b3beac5107.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": " stands for the elementwise product. In words, " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": " is a perturbation (masked version) of " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": ", according to the mask " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": ", where all masked elements of " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": " are replaced with their respective elements from the baseline " + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 174, + 554, + 221 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "content": "In this work, we consider binary masks constructed over patches, where each patch is randomly assigned a Bernoulli outcome with probability " + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "content": ". We explore two methodologies for automatically selecting " + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "content": ": one that determines " + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "content": " dynamically per input image, and another that uses a fixed value per model. The reader is referred to Appendix D for the exact details of the masks construction process, including the methodologies used for determining " + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 221, + 554, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "text", + "content": "Given an initial attribution map " + }, + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "text", + "content": " and a set of masks " + }, + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\{0,1\\}^n" + }, + { + "bbox": [ + 313, + 317, + 554, + 342 + ], + "type": "text", + "content": ", the SLOC loss is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 348, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 348, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 322, + 348, + 553, + 392 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}; \\mathcal {M}\\right) = \\frac {1}{2 | \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} \\underbrace {(r \\left(\\mathbf {x} ^ {\\mathbf {m}}\\right) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {c o m p l e t e n e s s g a p}} ^ {2}, \\tag {2}", + "image_path": "15721fae54a0ae49c2fca717764ee62fc62cca7c27485c71052f292bdf0be423.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": " is the dot-product operator, and " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": " denotes the L1 norm. As observed in Eq. 2, " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": " promotes the minimization of the completeness gap, a soft measure of completeness, locally, for all sub-maps induced by " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": ". The rationale behind the local minimization of the completeness gap is that attributions within a sub-map should be strengthened or reduced so that their sum closely matches " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "r(\\mathbf{x}^{\\mathbf{m}})[y]" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": " - the model's response to the respective regions exposed in " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": ". This behavior is achieved by applying gradient descent on " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 313, + 393, + 554, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 513, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 536 + ], + "type": "text", + "content": "To better understand the optimization process, one can inspect the gradient:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 543, + 553, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 543, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 326, + 543, + 553, + 572 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y}} = - \\frac {1}{| \\mathcal {M} |} \\sum_ {\\mathbf {m} \\in \\mathcal {M}} \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) \\mathbf {m}. \\tag {3}", + "image_path": "71e91a149f8c0dfff75ac35cf99530615f1ab63428c03c660eae79a5078b587d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": "For simplicity, assume " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": " contains a single mask " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": ". Then, for each element " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}[i]" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": " that is exposed in " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}[i] = 1" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": "), the gradient is the negative of the completeness gap normalized by the number of elements in the sub-map, i.e., " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}_c}{\\partial\\mathbf{a}_x^y[i]} = -\\frac{1}{|\\mathbf{m}|} (r(\\mathbf{x}^{\\mathbf{m}})[y] - \\mathbf{a}_{\\mathbf{x}}^y\\cdot \\mathbf{m})" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": ", and for each " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m}[i] = 0" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": ", the gradient vanishes. Accordingly, the gradient descent update ensures an equal increase (decrease) for all elements in the sub-map when the sum of the attributions falls short of (exceeds) the model's response " + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "inline_equation", + "content": "r(\\mathbf{x}^{\\mathbf{m}})[y]" + }, + { + "bbox": [ + 313, + 580, + 555, + 715 + ], + "type": "text", + "content": ". Assuming a learning rate of 1, it follows that after the gradient descent update, the completeness gap vanishes, resulting in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19797" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "an updated sub-map satisfying completeness. This can be seen mathematically as follows: for a sub-map induced by the mask " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ", the sum of its elements after the gradient update is given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 289, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 289, + 251 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] - \\frac {\\partial \\mathcal {L} _ {c}}{\\partial \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ]} \\\\ = \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] + \\underbrace {\\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m})} _ {\\text {i n d e p e n d e n t o f} i} \\\\ = | \\mathbf {m} | \\frac {1}{| \\mathbf {m} |} (r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m}) + \\sum_ {i: \\mathbf {m} [ i ] = 1} \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i ] \\\\ = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} + \\mathbf {a} _ {\\mathbf {x}} ^ {y} \\cdot \\mathbf {m} = r (\\mathbf {x} ^ {\\mathbf {m}}) [ y ], \\\\ \\end{array}", + "image_path": "51451c7071da3501bcdc42492d35912e7df4832ce7184774acb97340078ee5b8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": "where the second transition follows from the fact that the normalized completeness gap is independent of " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": " and is therefore summed " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "|\\mathbf{m}|" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": " times, as the summation is taken over every " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{m}[i] = 1" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": " being a binary tensor. The penultimate transition then arises from expressing the sum of elements in the sub-map as the dot product, " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y} \\cdot \\mathbf{m}" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": ". This derivation justifies the normalization of the squared completeness gap by " + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "inline_equation", + "content": "|\\mathbf{m}|" + }, + { + "bbox": [ + 55, + 258, + 296, + 378 + ], + "type": "text", + "content": " in Eq. 2, as it ensures that after the gradient update, the sub-map satisfies local completeness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "text", + "content": "To give further intuition for using the normalized completeness gap in Eq. 2, consider two different masks: one that exposes a single pixel (element), and another that exposes a large portion of the input. Given that the gradient update of elements in the sub-map is equal for all elements, it makes sense to distribute the completeness gap among all elements. In the case of a mask exposing a single pixel, we have full confidence that this specific element is responsible for the model's response " + }, + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "inline_equation", + "content": "r(\\mathbf{x}^{\\mathrm{m}})[y]" + }, + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "text", + "content": ". Therefore, it is reasonable to expect that the gradient update for this element will match the size of the completeness gap. However, in the case of a sub-map associated with a mask exposing large portions of the input, we cannot differentiate the accountability of individual elements within the exposed portions for the model's response. Thus, without further information, it is reasonable to apply an update step that evenly distributes the completeness gap across all exposed elements. Hence, the squared completeness gap is divided by " + }, + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "inline_equation", + "content": "|\\mathbf{m}|" + }, + { + "bbox": [ + 55, + 379, + 296, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "So far, the discussion and analysis have been limited to the case of a single mask, i.e., " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}| = 1" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". When considering a set of masks, the update for " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}[i]" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " is determined by the accumulation of gradients from all the sub-maps it is associated with. Specifically, sub-maps that include the " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "-th element and whose total sum of elements is below (exceeds) the model's response " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "r(\\mathbf{x}^{\\mathrm{m}})[y]" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " will contribute to the intensification (reduction) of " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}[i]" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", with the contribution from each individual mask weighted according to the normalized completeness gap it induces." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 83, + 553, + 262 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 463, + 83 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 463, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 463, + 83 + ], + "type": "text", + "content": "Masks & Responses Generation Phase" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 83, + 553, + 262 + ], + "lines": [ + { + "bbox": [ + 316, + 83, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 83, + 553, + 262 + ], + "type": "image", + "image_path": "4249ef80320f2516bc639b72cde85f1ac389b5da773249ce5d9982fa591efaeb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "type": "text", + "content": "Figure 2. SLOC overview: The upper section illustrates the Masks & Responses Generation Phase. In this phase, (1) " + }, + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "type": "text", + "content": " masks are randomly generated. (2) Perturbations are produced by combining the original image and baseline according to the masks, as defined in Eq. (1) (3) These perturbations are passed as inputs to the model, and the corresponding outputs are obtained. (4) Subtracting the model's output for the baseline from these outputs produces the model responses. (5) These masks and their corresponding responses are then passed to the Optimization Phase. (6) Starting with an initial attribution. (7) Gradient descent is used to iteratively update the attribution. The loss function " + }, + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 269, + 555, + 422 + ], + "type": "text", + "content": " (Eq. 4) consists of three components: the completeness gap (computed using the masks, responses, and attributions), as well as TV and L1 regularization. See Sec. 3.2 for details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 441, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 502 + ], + "type": "text", + "content": "To promote smooth and focused attribution maps, we combine " + }, + { + "bbox": [ + 313, + 441, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 313, + 441, + 555, + 502 + ], + "type": "text", + "content": " with two types of regularization applied to the attribution map: Total Variation (TV) and L1 regularization that encourage smoothness and sparsity, respectively. The final loss function is then defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 508, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 508, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 328, + 508, + 555, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) = \\mathcal {L} _ {c} (\\mathbf {a _ {x} ^ {y}}; \\mathcal {M}) + \\lambda_ {1} | \\mathbf {a _ {x} ^ {y}} | + \\lambda_ {2} \\mathrm {T V} (\\mathbf {a _ {x} ^ {y}}), (4)", + "image_path": "817bdec26327be2f706a9f8d6cf8057c0f4af7590d8bb7e843771e2615c1fb06.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 529, + 336, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 336, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 336, + 539 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 568, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 568, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 568, + 582 + ], + "type": "interline_equation", + "content": "\\operatorname {T V} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y}\\right) := \\sum_ {i, j} \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i + 1, j ]\\right) ^ {2} + \\left(\\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j ] - \\mathbf {a} _ {\\mathbf {x}} ^ {y} [ i, j + 1 ]\\right) ^ {2} \\tag {5}", + "image_path": "59b9fa1546e0706804702d16f979ed3192fc11b197561a727337c58389dce296.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "text", + "content": " in Eq. 5 is reshaped to match the spatial dimensions of the final attribution map. Additionally, " + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 582, + 555, + 641 + ], + "type": "text", + "content": " are hyperparameters controlling the strength of the regularization terms. In our experiments, these hyperparameters are set based on a separate validation set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "The optimization of " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathbf{x}}^{y}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is carried out via gradient descent on " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". The optimization process allows for tracking the performance of the metric during the optimization and selecting the best-performing attribution map based on this metric from the resulting maps [12]. A diagram of the SLOC method is depicted in Fig. 2." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19798" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 140 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 140 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 140 + ], + "type": "image", + "image_path": "35fc463f0d7766e6c4af7289daae1f6ac631fc65990a80dc1d83d2907e3e4592.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "type": "text", + "content": "Figure 3. SLOC attribution maps across training steps. Faithful explanations emerge after a few hundred gradient updates, with an appropriate learning rate decay." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 197, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 197, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 295, + 292 + ], + "type": "text", + "content": "In practice, we observed that a few hundred gradient updates, combined with an appropriate learning rate decay, were sufficient to produce faithful and meaningful explanations. Therefore, in our experiments, we report SLOC results for " + }, + { + "bbox": [ + 54, + 197, + 295, + 292 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 54, + 197, + 295, + 292 + ], + "type": "text", + "content": ". An example for SLOC's attribution across gradient update steps is shown in Fig. 3. For the exact implementation details and hyperparameter configuration, the reader is referred to Appendix D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 293, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 387 + ], + "type": "text", + "content": "Due to space limitations, Appendix G provides a detailed discussion of SLOC's computational complexity. Appendix H presents a runtime analysis, comparing SLOC against various methods and demonstrating its efficiency over other state-of-the-art learning-based perturbation methods. Additionally, Appendix L examines additional axioms and explains which of them are satisfied by SLOC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "type": "text", + "content": "Finally, it is important to clarify that the SLOC optimization procedure does not guarantee convergence to an attribution map that fully satisfies completeness. Instead, it leverages the completeness gap as a guiding measure to update subregions within the attribution map, thereby promoting local completeness for these subregions in a soft manner. Nevertheless, as demonstrated in Sec. 4, SLOC has proven effective in generating attribution maps that outperform state-of-the-art explanation methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 198, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 198, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 198, + 520 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 170, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 170, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 170, + 539 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 543, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 543, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 543, + 295, + 662 + ], + "type": "text", + "content": "The experiments were conducted on an NVIDIA DGX 8xA100 Server, utilizing the PyTorch package. As explanations are multifaceted by nature, no single evaluation metric can fully capture their quality, and no definitive metric exists [21, 23, 35, 45, 47, 51, 68]. Therefore, our evaluation encompasses several distinct protocols designed to assess explanation quality from multiple complementary perspectives, on different datasets, and across four model architectures: ResNet50 (RN) [43], DenseNet201 (DN) [46], ViT-Base (ViT-B) and ViT-Small (ViT-S) [33]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "content": "Experiment 1: Faithfulness evaluation This evaluation assesses the faithfulness (also known as correctness or fidelity) of the explanation via a set of perturbation-based" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 360 + ], + "type": "text", + "content": "metrics. These metrics are designed to reveal the actual input elements the model relies on for its predictions. Following prior works [28, 49, 50, 69], we conducted an extensive evaluation using a comprehensive set of perturbation-based explanation metrics to assess the faithfulness of the generated explanations: the Area Under the Curve (AUC) of Positive (POS) and Negative (NEG) perturbation tests [28], AUC of the Insertion (INS) and Deletion (DEL) tests [57], and AUC of the Softmax Information Curve (SIC) and Accuracy Information Curve (AIC) [49]. For POS and DEL, lower values indicate better performance, while for NEG, INS, SIC, and AIC, higher values are preferred. Since NEG and POS, as well as INS and DEL, are complementary metrics, we also report the NEG-POS Difference (NPD) and the INS-DEL Difference (IDD) as single summaries of their respective complementary performances. Faithfulness performance is reported on a sample from the ImageNet [31] ILSVRC 2012 (IN) validation set, which consists of 10,000 images across 1,000 classes. Since most images in the IN dataset contain a single, centrally focused object, we extend our evaluation to multi-object scenes by assessing explanation faithfulness on a sample from the Pascal VOC 2012 [37] (VOC) test set, which includes 1,000 images spanning 20 object classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 639 + ], + "type": "text", + "content": "Experiment 2: FunnyBirds evaluation While perturbation-based protocols are a standard approach for assessing explanation faithfulness, they are criticized for introducing domain shifts that undermine the validity of the evaluation [45]. To address this, we additionally report results on the FunnyBirds (FB) dataset which consists of 500 images from 50 classes, following the evaluation protocol from [44]. The FB evaluation employs controlled interventions to estimate ground-truth (GT) importance scores at the level of object parts rather than individual pixels, mitigating issues associated with pixelwise perturbation-based evaluations [68]. Furthermore, FB avoids domain shifts by incorporating semantically meaningful interventions during training. By evaluating explanations at the part level—closer to human visual understanding—FB reduces the gap between automatic XAI evaluation (e.g., faithfulness metrics) and human-centric studies, providing a more interpretable and robust assessment of the explanations. FB evaluates explainability in three aspects: Completeness, Correctness, and Contrastivity - and provides a combined overall score (higher is better)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Experiment 3: Segmentation evaluation To ensure a comprehensive comparison with prior works [27, 28, 48], we conduct an extensive evaluation using segmentation tests on the ImageNet-Seg dataset (IN-Seg) [42], a subset of the ImageNet validation set comprising 4,276 human-annotated" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19799" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "ground-truth (GT) segmentation maps across 445 classes. As detailed in [28], these tests generalize the \"Pointing Game\" [45] and serve as a human-grounded evaluation [56] by measuring the alignment between GT segmentation maps and the attributions generated by the explanation method. While higher segmentation accuracy does not necessarily indicate greater explanatory power [57], segmentation tests assess the alignment between explanations and human-annotated GT, hence providing a complementary perspective to other evaluations. We follow the evaluation protocol from [28], assessing segmentation performance using Pixel Accuracy (PA), mean Intersection-over-Union (mIoU), and mean Average Precision (mAP), higher is better for all metrics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 277 + ], + "type": "text", + "content": "Detailed descriptions of the evaluation protocols and metrics used in Experiments 1, 2, and 3 are provided in Appendices B.1, B.2, and B.3, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": "Evaluated methods We evaluate SLOC against a comprehensive suite of 14 explanation methods encompassing gradient-based, path-integration, perturbation, and CAM methods. For CNN models, we include the following methods: Grad-CAM (GC) [59], Grad-CAM++ " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "(\\mathbf{GC} + +)" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": " [26], Deep Integrated Explanations (DIX) [14], FullGrad (FG) [63], Ablation-CAM (AC) [32], Layer-CAM (LC) [48], Learning To Explain (LTX) [12], RISE [57], Meaningful Perturbation (MP) [39], Extremal Perturbations (EP) [38], Integrated Gradients (IG) [64], and Guided IG (GIG) [50]. For ViT models, we considered the following methods: Transformer Attribution (T-Attr) [28] and Generic Attention Explainability (GAE) [27], alongside DIX, RISE, LTX, EP, and MP, which are applicable to both CNN and ViT architectures. Hyperparameters for all methods were set according to the recommended settings published by the authors, unless a better configuration was found. A detailed description of all explanation methods is provided in Appendix C. Finally, we evaluated three different versions of our SLOC method on both CNN and ViT models: SLOC, which tunes " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": " dynamically per input, " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{xp}}" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": " is calibrated per model (i.e., fixed across all inputs for a given model) based on a validation dataset, and " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{m}}" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": ", which is identical to SLOC but additionally monitors the IDD metric on the specific input and selects the explanation with the best performance according to this metric. All versions were run with " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "T = 500" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": ". The precise implementation details, including the monitoring procedure, the " + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 56, + 281, + 296, + 653 + ], + "type": "text", + "content": " tuning procedures, optimization process, and all hyperparameter configurations, are provided in Appendix D and in our GitHub repository." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 659, + 113, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 113, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 113, + 672 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Tables 1 and 2 present a quantitative comparison of SLOC with other state-of-the-art explanation methods across multiple faithfulness metrics, for DN and ViT-S, respectively." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 566, + 208 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 566, + 208 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 566, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 566, + 208 + ], + "type": "table", + "html": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)11.6176.549.3965.7964.9356.479.2178.35
SLOC11.5270.979.1260.5359.4551.4177.7977.72
\\( SLOC_{xp} \\)11.7870.209.3459.7558.4350.4177.3376.80
AC17.2467.6813.2757.1850.4443.9177.7875.41
DIX13.3662.9510.3652.4349.5942.0774.6271.47
EP16.1265.6812.9455.049.5542.0677.3874.97
FG19.0644.6615.2637.6225.622.3758.2353.86
GC16.8868.5413.0457.9551.6644.9178.3876.01
GC++17.3667.2913.3456.7549.9343.4178.0475.62
GIG14.8149.9612.2841.9335.1629.6561.1257.6
IG14.1451.9911.244.0837.8532.8861.2658.48
LC17.2867.2713.3156.7149.9943.477.9575.42
LTX16.2471.0912.9559.6954.8446.7478.9276.25
MP18.5453.2414.8743.7934.728.9266.9864.13
RISE18.4262.7514.2652.9944.3338.7376.8274.24
", + "image_path": "b34f2d404e7a71d8bfc772af98fc76d041ea4179f6613b74e8afd3377dbb3026.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 244, + 569, + 340 + ], + "blocks": [ + { + "bbox": [ + 314, + 209, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 314, + 209, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 555, + 232 + ], + "type": "text", + "content": "Table 1. Faithfulness results for all combinations of method and metric, using the DN model on the IN dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 244, + 569, + 340 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 569, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 569, + 340 + ], + "type": "table", + "html": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)14.8381.8112.2770.3166.9858.0483.7883.34
SLOC15.2577.9712.4766.5162.7254.0483.1982.88
\\( SLOC_{xp} \\)15.7977.8512.8566.1862.0653.3383.0681.97
DIX18.6968.1714.8656.8349.4841.9776.8875.05
EP27.3772.5221.6660.845.1439.1479.4577.56
GAE19.9866.9315.9355.7246.9539.7975.4273.9
LTX20.8468.516.6356.8947.6640.2774.2271.56
MP27.7263.2522.4652.2835.5329.8174.2271.25
RISE29.5169.5223.4757.9340.0234.4679.9377.23
TATTR19.0667.5214.9156.4148.4641.578.0375.69
", + "image_path": "cfd6aae8bc9b8a0dc1af31383467bd0143a3e6682a136976cb7912e76a99399b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 343, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 365 + ], + "type": "text", + "content": "Table 2. Faithfulness results for all combinations of method and metric, using the ViT-S model on the IN dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "content": "Results for RN and ViT-B are available in Tabs. 7 and 8 in Appendix E, respectively. Overall, the results indicate that all SLOC versions are the best-performing methods, with LTX, DIX, RISE, and EP as runners-up, depending on the combination of architecture and evaluation metric." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": "Among the three versions of SLOC, " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " performs best. This can be attributed to its use of the IDD metric, which measures the difference between INS and DEL performance. The fact that " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " outperforms across the vast majority of metrics suggests that the IDD metric correlates well with other faithfulness metrics. SLOC and " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{xp}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " follow, with SLOC performing slightly better. This suggests that tuning " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " per instance offers a modest advantage in metrics, at the cost of increased inference complexity. When a representative sample of data is available, " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " can be calibrated per model using this data. Otherwise, " + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 449, + 556, + 592 + ], + "type": "text", + "content": " should be adjusted dynamically per input during inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 593, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 700 + ], + "type": "text", + "content": "Notably, both SLOC and " + }, + { + "bbox": [ + 313, + 593, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{SLOC}_{\\mathrm{xp}}" + }, + { + "bbox": [ + 313, + 593, + 556, + 700 + ], + "type": "text", + "content": " surpass learning-based approaches such as LTX, MP, and EP, achieving state-of-the-art results without requiring gradient backpropagation through the model (see the discussion on SLOC complexity in Appendix G) or metric monitoring (as in LTX). Figures 6 and 5 present comparative examples of the top-performing methods for DN and ViT-S. Arguably, SLOC produces focused explanation maps highlighting relevant class-discriminative features in the image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "Table 3 presents faithfulness results on the VOC dataset" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19800" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 301, + 201 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 301, + 201 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 301, + 201 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 301, + 201 + ], + "type": "image", + "image_path": "21c8ed707eccc1e9584f1e9e63e06c9b1ef8611b7710b02765014f79d133d6d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 296, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 237 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of attributions produced by different methods, using the DN model w.r.t. the classes (top to bottom): 'sea anemone', 'French horn', 'airliner', 'brown bear'." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 249, + 301, + 380 + ], + "blocks": [ + { + "bbox": [ + 57, + 249, + 301, + 380 + ], + "lines": [ + { + "bbox": [ + 57, + 249, + 301, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 249, + 301, + 380 + ], + "type": "image", + "image_path": "1d4d3509e74a0afee2ec14a636adfc58791fef98b47e496f1499d49e77d090a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 384, + 296, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 429 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of attributions produced by different methods, using the ViT-S model w.r.t. the classes (top to bottom): 'American aligator', 'spiny lobster', 'black and gold garden spider', 'promontory'." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 445, + 304, + 533 + ], + "blocks": [ + { + "bbox": [ + 58, + 445, + 304, + 533 + ], + "lines": [ + { + "bbox": [ + 58, + 445, + 304, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 445, + 304, + 533 + ], + "type": "table", + "html": "
MethodPOS↓NEG↑DEL↓INS↑NPD↑IDD↑AIC↑SIC↑
\\( SLOC_m \\)7.3169.25.0748.8361.8943.7675.6379.09
SLOC7.3263.565.0944.5856.2539.574.9677.84
\\( SLOC_{xp} \\)7.6163.235.2243.7255.6538.4873.6977.59
DIX10.1948.666.7431.7338.4724.9964.5168.66
EP14.0655.969.2636.5541.8927.2968.4372.04
GAE11.247.47.4730.9236.2023.4464.2267.95
LTX12.5649.288.232.2636.7124.0656.5563.16
RISE15.6753.2110.2635.3537.5325.0966.9170.5
T-Attr10.3747.566.8131.3537.1924.5465.6269.53
", + "image_path": "76af45be9a434eb5bc482e68f05f7c68cbd7525042adf56306a5aaf55495384f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "content": "using ViT-S. The VOC dataset contains multi-object images, allowing multiple classes to appear within the same image. These results demonstrate the superior performance of all SLOC variants not only on single-object images but also on images containing multiple objects from different classes. Results for the DN model are provided in Tab. 9 in Appendix E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Tables 4 and 5 present the overall summarized performance scores (higher is better) on the FB benchmark for the RN and ViT-B models, respectively. Detailed fine-grained results are provided in Appendix E. Notably, SLOC" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 574, + 88 + ], + "blocks": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 559 + ], + "type": "text", + "content": "Table 3. Faithfulness results for combinations of method and metric, using the ViT-S model on the VOC dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 69, + 574, + 88 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 574, + 88 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 574, + 88 + ], + "type": "table", + "html": "
MethodSLOCACDIXEPFGGCGC++GIGIGLCRISE
Score0.780.700.720.730.700.720.720.560.630.720.62
", + "image_path": "3f85b8bbd97eb43a97563c7a4468fc84af137ba9d3c5b2179165c9c874b96948.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 349, + 114, + 520, + 137 + ], + "blocks": [ + { + "bbox": [ + 349, + 91, + 518, + 102 + ], + "lines": [ + { + "bbox": [ + 349, + 91, + 518, + 102 + ], + "spans": [ + { + "bbox": [ + 349, + 91, + 518, + 102 + ], + "type": "text", + "content": "Table 4. FunnyBirds results for the RN model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 114, + 520, + 137 + ], + "lines": [ + { + "bbox": [ + 349, + 114, + 520, + 137 + ], + "spans": [ + { + "bbox": [ + 349, + 114, + 520, + 137 + ], + "type": "table", + "html": "
MethodSLOCDIXEPRISET-Attr
Score0.880.870.790.770.87
", + "image_path": "d945241084ec291498353bfb3bb4d01880017fe4f2d4cc7d885c21da0d9cf475.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 161, + 582, + 196 + ], + "blocks": [ + { + "bbox": [ + 345, + 139, + 523, + 149 + ], + "lines": [ + { + "bbox": [ + 345, + 139, + 523, + 149 + ], + "spans": [ + { + "bbox": [ + 345, + 139, + 523, + 149 + ], + "type": "text", + "content": "Table 5. FunnyBirds results for the ViT-B model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 161, + 582, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 161, + 582, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 582, + 196 + ], + "type": "table", + "html": "
MethodSLOCACDIXEPGCGC++GIGIGLCLTXRISE
mIoU↑0.560.550.660.520.550.560.510.480.550.560.51
mAP↑0.800.860.840.760.850.850.780.760.850.830.79
PA↑0.770.720.820.720.730.730.740.790.730.510.7
", + "image_path": "b5ba4108174f1d0242e0c781bd1515ca4918820960b7864d9b3f2b57441925d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 198, + 531, + 209 + ], + "lines": [ + { + "bbox": [ + 336, + 198, + 531, + 209 + ], + "spans": [ + { + "bbox": [ + 336, + 198, + 531, + 209 + ], + "type": "text", + "content": "Table 6. Segmentation tests results for the RN model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 231, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 290 + ], + "type": "text", + "content": "achieves state-of-the-art performance on the FB benchmark as well. The distinct evaluation protocol employed in FB offers a complementary perspective to the faithfulness evaluation conducted on the IN dataset, further reinforcing our confidence in the robustness of SLOC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 291, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 458 + ], + "type": "text", + "content": "Table 6 presents segmentation tests results on the IN-Seg dataset using the RN model. We observe that DIX achieves the best average performance across all metrics, while the runner-up varies depending on the specific metric. Overall, SLOC is found to be competitive, on average, with all runner-up methods. We note that higher segmentation accuracy may reflect a method's ability to facilitate strong object detection rather than to identify the most informative features that explain the model's prediction. In reality, the most explanatory features do not always cover the entire object; instead, they may correspond to a subset of features that are critical to the model's decision. Consequently, higher segmentation accuracy does not necessarily indicate superior explanatory value." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 518 + ], + "type": "text", + "content": "Due to space limitation, ablation studies, runtime comparison, sanity checks [3], and additional qualitative analyses, are provided in Appendices F, H, J, and I, respectively. These experiments offer further insights into the effectiveness and efficiency of SLOC." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 388, + 555 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 563, + 555, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 695 + ], + "type": "text", + "content": "This work introduced SLOC, a novel and efficient explainability method. By rethinking completeness as a guiding principle promoted locally rather than as a strict global constraint, SLOC exhibits state-of-the-art performance across various explainability benchmarks. These findings suggest that completeness, when promoted in a soft and local manner, provides a robust foundation for generating high-quality attribution maps that closely align with the model's predictive behavior and human comprehension. Finally, discussions on the limitations of SLOC and potential avenues for future research are provided in Appendix M." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19801" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 122 + ], + "type": "text", + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXiv preprint arXiv:2005.00928, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "text", + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems, pages 9505-9515, 2018. 1, 8, 12, 25, 26" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International conference on machine learning, pages 435-451. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 349 + ], + "type": "text", + "content": "[5] Marco Ancona, Enea Ceolini, Cengiz Oztireli, and Markus Gross. Towards better understanding of gradient-based attribution methods for deep neural networks. In Proceedings of the International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "text", + "content": "[6] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. *PloS one*, 10 (7):e0130140, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 462 + ], + "type": "text", + "content": "[7] Oren Barkan, Yonatan Fuchs, Avi Caciularu, and Noam Koenigstein. Explainable recommendations via attentive multi-persona collaborative filtering. In Proceedings of the 14th ACM Conference on Recommender Systems, pages 468-473, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 509 + ], + "type": "text", + "content": "[8] Oren Barkan, Ori Katz, and Noam Koenigstein. Neural attentive multiview machines. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3357-3361. IEEE, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 295, + 544 + ], + "type": "text", + "content": "[9] Oren Barkan, Idan Rejwan, Avi Caciularu, and Noam Koenigstein. Bayesian hierarchical words representation learning. arXiv preprint arXiv:2004.07126, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 610 + ], + "type": "text", + "content": "[10] Oren Barkan, Omri Armstrong, Amir Hertz, Avi Caciularu, Ori Katz, Itzik Malkiel, and Noam Koenigstein. Gam: Explainable visual similarity and classification via gradient activation maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 68-77, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 677 + ], + "type": "text", + "content": "[11] Oren Barkan, Edan Hauon, Avi Caciularu, Ori Katz, Itzik Malkiel, Omri Armstrong, and Noam Koenigstein. Gradsam: Explaining transformers via gradient self-attention maps. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 2882–2887, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Oren Barkan, Yuval Asher, Amit Eshel, Yehonatan Elisha, and Noam Koenigstein. Learning to explain: A model-agnostic framework for explaining black box mod" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "els. In 2023 IEEE International Conference on Data Mining (ICDM), pages 944-949. IEEE, 2023. 2, 5, 7, 15, 23" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[13] Oren Barkan, Yehonatan Elisha, Yuval Asher, Amit Eshel, and Noam Koenigstein. Visual explanations via iterated integrated attributions. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 2073-2084, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 196 + ], + "type": "text", + "content": "[14] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Deep integrated explanations. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 57-67, 2023. 2, 7, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 251 + ], + "type": "text", + "content": "[15] Oren Barkan, Yehonatan Elisha, Jonathan Weill, Yuval Asher, Amit Eshel, and Noam Koenigstein. Stochastic integrated explanations for vision models. In 2023 IEEE International Conference on Data Mining (ICDM), pages 938-943. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 255, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 298 + ], + "type": "text", + "content": "[16] Oren Barkan, Tom Shaked, Yonatan Fuchs, and Noam Koenigstein. Modeling users' heterogeneous taste with diversified attentive user profiles. User Modeling and User-Adapted Interaction, pages 1–31, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 354 + ], + "type": "text", + "content": "[17] Oren Barkan, Veronika Bogina, Liya Gurevitch, Yuval Asher, and Noam Koenigstein. A counterfactual framework for learning and evaluating explanations for recommender systems. In Proceedings of the ACM Web Conference 2024, pages 3723-3733, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 356, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 356, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 553, + 409 + ], + "type": "text", + "content": "[18] Oren Barkan, Yehonatan Elisha, Yonatan Toib, Jonathan Weill, and Noam Koenigstein. Improving llm attributions with randomized path-integration. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9430-9446, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 412, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 553, + 466 + ], + "type": "text", + "content": "[19] Oren Barkan, Yonatan Toib, Yehonatan Elisha, and Noam Koenigstein. A learning-based approach for explaining language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 98-108, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[20] Oren Barkan, Yonatan Toib, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Llm explainability via attributive masking learning. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9522-9537, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "type": "text", + "content": "[21] Oren Barkan, Yehonatan Elisha, Jonathan Weill, and Noam Koenigstein. Bee: Metric-adapted explanations via baseline exploration-exploitation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1835-1843, 2025. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "text", + "content": "[22] Alexander Binder, Gregoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In International Conference on Artificial Neural Networks, pages 63-71. Springer, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 712 + ], + "type": "text", + "content": "[23] Alexander Binder, Leander Weber, Sebastian Lapuschkin, Grégoire Montavon, Klaus-Robert Müller, and Wojciech Samek. Shortcomings of top-down randomization-based sanity checks for evaluations of deep neural network explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16143-16152, 2023. 6, 25" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19802" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[24] Moritz Böhle, Mario Fritz, and Bernt Schiele. B-cos networks: Alignment is all we need for interpretability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10329-10338, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[25] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "text", + "content": "[26] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847. IEEE, 2018. 2, 7, 12, 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[27] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 331 + ], + "type": "text", + "content": "[28] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 2, 6, 7, 12, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[29] Jianbo Chen, Le Song, Martin J. Wainwright, and Michael I. Jordan. L-shapley and c-shapley: Efficient model interpretation for structured data. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 410 + ], + "type": "text", + "content": "[30] Piotr Dabkowski and Yarin Gal. Real time image saliency for black box classifiers. In Advances in Neural Information Processing Systems, pages 6970-6979, 2017. 2, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[31] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In Computer Vision and Pattern Recognition (CVPR), 2009. 6, 25" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[32] Saurabh Satish Desai and H. G. Ramaswamy. Ablationcam: Visual explanations for deep convolutional network via gradient-free localization. 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 972-980, 2020. 1, 7, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[33] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[34] Rudresh Dwivedi, Devam Dave, Het Naik, Smiti Singhal, Rana Omer, Pankesh Patel, Bin Qian, Zhenyu Wen, Tejal Shah, Graham Morgan, et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9):1-33, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Yehonatan Elisha, Oren Barkan, and Noam Koenigstein. Probabilistic path integration with mixture of baseline distributions. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 570-580, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 715 + ], + "type": "text", + "content": "[36] Gabriel Erion, Joseph D Janizek, Pascal Sturmfels, Scott M Lundberg, and Su-In Lee. Improving performance of deep" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "learning models with axiomatic attribution priors and expected gradients. Nature machine intelligence, 3(7):620-631, 2021. 3, 30" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[37] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[38] Ruth Fong, Mandela Patrick, and Andrea Vedaldi. Understanding deep networks via extremal perturbations and smooth masks. In Proceedings of the IEEE International Conference on Computer Vision, pages 2950-2958, 2019. 2, 7, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 263 + ], + "type": "text", + "content": "[39] Ruth C Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3429-3437, 2017. 1, 2, 7, 15, 23" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[40] Keren Gaiger, Oren Barkan, Shir Tsipory-Samuel, and Noam Koenigstein. Not all memories created equal: Dynamic user representations for collaborative filtering. IEEE Access, 11: 34746-34763, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[41] Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR), 51(5):1-42, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[42] Matthieu Guillaumin, Daniel Kuttel, and Vittorio Ferrari. Imagenet auto-annotation with segmentation propagation. International Journal of Computer Vision, 110(3):328-348, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "text", + "content": "[44] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods, 2023. 6, 13, 14, 16" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 521 + ], + "type": "text", + "content": "[45] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in neural information processing systems, 32, 2019. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 554, + 567 + ], + "type": "text", + "content": "[46] Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 622 + ], + "type": "text", + "content": "[47] Alon Jacovi and Yoav Goldberg. Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4198-4205, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[48] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 7, 15" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 715 + ], + "type": "text", + "content": "[49] Andrei Kapishnikov, Tolga Bolukbasi, Fernanda Viégas, and Michael Terry. Xrai: Better attributions through regions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4948-4957, 2019. 6, 12, 13" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19803" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[50] Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050-5058, 2021. 2, 6, 7, 14" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[51] Sunnie SY Kim, Nicole Meister, Vikram V Ramaswamy, Ruth Fong, and Olga Russakovsky. Hive: Evaluating the human interpretability of visual explanations. In European Conference on Computer Vision, pages 280-298. Springer, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "text", + "content": "[52] D Liu, R Nicoescu, and R Klette. Stereo-based bokeh effects for photography. Machine Vision and Applications, pages 1-13, 2016. 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[53] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pages 4765-4774, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[54] Itzik Malkiel, Dvir Ginzburg, Oren Barkan, Avi Caciularu, Jonathan Weill, and Noam Koenigstein. Interpreting bert-based text similarity via activation and saliency maps. In Proceedings of the ACM Web Conference 2022, pages 3259-3268, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[55] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern Recognition, 65:211-222, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "text", + "content": "[56] Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, Jörg Schlötterer, Maurice Van Keulen, and Christin Seifert. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai. ACM Computing Surveys, 55(13s): 1-42, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "text", + "content": "[57] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. arXiv preprint arXiv:1806.07421, 2018. 2, 3, 6, 7, 12, 15, 16, 24" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[58] Harish Guruprasad Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 983-991, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 600 + ], + "type": "text", + "content": "[59] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 7, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "text", + "content": "[60] Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. Learning important features through propagating activation differences. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3145-3153. JMLR.org, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[61] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. 25" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 715 + ], + "type": "text", + "content": "[62] Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 398 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "classification models and saliency maps. arXiv preprint arXiv:1312.6034, 2013. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[63] Suraj Srinivas and François Fleuret. Full-gradient representation for neural network visualization. In NeurIPS, 2019. 2, 7, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[64] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, pages 3319-3328, 2017. 7, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[65] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319-3328. JMLR.org, 2017. 1, 2, 3, 30" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[67] Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[68] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "text", + "content": "[69] Shawn Xu, Subhashini Venugopalan, and Mukund Sundararajan. Attribution in scale and space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9680-9689, 2020. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19804" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_content_list.json b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd4b37ea983c801d450109c7a6a061c4a9ef7818 --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_content_list.json @@ -0,0 +1,1638 @@ +[ + { + "type": "text", + "text": "Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning", + "text_level": 1, + "bbox": [ + 184, + 128, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hung-Chieh Fang $^{1,*}$ Hsuan-Tien Lin $^{1}$ Irwin King $^{2}$ Yifei Zhang $^{2,\\dagger}$", + "bbox": [ + 215, + 202, + 784, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Taiwan University $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 223, + 238, + 767, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated Unsupervised Learning (FUL) aims to learn expressive representations in federated and self-supervised settings. The quality of representations learned in FUL is usually determined by uniformity, a measure of how uniformly representations are distributed in the embedding space. However, existing solutions perform well in achieving intra-client (local) uniformity for local models while failing to achieve inter-client (global) uniformity after aggregation due to non-IID data distributions and the decentralized nature of FUL. To address this issue, we propose Soft Separation and Distillation (SSD), a novel approach that preserves inter-client uniformity by encouraging client representations to spread toward different directions. This design reduces interference during client model aggregation, thereby improving global uniformity while preserving local representation expressiveness. We further enhance this effect by introducing a projector distillation module to address the discrepancy between loss optimization and representation quality. We evaluate SSD in both cross-silo and cross-device federated settings, demonstrating consistent improvements in representation quality and task performance across various training scenarios. Our results highlight the importance of inter-client uniformity in FUL and establish SSD as an effective solution to this challenge. Project page: https://ssd-uniformity.github.io/.", + "bbox": [ + 86, + 323, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved remarkable success across a wide spectrum of applications, from computer vision and natural language processing to speech recognition and reinforcement learning [1, 11, 18, 29]. This progress can be largely attributed to the availability of massive labeled datasets, particularly since the emergence of ImageNet [5],", + "bbox": [ + 89, + 772, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9de7bcfe6166e3e1f81d9057d7b69aedb3259c202278684a2c16a5c780909ccb.jpg", + "image_caption": [ + "Figure 1. Comparison with state-of-the-art methods. Our SSD achieves the highest performance (marked in brackets) and the best representation quality, measured by uniformity [31] and effective rank [27], among existing FUL approaches." + ], + "image_footnote": [], + "bbox": [ + 550, + 323, + 836, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which has enabled the training of increasingly powerful neural networks. However, the practical deployment of deep learning in real-world scenarios faces two critical challenges. First, data in real-world applications is often non-Independent and Identically Distributed (non-IID) and cannot be freely shared due to privacy concerns, regulatory restrictions, or proprietary constraints. This data is typically generated across diverse sources, including user devices, healthcare institutions, and industrial settings. The distributed nature and sharing limitations of such data make centralized training approaches infeasible, leading to the development of Federated Learning (FL) [15, 20, 24] as a privacy-preserving distributed learning paradigm. Second, a vast proportion of available data remains unlabeled, necessitating effective methods for unsupervised representation learning [2, 3, 8, 12]. These methods aim to learn meaningful feature representations without relying on explicit la", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Project was partially completed during the author's visit to the CUHK", + "bbox": [ + 107, + 875, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2971", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bels, enabling models to capture underlying data structures and semantic relationships.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated Unsupervised Learning (FUL) emerges at the intersection of these two challenges, aiming to learn expressive representations in settings where data is both distributed and unlabeled. In FUL, the quality of learned representations is critically determined by two properties: alignment and uniformity [31]. Alignment measures how close similar data points are positioned in the representation space, relating to the model's ability to group semantically similar objects. Uniformity quantifies how evenly representations are distributed across the unit hypersphere, essentially measuring the entropy of the representation distribution. High uniformity prevents representation collapse and ensures that the learned features effectively utilize the available embedding dimensions.", + "bbox": [ + 88, + 121, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing FUL methods face a significant challenge that remains inadequately addressed. While current approaches successfully achieve good intra-client (local) uniformity for representations within each client, they struggle to maintain inter-client (global) uniformity after model aggregation. This limitation stems from two key factors: (1) the non-IID distribution of data across clients, which naturally leads to divergent updates, and (2) the decentralized nature of FL, where the server lacks direct access to raw data, preventing the application of explicit uniformity constraints across clients.", + "bbox": [ + 89, + 333, + 480, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing approaches have focused primarily on two aspects of this challenge. One line of research builds upon the framework established by FedProx [21], introducing proximal terms to constrain local updates within a bounded neighborhood of the global model. For instance, FedU [36] and FedEMA [37] dynamically adjust aggregation weights based on inter-client model divergence, while FedX [10] incorporates a global relational loss to align pairwise sample relationships across clients. These methods aim to maintain global model consistency but do not explicitly address representation uniformity.", + "bbox": [ + 89, + 500, + 480, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another research direction tackles the issue of dimensional collapse [14] caused by lower uniformity in local representations. FedDecorr [28] demonstrates that embeddings of local clients are often less uniformly distributed and mitigates this issue by decorrelating local features. Similarly, FedU2 [22] regularizes local updates to approximate a spherical Gaussian distribution, encouraging isotropic feature spaces. While these approaches enhance local uniformity, the improvements at the client level do not inherently translate to better global uniformity during model aggregation. This leads to a crucial question:", + "bbox": [ + 89, + 667, + 480, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "How can we effectively improve inter-client uniformity in Federated Unsupervised Learning?", + "bbox": [ + 98, + 844, + 475, + 873 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose Soft Separation", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and Distillation (SSD), a novel method that enhances inter-client uniformity without compromising local representation expressiveness. Our approach employs a dimension-scaled regularization strategy that softly separates each client's feature space, encouraging client representations to spread toward different directions. As illustrated in Figure 2, this technique reduces interference during client model aggregation, thereby improving global uniformity without imposing rigid boundaries that could distort the underlying feature distributions.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we empirically observe that the regularization effect of SSD may not always effectively transfer from loss optimization to the representation space due to the presence of a projector between the loss function and the encoder. While removing the projector might seem like a straightforward solution, doing so often degrades performance because the projector plays a crucial role in separating optimization objectives from feature representations [2, 9, 32]. To bridge this gap, we introduce a projector distillation module that minimizes the KL divergence between representations and embeddings, effectively encouraging the encoder to internalize the learned structure while preserving the projector's role in loss optimization.", + "bbox": [ + 511, + 243, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate SSD across diverse federated learning scenarios, including cross-silo (few clients with large datasets) and cross-device (many clients with small datasets) settings. Our experiments span both in-distribution tasks and out-of-distribution (OOD) datasets to assess generalizability. Results demonstrate that SSD consistently outperforms existing methods in both downstream task performance (measured via linear probing and fine-tuning accuracy) and representation quality (quantified using effective rank [7] and uniformity [31] metrics).", + "bbox": [ + 511, + 441, + 903, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are threefold:", + "bbox": [ + 532, + 593, + 781, + 606 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify and formalize the challenge of inter-client uniformity in FUL, establishing it as a critical direction for decentralized unsupervised representation learning.", + "- We propose Soft Separation and Distillation (SSD), a simple yet effective framework that addresses inter-client uniformity without additional communication overhead or compromising privacy.", + "- We conduct extensive experiments across varied FL settings and tasks, confirming SSD's superiority over baseline methods in both performance and robustness." + ], + "bbox": [ + 511, + 608, + 903, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 777, + 651, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Federated unsupervised learning (FUL) is a learning paradigm in which multiple clients collaboratively train a model without sharing new raw data, and where the local datasets contain unlabeled data. The goal is to learn a global representation that generalizes across all participating clients. Consider a federated learning system with $K$", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2972", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8263cb1e601bf03ab9cf5331c720d3c6a1a01f86d271957f17eba7846c89dc31.jpg", + "image_caption": [ + "Figure 2. Illustration of Intra-Client and Inter-Client Uniformity with Soft Separation. Intra-client uniformity ensures that representations within each client are well-distributed, while inter-client uniformity promotes global representation consistency across clients. Our proposed Soft Separation method encourages each client's representations to spread in distinct directions, mitigating interference during model aggregation and improving global uniformity." + ], + "image_footnote": [], + "bbox": [ + 174, + 93, + 379, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2fcd06304674a14763cc8f20b16021d5ae6548320739da7837754e1ae66ba792.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 93, + 820, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "local clients and a central server. Each client $k$ has access to an unlabeled local dataset $\\mathcal{D}_k$ , modeled as samples from a client-specific data distribution $p_k(\\mathbf{x})$ :", + "bbox": [ + 89, + 351, + 482, + 397 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {k} = \\left\\{\\mathbf {x} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {n _ {k}} \\sim p _ {k} (\\mathbf {x}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 404, + 480, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n_k$ denotes the number of samples in $\\mathcal{D}_k$ . The global objective in FUL can be formulated as:", + "bbox": [ + 89, + 430, + 482, + 460 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\sum_ {k = 1} ^ {K} P _ {k} \\mathbb {E} _ {\\mathbf {x} \\sim p _ {k} (\\mathbf {x})} \\left[ \\mathcal {L} _ {k} (\\theta ; \\mathbf {x}) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 468, + 480, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P_{k}$ is the probability of drawing a data from client $k$ , and $\\mathcal{L}_k(\\cdot)$ denotes the local unsupervised loss of client $k$ .", + "bbox": [ + 89, + 517, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Representation Learning aims to learn meaningful representations that capture the underlying structure of data. A typical framework consists of the following components:", + "bbox": [ + 89, + 553, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Encoder $f(\\cdot)$ : Maps (augmented) input data $\\tilde{\\mathbf{x}} \\sim \\mathcal{T}(\\mathbf{x})$ into a representation $\\mathbf{h} = f(\\tilde{\\mathbf{x}}) \\in \\mathbb{R}^d$ , where $d$ is the hidden dimension. It is common to use a deep neural network (e.g., ResNet50) as the encoder.", + "- **Projector $g(\\cdot)$ : Transforms the representation $\\mathbf{h}$ into an embedding $\\mathbf{z} = g(\\mathbf{h}) \\in \\mathbb{R}^d$ . The training loss function is usually applied on this embedding space, and the projector is often a small network such as a multilayer perceptron." + ], + "bbox": [ + 89, + 614, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A fundamental principle in self-supervised representation learning is that similar samples should have similar representations. In other words, representations should be invariant to minor variations in input data. This is typically achieved by aligning the representations of two augmented versions of the same image (referred to as a positive pair) using the following alignment loss:", + "bbox": [ + 89, + 766, + 482, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\tilde {\\mathbf {x}}, \\tilde {\\mathbf {x}} ^ {+} \\sim \\mathcal {T} (\\mathbf {x})} | | \\mathbf {z} - \\mathbf {z} ^ {+} | | _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 878, + 482, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}},\\tilde{\\mathbf{x}}^{+}\\sim \\mathcal{T}(\\mathbf{x})$ denotes two independent augmentations of the same sample $\\mathbf{x}$ and $\\mathbf{z} = g(f(\\tilde{\\mathbf{x}})),\\mathbf{z}^{+} = g(f(\\tilde{\\mathbf{x}}^{+}))$ are the corresponding embeddings.", + "bbox": [ + 511, + 351, + 903, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although aligning positive pairs encourages similarity, it can lead to a degenerate collapse where all representations converge to a single point. To circumvent this, Wang and Isola [31] introduces a uniformity objective to ensure that features are evenly dispersed on the unit hypersphere. This uniformity is often measured using the log of the average Gaussian potential [4]:", + "bbox": [ + 511, + 397, + 905, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\stackrel {\\mathrm {i . i . d}} {\\sim} p (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} _ {i} - \\mathbf {z} _ {j} | | _ {2} ^ {2}} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 515, + 903, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\mathbf{z})$ is the distribution of embeddings obtained by mapping data samples through the encoder $f$ and projector $g$ and $t$ is the temperature hyperparameter.", + "bbox": [ + 511, + 551, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 613, + 648, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce Soft Separation and Distillation (SSD), a novel approach designed to enhance representation uniformity in federated learning (FL). We first analyze the limitations of uniformity in non-IID FL settings, then propose dimension-scaled regularization to softly separate client features, and finally introduce projector distillation to effectively transfer improved embedding uniformity to the representations.", + "bbox": [ + 511, + 638, + 905, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limited Inter-Client Uniformity of FL. Uniformity is a critical metric in representation learning [6, 30, 31], where higher uniformity indicates better preservation of information in the learned representations. In centralized training, jointly optimizing alignment loss $\\mathcal{L}_{\\mathrm{align}}$ and uniformity loss $\\mathcal{L}_{\\mathrm{uniform}}$ yields high-quality representations. However, the distributed nature of FL introduces a fundamental challenge to uniformity optimization. Specifically, the uniformity metric can be decomposed into intra-client and inter", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2973", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6d799e2047d6309e53442249fc68e3e14073f9757fd68eccb1928973d710f04.jpg", + "image_caption": [ + "Figure 3. Overview of our Training Pipeline. $①$ The server initially assigns distinct weight vectors to each client. $②$ Clients perform local training by optimizing their respective loss functions on their own data. $③$ The server aggregates updated parameters from all clients. Steps $②$ and $③$ are iteratively repeated until convergence." + ], + "image_footnote": [], + "bbox": [ + 117, + 90, + 898, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "client components:", + "bbox": [ + 89, + 295, + 220, + 310 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\left(\\underbrace {\\sum_ {k = 1} ^ {K} \\mathbb {E} _ {\\mathbf {z} , \\mathbf {z} ^ {\\prime} \\sim p _ {k} (\\mathbf {z})} \\left[ e ^ {- t \\left| \\mathbf {z} - \\mathbf {z} ^ {\\prime} \\right| / 2} \\right]} _ {\\text {i n t r a - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k}}\\right) \\tag {5} \\\\ \\left. + \\underbrace {\\sum_ {i \\neq j} \\mathbb {E} _ {\\mathbf {z} \\sim p _ {i} (\\mathbf {z}) , \\mathbf {z} ^ {\\prime} \\sim p _ {j} (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | | _ {2} ^ {2}} \\right]} _ {\\text {i n t e r - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {(i, j)}}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 318, + 483, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term represents intra-client uniformity $\\mathcal{L}_{\\mathrm{uniform}}^k$ and the second term represents inter-client uniformity $\\mathcal{L}_{\\mathrm{uniform}}^{(i,j)}$ .", + "bbox": [ + 89, + 455, + 482, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In federated learning, each client $i$ optimizes its local loss function using samples drawn from its own local distribution $p_i$ , which means that only the intra-client term in the above equation is explicitly optimized. In a homogeneous setting, when client distributions are similar (i.e., $p_i \\approx p_j$ ), optimizing intra-client uniformity naturally promotes good inter-client uniformity. However, in non-IID settings where client distributions differ significantly, optimizing only intra-client uniformity does not guarantee good inter-client uniformity, potentially limiting the quality of the globally aggregated model. The core challenge in enhancing inter-client uniformity lies in the federated learning constraint that the server has no access to raw client data or embeddings, making it impossible to directly impose a loss function that operates across different clients.", + "bbox": [ + 88, + 505, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dimension-Scaled Regularization (DSR). To address the inter-client uniformity challenge, we propose Dimension-Scaled Regularization (DSR), which improves global uniformity by encouraging client embeddings to spread in different directions of the representation space. This approach effectively increases the separation between clients without enforcing rigid boundaries that might distort the intrinsic data structure.", + "bbox": [ + 88, + 734, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each client $k$ , we define a dimension-scaling vector $\\mathbf{d}_k \\in \\mathbb{R}^d$ , where $d$ is the dimensionality of the embedding space. This vector applies selective scaling to specific di", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mensions, with some dimensions scaled by a factor $\\alpha \\neq 1$ in a client-specific manner:", + "bbox": [ + 511, + 295, + 903, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i, k} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} i \\in \\mathcal {S} _ {k} \\\\ 1, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 332, + 903, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{k}$ is a set of dimensions uniquely assigned to client $k$ , ensuring that the scaled dimensions are nonoverlapping across clients (i.e., $S_{i} \\cap S_{j} = \\emptyset$ for $i \\neq j$ ). The size of each set $S_{k}$ is approximately $\\lfloor d / K \\rfloor$ , where $K$ is the number of clients.", + "bbox": [ + 511, + 378, + 906, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then regularize each client's embeddings by encouraging them to move toward their dimension-scaled versions through the following loss:", + "bbox": [ + 511, + 455, + 906, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D S R}} ^ {k} = \\mathbb {E} _ {\\mathbf {z} \\sim p _ {k} (\\mathbf {z})} \\left[ \\| \\mathbf {z} - \\operatorname {s t o p g r a d} \\left(\\mathbf {z} \\odot \\mathbf {d} _ {k}\\right) \\| _ {2} ^ {2} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 508, + 903, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ represents element-wise multiplication, and stopgrad $(\\cdot)$ prevents gradient flow through the scaled target, ensuring that we pull the original embedding toward the scaled version rather than vice versa.", + "bbox": [ + 511, + 534, + 905, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To understand why DSR enhances inter-client uniformity, we can analyze its effect on the representation distribution. When a vector $\\mathbf{z}$ is scaled along specific dimensions by a factor $\\alpha > 1$ and then normalized to the unit hypersphere, the resulting vector shifts toward those scaled dimensions. By assigning different scaling dimensions to each client, DSR effectively encourages client representations to occupy different regions of the unit hypersphere, as illustrated in Figure 5. Mathematically, if clients $i$ and $j$ have scaling vectors $\\mathbf{d}_i$ and $\\mathbf{d}_j$ with non-overlapping scaled dimensions, their optimized embeddings will tend to have a smaller dot product:", + "bbox": [ + 511, + 595, + 906, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathbf {z} _ {i} \\sim p _ {i} (\\mathbf {z}), \\mathbf {z} _ {j} \\sim p _ {j} (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right] < \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\sim p (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 782, + 903, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p(\\mathbf{z})$ represents the distribution without DSR. This reduced dot product corresponds to increased angular separation, directly contributing to improved uniformity across the global distribution.", + "bbox": [ + 511, + 809, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike a hard separation approach that would restrict each client to entirely separate subspaces (which could", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/293ad20cd014bd41b353af42752f8e3d426a213b03895cd37947c77c2dbdd256.jpg", + "image_caption": [ + "Figure 4. Representation norm, embedding norm, and DSR loss during training. Adding DSR primarily affects the embedding space, as reflected in the embedding norm, while leaving the representation norm unchanged. In contrast, incorporating PD directly influences the representation norm, and when combined with DSR, it leads to a more pronounced decrease in representation norm. Additionally, PD enhances the effect of DSR, resulting in a lower $\\mathcal{L}_{\\mathrm{DSR}}$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 359, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b7786b85c457eaf192d008aa99161c863902d0e279ff1f63ef93bdf2674a26a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 629, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/73b86710d64cc8088f62a4242cbbb70349561247d702a8957aefe969378797be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 90, + 901, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a23779c534c4f356fc90616c05d3f9f40a437194fad3466f0309737cf1a7538.jpg", + "image_caption": [ + "Figure 5. Illustration of the effect of DSR. The left figure shows the original data, where arrows indicate the transformation direction after applying DSR. The right figure presents the data after transformation. DSR increases separation and enhances inter-client uniformity." + ], + "image_footnote": [], + "bbox": [ + 130, + 321, + 279, + 450 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d4c462cabe02ba7b71971ac33fa5cf9fea5f3d464edbfb9f1c06e5ccf907ea1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 321, + 442, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "severely constrain representation capacity), our soft separation allows clients to share most dimensions while gently pushing them toward different directions. This balanced approach preserves the flexibility needed for effective representation learning while significantly improving inter-client uniformity.", + "bbox": [ + 89, + 558, + 483, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An extreme example of fully separating embeddings would be assigning each client a dedicated subspace (e.g., if $d = 500$ and $K = 10$ , each client could occupy 50 distinct dimensions). Although this configuration maximizes inter-client uniformity, it may disrupt the intrinsic data structure and miss out on the collaborative benefits of FL. In contrast, our soft DSR loss encourages some degree of separation without completely isolating the clients' embeddings. We analyze this tradeoff further in Section 4.2.", + "bbox": [ + 89, + 652, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projector Distillation (PD). While DSR effectively enhances uniformity at the embedding level, we observe that this improvement may not fully transfer to the representation level. This discrepancy occurs because the projector $g(\\cdot)$ placed between the encoder $f(\\cdot)$ and the loss function can absorb much of the optimization effect, as shown in Figure 4. Although completely removing the projector", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "might seem like a direct solution, prior work [2, 9, 32] has demonstrated that the projector plays a crucial role in separating optimization objectives from feature representations, thereby preventing overfitting to specific self-supervised tasks and improving downstream performance.", + "bbox": [ + 511, + 310, + 903, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To bridge this gap, we introduce Projector Distillation (PD), which explicitly aligns the encoder's representations with the projector's embeddings:", + "bbox": [ + 511, + 386, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} ^ {k} = \\mathbb {E} _ {x \\sim p _ {k} (\\mathbf {x})} \\left[ D _ {\\mathrm {K L}} \\left(\\sigma (\\mathbf {h}) \\| \\sigma (\\mathbf {z})\\right) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 440, + 903, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h} = f(\\mathbf{x})$ is the representation from the encoder, $\\mathbf{z} = g(\\mathbf{h})$ is the embedding from the projector, $\\sigma(\\cdot)$ denotes the softmax function, and $D_{\\mathrm{KL}}$ is the Kullback-Leibler divergence.", + "bbox": [ + 511, + 468, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This distillation mechanism encourages the encoder to internalize the beneficial structure learned in the embedding space, effectively transferring the improved uniformity from embeddings to representations. By minimizing the KL divergence between these distributions, we ensure that the uniformity improvements achieved through DSR are reflected in the representations used for downstream tasks.", + "bbox": [ + 511, + 529, + 905, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Overview of Training Pipeline", + "text_level": 1, + "bbox": [ + 511, + 643, + 782, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete SSD algorithm is outlined in Algorithm 1, with our key design components highlighted in red. The server initially assigns unique weight vectors to all clients for dimension-scaled regularization. During local training, each client optimizes a combination of the standard alignment and uniformity losses for self-supervised learning, augmented with our DSR and PD terms. The local training objective for each client $k$ is formulated as:", + "bbox": [ + 511, + 666, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {k} = \\mathcal {L} _ {\\text {a l i g n}} ^ {k} + \\beta \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k} + \\gamma \\mathcal {L} _ {\\text {D S R}} ^ {k} + \\delta \\mathcal {L} _ {\\text {d i s t i l l}} ^ {k}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 811, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta, \\gamma$ , and $\\delta$ are hyperparameters that balance the contribution of each loss term.", + "bbox": [ + 511, + 840, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After local training, clients upload their updated models to the server, which aggregates them using standard", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2975", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1: Soft Separation and Distillation (SSD) \nInput: communication rounds $T$ , local epochs $E$ , number of clients $K$ , dataset $\\mathcal{D} = \\cup_{k \\in [K]} \\mathcal{D}_k$ , augmentation function $\\mathcal{T}(\\cdot)$ \nOutput: Final global encoder $f_\\theta^T$ \n1 Server: \n2 Initialize global encoder $f_\\theta^0$ and projector $g_\\theta^0$ ; \n3 Assign weight vector $\\{\\mathbf{d}_k\\}_{k \\in [K]}$ for all clients; \n4 for $t = 0$ to $T - 1$ do \n5 $S_t \\gets$ (randomly select a set of clients); \n6 for each client $k \\in S_t$ in parallel do \n7 $f_\\theta^{(t,k)}, g_\\theta^{(t,k)} \\gets \\text{Client}(f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)$ ; \n8 end \n9 // Model aggregation; \n10 for each parameter $\\theta$ in $f, g$ do \n11 $\\theta^{t+1} \\gets \\sum_{k \\in S_t} w_k \\theta^{(t,k)}$ \n12 end \n13 end \n14 return $f_\\theta^T$ ; \n15 Client $(f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)$ : \n16 // Initialize local models; \n17 $f \\gets f_\\theta^t, g \\gets g_\\theta^t$ ; \n18 for each local epoch $e = 0$ to $E - 1$ do \n19 for each batch $\\{\\mathbf{x}_i\\}_{i=1}^{n_B} \\in \\mathcal{D}_k$ do \n20 // Data augmentation; \n21 $\\tilde{\\mathbf{x}}_i, \\tilde{\\mathbf{x}}_i^+ = \\mathcal{T}(\\mathbf{x}_i), \\mathcal{T}(\\mathbf{x}_i)$ ; \n22 // Get representations and embeddings; \n23 $\\mathbf{h}_i, \\mathbf{h}_i^+ = f(\\tilde{\\mathbf{x}}_i), f(\\tilde{\\mathbf{x}}_i^+)$ ; \n24 $\\mathbf{z}_i, \\mathbf{z}_i^+ = g(\\mathbf{h}_i), g(\\mathbf{h}_i^+)$ ; \n25 // Optimize model using combined loss; \n26 Update $f, g$ by minimizing $\\mathcal{L}^k$ (Eq (10)); \n27 end \n28 end \n29 return $f, g$ // Model upload; \n30 return", + "bbox": [ + 89, + 92, + 480, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "federated averaging [24]: $\\theta^{t + 1} = \\sum_{k\\in S_t}w_k\\theta^{(t,k)}$ where $w_{k} = \\frac{n_{k}}{\\sum_{i = 1}^{K}n_{i}}$ represents the aggregation weight based on the client's data size.", + "bbox": [ + 89, + 676, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Importantly, our method preserves privacy since it requires no sharing of raw data. The only additional communication is the initial weight vectors $\\{\\mathbf{d}_k\\}_{k=1}^K$ , which are lightweight and contain no sensitive information. The computational overhead of SSD is minimal, making it highly practical for real-world federated learning deployments.", + "bbox": [ + 89, + 726, + 482, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 220, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our approach on CIFAR-10 and CIFAR-100, which contain 50K training images and 10K testing images, with 10 and 100 classes, respectively. Fol", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lowing prior work [22, 37], we simulate non-IID data across $K$ clients using a Dirichlet prior $\\mathrm{Dir}(\\alpha)$ , where a smaller $\\alpha$ indicates a higher degree of data heterogeneity. We conduct experiments in two federated learning settings: cross-silo $(K = 10)$ with full participation and cross-device $(K = 50)$ with a participation rate of 0.2.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate our approach based on downstream performance and representation quality. For downstream performance, we follow prior work [10, 22, 36], using linear probing and fine-tuning with $1\\%$ and $10\\%$ of the data. For representation quality, we assess uniformity [6, 30, 31] as defined in Equation (4) and effective rank [7, 27, 34] (see the definition in Appendix A.2).", + "bbox": [ + 511, + 186, + 903, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method against three categories of baselines. (1) Adapting a centralized algorithm to the federated setting: FedAlignUniform [31]. (2) FUL methods that mitigate divergence from the global model: FedX [10] and L-DAWA [25]. (3) FUL methods addressing dimensional collapse in local clients: FedDecorr [28]. FedU2 [22] incorporates modules that tackle both divergence control and dimensional collapse.", + "bbox": [ + 511, + 296, + 903, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We follow prior work [22] for image augmentations, training configurations, and model architecture, using ResNet-18 [11] as the encoder and a two-layer linear projector. The hyperparameters $\\beta$ , $\\gamma$ , and $\\delta$ are set to 1.0, 1.0, and 0.1, respectively, and the scaling factor $\\alpha$ is set to 10.0. Further details are provided in Appendix A.1.", + "bbox": [ + 511, + 422, + 903, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 513, + 540, + 651, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on different FL settings. Table 1 compares the performance of various methods in cross-silo and crossdevice federated learning. SSD consistently achieves the highest accuracy and excels in fine-tuning tasks. FedAlignUniform, FedDecorr, and FedU2 are competitive but generally lag behind SSD, while FedX and L-DAWA—aimed at ensuring global model consistency—perform the worst across all settings. Although FedDecorr and FedU2 enhance intra-client uniformity (with FedU2 adding a consistency module), they still trail SSD. By improving inter-client uniformity, SSD underscores the importance of addressing inter-client variations in federated learning.", + "bbox": [ + 511, + 563, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Representation quality. Figure 1 compares different methods based on effective rank [27] and uniformity [31], two key metrics for representation quality. Among state-of-the-art methods, FedX and L-DAWA, which emphasize global model consistency, achieve strong representation quality. In contrast, methods focusing on local uniformity, such as FedDecorr and FedU2, exhibit lower effective rank and uniformity, indicating that they fail to address global uniformity. Our method, SSD, surpasses all other federated approaches, achieving the highest effective rank and the best uniformity.", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2976", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/76ffd5e441574799aac1328380d9095a6e29b211f2923076f69efb2ec347cc66.jpg", + "table_caption": [ + "Table 1. Results on Cross-Silo and Cross-Device Settings. Accuracy (%) of linear probing (LP), fine-tuning (FT) $1\\%$ , and $10\\%$ labeled data on CIFAR10 and CIFAR100 ( $\\alpha = 0.1$ ). The highest score is highlighted in bold, and the second-highest score is underlined." + ], + "table_footnote": [], + "table_body": "
CIFAR10CIFAR100
Cross-Silo (K=10)Cross-Device (K=50)Cross-Silo (K=10)Cross-Device (K=50)
LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%
FedAlignUniform [31]80.8469.9981.0071.2857.4173.7757.2528.9748.9943.0316.3736.64
FedX [10]78.466.7880.0171.0156.9173.2457.3427.4649.5043.0716.0435.46
L-DAWA [25]77.6765.9479.3467.6553.7571.2256.9027.0849.5742.5815.5434.82
FedDecorr [28]80.1369.0980.3371.4958.1973.9757.2529.3849.5344.7417.6736.68
FedU2 [22]81.0169.6281.0171.0957.1574.2157.4029.3949.6442.9016.0835.48
SSD81.3270.7481.6771.8357.7774.6157.3829.5749.8745.2117.7036.82
", + "bbox": [ + 94, + 128, + 903, + 260 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c37d11505025b02e202e709a3ea7e8b736218ddb41660ffec745a71be36aee53.jpg", + "table_caption": [ + "Table 2. Generalization on OOD datasets. Accuracy $(\\%)$ of linear probing, uniformity, and effective rank when trained on CIFAR-100 and TinyImageNet-200, and evaluated on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
CIRAR100 → CIFAR10TinyImageNet → CIFAR10
LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)
FedAlignUniform [31]77.663.6566.9979.863.7176.31
FedX [10]78.023.7384.8879.873.7693.48
L-DAWA [25]77.463.7184.7579.623.7493.41
FedDecorr [28]77.623.6674.2079.793.7284.93
FedU2 [22]77.743.6668.379.743.6975.97
SSD78.483.7386.9580.003.7798.45
", + "bbox": [ + 94, + 337, + 480, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This demonstrates that SSD effectively enhances representation quality, outperforming existing methods.", + "bbox": [ + 89, + 441, + 482, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization on OOD datasets. Evaluating generalization to out-of-distribution (OOD) datasets helps assess whether models can learn transferable representations that perform well beyond their training distribution. Table 2 presents the generalization performance of different methods, where models trained on CIFAR-100 and TinyImageNet-200 are evaluated on CIFAR-10.", + "bbox": [ + 89, + 477, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among state-of-the-art methods, FedX, which exhibits high representation quality but performs poorly in indistribution settings, achieves the best generalization in OOD scenarios. However, our method, SSD, consistently outperforms all baselines, achieving the highest accuracy in both settings. SSD also demonstrates the best uniformity [31] and effective rank [7], indicating superior feature representation quality. While FedAlignUniform, FedDecorr, L-DAWA and FedU2 perform competitively, they generally fall short of SSD, further reinforcing its effectiveness in handling distribution shifts.", + "bbox": [ + 89, + 582, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Analysis", + "text_level": 1, + "bbox": [ + 89, + 757, + 192, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. We provide the ablation study in Table 3. It shows that adding projector distillation (PD) alone does not provide noticeable improvements. In contrast, adding dimension-scaled regularization (DSR) alone leads to slight improvements in LP and FT $10\\%$ , but the uniformity enhancement is minimal. However, when combining both DSR and PD (SSD), the method effectively transfers optimization benefits to representation quality, leading to a", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3048135ca9698ca1158e26a1a0c148a530940fca638a8ab62ab0ffa28922b3e6.jpg", + "table_caption": [ + "Table 3. Ablation Study. PD alone has minimal impact on performance. Adding DSR alone provides a slight improvement in both performance and uniformity. Combining both DSR and PD leads to a significant boost in uniformity and achieves the best performance." + ], + "table_footnote": [], + "table_body": "
LPFT 1%FT 10%-Luniform(↑)
FedAlignUniform [31]80.8469.9981.003.79
+PD80.7469.7880.713.80
+DSR81.0569.7781.153.81
+ DSR + PD (SSD)81.3270.7481.673.84
", + "bbox": [ + 517, + 364, + 903, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significant uniformity improvement and achieving the best overall performance.", + "bbox": [ + 511, + 468, + 906, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness of DSR. We evaluate our method from two perspectives: the scaled factor $\\alpha$ and the selection of scaled dimensions for each client. The scaled factor $\\alpha$ influences the overall discrepancy, while the selection of scaled dimensions affects the direction each client is guided toward. To assess robustness, each $\\alpha$ is evaluated with three different selections of scaled dimensions. As shown in Figure 6a, our method maintains stable performance across different $\\alpha$ values, introducing minimal variation while consistently outperforming FedAlignUniform. This demonstrates that our approach effectively balances alignment and uniformity without being overly sensitive to the choice of $\\alpha$ or the specific dimension selection.", + "bbox": [ + 511, + 503, + 908, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft vs. hard client separation. We examine the impact of client feature separation on alignment, uniformity, and overall performance. Our intuition is that increasing separation between clients enhances global uniformity. One way to enforce this separation is by restricting each client to its own subspace without sharing information with others. To illustrate the effects of such strict partitioning, we consider Hard Separation and Distillation (HSD) as a baseline. As shown in Figure 6b, while HSD achieves the highest uniformity, it does so at the cost of severely reduced alignment, ultimately leading to poor downstream performance. This performance drop can be attributed to the disruption of intrinsic feature structures. In contrast, our proposed SSD balances", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2543d03136fed646838407ea320da77c9d65b100fc9fc8549463b4ff2ff7c37e.jpg", + "image_caption": [ + "(a) Robustness", + "Figure 6. (a) Robustness of DSR across different scaled factors $\\alpha$ , where each $\\alpha$ is tested with three different randomly selected scaled dimensions. (b) Soft vs. hard separation. HSD achieves the highest uniformity but at the cost of reduced alignment, resulting in poor overall performance." + ], + "image_footnote": [], + "bbox": [ + 89, + 104, + 277, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/08d31bfc5fc52f049028f3a49080672d2c9485c1cda065c7337028517ca1696f.jpg", + "image_caption": [ + "(b) Separation" + ], + "image_footnote": [], + "bbox": [ + 290, + 104, + 475, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/da8fa3307aac4428dcdfbc4819d36def8d842ad2a978770f02eb622a41f0eb0c.jpg", + "table_caption": [ + "Table 4. The Effect of the Projector. Without the projector, DSR significantly improves uniformity and performance. Yet, the overall results remain inferior to those achieved with the projector." + ], + "table_footnote": [], + "table_body": "
ProjectileLP-Luniform(↑)
FedAlignUniformX73.163.72
+ DSRX76.14 (+2.98)3.77 (+0.05)
FedAlignUniform✓80.843.79
+ DSR✓81.05 (+0.21)3.81 (+0.02)
", + "bbox": [ + 133, + 407, + 442, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "both alignment and uniformity, mitigating the drawbacks of hard separation while enhancing uniformity, resulting in improved overall performance.", + "bbox": [ + 89, + 503, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Why not remove projector or apply loss on representations? A natural idea for addressing the limited transfer of embedding-level loss optimization (in terms of representation uniformity) is to remove the projector or apply the loss directly to the representations. However, prior work [2, 9, 32] has demonstrated that the projector plays a critical role in preventing the encoder from overfitting on the upstream task. Indeed, as shown in Table 4, removing the projector and adding DSR does substantially improve uniformity and downstream performance, but these results still lag behind those achieved when the projector is retained. Therefore, it is crucial to develop a method that effectively promotes uniformity while preserving the benefits offered by the projector.", + "bbox": [ + 89, + 554, + 482, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 777, + 232, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Federated Learning with Non-IID Data", + "text_level": 1, + "bbox": [ + 89, + 801, + 428, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Federated learning (FL) enables collaborative model training across multiple clients without centralizing data. Since client data distributions are typically non-IID, model convergence and performance are challenging. FedAvg [24], which aggregates local model updates using simple averag", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing, is the foundational framework but suffers under heterogeneous data and imbalanced client participation.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address these issues, various aggregation improvements have been proposed. FedProx [21] reduces client drift by adding a proximal term to limit local updates from straying from the global model. SCAFFOLD [16] corrects client drift using control variates to reduce variance in local updates. MOON [19] improves representation consistency with a contrastive loss. Other methods use globally shared data to improve generalization; FedShare [35] aligns client models using shared data, while FedDistill [13] distills knowledge from the global model to local models using shared data.", + "bbox": [ + 511, + 121, + 906, + 286 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Federated Unsupervised Learning", + "text_level": 1, + "bbox": [ + 511, + 296, + 812, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Federated Unsupervised Learning (FUL) combines federated learning with un-/self-supervised learning, primarily addressing (1) non-IID data distribution and (2) representation collapse.", + "bbox": [ + 511, + 318, + 903, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To handle non-IID data, FedCA [33] uses a shared dictionary module for better aggregation but risks privacy leakage. FedU [36] reduces this risk by selectively uploading the online network's encoder and deciding predictor updates based on divergence. FedEMA [37] extends this with exponential moving average updates. FedX [10] adds an alignment/contrastive term with the global model, while Orchestra [23] preserves global structural consistency using global centroids. FedU2 [22] ensures balanced updates across clients for better alignment but does not guarantee global representation uniformity under non-IID settings.", + "bbox": [ + 511, + 378, + 905, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For representation collapse, FedDecorr [28] shows that local clients suffer from dimensional collapse, which propagates to the global model, and addresses it with a local decorrelation loss. FedU2 [22] encourages local representation uniformity by minimizing divergence with a spherical Gaussian. Our work extends this by showing that intra-client uniformity alone is insufficient inter-client uniformity must also be explicitly addressed.", + "bbox": [ + 511, + 545, + 905, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Soft Separation & Distillation (SSD), a framework designed to enhance representation quality by improving inter-client uniformity in federated learning. SSD consists of a dimension-scaled regularization term that softly separates client embeddings while preserving the intrinsic data structure, and a projector distillation term that transfers the optimization benefits of the projector to the encoder, thereby improving representation quality. SSD achieves state-of-the-art performance in both representation learning and downstream tasks across various training and FL settings. Our work highlights the importance of global representation quality in federated unsupervised learning, opening new directions for future research.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2978", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers and members of CLLab for their constructive feedback. This work is partially supported by the National Science and Technology Council in Taiwan via NSTC 113-2634-F-002-008, 114-2221-E002-102-MY3, and the National Taiwan University Center for Data Intelligence via NTU-114L900901 and Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410072, RGC R1015-23). We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. H.-T. Lin is honored to be supported by the Leap Fellowship of the Foundation for the Advancement of Outstanding Scholarship in Taiwan since 2025.", + "bbox": [ + 89, + 116, + 485, + 327 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 344, + 187, + 359 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 1", + "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 1, 2, 5, 8, 11", + "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1", + "[4] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1", + "[6] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. In The Twelfth International Conference on Learning Representations, 2024. 3, 6", + "[7] Quentin Garrido, Randall Balestriero, Laurent Najman, and Yann Lecun. Rankme: Assessing the downstream performance of pretrained self-supervised representations by their rank. In International conference on machine learning, pages 10929-10974. PMLR, 2023. 2, 6, 7", + "[8] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1", + "[9] Kartik Gupta, Thalaiyasingam Ajanthan, Anton van den Hengel, and Stephen Gould. Understanding and improving" + ], + "bbox": [ + 99, + 369, + 480, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the role of projection head in self-supervised learning. arXiv preprint arXiv:2212.11491, 2022. 2, 5, 8", + "[10] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xing Xie, and Meeyoung Cha. Fedx: Unsupervised federated learning with cross knowledge distillation. In European Conference on Computer Vision, pages 691-707. Springer, 2022. 2, 6, 7, 8", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6, 11", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1", + "[13] Eunjeeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non- $\\mathrm{id}$ private data. arXiv preprint arXiv:1811.11479, 2018. 8", + "[14] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. In ICLR, 2022. 2", + "[15] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1", + "[16] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 8", + "[17] Taehyeon Kim, Jaehoon Oh, Nak Yil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2628–2635. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 11", + "[18] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1-40, 2016. 1", + "[19] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 8", + "[20] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1", + "[21] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 2, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2979", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Xinting Liao, Weiming Liu, Chaochao Chen, Pengyang Zhou, Fengyuan Yu, Huabin Zhu, Binhui Yao, Tao Wang, Xiaolin Zheng, and Yanchao Tan. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22841-22850, 2024. 2, 6, 7, 8", + "[23] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In Proceedings of the 39th International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 8", + "[24] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6, 8", + "[25] Yasar Abbas Ur Rehman, Yan Gao, Pedro Porto Buarque De Gusmão, Mina Alibeigi, Jiajun Shen, and Nicholas D Lane. L-dawa: Layer-wise divergence aware weight aggregation in federated self-supervised visual representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16464-16473, 2023. 6, 7", + "[26] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22 (3):400 - 407, 1951. 11", + "[27] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007. 1, 6", + "[28] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7, 8", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1", + "[30] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3, 6", + "[31] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, 2020. 1, 2, 3, 6, 7, 11", + "[32] Yihao Xue, Eric Gan, Jiayi Ni, Siddharth Joshi, and Baharan Mirzasoleiman. Investigating the benefits of projection head for representation learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 8", + "[33] Fengda Zhang, Kun Kuang, Long Chen, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Fei Wu, Yueting Zhuang, et al. Federated unsupervised representation learning. Frontiers of Information Technology & Electronic Engineering, 24(8):1181-1193, 2023. 8", + "[34] Yifei Zhang, Hao Zhu, Yankai Chen, Zixing Song, Piotr Koniusz, and Irwin King. Mitigating the popularity bias of" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph collaborative filtering: A dimensional collapse perspective. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 6", + "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 8", + "[36] Weiming Zhuang, Xin Gan, Yonggang Wen, Shuai Zhang, and Shuai Yi. Collaborative unsupervised visual representation learning from decentralized data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4912-4921, 2021. 2, 6, 8", + "[37] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In International Conference on Learning Representations, 2022. 2, 6, 8" + ], + "bbox": [ + 516, + 92, + 906, + 301 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2980", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_model.json b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9eadbe6732acd7ed5cee28fc702813e6d8e783cd --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.13, + 0.817, + 0.177 + ], + "angle": 0, + "content": "Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.203, + 0.785, + 0.223 + ], + "angle": 0, + "content": "Hung-Chieh Fang\\(^{1,*}\\) Hsuan-Tien Lin\\(^{1}\\) Irwin King\\(^{2}\\) Yifei Zhang\\(^{2,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.239, + 0.768, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)National Taiwan University \\(^{2}\\)The Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Federated Unsupervised Learning (FUL) aims to learn expressive representations in federated and self-supervised settings. The quality of representations learned in FUL is usually determined by uniformity, a measure of how uniformly representations are distributed in the embedding space. However, existing solutions perform well in achieving intra-client (local) uniformity for local models while failing to achieve inter-client (global) uniformity after aggregation due to non-IID data distributions and the decentralized nature of FUL. To address this issue, we propose Soft Separation and Distillation (SSD), a novel approach that preserves inter-client uniformity by encouraging client representations to spread toward different directions. This design reduces interference during client model aggregation, thereby improving global uniformity while preserving local representation expressiveness. We further enhance this effect by introducing a projector distillation module to address the discrepancy between loss optimization and representation quality. We evaluate SSD in both cross-silo and cross-device federated settings, demonstrating consistent improvements in representation quality and task performance across various training scenarios. Our results highlight the importance of inter-client uniformity in FUL and establish SSD as an effective solution to this challenge. Project page: https://ssd-uniformity.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Deep learning has achieved remarkable success across a wide spectrum of applications, from computer vision and natural language processing to speech recognition and reinforcement learning [1, 11, 18, 29]. This progress can be largely attributed to the availability of massive labeled datasets, particularly since the emergence of ImageNet [5]," + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.324, + 0.837, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.577, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Figure 1. Comparison with state-of-the-art methods. Our SSD achieves the highest performance (marked in brackets) and the best representation quality, measured by uniformity [31] and effective rank [27], among existing FUL approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "which has enabled the training of increasingly powerful neural networks. However, the practical deployment of deep learning in real-world scenarios faces two critical challenges. First, data in real-world applications is often non-Independent and Identically Distributed (non-IID) and cannot be freely shared due to privacy concerns, regulatory restrictions, or proprietary constraints. This data is typically generated across diverse sources, including user devices, healthcare institutions, and industrial settings. The distributed nature and sharing limitations of such data make centralized training approaches infeasible, leading to the development of Federated Learning (FL) [15, 20, 24] as a privacy-preserving distributed learning paradigm. Second, a vast proportion of available data remains unlabeled, necessitating effective methods for unsupervised representation learning [2, 3, 8, 12]. These methods aim to learn meaningful feature representations without relying on explicit la" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.483, + 0.888 + ], + "angle": 0, + "content": "* Project was partially completed during the author's visit to the CUHK" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.243, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "bels, enabling models to capture underlying data structures and semantic relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.482, + 0.333 + ], + "angle": 0, + "content": "Federated Unsupervised Learning (FUL) emerges at the intersection of these two challenges, aiming to learn expressive representations in settings where data is both distributed and unlabeled. In FUL, the quality of learned representations is critically determined by two properties: alignment and uniformity [31]. Alignment measures how close similar data points are positioned in the representation space, relating to the model's ability to group semantically similar objects. Uniformity quantifies how evenly representations are distributed across the unit hypersphere, essentially measuring the entropy of the representation distribution. High uniformity prevents representation collapse and ensures that the learned features effectively utilize the available embedding dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.482, + 0.499 + ], + "angle": 0, + "content": "However, existing FUL methods face a significant challenge that remains inadequately addressed. While current approaches successfully achieve good intra-client (local) uniformity for representations within each client, they struggle to maintain inter-client (global) uniformity after model aggregation. This limitation stems from two key factors: (1) the non-IID distribution of data across clients, which naturally leads to divergent updates, and (2) the decentralized nature of FL, where the server lacks direct access to raw data, preventing the application of explicit uniformity constraints across clients." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.482, + 0.667 + ], + "angle": 0, + "content": "Most existing approaches have focused primarily on two aspects of this challenge. One line of research builds upon the framework established by FedProx [21], introducing proximal terms to constrain local updates within a bounded neighborhood of the global model. For instance, FedU [36] and FedEMA [37] dynamically adjust aggregation weights based on inter-client model divergence, while FedX [10] incorporates a global relational loss to align pairwise sample relationships across clients. These methods aim to maintain global model consistency but do not explicitly address representation uniformity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.482, + 0.833 + ], + "angle": 0, + "content": "Another research direction tackles the issue of dimensional collapse [14] caused by lower uniformity in local representations. FedDecorr [28] demonstrates that embeddings of local clients are often less uniformly distributed and mitigates this issue by decorrelating local features. Similarly, FedU2 [22] regularizes local updates to approximate a spherical Gaussian distribution, encouraging isotropic feature spaces. While these approaches enhance local uniformity, the improvements at the client level do not inherently translate to better global uniformity during model aggregation. This leads to a crucial question:" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.845, + 0.476, + 0.874 + ], + "angle": 0, + "content": "How can we effectively improve inter-client uniformity in Federated Unsupervised Learning?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, we propose Soft Separation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "and Distillation (SSD), a novel method that enhances inter-client uniformity without compromising local representation expressiveness. Our approach employs a dimension-scaled regularization strategy that softly separates each client's feature space, encouraging client representations to spread toward different directions. As illustrated in Figure 2, this technique reduces interference during client model aggregation, thereby improving global uniformity without imposing rigid boundaries that could distort the underlying feature distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.44 + ], + "angle": 0, + "content": "Furthermore, we empirically observe that the regularization effect of SSD may not always effectively transfer from loss optimization to the representation space due to the presence of a projector between the loss function and the encoder. While removing the projector might seem like a straightforward solution, doing so often degrades performance because the projector plays a crucial role in separating optimization objectives from feature representations [2, 9, 32]. To bridge this gap, we introduce a projector distillation module that minimizes the KL divergence between representations and embeddings, effectively encouraging the encoder to internalize the learned structure while preserving the projector's role in loss optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.905, + 0.592 + ], + "angle": 0, + "content": "We evaluate SSD across diverse federated learning scenarios, including cross-silo (few clients with large datasets) and cross-device (many clients with small datasets) settings. Our experiments span both in-distribution tasks and out-of-distribution (OOD) datasets to assess generalizability. Results demonstrate that SSD consistently outperforms existing methods in both downstream task performance (measured via linear probing and fine-tuning accuracy) and representation quality (quantified using effective rank [7] and uniformity [31] metrics)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.594, + 0.782, + 0.607 + ], + "angle": 0, + "content": "Our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.905, + 0.654 + ], + "angle": 0, + "content": "- We identify and formalize the challenge of inter-client uniformity in FUL, establishing it as a critical direction for decentralized unsupervised representation learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.905, + 0.715 + ], + "angle": 0, + "content": "- We propose Soft Separation and Distillation (SSD), a simple yet effective framework that addresses inter-client uniformity without additional communication overhead or compromising privacy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.715, + 0.905, + 0.76 + ], + "angle": 0, + "content": "- We conduct extensive experiments across varied FL settings and tasks, confirming SSD's superiority over baseline methods in both performance and robustness." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.609, + 0.905, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.778, + 0.652, + 0.793 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Federated unsupervised learning (FUL) is a learning paradigm in which multiple clients collaboratively train a model without sharing new raw data, and where the local datasets contain unlabeled data. The goal is to learn a global representation that generalizes across all participating clients. Consider a federated learning system with \\( K \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2972" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.094, + 0.38, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.094, + 0.821, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.269, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Illustration of Intra-Client and Inter-Client Uniformity with Soft Separation. Intra-client uniformity ensures that representations within each client are well-distributed, while inter-client uniformity promotes global representation consistency across clients. Our proposed Soft Separation method encourages each client's representations to spread in distinct directions, mitigating interference during model aggregation and improving global uniformity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.398 + ], + "angle": 0, + "content": "local clients and a central server. Each client \\( k \\) has access to an unlabeled local dataset \\( \\mathcal{D}_k \\), modeled as samples from a client-specific data distribution \\( p_k(\\mathbf{x}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.405, + 0.482, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {k} = \\left\\{\\mathbf {x} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {n _ {k}} \\sim p _ {k} (\\mathbf {x}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.461 + ], + "angle": 0, + "content": "where \\(n_k\\) denotes the number of samples in \\(\\mathcal{D}_k\\). The global objective in FUL can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.469, + 0.482, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\sum_ {k = 1} ^ {K} P _ {k} \\mathbb {E} _ {\\mathbf {x} \\sim p _ {k} (\\mathbf {x})} \\left[ \\mathcal {L} _ {k} (\\theta ; \\mathbf {x}) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.549 + ], + "angle": 0, + "content": "where \\(P_{k}\\) is the probability of drawing a data from client \\(k\\), and \\(\\mathcal{L}_k(\\cdot)\\) denotes the local unsupervised loss of client \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Self-Supervised Representation Learning aims to learn meaningful representations that capture the underlying structure of data. A typical framework consists of the following components:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.616, + 0.483, + 0.677 + ], + "angle": 0, + "content": "- Encoder \\( f(\\cdot) \\): Maps (augmented) input data \\( \\tilde{\\mathbf{x}} \\sim \\mathcal{T}(\\mathbf{x}) \\) into a representation \\( \\mathbf{h} = f(\\tilde{\\mathbf{x}}) \\in \\mathbb{R}^d \\), where \\( d \\) is the hidden dimension. It is common to use a deep neural network (e.g., ResNet50) as the encoder." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- **Projector \\( g(\\cdot) \\): Transforms the representation \\( \\mathbf{h} \\) into an embedding \\( \\mathbf{z} = g(\\mathbf{h}) \\in \\mathbb{R}^d \\). The training loss function is usually applied on this embedding space, and the projector is often a small network such as a multilayer perceptron." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.616, + 0.483, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.767, + 0.483, + 0.873 + ], + "angle": 0, + "content": "A fundamental principle in self-supervised representation learning is that similar samples should have similar representations. In other words, representations should be invariant to minor variations in input data. This is typically achieved by aligning the representations of two augmented versions of the same image (referred to as a positive pair) using the following alignment loss:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.88, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\tilde {\\mathbf {x}}, \\tilde {\\mathbf {x}} ^ {+} \\sim \\mathcal {T} (\\mathbf {x})} | | \\mathbf {z} - \\mathbf {z} ^ {+} | | _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.905, + 0.398 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}},\\tilde{\\mathbf{x}}^{+}\\sim \\mathcal{T}(\\mathbf{x})\\) denotes two independent augmentations of the same sample \\(\\mathbf{x}\\) and \\(\\mathbf{z} = g(f(\\tilde{\\mathbf{x}})),\\mathbf{z}^{+} = g(f(\\tilde{\\mathbf{x}}^{+}))\\) are the corresponding embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Although aligning positive pairs encourages similarity, it can lead to a degenerate collapse where all representations converge to a single point. To circumvent this, Wang and Isola [31] introduces a uniformity objective to ensure that features are evenly dispersed on the unit hypersphere. This uniformity is often measured using the log of the average Gaussian potential [4]:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.516, + 0.905, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\stackrel {\\mathrm {i . i . d}} {\\sim} p (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} _ {i} - \\mathbf {z} _ {j} | | _ {2} ^ {2}} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.6 + ], + "angle": 0, + "content": "where \\( p(\\mathbf{z}) \\) is the distribution of embeddings obtained by mapping data samples through the encoder \\( f \\) and projector \\( g \\) and \\( t \\) is the temperature hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.649, + 0.632 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.761 + ], + "angle": 0, + "content": "In this section, we introduce Soft Separation and Distillation (SSD), a novel approach designed to enhance representation uniformity in federated learning (FL). We first analyze the limitations of uniformity in non-IID FL settings, then propose dimension-scaled regularization to softly separate client features, and finally introduce projector distillation to effectively transfer improved embedding uniformity to the representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Limited Inter-Client Uniformity of FL. Uniformity is a critical metric in representation learning [6, 30, 31], where higher uniformity indicates better preservation of information in the learned representations. In centralized training, jointly optimizing alignment loss \\(\\mathcal{L}_{\\mathrm{align}}\\) and uniformity loss \\(\\mathcal{L}_{\\mathrm{uniform}}\\) yields high-quality representations. However, the distributed nature of FL introduces a fundamental challenge to uniformity optimization. Specifically, the uniformity metric can be decomposed into intra-client and inter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2973" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.092, + 0.9, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.239, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Overview of our Training Pipeline. \\(①\\) The server initially assigns distinct weight vectors to each client. \\(②\\) Clients perform local training by optimizing their respective loss functions on their own data. \\(③\\) The server aggregates updated parameters from all clients. Steps \\(②\\) and \\(③\\) are iteratively repeated until convergence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.296, + 0.221, + 0.311 + ], + "angle": 0, + "content": "client components:" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.319, + 0.484, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\left(\\underbrace {\\sum_ {k = 1} ^ {K} \\mathbb {E} _ {\\mathbf {z} , \\mathbf {z} ^ {\\prime} \\sim p _ {k} (\\mathbf {z})} \\left[ e ^ {- t \\left| \\mathbf {z} - \\mathbf {z} ^ {\\prime} \\right| / 2} \\right]} _ {\\text {i n t r a - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k}}\\right) \\tag {5} \\\\ \\left. + \\underbrace {\\sum_ {i \\neq j} \\mathbb {E} _ {\\mathbf {z} \\sim p _ {i} (\\mathbf {z}) , \\mathbf {z} ^ {\\prime} \\sim p _ {j} (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | | _ {2} ^ {2}} \\right]} _ {\\text {i n t e r - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {(i, j)}}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.507 + ], + "angle": 0, + "content": "where the first term represents intra-client uniformity \\(\\mathcal{L}_{\\mathrm{uniform}}^k\\) and the second term represents inter-client uniformity \\(\\mathcal{L}_{\\mathrm{uniform}}^{(i,j)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.506, + 0.484, + 0.732 + ], + "angle": 0, + "content": "In federated learning, each client \\( i \\) optimizes its local loss function using samples drawn from its own local distribution \\( p_i \\), which means that only the intra-client term in the above equation is explicitly optimized. In a homogeneous setting, when client distributions are similar (i.e., \\( p_i \\approx p_j \\)), optimizing intra-client uniformity naturally promotes good inter-client uniformity. However, in non-IID settings where client distributions differ significantly, optimizing only intra-client uniformity does not guarantee good inter-client uniformity, potentially limiting the quality of the globally aggregated model. The core challenge in enhancing inter-client uniformity lies in the federated learning constraint that the server has no access to raw client data or embeddings, making it impossible to directly impose a loss function that operates across different clients." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Dimension-Scaled Regularization (DSR). To address the inter-client uniformity challenge, we propose Dimension-Scaled Regularization (DSR), which improves global uniformity by encouraging client embeddings to spread in different directions of the representation space. This approach effectively increases the separation between clients without enforcing rigid boundaries that might distort the intrinsic data structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For each client \\(k\\), we define a dimension-scaling vector \\(\\mathbf{d}_k \\in \\mathbb{R}^d\\), where \\(d\\) is the dimensionality of the embedding space. This vector applies selective scaling to specific di" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.905, + 0.326 + ], + "angle": 0, + "content": "mensions, with some dimensions scaled by a factor \\(\\alpha \\neq 1\\) in a client-specific manner:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.333, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\[\nd _ {i, k} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} i \\in \\mathcal {S} _ {k} \\\\ 1, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.455 + ], + "angle": 0, + "content": "where \\(S_{k}\\) is a set of dimensions uniquely assigned to client \\(k\\), ensuring that the scaled dimensions are nonoverlapping across clients (i.e., \\(S_{i} \\cap S_{j} = \\emptyset\\) for \\(i \\neq j\\)). The size of each set \\(S_{k}\\) is approximately \\(\\lfloor d / K \\rfloor\\), where \\(K\\) is the number of clients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.502 + ], + "angle": 0, + "content": "We then regularize each client's embeddings by encouraging them to move toward their dimension-scaled versions through the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.509, + 0.905, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D S R}} ^ {k} = \\mathbb {E} _ {\\mathbf {z} \\sim p _ {k} (\\mathbf {z})} \\left[ \\| \\mathbf {z} - \\operatorname {s t o p g r a d} \\left(\\mathbf {z} \\odot \\mathbf {d} _ {k}\\right) \\| _ {2} ^ {2} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.594 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents element-wise multiplication, and stopgrad \\((\\cdot)\\) prevents gradient flow through the scaled target, ensuring that we pull the original embedding toward the scaled version rather than vice versa." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.777 + ], + "angle": 0, + "content": "To understand why DSR enhances inter-client uniformity, we can analyze its effect on the representation distribution. When a vector \\(\\mathbf{z}\\) is scaled along specific dimensions by a factor \\(\\alpha > 1\\) and then normalized to the unit hypersphere, the resulting vector shifts toward those scaled dimensions. By assigning different scaling dimensions to each client, DSR effectively encourages client representations to occupy different regions of the unit hypersphere, as illustrated in Figure 5. Mathematically, if clients \\(i\\) and \\(j\\) have scaling vectors \\(\\mathbf{d}_i\\) and \\(\\mathbf{d}_j\\) with non-overlapping scaled dimensions, their optimized embeddings will tend to have a smaller dot product:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.784, + 0.905, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathbf {z} _ {i} \\sim p _ {i} (\\mathbf {z}), \\mathbf {z} _ {j} \\sim p _ {j} (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right] < \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\sim p (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\( p(\\mathbf{z}) \\) represents the distribution without DSR. This reduced dot product corresponds to increased angular separation, directly contributing to improved uniformity across the global distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Unlike a hard separation approach that would restrict each client to entirely separate subspaces (which could" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.36, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.092, + 0.63, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.092, + 0.902, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.241, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Representation norm, embedding norm, and DSR loss during training. Adding DSR primarily affects the embedding space, as reflected in the embedding norm, while leaving the representation norm unchanged. In contrast, incorporating PD directly influences the representation norm, and when combined with DSR, it leads to a more pronounced decrease in representation norm. Additionally, PD enhances the effect of DSR, resulting in a lower \\(\\mathcal{L}_{\\mathrm{DSR}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.322, + 0.28, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.322, + 0.444, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.537 + ], + "angle": 0, + "content": "Figure 5. Illustration of the effect of DSR. The left figure shows the original data, where arrows indicate the transformation direction after applying DSR. The right figure presents the data after transformation. DSR increases separation and enhances inter-client uniformity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.484, + 0.65 + ], + "angle": 0, + "content": "severely constrain representation capacity), our soft separation allows clients to share most dimensions while gently pushing them toward different directions. This balanced approach preserves the flexibility needed for effective representation learning while significantly improving inter-client uniformity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.789 + ], + "angle": 0, + "content": "An extreme example of fully separating embeddings would be assigning each client a dedicated subspace (e.g., if \\( d = 500 \\) and \\( K = 10 \\), each client could occupy 50 distinct dimensions). Although this configuration maximizes inter-client uniformity, it may disrupt the intrinsic data structure and miss out on the collaborative benefits of FL. In contrast, our soft DSR loss encourages some degree of separation without completely isolating the clients' embeddings. We analyze this tradeoff further in Section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Projector Distillation (PD). While DSR effectively enhances uniformity at the embedding level, we observe that this improvement may not fully transfer to the representation level. This discrepancy occurs because the projector \\( g(\\cdot) \\) placed between the encoder \\( f(\\cdot) \\) and the loss function can absorb much of the optimization effect, as shown in Figure 4. Although completely removing the projector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.905, + 0.386 + ], + "angle": 0, + "content": "might seem like a direct solution, prior work [2, 9, 32] has demonstrated that the projector plays a crucial role in separating optimization objectives from feature representations, thereby preventing overfitting to specific self-supervised tasks and improving downstream performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.905, + 0.432 + ], + "angle": 0, + "content": "To bridge this gap, we introduce Projector Distillation (PD), which explicitly aligns the encoder's representations with the projector's embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.441, + 0.905, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} ^ {k} = \\mathbb {E} _ {x \\sim p _ {k} (\\mathbf {x})} \\left[ D _ {\\mathrm {K L}} \\left(\\sigma (\\mathbf {h}) \\| \\sigma (\\mathbf {z})\\right) \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.905, + 0.529 + ], + "angle": 0, + "content": "where \\(\\mathbf{h} = f(\\mathbf{x})\\) is the representation from the encoder, \\(\\mathbf{z} = g(\\mathbf{h})\\) is the embedding from the projector, \\(\\sigma(\\cdot)\\) denotes the softmax function, and \\(D_{\\mathrm{KL}}\\) is the Kullback-Leibler divergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.636 + ], + "angle": 0, + "content": "This distillation mechanism encourages the encoder to internalize the beneficial structure learned in the embedding space, effectively transferring the improved uniformity from embeddings to representations. By minimizing the KL divergence between these distributions, we ensure that the uniformity improvements achieved through DSR are reflected in the representations used for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.783, + 0.661 + ], + "angle": 0, + "content": "3.1. Overview of Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.788 + ], + "angle": 0, + "content": "The complete SSD algorithm is outlined in Algorithm 1, with our key design components highlighted in red. The server initially assigns unique weight vectors to all clients for dimension-scaled regularization. During local training, each client optimizes a combination of the standard alignment and uniformity losses for self-supervised learning, augmented with our DSR and PD terms. The local training objective for each client \\( k \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.812, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {k} = \\mathcal {L} _ {\\text {a l i g n}} ^ {k} + \\beta \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k} + \\gamma \\mathcal {L} _ {\\text {D S R}} ^ {k} + \\delta \\mathcal {L} _ {\\text {d i s t i l l}} ^ {k}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.87 + ], + "angle": 0, + "content": "where \\(\\beta, \\gamma\\), and \\(\\delta\\) are hyperparameters that balance the contribution of each loss term." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "After local training, clients upload their updated models to the server, which aggregates them using standard" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2975" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.091, + 0.093, + 0.482, + 0.645 + ], + "angle": 0, + "content": "Algorithm 1: Soft Separation and Distillation (SSD) \nInput: communication rounds \\(T\\), local epochs \\(E\\), number of clients \\(K\\), dataset \\(\\mathcal{D} = \\cup_{k \\in [K]} \\mathcal{D}_k\\), augmentation function \\(\\mathcal{T}(\\cdot)\\) \nOutput: Final global encoder \\(f_\\theta^T\\) \n1 Server: \n2 Initialize global encoder \\(f_\\theta^0\\) and projector \\(g_\\theta^0\\); \n3 Assign weight vector \\(\\{\\mathbf{d}_k\\}_{k \\in [K]}\\) for all clients; \n4 for \\(t = 0\\) to \\(T - 1\\) do \n5 \\(S_t \\gets\\) (randomly select a set of clients); \n6 for each client \\(k \\in S_t\\) in parallel do \n7 \\(f_\\theta^{(t,k)}, g_\\theta^{(t,k)} \\gets \\text{Client}(f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)\\); \n8 end \n9 // Model aggregation; \n10 for each parameter \\(\\theta\\) in \\(f, g\\) do \n11 \\(\\theta^{t+1} \\gets \\sum_{k \\in S_t} w_k \\theta^{(t,k)}\\) \n12 end \n13 end \n14 return \\(f_\\theta^T\\); \n15 Client \\((f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)\\): \n16 // Initialize local models; \n17 \\(f \\gets f_\\theta^t, g \\gets g_\\theta^t\\); \n18 for each local epoch \\(e = 0\\) to \\(E - 1\\) do \n19 for each batch \\(\\{\\mathbf{x}_i\\}_{i=1}^{n_B} \\in \\mathcal{D}_k\\) do \n20 // Data augmentation; \n21 \\(\\tilde{\\mathbf{x}}_i, \\tilde{\\mathbf{x}}_i^+ = \\mathcal{T}(\\mathbf{x}_i), \\mathcal{T}(\\mathbf{x}_i)\\); \n22 // Get representations and embeddings; \n23 \\(\\mathbf{h}_i, \\mathbf{h}_i^+ = f(\\tilde{\\mathbf{x}}_i), f(\\tilde{\\mathbf{x}}_i^+)\\); \n24 \\(\\mathbf{z}_i, \\mathbf{z}_i^+ = g(\\mathbf{h}_i), g(\\mathbf{h}_i^+)\\); \n25 // Optimize model using combined loss; \n26 Update \\(f, g\\) by minimizing \\(\\mathcal{L}^k\\) (Eq (10)); \n27 end \n28 end \n29 return \\(f, g\\) // Model upload; \n30 return" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.725 + ], + "angle": 0, + "content": "federated averaging [24]: \\(\\theta^{t + 1} = \\sum_{k\\in S_t}w_k\\theta^{(t,k)}\\) where \\(w_{k} = \\frac{n_{k}}{\\sum_{i = 1}^{K}n_{i}}\\) represents the aggregation weight based on the client's data size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Importantly, our method preserves privacy since it requires no sharing of raw data. The only additional communication is the initial weight vectors \\(\\{\\mathbf{d}_k\\}_{k=1}^K\\), which are lightweight and contain no sensitive information. The computational overhead of SSD is minimal, making it highly practical for real-world federated learning deployments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.222, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our approach on CIFAR-10 and CIFAR-100, which contain 50K training images and 10K testing images, with 10 and 100 classes, respectively. Fol" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "lowing prior work [22, 37], we simulate non-IID data across \\( K \\) clients using a Dirichlet prior \\( \\mathrm{Dir}(\\alpha) \\), where a smaller \\( \\alpha \\) indicates a higher degree of data heterogeneity. We conduct experiments in two federated learning settings: cross-silo \\( (K = 10) \\) with full participation and cross-device \\( (K = 50) \\) with a participation rate of 0.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.187, + 0.905, + 0.293 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate our approach based on downstream performance and representation quality. For downstream performance, we follow prior work [10, 22, 36], using linear probing and fine-tuning with \\(1\\%\\) and \\(10\\%\\) of the data. For representation quality, we assess uniformity [6, 30, 31] as defined in Equation (4) and effective rank [7, 27, 34] (see the definition in Appendix A.2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Baselines. We compare our method against three categories of baselines. (1) Adapting a centralized algorithm to the federated setting: FedAlignUniform [31]. (2) FUL methods that mitigate divergence from the global model: FedX [10] and L-DAWA [25]. (3) FUL methods addressing dimensional collapse in local clients: FedDecorr [28]. FedU2 [22] incorporates modules that tackle both divergence control and dimensional collapse." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.905, + 0.528 + ], + "angle": 0, + "content": "Implementation Details. We follow prior work [22] for image augmentations, training configurations, and model architecture, using ResNet-18 [11] as the encoder and a two-layer linear projector. The hyperparameters \\(\\beta\\), \\(\\gamma\\), and \\(\\delta\\) are set to 1.0, 1.0, and 0.1, respectively, and the scaling factor \\(\\alpha\\) is set to 10.0. Further details are provided in Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.541, + 0.652, + 0.556 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.906, + 0.745 + ], + "angle": 0, + "content": "Performance on different FL settings. Table 1 compares the performance of various methods in cross-silo and crossdevice federated learning. SSD consistently achieves the highest accuracy and excels in fine-tuning tasks. FedAlignUniform, FedDecorr, and FedU2 are competitive but generally lag behind SSD, while FedX and L-DAWA—aimed at ensuring global model consistency—perform the worst across all settings. Although FedDecorr and FedU2 enhance intra-client uniformity (with FedU2 adding a consistency module), they still trail SSD. By improving inter-client uniformity, SSD underscores the importance of addressing inter-client variations in federated learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Representation quality. Figure 1 compares different methods based on effective rank [27] and uniformity [31], two key metrics for representation quality. Among state-of-the-art methods, FedX and L-DAWA, which emphasize global model consistency, achieve strong representation quality. In contrast, methods focusing on local uniformity, such as FedDecorr and FedU2, exhibit lower effective rank and uniformity, indicating that they fail to address global uniformity. Our method, SSD, surpasses all other federated approaches, achieving the highest effective rank and the best uniformity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2976" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Results on Cross-Silo and Cross-Device Settings. Accuracy (%) of linear probing (LP), fine-tuning (FT) \\(1\\%\\), and \\(10\\%\\) labeled data on CIFAR10 and CIFAR100 (\\(\\alpha = 0.1\\)). The highest score is highlighted in bold, and the second-highest score is underlined." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.905, + 0.261 + ], + "angle": 0, + "content": "
CIFAR10CIFAR100
Cross-Silo (K=10)Cross-Device (K=50)Cross-Silo (K=10)Cross-Device (K=50)
LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%
FedAlignUniform [31]80.8469.9981.0071.2857.4173.7757.2528.9748.9943.0316.3736.64
FedX [10]78.466.7880.0171.0156.9173.2457.3427.4649.5043.0716.0435.46
L-DAWA [25]77.6765.9479.3467.6553.7571.2256.9027.0849.5742.5815.5434.82
FedDecorr [28]80.1369.0980.3371.4958.1973.9757.2529.3849.5344.7417.6736.68
FedU2 [22]81.0169.6281.0171.0957.1574.2157.4029.3949.6442.9016.0835.48
SSD81.3270.7481.6771.8357.7774.6157.3829.5749.8745.2117.7036.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.284, + 0.486, + 0.327 + ], + "angle": 0, + "content": "Table 2. Generalization on OOD datasets. Accuracy \\((\\%)\\) of linear probing, uniformity, and effective rank when trained on CIFAR-100 and TinyImageNet-200, and evaluated on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.338, + 0.482, + 0.419 + ], + "angle": 0, + "content": "
CIRAR100 → CIFAR10TinyImageNet → CIFAR10
LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)
FedAlignUniform [31]77.663.6566.9979.863.7176.31
FedX [10]78.023.7384.8879.873.7693.48
L-DAWA [25]77.463.7184.7579.623.7493.41
FedDecorr [28]77.623.6674.2079.793.7284.93
FedU2 [22]77.743.6668.379.743.6975.97
SSD78.483.7386.9580.003.7798.45
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.473 + ], + "angle": 0, + "content": "This demonstrates that SSD effectively enhances representation quality, outperforming existing methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.582 + ], + "angle": 0, + "content": "Generalization on OOD datasets. Evaluating generalization to out-of-distribution (OOD) datasets helps assess whether models can learn transferable representations that perform well beyond their training distribution. Table 2 presents the generalization performance of different methods, where models trained on CIFAR-100 and TinyImageNet-200 are evaluated on CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Among state-of-the-art methods, FedX, which exhibits high representation quality but performs poorly in indistribution settings, achieves the best generalization in OOD scenarios. However, our method, SSD, consistently outperforms all baselines, achieving the highest accuracy in both settings. SSD also demonstrates the best uniformity [31] and effective rank [7], indicating superior feature representation quality. While FedAlignUniform, FedDecorr, L-DAWA and FedU2 perform competitively, they generally fall short of SSD, further reinforcing its effectiveness in handling distribution shifts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.194, + 0.774 + ], + "angle": 0, + "content": "4.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. We provide the ablation study in Table 3. It shows that adding projector distillation (PD) alone does not provide noticeable improvements. In contrast, adding dimension-scaled regularization (DSR) alone leads to slight improvements in LP and FT \\(10\\%\\), but the uniformity enhancement is minimal. However, when combining both DSR and PD (SSD), the method effectively transfers optimization benefits to representation quality, leading to a" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.284, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Table 3. Ablation Study. PD alone has minimal impact on performance. Adding DSR alone provides a slight improvement in both performance and uniformity. Combining both DSR and PD leads to a significant boost in uniformity and achieves the best performance." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.365, + 0.905, + 0.442 + ], + "angle": 0, + "content": "
LPFT 1%FT 10%-Luniform(↑)
FedAlignUniform [31]80.8469.9981.003.79
+PD80.7469.7880.713.80
+DSR81.0569.7781.153.81
+ DSR + PD (SSD)81.3270.7481.673.84
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.498 + ], + "angle": 0, + "content": "significant uniformity improvement and achieving the best overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.7 + ], + "angle": 0, + "content": "Robustness of DSR. We evaluate our method from two perspectives: the scaled factor \\(\\alpha\\) and the selection of scaled dimensions for each client. The scaled factor \\(\\alpha\\) influences the overall discrepancy, while the selection of scaled dimensions affects the direction each client is guided toward. To assess robustness, each \\(\\alpha\\) is evaluated with three different selections of scaled dimensions. As shown in Figure 6a, our method maintains stable performance across different \\(\\alpha\\) values, introducing minimal variation while consistently outperforming FedAlignUniform. This demonstrates that our approach effectively balances alignment and uniformity without being overly sensitive to the choice of \\(\\alpha\\) or the specific dimension selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Soft vs. hard client separation. We examine the impact of client feature separation on alignment, uniformity, and overall performance. Our intuition is that increasing separation between clients enhances global uniformity. One way to enforce this separation is by restricting each client to its own subspace without sharing information with others. To illustrate the effects of such strict partitioning, we consider Hard Separation and Distillation (HSD) as a baseline. As shown in Figure 6b, while HSD achieves the highest uniformity, it does so at the cost of severely reduced alignment, ultimately leading to poor downstream performance. This performance drop can be attributed to the disruption of intrinsic feature structures. In contrast, our proposed SSD balances" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.106, + 0.278, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.248, + 0.227, + 0.26 + ], + "angle": 0, + "content": "(a) Robustness" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.106, + 0.476, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.248, + 0.427, + 0.26 + ], + "angle": 0, + "content": "(b) Separation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Figure 6. (a) Robustness of DSR across different scaled factors \\(\\alpha\\), where each \\(\\alpha\\) is tested with three different randomly selected scaled dimensions. (b) Soft vs. hard separation. HSD achieves the highest uniformity but at the cost of reduced alignment, resulting in poor overall performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.398 + ], + "angle": 0, + "content": "Table 4. The Effect of the Projector. Without the projector, DSR significantly improves uniformity and performance. Yet, the overall results remain inferior to those achieved with the projector." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.409, + 0.443, + 0.482 + ], + "angle": 0, + "content": "
ProjectileLP-Luniform(↑)
FedAlignUniformX73.163.72
+ DSRX76.14 (+2.98)3.77 (+0.05)
FedAlignUniform✓80.843.79
+ DSR✓81.05 (+0.21)3.81 (+0.02)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.55 + ], + "angle": 0, + "content": "both alignment and uniformity, mitigating the drawbacks of hard separation while enhancing uniformity, resulting in improved overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.766 + ], + "angle": 0, + "content": "Why not remove projector or apply loss on representations? A natural idea for addressing the limited transfer of embedding-level loss optimization (in terms of representation uniformity) is to remove the projector or apply the loss directly to the representations. However, prior work [2, 9, 32] has demonstrated that the projector plays a critical role in preventing the encoder from overfitting on the upstream task. Indeed, as shown in Table 4, removing the projector and adding DSR does substantially improve uniformity and downstream performance, but these results still lag behind those achieved when the projector is retained. Therefore, it is crucial to develop a method that effectively promotes uniformity while preserving the benefits offered by the projector." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.779, + 0.233, + 0.794 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.429, + 0.819 + ], + "angle": 0, + "content": "5.1. Federated Learning with Non-IID Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Federated learning (FL) enables collaborative model training across multiple clients without centralizing data. Since client data distributions are typically non-IID, model convergence and performance are challenging. FedAvg [24], which aggregates local model updates using simple averag" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ing, is the foundational framework but suffers under heterogeneous data and imbalanced client participation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.907, + 0.287 + ], + "angle": 0, + "content": "To address these issues, various aggregation improvements have been proposed. FedProx [21] reduces client drift by adding a proximal term to limit local updates from straying from the global model. SCAFFOLD [16] corrects client drift using control variates to reduce variance in local updates. MOON [19] improves representation consistency with a contrastive loss. Other methods use globally shared data to improve generalization; FedShare [35] aligns client models using shared data, while FedDistill [13] distills knowledge from the global model to local models using shared data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.813, + 0.314 + ], + "angle": 0, + "content": "5.2. Federated Unsupervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Federated Unsupervised Learning (FUL) combines federated learning with un-/self-supervised learning, primarily addressing (1) non-IID data distribution and (2) representation collapse." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.906, + 0.546 + ], + "angle": 0, + "content": "To handle non-IID data, FedCA [33] uses a shared dictionary module for better aggregation but risks privacy leakage. FedU [36] reduces this risk by selectively uploading the online network's encoder and deciding predictor updates based on divergence. FedEMA [37] extends this with exponential moving average updates. FedX [10] adds an alignment/contrastive term with the global model, while Orchestra [23] preserves global structural consistency using global centroids. FedU2 [22] ensures balanced updates across clients for better alignment but does not guarantee global representation uniformity under non-IID settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.667 + ], + "angle": 0, + "content": "For representation collapse, FedDecorr [28] shows that local clients suffer from dimensional collapse, which propagates to the global model, and addresses it with a local decorrelation loss. FedU2 [22] encourages local representation uniformity by minimizing divergence with a spherical Gaussian. Our work extends this by showing that intra-client uniformity alone is insufficient inter-client uniformity must also be explicitly addressed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce Soft Separation & Distillation (SSD), a framework designed to enhance representation quality by improving inter-client uniformity in federated learning. SSD consists of a dimension-scaled regularization term that softly separates client embeddings while preserving the intrinsic data structure, and a projector distillation term that transfers the optimization benefits of the projector to the encoder, thereby improving representation quality. SSD achieves state-of-the-art performance in both representation learning and downstream tasks across various training and FL settings. Our work highlights the importance of global representation quality in federated unsupervised learning, opening new directions for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2978" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.328 + ], + "angle": 0, + "content": "We thank the anonymous reviewers and members of CLLab for their constructive feedback. This work is partially supported by the National Science and Technology Council in Taiwan via NSTC 113-2634-F-002-008, 114-2221-E002-102-MY3, and the National Taiwan University Center for Data Intelligence via NTU-114L900901 and Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410072, RGC R1015-23). We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. H.-T. Lin is honored to be supported by the Leap Fellowship of the Foundation for the Advancement of Outstanding Scholarship in Taiwan since 2025." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.345, + 0.188, + 0.36 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.37, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 1, 2, 5, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.482, + 0.554 + ], + "angle": 0, + "content": "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[4] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[6] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. In The Twelfth International Conference on Learning Representations, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[7] Quentin Garrido, Randall Balestriero, Laurent Najman, and Yann Lecun. Rankme: Assessing the downstream performance of pretrained self-supervised representations by their rank. In International conference on machine learning, pages 10929-10974. PMLR, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[8] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[9] Kartik Gupta, Thalaiyasingam Ajanthan, Anton van den Hengel, and Stephen Gould. Understanding and improving" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.37, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the role of projection head in self-supervised learning. arXiv preprint arXiv:2212.11491, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[10] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xing Xie, and Meeyoung Cha. Fedx: Unsupervised federated learning with cross knowledge distillation. In European Conference on Computer Vision, pages 691-707. Springer, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[13] Eunjeeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non- \\(\\mathrm{id}\\) private data. arXiv preprint arXiv:1811.11479, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.427 + ], + "angle": 0, + "content": "[14] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[15] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[16] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[17] Taehyeon Kim, Jaehoon Oh, Nak Yil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2628–2635. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.678, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[18] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1-40, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[19] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[20] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2979" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.188 + ], + "angle": 0, + "content": "[22] Xinting Liao, Weiming Liu, Chaochao Chen, Pengyang Zhou, Fengyuan Yu, Huabin Zhu, Binhui Yao, Tao Wang, Xiaolin Zheng, and Yanchao Tan. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22841-22850, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[23] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In Proceedings of the 39th International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[24] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[25] Yasar Abbas Ur Rehman, Yan Gao, Pedro Porto Buarque De Gusmão, Mina Alibeigi, Jiajun Shen, and Nicholas D Lane. L-dawa: Layer-wise divergence aware weight aggregation in federated self-supervised visual representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16464-16473, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[26] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22 (3):400 - 407, 1951. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[27] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[28] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[30] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[31] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, 2020. 1, 2, 3, 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[32] Yihao Xue, Eric Gan, Jiayi Ni, Siddharth Joshi, and Baharan Mirzasoleiman. Investigating the benefits of projection head for representation learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[33] Fengda Zhang, Kun Kuang, Long Chen, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Fei Wu, Yueting Zhuang, et al. Federated unsupervised representation learning. Frontiers of Information Technology & Electronic Engineering, 24(8):1181-1193, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Yifei Zhang, Hao Zhu, Yankai Chen, Zixing Song, Piotr Koniusz, and Irwin King. Mitigating the popularity bias of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "graph collaborative filtering: A dimensional collapse perspective. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[36] Weiming Zhuang, Xin Gan, Yonggang Wen, Shuai Zhang, and Shuai Yi. Collaborative unsupervised visual representation learning from decentralized data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4912-4921, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[37] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In International Conference on Learning Representations, 2022. 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2980" + } + ] +] \ No newline at end of file diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_origin.pdf b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f928433f2fa530a55e436c8d9f133d4bd523eca8 --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/a7c7f9f2-5e8e-4c1b-b520-1318204d80e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1e6c5614aab535cf28534ab265d44031741b32a910c259fa46696e151692f47 +size 2363568 diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/full.md b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2ba6e292da4336ef070cd009cda231b4496f6792 --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/full.md @@ -0,0 +1,349 @@ +# Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning + +Hung-Chieh Fang $^{1,*}$ Hsuan-Tien Lin $^{1}$ Irwin King $^{2}$ Yifei Zhang $^{2,\dagger}$ + +$^{1}$ National Taiwan University $^{2}$ The Chinese University of Hong Kong + +# Abstract + +Federated Unsupervised Learning (FUL) aims to learn expressive representations in federated and self-supervised settings. The quality of representations learned in FUL is usually determined by uniformity, a measure of how uniformly representations are distributed in the embedding space. However, existing solutions perform well in achieving intra-client (local) uniformity for local models while failing to achieve inter-client (global) uniformity after aggregation due to non-IID data distributions and the decentralized nature of FUL. To address this issue, we propose Soft Separation and Distillation (SSD), a novel approach that preserves inter-client uniformity by encouraging client representations to spread toward different directions. This design reduces interference during client model aggregation, thereby improving global uniformity while preserving local representation expressiveness. We further enhance this effect by introducing a projector distillation module to address the discrepancy between loss optimization and representation quality. We evaluate SSD in both cross-silo and cross-device federated settings, demonstrating consistent improvements in representation quality and task performance across various training scenarios. Our results highlight the importance of inter-client uniformity in FUL and establish SSD as an effective solution to this challenge. Project page: https://ssd-uniformity.github.io/. + +# 1. Introduction + +Deep learning has achieved remarkable success across a wide spectrum of applications, from computer vision and natural language processing to speech recognition and reinforcement learning [1, 11, 18, 29]. This progress can be largely attributed to the availability of massive labeled datasets, particularly since the emergence of ImageNet [5], + +![](images/9de7bcfe6166e3e1f81d9057d7b69aedb3259c202278684a2c16a5c780909ccb.jpg) +Figure 1. Comparison with state-of-the-art methods. Our SSD achieves the highest performance (marked in brackets) and the best representation quality, measured by uniformity [31] and effective rank [27], among existing FUL approaches. + +which has enabled the training of increasingly powerful neural networks. However, the practical deployment of deep learning in real-world scenarios faces two critical challenges. First, data in real-world applications is often non-Independent and Identically Distributed (non-IID) and cannot be freely shared due to privacy concerns, regulatory restrictions, or proprietary constraints. This data is typically generated across diverse sources, including user devices, healthcare institutions, and industrial settings. The distributed nature and sharing limitations of such data make centralized training approaches infeasible, leading to the development of Federated Learning (FL) [15, 20, 24] as a privacy-preserving distributed learning paradigm. Second, a vast proportion of available data remains unlabeled, necessitating effective methods for unsupervised representation learning [2, 3, 8, 12]. These methods aim to learn meaningful feature representations without relying on explicit la + +bels, enabling models to capture underlying data structures and semantic relationships. + +Federated Unsupervised Learning (FUL) emerges at the intersection of these two challenges, aiming to learn expressive representations in settings where data is both distributed and unlabeled. In FUL, the quality of learned representations is critically determined by two properties: alignment and uniformity [31]. Alignment measures how close similar data points are positioned in the representation space, relating to the model's ability to group semantically similar objects. Uniformity quantifies how evenly representations are distributed across the unit hypersphere, essentially measuring the entropy of the representation distribution. High uniformity prevents representation collapse and ensures that the learned features effectively utilize the available embedding dimensions. + +However, existing FUL methods face a significant challenge that remains inadequately addressed. While current approaches successfully achieve good intra-client (local) uniformity for representations within each client, they struggle to maintain inter-client (global) uniformity after model aggregation. This limitation stems from two key factors: (1) the non-IID distribution of data across clients, which naturally leads to divergent updates, and (2) the decentralized nature of FL, where the server lacks direct access to raw data, preventing the application of explicit uniformity constraints across clients. + +Most existing approaches have focused primarily on two aspects of this challenge. One line of research builds upon the framework established by FedProx [21], introducing proximal terms to constrain local updates within a bounded neighborhood of the global model. For instance, FedU [36] and FedEMA [37] dynamically adjust aggregation weights based on inter-client model divergence, while FedX [10] incorporates a global relational loss to align pairwise sample relationships across clients. These methods aim to maintain global model consistency but do not explicitly address representation uniformity. + +Another research direction tackles the issue of dimensional collapse [14] caused by lower uniformity in local representations. FedDecorr [28] demonstrates that embeddings of local clients are often less uniformly distributed and mitigates this issue by decorrelating local features. Similarly, FedU2 [22] regularizes local updates to approximate a spherical Gaussian distribution, encouraging isotropic feature spaces. While these approaches enhance local uniformity, the improvements at the client level do not inherently translate to better global uniformity during model aggregation. This leads to a crucial question: + +How can we effectively improve inter-client uniformity in Federated Unsupervised Learning? + +To address this challenge, we propose Soft Separation + +and Distillation (SSD), a novel method that enhances inter-client uniformity without compromising local representation expressiveness. Our approach employs a dimension-scaled regularization strategy that softly separates each client's feature space, encouraging client representations to spread toward different directions. As illustrated in Figure 2, this technique reduces interference during client model aggregation, thereby improving global uniformity without imposing rigid boundaries that could distort the underlying feature distributions. + +Furthermore, we empirically observe that the regularization effect of SSD may not always effectively transfer from loss optimization to the representation space due to the presence of a projector between the loss function and the encoder. While removing the projector might seem like a straightforward solution, doing so often degrades performance because the projector plays a crucial role in separating optimization objectives from feature representations [2, 9, 32]. To bridge this gap, we introduce a projector distillation module that minimizes the KL divergence between representations and embeddings, effectively encouraging the encoder to internalize the learned structure while preserving the projector's role in loss optimization. + +We evaluate SSD across diverse federated learning scenarios, including cross-silo (few clients with large datasets) and cross-device (many clients with small datasets) settings. Our experiments span both in-distribution tasks and out-of-distribution (OOD) datasets to assess generalizability. Results demonstrate that SSD consistently outperforms existing methods in both downstream task performance (measured via linear probing and fine-tuning accuracy) and representation quality (quantified using effective rank [7] and uniformity [31] metrics). + +Our main contributions are threefold: + +- We identify and formalize the challenge of inter-client uniformity in FUL, establishing it as a critical direction for decentralized unsupervised representation learning. +- We propose Soft Separation and Distillation (SSD), a simple yet effective framework that addresses inter-client uniformity without additional communication overhead or compromising privacy. +- We conduct extensive experiments across varied FL settings and tasks, confirming SSD's superiority over baseline methods in both performance and robustness. + +# 2. Preliminaries + +Federated unsupervised learning (FUL) is a learning paradigm in which multiple clients collaboratively train a model without sharing new raw data, and where the local datasets contain unlabeled data. The goal is to learn a global representation that generalizes across all participating clients. Consider a federated learning system with $K$ + +![](images/8263cb1e601bf03ab9cf5331c720d3c6a1a01f86d271957f17eba7846c89dc31.jpg) +Figure 2. Illustration of Intra-Client and Inter-Client Uniformity with Soft Separation. Intra-client uniformity ensures that representations within each client are well-distributed, while inter-client uniformity promotes global representation consistency across clients. Our proposed Soft Separation method encourages each client's representations to spread in distinct directions, mitigating interference during model aggregation and improving global uniformity. + +![](images/2fcd06304674a14763cc8f20b16021d5ae6548320739da7837754e1ae66ba792.jpg) + +local clients and a central server. Each client $k$ has access to an unlabeled local dataset $\mathcal{D}_k$ , modeled as samples from a client-specific data distribution $p_k(\mathbf{x})$ : + +$$ +\mathcal {D} _ {k} = \left\{\mathbf {x} _ {i} ^ {k} \right\} _ {i = 1} ^ {n _ {k}} \sim p _ {k} (\mathbf {x}), \tag {1} +$$ + +where $n_k$ denotes the number of samples in $\mathcal{D}_k$ . The global objective in FUL can be formulated as: + +$$ +\min _ {\theta} \sum_ {k = 1} ^ {K} P _ {k} \mathbb {E} _ {\mathbf {x} \sim p _ {k} (\mathbf {x})} \left[ \mathcal {L} _ {k} (\theta ; \mathbf {x}) \right], \tag {2} +$$ + +where $P_{k}$ is the probability of drawing a data from client $k$ , and $\mathcal{L}_k(\cdot)$ denotes the local unsupervised loss of client $k$ . + +Self-Supervised Representation Learning aims to learn meaningful representations that capture the underlying structure of data. A typical framework consists of the following components: + +- Encoder $f(\cdot)$ : Maps (augmented) input data $\tilde{\mathbf{x}} \sim \mathcal{T}(\mathbf{x})$ into a representation $\mathbf{h} = f(\tilde{\mathbf{x}}) \in \mathbb{R}^d$ , where $d$ is the hidden dimension. It is common to use a deep neural network (e.g., ResNet50) as the encoder. +- **Projector $g(\cdot)$ : Transforms the representation $\mathbf{h}$ into an embedding $\mathbf{z} = g(\mathbf{h}) \in \mathbb{R}^d$ . The training loss function is usually applied on this embedding space, and the projector is often a small network such as a multilayer perceptron. + +A fundamental principle in self-supervised representation learning is that similar samples should have similar representations. In other words, representations should be invariant to minor variations in input data. This is typically achieved by aligning the representations of two augmented versions of the same image (referred to as a positive pair) using the following alignment loss: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \mathbb {E} _ {\mathbf {x} \sim p (\mathbf {x}), \tilde {\mathbf {x}}, \tilde {\mathbf {x}} ^ {+} \sim \mathcal {T} (\mathbf {x})} | | \mathbf {z} - \mathbf {z} ^ {+} | | _ {2} ^ {2}, \tag {3} +$$ + +where $\tilde{\mathbf{x}},\tilde{\mathbf{x}}^{+}\sim \mathcal{T}(\mathbf{x})$ denotes two independent augmentations of the same sample $\mathbf{x}$ and $\mathbf{z} = g(f(\tilde{\mathbf{x}})),\mathbf{z}^{+} = g(f(\tilde{\mathbf{x}}^{+}))$ are the corresponding embeddings. + +Although aligning positive pairs encourages similarity, it can lead to a degenerate collapse where all representations converge to a single point. To circumvent this, Wang and Isola [31] introduces a uniformity objective to ensure that features are evenly dispersed on the unit hypersphere. This uniformity is often measured using the log of the average Gaussian potential [4]: + +$$ +\mathcal {L} _ {\text {u n i f o r m}} = - \log \mathbb {E} _ {\mathbf {z} _ {i}, \mathbf {z} _ {j} \stackrel {\mathrm {i . i . d}} {\sim} p (\mathbf {z})} \left[ e ^ {- t | | \mathbf {z} _ {i} - \mathbf {z} _ {j} | | _ {2} ^ {2}} \right] \tag {4} +$$ + +where $p(\mathbf{z})$ is the distribution of embeddings obtained by mapping data samples through the encoder $f$ and projector $g$ and $t$ is the temperature hyperparameter. + +# 3. Methodology + +In this section, we introduce Soft Separation and Distillation (SSD), a novel approach designed to enhance representation uniformity in federated learning (FL). We first analyze the limitations of uniformity in non-IID FL settings, then propose dimension-scaled regularization to softly separate client features, and finally introduce projector distillation to effectively transfer improved embedding uniformity to the representations. + +Limited Inter-Client Uniformity of FL. Uniformity is a critical metric in representation learning [6, 30, 31], where higher uniformity indicates better preservation of information in the learned representations. In centralized training, jointly optimizing alignment loss $\mathcal{L}_{\mathrm{align}}$ and uniformity loss $\mathcal{L}_{\mathrm{uniform}}$ yields high-quality representations. However, the distributed nature of FL introduces a fundamental challenge to uniformity optimization. Specifically, the uniformity metric can be decomposed into intra-client and inter + +![](images/e6d799e2047d6309e53442249fc68e3e14073f9757fd68eccb1928973d710f04.jpg) +Figure 3. Overview of our Training Pipeline. $①$ The server initially assigns distinct weight vectors to each client. $②$ Clients perform local training by optimizing their respective loss functions on their own data. $③$ The server aggregates updated parameters from all clients. Steps $②$ and $③$ are iteratively repeated until convergence. + +client components: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {u n i f o r m}} = - \log \left(\underbrace {\sum_ {k = 1} ^ {K} \mathbb {E} _ {\mathbf {z} , \mathbf {z} ^ {\prime} \sim p _ {k} (\mathbf {z})} \left[ e ^ {- t \left| \mathbf {z} - \mathbf {z} ^ {\prime} \right| / 2} \right]} _ {\text {i n t r a - c l i e n t} \mathcal {L} _ {\text {u n i f o r m}} ^ {k}}\right) \tag {5} \\ \left. + \underbrace {\sum_ {i \neq j} \mathbb {E} _ {\mathbf {z} \sim p _ {i} (\mathbf {z}) , \mathbf {z} ^ {\prime} \sim p _ {j} (\mathbf {z})} \left[ e ^ {- t | | \mathbf {z} - \mathbf {z} ^ {\prime} | | _ {2} ^ {2}} \right]} _ {\text {i n t e r - c l i e n t} \mathcal {L} _ {\text {u n i f o r m}} ^ {(i, j)}}\right), \\ \end{array} +$$ + +where the first term represents intra-client uniformity $\mathcal{L}_{\mathrm{uniform}}^k$ and the second term represents inter-client uniformity $\mathcal{L}_{\mathrm{uniform}}^{(i,j)}$ . + +In federated learning, each client $i$ optimizes its local loss function using samples drawn from its own local distribution $p_i$ , which means that only the intra-client term in the above equation is explicitly optimized. In a homogeneous setting, when client distributions are similar (i.e., $p_i \approx p_j$ ), optimizing intra-client uniformity naturally promotes good inter-client uniformity. However, in non-IID settings where client distributions differ significantly, optimizing only intra-client uniformity does not guarantee good inter-client uniformity, potentially limiting the quality of the globally aggregated model. The core challenge in enhancing inter-client uniformity lies in the federated learning constraint that the server has no access to raw client data or embeddings, making it impossible to directly impose a loss function that operates across different clients. + +Dimension-Scaled Regularization (DSR). To address the inter-client uniformity challenge, we propose Dimension-Scaled Regularization (DSR), which improves global uniformity by encouraging client embeddings to spread in different directions of the representation space. This approach effectively increases the separation between clients without enforcing rigid boundaries that might distort the intrinsic data structure. + +For each client $k$ , we define a dimension-scaling vector $\mathbf{d}_k \in \mathbb{R}^d$ , where $d$ is the dimensionality of the embedding space. This vector applies selective scaling to specific di + +mensions, with some dimensions scaled by a factor $\alpha \neq 1$ in a client-specific manner: + +$$ +d _ {i, k} = \left\{ \begin{array}{l l} \alpha , & \text {i f} i \in \mathcal {S} _ {k} \\ 1, & \text {o t h e r w i s e}, \end{array} \right. \tag {6} +$$ + +where $S_{k}$ is a set of dimensions uniquely assigned to client $k$ , ensuring that the scaled dimensions are nonoverlapping across clients (i.e., $S_{i} \cap S_{j} = \emptyset$ for $i \neq j$ ). The size of each set $S_{k}$ is approximately $\lfloor d / K \rfloor$ , where $K$ is the number of clients. + +We then regularize each client's embeddings by encouraging them to move toward their dimension-scaled versions through the following loss: + +$$ +\mathcal {L} _ {\mathrm {D S R}} ^ {k} = \mathbb {E} _ {\mathbf {z} \sim p _ {k} (\mathbf {z})} \left[ \| \mathbf {z} - \operatorname {s t o p g r a d} \left(\mathbf {z} \odot \mathbf {d} _ {k}\right) \| _ {2} ^ {2} \right], \tag {7} +$$ + +where $\odot$ represents element-wise multiplication, and stopgrad $(\cdot)$ prevents gradient flow through the scaled target, ensuring that we pull the original embedding toward the scaled version rather than vice versa. + +To understand why DSR enhances inter-client uniformity, we can analyze its effect on the representation distribution. When a vector $\mathbf{z}$ is scaled along specific dimensions by a factor $\alpha > 1$ and then normalized to the unit hypersphere, the resulting vector shifts toward those scaled dimensions. By assigning different scaling dimensions to each client, DSR effectively encourages client representations to occupy different regions of the unit hypersphere, as illustrated in Figure 5. Mathematically, if clients $i$ and $j$ have scaling vectors $\mathbf{d}_i$ and $\mathbf{d}_j$ with non-overlapping scaled dimensions, their optimized embeddings will tend to have a smaller dot product: + +$$ +\mathbb {E} _ {\mathbf {z} _ {i} \sim p _ {i} (\mathbf {z}), \mathbf {z} _ {j} \sim p _ {j} (\mathbf {z})} \left[ \mathbf {z} _ {i} ^ {\top} \mathbf {z} _ {j} \right] < \mathbb {E} _ {\mathbf {z} _ {i}, \mathbf {z} _ {j} \sim p (\mathbf {z})} \left[ \mathbf {z} _ {i} ^ {\top} \mathbf {z} _ {j} \right], \tag {8} +$$ + +where $p(\mathbf{z})$ represents the distribution without DSR. This reduced dot product corresponds to increased angular separation, directly contributing to improved uniformity across the global distribution. + +Unlike a hard separation approach that would restrict each client to entirely separate subspaces (which could + +![](images/293ad20cd014bd41b353af42752f8e3d426a213b03895cd37947c77c2dbdd256.jpg) +Figure 4. Representation norm, embedding norm, and DSR loss during training. Adding DSR primarily affects the embedding space, as reflected in the embedding norm, while leaving the representation norm unchanged. In contrast, incorporating PD directly influences the representation norm, and when combined with DSR, it leads to a more pronounced decrease in representation norm. Additionally, PD enhances the effect of DSR, resulting in a lower $\mathcal{L}_{\mathrm{DSR}}$ . + +![](images/4b7786b85c457eaf192d008aa99161c863902d0e279ff1f63ef93bdf2674a26a.jpg) + +![](images/73b86710d64cc8088f62a4242cbbb70349561247d702a8957aefe969378797be.jpg) + +![](images/9a23779c534c4f356fc90616c05d3f9f40a437194fad3466f0309737cf1a7538.jpg) +Figure 5. Illustration of the effect of DSR. The left figure shows the original data, where arrows indicate the transformation direction after applying DSR. The right figure presents the data after transformation. DSR increases separation and enhances inter-client uniformity. + +![](images/d4c462cabe02ba7b71971ac33fa5cf9fea5f3d464edbfb9f1c06e5ccf907ea1e.jpg) + +severely constrain representation capacity), our soft separation allows clients to share most dimensions while gently pushing them toward different directions. This balanced approach preserves the flexibility needed for effective representation learning while significantly improving inter-client uniformity. + +An extreme example of fully separating embeddings would be assigning each client a dedicated subspace (e.g., if $d = 500$ and $K = 10$ , each client could occupy 50 distinct dimensions). Although this configuration maximizes inter-client uniformity, it may disrupt the intrinsic data structure and miss out on the collaborative benefits of FL. In contrast, our soft DSR loss encourages some degree of separation without completely isolating the clients' embeddings. We analyze this tradeoff further in Section 4.2. + +Projector Distillation (PD). While DSR effectively enhances uniformity at the embedding level, we observe that this improvement may not fully transfer to the representation level. This discrepancy occurs because the projector $g(\cdot)$ placed between the encoder $f(\cdot)$ and the loss function can absorb much of the optimization effect, as shown in Figure 4. Although completely removing the projector + +might seem like a direct solution, prior work [2, 9, 32] has demonstrated that the projector plays a crucial role in separating optimization objectives from feature representations, thereby preventing overfitting to specific self-supervised tasks and improving downstream performance. + +To bridge this gap, we introduce Projector Distillation (PD), which explicitly aligns the encoder's representations with the projector's embeddings: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} ^ {k} = \mathbb {E} _ {x \sim p _ {k} (\mathbf {x})} \left[ D _ {\mathrm {K L}} \left(\sigma (\mathbf {h}) \| \sigma (\mathbf {z})\right) \right], \tag {9} +$$ + +where $\mathbf{h} = f(\mathbf{x})$ is the representation from the encoder, $\mathbf{z} = g(\mathbf{h})$ is the embedding from the projector, $\sigma(\cdot)$ denotes the softmax function, and $D_{\mathrm{KL}}$ is the Kullback-Leibler divergence. + +This distillation mechanism encourages the encoder to internalize the beneficial structure learned in the embedding space, effectively transferring the improved uniformity from embeddings to representations. By minimizing the KL divergence between these distributions, we ensure that the uniformity improvements achieved through DSR are reflected in the representations used for downstream tasks. + +# 3.1. Overview of Training Pipeline + +The complete SSD algorithm is outlined in Algorithm 1, with our key design components highlighted in red. The server initially assigns unique weight vectors to all clients for dimension-scaled regularization. During local training, each client optimizes a combination of the standard alignment and uniformity losses for self-supervised learning, augmented with our DSR and PD terms. The local training objective for each client $k$ is formulated as: + +$$ +\mathcal {L} ^ {k} = \mathcal {L} _ {\text {a l i g n}} ^ {k} + \beta \mathcal {L} _ {\text {u n i f o r m}} ^ {k} + \gamma \mathcal {L} _ {\text {D S R}} ^ {k} + \delta \mathcal {L} _ {\text {d i s t i l l}} ^ {k}, \tag {10} +$$ + +where $\beta, \gamma$ , and $\delta$ are hyperparameters that balance the contribution of each loss term. + +After local training, clients upload their updated models to the server, which aggregates them using standard + +Algorithm 1: Soft Separation and Distillation (SSD) +Input: communication rounds $T$ , local epochs $E$ , number of clients $K$ , dataset $\mathcal{D} = \cup_{k \in [K]} \mathcal{D}_k$ , augmentation function $\mathcal{T}(\cdot)$ +Output: Final global encoder $f_\theta^T$ +1 Server: +2 Initialize global encoder $f_\theta^0$ and projector $g_\theta^0$ ; +3 Assign weight vector $\{\mathbf{d}_k\}_{k \in [K]}$ for all clients; +4 for $t = 0$ to $T - 1$ do +5 $S_t \gets$ (randomly select a set of clients); +6 for each client $k \in S_t$ in parallel do +7 $f_\theta^{(t,k)}, g_\theta^{(t,k)} \gets \text{Client}(f_\theta^t, g_\theta^t, \mathbf{d}_k)$ ; +8 end +9 // Model aggregation; +10 for each parameter $\theta$ in $f, g$ do +11 $\theta^{t+1} \gets \sum_{k \in S_t} w_k \theta^{(t,k)}$ +12 end +13 end +14 return $f_\theta^T$ ; +15 Client $(f_\theta^t, g_\theta^t, \mathbf{d}_k)$ : +16 // Initialize local models; +17 $f \gets f_\theta^t, g \gets g_\theta^t$ ; +18 for each local epoch $e = 0$ to $E - 1$ do +19 for each batch $\{\mathbf{x}_i\}_{i=1}^{n_B} \in \mathcal{D}_k$ do +20 // Data augmentation; +21 $\tilde{\mathbf{x}}_i, \tilde{\mathbf{x}}_i^+ = \mathcal{T}(\mathbf{x}_i), \mathcal{T}(\mathbf{x}_i)$ ; +22 // Get representations and embeddings; +23 $\mathbf{h}_i, \mathbf{h}_i^+ = f(\tilde{\mathbf{x}}_i), f(\tilde{\mathbf{x}}_i^+)$ ; +24 $\mathbf{z}_i, \mathbf{z}_i^+ = g(\mathbf{h}_i), g(\mathbf{h}_i^+)$ ; +25 // Optimize model using combined loss; +26 Update $f, g$ by minimizing $\mathcal{L}^k$ (Eq (10)); +27 end +28 end +29 return $f, g$ // Model upload; +30 return + +federated averaging [24]: $\theta^{t + 1} = \sum_{k\in S_t}w_k\theta^{(t,k)}$ where $w_{k} = \frac{n_{k}}{\sum_{i = 1}^{K}n_{i}}$ represents the aggregation weight based on the client's data size. + +Importantly, our method preserves privacy since it requires no sharing of raw data. The only additional communication is the initial weight vectors $\{\mathbf{d}_k\}_{k=1}^K$ , which are lightweight and contain no sensitive information. The computational overhead of SSD is minimal, making it highly practical for real-world federated learning deployments. + +# 4. Experiments + +Datasets. We evaluate our approach on CIFAR-10 and CIFAR-100, which contain 50K training images and 10K testing images, with 10 and 100 classes, respectively. Fol + +lowing prior work [22, 37], we simulate non-IID data across $K$ clients using a Dirichlet prior $\mathrm{Dir}(\alpha)$ , where a smaller $\alpha$ indicates a higher degree of data heterogeneity. We conduct experiments in two federated learning settings: cross-silo $(K = 10)$ with full participation and cross-device $(K = 50)$ with a participation rate of 0.2. + +Evaluation Metrics. We evaluate our approach based on downstream performance and representation quality. For downstream performance, we follow prior work [10, 22, 36], using linear probing and fine-tuning with $1\%$ and $10\%$ of the data. For representation quality, we assess uniformity [6, 30, 31] as defined in Equation (4) and effective rank [7, 27, 34] (see the definition in Appendix A.2). + +Baselines. We compare our method against three categories of baselines. (1) Adapting a centralized algorithm to the federated setting: FedAlignUniform [31]. (2) FUL methods that mitigate divergence from the global model: FedX [10] and L-DAWA [25]. (3) FUL methods addressing dimensional collapse in local clients: FedDecorr [28]. FedU2 [22] incorporates modules that tackle both divergence control and dimensional collapse. + +Implementation Details. We follow prior work [22] for image augmentations, training configurations, and model architecture, using ResNet-18 [11] as the encoder and a two-layer linear projector. The hyperparameters $\beta$ , $\gamma$ , and $\delta$ are set to 1.0, 1.0, and 0.1, respectively, and the scaling factor $\alpha$ is set to 10.0. Further details are provided in Appendix A.1. + +# 4.1. Main Results + +Performance on different FL settings. Table 1 compares the performance of various methods in cross-silo and crossdevice federated learning. SSD consistently achieves the highest accuracy and excels in fine-tuning tasks. FedAlignUniform, FedDecorr, and FedU2 are competitive but generally lag behind SSD, while FedX and L-DAWA—aimed at ensuring global model consistency—perform the worst across all settings. Although FedDecorr and FedU2 enhance intra-client uniformity (with FedU2 adding a consistency module), they still trail SSD. By improving inter-client uniformity, SSD underscores the importance of addressing inter-client variations in federated learning. + +Representation quality. Figure 1 compares different methods based on effective rank [27] and uniformity [31], two key metrics for representation quality. Among state-of-the-art methods, FedX and L-DAWA, which emphasize global model consistency, achieve strong representation quality. In contrast, methods focusing on local uniformity, such as FedDecorr and FedU2, exhibit lower effective rank and uniformity, indicating that they fail to address global uniformity. Our method, SSD, surpasses all other federated approaches, achieving the highest effective rank and the best uniformity. + +Table 1. Results on Cross-Silo and Cross-Device Settings. Accuracy (%) of linear probing (LP), fine-tuning (FT) $1\%$ , and $10\%$ labeled data on CIFAR10 and CIFAR100 ( $\alpha = 0.1$ ). The highest score is highlighted in bold, and the second-highest score is underlined. + +
CIFAR10CIFAR100
Cross-Silo (K=10)Cross-Device (K=50)Cross-Silo (K=10)Cross-Device (K=50)
LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%
FedAlignUniform [31]80.8469.9981.0071.2857.4173.7757.2528.9748.9943.0316.3736.64
FedX [10]78.466.7880.0171.0156.9173.2457.3427.4649.5043.0716.0435.46
L-DAWA [25]77.6765.9479.3467.6553.7571.2256.9027.0849.5742.5815.5434.82
FedDecorr [28]80.1369.0980.3371.4958.1973.9757.2529.3849.5344.7417.6736.68
FedU2 [22]81.0169.6281.0171.0957.1574.2157.4029.3949.6442.9016.0835.48
SSD81.3270.7481.6771.8357.7774.6157.3829.5749.8745.2117.7036.82
+ +Table 2. Generalization on OOD datasets. Accuracy $(\%)$ of linear probing, uniformity, and effective rank when trained on CIFAR-100 and TinyImageNet-200, and evaluated on CIFAR-10. + +
CIRAR100 → CIFAR10TinyImageNet → CIFAR10
LP- \( \mathcal{L}_{\text{uniform}}(\uparrow) \)ERank (↑)LP- \( \mathcal{L}_{\text{uniform}}(\uparrow) \)ERank (↑)
FedAlignUniform [31]77.663.6566.9979.863.7176.31
FedX [10]78.023.7384.8879.873.7693.48
L-DAWA [25]77.463.7184.7579.623.7493.41
FedDecorr [28]77.623.6674.2079.793.7284.93
FedU2 [22]77.743.6668.379.743.6975.97
SSD78.483.7386.9580.003.7798.45
+ +This demonstrates that SSD effectively enhances representation quality, outperforming existing methods. + +Generalization on OOD datasets. Evaluating generalization to out-of-distribution (OOD) datasets helps assess whether models can learn transferable representations that perform well beyond their training distribution. Table 2 presents the generalization performance of different methods, where models trained on CIFAR-100 and TinyImageNet-200 are evaluated on CIFAR-10. + +Among state-of-the-art methods, FedX, which exhibits high representation quality but performs poorly in indistribution settings, achieves the best generalization in OOD scenarios. However, our method, SSD, consistently outperforms all baselines, achieving the highest accuracy in both settings. SSD also demonstrates the best uniformity [31] and effective rank [7], indicating superior feature representation quality. While FedAlignUniform, FedDecorr, L-DAWA and FedU2 perform competitively, they generally fall short of SSD, further reinforcing its effectiveness in handling distribution shifts. + +# 4.2. Analysis + +Ablation Study. We provide the ablation study in Table 3. It shows that adding projector distillation (PD) alone does not provide noticeable improvements. In contrast, adding dimension-scaled regularization (DSR) alone leads to slight improvements in LP and FT $10\%$ , but the uniformity enhancement is minimal. However, when combining both DSR and PD (SSD), the method effectively transfers optimization benefits to representation quality, leading to a + +Table 3. Ablation Study. PD alone has minimal impact on performance. Adding DSR alone provides a slight improvement in both performance and uniformity. Combining both DSR and PD leads to a significant boost in uniformity and achieves the best performance. + +
LPFT 1%FT 10%-Luniform(↑)
FedAlignUniform [31]80.8469.9981.003.79
+PD80.7469.7880.713.80
+DSR81.0569.7781.153.81
+ DSR + PD (SSD)81.3270.7481.673.84
+ +significant uniformity improvement and achieving the best overall performance. + +Robustness of DSR. We evaluate our method from two perspectives: the scaled factor $\alpha$ and the selection of scaled dimensions for each client. The scaled factor $\alpha$ influences the overall discrepancy, while the selection of scaled dimensions affects the direction each client is guided toward. To assess robustness, each $\alpha$ is evaluated with three different selections of scaled dimensions. As shown in Figure 6a, our method maintains stable performance across different $\alpha$ values, introducing minimal variation while consistently outperforming FedAlignUniform. This demonstrates that our approach effectively balances alignment and uniformity without being overly sensitive to the choice of $\alpha$ or the specific dimension selection. + +Soft vs. hard client separation. We examine the impact of client feature separation on alignment, uniformity, and overall performance. Our intuition is that increasing separation between clients enhances global uniformity. One way to enforce this separation is by restricting each client to its own subspace without sharing information with others. To illustrate the effects of such strict partitioning, we consider Hard Separation and Distillation (HSD) as a baseline. As shown in Figure 6b, while HSD achieves the highest uniformity, it does so at the cost of severely reduced alignment, ultimately leading to poor downstream performance. This performance drop can be attributed to the disruption of intrinsic feature structures. In contrast, our proposed SSD balances + +![](images/2543d03136fed646838407ea320da77c9d65b100fc9fc8549463b4ff2ff7c37e.jpg) +(a) Robustness +Figure 6. (a) Robustness of DSR across different scaled factors $\alpha$ , where each $\alpha$ is tested with three different randomly selected scaled dimensions. (b) Soft vs. hard separation. HSD achieves the highest uniformity but at the cost of reduced alignment, resulting in poor overall performance. + +![](images/08d31bfc5fc52f049028f3a49080672d2c9485c1cda065c7337028517ca1696f.jpg) +(b) Separation + +Table 4. The Effect of the Projector. Without the projector, DSR significantly improves uniformity and performance. Yet, the overall results remain inferior to those achieved with the projector. + +
ProjectileLP-Luniform(↑)
FedAlignUniformX73.163.72
+ DSRX76.14 (+2.98)3.77 (+0.05)
FedAlignUniform✓80.843.79
+ DSR✓81.05 (+0.21)3.81 (+0.02)
+ +both alignment and uniformity, mitigating the drawbacks of hard separation while enhancing uniformity, resulting in improved overall performance. + +Why not remove projector or apply loss on representations? A natural idea for addressing the limited transfer of embedding-level loss optimization (in terms of representation uniformity) is to remove the projector or apply the loss directly to the representations. However, prior work [2, 9, 32] has demonstrated that the projector plays a critical role in preventing the encoder from overfitting on the upstream task. Indeed, as shown in Table 4, removing the projector and adding DSR does substantially improve uniformity and downstream performance, but these results still lag behind those achieved when the projector is retained. Therefore, it is crucial to develop a method that effectively promotes uniformity while preserving the benefits offered by the projector. + +# 5. Related Work + +# 5.1. Federated Learning with Non-IID Data + +Federated learning (FL) enables collaborative model training across multiple clients without centralizing data. Since client data distributions are typically non-IID, model convergence and performance are challenging. FedAvg [24], which aggregates local model updates using simple averag + +ing, is the foundational framework but suffers under heterogeneous data and imbalanced client participation. + +To address these issues, various aggregation improvements have been proposed. FedProx [21] reduces client drift by adding a proximal term to limit local updates from straying from the global model. SCAFFOLD [16] corrects client drift using control variates to reduce variance in local updates. MOON [19] improves representation consistency with a contrastive loss. Other methods use globally shared data to improve generalization; FedShare [35] aligns client models using shared data, while FedDistill [13] distills knowledge from the global model to local models using shared data. + +# 5.2. Federated Unsupervised Learning + +Federated Unsupervised Learning (FUL) combines federated learning with un-/self-supervised learning, primarily addressing (1) non-IID data distribution and (2) representation collapse. + +To handle non-IID data, FedCA [33] uses a shared dictionary module for better aggregation but risks privacy leakage. FedU [36] reduces this risk by selectively uploading the online network's encoder and deciding predictor updates based on divergence. FedEMA [37] extends this with exponential moving average updates. FedX [10] adds an alignment/contrastive term with the global model, while Orchestra [23] preserves global structural consistency using global centroids. FedU2 [22] ensures balanced updates across clients for better alignment but does not guarantee global representation uniformity under non-IID settings. + +For representation collapse, FedDecorr [28] shows that local clients suffer from dimensional collapse, which propagates to the global model, and addresses it with a local decorrelation loss. FedU2 [22] encourages local representation uniformity by minimizing divergence with a spherical Gaussian. Our work extends this by showing that intra-client uniformity alone is insufficient inter-client uniformity must also be explicitly addressed. + +# 6. Conclusion + +We introduce Soft Separation & Distillation (SSD), a framework designed to enhance representation quality by improving inter-client uniformity in federated learning. SSD consists of a dimension-scaled regularization term that softly separates client embeddings while preserving the intrinsic data structure, and a projector distillation term that transfers the optimization benefits of the projector to the encoder, thereby improving representation quality. SSD achieves state-of-the-art performance in both representation learning and downstream tasks across various training and FL settings. Our work highlights the importance of global representation quality in federated unsupervised learning, opening new directions for future research. + +# Acknowledgment + +We thank the anonymous reviewers and members of CLLab for their constructive feedback. This work is partially supported by the National Science and Technology Council in Taiwan via NSTC 113-2634-F-002-008, 114-2221-E002-102-MY3, and the National Taiwan University Center for Data Intelligence via NTU-114L900901 and Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410072, RGC R1015-23). We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. H.-T. Lin is honored to be supported by the Leap Fellowship of the Foundation for the Advancement of Outstanding Scholarship in Taiwan since 2025. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 1 +[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 1, 2, 5, 8, 11 +[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1 +[4] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1 +[6] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. In The Twelfth International Conference on Learning Representations, 2024. 3, 6 +[7] Quentin Garrido, Randall Balestriero, Laurent Najman, and Yann Lecun. Rankme: Assessing the downstream performance of pretrained self-supervised representations by their rank. In International conference on machine learning, pages 10929-10974. PMLR, 2023. 2, 6, 7 +[8] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1 +[9] Kartik Gupta, Thalaiyasingam Ajanthan, Anton van den Hengel, and Stephen Gould. Understanding and improving + +the role of projection head in self-supervised learning. arXiv preprint arXiv:2212.11491, 2022. 2, 5, 8 +[10] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xing Xie, and Meeyoung Cha. Fedx: Unsupervised federated learning with cross knowledge distillation. In European Conference on Computer Vision, pages 691-707. Springer, 2022. 2, 6, 7, 8 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6, 11 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1 +[13] Eunjeeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non- $\mathrm{id}$ private data. arXiv preprint arXiv:1811.11479, 2018. 8 +[14] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. In ICLR, 2022. 2 +[15] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1 +[16] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 8 +[17] Taehyeon Kim, Jaehoon Oh, Nak Yil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2628–2635. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 11 +[18] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1-40, 2016. 1 +[19] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 8 +[20] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1 +[21] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 2, 8 + +[22] Xinting Liao, Weiming Liu, Chaochao Chen, Pengyang Zhou, Fengyuan Yu, Huabin Zhu, Binhui Yao, Tao Wang, Xiaolin Zheng, and Yanchao Tan. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22841-22850, 2024. 2, 6, 7, 8 +[23] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In Proceedings of the 39th International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 8 +[24] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6, 8 +[25] Yasar Abbas Ur Rehman, Yan Gao, Pedro Porto Buarque De Gusmão, Mina Alibeigi, Jiajun Shen, and Nicholas D Lane. L-dawa: Layer-wise divergence aware weight aggregation in federated self-supervised visual representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16464-16473, 2023. 6, 7 +[26] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22 (3):400 - 407, 1951. 11 +[27] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007. 1, 6 +[28] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7, 8 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1 +[30] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3, 6 +[31] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, 2020. 1, 2, 3, 6, 7, 11 +[32] Yihao Xue, Eric Gan, Jiayi Ni, Siddharth Joshi, and Baharan Mirzasoleiman. Investigating the benefits of projection head for representation learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 8 +[33] Fengda Zhang, Kun Kuang, Long Chen, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Fei Wu, Yueting Zhuang, et al. Federated unsupervised representation learning. Frontiers of Information Technology & Electronic Engineering, 24(8):1181-1193, 2023. 8 +[34] Yifei Zhang, Hao Zhu, Yankai Chen, Zixing Song, Piotr Koniusz, and Irwin King. Mitigating the popularity bias of + +graph collaborative filtering: A dimensional collapse perspective. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 6 +[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 8 +[36] Weiming Zhuang, Xin Gan, Yonggang Wen, Shuai Zhang, and Shuai Yi. Collaborative unsupervised visual representation learning from decentralized data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4912-4921, 2021. 2, 6, 8 +[37] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In International Conference on Learning Representations, 2022. 2, 6, 8 \ No newline at end of file diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/images.zip b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..288afe3b08d0f2ce2d57c96af99f8520d251db73 --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3643344c0633561305e2a13231a04c19a410b33db3500736562ee8714cda79d +size 450358 diff --git a/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/layout.json b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..77d5b4e20745e9b3581ee3717d1a144708285439 --- /dev/null +++ b/2025/Soft Separation and Distillation_ Toward Global Uniformity in Federated Unsupervised Learning/layout.json @@ -0,0 +1,9115 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 102, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 102, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 113, + 102, + 500, + 140 + ], + "type": "text", + "content": "Soft Separation and Distillation: Toward Global Uniformity in Federated Unsupervised Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "text", + "content": "Hung-Chieh Fang" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "text", + "content": " Hsuan-Tien Lin" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "text", + "content": " Irwin King" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "text", + "content": " Yifei Zhang" + }, + { + "bbox": [ + 132, + 160, + 480, + 176 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "spans": [ + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "type": "text", + "content": "National Taiwan University " + }, + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 189, + 470, + 205 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 256, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 558 + ], + "type": "text", + "content": "Federated Unsupervised Learning (FUL) aims to learn expressive representations in federated and self-supervised settings. The quality of representations learned in FUL is usually determined by uniformity, a measure of how uniformly representations are distributed in the embedding space. However, existing solutions perform well in achieving intra-client (local) uniformity for local models while failing to achieve inter-client (global) uniformity after aggregation due to non-IID data distributions and the decentralized nature of FUL. To address this issue, we propose Soft Separation and Distillation (SSD), a novel approach that preserves inter-client uniformity by encouraging client representations to spread toward different directions. This design reduces interference during client model aggregation, thereby improving global uniformity while preserving local representation expressiveness. We further enhance this effect by introducing a projector distillation module to address the discrepancy between loss optimization and representation quality. We evaluate SSD in both cross-silo and cross-device federated settings, demonstrating consistent improvements in representation quality and task performance across various training scenarios. Our results highlight the importance of inter-client uniformity in FUL and establish SSD as an effective solution to this challenge. Project page: https://ssd-uniformity.github.io/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 296, + 685 + ], + "type": "text", + "content": "Deep learning has achieved remarkable success across a wide spectrum of applications, from computer vision and natural language processing to speech recognition and reinforcement learning [1, 11, 18, 29]. This progress can be largely attributed to the availability of massive labeled datasets, particularly since the emergence of ImageNet [5]," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 337, + 256, + 512, + 451 + ], + "blocks": [ + { + "bbox": [ + 337, + 256, + 512, + 451 + ], + "lines": [ + { + "bbox": [ + 337, + 256, + 512, + 451 + ], + "spans": [ + { + "bbox": [ + 337, + 256, + 512, + 451 + ], + "type": "image", + "image_path": "9de7bcfe6166e3e1f81d9057d7b69aedb3259c202278684a2c16a5c780909ccb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Comparison with state-of-the-art methods. Our SSD achieves the highest performance (marked in brackets) and the best representation quality, measured by uniformity [31] and effective rank [27], among existing FUL approaches." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "which has enabled the training of increasingly powerful neural networks. However, the practical deployment of deep learning in real-world scenarios faces two critical challenges. First, data in real-world applications is often non-Independent and Identically Distributed (non-IID) and cannot be freely shared due to privacy concerns, regulatory restrictions, or proprietary constraints. This data is typically generated across diverse sources, including user devices, healthcare institutions, and industrial settings. The distributed nature and sharing limitations of such data make centralized training approaches infeasible, leading to the development of Federated Learning (FL) [15, 20, 24] as a privacy-preserving distributed learning paradigm. Second, a vast proportion of available data remains unlabeled, necessitating effective methods for unsupervised representation learning [2, 3, 8, 12]. These methods aim to learn meaningful feature representations without relying on explicit la" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "type": "text", + "content": "* Project was partially completed during the author's visit to the CUHK" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2971" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "bels, enabling models to capture underlying data structures and semantic relationships." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 294, + 263 + ], + "type": "text", + "content": "Federated Unsupervised Learning (FUL) emerges at the intersection of these two challenges, aiming to learn expressive representations in settings where data is both distributed and unlabeled. In FUL, the quality of learned representations is critically determined by two properties: alignment and uniformity [31]. Alignment measures how close similar data points are positioned in the representation space, relating to the model's ability to group semantically similar objects. Uniformity quantifies how evenly representations are distributed across the unit hypersphere, essentially measuring the entropy of the representation distribution. High uniformity prevents representation collapse and ensures that the learned features effectively utilize the available embedding dimensions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 294, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 395 + ], + "type": "text", + "content": "However, existing FUL methods face a significant challenge that remains inadequately addressed. While current approaches successfully achieve good intra-client (local) uniformity for representations within each client, they struggle to maintain inter-client (global) uniformity after model aggregation. This limitation stems from two key factors: (1) the non-IID distribution of data across clients, which naturally leads to divergent updates, and (2) the decentralized nature of FL, where the server lacks direct access to raw data, preventing the application of explicit uniformity constraints across clients." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 294, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 294, + 528 + ], + "type": "text", + "content": "Most existing approaches have focused primarily on two aspects of this challenge. One line of research builds upon the framework established by FedProx [21], introducing proximal terms to constrain local updates within a bounded neighborhood of the global model. For instance, FedU [36] and FedEMA [37] dynamically adjust aggregation weights based on inter-client model divergence, while FedX [10] incorporates a global relational loss to align pairwise sample relationships across clients. These methods aim to maintain global model consistency but do not explicitly address representation uniformity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 294, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 294, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 294, + 659 + ], + "type": "text", + "content": "Another research direction tackles the issue of dimensional collapse [14] caused by lower uniformity in local representations. FedDecorr [28] demonstrates that embeddings of local clients are often less uniformly distributed and mitigates this issue by decorrelating local features. Similarly, FedU2 [22] regularizes local updates to approximate a spherical Gaussian distribution, encouraging isotropic feature spaces. While these approaches enhance local uniformity, the improvements at the client level do not inherently translate to better global uniformity during model aggregation. This leads to a crucial question:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 669, + 291, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 669, + 291, + 692 + ], + "spans": [ + { + "bbox": [ + 60, + 669, + 291, + 692 + ], + "type": "text", + "content": "How can we effectively improve inter-client uniformity in Federated Unsupervised Learning?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "To address this challenge, we propose Soft Separation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "and Distillation (SSD), a novel method that enhances inter-client uniformity without compromising local representation expressiveness. Our approach employs a dimension-scaled regularization strategy that softly separates each client's feature space, encouraging client representations to spread toward different directions. As illustrated in Figure 2, this technique reduces interference during client model aggregation, thereby improving global uniformity without imposing rigid boundaries that could distort the underlying feature distributions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 348 + ], + "type": "text", + "content": "Furthermore, we empirically observe that the regularization effect of SSD may not always effectively transfer from loss optimization to the representation space due to the presence of a projector between the loss function and the encoder. While removing the projector might seem like a straightforward solution, doing so often degrades performance because the projector plays a crucial role in separating optimization objectives from feature representations [2, 9, 32]. To bridge this gap, we introduce a projector distillation module that minimizes the KL divergence between representations and embeddings, effectively encouraging the encoder to internalize the learned structure while preserving the projector's role in loss optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 468 + ], + "type": "text", + "content": "We evaluate SSD across diverse federated learning scenarios, including cross-silo (few clients with large datasets) and cross-device (many clients with small datasets) settings. Our experiments span both in-distribution tasks and out-of-distribution (OOD) datasets to assess generalizability. Results demonstrate that SSD consistently outperforms existing methods in both downstream task performance (measured via linear probing and fine-tuning accuracy) and representation quality (quantified using effective rank [7] and uniformity [31] metrics)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 470, + 478, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 470, + 478, + 480 + ], + "spans": [ + { + "bbox": [ + 326, + 470, + 478, + 480 + ], + "type": "text", + "content": "Our main contributions are threefold:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 482, + 553, + 601 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 482, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 517 + ], + "type": "text", + "content": "- We identify and formalize the challenge of inter-client uniformity in FUL, establishing it as a critical direction for decentralized unsupervised representation learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 518, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 553, + 566 + ], + "type": "text", + "content": "- We propose Soft Separation and Distillation (SSD), a simple yet effective framework that addresses inter-client uniformity without additional communication overhead or compromising privacy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 566, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 553, + 601 + ], + "type": "text", + "content": "- We conduct extensive experiments across varied FL settings and tasks, confirming SSD's superiority over baseline methods in both performance and robustness." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 616, + 399, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 399, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 399, + 628 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Federated unsupervised learning (FUL) is a learning paradigm in which multiple clients collaboratively train a model without sharing new raw data, and where the local datasets contain unlabeled data. The goal is to learn a global representation that generalizes across all participating clients. Consider a federated learning system with " + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2972" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 74, + 232, + 201 + ], + "blocks": [ + { + "bbox": [ + 107, + 74, + 232, + 201 + ], + "lines": [ + { + "bbox": [ + 107, + 74, + 232, + 201 + ], + "spans": [ + { + "bbox": [ + 107, + 74, + 232, + 201 + ], + "type": "image", + "image_path": "8263cb1e601bf03ab9cf5331c720d3c6a1a01f86d271957f17eba7846c89dc31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Illustration of Intra-Client and Inter-Client Uniformity with Soft Separation. Intra-client uniformity ensures that representations within each client are well-distributed, while inter-client uniformity promotes global representation consistency across clients. Our proposed Soft Separation method encourages each client's representations to spread in distinct directions, mitigating interference during model aggregation and improving global uniformity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 74, + 502, + 201 + ], + "blocks": [ + { + "bbox": [ + 244, + 74, + 502, + 201 + ], + "lines": [ + { + "bbox": [ + 244, + 74, + 502, + 201 + ], + "spans": [ + { + "bbox": [ + 244, + 74, + 502, + 201 + ], + "type": "image", + "image_path": "2fcd06304674a14763cc8f20b16021d5ae6548320739da7837754e1ae66ba792.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "text", + "content": "local clients and a central server. Each client " + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "text", + "content": " has access to an unlabeled local dataset " + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "text", + "content": ", modeled as samples from a client-specific data distribution " + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "inline_equation", + "content": "p_k(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 278, + 295, + 315 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 320, + 294, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 320, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 123, + 320, + 294, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {k} = \\left\\{\\mathbf {x} _ {i} ^ {k} \\right\\} _ {i = 1} ^ {n _ {k}} \\sim p _ {k} (\\mathbf {x}), \\tag {1}", + "image_path": "d77026293ee0c2c7dffed82354df2c93f924d1f6ff008c6d5c433e08b6b29654.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "inline_equation", + "content": "n_k" + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "content": " denotes the number of samples in " + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 341, + 295, + 365 + ], + "type": "text", + "content": ". The global objective in FUL can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 371, + 294, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 371, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 108, + 371, + 294, + 403 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\sum_ {k = 1} ^ {K} P _ {k} \\mathbb {E} _ {\\mathbf {x} \\sim p _ {k} (\\mathbf {x})} \\left[ \\mathcal {L} _ {k} (\\theta ; \\mathbf {x}) \\right], \\tag {2}", + "image_path": "4555b7e71c0acc8b96bc2fa43a4ffa75b4c4ed4ecbedbd3875b65cf0001b47ea.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "inline_equation", + "content": "P_{k}" + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": " is the probability of drawing a data from client " + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k(\\cdot)" + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": " denotes the local unsupervised loss of client " + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 438, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 486 + ], + "type": "text", + "content": "Self-Supervised Representation Learning aims to learn meaningful representations that capture the underlying structure of data. A typical framework consists of the following components:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 487, + 295, + 604 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": "- Encoder " + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": ": Maps (augmented) input data " + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} \\sim \\mathcal{T}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": " into a representation " + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = f(\\tilde{\\mathbf{x}}) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": " is the hidden dimension. It is common to use a deep neural network (e.g., ResNet50) as the encoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "text", + "content": "- **Projector " + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "text", + "content": ": Transforms the representation " + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "text", + "content": " into an embedding " + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g(\\mathbf{h}) \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 545, + 295, + 604 + ], + "type": "text", + "content": ". The training loss function is usually applied on this embedding space, and the projector is often a small network such as a multilayer perceptron." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 607, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 691 + ], + "type": "text", + "content": "A fundamental principle in self-supervised representation learning is that similar samples should have similar representations. In other words, representations should be invariant to minor variations in input data. This is typically achieved by aligning the representations of two augmented versions of the same image (referred to as a positive pair) using the following alignment loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 94, + 696, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 696, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 94, + 696, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x}), \\tilde {\\mathbf {x}}, \\tilde {\\mathbf {x}} ^ {+} \\sim \\mathcal {T} (\\mathbf {x})} | | \\mathbf {z} - \\mathbf {z} ^ {+} | | _ {2} ^ {2}, \\tag {3}", + "image_path": "b22b4683dfb7c17dd4dcb29dd3481668e7402ffa46322746c60c81c6ae2b49c0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}},\\tilde{\\mathbf{x}}^{+}\\sim \\mathcal{T}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "text", + "content": " denotes two independent augmentations of the same sample " + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g(f(\\tilde{\\mathbf{x}})),\\mathbf{z}^{+} = g(f(\\tilde{\\mathbf{x}}^{+}))" + }, + { + "bbox": [ + 313, + 278, + 553, + 315 + ], + "type": "text", + "content": " are the corresponding embeddings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 315, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 554, + 399 + ], + "type": "text", + "content": "Although aligning positive pairs encourages similarity, it can lead to a degenerate collapse where all representations converge to a single point. To circumvent this, Wang and Isola [31] introduces a uniformity objective to ensure that features are evenly dispersed on the unit hypersphere. This uniformity is often measured using the log of the average Gaussian potential [4]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 408, + 553, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 408, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 347, + 408, + 553, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\stackrel {\\mathrm {i . i . d}} {\\sim} p (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} _ {i} - \\mathbf {z} _ {j} | | _ {2} ^ {2}} \\right] \\tag {4}", + "image_path": "283edfa6badaa929c0216b66e500f4a30afdc1fd94e014082f4fbdd0e10ca4c4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z})" + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "content": " is the distribution of embeddings obtained by mapping data samples through the encoder " + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "content": " and projector " + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 554, + 475 + ], + "type": "text", + "content": " is the temperature hyperparameter." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 397, + 500 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "content": "In this section, we introduce Soft Separation and Distillation (SSD), a novel approach designed to enhance representation uniformity in federated learning (FL). We first analyze the limitations of uniformity in non-IID FL settings, then propose dimension-scaled regularization to softly separate client features, and finally introduce projector distillation to effectively transfer improved embedding uniformity to the representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Limited Inter-Client Uniformity of FL. Uniformity is a critical metric in representation learning [6, 30, 31], where higher uniformity indicates better preservation of information in the learned representations. In centralized training, jointly optimizing alignment loss " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and uniformity loss " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{uniform}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " yields high-quality representations. However, the distributed nature of FL introduces a fundamental challenge to uniformity optimization. Specifically, the uniformity metric can be decomposed into intra-client and inter" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2973" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 72, + 550, + 178 + ], + "blocks": [ + { + "bbox": [ + 72, + 72, + 550, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 72, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 550, + 178 + ], + "type": "image", + "image_path": "e6d799e2047d6309e53442249fc68e3e14073f9757fd68eccb1928973d710f04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": "Figure 3. Overview of our Training Pipeline. " + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": " The server initially assigns distinct weight vectors to each client. " + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": " Clients perform local training by optimizing their respective loss functions on their own data. " + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": " The server aggregates updated parameters from all clients. Steps " + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 54, + 189, + 555, + 224 + ], + "type": "text", + "content": " are iteratively repeated until convergence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 135, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 135, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 135, + 246 + ], + "type": "text", + "content": "client components:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 252, + 296, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 252, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 80, + 252, + 296, + 356 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {u n i f o r m}} = - \\log \\left(\\underbrace {\\sum_ {k = 1} ^ {K} \\mathbb {E} _ {\\mathbf {z} , \\mathbf {z} ^ {\\prime} \\sim p _ {k} (\\mathbf {z})} \\left[ e ^ {- t \\left| \\mathbf {z} - \\mathbf {z} ^ {\\prime} \\right| / 2} \\right]} _ {\\text {i n t r a - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k}}\\right) \\tag {5} \\\\ \\left. + \\underbrace {\\sum_ {i \\neq j} \\mathbb {E} _ {\\mathbf {z} \\sim p _ {i} (\\mathbf {z}) , \\mathbf {z} ^ {\\prime} \\sim p _ {j} (\\mathbf {z})} \\left[ e ^ {- t | | \\mathbf {z} - \\mathbf {z} ^ {\\prime} | | _ {2} ^ {2}} \\right]} _ {\\text {i n t e r - c l i e n t} \\mathcal {L} _ {\\text {u n i f o r m}} ^ {(i, j)}}\\right), \\\\ \\end{array}", + "image_path": "07c6e0856acd3ebba0929f42349a8c669c53bb706b5cb5ca9eaeb9ebf718cae0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "text", + "content": "where the first term represents intra-client uniformity " + }, + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{uniform}}^k" + }, + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "text", + "content": " and the second term represents inter-client uniformity " + }, + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{uniform}}^{(i,j)}" + }, + { + "bbox": [ + 55, + 361, + 295, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "text", + "content": "In federated learning, each client " + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "text", + "content": " optimizes its local loss function using samples drawn from its own local distribution " + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "text", + "content": ", which means that only the intra-client term in the above equation is explicitly optimized. In a homogeneous setting, when client distributions are similar (i.e., " + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "inline_equation", + "content": "p_i \\approx p_j" + }, + { + "bbox": [ + 54, + 400, + 296, + 579 + ], + "type": "text", + "content": "), optimizing intra-client uniformity naturally promotes good inter-client uniformity. However, in non-IID settings where client distributions differ significantly, optimizing only intra-client uniformity does not guarantee good inter-client uniformity, potentially limiting the quality of the globally aggregated model. The core challenge in enhancing inter-client uniformity lies in the federated learning constraint that the server has no access to raw client data or embeddings, making it impossible to directly impose a loss function that operates across different clients." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 676 + ], + "type": "text", + "content": "Dimension-Scaled Regularization (DSR). To address the inter-client uniformity challenge, we propose Dimension-Scaled Regularization (DSR), which improves global uniformity by encouraging client embeddings to spread in different directions of the representation space. This approach effectively increases the separation between clients without enforcing rigid boundaries that might distort the intrinsic data structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "For each client " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we define a dimension-scaling vector " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_k \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is the dimensionality of the embedding space. This vector applies selective scaling to specific di" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 234, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 258 + ], + "type": "text", + "content": "mensions, with some dimensions scaled by a factor " + }, + { + "bbox": [ + 313, + 234, + 553, + 258 + ], + "type": "inline_equation", + "content": "\\alpha \\neq 1" + }, + { + "bbox": [ + 313, + 234, + 553, + 258 + ], + "type": "text", + "content": " in a client-specific manner:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 383, + 263, + 553, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 263, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 383, + 263, + 553, + 297 + ], + "type": "interline_equation", + "content": "d _ {i, k} = \\left\\{ \\begin{array}{l l} \\alpha , & \\text {i f} i \\in \\mathcal {S} _ {k} \\\\ 1, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {6}", + "image_path": "07492799c0ac4b828ca2270f94dea80c974fd9931bf66ba0a06027173de0cea1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": " is a set of dimensions uniquely assigned to client " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": ", ensuring that the scaled dimensions are nonoverlapping across clients (i.e., " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "S_{i} \\cap S_{j} = \\emptyset" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": "). The size of each set " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": " is approximately " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\lfloor d / K \\rfloor" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 300, + 555, + 360 + ], + "type": "text", + "content": " is the number of clients." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "content": "We then regularize each client's embeddings by encouraging them to move toward their dimension-scaled versions through the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 403, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 403, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 339, + 403, + 553, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D S R}} ^ {k} = \\mathbb {E} _ {\\mathbf {z} \\sim p _ {k} (\\mathbf {z})} \\left[ \\| \\mathbf {z} - \\operatorname {s t o p g r a d} \\left(\\mathbf {z} \\odot \\mathbf {d} _ {k}\\right) \\| _ {2} ^ {2} \\right], \\tag {7}", + "image_path": "1ff05b6e9a6755856264d872ea8d70a37143068553ee4257839dfcfb4dad1c53.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "text", + "content": " represents element-wise multiplication, and stopgrad " + }, + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 423, + 554, + 470 + ], + "type": "text", + "content": " prevents gradient flow through the scaled target, ensuring that we pull the original embedding toward the scaled version rather than vice versa." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": "To understand why DSR enhances inter-client uniformity, we can analyze its effect on the representation distribution. When a vector " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " is scaled along specific dimensions by a factor " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\alpha > 1" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " and then normalized to the unit hypersphere, the resulting vector shifts toward those scaled dimensions. By assigning different scaling dimensions to each client, DSR effectively encourages client representations to occupy different regions of the unit hypersphere, as illustrated in Figure 5. Mathematically, if clients " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " have scaling vectors " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_j" + }, + { + "bbox": [ + 313, + 472, + 555, + 615 + ], + "type": "text", + "content": " with non-overlapping scaled dimensions, their optimized embeddings will tend to have a smaller dot product:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 620, + 553, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 620, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 339, + 620, + 553, + 636 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathbf {z} _ {i} \\sim p _ {i} (\\mathbf {z}), \\mathbf {z} _ {j} \\sim p _ {j} (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right] < \\mathbb {E} _ {\\mathbf {z} _ {i}, \\mathbf {z} _ {j} \\sim p (\\mathbf {z})} \\left[ \\mathbf {z} _ {i} ^ {\\top} \\mathbf {z} _ {j} \\right], \\tag {8}", + "image_path": "c79d04eaac18ecb149ea4e4cdc33303dd5b68c828b0500f39ae86ce290c7e6cd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z})" + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": " represents the distribution without DSR. This reduced dot product corresponds to increased angular separation, directly contributing to improved uniformity across the global distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Unlike a hard separation approach that would restrict each client to entirely separate subspaces (which could" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2974" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 220, + 178 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 220, + 178 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 220, + 178 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 220, + 178 + ], + "type": "image", + "image_path": "293ad20cd014bd41b353af42752f8e3d426a213b03895cd37947c77c2dbdd256.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": "Figure 4. Representation norm, embedding norm, and DSR loss during training. Adding DSR primarily affects the embedding space, as reflected in the embedding norm, while leaving the representation norm unchanged. In contrast, incorporating PD directly influences the representation norm, and when combined with DSR, it leads to a more pronounced decrease in representation norm. Additionally, PD enhances the effect of DSR, resulting in a lower " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DSR}}" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 385, + 178 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 385, + 178 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 385, + 178 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 385, + 178 + ], + "type": "image", + "image_path": "4b7786b85c457eaf192d008aa99161c863902d0e279ff1f63ef93bdf2674a26a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 552, + 178 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 552, + 178 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 552, + 178 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 552, + 178 + ], + "type": "image", + "image_path": "73b86710d64cc8088f62a4242cbbb70349561247d702a8957aefe969378797be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 255, + 171, + 357 + ], + "blocks": [ + { + "bbox": [ + 80, + 255, + 171, + 357 + ], + "lines": [ + { + "bbox": [ + 80, + 255, + 171, + 357 + ], + "spans": [ + { + "bbox": [ + 80, + 255, + 171, + 357 + ], + "type": "image", + "image_path": "9a23779c534c4f356fc90616c05d3f9f40a437194fad3466f0309737cf1a7538.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 296, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 425 + ], + "type": "text", + "content": "Figure 5. Illustration of the effect of DSR. The left figure shows the original data, where arrows indicate the transformation direction after applying DSR. The right figure presents the data after transformation. DSR increases separation and enhances inter-client uniformity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 255, + 271, + 357 + ], + "blocks": [ + { + "bbox": [ + 176, + 255, + 271, + 357 + ], + "lines": [ + { + "bbox": [ + 176, + 255, + 271, + 357 + ], + "spans": [ + { + "bbox": [ + 176, + 255, + 271, + 357 + ], + "type": "image", + "image_path": "d4c462cabe02ba7b71971ac33fa5cf9fea5f3d464edbfb9f1c06e5ccf907ea1e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 442, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 296, + 514 + ], + "type": "text", + "content": "severely constrain representation capacity), our soft separation allows clients to share most dimensions while gently pushing them toward different directions. This balanced approach preserves the flexibility needed for effective representation learning while significantly improving inter-client uniformity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "text", + "content": "An extreme example of fully separating embeddings would be assigning each client a dedicated subspace (e.g., if " + }, + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "inline_equation", + "content": "d = 500" + }, + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 55, + 517, + 296, + 624 + ], + "type": "text", + "content": ", each client could occupy 50 distinct dimensions). Although this configuration maximizes inter-client uniformity, it may disrupt the intrinsic data structure and miss out on the collaborative benefits of FL. In contrast, our soft DSR loss encourages some degree of separation without completely isolating the clients' embeddings. We analyze this tradeoff further in Section 4.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "Projector Distillation (PD). While DSR effectively enhances uniformity at the embedding level, we observe that this improvement may not fully transfer to the representation level. This discrepancy occurs because the projector " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " placed between the encoder " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " and the loss function can absorb much of the optimization effect, as shown in Figure 4. Although completely removing the projector" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 305 + ], + "type": "text", + "content": "might seem like a direct solution, prior work [2, 9, 32] has demonstrated that the projector plays a crucial role in separating optimization objectives from feature representations, thereby preventing overfitting to specific self-supervised tasks and improving downstream performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": "To bridge this gap, we introduce Projector Distillation (PD), which explicitly aligns the encoder's representations with the projector's embeddings:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 349, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 349, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 353, + 349, + 553, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} ^ {k} = \\mathbb {E} _ {x \\sim p _ {k} (\\mathbf {x})} \\left[ D _ {\\mathrm {K L}} \\left(\\sigma (\\mathbf {h}) \\| \\sigma (\\mathbf {z})\\right) \\right], \\tag {9}", + "image_path": "323d47fa91a91c661693b267a05f67e451731f115cb7ec2001e3be811fcd4055.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{h} = f(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "content": " is the representation from the encoder, " + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g(\\mathbf{h})" + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "content": " is the embedding from the projector, " + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "content": " denotes the softmax function, and " + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 371, + 553, + 418 + ], + "type": "text", + "content": " is the Kullback-Leibler divergence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 419, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 503 + ], + "type": "text", + "content": "This distillation mechanism encourages the encoder to internalize the beneficial structure learned in the embedding space, effectively transferring the improved uniformity from embeddings to representations. By minimizing the KL divergence between these distributions, we ensure that the uniformity improvements achieved through DSR are reflected in the representations used for downstream tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 479, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 479, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 479, + 523 + ], + "type": "text", + "content": "3.1. Overview of Training Pipeline" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 528, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 624 + ], + "type": "text", + "content": "The complete SSD algorithm is outlined in Algorithm 1, with our key design components highlighted in red. The server initially assigns unique weight vectors to all clients for dimension-scaled regularization. During local training, each client optimizes a combination of the standard alignment and uniformity losses for self-supervised learning, augmented with our DSR and PD terms. The local training objective for each client " + }, + { + "bbox": [ + 313, + 528, + 554, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 528, + 554, + 624 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 643, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 643, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 336, + 643, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {k} = \\mathcal {L} _ {\\text {a l i g n}} ^ {k} + \\beta \\mathcal {L} _ {\\text {u n i f o r m}} ^ {k} + \\gamma \\mathcal {L} _ {\\text {D S R}} ^ {k} + \\delta \\mathcal {L} _ {\\text {d i s t i l l}} ^ {k}, \\tag {10}", + "image_path": "76a146aecfa4d8ab2398dbe0406894a3798fa5d39251ddb59e05ec4664719e05.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\beta, \\gamma" + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 666, + 553, + 689 + ], + "type": "text", + "content": " are hyperparameters that balance the contribution of each loss term." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "After local training, clients upload their updated models to the server, which aggregates them using standard" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2975" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 55, + 73, + 294, + 510 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "Algorithm 1: Soft Separation and Distillation (SSD) \nInput: communication rounds " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": ", local epochs " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": ", number of clients " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": ", dataset " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\cup_{k \\in [K]} \\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": ", augmentation function " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " \nOutput: Final global encoder " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f_\\theta^T" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " \n1 Server: \n2 Initialize global encoder " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f_\\theta^0" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " and projector " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "g_\\theta^0" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n3 Assign weight vector " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_k\\}_{k \\in [K]}" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " for all clients; \n4 for " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "T - 1" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " do \n5 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "S_t \\gets" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " (randomly select a set of clients); \n6 for each client " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "k \\in S_t" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " in parallel do \n7 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f_\\theta^{(t,k)}, g_\\theta^{(t,k)} \\gets \\text{Client}(f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n8 end \n9 // Model aggregation; \n10 for each parameter " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " do \n11 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\theta^{t+1} \\gets \\sum_{k \\in S_t} w_k \\theta^{(t,k)}" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " \n12 end \n13 end \n14 return " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f_\\theta^T" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n15 Client " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "(f_\\theta^t, g_\\theta^t, \\mathbf{d}_k)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": ": \n16 // Initialize local models; \n17 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f \\gets f_\\theta^t, g \\gets g_\\theta^t" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n18 for each local epoch " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "e = 0" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "E - 1" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " do \n19 for each batch " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^{n_B} \\in \\mathcal{D}_k" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " do \n20 // Data augmentation; \n21 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i, \\tilde{\\mathbf{x}}_i^+ = \\mathcal{T}(\\mathbf{x}_i), \\mathcal{T}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n22 // Get representations and embeddings; \n23 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i, \\mathbf{h}_i^+ = f(\\tilde{\\mathbf{x}}_i), f(\\tilde{\\mathbf{x}}_i^+)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n24 " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i, \\mathbf{z}_i^+ = g(\\mathbf{h}_i), g(\\mathbf{h}_i^+)" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": "; \n25 // Optimize model using combined loss; \n26 Update " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^k" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " (Eq (10)); \n27 end \n28 end \n29 return " + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "inline_equation", + "content": "f, g" + }, + { + "bbox": [ + 55, + 73, + 294, + 510 + ], + "type": "text", + "content": " // Model upload; \n30 return" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "text", + "content": "federated averaging [24]: " + }, + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\theta^{t + 1} = \\sum_{k\\in S_t}w_k\\theta^{(t,k)}" + }, + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "inline_equation", + "content": "w_{k} = \\frac{n_{k}}{\\sum_{i = 1}^{K}n_{i}}" + }, + { + "bbox": [ + 55, + 536, + 295, + 574 + ], + "type": "text", + "content": " represents the aggregation weight based on the client's data size." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 575, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 647 + ], + "type": "text", + "content": "Importantly, our method preserves privacy since it requires no sharing of raw data. The only additional communication is the initial weight vectors " + }, + { + "bbox": [ + 55, + 575, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 55, + 575, + 295, + 647 + ], + "type": "text", + "content": ", which are lightweight and contain no sensitive information. The computational overhead of SSD is minimal, making it highly practical for real-world federated learning deployments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 658, + 135, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 135, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 135, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our approach on CIFAR-10 and CIFAR-100, which contain 50K training images and 10K testing images, with 10 and 100 classes, respectively. Fol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "lowing prior work [22, 37], we simulate non-IID data across " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " clients using a Dirichlet prior " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{Dir}(\\alpha)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", where a smaller " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " indicates a higher degree of data heterogeneity. We conduct experiments in two federated learning settings: cross-silo " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(K = 10)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " with full participation and cross-device " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(K = 50)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " with a participation rate of 0.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate our approach based on downstream performance and representation quality. For downstream performance, we follow prior work [10, 22, 36], using linear probing and fine-tuning with " + }, + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 148, + 553, + 232 + ], + "type": "text", + "content": " of the data. For representation quality, we assess uniformity [6, 30, 31] as defined in Equation (4) and effective rank [7, 27, 34] (see the definition in Appendix A.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 235, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 331 + ], + "type": "text", + "content": "Baselines. We compare our method against three categories of baselines. (1) Adapting a centralized algorithm to the federated setting: FedAlignUniform [31]. (2) FUL methods that mitigate divergence from the global model: FedX [10] and L-DAWA [25]. (3) FUL methods addressing dimensional collapse in local clients: FedDecorr [28]. FedU2 [22] incorporates modules that tackle both divergence control and dimensional collapse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "content": "Implementation Details. We follow prior work [22] for image augmentations, training configurations, and model architecture, using ResNet-18 [11] as the encoder and a two-layer linear projector. The hyperparameters " + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "content": " are set to 1.0, 1.0, and 0.1, respectively, and the scaling factor " + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 335, + 553, + 418 + ], + "type": "text", + "content": " is set to 10.0. Further details are provided in Appendix A.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 428, + 399, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 399, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 399, + 440 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 446, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 590 + ], + "type": "text", + "content": "Performance on different FL settings. Table 1 compares the performance of various methods in cross-silo and crossdevice federated learning. SSD consistently achieves the highest accuracy and excels in fine-tuning tasks. FedAlignUniform, FedDecorr, and FedU2 are competitive but generally lag behind SSD, while FedX and L-DAWA—aimed at ensuring global model consistency—perform the worst across all settings. Although FedDecorr and FedU2 enhance intra-client uniformity (with FedU2 adding a consistency module), they still trail SSD. By improving inter-client uniformity, SSD underscores the importance of addressing inter-client variations in federated learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Representation quality. Figure 1 compares different methods based on effective rank [27] and uniformity [31], two key metrics for representation quality. Among state-of-the-art methods, FedX and L-DAWA, which emphasize global model consistency, achieve strong representation quality. In contrast, methods focusing on local uniformity, such as FedDecorr and FedU2, exhibit lower effective rank and uniformity, indicating that they fail to address global uniformity. Our method, SSD, surpasses all other federated approaches, achieving the highest effective rank and the best uniformity." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2976" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 553, + 206 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Results on Cross-Silo and Cross-Device Settings. Accuracy (%) of linear probing (LP), fine-tuning (FT) " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " labeled data on CIFAR10 and CIFAR100 (" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "). The highest score is highlighted in bold, and the second-highest score is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 206 + ], + "type": "table", + "html": "
CIFAR10CIFAR100
Cross-Silo (K=10)Cross-Device (K=50)Cross-Silo (K=10)Cross-Device (K=50)
LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%LPFT 1%FT 10%
FedAlignUniform [31]80.8469.9981.0071.2857.4173.7757.2528.9748.9943.0316.3736.64
FedX [10]78.466.7880.0171.0156.9173.2457.3427.4649.5043.0716.0435.46
L-DAWA [25]77.6765.9479.3467.6553.7571.2256.9027.0849.5742.5815.5434.82
FedDecorr [28]80.1369.0980.3371.4958.1973.9757.2529.3849.5344.7417.6736.68
FedU2 [22]81.0169.6281.0171.0957.1574.2157.4029.3949.6442.9016.0835.48
SSD81.3270.7481.6771.8357.7774.6157.3829.5749.8745.2117.7036.82
", + "image_path": "76ffd5e441574799aac1328380d9095a6e29b211f2923076f69efb2ec347cc66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 267, + 294, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 297, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 297, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 297, + 258 + ], + "type": "text", + "content": "Table 2. Generalization on OOD datasets. Accuracy " + }, + { + "bbox": [ + 55, + 224, + 297, + 258 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 224, + 297, + 258 + ], + "type": "text", + "content": " of linear probing, uniformity, and effective rank when trained on CIFAR-100 and TinyImageNet-200, and evaluated on CIFAR-10." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 267, + 294, + 331 + ], + "lines": [ + { + "bbox": [ + 58, + 267, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 58, + 267, + 294, + 331 + ], + "type": "table", + "html": "
CIRAR100 → CIFAR10TinyImageNet → CIFAR10
LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)LP- \\( \\mathcal{L}_{\\text{uniform}}(\\uparrow) \\)ERank (↑)
FedAlignUniform [31]77.663.6566.9979.863.7176.31
FedX [10]78.023.7384.8879.873.7693.48
L-DAWA [25]77.463.7184.7579.623.7493.41
FedDecorr [28]77.623.6674.2079.793.7284.93
FedU2 [22]77.743.6668.379.743.6975.97
SSD78.483.7386.9580.003.7798.45
", + "image_path": "c37d11505025b02e202e709a3ea7e8b736218ddb41660ffec745a71be36aee53.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 374 + ], + "type": "text", + "content": "This demonstrates that SSD effectively enhances representation quality, outperforming existing methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 460 + ], + "type": "text", + "content": "Generalization on OOD datasets. Evaluating generalization to out-of-distribution (OOD) datasets helps assess whether models can learn transferable representations that perform well beyond their training distribution. Table 2 presents the generalization performance of different methods, where models trained on CIFAR-100 and TinyImageNet-200 are evaluated on CIFAR-10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "type": "text", + "content": "Among state-of-the-art methods, FedX, which exhibits high representation quality but performs poorly in indistribution settings, achieves the best generalization in OOD scenarios. However, our method, SSD, consistently outperforms all baselines, achieving the highest accuracy in both settings. SSD also demonstrates the best uniformity [31] and effective rank [7], indicating superior feature representation quality. While FedAlignUniform, FedDecorr, L-DAWA and FedU2 perform competitively, they generally fall short of SSD, further reinforcing its effectiveness in handling distribution shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 118, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 118, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 118, + 613 + ], + "type": "text", + "content": "4.2. Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Ablation Study. We provide the ablation study in Table 3. It shows that adding projector distillation (PD) alone does not provide noticeable improvements. In contrast, adding dimension-scaled regularization (DSR) alone leads to slight improvements in LP and FT " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", but the uniformity enhancement is minimal. However, when combining both DSR and PD (SSD), the method effectively transfers optimization benefits to representation quality, leading to a" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 289, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 278 + ], + "type": "text", + "content": "Table 3. Ablation Study. PD alone has minimal impact on performance. Adding DSR alone provides a slight improvement in both performance and uniformity. Combining both DSR and PD leads to a significant boost in uniformity and achieves the best performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 289, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 289, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 553, + 350 + ], + "type": "table", + "html": "
LPFT 1%FT 10%-Luniform(↑)
FedAlignUniform [31]80.8469.9981.003.79
+PD80.7469.7880.713.80
+DSR81.0569.7781.153.81
+ DSR + PD (SSD)81.3270.7481.673.84
", + "image_path": "3048135ca9698ca1158e26a1a0c148a530940fca638a8ab62ab0ffa28922b3e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 394 + ], + "type": "text", + "content": "significant uniformity improvement and achieving the best overall performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": "Robustness of DSR. We evaluate our method from two perspectives: the scaled factor " + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": " and the selection of scaled dimensions for each client. The scaled factor " + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": " influences the overall discrepancy, while the selection of scaled dimensions affects the direction each client is guided toward. To assess robustness, each " + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": " is evaluated with three different selections of scaled dimensions. As shown in Figure 6a, our method maintains stable performance across different " + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": " values, introducing minimal variation while consistently outperforming FedAlignUniform. This demonstrates that our approach effectively balances alignment and uniformity without being overly sensitive to the choice of " + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 556, + 554 + ], + "type": "text", + "content": " or the specific dimension selection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "Soft vs. hard client separation. We examine the impact of client feature separation on alignment, uniformity, and overall performance. Our intuition is that increasing separation between clients enhances global uniformity. One way to enforce this separation is by restricting each client to its own subspace without sharing information with others. To illustrate the effects of such strict partitioning, we consider Hard Separation and Distillation (HSD) as a baseline. As shown in Figure 6b, while HSD achieves the highest uniformity, it does so at the cost of severely reduced alignment, ultimately leading to poor downstream performance. This performance drop can be attributed to the disruption of intrinsic feature structures. In contrast, our proposed SSD balances" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2977" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 83, + 170, + 192 + ], + "blocks": [ + { + "bbox": [ + 55, + 83, + 170, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 83, + 170, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 83, + 170, + 192 + ], + "type": "image", + "image_path": "2543d03136fed646838407ea320da77c9d65b100fc9fc8549463b4ff2ff7c37e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 196, + 138, + 205 + ], + "lines": [ + { + "bbox": [ + 90, + 196, + 138, + 205 + ], + "spans": [ + { + "bbox": [ + 90, + 196, + 138, + 205 + ], + "type": "text", + "content": "(a) Robustness" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "type": "text", + "content": "Figure 6. (a) Robustness of DSR across different scaled factors " + }, + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 215, + 295, + 270 + ], + "type": "text", + "content": " is tested with three different randomly selected scaled dimensions. (b) Soft vs. hard separation. HSD achieves the highest uniformity but at the cost of reduced alignment, resulting in poor overall performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 83, + 291, + 192 + ], + "blocks": [ + { + "bbox": [ + 178, + 83, + 291, + 192 + ], + "lines": [ + { + "bbox": [ + 178, + 83, + 291, + 192 + ], + "spans": [ + { + "bbox": [ + 178, + 83, + 291, + 192 + ], + "type": "image", + "image_path": "08d31bfc5fc52f049028f3a49080672d2c9485c1cda065c7337028517ca1696f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 196, + 261, + 205 + ], + "lines": [ + { + "bbox": [ + 214, + 196, + 261, + 205 + ], + "spans": [ + { + "bbox": [ + 214, + 196, + 261, + 205 + ], + "type": "text", + "content": "(b) Separation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 82, + 323, + 271, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 295, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 315 + ], + "type": "text", + "content": "Table 4. The Effect of the Projector. Without the projector, DSR significantly improves uniformity and performance. Yet, the overall results remain inferior to those achieved with the projector." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 323, + 271, + 381 + ], + "lines": [ + { + "bbox": [ + 82, + 323, + 271, + 381 + ], + "spans": [ + { + "bbox": [ + 82, + 323, + 271, + 381 + ], + "type": "table", + "html": "
ProjectileLP-Luniform(↑)
FedAlignUniformX73.163.72
+ DSRX76.14 (+2.98)3.77 (+0.05)
FedAlignUniform✓80.843.79
+ DSR✓81.05 (+0.21)3.81 (+0.02)
", + "image_path": "da8fa3307aac4428dcdfbc4819d36def8d842ad2a978770f02eb622a41f0eb0c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 435 + ], + "type": "text", + "content": "both alignment and uniformity, mitigating the drawbacks of hard separation while enhancing uniformity, resulting in improved overall performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 439, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 606 + ], + "type": "text", + "content": "Why not remove projector or apply loss on representations? A natural idea for addressing the limited transfer of embedding-level loss optimization (in terms of representation uniformity) is to remove the projector or apply the loss directly to the representations. However, prior work [2, 9, 32] has demonstrated that the projector plays a critical role in preventing the encoder from overfitting on the upstream task. Indeed, as shown in Table 4, removing the projector and adding DSR does substantially improve uniformity and downstream performance, but these results still lag behind those achieved when the projector is retained. Therefore, it is crucial to develop a method that effectively promotes uniformity while preserving the benefits offered by the projector." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 142, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 142, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 142, + 628 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 262, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 262, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 262, + 648 + ], + "type": "text", + "content": "5.1. Federated Learning with Non-IID Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Federated learning (FL) enables collaborative model training across multiple clients without centralizing data. Since client data distributions are typically non-IID, model convergence and performance are challenging. FedAvg [24], which aggregates local model updates using simple averag" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "ing, is the foundational framework but suffers under heterogeneous data and imbalanced client participation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 96, + 555, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 227 + ], + "type": "text", + "content": "To address these issues, various aggregation improvements have been proposed. FedProx [21] reduces client drift by adding a proximal term to limit local updates from straying from the global model. SCAFFOLD [16] corrects client drift using control variates to reduce variance in local updates. MOON [19] improves representation consistency with a contrastive loss. Other methods use globally shared data to improve generalization; FedShare [35] aligns client models using shared data, while FedDistill [13] distills knowledge from the global model to local models using shared data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 235, + 497, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 497, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 497, + 248 + ], + "type": "text", + "content": "5.2. Federated Unsupervised Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 553, + 300 + ], + "type": "text", + "content": "Federated Unsupervised Learning (FUL) combines federated learning with un-/self-supervised learning, primarily addressing (1) non-IID data distribution and (2) representation collapse." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 300, + 554, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 432 + ], + "type": "text", + "content": "To handle non-IID data, FedCA [33] uses a shared dictionary module for better aggregation but risks privacy leakage. FedU [36] reduces this risk by selectively uploading the online network's encoder and deciding predictor updates based on divergence. FedEMA [37] extends this with exponential moving average updates. FedX [10] adds an alignment/contrastive term with the global model, while Orchestra [23] preserves global structural consistency using global centroids. FedU2 [22] ensures balanced updates across clients for better alignment but does not guarantee global representation uniformity under non-IID settings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 432, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 528 + ], + "type": "text", + "content": "For representation collapse, FedDecorr [28] shows that local clients suffer from dimensional collapse, which propagates to the global model, and addresses it with a local decorrelation loss. FedU2 [22] encourages local representation uniformity by minimizing divergence with a spherical Gaussian. Our work extends this by showing that intra-client uniformity alone is insufficient inter-client uniformity must also be explicitly addressed." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "We introduce Soft Separation & Distillation (SSD), a framework designed to enhance representation quality by improving inter-client uniformity in federated learning. SSD consists of a dimension-scaled regularization term that softly separates client embeddings while preserving the intrinsic data structure, and a projector distillation term that transfers the optimization benefits of the projector to the encoder, thereby improving representation quality. SSD achieves state-of-the-art performance in both representation learning and downstream tasks across various training and FL settings. Our work highlights the importance of global representation quality in federated unsupervised learning, opening new directions for future research." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2978" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 259 + ], + "type": "text", + "content": "We thank the anonymous reviewers and members of CLLab for their constructive feedback. This work is partially supported by the National Science and Technology Council in Taiwan via NSTC 113-2634-F-002-008, 114-2221-E002-102-MY3, and the National Taiwan University Center for Data Intelligence via NTU-114L900901 and Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410072, RGC R1015-23). We thank to National Center for High-performance Computing (NCHC) of National Applied Research Laboratories (NARLabs) in Taiwan for providing computational and storage resources. H.-T. Lin is honored to be supported by the Leap Fellowship of the Foundation for the Advancement of Outstanding Scholarship in Taiwan since 2025." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 115, + 285 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 293, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 293, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 294, + 336 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 294, + 392 + ], + "type": "text", + "content": "[2] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 1, 2, 5, 8, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 395, + 294, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 294, + 438 + ], + "type": "text", + "content": "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 441, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 294, + 474 + ], + "type": "text", + "content": "[4] Henry Cohn and Abhinav Kumar. Universally optimal distribution of points on spheres. Journal of the American Mathematical Society, 20(1):99-148, 2007. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 294, + 520 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 294, + 564 + ], + "type": "text", + "content": "[6] Xianghong Fang, Jian Li, Qiang Sun, and Benyou Wang. Rethinking the uniformity metric in self-supervised learning. In The Twelfth International Conference on Learning Representations, 2024. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 621 + ], + "type": "text", + "content": "[7] Quentin Garrido, Randall Balestriero, Laurent Najman, and Yann Lecun. Rankme: Assessing the downstream performance of pretrained self-supervised representations by their rank. In International conference on machine learning, pages 10929-10974. PMLR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 689 + ], + "type": "text", + "content": "[8] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "text", + "content": "[9] Kartik Gupta, Thalaiyasingam Ajanthan, Anton van den Hengel, and Stephen Gould. Understanding and improving" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the role of projection head in self-supervised learning. arXiv preprint arXiv:2212.11491, 2022. 2, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 149 + ], + "type": "text", + "content": "[10] Sungwon Han, Sungwon Park, Fangzhao Wu, Sundong Kim, Chuhan Wu, Xing Xie, and Meeyoung Cha. Fedx: Unsupervised federated learning with cross knowledge distillation. In European Conference on Computer Vision, pages 691-707. Springer, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 248 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": "[13] Eunjeeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non- " + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{id}" + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": " private data. arXiv preprint arXiv:1811.11479, 2018. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 338 + ], + "type": "text", + "content": "[14] Li Jing, Pascal Vincent, Yann LeCun, and Yuandong Tian. Understanding dimensional collapse in contrastive self-supervised learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 403 + ], + "type": "text", + "content": "[15] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "text", + "content": "[16] Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In International conference on machine learning, pages 5132-5143. PMLR, 2020. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 535 + ], + "type": "text", + "content": "[17] Taehyeon Kim, Jaehoon Oh, Nak Yil Kim, Sangwook Cho, and Se-Young Yun. Comparing kullback-leibler divergence and mean squared error loss in knowledge distillation. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 2628–2635. International Joint Conferences on Artificial Intelligence Organization, 2021. Main Track. 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 536, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 536, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 536, + 553, + 579 + ], + "type": "text", + "content": "[18] Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. Journal of Machine Learning Research, 17(39):1-40, 2016. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 624 + ], + "type": "text", + "content": "[19] Qinbin Li, Bingsheng He, and Dawn Song. Model-contrastive federated learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10713–10722, 2021. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "text", + "content": "[20] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[21] Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429-450, 2020. 2, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2979" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 148 + ], + "type": "text", + "content": "[22] Xinting Liao, Weiming Liu, Chaochao Chen, Pengyang Zhou, Fengyuan Yu, Huabin Zhu, Binhui Yao, Tao Wang, Xiaolin Zheng, and Yanchao Tan. Rethinking the representation in federated unsupervised learning with non-iid data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22841-22850, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "text", + "content": "[23] Ekdeep Lubana, Chi Ian Tang, Fahim Kawsar, Robert Dick, and Akhil Mathur. Orchestra: Unsupervised federated learning via globally consistent clustering. In Proceedings of the 39th International Conference on Machine Learning, pages 14461-14484. PMLR, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "text", + "content": "[24] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "text", + "content": "[25] Yasar Abbas Ur Rehman, Yan Gao, Pedro Porto Buarque De Gusmão, Mina Alibeigi, Jiajun Shen, and Nicholas D Lane. L-dawa: Layer-wise divergence aware weight aggregation in federated self-supervised visual representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16464-16473, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[26] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Mathematical Statistics, 22 (3):400 - 407, 1951. 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 402 + ], + "type": "text", + "content": "[27] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In 2007 15th European signal processing conference, pages 606-610. IEEE, 2007. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[28] Yujun Shi, Jian Liang, Wenqing Zhang, Vincent Tan, and Song Bai. Towards understanding and mitigating dimensional collapse in heterogeneous federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "text", + "content": "[30] Feng Wang and Huaping Liu. Understanding the behaviour of contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2021. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[31] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, 2020. 1, 2, 3, 6, 7, 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[32] Yihao Xue, Eric Gan, Jiayi Ni, Siddharth Joshi, and Baharan Mirzasoleiman. Investigating the benefits of projection head for representation learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[33] Fengda Zhang, Kun Kuang, Long Chen, Zhaoyang You, Tao Shen, Jun Xiao, Yin Zhang, Chao Wu, Fei Wu, Yueting Zhuang, et al. Federated unsupervised representation learning. Frontiers of Information Technology & Electronic Engineering, 24(8):1181-1193, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[34] Yifei Zhang, Hao Zhu, Yankai Chen, Zixing Song, Piotr Koniusz, and Irwin King. Mitigating the popularity bias of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 239 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "graph collaborative filtering: A dimensional collapse perspective. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[35] Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. Federated learning with non-iid data. arXiv preprint arXiv:1806.00582, 2018. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[36] Weiming Zhuang, Xin Gan, Yonggang Wen, Shuai Zhang, and Shuai Yi. Collaborative unsupervised visual representation learning from decentralized data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4912-4921, 2021. 2, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[37] Weiming Zhuang, Yonggang Wen, and Shuai Zhang. Divergence-aware federated self-supervised learning. In International Conference on Learning Representations, 2022. 2, 6, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2980" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_content_list.json b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..26b027d7cf28323a62c3abef76e459c2753fc52f --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_content_list.json @@ -0,0 +1,1705 @@ +[ + { + "type": "text", + "text": "Sparfels: Fast Reconstruction from Sparse Unposed Imagery", + "text_level": 1, + "bbox": [ + 189, + 79, + 807, + 103 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shubhendu Jena*, Amine Ousafi*, Mae Younes, Adnane Boukhayma Inria, Univ. Rennes, CNRS, IRISA", + "bbox": [ + 225, + 130, + 772, + 165 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e4c4da61c26933d1d0060d5a82578777048f050ccd480691d2efac615dd4c02.jpg", + "image_caption": [ + "Figure 1. Examples of reconstruction meshes obtained within 3 minutes from sparse pose-free images via our method. We used here 6 (left) and 3 (right) input images respectively from scenes of datasets MVImgNet [97] and BlendedMVS [92]." + ], + "image_footnote": [], + "bbox": [ + 93, + 169, + 906, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 354, + 326, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a method for Sparse view reconstruction with surface element splatting that runs within 3 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate state-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view benchmarks based on established multi-view datasets. Code will be made available at https://shubhendujena.github.io/Sparfels-web/", + "bbox": [ + 88, + 387, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 756, + 220, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D reconstruction is a long standing problem at the intersection of computer vision, graphics and machine learning. It is core to a myriad of downstream applications in virtual and augmented reality, autonomous driving, medical imaging, film and animation, to name a few. Classical optimization based techniques such as structure from motion and", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multi-view stereo [70, 71], shape from visual cues such as shading or silhouette [28, 43], photometric stereo [83] etc., have been popular for decades in this regard. More recently, deep learning has been progressively disrupting 3D vision and top-down approaches to vision more broadly. Among the latest additions, large vision and foundation models (e.g. DUSt3R [79], MASt3R [44]) can offer robustness and fast feed-forward inference compared to classical photogrammetry [70, 71]. Given a few raw uncalibrated images of a scene, they are capable of inferring reasonable relative poses and a competitive coarse 3D structure. These predictions can lack precision and details, though, and cannot enable stand-alone novel view synthesis.", + "bbox": [ + 511, + 356, + 906, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Furthermore, state-of-the-art data-prior free scene specific learning methods for reconstruction can deliver highly detailed geometry and impressive novel view synthesis capabilities from dense calibrated image grids. While implicit and some grid shape representations (e.g. NeuS, VolSDF, HF-NeuS, Voxurf [78, 80, 85, 94]) trained via differentiable volumetric rendering (NeRF [55]) have been powerful tools for this task for a while, Gaussian Splatting [40] based alternatives emerged with comparable novel view and reconstruction [29, 100] quality, within a fraction of the training time, and with orders of magnitude higher interactive rendering frame rates. These methods typically fail however under uncalibrated and sparse training images. Contemporary literature attempts to hedge against the underconstrained nature of the sparse view problem through regularization with a multitude of deep geometric priors (monocular depth [18, 76, 104], normal [98], point cloud [67] etc.), or training new priors from scratch [50, 67, 87]. Such strategies can prove to be cumbersome, impractical or costly in terms of data, memory or compute.", + "bbox": [ + 511, + 553, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose to marry the best of both worlds efficiently — namely pre-existing large 3D vision feedforward models (MASt3R) and fast test-time optimization", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27476", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(2DGS) — with the aim of addressing the underexplored problem of 3D geometric reconstruction from sparse un-posed images within minutes (less than 3 on average in our experiments) at minimal cost.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to exploit a single pre-existing foundation model (MASt3R [44]) to the fullest to bootstrap our reconstruction, without needing to load or deploy any additional deep neural networks. Given a few sparse color images, MASt3R delivers a coarse point cloud and initial camera poses. We use these to initialize a bundle adjusting 2D Gaussian Splatting (2DGS) optimization via differentiable render-and-compare, where surface elements are instantiated at the point samples, oriented initially according to their normals. Differently from the closest competition [21, 22], we harness the MASt3R [44] produced image pair correspondences to guide camera optimization during 2DGS training. This is achieved via a splatted depth reprojection error loss. The resulting camera alignment improvement proves crucial for obtaining accurate shape representations.", + "bbox": [ + 89, + 154, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By analyzing the surface element splatting from a statistical moment perspective, rendered pixels can be seen as color expectation over fragments following the ray termination probability. Based on this, we derive an expression of splatted color variance along the ray, and propose a novel loss tasked with reducing it. This results in improved reconstruction metrics and qualitative sharper mesh details as witnessed by our ablation studies.", + "bbox": [ + 89, + 397, + 482, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The integration of these elements yields state-of-art performance in sparse pose-free 3D reconstruction, obtained in very short times on mid-range GPUs. We evaluate our method using established reconstruction and novel view datasets [1, 5, 42, 92, 97], showcasing both superior geometric reconstruction, camera pose estimation and novel view synthesis w.r.t. comparable state-of-the-art baselines.", + "bbox": [ + 89, + 521, + 482, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation of 3D shape reconstruction in the posed setting relies traditionally on mesh Chamfer distance to ground-truth, as used in multi-view reconstruction benchmarks like DTU [1]. This is a mesh pose and scale sensitive metric. In the unposed setting, reconstructions are obtained naturally up-to a rigid transformation. RANSAC powered alignments between predicted and ground-truth cameras have been shown to be acceptable for pre-evaluation reconstruction mesh alignment in the dense view setting [7], and good enough for camera pose evaluation for sparse sets [75]. However, we found these camera based alignments to be too inaccurate in the sparse setting to enable reliable mesh pre-alignments for Chamfer distance evaluation. Hence, inspired by the rigid-transform invariant evaluation in [79], we compare the reconstructed mesh depth and normal rasterizations to the groundtruth in screen space in our evaluation, and show considerable improvements compared to the competition in the standard DTU benchmark", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/60af467a4cb250858e54d2498e0e32ad7d04c0446ce61d2332cd7bc5c79e19ec.jpg", + "image_caption": [ + "Figure 2. 3D Foundation models meet differentiable perspective accurate Splating for fast reconstruction from pose-free sparse images. We bootstrap a bundle adjusting 2DGS with MAST3R [44] 3D point clouds and initial camera predictions. Besides standard 2DGS photometric reconstruction and geometric regularization losses, we leverage a camera reprojection error loss using MAST3R's correspondences and a novel splatted color variance reduction loss." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 906, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(Tab.1,2,3). Further qualitative comparison in Blended-MVS [92], MVImgNet [97], Mip-NeRF 360 [4] illustrate the quality of our reconstructions.", + "bbox": [ + 511, + 313, + 905, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 368, + 651, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NVS and 3D Reconstruction Novel view synthesis (NVS) aims to generate unseen views from input images [3, 54]. Neural Radiance Fields (NeRF) [55] revolutionized this field by representing 3D scenes as continuous implicit functions, though challenges remain in efficiency and quality [4-6, 66, 72, 73]. Inspired by the success of other explicit representations (e.g. meshes [33, 38, 77] and point clouds [2, 20]), recent advancements introduced primitive-based scene representations [13, 40, 90], notably 3D Gaussian Splitting (3DGS) [40], which uses anisotropic 3D Gaussians with differentiable splatting for efficient high-quality reconstruction. This has sparked extensions for surface reconstruction [29, 100, 101], neural representations [51], multi-resolution modeling [24, 99], and feedforward methods [10, 12, 23, 27, 89]. In parallel, implicit representations model 3D geometry as continuous functions [16, 37, 39, 57, 59]. Methods like IDR [93], NeuS [78], and VolSDF [94] enhance geometric accuracy but often suffer from slow rendering. In contrast, 2D Gaussian Splitting (2DGS) [29] offers real-time rendering while improving view consistency for surface reconstruction through planar Gaussian primitives.", + "bbox": [ + 511, + 393, + 906, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse-View NVs and 3D Reconstruction Traditional NeRFs and GS methods require carefully captured images and SfM preprocessing with COLMAP [69], which could limit practical applications. The few-shot setup has been addressed for both reconstruction from Point Cloud [11, 62, 64, 65] and images. Various approaches address the challenge of limited views through regularization techniques, including depth supervision [17, 46, 58, 76], foundation model supervision [32, 84, 88], specialized architectures [91, 96], while SparseCraft [95] uses learning-free", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27477", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MVS cues for SDF regularization. For Gaussian Splating, methods like FSGS [104] and SparseGS [86] incorporate external priors. Generalizable methods address sparse-view challenges by pretraining on large datasets, enabling feed-forward inference. Inputs include images (e.g. [10, 12, 35, 45, 49]) and 3D Points (e.g. [9, 30, 52, 60, 61, 63]). For surface reconstruction from color images, approaches like SparseNeuS [50], VolRecon [68], ReTR [47], UfoRecon [56], GeoTransfer [34], and Spurfies [67] leverage various mechanisms such as using image priors to form multiview cost volumes, using transformer architectures to enhance reconstruction from limited views, as well as using reconstruction priors trained on sparse point clouds.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-Free NVS and 3D Reconstruction Pose-free methods eliminate the need for accurate camera calibration. NeRFmm [82], BARF [48], SCNeRF [36], and GARF [15] jointly optimize camera parameters and scene representations. SPARF [75] and PoRF [7] address optimization with noisy poses, though often requiring lengthy training. Recent work like Nope-NeRF [8], Lu-NeRF [14], LocalRF [53], and CF-3DGS [25] leverage depth information but typically assume dense video sequences. InstantSplat [21, 22] addresses these limitations by using MASt3R [44] for initialization and jointly optimizing camera poses and 3D models. Our work builds on these foundations but focuses specifically on using 2D Gaussian Splatting for efficient 3D surface reconstruction from sparse, uncalibrated views. We initialize with MASt3R [44] and refine through joint optimization of Gaussian parameters and camera poses, incorporating a novel variance-based color loss for improved detail preservation. Unlike previous approaches requiring dense sequences, known intrinsic, or long optimization times, our method efficiently handles sparse, uncalibrated images while achieving high-quality reconstructions.", + "bbox": [ + 91, + 306, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 631, + 183, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a few unposed color images $\\{I^i\\}_{i = 1}^N$ , our goal is to recover a 3D triangle mesh $S$ and render novel views of the observed scene efficiently. We achieve this by fitting a bundle adjusting 2DGS [29] model to the images of the scene, bootstrapped by point map and correspondence feedforward predictions from a 3D foundation model (MASt3R [44]) and carefully designed regularizations. At test time, mesh $S$ can be extracted using the TSDF algorithm in Open3D [103] using the converged 2DGS model rendered depth, as recommended in [29]. We detail the workings of our method below.", + "bbox": [ + 89, + 656, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 832, + 232, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Gaussian Splatting Our method builds upon 2D Gaussian Splatting (2DGS) [29], which represents 3D scenes using planar disk primitives defined within local tan-", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gent planes. Unlike volumetric 3D Gaussians [40], 2DGS is a surface-based representation that improves multi-view consistency. Each 2D Gaussian is defined as:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {p}) = \\exp \\left(- \\frac {u (\\mathbf {p}) ^ {2} + v (\\mathbf {p}) ^ {2}}{2}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 148, + 906, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u(\\mathbf{p})$ and $v(\\mathbf{p})$ are local UV space coordinates of a pixel $\\mathbf{p}$ in the primitive's tangent plane. This ray-splat intersection can be expressed as a function of the camera parameters and splat attributes differentiably. Each primitive is parameterized by the center of the Gaussian in world coordinates $\\mathbf{x} \\in \\mathbb{R}^3$ , a quaternion representing the orientation of the primitive $\\mathbf{q} \\in \\mathbb{R}^4$ , scale factors for the two dimensions in the local UV tangent plane $\\mathbf{s} \\in \\mathbb{R}^2$ , opacity $\\alpha \\in \\mathbb{R}$ controlling the transparency of the primitive, and Spherical harmonic coefficients $c_{\\mathrm{sh}k} \\in \\mathbb{R}^3 | k = 1, \\dots, n$ encoding view-dependent appearance. Rotation $\\mathbf{q}$ defines a local coordinate frame $[\\mathbf{u}, \\mathbf{v}, \\mathbf{n}]$ , $\\mathbf{n}$ being the normal vector orthogonal to the primitive's plane, and $\\mathbf{u}$ and $\\mathbf{v}$ spanning the tangent plane.", + "bbox": [ + 511, + 191, + 906, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For rendering, 2DGS uses alpha-compositing for volumetric rendering similarly to 3DGS. Given a set of primitives along a ray sorted from front to back $G_{1}, G_{2}, \\ldots, G_{K}$ , the final color $C$ is computed as:", + "bbox": [ + 511, + 404, + 905, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 476, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}^{\\prime} = \\alpha_{i}G(\\mathbf{p})$ is the modified opacity of the $i$ -th primitive and $\\mathbf{c}_i$ is its color, computed from the spherical harmonic coefficients based on the viewing direction.", + "bbox": [ + 511, + 532, + 905, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MASt3R MASt3R [44] is a foundation model for 3D vision that extends DUSt3R [79] with more powerful feature extraction and matching capabilities. It tackles the multi-view stereo reconstruction task without requiring prior camera calibration or poses.", + "bbox": [ + 511, + 599, + 905, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pair of images $I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3}$ , MAS3R outputs two corresponding pointmaps $\\mathbf{P}_{1,1}, \\mathbf{P}_{2,1} \\in \\mathbb{R}^{W \\times H \\times 3}$ with associated confidence maps $\\mathbf{O}_{1,1}, \\mathbf{O}_{2,1} \\in \\mathbb{R}^{W \\times H}$ . Each pointmap assigns a 3D point to every pixel in the image. Importantly, both pointmaps are expressed in the same coordinate frame of the first image $I^1$ .", + "bbox": [ + 511, + 675, + 905, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of MASt3R consists of a Siamese ViT encoder that processes both images independently, Transformer decoders that exchange information via crossattention, and regression heads that output pointmaps and confidence maps. MASt3R is trained using a confidence-weighted regression loss:", + "bbox": [ + 511, + 767, + 905, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n f}} = \\sum_ {v \\in 1, 2} \\sum_ {i} \\mathbf {O} _ {v, 1} ^ {i} \\mathcal {L} _ {\\text {r e g}} (v, i) - \\alpha \\log \\mathbf {O} _ {v, 1} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27478", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{reg}}(v,i)$ measures the Euclidean distance between predicted and ground-truth 3D points for pixel $i$ , and $\\alpha$ is a regularization parameter.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides pointmaps, MASt3R also provides dense pixelwise correspondences between the input views. These correspondences establish reliable geometric constraints across different viewpoints, which we leverage in our optimization framework.", + "bbox": [ + 89, + 136, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Scene Initialization", + "text_level": 1, + "bbox": [ + 89, + 220, + 272, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Geometry Alignment Given a set of input images $I^1, I^2, \\ldots, I^N$ , we first construct a connectivity graph $\\mathcal{G}(\\mathcal{V}, \\mathcal{E})$ where each vertex $v \\in \\mathcal{V}$ represents an image, and each edge $e = (n, m) \\in \\mathcal{E}$ indicates a pair of images with shared visual content. For each edge in the graph, we run MASt3R to obtain pairwise pointmaps and correspondences. To align these pairwise predictions into a globally consistent reconstruction, we follow the optimization approach from MASt3R:", + "bbox": [ + 89, + 241, + 482, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\chi , P, \\sigma} \\sum_ {e \\in \\mathcal {E}} \\sum_ {v \\in e} \\sum_ {i = 1} ^ {H W} \\mathbf {O} _ {v, e} ^ {i} \\left| \\chi_ {i} ^ {v} - \\sigma_ {e} P _ {e} \\mathbf {P} _ {v, e} ^ {i} \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 385, + 480, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\chi_i^v$ is the globally aligned pointmap for view $v$ at pixel $i$ , $P_e \\in \\mathbb{R}^{3 \\times 4}$ the rigid transformation for edge $e$ , $\\sigma_e > 0$ is a scaling factor for edge $e$ , $\\mathbf{O}_{v,e}^i$ the confidence value for pixel $i$ in view $v$ for edge $e$ , and finally $\\mathbf{P}_{v,e}^i$ the original pointmap from MASt3R. This optimization produces globally aligned pointmaps $\\{\\chi^{v*}\\}_{v=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}$ and camera parameters (intrinsics and extrinsics for each view $v$ ). We stabilize the estimated focal lengths by averaging them across all training views: $\\bar{f} = \\frac{1}{N} \\sum_{v=1}^N f_i^*$ , assuming a single-camera setup.", + "bbox": [ + 89, + 435, + 483, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Primitive Initialization We leverage the merged globally aligned pointmap to initialize 2D Gaussian primitives with suitable orientation, which is crucial for accurately representing surface geometry. For each point $\\mathbf{p}_i$ in the globally aligned pointmap, we compute a local surface normal using Principal Component Analysis (PCA). First, we determine the $k$ -nearest neighbors $\\{\\mathbf{p}_j\\}_{j=1}^k$ and compute the covariance matrix $\\mathbf{C} = \\frac{1}{k}\\sum_{j=1}^k (\\mathbf{p}_j - \\bar{\\mathbf{p}})(\\mathbf{p}_j - \\bar{\\mathbf{p}})^T$ , where $\\bar{\\mathbf{p}}$ is the centroid of the neighborhood. We then perform eigendecomposition on $\\mathbf{C}$ to obtain the eigenvectors $\\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\}$ and corresponding eigenvalues $\\{\\lambda_1, \\lambda_2, \\lambda_3\\}$ , sorted such that $\\lambda_1 \\geq \\lambda_2 \\geq \\lambda_3$ . The eigenvector $\\mathbf{v}_3$ , associated with the smallest eigenvalue, is chosen as the normal direction, i.e., $\\mathbf{n} = \\mathbf{v}_3$ . To construct an orthonormal basis, we initialize an arbitrary vector $\\mathbf{a}$ orthogonal to $\\mathbf{n} = (n_x, n_y, n_z)$ , defined as $\\mathbf{a} = (-n_y, n_x, 0)^T$ . The tangent vector is then computed as $\\mathbf{u} = \\frac{\\mathbf{n} \\times \\mathbf{a}}{\\|\\mathbf{n} \\times \\mathbf{a}\\|}$ and the bitangent as $\\mathbf{v} = \\mathbf{n} \\times \\mathbf{u}$ . Finally, the local reference frame $[\\mathbf{u}, \\mathbf{v}, \\mathbf{n}]$ is used to define a rotation matrix, which can be converted to a quaternion $\\mathbf{q}_i$ for efficient representation. Next, we initialize each", + "bbox": [ + 89, + 592, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D Gaussian primitive with position $\\mathbf{x}_i$ directly from the pointmap, rotation $\\mathbf{q}_i$ , scale $\\mathbf{s}_i$ set proportionally to the local pointmap density, and opacity $\\alpha_i$ set to a default value (e.g. 0.8). Color for the lowest spherical harmonic band (DC component) is initialized from the original image colors projected onto the point, with higher-order coefficients set to zero. This initialization preserves the surface structure from the MASt3R reconstruction while enabling efficient optimization with 2D Gaussian primitives.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Joint Optimization", + "text_level": 1, + "bbox": [ + 511, + 234, + 694, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correspondence Loss Using the dense correspondences provided by MASt3R, and inspired by SPARF [75], we formulate a multi-view correspondence loss that enforces geometric consistency across different viewpoints:", + "bbox": [ + 511, + 257, + 903, + 318 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o r r}} = w _ {p _ {n}, p _ {m}} \\rho \\left(p _ {m} - \\pi \\left(P _ {m} ^ {- 1} P _ {n} \\pi^ {- 1} \\left(p _ {n}, d _ {n}\\right)\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 325, + 903, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(p_n, p_m)$ is a correspondence pair with $p_n \\in [0,1]^2$ a pixel in view $n$ and $p_m$ its corresponding pixel in view $m$ . $w_{p_n,p_m} \\in [0,1]$ is the confidence of the correspondence. $\\rho$ represents the Huber loss [31] function. $\\pi: \\mathbb{R}^3 \\to \\mathbb{R}^2$ is the projection operator mapping 3D points to pixel coordinates. $\\pi^{-1}: \\mathbb{R}^2 \\times \\mathbb{R} \\to \\mathbb{R}^3$ is the back-projection operator mapping pixel coordinates and depth to 3D points. $d_n$ is the 2DGS model splatted depth at pixel $p_n$ . $P_n$ and $P_m$ represent the camera parameters for views $n$ and $m$ . This loss enforces that corresponding pixels across views should project to the same 3D point when back-projected using their respective depths and camera parameters, and thus helps guide our bundle adjustment to a good minima.", + "bbox": [ + 511, + 354, + 906, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Photometric Loss Following 3DGS and 2DGS, we use a combination of $\\mathcal{L}_1$ image reconstruction loss and structural similarity (SSIM) to optimize the rendering of our system:", + "bbox": [ + 511, + 570, + 905, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p h o t o}} = (1 - \\lambda) \\cdot \\mathcal {L} _ {1} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {S S I M}} + \\mathcal {L} _ {\\mathrm {r e g}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 626, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notice that 2DGS comes also with geometric regularization term $\\mathcal{L}_{\\mathrm{reg}}$ binding pixel depth normals to splatted normals, and reducing depth distortion, as detailed in [29].", + "bbox": [ + 511, + 652, + 903, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Variance Regularization Loss A key contribution of our method is the introduction of a splatted color variance regularization loss. The fundamental insight is that volume rendering can be interpreted as computing the expected value of colors sampled along a ray [74]:", + "bbox": [ + 511, + 703, + 905, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 787, + 903, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c(\\mathbf{t})$ represents the color of a point $\\mathbf{t}$ along the ray, and the expectation is taken with respect to $p(t) = \\sigma(t)T(t)$ with $\\sigma(t)$ being the density at $t$ and $T(t)$ representing the transmittance function.", + "bbox": [ + 511, + 839, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27479", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hypothesize that stabilizing and robustifying our training would lead to more robust and improved reconstructions. In order to robustify the training, we would like to control how the rendered color deviates from the ground truth color when the geometry is perturbed, i.e. given divergence measure $\\mathcal{D}$ , we want to minimize the following worst-case loss along the ray:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\sup _ {q: D (q, p) < \\rho} | \\mathbb {E} _ {\\mathbf {t} \\sim q (\\mathbf {t})} [ c (\\mathbf {t}) ] - C _ {g t} |, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 203, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ controls the strength of the perturbation. By leveraging theorem 2 in [19], this loss can be upper bounded by:", + "bbox": [ + 89, + 234, + 483, + 266 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} + \\eta \\sqrt {\\operatorname {V a r} _ {\\boldsymbol {t} \\sim p (\\boldsymbol {t})} [ c (\\boldsymbol {t}) ]}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 273, + 483, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consequently, we consider minimizing the variance of the color along the ray under the distribution $p(t)$ that can be expressed as follows:", + "bbox": [ + 89, + 301, + 483, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {v a r}} = \\mathbb {V} a r _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] (10) \\\\ = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ (c (\\mathbf {t}) - C) ^ {2} ] = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ^ {2} ] - C ^ {2}. (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 354, + 482, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For efficient implementation, we compute the first term using a modified CUDA kernel of the 2DGS splatting algorithm, where we render the square of colors in addition to the rendering of the colors. This allows for efficient forward and backward passes during optimization. Intuitively, reducing rendered color variance reduces uncertainty in the rendering, which leads to sharper and more multi-view coherent reconstructions. This reflects positively on the qualitative (Fig 5) and quantitative (Tab 4 and Fig 6) performance of our reconstructions. We note that in practice, loss $\\mathcal{L}_{\\mathrm{var}}$ is averaged across pixels for a given view.", + "bbox": [ + 89, + 412, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective Our total optimization objective combines the three loss terms:", + "bbox": [ + 89, + 583, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {p h o t o}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\text {c o r r}} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 619, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{photo}}$ , $\\lambda_{\\mathrm{corr}}$ , and $\\lambda_{\\mathrm{var}}$ are weighting factors that balance the contribution of each term. We optimize all parameters simultaneously using the Adam [41] optimizer with a learning rate scheduling following [29]. Unlike multi-stage approaches [75], our single-stage optimization allows for faster convergence while maintaining high reconstruction quality. The optimization updates both camera extrinsics $\\{P\\}$ per input view and 2D Gaussian primitive parameters $\\{\\alpha, \\mathbf{c}, \\mathbf{q}, \\mathbf{s}\\}$ according to the gradient of the total loss.", + "bbox": [ + 89, + 643, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 784, + 223, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our approach across 3D reconstruction, novel view synthesis, and camera pose estimation. We outline the experimental setup, including datasets, baseline methods, and implementation details, followed by quantitative and qualitative comparisons. Ablation studies further analyze the impact of key components.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Configuration", + "text_level": 1, + "bbox": [ + 511, + 90, + 870, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is tested on benchmark datasets [1, 5, 42, 92, 97]. Further implementation details and additional qualitative results are provided in the supplementary material.", + "bbox": [ + 511, + 113, + 903, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-Shot 3D Reconstruction: We evaluate sparse 3D reconstruction on DTU [1], BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]. For 3-view reconstruction on DTU, we follow SparseNeuS [50], evaluating 15 scenes with two sets of 3-view inputs, reporting the mean per scene. We also provide qualitative comparisons for 3, 6, and 12-view settings on selected scenes from BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5].", + "bbox": [ + 511, + 159, + 905, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-Shot Novel View Synthesis: We conduct novel view synthesis (NVS) on Tanks and Temples [42] (8 scenes), MVImgNet [97] (7 scenes), and MipNeRF360 [5] (9 scenes), following InstantSplat [21, 22]. Results are reported under sparse input settings (3, 6, and 12 views).", + "bbox": [ + 511, + 281, + 905, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera Pose Estimation: Camera poses are evaluated on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5], following InstantSplat [21, 22]. Additionally, for few-shot reconstruction, we assess camera poses on DTU [1] using the same 3-view evaluation protocol as in the reconstruction task.", + "bbox": [ + 511, + 357, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For 3D reconstruction quality, we employ the Absolute Relative Error (Rel) and Normal Consistency (NC). Rel measures the accuracy of depth estimation following DUSt3R [79], while NC evaluates the geometric fidelity by measuring the alignment between predicted and ground truth surface normals. We apply these metrics to our final meshes. We compute ground truth normals from the gradients of the ground truth DTU depth maps.", + "bbox": [ + 511, + 459, + 905, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For novel view synthesis, we use standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [81], and Learned Perceptual Image Patch Similarity (LPIPS) [102]. We note that for this evaluation we follow InstantSplat[21, 22]: for novel test views without pre-computed camera poses, we perform an additional camera optimization step, which optimizes the test cameras while keeping the 2D Gaussian primitives fixed using the $\\mathcal{L}_1$ photometric reconstruction loss.", + "bbox": [ + 511, + 597, + 905, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For camera pose estimation, we report the Absolute Trajectory Error (ATE) as defined in [8, 21], using COLMAP-computed poses from dense views as ground-truth references.", + "bbox": [ + 511, + 733, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method against several state-of-the-art approaches:", + "bbox": [ + 511, + 805, + 903, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- NVS specific methods: Pose-free methods NoPe-NeRF [8], CF-3DGS [25], NeRFmm [82], SPARF [75] and InstantSplat [21, 22] (in both S and XL configurations). We also include comparisons with 3DGS [40] and", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27480", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a6a443dd589d0469fd40f93ab0ba1d98180304669f8cc90e02d554da7227f9a.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on DTU dataset from 3 input images." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FSGS [104], which use pre-computed camera parameters from Colmap [69].", + "bbox": [ + 102, + 325, + 482, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Reconstruction-specific methods: MASt3R [44], Colmap [69], UfoRecon [56], CasMVSNet [26], Spurfies [67], SpaRP [87], SparseCraft [95], InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22]. Of these, similar to our approach, InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22] further refine cameras.", + "bbox": [ + 89, + 357, + 482, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our framework is implemented in PyTorch. We set the optimization iterations to 1k for DTU dataset reconstruction experiments and between 2k-4k for other reconstruction and novel view synthesis experiments. For test-time optimization, camera parameters are jointly optimized for 1k iterations (while keeping the learned Gaussian parameters fixed). We initialize our pipeline with MASt3R [44] configured at a resolution of 512 pixels and use Open3D [103] to compute normals for the point cloud, which are used to initialize rotation parameters of the 2D Gaussians. All experiments are conducted on a single NVIDIA A6000 GPU. We set $\\lambda_{\\mathrm{photo}} = 1.0$ and $\\lambda_{\\mathrm{corr}} = 5 \\times 10^{-5}$ . The variance loss weight $\\lambda_{\\mathrm{var}}$ follows a cosine annealing schedule: $\\lambda_{\\mathrm{var}} = 0.5\\left(1 + \\cos \\left(\\frac{\\pi \\cdot t}{T}\\right)\\right)$ , where $t$ denotes the current training step and $T$ the total number of steps. This schedule gradually decreases $\\lambda_{\\mathrm{var}}$ from 1.0 to 0.0, ensuring an adaptive regularization throughout training.", + "bbox": [ + 89, + 473, + 483, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. 3D Reconstruction Evaluation", + "text_level": 1, + "bbox": [ + 89, + 756, + 359, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For few-shot reconstruction, we follow the evaluation protocol from prior work [50, 56, 95], cleaning the generated meshes with masks from training views. We evaluate the Absolute Relative Error (Rel) and Normal Consistency (NC) between predicted depths normals from the extracted meshes and the ground truth depths normals from the DTU dataset [1].", + "bbox": [ + 89, + 779, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 and 2 present the quantitative results on the DTU", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6305944af912b06f0aff59928c7261dad0296f90d6452780b12f6d36b2e9ea7e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on BlendedMVS dataset from 3 input images." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset for Rel and NC metrics, respectively. Fig 3, 4 and 7 show qualitative reconstruction comparisons, with additional qualitative comparisons with SpaRP [87] included in the supplementary material. Additionally, the supplementary material contains extended qualitative results, in the form of video comparisons demonstrating our reconstructions for several datasets. Our method achieves state-of-the-art performance, outperforming all competing approaches across the majority of scenes.", + "bbox": [ + 511, + 412, + 906, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several key observations can be made from these results. Pretrained generalizable methods (UfoRecon [56] and CasMVSNet [26]) generalize poorly when faced with noisy initial camera parameters from MASt3R [44], resulting in significantly worse performance. Despite being designed for sparse reconstruction, SparseCraft [95] struggles with the noisy camera poses and point clouds from MASt3R [44]. Spurfies [67], which employs an SDF prior trained on ShapeNet, performs particularly poorly with noisy camera poses, resulting in worse mesh extractions. InstantSplat [21, 22] variants (InstSplatGOF and InstSplat2DGS) perform better than other competing approaches, but still produce overly smooth meshes lacking fine details.", + "bbox": [ + 511, + 550, + 908, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method significantly outperforms these approaches due to several key advantages, including the variance regularization loss (Sec. 3.3) promoting sharper and more detailed reconstructions Fig 5, The correspondence loss (Sec. 3.3) and joint optimization framework effectively refine the initial noisy camera poses from MAsst3R [44]. Our method's ability to preserve fine geometric details is also evidenced by the superior Normal Consistency scores.", + "bbox": [ + 511, + 748, + 908, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents camera pose estimation results measured by ATE. Our approach achieves significant improve", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27481", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/83eeb0af8fa39f4896cc229207a561397ffb54314d1d90040ac181ccc2a489ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]2.4920.4812.025.398.866.34.719.017.7010.335.862.505.04.564.957.34
UFORecon [56]35.5557.1538.8135.4750.0244.9434.7053.5641.6747.9340.2239.3439.7533.3449.0442.77
CasMVSNet [26]9.7638.0037.0317.3019.3924.106.5729.5711.8414.9212.388.383.939.8713.4417.10
Spurfies [67]2.9652.1558.4220.2514.6619.244.9712.447.5010.647.366.2638.6125.8444.1921.70
InstSplatGOF [21]9.5423.1413.7415.6213.3012.3311.1912.2114.6915.6113.6112.3015.3112.9012.0513.84
Sparsecraft [95]2.9610.517.162.928.179.443.836.197.616.444.756.435.367.658.146.50
InstantSplat2DGS [22]2.4221.749.083.086.773.603.635.245.827.124.482.783.623.113.435.73
Ours2.4212.3137.93.156.053.5263.484.985.296.224.392.413.463.323.384.82
", + "bbox": [ + 93, + 89, + 903, + 183 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d7c7fa5a9cca8cd5ef106b8e59efa127987aeaa16aee610d4f71445ad439d7d4.jpg", + "table_caption": [ + "Table 1. 3-view reconstruction results on DTU dataset The relative error (rel↓) is reported for 15 scans. Best results are bold." + ], + "table_footnote": [], + "table_body": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122\\( \\uparrow \\) Mean
MASt3R [44]0.8880.6950.8440.8740.8370.8390.8510.7540.8240.7450.8710.8750.8840.8310.8460.830
UFORecon [56]0.3620.2780.5060.4690.2770.3720.3830.3610.3620.3380.4240.3660.3320.4020.3390.371
CasMVSNet [26]0.7180.5020.6190.7560.6930.6830.7400.5640.7040.6810.7850.7950.8590.7780.7550.709
Spurfies [67]0.8630.350.2970.7110.7420.6680.7710.6710.7770.6830.780.7710.5090.5720.4330.640
InstSplatGOF [21]0.7740.5810.6890.6960.7460.7500.7100.6580.6820.6530.6950.7240.7100.6630.6500.692
Sparsecraft [95]0.8410.6790.8270.8640.7280.7140.8340.7560.7460.7650.8370.6290.8010.7610.7350.768
InstSplat2DGS [22]0.8820.6160.860.8990.8340.8650.8450.7830.8100.7730.8820.8490.8950.8270.8550.832
Ours0.9040.750.8770.8960.8660.8870.8670.7990.8450.7890.8880.8740.9020.850.8650.857
", + "bbox": [ + 93, + 199, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. 3-view reconstruction results on DTU dataset Normal consistency (NC↑) is reported for 15 scans. Best results are bold.", + "bbox": [ + 109, + 297, + 883, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ments in camera pose estimation, with a $13.1\\%$ reduction in error compared to the best-performing baseline (InstSplatGOF [21]). This demonstrates the effectiveness of our correspondence loss in refining the initial noisy camera poses from MAST3R [44].", + "bbox": [ + 89, + 325, + 482, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Novel View Synthesis Evaluation", + "text_level": 1, + "bbox": [ + 89, + 411, + 377, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our method's capability for novel view synthesis on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5] datasets. Quantitative and qualitative result comparisons on all datasets are included in the supplementary material.", + "bbox": [ + 89, + 433, + 483, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 518, + 253, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to analyze the contribution of each component in our method, including an experiment to study the impact of varying the number of input views.", + "bbox": [ + 89, + 541, + 483, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component Ablation These experiments were performed using 3 input unposed views on 5 testing scenes from the DTU dataset using two sets of views, following the protocols established in [50, 56]. Tab 4 shows the impact of incorporating different components into our pipeline. Fig 5 shows a qualitative ablation comparison. Camera optimization using bundle adjustment improves all metrics. Incorporating the correspondence loss further enhances this alignment, resulting in improved reconstructions. Splat orientation initialization with pointmap normals improves results further. Notice that our variance loss contributes to both the final metrics and the qualitative aspect of our results. The reconstructions with variance loss exhibit significantly sharper details and better preservation of fine structures, validating the quantitative improvements observed in our metrics.", + "bbox": [ + 89, + 592, + 482, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Views Ablation We progressively increase the number of views and perform pose-free reconstruction with our method using scan37 from DTU. We evaluate by computing the mean metrics across both test view sets. As", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11ed57c145d7816deeda225b27f0acd20dad063b7773d2750cabe523f0300f74.jpg", + "image_caption": [ + "Figure 5. Qualitative ablation of the component of our method, showing also the effect of variance loss. Left blue box: w/o $\\mathcal{L}_{\\mathrm{var}}$ . Right red box: w $\\mathcal{L}_{\\mathrm{var}}$ . (3 input images)." + ], + "image_footnote": [], + "bbox": [ + 516, + 323, + 906, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shown in Figure 6, our method benefits favorably from increased views. The benefits of all components of our method are also maintained, if not increased. This includes our variance loss.", + "bbox": [ + 511, + 678, + 906, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Running time", + "text_level": 1, + "bbox": [ + 511, + 744, + 653, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our reconstruction experiments, we observed that the mean training time on DTU [1] for 1000 iterations (similar to InstantSplat-XL [21, 22]) was roughly 146s on a single Nvidia A6000 GPU. For a fair comparison, we also trained InstantSplat-XL [21, 22] on the same GPU and observed that the mean training time was around 128s. Hence, our training time is in the same range as InstantSplat [21, 22] while offering significantly higher reconstruction and novel-view quality.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27482", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/03861876aed40843d323b9101d0af3401ca5b3477a27d1b8d0e56a93aefa1ccc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]0.02020.06890.00940.02030.01710.03920.02910.02940.02710.03260.01670.02100.02890.01920.02420.0269
InstSplatGOF [21]0.01800.05360.00890.01360.00490.02730.02630.02390.01800.03230.01280.02380.02490.01750.01490.0214
InstSplat2DGS [22]0.01920.08950.00900.01540.01230.03380.02940.02720.02420.03220.01660.02100.02740.01630.01750.0261
Ours0.00630.04500.00690.01240.01750.02940.01980.02830.01460.02730.00820.01790.02420.01120.01010.0186
", + "bbox": [ + 93, + 89, + 903, + 147 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cd08e9161e9119c2c9dd436b19e0cd3b2230c9d4e020076148ddb4a0464703a1.jpg", + "table_caption": [ + "Table 3. Results on DTU dataset Results include ATE $\\downarrow$ metrics for 15 scans. Best results are bold." + ], + "table_footnote": [], + "table_body": "
Methodsrel ↓NC ↑ATE ↓
2DGS7.930.7850.0327
BA-2DGS7.790.7880.0284
BA-2DGS+Lcorr7.410.8040.0256
BA-2DGS+Lcorr+Norm. init.7.030.8290.0237
BA-2DGS+Lcorr+Norm. init.+Lvar6.790.8340.0238
", + "bbox": [ + 125, + 172, + 450, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation on DTU dataset, adding one component at a time. Results include rel and NC metrics. Best results are bold.", + "bbox": [ + 89, + 255, + 482, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47137a74a20088ba55944074e10e22f89f5ccd9f60fee40815dd7b750f92d2a6.jpg", + "image_caption": [ + "Figure 6. Effect of increasing the number of input views on different metrics." + ], + "image_footnote": [], + "bbox": [ + 96, + 311, + 282, + 445 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f13245cd30b3c8b573694d5338628d5c0380c6c9098ed2208f176a1322be650b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 311, + 480, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 89, + 513, + 362, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the promising results, several limitations remain, which present opportunities for future research. As foundation models keep on improving, better initializations could be available in the future. Furthermore, while our approach handles a variety of scenes across different datasets, performance may be limited for highly complex geometry, very thin structures, or semi-transparent objects (observed for the glass bottle in Fig 7 room scene in MipNeRF360 [5]), which are common challenges for surface based methods especially in the sparse setting. Additionally, our method assumes a static scene under consistent lighting. Extending it to more dynamic scenarios is one of the challenges we are looking forward to exploring.", + "bbox": [ + 88, + 541, + 482, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 753, + 209, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We addressed the underexplored problem of shape reconstruction from sparse uncalibrated multi-view images. Our method leverages foundation models for initial camera and geometry estimation, followed by joint optimization of 2D Gaussian primitives and camera parameters. The introduction of a variance regularization loss significantly improves the detail and sharpness of reconstructions, while our correspondence matching based framework enhances", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4d4b23f1482015766d47d7a0e2f26babd5824cb44868f579f50cc30f2f0f0d3.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on MVImgNet and MipNeRF360 datasets from 6 and 12 input images, respectively." + ], + "image_footnote": [], + "bbox": [ + 516, + 172, + 903, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "camera pose accuracy. Extensive evaluations on multiple datasets demonstrate that our approach consistently outperforms state-of-the-art methods in reconstruction quality, novel view synthesis, and camera pose estimation.", + "bbox": [ + 511, + 729, + 906, + 790 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27483", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 2, 5, 6, 7", + "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2", + "[3] Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1034-1040. IEEE, 1997. 2", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460-5469, 2022. 2", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 7, 8", + "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2", + "[7] Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, and Philip Torr. Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449, 2023. 2, 3", + "[8] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 3, 5", + "[9] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3", + "[10] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3", + "[11] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting" + ], + "bbox": [ + 99, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 3", + "[13] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4182-4194, 2023. 2", + "[14] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312-18321, 2023. 3", + "[15] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In European Conference on Computer Vision, pages 264-280. Springer, 2022. 3", + "[16] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2", + "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 2", + "[18] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1", + "[19] John Duchi, Peter Glynn, and Hongseok Namkoong. Statistics of robust optimization: A generalized empirical likelihood approach, 2018. 5", + "[20] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2", + "[21] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2(3):4, 2024. 2, 3, 5, 6, 7, 8", + "[22] Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. arXiv preprint arXiv:2403.20309, 2024. 2, 3, 5, 6, 7, 8", + "[23] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. arXiv preprint arXiv:2410.18956, 2024. 2", + "[24] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024. 2" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27484", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatt-ing. arXiv preprint arXiv:2312.07504, 2023. 3, 5", + "[26] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 6, 7", + "[27] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jisang Han, Jiaolong Yang, Chong Luo, and Seungryong Kim. Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv preprint arXiv:2410.22128, 2024. 2", + "[28] Berthold KP Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. 1970. 1", + "[29] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5", + "[30] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 3", + "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 4", + "[32] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2", + "[33] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Neural mesh-based graphics. In European Conference on Computer Vision, pages 739-757. Springer, 2022. 2", + "[34] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Geotransfer: Generalizable few-shot multi-view reconstruction via transfer learning. arXiv preprint arXiv:2408.14724, 2024. 3", + "[35] Shubhendu Jena, Shishir Reddy Vutukur, and Adnane Boukhayma. Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting. arXiv, 2025. 3", + "[36] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5846-5854, 2021. 3", + "[37] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2", + "[38] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2", + "[39] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 5", + "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[42] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2, 5, 7", + "[43] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 1", + "[44] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 6, 7, 8", + "[45] Qian Li, Franck Moulton, and Adnane Boukhayma. Learning generalizable light field networks from few images. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3", + "[46] Qian Li, Franck Multon, and Adnane Boukhayma. Regularizing neural radiance fields from sparse rgb-d inputs. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2320-2324. IEEE, 2023. 2", + "[47] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[48] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 3", + "[49] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Fast generalizable gaussian splattering reconstruction from multi-view stereo. arXiv preprint arXiv:2405.12218, 2024.3", + "[50] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 3, 5, 6, 7", + "[51] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2", + "[52] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6326-6337, 2022. 3", + "[53] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27485", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3", + "[54] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2", + "[55] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[56] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 3, 6, 7", + "[57] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2", + "[58] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 2", + "[59] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2", + "[60] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 3", + "[61] Amine Ouasfi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. 3DV, 2024. 3", + "[62] Amine Ouasfi and Adnane Boukhayma. Few-shot unsupervised implicit neural shape representation learning with spatial adversaries. arXiv preprint arXiv:2408.15114, 2024. 2", + "[63] Amine Ouasfi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[64] Amine Ouasfi and Adnane Boukhayma. Unsupervised occupancy learning from sparse point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21729-21739, 2024. 2", + "[65] Amine Ouasfi, Shubhendu Jena, Eric Marchand, and Adnane Boukhayma. Toward robust neural reconstruction from sparse point sets. arXiv preprint arXiv:2412.16361, 2024. 2" + ], + "bbox": [ + 99, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Eric Penner and Li Zhang. Soft 3D Reconstruction for View Synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 2", + "[67] Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Spurfies: Sparse surface reconstruction using local geometry priors. arXiv preprint arXiv:2408.16544, 2024. 1, 3, 6, 7", + "[68] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 3", + "[69] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 6", + "[70] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1", + "[71] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1", + "[72] Steven M Seitz and Charles R Dyer. Photorealistic Scene Reconstruction by Voxel Coloring, 2002. US Patent 6,363,170. 2", + "[73] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2", + "[74] Andrea Tagliaasacchi and Ben Mildenhall. Volume rendering digest (for nerf), 2022. 4", + "[75] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2, 3, 4, 5", + "[76] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. arXiv preprint arXiv:2303.16196, 2023. 1, 2", + "[77] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2", + "[78] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2", + "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 5" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27486", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35:1966-1978, 2022. 1", + "[81] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[82] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. 2021. 3, 5", + "[83] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 1", + "[84] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2", + "[85] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1", + "[86] Haolin Xiong, Sairishek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 3", + "[87] Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. In European Conference on Computer Vision, pages 143-163. Springer, 2024. 1, 6", + "[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 2", + "[89] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2", + "[90] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2", + "[91] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2", + "[92] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 1, 2, 5" + ], + "bbox": [ + 99, + 90, + 485, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2", + "[94] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2", + "[95] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6, 7", + "[96] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2", + "[97] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 1, 2, 5, 7", + "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 1", + "[99] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2", + "[100] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2", + "[101] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2", + "[102] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[103] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 3, 6", + "[104] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 866 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27487", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_model.json b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b638ed347d949d0bfa20323a469a7b1a2bc70421 --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_model.json @@ -0,0 +1,2798 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.08, + 0.808, + 0.104 + ], + "angle": 0, + "content": "Sparfels: Fast Reconstruction from Sparse Unposed Imagery" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.131, + 0.773, + 0.166 + ], + "angle": 0, + "content": "Shubhendu Jena*, Amine Ousafi*, Mae Younes, Adnane Boukhayma Inria, Univ. Rennes, CNRS, IRISA" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.17, + 0.907, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 1. Examples of reconstruction meshes obtained within 3 minutes from sparse pose-free images via our method. We used here 6 (left) and 3 (right) input images respectively from scenes of datasets MVImgNet [97] and BlendedMVS [92]." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.355, + 0.327, + 0.371 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.486, + 0.736 + ], + "angle": 0, + "content": "We present a method for Sparse view reconstruction with surface element splatting that runs within 3 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate state-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view benchmarks based on established multi-view datasets. Code will be made available at https://shubhendujena.github.io/Sparfels-web/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.222, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.485, + 0.875 + ], + "angle": 0, + "content": "3D reconstruction is a long standing problem at the intersection of computer vision, graphics and machine learning. It is core to a myriad of downstream applications in virtual and augmented reality, autonomous driving, medical imaging, film and animation, to name a few. Classical optimization based techniques such as structure from motion and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.553 + ], + "angle": 0, + "content": "multi-view stereo [70, 71], shape from visual cues such as shading or silhouette [28, 43], photometric stereo [83] etc., have been popular for decades in this regard. More recently, deep learning has been progressively disrupting 3D vision and top-down approaches to vision more broadly. Among the latest additions, large vision and foundation models (e.g. DUSt3R [79], MASt3R [44]) can offer robustness and fast feed-forward inference compared to classical photogrammetry [70, 71]. Given a few raw uncalibrated images of a scene, they are capable of inferring reasonable relative poses and a competitive coarse 3D structure. These predictions can lack precision and details, though, and cannot enable stand-alone novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Furthermore, state-of-the-art data-prior free scene specific learning methods for reconstruction can deliver highly detailed geometry and impressive novel view synthesis capabilities from dense calibrated image grids. While implicit and some grid shape representations (e.g. NeuS, VolSDF, HF-NeuS, Voxurf [78, 80, 85, 94]) trained via differentiable volumetric rendering (NeRF [55]) have been powerful tools for this task for a while, Gaussian Splatting [40] based alternatives emerged with comparable novel view and reconstruction [29, 100] quality, within a fraction of the training time, and with orders of magnitude higher interactive rendering frame rates. These methods typically fail however under uncalibrated and sparse training images. Contemporary literature attempts to hedge against the underconstrained nature of the sparse view problem through regularization with a multitude of deep geometric priors (monocular depth [18, 76, 104], normal [98], point cloud [67] etc.), or training new priors from scratch [50, 67, 87]. Such strategies can prove to be cumbersome, impractical or costly in terms of data, memory or compute." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose to marry the best of both worlds efficiently — namely pre-existing large 3D vision feedforward models (MASt3R) and fast test-time optimization" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27476" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "(2DGS) — with the aim of addressing the underexplored problem of 3D geometric reconstruction from sparse un-posed images within minutes (less than 3 on average in our experiments) at minimal cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.395 + ], + "angle": 0, + "content": "We propose to exploit a single pre-existing foundation model (MASt3R [44]) to the fullest to bootstrap our reconstruction, without needing to load or deploy any additional deep neural networks. Given a few sparse color images, MASt3R delivers a coarse point cloud and initial camera poses. We use these to initialize a bundle adjusting 2D Gaussian Splatting (2DGS) optimization via differentiable render-and-compare, where surface elements are instantiated at the point samples, oriented initially according to their normals. Differently from the closest competition [21, 22], we harness the MASt3R [44] produced image pair correspondences to guide camera optimization during 2DGS training. This is achieved via a splatted depth reprojection error loss. The resulting camera alignment improvement proves crucial for obtaining accurate shape representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.519 + ], + "angle": 0, + "content": "By analyzing the surface element splatting from a statistical moment perspective, rendered pixels can be seen as color expectation over fragments following the ray termination probability. Based on this, we derive an expression of splatted color variance along the ray, and propose a novel loss tasked with reducing it. This results in improved reconstruction metrics and qualitative sharper mesh details as witnessed by our ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.627 + ], + "angle": 0, + "content": "The integration of these elements yields state-of-art performance in sparse pose-free 3D reconstruction, obtained in very short times on mid-range GPUs. We evaluate our method using established reconstruction and novel view datasets [1, 5, 42, 92, 97], showcasing both superior geometric reconstruction, camera pose estimation and novel view synthesis w.r.t. comparable state-of-the-art baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation of 3D shape reconstruction in the posed setting relies traditionally on mesh Chamfer distance to ground-truth, as used in multi-view reconstruction benchmarks like DTU [1]. This is a mesh pose and scale sensitive metric. In the unposed setting, reconstructions are obtained naturally up-to a rigid transformation. RANSAC powered alignments between predicted and ground-truth cameras have been shown to be acceptable for pre-evaluation reconstruction mesh alignment in the dense view setting [7], and good enough for camera pose evaluation for sparse sets [75]. However, we found these camera based alignments to be too inaccurate in the sparse setting to enable reliable mesh pre-alignments for Chamfer distance evaluation. Hence, inspired by the rigid-transform invariant evaluation in [79], we compare the reconstructed mesh depth and normal rasterizations to the groundtruth in screen space in our evaluation, and show considerable improvements compared to the competition in the standard DTU benchmark" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.907, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.187, + 0.907, + 0.299 + ], + "angle": 0, + "content": "Figure 2. 3D Foundation models meet differentiable perspective accurate Splating for fast reconstruction from pose-free sparse images. We bootstrap a bundle adjusting 2DGS with MAST3R [44] 3D point clouds and initial camera predictions. Besides standard 2DGS photometric reconstruction and geometric regularization losses, we leverage a camera reprojection error loss using MAST3R's correspondences and a novel splatted color variance reduction loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.906, + 0.361 + ], + "angle": 0, + "content": "(Tab.1,2,3). Further qualitative comparison in Blended-MVS [92], MVImgNet [97], Mip-NeRF 360 [4] illustrate the quality of our reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.369, + 0.652, + 0.385 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.729 + ], + "angle": 0, + "content": "NVS and 3D Reconstruction Novel view synthesis (NVS) aims to generate unseen views from input images [3, 54]. Neural Radiance Fields (NeRF) [55] revolutionized this field by representing 3D scenes as continuous implicit functions, though challenges remain in efficiency and quality [4-6, 66, 72, 73]. Inspired by the success of other explicit representations (e.g. meshes [33, 38, 77] and point clouds [2, 20]), recent advancements introduced primitive-based scene representations [13, 40, 90], notably 3D Gaussian Splitting (3DGS) [40], which uses anisotropic 3D Gaussians with differentiable splatting for efficient high-quality reconstruction. This has sparked extensions for surface reconstruction [29, 100, 101], neural representations [51], multi-resolution modeling [24, 99], and feedforward methods [10, 12, 23, 27, 89]. In parallel, implicit representations model 3D geometry as continuous functions [16, 37, 39, 57, 59]. Methods like IDR [93], NeuS [78], and VolSDF [94] enhance geometric accuracy but often suffer from slow rendering. In contrast, 2D Gaussian Splitting (2DGS) [29] offers real-time rendering while improving view consistency for surface reconstruction through planar Gaussian primitives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Sparse-View NVs and 3D Reconstruction Traditional NeRFs and GS methods require carefully captured images and SfM preprocessing with COLMAP [69], which could limit practical applications. The few-shot setup has been addressed for both reconstruction from Point Cloud [11, 62, 64, 65] and images. Various approaches address the challenge of limited views through regularization techniques, including depth supervision [17, 46, 58, 76], foundation model supervision [32, 84, 88], specialized architectures [91, 96], while SparseCraft [95] uses learning-free" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27477" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "MVS cues for SDF regularization. For Gaussian Splating, methods like FSGS [104] and SparseGS [86] incorporate external priors. Generalizable methods address sparse-view challenges by pretraining on large datasets, enabling feed-forward inference. Inputs include images (e.g. [10, 12, 35, 45, 49]) and 3D Points (e.g. [9, 30, 52, 60, 61, 63]). For surface reconstruction from color images, approaches like SparseNeuS [50], VolRecon [68], ReTR [47], UfoRecon [56], GeoTransfer [34], and Spurfies [67] leverage various mechanisms such as using image priors to form multiview cost volumes, using transformer architectures to enhance reconstruction from limited views, as well as using reconstruction priors trained on sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.307, + 0.485, + 0.626 + ], + "angle": 0, + "content": "Pose-Free NVS and 3D Reconstruction Pose-free methods eliminate the need for accurate camera calibration. NeRFmm [82], BARF [48], SCNeRF [36], and GARF [15] jointly optimize camera parameters and scene representations. SPARF [75] and PoRF [7] address optimization with noisy poses, though often requiring lengthy training. Recent work like Nope-NeRF [8], Lu-NeRF [14], LocalRF [53], and CF-3DGS [25] leverage depth information but typically assume dense video sequences. InstantSplat [21, 22] addresses these limitations by using MASt3R [44] for initialization and jointly optimizing camera poses and 3D models. Our work builds on these foundations but focuses specifically on using 2D Gaussian Splatting for efficient 3D surface reconstruction from sparse, uncalibrated views. We initialize with MASt3R [44] and refine through joint optimization of Gaussian parameters and camera poses, incorporating a novel variance-based color loss for improved detail preservation. Unlike previous approaches requiring dense sequences, known intrinsic, or long optimization times, our method efficiently handles sparse, uncalibrated images while achieving high-quality reconstructions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.184, + 0.648 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Given a few unposed color images \\(\\{I^i\\}_{i = 1}^N\\), our goal is to recover a 3D triangle mesh \\(S\\) and render novel views of the observed scene efficiently. We achieve this by fitting a bundle adjusting 2DGS [29] model to the images of the scene, bootstrapped by point map and correspondence feedforward predictions from a 3D foundation model (MASt3R [44]) and carefully designed regularizations. At test time, mesh \\(S\\) can be extracted using the TSDF algorithm in Open3D [103] using the converged 2DGS model rendered depth, as recommended in [29]. We detail the workings of our method below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.233, + 0.849 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "2D Gaussian Splatting Our method builds upon 2D Gaussian Splatting (2DGS) [29], which represents 3D scenes using planar disk primitives defined within local tan-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "gent planes. Unlike volumetric 3D Gaussians [40], 2DGS is a surface-based representation that improves multi-view consistency. Each 2D Gaussian is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.149, + 0.907, + 0.185 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {p}) = \\exp \\left(- \\frac {u (\\mathbf {p}) ^ {2} + v (\\mathbf {p}) ^ {2}}{2}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.907, + 0.405 + ], + "angle": 0, + "content": "where \\( u(\\mathbf{p}) \\) and \\( v(\\mathbf{p}) \\) are local UV space coordinates of a pixel \\( \\mathbf{p} \\) in the primitive's tangent plane. This ray-splat intersection can be expressed as a function of the camera parameters and splat attributes differentiably. Each primitive is parameterized by the center of the Gaussian in world coordinates \\( \\mathbf{x} \\in \\mathbb{R}^3 \\), a quaternion representing the orientation of the primitive \\( \\mathbf{q} \\in \\mathbb{R}^4 \\), scale factors for the two dimensions in the local UV tangent plane \\( \\mathbf{s} \\in \\mathbb{R}^2 \\), opacity \\( \\alpha \\in \\mathbb{R} \\) controlling the transparency of the primitive, and Spherical harmonic coefficients \\( c_{\\mathrm{sh}k} \\in \\mathbb{R}^3 | k = 1, \\dots, n \\) encoding view-dependent appearance. Rotation \\( \\mathbf{q} \\) defines a local coordinate frame \\( [\\mathbf{u}, \\mathbf{v}, \\mathbf{n}] \\), \\( \\mathbf{n} \\) being the normal vector orthogonal to the primitive's plane, and \\( \\mathbf{u} \\) and \\( \\mathbf{v} \\) spanning the tangent plane." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.466 + ], + "angle": 0, + "content": "For rendering, 2DGS uses alpha-compositing for volumetric rendering similarly to 3DGS. Given a set of primitives along a ray sorted from front to back \\( G_{1}, G_{2}, \\ldots, G_{K} \\), the final color \\( C \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.477, + 0.907, + 0.52 + ], + "angle": 0, + "content": "\\[\nC (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.58 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}^{\\prime} = \\alpha_{i}G(\\mathbf{p})\\) is the modified opacity of the \\(i\\)-th primitive and \\(\\mathbf{c}_i\\) is its color, computed from the spherical harmonic coefficients based on the viewing direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.676 + ], + "angle": 0, + "content": "MASt3R MASt3R [44] is a foundation model for 3D vision that extends DUSt3R [79] with more powerful feature extraction and matching capabilities. It tackles the multi-view stereo reconstruction task without requiring prior camera calibration or poses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.906, + 0.767 + ], + "angle": 0, + "content": "Given a pair of images \\( I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3} \\), MAS3R outputs two corresponding pointmaps \\( \\mathbf{P}_{1,1}, \\mathbf{P}_{2,1} \\in \\mathbb{R}^{W \\times H \\times 3} \\) with associated confidence maps \\( \\mathbf{O}_{1,1}, \\mathbf{O}_{2,1} \\in \\mathbb{R}^{W \\times H} \\). Each pointmap assigns a 3D point to every pixel in the image. Importantly, both pointmaps are expressed in the same coordinate frame of the first image \\( I^1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.906, + 0.859 + ], + "angle": 0, + "content": "The architecture of MASt3R consists of a Siamese ViT encoder that processes both images independently, Transformer decoders that exchange information via crossattention, and regression heads that output pointmaps and confidence maps. MASt3R is trained using a confidence-weighted regression loss:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.871, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n f}} = \\sum_ {v \\in 1, 2} \\sum_ {i} \\mathbf {O} _ {v, 1} ^ {i} \\mathcal {L} _ {\\text {r e g}} (v, i) - \\alpha \\log \\mathbf {O} _ {v, 1} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27478" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{reg}}(v,i)\\) measures the Euclidean distance between predicted and ground-truth 3D points for pixel \\(i\\), and \\(\\alpha\\) is a regularization parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.211 + ], + "angle": 0, + "content": "Besides pointmaps, MASt3R also provides dense pixelwise correspondences between the input views. These correspondences establish reliable geometric constraints across different viewpoints, which we leverage in our optimization framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.221, + 0.274, + 0.235 + ], + "angle": 0, + "content": "3.2. Scene Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Global Geometry Alignment Given a set of input images \\( I^1, I^2, \\ldots, I^N \\), we first construct a connectivity graph \\( \\mathcal{G}(\\mathcal{V}, \\mathcal{E}) \\) where each vertex \\( v \\in \\mathcal{V} \\) represents an image, and each edge \\( e = (n, m) \\in \\mathcal{E} \\) indicates a pair of images with shared visual content. For each edge in the graph, we run MASt3R to obtain pairwise pointmaps and correspondences. To align these pairwise predictions into a globally consistent reconstruction, we follow the optimization approach from MASt3R:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.386, + 0.482, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\chi , P, \\sigma} \\sum_ {e \\in \\mathcal {E}} \\sum_ {v \\in e} \\sum_ {i = 1} ^ {H W} \\mathbf {O} _ {v, e} ^ {i} \\left| \\chi_ {i} ^ {v} - \\sigma_ {e} P _ {e} \\mathbf {P} _ {v, e} ^ {i} \\right|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.589 + ], + "angle": 0, + "content": "where \\(\\chi_i^v\\) is the globally aligned pointmap for view \\(v\\) at pixel \\(i\\), \\(P_e \\in \\mathbb{R}^{3 \\times 4}\\) the rigid transformation for edge \\(e\\), \\(\\sigma_e > 0\\) is a scaling factor for edge \\(e\\), \\(\\mathbf{O}_{v,e}^i\\) the confidence value for pixel \\(i\\) in view \\(v\\) for edge \\(e\\), and finally \\(\\mathbf{P}_{v,e}^i\\) the original pointmap from MASt3R. This optimization produces globally aligned pointmaps \\(\\{\\chi^{v*}\\}_{v=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}\\) and camera parameters (intrinsics and extrinsics for each view \\(v\\)). We stabilize the estimated focal lengths by averaging them across all training views: \\(\\bar{f} = \\frac{1}{N} \\sum_{v=1}^N f_i^*\\), assuming a single-camera setup." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Primitive Initialization We leverage the merged globally aligned pointmap to initialize 2D Gaussian primitives with suitable orientation, which is crucial for accurately representing surface geometry. For each point \\(\\mathbf{p}_i\\) in the globally aligned pointmap, we compute a local surface normal using Principal Component Analysis (PCA). First, we determine the \\(k\\)-nearest neighbors \\(\\{\\mathbf{p}_j\\}_{j=1}^k\\) and compute the covariance matrix \\(\\mathbf{C} = \\frac{1}{k}\\sum_{j=1}^k (\\mathbf{p}_j - \\bar{\\mathbf{p}})(\\mathbf{p}_j - \\bar{\\mathbf{p}})^T\\), where \\(\\bar{\\mathbf{p}}\\) is the centroid of the neighborhood. We then perform eigendecomposition on \\(\\mathbf{C}\\) to obtain the eigenvectors \\(\\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\}\\) and corresponding eigenvalues \\(\\{\\lambda_1, \\lambda_2, \\lambda_3\\}\\), sorted such that \\(\\lambda_1 \\geq \\lambda_2 \\geq \\lambda_3\\). The eigenvector \\(\\mathbf{v}_3\\), associated with the smallest eigenvalue, is chosen as the normal direction, i.e., \\(\\mathbf{n} = \\mathbf{v}_3\\). To construct an orthonormal basis, we initialize an arbitrary vector \\(\\mathbf{a}\\) orthogonal to \\(\\mathbf{n} = (n_x, n_y, n_z)\\), defined as \\(\\mathbf{a} = (-n_y, n_x, 0)^T\\). The tangent vector is then computed as \\(\\mathbf{u} = \\frac{\\mathbf{n} \\times \\mathbf{a}}{\\|\\mathbf{n} \\times \\mathbf{a}\\|}\\) and the bitangent as \\(\\mathbf{v} = \\mathbf{n} \\times \\mathbf{u}\\). Finally, the local reference frame \\([\\mathbf{u}, \\mathbf{v}, \\mathbf{n}]\\) is used to define a rotation matrix, which can be converted to a quaternion \\(\\mathbf{q}_i\\) for efficient representation. Next, we initialize each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "2D Gaussian primitive with position \\(\\mathbf{x}_i\\) directly from the pointmap, rotation \\(\\mathbf{q}_i\\), scale \\(\\mathbf{s}_i\\) set proportionally to the local pointmap density, and opacity \\(\\alpha_i\\) set to a default value (e.g. 0.8). Color for the lowest spherical harmonic band (DC component) is initialized from the original image colors projected onto the point, with higher-order coefficients set to zero. This initialization preserves the surface structure from the MASt3R reconstruction while enabling efficient optimization with 2D Gaussian primitives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.695, + 0.252 + ], + "angle": 0, + "content": "3.3. Joint Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.319 + ], + "angle": 0, + "content": "Correspondence Loss Using the dense correspondences provided by MASt3R, and inspired by SPARF [75], we formulate a multi-view correspondence loss that enforces geometric consistency across different viewpoints:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.327, + 0.905, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o r r}} = w _ {p _ {n}, p _ {m}} \\rho \\left(p _ {m} - \\pi \\left(P _ {m} ^ {- 1} P _ {n} \\pi^ {- 1} \\left(p _ {n}, d _ {n}\\right)\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.566 + ], + "angle": 0, + "content": "where \\((p_n, p_m)\\) is a correspondence pair with \\(p_n \\in [0,1]^2\\) a pixel in view \\(n\\) and \\(p_m\\) its corresponding pixel in view \\(m\\). \\(w_{p_n,p_m} \\in [0,1]\\) is the confidence of the correspondence. \\(\\rho\\) represents the Huber loss [31] function. \\(\\pi: \\mathbb{R}^3 \\to \\mathbb{R}^2\\) is the projection operator mapping 3D points to pixel coordinates. \\(\\pi^{-1}: \\mathbb{R}^2 \\times \\mathbb{R} \\to \\mathbb{R}^3\\) is the back-projection operator mapping pixel coordinates and depth to 3D points. \\(d_n\\) is the 2DGS model splatted depth at pixel \\(p_n\\). \\(P_n\\) and \\(P_m\\) represent the camera parameters for views \\(n\\) and \\(m\\). This loss enforces that corresponding pixels across views should project to the same 3D point when back-projected using their respective depths and camera parameters, and thus helps guide our bundle adjustment to a good minima." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.906, + 0.617 + ], + "angle": 0, + "content": "Photometric Loss Following 3DGS and 2DGS, we use a combination of \\(\\mathcal{L}_1\\) image reconstruction loss and structural similarity (SSIM) to optimize the rendering of our system:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.627, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p h o t o}} = (1 - \\lambda) \\cdot \\mathcal {L} _ {1} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {S S I M}} + \\mathcal {L} _ {\\mathrm {r e g}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.905, + 0.699 + ], + "angle": 0, + "content": "Notice that 2DGS comes also with geometric regularization term \\(\\mathcal{L}_{\\mathrm{reg}}\\) binding pixel depth normals to splatted normals, and reducing depth distortion, as detailed in [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Variance Regularization Loss A key contribution of our method is the introduction of a splatted color variance regularization loss. The fundamental insight is that volume rendering can be interpreted as computing the expected value of colors sampled along a ray [74]:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.788, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\nC = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.9 + ], + "angle": 0, + "content": "where \\( c(\\mathbf{t}) \\) represents the color of a point \\( \\mathbf{t} \\) along the ray, and the expectation is taken with respect to \\( p(t) = \\sigma(t)T(t) \\) with \\( \\sigma(t) \\) being the density at \\( t \\) and \\( T(t) \\) representing the transmittance function." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "We hypothesize that stabilizing and robustifying our training would lead to more robust and improved reconstructions. In order to robustify the training, we would like to control how the rendered color deviates from the ground truth color when the geometry is perturbed, i.e. given divergence measure \\(\\mathcal{D}\\), we want to minimize the following worst-case loss along the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.204, + 0.484, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\sup _ {q: D (q, p) < \\rho} | \\mathbb {E} _ {\\mathbf {t} \\sim q (\\mathbf {t})} [ c (\\mathbf {t}) ] - C _ {g t} |, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.267 + ], + "angle": 0, + "content": "where \\(\\rho\\) controls the strength of the perturbation. By leveraging theorem 2 in [19], this loss can be upper bounded by:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.274, + 0.484, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} + \\eta \\sqrt {\\operatorname {V a r} _ {\\boldsymbol {t} \\sim p (\\boldsymbol {t})} [ c (\\boldsymbol {t}) ]}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Consequently, we consider minimizing the variance of the color along the ray under the distribution \\( p(t) \\) that can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.355, + 0.483, + 0.407 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {v a r}} = \\mathbb {V} a r _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] (10) \\\\ = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ (c (\\mathbf {t}) - C) ^ {2} ] = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ^ {2} ] - C ^ {2}. (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.581 + ], + "angle": 0, + "content": "For efficient implementation, we compute the first term using a modified CUDA kernel of the 2DGS splatting algorithm, where we render the square of colors in addition to the rendering of the colors. This allows for efficient forward and backward passes during optimization. Intuitively, reducing rendered color variance reduces uncertainty in the rendering, which leads to sharper and more multi-view coherent reconstructions. This reflects positively on the qualitative (Fig 5) and quantitative (Tab 4 and Fig 6) performance of our reconstructions. We note that in practice, loss \\(\\mathcal{L}_{\\mathrm{var}}\\) is averaged across pixels for a given view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Training Objective Our total optimization objective combines the three loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.621, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {p h o t o}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\text {c o r r}} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.781 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{photo}}\\), \\(\\lambda_{\\mathrm{corr}}\\), and \\(\\lambda_{\\mathrm{var}}\\) are weighting factors that balance the contribution of each term. We optimize all parameters simultaneously using the Adam [41] optimizer with a learning rate scheduling following [29]. Unlike multi-stage approaches [75], our single-stage optimization allows for faster convergence while maintaining high reconstruction quality. The optimization updates both camera extrinsics \\(\\{P\\}\\) per input view and 2D Gaussian primitive parameters \\(\\{\\alpha, \\mathbf{c}, \\mathbf{q}, \\mathbf{s}\\}\\) according to the gradient of the total loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.224, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We evaluate our approach across 3D reconstruction, novel view synthesis, and camera pose estimation. We outline the experimental setup, including datasets, baseline methods, and implementation details, followed by quantitative and qualitative comparisons. Ablation studies further analyze the impact of key components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.871, + 0.108 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Configuration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.905, + 0.159 + ], + "angle": 0, + "content": "Our method is tested on benchmark datasets [1, 5, 42, 92, 97]. Further implementation details and additional qualitative results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.16, + 0.906, + 0.281 + ], + "angle": 0, + "content": "Few-Shot 3D Reconstruction: We evaluate sparse 3D reconstruction on DTU [1], BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]. For 3-view reconstruction on DTU, we follow SparseNeuS [50], evaluating 15 scenes with two sets of 3-view inputs, reporting the mean per scene. We also provide qualitative comparisons for 3, 6, and 12-view settings on selected scenes from BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.906, + 0.357 + ], + "angle": 0, + "content": "Few-Shot Novel View Synthesis: We conduct novel view synthesis (NVS) on Tanks and Temples [42] (8 scenes), MVImgNet [97] (7 scenes), and MipNeRF360 [5] (9 scenes), following InstantSplat [21, 22]. Results are reported under sparse input settings (3, 6, and 12 views)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.448 + ], + "angle": 0, + "content": "Camera Pose Estimation: Camera poses are evaluated on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5], following InstantSplat [21, 22]. Additionally, for few-shot reconstruction, we assess camera poses on DTU [1] using the same 3-view evaluation protocol as in the reconstruction task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.596 + ], + "angle": 0, + "content": "Evaluation Metrics. For 3D reconstruction quality, we employ the Absolute Relative Error (Rel) and Normal Consistency (NC). Rel measures the accuracy of depth estimation following DUSt3R [79], while NC evaluates the geometric fidelity by measuring the alignment between predicted and ground truth surface normals. We apply these metrics to our final meshes. We compute ground truth normals from the gradients of the ground truth DTU depth maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.733 + ], + "angle": 0, + "content": "For novel view synthesis, we use standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [81], and Learned Perceptual Image Patch Similarity (LPIPS) [102]. We note that for this evaluation we follow InstantSplat[21, 22]: for novel test views without pre-computed camera poses, we perform an additional camera optimization step, which optimizes the test cameras while keeping the 2D Gaussian primitives fixed using the \\(\\mathcal{L}_1\\) photometric reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.793 + ], + "angle": 0, + "content": "For camera pose estimation, we report the Absolute Trajectory Error (ATE) as defined in [8, 21], using COLMAP-computed poses from dense views as ground-truth references." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.806, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Baselines. We compare our method against several state-of-the-art approaches:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "- NVS specific methods: Pose-free methods NoPe-NeRF [8], CF-3DGS [25], NeRFmm [82], SPARF [75] and InstantSplat [21, 22] (in both S and XL configurations). We also include comparisons with 3DGS [40] and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on DTU dataset from 3 input images." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.327, + 0.483, + 0.357 + ], + "angle": 0, + "content": "FSGS [104], which use pre-computed camera parameters from Colmap [69]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.358, + 0.483, + 0.463 + ], + "angle": 0, + "content": "- Reconstruction-specific methods: MASt3R [44], Colmap [69], UfoRecon [56], CasMVSNet [26], Spurfies [67], SpaRP [87], SparseCraft [95], InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22]. Of these, similar to our approach, InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22] further refine cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.485, + 0.748 + ], + "angle": 0, + "content": "Implementation Details. Our framework is implemented in PyTorch. We set the optimization iterations to 1k for DTU dataset reconstruction experiments and between 2k-4k for other reconstruction and novel view synthesis experiments. For test-time optimization, camera parameters are jointly optimized for 1k iterations (while keeping the learned Gaussian parameters fixed). We initialize our pipeline with MASt3R [44] configured at a resolution of 512 pixels and use Open3D [103] to compute normals for the point cloud, which are used to initialize rotation parameters of the 2D Gaussians. All experiments are conducted on a single NVIDIA A6000 GPU. We set \\(\\lambda_{\\mathrm{photo}} = 1.0\\) and \\(\\lambda_{\\mathrm{corr}} = 5 \\times 10^{-5}\\). The variance loss weight \\(\\lambda_{\\mathrm{var}}\\) follows a cosine annealing schedule: \\(\\lambda_{\\mathrm{var}} = 0.5\\left(1 + \\cos \\left(\\frac{\\pi \\cdot t}{T}\\right)\\right)\\), where \\(t\\) denotes the current training step and \\(T\\) the total number of steps. This schedule gradually decreases \\(\\lambda_{\\mathrm{var}}\\) from 1.0 to 0.0, ensuring an adaptive regularization throughout training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.36, + 0.773 + ], + "angle": 0, + "content": "4.2. 3D Reconstruction Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.884 + ], + "angle": 0, + "content": "For few-shot reconstruction, we follow the evaluation protocol from prior work [50, 56, 95], cleaning the generated meshes with masks from training views. We evaluate the Absolute Relative Error (Rel) and Normal Consistency (NC) between predicted depths normals from the extracted meshes and the ground truth depths normals from the DTU dataset [1]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 and 2 present the quantitative results on the DTU" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.374, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on BlendedMVS dataset from 3 input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.549 + ], + "angle": 0, + "content": "dataset for Rel and NC metrics, respectively. Fig 3, 4 and 7 show qualitative reconstruction comparisons, with additional qualitative comparisons with SpaRP [87] included in the supplementary material. Additionally, the supplementary material contains extended qualitative results, in the form of video comparisons demonstrating our reconstructions for several datasets. Our method achieves state-of-the-art performance, outperforming all competing approaches across the majority of scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Several key observations can be made from these results. Pretrained generalizable methods (UfoRecon [56] and CasMVSNet [26]) generalize poorly when faced with noisy initial camera parameters from MASt3R [44], resulting in significantly worse performance. Despite being designed for sparse reconstruction, SparseCraft [95] struggles with the noisy camera poses and point clouds from MASt3R [44]. Spurfies [67], which employs an SDF prior trained on ShapeNet, performs particularly poorly with noisy camera poses, resulting in worse mesh extractions. InstantSplat [21, 22] variants (InstSplatGOF and InstSplat2DGS) perform better than other competing approaches, but still produce overly smooth meshes lacking fine details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Our method significantly outperforms these approaches due to several key advantages, including the variance regularization loss (Sec. 3.3) promoting sharper and more detailed reconstructions Fig 5, The correspondence loss (Sec. 3.3) and joint optimization framework effectively refine the initial noisy camera poses from MAsst3R [44]. Our method's ability to preserve fine geometric details is also evidenced by the superior Normal Consistency scores." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Table 3 presents camera pose estimation results measured by ATE. Our approach achieves significant improve" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27481" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.184 + ], + "angle": 0, + "content": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]2.4920.4812.025.398.866.34.719.017.7010.335.862.505.04.564.957.34
UFORecon [56]35.5557.1538.8135.4750.0244.9434.7053.5641.6747.9340.2239.3439.7533.3449.0442.77
CasMVSNet [26]9.7638.0037.0317.3019.3924.106.5729.5711.8414.9212.388.383.939.8713.4417.10
Spurfies [67]2.9652.1558.4220.2514.6619.244.9712.447.5010.647.366.2638.6125.8444.1921.70
InstSplatGOF [21]9.5423.1413.7415.6213.3012.3311.1912.2114.6915.6113.6112.3015.3112.9012.0513.84
Sparsecraft [95]2.9610.517.162.928.179.443.836.197.616.444.756.435.367.658.146.50
InstantSplat2DGS [22]2.4221.749.083.086.773.603.635.245.827.124.482.783.623.113.435.73
Ours2.4212.3137.93.156.053.5263.484.985.296.224.392.413.463.323.384.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.186, + 0.874, + 0.198 + ], + "angle": 0, + "content": "Table 1. 3-view reconstruction results on DTU dataset The relative error (rel↓) is reported for 15 scans. Best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.2, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122\\( \\uparrow \\) Mean
MASt3R [44]0.8880.6950.8440.8740.8370.8390.8510.7540.8240.7450.8710.8750.8840.8310.8460.830
UFORecon [56]0.3620.2780.5060.4690.2770.3720.3830.3610.3620.3380.4240.3660.3320.4020.3390.371
CasMVSNet [26]0.7180.5020.6190.7560.6930.6830.7400.5640.7040.6810.7850.7950.8590.7780.7550.709
Spurfies [67]0.8630.350.2970.7110.7420.6680.7710.6710.7770.6830.780.7710.5090.5720.4330.640
InstSplatGOF [21]0.7740.5810.6890.6960.7460.7500.7100.6580.6820.6530.6950.7240.7100.6630.6500.692
Sparsecraft [95]0.8410.6790.8270.8640.7280.7140.8340.7560.7460.7650.8370.6290.8010.7610.7350.768
InstSplat2DGS [22]0.8820.6160.860.8990.8340.8650.8450.7830.8100.7730.8820.8490.8950.8270.8550.832
Ours0.9040.750.8770.8960.8660.8870.8670.7990.8450.7890.8880.8740.9020.850.8650.857
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.299, + 0.884, + 0.312 + ], + "angle": 0, + "content": "Table 2. 3-view reconstruction results on DTU dataset Normal consistency (NC↑) is reported for 15 scans. Best results are bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.402 + ], + "angle": 0, + "content": "ments in camera pose estimation, with a \\(13.1\\%\\) reduction in error compared to the best-performing baseline (InstSplatGOF [21]). This demonstrates the effectiveness of our correspondence loss in refining the initial noisy camera poses from MAST3R [44]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.412, + 0.378, + 0.427 + ], + "angle": 0, + "content": "4.3. Novel View Synthesis Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.509 + ], + "angle": 0, + "content": "We evaluate our method's capability for novel view synthesis on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5] datasets. Quantitative and qualitative result comparisons on all datasets are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.519, + 0.254, + 0.533 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.587 + ], + "angle": 0, + "content": "We conduct ablation studies to analyze the contribution of each component in our method, including an experiment to study the impact of varying the number of input views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.833 + ], + "angle": 0, + "content": "Component Ablation These experiments were performed using 3 input unposed views on 5 testing scenes from the DTU dataset using two sets of views, following the protocols established in [50, 56]. Tab 4 shows the impact of incorporating different components into our pipeline. Fig 5 shows a qualitative ablation comparison. Camera optimization using bundle adjustment improves all metrics. Incorporating the correspondence loss further enhances this alignment, resulting in improved reconstructions. Splat orientation initialization with pointmap normals improves results further. Notice that our variance loss contributes to both the final metrics and the qualitative aspect of our results. The reconstructions with variance loss exhibit significantly sharper details and better preservation of fine structures, validating the quantitative improvements observed in our metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Number of Views Ablation We progressively increase the number of views and perform pose-free reconstruction with our method using scan37 from DTU. We evaluate by computing the mean metrics across both test view sets. As" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.324, + 0.907, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.629, + 0.907, + 0.671 + ], + "angle": 0, + "content": "Figure 5. Qualitative ablation of the component of our method, showing also the effect of variance loss. Left blue box: w/o \\(\\mathcal{L}_{\\mathrm{var}}\\). Right red box: w \\(\\mathcal{L}_{\\mathrm{var}}\\). (3 input images)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.907, + 0.739 + ], + "angle": 0, + "content": "shown in Figure 6, our method benefits favorably from increased views. The benefits of all components of our method are also maintained, if not increased. This includes our variance loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.745, + 0.655, + 0.762 + ], + "angle": 0, + "content": "5. Running time" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In our reconstruction experiments, we observed that the mean training time on DTU [1] for 1000 iterations (similar to InstantSplat-XL [21, 22]) was roughly 146s on a single Nvidia A6000 GPU. For a fair comparison, we also trained InstantSplat-XL [21, 22] on the same GPU and observed that the mean training time was around 128s. Hence, our training time is in the same range as InstantSplat [21, 22] while offering significantly higher reconstruction and novel-view quality." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27482" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.148 + ], + "angle": 0, + "content": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]0.02020.06890.00940.02030.01710.03920.02910.02940.02710.03260.01670.02100.02890.01920.02420.0269
InstSplatGOF [21]0.01800.05360.00890.01360.00490.02730.02630.02390.01800.03230.01280.02380.02490.01750.01490.0214
InstSplat2DGS [22]0.01920.08950.00900.01540.01230.03380.02940.02720.02420.03220.01660.02100.02740.01630.01750.0261
Ours0.00630.04500.00690.01240.01750.02940.01980.02830.01460.02730.00820.01790.02420.01120.01010.0186
" + }, + { + "type": "table_caption", + "bbox": [ + 0.203, + 0.149, + 0.794, + 0.162 + ], + "angle": 0, + "content": "Table 3. Results on DTU dataset Results include ATE \\( \\downarrow \\) metrics for 15 scans. Best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.174, + 0.451, + 0.245 + ], + "angle": 0, + "content": "
Methodsrel ↓NC ↑ATE ↓
2DGS7.930.7850.0327
BA-2DGS7.790.7880.0284
BA-2DGS+Lcorr7.410.8040.0256
BA-2DGS+Lcorr+Norm. init.7.030.8290.0237
BA-2DGS+Lcorr+Norm. init.+Lvar6.790.8340.0238
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.256, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Table 4. Ablation on DTU dataset, adding one component at a time. Results include rel and NC metrics. Best results are bold." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.313, + 0.283, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.313, + 0.482, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Effect of increasing the number of input views on different metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.515, + 0.364, + 0.531 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.542, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Despite the promising results, several limitations remain, which present opportunities for future research. As foundation models keep on improving, better initializations could be available in the future. Furthermore, while our approach handles a variety of scenes across different datasets, performance may be limited for highly complex geometry, very thin structures, or semi-transparent objects (observed for the glass bottle in Fig 7 room scene in MipNeRF360 [5]), which are common challenges for surface based methods especially in the sparse setting. Additionally, our method assumes a static scene under consistent lighting. Extending it to more dynamic scenarios is one of the challenges we are looking forward to exploring." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.21, + 0.769 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We addressed the underexplored problem of shape reconstruction from sparse uncalibrated multi-view images. Our method leverages foundation models for initial camera and geometry estimation, followed by joint optimization of 2D Gaussian primitives and camera parameters. The introduction of a variance regularization loss significantly improves the detail and sharpness of reconstructions, while our correspondence matching based framework enhances" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.174, + 0.905, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.676, + 0.905, + 0.704 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on MVImgNet and MipNeRF360 datasets from 6 and 12 input images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.791 + ], + "angle": 0, + "content": "camera pose accuracy. Extensive evaluations on multiple datasets demonstrate that our approach consistently outperforms state-of-the-art methods in reconstruction quality, novel view synthesis, and camera pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27483" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.245, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[3] Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1034-1040. IEEE, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.302, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.373, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.445, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.516, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[7] Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, and Philip Torr. Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.574, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[8] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.645, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[9] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.703, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[10] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4182-4194, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.191, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[14] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312-18321, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In European Conference on Computer Vision, pages 264-280. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[16] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.455, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[18] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[19] John Duchi, Peter Glynn, and Hongseok Namkoong. Statistics of robust optimization: A generalized empirical likelihood approach, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.568, + 0.905, + 0.609 + ], + "angle": 0, + "content": "[20] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.61, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[21] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2(3):4, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[22] Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. arXiv preprint arXiv:2403.20309, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.763, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[23] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. arXiv preprint arXiv:2410.18956, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27484" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[25] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatt-ing. arXiv preprint arXiv:2312.07504, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[26] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[27] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jisang Han, Jiaolong Yang, Chong Luo, and Seungryong Kim. Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv preprint arXiv:2410.22128, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[28] Berthold KP Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. 1970. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[30] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.427, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[33] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Neural mesh-based graphics. In European Conference on Computer Vision, pages 739-757. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[34] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Geotransfer: Generalizable few-shot multi-view reconstruction via transfer learning. arXiv preprint arXiv:2408.14724, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[35] Shubhendu Jena, Shishir Reddy Vutukur, and Adnane Boukhayma. Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting. arXiv, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[36] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5846-5854, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[37] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[38] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[39] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[40] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[42] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[43] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[44] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.336, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[45] Qian Li, Franck Moulton, and Adnane Boukhayma. Learning generalizable light field networks from few images. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.406, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[46] Qian Li, Franck Multon, and Adnane Boukhayma. Regularizing neural radiance fields from sparse rgb-d inputs. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2320-2324. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.463, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[47] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.52, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[48] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.577, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[49] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Fast generalizable gaussian splattering reconstruction from multi-view stereo. arXiv preprint arXiv:2405.12218, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.647, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[50] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[51] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[52] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6326-6337, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27485" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.137, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[54] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.209, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[55] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.281, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[56] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.367, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[57] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[58] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[59] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.483, + 0.638 + ], + "angle": 0, + "content": "[60] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.641, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[61] Amine Ouasfi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. 3DV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.483, + 0.724 + ], + "angle": 0, + "content": "[62] Amine Ouasfi and Adnane Boukhayma. Few-shot unsupervised implicit neural shape representation learning with spatial adversaries. arXiv preprint arXiv:2408.15114, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[63] Amine Ouasfi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[64] Amine Ouasfi and Adnane Boukhayma. Unsupervised occupancy learning from sparse point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21729-21739, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[65] Amine Ouasfi, Shubhendu Jena, Eric Marchand, and Adnane Boukhayma. Toward robust neural reconstruction from sparse point sets. arXiv preprint arXiv:2412.16361, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[66] Eric Penner and Li Zhang. Soft 3D Reconstruction for View Synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[67] Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Spurfies: Sparse surface reconstruction using local geometry priors. arXiv preprint arXiv:2408.16544, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.192, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[68] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.274, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[69] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.331, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[70] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.387, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[71] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[72] Steven M Seitz and Charles R Dyer. Photorealistic Scene Reconstruction by Voxel Coloring, 2002. US Patent 6,363,170. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.484, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[73] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.568, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[74] Andrea Tagliaasacchi and Ben Mildenhall. Volume rendering digest (for nerf), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.596, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[75] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[76] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. arXiv preprint arXiv:2303.16196, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.721, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[77] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.763, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[78] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27486" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[80] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35:1966-1978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[81] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[82] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[83] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[84] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.364, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[85] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[86] Haolin Xiong, Sairishek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[87] Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. In European Conference on Computer Vision, pages 143-163. Springer, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[89] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[90] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[91] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[92] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[93] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[94] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[95] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[96] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.361, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[97] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.446, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.516, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[99] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[100] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[101] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[102] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.905, + 0.81 + ], + "angle": 0, + "content": "[103] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.813, + 0.905, + 0.867 + ], + "angle": 0, + "content": "[104] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27487" + } + ] +] \ No newline at end of file diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_origin.pdf b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a53544fb21ddc8ff7c27c69f99afb74d4db0895a --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/e0c26abe-dde6-464c-bb5e-932b397213e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dcc85450bacce25d7b85438e7453a1953e6590a6105bf43fc10468a09ea0770 +size 24151733 diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/full.md b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b0a30531e1b8244a8d740250d2791aa69996a0a --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/full.md @@ -0,0 +1,374 @@ +# Sparfels: Fast Reconstruction from Sparse Unposed Imagery + +Shubhendu Jena*, Amine Ousafi*, Mae Younes, Adnane Boukhayma Inria, Univ. Rennes, CNRS, IRISA + +![](images/2e4c4da61c26933d1d0060d5a82578777048f050ccd480691d2efac615dd4c02.jpg) +Figure 1. Examples of reconstruction meshes obtained within 3 minutes from sparse pose-free images via our method. We used here 6 (left) and 3 (right) input images respectively from scenes of datasets MVImgNet [97] and BlendedMVS [92]. + +# Abstract + +We present a method for Sparse view reconstruction with surface element splatting that runs within 3 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate state-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view benchmarks based on established multi-view datasets. Code will be made available at https://shubhendujena.github.io/Sparfels-web/ + +# 1. Introduction + +3D reconstruction is a long standing problem at the intersection of computer vision, graphics and machine learning. It is core to a myriad of downstream applications in virtual and augmented reality, autonomous driving, medical imaging, film and animation, to name a few. Classical optimization based techniques such as structure from motion and + +multi-view stereo [70, 71], shape from visual cues such as shading or silhouette [28, 43], photometric stereo [83] etc., have been popular for decades in this regard. More recently, deep learning has been progressively disrupting 3D vision and top-down approaches to vision more broadly. Among the latest additions, large vision and foundation models (e.g. DUSt3R [79], MASt3R [44]) can offer robustness and fast feed-forward inference compared to classical photogrammetry [70, 71]. Given a few raw uncalibrated images of a scene, they are capable of inferring reasonable relative poses and a competitive coarse 3D structure. These predictions can lack precision and details, though, and cannot enable stand-alone novel view synthesis. + +Furthermore, state-of-the-art data-prior free scene specific learning methods for reconstruction can deliver highly detailed geometry and impressive novel view synthesis capabilities from dense calibrated image grids. While implicit and some grid shape representations (e.g. NeuS, VolSDF, HF-NeuS, Voxurf [78, 80, 85, 94]) trained via differentiable volumetric rendering (NeRF [55]) have been powerful tools for this task for a while, Gaussian Splatting [40] based alternatives emerged with comparable novel view and reconstruction [29, 100] quality, within a fraction of the training time, and with orders of magnitude higher interactive rendering frame rates. These methods typically fail however under uncalibrated and sparse training images. Contemporary literature attempts to hedge against the underconstrained nature of the sparse view problem through regularization with a multitude of deep geometric priors (monocular depth [18, 76, 104], normal [98], point cloud [67] etc.), or training new priors from scratch [50, 67, 87]. Such strategies can prove to be cumbersome, impractical or costly in terms of data, memory or compute. + +In this paper, we propose to marry the best of both worlds efficiently — namely pre-existing large 3D vision feedforward models (MASt3R) and fast test-time optimization + +(2DGS) — with the aim of addressing the underexplored problem of 3D geometric reconstruction from sparse un-posed images within minutes (less than 3 on average in our experiments) at minimal cost. + +We propose to exploit a single pre-existing foundation model (MASt3R [44]) to the fullest to bootstrap our reconstruction, without needing to load or deploy any additional deep neural networks. Given a few sparse color images, MASt3R delivers a coarse point cloud and initial camera poses. We use these to initialize a bundle adjusting 2D Gaussian Splatting (2DGS) optimization via differentiable render-and-compare, where surface elements are instantiated at the point samples, oriented initially according to their normals. Differently from the closest competition [21, 22], we harness the MASt3R [44] produced image pair correspondences to guide camera optimization during 2DGS training. This is achieved via a splatted depth reprojection error loss. The resulting camera alignment improvement proves crucial for obtaining accurate shape representations. + +By analyzing the surface element splatting from a statistical moment perspective, rendered pixels can be seen as color expectation over fragments following the ray termination probability. Based on this, we derive an expression of splatted color variance along the ray, and propose a novel loss tasked with reducing it. This results in improved reconstruction metrics and qualitative sharper mesh details as witnessed by our ablation studies. + +The integration of these elements yields state-of-art performance in sparse pose-free 3D reconstruction, obtained in very short times on mid-range GPUs. We evaluate our method using established reconstruction and novel view datasets [1, 5, 42, 92, 97], showcasing both superior geometric reconstruction, camera pose estimation and novel view synthesis w.r.t. comparable state-of-the-art baselines. + +Evaluation of 3D shape reconstruction in the posed setting relies traditionally on mesh Chamfer distance to ground-truth, as used in multi-view reconstruction benchmarks like DTU [1]. This is a mesh pose and scale sensitive metric. In the unposed setting, reconstructions are obtained naturally up-to a rigid transformation. RANSAC powered alignments between predicted and ground-truth cameras have been shown to be acceptable for pre-evaluation reconstruction mesh alignment in the dense view setting [7], and good enough for camera pose evaluation for sparse sets [75]. However, we found these camera based alignments to be too inaccurate in the sparse setting to enable reliable mesh pre-alignments for Chamfer distance evaluation. Hence, inspired by the rigid-transform invariant evaluation in [79], we compare the reconstructed mesh depth and normal rasterizations to the groundtruth in screen space in our evaluation, and show considerable improvements compared to the competition in the standard DTU benchmark + +![](images/60af467a4cb250858e54d2498e0e32ad7d04c0446ce61d2332cd7bc5c79e19ec.jpg) +Figure 2. 3D Foundation models meet differentiable perspective accurate Splating for fast reconstruction from pose-free sparse images. We bootstrap a bundle adjusting 2DGS with MAST3R [44] 3D point clouds and initial camera predictions. Besides standard 2DGS photometric reconstruction and geometric regularization losses, we leverage a camera reprojection error loss using MAST3R's correspondences and a novel splatted color variance reduction loss. + +(Tab.1,2,3). Further qualitative comparison in Blended-MVS [92], MVImgNet [97], Mip-NeRF 360 [4] illustrate the quality of our reconstructions. + +# 2. Related work + +NVS and 3D Reconstruction Novel view synthesis (NVS) aims to generate unseen views from input images [3, 54]. Neural Radiance Fields (NeRF) [55] revolutionized this field by representing 3D scenes as continuous implicit functions, though challenges remain in efficiency and quality [4-6, 66, 72, 73]. Inspired by the success of other explicit representations (e.g. meshes [33, 38, 77] and point clouds [2, 20]), recent advancements introduced primitive-based scene representations [13, 40, 90], notably 3D Gaussian Splitting (3DGS) [40], which uses anisotropic 3D Gaussians with differentiable splatting for efficient high-quality reconstruction. This has sparked extensions for surface reconstruction [29, 100, 101], neural representations [51], multi-resolution modeling [24, 99], and feedforward methods [10, 12, 23, 27, 89]. In parallel, implicit representations model 3D geometry as continuous functions [16, 37, 39, 57, 59]. Methods like IDR [93], NeuS [78], and VolSDF [94] enhance geometric accuracy but often suffer from slow rendering. In contrast, 2D Gaussian Splitting (2DGS) [29] offers real-time rendering while improving view consistency for surface reconstruction through planar Gaussian primitives. + +Sparse-View NVs and 3D Reconstruction Traditional NeRFs and GS methods require carefully captured images and SfM preprocessing with COLMAP [69], which could limit practical applications. The few-shot setup has been addressed for both reconstruction from Point Cloud [11, 62, 64, 65] and images. Various approaches address the challenge of limited views through regularization techniques, including depth supervision [17, 46, 58, 76], foundation model supervision [32, 84, 88], specialized architectures [91, 96], while SparseCraft [95] uses learning-free + +MVS cues for SDF regularization. For Gaussian Splating, methods like FSGS [104] and SparseGS [86] incorporate external priors. Generalizable methods address sparse-view challenges by pretraining on large datasets, enabling feed-forward inference. Inputs include images (e.g. [10, 12, 35, 45, 49]) and 3D Points (e.g. [9, 30, 52, 60, 61, 63]). For surface reconstruction from color images, approaches like SparseNeuS [50], VolRecon [68], ReTR [47], UfoRecon [56], GeoTransfer [34], and Spurfies [67] leverage various mechanisms such as using image priors to form multiview cost volumes, using transformer architectures to enhance reconstruction from limited views, as well as using reconstruction priors trained on sparse point clouds. + +Pose-Free NVS and 3D Reconstruction Pose-free methods eliminate the need for accurate camera calibration. NeRFmm [82], BARF [48], SCNeRF [36], and GARF [15] jointly optimize camera parameters and scene representations. SPARF [75] and PoRF [7] address optimization with noisy poses, though often requiring lengthy training. Recent work like Nope-NeRF [8], Lu-NeRF [14], LocalRF [53], and CF-3DGS [25] leverage depth information but typically assume dense video sequences. InstantSplat [21, 22] addresses these limitations by using MASt3R [44] for initialization and jointly optimizing camera poses and 3D models. Our work builds on these foundations but focuses specifically on using 2D Gaussian Splatting for efficient 3D surface reconstruction from sparse, uncalibrated views. We initialize with MASt3R [44] and refine through joint optimization of Gaussian parameters and camera poses, incorporating a novel variance-based color loss for improved detail preservation. Unlike previous approaches requiring dense sequences, known intrinsic, or long optimization times, our method efficiently handles sparse, uncalibrated images while achieving high-quality reconstructions. + +# 3. Method + +Given a few unposed color images $\{I^i\}_{i = 1}^N$ , our goal is to recover a 3D triangle mesh $S$ and render novel views of the observed scene efficiently. We achieve this by fitting a bundle adjusting 2DGS [29] model to the images of the scene, bootstrapped by point map and correspondence feedforward predictions from a 3D foundation model (MASt3R [44]) and carefully designed regularizations. At test time, mesh $S$ can be extracted using the TSDF algorithm in Open3D [103] using the converged 2DGS model rendered depth, as recommended in [29]. We detail the workings of our method below. + +# 3.1. Preliminaries + +2D Gaussian Splatting Our method builds upon 2D Gaussian Splatting (2DGS) [29], which represents 3D scenes using planar disk primitives defined within local tan- + +gent planes. Unlike volumetric 3D Gaussians [40], 2DGS is a surface-based representation that improves multi-view consistency. Each 2D Gaussian is defined as: + +$$ +G (\mathbf {p}) = \exp \left(- \frac {u (\mathbf {p}) ^ {2} + v (\mathbf {p}) ^ {2}}{2}\right) \tag {1} +$$ + +where $u(\mathbf{p})$ and $v(\mathbf{p})$ are local UV space coordinates of a pixel $\mathbf{p}$ in the primitive's tangent plane. This ray-splat intersection can be expressed as a function of the camera parameters and splat attributes differentiably. Each primitive is parameterized by the center of the Gaussian in world coordinates $\mathbf{x} \in \mathbb{R}^3$ , a quaternion representing the orientation of the primitive $\mathbf{q} \in \mathbb{R}^4$ , scale factors for the two dimensions in the local UV tangent plane $\mathbf{s} \in \mathbb{R}^2$ , opacity $\alpha \in \mathbb{R}$ controlling the transparency of the primitive, and Spherical harmonic coefficients $c_{\mathrm{sh}k} \in \mathbb{R}^3 | k = 1, \dots, n$ encoding view-dependent appearance. Rotation $\mathbf{q}$ defines a local coordinate frame $[\mathbf{u}, \mathbf{v}, \mathbf{n}]$ , $\mathbf{n}$ being the normal vector orthogonal to the primitive's plane, and $\mathbf{u}$ and $\mathbf{v}$ spanning the tangent plane. + +For rendering, 2DGS uses alpha-compositing for volumetric rendering similarly to 3DGS. Given a set of primitives along a ray sorted from front to back $G_{1}, G_{2}, \ldots, G_{K}$ , the final color $C$ is computed as: + +$$ +C (\mathbf {p}) = \sum_ {i = 1} ^ {K} \alpha_ {i} ^ {\prime} \mathbf {c} _ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} ^ {\prime}\right), \tag {2} +$$ + +where $\alpha_{i}^{\prime} = \alpha_{i}G(\mathbf{p})$ is the modified opacity of the $i$ -th primitive and $\mathbf{c}_i$ is its color, computed from the spherical harmonic coefficients based on the viewing direction. + +MASt3R MASt3R [44] is a foundation model for 3D vision that extends DUSt3R [79] with more powerful feature extraction and matching capabilities. It tackles the multi-view stereo reconstruction task without requiring prior camera calibration or poses. + +Given a pair of images $I^1, I^2 \in \mathbb{R}^{W \times H \times 3}$ , MAS3R outputs two corresponding pointmaps $\mathbf{P}_{1,1}, \mathbf{P}_{2,1} \in \mathbb{R}^{W \times H \times 3}$ with associated confidence maps $\mathbf{O}_{1,1}, \mathbf{O}_{2,1} \in \mathbb{R}^{W \times H}$ . Each pointmap assigns a 3D point to every pixel in the image. Importantly, both pointmaps are expressed in the same coordinate frame of the first image $I^1$ . + +The architecture of MASt3R consists of a Siamese ViT encoder that processes both images independently, Transformer decoders that exchange information via crossattention, and regression heads that output pointmaps and confidence maps. MASt3R is trained using a confidence-weighted regression loss: + +$$ +\mathcal {L} _ {\text {c o n f}} = \sum_ {v \in 1, 2} \sum_ {i} \mathbf {O} _ {v, 1} ^ {i} \mathcal {L} _ {\text {r e g}} (v, i) - \alpha \log \mathbf {O} _ {v, 1} ^ {i}, \tag {3} +$$ + +where $\mathcal{L}_{\mathrm{reg}}(v,i)$ measures the Euclidean distance between predicted and ground-truth 3D points for pixel $i$ , and $\alpha$ is a regularization parameter. + +Besides pointmaps, MASt3R also provides dense pixelwise correspondences between the input views. These correspondences establish reliable geometric constraints across different viewpoints, which we leverage in our optimization framework. + +# 3.2. Scene Initialization + +Global Geometry Alignment Given a set of input images $I^1, I^2, \ldots, I^N$ , we first construct a connectivity graph $\mathcal{G}(\mathcal{V}, \mathcal{E})$ where each vertex $v \in \mathcal{V}$ represents an image, and each edge $e = (n, m) \in \mathcal{E}$ indicates a pair of images with shared visual content. For each edge in the graph, we run MASt3R to obtain pairwise pointmaps and correspondences. To align these pairwise predictions into a globally consistent reconstruction, we follow the optimization approach from MASt3R: + +$$ +\min _ {\chi , P, \sigma} \sum_ {e \in \mathcal {E}} \sum_ {v \in e} \sum_ {i = 1} ^ {H W} \mathbf {O} _ {v, e} ^ {i} \left| \chi_ {i} ^ {v} - \sigma_ {e} P _ {e} \mathbf {P} _ {v, e} ^ {i} \right|, \tag {4} +$$ + +where $\chi_i^v$ is the globally aligned pointmap for view $v$ at pixel $i$ , $P_e \in \mathbb{R}^{3 \times 4}$ the rigid transformation for edge $e$ , $\sigma_e > 0$ is a scaling factor for edge $e$ , $\mathbf{O}_{v,e}^i$ the confidence value for pixel $i$ in view $v$ for edge $e$ , and finally $\mathbf{P}_{v,e}^i$ the original pointmap from MASt3R. This optimization produces globally aligned pointmaps $\{\chi^{v*}\}_{v=1}^N \in \mathbb{R}^{H \times W \times 3}$ and camera parameters (intrinsics and extrinsics for each view $v$ ). We stabilize the estimated focal lengths by averaging them across all training views: $\bar{f} = \frac{1}{N} \sum_{v=1}^N f_i^*$ , assuming a single-camera setup. + +Primitive Initialization We leverage the merged globally aligned pointmap to initialize 2D Gaussian primitives with suitable orientation, which is crucial for accurately representing surface geometry. For each point $\mathbf{p}_i$ in the globally aligned pointmap, we compute a local surface normal using Principal Component Analysis (PCA). First, we determine the $k$ -nearest neighbors $\{\mathbf{p}_j\}_{j=1}^k$ and compute the covariance matrix $\mathbf{C} = \frac{1}{k}\sum_{j=1}^k (\mathbf{p}_j - \bar{\mathbf{p}})(\mathbf{p}_j - \bar{\mathbf{p}})^T$ , where $\bar{\mathbf{p}}$ is the centroid of the neighborhood. We then perform eigendecomposition on $\mathbf{C}$ to obtain the eigenvectors $\{\mathbf{v}_1, \mathbf{v}_2, \mathbf{v}_3\}$ and corresponding eigenvalues $\{\lambda_1, \lambda_2, \lambda_3\}$ , sorted such that $\lambda_1 \geq \lambda_2 \geq \lambda_3$ . The eigenvector $\mathbf{v}_3$ , associated with the smallest eigenvalue, is chosen as the normal direction, i.e., $\mathbf{n} = \mathbf{v}_3$ . To construct an orthonormal basis, we initialize an arbitrary vector $\mathbf{a}$ orthogonal to $\mathbf{n} = (n_x, n_y, n_z)$ , defined as $\mathbf{a} = (-n_y, n_x, 0)^T$ . The tangent vector is then computed as $\mathbf{u} = \frac{\mathbf{n} \times \mathbf{a}}{\|\mathbf{n} \times \mathbf{a}\|}$ and the bitangent as $\mathbf{v} = \mathbf{n} \times \mathbf{u}$ . Finally, the local reference frame $[\mathbf{u}, \mathbf{v}, \mathbf{n}]$ is used to define a rotation matrix, which can be converted to a quaternion $\mathbf{q}_i$ for efficient representation. Next, we initialize each + +2D Gaussian primitive with position $\mathbf{x}_i$ directly from the pointmap, rotation $\mathbf{q}_i$ , scale $\mathbf{s}_i$ set proportionally to the local pointmap density, and opacity $\alpha_i$ set to a default value (e.g. 0.8). Color for the lowest spherical harmonic band (DC component) is initialized from the original image colors projected onto the point, with higher-order coefficients set to zero. This initialization preserves the surface structure from the MASt3R reconstruction while enabling efficient optimization with 2D Gaussian primitives. + +# 3.3. Joint Optimization + +Correspondence Loss Using the dense correspondences provided by MASt3R, and inspired by SPARF [75], we formulate a multi-view correspondence loss that enforces geometric consistency across different viewpoints: + +$$ +\mathcal {L} _ {\text {c o r r}} = w _ {p _ {n}, p _ {m}} \rho \left(p _ {m} - \pi \left(P _ {m} ^ {- 1} P _ {n} \pi^ {- 1} \left(p _ {n}, d _ {n}\right)\right)\right), \tag {5} +$$ + +where $(p_n, p_m)$ is a correspondence pair with $p_n \in [0,1]^2$ a pixel in view $n$ and $p_m$ its corresponding pixel in view $m$ . $w_{p_n,p_m} \in [0,1]$ is the confidence of the correspondence. $\rho$ represents the Huber loss [31] function. $\pi: \mathbb{R}^3 \to \mathbb{R}^2$ is the projection operator mapping 3D points to pixel coordinates. $\pi^{-1}: \mathbb{R}^2 \times \mathbb{R} \to \mathbb{R}^3$ is the back-projection operator mapping pixel coordinates and depth to 3D points. $d_n$ is the 2DGS model splatted depth at pixel $p_n$ . $P_n$ and $P_m$ represent the camera parameters for views $n$ and $m$ . This loss enforces that corresponding pixels across views should project to the same 3D point when back-projected using their respective depths and camera parameters, and thus helps guide our bundle adjustment to a good minima. + +Photometric Loss Following 3DGS and 2DGS, we use a combination of $\mathcal{L}_1$ image reconstruction loss and structural similarity (SSIM) to optimize the rendering of our system: + +$$ +\mathcal {L} _ {\text {p h o t o}} = (1 - \lambda) \cdot \mathcal {L} _ {1} + \lambda \cdot \mathcal {L} _ {\mathrm {S S I M}} + \mathcal {L} _ {\mathrm {r e g}}. \tag {6} +$$ + +Notice that 2DGS comes also with geometric regularization term $\mathcal{L}_{\mathrm{reg}}$ binding pixel depth normals to splatted normals, and reducing depth distortion, as detailed in [29]. + +Variance Regularization Loss A key contribution of our method is the introduction of a splatted color variance regularization loss. The fundamental insight is that volume rendering can be interpreted as computing the expected value of colors sampled along a ray [74]: + +$$ +C = \mathbb {E} _ {\mathbf {t} \sim p (\mathbf {t})} [ c (\mathbf {t}) ] = \sum_ {i = 1} ^ {K} \alpha_ {i} \mathbf {c} _ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {7} +$$ + +where $c(\mathbf{t})$ represents the color of a point $\mathbf{t}$ along the ray, and the expectation is taken with respect to $p(t) = \sigma(t)T(t)$ with $\sigma(t)$ being the density at $t$ and $T(t)$ representing the transmittance function. + +We hypothesize that stabilizing and robustifying our training would lead to more robust and improved reconstructions. In order to robustify the training, we would like to control how the rendered color deviates from the ground truth color when the geometry is perturbed, i.e. given divergence measure $\mathcal{D}$ , we want to minimize the following worst-case loss along the ray: + +$$ +\mathcal {L} _ {r} = \sup _ {q: D (q, p) < \rho} | \mathbb {E} _ {\mathbf {t} \sim q (\mathbf {t})} [ c (\mathbf {t}) ] - C _ {g t} |, \tag {8} +$$ + +where $\rho$ controls the strength of the perturbation. By leveraging theorem 2 in [19], this loss can be upper bounded by: + +$$ +\mathcal {L} _ {1} + \eta \sqrt {\operatorname {V a r} _ {\boldsymbol {t} \sim p (\boldsymbol {t})} [ c (\boldsymbol {t}) ]}. \tag {9} +$$ + +Consequently, we consider minimizing the variance of the color along the ray under the distribution $p(t)$ that can be expressed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {v a r}} = \mathbb {V} a r _ {\mathbf {t} \sim p (\mathbf {t})} [ c (\mathbf {t}) ] (10) \\ = \mathbb {E} _ {\mathbf {t} \sim p (\mathbf {t})} [ (c (\mathbf {t}) - C) ^ {2} ] = \mathbb {E} _ {\mathbf {t} \sim p (\mathbf {t})} [ c (\mathbf {t}) ^ {2} ] - C ^ {2}. (11) \\ \end{array} +$$ + +For efficient implementation, we compute the first term using a modified CUDA kernel of the 2DGS splatting algorithm, where we render the square of colors in addition to the rendering of the colors. This allows for efficient forward and backward passes during optimization. Intuitively, reducing rendered color variance reduces uncertainty in the rendering, which leads to sharper and more multi-view coherent reconstructions. This reflects positively on the qualitative (Fig 5) and quantitative (Tab 4 and Fig 6) performance of our reconstructions. We note that in practice, loss $\mathcal{L}_{\mathrm{var}}$ is averaged across pixels for a given view. + +Training Objective Our total optimization objective combines the three loss terms: + +$$ +\mathcal {L} = \lambda_ {\text {p h o t o}} \mathcal {L} _ {\text {p h o t o}} + \lambda_ {\text {c o r r}} \mathcal {L} _ {\text {c o r r}} + \lambda_ {\text {v a r}} \mathcal {L} _ {\text {v a r}}, \tag {12} +$$ + +where $\lambda_{\mathrm{photo}}$ , $\lambda_{\mathrm{corr}}$ , and $\lambda_{\mathrm{var}}$ are weighting factors that balance the contribution of each term. We optimize all parameters simultaneously using the Adam [41] optimizer with a learning rate scheduling following [29]. Unlike multi-stage approaches [75], our single-stage optimization allows for faster convergence while maintaining high reconstruction quality. The optimization updates both camera extrinsics $\{P\}$ per input view and 2D Gaussian primitive parameters $\{\alpha, \mathbf{c}, \mathbf{q}, \mathbf{s}\}$ according to the gradient of the total loss. + +# 4. Experiments + +We evaluate our approach across 3D reconstruction, novel view synthesis, and camera pose estimation. We outline the experimental setup, including datasets, baseline methods, and implementation details, followed by quantitative and qualitative comparisons. Ablation studies further analyze the impact of key components. + +# 4.1. Datasets and Experimental Configuration + +Our method is tested on benchmark datasets [1, 5, 42, 92, 97]. Further implementation details and additional qualitative results are provided in the supplementary material. + +Few-Shot 3D Reconstruction: We evaluate sparse 3D reconstruction on DTU [1], BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]. For 3-view reconstruction on DTU, we follow SparseNeuS [50], evaluating 15 scenes with two sets of 3-view inputs, reporting the mean per scene. We also provide qualitative comparisons for 3, 6, and 12-view settings on selected scenes from BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]. + +Few-Shot Novel View Synthesis: We conduct novel view synthesis (NVS) on Tanks and Temples [42] (8 scenes), MVImgNet [97] (7 scenes), and MipNeRF360 [5] (9 scenes), following InstantSplat [21, 22]. Results are reported under sparse input settings (3, 6, and 12 views). + +Camera Pose Estimation: Camera poses are evaluated on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5], following InstantSplat [21, 22]. Additionally, for few-shot reconstruction, we assess camera poses on DTU [1] using the same 3-view evaluation protocol as in the reconstruction task. + +Evaluation Metrics. For 3D reconstruction quality, we employ the Absolute Relative Error (Rel) and Normal Consistency (NC). Rel measures the accuracy of depth estimation following DUSt3R [79], while NC evaluates the geometric fidelity by measuring the alignment between predicted and ground truth surface normals. We apply these metrics to our final meshes. We compute ground truth normals from the gradients of the ground truth DTU depth maps. + +For novel view synthesis, we use standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [81], and Learned Perceptual Image Patch Similarity (LPIPS) [102]. We note that for this evaluation we follow InstantSplat[21, 22]: for novel test views without pre-computed camera poses, we perform an additional camera optimization step, which optimizes the test cameras while keeping the 2D Gaussian primitives fixed using the $\mathcal{L}_1$ photometric reconstruction loss. + +For camera pose estimation, we report the Absolute Trajectory Error (ATE) as defined in [8, 21], using COLMAP-computed poses from dense views as ground-truth references. + +Baselines. We compare our method against several state-of-the-art approaches: + +- NVS specific methods: Pose-free methods NoPe-NeRF [8], CF-3DGS [25], NeRFmm [82], SPARF [75] and InstantSplat [21, 22] (in both S and XL configurations). We also include comparisons with 3DGS [40] and + +![](images/9a6a443dd589d0469fd40f93ab0ba1d98180304669f8cc90e02d554da7227f9a.jpg) +Figure 3. Qualitative comparison on DTU dataset from 3 input images. + +FSGS [104], which use pre-computed camera parameters from Colmap [69]. + +- Reconstruction-specific methods: MASt3R [44], Colmap [69], UfoRecon [56], CasMVSNet [26], Spurfies [67], SpaRP [87], SparseCraft [95], InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22]. Of these, similar to our approach, InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22] further refine cameras. + +Implementation Details. Our framework is implemented in PyTorch. We set the optimization iterations to 1k for DTU dataset reconstruction experiments and between 2k-4k for other reconstruction and novel view synthesis experiments. For test-time optimization, camera parameters are jointly optimized for 1k iterations (while keeping the learned Gaussian parameters fixed). We initialize our pipeline with MASt3R [44] configured at a resolution of 512 pixels and use Open3D [103] to compute normals for the point cloud, which are used to initialize rotation parameters of the 2D Gaussians. All experiments are conducted on a single NVIDIA A6000 GPU. We set $\lambda_{\mathrm{photo}} = 1.0$ and $\lambda_{\mathrm{corr}} = 5 \times 10^{-5}$ . The variance loss weight $\lambda_{\mathrm{var}}$ follows a cosine annealing schedule: $\lambda_{\mathrm{var}} = 0.5\left(1 + \cos \left(\frac{\pi \cdot t}{T}\right)\right)$ , where $t$ denotes the current training step and $T$ the total number of steps. This schedule gradually decreases $\lambda_{\mathrm{var}}$ from 1.0 to 0.0, ensuring an adaptive regularization throughout training. + +# 4.2. 3D Reconstruction Evaluation + +For few-shot reconstruction, we follow the evaluation protocol from prior work [50, 56, 95], cleaning the generated meshes with masks from training views. We evaluate the Absolute Relative Error (Rel) and Normal Consistency (NC) between predicted depths normals from the extracted meshes and the ground truth depths normals from the DTU dataset [1]. + +Tab. 1 and 2 present the quantitative results on the DTU + +![](images/6305944af912b06f0aff59928c7261dad0296f90d6452780b12f6d36b2e9ea7e.jpg) +Figure 4. Qualitative comparison on BlendedMVS dataset from 3 input images. + +dataset for Rel and NC metrics, respectively. Fig 3, 4 and 7 show qualitative reconstruction comparisons, with additional qualitative comparisons with SpaRP [87] included in the supplementary material. Additionally, the supplementary material contains extended qualitative results, in the form of video comparisons demonstrating our reconstructions for several datasets. Our method achieves state-of-the-art performance, outperforming all competing approaches across the majority of scenes. + +Several key observations can be made from these results. Pretrained generalizable methods (UfoRecon [56] and CasMVSNet [26]) generalize poorly when faced with noisy initial camera parameters from MASt3R [44], resulting in significantly worse performance. Despite being designed for sparse reconstruction, SparseCraft [95] struggles with the noisy camera poses and point clouds from MASt3R [44]. Spurfies [67], which employs an SDF prior trained on ShapeNet, performs particularly poorly with noisy camera poses, resulting in worse mesh extractions. InstantSplat [21, 22] variants (InstSplatGOF and InstSplat2DGS) perform better than other competing approaches, but still produce overly smooth meshes lacking fine details. + +Our method significantly outperforms these approaches due to several key advantages, including the variance regularization loss (Sec. 3.3) promoting sharper and more detailed reconstructions Fig 5, The correspondence loss (Sec. 3.3) and joint optimization framework effectively refine the initial noisy camera poses from MAsst3R [44]. Our method's ability to preserve fine geometric details is also evidenced by the superior Normal Consistency scores. + +Table 3 presents camera pose estimation results measured by ATE. Our approach achieves significant improve + +
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]2.4920.4812.025.398.866.34.719.017.7010.335.862.505.04.564.957.34
UFORecon [56]35.5557.1538.8135.4750.0244.9434.7053.5641.6747.9340.2239.3439.7533.3449.0442.77
CasMVSNet [26]9.7638.0037.0317.3019.3924.106.5729.5711.8414.9212.388.383.939.8713.4417.10
Spurfies [67]2.9652.1558.4220.2514.6619.244.9712.447.5010.647.366.2638.6125.8444.1921.70
InstSplatGOF [21]9.5423.1413.7415.6213.3012.3311.1912.2114.6915.6113.6112.3015.3112.9012.0513.84
Sparsecraft [95]2.9610.517.162.928.179.443.836.197.616.444.756.435.367.658.146.50
InstantSplat2DGS [22]2.4221.749.083.086.773.603.635.245.827.124.482.783.623.113.435.73
Ours2.4212.3137.93.156.053.5263.484.985.296.224.392.413.463.323.384.82
+ +Table 1. 3-view reconstruction results on DTU dataset The relative error (rel↓) is reported for 15 scans. Best results are bold. + +
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122\( \uparrow \) Mean
MASt3R [44]0.8880.6950.8440.8740.8370.8390.8510.7540.8240.7450.8710.8750.8840.8310.8460.830
UFORecon [56]0.3620.2780.5060.4690.2770.3720.3830.3610.3620.3380.4240.3660.3320.4020.3390.371
CasMVSNet [26]0.7180.5020.6190.7560.6930.6830.7400.5640.7040.6810.7850.7950.8590.7780.7550.709
Spurfies [67]0.8630.350.2970.7110.7420.6680.7710.6710.7770.6830.780.7710.5090.5720.4330.640
InstSplatGOF [21]0.7740.5810.6890.6960.7460.7500.7100.6580.6820.6530.6950.7240.7100.6630.6500.692
Sparsecraft [95]0.8410.6790.8270.8640.7280.7140.8340.7560.7460.7650.8370.6290.8010.7610.7350.768
InstSplat2DGS [22]0.8820.6160.860.8990.8340.8650.8450.7830.8100.7730.8820.8490.8950.8270.8550.832
Ours0.9040.750.8770.8960.8660.8870.8670.7990.8450.7890.8880.8740.9020.850.8650.857
+ +Table 2. 3-view reconstruction results on DTU dataset Normal consistency (NC↑) is reported for 15 scans. Best results are bold. + +ments in camera pose estimation, with a $13.1\%$ reduction in error compared to the best-performing baseline (InstSplatGOF [21]). This demonstrates the effectiveness of our correspondence loss in refining the initial noisy camera poses from MAST3R [44]. + +# 4.3. Novel View Synthesis Evaluation + +We evaluate our method's capability for novel view synthesis on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5] datasets. Quantitative and qualitative result comparisons on all datasets are included in the supplementary material. + +# 4.4. Ablation Studies + +We conduct ablation studies to analyze the contribution of each component in our method, including an experiment to study the impact of varying the number of input views. + +Component Ablation These experiments were performed using 3 input unposed views on 5 testing scenes from the DTU dataset using two sets of views, following the protocols established in [50, 56]. Tab 4 shows the impact of incorporating different components into our pipeline. Fig 5 shows a qualitative ablation comparison. Camera optimization using bundle adjustment improves all metrics. Incorporating the correspondence loss further enhances this alignment, resulting in improved reconstructions. Splat orientation initialization with pointmap normals improves results further. Notice that our variance loss contributes to both the final metrics and the qualitative aspect of our results. The reconstructions with variance loss exhibit significantly sharper details and better preservation of fine structures, validating the quantitative improvements observed in our metrics. + +Number of Views Ablation We progressively increase the number of views and perform pose-free reconstruction with our method using scan37 from DTU. We evaluate by computing the mean metrics across both test view sets. As + +![](images/11ed57c145d7816deeda225b27f0acd20dad063b7773d2750cabe523f0300f74.jpg) +Figure 5. Qualitative ablation of the component of our method, showing also the effect of variance loss. Left blue box: w/o $\mathcal{L}_{\mathrm{var}}$ . Right red box: w $\mathcal{L}_{\mathrm{var}}$ . (3 input images). + +shown in Figure 6, our method benefits favorably from increased views. The benefits of all components of our method are also maintained, if not increased. This includes our variance loss. + +# 5. Running time + +In our reconstruction experiments, we observed that the mean training time on DTU [1] for 1000 iterations (similar to InstantSplat-XL [21, 22]) was roughly 146s on a single Nvidia A6000 GPU. For a fair comparison, we also trained InstantSplat-XL [21, 22] on the same GPU and observed that the mean training time was around 128s. Hence, our training time is in the same range as InstantSplat [21, 22] while offering significantly higher reconstruction and novel-view quality. + +
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]0.02020.06890.00940.02030.01710.03920.02910.02940.02710.03260.01670.02100.02890.01920.02420.0269
InstSplatGOF [21]0.01800.05360.00890.01360.00490.02730.02630.02390.01800.03230.01280.02380.02490.01750.01490.0214
InstSplat2DGS [22]0.01920.08950.00900.01540.01230.03380.02940.02720.02420.03220.01660.02100.02740.01630.01750.0261
Ours0.00630.04500.00690.01240.01750.02940.01980.02830.01460.02730.00820.01790.02420.01120.01010.0186
+ +Table 3. Results on DTU dataset Results include ATE $\downarrow$ metrics for 15 scans. Best results are bold. + +
Methodsrel ↓NC ↑ATE ↓
2DGS7.930.7850.0327
BA-2DGS7.790.7880.0284
BA-2DGS+Lcorr7.410.8040.0256
BA-2DGS+Lcorr+Norm. init.7.030.8290.0237
BA-2DGS+Lcorr+Norm. init.+Lvar6.790.8340.0238
+ +Table 4. Ablation on DTU dataset, adding one component at a time. Results include rel and NC metrics. Best results are bold. + +![](images/47137a74a20088ba55944074e10e22f89f5ccd9f60fee40815dd7b750f92d2a6.jpg) +Figure 6. Effect of increasing the number of input views on different metrics. + +![](images/f13245cd30b3c8b573694d5338628d5c0380c6c9098ed2208f176a1322be650b.jpg) + +# 6. Limitations and Future Work + +Despite the promising results, several limitations remain, which present opportunities for future research. As foundation models keep on improving, better initializations could be available in the future. Furthermore, while our approach handles a variety of scenes across different datasets, performance may be limited for highly complex geometry, very thin structures, or semi-transparent objects (observed for the glass bottle in Fig 7 room scene in MipNeRF360 [5]), which are common challenges for surface based methods especially in the sparse setting. Additionally, our method assumes a static scene under consistent lighting. Extending it to more dynamic scenarios is one of the challenges we are looking forward to exploring. + +# 7. Conclusion + +We addressed the underexplored problem of shape reconstruction from sparse uncalibrated multi-view images. Our method leverages foundation models for initial camera and geometry estimation, followed by joint optimization of 2D Gaussian primitives and camera parameters. The introduction of a variance regularization loss significantly improves the detail and sharpness of reconstructions, while our correspondence matching based framework enhances + +![](images/e4d4b23f1482015766d47d7a0e2f26babd5824cb44868f579f50cc30f2f0f0d3.jpg) +Figure 7. Qualitative comparison on MVImgNet and MipNeRF360 datasets from 6 and 12 input images, respectively. + +camera pose accuracy. Extensive evaluations on multiple datasets demonstrate that our approach consistently outperforms state-of-the-art methods in reconstruction quality, novel view synthesis, and camera pose estimation. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 2, 5, 6, 7 +[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2 +[3] Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1034-1040. IEEE, 1997. 2 +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460-5469, 2022. 2 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 7, 8 +[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[7] Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, and Philip Torr. Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449, 2023. 2, 3 +[8] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 3, 5 +[9] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3 +[10] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3 +[11] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting + +from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 3 +[13] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4182-4194, 2023. 2 +[14] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312-18321, 2023. 3 +[15] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In European Conference on Computer Vision, pages 264-280. Springer, 2022. 3 +[16] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2 +[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 2 +[18] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1 +[19] John Duchi, Peter Glynn, and Hongseok Namkoong. Statistics of robust optimization: A generalized empirical likelihood approach, 2018. 5 +[20] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2 +[21] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2(3):4, 2024. 2, 3, 5, 6, 7, 8 +[22] Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. arXiv preprint arXiv:2403.20309, 2024. 2, 3, 5, 6, 7, 8 +[23] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. arXiv preprint arXiv:2410.18956, 2024. 2 +[24] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024. 2 + +[25] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatt-ing. arXiv preprint arXiv:2312.07504, 2023. 3, 5 +[26] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 6, 7 +[27] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jisang Han, Jiaolong Yang, Chong Luo, and Seungryong Kim. Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv preprint arXiv:2410.22128, 2024. 2 +[28] Berthold KP Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. 1970. 1 +[29] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5 +[30] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 3 +[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 4 +[32] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2 +[33] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Neural mesh-based graphics. In European Conference on Computer Vision, pages 739-757. Springer, 2022. 2 +[34] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Geotransfer: Generalizable few-shot multi-view reconstruction via transfer learning. arXiv preprint arXiv:2408.14724, 2024. 3 +[35] Shubhendu Jena, Shishir Reddy Vutukur, and Adnane Boukhayma. Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting. arXiv, 2025. 3 +[36] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5846-5854, 2021. 3 +[37] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2 +[38] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2 +[39] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2 + +[40] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 5 +[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[42] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2, 5, 7 +[43] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 1 +[44] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 6, 7, 8 +[45] Qian Li, Franck Moulton, and Adnane Boukhayma. Learning generalizable light field networks from few images. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3 +[46] Qian Li, Franck Multon, and Adnane Boukhayma. Regularizing neural radiance fields from sparse rgb-d inputs. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2320-2324. IEEE, 2023. 2 +[47] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3 +[48] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 3 +[49] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Fast generalizable gaussian splattering reconstruction from multi-view stereo. arXiv preprint arXiv:2405.12218, 2024.3 +[50] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 3, 5, 6, 7 +[51] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[52] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6326-6337, 2022. 3 +[53] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view + +synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3 +[54] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2 +[55] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[56] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 3, 6, 7 +[57] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2 +[58] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 2 +[59] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2 +[60] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 3 +[61] Amine Ouasfi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. 3DV, 2024. 3 +[62] Amine Ouasfi and Adnane Boukhayma. Few-shot unsupervised implicit neural shape representation learning with spatial adversaries. arXiv preprint arXiv:2408.15114, 2024. 2 +[63] Amine Ouasfi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. Advances in Neural Information Processing Systems, 36, 2024. 3 +[64] Amine Ouasfi and Adnane Boukhayma. Unsupervised occupancy learning from sparse point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21729-21739, 2024. 2 +[65] Amine Ouasfi, Shubhendu Jena, Eric Marchand, and Adnane Boukhayma. Toward robust neural reconstruction from sparse point sets. arXiv preprint arXiv:2412.16361, 2024. 2 + +[66] Eric Penner and Li Zhang. Soft 3D Reconstruction for View Synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 2 +[67] Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Spurfies: Sparse surface reconstruction using local geometry priors. arXiv preprint arXiv:2408.16544, 2024. 1, 3, 6, 7 +[68] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 3 +[69] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 6 +[70] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1 +[71] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1 +[72] Steven M Seitz and Charles R Dyer. Photorealistic Scene Reconstruction by Voxel Coloring, 2002. US Patent 6,363,170. 2 +[73] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2 +[74] Andrea Tagliaasacchi and Ben Mildenhall. Volume rendering digest (for nerf), 2022. 4 +[75] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2, 3, 4, 5 +[76] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. arXiv preprint arXiv:2303.16196, 2023. 1, 2 +[77] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2 +[78] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2 +[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 5 + +[80] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35:1966-1978, 2022. 1 +[81] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[82] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. 2021. 3, 5 +[83] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 1 +[84] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2 +[85] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1 +[86] Haolin Xiong, Sairishek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 3 +[87] Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. In European Conference on Computer Vision, pages 143-163. Springer, 2024. 1, 6 +[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 2 +[89] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2 +[90] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2 +[91] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2 +[92] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 1, 2, 5 + +[93] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2 +[94] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2 +[95] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6, 7 +[96] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2 +[97] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 1, 2, 5, 7 +[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 1 +[99] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[100] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2 +[101] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2 +[102] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[103] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 3, 6 +[104] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 1, 3, 6 \ No newline at end of file diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/images.zip b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed507d32b7073556cf33a4cacd74dcf65f579645 --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d620005034c25b071aba50ebdfe8f20b1f9e45108c7431dc1ae544b68b4f78bd +size 684623 diff --git a/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/layout.json b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ba8fb91618323af34223f01b85368428fcc465dd --- /dev/null +++ b/2025/Sparfels_ Fast Reconstruction from Sparse Unposed Imagery/layout.json @@ -0,0 +1,10721 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 63, + 494, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 63, + 494, + 82 + ], + "spans": [ + { + "bbox": [ + 116, + 63, + 494, + 82 + ], + "type": "text", + "content": "Sparfels: Fast Reconstruction from Sparse Unposed Imagery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 103, + 473, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 473, + 131 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 473, + 131 + ], + "type": "text", + "content": "Shubhendu Jena*, Amine Ousafi*, Mae Younes, Adnane Boukhayma Inria, Univ. Rennes, CNRS, IRISA" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 134, + 555, + 249 + ], + "blocks": [ + { + "bbox": [ + 57, + 134, + 555, + 249 + ], + "lines": [ + { + "bbox": [ + 57, + 134, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 134, + 555, + 249 + ], + "type": "image", + "image_path": "2e4c4da61c26933d1d0060d5a82578777048f050ccd480691d2efac615dd4c02.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 273 + ], + "type": "text", + "content": "Figure 1. Examples of reconstruction meshes obtained within 3 minutes from sparse pose-free images via our method. We used here 6 (left) and 3 (right) input images respectively from scenes of datasets MVImgNet [97] and BlendedMVS [92]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 281, + 200, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 281, + 200, + 293 + ], + "spans": [ + { + "bbox": [ + 151, + 281, + 200, + 293 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 307, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 297, + 582 + ], + "type": "text", + "content": "We present a method for Sparse view reconstruction with surface element splatting that runs within 3 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate state-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view benchmarks based on established multi-view datasets. Code will be made available at https://shubhendujena.github.io/Sparfels-web/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "3D reconstruction is a long standing problem at the intersection of computer vision, graphics and machine learning. It is core to a myriad of downstream applications in virtual and augmented reality, autonomous driving, medical imaging, film and animation, to name a few. Classical optimization based techniques such as structure from motion and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 282, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 437 + ], + "type": "text", + "content": "multi-view stereo [70, 71], shape from visual cues such as shading or silhouette [28, 43], photometric stereo [83] etc., have been popular for decades in this regard. More recently, deep learning has been progressively disrupting 3D vision and top-down approaches to vision more broadly. Among the latest additions, large vision and foundation models (e.g. DUSt3R [79], MASt3R [44]) can offer robustness and fast feed-forward inference compared to classical photogrammetry [70, 71]. Given a few raw uncalibrated images of a scene, they are capable of inferring reasonable relative poses and a competitive coarse 3D structure. These predictions can lack precision and details, though, and cannot enable stand-alone novel view synthesis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 676 + ], + "type": "text", + "content": "Furthermore, state-of-the-art data-prior free scene specific learning methods for reconstruction can deliver highly detailed geometry and impressive novel view synthesis capabilities from dense calibrated image grids. While implicit and some grid shape representations (e.g. NeuS, VolSDF, HF-NeuS, Voxurf [78, 80, 85, 94]) trained via differentiable volumetric rendering (NeRF [55]) have been powerful tools for this task for a while, Gaussian Splatting [40] based alternatives emerged with comparable novel view and reconstruction [29, 100] quality, within a fraction of the training time, and with orders of magnitude higher interactive rendering frame rates. These methods typically fail however under uncalibrated and sparse training images. Contemporary literature attempts to hedge against the underconstrained nature of the sparse view problem through regularization with a multitude of deep geometric priors (monocular depth [18, 76, 104], normal [98], point cloud [67] etc.), or training new priors from scratch [50, 67, 87]. Such strategies can prove to be cumbersome, impractical or costly in terms of data, memory or compute." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose to marry the best of both worlds efficiently — namely pre-existing large 3D vision feedforward models (MASt3R) and fast test-time optimization" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27476" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "(2DGS) — with the aim of addressing the underexplored problem of 3D geometric reconstruction from sparse un-posed images within minutes (less than 3 on average in our experiments) at minimal cost." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 312 + ], + "type": "text", + "content": "We propose to exploit a single pre-existing foundation model (MASt3R [44]) to the fullest to bootstrap our reconstruction, without needing to load or deploy any additional deep neural networks. Given a few sparse color images, MASt3R delivers a coarse point cloud and initial camera poses. We use these to initialize a bundle adjusting 2D Gaussian Splatting (2DGS) optimization via differentiable render-and-compare, where surface elements are instantiated at the point samples, oriented initially according to their normals. Differently from the closest competition [21, 22], we harness the MASt3R [44] produced image pair correspondences to guide camera optimization during 2DGS training. This is achieved via a splatted depth reprojection error loss. The resulting camera alignment improvement proves crucial for obtaining accurate shape representations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 411 + ], + "type": "text", + "content": "By analyzing the surface element splatting from a statistical moment perspective, rendered pixels can be seen as color expectation over fragments following the ray termination probability. Based on this, we derive an expression of splatted color variance along the ray, and propose a novel loss tasked with reducing it. This results in improved reconstruction metrics and qualitative sharper mesh details as witnessed by our ablation studies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 496 + ], + "type": "text", + "content": "The integration of these elements yields state-of-art performance in sparse pose-free 3D reconstruction, obtained in very short times on mid-range GPUs. We evaluate our method using established reconstruction and novel view datasets [1, 5, 42, 92, 97], showcasing both superior geometric reconstruction, camera pose estimation and novel view synthesis w.r.t. comparable state-of-the-art baselines." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Evaluation of 3D shape reconstruction in the posed setting relies traditionally on mesh Chamfer distance to ground-truth, as used in multi-view reconstruction benchmarks like DTU [1]. This is a mesh pose and scale sensitive metric. In the unposed setting, reconstructions are obtained naturally up-to a rigid transformation. RANSAC powered alignments between predicted and ground-truth cameras have been shown to be acceptable for pre-evaluation reconstruction mesh alignment in the dense view setting [7], and good enough for camera pose evaluation for sparse sets [75]. However, we found these camera based alignments to be too inaccurate in the sparse setting to enable reliable mesh pre-alignments for Chamfer distance evaluation. Hence, inspired by the rigid-transform invariant evaluation in [79], we compare the reconstructed mesh depth and normal rasterizations to the groundtruth in screen space in our evaluation, and show considerable improvements compared to the competition in the standard DTU benchmark" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 555, + 139 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 555, + 139 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 555, + 139 + ], + "type": "image", + "image_path": "60af467a4cb250858e54d2498e0e32ad7d04c0446ce61d2332cd7bc5c79e19ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 148, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 236 + ], + "type": "text", + "content": "Figure 2. 3D Foundation models meet differentiable perspective accurate Splating for fast reconstruction from pose-free sparse images. We bootstrap a bundle adjusting 2DGS with MAST3R [44] 3D point clouds and initial camera predictions. Besides standard 2DGS photometric reconstruction and geometric regularization losses, we leverage a camera reprojection error loss using MAST3R's correspondences and a novel splatted color variance reduction loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 248, + 554, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 554, + 285 + ], + "type": "text", + "content": "(Tab.1,2,3). Further qualitative comparison in Blended-MVS [92], MVImgNet [97], Mip-NeRF 360 [4] illustrate the quality of our reconstructions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 292, + 399, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 292, + 399, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 292, + 399, + 304 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 577 + ], + "type": "text", + "content": "NVS and 3D Reconstruction Novel view synthesis (NVS) aims to generate unseen views from input images [3, 54]. Neural Radiance Fields (NeRF) [55] revolutionized this field by representing 3D scenes as continuous implicit functions, though challenges remain in efficiency and quality [4-6, 66, 72, 73]. Inspired by the success of other explicit representations (e.g. meshes [33, 38, 77] and point clouds [2, 20]), recent advancements introduced primitive-based scene representations [13, 40, 90], notably 3D Gaussian Splitting (3DGS) [40], which uses anisotropic 3D Gaussians with differentiable splatting for efficient high-quality reconstruction. This has sparked extensions for surface reconstruction [29, 100, 101], neural representations [51], multi-resolution modeling [24, 99], and feedforward methods [10, 12, 23, 27, 89]. In parallel, implicit representations model 3D geometry as continuous functions [16, 37, 39, 57, 59]. Methods like IDR [93], NeuS [78], and VolSDF [94] enhance geometric accuracy but often suffer from slow rendering. In contrast, 2D Gaussian Splitting (2DGS) [29] offers real-time rendering while improving view consistency for surface reconstruction through planar Gaussian primitives." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Sparse-View NVs and 3D Reconstruction Traditional NeRFs and GS methods require carefully captured images and SfM preprocessing with COLMAP [69], which could limit practical applications. The few-shot setup has been addressed for both reconstruction from Point Cloud [11, 62, 64, 65] and images. Various approaches address the challenge of limited views through regularization techniques, including depth supervision [17, 46, 58, 76], foundation model supervision [32, 84, 88], specialized architectures [91, 96], while SparseCraft [95] uses learning-free" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27477" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "MVS cues for SDF regularization. For Gaussian Splating, methods like FSGS [104] and SparseGS [86] incorporate external priors. Generalizable methods address sparse-view challenges by pretraining on large datasets, enabling feed-forward inference. Inputs include images (e.g. [10, 12, 35, 45, 49]) and 3D Points (e.g. [9, 30, 52, 60, 61, 63]). For surface reconstruction from color images, approaches like SparseNeuS [50], VolRecon [68], ReTR [47], UfoRecon [56], GeoTransfer [34], and Spurfies [67] leverage various mechanisms such as using image priors to form multiview cost volumes, using transformer architectures to enhance reconstruction from limited views, as well as using reconstruction priors trained on sparse point clouds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 495 + ], + "type": "text", + "content": "Pose-Free NVS and 3D Reconstruction Pose-free methods eliminate the need for accurate camera calibration. NeRFmm [82], BARF [48], SCNeRF [36], and GARF [15] jointly optimize camera parameters and scene representations. SPARF [75] and PoRF [7] address optimization with noisy poses, though often requiring lengthy training. Recent work like Nope-NeRF [8], Lu-NeRF [14], LocalRF [53], and CF-3DGS [25] leverage depth information but typically assume dense video sequences. InstantSplat [21, 22] addresses these limitations by using MASt3R [44] for initialization and jointly optimizing camera poses and 3D models. Our work builds on these foundations but focuses specifically on using 2D Gaussian Splatting for efficient 3D surface reconstruction from sparse, uncalibrated views. We initialize with MASt3R [44] and refine through joint optimization of Gaussian parameters and camera poses, incorporating a novel variance-based color loss for improved detail preservation. Unlike previous approaches requiring dense sequences, known intrinsic, or long optimization times, our method efficiently handles sparse, uncalibrated images while achieving high-quality reconstructions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 500, + 112, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 112, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 112, + 513 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "text", + "content": "Given a few unposed color images " + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\{I^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "text", + "content": ", our goal is to recover a 3D triangle mesh " + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "text", + "content": " and render novel views of the observed scene efficiently. We achieve this by fitting a bundle adjusting 2DGS [29] model to the images of the scene, bootstrapped by point map and correspondence feedforward predictions from a 3D foundation model (MASt3R [44]) and carefully designed regularizations. At test time, mesh " + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 520, + 296, + 651 + ], + "type": "text", + "content": " can be extracted using the TSDF algorithm in Open3D [103] using the converged 2DGS model rendered depth, as recommended in [29]. We detail the workings of our method below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 142, + 672 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "2D Gaussian Splatting Our method builds upon 2D Gaussian Splatting (2DGS) [29], which represents 3D scenes using planar disk primitives defined within local tan-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "gent planes. Unlike volumetric 3D Gaussians [40], 2DGS is a surface-based representation that improves multi-view consistency. Each 2D Gaussian is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 363, + 118, + 555, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 118, + 555, + 146 + ], + "spans": [ + { + "bbox": [ + 363, + 118, + 555, + 146 + ], + "type": "interline_equation", + "content": "G (\\mathbf {p}) = \\exp \\left(- \\frac {u (\\mathbf {p}) ^ {2} + v (\\mathbf {p}) ^ {2}}{2}\\right) \\tag {1}", + "image_path": "3da00862a4fe0e2d95d67e5b56afb9132a2958ee510fc1fc5dec6544ea238255.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "u(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "v(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " are local UV space coordinates of a pixel " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " in the primitive's tangent plane. This ray-splat intersection can be expressed as a function of the camera parameters and splat attributes differentiably. Each primitive is parameterized by the center of the Gaussian in world coordinates " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": ", a quaternion representing the orientation of the primitive " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": ", scale factors for the two dimensions in the local UV tangent plane " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " controlling the transparency of the primitive, and Spherical harmonic coefficients " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{sh}k} \\in \\mathbb{R}^3 | k = 1, \\dots, n" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " encoding view-dependent appearance. Rotation " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " defines a local coordinate frame " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "[\\mathbf{u}, \\mathbf{v}, \\mathbf{n}]" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " being the normal vector orthogonal to the primitive's plane, and " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 152, + 555, + 320 + ], + "type": "text", + "content": " spanning the tangent plane." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": "For rendering, 2DGS uses alpha-compositing for volumetric rendering similarly to 3DGS. Given a set of primitives along a ray sorted from front to back " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "G_{1}, G_{2}, \\ldots, G_{K}" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": ", the final color " + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 320, + 554, + 369 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 377, + 555, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 377, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 370, + 377, + 555, + 411 + ], + "type": "interline_equation", + "content": "C (\\mathbf {p}) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} ^ {\\prime} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j} ^ {\\prime}\\right), \\tag {2}", + "image_path": "0f230e719e8cda30ad6be91463d2e4ee18fad355b2a1dea2af232ce1045c8737.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{\\prime} = \\alpha_{i}G(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "text", + "content": " is the modified opacity of the " + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "text", + "content": "-th primitive and " + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 422, + 554, + 459 + ], + "type": "text", + "content": " is its color, computed from the spherical harmonic coefficients based on the viewing direction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 535 + ], + "type": "text", + "content": "MASt3R MASt3R [44] is a foundation model for 3D vision that extends DUSt3R [79] with more powerful feature extraction and matching capabilities. It tackles the multi-view stereo reconstruction task without requiring prior camera calibration or poses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "content": "Given a pair of images " + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "inline_equation", + "content": "I^1, I^2 \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "content": ", MAS3R outputs two corresponding pointmaps " + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{1,1}, \\mathbf{P}_{2,1} \\in \\mathbb{R}^{W \\times H \\times 3}" + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "content": " with associated confidence maps " + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{1,1}, \\mathbf{O}_{2,1} \\in \\mathbb{R}^{W \\times H}" + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "content": ". Each pointmap assigns a 3D point to every pixel in the image. Importantly, both pointmaps are expressed in the same coordinate frame of the first image " + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "inline_equation", + "content": "I^1" + }, + { + "bbox": [ + 313, + 535, + 554, + 607 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 608, + 554, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 554, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 554, + 680 + ], + "type": "text", + "content": "The architecture of MASt3R consists of a Siamese ViT encoder that processes both images independently, Transformer decoders that exchange information via crossattention, and regression heads that output pointmaps and confidence maps. MASt3R is trained using a confidence-weighted regression loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 689, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 689, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n f}} = \\sum_ {v \\in 1, 2} \\sum_ {i} \\mathbf {O} _ {v, 1} ^ {i} \\mathcal {L} _ {\\text {r e g}} (v, i) - \\alpha \\log \\mathbf {O} _ {v, 1} ^ {i}, \\tag {3}", + "image_path": "e15ecda8867f6dc31413d10b85664432f60b14b1f236e102204ec8bc71d5c433.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27478" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}(v,i)" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " measures the Euclidean distance between predicted and ground-truth 3D points for pixel " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " is a regularization parameter." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 167 + ], + "type": "text", + "content": "Besides pointmaps, MASt3R also provides dense pixelwise correspondences between the input views. These correspondences establish reliable geometric constraints across different viewpoints, which we leverage in our optimization framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 167, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 167, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 167, + 186 + ], + "type": "text", + "content": "3.2. Scene Initialization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "content": "Global Geometry Alignment Given a set of input images " + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "inline_equation", + "content": "I^1, I^2, \\ldots, I^N" + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "content": ", we first construct a connectivity graph " + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{V}, \\mathcal{E})" + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "content": " where each vertex " + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "content": " represents an image, and each edge " + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "inline_equation", + "content": "e = (n, m) \\in \\mathcal{E}" + }, + { + "bbox": [ + 55, + 191, + 295, + 299 + ], + "type": "text", + "content": " indicates a pair of images with shared visual content. For each edge in the graph, we run MASt3R to obtain pairwise pointmaps and correspondences. To align these pairwise predictions into a globally consistent reconstruction, we follow the optimization approach from MASt3R:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 305, + 294, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 305, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 93, + 305, + 294, + 338 + ], + "type": "interline_equation", + "content": "\\min _ {\\chi , P, \\sigma} \\sum_ {e \\in \\mathcal {E}} \\sum_ {v \\in e} \\sum_ {i = 1} ^ {H W} \\mathbf {O} _ {v, e} ^ {i} \\left| \\chi_ {i} ^ {v} - \\sigma_ {e} P _ {e} \\mathbf {P} _ {v, e} ^ {i} \\right|, \\tag {4}", + "image_path": "085793a4aa100e361f21805000cc85e4aa74ae4b70a9b52b2a0f65e56807f868.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\chi_i^v" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " is the globally aligned pointmap for view " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "P_e \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " the rigid transformation for edge " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\sigma_e > 0" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " is a scaling factor for edge " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{v,e}^i" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " the confidence value for pixel " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " for edge " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": ", and finally " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{v,e}^i" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " the original pointmap from MASt3R. This optimization produces globally aligned pointmaps " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\{\\chi^{v*}\\}_{v=1}^N \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": " and camera parameters (intrinsics and extrinsics for each view " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": "). We stabilize the estimated focal lengths by averaging them across all training views: " + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\bar{f} = \\frac{1}{N} \\sum_{v=1}^N f_i^*" + }, + { + "bbox": [ + 55, + 345, + 296, + 466 + ], + "type": "text", + "content": ", assuming a single-camera setup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": "Primitive Initialization We leverage the merged globally aligned pointmap to initialize 2D Gaussian primitives with suitable orientation, which is crucial for accurately representing surface geometry. For each point " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " in the globally aligned pointmap, we compute a local surface normal using Principal Component Analysis (PCA). First, we determine the " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": "-nearest neighbors " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_j\\}_{j=1}^k" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " and compute the covariance matrix " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = \\frac{1}{k}\\sum_{j=1}^k (\\mathbf{p}_j - \\bar{\\mathbf{p}})(\\mathbf{p}_j - \\bar{\\mathbf{p}})^T" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{p}}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " is the centroid of the neighborhood. We then perform eigendecomposition on " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " to obtain the eigenvectors " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_1, \\mathbf{v}_2, \\mathbf{v}_3\\}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " and corresponding eigenvalues " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\lambda_1, \\lambda_2, \\lambda_3\\}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ", sorted such that " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_1 \\geq \\lambda_2 \\geq \\lambda_3" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ". The eigenvector " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_3" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ", associated with the smallest eigenvalue, is chosen as the normal direction, i.e., " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = \\mathbf{v}_3" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ". To construct an orthonormal basis, we initialize an arbitrary vector " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " orthogonal to " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = (n_x, n_y, n_z)" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = (-n_y, n_x, 0)^T" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ". The tangent vector is then computed as " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\frac{\\mathbf{n} \\times \\mathbf{a}}{\\|\\mathbf{n} \\times \\mathbf{a}\\|}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " and the bitangent as " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\mathbf{n} \\times \\mathbf{u}" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": ". Finally, the local reference frame " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbf{u}, \\mathbf{v}, \\mathbf{n}]" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " is used to define a rotation matrix, which can be converted to a quaternion " + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 55, + 469, + 296, + 713 + ], + "type": "text", + "content": " for efficient representation. Next, we initialize each" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "2D Gaussian primitive with position " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " directly from the pointmap, rotation " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " set proportionally to the local pointmap density, and opacity " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " set to a default value (e.g. 0.8). Color for the lowest spherical harmonic band (DC component) is initialized from the original image colors projected onto the point, with higher-order coefficients set to zero. This initialization preserves the surface structure from the MASt3R reconstruction while enabling efficient optimization with 2D Gaussian primitives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 425, + 199 + ], + "type": "text", + "content": "3.3. Joint Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 204, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 252 + ], + "type": "text", + "content": "Correspondence Loss Using the dense correspondences provided by MASt3R, and inspired by SPARF [75], we formulate a multi-view correspondence loss that enforces geometric consistency across different viewpoints:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 258, + 553, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 258, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 324, + 258, + 553, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o r r}} = w _ {p _ {n}, p _ {m}} \\rho \\left(p _ {m} - \\pi \\left(P _ {m} ^ {- 1} P _ {n} \\pi^ {- 1} \\left(p _ {n}, d _ {n}\\right)\\right)\\right), \\tag {5}", + "image_path": "3d11e30266c82a0c34715451b88223c3b208d3f24da54def7d0651a57f05c22d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "(p_n, p_m)" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " is a correspondence pair with " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "p_n \\in [0,1]^2" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " a pixel in view " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "p_m" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " its corresponding pixel in view " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "w_{p_n,p_m} \\in [0,1]" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " is the confidence of the correspondence. " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " represents the Huber loss [31] function. " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\pi: \\mathbb{R}^3 \\to \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " is the projection operator mapping 3D points to pixel coordinates. " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\pi^{-1}: \\mathbb{R}^2 \\times \\mathbb{R} \\to \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " is the back-projection operator mapping pixel coordinates and depth to 3D points. " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "d_n" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " is the 2DGS model splatted depth at pixel " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "p_n" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "P_n" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "P_m" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " represent the camera parameters for views " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 281, + 555, + 448 + ], + "type": "text", + "content": ". This loss enforces that corresponding pixels across views should project to the same 3D point when back-projected using their respective depths and camera parameters, and thus helps guide our bundle adjustment to a good minima." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 488 + ], + "type": "text", + "content": "Photometric Loss Following 3DGS and 2DGS, we use a combination of " + }, + { + "bbox": [ + 313, + 452, + 554, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 452, + 554, + 488 + ], + "type": "text", + "content": " image reconstruction loss and structural similarity (SSIM) to optimize the rendering of our system:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 496, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 496, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 349, + 496, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p h o t o}} = (1 - \\lambda) \\cdot \\mathcal {L} _ {1} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {S S I M}} + \\mathcal {L} _ {\\mathrm {r e g}}. \\tag {6}", + "image_path": "7b783f4c9bb8744a585dac8e264444afa9efd9dd880d28459e163e4f357368d0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 553, + 553 + ], + "type": "text", + "content": "Notice that 2DGS comes also with geometric regularization term " + }, + { + "bbox": [ + 313, + 517, + 553, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 517, + 553, + 553 + ], + "type": "text", + "content": " binding pixel depth normals to splatted normals, and reducing depth distortion, as detailed in [29]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": "Variance Regularization Loss A key contribution of our method is the introduction of a splatted color variance regularization loss. The fundamental insight is that volume rendering can be interpreted as computing the expected value of colors sampled along a ray [74]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 624, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 624, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 344, + 624, + 553, + 658 + ], + "type": "interline_equation", + "content": "C = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathbf {c} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {7}", + "image_path": "3571bbe3b0a7736f22b23be4f81260e23eca2de071a971056cb0b589dc7afc4c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "c(\\mathbf{t})" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " represents the color of a point " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " along the ray, and the expectation is taken with respect to " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "p(t) = \\sigma(t)T(t)" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " being the density at " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "inline_equation", + "content": "T(t)" + }, + { + "bbox": [ + 313, + 665, + 553, + 712 + ], + "type": "text", + "content": " representing the transmittance function." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27479" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "We hypothesize that stabilizing and robustifying our training would lead to more robust and improved reconstructions. In order to robustify the training, we would like to control how the rendered color deviates from the ground truth color when the geometry is perturbed, i.e. given divergence measure " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", we want to minimize the following worst-case loss along the ray:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 161, + 296, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 161, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 96, + 161, + 296, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\sup _ {q: D (q, p) < \\rho} | \\mathbb {E} _ {\\mathbf {t} \\sim q (\\mathbf {t})} [ c (\\mathbf {t}) ] - C _ {g t} |, \\tag {8}", + "image_path": "9f8468b26408ce5aeb0b0494f739c5b9fdb6ad64bea21ea7bfc76de5e8922441.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 211 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 186, + 296, + 211 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 186, + 296, + 211 + ], + "type": "text", + "content": " controls the strength of the perturbation. By leveraging theorem 2 in [19], this loss can be upper bounded by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 217, + 296, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 217, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 123, + 217, + 296, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} + \\eta \\sqrt {\\operatorname {V a r} _ {\\boldsymbol {t} \\sim p (\\boldsymbol {t})} [ c (\\boldsymbol {t}) ]}. \\tag {9}", + "image_path": "ebe19859198cd40d00a768319a17575ab82d8fcf438241c8ca53b2453b2c21d3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "content": "Consequently, we consider minimizing the variance of the color along the ray under the distribution " + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "inline_equation", + "content": "p(t)" + }, + { + "bbox": [ + 55, + 239, + 296, + 275 + ], + "type": "text", + "content": " that can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 281, + 295, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 281, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 67, + 281, + 295, + 322 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {v a r}} = \\mathbb {V} a r _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ] (10) \\\\ = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ (c (\\mathbf {t}) - C) ^ {2} ] = \\mathbb {E} _ {\\mathbf {t} \\sim p (\\mathbf {t})} [ c (\\mathbf {t}) ^ {2} ] - C ^ {2}. (11) \\\\ \\end{array}", + "image_path": "bf7edf3e4b831bfa5773bf6ddd32768e03f02663acc10a251518c261d5406db9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": "For efficient implementation, we compute the first term using a modified CUDA kernel of the 2DGS splatting algorithm, where we render the square of colors in addition to the rendering of the colors. This allows for efficient forward and backward passes during optimization. Intuitively, reducing rendered color variance reduces uncertainty in the rendering, which leads to sharper and more multi-view coherent reconstructions. This reflects positively on the qualitative (Fig 5) and quantitative (Tab 4 and Fig 6) performance of our reconstructions. We note that in practice, loss " + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": " is averaged across pixels for a given view." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 486 + ], + "type": "text", + "content": "Training Objective Our total optimization objective combines the three loss terms:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 491, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 491, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 93, + 491, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {p h o t o}} \\mathcal {L} _ {\\text {p h o t o}} + \\lambda_ {\\text {c o r r}} \\mathcal {L} _ {\\text {c o r r}} + \\lambda_ {\\text {v a r}} \\mathcal {L} _ {\\text {v a r}}, \\tag {12}", + "image_path": "0db9533fb673505f9c713b34bde0001dc408001696dea8133214a3d854ca2815.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{photo}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{corr}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": " are weighting factors that balance the contribution of each term. We optimize all parameters simultaneously using the Adam [41] optimizer with a learning rate scheduling following [29]. Unlike multi-stage approaches [75], our single-stage optimization allows for faster convergence while maintaining high reconstruction quality. The optimization updates both camera extrinsics " + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{P\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": " per input view and 2D Gaussian primitive parameters " + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "inline_equation", + "content": "\\{\\alpha, \\mathbf{c}, \\mathbf{q}, \\mathbf{s}\\}" + }, + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": " according to the gradient of the total loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We evaluate our approach across 3D reconstruction, novel view synthesis, and camera pose estimation. We outline the experimental setup, including datasets, baseline methods, and implementation details, followed by quantitative and qualitative comparisons. Ablation studies further analyze the impact of key components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Configuration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 90, + 553, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 125 + ], + "type": "text", + "content": "Our method is tested on benchmark datasets [1, 5, 42, 92, 97]. Further implementation details and additional qualitative results are provided in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 126, + 554, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 554, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 554, + 222 + ], + "type": "text", + "content": "Few-Shot 3D Reconstruction: We evaluate sparse 3D reconstruction on DTU [1], BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]. For 3-view reconstruction on DTU, we follow SparseNeuS [50], evaluating 15 scenes with two sets of 3-view inputs, reporting the mean per scene. We also provide qualitative comparisons for 3, 6, and 12-view settings on selected scenes from BlendedMVS [92], MVImgNet [97], and MipNeRF360 [5]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 223, + 554, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 554, + 282 + ], + "type": "text", + "content": "Few-Shot Novel View Synthesis: We conduct novel view synthesis (NVS) on Tanks and Temples [42] (8 scenes), MVImgNet [97] (7 scenes), and MipNeRF360 [5] (9 scenes), following InstantSplat [21, 22]. Results are reported under sparse input settings (3, 6, and 12 views)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 283, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 354 + ], + "type": "text", + "content": "Camera Pose Estimation: Camera poses are evaluated on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5], following InstantSplat [21, 22]. Additionally, for few-shot reconstruction, we assess camera poses on DTU [1] using the same 3-view evaluation protocol as in the reconstruction task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 364, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 472 + ], + "type": "text", + "content": "Evaluation Metrics. For 3D reconstruction quality, we employ the Absolute Relative Error (Rel) and Normal Consistency (NC). Rel measures the accuracy of depth estimation following DUSt3R [79], while NC evaluates the geometric fidelity by measuring the alignment between predicted and ground truth surface normals. We apply these metrics to our final meshes. We compute ground truth normals from the gradients of the ground truth DTU depth maps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 473, + 554, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 580 + ], + "type": "text", + "content": "For novel view synthesis, we use standard metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [81], and Learned Perceptual Image Patch Similarity (LPIPS) [102]. We note that for this evaluation we follow InstantSplat[21, 22]: for novel test views without pre-computed camera poses, we perform an additional camera optimization step, which optimizes the test cameras while keeping the 2D Gaussian primitives fixed using the " + }, + { + "bbox": [ + 313, + 473, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 473, + 554, + 580 + ], + "type": "text", + "content": " photometric reconstruction loss." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 581, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 628 + ], + "type": "text", + "content": "For camera pose estimation, we report the Absolute Trajectory Error (ATE) as defined in [8, 21], using COLMAP-computed poses from dense views as ground-truth references." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 638, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 553, + 662 + ], + "type": "text", + "content": "Baselines. We compare our method against several state-of-the-art approaches:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "- NVS specific methods: Pose-free methods NoPe-NeRF [8], CF-3DGS [25], NeRFmm [82], SPARF [75] and InstantSplat [21, 22] (in both S and XL configurations). We also include comparisons with 3DGS [40] and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27480" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 224 + ], + "type": "image", + "image_path": "9a6a443dd589d0469fd40f93ab0ba1d98180304669f8cc90e02d554da7227f9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 251 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on DTU dataset from 3 input images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 258, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 258, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 63, + 258, + 295, + 282 + ], + "type": "text", + "content": "FSGS [104], which use pre-computed camera parameters from Colmap [69]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 295, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 366 + ], + "type": "text", + "content": "- Reconstruction-specific methods: MASt3R [44], Colmap [69], UfoRecon [56], CasMVSNet [26], Spurfies [67], SpaRP [87], SparseCraft [95], InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22]. Of these, similar to our approach, InstantSplatGOF (our implementation based on [22]) and InstantSplat2dns [22] further refine cameras." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": "Implementation Details. Our framework is implemented in PyTorch. We set the optimization iterations to 1k for DTU dataset reconstruction experiments and between 2k-4k for other reconstruction and novel view synthesis experiments. For test-time optimization, camera parameters are jointly optimized for 1k iterations (while keeping the learned Gaussian parameters fixed). We initialize our pipeline with MASt3R [44] configured at a resolution of 512 pixels and use Open3D [103] to compute normals for the point cloud, which are used to initialize rotation parameters of the 2D Gaussians. All experiments are conducted on a single NVIDIA A6000 GPU. We set " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{photo}} = 1.0" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{corr}} = 5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": ". The variance loss weight " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": " follows a cosine annealing schedule: " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{var}} = 0.5\\left(1 + \\cos \\left(\\frac{\\pi \\cdot t}{T}\\right)\\right)" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": " denotes the current training step and " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": " the total number of steps. This schedule gradually decreases " + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{var}}" + }, + { + "bbox": [ + 55, + 375, + 296, + 592 + ], + "type": "text", + "content": " from 1.0 to 0.0, ensuring an adaptive regularization throughout training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 220, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 220, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 220, + 612 + ], + "type": "text", + "content": "4.2. 3D Reconstruction Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 700 + ], + "type": "text", + "content": "For few-shot reconstruction, we follow the evaluation protocol from prior work [50, 56, 95], cleaning the generated meshes with masks from training views. We evaluate the Absolute Relative Error (Rel) and Normal Consistency (NC) between predicted depths normals from the extracted meshes and the ground truth depths normals from the DTU dataset [1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Tab. 1 and 2 present the quantitative results on the DTU" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 293 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 293 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 293 + ], + "type": "image", + "image_path": "6305944af912b06f0aff59928c7261dad0296f90d6452780b12f6d36b2e9ea7e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on BlendedMVS dataset from 3 input images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 327, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 434 + ], + "type": "text", + "content": "dataset for Rel and NC metrics, respectively. Fig 3, 4 and 7 show qualitative reconstruction comparisons, with additional qualitative comparisons with SpaRP [87] included in the supplementary material. Additionally, the supplementary material contains extended qualitative results, in the form of video comparisons demonstrating our reconstructions for several datasets. Our method achieves state-of-the-art performance, outperforming all competing approaches across the majority of scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 591 + ], + "type": "text", + "content": "Several key observations can be made from these results. Pretrained generalizable methods (UfoRecon [56] and CasMVSNet [26]) generalize poorly when faced with noisy initial camera parameters from MASt3R [44], resulting in significantly worse performance. Despite being designed for sparse reconstruction, SparseCraft [95] struggles with the noisy camera poses and point clouds from MASt3R [44]. Spurfies [67], which employs an SDF prior trained on ShapeNet, performs particularly poorly with noisy camera poses, resulting in worse mesh extractions. InstantSplat [21, 22] variants (InstSplatGOF and InstSplat2DGS) perform better than other competing approaches, but still produce overly smooth meshes lacking fine details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 689 + ], + "type": "text", + "content": "Our method significantly outperforms these approaches due to several key advantages, including the variance regularization loss (Sec. 3.3) promoting sharper and more detailed reconstructions Fig 5, The correspondence loss (Sec. 3.3) and joint optimization framework effectively refine the initial noisy camera poses from MAsst3R [44]. Our method's ability to preserve fine geometric details is also evidenced by the superior Normal Consistency scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Table 3 presents camera pose estimation results measured by ATE. Our approach achieves significant improve" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27481" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 145 + ], + "type": "table", + "html": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]2.4920.4812.025.398.866.34.719.017.7010.335.862.505.04.564.957.34
UFORecon [56]35.5557.1538.8135.4750.0244.9434.7053.5641.6747.9340.2239.3439.7533.3449.0442.77
CasMVSNet [26]9.7638.0037.0317.3019.3924.106.5729.5711.8414.9212.388.383.939.8713.4417.10
Spurfies [67]2.9652.1558.4220.2514.6619.244.9712.447.5010.647.366.2638.6125.8444.1921.70
InstSplatGOF [21]9.5423.1413.7415.6213.3012.3311.1912.2114.6915.6113.6112.3015.3112.9012.0513.84
Sparsecraft [95]2.9610.517.162.928.179.443.836.197.616.444.756.435.367.658.146.50
InstantSplat2DGS [22]2.4221.749.083.086.773.603.635.245.827.124.482.783.623.113.435.73
Ours2.4212.3137.93.156.053.5263.484.985.296.224.392.413.463.323.384.82
", + "image_path": "83eeb0af8fa39f4896cc229207a561397ffb54314d1d90040ac181ccc2a489ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 158, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 74, + 147, + 534, + 156 + ], + "lines": [ + { + "bbox": [ + 74, + 147, + 534, + 156 + ], + "spans": [ + { + "bbox": [ + 74, + 147, + 534, + 156 + ], + "type": "text", + "content": "Table 1. 3-view reconstruction results on DTU dataset The relative error (rel↓) is reported for 15 scans. Best results are bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 158, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 57, + 158, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 57, + 158, + 553, + 235 + ], + "type": "table", + "html": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122\\( \\uparrow \\) Mean
MASt3R [44]0.8880.6950.8440.8740.8370.8390.8510.7540.8240.7450.8710.8750.8840.8310.8460.830
UFORecon [56]0.3620.2780.5060.4690.2770.3720.3830.3610.3620.3380.4240.3660.3320.4020.3390.371
CasMVSNet [26]0.7180.5020.6190.7560.6930.6830.7400.5640.7040.6810.7850.7950.8590.7780.7550.709
Spurfies [67]0.8630.350.2970.7110.7420.6680.7710.6710.7770.6830.780.7710.5090.5720.4330.640
InstSplatGOF [21]0.7740.5810.6890.6960.7460.7500.7100.6580.6820.6530.6950.7240.7100.6630.6500.692
Sparsecraft [95]0.8410.6790.8270.8640.7280.7140.8340.7560.7460.7650.8370.6290.8010.7610.7350.768
InstSplat2DGS [22]0.8820.6160.860.8990.8340.8650.8450.7830.8100.7730.8820.8490.8950.8270.8550.832
Ours0.9040.750.8770.8960.8660.8870.8670.7990.8450.7890.8880.8740.9020.850.8650.857
", + "image_path": "d7c7fa5a9cca8cd5ef106b8e59efa127987aeaa16aee610d4f71445ad439d7d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 236, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 236, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 236, + 541, + 247 + ], + "type": "text", + "content": "Table 2. 3-view reconstruction results on DTU dataset Normal consistency (NC↑) is reported for 15 scans. Best results are bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 258, + 295, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 318 + ], + "type": "text", + "content": "ments in camera pose estimation, with a " + }, + { + "bbox": [ + 55, + 258, + 295, + 318 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 55, + 258, + 295, + 318 + ], + "type": "text", + "content": " reduction in error compared to the best-performing baseline (InstSplatGOF [21]). This demonstrates the effectiveness of our correspondence loss in refining the initial noisy camera poses from MAST3R [44]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 231, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 231, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 231, + 338 + ], + "type": "text", + "content": "4.3. Novel View Synthesis Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 403 + ], + "type": "text", + "content": "We evaluate our method's capability for novel view synthesis on Tanks and Temples [42], MVImgNet [97], and MipNeRF360 [5] datasets. Quantitative and qualitative result comparisons on all datasets are included in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 155, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 155, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 155, + 422 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "content": "We conduct ablation studies to analyze the contribution of each component in our method, including an experiment to study the impact of varying the number of input views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 469, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 659 + ], + "type": "text", + "content": "Component Ablation These experiments were performed using 3 input unposed views on 5 testing scenes from the DTU dataset using two sets of views, following the protocols established in [50, 56]. Tab 4 shows the impact of incorporating different components into our pipeline. Fig 5 shows a qualitative ablation comparison. Camera optimization using bundle adjustment improves all metrics. Incorporating the correspondence loss further enhances this alignment, resulting in improved reconstructions. Splat orientation initialization with pointmap normals improves results further. Notice that our variance loss contributes to both the final metrics and the qualitative aspect of our results. The reconstructions with variance loss exhibit significantly sharper details and better preservation of fine structures, validating the quantitative improvements observed in our metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Number of Views Ablation We progressively increase the number of views and perform pose-free reconstruction with our method using scan37 from DTU. We evaluate by computing the mean metrics across both test view sets. As" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 256, + 555, + 494 + ], + "blocks": [ + { + "bbox": [ + 316, + 256, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 316, + 256, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 555, + 494 + ], + "type": "image", + "image_path": "11ed57c145d7816deeda225b27f0acd20dad063b7773d2750cabe523f0300f74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "type": "text", + "content": "Figure 5. Qualitative ablation of the component of our method, showing also the effect of variance loss. Left blue box: w/o " + }, + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{var}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "type": "text", + "content": ". Right red box: w " + }, + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{var}}" + }, + { + "bbox": [ + 313, + 498, + 555, + 531 + ], + "type": "text", + "content": ". (3 input images)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 537, + 555, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 555, + 585 + ], + "type": "text", + "content": "shown in Figure 6, our method benefits favorably from increased views. The benefits of all components of our method are also maintained, if not increased. This includes our variance loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "text", + "content": "5. Running time" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "In our reconstruction experiments, we observed that the mean training time on DTU [1] for 1000 iterations (similar to InstantSplat-XL [21, 22]) was roughly 146s on a single Nvidia A6000 GPU. For a fair comparison, we also trained InstantSplat-XL [21, 22] on the same GPU and observed that the mean training time was around 128s. Hence, our training time is in the same range as InstantSplat [21, 22] while offering significantly higher reconstruction and novel-view quality." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27482" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 553, + 117 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 117 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 117 + ], + "type": "table", + "html": "
Methodsscan24scan37scan40scan55scan63scan65scan69scan83scan97scan105scan106scan110scan114scan118scan122↓Mean
MASt3R [44]0.02020.06890.00940.02030.01710.03920.02910.02940.02710.03260.01670.02100.02890.01920.02420.0269
InstSplatGOF [21]0.01800.05360.00890.01360.00490.02730.02630.02390.01800.03230.01280.02380.02490.01750.01490.0214
InstSplat2DGS [22]0.01920.08950.00900.01540.01230.03380.02940.02720.02420.03220.01660.02100.02740.01630.01750.0261
Ours0.00630.04500.00690.01240.01750.02940.01980.02830.01460.02730.00820.01790.02420.01120.01010.0186
", + "image_path": "03861876aed40843d323b9101d0af3401ca5b3477a27d1b8d0e56a93aefa1ccc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 137, + 276, + 194 + ], + "blocks": [ + { + "bbox": [ + 124, + 118, + 485, + 128 + ], + "lines": [ + { + "bbox": [ + 124, + 118, + 485, + 128 + ], + "spans": [ + { + "bbox": [ + 124, + 118, + 485, + 128 + ], + "type": "text", + "content": "Table 3. Results on DTU dataset Results include ATE " + }, + { + "bbox": [ + 124, + 118, + 485, + 128 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 124, + 118, + 485, + 128 + ], + "type": "text", + "content": " metrics for 15 scans. Best results are bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 137, + 276, + 194 + ], + "lines": [ + { + "bbox": [ + 77, + 137, + 276, + 194 + ], + "spans": [ + { + "bbox": [ + 77, + 137, + 276, + 194 + ], + "type": "table", + "html": "
Methodsrel ↓NC ↑ATE ↓
2DGS7.930.7850.0327
BA-2DGS7.790.7880.0284
BA-2DGS+Lcorr7.410.8040.0256
BA-2DGS+Lcorr+Norm. init.7.030.8290.0237
BA-2DGS+Lcorr+Norm. init.+Lvar6.790.8340.0238
", + "image_path": "cd08e9161e9119c2c9dd436b19e0cd3b2230c9d4e020076148ddb4a0464703a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 202, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 295, + 224 + ], + "type": "text", + "content": "Table 4. Ablation on DTU dataset, adding one component at a time. Results include rel and NC metrics. Best results are bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 247, + 173, + 353 + ], + "blocks": [ + { + "bbox": [ + 59, + 247, + 173, + 353 + ], + "lines": [ + { + "bbox": [ + 59, + 247, + 173, + 353 + ], + "spans": [ + { + "bbox": [ + 59, + 247, + 173, + 353 + ], + "type": "image", + "image_path": "47137a74a20088ba55944074e10e22f89f5ccd9f60fee40815dd7b750f92d2a6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 364, + 295, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 385 + ], + "type": "text", + "content": "Figure 6. Effect of increasing the number of input views on different metrics." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 247, + 294, + 353 + ], + "blocks": [ + { + "bbox": [ + 179, + 247, + 294, + 353 + ], + "lines": [ + { + "bbox": [ + 179, + 247, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 179, + 247, + 294, + 353 + ], + "type": "image", + "image_path": "f13245cd30b3c8b573694d5338628d5c0380c6c9098ed2208f176a1322be650b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 222, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 222, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 222, + 420 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 429, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 429, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 429, + 295, + 584 + ], + "type": "text", + "content": "Despite the promising results, several limitations remain, which present opportunities for future research. As foundation models keep on improving, better initializations could be available in the future. Furthermore, while our approach handles a variety of scenes across different datasets, performance may be limited for highly complex geometry, very thin structures, or semi-transparent objects (observed for the glass bottle in Fig 7 room scene in MipNeRF360 [5]), which are common challenges for surface based methods especially in the sparse setting. Additionally, our method assumes a static scene under consistent lighting. Extending it to more dynamic scenarios is one of the challenges we are looking forward to exploring." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "We addressed the underexplored problem of shape reconstruction from sparse uncalibrated multi-view images. Our method leverages foundation models for initial camera and geometry estimation, followed by joint optimization of 2D Gaussian primitives and camera parameters. The introduction of a variance regularization loss significantly improves the detail and sharpness of reconstructions, while our correspondence matching based framework enhances" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 137, + 553, + 534 + ], + "blocks": [ + { + "bbox": [ + 316, + 137, + 553, + 534 + ], + "lines": [ + { + "bbox": [ + 316, + 137, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 553, + 534 + ], + "type": "image", + "image_path": "e4d4b23f1482015766d47d7a0e2f26babd5824cb44868f579f50cc30f2f0f0d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 535, + 553, + 557 + ], + "lines": [ + { + "bbox": [ + 313, + 535, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 553, + 557 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on MVImgNet and MipNeRF360 datasets from 6 and 12 input images, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 578, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 626 + ], + "type": "text", + "content": "camera pose accuracy. Extensive evaluations on multiple datasets demonstrate that our approach consistently outperforms state-of-the-art methods in reconstruction quality, novel view synthesis, and camera pose estimation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27483" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 65, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 194, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 194, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 65, + 194, + 296, + 237 + ], + "type": "text", + "content": "[3] Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1034-1040. IEEE, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 239, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 239, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 65, + 239, + 296, + 294 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460-5469, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 65, + 295, + 296, + 350 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 352, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 352, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 65, + 352, + 296, + 407 + ], + "type": "text", + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 408, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 408, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 65, + 408, + 296, + 452 + ], + "type": "text", + "content": "[7] Jia-Wang Bian, Wenjing Bian, Victor Adrian Prisacariu, and Philip Torr. Porf: Pose residual field for accurate neural surface reconstruction. arXiv preprint arXiv:2310.07449, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 454, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 454, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 65, + 454, + 296, + 510 + ], + "type": "text", + "content": "[8] Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 510, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 510, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 65, + 510, + 296, + 555 + ], + "type": "text", + "content": "[9] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 556, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 556, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 61, + 556, + 296, + 621 + ], + "type": "text", + "content": "[10] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 678 + ], + "type": "text", + "content": "[11] Chao Chen, Zhizhong Han, and Yu-Shen Liu. Unsupervised inference of signed distance functions from single sparse point clouds without learning priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 680, + 296, + 714 + ], + "type": "text", + "content": "[12] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 150 + ], + "type": "text", + "content": "[13] Zhang Chen, Zhong Li, Liangchen Song, Lele Chen, Jingyi Yu, Junsong Yuan, and Yi Xu. Neurbf: A neural fields representation with adaptive radial basis functions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4182-4194, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 553, + 205 + ], + "type": "text", + "content": "[14] Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18312-18321, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 320, + 206, + 553, + 260 + ], + "type": "text", + "content": "[15] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Gaussian activated neural radiance fields for high fidelity reconstruction and pose estimation. In European Conference on Computer Vision, pages 264-280. Springer, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 553, + 316 + ], + "type": "text", + "content": "[16] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 553, + 358 + ], + "type": "text", + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. arXiv preprint arXiv:2107.02791, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 360, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 415 + ], + "type": "text", + "content": "[18] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 449 + ], + "type": "text", + "content": "[19] John Duchi, Peter Glynn, and Hongseok Namkoong. Statistics of robust optimization: A generalized empirical likelihood approach, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 449, + 553, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 449, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 320, + 449, + 553, + 482 + ], + "type": "text", + "content": "[20] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In CVPR, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 483, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 483, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 320, + 483, + 553, + 547 + ], + "type": "text", + "content": "[21] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2(3):4, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 553, + 603 + ], + "type": "text", + "content": "[22] Zhiwen Fan, Kairun Wen, Wenyan Cong, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Sparse-view sfm-free gaussian splatting in seconds. arXiv preprint arXiv:2403.20309, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 604, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 553, + 658 + ], + "type": "text", + "content": "[23] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, et al. Large spatial model: End-to-end unposed images to semantic 3d. arXiv preprint arXiv:2410.18956, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[24] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27484" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[25] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatt-ing. arXiv preprint arXiv:2312.07504, 2023. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 294, + 161 + ], + "type": "text", + "content": "[26] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495-2504, 2020. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 294, + 205 + ], + "type": "text", + "content": "[27] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jisang Han, Jiaolong Yang, Chong Luo, and Seungryong Kim. Pf3plat: Pose-free feed-forward 3d gaussian splatting. arXiv preprint arXiv:2410.22128, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "text", + "content": "[28] Berthold KP Horn. Shape from shading: A method for obtaining the shape of a smooth opaque object from one view. 1970. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[29] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 294, + 336 + ], + "type": "text", + "content": "[30] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4369-4379, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 370 + ], + "type": "text", + "content": "[31] Peter J Huber. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pages 492-518. Springer, 1992. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 415 + ], + "type": "text", + "content": "[32] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 448 + ], + "type": "text", + "content": "[33] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Neural mesh-based graphics. In European Conference on Computer Vision, pages 739-757. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "text", + "content": "[34] Shubhendu Jena, Franck Multon, and Adnane Boukhayma. Geotransfer: Generalizable few-shot multi-view reconstruction via transfer learning. arXiv preprint arXiv:2408.14724, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 294, + 525 + ], + "type": "text", + "content": "[35] Shubhendu Jena, Shishir Reddy Vutukur, and Adnane Boukhayma. Sparsplat: Fast multi-view reconstruction with generalizable 2d gaussian splatting. arXiv, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 579 + ], + "type": "text", + "content": "[36] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5846-5854, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "type": "text", + "content": "[37] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 636, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 657 + ], + "type": "text", + "content": "[38] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[39] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[40] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 150 + ], + "type": "text", + "content": "[41] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 195 + ], + "type": "text", + "content": "[42] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 553, + 229 + ], + "type": "text", + "content": "[43] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 553, + 263 + ], + "type": "text", + "content": "[44] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 320, + 266, + 553, + 319 + ], + "type": "text", + "content": "[45] Qian Li, Franck Moulton, and Adnane Boukhayma. Learning generalizable light field networks from few images. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 553, + 365 + ], + "type": "text", + "content": "[46] Qian Li, Franck Multon, and Adnane Boukhayma. Regularizing neural radiance fields from sparse rgb-d inputs. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2320-2324. IEEE, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "text", + "content": "[47] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 411, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 553, + 455 + ], + "type": "text", + "content": "[48] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741-5751, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "text", + "content": "[49] Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Fast generalizable gaussian splattering reconstruction from multi-view stereo. arXiv preprint arXiv:2405.12218, 2024.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 512, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 553, + 566 + ], + "type": "text", + "content": "[50] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 320, + 568, + 553, + 623 + ], + "type": "text", + "content": "[51] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 320, + 624, + 553, + 678 + ], + "type": "text", + "content": "[52] Baorui Ma, Yu-Shen Liu, Matthias Zwicker, and Zhizhong Han. Surface reconstruction from point clouds by learning predictive context priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6326-6337, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "text", + "content": "[53] Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27485" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "text", + "content": "synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539-16548, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 108, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 108, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 108, + 295, + 163 + ], + "type": "text", + "content": "[54] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 165, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 165, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 295, + 219 + ], + "type": "text", + "content": "[55] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 222, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 222, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 62, + 222, + 296, + 287 + ], + "type": "text", + "content": "[56] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. Uforecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 290, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 290, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 290, + 295, + 355 + ], + "type": "text", + "content": "[57] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 412 + ], + "type": "text", + "content": "[58] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. arXiv preprint arXiv:2112.00724, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 470 + ], + "type": "text", + "content": "[59] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 472, + 295, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 295, + 505 + ], + "type": "text", + "content": "[60] Amine Ouasfi and Adnane Boukhayma. Few'zero level set'-shot learning of shape signed distance functions in feature space. In ECCV, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 507, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 507, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 507, + 295, + 529 + ], + "type": "text", + "content": "[61] Amine Ouasfi and Adnane Boukhayma. Mixing-denoising generalizable occupancy networks. 3DV, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 532, + 295, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 295, + 573 + ], + "type": "text", + "content": "[62] Amine Ouasfi and Adnane Boukhayma. Few-shot unsupervised implicit neural shape representation learning with spatial adversaries. arXiv preprint arXiv:2408.15114, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "text", + "content": "[63] Amine Ouasfi and Adnane Boukhayma. Robustifying generalizable implicit shape networks with a tunable nonparametric model. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 667 + ], + "type": "text", + "content": "[64] Amine Ouasfi and Adnane Boukhayma. Unsupervised occupancy learning from sparse point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21729-21739, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 712 + ], + "type": "text", + "content": "[65] Amine Ouasfi, Shubhendu Jena, Eric Marchand, and Adnane Boukhayma. Toward robust neural reconstruction from sparse point sets. arXiv preprint arXiv:2412.16361, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[66] Eric Penner and Li Zhang. Soft 3D Reconstruction for View Synthesis. ACM Transactions on Graphics (TOG), 36(6): 1-11, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 149 + ], + "type": "text", + "content": "[67] Kevin Raj, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Spurfies: Sparse surface reconstruction using local geometry priors. arXiv preprint arXiv:2408.16544, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 152, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 215 + ], + "type": "text", + "content": "[68] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 217, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 555, + 260 + ], + "type": "text", + "content": "[69] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 262, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 262, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 262, + 555, + 304 + ], + "type": "text", + "content": "[70] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 306, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 306, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 321, + 306, + 553, + 349 + ], + "type": "text", + "content": "[71] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 553, + 381 + ], + "type": "text", + "content": "[72] Steven M Seitz and Charles R Dyer. Photorealistic Scene Reconstruction by Voxel Coloring, 2002. US Patent 6,363,170. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 383, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 383, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 321, + 383, + 553, + 448 + ], + "type": "text", + "content": "[73] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 449, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 449, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 321, + 449, + 553, + 470 + ], + "type": "text", + "content": "[74] Andrea Tagliaasacchi and Ben Mildenhall. Volume rendering digest (for nerf), 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 321, + 472, + 553, + 525 + ], + "type": "text", + "content": "[75] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 321, + 526, + 553, + 568 + ], + "type": "text", + "content": "[76] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. arXiv preprint arXiv:2303.16196, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 571, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 571, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 321, + 571, + 553, + 603 + ], + "type": "text", + "content": "[77] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 604, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 604, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 604, + 553, + 656 + ], + "type": "text", + "content": "[78] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "text", + "content": "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27486" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[80] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. Advances in Neural Information Processing Systems, 35:1966-1978, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 162 + ], + "type": "text", + "content": "[81] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 296, + 196 + ], + "type": "text", + "content": "[82] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. 2021. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "type": "text", + "content": "[83] Robert J Woodham. Photometric method for determining surface orientation from multiple images. Optical engineering, 19(1):139-144, 1980. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 285 + ], + "type": "text", + "content": "[84] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. arXiv preprint arXiv:2312.02981, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 288, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 288, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 295, + 331 + ], + "type": "text", + "content": "[85] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 333, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 376 + ], + "type": "text", + "content": "[86] Haolin Xiong, Sairishek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime 360 sparse view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00206, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 378, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 295, + 432 + ], + "type": "text", + "content": "[87] Chao Xu, Ang Li, Linghao Chen, Yulin Liu, Ruoxi Shi, Hao Su, and Minghua Liu. Sparp: Fast 3d object reconstruction and pose estimation from sparse views. In European Conference on Computer Vision, pages 143-163. Springer, 2024. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 489 + ], + "type": "text", + "content": "[88] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "text", + "content": "[89] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[90] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438-5448, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "text", + "content": "[91] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 712 + ], + "type": "text", + "content": "[92] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790-1799, 2020. 1, 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 686 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 127 + ], + "type": "text", + "content": "[93] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "text", + "content": "[94] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 228 + ], + "type": "text", + "content": "[95] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 320, + 230, + 553, + 283 + ], + "type": "text", + "content": "[96] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 285, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 285, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 320, + 285, + 553, + 351 + ], + "type": "text", + "content": "[97] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 1, 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 353, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 353, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 320, + 353, + 553, + 406 + ], + "type": "text", + "content": "[98] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 408, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 408, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 320, + 408, + 553, + 461 + ], + "type": "text", + "content": "[99] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "type": "text", + "content": "[100] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[101] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 553, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 553, + 608 + ], + "type": "text", + "content": "[102] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 609, + 553, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 641 + ], + "type": "text", + "content": "[103] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "text", + "content": "[104] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. arXiv preprint arXiv:2312.00451, 2023. 1, 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27487" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_content_list.json b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7fbde32150d411390cad8bab24be8ad98f6f18b1 --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_content_list.json @@ -0,0 +1,1500 @@ +[ + { + "type": "text", + "text": "Sparse Fine-Tuning of Transformers for Generative Tasks", + "text_level": 1, + "bbox": [ + 205, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Chen, Jingxi Yu, Zichen Miao, Qiang Qiu Purdue University, IN, USA", + "bbox": [ + 313, + 180, + 681, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen2732, yu667, miaoz, qqiu}@purdue.edu", + "bbox": [ + 256, + 217, + 736, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large pre-trained transformers have revolutionized artificial intelligence across various domains, and fine-tuning remains the dominant approach for adapting these models to downstream tasks due to the cost of training from scratch. However, in existing fine-tuning methods, the updated representations are formed as a dense combination of modified parameters, making it challenging to interpret their contributions and understand how the model adapts to new tasks. In this work, we introduce a fine-tuning framework inspired by sparse coding, where fine-tuned features are represented as a sparse combination of basic elements, i.e., feature dictionary atoms. The feature dictionary atoms function as fundamental building blocks of the representation, and tuning atoms allows for seamless adaptation to downstream tasks. Sparse coefficients then serve as indicators of atom importance, identifying the contribution of each atom to the updated representation. Leveraging the atom selection capability of sparse coefficients, we first demonstrate that our method enhances image editing performance by improving text alignment through the removal of unimportant feature dictionary atoms. Additionally, we validate the effectiveness of our approach in the text-to-image concept customization task, where our method efficiently constructs the target concept using a sparse combination of feature dictionary atoms, outperforming various baseline fine-tuning methods. We provide more details at our project page.", + "bbox": [ + 91, + 301, + 482, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 723, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, pre-trained large models have revolutionized artificial intelligence across domains like natural language processing and computer vision [11, 22, 46, 52]. Given the prohibitive computational resources required for training these models from scratch, fine-tuning has emerged as the dominant paradigm for adapting pre-trained models to specific downstream tasks, enabling efficient knowledge transfer while minimizing computational overhead.", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various parameter-efficient fine-tuning (PEFT) methods [17, 26, 43] have been developed to tune the large trans", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "formers by updating a small subset of parameters while keeping large pre-trained weights frozen. A representative family of these approaches [17, 26] utilizes lightweight parameterizations, such as low-rank decomposition, to adapt the weight matrices. However, despite being lightweight, these methods obtain the updated features with a dense matrix multiplication between the input and the adapted weights, making it challenging to interpret how the individual updated parameters contribute to the newly learned representations and to identify the new knowledge gained through fine-tuning.", + "bbox": [ + 511, + 268, + 903, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this challenge, we take inspiration from sparse coding [9, 30, 32, 33, 37, 39, 42, 54-56], in which signals are expressed as a sparse combination of basic elements, i.e., dictionary atoms. In this paper, we frame model fine-tuning as a sparse coding problem, where the adapted feature is constructed using a dictionary of feature atoms. These feature atoms are selectively combined with sparse coefficients, ensuring that only a small subset of atoms contributes to the updated representation. This sparsity not only disentangles interactions between features but also allows each atom to capture a distinct aspect of the downstream data.", + "bbox": [ + 511, + 439, + 903, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We first analyze with a toy experiment that the tuned feature dictionary atoms function as building blocks of the adapted representation, and tuning these atoms allows for effective adaptation to downstream tasks. We demonstrate that with pre-trained coefficients and atoms, fine-tuning the atoms effectively captures the representation of downstream tasks. Representing features through individual atoms enhances controllability in the adaptation process.", + "bbox": [ + 511, + 625, + 903, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We validate our method on generative fine-tuning tasks. We show that the target concept can be effectively represented using feature dictionary atoms after fine-tuning the pre-trained model within our framework. Compared to baseline methods, our approach allows the selection of specific atoms to refine the generated image. This adaptability improves performance on image editing tasks by enhancing editing results through the removal of unimportant atoms. We also apply our approach to customized text-to-image tasks, showing that fine-tuning atoms effectively captures", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18703", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77c1546029f1671b0b227341cd29ef583e75cc8e99d378dd460adf90c9fc0a3d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 359, + 310 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/729fbb1d304c1962ed75a5d40e2081addb6167c069c50549d9a1016b7b3cc437.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 366, + 88, + 630, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47aede7514ac4790cf5e35eef8a6dbe914fc5f711c44988e0614d3229d9bb343.jpg", + "image_caption": [ + "(c)", + "Figure 1. (a) Our method leverages sparse coding and enforces high sparsity in adapted feature $\\Delta \\mathbf{O}$ , ensuring that the output feature is constructed using only a few dictionary atoms $\\mathbf{D}$ . (b) The pre-trained model initially generates \"8\". It generates \"3\" after integrating the learned dictionary $\\mathbf{D}$ which is fine-tuned on \"3\". (c) The overall transformation of the generated digit from \"8\" to \"3\" using the dictionary atoms. The red indicates an increase in value at a specific location and blue indicates a decrease. The total influence is decomposed into contributions from atoms at different layers: $\\Delta \\mathbf{O}^{(1)}$ represents the combined effect of 8 atoms in the first layer, while $\\Delta \\mathbf{O}^{(2)}$ accounts for the 4 atoms in the second layer. The details of this experiment is provided in Section 3.3" + ], + "image_footnote": [], + "bbox": [ + 635, + 107, + 901, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "personal concepts while preserving the capabilities of pretrained models. Our method outperforms various baseline approaches, demonstrating its effectiveness in controlled adaptation for generative modeling.", + "bbox": [ + 88, + 439, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 109, + 500, + 401, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We formulate large transformer fine-tuning as a sparse coding problem, where the adapted feature representation is a sparse combination of a dictionary of feature atoms.", + "- We introduce a fine-tuning framework that leverages a combination of dictionary atoms and demonstrates that tuning atoms enhances control over the adaptation process.", + "- We validate the effectiveness of our approach in generative tasks by demonstrating the improvement in image editing and concept customization." + ], + "bbox": [ + 89, + 516, + 482, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 681, + 232, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse Representation and Attention Recent research has demonstrated the significant role of sparsity in enhancing both interpretability and computational efficiency in generative models through various approaches including sparse autoencoders for feature extraction [49], sparse feature circuits for improved generalization [31], transcoders for detailed circuit analysis [12], dictionary learning for monosemantic representations [5], and selective parameter tuning strategies [18]. Complementary work has addressed the quadratic computational complexity of standard attention mechanisms through sparse attention techniques that compute attention scores for only a subset of token pairs [3, 10, 50, 52], with optimizations including", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dynamic pattern adjustment [24], hardware-aligned implementations [60], and graph-based conceptualizations [51], collectively enabling transformers to process significantly longer sequences while maintaining model quality despite challenges in balancing efficiency with performance in complex long-range dependency scenarios. Compared to previous works, our method applies sparse coding for finetuning transformer-based models in generative tasks.", + "bbox": [ + 511, + 438, + 906, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representation Tuning Recent representation tuning approaches operate directly on model representations rather than weights, achieving parameter efficiency while enhancing controllability, and interpretability across various model types and tasks [1, 8, 27, 28, 36, 53, 57, 58, 64]. Unlike traditional PEFT approaches that impact representations indirectly through weight adjustments, representation tuning methods preserve the original pre-trained model weights while directly modifying the output representations. In comparison, our work studies the sparsity in the updated representation and frames it as a dictionary learning problem.", + "bbox": [ + 511, + 571, + 908, + 751 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Personalized Text-to-Image Generation Extensive research has been conducted on aligning text prompts and generated concepts [19, 20, 23, 25, 34, 35, 41, 47]. Zhang et al. [62] developed an object-conditioned energy-based attention map alignment technique in diffusion models, enhancing control over specific objects in generated images. Another study by Zhang et al. [61] improved semantic fidelity in text-to-image synthesis through attention regulation. Hao et al. [16] introduces ConceptExpress, an unsu", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18704", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pervised method for extracting personalized concepts from a single image. Ren et al. [45] addressed memorization issues in text-to-image diffusion models by analyzing cross-attention mechanisms, proposing solutions to reduce overfitting to user-specific data. We approach this problem by adopting the ideas from sparse coding and learning the target concept as a sparse combination of coefficients and dictionary atoms.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 224, + 181, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 250, + 220, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-head attention (MHA). The attention layer [52] transforms the input sequence $\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}$ to the output sequence $\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}$ , where $N$ denotes the sequence length, $C_i$ and $C_o$ are the dimension of input and output features. The attention layer projects the input as $\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q$ , $\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k$ , $\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v$ using the corresponding projection matrices $\\mathbf{W}_q$ , $\\mathbf{W}_k$ , $\\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}$ , and calculates the attention map,", + "bbox": [ + 89, + 271, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = \\operatorname {a t t n} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 404, + 482, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention map $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ captures the token-wise relationship by doing inner-product in a space transformed by $\\mathbf{W}_q, \\mathbf{W}_k$ . Then the output of the attention layer is,", + "bbox": [ + 89, + 433, + 483, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\mathbf {A X W} _ {v}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 492, + 482, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-head attention extends this by allowing multiple attention mechanisms to work in parallel, with each head independently learning attention patterns. The project matrices for each head are $\\mathbf{W}_q^{(h)},\\mathbf{W}_k^{(h)},\\mathbf{W}_v^{(h)}\\in \\mathbb{R}^{C_i\\times \\frac{C_o}{H}}$ where $h = 1,\\dots ,H$ and $H$ is the number of heads. The multi-head attention represents as,", + "bbox": [ + 89, + 521, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 623, + 482, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}$ and $\\mathbf{W}_{vo}^{(h)} = \\mathbf{W}_v^{(h)}\\mathbf{W}_o^{(h)\\intercal},\\mathbf{W}_o^{(h)}\\in$ $\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}$ $\\mathbf{W}_{vo}^{(h)}\\in \\mathbb{R}^{C_i\\times C_o}$", + "bbox": [ + 89, + 674, + 482, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse coding. Sparse coding is a signal representation technique that models a signal as a linear combination of a small number of basis functions from a dictionary. Given a signal $\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}$ , the goal of sparse coding is to find a dictionary $\\mathbf{D} \\in \\mathbb{R}^{m \\times C_i}$ (where $m > C_i$ is the number of atoms) and a sparse coefficient vector $\\mathbf{S} \\in \\mathbb{R}^{N \\times m}$ such that the signal can be approximated as:", + "bbox": [ + 89, + 727, + 483, + 834 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} \\approx \\mathbf {S D}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 251, + 844, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{S}$ has only a few nonzero elements, ensuring sparsity. The sparse coding problem is often formulated as an", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "optimization problem:", + "bbox": [ + 513, + 92, + 665, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {S}} \\| \\mathbf {X} - \\mathbf {S D} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {S} \\| _ {1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 111, + 903, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda$ is a regularization parameter controlling the tradeoff between sparsity and reconstruction error. The Proximal gradient method is a common approach to solving this optimization problem, including Iterative Shrinkage-Thresholding Algorithm (ISTA) and its accelerated version (FISTA) [2], which utilize proximal operators for efficient iterative updates. For sparse coding with $\\ell_1$ regularization, the corresponding proximal operator is the soft-thresholding function:", + "bbox": [ + 511, + 138, + 906, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {p r o x} _ {\\lambda \\| \\cdot \\| _ {1}} (\\mathbf {S}) = \\operatorname {s i g n} (\\mathbf {S}) \\odot \\max (| \\mathbf {S} | - \\lambda , 0), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 280, + 906, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\odot$ is Hadamard product. It encourages sparsity by shrinking small values of $\\mathbf{S}$ toward zero. Sparse coding is widely used in signal processing, machine learning, and neuroscience, providing efficient and interpretable representations of data.", + "bbox": [ + 511, + 301, + 905, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sparse Fine-Tuning", + "text_level": 1, + "bbox": [ + 513, + 383, + 702, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we conceptualize model fine-tuning as a sparse coding problem, i.e., the adapted features are constructed using a dictionary of feature atoms. As our method focuses on the transformer architecture, we first propose an interpretation of the attention operation by viewing the product between the attention map and value matrices as a combination of dictionary atoms with dictionary coefficients. We then propose a fine-tuning approach by representing the adapted features as a learned sparse combination of dictionary atoms with sparse coefficients, representing the fine-tuned representations' underlying sparsity nature.", + "bbox": [ + 511, + 406, + 906, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A dictionary view of attention. For single head attention (2), each row of $\\mathbf{A}$ can be interpreted as coefficients for combining $\\mathbf{V}$ into the output $\\mathbf{O}$ , i.e., $\\mathbf{O}_i = \\sum_j\\mathbf{A}_{ij}\\mathbf{V}_j$ . In the case of multi-head attention (3),", + "bbox": [ + 511, + 590, + 905, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)} = \\left[ \\begin{array}{l l l} \\mathbf {A} ^ {(1)} & \\dots & \\mathbf {A} ^ {(H)} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(1)} \\\\ \\vdots \\\\ \\mathbf {X} \\mathbf {W} _ {v o} ^ {(H)} \\end{array} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 654, + 903, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "it can be represented as a product of a composite coefficient $\\left[\\mathbf{A}^{(1)},\\dots ,\\mathbf{A}^{(H)}\\right]\\in \\mathbb{R}^{N\\times NH}$ , and an extensive dictionary $\\left[\\mathbf{XW}_{vo}^{(1)},\\dots ,\\mathbf{XW}_{vo}^{(H)}\\right]^{\\intercal}\\in \\mathbb{R}^{NH\\times C_o}$ . Each row of the output $\\mathbf{O}$ is a linear combination of $\\mathbf{XW}_{vo}$ , i.e., $\\mathbf{O}_i = \\sum_{h,j}\\mathbf{A}_{ij}^{(h)}(\\mathbf{XW}_{vo}^{(h)})_j$ . When represented as a linear combination of dictionary elements, multi-head attention and single-head attention follow the same formulation. In the following discussion, we focus on the single-head attention formulation, which can be seamlessly extended to the multi-head attention case.", + "bbox": [ + 511, + 738, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18705", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A sparse dictionary view of attention. From the perspective of sparse coding, the feature is expressed as SD, where S represents the sparse coefficients and D consists of dictionary atoms that remain data-independent once learned. Although attention can be expressed as a linear combination of atoms, it does not inherently exhibit sparsity. To fully leverage the advantages of sparse representations, we propose to represent the attention mechanism (2) in the form of sparse coding,", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\mathbf {S D} = \\sigma_ {\\lambda} \\left(\\mathbf {A X W} _ {s}\\right) \\mathbf {D}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 241, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{\\lambda}(\\cdot)$ can be soft-thresholding function or $\\mathrm{ReLU}(x - \\lambda)$ to ensure sparsity of $\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s) \\in \\mathbb{R}^{N \\times C_o}$ , $\\mathbf{W}_s \\in \\mathbb{R}^{C_i \\times M}$ and $\\mathbf{D} \\in \\mathbb{R}^{M \\times C_o}$ is the feature dictionary atoms, $M$ is the number of atoms.", + "bbox": [ + 89, + 271, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this formulation, we obtain the feature dictionary atoms $\\mathbf{D}$ (we use the term \"atoms\" for simplification in this paper), which are the building blocks for feature representation $\\mathbf{O}$ . We illustrate our formulation in Figure 1 (a). This formulation provides the following properties:", + "bbox": [ + 89, + 332, + 483, + 407 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The sparse coefficients $\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)$ ensure that the feature representation $\\mathbf{O}$ is combined by only a few $\\mathbf{D}$ . We present an extreme example to illustrate how atoms enhance interpretability: $\\mathbf{O}_i = \\sum_{j=1}^{M} \\mathbf{S}_{ij} \\mathbf{D}_j$ can be simplified as $\\mathbf{O}_i = \\mathbf{S}_{ij^*} \\mathbf{D}_{j^*}$ if all elements of $\\mathbf{S}_i$ is 0 except the element at index $j^*$ , and only $\\mathbf{D}_{j^*}$ contributes to the output feature $\\mathbf{O}_i$ .", + "- The sparse coefficients $\\mathbf{S}$ are data-dependent, while the feature dictionary atoms $\\mathbf{D}$ are data-independent. While different inputs generate different coefficients, the output features are always constructed using the same feature dictionary atoms.", + "- We can adjust the size of the feature dictionary by varying $M$ to handle tasks with different difficulties. Large $M$ leads to an overcomplete dictionary while small $M$ produces an undercomplete dictionary." + ], + "bbox": [ + 89, + 407, + 483, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tuning via sparse dictionary learning. The above formulation frames feature representation through the lens of sparse coding. In the context of large model fine-tuning, we freeze the pre-trained parameters to make feature representations $\\mathbf{O}$ of the pre-trained model fixed. Instead, we introduce a sparse coding perspective to encode the adapted feature representation $\\Delta \\mathbf{O}$ , which is learned from downstream tasks. Specifically, we have,", + "bbox": [ + 89, + 662, + 483, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} + \\Delta \\mathbf {O} = \\mathbf {O} + \\sigma_ {\\lambda} (\\mathbf {A X W} _ {s}) \\mathbf {D}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 796, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this formulation, the pre-trained weights remain unchanged, preserving the capabilities of the pre-trained model. The adapted feature $\\Delta \\mathbf{O}$ is represented as a linear combination of feature dictionary atoms $\\mathbf{D}$ , using sparse coefficients $\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)$ . This formulation offers", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/37034641eb61d11891b337d66289edce00d0e01b91ddbace63dcab804c2dc3a9.jpg", + "image_caption": [ + "Figure 2. (a) Adjusting only the coefficients fails to adapt the model to the downstream task, whereas (b) modifying only the atoms enables a perfect fit to the new task composed of high-frequency signals. (c) Fine-tuning the atoms effectively transitions the model from low-frequency to high-frequency representations. They encode the representation of the downstream task." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 883, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a framework for understanding how $\\mathbf{D}$ contributes to the adapted features in the fine-tuned model. We illustrate our method in Figure 1 (a).", + "bbox": [ + 511, + 401, + 906, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Analysis of Dictionary Atoms", + "text_level": 1, + "bbox": [ + 511, + 455, + 774, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we demonstrate distinct properties of atoms $\\mathbf{D}$ and sparse coefficients $\\mathbf{S}$ .", + "bbox": [ + 511, + 479, + 905, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Features atoms are key to fine-tuning. We analyze the distinct behaviors of coefficients and atoms, and discover that atoms play a crucial role in adapting to new tasks. To explore this within the context of an attention block, we perform a toy experiment where masked one-dimensional signals are reconstructed using an attention layer. The signals, with a length of 64, are synthesized by selecting 5 random Fourier bases, with a low frequency ranging from $\\frac{0}{64}$ to $\\frac{24}{64}$ . The objective is to reconstruct the signal after masking out $75\\%$ of its values with only 16 randomly sampled observations. After we have a pre-trained model on the low-frequency signals, we fine-tune the model to a new task where signals are constructed with high-frequency Fourier bases ranging from $\\frac{25}{64}$ to $\\frac{32}{64}$ . In this experiment, we test two different cases: (1) tuning only coefficients on the high-frequency data with atoms fixed, and (2) tuning only atoms on the high-frequency data with coefficients fixed. As illustrated in Figure 2, when adapting to high-frequency signals, modifying only the atoms while keeping the coefficients fixed allows the model to effectively generalize to the new task. In contrast, updating only the coefficients is insufficient for proper adaptation. Building on this observation, we hypothesize that adjusting atoms is more suitable than adjusting coefficients for adapting to a new task. We observe from Figure 2 (c) that atoms are capable of encoding", + "bbox": [ + 511, + 523, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18706", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the representation of the downstream task. The experiment details are in Appendix B.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Identifying the importance of feature atoms. We analyze the impact of atoms on the newly learned concept by visualizing both the influence of individual atoms and their combined effect. In this experiment, we adopt a transformer-based variational autoencoder (VAE) [21] with a two-layer decoder. The experiment process is illustrated in Figure 1 (b). We first pre-train the model on an image-denoising task using a subset of the MNIST dataset containing digits 5 to 9. We then fine-tune the model on a downstream denoising task with only digit 3 by learning a dictionary for each layer with $M = 100$ . In this scenario, the digit \"3\" is the newly learned concept. Given the same noisy latent, the decoder outputs an \"8\" without incorporating fine-tuned feature atoms, whereas it generates a \"3\" when the fine-tuned feature atoms are applied. The model requires only 12 atoms to shift the generated digits from 8 to 3. The influence of each atom is visualized in Figure 1 (c), with the overall impact represented as $\\Delta \\mathbf{O}$ , where the red area indicates an increase in pixel value at a specific location and the blue area represents a decrease. It is straightforward to see that these 12 feature atoms enhance the regions corresponding to \"3\" while suppressing certain portions of \"8\". Additionally, Figure 1 (c) illustrates the individual behavior of each atom, showing how different atoms affect distinct parts of the generated digit. The experiment details are in Appendix B.", + "bbox": [ + 89, + 147, + 485, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Design Variations", + "text_level": 1, + "bbox": [ + 89, + 556, + 264, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we outline the various design choices implemented in our experiments. Our method achieves sparse coefficient $\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)$ by utilizing non-linear function $\\sigma_{\\lambda}(\\cdot)$ . In the literature on sparse coding [4], there are three different methods: (i) soft-thresholding function, $\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{prox}_{\\lambda \\| \\cdot \\|_1}(\\mathbf{x}) = \\mathrm{sign}(\\mathbf{x}) \\odot \\max(|\\mathbf{x}| - \\lambda, 0)$ , (ii) ReLU, $\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{ReLU}(\\mathbf{x} - \\lambda) = \\max(\\mathbf{x} - \\lambda, 0)$ , and (iii) top- $k$ activation selection, which selects the most significant features of top $k$ contributing to the learned representation. In practice, we observe that top- $k$ activation selection maintains a consistent level of sparsity across different inputs. In contrast, methods (i) soft-thresholding function and (ii) ReLU exhibit sensitivity to input values, leading to variations in sparsity. However, methods (i) and (ii) offer control over the coefficient values, ensuring they stay above a specified threshold.", + "bbox": [ + 89, + 580, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we find that when $\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)$ is highly sparse, the resulting product $\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)$ also retains sparsity. In modern transformers, computation of $\\mathbf{A}$ and $\\mathbf{X}\\mathbf{W}_v$ are performed separately. For simplicity in implementation, we adopt the formulation $\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)$ .", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Analysis of Our Design", + "text_level": 1, + "bbox": [ + 513, + 90, + 728, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with LoRA. When representing attention mechanism (1-3) as a linear combination of dictionary atoms, updating $\\mathbf{W}_q$ , $\\mathbf{W}_k$ is equivalent to updating the coefficients while updating $\\mathbf{W}_v$ , $\\mathbf{W}_o$ is equivalent to updating the atoms. LoRA applies adaptation on $\\mathbf{W}_q$ , $\\mathbf{W}_k$ , $\\mathbf{W}_v$ , $\\mathbf{W}_o$ with a low-rank matrices $\\mathbf{W}_A \\in \\mathbb{R}^{C_i \\times r}$ , $\\mathbf{W}_B \\in \\mathbb{R}^{r \\times C_o}$ , where $r$ represents the rank. LoRA adapts both coefficients and atoms in a low-rank way. In comparison, our method adapts the coefficients with $\\mathbf{W}_s$ and obtains sparse coefficients $\\sigma_{\\lambda}(\\mathbf{AXW}_s)$ , while representing atoms with $\\mathbf{D}$ .", + "bbox": [ + 511, + 113, + 906, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Parameters and computational cost. To adapt the representation at each layer, we need additional parameters for $\\mathbf{W}_s$ and $\\mathbf{D}$ , which are $(C_i + C_o)M$ . After the fine-tuning, considering the density of the coefficients $\\rho$ , i.e., $1 - \\rho$ is sparsity, the storage required for the parameters is $C_i M + \\rho C_o M$ . In comparison, LoRA requires $4(C_i + C_o)r$ additional parameters. The FLOPs required by our method are $2N M C_i + 2\\rho N M C_o$ . LoRA requires $8C_i C_o r$ FLOPs. For example, if $C_i = C_o = 1024$ , $M = 256$ , $r = 16$ , $N = 1024$ , $\\rho = 0.01$ , our method requires 264k parameters for storage and 0.54GFLOPs for computation, while LoRA requires 131k parameters and 0.13GFLOps. Our method involves slightly higher computational costs than LoRA, as it accounts for feature interactions, whereas LoRA only updates the weights. Next, we demonstrate that by leveraging sparse feature representations, our approach achieves greater robustness compared to LoRA.", + "bbox": [ + 511, + 281, + 908, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method provides more stable adaptation. We compare an extreme case between LoRA and our method, by considering the attention in a dictionary form $\\Delta \\mathbf{O} = \\mathbf{A}\\mathbf{V}$ or $\\Delta \\mathbf{O} = \\mathbf{SD}$ . In this experiment, we examine the difference between updating coefficients using a low-rank approach versus a sparse approach. Both LoRA and our method produce the updated feature $\\Delta \\mathbf{O}$ with rank 1, where LoRA achieves this by setting $r = 1$ and our method enforces only one active atom in D. The key difference lies in how the coefficients are updated—LoRA adapts them in a low-rank manner, whereas our method updates them sparsely. As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. LoRA fails to produce meaningful images with a slightly modified prompt, while our method remains robust, consistently generating a concept that aligns with the updated text prompt. We attribute the effectiveness of our approach to the sparse coefficients produced by $\\sigma_{\\lambda}(\\mathbf{AXW}_s)$ . With a different text prompt, the values in S either remain zero or take on a significant value when receiving the proper input, ensuring that the generated result remains coherent. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs.", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18707", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbb7deb14fb6089e4f69d50f446103fb7166aca5d7a08551e59194b5ed405e56.jpg", + "image_caption": [ + "Figure 3. The pre-trained model can generate the target concept when integrated with the learned dictionary, even with about 40 atoms. With just four atoms, the core structure of the concept can still be generated. Adjusting the number of selected atoms influences the generated results, offering a controllable approach to image editing. Additional discussions of these figures are in Section 4.1." + ], + "image_footnote": [], + "bbox": [ + 155, + 89, + 844, + 343 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8216130cf24de06def538ebcb8bfbaa1681552147f8a4a0e26223a49d1451517.jpg", + "image_caption": [ + "Figure 4. The role of sparsity: Our method utilizes the same number of atoms as LoRA, yet LoRA struggles to adapt to the target prompt effectively. Sparse coding ensures that atoms are either activated or inactive, providing a structured and interpretable adaptation process. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs." + ], + "image_footnote": [], + "bbox": [ + 107, + 412, + 472, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 752, + 223, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of our method in image generation tasks. We demonstrate that feature dictionary atoms effectively represent the target concept after fine-tuning the pre-trained model. Compared to baselines, our approach enables selective atom adjustment, improving image editing by refining results and enhancing text-to-image concept customization while preserving pre-trained model capabilities.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Representing Concepts Using Atoms", + "text_level": 1, + "bbox": [ + 511, + 410, + 830, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental settings. In this experiment, we utilize PixArt- $\\Sigma$ [7], a Diffusion Transformer (DiT) [40] model consisting of 28 transformer blocks. We fine-tune the model on a single image for 60 epochs using our method and reconstruct the corresponding concept using the same prompt used during fine-tuning. The selected images are drawn from DreamBooth [47], Custom Diffusion [23]. We control the number of active atoms and visualize the variation of the generated results with different atoms.", + "bbox": [ + 511, + 431, + 906, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The concept is represented by a set of atoms. As shown in Figure 3, after fine-tuning our method on the target concept, its representation can be effectively captured using a small set of atoms, such as 40 atoms. Among these, just 4 atoms are sufficient to construct the core structure of the concept, including key attributes like pose and color. We provide details descriptions of these images in Appendix C.", + "bbox": [ + 511, + 580, + 908, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Image Editing", + "text_level": 1, + "bbox": [ + 511, + 696, + 661, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on our previous observation that adjusting the number of atoms influences the generated results, we explore leveraging this property for image editing. Image editing tasks aim to modify an image using a generative model while retaining as many details as possible from the original [63]. However, certain editing tasks pose challenges in maintaining most of the original details while achieving effective modifications. For example, altering the style can significantly change the overall appearance of the image. We argue that successful image editing should preserve the main components of the image while applying modifications selectively. Since our method represents the core", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18708", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36b33affdd79a29e0e81d0545200c0b59997033b1c86c600b662afd65bac9bf4.jpg", + "image_caption": [ + "Figure 5. Results on image editing. Our method shows improved results by better following editing prompts while preserving the key components of the source images." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 468, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "components of a concept using only a few atoms, it is well-suited for image editing tasks, ensuring that key structural elements remain intact while allowing for controlled transformations.", + "bbox": [ + 89, + 500, + 483, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental settings. In our experiment, we select data from previous work [63] and evaluate different editing tasks, including background modification, scene alteration, and pose adaptation. Previous works [15, 63] fine-tune models to learn a concept and modify its appearance by adjusting the prompt. However, these approaches have not yet been applied to DiT. Following their setup, we compare our method with fine-tuning approaches, such as LoRA [17], using them as baselines for evaluation. We adopt the DiT architecture and fine-tune both LoRA and our method on a single image for 60 epochs to achieve full reconstruction. After fine-tuning, we modify the image using a different prompt for editing.", + "bbox": [ + 89, + 579, + 483, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image editing results. The results are shown in Figure 5. Our method applies edits based on the test prompt while maintaining the core structure of the source images. In addition, we control the number of active atoms by adjusting the coefficient density $\\rho$ , where $1 - \\rho$ represents the sparsity level. In Figure 6, as sparsity increases, only the most essential atoms remain, preserving the core concept while al", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b93197756b6e23e6edfc00ca4b60b4d148bdfca649748376b52735d4f4218ba.jpg", + "image_caption": [ + "$\\rho = 0.01$" + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 648, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82b26ff313d9e7fb87aafc67cce219ae5790621996dc2649db2a7586d60fb039.jpg", + "image_caption": [ + "$\\rho = 0.02$" + ], + "image_footnote": [], + "bbox": [ + 651, + 90, + 769, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/351038a862d7318ed76948c56ed0396141491118b19eea0e699f3c98fac29398.jpg", + "image_caption": [ + "$\\rho = 0.04$", + "Figure 6. Adjusting the density, i.e., the number of activated atoms, can change the influence on the edited image." + ], + "image_footnote": [], + "bbox": [ + 769, + 90, + 887, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lowing the pre-trained model greater flexibility to integrate editing effects based on the modified text prompt.", + "bbox": [ + 511, + 258, + 906, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Concept Customization", + "text_level": 1, + "bbox": [ + 511, + 297, + 730, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the previous section, we demonstrated that our method can efficiently represent the target concept using only a small number of atoms. In this section, we further show that even with just a few atoms, our method achieves good performance in concept customization.", + "bbox": [ + 511, + 319, + 906, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental settings. We evaluated our approach on the DreamBooth [47] dataset, which consists of 30 subjects across 15 distinct classes. Among these, 9 are live subjects, while the remaining 21 are objects. Each subject has between 4 to 6 images, captured under varying conditions, in diverse environments, and from multiple angles. All 30 subjects were trained for 15 epochs and validated on 25 different prompts. For our experiments, we adopted the Pixart- $\\Sigma$ , a Diffusion Transformer (DiT) [7], as the base architecture and fine-tuned it using our proposed method. To verify the effectiveness of our approach, we compared its performance against several PEFT baselines for personalized text-to-image generation, including LoRA [17], DoRA [26], and OFT [43]. We use the AdamW optimizer with a learning rate of $1\\times 10^{-4}$ to fine-tune the Pixart- $\\Sigma$ [7] for 15 steps. For the baseline methods, LoRA and DoRA are assigned a rank of $r = 16$ , while OFT is set to $r = 4$ . We generated five images for each prompt at a resolution of $1024\\times 1024$ .", + "bbox": [ + 511, + 411, + 908, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We measure concept alignment by computing the mean cosine similarity between the generated image's CLIP embeddings [44] and those of the original concept images. We quantify image diversity using the Vendi score [13] with DINOv2 embeddings [38]. For text alignment, we calculate the average cosine similarity in the CLIP feature space [44]. ImageReward [59] is a general-purpose text-to-image human preference reward model that evaluates and ranks AI-generated images by scoring them based on human preferences.", + "bbox": [ + 511, + 700, + 908, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Customization results. As shown in Table 1, compared with baseline methods, our approach not only represents", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18709", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1a48b9147cf82ee65d3de01bd9dd04dafccc7a467ef22041d60b48be325123f0.jpg", + "table_caption": [ + "Table 1. Comparison with baseline methods on concept customization." + ], + "table_footnote": [], + "table_body": "
AlignmentDiversityFidelityImageReward
LoRA [17]0.2184.7070.7110.220
DoRA [26]0.2134.5400.7120.142
OFT [43]0.1903.9260.6630.137
Ours (ρ = 0.02)0.2355.2050.7020.182
Ours (ρ = 0.04)0.2315.1720.7130.431
", + "bbox": [ + 94, + 128, + 480, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the concept with a minimal set of learned atoms but also generalizes robustly to diverse scenarios. In particular, our method achieves higher text-to-image alignment and diversity scores than most other PEFT baselines, while still preserving a relatively high fidelity score. This indicates that our approach effectively captures essential features of each concept without sacrificing the quality of customization text-to-image generation.", + "bbox": [ + 89, + 247, + 483, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of parameters in different methods. We compare the trainable parameters of our method and the baselines. The pre-trained PixArt model has around 616M parameters in total. Among the fine-tuning methods, both LoRA and DoRA have around 4M trainable parameters. OFT uses around 37M trainable parameters, and our method trains around 17M parameters.", + "bbox": [ + 89, + 387, + 483, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Abalation Study", + "text_level": 1, + "bbox": [ + 89, + 503, + 253, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of sparsity in the coefficients. We analyze the impact of varying sparsity levels in the coefficients through a concept customization experiment, evaluating text alignment, image diversity, and image fidelity. As shown in Figure 7, higher sparsity (low density) enhances text alignment and image diversity but results in slightly lower image fidelity. In contrast, lower sparsity (high density) improves fidelity, indicating that using more atoms allows for a more precise representation of the concept from the source image. In our experiment, we adopt both $\\rho = 0.02$ and $\\rho = 0.04$ for either better diversity or better fidelity.", + "bbox": [ + 89, + 525, + 483, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of member numbers $M$ . We also examine the impact of the number of members, i.e., the dictionary size, on the generated results while maintaining the same level of sparsity. As shown in Table 2, increasing the number of members improves image fidelity, whereas a smaller dictionary size reduces fidelity but enhances text alignment and image diversity. In our experiment, we choose $M = 256$ to provide balanced fidelity and diversity.", + "bbox": [ + 89, + 720, + 483, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparsity coefficients provide more stable adaptation. In this experiment, we compare our method with LoRA on the image editing task. As we discussed in Section 3.5,", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/395defdb074916240743b6a3b1336baf7c6175f7c76b1b0811432ebf039c142c.jpg", + "image_caption": [ + "Figure 7. The influence of different density $\\rho$ on the image diversity, fidelity and text-to-image alignment." + ], + "image_footnote": [], + "bbox": [ + 522, + 95, + 898, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ede1e65e994380147004cdd7b9f00e113559aec4fdf0ed46cfe45365e263b20.jpg", + "table_caption": [ + "Table 2. The influence of different member numbers $M$ . As the number of members increases,alignment and diversity scores decrease, while the fidelity score improves." + ], + "table_footnote": [], + "table_body": "
ρ = 0.01AlignmentDiversityFidelity
M = 160.2485.8670.547
M = 2560.2494.3410.650
M = 10240.2413.7330.672
M = 20480.2383.5250.692
", + "bbox": [ + 575, + 359, + 836, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LoRA adapts the coefficients in a low-rank manner, whereas our method updates them sparsely. Both LoRA and our method produce the updated feature $\\Delta \\mathbf{O}$ with rank 1, where LoRA achieves this by setting $r = 1$ and our method enforces only one active atom in $\\mathbf{D}$ . As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. However, when applying the fine-tuned weights to generate images with a slightly modified prompt, LoRA becomes unstable, failing to produce meaningful images. In contrast, our method remains robust, consistently generating a concept that aligns with the updated text prompt.", + "bbox": [ + 511, + 453, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce to frame model fine-tuning as a sparse coding problem, where the adapted feature representation is constructed using a dictionary of feature atoms. We propose a fine-tuning framework that leverages a combination of feature dictionary atoms, demonstrating that tuning them enhances control over the adaptation process. We validate the effectiveness of our approach across various generative tasks. By representing concepts with only a few essential atoms, our method proves to be highly adaptable and can be effectively applied to image editing, style mixing, and concept personalization. Our study provides a new perspective on the mechanics of fine-tuning, paving the way for more controllable adaptation strategies in large-scale pretrained models.", + "bbox": [ + 511, + 688, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18710", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. Decomposing and interpreting image representations via text in vits beyond CLIP. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[2] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2009. 3", + "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2", + "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 5", + "[5] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Chris Olah. Towards monoseismicity: Decomposing language models with dictionary learning. 2023. Transformer Circuits Thread. 2", + "[6] Emmanuel J Candès et al. Compressive sampling. In Proceedings of the international congress of mathematicians, pages 1433-1452, 2006. 3", + "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91, 2024. 6, 7, 2", + "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. International Conference on Learning Representations, 2025. 2", + "[9] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 1", + "[10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1", + "[12] Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. Transcoders enable fine-grained interpretable circuit analysis for language models. 2024. AI Alignment Forum. 2", + "[13] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7", + "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 2023. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In CVPR, 2023. 7", + "[16] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In European Conference on Computer Vision, 2024. 2", + "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 7, 8", + "[18] Teng Hu, Jiangning Zhang, Ran Yi, Hongrui Huang, Yabiao Wang, and Lizhuang Ma. Sara: High-efficient diffusion model fine-tuning with progressive sparse low-rank adaptation. arXiv preprint arXiv:2409.06633, 2024. 2", + "[19] Jeeyung Kim, Erfan Esmaeili, and Qiang Qiu. Text embedding is not all you need: Attention control for text-to-image semantic alignment with text self-attention maps. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8031–8040, 2025. 2", + "[20] Taewook Kim, Wei Chen, and Qiang Qiu. Learning to customize text-to-image diffusion in diverse context. arXiv preprint arXiv:2410.10058, 2024. 2", + "[21] Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. Foundations and Trends® in Machine Learning, 2019. 5", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1", + "[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 6", + "[24] Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference. arXiv preprint arXiv:2502.20766, 2025. 2", + "[25] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 2", + "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 1, 7, 8", + "[27] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 2", + "[28] Yiyang Liu, James Chenhao Liang, Ruixiang Tang, Yugyung Lee, Majid Rabbani, Sohail Dianat, Raghuveer Rao, Lifu Huang, Dongfang Liu, Qifan Wang, et al. Re-imagining" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18711", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multimodal instruction tuning: A representation view. arXiv preprint arXiv:2503.00723, 2025. 2", + "[29] Yang Luo, Xiaozhe Ren, Zangwei Zheng, Zhuo Jiang, Xin Jiang, and Yang You. Came: Confidence-guided adaptive memory efficient optimization. arXiv preprint arXiv:2307.02047, 2023. 2", + "[30] Julien Mairal, Jean Ponce, Guillermo Sapiro, Andrew Zisserman, and Francis Bach. Supervised dictionary learning. Advances in neural information processing systems, 2008. 1", + "[31] Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647, 2024. 2", + "[32] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1", + "[33] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1", + "[34] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 2", + "[35] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 2", + "[36] Zichen Miao, Wei Chen, and Qiang Qiu. Coeff-tuning: A graph filter subspace view for tuning attention-based large models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20146-20157, 2025. 2", + "[37] Bruno A Olshausen and David J Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, pages 607-609, 1996. 1", + "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7", + "[39] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1", + "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In CVPR, 2023. 6", + "[41] Vishal Purohit, Matthew Repasky, Jianfeng Lu, Qiang Qiu, Yao Xie, and Xiuyuan Cheng. Posterior sampling via Langevin dynamics based on generative priors. arXiv preprint arXiv:2410.02078, 2024. 2", + "[42] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 1, 7, 8", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 7", + "[45] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, 2024. 3", + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1", + "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 6, 7, 3", + "[48] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuzhhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2024. 3", + "[49] Viacheslav Surkov, Chris Wendler, Mikhail Terekhov, Justin Deschenaux, Robert West, and Caglar Gulcehre. Unpacking sdxl turbo: Interpreting text-to-image models with sparse autoencoders. arXiv preprint arXiv:2410.22366, 2024. 2", + "[50] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International conference on machine learning, 2020. 2", + "[51] Nathaniel Tomczak and Sanmukh Kuppannagari. Longer attention span: Increasing transformer context length with sparse graph processing techniques. arXiv preprint arXiv:2502.01659, 2025. 2", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 1, 2, 3", + "[53] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 2", + "[54] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1", + "[55] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1", + "[57] Muling Wu, Wenhao Liu, Xiaohua Wang, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaqing Zheng, and Xuanjing Huang. Advancing parameter efficiency in fine-tuning via representation editing. 2024. 2", + "[58] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. 2024. 2", + "[59] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 7", + "[60] Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, and Wangding Zeng. Native sparse attention: Hardware-aligned and natively trainable sparse attention, 2025. 2", + "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. In European Conference on Computer Vision, 2024. 2", + "[62] Yasi Zhang, Peiyu Yu, and Ying Nian Wu. Object-conditioned energy-based attention map alignment in text-to-image diffusion models. In European Conference on Computer Vision, 2024. 2", + "[63] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 6, 7", + "[64] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 628 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18713", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_model.json b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0cb8c0dc40b53fe2025a9f7b49d42ed700f7490a --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.791, + 0.153 + ], + "angle": 0, + "content": "Sparse Fine-Tuning of Transformers for Generative Tasks" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.181, + 0.682, + 0.217 + ], + "angle": 0, + "content": "Wei Chen, Jingxi Yu, Zichen Miao, Qiang Qiu Purdue University, IN, USA" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.218, + 0.737, + 0.235 + ], + "angle": 0, + "content": "{chen2732, yu667, miaoz, qqiu}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.302, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Large pre-trained transformers have revolutionized artificial intelligence across various domains, and fine-tuning remains the dominant approach for adapting these models to downstream tasks due to the cost of training from scratch. However, in existing fine-tuning methods, the updated representations are formed as a dense combination of modified parameters, making it challenging to interpret their contributions and understand how the model adapts to new tasks. In this work, we introduce a fine-tuning framework inspired by sparse coding, where fine-tuned features are represented as a sparse combination of basic elements, i.e., feature dictionary atoms. The feature dictionary atoms function as fundamental building blocks of the representation, and tuning atoms allows for seamless adaptation to downstream tasks. Sparse coefficients then serve as indicators of atom importance, identifying the contribution of each atom to the updated representation. Leveraging the atom selection capability of sparse coefficients, we first demonstrate that our method enhances image editing performance by improving text alignment through the removal of unimportant feature dictionary atoms. Additionally, we validate the effectiveness of our approach in the text-to-image concept customization task, where our method efficiently constructs the target concept using a sparse combination of feature dictionary atoms, outperforming various baseline fine-tuning methods. We provide more details at our project page." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.724, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In recent years, pre-trained large models have revolutionized artificial intelligence across domains like natural language processing and computer vision [11, 22, 46, 52]. Given the prohibitive computational resources required for training these models from scratch, fine-tuning has emerged as the dominant paradigm for adapting pre-trained models to specific downstream tasks, enabling efficient knowledge transfer while minimizing computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Various parameter-efficient fine-tuning (PEFT) methods [17, 26, 43] have been developed to tune the large trans" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.905, + 0.437 + ], + "angle": 0, + "content": "formers by updating a small subset of parameters while keeping large pre-trained weights frozen. A representative family of these approaches [17, 26] utilizes lightweight parameterizations, such as low-rank decomposition, to adapt the weight matrices. However, despite being lightweight, these methods obtain the updated features with a dense matrix multiplication between the input and the adapted weights, making it challenging to interpret how the individual updated parameters contribute to the newly learned representations and to identify the new knowledge gained through fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.905, + 0.621 + ], + "angle": 0, + "content": "To tackle this challenge, we take inspiration from sparse coding [9, 30, 32, 33, 37, 39, 42, 54-56], in which signals are expressed as a sparse combination of basic elements, i.e., dictionary atoms. In this paper, we frame model fine-tuning as a sparse coding problem, where the adapted feature is constructed using a dictionary of feature atoms. These feature atoms are selectively combined with sparse coefficients, ensuring that only a small subset of atoms contributes to the updated representation. This sparsity not only disentangles interactions between features but also allows each atom to capture a distinct aspect of the downstream data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.905, + 0.747 + ], + "angle": 0, + "content": "We first analyze with a toy experiment that the tuned feature dictionary atoms function as building blocks of the adapted representation, and tuning these atoms allows for effective adaptation to downstream tasks. We demonstrate that with pre-trained coefficients and atoms, fine-tuning the atoms effectively captures the representation of downstream tasks. Representing features through individual atoms enhances controllability in the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We validate our method on generative fine-tuning tasks. We show that the target concept can be effectively represented using feature dictionary atoms after fine-tuning the pre-trained model within our framework. Compared to baseline methods, our approach allows the selection of specific atoms to refine the generated image. This adaptability improves performance on image editing tasks by enhancing editing results through the removal of unimportant atoms. We also apply our approach to customized text-to-image tasks, showing that fine-tuning atoms effectively captures" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.361, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.316, + 0.237, + 0.327 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.631, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.315, + 0.508, + 0.327 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.108, + 0.902, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.316, + 0.779, + 0.327 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 1. (a) Our method leverages sparse coding and enforces high sparsity in adapted feature \\(\\Delta \\mathbf{O}\\), ensuring that the output feature is constructed using only a few dictionary atoms \\(\\mathbf{D}\\). (b) The pre-trained model initially generates \"8\". It generates \"3\" after integrating the learned dictionary \\(\\mathbf{D}\\) which is fine-tuned on \"3\". (c) The overall transformation of the generated digit from \"8\" to \"3\" using the dictionary atoms. The red indicates an increase in value at a specific location and blue indicates a decrease. The total influence is decomposed into contributions from atoms at different layers: \\(\\Delta \\mathbf{O}^{(1)}\\) represents the combined effect of 8 atoms in the first layer, while \\(\\Delta \\mathbf{O}^{(2)}\\) accounts for the 4 atoms in the second layer. The details of this experiment is provided in Section 3.3" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.483, + 0.5 + ], + "angle": 0, + "content": "personal concepts while preserving the capabilities of pretrained models. Our method outperforms various baseline approaches, demonstrating its effectiveness in controlled adaptation for generative modeling." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.501, + 0.403, + 0.514 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.517, + 0.483, + 0.56 + ], + "angle": 0, + "content": "- We formulate large transformer fine-tuning as a sparse coding problem, where the adapted feature representation is a sparse combination of a dictionary of feature atoms." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.62 + ], + "angle": 0, + "content": "- We introduce a fine-tuning framework that leverages a combination of dictionary atoms and demonstrates that tuning atoms enhances control over the adaptation process." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.622, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- We validate the effectiveness of our approach in generative tasks by demonstrating the improvement in image editing and concept customization." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.517, + 0.483, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.233, + 0.699 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Sparse Representation and Attention Recent research has demonstrated the significant role of sparsity in enhancing both interpretability and computational efficiency in generative models through various approaches including sparse autoencoders for feature extraction [49], sparse feature circuits for improved generalization [31], transcoders for detailed circuit analysis [12], dictionary learning for monosemantic representations [5], and selective parameter tuning strategies [18]. Complementary work has addressed the quadratic computational complexity of standard attention mechanisms through sparse attention techniques that compute attention scores for only a subset of token pairs [3, 10, 50, 52], with optimizations including" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.561 + ], + "angle": 0, + "content": "dynamic pattern adjustment [24], hardware-aligned implementations [60], and graph-based conceptualizations [51], collectively enabling transformers to process significantly longer sequences while maintaining model quality despite challenges in balancing efficiency with performance in complex long-range dependency scenarios. Compared to previous works, our method applies sparse coding for finetuning transformer-based models in generative tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.909, + 0.752 + ], + "angle": 0, + "content": "Representation Tuning Recent representation tuning approaches operate directly on model representations rather than weights, achieving parameter efficiency while enhancing controllability, and interpretability across various model types and tasks [1, 8, 27, 28, 36, 53, 57, 58, 64]. Unlike traditional PEFT approaches that impact representations indirectly through weight adjustments, representation tuning methods preserve the original pre-trained model weights while directly modifying the output representations. In comparison, our work studies the sparsity in the updated representation and frames it as a dictionary learning problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Personalized Text-to-Image Generation Extensive research has been conducted on aligning text prompts and generated concepts [19, 20, 23, 25, 34, 35, 41, 47]. Zhang et al. [62] developed an object-conditioned energy-based attention map alignment technique in diffusion models, enhancing control over specific objects in generated images. Another study by Zhang et al. [61] improved semantic fidelity in text-to-image synthesis through attention regulation. Hao et al. [16] introduces ConceptExpress, an unsu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18704" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "pervised method for extracting personalized concepts from a single image. Ren et al. [45] addressed memorization issues in text-to-image diffusion models by analyzing cross-attention mechanisms, proposing solutions to reduce overfitting to user-specific data. We approach this problem by adopting the ideas from sparse coding and learning the target concept as a sparse combination of coefficients and dictionary atoms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.225, + 0.182, + 0.24 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.25, + 0.221, + 0.266 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.395 + ], + "angle": 0, + "content": "Multi-head attention (MHA). The attention layer [52] transforms the input sequence \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}\\) to the output sequence \\(\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}\\), where \\(N\\) denotes the sequence length, \\(C_i\\) and \\(C_o\\) are the dimension of input and output features. The attention layer projects the input as \\(\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q\\), \\(\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k\\), \\(\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v\\) using the corresponding projection matrices \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\), \\(\\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}\\), and calculates the attention map," + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.405, + 0.483, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = \\operatorname {a t t n} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.482 + ], + "angle": 0, + "content": "The attention map \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\) captures the token-wise relationship by doing inner-product in a space transformed by \\(\\mathbf{W}_q, \\mathbf{W}_k\\). Then the output of the attention layer is," + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.493, + 0.483, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\mathbf {A X W} _ {v}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Multi-head attention extends this by allowing multiple attention mechanisms to work in parallel, with each head independently learning attention patterns. The project matrices for each head are \\(\\mathbf{W}_q^{(h)},\\mathbf{W}_k^{(h)},\\mathbf{W}_v^{(h)}\\in \\mathbb{R}^{C_i\\times \\frac{C_o}{H}}\\) where \\(h = 1,\\dots ,H\\) and \\(H\\) is the number of heads. The multi-head attention represents as," + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.624, + 0.483, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.71 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}\\) and \\(\\mathbf{W}_{vo}^{(h)} = \\mathbf{W}_v^{(h)}\\mathbf{W}_o^{(h)\\intercal},\\mathbf{W}_o^{(h)}\\in\\) \\(\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}\\) \\(\\mathbf{W}_{vo}^{(h)}\\in \\mathbb{R}^{C_i\\times C_o}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Sparse coding. Sparse coding is a signal representation technique that models a signal as a linear combination of a small number of basis functions from a dictionary. Given a signal \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}\\), the goal of sparse coding is to find a dictionary \\(\\mathbf{D} \\in \\mathbb{R}^{m \\times C_i}\\) (where \\(m > C_i\\) is the number of atoms) and a sparse coefficient vector \\(\\mathbf{S} \\in \\mathbb{R}^{N \\times m}\\) such that the signal can be approximated as:" + }, + { + "type": "equation", + "bbox": [ + 0.252, + 0.845, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} \\approx \\mathbf {S D}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}\\) has only a few nonzero elements, ensuring sparsity. The sparse coding problem is often formulated as an" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.666, + 0.107 + ], + "angle": 0, + "content": "optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.112, + 0.905, + 0.135 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {S}} \\| \\mathbf {X} - \\mathbf {S D} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {S} \\| _ {1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.907, + 0.275 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a regularization parameter controlling the tradeoff between sparsity and reconstruction error. The Proximal gradient method is a common approach to solving this optimization problem, including Iterative Shrinkage-Thresholding Algorithm (ISTA) and its accelerated version (FISTA) [2], which utilize proximal operators for efficient iterative updates. For sparse coding with \\(\\ell_1\\) regularization, the corresponding proximal operator is the soft-thresholding function:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.281, + 0.907, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\operatorname {p r o x} _ {\\lambda \\| \\cdot \\| _ {1}} (\\mathbf {S}) = \\operatorname {s i g n} (\\mathbf {S}) \\odot \\max (| \\mathbf {S} | - \\lambda , 0), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.906, + 0.378 + ], + "angle": 0, + "content": "where \\(\\odot\\) is Hadamard product. It encourages sparsity by shrinking small values of \\(\\mathbf{S}\\) toward zero. Sparse coding is widely used in signal processing, machine learning, and neuroscience, providing efficient and interpretable representations of data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.385, + 0.704, + 0.401 + ], + "angle": 0, + "content": "3.2. Sparse Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.575 + ], + "angle": 0, + "content": "In this work, we conceptualize model fine-tuning as a sparse coding problem, i.e., the adapted features are constructed using a dictionary of feature atoms. As our method focuses on the transformer architecture, we first propose an interpretation of the attention operation by viewing the product between the attention map and value matrices as a combination of dictionary atoms with dictionary coefficients. We then propose a fine-tuning approach by representing the adapted features as a learned sparse combination of dictionary atoms with sparse coefficients, representing the fine-tuned representations' underlying sparsity nature." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.591, + 0.906, + 0.651 + ], + "angle": 0, + "content": "A dictionary view of attention. For single head attention (2), each row of \\(\\mathbf{A}\\) can be interpreted as coefficients for combining \\(\\mathbf{V}\\) into the output \\(\\mathbf{O}\\), i.e., \\(\\mathbf{O}_i = \\sum_j\\mathbf{A}_{ij}\\mathbf{V}_j\\). In the case of multi-head attention (3)," + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.655, + 0.905, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)} = \\left[ \\begin{array}{l l l} \\mathbf {A} ^ {(1)} & \\dots & \\mathbf {A} ^ {(H)} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(1)} \\\\ \\vdots \\\\ \\mathbf {X} \\mathbf {W} _ {v o} ^ {(H)} \\end{array} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.907, + 0.901 + ], + "angle": 0, + "content": "it can be represented as a product of a composite coefficient \\(\\left[\\mathbf{A}^{(1)},\\dots ,\\mathbf{A}^{(H)}\\right]\\in \\mathbb{R}^{N\\times NH}\\), and an extensive dictionary \\(\\left[\\mathbf{XW}_{vo}^{(1)},\\dots ,\\mathbf{XW}_{vo}^{(H)}\\right]^{\\intercal}\\in \\mathbb{R}^{NH\\times C_o}\\). Each row of the output \\(\\mathbf{O}\\) is a linear combination of \\(\\mathbf{XW}_{vo}\\), i.e., \\(\\mathbf{O}_i = \\sum_{h,j}\\mathbf{A}_{ij}^{(h)}(\\mathbf{XW}_{vo}^{(h)})_j\\). When represented as a linear combination of dictionary elements, multi-head attention and single-head attention follow the same formulation. In the following discussion, we focus on the single-head attention formulation, which can be seamlessly extended to the multi-head attention case." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18705" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "A sparse dictionary view of attention. From the perspective of sparse coding, the feature is expressed as SD, where S represents the sparse coefficients and D consists of dictionary atoms that remain data-independent once learned. Although attention can be expressed as a linear combination of atoms, it does not inherently exhibit sparsity. To fully leverage the advantages of sparse representations, we propose to represent the attention mechanism (2) in the form of sparse coding," + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.242, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\mathbf {S D} = \\sigma_ {\\lambda} \\left(\\mathbf {A X W} _ {s}\\right) \\mathbf {D}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.332 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\lambda}(\\cdot)\\) can be soft-thresholding function or \\(\\mathrm{ReLU}(x - \\lambda)\\) to ensure sparsity of \\(\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s) \\in \\mathbb{R}^{N \\times C_o}\\), \\(\\mathbf{W}_s \\in \\mathbb{R}^{C_i \\times M}\\) and \\(\\mathbf{D} \\in \\mathbb{R}^{M \\times C_o}\\) is the feature dictionary atoms, \\(M\\) is the number of atoms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.409 + ], + "angle": 0, + "content": "With this formulation, we obtain the feature dictionary atoms \\(\\mathbf{D}\\) (we use the term \"atoms\" for simplification in this paper), which are the building blocks for feature representation \\(\\mathbf{O}\\). We illustrate our formulation in Figure 1 (a). This formulation provides the following properties:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.409, + 0.484, + 0.514 + ], + "angle": 0, + "content": "- The sparse coefficients \\(\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)\\) ensure that the feature representation \\(\\mathbf{O}\\) is combined by only a few \\(\\mathbf{D}\\). We present an extreme example to illustrate how atoms enhance interpretability: \\(\\mathbf{O}_i = \\sum_{j=1}^{M} \\mathbf{S}_{ij} \\mathbf{D}_j\\) can be simplified as \\(\\mathbf{O}_i = \\mathbf{S}_{ij^*} \\mathbf{D}_{j^*}\\) if all elements of \\(\\mathbf{S}_i\\) is 0 except the element at index \\(j^*\\), and only \\(\\mathbf{D}_{j^*}\\) contributes to the output feature \\(\\mathbf{O}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.515, + 0.484, + 0.59 + ], + "angle": 0, + "content": "- The sparse coefficients \\(\\mathbf{S}\\) are data-dependent, while the feature dictionary atoms \\(\\mathbf{D}\\) are data-independent. While different inputs generate different coefficients, the output features are always constructed using the same feature dictionary atoms." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.59, + 0.484, + 0.651 + ], + "angle": 0, + "content": "- We can adjust the size of the feature dictionary by varying \\( M \\) to handle tasks with different difficulties. Large \\( M \\) leads to an overcomplete dictionary while small \\( M \\) produces an undercomplete dictionary." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.409, + 0.484, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.785 + ], + "angle": 0, + "content": "Fine-tuning via sparse dictionary learning. The above formulation frames feature representation through the lens of sparse coding. In the context of large model fine-tuning, we freeze the pre-trained parameters to make feature representations \\(\\mathbf{O}\\) of the pre-trained model fixed. Instead, we introduce a sparse coding perspective to encode the adapted feature representation \\(\\Delta \\mathbf{O}\\), which is learned from downstream tasks. Specifically, we have," + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.797, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} + \\Delta \\mathbf {O} = \\mathbf {O} + \\sigma_ {\\lambda} (\\mathbf {A X W} _ {s}) \\mathbf {D}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With this formulation, the pre-trained weights remain unchanged, preserving the capabilities of the pre-trained model. The adapted feature \\(\\Delta \\mathbf{O}\\) is represented as a linear combination of feature dictionary atoms \\(\\mathbf{D}\\), using sparse coefficients \\(\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)\\). This formulation offers" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 2. (a) Adjusting only the coefficients fails to adapt the model to the downstream task, whereas (b) modifying only the atoms enables a perfect fit to the new task composed of high-frequency signals. (c) Fine-tuning the atoms effectively transitions the model from low-frequency to high-frequency representations. They encode the representation of the downstream task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.448 + ], + "angle": 0, + "content": "a framework for understanding how \\(\\mathbf{D}\\) contributes to the adapted features in the fine-tuned model. We illustrate our method in Figure 1 (a)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.776, + 0.473 + ], + "angle": 0, + "content": "3.3. Analysis of Dictionary Atoms" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.48, + 0.906, + 0.51 + ], + "angle": 0, + "content": "In this section, we demonstrate distinct properties of atoms \\(\\mathbf{D}\\) and sparse coefficients \\(\\mathbf{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Features atoms are key to fine-tuning. We analyze the distinct behaviors of coefficients and atoms, and discover that atoms play a crucial role in adapting to new tasks. To explore this within the context of an attention block, we perform a toy experiment where masked one-dimensional signals are reconstructed using an attention layer. The signals, with a length of 64, are synthesized by selecting 5 random Fourier bases, with a low frequency ranging from \\(\\frac{0}{64}\\) to \\(\\frac{24}{64}\\). The objective is to reconstruct the signal after masking out \\(75\\%\\) of its values with only 16 randomly sampled observations. After we have a pre-trained model on the low-frequency signals, we fine-tune the model to a new task where signals are constructed with high-frequency Fourier bases ranging from \\(\\frac{25}{64}\\) to \\(\\frac{32}{64}\\). In this experiment, we test two different cases: (1) tuning only coefficients on the high-frequency data with atoms fixed, and (2) tuning only atoms on the high-frequency data with coefficients fixed. As illustrated in Figure 2, when adapting to high-frequency signals, modifying only the atoms while keeping the coefficients fixed allows the model to effectively generalize to the new task. In contrast, updating only the coefficients is insufficient for proper adaptation. Building on this observation, we hypothesize that adjusting atoms is more suitable than adjusting coefficients for adapting to a new task. We observe from Figure 2 (c) that atoms are capable of encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18706" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "the representation of the downstream task. The experiment details are in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.148, + 0.486, + 0.542 + ], + "angle": 0, + "content": "Identifying the importance of feature atoms. We analyze the impact of atoms on the newly learned concept by visualizing both the influence of individual atoms and their combined effect. In this experiment, we adopt a transformer-based variational autoencoder (VAE) [21] with a two-layer decoder. The experiment process is illustrated in Figure 1 (b). We first pre-train the model on an image-denoising task using a subset of the MNIST dataset containing digits 5 to 9. We then fine-tune the model on a downstream denoising task with only digit 3 by learning a dictionary for each layer with \\( M = 100 \\). In this scenario, the digit \"3\" is the newly learned concept. Given the same noisy latent, the decoder outputs an \"8\" without incorporating fine-tuned feature atoms, whereas it generates a \"3\" when the fine-tuned feature atoms are applied. The model requires only 12 atoms to shift the generated digits from 8 to 3. The influence of each atom is visualized in Figure 1 (c), with the overall impact represented as \\( \\Delta \\mathbf{O} \\), where the red area indicates an increase in pixel value at a specific location and the blue area represents a decrease. It is straightforward to see that these 12 feature atoms enhance the regions corresponding to \"3\" while suppressing certain portions of \"8\". Additionally, Figure 1 (c) illustrates the individual behavior of each atom, showing how different atoms affect distinct parts of the generated digit. The experiment details are in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.265, + 0.573 + ], + "angle": 0, + "content": "3.4. Design Variations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.485, + 0.822 + ], + "angle": 0, + "content": "In this section, we outline the various design choices implemented in our experiments. Our method achieves sparse coefficient \\(\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)\\) by utilizing non-linear function \\(\\sigma_{\\lambda}(\\cdot)\\). In the literature on sparse coding [4], there are three different methods: (i) soft-thresholding function, \\(\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{prox}_{\\lambda \\| \\cdot \\|_1}(\\mathbf{x}) = \\mathrm{sign}(\\mathbf{x}) \\odot \\max(|\\mathbf{x}| - \\lambda, 0)\\), (ii) ReLU, \\(\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{ReLU}(\\mathbf{x} - \\lambda) = \\max(\\mathbf{x} - \\lambda, 0)\\), and (iii) top-\\(k\\) activation selection, which selects the most significant features of top \\(k\\) contributing to the learned representation. In practice, we observe that top-\\(k\\) activation selection maintains a consistent level of sparsity across different inputs. In contrast, methods (i) soft-thresholding function and (ii) ReLU exhibit sensitivity to input values, leading to variations in sparsity. However, methods (i) and (ii) offer control over the coefficient values, ensuring they stay above a specified threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we find that when \\(\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)\\) is highly sparse, the resulting product \\(\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)\\) also retains sparsity. In modern transformers, computation of \\(\\mathbf{A}\\) and \\(\\mathbf{X}\\mathbf{W}_v\\) are performed separately. For simplicity in implementation, we adopt the formulation \\(\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.729, + 0.108 + ], + "angle": 0, + "content": "3.5. Analysis of Our Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Comparison with LoRA. When representing attention mechanism (1-3) as a linear combination of dictionary atoms, updating \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\) is equivalent to updating the coefficients while updating \\(\\mathbf{W}_v\\), \\(\\mathbf{W}_o\\) is equivalent to updating the atoms. LoRA applies adaptation on \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\), \\(\\mathbf{W}_v\\), \\(\\mathbf{W}_o\\) with a low-rank matrices \\(\\mathbf{W}_A \\in \\mathbb{R}^{C_i \\times r}\\), \\(\\mathbf{W}_B \\in \\mathbb{R}^{r \\times C_o}\\), where \\(r\\) represents the rank. LoRA adapts both coefficients and atoms in a low-rank way. In comparison, our method adapts the coefficients with \\(\\mathbf{W}_s\\) and obtains sparse coefficients \\(\\sigma_{\\lambda}(\\mathbf{AXW}_s)\\), while representing atoms with \\(\\mathbf{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.909, + 0.538 + ], + "angle": 0, + "content": "Parameters and computational cost. To adapt the representation at each layer, we need additional parameters for \\(\\mathbf{W}_s\\) and \\(\\mathbf{D}\\), which are \\((C_i + C_o)M\\). After the fine-tuning, considering the density of the coefficients \\(\\rho\\), i.e., \\(1 - \\rho\\) is sparsity, the storage required for the parameters is \\(C_i M + \\rho C_o M\\). In comparison, LoRA requires \\(4(C_i + C_o)r\\) additional parameters. The FLOPs required by our method are \\(2N M C_i + 2\\rho N M C_o\\). LoRA requires \\(8C_i C_o r\\) FLOPs. For example, if \\(C_i = C_o = 1024\\), \\(M = 256\\), \\(r = 16\\), \\(N = 1024\\), \\(\\rho = 0.01\\), our method requires 264k parameters for storage and 0.54GFLOPs for computation, while LoRA requires 131k parameters and 0.13GFLOps. Our method involves slightly higher computational costs than LoRA, as it accounts for feature interactions, whereas LoRA only updates the weights. Next, we demonstrate that by leveraging sparse feature representations, our approach achieves greater robustness compared to LoRA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our method provides more stable adaptation. We compare an extreme case between LoRA and our method, by considering the attention in a dictionary form \\(\\Delta \\mathbf{O} = \\mathbf{A}\\mathbf{V}\\) or \\(\\Delta \\mathbf{O} = \\mathbf{SD}\\). In this experiment, we examine the difference between updating coefficients using a low-rank approach versus a sparse approach. Both LoRA and our method produce the updated feature \\(\\Delta \\mathbf{O}\\) with rank 1, where LoRA achieves this by setting \\(r = 1\\) and our method enforces only one active atom in D. The key difference lies in how the coefficients are updated—LoRA adapts them in a low-rank manner, whereas our method updates them sparsely. As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. LoRA fails to produce meaningful images with a slightly modified prompt, while our method remains robust, consistently generating a concept that aligns with the updated text prompt. We attribute the effectiveness of our approach to the sparse coefficients produced by \\(\\sigma_{\\lambda}(\\mathbf{AXW}_s)\\). With a different text prompt, the values in S either remain zero or take on a significant value when receiving the proper input, ensuring that the generated result remains coherent. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18707" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.09, + 0.845, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 3. The pre-trained model can generate the target concept when integrated with the learned dictionary, even with about 40 atoms. With just four atoms, the core structure of the concept can still be generated. Adjusting the number of selected atoms influences the generated results, offering a controllable approach to image editing. Additional discussions of these figures are in Section 4.1." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.414, + 0.473, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.653, + 0.486, + 0.738 + ], + "angle": 0, + "content": "Figure 4. The role of sparsity: Our method utilizes the same number of atoms as LoRA, yet LoRA struggles to adapt to the target prompt effectively. Sparse coding ensures that atoms are either activated or inactive, providing a structured and interpretable adaptation process. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.224, + 0.77 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of our method in image generation tasks. We demonstrate that feature dictionary atoms effectively represent the target concept after fine-tuning the pre-trained model. Compared to baselines, our approach enables selective atom adjustment, improving image editing by refining results and enhancing text-to-image concept customization while preserving pre-trained model capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.831, + 0.427 + ], + "angle": 0, + "content": "4.1. Representing Concepts Using Atoms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Experimental settings. In this experiment, we utilize PixArt- \\(\\Sigma\\) [7], a Diffusion Transformer (DiT) [40] model consisting of 28 transformer blocks. We fine-tune the model on a single image for 60 epochs using our method and reconstruct the corresponding concept using the same prompt used during fine-tuning. The selected images are drawn from DreamBooth [47], Custom Diffusion [23]. We control the number of active atoms and visualize the variation of the generated results with different atoms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.689 + ], + "angle": 0, + "content": "The concept is represented by a set of atoms. As shown in Figure 3, after fine-tuning our method on the target concept, its representation can be effectively captured using a small set of atoms, such as 40 atoms. Among these, just 4 atoms are sufficient to construct the core structure of the concept, including key attributes like pose and color. We provide details descriptions of these images in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.663, + 0.714 + ], + "angle": 0, + "content": "4.2. Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Building on our previous observation that adjusting the number of atoms influences the generated results, we explore leveraging this property for image editing. Image editing tasks aim to modify an image using a generative model while retaining as many details as possible from the original [63]. However, certain editing tasks pose challenges in maintaining most of the original details while achieving effective modifications. For example, altering the style can significantly change the overall appearance of the image. We argue that successful image editing should preserve the main components of the image while applying modifications selectively. Since our method represents the core" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18708" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.091, + 0.47, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.438, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Figure 5. Results on image editing. Our method shows improved results by better following editing prompts while preserving the key components of the source images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.561 + ], + "angle": 0, + "content": "components of a concept using only a few atoms, it is well-suited for image editing tasks, ensuring that key structural elements remain intact while allowing for controlled transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Experimental settings. In our experiment, we select data from previous work [63] and evaluate different editing tasks, including background modification, scene alteration, and pose adaptation. Previous works [15, 63] fine-tune models to learn a concept and modify its appearance by adjusting the prompt. However, these approaches have not yet been applied to DiT. Following their setup, we compare our method with fine-tuning approaches, such as LoRA [17], using them as baselines for evaluation. We adopt the DiT architecture and fine-tune both LoRA and our method on a single image for 60 epochs to achieve full reconstruction. After fine-tuning, we modify the image using a different prompt for editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Image editing results. The results are shown in Figure 5. Our method applies edits based on the test prompt while maintaining the core structure of the source images. In addition, we control the number of active atoms by adjusting the coefficient density \\(\\rho\\), where \\(1 - \\rho\\) represents the sparsity level. In Figure 6, as sparsity increases, only the most essential atoms remain, preserving the core concept while al" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.65, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.187, + 0.625, + 0.2 + ], + "angle": 0, + "content": "\\(\\rho = 0.01\\)" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.091, + 0.771, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.187, + 0.743, + 0.2 + ], + "angle": 0, + "content": "\\(\\rho = 0.02\\)" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.091, + 0.888, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.802, + 0.187, + 0.858, + 0.199 + ], + "angle": 0, + "content": "\\(\\rho = 0.04\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.213, + 0.907, + 0.242 + ], + "angle": 0, + "content": "Figure 6. Adjusting the density, i.e., the number of activated atoms, can change the influence on the edited image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.29 + ], + "angle": 0, + "content": "lowing the pre-trained model greater flexibility to integrate editing effects based on the modified text prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.298, + 0.731, + 0.314 + ], + "angle": 0, + "content": "4.3. Concept Customization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.907, + 0.396 + ], + "angle": 0, + "content": "In the previous section, we demonstrated that our method can efficiently represent the target concept using only a small number of atoms. In this section, we further show that even with just a few atoms, our method achieves good performance in concept customization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Experimental settings. We evaluated our approach on the DreamBooth [47] dataset, which consists of 30 subjects across 15 distinct classes. Among these, 9 are live subjects, while the remaining 21 are objects. Each subject has between 4 to 6 images, captured under varying conditions, in diverse environments, and from multiple angles. All 30 subjects were trained for 15 epochs and validated on 25 different prompts. For our experiments, we adopted the Pixart- \\(\\Sigma\\), a Diffusion Transformer (DiT) [7], as the base architecture and fine-tuned it using our proposed method. To verify the effectiveness of our approach, we compared its performance against several PEFT baselines for personalized text-to-image generation, including LoRA [17], DoRA [26], and OFT [43]. We use the AdamW optimizer with a learning rate of \\(1\\times 10^{-4}\\) to fine-tune the Pixart- \\(\\Sigma\\) [7] for 15 steps. For the baseline methods, LoRA and DoRA are assigned a rank of \\(r = 16\\), while OFT is set to \\(r = 4\\). We generated five images for each prompt at a resolution of \\(1024\\times 1024\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Evaluation metrics. We measure concept alignment by computing the mean cosine similarity between the generated image's CLIP embeddings [44] and those of the original concept images. We quantify image diversity using the Vendi score [13] with DINOv2 embeddings [38]. For text alignment, we calculate the average cosine similarity in the CLIP feature space [44]. ImageReward [59] is a general-purpose text-to-image human preference reward model that evaluates and ranks AI-generated images by scoring them based on human preferences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Customization results. As shown in Table 1, compared with baseline methods, our approach not only represents" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18709" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison with baseline methods on concept customization." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.481, + 0.221 + ], + "angle": 0, + "content": "
AlignmentDiversityFidelityImageReward
LoRA [17]0.2184.7070.7110.220
DoRA [26]0.2134.5400.7120.142
OFT [43]0.1903.9260.6630.137
Ours (ρ = 0.02)0.2355.2050.7020.182
Ours (ρ = 0.04)0.2315.1720.7130.431
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.368 + ], + "angle": 0, + "content": "the concept with a minimal set of learned atoms but also generalizes robustly to diverse scenarios. In particular, our method achieves higher text-to-image alignment and diversity scores than most other PEFT baselines, while still preserving a relatively high fidelity score. This indicates that our approach effectively captures essential features of each concept without sacrificing the quality of customization text-to-image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.495 + ], + "angle": 0, + "content": "Number of parameters in different methods. We compare the trainable parameters of our method and the baselines. The pre-trained PixArt model has around 616M parameters in total. Among the fine-tuning methods, both LoRA and DoRA have around 4M trainable parameters. OFT uses around 37M trainable parameters, and our method trains around 17M parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.254, + 0.52 + ], + "angle": 0, + "content": "4.4. Abalation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.709 + ], + "angle": 0, + "content": "The impact of sparsity in the coefficients. We analyze the impact of varying sparsity levels in the coefficients through a concept customization experiment, evaluating text alignment, image diversity, and image fidelity. As shown in Figure 7, higher sparsity (low density) enhances text alignment and image diversity but results in slightly lower image fidelity. In contrast, lower sparsity (high density) improves fidelity, indicating that using more atoms allows for a more precise representation of the concept from the source image. In our experiment, we adopt both \\(\\rho = 0.02\\) and \\(\\rho = 0.04\\) for either better diversity or better fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.484, + 0.843 + ], + "angle": 0, + "content": "The impact of member numbers \\( M \\). We also examine the impact of the number of members, i.e., the dictionary size, on the generated results while maintaining the same level of sparsity. As shown in Table 2, increasing the number of members improves image fidelity, whereas a smaller dictionary size reduces fidelity but enhances text alignment and image diversity. In our experiment, we choose \\( M = 256 \\) to provide balanced fidelity and diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Sparsity coefficients provide more stable adaptation. In this experiment, we compare our method with LoRA on the image editing task. As we discussed in Section 3.5," + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.096, + 0.9, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Figure 7. The influence of different density \\(\\rho\\) on the image diversity, fidelity and text-to-image alignment." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.308, + 0.905, + 0.351 + ], + "angle": 0, + "content": "Table 2. The influence of different member numbers \\( M \\) . As the number of members increases,alignment and diversity scores decrease, while the fidelity score improves." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.361, + 0.838, + 0.436 + ], + "angle": 0, + "content": "
ρ = 0.01AlignmentDiversityFidelity
M = 160.2485.8670.547
M = 2560.2494.3410.650
M = 10240.2413.7330.672
M = 20480.2383.5250.692
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.651 + ], + "angle": 0, + "content": "LoRA adapts the coefficients in a low-rank manner, whereas our method updates them sparsely. Both LoRA and our method produce the updated feature \\(\\Delta \\mathbf{O}\\) with rank 1, where LoRA achieves this by setting \\(r = 1\\) and our method enforces only one active atom in \\(\\mathbf{D}\\). As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. However, when applying the fine-tuned weights to generate images with a slightly modified prompt, LoRA becomes unstable, failing to produce meaningful images. In contrast, our method remains robust, consistently generating a concept that aligns with the updated text prompt." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduce to frame model fine-tuning as a sparse coding problem, where the adapted feature representation is constructed using a dictionary of feature atoms. We propose a fine-tuning framework that leverages a combination of feature dictionary atoms, demonstrating that tuning them enhances control over the adaptation process. We validate the effectiveness of our approach across various generative tasks. By representing concepts with only a few essential atoms, our method proves to be highly adaptable and can be effectively applied to image editing, style mixing, and concept personalization. Our study provides a new perspective on the mechanics of fine-tuning, paving the way for more controllable adaptation strategies in large-scale pretrained models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18710" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. Decomposing and interpreting image representations via text in vits beyond CLIP. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Chris Olah. Towards monoseismicity: Decomposing language models with dictionary learning. 2023. Transformer Circuits Thread. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[6] Emmanuel J Candès et al. Compressive sampling. In Proceedings of the international congress of mathematicians, pages 1433-1452, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.576 + ], + "angle": 0, + "content": "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\sigma\\) : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91, 2024. 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[12] Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. Transcoders enable fine-grained interpretable circuit analysis for language models. 2024. AI Alignment Forum. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[13] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[16] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[18] Teng Hu, Jiangning Zhang, Ran Yi, Hongrui Huang, Yabiao Wang, and Lizhuang Ma. Sara: High-efficient diffusion model fine-tuning with progressive sparse low-rank adaptation. arXiv preprint arXiv:2409.06633, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[19] Jeeyung Kim, Erfan Esmaeili, and Qiang Qiu. Text embedding is not all you need: Attention control for text-to-image semantic alignment with text self-attention maps. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8031–8040, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[20] Taewook Kim, Wei Chen, and Qiang Qiu. Learning to customize text-to-image diffusion in diverse context. arXiv preprint arXiv:2410.10058, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[21] Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. Foundations and Trends® in Machine Learning, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[24] Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference. arXiv preprint arXiv:2502.20766, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[25] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[27] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Yiyang Liu, James Chenhao Liang, Ruixiang Tang, Yugyung Lee, Majid Rabbani, Sohail Dianat, Raghuveer Rao, Lifu Huang, Dongfang Liu, Qifan Wang, et al. Re-imagining" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18711" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "multimodal instruction tuning: A representation view. arXiv preprint arXiv:2503.00723, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[29] Yang Luo, Xiaozhe Ren, Zangwei Zheng, Zhuo Jiang, Xin Jiang, and Yang You. Came: Confidence-guided adaptive memory efficient optimization. arXiv preprint arXiv:2307.02047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[30] Julien Mairal, Jean Ponce, Guillermo Sapiro, Andrew Zisserman, and Francis Bach. Supervised dictionary learning. Advances in neural information processing systems, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[31] Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[32] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[33] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[34] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[35] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[36] Zichen Miao, Wei Chen, and Qiang Qiu. Coeff-tuning: A graph filter subspace view for tuning attention-based large models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20146-20157, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[37] Bruno A Olshausen and David J Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, pages 607-609, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[39] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[41] Vishal Purohit, Matthew Repasky, Jianfeng Lu, Qiang Qiu, Yao Xie, and Xiuyuan Cheng. Posterior sampling via Langevin dynamics based on generative priors. arXiv preprint arXiv:2410.02078, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[42] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.243 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[45] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[48] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuzhhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[49] Viacheslav Surkov, Chris Wendler, Mikhail Terekhov, Justin Deschenaux, Robert West, and Caglar Gulcehre. Unpacking sdxl turbo: Interpreting text-to-image models with sparse autoencoders. arXiv preprint arXiv:2410.22366, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[50] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International conference on machine learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[51] Nathaniel Tomczak and Sanmukh Kuppannagari. Longer attention span: Increasing transformer context length with sparse graph processing techniques. arXiv preprint arXiv:2502.01659, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[53] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[54] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[55] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18712" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[56] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[57] Muling Wu, Wenhao Liu, Xiaohua Wang, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaqing Zheng, and Xuanjing Huang. Advancing parameter efficiency in fine-tuning via representation editing. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[58] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[59] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[60] Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, and Wangding Zeng. Native sparse attention: Hardware-aligned and natively trainable sparse attention, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[62] Yasi Zhang, Peiyu Yu, and Ying Nian Wu. Object-conditioned energy-based attention map alignment in text-to-image diffusion models. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[63] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[64] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18713" + } + ] +] \ No newline at end of file diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_origin.pdf b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2f2c29fa21f01c3f9782ec5736f62e9e4f759a1b --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/93b9d533-60f0-4a0a-a330-dae67ada8155_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1901e5a12823d0fa0655de43d30e5a57fe7e6ed30a1a577b811f75b5d1784b8d +size 26098408 diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/full.md b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3755590ce40f982f290b046192e487ba6201a736 --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/full.md @@ -0,0 +1,314 @@ +# Sparse Fine-Tuning of Transformers for Generative Tasks + +Wei Chen, Jingxi Yu, Zichen Miao, Qiang Qiu Purdue University, IN, USA + +{chen2732, yu667, miaoz, qqiu}@purdue.edu + +# Abstract + +Large pre-trained transformers have revolutionized artificial intelligence across various domains, and fine-tuning remains the dominant approach for adapting these models to downstream tasks due to the cost of training from scratch. However, in existing fine-tuning methods, the updated representations are formed as a dense combination of modified parameters, making it challenging to interpret their contributions and understand how the model adapts to new tasks. In this work, we introduce a fine-tuning framework inspired by sparse coding, where fine-tuned features are represented as a sparse combination of basic elements, i.e., feature dictionary atoms. The feature dictionary atoms function as fundamental building blocks of the representation, and tuning atoms allows for seamless adaptation to downstream tasks. Sparse coefficients then serve as indicators of atom importance, identifying the contribution of each atom to the updated representation. Leveraging the atom selection capability of sparse coefficients, we first demonstrate that our method enhances image editing performance by improving text alignment through the removal of unimportant feature dictionary atoms. Additionally, we validate the effectiveness of our approach in the text-to-image concept customization task, where our method efficiently constructs the target concept using a sparse combination of feature dictionary atoms, outperforming various baseline fine-tuning methods. We provide more details at our project page. + +# 1. Introduction + +In recent years, pre-trained large models have revolutionized artificial intelligence across domains like natural language processing and computer vision [11, 22, 46, 52]. Given the prohibitive computational resources required for training these models from scratch, fine-tuning has emerged as the dominant paradigm for adapting pre-trained models to specific downstream tasks, enabling efficient knowledge transfer while minimizing computational overhead. + +Various parameter-efficient fine-tuning (PEFT) methods [17, 26, 43] have been developed to tune the large trans + +formers by updating a small subset of parameters while keeping large pre-trained weights frozen. A representative family of these approaches [17, 26] utilizes lightweight parameterizations, such as low-rank decomposition, to adapt the weight matrices. However, despite being lightweight, these methods obtain the updated features with a dense matrix multiplication between the input and the adapted weights, making it challenging to interpret how the individual updated parameters contribute to the newly learned representations and to identify the new knowledge gained through fine-tuning. + +To tackle this challenge, we take inspiration from sparse coding [9, 30, 32, 33, 37, 39, 42, 54-56], in which signals are expressed as a sparse combination of basic elements, i.e., dictionary atoms. In this paper, we frame model fine-tuning as a sparse coding problem, where the adapted feature is constructed using a dictionary of feature atoms. These feature atoms are selectively combined with sparse coefficients, ensuring that only a small subset of atoms contributes to the updated representation. This sparsity not only disentangles interactions between features but also allows each atom to capture a distinct aspect of the downstream data. + +We first analyze with a toy experiment that the tuned feature dictionary atoms function as building blocks of the adapted representation, and tuning these atoms allows for effective adaptation to downstream tasks. We demonstrate that with pre-trained coefficients and atoms, fine-tuning the atoms effectively captures the representation of downstream tasks. Representing features through individual atoms enhances controllability in the adaptation process. + +We validate our method on generative fine-tuning tasks. We show that the target concept can be effectively represented using feature dictionary atoms after fine-tuning the pre-trained model within our framework. Compared to baseline methods, our approach allows the selection of specific atoms to refine the generated image. This adaptability improves performance on image editing tasks by enhancing editing results through the removal of unimportant atoms. We also apply our approach to customized text-to-image tasks, showing that fine-tuning atoms effectively captures + +![](images/77c1546029f1671b0b227341cd29ef583e75cc8e99d378dd460adf90c9fc0a3d.jpg) +(a) + +![](images/729fbb1d304c1962ed75a5d40e2081addb6167c069c50549d9a1016b7b3cc437.jpg) +(b) + +![](images/47aede7514ac4790cf5e35eef8a6dbe914fc5f711c44988e0614d3229d9bb343.jpg) +(c) +Figure 1. (a) Our method leverages sparse coding and enforces high sparsity in adapted feature $\Delta \mathbf{O}$ , ensuring that the output feature is constructed using only a few dictionary atoms $\mathbf{D}$ . (b) The pre-trained model initially generates "8". It generates "3" after integrating the learned dictionary $\mathbf{D}$ which is fine-tuned on "3". (c) The overall transformation of the generated digit from "8" to "3" using the dictionary atoms. The red indicates an increase in value at a specific location and blue indicates a decrease. The total influence is decomposed into contributions from atoms at different layers: $\Delta \mathbf{O}^{(1)}$ represents the combined effect of 8 atoms in the first layer, while $\Delta \mathbf{O}^{(2)}$ accounts for the 4 atoms in the second layer. The details of this experiment is provided in Section 3.3 + +personal concepts while preserving the capabilities of pretrained models. Our method outperforms various baseline approaches, demonstrating its effectiveness in controlled adaptation for generative modeling. + +We summarize our contributions as follows: + +- We formulate large transformer fine-tuning as a sparse coding problem, where the adapted feature representation is a sparse combination of a dictionary of feature atoms. +- We introduce a fine-tuning framework that leverages a combination of dictionary atoms and demonstrates that tuning atoms enhances control over the adaptation process. +- We validate the effectiveness of our approach in generative tasks by demonstrating the improvement in image editing and concept customization. + +# 2. Related Work + +Sparse Representation and Attention Recent research has demonstrated the significant role of sparsity in enhancing both interpretability and computational efficiency in generative models through various approaches including sparse autoencoders for feature extraction [49], sparse feature circuits for improved generalization [31], transcoders for detailed circuit analysis [12], dictionary learning for monosemantic representations [5], and selective parameter tuning strategies [18]. Complementary work has addressed the quadratic computational complexity of standard attention mechanisms through sparse attention techniques that compute attention scores for only a subset of token pairs [3, 10, 50, 52], with optimizations including + +dynamic pattern adjustment [24], hardware-aligned implementations [60], and graph-based conceptualizations [51], collectively enabling transformers to process significantly longer sequences while maintaining model quality despite challenges in balancing efficiency with performance in complex long-range dependency scenarios. Compared to previous works, our method applies sparse coding for finetuning transformer-based models in generative tasks. + +Representation Tuning Recent representation tuning approaches operate directly on model representations rather than weights, achieving parameter efficiency while enhancing controllability, and interpretability across various model types and tasks [1, 8, 27, 28, 36, 53, 57, 58, 64]. Unlike traditional PEFT approaches that impact representations indirectly through weight adjustments, representation tuning methods preserve the original pre-trained model weights while directly modifying the output representations. In comparison, our work studies the sparsity in the updated representation and frames it as a dictionary learning problem. + +Personalized Text-to-Image Generation Extensive research has been conducted on aligning text prompts and generated concepts [19, 20, 23, 25, 34, 35, 41, 47]. Zhang et al. [62] developed an object-conditioned energy-based attention map alignment technique in diffusion models, enhancing control over specific objects in generated images. Another study by Zhang et al. [61] improved semantic fidelity in text-to-image synthesis through attention regulation. Hao et al. [16] introduces ConceptExpress, an unsu + +pervised method for extracting personalized concepts from a single image. Ren et al. [45] addressed memorization issues in text-to-image diffusion models by analyzing cross-attention mechanisms, proposing solutions to reduce overfitting to user-specific data. We approach this problem by adopting the ideas from sparse coding and learning the target concept as a sparse combination of coefficients and dictionary atoms. + +# 3. Method + +# 3.1. Preliminary + +Multi-head attention (MHA). The attention layer [52] transforms the input sequence $\mathbf{X} \in \mathbb{R}^{N \times C_i}$ to the output sequence $\mathbf{O} \in \mathbb{R}^{N \times C_o}$ , where $N$ denotes the sequence length, $C_i$ and $C_o$ are the dimension of input and output features. The attention layer projects the input as $\mathbf{Q} = \mathbf{X}\mathbf{W}_q$ , $\mathbf{K} = \mathbf{X}\mathbf{W}_k$ , $\mathbf{V} = \mathbf{X}\mathbf{W}_v$ using the corresponding projection matrices $\mathbf{W}_q$ , $\mathbf{W}_k$ , $\mathbf{W}_v \in \mathbb{R}^{C_i \times C_o}$ , and calculates the attention map, + +$$ +\mathbf {A} = \operatorname {a t t n} (\mathbf {X}) = \operatorname {s o f t m a x} \left(\mathbf {X} \mathbf {W} _ {q} \mathbf {W} _ {k} ^ {\intercal} \mathbf {X} ^ {\intercal}\right). \tag {1} +$$ + +The attention map $\mathbf{A} \in \mathbb{R}^{N \times N}$ captures the token-wise relationship by doing inner-product in a space transformed by $\mathbf{W}_q, \mathbf{W}_k$ . Then the output of the attention layer is, + +$$ +\mathbf {O} = \mathbf {A X W} _ {v}. \tag {2} +$$ + +Multi-head attention extends this by allowing multiple attention mechanisms to work in parallel, with each head independently learning attention patterns. The project matrices for each head are $\mathbf{W}_q^{(h)},\mathbf{W}_k^{(h)},\mathbf{W}_v^{(h)}\in \mathbb{R}^{C_i\times \frac{C_o}{H}}$ where $h = 1,\dots ,H$ and $H$ is the number of heads. The multi-head attention represents as, + +$$ +\mathbf {Z} = \sum_ {h = 1} ^ {H} \mathbf {A} ^ {(h)} \mathbf {X} \mathbf {W} _ {v o} ^ {(h)}, \tag {3} +$$ + +where $\mathbf{W}_o\in \mathbb{R}^{C_o\times C_o}$ and $\mathbf{W}_{vo}^{(h)} = \mathbf{W}_v^{(h)}\mathbf{W}_o^{(h)\intercal},\mathbf{W}_o^{(h)}\in$ $\mathbb{R}^{C_o\times \frac{C_o}{H}}$ $\mathbf{W}_{vo}^{(h)}\in \mathbb{R}^{C_i\times C_o}$ + +Sparse coding. Sparse coding is a signal representation technique that models a signal as a linear combination of a small number of basis functions from a dictionary. Given a signal $\mathbf{X} \in \mathbb{R}^{N \times C_i}$ , the goal of sparse coding is to find a dictionary $\mathbf{D} \in \mathbb{R}^{m \times C_i}$ (where $m > C_i$ is the number of atoms) and a sparse coefficient vector $\mathbf{S} \in \mathbb{R}^{N \times m}$ such that the signal can be approximated as: + +$$ +\mathbf {X} \approx \mathbf {S D}, \tag {4} +$$ + +where $\mathbf{S}$ has only a few nonzero elements, ensuring sparsity. The sparse coding problem is often formulated as an + +optimization problem: + +$$ +\min _ {\mathbf {S}} \| \mathbf {X} - \mathbf {S D} \| _ {2} ^ {2} + \lambda \| \mathbf {S} \| _ {1}, \tag {5} +$$ + +where $\lambda$ is a regularization parameter controlling the tradeoff between sparsity and reconstruction error. The Proximal gradient method is a common approach to solving this optimization problem, including Iterative Shrinkage-Thresholding Algorithm (ISTA) and its accelerated version (FISTA) [2], which utilize proximal operators for efficient iterative updates. For sparse coding with $\ell_1$ regularization, the corresponding proximal operator is the soft-thresholding function: + +$$ +\operatorname {p r o x} _ {\lambda \| \cdot \| _ {1}} (\mathbf {S}) = \operatorname {s i g n} (\mathbf {S}) \odot \max (| \mathbf {S} | - \lambda , 0), \tag {6} +$$ + +where $\odot$ is Hadamard product. It encourages sparsity by shrinking small values of $\mathbf{S}$ toward zero. Sparse coding is widely used in signal processing, machine learning, and neuroscience, providing efficient and interpretable representations of data. + +# 3.2. Sparse Fine-Tuning + +In this work, we conceptualize model fine-tuning as a sparse coding problem, i.e., the adapted features are constructed using a dictionary of feature atoms. As our method focuses on the transformer architecture, we first propose an interpretation of the attention operation by viewing the product between the attention map and value matrices as a combination of dictionary atoms with dictionary coefficients. We then propose a fine-tuning approach by representing the adapted features as a learned sparse combination of dictionary atoms with sparse coefficients, representing the fine-tuned representations' underlying sparsity nature. + +A dictionary view of attention. For single head attention (2), each row of $\mathbf{A}$ can be interpreted as coefficients for combining $\mathbf{V}$ into the output $\mathbf{O}$ , i.e., $\mathbf{O}_i = \sum_j\mathbf{A}_{ij}\mathbf{V}_j$ . In the case of multi-head attention (3), + +$$ +\sum_ {h = 1} ^ {H} \mathbf {A} ^ {(h)} \mathbf {X} \mathbf {W} _ {v o} ^ {(h)} = \left[ \begin{array}{l l l} \mathbf {A} ^ {(1)} & \dots & \mathbf {A} ^ {(H)} \end{array} \right] \left[ \begin{array}{c} \mathbf {X} \mathbf {W} _ {v o} ^ {(1)} \\ \vdots \\ \mathbf {X} \mathbf {W} _ {v o} ^ {(H)} \end{array} \right], \tag {7} +$$ + +it can be represented as a product of a composite coefficient $\left[\mathbf{A}^{(1)},\dots ,\mathbf{A}^{(H)}\right]\in \mathbb{R}^{N\times NH}$ , and an extensive dictionary $\left[\mathbf{XW}_{vo}^{(1)},\dots ,\mathbf{XW}_{vo}^{(H)}\right]^{\intercal}\in \mathbb{R}^{NH\times C_o}$ . Each row of the output $\mathbf{O}$ is a linear combination of $\mathbf{XW}_{vo}$ , i.e., $\mathbf{O}_i = \sum_{h,j}\mathbf{A}_{ij}^{(h)}(\mathbf{XW}_{vo}^{(h)})_j$ . When represented as a linear combination of dictionary elements, multi-head attention and single-head attention follow the same formulation. In the following discussion, we focus on the single-head attention formulation, which can be seamlessly extended to the multi-head attention case. + +A sparse dictionary view of attention. From the perspective of sparse coding, the feature is expressed as SD, where S represents the sparse coefficients and D consists of dictionary atoms that remain data-independent once learned. Although attention can be expressed as a linear combination of atoms, it does not inherently exhibit sparsity. To fully leverage the advantages of sparse representations, we propose to represent the attention mechanism (2) in the form of sparse coding, + +$$ +\mathbf {O} = \mathbf {S D} = \sigma_ {\lambda} \left(\mathbf {A X W} _ {s}\right) \mathbf {D}, \tag {8} +$$ + +where $\sigma_{\lambda}(\cdot)$ can be soft-thresholding function or $\mathrm{ReLU}(x - \lambda)$ to ensure sparsity of $\mathbf{S} = \sigma_{\lambda}(\mathbf{AXW}_s) \in \mathbb{R}^{N \times C_o}$ , $\mathbf{W}_s \in \mathbb{R}^{C_i \times M}$ and $\mathbf{D} \in \mathbb{R}^{M \times C_o}$ is the feature dictionary atoms, $M$ is the number of atoms. + +With this formulation, we obtain the feature dictionary atoms $\mathbf{D}$ (we use the term "atoms" for simplification in this paper), which are the building blocks for feature representation $\mathbf{O}$ . We illustrate our formulation in Figure 1 (a). This formulation provides the following properties: + +- The sparse coefficients $\mathbf{S} = \sigma_{\lambda}(\mathbf{AXW}_s)$ ensure that the feature representation $\mathbf{O}$ is combined by only a few $\mathbf{D}$ . We present an extreme example to illustrate how atoms enhance interpretability: $\mathbf{O}_i = \sum_{j=1}^{M} \mathbf{S}_{ij} \mathbf{D}_j$ can be simplified as $\mathbf{O}_i = \mathbf{S}_{ij^*} \mathbf{D}_{j^*}$ if all elements of $\mathbf{S}_i$ is 0 except the element at index $j^*$ , and only $\mathbf{D}_{j^*}$ contributes to the output feature $\mathbf{O}_i$ . +- The sparse coefficients $\mathbf{S}$ are data-dependent, while the feature dictionary atoms $\mathbf{D}$ are data-independent. While different inputs generate different coefficients, the output features are always constructed using the same feature dictionary atoms. +- We can adjust the size of the feature dictionary by varying $M$ to handle tasks with different difficulties. Large $M$ leads to an overcomplete dictionary while small $M$ produces an undercomplete dictionary. + +Fine-tuning via sparse dictionary learning. The above formulation frames feature representation through the lens of sparse coding. In the context of large model fine-tuning, we freeze the pre-trained parameters to make feature representations $\mathbf{O}$ of the pre-trained model fixed. Instead, we introduce a sparse coding perspective to encode the adapted feature representation $\Delta \mathbf{O}$ , which is learned from downstream tasks. Specifically, we have, + +$$ +\mathbf {O} + \Delta \mathbf {O} = \mathbf {O} + \sigma_ {\lambda} (\mathbf {A X W} _ {s}) \mathbf {D}. \tag {9} +$$ + +With this formulation, the pre-trained weights remain unchanged, preserving the capabilities of the pre-trained model. The adapted feature $\Delta \mathbf{O}$ is represented as a linear combination of feature dictionary atoms $\mathbf{D}$ , using sparse coefficients $\mathbf{S} = \sigma_{\lambda}(\mathbf{AXW}_s)$ . This formulation offers + +![](images/37034641eb61d11891b337d66289edce00d0e01b91ddbace63dcab804c2dc3a9.jpg) +Figure 2. (a) Adjusting only the coefficients fails to adapt the model to the downstream task, whereas (b) modifying only the atoms enables a perfect fit to the new task composed of high-frequency signals. (c) Fine-tuning the atoms effectively transitions the model from low-frequency to high-frequency representations. They encode the representation of the downstream task. + +a framework for understanding how $\mathbf{D}$ contributes to the adapted features in the fine-tuned model. We illustrate our method in Figure 1 (a). + +# 3.3. Analysis of Dictionary Atoms + +In this section, we demonstrate distinct properties of atoms $\mathbf{D}$ and sparse coefficients $\mathbf{S}$ . + +Features atoms are key to fine-tuning. We analyze the distinct behaviors of coefficients and atoms, and discover that atoms play a crucial role in adapting to new tasks. To explore this within the context of an attention block, we perform a toy experiment where masked one-dimensional signals are reconstructed using an attention layer. The signals, with a length of 64, are synthesized by selecting 5 random Fourier bases, with a low frequency ranging from $\frac{0}{64}$ to $\frac{24}{64}$ . The objective is to reconstruct the signal after masking out $75\%$ of its values with only 16 randomly sampled observations. After we have a pre-trained model on the low-frequency signals, we fine-tune the model to a new task where signals are constructed with high-frequency Fourier bases ranging from $\frac{25}{64}$ to $\frac{32}{64}$ . In this experiment, we test two different cases: (1) tuning only coefficients on the high-frequency data with atoms fixed, and (2) tuning only atoms on the high-frequency data with coefficients fixed. As illustrated in Figure 2, when adapting to high-frequency signals, modifying only the atoms while keeping the coefficients fixed allows the model to effectively generalize to the new task. In contrast, updating only the coefficients is insufficient for proper adaptation. Building on this observation, we hypothesize that adjusting atoms is more suitable than adjusting coefficients for adapting to a new task. We observe from Figure 2 (c) that atoms are capable of encoding + +the representation of the downstream task. The experiment details are in Appendix B. + +Identifying the importance of feature atoms. We analyze the impact of atoms on the newly learned concept by visualizing both the influence of individual atoms and their combined effect. In this experiment, we adopt a transformer-based variational autoencoder (VAE) [21] with a two-layer decoder. The experiment process is illustrated in Figure 1 (b). We first pre-train the model on an image-denoising task using a subset of the MNIST dataset containing digits 5 to 9. We then fine-tune the model on a downstream denoising task with only digit 3 by learning a dictionary for each layer with $M = 100$ . In this scenario, the digit "3" is the newly learned concept. Given the same noisy latent, the decoder outputs an "8" without incorporating fine-tuned feature atoms, whereas it generates a "3" when the fine-tuned feature atoms are applied. The model requires only 12 atoms to shift the generated digits from 8 to 3. The influence of each atom is visualized in Figure 1 (c), with the overall impact represented as $\Delta \mathbf{O}$ , where the red area indicates an increase in pixel value at a specific location and the blue area represents a decrease. It is straightforward to see that these 12 feature atoms enhance the regions corresponding to "3" while suppressing certain portions of "8". Additionally, Figure 1 (c) illustrates the individual behavior of each atom, showing how different atoms affect distinct parts of the generated digit. The experiment details are in Appendix B. + +# 3.4. Design Variations + +In this section, we outline the various design choices implemented in our experiments. Our method achieves sparse coefficient $\mathbf{S} = \sigma_{\lambda}(\mathbf{AXW}_s)$ by utilizing non-linear function $\sigma_{\lambda}(\cdot)$ . In the literature on sparse coding [4], there are three different methods: (i) soft-thresholding function, $\sigma_{\lambda}(\mathbf{x}) = \mathrm{prox}_{\lambda \| \cdot \|_1}(\mathbf{x}) = \mathrm{sign}(\mathbf{x}) \odot \max(|\mathbf{x}| - \lambda, 0)$ , (ii) ReLU, $\sigma_{\lambda}(\mathbf{x}) = \mathrm{ReLU}(\mathbf{x} - \lambda) = \max(\mathbf{x} - \lambda, 0)$ , and (iii) top- $k$ activation selection, which selects the most significant features of top $k$ contributing to the learned representation. In practice, we observe that top- $k$ activation selection maintains a consistent level of sparsity across different inputs. In contrast, methods (i) soft-thresholding function and (ii) ReLU exhibit sensitivity to input values, leading to variations in sparsity. However, methods (i) and (ii) offer control over the coefficient values, ensuring they stay above a specified threshold. + +Furthermore, we find that when $\sigma_{\lambda}(\mathbf{X}\mathbf{W}_s)$ is highly sparse, the resulting product $\mathbf{A}\sigma_{\lambda}(\mathbf{X}\mathbf{W}_s)$ also retains sparsity. In modern transformers, computation of $\mathbf{A}$ and $\mathbf{X}\mathbf{W}_v$ are performed separately. For simplicity in implementation, we adopt the formulation $\mathbf{A}\sigma_{\lambda}(\mathbf{X}\mathbf{W}_s)$ . + +# 3.5. Analysis of Our Design + +Comparison with LoRA. When representing attention mechanism (1-3) as a linear combination of dictionary atoms, updating $\mathbf{W}_q$ , $\mathbf{W}_k$ is equivalent to updating the coefficients while updating $\mathbf{W}_v$ , $\mathbf{W}_o$ is equivalent to updating the atoms. LoRA applies adaptation on $\mathbf{W}_q$ , $\mathbf{W}_k$ , $\mathbf{W}_v$ , $\mathbf{W}_o$ with a low-rank matrices $\mathbf{W}_A \in \mathbb{R}^{C_i \times r}$ , $\mathbf{W}_B \in \mathbb{R}^{r \times C_o}$ , where $r$ represents the rank. LoRA adapts both coefficients and atoms in a low-rank way. In comparison, our method adapts the coefficients with $\mathbf{W}_s$ and obtains sparse coefficients $\sigma_{\lambda}(\mathbf{AXW}_s)$ , while representing atoms with $\mathbf{D}$ . + +Parameters and computational cost. To adapt the representation at each layer, we need additional parameters for $\mathbf{W}_s$ and $\mathbf{D}$ , which are $(C_i + C_o)M$ . After the fine-tuning, considering the density of the coefficients $\rho$ , i.e., $1 - \rho$ is sparsity, the storage required for the parameters is $C_i M + \rho C_o M$ . In comparison, LoRA requires $4(C_i + C_o)r$ additional parameters. The FLOPs required by our method are $2N M C_i + 2\rho N M C_o$ . LoRA requires $8C_i C_o r$ FLOPs. For example, if $C_i = C_o = 1024$ , $M = 256$ , $r = 16$ , $N = 1024$ , $\rho = 0.01$ , our method requires 264k parameters for storage and 0.54GFLOPs for computation, while LoRA requires 131k parameters and 0.13GFLOps. Our method involves slightly higher computational costs than LoRA, as it accounts for feature interactions, whereas LoRA only updates the weights. Next, we demonstrate that by leveraging sparse feature representations, our approach achieves greater robustness compared to LoRA. + +Our method provides more stable adaptation. We compare an extreme case between LoRA and our method, by considering the attention in a dictionary form $\Delta \mathbf{O} = \mathbf{A}\mathbf{V}$ or $\Delta \mathbf{O} = \mathbf{SD}$ . In this experiment, we examine the difference between updating coefficients using a low-rank approach versus a sparse approach. Both LoRA and our method produce the updated feature $\Delta \mathbf{O}$ with rank 1, where LoRA achieves this by setting $r = 1$ and our method enforces only one active atom in D. The key difference lies in how the coefficients are updated—LoRA adapts them in a low-rank manner, whereas our method updates them sparsely. As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. LoRA fails to produce meaningful images with a slightly modified prompt, while our method remains robust, consistently generating a concept that aligns with the updated text prompt. We attribute the effectiveness of our approach to the sparse coefficients produced by $\sigma_{\lambda}(\mathbf{AXW}_s)$ . With a different text prompt, the values in S either remain zero or take on a significant value when receiving the proper input, ensuring that the generated result remains coherent. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs. + +![](images/cbb7deb14fb6089e4f69d50f446103fb7166aca5d7a08551e59194b5ed405e56.jpg) +Figure 3. The pre-trained model can generate the target concept when integrated with the learned dictionary, even with about 40 atoms. With just four atoms, the core structure of the concept can still be generated. Adjusting the number of selected atoms influences the generated results, offering a controllable approach to image editing. Additional discussions of these figures are in Section 4.1. + +![](images/8216130cf24de06def538ebcb8bfbaa1681552147f8a4a0e26223a49d1451517.jpg) +Figure 4. The role of sparsity: Our method utilizes the same number of atoms as LoRA, yet LoRA struggles to adapt to the target prompt effectively. Sparse coding ensures that atoms are either activated or inactive, providing a structured and interpretable adaptation process. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs. + +# 4. Experiments + +In this section, we evaluate the effectiveness of our method in image generation tasks. We demonstrate that feature dictionary atoms effectively represent the target concept after fine-tuning the pre-trained model. Compared to baselines, our approach enables selective atom adjustment, improving image editing by refining results and enhancing text-to-image concept customization while preserving pre-trained model capabilities. + +# 4.1. Representing Concepts Using Atoms + +Experimental settings. In this experiment, we utilize PixArt- $\Sigma$ [7], a Diffusion Transformer (DiT) [40] model consisting of 28 transformer blocks. We fine-tune the model on a single image for 60 epochs using our method and reconstruct the corresponding concept using the same prompt used during fine-tuning. The selected images are drawn from DreamBooth [47], Custom Diffusion [23]. We control the number of active atoms and visualize the variation of the generated results with different atoms. + +The concept is represented by a set of atoms. As shown in Figure 3, after fine-tuning our method on the target concept, its representation can be effectively captured using a small set of atoms, such as 40 atoms. Among these, just 4 atoms are sufficient to construct the core structure of the concept, including key attributes like pose and color. We provide details descriptions of these images in Appendix C. + +# 4.2. Image Editing + +Building on our previous observation that adjusting the number of atoms influences the generated results, we explore leveraging this property for image editing. Image editing tasks aim to modify an image using a generative model while retaining as many details as possible from the original [63]. However, certain editing tasks pose challenges in maintaining most of the original details while achieving effective modifications. For example, altering the style can significantly change the overall appearance of the image. We argue that successful image editing should preserve the main components of the image while applying modifications selectively. Since our method represents the core + +![](images/36b33affdd79a29e0e81d0545200c0b59997033b1c86c600b662afd65bac9bf4.jpg) +Figure 5. Results on image editing. Our method shows improved results by better following editing prompts while preserving the key components of the source images. + +components of a concept using only a few atoms, it is well-suited for image editing tasks, ensuring that key structural elements remain intact while allowing for controlled transformations. + +Experimental settings. In our experiment, we select data from previous work [63] and evaluate different editing tasks, including background modification, scene alteration, and pose adaptation. Previous works [15, 63] fine-tune models to learn a concept and modify its appearance by adjusting the prompt. However, these approaches have not yet been applied to DiT. Following their setup, we compare our method with fine-tuning approaches, such as LoRA [17], using them as baselines for evaluation. We adopt the DiT architecture and fine-tune both LoRA and our method on a single image for 60 epochs to achieve full reconstruction. After fine-tuning, we modify the image using a different prompt for editing. + +Image editing results. The results are shown in Figure 5. Our method applies edits based on the test prompt while maintaining the core structure of the source images. In addition, we control the number of active atoms by adjusting the coefficient density $\rho$ , where $1 - \rho$ represents the sparsity level. In Figure 6, as sparsity increases, only the most essential atoms remain, preserving the core concept while al + +![](images/5b93197756b6e23e6edfc00ca4b60b4d148bdfca649748376b52735d4f4218ba.jpg) +$\rho = 0.01$ + +![](images/82b26ff313d9e7fb87aafc67cce219ae5790621996dc2649db2a7586d60fb039.jpg) +$\rho = 0.02$ + +![](images/351038a862d7318ed76948c56ed0396141491118b19eea0e699f3c98fac29398.jpg) +$\rho = 0.04$ +Figure 6. Adjusting the density, i.e., the number of activated atoms, can change the influence on the edited image. + +lowing the pre-trained model greater flexibility to integrate editing effects based on the modified text prompt. + +# 4.3. Concept Customization + +In the previous section, we demonstrated that our method can efficiently represent the target concept using only a small number of atoms. In this section, we further show that even with just a few atoms, our method achieves good performance in concept customization. + +Experimental settings. We evaluated our approach on the DreamBooth [47] dataset, which consists of 30 subjects across 15 distinct classes. Among these, 9 are live subjects, while the remaining 21 are objects. Each subject has between 4 to 6 images, captured under varying conditions, in diverse environments, and from multiple angles. All 30 subjects were trained for 15 epochs and validated on 25 different prompts. For our experiments, we adopted the Pixart- $\Sigma$ , a Diffusion Transformer (DiT) [7], as the base architecture and fine-tuned it using our proposed method. To verify the effectiveness of our approach, we compared its performance against several PEFT baselines for personalized text-to-image generation, including LoRA [17], DoRA [26], and OFT [43]. We use the AdamW optimizer with a learning rate of $1\times 10^{-4}$ to fine-tune the Pixart- $\Sigma$ [7] for 15 steps. For the baseline methods, LoRA and DoRA are assigned a rank of $r = 16$ , while OFT is set to $r = 4$ . We generated five images for each prompt at a resolution of $1024\times 1024$ . + +Evaluation metrics. We measure concept alignment by computing the mean cosine similarity between the generated image's CLIP embeddings [44] and those of the original concept images. We quantify image diversity using the Vendi score [13] with DINOv2 embeddings [38]. For text alignment, we calculate the average cosine similarity in the CLIP feature space [44]. ImageReward [59] is a general-purpose text-to-image human preference reward model that evaluates and ranks AI-generated images by scoring them based on human preferences. + +Customization results. As shown in Table 1, compared with baseline methods, our approach not only represents + +Table 1. Comparison with baseline methods on concept customization. + +
AlignmentDiversityFidelityImageReward
LoRA [17]0.2184.7070.7110.220
DoRA [26]0.2134.5400.7120.142
OFT [43]0.1903.9260.6630.137
Ours (ρ = 0.02)0.2355.2050.7020.182
Ours (ρ = 0.04)0.2315.1720.7130.431
+ +the concept with a minimal set of learned atoms but also generalizes robustly to diverse scenarios. In particular, our method achieves higher text-to-image alignment and diversity scores than most other PEFT baselines, while still preserving a relatively high fidelity score. This indicates that our approach effectively captures essential features of each concept without sacrificing the quality of customization text-to-image generation. + +Number of parameters in different methods. We compare the trainable parameters of our method and the baselines. The pre-trained PixArt model has around 616M parameters in total. Among the fine-tuning methods, both LoRA and DoRA have around 4M trainable parameters. OFT uses around 37M trainable parameters, and our method trains around 17M parameters. + +# 4.4. Abalation Study + +The impact of sparsity in the coefficients. We analyze the impact of varying sparsity levels in the coefficients through a concept customization experiment, evaluating text alignment, image diversity, and image fidelity. As shown in Figure 7, higher sparsity (low density) enhances text alignment and image diversity but results in slightly lower image fidelity. In contrast, lower sparsity (high density) improves fidelity, indicating that using more atoms allows for a more precise representation of the concept from the source image. In our experiment, we adopt both $\rho = 0.02$ and $\rho = 0.04$ for either better diversity or better fidelity. + +The impact of member numbers $M$ . We also examine the impact of the number of members, i.e., the dictionary size, on the generated results while maintaining the same level of sparsity. As shown in Table 2, increasing the number of members improves image fidelity, whereas a smaller dictionary size reduces fidelity but enhances text alignment and image diversity. In our experiment, we choose $M = 256$ to provide balanced fidelity and diversity. + +Sparsity coefficients provide more stable adaptation. In this experiment, we compare our method with LoRA on the image editing task. As we discussed in Section 3.5, + +![](images/395defdb074916240743b6a3b1336baf7c6175f7c76b1b0811432ebf039c142c.jpg) +Figure 7. The influence of different density $\rho$ on the image diversity, fidelity and text-to-image alignment. + +Table 2. The influence of different member numbers $M$ . As the number of members increases,alignment and diversity scores decrease, while the fidelity score improves. + +
ρ = 0.01AlignmentDiversityFidelity
M = 160.2485.8670.547
M = 2560.2494.3410.650
M = 10240.2413.7330.672
M = 20480.2383.5250.692
+ +LoRA adapts the coefficients in a low-rank manner, whereas our method updates them sparsely. Both LoRA and our method produce the updated feature $\Delta \mathbf{O}$ with rank 1, where LoRA achieves this by setting $r = 1$ and our method enforces only one active atom in $\mathbf{D}$ . As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. However, when applying the fine-tuned weights to generate images with a slightly modified prompt, LoRA becomes unstable, failing to produce meaningful images. In contrast, our method remains robust, consistently generating a concept that aligns with the updated text prompt. + +# 5. Conclusion + +In this paper, we introduce to frame model fine-tuning as a sparse coding problem, where the adapted feature representation is constructed using a dictionary of feature atoms. We propose a fine-tuning framework that leverages a combination of feature dictionary atoms, demonstrating that tuning them enhances control over the adaptation process. We validate the effectiveness of our approach across various generative tasks. By representing concepts with only a few essential atoms, our method proves to be highly adaptable and can be effectively applied to image editing, style mixing, and concept personalization. Our study provides a new perspective on the mechanics of fine-tuning, paving the way for more controllable adaptation strategies in large-scale pretrained models. + +# References + +[1] Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. Decomposing and interpreting image representations via text in vits beyond CLIP. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[2] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2009. 3 +[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2 +[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 5 +[5] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Chris Olah. Towards monoseismicity: Decomposing language models with dictionary learning. 2023. Transformer Circuits Thread. 2 +[6] Emmanuel J Candès et al. Compressive sampling. In Proceedings of the international congress of mathematicians, pages 1433-1452, 2006. 3 +[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\sigma$ : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91, 2024. 6, 7, 2 +[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. International Conference on Learning Representations, 2025. 2 +[9] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 1 +[10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1 +[12] Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. Transcoders enable fine-grained interpretable circuit analysis for language models. 2024. AI Alignment Forum. 2 +[13] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7 +[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 2023. 3 + +[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In CVPR, 2023. 7 +[16] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In European Conference on Computer Vision, 2024. 2 +[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 7, 8 +[18] Teng Hu, Jiangning Zhang, Ran Yi, Hongrui Huang, Yabiao Wang, and Lizhuang Ma. Sara: High-efficient diffusion model fine-tuning with progressive sparse low-rank adaptation. arXiv preprint arXiv:2409.06633, 2024. 2 +[19] Jeeyung Kim, Erfan Esmaeili, and Qiang Qiu. Text embedding is not all you need: Attention control for text-to-image semantic alignment with text self-attention maps. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8031–8040, 2025. 2 +[20] Taewook Kim, Wei Chen, and Qiang Qiu. Learning to customize text-to-image diffusion in diverse context. arXiv preprint arXiv:2410.10058, 2024. 2 +[21] Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. Foundations and Trends® in Machine Learning, 2019. 5 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1 +[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 6 +[24] Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference. arXiv preprint arXiv:2502.20766, 2025. 2 +[25] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 2 +[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 1, 7, 8 +[27] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 2 +[28] Yiyang Liu, James Chenhao Liang, Ruixiang Tang, Yugyung Lee, Majid Rabbani, Sohail Dianat, Raghuveer Rao, Lifu Huang, Dongfang Liu, Qifan Wang, et al. Re-imagining + +multimodal instruction tuning: A representation view. arXiv preprint arXiv:2503.00723, 2025. 2 +[29] Yang Luo, Xiaozhe Ren, Zangwei Zheng, Zhuo Jiang, Xin Jiang, and Yang You. Came: Confidence-guided adaptive memory efficient optimization. arXiv preprint arXiv:2307.02047, 2023. 2 +[30] Julien Mairal, Jean Ponce, Guillermo Sapiro, Andrew Zisserman, and Francis Bach. Supervised dictionary learning. Advances in neural information processing systems, 2008. 1 +[31] Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647, 2024. 2 +[32] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1 +[33] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1 +[34] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 2 +[35] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 2 +[36] Zichen Miao, Wei Chen, and Qiang Qiu. Coeff-tuning: A graph filter subspace view for tuning attention-based large models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20146-20157, 2025. 2 +[37] Bruno A Olshausen and David J Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, pages 607-609, 1996. 1 +[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[39] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1 +[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In CVPR, 2023. 6 +[41] Vishal Purohit, Matthew Repasky, Jianfeng Lu, Qiang Qiu, Yao Xie, and Xiuyuan Cheng. Posterior sampling via Langevin dynamics based on generative priors. arXiv preprint arXiv:2410.02078, 2024. 2 +[42] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1 + +[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 1, 7, 8 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 7 +[45] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, 2024. 3 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1 +[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 6, 7, 3 +[48] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuzhhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2024. 3 +[49] Viacheslav Surkov, Chris Wendler, Mikhail Terekhov, Justin Deschenaux, Robert West, and Caglar Gulcehre. Unpacking sdxl turbo: Interpreting text-to-image models with sparse autoencoders. arXiv preprint arXiv:2410.22366, 2024. 2 +[50] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International conference on machine learning, 2020. 2 +[51] Nathaniel Tomczak and Sanmukh Kuppannagari. Longer attention span: Increasing transformer context length with sparse graph processing techniques. arXiv preprint arXiv:2502.01659, 2025. 2 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 1, 2, 3 +[53] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 2 +[54] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1 +[55] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021. + +[56] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1 +[57] Muling Wu, Wenhao Liu, Xiaohua Wang, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaqing Zheng, and Xuanjing Huang. Advancing parameter efficiency in fine-tuning via representation editing. 2024. 2 +[58] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. 2024. 2 +[59] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 7 +[60] Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, and Wangding Zeng. Native sparse attention: Hardware-aligned and natively trainable sparse attention, 2025. 2 +[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. In European Conference on Computer Vision, 2024. 2 +[62] Yasi Zhang, Peiyu Yu, and Ying Nian Wu. Object-conditioned energy-based attention map alignment in text-to-image diffusion models. In European Conference on Computer Vision, 2024. 2 +[63] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 6, 7 +[64] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 2 \ No newline at end of file diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/images.zip b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f7d17c723200bfd918fb05b686efd77ab0df7b2a --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:830d6b622c15261fd8dacd35f8f529b13388e92bccce725f2cadf07723d34141 +size 544571 diff --git a/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/layout.json b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e7f91c5fb853bf5df69cd51cb946ff55e04bd687 --- /dev/null +++ b/2025/Sparse Fine-Tuning of Transformers for Generative Tasks/layout.json @@ -0,0 +1,9842 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 484, + 121 + ], + "type": "text", + "content": "Sparse Fine-Tuning of Transformers for Generative Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 143, + 417, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 143, + 417, + 171 + ], + "spans": [ + { + "bbox": [ + 192, + 143, + 417, + 171 + ], + "type": "text", + "content": "Wei Chen, Jingxi Yu, Zichen Miao, Qiang Qiu Purdue University, IN, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 157, + 172, + 451, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 172, + 451, + 186 + ], + "spans": [ + { + "bbox": [ + 157, + 172, + 451, + 186 + ], + "type": "text", + "content": "{chen2732, yu667, miaoz, qqiu}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 239, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 548 + ], + "type": "text", + "content": "Large pre-trained transformers have revolutionized artificial intelligence across various domains, and fine-tuning remains the dominant approach for adapting these models to downstream tasks due to the cost of training from scratch. However, in existing fine-tuning methods, the updated representations are formed as a dense combination of modified parameters, making it challenging to interpret their contributions and understand how the model adapts to new tasks. In this work, we introduce a fine-tuning framework inspired by sparse coding, where fine-tuned features are represented as a sparse combination of basic elements, i.e., feature dictionary atoms. The feature dictionary atoms function as fundamental building blocks of the representation, and tuning atoms allows for seamless adaptation to downstream tasks. Sparse coefficients then serve as indicators of atom importance, identifying the contribution of each atom to the updated representation. Leveraging the atom selection capability of sparse coefficients, we first demonstrate that our method enhances image editing performance by improving text alignment through the removal of unimportant feature dictionary atoms. Additionally, we validate the effectiveness of our approach in the text-to-image concept customization task, where our method efficiently constructs the target concept using a sparse combination of feature dictionary atoms, outperforming various baseline fine-tuning methods. We provide more details at our project page." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "In recent years, pre-trained large models have revolutionized artificial intelligence across domains like natural language processing and computer vision [11, 22, 46, 52]. Given the prohibitive computational resources required for training these models from scratch, fine-tuning has emerged as the dominant paradigm for adapting pre-trained models to specific downstream tasks, enabling efficient knowledge transfer while minimizing computational overhead." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Various parameter-efficient fine-tuning (PEFT) methods [17, 26, 43] have been developed to tune the large trans" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 346 + ], + "type": "text", + "content": "formers by updating a small subset of parameters while keeping large pre-trained weights frozen. A representative family of these approaches [17, 26] utilizes lightweight parameterizations, such as low-rank decomposition, to adapt the weight matrices. However, despite being lightweight, these methods obtain the updated features with a dense matrix multiplication between the input and the adapted weights, making it challenging to interpret how the individual updated parameters contribute to the newly learned representations and to identify the new knowledge gained through fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 491 + ], + "type": "text", + "content": "To tackle this challenge, we take inspiration from sparse coding [9, 30, 32, 33, 37, 39, 42, 54-56], in which signals are expressed as a sparse combination of basic elements, i.e., dictionary atoms. In this paper, we frame model fine-tuning as a sparse coding problem, where the adapted feature is constructed using a dictionary of feature atoms. These feature atoms are selectively combined with sparse coefficients, ensuring that only a small subset of atoms contributes to the updated representation. This sparsity not only disentangles interactions between features but also allows each atom to capture a distinct aspect of the downstream data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 591 + ], + "type": "text", + "content": "We first analyze with a toy experiment that the tuned feature dictionary atoms function as building blocks of the adapted representation, and tuning these atoms allows for effective adaptation to downstream tasks. We demonstrate that with pre-trained coefficients and atoms, fine-tuning the atoms effectively captures the representation of downstream tasks. Representing features through individual atoms enhances controllability in the adaptation process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "We validate our method on generative fine-tuning tasks. We show that the target concept can be effectively represented using feature dictionary atoms after fine-tuning the pre-trained model within our framework. Compared to baseline methods, our approach allows the selection of specific atoms to refine the generated image. This adaptability improves performance on image editing tasks by enhancing editing results through the removal of unimportant atoms. We also apply our approach to customized text-to-image tasks, showing that fine-tuning atoms effectively captures" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18703" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 220, + 246 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 220, + 246 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 220, + 246 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 220, + 246 + ], + "type": "image", + "image_path": "77c1546029f1671b0b227341cd29ef583e75cc8e99d378dd460adf90c9fc0a3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 250, + 145, + 258 + ], + "lines": [ + { + "bbox": [ + 134, + 250, + 145, + 258 + ], + "spans": [ + { + "bbox": [ + 134, + 250, + 145, + 258 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 386, + 245 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 386, + 245 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 386, + 245 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 386, + 245 + ], + "type": "image", + "image_path": "729fbb1d304c1962ed75a5d40e2081addb6167c069c50549d9a1016b7b3cc437.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 249, + 310, + 258 + ], + "lines": [ + { + "bbox": [ + 299, + 249, + 310, + 258 + ], + "spans": [ + { + "bbox": [ + 299, + 249, + 310, + 258 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 85, + 552, + 248 + ], + "blocks": [ + { + "bbox": [ + 389, + 85, + 552, + 248 + ], + "lines": [ + { + "bbox": [ + 389, + 85, + 552, + 248 + ], + "spans": [ + { + "bbox": [ + 389, + 85, + 552, + 248 + ], + "type": "image", + "image_path": "47aede7514ac4790cf5e35eef8a6dbe914fc5f711c44988e0614d3229d9bb343.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 250, + 476, + 258 + ], + "lines": [ + { + "bbox": [ + 466, + 250, + 476, + 258 + ], + "spans": [ + { + "bbox": [ + 466, + 250, + 476, + 258 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": "Figure 1. (a) Our method leverages sparse coding and enforces high sparsity in adapted feature " + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": ", ensuring that the output feature is constructed using only a few dictionary atoms " + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": ". (b) The pre-trained model initially generates \"8\". It generates \"3\" after integrating the learned dictionary " + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": " which is fine-tuned on \"3\". (c) The overall transformation of the generated digit from \"8\" to \"3\" using the dictionary atoms. The red indicates an increase in value at a specific location and blue indicates a decrease. The total influence is decomposed into contributions from atoms at different layers: " + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}^{(1)}" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": " represents the combined effect of 8 atoms in the first layer, while " + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}^{(2)}" + }, + { + "bbox": [ + 54, + 268, + 555, + 335 + ], + "type": "text", + "content": " accounts for the 4 atoms in the second layer. The details of this experiment is provided in Section 3.3" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 348, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 295, + 396 + ], + "type": "text", + "content": "personal concepts while preserving the capabilities of pretrained models. Our method outperforms various baseline approaches, demonstrating its effectiveness in controlled adaptation for generative modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 396, + 246, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 246, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 246, + 407 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 409, + 295, + 528 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "type": "text", + "content": "- We formulate large transformer fine-tuning as a sparse coding problem, where the adapted feature representation is a sparse combination of a dictionary of feature atoms." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 491 + ], + "type": "text", + "content": "- We introduce a fine-tuning framework that leverages a combination of dictionary atoms and demonstrates that tuning atoms enhances control over the adaptation process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 492, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 528 + ], + "type": "text", + "content": "- We validate the effectiveness of our approach in generative tasks by demonstrating the improvement in image editing and concept customization." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 540, + 142, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 142, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 142, + 553 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "Sparse Representation and Attention Recent research has demonstrated the significant role of sparsity in enhancing both interpretability and computational efficiency in generative models through various approaches including sparse autoencoders for feature extraction [49], sparse feature circuits for improved generalization [31], transcoders for detailed circuit analysis [12], dictionary learning for monosemantic representations [5], and selective parameter tuning strategies [18]. Complementary work has addressed the quadratic computational complexity of standard attention mechanisms through sparse attention techniques that compute attention scores for only a subset of token pairs [3, 10, 50, 52], with optimizations including" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 347, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 444 + ], + "type": "text", + "content": "dynamic pattern adjustment [24], hardware-aligned implementations [60], and graph-based conceptualizations [51], collectively enabling transformers to process significantly longer sequences while maintaining model quality despite challenges in balancing efficiency with performance in complex long-range dependency scenarios. Compared to previous works, our method applies sparse coding for finetuning transformer-based models in generative tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 453, + 556, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 595 + ], + "type": "text", + "content": "Representation Tuning Recent representation tuning approaches operate directly on model representations rather than weights, achieving parameter efficiency while enhancing controllability, and interpretability across various model types and tasks [1, 8, 27, 28, 36, 53, 57, 58, 64]. Unlike traditional PEFT approaches that impact representations indirectly through weight adjustments, representation tuning methods preserve the original pre-trained model weights while directly modifying the output representations. In comparison, our work studies the sparsity in the updated representation and frames it as a dictionary learning problem." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Personalized Text-to-Image Generation Extensive research has been conducted on aligning text prompts and generated concepts [19, 20, 23, 25, 34, 35, 41, 47]. Zhang et al. [62] developed an object-conditioned energy-based attention map alignment technique in diffusion models, enhancing control over specific objects in generated images. Another study by Zhang et al. [61] improved semantic fidelity in text-to-image synthesis through attention regulation. Hao et al. [16] introduces ConceptExpress, an unsu" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18704" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "pervised method for extracting personalized concepts from a single image. Ren et al. [45] addressed memorization issues in text-to-image diffusion models by analyzing cross-attention mechanisms, proposing solutions to reduce overfitting to user-specific data. We approach this problem by adopting the ideas from sparse coding and learning the target concept as a sparse combination of coefficients and dictionary atoms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 135, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 135, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 135, + 210 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": "Multi-head attention (MHA). The attention layer [52] transforms the input sequence " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": " to the output sequence " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": " denotes the sequence length, " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": " are the dimension of input and output features. The attention layer projects the input as " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": " using the corresponding projection matrices " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}" + }, + { + "bbox": [ + 55, + 215, + 296, + 312 + ], + "type": "text", + "content": ", and calculates the attention map," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 320, + 295, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 320, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 91, + 320, + 295, + 336 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = \\operatorname {a t t n} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {1}", + "image_path": "01514875253aab43c3b1c5f4c5b462bcf7fa23a4927a0c689c69865e36801b20.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "text", + "content": "The attention map " + }, + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "text", + "content": " captures the token-wise relationship by doing inner-product in a space transformed by " + }, + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 343, + 296, + 381 + ], + "type": "text", + "content": ". Then the output of the attention layer is," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 390, + 295, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 390, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 144, + 390, + 295, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\mathbf {A X W} _ {v}. \\tag {2}", + "image_path": "79c14d9545533b5bc12b2d6ed1e0e12c07f5fcb7526ee5b5140079865e2eb649.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "text", + "content": "Multi-head attention extends this by allowing multiple attention mechanisms to work in parallel, with each head independently learning attention patterns. The project matrices for each head are " + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^{(h)},\\mathbf{W}_k^{(h)},\\mathbf{W}_v^{(h)}\\in \\mathbb{R}^{C_i\\times \\frac{C_o}{H}}" + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "inline_equation", + "content": "h = 1,\\dots ,H" + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 413, + 296, + 486 + ], + "type": "text", + "content": " is the number of heads. The multi-head attention represents as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 494, + 295, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 494, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 127, + 494, + 295, + 527 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)}, \\tag {3}", + "image_path": "63772d414dc6262d8935e625f7411dcc166d5efdaf5e185dcb3c2c74ea299096.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}" + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{vo}^{(h)} = \\mathbf{W}_v^{(h)}\\mathbf{W}_o^{(h)\\intercal},\\mathbf{W}_o^{(h)}\\in" + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}" + }, + { + "bbox": [ + 55, + 534, + 295, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{vo}^{(h)}\\in \\mathbb{R}^{C_i\\times C_o}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "content": "Sparse coding. Sparse coding is a signal representation technique that models a signal as a linear combination of a small number of basis functions from a dictionary. Given a signal " + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}" + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "content": ", the goal of sparse coding is to find a dictionary " + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{m \\times C_i}" + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "inline_equation", + "content": "m > C_i" + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "content": " is the number of atoms) and a sparse coefficient vector " + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{N \\times m}" + }, + { + "bbox": [ + 55, + 576, + 296, + 661 + ], + "type": "text", + "content": " such that the signal can be approximated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 154, + 669, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 669, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 154, + 669, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {X} \\approx \\mathbf {S D}, \\tag {4}", + "image_path": "4f24957683169c5b54228aa607ccef08fff0bedee36d27fe800c2fa024797a06.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " has only a few nonzero elements, ensuring sparsity. The sparse coding problem is often formulated as an" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 73, + 407, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 407, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 407, + 84 + ], + "type": "text", + "content": "optimization problem:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 88, + 553, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 88, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 377, + 88, + 553, + 106 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {S}} \\| \\mathbf {X} - \\mathbf {S D} \\| _ {2} ^ {2} + \\lambda \\| \\mathbf {S} \\| _ {1}, \\tag {5}", + "image_path": "b74b8251bb5f3ef4420daa70c222340f12d290339f909f2241d7a2fcb0851d58.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "content": " is a regularization parameter controlling the tradeoff between sparsity and reconstruction error. The Proximal gradient method is a common approach to solving this optimization problem, including Iterative Shrinkage-Thresholding Algorithm (ISTA) and its accelerated version (FISTA) [2], which utilize proximal operators for efficient iterative updates. For sparse coding with " + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 110, + 555, + 217 + ], + "type": "text", + "content": " regularization, the corresponding proximal operator is the soft-thresholding function:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 222, + 555, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 222, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 343, + 222, + 555, + 237 + ], + "type": "interline_equation", + "content": "\\operatorname {p r o x} _ {\\lambda \\| \\cdot \\| _ {1}} (\\mathbf {S}) = \\operatorname {s i g n} (\\mathbf {S}) \\odot \\max (| \\mathbf {S} | - \\lambda , 0), \\tag {6}", + "image_path": "5201e1bae83362d404687f2a402cf98e38950166dbfef2acd2e49fc36ab40faf.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "text", + "content": " is Hadamard product. It encourages sparsity by shrinking small values of " + }, + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 239, + 554, + 299 + ], + "type": "text", + "content": " toward zero. Sparse coding is widely used in signal processing, machine learning, and neuroscience, providing efficient and interpretable representations of data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 304, + 430, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 430, + 317 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 430, + 317 + ], + "type": "text", + "content": "3.2. Sparse Fine-Tuning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 322, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 455 + ], + "type": "text", + "content": "In this work, we conceptualize model fine-tuning as a sparse coding problem, i.e., the adapted features are constructed using a dictionary of feature atoms. As our method focuses on the transformer architecture, we first propose an interpretation of the attention operation by viewing the product between the attention map and value matrices as a combination of dictionary atoms with dictionary coefficients. We then propose a fine-tuning approach by representing the adapted features as a learned sparse combination of dictionary atoms with sparse coefficients, representing the fine-tuned representations' underlying sparsity nature." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "content": "A dictionary view of attention. For single head attention (2), each row of " + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "content": " can be interpreted as coefficients for combining " + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "content": " into the output " + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i = \\sum_j\\mathbf{A}_{ij}\\mathbf{V}_j" + }, + { + "bbox": [ + 313, + 468, + 554, + 515 + ], + "type": "text", + "content": ". In the case of multi-head attention (3)," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 322, + 518, + 553, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 518, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 322, + 518, + 553, + 581 + ], + "type": "interline_equation", + "content": "\\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {(h)} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(h)} = \\left[ \\begin{array}{l l l} \\mathbf {A} ^ {(1)} & \\dots & \\mathbf {A} ^ {(H)} \\end{array} \\right] \\left[ \\begin{array}{c} \\mathbf {X} \\mathbf {W} _ {v o} ^ {(1)} \\\\ \\vdots \\\\ \\mathbf {X} \\mathbf {W} _ {v o} ^ {(H)} \\end{array} \\right], \\tag {7}", + "image_path": "66c650d513dfaee76aec7a6a25c469bffa9323a750f58b3a4cab2d9f6e79ca6c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": "it can be represented as a product of a composite coefficient " + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{A}^{(1)},\\dots ,\\mathbf{A}^{(H)}\\right]\\in \\mathbb{R}^{N\\times NH}" + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": ", and an extensive dictionary " + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{XW}_{vo}^{(1)},\\dots ,\\mathbf{XW}_{vo}^{(H)}\\right]^{\\intercal}\\in \\mathbb{R}^{NH\\times C_o}" + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": ". Each row of the output " + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": " is a linear combination of " + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{XW}_{vo}" + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i = \\sum_{h,j}\\mathbf{A}_{ij}^{(h)}(\\mathbf{XW}_{vo}^{(h)})_j" + }, + { + "bbox": [ + 313, + 585, + 555, + 713 + ], + "type": "text", + "content": ". When represented as a linear combination of dictionary elements, multi-head attention and single-head attention follow the same formulation. In the following discussion, we focus on the single-head attention formulation, which can be seamlessly extended to the multi-head attention case." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18705" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "A sparse dictionary view of attention. From the perspective of sparse coding, the feature is expressed as SD, where S represents the sparse coefficients and D consists of dictionary atoms that remain data-independent once learned. Although attention can be expressed as a linear combination of atoms, it does not inherently exhibit sparsity. To fully leverage the advantages of sparse representations, we propose to represent the attention mechanism (2) in the form of sparse coding," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 191, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 191, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 116, + 191, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\mathbf {S D} = \\sigma_ {\\lambda} \\left(\\mathbf {A X W} _ {s}\\right) \\mathbf {D}, \\tag {8}", + "image_path": "859d72c47b42378a0d2522932d9949f4eaf574da05ef9d1844da6dcec95b74e7.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\cdot)" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": " can be soft-thresholding function or " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}(x - \\lambda)" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": " to ensure sparsity of " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s) \\in \\mathbb{R}^{N \\times C_o}" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s \\in \\mathbb{R}^{C_i \\times M}" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{M \\times C_o}" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": " is the feature dictionary atoms, " + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 215, + 296, + 262 + ], + "type": "text", + "content": " is the number of atoms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "text", + "content": "With this formulation, we obtain the feature dictionary atoms " + }, + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "text", + "content": " (we use the term \"atoms\" for simplification in this paper), which are the building blocks for feature representation " + }, + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 263, + 296, + 323 + ], + "type": "text", + "content": ". We illustrate our formulation in Figure 1 (a). This formulation provides the following properties:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 515 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": "- The sparse coefficients " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " ensure that the feature representation " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " is combined by only a few " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": ". We present an extreme example to illustrate how atoms enhance interpretability: " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i = \\sum_{j=1}^{M} \\mathbf{S}_{ij} \\mathbf{D}_j" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " can be simplified as " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i = \\mathbf{S}_{ij^*} \\mathbf{D}_{j^*}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " if all elements of " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_i" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " is 0 except the element at index " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "j^*" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": ", and only " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{j^*}" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": " contributes to the output feature " + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 55, + 323, + 296, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "content": "- The sparse coefficients " + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "content": " are data-dependent, while the feature dictionary atoms " + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 407, + 296, + 467 + ], + "type": "text", + "content": " are data-independent. While different inputs generate different coefficients, the output features are always constructed using the same feature dictionary atoms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "text", + "content": "- We can adjust the size of the feature dictionary by varying " + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "text", + "content": " to handle tasks with different difficulties. Large " + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "text", + "content": " leads to an overcomplete dictionary while small " + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 467, + 296, + 515 + ], + "type": "text", + "content": " produces an undercomplete dictionary." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "content": "Fine-tuning via sparse dictionary learning. The above formulation frames feature representation through the lens of sparse coding. In the context of large model fine-tuning, we freeze the pre-trained parameters to make feature representations " + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "content": " of the pre-trained model fixed. Instead, we introduce a sparse coding perspective to encode the adapted feature representation " + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "content": ", which is learned from downstream tasks. Specifically, we have," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 631, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 631, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {O} + \\Delta \\mathbf {O} = \\mathbf {O} + \\sigma_ {\\lambda} (\\mathbf {A X W} _ {s}) \\mathbf {D}. \\tag {9}", + "image_path": "125bd2f14f090ff41ee5f28790c7007cb5884f6897c130b6e3cb7f08d8932ff5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "With this formulation, the pre-trained weights remain unchanged, preserving the capabilities of the pre-trained model. The adapted feature " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is represented as a linear combination of feature dictionary atoms " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", using sparse coefficients " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ". This formulation offers" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 541, + 223 + ], + "type": "image", + "image_path": "37034641eb61d11891b337d66289edce00d0e01b91ddbace63dcab804c2dc3a9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 236, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 302 + ], + "type": "text", + "content": "Figure 2. (a) Adjusting only the coefficients fails to adapt the model to the downstream task, whereas (b) modifying only the atoms enables a perfect fit to the new task composed of high-frequency signals. (c) Fine-tuning the atoms effectively transitions the model from low-frequency to high-frequency representations. They encode the representation of the downstream task." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 354 + ], + "type": "text", + "content": "a framework for understanding how " + }, + { + "bbox": [ + 313, + 318, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 318, + 555, + 354 + ], + "type": "text", + "content": " contributes to the adapted features in the fine-tuned model. We illustrate our method in Figure 1 (a)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 361, + 474, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 474, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 474, + 374 + ], + "type": "text", + "content": "3.3. Analysis of Dictionary Atoms" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "text", + "content": "In this section, we demonstrate distinct properties of atoms " + }, + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "text", + "content": " and sparse coefficients " + }, + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 380, + 554, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": "Features atoms are key to fine-tuning. We analyze the distinct behaviors of coefficients and atoms, and discover that atoms play a crucial role in adapting to new tasks. To explore this within the context of an attention block, we perform a toy experiment where masked one-dimensional signals are reconstructed using an attention layer. The signals, with a length of 64, are synthesized by selecting 5 random Fourier bases, with a low frequency ranging from " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\frac{0}{64}" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\frac{24}{64}" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": ". The objective is to reconstruct the signal after masking out " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": " of its values with only 16 randomly sampled observations. After we have a pre-trained model on the low-frequency signals, we fine-tune the model to a new task where signals are constructed with high-frequency Fourier bases ranging from " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\frac{25}{64}" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\frac{32}{64}" + }, + { + "bbox": [ + 313, + 415, + 555, + 715 + ], + "type": "text", + "content": ". In this experiment, we test two different cases: (1) tuning only coefficients on the high-frequency data with atoms fixed, and (2) tuning only atoms on the high-frequency data with coefficients fixed. As illustrated in Figure 2, when adapting to high-frequency signals, modifying only the atoms while keeping the coefficients fixed allows the model to effectively generalize to the new task. In contrast, updating only the coefficients is insufficient for proper adaptation. Building on this observation, we hypothesize that adjusting atoms is more suitable than adjusting coefficients for adapting to a new task. We observe from Figure 2 (c) that atoms are capable of encoding" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18706" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "the representation of the downstream task. The experiment details are in Appendix B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "text", + "content": "Identifying the importance of feature atoms. We analyze the impact of atoms on the newly learned concept by visualizing both the influence of individual atoms and their combined effect. In this experiment, we adopt a transformer-based variational autoencoder (VAE) [21] with a two-layer decoder. The experiment process is illustrated in Figure 1 (b). We first pre-train the model on an image-denoising task using a subset of the MNIST dataset containing digits 5 to 9. We then fine-tune the model on a downstream denoising task with only digit 3 by learning a dictionary for each layer with " + }, + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "inline_equation", + "content": "M = 100" + }, + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "text", + "content": ". In this scenario, the digit \"3\" is the newly learned concept. Given the same noisy latent, the decoder outputs an \"8\" without incorporating fine-tuned feature atoms, whereas it generates a \"3\" when the fine-tuned feature atoms are applied. The model requires only 12 atoms to shift the generated digits from 8 to 3. The influence of each atom is visualized in Figure 1 (c), with the overall impact represented as " + }, + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 55, + 117, + 297, + 429 + ], + "type": "text", + "content": ", where the red area indicates an increase in pixel value at a specific location and the blue area represents a decrease. It is straightforward to see that these 12 feature atoms enhance the regions corresponding to \"3\" while suppressing certain portions of \"8\". Additionally, Figure 1 (c) illustrates the individual behavior of each atom, showing how different atoms affect distinct parts of the generated digit. The experiment details are in Appendix B." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 441, + 162, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 162, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 162, + 453 + ], + "type": "text", + "content": "3.4. Design Variations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": "In this section, we outline the various design choices implemented in our experiments. Our method achieves sparse coefficient " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\sigma_{\\lambda}(\\mathbf{AXW}_s)" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": " by utilizing non-linear function " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\cdot)" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": ". In the literature on sparse coding [4], there are three different methods: (i) soft-thresholding function, " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{prox}_{\\lambda \\| \\cdot \\|_1}(\\mathbf{x}) = \\mathrm{sign}(\\mathbf{x}) \\odot \\max(|\\mathbf{x}| - \\lambda, 0)" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": ", (ii) ReLU, " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\mathbf{x}) = \\mathrm{ReLU}(\\mathbf{x} - \\lambda) = \\max(\\mathbf{x} - \\lambda, 0)" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": ", and (iii) top-" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": " activation selection, which selects the most significant features of top " + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": " contributing to the learned representation. In practice, we observe that top-" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 460, + 296, + 651 + ], + "type": "text", + "content": " activation selection maintains a consistent level of sparsity across different inputs. In contrast, methods (i) soft-thresholding function and (ii) ReLU exhibit sensitivity to input values, leading to variations in sparsity. However, methods (i) and (ii) offer control over the coefficient values, ensuring they stay above a specified threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Furthermore, we find that when " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " is highly sparse, the resulting product " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " also retains sparsity. In modern transformers, computation of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " are performed separately. For simplicity in implementation, we adopt the formulation " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\sigma_{\\lambda}(\\mathbf{X}\\mathbf{W}_s)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 446, + 85 + ], + "type": "text", + "content": "3.5. Analysis of Our Design" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": "Comparison with LoRA. When representing attention mechanism (1-3) as a linear combination of dictionary atoms, updating " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": " is equivalent to updating the coefficients while updating " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": " is equivalent to updating the atoms. LoRA applies adaptation on " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": " with a low-rank matrices " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_A \\in \\mathbb{R}^{C_i \\times r}" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_B \\in \\mathbb{R}^{r \\times C_o}" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": " represents the rank. LoRA adapts both coefficients and atoms in a low-rank way. In comparison, our method adapts the coefficients with " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": " and obtains sparse coefficients " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\mathbf{AXW}_s)" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": ", while representing atoms with " + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 90, + 555, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": "Parameters and computational cost. To adapt the representation at each layer, we need additional parameters for " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", which are " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "(C_i + C_o)M" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ". After the fine-tuning, considering the density of the coefficients " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "1 - \\rho" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": " is sparsity, the storage required for the parameters is " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "C_i M + \\rho C_o M" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ". In comparison, LoRA requires " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "4(C_i + C_o)r" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": " additional parameters. The FLOPs required by our method are " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "2N M C_i + 2\\rho N M C_o" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ". LoRA requires " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "8C_i C_o r" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": " FLOPs. For example, if " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "C_i = C_o = 1024" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "M = 256" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "N = 1024" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "inline_equation", + "content": "\\rho = 0.01" + }, + { + "bbox": [ + 313, + 223, + 556, + 426 + ], + "type": "text", + "content": ", our method requires 264k parameters for storage and 0.54GFLOPs for computation, while LoRA requires 131k parameters and 0.13GFLOps. Our method involves slightly higher computational costs than LoRA, as it accounts for feature interactions, whereas LoRA only updates the weights. Next, we demonstrate that by leveraging sparse feature representations, our approach achieves greater robustness compared to LoRA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "Our method provides more stable adaptation. We compare an extreme case between LoRA and our method, by considering the attention in a dictionary form " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O} = \\mathbf{A}\\mathbf{V}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O} = \\mathbf{SD}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": ". In this experiment, we examine the difference between updating coefficients using a low-rank approach versus a sparse approach. Both LoRA and our method produce the updated feature " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " with rank 1, where LoRA achieves this by setting " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " and our method enforces only one active atom in D. The key difference lies in how the coefficients are updated—LoRA adapts them in a low-rank manner, whereas our method updates them sparsely. As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. LoRA fails to produce meaningful images with a slightly modified prompt, while our method remains robust, consistently generating a concept that aligns with the updated text prompt. We attribute the effectiveness of our approach to the sparse coefficients produced by " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{\\lambda}(\\mathbf{AXW}_s)" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": ". With a different text prompt, the values in S either remain zero or take on a significant value when receiving the proper input, ensuring that the generated result remains coherent. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18707" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 71, + 517, + 272 + ], + "blocks": [ + { + "bbox": [ + 95, + 71, + 517, + 272 + ], + "lines": [ + { + "bbox": [ + 95, + 71, + 517, + 272 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 517, + 272 + ], + "type": "image", + "image_path": "cbb7deb14fb6089e4f69d50f446103fb7166aca5d7a08551e59194b5ed405e56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "type": "text", + "content": "Figure 3. The pre-trained model can generate the target concept when integrated with the learned dictionary, even with about 40 atoms. With just four atoms, the core structure of the concept can still be generated. Adjusting the number of selected atoms influences the generated results, offering a controllable approach to image editing. Additional discussions of these figures are in Section 4.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 327, + 289, + 507 + ], + "blocks": [ + { + "bbox": [ + 66, + 327, + 289, + 507 + ], + "lines": [ + { + "bbox": [ + 66, + 327, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 66, + 327, + 289, + 507 + ], + "type": "image", + "image_path": "8216130cf24de06def538ebcb8bfbaa1681552147f8a4a0e26223a49d1451517.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 517, + 297, + 584 + ], + "lines": [ + { + "bbox": [ + 55, + 517, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 297, + 584 + ], + "type": "text", + "content": "Figure 4. The role of sparsity: Our method utilizes the same number of atoms as LoRA, yet LoRA struggles to adapt to the target prompt effectively. Sparse coding ensures that atoms are either activated or inactive, providing a structured and interpretable adaptation process. In contrast, LoRA produces noisy coefficients, leading to completely disrupted outputs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of our method in image generation tasks. We demonstrate that feature dictionary atoms effectively represent the target concept after fine-tuning the pre-trained model. Compared to baselines, our approach enables selective atom adjustment, improving image editing by refining results and enhancing text-to-image concept customization while preserving pre-trained model capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 325, + 508, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 508, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 508, + 338 + ], + "type": "text", + "content": "4.1. Representing Concepts Using Atoms" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": "Experimental settings. In this experiment, we utilize PixArt- " + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 342, + 555, + 450 + ], + "type": "text", + "content": " [7], a Diffusion Transformer (DiT) [40] model consisting of 28 transformer blocks. We fine-tune the model on a single image for 60 epochs using our method and reconstruct the corresponding concept using the same prompt used during fine-tuning. The selected images are drawn from DreamBooth [47], Custom Diffusion [23]. We control the number of active atoms and visualize the variation of the generated results with different atoms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 545 + ], + "type": "text", + "content": "The concept is represented by a set of atoms. As shown in Figure 3, after fine-tuning our method on the target concept, its representation can be effectively captured using a small set of atoms, such as 40 atoms. Among these, just 4 atoms are sufficient to construct the core structure of the concept, including key attributes like pose and color. We provide details descriptions of these images in Appendix C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 405, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 405, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 405, + 565 + ], + "type": "text", + "content": "4.2. Image Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Building on our previous observation that adjusting the number of atoms influences the generated results, we explore leveraging this property for image editing. Image editing tasks aim to modify an image using a generative model while retaining as many details as possible from the original [63]. However, certain editing tasks pose challenges in maintaining most of the original details while achieving effective modifications. For example, altering the style can significantly change the overall appearance of the image. We argue that successful image editing should preserve the main components of the image while applying modifications selectively. Since our method represents the core" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18708" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 287, + 335 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 335 + ], + "type": "image", + "image_path": "36b33affdd79a29e0e81d0545200c0b59997033b1c86c600b662afd65bac9bf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 381 + ], + "type": "text", + "content": "Figure 5. Results on image editing. Our method shows improved results by better following editing prompts while preserving the key components of the source images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 444 + ], + "type": "text", + "content": "components of a concept using only a few atoms, it is well-suited for image editing tasks, ensuring that key structural elements remain intact while allowing for controlled transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 616 + ], + "type": "text", + "content": "Experimental settings. In our experiment, we select data from previous work [63] and evaluate different editing tasks, including background modification, scene alteration, and pose adaptation. Previous works [15, 63] fine-tune models to learn a concept and modify its appearance by adjusting the prompt. However, these approaches have not yet been applied to DiT. Following their setup, we compare our method with fine-tuning approaches, such as LoRA [17], using them as baselines for evaluation. We adopt the DiT architecture and fine-tune both LoRA and our method on a single image for 60 epochs to achieve full reconstruction. After fine-tuning, we modify the image using a different prompt for editing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Image editing results. The results are shown in Figure 5. Our method applies edits based on the test prompt while maintaining the core structure of the source images. In addition, we control the number of active atoms by adjusting the coefficient density " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "1 - \\rho" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " represents the sparsity level. In Figure 6, as sparsity increases, only the most essential atoms remain, preserving the core concept while al" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 397, + 144 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 397, + 144 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 397, + 144 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 397, + 144 + ], + "type": "image", + "image_path": "5b93197756b6e23e6edfc00ca4b60b4d148bdfca649748376b52735d4f4218ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 148, + 382, + 158 + ], + "lines": [ + { + "bbox": [ + 348, + 148, + 382, + 158 + ], + "spans": [ + { + "bbox": [ + 348, + 148, + 382, + 158 + ], + "type": "inline_equation", + "content": "\\rho = 0.01" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 471, + 145 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 471, + 145 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 471, + 145 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 471, + 145 + ], + "type": "image", + "image_path": "82b26ff313d9e7fb87aafc67cce219ae5790621996dc2649db2a7586d60fb039.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 148, + 454, + 158 + ], + "lines": [ + { + "bbox": [ + 420, + 148, + 454, + 158 + ], + "spans": [ + { + "bbox": [ + 420, + 148, + 454, + 158 + ], + "type": "inline_equation", + "content": "\\rho = 0.02" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 543, + 145 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 543, + 145 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 543, + 145 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 543, + 145 + ], + "type": "image", + "image_path": "351038a862d7318ed76948c56ed0396141491118b19eea0e699f3c98fac29398.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 148, + 525, + 157 + ], + "lines": [ + { + "bbox": [ + 490, + 148, + 525, + 157 + ], + "spans": [ + { + "bbox": [ + 490, + 148, + 525, + 157 + ], + "type": "inline_equation", + "content": "\\rho = 0.04" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 168, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 191 + ], + "type": "text", + "content": "Figure 6. Adjusting the density, i.e., the number of activated atoms, can change the influence on the edited image." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "type": "text", + "content": "lowing the pre-trained model greater flexibility to integrate editing effects based on the modified text prompt." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 447, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 447, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 447, + 248 + ], + "type": "text", + "content": "4.3. Concept Customization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 313 + ], + "type": "text", + "content": "In the previous section, we demonstrated that our method can efficiently represent the target concept using only a small number of atoms. In this section, we further show that even with just a few atoms, our method achieves good performance in concept customization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": "Experimental settings. We evaluated our approach on the DreamBooth [47] dataset, which consists of 30 subjects across 15 distinct classes. Among these, 9 are live subjects, while the remaining 21 are objects. Each subject has between 4 to 6 images, captured under varying conditions, in diverse environments, and from multiple angles. All 30 subjects were trained for 15 epochs and validated on 25 different prompts. For our experiments, we adopted the Pixart- " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": ", a Diffusion Transformer (DiT) [7], as the base architecture and fine-tuned it using our proposed method. To verify the effectiveness of our approach, we compared its performance against several PEFT baselines for personalized text-to-image generation, including LoRA [17], DoRA [26], and OFT [43]. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": " to fine-tune the Pixart- " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": " [7] for 15 steps. For the baseline methods, LoRA and DoRA are assigned a rank of " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": ", while OFT is set to " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": ". We generated five images for each prompt at a resolution of " + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "inline_equation", + "content": "1024\\times 1024" + }, + { + "bbox": [ + 313, + 326, + 556, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 675 + ], + "type": "text", + "content": "Evaluation metrics. We measure concept alignment by computing the mean cosine similarity between the generated image's CLIP embeddings [44] and those of the original concept images. We quantify image diversity using the Vendi score [13] with DINOv2 embeddings [38]. For text alignment, we calculate the average cosine similarity in the CLIP feature space [44]. ImageReward [59] is a general-purpose text-to-image human preference reward model that evaluates and ranks AI-generated images by scoring them based on human preferences." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Customization results. As shown in Table 1, compared with baseline methods, our approach not only represents" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18709" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "type": "text", + "content": "Table 1. Comparison with baseline methods on concept customization." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 294, + 175 + ], + "type": "table", + "html": "
AlignmentDiversityFidelityImageReward
LoRA [17]0.2184.7070.7110.220
DoRA [26]0.2134.5400.7120.142
OFT [43]0.1903.9260.6630.137
Ours (ρ = 0.02)0.2355.2050.7020.182
Ours (ρ = 0.04)0.2315.1720.7130.431
", + "image_path": "1a48b9147cf82ee65d3de01bd9dd04dafccc7a467ef22041d60b48be325123f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 291 + ], + "type": "text", + "content": "the concept with a minimal set of learned atoms but also generalizes robustly to diverse scenarios. In particular, our method achieves higher text-to-image alignment and diversity scores than most other PEFT baselines, while still preserving a relatively high fidelity score. This indicates that our approach effectively captures essential features of each concept without sacrificing the quality of customization text-to-image generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "content": "Number of parameters in different methods. We compare the trainable parameters of our method and the baselines. The pre-trained PixArt model has around 616M parameters in total. Among the fine-tuning methods, both LoRA and DoRA have around 4M trainable parameters. OFT uses around 37M trainable parameters, and our method trains around 17M parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 155, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 155, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 155, + 411 + ], + "type": "text", + "content": "4.4. Abalation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "text", + "content": "The impact of sparsity in the coefficients. We analyze the impact of varying sparsity levels in the coefficients through a concept customization experiment, evaluating text alignment, image diversity, and image fidelity. As shown in Figure 7, higher sparsity (low density) enhances text alignment and image diversity but results in slightly lower image fidelity. In contrast, lower sparsity (high density) improves fidelity, indicating that using more atoms allows for a more precise representation of the concept from the source image. In our experiment, we adopt both " + }, + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\rho = 0.02" + }, + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "inline_equation", + "content": "\\rho = 0.04" + }, + { + "bbox": [ + 55, + 416, + 296, + 561 + ], + "type": "text", + "content": " for either better diversity or better fidelity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "text", + "content": "The impact of member numbers " + }, + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "text", + "content": ". We also examine the impact of the number of members, i.e., the dictionary size, on the generated results while maintaining the same level of sparsity. As shown in Table 2, increasing the number of members improves image fidelity, whereas a smaller dictionary size reduces fidelity but enhances text alignment and image diversity. In our experiment, we choose " + }, + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "inline_equation", + "content": "M = 256" + }, + { + "bbox": [ + 55, + 571, + 296, + 667 + ], + "type": "text", + "content": " to provide balanced fidelity and diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Sparsity coefficients provide more stable adaptation. In this experiment, we compare our method with LoRA on the image editing task. As we discussed in Section 3.5," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 76, + 550, + 194 + ], + "blocks": [ + { + "bbox": [ + 320, + 76, + 550, + 194 + ], + "lines": [ + { + "bbox": [ + 320, + 76, + 550, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 76, + 550, + 194 + ], + "type": "image", + "image_path": "395defdb074916240743b6a3b1336baf7c6175f7c76b1b0811432ebf039c142c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "text", + "content": "Figure 7. The influence of different density " + }, + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "text", + "content": " on the image diversity, fidelity and text-to-image alignment." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 352, + 285, + 512, + 345 + ], + "blocks": [ + { + "bbox": [ + 313, + 243, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 277 + ], + "type": "text", + "content": "Table 2. The influence of different member numbers " + }, + { + "bbox": [ + 313, + 243, + 553, + 277 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 243, + 553, + 277 + ], + "type": "text", + "content": " . As the number of members increases,alignment and diversity scores decrease, while the fidelity score improves." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 285, + 512, + 345 + ], + "lines": [ + { + "bbox": [ + 352, + 285, + 512, + 345 + ], + "spans": [ + { + "bbox": [ + 352, + 285, + 512, + 345 + ], + "type": "table", + "html": "
ρ = 0.01AlignmentDiversityFidelity
M = 160.2485.8670.547
M = 2560.2494.3410.650
M = 10240.2413.7330.672
M = 20480.2383.5250.692
", + "image_path": "9ede1e65e994380147004cdd7b9f00e113559aec4fdf0ed46cfe45365e263b20.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "text", + "content": "LoRA adapts the coefficients in a low-rank manner, whereas our method updates them sparsely. Both LoRA and our method produce the updated feature " + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{O}" + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "text", + "content": " with rank 1, where LoRA achieves this by setting " + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "text", + "content": " and our method enforces only one active atom in " + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 359, + 555, + 515 + ], + "type": "text", + "content": ". As shown in Figure 4, by fine-tuning the model on a single concept, both LoRA and our method successfully reconstruct the concept using the same prompt. However, when applying the fine-tuned weights to generate images with a slightly modified prompt, LoRA becomes unstable, failing to produce meaningful images. In contrast, our method remains robust, consistently generating a concept that aligns with the updated text prompt." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": "In this paper, we introduce to frame model fine-tuning as a sparse coding problem, where the adapted feature representation is constructed using a dictionary of feature atoms. We propose a fine-tuning framework that leverages a combination of feature dictionary atoms, demonstrating that tuning them enhances control over the adaptation process. We validate the effectiveness of our approach across various generative tasks. By representing concepts with only a few essential atoms, our method proves to be highly adaptable and can be effectively applied to image editing, style mixing, and concept personalization. Our study provides a new perspective on the mechanics of fine-tuning, paving the way for more controllable adaptation strategies in large-scale pretrained models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18710" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. Decomposing and interpreting image representations via text in vits beyond CLIP. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 179 + ], + "type": "text", + "content": "[2] Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM journal on imaging sciences, 2009. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 212 + ], + "type": "text", + "content": "[3] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 256 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35(8):1798-1828, 2013. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 356 + ], + "type": "text", + "content": "[5] Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nicholas L Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Alex Tamkin, Karina Nguyen, Brayden McLean, Josiah E Burke, Tristan Hume, Shan Carter, Tom Henighan, and Chris Olah. Towards monoseismicity: Decomposing language models with dictionary learning. 2023. Transformer Circuits Thread. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 294, + 390 + ], + "type": "text", + "content": "[6] Emmanuel J Candès et al. Compressive sampling. In Proceedings of the international congress of mathematicians, pages 1433-1452, 2006. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "text", + "content": "[7] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 62, + 392, + 294, + 456 + ], + "type": "text", + "content": " : Weak-to-strong training of diffusion transformer for 4k text-to-image generation. In European Conference on Computer Vision, pages 74–91, 2024. 6, 7, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 294, + 490 + ], + "type": "text", + "content": "[8] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 523 + ], + "type": "text", + "content": "[9] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "text", + "content": "[10] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 580 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 613 + ], + "type": "text", + "content": "[12] Jacob Dunefsky, Philippe Chlenski, and Neel Nanda. Transcoders enable fine-grained interpretable circuit analysis for language models. 2024. AI Alignment Forum. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 647 + ], + "type": "text", + "content": "[13] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 712 + ], + "type": "text", + "content": "[14] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models. Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In CVPR, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[16] Shaozhe Hao, Kai Han, Zhengyao Lv, Shihao Zhao, and Kwan-Yee K Wong. Conceptexpress: Harnessing diffusion models for single-image unsupervised concept extraction. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 196 + ], + "type": "text", + "content": "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[18] Teng Hu, Jiangning Zhang, Ran Yi, Hongrui Huang, Yabiao Wang, and Lizhuang Ma. Sara: High-efficient diffusion model fine-tuning with progressive sparse low-rank adaptation. arXiv preprint arXiv:2409.06633, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "text", + "content": "[19] Jeeyung Kim, Erfan Esmaeili, and Qiang Qiu. Text embedding is not all you need: Attention control for text-to-image semantic alignment with text self-attention maps. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8031–8040, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[20] Taewook Kim, Wei Chen, and Qiang Qiu. Learning to customize text-to-image diffusion in diverse context. arXiv preprint arXiv:2410.10058, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[21] Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. Foundations and Trends® in Machine Learning, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 422 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "type": "text", + "content": "[23] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[24] Xunhao Lai, Jianqiao Lu, Yao Luo, Yiyuan Ma, and Xun Zhou. Flexprefill: A context-aware sparse attention mechanism for efficient long-sequence inference. arXiv preprint arXiv:2502.20766, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[25] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 624 + ], + "type": "text", + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 1, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[27] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "text", + "content": "[28] Yiyang Liu, James Chenhao Liang, Ruixiang Tang, Yugyung Lee, Majid Rabbani, Sohail Dianat, Raghuveer Rao, Lifu Huang, Dongfang Liu, Qifan Wang, et al. Re-imagining" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18711" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "multimodal instruction tuning: A representation view. arXiv preprint arXiv:2503.00723, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[29] Yang Luo, Xiaozhe Ren, Zangwei Zheng, Zhuo Jiang, Xin Jiang, and Yang You. Came: Confidence-guided adaptive memory efficient optimization. arXiv preprint arXiv:2307.02047, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 174 + ], + "type": "text", + "content": "[30] Julien Mairal, Jean Ponce, Guillermo Sapiro, Andrew Zisserman, and Francis Bach. Supervised dictionary learning. Advances in neural information processing systems, 2008. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 219 + ], + "type": "text", + "content": "[31] Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 253 + ], + "type": "text", + "content": "[32] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 255, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 297 + ], + "type": "text", + "content": "[33] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 364 + ], + "type": "text", + "content": "[34] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[35] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 455 + ], + "type": "text", + "content": "[36] Zichen Miao, Wei Chen, and Qiang Qiu. Coeff-tuning: A graph filter subspace view for tuning attention-based large models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20146-20157, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 489 + ], + "type": "text", + "content": "[37] Bruno A Olshausen and David J Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, pages 607-609, 1996. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 544 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 599 + ], + "type": "text", + "content": "[39] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 623 + ], + "type": "text", + "content": "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In CVPR, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[41] Vishal Purohit, Matthew Repasky, Jianfeng Lu, Qiang Qiu, Yao Xie, and Xiuyuan Cheng. Posterior sampling via Langevin dynamics based on generative priors. arXiv preprint arXiv:2410.02078, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[42] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 1, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 192 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 194, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 247 + ], + "type": "text", + "content": "[45] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 358 + ], + "type": "text", + "content": "[47] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 6, 7, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 415 + ], + "type": "text", + "content": "[48] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuzhhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In European Conference on Computer Vision, pages 422-438. Springer, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "text", + "content": "[49] Viacheslav Surkov, Chris Wendler, Mikhail Terekhov, Justin Deschenaux, Robert West, and Caglar Gulcehre. Unpacking sdxl turbo: Interpreting text-to-image models with sparse autoencoders. arXiv preprint arXiv:2410.22366, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[50] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International conference on machine learning, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 494, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 535 + ], + "type": "text", + "content": "[51] Nathaniel Tomczak and Sanmukh Kuppannagari. Longer attention span: Increasing transformer context length with sparse graph processing techniques. arXiv preprint arXiv:2502.01659, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 581 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 2017. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[53] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[54] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[55] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18712" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[56] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[57] Muling Wu, Wenhao Liu, Xiaohua Wang, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaqing Zheng, and Xuanjing Huang. Advancing parameter efficiency in fine-tuning via representation editing. 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[58] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[59] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 328 + ], + "type": "text", + "content": "[60] Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, and Wangding Zeng. Native sparse attention: Hardware-aligned and natively trainable sparse attention, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[61] Yang Zhang, Teoh Tze Tzun, Lim Wei Hern, and Kenji Kawaguchi. Enhancing semantic fidelity in text-to-image synthesis: Attention regulation in diffusion models. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[62] Yasi Zhang, Peiyu Yu, and Ying Nian Wu. Object-conditioned energy-based attention map alignment in text-to-image diffusion models. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "text", + "content": "[63] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "text", + "content": "[64] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18713" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_content_list.json b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9ec86cb5026965731fa325e86f5c608961cdcb7a --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_content_list.json @@ -0,0 +1,1660 @@ +[ + { + "type": "text", + "text": "Sparse-Dense Side-Tuner for efficient Video Temporal Grounding", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "David Pujol-Perich, Sergio Escalera, Albert Clapés \nUniversitat de Barcelona and Computer Vision Center, Barcelona, Spain \n{david.pujolperich, sescalera, aclapes}@ub.edu", + "bbox": [ + 212, + 180, + 782, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full finetuning is often impractical, parameter-efficient fine-tuning – and particularly side-tuning (ST)– has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention –a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a $73\\%$ the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST.", + "bbox": [ + 89, + 301, + 485, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 676, + 220, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently the field of Video Understanding has gained attention due to its potential in applications like search engines or recommendation systems. A key task in this field is Video Temporal Grounding (VTG), which localizes moments within videos based on textual descriptions. Typically, this involves doing both Moment Retrieval (MR) [2, 5, 7, 11, 15] and Highlight Detection (HD) [1, 8, 10, 33, 40]. More specifically, MR predicts moment boundaries w.r.t. textual queries, while HD offers a more interpretable perspective, predicting frame-level saliency scores.", + "bbox": [ + 89, + 703, + 482, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past, existing works focused on either MR or HD, given the lack of datasets that combined both tasks. QVHighlights [15] shifted this paradigm, proposing the first", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/245f34659e271e7f11ba68cd7ae960c6512b826dafc8cbc103298dea05a31ace.jpg", + "image_caption": [ + "Figure 1. Comparison of our proposed method and the main MR (left) and HD (right) baselines. These are all evaluated on QVHighlights val split and use InternVideo2-1B features. These results show that our method improves existing side-tuning works and attains SOTA results while significantly reducing the number of trainable parameters." + ], + "image_footnote": [], + "bbox": [ + 514, + 267, + 709, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa0fc1c28d03ed50e28f05ad4f339fc3ab5441bfbbe68f19b090b04042d9d9ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 268, + 903, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dataset suitable for this multi-task setup. Since then, various works have been proposed following anchor-based [2, 17, 21, 22], anchor-free [7, 15, 25] and even LLM-based [23, 24] approaches. All of these, however, incur in a common critical limitation, which is relying on the final-layer features of frozen backbones. This is especially limiting when facing large distribution shifts between the pre-training distribution and that of the downstream task [29]. This issue is particularly pronounced in VTG, where an image-domain backbone [31] is transferred to the video domain.", + "bbox": [ + 511, + 489, + 908, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A typical solution is fine-tuning; however, full fine-tuning is impractical due to its high computational cost. Doing this more efficiently has thus become critical, motivating the interest in parameter-efficient fine-tuning (PEFT) [9] methods -e.g., Prompting or Adapters- which optimize only a small subset of parameters. Unfortunately, these still require full back-propagation through the backbone, making it still memory-intensive. To address this, [34] introduced side-tuning (ST), an effective PEFT, and memory-efficient fine-tuning (MEFT) method. ST creates a parallel pathway to refine intermediate features, thus backpropagating over a minimal set of parameters only. In this regard, we highlight the work of $\\mathbb{R}^2$ -Tuning [21], the first ST approach for VTG, which recursively fuses multi-modal intermediate CLIP [31] embeddings. [21] also proposes adapting these frame-level embeddings for MR by generating a dense set of anchors. As shown in Fig. 1, this", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21515", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "method proves ineffective for MR - a highly sparse task where videos may contain very few ground-truth action.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This motivates our dual-stream Sparse-Dense Side-Tuner (SDST), the first proposal-free ST method, carefully designed for multi-task learning across sparse (MR) and dense (HD) tasks. For this, SDST jointly refines frame-level embeddings -suitable for HD- while learning what we call the recurrent decoder queries for MR. We also identify an implicit contextual limitation of the deformable attention [45] module of anchor-free architectures like ours, which in our task results in collapsing offsets around their initialization values. This also impedes the selection of keys outside of the current estimated moment boundaries, critical to potentially refine the boundaries of longer moments. Consequently, we propose the alternative Reference-based Deformable Self-Attention (RDSA), which naturally addresses this issue by reformulating the CA into a SA-based mechanism. Finally, we tackle the performance degradation of ST methods derived from the use of image-based CLIP, over more advanced spatio-temporal VLMs [36], as noted in [7]. A key challenge of exploiting [36] is, however, defining an effective token pooling strategy, as naive strategies like CLS pooling yield significant performance drops. This motivates our proposed module re-utilization scheme, which results in the first successful integration of InternVideo2 [36] into an ST framework.", + "bbox": [ + 91, + 123, + 482, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are threefold: 1) We propose SDST, the first anchor-free ST architecture, specifically tailored for complex sparse-dense multi-task setups like VTG. Our method significantly outperforms existing ST architectures on QVHighlights [15], TACoS [32] and Charades-STA [5] (see Fig. 1). SDST also performs competitively and even surpasses existing SOTA while reducing its learnable parameter count by up to $73\\%$ , and incurring in a minimal memory overhead. 2) We identify a key contextual limitation of the deformable attention mechanism that, in our task, results in collapsing offsets. Consequently, we propose RDSA, which tackles this issue by allowing more complex key-selection strategies. 3) We address the limitations of transfer learning from an image-domain backbone to a video domain by integrating, for the first time, the more advanced spatio-temporal backbone [36] into an ST framework. This proves to be a non-trivial challenge with massive performance implications.", + "bbox": [ + 89, + 484, + 482, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 768, + 227, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video temporal grounding (VTG): One relevant task in Video Understanding is VTG [17] which aims to identify actions from text queries. Traditionally, this task was approached either from a MR [2, 5, 7, 11, 15, 25, 26] or HD [1, 8, 10, 33, 40] perspective. MR focuses on predicting specific action proposals while HD aims to compute the saliency scores for each of the frames w.r.t. the query. Nev", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ertheless, after the proposal of QVHighlights [15], the literature started approaching this from a multi-task learning perspective. In this regard, arguably the most notable family of methods is that based on DETR [7, 14, 15, 25, 26, 39] given the sparse nature of MR. These methods normally fuse both video and textual modality into a final embedding which is used for HD and as a memory of a Transformer decoder that refines a set of learnable queries. More in detail, Moment-DETR [15] constitutes the first baseline, which uses a standard Transformer encoder to process both modalities simultaneously. QD-DETR [26] proposes instead to inject the textual information to the video modality with cross attention (CA), followed by a temporal modeling module. CG-DETR [25] reduces the negative impact of irrelevant textual tokens by adding an Adaptive CA module that leverages dummy tokens to redirect attention weight. In this same line, SG-DETR [7] introduces a saliency-guided CA, weighting the contribution of irrelevant textual tokens based on the computed saliency scores. Despite the notable success of these proposal-free methods, one core limitation that they all share is that they rely solely on the last-layer features of a frozen CLIP [31] and/or Slowfast [4] backbones. [41] overcomes this limitation via the full finetuning of CLIP, even though this becomes nearly intractable given the resources that this requires. $\\mathbb{R}^2$ -Tuning [21], the most similar method to our work, overcomes this limitation by proposing the first PEFT and MEFT method for VTG, which recursively applies a multi-modal fusion ST. In this work, we argue that one of its core drawbacks, however, is its anchor-based nature, remaining oblivious to the shown benefits of DETR-based architectures in very sparse tasks like MR [7, 16, 20]. This motivates our proposed SDST, which incorporates the best of both PEFT and MEFT.", + "bbox": [ + 511, + 90, + 906, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-and-memory-efficient fine-tuning: Foundation models and VLMs [31, 37] have become the cornerstone of recent advances in Video Understanding applications. The use of pre-extracted features remains, however, an important limitation to its application on downstream tasks like VTG [17]. Fine-tuning these huge models is often simply infeasible given the large resources that these require. This has motivated the rise of PEFT [9], which aims to democratize the fine-tuning process by restricting the tunable parameters to the bare minimum. In this regard, some of the most prominent approaches are Prompting [12, 44] and the use of Adapters [6, 28]. The first keeps the backbone frozen while learning prompts that are appended to the input, bridging the gap between the backbone's expected distribution and that of the downstream task. Adapters, in contrast, incorporate small trainable modules in the backbone while keeping the rest unchanged. These methods, however, still require full backpropagation through the frozen model, making them memory-inefficient. Recently, the new paradigm of $ST$ [19, 30, 34] –and particularly [21]", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21516", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for VTG- has gained attraction creating a parallel pathway to exploit intermediate backbone representations while guaranteeing that backpropagation is only applied to this small parallel module. Interestingly, these methods normally rely on pre-extracted CLIP representations for tasks like HD/MR. As shown by [7], this is a very limiting aspect given the lack of temporal reasoning of CLIP or its difficulties in understanding textual queries beyond simple spatial descriptions. For this reason, in this work, we propose a novel ST architecture that, for the first time, relies on the more advanced InternVideo2 [36] backbone. Unlike CLIP, this backbone is trained on video-domain inputs, resulting in enhanced spatial and temporal modeling capabilities.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deformable attention: A key component of most SOTA methods for VTG like SG-DETR [7] is their deformable attention module [38, 45]. This module addresses the well-studied slow convergence of DETR [3]. It does so by limiting the selectable keys based on the predictions of the learnable queries, completely decoupling both query and key spaces. Despite desirable in terms of efficiency, it has been previously noted in the literature [43] that the lack of context of the queries w.r.t. the key/value space can lead to suboptimal attendable key selection. VTG methods often mitigate this effect by partially or completely initializing the learnable queries based on the DETR-memory [43] -i.e., frame-level representations in our case. In this work, we empirically demonstrate the limitations of these initialization-based methods for VTG tasks, which motivates our proposed simple yet effective alternative, the Reference-based Deformable Self-Attention (RDSA). This reformulates the deformable CA into a deformable SA, naturally solving the aforementioned issue.", + "bbox": [ + 91, + 287, + 483, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 582, + 181, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem definition", + "text_level": 1, + "bbox": [ + 89, + 606, + 269, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we address the problem of MR and HD based on textual descriptions. For this, we consider an arbitrary input-query pair $(\\mathbf{X}^v,\\mathbf{X}^t)$ where $\\mathbf{X}^v\\in \\mathbb{R}^{T\\times H\\times W\\times 3}$ and $\\mathbf{X}^t\\in \\mathbb{R}^{L\\times F_e}$ . Here $T$ and $L$ are the number of frames and tokens, and $F_{e}$ the textual encoding dimension. Our goal is to predict, for the given video-query pair, its respective frame-wise saliency scores $\\mathbf{Y}^s\\in \\mathbb{R}^T$ for HD, and the $M$ action moments $\\mathbf{Y}^m\\in \\mathbb{R}^{M\\times 2}$ for MR.", + "bbox": [ + 89, + 628, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 89, + 757, + 200, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work we propose a dual-stream ST architecture that includes a dense (frame-level) and a sparse (segment-level) stream for MR and HD respectively (see Fig. 2a). More specifically, we first leverage a frozen InternVideo2-1B [36] backbone, a model that possesses powerful spatio-temporal modeling capabilities, to extract $K$ intermediate visual-textual representations. These are then processed by multiple weight-sharing SDST layers, which refine dense rep", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentations and recurrent decoder queries. More in detail, at each level, the model refines the dense embeddings via multi-modal textual conditioning as well as temporal modeling. This signal is then used to condition the sparse stream $S$ , that builds on [16] while incorporating our novel deformable attention mechanism, namely Reference-based Deformable Self-Attention (see Fig. 2b). This mitigates the shortcomings of the deformable CA mechanism, enhancing the contextual information of the decoder queries and thus improving the quality of the selected keys. After unrolling this process $K$ times, we apply the dense and the sparse prediction heads to compute the saliency scores and the predicted segment boundaries, respectively.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Sparse-Dense Side-Tuner (SDST)", + "text_level": 1, + "bbox": [ + 511, + 306, + 805, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Recurrent refinement with side-tuners", + "text_level": 1, + "bbox": [ + 511, + 330, + 825, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work builds on an ST framework, which first requires extracting $K$ intermediate video and textual features $\\tilde{\\mathbf{V}}\\in$ $\\mathbb{R}^{K\\times T\\times F_v}$ and $\\tilde{\\mathbf{T}}\\in \\mathbb{R}^{K\\times L\\times F_t}$ (find the details in Sec. 4). Here $F_{v}$ and $F_{t}$ are their respective dimensionalities. We then define a zero-initialized dense embedding $\\mathbf{D}^0\\in \\mathbb{R}^{T\\times F}$ and a set of $M$ learnable recurrent moment proposals inspired by [20], which we refer to as recurrent decoder queries. The queries include the learnable center-width moment references $\\mathbf{R}^0\\in \\mathbb{R}^{M\\times 2}$ , and their corresponding latent embeddings $\\mathbf{H}^0\\in \\mathbb{R}^{M\\times F}$ . We then define the recurrence for a given level $1\\leq \\ell \\leq K$ as follows:", + "bbox": [ + 511, + 353, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {\\ell + 1}, \\mathbf {R} ^ {\\ell + 1}, \\mathbf {H} ^ {\\ell + 1} = S D S T \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\tilde {\\mathbf {V}} ^ {\\ell}, \\tilde {\\mathbf {T}} ^ {\\ell}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 521, + 903, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where SDST represents the shared-weight layers of our model, progressively refining the representations across two streams: the 1) dense learning stream $\\mathcal{D}$ and the 2) sparse learning stream $S$ , that we describe below.", + "bbox": [ + 511, + 544, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.2. Dense learning stream $\\mathcal{D}$", + "text_level": 1, + "bbox": [ + 511, + 622, + 733, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [21], this stream first refines the dense embeddings $\\mathbf{D}^{\\ell}$ by conditioning it to the video-textual embeddings $\\tilde{\\mathbf{V}}^{\\ell}$ and $\\tilde{\\mathbf{T}}^{\\ell}$ . For this, both of these multi-modal embeddings are projected to a shared $F$ -dimensional space, using two MLPs $(\\mathcal{F}_v:F_v\\to F$ and $\\mathcal{F}_t:F_t\\to F)$ :", + "bbox": [ + 511, + 643, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} ^ {\\ell} = \\mathcal {F} _ {v} \\left(\\tilde {\\mathbf {V}} ^ {\\ell}\\right), \\mathbf {T} ^ {\\ell} = \\mathcal {F} _ {t} \\left(\\tilde {\\mathbf {T}} ^ {\\ell}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 722, + 903, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then we incorporate the clip-wise visual information to $\\mathbf{D}^{\\ell}$ via a weighted sum modulated by $\\beta^{\\ell} \\in [0,1]$ , a zero-initialized layer-dependent parameter that yields:", + "bbox": [ + 511, + 744, + 905, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {\\ell} := \\beta^ {\\ell} \\mathbf {D} ^ {\\ell} + (1 - \\beta^ {\\ell}) \\mathbf {V} ^ {\\ell}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 792, + 903, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thereafter, we inject the textual information with CA, where $\\mathbf{T}^{\\ell}$ does not include the CLS token. We also apply a temporal modeling module defined as a SA and a subsequent point-wise feed forward network (PFFN):", + "bbox": [ + 511, + 815, + 905, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {\\ell + 1} = P F F N \\left(S A \\left(C A \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 878, + 903, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70f0a1d87307015aebb2100906a00ad80a96633368a9a88475129bdd570cf72e.jpg", + "image_caption": [ + "(a) Overview of our proposed architecture." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 604, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d94d93dd84419ed993f4e092c9c72f14e9c5a29a2ee0cf177334e97d4a0dd41f.jpg", + "image_caption": [ + "(b) RDSA vs Def.CA[38] for a proposal $m$ .", + "Figure 2. Our method (left) first processes the video and textual inputs using [36], and then recursively applies a shared SDST, a recurrent dual-stream model attached to the last $K$ layers. This iteratively refines dense embeddings and learnable recurrent decoder queries. The SDST consists of a dense stream $\\mathcal{D}$ for temporal and multi-modal dense refinement, and a sparse stream $\\mathcal{S}$ that refines the recurrent queries conditioned on the dense signal. $\\mathcal{S}$ includes the RDSA module (right), a context enhanced deformable attention mechanism that restricts the selectable keys based on the center, left-, and right-most action embeddings. The final dense embedding solves HD, while the recurrent queries address MR." + ], + "image_footnote": [], + "bbox": [ + 629, + 103, + 898, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Sparse learning stream $S$", + "text_level": 1, + "bbox": [ + 89, + 344, + 313, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This stream can be seen as a recurrent DETR-based mechanism that exploits the complementarity of both sparse (MR) and dense (HD) tasks to refine the recurrent decoder queries. This is, refining the center-width references $\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}$ and their latent embeddings $\\mathbf{H}^{\\ell}\\in \\mathbb{R}^{M\\times F}$ . Similar to the stream $\\mathcal{D}$ , we first incorporate a CA module that conditions $\\mathbf{H}^{\\ell}$ to the textual query. This is followed by a SA module to enable the information flow between different moment proposals:", + "bbox": [ + 89, + 364, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} ^ {\\ell} = S A \\left(C A \\left(\\mathbf {H} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 503, + 480, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another essential aspect of this stream is effectively conditioning the recurrent decoder queries to the video modality. Existing works [7, 16, 20] typically rely on the use of deformable attention due to its convergence and performance benefits [38, 45]. This mechanism leverages attention queries to predict offsets that define a limited set of selectable keys. However, in this task we observe that 1) these offsets collapse near their initialization and 2) are unable to look beyond the current estimated boundaries (see Sec. 6.2). We attribute this to contextual limitations, which motivates our proposed alternative Reference-based Deformable Self-Attention (RDSA), enabling an enhanced cross-modality injection (see Sec. 3.3.4). We also incorporate a simple PFFN to enhance the model's expressivity:", + "bbox": [ + 88, + 526, + 483, + 737 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} ^ {\\ell} = P F F N \\left(R D S A \\left(\\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\mathbf {D} ^ {\\ell + 1}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 739, + 480, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, all of these modules are shared across the $K$ levels for parameter efficiency, and include residual connections and DropPath [13] for stability and regularization.", + "bbox": [ + 89, + 762, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.4. Reference-based deformable self-attention", + "text_level": 1, + "bbox": [ + 89, + 820, + 429, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deformable attention in the context of existing anchor-free methods: One of the keys to the success of recent anchor-free methods like DETR -including works tackling VTG [7]- is the use of deformable attention mecha", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nisms [45] for cross-modal interaction. This mechanism is known to provide faster convergence speed, reduced time complexity, and as shown in Sec. K, overall better performance than standard CA. As depicted in Fig. 2b (down), the key of this module is the lack of explicit interaction between queries and keys, which are defined as", + "bbox": [ + 511, + 344, + 905, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\mathbf {K} = \\mathbf {X} _ {\\mathcal {K}} \\mathbf {W} _ {\\mathcal {K}} ^ {d e f}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 436, + 903, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathcal{Q}}^{def}$ , $\\mathbf{W}_{\\mathcal{K}}^{def}$ are two linear projections. The deformable attention thus avoids computing the $\\mathbf{Q}\\mathbf{K}^T$ similarity matrix, and instead employs an offset and attention-score predictors, $\\mathcal{G}_{\\Delta}$ and $\\mathcal{G}_{\\mathbf{A}}$ to select and weight -only based on the queries - a small subset $P$ of selectable keys. The output of this mechanism, the weighted aggregation of the selected keys, namely $\\mathbf{S}$ , is computed as follows", + "bbox": [ + 511, + 460, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta = \\mathcal {G} _ {\\Delta} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\mathbf {A} = \\mathcal {G} _ {\\mathbf {A}} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 574, + 903, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\sum_ {p = 1} ^ {P} \\left(\\mathbf {A} _ {:, p} \\mathbf {K} [ \\mathbf {c} + \\mathbf {w} \\odot \\Delta_ {:, p} ]\\right) \\in \\mathbb {R} ^ {M \\times F} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 595, + 903, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_{\\Delta}$ , $\\mathcal{G}_{\\mathbf{A}}$ are often modeled as CNNs and $x[y]$ is the bilinear sampling of $x$ on index(es) $y$ . Notably, in our 2-dimensional setup, the offsets $\\Delta_{:,p}$ are added to center reference $\\mathbf{c} = \\mathbf{R}_{:,0}^{\\ell} \\in \\mathbb{R}^{M}$ , and weighted over the action width $\\mathbf{w} = \\mathbf{R}_{:,1}^{\\ell} \\in \\mathbb{R}^{M}$ . See Sec. K for additional details.", + "bbox": [ + 511, + 642, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations: Interestingly, in the literature we find that when it comes to the deformable attention, most works operate with a certain independence to whether they deal with a deformable self-attention or cross-attention, assuming that the benefits of this mechanism extrapolate to both scenarios. In this work, however, we find that in fact, the deformable attention is only naturally suited for self-attention scenarios as its effectiveness inherently relies on implicit interactions between $\\mathbf{Q}$ and $\\mathbf{K}$ -not present in the cross-attention case. Concretely, in a self-attention setup, $\\mathbf{X}_{\\mathcal{Q}} = \\mathbf{X}_{\\mathcal{K}}$ (see Eq. 7). Thus, modeling $\\mathcal{G}_{\\Delta}$ and $\\mathcal{G}_{\\mathbf{A}}$ as CNNs naturally al", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21518", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lows queries to gain context of the local neighborhood of queries, and consequently, of the key/value space. This breaks their independence assumption, giving the queries critical information to decide where to look.", + "bbox": [ + 88, + 90, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, this is not the case with the deformable cross attention where we try to inject cross-modal information from the keys to the queries, which is our goal in this work. In this scenario, $\\mathbf{X}_{\\mathcal{Q}} \\neq \\mathbf{X}_{\\mathcal{K}}$ , and thus, using CNN-based offset predictors does not inject knowledge of the key/value space, making its guesses blind or based on the blind predictions of previous iterations. Concretely, in Sec. 6.2 we show that in our task, the drawbacks of this naive deformable cross-attention have various empirical implications such as the offset collapse near the offset initialization values, or the inability to select keys from frames beyond the estimated action boundaries, critical to refine segments into longer actions. This motivates our proposed Reference-based Deformable Self-Attention, an alternative mechanism that allows reformulating the deformable CA module into a deformable SA based on the references $\\mathbf{R}^{\\ell}$ .", + "bbox": [ + 93, + 152, + 485, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed alternative (RDSA): Our proposed RDSA (see Fig. 2b) takes as input the center-width references $\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}$ and the dense embeddings $\\mathbf{D}^{\\ell}\\in \\mathbb{R}^{T\\times F}$ . Then, differently from other methods that leverage the learnable queries $\\mathbf{Q} = \\mathbf{H}^{\\ell}$ to compute the offsets and attention scores (see Eq. 8), we propose using an alternative query embedding. Concretely, we first refine the dense embedding $\\mathbf{D}_l$ with a simple CNN, allowing frames to gain local context of their surroundings. Then, we use bilinear sampling to extract three key action embeddings: left-most (l), center (c), and right-most (r). The center provides highly informative action features, while the extremes help refine the moment boundaries. This yields,", + "bbox": [ + 89, + 393, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Q}} = \\hat {\\mathbf {X}} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\hat {\\mathbf {X}} _ {\\mathcal {Q}} = C N N (\\mathbf {D} ^ {\\ell}) [ l, c, r ]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 592, + 480, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}_{\\mathcal{Q}}\\in \\mathbb{R}^{M\\times 3F}$ . This rewrites Eq. 8 to", + "bbox": [ + 89, + 614, + 388, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Delta} = \\mathcal {G} _ {\\Delta} (\\hat {\\mathbf {Q}}), \\hat {\\mathbf {A}} = \\mathcal {G} _ {\\mathbf {A}} (\\hat {\\mathbf {Q}}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 635, + 480, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we apply the deformable attention from Eq. 9 using the new offsets and attention scores $\\hat{\\Delta}$ and $\\hat{\\mathbf{A}}$ . This can be viewed as asking the key spots of the action where to look based on the textual information that we previously injected in Eq. 5. This naturally solves the context limitation of the standard deformable CA given that $\\hat{\\mathbf{X}}_{\\mathcal{Q}}$ and $\\mathbf{X}_{\\mathcal{K}}$ now live in the same latent space -as they are both derived from $\\mathbf{D}^{\\ell}$ . This allows the model to make informed decisions about which keys to attend to while keeping all its core computational advantages.", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Prediction heads and training objectives", + "text_level": 1, + "bbox": [ + 89, + 816, + 434, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After recursively applying the SDST over $K$ intermediate layers, we compute the respective saliency scores and action segment proposals, as well as define their corresponding objective functions. We briefly describe these below,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "but refer interested readers to Sec. C for more details.", + "bbox": [ + 511, + 90, + 867, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Highlight Detection: To solve HD, we define the saliency scores as the cosine similarity between the dense embedding $\\mathbf{D}^K$ and a pooled representation of $\\mathbf{T}^K$ . We then apply an InfoNCE loss [27] that learns a ranking of these scores", + "bbox": [ + 511, + 107, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {H D} = \\lambda_ {0} \\mathcal {L} _ {\\text {I n f o N C E}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 172, + 903, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moment Retrieval: To solve MR we first apply the Hungarian algorithm to obtain a one-to-one matching between the predicted queries and the GT actions. We then project $\\mathbf{H}^{\\ell}$ to compute the action probability $\\mathbf{p} \\in \\mathbb{R}^{M}$ of the $M$ different queries, which we train with a FocalLoss [18]. We also compute the moment boundaries $\\hat{\\mathbf{Y}}^{m} \\in \\mathbb{R}^{M \\times 2}$ which we optimize using an L1 and an IoU loss. Finally, we use an L1 loss to learn an actionness score, which predicts the maximum IoU of each query w.r.t. the GT actions", + "bbox": [ + 511, + 191, + 906, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {M R} = \\lambda_ {1} \\mathcal {L} _ {a c t} + \\sum_ {l = 1} ^ {\\ell} \\lambda_ {2} \\mathcal {L} _ {c l s} ^ {\\ell} + \\lambda_ {3} \\mathcal {L} _ {l 1} ^ {\\ell} + \\lambda_ {4} \\mathcal {L} _ {I o U} ^ {\\ell}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 333, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that all but the $\\mathcal{L}_{act}$ term are optimized across different refinement levels to promote a faster convergence.", + "bbox": [ + 511, + 377, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment losses: To improve the video-textual alignment, we define $\\mathcal{L}_{\\text {align }}$ based on two SampledNCE losses [42] to align $\\mathbf{V}$ and $\\mathbf{T}$ embeddings across the different refinement levels. We apply this along both the batch and the intermediate-layer dimensions. This pulls the action embeddings closer to the textual representations while learning complementary information across intermediate layers.", + "bbox": [ + 511, + 407, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final loss: The final loss is defined as", + "bbox": [ + 511, + 513, + 767, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {5} \\mathcal {L} _ {H D} + \\lambda_ {6} \\mathcal {L} _ {M R} + \\lambda_ {7} \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 535, + 903, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Extracting intermediate representations of InternVideo2", + "text_level": 1, + "bbox": [ + 511, + 561, + 906, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The cornerstone of ST is leveraging intermediate multimodal representations from frozen VLMs. While previous ST approaches predominantly rely on CLIP, we argue that CLIP has notable limitations in temporal modeling and aligning visual features with textual queries beyond simple static descriptions. To overcome these limitations, we incorporate InternVideo2 [36] into our framework, leveraging its advanced spatio-temporal modeling capabilities [7]. However, extracting intermediate representations from InternVideo2 –particularly from its visual encoder $\\mathcal{E}_v$ – remains a challenging and underexplored task in the literature, despite its significant impact on performance.", + "bbox": [ + 511, + 604, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, the main challenge arises from how to pool the spatio-temporal token dimension $L_{v}$ of $\\hat{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times L_v \\times F}$ , the output embedding of the $\\ell$ -th layer of $\\mathcal{E}_v$ into the final intermediate features $\\tilde{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times F}$ (see Eq. 1). Prior CLIP-based methods simply use the CLS token as \"summaries\" over $L_{v}$ . However, we find this strategy to be suboptimal for InternVideo2 (see Sec. 6.1) as this limits its", + "bbox": [ + 511, + 787, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21519", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "spatial-aggregation capabilities. Notably, [36] optimizes an AdaptivePool module that computes the final-layer embeddings, which are then aligned with the text. We conjecture that leveraging this enhanced multi-modal alignment is key to performance. Unfortunately, such pooled representation is only computed at the last layer, leaving the question of how to pool the remaining intermediate layers. Ideally, we would optimize layer-independent AdaptivePool modules, but this would require full back-propagation through the entire model. A more efficient alternative would be computing gradients of the AdaptivePool modules only. This is, however, still computationally infeasible, not for the gradients themselves, but because of the need to load $L_{v}$ spatiotemporal tokens per frame. This increase in the input memory size translates, for instance, to a $15 \\times$ memory increase for QVHighlights. We thus hypothesize that re-using this frozen pooling module across the $K$ intermediate layers allows a memory efficient training –no additional backpropagation nor memory requirements are needed– while leveraging its enhanced pooling capabilities, even despite their respective distribution shifts. Formally, we compute", + "bbox": [ + 89, + 90, + 480, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {V}} ^ {\\ell} = \\text {A d a p t i v e P o o l} \\left(\\hat {\\mathbf {V}} ^ {\\ell}\\right) \\in \\mathbb {R} ^ {T \\times F}, 1 \\leq \\ell \\leq K. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 411, + 480, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Sec. 6.1, this pooling strategy considerably improves other alternatives, demonstrating the benefits of this module re-utilization.", + "bbox": [ + 89, + 439, + 482, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experimentation", + "text_level": 1, + "bbox": [ + 89, + 492, + 256, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 517, + 277, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We test our proposal on various datasets for MR and HD, namely QVHighlights [15], TACoS [32], and the Charades-STA [5] dataset. Further details of the datasets can be found in Sec. B. For the task of MR on QVHighlights, we compute Recall@1 with two IoU thresholds, 0.5 and 0.7, and the mean average precision (mAP) at IoU thresholds between 0.5 and 0.95 with a step of 0.05 -i.e., [0.5:0.05:0.95]. For HD, we report mAP and HIT@1 over the positive frames -i.e., the most salient ones classified as \"VeryGood\". For TACoS and Charades-STA, we compute Recall@1 at 0.3, 0.5, and 0.7 IoU thresholds, as well as mIoU.", + "bbox": [ + 89, + 539, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main experimental results", + "text_level": 1, + "bbox": [ + 89, + 712, + 328, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by comparing the performance of SDST over various relevant baselines. Concretely, observe in Tab. 1 the evaluation of our method on the test and val splits of QVHighlights. These results indicate that our method considerably outperforms $\\mathbb{R}^2$ -Tuning. Similarly, SDST performs very competitively and even surpasses on several metrics the current SOTA method, namely SG-DETR [7]. Note however that our method attains this performance while having only a $27\\%$ of the parameter count of SG-DETR. We refer interested readers to Sec. M for the statistical significance analysis.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Most of the presented methods in Tab. 1 rely on alternative backbones different from InternVideo2 -e.g., CLIP or Slowfast. To establish a more fair comparison, in Fig. 1 we follow the work of [7] and evaluate a set of relevant baselines when using InternVideo2 features -find the complete ablation in Sec. E. Observe that our work performs on par with the SG-DETR and even more importantly to our work, the SDST very considerably outperforms $\\mathbb{R}^2$ -Tuning, improving the MR performance by a $3.82\\%$ average mAP or the HD by a $2.21\\%$ mAP. This empirically demonstrates the benefits of leveraging a proposal-free architecture over a more rigid anchor-based method. Also find in Sec. H an extended comparison with other ST works.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To complement our findings, in Tab. 2 we evaluate our method on the Charades-STA and TACoS benchmarks. Observe that our method obtains SOTA in both of these datasets. In Charades-STA, for instance, improving by $2.71\\%$ of R1@0.7 and $2.06\\%$ of mIoU, the existing SOTA. Similarly, in TACoS, SDST improves SG-DETR by $2.39\\%$ R1@0.7 and $1.27\\%$ mIoU.", + "bbox": [ + 511, + 287, + 903, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 404, + 671, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we ablate over various relevant aspects of our main contributions. Note that unless states otherwise, we follow the literature and evaluate them on the val split of QVHighlights. Also, we write write in bold the best results.", + "bbox": [ + 511, + 430, + 906, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Leveraging InternVideo2 features for ST", + "text_level": 1, + "bbox": [ + 511, + 500, + 861, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) Does pooling matter? As argued in Sec. 4, one of the main challenges in using the InternVideo2 backbone for ST is pooling the spatio-temporal clip-wise intermediate representations. In Tab. 3 we compare different pooling strategies. Concretely, we compare the standard CLS-pooling, average-pooling, and our proposed re-utilization of the frozen AdaptivePool (Sec. 4). Observe that CLS-pooling results in a considerable performance degradation of up to $5.07\\%$ average mAP on MR or $7.74\\%$ HIT@1 on HD w.r.t. the adaptive pooling strategy. The average pooling partially mitigates this degradation, but still obtains a considerably decreased performance. This confirms the significant impact of a carefully chosen pooling strategy.", + "bbox": [ + 511, + 522, + 903, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2) Feature refinement vs. feature sampling: One of the most critical decisions of ST is determining the number of intermediate levels that should be used. Normally, doing so involves evaluating the performance when refining the last $K$ intermediate layers [21]. As shown in Fig. 3 (red) this indicates that in our case, the best performing $K$ is 4. The literature often suggests this argument to be compelling enough to claim that using intermediate representations is beneficial over using only last-layer features. In this work, however, we find it reasonable to wonder: Does this analysis depict the importance of intermediate features, or is it simply evidencing the need to do multiple refinement steps,", + "bbox": [ + 511, + 719, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21520", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efee25b1a87dbe08762874ef64ea14de9764ad68561a9a3d44e19a64c32c6ca5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodtestval#Params
MRHDMRHD
R1@0.5@0.7mAP@0.75Avg.≥ Very good mAP HIT@1R1@0.5@0.7@0.5mAP@0.75Avg.≥ Very good mAP HIT@1
BeautyThumb----14.3620.88--------
DVSE----18.7521.79--------
MCN11.412.7124.948.2210.67---------
CAL25.4911.5423.407.659.89---------
XML+46.6933.4647.8934.6734.9035.3855.06-------
Moment-DETR52.8933.0254.8229.4030.7335.6955.6053.9434.84-32.2035.6535.654.8M
UMT56.2341.1853.8337.0136.1238.1859.9960.2644.2656.7039.9038.5939.9064.20
MomentDiff58.2141.4854.5737.2136.84---------
QD-DETR62.4044.9862.5239.8839.8638.9462.4062.6846.6662.2341.8241.2239.1363.03
MH-DETR60.0542.4860.7538.1338.3838.2260.5160.8444.9060.7639.6439.2638.7761.74
UniVTG58.8640.8657.6035.5935.4738.2060.9659.74---36.1338.8061.8
TR-DETR64.6648.9663.9843.7342.6239.9163.4267.1051.4866.2746.4245.09--
CG-DETR65.4348.3864.5142.7742.8640.3366.2167.3552.0665.5745.7344.9340.8066.70
BAM-DETR62.7148.6464.5746.3345.36--65.1051.6165.4148.5647.61--
EaTR------61.3645.7961.8641.9141.7437.1558.659.0 M
Mr. BLIP74.7760.5168.1253.38--76.1363.3569.3955.78---19.0 M
LLaVA-MR76.5961.4869.4154.40--78.1364.1369.6456.32---17.0 M
HL-CLIP-----41.9470.60----42.3772.402.0 M
R²-Tuning68.0349.3569.0447.5646.1740.7564.2068.7152.06--47.5940.5964.32
SG-DETR†72.2056.6073.2055.8054.1043.7669.13--73.5257.9155.6443.9171.47
Flash-VTG†70.6953.9672.3353.8552.00--73.1057.2972.7554.3352.84--
Ours†70.8256.2371.3154.9953.3143.4069.1373.6860.9073.5257.4255.6044.0072.00
", + "bbox": [ + 140, + 88, + 857, + 362 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6fc235b80c13d679b09c7d43a688cf8f481c19a71758e87388e434f7ea7a0db0.jpg", + "table_caption": [ + "Table 1. MR and HD results for QVHighlights test and val split.† indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best, considering only works supporting both MR and HD. The rest are marked in gray." + ], + "table_footnote": [], + "table_body": "
MethodCharades-STTACoS
R@0.5R@0.7mIoUR@0.5R@0.7mIoU
M-DETR53.631.4-24.712.025.5
UMT48.329.3----
UniVTG58.035.750.135.017.433.6
QD-DETR57.332.6-36.821.135.8
CG-DETR58.436.350.139.622.236.5
BAM-DETR60.039.452.341.526.839.3
TR-DETR57.633.5----
MR_BLIP69.349.358.6---
LLaVA-MR70.649.659.8---
R2-Tuning59.837.050.938.725.135.9
SG-DETR†70.249.559.144.729.940.9
FlashVTG†70.349.9-41.824.737.6
Ours†72.052.661.244.532.342.2
", + "bbox": [ + 142, + 401, + 434, + 551 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/059358abaff7add68b3019bda699ae6a1dda0ef836fd988a840b98c0287bc867.jpg", + "table_caption": [ + "Table 2. Comparison on Charades-STA and TACoS datasets. ${}^{ \\dagger }$ indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best." + ], + "table_footnote": [], + "table_body": "
Pool strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
CLS66.4552.6567.9651.3250.5341.0164.26
Avg. pool70.6556.970.4354.5853.4443.0669.68
Adapt. pool73.6860.9073.5257.4255.6044.0072.00
", + "bbox": [ + 102, + 599, + 470, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "regardless of the chosen features?", + "bbox": [ + 89, + 686, + 316, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To shed light on this issue, in Fig. 3 (blue) we evaluate the performance when doing $K$ different refinement steps, but importantly, we always use the last-layer features only. Observe that for $K = 2$ and $K = 3$ , actually, this improves the homologous experiments with intermediate features. Differently, for $K = 4$ and $K = 5$ , the use of last-layer features harms the performance. These experiments thus suggest that proving the usefulness of intermediate features is not as straightforward as previously thought. In fact, these indicate that the advantages of using intermediate features arise only as we consider shallower layers.", + "bbox": [ + 88, + 703, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3) Why not sampling features from even shallower layers? As shown in Tab. 4, this is not necessarily helpful", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e3c9c98f97f220cfc053f685d964636bdbe6c1c2850c327052a92b3e5ff9093.jpg", + "image_caption": [ + "Figure 3. Ablation of the number of refinement levels and of their use of intermediate or last-layer features only." + ], + "image_footnote": [], + "bbox": [ + 514, + 401, + 709, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/618e30d98c0ccde30948435dac14d414063400100c51eea9f2de997e93d982d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 400, + 906, + 492 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/44aec9307beb82307efeb65d303ab11b6026c12516eb471c57d6731110f2f18b.jpg", + "table_caption": [ + "Table 3. Effect of using different pooling strategies." + ], + "table_footnote": [], + "table_body": "
LevelsMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
[37,38,39,40]73.6860.9073.5257.4255.6044.072.00
[25,30,35,40]71.4859.4871.4556.3754.8343.4270.19
[10,20,30,40]69.0355.3569.5653.3752.4742.7368.39
", + "bbox": [ + 516, + 526, + 908, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of different sampling strategies of intermediate features. The first column indicates the $K = 4$ sampled layers –of a total of 40 layers.", + "bbox": [ + 511, + 584, + 906, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "because of what we call, depth-pooling trade-off. We conjecture that sampling from shallower layers does in fact provide additional complementary information. Nevertheless, this inevitably results in a distribution shift w.r.t. the last-layer feature distribution, hindering the effectiveness of the frozen AdaptivePool (see Eq. 15). The infeasibility of retraining this module for computational reasons, thus, leaves us with a delicate trade-off between the quality of the pooling and the depth of the features that must be accounted for.", + "bbox": [ + 511, + 637, + 906, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Study of deformable attention", + "text_level": 1, + "bbox": [ + 511, + 786, + 779, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1) Comparison w.r.t. other baselines: Here we empirically evaluate the benefits of our proposed RDSA. Concretely, in Tab. 5 we empirically compare the performance of the standard CA [35] and Def. CA [45], as well as different decoder-query initialization strategies [43] for the latter. Notice the prominent performance degradation de", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2d1dd3ab546a7b48a124d8d76ffe9d83d8ac01f00e93b285bfee3af7db5d358.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Att. strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Stand. CA [35]70.7751.7466.7845.1942.7243.1669.87
Def. CA [45]72.5858.1971.8255.8054.2743.2670.58
Def. CA +PureInit [43]72.1357.8770.6854.8552.9243.1970.32
Def. CA +MixedInit [43]72.2658.7171.9456.4354.9443.1069.86
Ours73.6860.9073.5357.4255.6044.072.00
", + "bbox": [ + 93, + 88, + 485, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb63f808cbe9a3924b16163b2f897ae83feb7f744fcc3010ab9e4416e6eca1df.jpg", + "image_caption": [ + "Figure 4. Average of the weighted offsets across $M$ decoder queries and $N$ batch elements for the $K = 4$ refinement levels. Here head 0 (left) is initialized near the left boundary, and head 1 (right) near the right boundary." + ], + "image_footnote": [], + "bbox": [ + 99, + 193, + 281, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0dc62f156c17fe0ccc319e3ccfc324a86027ac54bfe06751d2ff0dc41b009e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 193, + 468, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rived from the use of the CA, or how the query initialization techniques –far from improving the performance of the Def. CA– in some cases are even harmful. In contrast, our method consistently outperforms all these tested baselines, improving by $2.71\\%$ R1@0.7 or $1.62\\%$ mAP w.r.t. the original Def. CA. We refer to Sec. K for the ablation of the different sampling strategies of RDSA.", + "bbox": [ + 89, + 351, + 483, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2) Where are the offsets pointing to? In Fig. 4 we depict the refinement of the weighted offset of a given query $q$ , defined as $d_{q} = \\sum_{p=1}^{P} A_{q,p} \\Delta_{q,p}$ , across the $K$ different refinement steps. Note that as depicted in Eq. 9, these offsets are relative to the estimated moment widths, with -1 and +1 indicating the estimated left- and right-most boundaries. The Def. CA lacks context awareness, causing its offset predictions to stay near the offset initializations. In contrast, RDSA leverages an enhanced contextualization, thus reducing its bias towards the offset initialization. Notably, RDSA learns to look beyond the current action boundaries (offsets < -1 and > +1) capturing near-boundary information, crucial for refining moment boundaries. This contrasts with [45] which remains constrained within the predicted moment boundaries. We consider this the reason why RDSA is especially useful to localize long actions (see Sec. K).", + "bbox": [ + 89, + 458, + 483, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Conditioning signal for the sparse stream $\\mathcal{S}$", + "text_level": 1, + "bbox": [ + 89, + 710, + 459, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To effectively integrate RDSA, it is imperative to obtain an expressive video-based conditioning signal -i.e., the red arrow in Fig. 2a. As shown in Tab. 6, using the raw video representation $\\mathbf{V}^{\\ell}$ significantly degrades performance, highlighting the need for a richer signal incorporating textual and temporal information. While conditioning on the outputs of the CA or SA modules of the dense stream $\\mathcal{D}$ improves results, these lack non-linearities. Instead, using the non-linear PFFN output provides greater flexibility, leading to superior expressiveness. Notably, this choice also enhances HD performance, underscoring the benefits of task", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/94a582ac2d00b618c3f4362957f8c5718096f46f1dce3a16a92e57c34c6d6c72.jpg", + "table_caption": [ + "Table 5. Comparison across different attention strategies as well as various decoder query initializations." + ], + "table_footnote": [], + "table_body": "
Type of featuresMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Raw. rep72.5856.3770.1250.9950.7444.1371.37
Post CA72.2658.3972.2055.9254.8343.6470.65
Post CA+SA72.1358.3272.0056.5754.9443.7870.06
Post CA+SA+FFN73.6860.9073.5357.4255.6044.0072.00
", + "bbox": [ + 517, + 88, + 906, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3ed9da65d866c5d267e2b8d9b396bfb0d11f78812bfd1d128ba7376ebdbeaff7.jpg", + "table_caption": [ + "Table 6. Evaluation of various conditioning signals." + ], + "table_footnote": [], + "table_body": "
CASARDSAFFNMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
✓✓✓71.4257.7471.7356.1854.4743.5371.35
✓✓✓71.8758.7172.1355.9954.3243.2869.55
✓✓✓59.6830.7159.0224.629.0943.5870.39
✓✓✓72.8459.2973.1356.9155.7243.8572.58
✓✓✓✓73.6860.9073.5257.4255.6044.072.00
", + "bbox": [ + 517, + 181, + 906, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Importance of the modules of the sparse stream $S$ .", + "bbox": [ + 532, + 257, + 883, + 270 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "interaction between MR and HD.", + "bbox": [ + 511, + 284, + 733, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Effect of each module of the sparse stream $\\mathcal{S}$", + "text_level": 1, + "bbox": [ + 511, + 315, + 888, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this ablation, we empirically evaluate the contribution of each of the components of $S$ . Concretely, we evaluate the influence of the CA, the SA, the RDSA, and the final PFFN. Observe in Tab. 7 that the initial textual injection CA seems to be particularly relevant for the R1 metric, while the reasoning SA module is important for the mAP. Moreover, as expected, the RDSA module is critical. This is to be expected, as without it, the queries become oblivious to the input video, leaving no other option but that of randomly guessing. Again, the additional expressivity that the final PFFN brings via its non-linearities proves to be beneficial, despite affecting less than the other modules.", + "bbox": [ + 511, + 339, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 537, + 638, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced the Sparse-Dense Side-Tuner, the first proposal-free side-tuning method for VTG. Our dual-stream architecture jointly optimizes dense embeddings and recurrent decoder queries, carefully adapting to the sparse and dense nature of MR and HD. We also identify an inherent context limitation of the Deformable CA-a key component of proposal-free architectures like ourswhich motivates our proposed alternative, the Referencebased Deformable Self-Attention. Finally, we provide the first effective integration of InternVideo2 to a side-tuning framework, leading to a substantial performance boost. We evaluate our method on QVHighlights, TACoS, and Charades-STA, surpassing previous ST works and attaining either SOTA or near-SOTA, all while maintaining a minimal parameter count and a low memory overhead. We leave as future work studying the often overlooked problem of the cross-domain generalization capabilities of these methods.", + "bbox": [ + 511, + 566, + 906, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments: This work has been partially supported by the Spanish project PID2022-136436NB-I00 and by ICREA under the ICREA Academia programme.", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21522", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 1, 2", + "[2] Zhuo Cao, Bingqing Zhang, Heming Du, Xin Yu, Xue Li, and Sen Wang. Flashvtg: Feature layering and adaptive score handling network for video temporal grounding. arXiv preprint arXiv:2412.13441, 2024. 1, 2", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2", + "[5] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 6", + "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 2", + "[7] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 1, 2, 3, 4, 5, 6", + "[8] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. arXiv preprint arXiv:2404.01745, 2024. 1, 2", + "[9] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2", + "[10] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 1, 2", + "[11] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2", + "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2", + "[13] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. 4" + ], + "bbox": [ + 93, + 114, + 483, + 886 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 2", + "[15] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 1, 2, 6", + "[16] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 2, 3, 4", + "[17] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2", + "[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[19] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 2", + "[20] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3, 4", + "[21] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. r2-tuning: Efficient image-to-video transfer learning for video temporal grounding. arXiv preprint arXiv:2404.00801, 2024. 1, 2, 3, 6", + "[22] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1", + "[23] Weiheng Lu, Jian Li, An Yu, Ming-Ching Chang, Shengpeng Ji, and Min Xia. Llava-mr: Large language-and- vision assistant for video moment retrieval. arXiv preprint arXiv:2411.14505, 2024. 1", + "[24] Boris Meinardus, Anil Batra, Anna Rohrbach, and Marcus Rohrbach. The surprising effectiveness of multimodal large language models for video moment retrieval. arXiv preprint arXiv:2406.18113, 2024. 1", + "[25] WonJun Moon, Sangeek Hyun, Su Been Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, 2023. 1, 2", + "[26] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 2" + ], + "bbox": [ + 516, + 92, + 905, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21523", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[28] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 2", + "[29] David Pujol-Perich, Albert Clapés, and Sergio Escalera. Sada: Semantic adversarial unsupervised domain adaptation for temporal action localization. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9237–9247. IEEE, 2025. 1", + "[30] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13934-13944, 2023. 2", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[32] Anna Rohrbach, Marcus Rohrbach, Wei Qiu, Annemarie Friedrich, Manfred Pinkal, and Bernt Schiele. Coherent multi-sentence video description with variable level of detail. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 184-195. Springer, 2014. 2, 6", + "[33] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2", + "[34] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35:12991-13005, 2022. 1, 2", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7, 8", + "[36] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2025. 2, 3, 4, 5, 6", + "[37] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2", + "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 3, 4", + "[39] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Heng-" + ], + "bbox": [ + 89, + 92, + 482, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2", + "[40] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 1, 2", + "[41] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623-13633, 2023. 2", + "[42] Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. Understanding negative sampling in graph representation learning. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1666-1676, 2020. 5", + "[43] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 7, 8", + "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 2", + "[45] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3, 4, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 554 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21524", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_model.json b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2bfc1d8eac26fdf3e7f93f644a0d7771b6bfba76 --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.833, + 0.155 + ], + "angle": 0, + "content": "Sparse-Dense Side-Tuner for efficient Video Temporal Grounding" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.784, + 0.234 + ], + "angle": 0, + "content": "David Pujol-Perich, Sergio Escalera, Albert Clapés \nUniversitat de Barcelona and Computer Vision Center, Barcelona, Spain \n{david.pujolperich, sescalera, aclapes}@ub.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.662 + ], + "angle": 0, + "content": "Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full finetuning is often impractical, parameter-efficient fine-tuning – and particularly side-tuning (ST)– has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention –a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a \\(73\\%\\) the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.678, + 0.222, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Recently the field of Video Understanding has gained attention due to its potential in applications like search engines or recommendation systems. A key task in this field is Video Temporal Grounding (VTG), which localizes moments within videos based on textual descriptions. Typically, this involves doing both Moment Retrieval (MR) [2, 5, 7, 11, 15] and Highlight Detection (HD) [1, 8, 10, 33, 40]. More specifically, MR predicts moment boundaries w.r.t. textual queries, while HD offers a more interpretable perspective, predicting frame-level saliency scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the past, existing works focused on either MR or HD, given the lack of datasets that combined both tasks. QVHighlights [15] shifted this paradigm, proposing the first" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.268, + 0.71, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.269, + 0.905, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Figure 1. Comparison of our proposed method and the main MR (left) and HD (right) baselines. These are all evaluated on QVHighlights val split and use InternVideo2-1B features. These results show that our method improves existing side-tuning works and attains SOTA results while significantly reducing the number of trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.642 + ], + "angle": 0, + "content": "dataset suitable for this multi-task setup. Since then, various works have been proposed following anchor-based [2, 17, 21, 22], anchor-free [7, 15, 25] and even LLM-based [23, 24] approaches. All of these, however, incur in a common critical limitation, which is relying on the final-layer features of frozen backbones. This is especially limiting when facing large distribution shifts between the pre-training distribution and that of the downstream task [29]. This issue is particularly pronounced in VTG, where an image-domain backbone [31] is transferred to the video domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A typical solution is fine-tuning; however, full fine-tuning is impractical due to its high computational cost. Doing this more efficiently has thus become critical, motivating the interest in parameter-efficient fine-tuning (PEFT) [9] methods -e.g., Prompting or Adapters- which optimize only a small subset of parameters. Unfortunately, these still require full back-propagation through the backbone, making it still memory-intensive. To address this, [34] introduced side-tuning (ST), an effective PEFT, and memory-efficient fine-tuning (MEFT) method. ST creates a parallel pathway to refine intermediate features, thus backpropagating over a minimal set of parameters only. In this regard, we highlight the work of \\(\\mathbb{R}^2\\)-Tuning [21], the first ST approach for VTG, which recursively fuses multi-modal intermediate CLIP [31] embeddings. [21] also proposes adapting these frame-level embeddings for MR by generating a dense set of anchors. As shown in Fig. 1, this" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21515" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "method proves ineffective for MR - a highly sparse task where videos may contain very few ground-truth action." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.124, + 0.483, + 0.484 + ], + "angle": 0, + "content": "This motivates our dual-stream Sparse-Dense Side-Tuner (SDST), the first proposal-free ST method, carefully designed for multi-task learning across sparse (MR) and dense (HD) tasks. For this, SDST jointly refines frame-level embeddings -suitable for HD- while learning what we call the recurrent decoder queries for MR. We also identify an implicit contextual limitation of the deformable attention [45] module of anchor-free architectures like ours, which in our task results in collapsing offsets around their initialization values. This also impedes the selection of keys outside of the current estimated moment boundaries, critical to potentially refine the boundaries of longer moments. Consequently, we propose the alternative Reference-based Deformable Self-Attention (RDSA), which naturally addresses this issue by reformulating the CA into a SA-based mechanism. Finally, we tackle the performance degradation of ST methods derived from the use of image-based CLIP, over more advanced spatio-temporal VLMs [36], as noted in [7]. A key challenge of exploiting [36] is, however, defining an effective token pooling strategy, as naive strategies like CLS pooling yield significant performance drops. This motivates our proposed module re-utilization scheme, which results in the first successful integration of InternVideo2 [36] into an ST framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.758 + ], + "angle": 0, + "content": "In summary, our main contributions are threefold: 1) We propose SDST, the first anchor-free ST architecture, specifically tailored for complex sparse-dense multi-task setups like VTG. Our method significantly outperforms existing ST architectures on QVHighlights [15], TACoS [32] and Charades-STA [5] (see Fig. 1). SDST also performs competitively and even surpasses existing SOTA while reducing its learnable parameter count by up to \\(73\\%\\), and incurring in a minimal memory overhead. 2) We identify a key contextual limitation of the deformable attention mechanism that, in our task, results in collapsing offsets. Consequently, we propose RDSA, which tackles this issue by allowing more complex key-selection strategies. 3) We address the limitations of transfer learning from an image-domain backbone to a video domain by integrating, for the first time, the more advanced spatio-temporal backbone [36] into an ST framework. This proves to be a non-trivial challenge with massive performance implications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.228, + 0.785 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Video temporal grounding (VTG): One relevant task in Video Understanding is VTG [17] which aims to identify actions from text queries. Traditionally, this task was approached either from a MR [2, 5, 7, 11, 15, 25, 26] or HD [1, 8, 10, 33, 40] perspective. MR focuses on predicting specific action proposals while HD aims to compute the saliency scores for each of the frames w.r.t. the query. Nev" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.591 + ], + "angle": 0, + "content": "ertheless, after the proposal of QVHighlights [15], the literature started approaching this from a multi-task learning perspective. In this regard, arguably the most notable family of methods is that based on DETR [7, 14, 15, 25, 26, 39] given the sparse nature of MR. These methods normally fuse both video and textual modality into a final embedding which is used for HD and as a memory of a Transformer decoder that refines a set of learnable queries. More in detail, Moment-DETR [15] constitutes the first baseline, which uses a standard Transformer encoder to process both modalities simultaneously. QD-DETR [26] proposes instead to inject the textual information to the video modality with cross attention (CA), followed by a temporal modeling module. CG-DETR [25] reduces the negative impact of irrelevant textual tokens by adding an Adaptive CA module that leverages dummy tokens to redirect attention weight. In this same line, SG-DETR [7] introduces a saliency-guided CA, weighting the contribution of irrelevant textual tokens based on the computed saliency scores. Despite the notable success of these proposal-free methods, one core limitation that they all share is that they rely solely on the last-layer features of a frozen CLIP [31] and/or Slowfast [4] backbones. [41] overcomes this limitation via the full finetuning of CLIP, even though this becomes nearly intractable given the resources that this requires. \\(\\mathbb{R}^2\\)-Tuning [21], the most similar method to our work, overcomes this limitation by proposing the first PEFT and MEFT method for VTG, which recursively applies a multi-modal fusion ST. In this work, we argue that one of its core drawbacks, however, is its anchor-based nature, remaining oblivious to the shown benefits of DETR-based architectures in very sparse tasks like MR [7, 16, 20]. This motivates our proposed SDST, which incorporates the best of both PEFT and MEFT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Parameter-and-memory-efficient fine-tuning: Foundation models and VLMs [31, 37] have become the cornerstone of recent advances in Video Understanding applications. The use of pre-extracted features remains, however, an important limitation to its application on downstream tasks like VTG [17]. Fine-tuning these huge models is often simply infeasible given the large resources that these require. This has motivated the rise of PEFT [9], which aims to democratize the fine-tuning process by restricting the tunable parameters to the bare minimum. In this regard, some of the most prominent approaches are Prompting [12, 44] and the use of Adapters [6, 28]. The first keeps the backbone frozen while learning prompts that are appended to the input, bridging the gap between the backbone's expected distribution and that of the downstream task. Adapters, in contrast, incorporate small trainable modules in the backbone while keeping the rest unchanged. These methods, however, still require full backpropagation through the frozen model, making them memory-inefficient. Recently, the new paradigm of \\(ST\\) [19, 30, 34] –and particularly [21]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "for VTG- has gained attraction creating a parallel pathway to exploit intermediate backbone representations while guaranteeing that backpropagation is only applied to this small parallel module. Interestingly, these methods normally rely on pre-extracted CLIP representations for tasks like HD/MR. As shown by [7], this is a very limiting aspect given the lack of temporal reasoning of CLIP or its difficulties in understanding textual queries beyond simple spatial descriptions. For this reason, in this work, we propose a novel ST architecture that, for the first time, relies on the more advanced InternVideo2 [36] backbone. Unlike CLIP, this backbone is trained on video-domain inputs, resulting in enhanced spatial and temporal modeling capabilities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.288, + 0.485, + 0.576 + ], + "angle": 0, + "content": "Deformable attention: A key component of most SOTA methods for VTG like SG-DETR [7] is their deformable attention module [38, 45]. This module addresses the well-studied slow convergence of DETR [3]. It does so by limiting the selectable keys based on the predictions of the learnable queries, completely decoupling both query and key spaces. Despite desirable in terms of efficiency, it has been previously noted in the literature [43] that the lack of context of the queries w.r.t. the key/value space can lead to suboptimal attendable key selection. VTG methods often mitigate this effect by partially or completely initializing the learnable queries based on the DETR-memory [43] -i.e., frame-level representations in our case. In this work, we empirically demonstrate the limitations of these initialization-based methods for VTG tasks, which motivates our proposed simple yet effective alternative, the Reference-based Deformable Self-Attention (RDSA). This reformulates the deformable CA into a deformable SA, naturally solving the aforementioned issue." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.182, + 0.599 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.271, + 0.623 + ], + "angle": 0, + "content": "3.1. Problem definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.751 + ], + "angle": 0, + "content": "In this paper, we address the problem of MR and HD based on textual descriptions. For this, we consider an arbitrary input-query pair \\((\\mathbf{X}^v,\\mathbf{X}^t)\\) where \\(\\mathbf{X}^v\\in \\mathbb{R}^{T\\times H\\times W\\times 3}\\) and \\(\\mathbf{X}^t\\in \\mathbb{R}^{L\\times F_e}\\). Here \\(T\\) and \\(L\\) are the number of frames and tokens, and \\(F_{e}\\) the textual encoding dimension. Our goal is to predict, for the given video-query pair, its respective frame-wise saliency scores \\(\\mathbf{Y}^s\\in \\mathbb{R}^T\\) for HD, and the \\(M\\) action moments \\(\\mathbf{Y}^m\\in \\mathbb{R}^{M\\times 2}\\) for MR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.202, + 0.773 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our work we propose a dual-stream ST architecture that includes a dense (frame-level) and a sparse (segment-level) stream for MR and HD respectively (see Fig. 2a). More specifically, we first leverage a frozen InternVideo2-1B [36] backbone, a model that possesses powerful spatio-temporal modeling capabilities, to extract \\(K\\) intermediate visual-textual representations. These are then processed by multiple weight-sharing SDST layers, which refine dense rep" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "resentations and recurrent decoder queries. More in detail, at each level, the model refines the dense embeddings via multi-modal textual conditioning as well as temporal modeling. This signal is then used to condition the sparse stream \\( S \\), that builds on [16] while incorporating our novel deformable attention mechanism, namely Reference-based Deformable Self-Attention (see Fig. 2b). This mitigates the shortcomings of the deformable CA mechanism, enhancing the contextual information of the decoder queries and thus improving the quality of the selected keys. After unrolling this process \\( K \\) times, we apply the dense and the sparse prediction heads to compute the saliency scores and the predicted segment boundaries, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.806, + 0.323 + ], + "angle": 0, + "content": "3.3. Sparse-Dense Side-Tuner (SDST)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.332, + 0.826, + 0.346 + ], + "angle": 0, + "content": "3.3.1. Recurrent refinement with side-tuners" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Our work builds on an ST framework, which first requires extracting \\(K\\) intermediate video and textual features \\(\\tilde{\\mathbf{V}}\\in\\) \\(\\mathbb{R}^{K\\times T\\times F_v}\\) and \\(\\tilde{\\mathbf{T}}\\in \\mathbb{R}^{K\\times L\\times F_t}\\) (find the details in Sec. 4). Here \\(F_{v}\\) and \\(F_{t}\\) are their respective dimensionalities. We then define a zero-initialized dense embedding \\(\\mathbf{D}^0\\in \\mathbb{R}^{T\\times F}\\) and a set of \\(M\\) learnable recurrent moment proposals inspired by [20], which we refer to as recurrent decoder queries. The queries include the learnable center-width moment references \\(\\mathbf{R}^0\\in \\mathbb{R}^{M\\times 2}\\), and their corresponding latent embeddings \\(\\mathbf{H}^0\\in \\mathbb{R}^{M\\times F}\\). We then define the recurrence for a given level \\(1\\leq \\ell \\leq K\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.522, + 0.905, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {\\ell + 1}, \\mathbf {R} ^ {\\ell + 1}, \\mathbf {H} ^ {\\ell + 1} = S D S T \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\tilde {\\mathbf {V}} ^ {\\ell}, \\tilde {\\mathbf {T}} ^ {\\ell}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.605 + ], + "angle": 0, + "content": "where SDST represents the shared-weight layers of our model, progressively refining the representations across two streams: the 1) dense learning stream \\(\\mathcal{D}\\) and the 2) sparse learning stream \\(S\\), that we describe below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.734, + 0.638 + ], + "angle": 0, + "content": "3.3.2. Dense learning stream \\(\\mathcal{D}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.72 + ], + "angle": 0, + "content": "Following [21], this stream first refines the dense embeddings \\(\\mathbf{D}^{\\ell}\\) by conditioning it to the video-textual embeddings \\(\\tilde{\\mathbf{V}}^{\\ell}\\) and \\(\\tilde{\\mathbf{T}}^{\\ell}\\). For this, both of these multi-modal embeddings are projected to a shared \\(F\\)-dimensional space, using two MLPs \\((\\mathcal{F}_v:F_v\\to F\\) and \\(\\mathcal{F}_t:F_t\\to F)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.723, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} ^ {\\ell} = \\mathcal {F} _ {v} \\left(\\tilde {\\mathbf {V}} ^ {\\ell}\\right), \\mathbf {T} ^ {\\ell} = \\mathcal {F} _ {t} \\left(\\tilde {\\mathbf {T}} ^ {\\ell}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Then we incorporate the clip-wise visual information to \\(\\mathbf{D}^{\\ell}\\) via a weighted sum modulated by \\(\\beta^{\\ell} \\in [0,1]\\), a zero-initialized layer-dependent parameter that yields:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.794, + 0.905, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {\\ell} := \\beta^ {\\ell} \\mathbf {D} ^ {\\ell} + (1 - \\beta^ {\\ell}) \\mathbf {V} ^ {\\ell}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.816, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Thereafter, we inject the textual information with CA, where \\(\\mathbf{T}^{\\ell}\\) does not include the CLS token. We also apply a temporal modeling module defined as a SA and a subsequent point-wise feed forward network (PFFN):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.879, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {\\ell + 1} = P F F N \\left(S A \\left(C A \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.605, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.237, + 0.462, + 0.248 + ], + "angle": 0, + "content": "(a) Overview of our proposed architecture." + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.104, + 0.9, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.237, + 0.881, + 0.248 + ], + "angle": 0, + "content": "(b) RDSA vs Def.CA[38] for a proposal \\( m \\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Our method (left) first processes the video and textual inputs using [36], and then recursively applies a shared SDST, a recurrent dual-stream model attached to the last \\(K\\) layers. This iteratively refines dense embeddings and learnable recurrent decoder queries. The SDST consists of a dense stream \\(\\mathcal{D}\\) for temporal and multi-modal dense refinement, and a sparse stream \\(\\mathcal{S}\\) that refines the recurrent queries conditioned on the dense signal. \\(\\mathcal{S}\\) includes the RDSA module (right), a context enhanced deformable attention mechanism that restricts the selectable keys based on the center, left-, and right-most action embeddings. The final dense embedding solves HD, while the recurrent queries address MR." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.345, + 0.314, + 0.361 + ], + "angle": 0, + "content": "3.3.3. Sparse learning stream \\(S\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.501 + ], + "angle": 0, + "content": "This stream can be seen as a recurrent DETR-based mechanism that exploits the complementarity of both sparse (MR) and dense (HD) tasks to refine the recurrent decoder queries. This is, refining the center-width references \\(\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}\\) and their latent embeddings \\(\\mathbf{H}^{\\ell}\\in \\mathbb{R}^{M\\times F}\\). Similar to the stream \\(\\mathcal{D}\\), we first incorporate a CA module that conditions \\(\\mathbf{H}^{\\ell}\\) to the textual query. This is followed by a SA module to enable the information flow between different moment proposals:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.505, + 0.482, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} ^ {\\ell} = S A \\left(C A \\left(\\mathbf {H} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.527, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Another essential aspect of this stream is effectively conditioning the recurrent decoder queries to the video modality. Existing works [7, 16, 20] typically rely on the use of deformable attention due to its convergence and performance benefits [38, 45]. This mechanism leverages attention queries to predict offsets that define a limited set of selectable keys. However, in this task we observe that 1) these offsets collapse near their initialization and 2) are unable to look beyond the current estimated boundaries (see Sec. 6.2). We attribute this to contextual limitations, which motivates our proposed alternative Reference-based Deformable Self-Attention (RDSA), enabling an enhanced cross-modality injection (see Sec. 3.3.4). We also incorporate a simple PFFN to enhance the model's expressivity:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.741, + 0.482, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} ^ {\\ell} = P F F N \\left(R D S A \\left(\\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\mathbf {D} ^ {\\ell + 1}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Importantly, all of these modules are shared across the \\(K\\) levels for parameter efficiency, and include residual connections and DropPath [13] for stability and regularization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.431, + 0.835 + ], + "angle": 0, + "content": "3.3.4. Reference-based deformable self-attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Deformable attention in the context of existing anchor-free methods: One of the keys to the success of recent anchor-free methods like DETR -including works tackling VTG [7]- is the use of deformable attention mecha" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.906, + 0.435 + ], + "angle": 0, + "content": "nisms [45] for cross-modal interaction. This mechanism is known to provide faster convergence speed, reduced time complexity, and as shown in Sec. K, overall better performance than standard CA. As depicted in Fig. 2b (down), the key of this module is the lack of explicit interaction between queries and keys, which are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.438, + 0.905, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\mathbf {K} = \\mathbf {X} _ {\\mathcal {K}} \\mathbf {W} _ {\\mathcal {K}} ^ {d e f}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.572 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathcal{Q}}^{def}\\), \\(\\mathbf{W}_{\\mathcal{K}}^{def}\\) are two linear projections. The deformable attention thus avoids computing the \\(\\mathbf{Q}\\mathbf{K}^T\\) similarity matrix, and instead employs an offset and attention-score predictors, \\(\\mathcal{G}_{\\Delta}\\) and \\(\\mathcal{G}_{\\mathbf{A}}\\) to select and weight -only based on the queries - a small subset \\(P\\) of selectable keys. The output of this mechanism, the weighted aggregation of the selected keys, namely \\(\\mathbf{S}\\), is computed as follows" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.575, + 0.905, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\Delta = \\mathcal {G} _ {\\Delta} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\mathbf {A} = \\mathcal {G} _ {\\mathbf {A}} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.597, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\sum_ {p = 1} ^ {P} \\left(\\mathbf {A} _ {:, p} \\mathbf {K} [ \\mathbf {c} + \\mathbf {w} \\odot \\Delta_ {:, p} ]\\right) \\in \\mathbb {R} ^ {M \\times F} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.721 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_{\\Delta}\\), \\(\\mathcal{G}_{\\mathbf{A}}\\) are often modeled as CNNs and \\(x[y]\\) is the bilinear sampling of \\(x\\) on index(es) \\(y\\). Notably, in our 2-dimensional setup, the offsets \\(\\Delta_{:,p}\\) are added to center reference \\(\\mathbf{c} = \\mathbf{R}_{:,0}^{\\ell} \\in \\mathbb{R}^{M}\\), and weighted over the action width \\(\\mathbf{w} = \\mathbf{R}_{:,1}^{\\ell} \\in \\mathbb{R}^{M}\\). See Sec. K for additional details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Limitations: Interestingly, in the literature we find that when it comes to the deformable attention, most works operate with a certain independence to whether they deal with a deformable self-attention or cross-attention, assuming that the benefits of this mechanism extrapolate to both scenarios. In this work, however, we find that in fact, the deformable attention is only naturally suited for self-attention scenarios as its effectiveness inherently relies on implicit interactions between \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) -not present in the cross-attention case. Concretely, in a self-attention setup, \\(\\mathbf{X}_{\\mathcal{Q}} = \\mathbf{X}_{\\mathcal{K}}\\) (see Eq. 7). Thus, modeling \\(\\mathcal{G}_{\\Delta}\\) and \\(\\mathcal{G}_{\\mathbf{A}}\\) as CNNs naturally al" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21518" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.151 + ], + "angle": 0, + "content": "lows queries to gain context of the local neighborhood of queries, and consequently, of the key/value space. This breaks their independence assumption, giving the queries critical information to decide where to look." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.153, + 0.486, + 0.394 + ], + "angle": 0, + "content": "Importantly, this is not the case with the deformable cross attention where we try to inject cross-modal information from the keys to the queries, which is our goal in this work. In this scenario, \\(\\mathbf{X}_{\\mathcal{Q}} \\neq \\mathbf{X}_{\\mathcal{K}}\\), and thus, using CNN-based offset predictors does not inject knowledge of the key/value space, making its guesses blind or based on the blind predictions of previous iterations. Concretely, in Sec. 6.2 we show that in our task, the drawbacks of this naive deformable cross-attention have various empirical implications such as the offset collapse near the offset initialization values, or the inability to select keys from frames beyond the estimated action boundaries, critical to refine segments into longer actions. This motivates our proposed Reference-based Deformable Self-Attention, an alternative mechanism that allows reformulating the deformable CA module into a deformable SA based on the references \\(\\mathbf{R}^{\\ell}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Our proposed alternative (RDSA): Our proposed RDSA (see Fig. 2b) takes as input the center-width references \\(\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}\\) and the dense embeddings \\(\\mathbf{D}^{\\ell}\\in \\mathbb{R}^{T\\times F}\\). Then, differently from other methods that leverage the learnable queries \\(\\mathbf{Q} = \\mathbf{H}^{\\ell}\\) to compute the offsets and attention scores (see Eq. 8), we propose using an alternative query embedding. Concretely, we first refine the dense embedding \\(\\mathbf{D}_l\\) with a simple CNN, allowing frames to gain local context of their surroundings. Then, we use bilinear sampling to extract three key action embeddings: left-most (l), center (c), and right-most (r). The center provides highly informative action features, while the extremes help refine the moment boundaries. This yields," + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.593, + 0.482, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Q}} = \\hat {\\mathbf {X}} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\hat {\\mathbf {X}} _ {\\mathcal {Q}} = C N N (\\mathbf {D} ^ {\\ell}) [ l, c, r ]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.616, + 0.39, + 0.633 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}_{\\mathcal{Q}}\\in \\mathbb{R}^{M\\times 3F}\\). This rewrites Eq. 8 to" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.636, + 0.482, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Delta} = \\mathcal {G} _ {\\Delta} (\\hat {\\mathbf {Q}}), \\hat {\\mathbf {A}} = \\mathcal {G} _ {\\mathbf {A}} (\\hat {\\mathbf {Q}}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Finally, we apply the deformable attention from Eq. 9 using the new offsets and attention scores \\(\\hat{\\Delta}\\) and \\(\\hat{\\mathbf{A}}\\). This can be viewed as asking the key spots of the action where to look based on the textual information that we previously injected in Eq. 5. This naturally solves the context limitation of the standard deformable CA given that \\(\\hat{\\mathbf{X}}_{\\mathcal{Q}}\\) and \\(\\mathbf{X}_{\\mathcal{K}}\\) now live in the same latent space -as they are both derived from \\(\\mathbf{D}^{\\ell}\\). This allows the model to make informed decisions about which keys to attend to while keeping all its core computational advantages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.436, + 0.835 + ], + "angle": 0, + "content": "3.4. Prediction heads and training objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After recursively applying the SDST over \\( K \\) intermediate layers, we compute the respective saliency scores and action segment proposals, as well as define their corresponding objective functions. We briefly describe these below," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.868, + 0.106 + ], + "angle": 0, + "content": "but refer interested readers to Sec. C for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Highlight Detection: To solve HD, we define the saliency scores as the cosine similarity between the dense embedding \\(\\mathbf{D}^K\\) and a pooled representation of \\(\\mathbf{T}^K\\). We then apply an InfoNCE loss [27] that learns a ranking of these scores" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.173, + 0.905, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {H D} = \\lambda_ {0} \\mathcal {L} _ {\\text {I n f o N C E}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Moment Retrieval: To solve MR we first apply the Hungarian algorithm to obtain a one-to-one matching between the predicted queries and the GT actions. We then project \\(\\mathbf{H}^{\\ell}\\) to compute the action probability \\(\\mathbf{p} \\in \\mathbb{R}^{M}\\) of the \\(M\\) different queries, which we train with a FocalLoss [18]. We also compute the moment boundaries \\(\\hat{\\mathbf{Y}}^{m} \\in \\mathbb{R}^{M \\times 2}\\) which we optimize using an L1 and an IoU loss. Finally, we use an L1 loss to learn an actionness score, which predicts the maximum IoU of each query w.r.t. the GT actions" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {M R} = \\lambda_ {1} \\mathcal {L} _ {a c t} + \\sum_ {l = 1} ^ {\\ell} \\lambda_ {2} \\mathcal {L} _ {c l s} ^ {\\ell} + \\lambda_ {3} \\mathcal {L} _ {l 1} ^ {\\ell} + \\lambda_ {4} \\mathcal {L} _ {I o U} ^ {\\ell}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.905, + 0.408 + ], + "angle": 0, + "content": "Note that all but the \\(\\mathcal{L}_{act}\\) term are optimized across different refinement levels to promote a faster convergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Alignment losses: To improve the video-textual alignment, we define \\(\\mathcal{L}_{\\text {align }}\\) based on two SampledNCE losses [42] to align \\(\\mathbf{V}\\) and \\(\\mathbf{T}\\) embeddings across the different refinement levels. We apply this along both the batch and the intermediate-layer dimensions. This pulls the action embeddings closer to the textual representations while learning complementary information across intermediate layers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.768, + 0.529 + ], + "angle": 0, + "content": "Final loss: The final loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.536, + 0.905, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {5} \\mathcal {L} _ {H D} + \\lambda_ {6} \\mathcal {L} _ {M R} + \\lambda_ {7} \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.907, + 0.595 + ], + "angle": 0, + "content": "4. Extracting intermediate representations of InternVideo2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.788 + ], + "angle": 0, + "content": "The cornerstone of ST is leveraging intermediate multimodal representations from frozen VLMs. While previous ST approaches predominantly rely on CLIP, we argue that CLIP has notable limitations in temporal modeling and aligning visual features with textual queries beyond simple static descriptions. To overcome these limitations, we incorporate InternVideo2 [36] into our framework, leveraging its advanced spatio-temporal modeling capabilities [7]. However, extracting intermediate representations from InternVideo2 –particularly from its visual encoder \\(\\mathcal{E}_v\\) – remains a challenging and underexplored task in the literature, despite its significant impact on performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In particular, the main challenge arises from how to pool the spatio-temporal token dimension \\(L_{v}\\) of \\(\\hat{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times L_v \\times F}\\), the output embedding of the \\(\\ell\\)-th layer of \\(\\mathcal{E}_v\\) into the final intermediate features \\(\\tilde{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times F}\\) (see Eq. 1). Prior CLIP-based methods simply use the CLS token as \"summaries\" over \\(L_{v}\\). However, we find this strategy to be suboptimal for InternVideo2 (see Sec. 6.1) as this limits its" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21519" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.409 + ], + "angle": 0, + "content": "spatial-aggregation capabilities. Notably, [36] optimizes an AdaptivePool module that computes the final-layer embeddings, which are then aligned with the text. We conjecture that leveraging this enhanced multi-modal alignment is key to performance. Unfortunately, such pooled representation is only computed at the last layer, leaving the question of how to pool the remaining intermediate layers. Ideally, we would optimize layer-independent AdaptivePool modules, but this would require full back-propagation through the entire model. A more efficient alternative would be computing gradients of the AdaptivePool modules only. This is, however, still computationally infeasible, not for the gradients themselves, but because of the need to load \\( L_{v} \\) spatiotemporal tokens per frame. This increase in the input memory size translates, for instance, to a \\( 15 \\times \\) memory increase for QVHighlights. We thus hypothesize that re-using this frozen pooling module across the \\( K \\) intermediate layers allows a memory efficient training –no additional backpropagation nor memory requirements are needed– while leveraging its enhanced pooling capabilities, even despite their respective distribution shifts. Formally, we compute" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.412, + 0.482, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {V}} ^ {\\ell} = \\text {A d a p t i v e P o o l} \\left(\\hat {\\mathbf {V}} ^ {\\ell}\\right) \\in \\mathbb {R} ^ {T \\times F}, 1 \\leq \\ell \\leq K. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.485 + ], + "angle": 0, + "content": "As shown in Sec. 6.1, this pooling strategy considerably improves other alternatives, demonstrating the benefits of this module re-utilization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.493, + 0.258, + 0.51 + ], + "angle": 0, + "content": "5. Experimentation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.518, + 0.279, + 0.535 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.706 + ], + "angle": 0, + "content": "We test our proposal on various datasets for MR and HD, namely QVHighlights [15], TACoS [32], and the Charades-STA [5] dataset. Further details of the datasets can be found in Sec. B. For the task of MR on QVHighlights, we compute Recall@1 with two IoU thresholds, 0.5 and 0.7, and the mean average precision (mAP) at IoU thresholds between 0.5 and 0.95 with a step of 0.05 -i.e., [0.5:0.05:0.95]. For HD, we report mAP and HIT@1 over the positive frames -i.e., the most salient ones classified as \"VeryGood\". For TACoS and Charades-STA, we compute Recall@1 at 0.3, 0.5, and 0.7 IoU thresholds, as well as mIoU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.33, + 0.729 + ], + "angle": 0, + "content": "5.2. Main experimental results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We begin by comparing the performance of SDST over various relevant baselines. Concretely, observe in Tab. 1 the evaluation of our method on the test and val splits of QVHighlights. These results indicate that our method considerably outperforms \\(\\mathbb{R}^2\\)-Tuning. Similarly, SDST performs very competitively and even surpasses on several metrics the current SOTA method, namely SG-DETR [7]. Note however that our method attains this performance while having only a \\(27\\%\\) of the parameter count of SG-DETR. We refer interested readers to Sec. M for the statistical significance analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Most of the presented methods in Tab. 1 rely on alternative backbones different from InternVideo2 -e.g., CLIP or Slowfast. To establish a more fair comparison, in Fig. 1 we follow the work of [7] and evaluate a set of relevant baselines when using InternVideo2 features -find the complete ablation in Sec. E. Observe that our work performs on par with the SG-DETR and even more importantly to our work, the SDST very considerably outperforms \\(\\mathbb{R}^2\\)-Tuning, improving the MR performance by a \\(3.82\\%\\) average mAP or the HD by a \\(2.21\\%\\) mAP. This empirically demonstrates the benefits of leveraging a proposal-free architecture over a more rigid anchor-based method. Also find in Sec. H an extended comparison with other ST works." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.393 + ], + "angle": 0, + "content": "To complement our findings, in Tab. 2 we evaluate our method on the Charades-STA and TACoS benchmarks. Observe that our method obtains SOTA in both of these datasets. In Charades-STA, for instance, improving by \\(2.71\\%\\) of R1@0.7 and \\(2.06\\%\\) of mIoU, the existing SOTA. Similarly, in TACoS, SDST improves SG-DETR by \\(2.39\\%\\) R1@0.7 and \\(1.27\\%\\) mIoU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.673, + 0.421 + ], + "angle": 0, + "content": "6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.492 + ], + "angle": 0, + "content": "In this section, we ablate over various relevant aspects of our main contributions. Note that unless states otherwise, we follow the literature and evaluate them on the val split of QVHighlights. Also, we write write in bold the best results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.862, + 0.517 + ], + "angle": 0, + "content": "6.1. Leveraging InternVideo2 features for ST" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.905, + 0.72 + ], + "angle": 0, + "content": "1) Does pooling matter? As argued in Sec. 4, one of the main challenges in using the InternVideo2 backbone for ST is pooling the spatio-temporal clip-wise intermediate representations. In Tab. 3 we compare different pooling strategies. Concretely, we compare the standard CLS-pooling, average-pooling, and our proposed re-utilization of the frozen AdaptivePool (Sec. 4). Observe that CLS-pooling results in a considerable performance degradation of up to \\(5.07\\%\\) average mAP on MR or \\(7.74\\%\\) HIT@1 on HD w.r.t. the adaptive pooling strategy. The average pooling partially mitigates this degradation, but still obtains a considerably decreased performance. This confirms the significant impact of a carefully chosen pooling strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.902 + ], + "angle": 0, + "content": "2) Feature refinement vs. feature sampling: One of the most critical decisions of ST is determining the number of intermediate levels that should be used. Normally, doing so involves evaluating the performance when refining the last \\( K \\) intermediate layers [21]. As shown in Fig. 3 (red) this indicates that in our case, the best performing \\( K \\) is 4. The literature often suggests this argument to be compelling enough to claim that using intermediate representations is beneficial over using only last-layer features. In this work, however, we find it reasonable to wonder: Does this analysis depict the importance of intermediate features, or is it simply evidencing the need to do multiple refinement steps," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21520" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.141, + 0.089, + 0.858, + 0.363 + ], + "angle": 0, + "content": "
Methodtestval#Params
MRHDMRHD
R1@0.5@0.7mAP@0.75Avg.≥ Very good mAP HIT@1R1@0.5@0.7@0.5mAP@0.75Avg.≥ Very good mAP HIT@1
BeautyThumb----14.3620.88--------
DVSE----18.7521.79--------
MCN11.412.7124.948.2210.67---------
CAL25.4911.5423.407.659.89---------
XML+46.6933.4647.8934.6734.9035.3855.06-------
Moment-DETR52.8933.0254.8229.4030.7335.6955.6053.9434.84-32.2035.6535.654.8M
UMT56.2341.1853.8337.0136.1238.1859.9960.2644.2656.7039.9038.5939.9064.20
MomentDiff58.2141.4854.5737.2136.84---------
QD-DETR62.4044.9862.5239.8839.8638.9462.4062.6846.6662.2341.8241.2239.1363.03
MH-DETR60.0542.4860.7538.1338.3838.2260.5160.8444.9060.7639.6439.2638.7761.74
UniVTG58.8640.8657.6035.5935.4738.2060.9659.74---36.1338.8061.8
TR-DETR64.6648.9663.9843.7342.6239.9163.4267.1051.4866.2746.4245.09--
CG-DETR65.4348.3864.5142.7742.8640.3366.2167.3552.0665.5745.7344.9340.8066.70
BAM-DETR62.7148.6464.5746.3345.36--65.1051.6165.4148.5647.61--
EaTR------61.3645.7961.8641.9141.7437.1558.659.0 M
Mr. BLIP74.7760.5168.1253.38--76.1363.3569.3955.78---19.0 M
LLaVA-MR76.5961.4869.4154.40--78.1364.1369.6456.32---17.0 M
HL-CLIP-----41.9470.60----42.3772.402.0 M
R²-Tuning68.0349.3569.0447.5646.1740.7564.2068.7152.06--47.5940.5964.32
SG-DETR†72.2056.6073.2055.8054.1043.7669.13--73.5257.9155.6443.9171.47
Flash-VTG†70.6953.9672.3353.8552.00--73.1057.2972.7554.3352.84--
Ours†70.8256.2371.3154.9953.3143.4069.1373.6860.9073.5257.4255.6044.0072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.367, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Table 1. MR and HD results for QVHighlights test and val split.† indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best, considering only works supporting both MR and HD. The rest are marked in gray." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.402, + 0.436, + 0.552 + ], + "angle": 0, + "content": "
MethodCharades-STTACoS
R@0.5R@0.7mIoUR@0.5R@0.7mIoU
M-DETR53.631.4-24.712.025.5
UMT48.329.3----
UniVTG58.035.750.135.017.433.6
QD-DETR57.332.6-36.821.135.8
CG-DETR58.436.350.139.622.236.5
BAM-DETR60.039.452.341.526.839.3
TR-DETR57.633.5----
MR_BLIP69.349.358.6---
LLaVA-MR70.649.659.8---
R2-Tuning59.837.050.938.725.135.9
SG-DETR†70.249.559.144.729.940.9
FlashVTG†70.349.9-41.824.737.6
Ours†72.052.661.244.532.342.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Table 2. Comparison on Charades-STA and TACoS datasets. \\( {}^{ \\dagger } \\) indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.6, + 0.472, + 0.659 + ], + "angle": 0, + "content": "
Pool strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
CLS66.4552.6567.9651.3250.5341.0164.26
Avg. pool70.6556.970.4354.5853.4443.0669.68
Adapt. pool73.6860.9073.5257.4255.6044.0072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.661, + 0.441, + 0.675 + ], + "angle": 0, + "content": "Table 3. Effect of using different pooling strategies." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.687, + 0.318, + 0.702 + ], + "angle": 0, + "content": "regardless of the chosen features?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To shed light on this issue, in Fig. 3 (blue) we evaluate the performance when doing \\( K \\) different refinement steps, but importantly, we always use the last-layer features only. Observe that for \\( K = 2 \\) and \\( K = 3 \\), actually, this improves the homologous experiments with intermediate features. Differently, for \\( K = 4 \\) and \\( K = 5 \\), the use of last-layer features harms the performance. These experiments thus suggest that proving the usefulness of intermediate features is not as straightforward as previously thought. In fact, these indicate that the advantages of using intermediate features arise only as we consider shallower layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "3) Why not sampling features from even shallower layers? As shown in Tab. 4, this is not necessarily helpful" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.402, + 0.71, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.401, + 0.907, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.496, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 3. Ablation of the number of refinement levels and of their use of intermediate or last-layer features only." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.527, + 0.909, + 0.584 + ], + "angle": 0, + "content": "
LevelsMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
[37,38,39,40]73.6860.9073.5257.4255.6044.072.00
[25,30,35,40]71.4859.4871.4556.3754.8343.4270.19
[10,20,30,40]69.0355.3569.5653.3752.4742.7368.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.585, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Table 4. Ablation of different sampling strategies of intermediate features. The first column indicates the \\( K = 4 \\) sampled layers –of a total of 40 layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.775 + ], + "angle": 0, + "content": "because of what we call, depth-pooling trade-off. We conjecture that sampling from shallower layers does in fact provide additional complementary information. Nevertheless, this inevitably results in a distribution shift w.r.t. the last-layer feature distribution, hindering the effectiveness of the frozen AdaptivePool (see Eq. 15). The infeasibility of retraining this module for computational reasons, thus, leaves us with a delicate trade-off between the quality of the pooling and the depth of the features that must be accounted for." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.78, + 0.803 + ], + "angle": 0, + "content": "6.2. Study of deformable attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "1) Comparison w.r.t. other baselines: Here we empirically evaluate the benefits of our proposed RDSA. Concretely, in Tab. 5 we empirically compare the performance of the standard CA [35] and Def. CA [45], as well as different decoder-query initialization strategies [43] for the latter. Notice the prominent performance degradation de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21521" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.486, + 0.159 + ], + "angle": 0, + "content": "
Att. strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Stand. CA [35]70.7751.7466.7845.1942.7243.1669.87
Def. CA [45]72.5858.1971.8255.8054.2743.2670.58
Def. CA +PureInit [43]72.1357.8770.6854.8552.9243.1970.32
Def. CA +MixedInit [43]72.2658.7171.9456.4354.9443.1069.86
Ours73.6860.9073.5357.4255.6044.072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.162, + 0.483, + 0.19 + ], + "angle": 0, + "content": "Table 5. Comparison across different attention strategies as well as various decoder query initializations." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.194, + 0.282, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.194, + 0.47, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Average of the weighted offsets across \\( M \\) decoder queries and \\( N \\) batch elements for the \\( K = 4 \\) refinement levels. Here head 0 (left) is initialized near the left boundary, and head 1 (right) near the right boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.457 + ], + "angle": 0, + "content": "rived from the use of the CA, or how the query initialization techniques –far from improving the performance of the Def. CA– in some cases are even harmful. In contrast, our method consistently outperforms all these tested baselines, improving by \\(2.71\\%\\) R1@0.7 or \\(1.62\\%\\) mAP w.r.t. the original Def. CA. We refer to Sec. K for the ablation of the different sampling strategies of RDSA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.701 + ], + "angle": 0, + "content": "2) Where are the offsets pointing to? In Fig. 4 we depict the refinement of the weighted offset of a given query \\( q \\), defined as \\( d_{q} = \\sum_{p=1}^{P} A_{q,p} \\Delta_{q,p} \\), across the \\( K \\) different refinement steps. Note that as depicted in Eq. 9, these offsets are relative to the estimated moment widths, with -1 and +1 indicating the estimated left- and right-most boundaries. The Def. CA lacks context awareness, causing its offset predictions to stay near the offset initializations. In contrast, RDSA leverages an enhanced contextualization, thus reducing its bias towards the offset initialization. Notably, RDSA learns to look beyond the current action boundaries (offsets < -1 and > +1) capturing near-boundary information, crucial for refining moment boundaries. This contrasts with [45] which remains constrained within the predicted moment boundaries. We consider this the reason why RDSA is especially useful to localize long actions (see Sec. K)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.712, + 0.46, + 0.729 + ], + "angle": 0, + "content": "6.3. Conditioning signal for the sparse stream \\(\\mathcal{S}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To effectively integrate RDSA, it is imperative to obtain an expressive video-based conditioning signal -i.e., the red arrow in Fig. 2a. As shown in Tab. 6, using the raw video representation \\(\\mathbf{V}^{\\ell}\\) significantly degrades performance, highlighting the need for a richer signal incorporating textual and temporal information. While conditioning on the outputs of the CA or SA modules of the dense stream \\(\\mathcal{D}\\) improves results, these lack non-linearities. Instead, using the non-linear PFFN output provides greater flexibility, leading to superior expressiveness. Notably, this choice also enhances HD performance, underscoring the benefits of task" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.908, + 0.156 + ], + "angle": 0, + "content": "
Type of featuresMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Raw. rep72.5856.3770.1250.9950.7444.1371.37
Post CA72.2658.3972.2055.9254.8343.6470.65
Post CA+SA72.1358.3272.0056.5754.9443.7870.06
Post CA+SA+FFN73.6860.9073.5357.4255.6044.0072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.556, + 0.161, + 0.862, + 0.175 + ], + "angle": 0, + "content": "Table 6. Evaluation of various conditioning signals." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.182, + 0.908, + 0.253 + ], + "angle": 0, + "content": "
CASARDSAFFNMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
✓✓✓71.4257.7471.7356.1854.4743.5371.35
✓✓✓71.8758.7172.1355.9954.3243.2869.55
✓✓✓59.6830.7159.0224.629.0943.5870.39
✓✓✓72.8459.2973.1356.9155.7243.8572.58
✓✓✓✓73.6860.9073.5257.4255.6044.072.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.258, + 0.885, + 0.271 + ], + "angle": 0, + "content": "Table 7. Importance of the modules of the sparse stream \\( S \\) ." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.285, + 0.735, + 0.299 + ], + "angle": 0, + "content": "interaction between MR and HD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.316, + 0.89, + 0.333 + ], + "angle": 0, + "content": "6.4. Effect of each module of the sparse stream \\(\\mathcal{S}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.522 + ], + "angle": 0, + "content": "In this ablation, we empirically evaluate the contribution of each of the components of \\( S \\). Concretely, we evaluate the influence of the CA, the SA, the RDSA, and the final PFFN. Observe in Tab. 7 that the initial textual injection CA seems to be particularly relevant for the R1 metric, while the reasoning SA module is important for the mAP. Moreover, as expected, the RDSA module is critical. This is to be expected, as without it, the queries become oblivious to the input video, leaving no other option but that of randomly guessing. Again, the additional expressivity that the final PFFN brings via its non-linearities proves to be beneficial, despite affecting less than the other modules." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.539, + 0.64, + 0.555 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.838 + ], + "angle": 0, + "content": "In this work, we introduced the Sparse-Dense Side-Tuner, the first proposal-free side-tuning method for VTG. Our dual-stream architecture jointly optimizes dense embeddings and recurrent decoder queries, carefully adapting to the sparse and dense nature of MR and HD. We also identify an inherent context limitation of the Deformable CA-a key component of proposal-free architectures like ourswhich motivates our proposed alternative, the Referencebased Deformable Self-Attention. Finally, we provide the first effective integration of InternVideo2 to a side-tuning framework, leading to a substantial performance boost. We evaluate our method on QVHighlights, TACoS, and Charades-STA, surpassing previous ST works and attaining either SOTA or near-SOTA, all while maintaining a minimal parameter count and a low memory overhead. We leave as future work studying the often overlooked problem of the cross-domain generalization capabilities of these methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments: This work has been partially supported by the Spanish project PID2022-136436NB-I00 and by ICREA under the ICREA Academia programme." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21522" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.484, + 0.238 + ], + "angle": 0, + "content": "[2] Zhuo Cao, Bingqing Zhang, Heming Du, Xin Yu, Xue Li, and Sen Wang. Flashvtg: Feature layering and adaptive score handling network for video temporal grounding. arXiv preprint arXiv:2412.13441, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[7] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[8] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. arXiv preprint arXiv:2404.01745, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[9] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[10] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.713, + 0.483, + 0.781 + ], + "angle": 0, + "content": "[11] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.781, + 0.483, + 0.835 + ], + "angle": 0, + "content": "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.835, + 0.483, + 0.887 + ], + "angle": 0, + "content": "[13] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[14] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.148, + 0.906, + 0.201 + ], + "angle": 0, + "content": "[15] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.202, + 0.906, + 0.269 + ], + "angle": 0, + "content": "[16] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.27, + 0.906, + 0.351 + ], + "angle": 0, + "content": "[17] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.906, + 0.405 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.906, + 0.473 + ], + "angle": 0, + "content": "[19] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.906, + 0.527 + ], + "angle": 0, + "content": "[20] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[21] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. r2-tuning: Efficient image-to-video transfer learning for video temporal grounding. arXiv preprint arXiv:2404.00801, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[22] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[23] Weiheng Lu, Jian Li, An Yu, Ming-Ching Chang, Shengpeng Ji, and Min Xia. Llava-mr: Large language-and- vision assistant for video moment retrieval. arXiv preprint arXiv:2411.14505, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[24] Boris Meinardus, Anil Batra, Anna Rohrbach, and Marcus Rohrbach. The surprising effectiveness of multimodal large language models for video moment retrieval. arXiv preprint arXiv:2406.18113, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.772, + 0.906, + 0.825 + ], + "angle": 0, + "content": "[25] WonJun Moon, Sangeek Hyun, Su Been Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.906, + 0.895 + ], + "angle": 0, + "content": "[26] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21523" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[28] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.188, + 0.482, + 0.255 + ], + "angle": 0, + "content": "[29] David Pujol-Perich, Albert Clapés, and Sergio Escalera. Sada: Semantic adversarial unsupervised domain adaptation for temporal action localization. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9237–9247. IEEE, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.256, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[30] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13934-13944, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.338, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.42, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[32] Anna Rohrbach, Marcus Rohrbach, Wei Qiu, Annemarie Friedrich, Manfred Pinkal, and Bernt Schiele. Coherent multi-sentence video description with variable level of detail. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 184-195. Springer, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.501, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[33] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.569, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[34] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35:12991-13005, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[36] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2025. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[37] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.828, + 0.482, + 0.881 + ], + "angle": 0, + "content": "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.882, + 0.482, + 0.894 + ], + "angle": 0, + "content": "[39] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Heng-" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "wei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.906, + 0.229 + ], + "angle": 0, + "content": "[40] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.231, + 0.906, + 0.309 + ], + "angle": 0, + "content": "[41] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623-13633, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[42] Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. Understanding negative sampling in graph representation learning. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1666-1676, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[43] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.501, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[45] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3, 4, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21524" + } + ] +] \ No newline at end of file diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_origin.pdf b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7f2e97da976abb51d1623e3f2833da63e60e089 --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/87762bca-719d-433c-abf0-05e54d888e6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90463d08c437111df64109618234eb3fed7b69269a8dcf8a6b9917d9adf518e7 +size 2188312 diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/full.md b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a967181a3bc14c36157821a40ebcaf6cf06a2694 --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/full.md @@ -0,0 +1,339 @@ +# Sparse-Dense Side-Tuner for efficient Video Temporal Grounding + +David Pujol-Perich, Sergio Escalera, Albert Clapés +Universitat de Barcelona and Computer Vision Center, Barcelona, Spain +{david.pujolperich, sescalera, aclapes}@ub.edu + +# Abstract + +Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full finetuning is often impractical, parameter-efficient fine-tuning – and particularly side-tuning (ST)– has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention –a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a $73\%$ the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST. + +# 1. Introduction + +Recently the field of Video Understanding has gained attention due to its potential in applications like search engines or recommendation systems. A key task in this field is Video Temporal Grounding (VTG), which localizes moments within videos based on textual descriptions. Typically, this involves doing both Moment Retrieval (MR) [2, 5, 7, 11, 15] and Highlight Detection (HD) [1, 8, 10, 33, 40]. More specifically, MR predicts moment boundaries w.r.t. textual queries, while HD offers a more interpretable perspective, predicting frame-level saliency scores. + +In the past, existing works focused on either MR or HD, given the lack of datasets that combined both tasks. QVHighlights [15] shifted this paradigm, proposing the first + +![](images/245f34659e271e7f11ba68cd7ae960c6512b826dafc8cbc103298dea05a31ace.jpg) +Figure 1. Comparison of our proposed method and the main MR (left) and HD (right) baselines. These are all evaluated on QVHighlights val split and use InternVideo2-1B features. These results show that our method improves existing side-tuning works and attains SOTA results while significantly reducing the number of trainable parameters. + +![](images/fa0fc1c28d03ed50e28f05ad4f339fc3ab5441bfbbe68f19b090b04042d9d9ff.jpg) + +dataset suitable for this multi-task setup. Since then, various works have been proposed following anchor-based [2, 17, 21, 22], anchor-free [7, 15, 25] and even LLM-based [23, 24] approaches. All of these, however, incur in a common critical limitation, which is relying on the final-layer features of frozen backbones. This is especially limiting when facing large distribution shifts between the pre-training distribution and that of the downstream task [29]. This issue is particularly pronounced in VTG, where an image-domain backbone [31] is transferred to the video domain. + +A typical solution is fine-tuning; however, full fine-tuning is impractical due to its high computational cost. Doing this more efficiently has thus become critical, motivating the interest in parameter-efficient fine-tuning (PEFT) [9] methods -e.g., Prompting or Adapters- which optimize only a small subset of parameters. Unfortunately, these still require full back-propagation through the backbone, making it still memory-intensive. To address this, [34] introduced side-tuning (ST), an effective PEFT, and memory-efficient fine-tuning (MEFT) method. ST creates a parallel pathway to refine intermediate features, thus backpropagating over a minimal set of parameters only. In this regard, we highlight the work of $\mathbb{R}^2$ -Tuning [21], the first ST approach for VTG, which recursively fuses multi-modal intermediate CLIP [31] embeddings. [21] also proposes adapting these frame-level embeddings for MR by generating a dense set of anchors. As shown in Fig. 1, this + +method proves ineffective for MR - a highly sparse task where videos may contain very few ground-truth action. + +This motivates our dual-stream Sparse-Dense Side-Tuner (SDST), the first proposal-free ST method, carefully designed for multi-task learning across sparse (MR) and dense (HD) tasks. For this, SDST jointly refines frame-level embeddings -suitable for HD- while learning what we call the recurrent decoder queries for MR. We also identify an implicit contextual limitation of the deformable attention [45] module of anchor-free architectures like ours, which in our task results in collapsing offsets around their initialization values. This also impedes the selection of keys outside of the current estimated moment boundaries, critical to potentially refine the boundaries of longer moments. Consequently, we propose the alternative Reference-based Deformable Self-Attention (RDSA), which naturally addresses this issue by reformulating the CA into a SA-based mechanism. Finally, we tackle the performance degradation of ST methods derived from the use of image-based CLIP, over more advanced spatio-temporal VLMs [36], as noted in [7]. A key challenge of exploiting [36] is, however, defining an effective token pooling strategy, as naive strategies like CLS pooling yield significant performance drops. This motivates our proposed module re-utilization scheme, which results in the first successful integration of InternVideo2 [36] into an ST framework. + +In summary, our main contributions are threefold: 1) We propose SDST, the first anchor-free ST architecture, specifically tailored for complex sparse-dense multi-task setups like VTG. Our method significantly outperforms existing ST architectures on QVHighlights [15], TACoS [32] and Charades-STA [5] (see Fig. 1). SDST also performs competitively and even surpasses existing SOTA while reducing its learnable parameter count by up to $73\%$ , and incurring in a minimal memory overhead. 2) We identify a key contextual limitation of the deformable attention mechanism that, in our task, results in collapsing offsets. Consequently, we propose RDSA, which tackles this issue by allowing more complex key-selection strategies. 3) We address the limitations of transfer learning from an image-domain backbone to a video domain by integrating, for the first time, the more advanced spatio-temporal backbone [36] into an ST framework. This proves to be a non-trivial challenge with massive performance implications. + +# 2. Related work + +Video temporal grounding (VTG): One relevant task in Video Understanding is VTG [17] which aims to identify actions from text queries. Traditionally, this task was approached either from a MR [2, 5, 7, 11, 15, 25, 26] or HD [1, 8, 10, 33, 40] perspective. MR focuses on predicting specific action proposals while HD aims to compute the saliency scores for each of the frames w.r.t. the query. Nev + +ertheless, after the proposal of QVHighlights [15], the literature started approaching this from a multi-task learning perspective. In this regard, arguably the most notable family of methods is that based on DETR [7, 14, 15, 25, 26, 39] given the sparse nature of MR. These methods normally fuse both video and textual modality into a final embedding which is used for HD and as a memory of a Transformer decoder that refines a set of learnable queries. More in detail, Moment-DETR [15] constitutes the first baseline, which uses a standard Transformer encoder to process both modalities simultaneously. QD-DETR [26] proposes instead to inject the textual information to the video modality with cross attention (CA), followed by a temporal modeling module. CG-DETR [25] reduces the negative impact of irrelevant textual tokens by adding an Adaptive CA module that leverages dummy tokens to redirect attention weight. In this same line, SG-DETR [7] introduces a saliency-guided CA, weighting the contribution of irrelevant textual tokens based on the computed saliency scores. Despite the notable success of these proposal-free methods, one core limitation that they all share is that they rely solely on the last-layer features of a frozen CLIP [31] and/or Slowfast [4] backbones. [41] overcomes this limitation via the full finetuning of CLIP, even though this becomes nearly intractable given the resources that this requires. $\mathbb{R}^2$ -Tuning [21], the most similar method to our work, overcomes this limitation by proposing the first PEFT and MEFT method for VTG, which recursively applies a multi-modal fusion ST. In this work, we argue that one of its core drawbacks, however, is its anchor-based nature, remaining oblivious to the shown benefits of DETR-based architectures in very sparse tasks like MR [7, 16, 20]. This motivates our proposed SDST, which incorporates the best of both PEFT and MEFT. + +Parameter-and-memory-efficient fine-tuning: Foundation models and VLMs [31, 37] have become the cornerstone of recent advances in Video Understanding applications. The use of pre-extracted features remains, however, an important limitation to its application on downstream tasks like VTG [17]. Fine-tuning these huge models is often simply infeasible given the large resources that these require. This has motivated the rise of PEFT [9], which aims to democratize the fine-tuning process by restricting the tunable parameters to the bare minimum. In this regard, some of the most prominent approaches are Prompting [12, 44] and the use of Adapters [6, 28]. The first keeps the backbone frozen while learning prompts that are appended to the input, bridging the gap between the backbone's expected distribution and that of the downstream task. Adapters, in contrast, incorporate small trainable modules in the backbone while keeping the rest unchanged. These methods, however, still require full backpropagation through the frozen model, making them memory-inefficient. Recently, the new paradigm of $ST$ [19, 30, 34] –and particularly [21] + +for VTG- has gained attraction creating a parallel pathway to exploit intermediate backbone representations while guaranteeing that backpropagation is only applied to this small parallel module. Interestingly, these methods normally rely on pre-extracted CLIP representations for tasks like HD/MR. As shown by [7], this is a very limiting aspect given the lack of temporal reasoning of CLIP or its difficulties in understanding textual queries beyond simple spatial descriptions. For this reason, in this work, we propose a novel ST architecture that, for the first time, relies on the more advanced InternVideo2 [36] backbone. Unlike CLIP, this backbone is trained on video-domain inputs, resulting in enhanced spatial and temporal modeling capabilities. + +Deformable attention: A key component of most SOTA methods for VTG like SG-DETR [7] is their deformable attention module [38, 45]. This module addresses the well-studied slow convergence of DETR [3]. It does so by limiting the selectable keys based on the predictions of the learnable queries, completely decoupling both query and key spaces. Despite desirable in terms of efficiency, it has been previously noted in the literature [43] that the lack of context of the queries w.r.t. the key/value space can lead to suboptimal attendable key selection. VTG methods often mitigate this effect by partially or completely initializing the learnable queries based on the DETR-memory [43] -i.e., frame-level representations in our case. In this work, we empirically demonstrate the limitations of these initialization-based methods for VTG tasks, which motivates our proposed simple yet effective alternative, the Reference-based Deformable Self-Attention (RDSA). This reformulates the deformable CA into a deformable SA, naturally solving the aforementioned issue. + +# 3. Method + +# 3.1. Problem definition + +In this paper, we address the problem of MR and HD based on textual descriptions. For this, we consider an arbitrary input-query pair $(\mathbf{X}^v,\mathbf{X}^t)$ where $\mathbf{X}^v\in \mathbb{R}^{T\times H\times W\times 3}$ and $\mathbf{X}^t\in \mathbb{R}^{L\times F_e}$ . Here $T$ and $L$ are the number of frames and tokens, and $F_{e}$ the textual encoding dimension. Our goal is to predict, for the given video-query pair, its respective frame-wise saliency scores $\mathbf{Y}^s\in \mathbb{R}^T$ for HD, and the $M$ action moments $\mathbf{Y}^m\in \mathbb{R}^{M\times 2}$ for MR. + +# 3.2. Overview + +In our work we propose a dual-stream ST architecture that includes a dense (frame-level) and a sparse (segment-level) stream for MR and HD respectively (see Fig. 2a). More specifically, we first leverage a frozen InternVideo2-1B [36] backbone, a model that possesses powerful spatio-temporal modeling capabilities, to extract $K$ intermediate visual-textual representations. These are then processed by multiple weight-sharing SDST layers, which refine dense rep + +resentations and recurrent decoder queries. More in detail, at each level, the model refines the dense embeddings via multi-modal textual conditioning as well as temporal modeling. This signal is then used to condition the sparse stream $S$ , that builds on [16] while incorporating our novel deformable attention mechanism, namely Reference-based Deformable Self-Attention (see Fig. 2b). This mitigates the shortcomings of the deformable CA mechanism, enhancing the contextual information of the decoder queries and thus improving the quality of the selected keys. After unrolling this process $K$ times, we apply the dense and the sparse prediction heads to compute the saliency scores and the predicted segment boundaries, respectively. + +# 3.3. Sparse-Dense Side-Tuner (SDST) + +# 3.3.1. Recurrent refinement with side-tuners + +Our work builds on an ST framework, which first requires extracting $K$ intermediate video and textual features $\tilde{\mathbf{V}}\in$ $\mathbb{R}^{K\times T\times F_v}$ and $\tilde{\mathbf{T}}\in \mathbb{R}^{K\times L\times F_t}$ (find the details in Sec. 4). Here $F_{v}$ and $F_{t}$ are their respective dimensionalities. We then define a zero-initialized dense embedding $\mathbf{D}^0\in \mathbb{R}^{T\times F}$ and a set of $M$ learnable recurrent moment proposals inspired by [20], which we refer to as recurrent decoder queries. The queries include the learnable center-width moment references $\mathbf{R}^0\in \mathbb{R}^{M\times 2}$ , and their corresponding latent embeddings $\mathbf{H}^0\in \mathbb{R}^{M\times F}$ . We then define the recurrence for a given level $1\leq \ell \leq K$ as follows: + +$$ +\mathbf {D} ^ {\ell + 1}, \mathbf {R} ^ {\ell + 1}, \mathbf {H} ^ {\ell + 1} = S D S T \left(\mathbf {D} ^ {\ell}, \mathbf {R} ^ {\ell}, \mathbf {H} ^ {\ell}, \tilde {\mathbf {V}} ^ {\ell}, \tilde {\mathbf {T}} ^ {\ell}\right), \tag {1} +$$ + +where SDST represents the shared-weight layers of our model, progressively refining the representations across two streams: the 1) dense learning stream $\mathcal{D}$ and the 2) sparse learning stream $S$ , that we describe below. + +# 3.3.2. Dense learning stream $\mathcal{D}$ + +Following [21], this stream first refines the dense embeddings $\mathbf{D}^{\ell}$ by conditioning it to the video-textual embeddings $\tilde{\mathbf{V}}^{\ell}$ and $\tilde{\mathbf{T}}^{\ell}$ . For this, both of these multi-modal embeddings are projected to a shared $F$ -dimensional space, using two MLPs $(\mathcal{F}_v:F_v\to F$ and $\mathcal{F}_t:F_t\to F)$ : + +$$ +\mathbf {V} ^ {\ell} = \mathcal {F} _ {v} \left(\tilde {\mathbf {V}} ^ {\ell}\right), \mathbf {T} ^ {\ell} = \mathcal {F} _ {t} \left(\tilde {\mathbf {T}} ^ {\ell}\right). \tag {2} +$$ + +Then we incorporate the clip-wise visual information to $\mathbf{D}^{\ell}$ via a weighted sum modulated by $\beta^{\ell} \in [0,1]$ , a zero-initialized layer-dependent parameter that yields: + +$$ +\mathbf {D} ^ {\ell} := \beta^ {\ell} \mathbf {D} ^ {\ell} + (1 - \beta^ {\ell}) \mathbf {V} ^ {\ell}. \tag {3} +$$ + +Thereafter, we inject the textual information with CA, where $\mathbf{T}^{\ell}$ does not include the CLS token. We also apply a temporal modeling module defined as a SA and a subsequent point-wise feed forward network (PFFN): + +$$ +\mathbf {D} ^ {\ell + 1} = P F F N \left(S A \left(C A \left(\mathbf {D} ^ {\ell}, \mathbf {T} ^ {\ell}, \mathbf {T} ^ {\ell}\right)\right)\right), \tag {4} +$$ + +![](images/70f0a1d87307015aebb2100906a00ad80a96633368a9a88475129bdd570cf72e.jpg) +(a) Overview of our proposed architecture. + +![](images/d94d93dd84419ed993f4e092c9c72f14e9c5a29a2ee0cf177334e97d4a0dd41f.jpg) +(b) RDSA vs Def.CA[38] for a proposal $m$ . +Figure 2. Our method (left) first processes the video and textual inputs using [36], and then recursively applies a shared SDST, a recurrent dual-stream model attached to the last $K$ layers. This iteratively refines dense embeddings and learnable recurrent decoder queries. The SDST consists of a dense stream $\mathcal{D}$ for temporal and multi-modal dense refinement, and a sparse stream $\mathcal{S}$ that refines the recurrent queries conditioned on the dense signal. $\mathcal{S}$ includes the RDSA module (right), a context enhanced deformable attention mechanism that restricts the selectable keys based on the center, left-, and right-most action embeddings. The final dense embedding solves HD, while the recurrent queries address MR. + +# 3.3.3. Sparse learning stream $S$ + +This stream can be seen as a recurrent DETR-based mechanism that exploits the complementarity of both sparse (MR) and dense (HD) tasks to refine the recurrent decoder queries. This is, refining the center-width references $\mathbf{R}^{\ell}\in \mathbb{R}^{M\times 2}$ and their latent embeddings $\mathbf{H}^{\ell}\in \mathbb{R}^{M\times F}$ . Similar to the stream $\mathcal{D}$ , we first incorporate a CA module that conditions $\mathbf{H}^{\ell}$ to the textual query. This is followed by a SA module to enable the information flow between different moment proposals: + +$$ +\mathbf {H} ^ {\ell} = S A \left(C A \left(\mathbf {H} ^ {\ell}, \mathbf {T} ^ {\ell}, \mathbf {T} ^ {\ell}\right)\right). \tag {5} +$$ + +Another essential aspect of this stream is effectively conditioning the recurrent decoder queries to the video modality. Existing works [7, 16, 20] typically rely on the use of deformable attention due to its convergence and performance benefits [38, 45]. This mechanism leverages attention queries to predict offsets that define a limited set of selectable keys. However, in this task we observe that 1) these offsets collapse near their initialization and 2) are unable to look beyond the current estimated boundaries (see Sec. 6.2). We attribute this to contextual limitations, which motivates our proposed alternative Reference-based Deformable Self-Attention (RDSA), enabling an enhanced cross-modality injection (see Sec. 3.3.4). We also incorporate a simple PFFN to enhance the model's expressivity: + +$$ +\mathbf {H} ^ {\ell} = P F F N \left(R D S A \left(\mathbf {R} ^ {\ell}, \mathbf {H} ^ {\ell}, \mathbf {D} ^ {\ell + 1}\right)\right). \tag {6} +$$ + +Importantly, all of these modules are shared across the $K$ levels for parameter efficiency, and include residual connections and DropPath [13] for stability and regularization. + +# 3.3.4. Reference-based deformable self-attention + +Deformable attention in the context of existing anchor-free methods: One of the keys to the success of recent anchor-free methods like DETR -including works tackling VTG [7]- is the use of deformable attention mecha + +nisms [45] for cross-modal interaction. This mechanism is known to provide faster convergence speed, reduced time complexity, and as shown in Sec. K, overall better performance than standard CA. As depicted in Fig. 2b (down), the key of this module is the lack of explicit interaction between queries and keys, which are defined as + +$$ +\mathbf {Q} = \mathbf {X} _ {\mathcal {Q}} \mathbf {W} _ {\mathcal {Q}} ^ {d e f}, \mathbf {K} = \mathbf {X} _ {\mathcal {K}} \mathbf {W} _ {\mathcal {K}} ^ {d e f}, \tag {7} +$$ + +where $\mathbf{W}_{\mathcal{Q}}^{def}$ , $\mathbf{W}_{\mathcal{K}}^{def}$ are two linear projections. The deformable attention thus avoids computing the $\mathbf{Q}\mathbf{K}^T$ similarity matrix, and instead employs an offset and attention-score predictors, $\mathcal{G}_{\Delta}$ and $\mathcal{G}_{\mathbf{A}}$ to select and weight -only based on the queries - a small subset $P$ of selectable keys. The output of this mechanism, the weighted aggregation of the selected keys, namely $\mathbf{S}$ , is computed as follows + +$$ +\Delta = \mathcal {G} _ {\Delta} (\mathbf {Q}) \in \mathbb {R} ^ {M \times P}, \mathbf {A} = \mathcal {G} _ {\mathbf {A}} (\mathbf {Q}) \in \mathbb {R} ^ {M \times P}, \tag {8} +$$ + +$$ +\mathbf {S} = \sum_ {p = 1} ^ {P} \left(\mathbf {A} _ {:, p} \mathbf {K} [ \mathbf {c} + \mathbf {w} \odot \Delta_ {:, p} ]\right) \in \mathbb {R} ^ {M \times F} \tag {9} +$$ + +where $\mathcal{G}_{\Delta}$ , $\mathcal{G}_{\mathbf{A}}$ are often modeled as CNNs and $x[y]$ is the bilinear sampling of $x$ on index(es) $y$ . Notably, in our 2-dimensional setup, the offsets $\Delta_{:,p}$ are added to center reference $\mathbf{c} = \mathbf{R}_{:,0}^{\ell} \in \mathbb{R}^{M}$ , and weighted over the action width $\mathbf{w} = \mathbf{R}_{:,1}^{\ell} \in \mathbb{R}^{M}$ . See Sec. K for additional details. + +Limitations: Interestingly, in the literature we find that when it comes to the deformable attention, most works operate with a certain independence to whether they deal with a deformable self-attention or cross-attention, assuming that the benefits of this mechanism extrapolate to both scenarios. In this work, however, we find that in fact, the deformable attention is only naturally suited for self-attention scenarios as its effectiveness inherently relies on implicit interactions between $\mathbf{Q}$ and $\mathbf{K}$ -not present in the cross-attention case. Concretely, in a self-attention setup, $\mathbf{X}_{\mathcal{Q}} = \mathbf{X}_{\mathcal{K}}$ (see Eq. 7). Thus, modeling $\mathcal{G}_{\Delta}$ and $\mathcal{G}_{\mathbf{A}}$ as CNNs naturally al + +lows queries to gain context of the local neighborhood of queries, and consequently, of the key/value space. This breaks their independence assumption, giving the queries critical information to decide where to look. + +Importantly, this is not the case with the deformable cross attention where we try to inject cross-modal information from the keys to the queries, which is our goal in this work. In this scenario, $\mathbf{X}_{\mathcal{Q}} \neq \mathbf{X}_{\mathcal{K}}$ , and thus, using CNN-based offset predictors does not inject knowledge of the key/value space, making its guesses blind or based on the blind predictions of previous iterations. Concretely, in Sec. 6.2 we show that in our task, the drawbacks of this naive deformable cross-attention have various empirical implications such as the offset collapse near the offset initialization values, or the inability to select keys from frames beyond the estimated action boundaries, critical to refine segments into longer actions. This motivates our proposed Reference-based Deformable Self-Attention, an alternative mechanism that allows reformulating the deformable CA module into a deformable SA based on the references $\mathbf{R}^{\ell}$ . + +Our proposed alternative (RDSA): Our proposed RDSA (see Fig. 2b) takes as input the center-width references $\mathbf{R}^{\ell}\in \mathbb{R}^{M\times 2}$ and the dense embeddings $\mathbf{D}^{\ell}\in \mathbb{R}^{T\times F}$ . Then, differently from other methods that leverage the learnable queries $\mathbf{Q} = \mathbf{H}^{\ell}$ to compute the offsets and attention scores (see Eq. 8), we propose using an alternative query embedding. Concretely, we first refine the dense embedding $\mathbf{D}_l$ with a simple CNN, allowing frames to gain local context of their surroundings. Then, we use bilinear sampling to extract three key action embeddings: left-most (l), center (c), and right-most (r). The center provides highly informative action features, while the extremes help refine the moment boundaries. This yields, + +$$ +\hat {\mathbf {Q}} = \hat {\mathbf {X}} _ {\mathcal {Q}} \mathbf {W} _ {\mathcal {Q}} ^ {d e f}, \hat {\mathbf {X}} _ {\mathcal {Q}} = C N N (\mathbf {D} ^ {\ell}) [ l, c, r ]. \tag {10} +$$ + +where $\hat{\mathbf{X}}_{\mathcal{Q}}\in \mathbb{R}^{M\times 3F}$ . This rewrites Eq. 8 to + +$$ +\hat {\Delta} = \mathcal {G} _ {\Delta} (\hat {\mathbf {Q}}), \hat {\mathbf {A}} = \mathcal {G} _ {\mathbf {A}} (\hat {\mathbf {Q}}). \tag {11} +$$ + +Finally, we apply the deformable attention from Eq. 9 using the new offsets and attention scores $\hat{\Delta}$ and $\hat{\mathbf{A}}$ . This can be viewed as asking the key spots of the action where to look based on the textual information that we previously injected in Eq. 5. This naturally solves the context limitation of the standard deformable CA given that $\hat{\mathbf{X}}_{\mathcal{Q}}$ and $\mathbf{X}_{\mathcal{K}}$ now live in the same latent space -as they are both derived from $\mathbf{D}^{\ell}$ . This allows the model to make informed decisions about which keys to attend to while keeping all its core computational advantages. + +# 3.4. Prediction heads and training objectives + +After recursively applying the SDST over $K$ intermediate layers, we compute the respective saliency scores and action segment proposals, as well as define their corresponding objective functions. We briefly describe these below, + +but refer interested readers to Sec. C for more details. + +Highlight Detection: To solve HD, we define the saliency scores as the cosine similarity between the dense embedding $\mathbf{D}^K$ and a pooled representation of $\mathbf{T}^K$ . We then apply an InfoNCE loss [27] that learns a ranking of these scores + +$$ +\mathcal {L} _ {H D} = \lambda_ {0} \mathcal {L} _ {\text {I n f o N C E}}. \tag {12} +$$ + +Moment Retrieval: To solve MR we first apply the Hungarian algorithm to obtain a one-to-one matching between the predicted queries and the GT actions. We then project $\mathbf{H}^{\ell}$ to compute the action probability $\mathbf{p} \in \mathbb{R}^{M}$ of the $M$ different queries, which we train with a FocalLoss [18]. We also compute the moment boundaries $\hat{\mathbf{Y}}^{m} \in \mathbb{R}^{M \times 2}$ which we optimize using an L1 and an IoU loss. Finally, we use an L1 loss to learn an actionness score, which predicts the maximum IoU of each query w.r.t. the GT actions + +$$ +\mathcal {L} _ {M R} = \lambda_ {1} \mathcal {L} _ {a c t} + \sum_ {l = 1} ^ {\ell} \lambda_ {2} \mathcal {L} _ {c l s} ^ {\ell} + \lambda_ {3} \mathcal {L} _ {l 1} ^ {\ell} + \lambda_ {4} \mathcal {L} _ {I o U} ^ {\ell}. \tag {13} +$$ + +Note that all but the $\mathcal{L}_{act}$ term are optimized across different refinement levels to promote a faster convergence. + +Alignment losses: To improve the video-textual alignment, we define $\mathcal{L}_{\text {align }}$ based on two SampledNCE losses [42] to align $\mathbf{V}$ and $\mathbf{T}$ embeddings across the different refinement levels. We apply this along both the batch and the intermediate-layer dimensions. This pulls the action embeddings closer to the textual representations while learning complementary information across intermediate layers. + +Final loss: The final loss is defined as + +$$ +\mathcal {L} = \lambda_ {5} \mathcal {L} _ {H D} + \lambda_ {6} \mathcal {L} _ {M R} + \lambda_ {7} \mathcal {L} _ {\text {a l i g n}}. \tag {14} +$$ + +# 4. Extracting intermediate representations of InternVideo2 + +The cornerstone of ST is leveraging intermediate multimodal representations from frozen VLMs. While previous ST approaches predominantly rely on CLIP, we argue that CLIP has notable limitations in temporal modeling and aligning visual features with textual queries beyond simple static descriptions. To overcome these limitations, we incorporate InternVideo2 [36] into our framework, leveraging its advanced spatio-temporal modeling capabilities [7]. However, extracting intermediate representations from InternVideo2 –particularly from its visual encoder $\mathcal{E}_v$ – remains a challenging and underexplored task in the literature, despite its significant impact on performance. + +In particular, the main challenge arises from how to pool the spatio-temporal token dimension $L_{v}$ of $\hat{\mathbf{V}}^{\ell} \in \mathbb{R}^{T \times L_v \times F}$ , the output embedding of the $\ell$ -th layer of $\mathcal{E}_v$ into the final intermediate features $\tilde{\mathbf{V}}^{\ell} \in \mathbb{R}^{T \times F}$ (see Eq. 1). Prior CLIP-based methods simply use the CLS token as "summaries" over $L_{v}$ . However, we find this strategy to be suboptimal for InternVideo2 (see Sec. 6.1) as this limits its + +spatial-aggregation capabilities. Notably, [36] optimizes an AdaptivePool module that computes the final-layer embeddings, which are then aligned with the text. We conjecture that leveraging this enhanced multi-modal alignment is key to performance. Unfortunately, such pooled representation is only computed at the last layer, leaving the question of how to pool the remaining intermediate layers. Ideally, we would optimize layer-independent AdaptivePool modules, but this would require full back-propagation through the entire model. A more efficient alternative would be computing gradients of the AdaptivePool modules only. This is, however, still computationally infeasible, not for the gradients themselves, but because of the need to load $L_{v}$ spatiotemporal tokens per frame. This increase in the input memory size translates, for instance, to a $15 \times$ memory increase for QVHighlights. We thus hypothesize that re-using this frozen pooling module across the $K$ intermediate layers allows a memory efficient training –no additional backpropagation nor memory requirements are needed– while leveraging its enhanced pooling capabilities, even despite their respective distribution shifts. Formally, we compute + +$$ +\tilde {\mathbf {V}} ^ {\ell} = \text {A d a p t i v e P o o l} \left(\hat {\mathbf {V}} ^ {\ell}\right) \in \mathbb {R} ^ {T \times F}, 1 \leq \ell \leq K. \tag {15} +$$ + +As shown in Sec. 6.1, this pooling strategy considerably improves other alternatives, demonstrating the benefits of this module re-utilization. + +# 5. Experimentation + +# 5.1. Experimental setup + +We test our proposal on various datasets for MR and HD, namely QVHighlights [15], TACoS [32], and the Charades-STA [5] dataset. Further details of the datasets can be found in Sec. B. For the task of MR on QVHighlights, we compute Recall@1 with two IoU thresholds, 0.5 and 0.7, and the mean average precision (mAP) at IoU thresholds between 0.5 and 0.95 with a step of 0.05 -i.e., [0.5:0.05:0.95]. For HD, we report mAP and HIT@1 over the positive frames -i.e., the most salient ones classified as "VeryGood". For TACoS and Charades-STA, we compute Recall@1 at 0.3, 0.5, and 0.7 IoU thresholds, as well as mIoU. + +# 5.2. Main experimental results + +We begin by comparing the performance of SDST over various relevant baselines. Concretely, observe in Tab. 1 the evaluation of our method on the test and val splits of QVHighlights. These results indicate that our method considerably outperforms $\mathbb{R}^2$ -Tuning. Similarly, SDST performs very competitively and even surpasses on several metrics the current SOTA method, namely SG-DETR [7]. Note however that our method attains this performance while having only a $27\%$ of the parameter count of SG-DETR. We refer interested readers to Sec. M for the statistical significance analysis. + +Most of the presented methods in Tab. 1 rely on alternative backbones different from InternVideo2 -e.g., CLIP or Slowfast. To establish a more fair comparison, in Fig. 1 we follow the work of [7] and evaluate a set of relevant baselines when using InternVideo2 features -find the complete ablation in Sec. E. Observe that our work performs on par with the SG-DETR and even more importantly to our work, the SDST very considerably outperforms $\mathbb{R}^2$ -Tuning, improving the MR performance by a $3.82\%$ average mAP or the HD by a $2.21\%$ mAP. This empirically demonstrates the benefits of leveraging a proposal-free architecture over a more rigid anchor-based method. Also find in Sec. H an extended comparison with other ST works. + +To complement our findings, in Tab. 2 we evaluate our method on the Charades-STA and TACoS benchmarks. Observe that our method obtains SOTA in both of these datasets. In Charades-STA, for instance, improving by $2.71\%$ of R1@0.7 and $2.06\%$ of mIoU, the existing SOTA. Similarly, in TACoS, SDST improves SG-DETR by $2.39\%$ R1@0.7 and $1.27\%$ mIoU. + +# 6. Ablation studies + +In this section, we ablate over various relevant aspects of our main contributions. Note that unless states otherwise, we follow the literature and evaluate them on the val split of QVHighlights. Also, we write write in bold the best results. + +# 6.1. Leveraging InternVideo2 features for ST + +1) Does pooling matter? As argued in Sec. 4, one of the main challenges in using the InternVideo2 backbone for ST is pooling the spatio-temporal clip-wise intermediate representations. In Tab. 3 we compare different pooling strategies. Concretely, we compare the standard CLS-pooling, average-pooling, and our proposed re-utilization of the frozen AdaptivePool (Sec. 4). Observe that CLS-pooling results in a considerable performance degradation of up to $5.07\%$ average mAP on MR or $7.74\%$ HIT@1 on HD w.r.t. the adaptive pooling strategy. The average pooling partially mitigates this degradation, but still obtains a considerably decreased performance. This confirms the significant impact of a carefully chosen pooling strategy. + +2) Feature refinement vs. feature sampling: One of the most critical decisions of ST is determining the number of intermediate levels that should be used. Normally, doing so involves evaluating the performance when refining the last $K$ intermediate layers [21]. As shown in Fig. 3 (red) this indicates that in our case, the best performing $K$ is 4. The literature often suggests this argument to be compelling enough to claim that using intermediate representations is beneficial over using only last-layer features. In this work, however, we find it reasonable to wonder: Does this analysis depict the importance of intermediate features, or is it simply evidencing the need to do multiple refinement steps, + +
Methodtestval#Params
MRHDMRHD
R1@0.5@0.7mAP@0.75Avg.≥ Very good mAP HIT@1R1@0.5@0.7@0.5mAP@0.75Avg.≥ Very good mAP HIT@1
BeautyThumb----14.3620.88--------
DVSE----18.7521.79--------
MCN11.412.7124.948.2210.67---------
CAL25.4911.5423.407.659.89---------
XML+46.6933.4647.8934.6734.9035.3855.06-------
Moment-DETR52.8933.0254.8229.4030.7335.6955.6053.9434.84-32.2035.6535.654.8M
UMT56.2341.1853.8337.0136.1238.1859.9960.2644.2656.7039.9038.5939.9064.20
MomentDiff58.2141.4854.5737.2136.84---------
QD-DETR62.4044.9862.5239.8839.8638.9462.4062.6846.6662.2341.8241.2239.1363.03
MH-DETR60.0542.4860.7538.1338.3838.2260.5160.8444.9060.7639.6439.2638.7761.74
UniVTG58.8640.8657.6035.5935.4738.2060.9659.74---36.1338.8061.8
TR-DETR64.6648.9663.9843.7342.6239.9163.4267.1051.4866.2746.4245.09--
CG-DETR65.4348.3864.5142.7742.8640.3366.2167.3552.0665.5745.7344.9340.8066.70
BAM-DETR62.7148.6464.5746.3345.36--65.1051.6165.4148.5647.61--
EaTR------61.3645.7961.8641.9141.7437.1558.659.0 M
Mr. BLIP74.7760.5168.1253.38--76.1363.3569.3955.78---19.0 M
LLaVA-MR76.5961.4869.4154.40--78.1364.1369.6456.32---17.0 M
HL-CLIP-----41.9470.60----42.3772.402.0 M
R²-Tuning68.0349.3569.0447.5646.1740.7564.2068.7152.06--47.5940.5964.32
SG-DETR†72.2056.6073.2055.8054.1043.7669.13--73.5257.9155.6443.9171.47
Flash-VTG†70.6953.9672.3353.8552.00--73.1057.2972.7554.3352.84--
Ours†70.8256.2371.3154.9953.3143.4069.1373.6860.9073.5257.4255.6044.0072.00
+ +Table 1. MR and HD results for QVHighlights test and val split.† indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best, considering only works supporting both MR and HD. The rest are marked in gray. + +
MethodCharades-STTACoS
R@0.5R@0.7mIoUR@0.5R@0.7mIoU
M-DETR53.631.4-24.712.025.5
UMT48.329.3----
UniVTG58.035.750.135.017.433.6
QD-DETR57.332.6-36.821.135.8
CG-DETR58.436.350.139.622.236.5
BAM-DETR60.039.452.341.526.839.3
TR-DETR57.633.5----
MR_BLIP69.349.358.6---
LLaVA-MR70.649.659.8---
R2-Tuning59.837.050.938.725.135.9
SG-DETR†70.249.559.144.729.940.9
FlashVTG†70.349.9-41.824.737.6
Ours†72.052.661.244.532.342.2
+ +Table 2. Comparison on Charades-STA and TACoS datasets. ${}^{ \dagger }$ indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best. + +
Pool strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
CLS66.4552.6567.9651.3250.5341.0164.26
Avg. pool70.6556.970.4354.5853.4443.0669.68
Adapt. pool73.6860.9073.5257.4255.6044.0072.00
+ +regardless of the chosen features? + +To shed light on this issue, in Fig. 3 (blue) we evaluate the performance when doing $K$ different refinement steps, but importantly, we always use the last-layer features only. Observe that for $K = 2$ and $K = 3$ , actually, this improves the homologous experiments with intermediate features. Differently, for $K = 4$ and $K = 5$ , the use of last-layer features harms the performance. These experiments thus suggest that proving the usefulness of intermediate features is not as straightforward as previously thought. In fact, these indicate that the advantages of using intermediate features arise only as we consider shallower layers. + +3) Why not sampling features from even shallower layers? As shown in Tab. 4, this is not necessarily helpful + +![](images/3e3c9c98f97f220cfc053f685d964636bdbe6c1c2850c327052a92b3e5ff9093.jpg) +Figure 3. Ablation of the number of refinement levels and of their use of intermediate or last-layer features only. + +![](images/618e30d98c0ccde30948435dac14d414063400100c51eea9f2de997e93d982d7.jpg) + +Table 3. Effect of using different pooling strategies. + +
LevelsMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
[37,38,39,40]73.6860.9073.5257.4255.6044.072.00
[25,30,35,40]71.4859.4871.4556.3754.8343.4270.19
[10,20,30,40]69.0355.3569.5653.3752.4742.7368.39
+ +Table 4. Ablation of different sampling strategies of intermediate features. The first column indicates the $K = 4$ sampled layers –of a total of 40 layers. + +because of what we call, depth-pooling trade-off. We conjecture that sampling from shallower layers does in fact provide additional complementary information. Nevertheless, this inevitably results in a distribution shift w.r.t. the last-layer feature distribution, hindering the effectiveness of the frozen AdaptivePool (see Eq. 15). The infeasibility of retraining this module for computational reasons, thus, leaves us with a delicate trade-off between the quality of the pooling and the depth of the features that must be accounted for. + +# 6.2. Study of deformable attention + +1) Comparison w.r.t. other baselines: Here we empirically evaluate the benefits of our proposed RDSA. Concretely, in Tab. 5 we empirically compare the performance of the standard CA [35] and Def. CA [45], as well as different decoder-query initialization strategies [43] for the latter. Notice the prominent performance degradation de + +
Att. strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Stand. CA [35]70.7751.7466.7845.1942.7243.1669.87
Def. CA [45]72.5858.1971.8255.8054.2743.2670.58
Def. CA +PureInit [43]72.1357.8770.6854.8552.9243.1970.32
Def. CA +MixedInit [43]72.2658.7171.9456.4354.9443.1069.86
Ours73.6860.9073.5357.4255.6044.072.00
+ +![](images/eb63f808cbe9a3924b16163b2f897ae83feb7f744fcc3010ab9e4416e6eca1df.jpg) +Figure 4. Average of the weighted offsets across $M$ decoder queries and $N$ batch elements for the $K = 4$ refinement levels. Here head 0 (left) is initialized near the left boundary, and head 1 (right) near the right boundary. + +![](images/f0dc62f156c17fe0ccc319e3ccfc324a86027ac54bfe06751d2ff0dc41b009e4.jpg) + +rived from the use of the CA, or how the query initialization techniques –far from improving the performance of the Def. CA– in some cases are even harmful. In contrast, our method consistently outperforms all these tested baselines, improving by $2.71\%$ R1@0.7 or $1.62\%$ mAP w.r.t. the original Def. CA. We refer to Sec. K for the ablation of the different sampling strategies of RDSA. + +2) Where are the offsets pointing to? In Fig. 4 we depict the refinement of the weighted offset of a given query $q$ , defined as $d_{q} = \sum_{p=1}^{P} A_{q,p} \Delta_{q,p}$ , across the $K$ different refinement steps. Note that as depicted in Eq. 9, these offsets are relative to the estimated moment widths, with -1 and +1 indicating the estimated left- and right-most boundaries. The Def. CA lacks context awareness, causing its offset predictions to stay near the offset initializations. In contrast, RDSA leverages an enhanced contextualization, thus reducing its bias towards the offset initialization. Notably, RDSA learns to look beyond the current action boundaries (offsets < -1 and > +1) capturing near-boundary information, crucial for refining moment boundaries. This contrasts with [45] which remains constrained within the predicted moment boundaries. We consider this the reason why RDSA is especially useful to localize long actions (see Sec. K). + +# 6.3. Conditioning signal for the sparse stream $\mathcal{S}$ + +To effectively integrate RDSA, it is imperative to obtain an expressive video-based conditioning signal -i.e., the red arrow in Fig. 2a. As shown in Tab. 6, using the raw video representation $\mathbf{V}^{\ell}$ significantly degrades performance, highlighting the need for a richer signal incorporating textual and temporal information. While conditioning on the outputs of the CA or SA modules of the dense stream $\mathcal{D}$ improves results, these lack non-linearities. Instead, using the non-linear PFFN output provides greater flexibility, leading to superior expressiveness. Notably, this choice also enhances HD performance, underscoring the benefits of task + +Table 5. Comparison across different attention strategies as well as various decoder query initializations. + +
Type of featuresMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Raw. rep72.5856.3770.1250.9950.7444.1371.37
Post CA72.2658.3972.2055.9254.8343.6470.65
Post CA+SA72.1358.3272.0056.5754.9443.7870.06
Post CA+SA+FFN73.6860.9073.5357.4255.6044.0072.00
+ +Table 6. Evaluation of various conditioning signals. + +
CASARDSAFFNMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
✓✓✓71.4257.7471.7356.1854.4743.5371.35
✓✓✓71.8758.7172.1355.9954.3243.2869.55
✓✓✓59.6830.7159.0224.629.0943.5870.39
✓✓✓72.8459.2973.1356.9155.7243.8572.58
✓✓✓✓73.6860.9073.5257.4255.6044.072.00
+ +Table 7. Importance of the modules of the sparse stream $S$ . + +interaction between MR and HD. + +# 6.4. Effect of each module of the sparse stream $\mathcal{S}$ + +In this ablation, we empirically evaluate the contribution of each of the components of $S$ . Concretely, we evaluate the influence of the CA, the SA, the RDSA, and the final PFFN. Observe in Tab. 7 that the initial textual injection CA seems to be particularly relevant for the R1 metric, while the reasoning SA module is important for the mAP. Moreover, as expected, the RDSA module is critical. This is to be expected, as without it, the queries become oblivious to the input video, leaving no other option but that of randomly guessing. Again, the additional expressivity that the final PFFN brings via its non-linearities proves to be beneficial, despite affecting less than the other modules. + +# 7. Conclusions + +In this work, we introduced the Sparse-Dense Side-Tuner, the first proposal-free side-tuning method for VTG. Our dual-stream architecture jointly optimizes dense embeddings and recurrent decoder queries, carefully adapting to the sparse and dense nature of MR and HD. We also identify an inherent context limitation of the Deformable CA-a key component of proposal-free architectures like ourswhich motivates our proposed alternative, the Referencebased Deformable Self-Attention. Finally, we provide the first effective integration of InternVideo2 to a side-tuning framework, leading to a substantial performance boost. We evaluate our method on QVHighlights, TACoS, and Charades-STA, surpassing previous ST works and attaining either SOTA or near-SOTA, all while maintaining a minimal parameter count and a low memory overhead. We leave as future work studying the often overlooked problem of the cross-domain generalization capabilities of these methods. + +Acknowledgments: This work has been partially supported by the Spanish project PID2022-136436NB-I00 and by ICREA under the ICREA Academia programme. + +# References + +[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 1, 2 +[2] Zhuo Cao, Bingqing Zhang, Heming Du, Xin Yu, Xue Li, and Sen Wang. Flashvtg: Feature layering and adaptive score handling network for video temporal grounding. arXiv preprint arXiv:2412.13441, 2024. 1, 2 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2 +[5] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 6 +[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 2 +[7] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 1, 2, 3, 4, 5, 6 +[8] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. arXiv preprint arXiv:2404.01745, 2024. 1, 2 +[9] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2 +[10] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 1, 2 +[11] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2 +[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2 +[13] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. 4 + +[14] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 2 +[15] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 1, 2, 6 +[16] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 2, 3, 4 +[17] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2 +[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[19] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 2 +[20] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3, 4 +[21] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. r2-tuning: Efficient image-to-video transfer learning for video temporal grounding. arXiv preprint arXiv:2404.00801, 2024. 1, 2, 3, 6 +[22] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1 +[23] Weiheng Lu, Jian Li, An Yu, Ming-Ching Chang, Shengpeng Ji, and Min Xia. Llava-mr: Large language-and- vision assistant for video moment retrieval. arXiv preprint arXiv:2411.14505, 2024. 1 +[24] Boris Meinardus, Anil Batra, Anna Rohrbach, and Marcus Rohrbach. The surprising effectiveness of multimodal large language models for video moment retrieval. arXiv preprint arXiv:2406.18113, 2024. 1 +[25] WonJun Moon, Sangeek Hyun, Su Been Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, 2023. 1, 2 +[26] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 2 + +[27] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[28] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 2 +[29] David Pujol-Perich, Albert Clapés, and Sergio Escalera. Sada: Semantic adversarial unsupervised domain adaptation for temporal action localization. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9237–9247. IEEE, 2025. 1 +[30] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13934-13944, 2023. 2 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[32] Anna Rohrbach, Marcus Rohrbach, Wei Qiu, Annemarie Friedrich, Manfred Pinkal, and Bernt Schiele. Coherent multi-sentence video description with variable level of detail. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 184-195. Springer, 2014. 2, 6 +[33] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2 +[34] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35:12991-13005, 2022. 1, 2 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7, 8 +[36] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2025. 2, 3, 4, 5, 6 +[37] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2 +[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 3, 4 +[39] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Heng- + +wei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2 +[40] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 1, 2 +[41] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623-13633, 2023. 2 +[42] Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. Understanding negative sampling in graph representation learning. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1666-1676, 2020. 5 +[43] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 7, 8 +[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 2 +[45] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3, 4, 7, 8 \ No newline at end of file diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/images.zip b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f3ab3198dd3951e77d866b339b4c69a60b0d85e --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:369a5f7c4e3a1092b5066e688ad582476cd6619f93444b8447d10228bcf9dfab +size 474965 diff --git a/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/layout.json b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..10def3136b5cd2d2ceec3295480f98da9662ceb9 --- /dev/null +++ b/2025/Sparse-Dense Side-Tuner for efficient Video Temporal Grounding/layout.json @@ -0,0 +1,9354 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 122 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 122 + ], + "type": "text", + "content": "Sparse-Dense Side-Tuner for efficient Video Temporal Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 479, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 479, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 479, + 185 + ], + "type": "text", + "content": "David Pujol-Perich, Sergio Escalera, Albert Clapés \nUniversitat de Barcelona and Computer Vision Center, Barcelona, Spain \n{david.pujolperich, sescalera, aclapes}@ub.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 297, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 524 + ], + "type": "text", + "content": "Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full finetuning is often impractical, parameter-efficient fine-tuning – and particularly side-tuning (ST)– has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention –a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a " + }, + { + "bbox": [ + 55, + 239, + 297, + 524 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 55, + 239, + 297, + 524 + ], + "type": "text", + "content": " the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 676 + ], + "type": "text", + "content": "Recently the field of Video Understanding has gained attention due to its potential in applications like search engines or recommendation systems. A key task in this field is Video Temporal Grounding (VTG), which localizes moments within videos based on textual descriptions. Typically, this involves doing both Moment Retrieval (MR) [2, 5, 7, 11, 15] and Highlight Detection (HD) [1, 8, 10, 33, 40]. More specifically, MR predicts moment boundaries w.r.t. textual queries, while HD offers a more interpretable perspective, predicting frame-level saliency scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In the past, existing works focused on either MR or HD, given the lack of datasets that combined both tasks. QVHighlights [15] shifted this paradigm, proposing the first" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 212, + 434, + 308 + ], + "blocks": [ + { + "bbox": [ + 315, + 212, + 434, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 212, + 434, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 434, + 308 + ], + "type": "image", + "image_path": "245f34659e271e7f11ba68cd7ae960c6512b826dafc8cbc103298dea05a31ace.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 311, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 376 + ], + "type": "text", + "content": "Figure 1. Comparison of our proposed method and the main MR (left) and HD (right) baselines. These are all evaluated on QVHighlights val split and use InternVideo2-1B features. These results show that our method improves existing side-tuning works and attains SOTA results while significantly reducing the number of trainable parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 213, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 434, + 213, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 434, + 213, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 434, + 213, + 553, + 308 + ], + "type": "image", + "image_path": "fa0fc1c28d03ed50e28f05ad4f339fc3ab5441bfbbe68f19b090b04042d9d9ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": "dataset suitable for this multi-task setup. Since then, various works have been proposed following anchor-based [2, 17, 21, 22], anchor-free [7, 15, 25] and even LLM-based [23, 24] approaches. All of these, however, incur in a common critical limitation, which is relying on the final-layer features of frozen backbones. This is especially limiting when facing large distribution shifts between the pre-training distribution and that of the downstream task [29]. This issue is particularly pronounced in VTG, where an image-domain backbone [31] is transferred to the video domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "A typical solution is fine-tuning; however, full fine-tuning is impractical due to its high computational cost. Doing this more efficiently has thus become critical, motivating the interest in parameter-efficient fine-tuning (PEFT) [9] methods -e.g., Prompting or Adapters- which optimize only a small subset of parameters. Unfortunately, these still require full back-propagation through the backbone, making it still memory-intensive. To address this, [34] introduced side-tuning (ST), an effective PEFT, and memory-efficient fine-tuning (MEFT) method. ST creates a parallel pathway to refine intermediate features, thus backpropagating over a minimal set of parameters only. In this regard, we highlight the work of " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "-Tuning [21], the first ST approach for VTG, which recursively fuses multi-modal intermediate CLIP [31] embeddings. [21] also proposes adapting these frame-level embeddings for MR by generating a dense set of anchors. As shown in Fig. 1, this" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21515" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "method proves ineffective for MR - a highly sparse task where videos may contain very few ground-truth action." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 295, + 383 + ], + "type": "text", + "content": "This motivates our dual-stream Sparse-Dense Side-Tuner (SDST), the first proposal-free ST method, carefully designed for multi-task learning across sparse (MR) and dense (HD) tasks. For this, SDST jointly refines frame-level embeddings -suitable for HD- while learning what we call the recurrent decoder queries for MR. We also identify an implicit contextual limitation of the deformable attention [45] module of anchor-free architectures like ours, which in our task results in collapsing offsets around their initialization values. This also impedes the selection of keys outside of the current estimated moment boundaries, critical to potentially refine the boundaries of longer moments. Consequently, we propose the alternative Reference-based Deformable Self-Attention (RDSA), which naturally addresses this issue by reformulating the CA into a SA-based mechanism. Finally, we tackle the performance degradation of ST methods derived from the use of image-based CLIP, over more advanced spatio-temporal VLMs [36], as noted in [7]. A key challenge of exploiting [36] is, however, defining an effective token pooling strategy, as naive strategies like CLS pooling yield significant performance drops. This motivates our proposed module re-utilization scheme, which results in the first successful integration of InternVideo2 [36] into an ST framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 384, + 295, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 600 + ], + "type": "text", + "content": "In summary, our main contributions are threefold: 1) We propose SDST, the first anchor-free ST architecture, specifically tailored for complex sparse-dense multi-task setups like VTG. Our method significantly outperforms existing ST architectures on QVHighlights [15], TACoS [32] and Charades-STA [5] (see Fig. 1). SDST also performs competitively and even surpasses existing SOTA while reducing its learnable parameter count by up to " + }, + { + "bbox": [ + 55, + 384, + 295, + 600 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 55, + 384, + 295, + 600 + ], + "type": "text", + "content": ", and incurring in a minimal memory overhead. 2) We identify a key contextual limitation of the deformable attention mechanism that, in our task, results in collapsing offsets. Consequently, we propose RDSA, which tackles this issue by allowing more complex key-selection strategies. 3) We address the limitations of transfer learning from an image-domain backbone to a video domain by integrating, for the first time, the more advanced spatio-temporal backbone [36] into an ST framework. This proves to be a non-trivial challenge with massive performance implications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 139, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 139, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 139, + 621 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Video temporal grounding (VTG): One relevant task in Video Understanding is VTG [17] which aims to identify actions from text queries. Traditionally, this task was approached either from a MR [2, 5, 7, 11, 15, 25, 26] or HD [1, 8, 10, 33, 40] perspective. MR focuses on predicting specific action proposals while HD aims to compute the saliency scores for each of the frames w.r.t. the query. Nev" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 468 + ], + "type": "text", + "content": "ertheless, after the proposal of QVHighlights [15], the literature started approaching this from a multi-task learning perspective. In this regard, arguably the most notable family of methods is that based on DETR [7, 14, 15, 25, 26, 39] given the sparse nature of MR. These methods normally fuse both video and textual modality into a final embedding which is used for HD and as a memory of a Transformer decoder that refines a set of learnable queries. More in detail, Moment-DETR [15] constitutes the first baseline, which uses a standard Transformer encoder to process both modalities simultaneously. QD-DETR [26] proposes instead to inject the textual information to the video modality with cross attention (CA), followed by a temporal modeling module. CG-DETR [25] reduces the negative impact of irrelevant textual tokens by adding an Adaptive CA module that leverages dummy tokens to redirect attention weight. In this same line, SG-DETR [7] introduces a saliency-guided CA, weighting the contribution of irrelevant textual tokens based on the computed saliency scores. Despite the notable success of these proposal-free methods, one core limitation that they all share is that they rely solely on the last-layer features of a frozen CLIP [31] and/or Slowfast [4] backbones. [41] overcomes this limitation via the full finetuning of CLIP, even though this becomes nearly intractable given the resources that this requires. " + }, + { + "bbox": [ + 313, + 72, + 555, + 468 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 468 + ], + "type": "text", + "content": "-Tuning [21], the most similar method to our work, overcomes this limitation by proposing the first PEFT and MEFT method for VTG, which recursively applies a multi-modal fusion ST. In this work, we argue that one of its core drawbacks, however, is its anchor-based nature, remaining oblivious to the shown benefits of DETR-based architectures in very sparse tasks like MR [7, 16, 20]. This motivates our proposed SDST, which incorporates the best of both PEFT and MEFT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Parameter-and-memory-efficient fine-tuning: Foundation models and VLMs [31, 37] have become the cornerstone of recent advances in Video Understanding applications. The use of pre-extracted features remains, however, an important limitation to its application on downstream tasks like VTG [17]. Fine-tuning these huge models is often simply infeasible given the large resources that these require. This has motivated the rise of PEFT [9], which aims to democratize the fine-tuning process by restricting the tunable parameters to the bare minimum. In this regard, some of the most prominent approaches are Prompting [12, 44] and the use of Adapters [6, 28]. The first keeps the backbone frozen while learning prompts that are appended to the input, bridging the gap between the backbone's expected distribution and that of the downstream task. Adapters, in contrast, incorporate small trainable modules in the backbone while keeping the rest unchanged. These methods, however, still require full backpropagation through the frozen model, making them memory-inefficient. Recently, the new paradigm of " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "ST" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " [19, 30, 34] –and particularly [21]" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21516" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "for VTG- has gained attraction creating a parallel pathway to exploit intermediate backbone representations while guaranteeing that backpropagation is only applied to this small parallel module. Interestingly, these methods normally rely on pre-extracted CLIP representations for tasks like HD/MR. As shown by [7], this is a very limiting aspect given the lack of temporal reasoning of CLIP or its difficulties in understanding textual queries beyond simple spatial descriptions. For this reason, in this work, we propose a novel ST architecture that, for the first time, relies on the more advanced InternVideo2 [36] backbone. Unlike CLIP, this backbone is trained on video-domain inputs, resulting in enhanced spatial and temporal modeling capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 456 + ], + "type": "text", + "content": "Deformable attention: A key component of most SOTA methods for VTG like SG-DETR [7] is their deformable attention module [38, 45]. This module addresses the well-studied slow convergence of DETR [3]. It does so by limiting the selectable keys based on the predictions of the learnable queries, completely decoupling both query and key spaces. Despite desirable in terms of efficiency, it has been previously noted in the literature [43] that the lack of context of the queries w.r.t. the key/value space can lead to suboptimal attendable key selection. VTG methods often mitigate this effect by partially or completely initializing the learnable queries based on the DETR-memory [43] -i.e., frame-level representations in our case. In this work, we empirically demonstrate the limitations of these initialization-based methods for VTG tasks, which motivates our proposed simple yet effective alternative, the Reference-based Deformable Self-Attention (RDSA). This reformulates the deformable CA into a deformable SA, naturally solving the aforementioned issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 461, + 111, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 111, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 111, + 474 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 165, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 165, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 165, + 493 + ], + "type": "text", + "content": "3.1. Problem definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "In this paper, we address the problem of MR and HD based on textual descriptions. For this, we consider an arbitrary input-query pair " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "(\\mathbf{X}^v,\\mathbf{X}^t)" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^v\\in \\mathbb{R}^{T\\times H\\times W\\times 3}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t\\in \\mathbb{R}^{L\\times F_e}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " are the number of frames and tokens, and " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "F_{e}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " the textual encoding dimension. Our goal is to predict, for the given video-query pair, its respective frame-wise saliency scores " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^s\\in \\mathbb{R}^T" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " for HD, and the " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " action moments " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^m\\in \\mathbb{R}^{M\\times 2}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " for MR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 123, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 123, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 123, + 612 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "In our work we propose a dual-stream ST architecture that includes a dense (frame-level) and a sparse (segment-level) stream for MR and HD respectively (see Fig. 2a). More specifically, we first leverage a frozen InternVideo2-1B [36] backbone, a model that possesses powerful spatio-temporal modeling capabilities, to extract " + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": " intermediate visual-textual representations. These are then processed by multiple weight-sharing SDST layers, which refine dense rep" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "resentations and recurrent decoder queries. More in detail, at each level, the model refines the dense embeddings via multi-modal textual conditioning as well as temporal modeling. This signal is then used to condition the sparse stream " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ", that builds on [16] while incorporating our novel deformable attention mechanism, namely Reference-based Deformable Self-Attention (see Fig. 2b). This mitigates the shortcomings of the deformable CA mechanism, enhancing the contextual information of the decoder queries and thus improving the quality of the selected keys. After unrolling this process " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " times, we apply the dense and the sparse prediction heads to compute the saliency scores and the predicted segment boundaries, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 493, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 493, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 493, + 255 + ], + "type": "text", + "content": "3.3. Sparse-Dense Side-Tuner (SDST)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 262, + 505, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 505, + 274 + ], + "type": "text", + "content": "3.3.1. Recurrent refinement with side-tuners" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": "Our work builds on an ST framework, which first requires extracting " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " intermediate video and textual features " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}\\in" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{K\\times T\\times F_v}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}\\in \\mathbb{R}^{K\\times L\\times F_t}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " (find the details in Sec. 4). Here " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " are their respective dimensionalities. We then define a zero-initialized dense embedding " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^0\\in \\mathbb{R}^{T\\times F}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " learnable recurrent moment proposals inspired by [20], which we refer to as recurrent decoder queries. The queries include the learnable center-width moment references " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^0\\in \\mathbb{R}^{M\\times 2}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": ", and their corresponding latent embeddings " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^0\\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": ". We then define the recurrence for a given level " + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "inline_equation", + "content": "1\\leq \\ell \\leq K" + }, + { + "bbox": [ + 313, + 280, + 555, + 411 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 413, + 553, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 413, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 321, + 413, + 553, + 429 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {\\ell + 1}, \\mathbf {R} ^ {\\ell + 1}, \\mathbf {H} ^ {\\ell + 1} = S D S T \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\tilde {\\mathbf {V}} ^ {\\ell}, \\tilde {\\mathbf {T}} ^ {\\ell}\\right), \\tag {1}", + "image_path": "57bcc005bd43cf498049eac8461ced25b3342ae8e5eb9c3735f82836ef956037.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "content": "where SDST represents the shared-weight layers of our model, progressively refining the representations across two streams: the 1) dense learning stream " + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "content": " and the 2) sparse learning stream " + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "content": ", that we describe below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 493, + 449, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 449, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 449, + 505 + ], + "type": "text", + "content": "3.3.2. Dense learning stream " + }, + { + "bbox": [ + 313, + 493, + 449, + 505 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": "Following [21], this stream first refines the dense embeddings " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\ell}" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": " by conditioning it to the video-textual embeddings " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}^{\\ell}" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{T}}^{\\ell}" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": ". For this, both of these multi-modal embeddings are projected to a shared " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": "-dimensional space, using two MLPs " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "(\\mathcal{F}_v:F_v\\to F" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t:F_t\\to F)" + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 572, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 572, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 369, + 572, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\mathbf {V} ^ {\\ell} = \\mathcal {F} _ {v} \\left(\\tilde {\\mathbf {V}} ^ {\\ell}\\right), \\mathbf {T} ^ {\\ell} = \\mathcal {F} _ {t} \\left(\\tilde {\\mathbf {T}} ^ {\\ell}\\right). \\tag {2}", + "image_path": "c53173c2c143a383ee6a1176e04529a37ebe7236fa64913532ce898be49201d1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "text", + "content": "Then we incorporate the clip-wise visual information to " + }, + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\ell}" + }, + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "text", + "content": " via a weighted sum modulated by " + }, + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "inline_equation", + "content": "\\beta^{\\ell} \\in [0,1]" + }, + { + "bbox": [ + 313, + 590, + 554, + 626 + ], + "type": "text", + "content": ", a zero-initialized layer-dependent parameter that yields:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {\\ell} := \\beta^ {\\ell} \\mathbf {D} ^ {\\ell} + (1 - \\beta^ {\\ell}) \\mathbf {V} ^ {\\ell}. \\tag {3}", + "image_path": "082acb4ce4f27f7fb509ec2e998e96ee112063f1f7f8cb76f8d3a6aa698ff402.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "content": "Thereafter, we inject the textual information with CA, where " + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{\\ell}" + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "content": " does not include the CLS token. We also apply a temporal modeling module defined as a SA and a subsequent point-wise feed forward network (PFFN):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 696, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 696, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 347, + 696, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {\\ell + 1} = P F F N \\left(S A \\left(C A \\left(\\mathbf {D} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right)\\right), \\tag {4}", + "image_path": "b178e0d4b050c9a7fbe324ff420879635e1124e7473493604d535a18c2b99554.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21517" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 370, + 186 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 370, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 370, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 370, + 186 + ], + "type": "image", + "image_path": "70f0a1d87307015aebb2100906a00ad80a96633368a9a88475129bdd570cf72e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 187, + 282, + 196 + ], + "lines": [ + { + "bbox": [ + 145, + 187, + 282, + 196 + ], + "spans": [ + { + "bbox": [ + 145, + 187, + 282, + 196 + ], + "type": "text", + "content": "(a) Overview of our proposed architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 385, + 82, + 550, + 186 + ], + "blocks": [ + { + "bbox": [ + 385, + 82, + 550, + 186 + ], + "lines": [ + { + "bbox": [ + 385, + 82, + 550, + 186 + ], + "spans": [ + { + "bbox": [ + 385, + 82, + 550, + 186 + ], + "type": "image", + "image_path": "d94d93dd84419ed993f4e092c9c72f14e9c5a29a2ee0cf177334e97d4a0dd41f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 187, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 398, + 187, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 398, + 187, + 539, + 196 + ], + "type": "text", + "content": "(b) RDSA vs Def.CA[38] for a proposal " + }, + { + "bbox": [ + 398, + 187, + 539, + 196 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 398, + 187, + 539, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. Our method (left) first processes the video and textual inputs using [36], and then recursively applies a shared SDST, a recurrent dual-stream model attached to the last " + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "text", + "content": " layers. This iteratively refines dense embeddings and learnable recurrent decoder queries. The SDST consists of a dense stream " + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "text", + "content": " for temporal and multi-modal dense refinement, and a sparse stream " + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "text", + "content": " that refines the recurrent queries conditioned on the dense signal. " + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 54, + 200, + 555, + 266 + ], + "type": "text", + "content": " includes the RDSA module (right), a context enhanced deformable attention mechanism that restricts the selectable keys based on the center, left-, and right-most action embeddings. The final dense embedding solves HD, while the recurrent queries address MR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 192, + 285 + ], + "type": "text", + "content": "3.3.3. Sparse learning stream " + }, + { + "bbox": [ + 55, + 273, + 192, + 285 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "content": "This stream can be seen as a recurrent DETR-based mechanism that exploits the complementarity of both sparse (MR) and dense (HD) tasks to refine the recurrent decoder queries. This is, refining the center-width references " + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}" + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "content": " and their latent embeddings " + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{\\ell}\\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "content": ". Similar to the stream " + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "content": ", we first incorporate a CA module that conditions " + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{\\ell}" + }, + { + "bbox": [ + 55, + 289, + 296, + 396 + ], + "type": "text", + "content": " to the textual query. This is followed by a SA module to enable the information flow between different moment proposals:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 399, + 294, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 399, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 113, + 399, + 294, + 414 + ], + "type": "interline_equation", + "content": "\\mathbf {H} ^ {\\ell} = S A \\left(C A \\left(\\mathbf {H} ^ {\\ell}, \\mathbf {T} ^ {\\ell}, \\mathbf {T} ^ {\\ell}\\right)\\right). \\tag {5}", + "image_path": "c09b548a0b948d32c4273e7fb360fbde1f30c8462a1b7d893aa16bd1a3dce49b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 417, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 417, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 417, + 296, + 584 + ], + "type": "text", + "content": "Another essential aspect of this stream is effectively conditioning the recurrent decoder queries to the video modality. Existing works [7, 16, 20] typically rely on the use of deformable attention due to its convergence and performance benefits [38, 45]. This mechanism leverages attention queries to predict offsets that define a limited set of selectable keys. However, in this task we observe that 1) these offsets collapse near their initialization and 2) are unable to look beyond the current estimated boundaries (see Sec. 6.2). We attribute this to contextual limitations, which motivates our proposed alternative Reference-based Deformable Self-Attention (RDSA), enabling an enhanced cross-modality injection (see Sec. 3.3.4). We also incorporate a simple PFFN to enhance the model's expressivity:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 586, + 294, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 586, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 91, + 586, + 294, + 601 + ], + "type": "interline_equation", + "content": "\\mathbf {H} ^ {\\ell} = P F F N \\left(R D S A \\left(\\mathbf {R} ^ {\\ell}, \\mathbf {H} ^ {\\ell}, \\mathbf {D} ^ {\\ell + 1}\\right)\\right). \\tag {6}", + "image_path": "6e434cc37ff28791a37838a71973c1f69f18edcee1cf6703be3c930649609294.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 604, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 640 + ], + "type": "text", + "content": "Importantly, all of these modules are shared across the " + }, + { + "bbox": [ + 55, + 604, + 295, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 604, + 295, + 640 + ], + "type": "text", + "content": " levels for parameter efficiency, and include residual connections and DropPath [13] for stability and regularization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 650, + 263, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 263, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 263, + 661 + ], + "type": "text", + "content": "3.3.4. Reference-based deformable self-attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Deformable attention in the context of existing anchor-free methods: One of the keys to the success of recent anchor-free methods like DETR -including works tackling VTG [7]- is the use of deformable attention mecha" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 273, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 344 + ], + "type": "text", + "content": "nisms [45] for cross-modal interaction. This mechanism is known to provide faster convergence speed, reduced time complexity, and as shown in Sec. K, overall better performance than standard CA. As depicted in Fig. 2b (down), the key of this module is the lack of explicit interaction between queries and keys, which are defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 363, + 346, + 553, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 346, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 363, + 346, + 553, + 363 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\mathbf {K} = \\mathbf {X} _ {\\mathcal {K}} \\mathbf {W} _ {\\mathcal {K}} ^ {d e f}, \\tag {7}", + "image_path": "7ee3fca8173d217fab44849d30591d9d477b905a162040257862bcd58a71c028.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathcal{Q}}^{def}" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathcal{K}}^{def}" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": " are two linear projections. The deformable attention thus avoids computing the " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^T" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": " similarity matrix, and instead employs an offset and attention-score predictors, " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\Delta}" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": " to select and weight -only based on the queries - a small subset " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": " of selectable keys. The output of this mechanism, the weighted aggregation of the selected keys, namely " + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 365, + 554, + 453 + ], + "type": "text", + "content": ", is computed as follows" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 455, + 553, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 455, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 330, + 455, + 553, + 470 + ], + "type": "interline_equation", + "content": "\\Delta = \\mathcal {G} _ {\\Delta} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\mathbf {A} = \\mathcal {G} _ {\\mathbf {A}} (\\mathbf {Q}) \\in \\mathbb {R} ^ {M \\times P}, \\tag {8}", + "image_path": "cbc538709a5e04988be2207f76bc78788b6ff3cbd79ab924b0e532098903db28.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 472, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 472, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 350, + 472, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\sum_ {p = 1} ^ {P} \\left(\\mathbf {A} _ {:, p} \\mathbf {K} [ \\mathbf {c} + \\mathbf {w} \\odot \\Delta_ {:, p} ]\\right) \\in \\mathbb {R} ^ {M \\times F} \\tag {9}", + "image_path": "82b67501814279d7023265b8d223965b85d426b4841c51875410b292d01cbf1d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\Delta}" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": " are often modeled as CNNs and " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "x[y]" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": " is the bilinear sampling of " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": " on index(es) " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": ". Notably, in our 2-dimensional setup, the offsets " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\Delta_{:,p}" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": " are added to center reference " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\mathbf{R}_{:,0}^{\\ell} \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": ", and weighted over the action width " + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = \\mathbf{R}_{:,1}^{\\ell} \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 509, + 554, + 571 + ], + "type": "text", + "content": ". See Sec. K for additional details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Limitations: Interestingly, in the literature we find that when it comes to the deformable attention, most works operate with a certain independence to whether they deal with a deformable self-attention or cross-attention, assuming that the benefits of this mechanism extrapolate to both scenarios. In this work, however, we find that in fact, the deformable attention is only naturally suited for self-attention scenarios as its effectiveness inherently relies on implicit interactions between " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " -not present in the cross-attention case. Concretely, in a self-attention setup, " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathcal{Q}} = \\mathbf{X}_{\\mathcal{K}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " (see Eq. 7). Thus, modeling " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\Delta}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " as CNNs naturally al" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21518" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 119 + ], + "type": "text", + "content": "lows queries to gain context of the local neighborhood of queries, and consequently, of the key/value space. This breaks their independence assumption, giving the queries critical information to decide where to look." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "text", + "content": "Importantly, this is not the case with the deformable cross attention where we try to inject cross-modal information from the keys to the queries, which is our goal in this work. In this scenario, " + }, + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathcal{Q}} \\neq \\mathbf{X}_{\\mathcal{K}}" + }, + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "text", + "content": ", and thus, using CNN-based offset predictors does not inject knowledge of the key/value space, making its guesses blind or based on the blind predictions of previous iterations. Concretely, in Sec. 6.2 we show that in our task, the drawbacks of this naive deformable cross-attention have various empirical implications such as the offset collapse near the offset initialization values, or the inability to select keys from frames beyond the estimated action boundaries, critical to refine segments into longer actions. This motivates our proposed Reference-based Deformable Self-Attention, an alternative mechanism that allows reformulating the deformable CA module into a deformable SA based on the references " + }, + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{\\ell}" + }, + { + "bbox": [ + 57, + 121, + 297, + 312 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": "Our proposed alternative (RDSA): Our proposed RDSA (see Fig. 2b) takes as input the center-width references " + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^{\\ell}\\in \\mathbb{R}^{M\\times 2}" + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": " and the dense embeddings " + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\ell}\\in \\mathbb{R}^{T\\times F}" + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": ". Then, differently from other methods that leverage the learnable queries " + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{H}^{\\ell}" + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": " to compute the offsets and attention scores (see Eq. 8), we propose using an alternative query embedding. Concretely, we first refine the dense embedding " + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_l" + }, + { + "bbox": [ + 55, + 312, + 296, + 467 + ], + "type": "text", + "content": " with a simple CNN, allowing frames to gain local context of their surroundings. Then, we use bilinear sampling to extract three key action embeddings: left-most (l), center (c), and right-most (r). The center provides highly informative action features, while the extremes help refine the moment boundaries. This yields," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 469, + 294, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 469, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 77, + 469, + 294, + 485 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Q}} = \\hat {\\mathbf {X}} _ {\\mathcal {Q}} \\mathbf {W} _ {\\mathcal {Q}} ^ {d e f}, \\hat {\\mathbf {X}} _ {\\mathcal {Q}} = C N N (\\mathbf {D} ^ {\\ell}) [ l, c, r ]. \\tag {10}", + "image_path": "4f265684ce129087d51ae65e7fafe56fa030953761993261fb3da634b1900813.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 238, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 238, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 238, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 487, + 238, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathcal{Q}}\\in \\mathbb{R}^{M\\times 3F}" + }, + { + "bbox": [ + 55, + 487, + 238, + 501 + ], + "type": "text", + "content": ". This rewrites Eq. 8 to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 503, + 294, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 503, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 118, + 503, + 294, + 517 + ], + "type": "interline_equation", + "content": "\\hat {\\Delta} = \\mathcal {G} _ {\\Delta} (\\hat {\\mathbf {Q}}), \\hat {\\mathbf {A}} = \\mathcal {G} _ {\\mathbf {A}} (\\hat {\\mathbf {Q}}). \\tag {11}", + "image_path": "7aa7bccb4e91ae6fc379420b4efa548b63ccff2b8d158060a7790a360b388540.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "Finally, we apply the deformable attention from Eq. 9 using the new offsets and attention scores " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\Delta}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": ". This can be viewed as asking the key spots of the action where to look based on the textual information that we previously injected in Eq. 5. This naturally solves the context limitation of the standard deformable CA given that " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathcal{Q}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathcal{K}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " now live in the same latent space -as they are both derived from " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{\\ell}" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": ". This allows the model to make informed decisions about which keys to attend to while keeping all its core computational advantages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 266, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 266, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 266, + 661 + ], + "type": "text", + "content": "3.4. Prediction heads and training objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "After recursively applying the SDST over " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " intermediate layers, we compute the respective saliency scores and action segment proposals, as well as define their corresponding objective functions. We briefly describe these below," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 531, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 531, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 531, + 83 + ], + "type": "text", + "content": "but refer interested readers to Sec. C for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "text", + "content": "Highlight Detection: To solve HD, we define the saliency scores as the cosine similarity between the dense embedding " + }, + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^K" + }, + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "text", + "content": " and a pooled representation of " + }, + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^K" + }, + { + "bbox": [ + 313, + 85, + 553, + 133 + ], + "type": "text", + "content": ". We then apply an InfoNCE loss [27] that learns a ranking of these scores" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 137, + 553, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 137, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 386, + 137, + 553, + 149 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {H D} = \\lambda_ {0} \\mathcal {L} _ {\\text {I n f o N C E}}. \\tag {12}", + "image_path": "1604f22360b0e628d08d950062c2a735730a940eee1fdf154f300133c56dcf51.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "content": "Moment Retrieval: To solve MR we first apply the Hungarian algorithm to obtain a one-to-one matching between the predicted queries and the GT actions. We then project " + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^{\\ell}" + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "content": " to compute the action probability " + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{M}" + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "content": " different queries, which we train with a FocalLoss [18]. We also compute the moment boundaries " + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}^{m} \\in \\mathbb{R}^{M \\times 2}" + }, + { + "bbox": [ + 313, + 152, + 555, + 260 + ], + "type": "text", + "content": " which we optimize using an L1 and an IoU loss. Finally, we use an L1 loss to learn an actionness score, which predicts the maximum IoU of each query w.r.t. the GT actions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 264, + 553, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 323, + 264, + 553, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {M R} = \\lambda_ {1} \\mathcal {L} _ {a c t} + \\sum_ {l = 1} ^ {\\ell} \\lambda_ {2} \\mathcal {L} _ {c l s} ^ {\\ell} + \\lambda_ {3} \\mathcal {L} _ {l 1} ^ {\\ell} + \\lambda_ {4} \\mathcal {L} _ {I o U} ^ {\\ell}. \\tag {13}", + "image_path": "79be4195466f5402813f78e1cb26d762d1f7f65f913c34b820718b2c5f2036a8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 299, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 553, + 323 + ], + "type": "text", + "content": "Note that all but the " + }, + { + "bbox": [ + 313, + 299, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{act}" + }, + { + "bbox": [ + 313, + 299, + 553, + 323 + ], + "type": "text", + "content": " term are optimized across different refinement levels to promote a faster convergence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "content": "Alignment losses: To improve the video-textual alignment, we define " + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {align }}" + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "content": " based on two SampledNCE losses [42] to align " + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 323, + 554, + 407 + ], + "type": "text", + "content": " embeddings across the different refinement levels. We apply this along both the batch and the intermediate-layer dimensions. This pulls the action embeddings closer to the textual representations while learning complementary information across intermediate layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 470, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 470, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 470, + 418 + ], + "type": "text", + "content": "Final loss: The final loss is defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 424, + 553, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 424, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 359, + 424, + 553, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {5} \\mathcal {L} _ {H D} + \\lambda_ {6} \\mathcal {L} _ {M R} + \\lambda_ {7} \\mathcal {L} _ {\\text {a l i g n}}. \\tag {14}", + "image_path": "d70eb1a794bb3e8c156c1914d458f48a5feb7df14e7fa5b55e12e91610ed6e30.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 445, + 555, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 471 + ], + "type": "text", + "content": "4. Extracting intermediate representations of InternVideo2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 479, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 624 + ], + "type": "text", + "content": "The cornerstone of ST is leveraging intermediate multimodal representations from frozen VLMs. While previous ST approaches predominantly rely on CLIP, we argue that CLIP has notable limitations in temporal modeling and aligning visual features with textual queries beyond simple static descriptions. To overcome these limitations, we incorporate InternVideo2 [36] into our framework, leveraging its advanced spatio-temporal modeling capabilities [7]. However, extracting intermediate representations from InternVideo2 –particularly from its visual encoder " + }, + { + "bbox": [ + 313, + 479, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 479, + 554, + 624 + ], + "type": "text", + "content": " – remains a challenging and underexplored task in the literature, despite its significant impact on performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": "In particular, the main challenge arises from how to pool the spatio-temporal token dimension " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{v}" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times L_v \\times F}" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": ", the output embedding of the " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": "-th layer of " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": " into the final intermediate features " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}^{\\ell} \\in \\mathbb{R}^{T \\times F}" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": " (see Eq. 1). Prior CLIP-based methods simply use the CLS token as \"summaries\" over " + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{v}" + }, + { + "bbox": [ + 313, + 624, + 555, + 713 + ], + "type": "text", + "content": ". However, we find this strategy to be suboptimal for InternVideo2 (see Sec. 6.1) as this limits its" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21519" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": "spatial-aggregation capabilities. Notably, [36] optimizes an AdaptivePool module that computes the final-layer embeddings, which are then aligned with the text. We conjecture that leveraging this enhanced multi-modal alignment is key to performance. Unfortunately, such pooled representation is only computed at the last layer, leaving the question of how to pool the remaining intermediate layers. Ideally, we would optimize layer-independent AdaptivePool modules, but this would require full back-propagation through the entire model. A more efficient alternative would be computing gradients of the AdaptivePool modules only. This is, however, still computationally infeasible, not for the gradients themselves, but because of the need to load " + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "inline_equation", + "content": "L_{v}" + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": " spatiotemporal tokens per frame. This increase in the input memory size translates, for instance, to a " + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": " memory increase for QVHighlights. We thus hypothesize that re-using this frozen pooling module across the " + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": " intermediate layers allows a memory efficient training –no additional backpropagation nor memory requirements are needed– while leveraging its enhanced pooling capabilities, even despite their respective distribution shifts. Formally, we compute" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 326, + 294, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 326, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 64, + 326, + 294, + 346 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {V}} ^ {\\ell} = \\text {A d a p t i v e P o o l} \\left(\\hat {\\mathbf {V}} ^ {\\ell}\\right) \\in \\mathbb {R} ^ {T \\times F}, 1 \\leq \\ell \\leq K. \\tag {15}", + "image_path": "d668309bc5c0c52c5cfb3d5df8361df28a72e451c7c354c970d041f77f9081a2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 384 + ], + "type": "text", + "content": "As shown in Sec. 6.1, this pooling strategy considerably improves other alternatives, demonstrating the benefits of this module re-utilization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 157, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 157, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 157, + 403 + ], + "type": "text", + "content": "5. Experimentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 410, + 170, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 170, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 170, + 423 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 427, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 559 + ], + "type": "text", + "content": "We test our proposal on various datasets for MR and HD, namely QVHighlights [15], TACoS [32], and the Charades-STA [5] dataset. Further details of the datasets can be found in Sec. B. For the task of MR on QVHighlights, we compute Recall@1 with two IoU thresholds, 0.5 and 0.7, and the mean average precision (mAP) at IoU thresholds between 0.5 and 0.95 with a step of 0.05 -i.e., [0.5:0.05:0.95]. For HD, we report mAP and HIT@1 over the positive frames -i.e., the most salient ones classified as \"VeryGood\". For TACoS and Charades-STA, we compute Recall@1 at 0.3, 0.5, and 0.7 IoU thresholds, as well as mIoU." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 201, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 201, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 201, + 577 + ], + "type": "text", + "content": "5.2. Main experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "We begin by comparing the performance of SDST over various relevant baselines. Concretely, observe in Tab. 1 the evaluation of our method on the test and val splits of QVHighlights. These results indicate that our method considerably outperforms " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "-Tuning. Similarly, SDST performs very competitively and even surpasses on several metrics the current SOTA method, namely SG-DETR [7]. Note however that our method attains this performance while having only a " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " of the parameter count of SG-DETR. We refer interested readers to Sec. M for the statistical significance analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "Most of the presented methods in Tab. 1 rely on alternative backbones different from InternVideo2 -e.g., CLIP or Slowfast. To establish a more fair comparison, in Fig. 1 we follow the work of [7] and evaluate a set of relevant baselines when using InternVideo2 features -find the complete ablation in Sec. E. Observe that our work performs on par with the SG-DETR and even more importantly to our work, the SDST very considerably outperforms " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "-Tuning, improving the MR performance by a " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "3.82\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " average mAP or the HD by a " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "2.21\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " mAP. This empirically demonstrates the benefits of leveraging a proposal-free architecture over a more rigid anchor-based method. Also find in Sec. H an extended comparison with other ST works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": "To complement our findings, in Tab. 2 we evaluate our method on the Charades-STA and TACoS benchmarks. Observe that our method obtains SOTA in both of these datasets. In Charades-STA, for instance, improving by " + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": " of R1@0.7 and " + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "inline_equation", + "content": "2.06\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": " of mIoU, the existing SOTA. Similarly, in TACoS, SDST improves SG-DETR by " + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": " R1@0.7 and " + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "inline_equation", + "content": "1.27\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 311 + ], + "type": "text", + "content": " mIoU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 320, + 411, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 411, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 411, + 333 + ], + "type": "text", + "content": "6. Ablation studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 389 + ], + "type": "text", + "content": "In this section, we ablate over various relevant aspects of our main contributions. Note that unless states otherwise, we follow the literature and evaluate them on the val split of QVHighlights. Also, we write write in bold the best results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 527, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 527, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 527, + 409 + ], + "type": "text", + "content": "6.1. Leveraging InternVideo2 features for ST" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "text", + "content": "1) Does pooling matter? As argued in Sec. 4, one of the main challenges in using the InternVideo2 backbone for ST is pooling the spatio-temporal clip-wise intermediate representations. In Tab. 3 we compare different pooling strategies. Concretely, we compare the standard CLS-pooling, average-pooling, and our proposed re-utilization of the frozen AdaptivePool (Sec. 4). Observe that CLS-pooling results in a considerable performance degradation of up to " + }, + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "inline_equation", + "content": "5.07\\%" + }, + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "text", + "content": " average mAP on MR or " + }, + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "inline_equation", + "content": "7.74\\%" + }, + { + "bbox": [ + 313, + 414, + 553, + 570 + ], + "type": "text", + "content": " HIT@1 on HD w.r.t. the adaptive pooling strategy. The average pooling partially mitigates this degradation, but still obtains a considerably decreased performance. This confirms the significant impact of a carefully chosen pooling strategy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "content": "2) Feature refinement vs. feature sampling: One of the most critical decisions of ST is determining the number of intermediate levels that should be used. Normally, doing so involves evaluating the performance when refining the last " + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "content": " intermediate layers [21]. As shown in Fig. 3 (red) this indicates that in our case, the best performing " + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "content": " is 4. The literature often suggests this argument to be compelling enough to claim that using intermediate representations is beneficial over using only last-layer features. In this work, however, we find it reasonable to wonder: Does this analysis depict the importance of intermediate features, or is it simply evidencing the need to do multiple refinement steps," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21520" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 70, + 525, + 287 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 525, + 287 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 525, + 287 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 525, + 287 + ], + "type": "table", + "html": "
Methodtestval#Params
MRHDMRHD
R1@0.5@0.7mAP@0.75Avg.≥ Very good mAP HIT@1R1@0.5@0.7@0.5mAP@0.75Avg.≥ Very good mAP HIT@1
BeautyThumb----14.3620.88--------
DVSE----18.7521.79--------
MCN11.412.7124.948.2210.67---------
CAL25.4911.5423.407.659.89---------
XML+46.6933.4647.8934.6734.9035.3855.06-------
Moment-DETR52.8933.0254.8229.4030.7335.6955.6053.9434.84-32.2035.6535.654.8M
UMT56.2341.1853.8337.0136.1238.1859.9960.2644.2656.7039.9038.5939.9064.20
MomentDiff58.2141.4854.5737.2136.84---------
QD-DETR62.4044.9862.5239.8839.8638.9462.4062.6846.6662.2341.8241.2239.1363.03
MH-DETR60.0542.4860.7538.1338.3838.2260.5160.8444.9060.7639.6439.2638.7761.74
UniVTG58.8640.8657.6035.5935.4738.2060.9659.74---36.1338.8061.8
TR-DETR64.6648.9663.9843.7342.6239.9163.4267.1051.4866.2746.4245.09--
CG-DETR65.4348.3864.5142.7742.8640.3366.2167.3552.0665.5745.7344.9340.8066.70
BAM-DETR62.7148.6464.5746.3345.36--65.1051.6165.4148.5647.61--
EaTR------61.3645.7961.8641.9141.7437.1558.659.0 M
Mr. BLIP74.7760.5168.1253.38--76.1363.3569.3955.78---19.0 M
LLaVA-MR76.5961.4869.4154.40--78.1364.1369.6456.32---17.0 M
HL-CLIP-----41.9470.60----42.3772.402.0 M
R²-Tuning68.0349.3569.0447.5646.1740.7564.2068.7152.06--47.5940.5964.32
SG-DETR†72.2056.6073.2055.8054.1043.7669.13--73.5257.9155.6443.9171.47
Flash-VTG†70.6953.9672.3353.8552.00--73.1057.2972.7554.3352.84--
Ours†70.8256.2371.3154.9953.3143.4069.1373.6860.9073.5257.4255.6044.0072.00
", + "image_path": "efee25b1a87dbe08762874ef64ea14de9764ad68561a9a3d44e19a64c32c6ca5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 318, + 266, + 437 + ], + "blocks": [ + { + "bbox": [ + 55, + 290, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 555, + 314 + ], + "type": "text", + "content": "Table 1. MR and HD results for QVHighlights test and val split.† indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best, considering only works supporting both MR and HD. The rest are marked in gray." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 318, + 266, + 437 + ], + "lines": [ + { + "bbox": [ + 87, + 318, + 266, + 437 + ], + "spans": [ + { + "bbox": [ + 87, + 318, + 266, + 437 + ], + "type": "table", + "html": "
MethodCharades-STTACoS
R@0.5R@0.7mIoUR@0.5R@0.7mIoU
M-DETR53.631.4-24.712.025.5
UMT48.329.3----
UniVTG58.035.750.135.017.433.6
QD-DETR57.332.6-36.821.135.8
CG-DETR58.436.350.139.622.236.5
BAM-DETR60.039.452.341.526.839.3
TR-DETR57.633.5----
MR_BLIP69.349.358.6---
LLaVA-MR70.649.659.8---
R2-Tuning59.837.050.938.725.135.9
SG-DETR†70.249.559.144.729.940.9
FlashVTG†70.349.9-41.824.737.6
Ours†72.052.661.244.532.342.2
", + "image_path": "6fc235b80c13d679b09c7d43a688cf8f481c19a71758e87388e434f7ea7a0db0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 475, + 288, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 438, + 295, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 472 + ], + "type": "text", + "content": "Table 2. Comparison on Charades-STA and TACoS datasets. " + }, + { + "bbox": [ + 55, + 438, + 295, + 472 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 438, + 295, + 472 + ], + "type": "text", + "content": " indicates the use of InternVideo2 features. Bold stands for the best and underline for the second-best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 475, + 288, + 521 + ], + "lines": [ + { + "bbox": [ + 63, + 475, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 63, + 475, + 288, + 521 + ], + "type": "table", + "html": "
Pool strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
CLS66.4552.6567.9651.3250.5341.0164.26
Avg. pool70.6556.970.4354.5853.4443.0669.68
Adapt. pool73.6860.9073.5257.4255.6044.0072.00
", + "image_path": "059358abaff7add68b3019bda699ae6a1dda0ef836fd988a840b98c0287bc867.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 194, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 194, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 194, + 555 + ], + "type": "text", + "content": "regardless of the chosen features?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": "To shed light on this issue, in Fig. 3 (blue) we evaluate the performance when doing " + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": " different refinement steps, but importantly, we always use the last-layer features only. Observe that for " + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": ", actually, this improves the homologous experiments with intermediate features. Differently, for " + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 54, + 557, + 295, + 688 + ], + "type": "text", + "content": ", the use of last-layer features harms the performance. These experiments thus suggest that proving the usefulness of intermediate features is not as straightforward as previously thought. In fact, these indicate that the advantages of using intermediate features arise only as we consider shallower layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "3) Why not sampling features from even shallower layers? As shown in Tab. 4, this is not necessarily helpful" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 318, + 434, + 390 + ], + "blocks": [ + { + "bbox": [ + 315, + 318, + 434, + 390 + ], + "lines": [ + { + "bbox": [ + 315, + 318, + 434, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 318, + 434, + 390 + ], + "type": "image", + "image_path": "3e3c9c98f97f220cfc053f685d964636bdbe6c1c2850c327052a92b3e5ff9093.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 3. Ablation of the number of refinement levels and of their use of intermediate or last-layer features only." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 317, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 436, + 317, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 436, + 317, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 436, + 317, + 555, + 390 + ], + "type": "image", + "image_path": "618e30d98c0ccde30948435dac14d414063400100c51eea9f2de997e93d982d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 316, + 417, + 556, + 462 + ], + "blocks": [ + { + "bbox": [ + 80, + 523, + 269, + 534 + ], + "lines": [ + { + "bbox": [ + 80, + 523, + 269, + 534 + ], + "spans": [ + { + "bbox": [ + 80, + 523, + 269, + 534 + ], + "type": "text", + "content": "Table 3. Effect of using different pooling strategies." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 417, + 556, + 462 + ], + "lines": [ + { + "bbox": [ + 316, + 417, + 556, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 556, + 462 + ], + "type": "table", + "html": "
LevelsMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
[37,38,39,40]73.6860.9073.5257.4255.6044.072.00
[25,30,35,40]71.4859.4871.4556.3754.8343.4270.19
[10,20,30,40]69.0355.3569.5653.3752.4742.7368.39
", + "image_path": "44aec9307beb82307efeb65d303ab11b6026c12516eb471c57d6731110f2f18b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "type": "text", + "content": "Table 4. Ablation of different sampling strategies of intermediate features. The first column indicates the " + }, + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 313, + 463, + 555, + 496 + ], + "type": "text", + "content": " sampled layers –of a total of 40 layers." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "type": "text", + "content": "because of what we call, depth-pooling trade-off. We conjecture that sampling from shallower layers does in fact provide additional complementary information. Nevertheless, this inevitably results in a distribution shift w.r.t. the last-layer feature distribution, hindering the effectiveness of the frozen AdaptivePool (see Eq. 15). The infeasibility of retraining this module for computational reasons, thus, leaves us with a delicate trade-off between the quality of the pooling and the depth of the features that must be accounted for." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 477, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 477, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 477, + 635 + ], + "type": "text", + "content": "6.2. Study of deformable attention" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "1) Comparison w.r.t. other baselines: Here we empirically evaluate the benefits of our proposed RDSA. Concretely, in Tab. 5 we empirically compare the performance of the standard CA [35] and Def. CA [45], as well as different decoder-query initialization strategies [43] for the latter. Notice the prominent performance degradation de" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21521" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 297, + 125 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 297, + 125 + ], + "type": "table", + "html": "
Att. strat.MRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Stand. CA [35]70.7751.7466.7845.1942.7243.1669.87
Def. CA [45]72.5858.1971.8255.8054.2743.2670.58
Def. CA +PureInit [43]72.1357.8770.6854.8552.9243.1970.32
Def. CA +MixedInit [43]72.2658.7171.9456.4354.9443.1069.86
Ours73.6860.9073.5357.4255.6044.072.00
", + "image_path": "c2d1dd3ab546a7b48a124d8d76ffe9d83d8ac01f00e93b285bfee3af7db5d358.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 153, + 172, + 225 + ], + "blocks": [ + { + "bbox": [ + 61, + 153, + 172, + 225 + ], + "lines": [ + { + "bbox": [ + 61, + 153, + 172, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 172, + 225 + ], + "type": "image", + "image_path": "eb63f808cbe9a3924b16163b2f897ae83feb7f744fcc3010ab9e4416e6eca1df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "text", + "content": "Figure 4. Average of the weighted offsets across " + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "text", + "content": " decoder queries and " + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "text", + "content": " batch elements for the " + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 55, + 226, + 295, + 270 + ], + "type": "text", + "content": " refinement levels. Here head 0 (left) is initialized near the left boundary, and head 1 (right) near the right boundary." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 153, + 287, + 225 + ], + "blocks": [ + { + "bbox": [ + 175, + 153, + 287, + 225 + ], + "lines": [ + { + "bbox": [ + 175, + 153, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 175, + 153, + 287, + 225 + ], + "type": "image", + "image_path": "f0dc62f156c17fe0ccc319e3ccfc324a86027ac54bfe06751d2ff0dc41b009e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "text", + "content": "rived from the use of the CA, or how the query initialization techniques –far from improving the performance of the Def. CA– in some cases are even harmful. In contrast, our method consistently outperforms all these tested baselines, improving by " + }, + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "inline_equation", + "content": "2.71\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "text", + "content": " R1@0.7 or " + }, + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "inline_equation", + "content": "1.62\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 361 + ], + "type": "text", + "content": " mAP w.r.t. the original Def. CA. We refer to Sec. K for the ablation of the different sampling strategies of RDSA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "text", + "content": "2) Where are the offsets pointing to? In Fig. 4 we depict the refinement of the weighted offset of a given query " + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "inline_equation", + "content": "d_{q} = \\sum_{p=1}^{P} A_{q,p} \\Delta_{q,p}" + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "text", + "content": ", across the " + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 363, + 296, + 555 + ], + "type": "text", + "content": " different refinement steps. Note that as depicted in Eq. 9, these offsets are relative to the estimated moment widths, with -1 and +1 indicating the estimated left- and right-most boundaries. The Def. CA lacks context awareness, causing its offset predictions to stay near the offset initializations. In contrast, RDSA leverages an enhanced contextualization, thus reducing its bias towards the offset initialization. Notably, RDSA learns to look beyond the current action boundaries (offsets < -1 and > +1) capturing near-boundary information, crucial for refining moment boundaries. This contrasts with [45] which remains constrained within the predicted moment boundaries. We consider this the reason why RDSA is especially useful to localize long actions (see Sec. K)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 281, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 281, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 281, + 577 + ], + "type": "text", + "content": "6.3. Conditioning signal for the sparse stream " + }, + { + "bbox": [ + 55, + 563, + 281, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "To effectively integrate RDSA, it is imperative to obtain an expressive video-based conditioning signal -i.e., the red arrow in Fig. 2a. As shown in Tab. 6, using the raw video representation " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\ell}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " significantly degrades performance, highlighting the need for a richer signal incorporating textual and temporal information. While conditioning on the outputs of the CA or SA modules of the dense stream " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " improves results, these lack non-linearities. Instead, using the non-linear PFFN output provides greater flexibility, leading to superior expressiveness. Notably, this choice also enhances HD performance, underscoring the benefits of task" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 555, + 123 + ], + "blocks": [ + { + "bbox": [ + 55, + 128, + 295, + 150 + ], + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 150 + ], + "type": "text", + "content": "Table 5. Comparison across different attention strategies as well as various decoder query initializations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 555, + 123 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 123 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 123 + ], + "type": "table", + "html": "
Type of featuresMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
Raw. rep72.5856.3770.1250.9950.7444.1371.37
Post CA72.2658.3972.2055.9254.8343.6470.65
Post CA+SA72.1358.3272.0056.5754.9443.7870.06
Post CA+SA+FFN73.6860.9073.5357.4255.6044.0072.00
", + "image_path": "94a582ac2d00b618c3f4362957f8c5718096f46f1dce3a16a92e57c34c6d6c72.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 144, + 555, + 200 + ], + "blocks": [ + { + "bbox": [ + 340, + 127, + 527, + 138 + ], + "lines": [ + { + "bbox": [ + 340, + 127, + 527, + 138 + ], + "spans": [ + { + "bbox": [ + 340, + 127, + 527, + 138 + ], + "type": "text", + "content": "Table 6. Evaluation of various conditioning signals." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 144, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 144, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 555, + 200 + ], + "type": "table", + "html": "
CASARDSAFFNMRHD
R1@0.5R1@0.7mAP@0.5mAP@0.75mAPmAPHIT@1
✓✓✓71.4257.7471.7356.1854.4743.5371.35
✓✓✓71.8758.7172.1355.9954.3243.2869.55
✓✓✓59.6830.7159.0224.629.0943.5870.39
✓✓✓72.8459.2973.1356.9155.7243.8572.58
✓✓✓✓73.6860.9073.5257.4255.6044.072.00
", + "image_path": "3ed9da65d866c5d267e2b8d9b396bfb0d11f78812bfd1d128ba7376ebdbeaff7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 204, + 541, + 214 + ], + "lines": [ + { + "bbox": [ + 326, + 204, + 541, + 214 + ], + "spans": [ + { + "bbox": [ + 326, + 204, + 541, + 214 + ], + "type": "text", + "content": "Table 7. Importance of the modules of the sparse stream " + }, + { + "bbox": [ + 326, + 204, + 541, + 214 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 326, + 204, + 541, + 214 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 225, + 449, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 449, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 449, + 236 + ], + "type": "text", + "content": "interaction between MR and HD." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 250, + 544, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 544, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 544, + 263 + ], + "type": "text", + "content": "6.4. Effect of each module of the sparse stream " + }, + { + "bbox": [ + 313, + 250, + 544, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 269, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 413 + ], + "type": "text", + "content": "In this ablation, we empirically evaluate the contribution of each of the components of " + }, + { + "bbox": [ + 313, + 269, + 555, + 413 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 269, + 555, + 413 + ], + "type": "text", + "content": ". Concretely, we evaluate the influence of the CA, the SA, the RDSA, and the final PFFN. Observe in Tab. 7 that the initial textual injection CA seems to be particularly relevant for the R1 metric, while the reasoning SA module is important for the mAP. Moreover, as expected, the RDSA module is critical. This is to be expected, as without it, the queries become oblivious to the input video, leaving no other option but that of randomly guessing. Again, the additional expressivity that the final PFFN brings via its non-linearities proves to be beneficial, despite affecting less than the other modules." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 391, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 391, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 391, + 439 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 449, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 663 + ], + "type": "text", + "content": "In this work, we introduced the Sparse-Dense Side-Tuner, the first proposal-free side-tuning method for VTG. Our dual-stream architecture jointly optimizes dense embeddings and recurrent decoder queries, carefully adapting to the sparse and dense nature of MR and HD. We also identify an inherent context limitation of the Deformable CA-a key component of proposal-free architectures like ourswhich motivates our proposed alternative, the Referencebased Deformable Self-Attention. Finally, we provide the first effective integration of InternVideo2 to a side-tuning framework, leading to a substantial performance boost. We evaluate our method on QVHighlights, TACoS, and Charades-STA, surpassing previous ST works and attaining either SOTA or near-SOTA, all while maintaining a minimal parameter count and a low memory overhead. We leave as future work studying the often overlooked problem of the cross-domain generalization capabilities of these methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Acknowledgments: This work has been partially supported by the Spanish project PID2022-136436NB-I00 and by ICREA under the ICREA Academia programme." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21522" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 702 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Taivanbat Badamdorj, Mrigank Rochan, Yang Wang, and Li Cheng. Joint visual and audio learning for video highlight detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8127-8137, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 145, + 296, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 296, + 188 + ], + "type": "text", + "content": "[2] Zhuo Cao, Bingqing Zhang, Heming Du, Xin Yu, Xue Li, and Sen Wang. Flashvtg: Feature layering and adaptive score handling network for video temporal grounding. arXiv preprint arXiv:2412.13441, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "text", + "content": "[4] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[5] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 370 + ], + "type": "text", + "content": "[6] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 295, + 413 + ], + "type": "text", + "content": "[7] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 456 + ], + "type": "text", + "content": "[8] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. arXiv preprint arXiv:2404.01745, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 498 + ], + "type": "text", + "content": "[9] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 295, + 563 + ], + "type": "text", + "content": "[10] Fa-Ting Hong, Xuanteng Huang, Wei-Hong Li, and Wei-Shi Zheng. Mini-net: Multiple instance ranking network for video highlight detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 345-360. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 564, + 295, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 564, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 57, + 564, + 295, + 618 + ], + "type": "text", + "content": "[11] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13846-13856, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 618, + 295, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 618, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 57, + 618, + 295, + 661 + ], + "type": "text", + "content": "[12] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 661, + 295, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 661, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 57, + 661, + 295, + 702 + ], + "type": "text", + "content": "[13] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648, 2016. 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 708 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[14] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In European Conference on Computer Vision, pages 220-238. Springer, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 117, + 554, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 117, + 554, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 117, + 554, + 159 + ], + "type": "text", + "content": "[15] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 159, + 554, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 159, + 554, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 159, + 554, + 213 + ], + "type": "text", + "content": "[16] Feng Li, Hao Zhang, Shilong Liu, Jian Guo, Lionel M Ni, and Lei Zhang. Dn-detr: Accelerate detr training by introducing query denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13619-13627, 2022. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 213, + 554, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 554, + 277 + ], + "type": "text", + "content": "[17] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 278, + 554, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 554, + 320 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 554, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 554, + 374 + ], + "type": "text", + "content": "[19] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 374, + 554, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 554, + 417 + ], + "type": "text", + "content": "[20] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 418, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 554, + 470 + ], + "type": "text", + "content": "[21] Ye Liu, Jixuan He, Wanhua Li, Junsik Kim, Donglai Wei, Hanspeter Pfister, and Chang Wen Chen. r2-tuning: Efficient image-to-video transfer learning for video temporal grounding. arXiv preprint arXiv:2404.00801, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 472, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 525 + ], + "type": "text", + "content": "[22] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 525, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 554, + 567 + ], + "type": "text", + "content": "[23] Weiheng Lu, Jian Li, An Yu, Ming-Ching Chang, Shengpeng Ji, and Min Xia. Llava-mr: Large language-and- vision assistant for video moment retrieval. arXiv preprint arXiv:2411.14505, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 610 + ], + "type": "text", + "content": "[24] Boris Meinardus, Anil Batra, Anna Rohrbach, and Marcus Rohrbach. The surprising effectiveness of multimodal large language models for video moment retrieval. arXiv preprint arXiv:2406.18113, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 611, + 554, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 611, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 317, + 611, + 554, + 653 + ], + "type": "text", + "content": "[25] WonJun Moon, Sangeek Hyun, Su Been Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 654, + 554, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 554, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 554, + 708 + ], + "type": "text", + "content": "[26] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23023-23033, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21523" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[27] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "type": "text", + "content": "[28] Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hongsheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. Advances in Neural Information Processing Systems, 35:26462-26477, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 148, + 294, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 148, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 57, + 148, + 294, + 201 + ], + "type": "text", + "content": "[29] David Pujol-Perich, Albert Clapés, and Sergio Escalera. Sada: Semantic adversarial unsupervised domain adaptation for temporal action localization. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9237–9247. IEEE, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 202, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 202, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 202, + 294, + 266 + ], + "type": "text", + "content": "[30] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yingya Zhang, Changxin Gao, Deli Zhao, and Nong Sang. Disentangling spatial and temporal learning for efficient image-to-video transfer learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13934-13944, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 267, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 267, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 294, + 331 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 332, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 294, + 396 + ], + "type": "text", + "content": "[32] Anna Rohrbach, Marcus Rohrbach, Wei Qiu, Annemarie Friedrich, Manfred Pinkal, and Bernt Schiele. Coherent multi-sentence video description with variable level of detail. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 184-195. Springer, 2014. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 396, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 396, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 396, + 294, + 450 + ], + "type": "text", + "content": "[33] Min Sun, Ali Farhadi, and Steve Seitz. Ranking domain-specific highlights by analyzing edited videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, pages 787-802. Springer, 2014. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 450, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 450, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 57, + 450, + 294, + 492 + ], + "type": "text", + "content": "[34] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Lst: Ladder side-tuning for parameter and memory efficient transfer learning. Advances in Neural Information Processing Systems, 35:12991-13005, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 57, + 493, + 294, + 536 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 601 + ], + "type": "text", + "content": "[36] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396-416. Springer, 2025. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "text", + "content": "[37] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 655, + 294, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 655, + 294, + 697 + ], + "spans": [ + { + "bbox": [ + 57, + 655, + 294, + 697 + ], + "type": "text", + "content": "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 698, + 294, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 698, + 294, + 708 + ], + "spans": [ + { + "bbox": [ + 57, + 698, + 294, + 708 + ], + "type": "text", + "content": "[39] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Heng-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 439 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 127 + ], + "type": "text", + "content": "wei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18709-18719, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 554, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 554, + 181 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 554, + 181 + ], + "type": "text", + "content": "[40] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 182, + 554, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 554, + 244 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 554, + 244 + ], + "type": "text", + "content": "[41] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623-13633, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 246, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 300 + ], + "type": "text", + "content": "[42] Zhen Yang, Ming Ding, Chang Zhou, Hongxia Yang, Jingren Zhou, and Jie Tang. Understanding negative sampling in graph representation learning. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1666-1676, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 342 + ], + "type": "text", + "content": "[43] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 396 + ], + "type": "text", + "content": "[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 396, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 439 + ], + "type": "text", + "content": "[45] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21524" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_content_list.json b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9f02c563bbdd8af2b57fb2e697ca8a64b9eec37c --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling", + "text_level": 1, + "bbox": [ + 125, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianglong He $^{1,2\\star}$ Zi-Xin Zou $^{2\\star}$ Chia-Hao Chen $^{1,2}$ Yuan-Chen Guo $^{2}$ Ding Liang $^{2}$ Chun Yuan $^{1\\dagger}$ Wanli Ouyang $^{3}$ Yan-Pei Cao $^{2}$ Yangguang Li $^{3,2\\dagger}$ Tsinghua University ${}^{2}\\mathrm{VAST}$ ${}^{3}\\mathrm{The~Chinese~University~of~Hong~Kong}$", + "bbox": [ + 99, + 178, + 895, + 236 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74ad2e45a4a9ab7a764ae131195c7a807ffba864067c7cda15dbfaf7d1934056.jpg", + "image_caption": [ + "Figure 1. SparseFlex VAE achieves high-fidelity reconstruction and generalization from point clouds. Benefiting from a sparse-structured differentiable isosurface surface representation and an efficient frustum-aware sectional voxel training strategy, our SparseFlex VAE demonstrates the state-of-the-art performance on complex geometries (left), open surfaces (top right), and even interior structures (bottom right), facilitating the high-quality image-to-3D generation with arbitrary topology." + ], + "image_footnote": [], + "bbox": [ + 98, + 250, + 906, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 577, + 326, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating high-fidelity 3D meshes with arbitrary topology, including open surfaces and complex interiors, remains a significant challenge. Existing implicit field methods often require costly and detail-degrading watertight conversion, while other approaches struggle with high resolutions. This paper introduces SparseFlex, a novel sparse-structured isosurface representation that enables differentiable mesh reconstruction at resolutions up to $1024^{3}$ directly from rendering losses. SparseFlex combines the accuracy of Flexicubes with a sparse voxel structure, focusing computation on surface-adjacent regions and efficiently handling open surfaces. Crucially, we introduce a frustum-aware sectional voxel training strategy that activates only relevant voxels during rendering, dramatically reducing memory consumption and enabling high-resolution training. This also allows, for the first time, the reconstruction of mesh interiors using only rendering supervision. Building upon this, we demonstrate", + "bbox": [ + 88, + 609, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a complete shape modeling pipeline by training a variational autoencoder (VAE) and a rectified flow transformer for high-quality 3D shape generation. Our experiments show state-of-the-art reconstruction accuracy, with a $\\sim 82\\%$ reduction in Chamfer Distance and a $\\sim 88\\%$ increase in $F$ -score compared to previous methods, and demonstrate the generation of high-resolution, detailed 3D shapes with arbitrary topology. By enabling high-resolution, differentiable mesh reconstruction and generation with rendering losses, Sparse-Flex significantly advances the state-of-the-art in 3D shape representation and modeling.", + "bbox": [ + 511, + 579, + 908, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 767, + 645, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Generative AI is rapidly advancing, with applications spanning entertainment, design, and robotics. However, generating 3D content is fundamentally more challenging than 2D image or text generation due to the inherent complexity of representing and manipulating 3D geometry. Achieving high fidelity and supporting arbitrary topology—including open surfaces and complex interiors—presents particularly", + "bbox": [ + 511, + 795, + 908, + 905 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\star$ Equal contribution.", + "bbox": [ + 91, + 876, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 91, + 888, + 227, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14822", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "significant hurdles.", + "bbox": [ + 89, + 90, + 220, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent progress in 3D generative models has explored various representations, including point clouds [45, 46, 52, 69, 82], meshes [51, 63], 3DGS [5, 16, 19, 34, 79, 90], and implicit fields [9, 10, 36, 37, 59, 78, 79, 89, 93, 98, 100]. Implicit field representations, such as Signed Distance Functions (SDFs) and occupancy fields, have shown favorable results [33, 78, 79, 87, 89, 100]. However, creating training data for these methods typically involves a two-step process that limits their effectiveness. First, raw 3D mesh data must be converted into watertight representations [80, 93] to calculate SDF or occupancy values—a computationally expensive preprocessing step that often degrades fine details. Second, isosurfacing techniques (e.g., Marching Cubes [43], Dual Contouring [28]) are used to extract meshes from the learned continuous field, which can introduce further inaccuracies and artifacts. While Unsigned Distance Fields (UDFs) offer a potential way to model open surfaces [7, 11, 17, 40, 42, 86], they often suffer from instability and struggle to capture geometric fine details.", + "bbox": [ + 89, + 108, + 485, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rendering-based supervision offers a powerful, differentiable alternative for training 3D representations and 3D generative models [60, 61, 74, 75, 79, 81]. By directly comparing rendered images of a generated mesh to ground-truth data, rendering losses avoid the need for the initial watertight preprocessing step and better preserve fine details. However, a critical bottleneck arises: when used with dense implicit fields, rendering supervision requires extremely high memory consumption at high resolutions, severely limiting the achievable fidelity.", + "bbox": [ + 89, + 397, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces SparseFlex, a new sparse-structured isosurface representation that addresses these limitations and unlocks high-resolution, differentiable mesh reconstruction and generation using rendering supervisions. SparseFlex is built upon Flexicubes [61], providing accurate and differentiable isosurface extraction. The key design is the use of a sparse voxel structure instead of a conventional dense grid. This sparsity is crucial for two primary reasons: (1) it dramatically reduces memory consumption, enabling high-resolution modeling, and (2) it allows for the effective pruning of voxels near open boundaries, enabling the representation of open surfaces.", + "bbox": [ + 89, + 551, + 485, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully leverage the capabilities of SparseFlex, we propose frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering [1], this approach activates only the subset of SparseFlex voxels that reside within the camera's viewing frustum during each training iteration. We also introduce an adaptive strategy to control the frustum's parameters, further optimizing memory usage. This not only substantially reduces computational and memory overhead but also enables, for the first time, the reconstruction of mesh interiors using only rendering supervision by appropriately positioning the camera. Fig. 3 illustrates our", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SparseFlex representation and the efficient sectional voxel training strategy.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on SparseFlex and our frustum-aware training, we present a complete 3D shape modeling pipeline. We employ a variational autoencoder (VAE) architecture, drawing inspiration from TRELLIS [79] but with key modifications. Firstly, because our focus is on high-fidelity geometry, we use point clouds as input to the VAE, providing a direct and detailed representation of the shape's surface. Furthermore, we introduce a self-pruning upsampling module within the decoder to further refine the sparse voxel structure, which is particularly beneficial for representing open surfaces. A rectified flow transformer is then trained on the learned latent space for high-quality, image-conditioned 3D shape generation. Through extensive experiments on Toy4k [66], ABO [12], GSO [15], Meta [47], Objaverse [14], and Deep fashion3D [21], our method demonstrates state-of-the-art shape reconstruction accuracy with minimal detail degradation, and high-quality single-image 3D shape generation.", + "bbox": [ + 511, + 123, + 908, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are:", + "bbox": [ + 532, + 382, + 718, + 396 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SparseFlex, a new sparse-structured isosurface representation enabling efficient, high-resolution, and differentiable 3D shape modeling, with natural handling of open surfaces.", + "- We introduce a novel sectional voxel training strategy with adaptive view frustum control, dramatically reducing memory consumption and enabling high-resolution mesh reconstruction and generation, including interiors, using rendering losses.", + "- We demonstrate state-of-the-art reconstruction accuracy and the generation of high-resolution, detailed 3D shapes with arbitrary topology, representing a significant advance in the field." + ], + "bbox": [ + 513, + 398, + 906, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 607, + 653, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Shape Representations for Generation", + "text_level": 1, + "bbox": [ + 511, + 633, + 870, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud. Point clouds are a flexible 3D representation and can be easily acquired using depth sensors and LiDAR. Research has focused on point cloud processing [57, 58, 97] and generation [45, 46, 52, 69], with recent approaches treating point clouds as a distribution, directly sampling point clouds from noise via generative models [4, 91]. However, due to their limitation in representing solid surfaces, an additional surface reconstruction step [24, 29, 56] is required.", + "bbox": [ + 511, + 657, + 908, + 780 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Triangle Mesh. Triangle meshes are the primary representation for 3D assets in modern industrial pipelines. Recent works employ auto-regressive models to generate triangle faces sequentially [20, 51, 63, 68], improving resemblance to artist-created meshes. While effective for low-poly meshes, these methods struggle to produce high-quality meshes with a high face count.", + "bbox": [ + 511, + 794, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af5f77348e825581735aa3ecfbe62036e2e7048c1d469b53250f0ec3bdad936e.jpg", + "image_caption": [ + "Figure 2. Overview of the SparseFlex VAE pipeline. SparseFlex VAE takes point clouds sampled from a mesh as input, voxelizes them, and aggregates their features into each voxel. A sparse transformer encoder-decoder compresses the structured feature into a more compact latent space, followed by a self-pruning upsampling for higher resolution. Finally, the structured features are decoded to SparseFlex through a linear layer. Using the frustum-aware section voxel training strategy, we can train the entire pipeline more efficiently by rendering loss." + ], + "image_footnote": [], + "bbox": [ + 112, + 65, + 880, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Implicit Field. Implicit fields (SDF or occupancy) are widely used in geometry learning, particularly for 3D reconstruction [17, 24, 55, 61, 72, 101] and generation [9, 10, 36, 37, 50, 59, 78, 79, 83, 89, 93-96, 98, 100], producing high-quality meshes. Enhancements with triplane [78], vector set [87, 89, 98] sparse voxels [79] and sparse hierarchical voxels [59] allow neural networks to decode the field more effectively. These methods typically use the isosurface techniques [28, 43, 53] for surface extraction but struggle with open-surface shapes like clothing and flowers. Additionally, many require a time-consuming watertight conversion that degrades details, while rendering supervised approaches [79] suffer from high memory consumption when training at high resolutions.", + "bbox": [ + 89, + 348, + 485, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open Surfaces. Meshes with open surfaces are common but remain challenging to process. The Unsigned Distance Field (UDF) is widely used for open-surface modeling from point clouds [11] or multiview images [40, 42]. Surf-D [86] introduces a UDF-based diffusion model for generating meshes with arbitrary topology. However, UDF-based shape modeling is more difficult than SDF, and its surface extraction [18] is prone to instability due to inaccuracy in neural network gradient estimations. As a result, achieving high-quality open-surface meshes remain challenging. 3PSDF [7] introduces a three-pole sign to distinguish surface-adjacent regions but relies on binary occupancy grids for surface extraction, leading to discontinuities and artifacts on the surface. Slice3D [73] introduces a slicing representation to reconstruct the interior structure with axis-aligned multislice images. DMesh series [64, 65] introduce differentiable mesh representations to represent general triangular meshes, while directly utilizing the representation for VAE remains a nontrivial challenge. In this paper, we incorporate sparse structure into Flexicubes [61] to efficient high-quality isosurface representation for open faces.", + "bbox": [ + 91, + 583, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Generative Models and VAE", + "text_level": 1, + "bbox": [ + 511, + 347, + 795, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing 3D generation studies are primarily classified into two categories: large 3D reconstruction model following 2D multiview diffusion models and native 3D generation models. The first category uses multiview diffusion models to generate multiview images from text or an image [25, 26, 39, 62, 71], followed by a large 3D reconstruction model [23, 35, 67, 81, 84, 92, 102] is employed to reconstruct the 3D representation from these images in seconds. However, inconsistencies between generated multiview images significantly often degrade the result quality. The second category focuses on native 3D generation models that directly generate 3D models through generative models, including GAN [3, 8, 77, 99], auto-regressive [49, 88], diffusion [27, 69, 89, 93] and rectified flow [37, 79].", + "bbox": [ + 511, + 369, + 908, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the diverse and non-compact nature of 3D representation, many approaches use Variational Auto-Encoder (VAE) [31] or Vector Quantized VAE (VQ-VAE) [70] to encode 3D shapes into latent spaces for the generative models. Geometry-focused methods often input point clouds uniformly sampled from mesh surfaces into the VAE [6, 89]. When aim at decoding both geometry and texture, some methods [33, 34, 79] encode multiview image features into latent spaces. This two-stage process makes the VAE's reconstruction quality crucial for subsequent generation performance. Some works [6, 87] have improved generation quality by enhancing the VAE's shape encoding-decoding capabilities. In this paper, we aim to develop a foundational VAE that encodes 3D shapes into latent spaces and reconstructs them with arbitrary topology while preserving the raw 3D shape's details.", + "bbox": [ + 511, + 582, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 829, + 604, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present a method for high-resolution 3D shape modeling based on a novel sparse-structured isosurface representation, SparseFlex. Our approach leverages differentiable rendering", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14824", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54d6c885ed078716d2714c223431bdae9fb7e3b2e736671b67a34167fc775652.jpg", + "image_caption": [ + "Figure 3. Frustum-aware sectional voxel training. The previous mesh-based rendering training strategy (left) requires activating the entire dense grid to extract the mesh surface, even though only a few voxels are necessary during rendering. In contrast, our approach (right) adaptively activates the relevant voxels and enables the reconstruction of mesh interiors only using rendering supervision." + ], + "image_footnote": [], + "bbox": [ + 89, + 66, + 477, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for training, enabling accurate reconstruction of complex geometries, including open surfaces and interiors (see Fig. 1. Fig. 2 illustrates our variational autoencoder (VAE) based shape modeling pipeline, which utilizes the SparseFlex representation to learn a compact latent space of 3D shapes. The VAE decoder outputs the parameters of a SparseFlex instance, facilitating high-resolution mesh reconstruction. A key component of our approach is frustum-aware sectional voxel training, which significantly reduces memory consumption, allowing for training at resolutions up to $1024^3$ . Details of the SparseFlex representation are provided in Sec. 3.1, followed by a description of the VAE architecture in Sec. 3.2 and the training procedure in Sec. 3.3. Sec. 3.4 describes image-conditioned 3D shape generation using a rectified flow transformer on the learned latent space.", + "bbox": [ + 91, + 314, + 486, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. SparseFlex Representation", + "text_level": 1, + "bbox": [ + 89, + 550, + 333, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary. To achieve differentiable mesh extraction while preserving sharp features, we build upon Flexicubes [61], a method based on Dual Marching Cubes [53] (DMC). DMC places vertices at the center of voxels rather than edges/corners, leading to better feature preservation. Flexicubes constructs a dense voxel with a resolution of $N_r^3$ , where an SDF scalar grid $s \\in \\mathbb{R}^{N_g^3}$ is assigned to the voxel corner points, where $N_g = N_r + 1$ . For each voxel, it incorporates interpolation weights $\\alpha \\in \\mathbb{R}_{>0}^{N_r^3 \\times 8}, \\beta \\in \\mathbb{R}_{>0}^{N_r^3 \\times 12}$ to each voxel cell, and deformation vectors $\\delta \\in \\mathbb{R}^{N_g^3}$ to each SDF grid. The underlying surface mesh can be effectively optimized via differentiable rendering [32].", + "bbox": [ + 89, + 568, + 483, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse Structured Flexicubes. The core design of SparseFlex is the introduction of a sparse voxel structure, enabling high-resolution shape representation while drastically reducing memory consumption. Instead of a dense grid, SparseFlex represents a shape using a significantly smaller set of voxels, $\\mathcal{V}$ , concentrated near the surface. This sparsity is crucial for two reasons: (1) it allows us to achieve much higher resolutions than would be possible with a dense grid, and (2) it enables the natural representation of open surfaces by", + "bbox": [ + 89, + 763, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "simply omitting voxels in empty regions.", + "bbox": [ + 511, + 92, + 785, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the SparseFlex is defined by a set of $N_{v}$ voxels, $\\mathcal{V} = \\{v_{i} = (x_{i},y_{i},z_{i})\\}$ , where $v_{i}$ represents the 3D coordinates of the center of the $i$ -th voxel. Let $N_{c}$ be the number of corner grids associated with these voxels, where $N_{v} = |\\mathcal{V}|$ . Each voxel is associated with interpolation weights $\\{\\alpha_{i}\\in \\mathbb{R}_{>0}^{8},\\beta_{i}\\in \\mathbb{R}_{>0}^{12}|0\\leq i < N_{v}\\}$ . Each corner grid is associated with an SDF value $\\{s_j|0\\leq j < N_c\\}$ and deformation vectors $\\{\\delta_j|0\\leq j < N_c\\}$ . Due to the sparsity, $N_{v}\\ll N_{r}^{3}$ and $N_{c}\\ll N_{g}^{3}$ , representing a significant reduction in memory usage compared to the dense Flexicubes representation. We only apply Dual Marching Cubes on these sparse voxels to extract the underlying surface. Formally, the SparseFlex representation, $\\mathcal{S}$ , is defined as:", + "bbox": [ + 511, + 106, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left(\\mathcal {V}, \\mathcal {F} _ {c}, \\mathcal {F} _ {v}\\right), \\quad \\mathcal {F} _ {c} = \\left\\{s _ {j}, \\delta_ {j} \\right\\}, \\quad \\mathcal {F} _ {v} = \\left\\{\\alpha_ {i}, \\beta_ {i} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 299, + 906, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{V}$ represents the voxel centers, $\\mathcal{F}_c$ contains the SDF values and deformations at the corner grids, and $\\mathcal{F}_v$ contains the interpolation weights for each voxel.", + "bbox": [ + 511, + 325, + 906, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SparseFlex inherits the differentiability of Flexicubes, allowing for end-to-end optimization using rendering losses. This eliminates the need for watertight mesh pre-processing, preserving fine details. Furthermore, the sparse structure, combined with the continuous and deformable nature of the SDF, allows for accurate and efficient representation of high-quality open-surface meshes. The sparsity also paves the way for our efficient frustum-aware training strategy, described in Sec. 3.3.", + "bbox": [ + 511, + 372, + 908, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SparseFlex VAE for Shape Modeling", + "text_level": 1, + "bbox": [ + 511, + 518, + 831, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn a compact and disentangled latent space of 3D shapes, we employ a variational autoencoder (VAE) [31] that utilizes the SparseFlex representation. A VAE learns a probabilistic mapping between an input space (in our case, 3D shapes represented as point clouds) and a lower-dimensional latent space, enabling both reconstruction and generation of shapes. Fig. 2 also illustrates our VAE architecture. Our architecture draws inspiration from TRELLIS [79], but with key modifications to leverage the strengths of SparseFlex.", + "bbox": [ + 511, + 542, + 908, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoder. The input to our encoder is a point cloud $\\mathcal{P} = \\{p_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}$ , uniformly sampled from the surface of a 3D mesh, along with corresponding normals $\\mathcal{N} = \\{n_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}$ . We first voxelize the point cloud to obtain the sparse structure $\\mathcal{V}$ of the SparseFlex representation $\\mathcal{S}$ . We then employ a shallow PointNet [57] to aggregate local geometric features within each voxel. Specifically, for each voxel $v_{i}\\in \\mathcal{V}$ , we apply a local max-pooling operation [55] to the points contained within that voxel, producing a feature vector $f_{i}$ . These voxel features $\\mathcal{F} = \\{f_i\\}$ , along with the sparse structure $\\mathcal{V}$ , are then fed into a sparse transformer backbone. This backbone utilizes shifted window attention [41, 85], similar to TRELLIS [79], but is adapted to operate directly on the sparse voxel features $\\mathcal{F}$ and structure $\\mathcal{V}$ . The trans", + "bbox": [ + 511, + 685, + 910, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "former outputs a latent code $z \\in \\mathbb{R}^{d_z}$ , which represents the encoded 3D shape.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder. The decoder takes the latent code $z$ as input and predicts the parameters of a SparseFlex instance, $S = (\\mathcal{V},\\mathcal{F}_c,\\mathcal{F}_v)$ . We use a series of transformer layers, culminating in a final linear layer, to predict the SDF values $(s_j)$ and deformations $(\\delta_{j})$ for each corner grid, as well as the interpolation weights $(\\alpha_{i},\\beta_{i})$ for each voxel.", + "bbox": [ + 89, + 128, + 483, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Upsampling Modules. To achieve high-resolution reconstructions, we incorporate two convolutional, self-pruning upsampling modules within the decoder, following the transformer. These modules progressively increase the resolution of the SparseFlex representation. Each upsampling module subdivides existing voxels into smaller voxels (increasing the resolution by a factor of 4). Crucially, each module also prunes redundant voxels based on a predicted occupancy value. A voxel is considered occupied if, after the subdivision, it contains any points from the input point cloud $\\mathcal{P}$ . This pruning process, inspired by [59], is essential for maintaining the sparsity of the SparseFlex representation and is particularly beneficial for accurately representing open surfaces, as it removes unnecessary voxels in empty regions.", + "bbox": [ + 89, + 224, + 486, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training SparseFlex VAE", + "text_level": 1, + "bbox": [ + 89, + 446, + 323, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our SparseFlex VAE end-to-end using rendering losses, leveraging the differentiability of the SparseFlex representation and a new training strategy called frustum-aware sectional voxel training. This strategy dramatically reduces memory consumption during training, enabling us to achieve high resolutions (up to $1024^{3}$ ) that would be infeasible with traditional approaches.", + "bbox": [ + 89, + 470, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frustum-aware Sectional Voxel Training. Even with the sparse structure of SparseFlex, directly rendering the entire representation at high resolutions can be computationally expensive. Furthermore, standard rendering supervision typically focuses only on the visible surface, neglecting the interior of the shape. Besides, recent methods [38, 74, 76, 79, 81] relying on rendering supervision from mesh typically require extracting the entire mesh because a dense representation doesn't trivially allow for partial extraction. In contrast, our sparse representation naturally enables partial extraction. To address these issues, we introduce frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering for efficient visibility culling [1], this approach activates only the voxels within the camera's viewing frustum during each training iteration. \"Activating\" a voxel means including it in the isosurface extraction and rendering process. This \"sectional\" approach means we only process a portion of the 3D space at a time, markedly reducing memory usage.", + "bbox": [ + 89, + 583, + 485, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, given the camera's extrinsic $\\pi$ , intrinsics $K$ , and the near $(n)$ and far $(f)$ clipping planes of the viewing frustum, we compute the Model-View-Projection", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(MVP) matrix MVP. We then use a boolean operator to check whether the center of each voxel $v_{i}$ lies within the viewing frustum defined by the MVP matrix. We use $I(v_{i}\\in \\mathrm{Frustum}(\\mathbf{MVP}))$ to represent this check, where $I(\\cdot)$ is an indicator function. The set of active voxels, $\\mathcal{V}_{\\mathrm{active}}$ , is then defined as:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} _ {\\text {a c t i v e}} = \\left\\{v _ {i} \\mid I \\left(v _ {i} \\in \\operatorname {F r u s t u m} (\\mathbf {M V P})\\right) = 1, v _ {i} \\in \\mathcal {V} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 186, + 906, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Frustum and Interior Reconstruction. We introduce a visibility ratio $\\alpha$ ( $0 < \\alpha \\leq 1$ ) controlling the proportion of active voxels in SparseFlex. We adaptively adjust the near and far clipping planes to ensure that approximately $\\alpha N_v$ voxels are within the frustum. This is achieved through an iterative process: we initially set the near and far planes and iteratively adjust them based on the number of active voxels until the desired proportion is reached.", + "bbox": [ + 511, + 204, + 906, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This adaptive frustum also enables a novel capability: reconstructing mesh interiors using only rendering supervision. By positioning a virtual camera inside the object or adjusting the near clipping plane to intersect the mesh, we can render and supervise the internal structure (see Fig. 4). Moreover, the zoom-in camera viewpoint can render the mesh surface with greater details, making it better for higher-resolution training. This is a significant advantage over methods that rely on watertight representations, which cannot capture interior details.", + "bbox": [ + 511, + 325, + 908, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. We train our VAE in an end-to-end manner, with an objective function comprising four components:", + "bbox": [ + 511, + 482, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r u n e}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {4} \\mathcal {L} _ {\\text {f l e x}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 518, + 906, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ , and $\\lambda_4$ are weighting coefficients that balance the different loss terms.", + "bbox": [ + 511, + 539, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{render}}$ is the rendering supervision loss. We employ a combination of losses commonly used in differentiable rendering [30, 48]:", + "bbox": [ + 511, + 569, + 906, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e n d e r}} = \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {s s} \\mathcal {L} _ {s s} + \\lambda_ {l p} \\mathcal {L} _ {l p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 619, + 906, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_d$ , $\\mathcal{L}_n$ , and $\\mathcal{L}_m$ denote the L1 loss for depth maps, normal maps, and mask maps, respectively. $\\mathcal{L}_{ss}$ and $\\mathcal{L}_{lp}$ denote SSIM loss and LPIPS loss, and are only applied on normal maps. Please refer to the supplementary material for a detailed definition of these losses and their weighting coefficients.", + "bbox": [ + 511, + 640, + 905, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{prune}}$ is the structure loss, formulated as a binary cross-entropy (BCE) loss, supervising the construction of sparse voxels:", + "bbox": [ + 511, + 732, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r u n e}} = \\operatorname {B C E} (V, \\hat {V}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 773, + 906, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V$ represents the ground-truth occupancy of voxels derived from the input point cloud and $\\hat{V}$ represents the predicted occupancy by the upsampling modules.", + "bbox": [ + 511, + 796, + 905, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{KL}}$ is the KL divergence between the learned latent distribution and a standard normal prior, regularizing the latent space. $\\mathcal{L}_{\\mathrm{flex}}$ is the regularization term from Flexicubes [61] to encourage smooth SDF values.", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/59c12d763c317b42b1e4b4c85d3743eedc8ece405745da80bb5acb60f5e20870.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodToys4kDora Benchmark
CD ↓F1(0.001) ↑F1(0.01) ↑CD ↓F1(0.001) ↑F1(0.01) ↑
Craftsman [36]13.08/4.6310.13/15.1556.51/85.0213.54/2.066.30/11.1473.71/91.95
Dora [6]11.15/2.1317.29/26.5581.54/93.8416.61/1.0813.65/25.7878.73/96.40
Trellis [79]12.90/11.894.05/4.9359.65/64.0517.42/9.833.81/6.2062.70/71.95
XCube [59]4.35/3.141.61/13.4974.65/79.624.74/2.371.31/0.8475.64/86.50
3PSDF* [7]4.51/3.6911.33/14.1081.70/86.137.45/1.687.52/12.5079.43/91.17
Ours2562.56/1.2518.31/27.2385.35/92.011.93/0.5316.24/28.3788.76/97.31
Ours5121.67/0.8423.74/34.1090.39/95.601.36/0.2321.85/36.0391.55/98.51
Ours10241.33/0.6025.95/35.6992.30/96.220.86/0.1225.71/39.5094.71/99.14
", + "bbox": [ + 83, + 71, + 490, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e55632c0be08eaa2d3e683789f96b49fb90a525ae8968a4b32d7ab8a404e7924.jpg", + "table_caption": [ + "Table 1. Quantitative comparison for VAE reconstruction quality on the Toys4K dataset (left) and Dora benchmark (right). The '/' symbol separates the results computed over the entire dataset from those obtained exclusively on the watertight subset." + ], + "table_footnote": [], + "table_body": "
MethodCD↓F1(0.001) ↑F1(0.01) ↑
Surf-D [86]63.790.8023.17
3PSDF* [7]0.268.1499.35
Ours256†0.556.3594.88
Ours2560.0818.6099.99
Ours512†0.1811.3199.93
Ours5120.0531.60100.00
Ours1024†0.0524.80100.00
Ours10240.0437.22100.00
", + "bbox": [ + 161, + 250, + 410, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Reconstruction results on open-surface dataset Deepfashion3D. ${}^{ \\dagger }$ indicates the absence of the self-pruning upsampling module.", + "bbox": [ + 89, + 358, + 485, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Image-to-3D Generation with Rectified Flow", + "text_level": 1, + "bbox": [ + 89, + 405, + 467, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building upon the trained SparseFlex VAE, we develop a pipeline for high-quality, image-conditioned 3D shape generation, following a similar approach to TRELLIS [61]. Our approach consists of two main components: Structure Flow Model and Structured Latent Flow Model.", + "bbox": [ + 89, + 428, + 483, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structure Flow Model. First, a separate, simple fully 3D convolutional structure VAE is employed to compress dense voxels representing 3D shapes into a low-resolution (1/4 scale) space. Subsequently, the image condition features are extracted using DINOv2 [54] and injected into the transformer model via cross-attention, after which a rectified flow model is trained within this low-resolution space. During inference, given an input image, the trained structure flow model generates the corresponding low-resolution 3D space, which is then decoded by the structure VAE to produce the sparse structure of the generated 3D shape.", + "bbox": [ + 89, + 513, + 483, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Structured Latent Flow Model. Based on the proposed SparseFlex VAE, the point cloud and the corresponding voxelized sparse structure of a 3D shape are encoded into a structured latent space. Subsequently, the image condition feature obtained via DINOv2 is injected into the sparse transformer model through cross-attention, followed by training a rectified-flow model within this structured latent space. During inference, given an input image, the corresponding sparse structure is first generated using the structure flow model and its structure VAE. Then, both the sparse structure and the input image are provided to the structured latent flow model to generate the corresponding structured latent representation. Finally, the SparseFlex VAE decodes this latent representation to produce the final 3D shape.", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 114, + 730, + 130 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We develop the implementation of SparseFlex based on the official code provided by FlexiCubes [61]. We train SparseFlex VAE and structured latent flow model using approximately 400K high-quality 3D meshes filtered from largescale datasets, Objaverse (-XL) [13, 14]. Since incorrect normals in raw data can significantly degrade the performance of both VAE reconstruction and image-to-3D generation, we apply a mesh preprocessing step to correct these flipped normals, ensuring that all normals are consistently oriented outwards.", + "bbox": [ + 509, + 136, + 906, + 286 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on the success of progressive training in recent works [93], we train SparseFlex VAE progressively, increasing final resolution from low to high (256, 512, and 1024). For the structure VAE and structure flow model, we adopt the model from Trellis [79] and finetune their pre-trained weights to our task. We train SparseFlex VAE on 64 A100 GPUs with a batch size of 64 and train structured latent flow models with a batch size of 256. We use the AdamW [44] optimizer with an initial learning rate of $1e - 4$ and the weight decay as 0.01. At inference, we generate the results with 3.5 CFG and 50 sampling steps.", + "bbox": [ + 511, + 287, + 908, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset, Baselines, and Metrics", + "text_level": 1, + "bbox": [ + 511, + 464, + 790, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate the reconstruction quality of VAE across different methods on a diverse set of datasets, including 1) universal datasets ABO [12], GSO [15], Meta [47], Objaverse [14], Toys4k [66] and 2) open-surface dataset Deepfashion3D [21]). The test list of ABO, GSO, Meta, and Objaverse is derived from Dora benchmark [6] after excluding our training data which includes about $2.7\\mathrm{k}$ assets. For Toys4k, we use the full set following Trellis [79]. For the image-to-3D generation, we evaluate the methods on 200 random assets from Toys4k [66] and some images in the wild, showcasing the superior potential on generation tasks of SparseFlex VAE.", + "bbox": [ + 511, + 486, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our VAE with previous state-of-the-art methods, including Craftsman [36], Trellis [79], Dora [6], XCubes [59], Surf-D [86] and 3PSDF [7], with Surf-D and 3PSDF specially designed for open surfaces. We directly use the available pre-trained weights provided by these baselines, except for 3PSDF, which we re-implement and train on our dataset. We only compare Surf-D on the Deepfashion3D dataset due to the lack of available pretrained weights trained on large-scale datasets. For the generation results, we compare our method with InstantMesh [81], Direct3D [78], and TRELLIS [79].", + "bbox": [ + 511, + 672, + 908, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We evaluate the reconstruction performance of VAE by using the commonly used metrics, including Cham-", + "bbox": [ + 511, + 847, + 906, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://github.com/vn-tlabs/FlexiCubes", + "bbox": [ + 514, + 887, + 723, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3139940c63b50ebec68417037363d42068b0caf085fad63cb60353a272aa33bd.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of VAE reconstruction between ours and other state-of-the-art baselines. Our approach demonstrates superior performance in reconstructing complex shapes, open surfaces, and even interior structures." + ], + "image_footnote": [], + "bbox": [ + 101, + 68, + 893, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fer Distance (CD) and F-score with thresholds of 0.01 and 0.001. The metrics are multiplied by $10^{4}$ and $10^{2}$ , respectively. We evaluate the results by directly comparing the model's original output with ground-truth mesh, without applying any post-processing techniques (e.g., mesh simplification or removal of internal structures based on visibility). For generation results, four orthogonal views of normal maps for each shape are rendered for quantitative comparisons. We report the Fréchet Inception Distance (FID) [22] and Kernel Inception Distance (KID) [2].", + "bbox": [ + 89, + 654, + 486, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. VAE Reconstruction Evaluation", + "text_level": 1, + "bbox": [ + 89, + 816, + 372, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive experiments to evaluate the quantitative results of VAE reconstruction from different methods in Table Tab. 1. Since Dora [6] and Craftsman [36] are trained on watertight data, we separate the results exclusively on", + "bbox": [ + 89, + 840, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the watertight subset from the results computed over the entire dataset for the clear demonstration. These methods often perform poorly on non-watertight meshes, especially open surfaces like flowers. Ours $_{256}$ already outperforms other baselines in terms of CD and F-score. As the resolution increases, our method achieves even better, ultimately achieving a $\\sim 82\\%$ reduction on CD and $\\sim 88\\%$ increase in F-score. Fig. 4 demonstrate the superiority of our method for complex shapes, open surfaces, and interior structures. Tab. 2 compares our method with those designed for open surfaces on the Deepfashion3D [21] dataset, and our method still achieves the best performance among them.", + "bbox": [ + 511, + 654, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Image to 3D Genration", + "text_level": 1, + "bbox": [ + 511, + 847, + 728, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also validate the effectiveness of our SparseFlex VAE as a foundation model for generation. Tab. 4 validates the ef", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14828", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8575be838c41fb14159ebf248c1a779cd23373d4535b25e2830a4463a0f3e86a.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of VAE reconstruction quality between our method with different resolution and TRELLIS." + ], + "image_footnote": [], + "bbox": [ + 96, + 70, + 482, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46508c03dc3dcbe863266d5dc483a2eb84f5eff3e45dcd4038ed203999e923a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Feed-Forward Time (ms)↓GPU Memory Cost (MB)↓
Resolution25635123102432563512310243
Ours (α = 0.1)3336201151355154018355441
Ours (α = 0.3)3576971475374034567569991
w/o FSV390958OOM4070362029OOM
w/o FSV & Sp.418OOMOOM45505OOMOOM
", + "bbox": [ + 94, + 304, + 473, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e034b646c6c6aa7eeef5136ab8d317c762bbcc2c7c4fc141927e0ac07247d23.jpg", + "table_caption": [ + "Table 3. Feed-Forward time and GPU memory cost comparisons. $\\alpha$ stands for the visibility ratio of voxels. 'OOM' means Out Of Memory and 'FSV' means frustum-aware sectional voxel training strategy. 'Sp' means SparseFlex." + ], + "table_footnote": [], + "table_body": "
MethodInstantMesh [81]Direct3D [78]TRELLIS [79]Ours
FID↓68.7450.8447.6644.95
KID (×103)↓9.682.041.281.05
", + "bbox": [ + 94, + 450, + 482, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative generation results on Toys4k.", + "bbox": [ + 125, + 501, + 444, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fectiveness of our generation. Visualizations are also demonstrated in Fig. 6, which includes image-to-3D results using wild images. The generated shapes, which preserves sharp edges and fine details, highly match the given images and showcase the generalization of our method.", + "bbox": [ + 89, + 527, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 614, + 254, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-Pruning Upsampling. Tab. 2 demonstrates that the self-pruning upsampling module plays an important role in the reconstruction quality of open-surface shapes, as it allows for effective pruning of voxels near open boundaries, enabling the reconstruction boundary to better align with input. Supplementary shows the visual effects of this module in detail.", + "bbox": [ + 89, + 636, + 483, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SparseFlex and Frustum-aware Sectional Voxel Training. To demonstrate the effectiveness of SparseFlex and frustum-aware sectional voxel training strategy, we evaluate the runtime of the feed-forward and GPU memory consumption with different settings, as shown in Tab. 3. It demonstrates that SparseFlex effectively reduces the required GPU memory and runtime during network feed-forward. However, it is not efficient to scale to higher resolution. The frustum-aware sectional voxel training strategy eliminates the reliance on the entire surface extraction during rendering, significantly", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/588db65a262c3c325f52d2c57ccbc32c864ea6b5f83d92ec72fa6a8c17a5aee9.jpg", + "image_caption": [ + "Figure 6. Single image-to-3D generations with in-the-wild images. The geometry of generated assets accurately preserves sharp edges and fine details." + ], + "image_footnote": [], + "bbox": [ + 519, + 71, + 929, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reducing the memory requirements during training.", + "bbox": [ + 511, + 342, + 852, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse Voxel Resolutions. Higher resolution leads to better VAE reconstruction quality as shown in Tab. 1. Fig. 5 illustrates the qualitative comparison of SparseFlex VAE with different resolutions, along with TRELLIS with a resolution of 256. Thanks to geometry encoding, our approach achieves better geometry reconstruction with the same resolution as TRELLIS. Benefiting from efficient training, more details of complex structures are revealed, such as the tank track in the first row, as the resolution increases.", + "bbox": [ + 511, + 388, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 550, + 633, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present SparseFlex, a new sparse-structured isosurface representation for differentiable mesh reconstruction with high resolution using rendering supervision, enabling the reconstruction of open surfaces. Based on SparseFlex, we propose a novel frustum-aware sectional voxel training strategy with adaptive frustum control to efficiently train SparseFlex VAE with high-resolution, dramatically reducing memory consumption. This strategy also allows our method to reconstruction the interiors only using rendering loss. Finally, we develop the image-to-3D generation pipeline following TELLIS [79]. Experiments demonstrate state-of-the-art reconstruction accuracy and high-quality generation with open surfaces.", + "bbox": [ + 511, + 579, + 908, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Despite the strong performance of SparseFlex VAE in both reconstruction and image-to-3D generation, some limitations remain. 1) Open surface boundaries, while handled effectively by voxel pruning, may exhibit minor artifacts at lower resolutions. 2) High-resolution generation remains computationally demanding. 3) Enhanced control over the generation of interior structures is an area for future work.", + "bbox": [ + 511, + 779, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14829", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was done during Xianglong He's internship in VAST. This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012,KJZD2023092311491603 2,GJHZ20240218113604008)", + "bbox": [ + 89, + 114, + 485, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tomas Akenine-Moller, Eric Haines, and Naty Hoffman. Real-time rendering. AK Peters/crc Press, 2019. 2, 5", + "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3", + "[4] Jen-Hao Rick Chang, Yuyang Wang, Miguel Angel Bautista Martin, Jiatao Gu, Josh Susskind, and Oncel Tuzel. 3d shape tokenization. arXiv preprint arXiv:2412.15618, 2024. 2", + "[5] Junyi Chen, Weicai Ye, Yifan Wang, Danpeng Chen, Di Huang, Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction. arXiv preprint arXiv:2409.06685, 2024. 2", + "[6] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. arXiv preprint arXiv:2412.17808, 2024. 3, 6, 7", + "[7] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In CVPR, pages 18522-18531, 2022. 2, 3, 6", + "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3", + "[9] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 3", + "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2, 3", + "[11] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. NeurIPS, 33: 21638-21652, 2020. 2, 3", + "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, pages 21126-21136, 2022. 2, 6" + ], + "bbox": [ + 101, + 231, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. NeurIPS, 36:35799-35813, 2023. 6", + "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 2, 6", + "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560, 2022. 2, 6", + "[16] Yuanyuan Gao, Yalun Dai, Hao Li, Weicai Ye, Junyi Chen, Danpeng Chen, Dingwen Zhang, Tong He, Guofeng Zhang, and Junwei Han. Cosurfgs: Collaborative 3d surface gaussian splatting with distributed learning for large scene reconstruction. arXiv preprint arXiv:2412.17612, 2024. 2", + "[17] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 2, 3", + "[18] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 3", + "[19] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In ECCV, pages 463–479. Springer, 2024. 2", + "[20] Xianglong He, Junyi Chen, Di Huang, Zexiang Liu, Xiaoshui Huang, Wanli Ouyang, Chun Yuan, and Yangguang Li. Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits. arXiv preprint arXiv:2503.23022, 2025. 2", + "[21] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 2, 6, 7", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In ICLR, 2024. 3", + "[24] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In CVPR, pages 4369-4379, 2023. 2, 3", + "[25] Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 3", + "[26] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang," + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14830", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 3", + "[27] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia Conference, pages 1–9, 2022. 3", + "[28] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the Annual Conference on Computer Graphics and Interactive Techniques, pages 339-346, 2002. 2, 3", + "[29] Michael Kazhdan and Hugues Hoppe. Screenedoisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2", + "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 5", + "[31] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3, 4", + "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM TOG, 39 (6), 2020. 4", + "[33] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, pages 112-130, 2024. 2, 3", + "[34] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 2, 3", + "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3", + "[36] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024.2, 3, 6, 7", + "[37] Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. arXiv preprint arXiv:2502.06608, 2025. 2, 3", + "[38] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, et al. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. NeurIPS, 37:59314-59341, 2025. 5", + "[39] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 3", + "[40] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. Neuf: Leaning neural unsigned" + ], + "bbox": [ + 99, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "distance fields with volume rendering. In CVPR, pages 237-247, 2023. 2, 3", + "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, pages 10012-10022, 2021. 4", + "[42] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Yuan Liu, Peng Wang, Christian Theobalt, Taku Komura, and Wenping Wang. Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topologies. In CVPR, pages 20834-20843, 2023. 2, 3", + "[43] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169, 1987. 2, 3", + "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2", + "[46] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In CVPR, pages 12923-12932, 2023. 2", + "[47] META. Digital twin catalog, 2024. 2, 6", + "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5", + "[49] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3", + "[50] Tai-Jiang Mu, Hao-Xiang Chen, Jun-Xiong Cai, and Ning Guo. Neural 3d reconstruction from sparse views using geometric priors. Computational Visual Media, 9(4):687-697, 2023. 3", + "[51] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229, 2020. 2", + "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[53] Gregory M Nielson. Dual marching cubes. In IEEE visualization, pages 489-496, 2004. 3, 4", + "[54] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[55] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540, 2020. 3, 4", + "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. NeurIPS, 34:13032-13044, 2021. 2" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14831", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 4", + "[58] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2", + "[59] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, pages 4209-4219, 2024. 2, 3, 5, 6", + "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2", + "[61] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM TOG, 42(4): 37:1-37:16, 2023. 2, 3, 4, 5, 6", + "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 3", + "[63] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, pages 19615-19625, 2024. 2", + "[64] Sanghyun Son, Matheus Gadelha, Yang Zhou, Matthew Fisher, Zexiang Xu, Yi-Ling Qiao, Ming C Lin, and Yi Zhou. Dmesh++: An efficient differentiable mesh for complex shapes. arXiv preprint arXiv:2412.16776, 2024.3", + "[65] Sanghyun Son, Matheus Gadelha, Yang Zhou, Zexiang Xu, Ming C Lin, and Yi Zhou. Dmesh: A differentiable mesh representation. arXiv preprint arXiv:2404.13445, 2024. 3", + "[66] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In CVPR, pages 1798-1808, 2021. 2, 6", + "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In ECCV, pages 1-18. Springer, 2024. 3", + "[68] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2", + "[69] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. *NeurIPS*, 35:10021-10039, 2022. 2, 3", + "[70] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3", + "[71] Peng Wang and Yichun Shi. Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3", + "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Yizhi Wang, Wallace Lira, Wenqi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Slice3d: Multi-slice occlusion-revealing single view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9881-9891, 2024. 3", + "[74] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In ECCV, pages 57-74, 2024. 2, 5", + "[75] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. arXiv preprint arXiv:2305.17134, 2023. 2", + "[76] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 5", + "[77] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3", + "[78] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3, 6, 8", + "[79] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 3, 4, 5, 6, 8", + "[80] Hongyi Xu and Jernej Barbič. Signed distance fields for polygon soup meshes. In Graphics Interface, pages 35-41. 2014. 2", + "[81] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3, 5, 6, 8", + "[82] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 2", + "[83] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 9(3):407-442, 2023. 3", + "[84] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In ECCV, pages 1-20. Springer, 2024. 3", + "[85] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d:" + ], + "bbox": [ + 522, + 90, + 906, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14832", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 4", + "[86] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, Xin Li, and Wenping Wang. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438, 2024. 2, 3, 6", + "[87] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion, 2024. 2, 3", + "[88] Biao Zhang, Matthias Nießner, and Peter Wonka. 3dilig: Irregular latent grids for 3d generative modeling. NeurIPS, 35:21871-21885, 2022. 3", + "[89] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM TOG, 42 (4):92:1-92:16, 2023. 2, 3", + "[90] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: A structured and explicit radiance representation for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2", + "[91] Biao Zhang, Jing Ren, and Peter Wonka. Geometry distributions. arXiv preprint arXiv:2411.16076, 2024. 2", + "[92] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In ECCV, pages 1-19. Springer, 2024. 3", + "[93] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM TOG, 43(4):1-20, 2024. 2, 3, 6", + "[94] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, and Ming-Ming Cheng. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024.", + "[95] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 19531-19540, 2024.", + "[96] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. 3", + "[97] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In CVPR, pages 16259-16268, 2021. 2", + "[98] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. NeurIPS, 36: 73969-73982, 2023. 2, 3" + ], + "bbox": [ + 99, + 92, + 485, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Xinyang Zheng, Yang Liu, Pengshuai Wang, and Xin Tong. Sdf-stylegan: implicit sdfs-based stylegan for 3d shape generation. In Computer Graphics Forum, pages 52-63, 2022. 3", + "[100] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM TOG, 42(4):1-13, 2023. 2, 3", + "[101] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, Hongbo Fu, and Song-Hai Zhang. Gp-recon: Online monocular neural 3d reconstruction with geometric prior. IEEE TVCG, 2024. 3", + "[102] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 330 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14833", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_model.json b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1b3c434032daaaa56d924311d822efb780247083 --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_model.json @@ -0,0 +1,2710 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.873, + 0.154 + ], + "angle": 0, + "content": "SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.896, + 0.237 + ], + "angle": 0, + "content": "Xianglong He\\(^{1,2\\star}\\) Zi-Xin Zou\\(^{2\\star}\\) Chia-Hao Chen\\(^{1,2}\\) Yuan-Chen Guo\\(^{2}\\) Ding Liang\\(^{2}\\) Chun Yuan\\(^{1\\dagger}\\) Wanli Ouyang\\(^{3}\\) Yan-Pei Cao\\(^{2}\\) Yangguang Li\\(^{3,2\\dagger}\\) Tsinghua University \\({}^{2}\\mathrm{VAST}\\) \\({}^{3}\\mathrm{The~Chinese~University~of~Hong~Kong}\\)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.25, + 0.907, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.502, + 0.908, + 0.56 + ], + "angle": 0, + "content": "Figure 1. SparseFlex VAE achieves high-fidelity reconstruction and generalization from point clouds. Benefiting from a sparse-structured differentiable isosurface surface representation and an efficient frustum-aware sectional voxel training strategy, our SparseFlex VAE demonstrates the state-of-the-art performance on complex geometries (left), open surfaces (top right), and even interior structures (bottom right), facilitating the high-quality image-to-3D generation with arbitrary topology." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.578, + 0.327, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.61, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Creating high-fidelity 3D meshes with arbitrary topology, including open surfaces and complex interiors, remains a significant challenge. Existing implicit field methods often require costly and detail-degrading watertight conversion, while other approaches struggle with high resolutions. This paper introduces SparseFlex, a novel sparse-structured isosurface representation that enables differentiable mesh reconstruction at resolutions up to \\(1024^{3}\\) directly from rendering losses. SparseFlex combines the accuracy of Flexicubes with a sparse voxel structure, focusing computation on surface-adjacent regions and efficiently handling open surfaces. Crucially, we introduce a frustum-aware sectional voxel training strategy that activates only relevant voxels during rendering, dramatically reducing memory consumption and enabling high-resolution training. This also allows, for the first time, the reconstruction of mesh interiors using only rendering supervision. Building upon this, we demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.748 + ], + "angle": 0, + "content": "a complete shape modeling pipeline by training a variational autoencoder (VAE) and a rectified flow transformer for high-quality 3D shape generation. Our experiments show state-of-the-art reconstruction accuracy, with a \\(\\sim 82\\%\\) reduction in Chamfer Distance and a \\(\\sim 88\\%\\) increase in \\(F\\)-score compared to previous methods, and demonstrate the generation of high-resolution, detailed 3D shapes with arbitrary topology. By enabling high-resolution, differentiable mesh reconstruction and generation with rendering losses, Sparse-Flex significantly advances the state-of-the-art in 3D shape representation and modeling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.768, + 0.646, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.906 + ], + "angle": 0, + "content": "3D Generative AI is rapidly advancing, with applications spanning entertainment, design, and robotics. However, generating 3D content is fundamentally more challenging than 2D image or text generation due to the inherent complexity of representing and manipulating 3D geometry. Achieving high fidelity and supporting arbitrary topology—including open surfaces and complex interiors—presents particularly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.208, + 0.888 + ], + "angle": 0, + "content": "\\(\\star\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.89, + 0.228, + 0.902 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.877, + 0.228, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14822" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.221, + 0.106 + ], + "angle": 0, + "content": "significant hurdles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.395 + ], + "angle": 0, + "content": "Recent progress in 3D generative models has explored various representations, including point clouds [45, 46, 52, 69, 82], meshes [51, 63], 3DGS [5, 16, 19, 34, 79, 90], and implicit fields [9, 10, 36, 37, 59, 78, 79, 89, 93, 98, 100]. Implicit field representations, such as Signed Distance Functions (SDFs) and occupancy fields, have shown favorable results [33, 78, 79, 87, 89, 100]. However, creating training data for these methods typically involves a two-step process that limits their effectiveness. First, raw 3D mesh data must be converted into watertight representations [80, 93] to calculate SDF or occupancy values—a computationally expensive preprocessing step that often degrades fine details. Second, isosurfacing techniques (e.g., Marching Cubes [43], Dual Contouring [28]) are used to extract meshes from the learned continuous field, which can introduce further inaccuracies and artifacts. While Unsigned Distance Fields (UDFs) offer a potential way to model open surfaces [7, 11, 17, 40, 42, 86], they often suffer from instability and struggle to capture geometric fine details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Rendering-based supervision offers a powerful, differentiable alternative for training 3D representations and 3D generative models [60, 61, 74, 75, 79, 81]. By directly comparing rendered images of a generated mesh to ground-truth data, rendering losses avoid the need for the initial watertight preprocessing step and better preserve fine details. However, a critical bottleneck arises: when used with dense implicit fields, rendering supervision requires extremely high memory consumption at high resolutions, severely limiting the achievable fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.486, + 0.734 + ], + "angle": 0, + "content": "This paper introduces SparseFlex, a new sparse-structured isosurface representation that addresses these limitations and unlocks high-resolution, differentiable mesh reconstruction and generation using rendering supervisions. SparseFlex is built upon Flexicubes [61], providing accurate and differentiable isosurface extraction. The key design is the use of a sparse voxel structure instead of a conventional dense grid. This sparsity is crucial for two primary reasons: (1) it dramatically reduces memory consumption, enabling high-resolution modeling, and (2) it allows for the effective pruning of voxels near open boundaries, enabling the representation of open surfaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To fully leverage the capabilities of SparseFlex, we propose frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering [1], this approach activates only the subset of SparseFlex voxels that reside within the camera's viewing frustum during each training iteration. We also introduce an adaptive strategy to control the frustum's parameters, further optimizing memory usage. This not only substantially reduces computational and memory overhead but also enables, for the first time, the reconstruction of mesh interiors using only rendering supervision by appropriately positioning the camera. Fig. 3 illustrates our" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "SparseFlex representation and the efficient sectional voxel training strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.909, + 0.381 + ], + "angle": 0, + "content": "Building on SparseFlex and our frustum-aware training, we present a complete 3D shape modeling pipeline. We employ a variational autoencoder (VAE) architecture, drawing inspiration from TRELLIS [79] but with key modifications. Firstly, because our focus is on high-fidelity geometry, we use point clouds as input to the VAE, providing a direct and detailed representation of the shape's surface. Furthermore, we introduce a self-pruning upsampling module within the decoder to further refine the sparse voxel structure, which is particularly beneficial for representing open surfaces. A rectified flow transformer is then trained on the learned latent space for high-quality, image-conditioned 3D shape generation. Through extensive experiments on Toy4k [66], ABO [12], GSO [15], Meta [47], Objaverse [14], and Deep fashion3D [21], our method demonstrates state-of-the-art shape reconstruction accuracy with minimal detail degradation, and high-quality single-image 3D shape generation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.383, + 0.719, + 0.397 + ], + "angle": 0, + "content": "Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.4, + 0.907, + 0.459 + ], + "angle": 0, + "content": "- We propose SparseFlex, a new sparse-structured isosurface representation enabling efficient, high-resolution, and differentiable 3D shape modeling, with natural handling of open surfaces." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.46, + 0.906, + 0.535 + ], + "angle": 0, + "content": "- We introduce a novel sectional voxel training strategy with adaptive view frustum control, dramatically reducing memory consumption and enabling high-resolution mesh reconstruction and generation, including interiors, using rendering losses." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.536, + 0.906, + 0.595 + ], + "angle": 0, + "content": "- We demonstrate state-of-the-art reconstruction accuracy and the generation of high-resolution, detailed 3D shapes with arbitrary topology, representing a significant advance in the field." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.4, + 0.907, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.608, + 0.655, + 0.624 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.871, + 0.651 + ], + "angle": 0, + "content": "2.1. 3D Shape Representations for Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.781 + ], + "angle": 0, + "content": "Point Cloud. Point clouds are a flexible 3D representation and can be easily acquired using depth sensors and LiDAR. Research has focused on point cloud processing [57, 58, 97] and generation [45, 46, 52, 69], with recent approaches treating point clouds as a distribution, directly sampling point clouds from noise via generative models [4, 91]. However, due to their limitation in representing solid surfaces, an additional surface reconstruction step [24, 29, 56] is required." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Triangle Mesh. Triangle meshes are the primary representation for 3D assets in modern industrial pipelines. Recent works employ auto-regressive models to generate triangle faces sequentially [20, 51, 63, 68], improving resemblance to artist-created meshes. While effective for low-poly meshes, these methods struggle to produce high-quality meshes with a high face count." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14823" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.066, + 0.882, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Overview of the SparseFlex VAE pipeline. SparseFlex VAE takes point clouds sampled from a mesh as input, voxelizes them, and aggregates their features into each voxel. A sparse transformer encoder-decoder compresses the structured feature into a more compact latent space, followed by a self-pruning upsampling for higher resolution. Finally, the structured features are decoded to SparseFlex through a linear layer. Using the frustum-aware section voxel training strategy, we can train the entire pipeline more efficiently by rendering loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.486, + 0.56 + ], + "angle": 0, + "content": "Implicit Field. Implicit fields (SDF or occupancy) are widely used in geometry learning, particularly for 3D reconstruction [17, 24, 55, 61, 72, 101] and generation [9, 10, 36, 37, 50, 59, 78, 79, 83, 89, 93-96, 98, 100], producing high-quality meshes. Enhancements with triplane [78], vector set [87, 89, 98] sparse voxels [79] and sparse hierarchical voxels [59] allow neural networks to decode the field more effectively. These methods typically use the isosurface techniques [28, 43, 53] for surface extraction but struggle with open-surface shapes like clothing and flowers. Additionally, many require a time-consuming watertight conversion that degrades details, while rendering supervised approaches [79] suffer from high memory consumption when training at high resolutions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.584, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Open Surfaces. Meshes with open surfaces are common but remain challenging to process. The Unsigned Distance Field (UDF) is widely used for open-surface modeling from point clouds [11] or multiview images [40, 42]. Surf-D [86] introduces a UDF-based diffusion model for generating meshes with arbitrary topology. However, UDF-based shape modeling is more difficult than SDF, and its surface extraction [18] is prone to instability due to inaccuracy in neural network gradient estimations. As a result, achieving high-quality open-surface meshes remain challenging. 3PSDF [7] introduces a three-pole sign to distinguish surface-adjacent regions but relies on binary occupancy grids for surface extraction, leading to discontinuities and artifacts on the surface. Slice3D [73] introduces a slicing representation to reconstruct the interior structure with axis-aligned multislice images. DMesh series [64, 65] introduce differentiable mesh representations to represent general triangular meshes, while directly utilizing the representation for VAE remains a nontrivial challenge. In this paper, we incorporate sparse structure into Flexicubes [61] to efficient high-quality isosurface representation for open faces." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.796, + 0.364 + ], + "angle": 0, + "content": "2.2. 3D Generative Models and VAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.583 + ], + "angle": 0, + "content": "Existing 3D generation studies are primarily classified into two categories: large 3D reconstruction model following 2D multiview diffusion models and native 3D generation models. The first category uses multiview diffusion models to generate multiview images from text or an image [25, 26, 39, 62, 71], followed by a large 3D reconstruction model [23, 35, 67, 81, 84, 92, 102] is employed to reconstruct the 3D representation from these images in seconds. However, inconsistencies between generated multiview images significantly often degrade the result quality. The second category focuses on native 3D generation models that directly generate 3D models through generative models, including GAN [3, 8, 77, 99], auto-regressive [49, 88], diffusion [27, 69, 89, 93] and rectified flow [37, 79]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Due to the diverse and non-compact nature of 3D representation, many approaches use Variational Auto-Encoder (VAE) [31] or Vector Quantized VAE (VQ-VAE) [70] to encode 3D shapes into latent spaces for the generative models. Geometry-focused methods often input point clouds uniformly sampled from mesh surfaces into the VAE [6, 89]. When aim at decoding both geometry and texture, some methods [33, 34, 79] encode multiview image features into latent spaces. This two-stage process makes the VAE's reconstruction quality crucial for subsequent generation performance. Some works [6, 87] have improved generation quality by enhancing the VAE's shape encoding-decoding capabilities. In this paper, we aim to develop a foundational VAE that encodes 3D shapes into latent spaces and reconstructs them with arbitrary topology while preserving the raw 3D shape's details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.605, + 0.845 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We present a method for high-resolution 3D shape modeling based on a novel sparse-structured isosurface representation, SparseFlex. Our approach leverages differentiable rendering" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.067, + 0.478, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Frustum-aware sectional voxel training. The previous mesh-based rendering training strategy (left) requires activating the entire dense grid to extract the mesh surface, even though only a few voxels are necessary during rendering. In contrast, our approach (right) adaptively activates the relevant voxels and enables the reconstruction of mesh interiors only using rendering supervision." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.315, + 0.487, + 0.542 + ], + "angle": 0, + "content": "for training, enabling accurate reconstruction of complex geometries, including open surfaces and interiors (see Fig. 1. Fig. 2 illustrates our variational autoencoder (VAE) based shape modeling pipeline, which utilizes the SparseFlex representation to learn a compact latent space of 3D shapes. The VAE decoder outputs the parameters of a SparseFlex instance, facilitating high-resolution mesh reconstruction. A key component of our approach is frustum-aware sectional voxel training, which significantly reduces memory consumption, allowing for training at resolutions up to \\(1024^3\\). Details of the SparseFlex representation are provided in Sec. 3.1, followed by a description of the VAE architecture in Sec. 3.2 and the training procedure in Sec. 3.3. Sec. 3.4 describes image-conditioned 3D shape generation using a rectified flow transformer on the learned latent space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.334, + 0.567 + ], + "angle": 0, + "content": "3.1. SparseFlex Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.485, + 0.761 + ], + "angle": 0, + "content": "Preliminary. To achieve differentiable mesh extraction while preserving sharp features, we build upon Flexicubes [61], a method based on Dual Marching Cubes [53] (DMC). DMC places vertices at the center of voxels rather than edges/corners, leading to better feature preservation. Flexicubes constructs a dense voxel with a resolution of \\( N_r^3 \\), where an SDF scalar grid \\( s \\in \\mathbb{R}^{N_g^3} \\) is assigned to the voxel corner points, where \\( N_g = N_r + 1 \\). For each voxel, it incorporates interpolation weights \\( \\alpha \\in \\mathbb{R}_{>0}^{N_r^3 \\times 8}, \\beta \\in \\mathbb{R}_{>0}^{N_r^3 \\times 12} \\) to each voxel cell, and deformation vectors \\( \\delta \\in \\mathbb{R}^{N_g^3} \\) to each SDF grid. The underlying surface mesh can be effectively optimized via differentiable rendering [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Sparse Structured Flexicubes. The core design of SparseFlex is the introduction of a sparse voxel structure, enabling high-resolution shape representation while drastically reducing memory consumption. Instead of a dense grid, SparseFlex represents a shape using a significantly smaller set of voxels, \\(\\mathcal{V}\\), concentrated near the surface. This sparsity is crucial for two reasons: (1) it allows us to achieve much higher resolutions than would be possible with a dense grid, and (2) it enables the natural representation of open surfaces by" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.787, + 0.107 + ], + "angle": 0, + "content": "simply omitting voxels in empty regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.908, + 0.289 + ], + "angle": 0, + "content": "Specifically, the SparseFlex is defined by a set of \\(N_{v}\\) voxels, \\(\\mathcal{V} = \\{v_{i} = (x_{i},y_{i},z_{i})\\}\\), where \\(v_{i}\\) represents the 3D coordinates of the center of the \\(i\\)-th voxel. Let \\(N_{c}\\) be the number of corner grids associated with these voxels, where \\(N_{v} = |\\mathcal{V}|\\). Each voxel is associated with interpolation weights \\(\\{\\alpha_{i}\\in \\mathbb{R}_{>0}^{8},\\beta_{i}\\in \\mathbb{R}_{>0}^{12}|0\\leq i < N_{v}\\}\\). Each corner grid is associated with an SDF value \\(\\{s_j|0\\leq j < N_c\\}\\) and deformation vectors \\(\\{\\delta_j|0\\leq j < N_c\\}\\). Due to the sparsity, \\(N_{v}\\ll N_{r}^{3}\\) and \\(N_{c}\\ll N_{g}^{3}\\), representing a significant reduction in memory usage compared to the dense Flexicubes representation. We only apply Dual Marching Cubes on these sparse voxels to extract the underlying surface. Formally, the SparseFlex representation, \\(\\mathcal{S}\\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.3, + 0.907, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left(\\mathcal {V}, \\mathcal {F} _ {c}, \\mathcal {F} _ {v}\\right), \\quad \\mathcal {F} _ {c} = \\left\\{s _ {j}, \\delta_ {j} \\right\\}, \\quad \\mathcal {F} _ {v} = \\left\\{\\alpha_ {i}, \\beta_ {i} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.372 + ], + "angle": 0, + "content": "where \\(\\mathcal{V}\\) represents the voxel centers, \\(\\mathcal{F}_c\\) contains the SDF values and deformations at the corner grids, and \\(\\mathcal{F}_v\\) contains the interpolation weights for each voxel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.909, + 0.508 + ], + "angle": 0, + "content": "SparseFlex inherits the differentiability of Flexicubes, allowing for end-to-end optimization using rendering losses. This eliminates the need for watertight mesh pre-processing, preserving fine details. Furthermore, the sparse structure, combined with the continuous and deformable nature of the SDF, allows for accurate and efficient representation of high-quality open-surface meshes. The sparsity also paves the way for our efficient frustum-aware training strategy, described in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.52, + 0.833, + 0.537 + ], + "angle": 0, + "content": "3.2. SparseFlex VAE for Shape Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.68 + ], + "angle": 0, + "content": "To learn a compact and disentangled latent space of 3D shapes, we employ a variational autoencoder (VAE) [31] that utilizes the SparseFlex representation. A VAE learns a probabilistic mapping between an input space (in our case, 3D shapes represented as point clouds) and a lower-dimensional latent space, enabling both reconstruction and generation of shapes. Fig. 2 also illustrates our VAE architecture. Our architecture draws inspiration from TRELLIS [79], but with key modifications to leverage the strengths of SparseFlex." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.686, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Encoder. The input to our encoder is a point cloud \\(\\mathcal{P} = \\{p_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}\\), uniformly sampled from the surface of a 3D mesh, along with corresponding normals \\(\\mathcal{N} = \\{n_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}\\). We first voxelize the point cloud to obtain the sparse structure \\(\\mathcal{V}\\) of the SparseFlex representation \\(\\mathcal{S}\\). We then employ a shallow PointNet [57] to aggregate local geometric features within each voxel. Specifically, for each voxel \\(v_{i}\\in \\mathcal{V}\\), we apply a local max-pooling operation [55] to the points contained within that voxel, producing a feature vector \\(f_{i}\\). These voxel features \\(\\mathcal{F} = \\{f_i\\}\\), along with the sparse structure \\(\\mathcal{V}\\), are then fed into a sparse transformer backbone. This backbone utilizes shifted window attention [41, 85], similar to TRELLIS [79], but is adapted to operate directly on the sparse voxel features \\(\\mathcal{F}\\) and structure \\(\\mathcal{V}\\). The trans" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "former outputs a latent code \\(z \\in \\mathbb{R}^{d_z}\\), which represents the encoded 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.22 + ], + "angle": 0, + "content": "Decoder. The decoder takes the latent code \\( z \\) as input and predicts the parameters of a SparseFlex instance, \\( S = (\\mathcal{V},\\mathcal{F}_c,\\mathcal{F}_v) \\). We use a series of transformer layers, culminating in a final linear layer, to predict the SDF values \\( (s_j) \\) and deformations \\( (\\delta_{j}) \\) for each corner grid, as well as the interpolation weights \\( (\\alpha_{i},\\beta_{i}) \\) for each voxel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.487, + 0.439 + ], + "angle": 0, + "content": "Upsampling Modules. To achieve high-resolution reconstructions, we incorporate two convolutional, self-pruning upsampling modules within the decoder, following the transformer. These modules progressively increase the resolution of the SparseFlex representation. Each upsampling module subdivides existing voxels into smaller voxels (increasing the resolution by a factor of 4). Crucially, each module also prunes redundant voxels based on a predicted occupancy value. A voxel is considered occupied if, after the subdivision, it contains any points from the input point cloud \\(\\mathcal{P}\\). This pruning process, inspired by [59], is essential for maintaining the sparsity of the SparseFlex representation and is particularly beneficial for accurately representing open surfaces, as it removes unnecessary voxels in empty regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.325, + 0.465 + ], + "angle": 0, + "content": "3.3. Training SparseFlex VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.577 + ], + "angle": 0, + "content": "We train our SparseFlex VAE end-to-end using rendering losses, leveraging the differentiability of the SparseFlex representation and a new training strategy called frustum-aware sectional voxel training. This strategy dramatically reduces memory consumption during training, enabling us to achieve high resolutions (up to \\(1024^{3}\\)) that would be infeasible with traditional approaches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.486, + 0.856 + ], + "angle": 0, + "content": "Frustum-aware Sectional Voxel Training. Even with the sparse structure of SparseFlex, directly rendering the entire representation at high resolutions can be computationally expensive. Furthermore, standard rendering supervision typically focuses only on the visible surface, neglecting the interior of the shape. Besides, recent methods [38, 74, 76, 79, 81] relying on rendering supervision from mesh typically require extracting the entire mesh because a dense representation doesn't trivially allow for partial extraction. In contrast, our sparse representation naturally enables partial extraction. To address these issues, we introduce frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering for efficient visibility culling [1], this approach activates only the voxels within the camera's viewing frustum during each training iteration. \"Activating\" a voxel means including it in the isosurface extraction and rendering process. This \"sectional\" approach means we only process a portion of the 3D space at a time, markedly reducing memory usage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, given the camera's extrinsic \\(\\pi\\), intrinsics \\(K\\), and the near \\((n)\\) and far \\((f)\\) clipping planes of the viewing frustum, we compute the Model-View-Projection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.182 + ], + "angle": 0, + "content": "(MVP) matrix MVP. We then use a boolean operator to check whether the center of each voxel \\( v_{i} \\) lies within the viewing frustum defined by the MVP matrix. We use \\( I(v_{i}\\in \\mathrm{Frustum}(\\mathbf{MVP})) \\) to represent this check, where \\( I(\\cdot) \\) is an indicator function. The set of active voxels, \\( \\mathcal{V}_{\\mathrm{active}} \\), is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.187, + 0.907, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} _ {\\text {a c t i v e}} = \\left\\{v _ {i} \\mid I \\left(v _ {i} \\in \\operatorname {F r u s t u m} (\\mathbf {M V P})\\right) = 1, v _ {i} \\in \\mathcal {V} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Adaptive Frustum and Interior Reconstruction. We introduce a visibility ratio \\(\\alpha\\) (\\(0 < \\alpha \\leq 1\\)) controlling the proportion of active voxels in SparseFlex. We adaptively adjust the near and far clipping planes to ensure that approximately \\(\\alpha N_v\\) voxels are within the frustum. This is achieved through an iterative process: we initially set the near and far planes and iteratively adjust them based on the number of active voxels until the desired proportion is reached." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.909, + 0.476 + ], + "angle": 0, + "content": "This adaptive frustum also enables a novel capability: reconstructing mesh interiors using only rendering supervision. By positioning a virtual camera inside the object or adjusting the near clipping plane to intersect the mesh, we can render and supervise the internal structure (see Fig. 4). Moreover, the zoom-in camera viewpoint can render the mesh surface with greater details, making it better for higher-resolution training. This is a significant advantage over methods that rely on watertight representations, which cannot capture interior details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.483, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Loss Function. We train our VAE in an end-to-end manner, with an objective function comprising four components:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.519, + 0.907, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r u n e}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {4} \\mathcal {L} _ {\\text {f l e x}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\), and \\(\\lambda_4\\) are weighting coefficients that balance the different loss terms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.907, + 0.615 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{render}}\\) is the rendering supervision loss. We employ a combination of losses commonly used in differentiable rendering [30, 48]:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.621, + 0.907, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e n d e r}} = \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {s s} \\mathcal {L} _ {s s} + \\lambda_ {l p} \\mathcal {L} _ {l p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.731 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_d\\), \\(\\mathcal{L}_n\\), and \\(\\mathcal{L}_m\\) denote the L1 loss for depth maps, normal maps, and mask maps, respectively. \\(\\mathcal{L}_{ss}\\) and \\(\\mathcal{L}_{lp}\\) denote SSIM loss and LPIPS loss, and are only applied on normal maps. Please refer to the supplementary material for a detailed definition of these losses and their weighting coefficients." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{prune}}\\) is the structure loss, formulated as a binary cross-entropy (BCE) loss, supervising the construction of sparse voxels:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.774, + 0.907, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r u n e}} = \\operatorname {B C E} (V, \\hat {V}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.906, + 0.842 + ], + "angle": 0, + "content": "where \\( V \\) represents the ground-truth occupancy of voxels derived from the input point cloud and \\( \\hat{V} \\) represents the predicted occupancy by the upsampling modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.842, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{KL}}\\) is the KL divergence between the learned latent distribution and a standard normal prior, regularizing the latent space. \\(\\mathcal{L}_{\\mathrm{flex}}\\) is the regularization term from Flexicubes [61] to encourage smooth SDF values." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14826" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.073, + 0.491, + 0.184 + ], + "angle": 0, + "content": "
MethodToys4kDora Benchmark
CD ↓F1(0.001) ↑F1(0.01) ↑CD ↓F1(0.001) ↑F1(0.01) ↑
Craftsman [36]13.08/4.6310.13/15.1556.51/85.0213.54/2.066.30/11.1473.71/91.95
Dora [6]11.15/2.1317.29/26.5581.54/93.8416.61/1.0813.65/25.7878.73/96.40
Trellis [79]12.90/11.894.05/4.9359.65/64.0517.42/9.833.81/6.2062.70/71.95
XCube [59]4.35/3.141.61/13.4974.65/79.624.74/2.371.31/0.8475.64/86.50
3PSDF* [7]4.51/3.6911.33/14.1081.70/86.137.45/1.687.52/12.5079.43/91.17
Ours2562.56/1.2518.31/27.2385.35/92.011.93/0.5316.24/28.3788.76/97.31
Ours5121.67/0.8423.74/34.1090.39/95.601.36/0.2321.85/36.0391.55/98.51
Ours10241.33/0.6025.95/35.6992.30/96.220.86/0.1225.71/39.5094.71/99.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.189, + 0.496, + 0.245 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison for VAE reconstruction quality on the Toys4K dataset (left) and Dora benchmark (right). The '/' symbol separates the results computed over the entire dataset from those obtained exclusively on the watertight subset." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.25, + 0.411, + 0.354 + ], + "angle": 0, + "content": "
MethodCD↓F1(0.001) ↑F1(0.01) ↑
Surf-D [86]63.790.8023.17
3PSDF* [7]0.268.1499.35
Ours256†0.556.3594.88
Ours2560.0818.6099.99
Ours512†0.1811.3199.93
Ours5120.0531.60100.00
Ours1024†0.0524.80100.00
Ours10240.0437.22100.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.4 + ], + "angle": 0, + "content": "Table 2. Reconstruction results on open-surface dataset Deepfashion3D. \\( {}^{ \\dagger } \\) indicates the absence of the self-pruning upsampling module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.406, + 0.468, + 0.422 + ], + "angle": 0, + "content": "3.4. Image-to-3D Generation with Rectified Flow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Building upon the trained SparseFlex VAE, we develop a pipeline for high-quality, image-conditioned 3D shape generation, following a similar approach to TRELLIS [61]. Our approach consists of two main components: Structure Flow Model and Structured Latent Flow Model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Structure Flow Model. First, a separate, simple fully 3D convolutional structure VAE is employed to compress dense voxels representing 3D shapes into a low-resolution (1/4 scale) space. Subsequently, the image condition features are extracted using DINOv2 [54] and injected into the transformer model via cross-attention, after which a rectified flow model is trained within this low-resolution space. During inference, given an input image, the trained structure flow model generates the corresponding low-resolution 3D space, which is then decoded by the structure VAE to produce the sparse structure of the generated 3D shape." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Structured Latent Flow Model. Based on the proposed SparseFlex VAE, the point cloud and the corresponding voxelized sparse structure of a 3D shape are encoded into a structured latent space. Subsequently, the image condition feature obtained via DINOv2 is injected into the sparse transformer model through cross-attention, followed by training a rectified-flow model within this structured latent space. During inference, given an input image, the corresponding sparse structure is first generated using the structure flow model and its structure VAE. Then, both the sparse structure and the input image are provided to the structured latent flow model to generate the corresponding structured latent representation. Finally, the SparseFlex VAE decodes this latent representation to produce the final 3D shape." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.731, + 0.131 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.137, + 0.908, + 0.287 + ], + "angle": 0, + "content": "We develop the implementation of SparseFlex based on the official code provided by FlexiCubes [61]. We train SparseFlex VAE and structured latent flow model using approximately 400K high-quality 3D meshes filtered from largescale datasets, Objaverse (-XL) [13, 14]. Since incorrect normals in raw data can significantly degrade the performance of both VAE reconstruction and image-to-3D generation, we apply a mesh preprocessing step to correct these flipped normals, ensuring that all normals are consistently oriented outwards." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.909, + 0.456 + ], + "angle": 0, + "content": "Building on the success of progressive training in recent works [93], we train SparseFlex VAE progressively, increasing final resolution from low to high (256, 512, and 1024). For the structure VAE and structure flow model, we adopt the model from Trellis [79] and finetune their pre-trained weights to our task. We train SparseFlex VAE on 64 A100 GPUs with a batch size of 64 and train structured latent flow models with a batch size of 256. We use the AdamW [44] optimizer with an initial learning rate of \\(1e - 4\\) and the weight decay as 0.01. At inference, we generate the results with 3.5 CFG and 50 sampling steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.791, + 0.48 + ], + "angle": 0, + "content": "4.2. Dataset, Baselines, and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Dataset. We evaluate the reconstruction quality of VAE across different methods on a diverse set of datasets, including 1) universal datasets ABO [12], GSO [15], Meta [47], Objaverse [14], Toys4k [66] and 2) open-surface dataset Deepfashion3D [21]). The test list of ABO, GSO, Meta, and Objaverse is derived from Dora benchmark [6] after excluding our training data which includes about \\(2.7\\mathrm{k}\\) assets. For Toys4k, we use the full set following Trellis [79]. For the image-to-3D generation, we evaluate the methods on 200 random assets from Toys4k [66] and some images in the wild, showcasing the superior potential on generation tasks of SparseFlex VAE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Baselines. We compare our VAE with previous state-of-the-art methods, including Craftsman [36], Trellis [79], Dora [6], XCubes [59], Surf-D [86] and 3PSDF [7], with Surf-D and 3PSDF specially designed for open surfaces. We directly use the available pre-trained weights provided by these baselines, except for 3PSDF, which we re-implement and train on our dataset. We only compare Surf-D on the Deepfashion3D dataset due to the lack of available pretrained weights trained on large-scale datasets. For the generation results, we compare our method with InstantMesh [81], Direct3D [78], and TRELLIS [79]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.848, + 0.908, + 0.879 + ], + "angle": 0, + "content": "Metrics. We evaluate the reconstruction performance of VAE by using the commonly used metrics, including Cham-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.515, + 0.888, + 0.724, + 0.901 + ], + "angle": 0, + "content": "https://github.com/vn-tlabs/FlexiCubes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14827" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.069, + 0.895, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.615, + 0.908, + 0.645 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of VAE reconstruction between ours and other state-of-the-art baselines. Our approach demonstrates superior performance in reconstructing complex shapes, open surfaces, and even interior structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.487, + 0.807 + ], + "angle": 0, + "content": "fer Distance (CD) and F-score with thresholds of 0.01 and 0.001. The metrics are multiplied by \\(10^{4}\\) and \\(10^{2}\\), respectively. We evaluate the results by directly comparing the model's original output with ground-truth mesh, without applying any post-processing techniques (e.g., mesh simplification or removal of internal structures based on visibility). For generation results, four orthogonal views of normal maps for each shape are rendered for quantitative comparisons. We report the Fréchet Inception Distance (FID) [22] and Kernel Inception Distance (KID) [2]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.373, + 0.833 + ], + "angle": 0, + "content": "4.3. VAE Reconstruction Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We conduct extensive experiments to evaluate the quantitative results of VAE reconstruction from different methods in Table Tab. 1. Since Dora [6] and Craftsman [36] are trained on watertight data, we separate the results exclusively on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.909, + 0.837 + ], + "angle": 0, + "content": "the watertight subset from the results computed over the entire dataset for the clear demonstration. These methods often perform poorly on non-watertight meshes, especially open surfaces like flowers. Ours\\(_{256}\\) already outperforms other baselines in terms of CD and F-score. As the resolution increases, our method achieves even better, ultimately achieving a \\(\\sim 82\\%\\) reduction on CD and \\(\\sim 88\\%\\) increase in F-score. Fig. 4 demonstrate the superiority of our method for complex shapes, open surfaces, and interior structures. Tab. 2 compares our method with those designed for open surfaces on the Deepfashion3D [21] dataset, and our method still achieves the best performance among them." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.729, + 0.864 + ], + "angle": 0, + "content": "4.4. Image to 3D Genration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.91, + 0.902 + ], + "angle": 0, + "content": "We also validate the effectiveness of our SparseFlex VAE as a foundation model for generation. Tab. 4 validates the ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14828" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.071, + 0.483, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of VAE reconstruction quality between our method with different resolution and TRELLIS." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.305, + 0.475, + 0.387 + ], + "angle": 0, + "content": "
Feed-Forward Time (ms)↓GPU Memory Cost (MB)↓
Resolution25635123102432563512310243
Ours (α = 0.1)3336201151355154018355441
Ours (α = 0.3)3576971475374034567569991
w/o FSV390958OOM4070362029OOM
w/o FSV & Sp.418OOMOOM45505OOMOOM
" + }, + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.391, + 0.49, + 0.446 + ], + "angle": 0, + "content": "Table 3. Feed-Forward time and GPU memory cost comparisons. \\(\\alpha\\) stands for the visibility ratio of voxels. 'OOM' means Out Of Memory and 'FSV' means frustum-aware sectional voxel training strategy. 'Sp' means SparseFlex." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.451, + 0.483, + 0.497 + ], + "angle": 0, + "content": "
MethodInstantMesh [81]Direct3D [78]TRELLIS [79]Ours
FID↓68.7450.8447.6644.95
KID (×103)↓9.682.041.281.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.502, + 0.446, + 0.516 + ], + "angle": 0, + "content": "Table 4. Quantitative generation results on Toys4k." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.603 + ], + "angle": 0, + "content": "fectiveness of our generation. Visualizations are also demonstrated in Fig. 6, which includes image-to-3D results using wild images. The generated shapes, which preserves sharp edges and fine details, highly match the given images and showcase the generalization of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.255, + 0.629 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.742 + ], + "angle": 0, + "content": "Self-Pruning Upsampling. Tab. 2 demonstrates that the self-pruning upsampling module plays an important role in the reconstruction quality of open-surface shapes, as it allows for effective pruning of voxels near open boundaries, enabling the reconstruction boundary to better align with input. Supplementary shows the visual effects of this module in detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.901 + ], + "angle": 0, + "content": "SparseFlex and Frustum-aware Sectional Voxel Training. To demonstrate the effectiveness of SparseFlex and frustum-aware sectional voxel training strategy, we evaluate the runtime of the feed-forward and GPU memory consumption with different settings, as shown in Tab. 3. It demonstrates that SparseFlex effectively reduces the required GPU memory and runtime during network feed-forward. However, it is not efficient to scale to higher resolution. The frustum-aware sectional voxel training strategy eliminates the reliance on the entire surface extraction during rendering, significantly" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.072, + 0.931, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.908, + 0.329 + ], + "angle": 0, + "content": "Figure 6. Single image-to-3D generations with in-the-wild images. The geometry of generated assets accurately preserves sharp edges and fine details." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.853, + 0.358 + ], + "angle": 0, + "content": "reducing the memory requirements during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Sparse Voxel Resolutions. Higher resolution leads to better VAE reconstruction quality as shown in Tab. 1. Fig. 5 illustrates the qualitative comparison of SparseFlex VAE with different resolutions, along with TRELLIS with a resolution of 256. Thanks to geometry encoding, our approach achieves better geometry reconstruction with the same resolution as TRELLIS. Benefiting from efficient training, more details of complex structures are revealed, such as the tank track in the first row, as the resolution increases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.634, + 0.566 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.777 + ], + "angle": 0, + "content": "In this paper, we present SparseFlex, a new sparse-structured isosurface representation for differentiable mesh reconstruction with high resolution using rendering supervision, enabling the reconstruction of open surfaces. Based on SparseFlex, we propose a novel frustum-aware sectional voxel training strategy with adaptive frustum control to efficiently train SparseFlex VAE with high-resolution, dramatically reducing memory consumption. This strategy also allows our method to reconstruction the interiors only using rendering loss. Finally, we develop the image-to-3D generation pipeline following TELLIS [79]. Experiments demonstrate state-of-the-art reconstruction accuracy and high-quality generation with open surfaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitations: Despite the strong performance of SparseFlex VAE in both reconstruction and image-to-3D generation, some limitations remain. 1) Open surface boundaries, while handled effectively by voxel pruning, may exhibit minor artifacts at lower resolutions. 2) High-resolution generation remains computationally demanding. 3) Enhanced control over the generation of interior structures is an area for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14829" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.192 + ], + "angle": 0, + "content": "This work was done during Xianglong He's internship in VAST. This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012,KJZD2023092311491603 2,GJHZ20240218113604008)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.232, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[1] Tomas Akenine-Moller, Eric Haines, and Naty Hoffman. Real-time rendering. AK Peters/crc Press, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.261, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.376, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[4] Jen-Hao Rick Chang, Yuyang Wang, Miguel Angel Bautista Martin, Jiatao Gu, Josh Susskind, and Oncel Tuzel. 3d shape tokenization. arXiv preprint arXiv:2412.15618, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.418, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[5] Junyi Chen, Weicai Ye, Yifan Wang, Danpeng Chen, Di Huang, Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction. arXiv preprint arXiv:2409.06685, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.489, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[6] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. arXiv preprint arXiv:2412.17808, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.561, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[7] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In CVPR, pages 18522-18531, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.617, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.661, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[9] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[11] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. NeurIPS, 33: 21638-21652, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, pages 21126-21136, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.232, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. NeurIPS, 36:35799-35813, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.163, + 0.908, + 0.231 + ], + "angle": 0, + "content": "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.233, + 0.907, + 0.301 + ], + "angle": 0, + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.302, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[16] Yuanyuan Gao, Yalun Dai, Hao Li, Weicai Ye, Junyi Chen, Danpeng Chen, Dingwen Zhang, Tong He, Guofeng Zhang, and Junwei Han. Cosurfgs: Collaborative 3d surface gaussian splatting with distributed learning for large scene reconstruction. arXiv preprint arXiv:2412.17612, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.372, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[17] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.414, + 0.906, + 0.454 + ], + "angle": 0, + "content": "[18] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.456, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[19] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In ECCV, pages 463–479. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.511, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[20] Xianglong He, Junyi Chen, Di Huang, Zexiang Liu, Xiaoshui Huang, Wanli Ouyang, Chun Yuan, and Yangguang Li. Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits. arXiv preprint arXiv:2503.23022, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.581, + 0.907, + 0.648 + ], + "angle": 0, + "content": "[21] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.65, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.72, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.776, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[24] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In CVPR, pages 4369-4379, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[25] Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang," + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14830" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[27] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia Conference, pages 1–9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[28] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the Annual Conference on Computer Graphics and Interactive Techniques, pages 339-346, 2002. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[29] Michael Kazhdan and Hugues Hoppe. Screenedoisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.279, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.322, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[31] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.351, + 0.483, + 0.405 + ], + "angle": 0, + "content": "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM TOG, 39 (6), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.483, + 0.463 + ], + "angle": 0, + "content": "[33] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, pages 112-130, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[34] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[36] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024.2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[37] Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. arXiv preprint arXiv:2502.06608, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[38] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, et al. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. NeurIPS, 37:59314-59341, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[39] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. Neuf: Leaning neural unsigned" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "distance fields with volume rendering. In CVPR, pages 237-247, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.121, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.177, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[42] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Yuan Liu, Peng Wang, Christian Theobalt, Taku Komura, and Wenping Wang. Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topologies. In CVPR, pages 20834-20843, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.247, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[43] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169, 1987. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.289, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.317, + 0.907, + 0.355 + ], + "angle": 0, + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.358, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[46] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In CVPR, pages 12923-12932, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.414, + 0.792, + 0.427 + ], + "angle": 0, + "content": "[47] META. Digital twin catalog, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.429, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.497, + 0.907, + 0.551 + ], + "angle": 0, + "content": "[49] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[50] Tai-Jiang Mu, Hao-Xiang Chen, Jun-Xiong Cai, and Ning Guo. Neural 3d reconstruction from sparse views using geometric priors. Computational Visual Media, 9(4):687-697, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[51] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.707, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[53] Gregory M Nielson. Dual marching cubes. In IEEE visualization, pages 489-496, 2004. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.735, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[54] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.805, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[55] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.847, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. NeurIPS, 34:13032-13044, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14831" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[57] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.485, + 0.177 + ], + "angle": 0, + "content": "[58] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.177, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[59] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, pages 4209-4219, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[61] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM TOG, 42(4): 37:1-37:16, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[63] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, pages 19615-19625, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.525 + ], + "angle": 0, + "content": "[64] Sanghyun Son, Matheus Gadelha, Yang Zhou, Matthew Fisher, Zexiang Xu, Yi-Ling Qiao, Ming C Lin, and Yi Zhou. Dmesh++: An efficient differentiable mesh for complex shapes. arXiv preprint arXiv:2412.16776, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.567 + ], + "angle": 0, + "content": "[65] Sanghyun Son, Matheus Gadelha, Yang Zhou, Zexiang Xu, Ming C Lin, and Yi Zhou. Dmesh: A differentiable mesh representation. arXiv preprint arXiv:2404.13445, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[66] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In CVPR, pages 1798-1808, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In ECCV, pages 1-18. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[68] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[69] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. *NeurIPS*, 35:10021-10039, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.804 + ], + "angle": 0, + "content": "[70] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[71] Peng Wang and Yichun Shi. Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.092, + 0.908, + 0.163 + ], + "angle": 0, + "content": "[73] Yizhi Wang, Wallace Lira, Wenqi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Slice3d: Multi-slice occlusion-revealing single view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9881-9891, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[74] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In ECCV, pages 57-74, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.234, + 0.908, + 0.303 + ], + "angle": 0, + "content": "[75] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. arXiv preprint arXiv:2305.17134, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.305, + 0.908, + 0.374 + ], + "angle": 0, + "content": "[76] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.375, + 0.908, + 0.445 + ], + "angle": 0, + "content": "[77] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.446, + 0.908, + 0.503 + ], + "angle": 0, + "content": "[78] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.504, + 0.908, + 0.573 + ], + "angle": 0, + "content": "[79] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.574, + 0.908, + 0.615 + ], + "angle": 0, + "content": "[80] Hongyi Xu and Jernej Barbič. Signed distance fields for polygon soup meshes. In Graphics Interface, pages 35-41. 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.616, + 0.908, + 0.687 + ], + "angle": 0, + "content": "[81] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.688, + 0.908, + 0.757 + ], + "angle": 0, + "content": "[82] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.758, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[83] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 9(3):407-442, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.802, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[84] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In ECCV, pages 1-20. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.873, + 0.908, + 0.903 + ], + "angle": 0, + "content": "[85] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d:" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.092, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14832" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.485, + 0.12 + ], + "angle": 0, + "content": "A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.122, + 0.486, + 0.205 + ], + "angle": 0, + "content": "[86] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, Xin Li, and Wenping Wang. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[87] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.485, + 0.29 + ], + "angle": 0, + "content": "[88] Biao Zhang, Matthias Nießner, and Peter Wonka. 3dilig: Irregular latent grids for 3d generative modeling. NeurIPS, 35:21871-21885, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.292, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[89] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM TOG, 42 (4):92:1-92:16, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.484, + 0.417 + ], + "angle": 0, + "content": "[90] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: A structured and explicit radiance representation for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[91] Biao Zhang, Jing Ren, and Peter Wonka. Geometry distributions. arXiv preprint arXiv:2411.16076, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.45, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[92] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In ECCV, pages 1-19. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.484, + 0.574 + ], + "angle": 0, + "content": "[93] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM TOG, 43(4):1-20, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[94] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, and Ming-Ming Cheng. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[95] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 19531-19540, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[96] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[97] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In CVPR, pages 16259-16268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[98] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. NeurIPS, 36: 73969-73982, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[99] Xinyang Zheng, Yang Liu, Pengshuai Wang, and Xin Tong. Sdf-stylegan: implicit sdfs-based stylegan for 3d shape generation. In Computer Graphics Forum, pages 52-63, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.907, + 0.204 + ], + "angle": 0, + "content": "[100] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM TOG, 42(4):1-13, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[101] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, Hongbo Fu, and Song-Hai Zhang. Gp-recon: Online monocular neural 3d reconstruction with geometric prior. IEEE TVCG, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[102] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14833" + } + ] +] \ No newline at end of file diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_origin.pdf b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9971f215fe8eef2dcc6b8a08f357758314bdb986 --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/fcf007d6-72eb-465a-b05f-460f5649b8d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d1fb5b0fa39ed3d738210d40a3169477ffbdfac36cc0f0930b4b77134aec8b6 +size 20836246 diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/full.md b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..262a22f3010d7c2d8bd7315f43a333dfe85123a8 --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/full.md @@ -0,0 +1,338 @@ +# SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling + +Xianglong He $^{1,2\star}$ Zi-Xin Zou $^{2\star}$ Chia-Hao Chen $^{1,2}$ Yuan-Chen Guo $^{2}$ Ding Liang $^{2}$ Chun Yuan $^{1\dagger}$ Wanli Ouyang $^{3}$ Yan-Pei Cao $^{2}$ Yangguang Li $^{3,2\dagger}$ Tsinghua University ${}^{2}\mathrm{VAST}$ ${}^{3}\mathrm{The~Chinese~University~of~Hong~Kong}$ + +![](images/74ad2e45a4a9ab7a764ae131195c7a807ffba864067c7cda15dbfaf7d1934056.jpg) +Figure 1. SparseFlex VAE achieves high-fidelity reconstruction and generalization from point clouds. Benefiting from a sparse-structured differentiable isosurface surface representation and an efficient frustum-aware sectional voxel training strategy, our SparseFlex VAE demonstrates the state-of-the-art performance on complex geometries (left), open surfaces (top right), and even interior structures (bottom right), facilitating the high-quality image-to-3D generation with arbitrary topology. + +# Abstract + +Creating high-fidelity 3D meshes with arbitrary topology, including open surfaces and complex interiors, remains a significant challenge. Existing implicit field methods often require costly and detail-degrading watertight conversion, while other approaches struggle with high resolutions. This paper introduces SparseFlex, a novel sparse-structured isosurface representation that enables differentiable mesh reconstruction at resolutions up to $1024^{3}$ directly from rendering losses. SparseFlex combines the accuracy of Flexicubes with a sparse voxel structure, focusing computation on surface-adjacent regions and efficiently handling open surfaces. Crucially, we introduce a frustum-aware sectional voxel training strategy that activates only relevant voxels during rendering, dramatically reducing memory consumption and enabling high-resolution training. This also allows, for the first time, the reconstruction of mesh interiors using only rendering supervision. Building upon this, we demonstrate + +a complete shape modeling pipeline by training a variational autoencoder (VAE) and a rectified flow transformer for high-quality 3D shape generation. Our experiments show state-of-the-art reconstruction accuracy, with a $\sim 82\%$ reduction in Chamfer Distance and a $\sim 88\%$ increase in $F$ -score compared to previous methods, and demonstrate the generation of high-resolution, detailed 3D shapes with arbitrary topology. By enabling high-resolution, differentiable mesh reconstruction and generation with rendering losses, Sparse-Flex significantly advances the state-of-the-art in 3D shape representation and modeling. + +# 1. Introduction + +3D Generative AI is rapidly advancing, with applications spanning entertainment, design, and robotics. However, generating 3D content is fundamentally more challenging than 2D image or text generation due to the inherent complexity of representing and manipulating 3D geometry. Achieving high fidelity and supporting arbitrary topology—including open surfaces and complex interiors—presents particularly + +significant hurdles. + +Recent progress in 3D generative models has explored various representations, including point clouds [45, 46, 52, 69, 82], meshes [51, 63], 3DGS [5, 16, 19, 34, 79, 90], and implicit fields [9, 10, 36, 37, 59, 78, 79, 89, 93, 98, 100]. Implicit field representations, such as Signed Distance Functions (SDFs) and occupancy fields, have shown favorable results [33, 78, 79, 87, 89, 100]. However, creating training data for these methods typically involves a two-step process that limits their effectiveness. First, raw 3D mesh data must be converted into watertight representations [80, 93] to calculate SDF or occupancy values—a computationally expensive preprocessing step that often degrades fine details. Second, isosurfacing techniques (e.g., Marching Cubes [43], Dual Contouring [28]) are used to extract meshes from the learned continuous field, which can introduce further inaccuracies and artifacts. While Unsigned Distance Fields (UDFs) offer a potential way to model open surfaces [7, 11, 17, 40, 42, 86], they often suffer from instability and struggle to capture geometric fine details. + +Rendering-based supervision offers a powerful, differentiable alternative for training 3D representations and 3D generative models [60, 61, 74, 75, 79, 81]. By directly comparing rendered images of a generated mesh to ground-truth data, rendering losses avoid the need for the initial watertight preprocessing step and better preserve fine details. However, a critical bottleneck arises: when used with dense implicit fields, rendering supervision requires extremely high memory consumption at high resolutions, severely limiting the achievable fidelity. + +This paper introduces SparseFlex, a new sparse-structured isosurface representation that addresses these limitations and unlocks high-resolution, differentiable mesh reconstruction and generation using rendering supervisions. SparseFlex is built upon Flexicubes [61], providing accurate and differentiable isosurface extraction. The key design is the use of a sparse voxel structure instead of a conventional dense grid. This sparsity is crucial for two primary reasons: (1) it dramatically reduces memory consumption, enabling high-resolution modeling, and (2) it allows for the effective pruning of voxels near open boundaries, enabling the representation of open surfaces. + +To fully leverage the capabilities of SparseFlex, we propose frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering [1], this approach activates only the subset of SparseFlex voxels that reside within the camera's viewing frustum during each training iteration. We also introduce an adaptive strategy to control the frustum's parameters, further optimizing memory usage. This not only substantially reduces computational and memory overhead but also enables, for the first time, the reconstruction of mesh interiors using only rendering supervision by appropriately positioning the camera. Fig. 3 illustrates our + +SparseFlex representation and the efficient sectional voxel training strategy. + +Building on SparseFlex and our frustum-aware training, we present a complete 3D shape modeling pipeline. We employ a variational autoencoder (VAE) architecture, drawing inspiration from TRELLIS [79] but with key modifications. Firstly, because our focus is on high-fidelity geometry, we use point clouds as input to the VAE, providing a direct and detailed representation of the shape's surface. Furthermore, we introduce a self-pruning upsampling module within the decoder to further refine the sparse voxel structure, which is particularly beneficial for representing open surfaces. A rectified flow transformer is then trained on the learned latent space for high-quality, image-conditioned 3D shape generation. Through extensive experiments on Toy4k [66], ABO [12], GSO [15], Meta [47], Objaverse [14], and Deep fashion3D [21], our method demonstrates state-of-the-art shape reconstruction accuracy with minimal detail degradation, and high-quality single-image 3D shape generation. + +Our main contributions are: + +- We propose SparseFlex, a new sparse-structured isosurface representation enabling efficient, high-resolution, and differentiable 3D shape modeling, with natural handling of open surfaces. +- We introduce a novel sectional voxel training strategy with adaptive view frustum control, dramatically reducing memory consumption and enabling high-resolution mesh reconstruction and generation, including interiors, using rendering losses. +- We demonstrate state-of-the-art reconstruction accuracy and the generation of high-resolution, detailed 3D shapes with arbitrary topology, representing a significant advance in the field. + +# 2. Related Work + +# 2.1. 3D Shape Representations for Generation + +Point Cloud. Point clouds are a flexible 3D representation and can be easily acquired using depth sensors and LiDAR. Research has focused on point cloud processing [57, 58, 97] and generation [45, 46, 52, 69], with recent approaches treating point clouds as a distribution, directly sampling point clouds from noise via generative models [4, 91]. However, due to their limitation in representing solid surfaces, an additional surface reconstruction step [24, 29, 56] is required. + +Triangle Mesh. Triangle meshes are the primary representation for 3D assets in modern industrial pipelines. Recent works employ auto-regressive models to generate triangle faces sequentially [20, 51, 63, 68], improving resemblance to artist-created meshes. While effective for low-poly meshes, these methods struggle to produce high-quality meshes with a high face count. + +![](images/af5f77348e825581735aa3ecfbe62036e2e7048c1d469b53250f0ec3bdad936e.jpg) +Figure 2. Overview of the SparseFlex VAE pipeline. SparseFlex VAE takes point clouds sampled from a mesh as input, voxelizes them, and aggregates their features into each voxel. A sparse transformer encoder-decoder compresses the structured feature into a more compact latent space, followed by a self-pruning upsampling for higher resolution. Finally, the structured features are decoded to SparseFlex through a linear layer. Using the frustum-aware section voxel training strategy, we can train the entire pipeline more efficiently by rendering loss. + +Implicit Field. Implicit fields (SDF or occupancy) are widely used in geometry learning, particularly for 3D reconstruction [17, 24, 55, 61, 72, 101] and generation [9, 10, 36, 37, 50, 59, 78, 79, 83, 89, 93-96, 98, 100], producing high-quality meshes. Enhancements with triplane [78], vector set [87, 89, 98] sparse voxels [79] and sparse hierarchical voxels [59] allow neural networks to decode the field more effectively. These methods typically use the isosurface techniques [28, 43, 53] for surface extraction but struggle with open-surface shapes like clothing and flowers. Additionally, many require a time-consuming watertight conversion that degrades details, while rendering supervised approaches [79] suffer from high memory consumption when training at high resolutions. + +Open Surfaces. Meshes with open surfaces are common but remain challenging to process. The Unsigned Distance Field (UDF) is widely used for open-surface modeling from point clouds [11] or multiview images [40, 42]. Surf-D [86] introduces a UDF-based diffusion model for generating meshes with arbitrary topology. However, UDF-based shape modeling is more difficult than SDF, and its surface extraction [18] is prone to instability due to inaccuracy in neural network gradient estimations. As a result, achieving high-quality open-surface meshes remain challenging. 3PSDF [7] introduces a three-pole sign to distinguish surface-adjacent regions but relies on binary occupancy grids for surface extraction, leading to discontinuities and artifacts on the surface. Slice3D [73] introduces a slicing representation to reconstruct the interior structure with axis-aligned multislice images. DMesh series [64, 65] introduce differentiable mesh representations to represent general triangular meshes, while directly utilizing the representation for VAE remains a nontrivial challenge. In this paper, we incorporate sparse structure into Flexicubes [61] to efficient high-quality isosurface representation for open faces. + +# 2.2. 3D Generative Models and VAE + +Existing 3D generation studies are primarily classified into two categories: large 3D reconstruction model following 2D multiview diffusion models and native 3D generation models. The first category uses multiview diffusion models to generate multiview images from text or an image [25, 26, 39, 62, 71], followed by a large 3D reconstruction model [23, 35, 67, 81, 84, 92, 102] is employed to reconstruct the 3D representation from these images in seconds. However, inconsistencies between generated multiview images significantly often degrade the result quality. The second category focuses on native 3D generation models that directly generate 3D models through generative models, including GAN [3, 8, 77, 99], auto-regressive [49, 88], diffusion [27, 69, 89, 93] and rectified flow [37, 79]. + +Due to the diverse and non-compact nature of 3D representation, many approaches use Variational Auto-Encoder (VAE) [31] or Vector Quantized VAE (VQ-VAE) [70] to encode 3D shapes into latent spaces for the generative models. Geometry-focused methods often input point clouds uniformly sampled from mesh surfaces into the VAE [6, 89]. When aim at decoding both geometry and texture, some methods [33, 34, 79] encode multiview image features into latent spaces. This two-stage process makes the VAE's reconstruction quality crucial for subsequent generation performance. Some works [6, 87] have improved generation quality by enhancing the VAE's shape encoding-decoding capabilities. In this paper, we aim to develop a foundational VAE that encodes 3D shapes into latent spaces and reconstructs them with arbitrary topology while preserving the raw 3D shape's details. + +# 3. Method + +We present a method for high-resolution 3D shape modeling based on a novel sparse-structured isosurface representation, SparseFlex. Our approach leverages differentiable rendering + +![](images/54d6c885ed078716d2714c223431bdae9fb7e3b2e736671b67a34167fc775652.jpg) +Figure 3. Frustum-aware sectional voxel training. The previous mesh-based rendering training strategy (left) requires activating the entire dense grid to extract the mesh surface, even though only a few voxels are necessary during rendering. In contrast, our approach (right) adaptively activates the relevant voxels and enables the reconstruction of mesh interiors only using rendering supervision. + +for training, enabling accurate reconstruction of complex geometries, including open surfaces and interiors (see Fig. 1. Fig. 2 illustrates our variational autoencoder (VAE) based shape modeling pipeline, which utilizes the SparseFlex representation to learn a compact latent space of 3D shapes. The VAE decoder outputs the parameters of a SparseFlex instance, facilitating high-resolution mesh reconstruction. A key component of our approach is frustum-aware sectional voxel training, which significantly reduces memory consumption, allowing for training at resolutions up to $1024^3$ . Details of the SparseFlex representation are provided in Sec. 3.1, followed by a description of the VAE architecture in Sec. 3.2 and the training procedure in Sec. 3.3. Sec. 3.4 describes image-conditioned 3D shape generation using a rectified flow transformer on the learned latent space. + +# 3.1. SparseFlex Representation + +Preliminary. To achieve differentiable mesh extraction while preserving sharp features, we build upon Flexicubes [61], a method based on Dual Marching Cubes [53] (DMC). DMC places vertices at the center of voxels rather than edges/corners, leading to better feature preservation. Flexicubes constructs a dense voxel with a resolution of $N_r^3$ , where an SDF scalar grid $s \in \mathbb{R}^{N_g^3}$ is assigned to the voxel corner points, where $N_g = N_r + 1$ . For each voxel, it incorporates interpolation weights $\alpha \in \mathbb{R}_{>0}^{N_r^3 \times 8}, \beta \in \mathbb{R}_{>0}^{N_r^3 \times 12}$ to each voxel cell, and deformation vectors $\delta \in \mathbb{R}^{N_g^3}$ to each SDF grid. The underlying surface mesh can be effectively optimized via differentiable rendering [32]. + +Sparse Structured Flexicubes. The core design of SparseFlex is the introduction of a sparse voxel structure, enabling high-resolution shape representation while drastically reducing memory consumption. Instead of a dense grid, SparseFlex represents a shape using a significantly smaller set of voxels, $\mathcal{V}$ , concentrated near the surface. This sparsity is crucial for two reasons: (1) it allows us to achieve much higher resolutions than would be possible with a dense grid, and (2) it enables the natural representation of open surfaces by + +simply omitting voxels in empty regions. + +Specifically, the SparseFlex is defined by a set of $N_{v}$ voxels, $\mathcal{V} = \{v_{i} = (x_{i},y_{i},z_{i})\}$ , where $v_{i}$ represents the 3D coordinates of the center of the $i$ -th voxel. Let $N_{c}$ be the number of corner grids associated with these voxels, where $N_{v} = |\mathcal{V}|$ . Each voxel is associated with interpolation weights $\{\alpha_{i}\in \mathbb{R}_{>0}^{8},\beta_{i}\in \mathbb{R}_{>0}^{12}|0\leq i < N_{v}\}$ . Each corner grid is associated with an SDF value $\{s_j|0\leq j < N_c\}$ and deformation vectors $\{\delta_j|0\leq j < N_c\}$ . Due to the sparsity, $N_{v}\ll N_{r}^{3}$ and $N_{c}\ll N_{g}^{3}$ , representing a significant reduction in memory usage compared to the dense Flexicubes representation. We only apply Dual Marching Cubes on these sparse voxels to extract the underlying surface. Formally, the SparseFlex representation, $\mathcal{S}$ , is defined as: + +$$ +\mathcal {S} = \left(\mathcal {V}, \mathcal {F} _ {c}, \mathcal {F} _ {v}\right), \quad \mathcal {F} _ {c} = \left\{s _ {j}, \delta_ {j} \right\}, \quad \mathcal {F} _ {v} = \left\{\alpha_ {i}, \beta_ {i} \right\}, \tag {1} +$$ + +where $\mathcal{V}$ represents the voxel centers, $\mathcal{F}_c$ contains the SDF values and deformations at the corner grids, and $\mathcal{F}_v$ contains the interpolation weights for each voxel. + +SparseFlex inherits the differentiability of Flexicubes, allowing for end-to-end optimization using rendering losses. This eliminates the need for watertight mesh pre-processing, preserving fine details. Furthermore, the sparse structure, combined with the continuous and deformable nature of the SDF, allows for accurate and efficient representation of high-quality open-surface meshes. The sparsity also paves the way for our efficient frustum-aware training strategy, described in Sec. 3.3. + +# 3.2. SparseFlex VAE for Shape Modeling + +To learn a compact and disentangled latent space of 3D shapes, we employ a variational autoencoder (VAE) [31] that utilizes the SparseFlex representation. A VAE learns a probabilistic mapping between an input space (in our case, 3D shapes represented as point clouds) and a lower-dimensional latent space, enabling both reconstruction and generation of shapes. Fig. 2 also illustrates our VAE architecture. Our architecture draws inspiration from TRELLIS [79], but with key modifications to leverage the strengths of SparseFlex. + +Encoder. The input to our encoder is a point cloud $\mathcal{P} = \{p_i\in \mathbb{R}^3\}_{i = 1}^{N_p}$ , uniformly sampled from the surface of a 3D mesh, along with corresponding normals $\mathcal{N} = \{n_i\in \mathbb{R}^3\}_{i = 1}^{N_p}$ . We first voxelize the point cloud to obtain the sparse structure $\mathcal{V}$ of the SparseFlex representation $\mathcal{S}$ . We then employ a shallow PointNet [57] to aggregate local geometric features within each voxel. Specifically, for each voxel $v_{i}\in \mathcal{V}$ , we apply a local max-pooling operation [55] to the points contained within that voxel, producing a feature vector $f_{i}$ . These voxel features $\mathcal{F} = \{f_i\}$ , along with the sparse structure $\mathcal{V}$ , are then fed into a sparse transformer backbone. This backbone utilizes shifted window attention [41, 85], similar to TRELLIS [79], but is adapted to operate directly on the sparse voxel features $\mathcal{F}$ and structure $\mathcal{V}$ . The trans + +former outputs a latent code $z \in \mathbb{R}^{d_z}$ , which represents the encoded 3D shape. + +Decoder. The decoder takes the latent code $z$ as input and predicts the parameters of a SparseFlex instance, $S = (\mathcal{V},\mathcal{F}_c,\mathcal{F}_v)$ . We use a series of transformer layers, culminating in a final linear layer, to predict the SDF values $(s_j)$ and deformations $(\delta_{j})$ for each corner grid, as well as the interpolation weights $(\alpha_{i},\beta_{i})$ for each voxel. + +Upsampling Modules. To achieve high-resolution reconstructions, we incorporate two convolutional, self-pruning upsampling modules within the decoder, following the transformer. These modules progressively increase the resolution of the SparseFlex representation. Each upsampling module subdivides existing voxels into smaller voxels (increasing the resolution by a factor of 4). Crucially, each module also prunes redundant voxels based on a predicted occupancy value. A voxel is considered occupied if, after the subdivision, it contains any points from the input point cloud $\mathcal{P}$ . This pruning process, inspired by [59], is essential for maintaining the sparsity of the SparseFlex representation and is particularly beneficial for accurately representing open surfaces, as it removes unnecessary voxels in empty regions. + +# 3.3. Training SparseFlex VAE + +We train our SparseFlex VAE end-to-end using rendering losses, leveraging the differentiability of the SparseFlex representation and a new training strategy called frustum-aware sectional voxel training. This strategy dramatically reduces memory consumption during training, enabling us to achieve high resolutions (up to $1024^{3}$ ) that would be infeasible with traditional approaches. + +Frustum-aware Sectional Voxel Training. Even with the sparse structure of SparseFlex, directly rendering the entire representation at high resolutions can be computationally expensive. Furthermore, standard rendering supervision typically focuses only on the visible surface, neglecting the interior of the shape. Besides, recent methods [38, 74, 76, 79, 81] relying on rendering supervision from mesh typically require extracting the entire mesh because a dense representation doesn't trivially allow for partial extraction. In contrast, our sparse representation naturally enables partial extraction. To address these issues, we introduce frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering for efficient visibility culling [1], this approach activates only the voxels within the camera's viewing frustum during each training iteration. "Activating" a voxel means including it in the isosurface extraction and rendering process. This "sectional" approach means we only process a portion of the 3D space at a time, markedly reducing memory usage. + +As illustrated in Fig. 3, given the camera's extrinsic $\pi$ , intrinsics $K$ , and the near $(n)$ and far $(f)$ clipping planes of the viewing frustum, we compute the Model-View-Projection + +(MVP) matrix MVP. We then use a boolean operator to check whether the center of each voxel $v_{i}$ lies within the viewing frustum defined by the MVP matrix. We use $I(v_{i}\in \mathrm{Frustum}(\mathbf{MVP}))$ to represent this check, where $I(\cdot)$ is an indicator function. The set of active voxels, $\mathcal{V}_{\mathrm{active}}$ , is then defined as: + +$$ +\mathcal {V} _ {\text {a c t i v e}} = \left\{v _ {i} \mid I \left(v _ {i} \in \operatorname {F r u s t u m} (\mathbf {M V P})\right) = 1, v _ {i} \in \mathcal {V} \right\}. \tag {2} +$$ + +Adaptive Frustum and Interior Reconstruction. We introduce a visibility ratio $\alpha$ ( $0 < \alpha \leq 1$ ) controlling the proportion of active voxels in SparseFlex. We adaptively adjust the near and far clipping planes to ensure that approximately $\alpha N_v$ voxels are within the frustum. This is achieved through an iterative process: we initially set the near and far planes and iteratively adjust them based on the number of active voxels until the desired proportion is reached. + +This adaptive frustum also enables a novel capability: reconstructing mesh interiors using only rendering supervision. By positioning a virtual camera inside the object or adjusting the near clipping plane to intersect the mesh, we can render and supervise the internal structure (see Fig. 4). Moreover, the zoom-in camera viewpoint can render the mesh surface with greater details, making it better for higher-resolution training. This is a significant advantage over methods that rely on watertight representations, which cannot capture interior details. + +Loss Function. We train our VAE in an end-to-end manner, with an objective function comprising four components: + +$$ +\mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {r e n d e r}} + \lambda_ {2} \mathcal {L} _ {\text {p r u n e}} + \lambda_ {3} \mathcal {L} _ {\mathrm {K L}} + \lambda_ {4} \mathcal {L} _ {\text {f l e x}}, \tag {3} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ , and $\lambda_4$ are weighting coefficients that balance the different loss terms. + +$\mathcal{L}_{\mathrm{render}}$ is the rendering supervision loss. We employ a combination of losses commonly used in differentiable rendering [30, 48]: + +$$ +\mathcal {L} _ {\text {r e n d e r}} = \lambda_ {d} \mathcal {L} _ {d} + \lambda_ {n} \mathcal {L} _ {n} + \lambda_ {m} \mathcal {L} _ {m} + \lambda_ {s s} \mathcal {L} _ {s s} + \lambda_ {l p} \mathcal {L} _ {l p}, \tag {4} +$$ + +where $\mathcal{L}_d$ , $\mathcal{L}_n$ , and $\mathcal{L}_m$ denote the L1 loss for depth maps, normal maps, and mask maps, respectively. $\mathcal{L}_{ss}$ and $\mathcal{L}_{lp}$ denote SSIM loss and LPIPS loss, and are only applied on normal maps. Please refer to the supplementary material for a detailed definition of these losses and their weighting coefficients. + +$\mathcal{L}_{\mathrm{prune}}$ is the structure loss, formulated as a binary cross-entropy (BCE) loss, supervising the construction of sparse voxels: + +$$ +\mathcal {L} _ {\text {p r u n e}} = \operatorname {B C E} (V, \hat {V}), \tag {5} +$$ + +where $V$ represents the ground-truth occupancy of voxels derived from the input point cloud and $\hat{V}$ represents the predicted occupancy by the upsampling modules. + +$\mathcal{L}_{\mathrm{KL}}$ is the KL divergence between the learned latent distribution and a standard normal prior, regularizing the latent space. $\mathcal{L}_{\mathrm{flex}}$ is the regularization term from Flexicubes [61] to encourage smooth SDF values. + +
MethodToys4kDora Benchmark
CD ↓F1(0.001) ↑F1(0.01) ↑CD ↓F1(0.001) ↑F1(0.01) ↑
Craftsman [36]13.08/4.6310.13/15.1556.51/85.0213.54/2.066.30/11.1473.71/91.95
Dora [6]11.15/2.1317.29/26.5581.54/93.8416.61/1.0813.65/25.7878.73/96.40
Trellis [79]12.90/11.894.05/4.9359.65/64.0517.42/9.833.81/6.2062.70/71.95
XCube [59]4.35/3.141.61/13.4974.65/79.624.74/2.371.31/0.8475.64/86.50
3PSDF* [7]4.51/3.6911.33/14.1081.70/86.137.45/1.687.52/12.5079.43/91.17
Ours2562.56/1.2518.31/27.2385.35/92.011.93/0.5316.24/28.3788.76/97.31
Ours5121.67/0.8423.74/34.1090.39/95.601.36/0.2321.85/36.0391.55/98.51
Ours10241.33/0.6025.95/35.6992.30/96.220.86/0.1225.71/39.5094.71/99.14
+ +Table 1. Quantitative comparison for VAE reconstruction quality on the Toys4K dataset (left) and Dora benchmark (right). The '/' symbol separates the results computed over the entire dataset from those obtained exclusively on the watertight subset. + +
MethodCD↓F1(0.001) ↑F1(0.01) ↑
Surf-D [86]63.790.8023.17
3PSDF* [7]0.268.1499.35
Ours256†0.556.3594.88
Ours2560.0818.6099.99
Ours512†0.1811.3199.93
Ours5120.0531.60100.00
Ours1024†0.0524.80100.00
Ours10240.0437.22100.00
+ +Table 2. Reconstruction results on open-surface dataset Deepfashion3D. ${}^{ \dagger }$ indicates the absence of the self-pruning upsampling module. + +# 3.4. Image-to-3D Generation with Rectified Flow + +Building upon the trained SparseFlex VAE, we develop a pipeline for high-quality, image-conditioned 3D shape generation, following a similar approach to TRELLIS [61]. Our approach consists of two main components: Structure Flow Model and Structured Latent Flow Model. + +Structure Flow Model. First, a separate, simple fully 3D convolutional structure VAE is employed to compress dense voxels representing 3D shapes into a low-resolution (1/4 scale) space. Subsequently, the image condition features are extracted using DINOv2 [54] and injected into the transformer model via cross-attention, after which a rectified flow model is trained within this low-resolution space. During inference, given an input image, the trained structure flow model generates the corresponding low-resolution 3D space, which is then decoded by the structure VAE to produce the sparse structure of the generated 3D shape. + +Structured Latent Flow Model. Based on the proposed SparseFlex VAE, the point cloud and the corresponding voxelized sparse structure of a 3D shape are encoded into a structured latent space. Subsequently, the image condition feature obtained via DINOv2 is injected into the sparse transformer model through cross-attention, followed by training a rectified-flow model within this structured latent space. During inference, given an input image, the corresponding sparse structure is first generated using the structure flow model and its structure VAE. Then, both the sparse structure and the input image are provided to the structured latent flow model to generate the corresponding structured latent representation. Finally, the SparseFlex VAE decodes this latent representation to produce the final 3D shape. + +# 4. Experiments + +# 4.1. Implementation Details + +We develop the implementation of SparseFlex based on the official code provided by FlexiCubes [61]. We train SparseFlex VAE and structured latent flow model using approximately 400K high-quality 3D meshes filtered from largescale datasets, Objaverse (-XL) [13, 14]. Since incorrect normals in raw data can significantly degrade the performance of both VAE reconstruction and image-to-3D generation, we apply a mesh preprocessing step to correct these flipped normals, ensuring that all normals are consistently oriented outwards. + +Building on the success of progressive training in recent works [93], we train SparseFlex VAE progressively, increasing final resolution from low to high (256, 512, and 1024). For the structure VAE and structure flow model, we adopt the model from Trellis [79] and finetune their pre-trained weights to our task. We train SparseFlex VAE on 64 A100 GPUs with a batch size of 64 and train structured latent flow models with a batch size of 256. We use the AdamW [44] optimizer with an initial learning rate of $1e - 4$ and the weight decay as 0.01. At inference, we generate the results with 3.5 CFG and 50 sampling steps. + +# 4.2. Dataset, Baselines, and Metrics + +Dataset. We evaluate the reconstruction quality of VAE across different methods on a diverse set of datasets, including 1) universal datasets ABO [12], GSO [15], Meta [47], Objaverse [14], Toys4k [66] and 2) open-surface dataset Deepfashion3D [21]). The test list of ABO, GSO, Meta, and Objaverse is derived from Dora benchmark [6] after excluding our training data which includes about $2.7\mathrm{k}$ assets. For Toys4k, we use the full set following Trellis [79]. For the image-to-3D generation, we evaluate the methods on 200 random assets from Toys4k [66] and some images in the wild, showcasing the superior potential on generation tasks of SparseFlex VAE. + +Baselines. We compare our VAE with previous state-of-the-art methods, including Craftsman [36], Trellis [79], Dora [6], XCubes [59], Surf-D [86] and 3PSDF [7], with Surf-D and 3PSDF specially designed for open surfaces. We directly use the available pre-trained weights provided by these baselines, except for 3PSDF, which we re-implement and train on our dataset. We only compare Surf-D on the Deepfashion3D dataset due to the lack of available pretrained weights trained on large-scale datasets. For the generation results, we compare our method with InstantMesh [81], Direct3D [78], and TRELLIS [79]. + +Metrics. We evaluate the reconstruction performance of VAE by using the commonly used metrics, including Cham- + +![](images/3139940c63b50ebec68417037363d42068b0caf085fad63cb60353a272aa33bd.jpg) +Figure 4. Qualitative comparison of VAE reconstruction between ours and other state-of-the-art baselines. Our approach demonstrates superior performance in reconstructing complex shapes, open surfaces, and even interior structures. + +fer Distance (CD) and F-score with thresholds of 0.01 and 0.001. The metrics are multiplied by $10^{4}$ and $10^{2}$ , respectively. We evaluate the results by directly comparing the model's original output with ground-truth mesh, without applying any post-processing techniques (e.g., mesh simplification or removal of internal structures based on visibility). For generation results, four orthogonal views of normal maps for each shape are rendered for quantitative comparisons. We report the Fréchet Inception Distance (FID) [22] and Kernel Inception Distance (KID) [2]. + +# 4.3. VAE Reconstruction Evaluation + +We conduct extensive experiments to evaluate the quantitative results of VAE reconstruction from different methods in Table Tab. 1. Since Dora [6] and Craftsman [36] are trained on watertight data, we separate the results exclusively on + +the watertight subset from the results computed over the entire dataset for the clear demonstration. These methods often perform poorly on non-watertight meshes, especially open surfaces like flowers. Ours $_{256}$ already outperforms other baselines in terms of CD and F-score. As the resolution increases, our method achieves even better, ultimately achieving a $\sim 82\%$ reduction on CD and $\sim 88\%$ increase in F-score. Fig. 4 demonstrate the superiority of our method for complex shapes, open surfaces, and interior structures. Tab. 2 compares our method with those designed for open surfaces on the Deepfashion3D [21] dataset, and our method still achieves the best performance among them. + +# 4.4. Image to 3D Genration + +We also validate the effectiveness of our SparseFlex VAE as a foundation model for generation. Tab. 4 validates the ef + +![](images/8575be838c41fb14159ebf248c1a779cd23373d4535b25e2830a4463a0f3e86a.jpg) +Figure 5. Qualitative comparison of VAE reconstruction quality between our method with different resolution and TRELLIS. + +
Feed-Forward Time (ms)↓GPU Memory Cost (MB)↓
Resolution25635123102432563512310243
Ours (α = 0.1)3336201151355154018355441
Ours (α = 0.3)3576971475374034567569991
w/o FSV390958OOM4070362029OOM
w/o FSV & Sp.418OOMOOM45505OOMOOM
+ +Table 3. Feed-Forward time and GPU memory cost comparisons. $\alpha$ stands for the visibility ratio of voxels. 'OOM' means Out Of Memory and 'FSV' means frustum-aware sectional voxel training strategy. 'Sp' means SparseFlex. + +
MethodInstantMesh [81]Direct3D [78]TRELLIS [79]Ours
FID↓68.7450.8447.6644.95
KID (×103)↓9.682.041.281.05
+ +Table 4. Quantitative generation results on Toys4k. + +fectiveness of our generation. Visualizations are also demonstrated in Fig. 6, which includes image-to-3D results using wild images. The generated shapes, which preserves sharp edges and fine details, highly match the given images and showcase the generalization of our method. + +# 4.5. Ablation Studies + +Self-Pruning Upsampling. Tab. 2 demonstrates that the self-pruning upsampling module plays an important role in the reconstruction quality of open-surface shapes, as it allows for effective pruning of voxels near open boundaries, enabling the reconstruction boundary to better align with input. Supplementary shows the visual effects of this module in detail. + +SparseFlex and Frustum-aware Sectional Voxel Training. To demonstrate the effectiveness of SparseFlex and frustum-aware sectional voxel training strategy, we evaluate the runtime of the feed-forward and GPU memory consumption with different settings, as shown in Tab. 3. It demonstrates that SparseFlex effectively reduces the required GPU memory and runtime during network feed-forward. However, it is not efficient to scale to higher resolution. The frustum-aware sectional voxel training strategy eliminates the reliance on the entire surface extraction during rendering, significantly + +![](images/588db65a262c3c325f52d2c57ccbc32c864ea6b5f83d92ec72fa6a8c17a5aee9.jpg) +Figure 6. Single image-to-3D generations with in-the-wild images. The geometry of generated assets accurately preserves sharp edges and fine details. + +reducing the memory requirements during training. + +Sparse Voxel Resolutions. Higher resolution leads to better VAE reconstruction quality as shown in Tab. 1. Fig. 5 illustrates the qualitative comparison of SparseFlex VAE with different resolutions, along with TRELLIS with a resolution of 256. Thanks to geometry encoding, our approach achieves better geometry reconstruction with the same resolution as TRELLIS. Benefiting from efficient training, more details of complex structures are revealed, such as the tank track in the first row, as the resolution increases. + +# 5. Conclusion + +In this paper, we present SparseFlex, a new sparse-structured isosurface representation for differentiable mesh reconstruction with high resolution using rendering supervision, enabling the reconstruction of open surfaces. Based on SparseFlex, we propose a novel frustum-aware sectional voxel training strategy with adaptive frustum control to efficiently train SparseFlex VAE with high-resolution, dramatically reducing memory consumption. This strategy also allows our method to reconstruction the interiors only using rendering loss. Finally, we develop the image-to-3D generation pipeline following TELLIS [79]. Experiments demonstrate state-of-the-art reconstruction accuracy and high-quality generation with open surfaces. + +Limitations: Despite the strong performance of SparseFlex VAE in both reconstruction and image-to-3D generation, some limitations remain. 1) Open surface boundaries, while handled effectively by voxel pruning, may exhibit minor artifacts at lower resolutions. 2) High-resolution generation remains computationally demanding. 3) Enhanced control over the generation of interior structures is an area for future work. + +# 6. Acknowledgement + +This work was done during Xianglong He's internship in VAST. This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012,KJZD2023092311491603 2,GJHZ20240218113604008) + +# References + +[1] Tomas Akenine-Moller, Eric Haines, and Naty Hoffman. Real-time rendering. AK Peters/crc Press, 2019. 2, 5 +[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3 +[4] Jen-Hao Rick Chang, Yuyang Wang, Miguel Angel Bautista Martin, Jiatao Gu, Josh Susskind, and Oncel Tuzel. 3d shape tokenization. arXiv preprint arXiv:2412.15618, 2024. 2 +[5] Junyi Chen, Weicai Ye, Yifan Wang, Danpeng Chen, Di Huang, Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction. arXiv preprint arXiv:2409.06685, 2024. 2 +[6] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. arXiv preprint arXiv:2412.17808, 2024. 3, 6, 7 +[7] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In CVPR, pages 18522-18531, 2022. 2, 3, 6 +[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3 +[9] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 3 +[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2, 3 +[11] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. NeurIPS, 33: 21638-21652, 2020. 2, 3 +[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, pages 21126-21136, 2022. 2, 6 + +[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. NeurIPS, 36:35799-35813, 2023. 6 +[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 2, 6 +[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560, 2022. 2, 6 +[16] Yuanyuan Gao, Yalun Dai, Hao Li, Weicai Ye, Junyi Chen, Danpeng Chen, Dingwen Zhang, Tong He, Guofeng Zhang, and Junwei Han. Cosurfgs: Collaborative 3d surface gaussian splatting with distributed learning for large scene reconstruction. arXiv preprint arXiv:2412.17612, 2024. 2 +[17] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 2, 3 +[18] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 3 +[19] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In ECCV, pages 463–479. Springer, 2024. 2 +[20] Xianglong He, Junyi Chen, Di Huang, Zexiang Liu, Xiaoshui Huang, Wanli Ouyang, Chun Yuan, and Yangguang Li. Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits. arXiv preprint arXiv:2503.23022, 2025. 2 +[21] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 2, 6, 7 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In ICLR, 2024. 3 +[24] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In CVPR, pages 4369-4379, 2023. 2, 3 +[25] Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 3 +[26] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, + +Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 3 +[27] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia Conference, pages 1–9, 2022. 3 +[28] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the Annual Conference on Computer Graphics and Interactive Techniques, pages 339-346, 2002. 2, 3 +[29] Michael Kazhdan and Hugues Hoppe. Screenedoisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2 +[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 5 +[31] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3, 4 +[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM TOG, 39 (6), 2020. 4 +[33] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, pages 112-130, 2024. 2, 3 +[34] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 2, 3 +[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3 +[36] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024.2, 3, 6, 7 +[37] Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. arXiv preprint arXiv:2502.06608, 2025. 2, 3 +[38] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, et al. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. NeurIPS, 37:59314-59341, 2025. 5 +[39] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 3 +[40] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. Neuf: Leaning neural unsigned + +distance fields with volume rendering. In CVPR, pages 237-247, 2023. 2, 3 +[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, pages 10012-10022, 2021. 4 +[42] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Yuan Liu, Peng Wang, Christian Theobalt, Taku Komura, and Wenping Wang. Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topologies. In CVPR, pages 20834-20843, 2023. 2, 3 +[43] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169, 1987. 2, 3 +[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2 +[46] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In CVPR, pages 12923-12932, 2023. 2 +[47] META. Digital twin catalog, 2024. 2, 6 +[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5 +[49] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3 +[50] Tai-Jiang Mu, Hao-Xiang Chen, Jun-Xiong Cai, and Ning Guo. Neural 3d reconstruction from sparse views using geometric priors. Computational Visual Media, 9(4):687-697, 2023. 3 +[51] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229, 2020. 2 +[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[53] Gregory M Nielson. Dual marching cubes. In IEEE visualization, pages 489-496, 2004. 3, 4 +[54] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[55] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540, 2020. 3, 4 +[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. NeurIPS, 34:13032-13044, 2021. 2 + +[57] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 4 +[58] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2 +[59] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, pages 4209-4219, 2024. 2, 3, 5, 6 +[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2 +[61] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM TOG, 42(4): 37:1-37:16, 2023. 2, 3, 4, 5, 6 +[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 3 +[63] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, pages 19615-19625, 2024. 2 +[64] Sanghyun Son, Matheus Gadelha, Yang Zhou, Matthew Fisher, Zexiang Xu, Yi-Ling Qiao, Ming C Lin, and Yi Zhou. Dmesh++: An efficient differentiable mesh for complex shapes. arXiv preprint arXiv:2412.16776, 2024.3 +[65] Sanghyun Son, Matheus Gadelha, Yang Zhou, Zexiang Xu, Ming C Lin, and Yi Zhou. Dmesh: A differentiable mesh representation. arXiv preprint arXiv:2404.13445, 2024. 3 +[66] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In CVPR, pages 1798-1808, 2021. 2, 6 +[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In ECCV, pages 1-18. Springer, 2024. 3 +[68] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2 +[69] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. *NeurIPS*, 35:10021-10039, 2022. 2, 3 +[70] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3 +[71] Peng Wang and Yichun Shi. Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3 +[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 + +[73] Yizhi Wang, Wallace Lira, Wenqi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Slice3d: Multi-slice occlusion-revealing single view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9881-9891, 2024. 3 +[74] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In ECCV, pages 57-74, 2024. 2, 5 +[75] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. arXiv preprint arXiv:2305.17134, 2023. 2 +[76] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 5 +[77] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3 +[78] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3, 6, 8 +[79] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 3, 4, 5, 6, 8 +[80] Hongyi Xu and Jernej Barbič. Signed distance fields for polygon soup meshes. In Graphics Interface, pages 35-41. 2014. 2 +[81] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3, 5, 6, 8 +[82] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 2 +[83] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 9(3):407-442, 2023. 3 +[84] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In ECCV, pages 1-20. Springer, 2024. 3 +[85] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: + +A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 4 +[86] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, Xin Li, and Wenping Wang. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438, 2024. 2, 3, 6 +[87] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion, 2024. 2, 3 +[88] Biao Zhang, Matthias Nießner, and Peter Wonka. 3dilig: Irregular latent grids for 3d generative modeling. NeurIPS, 35:21871-21885, 2022. 3 +[89] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM TOG, 42 (4):92:1-92:16, 2023. 2, 3 +[90] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: A structured and explicit radiance representation for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2 +[91] Biao Zhang, Jing Ren, and Peter Wonka. Geometry distributions. arXiv preprint arXiv:2411.16076, 2024. 2 +[92] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In ECCV, pages 1-19. Springer, 2024. 3 +[93] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM TOG, 43(4):1-20, 2024. 2, 3, 6 +[94] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, and Ming-Ming Cheng. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. +[95] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 19531-19540, 2024. +[96] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. 3 +[97] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In CVPR, pages 16259-16268, 2021. 2 +[98] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. NeurIPS, 36: 73969-73982, 2023. 2, 3 + +[99] Xinyang Zheng, Yang Liu, Pengshuai Wang, and Xin Tong. Sdf-stylegan: implicit sdfs-based stylegan for 3d shape generation. In Computer Graphics Forum, pages 52-63, 2022. 3 +[100] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM TOG, 42(4):1-13, 2023. 2, 3 +[101] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, Hongbo Fu, and Song-Hai Zhang. Gp-recon: Online monocular neural 3d reconstruction with geometric prior. IEEE TVCG, 2024. 3 +[102] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 3 \ No newline at end of file diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/images.zip b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a9b11bd81f8e8947c16fb56212e7995fb63c0fc --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87a27799438887fa80c19f5941b8e6e9d004e8a88bf0a057b4729130ee3f792a +size 671662 diff --git a/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/layout.json b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d0a4547dadf1b416daf4f8331420adba3369acfc --- /dev/null +++ b/2025/SparseFlex_ High-Resolution and Arbitrary-Topology 3D Shape Modeling/layout.json @@ -0,0 +1,9855 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "text", + "content": "SparseFlex: High-Resolution and Arbitrary-Topology 3D Shape Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": "Xianglong He" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{1,2\\star}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Zi-Xin Zou" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Chia-Hao Chen" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Yuan-Chen Guo" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Ding Liang" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Chun Yuan" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Wanli Ouyang" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Yan-Pei Cao" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Yangguang Li" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "^{3,2\\dagger}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "{}^{2}\\mathrm{VAST}" + }, + { + "bbox": [ + 61, + 141, + 548, + 187 + ], + "type": "inline_equation", + "content": "{}^{3}\\mathrm{The~Chinese~University~of~Hong~Kong}" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 198, + 555, + 392 + ], + "blocks": [ + { + "bbox": [ + 60, + 198, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 60, + 198, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 60, + 198, + 555, + 392 + ], + "type": "image", + "image_path": "74ad2e45a4a9ab7a764ae131195c7a807ffba864067c7cda15dbfaf7d1934056.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 397, + 555, + 443 + ], + "lines": [ + { + "bbox": [ + 54, + 397, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 397, + 555, + 443 + ], + "type": "text", + "content": "Figure 1. SparseFlex VAE achieves high-fidelity reconstruction and generalization from point clouds. Benefiting from a sparse-structured differentiable isosurface surface representation and an efficient frustum-aware sectional voxel training strategy, our SparseFlex VAE demonstrates the state-of-the-art performance on complex geometries (left), open surfaces (top right), and even interior structures (bottom right), facilitating the high-quality image-to-3D generation with arbitrary topology." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "spans": [ + { + "bbox": [ + 151, + 457, + 200, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "text", + "content": "Creating high-fidelity 3D meshes with arbitrary topology, including open surfaces and complex interiors, remains a significant challenge. Existing implicit field methods often require costly and detail-degrading watertight conversion, while other approaches struggle with high resolutions. This paper introduces SparseFlex, a novel sparse-structured isosurface representation that enables differentiable mesh reconstruction at resolutions up to " + }, + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "inline_equation", + "content": "1024^{3}" + }, + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "text", + "content": " directly from rendering losses. SparseFlex combines the accuracy of Flexicubes with a sparse voxel structure, focusing computation on surface-adjacent regions and efficiently handling open surfaces. Crucially, we introduce a frustum-aware sectional voxel training strategy that activates only relevant voxels during rendering, dramatically reducing memory consumption and enabling high-resolution training. This also allows, for the first time, the reconstruction of mesh interiors using only rendering supervision. Building upon this, we demonstrate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "text", + "content": "a complete shape modeling pipeline by training a variational autoencoder (VAE) and a rectified flow transformer for high-quality 3D shape generation. Our experiments show state-of-the-art reconstruction accuracy, with a " + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\sim 82\\%" + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "text", + "content": " reduction in Chamfer Distance and a " + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\sim 88\\%" + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "text", + "content": " increase in " + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 459, + 556, + 592 + ], + "type": "text", + "content": "-score compared to previous methods, and demonstrate the generation of high-resolution, detailed 3D shapes with arbitrary topology. By enabling high-resolution, differentiable mesh reconstruction and generation with rendering losses, Sparse-Flex significantly advances the state-of-the-art in 3D shape representation and modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 608, + 395, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 717 + ], + "type": "text", + "content": "3D Generative AI is rapidly advancing, with applications spanning entertainment, design, and robotics. However, generating 3D content is fundamentally more challenging than 2D image or text generation due to the inherent complexity of representing and manipulating 3D geometry. Achieving high fidelity and supporting arbitrary topology—including open surfaces and complex interiors—presents particularly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 694, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 127, + 703 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 56, + 694, + 127, + 703 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 704, + 139, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 139, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 139, + 714 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14822" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 135, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 135, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 135, + 83 + ], + "type": "text", + "content": "significant hurdles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 312 + ], + "type": "text", + "content": "Recent progress in 3D generative models has explored various representations, including point clouds [45, 46, 52, 69, 82], meshes [51, 63], 3DGS [5, 16, 19, 34, 79, 90], and implicit fields [9, 10, 36, 37, 59, 78, 79, 89, 93, 98, 100]. Implicit field representations, such as Signed Distance Functions (SDFs) and occupancy fields, have shown favorable results [33, 78, 79, 87, 89, 100]. However, creating training data for these methods typically involves a two-step process that limits their effectiveness. First, raw 3D mesh data must be converted into watertight representations [80, 93] to calculate SDF or occupancy values—a computationally expensive preprocessing step that often degrades fine details. Second, isosurfacing techniques (e.g., Marching Cubes [43], Dual Contouring [28]) are used to extract meshes from the learned continuous field, which can introduce further inaccuracies and artifacts. While Unsigned Distance Fields (UDFs) offer a potential way to model open surfaces [7, 11, 17, 40, 42, 86], they often suffer from instability and struggle to capture geometric fine details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 434 + ], + "type": "text", + "content": "Rendering-based supervision offers a powerful, differentiable alternative for training 3D representations and 3D generative models [60, 61, 74, 75, 79, 81]. By directly comparing rendered images of a generated mesh to ground-truth data, rendering losses avoid the need for the initial watertight preprocessing step and better preserve fine details. However, a critical bottleneck arises: when used with dense implicit fields, rendering supervision requires extremely high memory consumption at high resolutions, severely limiting the achievable fidelity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 297, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 297, + 581 + ], + "type": "text", + "content": "This paper introduces SparseFlex, a new sparse-structured isosurface representation that addresses these limitations and unlocks high-resolution, differentiable mesh reconstruction and generation using rendering supervisions. SparseFlex is built upon Flexicubes [61], providing accurate and differentiable isosurface extraction. The key design is the use of a sparse voxel structure instead of a conventional dense grid. This sparsity is crucial for two primary reasons: (1) it dramatically reduces memory consumption, enabling high-resolution modeling, and (2) it allows for the effective pruning of voxels near open boundaries, enabling the representation of open surfaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "To fully leverage the capabilities of SparseFlex, we propose frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering [1], this approach activates only the subset of SparseFlex voxels that reside within the camera's viewing frustum during each training iteration. We also introduce an adaptive strategy to control the frustum's parameters, further optimizing memory usage. This not only substantially reduces computational and memory overhead but also enables, for the first time, the reconstruction of mesh interiors using only rendering supervision by appropriately positioning the camera. Fig. 3 illustrates our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "SparseFlex representation and the efficient sectional voxel training strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 98, + 556, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 556, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 556, + 301 + ], + "type": "text", + "content": "Building on SparseFlex and our frustum-aware training, we present a complete 3D shape modeling pipeline. We employ a variational autoencoder (VAE) architecture, drawing inspiration from TRELLIS [79] but with key modifications. Firstly, because our focus is on high-fidelity geometry, we use point clouds as input to the VAE, providing a direct and detailed representation of the shape's surface. Furthermore, we introduce a self-pruning upsampling module within the decoder to further refine the sparse voxel structure, which is particularly beneficial for representing open surfaces. A rectified flow transformer is then trained on the learned latent space for high-quality, image-conditioned 3D shape generation. Through extensive experiments on Toy4k [66], ABO [12], GSO [15], Meta [47], Objaverse [14], and Deep fashion3D [21], our method demonstrates state-of-the-art shape reconstruction accuracy with minimal detail degradation, and high-quality single-image 3D shape generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 303, + 440, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 303, + 440, + 314 + ], + "spans": [ + { + "bbox": [ + 326, + 303, + 440, + 314 + ], + "type": "text", + "content": "Our main contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 316, + 555, + 471 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 316, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 555, + 363 + ], + "type": "text", + "content": "- We propose SparseFlex, a new sparse-structured isosurface representation enabling efficient, high-resolution, and differentiable 3D shape modeling, with natural handling of open surfaces." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 364, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 364, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 364, + 554, + 423 + ], + "type": "text", + "content": "- We introduce a novel sectional voxel training strategy with adaptive view frustum control, dramatically reducing memory consumption and enabling high-resolution mesh reconstruction and generation, including interiors, using rendering losses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 424, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 554, + 471 + ], + "type": "text", + "content": "- We demonstrate state-of-the-art reconstruction accuracy and the generation of high-resolution, detailed 3D shapes with arbitrary topology, representing a significant advance in the field." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 481, + 400, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 400, + 494 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 400, + 494 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 533, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 533, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 533, + 515 + ], + "type": "text", + "content": "2.1. 3D Shape Representations for Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 618 + ], + "type": "text", + "content": "Point Cloud. Point clouds are a flexible 3D representation and can be easily acquired using depth sensors and LiDAR. Research has focused on point cloud processing [57, 58, 97] and generation [45, 46, 52, 69], with recent approaches treating point clouds as a distribution, directly sampling point clouds from noise via generative models [4, 91]. However, due to their limitation in representing solid surfaces, an additional surface reconstruction step [24, 29, 56] is required." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 557, + 715 + ], + "type": "text", + "content": "Triangle Mesh. Triangle meshes are the primary representation for 3D assets in modern industrial pipelines. Recent works employ auto-regressive models to generate triangle faces sequentially [20, 51, 63, 68], improving resemblance to artist-created meshes. While effective for low-poly meshes, these methods struggle to produce high-quality meshes with a high face count." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14823" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 52, + 539, + 218 + ], + "blocks": [ + { + "bbox": [ + 69, + 52, + 539, + 218 + ], + "lines": [ + { + "bbox": [ + 69, + 52, + 539, + 218 + ], + "spans": [ + { + "bbox": [ + 69, + 52, + 539, + 218 + ], + "type": "image", + "image_path": "af5f77348e825581735aa3ecfbe62036e2e7048c1d469b53250f0ec3bdad936e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 268 + ], + "type": "text", + "content": "Figure 2. Overview of the SparseFlex VAE pipeline. SparseFlex VAE takes point clouds sampled from a mesh as input, voxelizes them, and aggregates their features into each voxel. A sparse transformer encoder-decoder compresses the structured feature into a more compact latent space, followed by a self-pruning upsampling for higher resolution. Finally, the structured features are decoded to SparseFlex through a linear layer. Using the frustum-aware section voxel training strategy, we can train the entire pipeline more efficiently by rendering loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 297, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 297, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 297, + 443 + ], + "type": "text", + "content": "Implicit Field. Implicit fields (SDF or occupancy) are widely used in geometry learning, particularly for 3D reconstruction [17, 24, 55, 61, 72, 101] and generation [9, 10, 36, 37, 50, 59, 78, 79, 83, 89, 93-96, 98, 100], producing high-quality meshes. Enhancements with triplane [78], vector set [87, 89, 98] sparse voxels [79] and sparse hierarchical voxels [59] allow neural networks to decode the field more effectively. These methods typically use the isosurface techniques [28, 43, 53] for surface extraction but struggle with open-surface shapes like clothing and flowers. Additionally, many require a time-consuming watertight conversion that degrades details, while rendering supervised approaches [79] suffer from high memory consumption when training at high resolutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 462, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 296, + 714 + ], + "type": "text", + "content": "Open Surfaces. Meshes with open surfaces are common but remain challenging to process. The Unsigned Distance Field (UDF) is widely used for open-surface modeling from point clouds [11] or multiview images [40, 42]. Surf-D [86] introduces a UDF-based diffusion model for generating meshes with arbitrary topology. However, UDF-based shape modeling is more difficult than SDF, and its surface extraction [18] is prone to instability due to inaccuracy in neural network gradient estimations. As a result, achieving high-quality open-surface meshes remain challenging. 3PSDF [7] introduces a three-pole sign to distinguish surface-adjacent regions but relies on binary occupancy grids for surface extraction, leading to discontinuities and artifacts on the surface. Slice3D [73] introduces a slicing representation to reconstruct the interior structure with axis-aligned multislice images. DMesh series [64, 65] introduce differentiable mesh representations to represent general triangular meshes, while directly utilizing the representation for VAE remains a nontrivial challenge. In this paper, we incorporate sparse structure into Flexicubes [61] to efficient high-quality isosurface representation for open faces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 275, + 487, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 487, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 487, + 288 + ], + "type": "text", + "content": "2.2. 3D Generative Models and VAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 293, + 556, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 461 + ], + "type": "text", + "content": "Existing 3D generation studies are primarily classified into two categories: large 3D reconstruction model following 2D multiview diffusion models and native 3D generation models. The first category uses multiview diffusion models to generate multiview images from text or an image [25, 26, 39, 62, 71], followed by a large 3D reconstruction model [23, 35, 67, 81, 84, 92, 102] is employed to reconstruct the 3D representation from these images in seconds. However, inconsistencies between generated multiview images significantly often degrade the result quality. The second category focuses on native 3D generation models that directly generate 3D models through generative models, including GAN [3, 8, 77, 99], auto-regressive [49, 88], diffusion [27, 69, 89, 93] and rectified flow [37, 79]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 461, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 653 + ], + "type": "text", + "content": "Due to the diverse and non-compact nature of 3D representation, many approaches use Variational Auto-Encoder (VAE) [31] or Vector Quantized VAE (VQ-VAE) [70] to encode 3D shapes into latent spaces for the generative models. Geometry-focused methods often input point clouds uniformly sampled from mesh surfaces into the VAE [6, 89]. When aim at decoding both geometry and texture, some methods [33, 34, 79] encode multiview image features into latent spaces. This two-stage process makes the VAE's reconstruction quality crucial for subsequent generation performance. Some works [6, 87] have improved generation quality by enhancing the VAE's shape encoding-decoding capabilities. In this paper, we aim to develop a foundational VAE that encodes 3D shapes into latent spaces and reconstructs them with arbitrary topology while preserving the raw 3D shape's details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 370, + 669 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We present a method for high-resolution 3D shape modeling based on a novel sparse-structured isosurface representation, SparseFlex. Our approach leverages differentiable rendering" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14824" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 53, + 292, + 171 + ], + "blocks": [ + { + "bbox": [ + 55, + 53, + 292, + 171 + ], + "lines": [ + { + "bbox": [ + 55, + 53, + 292, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 53, + 292, + 171 + ], + "type": "image", + "image_path": "54d6c885ed078716d2714c223431bdae9fb7e3b2e736671b67a34167fc775652.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 242 + ], + "type": "text", + "content": "Figure 3. Frustum-aware sectional voxel training. The previous mesh-based rendering training strategy (left) requires activating the entire dense grid to extract the mesh surface, even though only a few voxels are necessary during rendering. In contrast, our approach (right) adaptively activates the relevant voxels and enables the reconstruction of mesh interiors only using rendering supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 249, + 298, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 298, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 298, + 429 + ], + "type": "text", + "content": "for training, enabling accurate reconstruction of complex geometries, including open surfaces and interiors (see Fig. 1. Fig. 2 illustrates our variational autoencoder (VAE) based shape modeling pipeline, which utilizes the SparseFlex representation to learn a compact latent space of 3D shapes. The VAE decoder outputs the parameters of a SparseFlex instance, facilitating high-resolution mesh reconstruction. A key component of our approach is frustum-aware sectional voxel training, which significantly reduces memory consumption, allowing for training at resolutions up to " + }, + { + "bbox": [ + 56, + 249, + 298, + 429 + ], + "type": "inline_equation", + "content": "1024^3" + }, + { + "bbox": [ + 56, + 249, + 298, + 429 + ], + "type": "text", + "content": ". Details of the SparseFlex representation are provided in Sec. 3.1, followed by a description of the VAE architecture in Sec. 3.2 and the training procedure in Sec. 3.3. Sec. 3.4 describes image-conditioned 3D shape generation using a rectified flow transformer on the learned latent space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 204, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 204, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 204, + 449 + ], + "type": "text", + "content": "3.1. SparseFlex Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": "Preliminary. To achieve differentiable mesh extraction while preserving sharp features, we build upon Flexicubes [61], a method based on Dual Marching Cubes [53] (DMC). DMC places vertices at the center of voxels rather than edges/corners, leading to better feature preservation. Flexicubes constructs a dense voxel with a resolution of " + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "inline_equation", + "content": "N_r^3" + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": ", where an SDF scalar grid " + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^{N_g^3}" + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": " is assigned to the voxel corner points, where " + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "inline_equation", + "content": "N_g = N_r + 1" + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": ". For each voxel, it incorporates interpolation weights " + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}_{>0}^{N_r^3 \\times 8}, \\beta \\in \\mathbb{R}_{>0}^{N_r^3 \\times 12}" + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": " to each voxel cell, and deformation vectors " + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "inline_equation", + "content": "\\delta \\in \\mathbb{R}^{N_g^3}" + }, + { + "bbox": [ + 55, + 450, + 296, + 602 + ], + "type": "text", + "content": " to each SDF grid. The underlying surface mesh can be effectively optimized via differentiable rendering [32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "content": "Sparse Structured Flexicubes. The core design of SparseFlex is the introduction of a sparse voxel structure, enabling high-resolution shape representation while drastically reducing memory consumption. Instead of a dense grid, SparseFlex represents a shape using a significantly smaller set of voxels, " + }, + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "content": ", concentrated near the surface. This sparsity is crucial for two reasons: (1) it allows us to achieve much higher resolutions than would be possible with a dense grid, and (2) it enables the natural representation of open surfaces by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 73, + 481, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 481, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 481, + 84 + ], + "type": "text", + "content": "simply omitting voxels in empty regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": "Specifically, the SparseFlex is defined by a set of " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": " voxels, " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{i} = (x_{i},y_{i},z_{i})\\}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": " represents the 3D coordinates of the center of the " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": "-th voxel. Let " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": " be the number of corner grids associated with these voxels, where " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "N_{v} = |\\mathcal{V}|" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ". Each voxel is associated with interpolation weights " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{i}\\in \\mathbb{R}_{>0}^{8},\\beta_{i}\\in \\mathbb{R}_{>0}^{12}|0\\leq i < N_{v}\\}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ". Each corner grid is associated with an SDF value " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\{s_j|0\\leq j < N_c\\}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": " and deformation vectors " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\{\\delta_j|0\\leq j < N_c\\}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ". Due to the sparsity, " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "N_{v}\\ll N_{r}^{3}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "N_{c}\\ll N_{g}^{3}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ", representing a significant reduction in memory usage compared to the dense Flexicubes representation. We only apply Dual Marching Cubes on these sparse voxels to extract the underlying surface. Formally, the SparseFlex representation, " + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 84, + 555, + 228 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 333, + 237, + 555, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 237, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 333, + 237, + 555, + 250 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left(\\mathcal {V}, \\mathcal {F} _ {c}, \\mathcal {F} _ {v}\\right), \\quad \\mathcal {F} _ {c} = \\left\\{s _ {j}, \\delta_ {j} \\right\\}, \\quad \\mathcal {F} _ {v} = \\left\\{\\alpha_ {i}, \\beta_ {i} \\right\\}, \\tag {1}", + "image_path": "8eae0b20d7e92dbfee86ef359b908b252a5850183896c20cb8cf766b8a8fe024.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "content": " represents the voxel centers, " + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c" + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "content": " contains the SDF values and deformations at the corner grids, and " + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v" + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "content": " contains the interpolation weights for each voxel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 556, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 402 + ], + "type": "text", + "content": "SparseFlex inherits the differentiability of Flexicubes, allowing for end-to-end optimization using rendering losses. This eliminates the need for watertight mesh pre-processing, preserving fine details. Furthermore, the sparse structure, combined with the continuous and deformable nature of the SDF, allows for accurate and efficient representation of high-quality open-surface meshes. The sparsity also paves the way for our efficient frustum-aware training strategy, described in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 509, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 509, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 509, + 425 + ], + "type": "text", + "content": "3.2. SparseFlex VAE for Shape Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "content": "To learn a compact and disentangled latent space of 3D shapes, we employ a variational autoencoder (VAE) [31] that utilizes the SparseFlex representation. A VAE learns a probabilistic mapping between an input space (in our case, 3D shapes represented as point clouds) and a lower-dimensional latent space, enabling both reconstruction and generation of shapes. Fig. 2 also illustrates our VAE architecture. Our architecture draws inspiration from TRELLIS [79], but with key modifications to leverage the strengths of SparseFlex." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": "Encoder. The input to our encoder is a point cloud " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ", uniformly sampled from the surface of a 3D mesh, along with corresponding normals " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N} = \\{n_i\\in \\mathbb{R}^3\\}_{i = 1}^{N_p}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ". We first voxelize the point cloud to obtain the sparse structure " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": " of the SparseFlex representation " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ". We then employ a shallow PointNet [57] to aggregate local geometric features within each voxel. Specifically, for each voxel " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "v_{i}\\in \\mathcal{V}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ", we apply a local max-pooling operation [55] to the points contained within that voxel, producing a feature vector " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ". These voxel features " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{f_i\\}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ", along with the sparse structure " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ", are then fed into a sparse transformer backbone. This backbone utilizes shifted window attention [41, 85], similar to TRELLIS [79], but is adapted to operate directly on the sparse voxel features " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": " and structure " + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 543, + 557, + 713 + ], + "type": "text", + "content": ". The trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14825" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "former outputs a latent code " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", which represents the encoded 3D shape." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": "Decoder. The decoder takes the latent code " + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": " as input and predicts the parameters of a SparseFlex instance, " + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "inline_equation", + "content": "S = (\\mathcal{V},\\mathcal{F}_c,\\mathcal{F}_v)" + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": ". We use a series of transformer layers, culminating in a final linear layer, to predict the SDF values " + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "inline_equation", + "content": "(s_j)" + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": " and deformations " + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "inline_equation", + "content": "(\\delta_{j})" + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": " for each corner grid, as well as the interpolation weights " + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "inline_equation", + "content": "(\\alpha_{i},\\beta_{i})" + }, + { + "bbox": [ + 55, + 102, + 296, + 174 + ], + "type": "text", + "content": " for each voxel." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 298, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 298, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 298, + 347 + ], + "type": "text", + "content": "Upsampling Modules. To achieve high-resolution reconstructions, we incorporate two convolutional, self-pruning upsampling modules within the decoder, following the transformer. These modules progressively increase the resolution of the SparseFlex representation. Each upsampling module subdivides existing voxels into smaller voxels (increasing the resolution by a factor of 4). Crucially, each module also prunes redundant voxels based on a predicted occupancy value. A voxel is considered occupied if, after the subdivision, it contains any points from the input point cloud " + }, + { + "bbox": [ + 55, + 178, + 298, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 178, + 298, + 347 + ], + "type": "text", + "content": ". This pruning process, inspired by [59], is essential for maintaining the sparsity of the SparseFlex representation and is particularly beneficial for accurately representing open surfaces, as it removes unnecessary voxels in empty regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 198, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 198, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 198, + 368 + ], + "type": "text", + "content": "3.3. Training SparseFlex VAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 456 + ], + "type": "text", + "content": "We train our SparseFlex VAE end-to-end using rendering losses, leveraging the differentiability of the SparseFlex representation and a new training strategy called frustum-aware sectional voxel training. This strategy dramatically reduces memory consumption during training, enabling us to achieve high resolutions (up to " + }, + { + "bbox": [ + 55, + 373, + 296, + 456 + ], + "type": "inline_equation", + "content": "1024^{3}" + }, + { + "bbox": [ + 55, + 373, + 296, + 456 + ], + "type": "text", + "content": ") that would be infeasible with traditional approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 297, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 297, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 297, + 677 + ], + "type": "text", + "content": "Frustum-aware Sectional Voxel Training. Even with the sparse structure of SparseFlex, directly rendering the entire representation at high resolutions can be computationally expensive. Furthermore, standard rendering supervision typically focuses only on the visible surface, neglecting the interior of the shape. Besides, recent methods [38, 74, 76, 79, 81] relying on rendering supervision from mesh typically require extracting the entire mesh because a dense representation doesn't trivially allow for partial extraction. In contrast, our sparse representation naturally enables partial extraction. To address these issues, we introduce frustum-aware sectional voxel training. Inspired by techniques used in real-time rendering for efficient visibility culling [1], this approach activates only the voxels within the camera's viewing frustum during each training iteration. \"Activating\" a voxel means including it in the isosurface extraction and rendering process. This \"sectional\" approach means we only process a portion of the 3D space at a time, markedly reducing memory usage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 3, given the camera's extrinsic " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", intrinsics " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", and the near " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "(n)" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and far " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " clipping planes of the viewing frustum, we compute the Model-View-Projection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "(MVP) matrix MVP. We then use a boolean operator to check whether the center of each voxel " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " lies within the viewing frustum defined by the MVP matrix. We use " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "I(v_{i}\\in \\mathrm{Frustum}(\\mathbf{MVP}))" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " to represent this check, where " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "I(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is an indicator function. The set of active voxels, " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{active}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", is then defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 334, + 148, + 555, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 148, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 334, + 148, + 555, + 160 + ], + "type": "interline_equation", + "content": "\\mathcal {V} _ {\\text {a c t i v e}} = \\left\\{v _ {i} \\mid I \\left(v _ {i} \\in \\operatorname {F r u s t u m} (\\mathbf {M V P})\\right) = 1, v _ {i} \\in \\mathcal {V} \\right\\}. \\tag {2}", + "image_path": "39c6a5e8a68efdb1adfb5f382d766a0ce382bbb894de2b06a5b76c0a230a4193.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "text", + "content": "Adaptive Frustum and Interior Reconstruction. We introduce a visibility ratio " + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "inline_equation", + "content": "0 < \\alpha \\leq 1" + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "text", + "content": ") controlling the proportion of active voxels in SparseFlex. We adaptively adjust the near and far clipping planes to ensure that approximately " + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\alpha N_v" + }, + { + "bbox": [ + 313, + 162, + 555, + 257 + ], + "type": "text", + "content": " voxels are within the frustum. This is achieved through an iterative process: we initially set the near and far planes and iteratively adjust them based on the number of active voxels until the desired proportion is reached." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 556, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 556, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 556, + 376 + ], + "type": "text", + "content": "This adaptive frustum also enables a novel capability: reconstructing mesh interiors using only rendering supervision. By positioning a virtual camera inside the object or adjusting the near clipping plane to intersect the mesh, we can render and supervise the internal structure (see Fig. 4). Moreover, the zoom-in camera viewpoint can render the mesh surface with greater details, making it better for higher-resolution training. This is a significant advantage over methods that rely on watertight representations, which cannot capture interior details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 382, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 407 + ], + "type": "text", + "content": "Loss Function. We train our VAE in an end-to-end manner, with an objective function comprising four components:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 411, + 555, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 411, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 345, + 411, + 555, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e n d e r}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r u n e}} + \\lambda_ {3} \\mathcal {L} _ {\\mathrm {K L}} + \\lambda_ {4} \\mathcal {L} _ {\\text {f l e x}}, \\tag {3}", + "image_path": "7ae59b74392443edfa97321b21a4687fab0198c5a277c9277d901cdc92df8c87.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "content": " are weighting coefficients that balance the different loss terms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{render}}" + }, + { + "bbox": [ + 313, + 451, + 555, + 487 + ], + "type": "text", + "content": " is the rendering supervision loss. We employ a combination of losses commonly used in differentiable rendering [30, 48]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 491, + 555, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 491, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 325, + 491, + 555, + 504 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e n d e r}} = \\lambda_ {d} \\mathcal {L} _ {d} + \\lambda_ {n} \\mathcal {L} _ {n} + \\lambda_ {m} \\mathcal {L} _ {m} + \\lambda_ {s s} \\mathcal {L} _ {s s} + \\lambda_ {l p} \\mathcal {L} _ {l p}, \\tag {4}", + "image_path": "5f93aab4d4004a455c919e90b8d1d77a709a10902bfeb83b437414eb452c2e82.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": " denote the L1 loss for depth maps, normal maps, and mask maps, respectively. " + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ss}" + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lp}" + }, + { + "bbox": [ + 313, + 507, + 554, + 578 + ], + "type": "text", + "content": " denote SSIM loss and LPIPS loss, and are only applied on normal maps. Please refer to the supplementary material for a detailed definition of these losses and their weighting coefficients." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 580, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prune}}" + }, + { + "bbox": [ + 313, + 580, + 555, + 613 + ], + "type": "text", + "content": " is the structure loss, formulated as a binary cross-entropy (BCE) loss, supervising the construction of sparse voxels:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 389, + 613, + 555, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 613, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 389, + 613, + 555, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r u n e}} = \\operatorname {B C E} (V, \\hat {V}), \\tag {5}", + "image_path": "6af6de4f0039464092b55eae4af0fe8b8826add21caa7cb1b3e9e83eeb18d3db.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "text", + "content": " represents the ground-truth occupancy of voxels derived from the input point cloud and " + }, + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 313, + 631, + 554, + 666 + ], + "type": "text", + "content": " represents the predicted occupancy by the upsampling modules." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " is the KL divergence between the learned latent distribution and a standard normal prior, regularizing the latent space. " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{flex}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " is the regularization term from Flexicubes [61] to encourage smooth SDF values." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14826" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 57, + 300, + 145 + ], + "blocks": [ + { + "bbox": [ + 51, + 57, + 300, + 145 + ], + "lines": [ + { + "bbox": [ + 51, + 57, + 300, + 145 + ], + "spans": [ + { + "bbox": [ + 51, + 57, + 300, + 145 + ], + "type": "table", + "html": "
MethodToys4kDora Benchmark
CD ↓F1(0.001) ↑F1(0.01) ↑CD ↓F1(0.001) ↑F1(0.01) ↑
Craftsman [36]13.08/4.6310.13/15.1556.51/85.0213.54/2.066.30/11.1473.71/91.95
Dora [6]11.15/2.1317.29/26.5581.54/93.8416.61/1.0813.65/25.7878.73/96.40
Trellis [79]12.90/11.894.05/4.9359.65/64.0517.42/9.833.81/6.2062.70/71.95
XCube [59]4.35/3.141.61/13.4974.65/79.624.74/2.371.31/0.8475.64/86.50
3PSDF* [7]4.51/3.6911.33/14.1081.70/86.137.45/1.687.52/12.5079.43/91.17
Ours2562.56/1.2518.31/27.2385.35/92.011.93/0.5316.24/28.3788.76/97.31
Ours5121.67/0.8423.74/34.1090.39/95.601.36/0.2321.85/36.0391.55/98.51
Ours10241.33/0.6025.95/35.6992.30/96.220.86/0.1225.71/39.5094.71/99.14
", + "image_path": "59c12d763c317b42b1e4b4c85d3743eedc8ece405745da80bb5acb60f5e20870.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 198, + 251, + 280 + ], + "blocks": [ + { + "bbox": [ + 48, + 149, + 303, + 194 + ], + "lines": [ + { + "bbox": [ + 48, + 149, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 303, + 194 + ], + "type": "text", + "content": "Table 1. Quantitative comparison for VAE reconstruction quality on the Toys4K dataset (left) and Dora benchmark (right). The '/' symbol separates the results computed over the entire dataset from those obtained exclusively on the watertight subset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 198, + 251, + 280 + ], + "lines": [ + { + "bbox": [ + 99, + 198, + 251, + 280 + ], + "spans": [ + { + "bbox": [ + 99, + 198, + 251, + 280 + ], + "type": "table", + "html": "
MethodCD↓F1(0.001) ↑F1(0.01) ↑
Surf-D [86]63.790.8023.17
3PSDF* [7]0.268.1499.35
Ours256†0.556.3594.88
Ours2560.0818.6099.99
Ours512†0.1811.3199.93
Ours5120.0531.60100.00
Ours1024†0.0524.80100.00
Ours10240.0437.22100.00
", + "image_path": "e55632c0be08eaa2d3e683789f96b49fb90a525ae8968a4b32d7ab8a404e7924.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 297, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 316 + ], + "type": "text", + "content": "Table 2. Reconstruction results on open-surface dataset Deepfashion3D. " + }, + { + "bbox": [ + 55, + 284, + 297, + 316 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 284, + 297, + 316 + ], + "type": "text", + "content": " indicates the absence of the self-pruning upsampling module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 321, + 286, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 286, + 334 + ], + "type": "text", + "content": "3.4. Image-to-3D Generation with Rectified Flow" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": "Building upon the trained SparseFlex VAE, we develop a pipeline for high-quality, image-conditioned 3D shape generation, following a similar approach to TRELLIS [61]. Our approach consists of two main components: Structure Flow Model and Structured Latent Flow Model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 539 + ], + "type": "text", + "content": "Structure Flow Model. First, a separate, simple fully 3D convolutional structure VAE is employed to compress dense voxels representing 3D shapes into a low-resolution (1/4 scale) space. Subsequently, the image condition features are extracted using DINOv2 [54] and injected into the transformer model via cross-attention, after which a rectified flow model is trained within this low-resolution space. During inference, given an input image, the trained structure flow model generates the corresponding low-resolution 3D space, which is then decoded by the structure VAE to produce the sparse structure of the generated 3D shape." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Structured Latent Flow Model. Based on the proposed SparseFlex VAE, the point cloud and the corresponding voxelized sparse structure of a 3D shape are encoded into a structured latent space. Subsequently, the image condition feature obtained via DINOv2 is injected into the sparse transformer model through cross-attention, followed by training a rectified-flow model within this structured latent space. During inference, given an input image, the corresponding sparse structure is first generated using the structure flow model and its structure VAE. Then, both the sparse structure and the input image are provided to the structured latent flow model to generate the corresponding structured latent representation. Finally, the SparseFlex VAE decodes this latent representation to produce the final 3D shape." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 447, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 447, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 447, + 103 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 108, + 555, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 555, + 227 + ], + "type": "text", + "content": "We develop the implementation of SparseFlex based on the official code provided by FlexiCubes [61]. We train SparseFlex VAE and structured latent flow model using approximately 400K high-quality 3D meshes filtered from largescale datasets, Objaverse (-XL) [13, 14]. Since incorrect normals in raw data can significantly degrade the performance of both VAE reconstruction and image-to-3D generation, we apply a mesh preprocessing step to correct these flipped normals, ensuring that all normals are consistently oriented outwards." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 556, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 361 + ], + "type": "text", + "content": "Building on the success of progressive training in recent works [93], we train SparseFlex VAE progressively, increasing final resolution from low to high (256, 512, and 1024). For the structure VAE and structure flow model, we adopt the model from Trellis [79] and finetune their pre-trained weights to our task. We train SparseFlex VAE on 64 A100 GPUs with a batch size of 64 and train structured latent flow models with a batch size of 256. We use the AdamW [44] optimizer with an initial learning rate of " + }, + { + "bbox": [ + 313, + 228, + 556, + 361 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 228, + 556, + 361 + ], + "type": "text", + "content": " and the weight decay as 0.01. At inference, we generate the results with 3.5 CFG and 50 sampling steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 368, + 484, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 484, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 484, + 380 + ], + "type": "text", + "content": "4.2. Dataset, Baselines, and Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 556, + 529 + ], + "type": "text", + "content": "Dataset. We evaluate the reconstruction quality of VAE across different methods on a diverse set of datasets, including 1) universal datasets ABO [12], GSO [15], Meta [47], Objaverse [14], Toys4k [66] and 2) open-surface dataset Deepfashion3D [21]). The test list of ABO, GSO, Meta, and Objaverse is derived from Dora benchmark [6] after excluding our training data which includes about " + }, + { + "bbox": [ + 313, + 385, + 556, + 529 + ], + "type": "inline_equation", + "content": "2.7\\mathrm{k}" + }, + { + "bbox": [ + 313, + 385, + 556, + 529 + ], + "type": "text", + "content": " assets. For Toys4k, we use the full set following Trellis [79]. For the image-to-3D generation, we evaluate the methods on 200 random assets from Toys4k [66] and some images in the wild, showcasing the superior potential on generation tasks of SparseFlex VAE." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": "Baselines. We compare our VAE with previous state-of-the-art methods, including Craftsman [36], Trellis [79], Dora [6], XCubes [59], Surf-D [86] and 3PSDF [7], with Surf-D and 3PSDF specially designed for open surfaces. We directly use the available pre-trained weights provided by these baselines, except for 3PSDF, which we re-implement and train on our dataset. We only compare Surf-D on the Deepfashion3D dataset due to the lack of available pretrained weights trained on large-scale datasets. For the generation results, we compare our method with InstantMesh [81], Direct3D [78], and TRELLIS [79]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 555, + 696 + ], + "type": "text", + "content": "Metrics. We evaluate the reconstruction performance of VAE by using the commonly used metrics, including Cham-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 703, + 443, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 703, + 443, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 703, + 443, + 713 + ], + "type": "text", + "content": "https://github.com/vn-tlabs/FlexiCubes" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14827" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 54, + 547, + 485 + ], + "blocks": [ + { + "bbox": [ + 62, + 54, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 62, + 54, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 54, + 547, + 485 + ], + "type": "image", + "image_path": "3139940c63b50ebec68417037363d42068b0caf085fad63cb60353a272aa33bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 487, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 487, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 555, + 510 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of VAE reconstruction between ours and other state-of-the-art baselines. Our approach demonstrates superior performance in reconstructing complex shapes, open surfaces, and even interior structures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "text", + "content": "fer Distance (CD) and F-score with thresholds of 0.01 and 0.001. The metrics are multiplied by " + }, + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "inline_equation", + "content": "10^{4}" + }, + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "inline_equation", + "content": "10^{2}" + }, + { + "bbox": [ + 55, + 518, + 298, + 639 + ], + "type": "text", + "content": ", respectively. We evaluate the results by directly comparing the model's original output with ground-truth mesh, without applying any post-processing techniques (e.g., mesh simplification or removal of internal structures based on visibility). For generation results, four orthogonal views of normal maps for each shape are rendered for quantitative comparisons. We report the Fréchet Inception Distance (FID) [22] and Kernel Inception Distance (KID) [2]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 647, + 228, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 228, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 228, + 659 + ], + "type": "text", + "content": "4.3. VAE Reconstruction Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "content": "We conduct extensive experiments to evaluate the quantitative results of VAE reconstruction from different methods in Table Tab. 1. Since Dora [6] and Craftsman [36] are trained on watertight data, we separate the results exclusively on" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "content": "the watertight subset from the results computed over the entire dataset for the clear demonstration. These methods often perform poorly on non-watertight meshes, especially open surfaces like flowers. Ours" + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "inline_equation", + "content": "_{256}" + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "content": " already outperforms other baselines in terms of CD and F-score. As the resolution increases, our method achieves even better, ultimately achieving a " + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\sim 82\\%" + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "content": " reduction on CD and " + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\sim 88\\%" + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "content": " increase in F-score. Fig. 4 demonstrate the superiority of our method for complex shapes, open surfaces, and interior structures. Tab. 2 compares our method with those designed for open surfaces on the Deepfashion3D [21] dataset, and our method still achieves the best performance among them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 671, + 446, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 446, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 446, + 684 + ], + "type": "text", + "content": "4.4. Image to 3D Genration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "We also validate the effectiveness of our SparseFlex VAE as a foundation model for generation. Tab. 4 validates the ef" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14828" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 56, + 295, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 56, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 56, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 56, + 295, + 205 + ], + "type": "image", + "image_path": "8575be838c41fb14159ebf248c1a779cd23373d4535b25e2830a4463a0f3e86a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 235 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of VAE reconstruction quality between our method with different resolution and TRELLIS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 241, + 290, + 306 + ], + "blocks": [ + { + "bbox": [ + 58, + 241, + 290, + 306 + ], + "lines": [ + { + "bbox": [ + 58, + 241, + 290, + 306 + ], + "spans": [ + { + "bbox": [ + 58, + 241, + 290, + 306 + ], + "type": "table", + "html": "
Feed-Forward Time (ms)↓GPU Memory Cost (MB)↓
Resolution25635123102432563512310243
Ours (α = 0.1)3336201151355154018355441
Ours (α = 0.3)3576971475374034567569991
w/o FSV390958OOM4070362029OOM
w/o FSV & Sp.418OOMOOM45505OOMOOM
", + "image_path": "46508c03dc3dcbe863266d5dc483a2eb84f5eff3e45dcd4038ed203999e923a3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 357, + 295, + 393 + ], + "blocks": [ + { + "bbox": [ + 48, + 309, + 299, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 309, + 299, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 299, + 353 + ], + "type": "text", + "content": "Table 3. Feed-Forward time and GPU memory cost comparisons. " + }, + { + "bbox": [ + 48, + 309, + 299, + 353 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 48, + 309, + 299, + 353 + ], + "type": "text", + "content": " stands for the visibility ratio of voxels. 'OOM' means Out Of Memory and 'FSV' means frustum-aware sectional voxel training strategy. 'Sp' means SparseFlex." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 357, + 295, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 357, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 357, + 295, + 393 + ], + "type": "table", + "html": "
MethodInstantMesh [81]Direct3D [78]TRELLIS [79]Ours
FID↓68.7450.8447.6644.95
KID (×103)↓9.682.041.281.05
", + "image_path": "5e034b646c6c6aa7eeef5136ab8d317c762bbcc2c7c4fc141927e0ac07247d23.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 397, + 272, + 408 + ], + "lines": [ + { + "bbox": [ + 77, + 397, + 272, + 408 + ], + "spans": [ + { + "bbox": [ + 77, + 397, + 272, + 408 + ], + "type": "text", + "content": "Table 4. Quantitative generation results on Toys4k." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 418, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 477 + ], + "type": "text", + "content": "fectiveness of our generation. Visualizations are also demonstrated in Fig. 6, which includes image-to-3D results using wild images. The generated shapes, which preserves sharp edges and fine details, highly match the given images and showcase the generalization of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 156, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 156, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 156, + 498 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 587 + ], + "type": "text", + "content": "Self-Pruning Upsampling. Tab. 2 demonstrates that the self-pruning upsampling module plays an important role in the reconstruction quality of open-surface shapes, as it allows for effective pruning of voxels near open boundaries, enabling the reconstruction boundary to better align with input. Supplementary shows the visual effects of this module in detail." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "SparseFlex and Frustum-aware Sectional Voxel Training. To demonstrate the effectiveness of SparseFlex and frustum-aware sectional voxel training strategy, we evaluate the runtime of the feed-forward and GPU memory consumption with different settings, as shown in Tab. 3. It demonstrates that SparseFlex effectively reduces the required GPU memory and runtime during network feed-forward. However, it is not efficient to scale to higher resolution. The frustum-aware sectional voxel training strategy eliminates the reliance on the entire surface extraction during rendering, significantly" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 57, + 569, + 223 + ], + "blocks": [ + { + "bbox": [ + 318, + 57, + 569, + 223 + ], + "lines": [ + { + "bbox": [ + 318, + 57, + 569, + 223 + ], + "spans": [ + { + "bbox": [ + 318, + 57, + 569, + 223 + ], + "type": "image", + "image_path": "588db65a262c3c325f52d2c57ccbc32c864ea6b5f83d92ec72fa6a8c17a5aee9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 260 + ], + "type": "text", + "content": "Figure 6. Single image-to-3D generations with in-the-wild images. The geometry of generated assets accurately preserves sharp edges and fine details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 271, + 522, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 522, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 522, + 283 + ], + "type": "text", + "content": "reducing the memory requirements during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 308, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 415 + ], + "type": "text", + "content": "Sparse Voxel Resolutions. Higher resolution leads to better VAE reconstruction quality as shown in Tab. 1. Fig. 5 illustrates the qualitative comparison of SparseFlex VAE with different resolutions, along with TRELLIS with a resolution of 256. Thanks to geometry encoding, our approach achieves better geometry reconstruction with the same resolution as TRELLIS. Benefiting from efficient training, more details of complex structures are revealed, such as the tank track in the first row, as the resolution increases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 436, + 388, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 388, + 448 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 388, + 448 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 615 + ], + "type": "text", + "content": "In this paper, we present SparseFlex, a new sparse-structured isosurface representation for differentiable mesh reconstruction with high resolution using rendering supervision, enabling the reconstruction of open surfaces. Based on SparseFlex, we propose a novel frustum-aware sectional voxel training strategy with adaptive frustum control to efficiently train SparseFlex VAE with high-resolution, dramatically reducing memory consumption. This strategy also allows our method to reconstruction the interiors only using rendering loss. Finally, we develop the image-to-3D generation pipeline following TELLIS [79]. Experiments demonstrate state-of-the-art reconstruction accuracy and high-quality generation with open surfaces." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 712 + ], + "type": "text", + "content": "Limitations: Despite the strong performance of SparseFlex VAE in both reconstruction and image-to-3D generation, some limitations remain. 1) Open surface boundaries, while handled effectively by voxel pruning, may exhibit minor artifacts at lower resolutions. 2) High-resolution generation remains computationally demanding. 3) Enhanced control over the generation of interior structures is an area for future work." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14829" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "content": "This work was done during Xianglong He's internship in VAST. This work was supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012,KJZD2023092311491603 2,GJHZ20240218113604008)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 183, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 183, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 183, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 66, + 183, + 296, + 205 + ], + "type": "text", + "content": "[1] Tomas Akenine-Moller, Eric Haines, and Naty Hoffman. Real-time rendering. AK Peters/crc Press, 2019. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 66, + 206, + 296, + 239 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 66, + 240, + 296, + 295 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, pages 16123-16133, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 297, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 297, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 66, + 297, + 294, + 329 + ], + "type": "text", + "content": "[4] Jen-Hao Rick Chang, Yuyang Wang, Miguel Angel Bautista Martin, Jiatao Gu, Josh Susskind, and Oncel Tuzel. 3d shape tokenization. arXiv preprint arXiv:2412.15618, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 331, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 331, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 66, + 331, + 296, + 385 + ], + "type": "text", + "content": "[5] Junyi Chen, Weicai Ye, Yifan Wang, Danpeng Chen, Di Huang, Wanli Ouyang, Guofeng Zhang, Yu Qiao, and Tong He. Gigags: Scaling up planar-based 3d gaussians for large scene surface reconstruction. arXiv preprint arXiv:2409.06685, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 387, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 387, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 387, + 296, + 441 + ], + "type": "text", + "content": "[6] Rui Chen, Jianfeng Zhang, Yixun Liang, Guan Luo, Weiyu Li, Jiarui Liu, Xiu Li, Xiaoxiao Long, Jiashi Feng, and Ping Tan. Dora: Sampling and benchmarking for 3d shape variational auto-encoders. arXiv preprint arXiv:2412.17808, 2024. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 444, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 444, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 66, + 444, + 296, + 486 + ], + "type": "text", + "content": "[7] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In CVPR, pages 18522-18531, 2022. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 488, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 488, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 66, + 488, + 296, + 520 + ], + "type": "text", + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 523, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 523, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 66, + 523, + 296, + 577 + ], + "type": "text", + "content": "[9] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 579, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 623 + ], + "type": "text", + "content": "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 624, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 656 + ], + "type": "text", + "content": "[11] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. NeurIPS, 33: 21638-21652, 2020. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 713 + ], + "type": "text", + "content": "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, pages 21126-21136, 2022. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "text", + "content": "[13] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 320, + 73, + 555, + 126 + ], + "type": "text", + "content": " objects. NeurIPS, 36:35799-35813, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 129, + 555, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 129, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 129, + 555, + 182 + ], + "type": "text", + "content": "[14] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, pages 13142-13153, 2023. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 184, + 555, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 555, + 238 + ], + "type": "text", + "content": "[15] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 555, + 293 + ], + "type": "text", + "content": "[16] Yuanyuan Gao, Yalun Dai, Hao Li, Weicai Ye, Junyi Chen, Danpeng Chen, Dingwen Zhang, Tong He, Guofeng Zhang, and Junwei Han. Cosurfgs: Collaborative 3d surface gaussian splatting with distributed learning for large scene reconstruction. arXiv preprint arXiv:2412.17612, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 294, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 294, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 294, + 554, + 326 + ], + "type": "text", + "content": "[17] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 327, + 554, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 554, + 359 + ], + "type": "text", + "content": "[18] Benoit Guillard, Federico Stella, and Pascal Fua. Meshudf: Fast and differentiable meshing of unsigned distance field networks. In ECCV, pages 576-592, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 361, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 361, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 321, + 361, + 555, + 404 + ], + "type": "text", + "content": "[19] Xianglong He, Junyi Chen, Sida Peng, Di Huang, Yangguang Li, Xiaoshui Huang, Chun Yuan, Wanli Ouyang, and Tong He. Gvgen: Text-to-3d generation with volumetric representation. In ECCV, pages 463–479. Springer, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 404, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 404, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 321, + 404, + 555, + 458 + ], + "type": "text", + "content": "[20] Xianglong He, Junyi Chen, Di Huang, Zexiang Liu, Xiaoshui Huang, Wanli Ouyang, Chun Yuan, and Yangguang Li. Meshcraft: Exploring efficient and controllable mesh generation with flow-based dits. arXiv preprint arXiv:2503.23022, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 460, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 460, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 321, + 460, + 555, + 513 + ], + "type": "text", + "content": "[21] Zhu Heming, Cao Yu, Jin Hang, Chen Weikai, Du Dong, Wang Zhangye, Cui Shuguang, and Han Xiaoguang. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In ECCV, pages 512-530, 2020. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 514, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 514, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 321, + 514, + 555, + 568 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 570, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 570, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 570, + 554, + 613 + ], + "type": "text", + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. LRM: large reconstruction model for single image to 3d. In ICLR, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 614, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 555, + 646 + ], + "type": "text", + "content": "[24] Jiahui Huang, Zan Gojcic, Matan Atzmon, Or Litany, Sanja Fidler, and Francis Williams. Neural kernel surface reconstruction. In CVPR, pages 4369-4379, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 647, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 555, + 690 + ], + "type": "text", + "content": "[25] Zehuan Huang, Yuan-Chen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 555, + 713 + ], + "type": "text", + "content": "[26] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14830" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 296, + 106 + ], + "type": "text", + "content": "Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 296, + 140 + ], + "type": "text", + "content": "[27] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia Conference, pages 1–9, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 296, + 185 + ], + "type": "text", + "content": "[28] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the Annual Conference on Computer Graphics and Interactive Techniques, pages 339-346, 2002. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 219 + ], + "type": "text", + "content": "[29] Michael Kazhdan and Hugues Hoppe. Screenedoisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 220, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 295, + 253 + ], + "type": "text", + "content": "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 42(4):139-1, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "type": "text", + "content": "[31] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 295, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 320 + ], + "type": "text", + "content": "[32] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM TOG, 39 (6), 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 322, + 295, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 295, + 366 + ], + "type": "text", + "content": "[33] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In ECCV, pages 112-130, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 295, + 421 + ], + "type": "text", + "content": "[34] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 423, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 295, + 477 + ], + "type": "text", + "content": "[35] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 533 + ], + "type": "text", + "content": "[36] Weiyu Li, Jiarui Liu, Rui Chen, Yixun Liang, Xuelin Chen, Ping Tan, and Xiaoxiao Long. Craftsman: High-fidelity mesh generation with 3d native generation and interactive geometry refiner. arXiv preprint arXiv:2405.14979, 2024.2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 535, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 590 + ], + "type": "text", + "content": "[37] Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, et al. Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. arXiv preprint arXiv:2502.06608, 2025. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 591, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 644 + ], + "type": "text", + "content": "[38] Minghua Liu, Chong Zeng, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, et al. Meshformer: High-quality mesh generation with 3d-guided reconstruction model. NeurIPS, 37:59314-59341, 2025. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 689 + ], + "type": "text", + "content": "[39] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "text", + "content": "[40] Yu-Tao Liu, Li Wang, Jie Yang, Weikai Chen, Xiaoxu Meng, Bo Yang, and Lin Gao. Neuf: Leaning neural unsigned" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "distance fields with volume rendering. In CVPR, pages 237-247, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 95, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 95, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 95, + 555, + 139 + ], + "type": "text", + "content": "[41] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In CVPR, pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 140, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 555, + 194 + ], + "type": "text", + "content": "[42] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Yuan Liu, Peng Wang, Christian Theobalt, Taku Komura, and Wenping Wang. Neuraludf: Learning unsigned distance fields for multi-view reconstruction of surfaces with arbitrary topologies. In CVPR, pages 20834-20843, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "type": "text", + "content": "[43] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In SIGGRAPH, pages 163-169, 1987. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "type": "text", + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 251, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 555, + 281 + ], + "type": "text", + "content": "[45] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "type": "text", + "content": "[46] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In CVPR, pages 12923-12932, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 327, + 484, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 484, + 338 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 484, + 338 + ], + "type": "text", + "content": "[47] META. Digital twin catalog, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 339, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 339, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 339, + 555, + 392 + ], + "type": "text", + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 393, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 393, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 393, + 555, + 436 + ], + "type": "text", + "content": "[49] Paritosh Mittal, Yen-Chi Cheng, Maneesh Singh, and Shubham Tulsiani. Autosdf: Shape priors for 3d completion, reconstruction and generation. In CVPR, pages 306-315, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 437, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 555, + 480 + ], + "type": "text", + "content": "[50] Tai-Jiang Mu, Hao-Xiang Chen, Jun-Xiong Cai, and Ning Guo. Neural 3d reconstruction from sparse views using geometric priors. Computational Visual Media, 9(4):687-697, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 482, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 554, + 514 + ], + "type": "text", + "content": "[51] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In ICML, pages 7220-7229, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 515, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 554, + 557 + ], + "type": "text", + "content": "[52] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 559, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 554, + 581 + ], + "type": "text", + "content": "[53] Gregory M Nielson. Dual marching cubes. In IEEE visualization, pages 489-496, 2004. 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 582, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 582, + 555, + 635 + ], + "type": "text", + "content": "[54] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 637, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 637, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 637, + 554, + 669 + ], + "type": "text", + "content": "[55] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In ECCV, pages 523-540, 2020. 3, 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 554, + 712 + ], + "type": "text", + "content": "[56] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape as points: A differentiable poisson solver. NeurIPS, 34:13032-13044, 2021. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14831" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "type": "text", + "content": "[57] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 140 + ], + "type": "text", + "content": "[58] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 140, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 296, + 183 + ], + "type": "text", + "content": "[59] Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, pages 4209-4219, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 227 + ], + "type": "text", + "content": "[60] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. NeurIPS, 34: 6087-6101, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 282 + ], + "type": "text", + "content": "[61] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM TOG, 42(4): 37:1-37:16, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 296, + 316 + ], + "type": "text", + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 296, + 370 + ], + "type": "text", + "content": "[63] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In CVPR, pages 19615-19625, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 415 + ], + "type": "text", + "content": "[64] Sanghyun Son, Matheus Gadelha, Yang Zhou, Matthew Fisher, Zexiang Xu, Yi-Ling Qiao, Ming C Lin, and Yi Zhou. Dmesh++: An efficient differentiable mesh for complex shapes. arXiv preprint arXiv:2412.16776, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 449 + ], + "type": "text", + "content": "[65] Sanghyun Son, Matheus Gadelha, Yang Zhou, Zexiang Xu, Ming C Lin, and Yi Zhou. Dmesh: A differentiable mesh representation. arXiv preprint arXiv:2404.13445, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 482 + ], + "type": "text", + "content": "[66] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In CVPR, pages 1798-1808, 2021. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 296, + 525 + ], + "type": "text", + "content": "[67] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In ECCV, pages 1-18. Springer, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 570 + ], + "type": "text", + "content": "[68] Jiaxiang Tang, Zhaoshuo Li, Zekun Hao, Xian Liu, Gang Zeng, Ming-Yu Liu, and Qinsheng Zhang. Edgerunner: Auto-regressive auto-encoder for artistic mesh generation. arXiv preprint arXiv:2409.18114, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "text", + "content": "[69] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. *NeurIPS*, 35:10021-10039, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 614, + 296, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 636 + ], + "type": "text", + "content": "[70] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 637, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 296, + 669 + ], + "type": "text", + "content": "[71] Peng Wang and Yichun Shi. Imaginedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[72] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 72, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 72, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 72, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 555, + 129 + ], + "type": "text", + "content": "[73] Yizhi Wang, Wallace Lira, Wenqi Wang, Ali Mahdavi-Amiri, and Hao Zhang. Slice3d: Multi-slice occlusion-revealing single view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9881-9891, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "type": "text", + "content": "[74] Zhengyi Wang, Yikai Wang, Yifei Chen, Chengdong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. Crm: Single image to 3d textured mesh with convolutional reconstruction model. In ECCV, pages 57-74, 2024. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 555, + 239 + ], + "type": "text", + "content": "[75] Xinyue Wei, Fanbo Xiang, Sai Bi, Anpei Chen, Kalyan Sunkavalli, Zexiang Xu, and Hao Su. Neumanifold: Neural watertight manifold reconstruction with efficient and high-quality rendering support. arXiv preprint arXiv:2305.17134, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 241, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 241, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 320, + 241, + 555, + 296 + ], + "type": "text", + "content": "[76] Xinyue Wei, Kai Zhang, Sai Bi, Hao Tan, Fujun Luan, Valentin Deschaintre, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Meshlrm: Large reconstruction model for high-quality meshes. arXiv preprint arXiv:2404.12385, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 297, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 297, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 555, + 352 + ], + "type": "text", + "content": "[77] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 353, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 353, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 353, + 555, + 398 + ], + "type": "text", + "content": "[78] Shuang Wu, Youtian Lin, Feihu Zhang, Yifei Zeng, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. arXiv preprint arXiv:2405.14832, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 399, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 555, + 453 + ], + "type": "text", + "content": "[79] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 454, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 454, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 454, + 555, + 487 + ], + "type": "text", + "content": "[80] Hongyi Xu and Jernej Barbič. Signed distance fields for polygon soup meshes. In Graphics Interface, pages 35-41. 2014. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 487, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 487, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 487, + 555, + 544 + ], + "type": "text", + "content": "[81] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 544, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 544, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 320, + 544, + 555, + 599 + ], + "type": "text", + "content": "[82] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 600, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 600, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 320, + 600, + 555, + 634 + ], + "type": "text", + "content": "[83] Qun-Ce Xu, Tai-Jiang Mu, and Yong-Liang Yang. A survey of deep learning-based 3d shape generation. Computational Visual Media, 9(3):407-442, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 635, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 555, + 690 + ], + "type": "text", + "content": "[84] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. In ECCV, pages 1-20. Springer, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 691, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 555, + 715 + ], + "type": "text", + "content": "[85] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d:" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14832" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 95 + ], + "type": "text", + "content": "A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 96, + 297, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 96, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 96, + 297, + 162 + ], + "type": "text", + "content": "[86] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, Xin Li, and Wenping Wang. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438, 2024. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "type": "text", + "content": "[87] Biao Zhang and Peter Wonka. Lagem: A large geometry model for 3d representation learning and diffusion, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 296, + 229 + ], + "type": "text", + "content": "[88] Biao Zhang, Matthias Nießner, and Peter Wonka. 3dilig: Irregular latent grids for 3d generative modeling. NeurIPS, 35:21871-21885, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 231, + 296, + 274 + ], + "type": "text", + "content": "[89] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM TOG, 42 (4):92:1-92:16, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 296, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 296, + 330 + ], + "type": "text", + "content": "[90] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: A structured and explicit radiance representation for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 296, + 354 + ], + "type": "text", + "content": "[91] Biao Zhang, Jing Ren, and Peter Wonka. Geometry distributions. arXiv preprint arXiv:2411.16076, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 356, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 356, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 296, + 399 + ], + "type": "text", + "content": "[92] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splattering. In ECCV, pages 1-19. Springer, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 401, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 296, + 454 + ], + "type": "text", + "content": "[93] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM TOG, 43(4):1-20, 2024. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 511 + ], + "type": "text", + "content": "[94] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, and Ming-Ming Cheng. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 513, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 296, + 567 + ], + "type": "text", + "content": "[95] Xuying Zhang, Bo-Wen Yin, Yuming Chen, Zheng Lin, Yunheng Li, Qibin Hou, and Ming-Ming Cheng. Temo: Towards text-driven 3d stylization for multi-object meshes. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 19531-19540, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "type": "text", + "content": "[96] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 657 + ], + "type": "text", + "content": "[97] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In CVPR, pages 16259-16268, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[98] Zibo Zhao, Wen Liu, Xin Chen, Xianfang Zeng, Rui Wang, Pei Cheng, Bin Fu, Tao Chen, Gang Yu, and Shenghua Gao. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation. NeurIPS, 36: 73969-73982, 2023. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 262 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[99] Xinyang Zheng, Yang Liu, Pengshuai Wang, and Xin Tong. Sdf-stylegan: implicit sdfs-based stylegan for 3d shape generation. In Computer Graphics Forum, pages 52-63, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 161 + ], + "type": "text", + "content": "[100] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM TOG, 42(4):1-13, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "text", + "content": "[101] Zi-Xin Zou, Shi-Sheng Huang, Yan-Pei Cao, Tai-Jiang Mu, Ying Shan, Hongbo Fu, and Song-Hai Zhang. Gp-recon: Online monocular neural 3d reconstruction with geometric prior. IEEE TVCG, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "text", + "content": "[102] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splattering: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14833" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_content_list.json b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1173227b5970483caf58bb547c932593b908f072 --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "SparseLaneSTP: Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maximilian Pittner1,2, Joel Janai1, Mario Faigle*1,3, Alexandru Paul Condurache1,2", + "bbox": [ + 173, + 202, + 821, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Bosch Mobility Solutions, Robert Bosch GmbH", + "bbox": [ + 305, + 220, + 692, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Neuro- and Bioinformatics, University of Lübeck", + "bbox": [ + 254, + 238, + 741, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute for Signal Processing and System Theory, University of Stuttgart", + "bbox": [ + 204, + 256, + 792, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Maximilian.Pittner, Joel.Janai, Mario.Faigle, AlexandruPaul.Condurache}@de.bosch.com", + "bbox": [ + 120, + 276, + 870, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D lane detection has emerged as a critical challenge in autonomous driving, encompassing identification and localization of lane markings and the 3D road surface. Conventional 3D methods detect lanes from dense birds-eyeviewed (BEV) features, though erroneous transformations often result in a poor feature representation misaligned with the true 3D road surface. While recent sparse lane detectors have surpassed dense BEV approaches, they completely disregard valuable lane-specific priors. Furthermore, existing methods fail to utilize historic lane observations, which yield the potential to resolve ambiguities in situations of poor visibility. To address these challenges, we present SparseLaneSTP, a novel method that integrates both geometric properties of the lane structure and temporal information into a sparse lane transformer. It introduces a new lane-specific spatio-temporal attention mechanism, a continuous lane representation tailored for sparse architectures as well as temporal regularization.", + "bbox": [ + 89, + 357, + 485, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Identifying weaknesses of existing 3D lane datasets, we also introduce a precise and consistent 3D lane dataset using a simple yet effective auto-labeling strategy. Our experimental section proves the benefits of our contributions and demonstrates state-of-the-art performance across all detection and error metrics on existing 3D lane detection benchmarks as well as on our novel dataset.", + "bbox": [ + 89, + 630, + 483, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate and robust 3D lane detection forms a pivotal task in autonomous driving addressing safe and reliable identification and localization of the road surface and lane markings. While significant progress has been made in 2D lane detection [10, 13, 17-19, 21, 35, 36, 40, 41, 46, 48, 55], the prediction output is only provided in image space lacking", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth information, which is crucial for autonomous navigation in the 3D world. In contrast, 3D lane detection jointly estimates lane markings and the road surface, directly producing 3D lanes in a vehicle-centered coordinate system. Conventional methods operate on dense feature maps that are transformed from the front-viewed (FV) to the birds-eye-viewed (BEV) perspective using inverse perspective mapping (IPM) [9, 11, 38] or learned mappings [6, 39, 49]. Although this step yields a suitable intermediate representation to capture the road, potential errors in the transformation cause misalignment between the resulting BEV and the true road surface that can hardly be compensated in the subsequent lane estimation step.", + "bbox": [ + 511, + 328, + 906, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As opposed to this two-stage approach, sparse detection methods have gained popularity in object detection [34] and have recently been adapted to 3D lane detection [2, 31]. These methods model 3D lane points as queries and associate them with uncorrupted FV image features - avoiding the necessity of error-prone BEV representations. Queries composed of learned context embedding vectors are integrated into transformer architectures [3, 54] leveraging their ability to learn global context. However, current approaches completely ignore valuable well-known priors about lane and road geometry. Spatial priors have already been utilized in [39]. Based on lane continuity and other properties like line parallelism, a continuous lane representation was introduced together with an efficient regularization. Despite notable advances brought by this approach, the architecture is based on dense BEV representations and the proposed priors have not been successfully customized for sparse architectures. Another class of prior knowledge is grounded in the history of observations. Among various temporal fusion paradigms, object-centric query propagation [50] has recently been proposed for 3D object detection in sparse architectures. In lane detection, temporal information has the potential to resolve ambiguities, particularly in scenarios with limited visibility or occlusions. Though, it has not", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done in context of master's thesis at Bosch.", + "bbox": [ + 112, + 886, + 382, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29099", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "been applied effectively so far.", + "bbox": [ + 89, + 90, + 294, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We therefore present a new method SparseLaneSTP that combines spatial and temporal priors with a sparse 3D lane transformer architecture. Our model features a transformertailored continuous lane representation providing smooth curves directly. This, combined with robust regularization using additional spatial and temporal objectives, enables the model to learn regression and visibility estimation more accurately. Additionally, spatial and temporal knowledge is integrated into the attention mechanism in a novel way. Instead of standard global self-attention, we design the layers to focus on relevant relations based on lane-structure. Finally, lane queries are temporally propagated and incorporated into the attention to leverage valuable historic keys.", + "bbox": [ + 89, + 106, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On top of that, we provide a novel 3D lane dataset using a simple yet efficient auto-labeling pipeline. Existing datasets often yield inaccurate labels containing noise and outliers, particularly in the far-range, since they rely on Li-DAR to recover 3D information. Our auto-labeling instead uses temporal aggregation along video-sequences resulting in consistent and accurate 3D lane labels up to 250 meters, making a valuable contribution to the research community.", + "bbox": [ + 89, + 301, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 89, + 424, + 413, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel 3D lane detector that integrates spatial and temporal knowledge into a sparse transformer.", + "- We propose a new attention mechanism that focuses on learning lane-structure and successfully leverages temporal keys from past lane observations.", + "- We formulate a continuous lane representation tailored for sparse architectures and a temporal regularization.", + "- Our method achieves state-of-the-art performance on all 3D lane benchmarks.", + "- We provide a new dataset for 3D lane detection providing accurate and consistent long-range labels." + ], + "bbox": [ + 89, + 439, + 480, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 618, + 227, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense BEV vs. sparse query-based detection. A common strategy in 3D lane detection involves operating on an intermediate dense BEV representation generated from the FV. Early methods, such as 3D-LaneNet [9], GenLaneNet [11] and 3D-SpLineNet [38], applied IPM to project FV features onto a flat ground plane. However, the assumption of a perfectly flat road is often violated, leading to degraded BEV features. Alternatively, methods from related fields like 3D object detection [14, 15, 23] and BEV segmentation [37] proposed to learn the BEV representation. These concepts were applied to 3D lane detection with PersFormer [6] leveraging deformable cross-attention (DCA) [54], whereas BEV-LaneDet [49] relies on simple multi-layer perceptrons (MLPs). LaneCPP [39] incorporates surface priors to learn 3D features using depth classification inspired by Lift-Splat-Shoot (LSS) [37]. Despite these advancements, all methods remain constrained by an", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "intermediate error-prone BEV representation.", + "bbox": [ + 511, + 90, + 813, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, sparse methods based on DETR [3, 54] avoid BEV representation by modeling 3D objects as queries that exchange information via attention mechanisms. In 3D object detection, 3D position-aware object queries were introduced that are directly associated with the perspective view through DCA [26, 34]. Similarly, in 3D lane detection, CurveFormer [2] adapts the sparse query design to lanes using polynomial-based curve queries, though these lack the flexibility to capture diverse lane structures in real-world scenarios. LATR [31] refines this approach by representing lanes as individual points but fails to incorporate crucial priors in both the attention mechanism and lane representation.", + "bbox": [ + 511, + 109, + 903, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Priors in lane detection. An indisputable property about lanes lies in their smoothness and continuity. Several methods have proposed continuous lane representations in both 2D [8, 29, 47] and 3D [2, 28, 38, 39] lane detection. In contrast to discrete approaches [6, 9, 11, 16, 49], continuous methods directly provide smooth curves, requiring almost no post-processing and exploit the entire available dense ground truth during training [38, 39]. While CLGO [28] and CurveFormer [2] use simple polynomials, 3D-SpLineNet [38] and LaneCPP [39] use a B-Spline [7] representation, which offers local control over curve segments enabling it to model complex shapes. Additionally, lane structures exhibit inherent properties, such as parallelism between lanes. E.g. SGNet [44] introduces a penalty term to enforce a fixed lane width, but assumes a flat ground plane. GP [20] employs a parallelism loss to maintain a constant local distance between neighboring points, though its effectiveness is influenced by the number of anchor points used. In contrast, LaneCPP [39] demonstrates a more elegant way to encourage parallelism as well as other spatial properties like surface smoothness and curvature restriction using regularization losses. Motivated by LaneCPP, we propose a more suitable Catmull-Rom spline [4] representation tailored for the transformer's attention with sparse queries. We further present more sophisticated loss formulations and integrate prior knowledge into the transformer attention mechanism enhancing the focus on relevant relations. While a concurrent work also uses priors in attention [5], our approach focuses on explicit modeling of structural lane relations and even effectively incorporates temporal knowledge in lane detection for the first time.", + "bbox": [ + 511, + 292, + 903, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal modeling. Another source of prior knowledge is grounded in the history of observations. For video-based 3D object detection various approaches exist to model temporal interactions. BEV-based methods [23] transform past BEV features according to the ego-motion and fuse them with the current representation. More advanced perspective view (PV) methods [26, 30, 32] instead follow the sparse query design but require storing dense historic PV feature maps to interact with object queries. Recently,", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29100", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58a284d290683b79f2026809a8d222e2fc12b28edc9cc3ac10e419e411115b1f.jpg", + "image_caption": [ + "Figure 1. Overview of our method SparseLaneSTP. Sparse lane queries are processed by a transformer integrating temporal knowledge and spatial lane structure priors in a novel spatio-temporal attention. Based on these priors, we formulate spatial and temporally consistent regularization. Finally, our network predicts control points defining our new continuous 3D lane representation." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 908, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "object-centric temporal modeling [27, 50] has been proposed, which only stores sparse historic object queries and propagates them according to ego- and object-motion in 3D space. In particular, Sparse4Dv2 [27] and StreamPETR [50] leverage the sparse query design and model object queries as hidden states of a transformer, where the relation between current and historic queries is learned via attention.", + "bbox": [ + 88, + 398, + 480, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In 3D lane detection, only few approaches exist that use temporal modeling. STLane3D [51] follows a BEV-based method, PETRv2 [30] fuses historic PV features together with propagated 3D positional embeddings. On the other hand, Anchor3DLane-T [16] leverages temporal information by associating 3D anchors from the current frame with features from previous frames. In contrast, our sparse query-based architecture is more suitable for the effective memory-efficient object-centric paradigm. Consequently, we propose a temporal attention mechanism inspired by [27, 50], which exploits lane-specific properties by reinforcing most relevant query interactions. The static nature of lane markings also allows us to formulate regularization to encourage temporally consistent perception behavior.", + "bbox": [ + 88, + 507, + 482, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 737, + 225, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we describe our 3D lane detection method illustrated in Fig. 1. Given an image $\\mathbf{I}$ , the model predicts a set of $N$ lanes as continuous curves in 3D space, parameterized by our novel tailored spline representation with $M$ control points $\\mathbf{P}$ (see Section 3.1) and lane categories $\\mathbf{C}$ . The core component of our architecture is the transformer decoder, where each control point, inspired by sparse query design, is paired with a learned context embedding and is iteratively processed through $L$ decoder layers.", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8a5ac857e3e0ced946926e3dd3ec2b1c5b5dfd232cea74fe234a70ddcef4a7c.jpg", + "image_caption": [ + "(a) Front-view" + ], + "image_footnote": [], + "bbox": [ + 514, + 396, + 717, + 492 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a683ecd4aee8f9635024464c782e4c9ad2a3abddbf8cc2e707a62a96e4c54e2.jpg", + "image_caption": [ + "(b) BEV", + "Figure 2. CR spline and B-Spline control points in comparison, both fit to the ground truth. B-Spline control points do not align with the curve and are therefore not suited the sparse query design, whereas CR control points exactly match curve geometry." + ], + "image_footnote": [], + "bbox": [ + 750, + 398, + 883, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to previous work [2, 24, 31] that simply adapt the DETR [3, 54] framework, our transformer architecture introduces a novel spatio-temporal attention layer to capture spatial intra- and inter-lane relations, while a temporal cross-attention mechanism integrates past observations via propagated lane queries from a memory queue. Finally, besides standard detection losses, our model is cautiously regularized by spatial and temporal objectives to enhance robust and consistent detection behavior.", + "bbox": [ + 511, + 604, + 906, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D lane representation", + "text_level": 1, + "bbox": [ + 511, + 755, + 723, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the 3D lane representation, we take inspiration from prior work [38, 39] that demonstrated significant benefits of continuous curves. In these frameworks, the model output serves as control points of B-Spline curves that do not necessarily align with the curve geometry. However, using the sparse query design, the 3D position of control points serve as the model's internal state and should therefore correspond to the exact 3D position of lane points. Thus, spline", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "control points should directly lie on the curve, which is not the case for B-Splines (see Fig. 2). Therefore, we choose Catmull-Rom (CR) splines [4], a class of piece-wise defined smooth third-order polynomial splines, for which the curve inherently passes through its control points.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adapt this spline-model to define a curve representation $\\mathbf{f}_i(s) \\in \\mathbb{R}^4$ representing the entire $i^{\\mathrm{th}}$ line proposal parameterized by $M$ control points with 3 spatial $(x_{ij}, y_{ij}, z_{ij})$ and 1 visibility $(v_{ij})$ dimension as", + "bbox": [ + 89, + 167, + 483, + 228 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i} (s) = \\left[ \\begin{array}{l l l l} s ^ {3} & s ^ {2} & s & 1 \\end{array} \\right] \\cdot \\mathbf {M} _ {\\mathrm {C R}} \\cdot \\mathbf {P} _ {i} \\quad \\text {w i t h} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 238, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i} = \\left[ \\mathbf {P} _ {\\mathrm {3 D}, i} \\mid \\mathbf {P} _ {v, i} \\right] = \\left[ \\begin{array}{c c c c} x _ {i 1} & y _ {i 1} & z _ {i 1} & v _ {i 1} \\\\ & \\dots & & \\\\ x _ {i j} & y _ {i j} & z _ {i j} & v _ {i j} \\\\ & \\dots & & \\\\ x _ {i M} & y _ {i M} & z _ {i M} & v _ {i M} \\end{array} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 258, + 482, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with curve argument $s\\in [0,1]$ $\\mathbf{P}_i\\in \\mathbb{R}^{M\\times 4}$ denotes the control point matrix composed of $\\mathbf{P}_{3\\mathrm{D},i}\\in \\mathbb{R}^{M\\times 3}$ defining the curve 3D shape and visibility $\\mathbf{P}_{v,i}\\in \\mathbb{R}^{M\\times 1}$ $[\\cdot |\\cdot ]$ is the concatenation operator and $\\mathbf{M}_{\\mathrm{CR}}$ the CR coefficient matrix. Note that we can eventually express the curve computation as a simple matrix multiplication between the pre-computed spline arguments and the entire predicted control points $\\mathbf{P}$", + "bbox": [ + 89, + 349, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall architecture", + "text_level": 1, + "bbox": [ + 89, + 467, + 282, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Backbone and query initialization. The RGB input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ is first processed by a CNN backbone (e.g. ResNet [12]) to extract image features of $\\mathbf{F} \\in \\mathbb{R}^{H_{\\mathbf{F}} \\times W_{\\mathbf{F}} \\times C}$ , where $C$ is the channel size. Similar to [31], we then feed $\\mathbf{F}$ through a lane instance segmentation branch to obtain initial query embedding vectors ${}^0\\mathbf{Q} \\in \\mathbb{R}^{N \\times M \\times C}$ , where $N$ denotes the number of line proposals and $M$ the number of spline control points. An MLP then predicts initial control points ${}^0\\mathbf{P} \\in \\mathbb{R}^{N \\times M \\times 4}$ from ${}^0\\mathbf{Q}$ .", + "bbox": [ + 89, + 489, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer decoder. In each transformer decoder layer $l$ , queries are processed by two attention layers", + "bbox": [ + 89, + 626, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } = \\operatorname { S T A } \\left( { } ^ { l - 1 } \\mathbf { Q } , { } ^ { l - 1 } \\mathbf { P } , \\mathbf { Q } _ { \\text { M e m } } , \\mathbf { P } _ { \\text { M e m } } \\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 666, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { l } \\mathbf { Q } _ { \\mathrm { D C A } } = \\mathrm { D C A } \\big ( { } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } , \\mathbf { F } , { } ^ { l - 1 } \\mathbf { P } _ { 2 \\mathrm { D } } \\big ) , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 688, + 480, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathrm{STA}(\\cdot)$ denotes our proposed spatio-temporal attention, which is explained in Section 3.3, and $\\mathrm{DCA}(\\cdot)$ the standard deformable cross-attention presented in [54]. $\\mathbf{P}_{2\\mathrm{D}}$ are the 2D spline control points in image coordinates resulting from re-projecting $\\mathbf{P}_{3\\mathrm{D}}$ . Finally, the query embeddings are fed through a feed-forward network $^l\\mathbf{Q} = \\mathrm{FFN}(^l\\mathbf{Q}_{\\mathrm{DCA}})$ composed of an MLP and layer normalization.", + "bbox": [ + 89, + 719, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction layer. After each transformer attention layer, the emerging query embeddings ${}^{l}\\mathbf{Q}$ are fed through a prediction layer that outputs the spline control points ${}^{l}\\mathbf{P}$ as well as classification probabilities $\\mathbf{C} \\in \\mathbb{R}^{N \\times (K + 1)}$ for $K$ lane categories and one background class per line. Particularly,", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d36f3edc002eec1903da03a9d36119061be4262d0382ab5a775fe04e10924fc2.jpg", + "image_caption": [ + "Figure 3. Overview of our spatio-temporal attention module." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each layer $l$ and query ${}^{l}\\mathbf{Q}_{ij}$ three MLPs with shared weights among all layers are used to predict the $x-$ , $z-$ and $v$ -component of the control point matrix", + "bbox": [ + 511, + 319, + 906, + 366 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { l } \\mathbf { P } _ { x , i j } = \\sigma ( \\mathrm { M L P } _ { x } \\left( { } ^ { l } \\mathbf { Q } _ { i j } \\right) ) \\cdot \\left( x _ { e } - x _ { s } \\right) + x _ { s } \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 373, + 903, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { l } \\mathbf { P } _ { z , i j } = \\sigma ( \\mathrm { M L P } _ { z } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) \\cdot ( z _ { e } - z _ { s } ) + z _ { s } \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 393, + 903, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n{ } ^ { l } \\mathbf { P } _ { v , i j } = \\sigma ( \\mathrm { M L P } _ { v } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) , ( 7 )\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 416, + 903, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the sigmoid $\\sigma (\\cdot)$ normalizes the MLP output to [0, 1]. The start-/end- of $x$ - and $z$ -range $x_{s / e}$ , $z_{s / e}$ scales up the normalized output to the desired range. The longitudinal $y$ -component ${}^l\\mathbf{P}_{y,ij}$ is pre-defined and uniformly distributed between $y_{s}$ and $y_{e}$ to avoid over-parameterization and potential overfitting as argued in [39]. For classification, all queries belonging to the same line proposal are averaged and fed into an MLP followed by a softmax layer.", + "bbox": [ + 511, + 441, + 906, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the formulation of our supervised losses is inspired by [39], with L1 loss for regression of $x$ and $z$ , binary cross-entropy for visibility $v$ and focal loss [25] for classification.", + "bbox": [ + 511, + 564, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Integrating spatial and temporal knowledge", + "text_level": 1, + "bbox": [ + 511, + 617, + 883, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory queue. To facilitate attention to past observations, queries are stored within a memory queue. Following a similar approach to [50], this queue is recursively updated after each prediction step and retains queries from the past $T$ frames. Therefore, embedding vectors and their associated control points are recursively pushed into the queue using a first-in, first-out (FIFO) strategy, which yields", + "bbox": [ + 511, + 638, + 906, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\text {M e m}} = \\left[ \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 1)} \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - T)} \\right] \\quad \\text {a n d} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 753, + 903, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\mathrm {M e m}} = \\left[ \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 1)} \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - T)} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 777, + 903, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_{:N_{\\mathrm{Mem}}}^{(t)}\\in \\mathbb{R}^{(N_{\\mathrm{Mem}}\\times M\\times C)}$ denotes the embedding vectors from the $N_{\\mathrm{Mem}}$ most confident predictions of frame $t$ and $\\mathbf{Q}_{\\mathrm{Mem}}\\in \\mathbb{R}^{((T\\cdot N_{\\mathrm{Mem}})\\times M\\times C)}$ the full memory queue.", + "bbox": [ + 511, + 806, + 903, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal propagation. Since the model is supposed to learn spatio-temporal context, it must incorporate the positional information of 3D control points from past frames.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29102", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, due to the vehicle's ego-motion, lane positions shift relative to the vehicle. Therefore, the 3D coordinates of past lane control points stored in the memory queue must be transformed according to the ego-pose $\\mathbf{E}^{(t)}$ of frame $t$ , which is represented as a homogeneous transformation matrix. Consequently, a control point $\\mathbf{P}_{ij}^{(t - k)\\rightarrow (t)}$ propagated from frame $t - k$ to present frame $t$ is given as", + "bbox": [ + 89, + 90, + 483, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\left[ \\mathbf {E} _ {i n v} ^ {(t)} \\cdot \\mathbf {E} ^ {(t - k)} \\cdot \\mathbf {P} _ {\\mathrm {3 D}, i j} ^ {(t - k)} \\mid \\mathbf {P} _ {v, i j} ^ {(t - k)} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 204, + 482, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_{inv}$ denotes the inverted matrix $\\mathbf{E}$ . Obviously, only the geometry component is affected by the propagation, whereas the visibility component remains unchanged. To incorporate spatio-temporal information from propagated control points into queries, we apply a positional encoding (PE) based on 3D geometry and visibility (see Fig. 3). This process yields position-informed embedding vectors", + "bbox": [ + 89, + 232, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {Q}} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\mathbf {Q} _ {i j} ^ {(t - k) \\rightarrow (t)} + \\mathrm {P E} \\left(\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 343, + 482, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which are processed by the spatio-temporal attention layer. Spatio-temporal attention. Most methods [2, 31] rely solely on global self-attention, which contains a majority of redundant lane query interactions that can distract the model from learning relevant context. To address this, we leverage spatial and temporal priors, distinguishing three key relation types: intra-lane (between adjacent points on the same lane), inter-lane (between parallel lanes), and historic (between current and past queries in the memory queue).", + "bbox": [ + 89, + 371, + 482, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We therefore introduce same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA). Applying specific masking, SLA restricts interactions to queries $\\tilde{\\mathbf{Q}}_{\\mathrm{SLA}}$ within the same line, PNA facilitates interactions with queries $\\tilde{\\mathbf{Q}}_{\\mathrm{PNA}}$ from neighboring lines, and TCA enables interactions with closest past queries $\\tilde{\\mathbf{Q}}_{\\mathrm{TCA}}$ in the memory queue (see Fig. 3). Such restriction of query interactions reduces redundancy, encourages the model to focus on critical lane-specific spatial context and integrates valuable temporal information from past observations.", + "bbox": [ + 89, + 507, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regularization. Based on spatial and temporal priors we can formulate regularization techniques that promote robust and consistent detection behavior. For the spatial regularization $\\mathcal{L}_{\\text {spatial }}$ we adopt the scheme proposed in [39] encouraging lane parallelism, surface smoothness and suppressing excessive curvature. Besides, temporal consistency can be achieved by optimizing a smoothness loss", + "bbox": [ + 89, + 657, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t e m p} = \\frac {1}{N} \\sum_ {i} ^ {N} \\int_ {s} \\overline {{\\mathbf {f}}} _ {v, i} ^ {(t)} (s) \\cdot \\left| \\left| \\mathbf {f} _ {\\mathrm {3 D}, i} (s) - \\overline {{\\mathbf {f}}} _ {\\mathrm {3 D}, i} ^ {(t)} (s) \\right| \\right| _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 770, + 482, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{\\mathbf{f}}_{3\\mathrm{D},i}^{(t)}(s)$ or $\\overline{\\mathbf{f}}_{v,i}^{(t)}(s)$ is an exponentially moving average of the past predictions", + "bbox": [ + 89, + 819, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {f}}} _ {i} ^ {(t)} = \\alpha \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t)} + (1 - \\alpha) \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t - 1)} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 858, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with smoothing factor $\\alpha \\in [0,1]$ .", + "bbox": [ + 89, + 885, + 310, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fda28af3d24d1f53837be1f37fc3ab7a591e3c4296b656c452cc522035b8533a.jpg", + "image_caption": [ + "Figure 4. Our auto-labeling strategy." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ecde0324ebefa099ca401ef2b608f2e31a5ccbea365a1f208f0eebf6f577adf9.jpg", + "image_caption": [ + "(a) OpenLane", + "Figure 5. Examples from OpenLane compared to ours." + ], + "image_footnote": [], + "bbox": [ + 517, + 273, + 709, + 407 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b69ab776c619896e7369f6884e623b91a59630cd36aa4981715490cff32ede5d.jpg", + "image_caption": [ + "(b) Our 3D lane dataset" + ], + "image_footnote": [], + "bbox": [ + 712, + 273, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Auto-labeling and our 3D lane dataset", + "text_level": 1, + "bbox": [ + 511, + 474, + 856, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe our auto-labeling pipeline and provide details about our new 3D lane dataset.", + "bbox": [ + 511, + 502, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Auto-labeling strategy", + "text_level": 1, + "bbox": [ + 511, + 545, + 718, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While 3D lane datasets like OpenLane [6] and ONCE-3DLanes [52] have contributed significantly to the field, they possess specific shortcomings that our auto-labeling strategy addresses. As illustrated in Fig. 5a the LiDAR-based annotation approach, which recovers 3D information of 2D lane labels, introduces noise and outliers, particularly in the far range due to sparse scan points, while also posing calibration and synchronization challenges. To overcome these limitations, we propose an automated labeling approach (see Fig. 4) that leverages confident near-range lane point estimates from a 2D detector, which are temporally accumulated in 3D space along an accurate ego-trajectory. More precisely, common visual odometry methods [1, 42, 43] are applied to recover the vehicle ego-trajectory from the input video sequence. Our road surface generator then fits a spline surface to the vehicle's trajectory and orientations. On the other hand, lane pseudo-labels in image space are obtained from a state-of-the-art 2D lane detector [21, 46], providing particularly reliable detection results in the near-range. Lane and surface information are then combined in 3D by projecting near-range detection results to the road surface and accumulating line points frame", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29103", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/45007fe7c541efa29d5b0b60c92d84b92fdfdd84020bbd290e108e3adf4c121f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 222, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/966afd6af52639b3b08bf7c93796f58a06b8177ddbae6d2dec1674f6a72ebf3f.jpg", + "image_caption": [ + "(a) $t = 1$" + ], + "image_footnote": [], + "bbox": [ + 99, + 184, + 210, + 231 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f09b95709d6a03753724bf7ca9e067d9c2bd2913f359c8bf0ba7d9ce109a5e9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 88, + 349, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8cae87e5cbccb743668b6d9968a55f484020858cff834198213833a2bb68383.jpg", + "image_caption": [ + "(b) $t = 2$" + ], + "image_footnote": [], + "bbox": [ + 225, + 184, + 338, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13099c230a58274e3bdfe6fb7038a3cc0cbec47c5a9c24c3127a5ba81df1aaf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 88, + 477, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bcf7f0b866de407cb7f9a0ddc78bda559a27cd0223ebaf863f8b59a5e2a2792.jpg", + "image_caption": [ + "(c) $t = 3$", + "Figure 6. Qualitative comparison of two models, one without temporal and one with temporal for three consecutive frames." + ], + "image_footnote": [], + "bbox": [ + 354, + 184, + 467, + 231 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by frame along the ego-trajectory of the sequence. Finally, we apply an accurate semantic segmentation network to automatically assign occlusion labels to each lane point.", + "bbox": [ + 89, + 313, + 483, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset properties", + "text_level": 1, + "bbox": [ + 89, + 368, + 267, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the approach described in 4.1, we are able to automatically generate temporally consistent, high-quality 3D lane labels in a range up to $250\\mathrm{m}$ (see Fig. 5b). Besides 3D lane geometry, we provide camera intrinsics and extrinsics, lane categories, lane visibility information and consistent track IDs. We further supply relational information about lanes (like line parallelism), local 3D ego-trajectory and global pose information for each frame. Important to mention is also that in contrast to former datasets, our visibility information contains occlusions caused by other traffic participants, enabling models to consistently learn lane structure independent of interruptions and lane occlusions. We refer to the supplementary material for additional information.", + "bbox": [ + 89, + 391, + 483, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 599, + 220, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first outline our experimental setup, followed by a comprehensive analysis of our approach on two public and our own novel 3D lane dataset.", + "bbox": [ + 89, + 626, + 483, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and evaluation metrics", + "text_level": 1, + "bbox": [ + 89, + 681, + 370, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OpenLane [6] based on the Waymo Open dataset [45] remains the most prominent real-world 3D lane dataset to date. It is comprised of 200K images from 1000 sequences recorded in several cities in the USA.", + "bbox": [ + 89, + 703, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ONCE-3DLanes [52] is another real-world dataset derived from the ONCE dataset [33]. It consists of 211K images collected from multiple cities in China.", + "bbox": [ + 89, + 763, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our 3D lane dataset, which was introduced in Section 4, was gathered from multiple regions and countries across the globe. It consists of 511K images with a $90\\%$ / $10\\%$ train/test split.", + "bbox": [ + 89, + 810, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. The OpenLane evaluation scheme [6, 11] uses the euclidean distance at uniformly distributed", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8d9996ddeae8a02bbfa50f4a47bf46679370ddbffa55984c1d7dc49beb15b97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Continuous Rep.STARegularizationF1(%)↑Gain(%)
61.8(baseline)
✓62.9+1.1
✓✓65.0+2.1
✓✓✓65.3+0.3
", + "bbox": [ + 517, + 88, + 903, + 161 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5980bb39de87b938b580cbcfb4bda04813643ef7ec8057d7c263c43eb8f1dba0.jpg", + "table_caption": [ + "Table 1. Performance gain for different contributions on OpenLane using our novel Continuous Lane Representation, Spatio-Temporal Attention (STA) and Regularization." + ], + "table_footnote": [], + "table_body": "
AttentionGlobalSLAPNASLA + PNASLA + PNA + TCA
F1-Score62.962.863.563.865.0
", + "bbox": [ + 517, + 232, + 903, + 263 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb510596631316ddc44b66109bb556746b5b68fa428381c5918c008e003997d4.jpg", + "table_caption": [ + "Table 2. Analysis of spatio-temporal attention. Same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA) and combinations compared to global attention." + ], + "table_footnote": [], + "table_body": "
Num. Frames (T)012345
F1(%)↑63.864.164.665.064.563.9
", + "bbox": [ + 517, + 345, + 903, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Analysis of number of frames in the memory for STA.", + "bbox": [ + 519, + 393, + 895, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "points and a range-IoU to compute the F1-Score. Moreover, x- and z-errors in near- $(0 - 40\\mathrm{m})$ and far-range $(40 - 100\\mathrm{m})$ evaluate the geometric accuracy. ONCE-3DLanes instead uses the uni-lateral chamfer distance (CD) and computes Precision (P), Recall (R) and F1-Score based on it.", + "bbox": [ + 511, + 436, + 903, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our dataset, we extend the well-established OpenLane evaluation framework [6, 11]. Due to the long-range labels, we add x-/z-error intervals for the ranges $100 - 150\\mathrm{m}$ and $150 - 200\\mathrm{m}$ . Since our dataset offers visibility labels for occlusions, we also add a visibility Intersection-over-Union metric (Vis-IoU). This score helps to gain understanding about the model's capability of distinguishing visible lane segments from parts occluded by other traffic participants and provides a measure for start-end range detection.", + "bbox": [ + 511, + 515, + 905, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 665, + 725, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use an image input size $720 \\times 960$ and ResNet-50 [12] as our standard backbone. The query generator is trained on the auxiliary task of lane instance segmentation similar to [31]. The transformer decoder applied in the large model used for the evaluation in Section 5.4 consists of $L = 6$ layers, whereas the lighter model used for ablation studies in Sec. 5.3 has 2 layers. The lane representation uses $M = 20$ control points on OpenLane and ONCE-3DLanes and 30 control points on our dataset due to the larger range. The number of lane proposals is set to $N = 40$ for OpenLane and ours and 10 for ONCE-3DLanes, which is sufficient for the small number of occurring lanes per frame. Regarding temporal fusion, our best model uses propagated queries from $T = 3$ historic frames.", + "bbox": [ + 511, + 688, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29104", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9972d0ed71a3c90c0ce3f0d7e9d2cff87b91be34b25a26cbad3a46a4531f5f48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneResolutionF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [9]VGG-16360 x 48044.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [11]ERFNet360 x 48032.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [6]EfficientNet-B7360 x 48050.50.4850.5530.3640.43142.455.648.646.640.050.7
CurveFormer [2]EfficientNet-B7360 x 48050.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [49]ResNet-34360 x 48058.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane-T [16]ResNet-18360 x 48054.30.2750.3100.1050.13547.258.052.748.745.851.7
PETRv2 [30]VoV-99360 x 48061.20.4000.5730.2650.413------
LATR [31]ResNet-50720 x 96061.90.2190.2590.0750.10455.268.257.155.452.361.5
LaneCPP [39]EfficientNet-B7360 x 48060.30.2640.3100.0770.11753.664.456.754.952.058.7
PVALane [53]Swin-B720 x 96063.40.2260.2570.0930.11956.167.764.058.653.660.8
GroupLane [22]ConvNext-Basenot specified64.10.3200.4410.2330.402------
SparseLaneSTP (Ours)ResNet-50720 x 96066.10.2030.2400.0660.09257.373.060.158.358.266.5
", + "bbox": [ + 94, + 88, + 906, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/14840fbe39534a1f66d702d98ca73ae90a67205e9397384a485d38242e81be0e.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on OpenLane [6]. Best performance and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodF1-Score(%)↑X-error / Z-error (m)↓Vis-IoU(%)↑
0 m - 40 m40 m - 100 m100 m - 150 m150 m - 200 m
PersFormer [6]59.20.191 / 0.0720.344 / 0.1450.671 / 0.2200.845 / 0.37970.1
LaneCPP [39]63.10.144 / 0.0400.298 / 0.0930.494 / 0.1880.761 / 0.29377.5
Baseline (LATR) [31]65.10.122 / 0.0380.263 / 0.0980.468 / 0.1850.700 / 0.26478.3
SparseLaneSTP (ours)68.20.109 / 0.0330.233 / 0.0920.443 / 0.1820.646 / 0.26381.4
", + "bbox": [ + 96, + 285, + 906, + 382 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59add4d0edbc3287de048c01130ac5cceeeecf0846d50a479317c0862b580d40.jpg", + "image_caption": [ + "Figure 7. Efficiency analysis of SparseLaneSTP." + ], + "image_footnote": [], + "bbox": [ + 148, + 430, + 424, + 517 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/624221499df66012ed5d127dcb1fcd873988ffcbbe512e5ffcd64711f73b1882.jpg", + "table_caption": [ + "Table 5. Quantitative comparison on our 3D lane dataset. Best performance and second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodF1(%)↑P(%)↑R(%)↑CD(m)↓
3D-LaneNet [9]44.7361.4635.160.127
Gen-LaneNet [11]45.5963.9535.420.121
SALAD [52]64.0775.9055.420.098
PersFormer [6]74.3380.3069.180.074
Anchor3DLane [16]74.8780.8569.710.060
LATR [31]80.5986.1275.730.052
PVALane [53]76.3582.8170.830.059
GroupLane [22]80.7382.5678.900.053
SparseLaneSTP (ours)82.7586.4779.330.048
", + "bbox": [ + 96, + 560, + 482, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Quantitative comparison on ONCE-3DLanes [52].", + "bbox": [ + 109, + 708, + 460, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 753, + 251, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform a comprehensive analysis validating the effectiveness of our design choices in SparseLaneSTP. For these experiments, we used a smaller model consisting of two decoder layers evaluated on OpenLane.", + "bbox": [ + 89, + 777, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 summarizes the impact of our contributions. Our CR spline based lane representation improves the F1-Score by over 1 percentage point compared to the discrete one (baseline). The most significant gain in F1-Score is", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieved by the novel STA component, which encourages focus on relevant relations based on lane structure and, more importantly, on related past observations by integrating temporal cross-attention. Eventually, regularization enhances robustness and consistency, leading to improved generalization, as evidenced by the additional F1-Score gain.", + "bbox": [ + 511, + 433, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, global self-attention apparently fails to capture important relations of lane structures. The combination of our proposed SLA and PNA demonstrates the benefits of enhancing the focus on intra- and interlane relations by an F1-Score increase $+0.9\\%$ . Integrating knowledge about past observations via our suggested TCA yields the greatest improvement of additional $+1.2\\%$ . As indicated by Table 3, $T = 3$ frames in the memory produces the best results. Using too few frames provides insufficient temporal context, whereas too many frames apparently introduce redundancy and distract the model from relevant information. We further qualitatively demonstrate the beneficial impact of our primary contribution on temporal fusion in Fig. 6. The non-temporal model fails to retain the detection from frame $t = 1$ in subsequent frames as lane visibility deteriorates. In contrast, the temporally-aware model maintains the detection despite partial occlusions, demonstrating improved robustness. Regularization further preserves temporal consistency mitigating the gradual disappearing and drifting in the non-temporal model.", + "bbox": [ + 511, + 530, + 908, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, Fig. 7 reveals efficiency benefits of our approach. SparseLaneSTP with six decoder layers achieves 11.0 FPS, comparable to LATR at 12.1 FPS. While our proposed temporal integration (memory queue and STA) in", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29105", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e60b3b80c7c73572554363de529587302b91ef5c04f758e9b59908a04491d22.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 205, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dff13bace1f0129b4c3e707079de744a9e73e04046c362ddb98a09018cff4c3e.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison of LATR and SparseLaneSTP on OpenLane with ground truth. Main differences are highlighted." + ], + "image_footnote": [], + "bbox": [ + 205, + 88, + 297, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a52c94e084f557f206f90e78bed8b67d07d194a86863776cf820913af80061a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 297, + 88, + 406, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/751ce85b9ff1b7970c5939aef8016806f8ab1ebef5cf3997a2520a4733c9491e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 90, + 500, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78d9e261e7932fc92997426db18bc46ebaf145014f211d2754e5a7f988c6aff1.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 607, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef425db9dba7b94391769e8a6def0d3476c889caa25df0ea86ad485e2cc8c8e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 90, + 700, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a69dd879bf2780b9f47cbab5b2ddc8dacadb901e0e2b6879c4d34fe2f278fe1.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 700, + 88, + 810, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b69532c9906cecac45d277e2297489022d81ce5eae3f58bd375ce8cbea91d2a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 90, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "produces a modest $9\\%$ overhead, it also enables a smaller model with only two layers to reach $65.3\\%$ F1-Score $(+3.4\\%$ over six-layer LATR) at a significantly higher speed of 16.5 FPS.", + "bbox": [ + 89, + 292, + 483, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Main results", + "text_level": 1, + "bbox": [ + 89, + 364, + 223, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we compare our method to other 3D lane detectors on public benchmarks as well as on our own dataset. Results on OpenLane. Table 4 presents a quantitative comparison demonstrating the superiority of our approach. SparseLaneSTP reaches the highest F1-Score, exceeding $66\\%$ , significantly surpassing the second-best model despite its larger backbone and even achieves the lowest geometric errors.", + "bbox": [ + 89, + 386, + 483, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 illustrates the advantages of our approach qualitatively across multiple aspects. We compare our SparseLaneSTP to LATR [31], which is most related to our baseline. As shown in Fig. 8a, our continuous lane representation enables precise visibility estimation with accurate detection of start/end of markings, whereas LATR - constrained by fixed discrete anchor points - produces erroneous estimates. In Fig. 8b, our temporal fusion preserves the line despite partial occlusions, whereas non-temporal LATR fails. It also struggles to estimate faint lane markings on the far side of the junction shown in Fig. 8c, while with STA our model captures lane structures more effectively. Eventually, in Fig. 8d our model exhibits a stronger reliance on priors resulting in robust and consistent behavior, most likely due to the incorporated regularization.", + "bbox": [ + 89, + 508, + 483, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on ONCE-3DLanes. As shown in Table 6, our method achieves the highest detection scores and demonstrates accurate regression with a significantly lower CD, despite the inaccurate camera parameters provided by the dataset. Notably, even though ONCE-3DLanes has a lower frame rate, our temporal approach performs effectively.", + "bbox": [ + 89, + 734, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on our 3D lane dataset. Initial results on our novel dataset are presented in Table 5. Our proposed extension of the evaluation metrics provides a more comprehensive assessment of overall performance, particularly in the far-range and for occlusion detection. To facilitate a fair", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31f9551c46bb17502bde5d110bc06b96fc2a9d14e306edb0acd1e72e614aa638.jpg", + "image_caption": [ + "Front-view" + ], + "image_footnote": [], + "bbox": [ + 535, + 290, + 741, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c773caa969f771df28e62a6d442b824794f8eeb5ef38f8c74e42a6c3b01b53c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 290, + 883, + 343 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/798bbab2a95bb344bb0a6b015891784b153d81168102718a5c417f74c6927f5c.jpg", + "image_caption": [ + "Figure 9. Qualitative example of SparseLaneSTP on our data." + ], + "image_footnote": [], + "bbox": [ + 745, + 343, + 883, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison, we train the baseline and other methods with specific modifications to range and capacity. Our method excels in F1-Score, error metrics and notably in the VisIoU, highlighting the superior performance of our temporal model in detecting occlusions. Fig. 9 shows a qualitative result on our dataset, showcasing accurate long-range lane estimation $(200\\mathrm{m})$ . For more qualitative results with comparisons to the baseline we refer to the supplementary.", + "bbox": [ + 511, + 449, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 584, + 633, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented a novel 3D lane detection approach that combines sparse lane transformers with spatial and temporal priors. We proposed a continuous lane representation, tailored for the sparse query design, thereby overcoming limitations of discrete representations. Additionally, we developed a novel attention layer that effectively captures intra- and inter-lane relations while integrating temporal information through propagated lane queries. Furthermore, we introduced a dedicated regularization scheme enhancing temporal consistency. Our experiments validate the effective integration of temporal information and spatial priors, as well as the benefits of our lane representation, leading to state-of-the-art performance across all evaluated 3D lane datasets. Moreover, we presented a new 3D lane dataset, constructed using a simple yet effective auto-labeling pipeline, providing geometrically accurate labels that span long distances and explicitly annotate occlusions, fostering further advancements in the field. In future, we will explore an extension to 3D lane tracking to leverage the full potential of temporal information.", + "bbox": [ + 511, + 608, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29106", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lucas R Agostinho, Nuno M Ricardo, Maria I Pereira, Antoine Hiolle, and Andry M Pinto. A practical survey on visual odometry for autonomous driving in challenging scenarios and conditions. IEEE Access, 2022. 5", + "[2] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 5, 7", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3", + "[4] Edwin Catmull and Raphael Rom. A class of local interpolating splines. Computer aided geometric design, 1972. 2, 4", + "[5] Yifan Chang, Junjie Huang, Xiaofeng Wang, Yun Ye, Zhujin Liang, Yi Shan, Dalong Du, and Xingang Wang. Rethinking lanes and points in complex scenarios for monocular 3d lane detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2", + "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 1, 2, 5, 6, 7", + "[7] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2", + "[8] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[9] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 7", + "[10] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1", + "[11] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 6, 7", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6", + "[13] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1", + "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. CoRR," + ], + "bbox": [ + 93, + 114, + 483, + 887 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "abs/2203.17054, 2022. 2", + "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR, abs/2112.11790, 2021. 2", + "[16] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7", + "[17] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1", + "[18] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020.", + "[19] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1", + "[20] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[21] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 5", + "[22] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Groupplane: End-to-end 3d lane detection with channel-wise grouping. IEEE Robotics and Automation Letters (RA-L), 2024. 7", + "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2", + "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In Proc. of the International Conf. on Learning Representations (ICLR), 2023. 3", + "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 4", + "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. CoRR, abs/2211.10581, 2022. 2", + "[27] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. CoRR, abs/2305.14018, 2023. 3", + "[28] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Ze" + ], + "bbox": [ + 516, + 93, + 903, + 891 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "jian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 2", + "[29] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petr2: A unified framework for 3d perception from multi-camera images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 2, 3, 7", + "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[32] Zhipeng Luo, Changqing Zhou, Gongjie Zhang, and Shijian Lu. Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849, 2022. 2", + "[33] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, and Hang Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 6", + "[34] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 1, 2", + "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1", + "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1", + "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2", + "[38] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3", + "[39] Maximilian Pittner, Joel Janai, and Alexandru P Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 7", + "[40] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1", + "[41] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[42] Davide Scaramuzza, Friedrich Fraundorfer, and Roland Siegwart. Real-time monocular visual odometry for on-road" + ], + "bbox": [ + 91, + 90, + 483, + 891 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vehicles with 1-point ransac. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2009. 5", + "[43] Frank Steinbrücker, Jürgen Sturm, and Daniel Cremers. Real-time visual odometry from dense rgb-d images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV) Workshops, 2011. 5", + "[44] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2", + "[45] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 5", + "[47] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2", + "[48] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[49] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 7", + "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 3, 4", + "[51] Yin Wang, Qiuyi Guo, Peiwen Lin, Guangliang Cheng, and Jian Wu. Spatio-temporal fusion-based monocular 3d lane detection. In Proc. of the British Machine Vision Conf. (BMVC), 2022. 3", + "[52] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7", + "[53] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2024. 7", + "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Proc. of the International Conf. on Learning Representations (ICLR), 2021. 1, 2, 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29108", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[55] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1", + "bbox": [ + 91, + 90, + 482, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29109", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_model.json b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35276665edef562bee54a40ff3ae9aea76c767a0 --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "SparseLaneSTP: Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.203, + 0.823, + 0.221 + ], + "angle": 0, + "content": "Maximilian Pittner1,2, Joel Janai1, Mario Faigle*1,3, Alexandru Paul Condurache1,2" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.222, + 0.693, + 0.239 + ], + "angle": 0, + "content": "1Bosch Mobility Solutions, Robert Bosch GmbH" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.239, + 0.743, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Institute of Neuro- and Bioinformatics, University of Lübeck" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.257, + 0.793, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)Institute for Signal Processing and System Theory, University of Stuttgart" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.277, + 0.871, + 0.293 + ], + "angle": 0, + "content": "{Maximilian.Pittner, Joel.Janai, Mario.Faigle, AlexandruPaul.Condurache}@de.bosch.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.63 + ], + "angle": 0, + "content": "3D lane detection has emerged as a critical challenge in autonomous driving, encompassing identification and localization of lane markings and the 3D road surface. Conventional 3D methods detect lanes from dense birds-eyeviewed (BEV) features, though erroneous transformations often result in a poor feature representation misaligned with the true 3D road surface. While recent sparse lane detectors have surpassed dense BEV approaches, they completely disregard valuable lane-specific priors. Furthermore, existing methods fail to utilize historic lane observations, which yield the potential to resolve ambiguities in situations of poor visibility. To address these challenges, we present SparseLaneSTP, a novel method that integrates both geometric properties of the lane structure and temporal information into a sparse lane transformer. It introduces a new lane-specific spatio-temporal attention mechanism, a continuous lane representation tailored for sparse architectures as well as temporal regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.485, + 0.736 + ], + "angle": 0, + "content": "Identifying weaknesses of existing 3D lane datasets, we also introduce a precise and consistent 3D lane dataset using a simple yet effective auto-labeling strategy. Our experimental section proves the benefits of our contributions and demonstrates state-of-the-art performance across all detection and error metrics on existing 3D lane detection benchmarks as well as on our novel dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Accurate and robust 3D lane detection forms a pivotal task in autonomous driving addressing safe and reliable identification and localization of the road surface and lane markings. While significant progress has been made in 2D lane detection [10, 13, 17-19, 21, 35, 36, 40, 41, 46, 48, 55], the prediction output is only provided in image space lacking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.526 + ], + "angle": 0, + "content": "depth information, which is crucial for autonomous navigation in the 3D world. In contrast, 3D lane detection jointly estimates lane markings and the road surface, directly producing 3D lanes in a vehicle-centered coordinate system. Conventional methods operate on dense feature maps that are transformed from the front-viewed (FV) to the birds-eye-viewed (BEV) perspective using inverse perspective mapping (IPM) [9, 11, 38] or learned mappings [6, 39, 49]. Although this step yields a suitable intermediate representation to capture the road, potential errors in the transformation cause misalignment between the resulting BEV and the true road surface that can hardly be compensated in the subsequent lane estimation step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As opposed to this two-stage approach, sparse detection methods have gained popularity in object detection [34] and have recently been adapted to 3D lane detection [2, 31]. These methods model 3D lane points as queries and associate them with uncorrupted FV image features - avoiding the necessity of error-prone BEV representations. Queries composed of learned context embedding vectors are integrated into transformer architectures [3, 54] leveraging their ability to learn global context. However, current approaches completely ignore valuable well-known priors about lane and road geometry. Spatial priors have already been utilized in [39]. Based on lane continuity and other properties like line parallelism, a continuous lane representation was introduced together with an efficient regularization. Despite notable advances brought by this approach, the architecture is based on dense BEV representations and the proposed priors have not been successfully customized for sparse architectures. Another class of prior knowledge is grounded in the history of observations. Among various temporal fusion paradigms, object-centric query propagation [50] has recently been proposed for 3D object detection in sparse architectures. In lane detection, temporal information has the potential to resolve ambiguities, particularly in scenarios with limited visibility or occlusions. Though, it has not" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.383, + 0.9 + ], + "angle": 0, + "content": "*Work done in context of master's thesis at Bosch." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "29099" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.295, + 0.106 + ], + "angle": 0, + "content": "been applied effectively so far." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.482, + 0.303 + ], + "angle": 0, + "content": "We therefore present a new method SparseLaneSTP that combines spatial and temporal priors with a sparse 3D lane transformer architecture. Our model features a transformertailored continuous lane representation providing smooth curves directly. This, combined with robust regularization using additional spatial and temporal objectives, enables the model to learn regression and visibility estimation more accurately. Additionally, spatial and temporal knowledge is integrated into the attention mechanism in a novel way. Instead of standard global self-attention, we design the layers to focus on relevant relations based on lane-structure. Finally, lane queries are temporally propagated and incorporated into the attention to leverage valuable historic keys." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.424 + ], + "angle": 0, + "content": "On top of that, we provide a novel 3D lane dataset using a simple yet efficient auto-labeling pipeline. Existing datasets often yield inaccurate labels containing noise and outliers, particularly in the far-range, since they rely on Li-DAR to recover 3D information. Our auto-labeling instead uses temporal aggregation along video-sequences resulting in consistent and accurate 3D lane labels up to 250 meters, making a valuable contribution to the research community." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.425, + 0.415, + 0.438 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.481, + 0.469 + ], + "angle": 0, + "content": "- We present a novel 3D lane detector that integrates spatial and temporal knowledge into a sparse transformer." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.47, + 0.481, + 0.514 + ], + "angle": 0, + "content": "- We propose a new attention mechanism that focuses on learning lane-structure and successfully leverages temporal keys from past lane observations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.481, + 0.544 + ], + "angle": 0, + "content": "- We formulate a continuous lane representation tailored for sparse architectures and a temporal regularization." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.545, + 0.481, + 0.573 + ], + "angle": 0, + "content": "- Our method achieves state-of-the-art performance on all 3D lane benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.575, + 0.481, + 0.605 + ], + "angle": 0, + "content": "- We provide a new dataset for 3D lane detection providing accurate and consistent long-range labels." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.44, + 0.481, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.228, + 0.635 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dense BEV vs. sparse query-based detection. A common strategy in 3D lane detection involves operating on an intermediate dense BEV representation generated from the FV. Early methods, such as 3D-LaneNet [9], GenLaneNet [11] and 3D-SpLineNet [38], applied IPM to project FV features onto a flat ground plane. However, the assumption of a perfectly flat road is often violated, leading to degraded BEV features. Alternatively, methods from related fields like 3D object detection [14, 15, 23] and BEV segmentation [37] proposed to learn the BEV representation. These concepts were applied to 3D lane detection with PersFormer [6] leveraging deformable cross-attention (DCA) [54], whereas BEV-LaneDet [49] relies on simple multi-layer perceptrons (MLPs). LaneCPP [39] incorporates surface priors to learn 3D features using depth classification inspired by Lift-Splat-Shoot (LSS) [37]. Despite these advancements, all methods remain constrained by an" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.814, + 0.107 + ], + "angle": 0, + "content": "intermediate error-prone BEV representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.905, + 0.291 + ], + "angle": 0, + "content": "In contrast, sparse methods based on DETR [3, 54] avoid BEV representation by modeling 3D objects as queries that exchange information via attention mechanisms. In 3D object detection, 3D position-aware object queries were introduced that are directly associated with the perspective view through DCA [26, 34]. Similarly, in 3D lane detection, CurveFormer [2] adapts the sparse query design to lanes using polynomial-based curve queries, though these lack the flexibility to capture diverse lane structures in real-world scenarios. LATR [31] refines this approach by representing lanes as individual points but fails to incorporate crucial priors in both the attention mechanism and lane representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.905, + 0.761 + ], + "angle": 0, + "content": "Priors in lane detection. An indisputable property about lanes lies in their smoothness and continuity. Several methods have proposed continuous lane representations in both 2D [8, 29, 47] and 3D [2, 28, 38, 39] lane detection. In contrast to discrete approaches [6, 9, 11, 16, 49], continuous methods directly provide smooth curves, requiring almost no post-processing and exploit the entire available dense ground truth during training [38, 39]. While CLGO [28] and CurveFormer [2] use simple polynomials, 3D-SpLineNet [38] and LaneCPP [39] use a B-Spline [7] representation, which offers local control over curve segments enabling it to model complex shapes. Additionally, lane structures exhibit inherent properties, such as parallelism between lanes. E.g. SGNet [44] introduces a penalty term to enforce a fixed lane width, but assumes a flat ground plane. GP [20] employs a parallelism loss to maintain a constant local distance between neighboring points, though its effectiveness is influenced by the number of anchor points used. In contrast, LaneCPP [39] demonstrates a more elegant way to encourage parallelism as well as other spatial properties like surface smoothness and curvature restriction using regularization losses. Motivated by LaneCPP, we propose a more suitable Catmull-Rom spline [4] representation tailored for the transformer's attention with sparse queries. We further present more sophisticated loss formulations and integrate prior knowledge into the transformer attention mechanism enhancing the focus on relevant relations. While a concurrent work also uses priors in attention [5], our approach focuses on explicit modeling of structural lane relations and even effectively incorporates temporal knowledge in lane detection for the first time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Temporal modeling. Another source of prior knowledge is grounded in the history of observations. For video-based 3D object detection various approaches exist to model temporal interactions. BEV-based methods [23] transform past BEV features according to the ego-motion and fuse them with the current representation. More advanced perspective view (PV) methods [26, 30, 32] instead follow the sparse query design but require storing dense historic PV feature maps to interact with object queries. Recently," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29100" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.086, + 0.91, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.331, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 1. Overview of our method SparseLaneSTP. Sparse lane queries are processed by a transformer integrating temporal knowledge and spatial lane structure priors in a novel spatio-temporal attention. Based on these priors, we formulate spatial and temporally consistent regularization. Finally, our network predicts control points defining our new continuous 3D lane representation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.482, + 0.506 + ], + "angle": 0, + "content": "object-centric temporal modeling [27, 50] has been proposed, which only stores sparse historic object queries and propagates them according to ego- and object-motion in 3D space. In particular, Sparse4Dv2 [27] and StreamPETR [50] leverage the sparse query design and model object queries as hidden states of a transformer, where the relation between current and historic queries is learned via attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.721 + ], + "angle": 0, + "content": "In 3D lane detection, only few approaches exist that use temporal modeling. STLane3D [51] follows a BEV-based method, PETRv2 [30] fuses historic PV features together with propagated 3D positional embeddings. On the other hand, Anchor3DLane-T [16] leverages temporal information by associating 3D anchors from the current frame with features from previous frames. In contrast, our sparse query-based architecture is more suitable for the effective memory-efficient object-centric paradigm. Consequently, we propose a temporal attention mechanism inspired by [27, 50], which exploits lane-specific properties by reinforcing most relevant query interactions. The static nature of lane markings also allows us to formulate regularization to encourage temporally consistent perception behavior." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.226, + 0.755 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In the following, we describe our 3D lane detection method illustrated in Fig. 1. Given an image \\(\\mathbf{I}\\), the model predicts a set of \\(N\\) lanes as continuous curves in 3D space, parameterized by our novel tailored spline representation with \\(M\\) control points \\(\\mathbf{P}\\) (see Section 3.1) and lane categories \\(\\mathbf{C}\\). The core component of our architecture is the transformer decoder, where each control point, inspired by sparse query design, is paired with a learned context embedding and is iteratively processed through \\(L\\) decoder layers." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.397, + 0.718, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.495, + 0.672, + 0.506 + ], + "angle": 0, + "content": "(a) Front-view" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.399, + 0.884, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.494, + 0.85, + 0.506 + ], + "angle": 0, + "content": "(b) BEV" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 2. CR spline and B-Spline control points in comparison, both fit to the ground truth. B-Spline control points do not align with the curve and are therefore not suited the sparse query design, whereas CR control points exactly match curve geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.741 + ], + "angle": 0, + "content": "In contrast to previous work [2, 24, 31] that simply adapt the DETR [3, 54] framework, our transformer architecture introduces a novel spatio-temporal attention layer to capture spatial intra- and inter-lane relations, while a temporal cross-attention mechanism integrates past observations via propagated lane queries from a memory queue. Finally, besides standard detection losses, our model is cautiously regularized by spatial and temporal objectives to enhance robust and consistent detection behavior." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.725, + 0.772 + ], + "angle": 0, + "content": "3.1. 3D lane representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "For the 3D lane representation, we take inspiration from prior work [38, 39] that demonstrated significant benefits of continuous curves. In these frameworks, the model output serves as control points of B-Spline curves that do not necessarily align with the curve geometry. However, using the sparse query design, the 3D position of control points serve as the model's internal state and should therefore correspond to the exact 3D position of lane points. Thus, spline" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "control points should directly lie on the curve, which is not the case for B-Splines (see Fig. 2). Therefore, we choose Catmull-Rom (CR) splines [4], a class of piece-wise defined smooth third-order polynomial splines, for which the curve inherently passes through its control points." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.484, + 0.229 + ], + "angle": 0, + "content": "We adapt this spline-model to define a curve representation \\(\\mathbf{f}_i(s) \\in \\mathbb{R}^4\\) representing the entire \\(i^{\\mathrm{th}}\\) line proposal parameterized by \\(M\\) control points with 3 spatial \\((x_{ij}, y_{ij}, z_{ij})\\) and 1 visibility \\((v_{ij})\\) dimension as" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.239, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i} (s) = \\left[ \\begin{array}{l l l l} s ^ {3} & s ^ {2} & s & 1 \\end{array} \\right] \\cdot \\mathbf {M} _ {\\mathrm {C R}} \\cdot \\mathbf {P} _ {i} \\quad \\text {w i t h} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.26, + 0.483, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i} = \\left[ \\mathbf {P} _ {\\mathrm {3 D}, i} \\mid \\mathbf {P} _ {v, i} \\right] = \\left[ \\begin{array}{c c c c} x _ {i 1} & y _ {i 1} & z _ {i 1} & v _ {i 1} \\\\ & \\dots & & \\\\ x _ {i j} & y _ {i j} & z _ {i j} & v _ {i j} \\\\ & \\dots & & \\\\ x _ {i M} & y _ {i M} & z _ {i M} & v _ {i M} \\end{array} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.457 + ], + "angle": 0, + "content": "with curve argument \\(s\\in [0,1]\\) \\(\\mathbf{P}_i\\in \\mathbb{R}^{M\\times 4}\\) denotes the control point matrix composed of \\(\\mathbf{P}_{3\\mathrm{D},i}\\in \\mathbb{R}^{M\\times 3}\\) defining the curve 3D shape and visibility \\(\\mathbf{P}_{v,i}\\in \\mathbb{R}^{M\\times 1}\\) \\([\\cdot |\\cdot ]\\) is the concatenation operator and \\(\\mathbf{M}_{\\mathrm{CR}}\\) the CR coefficient matrix. Note that we can eventually express the curve computation as a simple matrix multiplication between the pre-computed spline arguments and the entire predicted control points \\(\\mathbf{P}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.284, + 0.482 + ], + "angle": 0, + "content": "3.2. Overall architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Backbone and query initialization. The RGB input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is first processed by a CNN backbone (e.g. ResNet [12]) to extract image features of \\(\\mathbf{F} \\in \\mathbb{R}^{H_{\\mathbf{F}} \\times W_{\\mathbf{F}} \\times C}\\), where \\(C\\) is the channel size. Similar to [31], we then feed \\(\\mathbf{F}\\) through a lane instance segmentation branch to obtain initial query embedding vectors \\({}^0\\mathbf{Q} \\in \\mathbb{R}^{N \\times M \\times C}\\), where \\(N\\) denotes the number of line proposals and \\(M\\) the number of spline control points. An MLP then predicts initial control points \\({}^0\\mathbf{P} \\in \\mathbb{R}^{N \\times M \\times 4}\\) from \\({}^0\\mathbf{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Transformer decoder. In each transformer decoder layer \\(l\\), queries are processed by two attention layers" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.667, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\n{ } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } = \\operatorname { S T A } \\left( { } ^ { l - 1 } \\mathbf { Q } , { } ^ { l - 1 } \\mathbf { P } , \\mathbf { Q } _ { \\text { M e m } } , \\mathbf { P } _ { \\text { M e m } } \\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.689, + 0.482, + 0.709 + ], + "angle": 0, + "content": "\\[\n{ } ^ { l } \\mathbf { Q } _ { \\mathrm { D C A } } = \\mathrm { D C A } \\big ( { } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } , \\mathbf { F } , { } ^ { l - 1 } \\mathbf { P } _ { 2 \\mathrm { D } } \\big ) , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.825 + ], + "angle": 0, + "content": "\\(\\mathrm{STA}(\\cdot)\\) denotes our proposed spatio-temporal attention, which is explained in Section 3.3, and \\(\\mathrm{DCA}(\\cdot)\\) the standard deformable cross-attention presented in [54]. \\(\\mathbf{P}_{2\\mathrm{D}}\\) are the 2D spline control points in image coordinates resulting from re-projecting \\(\\mathbf{P}_{3\\mathrm{D}}\\). Finally, the query embeddings are fed through a feed-forward network \\(^l\\mathbf{Q} = \\mathrm{FFN}(^l\\mathbf{Q}_{\\mathrm{DCA}})\\) composed of an MLP and layer normalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Prediction layer. After each transformer attention layer, the emerging query embeddings \\({}^{l}\\mathbf{Q}\\) are fed through a prediction layer that outputs the spline control points \\({}^{l}\\mathbf{P}\\) as well as classification probabilities \\(\\mathbf{C} \\in \\mathbb{R}^{N \\times (K + 1)}\\) for \\(K\\) lane categories and one background class per line. Particularly," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.282, + 0.89, + 0.296 + ], + "angle": 0, + "content": "Figure 3. Overview of our spatio-temporal attention module." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.907, + 0.367 + ], + "angle": 0, + "content": "for each layer \\(l\\) and query \\({}^{l}\\mathbf{Q}_{ij}\\) three MLPs with shared weights among all layers are used to predict the \\(x-\\), \\(z-\\) and \\(v\\)-component of the control point matrix" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.374, + 0.905, + 0.393 + ], + "angle": 0, + "content": "\\[\n{ } ^ { l } \\mathbf { P } _ { x , i j } = \\sigma ( \\mathrm { M L P } _ { x } \\left( { } ^ { l } \\mathbf { Q } _ { i j } \\right) ) \\cdot \\left( x _ { e } - x _ { s } \\right) + x _ { s } \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.395, + 0.905, + 0.414 + ], + "angle": 0, + "content": "\\[\n{ } ^ { l } \\mathbf { P } _ { z , i j } = \\sigma ( \\mathrm { M L P } _ { z } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) \\cdot ( z _ { e } - z _ { s } ) + z _ { s } \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.417, + 0.905, + 0.436 + ], + "angle": 0, + "content": "\\[\n{ } ^ { l } \\mathbf { P } _ { v , i j } = \\sigma ( \\mathrm { M L P } _ { v } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) , ( 7 )\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.564 + ], + "angle": 0, + "content": "where the sigmoid \\(\\sigma (\\cdot)\\) normalizes the MLP output to [0, 1]. The start-/end- of \\(x\\)- and \\(z\\)-range \\(x_{s / e}\\), \\(z_{s / e}\\) scales up the normalized output to the desired range. The longitudinal \\(y\\)-component \\({}^l\\mathbf{P}_{y,ij}\\) is pre-defined and uniformly distributed between \\(y_{s}\\) and \\(y_{e}\\) to avoid over-parameterization and potential overfitting as argued in [39]. For classification, all queries belonging to the same line proposal are averaged and fed into an MLP followed by a softmax layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Finally, the formulation of our supervised losses is inspired by [39], with L1 loss for regression of \\( x \\) and \\( z \\), binary cross-entropy for visibility \\( v \\) and focal loss [25] for classification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.885, + 0.635 + ], + "angle": 0, + "content": "3.3. Integrating spatial and temporal knowledge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.746 + ], + "angle": 0, + "content": "Memory queue. To facilitate attention to past observations, queries are stored within a memory queue. Following a similar approach to [50], this queue is recursively updated after each prediction step and retains queries from the past \\( T \\) frames. Therefore, embedding vectors and their associated control points are recursively pushed into the queue using a first-in, first-out (FIFO) strategy, which yields" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.754, + 0.905, + 0.776 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\text {M e m}} = \\left[ \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 1)} \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - T)} \\right] \\quad \\text {a n d} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.778, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\mathrm {M e m}} = \\left[ \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 1)} \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - T)} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.905, + 0.856 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_{:N_{\\mathrm{Mem}}}^{(t)}\\in \\mathbb{R}^{(N_{\\mathrm{Mem}}\\times M\\times C)}\\) denotes the embedding vectors from the \\(N_{\\mathrm{Mem}}\\) most confident predictions of frame \\(t\\) and \\(\\mathbf{Q}_{\\mathrm{Mem}}\\in \\mathbb{R}^{((T\\cdot N_{\\mathrm{Mem}})\\times M\\times C)}\\) the full memory queue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Temporal propagation. Since the model is supposed to learn spatio-temporal context, it must incorporate the positional information of 3D control points from past frames." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.199 + ], + "angle": 0, + "content": "However, due to the vehicle's ego-motion, lane positions shift relative to the vehicle. Therefore, the 3D coordinates of past lane control points stored in the memory queue must be transformed according to the ego-pose \\(\\mathbf{E}^{(t)}\\) of frame \\(t\\), which is represented as a homogeneous transformation matrix. Consequently, a control point \\(\\mathbf{P}_{ij}^{(t - k)\\rightarrow (t)}\\) propagated from frame \\(t - k\\) to present frame \\(t\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.205, + 0.483, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\left[ \\mathbf {E} _ {i n v} ^ {(t)} \\cdot \\mathbf {E} ^ {(t - k)} \\cdot \\mathbf {P} _ {\\mathrm {3 D}, i j} ^ {(t - k)} \\mid \\mathbf {P} _ {v, i j} ^ {(t - k)} \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.484, + 0.339 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_{inv}\\) denotes the inverted matrix \\(\\mathbf{E}\\). Obviously, only the geometry component is affected by the propagation, whereas the visibility component remains unchanged. To incorporate spatio-temporal information from propagated control points into queries, we apply a positional encoding (PE) based on 3D geometry and visibility (see Fig. 3). This process yields position-informed embedding vectors" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.344, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {Q}} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\mathbf {Q} _ {i j} ^ {(t - k) \\rightarrow (t)} + \\mathrm {P E} \\left(\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.507 + ], + "angle": 0, + "content": "which are processed by the spatio-temporal attention layer. Spatio-temporal attention. Most methods [2, 31] rely solely on global self-attention, which contains a majority of redundant lane query interactions that can distract the model from learning relevant context. To address this, we leverage spatial and temporal priors, distinguishing three key relation types: intra-lane (between adjacent points on the same lane), inter-lane (between parallel lanes), and historic (between current and past queries in the memory queue)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.483, + 0.658 + ], + "angle": 0, + "content": "We therefore introduce same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA). Applying specific masking, SLA restricts interactions to queries \\(\\tilde{\\mathbf{Q}}_{\\mathrm{SLA}}\\) within the same line, PNA facilitates interactions with queries \\(\\tilde{\\mathbf{Q}}_{\\mathrm{PNA}}\\) from neighboring lines, and TCA enables interactions with closest past queries \\(\\tilde{\\mathbf{Q}}_{\\mathrm{TCA}}\\) in the memory queue (see Fig. 3). Such restriction of query interactions reduces redundancy, encourages the model to focus on critical lane-specific spatial context and integrates valuable temporal information from past observations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Regularization. Based on spatial and temporal priors we can formulate regularization techniques that promote robust and consistent detection behavior. For the spatial regularization \\(\\mathcal{L}_{\\text {spatial }}\\) we adopt the scheme proposed in [39] encouraging lane parallelism, surface smoothness and suppressing excessive curvature. Besides, temporal consistency can be achieved by optimizing a smoothness loss" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.771, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t e m p} = \\frac {1}{N} \\sum_ {i} ^ {N} \\int_ {s} \\overline {{\\mathbf {f}}} _ {v, i} ^ {(t)} (s) \\cdot \\left| \\left| \\mathbf {f} _ {\\mathrm {3 D}, i} (s) - \\overline {{\\mathbf {f}}} _ {\\mathrm {3 D}, i} ^ {(t)} (s) \\right| \\right| _ {1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.483, + 0.854 + ], + "angle": 0, + "content": "where \\(\\overline{\\mathbf{f}}_{3\\mathrm{D},i}^{(t)}(s)\\) or \\(\\overline{\\mathbf{f}}_{v,i}^{(t)}(s)\\) is an exponentially moving average of the past predictions" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.859, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {f}}} _ {i} ^ {(t)} = \\alpha \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t)} + (1 - \\alpha) \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t - 1)} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.311, + 0.902 + ], + "angle": 0, + "content": "with smoothing factor \\(\\alpha \\in [0,1]\\)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.246, + 0.819, + 0.261 + ], + "angle": 0, + "content": "Figure 4. Our auto-labeling strategy." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.275, + 0.71, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.411, + 0.652, + 0.423 + ], + "angle": 0, + "content": "(a) OpenLane" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.275, + 0.905, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.411, + 0.871, + 0.423 + ], + "angle": 0, + "content": "(b) Our 3D lane dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.435, + 0.872, + 0.449 + ], + "angle": 0, + "content": "Figure 5. Examples from OpenLane compared to ours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.476, + 0.857, + 0.493 + ], + "angle": 0, + "content": "4. Auto-labeling and our 3D lane dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.533 + ], + "angle": 0, + "content": "In this section, we describe our auto-labeling pipeline and provide details about our new 3D lane dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.72, + 0.563 + ], + "angle": 0, + "content": "4.1. Auto-labeling strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "While 3D lane datasets like OpenLane [6] and ONCE-3DLanes [52] have contributed significantly to the field, they possess specific shortcomings that our auto-labeling strategy addresses. As illustrated in Fig. 5a the LiDAR-based annotation approach, which recovers 3D information of 2D lane labels, introduces noise and outliers, particularly in the far range due to sparse scan points, while also posing calibration and synchronization challenges. To overcome these limitations, we propose an automated labeling approach (see Fig. 4) that leverages confident near-range lane point estimates from a 2D detector, which are temporally accumulated in 3D space along an accurate ego-trajectory. More precisely, common visual odometry methods [1, 42, 43] are applied to recover the vehicle ego-trajectory from the input video sequence. Our road surface generator then fits a spline surface to the vehicle's trajectory and orientations. On the other hand, lane pseudo-labels in image space are obtained from a state-of-the-art 2D lane detector [21, 46], providing particularly reliable detection results in the near-range. Lane and surface information are then combined in 3D by projecting near-range detection results to the road surface and accumulating line points frame" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.223, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.185, + 0.212, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.235, + 0.183, + 0.245 + ], + "angle": 0, + "content": "(a) \\(t = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.089, + 0.351, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.185, + 0.339, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.235, + 0.312, + 0.245 + ], + "angle": 0, + "content": "(b) \\(t = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.089, + 0.478, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.185, + 0.468, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.235, + 0.439, + 0.245 + ], + "angle": 0, + "content": "(c) \\(t = 3\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of two models, one without temporal and one with temporal for three consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.36 + ], + "angle": 0, + "content": "by frame along the ego-trajectory of the sequence. Finally, we apply an accurate semantic segmentation network to automatically assign occlusion labels to each lane point." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.369, + 0.269, + 0.385 + ], + "angle": 0, + "content": "4.2. Dataset properties" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.588 + ], + "angle": 0, + "content": "Using the approach described in 4.1, we are able to automatically generate temporally consistent, high-quality 3D lane labels in a range up to \\(250\\mathrm{m}\\) (see Fig. 5b). Besides 3D lane geometry, we provide camera intrinsics and extrinsics, lane categories, lane visibility information and consistent track IDs. We further supply relational information about lanes (like line parallelism), local 3D ego-trajectory and global pose information for each frame. Important to mention is also that in contrast to former datasets, our visibility information contains occlusions caused by other traffic participants, enabling models to consistently learn lane structure independent of interruptions and lane occlusions. We refer to the supplementary material for additional information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.601, + 0.222, + 0.618 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.673 + ], + "angle": 0, + "content": "In this section, we first outline our experimental setup, followed by a comprehensive analysis of our approach on two public and our own novel 3D lane dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.371, + 0.697 + ], + "angle": 0, + "content": "5.1. Datasets and evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.763 + ], + "angle": 0, + "content": "OpenLane [6] based on the Waymo Open dataset [45] remains the most prominent real-world 3D lane dataset to date. It is comprised of 200K images from 1000 sequences recorded in several cities in the USA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.81 + ], + "angle": 0, + "content": "ONCE-3DLanes [52] is another real-world dataset derived from the ONCE dataset [33]. It consists of 211K images collected from multiple cities in China." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Our 3D lane dataset, which was introduced in Section 4, was gathered from multiple regions and countries across the globe. It consists of 511K images with a \\(90\\%\\) / \\(10\\%\\) train/test split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. The OpenLane evaluation scheme [6, 11] uses the euclidean distance at uniformly distributed" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.162 + ], + "angle": 0, + "content": "
Continuous Rep.STARegularizationF1(%)↑Gain(%)
61.8(baseline)
✓62.9+1.1
✓✓65.0+2.1
✓✓✓65.3+0.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.173, + 0.905, + 0.215 + ], + "angle": 0, + "content": "Table 1. Performance gain for different contributions on OpenLane using our novel Continuous Lane Representation, Spatio-Temporal Attention (STA) and Regularization." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.264 + ], + "angle": 0, + "content": "
AttentionGlobalSLAPNASLA + PNASLA + PNA + TCA
F1-Score62.962.863.563.865.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.316 + ], + "angle": 0, + "content": "Table 2. Analysis of spatio-temporal attention. Same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA) and combinations compared to global attention." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.347, + 0.905, + 0.383 + ], + "angle": 0, + "content": "
Num. Frames (T)012345
F1(%)↑63.864.164.665.064.563.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.394, + 0.897, + 0.408 + ], + "angle": 0, + "content": "Table 3. Analysis of number of frames in the memory for STA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.905, + 0.514 + ], + "angle": 0, + "content": "points and a range-IoU to compute the F1-Score. Moreover, x- and z-errors in near- \\((0 - 40\\mathrm{m})\\) and far-range \\((40 - 100\\mathrm{m})\\) evaluate the geometric accuracy. ONCE-3DLanes instead uses the uni-lateral chamfer distance (CD) and computes Precision (P), Recall (R) and F1-Score based on it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.652 + ], + "angle": 0, + "content": "For our dataset, we extend the well-established OpenLane evaluation framework [6, 11]. Due to the long-range labels, we add x-/z-error intervals for the ranges \\(100 - 150\\mathrm{m}\\) and \\(150 - 200\\mathrm{m}\\). Since our dataset offers visibility labels for occlusions, we also add a visibility Intersection-over-Union metric (Vis-IoU). This score helps to gain understanding about the model's capability of distinguishing visible lane segments from parts occluded by other traffic participants and provides a measure for start-end range detection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.727, + 0.682 + ], + "angle": 0, + "content": "5.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We use an image input size \\(720 \\times 960\\) and ResNet-50 [12] as our standard backbone. The query generator is trained on the auxiliary task of lane instance segmentation similar to [31]. The transformer decoder applied in the large model used for the evaluation in Section 5.4 consists of \\(L = 6\\) layers, whereas the lighter model used for ablation studies in Sec. 5.3 has 2 layers. The lane representation uses \\(M = 20\\) control points on OpenLane and ONCE-3DLanes and 30 control points on our dataset due to the larger range. The number of lane proposals is set to \\(N = 40\\) for OpenLane and ours and 10 for ONCE-3DLanes, which is sufficient for the small number of occurring lanes per frame. Regarding temporal fusion, our best model uses propagated queries from \\(T = 3\\) historic frames." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29104" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.248 + ], + "angle": 0, + "content": "
MethodBackboneResolutionF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [9]VGG-16360 x 48044.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [11]ERFNet360 x 48032.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [6]EfficientNet-B7360 x 48050.50.4850.5530.3640.43142.455.648.646.640.050.7
CurveFormer [2]EfficientNet-B7360 x 48050.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [49]ResNet-34360 x 48058.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane-T [16]ResNet-18360 x 48054.30.2750.3100.1050.13547.258.052.748.745.851.7
PETRv2 [30]VoV-99360 x 48061.20.4000.5730.2650.413------
LATR [31]ResNet-50720 x 96061.90.2190.2590.0750.10455.268.257.155.452.361.5
LaneCPP [39]EfficientNet-B7360 x 48060.30.2640.3100.0770.11753.664.456.754.952.058.7
PVALane [53]Swin-B720 x 96063.40.2260.2570.0930.11956.167.764.058.653.660.8
GroupLane [22]ConvNext-Basenot specified64.10.3200.4410.2330.402------
SparseLaneSTP (Ours)ResNet-50720 x 96066.10.2030.2400.0660.09257.373.060.158.358.266.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.19, + 0.258, + 0.807, + 0.273 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on OpenLane [6]. Best performance and second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.286, + 0.907, + 0.383 + ], + "angle": 0, + "content": "
MethodF1-Score(%)↑X-error / Z-error (m)↓Vis-IoU(%)↑
0 m - 40 m40 m - 100 m100 m - 150 m150 m - 200 m
PersFormer [6]59.20.191 / 0.0720.344 / 0.1450.671 / 0.2200.845 / 0.37970.1
LaneCPP [39]63.10.144 / 0.0400.298 / 0.0930.494 / 0.1880.761 / 0.29377.5
Baseline (LATR) [31]65.10.122 / 0.0380.263 / 0.0980.468 / 0.1850.700 / 0.26478.3
SparseLaneSTP (ours)68.20.109 / 0.0330.233 / 0.0920.443 / 0.1820.646 / 0.26381.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.174, + 0.393, + 0.822, + 0.407 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison on our 3D lane dataset. Best performance and second best are highlighted." + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.431, + 0.426, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.529, + 0.431, + 0.543 + ], + "angle": 0, + "content": "Figure 7. Efficiency analysis of SparseLaneSTP." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.561, + 0.483, + 0.699 + ], + "angle": 0, + "content": "
MethodF1(%)↑P(%)↑R(%)↑CD(m)↓
3D-LaneNet [9]44.7361.4635.160.127
Gen-LaneNet [11]45.5963.9535.420.121
SALAD [52]64.0775.9055.420.098
PersFormer [6]74.3380.3069.180.074
Anchor3DLane [16]74.8780.8569.710.060
LATR [31]80.5986.1275.730.052
PVALane [53]76.3582.8170.830.059
GroupLane [22]80.7382.5678.900.053
SparseLaneSTP (ours)82.7586.4779.330.048
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.709, + 0.462, + 0.723 + ], + "angle": 0, + "content": "Table 6. Quantitative comparison on ONCE-3DLanes [52]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.252, + 0.768 + ], + "angle": 0, + "content": "5.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.838 + ], + "angle": 0, + "content": "We perform a comprehensive analysis validating the effectiveness of our design choices in SparseLaneSTP. For these experiments, we used a smaller model consisting of two decoder layers evaluated on OpenLane." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Table 1 summarizes the impact of our contributions. Our CR spline based lane representation improves the F1-Score by over 1 percentage point compared to the discrete one (baseline). The most significant gain in F1-Score is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.907, + 0.525 + ], + "angle": 0, + "content": "achieved by the novel STA component, which encourages focus on relevant relations based on lane structure and, more importantly, on related past observations by integrating temporal cross-attention. Eventually, regularization enhances robustness and consistency, leading to improved generalization, as evidenced by the additional F1-Score gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.835 + ], + "angle": 0, + "content": "As shown in Table 2, global self-attention apparently fails to capture important relations of lane structures. The combination of our proposed SLA and PNA demonstrates the benefits of enhancing the focus on intra- and interlane relations by an F1-Score increase \\(+0.9\\%\\). Integrating knowledge about past observations via our suggested TCA yields the greatest improvement of additional \\(+1.2\\%\\). As indicated by Table 3, \\(T = 3\\) frames in the memory produces the best results. Using too few frames provides insufficient temporal context, whereas too many frames apparently introduce redundancy and distract the model from relevant information. We further qualitatively demonstrate the beneficial impact of our primary contribution on temporal fusion in Fig. 6. The non-temporal model fails to retain the detection from frame \\(t = 1\\) in subsequent frames as lane visibility deteriorates. In contrast, the temporally-aware model maintains the detection despite partial occlusions, demonstrating improved robustness. Regularization further preserves temporal consistency mitigating the gradual disappearing and drifting in the non-temporal model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Finally, Fig. 7 reveals efficiency benefits of our approach. SparseLaneSTP with six decoder layers achieves 11.0 FPS, comparable to LATR at 12.1 FPS. While our proposed temporal integration (memory queue and STA) in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29105" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.206, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.228, + 0.205, + 0.24 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.089, + 0.298, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.089, + 0.408, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.228, + 0.407, + 0.24 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.091, + 0.5, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.609, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.228, + 0.61, + 0.24 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.091, + 0.702, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.089, + 0.812, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.091, + 0.904, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.228, + 0.812, + 0.24 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.253, + 0.88, + 0.267 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of LATR and SparseLaneSTP on OpenLane with ground truth. Main differences are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.355 + ], + "angle": 0, + "content": "produces a modest \\(9\\%\\) overhead, it also enables a smaller model with only two layers to reach \\(65.3\\%\\) F1-Score \\((+3.4\\%\\) over six-layer LATR) at a significantly higher speed of 16.5 FPS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.224, + 0.379 + ], + "angle": 0, + "content": "5.4. Main results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.506 + ], + "angle": 0, + "content": "In this section, we compare our method to other 3D lane detectors on public benchmarks as well as on our own dataset. Results on OpenLane. Table 4 presents a quantitative comparison demonstrating the superiority of our approach. SparseLaneSTP reaches the highest F1-Score, exceeding \\(66\\%\\), significantly surpassing the second-best model despite its larger backbone and even achieves the lowest geometric errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Fig. 8 illustrates the advantages of our approach qualitatively across multiple aspects. We compare our SparseLaneSTP to LATR [31], which is most related to our baseline. As shown in Fig. 8a, our continuous lane representation enables precise visibility estimation with accurate detection of start/end of markings, whereas LATR - constrained by fixed discrete anchor points - produces erroneous estimates. In Fig. 8b, our temporal fusion preserves the line despite partial occlusions, whereas non-temporal LATR fails. It also struggles to estimate faint lane markings on the far side of the junction shown in Fig. 8c, while with STA our model captures lane structures more effectively. Eventually, in Fig. 8d our model exhibits a stronger reliance on priors resulting in robust and consistent behavior, most likely due to the incorporated regularization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Results on ONCE-3DLanes. As shown in Table 6, our method achieves the highest detection scores and demonstrates accurate regression with a significantly lower CD, despite the inaccurate camera parameters provided by the dataset. Notably, even though ONCE-3DLanes has a lower frame rate, our temporal approach performs effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Results on our 3D lane dataset. Initial results on our novel dataset are presented in Table 5. Our proposed extension of the evaluation metrics provides a more comprehensive assessment of overall performance, particularly in the far-range and for occlusion detection. To facilitate a fair" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.291, + 0.743, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.39, + 0.668, + 0.398 + ], + "angle": 0, + "content": "Front-view" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.291, + 0.885, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.344, + 0.885, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.408, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Figure 9. Qualitative example of SparseLaneSTP on our data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.572 + ], + "angle": 0, + "content": "comparison, we train the baseline and other methods with specific modifications to range and capacity. Our method excels in F1-Score, error metrics and notably in the VisIoU, highlighting the superior performance of our temporal model in detecting occlusions. Fig. 9 shows a qualitative result on our dataset, showcasing accurate long-range lane estimation \\((200\\mathrm{m})\\). For more qualitative results with comparisons to the baseline we refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.634, + 0.601 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.889 + ], + "angle": 0, + "content": "In this work, we presented a novel 3D lane detection approach that combines sparse lane transformers with spatial and temporal priors. We proposed a continuous lane representation, tailored for the sparse query design, thereby overcoming limitations of discrete representations. Additionally, we developed a novel attention layer that effectively captures intra- and inter-lane relations while integrating temporal information through propagated lane queries. Furthermore, we introduced a dedicated regularization scheme enhancing temporal consistency. Our experiments validate the effective integration of temporal information and spatial priors, as well as the benefits of our lane representation, leading to state-of-the-art performance across all evaluated 3D lane datasets. Moreover, we presented a new 3D lane dataset, constructed using a simple yet effective auto-labeling pipeline, providing geometrically accurate labels that span long distances and explicitly annotate occlusions, fostering further advancements in the field. In future, we will explore an extension to 3D lane tracking to leverage the full potential of temporal information." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29106" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Lucas R Agostinho, Nuno M Ricardo, Maria I Pereira, Antoine Hiolle, and Andry M Pinto. A practical survey on visual odometry for autonomous driving in challenging scenarios and conditions. IEEE Access, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.237 + ], + "angle": 0, + "content": "[2] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.239, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.294, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[4] Edwin Catmull and Raphael Rom. A class of local interpolating splines. Computer aided geometric design, 1972. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.483, + 0.4 + ], + "angle": 0, + "content": "[5] Yifan Chang, Junjie Huang, Xiaofeng Wang, Yun Ye, Zhujin Liang, Yi Shan, Dalong Du, and Xingang Wang. Rethinking lanes and points in complex scenarios for monocular 3d lane detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[7] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[8] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[9] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.683 + ], + "angle": 0, + "content": "[10] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.685, + 0.483, + 0.751 + ], + "angle": 0, + "content": "[11] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.753, + 0.483, + 0.806 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.807, + 0.483, + 0.86 + ], + "angle": 0, + "content": "[13] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.862, + 0.483, + 0.888 + ], + "angle": 0, + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. CoRR," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.698, + 0.105 + ], + "angle": 0, + "content": "abs/2203.17054, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.107, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR, abs/2112.11790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.147, + 0.905, + 0.214 + ], + "angle": 0, + "content": "[16] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.215, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[17] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.297, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[18] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.337, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[19] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[20] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[21] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.527, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[22] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Groupplane: End-to-end 3d lane detection with channel-wise grouping. IEEE Robotics and Automation Letters (RA-L), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.663, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In Proc. of the International Conf. on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.785, + 0.905, + 0.837 + ], + "angle": 0, + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. CoRR, abs/2211.10581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.839, + 0.905, + 0.878 + ], + "angle": 0, + "content": "[27] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. CoRR, abs/2305.14018, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.879, + 0.905, + 0.892 + ], + "angle": 0, + "content": "[28] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Ze" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29107" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "jian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.187 + ], + "angle": 0, + "content": "[29] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.187, + 0.482, + 0.255 + ], + "angle": 0, + "content": "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petr2: A unified framework for 3d perception from multi-camera images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.255, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[32] Zhipeng Luo, Changqing Zhou, Gongjie Zhang, and Shijian Lu. Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[33] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, and Hang Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[34] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.594, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[38] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[39] Maximilian Pittner, Joel Janai, and Alexandru P Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[40] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.811, + 0.482, + 0.864 + ], + "angle": 0, + "content": "[41] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.864, + 0.482, + 0.892 + ], + "angle": 0, + "content": "[42] Davide Scaramuzza, Friedrich Fraundorfer, and Roland Siegwart. Real-time monocular visual odometry for on-road" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "vehicles with 1-point ransac. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.12, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[43] Frank Steinbrücker, Jürgen Sturm, and Daniel Cremers. Real-time visual odometry from dense rgb-d images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV) Workshops, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.174, + 0.905, + 0.228 + ], + "angle": 0, + "content": "[44] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.905, + 0.309 + ], + "angle": 0, + "content": "[45] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[47] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[48] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[49] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[51] Yin Wang, Qiuyi Guo, Peiwen Lin, Guangliang Cheng, and Jian Wu. Spatio-temporal fusion-based monocular 3d lane detection. In Proc. of the British Machine Vision Conf. (BMVC), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[52] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.771, + 0.905, + 0.84 + ], + "angle": 0, + "content": "[53] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.84, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Proc. of the International Conf. on Learning Representations (ICLR), 2021. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29108" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[55] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29109" + } + ] +] \ No newline at end of file diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_origin.pdf b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a29350bebc431bc60a7dd02ae5f4743c75c42f4 --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/1eda17b7-8ae2-4be3-89f2-07bc44a3e4eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9039fbf5e92feccac61d18eff0f2cf39c43a3b47140f5e465a9c1f46e98275fb +size 20296987 diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/full.md b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..29657355a37c241fada22565e3c700cf6d63ebba --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/full.md @@ -0,0 +1,387 @@ +# SparseLaneSTP: Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection + +Maximilian Pittner1,2, Joel Janai1, Mario Faigle*1,3, Alexandru Paul Condurache1,2 + +1Bosch Mobility Solutions, Robert Bosch GmbH + +$^{2}$ Institute of Neuro- and Bioinformatics, University of Lübeck + +$^{3}$ Institute for Signal Processing and System Theory, University of Stuttgart + +{Maximilian.Pittner, Joel.Janai, Mario.Faigle, AlexandruPaul.Condurache}@de.bosch.com + +# Abstract + +3D lane detection has emerged as a critical challenge in autonomous driving, encompassing identification and localization of lane markings and the 3D road surface. Conventional 3D methods detect lanes from dense birds-eyeviewed (BEV) features, though erroneous transformations often result in a poor feature representation misaligned with the true 3D road surface. While recent sparse lane detectors have surpassed dense BEV approaches, they completely disregard valuable lane-specific priors. Furthermore, existing methods fail to utilize historic lane observations, which yield the potential to resolve ambiguities in situations of poor visibility. To address these challenges, we present SparseLaneSTP, a novel method that integrates both geometric properties of the lane structure and temporal information into a sparse lane transformer. It introduces a new lane-specific spatio-temporal attention mechanism, a continuous lane representation tailored for sparse architectures as well as temporal regularization. + +Identifying weaknesses of existing 3D lane datasets, we also introduce a precise and consistent 3D lane dataset using a simple yet effective auto-labeling strategy. Our experimental section proves the benefits of our contributions and demonstrates state-of-the-art performance across all detection and error metrics on existing 3D lane detection benchmarks as well as on our novel dataset. + +# 1. Introduction + +Accurate and robust 3D lane detection forms a pivotal task in autonomous driving addressing safe and reliable identification and localization of the road surface and lane markings. While significant progress has been made in 2D lane detection [10, 13, 17-19, 21, 35, 36, 40, 41, 46, 48, 55], the prediction output is only provided in image space lacking + +depth information, which is crucial for autonomous navigation in the 3D world. In contrast, 3D lane detection jointly estimates lane markings and the road surface, directly producing 3D lanes in a vehicle-centered coordinate system. Conventional methods operate on dense feature maps that are transformed from the front-viewed (FV) to the birds-eye-viewed (BEV) perspective using inverse perspective mapping (IPM) [9, 11, 38] or learned mappings [6, 39, 49]. Although this step yields a suitable intermediate representation to capture the road, potential errors in the transformation cause misalignment between the resulting BEV and the true road surface that can hardly be compensated in the subsequent lane estimation step. + +As opposed to this two-stage approach, sparse detection methods have gained popularity in object detection [34] and have recently been adapted to 3D lane detection [2, 31]. These methods model 3D lane points as queries and associate them with uncorrupted FV image features - avoiding the necessity of error-prone BEV representations. Queries composed of learned context embedding vectors are integrated into transformer architectures [3, 54] leveraging their ability to learn global context. However, current approaches completely ignore valuable well-known priors about lane and road geometry. Spatial priors have already been utilized in [39]. Based on lane continuity and other properties like line parallelism, a continuous lane representation was introduced together with an efficient regularization. Despite notable advances brought by this approach, the architecture is based on dense BEV representations and the proposed priors have not been successfully customized for sparse architectures. Another class of prior knowledge is grounded in the history of observations. Among various temporal fusion paradigms, object-centric query propagation [50] has recently been proposed for 3D object detection in sparse architectures. In lane detection, temporal information has the potential to resolve ambiguities, particularly in scenarios with limited visibility or occlusions. Though, it has not + +been applied effectively so far. + +We therefore present a new method SparseLaneSTP that combines spatial and temporal priors with a sparse 3D lane transformer architecture. Our model features a transformertailored continuous lane representation providing smooth curves directly. This, combined with robust regularization using additional spatial and temporal objectives, enables the model to learn regression and visibility estimation more accurately. Additionally, spatial and temporal knowledge is integrated into the attention mechanism in a novel way. Instead of standard global self-attention, we design the layers to focus on relevant relations based on lane-structure. Finally, lane queries are temporally propagated and incorporated into the attention to leverage valuable historic keys. + +On top of that, we provide a novel 3D lane dataset using a simple yet efficient auto-labeling pipeline. Existing datasets often yield inaccurate labels containing noise and outliers, particularly in the far-range, since they rely on Li-DAR to recover 3D information. Our auto-labeling instead uses temporal aggregation along video-sequences resulting in consistent and accurate 3D lane labels up to 250 meters, making a valuable contribution to the research community. + +Our contributions can be summarized as follows: + +- We present a novel 3D lane detector that integrates spatial and temporal knowledge into a sparse transformer. +- We propose a new attention mechanism that focuses on learning lane-structure and successfully leverages temporal keys from past lane observations. +- We formulate a continuous lane representation tailored for sparse architectures and a temporal regularization. +- Our method achieves state-of-the-art performance on all 3D lane benchmarks. +- We provide a new dataset for 3D lane detection providing accurate and consistent long-range labels. + +# 2. Related work + +Dense BEV vs. sparse query-based detection. A common strategy in 3D lane detection involves operating on an intermediate dense BEV representation generated from the FV. Early methods, such as 3D-LaneNet [9], GenLaneNet [11] and 3D-SpLineNet [38], applied IPM to project FV features onto a flat ground plane. However, the assumption of a perfectly flat road is often violated, leading to degraded BEV features. Alternatively, methods from related fields like 3D object detection [14, 15, 23] and BEV segmentation [37] proposed to learn the BEV representation. These concepts were applied to 3D lane detection with PersFormer [6] leveraging deformable cross-attention (DCA) [54], whereas BEV-LaneDet [49] relies on simple multi-layer perceptrons (MLPs). LaneCPP [39] incorporates surface priors to learn 3D features using depth classification inspired by Lift-Splat-Shoot (LSS) [37]. Despite these advancements, all methods remain constrained by an + +intermediate error-prone BEV representation. + +In contrast, sparse methods based on DETR [3, 54] avoid BEV representation by modeling 3D objects as queries that exchange information via attention mechanisms. In 3D object detection, 3D position-aware object queries were introduced that are directly associated with the perspective view through DCA [26, 34]. Similarly, in 3D lane detection, CurveFormer [2] adapts the sparse query design to lanes using polynomial-based curve queries, though these lack the flexibility to capture diverse lane structures in real-world scenarios. LATR [31] refines this approach by representing lanes as individual points but fails to incorporate crucial priors in both the attention mechanism and lane representation. + +Priors in lane detection. An indisputable property about lanes lies in their smoothness and continuity. Several methods have proposed continuous lane representations in both 2D [8, 29, 47] and 3D [2, 28, 38, 39] lane detection. In contrast to discrete approaches [6, 9, 11, 16, 49], continuous methods directly provide smooth curves, requiring almost no post-processing and exploit the entire available dense ground truth during training [38, 39]. While CLGO [28] and CurveFormer [2] use simple polynomials, 3D-SpLineNet [38] and LaneCPP [39] use a B-Spline [7] representation, which offers local control over curve segments enabling it to model complex shapes. Additionally, lane structures exhibit inherent properties, such as parallelism between lanes. E.g. SGNet [44] introduces a penalty term to enforce a fixed lane width, but assumes a flat ground plane. GP [20] employs a parallelism loss to maintain a constant local distance between neighboring points, though its effectiveness is influenced by the number of anchor points used. In contrast, LaneCPP [39] demonstrates a more elegant way to encourage parallelism as well as other spatial properties like surface smoothness and curvature restriction using regularization losses. Motivated by LaneCPP, we propose a more suitable Catmull-Rom spline [4] representation tailored for the transformer's attention with sparse queries. We further present more sophisticated loss formulations and integrate prior knowledge into the transformer attention mechanism enhancing the focus on relevant relations. While a concurrent work also uses priors in attention [5], our approach focuses on explicit modeling of structural lane relations and even effectively incorporates temporal knowledge in lane detection for the first time. + +Temporal modeling. Another source of prior knowledge is grounded in the history of observations. For video-based 3D object detection various approaches exist to model temporal interactions. BEV-based methods [23] transform past BEV features according to the ego-motion and fuse them with the current representation. More advanced perspective view (PV) methods [26, 30, 32] instead follow the sparse query design but require storing dense historic PV feature maps to interact with object queries. Recently, + +![](images/58a284d290683b79f2026809a8d222e2fc12b28edc9cc3ac10e419e411115b1f.jpg) +Figure 1. Overview of our method SparseLaneSTP. Sparse lane queries are processed by a transformer integrating temporal knowledge and spatial lane structure priors in a novel spatio-temporal attention. Based on these priors, we formulate spatial and temporally consistent regularization. Finally, our network predicts control points defining our new continuous 3D lane representation. + +object-centric temporal modeling [27, 50] has been proposed, which only stores sparse historic object queries and propagates them according to ego- and object-motion in 3D space. In particular, Sparse4Dv2 [27] and StreamPETR [50] leverage the sparse query design and model object queries as hidden states of a transformer, where the relation between current and historic queries is learned via attention. + +In 3D lane detection, only few approaches exist that use temporal modeling. STLane3D [51] follows a BEV-based method, PETRv2 [30] fuses historic PV features together with propagated 3D positional embeddings. On the other hand, Anchor3DLane-T [16] leverages temporal information by associating 3D anchors from the current frame with features from previous frames. In contrast, our sparse query-based architecture is more suitable for the effective memory-efficient object-centric paradigm. Consequently, we propose a temporal attention mechanism inspired by [27, 50], which exploits lane-specific properties by reinforcing most relevant query interactions. The static nature of lane markings also allows us to formulate regularization to encourage temporally consistent perception behavior. + +# 3. Methodology + +In the following, we describe our 3D lane detection method illustrated in Fig. 1. Given an image $\mathbf{I}$ , the model predicts a set of $N$ lanes as continuous curves in 3D space, parameterized by our novel tailored spline representation with $M$ control points $\mathbf{P}$ (see Section 3.1) and lane categories $\mathbf{C}$ . The core component of our architecture is the transformer decoder, where each control point, inspired by sparse query design, is paired with a learned context embedding and is iteratively processed through $L$ decoder layers. + +![](images/a8a5ac857e3e0ced946926e3dd3ec2b1c5b5dfd232cea74fe234a70ddcef4a7c.jpg) +(a) Front-view + +![](images/8a683ecd4aee8f9635024464c782e4c9ad2a3abddbf8cc2e707a62a96e4c54e2.jpg) +(b) BEV +Figure 2. CR spline and B-Spline control points in comparison, both fit to the ground truth. B-Spline control points do not align with the curve and are therefore not suited the sparse query design, whereas CR control points exactly match curve geometry. + +In contrast to previous work [2, 24, 31] that simply adapt the DETR [3, 54] framework, our transformer architecture introduces a novel spatio-temporal attention layer to capture spatial intra- and inter-lane relations, while a temporal cross-attention mechanism integrates past observations via propagated lane queries from a memory queue. Finally, besides standard detection losses, our model is cautiously regularized by spatial and temporal objectives to enhance robust and consistent detection behavior. + +# 3.1. 3D lane representation + +For the 3D lane representation, we take inspiration from prior work [38, 39] that demonstrated significant benefits of continuous curves. In these frameworks, the model output serves as control points of B-Spline curves that do not necessarily align with the curve geometry. However, using the sparse query design, the 3D position of control points serve as the model's internal state and should therefore correspond to the exact 3D position of lane points. Thus, spline + +control points should directly lie on the curve, which is not the case for B-Splines (see Fig. 2). Therefore, we choose Catmull-Rom (CR) splines [4], a class of piece-wise defined smooth third-order polynomial splines, for which the curve inherently passes through its control points. + +We adapt this spline-model to define a curve representation $\mathbf{f}_i(s) \in \mathbb{R}^4$ representing the entire $i^{\mathrm{th}}$ line proposal parameterized by $M$ control points with 3 spatial $(x_{ij}, y_{ij}, z_{ij})$ and 1 visibility $(v_{ij})$ dimension as + +$$ +\mathbf {f} _ {i} (s) = \left[ \begin{array}{l l l l} s ^ {3} & s ^ {2} & s & 1 \end{array} \right] \cdot \mathbf {M} _ {\mathrm {C R}} \cdot \mathbf {P} _ {i} \quad \text {w i t h} \tag {1} +$$ + +$$ +\mathbf {P} _ {i} = \left[ \mathbf {P} _ {\mathrm {3 D}, i} \mid \mathbf {P} _ {v, i} \right] = \left[ \begin{array}{c c c c} x _ {i 1} & y _ {i 1} & z _ {i 1} & v _ {i 1} \\ & \dots & & \\ x _ {i j} & y _ {i j} & z _ {i j} & v _ {i j} \\ & \dots & & \\ x _ {i M} & y _ {i M} & z _ {i M} & v _ {i M} \end{array} \right], \tag {2} +$$ + +with curve argument $s\in [0,1]$ $\mathbf{P}_i\in \mathbb{R}^{M\times 4}$ denotes the control point matrix composed of $\mathbf{P}_{3\mathrm{D},i}\in \mathbb{R}^{M\times 3}$ defining the curve 3D shape and visibility $\mathbf{P}_{v,i}\in \mathbb{R}^{M\times 1}$ $[\cdot |\cdot ]$ is the concatenation operator and $\mathbf{M}_{\mathrm{CR}}$ the CR coefficient matrix. Note that we can eventually express the curve computation as a simple matrix multiplication between the pre-computed spline arguments and the entire predicted control points $\mathbf{P}$ + +# 3.2. Overall architecture + +Backbone and query initialization. The RGB input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ is first processed by a CNN backbone (e.g. ResNet [12]) to extract image features of $\mathbf{F} \in \mathbb{R}^{H_{\mathbf{F}} \times W_{\mathbf{F}} \times C}$ , where $C$ is the channel size. Similar to [31], we then feed $\mathbf{F}$ through a lane instance segmentation branch to obtain initial query embedding vectors ${}^0\mathbf{Q} \in \mathbb{R}^{N \times M \times C}$ , where $N$ denotes the number of line proposals and $M$ the number of spline control points. An MLP then predicts initial control points ${}^0\mathbf{P} \in \mathbb{R}^{N \times M \times 4}$ from ${}^0\mathbf{Q}$ . + +Transformer decoder. In each transformer decoder layer $l$ , queries are processed by two attention layers + +$$ +{ } ^ { l } \mathbf { Q } _ { \mathrm { S T A } } = \operatorname { S T A } \left( { } ^ { l - 1 } \mathbf { Q } , { } ^ { l - 1 } \mathbf { P } , \mathbf { Q } _ { \text { M e m } } , \mathbf { P } _ { \text { M e m } } \right) \tag {3} +$$ + +$$ +{ } ^ { l } \mathbf { Q } _ { \mathrm { D C A } } = \mathrm { D C A } \big ( { } ^ { l } \mathbf { Q } _ { \mathrm { S T A } } , \mathbf { F } , { } ^ { l - 1 } \mathbf { P } _ { 2 \mathrm { D } } \big ) , \tag {4} +$$ + +$\mathrm{STA}(\cdot)$ denotes our proposed spatio-temporal attention, which is explained in Section 3.3, and $\mathrm{DCA}(\cdot)$ the standard deformable cross-attention presented in [54]. $\mathbf{P}_{2\mathrm{D}}$ are the 2D spline control points in image coordinates resulting from re-projecting $\mathbf{P}_{3\mathrm{D}}$ . Finally, the query embeddings are fed through a feed-forward network $^l\mathbf{Q} = \mathrm{FFN}(^l\mathbf{Q}_{\mathrm{DCA}})$ composed of an MLP and layer normalization. + +Prediction layer. After each transformer attention layer, the emerging query embeddings ${}^{l}\mathbf{Q}$ are fed through a prediction layer that outputs the spline control points ${}^{l}\mathbf{P}$ as well as classification probabilities $\mathbf{C} \in \mathbb{R}^{N \times (K + 1)}$ for $K$ lane categories and one background class per line. Particularly, + +![](images/d36f3edc002eec1903da03a9d36119061be4262d0382ab5a775fe04e10924fc2.jpg) +Figure 3. Overview of our spatio-temporal attention module. + +for each layer $l$ and query ${}^{l}\mathbf{Q}_{ij}$ three MLPs with shared weights among all layers are used to predict the $x-$ , $z-$ and $v$ -component of the control point matrix + +$$ +{ } ^ { l } \mathbf { P } _ { x , i j } = \sigma ( \mathrm { M L P } _ { x } \left( { } ^ { l } \mathbf { Q } _ { i j } \right) ) \cdot \left( x _ { e } - x _ { s } \right) + x _ { s } \tag {5} +$$ + +$$ +{ } ^ { l } \mathbf { P } _ { z , i j } = \sigma ( \mathrm { M L P } _ { z } ( { } ^ { l } \mathbf { Q } _ { i j } ) ) \cdot ( z _ { e } - z _ { s } ) + z _ { s } \tag {6} +$$ + +$$ +{ } ^ { l } \mathbf { P } _ { v , i j } = \sigma ( \mathrm { M L P } _ { v } ( { } ^ { l } \mathbf { Q } _ { i j } ) ) , ( 7 ) +$$ + +where the sigmoid $\sigma (\cdot)$ normalizes the MLP output to [0, 1]. The start-/end- of $x$ - and $z$ -range $x_{s / e}$ , $z_{s / e}$ scales up the normalized output to the desired range. The longitudinal $y$ -component ${}^l\mathbf{P}_{y,ij}$ is pre-defined and uniformly distributed between $y_{s}$ and $y_{e}$ to avoid over-parameterization and potential overfitting as argued in [39]. For classification, all queries belonging to the same line proposal are averaged and fed into an MLP followed by a softmax layer. + +Finally, the formulation of our supervised losses is inspired by [39], with L1 loss for regression of $x$ and $z$ , binary cross-entropy for visibility $v$ and focal loss [25] for classification. + +# 3.3. Integrating spatial and temporal knowledge + +Memory queue. To facilitate attention to past observations, queries are stored within a memory queue. Following a similar approach to [50], this queue is recursively updated after each prediction step and retains queries from the past $T$ frames. Therefore, embedding vectors and their associated control points are recursively pushed into the queue using a first-in, first-out (FIFO) strategy, which yields + +$$ +\mathbf {Q} _ {\text {M e m}} = \left[ \mathbf {Q} _ {: N _ {\text {M e m}}} ^ {(t - 1)} \mid \mathbf {Q} _ {: N _ {\text {M e m}}} ^ {(t - 2)} \mid \dots \mid \mathbf {Q} _ {: N _ {\text {M e m}}} ^ {(t - T)} \right] \quad \text {a n d} \tag {8} +$$ + +$$ +\mathbf {P} _ {\mathrm {M e m}} = \left[ \mathbf {P} _ {: N _ {\mathrm {M e m}}} ^ {(t - 1)} \mid \mathbf {P} _ {: N _ {\mathrm {M e m}}} ^ {(t - 2)} \mid \dots \mid \mathbf {P} _ {: N _ {\mathrm {M e m}}} ^ {(t - T)} \right], \tag {9} +$$ + +where $\mathbf{Q}_{:N_{\mathrm{Mem}}}^{(t)}\in \mathbb{R}^{(N_{\mathrm{Mem}}\times M\times C)}$ denotes the embedding vectors from the $N_{\mathrm{Mem}}$ most confident predictions of frame $t$ and $\mathbf{Q}_{\mathrm{Mem}}\in \mathbb{R}^{((T\cdot N_{\mathrm{Mem}})\times M\times C)}$ the full memory queue. + +Temporal propagation. Since the model is supposed to learn spatio-temporal context, it must incorporate the positional information of 3D control points from past frames. + +However, due to the vehicle's ego-motion, lane positions shift relative to the vehicle. Therefore, the 3D coordinates of past lane control points stored in the memory queue must be transformed according to the ego-pose $\mathbf{E}^{(t)}$ of frame $t$ , which is represented as a homogeneous transformation matrix. Consequently, a control point $\mathbf{P}_{ij}^{(t - k)\rightarrow (t)}$ propagated from frame $t - k$ to present frame $t$ is given as + +$$ +\mathbf {P} _ {i j} ^ {(t - k) \rightarrow (t)} = \left[ \mathbf {E} _ {i n v} ^ {(t)} \cdot \mathbf {E} ^ {(t - k)} \cdot \mathbf {P} _ {\mathrm {3 D}, i j} ^ {(t - k)} \mid \mathbf {P} _ {v, i j} ^ {(t - k)} \right], \tag {10} +$$ + +where $\mathbf{E}_{inv}$ denotes the inverted matrix $\mathbf{E}$ . Obviously, only the geometry component is affected by the propagation, whereas the visibility component remains unchanged. To incorporate spatio-temporal information from propagated control points into queries, we apply a positional encoding (PE) based on 3D geometry and visibility (see Fig. 3). This process yields position-informed embedding vectors + +$$ +\tilde {\mathbf {Q}} _ {i j} ^ {(t - k) \rightarrow (t)} = \mathbf {Q} _ {i j} ^ {(t - k) \rightarrow (t)} + \mathrm {P E} \left(\mathbf {P} _ {i j} ^ {(t - k) \rightarrow (t)}\right), \tag {11} +$$ + +which are processed by the spatio-temporal attention layer. Spatio-temporal attention. Most methods [2, 31] rely solely on global self-attention, which contains a majority of redundant lane query interactions that can distract the model from learning relevant context. To address this, we leverage spatial and temporal priors, distinguishing three key relation types: intra-lane (between adjacent points on the same lane), inter-lane (between parallel lanes), and historic (between current and past queries in the memory queue). + +We therefore introduce same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA). Applying specific masking, SLA restricts interactions to queries $\tilde{\mathbf{Q}}_{\mathrm{SLA}}$ within the same line, PNA facilitates interactions with queries $\tilde{\mathbf{Q}}_{\mathrm{PNA}}$ from neighboring lines, and TCA enables interactions with closest past queries $\tilde{\mathbf{Q}}_{\mathrm{TCA}}$ in the memory queue (see Fig. 3). Such restriction of query interactions reduces redundancy, encourages the model to focus on critical lane-specific spatial context and integrates valuable temporal information from past observations. + +Regularization. Based on spatial and temporal priors we can formulate regularization techniques that promote robust and consistent detection behavior. For the spatial regularization $\mathcal{L}_{\text {spatial }}$ we adopt the scheme proposed in [39] encouraging lane parallelism, surface smoothness and suppressing excessive curvature. Besides, temporal consistency can be achieved by optimizing a smoothness loss + +$$ +\mathcal {L} _ {t e m p} = \frac {1}{N} \sum_ {i} ^ {N} \int_ {s} \overline {{\mathbf {f}}} _ {v, i} ^ {(t)} (s) \cdot \left| \left| \mathbf {f} _ {\mathrm {3 D}, i} (s) - \overline {{\mathbf {f}}} _ {\mathrm {3 D}, i} ^ {(t)} (s) \right| \right| _ {1}, \tag {12} +$$ + +where $\overline{\mathbf{f}}_{3\mathrm{D},i}^{(t)}(s)$ or $\overline{\mathbf{f}}_{v,i}^{(t)}(s)$ is an exponentially moving average of the past predictions + +$$ +\overline {{\mathbf {f}}} _ {i} ^ {(t)} = \alpha \cdot \overline {{\mathbf {f}}} _ {i} ^ {(t)} + (1 - \alpha) \cdot \overline {{\mathbf {f}}} _ {i} ^ {(t - 1)} \tag {13} +$$ + +with smoothing factor $\alpha \in [0,1]$ . + +![](images/fda28af3d24d1f53837be1f37fc3ab7a591e3c4296b656c452cc522035b8533a.jpg) +Figure 4. Our auto-labeling strategy. + +![](images/ecde0324ebefa099ca401ef2b608f2e31a5ccbea365a1f208f0eebf6f577adf9.jpg) +(a) OpenLane +Figure 5. Examples from OpenLane compared to ours. + +![](images/b69ab776c619896e7369f6884e623b91a59630cd36aa4981715490cff32ede5d.jpg) +(b) Our 3D lane dataset + +# 4. Auto-labeling and our 3D lane dataset + +In this section, we describe our auto-labeling pipeline and provide details about our new 3D lane dataset. + +# 4.1. Auto-labeling strategy + +While 3D lane datasets like OpenLane [6] and ONCE-3DLanes [52] have contributed significantly to the field, they possess specific shortcomings that our auto-labeling strategy addresses. As illustrated in Fig. 5a the LiDAR-based annotation approach, which recovers 3D information of 2D lane labels, introduces noise and outliers, particularly in the far range due to sparse scan points, while also posing calibration and synchronization challenges. To overcome these limitations, we propose an automated labeling approach (see Fig. 4) that leverages confident near-range lane point estimates from a 2D detector, which are temporally accumulated in 3D space along an accurate ego-trajectory. More precisely, common visual odometry methods [1, 42, 43] are applied to recover the vehicle ego-trajectory from the input video sequence. Our road surface generator then fits a spline surface to the vehicle's trajectory and orientations. On the other hand, lane pseudo-labels in image space are obtained from a state-of-the-art 2D lane detector [21, 46], providing particularly reliable detection results in the near-range. Lane and surface information are then combined in 3D by projecting near-range detection results to the road surface and accumulating line points frame + +![](images/45007fe7c541efa29d5b0b60c92d84b92fdfdd84020bbd290e108e3adf4c121f.jpg) + +![](images/966afd6af52639b3b08bf7c93796f58a06b8177ddbae6d2dec1674f6a72ebf3f.jpg) +(a) $t = 1$ + +![](images/f09b95709d6a03753724bf7ca9e067d9c2bd2913f359c8bf0ba7d9ce109a5e9f.jpg) + +![](images/b8cae87e5cbccb743668b6d9968a55f484020858cff834198213833a2bb68383.jpg) +(b) $t = 2$ + +![](images/13099c230a58274e3bdfe6fb7038a3cc0cbec47c5a9c24c3127a5ba81df1aaf5.jpg) + +![](images/2bcf7f0b866de407cb7f9a0ddc78bda559a27cd0223ebaf863f8b59a5e2a2792.jpg) +(c) $t = 3$ +Figure 6. Qualitative comparison of two models, one without temporal and one with temporal for three consecutive frames. + +by frame along the ego-trajectory of the sequence. Finally, we apply an accurate semantic segmentation network to automatically assign occlusion labels to each lane point. + +# 4.2. Dataset properties + +Using the approach described in 4.1, we are able to automatically generate temporally consistent, high-quality 3D lane labels in a range up to $250\mathrm{m}$ (see Fig. 5b). Besides 3D lane geometry, we provide camera intrinsics and extrinsics, lane categories, lane visibility information and consistent track IDs. We further supply relational information about lanes (like line parallelism), local 3D ego-trajectory and global pose information for each frame. Important to mention is also that in contrast to former datasets, our visibility information contains occlusions caused by other traffic participants, enabling models to consistently learn lane structure independent of interruptions and lane occlusions. We refer to the supplementary material for additional information. + +# 5. Experiments + +In this section, we first outline our experimental setup, followed by a comprehensive analysis of our approach on two public and our own novel 3D lane dataset. + +# 5.1. Datasets and evaluation metrics + +OpenLane [6] based on the Waymo Open dataset [45] remains the most prominent real-world 3D lane dataset to date. It is comprised of 200K images from 1000 sequences recorded in several cities in the USA. + +ONCE-3DLanes [52] is another real-world dataset derived from the ONCE dataset [33]. It consists of 211K images collected from multiple cities in China. + +Our 3D lane dataset, which was introduced in Section 4, was gathered from multiple regions and countries across the globe. It consists of 511K images with a $90\%$ / $10\%$ train/test split. + +Evaluation metrics. The OpenLane evaluation scheme [6, 11] uses the euclidean distance at uniformly distributed + +
Continuous Rep.STARegularizationF1(%)↑Gain(%)
61.8(baseline)
✓62.9+1.1
✓✓65.0+2.1
✓✓✓65.3+0.3
+ +Table 1. Performance gain for different contributions on OpenLane using our novel Continuous Lane Representation, Spatio-Temporal Attention (STA) and Regularization. + +
AttentionGlobalSLAPNASLA + PNASLA + PNA + TCA
F1-Score62.962.863.563.865.0
+ +Table 2. Analysis of spatio-temporal attention. Same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA) and combinations compared to global attention. + +
Num. Frames (T)012345
F1(%)↑63.864.164.665.064.563.9
+ +Table 3. Analysis of number of frames in the memory for STA. + +points and a range-IoU to compute the F1-Score. Moreover, x- and z-errors in near- $(0 - 40\mathrm{m})$ and far-range $(40 - 100\mathrm{m})$ evaluate the geometric accuracy. ONCE-3DLanes instead uses the uni-lateral chamfer distance (CD) and computes Precision (P), Recall (R) and F1-Score based on it. + +For our dataset, we extend the well-established OpenLane evaluation framework [6, 11]. Due to the long-range labels, we add x-/z-error intervals for the ranges $100 - 150\mathrm{m}$ and $150 - 200\mathrm{m}$ . Since our dataset offers visibility labels for occlusions, we also add a visibility Intersection-over-Union metric (Vis-IoU). This score helps to gain understanding about the model's capability of distinguishing visible lane segments from parts occluded by other traffic participants and provides a measure for start-end range detection. + +# 5.2. Implementation details + +We use an image input size $720 \times 960$ and ResNet-50 [12] as our standard backbone. The query generator is trained on the auxiliary task of lane instance segmentation similar to [31]. The transformer decoder applied in the large model used for the evaluation in Section 5.4 consists of $L = 6$ layers, whereas the lighter model used for ablation studies in Sec. 5.3 has 2 layers. The lane representation uses $M = 20$ control points on OpenLane and ONCE-3DLanes and 30 control points on our dataset due to the larger range. The number of lane proposals is set to $N = 40$ for OpenLane and ours and 10 for ONCE-3DLanes, which is sufficient for the small number of occurring lanes per frame. Regarding temporal fusion, our best model uses propagated queries from $T = 3$ historic frames. + +
MethodBackboneResolutionF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [9]VGG-16360 x 48044.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [11]ERFNet360 x 48032.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [6]EfficientNet-B7360 x 48050.50.4850.5530.3640.43142.455.648.646.640.050.7
CurveFormer [2]EfficientNet-B7360 x 48050.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [49]ResNet-34360 x 48058.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane-T [16]ResNet-18360 x 48054.30.2750.3100.1050.13547.258.052.748.745.851.7
PETRv2 [30]VoV-99360 x 48061.20.4000.5730.2650.413------
LATR [31]ResNet-50720 x 96061.90.2190.2590.0750.10455.268.257.155.452.361.5
LaneCPP [39]EfficientNet-B7360 x 48060.30.2640.3100.0770.11753.664.456.754.952.058.7
PVALane [53]Swin-B720 x 96063.40.2260.2570.0930.11956.167.764.058.653.660.8
GroupLane [22]ConvNext-Basenot specified64.10.3200.4410.2330.402------
SparseLaneSTP (Ours)ResNet-50720 x 96066.10.2030.2400.0660.09257.373.060.158.358.266.5
+ +Table 4. Quantitative comparison on OpenLane [6]. Best performance and second best are highlighted. + +
MethodF1-Score(%)↑X-error / Z-error (m)↓Vis-IoU(%)↑
0 m - 40 m40 m - 100 m100 m - 150 m150 m - 200 m
PersFormer [6]59.20.191 / 0.0720.344 / 0.1450.671 / 0.2200.845 / 0.37970.1
LaneCPP [39]63.10.144 / 0.0400.298 / 0.0930.494 / 0.1880.761 / 0.29377.5
Baseline (LATR) [31]65.10.122 / 0.0380.263 / 0.0980.468 / 0.1850.700 / 0.26478.3
SparseLaneSTP (ours)68.20.109 / 0.0330.233 / 0.0920.443 / 0.1820.646 / 0.26381.4
+ +![](images/59add4d0edbc3287de048c01130ac5cceeeecf0846d50a479317c0862b580d40.jpg) +Figure 7. Efficiency analysis of SparseLaneSTP. + +Table 5. Quantitative comparison on our 3D lane dataset. Best performance and second best are highlighted. + +
MethodF1(%)↑P(%)↑R(%)↑CD(m)↓
3D-LaneNet [9]44.7361.4635.160.127
Gen-LaneNet [11]45.5963.9535.420.121
SALAD [52]64.0775.9055.420.098
PersFormer [6]74.3380.3069.180.074
Anchor3DLane [16]74.8780.8569.710.060
LATR [31]80.5986.1275.730.052
PVALane [53]76.3582.8170.830.059
GroupLane [22]80.7382.5678.900.053
SparseLaneSTP (ours)82.7586.4779.330.048
+ +Table 6. Quantitative comparison on ONCE-3DLanes [52]. + +# 5.3. Ablation studies + +We perform a comprehensive analysis validating the effectiveness of our design choices in SparseLaneSTP. For these experiments, we used a smaller model consisting of two decoder layers evaluated on OpenLane. + +Table 1 summarizes the impact of our contributions. Our CR spline based lane representation improves the F1-Score by over 1 percentage point compared to the discrete one (baseline). The most significant gain in F1-Score is + +achieved by the novel STA component, which encourages focus on relevant relations based on lane structure and, more importantly, on related past observations by integrating temporal cross-attention. Eventually, regularization enhances robustness and consistency, leading to improved generalization, as evidenced by the additional F1-Score gain. + +As shown in Table 2, global self-attention apparently fails to capture important relations of lane structures. The combination of our proposed SLA and PNA demonstrates the benefits of enhancing the focus on intra- and interlane relations by an F1-Score increase $+0.9\%$ . Integrating knowledge about past observations via our suggested TCA yields the greatest improvement of additional $+1.2\%$ . As indicated by Table 3, $T = 3$ frames in the memory produces the best results. Using too few frames provides insufficient temporal context, whereas too many frames apparently introduce redundancy and distract the model from relevant information. We further qualitatively demonstrate the beneficial impact of our primary contribution on temporal fusion in Fig. 6. The non-temporal model fails to retain the detection from frame $t = 1$ in subsequent frames as lane visibility deteriorates. In contrast, the temporally-aware model maintains the detection despite partial occlusions, demonstrating improved robustness. Regularization further preserves temporal consistency mitigating the gradual disappearing and drifting in the non-temporal model. + +Finally, Fig. 7 reveals efficiency benefits of our approach. SparseLaneSTP with six decoder layers achieves 11.0 FPS, comparable to LATR at 12.1 FPS. While our proposed temporal integration (memory queue and STA) in + +![](images/5e60b3b80c7c73572554363de529587302b91ef5c04f758e9b59908a04491d22.jpg) +(a) + +![](images/dff13bace1f0129b4c3e707079de744a9e73e04046c362ddb98a09018cff4c3e.jpg) +Figure 8. Qualitative comparison of LATR and SparseLaneSTP on OpenLane with ground truth. Main differences are highlighted. + +![](images/0a52c94e084f557f206f90e78bed8b67d07d194a86863776cf820913af80061a.jpg) +(b) + +![](images/751ce85b9ff1b7970c5939aef8016806f8ab1ebef5cf3997a2520a4733c9491e.jpg) + +![](images/78d9e261e7932fc92997426db18bc46ebaf145014f211d2754e5a7f988c6aff1.jpg) +(c) + +![](images/ef425db9dba7b94391769e8a6def0d3476c889caa25df0ea86ad485e2cc8c8e2.jpg) + +![](images/9a69dd879bf2780b9f47cbab5b2ddc8dacadb901e0e2b6879c4d34fe2f278fe1.jpg) +(d) + +![](images/b69532c9906cecac45d277e2297489022d81ce5eae3f58bd375ce8cbea91d2a7.jpg) + +produces a modest $9\%$ overhead, it also enables a smaller model with only two layers to reach $65.3\%$ F1-Score $(+3.4\%$ over six-layer LATR) at a significantly higher speed of 16.5 FPS. + +# 5.4. Main results + +In this section, we compare our method to other 3D lane detectors on public benchmarks as well as on our own dataset. Results on OpenLane. Table 4 presents a quantitative comparison demonstrating the superiority of our approach. SparseLaneSTP reaches the highest F1-Score, exceeding $66\%$ , significantly surpassing the second-best model despite its larger backbone and even achieves the lowest geometric errors. + +Fig. 8 illustrates the advantages of our approach qualitatively across multiple aspects. We compare our SparseLaneSTP to LATR [31], which is most related to our baseline. As shown in Fig. 8a, our continuous lane representation enables precise visibility estimation with accurate detection of start/end of markings, whereas LATR - constrained by fixed discrete anchor points - produces erroneous estimates. In Fig. 8b, our temporal fusion preserves the line despite partial occlusions, whereas non-temporal LATR fails. It also struggles to estimate faint lane markings on the far side of the junction shown in Fig. 8c, while with STA our model captures lane structures more effectively. Eventually, in Fig. 8d our model exhibits a stronger reliance on priors resulting in robust and consistent behavior, most likely due to the incorporated regularization. + +Results on ONCE-3DLanes. As shown in Table 6, our method achieves the highest detection scores and demonstrates accurate regression with a significantly lower CD, despite the inaccurate camera parameters provided by the dataset. Notably, even though ONCE-3DLanes has a lower frame rate, our temporal approach performs effectively. + +Results on our 3D lane dataset. Initial results on our novel dataset are presented in Table 5. Our proposed extension of the evaluation metrics provides a more comprehensive assessment of overall performance, particularly in the far-range and for occlusion detection. To facilitate a fair + +![](images/31f9551c46bb17502bde5d110bc06b96fc2a9d14e306edb0acd1e72e614aa638.jpg) +Front-view + +![](images/c773caa969f771df28e62a6d442b824794f8eeb5ef38f8c74e42a6c3b01b53c6.jpg) + +![](images/798bbab2a95bb344bb0a6b015891784b153d81168102718a5c417f74c6927f5c.jpg) +Figure 9. Qualitative example of SparseLaneSTP on our data. + +comparison, we train the baseline and other methods with specific modifications to range and capacity. Our method excels in F1-Score, error metrics and notably in the VisIoU, highlighting the superior performance of our temporal model in detecting occlusions. Fig. 9 shows a qualitative result on our dataset, showcasing accurate long-range lane estimation $(200\mathrm{m})$ . For more qualitative results with comparisons to the baseline we refer to the supplementary. + +# 6. Conclusion + +In this work, we presented a novel 3D lane detection approach that combines sparse lane transformers with spatial and temporal priors. We proposed a continuous lane representation, tailored for the sparse query design, thereby overcoming limitations of discrete representations. Additionally, we developed a novel attention layer that effectively captures intra- and inter-lane relations while integrating temporal information through propagated lane queries. Furthermore, we introduced a dedicated regularization scheme enhancing temporal consistency. Our experiments validate the effective integration of temporal information and spatial priors, as well as the benefits of our lane representation, leading to state-of-the-art performance across all evaluated 3D lane datasets. Moreover, we presented a new 3D lane dataset, constructed using a simple yet effective auto-labeling pipeline, providing geometrically accurate labels that span long distances and explicitly annotate occlusions, fostering further advancements in the field. In future, we will explore an extension to 3D lane tracking to leverage the full potential of temporal information. + +# References + +[1] Lucas R Agostinho, Nuno M Ricardo, Maria I Pereira, Antoine Hiolle, and Andry M Pinto. A practical survey on visual odometry for autonomous driving in challenging scenarios and conditions. IEEE Access, 2022. 5 +[2] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 5, 7 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3 +[4] Edwin Catmull and Raphael Rom. A class of local interpolating splines. Computer aided geometric design, 1972. 2, 4 +[5] Yifan Chang, Junjie Huang, Xiaofeng Wang, Yun Ye, Zhujin Liang, Yi Shan, Dalong Du, and Xingang Wang. Rethinking lanes and points in complex scenarios for monocular 3d lane detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2 +[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 1, 2, 5, 6, 7 +[7] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2 +[8] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[9] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 7 +[10] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1 +[11] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 6, 7 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6 +[13] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1 +[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. CoRR, + +abs/2203.17054, 2022. 2 +[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR, abs/2112.11790, 2021. 2 +[16] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7 +[17] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1 +[18] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020. +[19] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1 +[20] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[21] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 5 +[22] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Groupplane: End-to-end 3d lane detection with channel-wise grouping. IEEE Robotics and Automation Letters (RA-L), 2024. 7 +[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2 +[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In Proc. of the International Conf. on Learning Representations (ICLR), 2023. 3 +[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 4 +[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. CoRR, abs/2211.10581, 2022. 2 +[27] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. CoRR, abs/2305.14018, 2023. 3 +[28] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Ze + +jian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 2 +[29] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petr2: A unified framework for 3d perception from multi-camera images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 2, 3, 7 +[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[32] Zhipeng Luo, Changqing Zhou, Gongjie Zhang, and Shijian Lu. Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849, 2022. 2 +[33] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, and Hang Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 6 +[34] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 1, 2 +[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1 +[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1 +[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2 +[38] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3 +[39] Maximilian Pittner, Joel Janai, and Alexandru P Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 7 +[40] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1 +[41] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[42] Davide Scaramuzza, Friedrich Fraundorfer, and Roland Siegwart. Real-time monocular visual odometry for on-road + +vehicles with 1-point ransac. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2009. 5 +[43] Frank Steinbrücker, Jürgen Sturm, and Daniel Cremers. Real-time visual odometry from dense rgb-d images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV) Workshops, 2011. 5 +[44] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2 +[45] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 5 +[47] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2 +[48] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[49] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 7 +[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 3, 4 +[51] Yin Wang, Qiuyi Guo, Peiwen Lin, Guangliang Cheng, and Jian Wu. Spatio-temporal fusion-based monocular 3d lane detection. In Proc. of the British Machine Vision Conf. (BMVC), 2022. 3 +[52] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7 +[53] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2024. 7 +[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Proc. of the International Conf. on Learning Representations (ICLR), 2021. 1, 2, 3, 4 + +[55] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1 \ No newline at end of file diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/images.zip b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f5705dc1d2fb8276befe36cb57c71b3ed60600e0 --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f9242ec93907bd6b7692668a23a7cc140cd84b110275d88be207e3902bd61bd +size 699937 diff --git a/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/layout.json b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d78d841785e1c3dce03d82633a917cca70ed9557 --- /dev/null +++ b/2025/SparseLaneSTP_ Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection/layout.json @@ -0,0 +1,10200 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "text", + "content": "SparseLaneSTP: Leveraging Spatio-Temporal Priors with Sparse Transformers for 3D Lane Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 160, + 503, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 160, + 503, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 503, + 175 + ], + "type": "text", + "content": "Maximilian Pittner1,2, Joel Janai1, Mario Faigle*1,3, Alexandru Paul Condurache1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 175, + 424, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 175, + 424, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 424, + 189 + ], + "type": "text", + "content": "1Bosch Mobility Solutions, Robert Bosch GmbH" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 189, + 454, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 189, + 454, + 203 + ], + "spans": [ + { + "bbox": [ + 156, + 189, + 454, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 156, + 189, + 454, + 203 + ], + "type": "text", + "content": "Institute of Neuro- and Bioinformatics, University of Lübeck" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 203, + 485, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 203, + 485, + 217 + ], + "spans": [ + { + "bbox": [ + 125, + 203, + 485, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 125, + 203, + 485, + 217 + ], + "type": "text", + "content": "Institute for Signal Processing and System Theory, University of Stuttgart" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 219, + 533, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 219, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 74, + 219, + 533, + 232 + ], + "type": "text", + "content": "{Maximilian.Pittner, Joel.Janai, Mario.Faigle, AlexandruPaul.Condurache}@de.bosch.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 283, + 297, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 498 + ], + "type": "text", + "content": "3D lane detection has emerged as a critical challenge in autonomous driving, encompassing identification and localization of lane markings and the 3D road surface. Conventional 3D methods detect lanes from dense birds-eyeviewed (BEV) features, though erroneous transformations often result in a poor feature representation misaligned with the true 3D road surface. While recent sparse lane detectors have surpassed dense BEV approaches, they completely disregard valuable lane-specific priors. Furthermore, existing methods fail to utilize historic lane observations, which yield the potential to resolve ambiguities in situations of poor visibility. To address these challenges, we present SparseLaneSTP, a novel method that integrates both geometric properties of the lane structure and temporal information into a sparse lane transformer. It introduces a new lane-specific spatio-temporal attention mechanism, a continuous lane representation tailored for sparse architectures as well as temporal regularization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 499, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 582 + ], + "type": "text", + "content": "Identifying weaknesses of existing 3D lane datasets, we also introduce a precise and consistent 3D lane dataset using a simple yet effective auto-labeling strategy. Our experimental section proves the benefits of our contributions and demonstrates state-of-the-art performance across all detection and error metrics on existing 3D lane detection benchmarks as well as on our novel dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 604, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "Accurate and robust 3D lane detection forms a pivotal task in autonomous driving addressing safe and reliable identification and localization of the road surface and lane markings. While significant progress has been made in 2D lane detection [10, 13, 17-19, 21, 35, 36, 40, 41, 46, 48, 55], the prediction output is only provided in image space lacking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 416 + ], + "type": "text", + "content": "depth information, which is crucial for autonomous navigation in the 3D world. In contrast, 3D lane detection jointly estimates lane markings and the road surface, directly producing 3D lanes in a vehicle-centered coordinate system. Conventional methods operate on dense feature maps that are transformed from the front-viewed (FV) to the birds-eye-viewed (BEV) perspective using inverse perspective mapping (IPM) [9, 11, 38] or learned mappings [6, 39, 49]. Although this step yields a suitable intermediate representation to capture the road, potential errors in the transformation cause misalignment between the resulting BEV and the true road surface that can hardly be compensated in the subsequent lane estimation step." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "As opposed to this two-stage approach, sparse detection methods have gained popularity in object detection [34] and have recently been adapted to 3D lane detection [2, 31]. These methods model 3D lane points as queries and associate them with uncorrupted FV image features - avoiding the necessity of error-prone BEV representations. Queries composed of learned context embedding vectors are integrated into transformer architectures [3, 54] leveraging their ability to learn global context. However, current approaches completely ignore valuable well-known priors about lane and road geometry. Spatial priors have already been utilized in [39]. Based on lane continuity and other properties like line parallelism, a continuous lane representation was introduced together with an efficient regularization. Despite notable advances brought by this approach, the architecture is based on dense BEV representations and the proposed priors have not been successfully customized for sparse architectures. Another class of prior knowledge is grounded in the history of observations. Among various temporal fusion paradigms, object-centric query propagation [50] has recently been proposed for 3D object detection in sparse architectures. In lane detection, temporal information has the potential to resolve ambiguities, particularly in scenarios with limited visibility or occlusions. Though, it has not" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 234, + 712 + ], + "type": "text", + "content": "*Work done in context of master's thesis at Bosch." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29099" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 180, + 83 + ], + "type": "text", + "content": "been applied effectively so far." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "type": "text", + "content": "We therefore present a new method SparseLaneSTP that combines spatial and temporal priors with a sparse 3D lane transformer architecture. Our model features a transformertailored continuous lane representation providing smooth curves directly. This, combined with robust regularization using additional spatial and temporal objectives, enables the model to learn regression and visibility estimation more accurately. Additionally, spatial and temporal knowledge is integrated into the attention mechanism in a novel way. Instead of standard global self-attention, we design the layers to focus on relevant relations based on lane-structure. Finally, lane queries are temporally propagated and incorporated into the attention to leverage valuable historic keys." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 335 + ], + "type": "text", + "content": "On top of that, we provide a novel 3D lane dataset using a simple yet efficient auto-labeling pipeline. Existing datasets often yield inaccurate labels containing noise and outliers, particularly in the far-range, since they rely on Li-DAR to recover 3D information. Our auto-labeling instead uses temporal aggregation along video-sequences resulting in consistent and accurate 3D lane labels up to 250 meters, making a valuable contribution to the research community." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 253, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 253, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 253, + 346 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 294, + 479 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 348, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 294, + 371 + ], + "type": "text", + "content": "- We present a novel 3D lane detector that integrates spatial and temporal knowledge into a sparse transformer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 372, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 407 + ], + "type": "text", + "content": "- We propose a new attention mechanism that focuses on learning lane-structure and successfully leverages temporal keys from past lane observations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 407, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 430 + ], + "type": "text", + "content": "- We formulate a continuous lane representation tailored for sparse architectures and a temporal regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 453 + ], + "type": "text", + "content": "- Our method achieves state-of-the-art performance on all 3D lane benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 479 + ], + "type": "text", + "content": "- We provide a new dataset for 3D lane detection providing accurate and consistent long-range labels." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 490, + 139, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 139, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 139, + 502 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Dense BEV vs. sparse query-based detection. A common strategy in 3D lane detection involves operating on an intermediate dense BEV representation generated from the FV. Early methods, such as 3D-LaneNet [9], GenLaneNet [11] and 3D-SpLineNet [38], applied IPM to project FV features onto a flat ground plane. However, the assumption of a perfectly flat road is often violated, leading to degraded BEV features. Alternatively, methods from related fields like 3D object detection [14, 15, 23] and BEV segmentation [37] proposed to learn the BEV representation. These concepts were applied to 3D lane detection with PersFormer [6] leveraging deformable cross-attention (DCA) [54], whereas BEV-LaneDet [49] relies on simple multi-layer perceptrons (MLPs). LaneCPP [39] incorporates surface priors to learn 3D features using depth classification inspired by Lift-Splat-Shoot (LSS) [37]. Despite these advancements, all methods remain constrained by an" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 498, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 498, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 498, + 84 + ], + "type": "text", + "content": "intermediate error-prone BEV representation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 87, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 230 + ], + "type": "text", + "content": "In contrast, sparse methods based on DETR [3, 54] avoid BEV representation by modeling 3D objects as queries that exchange information via attention mechanisms. In 3D object detection, 3D position-aware object queries were introduced that are directly associated with the perspective view through DCA [26, 34]. Similarly, in 3D lane detection, CurveFormer [2] adapts the sparse query design to lanes using polynomial-based curve queries, though these lack the flexibility to capture diverse lane structures in real-world scenarios. LATR [31] refines this approach by representing lanes as individual points but fails to incorporate crucial priors in both the attention mechanism and lane representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 232, + 553, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 602 + ], + "type": "text", + "content": "Priors in lane detection. An indisputable property about lanes lies in their smoothness and continuity. Several methods have proposed continuous lane representations in both 2D [8, 29, 47] and 3D [2, 28, 38, 39] lane detection. In contrast to discrete approaches [6, 9, 11, 16, 49], continuous methods directly provide smooth curves, requiring almost no post-processing and exploit the entire available dense ground truth during training [38, 39]. While CLGO [28] and CurveFormer [2] use simple polynomials, 3D-SpLineNet [38] and LaneCPP [39] use a B-Spline [7] representation, which offers local control over curve segments enabling it to model complex shapes. Additionally, lane structures exhibit inherent properties, such as parallelism between lanes. E.g. SGNet [44] introduces a penalty term to enforce a fixed lane width, but assumes a flat ground plane. GP [20] employs a parallelism loss to maintain a constant local distance between neighboring points, though its effectiveness is influenced by the number of anchor points used. In contrast, LaneCPP [39] demonstrates a more elegant way to encourage parallelism as well as other spatial properties like surface smoothness and curvature restriction using regularization losses. Motivated by LaneCPP, we propose a more suitable Catmull-Rom spline [4] representation tailored for the transformer's attention with sparse queries. We further present more sophisticated loss formulations and integrate prior knowledge into the transformer attention mechanism enhancing the focus on relevant relations. While a concurrent work also uses priors in attention [5], our approach focuses on explicit modeling of structural lane relations and even effectively incorporates temporal knowledge in lane detection for the first time." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Temporal modeling. Another source of prior knowledge is grounded in the history of observations. For video-based 3D object detection various approaches exist to model temporal interactions. BEV-based methods [23] transform past BEV features according to the ego-motion and fuse them with the current representation. More advanced perspective view (PV) methods [26, 30, 32] instead follow the sparse query design but require storing dense historic PV feature maps to interact with object queries. Recently," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29100" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 556, + 256 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 556, + 256 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 556, + 256 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 556, + 256 + ], + "type": "image", + "image_path": "58a284d290683b79f2026809a8d222e2fc12b28edc9cc3ac10e419e411115b1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 298 + ], + "type": "text", + "content": "Figure 1. Overview of our method SparseLaneSTP. Sparse lane queries are processed by a transformer integrating temporal knowledge and spatial lane structure priors in a novel spatio-temporal attention. Based on these priors, we formulate spatial and temporally consistent regularization. Finally, our network predicts control points defining our new continuous 3D lane representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 316, + 294, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 400 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 400 + ], + "type": "text", + "content": "object-centric temporal modeling [27, 50] has been proposed, which only stores sparse historic object queries and propagates them according to ego- and object-motion in 3D space. In particular, Sparse4Dv2 [27] and StreamPETR [50] leverage the sparse query design and model object queries as hidden states of a transformer, where the relation between current and historic queries is learned via attention." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 571 + ], + "type": "text", + "content": "In 3D lane detection, only few approaches exist that use temporal modeling. STLane3D [51] follows a BEV-based method, PETRv2 [30] fuses historic PV features together with propagated 3D positional embeddings. On the other hand, Anchor3DLane-T [16] leverages temporal information by associating 3D anchors from the current frame with features from previous frames. In contrast, our sparse query-based architecture is more suitable for the effective memory-efficient object-centric paradigm. Consequently, we propose a temporal attention mechanism inspired by [27, 50], which exploits lane-specific properties by reinforcing most relevant query interactions. The static nature of lane markings also allows us to formulate regularization to encourage temporally consistent perception behavior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 584, + 138, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 138, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 138, + 597 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "In the following, we describe our 3D lane detection method illustrated in Fig. 1. Given an image " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": ", the model predicts a set of " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " lanes as continuous curves in 3D space, parameterized by our novel tailored spline representation with " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " control points " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " (see Section 3.1) and lane categories " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": ". The core component of our architecture is the transformer decoder, where each control point, inspired by sparse query design, is paired with a learned context embedding and is iteratively processed through " + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": " decoder layers." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 314, + 439, + 390 + ], + "blocks": [ + { + "bbox": [ + 315, + 314, + 439, + 390 + ], + "lines": [ + { + "bbox": [ + 315, + 314, + 439, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 314, + 439, + 390 + ], + "type": "image", + "image_path": "a8a5ac857e3e0ced946926e3dd3ec2b1c5b5dfd232cea74fe234a70ddcef4a7c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 392, + 411, + 400 + ], + "lines": [ + { + "bbox": [ + 363, + 392, + 411, + 400 + ], + "spans": [ + { + "bbox": [ + 363, + 392, + 411, + 400 + ], + "type": "text", + "content": "(a) Front-view" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 459, + 316, + 541, + 389 + ], + "blocks": [ + { + "bbox": [ + 459, + 316, + 541, + 389 + ], + "lines": [ + { + "bbox": [ + 459, + 316, + 541, + 389 + ], + "spans": [ + { + "bbox": [ + 459, + 316, + 541, + 389 + ], + "type": "image", + "image_path": "8a683ecd4aee8f9635024464c782e4c9ad2a3abddbf8cc2e707a62a96e4c54e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 391, + 520, + 400 + ], + "lines": [ + { + "bbox": [ + 492, + 391, + 520, + 400 + ], + "spans": [ + { + "bbox": [ + 492, + 391, + 520, + 400 + ], + "type": "text", + "content": "(b) BEV" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "text", + "content": "Figure 2. CR spline and B-Spline control points in comparison, both fit to the ground truth. B-Spline control points do not align with the curve and are therefore not suited the sparse query design, whereas CR control points exactly match curve geometry." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 586 + ], + "type": "text", + "content": "In contrast to previous work [2, 24, 31] that simply adapt the DETR [3, 54] framework, our transformer architecture introduces a novel spatio-temporal attention layer to capture spatial intra- and inter-lane relations, while a temporal cross-attention mechanism integrates past observations via propagated lane queries from a memory queue. Finally, besides standard detection losses, our model is cautiously regularized by spatial and temporal objectives to enhance robust and consistent detection behavior." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 598, + 443, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 443, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 443, + 611 + ], + "type": "text", + "content": "3.1. 3D lane representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "For the 3D lane representation, we take inspiration from prior work [38, 39] that demonstrated significant benefits of continuous curves. In these frameworks, the model output serves as control points of B-Spline curves that do not necessarily align with the curve geometry. However, using the sparse query design, the 3D position of control points serve as the model's internal state and should therefore correspond to the exact 3D position of lane points. Thus, spline" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29101" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "control points should directly lie on the curve, which is not the case for B-Splines (see Fig. 2). Therefore, we choose Catmull-Rom (CR) splines [4], a class of piece-wise defined smooth third-order polynomial splines, for which the curve inherently passes through its control points." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": "We adapt this spline-model to define a curve representation " + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i(s) \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": " representing the entire " + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": " line proposal parameterized by " + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": " control points with 3 spatial " + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "inline_equation", + "content": "(x_{ij}, y_{ij}, z_{ij})" + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": " and 1 visibility " + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "inline_equation", + "content": "(v_{ij})" + }, + { + "bbox": [ + 55, + 133, + 296, + 181 + ], + "type": "text", + "content": " dimension as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 189, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 189, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 189, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i} (s) = \\left[ \\begin{array}{l l l l} s ^ {3} & s ^ {2} & s & 1 \\end{array} \\right] \\cdot \\mathbf {M} _ {\\mathrm {C R}} \\cdot \\mathbf {P} _ {i} \\quad \\text {w i t h} \\tag {1}", + "image_path": "a4be4a0140228d5441d12ef6cf4a114a30f29b6743bfec6673599ae7cfa4eb09.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 205, + 295, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 205, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 70, + 205, + 295, + 269 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i} = \\left[ \\mathbf {P} _ {\\mathrm {3 D}, i} \\mid \\mathbf {P} _ {v, i} \\right] = \\left[ \\begin{array}{c c c c} x _ {i 1} & y _ {i 1} & z _ {i 1} & v _ {i 1} \\\\ & \\dots & & \\\\ x _ {i j} & y _ {i j} & z _ {i j} & v _ {i j} \\\\ & \\dots & & \\\\ x _ {i M} & y _ {i M} & z _ {i M} & v _ {i M} \\end{array} \\right], \\tag {2}", + "image_path": "c7dc80ea7ce3ef64192fa0e1e84b7cc32b7560050b8a1c949a85af18d204b785.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": "with curve argument " + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "s\\in [0,1]" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i\\in \\mathbb{R}^{M\\times 4}" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": " denotes the control point matrix composed of " + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{3\\mathrm{D},i}\\in \\mathbb{R}^{M\\times 3}" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": " defining the curve 3D shape and visibility " + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{v,i}\\in \\mathbb{R}^{M\\times 1}" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "[\\cdot |\\cdot ]" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": " is the concatenation operator and " + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{CR}}" + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "text", + "content": " the CR coefficient matrix. Note that we can eventually express the curve computation as a simple matrix multiplication between the pre-computed spline arguments and the entire predicted control points " + }, + { + "bbox": [ + 55, + 277, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 173, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 173, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 173, + 381 + ], + "type": "text", + "content": "3.2. Overall architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": "Backbone and query initialization. The RGB input image " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " is first processed by a CNN backbone (e.g. ResNet [12]) to extract image features of " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{H_{\\mathbf{F}} \\times W_{\\mathbf{F}} \\times C}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " is the channel size. Similar to [31], we then feed " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " through a lane instance segmentation branch to obtain initial query embedding vectors " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "{}^0\\mathbf{Q} \\in \\mathbb{R}^{N \\times M \\times C}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " denotes the number of line proposals and " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " the number of spline control points. An MLP then predicts initial control points " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "{}^0\\mathbf{P} \\in \\mathbb{R}^{N \\times M \\times 4}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "inline_equation", + "content": "{}^0\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 388, + 296, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "text", + "content": "Transformer decoder. In each transformer decoder layer " + }, + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "text", + "content": ", queries are processed by two attention layers" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 528, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 528, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 82, + 528, + 295, + 544 + ], + "type": "interline_equation", + "content": "{ } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } = \\operatorname { S T A } \\left( { } ^ { l - 1 } \\mathbf { Q } , { } ^ { l - 1 } \\mathbf { P } , \\mathbf { Q } _ { \\text { M e m } } , \\mathbf { P } _ { \\text { M e m } } \\right) \\tag {3}", + "image_path": "61b495015ecbc2da02235ef0745898583257c162a0551a242d9141def0670c87.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 545, + 294, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 545, + 294, + 561 + ], + "spans": [ + { + "bbox": [ + 82, + 545, + 294, + 561 + ], + "type": "interline_equation", + "content": "{ } ^ { l } \\mathbf { Q } _ { \\mathrm { D C A } } = \\mathrm { D C A } \\big ( { } ^ { l } \\mathbf { Q } _ { \\mathrm { S T A } } , \\mathbf { F } , { } ^ { l - 1 } \\mathbf { P } _ { 2 \\mathrm { D } } \\big ) , \\tag {4}", + "image_path": "bf59c669baf460ef87ca3fa093c2c66f313137b84be63669247ed57d2e22076f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{STA}(\\cdot)" + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": " denotes our proposed spatio-temporal attention, which is explained in Section 3.3, and " + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{DCA}(\\cdot)" + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": " the standard deformable cross-attention presented in [54]. " + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": " are the 2D spline control points in image coordinates resulting from re-projecting " + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": ". Finally, the query embeddings are fed through a feed-forward network " + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "inline_equation", + "content": "^l\\mathbf{Q} = \\mathrm{FFN}(^l\\mathbf{Q}_{\\mathrm{DCA}})" + }, + { + "bbox": [ + 55, + 570, + 295, + 653 + ], + "type": "text", + "content": " composed of an MLP and layer normalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Prediction layer. After each transformer attention layer, the emerging query embeddings " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "{}^{l}\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " are fed through a prediction layer that outputs the spline control points " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "{}^{l}\\mathbf{P}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " as well as classification probabilities " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{N \\times (K + 1)}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " lane categories and one background class per line. Particularly," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 216 + ], + "type": "image", + "image_path": "d36f3edc002eec1903da03a9d36119061be4262d0382ab5a775fe04e10924fc2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 223, + 544, + 234 + ], + "lines": [ + { + "bbox": [ + 323, + 223, + 544, + 234 + ], + "spans": [ + { + "bbox": [ + 323, + 223, + 544, + 234 + ], + "type": "text", + "content": "Figure 3. Overview of our spatio-temporal attention module." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": "for each layer " + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "inline_equation", + "content": "{}^{l}\\mathbf{Q}_{ij}" + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": " three MLPs with shared weights among all layers are used to predict the " + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "inline_equation", + "content": "z-" + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 253, + 555, + 290 + ], + "type": "text", + "content": "-component of the control point matrix" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 296, + 553, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 296, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 344, + 296, + 553, + 311 + ], + "type": "interline_equation", + "content": "{ } ^ { l } \\mathbf { P } _ { x , i j } = \\sigma ( \\mathrm { M L P } _ { x } \\left( { } ^ { l } \\mathbf { Q } _ { i j } \\right) ) \\cdot \\left( x _ { e } - x _ { s } \\right) + x _ { s } \\tag {5}", + "image_path": "0dd1d96fb483d73f056c412d35490222afcd4b0bab85390dbbb413696e9d6cb5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 312, + 553, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 312, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 345, + 312, + 553, + 327 + ], + "type": "interline_equation", + "content": "{ } ^ { l } \\mathbf { P } _ { z , i j } = \\sigma ( \\mathrm { M L P } _ { z } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) \\cdot ( z _ { e } - z _ { s } ) + z _ { s } \\tag {6}", + "image_path": "484ddf465309432c432392200767e69f65807a766bd70e0ffc4d00b8ca19bb44.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 330, + 553, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 330, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 345, + 330, + 553, + 345 + ], + "type": "interline_equation", + "content": "{ } ^ { l } \\mathbf { P } _ { v , i j } = \\sigma ( \\mathrm { M L P } _ { v } ( { } ^ { l } \\mathbf { Q } _ { i j } ) ) , ( 7 )", + "image_path": "43a4dd53211276261644def4d8c8cd80941a020fcca0d4c45a9f3b6d1d55ee85.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": "where the sigmoid " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": " normalizes the MLP output to [0, 1]. The start-/end- of " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": "-range " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "x_{s / e}" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "z_{s / e}" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": " scales up the normalized output to the desired range. The longitudinal " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": "-component " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "{}^l\\mathbf{P}_{y,ij}" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": " is pre-defined and uniformly distributed between " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "y_{s}" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "inline_equation", + "content": "y_{e}" + }, + { + "bbox": [ + 313, + 350, + 555, + 446 + ], + "type": "text", + "content": " to avoid over-parameterization and potential overfitting as argued in [39]. For classification, all queries belonging to the same line proposal are averaged and fed into an MLP followed by a softmax layer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "content": "Finally, the formulation of our supervised losses is inspired by [39], with L1 loss for regression of " + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "content": ", binary cross-entropy for visibility " + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "content": " and focal loss [25] for classification." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 489, + 541, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 541, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 541, + 502 + ], + "type": "text", + "content": "3.3. Integrating spatial and temporal knowledge" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "content": "Memory queue. To facilitate attention to past observations, queries are stored within a memory queue. Following a similar approach to [50], this queue is recursively updated after each prediction step and retains queries from the past " + }, + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "content": " frames. Therefore, embedding vectors and their associated control points are recursively pushed into the queue using a first-in, first-out (FIFO) strategy, which yields" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 330, + 597, + 553, + 614 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 597, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 330, + 597, + 553, + 614 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\text {M e m}} = \\left[ \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 1)} \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {Q} _ {: N _ {\\text {M e m}}} ^ {(t - T)} \\right] \\quad \\text {a n d} \\tag {8}", + "image_path": "fa489b227c6ae4e953b4e2b9746aef4e6f8bfb3b77e8f9b7c3fa8f885648a4f9.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 331, + 616, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 616, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 331, + 616, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\mathrm {M e m}} = \\left[ \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 1)} \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - 2)} \\mid \\dots \\mid \\mathbf {P} _ {: N _ {\\mathrm {M e m}}} ^ {(t - T)} \\right], \\tag {9}", + "image_path": "0ed9c3c5a697c3262760bda97379722a3cd2c5e5349cc79c4f34e1eea8c62869.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{:N_{\\mathrm{Mem}}}^{(t)}\\in \\mathbb{R}^{(N_{\\mathrm{Mem}}\\times M\\times C)}" + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "content": " denotes the embedding vectors from the " + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{Mem}}" + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "content": " most confident predictions of frame " + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathrm{Mem}}\\in \\mathbb{R}^{((T\\cdot N_{\\mathrm{Mem}})\\times M\\times C)}" + }, + { + "bbox": [ + 313, + 639, + 553, + 677 + ], + "type": "text", + "content": " the full memory queue." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Temporal propagation. Since the model is supposed to learn spatio-temporal context, it must incorporate the positional information of 3D control points from past frames." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29102" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": "However, due to the vehicle's ego-motion, lane positions shift relative to the vehicle. Therefore, the 3D coordinates of past lane control points stored in the memory queue must be transformed according to the ego-pose " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{(t)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " of frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": ", which is represented as a homogeneous transformation matrix. Consequently, a control point " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{ij}^{(t - k)\\rightarrow (t)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " propagated from frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "t - k" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " to present frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 162, + 295, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 162, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 65, + 162, + 295, + 180 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\left[ \\mathbf {E} _ {i n v} ^ {(t)} \\cdot \\mathbf {E} ^ {(t - k)} \\cdot \\mathbf {P} _ {\\mathrm {3 D}, i j} ^ {(t - k)} \\mid \\mathbf {P} _ {v, i j} ^ {(t - k)} \\right], \\tag {10}", + "image_path": "9e857bb66258c3727b2f40ac69fd4eb86cf1bc6580b8cba09f4971d1fb543471.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{inv}" + }, + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "text", + "content": " denotes the inverted matrix " + }, + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 184, + 296, + 268 + ], + "type": "text", + "content": ". Obviously, only the geometry component is affected by the propagation, whereas the visibility component remains unchanged. To incorporate spatio-temporal information from propagated control points into queries, we apply a positional encoding (PE) based on 3D geometry and visibility (see Fig. 3). This process yields position-informed embedding vectors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 272, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 272, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 70, + 272, + 295, + 290 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {Q}} _ {i j} ^ {(t - k) \\rightarrow (t)} = \\mathbf {Q} _ {i j} ^ {(t - k) \\rightarrow (t)} + \\mathrm {P E} \\left(\\mathbf {P} _ {i j} ^ {(t - k) \\rightarrow (t)}\\right), \\tag {11}", + "image_path": "b6908cacc9a45a52a5d87631f076b165b175000cbfb956e09f126c6c55eaeebe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 294, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 401 + ], + "type": "text", + "content": "which are processed by the spatio-temporal attention layer. Spatio-temporal attention. Most methods [2, 31] rely solely on global self-attention, which contains a majority of redundant lane query interactions that can distract the model from learning relevant context. To address this, we leverage spatial and temporal priors, distinguishing three key relation types: intra-lane (between adjacent points on the same lane), inter-lane (between parallel lanes), and historic (between current and past queries in the memory queue)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "text", + "content": "We therefore introduce same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA). Applying specific masking, SLA restricts interactions to queries " + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{\\mathrm{SLA}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "text", + "content": " within the same line, PNA facilitates interactions with queries " + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{\\mathrm{PNA}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "text", + "content": " from neighboring lines, and TCA enables interactions with closest past queries " + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{\\mathrm{TCA}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 521 + ], + "type": "text", + "content": " in the memory queue (see Fig. 3). Such restriction of query interactions reduces redundancy, encourages the model to focus on critical lane-specific spatial context and integrates valuable temporal information from past observations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "text", + "content": "Regularization. Based on spatial and temporal priors we can formulate regularization techniques that promote robust and consistent detection behavior. For the spatial regularization " + }, + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {spatial }}" + }, + { + "bbox": [ + 55, + 521, + 295, + 605 + ], + "type": "text", + "content": " we adopt the scheme proposed in [39] encouraging lane parallelism, surface smoothness and suppressing excessive curvature. Besides, temporal consistency can be achieved by optimizing a smoothness loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 610, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 610, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 61, + 610, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t e m p} = \\frac {1}{N} \\sum_ {i} ^ {N} \\int_ {s} \\overline {{\\mathbf {f}}} _ {v, i} ^ {(t)} (s) \\cdot \\left| \\left| \\mathbf {f} _ {\\mathrm {3 D}, i} (s) - \\overline {{\\mathbf {f}}} _ {\\mathrm {3 D}, i} ^ {(t)} (s) \\right| \\right| _ {1}, \\tag {12}", + "image_path": "f73841ac161d135da313358145f93a43bcfc35b1e9bc15dbedebf1c724f15614.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{f}}_{3\\mathrm{D},i}^{(t)}(s)" + }, + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{f}}_{v,i}^{(t)}(s)" + }, + { + "bbox": [ + 55, + 649, + 295, + 676 + ], + "type": "text", + "content": " is an exponentially moving average of the past predictions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 680, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 680, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 107, + 680, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {f}}} _ {i} ^ {(t)} = \\alpha \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t)} + (1 - \\alpha) \\cdot \\overline {{\\mathbf {f}}} _ {i} ^ {(t - 1)} \\tag {13}", + "image_path": "3d0e47569adfd205341f07cfd09ab74bd9b5e351c12ec6adf70fcd82b32685cf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 190, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 190, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 190, + 714 + ], + "type": "text", + "content": "with smoothing factor " + }, + { + "bbox": [ + 55, + 701, + 190, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 55, + 701, + 190, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 187 + ], + "type": "image", + "image_path": "fda28af3d24d1f53837be1f37fc3ab7a591e3c4296b656c452cc522035b8533a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 194, + 501, + 206 + ], + "lines": [ + { + "bbox": [ + 367, + 194, + 501, + 206 + ], + "spans": [ + { + "bbox": [ + 367, + 194, + 501, + 206 + ], + "type": "text", + "content": "Figure 4. Our auto-labeling strategy." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 217, + 434, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 217, + 434, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 434, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 434, + 323 + ], + "type": "image", + "image_path": "ecde0324ebefa099ca401ef2b608f2e31a5ccbea365a1f208f0eebf6f577adf9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 325, + 399, + 335 + ], + "lines": [ + { + "bbox": [ + 353, + 325, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 353, + 325, + 399, + 335 + ], + "type": "text", + "content": "(a) OpenLane" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 344, + 533, + 355 + ], + "lines": [ + { + "bbox": [ + 334, + 344, + 533, + 355 + ], + "spans": [ + { + "bbox": [ + 334, + 344, + 533, + 355 + ], + "type": "text", + "content": "Figure 5. Examples from OpenLane compared to ours." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 217, + 553, + 323 + ], + "blocks": [ + { + "bbox": [ + 436, + 217, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 436, + 217, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 436, + 217, + 553, + 323 + ], + "type": "image", + "image_path": "b69ab776c619896e7369f6884e623b91a59630cd36aa4981715490cff32ede5d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 455, + 325, + 533, + 335 + ], + "lines": [ + { + "bbox": [ + 455, + 325, + 533, + 335 + ], + "spans": [ + { + "bbox": [ + 455, + 325, + 533, + 335 + ], + "type": "text", + "content": "(b) Our 3D lane dataset" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 376, + 524, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 524, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 524, + 390 + ], + "type": "text", + "content": "4. Auto-labeling and our 3D lane dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 422 + ], + "type": "text", + "content": "In this section, we describe our auto-labeling pipeline and provide details about our new 3D lane dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 432, + 440, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 440, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 440, + 445 + ], + "type": "text", + "content": "4.1. Auto-labeling strategy" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "While 3D lane datasets like OpenLane [6] and ONCE-3DLanes [52] have contributed significantly to the field, they possess specific shortcomings that our auto-labeling strategy addresses. As illustrated in Fig. 5a the LiDAR-based annotation approach, which recovers 3D information of 2D lane labels, introduces noise and outliers, particularly in the far range due to sparse scan points, while also posing calibration and synchronization challenges. To overcome these limitations, we propose an automated labeling approach (see Fig. 4) that leverages confident near-range lane point estimates from a 2D detector, which are temporally accumulated in 3D space along an accurate ego-trajectory. More precisely, common visual odometry methods [1, 42, 43] are applied to recover the vehicle ego-trajectory from the input video sequence. Our road surface generator then fits a spline surface to the vehicle's trajectory and orientations. On the other hand, lane pseudo-labels in image space are obtained from a state-of-the-art 2D lane detector [21, 46], providing particularly reliable detection results in the near-range. Lane and surface information are then combined in 3D by projecting near-range detection results to the road surface and accumulating line points frame" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29103" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 136, + 145 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 136, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 136, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 136, + 145 + ], + "type": "image", + "image_path": "45007fe7c541efa29d5b0b60c92d84b92fdfdd84020bbd290e108e3adf4c121f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 146, + 129, + 183 + ], + "blocks": [ + { + "bbox": [ + 61, + 146, + 129, + 183 + ], + "lines": [ + { + "bbox": [ + 61, + 146, + 129, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 129, + 183 + ], + "type": "image", + "image_path": "966afd6af52639b3b08bf7c93796f58a06b8177ddbae6d2dec1674f6a72ebf3f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 186, + 111, + 194 + ], + "lines": [ + { + "bbox": [ + 81, + 186, + 111, + 194 + ], + "spans": [ + { + "bbox": [ + 81, + 186, + 111, + 194 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 81, + 186, + 111, + 194 + ], + "type": "inline_equation", + "content": "t = 1" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 137, + 70, + 214, + 145 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 214, + 145 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 214, + 145 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 214, + 145 + ], + "type": "image", + "image_path": "f09b95709d6a03753724bf7ca9e067d9c2bd2913f359c8bf0ba7d9ce109a5e9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 138, + 146, + 207, + 182 + ], + "blocks": [ + { + "bbox": [ + 138, + 146, + 207, + 182 + ], + "lines": [ + { + "bbox": [ + 138, + 146, + 207, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 146, + 207, + 182 + ], + "type": "image", + "image_path": "b8cae87e5cbccb743668b6d9968a55f484020858cff834198213833a2bb68383.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 186, + 190, + 194 + ], + "lines": [ + { + "bbox": [ + 159, + 186, + 190, + 194 + ], + "spans": [ + { + "bbox": [ + 159, + 186, + 190, + 194 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 159, + 186, + 190, + 194 + ], + "type": "inline_equation", + "content": "t = 2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 292, + 145 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 292, + 145 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 292, + 145 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 292, + 145 + ], + "type": "image", + "image_path": "13099c230a58274e3bdfe6fb7038a3cc0cbec47c5a9c24c3127a5ba81df1aaf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 217, + 146, + 286, + 183 + ], + "blocks": [ + { + "bbox": [ + 217, + 146, + 286, + 183 + ], + "lines": [ + { + "bbox": [ + 217, + 146, + 286, + 183 + ], + "spans": [ + { + "bbox": [ + 217, + 146, + 286, + 183 + ], + "type": "image", + "image_path": "2bcf7f0b866de407cb7f9a0ddc78bda559a27cd0223ebaf863f8b59a5e2a2792.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 186, + 268, + 194 + ], + "lines": [ + { + "bbox": [ + 237, + 186, + 268, + 194 + ], + "spans": [ + { + "bbox": [ + 237, + 186, + 268, + 194 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 237, + 186, + 268, + 194 + ], + "type": "inline_equation", + "content": "t = 3" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 227 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of two models, one without temporal and one with temporal for three consecutive frames." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 248, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 285 + ], + "type": "text", + "content": "by frame along the ego-trajectory of the sequence. Finally, we apply an accurate semantic segmentation network to automatically assign occlusion labels to each lane point." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 292, + 164, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 164, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 164, + 304 + ], + "type": "text", + "content": "4.2. Dataset properties" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 310, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 465 + ], + "type": "text", + "content": "Using the approach described in 4.1, we are able to automatically generate temporally consistent, high-quality 3D lane labels in a range up to " + }, + { + "bbox": [ + 55, + 310, + 296, + 465 + ], + "type": "inline_equation", + "content": "250\\mathrm{m}" + }, + { + "bbox": [ + 55, + 310, + 296, + 465 + ], + "type": "text", + "content": " (see Fig. 5b). Besides 3D lane geometry, we provide camera intrinsics and extrinsics, lane categories, lane visibility information and consistent track IDs. We further supply relational information about lanes (like line parallelism), local 3D ego-trajectory and global pose information for each frame. Important to mention is also that in contrast to former datasets, our visibility information contains occlusions caused by other traffic participants, enabling models to consistently learn lane structure independent of interruptions and lane occlusions. We refer to the supplementary material for additional information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 475, + 135, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 135, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 135, + 489 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 496, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 533 + ], + "type": "text", + "content": "In this section, we first outline our experimental setup, followed by a comprehensive analysis of our approach on two public and our own novel 3D lane dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 540, + 227, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 227, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 227, + 552 + ], + "type": "text", + "content": "5.1. Datasets and evaluation metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 557, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 604 + ], + "type": "text", + "content": "OpenLane [6] based on the Waymo Open dataset [45] remains the most prominent real-world 3D lane dataset to date. It is comprised of 200K images from 1000 sequences recorded in several cities in the USA." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": "ONCE-3DLanes [52] is another real-world dataset derived from the ONCE dataset [33]. It consists of 211K images collected from multiple cities in China." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": "Our 3D lane dataset, which was introduced in Section 4, was gathered from multiple regions and countries across the globe. It consists of 511K images with a " + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": " train/test split." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Evaluation metrics. The OpenLane evaluation scheme [6, 11] uses the euclidean distance at uniformly distributed" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 128 + ], + "type": "table", + "html": "
Continuous Rep.STARegularizationF1(%)↑Gain(%)
61.8(baseline)
✓62.9+1.1
✓✓65.0+2.1
✓✓✓65.3+0.3
", + "image_path": "f8d9996ddeae8a02bbfa50f4a47bf46679370ddbffa55984c1d7dc49beb15b97.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 317, + 184, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 170 + ], + "type": "text", + "content": "Table 1. Performance gain for different contributions on OpenLane using our novel Continuous Lane Representation, Spatio-Temporal Attention (STA) and Regularization." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 184, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 209 + ], + "type": "table", + "html": "
AttentionGlobalSLAPNASLA + PNASLA + PNA + TCA
F1-Score62.962.863.563.865.0
", + "image_path": "5980bb39de87b938b580cbcfb4bda04813643ef7ec8057d7c263c43eb8f1dba0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 317, + 274, + 553, + 303 + ], + "blocks": [ + { + "bbox": [ + 313, + 217, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 250 + ], + "type": "text", + "content": "Table 2. Analysis of spatio-temporal attention. Same line attention (SLA), parallel neighbor attention (PNA) and temporal cross-attention (TCA) and combinations compared to global attention." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 274, + 553, + 303 + ], + "lines": [ + { + "bbox": [ + 317, + 274, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 553, + 303 + ], + "type": "table", + "html": "
Num. Frames (T)012345
F1(%)↑63.864.164.665.064.563.9
", + "image_path": "bb510596631316ddc44b66109bb556746b5b68fa428381c5918c008e003997d4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 312, + 548, + 323 + ], + "lines": [ + { + "bbox": [ + 318, + 312, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 318, + 312, + 548, + 323 + ], + "type": "text", + "content": "Table 3. Analysis of number of frames in the memory for STA." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "text", + "content": "points and a range-IoU to compute the F1-Score. Moreover, x- and z-errors in near- " + }, + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "inline_equation", + "content": "(0 - 40\\mathrm{m})" + }, + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "text", + "content": " and far-range " + }, + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "inline_equation", + "content": "(40 - 100\\mathrm{m})" + }, + { + "bbox": [ + 313, + 346, + 553, + 407 + ], + "type": "text", + "content": " evaluate the geometric accuracy. ONCE-3DLanes instead uses the uni-lateral chamfer distance (CD) and computes Precision (P), Recall (R) and F1-Score based on it." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "text", + "content": "For our dataset, we extend the well-established OpenLane evaluation framework [6, 11]. Due to the long-range labels, we add x-/z-error intervals for the ranges " + }, + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "inline_equation", + "content": "100 - 150\\mathrm{m}" + }, + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "inline_equation", + "content": "150 - 200\\mathrm{m}" + }, + { + "bbox": [ + 313, + 408, + 554, + 516 + ], + "type": "text", + "content": ". Since our dataset offers visibility labels for occlusions, we also add a visibility Intersection-over-Union metric (Vis-IoU). This score helps to gain understanding about the model's capability of distinguishing visible lane segments from parts occluded by other traffic participants and provides a measure for start-end range detection." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 527, + 444, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 444, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 444, + 540 + ], + "type": "text", + "content": "5.2. Implementation details" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": "We use an image input size " + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "inline_equation", + "content": "720 \\times 960" + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": " and ResNet-50 [12] as our standard backbone. The query generator is trained on the auxiliary task of lane instance segmentation similar to [31]. The transformer decoder applied in the large model used for the evaluation in Section 5.4 consists of " + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": " layers, whereas the lighter model used for ablation studies in Sec. 5.3 has 2 layers. The lane representation uses " + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "inline_equation", + "content": "M = 20" + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": " control points on OpenLane and ONCE-3DLanes and 30 control points on our dataset due to the larger range. The number of lane proposals is set to " + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "inline_equation", + "content": "N = 40" + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": " for OpenLane and ours and 10 for ONCE-3DLanes, which is sufficient for the small number of occurring lanes per frame. Regarding temporal fusion, our best model uses propagated queries from " + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": " historic frames." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29104" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 196 + ], + "type": "table", + "html": "
MethodBackboneResolutionF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [9]VGG-16360 x 48044.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [11]ERFNet360 x 48032.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [6]EfficientNet-B7360 x 48050.50.4850.5530.3640.43142.455.648.646.640.050.7
CurveFormer [2]EfficientNet-B7360 x 48050.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [49]ResNet-34360 x 48058.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane-T [16]ResNet-18360 x 48054.30.2750.3100.1050.13547.258.052.748.745.851.7
PETRv2 [30]VoV-99360 x 48061.20.4000.5730.2650.413------
LATR [31]ResNet-50720 x 96061.90.2190.2590.0750.10455.268.257.155.452.361.5
LaneCPP [39]EfficientNet-B7360 x 48060.30.2640.3100.0770.11753.664.456.754.952.058.7
PVALane [53]Swin-B720 x 96063.40.2260.2570.0930.11956.167.764.058.653.660.8
GroupLane [22]ConvNext-Basenot specified64.10.3200.4410.2330.402------
SparseLaneSTP (Ours)ResNet-50720 x 96066.10.2030.2400.0660.09257.373.060.158.358.266.5
", + "image_path": "9972d0ed71a3c90c0ce3f0d7e9d2cff87b91be34b25a26cbad3a46a4531f5f48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 226, + 555, + 303 + ], + "blocks": [ + { + "bbox": [ + 116, + 204, + 493, + 216 + ], + "lines": [ + { + "bbox": [ + 116, + 204, + 493, + 216 + ], + "spans": [ + { + "bbox": [ + 116, + 204, + 493, + 216 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on OpenLane [6]. Best performance and second best are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 226, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 59, + 226, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 59, + 226, + 555, + 303 + ], + "type": "table", + "html": "
MethodF1-Score(%)↑X-error / Z-error (m)↓Vis-IoU(%)↑
0 m - 40 m40 m - 100 m100 m - 150 m150 m - 200 m
PersFormer [6]59.20.191 / 0.0720.344 / 0.1450.671 / 0.2200.845 / 0.37970.1
LaneCPP [39]63.10.144 / 0.0400.298 / 0.0930.494 / 0.1880.761 / 0.29377.5
Baseline (LATR) [31]65.10.122 / 0.0380.263 / 0.0980.468 / 0.1850.700 / 0.26478.3
SparseLaneSTP (ours)68.20.109 / 0.0330.233 / 0.0920.443 / 0.1820.646 / 0.26381.4
", + "image_path": "14840fbe39534a1f66d702d98ca73ae90a67205e9397384a485d38242e81be0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 91, + 341, + 260, + 410 + ], + "blocks": [ + { + "bbox": [ + 91, + 341, + 260, + 410 + ], + "lines": [ + { + "bbox": [ + 91, + 341, + 260, + 410 + ], + "spans": [ + { + "bbox": [ + 91, + 341, + 260, + 410 + ], + "type": "image", + "image_path": "59add4d0edbc3287de048c01130ac5cceeeecf0846d50a479317c0862b580d40.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 418, + 263, + 430 + ], + "lines": [ + { + "bbox": [ + 87, + 418, + 263, + 430 + ], + "spans": [ + { + "bbox": [ + 87, + 418, + 263, + 430 + ], + "type": "text", + "content": "Figure 7. Efficiency analysis of SparseLaneSTP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 59, + 444, + 295, + 553 + ], + "blocks": [ + { + "bbox": [ + 106, + 311, + 503, + 322 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 503, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 503, + 322 + ], + "type": "text", + "content": "Table 5. Quantitative comparison on our 3D lane dataset. Best performance and second best are highlighted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 444, + 295, + 553 + ], + "lines": [ + { + "bbox": [ + 59, + 444, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 59, + 444, + 295, + 553 + ], + "type": "table", + "html": "
MethodF1(%)↑P(%)↑R(%)↑CD(m)↓
3D-LaneNet [9]44.7361.4635.160.127
Gen-LaneNet [11]45.5963.9535.420.121
SALAD [52]64.0775.9055.420.098
PersFormer [6]74.3380.3069.180.074
Anchor3DLane [16]74.8780.8569.710.060
LATR [31]80.5986.1275.730.052
PVALane [53]76.3582.8170.830.059
GroupLane [22]80.7382.5678.900.053
SparseLaneSTP (ours)82.7586.4779.330.048
", + "image_path": "624221499df66012ed5d127dcb1fcd873988ffcbbe512e5ffcd64711f73b1882.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 561, + 282, + 572 + ], + "lines": [ + { + "bbox": [ + 67, + 561, + 282, + 572 + ], + "spans": [ + { + "bbox": [ + 67, + 561, + 282, + 572 + ], + "type": "text", + "content": "Table 6. Quantitative comparison on ONCE-3DLanes [52]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 597, + 154, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 154, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 154, + 608 + ], + "type": "text", + "content": "5.3. Ablation studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 663 + ], + "type": "text", + "content": "We perform a comprehensive analysis validating the effectiveness of our design choices in SparseLaneSTP. For these experiments, we used a smaller model consisting of two decoder layers evaluated on OpenLane." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Table 1 summarizes the impact of our contributions. Our CR spline based lane representation improves the F1-Score by over 1 percentage point compared to the discrete one (baseline). The most significant gain in F1-Score is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 415 + ], + "type": "text", + "content": "achieved by the novel STA component, which encourages focus on relevant relations based on lane structure and, more importantly, on related past observations by integrating temporal cross-attention. Eventually, regularization enhances robustness and consistency, leading to improved generalization, as evidenced by the additional F1-Score gain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "content": "As shown in Table 2, global self-attention apparently fails to capture important relations of lane structures. The combination of our proposed SLA and PNA demonstrates the benefits of enhancing the focus on intra- and interlane relations by an F1-Score increase " + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "inline_equation", + "content": "+0.9\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "content": ". Integrating knowledge about past observations via our suggested TCA yields the greatest improvement of additional " + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "inline_equation", + "content": "+1.2\\%" + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "content": ". As indicated by Table 3, " + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "content": " frames in the memory produces the best results. Using too few frames provides insufficient temporal context, whereas too many frames apparently introduce redundancy and distract the model from relevant information. We further qualitatively demonstrate the beneficial impact of our primary contribution on temporal fusion in Fig. 6. The non-temporal model fails to retain the detection from frame " + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 420, + 556, + 661 + ], + "type": "text", + "content": " in subsequent frames as lane visibility deteriorates. In contrast, the temporally-aware model maintains the detection despite partial occlusions, demonstrating improved robustness. Regularization further preserves temporal consistency mitigating the gradual disappearing and drifting in the non-temporal model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Finally, Fig. 7 reveals efficiency benefits of our approach. SparseLaneSTP with six decoder layers achieves 11.0 FPS, comparable to LATR at 12.1 FPS. While our proposed temporal integration (memory queue and STA) in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29105" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 126, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 126, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 126, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 126, + 178 + ], + "type": "image", + "image_path": "5e60b3b80c7c73572554363de529587302b91ef5c04f758e9b59908a04491d22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 180, + 125, + 190 + ], + "lines": [ + { + "bbox": [ + 115, + 180, + 125, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 180, + 125, + 190 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 70, + 182, + 178 + ], + "blocks": [ + { + "bbox": [ + 126, + 70, + 182, + 178 + ], + "lines": [ + { + "bbox": [ + 126, + 70, + 182, + 178 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 182, + 178 + ], + "type": "image", + "image_path": "dff13bace1f0129b4c3e707079de744a9e73e04046c362ddb98a09018cff4c3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 200, + 538, + 211 + ], + "lines": [ + { + "bbox": [ + 67, + 200, + 538, + 211 + ], + "spans": [ + { + "bbox": [ + 67, + 200, + 538, + 211 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of LATR and SparseLaneSTP on OpenLane with ground truth. Main differences are highlighted." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 70, + 249, + 178 + ], + "blocks": [ + { + "bbox": [ + 182, + 70, + 249, + 178 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 249, + 178 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 249, + 178 + ], + "type": "image", + "image_path": "0a52c94e084f557f206f90e78bed8b67d07d194a86863776cf820913af80061a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 180, + 249, + 190 + ], + "lines": [ + { + "bbox": [ + 238, + 180, + 249, + 190 + ], + "spans": [ + { + "bbox": [ + 238, + 180, + 249, + 190 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 250, + 72, + 306, + 178 + ], + "blocks": [ + { + "bbox": [ + 250, + 72, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 250, + 72, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 250, + 72, + 306, + 178 + ], + "type": "image", + "image_path": "751ce85b9ff1b7970c5939aef8016806f8ab1ebef5cf3997a2520a4733c9491e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 372, + 178 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 372, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 372, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 372, + 178 + ], + "type": "image", + "image_path": "78d9e261e7932fc92997426db18bc46ebaf145014f211d2754e5a7f988c6aff1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 180, + 373, + 190 + ], + "lines": [ + { + "bbox": [ + 361, + 180, + 373, + 190 + ], + "spans": [ + { + "bbox": [ + 361, + 180, + 373, + 190 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 373, + 72, + 429, + 178 + ], + "blocks": [ + { + "bbox": [ + 373, + 72, + 429, + 178 + ], + "lines": [ + { + "bbox": [ + 373, + 72, + 429, + 178 + ], + "spans": [ + { + "bbox": [ + 373, + 72, + 429, + 178 + ], + "type": "image", + "image_path": "ef425db9dba7b94391769e8a6def0d3476c889caa25df0ea86ad485e2cc8c8e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 429, + 70, + 496, + 178 + ], + "blocks": [ + { + "bbox": [ + 429, + 70, + 496, + 178 + ], + "lines": [ + { + "bbox": [ + 429, + 70, + 496, + 178 + ], + "spans": [ + { + "bbox": [ + 429, + 70, + 496, + 178 + ], + "type": "image", + "image_path": "9a69dd879bf2780b9f47cbab5b2ddc8dacadb901e0e2b6879c4d34fe2f278fe1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 180, + 496, + 190 + ], + "lines": [ + { + "bbox": [ + 485, + 180, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 485, + 180, + 496, + 190 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 496, + 72, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 496, + 72, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 496, + 72, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 496, + 72, + 553, + 178 + ], + "type": "image", + "image_path": "b69532c9906cecac45d277e2297489022d81ce5eae3f58bd375ce8cbea91d2a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "text", + "content": "produces a modest " + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "text", + "content": " overhead, it also enables a smaller model with only two layers to reach " + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "inline_equation", + "content": "65.3\\%" + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "text", + "content": " F1-Score " + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "inline_equation", + "content": "(+3.4\\%" + }, + { + "bbox": [ + 55, + 232, + 296, + 281 + ], + "type": "text", + "content": " over six-layer LATR) at a significantly higher speed of 16.5 FPS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 289, + 137, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 137, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 137, + 300 + ], + "type": "text", + "content": "5.4. Main results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 306, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 400 + ], + "type": "text", + "content": "In this section, we compare our method to other 3D lane detectors on public benchmarks as well as on our own dataset. Results on OpenLane. Table 4 presents a quantitative comparison demonstrating the superiority of our approach. SparseLaneSTP reaches the highest F1-Score, exceeding " + }, + { + "bbox": [ + 55, + 306, + 296, + 400 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 306, + 296, + 400 + ], + "type": "text", + "content": ", significantly surpassing the second-best model despite its larger backbone and even achieves the lowest geometric errors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 403, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 581 + ], + "type": "text", + "content": "Fig. 8 illustrates the advantages of our approach qualitatively across multiple aspects. We compare our SparseLaneSTP to LATR [31], which is most related to our baseline. As shown in Fig. 8a, our continuous lane representation enables precise visibility estimation with accurate detection of start/end of markings, whereas LATR - constrained by fixed discrete anchor points - produces erroneous estimates. In Fig. 8b, our temporal fusion preserves the line despite partial occlusions, whereas non-temporal LATR fails. It also struggles to estimate faint lane markings on the far side of the junction shown in Fig. 8c, while with STA our model captures lane structures more effectively. Eventually, in Fig. 8d our model exhibits a stronger reliance on priors resulting in robust and consistent behavior, most likely due to the incorporated regularization." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 653 + ], + "type": "text", + "content": "Results on ONCE-3DLanes. As shown in Table 6, our method achieves the highest detection scores and demonstrates accurate regression with a significantly lower CD, despite the inaccurate camera parameters provided by the dataset. Notably, even though ONCE-3DLanes has a lower frame rate, our temporal approach performs effectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "Results on our 3D lane dataset. Initial results on our novel dataset are presented in Table 5. Our proposed extension of the evaluation metrics provides a more comprehensive assessment of overall performance, particularly in the far-range and for occlusion detection. To facilitate a fair" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 328, + 230, + 454, + 308 + ], + "blocks": [ + { + "bbox": [ + 328, + 230, + 454, + 308 + ], + "lines": [ + { + "bbox": [ + 328, + 230, + 454, + 308 + ], + "spans": [ + { + "bbox": [ + 328, + 230, + 454, + 308 + ], + "type": "image", + "image_path": "31f9551c46bb17502bde5d110bc06b96fc2a9d14e306edb0acd1e72e614aa638.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 308, + 408, + 315 + ], + "lines": [ + { + "bbox": [ + 381, + 308, + 408, + 315 + ], + "spans": [ + { + "bbox": [ + 381, + 308, + 408, + 315 + ], + "type": "text", + "content": "Front-view" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 456, + 230, + 541, + 272 + ], + "blocks": [ + { + "bbox": [ + 456, + 230, + 541, + 272 + ], + "lines": [ + { + "bbox": [ + 456, + 230, + 541, + 272 + ], + "spans": [ + { + "bbox": [ + 456, + 230, + 541, + 272 + ], + "type": "image", + "image_path": "c773caa969f771df28e62a6d442b824794f8eeb5ef38f8c74e42a6c3b01b53c6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 456, + 272, + 541, + 314 + ], + "blocks": [ + { + "bbox": [ + 456, + 272, + 541, + 314 + ], + "lines": [ + { + "bbox": [ + 456, + 272, + 541, + 314 + ], + "spans": [ + { + "bbox": [ + 456, + 272, + 541, + 314 + ], + "type": "image", + "image_path": "798bbab2a95bb344bb0a6b015891784b153d81168102718a5c417f74c6927f5c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 323, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 321, + 323, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 321, + 323, + 545, + 334 + ], + "type": "text", + "content": "Figure 9. Qualitative example of SparseLaneSTP on our data." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "content": "comparison, we train the baseline and other methods with specific modifications to range and capacity. Our method excels in F1-Score, error metrics and notably in the VisIoU, highlighting the superior performance of our temporal model in detecting occlusions. Fig. 9 shows a qualitative result on our dataset, showcasing accurate long-range lane estimation " + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "inline_equation", + "content": "(200\\mathrm{m})" + }, + { + "bbox": [ + 313, + 356, + 555, + 453 + ], + "type": "text", + "content": ". For more qualitative results with comparisons to the baseline we refer to the supplementary." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 388, + 475 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 482, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 704 + ], + "type": "text", + "content": "In this work, we presented a novel 3D lane detection approach that combines sparse lane transformers with spatial and temporal priors. We proposed a continuous lane representation, tailored for the sparse query design, thereby overcoming limitations of discrete representations. Additionally, we developed a novel attention layer that effectively captures intra- and inter-lane relations while integrating temporal information through propagated lane queries. Furthermore, we introduced a dedicated regularization scheme enhancing temporal consistency. Our experiments validate the effective integration of temporal information and spatial priors, as well as the benefits of our lane representation, leading to state-of-the-art performance across all evaluated 3D lane datasets. Moreover, we presented a new 3D lane dataset, constructed using a simple yet effective auto-labeling pipeline, providing geometrically accurate labels that span long distances and explicitly annotate occlusions, fostering further advancements in the field. In future, we will explore an extension to 3D lane tracking to leverage the full potential of temporal information." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29106" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 703 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Lucas R Agostinho, Nuno M Ricardo, Maria I Pereira, Antoine Hiolle, and Andry M Pinto. A practical survey on visual odometry for autonomous driving in challenging scenarios and conditions. IEEE Access, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 187 + ], + "type": "text", + "content": "[2] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 295, + 231 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 294, + 261 + ], + "type": "text", + "content": "[4] Edwin Catmull and Raphael Rom. A class of local interpolating splines. Computer aided geometric design, 1972. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 263, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 295, + 316 + ], + "type": "text", + "content": "[5] Yifan Chang, Junjie Huang, Xiaofeng Wang, Yun Ye, Zhujin Liang, Yi Shan, Dalong Du, and Xingang Wang. Rethinking lanes and points in complex scenarios for monocular 3d lane detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 380 + ], + "type": "text", + "content": "[6] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 382, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 294, + 403 + ], + "type": "text", + "content": "[7] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "text", + "content": "[8] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 487 + ], + "type": "text", + "content": "[9] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 540 + ], + "type": "text", + "content": "[10] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 542, + 295, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 542, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 542, + 295, + 594 + ], + "type": "text", + "content": "[11] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 596, + 295, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 596, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 57, + 596, + 295, + 638 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 639, + 295, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 639, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 57, + 639, + 295, + 681 + ], + "type": "text", + "content": "[13] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 682, + 295, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 682, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 682, + 295, + 703 + ], + "type": "text", + "content": "[14] Junjie Huang and Guan Huang. Bevdet4d: Exploit temporal cues in multi-camera 3d object detection. CoRR," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 706 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 427, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 427, + 83 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 427, + 83 + ], + "type": "text", + "content": "abs/2203.17054, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 84, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 84, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 84, + 553, + 115 + ], + "type": "text", + "content": "[15] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. CoRR, abs/2112.11790, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 116, + 553, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 116, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 317, + 116, + 553, + 169 + ], + "type": "text", + "content": "[16] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 170, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 170, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 553, + 233 + ], + "type": "text", + "content": "[17] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 235, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 235, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 235, + 553, + 266 + ], + "type": "text", + "content": "[18] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 266, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 331 + ], + "type": "text", + "content": "[19] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 373 + ], + "type": "text", + "content": "[20] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 374, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 415 + ], + "type": "text", + "content": "[21] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 417, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 417, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 417, + 553, + 469 + ], + "type": "text", + "content": "[22] Zhuoling Li, Chunrui Han, Zheng Ge, Jinrong Yang, En Yu, Haoqian Wang, Xiangyu Zhang, and Hengshuang Zhao. Groupplane: End-to-end 3d lane detection with channel-wise grouping. IEEE Robotics and Automation Letters (RA-L), 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 471, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 524 + ], + "type": "text", + "content": "[23] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 553, + 578 + ], + "type": "text", + "content": "[24] Benchcheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized HD map construction. In Proc. of the International Conf. on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 578, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 620 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 621, + 553, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 621, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 317, + 621, + 553, + 662 + ], + "type": "text", + "content": "[26] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion. CoRR, abs/2211.10581, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 664, + 553, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 664, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 317, + 664, + 553, + 695 + ], + "type": "text", + "content": "[27] Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d v2: Recurrent temporal fusion with sparse model. CoRR, abs/2305.14018, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 696, + 553, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 696, + 553, + 706 + ], + "spans": [ + { + "bbox": [ + 316, + 696, + 553, + 706 + ], + "type": "text", + "content": "[28] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Ze" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29107" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "jian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "text", + "content": "[29] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 294, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 294, + 201 + ], + "type": "text", + "content": "[30] Yingfei Liu, Junjie Yan, Fan Jia, Shuai Lin, Aqi Gao, Tiancai Wang, and Xiangyu Zhang. Petr2: A unified framework for 3d perception from multi-camera images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 201, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 294, + 255 + ], + "type": "text", + "content": "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 289 + ], + "type": "text", + "content": "[32] Zhipeng Luo, Changqing Zhou, Gongjie Zhang, and Shijian Lu. Detr4d: Direct multi-view 3d object detection with sparse attention. arXiv preprint arXiv:2212.07849, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 289, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 294, + 342 + ], + "type": "text", + "content": "[33] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Chunjing Xu, and Hang Xu. One million scenes for autonomous driving: ONCE dataset. In NeurIPS Datasets and Benchmarks, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 384 + ], + "type": "text", + "content": "[34] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 427 + ], + "type": "text", + "content": "[35] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 427, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 470 + ], + "type": "text", + "content": "[36] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 470, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 470, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 294, + 513 + ], + "type": "text", + "content": "[37] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[38] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "text", + "content": "[39] Maximilian Pittner, Joel Janai, and Alexandru P Condurache. Lanecpp: Continuous 3d lane detection using physical priors. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 599, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 642 + ], + "type": "text", + "content": "[40] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 642, + 294, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 294, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 294, + 684 + ], + "type": "text", + "content": "[41] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 684, + 294, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 294, + 706 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 294, + 706 + ], + "type": "text", + "content": "[42] Davide Scaramuzza, Friedrich Fraundorfer, and Roland Siegwart. Real-time monocular visual odometry for on-road" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 708 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "vehicles with 1-point ransac. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2009. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 137 + ], + "type": "text", + "content": "[43] Frank Steinbrücker, Jürgen Sturm, and Daniel Cremers. Real-time visual odometry from dense rgb-d images. In Proc. of the IEEE International Conf. on Computer Vision (ICCV) Workshops, 2011. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 137, + 553, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 137, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 137, + 553, + 180 + ], + "type": "text", + "content": "[44] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 180, + 553, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 553, + 244 + ], + "type": "text", + "content": "[45] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 299 + ], + "type": "text", + "content": "[46] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[47] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "text", + "content": "[48] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "text", + "content": "[49] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 514 + ], + "type": "text", + "content": "[50] Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi-angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2023. 1, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 514, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 556 + ], + "type": "text", + "content": "[51] Yin Wang, Qiuyi Guo, Peiwen Lin, Guangliang Cheng, and Jian Wu. Spatio-temporal fusion-based monocular 3d lane detection. In Proc. of the British Machine Vision Conf. (BMVC), 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 556, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 610 + ], + "type": "text", + "content": "[52] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 610, + 553, + 665 + ], + "type": "text", + "content": "[53] Zewen Zheng, Xuemin Zhang, Yongqiang Mou, Xiang Gao, Chengxin Li, Guoheng Huang, Chi-Man Pun, and Xiaochen Yuan. Pvalane: Prior-guided 3d lane detection with view-agnostic feature alignment. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2024. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "type": "text", + "content": "[54] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In Proc. of the International Conf. on Learning Representations (ICLR), 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29108" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "text", + "content": "[55] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29109" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_content_list.json b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..04bff5dab665c2335a5d80e35b21f3f38cb05601 --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_content_list.json @@ -0,0 +1,1707 @@ +[ + { + "type": "text", + "text": "SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs", + "text_level": 1, + "bbox": [ + 94, + 128, + 901, + 155 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahui Wang $^{1*}$ , Zuyan Liu $^{1,2*}$ , Yongming Rao $^{2,1}$ , Jiwen Lu $^{1\\dagger}$ $^{1}$ Tsinghua University $^{2}$ Tencent Hunyuan X", + "bbox": [ + 259, + 179, + 736, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 251, + 326, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than $5\\%$ ) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparsity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers $1.38 \\times$ real-time acceleration and $52\\%$ memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM.", + "bbox": [ + 88, + 282, + 485, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 694, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autoregressive large language models (LLMs) [6, 14, 37, 39, 45] have revolutionized artificial intelligence with their exceptional instruction-following capabilities and expansive knowledge repositories. Building upon this foundation, researchers have extended LLMs to multimodal domains, particularly in vision-language integration, creating multimodal large language models (MLLMs) [3, 7, 21, 26, 46, 53] that process both textual and visual inputs. Current approaches typically augment pre-trained LLMs by incorporating visual encoders (e.g., CLIP [42] or SigLIP [55]) paired with", + "bbox": [ + 89, + 720, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b63af9698478eebef7190b66803074286ef786d4969a68dfcc0abe71adfb790a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 252, + 924, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2106dd222bf1efb38776519850f029b2534a612d9a38e21c6e74cf306382324f.jpg", + "image_caption": [ + "Figure 1. Head Sparsity Emerges from Visual Concept Responses. We observe the visual-relevant heads are sparse in various MLLMs. Based on this observation, we devise a KV-Cache optimization strategy that allocates asymmetric budgets to LLM heads based on their importance for visual tokens, achieving better trade-off under limited computational resources." + ], + "image_footnote": [], + "bbox": [ + 521, + 468, + 921, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lightweight adapters to project visual features into the language model's hidden space. While these architectures demonstrate remarkable multimodal reasoning abilities, how LLMs fundamentally acquire visual comprehension during supervised fine-tuning remains poorly understood. This knowledge gap constrains our ability to recognize cross-modal alignment and risks undervaluing visual semantics during multi-modal relevant tasks and applications, which", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally to this research. † Corresponding author.", + "bbox": [ + 107, + 886, + 480, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23177", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "may potentially leading to suboptimal architecture designs and inefficient computational resource allocation.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present the first systematic investigation into how visual concepts are processed within LLMs. Through rigorous analysis of attention mechanisms, we uncover a critical phenomenon that only a small subset of attention heads (termed visual heads) drive visual content understanding, while the majority remain text-specialized. Specifically, our experiments reveal two critical properties of visual heads: (1) Sparsity: Less than $5\\%$ of attention heads are intrinsically visual-active across layers, even in models trained with extensive multimodal data; (2) Universality: Visual heads emerge consistently across diverse LLM architectures(e.g., Vicuna [8] and Qwen2 [41]) and generalize to multiple attention paradigms such as multi-head attention (MHA) [49] and grouped query attention (GQA) [2].", + "bbox": [ + 88, + 123, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To systematically identify these visual heads, we propose a training-free framework that quantifies the visual relevance of attention heads through targeted cross-modal response analysis. Specifically, our approach leverages OCR as an anchor task to establish precise correspondence between text outputs and visual inputs: for each generated word, we trace its activation back to spatially aligned image patches, enabling direct measurement of how specific attention heads mediate visual-text alignment. By analyzing and recoding the attention score of all the attention heads across a certain amount of samples, we compute visual scores that rank heads by their visual responsiveness. Crucially, while our identification mechanism relies on OCR's granular spatial grounding, we demonstrate that the detected visual heads exhibit task-agnostic generalizability—they remain dominant in diverse vision-language tasks including object recognition and scene understanding.", + "bbox": [ + 91, + 337, + 485, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, to demonstrate the effectiveness of visual heads on practical multi-modal tasks, we introduce SparseMM, a KV-Cache optimization framework that exploits visual head sparsity to achieve accelerated inference. As multimodal inputs grow in complexity—spanning multiturn dialogues [19, 20, 52], high-resolution interleaved images [9, 51], and dense video/3D sequences [12, 16, 23]—the computational overhead of maintaining full KV-Caches becomes prohibitive. Existing compression methods, however, treat all attention heads uniformly, disregarding the critical role of sparse visual heads in encoding visual semantics.", + "bbox": [ + 89, + 595, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SparseMM addresses this by asymmetrically allocating KV-Cache budgets: visual heads receive prioritized retention based on their precomputed visual scores, while nonvisual heads undergo aggressive compression via a hybrid strategy combining 1) Score-Preferred Cache (allocating cache budget based on visual head scores), 2) Uniform-Based Cache (preserving minimal budget for all the heads), and 3) Local Window Cache (preserving cache budget for recent tokens). This mixed approach ensures better accuracy", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "efficiency trade-offs, such that visual heads retain more computational cost while other heads are dynamically throttled.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experimental results demonstrate that SparseMM outperforms other strong baselines across multiple datasets, including DocVQA [35], OCRBench [29], TextVQA [44], MMBench [28], etc. For instance, on DocVQA, LLaVA-NeXT-Vicuna-7B [27] achieves the same level of accuracy while using only $20\\%$ of the cache, and Qwen2-VL-7B-Instruct [41] achieves equivalent performance with just $5.3\\%$ of the cache. These findings suggest that our method effectively captures visual information while compressing redundancies. Furthermore, the reduction in cache requirements enables our method to achieve lower decoding latency and peak memory usage. For example, LLaVA-NeXT-Mistral-7B [27] maintains nearly constant decoding latency with 32K input tokens, resulting in almost a $50\\%$ acceleration compared to the full model, and reduces memory usage by 5GB.", + "bbox": [ + 511, + 123, + 908, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 381, + 663, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architectures in MLLMs. The predominant architecture for Multi-Modal Large Language Models (MLLMs) consists of three key components: a visual encoder, an adapter, and a LLM. By leveraging alignment training techniques and subsequent fine-tuning, this integrated framework has achieved remarkable performance on various multi-modal understanding tasks [1, 10, 19, 24, 30, 32, 38]. In typical implementations, the visual encoder is realized using models like CLIP [42] or SigLIP [55], which are adept at extracting rich visual representations. The adapter component serves as an intermediary, bridging the gap between the visual features and the language domain; it is often instantiated as a multi-layer perceptron (MLP) or through more complex structures [33]. For the LLM part, architectures such as LLaMA [47, 48] and Qwen [40] series are commonly employed. Notably, previous LLMs such as LLaMA2 [48] utilizes a multi-head attention (MHA) mechanism, while more recent LLMs such as LLaMA3 [11] incorporate a grouped query attention (GQA) [2] design. The GQA approach aggregates multiple queries into a single group that corresponds to one key and one value, thereby substantially reducing memory usage without compromising model performance. In this paper, we observe a universal phenomenon of visual head in MLLMs across LLM architectures and attention implementations. The applications are demonstrated to be effective across the abundant MLLM series.", + "bbox": [ + 511, + 412, + 908, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model Acceleration in MLLMs. With the rapid growth of model size and input sequence length, model acceleration has become an urgent research focus in both language and multimodal domains. In the context of Large Language Models (LLMs), significant efforts have been dedicated to optimizing the prompt encoding phase through efficient compression", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23178", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f454ed261ce1b189df4b39f0ce7847335d8d89395b22cccaed64faa5c14ecf7.jpg", + "image_caption": [ + "Figure 2. Visual Heads are Sparse in MLLMs. We use OCR tasks to obtain visual scores for all heads. Upon visualizing these scores, we discovered that high-scoring heads, which we refer to as visual heads, are quite sparse within the MLLM, comprising only about $5\\%$ . The majority of heads have very low scores, indicating that most heads in LLMs do not focus on visual information." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of the KV-Cache. For instance, StreamingLLM [50] identifies attention sinks to stabilize long-context inference, while H2O [56] introduces a token-level importance scoring mechanism for adaptive KV-Cache eviction. Subsequent works, such as SnapKV [22], PyramidKV [4], and AdaKV [13], further refine the KV-Cache selection strategy by incorporating spatial-temporal redundancy reduction, hierarchical token retention, or dynamic eviction policies. However, these methods, primarily designed for text-only inputs, face limitations when applied to multi-modal scenarios. In MLLMs, acceleration challenges are exacerbated by the increasing complexity of multi-modal prompts (e.g., high-resolution images, videos) and cross-modal fusion mechanisms. Recent attempts, such as FastV [5], accelerate inference via layer-wise pruning of redundant visual tokens. ElasticCache [31], optimizes KV-Cache management during the generation phase. Despite these advances, head-wise acceleration strategies, particularly those targeting modality-specific attention heads, remain underexplored. In this work, we address this gap by proposing a systematic framework based on our findings about visual heads, enabling efficient deployment of MLLMs in resource-constrained environments.", + "bbox": [ + 91, + 276, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visual Heads are Sparse in MLLMs", + "text_level": 1, + "bbox": [ + 89, + 626, + 415, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our exploration of head sparsity in multi-modal large language models. To start with, we provide the preliminaries on the relations from LLMs to visual instruction tuning. Then we describe our approach for identifying sparse visual heads in MLLMs in Sec. 3.2, then introduce the deployment of visual heads in model acceleration in Sec. 3.3.", + "bbox": [ + 89, + 652, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. What is Learned during Visual Instruct Tuning", + "text_level": 1, + "bbox": [ + 89, + 771, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Extending a Large Language Model to a Multimodal Large Language Model is achieved by integrating a visual encoder $E$ , an adapter $H$ , and the LLM $p_{\\theta}$ . The original LLM is trained solely on textual tasks to model the distribution of text sequences as $p_{\\theta}(\\mathbf{x}) = \\prod_{i=1}^{N} p_{\\theta}(\\mathbf{x}_i | \\mathbf{x}_{ i \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 511, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This strategy effectively reduces the computational complexity from $O(N^2)$ to $O(N \\times L)$ , where $L = 32$ , thereby decreasing the runtime during the prefilling stage. To evaluate the attention for keys outside the local window, we compute the average attention weight:", + "bbox": [ + 89, + 556, + 485, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} _ {j} = \\frac {1}{w} \\sum_ {i = 1} ^ {w} \\hat {A} _ {i, j}, \\quad \\text {f o r} j \\in \\{0, \\dots , L - w - 1 \\} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 643, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ultimately, we select the top $K$ KV Caches based on the computed attention scores, where the value of $K$ is given by", + "bbox": [ + 89, + 693, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK = r + b _ {i j} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 736, + 483, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our three-part allocation mechanism leverages head sparsity to significantly reduce the computational and memory overhead in multi-modal models. It ensures that each head receives a guaranteed minimum cache allocation through both the nearest neighbor and uniform baseline allocations. The remaining cache is then adaptively distributed based on the visual attention scores, thereby achieving an efficient balance between computational efficiency and overall model performance.", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct extensive experiments to validate the effectiveness of Visual Head. We first introduce our experimental settings in Section 4.1. Then we present a comparison with state-of-the-art KV Cache Compression method, demonstrating that our method maintains strong performance in Section 4.2 while maintaining computational efficiency in Section 4.3. Moreover, we provide some analytical experiments to illustrate the importance of visual head in Section 4.4.", + "bbox": [ + 511, + 119, + 906, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 513, + 263, + 720, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. We employ three multi-modal models: LLaVA-NeXT-Vicuna-7B [27], LLaVA-NeXT-Mistral-7B [27], and Qwen2-VL-7B [41]. LLaVA-NeXT-Vicuna-7B is derived from Vicuna-7B [8], a model based on Multi-Head Attention (MHA) [49], and comprises 32 layers with 32 attention heads per layer. In contrast, LLaVA-NeXT-Mistral-7B is built upon Mistral-7B [17] and utilizes Grouped-Query Attention (GQA) [2]. This model features 32 layers, with each layer consisting of 32 query heads and 8 key-value heads. Similarly, Qwen2-VL-7B [41] is based on Qwen2, another GQA model, and is composed of 28 layers, with each layer containing 28 query heads and 4 key-value heads.", + "bbox": [ + 511, + 294, + 908, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We adopt SnapKV [22], PyramidKV [4], and AdaKV [13] as our baseline methods, as they represent the latest and state-of-the-art in KV Cache compression. SnapKV [22] utilizes an \"observation window\" mechanism to identify and preserve the most critical KV caches. PyramidKV [4] implements a hierarchical allocation strategy that distributes the KV cache budget in a pyramidal manner. More budget is allocated to lower layers with dispersed attention, and less to higher layers with focused patterns. Meanwhile, AdaKV [13] proposes a dynamic allocation framework that assigns varying cache budgets to different attention heads within a layer, based on the intra-layer attention distributions. In addition, to validate the Visual Head's effectiveness, we compare it with a Random Head baseline, where head scores are randomly initialized.", + "bbox": [ + 511, + 484, + 908, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Benchmarks.** To comprehensively assess the effectiveness of Visual Head in visual perception, we conduct evaluations on five widely used benchmarks covering both visual question answering (VQA) and image captioning tasks. Specifically, we utilize DocVQA [35], OCRBench [29], TextVQA [44], ChartQA [34], and TextCaps [43], which collectively encompass a diverse set of challenges, including document understanding, OCR-based question answering, chart interpretation, and text-based image captioning. Additionally, we also select the mainstream multiple-choice visual benchmarks, including MMBench [28] and VQAv2 [15] for a comprehensive evaluation.", + "bbox": [ + 511, + 719, + 910, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa47fa8748f9d80ddb92335408ccceb5d7c69f3425aed590f3712f956132697a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 90, + 710, + 104 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6573657b84bac5246ae74f7048fb5d2f6a3248ad5deb1f1e6a917f906cfef68f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 111, + 264, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89c8417768106a4c3cfb0f5d3ec0cbc47465b5064fb82293c9012c3ddff3a6d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 188, + 264, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aeaf95f78375fc875f73175624aaff5b79060624deb03d9c167b55bd9bc0b149.jpg", + "image_caption": [ + "Figure 4. Main Results on Multi-Modal Benchmarks. We evaluate SparseMM and other baselines on several multimodal benchmarks, and conduct experiments on a series of backbones. Our SparseMM consistently outperforms the other baselines." + ], + "image_footnote": [], + "bbox": [ + 91, + 261, + 264, + 337 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d5f14414c5ad7cdb902b9f4a9488a134a5883be30edbbbf9541b67d09225ad8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 112, + 424, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c4792824b23c298c8fa3dd550aa230fd83c39d6b1e871b2bbf1b81a696dd9445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 188, + 424, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2d16304559653f36b0e331556a7a5288ca6a90352fb08d41b0d37dd104adeb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 264, + 255, + 424, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/980da7ecb011f07d969120469ab4c558b24f4e0409f42b37fbf479596804f4d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 112, + 583, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eaa819719db344dd1e0579e5248f1f5fbd86de82f6b5cce5f89c7e5ff0634cca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 188, + 583, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e359a35cc745a7b188bd6e2fa8ac435b8707102659408b805d7a8a0c740145f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 255, + 583, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f06e602eadebeebd2794b611a5940bb5b024e63d6aa89bc0eda09caad7b192cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 112, + 743, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33d350ac77b28bb09814b2b519f6f3886b0678956bfe4da838fb5668c9028a4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 188, + 743, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c752a07eaa6613efa074e05e3e7e4d0d6a61c119d7dc82f55a0098afbcbf67e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 261, + 743, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da689e7f680bd9fc7b9ffad7ed7fcf8f67003e9b7b81dc058c8ba71455d74c92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 112, + 903, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/044716afad92b715ce54869ff2309ed9b093a7f1e96d34aa47346459eab2b537.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 188, + 903, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a2e0d55b9daea412c215f6e19b966fed7edbbb6f47836ca862a169305e1304e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 261, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0191b3d54c24619896de8093ef5c1e5d594fad8891f1cccd5aa20fa2365021cd.jpg", + "table_caption": [ + "Table 1. Average Number of Input Tokens. We analyzed the average length of input tokens across various benchmarks. Considering that text instructions are typically very short(fewer than 50 tokens), so visual tokens constitute the majority of the input sequence, accounting for $90\\%$ to $99\\%$ of the input tokens." + ], + "table_footnote": [], + "table_body": "
DatasetDocVQAOCRBenchTextVQAChartQATextCaps
LLaVA-Series24331700237622702376
Qwen2-VL-7B-Instruct4830124510246421024
", + "bbox": [ + 93, + 473, + 480, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Multi-Modal Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 546, + 410, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. To determine an appropriate budget allocation, we first measure the average length of input tokens for each benchmark, as reported in Table 1. Given that text instructions typically consist of no more than 50 tokens, the majority of input tokens are attributed to visual tokens. Considering the varying input sequence lengths across different datasets, we select a range of each head's KV Cache budget for evaluation: \\{64, 128, 256, 512, 1024, 2048\\}. Since general visual benchmarks utilize lower image resolutions, we adjust the input token budget range correspondingly: \\{48, 64, 96, 128, 256, 512\\}. This allows us to systematically analyze the impact of different cache sizes on performance and efficiency across various benchmarks.", + "bbox": [ + 88, + 575, + 485, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. Fig. 4 presents the evaluation results for three models and five benchmarks. Our experimental results demonstrate that our proposed method consistently outperforms baseline approaches, particularly under extreme cache budget constraints (e.g., 128 or 256). Under these conditions, our approach maintains performance levels close to those achieved with full cache utilization, significantly outperforming the competing baselines. For instance, on the", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e10d077dc1df307a8538d230ea6e958d18488cf6f13a2a1d90b4bcb5e5f62878.jpg", + "image_caption": [ + "Figure 5. Results on Multiple-choice Benchmarks. We evaluate SparseMM and other baselines on multiple-choice visual benchmarks with Qwen2-VL-7B-Instruct as the backbone model. Our SparseMM consistently outperforms the other baselines." + ], + "image_footnote": [], + "bbox": [ + 514, + 393, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TextVQA [44] task using LLaVA-NeXT-Vicuna-7B, a KV Cache budget of 256—which constitutes only approximately $10.77\\%$ of the average 2376 tokens—yields performance equivalent to the full-cache model, whereas AdaKV [13] and similar methods experience an accuracy drop of roughly $3\\%$ . Similarly, on OCRBench [29], LLaVA-NeXT-Mistral-7B demonstrates only a slight performance degradation at a KV Cache budget of 128 (about $7.5\\%$ of the average 1700 tokens), in contrast to a decline exceeding $10\\%$ observed with other methods. In addition, Qwen2-VL-7B-Instruct on DocVQA [35] maintains performance when operating with a KV Cache budget of 256 (merely $5.3\\%$ of the average 4830 tokens), while alternative approaches suffer performance drops between $5\\%$ and $17\\%$ . These results validate the effectiveness of our method in VQA tasks.", + "bbox": [ + 511, + 595, + 908, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 presents the evaluation results on multiple-choice benchmarks. Our method demonstrates competitive performance on multi-choice benchmarks compared with existing baselines. For instance, with only 96 token budget, our approach retains full performance on MMBench while ex", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23182", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "periencing only a minimal performance degradation ( $< 1\\%$ ) on GQA and VQAv2. These findings substantiate that our method can effectively recognize visual content while exhibiting strong generalizability towards diverse tasks.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our findings in Fig. 4 indicate that the random head baseline consistently produces the poorest performance across nearly all experiments, whereas our method achieves superior outcomes by utilizing Visual Head. This pronounced contrast underscores the efficacy of our approach in accurately capturing the manner in which multimodal language models attend to visual information. It is important to note that the performance of the random head method is comparable to that of SnapKV [22], particularly in the case of the MHA model. This similarity is attributable to the fact that when the scores of all heads are randomly initialized, the cache budget allocated to each head is statistically equivalent, effectively causing the method to revert to the behavior observed with SnapKV [22].", + "bbox": [ + 89, + 157, + 486, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Efficiency Evaluation", + "text_level": 1, + "bbox": [ + 89, + 396, + 292, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup In this subsection, we evaluate the computational efficiency of our proposed method, which holds significant practical value. Accordingly, our efficiency tests are conducted across a range of input token lengths $\\{2\\mathrm{K}, 4\\mathrm{K}, 8\\mathrm{K}, 16\\mathrm{K}, 32\\mathrm{K}\\}$ . For each experiment, the output sequence length was fixed at 100 tokens, with a KV Cache budget set to 256. We computed the average decoding latency and peak memory consumption for each configuration. Notably, all experiments are done using FlashAttention.", + "bbox": [ + 89, + 428, + 485, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoding Latency Fig. 6 illustrates that the reduction in KV Cache in our method substantially decreases the computational load during inference, thereby enhancing inference speed. For instance, when the input sequence length is $8\\mathrm{K}$ , the LLaVA-NeXT-Vicuna-7B model exhibits a speedup of $1.16\\times$ , while at a $32\\mathrm{K}$ input length, the speedup increases to $1.87\\times$ . These findings indicate that our approach significantly accelerates token generation, particularly in high-resolution or long video contexts.", + "bbox": [ + 89, + 573, + 486, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory Cost Our method also offers a marked reduction in peak memory usage, primarily by diminishing the memory footprint associated with the KV Cache. This reduction is especially pronounced in LLaVA-Series models. For example, with an input sequence length of 32K, LLaVA-NeXT-Vicuna-7B with full KV Cache requires 32.87 GB of memory, whereas our method reduces the requirement to 17.38 GB, thereby achieving an approximate $50\\%$ reduction in memory overhead. It is noteworthy that even for the Qwen2-VL-7B-Instruct model, which employs an aggressive compression technique in its GQA framework, we can still reduce the cache by nearly 2GB with 32k inputs.", + "bbox": [ + 89, + 719, + 486, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/200338dabfd258d11d8a56382d7650fb1f9c66893b67a95584a6b413768a56f1.jpg", + "image_caption": [ + "Figure 6. Efficiency Evaluation for SparseMM. Benefiting from the reduction in KV cache, SparseMM can maintain nearly constant decoding latency, achieving up to a $50\\%$ acceleration. Additionally, it effectively reduces peak memory usage." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 906, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 513, + 486, + 616, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance Influence of Visual Heads. To further elucidate the impact of visual heads on the visual perception capabilities of multimodal models, we conducted a series of masking experiments. In these experiments, we selectively masked a specific proportion of the visual heads and, for comparison, randomly masked an equivalent proportion of attention heads. The evaluation was performed on OCR-Bench [29] and TextVQA [44], with performance measured relative to the baseline unmasked model. The results, as illustrated in Fig. 7, reveal that masking visual heads leads to a significant performance decline. In contrast, randomly masking the same proportion of attention heads produced a much smaller impact—for instance. These findings underscore the critical role that visual heads play in enabling MLLMs to effectively capture and process visual information. Moreover, masking the top $5\\%$ of high-scoring visual heads causes a considerably greater performance loss than the additional impact of masking another $5\\%$ , which highlights the sparse yet indispensable distribution of visual heads.", + "bbox": [ + 511, + 515, + 908, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness of Visual Head Identification. To evaluate the robustness of our visual head detecting approach, we show the distribution of the detected visual heads under different datasets and tasks in Fig. 8 and show the accuracy curve in Fig. 9. For the OCR task, we used the Multi-Lingual Text(MLT) [36] and Chinese Text in the Wild(CTW) [54]", + "bbox": [ + 511, + 809, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d523c453f260dcbbb487f8afe152dc6adea0dc1fdd6c6670798d8a6b5730d869.jpg", + "image_caption": [ + "Figure 7. Comparisons of Mmasking Visual Head and Random Head. The left figure is the result on OCRBench, and the right figure is the result on TextVQA." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2a3b587fbb5035d10fe962ef6c5868e2217d23345bfd03bc3f7b7d204b182b5.jpg", + "image_caption": [ + "Figure 8. Visualizations of Visual Heads using Different Datasets. We visualized the attention distribution identified on MLT, CTW, and COCO datasets." + ], + "image_footnote": [], + "bbox": [ + 93, + 265, + 480, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/048a77182e77273234aa4b2489a60e153279a26d29b3f760880f1b2c9e40b950.jpg", + "image_caption": [ + "Figure 9. Results with Different Visual Head Identification Approaches and Datasets. We conduct an evaluation on visual heads identified on different datasets. The results on OCR datasets are similar and better than those on the detection dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 439, + 480, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets. In addition, we consider the object detection task and choose the COCO dataset [25], where the model is required to identify objects present in the images. We then localized the visual heads based on the correspondence between the model's answers and the relevant objects. As shown, the distribution of visual heads is relatively consistent across the OCR datasets, whereas there is greater variation on the COCO dataset. Moreover, experimental results demonstrate that the visual heads identified from OCR tasks are dataset-agnostic and exhibit strong generalizability, with better results than detection tasks. This is because OCR tasks establish an exact one-to-one mapping between the model's output and the visual content, whereas the COCO task, which focuses on larger bounding boxes, introduces more noise and results in less robustness.", + "bbox": [ + 88, + 625, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accuracy and Speed Trade-off. We compared the accuracy and speed of SparseMM with other baselines in Tab. 2. We conducted an experiment on LLaVA-NeXT-Vicuna-7B with", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f9eb4287ac4b7b6367e379ef9ef31e01fe04f9692b621467c17887cd5ade2c91.jpg", + "table_caption": [ + "Table 2. Comparison of Accuracy-speed Trade-off among Different Methods. We compare the speed of all methods under 256 KV Cache budget and 16K input tokens." + ], + "table_footnote": [], + "table_body": "
MethodsDocVQAOCRBenchTextVQAChartQATextCapsLatency(ms)
FullKV0.680.520.650.550.7352.9
SparseMM0.68(-0.00)0.52(-0.00)0.65(-0.00)0.54(-0.01)0.73(-0.00)37.1(-30%)
SnapKV0.64(-0.04)0.46(-0.06)0.62(-0.03)0.50(-0.05)0.65(-0.08)35.3(-33%)
PyramidKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.53(-0.02)0.65(-0.08)34.9(-34%)
AdaKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.49(-0.06)0.66(-0.07)37.3(-29%)
", + "bbox": [ + 516, + 141, + 903, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e18760383c881c12aa28c733f874edb4411ddd624268b8bf6683ecb82532f0ba.jpg", + "image_caption": [ + "Figure 10. Visualizations of Visual Heads. We visualized the attention distribution of several heads. The visual heads are able to accurately capture text or objects within the images, whereas the non-visual heads provide random results." + ], + "image_footnote": [], + "bbox": [ + 514, + 224, + 903, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a budget of 256 KV Cache. With the support of FlashAttention, our decoding latency is comparable to that of other methods, significantly lower than the FullKV method. However, our method outperforms others in terms of performance under the same budget. This effectively demonstrates the efficacy of SparseMM based on visual heads in MLLMs.", + "bbox": [ + 511, + 431, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Visual Heads. To gain a more intuitive understanding of how visual heads process and interpret visual information, we conducted a visualization analysis of visual heads and non-visual heads on LLaVA-NeXT-Vicuna7B. As illustrated in Fig. 10, our observations indicate that non-visual heads often either neglect the image entirely or allocate a disproportionate amount of attention to visually insignificant regions. In contrast, visual heads accurately pinpoint regions of interest, allocating a substantial proportion of attention to these critical areas.", + "bbox": [ + 511, + 526, + 908, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a systematic exploration of the visual processing characteristics inherent in MLLMs. Our analysis reveals a critical sparsity phenomenon that only a small fraction of attention heads actively engage in visual understanding. Leveraging this insight, we propose SparseMM, a novel KV-Cache optimization framework that dynamically allocates asymmetric computation budgets to attention heads based on their visual relevance. SparseMM prioritizes preserving vision-critical information during decoding, thereby achieving a more balanced accuracy-efficiency trade-off. We hope this study inspires deeper investigations into the principles governing multimodal learning.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23184", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62336004, Grant 62125603, Grant 62321005, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 223 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2", + "[2] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 2, 5", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1", + "[4] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information tunneling. arXiv preprint arXiv:2406.02069, 2024.3,5", + "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 3", + "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1", + "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 5", + "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. Advances in Neural Information Processing Systems, 37:42566-42592, 2024. 2" + ], + "bbox": [ + 99, + 232, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 2", + "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2", + "[12] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 2", + "[13] Yuan Feng, Junlin Lv, Yukun Cao, Xike Xie, and S Kevin Zhou. Ada-kv: Optimizing kv cache eviction by adaptive budget allocation for efficient llm inference. arXiv preprint arXiv:2407.11550, 2024. 3, 4, 5, 6", + "[14] GeminiTeam. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1", + "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5", + "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 36:20482-20494, 2023. 2", + "[17] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 5", + "[18] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7 (15):2, 2021. 4", + "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2", + "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-Interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 1", + "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation. Advances in Neural Information Processing Systems, 37:22947-22970, 2024. 3, 5, 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2", + "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 8", + "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,5", + "[28] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5", + "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 2, 5, 6, 7", + "[30] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 2", + "[31] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 3", + "[32] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025. 2", + "[33] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 2", + "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5", + "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6", + "[36] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In 2019 International conference on document analysis and recognition (ICDAR), pages 1582–1587. IEEE, 2019. 7", + "[37] OpenAI. Openai gpt-3.5 api. OpenAI API, 2023. 1", + "[38] OpenAI. Gpt-4v(ison) system card. OpenAI Blog, 2023. 2", + "[39] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1", + "[40] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2", + "[41] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 2, 5", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3", + "[43] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 5", + "[44] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 2, 5, 6, 7", + "[45] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1", + "[46] Qwen Team. Qwen2.5-v1, 2025. 1", + "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5", + "[50] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3", + "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2", + "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23186", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 1", + "[54] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. Journal of Computer Science and Technology, 34(3):509-521, 2019. 7", + "[55] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1, 2, 3", + "[56] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3" + ], + "bbox": [ + 91, + 92, + 486, + 371 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23187", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_model.json b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfd580b623ac001c1d55226e260b6bfc81ff1aaa --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.903, + 0.156 + ], + "angle": 0, + "content": "SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.18, + 0.737, + 0.219 + ], + "angle": 0, + "content": "Jiahui Wang\\(^{1*}\\), Zuyan Liu\\(^{1,2*}\\), Yongming Rao\\(^{2,1}\\), Jiwen Lu\\(^{1\\dagger}\\) \n\\(^{1}\\) Tsinghua University \\(^{2}\\) Tencent Hunyuan X" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.252, + 0.327, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.486, + 0.662 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than \\(5\\%\\)) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparsity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers \\(1.38 \\times\\) real-time acceleration and \\(52\\%\\) memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Autoregressive large language models (LLMs) [6, 14, 37, 39, 45] have revolutionized artificial intelligence with their exceptional instruction-following capabilities and expansive knowledge repositories. Building upon this foundation, researchers have extended LLMs to multimodal domains, particularly in vision-language integration, creating multimodal large language models (MLLMs) [3, 7, 21, 26, 46, 53] that process both textual and visual inputs. Current approaches typically augment pre-trained LLMs by incorporating visual encoders (e.g., CLIP [42] or SigLIP [55]) paired with" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.253, + 0.925, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.469, + 0.922, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.676, + 0.911, + 0.761 + ], + "angle": 0, + "content": "Figure 1. Head Sparsity Emerges from Visual Concept Responses. We observe the visual-relevant heads are sparse in various MLLMs. Based on this observation, we devise a KV-Cache optimization strategy that allocates asymmetric budgets to LLM heads based on their importance for visual tokens, achieving better trade-off under limited computational resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "lightweight adapters to project visual features into the language model's hidden space. While these architectures demonstrate remarkable multimodal reasoning abilities, how LLMs fundamentally acquire visual comprehension during supervised fine-tuning remains poorly understood. This knowledge gap constrains our ability to recognize cross-modal alignment and risks undervaluing visual semantics during multi-modal relevant tasks and applications, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.482, + 0.902 + ], + "angle": 0, + "content": "*Authors contributed equally to this research. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "may potentially leading to suboptimal architecture designs and inefficient computational resource allocation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.336 + ], + "angle": 0, + "content": "To this end, we present the first systematic investigation into how visual concepts are processed within LLMs. Through rigorous analysis of attention mechanisms, we uncover a critical phenomenon that only a small subset of attention heads (termed visual heads) drive visual content understanding, while the majority remain text-specialized. Specifically, our experiments reveal two critical properties of visual heads: (1) Sparsity: Less than \\(5\\%\\) of attention heads are intrinsically visual-active across layers, even in models trained with extensive multimodal data; (2) Universality: Visual heads emerge consistently across diverse LLM architectures(e.g., Vicuna [8] and Qwen2 [41]) and generalize to multiple attention paradigms such as multi-head attention (MHA) [49] and grouped query attention (GQA) [2]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.486, + 0.595 + ], + "angle": 0, + "content": "To systematically identify these visual heads, we propose a training-free framework that quantifies the visual relevance of attention heads through targeted cross-modal response analysis. Specifically, our approach leverages OCR as an anchor task to establish precise correspondence between text outputs and visual inputs: for each generated word, we trace its activation back to spatially aligned image patches, enabling direct measurement of how specific attention heads mediate visual-text alignment. By analyzing and recoding the attention score of all the attention heads across a certain amount of samples, we compute visual scores that rank heads by their visual responsiveness. Crucially, while our identification mechanism relies on OCR's granular spatial grounding, we demonstrate that the detected visual heads exhibit task-agnostic generalizability—they remain dominant in diverse vision-language tasks including object recognition and scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Building on these insights, to demonstrate the effectiveness of visual heads on practical multi-modal tasks, we introduce SparseMM, a KV-Cache optimization framework that exploits visual head sparsity to achieve accelerated inference. As multimodal inputs grow in complexity—spanning multiturn dialogues [19, 20, 52], high-resolution interleaved images [9, 51], and dense video/3D sequences [12, 16, 23]—the computational overhead of maintaining full KV-Caches becomes prohibitive. Existing compression methods, however, treat all attention heads uniformly, disregarding the critical role of sparse visual heads in encoding visual semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "SparseMM addresses this by asymmetrically allocating KV-Cache budgets: visual heads receive prioritized retention based on their precomputed visual scores, while nonvisual heads undergo aggressive compression via a hybrid strategy combining 1) Score-Preferred Cache (allocating cache budget based on visual head scores), 2) Uniform-Based Cache (preserving minimal budget for all the heads), and 3) Local Window Cache (preserving cache budget for recent tokens). This mixed approach ensures better accuracy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.122 + ], + "angle": 0, + "content": "efficiency trade-offs, such that visual heads retain more computational cost while other heads are dynamically throttled." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.909, + 0.365 + ], + "angle": 0, + "content": "Extensive experimental results demonstrate that SparseMM outperforms other strong baselines across multiple datasets, including DocVQA [35], OCRBench [29], TextVQA [44], MMBench [28], etc. For instance, on DocVQA, LLaVA-NeXT-Vicuna-7B [27] achieves the same level of accuracy while using only \\(20\\%\\) of the cache, and Qwen2-VL-7B-Instruct [41] achieves equivalent performance with just \\(5.3\\%\\) of the cache. These findings suggest that our method effectively captures visual information while compressing redundancies. Furthermore, the reduction in cache requirements enables our method to achieve lower decoding latency and peak memory usage. For example, LLaVA-NeXT-Mistral-7B [27] maintains nearly constant decoding latency with 32K input tokens, resulting in almost a \\(50\\%\\) acceleration compared to the full model, and reduces memory usage by 5GB." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.382, + 0.664, + 0.398 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.805 + ], + "angle": 0, + "content": "Architectures in MLLMs. The predominant architecture for Multi-Modal Large Language Models (MLLMs) consists of three key components: a visual encoder, an adapter, and a LLM. By leveraging alignment training techniques and subsequent fine-tuning, this integrated framework has achieved remarkable performance on various multi-modal understanding tasks [1, 10, 19, 24, 30, 32, 38]. In typical implementations, the visual encoder is realized using models like CLIP [42] or SigLIP [55], which are adept at extracting rich visual representations. The adapter component serves as an intermediary, bridging the gap between the visual features and the language domain; it is often instantiated as a multi-layer perceptron (MLP) or through more complex structures [33]. For the LLM part, architectures such as LLaMA [47, 48] and Qwen [40] series are commonly employed. Notably, previous LLMs such as LLaMA2 [48] utilizes a multi-head attention (MHA) mechanism, while more recent LLMs such as LLaMA3 [11] incorporate a grouped query attention (GQA) [2] design. The GQA approach aggregates multiple queries into a single group that corresponds to one key and one value, thereby substantially reducing memory usage without compromising model performance. In this paper, we observe a universal phenomenon of visual head in MLLMs across LLM architectures and attention implementations. The applications are demonstrated to be effective across the abundant MLLM series." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Model Acceleration in MLLMs. With the rapid growth of model size and input sequence length, model acceleration has become an urgent research focus in both language and multimodal domains. In the context of Large Language Models (LLMs), significant efforts have been dedicated to optimizing the prompt encoding phase through efficient compression" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.221, + 0.908, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Visual Heads are Sparse in MLLMs. We use OCR tasks to obtain visual scores for all heads. Upon visualizing these scores, we discovered that high-scoring heads, which we refer to as visual heads, are quite sparse within the MLLM, comprising only about \\(5\\%\\). The majority of heads have very low scores, indicating that most heads in LLMs do not focus on visual information." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.61 + ], + "angle": 0, + "content": "of the KV-Cache. For instance, StreamingLLM [50] identifies attention sinks to stabilize long-context inference, while H2O [56] introduces a token-level importance scoring mechanism for adaptive KV-Cache eviction. Subsequent works, such as SnapKV [22], PyramidKV [4], and AdaKV [13], further refine the KV-Cache selection strategy by incorporating spatial-temporal redundancy reduction, hierarchical token retention, or dynamic eviction policies. However, these methods, primarily designed for text-only inputs, face limitations when applied to multi-modal scenarios. In MLLMs, acceleration challenges are exacerbated by the increasing complexity of multi-modal prompts (e.g., high-resolution images, videos) and cross-modal fusion mechanisms. Recent attempts, such as FastV [5], accelerate inference via layer-wise pruning of redundant visual tokens. ElasticCache [31], optimizes KV-Cache management during the generation phase. Despite these advances, head-wise acceleration strategies, particularly those targeting modality-specific attention heads, remain underexplored. In this work, we address this gap by proposing a systematic framework based on our findings about visual heads, enabling efficient deployment of MLLMs in resource-constrained environments." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.627, + 0.416, + 0.645 + ], + "angle": 0, + "content": "3. Visual Heads are Sparse in MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.759 + ], + "angle": 0, + "content": "In this section, we present our exploration of head sparsity in multi-modal large language models. To start with, we provide the preliminaries on the relations from LLMs to visual instruction tuning. Then we describe our approach for identifying sparse visual heads in MLLMs in Sec. 3.2, then introduce the deployment of visual heads in model acceleration in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.789 + ], + "angle": 0, + "content": "3.1. What is Learned during Visual Instruct Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Extending a Large Language Model to a Multimodal Large Language Model is achieved by integrating a visual encoder \\( E \\), an adapter \\( H \\), and the LLM \\( p_{\\theta} \\). The original LLM is trained solely on textual tasks to model the distribution of text sequences as \\( p_{\\theta}(\\mathbf{x}) = \\prod_{i=1}^{N} p_{\\theta}(\\mathbf{x}_i | \\mathbf{x}_{ i \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.486, + 0.635 + ], + "angle": 0, + "content": "This strategy effectively reduces the computational complexity from \\( O(N^2) \\) to \\( O(N \\times L) \\), where \\( L = 32 \\), thereby decreasing the runtime during the prefilling stage. To evaluate the attention for keys outside the local window, we compute the average attention weight:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.645, + 0.484, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\bar {A} _ {j} = \\frac {1}{w} \\sum_ {i = 1} ^ {w} \\hat {A} _ {i, j}, \\quad \\text {f o r} j \\in \\{0, \\dots , L - w - 1 \\} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.694, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Ultimately, we select the top \\( K \\) KV Caches based on the computed attention scores, where the value of \\( K \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.244, + 0.737, + 0.484, + 0.755 + ], + "angle": 0, + "content": "\\[\nK = r + b _ {i j} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Our three-part allocation mechanism leverages head sparsity to significantly reduce the computational and memory overhead in multi-modal models. It ensures that each head receives a guaranteed minimum cache allocation through both the nearest neighbor and uniform baseline allocations. The remaining cache is then adaptively distributed based on the visual attention scores, thereby achieving an efficient balance between computational efficiency and overall model performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.12, + 0.908, + 0.242 + ], + "angle": 0, + "content": "We conduct extensive experiments to validate the effectiveness of Visual Head. We first introduce our experimental settings in Section 4.1. Then we present a comparison with state-of-the-art KV Cache Compression method, demonstrating that our method maintains strong performance in Section 4.2 while maintaining computational efficiency in Section 4.3. Moreover, we provide some analytical experiments to illustrate the importance of visual head in Section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.265, + 0.722, + 0.281 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.909, + 0.477 + ], + "angle": 0, + "content": "Models. We employ three multi-modal models: LLaVA-NeXT-Vicuna-7B [27], LLaVA-NeXT-Mistral-7B [27], and Qwen2-VL-7B [41]. LLaVA-NeXT-Vicuna-7B is derived from Vicuna-7B [8], a model based on Multi-Head Attention (MHA) [49], and comprises 32 layers with 32 attention heads per layer. In contrast, LLaVA-NeXT-Mistral-7B is built upon Mistral-7B [17] and utilizes Grouped-Query Attention (GQA) [2]. This model features 32 layers, with each layer consisting of 32 query heads and 8 key-value heads. Similarly, Qwen2-VL-7B [41] is based on Qwen2, another GQA model, and is composed of 28 layers, with each layer containing 28 query heads and 4 key-value heads." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Baselines. We adopt SnapKV [22], PyramidKV [4], and AdaKV [13] as our baseline methods, as they represent the latest and state-of-the-art in KV Cache compression. SnapKV [22] utilizes an \"observation window\" mechanism to identify and preserve the most critical KV caches. PyramidKV [4] implements a hierarchical allocation strategy that distributes the KV cache budget in a pyramidal manner. More budget is allocated to lower layers with dispersed attention, and less to higher layers with focused patterns. Meanwhile, AdaKV [13] proposes a dynamic allocation framework that assigns varying cache budgets to different attention heads within a layer, based on the intra-layer attention distributions. In addition, to validate the Visual Head's effectiveness, we compare it with a Random Head baseline, where head scores are randomly initialized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "**Benchmarks.** To comprehensively assess the effectiveness of Visual Head in visual perception, we conduct evaluations on five widely used benchmarks covering both visual question answering (VQA) and image captioning tasks. Specifically, we utilize DocVQA [35], OCRBench [29], TextVQA [44], ChartQA [34], and TextCaps [43], which collectively encompass a diverse set of challenges, including document understanding, OCR-based question answering, chart interpretation, and text-based image captioning. Additionally, we also select the mainstream multiple-choice visual benchmarks, including MMBench [28] and VQAv2 [15] for a comprehensive evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.295, + 0.091, + 0.711, + 0.105 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.112, + 0.266, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.189, + 0.265, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.262, + 0.265, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.113, + 0.425, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.189, + 0.425, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.256, + 0.425, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.113, + 0.584, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.189, + 0.584, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.256, + 0.584, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.113, + 0.744, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.189, + 0.744, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.262, + 0.744, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.113, + 0.904, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.189, + 0.904, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.262, + 0.904, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.383 + ], + "angle": 0, + "content": "Figure 4. Main Results on Multi-Modal Benchmarks. We evaluate SparseMM and other baselines on several multimodal benchmarks, and conduct experiments on a series of backbones. Our SparseMM consistently outperforms the other baselines." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.394, + 0.486, + 0.464 + ], + "angle": 0, + "content": "Table 1. Average Number of Input Tokens. We analyzed the average length of input tokens across various benchmarks. Considering that text instructions are typically very short(fewer than 50 tokens), so visual tokens constitute the majority of the input sequence, accounting for \\(90\\%\\) to \\(99\\%\\) of the input tokens." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.474, + 0.482, + 0.517 + ], + "angle": 0, + "content": "
DatasetDocVQAOCRBenchTextVQAChartQATextCaps
LLaVA-Series24331700237622702376
Qwen2-VL-7B-Instruct4830124510246421024
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.411, + 0.563 + ], + "angle": 0, + "content": "4.2. Results on Multi-Modal Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.577, + 0.486, + 0.773 + ], + "angle": 0, + "content": "Setup. To determine an appropriate budget allocation, we first measure the average length of input tokens for each benchmark, as reported in Table 1. Given that text instructions typically consist of no more than 50 tokens, the majority of input tokens are attributed to visual tokens. Considering the varying input sequence lengths across different datasets, we select a range of each head's KV Cache budget for evaluation: \\{64, 128, 256, 512, 1024, 2048\\}. Since general visual benchmarks utilize lower image resolutions, we adjust the input token budget range correspondingly: \\{48, 64, 96, 128, 256, 512\\}. This allows us to systematically analyze the impact of different cache sizes on performance and efficiency across various benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Results. Fig. 4 presents the evaluation results for three models and five benchmarks. Our experimental results demonstrate that our proposed method consistently outperforms baseline approaches, particularly under extreme cache budget constraints (e.g., 128 or 256). Under these conditions, our approach maintains performance levels close to those achieved with full cache utilization, significantly outperforming the competing baselines. For instance, on the" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.395, + 0.907, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.521, + 0.909, + 0.577 + ], + "angle": 0, + "content": "Figure 5. Results on Multiple-choice Benchmarks. We evaluate SparseMM and other baselines on multiple-choice visual benchmarks with Qwen2-VL-7B-Instruct as the backbone model. Our SparseMM consistently outperforms the other baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.823 + ], + "angle": 0, + "content": "TextVQA [44] task using LLaVA-NeXT-Vicuna-7B, a KV Cache budget of 256—which constitutes only approximately \\(10.77\\%\\) of the average 2376 tokens—yields performance equivalent to the full-cache model, whereas AdaKV [13] and similar methods experience an accuracy drop of roughly \\(3\\%\\). Similarly, on OCRBench [29], LLaVA-NeXT-Mistral-7B demonstrates only a slight performance degradation at a KV Cache budget of 128 (about \\(7.5\\%\\) of the average 1700 tokens), in contrast to a decline exceeding \\(10\\%\\) observed with other methods. In addition, Qwen2-VL-7B-Instruct on DocVQA [35] maintains performance when operating with a KV Cache budget of 256 (merely \\(5.3\\%\\) of the average 4830 tokens), while alternative approaches suffer performance drops between \\(5\\%\\) and \\(17\\%\\). These results validate the effectiveness of our method in VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.91, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 presents the evaluation results on multiple-choice benchmarks. Our method demonstrates competitive performance on multi-choice benchmarks compared with existing baselines. For instance, with only 96 token budget, our approach retains full performance on MMBench while ex" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "periencing only a minimal performance degradation (\\(< 1\\%\\)) on GQA and VQAv2. These findings substantiate that our method can effectively recognize visual content while exhibiting strong generalizability towards diverse tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.158, + 0.487, + 0.371 + ], + "angle": 0, + "content": "Furthermore, our findings in Fig. 4 indicate that the random head baseline consistently produces the poorest performance across nearly all experiments, whereas our method achieves superior outcomes by utilizing Visual Head. This pronounced contrast underscores the efficacy of our approach in accurately capturing the manner in which multimodal language models attend to visual information. It is important to note that the performance of the random head method is comparable to that of SnapKV [22], particularly in the case of the MHA model. This similarity is attributable to the fact that when the scores of all heads are randomly initialized, the cache budget allocated to each head is statistically equivalent, effectively causing the method to revert to the behavior observed with SnapKV [22]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.397, + 0.293, + 0.414 + ], + "angle": 0, + "content": "4.3. Efficiency Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.486, + 0.566 + ], + "angle": 0, + "content": "Setup In this subsection, we evaluate the computational efficiency of our proposed method, which holds significant practical value. Accordingly, our efficiency tests are conducted across a range of input token lengths \\(\\{2\\mathrm{K}, 4\\mathrm{K}, 8\\mathrm{K}, 16\\mathrm{K}, 32\\mathrm{K}\\}\\). For each experiment, the output sequence length was fixed at 100 tokens, with a KV Cache budget set to 256. We computed the average decoding latency and peak memory consumption for each configuration. Notably, all experiments are done using FlashAttention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.487, + 0.71 + ], + "angle": 0, + "content": "Decoding Latency Fig. 6 illustrates that the reduction in KV Cache in our method substantially decreases the computational load during inference, thereby enhancing inference speed. For instance, when the input sequence length is \\(8\\mathrm{K}\\), the LLaVA-NeXT-Vicuna-7B model exhibits a speedup of \\(1.16\\times\\), while at a \\(32\\mathrm{K}\\) input length, the speedup increases to \\(1.87\\times\\). These findings indicate that our approach significantly accelerates token generation, particularly in high-resolution or long video contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Memory Cost Our method also offers a marked reduction in peak memory usage, primarily by diminishing the memory footprint associated with the KV Cache. This reduction is especially pronounced in LLaVA-Series models. For example, with an input sequence length of 32K, LLaVA-NeXT-Vicuna-7B with full KV Cache requires 32.87 GB of memory, whereas our method reduces the requirement to 17.38 GB, thereby achieving an approximate \\(50\\%\\) reduction in memory overhead. It is noteworthy that even for the Qwen2-VL-7B-Instruct model, which employs an aggressive compression technique in its GQA framework, we can still reduce the cache by nearly 2GB with 32k inputs." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.412, + 0.91, + 0.468 + ], + "angle": 0, + "content": "Figure 6. Efficiency Evaluation for SparseMM. Benefiting from the reduction in KV cache, SparseMM can maintain nearly constant decoding latency, achieving up to a \\(50\\%\\) acceleration. Additionally, it effectively reduces peak memory usage." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.617, + 0.504 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Performance Influence of Visual Heads. To further elucidate the impact of visual heads on the visual perception capabilities of multimodal models, we conducted a series of masking experiments. In these experiments, we selectively masked a specific proportion of the visual heads and, for comparison, randomly masked an equivalent proportion of attention heads. The evaluation was performed on OCR-Bench [29] and TextVQA [44], with performance measured relative to the baseline unmasked model. The results, as illustrated in Fig. 7, reveal that masking visual heads leads to a significant performance decline. In contrast, randomly masking the same proportion of attention heads produced a much smaller impact—for instance. These findings underscore the critical role that visual heads play in enabling MLLMs to effectively capture and process visual information. Moreover, masking the top \\(5\\%\\) of high-scoring visual heads causes a considerably greater performance loss than the additional impact of masking another \\(5\\%\\), which highlights the sparse yet indispensable distribution of visual heads." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Robustness of Visual Head Identification. To evaluate the robustness of our visual head detecting approach, we show the distribution of the detected visual heads under different datasets and tasks in Fig. 8 and show the accuracy curve in Fig. 9. For the OCR task, we used the Multi-Lingual Text(MLT) [36] and Chinese Text in the Wild(CTW) [54]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.483, + 0.259 + ], + "angle": 0, + "content": "Figure 7. Comparisons of Mmasking Visual Head and Random Head. The left figure is the result on OCRBench, and the right figure is the result on TextVQA." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.266, + 0.482, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.395, + 0.485, + 0.436 + ], + "angle": 0, + "content": "Figure 8. Visualizations of Visual Heads using Different Datasets. We visualized the attention distribution identified on MLT, CTW, and COCO datasets." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.44, + 0.482, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Figure 9. Results with Different Visual Head Identification Approaches and Datasets. We conduct an evaluation on visual heads identified on different datasets. The results on OCR datasets are similar and better than those on the detection dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.483, + 0.852 + ], + "angle": 0, + "content": "datasets. In addition, we consider the object detection task and choose the COCO dataset [25], where the model is required to identify objects present in the images. We then localized the visual heads based on the correspondence between the model's answers and the relevant objects. As shown, the distribution of visual heads is relatively consistent across the OCR datasets, whereas there is greater variation on the COCO dataset. Moreover, experimental results demonstrate that the visual heads identified from OCR tasks are dataset-agnostic and exhibit strong generalizability, with better results than detection tasks. This is because OCR tasks establish an exact one-to-one mapping between the model's output and the visual content, whereas the COCO task, which focuses on larger bounding boxes, introduces more noise and results in less robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Accuracy and Speed Trade-off. We compared the accuracy and speed of SparseMM with other baselines in Tab. 2. We conducted an experiment on LLaVA-NeXT-Vicuna-7B with" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.908, + 0.132 + ], + "angle": 0, + "content": "Table 2. Comparison of Accuracy-speed Trade-off among Different Methods. We compare the speed of all methods under 256 KV Cache budget and 16K input tokens." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.142, + 0.905, + 0.21 + ], + "angle": 0, + "content": "
MethodsDocVQAOCRBenchTextVQAChartQATextCapsLatency(ms)
FullKV0.680.520.650.550.7352.9
SparseMM0.68(-0.00)0.52(-0.00)0.65(-0.00)0.54(-0.01)0.73(-0.00)37.1(-30%)
SnapKV0.64(-0.04)0.46(-0.06)0.62(-0.03)0.50(-0.05)0.65(-0.08)35.3(-33%)
PyramidKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.53(-0.02)0.65(-0.08)34.9(-34%)
AdaKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.49(-0.06)0.66(-0.07)37.3(-29%)
" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.226, + 0.905, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.361, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 10. Visualizations of Visual Heads. We visualized the attention distribution of several heads. The visual heads are able to accurately capture text or objects within the images, whereas the non-visual heads provide random results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.908, + 0.522 + ], + "angle": 0, + "content": "a budget of 256 KV Cache. With the support of FlashAttention, our decoding latency is comparable to that of other methods, significantly lower than the FullKV method. However, our method outperforms others in terms of performance under the same budget. This effectively demonstrates the efficacy of SparseMM based on visual heads in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.909, + 0.678 + ], + "angle": 0, + "content": "Visualization of Visual Heads. To gain a more intuitive understanding of how visual heads process and interpret visual information, we conducted a visualization analysis of visual heads and non-visual heads on LLaVA-NeXT-Vicuna7B. As illustrated in Fig. 10, our observations indicate that non-visual heads often either neglect the image entirely or allocate a disproportionate amount of attention to visually insignificant regions. In contrast, visual heads accurately pinpoint regions of interest, allocating a substantial proportion of attention to these critical areas." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a systematic exploration of the visual processing characteristics inherent in MLLMs. Our analysis reveals a critical sparsity phenomenon that only a small fraction of attention heads actively engage in visual understanding. Leveraging this insight, we propose SparseMM, a novel KV-Cache optimization framework that dynamically allocates asymmetric computation budgets to attention heads based on their visual relevance. SparseMM prioritizes preserving vision-critical information during decoding, thereby achieving a more balanced accuracy-efficiency trade-off. We hope this study inspires deeper investigations into the principles governing multimodal learning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23184" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62336004, Grant 62125603, Grant 62321005, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.224 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.376, + 0.485, + 0.445 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[4] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information tunneling. arXiv preprint arXiv:2406.02069, 2024.3,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. Advances in Neural Information Processing Systems, 37:42566-42592, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.233, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.907, + 0.278 + ], + "angle": 0, + "content": "[12] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.906, + 0.336 + ], + "angle": 0, + "content": "[13] Yuan Feng, Junlin Lv, Yukun Cao, Xike Xie, and S Kevin Zhou. Ada-kv: Optimizing kv cache eviction by adaptive budget allocation for efficient llm inference. arXiv preprint arXiv:2407.11550, 2024. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.907, + 0.38 + ], + "angle": 0, + "content": "[14] GeminiTeam. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.908, + 0.51 + ], + "angle": 0, + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 36:20482-20494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.908, + 0.581 + ], + "angle": 0, + "content": "[17] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.907, + 0.655 + ], + "angle": 0, + "content": "[18] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7 (15):2, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.907, + 0.713 + ], + "angle": 0, + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.907, + 0.77 + ], + "angle": 0, + "content": "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-Interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation. Advances in Neural Information Processing Systems, 37:22947-22970, 2024. 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[23] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[28] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[30] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[31] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[32] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[33] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In 2019 International conference on document analysis and recognition (ICDAR), pages 1582–1587. IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.852, + 0.164 + ], + "angle": 0, + "content": "[37] OpenAI. Openai gpt-3.5 api. OpenAI API, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.166, + 0.9, + 0.18 + ], + "angle": 0, + "content": "[38] OpenAI. Gpt-4v(ison) system card. OpenAI Blog, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.182, + 0.881, + 0.196 + ], + "angle": 0, + "content": "[39] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.198, + 0.906, + 0.224 + ], + "angle": 0, + "content": "[40] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.226, + 0.906, + 0.254 + ], + "angle": 0, + "content": "[41] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.257, + 0.908, + 0.327 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.328, + 0.908, + 0.409 + ], + "angle": 0, + "content": "[43] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.908, + 0.483 + ], + "angle": 0, + "content": "[44] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.908, + 0.512 + ], + "angle": 0, + "content": "[45] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.754, + 0.528 + ], + "angle": 0, + "content": "[46] Qwen Team. Qwen2.5-v1, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.908, + 0.599 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.908, + 0.67 + ], + "angle": 0, + "content": "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.908, + 0.729 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.908, + 0.772 + ], + "angle": 0, + "content": "[50] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23186" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.174 + ], + "angle": 0, + "content": "[53] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.231 + ], + "angle": 0, + "content": "[54] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. Journal of Computer Science and Technology, 34(3):509-521, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[55] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[56] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23187" + } + ] +] \ No newline at end of file diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_origin.pdf b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..020982be6d14ec82fb0ffb01935b4f0c4d89d6ed --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/9ba0adfc-c24a-4716-85c4-644efaf54a4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eca9fac54829e1c1b9b5d62c5c591bc588504a10c25ae84f331c6bf1db5ff02b +size 4312210 diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/full.md b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4162b37fa960e8338a13b297821023a8dc383f09 --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/full.md @@ -0,0 +1,356 @@ +# SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs + +Jiahui Wang $^{1*}$ , Zuyan Liu $^{1,2*}$ , Yongming Rao $^{2,1}$ , Jiwen Lu $^{1\dagger}$ $^{1}$ Tsinghua University $^{2}$ Tencent Hunyuan X + +# Abstract + +Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than $5\%$ ) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparsity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers $1.38 \times$ real-time acceleration and $52\%$ memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM. + +# 1. Introduction + +Autoregressive large language models (LLMs) [6, 14, 37, 39, 45] have revolutionized artificial intelligence with their exceptional instruction-following capabilities and expansive knowledge repositories. Building upon this foundation, researchers have extended LLMs to multimodal domains, particularly in vision-language integration, creating multimodal large language models (MLLMs) [3, 7, 21, 26, 46, 53] that process both textual and visual inputs. Current approaches typically augment pre-trained LLMs by incorporating visual encoders (e.g., CLIP [42] or SigLIP [55]) paired with + +![](images/b63af9698478eebef7190b66803074286ef786d4969a68dfcc0abe71adfb790a.jpg) + +![](images/2106dd222bf1efb38776519850f029b2534a612d9a38e21c6e74cf306382324f.jpg) +Figure 1. Head Sparsity Emerges from Visual Concept Responses. We observe the visual-relevant heads are sparse in various MLLMs. Based on this observation, we devise a KV-Cache optimization strategy that allocates asymmetric budgets to LLM heads based on their importance for visual tokens, achieving better trade-off under limited computational resources. + +lightweight adapters to project visual features into the language model's hidden space. While these architectures demonstrate remarkable multimodal reasoning abilities, how LLMs fundamentally acquire visual comprehension during supervised fine-tuning remains poorly understood. This knowledge gap constrains our ability to recognize cross-modal alignment and risks undervaluing visual semantics during multi-modal relevant tasks and applications, which + +may potentially leading to suboptimal architecture designs and inefficient computational resource allocation. + +To this end, we present the first systematic investigation into how visual concepts are processed within LLMs. Through rigorous analysis of attention mechanisms, we uncover a critical phenomenon that only a small subset of attention heads (termed visual heads) drive visual content understanding, while the majority remain text-specialized. Specifically, our experiments reveal two critical properties of visual heads: (1) Sparsity: Less than $5\%$ of attention heads are intrinsically visual-active across layers, even in models trained with extensive multimodal data; (2) Universality: Visual heads emerge consistently across diverse LLM architectures(e.g., Vicuna [8] and Qwen2 [41]) and generalize to multiple attention paradigms such as multi-head attention (MHA) [49] and grouped query attention (GQA) [2]. + +To systematically identify these visual heads, we propose a training-free framework that quantifies the visual relevance of attention heads through targeted cross-modal response analysis. Specifically, our approach leverages OCR as an anchor task to establish precise correspondence between text outputs and visual inputs: for each generated word, we trace its activation back to spatially aligned image patches, enabling direct measurement of how specific attention heads mediate visual-text alignment. By analyzing and recoding the attention score of all the attention heads across a certain amount of samples, we compute visual scores that rank heads by their visual responsiveness. Crucially, while our identification mechanism relies on OCR's granular spatial grounding, we demonstrate that the detected visual heads exhibit task-agnostic generalizability—they remain dominant in diverse vision-language tasks including object recognition and scene understanding. + +Building on these insights, to demonstrate the effectiveness of visual heads on practical multi-modal tasks, we introduce SparseMM, a KV-Cache optimization framework that exploits visual head sparsity to achieve accelerated inference. As multimodal inputs grow in complexity—spanning multiturn dialogues [19, 20, 52], high-resolution interleaved images [9, 51], and dense video/3D sequences [12, 16, 23]—the computational overhead of maintaining full KV-Caches becomes prohibitive. Existing compression methods, however, treat all attention heads uniformly, disregarding the critical role of sparse visual heads in encoding visual semantics. + +SparseMM addresses this by asymmetrically allocating KV-Cache budgets: visual heads receive prioritized retention based on their precomputed visual scores, while nonvisual heads undergo aggressive compression via a hybrid strategy combining 1) Score-Preferred Cache (allocating cache budget based on visual head scores), 2) Uniform-Based Cache (preserving minimal budget for all the heads), and 3) Local Window Cache (preserving cache budget for recent tokens). This mixed approach ensures better accuracy + +efficiency trade-offs, such that visual heads retain more computational cost while other heads are dynamically throttled. + +Extensive experimental results demonstrate that SparseMM outperforms other strong baselines across multiple datasets, including DocVQA [35], OCRBench [29], TextVQA [44], MMBench [28], etc. For instance, on DocVQA, LLaVA-NeXT-Vicuna-7B [27] achieves the same level of accuracy while using only $20\%$ of the cache, and Qwen2-VL-7B-Instruct [41] achieves equivalent performance with just $5.3\%$ of the cache. These findings suggest that our method effectively captures visual information while compressing redundancies. Furthermore, the reduction in cache requirements enables our method to achieve lower decoding latency and peak memory usage. For example, LLaVA-NeXT-Mistral-7B [27] maintains nearly constant decoding latency with 32K input tokens, resulting in almost a $50\%$ acceleration compared to the full model, and reduces memory usage by 5GB. + +# 2. Related Works + +Architectures in MLLMs. The predominant architecture for Multi-Modal Large Language Models (MLLMs) consists of three key components: a visual encoder, an adapter, and a LLM. By leveraging alignment training techniques and subsequent fine-tuning, this integrated framework has achieved remarkable performance on various multi-modal understanding tasks [1, 10, 19, 24, 30, 32, 38]. In typical implementations, the visual encoder is realized using models like CLIP [42] or SigLIP [55], which are adept at extracting rich visual representations. The adapter component serves as an intermediary, bridging the gap between the visual features and the language domain; it is often instantiated as a multi-layer perceptron (MLP) or through more complex structures [33]. For the LLM part, architectures such as LLaMA [47, 48] and Qwen [40] series are commonly employed. Notably, previous LLMs such as LLaMA2 [48] utilizes a multi-head attention (MHA) mechanism, while more recent LLMs such as LLaMA3 [11] incorporate a grouped query attention (GQA) [2] design. The GQA approach aggregates multiple queries into a single group that corresponds to one key and one value, thereby substantially reducing memory usage without compromising model performance. In this paper, we observe a universal phenomenon of visual head in MLLMs across LLM architectures and attention implementations. The applications are demonstrated to be effective across the abundant MLLM series. + +Model Acceleration in MLLMs. With the rapid growth of model size and input sequence length, model acceleration has become an urgent research focus in both language and multimodal domains. In the context of Large Language Models (LLMs), significant efforts have been dedicated to optimizing the prompt encoding phase through efficient compression + +![](images/7f454ed261ce1b189df4b39f0ce7847335d8d89395b22cccaed64faa5c14ecf7.jpg) +Figure 2. Visual Heads are Sparse in MLLMs. We use OCR tasks to obtain visual scores for all heads. Upon visualizing these scores, we discovered that high-scoring heads, which we refer to as visual heads, are quite sparse within the MLLM, comprising only about $5\%$ . The majority of heads have very low scores, indicating that most heads in LLMs do not focus on visual information. + +of the KV-Cache. For instance, StreamingLLM [50] identifies attention sinks to stabilize long-context inference, while H2O [56] introduces a token-level importance scoring mechanism for adaptive KV-Cache eviction. Subsequent works, such as SnapKV [22], PyramidKV [4], and AdaKV [13], further refine the KV-Cache selection strategy by incorporating spatial-temporal redundancy reduction, hierarchical token retention, or dynamic eviction policies. However, these methods, primarily designed for text-only inputs, face limitations when applied to multi-modal scenarios. In MLLMs, acceleration challenges are exacerbated by the increasing complexity of multi-modal prompts (e.g., high-resolution images, videos) and cross-modal fusion mechanisms. Recent attempts, such as FastV [5], accelerate inference via layer-wise pruning of redundant visual tokens. ElasticCache [31], optimizes KV-Cache management during the generation phase. Despite these advances, head-wise acceleration strategies, particularly those targeting modality-specific attention heads, remain underexplored. In this work, we address this gap by proposing a systematic framework based on our findings about visual heads, enabling efficient deployment of MLLMs in resource-constrained environments. + +# 3. Visual Heads are Sparse in MLLMs + +In this section, we present our exploration of head sparsity in multi-modal large language models. To start with, we provide the preliminaries on the relations from LLMs to visual instruction tuning. Then we describe our approach for identifying sparse visual heads in MLLMs in Sec. 3.2, then introduce the deployment of visual heads in model acceleration in Sec. 3.3. + +# 3.1. What is Learned during Visual Instruct Tuning + +Extending a Large Language Model to a Multimodal Large Language Model is achieved by integrating a visual encoder $E$ , an adapter $H$ , and the LLM $p_{\theta}$ . The original LLM is trained solely on textual tasks to model the distribution of text sequences as $p_{\theta}(\mathbf{x}) = \prod_{i=1}^{N} p_{\theta}(\mathbf{x}_i | \mathbf{x}_{ i \end{array} \right. \tag {12} +$$ + +This strategy effectively reduces the computational complexity from $O(N^2)$ to $O(N \times L)$ , where $L = 32$ , thereby decreasing the runtime during the prefilling stage. To evaluate the attention for keys outside the local window, we compute the average attention weight: + +$$ +\bar {A} _ {j} = \frac {1}{w} \sum_ {i = 1} ^ {w} \hat {A} _ {i, j}, \quad \text {f o r} j \in \{0, \dots , L - w - 1 \} \tag {13} +$$ + +Ultimately, we select the top $K$ KV Caches based on the computed attention scores, where the value of $K$ is given by + +$$ +K = r + b _ {i j} \tag {14} +$$ + +Our three-part allocation mechanism leverages head sparsity to significantly reduce the computational and memory overhead in multi-modal models. It ensures that each head receives a guaranteed minimum cache allocation through both the nearest neighbor and uniform baseline allocations. The remaining cache is then adaptively distributed based on the visual attention scores, thereby achieving an efficient balance between computational efficiency and overall model performance. + +# 4. Experiments + +We conduct extensive experiments to validate the effectiveness of Visual Head. We first introduce our experimental settings in Section 4.1. Then we present a comparison with state-of-the-art KV Cache Compression method, demonstrating that our method maintains strong performance in Section 4.2 while maintaining computational efficiency in Section 4.3. Moreover, we provide some analytical experiments to illustrate the importance of visual head in Section 4.4. + +# 4.1. Experimental Settings + +Models. We employ three multi-modal models: LLaVA-NeXT-Vicuna-7B [27], LLaVA-NeXT-Mistral-7B [27], and Qwen2-VL-7B [41]. LLaVA-NeXT-Vicuna-7B is derived from Vicuna-7B [8], a model based on Multi-Head Attention (MHA) [49], and comprises 32 layers with 32 attention heads per layer. In contrast, LLaVA-NeXT-Mistral-7B is built upon Mistral-7B [17] and utilizes Grouped-Query Attention (GQA) [2]. This model features 32 layers, with each layer consisting of 32 query heads and 8 key-value heads. Similarly, Qwen2-VL-7B [41] is based on Qwen2, another GQA model, and is composed of 28 layers, with each layer containing 28 query heads and 4 key-value heads. + +Baselines. We adopt SnapKV [22], PyramidKV [4], and AdaKV [13] as our baseline methods, as they represent the latest and state-of-the-art in KV Cache compression. SnapKV [22] utilizes an "observation window" mechanism to identify and preserve the most critical KV caches. PyramidKV [4] implements a hierarchical allocation strategy that distributes the KV cache budget in a pyramidal manner. More budget is allocated to lower layers with dispersed attention, and less to higher layers with focused patterns. Meanwhile, AdaKV [13] proposes a dynamic allocation framework that assigns varying cache budgets to different attention heads within a layer, based on the intra-layer attention distributions. In addition, to validate the Visual Head's effectiveness, we compare it with a Random Head baseline, where head scores are randomly initialized. + +**Benchmarks.** To comprehensively assess the effectiveness of Visual Head in visual perception, we conduct evaluations on five widely used benchmarks covering both visual question answering (VQA) and image captioning tasks. Specifically, we utilize DocVQA [35], OCRBench [29], TextVQA [44], ChartQA [34], and TextCaps [43], which collectively encompass a diverse set of challenges, including document understanding, OCR-based question answering, chart interpretation, and text-based image captioning. Additionally, we also select the mainstream multiple-choice visual benchmarks, including MMBench [28] and VQAv2 [15] for a comprehensive evaluation. + +![](images/fa47fa8748f9d80ddb92335408ccceb5d7c69f3425aed590f3712f956132697a.jpg) + +![](images/6573657b84bac5246ae74f7048fb5d2f6a3248ad5deb1f1e6a917f906cfef68f.jpg) + +![](images/89c8417768106a4c3cfb0f5d3ec0cbc47465b5064fb82293c9012c3ddff3a6d8.jpg) + +![](images/aeaf95f78375fc875f73175624aaff5b79060624deb03d9c167b55bd9bc0b149.jpg) +Figure 4. Main Results on Multi-Modal Benchmarks. We evaluate SparseMM and other baselines on several multimodal benchmarks, and conduct experiments on a series of backbones. Our SparseMM consistently outperforms the other baselines. + +![](images/5d5f14414c5ad7cdb902b9f4a9488a134a5883be30edbbbf9541b67d09225ad8.jpg) + +![](images/c4792824b23c298c8fa3dd550aa230fd83c39d6b1e871b2bbf1b81a696dd9445.jpg) + +![](images/a2d16304559653f36b0e331556a7a5288ca6a90352fb08d41b0d37dd104adeb8.jpg) + +![](images/980da7ecb011f07d969120469ab4c558b24f4e0409f42b37fbf479596804f4d9.jpg) + +![](images/eaa819719db344dd1e0579e5248f1f5fbd86de82f6b5cce5f89c7e5ff0634cca.jpg) + +![](images/1e359a35cc745a7b188bd6e2fa8ac435b8707102659408b805d7a8a0c740145f.jpg) + +![](images/f06e602eadebeebd2794b611a5940bb5b024e63d6aa89bc0eda09caad7b192cb.jpg) + +![](images/33d350ac77b28bb09814b2b519f6f3886b0678956bfe4da838fb5668c9028a4b.jpg) + +![](images/c752a07eaa6613efa074e05e3e7e4d0d6a61c119d7dc82f55a0098afbcbf67e1.jpg) + +![](images/da689e7f680bd9fc7b9ffad7ed7fcf8f67003e9b7b81dc058c8ba71455d74c92.jpg) + +![](images/044716afad92b715ce54869ff2309ed9b093a7f1e96d34aa47346459eab2b537.jpg) + +![](images/4a2e0d55b9daea412c215f6e19b966fed7edbbb6f47836ca862a169305e1304e.jpg) + +Table 1. Average Number of Input Tokens. We analyzed the average length of input tokens across various benchmarks. Considering that text instructions are typically very short(fewer than 50 tokens), so visual tokens constitute the majority of the input sequence, accounting for $90\%$ to $99\%$ of the input tokens. + +
DatasetDocVQAOCRBenchTextVQAChartQATextCaps
LLaVA-Series24331700237622702376
Qwen2-VL-7B-Instruct4830124510246421024
+ +# 4.2. Results on Multi-Modal Benchmarks + +Setup. To determine an appropriate budget allocation, we first measure the average length of input tokens for each benchmark, as reported in Table 1. Given that text instructions typically consist of no more than 50 tokens, the majority of input tokens are attributed to visual tokens. Considering the varying input sequence lengths across different datasets, we select a range of each head's KV Cache budget for evaluation: \{64, 128, 256, 512, 1024, 2048\}. Since general visual benchmarks utilize lower image resolutions, we adjust the input token budget range correspondingly: \{48, 64, 96, 128, 256, 512\}. This allows us to systematically analyze the impact of different cache sizes on performance and efficiency across various benchmarks. + +Results. Fig. 4 presents the evaluation results for three models and five benchmarks. Our experimental results demonstrate that our proposed method consistently outperforms baseline approaches, particularly under extreme cache budget constraints (e.g., 128 or 256). Under these conditions, our approach maintains performance levels close to those achieved with full cache utilization, significantly outperforming the competing baselines. For instance, on the + +![](images/e10d077dc1df307a8538d230ea6e958d18488cf6f13a2a1d90b4bcb5e5f62878.jpg) +Figure 5. Results on Multiple-choice Benchmarks. We evaluate SparseMM and other baselines on multiple-choice visual benchmarks with Qwen2-VL-7B-Instruct as the backbone model. Our SparseMM consistently outperforms the other baselines. + +TextVQA [44] task using LLaVA-NeXT-Vicuna-7B, a KV Cache budget of 256—which constitutes only approximately $10.77\%$ of the average 2376 tokens—yields performance equivalent to the full-cache model, whereas AdaKV [13] and similar methods experience an accuracy drop of roughly $3\%$ . Similarly, on OCRBench [29], LLaVA-NeXT-Mistral-7B demonstrates only a slight performance degradation at a KV Cache budget of 128 (about $7.5\%$ of the average 1700 tokens), in contrast to a decline exceeding $10\%$ observed with other methods. In addition, Qwen2-VL-7B-Instruct on DocVQA [35] maintains performance when operating with a KV Cache budget of 256 (merely $5.3\%$ of the average 4830 tokens), while alternative approaches suffer performance drops between $5\%$ and $17\%$ . These results validate the effectiveness of our method in VQA tasks. + +Fig. 5 presents the evaluation results on multiple-choice benchmarks. Our method demonstrates competitive performance on multi-choice benchmarks compared with existing baselines. For instance, with only 96 token budget, our approach retains full performance on MMBench while ex + +periencing only a minimal performance degradation ( $< 1\%$ ) on GQA and VQAv2. These findings substantiate that our method can effectively recognize visual content while exhibiting strong generalizability towards diverse tasks. + +Furthermore, our findings in Fig. 4 indicate that the random head baseline consistently produces the poorest performance across nearly all experiments, whereas our method achieves superior outcomes by utilizing Visual Head. This pronounced contrast underscores the efficacy of our approach in accurately capturing the manner in which multimodal language models attend to visual information. It is important to note that the performance of the random head method is comparable to that of SnapKV [22], particularly in the case of the MHA model. This similarity is attributable to the fact that when the scores of all heads are randomly initialized, the cache budget allocated to each head is statistically equivalent, effectively causing the method to revert to the behavior observed with SnapKV [22]. + +# 4.3. Efficiency Evaluation + +Setup In this subsection, we evaluate the computational efficiency of our proposed method, which holds significant practical value. Accordingly, our efficiency tests are conducted across a range of input token lengths $\{2\mathrm{K}, 4\mathrm{K}, 8\mathrm{K}, 16\mathrm{K}, 32\mathrm{K}\}$ . For each experiment, the output sequence length was fixed at 100 tokens, with a KV Cache budget set to 256. We computed the average decoding latency and peak memory consumption for each configuration. Notably, all experiments are done using FlashAttention. + +Decoding Latency Fig. 6 illustrates that the reduction in KV Cache in our method substantially decreases the computational load during inference, thereby enhancing inference speed. For instance, when the input sequence length is $8\mathrm{K}$ , the LLaVA-NeXT-Vicuna-7B model exhibits a speedup of $1.16\times$ , while at a $32\mathrm{K}$ input length, the speedup increases to $1.87\times$ . These findings indicate that our approach significantly accelerates token generation, particularly in high-resolution or long video contexts. + +Memory Cost Our method also offers a marked reduction in peak memory usage, primarily by diminishing the memory footprint associated with the KV Cache. This reduction is especially pronounced in LLaVA-Series models. For example, with an input sequence length of 32K, LLaVA-NeXT-Vicuna-7B with full KV Cache requires 32.87 GB of memory, whereas our method reduces the requirement to 17.38 GB, thereby achieving an approximate $50\%$ reduction in memory overhead. It is noteworthy that even for the Qwen2-VL-7B-Instruct model, which employs an aggressive compression technique in its GQA framework, we can still reduce the cache by nearly 2GB with 32k inputs. + +![](images/200338dabfd258d11d8a56382d7650fb1f9c66893b67a95584a6b413768a56f1.jpg) +Figure 6. Efficiency Evaluation for SparseMM. Benefiting from the reduction in KV cache, SparseMM can maintain nearly constant decoding latency, achieving up to a $50\%$ acceleration. Additionally, it effectively reduces peak memory usage. + +# 4.4. Analysis + +Performance Influence of Visual Heads. To further elucidate the impact of visual heads on the visual perception capabilities of multimodal models, we conducted a series of masking experiments. In these experiments, we selectively masked a specific proportion of the visual heads and, for comparison, randomly masked an equivalent proportion of attention heads. The evaluation was performed on OCR-Bench [29] and TextVQA [44], with performance measured relative to the baseline unmasked model. The results, as illustrated in Fig. 7, reveal that masking visual heads leads to a significant performance decline. In contrast, randomly masking the same proportion of attention heads produced a much smaller impact—for instance. These findings underscore the critical role that visual heads play in enabling MLLMs to effectively capture and process visual information. Moreover, masking the top $5\%$ of high-scoring visual heads causes a considerably greater performance loss than the additional impact of masking another $5\%$ , which highlights the sparse yet indispensable distribution of visual heads. + +Robustness of Visual Head Identification. To evaluate the robustness of our visual head detecting approach, we show the distribution of the detected visual heads under different datasets and tasks in Fig. 8 and show the accuracy curve in Fig. 9. For the OCR task, we used the Multi-Lingual Text(MLT) [36] and Chinese Text in the Wild(CTW) [54] + +![](images/d523c453f260dcbbb487f8afe152dc6adea0dc1fdd6c6670798d8a6b5730d869.jpg) +Figure 7. Comparisons of Mmasking Visual Head and Random Head. The left figure is the result on OCRBench, and the right figure is the result on TextVQA. + +![](images/e2a3b587fbb5035d10fe962ef6c5868e2217d23345bfd03bc3f7b7d204b182b5.jpg) +Figure 8. Visualizations of Visual Heads using Different Datasets. We visualized the attention distribution identified on MLT, CTW, and COCO datasets. + +![](images/048a77182e77273234aa4b2489a60e153279a26d29b3f760880f1b2c9e40b950.jpg) +Figure 9. Results with Different Visual Head Identification Approaches and Datasets. We conduct an evaluation on visual heads identified on different datasets. The results on OCR datasets are similar and better than those on the detection dataset. + +datasets. In addition, we consider the object detection task and choose the COCO dataset [25], where the model is required to identify objects present in the images. We then localized the visual heads based on the correspondence between the model's answers and the relevant objects. As shown, the distribution of visual heads is relatively consistent across the OCR datasets, whereas there is greater variation on the COCO dataset. Moreover, experimental results demonstrate that the visual heads identified from OCR tasks are dataset-agnostic and exhibit strong generalizability, with better results than detection tasks. This is because OCR tasks establish an exact one-to-one mapping between the model's output and the visual content, whereas the COCO task, which focuses on larger bounding boxes, introduces more noise and results in less robustness. + +Accuracy and Speed Trade-off. We compared the accuracy and speed of SparseMM with other baselines in Tab. 2. We conducted an experiment on LLaVA-NeXT-Vicuna-7B with + +Table 2. Comparison of Accuracy-speed Trade-off among Different Methods. We compare the speed of all methods under 256 KV Cache budget and 16K input tokens. + +
MethodsDocVQAOCRBenchTextVQAChartQATextCapsLatency(ms)
FullKV0.680.520.650.550.7352.9
SparseMM0.68(-0.00)0.52(-0.00)0.65(-0.00)0.54(-0.01)0.73(-0.00)37.1(-30%)
SnapKV0.64(-0.04)0.46(-0.06)0.62(-0.03)0.50(-0.05)0.65(-0.08)35.3(-33%)
PyramidKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.53(-0.02)0.65(-0.08)34.9(-34%)
AdaKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.49(-0.06)0.66(-0.07)37.3(-29%)
+ +![](images/e18760383c881c12aa28c733f874edb4411ddd624268b8bf6683ecb82532f0ba.jpg) +Figure 10. Visualizations of Visual Heads. We visualized the attention distribution of several heads. The visual heads are able to accurately capture text or objects within the images, whereas the non-visual heads provide random results. + +a budget of 256 KV Cache. With the support of FlashAttention, our decoding latency is comparable to that of other methods, significantly lower than the FullKV method. However, our method outperforms others in terms of performance under the same budget. This effectively demonstrates the efficacy of SparseMM based on visual heads in MLLMs. + +Visualization of Visual Heads. To gain a more intuitive understanding of how visual heads process and interpret visual information, we conducted a visualization analysis of visual heads and non-visual heads on LLaVA-NeXT-Vicuna7B. As illustrated in Fig. 10, our observations indicate that non-visual heads often either neglect the image entirely or allocate a disproportionate amount of attention to visually insignificant regions. In contrast, visual heads accurately pinpoint regions of interest, allocating a substantial proportion of attention to these critical areas. + +# 5. Conclusion + +In this paper, we present a systematic exploration of the visual processing characteristics inherent in MLLMs. Our analysis reveals a critical sparsity phenomenon that only a small fraction of attention heads actively engage in visual understanding. Leveraging this insight, we propose SparseMM, a novel KV-Cache optimization framework that dynamically allocates asymmetric computation budgets to attention heads based on their visual relevance. SparseMM prioritizes preserving vision-critical information during decoding, thereby achieving a more balanced accuracy-efficiency trade-off. We hope this study inspires deeper investigations into the principles governing multimodal learning. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China under Grant 62336004, Grant 62125603, Grant 62321005, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009. + +# References + +[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2 +[2] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 2, 5 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1 +[4] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information tunneling. arXiv preprint arXiv:2406.02069, 2024.3,5 +[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 3 +[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1 +[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 5 +[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. Advances in Neural Information Processing Systems, 37:42566-42592, 2024. 2 + +[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 2 +[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 +[12] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 2 +[13] Yuan Feng, Junlin Lv, Yukun Cao, Xike Xie, and S Kevin Zhou. Ada-kv: Optimizing kv cache eviction by adaptive budget allocation for efficient llm inference. arXiv preprint arXiv:2407.11550, 2024. 3, 4, 5, 6 +[14] GeminiTeam. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1 +[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5 +[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 36:20482-20494, 2023. 2 +[17] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 5 +[18] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7 (15):2, 2021. 4 +[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2 +[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-Interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 1 +[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation. Advances in Neural Information Processing Systems, 37:22947-22970, 2024. 3, 5, 7 + +[23] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2 +[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 8 +[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,5 +[28] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5 +[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 2, 5, 6, 7 +[30] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 2 +[31] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 3 +[32] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025. 2 +[33] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 2 +[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5 +[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6 +[36] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019 + +robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In 2019 International conference on document analysis and recognition (ICDAR), pages 1582–1587. IEEE, 2019. 7 +[37] OpenAI. Openai gpt-3.5 api. OpenAI API, 2023. 1 +[38] OpenAI. Gpt-4v(ison) system card. OpenAI Blog, 2023. 2 +[39] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1 +[40] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2 +[41] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 2, 5 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3 +[43] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 5 +[44] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 2, 5, 6, 7 +[45] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1 +[46] Qwen Team. Qwen2.5-v1, 2025. 1 +[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5 +[50] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3 +[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2 +[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2 + +[53] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 1 +[54] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. Journal of Computer Science and Technology, 34(3):509-521, 2019. 7 +[55] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1, 2, 3 +[56] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3 \ No newline at end of file diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/images.zip b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c872976b12f73d8db7d9b35b9465059489f4c8d --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6177af467955473c07f2e98e401a87cbb4c03b88f56e12a97b381df514ca05e1 +size 617630 diff --git a/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/layout.json b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..97920078c07c8b4e2c7158118d5be75ebbe7d7a5 --- /dev/null +++ b/2025/SparseMM_ Head Sparsity Emerges from Visual Concept Responses in MLLMs/layout.json @@ -0,0 +1,8868 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 552, + 123 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 123 + ], + "type": "text", + "content": "SparseMM: Head Sparsity Emerges from Visual Concept Responses in MLLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "spans": [ + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": "Jiahui Wang" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Zuyan Liu" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Yongming Rao" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": ", Jiwen Lu" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": " Tsinghua University " + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 159, + 142, + 451, + 173 + ], + "type": "text", + "content": " Tencent Hunyuan X" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "spans": [ + { + "bbox": [ + 151, + 199, + 200, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than " + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "text", + "content": ") of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparsity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers " + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "inline_equation", + "content": "1.38 \\times" + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "text", + "content": " real-time acceleration and " + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "inline_equation", + "content": "52\\%" + }, + { + "bbox": [ + 54, + 224, + 297, + 524 + ], + "type": "text", + "content": " memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "type": "text", + "content": "Autoregressive large language models (LLMs) [6, 14, 37, 39, 45] have revolutionized artificial intelligence with their exceptional instruction-following capabilities and expansive knowledge repositories. Building upon this foundation, researchers have extended LLMs to multimodal domains, particularly in vision-language integration, creating multimodal large language models (MLLMs) [3, 7, 21, 26, 46, 53] that process both textual and visual inputs. Current approaches typically augment pre-trained LLMs by incorporating visual encoders (e.g., CLIP [42] or SigLIP [55]) paired with" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 200, + 566, + 369 + ], + "blocks": [ + { + "bbox": [ + 318, + 200, + 566, + 369 + ], + "lines": [ + { + "bbox": [ + 318, + 200, + 566, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 200, + 566, + 369 + ], + "type": "image", + "image_path": "b63af9698478eebef7190b66803074286ef786d4969a68dfcc0abe71adfb790a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 371, + 564, + 521 + ], + "blocks": [ + { + "bbox": [ + 319, + 371, + 564, + 521 + ], + "lines": [ + { + "bbox": [ + 319, + 371, + 564, + 521 + ], + "spans": [ + { + "bbox": [ + 319, + 371, + 564, + 521 + ], + "type": "image", + "image_path": "2106dd222bf1efb38776519850f029b2534a612d9a38e21c6e74cf306382324f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 535, + 557, + 602 + ], + "lines": [ + { + "bbox": [ + 313, + 535, + 557, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 557, + 602 + ], + "type": "text", + "content": "Figure 1. Head Sparsity Emerges from Visual Concept Responses. We observe the visual-relevant heads are sparse in various MLLMs. Based on this observation, we devise a KV-Cache optimization strategy that allocates asymmetric budgets to LLM heads based on their importance for visual tokens, achieving better trade-off under limited computational resources." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "lightweight adapters to project visual features into the language model's hidden space. While these architectures demonstrate remarkable multimodal reasoning abilities, how LLMs fundamentally acquire visual comprehension during supervised fine-tuning remains poorly understood. This knowledge gap constrains our ability to recognize cross-modal alignment and risks undervaluing visual semantics during multi-modal relevant tasks and applications, which" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 294, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 294, + 714 + ], + "type": "text", + "content": "*Authors contributed equally to this research. † Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23177" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "may potentially leading to suboptimal architecture designs and inefficient computational resource allocation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "text", + "content": "To this end, we present the first systematic investigation into how visual concepts are processed within LLMs. Through rigorous analysis of attention mechanisms, we uncover a critical phenomenon that only a small subset of attention heads (termed visual heads) drive visual content understanding, while the majority remain text-specialized. Specifically, our experiments reveal two critical properties of visual heads: (1) Sparsity: Less than " + }, + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "text", + "content": " of attention heads are intrinsically visual-active across layers, even in models trained with extensive multimodal data; (2) Universality: Visual heads emerge consistently across diverse LLM architectures(e.g., Vicuna [8] and Qwen2 [41]) and generalize to multiple attention paradigms such as multi-head attention (MHA) [49] and grouped query attention (GQA) [2]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 267, + 297, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 297, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 297, + 471 + ], + "type": "text", + "content": "To systematically identify these visual heads, we propose a training-free framework that quantifies the visual relevance of attention heads through targeted cross-modal response analysis. Specifically, our approach leverages OCR as an anchor task to establish precise correspondence between text outputs and visual inputs: for each generated word, we trace its activation back to spatially aligned image patches, enabling direct measurement of how specific attention heads mediate visual-text alignment. By analyzing and recoding the attention score of all the attention heads across a certain amount of samples, we compute visual scores that rank heads by their visual responsiveness. Crucially, while our identification mechanism relies on OCR's granular spatial grounding, we demonstrate that the detected visual heads exhibit task-agnostic generalizability—they remain dominant in diverse vision-language tasks including object recognition and scene understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": "Building on these insights, to demonstrate the effectiveness of visual heads on practical multi-modal tasks, we introduce SparseMM, a KV-Cache optimization framework that exploits visual head sparsity to achieve accelerated inference. As multimodal inputs grow in complexity—spanning multiturn dialogues [19, 20, 52], high-resolution interleaved images [9, 51], and dense video/3D sequences [12, 16, 23]—the computational overhead of maintaining full KV-Caches becomes prohibitive. Existing compression methods, however, treat all attention heads uniformly, disregarding the critical role of sparse visual heads in encoding visual semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "SparseMM addresses this by asymmetrically allocating KV-Cache budgets: visual heads receive prioritized retention based on their precomputed visual scores, while nonvisual heads undergo aggressive compression via a hybrid strategy combining 1) Score-Preferred Cache (allocating cache budget based on visual head scores), 2) Uniform-Based Cache (preserving minimal budget for all the heads), and 3) Local Window Cache (preserving cache budget for recent tokens). This mixed approach ensures better accuracy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "efficiency trade-offs, such that visual heads retain more computational cost while other heads are dynamically throttled." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "text", + "content": "Extensive experimental results demonstrate that SparseMM outperforms other strong baselines across multiple datasets, including DocVQA [35], OCRBench [29], TextVQA [44], MMBench [28], etc. For instance, on DocVQA, LLaVA-NeXT-Vicuna-7B [27] achieves the same level of accuracy while using only " + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "text", + "content": " of the cache, and Qwen2-VL-7B-Instruct [41] achieves equivalent performance with just " + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "text", + "content": " of the cache. These findings suggest that our method effectively captures visual information while compressing redundancies. Furthermore, the reduction in cache requirements enables our method to achieve lower decoding latency and peak memory usage. For example, LLaVA-NeXT-Mistral-7B [27] maintains nearly constant decoding latency with 32K input tokens, resulting in almost a " + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 98, + 556, + 289 + ], + "type": "text", + "content": " acceleration compared to the full model, and reduces memory usage by 5GB." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 302, + 406, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 406, + 315 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 406, + 315 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 556, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 637 + ], + "type": "text", + "content": "Architectures in MLLMs. The predominant architecture for Multi-Modal Large Language Models (MLLMs) consists of three key components: a visual encoder, an adapter, and a LLM. By leveraging alignment training techniques and subsequent fine-tuning, this integrated framework has achieved remarkable performance on various multi-modal understanding tasks [1, 10, 19, 24, 30, 32, 38]. In typical implementations, the visual encoder is realized using models like CLIP [42] or SigLIP [55], which are adept at extracting rich visual representations. The adapter component serves as an intermediary, bridging the gap between the visual features and the language domain; it is often instantiated as a multi-layer perceptron (MLP) or through more complex structures [33]. For the LLM part, architectures such as LLaMA [47, 48] and Qwen [40] series are commonly employed. Notably, previous LLMs such as LLaMA2 [48] utilizes a multi-head attention (MHA) mechanism, while more recent LLMs such as LLaMA3 [11] incorporate a grouped query attention (GQA) [2] design. The GQA approach aggregates multiple queries into a single group that corresponds to one key and one value, thereby substantially reducing memory usage without compromising model performance. In this paper, we observe a universal phenomenon of visual head in MLLMs across LLM architectures and attention implementations. The applications are demonstrated to be effective across the abundant MLLM series." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Model Acceleration in MLLMs. With the rapid growth of model size and input sequence length, model acceleration has become an urgent research focus in both language and multimodal domains. In the context of Large Language Models (LLMs), significant efforts have been dedicated to optimizing the prompt encoding phase through efficient compression" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23178" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 167 + ], + "type": "image", + "image_path": "7f454ed261ce1b189df4b39f0ce7847335d8d89395b22cccaed64faa5c14ecf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 175, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 54, + 175, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 555, + 209 + ], + "type": "text", + "content": "Figure 2. Visual Heads are Sparse in MLLMs. We use OCR tasks to obtain visual scores for all heads. Upon visualizing these scores, we discovered that high-scoring heads, which we refer to as visual heads, are quite sparse within the MLLM, comprising only about " + }, + { + "bbox": [ + 54, + 175, + 555, + 209 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 175, + 555, + 209 + ], + "type": "text", + "content": ". The majority of heads have very low scores, indicating that most heads in LLMs do not focus on visual information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 483 + ], + "type": "text", + "content": "of the KV-Cache. For instance, StreamingLLM [50] identifies attention sinks to stabilize long-context inference, while H2O [56] introduces a token-level importance scoring mechanism for adaptive KV-Cache eviction. Subsequent works, such as SnapKV [22], PyramidKV [4], and AdaKV [13], further refine the KV-Cache selection strategy by incorporating spatial-temporal redundancy reduction, hierarchical token retention, or dynamic eviction policies. However, these methods, primarily designed for text-only inputs, face limitations when applied to multi-modal scenarios. In MLLMs, acceleration challenges are exacerbated by the increasing complexity of multi-modal prompts (e.g., high-resolution images, videos) and cross-modal fusion mechanisms. Recent attempts, such as FastV [5], accelerate inference via layer-wise pruning of redundant visual tokens. ElasticCache [31], optimizes KV-Cache management during the generation phase. Despite these advances, head-wise acceleration strategies, particularly those targeting modality-specific attention heads, remain underexplored. In this work, we address this gap by proposing a systematic framework based on our findings about visual heads, enabling efficient deployment of MLLMs in resource-constrained environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 254, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 254, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 254, + 510 + ], + "type": "text", + "content": "3. Visual Heads are Sparse in MLLMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "content": "In this section, we present our exploration of head sparsity in multi-modal large language models. To start with, we provide the preliminaries on the relations from LLMs to visual instruction tuning. Then we describe our approach for identifying sparse visual heads in MLLMs in Sec. 3.2, then introduce the deployment of visual heads in model acceleration in Sec. 3.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 296, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 624 + ], + "type": "text", + "content": "3.1. What is Learned during Visual Instruct Tuning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Extending a Large Language Model to a Multimodal Large Language Model is achieved by integrating a visual encoder " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", an adapter " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", and the LLM " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". The original LLM is trained solely on textual tasks to model the distribution of text sequences as " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}) = \\prod_{i=1}^{N} p_{\\theta}(\\mathbf{x}_i | \\mathbf{x}_{ i \\end{array} \\right. \\tag {12}", + "image_path": "ef1de37265030407592bccd6b641cd314a5f2eee54ab0af05be82b930631c5e0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "text", + "content": "This strategy effectively reduces the computational complexity from " + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "inline_equation", + "content": "O(N^2)" + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "inline_equation", + "content": "O(N \\times L)" + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "inline_equation", + "content": "L = 32" + }, + { + "bbox": [ + 55, + 441, + 297, + 502 + ], + "type": "text", + "content": ", thereby decreasing the runtime during the prefilling stage. To evaluate the attention for keys outside the local window, we compute the average attention weight:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 510, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 510, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\bar {A} _ {j} = \\frac {1}{w} \\sum_ {i = 1} ^ {w} \\hat {A} _ {i, j}, \\quad \\text {f o r} j \\in \\{0, \\dots , L - w - 1 \\} \\tag {13}", + "image_path": "2c83f664016fc3e621589d0cf7032ec3d83b3c542a36edb63e5bd58e845a8f83.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "text", + "content": "Ultimately, we select the top " + }, + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "text", + "content": " KV Caches based on the computed attention scores, where the value of " + }, + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 549, + 296, + 574 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 149, + 583, + 296, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 583, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 149, + 583, + 296, + 597 + ], + "type": "interline_equation", + "content": "K = r + b _ {i j} \\tag {14}", + "image_path": "e44b9c4abba4fc6b2d1ecf0a1ce18140c869105353483438d70257d732c1b8ed.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "Our three-part allocation mechanism leverages head sparsity to significantly reduce the computational and memory overhead in multi-modal models. It ensures that each head receives a guaranteed minimum cache allocation through both the nearest neighbor and uniform baseline allocations. The remaining cache is then adaptively distributed based on the visual attention scores, thereby achieving an efficient balance between computational efficiency and overall model performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 95, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 555, + 191 + ], + "type": "text", + "content": "We conduct extensive experiments to validate the effectiveness of Visual Head. We first introduce our experimental settings in Section 4.1. Then we present a comparison with state-of-the-art KV Cache Compression method, demonstrating that our method maintains strong performance in Section 4.2 while maintaining computational efficiency in Section 4.3. Moreover, we provide some analytical experiments to illustrate the importance of visual head in Section 4.4." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 209, + 441, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 441, + 222 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 441, + 222 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 233, + 556, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 556, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 556, + 377 + ], + "type": "text", + "content": "Models. We employ three multi-modal models: LLaVA-NeXT-Vicuna-7B [27], LLaVA-NeXT-Mistral-7B [27], and Qwen2-VL-7B [41]. LLaVA-NeXT-Vicuna-7B is derived from Vicuna-7B [8], a model based on Multi-Head Attention (MHA) [49], and comprises 32 layers with 32 attention heads per layer. In contrast, LLaVA-NeXT-Mistral-7B is built upon Mistral-7B [17] and utilizes Grouped-Query Attention (GQA) [2]. This model features 32 layers, with each layer consisting of 32 query heads and 8 key-value heads. Similarly, Qwen2-VL-7B [41] is based on Qwen2, another GQA model, and is composed of 28 layers, with each layer containing 28 query heads and 4 key-value heads." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 564 + ], + "type": "text", + "content": "Baselines. We adopt SnapKV [22], PyramidKV [4], and AdaKV [13] as our baseline methods, as they represent the latest and state-of-the-art in KV Cache compression. SnapKV [22] utilizes an \"observation window\" mechanism to identify and preserve the most critical KV caches. PyramidKV [4] implements a hierarchical allocation strategy that distributes the KV cache budget in a pyramidal manner. More budget is allocated to lower layers with dispersed attention, and less to higher layers with focused patterns. Meanwhile, AdaKV [13] proposes a dynamic allocation framework that assigns varying cache budgets to different attention heads within a layer, based on the intra-layer attention distributions. In addition, to validate the Visual Head's effectiveness, we compare it with a Random Head baseline, where head scores are randomly initialized." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": "**Benchmarks.** To comprehensively assess the effectiveness of Visual Head in visual perception, we conduct evaluations on five widely used benchmarks covering both visual question answering (VQA) and image captioning tasks. Specifically, we utilize DocVQA [35], OCRBench [29], TextVQA [44], ChartQA [34], and TextCaps [43], which collectively encompass a diverse set of challenges, including document understanding, OCR-based question answering, chart interpretation, and text-based image captioning. Additionally, we also select the mainstream multiple-choice visual benchmarks, including MMBench [28] and VQAv2 [15] for a comprehensive evaluation." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23181" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 180, + 72, + 435, + 83 + ], + "blocks": [ + { + "bbox": [ + 180, + 72, + 435, + 83 + ], + "lines": [ + { + "bbox": [ + 180, + 72, + 435, + 83 + ], + "spans": [ + { + "bbox": [ + 180, + 72, + 435, + 83 + ], + "type": "image", + "image_path": "fa47fa8748f9d80ddb92335408ccceb5d7c69f3425aed590f3712f956132697a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 88, + 162, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 88, + 162, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 162, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 162, + 149 + ], + "type": "image", + "image_path": "6573657b84bac5246ae74f7048fb5d2f6a3248ad5deb1f1e6a917f906cfef68f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 149, + 162, + 206 + ], + "blocks": [ + { + "bbox": [ + 56, + 149, + 162, + 206 + ], + "lines": [ + { + "bbox": [ + 56, + 149, + 162, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 162, + 206 + ], + "type": "image", + "image_path": "89c8417768106a4c3cfb0f5d3ec0cbc47465b5064fb82293c9012c3ddff3a6d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 207, + 162, + 267 + ], + "blocks": [ + { + "bbox": [ + 56, + 207, + 162, + 267 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 162, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 162, + 267 + ], + "type": "image", + "image_path": "aeaf95f78375fc875f73175624aaff5b79060624deb03d9c167b55bd9bc0b149.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 303 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 303 + ], + "type": "text", + "content": "Figure 4. Main Results on Multi-Modal Benchmarks. We evaluate SparseMM and other baselines on several multimodal benchmarks, and conduct experiments on a series of backbones. Our SparseMM consistently outperforms the other baselines." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 162, + 89, + 260, + 149 + ], + "blocks": [ + { + "bbox": [ + 162, + 89, + 260, + 149 + ], + "lines": [ + { + "bbox": [ + 162, + 89, + 260, + 149 + ], + "spans": [ + { + "bbox": [ + 162, + 89, + 260, + 149 + ], + "type": "image", + "image_path": "5d5f14414c5ad7cdb902b9f4a9488a134a5883be30edbbbf9541b67d09225ad8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 162, + 149, + 260, + 202 + ], + "blocks": [ + { + "bbox": [ + 162, + 149, + 260, + 202 + ], + "lines": [ + { + "bbox": [ + 162, + 149, + 260, + 202 + ], + "spans": [ + { + "bbox": [ + 162, + 149, + 260, + 202 + ], + "type": "image", + "image_path": "c4792824b23c298c8fa3dd550aa230fd83c39d6b1e871b2bbf1b81a696dd9445.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 162, + 202, + 260, + 266 + ], + "blocks": [ + { + "bbox": [ + 162, + 202, + 260, + 266 + ], + "lines": [ + { + "bbox": [ + 162, + 202, + 260, + 266 + ], + "spans": [ + { + "bbox": [ + 162, + 202, + 260, + 266 + ], + "type": "image", + "image_path": "a2d16304559653f36b0e331556a7a5288ca6a90352fb08d41b0d37dd104adeb8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 89, + 357, + 149 + ], + "blocks": [ + { + "bbox": [ + 260, + 89, + 357, + 149 + ], + "lines": [ + { + "bbox": [ + 260, + 89, + 357, + 149 + ], + "spans": [ + { + "bbox": [ + 260, + 89, + 357, + 149 + ], + "type": "image", + "image_path": "980da7ecb011f07d969120469ab4c558b24f4e0409f42b37fbf479596804f4d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 260, + 149, + 357, + 202 + ], + "blocks": [ + { + "bbox": [ + 260, + 149, + 357, + 202 + ], + "lines": [ + { + "bbox": [ + 260, + 149, + 357, + 202 + ], + "spans": [ + { + "bbox": [ + 260, + 149, + 357, + 202 + ], + "type": "image", + "image_path": "eaa819719db344dd1e0579e5248f1f5fbd86de82f6b5cce5f89c7e5ff0634cca.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 260, + 202, + 357, + 266 + ], + "blocks": [ + { + "bbox": [ + 260, + 202, + 357, + 266 + ], + "lines": [ + { + "bbox": [ + 260, + 202, + 357, + 266 + ], + "spans": [ + { + "bbox": [ + 260, + 202, + 357, + 266 + ], + "type": "image", + "image_path": "1e359a35cc745a7b188bd6e2fa8ac435b8707102659408b805d7a8a0c740145f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 357, + 89, + 455, + 149 + ], + "blocks": [ + { + "bbox": [ + 357, + 89, + 455, + 149 + ], + "lines": [ + { + "bbox": [ + 357, + 89, + 455, + 149 + ], + "spans": [ + { + "bbox": [ + 357, + 89, + 455, + 149 + ], + "type": "image", + "image_path": "f06e602eadebeebd2794b611a5940bb5b024e63d6aa89bc0eda09caad7b192cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 357, + 149, + 455, + 206 + ], + "blocks": [ + { + "bbox": [ + 357, + 149, + 455, + 206 + ], + "lines": [ + { + "bbox": [ + 357, + 149, + 455, + 206 + ], + "spans": [ + { + "bbox": [ + 357, + 149, + 455, + 206 + ], + "type": "image", + "image_path": "33d350ac77b28bb09814b2b519f6f3886b0678956bfe4da838fb5668c9028a4b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 357, + 207, + 455, + 266 + ], + "blocks": [ + { + "bbox": [ + 357, + 207, + 455, + 266 + ], + "lines": [ + { + "bbox": [ + 357, + 207, + 455, + 266 + ], + "spans": [ + { + "bbox": [ + 357, + 207, + 455, + 266 + ], + "type": "image", + "image_path": "c752a07eaa6613efa074e05e3e7e4d0d6a61c119d7dc82f55a0098afbcbf67e1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 455, + 89, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 455, + 89, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 455, + 89, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 455, + 89, + 553, + 149 + ], + "type": "image", + "image_path": "da689e7f680bd9fc7b9ffad7ed7fcf8f67003e9b7b81dc058c8ba71455d74c92.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 455, + 149, + 553, + 206 + ], + "blocks": [ + { + "bbox": [ + 455, + 149, + 553, + 206 + ], + "lines": [ + { + "bbox": [ + 455, + 149, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 455, + 149, + 553, + 206 + ], + "type": "image", + "image_path": "044716afad92b715ce54869ff2309ed9b093a7f1e96d34aa47346459eab2b537.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 455, + 207, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 455, + 207, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 455, + 207, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 455, + 207, + 553, + 266 + ], + "type": "image", + "image_path": "4a2e0d55b9daea412c215f6e19b966fed7edbbb6f47836ca862a169305e1304e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 57, + 375, + 294, + 409 + ], + "blocks": [ + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "type": "text", + "content": "Table 1. Average Number of Input Tokens. We analyzed the average length of input tokens across various benchmarks. Considering that text instructions are typically very short(fewer than 50 tokens), so visual tokens constitute the majority of the input sequence, accounting for " + }, + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 55, + 312, + 297, + 367 + ], + "type": "text", + "content": " of the input tokens." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 375, + 294, + 409 + ], + "lines": [ + { + "bbox": [ + 57, + 375, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 294, + 409 + ], + "type": "table", + "html": "
DatasetDocVQAOCRBenchTextVQAChartQATextCaps
LLaVA-Series24331700237622702376
Qwen2-VL-7B-Instruct4830124510246421024
", + "image_path": "0191b3d54c24619896de8093ef5c1e5d594fad8891f1cccd5aa20fa2365021cd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 251, + 445 + ], + "type": "text", + "content": "4.2. Results on Multi-Modal Benchmarks" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 456, + 297, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 297, + 612 + ], + "type": "text", + "content": "Setup. To determine an appropriate budget allocation, we first measure the average length of input tokens for each benchmark, as reported in Table 1. Given that text instructions typically consist of no more than 50 tokens, the majority of input tokens are attributed to visual tokens. Considering the varying input sequence lengths across different datasets, we select a range of each head's KV Cache budget for evaluation: \\{64, 128, 256, 512, 1024, 2048\\}. Since general visual benchmarks utilize lower image resolutions, we adjust the input token budget range correspondingly: \\{48, 64, 96, 128, 256, 512\\}. This allows us to systematically analyze the impact of different cache sizes on performance and efficiency across various benchmarks." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Results. Fig. 4 presents the evaluation results for three models and five benchmarks. Our experimental results demonstrate that our proposed method consistently outperforms baseline approaches, particularly under extreme cache budget constraints (e.g., 128 or 256). Under these conditions, our approach maintains performance levels close to those achieved with full cache utilization, significantly outperforming the competing baselines. For instance, on the" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 312, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 315, + 312, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 315, + 312, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 312, + 555, + 407 + ], + "type": "image", + "image_path": "e10d077dc1df307a8538d230ea6e958d18488cf6f13a2a1d90b4bcb5e5f62878.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 412, + 556, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 456 + ], + "type": "text", + "content": "Figure 5. Results on Multiple-choice Benchmarks. We evaluate SparseMM and other baselines on multiple-choice visual benchmarks with Qwen2-VL-7B-Instruct as the backbone model. Our SparseMM consistently outperforms the other baselines." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": "TextVQA [44] task using LLaVA-NeXT-Vicuna-7B, a KV Cache budget of 256—which constitutes only approximately " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "10.77\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": " of the average 2376 tokens—yields performance equivalent to the full-cache model, whereas AdaKV [13] and similar methods experience an accuracy drop of roughly " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": ". Similarly, on OCRBench [29], LLaVA-NeXT-Mistral-7B demonstrates only a slight performance degradation at a KV Cache budget of 128 (about " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": " of the average 1700 tokens), in contrast to a decline exceeding " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": " observed with other methods. In addition, Qwen2-VL-7B-Instruct on DocVQA [35] maintains performance when operating with a KV Cache budget of 256 (merely " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": " of the average 4830 tokens), while alternative approaches suffer performance drops between " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 651 + ], + "type": "text", + "content": ". These results validate the effectiveness of our method in VQA tasks." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Fig. 5 presents the evaluation results on multiple-choice benchmarks. Our method demonstrates competitive performance on multi-choice benchmarks compared with existing baselines. For instance, with only 96 token budget, our approach retains full performance on MMBench while ex" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23182" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "periencing only a minimal performance degradation (" + }, + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": ") on GQA and VQAv2. These findings substantiate that our method can effectively recognize visual content while exhibiting strong generalizability towards diverse tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 125, + 298, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 298, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 298, + 293 + ], + "type": "text", + "content": "Furthermore, our findings in Fig. 4 indicate that the random head baseline consistently produces the poorest performance across nearly all experiments, whereas our method achieves superior outcomes by utilizing Visual Head. This pronounced contrast underscores the efficacy of our approach in accurately capturing the manner in which multimodal language models attend to visual information. It is important to note that the performance of the random head method is comparable to that of SnapKV [22], particularly in the case of the MHA model. This similarity is attributable to the fact that when the scores of all heads are randomly initialized, the cache budget allocated to each head is statistically equivalent, effectively causing the method to revert to the behavior observed with SnapKV [22]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 314, + 179, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 179, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 179, + 327 + ], + "type": "text", + "content": "4.3. Efficiency Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 297, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 297, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 297, + 448 + ], + "type": "text", + "content": "Setup In this subsection, we evaluate the computational efficiency of our proposed method, which holds significant practical value. Accordingly, our efficiency tests are conducted across a range of input token lengths " + }, + { + "bbox": [ + 55, + 339, + 297, + 448 + ], + "type": "inline_equation", + "content": "\\{2\\mathrm{K}, 4\\mathrm{K}, 8\\mathrm{K}, 16\\mathrm{K}, 32\\mathrm{K}\\}" + }, + { + "bbox": [ + 55, + 339, + 297, + 448 + ], + "type": "text", + "content": ". For each experiment, the output sequence length was fixed at 100 tokens, with a KV Cache budget set to 256. We computed the average decoding latency and peak memory consumption for each configuration. Notably, all experiments are done using FlashAttention." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "content": "Decoding Latency Fig. 6 illustrates that the reduction in KV Cache in our method substantially decreases the computational load during inference, thereby enhancing inference speed. For instance, when the input sequence length is " + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "inline_equation", + "content": "8\\mathrm{K}" + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "content": ", the LLaVA-NeXT-Vicuna-7B model exhibits a speedup of " + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "inline_equation", + "content": "1.16\\times" + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "content": ", while at a " + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "inline_equation", + "content": "32\\mathrm{K}" + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "content": " input length, the speedup increases to " + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "inline_equation", + "content": "1.87\\times" + }, + { + "bbox": [ + 55, + 454, + 298, + 562 + ], + "type": "text", + "content": ". These findings indicate that our approach significantly accelerates token generation, particularly in high-resolution or long video contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "text", + "content": "Memory Cost Our method also offers a marked reduction in peak memory usage, primarily by diminishing the memory footprint associated with the KV Cache. This reduction is especially pronounced in LLaVA-Series models. For example, with an input sequence length of 32K, LLaVA-NeXT-Vicuna-7B with full KV Cache requires 32.87 GB of memory, whereas our method reduces the requirement to 17.38 GB, thereby achieving an approximate " + }, + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "text", + "content": " reduction in memory overhead. It is noteworthy that even for the Qwen2-VL-7B-Instruct model, which employs an aggressive compression technique in its GQA framework, we can still reduce the cache by nearly 2GB with 32k inputs." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 555, + 321 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 321 + ], + "type": "image", + "image_path": "200338dabfd258d11d8a56382d7650fb1f9c66893b67a95584a6b413768a56f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 326, + 556, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 370 + ], + "type": "text", + "content": "Figure 6. Efficiency Evaluation for SparseMM. Benefiting from the reduction in KV cache, SparseMM can maintain nearly constant decoding latency, achieving up to a " + }, + { + "bbox": [ + 313, + 326, + 556, + 370 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 326, + 556, + 370 + ], + "type": "text", + "content": " acceleration. Additionally, it effectively reduces peak memory usage." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 385, + 377, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 377, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 377, + 399 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "text", + "content": "Performance Influence of Visual Heads. To further elucidate the impact of visual heads on the visual perception capabilities of multimodal models, we conducted a series of masking experiments. In these experiments, we selectively masked a specific proportion of the visual heads and, for comparison, randomly masked an equivalent proportion of attention heads. The evaluation was performed on OCR-Bench [29] and TextVQA [44], with performance measured relative to the baseline unmasked model. The results, as illustrated in Fig. 7, reveal that masking visual heads leads to a significant performance decline. In contrast, randomly masking the same proportion of attention heads produced a much smaller impact—for instance. These findings underscore the critical role that visual heads play in enabling MLLMs to effectively capture and process visual information. Moreover, masking the top " + }, + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "text", + "content": " of high-scoring visual heads causes a considerably greater performance loss than the additional impact of masking another " + }, + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 408, + 556, + 635 + ], + "type": "text", + "content": ", which highlights the sparse yet indispensable distribution of visual heads." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 715 + ], + "type": "text", + "content": "Robustness of Visual Head Identification. To evaluate the robustness of our visual head detecting approach, we show the distribution of the detected visual heads under different datasets and tasks in Fig. 8 and show the accuracy curve in Fig. 9. For the OCR task, we used the Multi-Lingual Text(MLT) [36] and Chinese Text in the Wild(CTW) [54]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23183" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 167 + ], + "type": "image", + "image_path": "d523c453f260dcbbb487f8afe152dc6adea0dc1fdd6c6670798d8a6b5730d869.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 295, + 205 + ], + "type": "text", + "content": "Figure 7. Comparisons of Mmasking Visual Head and Random Head. The left figure is the result on OCRBench, and the right figure is the result on TextVQA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 210, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 57, + 210, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 210, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 210, + 294, + 308 + ], + "type": "image", + "image_path": "e2a3b587fbb5035d10fe962ef6c5868e2217d23345bfd03bc3f7b7d204b182b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 296, + 345 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 345 + ], + "type": "text", + "content": "Figure 8. Visualizations of Visual Heads using Different Datasets. We visualized the attention distribution identified on MLT, CTW, and COCO datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 348, + 294, + 434 + ], + "blocks": [ + { + "bbox": [ + 57, + 348, + 294, + 434 + ], + "lines": [ + { + "bbox": [ + 57, + 348, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 57, + 348, + 294, + 434 + ], + "type": "image", + "image_path": "048a77182e77273234aa4b2489a60e153279a26d29b3f760880f1b2c9e40b950.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 439, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 483 + ], + "type": "text", + "content": "Figure 9. Results with Different Visual Head Identification Approaches and Datasets. We conduct an evaluation on visual heads identified on different datasets. The results on OCR datasets are similar and better than those on the detection dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 495, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 674 + ], + "type": "text", + "content": "datasets. In addition, we consider the object detection task and choose the COCO dataset [25], where the model is required to identify objects present in the images. We then localized the visual heads based on the correspondence between the model's answers and the relevant objects. As shown, the distribution of visual heads is relatively consistent across the OCR datasets, whereas there is greater variation on the COCO dataset. Moreover, experimental results demonstrate that the visual heads identified from OCR tasks are dataset-agnostic and exhibit strong generalizability, with better results than detection tasks. This is because OCR tasks establish an exact one-to-one mapping between the model's output and the visual content, whereas the COCO task, which focuses on larger bounding boxes, introduces more noise and results in less robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Accuracy and Speed Trade-off. We compared the accuracy and speed of SparseMM with other baselines in Tab. 2. We conducted an experiment on LLaVA-NeXT-Vicuna-7B with" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 112, + 553, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 2. Comparison of Accuracy-speed Trade-off among Different Methods. We compare the speed of all methods under 256 KV Cache budget and 16K input tokens." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 112, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 316, + 112, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 316, + 112, + 553, + 166 + ], + "type": "table", + "html": "
MethodsDocVQAOCRBenchTextVQAChartQATextCapsLatency(ms)
FullKV0.680.520.650.550.7352.9
SparseMM0.68(-0.00)0.52(-0.00)0.65(-0.00)0.54(-0.01)0.73(-0.00)37.1(-30%)
SnapKV0.64(-0.04)0.46(-0.06)0.62(-0.03)0.50(-0.05)0.65(-0.08)35.3(-33%)
PyramidKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.53(-0.02)0.65(-0.08)34.9(-34%)
AdaKV0.65(-0.03)0.48(-0.04)0.62(-0.03)0.49(-0.06)0.66(-0.07)37.3(-29%)
", + "image_path": "f9eb4287ac4b7b6367e379ef9ef31e01fe04f9692b621467c17887cd5ade2c91.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 178, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 553, + 281 + ], + "type": "image", + "image_path": "e18760383c881c12aa28c733f874edb4411ddd624268b8bf6683ecb82532f0ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 329 + ], + "type": "text", + "content": "Figure 10. Visualizations of Visual Heads. We visualized the attention distribution of several heads. The visual heads are able to accurately capture text or objects within the images, whereas the non-visual heads provide random results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 413 + ], + "type": "text", + "content": "a budget of 256 KV Cache. With the support of FlashAttention, our decoding latency is comparable to that of other methods, significantly lower than the FullKV method. However, our method outperforms others in terms of performance under the same budget. This effectively demonstrates the efficacy of SparseMM based on visual heads in MLLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 417, + 556, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 556, + 536 + ], + "type": "text", + "content": "Visualization of Visual Heads. To gain a more intuitive understanding of how visual heads process and interpret visual information, we conducted a visualization analysis of visual heads and non-visual heads on LLaVA-NeXT-Vicuna7B. As illustrated in Fig. 10, our observations indicate that non-visual heads often either neglect the image entirely or allocate a disproportionate amount of attention to visually insignificant regions. In contrast, visual heads accurately pinpoint regions of interest, allocating a substantial proportion of attention to these critical areas." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we present a systematic exploration of the visual processing characteristics inherent in MLLMs. Our analysis reveals a critical sparsity phenomenon that only a small fraction of attention heads actively engage in visual understanding. Leveraging this insight, we propose SparseMM, a novel KV-Cache optimization framework that dynamically allocates asymmetric computation budgets to attention heads based on their visual relevance. SparseMM prioritizes preserving vision-critical information during decoding, thereby achieving a more balanced accuracy-efficiency trade-off. We hope this study inspires deeper investigations into the principles governing multimodal learning." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23184" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62336004, Grant 62125603, Grant 62321005, and Grant 62441616, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 177 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 184, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[1] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, et al. Pixtral 12b. arXiv preprint arXiv:2410.07073, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[2] Joshua Ainslie, James Lee-Thorp, Michiel De Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 352 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 408 + ], + "type": "text", + "content": "[4] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information tunneling. arXiv preprint arXiv:2406.02069, 2024.3,5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 465 + ], + "type": "text", + "content": "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 521 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 523, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 579 + ], + "type": "text", + "content": "[7] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 580, + 296, + 645 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. Advances in Neural Information Processing Systems, 37:42566-42592, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 176, + 555, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 555, + 220 + ], + "type": "text", + "content": "[12] Jiajun Fei, Dian Li, Zhidong Deng, Zekun Wang, Gang Liu, and Hui Wang. Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos. arXiv preprint arXiv:2408.14023, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 222, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 554, + 266 + ], + "type": "text", + "content": "[13] Yuan Feng, Junlin Lv, Yukun Cao, Xike Xie, and S Kevin Zhou. Ada-kv: Optimizing kv cache eviction by adaptive budget allocation for efficient llm inference. arXiv preprint arXiv:2407.11550, 2024. 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 268, + 555, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 555, + 300 + ], + "type": "text", + "content": "[14] GeminiTeam. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 303, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 555, + 358 + ], + "type": "text", + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 555, + 403 + ], + "type": "text", + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. NeurIPS, 36:20482-20494, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 406, + 555, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 555, + 460 + ], + "type": "text", + "content": "[17] Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 462, + 555, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 555, + 518 + ], + "type": "text", + "content": "[18] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Donut: Document understanding transformer withoutOCR.arXiv preprint arXiv:2111.15664, 7 (15):2, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 555, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 555, + 564 + ], + "type": "text", + "content": "[19] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 609 + ], + "type": "text", + "content": "[20] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-Interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 555, + 656 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation. Advances in Neural Information Processing Systems, 37:22947-22970, 2024. 3, 5, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23185" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[23] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 262 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.2,5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 352 + ], + "type": "text", + "content": "[28] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 398 + ], + "type": "text", + "content": "[29] Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models. arXiv preprint arXiv:2305.07895, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 442 + ], + "type": "text", + "content": "[30] Zuyan Liu, Yuhao Dong, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Oryx mllm: On-demand spatial-temporal understanding at arbitrary resolution. arXiv preprint arXiv:2409.12961, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 498 + ], + "type": "text", + "content": "[31] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "text", + "content": "[32] Zuyan Liu, Yuhao Dong, Jiahui Wang, Ziwei Liu, Winston Hu, Jiwen Lu, and Yongming Rao. Ola: Pushing the frontiers of omni-modal language model. arXiv preprint arXiv:2502.04328, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 589 + ], + "type": "text", + "content": "[33] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. arXiv preprint arXiv:2405.20797, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[36] Nibal Nayef, Yash Patel, Michal Busta, Pinaki Nath Chowdhury, Dimosthenis Karatzas, Wafa Khlif, Jiri Matas, Umapada Pal, Jean-Christophe Burie, Cheng-lin Liu, et al. Icdar2019" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "robust reading challenge on multi-lingual scene text detection and recognition—rrc-mlt-2019. In 2019 International conference on document analysis and recognition (ICDAR), pages 1582–1587. IEEE, 2019. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 521, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 521, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 521, + 129 + ], + "type": "text", + "content": "[37] OpenAI. Openai gpt-3.5 api. OpenAI API, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 131, + 550, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 131, + 550, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 550, + 142 + ], + "type": "text", + "content": "[38] OpenAI. Gpt-4v(ison) system card. OpenAI Blog, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 144, + 539, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 539, + 155 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 539, + 155 + ], + "type": "text", + "content": "[39] OpenAI. Hello gpt-4o — openai. OpenAI Blog, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 156, + 554, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 156, + 554, + 177 + ], + "spans": [ + { + "bbox": [ + 316, + 156, + 554, + 177 + ], + "type": "text", + "content": "[40] QwenTeam. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 178, + 554, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 178, + 554, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 178, + 554, + 201 + ], + "type": "text", + "content": "[41] QwenTeam. Qwen2-vl: To see the world more clearly. Wwen Blog, 2024. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 203, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 203, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 203, + 555, + 258 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 259, + 555, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 259, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 259, + 555, + 323 + ], + "type": "text", + "content": "[43] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 742-758. Springer, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 382 + ], + "type": "text", + "content": "[44] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 384, + 555, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 555, + 405 + ], + "type": "text", + "content": "[45] Qwen Team. Qwen2.5: A party of foundation models, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 407, + 461, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 461, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 461, + 418 + ], + "type": "text", + "content": "[46] Qwen Team. Qwen2.5-v1, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 419, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 555, + 474 + ], + "type": "text", + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 476, + 555, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 555, + 530 + ], + "type": "text", + "content": "[48] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 533, + 555, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 577 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 578, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 611 + ], + "type": "text", + "content": "[50] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. arXiv preprint arXiv:2309.17453, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "text", + "content": "[51] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an lmm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "text", + "content": "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23186" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 294 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 137 + ], + "type": "text", + "content": "[53] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Anwen Hu, Haowei Liu, Qi Qian, Ji Zhang, and Fei Huang. mplugowl2: Revolutionizing multi-modal large language model with modality collaboration. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 13040-13051, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 182 + ], + "type": "text", + "content": "[54] Tai-Ling Yuan, Zhe Zhu, Kun Xu, Cheng-Jun Li, Tai-Jiang Mu, and Shi-Min Hu. A large chinese text dataset in the wild. Journal of Computer Science and Technology, 34(3):509-521, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "text", + "content": "[55] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 294 + ], + "type": "text", + "content": "[56] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36:34661-34710, 2023. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23187" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_content_list.json b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cfb5cdbbf15b4a1903e57d9629107ee945193994 --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_content_list.json @@ -0,0 +1,2308 @@ +[ + { + "type": "text", + "text": "SparseRecon: Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies", + "text_level": 1, + "bbox": [ + 104, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liang Han $^{1}$ , Xu Zhang $^{3}$ , Haichuan Song $^{2*}$ , Kanle Shi $^{4}$ , Yu-Shen Liu $^{1*}$ , Zhizhong Han $^{5}$ $^{1}$ School of Software, Tsinghua University, Beijing, China \n $^{2}$ Computer Science and Technology, East China Normal University, Shanghai, China \n $^{3}$ China Telecom ${}^{4}$ Kuaishou Technology, Beijing, China \n $^{5}$ Department of Computer Science, Wayne State University, Detroit, USA", + "bbox": [ + 153, + 202, + 841, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hanl23@mails.tsinghua.edu.cn, zhangxu@chinatelecom.cn, hcsong@cs.ecnu.edu.cn shikanle@kuaishou.com, liuyushen@tsinghua.edu.cn, h312h@wayne.edu", + "bbox": [ + 158, + 295, + 834, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surface reconstruction from sparse views aims to reconstruct a 3D shape or scene from few RGB images. The latest methods are either generalization-based or overfitting-based. However, the generalization-based methods do not generalize well on views that were unseen during training, while the reconstruction quality of overfitting-based methods is still limited by the limited geometry clues. To address this issue, we propose SparseRecon, a novel neural implicit reconstruction method for sparse views with volume rendering-based feature consistency and uncertainty-guided depth constraint. Firstly, we introduce a feature consistency loss across views to constrain the neural implicit field. This design alleviates the ambiguity caused by insufficient consistency information of views and ensures completeness and smoothness in the reconstruction results. Secondly, we employ an uncertainty-guided depth constraint to back up the feature consistency loss in areas with occlusion and insignificant features, which recovers geometry details for better reconstruction quality. Experimental results demonstrate that our method outperforms the state-of-the-art methods, which can produce high-quality geometry with sparse-view input, especially in the scenarios with small overlapping views. Project page: https://hanl2010.github.io/SparseRecon/.", + "bbox": [ + 86, + 395, + 483, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As one of the important tasks in computer vision, 3D reconstruction has attracted lots of research attentions in recent years. With the advancement of deep learning, 3D reconstruction based on neural implicit representations learned", + "bbox": [ + 89, + 814, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d2d98f09d624cff078923207d92e2931e107eb3db4bff04f48aa31c9e6f7363.jpg", + "image_caption": [ + "Figure 1. Given only 3 input images with large view angle change, our method can reconstruct a smoother surface compared to the state-of-the-art methods, such as UFOREcon [28], S-VolSDF [40] and NeuSurf [13]. The details of each surface are shown in the colored boxes." + ], + "image_footnote": [], + "bbox": [ + 516, + 362, + 905, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from point clouds [26, 30, 34, 58, 59] or images [23, 37, 45, 60] becomes a popular research topic. Although recent multi-view reconstruction methods[5, 37, 38, 41, 45, 53] have made great progress in terms of the reconstruction quality and reconstruction speed, they require a large number of dense views as supervision. When limited number of views are available, current reconstruction methods usually struggle to reconstruct high-quality surfaces.", + "bbox": [ + 511, + 638, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for sparse view reconstruction can be mainly classified into two categories: generalization-based methods and overfitting-based methods. The generalization-based methods [22, 24, 28, 31, 32] emphasize the generalization of sparse-view reconstruction, but they are mainly effective in scenarios with large view overlaps. In cases with views that were unseen during training, the quality of the reconstructed surface degenerates significantly, as shown in Figure 1. Meanwhile, it takes a long", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Yu-Shen Liu and Haichuan Song.", + "bbox": [ + 107, + 886, + 415, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28514", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time to pre-train these methods on large-scale data, but the generalization is still not good enough. Instead, overfitting-based methods [13, 14, 40, 46, 47] typically fit the 3D geometry directly from the sparse views by leveraging geometry clues. They show promising capability of reconstructing higher-quality geometric surfaces with small-overlapping views. However, the reconstruction quality of the existing methods is still unsatisfactory.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a multi-view feature consistency loss based on volume rendering and an uncertainty-guided depth constraint to learn neural signed distance functions. This approach allows us to achieve high-quality mesh reconstruction on more challenging sparse views with small overlaps.", + "bbox": [ + 89, + 213, + 480, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the feature consistency loss, we first employ the pre-trained Vis-MVSNet [50] to obtain depth features from the input images. Then, within a neural implicit rendering framework, the sampled 3D points along the rays emitted from the reference image are projected to the source image and the reference image. This allows us to acquire source features and reference features of each 3D point and measure the similarity between these two kinds of features. Finally, the feature similarity for each 3D point along the rays is accumulated through volume rendering, thus yielding the feature similarity associated with the rays. During optimization, we pursue higher feature similarity along the rays. Since the depth information is implicitly encoded with image features, feature consistency constraint can significantly alleviate the ambiguity issues arising from insufficient consistency of sparse views and low-texture during reconstruction.", + "bbox": [ + 89, + 306, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the uncertainty-guided depth prior constraint, we follow MonoSDF [47], utilizing a pre-trained network to acquire depth priors for each image, and then use it to constrain the regions with uncertain depth. However, monocular depth priors do not have consistent scales to the ground truth depth, which are hard to get calibrated to ground truth either due to the distortion. To effectively leverage the depth priors and provide proper supervision for occluded or under-constrained regions, we propose an uncertainty-guided depth prior constraint. First, we calibrate the depth priors using sparse point clouds obtained from COLMAP [33]. Then, during training, we compute the depth confidence from the rendered depth and impose the depth prior constraint only in regions with low confidence. This constraint helps infer more accurate geometry in occluded or under-constrained regions, minimizing the negative impact of depth prior errors on well-constrained regions.", + "bbox": [ + 89, + 564, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our methods on several widely used benchmarks and report the state-of-the-art results. In summary, our main contributions are as follows.", + "bbox": [ + 89, + 823, + 480, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel feature consistency loss based on volume rendering. It can effectively constrain the neural", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "radiance field by leveraging feature consistency among multiple views, improving the performance in sparse-view reconstruction tasks.", + "bbox": [ + 526, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- By incorporating depth confidence, we utilize the calibrated depth prior more effectively to enhance geometric constraints, further improving the reconstruction quality.", + "- Extensive experiments on the well-known datasets, such as DTU [16] and BlendedMVS[44], demonstrate that our method outperforms existing sparse-view reconstruction methods and achieves the state-of-the-art results." + ], + "bbox": [ + 513, + 136, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 258, + 653, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Implicit Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 285, + 787, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural implicit reconstruction methods [5, 7, 21, 37, 41, 47, 55, 56] have been rapidly developed based on neural volume rendering [27]. These methods introduce the Signed Distance Function (SDF) as the implicit representation of 3D surfaces in volume rendering in multi-view 3D reconstruction. While these methods have made significant improvements in both reconstruction quality and speed, it is important to note that they heavily rely on multiple views with large overlaps during the optimization.", + "bbox": [ + 511, + 308, + 903, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalization-based surface reconstruction with sparse views. In order to directly generalize the reconstruction results on sparse views, methods [22, 24, 28, 31, 32, 42] adopt the strategy of aggregating features from multiple view images to construct a feature volume, which is then used to predict the SDF for reconstructing the surface. VolRecon [32] uses transformers [17] to aggregate multi-view features, C2F2NeUS [42] employs cascade architecture to construct a volume pyramid, while ReTR [22] and UFORecon [28] aggregates multi-level features. These methods require pretraining on large-scale datasets, which typically takes several days. However, when there is a significant domain gap between the testing and training data, they all fail to reconstruct shapes effectively.", + "bbox": [ + 511, + 446, + 903, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overfitting-based surface reconstruction with sparse views. In contrast, overfitting-based methods directly fit the 3D geometry from the sparse images by geometric prior constraints. MonoSDF [47] employs depth and normal priors to achieve sparse reconstruction with small-overlapping views. However, such priors come with errors, and it does not fully leverage inter-view consistency, resulting in lower reconstruction quality. S-VolSDF [40] employs probability volumes obtained from MVS [9] models to guide the rendering weight estimated by VolSDF [45]. This improves the reconstruction results in sparse views with small overlap. However, the uncertainties in volumes make negative impact on the reconstructed surface, leading to surface roughness or significant defects. NeuSurf [13] leverages sparse point clouds and employs CAP-UDF [58] to construct an implicit geometric prior to improve the reconstruction qual", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28515", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bcd4dac80121e96e064f3d1eef9554e5730b2d82f3481f549bbb0abf8d6f5194.jpg", + "image_caption": [ + "Figure 2. SparseRecon consists of two main parts. (a) Volume rendering-based feature consistency constraint. We extract features from the reference image and source images. For a ray emitted from the reference image, we project each sampled point on the ray onto the source images to obtain the corresponding features. Then, the volume rendering-based feature consistency loss is calculated using the corresponding features on the reference image. (b) Uncertainty-guided depth prior constraint. We use another pre-trained network to obtain the depth prior of the reference image and calibrate it with the sparse point cloud obtained by COLMAP. Then, we calculate the confidence of the rendered depth, so that the calibrated depth prior only constrains areas with low confidence." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 867, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity from sparse views. However, when the sparse point cloud fails to cover the majority of the object surface, implicit geometric prior information cannot be useful. In contrast, our method employs more robust feature priors, calculates feature consistency based on volume rendering, and simultaneously utilizes depth priors to optimize the occluded regions, ultimately resulting in high-quality surface.", + "bbox": [ + 88, + 454, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 571, + 272, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian Splatting [18] has achieved unprecedented optimization speed and rendering quality in the task of novel view synthesis [11, 19, 57, 61]. However, since the Gaussians are unorganized, the discrete and unstructured points make it difficult to extract 3D surfaces through postprocessing. To address this issue, some methods introduce to use regularization terms [10], convert 3D Gaussians to 2D surfels [4, 12], acquire opacity fields through rays [48], improve the depth rendering algorithm [1] of 3DGS, or jointly optimize 3DGS with neural radiance fields [2, 20, 25, 54]. However, these methods require dense views to work well. Recently, FatesGS [14] achieves fast sparse-view reconstruction by leveraging depth priors and on-surface feature consistency constraints. Nevertheless, the reconstruction results still exhibit roughness or noticeable defects.", + "bbox": [ + 89, + 594, + 483, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Sparse View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 847, + 299, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, the novel view synthesis from sparse views is another category of work closely related to sparse view", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reconstruction. Depending on the technical framework, these works can be categorized into NeRF-based methods [6, 15, 29, 35, 36, 43, 49] and Gaussian Splitting-based methods [3, 11, 19, 52, 61]. This line of research also employs a limited number of views as input. However, they solely focus on the rendering quality of novel views rather than surface reconstruction, which are not designed specifically for the accurate geometric surface reconstruction. Due to the discernible bias (i.e. inherent geometric errors) [37] caused by the conventional volume rendering or inconsistencies in depth that appear in Gaussian rendering, current sparse view synthesis methods still fail to correctly reconstruct high-fidelity geometric surfaces.", + "bbox": [ + 511, + 454, + 906, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 662, + 606, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of our method is depicted in Figure 2. We introduce a novel feature consistency loss and an uncertainty-guided depth constraint based on the NeuS [37] framework. In this section, we first explain how to compute feature consistency for sampled points along rays. Then we explain how to enhance geometric constraints using depth priors and depth uncertainty. Thirdly, we introduce the color consistency loss. Finally, we present the overall loss function for optimization.", + "bbox": [ + 511, + 688, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Volume Rendering-based Feature Consistency", + "text_level": 1, + "bbox": [ + 511, + 833, + 903, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we use a pre-trained MVS network [50] to extract the features from both the reference image and the source image. Given a ray emitted from the reference image, let $p_r(0)$", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28516", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "denote the point where a ray intersects the reference image. And for each point $x_{i}$ along the ray, we denote its projection on the source image as $p_{s}(i)$ . Then, we bilinearly interpolate $F_{r}(0)$ and $F_{s}(i)$ at points $p_{r}(0)$ and $p_{s}(i)$ on image features, respectively. Formally, we define the feature consistency loss function as follows,", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {f e a t}} = M ^ {\\text {o c c}} \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} w _ {i} f _ {\\text {c o s}} \\left(F _ {r} \\left(p _ {r} (0)\\right), F _ {s} \\left(p _ {s} (i)\\right)\\right), \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 194, + 483, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{cos}$ is the cosine similarity, and $w_{i}$ corresponds to the weight for each point along the ray. $p_{s}(i) = K(Rx_{i} + t)$ is the projection of $x_{i}$ in source view, and $[K;R;t]$ is the camera parameters of source view. $M^{occ}$ is the occlusion mask.", + "bbox": [ + 89, + 248, + 483, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although MVsDF [51], NeuSurf [13] and FatesGS [14] also employ feature consistency constraints, they just leverage the intersection point between a camera ray and the object's surface. Then, this intersection point gets projected onto adjacent views to obtain the corresponding image features for the purpose of comparing features at this point across multiple views. In sparse view scenarios, the estimated positions of surface points can easily deviate significantly, making the on-surface feature consistency loss not converge. NeuSurf [13] and FatesGS [14] utilize sparse point clouds generated by COLMAP [33] as priors, enabling it to obtain partially accurate positions of surface points, thereby allowing the on-surface feature consistency loss to be more effectively leveraged. However, in regions of lacking surface points, the on-surface feature consistency loss cannot ensure the attainment of high-quality geometric surfaces.", + "bbox": [ + 89, + 323, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 illustrates the difference between on-surface feature consistency and volume rendering-based feature consistency. Due to the uncertainty of gradient direction, the constraint solely relying on surface point features is challenging to be optimized. In contrast, our method does not require the prior estimation of surface points, it calculates feature consistency on all sampling points along the ray, and provides more reasonable and comprehensive supervision to the implicit field, thereby addressing the convergence issue that may arise in sparse reconstruction for MVsDF [51] and NeuSurf [13].", + "bbox": [ + 89, + 580, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Uncertainty-Guided Depth Constraint", + "text_level": 1, + "bbox": [ + 89, + 757, + 419, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although multi-view features offer more robust constraint than image colors, they are ineffective for occluded regions. Due to the limited number of views, some regions may only be visible from a single viewpoint. To enhance geometric constraints, we employ depth priors to supervise the radiance field. However, monocular depth priors are not perfect and accurate. Although MonoSDF [47] has already taken the inaccuracy of depth priors into account, i.e., it aligns", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65ace355b98a0bfb4bd893044bbddcc65d9f294e9ecc0f00640d3e80b9a2a34b.jpg", + "image_caption": [ + "(a) Surface Feature Consistency" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 625, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f37d7a6ca0870534856c8febb92469a26e0942dde5a49d2c566314facffb4605.jpg", + "image_caption": [ + "(b) Volume Rendering-based Feature Consistency" + ], + "image_footnote": [ + "Sample point", + "Gradient direction", + "Ray", + "Estimated surface", + "Real surface" + ], + "bbox": [ + 627, + 88, + 790, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66beb6e5d38527afcec9dbbf16ca9c447fabacbfbe81f20dc153aff7fdaeef00.jpg", + "image_caption": [ + "Figure 3. The illustration of (a) on-surface feature consistency and (b) feature consistency with volume rendering.", + "Figure 4. The illustration of rendered depth produced by different depth prior utilization methods, along with the corresponding error maps. (a) Calibrate the depth prior using the rendered depth during training. (b) Calibrate the depth prior using the COLMAP sparse point cloud." + ], + "image_footnote": [], + "bbox": [ + 514, + 255, + 900, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depth priors using rendered depth during training. However, the rendered depth during training is inaccurate, resulting in significant errors in the calibrated depth priors. This ultimately leads to the accumulation of errors during training, which results in inaccurate reconstructions. Figure 4 (a) shows the calibrated depth prior and rendered depth obtained by MonoSDF [47], as well as their error maps compared to the ground truth depth. It can be seen that both the calibrated depth prior and the rendered depth are with large errors. Therefore, MonoSDF [47] uses a weight annealing strategy to anneal the weight of depth loss to 0 during the first 200 training epochs.", + "bbox": [ + 511, + 520, + 906, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another trivial approach is to calibrate the depth priors using the sparse point cloud obtained from COLMAP [33]. Since the sparse points are generally located on the geometric surface of the object, their depth is relatively accurate. Therefore, calibrating the depth priors using the sparse point cloud can lead to more accurate depth priors. Figure 4 (b) shows the depth priors calibrated with the sparse point cloud, and the rendered depth, as well as their error maps compared to the ground truth depth. It indicates that the depth priors calibrated to the point cloud from the COLMAP [33] are more accurate. Therefore, we can use them as an constraint leads to more precise rendered depth.", + "bbox": [ + 511, + 703, + 908, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, due to the distortions in monocular depth pri", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9fad56129eeb94608cdb0ba400507bb5d1df0d83aff02e24730ee2c329fa0804.jpg", + "image_caption": [ + "Figure 5. Left: the method of obtaining the confidence of rendered depth. Right: the rendered depth and the depth confidence." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 483, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ors, it is impossible to perfectly align them with the ground truth depth. Even after calibration, the depth priors still exhibit noticeable errors when compared to the ground truth depth. In sparse view scenarios, occlusions and insufficient constraints are more common, leading to significant discrepancies between the geometry of occluded regions and the real surface. Therefore, to achieve more accurate geometry in these under-constrained regions while avoiding the negative impact of depth prior errors on well-constrained regions, we propose an uncertainty-guided depth prior constraint method to more effectively utilize the depth priors. Specifically, we apply depth prior constraints in regions with depth uncertainty, while refraining from using them in regions with high depth confidence.", + "bbox": [ + 88, + 260, + 482, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain the confidence of the rendered depth, we employ a method to evaluate the multi-view depth projection consistency. As shown in Figure 5, for a specific pixel $r_{uv}$ in the reference image with depth $d_r$ , it can be mapped to a neighboring image through the homography matrix $H_{rs}$ , leading to a pixel $s_{uv}$ ,", + "bbox": [ + 89, + 473, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {u v} = H _ {r s} r _ {u v}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 580, + 480, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {r s} = M _ {s} M _ {r} ^ {- 1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 603, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_r$ and $M_s$ are the projection matrices corresponding to the reference and source views, respectively. Similarly, we can map the pixel $s_{uv}$ in the source view to the reference view using the projection matrix $H_{sr}$ and its corresponding depth $d_s$ , resulting in $\\hat{r}_{uv}$ . The forward and backward projection distance error reflect the accuracy of depth predictions, so we take it as the depth confidence, which is defined as", + "bbox": [ + 89, + 628, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {d} = \\left\\{ \\begin{array}{l l} \\frac {1}{e ^ {\\| r _ {u v} - \\hat {r} _ {u v} \\|}}, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| \\leqslant 1 \\\\ 0, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| > 1 \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 762, + 482, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The right side of Figure 5 shows the rendered depth and the corresponding depth confidence.", + "bbox": [ + 89, + 809, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondingly, the depth uncertainty is defined as $U_{d} = 1 - C_{d}$ . Meanwhile, we can set a threshold $\\tau$ for depth confidence $C_d$ to obtain the occlusion mask $M^{occ} = \\{C_d > \\tau\\}$ .", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For depth calibration, we leverage COLMAP [33] to obtain a sparse point cloud $\\{X : x_1, x_2 \\ldots x_i \\in R\\}$ and visibility flags indicating which keypoints are visible from view $I$ . Given the camera parameters $P$ of view $I$ , we estimate the depth $\\bar{D}_i$ of keypoints by computing the distance from the visible keypoints $x_i$ to the camera center $o$ . Then, we calibrate the monocular depth prior $\\hat{D}$ with $\\bar{D}_i$ , it can be defined as $\\bar{D} \\approx a\\hat{D} + b$ , where $a$ is the scale factor and $b$ is the shift factor, obtained through the least squares method. Formally, the depth constraint loss is defined as,", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d e p t h} = \\sum_ {r \\in R} U _ {d} \\left\\| (a \\hat {D} + b) - D _ {p r e d} \\right\\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 248, + 906, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Color Consistency Constraint", + "text_level": 1, + "bbox": [ + 511, + 291, + 777, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although feature consistency constraint can ensure that the reconstruction does not suffer from severe artifacts, it does not provide sufficient supervision to reconstruct fine geometric details. Conversely, in cases with rich textures, image color constraint can refine the geometric details. Therefore, following the NeuralWarp [5], pixel warping loss and patch warping loss are used in our method as multi-view color consistency loss functions,", + "bbox": [ + 511, + 313, + 906, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {c o l o r} = \\sum_ {r \\in R} M ^ {o c c} d _ {p i x e l} (I (r), I _ {s} (r)) \\\\ + \\sum_ {r \\in R} M ^ {\\text {o c c}} d _ {\\text {p a t c h}} (P (r), P _ {s} (r)), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 441, + 903, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I(r)$ and $I_s(r)$ are the ground truth color of the pixel from which the ray emits and the rendered color, respectively, $P(r)$ and $P_s(r)$ are the ground truth color of the patch corresponding to the ray and the rendered patch color, respectively. $d_{pixel}$ is the loss metric for pixel color, where we use $L1$ loss as $d_{pixel}$ . $d_{patch}$ is the loss metric for patch color, where we use the Structural Similarity Index Measure (SSIM [39]) as $d_{patch}$ .", + "bbox": [ + 511, + 515, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Loss", + "text_level": 1, + "bbox": [ + 511, + 643, + 656, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the above-mentioned three loss functions, we also use the Eikonal loss [8] used in NeuS [37]. We define the overall loss function as follows:", + "bbox": [ + 511, + 666, + 905, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {f e a t}} + \\alpha L _ {\\text {d e p t h}} + L _ {\\text {c o l o r}} + \\beta L _ {\\text {e i k}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 719, + 903, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$L_{eki}$ is the Eikonal loss [8], used to regularize the SDF values of sampled points, defined as", + "bbox": [ + 511, + 742, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e k i.} = \\frac {1}{m n} \\sum_ {i, k} \\left(\\| \\nabla f \\left(x _ {i, k}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 779, + 903, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 848, + 609, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on DTU [16] and BlendedMVS [44] dataset. For the DTU [16] dataset, to avoid using the", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28518", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3f7d3fa13c00ccba9828aec1d78ba30d079e344df17c1eb48d554ad011882347.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods21243437384082106110114118Mean CD ↓
NeuS [37]5.633.586.004.602.574.531.914.185.461.194.163.98
NeuralWarp [5]2.531.880.741.800.8411.502.642.104.371.192.632.93
MonoSDF [47]4.145.921.394.552.192.142.365.624.581.633.023.41
Vis-MVSNet [50]3.394.440.853.361.693.353.352.342.160.741.832.50
MVSDF [51]4.314.711.656.371.774.473.611.871.671.251.693.03
SparseNeuSft [24]3.484.372.924.762.793.732.801.863.101.152.293.02
VolRecon [32]2.723.071.824.322.143.043.002.562.811.493.222.75
GenSft [31]5.867.673.628.575.375.415.486.045.294.694.355.67
ReTR [22]2.673.371.623.681.873.403.672.842.851.562.352.72
UFORecon [28]1.841.520.792.581.001.821.721.200.930.661.261.39
S-VolSDF [40]2.453.081.333.091.223.211.911.511.230.741.21.91
SparseCraft [46]2.882.420.922.971.582.782.511.105.240.650.882.16
NeuSurf [13]7.601.432.933.181.532.861.861.091.410.370.622.26
FatesGS [14]3.981.322.532.853.362.713.761.490.850.471.062.22
Ours2.141.260.721.460.861.391.370.940.770.440.831.11
", + "bbox": [ + 117, + 88, + 879, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of Chamfer Distance (CD↓) on DTU dataset with 3 small-overlapping images. The methods are divided into three categories, from top to bottom: (1) dense-view reconstruction methods related to ours, (2) generalization-based sparse-view reconstruction methods, and (3) overfitting-based sparse-view reconstruction methods. the best results are in bold, the second best are underlined.", + "bbox": [ + 89, + 357, + 906, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scenes that have already been used as training data on the pretrained Vis-MVSNet [50] model, we select the same 11 scenes as in S-VolSDF [40]. The image resolution is set to $1600 \\times 1200$ . Similar to the S-VolSDF [40] and NeuSurf [13] methods, we select the views 22, 25, and 28 for the more challenging reconstruction of small overlaps.", + "bbox": [ + 89, + 440, + 482, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the BlendedMVS [44] dataset, we follow the S-VolSDF [40] to use the same 9 challenging scenes, with 3 small-overlapping views for each scene. The image resolution is set to $768 \\times 576$ .", + "bbox": [ + 89, + 531, + 483, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 603, + 307, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the same network architecture and initialization strategy as NeuS [37] and incorporated our volume rendering feature consistency loss, uncertainty-guided depth constraint loss, and color consistency loss. For the weight factors in the loss functions Eq. 7, we set the $\\alpha$ for the uncertainty-guided depth prior constraint loss $L_{depth}$ to 0.5 and the $\\beta$ for the Eikonal loss $L_{eik}$ to 0.1. Each scene is trained 100K iterations on a RTX3090 GPU. The patch warping term in the color consistency loss requires the surface point normals to calculate homographies, but the initial normals are too noisy [5], therefore, the patch warping loss is applied after 20k training steps. The threshold $\\tau$ of the occlusion mask is set to 0.", + "bbox": [ + 89, + 625, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Baseline", + "text_level": 1, + "bbox": [ + 89, + 832, + 191, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our approach with three categories of methods including dense-view methods: NeuS [37], NeuralWarp [5], Vis-MVSNet [50], MVsDF [51], generalization-based", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "methods: SparseNeuS [24], VolRecon [32], GenS [31], ReTR [22] and UFORecon [28], overfitting-based methods: S-VolSDF [40], SparseCraft [46], NeuSurf [13] and FatesGS [14]. The reconstruction results for SparseNeuS [24] and GenS [31] are fine-tuned using 3 views for each scene.", + "bbox": [ + 511, + 440, + 906, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 544, + 653, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction on DTU. For a comprehensive comparison, we evaluate the baselines and our method on both small-overlapping and large-overlapping views. Following baselines [13, 14, 24], we report the Chamfer Distance (CD) between the reconstruction surfaces and the ground truth point clouds. Since the scenes in our experiments differ from those used in NeuSurf [13] and FatesGS [14], we reran the official implementations of NeuSurf and FatesGS. The CD results with small overlapping views are shown in Table 1. The meshes reconstructed by several methods using 3 views with small overlapping are shown in Fig. 6. For the generalization-based sparse reconstruction methods, we only show the reconstruction results of the latest UFORecon [28], as the reconstruction quality of other methods is lower than that of UFORecon [28]. The experimental results show that our method significantly improves the mesh quality with small overlap views, compared to the state-of-the-art sparse-view reconstruction methods. The results of large overlapping views are presented in the supplementary materials.", + "bbox": [ + 511, + 566, + 906, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 6, when input sparse views with small overlap, both MonoSDF [47] and SparseCraft [46]", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28519", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c7d9991dce4278a6c28be20aaa40925545f12143fc053d9aa8200959e6fccdc.jpg", + "image_caption": [ + "GT images" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 236, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e80e85c5aa13106c64819e8da04473dedc21a31942dea52021d9a21c8aae4d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 87, + 333, + 145 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fbb0598f2f8d14ac6a0ed644a7e8c4b8c2ce8b19d63b2a3689a3c961c628930.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 146, + 328, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4bd6a765902a982d98f947cd0ab0c2a58c89aa660ba68e77994c60f7e7cd50fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 200, + 321, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1116ebbb9d339ecabb444f30535805bd4b39de437914b0c9e5d1367dcfce12c4.jpg", + "image_caption": [ + "MonoSDF" + ], + "image_footnote": [], + "bbox": [ + 246, + 256, + 321, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3a528cc2808cecba63d53848c12e893630e392b85028ddc6908d52ff3461797.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 88, + 424, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3484fc088d7f31f8c9fe0aa3866bb47d3e6d6d3e7528fb800423fb3095cd858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 145, + 423, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8907d9a022d8aaa6107e5f971513366816ed5754232ab0f251c9b9c50bbacd4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 200, + 416, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a1c1de873c88483e7cebb3b4f36e387a2812f183d5661e1c34f60cfac90b2cb.jpg", + "image_caption": [ + "UFORecon" + ], + "image_footnote": [], + "bbox": [ + 339, + 255, + 416, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/540dfc52c2099175311b82afa4f5b4d9683fb39f32d0313078fe3e5de6a98458.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 88, + 517, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd37693c17a82b46234fe34f1b828e604e672a3fb66a1e5a2f00f0a696663537.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 145, + 517, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32f99fe08d624837f5888372013a4326247c67c772c35d3616c4be9847ff45db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 200, + 501, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9b515cb0727f709de61082d891b069bf040e1b1f6d7e427cf732930558239a8.jpg", + "image_caption": [ + "S-VolSDF", + "Figure 6. Visual comparison on DTU dataset with 3 small-overlapping images." + ], + "image_footnote": [], + "bbox": [ + 437, + 255, + 513, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5234037b56fba1b421cdb8b9b1fae1d7cda56a6ade0e3544079022b7bae265b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 89, + 614, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/848987358e9abd5d35bfdc9b94b9ef5f00d540bca6628b79591c13c32dee55b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 145, + 612, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6e8e3a59d606d08d455d4f5b0230b9cce792124eccbe0b31d5b877e7c34b09f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 200, + 596, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca6844839871e6b2bdd44366b4dd4fac7c6149c8cb76f5a3063c3e5c21bf613e.jpg", + "image_caption": [ + "SparseCraft" + ], + "image_footnote": [], + "bbox": [ + 522, + 255, + 617, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e4b9c45912f4a7bd4e3c203ec244f3266ace4c66a3733335df1298ddae5dd00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 88, + 710, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6aa118f1b7278b997ae3e977b98efc503b036a22ba3b7a66cb36222f03d3cfba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 145, + 709, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b9823d87660d692fabfc045ba6886fcdc515d61174d50d866df8354c48860c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 200, + 702, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d02057027f2e0cd53180fdf93cac62c0207c63130c08860a6de56764af03907.jpg", + "image_caption": [ + "NeuSurf" + ], + "image_footnote": [], + "bbox": [ + 625, + 255, + 702, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d199e8d4f230f4b9bcdd699e4fbc14914d76a25d0d994cf86fcce866656d79b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 89, + 803, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb57d9f67da2ae741fdc296abd71c0ea90e08c5d96200ba512d3bf2fb77cb02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 145, + 802, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d718d6e306504646c416f10ee17338571dc051bf6e5e69c8019513faf8a71796.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 200, + 790, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b18ff085e23afa7201f5e9ddfe05bc75c8d20425c0e038d70e11913565b9cdb.jpg", + "image_caption": [ + "FatesGS" + ], + "image_footnote": [], + "bbox": [ + 723, + 256, + 795, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/804ec0d600d7f20b65e8c4949a378ffb9eec19698a2cbfeb4fee6a14a29c276f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 90, + 900, + 143 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4733198d95d183b85f599307f0997b589da03ae02e2c4067e0dce0d43c94563f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 145, + 900, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd45dd827ccc96796c18ae86319e8ec97292e481c27170d7b591e38571f10175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 200, + 885, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ff8fffec39e575eb40b82320151dc22e8a07cf210405c0822065e8c92528ae0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 818, + 255, + 893, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "suffer from reconstruction ambiguity and failures, highlighting that relying solely on simplistic geometric prior constraints is insufficient to obtain complete and accurate meshes. UFORecon [28] shows significant roughness in its reconstruction results. S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit noticeable reconstruction defects. Experimental results demonstrate that our method is effective in alleviating geometric and appearance ambiguities during the optimization process. This significantly enhances the quality of mesh reconstruction, especially in scenarios with small overlapping views and low texture.", + "bbox": [ + 89, + 380, + 483, + 546 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e139c7de8fd0b69d91c91309f8ce88d8784fdd13540c9ce32b5df4390519c5a2.jpg", + "image_caption": [ + "GT Images", + "MonoSDF", + "Figure 7. Visual comparison on BlendedMVS dataset. $\\times$ indicates reconstruction failure." + ], + "image_footnote": [], + "bbox": [ + 91, + 560, + 184, + 700 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/750bf527000ee7ffc93382c4ca91b44018762741e3c4454082176563084aa809.jpg", + "image_caption": [ + "X" + ], + "image_footnote": [], + "bbox": [ + 192, + 560, + 233, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f2e5535ecc3df2c42535293861601229deaaa7c6e8914b0a340aab6aae2e3ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 561, + 299, + 660 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/284d239afac530b1d5468a34271c5f844108913ea5d07e4b7199b8983a8048e6.jpg", + "image_caption": [ + "S-VolSDF" + ], + "image_footnote": [], + "bbox": [ + 245, + 662, + 300, + 700 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42d69eecfa0ada455d2303e7fe3238e2b683c9bfa70f4c20414cc4d6df2f659d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 561, + 349, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8a2fd7e623536b8a49b54eea3bfe7f3c66b1e2cf39d446421535349d055d45e.jpg", + "image_caption": [ + "NeuSurf" + ], + "image_footnote": [], + "bbox": [ + 303, + 662, + 359, + 700 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b37806df4a5809c1f843a618567856c08790fb2825d6a72235962bedef785f09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 561, + 416, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c7c1bacb02c966cbdb4d82cbe62a27903985040aefd53f75e0f775235f75497.jpg", + "image_caption": [ + "FatesGS" + ], + "image_footnote": [], + "bbox": [ + 362, + 662, + 418, + 700 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6b0709656a8a4838036c97717cbcde6d23bbaa83161e4c36eb552356b313bb9.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 421, + 560, + 480, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction on BlendedMVS. Figure 7 presents the visual comparison of reconstructed mesh for overfitting-based methods. With only 3 small-overlapping views provided, all of the generalization-based methods completely fail to reconstruct in the sparse setting of BlendedMVS[44] dataset, even if SparseNeuS [24] is fine-tuned. Therefore, the reconstruction results of these methods are not included", + "bbox": [ + 89, + 794, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in Figure 7. Compared to other methods, our approach can generate more complete and detailed meshes. Similarly, MonoSDF [47] fails to reconstruct surfaces either. The meshes generated by S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit significant defects. Both NeurSurf [13] and FatesGS [14] use on-surface feature consistency constraints, but the reconstruction results are still not good enough. In contrast, our method achieves more comprehensive geometry and finer details by employing volume rendering-based feature consistency constraints. This highlights the advantages of our approach in geometric consistency. More visualizations are presented in the supplementary materials.", + "bbox": [ + 511, + 380, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2e14c2d9288456a7f4f6b8f98b5b1a414bcb81b7b85d4f82a0263d256de7a2b.jpg", + "image_caption": [ + "Scan 40" + ], + "image_footnote": [], + "bbox": [ + 542, + 588, + 614, + 640 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85612e3143694ee3e593587cac774d085724692cdf55ddec13def2e89ae45fbf.jpg", + "image_caption": [ + "Scan 110", + "Mesh w/ $L_{\\text{feat}}$ by ours", + "Figure 8. Reconstructed meshes and error maps on DTU dataset with different feature consistency losses." + ], + "image_footnote": [], + "bbox": [ + 542, + 641, + 617, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d82f1550c1a04c064fbe70b6f71c81e668383cbf078cb3ba61675481fe153e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 588, + 702, + 641 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2be16916de4bde2d6d84c72f7d19cc87bbb77f713247442bdeba7c52ccb2afa.jpg", + "image_caption": [ + "Error map w/ $L_{\\text{feat}}$ by ours", + "at Mesh w/ surf $f_{\\text{feat}}$ by MVSDF" + ], + "image_footnote": [], + "bbox": [ + 633, + 642, + 709, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6471b5b4c7c6ca2584c4b631c3c70db1768c4c07689cc7ed8de0cfa5d45a517d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 588, + 800, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/27bb6e20c01eae7f46cd2166eaa1dbc0b9ed988c3c36c4be0f96a244aef12996.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 643, + 802, + 695 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2586dba55b8aec8296c88d3413c0f39acc451fa281d08ee53b14c18c10168ffa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 588, + 893, + 642 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9166b28abdee442f792d892293b2e13a1fa28d43ff64575d719b8449e61a47ee.jpg", + "image_caption": [ + "Error map w/ $L_{\\text{feat}}^{\\text{surf}}$ by MVSDF" + ], + "image_footnote": [], + "bbox": [ + 807, + 643, + 895, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 787, + 666, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the components of our method with 3 small-overlapping views by an ablation study on the DTU [16] dataset. To compare the depth loss $L_{depth}^{mono}$ calibrated by rendered depth in MonoSDF [47] with our depth loss $L_{depth}$ , we replace $L_{depth}$ with $L_{depth}^{mono}$ to evaluate it in our method. We also compare the volume rendering-based", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28520", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b024ef5a3e4040cfabf5ed8069817cd1e0b120cfba75112561a07d647fdc259.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLcolorLfeatLdepthLmono depthL1featL2featsurffeatCD↓
Baseline3.35
✓1.76
✓✓1.47
✓✓1.62
✓✓✓1.11
✓✓✓1.59
✓✓✓2.36
✓✓✓1.81
✓✓✓2.93
", + "bbox": [ + 94, + 87, + 482, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on DTU dataset with 3 small-overlapping images.", + "bbox": [ + 89, + 210, + 482, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/451bb413ec3309041737c2a5b784522cd14538751a6dbdfd7c5ba2c2f15ef45c.jpg", + "image_caption": [ + "Step 10k" + ], + "image_footnote": [], + "bbox": [ + 93, + 255, + 189, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/131b2eec1aaa987c6c2fc44b3bf9e8506d1202bfe6add5f806c15e13e8bceda7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 255, + 287, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dc8b7ee7873229fbc42d7b82e09bfb5814477cbf41799d23abe17dc7f476241.jpg", + "image_caption": [ + "Step 20k" + ], + "image_footnote": [], + "bbox": [ + 189, + 313, + 287, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/565976665936a09e43f0998137701f1ac6ad3b225dd18998d44903bee55a0d4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 255, + 383, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18d4841a6253c82ef59dfb7b19add450704ec4221adc3f4d857b2be65c3111fb.jpg", + "image_caption": [ + "Step 40k" + ], + "image_footnote": [], + "bbox": [ + 289, + 313, + 383, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/238742f5599328f8c295cb67b70794054114a20205eb71cbb7a81fa4c44b8a2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 255, + 480, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bdd4739c5e95bcac5bb70a44287df07815f86a7561776eafecbab25c24ce8632.jpg", + "image_caption": [ + "Step 60k" + ], + "image_footnote": [], + "bbox": [ + 385, + 313, + 480, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b585ad7b38d2901e1288f5bfda5791d6ec524df280ec939f83997a0f3f50aefd.jpg", + "image_caption": [ + "Figure 9. The variation of weighted feature similarity during training, brighter colors indicate higher feature similarity.", + "Figure 10. Visualization of reconstruction and error maps for scene scan24 and scan37 in DTU dataset with different losses. The differences of error maps are highlighted." + ], + "image_footnote": [], + "bbox": [ + 98, + 440, + 480, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feature consistency loss calculated using L1 distance (denoted as $L_{\\text{feat}}^{L1}$ ) and L2 distance (denoted as $L_{\\text{feat}}^{L2}$ ) with our method using feature similarity distance. We found that feature similarity distance is better than both L1 and L2 distance, as shown in Table 2.", + "bbox": [ + 89, + 761, + 482, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we replace our volume rendering-based feature consistency loss $L_{feat}$ with the on-surface feature consistency loss $L_{feat}^{surf}$ used in MVSDF [51] to compare the effectiveness of two different loss functions. Figure 8 illustrates the performance of these two different loss functions.", + "bbox": [ + 89, + 838, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a93de04f8c54b87ca5aa86f8524b25e96821ab93218da111d0e9957273a690e7.jpg", + "image_caption": [ + "GT Images" + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 624, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5e2d2251d0b4eec902efa1ea1f7eef70213e789d50ec49e4c3e8d9414f4d6a9.jpg", + "image_caption": [ + "S-VolSDF" + ], + "image_footnote": [], + "bbox": [ + 635, + 90, + 692, + 127 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03dd7a95ec283cd2a379c445e8d07dd8229d276449f1c929f297781bca1804df.jpg", + "image_caption": [ + "NeuSurf", + "Figure 11. Failure case. For specular objects, the ambiguity in the color consistency constraint may lead to a rough surface." + ], + "image_footnote": [], + "bbox": [ + 694, + 90, + 761, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7fcc5f9a7a10ef9f527e262beedd3796d6f6220251f49abb4ceaff1f7337da73.jpg", + "image_caption": [ + "FatesGS" + ], + "image_footnote": [], + "bbox": [ + 764, + 90, + 831, + 130 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6f12d9fe4bb0baee019f107fb50213f5860956a32c6570b65995cf268da237f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 834, + 90, + 903, + 130 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trates the reconstruction results and error maps on the DTU dataset when using different feature consistency losses, under the on-surface feature consistency loss $L_{feat}^{surf}$ , the meshes show large artifacts.", + "bbox": [ + 511, + 212, + 903, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows the average Chamfer Distance over all 11 scenes on DTU dataset using different losses. The experimental results indicate that both feature consistency loss and uncertainty-guided depth constraint improve the surface reconstruction.", + "bbox": [ + 511, + 272, + 903, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 9 illustrates the variation of the weighted feature similarity map during the training process. Brighter colors indicate higher feature similarity, demonstrating that our volume rendering-based feature consistency loss can provide effective constraints.", + "bbox": [ + 511, + 347, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 10 shows the reconstructed meshes and error maps for scene scan24 and scan37 on the DTU [16] dataset when using different losses. It can be observed that the mesh deteriorates without the volume rendering-based feature consistency loss or the uncertainty-guided depth constraint loss, and the reconstruction quality drops when using the depth loss $L_{depth}^{mono}$ in MonoSDF [47].", + "bbox": [ + 511, + 422, + 905, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 542, + 638, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel method for learning implicit representations from sparse views with small overlaps. Our novelty lies in a novel volume rendering-based feature consistency loss and an uncertainty-guided depth constraint. Extensive experiments on the DTU [16] and BlendedMVS [44] datasets show that our method surpasses existing state-of-the-art sparse-view reconstruction methods in terms of reconstruction quality.", + "bbox": [ + 511, + 568, + 905, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our method shows significant improvement over other sparse view reconstruction methods, there are still some limitations. Firstly, for specular objects, the ambiguity in the color consistency constraint may lead to a rough surface, as shown in Figure 11. Secondly, Following previous studies [13, 24, 40], the camera poses of sparse views are obtained from the training dataset. However, in some cases, it may not be possible to obtain accurate camera poses using SfM methods like COLMAP [33] due to the lack of texture in the images or excessive viewing angles. Additionally, the feature consistency constraint method requires a pre-trained network to extract image features. The accuracy of the features determines the performance of the feature consistency constraint.", + "bbox": [ + 511, + 689, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 261, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The corresponding authors are Yu-Shen Liu and Haichuan Song. This work was supported by National Key R&D Program of China (2022YFC3800600), and the National Natural Science Foundation of China (62272263), and in part by Tsinghua-Kuaishou Institute of Future Media Data.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, PP: 1-12, 2024. 3", + "[2] Hanlin Chen, Chen Li, and Gim Hee Lee. NeuSG: Neural implicit surface reconstruction with 3D gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 3", + "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. MVSplat: Efficient 3D gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 3", + "[4] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3", + "[5] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 1, 2, 5, 6", + "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3", + "[7] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2", + "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 5", + "[9] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2", + "[10] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 3" + ], + "bbox": [ + 93, + 232, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3D gaussian splattering with view consistency for sparse view synthesis. In Advances in Neural Information Processing Systems, 2024. 3", + "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3", + "[13] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. NeuSurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 1, 2, 4, 6, 7, 8", + "[14] Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2, 3, 4, 6, 7", + "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a Diet: Semantically consistent few-shot view synthesis implementation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3", + "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanas. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 2, 5, 7, 8", + "[17] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165, 2020. 2", + "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 3", + "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3", + "[20] Shujuan Li, Yu-Shen Liu, and Zhizhong Han. GaussianUDF: Inferring unsigned distance functions through 3D gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 27113-27123, 2025. 3", + "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2", + "[22] Yixun Liang, Hao He, and Yingcong Chen. ReTR: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6", + "[23] Xinqi Liu, Jituo Li, and Guodong Lu. Reconstructing complex shaped clothing from a single image with feature stable" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28522", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "unsigned distance fields. IEEE Transactions on Visualization and Computer Graphics, 2024. 1", + "[24] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. SparseNeuS: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 6, 7, 8", + "[25] Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 3DGSR: Implicit surface reconstruction with 3D gaussian splatting. ACM Transactions on Graphics, 43(6):1-12, 2024. 3", + "[26] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3D point clouds via noise to noise mapping. In International Conference on Machine Learning, 2023. 1", + "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[28] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. UFORecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 1, 2, 6, 7", + "[29] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3", + "[30] Takeshi Noda, Chao Chen, Junsheng Zhou, Weiqi Zhang, Yu-Shen Liu, and Zhizhong Han. Learning bijective surface parameterization for inferring signed distance functions from sparse point clouds with grid deformation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22139-22149, 2025. 1", + "[31] Rui Peng, Xiaodong Gu, Luyang Tang, Shihe Shen, Fanqi Yu, and Ronggang Wang. GenS: Generalizable neural surface reconstruction from multi-view images. In Advances in Neural Information Processing Systems, pages 56932-56945, 2023. 1, 2, 6", + "[32] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. VolRecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 6", + "[33] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, 2016. 2, 4, 5, 8", + "[34] Hui Tian, Chenyang Zhu, Yifei Shi, and Kai Xu. SuperUDF: Self-supervised udf estimation for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2023. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. SPARF: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 3", + "[36] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. SparseNeRF: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3", + "[37] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 1, 2, 3, 5, 6", + "[38] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 1", + "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[40] Haoyu Wu, Alexandros Graikos, and Dimitris Samaras. SVolSDF: Sparse multi-view stereo regularization of neural implicit surfaces. International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8", + "[41] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. International Conference on Learning Representations, 2022. 1, 2", + "[42] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2F2NeUS: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 2", + "[43] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3", + "[44] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 2, 5, 6, 7, 8", + "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2", + "[46] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. SparseCraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28523", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 35:25018-25032, 2022. 2, 4, 6, 7, 8", + "[48] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6): 1-13, 2024. 3", + "[49] Yu-Jie Yuan, Yu-Kun Lai, Yi-Hua Huang, Leif Kobbelt, and Lin Gao. Neural radiance fields from sparse rgb-d images for high-quality view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8713-8728, 2022. 3", + "[50] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. The British Machine Vision Conference, 2020. 2, 3, 6", + "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 4, 6, 8", + "[52] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. CoR-GS: sparse-view 3D gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2024. 3", + "[53] Wenyuan Zhang, Ruofan Xing, Yunfan Zeng, Yu-Shen Liu, Kanle Shi, and Zhizhong Han. Fast learning radiance fields by shooting much fewer rays. IEEE Transactions on Image Processing, 2023. 1", + "[54] Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. Neural signed distance function inference through splatting 3D gaussians pulled on zero-level set. In Advances in Neural Information Processing Systems, 2024. 3", + "[55] Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou, Baorui Ma, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. NeRFPrior: Learning neural radiance field as a prior for indoor scene reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11317–11327, 2025. 2", + "[56] Wenyuan Zhang, Yixiao Yang, Han Huang, Liang Han, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. MonoInstance: Enhancing monocular priors via multi-view instance alignment for neural rendering and reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21642-21653, 2025. 2", + "[57] Weiqi Zhang, Junsheng Zhou, Haotian Geng, Wenyuan Zhang, and Yu-Shen Liu. GAP: Gaussianize any point clouds with text guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3", + "[58] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, Yi Fang, and Zhizhong Han. CAP-UDF: Learning unsigned distance functions progressively from raw point clouds with consistency-aware field optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2", + "[59] Junsheng Zhou, Baorui Ma, and Liu Yu-Shen. Fast learning of signed distance functions from noisy point clouds via noise to noise mapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Tiansong Zhou, Jing Huang, Tao Yu, Ruizhi Shao, and Kun Li. HDhuman: High-quality human novel-view rendering from sparse views. IEEE Transactions on Visualization and Computer Graphics, 2023. 1", + "[61] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. FSGS: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 3" + ], + "bbox": [ + 516, + 92, + 905, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28524", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_model.json b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..38ac1fd21acaa9d43dcb86f72b4ad4d1cd00f2aa --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_model.json @@ -0,0 +1,3247 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "SparseRecon: Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.203, + 0.842, + 0.293 + ], + "angle": 0, + "content": "Liang Han\\(^{1}\\), Xu Zhang\\(^{3}\\), Haichuan Song\\(^{2*}\\), Kanle Shi\\(^{4}\\), Yu-Shen Liu\\(^{1*}\\), Zhizhong Han\\(^{5}\\) \n\\(^{1}\\)School of Software, Tsinghua University, Beijing, China \n\\(^{2}\\)Computer Science and Technology, East China Normal University, Shanghai, China \n\\(^{3}\\)China Telecom \\({}^{4}\\)Kuaishou Technology, Beijing, China \n\\(^{5}\\)Department of Computer Science, Wayne State University, Detroit, USA" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.296, + 0.836, + 0.327 + ], + "angle": 0, + "content": "hanl23@mails.tsinghua.edu.cn, zhangxu@chinatelecom.cn, hcsong@cs.ecnu.edu.cn shikanle@kuaishou.com, liuyushen@tsinghua.edu.cn, h312h@wayne.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.396, + 0.485, + 0.759 + ], + "angle": 0, + "content": "Surface reconstruction from sparse views aims to reconstruct a 3D shape or scene from few RGB images. The latest methods are either generalization-based or overfitting-based. However, the generalization-based methods do not generalize well on views that were unseen during training, while the reconstruction quality of overfitting-based methods is still limited by the limited geometry clues. To address this issue, we propose SparseRecon, a novel neural implicit reconstruction method for sparse views with volume rendering-based feature consistency and uncertainty-guided depth constraint. Firstly, we introduce a feature consistency loss across views to constrain the neural implicit field. This design alleviates the ambiguity caused by insufficient consistency information of views and ensures completeness and smoothness in the reconstruction results. Secondly, we employ an uncertainty-guided depth constraint to back up the feature consistency loss in areas with occlusion and insignificant features, which recovers geometry details for better reconstruction quality. Experimental results demonstrate that our method outperforms the state-of-the-art methods, which can produce high-quality geometry with sparse-view input, especially in the scenarios with small overlapping views. Project page: https://hanl2010.github.io/SparseRecon/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.222, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.876 + ], + "angle": 0, + "content": "As one of the important tasks in computer vision, 3D reconstruction has attracted lots of research attentions in recent years. With the advancement of deep learning, 3D reconstruction based on neural implicit representations learned" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Given only 3 input images with large view angle change, our method can reconstruct a smoother surface compared to the state-of-the-art methods, such as UFOREcon [28], S-VolSDF [40] and NeuSurf [13]. The details of each surface are shown in the colored boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.761 + ], + "angle": 0, + "content": "from point clouds [26, 30, 34, 58, 59] or images [23, 37, 45, 60] becomes a popular research topic. Although recent multi-view reconstruction methods[5, 37, 38, 41, 45, 53] have made great progress in terms of the reconstruction quality and reconstruction speed, they require a large number of dense views as supervision. When limited number of views are available, current reconstruction methods usually struggle to reconstruct high-quality surfaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Existing methods for sparse view reconstruction can be mainly classified into two categories: generalization-based methods and overfitting-based methods. The generalization-based methods [22, 24, 28, 31, 32] emphasize the generalization of sparse-view reconstruction, but they are mainly effective in scenarios with large view overlaps. In cases with views that were unseen during training, the quality of the reconstructed surface degenerates significantly, as shown in Figure 1. Meanwhile, it takes a long" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.416, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Yu-Shen Liu and Haichuan Song." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "time to pre-train these methods on large-scale data, but the generalization is still not good enough. Instead, overfitting-based methods [13, 14, 40, 46, 47] typically fit the 3D geometry directly from the sparse views by leveraging geometry clues. They show promising capability of reconstructing higher-quality geometric surfaces with small-overlapping views. However, the reconstruction quality of the existing methods is still unsatisfactory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.305 + ], + "angle": 0, + "content": "In this paper, we introduce a multi-view feature consistency loss based on volume rendering and an uncertainty-guided depth constraint to learn neural signed distance functions. This approach allows us to achieve high-quality mesh reconstruction on more challenging sparse views with small overlaps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.562 + ], + "angle": 0, + "content": "For the feature consistency loss, we first employ the pre-trained Vis-MVSNet [50] to obtain depth features from the input images. Then, within a neural implicit rendering framework, the sampled 3D points along the rays emitted from the reference image are projected to the source image and the reference image. This allows us to acquire source features and reference features of each 3D point and measure the similarity between these two kinds of features. Finally, the feature similarity for each 3D point along the rays is accumulated through volume rendering, thus yielding the feature similarity associated with the rays. During optimization, we pursue higher feature similarity along the rays. Since the depth information is implicitly encoded with image features, feature consistency constraint can significantly alleviate the ambiguity issues arising from insufficient consistency of sparse views and low-texture during reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.822 + ], + "angle": 0, + "content": "For the uncertainty-guided depth prior constraint, we follow MonoSDF [47], utilizing a pre-trained network to acquire depth priors for each image, and then use it to constrain the regions with uncertain depth. However, monocular depth priors do not have consistent scales to the ground truth depth, which are hard to get calibrated to ground truth either due to the distortion. To effectively leverage the depth priors and provide proper supervision for occluded or under-constrained regions, we propose an uncertainty-guided depth prior constraint. First, we calibrate the depth priors using sparse point clouds obtained from COLMAP [33]. Then, during training, we compute the depth confidence from the rendered depth and impose the depth prior constraint only in regions with low confidence. This constraint helps infer more accurate geometry in occluded or under-constrained regions, minimizing the negative impact of depth prior errors on well-constrained regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.482, + 0.868 + ], + "angle": 0, + "content": "We evaluate our methods on several widely used benchmarks and report the state-of-the-art results. In summary, our main contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel feature consistency loss based on volume rendering. It can effectively constrain the neural" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "radiance field by leveraging feature consistency among multiple views, improving the performance in sparse-view reconstruction tasks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- By incorporating depth confidence, we utilize the calibrated depth prior more effectively to enhance geometric constraints, further improving the reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- Extensive experiments on the well-known datasets, such as DTU [16] and BlendedMVS[44], demonstrate that our method outperforms existing sparse-view reconstruction methods and achieves the state-of-the-art results." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.26, + 0.655, + 0.275 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.286, + 0.789, + 0.302 + ], + "angle": 0, + "content": "2.1. Neural Implicit Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.446 + ], + "angle": 0, + "content": "Neural implicit reconstruction methods [5, 7, 21, 37, 41, 47, 55, 56] have been rapidly developed based on neural volume rendering [27]. These methods introduce the Signed Distance Function (SDF) as the implicit representation of 3D surfaces in volume rendering in multi-view 3D reconstruction. While these methods have made significant improvements in both reconstruction quality and speed, it is important to note that they heavily rely on multiple views with large overlaps during the optimization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.905, + 0.658 + ], + "angle": 0, + "content": "Generalization-based surface reconstruction with sparse views. In order to directly generalize the reconstruction results on sparse views, methods [22, 24, 28, 31, 32, 42] adopt the strategy of aggregating features from multiple view images to construct a feature volume, which is then used to predict the SDF for reconstructing the surface. VolRecon [32] uses transformers [17] to aggregate multi-view features, C2F2NeUS [42] employs cascade architecture to construct a volume pyramid, while ReTR [22] and UFORecon [28] aggregates multi-level features. These methods require pretraining on large-scale datasets, which typically takes several days. However, when there is a significant domain gap between the testing and training data, they all fail to reconstruct shapes effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Overfitting-based surface reconstruction with sparse views. In contrast, overfitting-based methods directly fit the 3D geometry from the sparse images by geometric prior constraints. MonoSDF [47] employs depth and normal priors to achieve sparse reconstruction with small-overlapping views. However, such priors come with errors, and it does not fully leverage inter-view consistency, resulting in lower reconstruction quality. S-VolSDF [40] employs probability volumes obtained from MVS [9] models to guide the rendering weight estimated by VolSDF [45]. This improves the reconstruction results in sparse views with small overlap. However, the uncertainties in volumes make negative impact on the reconstructed surface, leading to surface roughness or significant defects. NeuSurf [13] leverages sparse point clouds and employs CAP-UDF [58] to construct an implicit geometric prior to improve the reconstruction qual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.868, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. SparseRecon consists of two main parts. (a) Volume rendering-based feature consistency constraint. We extract features from the reference image and source images. For a ray emitted from the reference image, we project each sampled point on the ray onto the source images to obtain the corresponding features. Then, the volume rendering-based feature consistency loss is calculated using the corresponding features on the reference image. (b) Uncertainty-guided depth prior constraint. We use another pre-trained network to obtain the depth prior of the reference image and calibrate it with the sparse point cloud obtained by COLMAP. Then, we calculate the confidence of the rendered depth, so that the calibrated depth prior only constrains areas with low confidence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.484, + 0.562 + ], + "angle": 0, + "content": "ity from sparse views. However, when the sparse point cloud fails to cover the majority of the object surface, implicit geometric prior information cannot be useful. In contrast, our method employs more robust feature priors, calculates feature consistency based on volume rendering, and simultaneously utilizes depth priors to optimize the occluded regions, ultimately resulting in high-quality surface." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.274, + 0.589 + ], + "angle": 0, + "content": "2.2. Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Gaussian Splatting [18] has achieved unprecedented optimization speed and rendering quality in the task of novel view synthesis [11, 19, 57, 61]. However, since the Gaussians are unorganized, the discrete and unstructured points make it difficult to extract 3D surfaces through postprocessing. To address this issue, some methods introduce to use regularization terms [10], convert 3D Gaussians to 2D surfels [4, 12], acquire opacity fields through rays [48], improve the depth rendering algorithm [1] of 3DGS, or jointly optimize 3DGS with neural radiance fields [2, 20, 25, 54]. However, these methods require dense views to work well. Recently, FatesGS [14] achieves fast sparse-view reconstruction by leveraging depth priors and on-surface feature consistency constraints. Nevertheless, the reconstruction results still exhibit roughness or noticeable defects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.3, + 0.865 + ], + "angle": 0, + "content": "2.3. Sparse View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In addition, the novel view synthesis from sparse views is another category of work closely related to sparse view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.651 + ], + "angle": 0, + "content": "reconstruction. Depending on the technical framework, these works can be categorized into NeRF-based methods [6, 15, 29, 35, 36, 43, 49] and Gaussian Splitting-based methods [3, 11, 19, 52, 61]. This line of research also employs a limited number of views as input. However, they solely focus on the rendering quality of novel views rather than surface reconstruction, which are not designed specifically for the accurate geometric surface reconstruction. Due to the discernible bias (i.e. inherent geometric errors) [37] caused by the conventional volume rendering or inconsistencies in depth that appear in Gaussian rendering, current sparse view synthesis methods still fail to correctly reconstruct high-fidelity geometric surfaces." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.607, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.827 + ], + "angle": 0, + "content": "The overview of our method is depicted in Figure 2. We introduce a novel feature consistency loss and an uncertainty-guided depth constraint based on the NeuS [37] framework. In this section, we first explain how to compute feature consistency for sampled points along rays. Then we explain how to enhance geometric constraints using depth priors and depth uncertainty. Thirdly, we introduce the color consistency loss. Finally, we present the overall loss function for optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.905, + 0.851 + ], + "angle": 0, + "content": "3.1. Volume Rendering-based Feature Consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "First, we use a pre-trained MVS network [50] to extract the features from both the reference image and the source image. Given a ray emitted from the reference image, let \\( p_r(0) \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "denote the point where a ray intersects the reference image. And for each point \\( x_{i} \\) along the ray, we denote its projection on the source image as \\( p_{s}(i) \\). Then, we bilinearly interpolate \\( F_{r}(0) \\) and \\( F_{s}(i) \\) at points \\( p_{r}(0) \\) and \\( p_{s}(i) \\) on image features, respectively. Formally, we define the feature consistency loss function as follows," + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.195, + 0.484, + 0.248 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {f e a t}} = M ^ {\\text {o c c}} \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} w _ {i} f _ {\\text {c o s}} \\left(F _ {r} \\left(p _ {r} (0)\\right), F _ {s} \\left(p _ {s} (i)\\right)\\right), \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.323 + ], + "angle": 0, + "content": "where \\( f_{cos} \\) is the cosine similarity, and \\( w_{i} \\) corresponds to the weight for each point along the ray. \\( p_{s}(i) = K(Rx_{i} + t) \\) is the projection of \\( x_{i} \\) in source view, and \\( [K;R;t] \\) is the camera parameters of source view. \\( M^{occ} \\) is the occlusion mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.58 + ], + "angle": 0, + "content": "Although MVsDF [51], NeuSurf [13] and FatesGS [14] also employ feature consistency constraints, they just leverage the intersection point between a camera ray and the object's surface. Then, this intersection point gets projected onto adjacent views to obtain the corresponding image features for the purpose of comparing features at this point across multiple views. In sparse view scenarios, the estimated positions of surface points can easily deviate significantly, making the on-surface feature consistency loss not converge. NeuSurf [13] and FatesGS [14] utilize sparse point clouds generated by COLMAP [33] as priors, enabling it to obtain partially accurate positions of surface points, thereby allowing the on-surface feature consistency loss to be more effectively leveraged. However, in regions of lacking surface points, the on-surface feature consistency loss cannot ensure the attainment of high-quality geometric surfaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Figure 3 illustrates the difference between on-surface feature consistency and volume rendering-based feature consistency. Due to the uncertainty of gradient direction, the constraint solely relying on surface point features is challenging to be optimized. In contrast, our method does not require the prior estimation of surface points, it calculates feature consistency on all sampling points along the ray, and provides more reasonable and comprehensive supervision to the implicit field, thereby addressing the convergence issue that may arise in sparse reconstruction for MVsDF [51] and NeuSurf [13]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.421, + 0.775 + ], + "angle": 0, + "content": "3.2. Uncertainty-Guided Depth Constraint" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Although multi-view features offer more robust constraint than image colors, they are ineffective for occluded regions. Due to the limited number of views, some regions may only be visible from a single viewpoint. To enhance geometric constraints, we employ depth priors to supervise the radiance field. However, monocular depth priors are not perfect and accurate. Although MonoSDF [47] has already taken the inaccuracy of depth priors into account, i.e., it aligns" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.626, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.174, + 0.609, + 0.194 + ], + "angle": 0, + "content": "(a) Surface Feature Consistency" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.089, + 0.791, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.174, + 0.766, + 0.194 + ], + "angle": 0, + "content": "(b) Volume Rendering-based Feature Consistency" + }, + { + "type": "image_footnote", + "bbox": [ + 0.8, + 0.112, + 0.876, + 0.121 + ], + "angle": 0, + "content": "Sample point" + }, + { + "type": "image_footnote", + "bbox": [ + 0.798, + 0.121, + 0.897, + 0.13 + ], + "angle": 0, + "content": "Gradient direction" + }, + { + "type": "image_footnote", + "bbox": [ + 0.797, + 0.131, + 0.842, + 0.138 + ], + "angle": 0, + "content": "Ray" + }, + { + "type": "image_footnote", + "bbox": [ + 0.797, + 0.139, + 0.896, + 0.148 + ], + "angle": 0, + "content": "Estimated surface" + }, + { + "type": "image_footnote", + "bbox": [ + 0.797, + 0.148, + 0.875, + 0.156 + ], + "angle": 0, + "content": "Real surface" + }, + { + "type": "list", + "bbox": [ + 0.797, + 0.112, + 0.897, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.209, + 0.907, + 0.238 + ], + "angle": 0, + "content": "Figure 3. The illustration of (a) on-surface feature consistency and (b) feature consistency with volume rendering." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.256, + 0.901, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Figure 4. The illustration of rendered depth produced by different depth prior utilization methods, along with the corresponding error maps. (a) Calibrate the depth prior using the rendered depth during training. (b) Calibrate the depth prior using the COLMAP sparse point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.702 + ], + "angle": 0, + "content": "depth priors using rendered depth during training. However, the rendered depth during training is inaccurate, resulting in significant errors in the calibrated depth priors. This ultimately leads to the accumulation of errors during training, which results in inaccurate reconstructions. Figure 4 (a) shows the calibrated depth prior and rendered depth obtained by MonoSDF [47], as well as their error maps compared to the ground truth depth. It can be seen that both the calibrated depth prior and the rendered depth are with large errors. Therefore, MonoSDF [47] uses a weight annealing strategy to anneal the weight of depth loss to 0 during the first 200 training epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Another trivial approach is to calibrate the depth priors using the sparse point cloud obtained from COLMAP [33]. Since the sparse points are generally located on the geometric surface of the object, their depth is relatively accurate. Therefore, calibrating the depth priors using the sparse point cloud can lead to more accurate depth priors. Figure 4 (b) shows the depth priors calibrated with the sparse point cloud, and the rendered depth, as well as their error maps compared to the ground truth depth. It indicates that the depth priors calibrated to the point cloud from the COLMAP [33] are more accurate. Therefore, we can use them as an constraint leads to more precise rendered depth." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "However, due to the distortions in monocular depth pri" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28517" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.485, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.234 + ], + "angle": 0, + "content": "Figure 5. Left: the method of obtaining the confidence of rendered depth. Right: the rendered depth and the depth confidence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.261, + 0.483, + 0.473 + ], + "angle": 0, + "content": "ors, it is impossible to perfectly align them with the ground truth depth. Even after calibration, the depth priors still exhibit noticeable errors when compared to the ground truth depth. In sparse view scenarios, occlusions and insufficient constraints are more common, leading to significant discrepancies between the geometry of occluded regions and the real surface. Therefore, to achieve more accurate geometry in these under-constrained regions while avoiding the negative impact of depth prior errors on well-constrained regions, we propose an uncertainty-guided depth prior constraint method to more effectively utilize the depth priors. Specifically, we apply depth prior constraints in regions with depth uncertainty, while refraining from using them in regions with high depth confidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.566 + ], + "angle": 0, + "content": "To obtain the confidence of the rendered depth, we employ a method to evaluate the multi-view depth projection consistency. As shown in Figure 5, for a specific pixel \\( r_{uv} \\) in the reference image with depth \\( d_r \\), it can be mapped to a neighboring image through the homography matrix \\( H_{rs} \\), leading to a pixel \\( s_{uv} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.581, + 0.482, + 0.597 + ], + "angle": 0, + "content": "\\[\ns _ {u v} = H _ {r s} r _ {u v}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.604, + 0.483, + 0.621 + ], + "angle": 0, + "content": "\\[\nH _ {r s} = M _ {s} M _ {r} ^ {- 1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.75 + ], + "angle": 0, + "content": "where \\( M_r \\) and \\( M_s \\) are the projection matrices corresponding to the reference and source views, respectively. Similarly, we can map the pixel \\( s_{uv} \\) in the source view to the reference view using the projection matrix \\( H_{sr} \\) and its corresponding depth \\( d_s \\), resulting in \\( \\hat{r}_{uv} \\). The forward and backward projection distance error reflect the accuracy of depth predictions, so we take it as the depth confidence, which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.763, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\nC _ {d} = \\left\\{ \\begin{array}{l l} \\frac {1}{e ^ {\\| r _ {u v} - \\hat {r} _ {u v} \\|}}, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| \\leqslant 1 \\\\ 0, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| > 1 \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.841 + ], + "angle": 0, + "content": "The right side of Figure 5 shows the rendered depth and the corresponding depth confidence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Correspondingly, the depth uncertainty is defined as \\( U_{d} = 1 - C_{d} \\). Meanwhile, we can set a threshold \\( \\tau \\) for depth confidence \\( C_d \\) to obtain the occlusion mask \\( M^{occ} = \\{C_d > \\tau\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "For depth calibration, we leverage COLMAP [33] to obtain a sparse point cloud \\(\\{X : x_1, x_2 \\ldots x_i \\in R\\}\\) and visibility flags indicating which keypoints are visible from view \\(I\\). Given the camera parameters \\(P\\) of view \\(I\\), we estimate the depth \\(\\bar{D}_i\\) of keypoints by computing the distance from the visible keypoints \\(x_i\\) to the camera center \\(o\\). Then, we calibrate the monocular depth prior \\(\\hat{D}\\) with \\(\\bar{D}_i\\), it can be defined as \\(\\bar{D} \\approx a\\hat{D} + b\\), where \\(a\\) is the scale factor and \\(b\\) is the shift factor, obtained through the least squares method. Formally, the depth constraint loss is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.249, + 0.907, + 0.285 + ], + "angle": 0, + "content": "\\[\nL _ {d e p t h} = \\sum_ {r \\in R} U _ {d} \\left\\| (a \\hat {D} + b) - D _ {p r e d} \\right\\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.292, + 0.779, + 0.308 + ], + "angle": 0, + "content": "3.3. Color Consistency Constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Although feature consistency constraint can ensure that the reconstruction does not suffer from severe artifacts, it does not provide sufficient supervision to reconstruct fine geometric details. Conversely, in cases with rich textures, image color constraint can refine the geometric details. Therefore, following the NeuralWarp [5], pixel warping loss and patch warping loss are used in our method as multi-view color consistency loss functions," + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {c o l o r} = \\sum_ {r \\in R} M ^ {o c c} d _ {p i x e l} (I (r), I _ {s} (r)) \\\\ + \\sum_ {r \\in R} M ^ {\\text {o c c}} d _ {\\text {p a t c h}} (P (r), P _ {s} (r)), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.639 + ], + "angle": 0, + "content": "where \\( I(r) \\) and \\( I_s(r) \\) are the ground truth color of the pixel from which the ray emits and the rendered color, respectively, \\( P(r) \\) and \\( P_s(r) \\) are the ground truth color of the patch corresponding to the ray and the rendered patch color, respectively. \\( d_{pixel} \\) is the loss metric for pixel color, where we use \\( L1 \\) loss as \\( d_{pixel} \\). \\( d_{patch} \\) is the loss metric for patch color, where we use the Structural Similarity Index Measure (SSIM [39]) as \\( d_{patch} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.658, + 0.661 + ], + "angle": 0, + "content": "3.4. Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.711 + ], + "angle": 0, + "content": "In addition to the above-mentioned three loss functions, we also use the Eikonal loss [8] used in NeuS [37]. We define the overall loss function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.72, + 0.905, + 0.737 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {f e a t}} + \\alpha L _ {\\text {d e p t h}} + L _ {\\text {c o l o r}} + \\beta L _ {\\text {e i k}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.743, + 0.907, + 0.773 + ], + "angle": 0, + "content": "\\(L_{eki}\\) is the Eikonal loss [8], used to regularize the SDF values of sampled points, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.78, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\nL _ {e k i.} = \\frac {1}{m n} \\sum_ {i, k} \\left(\\| \\nabla f \\left(x _ {i, k}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.61, + 0.863 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method on DTU [16] and BlendedMVS [44] dataset. For the DTU [16] dataset, to avoid using the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28518" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.089, + 0.88, + 0.348 + ], + "angle": 0, + "content": "
Methods21243437384082106110114118Mean CD ↓
NeuS [37]5.633.586.004.602.574.531.914.185.461.194.163.98
NeuralWarp [5]2.531.880.741.800.8411.502.642.104.371.192.632.93
MonoSDF [47]4.145.921.394.552.192.142.365.624.581.633.023.41
Vis-MVSNet [50]3.394.440.853.361.693.353.352.342.160.741.832.50
MVSDF [51]4.314.711.656.371.774.473.611.871.671.251.693.03
SparseNeuSft [24]3.484.372.924.762.793.732.801.863.101.152.293.02
VolRecon [32]2.723.071.824.322.143.043.002.562.811.493.222.75
GenSft [31]5.867.673.628.575.375.415.486.045.294.694.355.67
ReTR [22]2.673.371.623.681.873.403.672.842.851.562.352.72
UFORecon [28]1.841.520.792.581.001.821.721.200.930.661.261.39
S-VolSDF [40]2.453.081.333.091.223.211.911.511.230.741.21.91
SparseCraft [46]2.882.420.922.971.582.782.511.105.240.650.882.16
NeuSurf [13]7.601.432.933.181.532.861.861.091.410.370.622.26
FatesGS [14]3.981.322.532.853.362.713.761.490.850.471.062.22
Ours2.141.260.721.460.861.391.370.940.770.440.831.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.358, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Table 1. Quantitative results of Chamfer Distance (CD↓) on DTU dataset with 3 small-overlapping images. The methods are divided into three categories, from top to bottom: (1) dense-view reconstruction methods related to ours, (2) generalization-based sparse-view reconstruction methods, and (3) overfitting-based sparse-view reconstruction methods. the best results are in bold, the second best are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.531 + ], + "angle": 0, + "content": "scenes that have already been used as training data on the pretrained Vis-MVSNet [50] model, we select the same 11 scenes as in S-VolSDF [40]. The image resolution is set to \\(1600 \\times 1200\\). Similar to the S-VolSDF [40] and NeuSurf [13] methods, we select the views 22, 25, and 28 for the more challenging reconstruction of small overlaps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.593 + ], + "angle": 0, + "content": "For the BlendedMVS [44] dataset, we follow the S-VolSDF [40] to use the same 9 challenging scenes, with 3 small-overlapping views for each scene. The image resolution is set to \\(768 \\times 576\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.308, + 0.62 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We use the same network architecture and initialization strategy as NeuS [37] and incorporated our volume rendering feature consistency loss, uncertainty-guided depth constraint loss, and color consistency loss. For the weight factors in the loss functions Eq. 7, we set the \\(\\alpha\\) for the uncertainty-guided depth prior constraint loss \\(L_{depth}\\) to 0.5 and the \\(\\beta\\) for the Eikonal loss \\(L_{eik}\\) to 0.1. Each scene is trained 100K iterations on a RTX3090 GPU. The patch warping term in the color consistency loss requires the surface point normals to calculate homographies, but the initial normals are too noisy [5], therefore, the patch warping loss is applied after 20k training steps. The threshold \\(\\tau\\) of the occlusion mask is set to 0." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.192, + 0.848 + ], + "angle": 0, + "content": "4.3. Baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We compare our approach with three categories of methods including dense-view methods: NeuS [37], NeuralWarp [5], Vis-MVSNet [50], MVsDF [51], generalization-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.532 + ], + "angle": 0, + "content": "methods: SparseNeuS [24], VolRecon [32], GenS [31], ReTR [22] and UFORecon [28], overfitting-based methods: S-VolSDF [40], SparseCraft [46], NeuSurf [13] and FatesGS [14]. The reconstruction results for SparseNeuS [24] and GenS [31] are fine-tuned using 3 views for each scene." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.655, + 0.56 + ], + "angle": 0, + "content": "4.4. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Reconstruction on DTU. For a comprehensive comparison, we evaluate the baselines and our method on both small-overlapping and large-overlapping views. Following baselines [13, 14, 24], we report the Chamfer Distance (CD) between the reconstruction surfaces and the ground truth point clouds. Since the scenes in our experiments differ from those used in NeuSurf [13] and FatesGS [14], we reran the official implementations of NeuSurf and FatesGS. The CD results with small overlapping views are shown in Table 1. The meshes reconstructed by several methods using 3 views with small overlapping are shown in Fig. 6. For the generalization-based sparse reconstruction methods, we only show the reconstruction results of the latest UFORecon [28], as the reconstruction quality of other methods is lower than that of UFORecon [28]. The experimental results show that our method significantly improves the mesh quality with small overlap views, compared to the state-of-the-art sparse-view reconstruction methods. The results of large overlapping views are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 6, when input sparse views with small overlap, both MonoSDF [47] and SparseCraft [46]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.237, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.314, + 0.185, + 0.324 + ], + "angle": 0, + "content": "GT images" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.088, + 0.334, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.147, + 0.329, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.201, + 0.323, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.257, + 0.323, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.313, + 0.307, + 0.323 + ], + "angle": 0, + "content": "MonoSDF" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.425, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.146, + 0.424, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.201, + 0.418, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.256, + 0.417, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.313, + 0.405, + 0.323 + ], + "angle": 0, + "content": "UFORecon" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.089, + 0.519, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.146, + 0.519, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.201, + 0.502, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.438, + 0.256, + 0.514, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.313, + 0.499, + 0.322 + ], + "angle": 0, + "content": "S-VolSDF" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.09, + 0.616, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.146, + 0.613, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.201, + 0.598, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.256, + 0.619, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.313, + 0.6, + 0.323 + ], + "angle": 0, + "content": "SparseCraft" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.089, + 0.712, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.146, + 0.71, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.201, + 0.703, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.256, + 0.703, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.313, + 0.686, + 0.322 + ], + "angle": 0, + "content": "NeuSurf" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.09, + 0.805, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.146, + 0.803, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.201, + 0.791, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.257, + 0.797, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.313, + 0.78, + 0.322 + ], + "angle": 0, + "content": "FatesGS" + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.091, + 0.901, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.146, + 0.901, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.201, + 0.886, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.256, + 0.895, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.314, + 0.868, + 0.323 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.34, + 0.732, + 0.355 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on DTU dataset with 3 small-overlapping images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.547 + ], + "angle": 0, + "content": "suffer from reconstruction ambiguity and failures, highlighting that relying solely on simplistic geometric prior constraints is insufficient to obtain complete and accurate meshes. UFORecon [28] shows significant roughness in its reconstruction results. S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit noticeable reconstruction defects. Experimental results demonstrate that our method is effective in alleviating geometric and appearance ambiguities during the optimization process. This significantly enhances the quality of mesh reconstruction, especially in scenarios with small overlapping views and low texture." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.561, + 0.186, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.702, + 0.163, + 0.711 + ], + "angle": 0, + "content": "GT Images" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.561, + 0.234, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.677, + 0.223, + 0.687 + ], + "angle": 0, + "content": "X" + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.699, + 0.235, + 0.71 + ], + "angle": 0, + "content": "MonoSDF" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.562, + 0.3, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.664, + 0.301, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.702, + 0.294, + 0.711 + ], + "angle": 0, + "content": "S-VolSDF" + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.562, + 0.35, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.664, + 0.361, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.702, + 0.351, + 0.711 + ], + "angle": 0, + "content": "NeuSurf" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.562, + 0.418, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.664, + 0.419, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.702, + 0.41, + 0.711 + ], + "angle": 0, + "content": "FatesGS" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.561, + 0.482, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.702, + 0.462, + 0.711 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.755 + ], + "angle": 0, + "content": "Figure 7. Visual comparison on BlendedMVS dataset. \\(\\times\\) indicates reconstruction failure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Reconstruction on BlendedMVS. Figure 7 presents the visual comparison of reconstructed mesh for overfitting-based methods. With only 3 small-overlapping views provided, all of the generalization-based methods completely fail to reconstruct in the sparse setting of BlendedMVS[44] dataset, even if SparseNeuS [24] is fine-tuned. Therefore, the reconstruction results of these methods are not included" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.577 + ], + "angle": 0, + "content": "in Figure 7. Compared to other methods, our approach can generate more complete and detailed meshes. Similarly, MonoSDF [47] fails to reconstruct surfaces either. The meshes generated by S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit significant defects. Both NeurSurf [13] and FatesGS [14] use on-surface feature consistency constraints, but the reconstruction results are still not good enough. In contrast, our method achieves more comprehensive geometry and finer details by employing volume rendering-based feature consistency constraints. This highlights the advantages of our approach in geometric consistency. More visualizations are presented in the supplementary materials." + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.602, + 0.534, + 0.631 + ], + "angle": 270, + "content": "Scan 40" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.589, + 0.615, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.653, + 0.534, + 0.686 + ], + "angle": 270, + "content": "Scan 110" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.642, + 0.618, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.698, + 0.614, + 0.72 + ], + "angle": 0, + "content": "Mesh w/ \\(L_{\\text{feat}}\\) by ours" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.589, + 0.703, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.643, + 0.71, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.698, + 0.711, + 0.72 + ], + "angle": 0, + "content": "Error map w/ \\(L_{\\text{feat}}\\) by ours" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.589, + 0.801, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.644, + 0.803, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.702, + 0.697, + 0.795, + 0.721 + ], + "angle": 0, + "content": "at Mesh w/ surf \\(f_{\\text{feat}}\\) by MVSDF" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.589, + 0.894, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.644, + 0.896, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.697, + 0.897, + 0.721 + ], + "angle": 0, + "content": "Error map w/ \\(L_{\\text{feat}}^{\\text{surf}}\\) by MVSDF" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.737, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Figure 8. Reconstructed meshes and error maps on DTU dataset with different feature consistency losses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.667, + 0.805 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We evaluate the components of our method with 3 small-overlapping views by an ablation study on the DTU [16] dataset. To compare the depth loss \\( L_{depth}^{mono} \\) calibrated by rendered depth in MonoSDF [47] with our depth loss \\( L_{depth} \\), we replace \\( L_{depth} \\) with \\( L_{depth}^{mono} \\) to evaluate it in our method. We also compare the volume rendering-based" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28520" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.483, + 0.201 + ], + "angle": 0, + "content": "
MethodLcolorLfeatLdepthLmono depthL1featL2featsurffeatCD↓
Baseline3.35
✓1.76
✓✓1.47
✓✓1.62
✓✓✓1.11
✓✓✓1.59
✓✓✓2.36
✓✓✓1.81
✓✓✓2.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.211, + 0.483, + 0.24 + ], + "angle": 0, + "content": "Table 2. Ablation studies on DTU dataset with 3 small-overlapping images." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.256, + 0.19, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.372, + 0.16, + 0.381 + ], + "angle": 0, + "content": "Step 10k" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.256, + 0.288, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.314, + 0.288, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.372, + 0.258, + 0.381 + ], + "angle": 0, + "content": "Step 20k" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.256, + 0.385, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.314, + 0.385, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.372, + 0.354, + 0.381 + ], + "angle": 0, + "content": "Step 40k" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.256, + 0.482, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.314, + 0.482, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.372, + 0.451, + 0.381 + ], + "angle": 0, + "content": "Step 60k" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.425 + ], + "angle": 0, + "content": "Figure 9. The variation of weighted feature similarity during training, brighter colors indicate higher feature similarity." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.441, + 0.482, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Figure 10. Visualization of reconstruction and error maps for scene scan24 and scan37 in DTU dataset with different losses. The differences of error maps are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.837 + ], + "angle": 0, + "content": "feature consistency loss calculated using L1 distance (denoted as \\(L_{\\text{feat}}^{L1}\\)) and L2 distance (denoted as \\(L_{\\text{feat}}^{L2}\\)) with our method using feature similarity distance. We found that feature similarity distance is better than both L1 and L2 distance, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.839, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In addition, we replace our volume rendering-based feature consistency loss \\( L_{feat} \\) with the on-surface feature consistency loss \\( L_{feat}^{surf} \\) used in MVSDF [51] to compare the effectiveness of two different loss functions. Figure 8 illustrates the performance of these two different loss functions." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.625, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.133, + 0.594, + 0.143 + ], + "angle": 0, + "content": "GT Images" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.693, + 0.128 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.133, + 0.68, + 0.142 + ], + "angle": 0, + "content": "S-VolSDF" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.092, + 0.763, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.133, + 0.746, + 0.142 + ], + "angle": 0, + "content": "NeuSurf" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.092, + 0.833, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.133, + 0.816, + 0.142 + ], + "angle": 0, + "content": "FatesGS" + }, + { + "type": "image", + "bbox": [ + 0.835, + 0.092, + 0.904, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.858, + 0.133, + 0.88, + 0.142 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.157, + 0.905, + 0.187 + ], + "angle": 0, + "content": "Figure 11. Failure case. For specular objects, the ambiguity in the color consistency constraint may lead to a rough surface." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "trates the reconstruction results and error maps on the DTU dataset when using different feature consistency losses, under the on-surface feature consistency loss \\( L_{feat}^{surf} \\), the meshes show large artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Table 2 shows the average Chamfer Distance over all 11 scenes on DTU dataset using different losses. The experimental results indicate that both feature consistency loss and uncertainty-guided depth constraint improve the surface reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.905, + 0.423 + ], + "angle": 0, + "content": "Figure 9 illustrates the variation of the weighted feature similarity map during the training process. Brighter colors indicate higher feature similarity, demonstrating that our volume rendering-based feature consistency loss can provide effective constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.533 + ], + "angle": 0, + "content": "Figure 10 shows the reconstructed meshes and error maps for scene scan24 and scan37 on the DTU [16] dataset when using different losses. It can be observed that the mesh deteriorates without the volume rendering-based feature consistency loss or the uncertainty-guided depth constraint loss, and the reconstruction quality drops when using the depth loss \\( L_{depth}^{mono} \\) in MonoSDF [47]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.64, + 0.558 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.689 + ], + "angle": 0, + "content": "We propose a novel method for learning implicit representations from sparse views with small overlaps. Our novelty lies in a novel volume rendering-based feature consistency loss and an uncertainty-guided depth constraint. Extensive experiments on the DTU [16] and BlendedMVS [44] datasets show that our method surpasses existing state-of-the-art sparse-view reconstruction methods in terms of reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitations. Although our method shows significant improvement over other sparse view reconstruction methods, there are still some limitations. Firstly, for specular objects, the ambiguity in the color consistency constraint may lead to a rough surface, as shown in Figure 11. Secondly, Following previous studies [13, 24, 40], the camera poses of sparse views are obtained from the training dataset. However, in some cases, it may not be possible to obtain accurate camera poses using SfM methods like COLMAP [33] due to the lack of texture in the images or excessive viewing angles. Additionally, the feature consistency constraint method requires a pre-trained network to extract image features. The accuracy of the features determines the performance of the feature consistency constraint." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28521" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.262, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "The corresponding authors are Yu-Shen Liu and Haichuan Song. This work was supported by National Key R&D Program of China (2022YFC3800600), and the National Natural Science Foundation of China (62272263), and in part by Tsinghua-Kuaishou Institute of Future Media Data." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[1] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, PP: 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.318, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[2] Hanlin Chen, Chen Li, and Gim Hee Lee. NeuSG: Neural implicit surface reconstruction with 3D gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.361, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. MVSplat: Efficient 3D gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[4] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[5] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[7] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[9] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[11] Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3D gaussian splattering with view consistency for sparse view synthesis. In Advances in Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[13] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. NeuSurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[14] Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a Diet: Semantically consistent few-shot view synthesis implementation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanas. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[17] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.906, + 0.685 + ], + "angle": 0, + "content": "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[20] Shujuan Li, Yu-Shen Liu, and Zhizhong Han. GaussianUDF: Inferring unsigned distance functions through 3D gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 27113-27123, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[22] Yixun Liang, Hao He, and Yingcong Chen. ReTR: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Xinqi Liu, Jituo Li, and Guodong Lu. Reconstructing complex shaped clothing from a single image with feature stable" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28522" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "unsigned distance fields. IEEE Transactions on Visualization and Computer Graphics, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. SparseNeuS: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[25] Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 3DGSR: Implicit surface reconstruction with 3D gaussian splatting. ACM Transactions on Graphics, 43(6):1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[26] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3D point clouds via noise to noise mapping. In International Conference on Machine Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[28] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. UFORecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[29] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[30] Takeshi Noda, Chao Chen, Junsheng Zhou, Weiqi Zhang, Yu-Shen Liu, and Zhizhong Han. Learning bijective surface parameterization for inferring signed distance functions from sparse point clouds with grid deformation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22139-22149, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[31] Rui Peng, Xiaodong Gu, Luyang Tang, Shihe Shen, Fanqi Yu, and Ronggang Wang. GenS: Generalizable neural surface reconstruction from multi-view images. In Advances in Neural Information Processing Systems, pages 56932-56945, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[32] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. VolRecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[33] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, 2016. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Hui Tian, Chenyang Zhu, Yifei Shi, and Kai Xu. SuperUDF: Self-supervised udf estimation for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. SPARF: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[36] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. SparseNeRF: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[37] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[38] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[40] Haoyu Wu, Alexandros Graikos, and Dimitris Samaras. SVolSDF: Sparse multi-view stereo regularization of neural implicit surfaces. International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[41] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. International Conference on Learning Representations, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[42] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2F2NeUS: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[43] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. SparseCraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28523" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[47] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 35:25018-25032, 2022. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.162, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[48] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6): 1-13, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[49] Yu-Jie Yuan, Yu-Kun Lai, Yi-Hua Huang, Leif Kobbelt, and Lin Gao. Neural radiance fields from sparse rgb-d images for high-quality view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8713-8728, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.289, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[50] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. The British Machine Vision Conference, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.33, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.386, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[52] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. CoR-GS: sparse-view 3D gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.442, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[53] Wenyuan Zhang, Ruofan Xing, Yunfan Zeng, Yu-Shen Liu, Kanle Shi, and Zhizhong Han. Fast learning radiance fields by shooting much fewer rays. IEEE Transactions on Image Processing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.498, + 0.483, + 0.553 + ], + "angle": 0, + "content": "[54] Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. Neural signed distance function inference through splatting 3D gaussians pulled on zero-level set. In Advances in Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.554, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[55] Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou, Baorui Ma, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. NeRFPrior: Learning neural radiance field as a prior for indoor scene reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11317–11327, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.636, + 0.483, + 0.72 + ], + "angle": 0, + "content": "[56] Wenyuan Zhang, Yixiao Yang, Han Huang, Liang Han, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. MonoInstance: Enhancing monocular priors via multi-view instance alignment for neural rendering and reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21642-21653, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[57] Weiqi Zhang, Junsheng Zhou, Haotian Geng, Wenyuan Zhang, and Yu-Shen Liu. GAP: Gaussianize any point clouds with text guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[58] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, Yi Fang, and Zhizhong Han. CAP-UDF: Learning unsigned distance functions progressively from raw point clouds with consistency-aware field optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[59] Junsheng Zhou, Baorui Ma, and Liu Yu-Shen. Fast learning of signed distance functions from noisy point clouds via noise to noise mapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.149 + ], + "angle": 0, + "content": "[60] Tiansong Zhou, Jing Huang, Tao Yu, Ruizhi Shao, and Kun Li. HDhuman: High-quality human novel-view rendering from sparse views. IEEE Transactions on Visualization and Computer Graphics, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[61] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. FSGS: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28524" + } + ] +] \ No newline at end of file diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_origin.pdf b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3babda129d9530af3f00ed7ddb5f3a21d9f78c3 --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/0728af14-5300-46a3-bf64-7d187d835efb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bb4909c0484da20a0c8ab26a942bbd98bc425f2c5579c58b56aa9d55f4f2713 +size 31336952 diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/full.md b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f656a8bf5384bbf22fb192aa53765dbda1a42d56 --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/full.md @@ -0,0 +1,455 @@ +# SparseRecon: Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies + +Liang Han $^{1}$ , Xu Zhang $^{3}$ , Haichuan Song $^{2*}$ , Kanle Shi $^{4}$ , Yu-Shen Liu $^{1*}$ , Zhizhong Han $^{5}$ $^{1}$ School of Software, Tsinghua University, Beijing, China + $^{2}$ Computer Science and Technology, East China Normal University, Shanghai, China + $^{3}$ China Telecom ${}^{4}$ Kuaishou Technology, Beijing, China + $^{5}$ Department of Computer Science, Wayne State University, Detroit, USA + +hanl23@mails.tsinghua.edu.cn, zhangxu@chinatelecom.cn, hcsong@cs.ecnu.edu.cn shikanle@kuaishou.com, liuyushen@tsinghua.edu.cn, h312h@wayne.edu + +# Abstract + +Surface reconstruction from sparse views aims to reconstruct a 3D shape or scene from few RGB images. The latest methods are either generalization-based or overfitting-based. However, the generalization-based methods do not generalize well on views that were unseen during training, while the reconstruction quality of overfitting-based methods is still limited by the limited geometry clues. To address this issue, we propose SparseRecon, a novel neural implicit reconstruction method for sparse views with volume rendering-based feature consistency and uncertainty-guided depth constraint. Firstly, we introduce a feature consistency loss across views to constrain the neural implicit field. This design alleviates the ambiguity caused by insufficient consistency information of views and ensures completeness and smoothness in the reconstruction results. Secondly, we employ an uncertainty-guided depth constraint to back up the feature consistency loss in areas with occlusion and insignificant features, which recovers geometry details for better reconstruction quality. Experimental results demonstrate that our method outperforms the state-of-the-art methods, which can produce high-quality geometry with sparse-view input, especially in the scenarios with small overlapping views. Project page: https://hanl2010.github.io/SparseRecon/. + +# 1. Introduction + +As one of the important tasks in computer vision, 3D reconstruction has attracted lots of research attentions in recent years. With the advancement of deep learning, 3D reconstruction based on neural implicit representations learned + +![](images/9d2d98f09d624cff078923207d92e2931e107eb3db4bff04f48aa31c9e6f7363.jpg) +Figure 1. Given only 3 input images with large view angle change, our method can reconstruct a smoother surface compared to the state-of-the-art methods, such as UFOREcon [28], S-VolSDF [40] and NeuSurf [13]. The details of each surface are shown in the colored boxes. + +from point clouds [26, 30, 34, 58, 59] or images [23, 37, 45, 60] becomes a popular research topic. Although recent multi-view reconstruction methods[5, 37, 38, 41, 45, 53] have made great progress in terms of the reconstruction quality and reconstruction speed, they require a large number of dense views as supervision. When limited number of views are available, current reconstruction methods usually struggle to reconstruct high-quality surfaces. + +Existing methods for sparse view reconstruction can be mainly classified into two categories: generalization-based methods and overfitting-based methods. The generalization-based methods [22, 24, 28, 31, 32] emphasize the generalization of sparse-view reconstruction, but they are mainly effective in scenarios with large view overlaps. In cases with views that were unseen during training, the quality of the reconstructed surface degenerates significantly, as shown in Figure 1. Meanwhile, it takes a long + +time to pre-train these methods on large-scale data, but the generalization is still not good enough. Instead, overfitting-based methods [13, 14, 40, 46, 47] typically fit the 3D geometry directly from the sparse views by leveraging geometry clues. They show promising capability of reconstructing higher-quality geometric surfaces with small-overlapping views. However, the reconstruction quality of the existing methods is still unsatisfactory. + +In this paper, we introduce a multi-view feature consistency loss based on volume rendering and an uncertainty-guided depth constraint to learn neural signed distance functions. This approach allows us to achieve high-quality mesh reconstruction on more challenging sparse views with small overlaps. + +For the feature consistency loss, we first employ the pre-trained Vis-MVSNet [50] to obtain depth features from the input images. Then, within a neural implicit rendering framework, the sampled 3D points along the rays emitted from the reference image are projected to the source image and the reference image. This allows us to acquire source features and reference features of each 3D point and measure the similarity between these two kinds of features. Finally, the feature similarity for each 3D point along the rays is accumulated through volume rendering, thus yielding the feature similarity associated with the rays. During optimization, we pursue higher feature similarity along the rays. Since the depth information is implicitly encoded with image features, feature consistency constraint can significantly alleviate the ambiguity issues arising from insufficient consistency of sparse views and low-texture during reconstruction. + +For the uncertainty-guided depth prior constraint, we follow MonoSDF [47], utilizing a pre-trained network to acquire depth priors for each image, and then use it to constrain the regions with uncertain depth. However, monocular depth priors do not have consistent scales to the ground truth depth, which are hard to get calibrated to ground truth either due to the distortion. To effectively leverage the depth priors and provide proper supervision for occluded or under-constrained regions, we propose an uncertainty-guided depth prior constraint. First, we calibrate the depth priors using sparse point clouds obtained from COLMAP [33]. Then, during training, we compute the depth confidence from the rendered depth and impose the depth prior constraint only in regions with low confidence. This constraint helps infer more accurate geometry in occluded or under-constrained regions, minimizing the negative impact of depth prior errors on well-constrained regions. + +We evaluate our methods on several widely used benchmarks and report the state-of-the-art results. In summary, our main contributions are as follows. + +- We propose a novel feature consistency loss based on volume rendering. It can effectively constrain the neural + +radiance field by leveraging feature consistency among multiple views, improving the performance in sparse-view reconstruction tasks. + +- By incorporating depth confidence, we utilize the calibrated depth prior more effectively to enhance geometric constraints, further improving the reconstruction quality. +- Extensive experiments on the well-known datasets, such as DTU [16] and BlendedMVS[44], demonstrate that our method outperforms existing sparse-view reconstruction methods and achieves the state-of-the-art results. + +# 2. Related Work + +# 2.1. Neural Implicit Reconstruction + +Neural implicit reconstruction methods [5, 7, 21, 37, 41, 47, 55, 56] have been rapidly developed based on neural volume rendering [27]. These methods introduce the Signed Distance Function (SDF) as the implicit representation of 3D surfaces in volume rendering in multi-view 3D reconstruction. While these methods have made significant improvements in both reconstruction quality and speed, it is important to note that they heavily rely on multiple views with large overlaps during the optimization. + +Generalization-based surface reconstruction with sparse views. In order to directly generalize the reconstruction results on sparse views, methods [22, 24, 28, 31, 32, 42] adopt the strategy of aggregating features from multiple view images to construct a feature volume, which is then used to predict the SDF for reconstructing the surface. VolRecon [32] uses transformers [17] to aggregate multi-view features, C2F2NeUS [42] employs cascade architecture to construct a volume pyramid, while ReTR [22] and UFORecon [28] aggregates multi-level features. These methods require pretraining on large-scale datasets, which typically takes several days. However, when there is a significant domain gap between the testing and training data, they all fail to reconstruct shapes effectively. + +Overfitting-based surface reconstruction with sparse views. In contrast, overfitting-based methods directly fit the 3D geometry from the sparse images by geometric prior constraints. MonoSDF [47] employs depth and normal priors to achieve sparse reconstruction with small-overlapping views. However, such priors come with errors, and it does not fully leverage inter-view consistency, resulting in lower reconstruction quality. S-VolSDF [40] employs probability volumes obtained from MVS [9] models to guide the rendering weight estimated by VolSDF [45]. This improves the reconstruction results in sparse views with small overlap. However, the uncertainties in volumes make negative impact on the reconstructed surface, leading to surface roughness or significant defects. NeuSurf [13] leverages sparse point clouds and employs CAP-UDF [58] to construct an implicit geometric prior to improve the reconstruction qual + +![](images/bcd4dac80121e96e064f3d1eef9554e5730b2d82f3481f549bbb0abf8d6f5194.jpg) +Figure 2. SparseRecon consists of two main parts. (a) Volume rendering-based feature consistency constraint. We extract features from the reference image and source images. For a ray emitted from the reference image, we project each sampled point on the ray onto the source images to obtain the corresponding features. Then, the volume rendering-based feature consistency loss is calculated using the corresponding features on the reference image. (b) Uncertainty-guided depth prior constraint. We use another pre-trained network to obtain the depth prior of the reference image and calibrate it with the sparse point cloud obtained by COLMAP. Then, we calculate the confidence of the rendered depth, so that the calibrated depth prior only constrains areas with low confidence. + +ity from sparse views. However, when the sparse point cloud fails to cover the majority of the object surface, implicit geometric prior information cannot be useful. In contrast, our method employs more robust feature priors, calculates feature consistency based on volume rendering, and simultaneously utilizes depth priors to optimize the occluded regions, ultimately resulting in high-quality surface. + +# 2.2. Gaussian Splitting + +Gaussian Splatting [18] has achieved unprecedented optimization speed and rendering quality in the task of novel view synthesis [11, 19, 57, 61]. However, since the Gaussians are unorganized, the discrete and unstructured points make it difficult to extract 3D surfaces through postprocessing. To address this issue, some methods introduce to use regularization terms [10], convert 3D Gaussians to 2D surfels [4, 12], acquire opacity fields through rays [48], improve the depth rendering algorithm [1] of 3DGS, or jointly optimize 3DGS with neural radiance fields [2, 20, 25, 54]. However, these methods require dense views to work well. Recently, FatesGS [14] achieves fast sparse-view reconstruction by leveraging depth priors and on-surface feature consistency constraints. Nevertheless, the reconstruction results still exhibit roughness or noticeable defects. + +# 2.3. Sparse View Synthesis + +In addition, the novel view synthesis from sparse views is another category of work closely related to sparse view + +reconstruction. Depending on the technical framework, these works can be categorized into NeRF-based methods [6, 15, 29, 35, 36, 43, 49] and Gaussian Splitting-based methods [3, 11, 19, 52, 61]. This line of research also employs a limited number of views as input. However, they solely focus on the rendering quality of novel views rather than surface reconstruction, which are not designed specifically for the accurate geometric surface reconstruction. Due to the discernible bias (i.e. inherent geometric errors) [37] caused by the conventional volume rendering or inconsistencies in depth that appear in Gaussian rendering, current sparse view synthesis methods still fail to correctly reconstruct high-fidelity geometric surfaces. + +# 3. Method + +The overview of our method is depicted in Figure 2. We introduce a novel feature consistency loss and an uncertainty-guided depth constraint based on the NeuS [37] framework. In this section, we first explain how to compute feature consistency for sampled points along rays. Then we explain how to enhance geometric constraints using depth priors and depth uncertainty. Thirdly, we introduce the color consistency loss. Finally, we present the overall loss function for optimization. + +# 3.1. Volume Rendering-based Feature Consistency + +First, we use a pre-trained MVS network [50] to extract the features from both the reference image and the source image. Given a ray emitted from the reference image, let $p_r(0)$ + +denote the point where a ray intersects the reference image. And for each point $x_{i}$ along the ray, we denote its projection on the source image as $p_{s}(i)$ . Then, we bilinearly interpolate $F_{r}(0)$ and $F_{s}(i)$ at points $p_{r}(0)$ and $p_{s}(i)$ on image features, respectively. Formally, we define the feature consistency loss function as follows, + +$$ +L _ {\text {f e a t}} = M ^ {\text {o c c}} \left(1 - \frac {1}{N} \sum_ {i = 1} ^ {N} w _ {i} f _ {\text {c o s}} \left(F _ {r} \left(p _ {r} (0)\right), F _ {s} \left(p _ {s} (i)\right)\right), \right. \tag {1} +$$ + +where $f_{cos}$ is the cosine similarity, and $w_{i}$ corresponds to the weight for each point along the ray. $p_{s}(i) = K(Rx_{i} + t)$ is the projection of $x_{i}$ in source view, and $[K;R;t]$ is the camera parameters of source view. $M^{occ}$ is the occlusion mask. + +Although MVsDF [51], NeuSurf [13] and FatesGS [14] also employ feature consistency constraints, they just leverage the intersection point between a camera ray and the object's surface. Then, this intersection point gets projected onto adjacent views to obtain the corresponding image features for the purpose of comparing features at this point across multiple views. In sparse view scenarios, the estimated positions of surface points can easily deviate significantly, making the on-surface feature consistency loss not converge. NeuSurf [13] and FatesGS [14] utilize sparse point clouds generated by COLMAP [33] as priors, enabling it to obtain partially accurate positions of surface points, thereby allowing the on-surface feature consistency loss to be more effectively leveraged. However, in regions of lacking surface points, the on-surface feature consistency loss cannot ensure the attainment of high-quality geometric surfaces. + +Figure 3 illustrates the difference between on-surface feature consistency and volume rendering-based feature consistency. Due to the uncertainty of gradient direction, the constraint solely relying on surface point features is challenging to be optimized. In contrast, our method does not require the prior estimation of surface points, it calculates feature consistency on all sampling points along the ray, and provides more reasonable and comprehensive supervision to the implicit field, thereby addressing the convergence issue that may arise in sparse reconstruction for MVsDF [51] and NeuSurf [13]. + +# 3.2. Uncertainty-Guided Depth Constraint + +Although multi-view features offer more robust constraint than image colors, they are ineffective for occluded regions. Due to the limited number of views, some regions may only be visible from a single viewpoint. To enhance geometric constraints, we employ depth priors to supervise the radiance field. However, monocular depth priors are not perfect and accurate. Although MonoSDF [47] has already taken the inaccuracy of depth priors into account, i.e., it aligns + +![](images/65ace355b98a0bfb4bd893044bbddcc65d9f294e9ecc0f00640d3e80b9a2a34b.jpg) +(a) Surface Feature Consistency + +(b) Volume Rendering-based Feature Consistency +![](images/f37d7a6ca0870534856c8febb92469a26e0942dde5a49d2c566314facffb4605.jpg) +Sample point +Gradient direction +Ray +Estimated surface +Real surface + +![](images/66beb6e5d38527afcec9dbbf16ca9c447fabacbfbe81f20dc153aff7fdaeef00.jpg) +Figure 3. The illustration of (a) on-surface feature consistency and (b) feature consistency with volume rendering. +Figure 4. The illustration of rendered depth produced by different depth prior utilization methods, along with the corresponding error maps. (a) Calibrate the depth prior using the rendered depth during training. (b) Calibrate the depth prior using the COLMAP sparse point cloud. + +depth priors using rendered depth during training. However, the rendered depth during training is inaccurate, resulting in significant errors in the calibrated depth priors. This ultimately leads to the accumulation of errors during training, which results in inaccurate reconstructions. Figure 4 (a) shows the calibrated depth prior and rendered depth obtained by MonoSDF [47], as well as their error maps compared to the ground truth depth. It can be seen that both the calibrated depth prior and the rendered depth are with large errors. Therefore, MonoSDF [47] uses a weight annealing strategy to anneal the weight of depth loss to 0 during the first 200 training epochs. + +Another trivial approach is to calibrate the depth priors using the sparse point cloud obtained from COLMAP [33]. Since the sparse points are generally located on the geometric surface of the object, their depth is relatively accurate. Therefore, calibrating the depth priors using the sparse point cloud can lead to more accurate depth priors. Figure 4 (b) shows the depth priors calibrated with the sparse point cloud, and the rendered depth, as well as their error maps compared to the ground truth depth. It indicates that the depth priors calibrated to the point cloud from the COLMAP [33] are more accurate. Therefore, we can use them as an constraint leads to more precise rendered depth. + +However, due to the distortions in monocular depth pri + +![](images/9fad56129eeb94608cdb0ba400507bb5d1df0d83aff02e24730ee2c329fa0804.jpg) +Figure 5. Left: the method of obtaining the confidence of rendered depth. Right: the rendered depth and the depth confidence. + +ors, it is impossible to perfectly align them with the ground truth depth. Even after calibration, the depth priors still exhibit noticeable errors when compared to the ground truth depth. In sparse view scenarios, occlusions and insufficient constraints are more common, leading to significant discrepancies between the geometry of occluded regions and the real surface. Therefore, to achieve more accurate geometry in these under-constrained regions while avoiding the negative impact of depth prior errors on well-constrained regions, we propose an uncertainty-guided depth prior constraint method to more effectively utilize the depth priors. Specifically, we apply depth prior constraints in regions with depth uncertainty, while refraining from using them in regions with high depth confidence. + +To obtain the confidence of the rendered depth, we employ a method to evaluate the multi-view depth projection consistency. As shown in Figure 5, for a specific pixel $r_{uv}$ in the reference image with depth $d_r$ , it can be mapped to a neighboring image through the homography matrix $H_{rs}$ , leading to a pixel $s_{uv}$ , + +$$ +s _ {u v} = H _ {r s} r _ {u v}, \tag {2} +$$ + +$$ +H _ {r s} = M _ {s} M _ {r} ^ {- 1}, \tag {3} +$$ + +where $M_r$ and $M_s$ are the projection matrices corresponding to the reference and source views, respectively. Similarly, we can map the pixel $s_{uv}$ in the source view to the reference view using the projection matrix $H_{sr}$ and its corresponding depth $d_s$ , resulting in $\hat{r}_{uv}$ . The forward and backward projection distance error reflect the accuracy of depth predictions, so we take it as the depth confidence, which is defined as + +$$ +C _ {d} = \left\{ \begin{array}{l l} \frac {1}{e ^ {\| r _ {u v} - \hat {r} _ {u v} \|}}, & \text {i f} \| r _ {u v} - \hat {r} _ {u v} \| \leqslant 1 \\ 0, & \text {i f} \| r _ {u v} - \hat {r} _ {u v} \| > 1 \end{array} \right. \tag {4} +$$ + +The right side of Figure 5 shows the rendered depth and the corresponding depth confidence. + +Correspondingly, the depth uncertainty is defined as $U_{d} = 1 - C_{d}$ . Meanwhile, we can set a threshold $\tau$ for depth confidence $C_d$ to obtain the occlusion mask $M^{occ} = \{C_d > \tau\}$ . + +For depth calibration, we leverage COLMAP [33] to obtain a sparse point cloud $\{X : x_1, x_2 \ldots x_i \in R\}$ and visibility flags indicating which keypoints are visible from view $I$ . Given the camera parameters $P$ of view $I$ , we estimate the depth $\bar{D}_i$ of keypoints by computing the distance from the visible keypoints $x_i$ to the camera center $o$ . Then, we calibrate the monocular depth prior $\hat{D}$ with $\bar{D}_i$ , it can be defined as $\bar{D} \approx a\hat{D} + b$ , where $a$ is the scale factor and $b$ is the shift factor, obtained through the least squares method. Formally, the depth constraint loss is defined as, + +$$ +L _ {d e p t h} = \sum_ {r \in R} U _ {d} \left\| (a \hat {D} + b) - D _ {p r e d} \right\| ^ {2}. \tag {5} +$$ + +# 3.3. Color Consistency Constraint + +Although feature consistency constraint can ensure that the reconstruction does not suffer from severe artifacts, it does not provide sufficient supervision to reconstruct fine geometric details. Conversely, in cases with rich textures, image color constraint can refine the geometric details. Therefore, following the NeuralWarp [5], pixel warping loss and patch warping loss are used in our method as multi-view color consistency loss functions, + +$$ +\begin{array}{l} L _ {c o l o r} = \sum_ {r \in R} M ^ {o c c} d _ {p i x e l} (I (r), I _ {s} (r)) \\ + \sum_ {r \in R} M ^ {\text {o c c}} d _ {\text {p a t c h}} (P (r), P _ {s} (r)), \tag {6} \\ \end{array} +$$ + +where $I(r)$ and $I_s(r)$ are the ground truth color of the pixel from which the ray emits and the rendered color, respectively, $P(r)$ and $P_s(r)$ are the ground truth color of the patch corresponding to the ray and the rendered patch color, respectively. $d_{pixel}$ is the loss metric for pixel color, where we use $L1$ loss as $d_{pixel}$ . $d_{patch}$ is the loss metric for patch color, where we use the Structural Similarity Index Measure (SSIM [39]) as $d_{patch}$ . + +# 3.4. Training Loss + +In addition to the above-mentioned three loss functions, we also use the Eikonal loss [8] used in NeuS [37]. We define the overall loss function as follows: + +$$ +L = L _ {\text {f e a t}} + \alpha L _ {\text {d e p t h}} + L _ {\text {c o l o r}} + \beta L _ {\text {e i k}}, \tag {7} +$$ + +$L_{eki}$ is the Eikonal loss [8], used to regularize the SDF values of sampled points, defined as + +$$ +L _ {e k i.} = \frac {1}{m n} \sum_ {i, k} \left(\| \nabla f \left(x _ {i, k}\right) \| _ {2} - 1\right) ^ {2}. \tag {8} +$$ + +# 4. Experiments + +# 4.1. Dataset + +We evaluate our method on DTU [16] and BlendedMVS [44] dataset. For the DTU [16] dataset, to avoid using the + +
Methods21243437384082106110114118Mean CD ↓
NeuS [37]5.633.586.004.602.574.531.914.185.461.194.163.98
NeuralWarp [5]2.531.880.741.800.8411.502.642.104.371.192.632.93
MonoSDF [47]4.145.921.394.552.192.142.365.624.581.633.023.41
Vis-MVSNet [50]3.394.440.853.361.693.353.352.342.160.741.832.50
MVSDF [51]4.314.711.656.371.774.473.611.871.671.251.693.03
SparseNeuSft [24]3.484.372.924.762.793.732.801.863.101.152.293.02
VolRecon [32]2.723.071.824.322.143.043.002.562.811.493.222.75
GenSft [31]5.867.673.628.575.375.415.486.045.294.694.355.67
ReTR [22]2.673.371.623.681.873.403.672.842.851.562.352.72
UFORecon [28]1.841.520.792.581.001.821.721.200.930.661.261.39
S-VolSDF [40]2.453.081.333.091.223.211.911.511.230.741.21.91
SparseCraft [46]2.882.420.922.971.582.782.511.105.240.650.882.16
NeuSurf [13]7.601.432.933.181.532.861.861.091.410.370.622.26
FatesGS [14]3.981.322.532.853.362.713.761.490.850.471.062.22
Ours2.141.260.721.460.861.391.370.940.770.440.831.11
+ +Table 1. Quantitative results of Chamfer Distance (CD↓) on DTU dataset with 3 small-overlapping images. The methods are divided into three categories, from top to bottom: (1) dense-view reconstruction methods related to ours, (2) generalization-based sparse-view reconstruction methods, and (3) overfitting-based sparse-view reconstruction methods. the best results are in bold, the second best are underlined. + +scenes that have already been used as training data on the pretrained Vis-MVSNet [50] model, we select the same 11 scenes as in S-VolSDF [40]. The image resolution is set to $1600 \times 1200$ . Similar to the S-VolSDF [40] and NeuSurf [13] methods, we select the views 22, 25, and 28 for the more challenging reconstruction of small overlaps. + +For the BlendedMVS [44] dataset, we follow the S-VolSDF [40] to use the same 9 challenging scenes, with 3 small-overlapping views for each scene. The image resolution is set to $768 \times 576$ . + +# 4.2. Implementation Details + +We use the same network architecture and initialization strategy as NeuS [37] and incorporated our volume rendering feature consistency loss, uncertainty-guided depth constraint loss, and color consistency loss. For the weight factors in the loss functions Eq. 7, we set the $\alpha$ for the uncertainty-guided depth prior constraint loss $L_{depth}$ to 0.5 and the $\beta$ for the Eikonal loss $L_{eik}$ to 0.1. Each scene is trained 100K iterations on a RTX3090 GPU. The patch warping term in the color consistency loss requires the surface point normals to calculate homographies, but the initial normals are too noisy [5], therefore, the patch warping loss is applied after 20k training steps. The threshold $\tau$ of the occlusion mask is set to 0. + +# 4.3. Baseline + +We compare our approach with three categories of methods including dense-view methods: NeuS [37], NeuralWarp [5], Vis-MVSNet [50], MVsDF [51], generalization-based + +methods: SparseNeuS [24], VolRecon [32], GenS [31], ReTR [22] and UFORecon [28], overfitting-based methods: S-VolSDF [40], SparseCraft [46], NeuSurf [13] and FatesGS [14]. The reconstruction results for SparseNeuS [24] and GenS [31] are fine-tuned using 3 views for each scene. + +# 4.4. Comparisons + +Reconstruction on DTU. For a comprehensive comparison, we evaluate the baselines and our method on both small-overlapping and large-overlapping views. Following baselines [13, 14, 24], we report the Chamfer Distance (CD) between the reconstruction surfaces and the ground truth point clouds. Since the scenes in our experiments differ from those used in NeuSurf [13] and FatesGS [14], we reran the official implementations of NeuSurf and FatesGS. The CD results with small overlapping views are shown in Table 1. The meshes reconstructed by several methods using 3 views with small overlapping are shown in Fig. 6. For the generalization-based sparse reconstruction methods, we only show the reconstruction results of the latest UFORecon [28], as the reconstruction quality of other methods is lower than that of UFORecon [28]. The experimental results show that our method significantly improves the mesh quality with small overlap views, compared to the state-of-the-art sparse-view reconstruction methods. The results of large overlapping views are presented in the supplementary materials. + +As shown in Figure 6, when input sparse views with small overlap, both MonoSDF [47] and SparseCraft [46] + +![](images/8c7d9991dce4278a6c28be20aaa40925545f12143fc053d9aa8200959e6fccdc.jpg) +GT images + +![](images/e80e85c5aa13106c64819e8da04473dedc21a31942dea52021d9a21c8aae4d6b.jpg) + +![](images/1fbb0598f2f8d14ac6a0ed644a7e8c4b8c2ce8b19d63b2a3689a3c961c628930.jpg) + +![](images/4bd6a765902a982d98f947cd0ab0c2a58c89aa660ba68e77994c60f7e7cd50fe.jpg) + +![](images/1116ebbb9d339ecabb444f30535805bd4b39de437914b0c9e5d1367dcfce12c4.jpg) +MonoSDF + +![](images/e3a528cc2808cecba63d53848c12e893630e392b85028ddc6908d52ff3461797.jpg) + +![](images/a3484fc088d7f31f8c9fe0aa3866bb47d3e6d6d3e7528fb800423fb3095cd858.jpg) + +![](images/8907d9a022d8aaa6107e5f971513366816ed5754232ab0f251c9b9c50bbacd4a.jpg) + +![](images/3a1c1de873c88483e7cebb3b4f36e387a2812f183d5661e1c34f60cfac90b2cb.jpg) +UFORecon + +![](images/540dfc52c2099175311b82afa4f5b4d9683fb39f32d0313078fe3e5de6a98458.jpg) + +![](images/bd37693c17a82b46234fe34f1b828e604e672a3fb66a1e5a2f00f0a696663537.jpg) + +![](images/32f99fe08d624837f5888372013a4326247c67c772c35d3616c4be9847ff45db.jpg) + +![](images/f9b515cb0727f709de61082d891b069bf040e1b1f6d7e427cf732930558239a8.jpg) +S-VolSDF +Figure 6. Visual comparison on DTU dataset with 3 small-overlapping images. + +![](images/5234037b56fba1b421cdb8b9b1fae1d7cda56a6ade0e3544079022b7bae265b0.jpg) + +![](images/848987358e9abd5d35bfdc9b94b9ef5f00d540bca6628b79591c13c32dee55b0.jpg) + +![](images/c6e8e3a59d606d08d455d4f5b0230b9cce792124eccbe0b31d5b877e7c34b09f.jpg) + +![](images/ca6844839871e6b2bdd44366b4dd4fac7c6149c8cb76f5a3063c3e5c21bf613e.jpg) +SparseCraft + +![](images/3e4b9c45912f4a7bd4e3c203ec244f3266ace4c66a3733335df1298ddae5dd00.jpg) + +![](images/6aa118f1b7278b997ae3e977b98efc503b036a22ba3b7a66cb36222f03d3cfba.jpg) + +![](images/8b9823d87660d692fabfc045ba6886fcdc515d61174d50d866df8354c48860c3.jpg) + +![](images/9d02057027f2e0cd53180fdf93cac62c0207c63130c08860a6de56764af03907.jpg) +NeuSurf + +![](images/6d199e8d4f230f4b9bcdd699e4fbc14914d76a25d0d994cf86fcce866656d79b.jpg) + +![](images/ddb57d9f67da2ae741fdc296abd71c0ea90e08c5d96200ba512d3bf2fb77cb02.jpg) + +![](images/d718d6e306504646c416f10ee17338571dc051bf6e5e69c8019513faf8a71796.jpg) + +![](images/9b18ff085e23afa7201f5e9ddfe05bc75c8d20425c0e038d70e11913565b9cdb.jpg) +FatesGS + +![](images/804ec0d600d7f20b65e8c4949a378ffb9eec19698a2cbfeb4fee6a14a29c276f.jpg) + +![](images/4733198d95d183b85f599307f0997b589da03ae02e2c4067e0dce0d43c94563f.jpg) + +![](images/bd45dd827ccc96796c18ae86319e8ec97292e481c27170d7b591e38571f10175.jpg) + +![](images/3ff8fffec39e575eb40b82320151dc22e8a07cf210405c0822065e8c92528ae0.jpg) +Ours + +suffer from reconstruction ambiguity and failures, highlighting that relying solely on simplistic geometric prior constraints is insufficient to obtain complete and accurate meshes. UFORecon [28] shows significant roughness in its reconstruction results. S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit noticeable reconstruction defects. Experimental results demonstrate that our method is effective in alleviating geometric and appearance ambiguities during the optimization process. This significantly enhances the quality of mesh reconstruction, especially in scenarios with small overlapping views and low texture. + +![](images/e139c7de8fd0b69d91c91309f8ce88d8784fdd13540c9ce32b5df4390519c5a2.jpg) +GT Images +MonoSDF +Figure 7. Visual comparison on BlendedMVS dataset. $\times$ indicates reconstruction failure. + +![](images/750bf527000ee7ffc93382c4ca91b44018762741e3c4454082176563084aa809.jpg) +X + +![](images/2f2e5535ecc3df2c42535293861601229deaaa7c6e8914b0a340aab6aae2e3ae.jpg) + +![](images/284d239afac530b1d5468a34271c5f844108913ea5d07e4b7199b8983a8048e6.jpg) +S-VolSDF + +![](images/42d69eecfa0ada455d2303e7fe3238e2b683c9bfa70f4c20414cc4d6df2f659d.jpg) + +![](images/e8a2fd7e623536b8a49b54eea3bfe7f3c66b1e2cf39d446421535349d055d45e.jpg) +NeuSurf + +![](images/b37806df4a5809c1f843a618567856c08790fb2825d6a72235962bedef785f09.jpg) + +![](images/9c7c1bacb02c966cbdb4d82cbe62a27903985040aefd53f75e0f775235f75497.jpg) +FatesGS + +![](images/e6b0709656a8a4838036c97717cbcde6d23bbaa83161e4c36eb552356b313bb9.jpg) +Ours + +Reconstruction on BlendedMVS. Figure 7 presents the visual comparison of reconstructed mesh for overfitting-based methods. With only 3 small-overlapping views provided, all of the generalization-based methods completely fail to reconstruct in the sparse setting of BlendedMVS[44] dataset, even if SparseNeuS [24] is fine-tuned. Therefore, the reconstruction results of these methods are not included + +in Figure 7. Compared to other methods, our approach can generate more complete and detailed meshes. Similarly, MonoSDF [47] fails to reconstruct surfaces either. The meshes generated by S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit significant defects. Both NeurSurf [13] and FatesGS [14] use on-surface feature consistency constraints, but the reconstruction results are still not good enough. In contrast, our method achieves more comprehensive geometry and finer details by employing volume rendering-based feature consistency constraints. This highlights the advantages of our approach in geometric consistency. More visualizations are presented in the supplementary materials. + +![](images/b2e14c2d9288456a7f4f6b8f98b5b1a414bcb81b7b85d4f82a0263d256de7a2b.jpg) +Scan 40 + +![](images/85612e3143694ee3e593587cac774d085724692cdf55ddec13def2e89ae45fbf.jpg) +Scan 110 +Mesh w/ $L_{\text{feat}}$ by ours +Figure 8. Reconstructed meshes and error maps on DTU dataset with different feature consistency losses. + +![](images/0d82f1550c1a04c064fbe70b6f71c81e668383cbf078cb3ba61675481fe153e5.jpg) + +![](images/d2be16916de4bde2d6d84c72f7d19cc87bbb77f713247442bdeba7c52ccb2afa.jpg) +Error map w/ $L_{\text{feat}}$ by ours +at Mesh w/ surf $f_{\text{feat}}$ by MVSDF + +![](images/6471b5b4c7c6ca2584c4b631c3c70db1768c4c07689cc7ed8de0cfa5d45a517d.jpg) + +![](images/27bb6e20c01eae7f46cd2166eaa1dbc0b9ed988c3c36c4be0f96a244aef12996.jpg) + +![](images/2586dba55b8aec8296c88d3413c0f39acc451fa281d08ee53b14c18c10168ffa.jpg) + +![](images/9166b28abdee442f792d892293b2e13a1fa28d43ff64575d719b8449e61a47ee.jpg) +Error map w/ $L_{\text{feat}}^{\text{surf}}$ by MVSDF + +# 4.5. Ablation Study + +We evaluate the components of our method with 3 small-overlapping views by an ablation study on the DTU [16] dataset. To compare the depth loss $L_{depth}^{mono}$ calibrated by rendered depth in MonoSDF [47] with our depth loss $L_{depth}$ , we replace $L_{depth}$ with $L_{depth}^{mono}$ to evaluate it in our method. We also compare the volume rendering-based + +
MethodLcolorLfeatLdepthLmono depthL1featL2featsurffeatCD↓
Baseline3.35
✓1.76
✓✓1.47
✓✓1.62
✓✓✓1.11
✓✓✓1.59
✓✓✓2.36
✓✓✓1.81
✓✓✓2.93
+ +Table 2. Ablation studies on DTU dataset with 3 small-overlapping images. + +![](images/451bb413ec3309041737c2a5b784522cd14538751a6dbdfd7c5ba2c2f15ef45c.jpg) +Step 10k + +![](images/131b2eec1aaa987c6c2fc44b3bf9e8506d1202bfe6add5f806c15e13e8bceda7.jpg) + +![](images/3dc8b7ee7873229fbc42d7b82e09bfb5814477cbf41799d23abe17dc7f476241.jpg) +Step 20k + +![](images/565976665936a09e43f0998137701f1ac6ad3b225dd18998d44903bee55a0d4d.jpg) + +![](images/18d4841a6253c82ef59dfb7b19add450704ec4221adc3f4d857b2be65c3111fb.jpg) +Step 40k + +![](images/238742f5599328f8c295cb67b70794054114a20205eb71cbb7a81fa4c44b8a2d.jpg) + +![](images/bdd4739c5e95bcac5bb70a44287df07815f86a7561776eafecbab25c24ce8632.jpg) +Step 60k + +![](images/b585ad7b38d2901e1288f5bfda5791d6ec524df280ec939f83997a0f3f50aefd.jpg) +Figure 9. The variation of weighted feature similarity during training, brighter colors indicate higher feature similarity. +Figure 10. Visualization of reconstruction and error maps for scene scan24 and scan37 in DTU dataset with different losses. The differences of error maps are highlighted. + +feature consistency loss calculated using L1 distance (denoted as $L_{\text{feat}}^{L1}$ ) and L2 distance (denoted as $L_{\text{feat}}^{L2}$ ) with our method using feature similarity distance. We found that feature similarity distance is better than both L1 and L2 distance, as shown in Table 2. + +In addition, we replace our volume rendering-based feature consistency loss $L_{feat}$ with the on-surface feature consistency loss $L_{feat}^{surf}$ used in MVSDF [51] to compare the effectiveness of two different loss functions. Figure 8 illustrates the performance of these two different loss functions. + +![](images/a93de04f8c54b87ca5aa86f8524b25e96821ab93218da111d0e9957273a690e7.jpg) +GT Images + +![](images/e5e2d2251d0b4eec902efa1ea1f7eef70213e789d50ec49e4c3e8d9414f4d6a9.jpg) +S-VolSDF + +![](images/03dd7a95ec283cd2a379c445e8d07dd8229d276449f1c929f297781bca1804df.jpg) +NeuSurf +Figure 11. Failure case. For specular objects, the ambiguity in the color consistency constraint may lead to a rough surface. + +![](images/7fcc5f9a7a10ef9f527e262beedd3796d6f6220251f49abb4ceaff1f7337da73.jpg) +FatesGS + +![](images/a6f12d9fe4bb0baee019f107fb50213f5860956a32c6570b65995cf268da237f.jpg) +Ours + +trates the reconstruction results and error maps on the DTU dataset when using different feature consistency losses, under the on-surface feature consistency loss $L_{feat}^{surf}$ , the meshes show large artifacts. + +Table 2 shows the average Chamfer Distance over all 11 scenes on DTU dataset using different losses. The experimental results indicate that both feature consistency loss and uncertainty-guided depth constraint improve the surface reconstruction. + +Figure 9 illustrates the variation of the weighted feature similarity map during the training process. Brighter colors indicate higher feature similarity, demonstrating that our volume rendering-based feature consistency loss can provide effective constraints. + +Figure 10 shows the reconstructed meshes and error maps for scene scan24 and scan37 on the DTU [16] dataset when using different losses. It can be observed that the mesh deteriorates without the volume rendering-based feature consistency loss or the uncertainty-guided depth constraint loss, and the reconstruction quality drops when using the depth loss $L_{depth}^{mono}$ in MonoSDF [47]. + +# 5. Conclusions + +We propose a novel method for learning implicit representations from sparse views with small overlaps. Our novelty lies in a novel volume rendering-based feature consistency loss and an uncertainty-guided depth constraint. Extensive experiments on the DTU [16] and BlendedMVS [44] datasets show that our method surpasses existing state-of-the-art sparse-view reconstruction methods in terms of reconstruction quality. + +Limitations. Although our method shows significant improvement over other sparse view reconstruction methods, there are still some limitations. Firstly, for specular objects, the ambiguity in the color consistency constraint may lead to a rough surface, as shown in Figure 11. Secondly, Following previous studies [13, 24, 40], the camera poses of sparse views are obtained from the training dataset. However, in some cases, it may not be possible to obtain accurate camera poses using SfM methods like COLMAP [33] due to the lack of texture in the images or excessive viewing angles. Additionally, the feature consistency constraint method requires a pre-trained network to extract image features. The accuracy of the features determines the performance of the feature consistency constraint. + +# 6. Acknowledgment + +The corresponding authors are Yu-Shen Liu and Haichuan Song. This work was supported by National Key R&D Program of China (2022YFC3800600), and the National Natural Science Foundation of China (62272263), and in part by Tsinghua-Kuaishou Institute of Future Media Data. + +# References + +[1] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, PP: 1-12, 2024. 3 +[2] Hanlin Chen, Chen Li, and Gim Hee Lee. NeuSG: Neural implicit surface reconstruction with 3D gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 3 +[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. MVSplat: Efficient 3D gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 3 +[4] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3 +[5] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 1, 2, 5, 6 +[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3 +[7] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2 +[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 5 +[9] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2 +[10] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 3 + +[11] Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3D gaussian splattering with view consistency for sparse view synthesis. In Advances in Neural Information Processing Systems, 2024. 3 +[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3 +[13] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. NeuSurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 1, 2, 4, 6, 7, 8 +[14] Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2, 3, 4, 6, 7 +[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a Diet: Semantically consistent few-shot view synthesis implementation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3 +[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanas. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 2, 5, 7, 8 +[17] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165, 2020. 2 +[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 3 +[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3 +[20] Shujuan Li, Yu-Shen Liu, and Zhizhong Han. GaussianUDF: Inferring unsigned distance functions through 3D gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 27113-27123, 2025. 3 +[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2 +[22] Yixun Liang, Hao He, and Yingcong Chen. ReTR: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6 +[23] Xinqi Liu, Jituo Li, and Guodong Lu. Reconstructing complex shaped clothing from a single image with feature stable + +unsigned distance fields. IEEE Transactions on Visualization and Computer Graphics, 2024. 1 +[24] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. SparseNeuS: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 6, 7, 8 +[25] Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 3DGSR: Implicit surface reconstruction with 3D gaussian splatting. ACM Transactions on Graphics, 43(6):1-12, 2024. 3 +[26] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3D point clouds via noise to noise mapping. In International Conference on Machine Learning, 2023. 1 +[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[28] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. UFORecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 1, 2, 6, 7 +[29] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3 +[30] Takeshi Noda, Chao Chen, Junsheng Zhou, Weiqi Zhang, Yu-Shen Liu, and Zhizhong Han. Learning bijective surface parameterization for inferring signed distance functions from sparse point clouds with grid deformation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22139-22149, 2025. 1 +[31] Rui Peng, Xiaodong Gu, Luyang Tang, Shihe Shen, Fanqi Yu, and Ronggang Wang. GenS: Generalizable neural surface reconstruction from multi-view images. In Advances in Neural Information Processing Systems, pages 56932-56945, 2023. 1, 2, 6 +[32] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. VolRecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 6 +[33] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, 2016. 2, 4, 5, 8 +[34] Hui Tian, Chenyang Zhu, Yifei Shi, and Kai Xu. SuperUDF: Self-supervised udf estimation for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2023. 1 + +[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. SPARF: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 3 +[36] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. SparseNeRF: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3 +[37] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 1, 2, 3, 5, 6 +[38] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 1 +[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[40] Haoyu Wu, Alexandros Graikos, and Dimitris Samaras. SVolSDF: Sparse multi-view stereo regularization of neural implicit surfaces. International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8 +[41] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. International Conference on Learning Representations, 2022. 1, 2 +[42] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2F2NeUS: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 2 +[43] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3 +[44] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 2, 5, 6, 7, 8 +[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2 +[46] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. SparseCraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6 + +[47] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 35:25018-25032, 2022. 2, 4, 6, 7, 8 +[48] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6): 1-13, 2024. 3 +[49] Yu-Jie Yuan, Yu-Kun Lai, Yi-Hua Huang, Leif Kobbelt, and Lin Gao. Neural radiance fields from sparse rgb-d images for high-quality view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8713-8728, 2022. 3 +[50] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. The British Machine Vision Conference, 2020. 2, 3, 6 +[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 4, 6, 8 +[52] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. CoR-GS: sparse-view 3D gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2024. 3 +[53] Wenyuan Zhang, Ruofan Xing, Yunfan Zeng, Yu-Shen Liu, Kanle Shi, and Zhizhong Han. Fast learning radiance fields by shooting much fewer rays. IEEE Transactions on Image Processing, 2023. 1 +[54] Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. Neural signed distance function inference through splatting 3D gaussians pulled on zero-level set. In Advances in Neural Information Processing Systems, 2024. 3 +[55] Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou, Baorui Ma, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. NeRFPrior: Learning neural radiance field as a prior for indoor scene reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11317–11327, 2025. 2 +[56] Wenyuan Zhang, Yixiao Yang, Han Huang, Liang Han, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. MonoInstance: Enhancing monocular priors via multi-view instance alignment for neural rendering and reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21642-21653, 2025. 2 +[57] Weiqi Zhang, Junsheng Zhou, Haotian Geng, Wenyuan Zhang, and Yu-Shen Liu. GAP: Gaussianize any point clouds with text guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3 +[58] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, Yi Fang, and Zhizhong Han. CAP-UDF: Learning unsigned distance functions progressively from raw point clouds with consistency-aware field optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[59] Junsheng Zhou, Baorui Ma, and Liu Yu-Shen. Fast learning of signed distance functions from noisy point clouds via noise to noise mapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 + +[60] Tiansong Zhou, Jing Huang, Tao Yu, Ruizhi Shao, and Kun Li. HDhuman: High-quality human novel-view rendering from sparse views. IEEE Transactions on Visualization and Computer Graphics, 2023. 1 +[61] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. FSGS: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 3 \ No newline at end of file diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/images.zip b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1d212a2814a75140c86a5ec3297149783b874e2e --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c204c28874db71318512c6a2efc7e212aec4474aad418f8c4aaf23f8dc8ef5d4 +size 730057 diff --git a/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/layout.json b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..679a57870c3ae781819a53d66682e6872540f0dd --- /dev/null +++ b/2025/SparseRecon_ Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies/layout.json @@ -0,0 +1,12054 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "text", + "content": "SparseRecon: Neural Implicit Surface Reconstruction from Sparse Views with Feature and Depth Consistencies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "spans": [ + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "Liang Han" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": ", Xu Zhang" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": ", Haichuan Song" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": ", Kanle Shi" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": ", Yu-Shen Liu" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": ", Zhizhong Han" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "School of Software, Tsinghua University, Beijing, China \n" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "Computer Science and Technology, East China Normal University, Shanghai, China \n" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "China Telecom " + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "Kuaishou Technology, Beijing, China \n" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 94, + 160, + 515, + 232 + ], + "type": "text", + "content": "Department of Computer Science, Wayne State University, Detroit, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 234, + 511, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 234, + 511, + 258 + ], + "spans": [ + { + "bbox": [ + 97, + 234, + 511, + 258 + ], + "type": "text", + "content": "hanl23@mails.tsinghua.edu.cn, zhangxu@chinatelecom.cn, hcsong@cs.ecnu.edu.cn shikanle@kuaishou.com, liuyushen@tsinghua.edu.cn, h312h@wayne.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 313, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 296, + 601 + ], + "type": "text", + "content": "Surface reconstruction from sparse views aims to reconstruct a 3D shape or scene from few RGB images. The latest methods are either generalization-based or overfitting-based. However, the generalization-based methods do not generalize well on views that were unseen during training, while the reconstruction quality of overfitting-based methods is still limited by the limited geometry clues. To address this issue, we propose SparseRecon, a novel neural implicit reconstruction method for sparse views with volume rendering-based feature consistency and uncertainty-guided depth constraint. Firstly, we introduce a feature consistency loss across views to constrain the neural implicit field. This design alleviates the ambiguity caused by insufficient consistency information of views and ensures completeness and smoothness in the reconstruction results. Secondly, we employ an uncertainty-guided depth constraint to back up the feature consistency loss in areas with occlusion and insignificant features, which recovers geometry details for better reconstruction quality. Experimental results demonstrate that our method outperforms the state-of-the-art methods, which can produce high-quality geometry with sparse-view input, especially in the scenarios with small overlapping views. Project page: https://hanl2010.github.io/SparseRecon/." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "content": "As one of the important tasks in computer vision, 3D reconstruction has attracted lots of research attentions in recent years. With the advancement of deep learning, 3D reconstruction based on neural implicit representations learned" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 287, + 554, + 411 + ], + "blocks": [ + { + "bbox": [ + 316, + 287, + 554, + 411 + ], + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 411 + ], + "type": "image", + "image_path": "9d2d98f09d624cff078923207d92e2931e107eb3db4bff04f48aa31c9e6f7363.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 424, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 478 + ], + "type": "text", + "content": "Figure 1. Given only 3 input images with large view angle change, our method can reconstruct a smoother surface compared to the state-of-the-art methods, such as UFOREcon [28], S-VolSDF [40] and NeuSurf [13]. The details of each surface are shown in the colored boxes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 602 + ], + "type": "text", + "content": "from point clouds [26, 30, 34, 58, 59] or images [23, 37, 45, 60] becomes a popular research topic. Although recent multi-view reconstruction methods[5, 37, 38, 41, 45, 53] have made great progress in terms of the reconstruction quality and reconstruction speed, they require a large number of dense views as supervision. When limited number of views are available, current reconstruction methods usually struggle to reconstruct high-quality surfaces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Existing methods for sparse view reconstruction can be mainly classified into two categories: generalization-based methods and overfitting-based methods. The generalization-based methods [22, 24, 28, 31, 32] emphasize the generalization of sparse-view reconstruction, but they are mainly effective in scenarios with large view overlaps. In cases with views that were unseen during training, the quality of the reconstructed surface degenerates significantly, as shown in Figure 1. Meanwhile, it takes a long" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "type": "text", + "content": "*Corresponding author: Yu-Shen Liu and Haichuan Song." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28514" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "time to pre-train these methods on large-scale data, but the generalization is still not good enough. Instead, overfitting-based methods [13, 14, 40, 46, 47] typically fit the 3D geometry directly from the sparse views by leveraging geometry clues. They show promising capability of reconstructing higher-quality geometric surfaces with small-overlapping views. However, the reconstruction quality of the existing methods is still unsatisfactory." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 241 + ], + "type": "text", + "content": "In this paper, we introduce a multi-view feature consistency loss based on volume rendering and an uncertainty-guided depth constraint to learn neural signed distance functions. This approach allows us to achieve high-quality mesh reconstruction on more challenging sparse views with small overlaps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 445 + ], + "type": "text", + "content": "For the feature consistency loss, we first employ the pre-trained Vis-MVSNet [50] to obtain depth features from the input images. Then, within a neural implicit rendering framework, the sampled 3D points along the rays emitted from the reference image are projected to the source image and the reference image. This allows us to acquire source features and reference features of each 3D point and measure the similarity between these two kinds of features. Finally, the feature similarity for each 3D point along the rays is accumulated through volume rendering, thus yielding the feature similarity associated with the rays. During optimization, we pursue higher feature similarity along the rays. Since the depth information is implicitly encoded with image features, feature consistency constraint can significantly alleviate the ambiguity issues arising from insufficient consistency of sparse views and low-texture during reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 651 + ], + "type": "text", + "content": "For the uncertainty-guided depth prior constraint, we follow MonoSDF [47], utilizing a pre-trained network to acquire depth priors for each image, and then use it to constrain the regions with uncertain depth. However, monocular depth priors do not have consistent scales to the ground truth depth, which are hard to get calibrated to ground truth either due to the distortion. To effectively leverage the depth priors and provide proper supervision for occluded or under-constrained regions, we propose an uncertainty-guided depth prior constraint. First, we calibrate the depth priors using sparse point clouds obtained from COLMAP [33]. Then, during training, we compute the depth confidence from the rendered depth and impose the depth prior constraint only in regions with low confidence. This constraint helps infer more accurate geometry in occluded or under-constrained regions, minimizing the negative impact of depth prior errors on well-constrained regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 652, + 294, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 294, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 294, + 687 + ], + "type": "text", + "content": "We evaluate our methods on several widely used benchmarks and report the state-of-the-art results. In summary, our main contributions are as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose a novel feature consistency loss based on volume rendering. It can effectively constrain the neural" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 107 + ], + "type": "text", + "content": "radiance field by leveraging feature consistency among multiple views, improving the performance in sparse-view reconstruction tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 108, + 553, + 191 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 144 + ], + "type": "text", + "content": "- By incorporating depth confidence, we utilize the calibrated depth prior more effectively to enhance geometric constraints, further improving the reconstruction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "content": "- Extensive experiments on the well-known datasets, such as DTU [16] and BlendedMVS[44], demonstrate that our method outperforms existing sparse-view reconstruction methods and achieves the state-of-the-art results." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 205, + 400, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 400, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 226, + 482, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 226, + 482, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 226, + 482, + 239 + ], + "type": "text", + "content": "2.1. Neural Implicit Reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 244, + 553, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 353 + ], + "type": "text", + "content": "Neural implicit reconstruction methods [5, 7, 21, 37, 41, 47, 55, 56] have been rapidly developed based on neural volume rendering [27]. These methods introduce the Signed Distance Function (SDF) as the implicit representation of 3D surfaces in volume rendering in multi-view 3D reconstruction. While these methods have made significant improvements in both reconstruction quality and speed, it is important to note that they heavily rely on multiple views with large overlaps during the optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 521 + ], + "type": "text", + "content": "Generalization-based surface reconstruction with sparse views. In order to directly generalize the reconstruction results on sparse views, methods [22, 24, 28, 31, 32, 42] adopt the strategy of aggregating features from multiple view images to construct a feature volume, which is then used to predict the SDF for reconstructing the surface. VolRecon [32] uses transformers [17] to aggregate multi-view features, C2F2NeUS [42] employs cascade architecture to construct a volume pyramid, while ReTR [22] and UFORecon [28] aggregates multi-level features. These methods require pretraining on large-scale datasets, which typically takes several days. However, when there is a significant domain gap between the testing and training data, they all fail to reconstruct shapes effectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "Overfitting-based surface reconstruction with sparse views. In contrast, overfitting-based methods directly fit the 3D geometry from the sparse images by geometric prior constraints. MonoSDF [47] employs depth and normal priors to achieve sparse reconstruction with small-overlapping views. However, such priors come with errors, and it does not fully leverage inter-view consistency, resulting in lower reconstruction quality. S-VolSDF [40] employs probability volumes obtained from MVS [9] models to guide the rendering weight estimated by VolSDF [45]. This improves the reconstruction results in sparse views with small overlap. However, the uncertainties in volumes make negative impact on the reconstructed surface, leading to surface roughness or significant defects. NeuSurf [13] leverages sparse point clouds and employs CAP-UDF [58] to construct an implicit geometric prior to improve the reconstruction qual" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28515" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 531, + 266 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 531, + 266 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 531, + 266 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 531, + 266 + ], + "type": "image", + "image_path": "bcd4dac80121e96e064f3d1eef9554e5730b2d82f3481f549bbb0abf8d6f5194.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. SparseRecon consists of two main parts. (a) Volume rendering-based feature consistency constraint. We extract features from the reference image and source images. For a ray emitted from the reference image, we project each sampled point on the ray onto the source images to obtain the corresponding features. Then, the volume rendering-based feature consistency loss is calculated using the corresponding features on the reference image. (b) Uncertainty-guided depth prior constraint. We use another pre-trained network to obtain the depth prior of the reference image and calibrate it with the sparse point cloud obtained by COLMAP. Then, we calculate the confidence of the rendered depth, so that the calibrated depth prior only constrains areas with low confidence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 360, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 296, + 445 + ], + "type": "text", + "content": "ity from sparse views. However, when the sparse point cloud fails to cover the majority of the object surface, implicit geometric prior information cannot be useful. In contrast, our method employs more robust feature priors, calculates feature consistency based on volume rendering, and simultaneously utilizes depth priors to optimize the occluded regions, ultimately resulting in high-quality surface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 167, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 167, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 167, + 466 + ], + "type": "text", + "content": "2.2. Gaussian Splitting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 471, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 662 + ], + "type": "text", + "content": "Gaussian Splatting [18] has achieved unprecedented optimization speed and rendering quality in the task of novel view synthesis [11, 19, 57, 61]. However, since the Gaussians are unorganized, the discrete and unstructured points make it difficult to extract 3D surfaces through postprocessing. To address this issue, some methods introduce to use regularization terms [10], convert 3D Gaussians to 2D surfels [4, 12], acquire opacity fields through rays [48], improve the depth rendering algorithm [1] of 3DGS, or jointly optimize 3DGS with neural radiance fields [2, 20, 25, 54]. However, these methods require dense views to work well. Recently, FatesGS [14] achieves fast sparse-view reconstruction by leveraging depth priors and on-surface feature consistency constraints. Nevertheless, the reconstruction results still exhibit roughness or noticeable defects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 183, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 183, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 183, + 685 + ], + "type": "text", + "content": "2.3. Sparse View Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In addition, the novel view synthesis from sparse views is another category of work closely related to sparse view" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 360, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 515 + ], + "type": "text", + "content": "reconstruction. Depending on the technical framework, these works can be categorized into NeRF-based methods [6, 15, 29, 35, 36, 43, 49] and Gaussian Splitting-based methods [3, 11, 19, 52, 61]. This line of research also employs a limited number of views as input. However, they solely focus on the rendering quality of novel views rather than surface reconstruction, which are not designed specifically for the accurate geometric surface reconstruction. Due to the discernible bias (i.e. inherent geometric errors) [37] caused by the conventional volume rendering or inconsistencies in depth that appear in Gaussian rendering, current sparse view synthesis methods still fail to correctly reconstruct high-fidelity geometric surfaces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 525, + 371, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 371, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 371, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 654 + ], + "type": "text", + "content": "The overview of our method is depicted in Figure 2. We introduce a novel feature consistency loss and an uncertainty-guided depth constraint based on the NeuS [37] framework. In this section, we first explain how to compute feature consistency for sampled points along rays. Then we explain how to enhance geometric constraints using depth priors and depth uncertainty. Thirdly, we introduce the color consistency loss. Finally, we present the overall loss function for optimization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 660, + 553, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 553, + 673 + ], + "type": "text", + "content": "3.1. Volume Rendering-based Feature Consistency" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "First, we use a pre-trained MVS network [50] to extract the features from both the reference image and the source image. Given a ray emitted from the reference image, let " + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "inline_equation", + "content": "p_r(0)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28516" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "denote the point where a ray intersects the reference image. And for each point " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " along the ray, we denote its projection on the source image as " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "p_{s}(i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ". Then, we bilinearly interpolate " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "F_{r}(0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "F_{s}(i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " at points " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "p_{r}(0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "p_{s}(i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " on image features, respectively. Formally, we define the feature consistency loss function as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 154, + 296, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 154, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 154, + 296, + 196 + ], + "type": "interline_equation", + "content": "L _ {\\text {f e a t}} = M ^ {\\text {o c c}} \\left(1 - \\frac {1}{N} \\sum_ {i = 1} ^ {N} w _ {i} f _ {\\text {c o s}} \\left(F _ {r} \\left(p _ {r} (0)\\right), F _ {s} \\left(p _ {s} (i)\\right)\\right), \\right. \\tag {1}", + "image_path": "4ca9f0501dde1ceffee8eeaddfb8f3e440832e8d7dc9c46685af26149b86bb3d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "f_{cos}" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " is the cosine similarity, and " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " corresponds to the weight for each point along the ray. " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "p_{s}(i) = K(Rx_{i} + t)" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " is the projection of " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " in source view, and " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "[K;R;t]" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " is the camera parameters of source view. " + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "inline_equation", + "content": "M^{occ}" + }, + { + "bbox": [ + 55, + 197, + 296, + 255 + ], + "type": "text", + "content": " is the occlusion mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 256, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 459 + ], + "type": "text", + "content": "Although MVsDF [51], NeuSurf [13] and FatesGS [14] also employ feature consistency constraints, they just leverage the intersection point between a camera ray and the object's surface. Then, this intersection point gets projected onto adjacent views to obtain the corresponding image features for the purpose of comparing features at this point across multiple views. In sparse view scenarios, the estimated positions of surface points can easily deviate significantly, making the on-surface feature consistency loss not converge. NeuSurf [13] and FatesGS [14] utilize sparse point clouds generated by COLMAP [33] as priors, enabling it to obtain partially accurate positions of surface points, thereby allowing the on-surface feature consistency loss to be more effectively leveraged. However, in regions of lacking surface points, the on-surface feature consistency loss cannot ensure the attainment of high-quality geometric surfaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 593 + ], + "type": "text", + "content": "Figure 3 illustrates the difference between on-surface feature consistency and volume rendering-based feature consistency. Due to the uncertainty of gradient direction, the constraint solely relying on surface point features is challenging to be optimized. In contrast, our method does not require the prior estimation of surface points, it calculates feature consistency on all sampling points along the ray, and provides more reasonable and comprehensive supervision to the implicit field, thereby addressing the convergence issue that may arise in sparse reconstruction for MVsDF [51] and NeuSurf [13]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 257, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 257, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 257, + 613 + ], + "type": "text", + "content": "3.2. Uncertainty-Guided Depth Constraint" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "Although multi-view features offer more robust constraint than image colors, they are ineffective for occluded regions. Due to the limited number of views, some regions may only be visible from a single viewpoint. To enhance geometric constraints, we employ depth priors to supervise the radiance field. However, monocular depth priors are not perfect and accurate. Although MonoSDF [47] has already taken the inaccuracy of depth priors into account, i.e., it aligns" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 383, + 137 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 383, + 137 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 383, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 383, + 137 + ], + "type": "image", + "image_path": "65ace355b98a0bfb4bd893044bbddcc65d9f294e9ecc0f00640d3e80b9a2a34b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 137, + 372, + 153 + ], + "lines": [ + { + "bbox": [ + 322, + 137, + 372, + 153 + ], + "spans": [ + { + "bbox": [ + 322, + 137, + 372, + 153 + ], + "type": "text", + "content": "(a) Surface Feature Consistency" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 384, + 70, + 484, + 137 + ], + "blocks": [ + { + "bbox": [ + 384, + 70, + 484, + 137 + ], + "lines": [ + { + "bbox": [ + 384, + 70, + 484, + 137 + ], + "spans": [ + { + "bbox": [ + 384, + 70, + 484, + 137 + ], + "type": "image", + "image_path": "f37d7a6ca0870534856c8febb92469a26e0942dde5a49d2c566314facffb4605.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 137, + 468, + 153 + ], + "lines": [ + { + "bbox": [ + 393, + 137, + 468, + 153 + ], + "spans": [ + { + "bbox": [ + 393, + 137, + 468, + 153 + ], + "type": "text", + "content": "(b) Volume Rendering-based Feature Consistency" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 489, + 88, + 536, + 95 + ], + "lines": [ + { + "bbox": [ + 489, + 88, + 536, + 95 + ], + "spans": [ + { + "bbox": [ + 489, + 88, + 536, + 95 + ], + "type": "text", + "content": "Sample point" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 488, + 95, + 548, + 102 + ], + "lines": [ + { + "bbox": [ + 488, + 95, + 548, + 102 + ], + "spans": [ + { + "bbox": [ + 488, + 95, + 548, + 102 + ], + "type": "text", + "content": "Gradient direction" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 487, + 103, + 515, + 109 + ], + "lines": [ + { + "bbox": [ + 487, + 103, + 515, + 109 + ], + "spans": [ + { + "bbox": [ + 487, + 103, + 515, + 109 + ], + "type": "text", + "content": "Ray" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 487, + 110, + 548, + 117 + ], + "lines": [ + { + "bbox": [ + 487, + 110, + 548, + 117 + ], + "spans": [ + { + "bbox": [ + 487, + 110, + 548, + 117 + ], + "type": "text", + "content": "Estimated surface" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 487, + 117, + 535, + 123 + ], + "lines": [ + { + "bbox": [ + 487, + 117, + 535, + 123 + ], + "spans": [ + { + "bbox": [ + 487, + 117, + 535, + 123 + ], + "type": "text", + "content": "Real surface" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 202, + 551, + 323 + ], + "blocks": [ + { + "bbox": [ + 313, + 165, + 555, + 188 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 188 + ], + "type": "text", + "content": "Figure 3. The illustration of (a) on-surface feature consistency and (b) feature consistency with volume rendering." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 202, + 551, + 323 + ], + "lines": [ + { + "bbox": [ + 315, + 202, + 551, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 202, + 551, + 323 + ], + "type": "image", + "image_path": "66beb6e5d38527afcec9dbbf16ca9c447fabacbfbe81f20dc153aff7fdaeef00.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 334, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 390 + ], + "type": "text", + "content": "Figure 4. The illustration of rendered depth produced by different depth prior utilization methods, along with the corresponding error maps. (a) Calibrate the depth prior using the rendered depth during training. (b) Calibrate the depth prior using the COLMAP sparse point cloud." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 412, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 555 + ], + "type": "text", + "content": "depth priors using rendered depth during training. However, the rendered depth during training is inaccurate, resulting in significant errors in the calibrated depth priors. This ultimately leads to the accumulation of errors during training, which results in inaccurate reconstructions. Figure 4 (a) shows the calibrated depth prior and rendered depth obtained by MonoSDF [47], as well as their error maps compared to the ground truth depth. It can be seen that both the calibrated depth prior and the rendered depth are with large errors. Therefore, MonoSDF [47] uses a weight annealing strategy to anneal the weight of depth loss to 0 during the first 200 training epochs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 557, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 700 + ], + "type": "text", + "content": "Another trivial approach is to calibrate the depth priors using the sparse point cloud obtained from COLMAP [33]. Since the sparse points are generally located on the geometric surface of the object, their depth is relatively accurate. Therefore, calibrating the depth priors using the sparse point cloud can lead to more accurate depth priors. Figure 4 (b) shows the depth priors calibrated with the sparse point cloud, and the rendered depth, as well as their error maps compared to the ground truth depth. It indicates that the depth priors calibrated to the point cloud from the COLMAP [33] are more accurate. Therefore, we can use them as an constraint leads to more precise rendered depth." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "However, due to the distortions in monocular depth pri" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28517" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 296, + 155 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 296, + 155 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 296, + 155 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 296, + 155 + ], + "type": "image", + "image_path": "9fad56129eeb94608cdb0ba400507bb5d1df0d83aff02e24730ee2c329fa0804.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 163, + 296, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 185 + ], + "type": "text", + "content": "Figure 5. Left: the method of obtaining the confidence of rendered depth. Right: the rendered depth and the depth confidence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 206, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 295, + 374 + ], + "type": "text", + "content": "ors, it is impossible to perfectly align them with the ground truth depth. Even after calibration, the depth priors still exhibit noticeable errors when compared to the ground truth depth. In sparse view scenarios, occlusions and insufficient constraints are more common, leading to significant discrepancies between the geometry of occluded regions and the real surface. Therefore, to achieve more accurate geometry in these under-constrained regions while avoiding the negative impact of depth prior errors on well-constrained regions, we propose an uncertainty-guided depth prior constraint method to more effectively utilize the depth priors. Specifically, we apply depth prior constraints in regions with depth uncertainty, while refraining from using them in regions with high depth confidence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": "To obtain the confidence of the rendered depth, we employ a method to evaluate the multi-view depth projection consistency. As shown in Figure 5, for a specific pixel " + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "inline_equation", + "content": "r_{uv}" + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": " in the reference image with depth " + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "inline_equation", + "content": "d_r" + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": ", it can be mapped to a neighboring image through the homography matrix " + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "inline_equation", + "content": "H_{rs}" + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": ", leading to a pixel " + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "inline_equation", + "content": "s_{uv}" + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 460, + 294, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 460, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 143, + 460, + 294, + 472 + ], + "type": "interline_equation", + "content": "s _ {u v} = H _ {r s} r _ {u v}, \\tag {2}", + "image_path": "73669d16a980a23258b2811e8bcb6357b7ba87bce2f1369a97172ced7d306291.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 478, + 295, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 478, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 140, + 478, + 295, + 491 + ], + "type": "interline_equation", + "content": "H _ {r s} = M _ {s} M _ {r} ^ {- 1}, \\tag {3}", + "image_path": "66ee8747e88366d8d1119527fa64af913ec4cf0d55334195810a09cb84015df4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "M_s" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " are the projection matrices corresponding to the reference and source views, respectively. Similarly, we can map the pixel " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "s_{uv}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " in the source view to the reference view using the projection matrix " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "H_{sr}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": " and its corresponding depth " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "d_s" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\hat{r}_{uv}" + }, + { + "bbox": [ + 55, + 498, + 296, + 594 + ], + "type": "text", + "content": ". The forward and backward projection distance error reflect the accuracy of depth predictions, so we take it as the depth confidence, which is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 604, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 604, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 88, + 604, + 295, + 637 + ], + "type": "interline_equation", + "content": "C _ {d} = \\left\\{ \\begin{array}{l l} \\frac {1}{e ^ {\\| r _ {u v} - \\hat {r} _ {u v} \\|}}, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| \\leqslant 1 \\\\ 0, & \\text {i f} \\| r _ {u v} - \\hat {r} _ {u v} \\| > 1 \\end{array} \\right. \\tag {4}", + "image_path": "3c8a670e4e31c3f6c8f99d91e059d922bd951057919f651116ddf72113fd84f9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 666 + ], + "type": "text", + "content": "The right side of Figure 5 shows the rendered depth and the corresponding depth confidence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Correspondingly, the depth uncertainty is defined as " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "U_{d} = 1 - C_{d}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ". Meanwhile, we can set a threshold " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " for depth confidence " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " to obtain the occlusion mask " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "M^{occ} = \\{C_d > \\tau\\}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "For depth calibration, we leverage COLMAP [33] to obtain a sparse point cloud " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\{X : x_1, x_2 \\ldots x_i \\in R\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and visibility flags indicating which keypoints are visible from view " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". Given the camera parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " of view " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", we estimate the depth " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\bar{D}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " of keypoints by computing the distance from the visible keypoints " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " to the camera center " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". Then, we calibrate the monocular depth prior " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\hat{D}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\bar{D}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", it can be defined as " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\bar{D} \\approx a\\hat{D} + b" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " is the scale factor and " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " is the shift factor, obtained through the least squares method. Formally, the depth constraint loss is defined as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 349, + 197, + 555, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 197, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 349, + 197, + 555, + 225 + ], + "type": "interline_equation", + "content": "L _ {d e p t h} = \\sum_ {r \\in R} U _ {d} \\left\\| (a \\hat {D} + b) - D _ {p r e d} \\right\\| ^ {2}. \\tag {5}", + "image_path": "0269df46d23f79b7cb0126b7360aaab0d3d6cb9a5eb63dd4e60ab18edfd34bb1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 231, + 476, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 476, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 476, + 243 + ], + "type": "text", + "content": "3.3. Color Consistency Constraint" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 248, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 344 + ], + "type": "text", + "content": "Although feature consistency constraint can ensure that the reconstruction does not suffer from severe artifacts, it does not provide sufficient supervision to reconstruct fine geometric details. Conversely, in cases with rich textures, image color constraint can refine the geometric details. Therefore, following the NeuralWarp [5], pixel warping loss and patch warping loss are used in our method as multi-view color consistency loss functions," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 350, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 356, + 350, + 553, + 403 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {c o l o r} = \\sum_ {r \\in R} M ^ {o c c} d _ {p i x e l} (I (r), I _ {s} (r)) \\\\ + \\sum_ {r \\in R} M ^ {\\text {o c c}} d _ {\\text {p a t c h}} (P (r), P _ {s} (r)), \\tag {6} \\\\ \\end{array}", + "image_path": "4c87355e040ec445b46fa3285322f107349dc2d28483db045675d4755905ab18.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "I(r)" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "I_s(r)" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " are the ground truth color of the pixel from which the ray emits and the rendered color, respectively, " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "P(r)" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "P_s(r)" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " are the ground truth color of the patch corresponding to the ray and the rendered patch color, respectively. " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_{pixel}" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " is the loss metric for pixel color, where we use " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " loss as " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_{pixel}" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_{patch}" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": " is the loss metric for patch color, where we use the Structural Similarity Index Measure (SSIM [39]) as " + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "inline_equation", + "content": "d_{patch}" + }, + { + "bbox": [ + 313, + 408, + 555, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 402, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 402, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 402, + 523 + ], + "type": "text", + "content": "3.4. Training Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "type": "text", + "content": "In addition to the above-mentioned three loss functions, we also use the Eikonal loss [8] used in NeuS [37]. We define the overall loss function as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 350, + 570, + 553, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 570, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 350, + 570, + 553, + 583 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {f e a t}} + \\alpha L _ {\\text {d e p t h}} + L _ {\\text {c o l o r}} + \\beta L _ {\\text {e i k}}, \\tag {7}", + "image_path": "e343f8ddb2938e70bbdb42cc3e605c7f6a444ed1678721fad5155724924f0bec.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "type": "inline_equation", + "content": "L_{eki}" + }, + { + "bbox": [ + 313, + 588, + 555, + 612 + ], + "type": "text", + "content": " is the Eikonal loss [8], used to regularize the SDF values of sampled points, defined as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 356, + 617, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 617, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 356, + 617, + 553, + 646 + ], + "type": "interline_equation", + "content": "L _ {e k i.} = \\frac {1}{m n} \\sum_ {i, k} \\left(\\| \\nabla f \\left(x _ {i, k}\\right) \\| _ {2} - 1\\right) ^ {2}. \\tag {8}", + "image_path": "5ed7b0e941e464b8da7e0337c3b4bfdf23af49e6bb75fff0ba5a4561a93f932d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 672, + 373, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 373, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 373, + 683 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "We evaluate our method on DTU [16] and BlendedMVS [44] dataset. For the DTU [16] dataset, to avoid using the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28518" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 538, + 275 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 538, + 275 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 538, + 275 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 538, + 275 + ], + "type": "table", + "html": "
Methods21243437384082106110114118Mean CD ↓
NeuS [37]5.633.586.004.602.574.531.914.185.461.194.163.98
NeuralWarp [5]2.531.880.741.800.8411.502.642.104.371.192.632.93
MonoSDF [47]4.145.921.394.552.192.142.365.624.581.633.023.41
Vis-MVSNet [50]3.394.440.853.361.693.353.352.342.160.741.832.50
MVSDF [51]4.314.711.656.371.774.473.611.871.671.251.693.03
SparseNeuSft [24]3.484.372.924.762.793.732.801.863.101.152.293.02
VolRecon [32]2.723.071.824.322.143.043.002.562.811.493.222.75
GenSft [31]5.867.673.628.575.375.415.486.045.294.694.355.67
ReTR [22]2.673.371.623.681.873.403.672.842.851.562.352.72
UFORecon [28]1.841.520.792.581.001.821.721.200.930.661.261.39
S-VolSDF [40]2.453.081.333.091.223.211.911.511.230.741.21.91
SparseCraft [46]2.882.420.922.971.582.782.511.105.240.650.882.16
NeuSurf [13]7.601.432.933.181.532.861.861.091.410.370.622.26
FatesGS [14]3.981.322.532.853.362.713.761.490.850.471.062.22
Ours2.141.260.721.460.861.391.370.940.770.440.831.11
", + "image_path": "3f7d3fa13c00ccba9828aec1d78ba30d079e344df17c1eb48d554ad011882347.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 555, + 328 + ], + "type": "text", + "content": "Table 1. Quantitative results of Chamfer Distance (CD↓) on DTU dataset with 3 small-overlapping images. The methods are divided into three categories, from top to bottom: (1) dense-view reconstruction methods related to ours, (2) generalization-based sparse-view reconstruction methods, and (3) overfitting-based sparse-view reconstruction methods. the best results are in bold, the second best are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 349, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 420 + ], + "type": "text", + "content": "scenes that have already been used as training data on the pretrained Vis-MVSNet [50] model, we select the same 11 scenes as in S-VolSDF [40]. The image resolution is set to " + }, + { + "bbox": [ + 55, + 349, + 295, + 420 + ], + "type": "inline_equation", + "content": "1600 \\times 1200" + }, + { + "bbox": [ + 55, + 349, + 295, + 420 + ], + "type": "text", + "content": ". Similar to the S-VolSDF [40] and NeuSurf [13] methods, we select the views 22, 25, and 28 for the more challenging reconstruction of small overlaps." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": "For the BlendedMVS [44] dataset, we follow the S-VolSDF [40] to use the same 9 challenging scenes, with 3 small-overlapping views for each scene. The image resolution is set to " + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "inline_equation", + "content": "768 \\times 576" + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 478, + 188, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 188, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 188, + 491 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": "We use the same network architecture and initialization strategy as NeuS [37] and incorporated our volume rendering feature consistency loss, uncertainty-guided depth constraint loss, and color consistency loss. For the weight factors in the loss functions Eq. 7, we set the " + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": " for the uncertainty-guided depth prior constraint loss " + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "inline_equation", + "content": "L_{depth}" + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": " to 0.5 and the " + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": " for the Eikonal loss " + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "inline_equation", + "content": "L_{eik}" + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": " to 0.1. Each scene is trained 100K iterations on a RTX3090 GPU. The patch warping term in the color consistency loss requires the surface point normals to calculate homographies, but the initial normals are too noisy [5], therefore, the patch warping loss is applied after 20k training steps. The threshold " + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 495, + 295, + 651 + ], + "type": "text", + "content": " of the occlusion mask is set to 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 117, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 117, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 117, + 671 + ], + "type": "text", + "content": "4.3. Baseline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We compare our approach with three categories of methods including dense-view methods: NeuS [37], NeuralWarp [5], Vis-MVSNet [50], MVsDF [51], generalization-based" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 421 + ], + "type": "text", + "content": "methods: SparseNeuS [24], VolRecon [32], GenS [31], ReTR [22] and UFORecon [28], overfitting-based methods: S-VolSDF [40], SparseCraft [46], NeuSurf [13] and FatesGS [14]. The reconstruction results for SparseNeuS [24] and GenS [31] are fine-tuned using 3 views for each scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 431, + 400, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 400, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 400, + 443 + ], + "type": "text", + "content": "4.4. Comparisons" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 688 + ], + "type": "text", + "content": "Reconstruction on DTU. For a comprehensive comparison, we evaluate the baselines and our method on both small-overlapping and large-overlapping views. Following baselines [13, 14, 24], we report the Chamfer Distance (CD) between the reconstruction surfaces and the ground truth point clouds. Since the scenes in our experiments differ from those used in NeuSurf [13] and FatesGS [14], we reran the official implementations of NeuSurf and FatesGS. The CD results with small overlapping views are shown in Table 1. The meshes reconstructed by several methods using 3 views with small overlapping are shown in Fig. 6. For the generalization-based sparse reconstruction methods, we only show the reconstruction results of the latest UFORecon [28], as the reconstruction quality of other methods is lower than that of UFORecon [28]. The experimental results show that our method significantly improves the mesh quality with small overlap views, compared to the state-of-the-art sparse-view reconstruction methods. The results of large overlapping views are presented in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "As shown in Figure 6, when input sparse views with small overlap, both MonoSDF [47] and SparseCraft [46]" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28519" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 145, + 246 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 145, + 246 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 145, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 145, + 246 + ], + "type": "image", + "image_path": "8c7d9991dce4278a6c28be20aaa40925545f12143fc053d9aa8200959e6fccdc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 248, + 113, + 256 + ], + "lines": [ + { + "bbox": [ + 81, + 248, + 113, + 256 + ], + "spans": [ + { + "bbox": [ + 81, + 248, + 113, + 256 + ], + "type": "text", + "content": "GT images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 69, + 204, + 115 + ], + "blocks": [ + { + "bbox": [ + 145, + 69, + 204, + 115 + ], + "lines": [ + { + "bbox": [ + 145, + 69, + 204, + 115 + ], + "spans": [ + { + "bbox": [ + 145, + 69, + 204, + 115 + ], + "type": "image", + "image_path": "e80e85c5aa13106c64819e8da04473dedc21a31942dea52021d9a21c8aae4d6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 148, + 116, + 201, + 157 + ], + "blocks": [ + { + "bbox": [ + 148, + 116, + 201, + 157 + ], + "lines": [ + { + "bbox": [ + 148, + 116, + 201, + 157 + ], + "spans": [ + { + "bbox": [ + 148, + 116, + 201, + 157 + ], + "type": "image", + "image_path": "1fbb0598f2f8d14ac6a0ed644a7e8c4b8c2ce8b19d63b2a3689a3c961c628930.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 159, + 197, + 202 + ], + "blocks": [ + { + "bbox": [ + 151, + 159, + 197, + 202 + ], + "lines": [ + { + "bbox": [ + 151, + 159, + 197, + 202 + ], + "spans": [ + { + "bbox": [ + 151, + 159, + 197, + 202 + ], + "type": "image", + "image_path": "4bd6a765902a982d98f947cd0ab0c2a58c89aa660ba68e77994c60f7e7cd50fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 203, + 197, + 245 + ], + "blocks": [ + { + "bbox": [ + 151, + 203, + 197, + 245 + ], + "lines": [ + { + "bbox": [ + 151, + 203, + 197, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 203, + 197, + 245 + ], + "type": "image", + "image_path": "1116ebbb9d339ecabb444f30535805bd4b39de437914b0c9e5d1367dcfce12c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 247, + 187, + 255 + ], + "lines": [ + { + "bbox": [ + 158, + 247, + 187, + 255 + ], + "spans": [ + { + "bbox": [ + 158, + 247, + 187, + 255 + ], + "type": "text", + "content": "MonoSDF" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 260, + 114 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 260, + 114 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 260, + 114 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 260, + 114 + ], + "type": "image", + "image_path": "e3a528cc2808cecba63d53848c12e893630e392b85028ddc6908d52ff3461797.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 115, + 259, + 157 + ], + "blocks": [ + { + "bbox": [ + 206, + 115, + 259, + 157 + ], + "lines": [ + { + "bbox": [ + 206, + 115, + 259, + 157 + ], + "spans": [ + { + "bbox": [ + 206, + 115, + 259, + 157 + ], + "type": "image", + "image_path": "a3484fc088d7f31f8c9fe0aa3866bb47d3e6d6d3e7528fb800423fb3095cd858.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 159, + 255, + 200 + ], + "blocks": [ + { + "bbox": [ + 206, + 159, + 255, + 200 + ], + "lines": [ + { + "bbox": [ + 206, + 159, + 255, + 200 + ], + "spans": [ + { + "bbox": [ + 206, + 159, + 255, + 200 + ], + "type": "image", + "image_path": "8907d9a022d8aaa6107e5f971513366816ed5754232ab0f251c9b9c50bbacd4a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 208, + 202, + 255, + 246 + ], + "blocks": [ + { + "bbox": [ + 208, + 202, + 255, + 246 + ], + "lines": [ + { + "bbox": [ + 208, + 202, + 255, + 246 + ], + "spans": [ + { + "bbox": [ + 208, + 202, + 255, + 246 + ], + "type": "image", + "image_path": "3a1c1de873c88483e7cebb3b4f36e387a2812f183d5661e1c34f60cfac90b2cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 247, + 247, + 255 + ], + "lines": [ + { + "bbox": [ + 215, + 247, + 247, + 255 + ], + "spans": [ + { + "bbox": [ + 215, + 247, + 247, + 255 + ], + "type": "text", + "content": "UFORecon" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 263, + 70, + 317, + 114 + ], + "blocks": [ + { + "bbox": [ + 263, + 70, + 317, + 114 + ], + "lines": [ + { + "bbox": [ + 263, + 70, + 317, + 114 + ], + "spans": [ + { + "bbox": [ + 263, + 70, + 317, + 114 + ], + "type": "image", + "image_path": "540dfc52c2099175311b82afa4f5b4d9683fb39f32d0313078fe3e5de6a98458.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 264, + 115, + 317, + 157 + ], + "blocks": [ + { + "bbox": [ + 264, + 115, + 317, + 157 + ], + "lines": [ + { + "bbox": [ + 264, + 115, + 317, + 157 + ], + "spans": [ + { + "bbox": [ + 264, + 115, + 317, + 157 + ], + "type": "image", + "image_path": "bd37693c17a82b46234fe34f1b828e604e672a3fb66a1e5a2f00f0a696663537.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 272, + 159, + 307, + 201 + ], + "blocks": [ + { + "bbox": [ + 272, + 159, + 307, + 201 + ], + "lines": [ + { + "bbox": [ + 272, + 159, + 307, + 201 + ], + "spans": [ + { + "bbox": [ + 272, + 159, + 307, + 201 + ], + "type": "image", + "image_path": "32f99fe08d624837f5888372013a4326247c67c772c35d3616c4be9847ff45db.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 268, + 202, + 314, + 246 + ], + "blocks": [ + { + "bbox": [ + 268, + 202, + 314, + 246 + ], + "lines": [ + { + "bbox": [ + 268, + 202, + 314, + 246 + ], + "spans": [ + { + "bbox": [ + 268, + 202, + 314, + 246 + ], + "type": "image", + "image_path": "f9b515cb0727f709de61082d891b069bf040e1b1f6d7e427cf732930558239a8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 247, + 305, + 255 + ], + "lines": [ + { + "bbox": [ + 276, + 247, + 305, + 255 + ], + "spans": [ + { + "bbox": [ + 276, + 247, + 305, + 255 + ], + "type": "text", + "content": "S-VolSDF" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 161, + 269, + 447, + 281 + ], + "lines": [ + { + "bbox": [ + 161, + 269, + 447, + 281 + ], + "spans": [ + { + "bbox": [ + 161, + 269, + 447, + 281 + ], + "type": "text", + "content": "Figure 6. Visual comparison on DTU dataset with 3 small-overlapping images." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 323, + 71, + 376, + 114 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 376, + 114 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 376, + 114 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 376, + 114 + ], + "type": "image", + "image_path": "5234037b56fba1b421cdb8b9b1fae1d7cda56a6ade0e3544079022b7bae265b0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 115, + 375, + 157 + ], + "blocks": [ + { + "bbox": [ + 323, + 115, + 375, + 157 + ], + "lines": [ + { + "bbox": [ + 323, + 115, + 375, + 157 + ], + "spans": [ + { + "bbox": [ + 323, + 115, + 375, + 157 + ], + "type": "image", + "image_path": "848987358e9abd5d35bfdc9b94b9ef5f00d540bca6628b79591c13c32dee55b0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 332, + 159, + 365, + 201 + ], + "blocks": [ + { + "bbox": [ + 332, + 159, + 365, + 201 + ], + "lines": [ + { + "bbox": [ + 332, + 159, + 365, + 201 + ], + "spans": [ + { + "bbox": [ + 332, + 159, + 365, + 201 + ], + "type": "image", + "image_path": "c6e8e3a59d606d08d455d4f5b0230b9cce792124eccbe0b31d5b877e7c34b09f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 320, + 202, + 378, + 245 + ], + "blocks": [ + { + "bbox": [ + 320, + 202, + 378, + 245 + ], + "lines": [ + { + "bbox": [ + 320, + 202, + 378, + 245 + ], + "spans": [ + { + "bbox": [ + 320, + 202, + 378, + 245 + ], + "type": "image", + "image_path": "ca6844839871e6b2bdd44366b4dd4fac7c6149c8cb76f5a3063c3e5c21bf613e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 247, + 367, + 255 + ], + "lines": [ + { + "bbox": [ + 333, + 247, + 367, + 255 + ], + "spans": [ + { + "bbox": [ + 333, + 247, + 367, + 255 + ], + "type": "text", + "content": "SparseCraft" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 380, + 70, + 435, + 114 + ], + "blocks": [ + { + "bbox": [ + 380, + 70, + 435, + 114 + ], + "lines": [ + { + "bbox": [ + 380, + 70, + 435, + 114 + ], + "spans": [ + { + "bbox": [ + 380, + 70, + 435, + 114 + ], + "type": "image", + "image_path": "3e4b9c45912f4a7bd4e3c203ec244f3266ace4c66a3733335df1298ddae5dd00.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 380, + 115, + 434, + 157 + ], + "blocks": [ + { + "bbox": [ + 380, + 115, + 434, + 157 + ], + "lines": [ + { + "bbox": [ + 380, + 115, + 434, + 157 + ], + "spans": [ + { + "bbox": [ + 380, + 115, + 434, + 157 + ], + "type": "image", + "image_path": "6aa118f1b7278b997ae3e977b98efc503b036a22ba3b7a66cb36222f03d3cfba.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 383, + 159, + 430, + 202 + ], + "blocks": [ + { + "bbox": [ + 383, + 159, + 430, + 202 + ], + "lines": [ + { + "bbox": [ + 383, + 159, + 430, + 202 + ], + "spans": [ + { + "bbox": [ + 383, + 159, + 430, + 202 + ], + "type": "image", + "image_path": "8b9823d87660d692fabfc045ba6886fcdc515d61174d50d866df8354c48860c3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 383, + 202, + 430, + 245 + ], + "blocks": [ + { + "bbox": [ + 383, + 202, + 430, + 245 + ], + "lines": [ + { + "bbox": [ + 383, + 202, + 430, + 245 + ], + "spans": [ + { + "bbox": [ + 383, + 202, + 430, + 245 + ], + "type": "image", + "image_path": "9d02057027f2e0cd53180fdf93cac62c0207c63130c08860a6de56764af03907.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 247, + 419, + 255 + ], + "lines": [ + { + "bbox": [ + 395, + 247, + 419, + 255 + ], + "spans": [ + { + "bbox": [ + 395, + 247, + 419, + 255 + ], + "type": "text", + "content": "NeuSurf" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 438, + 71, + 492, + 114 + ], + "blocks": [ + { + "bbox": [ + 438, + 71, + 492, + 114 + ], + "lines": [ + { + "bbox": [ + 438, + 71, + 492, + 114 + ], + "spans": [ + { + "bbox": [ + 438, + 71, + 492, + 114 + ], + "type": "image", + "image_path": "6d199e8d4f230f4b9bcdd699e4fbc14914d76a25d0d994cf86fcce866656d79b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 440, + 115, + 491, + 159 + ], + "blocks": [ + { + "bbox": [ + 440, + 115, + 491, + 159 + ], + "lines": [ + { + "bbox": [ + 440, + 115, + 491, + 159 + ], + "spans": [ + { + "bbox": [ + 440, + 115, + 491, + 159 + ], + "type": "image", + "image_path": "ddb57d9f67da2ae741fdc296abd71c0ea90e08c5d96200ba512d3bf2fb77cb02.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 444, + 159, + 484, + 202 + ], + "blocks": [ + { + "bbox": [ + 444, + 159, + 484, + 202 + ], + "lines": [ + { + "bbox": [ + 444, + 159, + 484, + 202 + ], + "spans": [ + { + "bbox": [ + 444, + 159, + 484, + 202 + ], + "type": "image", + "image_path": "d718d6e306504646c416f10ee17338571dc051bf6e5e69c8019513faf8a71796.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 443, + 203, + 487, + 245 + ], + "blocks": [ + { + "bbox": [ + 443, + 203, + 487, + 245 + ], + "lines": [ + { + "bbox": [ + 443, + 203, + 487, + 245 + ], + "spans": [ + { + "bbox": [ + 443, + 203, + 487, + 245 + ], + "type": "image", + "image_path": "9b18ff085e23afa7201f5e9ddfe05bc75c8d20425c0e038d70e11913565b9cdb.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 247, + 477, + 255 + ], + "lines": [ + { + "bbox": [ + 453, + 247, + 477, + 255 + ], + "spans": [ + { + "bbox": [ + 453, + 247, + 477, + 255 + ], + "type": "text", + "content": "FatesGS" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 497, + 72, + 551, + 114 + ], + "blocks": [ + { + "bbox": [ + 497, + 72, + 551, + 114 + ], + "lines": [ + { + "bbox": [ + 497, + 72, + 551, + 114 + ], + "spans": [ + { + "bbox": [ + 497, + 72, + 551, + 114 + ], + "type": "image", + "image_path": "804ec0d600d7f20b65e8c4949a378ffb9eec19698a2cbfeb4fee6a14a29c276f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 497, + 115, + 551, + 158 + ], + "blocks": [ + { + "bbox": [ + 497, + 115, + 551, + 158 + ], + "lines": [ + { + "bbox": [ + 497, + 115, + 551, + 158 + ], + "spans": [ + { + "bbox": [ + 497, + 115, + 551, + 158 + ], + "type": "image", + "image_path": "4733198d95d183b85f599307f0997b589da03ae02e2c4067e0dce0d43c94563f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 504, + 159, + 542, + 201 + ], + "blocks": [ + { + "bbox": [ + 504, + 159, + 542, + 201 + ], + "lines": [ + { + "bbox": [ + 504, + 159, + 542, + 201 + ], + "spans": [ + { + "bbox": [ + 504, + 159, + 542, + 201 + ], + "type": "image", + "image_path": "bd45dd827ccc96796c18ae86319e8ec97292e481c27170d7b591e38571f10175.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 501, + 202, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 501, + 202, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 501, + 202, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 501, + 202, + 547, + 245 + ], + "type": "image", + "image_path": "3ff8fffec39e575eb40b82320151dc22e8a07cf210405c0822065e8c92528ae0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 517, + 248, + 531, + 255 + ], + "lines": [ + { + "bbox": [ + 517, + 248, + 531, + 255 + ], + "spans": [ + { + "bbox": [ + 517, + 248, + 531, + 255 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 301, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 433 + ], + "type": "text", + "content": "suffer from reconstruction ambiguity and failures, highlighting that relying solely on simplistic geometric prior constraints is insufficient to obtain complete and accurate meshes. UFORecon [28] shows significant roughness in its reconstruction results. S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit noticeable reconstruction defects. Experimental results demonstrate that our method is effective in alleviating geometric and appearance ambiguities during the optimization process. This significantly enhances the quality of mesh reconstruction, especially in scenarios with small overlapping views and low texture." + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 56, + 444, + 113, + 555 + ], + "blocks": [ + { + "bbox": [ + 56, + 444, + 113, + 555 + ], + "lines": [ + { + "bbox": [ + 56, + 444, + 113, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 113, + 555 + ], + "type": "image", + "image_path": "e139c7de8fd0b69d91c91309f8ce88d8784fdd13540c9ce32b5df4390519c5a2.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 555, + 99, + 563 + ], + "lines": [ + { + "bbox": [ + 72, + 555, + 99, + 563 + ], + "spans": [ + { + "bbox": [ + 72, + 555, + 99, + 563 + ], + "type": "text", + "content": "GT Images" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 553, + 143, + 562 + ], + "lines": [ + { + "bbox": [ + 117, + 553, + 143, + 562 + ], + "spans": [ + { + "bbox": [ + 117, + 553, + 143, + 562 + ], + "type": "text", + "content": "MonoSDF" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 575, + 295, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 597 + ], + "type": "text", + "content": "Figure 7. Visual comparison on BlendedMVS dataset. " + }, + { + "bbox": [ + 55, + 575, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 575, + 295, + 597 + ], + "type": "text", + "content": " indicates reconstruction failure." + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 118, + 444, + 143, + 525 + ], + "blocks": [ + { + "bbox": [ + 118, + 444, + 143, + 525 + ], + "lines": [ + { + "bbox": [ + 118, + 444, + 143, + 525 + ], + "spans": [ + { + "bbox": [ + 118, + 444, + 143, + 525 + ], + "type": "image", + "image_path": "750bf527000ee7ffc93382c4ca91b44018762741e3c4454082176563084aa809.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 536, + 136, + 544 + ], + "lines": [ + { + "bbox": [ + 128, + 536, + 136, + 544 + ], + "spans": [ + { + "bbox": [ + 128, + 536, + 136, + 544 + ], + "type": "text", + "content": "X" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 151, + 445, + 183, + 523 + ], + "blocks": [ + { + "bbox": [ + 151, + 445, + 183, + 523 + ], + "lines": [ + { + "bbox": [ + 151, + 445, + 183, + 523 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 183, + 523 + ], + "type": "image", + "image_path": "2f2e5535ecc3df2c42535293861601229deaaa7c6e8914b0a340aab6aae2e3ae.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 150, + 525, + 184, + 555 + ], + "blocks": [ + { + "bbox": [ + 150, + 525, + 184, + 555 + ], + "lines": [ + { + "bbox": [ + 150, + 525, + 184, + 555 + ], + "spans": [ + { + "bbox": [ + 150, + 525, + 184, + 555 + ], + "type": "image", + "image_path": "284d239afac530b1d5468a34271c5f844108913ea5d07e4b7199b8983a8048e6.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 555, + 179, + 563 + ], + "lines": [ + { + "bbox": [ + 154, + 555, + 179, + 563 + ], + "spans": [ + { + "bbox": [ + 154, + 555, + 179, + 563 + ], + "type": "text", + "content": "S-VolSDF" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 186, + 445, + 214, + 525 + ], + "blocks": [ + { + "bbox": [ + 186, + 445, + 214, + 525 + ], + "lines": [ + { + "bbox": [ + 186, + 445, + 214, + 525 + ], + "spans": [ + { + "bbox": [ + 186, + 445, + 214, + 525 + ], + "type": "image", + "image_path": "42d69eecfa0ada455d2303e7fe3238e2b683c9bfa70f4c20414cc4d6df2f659d.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 186, + 525, + 220, + 555 + ], + "blocks": [ + { + "bbox": [ + 186, + 525, + 220, + 555 + ], + "lines": [ + { + "bbox": [ + 186, + 525, + 220, + 555 + ], + "spans": [ + { + "bbox": [ + 186, + 525, + 220, + 555 + ], + "type": "image", + "image_path": "e8a2fd7e623536b8a49b54eea3bfe7f3c66b1e2cf39d446421535349d055d45e.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 555, + 214, + 563 + ], + "lines": [ + { + "bbox": [ + 192, + 555, + 214, + 563 + ], + "spans": [ + { + "bbox": [ + 192, + 555, + 214, + 563 + ], + "type": "text", + "content": "NeuSurf" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 222, + 445, + 255, + 525 + ], + "blocks": [ + { + "bbox": [ + 222, + 445, + 255, + 525 + ], + "lines": [ + { + "bbox": [ + 222, + 445, + 255, + 525 + ], + "spans": [ + { + "bbox": [ + 222, + 445, + 255, + 525 + ], + "type": "image", + "image_path": "b37806df4a5809c1f843a618567856c08790fb2825d6a72235962bedef785f09.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 222, + 525, + 256, + 555 + ], + "blocks": [ + { + "bbox": [ + 222, + 525, + 256, + 555 + ], + "lines": [ + { + "bbox": [ + 222, + 525, + 256, + 555 + ], + "spans": [ + { + "bbox": [ + 222, + 525, + 256, + 555 + ], + "type": "image", + "image_path": "9c7c1bacb02c966cbdb4d82cbe62a27903985040aefd53f75e0f775235f75497.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 555, + 250, + 563 + ], + "lines": [ + { + "bbox": [ + 228, + 555, + 250, + 563 + ], + "spans": [ + { + "bbox": [ + 228, + 555, + 250, + 563 + ], + "type": "text", + "content": "FatesGS" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 258, + 444, + 294, + 555 + ], + "blocks": [ + { + "bbox": [ + 258, + 444, + 294, + 555 + ], + "lines": [ + { + "bbox": [ + 258, + 444, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 258, + 444, + 294, + 555 + ], + "type": "image", + "image_path": "e6b0709656a8a4838036c97717cbcde6d23bbaa83161e4c36eb552356b313bb9.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 555, + 282, + 563 + ], + "lines": [ + { + "bbox": [ + 269, + 555, + 282, + 563 + ], + "spans": [ + { + "bbox": [ + 269, + 555, + 282, + 563 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": "Reconstruction on BlendedMVS. Figure 7 presents the visual comparison of reconstructed mesh for overfitting-based methods. With only 3 small-overlapping views provided, all of the generalization-based methods completely fail to reconstruct in the sparse setting of BlendedMVS[44] dataset, even if SparseNeuS [24] is fine-tuned. Therefore, the reconstruction results of these methods are not included" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 301, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 456 + ], + "type": "text", + "content": "in Figure 7. Compared to other methods, our approach can generate more complete and detailed meshes. Similarly, MonoSDF [47] fails to reconstruct surfaces either. The meshes generated by S-VolSDF [40], NeuSurf [13] and FatesGS [14] exhibit significant defects. Both NeurSurf [13] and FatesGS [14] use on-surface feature consistency constraints, but the reconstruction results are still not good enough. In contrast, our method achieves more comprehensive geometry and finer details by employing volume rendering-based feature consistency constraints. This highlights the advantages of our approach in geometric consistency. More visualizations are presented in the supplementary materials." + } + ] + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 332, + 466, + 376, + 507 + ], + "blocks": [ + { + "bbox": [ + 318, + 476, + 326, + 499 + ], + "lines": [ + { + "bbox": [ + 318, + 476, + 326, + 499 + ], + "spans": [ + { + "bbox": [ + 318, + 476, + 326, + 499 + ], + "type": "text", + "content": "Scan 40" + } + ] + } + ], + "index": 58, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 466, + 376, + 507 + ], + "lines": [ + { + "bbox": [ + 332, + 466, + 376, + 507 + ], + "spans": [ + { + "bbox": [ + 332, + 466, + 376, + 507 + ], + "type": "image", + "image_path": "b2e14c2d9288456a7f4f6b8f98b5b1a414bcb81b7b85d4f82a0263d256de7a2b.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 332, + 508, + 378, + 552 + ], + "blocks": [ + { + "bbox": [ + 318, + 517, + 326, + 543 + ], + "lines": [ + { + "bbox": [ + 318, + 517, + 326, + 543 + ], + "spans": [ + { + "bbox": [ + 318, + 517, + 326, + 543 + ], + "type": "text", + "content": "Scan 110" + } + ] + } + ], + "index": 60, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 508, + 378, + 552 + ], + "lines": [ + { + "bbox": [ + 332, + 508, + 378, + 552 + ], + "spans": [ + { + "bbox": [ + 332, + 508, + 378, + 552 + ], + "type": "image", + "image_path": "85612e3143694ee3e593587cac774d085724692cdf55ddec13def2e89ae45fbf.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 552, + 375, + 570 + ], + "lines": [ + { + "bbox": [ + 335, + 552, + 375, + 570 + ], + "spans": [ + { + "bbox": [ + 335, + 552, + 375, + 570 + ], + "type": "text", + "content": "Mesh w/ " + }, + { + "bbox": [ + 335, + 552, + 375, + 570 + ], + "type": "inline_equation", + "content": "L_{\\text{feat}}" + }, + { + "bbox": [ + 335, + 552, + 375, + 570 + ], + "type": "text", + "content": " by ours" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 583, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 313, + 583, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 555, + 605 + ], + "type": "text", + "content": "Figure 8. Reconstructed meshes and error maps on DTU dataset with different feature consistency losses." + } + ] + } + ], + "index": 72, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 388, + 466, + 430, + 508 + ], + "blocks": [ + { + "bbox": [ + 388, + 466, + 430, + 508 + ], + "lines": [ + { + "bbox": [ + 388, + 466, + 430, + 508 + ], + "spans": [ + { + "bbox": [ + 388, + 466, + 430, + 508 + ], + "type": "image", + "image_path": "0d82f1550c1a04c064fbe70b6f71c81e668383cbf078cb3ba61675481fe153e5.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 388, + 509, + 434, + 552 + ], + "blocks": [ + { + "bbox": [ + 388, + 509, + 434, + 552 + ], + "lines": [ + { + "bbox": [ + 388, + 509, + 434, + 552 + ], + "spans": [ + { + "bbox": [ + 388, + 509, + 434, + 552 + ], + "type": "image", + "image_path": "d2be16916de4bde2d6d84c72f7d19cc87bbb77f713247442bdeba7c52ccb2afa.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 552, + 435, + 570 + ], + "lines": [ + { + "bbox": [ + 383, + 552, + 435, + 570 + ], + "spans": [ + { + "bbox": [ + 383, + 552, + 435, + 570 + ], + "type": "text", + "content": "Error map w/ " + }, + { + "bbox": [ + 383, + 552, + 435, + 570 + ], + "type": "inline_equation", + "content": "L_{\\text{feat}}" + }, + { + "bbox": [ + 383, + 552, + 435, + 570 + ], + "type": "text", + "content": " by ours" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 552, + 486, + 571 + ], + "lines": [ + { + "bbox": [ + 429, + 552, + 486, + 571 + ], + "spans": [ + { + "bbox": [ + 429, + 552, + 486, + 571 + ], + "type": "text", + "content": "at Mesh w/ surf " + }, + { + "bbox": [ + 429, + 552, + 486, + 571 + ], + "type": "inline_equation", + "content": "f_{\\text{feat}}" + }, + { + "bbox": [ + 429, + 552, + 486, + 571 + ], + "type": "text", + "content": " by MVSDF" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 445, + 466, + 490, + 509 + ], + "blocks": [ + { + "bbox": [ + 445, + 466, + 490, + 509 + ], + "lines": [ + { + "bbox": [ + 445, + 466, + 490, + 509 + ], + "spans": [ + { + "bbox": [ + 445, + 466, + 490, + 509 + ], + "type": "image", + "image_path": "6471b5b4c7c6ca2584c4b631c3c70db1768c4c07689cc7ed8de0cfa5d45a517d.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 438, + 510, + 491, + 551 + ], + "blocks": [ + { + "bbox": [ + 438, + 510, + 491, + 551 + ], + "lines": [ + { + "bbox": [ + 438, + 510, + 491, + 551 + ], + "spans": [ + { + "bbox": [ + 438, + 510, + 491, + 551 + ], + "type": "image", + "image_path": "27bb6e20c01eae7f46cd2166eaa1dbc0b9ed988c3c36c4be0f96a244aef12996.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 494, + 466, + 547, + 509 + ], + "blocks": [ + { + "bbox": [ + 494, + 466, + 547, + 509 + ], + "lines": [ + { + "bbox": [ + 494, + 466, + 547, + 509 + ], + "spans": [ + { + "bbox": [ + 494, + 466, + 547, + 509 + ], + "type": "image", + "image_path": "2586dba55b8aec8296c88d3413c0f39acc451fa281d08ee53b14c18c10168ffa.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + } + ], + "index": 69 + }, + { + "type": "image", + "bbox": [ + 494, + 510, + 548, + 551 + ], + "blocks": [ + { + "bbox": [ + 494, + 510, + 548, + 551 + ], + "lines": [ + { + "bbox": [ + 494, + 510, + 548, + 551 + ], + "spans": [ + { + "bbox": [ + 494, + 510, + 548, + 551 + ], + "type": "image", + "image_path": "9166b28abdee442f792d892293b2e13a1fa28d43ff64575d719b8449e61a47ee.jpg" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 552, + 548, + 571 + ], + "lines": [ + { + "bbox": [ + 495, + 552, + 548, + 571 + ], + "spans": [ + { + "bbox": [ + 495, + 552, + 548, + 571 + ], + "type": "text", + "content": "Error map w/ " + }, + { + "bbox": [ + 495, + 552, + 548, + 571 + ], + "type": "inline_equation", + "content": "L_{\\text{feat}}^{\\text{surf}}" + }, + { + "bbox": [ + 495, + 552, + 548, + 571 + ], + "type": "text", + "content": " by MVSDF" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 70 + }, + { + "bbox": [ + 313, + 624, + 408, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 408, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 408, + 637 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "We evaluate the components of our method with 3 small-overlapping views by an ablation study on the DTU [16] dataset. To compare the depth loss " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{depth}^{mono}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " calibrated by rendered depth in MonoSDF [47] with our depth loss " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{depth}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": ", we replace " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{depth}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{depth}^{mono}" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " to evaluate it in our method. We also compare the volume rendering-based" + } + ] + } + ], + "index": 74 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28520" + } + ] + } + ], + "index": 75 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 295, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 295, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 295, + 159 + ], + "type": "table", + "html": "
MethodLcolorLfeatLdepthLmono depthL1featL2featsurffeatCD↓
Baseline3.35
✓1.76
✓✓1.47
✓✓1.62
✓✓✓1.11
✓✓✓1.59
✓✓✓2.36
✓✓✓1.81
✓✓✓2.93
", + "image_path": "0b024ef5a3e4040cfabf5ed8069817cd1e0b120cfba75112561a07d647fdc259.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 295, + 190 + ], + "type": "text", + "content": "Table 2. Ablation studies on DTU dataset with 3 small-overlapping images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 202, + 116, + 293 + ], + "blocks": [ + { + "bbox": [ + 57, + 202, + 116, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 202, + 116, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 202, + 116, + 293 + ], + "type": "image", + "image_path": "451bb413ec3309041737c2a5b784522cd14538751a6dbdfd7c5ba2c2f15ef45c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 294, + 97, + 301 + ], + "lines": [ + { + "bbox": [ + 76, + 294, + 97, + 301 + ], + "spans": [ + { + "bbox": [ + 76, + 294, + 97, + 301 + ], + "type": "text", + "content": "Step 10k" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 116, + 202, + 176, + 249 + ], + "blocks": [ + { + "bbox": [ + 116, + 202, + 176, + 249 + ], + "lines": [ + { + "bbox": [ + 116, + 202, + 176, + 249 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 176, + 249 + ], + "type": "image", + "image_path": "131b2eec1aaa987c6c2fc44b3bf9e8506d1202bfe6add5f806c15e13e8bceda7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 116, + 248, + 176, + 293 + ], + "blocks": [ + { + "bbox": [ + 116, + 248, + 176, + 293 + ], + "lines": [ + { + "bbox": [ + 116, + 248, + 176, + 293 + ], + "spans": [ + { + "bbox": [ + 116, + 248, + 176, + 293 + ], + "type": "image", + "image_path": "3dc8b7ee7873229fbc42d7b82e09bfb5814477cbf41799d23abe17dc7f476241.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 294, + 157, + 301 + ], + "lines": [ + { + "bbox": [ + 135, + 294, + 157, + 301 + ], + "spans": [ + { + "bbox": [ + 135, + 294, + 157, + 301 + ], + "type": "text", + "content": "Step 20k" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 177, + 202, + 235, + 249 + ], + "blocks": [ + { + "bbox": [ + 177, + 202, + 235, + 249 + ], + "lines": [ + { + "bbox": [ + 177, + 202, + 235, + 249 + ], + "spans": [ + { + "bbox": [ + 177, + 202, + 235, + 249 + ], + "type": "image", + "image_path": "565976665936a09e43f0998137701f1ac6ad3b225dd18998d44903bee55a0d4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 177, + 248, + 235, + 293 + ], + "blocks": [ + { + "bbox": [ + 177, + 248, + 235, + 293 + ], + "lines": [ + { + "bbox": [ + 177, + 248, + 235, + 293 + ], + "spans": [ + { + "bbox": [ + 177, + 248, + 235, + 293 + ], + "type": "image", + "image_path": "18d4841a6253c82ef59dfb7b19add450704ec4221adc3f4d857b2be65c3111fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 294, + 216, + 301 + ], + "lines": [ + { + "bbox": [ + 194, + 294, + 216, + 301 + ], + "spans": [ + { + "bbox": [ + 194, + 294, + 216, + 301 + ], + "type": "text", + "content": "Step 40k" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 202, + 294, + 249 + ], + "blocks": [ + { + "bbox": [ + 236, + 202, + 294, + 249 + ], + "lines": [ + { + "bbox": [ + 236, + 202, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 236, + 202, + 294, + 249 + ], + "type": "image", + "image_path": "238742f5599328f8c295cb67b70794054114a20205eb71cbb7a81fa4c44b8a2d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 236, + 248, + 294, + 293 + ], + "blocks": [ + { + "bbox": [ + 236, + 248, + 294, + 293 + ], + "lines": [ + { + "bbox": [ + 236, + 248, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 236, + 248, + 294, + 293 + ], + "type": "image", + "image_path": "bdd4739c5e95bcac5bb70a44287df07815f86a7561776eafecbab25c24ce8632.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 294, + 276, + 301 + ], + "lines": [ + { + "bbox": [ + 253, + 294, + 276, + 301 + ], + "spans": [ + { + "bbox": [ + 253, + 294, + 276, + 301 + ], + "type": "text", + "content": "Step 60k" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 60, + 349, + 294, + 534 + ], + "blocks": [ + { + "bbox": [ + 55, + 314, + 295, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 336 + ], + "type": "text", + "content": "Figure 9. The variation of weighted feature similarity during training, brighter colors indicate higher feature similarity." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 349, + 294, + 534 + ], + "lines": [ + { + "bbox": [ + 60, + 349, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 60, + 349, + 294, + 534 + ], + "type": "image", + "image_path": "b585ad7b38d2901e1288f5bfda5791d6ec524df280ec939f83997a0f3f50aefd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 579 + ], + "type": "text", + "content": "Figure 10. Visualization of reconstruction and error maps for scene scan24 and scan37 in DTU dataset with different losses. The differences of error maps are highlighted." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": "feature consistency loss calculated using L1 distance (denoted as " + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "inline_equation", + "content": "L_{\\text{feat}}^{L1}" + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": ") and L2 distance (denoted as " + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "inline_equation", + "content": "L_{\\text{feat}}^{L2}" + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": ") with our method using feature similarity distance. We found that feature similarity distance is better than both L1 and L2 distance, as shown in Table 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": "In addition, we replace our volume rendering-based feature consistency loss " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{feat}" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " with the on-surface feature consistency loss " + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{feat}^{surf}" + }, + { + "bbox": [ + 55, + 664, + 296, + 714 + ], + "type": "text", + "content": " used in MVSDF [51] to compare the effectiveness of two different loss functions. Figure 8 illustrates the performance of these two different loss functions." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 382, + 105 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 382, + 105 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 382, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 382, + 105 + ], + "type": "image", + "image_path": "a93de04f8c54b87ca5aa86f8524b25e96821ab93218da111d0e9957273a690e7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 105, + 363, + 113 + ], + "lines": [ + { + "bbox": [ + 335, + 105, + 363, + 113 + ], + "spans": [ + { + "bbox": [ + 335, + 105, + 363, + 113 + ], + "type": "text", + "content": "GT Images" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 424, + 101 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 424, + 101 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 424, + 101 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 424, + 101 + ], + "type": "image", + "image_path": "e5e2d2251d0b4eec902efa1ea1f7eef70213e789d50ec49e4c3e8d9414f4d6a9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 105, + 416, + 112 + ], + "lines": [ + { + "bbox": [ + 389, + 105, + 416, + 112 + ], + "spans": [ + { + "bbox": [ + 389, + 105, + 416, + 112 + ], + "type": "text", + "content": "S-VolSDF" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 466, + 103 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 466, + 103 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 466, + 103 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 466, + 103 + ], + "type": "image", + "image_path": "03dd7a95ec283cd2a379c445e8d07dd8229d276449f1c929f297781bca1804df.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 105, + 456, + 112 + ], + "lines": [ + { + "bbox": [ + 434, + 105, + 456, + 112 + ], + "spans": [ + { + "bbox": [ + 434, + 105, + 456, + 112 + ], + "type": "text", + "content": "NeuSurf" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 124, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 313, + 124, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 553, + 148 + ], + "type": "text", + "content": "Figure 11. Failure case. For specular objects, the ambiguity in the color consistency constraint may lead to a rough surface." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 468, + 72, + 509, + 103 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 509, + 103 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 509, + 103 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 509, + 103 + ], + "type": "image", + "image_path": "7fcc5f9a7a10ef9f527e262beedd3796d6f6220251f49abb4ceaff1f7337da73.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 105, + 499, + 112 + ], + "lines": [ + { + "bbox": [ + 477, + 105, + 499, + 112 + ], + "spans": [ + { + "bbox": [ + 477, + 105, + 499, + 112 + ], + "type": "text", + "content": "FatesGS" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 511, + 72, + 553, + 103 + ], + "blocks": [ + { + "bbox": [ + 511, + 72, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 511, + 72, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 511, + 72, + 553, + 103 + ], + "type": "image", + "image_path": "a6f12d9fe4bb0baee019f107fb50213f5860956a32c6570b65995cf268da237f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 525, + 105, + 538, + 112 + ], + "lines": [ + { + "bbox": [ + 525, + 105, + 538, + 112 + ], + "spans": [ + { + "bbox": [ + 525, + 105, + 538, + 112 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": "trates the reconstruction results and error maps on the DTU dataset when using different feature consistency losses, under the on-surface feature consistency loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "L_{feat}^{surf}" + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": ", the meshes show large artifacts." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "type": "text", + "content": "Table 2 shows the average Chamfer Distance over all 11 scenes on DTU dataset using different losses. The experimental results indicate that both feature consistency loss and uncertainty-guided depth constraint improve the surface reconstruction." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 335 + ], + "type": "text", + "content": "Figure 9 illustrates the variation of the weighted feature similarity map during the training process. Brighter colors indicate higher feature similarity, demonstrating that our volume rendering-based feature consistency loss can provide effective constraints." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 335, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 422 + ], + "type": "text", + "content": "Figure 10 shows the reconstructed meshes and error maps for scene scan24 and scan37 on the DTU [16] dataset when using different losses. It can be observed that the mesh deteriorates without the volume rendering-based feature consistency loss or the uncertainty-guided depth constraint loss, and the reconstruction quality drops when using the depth loss " + }, + { + "bbox": [ + 313, + 335, + 554, + 422 + ], + "type": "inline_equation", + "content": "L_{depth}^{mono}" + }, + { + "bbox": [ + 313, + 335, + 554, + 422 + ], + "type": "text", + "content": " in MonoSDF [47]." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 430, + 391, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 391, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 391, + 441 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 545 + ], + "type": "text", + "content": "We propose a novel method for learning implicit representations from sparse views with small overlaps. Our novelty lies in a novel volume rendering-based feature consistency loss and an uncertainty-guided depth constraint. Extensive experiments on the DTU [16] and BlendedMVS [44] datasets show that our method surpasses existing state-of-the-art sparse-view reconstruction methods in terms of reconstruction quality." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 713 + ], + "type": "text", + "content": "Limitations. Although our method shows significant improvement over other sparse view reconstruction methods, there are still some limitations. Firstly, for specular objects, the ambiguity in the color consistency constraint may lead to a rough surface, as shown in Figure 11. Secondly, Following previous studies [13, 24, 40], the camera poses of sparse views are obtained from the training dataset. However, in some cases, it may not be possible to obtain accurate camera poses using SfM methods like COLMAP [33] due to the lack of texture in the images or excessive viewing angles. Additionally, the feature consistency constraint method requires a pre-trained network to extract image features. The accuracy of the features determines the performance of the feature consistency constraint." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28521" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 160, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "The corresponding authors are Yu-Shen Liu and Haichuan Song. This work was supported by National Key R&D Program of China (2022YFC3800600), and the National Natural Science Foundation of China (62272263), and in part by Tsinghua-Kuaishou Institute of Future Media Data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 250 + ], + "type": "text", + "content": "[1] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, PP: 1-12, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 284 + ], + "type": "text", + "content": "[2] Hanlin Chen, Chen Li, and Gim Hee Lee. NeuSG: Neural implicit surface reconstruction with 3D gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 285, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 296, + 341 + ], + "type": "text", + "content": "[3] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. MVSplat: Efficient 3D gaussian splatting from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "type": "text", + "content": "[4] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "text", + "content": "[5] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "text", + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised NeRF: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 296, + 554 + ], + "type": "text", + "content": "[7] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-Neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 556, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 601 + ], + "type": "text", + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of Machine Learning and Systems 2020, pages 3569-3579. 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 657 + ], + "type": "text", + "content": "[9] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[11] Liang Han, Junsheng Zhou, Yu-Shen Liu, and Zhizhong Han. Binocular-guided 3D gaussian splattering with view consistency for sparse view synthesis. In Advances in Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 162 + ], + "type": "text", + "content": "[12] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Proceedings, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "type": "text", + "content": "[13] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. NeuSurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[14] Han Huang, Yulun Wu, Chao Deng, Ge Gao, Ming Gu, and Yu-Shen Liu. FatesGS: Fast and accurate sparse-view surface reconstruction using gaussian splatting with depth-feature consistency. In Proceedings of the AAAI Conference on Artificial Intelligence, 2025. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 330 + ], + "type": "text", + "content": "[15] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting NeRF on a Diet: Semantically consistent few-shot view synthesis implementation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 332, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 554, + 385 + ], + "type": "text", + "content": "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanas. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 2, 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[17] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International Conference on Machine Learning, pages 5156-5165, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 476 + ], + "type": "text", + "content": "[18] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 554, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 554, + 542 + ], + "type": "text", + "content": "[19] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 544, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 554, + 589 + ], + "type": "text", + "content": "[20] Shujuan Li, Yu-Shen Liu, and Zhizhong Han. GaussianUDF: Inferring unsigned distance functions through 3D gaussian splatting. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 27113-27123, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "text", + "content": "[21] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "text", + "content": "[22] Yixun Liang, Hao He, and Yingcong Chen. ReTR: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[23] Xinqi Liu, Jituo Li, and Guodong Lu. Reconstructing complex shaped clothing from a single image with feature stable" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28522" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "unsigned distance fields. IEEE Transactions on Visualization and Computer Graphics, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[24] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. SparseNeuS: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[25] Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 3DGSR: Implicit surface reconstruction with 3D gaussian splatting. ACM Transactions on Graphics, 43(6):1-12, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "text", + "content": "[26] Baorui Ma, Yu-Shen Liu, and Zhizhong Han. Learning signed distance functions from noisy 3D point clouds via noise to noise mapping. In International Conference on Machine Learning, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 370 + ], + "type": "text", + "content": "[28] Youngju Na, Woo Jae Kim, Kyu Beom Han, Suhyeon Ha, and Sung-Eui Yoon. UFORecon: Generalizable sparse-view surface reconstruction from arbitrary and unfavorable sets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5094-5104, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 435 + ], + "type": "text", + "content": "[29] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "text", + "content": "[30] Takeshi Noda, Chao Chen, Junsheng Zhou, Weiqi Zhang, Yu-Shen Liu, and Zhizhong Han. Learning bijective surface parameterization for inferring signed distance functions from sparse point clouds with grid deformation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 22139-22149, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[31] Rui Peng, Xiaodong Gu, Luyang Tang, Shihe Shen, Fanqi Yu, and Ronggang Wang. GenS: Generalizable neural surface reconstruction from multi-view images. In Advances in Neural Information Processing Systems, pages 56932-56945, 2023. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[32] Yufan Ren, Tong Zhang, Marc Pollefeys, Sabine Susstrunk, and Fangjinhua Wang. VolRecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[33] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, 2016. 2, 4, 5, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[34] Hui Tian, Chenyang Zhu, Yifei Shi, and Kai Xu. SuperUDF: Self-supervised udf estimation for surface reconstruction. IEEE Transactions on Visualization and Computer Graphics, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. SPARF: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[36] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. SparseNeRF: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "type": "text", + "content": "[37] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[38] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. NeuS2: Fast learning of neural implicit surfaces for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3295–3306, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "text", + "content": "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[40] Haoyu Wu, Alexandros Graikos, and Dimitris Samaras. SVolSDF: Sparse multi-view stereo regularization of neural implicit surfaces. International Conference on Computer Vision, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 435 + ], + "type": "text", + "content": "[41] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. International Conference on Learning Representations, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "text", + "content": "[42] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2F2NeUS: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[43] Jiawei Yang, Marco Pavone, and Yue Wang. FreeNeRF: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[44] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 667 + ], + "type": "text", + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[46] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. SparseCraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28523" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[47] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. MonoSDF: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in Neural Information Processing Systems, 35:25018-25032, 2022. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 128, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 172 + ], + "type": "text", + "content": "[48] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian Opacity Fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6): 1-13, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 228 + ], + "type": "text", + "content": "[49] Yu-Jie Yuan, Yu-Kun Lai, Yi-Hua Huang, Leif Kobbelt, and Lin Gao. Neural radiance fields from sparse rgb-d images for high-quality view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8713-8728, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 295, + 261 + ], + "type": "text", + "content": "[50] Jingyang Zhang, Yao Yao, Shiwei Li, Zixin Luo, and Tian Fang. Visibility-aware multi-view stereo network. The British Machine Vision Conference, 2020. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 57, + 261, + 295, + 304 + ], + "type": "text", + "content": "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 305, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 305, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 57, + 305, + 295, + 350 + ], + "type": "text", + "content": "[52] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. CoR-GS: sparse-view 3D gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 350, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 350, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 295, + 393 + ], + "type": "text", + "content": "[53] Wenyuan Zhang, Ruofan Xing, Yunfan Zeng, Yu-Shen Liu, Kanle Shi, and Zhizhong Han. Fast learning radiance fields by shooting much fewer rays. IEEE Transactions on Image Processing, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 394, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 394, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 394, + 295, + 437 + ], + "type": "text", + "content": "[54] Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. Neural signed distance function inference through splatting 3D gaussians pulled on zero-level set. In Advances in Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 438, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 295, + 502 + ], + "type": "text", + "content": "[55] Wenyuan Zhang, Emily Yue-ting Jia, Junsheng Zhou, Baorui Ma, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. NeRFPrior: Learning neural radiance field as a prior for indoor scene reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 11317–11327, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 503, + 295, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 503, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 57, + 503, + 295, + 570 + ], + "type": "text", + "content": "[56] Wenyuan Zhang, Yixiao Yang, Han Huang, Liang Han, Kanle Shi, Yu-Shen Liu, and Zhizhong Han. MonoInstance: Enhancing monocular priors via multi-view instance alignment for neural rendering and reconstruction. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21642-21653, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 571, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 295, + 613 + ], + "type": "text", + "content": "[57] Weiqi Zhang, Junsheng Zhou, Haotian Geng, Wenyuan Zhang, and Yu-Shen Liu. GAP: Gaussianize any point clouds with text guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 669 + ], + "type": "text", + "content": "[58] Junsheng Zhou, Baorui Ma, Shujuan Li, Yu-Shen Liu, Yi Fang, and Zhizhong Han. CAP-UDF: Learning unsigned distance functions progressively from raw point clouds with consistency-aware field optimization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[59] Junsheng Zhou, Baorui Ma, and Liu Yu-Shen. Fast learning of signed distance functions from noisy point clouds via noise to noise mapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 162 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 118 + ], + "type": "text", + "content": "[60] Tiansong Zhou, Jing Huang, Tao Yu, Ruizhi Shao, and Kun Li. HDhuman: High-quality human novel-view rendering from sparse views. IEEE Transactions on Visualization and Computer Graphics, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[61] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. FSGS: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision, pages 145-163. Springer, 2025. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28524" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_content_list.json b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18796d0523613a30be066acb99daf7e52b12d657 --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference", + "text_level": 1, + "bbox": [ + 145, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Samir Khaki", + "bbox": [ + 135, + 179, + 246, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Konstantinos N. Plataniotis4", + "bbox": [ + 184, + 200, + 408, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVIDIA", + "bbox": [ + 248, + 224, + 330, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ MIT", + "bbox": [ + 348, + 224, + 398, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaming Tang²", + "bbox": [ + 436, + 180, + 553, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yao Lu", + "bbox": [ + 450, + 200, + 517, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ UC San Diego", + "bbox": [ + 416, + 224, + 542, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shang Yang2", + "bbox": [ + 563, + 180, + 699, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Song Han1,2", + "bbox": [ + 558, + 200, + 656, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yukang Chen", + "bbox": [ + 738, + 180, + 854, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhijian Liu $^{1,3}$", + "bbox": [ + 697, + 200, + 805, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ University of Toronto", + "bbox": [ + 562, + 224, + 743, + 243 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fc9719806016f3661bfa5c364611f29c8b92b6c23c53ce8c39fb353eadfdd14.jpg", + "image_caption": [ + "Multi-Round Performance" + ], + "image_footnote": [], + "bbox": [ + 96, + 290, + 359, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21bf3170cfdc35fef1d36474a6efff25ddb4cae0791f21a9e123baf1fd2c06dd.jpg", + "image_caption": [ + "Image-Centric Evaluation", + "Figure 1. Overview of SparseVILA. Compared to existing approaches, SparseVILA exhibits advantages in the following aspects: 1. Multi-Round Performance. SparseVILA leverages a decoupled sparsity framework to achieve state-of-the-art performance at a $1.4 \\times$ speedup on LLaVA-NeXT [28]. 2. Image-centric evaluation. Leveraging decoupled sparsity, SparseVILA can retain higher context tokens resulting in a $5.9\\%$ margin over previous approaches. 3. SparseVILA scales to long-context and long-generation tasks, achieving lossless performance with up to a $3.6 \\times$ context, $1.7 \\times$ generation, and $2.2 \\times$ end-to-end speedup." + ], + "image_footnote": [], + "bbox": [ + 366, + 290, + 629, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/394567e0efa79343b86aa01a409fd220861792fd57d1bfadc8ac066c36f4b041.jpg", + "image_caption": [ + "Long-Context/Generation Performance" + ], + "image_footnote": [], + "bbox": [ + 633, + 290, + 901, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 547, + 326, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM, by pruning tokens in the context stage while attending to a sparse subset of visual tokens during the generation phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a $1.4 \\times$ speedup on image benchmarks. This approach leads to $+5.9\\%$ average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a $3.6 \\times$ and $1.7 \\times$ speedup in prefetch and decoding, respectively.", + "bbox": [ + 88, + 580, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 547, + 643, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs) have emerged as a state-of-the-art conversational tool enabling users to directly interact with Large Language Models (LLMs) using various visual features, including photographs, documents, and videos [2, 24, 27, 30, 47]. Unfortunately, this added modality comes at the expense of higher latency and memory associated with processing the visual tokens in the LLM. Hence, deploying LLMs efficiently at inference time remains a challenge.", + "bbox": [ + 509, + 575, + 906, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several works have aimed to reduce these associated costs through model pruning on the LLM or vision encoder [32, 41, 64], KV cache compression [13, 26, 29], and most recently, token sparsification [60]. By reducing the amount of computation in the inference pipeline, many of these methods can achieve significant context stage savings as this stage of the network is mainly compute-bound.", + "bbox": [ + 509, + 700, + 908, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Looking beyond the context, real-world applications often demand extensive generation. Tasks such as image captioning may require a few hundred tokens, meanwhile, video captioning or detailing easily requires more than a few thousand generated tokens. Hence in such applications, it is not sufficient to only focus on context stage optimization – effi", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23784", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cient implementations with real-world applications should focus on both context and decoding stage optimizations. One such example is multi-turn conversation.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-turn/round conversation is a practical use case for VLMs, as it is entirely possible for a user to pose multiple questions about a given visual input. In fact, most benchmarks inherently support multi-round conversation: the GQA dataset [15] has more than 90 questions for the same visual input. Despite this, most evaluation benchmarks run single-round evaluation (i.e. repetitive pre-filling), which is not only unrealistic but also inefficient as the context stage would be repeated for each generation round. In real-world scenarios, the visual input could span tens of thousands of context tokens, hence repetitive pre-filling would dramatically slow down the user's interaction with their VLM.", + "bbox": [ + 88, + 136, + 482, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to present a unified approach for tackling context and decoding latency in modern VLMs. Latency is a common challenge associated with VLMs due to the sheer amount of visual tokens to be processed. Existing methods for accelerating VLM inference primarily focus on token-wise pruning or merging techniques, with recent approaches leveraging textual priors to reduce visual token complexity in a query-aware manner. In practice, methods that permanently remove visual tokens during the context stage are quite lossy in multi-turn evaluations as visualized in Figure 1. Hence, in this paper, we introduce SparseVILA as a novel approach for accelerating VLM inference, while retaining multi-turn performance.", + "bbox": [ + 89, + 316, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key insight is a decoupled sparsity framework enabling SparseVILA to migrate sparsity from the context into the decoding stage. Further, SparseVILA leverages query-aware retrieval in the decoding stage, supporting multi-turn conversation as a different subset of context tokens can be retrieved per question. This decoupled approach allows SparseVILA to achieve significant performance improvements in image-centric benchmarks as shown in Figure 1 and outperforms previous methods in long-context/generation scaling.", + "bbox": [ + 89, + 513, + 482, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries and Related Work", + "text_level": 1, + "bbox": [ + 89, + 662, + 385, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Token pruning has proven effective in accelerating inference across a variety of tasks, including image classification [3, 17, 19] and semantic segmentation [5, 43]. With the rise of generative AI, these techniques have been further extended to diffusion models, large language models [18], and vision-language models [14, 55, 60]. We refer the readers to Section A.2 for a more detailed survey of related work.", + "bbox": [ + 89, + 688, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper focuses on token pruning for vision-language models (VLMs). The key idea is that not all visual tokens contribute equally to VLM's final prediction. By identifying and removing less informative tokens, it is possible to significantly reduce the computational cost of VLM inference and thereby improve efficiency. Existing work in this area largely differs in how these tokens are selected. We categorize prior", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods into two groups based on their dependence on the input query: (i) query-agnostic approaches, which identify unimportant tokens based solely on visual saliency or redundancy, and (ii) query-aware approaches, which incorporate the semantic relationship between visual and textual inputs to guide pruning. For each category, we highlight representative methods and their limitations, laying the groundwork for the approach proposed in this paper.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Query-Agnostic Sparsity", + "text_level": 1, + "bbox": [ + 511, + 223, + 741, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query-agnostic token pruning methods aim to reduce redundancy or select important visual tokens without relying on the textual input [49]. They prune tokens based solely on the visual context—either within or after the vision encoder. For example, PruMerge [39] clusters and discards less informative tokens using final-layer attention scores, while VisionZip [55] employs a token merging module, similar to ToMe [3], to compress redundant visual information.", + "bbox": [ + 511, + 244, + 906, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This visual-only focus, however, presents key limitations. First, these methods often sacrifice fine-grained visual details, especially under high sparsity, which can degrade performance. More importantly, they cannot adapt token selection based on the input query, leading to suboptimal results when task-relevant information is sparsely distributed. As shown in Figure 1, query-agnostic methods underperform on text-centric and structured visual understanding benchmarks such as DocVQA [35], InfoVQA [36], and ChartQA [34]. By treating all visual tokens uniformly, they risk discarding critical information necessary for accurate reasoning.", + "bbox": [ + 511, + 367, + 908, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Query-Aware Sparsity", + "text_level": 1, + "bbox": [ + 511, + 544, + 720, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Query-aware pruning improves visual token selection by explicitly modeling the relationship between textual queries and visual representations. For example, FastV [4] leverages attention maps from early LLM layers as a salience metric to guide token pruning during the prefetch stage. This contrasts with query-agnostic approaches, which perform token reduction prior to any LLM processing. Similarly, SparseVLM [60] utilizes query-to-vision attention scores to eliminate less relevant visual tokens during prefetch.", + "bbox": [ + 511, + 566, + 908, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While query-aware methods are effective at aligning token selection with the current query, they face limitations in multi-turn interactions. Pruning based solely on the initial query can eliminate tokens that are crucial for future queries, resulting in degraded performance. Empirically, such methods show a huge drop in accuracy across multi-turn conversations, often underperforming compared to query-agnostic methods (Figure 1). Re-prefilling at each turn can maintain accuracy, but it introduces significant latency (Section 4.4).", + "bbox": [ + 511, + 703, + 908, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To systematically evaluate the performance ceiling of query-aware pruning in multi-turn contexts, we introduce a query-aware oracle, which selects the optimal subset of visual tokens to maximize agreement with the original model's", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd42b7d99850455a33b4ce37211e8e9c187a67c13fcf2b81088a9d724017d339.jpg", + "image_caption": [ + "Figure 2. Multi-Round Query-Aware Comparisons with LLaVA-1.5 [27] on the POPE [21] dataset. Without re-profilling the context, the query-aware oracle degrades heavily indicating the inability of query-aware pruning to scale effectively in multi-turn conversation." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "response (see Section A.3 for more details). From Figure 2, pruning decisions made in early rounds often result in substantial degradation in later responses, underscoring the long-term cost of early token elimination. By incorporating both the current query and the ground-truth response, the oracle establishes an upper bound on the achievable performance of any query-aware pruning method. These findings reveal a fundamental limitation of query-aware pruning in sustaining performance over multi-turn dialogue.", + "bbox": [ + 88, + 330, + 485, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SparseVILA: Best of Both Worlds", + "text_level": 1, + "bbox": [ + 89, + 481, + 401, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both query-agnostic and query-aware approaches offer distinct advantages and face inherent limitations. Existing visual token pruning methods typically perform pruning exclusively during the prefetch stage, with the benefits of reduced computation carrying over naturally to the subsequent decode stage. In this paper, we introduce SparseVILA that decouples visual sparsity between the prefetch and decode stages. During the prefetch stage, we apply query-agnostic pruning to eliminate redundant tokens that are unlikely to be useful for any downstream queries. In the decode stage, we employ a query-aware retrieval mechanism to selectively fetch the most relevant entries from the KV cache based on the current query, ensuring that only tokens directly related to the question are used for generating response. This design combines the strengths of both paradigms, achieving efficient token reduction while preserving query relevance.", + "bbox": [ + 91, + 507, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Prefix Fill Phase: Query-Agnostic Pruning", + "text_level": 1, + "bbox": [ + 89, + 757, + 426, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the prefetch phase, the VLM processes the system, image, and potentially the first query tokens. To support multi-turn settings, pruning in this stage remains query-agnostic, targeting visual redundancy or salience rather than textual correlation. Since the visual content is only prefilled once, token reduction remains independent of specific queries to effectively support multi-turn conversations.", + "bbox": [ + 89, + 780, + 483, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We estimate token importance using attention scores", + "bbox": [ + 109, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from the visual encoder, providing an unbiased (query-independent) assessment of token salience, as shown in previous works [1, 39, 55]. This approach is detailed in Algorithm 2 under Section A.4 of the supplementary material. By default, we take the mean-column sum of the attention weights, making our method compatible with architectures like SigLIP [58] that don't produce a $cls$ token. For CLIP-style [37] visual encoders that do generate a $cls$ token, we compute the mean-column-sum relative to that token, leveraging the encoder's training that routes high information density to this specific token.", + "bbox": [ + 511, + 90, + 908, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Decode Phase: Query-Aware Retrieval", + "text_level": 1, + "bbox": [ + 511, + 266, + 846, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the generation phase, the VLM is typically memory-bound as it implements next-token prediction by computing attention with the existing pre-filled KV cache. By reducing the effective size of the KV cache we can accelerate this generation process while maintaining the necessary visual information to support multi-turn conversation.", + "bbox": [ + 511, + 287, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We identify which visual tokens from the KV cache can be contextually selected for all associated generation steps in answering the current query. During query prefilling at each conversation round, we calculate the mean-columnsum of the attention weights between the query and the remaining visual tokens from the context stage. Using a sparsity threshold, we select a compact subset of the KV cache to participate in all generation steps associated with the query, thereby achieving significant speedup in decoding attention. We detail the retrieval strategy in Algorithm 3 under Section A.4 of the supplementary materials.", + "bbox": [ + 511, + 378, + 908, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While prefilling the current round query, we run Algorithm 3 simultaneously to compute the indices of the KV Cache for generation, immediately after each layer's self-attention block. To avoid introducing complexity and overhead into the decoding stage, we create a smaller compact dense KV cache representing only the selected indices for the current generation round. We further discuss the efficiency benefits of this in Section 3.4.", + "bbox": [ + 511, + 545, + 908, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One interesting finding is that the selected tokens can be further divided as Visual Attention Sinks and Visual Retrieval Tokens. The sink tokens appear consistent over different queries, meanwhile, retrieval tokens are directly related to the given query. We discuss this further with qualitative examples in Section 4.5 and Section A.4 in our supplementary material. Similar findings on visual token sinks have been echoed in previous works, such as VisionZip [55] and Visual Attention Redistribution (VAR) [16].", + "bbox": [ + 511, + 666, + 908, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussions. Beyond VLM pruning approaches, KV cache compression has been heavily explored in the context of LLMs. In particular, methods such as SnapKV [22] and Quest [42] leverage salience metrics to select a smaller KV cache for decoding, thus similarly accelerating the subsequent decoding stage. In the case of SnapKV, after selecting", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23786", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/de8c37865d095ef0a18295c2212995323b6dd8cff930e39b771f9eee0a2b81e0.jpg", + "image_caption": [ + "Figure 3. Overview of the components in SparseVILA. Query agnostic pruning is run in the context stage using the salience map from the visual encoder to guide token selection. In the decoding stage, the question is used to retrieve a sparse set of critical visual tokens to participate in the generation stage, hence achieving decoding speedup." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 911, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/930cb03e8ed956c583846aa16aacf704f3c80be744de50ea9a8e4e9a3e3b28f8.jpg", + "image_caption": [ + "Figure 4. Comparing sparsity migration on VideoMME at a fixed latency speedup of $2.2 \\times$ . SparseVILA* uses a higher context sparsity, SparseVILA uses our default configuration." + ], + "image_footnote": [], + "bbox": [ + 91, + 391, + 486, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens relevant to the observation window, this method discards the original KV cache, making it ill-suited to multi-turn conversation. Meanwhile, Quest re-runs its selection mechanism for every generated token, significantly increasing the complexity of the decoding stage, while only achieving speedup in long-context cases. We provide detailed comparisons with each of these methods in Section A.4.3 of the supplementary material. SparseVILA introduces a novel combination of query-aware and query-agnostic sparsity, leveraging context and decoding stage improvement beyond any decoding-only or context-only method; hence, the novelty remains in the design of a unified framework for efficient VLM inference.", + "bbox": [ + 88, + 561, + 486, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Decoupled Prefix-Decode Visual Sparsity", + "text_level": 1, + "bbox": [ + 89, + 771, + 441, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SparseVILA's decoupled framework enables the migration of computational reduction from the context stage to the generation stage. This approach is based on two key insights. Firstly, the context stage should preserve enough information to support multi-turn conversation as it is only pre-filled once per set of queries. Secondly, the generation stage can leverage query-aware information to guide the pruning process", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and accelerate the decoding, without affecting subsequent turns of conversation.", + "bbox": [ + 511, + 395, + 906, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 1, migrating visual sparsity from prefetch stage into the decode stage enables SparseVILA to outperform previous pruning approaches while achieving better speedup. We further investigate this effect by comparing two variants of SparseVILA that both achieve a $2.2 \\times$ speedup on the VideoMME benchmark [9], as shown in Figure 4. By migrating sparsity from the context stage into the decoding stage (i.e., SparseVILA* → SparseVILA), we observe a $2.3\\%$ improvement in overall performance, surpassing the baseline implementation. Moreover, this shift allows the model to retain more tokens during the context stage, resulting in a significant $5.9\\%$ gain on image-centric benchmarks.", + "bbox": [ + 511, + 426, + 908, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Efficient Implementation", + "text_level": 1, + "bbox": [ + 511, + 635, + 743, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SparseVILA introduces a two-stage compression approach to distribute sparsity over both the context and decoding stages. This unique approach allows SparseVILA to deliver strong accuracy performance with measurable speedup. Our method, leverages a carefully designed implementation to support modern acceleration techniques. In particular, all parts of the language model leverage the latest FlashAttention2 [6] implementation.", + "bbox": [ + 511, + 657, + 908, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable KV-Cache retrieval, we dynamically recompute the QK attention score map once per question at each layer and compress it into a compact, dense KV cache. This design allows us to fully utilize efficient attention implementations across the network. While a naive approach to computing the mean column sum of the attention weights can introduce latency and increase peak memory usage – especially at longer context lengths – we address this with a custom Triton", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3619d9be254e7d1541af6629e0e877558ca87a8d1f5f7b2bcdb0852c46e49f68.jpg", + "image_caption": [ + "Figure 5. Illustrates the lightweight recomputation of SparseVILA alongside layer-wise FlashAttention." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "kernel that streams attention weight normalization efficiently. This kernel achieves exact results without materializing the full $S \\times S$ attention matrix, significantly improving both speed and memory efficiency. Furthermore, due to the high sparsity achieved during the generation phase, our compact KV cache incurs minimal memory overhead (up to 55MB), while reducing peak memory consumption by $39.3\\%$ on long-context video tasks with Qwen2-7B [54].", + "bbox": [ + 88, + 321, + 482, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building a compact KV cache avoids any non-contiguous \"on-the-fly indexing\" (sparse access patterns) during the generation stage. The overhead from this additional computation is aggregated with the packing of the KV Cache into a contiguous set as shown in Figure 5, accounting for less than $4\\%$ of the generation runtime for 20, and 100 tokens on Llama2 [45], and Llama3 [7]/Qwen2 [54], respectively.", + "bbox": [ + 88, + 443, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 561, + 223, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 585, + 171, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Turn Evaluation. Many existing benchmarks generate multiple question-answer pairs for each image: e.g., up to 18 in POPE [21]. However, evaluation protocols often remain confined to single-turn settings. In this work, we make use of the inherent multi-turn structure of such datasets to enable more efficient and accurate assessment of VLMs. We will release our multi-turn evaluation framework to support the broader multimodal research community.", + "bbox": [ + 89, + 608, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We organize questions associated with the same image into coherent multi-turn conversations, which allows visual tokens to be prefilled only once. This setup not only reduces computational overhead but also better reflects realistic VLM usage in interactive settings. However, a potential problem is information leakage between turns, where earlier questions can unintentionally reveal answers to later ones. For example, in the GQA dataset [15]:", + "bbox": [ + 89, + 729, + 482, + 849 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Q1: \"What is the person in front of the sky doing?\"", + "Q2: \"What's the person in front of?\"" + ], + "bbox": [ + 109, + 852, + 464, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, Q1 discloses information that effectively answers Q2,", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "enabling the model to respond correctly even without relying on the image. To mitigate this, we implement a partial KV cache eviction strategy: after each round, we remove only the KV entries corresponding to the previous question and answer. This preserves the efficiency gains of visual KV cache reuse while preventing unintended context carryover between turns. We provide more details in Section A.5.3.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparsity Ratio. Our sparsity configuration adopts a straightforward approach for both context and decoding, ensuring efficient implementation and cross-model support. Specifically, we set a constant context sparsity before the LLM, with a uniform decoding sparsity across all layers. More granular strategies, such as layer-wise or head-aware sparsity, may further optimize performance but introduce additional complexity and tuning. We prioritize simplicity and generalization, leaving these refinements for future work.", + "bbox": [ + 511, + 205, + 906, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results on Image Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 349, + 782, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 1, we evaluate our SparseVILA against a diverse set of query-aware and query-agnostic baselines across eight vision-language benchmarks: ChartQA [34], DocVQA [35], GQA [15], InfographicVQA [36], MME [8], POPE [21], ScienceQA [31], and TextVQA [40]. All results in this section are conducted on LLaVA-NeXT-7B [61]. To further evaluate the generality of SparseVILA, we extend our analysis to additional VLMs, including LLaVA-1.5 [27], VILA-1.5 [24], Qwen2-VL [46], and NVILA [30], in Table A5.", + "bbox": [ + 511, + 371, + 906, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our SparseVILA with existing query-aware pruning methods, including FastV [4], SparseVLM [60] and PDrop [52]. Our results show that at a given speedup, SparseVILA outperforms these prior approaches in multi-turn evaluation by migrating sparsity into the decode phase. Notably, SparseVILA achieves a $6.4\\%$ accuracy improvement on DocVQA, highlighting the effectiveness of our decoupled prefetch-decode sparsity approach.", + "bbox": [ + 511, + 507, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further evaluate SparseVILA against existing query-agnostic methods, including PruMerge [39], HIRED [1], and VisionZip [55]. At an end-to-end speedup of $1.44 \\times$ , SparseVILA achieves at least a $3.2\\%$ improvement on DocVQA. These results further demonstrate that, under a fixed speedup target, SparseVILA's decoupled sparsity approach consistently outperforms prior state-of-the-art pruning methods. In Section A.6.1.1, we also show that applying SparseVILA to these existing query-agnostic sparsity methods can further enhance their performance.", + "bbox": [ + 511, + 628, + 908, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Results on Video Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 787, + 779, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 2, we demonstrate that SparseVILA consistently outperforms existing approaches on long-context tasks. Additionally, query-aware works such as FastV [4] and SparseVLM [60] were unable to scale effectively to long context (64K) as they require access to the full visual context and query at the same time to compute their ranking scores, hence", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/69161a9615369ec0bd56886c5cd6ed72b392eb0d20fd6a935950beafd45c5658.jpg", + "table_caption": [ + "Table 1. Results on image benchmarks. We provide query-aware comparisons with SparseVLM [60] and FastV [4] and query-agnostic comparisons with LLaVA-PruMerge [39], HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.2. At a given latency, we show significant improvements over state-of-the-art approaches; hence, SparseVILA optimizes the latency-to-performance tradeoff. A generation length of 250 tokens is used to approximate the sequence length observed in image-c captioning tasks. TextVQA [40], DocVQA (validation set) [35], InfoVQA (validation set) [36], ChartQA [34], ScienceQA-Img [31], POPE [21], and GQA [15]." + ], + "table_footnote": [], + "table_body": "
Sparsity (%)SpeedupCQADVQAGQAIVQAMMEPOPESQATVQA
PrefillDecodePrefillDecodeOverall
LLaVA-NeXT-7B-----50.774.464.237.71846.185.970.361.2
+ FastV8303.26×1.24×1.37×37.354.460.631.81775.282.867.557.1
+ SparseVLM8303.26×1.24×1.37×37.754.460.831.41780.784.067.558.1
+ PDrop5301.89×1.16×1.24×41.544.862.926.51809.286.169.157.0
+ SparseVILA67502.59×1.27×1.38×41.860.863.433.31818.485.968.060.2
+ PruMerge8304.34×1.27×1.43×26.936.961.128.51753.283.666.754.6
+ HIRED8304.34×1.27×1.43×29.741.662.428.21777.181.767.755.9
+ VisionZip8304.34×1.27×1.43×34.151.261.230.61760.383.767.458.7
+ SparseVILA67902.59×1.33×1.44×39.954.463.430.81827.886.168.058.9
", + "bbox": [ + 91, + 183, + 903, + 371 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/45ca9b4d331c7dafc79599ef5b834929c6a697ef0cf9631ceaaf4020560af15d.jpg", + "table_caption": [ + "Table 2. Results on video benchmarks. We provide query-agnostic comparisons with HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.3. $^{\\dagger}$ FastV [4], SparseVLM [60], and PDrop [52] do not scale to 64K context length, hence resulting in $OOM -$ we include shorter context comparison in Section A.6.2 of our supplementary material. A generation length of 2000 tokens is used to approximate the sequence length observed in video captioning or reasoning tasks." + ], + "table_footnote": [], + "table_body": "
Sparsity (%)SpeedupMLVUVideo-MME (w/o subtitle)
PrefillDecodePrefillDecodeOverallm-avgShortMediumLongOverall
LongVILA-7B-----72.468.958.152.059.7
+ HIRED8604.9×1.5×2.1×70.268.054.649.657.4
+ VisionZip8604.9×1.5×2.1×73.166.655.150.957.5
+ SparseVILA75943.6×1.7×2.2×74.470.658.651.760.3
", + "bbox": [ + 135, + 450, + 857, + 563 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bfb0c836d8b084c6ea15a06d487208446dbf1817bedcbc9c7df0f3023949a573.jpg", + "table_caption": [ + "Table 3. Long-Context/Generation Performance on VideoChatGPT [33]. Scores are computed using the GPT4-mini assisted evaluation. SparseVILA offers lossless generation performance at a $2.2 \\times$ end-to-end speedup." + ], + "table_footnote": [], + "table_body": "
CIDOCUTUCMeanSpeedup
LongVILA-7B2.332.282.871.742.412.331.0×
+ SparseVILA2.362.342.921.882.562.412.2×
", + "bbox": [ + 91, + 652, + 485, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "preventing chunk pre-filling and adding significant memory overhead to their approach. This consequently limits their ability to scale. In contrast, SparseVILA efficiently scales to 64K context (LongVILA 256 Frames) by decoupling our sparsity in the context and decoding stage, achieving $3.6 \\times$ prefilling speedup, $1.7 \\times$ decoding speedup, and $2.2 \\times$ end-to-end speedup. We include additional results on 8K context in Section A.6.2.", + "bbox": [ + 89, + 746, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One of the key novelties in SparseVILA is the decoupling and migration of sparsity into the decoding stage. As", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "previously shown, this has a positive effect on image and long-context benchmarks. In this section, we want to validate that SparseVILA's heavy generation sparsity does not affect the model's ability to generate responses by testing on free-form text generation tasks. Our approach delivers significant efficiency gains, offering up to $2.2 \\times$ E2E speedup over the vanilla model while maintaining performance on the Video-ChatGPT [33] benchmark as shown in Table 3.", + "bbox": [ + 511, + 587, + 906, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to VQA and captioning tasks, we also evaluate the long-context retrieval performance with Visual Needle in a Haystack (V-NIAH) benchmark [53, 59]. The original V-NIAH only supports single-turn, so we extended this benchmark to a multi-turn setting. See Section A.7 of the supplementary material for more details. We compare the long-context retrieval capacity of our method with the vanilla implementation, SparseVLM [60], and FastV [4] on LongVILA [53] in Figure 6. Notably, SparseVLM and FastV do not scale beyond 32 Frames (approx. 8K context lengths) as they lack support for chunkedprefilling and incur a memory overhead in computing their salience metric. Sparse-", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23789", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc2bd110ef826cf0e7017e42b29f665117d37643ad70ad228b9f0336a952876b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b80bb9a085e40c7ea250e2b9187ef4b334bfac1f627160d17862154e35216d05.jpg", + "image_caption": [ + "Figure 6. Comparing the multi-turn performance of the vanilla implementation with SparseVLM [60], FastV [4] and SparseVILA on LongVILA-Qwen2 [53] for V-NIAH. FastV and SparseVLM reach OOM after 30 Frames (8K context) as they do not inherently support chunk prefilling. We include the full scaling up to 500 Frames (128K context length) in Section A.7.", + "Figure 7. Decoding attention kernel speed for video tasks and V-NIAH on LongVILA-7B [53] and image tasks on LLaVA-NeXT-7B [28]. We further explore the multi-turn scaling of the baseline, a query-aware approach with re-pre-filling and SparseVILA over conversation rounds 0, 1, and 3." + ], + "image_footnote": [], + "bbox": [ + 91, + 234, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VILA maintains comparable performance with the baseline. In contrast, query-aware context stage pruning strategies begin to degrade at only an 8K context length. In reference, to Figure 6, the vanilla implementation remains effective up to a context length of 89.6K, handling contexts up to $10 \\times$ as long. Meanwhile, SparseVLM [60] and FastV [4] exhibit degradation, indicating that these pruning frameworks significantly impair the model's capacity for long context retrieval in multi-turn evaluation. We find that our approach performs similarly to the vanilla baseline, indicating that SparseVILA can retain long context retrieval performance. We can attribute the retained performance to the query-aware nature of our decoding sparsity. We leverage a $90\\%$ decoding sparsity, translating to a $2 \\times$ decoding speedup. Additional details are provided in Section A.7 of the supplementary material.", + "bbox": [ + 88, + 593, + 485, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Efficiency Analysis", + "text_level": 1, + "bbox": [ + 89, + 832, + 272, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide details of our optimized inference pipeline and additional results. Efficiency metrics are reported relative to the optimized baseline, adding additional", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "benefit for efficient VLM inference.", + "bbox": [ + 511, + 594, + 751, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference Setting AWQ [25] is a widely adopted technique for reducing decoding latency by significantly decreasing the latency of linear layers. We developed an optimized inference pipeline leveraging AWQ [25] based on TinyChat. In the case of Llama2-7B [45] with a 32K context length, our dense configuration achieves a $1.55\\mathrm{x}$ speedup over a finely-tuned FP16 pipeline, with pre-allocated KV cache across the entire benchmark sequence, effectively eliminating the additional overhead associated with dynamic memory allocation. All latency measurements are obtained on an NVIDIA A6000 GPU with a batch size of 1.", + "bbox": [ + 511, + 618, + 908, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General Efficiency Savings According to Figure 7, we can gain up to $20 \\times$ and $6.3 \\times$ speedup in decoding attention kernel time on long-context video and V-NIAH with LongVILA [53]. Further, our multi-turn scaling exhibits compounded improvement, achieving a 1:1.7 ratio in conversation turn speed between the vanilla and SparseVILA. Beyond latency measurements, SparseVILA can cut down", + "bbox": [ + 511, + 794, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23790", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8ff59445fda0cb6fe4361fe8174c8cfb0453b95cc6b66081db64b9a4a879ca6.jpg", + "image_caption": [ + "\\\\n" + ], + "image_footnote": [ + "Prompt 1: Tell me about the car in the image." + ], + "bbox": [ + 94, + 107, + 241, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9972c6c1e7b6405fbde4fc285a7fde35cbfecf1724bf2c092a0c4172538318d8.jpg", + "image_caption": [ + "Context Pruning" + ], + "image_footnote": [ + "SparseVILA: \"The car in the image is a blue sports car, possibly a Mazda RX-8, driving down the road." + ], + "bbox": [ + 246, + 107, + 395, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e04d8c252e20bef311bbf75e04b7226c9de79c0ceaf96586a8c3cac1adc79829.jpg", + "image_caption": [ + "Visual Attention Sinks", + "Figure 8. Visualization of selected tokens from the KV cache for various queries in the multi-round setting. The heatmap illustrates the frequency of token selection from the KV Cache for the particular query of interest. We conduct said visualization on LLaVA-1.5 [27]" + ], + "image_footnote": [ + "Prompt 2: Tell me about the deer crossing the road" + ], + "bbox": [ + 400, + 106, + 547, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98c551e5cf89a5b3e0f7f4843c13cb45699456a2804a3f8858379569dfed9e0a.jpg", + "image_caption": [ + "+", + "Prompt 1 Retrieval" + ], + "image_footnote": [ + "SparseVILA: \"In the image, a deer is crossing a road in front of a car. The deer is walking across the street, and the car is waiting patiently behind it ... this moment highlights the importance of being cautious and aware of wildlife while driving...\"" + ], + "bbox": [ + 583, + 104, + 732, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/db1d650e701f4cf46fb7fc5aaa1582f2400b767e45e35cd1a6f84d8d1af11685.jpg", + "image_caption": [ + "Prompt 2 Retrieval" + ], + "image_footnote": [], + "bbox": [ + 750, + 104, + 900, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the peak memory cost by $39.3\\%$ for Qwen2-7B [54] on long context video tasks. Finally, we can decrease linear flops in the context stage by up to $87.6\\%$ .", + "bbox": [ + 89, + 353, + 482, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Round Efficiency In Figure 7 we plot the multiround scaling for the baseline, SparseVLM [60] with context reprefilling, and SparseVILA. Over multiple rounds, our method retains a consistent context latency enabled through our context sparsity approach. Further, the generation latency scales faster than both query-aware and vanilla implementations, offering larger savings at increased rounds. Notably, SparseVLM [60] will scale past the original latency cost of the vanilla baseline after round 3 on Llama2 [45] indicating the inefficiency of that approach for multi-turn settings.", + "bbox": [ + 89, + 438, + 483, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 616, + 282, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide a qualitative visualization of SparseVILA's token selection strategy in Figure 8 and Figure A5. We apply a $50\\%$ context and $75\\%$ generation sparsity for this visualization on LLaVA-1.5 [27]. During the decoding stage, tokens are selected based on their alignment with query information by thresholding attention weights. Since different tokens can be selected at each layer, the heatmap visualizes the frequency by which a token was selected over all LLM layers. Certain tokens were selected across most layers, and consistent over different queries, we coin these, visual sink tokens aligning with previous definitions [51]. The remainder of tokens selected are considered query-dependent, which we term as retrieval tokens. This division is quite intuitive in the qualitative samples shown in Figure A5. Our decoupled framework retains both sink and retrieval tokens in the KV cache. Further analysis and visualizations are provided in Section A.8 of the supplementary material.", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 352, + 633, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SparseVILA, a novel query-aware token retrieval method designed to optimize inference in Vision-Language Models (VLMs). Through a decoupled sparsity framework SparseVILA leverages query-agnostic context stage pruning with query-aware generation stage retrieval. By migrating sparsity into the decoding stage, SparseVILA is able to achieve state-of-the-art performance on multi-turn benchmarks, with a significant $+5.9\\%$ improvement margin on image-centric tasks. By focusing on both context and decoding stage optimization, SparseVILA can achieve $3.6\\times$ context stage speedup, with $1.7\\times$ decoding stage acceleration in long-context video tasks. Further SparseVILA is able to retain lossless performance in both long-context/generation tasks and needle-in-a-haystack retrieval. SparseVILA introduces a new paradigm in prefetch decoding sparsity decoupling that maintains performance and improves efficiency. This approach enables VLMs to efficiently handle longer contexts, accommodating hundreds of images and frames.", + "bbox": [ + 511, + 377, + 908, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 666, + 609, + 681 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments, 2024. 3, 5, 6, 1, 4, 7", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 1", + "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang" + ], + "bbox": [ + 521, + 691, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Context Stage Sparsity", + "bbox": [ + 192, + 88, + 294, + 98 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Decoding Stage Sparsity", + "bbox": [ + 602, + 89, + 710, + 99 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models, 2024. 2, 5, 6, 7, 1, 4", + "[5] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2061-2070, 2023. 2", + "[6] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 4, 2", + "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5", + "[8] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5", + "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 4", + "[10] Chaoyou Fu, Yi-Fan Zhang, Shukang Yin, Bo Li, Xinyu Fang, Sirui Zhao, Haodong Duan, Xing Sun, Ziwei Liu, Liang Wang, et al. Mme-survey: A comprehensive survey on evaluation of multimodal llms. arXiv preprint arXiv:2411.15296, 2024. 4", + "[11] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. Lazyllm: Dynamic token pruning for efficient long context llm inference, 2024. 1", + "[12] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 773-783, 2023. 1", + "[13] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun Sophia Shao, Kurt Keutzer, and Amir Gholami. Kvquant: Towards 10 million context length llm inference with kv cache quantization. arXiv preprint arXiv:2401.18079, 2024. 1", + "[14] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2025. 2", + "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 5, 6, 4", + "[16] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models, 2025. 3", + "[17] Samir Khaki and Konstantinos N Plataniotis. The need for speed: Pruning transformers with one recipe. In *The Twelfth" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Learning Representations, 2024. 2, 1", + "[18] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers, 2022. 2", + "[19] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, et al. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 2, 1", + "[20] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 6", + "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 3, 5, 6, 4", + "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation, 2024. 3, 1", + "[23] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 1", + "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 1, 5, 7", + "[25] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for lIm compression and acceleration, 2024. 7", + "[26] Yujun Lin, Haotian Tang, Shang Yang, Zhekai Zhang, Guangxuan Xiao, Chuang Gan, and Song Han. Qserve: W4a8kv4 quantization and system co-design for efficient llm serving. arXiv preprint arXiv:2405.04532, 2024. 1", + "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3, 5, 8", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,7,4", + "[29] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750, 2024.1", + "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1, 5, 6, 7", + "[31] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23792", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 4", + "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1", + "[33] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 6, 4, 5", + "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 5, 6, 4", + "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6, 4", + "[36] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2, 5, 6, 4", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 2", + "[38] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1", + "[39] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 3, 5, 6, 1, 4", + "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 4", + "[41] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 1", + "[42] Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference, 2024. 3, 1", + "[43] Quan Tang, Bowen Zhang, Jiajun Liu, Fagui Liu, and Yifan Liu. Dynamic token pruning in plain vision transformers for semantic segmentation, 2023. 2", + "[44] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165-12174, 2022. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 7, 8, 3", + "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 7", + "[47] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution, 2024. 1", + "[48] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 1", + "[49] Zichen Wen, Yifeng Gao, Weijia Li, Conghui He, and Linfeng Zhang. Token pruning in multimodal large language models: Are we solving the right problem? arXiv preprint arXiv:2502.11501, 2025. 2", + "[50] Guangxuan Xiao, Jiaming Tang, Jingwei Zuo, Junxian Guo, Shang Yang, Haotian Tang, Yao Fu, and Song Han. Duoattention: Efficient long-context llm inference with retrieval and streaming heads, 2024. 1", + "[51] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks, 2024. 8, 1", + "[52] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 5, 6, 7", + "[53] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, Yihui He, Hongxu Yin, Pavlo Molchanov, Jan Kautz, Linxi Fan, Yuke Zhu, Yao Lu, and Song Han. Longvila: Scaling long-context visual language models for long videos. 2024. 6, 7, 4, 5", + "[54] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23793", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report, 2024. 5, 8", + "[55] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2, 3, 5, 6, 1, 4, 7", + "[56] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models, 2025. 1", + "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1", + "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 2", + "[59] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024.6, 4", + "[60] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference arXiv preprint arXiv:2410.04417, 2024. 1, 2, 5, 6, 7, 8, 3, 4", + "[61] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 5", + "[62] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, Zhangyang Wang, and Beidi Chen. $\\mathrm{H}_{20}$ : Heavy-hitter oracle for efficient generative inference of large language models, 2023. 1", + "[63] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.4", + "[64] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning. arXiv preprint arXiv:2104.08500, 2021. 1" + ], + "bbox": [ + 91, + 90, + 483, + 713 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23794", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_model.json b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f1fe77dae49f44b0a16c8addd26f6d7b158d8b8 --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.851, + 0.153 + ], + "angle": 0, + "content": "SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.18, + 0.248, + 0.199 + ], + "angle": 0, + "content": "Samir Khaki" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.202, + 0.41, + 0.218 + ], + "angle": 0, + "content": "Konstantinos N. Plataniotis4" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.225, + 0.331, + 0.242 + ], + "angle": 0, + "content": "NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.225, + 0.4, + 0.241 + ], + "angle": 0, + "content": "\\(^{2}\\)MIT" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.181, + 0.554, + 0.2 + ], + "angle": 0, + "content": "Jiaming Tang²" + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.202, + 0.518, + 0.218 + ], + "angle": 0, + "content": "Yao Lu" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.225, + 0.543, + 0.244 + ], + "angle": 0, + "content": "\\(^{3}\\)UC San Diego" + }, + { + "type": "text", + "bbox": [ + 0.565, + 0.181, + 0.7, + 0.2 + ], + "angle": 0, + "content": "Shang Yang2" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.202, + 0.657, + 0.22 + ], + "angle": 0, + "content": "Song Han1,2" + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.181, + 0.855, + 0.199 + ], + "angle": 0, + "content": "Yukang Chen" + }, + { + "type": "text", + "bbox": [ + 0.698, + 0.201, + 0.807, + 0.22 + ], + "angle": 0, + "content": "Zhijian Liu\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.563, + 0.225, + 0.744, + 0.244 + ], + "angle": 0, + "content": "\\(^{4}\\)University of Toronto" + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.279, + 0.298, + 0.289 + ], + "angle": 0, + "content": "Multi-Round Performance" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.291, + 0.361, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.279, + 0.566, + 0.289 + ], + "angle": 0, + "content": "Image-Centric Evaluation" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.291, + 0.63, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.279, + 0.88, + 0.289 + ], + "angle": 0, + "content": "Long-Context/Generation Performance" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.291, + 0.903, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.453, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 1. Overview of SparseVILA. Compared to existing approaches, SparseVILA exhibits advantages in the following aspects: 1. Multi-Round Performance. SparseVILA leverages a decoupled sparsity framework to achieve state-of-the-art performance at a \\(1.4 \\times\\) speedup on LLaVA-NeXT [28]. 2. Image-centric evaluation. Leveraging decoupled sparsity, SparseVILA can retain higher context tokens resulting in a \\(5.9\\%\\) margin over previous approaches. 3. SparseVILA scales to long-context and long-generation tasks, achieving lossless performance with up to a \\(3.6 \\times\\) context, \\(1.7 \\times\\) generation, and \\(2.2 \\times\\) end-to-end speedup." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.549, + 0.327, + 0.564 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM, by pruning tokens in the context stage while attending to a sparse subset of visual tokens during the generation phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a \\(1.4 \\times\\) speedup on image benchmarks. This approach leads to \\(+5.9\\%\\) average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a \\(3.6 \\times\\) and \\(1.7 \\times\\) speedup in prefetch and decoding, respectively." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.549, + 0.645, + 0.564 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.577, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Vision Language Models (VLMs) have emerged as a state-of-the-art conversational tool enabling users to directly interact with Large Language Models (LLMs) using various visual features, including photographs, documents, and videos [2, 24, 27, 30, 47]. Unfortunately, this added modality comes at the expense of higher latency and memory associated with processing the visual tokens in the LLM. Hence, deploying LLMs efficiently at inference time remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.701, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Several works have aimed to reduce these associated costs through model pruning on the LLM or vision encoder [32, 41, 64], KV cache compression [13, 26, 29], and most recently, token sparsification [60]. By reducing the amount of computation in the inference pipeline, many of these methods can achieve significant context stage savings as this stage of the network is mainly compute-bound." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Looking beyond the context, real-world applications often demand extensive generation. Tasks such as image captioning may require a few hundred tokens, meanwhile, video captioning or detailing easily requires more than a few thousand generated tokens. Hence in such applications, it is not sufficient to only focus on context stage optimization – effi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "cient implementations with real-world applications should focus on both context and decoding stage optimizations. One such example is multi-turn conversation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.483, + 0.317 + ], + "angle": 0, + "content": "Multi-turn/round conversation is a practical use case for VLMs, as it is entirely possible for a user to pose multiple questions about a given visual input. In fact, most benchmarks inherently support multi-round conversation: the GQA dataset [15] has more than 90 questions for the same visual input. Despite this, most evaluation benchmarks run single-round evaluation (i.e. repetitive pre-filling), which is not only unrealistic but also inefficient as the context stage would be repeated for each generation round. In real-world scenarios, the visual input could span tens of thousands of context tokens, hence repetitive pre-filling would dramatically slow down the user's interaction with their VLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.514 + ], + "angle": 0, + "content": "In this work, we aim to present a unified approach for tackling context and decoding latency in modern VLMs. Latency is a common challenge associated with VLMs due to the sheer amount of visual tokens to be processed. Existing methods for accelerating VLM inference primarily focus on token-wise pruning or merging techniques, with recent approaches leveraging textual priors to reduce visual token complexity in a query-aware manner. In practice, methods that permanently remove visual tokens during the context stage are quite lossy in multi-turn evaluations as visualized in Figure 1. Hence, in this paper, we introduce SparseVILA as a novel approach for accelerating VLM inference, while retaining multi-turn performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Our key insight is a decoupled sparsity framework enabling SparseVILA to migrate sparsity from the context into the decoding stage. Further, SparseVILA leverages query-aware retrieval in the decoding stage, supporting multi-turn conversation as a different subset of context tokens can be retrieved per question. This decoupled approach allows SparseVILA to achieve significant performance improvements in image-centric benchmarks as shown in Figure 1 and outperforms previous methods in long-context/generation scaling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.387, + 0.68 + ], + "angle": 0, + "content": "2. Preliminaries and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Token pruning has proven effective in accelerating inference across a variety of tasks, including image classification [3, 17, 19] and semantic segmentation [5, 43]. With the rise of generative AI, these techniques have been further extended to diffusion models, large language models [18], and vision-language models [14, 55, 60]. We refer the readers to Section A.2 for a more detailed survey of related work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This paper focuses on token pruning for vision-language models (VLMs). The key idea is that not all visual tokens contribute equally to VLM's final prediction. By identifying and removing less informative tokens, it is possible to significantly reduce the computational cost of VLM inference and thereby improve efficiency. Existing work in this area largely differs in how these tokens are selected. We categorize prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.213 + ], + "angle": 0, + "content": "methods into two groups based on their dependence on the input query: (i) query-agnostic approaches, which identify unimportant tokens based solely on visual saliency or redundancy, and (ii) query-aware approaches, which incorporate the semantic relationship between visual and textual inputs to guide pruning. For each category, we highlight representative methods and their limitations, laying the groundwork for the approach proposed in this paper." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.224, + 0.742, + 0.24 + ], + "angle": 0, + "content": "2.1. Query-Agnostic Sparsity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.367 + ], + "angle": 0, + "content": "Query-agnostic token pruning methods aim to reduce redundancy or select important visual tokens without relying on the textual input [49]. They prune tokens based solely on the visual context—either within or after the vision encoder. For example, PruMerge [39] clusters and discards less informative tokens using final-layer attention scores, while VisionZip [55] employs a token merging module, similar to ToMe [3], to compress redundant visual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.909, + 0.534 + ], + "angle": 0, + "content": "This visual-only focus, however, presents key limitations. First, these methods often sacrifice fine-grained visual details, especially under high sparsity, which can degrade performance. More importantly, they cannot adapt token selection based on the input query, leading to suboptimal results when task-relevant information is sparsely distributed. As shown in Figure 1, query-agnostic methods underperform on text-centric and structured visual understanding benchmarks such as DocVQA [35], InfoVQA [36], and ChartQA [34]. By treating all visual tokens uniformly, they risk discarding critical information necessary for accurate reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.722, + 0.561 + ], + "angle": 0, + "content": "2.2. Query-Aware Sparsity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.704 + ], + "angle": 0, + "content": "Query-aware pruning improves visual token selection by explicitly modeling the relationship between textual queries and visual representations. For example, FastV [4] leverages attention maps from early LLM layers as a salience metric to guide token pruning during the prefetch stage. This contrasts with query-agnostic approaches, which perform token reduction prior to any LLM processing. Similarly, SparseVLM [60] utilizes query-to-vision attention scores to eliminate less relevant visual tokens during prefetch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.84 + ], + "angle": 0, + "content": "While query-aware methods are effective at aligning token selection with the current query, they face limitations in multi-turn interactions. Pruning based solely on the initial query can eliminate tokens that are crucial for future queries, resulting in degraded performance. Empirically, such methods show a huge drop in accuracy across multi-turn conversations, often underperforming compared to query-agnostic methods (Figure 1). Re-prefilling at each turn can maintain accuracy, but it introduces significant latency (Section 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To systematically evaluate the performance ceiling of query-aware pruning in multi-turn contexts, we introduce a query-aware oracle, which selects the optimal subset of visual tokens to maximize agreement with the original model's" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23785" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.482, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.308 + ], + "angle": 0, + "content": "Figure 2. Multi-Round Query-Aware Comparisons with LLaVA-1.5 [27] on the POPE [21] dataset. Without re-profilling the context, the query-aware oracle degrades heavily indicating the inability of query-aware pruning to scale effectively in multi-turn conversation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.486, + 0.471 + ], + "angle": 0, + "content": "response (see Section A.3 for more details). From Figure 2, pruning decisions made in early rounds often result in substantial degradation in later responses, underscoring the long-term cost of early token elimination. By incorporating both the current query and the ground-truth response, the oracle establishes an upper bound on the achievable performance of any query-aware pruning method. These findings reveal a fundamental limitation of query-aware pruning in sustaining performance over multi-turn dialogue." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.482, + 0.402, + 0.5 + ], + "angle": 0, + "content": "3. SparseVILA: Best of Both Worlds" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.508, + 0.485, + 0.75 + ], + "angle": 0, + "content": "Both query-agnostic and query-aware approaches offer distinct advantages and face inherent limitations. Existing visual token pruning methods typically perform pruning exclusively during the prefetch stage, with the benefits of reduced computation carrying over naturally to the subsequent decode stage. In this paper, we introduce SparseVILA that decouples visual sparsity between the prefetch and decode stages. During the prefetch stage, we apply query-agnostic pruning to eliminate redundant tokens that are unlikely to be useful for any downstream queries. In the decode stage, we employ a query-aware retrieval mechanism to selectively fetch the most relevant entries from the KV cache based on the current query, ensuring that only tokens directly related to the question are used for generating response. This design combines the strengths of both paradigms, achieving efficient token reduction while preserving query relevance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.427, + 0.775 + ], + "angle": 0, + "content": "3.1. Prefix Fill Phase: Query-Agnostic Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.886 + ], + "angle": 0, + "content": "During the prefetch phase, the VLM processes the system, image, and potentially the first query tokens. To support multi-turn settings, pruning in this stage remains query-agnostic, targeting visual redundancy or salience rather than textual correlation. Since the visual content is only prefilled once, token reduction remains independent of specific queries to effectively support multi-turn conversations." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We estimate token importance using attention scores" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.259 + ], + "angle": 0, + "content": "from the visual encoder, providing an unbiased (query-independent) assessment of token salience, as shown in previous works [1, 39, 55]. This approach is detailed in Algorithm 2 under Section A.4 of the supplementary material. By default, we take the mean-column sum of the attention weights, making our method compatible with architectures like SigLIP [58] that don't produce a \\(cls\\) token. For CLIP-style [37] visual encoders that do generate a \\(cls\\) token, we compute the mean-column-sum relative to that token, leveraging the encoder's training that routes high information density to this specific token." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.848, + 0.284 + ], + "angle": 0, + "content": "3.2. Decode Phase: Query-Aware Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.908, + 0.379 + ], + "angle": 0, + "content": "During the generation phase, the VLM is typically memory-bound as it implements next-token prediction by computing attention with the existing pre-filled KV cache. By reducing the effective size of the KV cache we can accelerate this generation process while maintaining the necessary visual information to support multi-turn conversation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.909, + 0.546 + ], + "angle": 0, + "content": "We identify which visual tokens from the KV cache can be contextually selected for all associated generation steps in answering the current query. During query prefilling at each conversation round, we calculate the mean-columnsum of the attention weights between the query and the remaining visual tokens from the context stage. Using a sparsity threshold, we select a compact subset of the KV cache to participate in all generation steps associated with the query, thereby achieving significant speedup in decoding attention. We detail the retrieval strategy in Algorithm 3 under Section A.4 of the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.909, + 0.666 + ], + "angle": 0, + "content": "While prefilling the current round query, we run Algorithm 3 simultaneously to compute the indices of the KV Cache for generation, immediately after each layer's self-attention block. To avoid introducing complexity and overhead into the decoding stage, we create a smaller compact dense KV cache representing only the selected indices for the current generation round. We further discuss the efficiency benefits of this in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.803 + ], + "angle": 0, + "content": "One interesting finding is that the selected tokens can be further divided as Visual Attention Sinks and Visual Retrieval Tokens. The sink tokens appear consistent over different queries, meanwhile, retrieval tokens are directly related to the given query. We discuss this further with qualitative examples in Section 4.5 and Section A.4 in our supplementary material. Similar findings on visual token sinks have been echoed in previous works, such as VisionZip [55] and Visual Attention Redistribution (VAR) [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Discussions. Beyond VLM pruning approaches, KV cache compression has been heavily explored in the context of LLMs. In particular, methods such as SnapKV [22] and Quest [42] leverage salience metrics to select a smaller KV cache for decoding, thus similarly accelerating the subsequent decoding stage. In the case of SnapKV, after selecting" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23786" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.912, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.327, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 3. Overview of the components in SparseVILA. Query agnostic pruning is run in the context stage using the salience map from the visual encoder to guide token selection. In the decoding stage, the question is used to retrieve a sparse set of critical visual tokens to participate in the generation stage, hence achieving decoding speedup." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.392, + 0.487, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Figure 4. Comparing sparsity migration on VideoMME at a fixed latency speedup of \\(2.2 \\times\\). SparseVILA* uses a higher context sparsity, SparseVILA uses our default configuration." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.563, + 0.487, + 0.759 + ], + "angle": 0, + "content": "tokens relevant to the observation window, this method discards the original KV cache, making it ill-suited to multi-turn conversation. Meanwhile, Quest re-runs its selection mechanism for every generated token, significantly increasing the complexity of the decoding stage, while only achieving speedup in long-context cases. We provide detailed comparisons with each of these methods in Section A.4.3 of the supplementary material. SparseVILA introduces a novel combination of query-aware and query-agnostic sparsity, leveraging context and decoding stage improvement beyond any decoding-only or context-only method; hence, the novelty remains in the design of a unified framework for efficient VLM inference." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.772, + 0.442, + 0.789 + ], + "angle": 0, + "content": "3.3. Decoupled Prefix-Decode Visual Sparsity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "SparseVILA's decoupled framework enables the migration of computational reduction from the context stage to the generation stage. This approach is based on two key insights. Firstly, the context stage should preserve enough information to support multi-turn conversation as it is only pre-filled once per set of queries. Secondly, the generation stage can leverage query-aware information to guide the pruning process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.424 + ], + "angle": 0, + "content": "and accelerate the decoding, without affecting subsequent turns of conversation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.909, + 0.622 + ], + "angle": 0, + "content": "As shown in Figure 1, migrating visual sparsity from prefetch stage into the decode stage enables SparseVILA to outperform previous pruning approaches while achieving better speedup. We further investigate this effect by comparing two variants of SparseVILA that both achieve a \\(2.2 \\times\\) speedup on the VideoMME benchmark [9], as shown in Figure 4. By migrating sparsity from the context stage into the decoding stage (i.e., SparseVILA* → SparseVILA), we observe a \\(2.3\\%\\) improvement in overall performance, surpassing the baseline implementation. Moreover, this shift allows the model to retain more tokens during the context stage, resulting in a significant \\(5.9\\%\\) gain on image-centric benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.744, + 0.652 + ], + "angle": 0, + "content": "3.4. Efficient Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.779 + ], + "angle": 0, + "content": "SparseVILA introduces a two-stage compression approach to distribute sparsity over both the context and decoding stages. This unique approach allows SparseVILA to deliver strong accuracy performance with measurable speedup. Our method, leverages a carefully designed implementation to support modern acceleration techniques. In particular, all parts of the language model leverage the latest FlashAttention2 [6] implementation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To enable KV-Cache retrieval, we dynamically recompute the QK attention score map once per question at each layer and compress it into a compact, dense KV cache. This design allows us to fully utilize efficient attention implementations across the network. While a naive approach to computing the mean column sum of the attention weights can introduce latency and increase peak memory usage – especially at longer context lengths – we address this with a custom Triton" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23787" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.485, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.299 + ], + "angle": 0, + "content": "Figure 5. Illustrates the lightweight recomputation of SparseVILA alongside layer-wise FlashAttention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.483, + 0.444 + ], + "angle": 0, + "content": "kernel that streams attention weight normalization efficiently. This kernel achieves exact results without materializing the full \\( S \\times S \\) attention matrix, significantly improving both speed and memory efficiency. Furthermore, due to the high sparsity achieved during the generation phase, our compact KV cache incurs minimal memory overhead (up to 55MB), while reducing peak memory consumption by \\( 39.3\\% \\) on long-context video tasks with Qwen2-7B [54]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.483, + 0.55 + ], + "angle": 0, + "content": "Building a compact KV cache avoids any non-contiguous \"on-the-fly indexing\" (sparse access patterns) during the generation stage. The overhead from this additional computation is aggregated with the packing of the KV Cache into a contiguous set as shown in Figure 5, accounting for less than \\(4\\%\\) of the generation runtime for 20, and 100 tokens on Llama2 [45], and Llama3 [7]/Qwen2 [54], respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.224, + 0.58 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.173, + 0.604 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Multi-Turn Evaluation. Many existing benchmarks generate multiple question-answer pairs for each image: e.g., up to 18 in POPE [21]. However, evaluation protocols often remain confined to single-turn settings. In this work, we make use of the inherent multi-turn structure of such datasets to enable more efficient and accurate assessment of VLMs. We will release our multi-turn evaluation framework to support the broader multimodal research community." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.85 + ], + "angle": 0, + "content": "We organize questions associated with the same image into coherent multi-turn conversations, which allows visual tokens to be prefilled only once. This setup not only reduces computational overhead but also better reflects realistic VLM usage in interactive settings. However, a potential problem is information leakage between turns, where earlier questions can unintentionally reveal answers to later ones. For example, in the GQA dataset [15]:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.853, + 0.465, + 0.868 + ], + "angle": 0, + "content": "Q1: \"What is the person in front of the sky doing?\"" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.868, + 0.37, + 0.884 + ], + "angle": 0, + "content": "Q2: \"What's the person in front of?\"" + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.853, + 0.465, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here, Q1 discloses information that effectively answers Q2," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "enabling the model to respond correctly even without relying on the image. To mitigate this, we implement a partial KV cache eviction strategy: after each round, we remove only the KV entries corresponding to the previous question and answer. This preserves the efficiency gains of visual KV cache reuse while preventing unintended context carryover between turns. We provide more details in Section A.5.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Sparsity Ratio. Our sparsity configuration adopts a straightforward approach for both context and decoding, ensuring efficient implementation and cross-model support. Specifically, we set a constant context sparsity before the LLM, with a uniform decoding sparsity across all layers. More granular strategies, such as layer-wise or head-aware sparsity, may further optimize performance but introduce additional complexity and tuning. We prioritize simplicity and generalization, leaving these refinements for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.35, + 0.783, + 0.366 + ], + "angle": 0, + "content": "4.2. Results on Image Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.908, + 0.508 + ], + "angle": 0, + "content": "In Table 1, we evaluate our SparseVILA against a diverse set of query-aware and query-agnostic baselines across eight vision-language benchmarks: ChartQA [34], DocVQA [35], GQA [15], InfographicVQA [36], MME [8], POPE [21], ScienceQA [31], and TextVQA [40]. All results in this section are conducted on LLaVA-NeXT-7B [61]. To further evaluate the generality of SparseVILA, we extend our analysis to additional VLMs, including LLaVA-1.5 [27], VILA-1.5 [24], Qwen2-VL [46], and NVILA [30], in Table A5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.908, + 0.629 + ], + "angle": 0, + "content": "We compare our SparseVILA with existing query-aware pruning methods, including FastV [4], SparseVLM [60] and PDrop [52]. Our results show that at a given speedup, SparseVILA outperforms these prior approaches in multi-turn evaluation by migrating sparsity into the decode phase. Notably, SparseVILA achieves a \\(6.4\\%\\) accuracy improvement on DocVQA, highlighting the effectiveness of our decoupled prefetch-decode sparsity approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.78 + ], + "angle": 0, + "content": "We further evaluate SparseVILA against existing query-agnostic methods, including PruMerge [39], HIRED [1], and VisionZip [55]. At an end-to-end speedup of \\(1.44 \\times\\), SparseVILA achieves at least a \\(3.2\\%\\) improvement on DocVQA. These results further demonstrate that, under a fixed speedup target, SparseVILA's decoupled sparsity approach consistently outperforms prior state-of-the-art pruning methods. In Section A.6.1.1, we also show that applying SparseVILA to these existing query-agnostic sparsity methods can further enhance their performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.78, + 0.803 + ], + "angle": 0, + "content": "4.3. Results on Video Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.908, + 0.902 + ], + "angle": 0, + "content": "In Table 2, we demonstrate that SparseVILA consistently outperforms existing approaches on long-context tasks. Additionally, query-aware works such as FastV [4] and SparseVLM [60] were unable to scale effectively to long context (64K) as they require access to the full visual context and query at the same time to compute their ranking scores, hence" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23788" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.175 + ], + "angle": 0, + "content": "Table 1. Results on image benchmarks. We provide query-aware comparisons with SparseVLM [60] and FastV [4] and query-agnostic comparisons with LLaVA-PruMerge [39], HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.2. At a given latency, we show significant improvements over state-of-the-art approaches; hence, SparseVILA optimizes the latency-to-performance tradeoff. A generation length of 250 tokens is used to approximate the sequence length observed in image-c captioning tasks. TextVQA [40], DocVQA (validation set) [35], InfoVQA (validation set) [36], ChartQA [34], ScienceQA-Img [31], POPE [21], and GQA [15]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.184, + 0.905, + 0.372 + ], + "angle": 0, + "content": "
Sparsity (%)SpeedupCQADVQAGQAIVQAMMEPOPESQATVQA
PrefillDecodePrefillDecodeOverall
LLaVA-NeXT-7B-----50.774.464.237.71846.185.970.361.2
+ FastV8303.26×1.24×1.37×37.354.460.631.81775.282.867.557.1
+ SparseVLM8303.26×1.24×1.37×37.754.460.831.41780.784.067.558.1
+ PDrop5301.89×1.16×1.24×41.544.862.926.51809.286.169.157.0
+ SparseVILA67502.59×1.27×1.38×41.860.863.433.31818.485.968.060.2
+ PruMerge8304.34×1.27×1.43×26.936.961.128.51753.283.666.754.6
+ HIRED8304.34×1.27×1.43×29.741.662.428.21777.181.767.755.9
+ VisionZip8304.34×1.27×1.43×34.151.261.230.61760.383.767.458.7
+ SparseVILA67902.59×1.33×1.44×39.954.463.430.81827.886.168.058.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.384, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Table 2. Results on video benchmarks. We provide query-agnostic comparisons with HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.3.\\(^{\\dagger}\\) FastV [4], SparseVLM [60], and PDrop [52] do not scale to 64K context length, hence resulting in \\(OOM - \\) we include shorter context comparison in Section A.6.2 of our supplementary material. A generation length of 2000 tokens is used to approximate the sequence length observed in video captioning or reasoning tasks." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.451, + 0.859, + 0.564 + ], + "angle": 0, + "content": "
Sparsity (%)SpeedupMLVUVideo-MME (w/o subtitle)
PrefillDecodePrefillDecodeOverallm-avgShortMediumLongOverall
LongVILA-7B-----72.468.958.152.059.7
+ HIRED8604.9×1.5×2.1×70.268.054.649.657.4
+ VisionZip8604.9×1.5×2.1×73.166.655.150.957.5
+ SparseVILA75943.6×1.7×2.2×74.470.658.651.760.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.585, + 0.486, + 0.643 + ], + "angle": 0, + "content": "Table 3. Long-Context/Generation Performance on VideoChatGPT [33]. Scores are computed using the GPT4-mini assisted evaluation. SparseVILA offers lossless generation performance at a \\(2.2 \\times\\) end-to-end speedup." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.653, + 0.486, + 0.718 + ], + "angle": 0, + "content": "
CIDOCUTUCMeanSpeedup
LongVILA-7B2.332.282.871.742.412.331.0×
+ SparseVILA2.362.342.921.882.562.412.2×
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.868 + ], + "angle": 0, + "content": "preventing chunk pre-filling and adding significant memory overhead to their approach. This consequently limits their ability to scale. In contrast, SparseVILA efficiently scales to 64K context (LongVILA 256 Frames) by decoupling our sparsity in the context and decoding stage, achieving \\(3.6 \\times\\) prefilling speedup, \\(1.7 \\times\\) decoding speedup, and \\(2.2 \\times\\) end-to-end speedup. We include additional results on 8K context in Section A.6.2." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.486, + 0.903 + ], + "angle": 0, + "content": "One of the key novelties in SparseVILA is the decoupling and migration of sparsity into the decoding stage. As" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.907, + 0.709 + ], + "angle": 0, + "content": "previously shown, this has a positive effect on image and long-context benchmarks. In this section, we want to validate that SparseVILA's heavy generation sparsity does not affect the model's ability to generate responses by testing on free-form text generation tasks. Our approach delivers significant efficiency gains, offering up to \\(2.2 \\times\\) E2E speedup over the vanilla model while maintaining performance on the Video-ChatGPT [33] benchmark as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In addition to VQA and captioning tasks, we also evaluate the long-context retrieval performance with Visual Needle in a Haystack (V-NIAH) benchmark [53, 59]. The original V-NIAH only supports single-turn, so we extended this benchmark to a multi-turn setting. See Section A.7 of the supplementary material for more details. We compare the long-context retrieval capacity of our method with the vanilla implementation, SparseVLM [60], and FastV [4] on LongVILA [53] in Figure 6. Notably, SparseVLM and FastV do not scale beyond 32 Frames (approx. 8K context lengths) as they lack support for chunkedprefilling and incur a memory overhead in computing their salience metric. Sparse-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.912, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.192, + 0.908, + 0.234 + ], + "angle": 0, + "content": "Figure 6. Comparing the multi-turn performance of the vanilla implementation with SparseVLM [60], FastV [4] and SparseVILA on LongVILA-Qwen2 [53] for V-NIAH. FastV and SparseVLM reach OOM after 30 Frames (8K context) as they do not inherently support chunk prefilling. We include the full scaling up to 500 Frames (128K context length) in Section A.7." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.236, + 0.907, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 7. Decoding attention kernel speed for video tasks and V-NIAH on LongVILA-7B [53] and image tasks on LLaVA-NeXT-7B [28]. We further explore the multi-turn scaling of the baseline, a query-aware approach with re-pre-filling and SparseVILA over conversation rounds 0, 1, and 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.594, + 0.486, + 0.822 + ], + "angle": 0, + "content": "VILA maintains comparable performance with the baseline. In contrast, query-aware context stage pruning strategies begin to degrade at only an 8K context length. In reference, to Figure 6, the vanilla implementation remains effective up to a context length of 89.6K, handling contexts up to \\(10 \\times\\) as long. Meanwhile, SparseVLM [60] and FastV [4] exhibit degradation, indicating that these pruning frameworks significantly impair the model's capacity for long context retrieval in multi-turn evaluation. We find that our approach performs similarly to the vanilla baseline, indicating that SparseVILA can retain long context retrieval performance. We can attribute the retained performance to the query-aware nature of our decoding sparsity. We leverage a \\(90\\%\\) decoding sparsity, translating to a \\(2 \\times\\) decoding speedup. Additional details are provided in Section A.7 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.273, + 0.849 + ], + "angle": 0, + "content": "4.4. Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this section, we provide details of our optimized inference pipeline and additional results. Efficiency metrics are reported relative to the optimized baseline, adding additional" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.752, + 0.609 + ], + "angle": 0, + "content": "benefit for efficient VLM inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.909, + 0.785 + ], + "angle": 0, + "content": "Inference Setting AWQ [25] is a widely adopted technique for reducing decoding latency by significantly decreasing the latency of linear layers. We developed an optimized inference pipeline leveraging AWQ [25] based on TinyChat. In the case of Llama2-7B [45] with a 32K context length, our dense configuration achieves a \\(1.55\\mathrm{x}\\) speedup over a finely-tuned FP16 pipeline, with pre-allocated KV cache across the entire benchmark sequence, effectively eliminating the additional overhead associated with dynamic memory allocation. All latency measurements are obtained on an NVIDIA A6000 GPU with a batch size of 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.911, + 0.902 + ], + "angle": 0, + "content": "General Efficiency Savings According to Figure 7, we can gain up to \\(20 \\times\\) and \\(6.3 \\times\\) speedup in decoding attention kernel time on long-context video and V-NIAH with LongVILA [53]. Further, our multi-turn scaling exhibits compounded improvement, achieving a 1:1.7 ratio in conversation turn speed between the vanilla and SparseVILA. Beyond latency measurements, SparseVILA can cut down" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23790" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.194, + 0.089, + 0.295, + 0.099 + ], + "angle": 0, + "content": "Context Stage Sparsity" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.108, + 0.242, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.224, + 0.194, + 0.234 + ], + "angle": 0, + "content": "\\\\n" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.108, + 0.396, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.224, + 0.361, + 0.234 + ], + "angle": 0, + "content": "Context Pruning" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.107, + 0.549, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.424, + 0.224, + 0.527, + 0.234 + ], + "angle": 0, + "content": "Visual Attention Sinks" + }, + { + "type": "header", + "bbox": [ + 0.604, + 0.09, + 0.712, + 0.1 + ], + "angle": 0, + "content": "Decoding Stage Sparsity" + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.16, + 0.577, + 0.179 + ], + "angle": 0, + "content": "+" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.105, + 0.733, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.223, + 0.702, + 0.233 + ], + "angle": 0, + "content": "Prompt 1 Retrieval" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.105, + 0.901, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.223, + 0.871, + 0.232 + ], + "angle": 0, + "content": "Prompt 2 Retrieval" + }, + { + "type": "image_footnote", + "bbox": [ + 0.097, + 0.242, + 0.294, + 0.254 + ], + "angle": 0, + "content": "Prompt 1: Tell me about the car in the image." + }, + { + "type": "image_footnote", + "bbox": [ + 0.099, + 0.256, + 0.439, + 0.274 + ], + "angle": 0, + "content": "SparseVILA: \"The car in the image is a blue sports car, possibly a Mazda RX-8, driving down the road." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.242, + 0.439, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.491, + 0.242, + 0.714, + 0.253 + ], + "angle": 0, + "content": "Prompt 2: Tell me about the deer crossing the road" + }, + { + "type": "image_footnote", + "bbox": [ + 0.492, + 0.256, + 0.872, + 0.282 + ], + "angle": 0, + "content": "SparseVILA: \"In the image, a deer is crossing a road in front of a car. The deer is walking across the street, and the car is waiting patiently behind it ... this moment highlights the importance of being cautious and aware of wildlife while driving...\"" + }, + { + "type": "list", + "bbox": [ + 0.491, + 0.242, + 0.872, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Figure 8. Visualization of selected tokens from the KV cache for various queries in the multi-round setting. The heatmap illustrates the frequency of token selection from the KV Cache for the particular query of interest. We conduct said visualization on LLaVA-1.5 [27]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.4 + ], + "angle": 0, + "content": "the peak memory cost by \\(39.3\\%\\) for Qwen2-7B [54] on long context video tasks. Finally, we can decrease linear flops in the context stage by up to \\(87.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.485, + 0.591 + ], + "angle": 0, + "content": "Multi-Round Efficiency In Figure 7 we plot the multiround scaling for the baseline, SparseVLM [60] with context reprefilling, and SparseVILA. Over multiple rounds, our method retains a consistent context latency enabled through our context sparsity approach. Further, the generation latency scales faster than both query-aware and vanilla implementations, offering larger savings at increased rounds. Notably, SparseVLM [60] will scale past the original latency cost of the vanilla baseline after round 3 on Llama2 [45] indicating the inefficiency of that approach for multi-turn settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.283, + 0.633 + ], + "angle": 0, + "content": "4.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We provide a qualitative visualization of SparseVILA's token selection strategy in Figure 8 and Figure A5. We apply a \\(50\\%\\) context and \\(75\\%\\) generation sparsity for this visualization on LLaVA-1.5 [27]. During the decoding stage, tokens are selected based on their alignment with query information by thresholding attention weights. Since different tokens can be selected at each layer, the heatmap visualizes the frequency by which a token was selected over all LLM layers. Certain tokens were selected across most layers, and consistent over different queries, we coin these, visual sink tokens aligning with previous definitions [51]. The remainder of tokens selected are considered query-dependent, which we term as retrieval tokens. This division is quite intuitive in the qualitative samples shown in Figure A5. Our decoupled framework retains both sink and retrieval tokens in the KV cache. Further analysis and visualizations are provided in Section A.8 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.353, + 0.634, + 0.369 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.909, + 0.642 + ], + "angle": 0, + "content": "We introduce SparseVILA, a novel query-aware token retrieval method designed to optimize inference in Vision-Language Models (VLMs). Through a decoupled sparsity framework SparseVILA leverages query-agnostic context stage pruning with query-aware generation stage retrieval. By migrating sparsity into the decoding stage, SparseVILA is able to achieve state-of-the-art performance on multi-turn benchmarks, with a significant \\(+5.9\\%\\) improvement margin on image-centric tasks. By focusing on both context and decoding stage optimization, SparseVILA can achieve \\(3.6\\times\\) context stage speedup, with \\(1.7\\times\\) decoding stage acceleration in long-context video tasks. Further SparseVILA is able to retain lossless performance in both long-context/generation tasks and needle-in-a-haystack retrieval. SparseVILA introduces a new paradigm in prefetch decoding sparsity decoupling that maintains performance and improves efficiency. This approach enables VLMs to efficiently handle longer contexts, accommodating hundreds of images and frames." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.667, + 0.611, + 0.683 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.692, + 0.909, + 0.761 + ], + "angle": 0, + "content": "[1] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments, 2024. 3, 5, 6, 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.761, + 0.908, + 0.83 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.831, + 0.908, + 0.885 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.887, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.692, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models, 2024. 2, 5, 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[5] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2061-2070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[6] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.302, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[8] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.372, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Yi-Fan Zhang, Shukang Yin, Bo Li, Xinyu Fang, Sirui Zhao, Haodong Duan, Xing Sun, Ziwei Liu, Liang Wang, et al. Mme-survey: A comprehensive survey on evaluation of multimodal llms. arXiv preprint arXiv:2411.15296, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.511, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[11] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. Lazyllm: Dynamic token pruning for efficient long context llm inference, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[12] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 773-783, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[13] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun Sophia Shao, Kurt Keutzer, and Amir Gholami. Kvquant: Towards 10 million context length llm inference with kv cache quantization. arXiv preprint arXiv:2401.18079, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[14] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[16] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[17] Samir Khaki and Konstantinos N Plataniotis. The need for speed: Pruning transformers with one recipe. In *The Twelfth" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "International Conference on Learning Representations, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.163 + ], + "angle": 0, + "content": "[18] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[19] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, et al. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[20] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 3, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation, 2024. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[23] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[25] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for lIm compression and acceleration, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[26] Yujun Lin, Haotian Tang, Shang Yang, Zhekai Zhang, Guangxuan Xiao, Chuang Gan, and Song Han. Qserve: W4a8kv4 quantization and system co-design for efficient llm serving. arXiv preprint arXiv:2405.04532, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.646 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.907, + 0.689 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,7,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.691, + 0.907, + 0.746 + ], + "angle": 0, + "content": "[29] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[31] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23792" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.133 + ], + "angle": 0, + "content": "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[33] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.439 + ], + "angle": 0, + "content": "[36] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[38] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[39] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 3, 5, 6, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[41] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[42] Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference, 2024. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[43] Quan Tang, Bowen Zhang, Jiajun Liu, Fagui Liu, and Yifan Liu. Dynamic token pruning in plain vision transformers for semantic segmentation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[44] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165-12174, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[47] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[48] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[49] Zichen Wen, Yifeng Gao, Weijia Li, Conghui He, and Linfeng Zhang. Token pruning in multimodal large language models: Are we solving the right problem? arXiv preprint arXiv:2502.11501, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.532 + ], + "angle": 0, + "content": "[50] Guangxuan Xiao, Jiaming Tang, Jingwei Zuo, Junxian Guo, Shang Yang, Haotian Tang, Yao Fu, and Song Han. Duoattention: Efficient long-context llm inference with retrieval and streaming heads, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.907, + 0.576 + ], + "angle": 0, + "content": "[51] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks, 2024. 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.907, + 0.648 + ], + "angle": 0, + "content": "[52] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.907, + 0.733 + ], + "angle": 0, + "content": "[53] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, Yihui He, Hongxu Yin, Pavlo Molchanov, Jan Kautz, Linxi Fan, Yuke Zhu, Yao Lu, and Song Han. Longvila: Scaling long-context visual language models for long videos. 2024. 6, 7, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[54] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23793" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[55] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2, 3, 5, 6, 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[56] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[59] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024.6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[60] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference arXiv preprint arXiv:2410.04417, 2024. 1, 2, 5, 6, 7, 8, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[61] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[62] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, Zhangyang Wang, and Beidi Chen. \\(\\mathrm{H}_{20}\\): Heavy-hitter oracle for efficient generative inference of large language models, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[63] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[64] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning. arXiv preprint arXiv:2104.08500, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.714 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23794" + } + ] +] \ No newline at end of file diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_origin.pdf b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec106efe3b59a82ab82262f8db8cbf905b500cad --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/9b15689c-1721-44a2-b85e-1b1667a45deb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bee1943695e5de20513326aca2c3fc54b02a8c06f4697e297a96df8bc6e0160 +size 6717609 diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/full.md b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1319b17005dfa7274983f5e9d85ec1490b51e5ae --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/full.md @@ -0,0 +1,311 @@ +# SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference + +Samir Khaki + +Konstantinos N. Plataniotis4 + +NVIDIA + +$^{2}$ MIT + +Jiaming Tang² + +Yao Lu + +$^{3}$ UC San Diego + +Shang Yang2 + +Song Han1,2 + +Yukang Chen + +Zhijian Liu $^{1,3}$ + +$^{4}$ University of Toronto + +![](images/8fc9719806016f3661bfa5c364611f29c8b92b6c23c53ce8c39fb353eadfdd14.jpg) +Multi-Round Performance + +![](images/21bf3170cfdc35fef1d36474a6efff25ddb4cae0791f21a9e123baf1fd2c06dd.jpg) +Image-Centric Evaluation +Figure 1. Overview of SparseVILA. Compared to existing approaches, SparseVILA exhibits advantages in the following aspects: 1. Multi-Round Performance. SparseVILA leverages a decoupled sparsity framework to achieve state-of-the-art performance at a $1.4 \times$ speedup on LLaVA-NeXT [28]. 2. Image-centric evaluation. Leveraging decoupled sparsity, SparseVILA can retain higher context tokens resulting in a $5.9\%$ margin over previous approaches. 3. SparseVILA scales to long-context and long-generation tasks, achieving lossless performance with up to a $3.6 \times$ context, $1.7 \times$ generation, and $2.2 \times$ end-to-end speedup. + +![](images/394567e0efa79343b86aa01a409fd220861792fd57d1bfadc8ac066c36f4b041.jpg) +Long-Context/Generation Performance + +# Abstract + +Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM, by pruning tokens in the context stage while attending to a sparse subset of visual tokens during the generation phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a $1.4 \times$ speedup on image benchmarks. This approach leads to $+5.9\%$ average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a $3.6 \times$ and $1.7 \times$ speedup in prefetch and decoding, respectively. + +# 1. Introduction + +Vision Language Models (VLMs) have emerged as a state-of-the-art conversational tool enabling users to directly interact with Large Language Models (LLMs) using various visual features, including photographs, documents, and videos [2, 24, 27, 30, 47]. Unfortunately, this added modality comes at the expense of higher latency and memory associated with processing the visual tokens in the LLM. Hence, deploying LLMs efficiently at inference time remains a challenge. + +Several works have aimed to reduce these associated costs through model pruning on the LLM or vision encoder [32, 41, 64], KV cache compression [13, 26, 29], and most recently, token sparsification [60]. By reducing the amount of computation in the inference pipeline, many of these methods can achieve significant context stage savings as this stage of the network is mainly compute-bound. + +Looking beyond the context, real-world applications often demand extensive generation. Tasks such as image captioning may require a few hundred tokens, meanwhile, video captioning or detailing easily requires more than a few thousand generated tokens. Hence in such applications, it is not sufficient to only focus on context stage optimization – effi + +cient implementations with real-world applications should focus on both context and decoding stage optimizations. One such example is multi-turn conversation. + +Multi-turn/round conversation is a practical use case for VLMs, as it is entirely possible for a user to pose multiple questions about a given visual input. In fact, most benchmarks inherently support multi-round conversation: the GQA dataset [15] has more than 90 questions for the same visual input. Despite this, most evaluation benchmarks run single-round evaluation (i.e. repetitive pre-filling), which is not only unrealistic but also inefficient as the context stage would be repeated for each generation round. In real-world scenarios, the visual input could span tens of thousands of context tokens, hence repetitive pre-filling would dramatically slow down the user's interaction with their VLM. + +In this work, we aim to present a unified approach for tackling context and decoding latency in modern VLMs. Latency is a common challenge associated with VLMs due to the sheer amount of visual tokens to be processed. Existing methods for accelerating VLM inference primarily focus on token-wise pruning or merging techniques, with recent approaches leveraging textual priors to reduce visual token complexity in a query-aware manner. In practice, methods that permanently remove visual tokens during the context stage are quite lossy in multi-turn evaluations as visualized in Figure 1. Hence, in this paper, we introduce SparseVILA as a novel approach for accelerating VLM inference, while retaining multi-turn performance. + +Our key insight is a decoupled sparsity framework enabling SparseVILA to migrate sparsity from the context into the decoding stage. Further, SparseVILA leverages query-aware retrieval in the decoding stage, supporting multi-turn conversation as a different subset of context tokens can be retrieved per question. This decoupled approach allows SparseVILA to achieve significant performance improvements in image-centric benchmarks as shown in Figure 1 and outperforms previous methods in long-context/generation scaling. + +# 2. Preliminaries and Related Work + +Token pruning has proven effective in accelerating inference across a variety of tasks, including image classification [3, 17, 19] and semantic segmentation [5, 43]. With the rise of generative AI, these techniques have been further extended to diffusion models, large language models [18], and vision-language models [14, 55, 60]. We refer the readers to Section A.2 for a more detailed survey of related work. + +This paper focuses on token pruning for vision-language models (VLMs). The key idea is that not all visual tokens contribute equally to VLM's final prediction. By identifying and removing less informative tokens, it is possible to significantly reduce the computational cost of VLM inference and thereby improve efficiency. Existing work in this area largely differs in how these tokens are selected. We categorize prior + +methods into two groups based on their dependence on the input query: (i) query-agnostic approaches, which identify unimportant tokens based solely on visual saliency or redundancy, and (ii) query-aware approaches, which incorporate the semantic relationship between visual and textual inputs to guide pruning. For each category, we highlight representative methods and their limitations, laying the groundwork for the approach proposed in this paper. + +# 2.1. Query-Agnostic Sparsity + +Query-agnostic token pruning methods aim to reduce redundancy or select important visual tokens without relying on the textual input [49]. They prune tokens based solely on the visual context—either within or after the vision encoder. For example, PruMerge [39] clusters and discards less informative tokens using final-layer attention scores, while VisionZip [55] employs a token merging module, similar to ToMe [3], to compress redundant visual information. + +This visual-only focus, however, presents key limitations. First, these methods often sacrifice fine-grained visual details, especially under high sparsity, which can degrade performance. More importantly, they cannot adapt token selection based on the input query, leading to suboptimal results when task-relevant information is sparsely distributed. As shown in Figure 1, query-agnostic methods underperform on text-centric and structured visual understanding benchmarks such as DocVQA [35], InfoVQA [36], and ChartQA [34]. By treating all visual tokens uniformly, they risk discarding critical information necessary for accurate reasoning. + +# 2.2. Query-Aware Sparsity + +Query-aware pruning improves visual token selection by explicitly modeling the relationship between textual queries and visual representations. For example, FastV [4] leverages attention maps from early LLM layers as a salience metric to guide token pruning during the prefetch stage. This contrasts with query-agnostic approaches, which perform token reduction prior to any LLM processing. Similarly, SparseVLM [60] utilizes query-to-vision attention scores to eliminate less relevant visual tokens during prefetch. + +While query-aware methods are effective at aligning token selection with the current query, they face limitations in multi-turn interactions. Pruning based solely on the initial query can eliminate tokens that are crucial for future queries, resulting in degraded performance. Empirically, such methods show a huge drop in accuracy across multi-turn conversations, often underperforming compared to query-agnostic methods (Figure 1). Re-prefilling at each turn can maintain accuracy, but it introduces significant latency (Section 4.4). + +To systematically evaluate the performance ceiling of query-aware pruning in multi-turn contexts, we introduce a query-aware oracle, which selects the optimal subset of visual tokens to maximize agreement with the original model's + +![](images/dd42b7d99850455a33b4ce37211e8e9c187a67c13fcf2b81088a9d724017d339.jpg) +Figure 2. Multi-Round Query-Aware Comparisons with LLaVA-1.5 [27] on the POPE [21] dataset. Without re-profilling the context, the query-aware oracle degrades heavily indicating the inability of query-aware pruning to scale effectively in multi-turn conversation. + +response (see Section A.3 for more details). From Figure 2, pruning decisions made in early rounds often result in substantial degradation in later responses, underscoring the long-term cost of early token elimination. By incorporating both the current query and the ground-truth response, the oracle establishes an upper bound on the achievable performance of any query-aware pruning method. These findings reveal a fundamental limitation of query-aware pruning in sustaining performance over multi-turn dialogue. + +# 3. SparseVILA: Best of Both Worlds + +Both query-agnostic and query-aware approaches offer distinct advantages and face inherent limitations. Existing visual token pruning methods typically perform pruning exclusively during the prefetch stage, with the benefits of reduced computation carrying over naturally to the subsequent decode stage. In this paper, we introduce SparseVILA that decouples visual sparsity between the prefetch and decode stages. During the prefetch stage, we apply query-agnostic pruning to eliminate redundant tokens that are unlikely to be useful for any downstream queries. In the decode stage, we employ a query-aware retrieval mechanism to selectively fetch the most relevant entries from the KV cache based on the current query, ensuring that only tokens directly related to the question are used for generating response. This design combines the strengths of both paradigms, achieving efficient token reduction while preserving query relevance. + +# 3.1. Prefix Fill Phase: Query-Agnostic Pruning + +During the prefetch phase, the VLM processes the system, image, and potentially the first query tokens. To support multi-turn settings, pruning in this stage remains query-agnostic, targeting visual redundancy or salience rather than textual correlation. Since the visual content is only prefilled once, token reduction remains independent of specific queries to effectively support multi-turn conversations. + +We estimate token importance using attention scores + +from the visual encoder, providing an unbiased (query-independent) assessment of token salience, as shown in previous works [1, 39, 55]. This approach is detailed in Algorithm 2 under Section A.4 of the supplementary material. By default, we take the mean-column sum of the attention weights, making our method compatible with architectures like SigLIP [58] that don't produce a $cls$ token. For CLIP-style [37] visual encoders that do generate a $cls$ token, we compute the mean-column-sum relative to that token, leveraging the encoder's training that routes high information density to this specific token. + +# 3.2. Decode Phase: Query-Aware Retrieval + +During the generation phase, the VLM is typically memory-bound as it implements next-token prediction by computing attention with the existing pre-filled KV cache. By reducing the effective size of the KV cache we can accelerate this generation process while maintaining the necessary visual information to support multi-turn conversation. + +We identify which visual tokens from the KV cache can be contextually selected for all associated generation steps in answering the current query. During query prefilling at each conversation round, we calculate the mean-columnsum of the attention weights between the query and the remaining visual tokens from the context stage. Using a sparsity threshold, we select a compact subset of the KV cache to participate in all generation steps associated with the query, thereby achieving significant speedup in decoding attention. We detail the retrieval strategy in Algorithm 3 under Section A.4 of the supplementary materials. + +While prefilling the current round query, we run Algorithm 3 simultaneously to compute the indices of the KV Cache for generation, immediately after each layer's self-attention block. To avoid introducing complexity and overhead into the decoding stage, we create a smaller compact dense KV cache representing only the selected indices for the current generation round. We further discuss the efficiency benefits of this in Section 3.4. + +One interesting finding is that the selected tokens can be further divided as Visual Attention Sinks and Visual Retrieval Tokens. The sink tokens appear consistent over different queries, meanwhile, retrieval tokens are directly related to the given query. We discuss this further with qualitative examples in Section 4.5 and Section A.4 in our supplementary material. Similar findings on visual token sinks have been echoed in previous works, such as VisionZip [55] and Visual Attention Redistribution (VAR) [16]. + +Discussions. Beyond VLM pruning approaches, KV cache compression has been heavily explored in the context of LLMs. In particular, methods such as SnapKV [22] and Quest [42] leverage salience metrics to select a smaller KV cache for decoding, thus similarly accelerating the subsequent decoding stage. In the case of SnapKV, after selecting + +![](images/de8c37865d095ef0a18295c2212995323b6dd8cff930e39b771f9eee0a2b81e0.jpg) +Figure 3. Overview of the components in SparseVILA. Query agnostic pruning is run in the context stage using the salience map from the visual encoder to guide token selection. In the decoding stage, the question is used to retrieve a sparse set of critical visual tokens to participate in the generation stage, hence achieving decoding speedup. + +![](images/930cb03e8ed956c583846aa16aacf704f3c80be744de50ea9a8e4e9a3e3b28f8.jpg) +Figure 4. Comparing sparsity migration on VideoMME at a fixed latency speedup of $2.2 \times$ . SparseVILA* uses a higher context sparsity, SparseVILA uses our default configuration. + +tokens relevant to the observation window, this method discards the original KV cache, making it ill-suited to multi-turn conversation. Meanwhile, Quest re-runs its selection mechanism for every generated token, significantly increasing the complexity of the decoding stage, while only achieving speedup in long-context cases. We provide detailed comparisons with each of these methods in Section A.4.3 of the supplementary material. SparseVILA introduces a novel combination of query-aware and query-agnostic sparsity, leveraging context and decoding stage improvement beyond any decoding-only or context-only method; hence, the novelty remains in the design of a unified framework for efficient VLM inference. + +# 3.3. Decoupled Prefix-Decode Visual Sparsity + +SparseVILA's decoupled framework enables the migration of computational reduction from the context stage to the generation stage. This approach is based on two key insights. Firstly, the context stage should preserve enough information to support multi-turn conversation as it is only pre-filled once per set of queries. Secondly, the generation stage can leverage query-aware information to guide the pruning process + +and accelerate the decoding, without affecting subsequent turns of conversation. + +As shown in Figure 1, migrating visual sparsity from prefetch stage into the decode stage enables SparseVILA to outperform previous pruning approaches while achieving better speedup. We further investigate this effect by comparing two variants of SparseVILA that both achieve a $2.2 \times$ speedup on the VideoMME benchmark [9], as shown in Figure 4. By migrating sparsity from the context stage into the decoding stage (i.e., SparseVILA* → SparseVILA), we observe a $2.3\%$ improvement in overall performance, surpassing the baseline implementation. Moreover, this shift allows the model to retain more tokens during the context stage, resulting in a significant $5.9\%$ gain on image-centric benchmarks. + +# 3.4. Efficient Implementation + +SparseVILA introduces a two-stage compression approach to distribute sparsity over both the context and decoding stages. This unique approach allows SparseVILA to deliver strong accuracy performance with measurable speedup. Our method, leverages a carefully designed implementation to support modern acceleration techniques. In particular, all parts of the language model leverage the latest FlashAttention2 [6] implementation. + +To enable KV-Cache retrieval, we dynamically recompute the QK attention score map once per question at each layer and compress it into a compact, dense KV cache. This design allows us to fully utilize efficient attention implementations across the network. While a naive approach to computing the mean column sum of the attention weights can introduce latency and increase peak memory usage – especially at longer context lengths – we address this with a custom Triton + +![](images/3619d9be254e7d1541af6629e0e877558ca87a8d1f5f7b2bcdb0852c46e49f68.jpg) +Figure 5. Illustrates the lightweight recomputation of SparseVILA alongside layer-wise FlashAttention. + +kernel that streams attention weight normalization efficiently. This kernel achieves exact results without materializing the full $S \times S$ attention matrix, significantly improving both speed and memory efficiency. Furthermore, due to the high sparsity achieved during the generation phase, our compact KV cache incurs minimal memory overhead (up to 55MB), while reducing peak memory consumption by $39.3\%$ on long-context video tasks with Qwen2-7B [54]. + +Building a compact KV cache avoids any non-contiguous "on-the-fly indexing" (sparse access patterns) during the generation stage. The overhead from this additional computation is aggregated with the packing of the KV Cache into a contiguous set as shown in Figure 5, accounting for less than $4\%$ of the generation runtime for 20, and 100 tokens on Llama2 [45], and Llama3 [7]/Qwen2 [54], respectively. + +# 4. Experiments + +# 4.1. Setup + +Multi-Turn Evaluation. Many existing benchmarks generate multiple question-answer pairs for each image: e.g., up to 18 in POPE [21]. However, evaluation protocols often remain confined to single-turn settings. In this work, we make use of the inherent multi-turn structure of such datasets to enable more efficient and accurate assessment of VLMs. We will release our multi-turn evaluation framework to support the broader multimodal research community. + +We organize questions associated with the same image into coherent multi-turn conversations, which allows visual tokens to be prefilled only once. This setup not only reduces computational overhead but also better reflects realistic VLM usage in interactive settings. However, a potential problem is information leakage between turns, where earlier questions can unintentionally reveal answers to later ones. For example, in the GQA dataset [15]: + +Q1: "What is the person in front of the sky doing?" +Q2: "What's the person in front of?" + +Here, Q1 discloses information that effectively answers Q2, + +enabling the model to respond correctly even without relying on the image. To mitigate this, we implement a partial KV cache eviction strategy: after each round, we remove only the KV entries corresponding to the previous question and answer. This preserves the efficiency gains of visual KV cache reuse while preventing unintended context carryover between turns. We provide more details in Section A.5.3. + +Sparsity Ratio. Our sparsity configuration adopts a straightforward approach for both context and decoding, ensuring efficient implementation and cross-model support. Specifically, we set a constant context sparsity before the LLM, with a uniform decoding sparsity across all layers. More granular strategies, such as layer-wise or head-aware sparsity, may further optimize performance but introduce additional complexity and tuning. We prioritize simplicity and generalization, leaving these refinements for future work. + +# 4.2. Results on Image Benchmarks + +In Table 1, we evaluate our SparseVILA against a diverse set of query-aware and query-agnostic baselines across eight vision-language benchmarks: ChartQA [34], DocVQA [35], GQA [15], InfographicVQA [36], MME [8], POPE [21], ScienceQA [31], and TextVQA [40]. All results in this section are conducted on LLaVA-NeXT-7B [61]. To further evaluate the generality of SparseVILA, we extend our analysis to additional VLMs, including LLaVA-1.5 [27], VILA-1.5 [24], Qwen2-VL [46], and NVILA [30], in Table A5. + +We compare our SparseVILA with existing query-aware pruning methods, including FastV [4], SparseVLM [60] and PDrop [52]. Our results show that at a given speedup, SparseVILA outperforms these prior approaches in multi-turn evaluation by migrating sparsity into the decode phase. Notably, SparseVILA achieves a $6.4\%$ accuracy improvement on DocVQA, highlighting the effectiveness of our decoupled prefetch-decode sparsity approach. + +We further evaluate SparseVILA against existing query-agnostic methods, including PruMerge [39], HIRED [1], and VisionZip [55]. At an end-to-end speedup of $1.44 \times$ , SparseVILA achieves at least a $3.2\%$ improvement on DocVQA. These results further demonstrate that, under a fixed speedup target, SparseVILA's decoupled sparsity approach consistently outperforms prior state-of-the-art pruning methods. In Section A.6.1.1, we also show that applying SparseVILA to these existing query-agnostic sparsity methods can further enhance their performance. + +# 4.3. Results on Video Benchmarks + +In Table 2, we demonstrate that SparseVILA consistently outperforms existing approaches on long-context tasks. Additionally, query-aware works such as FastV [4] and SparseVLM [60] were unable to scale effectively to long context (64K) as they require access to the full visual context and query at the same time to compute their ranking scores, hence + +Table 1. Results on image benchmarks. We provide query-aware comparisons with SparseVLM [60] and FastV [4] and query-agnostic comparisons with LLaVA-PruMerge [39], HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.2. At a given latency, we show significant improvements over state-of-the-art approaches; hence, SparseVILA optimizes the latency-to-performance tradeoff. A generation length of 250 tokens is used to approximate the sequence length observed in image-c captioning tasks. TextVQA [40], DocVQA (validation set) [35], InfoVQA (validation set) [36], ChartQA [34], ScienceQA-Img [31], POPE [21], and GQA [15]. + +
Sparsity (%)SpeedupCQADVQAGQAIVQAMMEPOPESQATVQA
PrefillDecodePrefillDecodeOverall
LLaVA-NeXT-7B-----50.774.464.237.71846.185.970.361.2
+ FastV8303.26×1.24×1.37×37.354.460.631.81775.282.867.557.1
+ SparseVLM8303.26×1.24×1.37×37.754.460.831.41780.784.067.558.1
+ PDrop5301.89×1.16×1.24×41.544.862.926.51809.286.169.157.0
+ SparseVILA67502.59×1.27×1.38×41.860.863.433.31818.485.968.060.2
+ PruMerge8304.34×1.27×1.43×26.936.961.128.51753.283.666.754.6
+ HIRED8304.34×1.27×1.43×29.741.662.428.21777.181.767.755.9
+ VisionZip8304.34×1.27×1.43×34.151.261.230.61760.383.767.458.7
+ SparseVILA67902.59×1.33×1.44×39.954.463.430.81827.886.168.058.9
+ +Table 2. Results on video benchmarks. We provide query-agnostic comparisons with HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.3. $^{\dagger}$ FastV [4], SparseVLM [60], and PDrop [52] do not scale to 64K context length, hence resulting in $OOM -$ we include shorter context comparison in Section A.6.2 of our supplementary material. A generation length of 2000 tokens is used to approximate the sequence length observed in video captioning or reasoning tasks. + +
Sparsity (%)SpeedupMLVUVideo-MME (w/o subtitle)
PrefillDecodePrefillDecodeOverallm-avgShortMediumLongOverall
LongVILA-7B-----72.468.958.152.059.7
+ HIRED8604.9×1.5×2.1×70.268.054.649.657.4
+ VisionZip8604.9×1.5×2.1×73.166.655.150.957.5
+ SparseVILA75943.6×1.7×2.2×74.470.658.651.760.3
+ +Table 3. Long-Context/Generation Performance on VideoChatGPT [33]. Scores are computed using the GPT4-mini assisted evaluation. SparseVILA offers lossless generation performance at a $2.2 \times$ end-to-end speedup. + +
CIDOCUTUCMeanSpeedup
LongVILA-7B2.332.282.871.742.412.331.0×
+ SparseVILA2.362.342.921.882.562.412.2×
+ +preventing chunk pre-filling and adding significant memory overhead to their approach. This consequently limits their ability to scale. In contrast, SparseVILA efficiently scales to 64K context (LongVILA 256 Frames) by decoupling our sparsity in the context and decoding stage, achieving $3.6 \times$ prefilling speedup, $1.7 \times$ decoding speedup, and $2.2 \times$ end-to-end speedup. We include additional results on 8K context in Section A.6.2. + +One of the key novelties in SparseVILA is the decoupling and migration of sparsity into the decoding stage. As + +previously shown, this has a positive effect on image and long-context benchmarks. In this section, we want to validate that SparseVILA's heavy generation sparsity does not affect the model's ability to generate responses by testing on free-form text generation tasks. Our approach delivers significant efficiency gains, offering up to $2.2 \times$ E2E speedup over the vanilla model while maintaining performance on the Video-ChatGPT [33] benchmark as shown in Table 3. + +In addition to VQA and captioning tasks, we also evaluate the long-context retrieval performance with Visual Needle in a Haystack (V-NIAH) benchmark [53, 59]. The original V-NIAH only supports single-turn, so we extended this benchmark to a multi-turn setting. See Section A.7 of the supplementary material for more details. We compare the long-context retrieval capacity of our method with the vanilla implementation, SparseVLM [60], and FastV [4] on LongVILA [53] in Figure 6. Notably, SparseVLM and FastV do not scale beyond 32 Frames (approx. 8K context lengths) as they lack support for chunkedprefilling and incur a memory overhead in computing their salience metric. Sparse- + +![](images/cc2bd110ef826cf0e7017e42b29f665117d37643ad70ad228b9f0336a952876b.jpg) + +![](images/b80bb9a085e40c7ea250e2b9187ef4b334bfac1f627160d17862154e35216d05.jpg) +Figure 6. Comparing the multi-turn performance of the vanilla implementation with SparseVLM [60], FastV [4] and SparseVILA on LongVILA-Qwen2 [53] for V-NIAH. FastV and SparseVLM reach OOM after 30 Frames (8K context) as they do not inherently support chunk prefilling. We include the full scaling up to 500 Frames (128K context length) in Section A.7. +Figure 7. Decoding attention kernel speed for video tasks and V-NIAH on LongVILA-7B [53] and image tasks on LLaVA-NeXT-7B [28]. We further explore the multi-turn scaling of the baseline, a query-aware approach with re-pre-filling and SparseVILA over conversation rounds 0, 1, and 3. + +VILA maintains comparable performance with the baseline. In contrast, query-aware context stage pruning strategies begin to degrade at only an 8K context length. In reference, to Figure 6, the vanilla implementation remains effective up to a context length of 89.6K, handling contexts up to $10 \times$ as long. Meanwhile, SparseVLM [60] and FastV [4] exhibit degradation, indicating that these pruning frameworks significantly impair the model's capacity for long context retrieval in multi-turn evaluation. We find that our approach performs similarly to the vanilla baseline, indicating that SparseVILA can retain long context retrieval performance. We can attribute the retained performance to the query-aware nature of our decoding sparsity. We leverage a $90\%$ decoding sparsity, translating to a $2 \times$ decoding speedup. Additional details are provided in Section A.7 of the supplementary material. + +# 4.4. Efficiency Analysis + +In this section, we provide details of our optimized inference pipeline and additional results. Efficiency metrics are reported relative to the optimized baseline, adding additional + +benefit for efficient VLM inference. + +Inference Setting AWQ [25] is a widely adopted technique for reducing decoding latency by significantly decreasing the latency of linear layers. We developed an optimized inference pipeline leveraging AWQ [25] based on TinyChat. In the case of Llama2-7B [45] with a 32K context length, our dense configuration achieves a $1.55\mathrm{x}$ speedup over a finely-tuned FP16 pipeline, with pre-allocated KV cache across the entire benchmark sequence, effectively eliminating the additional overhead associated with dynamic memory allocation. All latency measurements are obtained on an NVIDIA A6000 GPU with a batch size of 1. + +General Efficiency Savings According to Figure 7, we can gain up to $20 \times$ and $6.3 \times$ speedup in decoding attention kernel time on long-context video and V-NIAH with LongVILA [53]. Further, our multi-turn scaling exhibits compounded improvement, achieving a 1:1.7 ratio in conversation turn speed between the vanilla and SparseVILA. Beyond latency measurements, SparseVILA can cut down + +\\n +![](images/b8ff59445fda0cb6fe4361fe8174c8cfb0453b95cc6b66081db64b9a4a879ca6.jpg) +Prompt 1: Tell me about the car in the image. + +Context Pruning +![](images/9972c6c1e7b6405fbde4fc285a7fde35cbfecf1724bf2c092a0c4172538318d8.jpg) +SparseVILA: "The car in the image is a blue sports car, possibly a Mazda RX-8, driving down the road. + +Visual Attention Sinks +Figure 8. Visualization of selected tokens from the KV cache for various queries in the multi-round setting. The heatmap illustrates the frequency of token selection from the KV Cache for the particular query of interest. We conduct said visualization on LLaVA-1.5 [27] +![](images/e04d8c252e20bef311bbf75e04b7226c9de79c0ceaf96586a8c3cac1adc79829.jpg) +Prompt 2: Tell me about the deer crossing the road + ++ +Prompt 1 Retrieval +![](images/98c551e5cf89a5b3e0f7f4843c13cb45699456a2804a3f8858379569dfed9e0a.jpg) +SparseVILA: "In the image, a deer is crossing a road in front of a car. The deer is walking across the street, and the car is waiting patiently behind it ... this moment highlights the importance of being cautious and aware of wildlife while driving..." + +![](images/db1d650e701f4cf46fb7fc5aaa1582f2400b767e45e35cd1a6f84d8d1af11685.jpg) +Prompt 2 Retrieval + +the peak memory cost by $39.3\%$ for Qwen2-7B [54] on long context video tasks. Finally, we can decrease linear flops in the context stage by up to $87.6\%$ . + +Multi-Round Efficiency In Figure 7 we plot the multiround scaling for the baseline, SparseVLM [60] with context reprefilling, and SparseVILA. Over multiple rounds, our method retains a consistent context latency enabled through our context sparsity approach. Further, the generation latency scales faster than both query-aware and vanilla implementations, offering larger savings at increased rounds. Notably, SparseVLM [60] will scale past the original latency cost of the vanilla baseline after round 3 on Llama2 [45] indicating the inefficiency of that approach for multi-turn settings. + +# 4.5. Qualitative Analysis + +We provide a qualitative visualization of SparseVILA's token selection strategy in Figure 8 and Figure A5. We apply a $50\%$ context and $75\%$ generation sparsity for this visualization on LLaVA-1.5 [27]. During the decoding stage, tokens are selected based on their alignment with query information by thresholding attention weights. Since different tokens can be selected at each layer, the heatmap visualizes the frequency by which a token was selected over all LLM layers. Certain tokens were selected across most layers, and consistent over different queries, we coin these, visual sink tokens aligning with previous definitions [51]. The remainder of tokens selected are considered query-dependent, which we term as retrieval tokens. This division is quite intuitive in the qualitative samples shown in Figure A5. Our decoupled framework retains both sink and retrieval tokens in the KV cache. Further analysis and visualizations are provided in Section A.8 of the supplementary material. + +# 5. Conclusion + +We introduce SparseVILA, a novel query-aware token retrieval method designed to optimize inference in Vision-Language Models (VLMs). Through a decoupled sparsity framework SparseVILA leverages query-agnostic context stage pruning with query-aware generation stage retrieval. By migrating sparsity into the decoding stage, SparseVILA is able to achieve state-of-the-art performance on multi-turn benchmarks, with a significant $+5.9\%$ improvement margin on image-centric tasks. By focusing on both context and decoding stage optimization, SparseVILA can achieve $3.6\times$ context stage speedup, with $1.7\times$ decoding stage acceleration in long-context video tasks. Further SparseVILA is able to retain lossless performance in both long-context/generation tasks and needle-in-a-haystack retrieval. SparseVILA introduces a new paradigm in prefetch decoding sparsity decoupling that maintains performance and improves efficiency. This approach enables VLMs to efficiently handle longer contexts, accommodating hundreds of images and frames. + +# References + +[1] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments, 2024. 3, 5, 6, 1, 4, 7 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 1 +[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang + +Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models, 2024. 2, 5, 6, 7, 1, 4 +[5] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2061-2070, 2023. 2 +[6] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 4, 2 +[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5 +[8] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5 +[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 4 +[10] Chaoyou Fu, Yi-Fan Zhang, Shukang Yin, Bo Li, Xinyu Fang, Sirui Zhao, Haodong Duan, Xing Sun, Ziwei Liu, Liang Wang, et al. Mme-survey: A comprehensive survey on evaluation of multimodal llms. arXiv preprint arXiv:2411.15296, 2024. 4 +[11] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. Lazyllm: Dynamic token pruning for efficient long context llm inference, 2024. 1 +[12] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 773-783, 2023. 1 +[13] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun Sophia Shao, Kurt Keutzer, and Amir Gholami. Kvquant: Towards 10 million context length llm inference with kv cache quantization. arXiv preprint arXiv:2401.18079, 2024. 1 +[14] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2025. 2 +[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 5, 6, 4 +[16] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models, 2025. 3 +[17] Samir Khaki and Konstantinos N Plataniotis. The need for speed: Pruning transformers with one recipe. In *The Twelfth + +International Conference on Learning Representations, 2024. 2, 1 +[18] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers, 2022. 2 +[19] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, et al. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 2, 1 +[20] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 6 +[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 3, 5, 6, 4 +[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation, 2024. 3, 1 +[23] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 1 +[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 1, 5, 7 +[25] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for lIm compression and acceleration, 2024. 7 +[26] Yujun Lin, Haotian Tang, Shang Yang, Zhekai Zhang, Guangxuan Xiao, Chuang Gan, and Song Han. Qserve: W4a8kv4 quantization and system co-design for efficient llm serving. arXiv preprint arXiv:2405.04532, 2024. 1 +[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3, 5, 8 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,7,4 +[29] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750, 2024.1 +[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1, 5, 6, 7 +[31] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via + +thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 4 +[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1 +[33] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 6, 4, 5 +[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 5, 6, 4 +[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6, 4 +[36] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2, 5, 6, 4 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 2 +[38] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1 +[39] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 3, 5, 6, 1, 4 +[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 4 +[41] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 1 +[42] Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference, 2024. 3, 1 +[43] Quan Tang, Bowen Zhang, Jiajun Liu, Fagui Liu, and Yifan Liu. Dynamic token pruning in plain vision transformers for semantic segmentation, 2023. 2 +[44] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165-12174, 2022. 1 + +[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 7, 8, 3 +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 7 +[47] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution, 2024. 1 +[48] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 1 +[49] Zichen Wen, Yifeng Gao, Weijia Li, Conghui He, and Linfeng Zhang. Token pruning in multimodal large language models: Are we solving the right problem? arXiv preprint arXiv:2502.11501, 2025. 2 +[50] Guangxuan Xiao, Jiaming Tang, Jingwei Zuo, Junxian Guo, Shang Yang, Haotian Tang, Yao Fu, and Song Han. Duoattention: Efficient long-context llm inference with retrieval and streaming heads, 2024. 1 +[51] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks, 2024. 8, 1 +[52] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 5, 6, 7 +[53] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, Yihui He, Hongxu Yin, Pavlo Molchanov, Jan Kautz, Linxi Fan, Yuke Zhu, Yao Lu, and Song Han. Longvila: Scaling long-context visual language models for long videos. 2024. 6, 7, 4, 5 +[54] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, + +Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report, 2024. 5, 8 +[55] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2, 3, 5, 6, 1, 4, 7 +[56] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models, 2025. 1 +[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1 +[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 2 +[59] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024.6, 4 +[60] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference arXiv preprint arXiv:2410.04417, 2024. 1, 2, 5, 6, 7, 8, 3, 4 +[61] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 5 +[62] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, Zhangyang Wang, and Beidi Chen. $\mathrm{H}_{20}$ : Heavy-hitter oracle for efficient generative inference of large language models, 2023. 1 +[63] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.4 +[64] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning. arXiv preprint arXiv:2104.08500, 2021. 1 \ No newline at end of file diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/images.zip b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86a373e26a7740d745c8d30e88804da593c31ad4 --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0de3532aec7819c5c81aae1decee8e306448b2de03f5c86dd044465c97fb61ca +size 577350 diff --git a/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/layout.json b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..827f93c7733f4061566148480934ed6bb6f6f8e9 --- /dev/null +++ b/2025/SparseVILA_ Decoupling Visual Sparsity for Efficient VLM Inference/layout.json @@ -0,0 +1,8341 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 520, + 121 + ], + "type": "text", + "content": "SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 151, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 151, + 157 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 151, + 157 + ], + "type": "text", + "content": "Samir Khaki" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 159, + 250, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 159, + 250, + 172 + ], + "spans": [ + { + "bbox": [ + 113, + 159, + 250, + 172 + ], + "type": "text", + "content": "Konstantinos N. Plataniotis4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 178, + 202, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 178, + 202, + 191 + ], + "spans": [ + { + "bbox": [ + 152, + 178, + 202, + 191 + ], + "type": "text", + "content": "NVIDIA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 213, + 178, + 244, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 178, + 244, + 190 + ], + "spans": [ + { + "bbox": [ + 213, + 178, + 244, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 213, + 178, + 244, + 190 + ], + "type": "text", + "content": "MIT" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 267, + 143, + 339, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 143, + 339, + 158 + ], + "spans": [ + { + "bbox": [ + 267, + 143, + 339, + 158 + ], + "type": "text", + "content": "Jiaming Tang²" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 276, + 159, + 317, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 159, + 317, + 172 + ], + "spans": [ + { + "bbox": [ + 276, + 159, + 317, + 172 + ], + "type": "text", + "content": "Yao Lu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 255, + 178, + 332, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 178, + 332, + 193 + ], + "spans": [ + { + "bbox": [ + 255, + 178, + 332, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 255, + 178, + 332, + 193 + ], + "type": "text", + "content": "UC San Diego" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 143, + 428, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 143, + 428, + 158 + ], + "spans": [ + { + "bbox": [ + 345, + 143, + 428, + 158 + ], + "type": "text", + "content": "Shang Yang2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 159, + 402, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 159, + 402, + 174 + ], + "spans": [ + { + "bbox": [ + 342, + 159, + 402, + 174 + ], + "type": "text", + "content": "Song Han1,2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 452, + 143, + 523, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 143, + 523, + 157 + ], + "spans": [ + { + "bbox": [ + 452, + 143, + 523, + 157 + ], + "type": "text", + "content": "Yukang Chen" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 427, + 159, + 493, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 159, + 493, + 174 + ], + "spans": [ + { + "bbox": [ + 427, + 159, + 493, + 174 + ], + "type": "text", + "content": "Zhijian Liu" + }, + { + "bbox": [ + 427, + 159, + 493, + 174 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 178, + 455, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 178, + 455, + 193 + ], + "spans": [ + { + "bbox": [ + 344, + 178, + 455, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 344, + 178, + 455, + 193 + ], + "type": "text", + "content": "University of Toronto" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 59, + 230, + 220, + 347 + ], + "blocks": [ + { + "bbox": [ + 97, + 220, + 182, + 228 + ], + "lines": [ + { + "bbox": [ + 97, + 220, + 182, + 228 + ], + "spans": [ + { + "bbox": [ + 97, + 220, + 182, + 228 + ], + "type": "text", + "content": "Multi-Round Performance" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 230, + 220, + 347 + ], + "lines": [ + { + "bbox": [ + 59, + 230, + 220, + 347 + ], + "spans": [ + { + "bbox": [ + 59, + 230, + 220, + 347 + ], + "type": "image", + "image_path": "8fc9719806016f3661bfa5c364611f29c8b92b6c23c53ce8c39fb353eadfdd14.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 224, + 230, + 385, + 347 + ], + "blocks": [ + { + "bbox": [ + 264, + 220, + 346, + 228 + ], + "lines": [ + { + "bbox": [ + 264, + 220, + 346, + 228 + ], + "spans": [ + { + "bbox": [ + 264, + 220, + 346, + 228 + ], + "type": "text", + "content": "Image-Centric Evaluation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 230, + 385, + 347 + ], + "lines": [ + { + "bbox": [ + 224, + 230, + 385, + 347 + ], + "spans": [ + { + "bbox": [ + 224, + 230, + 385, + 347 + ], + "type": "image", + "image_path": "21bf3170cfdc35fef1d36474a6efff25ddb4cae0791f21a9e123baf1fd2c06dd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. Overview of SparseVILA. Compared to existing approaches, SparseVILA exhibits advantages in the following aspects: 1. Multi-Round Performance. SparseVILA leverages a decoupled sparsity framework to achieve state-of-the-art performance at a " + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": " speedup on LLaVA-NeXT [28]. 2. Image-centric evaluation. Leveraging decoupled sparsity, SparseVILA can retain higher context tokens resulting in a " + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": " margin over previous approaches. 3. SparseVILA scales to long-context and long-generation tasks, achieving lossless performance with up to a " + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "3.6 \\times" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": " context, " + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": " generation, and " + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 54, + 358, + 555, + 415 + ], + "type": "text", + "content": " end-to-end speedup." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 388, + 230, + 552, + 346 + ], + "blocks": [ + { + "bbox": [ + 411, + 220, + 538, + 228 + ], + "lines": [ + { + "bbox": [ + 411, + 220, + 538, + 228 + ], + "spans": [ + { + "bbox": [ + 411, + 220, + 538, + 228 + ], + "type": "text", + "content": "Long-Context/Generation Performance" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 230, + 552, + 346 + ], + "lines": [ + { + "bbox": [ + 388, + 230, + 552, + 346 + ], + "spans": [ + { + "bbox": [ + 388, + 230, + 552, + 346 + ], + "type": "image", + "image_path": "394567e0efa79343b86aa01a409fd220861792fd57d1bfadc8ac066c36f4b041.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 152, + 434, + 200, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 434, + 200, + 446 + ], + "spans": [ + { + "bbox": [ + 152, + 434, + 200, + 446 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "content": "Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM, by pruning tokens in the context stage while attending to a sparse subset of visual tokens during the generation phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a " + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "content": " speedup on image benchmarks. This approach leads to " + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "inline_equation", + "content": "+5.9\\%" + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "content": " average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a " + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "inline_equation", + "content": "3.6 \\times" + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 54, + 460, + 297, + 700 + ], + "type": "text", + "content": " speedup in prefetch and decoding, respectively." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 434, + 394, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 394, + 446 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 394, + 446 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 456, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 555, + 553 + ], + "type": "text", + "content": "Vision Language Models (VLMs) have emerged as a state-of-the-art conversational tool enabling users to directly interact with Large Language Models (LLMs) using various visual features, including photographs, documents, and videos [2, 24, 27, 30, 47]. Unfortunately, this added modality comes at the expense of higher latency and memory associated with processing the visual tokens in the LLM. Hence, deploying LLMs efficiently at inference time remains a challenge." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 555, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 556, + 639 + ], + "type": "text", + "content": "Several works have aimed to reduce these associated costs through model pruning on the LLM or vision encoder [32, 41, 64], KV cache compression [13, 26, 29], and most recently, token sparsification [60]. By reducing the amount of computation in the inference pipeline, many of these methods can achieve significant context stage savings as this stage of the network is mainly compute-bound." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "Looking beyond the context, real-world applications often demand extensive generation. Tasks such as image captioning may require a few hundred tokens, meanwhile, video captioning or detailing easily requires more than a few thousand generated tokens. Hence in such applications, it is not sufficient to only focus on context stage optimization – effi" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23784" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "cient implementations with real-world applications should focus on both context and decoding stage optimizations. One such example is multi-turn conversation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "type": "text", + "content": "Multi-turn/round conversation is a practical use case for VLMs, as it is entirely possible for a user to pose multiple questions about a given visual input. In fact, most benchmarks inherently support multi-round conversation: the GQA dataset [15] has more than 90 questions for the same visual input. Despite this, most evaluation benchmarks run single-round evaluation (i.e. repetitive pre-filling), which is not only unrealistic but also inefficient as the context stage would be repeated for each generation round. In real-world scenarios, the visual input could span tens of thousands of context tokens, hence repetitive pre-filling would dramatically slow down the user's interaction with their VLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 407 + ], + "type": "text", + "content": "In this work, we aim to present a unified approach for tackling context and decoding latency in modern VLMs. Latency is a common challenge associated with VLMs due to the sheer amount of visual tokens to be processed. Existing methods for accelerating VLM inference primarily focus on token-wise pruning or merging techniques, with recent approaches leveraging textual priors to reduce visual token complexity in a query-aware manner. In practice, methods that permanently remove visual tokens during the context stage are quite lossy in multi-turn evaluations as visualized in Figure 1. Hence, in this paper, we introduce SparseVILA as a novel approach for accelerating VLM inference, while retaining multi-turn performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 515 + ], + "type": "text", + "content": "Our key insight is a decoupled sparsity framework enabling SparseVILA to migrate sparsity from the context into the decoding stage. Further, SparseVILA leverages query-aware retrieval in the decoding stage, supporting multi-turn conversation as a different subset of context tokens can be retrieved per question. This decoupled approach allows SparseVILA to achieve significant performance improvements in image-centric benchmarks as shown in Figure 1 and outperforms previous methods in long-context/generation scaling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 236, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 236, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 236, + 538 + ], + "type": "text", + "content": "2. Preliminaries and Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 628 + ], + "type": "text", + "content": "Token pruning has proven effective in accelerating inference across a variety of tasks, including image classification [3, 17, 19] and semantic segmentation [5, 43]. With the rise of generative AI, these techniques have been further extended to diffusion models, large language models [18], and vision-language models [14, 55, 60]. We refer the readers to Section A.2 for a more detailed survey of related work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "This paper focuses on token pruning for vision-language models (VLMs). The key idea is that not all visual tokens contribute equally to VLM's final prediction. By identifying and removing less informative tokens, it is possible to significantly reduce the computational cost of VLM inference and thereby improve efficiency. Existing work in this area largely differs in how these tokens are selected. We categorize prior" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "methods into two groups based on their dependence on the input query: (i) query-agnostic approaches, which identify unimportant tokens based solely on visual saliency or redundancy, and (ii) query-aware approaches, which incorporate the semantic relationship between visual and textual inputs to guide pruning. For each category, we highlight representative methods and their limitations, laying the groundwork for the approach proposed in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 177, + 454, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 454, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 454, + 190 + ], + "type": "text", + "content": "2.1. Query-Agnostic Sparsity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 194, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 290 + ], + "type": "text", + "content": "Query-agnostic token pruning methods aim to reduce redundancy or select important visual tokens without relying on the textual input [49]. They prune tokens based solely on the visual context—either within or after the vision encoder. For example, PruMerge [39] clusters and discards less informative tokens using final-layer attention scores, while VisionZip [55] employs a token merging module, similar to ToMe [3], to compress redundant visual information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 291, + 556, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 556, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 556, + 422 + ], + "type": "text", + "content": "This visual-only focus, however, presents key limitations. First, these methods often sacrifice fine-grained visual details, especially under high sparsity, which can degrade performance. More importantly, they cannot adapt token selection based on the input query, leading to suboptimal results when task-relevant information is sparsely distributed. As shown in Figure 1, query-agnostic methods underperform on text-centric and structured visual understanding benchmarks such as DocVQA [35], InfoVQA [36], and ChartQA [34]. By treating all visual tokens uniformly, they risk discarding critical information necessary for accurate reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 431, + 441, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 441, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 441, + 444 + ], + "type": "text", + "content": "2.2. Query-Aware Sparsity" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 557 + ], + "type": "text", + "content": "Query-aware pruning improves visual token selection by explicitly modeling the relationship between textual queries and visual representations. For example, FastV [4] leverages attention maps from early LLM layers as a salience metric to guide token pruning during the prefetch stage. This contrasts with query-agnostic approaches, which perform token reduction prior to any LLM processing. Similarly, SparseVLM [60] utilizes query-to-vision attention scores to eliminate less relevant visual tokens during prefetch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "content": "While query-aware methods are effective at aligning token selection with the current query, they face limitations in multi-turn interactions. Pruning based solely on the initial query can eliminate tokens that are crucial for future queries, resulting in degraded performance. Empirically, such methods show a huge drop in accuracy across multi-turn conversations, often underperforming compared to query-agnostic methods (Figure 1). Re-prefilling at each turn can maintain accuracy, but it introduces significant latency (Section 4.4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To systematically evaluate the performance ceiling of query-aware pruning in multi-turn contexts, we introduce a query-aware oracle, which selects the optimal subset of visual tokens to maximize agreement with the original model's" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23785" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 190 + ], + "type": "image", + "image_path": "dd42b7d99850455a33b4ce37211e8e9c187a67c13fcf2b81088a9d724017d339.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 243 + ], + "type": "text", + "content": "Figure 2. Multi-Round Query-Aware Comparisons with LLaVA-1.5 [27] on the POPE [21] dataset. Without re-profilling the context, the query-aware oracle degrades heavily indicating the inability of query-aware pruning to scale effectively in multi-turn conversation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 262, + 297, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 297, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 297, + 373 + ], + "type": "text", + "content": "response (see Section A.3 for more details). From Figure 2, pruning decisions made in early rounds often result in substantial degradation in later responses, underscoring the long-term cost of early token elimination. By incorporating both the current query and the ground-truth response, the oracle establishes an upper bound on the achievable performance of any query-aware pruning method. These findings reveal a fundamental limitation of query-aware pruning in sustaining performance over multi-turn dialogue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 246, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 246, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 246, + 396 + ], + "type": "text", + "content": "3. SparseVILA: Best of Both Worlds" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 402, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 402, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 402, + 296, + 594 + ], + "type": "text", + "content": "Both query-agnostic and query-aware approaches offer distinct advantages and face inherent limitations. Existing visual token pruning methods typically perform pruning exclusively during the prefetch stage, with the benefits of reduced computation carrying over naturally to the subsequent decode stage. In this paper, we introduce SparseVILA that decouples visual sparsity between the prefetch and decode stages. During the prefetch stage, we apply query-agnostic pruning to eliminate redundant tokens that are unlikely to be useful for any downstream queries. In the decode stage, we employ a query-aware retrieval mechanism to selectively fetch the most relevant entries from the KV cache based on the current query, ensuring that only tokens directly related to the question are used for generating response. This design combines the strengths of both paradigms, achieving efficient token reduction while preserving query relevance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 261, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 261, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 261, + 613 + ], + "type": "text", + "content": "3.1. Prefix Fill Phase: Query-Agnostic Pruning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 701 + ], + "type": "text", + "content": "During the prefetch phase, the VLM processes the system, image, and potentially the first query tokens. To support multi-turn settings, pruning in this stage remains query-agnostic, targeting visual redundancy or salience rather than textual correlation. Since the visual content is only prefilled once, token reduction remains independent of specific queries to effectively support multi-turn conversations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "content": "We estimate token importance using attention scores" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "content": "from the visual encoder, providing an unbiased (query-independent) assessment of token salience, as shown in previous works [1, 39, 55]. This approach is detailed in Algorithm 2 under Section A.4 of the supplementary material. By default, we take the mean-column sum of the attention weights, making our method compatible with architectures like SigLIP [58] that don't produce a " + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "content": " token. For CLIP-style [37] visual encoders that do generate a " + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "inline_equation", + "content": "cls" + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "content": " token, we compute the mean-column-sum relative to that token, leveraging the encoder's training that routes high information density to this specific token." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 211, + 518, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 518, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 518, + 224 + ], + "type": "text", + "content": "3.2. Decode Phase: Query-Aware Retrieval" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 300 + ], + "type": "text", + "content": "During the generation phase, the VLM is typically memory-bound as it implements next-token prediction by computing attention with the existing pre-filled KV cache. By reducing the effective size of the KV cache we can accelerate this generation process while maintaining the necessary visual information to support multi-turn conversation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 556, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 556, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 556, + 432 + ], + "type": "text", + "content": "We identify which visual tokens from the KV cache can be contextually selected for all associated generation steps in answering the current query. During query prefilling at each conversation round, we calculate the mean-columnsum of the attention weights between the query and the remaining visual tokens from the context stage. Using a sparsity threshold, we select a compact subset of the KV cache to participate in all generation steps associated with the query, thereby achieving significant speedup in decoding attention. We detail the retrieval strategy in Algorithm 3 under Section A.4 of the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 556, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 556, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 556, + 527 + ], + "type": "text", + "content": "While prefilling the current round query, we run Algorithm 3 simultaneously to compute the indices of the KV Cache for generation, immediately after each layer's self-attention block. To avoid introducing complexity and overhead into the decoding stage, we create a smaller compact dense KV cache representing only the selected indices for the current generation round. We further discuss the efficiency benefits of this in Section 3.4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 528, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 635 + ], + "type": "text", + "content": "One interesting finding is that the selected tokens can be further divided as Visual Attention Sinks and Visual Retrieval Tokens. The sink tokens appear consistent over different queries, meanwhile, retrieval tokens are directly related to the given query. We discuss this further with qualitative examples in Section 4.5 and Section A.4 in our supplementary material. Similar findings on visual token sinks have been echoed in previous works, such as VisionZip [55] and Visual Attention Redistribution (VAR) [16]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Discussions. Beyond VLM pruning approaches, KV cache compression has been heavily explored in the context of LLMs. In particular, methods such as SnapKV [22] and Quest [42] leverage salience metrics to select a smaller KV cache for decoding, thus similarly accelerating the subsequent decoding stage. In the case of SnapKV, after selecting" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23786" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 558, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 558, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 558, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 558, + 251 + ], + "type": "image", + "image_path": "de8c37865d095ef0a18295c2212995323b6dd8cff930e39b771f9eee0a2b81e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "type": "text", + "content": "Figure 3. Overview of the components in SparseVILA. Query agnostic pruning is run in the context stage using the salience map from the visual encoder to guide token selection. In the decoding stage, the question is used to retrieve a sparse set of critical visual tokens to participate in the generation stage, hence achieving decoding speedup." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 310, + 298, + 381 + ], + "blocks": [ + { + "bbox": [ + 56, + 310, + 298, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 310, + 298, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 298, + 381 + ], + "type": "image", + "image_path": "930cb03e8ed956c583846aa16aacf704f3c80be744de50ea9a8e4e9a3e3b28f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 389, + 296, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 422 + ], + "type": "text", + "content": "Figure 4. Comparing sparsity migration on VideoMME at a fixed latency speedup of " + }, + { + "bbox": [ + 55, + 389, + 296, + 422 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 55, + 389, + 296, + 422 + ], + "type": "text", + "content": ". SparseVILA* uses a higher context sparsity, SparseVILA uses our default configuration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 445, + 298, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 298, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 298, + 601 + ], + "type": "text", + "content": "tokens relevant to the observation window, this method discards the original KV cache, making it ill-suited to multi-turn conversation. Meanwhile, Quest re-runs its selection mechanism for every generated token, significantly increasing the complexity of the decoding stage, while only achieving speedup in long-context cases. We provide detailed comparisons with each of these methods in Section A.4.3 of the supplementary material. SparseVILA introduces a novel combination of query-aware and query-agnostic sparsity, leveraging context and decoding stage improvement beyond any decoding-only or context-only method; hence, the novelty remains in the design of a unified framework for efficient VLM inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 270, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 270, + 624 + ], + "type": "text", + "content": "3.3. Decoupled Prefix-Decode Visual Sparsity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "SparseVILA's decoupled framework enables the migration of computational reduction from the context stage to the generation stage. This approach is based on two key insights. Firstly, the context stage should preserve enough information to support multi-turn conversation as it is only pre-filled once per set of queries. Secondly, the generation stage can leverage query-aware information to guide the pruning process" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 335 + ], + "type": "text", + "content": "and accelerate the decoding, without affecting subsequent turns of conversation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "text", + "content": "As shown in Figure 1, migrating visual sparsity from prefetch stage into the decode stage enables SparseVILA to outperform previous pruning approaches while achieving better speedup. We further investigate this effect by comparing two variants of SparseVILA that both achieve a " + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "text", + "content": " speedup on the VideoMME benchmark [9], as shown in Figure 4. By migrating sparsity from the context stage into the decoding stage (i.e., SparseVILA* → SparseVILA), we observe a " + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "text", + "content": " improvement in overall performance, surpassing the baseline implementation. Moreover, this shift allows the model to retain more tokens during the context stage, resulting in a significant " + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 338, + 556, + 492 + ], + "type": "text", + "content": " gain on image-centric benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 503, + 455, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 455, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 455, + 516 + ], + "type": "text", + "content": "3.4. Efficient Implementation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 616 + ], + "type": "text", + "content": "SparseVILA introduces a two-stage compression approach to distribute sparsity over both the context and decoding stages. This unique approach allows SparseVILA to deliver strong accuracy performance with measurable speedup. Our method, leverages a carefully designed implementation to support modern acceleration techniques. In particular, all parts of the language model leverage the latest FlashAttention2 [6] implementation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "To enable KV-Cache retrieval, we dynamically recompute the QK attention score map once per question at each layer and compress it into a compact, dense KV cache. This design allows us to fully utilize efficient attention implementations across the network. While a naive approach to computing the mean column sum of the attention weights can introduce latency and increase peak memory usage – especially at longer context lengths – we address this with a custom Triton" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23787" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 296, + 205 + ], + "type": "image", + "image_path": "3619d9be254e7d1541af6629e0e877558ca87a8d1f5f7b2bcdb0852c46e49f68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 236 + ], + "type": "text", + "content": "Figure 5. Illustrates the lightweight recomputation of SparseVILA alongside layer-wise FlashAttention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "text", + "content": "kernel that streams attention weight normalization efficiently. This kernel achieves exact results without materializing the full " + }, + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "inline_equation", + "content": "S \\times S" + }, + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "text", + "content": " attention matrix, significantly improving both speed and memory efficiency. Furthermore, due to the high sparsity achieved during the generation phase, our compact KV cache incurs minimal memory overhead (up to 55MB), while reducing peak memory consumption by " + }, + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "inline_equation", + "content": "39.3\\%" + }, + { + "bbox": [ + 54, + 255, + 295, + 351 + ], + "type": "text", + "content": " on long-context video tasks with Qwen2-7B [54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 351, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 295, + 435 + ], + "type": "text", + "content": "Building a compact KV cache avoids any non-contiguous \"on-the-fly indexing\" (sparse access patterns) during the generation stage. The overhead from this additional computation is aggregated with the packing of the KV Cache into a contiguous set as shown in Figure 5, accounting for less than " + }, + { + "bbox": [ + 54, + 351, + 295, + 435 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 54, + 351, + 295, + 435 + ], + "type": "text", + "content": " of the generation runtime for 20, and 100 tokens on Llama2 [45], and Llama3 [7]/Qwen2 [54], respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 105, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 105, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 105, + 478 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 482, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 578 + ], + "type": "text", + "content": "Multi-Turn Evaluation. Many existing benchmarks generate multiple question-answer pairs for each image: e.g., up to 18 in POPE [21]. However, evaluation protocols often remain confined to single-turn settings. In this work, we make use of the inherent multi-turn structure of such datasets to enable more efficient and accurate assessment of VLMs. We will release our multi-turn evaluation framework to support the broader multimodal research community." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": "We organize questions associated with the same image into coherent multi-turn conversations, which allows visual tokens to be prefilled only once. This setup not only reduces computational overhead but also better reflects realistic VLM usage in interactive settings. However, a potential problem is information leakage between turns, where earlier questions can unintentionally reveal answers to later ones. For example, in the GQA dataset [15]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 675, + 284, + 700 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 67, + 675, + 284, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 675, + 284, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 675, + 284, + 687 + ], + "type": "text", + "content": "Q1: \"What is the person in front of the sky doing?\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 687, + 226, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 687, + 226, + 700 + ], + "spans": [ + { + "bbox": [ + 67, + 687, + 226, + 700 + ], + "type": "text", + "content": "Q2: \"What's the person in front of?\"" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Here, Q1 discloses information that effectively answers Q2," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "enabling the model to respond correctly even without relying on the image. To mitigate this, we implement a partial KV cache eviction strategy: after each round, we remove only the KV entries corresponding to the previous question and answer. This preserves the efficiency gains of visual KV cache reuse while preventing unintended context carryover between turns. We provide more details in Section A.5.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 270 + ], + "type": "text", + "content": "Sparsity Ratio. Our sparsity configuration adopts a straightforward approach for both context and decoding, ensuring efficient implementation and cross-model support. Specifically, we set a constant context sparsity before the LLM, with a uniform decoding sparsity across all layers. More granular strategies, such as layer-wise or head-aware sparsity, may further optimize performance but introduce additional complexity and tuning. We prioritize simplicity and generalization, leaving these refinements for future work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 479, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 479, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 479, + 289 + ], + "type": "text", + "content": "4.2. Results on Image Benchmarks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 402 + ], + "type": "text", + "content": "In Table 1, we evaluate our SparseVILA against a diverse set of query-aware and query-agnostic baselines across eight vision-language benchmarks: ChartQA [34], DocVQA [35], GQA [15], InfographicVQA [36], MME [8], POPE [21], ScienceQA [31], and TextVQA [40]. All results in this section are conducted on LLaVA-NeXT-7B [61]. To further evaluate the generality of SparseVILA, we extend our analysis to additional VLMs, including LLaVA-1.5 [27], VILA-1.5 [24], Qwen2-VL [46], and NVILA [30], in Table A5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "content": "We compare our SparseVILA with existing query-aware pruning methods, including FastV [4], SparseVLM [60] and PDrop [52]. Our results show that at a given speedup, SparseVILA outperforms these prior approaches in multi-turn evaluation by migrating sparsity into the decode phase. Notably, SparseVILA achieves a " + }, + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "content": " accuracy improvement on DocVQA, highlighting the effectiveness of our decoupled prefetch-decode sparsity approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "text", + "content": "We further evaluate SparseVILA against existing query-agnostic methods, including PruMerge [39], HIRED [1], and VisionZip [55]. At an end-to-end speedup of " + }, + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "inline_equation", + "content": "1.44 \\times" + }, + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "text", + "content": ", SparseVILA achieves at least a " + }, + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 617 + ], + "type": "text", + "content": " improvement on DocVQA. These results further demonstrate that, under a fixed speedup target, SparseVILA's decoupled sparsity approach consistently outperforms prior state-of-the-art pruning methods. In Section A.6.1.1, we also show that applying SparseVILA to these existing query-agnostic sparsity methods can further enhance their performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 477, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 477, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 477, + 635 + ], + "type": "text", + "content": "4.3. Results on Video Benchmarks" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "In Table 2, we demonstrate that SparseVILA consistently outperforms existing approaches on long-context tasks. Additionally, query-aware works such as FastV [4] and SparseVLM [60] were unable to scale effectively to long context (64K) as they require access to the full visual context and query at the same time to compute their ranking scores, hence" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23788" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 145, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "type": "text", + "content": "Table 1. Results on image benchmarks. We provide query-aware comparisons with SparseVLM [60] and FastV [4] and query-agnostic comparisons with LLaVA-PruMerge [39], HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.2. At a given latency, we show significant improvements over state-of-the-art approaches; hence, SparseVILA optimizes the latency-to-performance tradeoff. A generation length of 250 tokens is used to approximate the sequence length observed in image-c captioning tasks. TextVQA [40], DocVQA (validation set) [35], InfoVQA (validation set) [36], ChartQA [34], ScienceQA-Img [31], POPE [21], and GQA [15]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 145, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 56, + 145, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 553, + 294 + ], + "type": "table", + "html": "
Sparsity (%)SpeedupCQADVQAGQAIVQAMMEPOPESQATVQA
PrefillDecodePrefillDecodeOverall
LLaVA-NeXT-7B-----50.774.464.237.71846.185.970.361.2
+ FastV8303.26×1.24×1.37×37.354.460.631.81775.282.867.557.1
+ SparseVLM8303.26×1.24×1.37×37.754.460.831.41780.784.067.558.1
+ PDrop5301.89×1.16×1.24×41.544.862.926.51809.286.169.157.0
+ SparseVILA67502.59×1.27×1.38×41.860.863.433.31818.485.968.060.2
+ PruMerge8304.34×1.27×1.43×26.936.961.128.51753.283.666.754.6
+ HIRED8304.34×1.27×1.43×29.741.662.428.21777.181.767.755.9
+ VisionZip8304.34×1.27×1.43×34.151.261.230.61760.383.767.458.7
+ SparseVILA67902.59×1.33×1.44×39.954.463.430.81827.886.168.058.9
", + "image_path": "69161a9615369ec0bd56886c5cd6ed72b392eb0d20fd6a935950beafd45c5658.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 357, + 525, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": "Table 2. Results on video benchmarks. We provide query-agnostic comparisons with HIRED [1], and VisionZip [55]. We report accuracy and efficiency per the setting described in Section 4.3." + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " FastV [4], SparseVLM [60], and PDrop [52] do not scale to 64K context length, hence resulting in " + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "inline_equation", + "content": "OOM -" + }, + { + "bbox": [ + 55, + 304, + 555, + 350 + ], + "type": "text", + "content": " we include shorter context comparison in Section A.6.2 of our supplementary material. A generation length of 2000 tokens is used to approximate the sequence length observed in video captioning or reasoning tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 357, + 525, + 446 + ], + "lines": [ + { + "bbox": [ + 83, + 357, + 525, + 446 + ], + "spans": [ + { + "bbox": [ + 83, + 357, + 525, + 446 + ], + "type": "table", + "html": "
Sparsity (%)SpeedupMLVUVideo-MME (w/o subtitle)
PrefillDecodePrefillDecodeOverallm-avgShortMediumLongOverall
LongVILA-7B-----72.468.958.152.059.7
+ HIRED8604.9×1.5×2.1×70.268.054.649.657.4
+ VisionZip8604.9×1.5×2.1×73.166.655.150.957.5
+ SparseVILA75943.6×1.7×2.2×74.470.658.651.760.3
", + "image_path": "45ca9b4d331c7dafc79599ef5b834929c6a697ef0cf9631ceaaf4020560af15d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 517, + 297, + 568 + ], + "blocks": [ + { + "bbox": [ + 55, + 463, + 297, + 509 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 297, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 297, + 509 + ], + "type": "text", + "content": "Table 3. Long-Context/Generation Performance on VideoChatGPT [33]. Scores are computed using the GPT4-mini assisted evaluation. SparseVILA offers lossless generation performance at a " + }, + { + "bbox": [ + 55, + 463, + 297, + 509 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 55, + 463, + 297, + 509 + ], + "type": "text", + "content": " end-to-end speedup." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 517, + 297, + 568 + ], + "lines": [ + { + "bbox": [ + 56, + 517, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 297, + 568 + ], + "type": "table", + "html": "
CIDOCUTUCMeanSpeedup
LongVILA-7B2.332.282.871.742.412.331.0×
+ SparseVILA2.362.342.921.882.562.412.2×
", + "image_path": "bfb0c836d8b084c6ea15a06d487208446dbf1817bedcbc9c7df0f3023949a573.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "text", + "content": "preventing chunk pre-filling and adding significant memory overhead to their approach. This consequently limits their ability to scale. In contrast, SparseVILA efficiently scales to 64K context (LongVILA 256 Frames) by decoupling our sparsity in the context and decoding stage, achieving " + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "inline_equation", + "content": "3.6 \\times" + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "text", + "content": " prefilling speedup, " + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "inline_equation", + "content": "1.7 \\times" + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "text", + "content": " decoding speedup, and " + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 55, + 591, + 295, + 687 + ], + "type": "text", + "content": " end-to-end speedup. We include additional results on 8K context in Section A.6.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "One of the key novelties in SparseVILA is the decoupling and migration of sparsity into the decoding stage. As" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 465, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 561 + ], + "type": "text", + "content": "previously shown, this has a positive effect on image and long-context benchmarks. In this section, we want to validate that SparseVILA's heavy generation sparsity does not affect the model's ability to generate responses by testing on free-form text generation tasks. Our approach delivers significant efficiency gains, offering up to " + }, + { + "bbox": [ + 313, + 465, + 555, + 561 + ], + "type": "inline_equation", + "content": "2.2 \\times" + }, + { + "bbox": [ + 313, + 465, + 555, + 561 + ], + "type": "text", + "content": " E2E speedup over the vanilla model while maintaining performance on the Video-ChatGPT [33] benchmark as shown in Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "In addition to VQA and captioning tasks, we also evaluate the long-context retrieval performance with Visual Needle in a Haystack (V-NIAH) benchmark [53, 59]. The original V-NIAH only supports single-turn, so we extended this benchmark to a multi-turn setting. See Section A.7 of the supplementary material for more details. We compare the long-context retrieval capacity of our method with the vanilla implementation, SparseVLM [60], and FastV [4] on LongVILA [53] in Figure 6. Notably, SparseVLM and FastV do not scale beyond 32 Frames (approx. 8K context lengths) as they lack support for chunkedprefilling and incur a memory overhead in computing their salience metric. Sparse-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23789" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 147 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 147 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 147 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 147 + ], + "type": "image", + "image_path": "cc2bd110ef826cf0e7017e42b29f665117d37643ad70ad228b9f0336a952876b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 186, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 555, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 555, + 185 + ], + "type": "text", + "content": "Figure 6. Comparing the multi-turn performance of the vanilla implementation with SparseVLM [60], FastV [4] and SparseVILA on LongVILA-Qwen2 [53] for V-NIAH. FastV and SparseVLM reach OOM after 30 Frames (8K context) as they do not inherently support chunk prefilling. We include the full scaling up to 500 Frames (128K context length) in Section A.7." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 186, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 555, + 407 + ], + "type": "image", + "image_path": "b80bb9a085e40c7ea250e2b9187ef4b334bfac1f627160d17862154e35216d05.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 449 + ], + "type": "text", + "content": "Figure 7. Decoding attention kernel speed for video tasks and V-NIAH on LongVILA-7B [53] and image tasks on LLaVA-NeXT-7B [28]. We further explore the multi-turn scaling of the baseline, a query-aware approach with re-pre-filling and SparseVILA over conversation rounds 0, 1, and 3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "text", + "content": "VILA maintains comparable performance with the baseline. In contrast, query-aware context stage pruning strategies begin to degrade at only an 8K context length. In reference, to Figure 6, the vanilla implementation remains effective up to a context length of 89.6K, handling contexts up to " + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "text", + "content": " as long. Meanwhile, SparseVLM [60] and FastV [4] exhibit degradation, indicating that these pruning frameworks significantly impair the model's capacity for long context retrieval in multi-turn evaluation. We find that our approach performs similarly to the vanilla baseline, indicating that SparseVILA can retain long context retrieval performance. We can attribute the retained performance to the query-aware nature of our decoding sparsity. We leverage a " + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "text", + "content": " decoding sparsity, translating to a " + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 54, + 470, + 297, + 651 + ], + "type": "text", + "content": " decoding speedup. Additional details are provided in Section A.7 of the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 167, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 167, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 167, + 672 + ], + "type": "text", + "content": "4.4. Efficiency Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In this section, we provide details of our optimized inference pipeline and additional results. Efficiency metrics are reported relative to the optimized baseline, adding additional" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 471, + 460, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 460, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 460, + 482 + ], + "type": "text", + "content": "benefit for efficient VLM inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 490, + 556, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 556, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 556, + 621 + ], + "type": "text", + "content": "Inference Setting AWQ [25] is a widely adopted technique for reducing decoding latency by significantly decreasing the latency of linear layers. We developed an optimized inference pipeline leveraging AWQ [25] based on TinyChat. In the case of Llama2-7B [45] with a 32K context length, our dense configuration achieves a " + }, + { + "bbox": [ + 313, + 490, + 556, + 621 + ], + "type": "inline_equation", + "content": "1.55\\mathrm{x}" + }, + { + "bbox": [ + 313, + 490, + 556, + 621 + ], + "type": "text", + "content": " speedup over a finely-tuned FP16 pipeline, with pre-allocated KV cache across the entire benchmark sequence, effectively eliminating the additional overhead associated with dynamic memory allocation. All latency measurements are obtained on an NVIDIA A6000 GPU with a batch size of 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "text", + "content": "General Efficiency Savings According to Figure 7, we can gain up to " + }, + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "inline_equation", + "content": "6.3 \\times" + }, + { + "bbox": [ + 313, + 629, + 557, + 714 + ], + "type": "text", + "content": " speedup in decoding attention kernel time on long-context video and V-NIAH with LongVILA [53]. Further, our multi-turn scaling exhibits compounded improvement, achieving a 1:1.7 ratio in conversation turn speed between the vanilla and SparseVILA. Beyond latency measurements, SparseVILA can cut down" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23790" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 85, + 148, + 174 + ], + "blocks": [ + { + "bbox": [ + 58, + 85, + 148, + 174 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 148, + 174 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 148, + 174 + ], + "type": "image", + "image_path": "b8ff59445fda0cb6fe4361fe8174c8cfb0453b95cc6b66081db64b9a4a879ca6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 177, + 118, + 185 + ], + "lines": [ + { + "bbox": [ + 87, + 177, + 118, + 185 + ], + "spans": [ + { + "bbox": [ + 87, + 177, + 118, + 185 + ], + "type": "text", + "content": "\\\\n" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 191, + 179, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 191, + 179, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 191, + 179, + 201 + ], + "type": "text", + "content": "Prompt 1: Tell me about the car in the image." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 85, + 242, + 175 + ], + "blocks": [ + { + "bbox": [ + 151, + 85, + 242, + 175 + ], + "lines": [ + { + "bbox": [ + 151, + 85, + 242, + 175 + ], + "spans": [ + { + "bbox": [ + 151, + 85, + 242, + 175 + ], + "type": "image", + "image_path": "9972c6c1e7b6405fbde4fc285a7fde35cbfecf1724bf2c092a0c4172538318d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 177, + 220, + 185 + ], + "lines": [ + { + "bbox": [ + 172, + 177, + 220, + 185 + ], + "spans": [ + { + "bbox": [ + 172, + 177, + 220, + 185 + ], + "type": "text", + "content": "Context Pruning" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 202, + 268, + 217 + ], + "lines": [ + { + "bbox": [ + 60, + 202, + 268, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 202, + 268, + 217 + ], + "type": "text", + "content": "SparseVILA: \"The car in the image is a blue sports car, possibly a Mazda RX-8, driving down the road." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 245, + 84, + 335, + 175 + ], + "blocks": [ + { + "bbox": [ + 245, + 84, + 335, + 175 + ], + "lines": [ + { + "bbox": [ + 245, + 84, + 335, + 175 + ], + "spans": [ + { + "bbox": [ + 245, + 84, + 335, + 175 + ], + "type": "image", + "image_path": "e04d8c252e20bef311bbf75e04b7226c9de79c0ceaf96586a8c3cac1adc79829.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 259, + 177, + 322, + 185 + ], + "lines": [ + { + "bbox": [ + 259, + 177, + 322, + 185 + ], + "spans": [ + { + "bbox": [ + 259, + 177, + 322, + 185 + ], + "type": "text", + "content": "Visual Attention Sinks" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 300, + 191, + 436, + 200 + ], + "lines": [ + { + "bbox": [ + 300, + 191, + 436, + 200 + ], + "spans": [ + { + "bbox": [ + 300, + 191, + 436, + 200 + ], + "type": "text", + "content": "Prompt 2: Tell me about the deer crossing the road" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 237, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 259 + ], + "type": "text", + "content": "Figure 8. Visualization of selected tokens from the KV cache for various queries in the multi-round setting. The heatmap illustrates the frequency of token selection from the KV Cache for the particular query of interest. We conduct said visualization on LLaVA-1.5 [27]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 357, + 83, + 448, + 174 + ], + "blocks": [ + { + "bbox": [ + 339, + 126, + 353, + 141 + ], + "lines": [ + { + "bbox": [ + 339, + 126, + 353, + 141 + ], + "spans": [ + { + "bbox": [ + 339, + 126, + 353, + 141 + ], + "type": "text", + "content": "+" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 83, + 448, + 174 + ], + "lines": [ + { + "bbox": [ + 357, + 83, + 448, + 174 + ], + "spans": [ + { + "bbox": [ + 357, + 83, + 448, + 174 + ], + "type": "image", + "image_path": "98c551e5cf89a5b3e0f7f4843c13cb45699456a2804a3f8858379569dfed9e0a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 176, + 429, + 184 + ], + "lines": [ + { + "bbox": [ + 375, + 176, + 429, + 184 + ], + "spans": [ + { + "bbox": [ + 375, + 176, + 429, + 184 + ], + "type": "text", + "content": "Prompt 1 Retrieval" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 202, + 533, + 223 + ], + "lines": [ + { + "bbox": [ + 301, + 202, + 533, + 223 + ], + "spans": [ + { + "bbox": [ + 301, + 202, + 533, + 223 + ], + "type": "text", + "content": "SparseVILA: \"In the image, a deer is crossing a road in front of a car. The deer is walking across the street, and the car is waiting patiently behind it ... this moment highlights the importance of being cautious and aware of wildlife while driving...\"" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 459, + 83, + 551, + 174 + ], + "blocks": [ + { + "bbox": [ + 459, + 83, + 551, + 174 + ], + "lines": [ + { + "bbox": [ + 459, + 83, + 551, + 174 + ], + "spans": [ + { + "bbox": [ + 459, + 83, + 551, + 174 + ], + "type": "image", + "image_path": "db1d650e701f4cf46fb7fc5aaa1582f2400b767e45e35cd1a6f84d8d1af11685.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 176, + 533, + 183 + ], + "lines": [ + { + "bbox": [ + 478, + 176, + 533, + 183 + ], + "spans": [ + { + "bbox": [ + 478, + 176, + 533, + 183 + ], + "type": "text", + "content": "Prompt 2 Retrieval" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "text", + "content": "the peak memory cost by " + }, + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "inline_equation", + "content": "39.3\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "text", + "content": " for Qwen2-7B [54] on long context video tasks. Finally, we can decrease linear flops in the context stage by up to " + }, + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "inline_equation", + "content": "87.6\\%" + }, + { + "bbox": [ + 55, + 280, + 295, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 347, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 468 + ], + "type": "text", + "content": "Multi-Round Efficiency In Figure 7 we plot the multiround scaling for the baseline, SparseVLM [60] with context reprefilling, and SparseVILA. Over multiple rounds, our method retains a consistent context latency enabled through our context sparsity approach. Further, the generation latency scales faster than both query-aware and vanilla implementations, offering larger savings at increased rounds. Notably, SparseVLM [60] will scale past the original latency cost of the vanilla baseline after round 3 on Llama2 [45] indicating the inefficiency of that approach for multi-turn settings." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 488, + 173, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 173, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 173, + 501 + ], + "type": "text", + "content": "4.5. Qualitative Analysis" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "We provide a qualitative visualization of SparseVILA's token selection strategy in Figure 8 and Figure A5. We apply a " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " context and " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " generation sparsity for this visualization on LLaVA-1.5 [27]. During the decoding stage, tokens are selected based on their alignment with query information by thresholding attention weights. Since different tokens can be selected at each layer, the heatmap visualizes the frequency by which a token was selected over all LLM layers. Certain tokens were selected across most layers, and consistent over different queries, we coin these, visual sink tokens aligning with previous definitions [51]. The remainder of tokens selected are considered query-dependent, which we term as retrieval tokens. This division is quite intuitive in the qualitative samples shown in Figure A5. Our decoupled framework retains both sink and retrieval tokens in the KV cache. Further analysis and visualizations are provided in Section A.8 of the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 279, + 388, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 279, + 388, + 292 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 388, + 292 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "text", + "content": "We introduce SparseVILA, a novel query-aware token retrieval method designed to optimize inference in Vision-Language Models (VLMs). Through a decoupled sparsity framework SparseVILA leverages query-agnostic context stage pruning with query-aware generation stage retrieval. By migrating sparsity into the decoding stage, SparseVILA is able to achieve state-of-the-art performance on multi-turn benchmarks, with a significant " + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "inline_equation", + "content": "+5.9\\%" + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "text", + "content": " improvement margin on image-centric tasks. By focusing on both context and decoding stage optimization, SparseVILA can achieve " + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "inline_equation", + "content": "3.6\\times" + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "text", + "content": " context stage speedup, with " + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "inline_equation", + "content": "1.7\\times" + }, + { + "bbox": [ + 313, + 299, + 556, + 508 + ], + "type": "text", + "content": " decoding stage acceleration in long-context video tasks. Further SparseVILA is able to retain lossless performance in both long-context/generation tasks and needle-in-a-haystack retrieval. SparseVILA introduces a new paradigm in prefetch decoding sparsity decoupling that maintains performance and improves efficiency. This approach enables VLMs to efficiently handle longer contexts, accommodating hundreds of images and frames." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 528, + 373, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 528, + 373, + 540 + ], + "spans": [ + { + "bbox": [ + 315, + 528, + 373, + 540 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 319, + 548, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 319, + 548, + 556, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 548, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 319, + 548, + 556, + 602 + ], + "type": "text", + "content": "[1] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models in resource-constrained environments, 2024. 3, 5, 6, 1, 4, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 319, + 602, + 555, + 657 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 319, + 658, + 555, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 658, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 319, + 658, + 555, + 700 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2, 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 319, + 702, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 702, + 555, + 714 + ], + "type": "text", + "content": "[4] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 118, + 70, + 180, + 78 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 70, + 180, + 78 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 180, + 78 + ], + "type": "text", + "content": "Context Stage Sparsity" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 369, + 71, + 435, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 71, + 435, + 79 + ], + "spans": [ + { + "bbox": [ + 369, + 71, + 435, + 79 + ], + "type": "text", + "content": "Decoding Stage Sparsity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23791" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models, 2024. 2, 5, 6, 7, 1, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "text", + "content": "[5] Xuanyao Chen, Zhijian Liu, Haotian Tang, Li Yi, Hang Zhao, and Song Han. Sparsevit: Revisiting activation sparsity for efficient high-resolution vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2061-2070, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 162, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 296, + 183 + ], + "type": "text", + "content": "[6] Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. 4, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 237 + ], + "type": "text", + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "text", + "content": "[8] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 294, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 296, + 348 + ], + "type": "text", + "content": "[9] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 350, + 296, + 403 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Yi-Fan Zhang, Shukang Yin, Bo Li, Xinyu Fang, Sirui Zhao, Haodong Duan, Xing Sun, Ziwei Liu, Liang Wang, et al. Mme-survey: A comprehensive survey on evaluation of multimodal llms. arXiv preprint arXiv:2411.15296, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 404, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 404, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 57, + 404, + 296, + 446 + ], + "type": "text", + "content": "[11] Qichen Fu, Minsik Cho, Thomas Merth, Sachin Mehta, Mohammad Rastegari, and Mahyar Najibi. Lazyllm: Dynamic token pruning for efficient long context llm inference, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "text", + "content": "[12] Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 773-783, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "type": "text", + "content": "[13] Coleman Hooper, Sehoon Kim, Hiva Mohammadzadeh, Michael W Mahoney, Yakun Sophia Shao, Kurt Keutzer, and Amir Gholami. Kvquant: Towards 10 million context length llm inference with kv cache quantization. arXiv preprint arXiv:2401.18079, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "text", + "content": "[14] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 657 + ], + "type": "text", + "content": "[15] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 5, 6, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 689 + ], + "type": "text", + "content": "[16] Seil Kang, Jinyeong Kim, Junhyeok Kim, and Seong Jae Hwang. See what you are told: Visual attention sink in large multimodal models, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[17] Samir Khaki and Konstantinos N Plataniotis. The need for speed: Pruning transformers with one recipe. In *The Twelfth" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 95 + ], + "type": "text", + "content": "International Conference on Learning Representations, 2024. 2, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "text", + "content": "[18] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "text", + "content": "[19] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Mengshu Sun, Wei Niu, Xuan Shen, Geng Yuan, Bin Ren, Minghai Qin, et al. Spvit: Enabling faster vision transformers via soft token pruning. arXiv preprint arXiv:2112.13890, 2021. 2, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "text", + "content": "[20] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 231, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 554, + 263 + ], + "type": "text", + "content": "[21] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models, 2023. 3, 5, 6, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 555, + 308 + ], + "type": "text", + "content": "[22] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: LIm knows what you are looking for before generation, 2024. 3, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 555, + 354 + ], + "type": "text", + "content": "[23] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. arXiv preprint arXiv:2202.07800, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "text", + "content": "[24] Ji Lin, Hongxu Yin, Wei Ping, Yao Lu, Pavlo Molchanov, Andrew Tao, Huizi Mao, Jan Kautz, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models, 2023. 1, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 443 + ], + "type": "text", + "content": "[25] Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for lIm compression and acceleration, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "text", + "content": "[26] Yujun Lin, Haotian Tang, Shang Yang, Zhekai Zhang, Guangxuan Xiao, Chuang Gan, and Song Han. Qserve: W4a8kv4 quantization and system co-design for efficient llm serving. arXiv preprint arXiv:2405.04532, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 490, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 511 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 3, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.1,7,4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 547, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 547, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 547, + 555, + 590 + ], + "type": "text", + "content": "[29] Zirui Liu, Jiayi Yuan, Hongye Jin, Shaochen Zhong, Zhaozhuo Xu, Vladimir Braverman, Beidi Chen, and Xia Hu. Kivi: A tuning-free asymmetric 2bit quantization for kv cache. arXiv preprint arXiv:2402.02750, 2024.1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 592, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 678 + ], + "type": "text", + "content": "[30] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[31] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23792" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 105 + ], + "type": "text", + "content": "thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "text", + "content": "[32] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. Advances in neural information processing systems, 36:21702-21720, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[33] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024), 2024. 6, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "text", + "content": "[34] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 5, 6, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 293 + ], + "type": "text", + "content": "[35] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2, 5, 6, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 347 + ], + "type": "text", + "content": "[36] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2, 5, 6, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 458 + ], + "type": "text", + "content": "[38] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 502 + ], + "type": "text", + "content": "[39] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 3, 5, 6, 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 558 + ], + "type": "text", + "content": "[40] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 591 + ], + "type": "text", + "content": "[41] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. arXiv preprint arXiv:2306.11695, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "text", + "content": "[42] Jiaming Tang, Yilong Zhao, Kan Zhu, Guangxuan Xiao, Baris Kasikci, and Song Han. Quest: Query-aware sparsity for efficient long-context llm inference, 2024. 3, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "text", + "content": "[43] Quan Tang, Bowen Zhang, Jiajun Liu, Fagui Liu, and Yifan Liu. Dynamic token pruning in plain vision transformers for semantic segmentation, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[44] Yehui Tang, Kai Han, Yunhe Wang, Chang Xu, Jianyuan Guo, Chao Xu, and Dacheng Tao. Patch slimming for efficient vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12165-12174, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 7, 8, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 205 + ], + "type": "text", + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 274 + ], + "type": "text", + "content": "[47] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-v1: Enhancing vision-language model's perception of the world at any resolution, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 275, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 555, + 330 + ], + "type": "text", + "content": "[48] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2092-2101, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "type": "text", + "content": "[49] Zichen Wen, Yifeng Gao, Weijia Li, Conghui He, and Linfeng Zhang. Token pruning in multimodal large language models: Are we solving the right problem? arXiv preprint arXiv:2502.11501, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 555, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 421 + ], + "type": "text", + "content": "[50] Guangxuan Xiao, Jiaming Tang, Jingwei Zuo, Junxian Guo, Shang Yang, Haotian Tang, Yao Fu, and Song Han. Duoattention: Efficient long-context llm inference with retrieval and streaming heads, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 456 + ], + "type": "text", + "content": "[51] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks, 2024. 8, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 457, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 555, + 513 + ], + "type": "text", + "content": "[52] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 555, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 555, + 580 + ], + "type": "text", + "content": "[53] Fuzhao Xue, Yukang Chen, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, Yihui He, Hongxu Yin, Pavlo Molchanov, Jan Kautz, Linxi Fan, Yuke Zhu, Yao Lu, and Song Han. Longvila: Scaling long-context visual language models for long videos. 2024. 6, 7, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 713 + ], + "type": "text", + "content": "[54] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23793" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 565 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report, 2024. 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[55] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2, 3, 5, 6, 1, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[56] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, and Yansong Tang. Voco-llama: Towards vision compression with large language models, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "text", + "content": "[59] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024.6, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 386 + ], + "type": "text", + "content": "[60] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference arXiv preprint arXiv:2410.04417, 2024. 1, 2, 5, 6, 7, 8, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[61] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 486 + ], + "type": "text", + "content": "[62] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, Zhangyang Wang, and Beidi Chen. " + }, + { + "bbox": [ + 56, + 432, + 294, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{H}_{20}" + }, + { + "bbox": [ + 56, + 432, + 294, + 486 + ], + "type": "text", + "content": ": Heavy-hitter oracle for efficient generative inference of large language models, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "text", + "content": "[63] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 543, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 565 + ], + "type": "text", + "content": "[64] Mingjian Zhu, Yehui Tang, and Kai Han. Vision transformer pruning. arXiv preprint arXiv:2104.08500, 2021. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23794" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_content_list.json b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..315c079b34e9fbff6391d32328a8a081a945ee5a --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_content_list.json @@ -0,0 +1,1377 @@ +[ + { + "type": "text", + "text": "Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation", + "text_level": 1, + "bbox": [ + 147, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nairouz Mrabah Nicolas Richet Ismail Ben Ayed Eric Granger LIVIA, ILLS, Department of Systems Engineering, ETS Montreal, Canada", + "bbox": [ + 200, + 181, + 795, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mrabah.nairouz@livia.etsmt1.ca, nicolas.richet.l@ens.etsmt1.ca, ismail.benayed@etsmt1.ca, eric.granger@etsmt1.ca", + "bbox": [ + 217, + 219, + 776, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adapting vision-language models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for local sparsity and global density, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for local randomness and global importance, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead. Code is available at: https://github.com/nairouz/SO.", + "bbox": [ + 89, + 316, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The performance of deep learning (DL) models is strongly influenced by their size, with larger architectures yielding superior results. For instance, the ViT-H/14 backbone outperforms ViT-B/32 by a significant margin across multiple benchmarks [6]. Despite recent architectural advancements, the best models remain highly data-dependent and require large-scale labeled datasets to achieve reliable generalization. While techniques such as data augmentation and regularization help mitigate overfitting in low-data scenarios, they are insufficient to fully resolve it [1]. In contrast, the human brain has more than 80 billion neurons. Each neuron forms, on average, only about 7,000 synaptic connec", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions, which results in an estimated total of 100 to 500 trillion synapses. Despite its tremendous capacity, the brain exhibits a remarkable ability to grasp new concepts from minimal exposure. This one-shot or few-shot learning capability might be partly attributed to the effective use of sparsity [17]. By selectively activating a fraction of connections at any given time, the brain optimizes both energy consumption and computational efficiency.", + "bbox": [ + 511, + 287, + 906, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among the prominent DL models, vision-language models (VLMs) learn visual concepts from natural language descriptions. In particular, the contrastive language-image pretraining (CLIP) [33] model is pretrained on large-scale datasets containing hundreds of millions of image-text pairs to align visual and textual embeddings within a shared semantic space. After training, VLMs show strong zero-shot capabilities across diverse visual recognition tasks [33]. Moreover, fine-tuning VLMs yields state-of-the-art performance in various domain adaptation scenarios [27, 39]. Despite their zero-shot capabilities, adapting VLMs to new tasks with small labeled datasets remains challenging due to pronounced overfitting and catastrophic forgetting problems. Furthermore, training such models requires vast resources, including significant GPU memory for storing millions of parameters, gradients, activations, and optimizer states. These challenges underscore the need for cost-effective optimization strategies that address overfitting.", + "bbox": [ + 511, + 411, + 908, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the limitations of full VLM fine-tuning in few-shot learning, state-of-the-art methods leverage parameter-efficient training (PET) [15, 16, 21]. This strategy consists of optimizing a small number of parameters compared to the full set of weights. Additive PET methods freeze the initial weights and attach trainable and unmergeable parameters to the model. Within this category, we find the adapter [11] and the prompt tuning variants [49]. The prevailing methods in few-shot VLMs have mainly focused on this category. Adjustable weights can be incorporated at various levels, either by embedding them directly into the input space as visual [18] or textual prompts [48], or by integrating adapters throughout the network architecture [50].", + "bbox": [ + 511, + 686, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate overfitting in few-shot adaptation and pre", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3143", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/998282f275fb5f80ced8c8975a2fe329bee4b2d3984092db3a153167a1ba6eef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 92, + 310, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c1e31495eeba8268f0f1be5f1d1e910acc283015936305e56824803b6132b55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 92, + 509, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd7189c5fb0e3252bf1035cbaf77742f34f9dbda8a8c2b54be96d5270f38c8c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 92, + 705, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c1248125ba560a9c08a7d18e181648eeedfd096700b2004eef961c29d8fbca5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 92, + 903, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bea01f790e9e1b4ad400c75b69eb5d83d0a3f1dd4d6b7a96ecea4496cafca844.jpg", + "image_caption": [ + "Figure 1. Test-set accuracy over training iterations for LoRA (with rank fixed at 3) and SO (with a sparsity ratio equal to $99.95\\%$ ) in a 1-shot setting on four datasets (Pets, Flowers, Caltech, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. Training stops when the loss falls below 0.01 or after 2000 iterations." + ], + "image_footnote": [], + "bbox": [ + 99, + 229, + 310, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/452e6c558c4b75c38d72e1227c9494fad1e750d7f5fc3002ebda03a024c0f17f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 233, + 509, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/53b4e9a77f4fb9bdaed33f0d375b7d53cc33afb4eec866f4dc11d629b952e4ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 233, + 705, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/af8f89d3882d67bcc788316703ca6fbcf88659c34eb898ce2b39a34441366178.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 229, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5218af9e9e4c7ca5c47bbbc554e75de01d4acddcebc631a7b0c8222cd4fd0693.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 424, + 310, + 555 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e559f2557e3b49eee358e923f88b9e8edb1b1105f44fdef72fc9d2db39947dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 425, + 509, + 555 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c9f5f5abccb0862c1bbd83c7dc801e35441702214c27b372be0898ad6c7477e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 425, + 705, + 555 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5b049a48fbd3758252415ed14fafab3997364ea8e5417b1c1f8be1da00f4a77f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 425, + 903, + 555 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/839bb7ad23c06b203f23082e233bc9504eff49688ff54685c8431df5df178874.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 561, + 310, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d0bb11896288fd91b0405849c9010a32bd9c8c63ff7e03dc8acfc516c302a54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 561, + 509, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22368532aeabaa75da0760d7dbe426cd4077df6b388fc62b89285f71013d898b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 561, + 705, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/653ac7451982321245105d471c82041e6c9935cff9bc915820edaf67285a88c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 559, + 903, + 676 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c37137fec9706b698d853b63f8e8b92887fcc698994fd0b701f28dd1e44e84c.jpg", + "image_caption": [ + "Figure 2. Test-set accuracy of LoRA in a 1-shot setting on three datasets (DTD, Oxford Pets, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. We vary the LoRA rank from 2 to 5 and train for at most 2000 iterations or until the loss drops below 0.01. Each subfigure shows test accuracy over training iterations for a specific rank." + ], + "image_footnote": [], + "bbox": [ + 99, + 681, + 310, + 797 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b0eec707c9caa65024f5b46af69bd6150e07d76842f0b0555575c617c03e88ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 681, + 509, + 797 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/65974d4476f4e198220419948cc9bb79554e39d06f6931ca6456c4f6b91932b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 681, + 705, + 797 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd46951795c2b021fcd660832629a4ce2a989c4f96164f8b14d021a989701f10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 681, + 903, + 797 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3144", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "serve the prior knowledge of VLMs, adapter-based methods typically employ three techniques. First, these small trainable modules are only integrated into the last layer of VLMs. Second, they use regularization techniques to penalize the divergence between the initial zero-shot class prototypes and the learnable class prototypes. Finally, they initialize the learnable class prototypes into a reliable region using the zero-shot prototypes. As a result, these methods have shown better results than the prompt tuning approaches [44], which often introduce computational overhead [45]. Despite these efforts, adapter-based methods are still prone to considerable overfitting [34]. Their success depends heavily on carefully tuning hyperparameters whose optimal values vary significantly across different tasks [34]. This sensitivity leads to performance degradation that undermines the generalizability of these approaches. It is worth noting that tuning these hyperparameters requires extensive grid searching with a validation set, which is not available in real-world few-shot scenarios.", + "bbox": [ + 89, + 90, + 480, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, low-rank reparametrization [16, 20, 26] has gained popularity for domain adaptation. These approaches reparameterize the initial weights with supplementary trainable and mergeable low-rank matrices. Low-rank adaptation (LoRA) [16] was the first to apply this strategy. It reformulates weight matrices $\\mathbf{W}$ as $\\mathbf{W} + \\mathbf{A} \\cdot \\mathbf{B}$ , where $\\mathbf{A}$ and $\\mathbf{B}$ are low-rank matrices. As a key advantage, this technique helps reduce overfitting and catastrophic forgetting because of two properties. First, LoRA enables fine-grained parallel adaptation by learning small perturbations to the frozen weights [13]. Second, the effective initialization of $\\mathbf{A}$ and $\\mathbf{B}$ helps in preserving prior knowledge [30, 43].", + "bbox": [ + 89, + 380, + 482, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on these advantages, recent studies [25, 45] show that low-rank reparameterization methods achieve state-of-the-art performance in few-shot adaptation of VLMs. However, these methods also have notable limitations. Compelling evidence suggests that restricting weight updates to low-rank subspaces reduces model expressiveness [24, 37, 47]. In this work, we argue that the low-rank projection technique exhibits severe overfitting in few-shot adaptation. Our systematic study of LoRA confirms this issue (see Fig. 1). Test-set accuracy initially improves but then declines sharply. At convergence, when the loss approaches zero, performance deteriorates significantly compared to its peak on test-set data. Moreover, we can see that the optimal number of training iterations varies widely across datasets. However, tuning hyperparameters, including the number of iterations for early stopping, on a validation set contradicts the principles of few-shot learning.", + "bbox": [ + 89, + 565, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The challenges of low-rank reparameterization methods extend beyond fixing the number of training iterations. Their performance is also highly sensitive to another hyperparameter: the rank of the low-rank projection. This sensitivity is shown in Fig. 2, which illustrates LoRA's perfor", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mance for different ranks on three datasets (DTD, Oxford Pets, UCF101) in a 1-shot setting. Each dataset reveals a distinct best rank, and higher ranks sometimes lead to an early spike in accuracy but then degrade as training continues. Further experiments examining these behaviors are presented in Appendix F of the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since rank is a discrete value that cannot be lower than 1, the low-rank methods lack the flexibility to adjust very few parameters. While setting the rank to a very low value (e.g., 1 or 2) can help mitigate overfitting, this choice might reduce the model's learning capacity. This limitation arises because the trainable parameters in low-rank methods remain fixed throughout training, potentially leading to underfitting the VLM capacity. Indeed, rank selection involves a trade-off between underfitting and overfitting (see Fig. 2). For instance, in the 1-shot DTD setting, LoRA at rank 5 briefly exceeds rank 2 but then drops below it as training continues. Overall, the test accuracy of LoRA exhibits oscillations and varies unpredictably w.r.t. rank and the number of training iterations. This instability makes it difficult to reliably tune these interdependent hyperparameters without access to a validation set.", + "bbox": [ + 511, + 185, + 903, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we deviate from the current long-standing trend of low-rank adaptation and introduce two new paradigms that make sparsity work in few-shot settings. Our approach is a simple yet effective optimization strategy, referred to as sparse optimization (SO), and aligns with these paradigms. As a first paradigm, we advocate for local sparsity and global density. Unlike the conventional low-rank approaches that update a fixed set of parameters, SO dynamically updates a minimal subset of original connections at each step. Specifically, high sparsity is enforced in both the gradient and moment updates while allowing the sparsity support to evolve dynamically throughout training. This paradigm ensures that, at any given iteration, only a minimal subset of parameters is updated to reduce the local learning capacity and prevent rapid overfitting. By dynamically changing the sparsity support, the model retains its overall expressiveness and ensures that different parameters contribute to learning over time.", + "bbox": [ + 511, + 429, + 903, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a second paradigm, we advocate for local randomness and global importance. Similar to Adam [19], we leverage the gradient and the first and second moments. However, we sparsify these components in a way that addresses overfitting and ensures efficiency in terms of memory consumption. In particular, we sparsify the gradient randomly and the first moment by importance. The gradient captures local and iteration-specific information. The first moment aggregates the gradients over the whole path and, thus, reflects long-term parameter importance. Random pruning of the gradient prevents the model from relying too much on short-term and local high-magnitude updates. The importance-based selection of the first moment", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3145", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ensures that connections with long-term significance are updated. By emphasizing local randomness and global importance, our strategy avoids severe overfitting.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely instantiating this paradigm, we fix a density rate $\\kappa$ and retain $M$ elements for both the gradient and moment updates at each iteration. First, we randomly select $M$ gradient values from the full set of parameter updates. Next, we compute the temporary first and second moments, which can contain up to $2M$ values in the worst-case scenario where no overlap exists between the selected gradients and the previous $M$ first moment values. We rank the $2M$ first moment values by importance using their magnitude. Only the top $M$ values are retained for the first moment. We then use the same indices to select the corresponding values for the second moment. This ensures alignment between the sparsified first and second moments. The selection strategy preserves the most relevant long-term updates while introducing randomness in the local updates.", + "bbox": [ + 89, + 137, + 482, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our contributions. (1) A novel framework is proposed that sparsifies both gradients and moments to improve generalization and mitigate overfitting in few-shot learning. Our approach differs from conventional low-rank reparameterization by focusing on sparsity. (2) We introduce the local sparsity and global density paradigm, where a dynamically evolving sparsity support updates only a minimal subset of model parameters at each iteration. This strategy preserves the overall model expressiveness while reducing its risk of overfitting. (3) We propose the local randomness and global importance paradigm, which selectively sparsifies the gradient using random selection while pruning the first moment based on importance. This selection strategy preserves the most relevant long-term updates while introducing local randomness to mitigate overfitting. (4) Our comprehensive empirical study on 11 diverse datasets compares our method to state-of-the-art low-rank approaches. Results indicate that SO can significantly improve few-shot generalization while achieving memory efficiency.", + "bbox": [ + 89, + 363, + 482, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 664, + 232, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(a) Low-Rank Methods. LoRA [16] approximates weight updates by decomposing them into two low-rank matrices. As a result, it reduces the number of trainable parameters while preserving the model's frozen pretrained backbone. DoRA [26] introduces a weight decomposition approach to improve LoRA's expressiveness by decoupling weight magnitude and direction and training them separately. VeRA [21] reduces the set of trainable parameters further than LoRA. In particular, VeRA freezes the pair of low-rank matrices to randomly initialized ones and only trains two small scaling vectors. PiSSA [30] initializes the two low-rank matrices using the most informative singular vectors and their corresponding singular values from the original weight matrix while freezing the less significant compo", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "nents in a residual matrix. ReLoRA [24] achieves high-rank weight updates by merging trained low-rank updates into the main model parameters at regular intervals and reinitializing the auxiliary low-rank matrices. Unlike these LoRA-based reparametrization methods, which constrain weight updates to low-rank subspaces, GaLore [47] performs full-parameter training. Specifically, GaLore projects gradients into a low-rank subspace to reduce memory consumption while allowing full updates to the model weights.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite their efficiency, these methods still face major limitations. Most low-rank techniques impose a fixed-rank constraint on the updates, which restricts adaptation flexibility and may lead to underfitting when the rank is too low. Even with minimal rank, these methods struggle with generalization in few-shot adaptation due to severe overfitting. Furthermore, these approaches require careful hyperparameter tuning, particularly in selecting the optimal rank and training iterations, which significantly impact performance. Our work challenges this strategy by showing that appropriately designed sparsity-based adaptation provides superior generalization while reducing memory consumption.", + "bbox": [ + 511, + 227, + 903, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Sparsity-Based Methods. These methods update only a subset of model parameters. The selection can be structured or unstructured. Structured selection methods [4, 9, 23, 28, 40] target entire modules (e.g., attention heads, layers), offering coarse-grained sparsity but often requiring architecture-specific designs that may degrade performance [5]. Unstructured selection methods [2, 12, 36, 38, 42, 46] prune individual weights for greater flexibility. However, most unstructured methods rely on static sparsity patterns and perform importance-based gradient pruning. Moreover, none of these methods incorporates sparse moments.", + "bbox": [ + 511, + 409, + 903, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, SO updates a minimal subset of parameters at each step. Furthermore, our approach uses random gradient selection to avoid local high-magnitude biases and importance-based moment pruning to retain long-term relevance. This combination enables stable adaptation in few-shot scenarios without sacrificing model expressiveness.", + "bbox": [ + 511, + 575, + 903, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Proposed Sparse Optimization Method", + "text_level": 1, + "bbox": [ + 511, + 679, + 862, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A comprehensive overview of few-shot adaptation in VLMs is provided in Appendix A. Briefly, we leverage a pretrained CLIP [33] model that aligns images and texts in a shared embedding space via contrastive learning. After pretraining, CLIP can perform zero-shot classification by comparing image embeddings to class-specific text prototypes. In the few-shot scenario, we only have a small support set of labeled examples (e.g., $K$ images per class). We adapt CLIP to this support set by minimizing the cross-entropy loss with respect to the class prototypes or learnable parameters, and we optimize the loss function using our proposed approach.", + "bbox": [ + 511, + 704, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the challenges inherent in few-shot adaptation, we introduce SO, a new optimizer designed to en", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3146", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hance memory efficiency while mitigating overfitting. Unlike standard optimization approaches that update all parameters at each iteration, SO dynamically updates a subset of parameters. This sparsification significantly reduces memory consumption, which makes SO suitable for training large-scale deep learning models like VLMs.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach builds on the Adam optimizer and behaves identically to it when the sparsity ratio is set to zero. However, SO deviates from Adam by following two paradigms. The first paradigm is local sparsity and global density, which ensures that a minimal subset of parameters is dynamically updated while maintaining overall model expressiveness. The second paradigm is local randomness and global importance, which applies random pruning to the gradient updates while retaining the most significant moment estimates. These principles prevent severe overfitting and make SO suitable for scenarios like few-shot learning.", + "bbox": [ + 89, + 183, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notation. Let $\\Theta \\in \\mathbb{R}^d$ represent the model parameters, where $d$ is the total number of parameters. At each iteration $t$ , the model updates some parameters to minimize a given loss function $\\mathcal{L}(\\Theta_t)$ . The gradient of the loss concerning all parameters is denoted as $g_{t} = \\nabla_{\\Theta}\\mathcal{L}(\\Theta_{t})$ . We use the first and second moments, denoted as $\\mu_t$ and $\\nu_{t}$ respectively, to capture long-term and global parameter significance throughout the optimization process. A small constant $\\epsilon$ is introduced for numerical stability. The learning rate, denoted as $\\eta$ , scales the parameter updates. We introduce a density ratio $\\kappa \\in [0,1]$ , which determines the fraction of gradient entries retained at each step. Let $\\mathcal{I}(\\cdot)$ and $\\mathcal{V}(\\cdot)$ be the functions that return the indices and values of a sparse input vector, respectively. Instead of updating all parameters, SO retains only a small subset of the gradient and generates sparse first and second moments.", + "bbox": [ + 89, + 351, + 482, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We enforce sparsity through two operations. The Top- $M$ operation selects $M$ elements with the highest magnitude. In contrast, Random- $M$ retains $M$ elements uniformly at random, preventing bias toward high-magnitude values.", + "bbox": [ + 89, + 595, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic and High-Sparsity Support. Conventional optimization methods update the full set of parameters throughout training, which accelerates overfitting. In contrast, our method dynamically updates a minimal subset of parameters. The high sparsity and dynamic selection provide two key advantages. First, updating a minimal subset of parameters reduces the local learning capacity and thus mitigates overfitting and catastrophic forgetting. Second, continuously changing the sparsity support ensures that different parameters receive updates throughout training. This avoids excessive reliance on a fixed set and makes the learning process benefit from the model's overall expressiveness.", + "bbox": [ + 89, + 657, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Randomness in Gradient Pruning. We sparsify the gradient vector to mitigate overfitting. The proposed optimizer applies a stochastic sampling mechanism that uniformly picks $M$ elements from the gradient $g_{t}$ . Given the density", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ratio $\\kappa$ , we define the number of retained gradient entries as $M = \\lfloor \\kappa d \\rfloor$ . Formally, we have $\\tilde{g}_t = \\text{Random-} M(g_t)$ .", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The motivation is to avoid biasing updates toward high-magnitude gradients. Enforcing stochastic sparsity prevents the model from over-relying on seemingly important local information. Our empirical results indicate that pruning the gradient by retaining only the largest values accelerates overfitting and significantly degrades performance.", + "bbox": [ + 511, + 121, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importance in Moment Pruning. To maintain efficiency, our goal is to construct sparse first and second moments, $\\tilde{\\mu}_t$ and $\\tilde{\\nu}_t$ , each of size $M$ . Since the sparse gradient changes its support dynamically, storing historical information becomes challenging. Without pruning, the first and second moments can become dense over time due to the accumulation of gradients with varying sparsity support.", + "bbox": [ + 511, + 212, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we introduce two fixed-size buffers that can store up to $2M$ values. This ensures that both new and past gradient values are considered. Let $\\mu_t$ and $\\nu_t$ be the temporary first and second moments, at iteration $t$ :", + "bbox": [ + 511, + 316, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {t} = \\beta_ {1} \\tilde {\\mu} _ {t - 1} + (1 - \\beta_ {1}) \\tilde {g} _ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 390, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nu_ {t} = \\beta_ {2} \\tilde {\\nu} _ {t - 1} + (1 - \\beta_ {2}) \\tilde {g} _ {t} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 409, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These moments may contain up to $2M$ values in the worst-case scenario when there is no overlap between the newly selected sparse gradient $\\tilde{g}_t$ and the previously retained sparse first moment $\\tilde{\\mu}_t$ . To ensure stable updates, we employ a Top- $M$ selection strategy for the first moment, where we rank the $2M$ values by magnitude and retain only the top $M$ ones. This guarantees that long-term significant parameters are consistently updated. For the second moment, we enforce alignment by retaining elements corresponding to the same indices selected for the first moment. The sparse first and second moments are expressed as:", + "bbox": [ + 511, + 439, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mu} _ {t} = \\operatorname {T o p} - M (\\mu_ {t}), \\quad \\tilde {\\nu} _ {t} = \\nu_ {t} [ \\mathcal {I} (\\tilde {\\mu} _ {t}) ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 617, + 903, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both sparse moments $\\tilde{\\mu}$ and $\\tilde{\\nu}$ are initialized at zero. Similar to Adam, we apply bias correction to $\\tilde{\\mu}_t$ and $\\tilde{\\nu}_t$ before parameter update. The corrected sparse first and second moments, denoted by $\\hat{\\mu}_t$ and $\\hat{\\nu}_t$ , are expressed as follows:", + "bbox": [ + 511, + 646, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} _ {t} = \\frac {\\tilde {\\mu} _ {t}}{1 - \\beta_ {1} ^ {t}}, \\quad \\hat {\\nu} _ {t} = \\frac {\\tilde {\\nu} _ {t}}{1 - \\beta_ {2} ^ {t}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 717, + 903, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the parameters $\\Theta$ are updated using the corrected sparse first and second moments, $\\hat{\\mu}_t$ and $\\hat{\\nu}_t$ , as provided in:", + "bbox": [ + 511, + 758, + 903, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {t + 1} = \\Theta_ {t} - \\frac {\\eta}{\\sqrt {\\hat {\\nu} _ {t}} + \\epsilon} \\hat {\\mu} _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 800, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate, and $\\beta_{1}$ and $\\beta_{2}$ are decay rates. Algorithm. Our algorithm is provided in Appendix B due to space limitations. We finetune CLIP until convergence using our optimization technique. To promote efficiency,", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3147", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7cfaa48a5aedc1ef162d9dbfc54d8c2fb8e7f5c8bbaaff316b0de98a55a8ee63.jpg", + "table_caption": [ + "Table 1. Few-shot classification performance on 11 datasets with ViT-B/16 backbone. Top-1 accuracy averaged over 3 random seeds is reported. Highest value is highlighted in bold, and the second highest is underlined." + ], + "table_footnote": [], + "table_body": "
ShotsMethodImageNetSUNAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
0CLIP (ICML '21)66.762.624.747.565.386.189.171.492.943.666.765.1
1Adam (ICLR '15)0.10.61.024.30.51.02.91.41.05.61.73.6
LoRA (ICLR '22)67.367.025.067.568.281.290.585.792.352.472.970.0
ReLoRA (ICLR '24)70.369.728.873.870.584.291.985.193.454.276.072.5
GaLoRE (ICML '24)65.165.521.165.263.674.185.280.591.448.969.966.4
PiSSA (NeurIPS '24)64.965.222.266.864.275.784.882.990.750.570.167.1
DoRA (ICML '24)66.966.925.371.468.381.389.386.292.351.972.970.2
VeRA (ICLR '24)68.467.327.569.367.486.291.073.193.952.270.169.7
SO (Ours)70.170.331.578.271.686.293.384.994.155.376.473.8
2Adam (ICLR '15)0.20.80.942.10.71.93.917.13.17.23.87.4
LoRA (ICLR '22)67.468.330.381.970.179.389.490.993.759.676.173.4
ReLoRA (ICLR '24)70.670.532.283.872.982.690.390.794.760.479.175.3
GaLoRE (ICML '24)66.167.025.882.267.272.681.585.691.556.173.469.9
PiSSA (NeurIPS '24)65.166.726.177.367.174.884.487.491.957.473.370.1
DoRA (ICML '24)67.268.229.883.470.379.489.691.293.659.775.973.5
VeRA (ICLR '24)69.469.031.781.669.886.792.679.694.459.374.773.5
SO (Ours)70.572.337.282.774.485.392.291.995.360.280.476.6
4Adam (ICLR '15)0.12.12.447.90.73.16.330.512.59.69.211.3
LoRA (ICLR '22)68.569.735.285.274.878.787.994.193.963.578.275.4
ReLoRA (ICLR '24)71.272.237.684.777.182.390.394.194.964.080.077.1
GaLoRE (ICML '24)67.569.034.182.173.175.281.091.193.260.075.472.9
PiSSA (NeurIPS '24)66.468.832.482.873.275.680.692.092.961.175.772.9
DoRA (ICML '24)68.469.734.985.575.179.089.594.194.064.178.575.7
VeRA (ICLR '24)70.571.535.685.673.786.793.189.895.364.579.576.9
SO (Ours)71.473.738.687.778.985.392.495.195.566.483.478.9
", + "bbox": [ + 107, + 128, + 893, + 482 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4d986dd91f21006405ef92679c972834ea6bce69737572fc2386d885ca38b58.jpg", + "table_caption": [ + "Table 2. Impact of gradient and moment sparsification in few-shot classification (ViT-B/16)." + ], + "table_footnote": [], + "table_body": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
DenseSparseDenseSparse
1SO (≡ Adam)✓✗✓✗1.024.30.51.02.91.41.05.61.74.4
SO✗✓✓✗29.473.969.683.993.286.493.755.777.073.6
SO✗✓✗✓31.578.271.686.293.384.994.155.376.474.6
2SO (≡ Adam)✓✗✓✗0.942.10.71.93.917.13.17.23.89.0
SO✗✓✓✗29.482.671.081.591.490.193.759.778.475.3
SO✗✓✗✓37.282.774.485.392.291.995.360.280.477.7
4SO (≡ Adam)✓✗✓✗2.447.90.73.16.330.512.59.69.213.6
SO✗✓✓✗34.585.473.578.991.192.794.266.278.377.2
SO✗✓✗✓38.687.778.985.392.495.195.566.483.480.4
", + "bbox": [ + 107, + 520, + 893, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we retain the gradient sparsity support fixed for $T$ iterations before reselecting the trainable parameters. Overall, our approach introduces two hyperparameters: the density ratio $\\kappa$ , which controls the fraction of retained gradient entries, and the update interval $T$ , which determines how frequently the sparsity support is refreshed. A smaller $\\kappa$ value improves generalization but results in slower convergence. In contrast, a larger $\\kappa$ accelerates optimization at the cost of reduced performance. Similarly, reducing $T$ leads to better results by enabling more frequent updates to the sparsity support, although increasing $T$ speeds up convergence.", + "bbox": [ + 88, + 699, + 480, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The role of $T$ in our method is conceptually similar to update intervals used in prior work such as GaLore [47]", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and ReLoRA [24]. In ReLoRA, the low-rank matrices are merged with the original parameters every $T$ interval. GaLore also updates the low-rank projectors every $T$ interval. Unlike the low-rank methods, where hyperparameters such as rank and number of training iterations involve a strong trade-off between overfitting and underfitting, our sparsity-based approach remains robust to overfitting.", + "bbox": [ + 511, + 699, + 906, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results and Discussion", + "text_level": 1, + "bbox": [ + 511, + 827, + 728, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) Experimental Methodology. Our optimizer is evaluated on 11 datasets commonly used in few-shot VLM adaptation [45, 49], including ImageNet [8], SUN397 [41], Air", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3148", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/409cc2d6695656b13026e69cad117f65bbfa07ff9975a917c3129da4ab742ba0.jpg", + "table_caption": [ + "Table 3. Impact of randomness and importance in gradient and moment sparsification in few-shot classification (ViT-B/16)." + ], + "table_footnote": [], + "table_body": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
RandImptRandImpt
1SOX✓X✓17.673.960.572.087.975.989.953.968.266.6
SOX✓✓X20.771.463.774.692.079.991.954.671.169.6
SO✓X✓X31.271.769.885.693.380.394.155.074.972.9
SO✓XX✓31.578.271.686.293.384.994.155.376.474.6
2SOX✓X✓21.480.862.667.782.884.688.757.668.268.3
SOX✓✓X22.882.465.373.487.184.291.360.374.571.3
SO✓X✓X33.282.172.885.792.191.294.860.979.676.9
SO✓XX✓37.282.774.485.392.291.995.360.280.477.7
4SOX✓X✓28.283.267.966.969.886.489.055.969.962.5
SOX✓✓X31.081.071.573.077.786.991.458.974.171.6
SO✓X✓X39.584.777.686.292.793.395.866.082.579.8
SO✓XX✓38.687.778.985.392.495.195.566.483.480.4
", + "bbox": [ + 107, + 114, + 895, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f6af10b2c31b62df36f187b1583418bdf0bcdd3c58d7e4dc708d8bf6a233d72.jpg", + "table_caption": [ + "Table 4. Impact of dynamic sparsity support for gradient pruning in few-shot classification (ViT-B/16)." + ], + "table_footnote": [], + "table_body": "
ShotsMethodSparsity supportAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
StaticDynamic
1SO✓✘28.272.468.185.993.173.494.552.672.671.2
SO✘✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✘31.083.069.785.891.683.094.355.775.874.4
SO✘✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✘34.883.672.486.292.588.095.162.577.877.0
SO✘✓38.687.778.985.392.495.195.566.483.480.4
", + "bbox": [ + 107, + 354, + 895, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "craft [29], EuroSAT [14], Stanford-Cars [22], Food101 [3], OxfordPets [32], Flowers102 [31], Caltech101 [10], DTD [7], and UCF101 [35]. These datasets provide a comprehensive assessment across a variety of image domains.", + "bbox": [ + 88, + 518, + 480, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate SO against state-of-the-art low-rank methods, which have shown strong performance in finetuning Large Language Models and Vision Language models. Specifically, we compare our method against LoRA, ReLoRA, VeRA, DoRA, PiSSA, and GaLoRE. Our comparison also includes full fine-tuning using Adam to showcase the effect of severe overfitting. A discussion of these baselines, including their limitations, is provided in Sec. 2.", + "bbox": [ + 88, + 580, + 482, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a fair comparison, we follow the same hyperparameter settings as CLIP-LoRA [45] for the low-rank baselines. In particular, we set the rank to 2, and we train for 500 iterations. Due to severe overfitting, we introduce an early stopping criterion for the low-rank methods. If the training loss reaches 0.01, training is halted. This is particularly necessary to obtain competitive results using the low-rank approaches for some datasets where the test accuracy peaks often within a few iterations before rapidly overfitting. For our method, we set the density ratio to $\\kappa = 0.05\\%$ , and we update the gradient sparsity support every $T = 10$ iterations. We train until convergence, defined as reaching a loss threshold of 0.01. If convergence is not achieved, we stop", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training at a maximum limit of 2000 iterations.", + "bbox": [ + 511, + 518, + 823, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following CLIP-LoRA [45], we adopt a learning rate of $2 \\times 10^{-4}$ with a cosine learning rate scheduler and a batch size of 32. We use the simple prompt template \"a photo of a [class name]\" without complex manual prompt engineering. All experiments are conducted using a ViT-B/16 backbone, and finetuning is applied to all layers of the 12 vision and text Transformer blocks. These settings were suggested by CLIP-LoRA and are kept the same for all methods, including ours. All hyperparameters remain fixed across all datasets for our approach and the baselines.", + "bbox": [ + 511, + 536, + 905, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments are conducted under identical hardware and software conditions. The details of these environments are provided in Appendix E. Additional experimental results showing CLIP's few-shot adaptation performance are presented in Appendix F. A comparative analysis of the memory efficiency of SO and state-of-the-art low-rank methods is presented in Appendix C. The sensitivity of our approach to the hyperparameters $\\kappa$ and $T$ is explored in Appendix D. Moreover, to further assess the effectiveness of our optimizer, we conduct both standard classification and few-shot learning experiments using a simple two-layer fully connected network. The results obtained with this small architecture are detailed in Appendix G.", + "bbox": [ + 511, + 688, + 905, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) Sparsity Outperforms Low-Rank Projections. The", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3149", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4617efd5f3608b24a9f4991687c7201ed405d79b26b7c98d4b84d9dc9ebdafd8.jpg", + "table_caption": [ + "Table 5. Impact of moments in few-shot classification (ViT-B/16)." + ], + "table_footnote": [], + "table_body": "
ShotsMethodMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
withoutwith
1SO✓✗20.974.751.976.585.782.992.554.570.267.8
SO✗✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✗24.984.660.277.589.590.992.961.176.073.0
SO✗✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✗32.289.172.179.089.193.593.864.477.376.7
SO✗✓38.687.778.985.392.495.195.566.483.480.4
", + "bbox": [ + 107, + 114, + 895, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results in Table 1 show the effectiveness of SO compared to state-of-the-art low-rank techniques. Across all shots (1, 2, and 4), our method achieves the highest average performance. Furthermore, SO yields the highest accuracy in most datasets, including EuroSAT, Cars, Flowers, and UCF101. Notably, it outperforms LoRA and ReLoRA on Aircraft by significant margins. Low-rank reparametrization methods constrain updates to a fixed subspace. The only exception is ReLoRA, which periodically merges and reinitializes the auxiliary low-rank matrices; this might explain why ReLoRA outperforms the other low-rank reparametrization methods. Moreover, these methods suffer from severe overfitting and require careful tuning of the rank and number of training iterations to achieve their highest performance. Unlike these methods, SO employs dynamic sparsity to update only a minimal subset of parameters, thereby mitigating overfitting while preserving the model's full representational capacity.", + "bbox": [ + 91, + 281, + 480, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(d) Effectiveness of Pruning in Few-Shot Optimization. Table 2 shows the impact of gradient and moment pruning on few-shot classification. Dense optimization (Adam) performs poorly in the few-shot setting due to severe overfitting. In contrast, gradient pruning leads to substantial performance improvement. Moment pruning further enhances performance by filtering out updates related to parameters that are only locally important. These findings validate that sparsity can improve generalization in few-shot learning.", + "bbox": [ + 91, + 553, + 480, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(e) Local Randomness Global Importance. Table 3 presents the performance of four sparse optimization variants based on randomness and importance in both gradient and moment pruning. Importance-based pruning is achieved by retaining the values with the highest magnitude. The results indicate that applying random selection for gradient pruning and importance-based selection for moment pruning achieves the highest average performance across all few-shot settings. Results in Appendix F confirm that importance-based gradient selection leads to rapid overfitting. Furthermore, Table 3 shows that importance-based moment pruning outperforms random moment pruning. This is because significant parameters over the long term are consistently updated, while locally influential ones", + "bbox": [ + 91, + 688, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are filtered out from the moment updates. Overall, our results validate the local randomness and global importance paradigm. Stochastic gradient selection prevents bias toward high-magnitude local updates and importance-based pruning preserves influential parameters over the long term. (f) Local Sparsity Global Density Table 4 shows that dynamic sparsity support consistently outperforms fixed sparsity support in gradient pruning across all few-shot settings. This improvement validates the local sparsity and global density paradigm. On the one hand, dynamic sparsity ensures that few parameters contribute to the updates to reduce the local learning capacity and thus mitigate overfitting. On the other hand, dynamic sparsity ensures a broader exploration of the parameter space while making the learning process benefit from the model's overall expressiveness. (g) Role of Moments in Few-Shot Adaptation. Table 5 shows that moment-based updates significantly improve performance. The advantage is more pronounced in datasets with higher complexity, such as Aircraft and Cars. Without moments, updates rely solely on randomly selected gradient values. These results provide evidence that gradient pruning is not sufficient. Random gradient pruning is more effective when combined with moments, particularly sparse ones.", + "bbox": [ + 511, + 281, + 903, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we shed light on sparsity as an effective mechanism to mitigate overfitting in few-shot adaptation of VLMs. Conventional low-rank methods suffer from severe overfitting and require extensive hyperparameter tuning. To address these limitations, we introduce two key paradigms. The first paradigm consists of dynamically selecting a minimal subset of parameters to reduce overfitting while preserving the overall model expressiveness. The second paradigm aims to prevent bias in gradient updates while retaining critical long-term information. Our experiments confirm that sparsity outperforms low-rank projections in few-shot learning and constitutes a strong mechanism against overfitting. As future work, we plan to evaluate our optimizer on large language models. Memory efficiency in this setting is even more critical.", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3150", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Digital Research Alliance of Canada.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Randall Balestriero, Leon Bottou, and Yann LeCun. The effects of regularization and data augmentation are class dependent. In Advances in Neural Information Processing Systems (NeurIPS), pages 37878-37891, 2022. 1", + "[2] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 1-9, 2022. 4", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446–461, 2014. 7", + "[4] Tianyi Chen, Tianyu Ding, Badal Yadav, Ilya Zharkov, and Luming Liang. Lorashear: Efficient large language model structured pruning and knowledge recovery. arXiv preprint arXiv:2310.18356, 2023. 4", + "[5] Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 10558-10578, 2024. 4", + "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2023. 1", + "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 7", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 6", + "[9] Lucio Dery, Steven Kolawole, Jean-François Kagy, Virginia Smith, Graham Neubig, and Ameet Talwalkar. Everybody prune now: Structured pruning of llms with only forward passes. arXiv preprint arXiv:2402.05406, 2024. 4", + "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 178–178, 2004. 7", + "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 1" + ], + "bbox": [ + 93, + 198, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, and Bamdev Mishra. SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4", + "[13] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations (ICLR), 2021. 3", + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7", + "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 1", + "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 1, 3, 4", + "[17] Iiro P Jaëskeläinen, Enrico Glerean, Vasily Klucharev, Anna Shestakova, and Jyrki Ahveninen. Do sparse brain activity patterns underlie human cognition? NeuroImage, 263: 119633, 2022. 1", + "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 1", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 3", + "[20] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3", + "[21] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M Asano. VeRA: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 1, 4", + "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops (WICCV), pages 554–561, 2013. 7", + "[23] François Lagunas, Ella Charlaix, Victor Sanh, and Alexander Rush. Block pruning for faster transformers. In Empirical Methods in Natural Language Processing (EMNLP), pages 10619-10629, 2021. 4", + "[24] Vladislav Lialin, Sherin Muckatira, Namrata Shivagunde, and Anna Rumshisky. Relora: High-rank training through low-rank updates. In International Conference on Learning Representations (ICLR), 2023. 3, 4, 6", + "[25] Cheng Lin, Lujun Li, Dezhi Li, Jie Zou, Wei Xue, and Yike Guo. Nora: Nested low-rank adaptation for efficient" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3151", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fine-tuning large models. arXiv preprint arXiv:2408.10280, 2024.3", + "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), pages 32100-32121, 2024. 3, 4", + "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning (ICML), pages 23033-23044, 2023. 1", + "[28] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4", + "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. In *British Machine Vision Conference (BMVC)*, 2013. 7", + "[30] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. PiSSA: Principal singular values and singular vectors adaptation of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4", + "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing (ICVGIP), pages 722-729, 2008. 7", + "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 4", + "[34] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3", + "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7", + "[36] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In International Conference on Learning Representations (ICLR), 2024. 4", + "[37] Xianghui Sun, Yunjie Ji, Baochang Ma, and Xiangang Li. A comparative study between full-parameter and lora-based fine-tuning on chinese instruction data for instruction following large language model. arXiv preprint arXiv:2304.08109, 2023. 3", + "[38] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. In Advances in Neural" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Information Processing Systems (NeurIPS), pages 24193-24205, 2021. 4", + "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1", + "[40] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In Annual Meeting of the Association for Computational Linguistics (ACL), 2022. 4", + "[41] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3485-3492, 2010. 6", + "[42] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In Empirical Methods in Natural Language Processing (EMNLP), pages 9514-9528, 2021. 4", + "[43] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. CorDA: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3", + "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10899-10909, 2023. 3", + "[45] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 1593-1603, 2024. 3, 6, 7", + "[46] Zhi Zhang, Qizhe Zhang, Zijun Gao, Renrui Zhang, Ekaterina Shutova, Shiji Zhou, and Shanghang Zhang. Gradient-based parameter selection for efficient fine-tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28566-28577, 2024. 4", + "[47] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient llm training by gradient low-rank projection. In International Conference on Machine Learning (ICML), 2024. 3, 4, 6", + "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 1", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 1, 6", + "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2605-2615, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3152", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_model.json b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..254836f4cec554d273575d07179f68c2ef7eff2c --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.182, + 0.796, + 0.217 + ], + "angle": 0, + "content": "Nairouz Mrabah Nicolas Richet Ismail Ben Ayed Eric Granger LIVIA, ILLS, Department of Systems Engineering, ETS Montreal, Canada" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.22, + 0.777, + 0.252 + ], + "angle": 0, + "content": "mrabah.nairouz@livia.etsmt1.ca, nicolas.richet.l@ens.etsmt1.ca, ismail.benayed@etsmt1.ca, eric.granger@etsmt1.ca" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Adapting vision-language models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for local sparsity and global density, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for local randomness and global importance, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead. Code is available at: https://github.com/nairouz/SO." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The performance of deep learning (DL) models is strongly influenced by their size, with larger architectures yielding superior results. For instance, the ViT-H/14 backbone outperforms ViT-B/32 by a significant margin across multiple benchmarks [6]. Despite recent architectural advancements, the best models remain highly data-dependent and require large-scale labeled datasets to achieve reliable generalization. While techniques such as data augmentation and regularization help mitigate overfitting in low-data scenarios, they are insufficient to fully resolve it [1]. In contrast, the human brain has more than 80 billion neurons. Each neuron forms, on average, only about 7,000 synaptic connec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.41 + ], + "angle": 0, + "content": "tions, which results in an estimated total of 100 to 500 trillion synapses. Despite its tremendous capacity, the brain exhibits a remarkable ability to grasp new concepts from minimal exposure. This one-shot or few-shot learning capability might be partly attributed to the effective use of sparsity [17]. By selectively activating a fraction of connections at any given time, the brain optimizes both energy consumption and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.684 + ], + "angle": 0, + "content": "Among the prominent DL models, vision-language models (VLMs) learn visual concepts from natural language descriptions. In particular, the contrastive language-image pretraining (CLIP) [33] model is pretrained on large-scale datasets containing hundreds of millions of image-text pairs to align visual and textual embeddings within a shared semantic space. After training, VLMs show strong zero-shot capabilities across diverse visual recognition tasks [33]. Moreover, fine-tuning VLMs yields state-of-the-art performance in various domain adaptation scenarios [27, 39]. Despite their zero-shot capabilities, adapting VLMs to new tasks with small labeled datasets remains challenging due to pronounced overfitting and catastrophic forgetting problems. Furthermore, training such models requires vast resources, including significant GPU memory for storing millions of parameters, gradients, activations, and optimizer states. These challenges underscore the need for cost-effective optimization strategies that address overfitting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.884 + ], + "angle": 0, + "content": "To address the limitations of full VLM fine-tuning in few-shot learning, state-of-the-art methods leverage parameter-efficient training (PET) [15, 16, 21]. This strategy consists of optimizing a small number of parameters compared to the full set of weights. Additive PET methods freeze the initial weights and attach trainable and unmergeable parameters to the model. Within this category, we find the adapter [11] and the prompt tuning variants [49]. The prevailing methods in few-shot VLMs have mainly focused on this category. Adjustable weights can be incorporated at various levels, either by embedding them directly into the input space as visual [18] or textual prompts [48], or by integrating adapters throughout the network architecture [50]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "To mitigate overfitting in few-shot adaptation and pre" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3143" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.093, + 0.312, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.093, + 0.51, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.093, + 0.707, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.093, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.23, + 0.312, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.234, + 0.51, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.234, + 0.707, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.231, + 0.904, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 1. Test-set accuracy over training iterations for LoRA (with rank fixed at 3) and SO (with a sparsity ratio equal to \\(99.95\\%\\)) in a 1-shot setting on four datasets (Pets, Flowers, Caltech, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. Training stops when the loss falls below 0.01 or after 2000 iterations." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.425, + 0.312, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.426, + 0.51, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.426, + 0.707, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.426, + 0.904, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.563, + 0.312, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.563, + 0.51, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.563, + 0.707, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.56, + 0.904, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.683, + 0.312, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.683, + 0.51, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.683, + 0.707, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.682, + 0.904, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.817, + 0.908, + 0.86 + ], + "angle": 0, + "content": "Figure 2. Test-set accuracy of LoRA in a 1-shot setting on three datasets (DTD, Oxford Pets, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. We vary the LoRA rank from 2 to 5 and train for at most 2000 iterations or until the loss drops below 0.01. Each subfigure shows test accuracy over training iterations for a specific rank." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3144" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.378 + ], + "angle": 0, + "content": "serve the prior knowledge of VLMs, adapter-based methods typically employ three techniques. First, these small trainable modules are only integrated into the last layer of VLMs. Second, they use regularization techniques to penalize the divergence between the initial zero-shot class prototypes and the learnable class prototypes. Finally, they initialize the learnable class prototypes into a reliable region using the zero-shot prototypes. As a result, these methods have shown better results than the prompt tuning approaches [44], which often introduce computational overhead [45]. Despite these efforts, adapter-based methods are still prone to considerable overfitting [34]. Their success depends heavily on carefully tuning hyperparameters whose optimal values vary significantly across different tasks [34]. This sensitivity leads to performance degradation that undermines the generalizability of these approaches. It is worth noting that tuning these hyperparameters requires extensive grid searching with a validation set, which is not available in real-world few-shot scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.563 + ], + "angle": 0, + "content": "Recently, low-rank reparametrization [16, 20, 26] has gained popularity for domain adaptation. These approaches reparameterize the initial weights with supplementary trainable and mergeable low-rank matrices. Low-rank adaptation (LoRA) [16] was the first to apply this strategy. It reformulates weight matrices \\(\\mathbf{W}\\) as \\(\\mathbf{W} + \\mathbf{A} \\cdot \\mathbf{B}\\), where \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) are low-rank matrices. As a key advantage, this technique helps reduce overfitting and catastrophic forgetting because of two properties. First, LoRA enables fine-grained parallel adaptation by learning small perturbations to the frozen weights [13]. Second, the effective initialization of \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) helps in preserving prior knowledge [30, 43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Building on these advantages, recent studies [25, 45] show that low-rank reparameterization methods achieve state-of-the-art performance in few-shot adaptation of VLMs. However, these methods also have notable limitations. Compelling evidence suggests that restricting weight updates to low-rank subspaces reduces model expressiveness [24, 37, 47]. In this work, we argue that the low-rank projection technique exhibits severe overfitting in few-shot adaptation. Our systematic study of LoRA confirms this issue (see Fig. 1). Test-set accuracy initially improves but then declines sharply. At convergence, when the loss approaches zero, performance deteriorates significantly compared to its peak on test-set data. Moreover, we can see that the optimal number of training iterations varies widely across datasets. However, tuning hyperparameters, including the number of iterations for early stopping, on a validation set contradicts the principles of few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The challenges of low-rank reparameterization methods extend beyond fixing the number of training iterations. Their performance is also highly sensitive to another hyperparameter: the rank of the low-rank projection. This sensitivity is shown in Fig. 2, which illustrates LoRA's perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "mance for different ranks on three datasets (DTD, Oxford Pets, UCF101) in a 1-shot setting. Each dataset reveals a distinct best rank, and higher ranks sometimes lead to an early spike in accuracy but then degrade as training continues. Further experiments examining these behaviors are presented in Appendix F of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Since rank is a discrete value that cannot be lower than 1, the low-rank methods lack the flexibility to adjust very few parameters. While setting the rank to a very low value (e.g., 1 or 2) can help mitigate overfitting, this choice might reduce the model's learning capacity. This limitation arises because the trainable parameters in low-rank methods remain fixed throughout training, potentially leading to underfitting the VLM capacity. Indeed, rank selection involves a trade-off between underfitting and overfitting (see Fig. 2). For instance, in the 1-shot DTD setting, LoRA at rank 5 briefly exceeds rank 2 but then drops below it as training continues. Overall, the test accuracy of LoRA exhibits oscillations and varies unpredictably w.r.t. rank and the number of training iterations. This instability makes it difficult to reliably tune these interdependent hyperparameters without access to a validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.701 + ], + "angle": 0, + "content": "In this paper, we deviate from the current long-standing trend of low-rank adaptation and introduce two new paradigms that make sparsity work in few-shot settings. Our approach is a simple yet effective optimization strategy, referred to as sparse optimization (SO), and aligns with these paradigms. As a first paradigm, we advocate for local sparsity and global density. Unlike the conventional low-rank approaches that update a fixed set of parameters, SO dynamically updates a minimal subset of original connections at each step. Specifically, high sparsity is enforced in both the gradient and moment updates while allowing the sparsity support to evolve dynamically throughout training. This paradigm ensures that, at any given iteration, only a minimal subset of parameters is updated to reduce the local learning capacity and prevent rapid overfitting. By dynamically changing the sparsity support, the model retains its overall expressiveness and ensures that different parameters contribute to learning over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As a second paradigm, we advocate for local randomness and global importance. Similar to Adam [19], we leverage the gradient and the first and second moments. However, we sparsify these components in a way that addresses overfitting and ensures efficiency in terms of memory consumption. In particular, we sparsify the gradient randomly and the first moment by importance. The gradient captures local and iteration-specific information. The first moment aggregates the gradients over the whole path and, thus, reflects long-term parameter importance. Random pruning of the gradient prevents the model from relying too much on short-term and local high-magnitude updates. The importance-based selection of the first moment" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3145" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "ensures that connections with long-term significance are updated. By emphasizing local randomness and global importance, our strategy avoids severe overfitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.483, + 0.363 + ], + "angle": 0, + "content": "Concretely instantiating this paradigm, we fix a density rate \\(\\kappa\\) and retain \\(M\\) elements for both the gradient and moment updates at each iteration. First, we randomly select \\(M\\) gradient values from the full set of parameter updates. Next, we compute the temporary first and second moments, which can contain up to \\(2M\\) values in the worst-case scenario where no overlap exists between the selected gradients and the previous \\(M\\) first moment values. We rank the \\(2M\\) first moment values by importance using their magnitude. Only the top \\(M\\) values are retained for the first moment. We then use the same indices to select the corresponding values for the second moment. This ensures alignment between the sparsified first and second moments. The selection strategy preserves the most relevant long-term updates while introducing randomness in the local updates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Our contributions. (1) A novel framework is proposed that sparsifies both gradients and moments to improve generalization and mitigate overfitting in few-shot learning. Our approach differs from conventional low-rank reparameterization by focusing on sparsity. (2) We introduce the local sparsity and global density paradigm, where a dynamically evolving sparsity support updates only a minimal subset of model parameters at each iteration. This strategy preserves the overall model expressiveness while reducing its risk of overfitting. (3) We propose the local randomness and global importance paradigm, which selectively sparsifies the gradient using random selection while pruning the first moment based on importance. This selection strategy preserves the most relevant long-term updates while introducing local randomness to mitigate overfitting. (4) Our comprehensive empirical study on 11 diverse datasets compares our method to state-of-the-art low-rank approaches. Results indicate that SO can significantly improve few-shot generalization while achieving memory efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.233, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(a) Low-Rank Methods. LoRA [16] approximates weight updates by decomposing them into two low-rank matrices. As a result, it reduces the number of trainable parameters while preserving the model's frozen pretrained backbone. DoRA [26] introduces a weight decomposition approach to improve LoRA's expressiveness by decoupling weight magnitude and direction and training them separately. VeRA [21] reduces the set of trainable parameters further than LoRA. In particular, VeRA freezes the pair of low-rank matrices to randomly initialized ones and only trains two small scaling vectors. PiSSA [30] initializes the two low-rank matrices using the most informative singular vectors and their corresponding singular values from the original weight matrix while freezing the less significant compo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "nents in a residual matrix. ReLoRA [24] achieves high-rank weight updates by merging trained low-rank updates into the main model parameters at regular intervals and reinitializing the auxiliary low-rank matrices. Unlike these LoRA-based reparametrization methods, which constrain weight updates to low-rank subspaces, GaLore [47] performs full-parameter training. Specifically, GaLore projects gradients into a low-rank subspace to reduce memory consumption while allowing full updates to the model weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.409 + ], + "angle": 0, + "content": "Despite their efficiency, these methods still face major limitations. Most low-rank techniques impose a fixed-rank constraint on the updates, which restricts adaptation flexibility and may lead to underfitting when the rank is too low. Even with minimal rank, these methods struggle with generalization in few-shot adaptation due to severe overfitting. Furthermore, these approaches require careful hyperparameter tuning, particularly in selecting the optimal rank and training iterations, which significantly impact performance. Our work challenges this strategy by showing that appropriately designed sparsity-based adaptation provides superior generalization while reducing memory consumption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.905, + 0.574 + ], + "angle": 0, + "content": "(b) Sparsity-Based Methods. These methods update only a subset of model parameters. The selection can be structured or unstructured. Structured selection methods [4, 9, 23, 28, 40] target entire modules (e.g., attention heads, layers), offering coarse-grained sparsity but often requiring architecture-specific designs that may degrade performance [5]. Unstructured selection methods [2, 12, 36, 38, 42, 46] prune individual weights for greater flexibility. However, most unstructured methods rely on static sparsity patterns and perform importance-based gradient pruning. Moreover, none of these methods incorporates sparse moments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.905, + 0.665 + ], + "angle": 0, + "content": "In contrast, SO updates a minimal subset of parameters at each step. Furthermore, our approach uses random gradient selection to avoid local high-magnitude biases and importance-based moment pruning to retain long-term relevance. This combination enables stable adaptation in few-shot scenarios without sacrificing model expressiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.864, + 0.696 + ], + "angle": 0, + "content": "3. Proposed Sparse Optimization Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.87 + ], + "angle": 0, + "content": "A comprehensive overview of few-shot adaptation in VLMs is provided in Appendix A. Briefly, we leverage a pretrained CLIP [33] model that aligns images and texts in a shared embedding space via contrastive learning. After pretraining, CLIP can perform zero-shot classification by comparing image embeddings to class-specific text prototypes. In the few-shot scenario, we only have a small support set of labeled examples (e.g., \\( K \\) images per class). We adapt CLIP to this support set by minimizing the cross-entropy loss with respect to the class prototypes or learnable parameters, and we optimize the loss function using our proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Motivated by the challenges inherent in few-shot adaptation, we introduce SO, a new optimizer designed to en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3146" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "hance memory efficiency while mitigating overfitting. Unlike standard optimization approaches that update all parameters at each iteration, SO dynamically updates a subset of parameters. This sparsification significantly reduces memory consumption, which makes SO suitable for training large-scale deep learning models like VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.35 + ], + "angle": 0, + "content": "Our approach builds on the Adam optimizer and behaves identically to it when the sparsity ratio is set to zero. However, SO deviates from Adam by following two paradigms. The first paradigm is local sparsity and global density, which ensures that a minimal subset of parameters is dynamically updated while maintaining overall model expressiveness. The second paradigm is local randomness and global importance, which applies random pruning to the gradient updates while retaining the most significant moment estimates. These principles prevent severe overfitting and make SO suitable for scenarios like few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.594 + ], + "angle": 0, + "content": "Notation. Let \\(\\Theta \\in \\mathbb{R}^d\\) represent the model parameters, where \\(d\\) is the total number of parameters. At each iteration \\(t\\), the model updates some parameters to minimize a given loss function \\(\\mathcal{L}(\\Theta_t)\\). The gradient of the loss concerning all parameters is denoted as \\(g_{t} = \\nabla_{\\Theta}\\mathcal{L}(\\Theta_{t})\\). We use the first and second moments, denoted as \\(\\mu_t\\) and \\(\\nu_{t}\\) respectively, to capture long-term and global parameter significance throughout the optimization process. A small constant \\(\\epsilon\\) is introduced for numerical stability. The learning rate, denoted as \\(\\eta\\), scales the parameter updates. We introduce a density ratio \\(\\kappa \\in [0,1]\\), which determines the fraction of gradient entries retained at each step. Let \\(\\mathcal{I}(\\cdot)\\) and \\(\\mathcal{V}(\\cdot)\\) be the functions that return the indices and values of a sparse input vector, respectively. Instead of updating all parameters, SO retains only a small subset of the gradient and generates sparse first and second moments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.656 + ], + "angle": 0, + "content": "We enforce sparsity through two operations. The Top-\\(M\\) operation selects \\(M\\) elements with the highest magnitude. In contrast, Random-\\(M\\) retains \\(M\\) elements uniformly at random, preventing bias toward high-magnitude values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Dynamic and High-Sparsity Support. Conventional optimization methods update the full set of parameters throughout training, which accelerates overfitting. In contrast, our method dynamically updates a minimal subset of parameters. The high sparsity and dynamic selection provide two key advantages. First, updating a minimal subset of parameters reduces the local learning capacity and thus mitigates overfitting and catastrophic forgetting. Second, continuously changing the sparsity support ensures that different parameters receive updates throughout training. This avoids excessive reliance on a fixed set and makes the learning process benefit from the model's overall expressiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Randomness in Gradient Pruning. We sparsify the gradient vector to mitigate overfitting. The proposed optimizer applies a stochastic sampling mechanism that uniformly picks \\( M \\) elements from the gradient \\( g_{t} \\). Given the density" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "ratio \\(\\kappa\\), we define the number of retained gradient entries as \\(M = \\lfloor \\kappa d \\rfloor\\). Formally, we have \\(\\tilde{g}_t = \\text{Random-} M(g_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.212 + ], + "angle": 0, + "content": "The motivation is to avoid biasing updates toward high-magnitude gradients. Enforcing stochastic sparsity prevents the model from over-relying on seemingly important local information. Our empirical results indicate that pruning the gradient by retaining only the largest values accelerates overfitting and significantly degrades performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Importance in Moment Pruning. To maintain efficiency, our goal is to construct sparse first and second moments, \\(\\tilde{\\mu}_t\\) and \\(\\tilde{\\nu}_t\\), each of size \\(M\\). Since the sparse gradient changes its support dynamically, storing historical information becomes challenging. Without pruning, the first and second moments can become dense over time due to the accumulation of gradients with varying sparsity support." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.379 + ], + "angle": 0, + "content": "To address this issue, we introduce two fixed-size buffers that can store up to \\(2M\\) values. This ensures that both new and past gradient values are considered. Let \\(\\mu_t\\) and \\(\\nu_t\\) be the temporary first and second moments, at iteration \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.391, + 0.905, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mu_ {t} = \\beta_ {1} \\tilde {\\mu} _ {t - 1} + (1 - \\beta_ {1}) \\tilde {g} _ {t}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.41, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\nu_ {t} = \\beta_ {2} \\tilde {\\nu} _ {t - 1} + (1 - \\beta_ {2}) \\tilde {g} _ {t} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.606 + ], + "angle": 0, + "content": "These moments may contain up to \\(2M\\) values in the worst-case scenario when there is no overlap between the newly selected sparse gradient \\(\\tilde{g}_t\\) and the previously retained sparse first moment \\(\\tilde{\\mu}_t\\). To ensure stable updates, we employ a Top-\\(M\\) selection strategy for the first moment, where we rank the \\(2M\\) values by magnitude and retain only the top \\(M\\) ones. This guarantees that long-term significant parameters are consistently updated. For the second moment, we enforce alignment by retaining elements corresponding to the same indices selected for the first moment. The sparse first and second moments are expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.618, + 0.905, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mu} _ {t} = \\operatorname {T o p} - M (\\mu_ {t}), \\quad \\tilde {\\nu} _ {t} = \\nu_ {t} [ \\mathcal {I} (\\tilde {\\mu} _ {t}) ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.905, + 0.707 + ], + "angle": 0, + "content": "Both sparse moments \\(\\tilde{\\mu}\\) and \\(\\tilde{\\nu}\\) are initialized at zero. Similar to Adam, we apply bias correction to \\(\\tilde{\\mu}_t\\) and \\(\\tilde{\\nu}_t\\) before parameter update. The corrected sparse first and second moments, denoted by \\(\\hat{\\mu}_t\\) and \\(\\hat{\\nu}_t\\), are expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.718, + 0.905, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} _ {t} = \\frac {\\tilde {\\mu} _ {t}}{1 - \\beta_ {1} ^ {t}}, \\quad \\hat {\\nu} _ {t} = \\frac {\\tilde {\\nu} _ {t}}{1 - \\beta_ {2} ^ {t}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.76, + 0.905, + 0.791 + ], + "angle": 0, + "content": "Finally, the parameters \\(\\Theta\\) are updated using the corrected sparse first and second moments, \\(\\hat{\\mu}_t\\) and \\(\\hat{\\nu}_t\\), as provided in:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.801, + 0.905, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {t + 1} = \\Theta_ {t} - \\frac {\\eta}{\\sqrt {\\hat {\\nu} _ {t}} + \\epsilon} \\hat {\\mu} _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate, and \\(\\beta_{1}\\) and \\(\\beta_{2}\\) are decay rates. Algorithm. Our algorithm is provided in Appendix B due to space limitations. We finetune CLIP until convergence using our optimization technique. To promote efficiency," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3147" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Few-shot classification performance on 11 datasets with ViT-B/16 backbone. Top-1 accuracy averaged over 3 random seeds is reported. Highest value is highlighted in bold, and the second highest is underlined." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.129, + 0.895, + 0.483 + ], + "angle": 0, + "content": "
ShotsMethodImageNetSUNAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
0CLIP (ICML '21)66.762.624.747.565.386.189.171.492.943.666.765.1
1Adam (ICLR '15)0.10.61.024.30.51.02.91.41.05.61.73.6
LoRA (ICLR '22)67.367.025.067.568.281.290.585.792.352.472.970.0
ReLoRA (ICLR '24)70.369.728.873.870.584.291.985.193.454.276.072.5
GaLoRE (ICML '24)65.165.521.165.263.674.185.280.591.448.969.966.4
PiSSA (NeurIPS '24)64.965.222.266.864.275.784.882.990.750.570.167.1
DoRA (ICML '24)66.966.925.371.468.381.389.386.292.351.972.970.2
VeRA (ICLR '24)68.467.327.569.367.486.291.073.193.952.270.169.7
SO (Ours)70.170.331.578.271.686.293.384.994.155.376.473.8
2Adam (ICLR '15)0.20.80.942.10.71.93.917.13.17.23.87.4
LoRA (ICLR '22)67.468.330.381.970.179.389.490.993.759.676.173.4
ReLoRA (ICLR '24)70.670.532.283.872.982.690.390.794.760.479.175.3
GaLoRE (ICML '24)66.167.025.882.267.272.681.585.691.556.173.469.9
PiSSA (NeurIPS '24)65.166.726.177.367.174.884.487.491.957.473.370.1
DoRA (ICML '24)67.268.229.883.470.379.489.691.293.659.775.973.5
VeRA (ICLR '24)69.469.031.781.669.886.792.679.694.459.374.773.5
SO (Ours)70.572.337.282.774.485.392.291.995.360.280.476.6
4Adam (ICLR '15)0.12.12.447.90.73.16.330.512.59.69.211.3
LoRA (ICLR '22)68.569.735.285.274.878.787.994.193.963.578.275.4
ReLoRA (ICLR '24)71.272.237.684.777.182.390.394.194.964.080.077.1
GaLoRE (ICML '24)67.569.034.182.173.175.281.091.193.260.075.472.9
PiSSA (NeurIPS '24)66.468.832.482.873.275.680.692.092.961.175.772.9
DoRA (ICML '24)68.469.734.985.575.179.089.594.194.064.178.575.7
VeRA (ICLR '24)70.571.535.685.673.786.793.189.895.364.579.576.9
SO (Ours)71.473.738.687.778.985.392.495.195.566.483.478.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.496, + 0.772, + 0.512 + ], + "angle": 0, + "content": "Table 2. Impact of gradient and moment sparsification in few-shot classification (ViT-B/16)." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.521, + 0.895, + 0.675 + ], + "angle": 0, + "content": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
DenseSparseDenseSparse
1SO (≡ Adam)✓✗✓✗1.024.30.51.02.91.41.05.61.74.4
SO✗✓✓✗29.473.969.683.993.286.493.755.777.073.6
SO✗✓✗✓31.578.271.686.293.384.994.155.376.474.6
2SO (≡ Adam)✓✗✓✗0.942.10.71.93.917.13.17.23.89.0
SO✗✓✓✗29.482.671.081.591.490.193.759.778.475.3
SO✗✓✗✓37.282.774.485.392.291.995.360.280.477.7
4SO (≡ Adam)✓✗✓✗2.447.90.73.16.330.512.59.69.213.6
SO✗✓✓✗34.585.473.578.991.192.794.266.278.377.2
SO✗✓✗✓38.687.778.985.392.495.195.566.483.480.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.7, + 0.482, + 0.866 + ], + "angle": 0, + "content": "we retain the gradient sparsity support fixed for \\( T \\) iterations before reselecting the trainable parameters. Overall, our approach introduces two hyperparameters: the density ratio \\( \\kappa \\), which controls the fraction of retained gradient entries, and the update interval \\( T \\), which determines how frequently the sparsity support is refreshed. A smaller \\( \\kappa \\) value improves generalization but results in slower convergence. In contrast, a larger \\( \\kappa \\) accelerates optimization at the cost of reduced performance. Similarly, reducing \\( T \\) leads to better results by enabling more frequent updates to the sparsity support, although increasing \\( T \\) speeds up convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The role of \\( T \\) in our method is conceptually similar to update intervals used in prior work such as GaLore [47]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.807 + ], + "angle": 0, + "content": "and ReLoRA [24]. In ReLoRA, the low-rank matrices are merged with the original parameters every \\(T\\) interval. GaLore also updates the low-rank projectors every \\(T\\) interval. Unlike the low-rank methods, where hyperparameters such as rank and number of training iterations involve a strong trade-off between overfitting and underfitting, our sparsity-based approach remains robust to overfitting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.828, + 0.73, + 0.844 + ], + "angle": 0, + "content": "4. Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "(a) Experimental Methodology. Our optimizer is evaluated on 11 datasets commonly used in few-shot VLM adaptation [45, 49], including ImageNet [8], SUN397 [41], Air" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3148" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.089, + 0.863, + 0.104 + ], + "angle": 0, + "content": "Table 3. Impact of randomness and importance in gradient and moment sparsification in few-shot classification (ViT-B/16)." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.115, + 0.896, + 0.32 + ], + "angle": 0, + "content": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
RandImptRandImpt
1SOX✓X✓17.673.960.572.087.975.989.953.968.266.6
SOX✓✓X20.771.463.774.692.079.991.954.671.169.6
SO✓X✓X31.271.769.885.693.380.394.155.074.972.9
SO✓XX✓31.578.271.686.293.384.994.155.376.474.6
2SOX✓X✓21.480.862.667.782.884.688.757.668.268.3
SOX✓✓X22.882.465.373.487.184.291.360.374.571.3
SO✓X✓X33.282.172.885.792.191.294.860.979.676.9
SO✓XX✓37.282.774.485.392.291.995.360.280.477.7
4SOX✓X✓28.283.267.966.969.886.489.055.969.962.5
SOX✓✓X31.081.071.573.077.786.991.458.974.171.6
SO✓X✓X39.584.777.686.292.793.395.866.082.579.8
SO✓XX✓38.687.778.985.392.495.195.566.483.480.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.331, + 0.802, + 0.345 + ], + "angle": 0, + "content": "Table 4. Impact of dynamic sparsity support for gradient pruning in few-shot classification (ViT-B/16)." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.356, + 0.896, + 0.496 + ], + "angle": 0, + "content": "
ShotsMethodSparsity supportAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
StaticDynamic
1SO✓✘28.272.468.185.993.173.494.552.672.671.2
SO✘✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✘31.083.069.785.891.683.094.355.775.874.4
SO✘✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✘34.883.672.486.292.588.095.162.577.877.0
SO✘✓38.687.778.985.392.495.195.566.483.480.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.482, + 0.581 + ], + "angle": 0, + "content": "craft [29], EuroSAT [14], Stanford-Cars [22], Food101 [3], OxfordPets [32], Flowers102 [31], Caltech101 [10], DTD [7], and UCF101 [35]. These datasets provide a comprehensive assessment across a variety of image domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.704 + ], + "angle": 0, + "content": "We evaluate SO against state-of-the-art low-rank methods, which have shown strong performance in finetuning Large Language Models and Vision Language models. Specifically, we compare our method against LoRA, ReLoRA, VeRA, DoRA, PiSSA, and GaLoRE. Our comparison also includes full fine-tuning using Adam to showcase the effect of severe overfitting. A discussion of these baselines, including their limitations, is provided in Sec. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For a fair comparison, we follow the same hyperparameter settings as CLIP-LoRA [45] for the low-rank baselines. In particular, we set the rank to 2, and we train for 500 iterations. Due to severe overfitting, we introduce an early stopping criterion for the low-rank methods. If the training loss reaches 0.01, training is halted. This is particularly necessary to obtain competitive results using the low-rank approaches for some datasets where the test accuracy peaks often within a few iterations before rapidly overfitting. For our method, we set the density ratio to \\(\\kappa = 0.05\\%\\), and we update the gradient sparsity support every \\(T = 10\\) iterations. We train until convergence, defined as reaching a loss threshold of 0.01. If convergence is not achieved, we stop" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.52, + 0.825, + 0.535 + ], + "angle": 0, + "content": "training at a maximum limit of 2000 iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Following CLIP-LoRA [45], we adopt a learning rate of \\(2 \\times 10^{-4}\\) with a cosine learning rate scheduler and a batch size of 32. We use the simple prompt template \"a photo of a [class name]\" without complex manual prompt engineering. All experiments are conducted using a ViT-B/16 backbone, and finetuning is applied to all layers of the 12 vision and text Transformer blocks. These settings were suggested by CLIP-LoRA and are kept the same for all methods, including ours. All hyperparameters remain fixed across all datasets for our approach and the baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.884 + ], + "angle": 0, + "content": "All experiments are conducted under identical hardware and software conditions. The details of these environments are provided in Appendix E. Additional experimental results showing CLIP's few-shot adaptation performance are presented in Appendix F. A comparative analysis of the memory efficiency of SO and state-of-the-art low-rank methods is presented in Appendix C. The sensitivity of our approach to the hyperparameters \\(\\kappa\\) and \\(T\\) is explored in Appendix D. Moreover, to further assess the effectiveness of our optimizer, we conduct both standard classification and few-shot learning experiments using a simple two-layer fully connected network. The results obtained with this small architecture are detailed in Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "(b) Sparsity Outperforms Low-Rank Projections. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3149" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.303, + 0.09, + 0.694, + 0.104 + ], + "angle": 0, + "content": "Table 5. Impact of moments in few-shot classification (ViT-B/16)." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.115, + 0.896, + 0.259 + ], + "angle": 0, + "content": "
ShotsMethodMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
withoutwith
1SO✓✗20.974.751.976.585.782.992.554.570.267.8
SO✗✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✗24.984.660.277.589.590.992.961.176.073.0
SO✗✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✗32.289.172.179.089.193.593.864.477.376.7
SO✗✓38.687.778.985.392.495.195.566.483.480.4
" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.553 + ], + "angle": 0, + "content": "results in Table 1 show the effectiveness of SO compared to state-of-the-art low-rank techniques. Across all shots (1, 2, and 4), our method achieves the highest average performance. Furthermore, SO yields the highest accuracy in most datasets, including EuroSAT, Cars, Flowers, and UCF101. Notably, it outperforms LoRA and ReLoRA on Aircraft by significant margins. Low-rank reparametrization methods constrain updates to a fixed subspace. The only exception is ReLoRA, which periodically merges and reinitializes the auxiliary low-rank matrices; this might explain why ReLoRA outperforms the other low-rank reparametrization methods. Moreover, these methods suffer from severe overfitting and require careful tuning of the rank and number of training iterations to achieve their highest performance. Unlike these methods, SO employs dynamic sparsity to update only a minimal subset of parameters, thereby mitigating overfitting while preserving the model's full representational capacity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.688 + ], + "angle": 0, + "content": "(d) Effectiveness of Pruning in Few-Shot Optimization. Table 2 shows the impact of gradient and moment pruning on few-shot classification. Dense optimization (Adam) performs poorly in the few-shot setting due to severe overfitting. In contrast, gradient pruning leads to substantial performance improvement. Moment pruning further enhances performance by filtering out updates related to parameters that are only locally important. These findings validate that sparsity can improve generalization in few-shot learning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.901 + ], + "angle": 0, + "content": "(e) Local Randomness Global Importance. Table 3 presents the performance of four sparse optimization variants based on randomness and importance in both gradient and moment pruning. Importance-based pruning is achieved by retaining the values with the highest magnitude. The results indicate that applying random selection for gradient pruning and importance-based selection for moment pruning achieves the highest average performance across all few-shot settings. Results in Appendix F confirm that importance-based gradient selection leads to rapid overfitting. Furthermore, Table 3 shows that importance-based moment pruning outperforms random moment pruning. This is because significant parameters over the long term are consistently updated, while locally influential ones" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.905, + 0.632 + ], + "angle": 0, + "content": "are filtered out from the moment updates. Overall, our results validate the local randomness and global importance paradigm. Stochastic gradient selection prevents bias toward high-magnitude local updates and importance-based pruning preserves influential parameters over the long term. (f) Local Sparsity Global Density Table 4 shows that dynamic sparsity support consistently outperforms fixed sparsity support in gradient pruning across all few-shot settings. This improvement validates the local sparsity and global density paradigm. On the one hand, dynamic sparsity ensures that few parameters contribute to the updates to reduce the local learning capacity and thus mitigate overfitting. On the other hand, dynamic sparsity ensures a broader exploration of the parameter space while making the learning process benefit from the model's overall expressiveness. (g) Role of Moments in Few-Shot Adaptation. Table 5 shows that moment-based updates significantly improve performance. The advantage is more pronounced in datasets with higher complexity, such as Aircraft and Cars. Without moments, updates rely solely on randomly selected gradient values. These results provide evidence that gradient pruning is not sufficient. Random gradient pruning is more effective when combined with moments, particularly sparse ones." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.634, + 0.664 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this paper, we shed light on sparsity as an effective mechanism to mitigate overfitting in few-shot adaptation of VLMs. Conventional low-rank methods suffer from severe overfitting and require extensive hyperparameter tuning. To address these limitations, we introduce two key paradigms. The first paradigm consists of dynamically selecting a minimal subset of parameters to reduce overfitting while preserving the overall model expressiveness. The second paradigm aims to prevent bias in gradient updates while retaining critical long-term information. Our experiments confirm that sparsity outperforms low-rank projections in few-shot learning and constitutes a strong mechanism against overfitting. As future work, we plan to evaluate our optimizer on large language models. Memory efficiency in this setting is even more critical." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3150" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Digital Research Alliance of Canada." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Randall Balestriero, Leon Bottou, and Yann LeCun. The effects of regularization and data augmentation are class dependent. In Advances in Neural Information Processing Systems (NeurIPS), pages 37878-37891, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[2] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 1-9, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446–461, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[4] Tianyi Chen, Tianyu Ding, Badal Yadav, Ilya Zharkov, and Luming Liang. Lorashear: Efficient large language model structured pruning and knowledge recovery. arXiv preprint arXiv:2310.18356, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[5] Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 10558-10578, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.591 + ], + "angle": 0, + "content": "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[9] Lucio Dery, Steven Kolawole, Jean-François Kagy, Virginia Smith, Graham Neubig, and Ameet Talwalkar. Everybody prune now: Structured pruning of llms with only forward passes. arXiv preprint arXiv:2402.05406, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 178–178, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[12] Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, and Bamdev Mishra. SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[13] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.362, + 0.904, + 0.43 + ], + "angle": 0, + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[17] Iiro P Jaëskeläinen, Enrico Glerean, Vasily Klucharev, Anna Shestakova, and Jyrki Ahveninen. Do sparse brain activity patterns underlie human cognition? NeuroImage, 263: 119633, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[20] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[21] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M Asano. VeRA: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops (WICCV), pages 554–561, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] François Lagunas, Ella Charlaix, Victor Sanh, and Alexander Rush. Block pruning for faster transformers. In Empirical Methods in Natural Language Processing (EMNLP), pages 10619-10629, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Vladislav Lialin, Sherin Muckatira, Namrata Shivagunde, and Anna Rumshisky. Relora: High-rank training through low-rank updates. In International Conference on Learning Representations (ICLR), 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Cheng Lin, Lujun Li, Dezhi Li, Jie Zou, Wei Xue, and Yike Guo. Nora: Nested low-rank adaptation for efficient" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "fine-tuning large models. arXiv preprint arXiv:2408.10280, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), pages 32100-32121, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning (ICML), pages 23033-23044, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[28] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. In *British Machine Vision Conference (BMVC)*, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[30] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. PiSSA: Principal singular values and singular vectors adaptation of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing (ICVGIP), pages 722-729, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[34] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[36] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In International Conference on Learning Representations (ICLR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[37] Xianghui Sun, Yunjie Ji, Baochang Ma, and Xiangang Li. A comparative study between full-parameter and lora-based fine-tuning on chinese instruction data for instruction following large language model. arXiv preprint arXiv:2304.08109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. In Advances in Neural" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Information Processing Systems (NeurIPS), pages 24193-24205, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[40] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In Annual Meeting of the Association for Computational Linguistics (ACL), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[41] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3485-3492, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[42] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In Empirical Methods in Natural Language Processing (EMNLP), pages 9514-9528, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[43] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. CorDA: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10899-10909, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[45] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 1593-1603, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[46] Zhi Zhang, Qizhe Zhang, Zijun Gao, Renrui Zhang, Ekaterina Shutova, Shiji Zhou, and Shanghang Zhang. Gradient-based parameter selection for efficient fine-tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28566-28577, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[47] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient llm training by gradient low-rank projection. In International Conference on Machine Learning (ICML), 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2605-2615, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3152" + } + ] +] \ No newline at end of file diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_origin.pdf b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a0ac98bf7c7fa8da0db23aee6acc8a4fa33c514 --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/6c4b6844-e0ee-4952-860e-16ecf63a3ae1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3e55c472bd3ef9d8a60c6126eb87867b405989f642ab5267aa4a00a21fae9d1 +size 5885756 diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/full.md b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95fafef1814194b26748f7e228cc973318a4b50a --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/full.md @@ -0,0 +1,269 @@ +# Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation + +Nairouz Mrabah Nicolas Richet Ismail Ben Ayed Eric Granger LIVIA, ILLS, Department of Systems Engineering, ETS Montreal, Canada + +mrabah.nairouz@livia.etsmt1.ca, nicolas.richet.l@ens.etsmt1.ca, ismail.benayed@etsmt1.ca, eric.granger@etsmt1.ca + +# Abstract + +Adapting vision-language models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for local sparsity and global density, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for local randomness and global importance, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead. Code is available at: https://github.com/nairouz/SO. + +# 1. Introduction + +The performance of deep learning (DL) models is strongly influenced by their size, with larger architectures yielding superior results. For instance, the ViT-H/14 backbone outperforms ViT-B/32 by a significant margin across multiple benchmarks [6]. Despite recent architectural advancements, the best models remain highly data-dependent and require large-scale labeled datasets to achieve reliable generalization. While techniques such as data augmentation and regularization help mitigate overfitting in low-data scenarios, they are insufficient to fully resolve it [1]. In contrast, the human brain has more than 80 billion neurons. Each neuron forms, on average, only about 7,000 synaptic connec + +tions, which results in an estimated total of 100 to 500 trillion synapses. Despite its tremendous capacity, the brain exhibits a remarkable ability to grasp new concepts from minimal exposure. This one-shot or few-shot learning capability might be partly attributed to the effective use of sparsity [17]. By selectively activating a fraction of connections at any given time, the brain optimizes both energy consumption and computational efficiency. + +Among the prominent DL models, vision-language models (VLMs) learn visual concepts from natural language descriptions. In particular, the contrastive language-image pretraining (CLIP) [33] model is pretrained on large-scale datasets containing hundreds of millions of image-text pairs to align visual and textual embeddings within a shared semantic space. After training, VLMs show strong zero-shot capabilities across diverse visual recognition tasks [33]. Moreover, fine-tuning VLMs yields state-of-the-art performance in various domain adaptation scenarios [27, 39]. Despite their zero-shot capabilities, adapting VLMs to new tasks with small labeled datasets remains challenging due to pronounced overfitting and catastrophic forgetting problems. Furthermore, training such models requires vast resources, including significant GPU memory for storing millions of parameters, gradients, activations, and optimizer states. These challenges underscore the need for cost-effective optimization strategies that address overfitting. + +To address the limitations of full VLM fine-tuning in few-shot learning, state-of-the-art methods leverage parameter-efficient training (PET) [15, 16, 21]. This strategy consists of optimizing a small number of parameters compared to the full set of weights. Additive PET methods freeze the initial weights and attach trainable and unmergeable parameters to the model. Within this category, we find the adapter [11] and the prompt tuning variants [49]. The prevailing methods in few-shot VLMs have mainly focused on this category. Adjustable weights can be incorporated at various levels, either by embedding them directly into the input space as visual [18] or textual prompts [48], or by integrating adapters throughout the network architecture [50]. + +To mitigate overfitting in few-shot adaptation and pre + +![](images/998282f275fb5f80ced8c8975a2fe329bee4b2d3984092db3a153167a1ba6eef.jpg) + +![](images/9c1e31495eeba8268f0f1be5f1d1e910acc283015936305e56824803b6132b55.jpg) + +![](images/dd7189c5fb0e3252bf1035cbaf77742f34f9dbda8a8c2b54be96d5270f38c8c5.jpg) + +![](images/2c1248125ba560a9c08a7d18e181648eeedfd096700b2004eef961c29d8fbca5.jpg) + +![](images/bea01f790e9e1b4ad400c75b69eb5d83d0a3f1dd4d6b7a96ecea4496cafca844.jpg) +Figure 1. Test-set accuracy over training iterations for LoRA (with rank fixed at 3) and SO (with a sparsity ratio equal to $99.95\%$ ) in a 1-shot setting on four datasets (Pets, Flowers, Caltech, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. Training stops when the loss falls below 0.01 or after 2000 iterations. + +![](images/452e6c558c4b75c38d72e1227c9494fad1e750d7f5fc3002ebda03a024c0f17f.jpg) + +![](images/53b4e9a77f4fb9bdaed33f0d375b7d53cc33afb4eec866f4dc11d629b952e4ce.jpg) + +![](images/af8f89d3882d67bcc788316703ca6fbcf88659c34eb898ce2b39a34441366178.jpg) + +![](images/5218af9e9e4c7ca5c47bbbc554e75de01d4acddcebc631a7b0c8222cd4fd0693.jpg) + +![](images/2e559f2557e3b49eee358e923f88b9e8edb1b1105f44fdef72fc9d2db39947dd.jpg) + +![](images/0c9f5f5abccb0862c1bbd83c7dc801e35441702214c27b372be0898ad6c7477e.jpg) + +![](images/5b049a48fbd3758252415ed14fafab3997364ea8e5417b1c1f8be1da00f4a77f.jpg) + +![](images/839bb7ad23c06b203f23082e233bc9504eff49688ff54685c8431df5df178874.jpg) + +![](images/7d0bb11896288fd91b0405849c9010a32bd9c8c63ff7e03dc8acfc516c302a54.jpg) + +![](images/22368532aeabaa75da0760d7dbe426cd4077df6b388fc62b89285f71013d898b.jpg) + +![](images/653ac7451982321245105d471c82041e6c9935cff9bc915820edaf67285a88c1.jpg) + +![](images/1c37137fec9706b698d853b63f8e8b92887fcc698994fd0b701f28dd1e44e84c.jpg) +Figure 2. Test-set accuracy of LoRA in a 1-shot setting on three datasets (DTD, Oxford Pets, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. We vary the LoRA rank from 2 to 5 and train for at most 2000 iterations or until the loss drops below 0.01. Each subfigure shows test accuracy over training iterations for a specific rank. + +![](images/b0eec707c9caa65024f5b46af69bd6150e07d76842f0b0555575c617c03e88ad.jpg) + +![](images/65974d4476f4e198220419948cc9bb79554e39d06f6931ca6456c4f6b91932b5.jpg) + +![](images/dd46951795c2b021fcd660832629a4ce2a989c4f96164f8b14d021a989701f10.jpg) + +serve the prior knowledge of VLMs, adapter-based methods typically employ three techniques. First, these small trainable modules are only integrated into the last layer of VLMs. Second, they use regularization techniques to penalize the divergence between the initial zero-shot class prototypes and the learnable class prototypes. Finally, they initialize the learnable class prototypes into a reliable region using the zero-shot prototypes. As a result, these methods have shown better results than the prompt tuning approaches [44], which often introduce computational overhead [45]. Despite these efforts, adapter-based methods are still prone to considerable overfitting [34]. Their success depends heavily on carefully tuning hyperparameters whose optimal values vary significantly across different tasks [34]. This sensitivity leads to performance degradation that undermines the generalizability of these approaches. It is worth noting that tuning these hyperparameters requires extensive grid searching with a validation set, which is not available in real-world few-shot scenarios. + +Recently, low-rank reparametrization [16, 20, 26] has gained popularity for domain adaptation. These approaches reparameterize the initial weights with supplementary trainable and mergeable low-rank matrices. Low-rank adaptation (LoRA) [16] was the first to apply this strategy. It reformulates weight matrices $\mathbf{W}$ as $\mathbf{W} + \mathbf{A} \cdot \mathbf{B}$ , where $\mathbf{A}$ and $\mathbf{B}$ are low-rank matrices. As a key advantage, this technique helps reduce overfitting and catastrophic forgetting because of two properties. First, LoRA enables fine-grained parallel adaptation by learning small perturbations to the frozen weights [13]. Second, the effective initialization of $\mathbf{A}$ and $\mathbf{B}$ helps in preserving prior knowledge [30, 43]. + +Building on these advantages, recent studies [25, 45] show that low-rank reparameterization methods achieve state-of-the-art performance in few-shot adaptation of VLMs. However, these methods also have notable limitations. Compelling evidence suggests that restricting weight updates to low-rank subspaces reduces model expressiveness [24, 37, 47]. In this work, we argue that the low-rank projection technique exhibits severe overfitting in few-shot adaptation. Our systematic study of LoRA confirms this issue (see Fig. 1). Test-set accuracy initially improves but then declines sharply. At convergence, when the loss approaches zero, performance deteriorates significantly compared to its peak on test-set data. Moreover, we can see that the optimal number of training iterations varies widely across datasets. However, tuning hyperparameters, including the number of iterations for early stopping, on a validation set contradicts the principles of few-shot learning. + +The challenges of low-rank reparameterization methods extend beyond fixing the number of training iterations. Their performance is also highly sensitive to another hyperparameter: the rank of the low-rank projection. This sensitivity is shown in Fig. 2, which illustrates LoRA's perfor + +mance for different ranks on three datasets (DTD, Oxford Pets, UCF101) in a 1-shot setting. Each dataset reveals a distinct best rank, and higher ranks sometimes lead to an early spike in accuracy but then degrade as training continues. Further experiments examining these behaviors are presented in Appendix F of the supplementary material. + +Since rank is a discrete value that cannot be lower than 1, the low-rank methods lack the flexibility to adjust very few parameters. While setting the rank to a very low value (e.g., 1 or 2) can help mitigate overfitting, this choice might reduce the model's learning capacity. This limitation arises because the trainable parameters in low-rank methods remain fixed throughout training, potentially leading to underfitting the VLM capacity. Indeed, rank selection involves a trade-off between underfitting and overfitting (see Fig. 2). For instance, in the 1-shot DTD setting, LoRA at rank 5 briefly exceeds rank 2 but then drops below it as training continues. Overall, the test accuracy of LoRA exhibits oscillations and varies unpredictably w.r.t. rank and the number of training iterations. This instability makes it difficult to reliably tune these interdependent hyperparameters without access to a validation set. + +In this paper, we deviate from the current long-standing trend of low-rank adaptation and introduce two new paradigms that make sparsity work in few-shot settings. Our approach is a simple yet effective optimization strategy, referred to as sparse optimization (SO), and aligns with these paradigms. As a first paradigm, we advocate for local sparsity and global density. Unlike the conventional low-rank approaches that update a fixed set of parameters, SO dynamically updates a minimal subset of original connections at each step. Specifically, high sparsity is enforced in both the gradient and moment updates while allowing the sparsity support to evolve dynamically throughout training. This paradigm ensures that, at any given iteration, only a minimal subset of parameters is updated to reduce the local learning capacity and prevent rapid overfitting. By dynamically changing the sparsity support, the model retains its overall expressiveness and ensures that different parameters contribute to learning over time. + +As a second paradigm, we advocate for local randomness and global importance. Similar to Adam [19], we leverage the gradient and the first and second moments. However, we sparsify these components in a way that addresses overfitting and ensures efficiency in terms of memory consumption. In particular, we sparsify the gradient randomly and the first moment by importance. The gradient captures local and iteration-specific information. The first moment aggregates the gradients over the whole path and, thus, reflects long-term parameter importance. Random pruning of the gradient prevents the model from relying too much on short-term and local high-magnitude updates. The importance-based selection of the first moment + +ensures that connections with long-term significance are updated. By emphasizing local randomness and global importance, our strategy avoids severe overfitting. + +Concretely instantiating this paradigm, we fix a density rate $\kappa$ and retain $M$ elements for both the gradient and moment updates at each iteration. First, we randomly select $M$ gradient values from the full set of parameter updates. Next, we compute the temporary first and second moments, which can contain up to $2M$ values in the worst-case scenario where no overlap exists between the selected gradients and the previous $M$ first moment values. We rank the $2M$ first moment values by importance using their magnitude. Only the top $M$ values are retained for the first moment. We then use the same indices to select the corresponding values for the second moment. This ensures alignment between the sparsified first and second moments. The selection strategy preserves the most relevant long-term updates while introducing randomness in the local updates. + +Our contributions. (1) A novel framework is proposed that sparsifies both gradients and moments to improve generalization and mitigate overfitting in few-shot learning. Our approach differs from conventional low-rank reparameterization by focusing on sparsity. (2) We introduce the local sparsity and global density paradigm, where a dynamically evolving sparsity support updates only a minimal subset of model parameters at each iteration. This strategy preserves the overall model expressiveness while reducing its risk of overfitting. (3) We propose the local randomness and global importance paradigm, which selectively sparsifies the gradient using random selection while pruning the first moment based on importance. This selection strategy preserves the most relevant long-term updates while introducing local randomness to mitigate overfitting. (4) Our comprehensive empirical study on 11 diverse datasets compares our method to state-of-the-art low-rank approaches. Results indicate that SO can significantly improve few-shot generalization while achieving memory efficiency. + +# 2. Related Work + +(a) Low-Rank Methods. LoRA [16] approximates weight updates by decomposing them into two low-rank matrices. As a result, it reduces the number of trainable parameters while preserving the model's frozen pretrained backbone. DoRA [26] introduces a weight decomposition approach to improve LoRA's expressiveness by decoupling weight magnitude and direction and training them separately. VeRA [21] reduces the set of trainable parameters further than LoRA. In particular, VeRA freezes the pair of low-rank matrices to randomly initialized ones and only trains two small scaling vectors. PiSSA [30] initializes the two low-rank matrices using the most informative singular vectors and their corresponding singular values from the original weight matrix while freezing the less significant compo + +nents in a residual matrix. ReLoRA [24] achieves high-rank weight updates by merging trained low-rank updates into the main model parameters at regular intervals and reinitializing the auxiliary low-rank matrices. Unlike these LoRA-based reparametrization methods, which constrain weight updates to low-rank subspaces, GaLore [47] performs full-parameter training. Specifically, GaLore projects gradients into a low-rank subspace to reduce memory consumption while allowing full updates to the model weights. + +Despite their efficiency, these methods still face major limitations. Most low-rank techniques impose a fixed-rank constraint on the updates, which restricts adaptation flexibility and may lead to underfitting when the rank is too low. Even with minimal rank, these methods struggle with generalization in few-shot adaptation due to severe overfitting. Furthermore, these approaches require careful hyperparameter tuning, particularly in selecting the optimal rank and training iterations, which significantly impact performance. Our work challenges this strategy by showing that appropriately designed sparsity-based adaptation provides superior generalization while reducing memory consumption. + +(b) Sparsity-Based Methods. These methods update only a subset of model parameters. The selection can be structured or unstructured. Structured selection methods [4, 9, 23, 28, 40] target entire modules (e.g., attention heads, layers), offering coarse-grained sparsity but often requiring architecture-specific designs that may degrade performance [5]. Unstructured selection methods [2, 12, 36, 38, 42, 46] prune individual weights for greater flexibility. However, most unstructured methods rely on static sparsity patterns and perform importance-based gradient pruning. Moreover, none of these methods incorporates sparse moments. + +In contrast, SO updates a minimal subset of parameters at each step. Furthermore, our approach uses random gradient selection to avoid local high-magnitude biases and importance-based moment pruning to retain long-term relevance. This combination enables stable adaptation in few-shot scenarios without sacrificing model expressiveness. + +# 3. Proposed Sparse Optimization Method + +A comprehensive overview of few-shot adaptation in VLMs is provided in Appendix A. Briefly, we leverage a pretrained CLIP [33] model that aligns images and texts in a shared embedding space via contrastive learning. After pretraining, CLIP can perform zero-shot classification by comparing image embeddings to class-specific text prototypes. In the few-shot scenario, we only have a small support set of labeled examples (e.g., $K$ images per class). We adapt CLIP to this support set by minimizing the cross-entropy loss with respect to the class prototypes or learnable parameters, and we optimize the loss function using our proposed approach. + +Motivated by the challenges inherent in few-shot adaptation, we introduce SO, a new optimizer designed to en + +hance memory efficiency while mitigating overfitting. Unlike standard optimization approaches that update all parameters at each iteration, SO dynamically updates a subset of parameters. This sparsification significantly reduces memory consumption, which makes SO suitable for training large-scale deep learning models like VLMs. + +Our approach builds on the Adam optimizer and behaves identically to it when the sparsity ratio is set to zero. However, SO deviates from Adam by following two paradigms. The first paradigm is local sparsity and global density, which ensures that a minimal subset of parameters is dynamically updated while maintaining overall model expressiveness. The second paradigm is local randomness and global importance, which applies random pruning to the gradient updates while retaining the most significant moment estimates. These principles prevent severe overfitting and make SO suitable for scenarios like few-shot learning. + +Notation. Let $\Theta \in \mathbb{R}^d$ represent the model parameters, where $d$ is the total number of parameters. At each iteration $t$ , the model updates some parameters to minimize a given loss function $\mathcal{L}(\Theta_t)$ . The gradient of the loss concerning all parameters is denoted as $g_{t} = \nabla_{\Theta}\mathcal{L}(\Theta_{t})$ . We use the first and second moments, denoted as $\mu_t$ and $\nu_{t}$ respectively, to capture long-term and global parameter significance throughout the optimization process. A small constant $\epsilon$ is introduced for numerical stability. The learning rate, denoted as $\eta$ , scales the parameter updates. We introduce a density ratio $\kappa \in [0,1]$ , which determines the fraction of gradient entries retained at each step. Let $\mathcal{I}(\cdot)$ and $\mathcal{V}(\cdot)$ be the functions that return the indices and values of a sparse input vector, respectively. Instead of updating all parameters, SO retains only a small subset of the gradient and generates sparse first and second moments. + +We enforce sparsity through two operations. The Top- $M$ operation selects $M$ elements with the highest magnitude. In contrast, Random- $M$ retains $M$ elements uniformly at random, preventing bias toward high-magnitude values. + +Dynamic and High-Sparsity Support. Conventional optimization methods update the full set of parameters throughout training, which accelerates overfitting. In contrast, our method dynamically updates a minimal subset of parameters. The high sparsity and dynamic selection provide two key advantages. First, updating a minimal subset of parameters reduces the local learning capacity and thus mitigates overfitting and catastrophic forgetting. Second, continuously changing the sparsity support ensures that different parameters receive updates throughout training. This avoids excessive reliance on a fixed set and makes the learning process benefit from the model's overall expressiveness. + +Randomness in Gradient Pruning. We sparsify the gradient vector to mitigate overfitting. The proposed optimizer applies a stochastic sampling mechanism that uniformly picks $M$ elements from the gradient $g_{t}$ . Given the density + +ratio $\kappa$ , we define the number of retained gradient entries as $M = \lfloor \kappa d \rfloor$ . Formally, we have $\tilde{g}_t = \text{Random-} M(g_t)$ . + +The motivation is to avoid biasing updates toward high-magnitude gradients. Enforcing stochastic sparsity prevents the model from over-relying on seemingly important local information. Our empirical results indicate that pruning the gradient by retaining only the largest values accelerates overfitting and significantly degrades performance. + +Importance in Moment Pruning. To maintain efficiency, our goal is to construct sparse first and second moments, $\tilde{\mu}_t$ and $\tilde{\nu}_t$ , each of size $M$ . Since the sparse gradient changes its support dynamically, storing historical information becomes challenging. Without pruning, the first and second moments can become dense over time due to the accumulation of gradients with varying sparsity support. + +To address this issue, we introduce two fixed-size buffers that can store up to $2M$ values. This ensures that both new and past gradient values are considered. Let $\mu_t$ and $\nu_t$ be the temporary first and second moments, at iteration $t$ : + +$$ +\mu_ {t} = \beta_ {1} \tilde {\mu} _ {t - 1} + (1 - \beta_ {1}) \tilde {g} _ {t}, \tag {1} +$$ + +$$ +\nu_ {t} = \beta_ {2} \tilde {\nu} _ {t - 1} + (1 - \beta_ {2}) \tilde {g} _ {t} ^ {2}. \tag {2} +$$ + +These moments may contain up to $2M$ values in the worst-case scenario when there is no overlap between the newly selected sparse gradient $\tilde{g}_t$ and the previously retained sparse first moment $\tilde{\mu}_t$ . To ensure stable updates, we employ a Top- $M$ selection strategy for the first moment, where we rank the $2M$ values by magnitude and retain only the top $M$ ones. This guarantees that long-term significant parameters are consistently updated. For the second moment, we enforce alignment by retaining elements corresponding to the same indices selected for the first moment. The sparse first and second moments are expressed as: + +$$ +\tilde {\mu} _ {t} = \operatorname {T o p} - M (\mu_ {t}), \quad \tilde {\nu} _ {t} = \nu_ {t} [ \mathcal {I} (\tilde {\mu} _ {t}) ]. \tag {3} +$$ + +Both sparse moments $\tilde{\mu}$ and $\tilde{\nu}$ are initialized at zero. Similar to Adam, we apply bias correction to $\tilde{\mu}_t$ and $\tilde{\nu}_t$ before parameter update. The corrected sparse first and second moments, denoted by $\hat{\mu}_t$ and $\hat{\nu}_t$ , are expressed as follows: + +$$ +\hat {\mu} _ {t} = \frac {\tilde {\mu} _ {t}}{1 - \beta_ {1} ^ {t}}, \quad \hat {\nu} _ {t} = \frac {\tilde {\nu} _ {t}}{1 - \beta_ {2} ^ {t}}. \tag {4} +$$ + +Finally, the parameters $\Theta$ are updated using the corrected sparse first and second moments, $\hat{\mu}_t$ and $\hat{\nu}_t$ , as provided in: + +$$ +\Theta_ {t + 1} = \Theta_ {t} - \frac {\eta}{\sqrt {\hat {\nu} _ {t}} + \epsilon} \hat {\mu} _ {t}, \tag {5} +$$ + +where $\eta$ is the learning rate, and $\beta_{1}$ and $\beta_{2}$ are decay rates. Algorithm. Our algorithm is provided in Appendix B due to space limitations. We finetune CLIP until convergence using our optimization technique. To promote efficiency, + +Table 1. Few-shot classification performance on 11 datasets with ViT-B/16 backbone. Top-1 accuracy averaged over 3 random seeds is reported. Highest value is highlighted in bold, and the second highest is underlined. + +
ShotsMethodImageNetSUNAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
0CLIP (ICML '21)66.762.624.747.565.386.189.171.492.943.666.765.1
1Adam (ICLR '15)0.10.61.024.30.51.02.91.41.05.61.73.6
LoRA (ICLR '22)67.367.025.067.568.281.290.585.792.352.472.970.0
ReLoRA (ICLR '24)70.369.728.873.870.584.291.985.193.454.276.072.5
GaLoRE (ICML '24)65.165.521.165.263.674.185.280.591.448.969.966.4
PiSSA (NeurIPS '24)64.965.222.266.864.275.784.882.990.750.570.167.1
DoRA (ICML '24)66.966.925.371.468.381.389.386.292.351.972.970.2
VeRA (ICLR '24)68.467.327.569.367.486.291.073.193.952.270.169.7
SO (Ours)70.170.331.578.271.686.293.384.994.155.376.473.8
2Adam (ICLR '15)0.20.80.942.10.71.93.917.13.17.23.87.4
LoRA (ICLR '22)67.468.330.381.970.179.389.490.993.759.676.173.4
ReLoRA (ICLR '24)70.670.532.283.872.982.690.390.794.760.479.175.3
GaLoRE (ICML '24)66.167.025.882.267.272.681.585.691.556.173.469.9
PiSSA (NeurIPS '24)65.166.726.177.367.174.884.487.491.957.473.370.1
DoRA (ICML '24)67.268.229.883.470.379.489.691.293.659.775.973.5
VeRA (ICLR '24)69.469.031.781.669.886.792.679.694.459.374.773.5
SO (Ours)70.572.337.282.774.485.392.291.995.360.280.476.6
4Adam (ICLR '15)0.12.12.447.90.73.16.330.512.59.69.211.3
LoRA (ICLR '22)68.569.735.285.274.878.787.994.193.963.578.275.4
ReLoRA (ICLR '24)71.272.237.684.777.182.390.394.194.964.080.077.1
GaLoRE (ICML '24)67.569.034.182.173.175.281.091.193.260.075.472.9
PiSSA (NeurIPS '24)66.468.832.482.873.275.680.692.092.961.175.772.9
DoRA (ICML '24)68.469.734.985.575.179.089.594.194.064.178.575.7
VeRA (ICLR '24)70.571.535.685.673.786.793.189.895.364.579.576.9
SO (Ours)71.473.738.687.778.985.392.495.195.566.483.478.9
+ +Table 2. Impact of gradient and moment sparsification in few-shot classification (ViT-B/16). + +
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
DenseSparseDenseSparse
1SO (≡ Adam)✓✗✓✗1.024.30.51.02.91.41.05.61.74.4
SO✗✓✓✗29.473.969.683.993.286.493.755.777.073.6
SO✗✓✗✓31.578.271.686.293.384.994.155.376.474.6
2SO (≡ Adam)✓✗✓✗0.942.10.71.93.917.13.17.23.89.0
SO✗✓✓✗29.482.671.081.591.490.193.759.778.475.3
SO✗✓✗✓37.282.774.485.392.291.995.360.280.477.7
4SO (≡ Adam)✓✗✓✗2.447.90.73.16.330.512.59.69.213.6
SO✗✓✓✗34.585.473.578.991.192.794.266.278.377.2
SO✗✓✗✓38.687.778.985.392.495.195.566.483.480.4
+ +we retain the gradient sparsity support fixed for $T$ iterations before reselecting the trainable parameters. Overall, our approach introduces two hyperparameters: the density ratio $\kappa$ , which controls the fraction of retained gradient entries, and the update interval $T$ , which determines how frequently the sparsity support is refreshed. A smaller $\kappa$ value improves generalization but results in slower convergence. In contrast, a larger $\kappa$ accelerates optimization at the cost of reduced performance. Similarly, reducing $T$ leads to better results by enabling more frequent updates to the sparsity support, although increasing $T$ speeds up convergence. + +The role of $T$ in our method is conceptually similar to update intervals used in prior work such as GaLore [47] + +and ReLoRA [24]. In ReLoRA, the low-rank matrices are merged with the original parameters every $T$ interval. GaLore also updates the low-rank projectors every $T$ interval. Unlike the low-rank methods, where hyperparameters such as rank and number of training iterations involve a strong trade-off between overfitting and underfitting, our sparsity-based approach remains robust to overfitting. + +# 4. Results and Discussion + +(a) Experimental Methodology. Our optimizer is evaluated on 11 datasets commonly used in few-shot VLM adaptation [45, 49], including ImageNet [8], SUN397 [41], Air + +Table 3. Impact of randomness and importance in gradient and moment sparsification in few-shot classification (ViT-B/16). + +
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
RandImptRandImpt
1SOX✓X✓17.673.960.572.087.975.989.953.968.266.6
SOX✓✓X20.771.463.774.692.079.991.954.671.169.6
SO✓X✓X31.271.769.885.693.380.394.155.074.972.9
SO✓XX✓31.578.271.686.293.384.994.155.376.474.6
2SOX✓X✓21.480.862.667.782.884.688.757.668.268.3
SOX✓✓X22.882.465.373.487.184.291.360.374.571.3
SO✓X✓X33.282.172.885.792.191.294.860.979.676.9
SO✓XX✓37.282.774.485.392.291.995.360.280.477.7
4SOX✓X✓28.283.267.966.969.886.489.055.969.962.5
SOX✓✓X31.081.071.573.077.786.991.458.974.171.6
SO✓X✓X39.584.777.686.292.793.395.866.082.579.8
SO✓XX✓38.687.778.985.392.495.195.566.483.480.4
+ +Table 4. Impact of dynamic sparsity support for gradient pruning in few-shot classification (ViT-B/16). + +
ShotsMethodSparsity supportAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
StaticDynamic
1SO✓✘28.272.468.185.993.173.494.552.672.671.2
SO✘✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✘31.083.069.785.891.683.094.355.775.874.4
SO✘✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✘34.883.672.486.292.588.095.162.577.877.0
SO✘✓38.687.778.985.392.495.195.566.483.480.4
+ +craft [29], EuroSAT [14], Stanford-Cars [22], Food101 [3], OxfordPets [32], Flowers102 [31], Caltech101 [10], DTD [7], and UCF101 [35]. These datasets provide a comprehensive assessment across a variety of image domains. + +We evaluate SO against state-of-the-art low-rank methods, which have shown strong performance in finetuning Large Language Models and Vision Language models. Specifically, we compare our method against LoRA, ReLoRA, VeRA, DoRA, PiSSA, and GaLoRE. Our comparison also includes full fine-tuning using Adam to showcase the effect of severe overfitting. A discussion of these baselines, including their limitations, is provided in Sec. 2. + +For a fair comparison, we follow the same hyperparameter settings as CLIP-LoRA [45] for the low-rank baselines. In particular, we set the rank to 2, and we train for 500 iterations. Due to severe overfitting, we introduce an early stopping criterion for the low-rank methods. If the training loss reaches 0.01, training is halted. This is particularly necessary to obtain competitive results using the low-rank approaches for some datasets where the test accuracy peaks often within a few iterations before rapidly overfitting. For our method, we set the density ratio to $\kappa = 0.05\%$ , and we update the gradient sparsity support every $T = 10$ iterations. We train until convergence, defined as reaching a loss threshold of 0.01. If convergence is not achieved, we stop + +training at a maximum limit of 2000 iterations. + +Following CLIP-LoRA [45], we adopt a learning rate of $2 \times 10^{-4}$ with a cosine learning rate scheduler and a batch size of 32. We use the simple prompt template "a photo of a [class name]" without complex manual prompt engineering. All experiments are conducted using a ViT-B/16 backbone, and finetuning is applied to all layers of the 12 vision and text Transformer blocks. These settings were suggested by CLIP-LoRA and are kept the same for all methods, including ours. All hyperparameters remain fixed across all datasets for our approach and the baselines. + +All experiments are conducted under identical hardware and software conditions. The details of these environments are provided in Appendix E. Additional experimental results showing CLIP's few-shot adaptation performance are presented in Appendix F. A comparative analysis of the memory efficiency of SO and state-of-the-art low-rank methods is presented in Appendix C. The sensitivity of our approach to the hyperparameters $\kappa$ and $T$ is explored in Appendix D. Moreover, to further assess the effectiveness of our optimizer, we conduct both standard classification and few-shot learning experiments using a simple two-layer fully connected network. The results obtained with this small architecture are detailed in Appendix G. + +(b) Sparsity Outperforms Low-Rank Projections. The + +Table 5. Impact of moments in few-shot classification (ViT-B/16). + +
ShotsMethodMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
withoutwith
1SO✓✗20.974.751.976.585.782.992.554.570.267.8
SO✗✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✗24.984.660.277.589.590.992.961.176.073.0
SO✗✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✗32.289.172.179.089.193.593.864.477.376.7
SO✗✓38.687.778.985.392.495.195.566.483.480.4
+ +results in Table 1 show the effectiveness of SO compared to state-of-the-art low-rank techniques. Across all shots (1, 2, and 4), our method achieves the highest average performance. Furthermore, SO yields the highest accuracy in most datasets, including EuroSAT, Cars, Flowers, and UCF101. Notably, it outperforms LoRA and ReLoRA on Aircraft by significant margins. Low-rank reparametrization methods constrain updates to a fixed subspace. The only exception is ReLoRA, which periodically merges and reinitializes the auxiliary low-rank matrices; this might explain why ReLoRA outperforms the other low-rank reparametrization methods. Moreover, these methods suffer from severe overfitting and require careful tuning of the rank and number of training iterations to achieve their highest performance. Unlike these methods, SO employs dynamic sparsity to update only a minimal subset of parameters, thereby mitigating overfitting while preserving the model's full representational capacity. + +(d) Effectiveness of Pruning in Few-Shot Optimization. Table 2 shows the impact of gradient and moment pruning on few-shot classification. Dense optimization (Adam) performs poorly in the few-shot setting due to severe overfitting. In contrast, gradient pruning leads to substantial performance improvement. Moment pruning further enhances performance by filtering out updates related to parameters that are only locally important. These findings validate that sparsity can improve generalization in few-shot learning. + +(e) Local Randomness Global Importance. Table 3 presents the performance of four sparse optimization variants based on randomness and importance in both gradient and moment pruning. Importance-based pruning is achieved by retaining the values with the highest magnitude. The results indicate that applying random selection for gradient pruning and importance-based selection for moment pruning achieves the highest average performance across all few-shot settings. Results in Appendix F confirm that importance-based gradient selection leads to rapid overfitting. Furthermore, Table 3 shows that importance-based moment pruning outperforms random moment pruning. This is because significant parameters over the long term are consistently updated, while locally influential ones + +are filtered out from the moment updates. Overall, our results validate the local randomness and global importance paradigm. Stochastic gradient selection prevents bias toward high-magnitude local updates and importance-based pruning preserves influential parameters over the long term. (f) Local Sparsity Global Density Table 4 shows that dynamic sparsity support consistently outperforms fixed sparsity support in gradient pruning across all few-shot settings. This improvement validates the local sparsity and global density paradigm. On the one hand, dynamic sparsity ensures that few parameters contribute to the updates to reduce the local learning capacity and thus mitigate overfitting. On the other hand, dynamic sparsity ensures a broader exploration of the parameter space while making the learning process benefit from the model's overall expressiveness. (g) Role of Moments in Few-Shot Adaptation. Table 5 shows that moment-based updates significantly improve performance. The advantage is more pronounced in datasets with higher complexity, such as Aircraft and Cars. Without moments, updates rely solely on randomly selected gradient values. These results provide evidence that gradient pruning is not sufficient. Random gradient pruning is more effective when combined with moments, particularly sparse ones. + +# 5. Conclusion + +In this paper, we shed light on sparsity as an effective mechanism to mitigate overfitting in few-shot adaptation of VLMs. Conventional low-rank methods suffer from severe overfitting and require extensive hyperparameter tuning. To address these limitations, we introduce two key paradigms. The first paradigm consists of dynamically selecting a minimal subset of parameters to reduce overfitting while preserving the overall model expressiveness. The second paradigm aims to prevent bias in gradient updates while retaining critical long-term information. Our experiments confirm that sparsity outperforms low-rank projections in few-shot learning and constitutes a strong mechanism against overfitting. As future work, we plan to evaluate our optimizer on large language models. Memory efficiency in this setting is even more critical. + +# Acknowledgement + +This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Digital Research Alliance of Canada. + +# References + +[1] Randall Balestriero, Leon Bottou, and Yann LeCun. The effects of regularization and data augmentation are class dependent. In Advances in Neural Information Processing Systems (NeurIPS), pages 37878-37891, 2022. 1 +[2] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 1-9, 2022. 4 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446–461, 2014. 7 +[4] Tianyi Chen, Tianyu Ding, Badal Yadav, Ilya Zharkov, and Luming Liang. Lorashear: Efficient large language model structured pruning and knowledge recovery. arXiv preprint arXiv:2310.18356, 2023. 4 +[5] Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 10558-10578, 2024. 4 +[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2023. 1 +[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 7 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 6 +[9] Lucio Dery, Steven Kolawole, Jean-François Kagy, Virginia Smith, Graham Neubig, and Ameet Talwalkar. Everybody prune now: Structured pruning of llms with only forward passes. arXiv preprint arXiv:2402.05406, 2024. 4 +[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 178–178, 2004. 7 +[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 1 + +[12] Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, and Bamdev Mishra. SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4 +[13] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations (ICLR), 2021. 3 +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7 +[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 1 +[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 1, 3, 4 +[17] Iiro P Jaëskeläinen, Enrico Glerean, Vasily Klucharev, Anna Shestakova, and Jyrki Ahveninen. Do sparse brain activity patterns underlie human cognition? NeuroImage, 263: 119633, 2022. 1 +[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 1 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 3 +[20] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3 +[21] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M Asano. VeRA: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 1, 4 +[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops (WICCV), pages 554–561, 2013. 7 +[23] François Lagunas, Ella Charlaix, Victor Sanh, and Alexander Rush. Block pruning for faster transformers. In Empirical Methods in Natural Language Processing (EMNLP), pages 10619-10629, 2021. 4 +[24] Vladislav Lialin, Sherin Muckatira, Namrata Shivagunde, and Anna Rumshisky. Relora: High-rank training through low-rank updates. In International Conference on Learning Representations (ICLR), 2023. 3, 4, 6 +[25] Cheng Lin, Lujun Li, Dezhi Li, Jie Zou, Wei Xue, and Yike Guo. Nora: Nested low-rank adaptation for efficient + +fine-tuning large models. arXiv preprint arXiv:2408.10280, 2024.3 +[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), pages 32100-32121, 2024. 3, 4 +[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning (ICML), pages 23033-23044, 2023. 1 +[28] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4 +[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. In *British Machine Vision Conference (BMVC)*, 2013. 7 +[30] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. PiSSA: Principal singular values and singular vectors adaptation of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4 +[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing (ICVGIP), pages 722-729, 2008. 7 +[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 4 +[34] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3 +[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7 +[36] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In International Conference on Learning Representations (ICLR), 2024. 4 +[37] Xianghui Sun, Yunjie Ji, Baochang Ma, and Xiangang Li. A comparative study between full-parameter and lora-based fine-tuning on chinese instruction data for instruction following large language model. arXiv preprint arXiv:2304.08109, 2023. 3 +[38] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. In Advances in Neural + +Information Processing Systems (NeurIPS), pages 24193-24205, 2021. 4 +[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1 +[40] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In Annual Meeting of the Association for Computational Linguistics (ACL), 2022. 4 +[41] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3485-3492, 2010. 6 +[42] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In Empirical Methods in Natural Language Processing (EMNLP), pages 9514-9528, 2021. 4 +[43] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. CorDA: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3 +[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10899-10909, 2023. 3 +[45] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 1593-1603, 2024. 3, 6, 7 +[46] Zhi Zhang, Qizhe Zhang, Zijun Gao, Renrui Zhang, Ekaterina Shutova, Shiji Zhou, and Shanghang Zhang. Gradient-based parameter selection for efficient fine-tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28566-28577, 2024. 4 +[47] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient llm training by gradient low-rank projection. In International Conference on Machine Learning (ICML), 2024. 3, 4, 6 +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 1 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 1, 6 +[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2605-2615, 2023. 1 \ No newline at end of file diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/images.zip b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d262b45347d07f8ff91def540c9bdfe25afa0dc5 --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b131ec0724e147c773b3a9fe8691b8085588df89e8fc532ff9bed027f1efd8f +size 702711 diff --git a/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/layout.json b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..90492ab0c2884311ae82497dd101c81241bf6ec0 --- /dev/null +++ b/2025/Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation/layout.json @@ -0,0 +1,7513 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 90, + 103, + 520, + 121 + ], + "type": "text", + "content": "Sparsity Outperforms Low-Rank Projections in Few-Shot Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 144, + 487, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 144, + 487, + 171 + ], + "spans": [ + { + "bbox": [ + 123, + 144, + 487, + 171 + ], + "type": "text", + "content": "Nairouz Mrabah Nicolas Richet Ismail Ben Ayed Eric Granger LIVIA, ILLS, Department of Systems Engineering, ETS Montreal, Canada" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 174, + 475, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 174, + 475, + 199 + ], + "spans": [ + { + "bbox": [ + 133, + 174, + 475, + 199 + ], + "type": "text", + "content": "mrabah.nairouz@livia.etsmt1.ca, nicolas.richet.l@ens.etsmt1.ca, ismail.benayed@etsmt1.ca, eric.granger@etsmt1.ca" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 515 + ], + "type": "text", + "content": "Adapting vision-language models (VLMs) to new domains with few labeled samples remains a significant challenge due to severe overfitting and computational constraints. State-of-the-art solutions, such as low-rank reparameterization, mitigate these issues but often struggle with generalization and require extensive hyperparameter tuning. In this paper, a novel Sparse Optimization (SO) framework is proposed. Unlike low-rank approaches that typically constrain updates to a fixed subspace, our SO method leverages high sparsity to dynamically adjust very few parameters. We introduce two key paradigms. First, we advocate for local sparsity and global density, which updates a minimal subset of parameters per iteration while maintaining overall model expressiveness. As a second paradigm, we advocate for local randomness and global importance, which sparsifies the gradient using random selection while pruning the first moment based on importance. This combination significantly mitigates overfitting and ensures stable adaptation in low-data regimes. Extensive experiments on 11 diverse datasets show that SO achieves state-of-the-art few-shot adaptation performance while reducing memory overhead. Code is available at: https://github.com/nairouz/SO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "The performance of deep learning (DL) models is strongly influenced by their size, with larger architectures yielding superior results. For instance, the ViT-H/14 backbone outperforms ViT-B/32 by a significant margin across multiple benchmarks [6]. Despite recent architectural advancements, the best models remain highly data-dependent and require large-scale labeled datasets to achieve reliable generalization. While techniques such as data augmentation and regularization help mitigate overfitting in low-data scenarios, they are insufficient to fully resolve it [1]. In contrast, the human brain has more than 80 billion neurons. Each neuron forms, on average, only about 7,000 synaptic connec" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": "tions, which results in an estimated total of 100 to 500 trillion synapses. Despite its tremendous capacity, the brain exhibits a remarkable ability to grasp new concepts from minimal exposure. This one-shot or few-shot learning capability might be partly attributed to the effective use of sparsity [17]. By selectively activating a fraction of connections at any given time, the brain optimizes both energy consumption and computational efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 326, + 556, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 541 + ], + "type": "text", + "content": "Among the prominent DL models, vision-language models (VLMs) learn visual concepts from natural language descriptions. In particular, the contrastive language-image pretraining (CLIP) [33] model is pretrained on large-scale datasets containing hundreds of millions of image-text pairs to align visual and textual embeddings within a shared semantic space. After training, VLMs show strong zero-shot capabilities across diverse visual recognition tasks [33]. Moreover, fine-tuning VLMs yields state-of-the-art performance in various domain adaptation scenarios [27, 39]. Despite their zero-shot capabilities, adapting VLMs to new tasks with small labeled datasets remains challenging due to pronounced overfitting and catastrophic forgetting problems. Furthermore, training such models requires vast resources, including significant GPU memory for storing millions of parameters, gradients, activations, and optimizer states. These challenges underscore the need for cost-effective optimization strategies that address overfitting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "content": "To address the limitations of full VLM fine-tuning in few-shot learning, state-of-the-art methods leverage parameter-efficient training (PET) [15, 16, 21]. This strategy consists of optimizing a small number of parameters compared to the full set of weights. Additive PET methods freeze the initial weights and attach trainable and unmergeable parameters to the model. Within this category, we find the adapter [11] and the prompt tuning variants [49]. The prevailing methods in few-shot VLMs have mainly focused on this category. Adjustable weights can be incorporated at various levels, either by embedding them directly into the input space as visual [18] or textual prompts [48], or by integrating adapters throughout the network architecture [50]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "To mitigate overfitting in few-shot adaptation and pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3143" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 190, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 190, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 190, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 190, + 178 + ], + "type": "image", + "image_path": "998282f275fb5f80ced8c8975a2fe329bee4b2d3984092db3a153167a1ba6eef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 200, + 73, + 312, + 178 + ], + "blocks": [ + { + "bbox": [ + 200, + 73, + 312, + 178 + ], + "lines": [ + { + "bbox": [ + 200, + 73, + 312, + 178 + ], + "spans": [ + { + "bbox": [ + 200, + 73, + 312, + 178 + ], + "type": "image", + "image_path": "9c1e31495eeba8268f0f1be5f1d1e910acc283015936305e56824803b6132b55.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 321, + 73, + 432, + 178 + ], + "blocks": [ + { + "bbox": [ + 321, + 73, + 432, + 178 + ], + "lines": [ + { + "bbox": [ + 321, + 73, + 432, + 178 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 432, + 178 + ], + "type": "image", + "image_path": "dd7189c5fb0e3252bf1035cbaf77742f34f9dbda8a8c2b54be96d5270f38c8c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 441, + 73, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 441, + 73, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 441, + 73, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 441, + 73, + 553, + 178 + ], + "type": "image", + "image_path": "2c1248125ba560a9c08a7d18e181648eeedfd096700b2004eef961c29d8fbca5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 182, + 190, + 275 + ], + "blocks": [ + { + "bbox": [ + 61, + 182, + 190, + 275 + ], + "lines": [ + { + "bbox": [ + 61, + 182, + 190, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 190, + 275 + ], + "type": "image", + "image_path": "bea01f790e9e1b4ad400c75b69eb5d83d0a3f1dd4d6b7a96ecea4496cafca844.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 322 + ], + "type": "text", + "content": "Figure 1. Test-set accuracy over training iterations for LoRA (with rank fixed at 3) and SO (with a sparsity ratio equal to " + }, + { + "bbox": [ + 55, + 289, + 555, + 322 + ], + "type": "inline_equation", + "content": "99.95\\%" + }, + { + "bbox": [ + 55, + 289, + 555, + 322 + ], + "type": "text", + "content": ") in a 1-shot setting on four datasets (Pets, Flowers, Caltech, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. Training stops when the loss falls below 0.01 or after 2000 iterations." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 200, + 185, + 312, + 275 + ], + "blocks": [ + { + "bbox": [ + 200, + 185, + 312, + 275 + ], + "lines": [ + { + "bbox": [ + 200, + 185, + 312, + 275 + ], + "spans": [ + { + "bbox": [ + 200, + 185, + 312, + 275 + ], + "type": "image", + "image_path": "452e6c558c4b75c38d72e1227c9494fad1e750d7f5fc3002ebda03a024c0f17f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 185, + 432, + 275 + ], + "blocks": [ + { + "bbox": [ + 321, + 185, + 432, + 275 + ], + "lines": [ + { + "bbox": [ + 321, + 185, + 432, + 275 + ], + "spans": [ + { + "bbox": [ + 321, + 185, + 432, + 275 + ], + "type": "image", + "image_path": "53b4e9a77f4fb9bdaed33f0d375b7d53cc33afb4eec866f4dc11d629b952e4ce.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 442, + 182, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 442, + 182, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 442, + 182, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 442, + 182, + 553, + 275 + ], + "type": "image", + "image_path": "af8f89d3882d67bcc788316703ca6fbcf88659c34eb898ce2b39a34441366178.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 336, + 190, + 440 + ], + "blocks": [ + { + "bbox": [ + 61, + 336, + 190, + 440 + ], + "lines": [ + { + "bbox": [ + 61, + 336, + 190, + 440 + ], + "spans": [ + { + "bbox": [ + 61, + 336, + 190, + 440 + ], + "type": "image", + "image_path": "5218af9e9e4c7ca5c47bbbc554e75de01d4acddcebc631a7b0c8222cd4fd0693.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 200, + 337, + 312, + 440 + ], + "blocks": [ + { + "bbox": [ + 200, + 337, + 312, + 440 + ], + "lines": [ + { + "bbox": [ + 200, + 337, + 312, + 440 + ], + "spans": [ + { + "bbox": [ + 200, + 337, + 312, + 440 + ], + "type": "image", + "image_path": "2e559f2557e3b49eee358e923f88b9e8edb1b1105f44fdef72fc9d2db39947dd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 337, + 432, + 440 + ], + "blocks": [ + { + "bbox": [ + 321, + 337, + 432, + 440 + ], + "lines": [ + { + "bbox": [ + 321, + 337, + 432, + 440 + ], + "spans": [ + { + "bbox": [ + 321, + 337, + 432, + 440 + ], + "type": "image", + "image_path": "0c9f5f5abccb0862c1bbd83c7dc801e35441702214c27b372be0898ad6c7477e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 442, + 337, + 553, + 440 + ], + "blocks": [ + { + "bbox": [ + 442, + 337, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 442, + 337, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 442, + 337, + 553, + 440 + ], + "type": "image", + "image_path": "5b049a48fbd3758252415ed14fafab3997364ea8e5417b1c1f8be1da00f4a77f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 61, + 445, + 190, + 536 + ], + "blocks": [ + { + "bbox": [ + 61, + 445, + 190, + 536 + ], + "lines": [ + { + "bbox": [ + 61, + 445, + 190, + 536 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 190, + 536 + ], + "type": "image", + "image_path": "839bb7ad23c06b203f23082e233bc9504eff49688ff54685c8431df5df178874.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 200, + 445, + 312, + 536 + ], + "blocks": [ + { + "bbox": [ + 200, + 445, + 312, + 536 + ], + "lines": [ + { + "bbox": [ + 200, + 445, + 312, + 536 + ], + "spans": [ + { + "bbox": [ + 200, + 445, + 312, + 536 + ], + "type": "image", + "image_path": "7d0bb11896288fd91b0405849c9010a32bd9c8c63ff7e03dc8acfc516c302a54.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 321, + 445, + 432, + 536 + ], + "blocks": [ + { + "bbox": [ + 321, + 445, + 432, + 536 + ], + "lines": [ + { + "bbox": [ + 321, + 445, + 432, + 536 + ], + "spans": [ + { + "bbox": [ + 321, + 445, + 432, + 536 + ], + "type": "image", + "image_path": "22368532aeabaa75da0760d7dbe426cd4077df6b388fc62b89285f71013d898b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 442, + 443, + 553, + 536 + ], + "blocks": [ + { + "bbox": [ + 442, + 443, + 553, + 536 + ], + "lines": [ + { + "bbox": [ + 442, + 443, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 442, + 443, + 553, + 536 + ], + "type": "image", + "image_path": "653ac7451982321245105d471c82041e6c9935cff9bc915820edaf67285a88c1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 61, + 540, + 190, + 632 + ], + "blocks": [ + { + "bbox": [ + 61, + 540, + 190, + 632 + ], + "lines": [ + { + "bbox": [ + 61, + 540, + 190, + 632 + ], + "spans": [ + { + "bbox": [ + 61, + 540, + 190, + 632 + ], + "type": "image", + "image_path": "1c37137fec9706b698d853b63f8e8b92887fcc698994fd0b701f28dd1e44e84c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 647, + 555, + 681 + ], + "lines": [ + { + "bbox": [ + 55, + 647, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 555, + 681 + ], + "type": "text", + "content": "Figure 2. Test-set accuracy of LoRA in a 1-shot setting on three datasets (DTD, Oxford Pets, and UCF101), using a pretrained CLIP (ViT-B/16) backbone. We vary the LoRA rank from 2 to 5 and train for at most 2000 iterations or until the loss drops below 0.01. Each subfigure shows test accuracy over training iterations for a specific rank." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 200, + 540, + 312, + 632 + ], + "blocks": [ + { + "bbox": [ + 200, + 540, + 312, + 632 + ], + "lines": [ + { + "bbox": [ + 200, + 540, + 312, + 632 + ], + "spans": [ + { + "bbox": [ + 200, + 540, + 312, + 632 + ], + "type": "image", + "image_path": "b0eec707c9caa65024f5b46af69bd6150e07d76842f0b0555575c617c03e88ad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 321, + 540, + 432, + 632 + ], + "blocks": [ + { + "bbox": [ + 321, + 540, + 432, + 632 + ], + "lines": [ + { + "bbox": [ + 321, + 540, + 432, + 632 + ], + "spans": [ + { + "bbox": [ + 321, + 540, + 432, + 632 + ], + "type": "image", + "image_path": "65974d4476f4e198220419948cc9bb79554e39d06f6931ca6456c4f6b91932b5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 442, + 540, + 553, + 632 + ], + "blocks": [ + { + "bbox": [ + 442, + 540, + 553, + 632 + ], + "lines": [ + { + "bbox": [ + 442, + 540, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 442, + 540, + 553, + 632 + ], + "type": "image", + "image_path": "dd46951795c2b021fcd660832629a4ce2a989c4f96164f8b14d021a989701f10.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3144" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "content": "serve the prior knowledge of VLMs, adapter-based methods typically employ three techniques. First, these small trainable modules are only integrated into the last layer of VLMs. Second, they use regularization techniques to penalize the divergence between the initial zero-shot class prototypes and the learnable class prototypes. Finally, they initialize the learnable class prototypes into a reliable region using the zero-shot prototypes. As a result, these methods have shown better results than the prompt tuning approaches [44], which often introduce computational overhead [45]. Despite these efforts, adapter-based methods are still prone to considerable overfitting [34]. Their success depends heavily on carefully tuning hyperparameters whose optimal values vary significantly across different tasks [34]. This sensitivity leads to performance degradation that undermines the generalizability of these approaches. It is worth noting that tuning these hyperparameters requires extensive grid searching with a validation set, which is not available in real-world few-shot scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": "Recently, low-rank reparametrization [16, 20, 26] has gained popularity for domain adaptation. These approaches reparameterize the initial weights with supplementary trainable and mergeable low-rank matrices. Low-rank adaptation (LoRA) [16] was the first to apply this strategy. It reformulates weight matrices " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{W} + \\mathbf{A} \\cdot \\mathbf{B}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": " are low-rank matrices. As a key advantage, this technique helps reduce overfitting and catastrophic forgetting because of two properties. First, LoRA enables fine-grained parallel adaptation by learning small perturbations to the frozen weights [13]. Second, the effective initialization of " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 301, + 295, + 445 + ], + "type": "text", + "content": " helps in preserving prior knowledge [30, 43]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 448, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 651 + ], + "type": "text", + "content": "Building on these advantages, recent studies [25, 45] show that low-rank reparameterization methods achieve state-of-the-art performance in few-shot adaptation of VLMs. However, these methods also have notable limitations. Compelling evidence suggests that restricting weight updates to low-rank subspaces reduces model expressiveness [24, 37, 47]. In this work, we argue that the low-rank projection technique exhibits severe overfitting in few-shot adaptation. Our systematic study of LoRA confirms this issue (see Fig. 1). Test-set accuracy initially improves but then declines sharply. At convergence, when the loss approaches zero, performance deteriorates significantly compared to its peak on test-set data. Moreover, we can see that the optimal number of training iterations varies widely across datasets. However, tuning hyperparameters, including the number of iterations for early stopping, on a validation set contradicts the principles of few-shot learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "The challenges of low-rank reparameterization methods extend beyond fixing the number of training iterations. Their performance is also highly sensitive to another hyperparameter: the rank of the low-rank projection. This sensitivity is shown in Fig. 2, which illustrates LoRA's perfor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "mance for different ranks on three datasets (DTD, Oxford Pets, UCF101) in a 1-shot setting. Each dataset reveals a distinct best rank, and higher ranks sometimes lead to an early spike in accuracy but then degrade as training continues. Further experiments examining these behaviors are presented in Appendix F of the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 147, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 337 + ], + "type": "text", + "content": "Since rank is a discrete value that cannot be lower than 1, the low-rank methods lack the flexibility to adjust very few parameters. While setting the rank to a very low value (e.g., 1 or 2) can help mitigate overfitting, this choice might reduce the model's learning capacity. This limitation arises because the trainable parameters in low-rank methods remain fixed throughout training, potentially leading to underfitting the VLM capacity. Indeed, rank selection involves a trade-off between underfitting and overfitting (see Fig. 2). For instance, in the 1-shot DTD setting, LoRA at rank 5 briefly exceeds rank 2 but then drops below it as training continues. Overall, the test accuracy of LoRA exhibits oscillations and varies unpredictably w.r.t. rank and the number of training iterations. This instability makes it difficult to reliably tune these interdependent hyperparameters without access to a validation set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 340, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 555 + ], + "type": "text", + "content": "In this paper, we deviate from the current long-standing trend of low-rank adaptation and introduce two new paradigms that make sparsity work in few-shot settings. Our approach is a simple yet effective optimization strategy, referred to as sparse optimization (SO), and aligns with these paradigms. As a first paradigm, we advocate for local sparsity and global density. Unlike the conventional low-rank approaches that update a fixed set of parameters, SO dynamically updates a minimal subset of original connections at each step. Specifically, high sparsity is enforced in both the gradient and moment updates while allowing the sparsity support to evolve dynamically throughout training. This paradigm ensures that, at any given iteration, only a minimal subset of parameters is updated to reduce the local learning capacity and prevent rapid overfitting. By dynamically changing the sparsity support, the model retains its overall expressiveness and ensures that different parameters contribute to learning over time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "As a second paradigm, we advocate for local randomness and global importance. Similar to Adam [19], we leverage the gradient and the first and second moments. However, we sparsify these components in a way that addresses overfitting and ensures efficiency in terms of memory consumption. In particular, we sparsify the gradient randomly and the first moment by importance. The gradient captures local and iteration-specific information. The first moment aggregates the gradients over the whole path and, thus, reflects long-term parameter importance. Random pruning of the gradient prevents the model from relying too much on short-term and local high-magnitude updates. The importance-based selection of the first moment" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3145" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "ensures that connections with long-term significance are updated. By emphasizing local randomness and global importance, our strategy avoids severe overfitting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": "Concretely instantiating this paradigm, we fix a density rate " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " and retain " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " elements for both the gradient and moment updates at each iteration. First, we randomly select " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " gradient values from the full set of parameter updates. Next, we compute the temporary first and second moments, which can contain up to " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " values in the worst-case scenario where no overlap exists between the selected gradients and the previous " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " first moment values. We rank the " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " first moment values by importance using their magnitude. Only the top " + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 109, + 295, + 287 + ], + "type": "text", + "content": " values are retained for the first moment. We then use the same indices to select the corresponding values for the second moment. This ensures alignment between the sparsified first and second moments. The selection strategy preserves the most relevant long-term updates while introducing randomness in the local updates." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 515 + ], + "type": "text", + "content": "Our contributions. (1) A novel framework is proposed that sparsifies both gradients and moments to improve generalization and mitigate overfitting in few-shot learning. Our approach differs from conventional low-rank reparameterization by focusing on sparsity. (2) We introduce the local sparsity and global density paradigm, where a dynamically evolving sparsity support updates only a minimal subset of model parameters at each iteration. This strategy preserves the overall model expressiveness while reducing its risk of overfitting. (3) We propose the local randomness and global importance paradigm, which selectively sparsifies the gradient using random selection while pruning the first moment based on importance. This selection strategy preserves the most relevant long-term updates while introducing local randomness to mitigate overfitting. (4) Our comprehensive empirical study on 11 diverse datasets compares our method to state-of-the-art low-rank approaches. Results indicate that SO can significantly improve few-shot generalization while achieving memory efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 142, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 142, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 142, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "(a) Low-Rank Methods. LoRA [16] approximates weight updates by decomposing them into two low-rank matrices. As a result, it reduces the number of trainable parameters while preserving the model's frozen pretrained backbone. DoRA [26] introduces a weight decomposition approach to improve LoRA's expressiveness by decoupling weight magnitude and direction and training them separately. VeRA [21] reduces the set of trainable parameters further than LoRA. In particular, VeRA freezes the pair of low-rank matrices to randomly initialized ones and only trains two small scaling vectors. PiSSA [30] initializes the two low-rank matrices using the most informative singular vectors and their corresponding singular values from the original weight matrix while freezing the less significant compo" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "nents in a residual matrix. ReLoRA [24] achieves high-rank weight updates by merging trained low-rank updates into the main model parameters at regular intervals and reinitializing the auxiliary low-rank matrices. Unlike these LoRA-based reparametrization methods, which constrain weight updates to low-rank subspaces, GaLore [47] performs full-parameter training. Specifically, GaLore projects gradients into a low-rank subspace to reduce memory consumption while allowing full updates to the model weights." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "type": "text", + "content": "Despite their efficiency, these methods still face major limitations. Most low-rank techniques impose a fixed-rank constraint on the updates, which restricts adaptation flexibility and may lead to underfitting when the rank is too low. Even with minimal rank, these methods struggle with generalization in few-shot adaptation due to severe overfitting. Furthermore, these approaches require careful hyperparameter tuning, particularly in selecting the optimal rank and training iterations, which significantly impact performance. Our work challenges this strategy by showing that appropriately designed sparsity-based adaptation provides superior generalization while reducing memory consumption." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 454 + ], + "type": "text", + "content": "(b) Sparsity-Based Methods. These methods update only a subset of model parameters. The selection can be structured or unstructured. Structured selection methods [4, 9, 23, 28, 40] target entire modules (e.g., attention heads, layers), offering coarse-grained sparsity but often requiring architecture-specific designs that may degrade performance [5]. Unstructured selection methods [2, 12, 36, 38, 42, 46] prune individual weights for greater flexibility. However, most unstructured methods rely on static sparsity patterns and perform importance-based gradient pruning. Moreover, none of these methods incorporates sparse moments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 526 + ], + "type": "text", + "content": "In contrast, SO updates a minimal subset of parameters at each step. Furthermore, our approach uses random gradient selection to avoid local high-magnitude biases and importance-based moment pruning to retain long-term relevance. This combination enables stable adaptation in few-shot scenarios without sacrificing model expressiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 538, + 528, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 528, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 528, + 551 + ], + "type": "text", + "content": "3. Proposed Sparse Optimization Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "text", + "content": "A comprehensive overview of few-shot adaptation in VLMs is provided in Appendix A. Briefly, we leverage a pretrained CLIP [33] model that aligns images and texts in a shared embedding space via contrastive learning. After pretraining, CLIP can perform zero-shot classification by comparing image embeddings to class-specific text prototypes. In the few-shot scenario, we only have a small support set of labeled examples (e.g., " + }, + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 558, + 553, + 689 + ], + "type": "text", + "content": " images per class). We adapt CLIP to this support set by minimizing the cross-entropy loss with respect to the class prototypes or learnable parameters, and we optimize the loss function using our proposed approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Motivated by the challenges inherent in few-shot adaptation, we introduce SO, a new optimizer designed to en" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3146" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "hance memory efficiency while mitigating overfitting. Unlike standard optimization approaches that update all parameters at each iteration, SO dynamically updates a subset of parameters. This sparsification significantly reduces memory consumption, which makes SO suitable for training large-scale deep learning models like VLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 277 + ], + "type": "text", + "content": "Our approach builds on the Adam optimizer and behaves identically to it when the sparsity ratio is set to zero. However, SO deviates from Adam by following two paradigms. The first paradigm is local sparsity and global density, which ensures that a minimal subset of parameters is dynamically updated while maintaining overall model expressiveness. The second paradigm is local randomness and global importance, which applies random pruning to the gradient updates while retaining the most significant moment estimates. These principles prevent severe overfitting and make SO suitable for scenarios like few-shot learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": "Notation. Let " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " represent the model parameters, where " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " is the total number of parameters. At each iteration " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": ", the model updates some parameters to minimize a given loss function " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\Theta_t)" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": ". The gradient of the loss concerning all parameters is denoted as " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "g_{t} = \\nabla_{\\Theta}\\mathcal{L}(\\Theta_{t})" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": ". We use the first and second moments, denoted as " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\nu_{t}" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " respectively, to capture long-term and global parameter significance throughout the optimization process. A small constant " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " is introduced for numerical stability. The learning rate, denoted as " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": ", scales the parameter updates. We introduce a density ratio " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\kappa \\in [0,1]" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": ", which determines the fraction of gradient entries retained at each step. Let " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(\\cdot)" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot)" + }, + { + "bbox": [ + 55, + 278, + 295, + 470 + ], + "type": "text", + "content": " be the functions that return the indices and values of a sparse input vector, respectively. Instead of updating all parameters, SO retains only a small subset of the gradient and generates sparse first and second moments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "content": "We enforce sparsity through two operations. The Top-" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "content": " operation selects " + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "content": " elements with the highest magnitude. In contrast, Random-" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "content": " retains " + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 472, + 295, + 519 + ], + "type": "text", + "content": " elements uniformly at random, preventing bias toward high-magnitude values." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 663 + ], + "type": "text", + "content": "Dynamic and High-Sparsity Support. Conventional optimization methods update the full set of parameters throughout training, which accelerates overfitting. In contrast, our method dynamically updates a minimal subset of parameters. The high sparsity and dynamic selection provide two key advantages. First, updating a minimal subset of parameters reduces the local learning capacity and thus mitigates overfitting and catastrophic forgetting. Second, continuously changing the sparsity support ensures that different parameters receive updates throughout training. This avoids excessive reliance on a fixed set and makes the learning process benefit from the model's overall expressiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Randomness in Gradient Pruning. We sparsify the gradient vector to mitigate overfitting. The proposed optimizer applies a stochastic sampling mechanism that uniformly picks " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": " elements from the gradient " + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": ". Given the density" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "ratio " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", we define the number of retained gradient entries as " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "M = \\lfloor \\kappa d \\rfloor" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ". Formally, we have " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t = \\text{Random-} M(g_t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "type": "text", + "content": "The motivation is to avoid biasing updates toward high-magnitude gradients. Enforcing stochastic sparsity prevents the model from over-relying on seemingly important local information. Our empirical results indicate that pruning the gradient by retaining only the largest values accelerates overfitting and significantly degrades performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": "Importance in Moment Pruning. To maintain efficiency, our goal is to construct sparse first and second moments, " + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t" + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu}_t" + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": ", each of size " + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": ". Since the sparse gradient changes its support dynamically, storing historical information becomes challenging. Without pruning, the first and second moments can become dense over time due to the accumulation of gradients with varying sparsity support." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": "To address this issue, we introduce two fixed-size buffers that can store up to " + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": " values. This ensures that both new and past gradient values are considered. Let " + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mu_t" + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\nu_t" + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": " be the temporary first and second moments, at iteration " + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 251, + 553, + 300 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 309, + 553, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 309, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 375, + 309, + 553, + 323 + ], + "type": "interline_equation", + "content": "\\mu_ {t} = \\beta_ {1} \\tilde {\\mu} _ {t - 1} + (1 - \\beta_ {1}) \\tilde {g} _ {t}, \\tag {1}", + "image_path": "66f40be36a63bb900a56bdd2c8085f3a3eb3eaa5593aeb64ee4edc57aada8241.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 324, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 324, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 376, + 324, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\nu_ {t} = \\beta_ {2} \\tilde {\\nu} _ {t - 1} + (1 - \\beta_ {2}) \\tilde {g} _ {t} ^ {2}. \\tag {2}", + "image_path": "a94c4835f071b17c3d1f1a4926f271c29ba9cc5ffc5292590ef4db5b5ad38e12.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": "These moments may contain up to " + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": " values in the worst-case scenario when there is no overlap between the newly selected sparse gradient " + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": " and the previously retained sparse first moment " + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": ". To ensure stable updates, we employ a Top-" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": " selection strategy for the first moment, where we rank the " + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "2M" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": " values by magnitude and retain only the top " + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 348, + 555, + 479 + ], + "type": "text", + "content": " ones. This guarantees that long-term significant parameters are consistently updated. For the second moment, we enforce alignment by retaining elements corresponding to the same indices selected for the first moment. The sparse first and second moments are expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 489, + 553, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 489, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 358, + 489, + 553, + 503 + ], + "type": "interline_equation", + "content": "\\tilde {\\mu} _ {t} = \\operatorname {T o p} - M (\\mu_ {t}), \\quad \\tilde {\\nu} _ {t} = \\nu_ {t} [ \\mathcal {I} (\\tilde {\\mu} _ {t}) ]. \\tag {3}", + "image_path": "d5d1536763317579164d45b285d4a834b1f504ad82d1d03ccfb05ca5319f4d5b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": "Both sparse moments " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu}" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": " are initialized at zero. Similar to Adam, we apply bias correction to " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mu}_t" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu}_t" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": " before parameter update. The corrected sparse first and second moments, denoted by " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}_t" + }, + { + "bbox": [ + 313, + 512, + 553, + 559 + ], + "type": "text", + "content": ", are expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 568, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 568, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 370, + 568, + 553, + 594 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} _ {t} = \\frac {\\tilde {\\mu} _ {t}}{1 - \\beta_ {1} ^ {t}}, \\quad \\hat {\\nu} _ {t} = \\frac {\\tilde {\\nu} _ {t}}{1 - \\beta_ {2} ^ {t}}. \\tag {4}", + "image_path": "47bdcc9dc81150f3d83affbe386bc1f63f8f8b1011df614ff3bbcb7a711a748f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "text", + "content": "Finally, the parameters " + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "text", + "content": " are updated using the corrected sparse first and second moments, " + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t" + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\nu}_t" + }, + { + "bbox": [ + 313, + 601, + 553, + 626 + ], + "type": "text", + "content": ", as provided in:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 379, + 634, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 634, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 379, + 634, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\Theta_ {t + 1} = \\Theta_ {t} - \\frac {\\eta}{\\sqrt {\\hat {\\nu} _ {t}} + \\epsilon} \\hat {\\mu} _ {t}, \\tag {5}", + "image_path": "a7cd53ac42b0ad5b1e9bc1f125900d6eaf528b1c761d75cd73994e2401616533.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " is the learning rate, and " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": " are decay rates. Algorithm. Our algorithm is provided in Appendix B due to space limitations. We finetune CLIP until convergence using our optimization technique. To promote efficiency," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3147" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 102, + 547, + 382 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Few-shot classification performance on 11 datasets with ViT-B/16 backbone. Top-1 accuracy averaged over 3 random seeds is reported. Highest value is highlighted in bold, and the second highest is underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 102, + 547, + 382 + ], + "lines": [ + { + "bbox": [ + 66, + 102, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 547, + 382 + ], + "type": "table", + "html": "
ShotsMethodImageNetSUNAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
0CLIP (ICML '21)66.762.624.747.565.386.189.171.492.943.666.765.1
1Adam (ICLR '15)0.10.61.024.30.51.02.91.41.05.61.73.6
LoRA (ICLR '22)67.367.025.067.568.281.290.585.792.352.472.970.0
ReLoRA (ICLR '24)70.369.728.873.870.584.291.985.193.454.276.072.5
GaLoRE (ICML '24)65.165.521.165.263.674.185.280.591.448.969.966.4
PiSSA (NeurIPS '24)64.965.222.266.864.275.784.882.990.750.570.167.1
DoRA (ICML '24)66.966.925.371.468.381.389.386.292.351.972.970.2
VeRA (ICLR '24)68.467.327.569.367.486.291.073.193.952.270.169.7
SO (Ours)70.170.331.578.271.686.293.384.994.155.376.473.8
2Adam (ICLR '15)0.20.80.942.10.71.93.917.13.17.23.87.4
LoRA (ICLR '22)67.468.330.381.970.179.389.490.993.759.676.173.4
ReLoRA (ICLR '24)70.670.532.283.872.982.690.390.794.760.479.175.3
GaLoRE (ICML '24)66.167.025.882.267.272.681.585.691.556.173.469.9
PiSSA (NeurIPS '24)65.166.726.177.367.174.884.487.491.957.473.370.1
DoRA (ICML '24)67.268.229.883.470.379.489.691.293.659.775.973.5
VeRA (ICLR '24)69.469.031.781.669.886.792.679.694.459.374.773.5
SO (Ours)70.572.337.282.774.485.392.291.995.360.280.476.6
4Adam (ICLR '15)0.12.12.447.90.73.16.330.512.59.69.211.3
LoRA (ICLR '22)68.569.735.285.274.878.787.994.193.963.578.275.4
ReLoRA (ICLR '24)71.272.237.684.777.182.390.394.194.964.080.077.1
GaLoRE (ICML '24)67.569.034.182.173.175.281.091.193.260.075.472.9
PiSSA (NeurIPS '24)66.468.832.482.873.275.680.692.092.961.175.772.9
DoRA (ICML '24)68.469.734.985.575.179.089.594.194.064.178.575.7
VeRA (ICLR '24)70.571.535.685.673.786.793.189.895.364.579.576.9
SO (Ours)71.473.738.687.778.985.392.495.195.566.483.478.9
", + "image_path": "7cfaa48a5aedc1ef162d9dbfc54d8c2fb8e7f5c8bbaaff316b0de98a55a8ee63.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 412, + 547, + 534 + ], + "blocks": [ + { + "bbox": [ + 138, + 392, + 472, + 405 + ], + "lines": [ + { + "bbox": [ + 138, + 392, + 472, + 405 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 472, + 405 + ], + "type": "text", + "content": "Table 2. Impact of gradient and moment sparsification in few-shot classification (ViT-B/16)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 412, + 547, + 534 + ], + "lines": [ + { + "bbox": [ + 66, + 412, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 66, + 412, + 547, + 534 + ], + "type": "table", + "html": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
DenseSparseDenseSparse
1SO (≡ Adam)✓✗✓✗1.024.30.51.02.91.41.05.61.74.4
SO✗✓✓✗29.473.969.683.993.286.493.755.777.073.6
SO✗✓✗✓31.578.271.686.293.384.994.155.376.474.6
2SO (≡ Adam)✓✗✓✗0.942.10.71.93.917.13.17.23.89.0
SO✗✓✓✗29.482.671.081.591.490.193.759.778.475.3
SO✗✓✗✓37.282.774.485.392.291.995.360.280.477.7
4SO (≡ Adam)✓✗✓✗2.447.90.73.16.330.512.59.69.213.6
SO✗✓✓✗34.585.473.578.991.192.794.266.278.377.2
SO✗✓✗✓38.687.778.985.392.495.195.566.483.480.4
", + "image_path": "a4d986dd91f21006405ef92679c972834ea6bce69737572fc2386d885ca38b58.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": "we retain the gradient sparsity support fixed for " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": " iterations before reselecting the trainable parameters. Overall, our approach introduces two hyperparameters: the density ratio " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": ", which controls the fraction of retained gradient entries, and the update interval " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": ", which determines how frequently the sparsity support is refreshed. A smaller " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": " value improves generalization but results in slower convergence. In contrast, a larger " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": " accelerates optimization at the cost of reduced performance. Similarly, reducing " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": " leads to better results by enabling more frequent updates to the sparsity support, although increasing " + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 554, + 294, + 685 + ], + "type": "text", + "content": " speeds up convergence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "The role of " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": " in our method is conceptually similar to update intervals used in prior work such as GaLore [47]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "content": "and ReLoRA [24]. In ReLoRA, the low-rank matrices are merged with the original parameters every " + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "content": " interval. GaLore also updates the low-rank projectors every " + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "content": " interval. Unlike the low-rank methods, where hyperparameters such as rank and number of training iterations involve a strong trade-off between overfitting and underfitting, our sparsity-based approach remains robust to overfitting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 655, + 446, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 446, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 446, + 668 + ], + "type": "text", + "content": "4. Results and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "(a) Experimental Methodology. Our optimizer is evaluated on 11 datasets commonly used in few-shot VLM adaptation [45, 49], including ImageNet [8], SUN397 [41], Air" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3148" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 91, + 548, + 253 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 528, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 528, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 528, + 82 + ], + "type": "text", + "content": "Table 3. Impact of randomness and importance in gradient and moment sparsification in few-shot classification (ViT-B/16)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 91, + 548, + 253 + ], + "lines": [ + { + "bbox": [ + 66, + 91, + 548, + 253 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 548, + 253 + ], + "type": "table", + "html": "
ShotsMethodGradientMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
RandImptRandImpt
1SOX✓X✓17.673.960.572.087.975.989.953.968.266.6
SOX✓✓X20.771.463.774.692.079.991.954.671.169.6
SO✓X✓X31.271.769.885.693.380.394.155.074.972.9
SO✓XX✓31.578.271.686.293.384.994.155.376.474.6
2SOX✓X✓21.480.862.667.782.884.688.757.668.268.3
SOX✓✓X22.882.465.373.487.184.291.360.374.571.3
SO✓X✓X33.282.172.885.792.191.294.860.979.676.9
SO✓XX✓37.282.774.485.392.291.995.360.280.477.7
4SOX✓X✓28.283.267.966.969.886.489.055.969.962.5
SOX✓✓X31.081.071.573.077.786.991.458.974.171.6
SO✓X✓X39.584.777.686.292.793.395.866.082.579.8
SO✓XX✓38.687.778.985.392.495.195.566.483.480.4
", + "image_path": "409cc2d6695656b13026e69cad117f65bbfa07ff9975a917c3129da4ab742ba0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 281, + 548, + 392 + ], + "blocks": [ + { + "bbox": [ + 118, + 262, + 490, + 273 + ], + "lines": [ + { + "bbox": [ + 118, + 262, + 490, + 273 + ], + "spans": [ + { + "bbox": [ + 118, + 262, + 490, + 273 + ], + "type": "text", + "content": "Table 4. Impact of dynamic sparsity support for gradient pruning in few-shot classification (ViT-B/16)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 281, + 548, + 392 + ], + "lines": [ + { + "bbox": [ + 66, + 281, + 548, + 392 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 548, + 392 + ], + "type": "table", + "html": "
ShotsMethodSparsity supportAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
StaticDynamic
1SO✓✘28.272.468.185.993.173.494.552.672.671.2
SO✘✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✘31.083.069.785.891.683.094.355.775.874.4
SO✘✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✘34.883.672.486.292.588.095.162.577.877.0
SO✘✓38.687.778.985.392.495.195.566.483.480.4
", + "image_path": "8f6af10b2c31b62df36f187b1583418bdf0bcdd3c58d7e4dc708d8bf6a233d72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 411, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 294, + 460 + ], + "type": "text", + "content": "craft [29], EuroSAT [14], Stanford-Cars [22], Food101 [3], OxfordPets [32], Flowers102 [31], Caltech101 [10], DTD [7], and UCF101 [35]. These datasets provide a comprehensive assessment across a variety of image domains." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 557 + ], + "type": "text", + "content": "We evaluate SO against state-of-the-art low-rank methods, which have shown strong performance in finetuning Large Language Models and Vision Language models. Specifically, we compare our method against LoRA, ReLoRA, VeRA, DoRA, PiSSA, and GaLoRE. Our comparison also includes full fine-tuning using Adam to showcase the effect of severe overfitting. A discussion of these baselines, including their limitations, is provided in Sec. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "For a fair comparison, we follow the same hyperparameter settings as CLIP-LoRA [45] for the low-rank baselines. In particular, we set the rank to 2, and we train for 500 iterations. Due to severe overfitting, we introduce an early stopping criterion for the low-rank methods. If the training loss reaches 0.01, training is halted. This is particularly necessary to obtain competitive results using the low-rank approaches for some datasets where the test accuracy peaks often within a few iterations before rapidly overfitting. For our method, we set the density ratio to " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\kappa = 0.05\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": ", and we update the gradient sparsity support every " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " iterations. We train until convergence, defined as reaching a loss threshold of 0.01. If convergence is not achieved, we stop" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 411, + 504, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 504, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 504, + 423 + ], + "type": "text", + "content": "training at a maximum limit of 2000 iterations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 425, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 544 + ], + "type": "text", + "content": "Following CLIP-LoRA [45], we adopt a learning rate of " + }, + { + "bbox": [ + 313, + 425, + 554, + 544 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 425, + 554, + 544 + ], + "type": "text", + "content": " with a cosine learning rate scheduler and a batch size of 32. We use the simple prompt template \"a photo of a [class name]\" without complex manual prompt engineering. All experiments are conducted using a ViT-B/16 backbone, and finetuning is applied to all layers of the 12 vision and text Transformer blocks. These settings were suggested by CLIP-LoRA and are kept the same for all methods, including ours. All hyperparameters remain fixed across all datasets for our approach and the baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "content": "All experiments are conducted under identical hardware and software conditions. The details of these environments are provided in Appendix E. Additional experimental results showing CLIP's few-shot adaptation performance are presented in Appendix F. A comparative analysis of the memory efficiency of SO and state-of-the-art low-rank methods is presented in Appendix C. The sensitivity of our approach to the hyperparameters " + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "content": " is explored in Appendix D. Moreover, to further assess the effectiveness of our optimizer, we conduct both standard classification and few-shot learning experiments using a simple two-layer fully connected network. The results obtained with this small architecture are detailed in Appendix G." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "(b) Sparsity Outperforms Low-Rank Projections. The" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3149" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 91, + 548, + 205 + ], + "blocks": [ + { + "bbox": [ + 185, + 71, + 424, + 82 + ], + "lines": [ + { + "bbox": [ + 185, + 71, + 424, + 82 + ], + "spans": [ + { + "bbox": [ + 185, + 71, + 424, + 82 + ], + "type": "text", + "content": "Table 5. Impact of moments in few-shot classification (ViT-B/16)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 91, + 548, + 205 + ], + "lines": [ + { + "bbox": [ + 66, + 91, + 548, + 205 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 548, + 205 + ], + "type": "table", + "html": "
ShotsMethodMomentsAircraftEuroSATCarsFoodPetsFlowersCaltechDTDUCFAverage
withoutwith
1SO✓✗20.974.751.976.585.782.992.554.570.267.8
SO✗✓31.578.271.686.293.384.994.155.376.474.6
2SO✓✗24.984.660.277.589.590.992.961.176.073.0
SO✗✓37.282.774.485.392.291.995.360.280.477.7
4SO✓✗32.289.172.179.089.193.593.864.477.376.7
SO✗✓38.687.778.985.392.495.195.566.483.480.4
", + "image_path": "4617efd5f3608b24a9f4991687c7201ed405d79b26b7c98d4b84d9dc9ebdafd8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 294, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 437 + ], + "type": "text", + "content": "results in Table 1 show the effectiveness of SO compared to state-of-the-art low-rank techniques. Across all shots (1, 2, and 4), our method achieves the highest average performance. Furthermore, SO yields the highest accuracy in most datasets, including EuroSAT, Cars, Flowers, and UCF101. Notably, it outperforms LoRA and ReLoRA on Aircraft by significant margins. Low-rank reparametrization methods constrain updates to a fixed subspace. The only exception is ReLoRA, which periodically merges and reinitializes the auxiliary low-rank matrices; this might explain why ReLoRA outperforms the other low-rank reparametrization methods. Moreover, these methods suffer from severe overfitting and require careful tuning of the rank and number of training iterations to achieve their highest performance. Unlike these methods, SO employs dynamic sparsity to update only a minimal subset of parameters, thereby mitigating overfitting while preserving the model's full representational capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 438, + 294, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 544 + ], + "type": "text", + "content": "(d) Effectiveness of Pruning in Few-Shot Optimization. Table 2 shows the impact of gradient and moment pruning on few-shot classification. Dense optimization (Adam) performs poorly in the few-shot setting due to severe overfitting. In contrast, gradient pruning leads to substantial performance improvement. Moment pruning further enhances performance by filtering out updates related to parameters that are only locally important. These findings validate that sparsity can improve generalization in few-shot learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 545, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 713 + ], + "type": "text", + "content": "(e) Local Randomness Global Importance. Table 3 presents the performance of four sparse optimization variants based on randomness and importance in both gradient and moment pruning. Importance-based pruning is achieved by retaining the values with the highest magnitude. The results indicate that applying random selection for gradient pruning and importance-based selection for moment pruning achieves the highest average performance across all few-shot settings. Results in Appendix F confirm that importance-based gradient selection leads to rapid overfitting. Furthermore, Table 3 shows that importance-based moment pruning outperforms random moment pruning. This is because significant parameters over the long term are consistently updated, while locally influential ones" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 223, + 553, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 500 + ], + "type": "text", + "content": "are filtered out from the moment updates. Overall, our results validate the local randomness and global importance paradigm. Stochastic gradient selection prevents bias toward high-magnitude local updates and importance-based pruning preserves influential parameters over the long term. (f) Local Sparsity Global Density Table 4 shows that dynamic sparsity support consistently outperforms fixed sparsity support in gradient pruning across all few-shot settings. This improvement validates the local sparsity and global density paradigm. On the one hand, dynamic sparsity ensures that few parameters contribute to the updates to reduce the local learning capacity and thus mitigate overfitting. On the other hand, dynamic sparsity ensures a broader exploration of the parameter space while making the learning process benefit from the model's overall expressiveness. (g) Role of Moments in Few-Shot Adaptation. Table 5 shows that moment-based updates significantly improve performance. The advantage is more pronounced in datasets with higher complexity, such as Aircraft and Cars. Without moments, updates rely solely on randomly selected gradient values. These results provide evidence that gradient pruning is not sufficient. Random gradient pruning is more effective when combined with moments, particularly sparse ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "In this paper, we shed light on sparsity as an effective mechanism to mitigate overfitting in few-shot adaptation of VLMs. Conventional low-rank methods suffer from severe overfitting and require extensive hyperparameter tuning. To address these limitations, we introduce two key paradigms. The first paradigm consists of dynamically selecting a minimal subset of parameters to reduce overfitting while preserving the overall model expressiveness. The second paradigm aims to prevent bias in gradient updates while retaining critical long-term information. Our experiments confirm that sparsity outperforms low-rank projections in few-shot learning and constitutes a strong mechanism against overfitting. As future work, we plan to evaluate our optimizer on large language models. Memory efficiency in this setting is even more critical." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3150" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Digital Research Alliance of Canada." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Randall Balestriero, Leon Bottou, and Yann LeCun. The effects of regularization and data augmentation are class dependent. In Advances in Neural Information Processing Systems (NeurIPS), pages 37878-37891, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 255 + ], + "type": "text", + "content": "[2] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 1-9, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 300 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In European Conference on Computer Vision (ECCV), pages 446–461, 2014. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 345 + ], + "type": "text", + "content": "[4] Tianyi Chen, Tianyu Ding, Badal Yadav, Ilya Zharkov, and Luming Liang. Lorashear: Efficient large language model structured pruning and knowledge recovery. arXiv preprint arXiv:2310.18356, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "text", + "content": "[5] Hongrong Cheng, Miao Zhang, and Javen Qinfeng Shi. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), pages 10558-10578, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 468 + ], + "type": "text", + "content": "[6] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 513 + ], + "type": "text", + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3606-3613, 2014. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248-255, 2009. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "type": "text", + "content": "[9] Lucio Dery, Steven Kolawole, Jean-François Kagy, Virginia Smith, Graham Neubig, and Ameet Talwalkar. Everybody prune now: Structured pruning of llms with only forward passes. arXiv preprint arXiv:2402.05406, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 295, + 658 + ], + "type": "text", + "content": "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 178–178, 2004. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[12] Andi Han, Jiaxiang Li, Wei Huang, Mingyi Hong, Akiko Takeda, Pratik Jawanpuria, and Bamdev Mishra. SLTrain: a sparse plus low-rank approach for parameter and memory efficient pretraining. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[13] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 230, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 285 + ], + "type": "text", + "content": "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning (ICML), pages 2790-2799, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 286, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 286, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 286, + 553, + 340 + ], + "type": "text", + "content": "[16] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. 1, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 385 + ], + "type": "text", + "content": "[17] Iiro P Jaëskeläinen, Enrico Glerean, Vasily Klucharev, Anna Shestakova, and Jyrki Ahveninen. Do sparse brain activity patterns underlie human cognition? NeuroImage, 263: 119633, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "text", + "content": "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), pages 709-727, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "text", + "content": "[20] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 555 + ], + "type": "text", + "content": "[21] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki M Asano. VeRA: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 600 + ], + "type": "text", + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops (WICCV), pages 554–561, 2013. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "text", + "content": "[23] François Lagunas, Ella Charlaix, Victor Sanh, and Alexander Rush. Block pruning for faster transformers. In Empirical Methods in Natural Language Processing (EMNLP), pages 10619-10629, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Vladislav Lialin, Sherin Muckatira, Namrata Shivagunde, and Anna Rumshisky. Relora: High-rank training through low-rank updates. In International Conference on Learning Representations (ICLR), 2023. 3, 4, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Cheng Lin, Lujun Li, Dezhi Li, Jie Zou, Wei Xue, and Yike Guo. Nora: Nested low-rank adaptation for efficient" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3151" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "fine-tuning large models. arXiv preprint arXiv:2408.10280, 2024.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), pages 32100-32121, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[27] Huaishao Luo, Junwei Bao, Youzheng Wu, Xiaodong He, and Tianrui Li. Segclip: Patch aggregation with learnable centers for open-vocabulary semantic segmentation. In International Conference on Machine Learning (ICML), pages 23033-23044, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[28] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Llm-pruner: On the structural pruning of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. In *British Machine Vision Conference (BMVC)*, 2013. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[30] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. PiSSA: Principal singular values and singular vectors adaptation of large language models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[31] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics & Image Processing (ICVGIP), pages 722-729, 2008. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[32] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3498-3505, 2012. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 1, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 555 + ], + "type": "text", + "content": "[34] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "text", + "content": "[35] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[36] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In International Conference on Learning Representations (ICLR), 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[37] Xianghui Sun, Yunjie Ji, Baochang Ma, and Xiangang Li. A comparative study between full-parameter and lora-based fine-tuning on chinese instruction data for instruction following large language model. arXiv preprint arXiv:2304.08109, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Yi-Lin Sung, Varun Nair, and Colin A Raffel. Training neural networks with fixed sparse masks. In Advances in Neural" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Information Processing Systems (NeurIPS), pages 24193-24205, 2021. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[39] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[40] Mengzhou Xia, Zexuan Zhong, and Danqi Chen. Structured pruning learns compact and accurate models. In Annual Meeting of the Association for Computational Linguistics (ACL), 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[41] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3485-3492, 2010. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[42] Runxin Xu, Fuli Luo, Zhiyuan Zhang, Chuanqi Tan, Baobao Chang, Songfang Huang, and Fei Huang. Raise a child in large language model: Towards effective and generalizable fine-tuning. In Empirical Methods in Natural Language Processing (EMNLP), pages 9514-9528, 2021. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "text", + "content": "[43] Yibo Yang, Xiaojie Li, Zhongzhu Zhou, Shuaiwen Leon Song, Jianlong Wu, Liqiang Nie, and Bernard Ghanem. CorDA: Context-oriented decomposition adaptation of large language models for task-aware parameter-efficient finetuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[44] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10899-10909, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[45] Maxime Zanella and Ismail Ben Ayed. Low-rank few-shot adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (WCVPR), pages 1593-1603, 2024. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[46] Zhi Zhang, Qizhe Zhang, Zijun Gao, Renrui Zhang, Ekaterina Shutova, Shiji Zhou, and Shanghang Zhang. Gradient-based parameter selection for efficient fine-tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28566-28577, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[47] Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. Galore: Memory-efficient llm training by gradient low-rank projection. In International Conference on Machine Learning (ICML), 2024. 3, 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16816-16825, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[50] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2605-2615, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3152" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_content_list.json b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a994dff09294442235bae2bd715b21a0f00f14b --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_content_list.json @@ -0,0 +1,1625 @@ +[ + { + "type": "text", + "text": "Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qian Liang $^{1}$ , Ruixu Geng $^{1}$ , Jinbo Chen $^{2}$ , Haoyu Wang $^{1}$ , Yan Chen $^{1}$ , Yang Hu $^{1*}$ $^{1}$ University of Science and Technology of China, $^{2}$ Nanyang Technological University \n{qianliang, gengruixu, why1999}@mail.ustc.edu.cn, jinbo.chen@ntu.edu.sg \n{eecyan, eeyhu}@ustc.edu.cn", + "bbox": [ + 161, + 202, + 836, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remote physiological measurement (RPM) based on video and radar has made significant progress in recent years. However, unimodal methods based solely on video or radar sensor have notable limitations due to their measurement principles, and multimodal RPM that combines these modalities has emerged as a promising direction. Despite its potential, the lack of large-scale multimodal data and the significant modality gap between video and radar pose substantial challenges in building robust videoradar RPM models. To handle these problems, we suggest leveraging unimodal pre-training and present the Spatial alignment and Temporal Matching (SATM) Adapter to effectively fine-tune pre-trained unimodal backbones into a multimodal RPM model. Given the distinct measurement principles of video- and radar-based methods, we propose Spatial Alignment to align the spatial distribution of their features. Furthermore, Temporal Matching is applied to mitigate waveform discrepancies between video and radar signals. By integrating these two modules into adapters, the unimodal backbones could retain their modality-specific knowledge while effectively extracting complementary features from each other. Extensive experiments across various challenging scenarios, including low light conditions and head motions, demonstrate that our approach significantly surpasses the state-of-the-art methods.", + "bbox": [ + 88, + 340, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 223, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remote physiological measurement has attracted a lot of attention recently owing to its ability to detect physiological signals without contact. A prominent line of research focuses on video-based remote photoplethysmography (rPPG), which uses ordinary cameras to extract vital signals by detecting periodic changes in skin color induced by heartbeats [4, 10, 24, 29, 49, 51]. Beyond video-based", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84f65275c7434545850e7505102f360569ce9f991d5cbf0f4a54965fb7b916f3.jpg", + "image_caption": [ + "Figure 1. Frameworks of video-radar RPM methods: (a) EquiPleth [41] trains unimodal backbones on the multimodal dataset and leverage a late fusion model to ensemble their predictions. (b) Fusion-Vital [6] and CardiacMamba [46] train their end-to-end models from scratch on the multimodal dataset. (c) Our method makes full use of unimodal pre-training and propose the SATM adapter to enable effective multimodal adaptation." + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 906, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods, an alternative approach involves the use of radar sensors to capture cardiac activities through subtle mechanical motions of the human chest [3, 35, 42, 43].", + "bbox": [ + 511, + 579, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, both video- and radar-based methods have notable limitations. Video-based methods are highly sensitive to light conditions and may be influenced by subjects' skin tones [41]. On the other hand, radar-based methods are much more susceptible to subject movements, and most existing methods are designed for strict-constraint scenarios. Therefore, it is necessary and promising to combine video and radar for RPM measurement to overcome the limitations of unimodal approaches.", + "bbox": [ + 511, + 626, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, constructing a robust video-radar RPM model remains a significant challenge. Due to the complexity of recording synchronized videos, radar data, and ground-truth PPG signals, video-radar RPM datasets are mostly limited in scale. Consequently, training a model from scratch on such datasets is prone to severe overfitting. Unfortunately, existing multimodal RPM works [6, 41, 46] overlook this problem, leading to limited robustness and generalization ability (See Sec. 4.3). In addition, video", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8623", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based methods focus primarily on color changes of facial skin, whereas radar-based methods capture mechanical displacements of the chest. This implies that these two modalities capture distinct signals from different regions of interest (ROIs), introducing a substantial modality gap and making it difficult to effectively fuse their features.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the data scarcity challenge, we suggest making full use of unimodal pre-training, as unimodal data are easier to collect and large-scale unimodal datasets for remote physiological measurement are already available [31, 40, 53]. As illustrated in Fig. 1 (c), we first utilize large-scale unimodal datasets to pre-train our backbones. Subsequently, instead of fully fine-tuning the entire network on limited multimodal data, we propose leveraging cross-modal adapters to fine-tune unimodal backbones into a multimodal model while preserving the pre-trained knowledge. Several previous studies have introduced adapter techniques into multimodal learning [2, 28, 48]. For instance, UniAdapter [28] and MMA [48] design parameter-sharing adapters to facilitate modality alignment in vision-language models, while BAT [2] proposes a bi-directional adapter to directly extract cross-modal features for multimodal tracking. However, multimodal adapter for video-radar RPM has not been explored before. Furthermore, the aforementioned methods are suboptimal for multimodal RPM for two reasons: (1) crude modality alignment would hamper the learning of modality-specific features in this context, and (2) the significant modality gap between video and radar, as discussed earlier, impedes the direct fusion of their features.", + "bbox": [ + 91, + 184, + 482, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel SATM adapter to enable effective multimodal fine-tuning on the video-radar RPM dataset. Specifically, the Spatial Alignment (SA) and Temporal Matching (TM) modules are devised to manage the large modality gap between video and radar. To mitigate the gap introduced by their distinct ROIs, Spatial Alignment projects latent features from different modalities into a unified time series space and then utilizes cross-attention to align their spatial distributions. Additionally, to handle the gap caused by different waveforms of detected signals from video and radar, Temporal Matching further applies a matched filter to refine the spatially aligned features along the temporal dimension. These two modules narrow the modality gap collaboratively and are integrated into adapters to effectively extract cross-modal features. To validate the effectiveness of our approach, we collect a challenging dataset, MMRPM, which includes diverse light conditions and subject head motions. Comprehensive experiments conducted on the MMRPM and EquiPleth [41] datasets demonstrate the superiority of our method.", + "bbox": [ + 89, + 531, + 482, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key contributions of this work are summarized as follows:", + "bbox": [ + 89, + 835, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To the best of our knowledge, we are the first to highlight the dilemma of data hungry and data scarcity in", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multimodal RPM. We propose leveraging unimodal pretraining and multimodal adapter to tackle this problem.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To handle the significant modality gap between video and radar, we design the Spatial Alignment and Temporal Matching modules. These modules are further integrated into a novel SATM adapter, enabling effective multimodal fine-tuning while mitigating overfitting.", + "- Extensive experiments across various challenging scenarios show the superiority of the proposed method compared to existing approaches." + ], + "bbox": [ + 513, + 121, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 257, + 663, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Unimodal Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 511, + 282, + 901, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing remote physiological measurement methods focus on video and radar sensors. Given that the optical absorption of hemoglobin molecules fluctuates along with the cardiac cycle, physiological signals can be remotely measured by detecting changes in skin color in RGB videos. Early works [7, 8, 36, 45] mainly utilize blind source separation and color space transformations to enhance the signal-to-noise ratio (SNR) of extracted rPPG signals. However, these methods are based on restrictive assumptions and struggle to handle complex scenarios. With the development of deep learning, researchers have employed 3D-CNN or modified vision transformer (ViT) to extract rPPG signals from raw videos [49, 51] or specifically designed STMaps [27, 32]. In addition to RGB videos, several studies have also explored rPPG measurement using near-infrared and thermal-infrared videos [5, 15, 34].", + "bbox": [ + 511, + 306, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, radar sensors can extract vital signals by detecting subtle movements of the human chest induced by physiological activities. [11] combines 4D beamforming and CNN to capture seismocardiography from radar sensors. [3] leverages the connection between cardiac mechanical activity and electrical activity to introduce the first radar-based electrocardiography (ECG) monitoring framework. [43] proposes a spatio-temporal network for radar-based Heart Rate Variability (HRV) detection. However, due to the minuscule scale of cardiac motions, radar-based physiological measurement often requires subjects to remain in a supine position and keep still, limiting its flexibility for real-world applications.", + "bbox": [ + 511, + 549, + 905, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video-Radar RPM Measurement", + "text_level": 1, + "bbox": [ + 513, + 756, + 803, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the complementary properties of vision and radar modalities, vision-radar fusion has been widely explored in tasks such as object detection [17, 23, 26, 30] and depth estimation [21, 25, 37]. In contrast, video-radar RPM measurement remains an understudied area, with only three related works reported recently [6, 41, 46]. EquiPleth [41] focuses on the skin tone bias problem of video-based methods. They design a late fusion model to ensemble predic", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions from video and radar backbones and propose an adversarial training strategy to enhance the fairness of the final predictions. Fusion-Vital [6] first projects video and radar inputs into a shared time-difference domain and then employs attention mechanism for cross-modal fusion, showcasing comparable performance even in modality corruption scenarios. Similarly, CardiacMamba [46] adopts a weight-sharing VisionMamba [20] to develop an end-to-end video-radar RPM model. However, these works train their models from scratch on small-scale video-radar datasets, making them prone to overfitting the training sets. To overcome this limitation, we propose leveraging the relatively large-scale unimodal datasets for pre-training and design the SATM adapter to facilitate effective multimodal finetuning while preserving pre-trained knowledge.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Adapter Fine-Tuning", + "text_level": 1, + "bbox": [ + 89, + 325, + 290, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The increasing scale of foundation models has brought about the challenge of prohibitive cost and substantial data requirements to fully fine-tune these models. Therefore, parameter-efficient fine-tuning (PEFT) has attracted growing attention [12, 13, 16, 52]. Adapter [12], comprising a pair of projection matrices and a nonlinear activation function, is one of the most representative PEFT methods. Although originally developed for natural language processing (NLP), recent works have extended its application to multimodal settings [14, 28, 44, 48]. [28, 48] leverage parameter-sharing adapters to fine-tune vision-language models and promote modality alignment, which is particularly crucial in vision-text tasks. [2] proposes a bi-directional adapter to foster modality complementarity in multimodal tracking. Considering cross-attention is an effective, yet computationally inefficient mechanism for cross-modal fusion, [22] utilizes latent tokens to efficiently integrate cross-attention into adapters. However, these existing approaches are not well-suited for video-radar RPM. This is primarily because naive modality alignment would hinder the learning of modality-specific features, and the substantial modality gap between video and radar further impedes direct fusion of their features. To this end, we present the SATM adapter, specifically tailored for video-radar RPM fine-tuning.", + "bbox": [ + 89, + 348, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 724, + 181, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we focus on constructing a robust multimodal RPM model on the video-radar RPM dataset. To this end, we first leverage existing large-scale unimodal datasets to pre-train our backbones. Afterwards, we freeze the backbones to preserve pre-trained knowledge and employ both vanilla and SATM adapters to fine-tune them into a video-radar RPM model. The proposed SATM adapter consists of two core components, Spatial Alignment and Temporal Matching, which can effectively narrow the modality gap and extract cross-modal features. The general framework", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for video-radar fine-tuning is depicted in Fig. 2. We will present our method in more detail in the following parts of this section.", + "bbox": [ + 511, + 90, + 906, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Backbone Design", + "text_level": 1, + "bbox": [ + 511, + 143, + 684, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Video Backbone", + "text_level": 1, + "bbox": [ + 511, + 165, + 671, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [19], we utilize the combination of DiffNorm and UniFormer [18] as our video backbone. DiffNorm is a lightweight module that utilizes difference operation and normalization to efficiently integrate the appearance and dynamic features of input videos. UniFormer is an enhanced version of the Vision Transformer (ViT) designed for video analysis. Each UniFormer block consists of Dynamic Position Embedding (DPE), Multi-Head Relation Aggregator (MHRA), and Feed Forward Network (FFN). MHRA is the core component that unifies convolution and attention in a concise transformer format. It employs convolution to tackle feature redundancy in shallow layers while utilizing self-attention for global modeling in deep layers.", + "bbox": [ + 511, + 184, + 906, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given an input video $X_{v} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}$ , the DiffNorm module first transforms it into $F_{v}^{(0)} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}$ as the input of UniFormer. Then the forward process of a UniFormer block can be expressed as:", + "bbox": [ + 511, + 380, + 906, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {v} ^ {(l)} = \\mathrm {D P E} \\left(F _ {v} ^ {(l - 1)}\\right) + F _ {v} ^ {(l - 1)}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 453, + 795, + 472 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(\\operatorname {N o r m} \\left(Y _ {v} ^ {(l)}\\right)\\right) + Y _ {v} ^ {(l)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 474, + 903, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v} ^ {(l)} = \\mathrm {F F N} (\\mathrm {N o r m} (Z _ {v} ^ {(l)})) + Z _ {v} ^ {(l)}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 496, + 813, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{v}^{(l - 1)}$ and $F_{v}^{(l)}$ are the output of the $l - 1$ -th and the $l$ -th UniFormer block, respectively. We will omit the normalization layers in the rest of the sections for brevity.", + "bbox": [ + 511, + 518, + 905, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Radar Preprocessing and Backbone", + "text_level": 1, + "bbox": [ + 511, + 571, + 808, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The fundamental principle of radar-based physiological measurement involves capturing the subtle chest movement caused by cardiac activity. To extract displacement information from the radar intermediate frequency (IF) signal, we first perform range-FFT and coarse range gating to generate a range-time matrix and filter out irrelevant reflections. Afterwards, given that the 2D antenna array enables signal sampling in the spatial domain, we leverage beamforming to separate the reflections from different spatial positions:", + "bbox": [ + 511, + 590, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {B} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\boldsymbol {W} ^ {H} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {M}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 732, + 816, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\left[ \\dots , \\exp \\left(- j \\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {d} _ {r} ^ {(i)}\\right), \\dots \\right] ^ {T} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 753, + 903, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\frac {2 \\pi}{\\lambda} \\left[ \\cos \\theta_ {E} \\cos \\theta_ {A}, \\cos \\theta_ {E} \\sin \\theta_ {A}, \\sin \\theta_ {E} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 775, + 875, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M \\in \\mathbb{R}^{N_C \\times R \\times T_2}$ is the range-time matrix, with $N_C, R, T_2$ indicating the number of virtual channels, range bins and frames, respectively. $\\pmb{W}(\\theta_A, \\theta_E) \\in \\mathbb{R}^{N_C}$ is the steering vector in the direction of azimuth angle $\\theta_A$ and elevation angle $\\theta_E$ , and $\\pmb{d}_r^{(i)}$ is the position of the i-th virtual channel. The shape of the beamformed output $\\pmb{B}$ is", + "bbox": [ + 511, + 808, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4622109d85be9d7d0bbc8cd42166a8a686f3a9afe126f93932a171eff27701d5.jpg", + "image_caption": [ + "Figure 2. The overall framework of our model: Both backbones are based on UniFormer to facilitate cross-modal fusion. Vanilla adapters and SATM adapters are inserted into the shallow and deep layers of UniFormers, respectively, for multimodal fine-tuning. Each SATM adapter comprises two key components: Spatial Alignment and Temporal Matching. The Spatial Alignment module serves to align the spatial distributions of video and radar features, while the Temporal Matching module further refines the aligned features along the temporal dimension. Finally, the refined cross-modal features are passed through a vanilla adapter and added to the corresponding backbone as complementary features." + ], + "image_footnote": [], + "bbox": [ + 132, + 85, + 867, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(\\Theta_A \\times \\Theta_E) \\times R \\times T_2$ , where $(\\Theta_A \\times \\Theta_E)$ is the number of angular bins. Following [3, 43], we further utilize a second-order differentiator to amplify cardiac movements while suppressing noise signals as follows:", + "bbox": [ + 89, + 465, + 482, + 527 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {r, t} = \\left(\\psi_ {t - 3} + \\psi_ {t + 3}\\right) + 2 \\left(\\psi_ {t - 2} + \\psi_ {t + 2}\\right) \\tag {3} \\\\ - \\left(\\psi_ {t - 1} + \\psi_ {t + 1}\\right) - 4 \\psi_ {t} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 537, + 480, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi_t$ is the phase of $\\pmb{B}$ at the t-th frame. The output $F_r^{(0)} \\in \\mathbb{R}^{(\\Theta_A \\times \\Theta_E) \\times R \\times T_2}$ is reshaped into $\\mathbb{R}^{1 \\times S \\times T_2}$ to serve as the input of the radar backbone, where $S = \\Theta_A \\times \\Theta_E \\times R$ is the spatial dimension of the radar modality.", + "bbox": [ + 89, + 580, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the necessity to model the mechanical motion of the human chest across both spatial and temporal dimensions [43], we adopt 2D UniFormer as our radar backbone. Overall, our video and radar backbones are symmetrically structured, which greatly facilitates cross-modal fusion.", + "bbox": [ + 89, + 643, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SATM Adapter", + "text_level": 1, + "bbox": [ + 89, + 742, + 246, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable effectively fine-tuning of the unimodal backbones and facilitate cross-modal modeling, we propose the SATM adapter. As illustrated in Fig. 2, Radar-to-Video SATM adapters and Video-to-Radar SATM adapters are inserted in parallel with the MHRA and FFN modules of the video and radar UniFormers respectively to extract complementary features from the other modality. As discussed in Sec. 1, the fundamental differences in measurement principles between video- and radar-based methods introduce", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a significant modality gap between their features and implicate cross-modal fusion. Considering the different ROIs of video and radar (face vs. chest), we present the Spatial Alignment (SA) module to align the spatial distributions of their features. As for the differences in captured signals (color changes vs. mechanical displacements), we design the Temporal Matching (TM) module to reduce waveform discrepancies and refine the spatially aligned features. Next, we will provide a detailed explanation of these modules.", + "bbox": [ + 511, + 467, + 906, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Spatial Alignment", + "text_level": 1, + "bbox": [ + 511, + 609, + 687, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The physiological information contained in the face and chest is inherently distributed in distinct forms, resulting in different spatial distributions of video and radar features. We would like to emphasize that this dissimilarity is not merely reflected in the different spatial dimensions of $F_{v}$ and $F_{r}$ ( $HW$ vs. $S$ ), and naively aligning their size through interpolation or linear projection is insufficient.", + "bbox": [ + 511, + 628, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this challenge, we design the Spatial Alignment module. Given the video feature $F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}$ and radar feature $F_{r} \\in \\mathbb{R}^{C \\times s \\times t_{2}}$ , the core idea involves utilizing cross-attention to align the spatial distributions of their spatial tokens, i.e. $f_{v} \\in \\mathbb{R}^{C \\times t_{1}}$ and $f_{r} \\in \\mathbb{R}^{C \\times t_{2}}$ . Without loss of generality, we use the Radar-to-Video Spatial Alignment as an example for explanation, with its overall pipeline illustrated in Fig. 3 (a). First, bilinear interpolation is applied to align the temporal dimensions of $F_{v}$ and $F_{r}$ . The primary challenge then lies in effectively transforming the 2D spatial tokens for the calculation of the attention", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8626", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/792d1c5ed9721d78fcf6a702d10073cd07d5ab30d9cc87a916cecadcf81003ca.jpg", + "image_caption": [ + "Figure 3. The pipelines of Radar-to-Video Spatial Alignment and Temporal Matching. \"MatMul\" and \"Hadamard\" denote matrix product and Hadamard product, respectively. $\\mathrm{FFT}_T$ represents FFT along the temporal dimension." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 483, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scores. Considering that the final predicted 1D PPG signals are generated by the integration of the spatial and channel dimensions, we propose treating spatial tokens as distinct candidates for PPG signal prediction at different spatial locations and then projecting them from the 2D space into a unified 1D time series space by compressing their channel dimensions. Specifically, we leverage linear projections to transform $F_{v}$ and $F_{r}$ into the query matrix $S_{v} \\in \\mathbb{R}^{hw\\times t_{1}}$ and the key matrix $S_{r} \\in \\mathbb{R}^{s\\times t_{1}}$ , respectively. Afterwards, the alignment between video and radar spatial tokens can be performed in the time series space and the resulting aligned radar features are computed as follows:", + "bbox": [ + 88, + 368, + 483, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {S _ {v} S _ {r} ^ {T}}{\\sqrt {t _ {1}}}\\right) F _ {r} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 559, + 482, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the matrix multiplication between the attention score and $F_{r}$ is performed channel-wise.", + "bbox": [ + 89, + 602, + 482, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now let us revisit the way to transform the spatial tokens. Two alternative approaches involve compressing the temporal dimension or flattening the channel and temporal dimensions of spatial tokens. However, the former variant risks irreversible information loss due to excessive compression of temporal signals, which is paramount for the prediction of PPG signals. The second approach, while straightforward, fails to account for the inherent differences between the channel and temporal dimensions. In contrast, our spatial alignment, which operates in the time series space, adheres to the paradigm of PPG signal regression and has the advantage of preserving temporal information. A quantitative comparison is provided in Sec. 4.4.", + "bbox": [ + 88, + 632, + 482, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Temporal Matching", + "text_level": 1, + "bbox": [ + 89, + 835, + 276, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although video and radar features are temporally synchronized and ideally encompass consistent frequency information, i.e. heart rate, the inherent disparity between wave", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "forms derived from skin color changes and chest mechanical movements introduces a significant modality gap in the temporal dimension. To this end, we further utilize Temporal Matching to refine the cross-modal features. According to the Matched Filter theory, given an input signal $r(t)$ composed of the target signal $s(t)$ and the noise $n(t)$ , an output signal with the highest SNR at time step $T$ could be obtained by applying the filter $h(t) = s(T - t)$ . Inspired by this, we formulate the spatially aligned features as $r(t)$ and the waveform discrepancy as $n(t)$ , and the refined feature could be obtained through matched filtering. Specifically, given the video feature $F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}$ and the spatially aligned radar feature $F_{r}^{\\prime} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}$ , the refined radar-to-video feature can be computed by:", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T M} _ {r 2 v} \\left(F _ {v}, F _ {r} ^ {\\prime}\\right) = \\mathrm {D W c o n v} \\left(\\mathrm {B N} \\left(\\mathcal {F} _ {T} ^ {- 1} \\left(\\overline {{\\mathcal {F} _ {T} \\left(F _ {v}\\right)}} \\mathcal {F} _ {T} \\left(F _ {r} ^ {\\prime}\\right)\\right)\\right) \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 308, + 903, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_T(\\cdot)$ denotes Fast Fourier Transform (FFT) along the temporal dimension and $\\overline{\\mathcal{F}_T(\\cdot)}$ represents FFT followed by the conjugate operation. DWConv and BN are depthwise convolution and batch normalization, respectively.", + "bbox": [ + 511, + 340, + 903, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Cross-Modal Fusion", + "text_level": 1, + "bbox": [ + 511, + 406, + 700, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the forward process of the Radar-to-Video SATM adapter can be represented as:", + "bbox": [ + 511, + 425, + 903, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {r 2 v} = \\mathrm {T M} _ {r 2 v} \\left(F _ {v}, \\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 462, + 846, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A T M} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = W _ {u p} \\left(\\operatorname {R e L U} \\left(W _ {\\text {d o w n}} \\left(F _ {v} + \\alpha \\cdot F _ {r 2 v}\\right)\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 481, + 911, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{down}, W_{up}$ are the down-projection and up-projection layers of the vanilla adapter [12] and $\\alpha$ is a learnable scaling factor. Afterward, the forward process of a video UniFormer block can be reformulated as:", + "bbox": [ + 511, + 511, + 906, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(Y _ {v} ^ {(l)}\\right) + Y _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Y _ {v} ^ {(l)}, Y _ {r} ^ {(l)}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 574, + 887, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {v} ^ {(l)} = \\operatorname {F F N} \\left(Z _ {v} ^ {(l)}\\right) + Z _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Z _ {v} ^ {(l)}, Z _ {r} ^ {(l)}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 595, + 903, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The forward processes of the Video-to-Radar SATM adapter and the fine-tuned radar UniFormer blocks can also be computed similarly to Eq. (6) and Eq. (7).", + "bbox": [ + 511, + 628, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, Spatial Alignment aligns the spatial distributions of video and radar features in a unified time series space, while Temporal Matching utilizes a matched filter to handle the waveform difference between their signals. They complement each other by processing the cross-modal features along spatial and temporal dimensions separately, thereby effectively narrowing the substantial modality gap between video and radar features. According to [9], it is essential to balance the learning of unimodal and cross-modal features. Consequently, we only insert our SATM adapters into the last two stages of UniFormers while utilizing vanilla adapters in the first two stages. The output features of both backbones are integrated by a lightweight fusion stem and subsequently passed through a regression module to generate the predicted PPG signal.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 222, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 114, + 187, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize the EquiPleth [41] dataset to evaluate our method, which contains 533 30-second samples collected from 28 light, 49 medium, and 14 dark skin tone volunteers. All samples are recorded under motionless conditions and with sufficient illumination.", + "bbox": [ + 89, + 136, + 482, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, considering that the EquiPleth dataset mainly focuses on constrained scenarios and shows limited practical diversity, we built a new dataset called MMRPM to promote a more comprehensive evaluation of our method. MMRPM comprises 180 60-second samples from 31 subjects, all of which are recorded under varying illumination conditions and subject motions.", + "bbox": [ + 88, + 212, + 482, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, MMRPM encompasses three illumination conditions: dark, medium, and light. The \"dark\" and \"medium\" settings correspond to illumination levels of approximately 2.5 and 6.3 lux respectively, while \"light\" represents samples are collected under sufficient sunlight and indoor illumination. Furthermore, two types of subject activities are considered: sitting still and free head movement. Overall, MMRPM covers six scenarios, where low-light conditions are challenging to the video modality and head motions have a greater impact on the radar modality. $^{1}$", + "bbox": [ + 89, + 318, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 479, + 307, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed method is implemented using PyTorch. For video preprocessing, we first employ MTCNN to crop the face region in the initial frame and fix the region across subsequent frames. Then, we sample each video into clips with a time window of 160 frames and a step size of 80 frames, which are then resized to $96 \\times 96$ pixels. In terms of radar preprocessing, the angular range selected for beamforming is $66^{\\circ} \\leq \\Theta_{A} \\leq 114^{\\circ}, -24^{\\circ} \\leq \\Theta_{E} \\leq 24^{\\circ}$ .", + "bbox": [ + 89, + 501, + 482, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize UniFormer-S [18] and MAM in MAGIC [55] as our feature extractor and fusion stem respectively. As for the regression module, we follow the design in ADDP [19]. The video backbone is jointly pre-trained on VIPL [31, 32], MMPD [40], UBFC [1], PURE [39] and BUAA [47], and we use the dataset introduced in [53] for radar pre-trianing.", + "bbox": [ + 89, + 622, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During multimodal fine-tuning, both backbones are frozen, and only the adapters, fusion stem and regression module are trained. We employ the loss functions introduced in [51] and utilize the Adam optimizer for training, with the learning rate and weight decay set to 5e-4 and 5e-5 respectively. The model is trained for 40 epochs with a batch size of 2. Following [38], we apply random horizontal flipping, spatially resized crop and random intensity noise to augment the input videos. For radar augmentation, we randomly shift the range bins of the range-time matrices.", + "bbox": [ + 89, + 714, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [41], we perform our evaluation of the PPG signals over 300 samples windows with a stride of 128 samples. The heart rate corresponding to a specific signal clip is calculated using a combination of band-pass filtering, detrending and peak detection in the frequency domain. Finally, the standard deviation of the error (Std), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (R) are adopted to evaluate the predicted HR error. The unit of Std, MAE and RMSE is beats per minute (bpm). We split the MMRPM dataset into 5 subject-exclusive folds and use the official split for the EquiPleth dataset.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 285, + 651, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare our SATM with existing video-radar RPM methods [6, 41, 46] and several state-of-the-art multimodal adapter methods [2, 22, 28, 48] on the EquiPleth [41] dataset. To ensure a fair comparison, we provide extra results of improved versions of existing RPM methods by applying unimodal pre-training and replacing their original backbones with UniFormers. Since the radar used in the EquiPleth dataset activates only 1 transmitter and 1 receiver, we do not perform beamforming during radar preprocessing for this dataset. For existing video-radar RPM methods, we use the official implementation of EquiPleth, and re-implement Fusion-Vital and CardiacMamba, as the codes have not been released by the authors. As shown in Tab. 1, all of these methods exhibit RMSEs exceeding 2.9 bpm when trained from scratch. Their performance consistently improves when unimodal pre-training is applied, underscoring the significance of unimodal pre-training when video-radar data is limited. However, they still display MAEs greater than 0.64 bpm, regardless of whether using their original backbones or UniFormers, indicating that their full fine-tuning-based fusion strategies are suboptimal for the small-scale multimodal RPM dataset. In contrast, our method employs SATM adapters for multimodal fine-tuning, effectively leveraging pre-trained knowledge and enabling superior cross-modal fusion.", + "bbox": [ + 511, + 308, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, our SATM also demonstrates superior performance compared to existing multimodal adapter methods. The subpar performance of UniAdapter and MMA suggests that their parameter-sharing strategies are ill-suited for multimodal RPM, as they impede the learning of modality-specific features. To adapt BAT for video-radar RPM, we employ an additional linear projection layer to align the token numbers of the video and radar branches. However, it fails to handle the significant modality gap between video and radar, thereby resulting in an RMSE exceeding 1.65 bpm. While the best baseline, LAVISH, achieves an RMSE of 1.498 bpm, our SATM yields a considerably lower RMSE of 1.242 bpm.", + "bbox": [ + 511, + 686, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We have also evaluated SATM on our self-collected", + "bbox": [ + 532, + 885, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "For more details about our MMRPM dataset and pre-training, please refer to the supplemental materials.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8628", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9eaa949fdcc92ac672026c95905a0c3993b9849deaf0108bec73e6d05c53f49b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsOriginal ModalityUnimodal Pre-trainedStd↓MAE↓RMSE↓R↑Trainable Params (M)↓
EquiPleth[41]Video-Radar×3.2741.0633.2830.9573.95
Fusion-Vital[6]Video-Radar×9.2883.7509.2970.709715.26
CardiacMamba[46]Video-Radar×2.9240.9012.9270.9759.93
EquiPlethVideo-Radar√2.9790.9682.9900.9653.95
Fusion-VitalVideo-Radar√9.1213.7509.1300.735715.26
CardiacMambaVideo-Radar√2.7970.8682.7710.9759.93
EquiPleth*Video-Radar√1.4350.6711.4360.99245.31
Fusion-Vital*Video-Radar√2.1430.6502.1570.983194.61
CardiacMamba*Video-Radar√2.3130.7442.3240.98033.62
UniAdapter[28]Vision-Language√1.6530.5901.6560.99010.60
MMA[48]Vision-Language√1.8350.6261.8420.98811.09
BAT[2]RGB-TIR√1.6790.6301.6840.99085.35
LAVISH[22]Audio-Visual√1.4930.5571.4980.99110.85
SATM (Ours)Video-Radar√1.2170.5401.2420.99410.78
", + "bbox": [ + 133, + 88, + 864, + 342 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/236595545519b7861ca36304020c4239d0b1469fe6c9f66cc14bb78548035d1e.jpg", + "table_caption": [ + "Table 1. Overall performance on the EquiPleth dataset. The best and second best results are in bold and underlined respectively. * denotes that we replace the methods' original backbones with UniFormers while preserving their fusion strategies. Blue and red indicate the existing video-radar RPM methods and existing adapter methods for other multimodal tasks, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth7.6032.4217.6030.715
Fusion-Vital9.1395.1659.3150.501
CardiacMamba6.0992.1606.1070.790
EquiPleth*5.9641.7395.9740.806
Fusion-Vital*5.5501.6625.5890.827
CardiacMamba*5.5261.8435.5640.829
UniAdapter5.5091.5885.5100.831
MMA5.6941.7225.6940.819
BAT6.4322.1516.4330.760
LAVISH5.6601.4545.3120.844
SATM (Ours)5.0631.2925.0740.858
", + "bbox": [ + 96, + 417, + 478, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MMRPM dataset, which covers more challenging and diverse scenarios than the EquiPleth dataset. As illustrated in Tab. 2, existing video-radar RPM methods exhibit poor results on this dataset, highlighting their limited capability to handle complex scenarios. In comparison, our SATM consistently outperforms the baseline methods in all metrics.", + "bbox": [ + 89, + 670, + 482, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conduct a cross-dataset evaluation by training the models on the EquiPleth dataset and evaluating them on the MMRPM dataset. As shown in Tab. 3, the improved versions of existing multimodal RPM methods exhibit subpar performance compared to adapter-based methods. This is primarily because they rely on full fine-tuning-based fusion strategies and are prone to overfitting small-scale training sets. In contrast, our method, built upon SATM adapters, achieves the best generalization performance and suffers", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/169b3ea8b28cd7e7ecf73632a2b9298223cfaced9ce02c2abd757a8175c31a73.jpg", + "table_caption": [ + "Table 2. Overall performance on the MMRPM dataset." + ], + "table_footnote": [], + "table_body": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth*8.7874.2569.0820.629
Fusion-Vital*8.5384.4578.8970.642
CardiacMamba*9.0485.0389.3080.603
LAVISH8.0263.6878.2320.712
SATM (Ours)7.5423.4237.6560.738
", + "bbox": [ + 519, + 417, + 901, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of the cross-dataset evaluation from the EquiPleth dataset to the MMRPM dataset.", + "bbox": [ + 511, + 531, + 906, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "minimal performance degradation compared to other baselines. Although it still lags behind the performance in the intra-dataset evaluation, we argue that this is reasonable, given the significant differences between these two datasets.[2]", + "bbox": [ + 511, + 587, + 906, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 672, + 668, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Components. Our SATM adapter mainly consists of two key components: Spatial Alignment and Temporal Matching. In Tab. 4, we conduct an ablation study on the MMRPM dataset to assess the effectiveness of these modules. First, compared to the vanilla adapter (the first row), the adapter with Spatial Alignment reduces the MAE and RMSE by 0.087 bpm and 0.149 bpm respectively. This can be attributed to its capability to facilitate modality complementarity by extracting spatially aligned cross-modal features to the corresponding backbones. The MAE and RMSE can be further reduced to 1.292 bpm and 5.074 bpm when Temporal Matching is inserted, highlighting the necessity", + "bbox": [ + 511, + 695, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Details about cross-dataset evaluation are provided the appendix.", + "bbox": [ + 529, + 886, + 880, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fe2bf6e8af76ae772070089209b1bf1bbc0dc7708066e34d8d1250731256ca0d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SATMStd↓MAE↓RMSE↓R↑
××5.4521.5415.4530.836
✓×5.2981.4545.3040.845
✓✓5.0631.2925.0740.858
", + "bbox": [ + 119, + 88, + 454, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91f6c66af950c80a533335e6b8c8954e745fdac5dd366c806f91b72f6e97de16.jpg", + "table_caption": [ + "Table 4. Ablation study on the key components. \"SA\" and \"TM\" denote Spatial Alignment and Temporal Matching respectively." + ], + "table_footnote": [], + "table_body": "
VariantsStd↓MAE↓RMSE↓R↑
Interpolate6.0411.8766.0520.795
Linear5.5791.6285.5820.828
Channel6.0871.8076.0910.792
Flatten5.5881.6935.5950826
Ours5.2981.4545.3040.845
", + "bbox": [ + 114, + 213, + 459, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to refine the aligned features along the temporal dimension.", + "bbox": [ + 89, + 369, + 482, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative Approaches for Spatial Alignment. We also investigate four alternative approaches for Spatial Alignment. Specifically, \"Interpolate\" and \"Linear\" refer to methods that directly align the sizes of video and radar features through interpolation and linear projection instead of cross-attention. The \"Channel\" and \"Flatten\" indicate that the alignment is performed in the channel space and on the flattened spatial tokens respectively. As shown in Tab. 5, both \"Interpolate\" and \"Linear\" exhibit suboptimal performance, achieving MAEs of 1.876 bpm and 1.628 bpm. These methods only align spatial sizes while neglecting inherent differences in spatial distributions between video and radar features, thereby failing to address the modality gap.", + "bbox": [ + 88, + 385, + 482, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The \"Channel\" variant compresses the temporal dimensions of spatial tokens for the computation of the attention scores, focusing solely on the overall temporal period while neglecting fine-grained temporal variability. This results in degraded performance with an RMSE exceeding 6 bpm. The \"Flatten\" variant preserves temporal information, but conflates temporal and channel dimensions, overlooking their intrinsic differences. Therefore, it ultimately yields subpar performance with an MAE of 1.693 bpm and an RMSE of 5.595 bpm. In contrast, Spatial Alignment performed in the time series space effectively avoids temporal information loss and dimensional distortion, showing superior performance compared to these four variants.", + "bbox": [ + 89, + 583, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Matching vs. Temporal Attention. Since attention is a straightforward method for cross-modal fusion, we compare our Temporal Matching (TM) with Temporal Attention (TA). Specifically, we replace the matched filtering operation with a cross-attention along the temporal dimension. As shown in Tab. 6, our TM consistently outperforms TA across all metrics. Moreover, since the time complexity of FFT is $\\mathcal{O}(N\\log N)$ , TM is also more com", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e2953e1bd1bfddaffef3be02d3511cafdadb7b0bbbc702191ca5c199621c3933.jpg", + "table_caption": [ + "Table 5. Ablation study on the design of Spatial Alignment." + ], + "table_footnote": [], + "table_body": "
TemporalStd↓MAE↓RMSE↓R↑
Attention5.1551.3675.1300.853
Matching5.0631.2925.0740.858
", + "bbox": [ + 542, + 88, + 879, + 145 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/600e2dfc610c370a073f7aa51e1a993c04ae9a037cbccadc30188fa239c7fba3.jpg", + "table_caption": [ + "Table 6. Comparison between Temporal Matching and Temporal Attention." + ], + "table_footnote": [], + "table_body": "
PositionsStd↓MAE↓RMSE↓R↑
SATM-45.1681.3865.1790.852
SATM-35.1071.3435.1090.857
SATM-25.0631.2925.0740.858
SATM-15.5261.6145.5260.830
", + "bbox": [ + 544, + 196, + 877, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the position to add SATM adapters.", + "bbox": [ + 522, + 292, + 893, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "putationally efficient compared to the quadratic complexity of TA.", + "bbox": [ + 511, + 334, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Position to Insert Our SATM Adapters. We also explore the impact of different positions to insert our SATM adapters. We present 4 positions for insertion, formulated as \"SATM-n\", where \"n\" denotes that we use SATM adapters in the last \"n\" stages of the UniFormers and employ vanilla adapters in earlier stages. As reported in Tab. 7, SATM-1 displays an MAE higher than the naive baseline (1.614 bpm vs. 1.541 bpm), indicating that inserting the cross-modal module too late would hinder performance due to the lack of cross-modal interaction. Meanwhile, the performance of SATM-3 and SATM-4 is also limited due to insufficient modality-specific modeling. In contrast, the results demonstrate that SATM-2 strikes a good trade-off between the learning of unimodal and corss-modal features.", + "bbox": [ + 511, + 364, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focus on video-radar RPM measurement. To overcome the data scarcity challenge, we make full use of unimodal pre-training and present the SATM adapter to effectively fine-tune unimodal backbones into a multimodal RPM model. The SATM adapter leverages Spatial Alignment and Temporal Matching to collaboratively narrow the modality gap between video and radar and effectively extract cross-modal features for both backbones. To thoroughly evaluate our method, we collect a challenging dataset that encompasses diverse illumination conditions and subject motions. Extensive experiments conducted on our self-collected and existing public datasets demonstrate that SATM outperforms existing methods.", + "bbox": [ + 511, + 613, + 906, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While SATM effectively constructs a robust video-radar RPM model, it primarily focuses on cross-modal interaction and still relies on vanilla adapters for unimodal modeling during fine-tuning. In the future, we hope to develop a unified framework that harmonizes cross-modal and unimodal modeling.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8630", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62172381.", + "bbox": [ + 89, + 113, + 485, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6, 1", + "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 2, 3, 6, 7", + "[3] Jinbo Chen, Dongheng Zhang, Zhi Wu, Fang Zhou, Qibin Sun, and Yan Chen. Contactless electrocardiogram monitoring with millimeter wave radar. IEEE Transactions on Mobile Computing, 23(1):270-285, 2022. 1, 2, 4", + "[4] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 349-365, 2018. 1", + "[5] Li-Wen Chiu, Yang-Ren Chou, Yi-Chiao Wu, and Bing-Fei Wu. Deep-learning based remote photoplethysmography measurement in driving scenarios with color and near-infrared images. IEEE Transactions on Instrumentation and Measurement, 2023. 2", + "[6] Jae-Ho Choi, Ki-Bong Kang, and Kyung-Tae Kim. Fusion-vital: Video-rf fusion transformer for advanced remote physiological measurement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1344-1352, 2024. 1, 2, 3, 6, 7", + "[7] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE transactions on biomedical engineering, 60(10):2878-2886, 2013. 2", + "[8] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiological measurement, 35(9):1913, 2014. 2", + "[9] Chenzhuang Du, Jiaye Teng, Tingle Li, Yichen Liu, Tianyuan Yuan, Yue Wang, Yang Yuan, and Hang Zhao. On uni-modal feature learning in supervised multi-modal learning. In International Conference on Machine Learning, pages 8632-8656. PMLR, 2023. 5", + "[10] Jingda Du, Si-Qi Liu, Bochao Zhang, and Pong C Yuen. Dual-bridging with adversarial noise generation for domain adaptive rpgg estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10355–10364, 2023. 1", + "[11] Unsoo Ha, Salah Assana, and Fadel Adib. Contactless seismocardiography via deep learning radars. In Proceedings of the 26th annual international conference on mobile computing and networking, pages 1-14, 2020. 2", + "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3, 5" + ], + "bbox": [ + 93, + 193, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[14] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, XueQing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 3", + "[15] Jitesh Joshi and Youngjun Cho. ibvp dataset: Rgb-thermal rppg dataset with high resolution signal quality labels. Electronics, 13(7):1334, 2024. 2", + "[16] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 3", + "[17] Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17615-17626, 2023. 2", + "[18] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 3, 6, 2", + "[19] Qian Liang, Yan Chen, and Yang Hu. Continual learning for remote physiological measurement: Minimize forgetting and simplify inference. In European conference on computer vision, pages 126-144. Springer, 2024. 3, 6", + "[20] Zhu Lianghui, Liao Bencheng, Zhang Qian, Wang Xinlong, Liu Wenyu, and Wang Xinggang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv. Org, pages arXiv-org, 2024. 3", + "[21] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 2", + "[22] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2299–2309, 2023. 3, 6, 7", + "[23] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2", + "[24] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 1", + "[25] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3343-3347. IEEE, 2021. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8631", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yunfei Long, Abhinav Kumar, Daniel Morris, Xiaoming Liu, Marcos Castro, and Punarjay Chakravarty. Radiant: Radar-image association network for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1808-1816, 2023. 2", + "[27] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 2", + "[28] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 2, 3, 6, 7", + "[29] Hao Lu, Zitong Yu, Xuesong Niu, and Ying-Cong Chen. Neuron structure modeling for generalizable remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18589-18599, 2023. 1", + "[30] Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2021. 2", + "[31] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14, pages 562–576. Springer, 2019. 2, 6, 1", + "[32] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 2, 6, 1", + "[33] Xuesong Niu, Xingyuan Zhao, Hu Han, Abhijit Das, Antitza Dantcheva, Shiguang Shan, and Xin Chen. Robust remote heart rate estimation from face utilizing spatial-temporal attention. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-8. IEEE, 2019. 1", + "[34] Soyeon Park, Bo-Kyeong Kim, and Suh-Yeon Dong. Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation. IEEE Transactions on Instrumentation and Measurement, 71:1-10, 2022. 2", + "[35] Vladimir L Petrović, Milica M Janković, Anita V Lupšić, Veljko R Mihajlović, and Jelena S Popovic-Božović. High-accuracy real-time monitoring of heart rate variability using 24 ghz continuous-wave doppler radar. *IEEE Access*, 7: 74721–74733, 2019. 1", + "[36] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10):10762-10774, 2010. 2", + "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 6", + "[39] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6, 1", + "[40] Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, and Xin Liu. Mmpd: multidomain mobile video physiology dataset. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5. IEEE, 2023. 2, 6, 1", + "[41] Alexander Vilesov, Pradyumna Chari, Adnan Armouti, Anirudh Bindiganavale Harish, Kimaya Kulkarni, Ananya Deoghare, Laleh Jalilian, and Achuta Kadambi. Blending camera and 77 ghz radar sensing for equitable, robust plethysmography. ACM Trans. Graph., 41(4):36-1, 2022. 1, 2, 6, 7", + "[42] Fengyu Wang, Xiaolu Zeng, Chenshu Wu, Beibei Wang, and KJ Ray Liu. mmhrv: Contactless heart rate variability monitoring using millimeter-wave radio. IEEE Internet of Things Journal, 8(22):16623-16636, 2021. 1", + "[43] Haoyu Wang, Jinbo Chen, Dongheng Zhang, Zhi Lu, Changwei Wu, Yang Hu, Qibin Sun, and Yan Chen. Contactless radar heart rate variability monitoring via deep spatiotemporal modeling. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 111-115. IEEE, 2024. 1, 2, 4", + "[44] Kai Wang, Yapeng Tian, and Dimitrios Hatzinakos. Towards efficient audio-visual learners via empowering pre-trained vision transformers with cross-modal adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1837–1846, 2024. 3", + "[45] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE transactions on Biomedical Engineering, 62(2): 415-425, 2014. 2", + "[46] Zheng Wu, Yiping Xie, Bo Zhao, Jiguang He, Fei Luo, Ning Deng, and Zitong Yu. Cardiacmamba: A multimodal rgb-rf fusion framework with state space models for remote physiological measurement. arXiv preprint arXiv:2502.13624, 2025. 1, 2, 3, 6, 7", + "[47] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6, 1", + "[48] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2, 3, 6, 7", + "[49] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from fa" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8632", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cial videos using spatio-temporal networks. arXiv preprint arXiv:1905.02419, 2019. 1, 2", + "[50] Zitong Yu, Xiaobai Li, Xuesong Niu, Jingang Shi, and Guoying Zhao. Autohr: A strong end-to-end baseline for remote heart rate measurement with neural searching. IEEE Signal Processing Letters, 27:1245-1249, 2020. 1", + "[51] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip HS Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4186-4196, 2022. 1, 2, 6", + "[52] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021.3", + "[53] Bin-Bin Zhang, Dongheng Zhang, Yadong Li, Zhi Lu, Jinbo Chen, Haoyu Wang, Fang Zhou, Yu Pu, Yang Hu, Li-Kun Ma, et al. Monitoring long-term cardiac activity with contactless radio frequency signals. Nature Communications, 15(1):1-11, 2024. 2, 6, 1", + "[54] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 1", + "[55] Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. Centering the value of every modality: Towards efficient and resilient modality-agnostic semantic segmentation. In European Conference on Computer Vision, pages 192–212. Springer, 2024. 6" + ], + "bbox": [ + 91, + 90, + 483, + 501 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8633", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_model.json b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_model.json new file mode 100644 index 0000000000000000000000000000000000000000..150157e9d6cee38e0001b43236c3b25da055d566 --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.203, + 0.837, + 0.277 + ], + "angle": 0, + "content": "Qian Liang\\(^{1}\\), Ruixu Geng\\(^{1}\\), Jinbo Chen\\(^{2}\\), Haoyu Wang\\(^{1}\\), Yan Chen\\(^{1}\\), Yang Hu\\(^{1*}\\) \n\\(^{1}\\)University of Science and Technology of China, \\(^{2}\\)Nanyang Technological University \n{qianliang, gengruixu, why1999}@mail.ustc.edu.cn, jinbo.chen@ntu.edu.sg \n{eecyan, eeyhu}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.328, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.721 + ], + "angle": 0, + "content": "Remote physiological measurement (RPM) based on video and radar has made significant progress in recent years. However, unimodal methods based solely on video or radar sensor have notable limitations due to their measurement principles, and multimodal RPM that combines these modalities has emerged as a promising direction. Despite its potential, the lack of large-scale multimodal data and the significant modality gap between video and radar pose substantial challenges in building robust videoradar RPM models. To handle these problems, we suggest leveraging unimodal pre-training and present the Spatial alignment and Temporal Matching (SATM) Adapter to effectively fine-tune pre-trained unimodal backbones into a multimodal RPM model. Given the distinct measurement principles of video- and radar-based methods, we propose Spatial Alignment to align the spatial distribution of their features. Furthermore, Temporal Matching is applied to mitigate waveform discrepancies between video and radar signals. By integrating these two modules into adapters, the unimodal backbones could retain their modality-specific knowledge while effectively extracting complementary features from each other. Extensive experiments across various challenging scenarios, including low light conditions and head motions, demonstrate that our approach significantly surpasses the state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.224, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Remote physiological measurement has attracted a lot of attention recently owing to its ability to detect physiological signals without contact. A prominent line of research focuses on video-based remote photoplethysmography (rPPG), which uses ordinary cameras to extract vital signals by detecting periodic changes in skin color induced by heartbeats [4, 10, 24, 29, 49, 51]. Beyond video-based" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.452, + 0.909, + 0.552 + ], + "angle": 0, + "content": "Figure 1. Frameworks of video-radar RPM methods: (a) EquiPleth [41] trains unimodal backbones on the multimodal dataset and leverage a late fusion model to ensemble their predictions. (b) Fusion-Vital [6] and CardiacMamba [46] train their end-to-end models from scratch on the multimodal dataset. (c) Our method makes full use of unimodal pre-training and propose the SATM adapter to enable effective multimodal adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.626 + ], + "angle": 0, + "content": "methods, an alternative approach involves the use of radar sensors to capture cardiac activities through subtle mechanical motions of the human chest [3, 35, 42, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.764 + ], + "angle": 0, + "content": "However, both video- and radar-based methods have notable limitations. Video-based methods are highly sensitive to light conditions and may be influenced by subjects' skin tones [41]. On the other hand, radar-based methods are much more susceptible to subject movements, and most existing methods are designed for strict-constraint scenarios. Therefore, it is necessary and promising to combine video and radar for RPM measurement to overcome the limitations of unimodal approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Nevertheless, constructing a robust video-radar RPM model remains a significant challenge. Due to the complexity of recording synchronized videos, radar data, and ground-truth PPG signals, video-radar RPM datasets are mostly limited in scale. Consequently, training a model from scratch on such datasets is prone to severe overfitting. Unfortunately, existing multimodal RPM works [6, 41, 46] overlook this problem, leading to limited robustness and generalization ability (See Sec. 4.3). In addition, video" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8623" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "based methods focus primarily on color changes of facial skin, whereas radar-based methods capture mechanical displacements of the chest. This implies that these two modalities capture distinct signals from different regions of interest (ROIs), introducing a substantial modality gap and making it difficult to effectively fuse their features." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.185, + 0.483, + 0.53 + ], + "angle": 0, + "content": "For the data scarcity challenge, we suggest making full use of unimodal pre-training, as unimodal data are easier to collect and large-scale unimodal datasets for remote physiological measurement are already available [31, 40, 53]. As illustrated in Fig. 1 (c), we first utilize large-scale unimodal datasets to pre-train our backbones. Subsequently, instead of fully fine-tuning the entire network on limited multimodal data, we propose leveraging cross-modal adapters to fine-tune unimodal backbones into a multimodal model while preserving the pre-trained knowledge. Several previous studies have introduced adapter techniques into multimodal learning [2, 28, 48]. For instance, UniAdapter [28] and MMA [48] design parameter-sharing adapters to facilitate modality alignment in vision-language models, while BAT [2] proposes a bi-directional adapter to directly extract cross-modal features for multimodal tracking. However, multimodal adapter for video-radar RPM has not been explored before. Furthermore, the aforementioned methods are suboptimal for multimodal RPM for two reasons: (1) crude modality alignment would hamper the learning of modality-specific features in this context, and (2) the significant modality gap between video and radar, as discussed earlier, impedes the direct fusion of their features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.834 + ], + "angle": 0, + "content": "To this end, we propose a novel SATM adapter to enable effective multimodal fine-tuning on the video-radar RPM dataset. Specifically, the Spatial Alignment (SA) and Temporal Matching (TM) modules are devised to manage the large modality gap between video and radar. To mitigate the gap introduced by their distinct ROIs, Spatial Alignment projects latent features from different modalities into a unified time series space and then utilizes cross-attention to align their spatial distributions. Additionally, to handle the gap caused by different waveforms of detected signals from video and radar, Temporal Matching further applies a matched filter to refine the spatially aligned features along the temporal dimension. These two modules narrow the modality gap collaboratively and are integrated into adapters to effectively extract cross-modal features. To validate the effectiveness of our approach, we collect a challenging dataset, MMRPM, which includes diverse light conditions and subject head motions. Comprehensive experiments conducted on the MMRPM and EquiPleth [41] datasets demonstrate the superiority of our method." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.836, + 0.483, + 0.865 + ], + "angle": 0, + "content": "The key contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to highlight the dilemma of data hungry and data scarcity in" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "multimodal RPM. We propose leveraging unimodal pretraining and multimodal adapter to tackle this problem." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- To handle the significant modality gap between video and radar, we design the Spatial Alignment and Temporal Matching modules. These modules are further integrated into a novel SATM adapter, enabling effective multimodal fine-tuning while mitigating overfitting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- Extensive experiments across various challenging scenarios show the superiority of the proposed method compared to existing approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.258, + 0.664, + 0.274 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.903, + 0.3 + ], + "angle": 0, + "content": "2.1. Unimodal Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.548 + ], + "angle": 0, + "content": "Most existing remote physiological measurement methods focus on video and radar sensors. Given that the optical absorption of hemoglobin molecules fluctuates along with the cardiac cycle, physiological signals can be remotely measured by detecting changes in skin color in RGB videos. Early works [7, 8, 36, 45] mainly utilize blind source separation and color space transformations to enhance the signal-to-noise ratio (SNR) of extracted rPPG signals. However, these methods are based on restrictive assumptions and struggle to handle complex scenarios. With the development of deep learning, researchers have employed 3D-CNN or modified vision transformer (ViT) to extract rPPG signals from raw videos [49, 51] or specifically designed STMaps [27, 32]. In addition to RGB videos, several studies have also explored rPPG measurement using near-infrared and thermal-infrared videos [5, 15, 34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.746 + ], + "angle": 0, + "content": "On the other hand, radar sensors can extract vital signals by detecting subtle movements of the human chest induced by physiological activities. [11] combines 4D beamforming and CNN to capture seismocardiography from radar sensors. [3] leverages the connection between cardiac mechanical activity and electrical activity to introduce the first radar-based electrocardiography (ECG) monitoring framework. [43] proposes a spatio-temporal network for radar-based Heart Rate Variability (HRV) detection. However, due to the minuscule scale of cardiac motions, radar-based physiological measurement often requires subjects to remain in a supine position and keep still, limiting its flexibility for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.757, + 0.805, + 0.772 + ], + "angle": 0, + "content": "2.2. Video-Radar RPM Measurement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Given the complementary properties of vision and radar modalities, vision-radar fusion has been widely explored in tasks such as object detection [17, 23, 26, 30] and depth estimation [21, 25, 37]. In contrast, video-radar RPM measurement remains an understudied area, with only three related works reported recently [6, 41, 46]. EquiPleth [41] focuses on the skin tone bias problem of video-based methods. They design a late fusion model to ensemble predic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "tions from video and radar backbones and propose an adversarial training strategy to enhance the fairness of the final predictions. Fusion-Vital [6] first projects video and radar inputs into a shared time-difference domain and then employs attention mechanism for cross-modal fusion, showcasing comparable performance even in modality corruption scenarios. Similarly, CardiacMamba [46] adopts a weight-sharing VisionMamba [20] to develop an end-to-end video-radar RPM model. However, these works train their models from scratch on small-scale video-radar datasets, making them prone to overfitting the training sets. To overcome this limitation, we propose leveraging the relatively large-scale unimodal datasets for pre-training and design the SATM adapter to facilitate effective multimodal finetuning while preserving pre-trained knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.292, + 0.344 + ], + "angle": 0, + "content": "2.3. Adapter Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.485, + 0.712 + ], + "angle": 0, + "content": "The increasing scale of foundation models has brought about the challenge of prohibitive cost and substantial data requirements to fully fine-tune these models. Therefore, parameter-efficient fine-tuning (PEFT) has attracted growing attention [12, 13, 16, 52]. Adapter [12], comprising a pair of projection matrices and a nonlinear activation function, is one of the most representative PEFT methods. Although originally developed for natural language processing (NLP), recent works have extended its application to multimodal settings [14, 28, 44, 48]. [28, 48] leverage parameter-sharing adapters to fine-tune vision-language models and promote modality alignment, which is particularly crucial in vision-text tasks. [2] proposes a bi-directional adapter to foster modality complementarity in multimodal tracking. Considering cross-attention is an effective, yet computationally inefficient mechanism for cross-modal fusion, [22] utilizes latent tokens to efficiently integrate cross-attention into adapters. However, these existing approaches are not well-suited for video-radar RPM. This is primarily because naive modality alignment would hinder the learning of modality-specific features, and the substantial modality gap between video and radar further impedes direct fusion of their features. To this end, we present the SATM adapter, specifically tailored for video-radar RPM fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.182, + 0.74 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we focus on constructing a robust multimodal RPM model on the video-radar RPM dataset. To this end, we first leverage existing large-scale unimodal datasets to pre-train our backbones. Afterwards, we freeze the backbones to preserve pre-trained knowledge and employ both vanilla and SATM adapters to fine-tune them into a video-radar RPM model. The proposed SATM adapter consists of two core components, Spatial Alignment and Temporal Matching, which can effectively narrow the modality gap and extract cross-modal features. The general framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.136 + ], + "angle": 0, + "content": "for video-radar fine-tuning is depicted in Fig. 2. We will present our method in more detail in the following parts of this section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.144, + 0.685, + 0.16 + ], + "angle": 0, + "content": "3.1. Backbone Design" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.166, + 0.673, + 0.18 + ], + "angle": 0, + "content": "3.1.1. Video Backbone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Following [19], we utilize the combination of DiffNorm and UniFormer [18] as our video backbone. DiffNorm is a lightweight module that utilizes difference operation and normalization to efficiently integrate the appearance and dynamic features of input videos. UniFormer is an enhanced version of the Vision Transformer (ViT) designed for video analysis. Each UniFormer block consists of Dynamic Position Embedding (DPE), Multi-Head Relation Aggregator (MHRA), and Feed Forward Network (FFN). MHRA is the core component that unifies convolution and attention in a concise transformer format. It employs convolution to tackle feature redundancy in shallow layers while utilizing self-attention for global modeling in deep layers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Formally, given an input video \\(X_{v} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}\\), the DiffNorm module first transforms it into \\(F_{v}^{(0)} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}\\) as the input of UniFormer. Then the forward process of a UniFormer block can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.454, + 0.796, + 0.473 + ], + "angle": 0, + "content": "\\[\nY _ {v} ^ {(l)} = \\mathrm {D P E} \\left(F _ {v} ^ {(l - 1)}\\right) + F _ {v} ^ {(l - 1)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.476, + 0.905, + 0.495 + ], + "angle": 0, + "content": "\\[\nZ _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(\\operatorname {N o r m} \\left(Y _ {v} ^ {(l)}\\right)\\right) + Y _ {v} ^ {(l)} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.497, + 0.815, + 0.516 + ], + "angle": 0, + "content": "\\[\nF _ {v} ^ {(l)} = \\mathrm {F F N} (\\mathrm {N o r m} (Z _ {v} ^ {(l)})) + Z _ {v} ^ {(l)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where \\( F_{v}^{(l - 1)} \\) and \\( F_{v}^{(l)} \\) are the output of the \\( l - 1 \\)-th and the \\( l \\)-th UniFormer block, respectively. We will omit the normalization layers in the rest of the sections for brevity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.81, + 0.588 + ], + "angle": 0, + "content": "3.1.2. Radar Preprocessing and Backbone" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.729 + ], + "angle": 0, + "content": "The fundamental principle of radar-based physiological measurement involves capturing the subtle chest movement caused by cardiac activity. To extract displacement information from the radar intermediate frequency (IF) signal, we first perform range-FFT and coarse range gating to generate a range-time matrix and filter out irrelevant reflections. Afterwards, given that the 2D antenna array enables signal sampling in the spatial domain, we leverage beamforming to separate the reflections from different spatial positions:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.733, + 0.817, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {B} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\boldsymbol {W} ^ {H} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {M}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.755, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\left[ \\dots , \\exp \\left(- j \\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {d} _ {r} ^ {(i)}\\right), \\dots \\right] ^ {T} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.776, + 0.876, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\frac {2 \\pi}{\\lambda} \\left[ \\cos \\theta_ {E} \\cos \\theta_ {A}, \\cos \\theta_ {E} \\sin \\theta_ {A}, \\sin \\theta_ {E} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.809, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(M \\in \\mathbb{R}^{N_C \\times R \\times T_2}\\) is the range-time matrix, with \\(N_C, R, T_2\\) indicating the number of virtual channels, range bins and frames, respectively. \\(\\pmb{W}(\\theta_A, \\theta_E) \\in \\mathbb{R}^{N_C}\\) is the steering vector in the direction of azimuth angle \\(\\theta_A\\) and elevation angle \\(\\theta_E\\), and \\(\\pmb{d}_r^{(i)}\\) is the position of the i-th virtual channel. The shape of the beamformed output \\(\\pmb{B}\\) is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8625" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.087, + 0.868, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.358, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our model: Both backbones are based on UniFormer to facilitate cross-modal fusion. Vanilla adapters and SATM adapters are inserted into the shallow and deep layers of UniFormers, respectively, for multimodal fine-tuning. Each SATM adapter comprises two key components: Spatial Alignment and Temporal Matching. The Spatial Alignment module serves to align the spatial distributions of video and radar features, while the Temporal Matching module further refines the aligned features along the temporal dimension. Finally, the refined cross-modal features are passed through a vanilla adapter and added to the corresponding backbone as complementary features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.528 + ], + "angle": 0, + "content": "\\((\\Theta_A \\times \\Theta_E) \\times R \\times T_2\\), where \\((\\Theta_A \\times \\Theta_E)\\) is the number of angular bins. Following [3, 43], we further utilize a second-order differentiator to amplify cardiac movements while suppressing noise signals as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.538, + 0.481, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {r, t} = \\left(\\psi_ {t - 3} + \\psi_ {t + 3}\\right) + 2 \\left(\\psi_ {t - 2} + \\psi_ {t + 2}\\right) \\tag {3} \\\\ - \\left(\\psi_ {t - 1} + \\psi_ {t + 1}\\right) - 4 \\psi_ {t} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.645 + ], + "angle": 0, + "content": "where \\(\\psi_t\\) is the phase of \\(\\pmb{B}\\) at the t-th frame. The output \\(F_r^{(0)} \\in \\mathbb{R}^{(\\Theta_A \\times \\Theta_E) \\times R \\times T_2}\\) is reshaped into \\(\\mathbb{R}^{1 \\times S \\times T_2}\\) to serve as the input of the radar backbone, where \\(S = \\Theta_A \\times \\Theta_E \\times R\\) is the spatial dimension of the radar modality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Considering the necessity to model the mechanical motion of the human chest across both spatial and temporal dimensions [43], we adopt 2D UniFormer as our radar backbone. Overall, our video and radar backbones are symmetrically structured, which greatly facilitates cross-modal fusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.248, + 0.759 + ], + "angle": 0, + "content": "3.2. SATM Adapter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To enable effectively fine-tuning of the unimodal backbones and facilitate cross-modal modeling, we propose the SATM adapter. As illustrated in Fig. 2, Radar-to-Video SATM adapters and Video-to-Radar SATM adapters are inserted in parallel with the MHRA and FFN modules of the video and radar UniFormers respectively to extract complementary features from the other modality. As discussed in Sec. 1, the fundamental differences in measurement principles between video- and radar-based methods introduce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.604 + ], + "angle": 0, + "content": "a significant modality gap between their features and implicate cross-modal fusion. Considering the different ROIs of video and radar (face vs. chest), we present the Spatial Alignment (SA) module to align the spatial distributions of their features. As for the differences in captured signals (color changes vs. mechanical displacements), we design the Temporal Matching (TM) module to reduce waveform discrepancies and refine the spatially aligned features. Next, we will provide a detailed explanation of these modules." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.688, + 0.626 + ], + "angle": 0, + "content": "3.2.1. Spatial Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.735 + ], + "angle": 0, + "content": "The physiological information contained in the face and chest is inherently distributed in distinct forms, resulting in different spatial distributions of video and radar features. We would like to emphasize that this dissimilarity is not merely reflected in the different spatial dimensions of \\( F_{v} \\) and \\( F_{r} \\) (\\( HW \\) vs. \\( S \\)), and naively aligning their size through interpolation or linear projection is insufficient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, we design the Spatial Alignment module. Given the video feature \\( F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w} \\) and radar feature \\( F_{r} \\in \\mathbb{R}^{C \\times s \\times t_{2}} \\), the core idea involves utilizing cross-attention to align the spatial distributions of their spatial tokens, i.e. \\( f_{v} \\in \\mathbb{R}^{C \\times t_{1}} \\) and \\( f_{r} \\in \\mathbb{R}^{C \\times t_{2}} \\). Without loss of generality, we use the Radar-to-Video Spatial Alignment as an example for explanation, with its overall pipeline illustrated in Fig. 3 (a). First, bilinear interpolation is applied to align the temporal dimensions of \\( F_{v} \\) and \\( F_{r} \\). The primary challenge then lies in effectively transforming the 2D spatial tokens for the calculation of the attention" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.343 + ], + "angle": 0, + "content": "Figure 3. The pipelines of Radar-to-Video Spatial Alignment and Temporal Matching. \"MatMul\" and \"Hadamard\" denote matrix product and Hadamard product, respectively. \\(\\mathrm{FFT}_T\\) represents FFT along the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.485, + 0.55 + ], + "angle": 0, + "content": "scores. Considering that the final predicted 1D PPG signals are generated by the integration of the spatial and channel dimensions, we propose treating spatial tokens as distinct candidates for PPG signal prediction at different spatial locations and then projecting them from the 2D space into a unified 1D time series space by compressing their channel dimensions. Specifically, we leverage linear projections to transform \\( F_{v} \\) and \\( F_{r} \\) into the query matrix \\( S_{v} \\in \\mathbb{R}^{hw\\times t_{1}} \\) and the key matrix \\( S_{r} \\in \\mathbb{R}^{s\\times t_{1}} \\), respectively. Afterwards, the alignment between video and radar spatial tokens can be performed in the time series space and the resulting aligned radar features are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.56, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {S _ {v} S _ {r} ^ {T}}{\\sqrt {t _ {1}}}\\right) F _ {r} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Note that the matrix multiplication between the attention score and \\(F_{r}\\) is performed channel-wise." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.633, + 0.483, + 0.83 + ], + "angle": 0, + "content": "Now let us revisit the way to transform the spatial tokens. Two alternative approaches involve compressing the temporal dimension or flattening the channel and temporal dimensions of spatial tokens. However, the former variant risks irreversible information loss due to excessive compression of temporal signals, which is paramount for the prediction of PPG signals. The second approach, while straightforward, fails to account for the inherent differences between the channel and temporal dimensions. In contrast, our spatial alignment, which operates in the time series space, adheres to the paradigm of PPG signal regression and has the advantage of preserving temporal information. A quantitative comparison is provided in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.277, + 0.852 + ], + "angle": 0, + "content": "3.2.2. Temporal Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although video and radar features are temporally synchronized and ideally encompass consistent frequency information, i.e. heart rate, the inherent disparity between wave" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "forms derived from skin color changes and chest mechanical movements introduces a significant modality gap in the temporal dimension. To this end, we further utilize Temporal Matching to refine the cross-modal features. According to the Matched Filter theory, given an input signal \\( r(t) \\) composed of the target signal \\( s(t) \\) and the noise \\( n(t) \\), an output signal with the highest SNR at time step \\( T \\) could be obtained by applying the filter \\( h(t) = s(T - t) \\). Inspired by this, we formulate the spatially aligned features as \\( r(t) \\) and the waveform discrepancy as \\( n(t) \\), and the refined feature could be obtained through matched filtering. Specifically, given the video feature \\( F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w} \\) and the spatially aligned radar feature \\( F_{r}^{\\prime} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w} \\), the refined radar-to-video feature can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T M} _ {r 2 v} \\left(F _ {v}, F _ {r} ^ {\\prime}\\right) = \\mathrm {D W c o n v} \\left(\\mathrm {B N} \\left(\\mathcal {F} _ {T} ^ {- 1} \\left(\\overline {{\\mathcal {F} _ {T} \\left(F _ {v}\\right)}} \\mathcal {F} _ {T} \\left(F _ {r} ^ {\\prime}\\right)\\right)\\right) \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.402 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_T(\\cdot)\\) denotes Fast Fourier Transform (FFT) along the temporal dimension and \\(\\overline{\\mathcal{F}_T(\\cdot)}\\) represents FFT followed by the conjugate operation. DWConv and BN are depthwise convolution and batch normalization, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.702, + 0.421 + ], + "angle": 0, + "content": "3.2.3. Cross-Modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.905, + 0.457 + ], + "angle": 0, + "content": "Overall, the forward process of the Radar-to-Video SATM adapter can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.463, + 0.847, + 0.48 + ], + "angle": 0, + "content": "\\[\nF _ {r 2 v} = \\mathrm {T M} _ {r 2 v} \\left(F _ {v}, \\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.482, + 0.912, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A T M} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = W _ {u p} \\left(\\operatorname {R e L U} \\left(W _ {\\text {d o w n}} \\left(F _ {v} + \\alpha \\cdot F _ {r 2 v}\\right)\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.572 + ], + "angle": 0, + "content": "where \\( W_{down}, W_{up} \\) are the down-projection and up-projection layers of the vanilla adapter [12] and \\( \\alpha \\) is a learnable scaling factor. Afterward, the forward process of a video UniFormer block can be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.575, + 0.888, + 0.595 + ], + "angle": 0, + "content": "\\[\nZ _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(Y _ {v} ^ {(l)}\\right) + Y _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Y _ {v} ^ {(l)}, Y _ {r} ^ {(l)}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.597, + 0.905, + 0.629 + ], + "angle": 0, + "content": "\\[\nF _ {v} ^ {(l)} = \\operatorname {F F N} \\left(Z _ {v} ^ {(l)}\\right) + Z _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Z _ {v} ^ {(l)}, Z _ {r} ^ {(l)}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.674 + ], + "angle": 0, + "content": "The forward processes of the Video-to-Radar SATM adapter and the fine-tuned radar UniFormer blocks can also be computed similarly to Eq. (6) and Eq. (7)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In summary, Spatial Alignment aligns the spatial distributions of video and radar features in a unified time series space, while Temporal Matching utilizes a matched filter to handle the waveform difference between their signals. They complement each other by processing the cross-modal features along spatial and temporal dimensions separately, thereby effectively narrowing the substantial modality gap between video and radar features. According to [9], it is essential to balance the learning of unimodal and cross-modal features. Consequently, we only insert our SATM adapters into the last two stages of UniFormers while utilizing vanilla adapters in the first two stages. The output features of both backbones are integrated by a lightweight fusion stem and subsequently passed through a regression module to generate the predicted PPG signal." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8627" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.223, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.188, + 0.13 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.212 + ], + "angle": 0, + "content": "We utilize the EquiPleth [41] dataset to evaluate our method, which contains 533 30-second samples collected from 28 light, 49 medium, and 14 dark skin tone volunteers. All samples are recorded under motionless conditions and with sufficient illumination." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.213, + 0.483, + 0.319 + ], + "angle": 0, + "content": "In addition, considering that the EquiPleth dataset mainly focuses on constrained scenarios and shows limited practical diversity, we built a new dataset called MMRPM to promote a more comprehensive evaluation of our method. MMRPM comprises 180 60-second samples from 31 subjects, all of which are recorded under varying illumination conditions and subject motions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.471 + ], + "angle": 0, + "content": "Specifically, MMRPM encompasses three illumination conditions: dark, medium, and light. The \"dark\" and \"medium\" settings correspond to illumination levels of approximately 2.5 and 6.3 lux respectively, while \"light\" represents samples are collected under sufficient sunlight and indoor illumination. Furthermore, two types of subject activities are considered: sitting still and free head movement. Overall, MMRPM covers six scenarios, where low-light conditions are challenging to the video modality and head motions have a greater impact on the radar modality.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.48, + 0.308, + 0.496 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.623 + ], + "angle": 0, + "content": "Our proposed method is implemented using PyTorch. For video preprocessing, we first employ MTCNN to crop the face region in the initial frame and fix the region across subsequent frames. Then, we sample each video into clips with a time window of 160 frames and a step size of 80 frames, which are then resized to \\(96 \\times 96\\) pixels. In terms of radar preprocessing, the angular range selected for beamforming is \\(66^{\\circ} \\leq \\Theta_{A} \\leq 114^{\\circ}, -24^{\\circ} \\leq \\Theta_{E} \\leq 24^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.715 + ], + "angle": 0, + "content": "We utilize UniFormer-S [18] and MAM in MAGIC [55] as our feature extractor and fusion stem respectively. As for the regression module, we follow the design in ADDP [19]. The video backbone is jointly pre-trained on VIPL [31, 32], MMPD [40], UBFC [1], PURE [39] and BUAA [47], and we use the dataset introduced in [53] for radar pre-trianing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.866 + ], + "angle": 0, + "content": "During multimodal fine-tuning, both backbones are frozen, and only the adapters, fusion stem and regression module are trained. We employ the loss functions introduced in [51] and utilize the Adam optimizer for training, with the learning rate and weight decay set to 5e-4 and 5e-5 respectively. The model is trained for 40 epochs with a batch size of 2. Following [38], we apply random horizontal flipping, spatially resized crop and random intensity noise to augment the input videos. For radar augmentation, we randomly shift the range bins of the range-time matrices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Following [41], we perform our evaluation of the PPG signals over 300 samples windows with a stride of 128 samples. The heart rate corresponding to a specific signal clip is calculated using a combination of band-pass filtering, detrending and peak detection in the frequency domain. Finally, the standard deviation of the error (Std), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (R) are adopted to evaluate the predicted HR error. The unit of Std, MAE and RMSE is beats per minute (bpm). We split the MMRPM dataset into 5 subject-exclusive folds and use the official split for the EquiPleth dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.286, + 0.653, + 0.301 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.687 + ], + "angle": 0, + "content": "We first compare our SATM with existing video-radar RPM methods [6, 41, 46] and several state-of-the-art multimodal adapter methods [2, 22, 28, 48] on the EquiPleth [41] dataset. To ensure a fair comparison, we provide extra results of improved versions of existing RPM methods by applying unimodal pre-training and replacing their original backbones with UniFormers. Since the radar used in the EquiPleth dataset activates only 1 transmitter and 1 receiver, we do not perform beamforming during radar preprocessing for this dataset. For existing video-radar RPM methods, we use the official implementation of EquiPleth, and re-implement Fusion-Vital and CardiacMamba, as the codes have not been released by the authors. As shown in Tab. 1, all of these methods exhibit RMSEs exceeding 2.9 bpm when trained from scratch. Their performance consistently improves when unimodal pre-training is applied, underscoring the significance of unimodal pre-training when video-radar data is limited. However, they still display MAEs greater than 0.64 bpm, regardless of whether using their original backbones or UniFormers, indicating that their full fine-tuning-based fusion strategies are suboptimal for the small-scale multimodal RPM dataset. In contrast, our method employs SATM adapters for multimodal fine-tuning, effectively leveraging pre-trained knowledge and enabling superior cross-modal fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Furthermore, our SATM also demonstrates superior performance compared to existing multimodal adapter methods. The subpar performance of UniAdapter and MMA suggests that their parameter-sharing strategies are ill-suited for multimodal RPM, as they impede the learning of modality-specific features. To adapt BAT for video-radar RPM, we employ an additional linear projection layer to align the token numbers of the video and radar branches. However, it fails to handle the significant modality gap between video and radar, thereby resulting in an RMSE exceeding 1.65 bpm. While the best baseline, LAVISH, achieves an RMSE of 1.498 bpm, our SATM yields a considerably lower RMSE of 1.242 bpm." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We have also evaluated SATM on our self-collected" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For more details about our MMRPM dataset and pre-training, please refer to the supplemental materials." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8628" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.089, + 0.865, + 0.343 + ], + "angle": 0, + "content": "
MethodsOriginal ModalityUnimodal Pre-trainedStd↓MAE↓RMSE↓R↑Trainable Params (M)↓
EquiPleth[41]Video-Radar×3.2741.0633.2830.9573.95
Fusion-Vital[6]Video-Radar×9.2883.7509.2970.709715.26
CardiacMamba[46]Video-Radar×2.9240.9012.9270.9759.93
EquiPlethVideo-Radar√2.9790.9682.9900.9653.95
Fusion-VitalVideo-Radar√9.1213.7509.1300.735715.26
CardiacMambaVideo-Radar√2.7970.8682.7710.9759.93
EquiPleth*Video-Radar√1.4350.6711.4360.99245.31
Fusion-Vital*Video-Radar√2.1430.6502.1570.983194.61
CardiacMamba*Video-Radar√2.3130.7442.3240.98033.62
UniAdapter[28]Vision-Language√1.6530.5901.6560.99010.60
MMA[48]Vision-Language√1.8350.6261.8420.98811.09
BAT[2]RGB-TIR√1.6790.6301.6840.99085.35
LAVISH[22]Audio-Visual√1.4930.5571.4980.99110.85
SATM (Ours)Video-Radar√1.2170.5401.2420.99410.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.353, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Table 1. Overall performance on the EquiPleth dataset. The best and second best results are in bold and underlined respectively. * denotes that we replace the methods' original backbones with UniFormers while preserving their fusion strategies. Blue and red indicate the existing video-radar RPM methods and existing adapter methods for other multimodal tasks, respectively." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.419, + 0.48, + 0.613 + ], + "angle": 0, + "content": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth7.6032.4217.6030.715
Fusion-Vital9.1395.1659.3150.501
CardiacMamba6.0992.1606.1070.790
EquiPleth*5.9641.7395.9740.806
Fusion-Vital*5.5501.6625.5890.827
CardiacMamba*5.5261.8435.5640.829
UniAdapter5.5091.5885.5100.831
MMA5.6941.7225.6940.819
BAT6.4322.1516.4330.760
LAVISH5.6601.4545.3120.844
SATM (Ours)5.0631.2925.0740.858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.623, + 0.45, + 0.637 + ], + "angle": 0, + "content": "Table 2. Overall performance on the MMRPM dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.762 + ], + "angle": 0, + "content": "MMRPM dataset, which covers more challenging and diverse scenarios than the EquiPleth dataset. As illustrated in Tab. 2, existing video-radar RPM methods exhibit poor results on this dataset, highlighting their limited capability to handle complex scenarios. In comparison, our SATM consistently outperforms the baseline methods in all metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We further conduct a cross-dataset evaluation by training the models on the EquiPleth dataset and evaluating them on the MMRPM dataset. As shown in Tab. 3, the improved versions of existing multimodal RPM methods exhibit subpar performance compared to adapter-based methods. This is primarily because they rely on full fine-tuning-based fusion strategies and are prone to overfitting small-scale training sets. In contrast, our method, built upon SATM adapters, achieves the best generalization performance and suffers" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.419, + 0.902, + 0.523 + ], + "angle": 0, + "content": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth*8.7874.2569.0820.629
Fusion-Vital*8.5384.4578.8970.642
CardiacMamba*9.0485.0389.3080.603
LAVISH8.0263.6878.2320.712
SATM (Ours)7.5423.4237.6560.738
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Table 3. Results of the cross-dataset evaluation from the EquiPleth dataset to the MMRPM dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.907, + 0.662 + ], + "angle": 0, + "content": "minimal performance degradation compared to other baselines. Although it still lags behind the performance in the intra-dataset evaluation, we argue that this is reasonable, given the significant differences between these two datasets.[2]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.673, + 0.669, + 0.69 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Key Components. Our SATM adapter mainly consists of two key components: Spatial Alignment and Temporal Matching. In Tab. 4, we conduct an ablation study on the MMRPM dataset to assess the effectiveness of these modules. First, compared to the vanilla adapter (the first row), the adapter with Spatial Alignment reduces the MAE and RMSE by 0.087 bpm and 0.149 bpm respectively. This can be attributed to its capability to facilitate modality complementarity by extracting spatially aligned cross-modal features to the corresponding backbones. The MAE and RMSE can be further reduced to 1.292 bpm and 5.074 bpm when Temporal Matching is inserted, highlighting the necessity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.881, + 0.902 + ], + "angle": 0, + "content": "2Details about cross-dataset evaluation are provided the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8629" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.455, + 0.16 + ], + "angle": 0, + "content": "
SATMStd↓MAE↓RMSE↓R↑
××5.4521.5415.4530.836
✓×5.2981.4545.3040.845
✓✓5.0631.2925.0740.858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.17, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Table 4. Ablation study on the key components. \"SA\" and \"TM\" denote Spatial Alignment and Temporal Matching respectively." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.214, + 0.46, + 0.317 + ], + "angle": 0, + "content": "
VariantsStd↓MAE↓RMSE↓R↑
Interpolate6.0411.8766.0520.795
Linear5.5791.6285.5820.828
Channel6.0871.8076.0910.792
Flatten5.5881.6935.5950826
Ours5.2981.4545.3040.845
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.327, + 0.465, + 0.341 + ], + "angle": 0, + "content": "Table 5. Ablation study on the design of Spatial Alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.385 + ], + "angle": 0, + "content": "to refine the aligned features along the temporal dimension." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.583 + ], + "angle": 0, + "content": "Alternative Approaches for Spatial Alignment. We also investigate four alternative approaches for Spatial Alignment. Specifically, \"Interpolate\" and \"Linear\" refer to methods that directly align the sizes of video and radar features through interpolation and linear projection instead of cross-attention. The \"Channel\" and \"Flatten\" indicate that the alignment is performed in the channel space and on the flattened spatial tokens respectively. As shown in Tab. 5, both \"Interpolate\" and \"Linear\" exhibit suboptimal performance, achieving MAEs of 1.876 bpm and 1.628 bpm. These methods only align spatial sizes while neglecting inherent differences in spatial distributions between video and radar features, thereby failing to address the modality gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.78 + ], + "angle": 0, + "content": "The \"Channel\" variant compresses the temporal dimensions of spatial tokens for the computation of the attention scores, focusing solely on the overall temporal period while neglecting fine-grained temporal variability. This results in degraded performance with an RMSE exceeding 6 bpm. The \"Flatten\" variant preserves temporal information, but conflates temporal and channel dimensions, overlooking their intrinsic differences. Therefore, it ultimately yields subpar performance with an MAE of 1.693 bpm and an RMSE of 5.595 bpm. In contrast, Spatial Alignment performed in the time series space effectively avoids temporal information loss and dimensional distortion, showing superior performance compared to these four variants." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Temporal Matching vs. Temporal Attention. Since attention is a straightforward method for cross-modal fusion, we compare our Temporal Matching (TM) with Temporal Attention (TA). Specifically, we replace the matched filtering operation with a cross-attention along the temporal dimension. As shown in Tab. 6, our TM consistently outperforms TA across all metrics. Moreover, since the time complexity of FFT is \\(\\mathcal{O}(N\\log N)\\), TM is also more com" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.88, + 0.146 + ], + "angle": 0, + "content": "
TemporalStd↓MAE↓RMSE↓R↑
Attention5.1551.3675.1300.853
Matching5.0631.2925.0740.858
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.156, + 0.905, + 0.183 + ], + "angle": 0, + "content": "Table 6. Comparison between Temporal Matching and Temporal Attention." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.197, + 0.878, + 0.284 + ], + "angle": 0, + "content": "
PositionsStd↓MAE↓RMSE↓R↑
SATM-45.1681.3865.1790.852
SATM-35.1071.3435.1090.857
SATM-25.0631.2925.0740.858
SATM-15.5261.6145.5260.830
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.294, + 0.895, + 0.309 + ], + "angle": 0, + "content": "Table 7. Ablation study on the position to add SATM adapters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.363 + ], + "angle": 0, + "content": "putationally efficient compared to the quadratic complexity of TA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Position to Insert Our SATM Adapters. We also explore the impact of different positions to insert our SATM adapters. We present 4 positions for insertion, formulated as \"SATM-n\", where \"n\" denotes that we use SATM adapters in the last \"n\" stages of the UniFormers and employ vanilla adapters in earlier stages. As reported in Tab. 7, SATM-1 displays an MAE higher than the naive baseline (1.614 bpm vs. 1.541 bpm), indicating that inserting the cross-modal module too late would hinder performance due to the lack of cross-modal interaction. Meanwhile, the performance of SATM-3 and SATM-4 is also limited due to insufficient modality-specific modeling. In contrast, the results demonstrate that SATM-2 strikes a good trade-off between the learning of unimodal and corss-modal features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.81 + ], + "angle": 0, + "content": "In this paper, we focus on video-radar RPM measurement. To overcome the data scarcity challenge, we make full use of unimodal pre-training and present the SATM adapter to effectively fine-tune unimodal backbones into a multimodal RPM model. The SATM adapter leverages Spatial Alignment and Temporal Matching to collaboratively narrow the modality gap between video and radar and effectively extract cross-modal features for both backbones. To thoroughly evaluate our method, we collect a challenging dataset that encompasses diverse illumination conditions and subject motions. Extensive experiments conducted on our self-collected and existing public datasets demonstrate that SATM outperforms existing methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitations. While SATM effectively constructs a robust video-radar RPM model, it primarily focuses on cross-modal interaction and still relies on vanilla adapters for unimodal modeling during fine-tuning. In the future, we hope to develop a unified framework that harmonizes cross-modal and unimodal modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8630" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.144 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62172381." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[3] Jinbo Chen, Dongheng Zhang, Zhi Wu, Fang Zhou, Qibin Sun, and Yan Chen. Contactless electrocardiogram monitoring with millimeter wave radar. IEEE Transactions on Mobile Computing, 23(1):270-285, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[4] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 349-365, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[5] Li-Wen Chiu, Yang-Ren Chou, Yi-Chiao Wu, and Bing-Fei Wu. Deep-learning based remote photoplethysmography measurement in driving scenarios with color and near-infrared images. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.485, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[6] Jae-Ho Choi, Ki-Bong Kang, and Kyung-Tae Kim. Fusion-vital: Video-rf fusion transformer for advanced remote physiological measurement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1344-1352, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.484, + 0.597 + ], + "angle": 0, + "content": "[7] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE transactions on biomedical engineering, 60(10):2878-2886, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.638 + ], + "angle": 0, + "content": "[8] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiological measurement, 35(9):1913, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.707 + ], + "angle": 0, + "content": "[9] Chenzhuang Du, Jiaye Teng, Tingle Li, Yichen Liu, Tianyuan Yuan, Yue Wang, Yang Yuan, and Hang Zhao. On uni-modal feature learning in supervised multi-modal learning. In International Conference on Machine Learning, pages 8632-8656. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.707, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[10] Jingda Du, Si-Qi Liu, Bochao Zhang, and Pong C Yuen. Dual-bridging with adversarial noise generation for domain adaptive rpgg estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10355–10364, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.832 + ], + "angle": 0, + "content": "[11] Unsoo Ha, Salah Assana, and Fadel Adib. Contactless seismocardiography via deep learning radars. In Proceedings of the 26th annual international conference on mobile computing and networking, pages 1-14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.194, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.216 + ], + "angle": 0, + "content": "[14] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, XueQing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[15] Jitesh Joshi and Youngjun Cho. ibvp dataset: Rgb-thermal rppg dataset with high resolution signal quality labels. Electronics, 13(7):1334, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[16] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[17] Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17615-17626, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[18] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[19] Qian Liang, Yan Chen, and Yang Hu. Continual learning for remote physiological measurement: Minimize forgetting and simplify inference. In European conference on computer vision, pages 126-144. Springer, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[20] Zhu Lianghui, Liao Bencheng, Zhang Qian, Wang Xinlong, Liu Wenyu, and Wang Xinggang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv. Org, pages arXiv-org, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[21] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[22] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2299–2309, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.776 + ], + "angle": 0, + "content": "[23] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.832 + ], + "angle": 0, + "content": "[24] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3343-3347. IEEE, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8631" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Yunfei Long, Abhinav Kumar, Daniel Morris, Xiaoming Liu, Marcos Castro, and Punarjay Chakravarty. Radiant: Radar-image association network for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1808-1816, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[27] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[28] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[29] Hao Lu, Zitong Yu, Xuesong Niu, and Ying-Cong Chen. Neuron structure modeling for generalizable remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18589-18599, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[30] Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14, pages 562–576. Springer, 2019. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[32] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Xuesong Niu, Xingyuan Zhao, Hu Han, Abhijit Das, Antitza Dantcheva, Shiguang Shan, and Xin Chen. Robust remote heart rate estimation from face utilizing spatial-temporal attention. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-8. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Soyeon Park, Bo-Kyeong Kim, and Suh-Yeon Dong. Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation. IEEE Transactions on Instrumentation and Measurement, 71:1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[35] Vladimir L Petrović, Milica M Janković, Anita V Lupšić, Veljko R Mihajlović, and Jelena S Popovic-Božović. High-accuracy real-time monitoring of heart rate variability using 24 ghz continuous-wave doppler radar. *IEEE Access*, 7: 74721–74733, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[36] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10):10762-10774, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[38] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[40] Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, and Xin Liu. Mmpd: multidomain mobile video physiology dataset. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5. IEEE, 2023. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.396 + ], + "angle": 0, + "content": "[41] Alexander Vilesov, Pradyumna Chari, Adnan Armouti, Anirudh Bindiganavale Harish, Kimaya Kulkarni, Ananya Deoghare, Laleh Jalilian, and Achuta Kadambi. Blending camera and 77 ghz radar sensing for equitable, robust plethysmography. ACM Trans. Graph., 41(4):36-1, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[42] Fengyu Wang, Xiaolu Zeng, Chenshu Wu, Beibei Wang, and KJ Ray Liu. mmhrv: Contactless heart rate variability monitoring using millimeter-wave radio. IEEE Internet of Things Journal, 8(22):16623-16636, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[43] Haoyu Wang, Jinbo Chen, Dongheng Zhang, Zhi Lu, Changwei Wu, Yang Hu, Qibin Sun, and Yan Chen. Contactless radar heart rate variability monitoring via deep spatiotemporal modeling. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 111-115. IEEE, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[44] Kai Wang, Yapeng Tian, and Dimitrios Hatzinakos. Towards efficient audio-visual learners via empowering pre-trained vision transformers with cross-modal adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1837–1846, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[45] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE transactions on Biomedical Engineering, 62(2): 415-425, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[46] Zheng Wu, Yiping Xie, Bo Zhao, Jiguang He, Fei Luo, Ning Deng, and Zitong Yu. Cardiacmamba: A multimodal rgb-rf fusion framework with state space models for remote physiological measurement. arXiv preprint arXiv:2502.13624, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[47] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from fa" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8632" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "cial videos using spatio-temporal networks. arXiv preprint arXiv:1905.02419, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[50] Zitong Yu, Xiaobai Li, Xuesong Niu, Jingang Shi, and Guoying Zhao. Autohr: A strong end-to-end baseline for remote heart rate measurement with neural searching. IEEE Signal Processing Letters, 27:1245-1249, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[51] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip HS Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4186-4196, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[52] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[53] Bin-Bin Zhang, Dongheng Zhang, Yadong Li, Zhi Lu, Jinbo Chen, Haoyu Wang, Fang Zhou, Yu Pu, Yang Hu, Li-Kun Ma, et al. Monitoring long-term cardiac activity with contactless radio frequency signals. Nature Communications, 15(1):1-11, 2024. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[54] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[55] Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. Centering the value of every modality: Towards efficient and resilient modality-agnostic semantic segmentation. In European Conference on Computer Vision, pages 192–212. Springer, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8633" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_origin.pdf b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..723f1b5bd1551196be1f389cbb4440960a1699d0 --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/2a28716b-580e-47e9-adc9-9d410d63b302_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98d1b4672d4970e204bb62f93f493c667854b1d8e7798212cf2591b5e263bdd9 +size 1087210 diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/full.md b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5dbdb10d877c219ef1d14a0ed440e9f4dbf4e9c4 --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/full.md @@ -0,0 +1,328 @@ +# Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement + +Qian Liang $^{1}$ , Ruixu Geng $^{1}$ , Jinbo Chen $^{2}$ , Haoyu Wang $^{1}$ , Yan Chen $^{1}$ , Yang Hu $^{1*}$ $^{1}$ University of Science and Technology of China, $^{2}$ Nanyang Technological University +{qianliang, gengruixu, why1999}@mail.ustc.edu.cn, jinbo.chen@ntu.edu.sg +{eecyan, eeyhu}@ustc.edu.cn + +# Abstract + +Remote physiological measurement (RPM) based on video and radar has made significant progress in recent years. However, unimodal methods based solely on video or radar sensor have notable limitations due to their measurement principles, and multimodal RPM that combines these modalities has emerged as a promising direction. Despite its potential, the lack of large-scale multimodal data and the significant modality gap between video and radar pose substantial challenges in building robust videoradar RPM models. To handle these problems, we suggest leveraging unimodal pre-training and present the Spatial alignment and Temporal Matching (SATM) Adapter to effectively fine-tune pre-trained unimodal backbones into a multimodal RPM model. Given the distinct measurement principles of video- and radar-based methods, we propose Spatial Alignment to align the spatial distribution of their features. Furthermore, Temporal Matching is applied to mitigate waveform discrepancies between video and radar signals. By integrating these two modules into adapters, the unimodal backbones could retain their modality-specific knowledge while effectively extracting complementary features from each other. Extensive experiments across various challenging scenarios, including low light conditions and head motions, demonstrate that our approach significantly surpasses the state-of-the-art methods. + +# 1. Introduction + +Remote physiological measurement has attracted a lot of attention recently owing to its ability to detect physiological signals without contact. A prominent line of research focuses on video-based remote photoplethysmography (rPPG), which uses ordinary cameras to extract vital signals by detecting periodic changes in skin color induced by heartbeats [4, 10, 24, 29, 49, 51]. Beyond video-based + +![](images/84f65275c7434545850e7505102f360569ce9f991d5cbf0f4a54965fb7b916f3.jpg) +Figure 1. Frameworks of video-radar RPM methods: (a) EquiPleth [41] trains unimodal backbones on the multimodal dataset and leverage a late fusion model to ensemble their predictions. (b) Fusion-Vital [6] and CardiacMamba [46] train their end-to-end models from scratch on the multimodal dataset. (c) Our method makes full use of unimodal pre-training and propose the SATM adapter to enable effective multimodal adaptation. + +methods, an alternative approach involves the use of radar sensors to capture cardiac activities through subtle mechanical motions of the human chest [3, 35, 42, 43]. + +However, both video- and radar-based methods have notable limitations. Video-based methods are highly sensitive to light conditions and may be influenced by subjects' skin tones [41]. On the other hand, radar-based methods are much more susceptible to subject movements, and most existing methods are designed for strict-constraint scenarios. Therefore, it is necessary and promising to combine video and radar for RPM measurement to overcome the limitations of unimodal approaches. + +Nevertheless, constructing a robust video-radar RPM model remains a significant challenge. Due to the complexity of recording synchronized videos, radar data, and ground-truth PPG signals, video-radar RPM datasets are mostly limited in scale. Consequently, training a model from scratch on such datasets is prone to severe overfitting. Unfortunately, existing multimodal RPM works [6, 41, 46] overlook this problem, leading to limited robustness and generalization ability (See Sec. 4.3). In addition, video + +based methods focus primarily on color changes of facial skin, whereas radar-based methods capture mechanical displacements of the chest. This implies that these two modalities capture distinct signals from different regions of interest (ROIs), introducing a substantial modality gap and making it difficult to effectively fuse their features. + +For the data scarcity challenge, we suggest making full use of unimodal pre-training, as unimodal data are easier to collect and large-scale unimodal datasets for remote physiological measurement are already available [31, 40, 53]. As illustrated in Fig. 1 (c), we first utilize large-scale unimodal datasets to pre-train our backbones. Subsequently, instead of fully fine-tuning the entire network on limited multimodal data, we propose leveraging cross-modal adapters to fine-tune unimodal backbones into a multimodal model while preserving the pre-trained knowledge. Several previous studies have introduced adapter techniques into multimodal learning [2, 28, 48]. For instance, UniAdapter [28] and MMA [48] design parameter-sharing adapters to facilitate modality alignment in vision-language models, while BAT [2] proposes a bi-directional adapter to directly extract cross-modal features for multimodal tracking. However, multimodal adapter for video-radar RPM has not been explored before. Furthermore, the aforementioned methods are suboptimal for multimodal RPM for two reasons: (1) crude modality alignment would hamper the learning of modality-specific features in this context, and (2) the significant modality gap between video and radar, as discussed earlier, impedes the direct fusion of their features. + +To this end, we propose a novel SATM adapter to enable effective multimodal fine-tuning on the video-radar RPM dataset. Specifically, the Spatial Alignment (SA) and Temporal Matching (TM) modules are devised to manage the large modality gap between video and radar. To mitigate the gap introduced by their distinct ROIs, Spatial Alignment projects latent features from different modalities into a unified time series space and then utilizes cross-attention to align their spatial distributions. Additionally, to handle the gap caused by different waveforms of detected signals from video and radar, Temporal Matching further applies a matched filter to refine the spatially aligned features along the temporal dimension. These two modules narrow the modality gap collaboratively and are integrated into adapters to effectively extract cross-modal features. To validate the effectiveness of our approach, we collect a challenging dataset, MMRPM, which includes diverse light conditions and subject head motions. Comprehensive experiments conducted on the MMRPM and EquiPleth [41] datasets demonstrate the superiority of our method. + +The key contributions of this work are summarized as follows: + +- To the best of our knowledge, we are the first to highlight the dilemma of data hungry and data scarcity in + +multimodal RPM. We propose leveraging unimodal pretraining and multimodal adapter to tackle this problem. + +- To handle the significant modality gap between video and radar, we design the Spatial Alignment and Temporal Matching modules. These modules are further integrated into a novel SATM adapter, enabling effective multimodal fine-tuning while mitigating overfitting. +- Extensive experiments across various challenging scenarios show the superiority of the proposed method compared to existing approaches. + +# 2. Related Works + +# 2.1. Unimodal Remote Physiological Measurement + +Most existing remote physiological measurement methods focus on video and radar sensors. Given that the optical absorption of hemoglobin molecules fluctuates along with the cardiac cycle, physiological signals can be remotely measured by detecting changes in skin color in RGB videos. Early works [7, 8, 36, 45] mainly utilize blind source separation and color space transformations to enhance the signal-to-noise ratio (SNR) of extracted rPPG signals. However, these methods are based on restrictive assumptions and struggle to handle complex scenarios. With the development of deep learning, researchers have employed 3D-CNN or modified vision transformer (ViT) to extract rPPG signals from raw videos [49, 51] or specifically designed STMaps [27, 32]. In addition to RGB videos, several studies have also explored rPPG measurement using near-infrared and thermal-infrared videos [5, 15, 34]. + +On the other hand, radar sensors can extract vital signals by detecting subtle movements of the human chest induced by physiological activities. [11] combines 4D beamforming and CNN to capture seismocardiography from radar sensors. [3] leverages the connection between cardiac mechanical activity and electrical activity to introduce the first radar-based electrocardiography (ECG) monitoring framework. [43] proposes a spatio-temporal network for radar-based Heart Rate Variability (HRV) detection. However, due to the minuscule scale of cardiac motions, radar-based physiological measurement often requires subjects to remain in a supine position and keep still, limiting its flexibility for real-world applications. + +# 2.2. Video-Radar RPM Measurement + +Given the complementary properties of vision and radar modalities, vision-radar fusion has been widely explored in tasks such as object detection [17, 23, 26, 30] and depth estimation [21, 25, 37]. In contrast, video-radar RPM measurement remains an understudied area, with only three related works reported recently [6, 41, 46]. EquiPleth [41] focuses on the skin tone bias problem of video-based methods. They design a late fusion model to ensemble predic + +tions from video and radar backbones and propose an adversarial training strategy to enhance the fairness of the final predictions. Fusion-Vital [6] first projects video and radar inputs into a shared time-difference domain and then employs attention mechanism for cross-modal fusion, showcasing comparable performance even in modality corruption scenarios. Similarly, CardiacMamba [46] adopts a weight-sharing VisionMamba [20] to develop an end-to-end video-radar RPM model. However, these works train their models from scratch on small-scale video-radar datasets, making them prone to overfitting the training sets. To overcome this limitation, we propose leveraging the relatively large-scale unimodal datasets for pre-training and design the SATM adapter to facilitate effective multimodal finetuning while preserving pre-trained knowledge. + +# 2.3. Adapter Fine-Tuning + +The increasing scale of foundation models has brought about the challenge of prohibitive cost and substantial data requirements to fully fine-tune these models. Therefore, parameter-efficient fine-tuning (PEFT) has attracted growing attention [12, 13, 16, 52]. Adapter [12], comprising a pair of projection matrices and a nonlinear activation function, is one of the most representative PEFT methods. Although originally developed for natural language processing (NLP), recent works have extended its application to multimodal settings [14, 28, 44, 48]. [28, 48] leverage parameter-sharing adapters to fine-tune vision-language models and promote modality alignment, which is particularly crucial in vision-text tasks. [2] proposes a bi-directional adapter to foster modality complementarity in multimodal tracking. Considering cross-attention is an effective, yet computationally inefficient mechanism for cross-modal fusion, [22] utilizes latent tokens to efficiently integrate cross-attention into adapters. However, these existing approaches are not well-suited for video-radar RPM. This is primarily because naive modality alignment would hinder the learning of modality-specific features, and the substantial modality gap between video and radar further impedes direct fusion of their features. To this end, we present the SATM adapter, specifically tailored for video-radar RPM fine-tuning. + +# 3. Method + +In this paper, we focus on constructing a robust multimodal RPM model on the video-radar RPM dataset. To this end, we first leverage existing large-scale unimodal datasets to pre-train our backbones. Afterwards, we freeze the backbones to preserve pre-trained knowledge and employ both vanilla and SATM adapters to fine-tune them into a video-radar RPM model. The proposed SATM adapter consists of two core components, Spatial Alignment and Temporal Matching, which can effectively narrow the modality gap and extract cross-modal features. The general framework + +for video-radar fine-tuning is depicted in Fig. 2. We will present our method in more detail in the following parts of this section. + +# 3.1. Backbone Design + +# 3.1.1. Video Backbone + +Following [19], we utilize the combination of DiffNorm and UniFormer [18] as our video backbone. DiffNorm is a lightweight module that utilizes difference operation and normalization to efficiently integrate the appearance and dynamic features of input videos. UniFormer is an enhanced version of the Vision Transformer (ViT) designed for video analysis. Each UniFormer block consists of Dynamic Position Embedding (DPE), Multi-Head Relation Aggregator (MHRA), and Feed Forward Network (FFN). MHRA is the core component that unifies convolution and attention in a concise transformer format. It employs convolution to tackle feature redundancy in shallow layers while utilizing self-attention for global modeling in deep layers. + +Formally, given an input video $X_{v} \in \mathbb{R}^{3 \times T_{1} \times H \times W}$ , the DiffNorm module first transforms it into $F_{v}^{(0)} \in \mathbb{R}^{3 \times T_{1} \times H \times W}$ as the input of UniFormer. Then the forward process of a UniFormer block can be expressed as: + +$$ +Y _ {v} ^ {(l)} = \mathrm {D P E} \left(F _ {v} ^ {(l - 1)}\right) + F _ {v} ^ {(l - 1)} +$$ + +$$ +Z _ {v} ^ {(l)} = \operatorname {M H R A} \left(\operatorname {N o r m} \left(Y _ {v} ^ {(l)}\right)\right) + Y _ {v} ^ {(l)} \tag {1} +$$ + +$$ +F _ {v} ^ {(l)} = \mathrm {F F N} (\mathrm {N o r m} (Z _ {v} ^ {(l)})) + Z _ {v} ^ {(l)} +$$ + +where $F_{v}^{(l - 1)}$ and $F_{v}^{(l)}$ are the output of the $l - 1$ -th and the $l$ -th UniFormer block, respectively. We will omit the normalization layers in the rest of the sections for brevity. + +# 3.1.2. Radar Preprocessing and Backbone + +The fundamental principle of radar-based physiological measurement involves capturing the subtle chest movement caused by cardiac activity. To extract displacement information from the radar intermediate frequency (IF) signal, we first perform range-FFT and coarse range gating to generate a range-time matrix and filter out irrelevant reflections. Afterwards, given that the 2D antenna array enables signal sampling in the spatial domain, we leverage beamforming to separate the reflections from different spatial positions: + +$$ +\boldsymbol {B} \left(\theta_ {A}, \theta_ {E}\right) = \boldsymbol {W} ^ {H} \left(\theta_ {A}, \theta_ {E}\right) \boldsymbol {M} +$$ + +$$ +\boldsymbol {W} \left(\theta_ {A}, \theta_ {E}\right) = \left[ \dots , \exp \left(- j \boldsymbol {k} \left(\theta_ {A}, \theta_ {E}\right) \boldsymbol {d} _ {r} ^ {(i)}\right), \dots \right] ^ {T} \tag {2} +$$ + +$$ +\boldsymbol {k} \left(\theta_ {A}, \theta_ {E}\right) = \frac {2 \pi}{\lambda} \left[ \cos \theta_ {E} \cos \theta_ {A}, \cos \theta_ {E} \sin \theta_ {A}, \sin \theta_ {E} \right] +$$ + +where $M \in \mathbb{R}^{N_C \times R \times T_2}$ is the range-time matrix, with $N_C, R, T_2$ indicating the number of virtual channels, range bins and frames, respectively. $\pmb{W}(\theta_A, \theta_E) \in \mathbb{R}^{N_C}$ is the steering vector in the direction of azimuth angle $\theta_A$ and elevation angle $\theta_E$ , and $\pmb{d}_r^{(i)}$ is the position of the i-th virtual channel. The shape of the beamformed output $\pmb{B}$ is + +![](images/4622109d85be9d7d0bbc8cd42166a8a686f3a9afe126f93932a171eff27701d5.jpg) +Figure 2. The overall framework of our model: Both backbones are based on UniFormer to facilitate cross-modal fusion. Vanilla adapters and SATM adapters are inserted into the shallow and deep layers of UniFormers, respectively, for multimodal fine-tuning. Each SATM adapter comprises two key components: Spatial Alignment and Temporal Matching. The Spatial Alignment module serves to align the spatial distributions of video and radar features, while the Temporal Matching module further refines the aligned features along the temporal dimension. Finally, the refined cross-modal features are passed through a vanilla adapter and added to the corresponding backbone as complementary features. + +$(\Theta_A \times \Theta_E) \times R \times T_2$ , where $(\Theta_A \times \Theta_E)$ is the number of angular bins. Following [3, 43], we further utilize a second-order differentiator to amplify cardiac movements while suppressing noise signals as follows: + +$$ +\begin{array}{l} f _ {r, t} = \left(\psi_ {t - 3} + \psi_ {t + 3}\right) + 2 \left(\psi_ {t - 2} + \psi_ {t + 2}\right) \tag {3} \\ - \left(\psi_ {t - 1} + \psi_ {t + 1}\right) - 4 \psi_ {t} \\ \end{array} +$$ + +where $\psi_t$ is the phase of $\pmb{B}$ at the t-th frame. The output $F_r^{(0)} \in \mathbb{R}^{(\Theta_A \times \Theta_E) \times R \times T_2}$ is reshaped into $\mathbb{R}^{1 \times S \times T_2}$ to serve as the input of the radar backbone, where $S = \Theta_A \times \Theta_E \times R$ is the spatial dimension of the radar modality. + +Considering the necessity to model the mechanical motion of the human chest across both spatial and temporal dimensions [43], we adopt 2D UniFormer as our radar backbone. Overall, our video and radar backbones are symmetrically structured, which greatly facilitates cross-modal fusion. + +# 3.2. SATM Adapter + +To enable effectively fine-tuning of the unimodal backbones and facilitate cross-modal modeling, we propose the SATM adapter. As illustrated in Fig. 2, Radar-to-Video SATM adapters and Video-to-Radar SATM adapters are inserted in parallel with the MHRA and FFN modules of the video and radar UniFormers respectively to extract complementary features from the other modality. As discussed in Sec. 1, the fundamental differences in measurement principles between video- and radar-based methods introduce + +a significant modality gap between their features and implicate cross-modal fusion. Considering the different ROIs of video and radar (face vs. chest), we present the Spatial Alignment (SA) module to align the spatial distributions of their features. As for the differences in captured signals (color changes vs. mechanical displacements), we design the Temporal Matching (TM) module to reduce waveform discrepancies and refine the spatially aligned features. Next, we will provide a detailed explanation of these modules. + +# 3.2.1. Spatial Alignment + +The physiological information contained in the face and chest is inherently distributed in distinct forms, resulting in different spatial distributions of video and radar features. We would like to emphasize that this dissimilarity is not merely reflected in the different spatial dimensions of $F_{v}$ and $F_{r}$ ( $HW$ vs. $S$ ), and naively aligning their size through interpolation or linear projection is insufficient. + +To address this challenge, we design the Spatial Alignment module. Given the video feature $F_{v} \in \mathbb{R}^{C \times t_{1} \times h \times w}$ and radar feature $F_{r} \in \mathbb{R}^{C \times s \times t_{2}}$ , the core idea involves utilizing cross-attention to align the spatial distributions of their spatial tokens, i.e. $f_{v} \in \mathbb{R}^{C \times t_{1}}$ and $f_{r} \in \mathbb{R}^{C \times t_{2}}$ . Without loss of generality, we use the Radar-to-Video Spatial Alignment as an example for explanation, with its overall pipeline illustrated in Fig. 3 (a). First, bilinear interpolation is applied to align the temporal dimensions of $F_{v}$ and $F_{r}$ . The primary challenge then lies in effectively transforming the 2D spatial tokens for the calculation of the attention + +![](images/792d1c5ed9721d78fcf6a702d10073cd07d5ab30d9cc87a916cecadcf81003ca.jpg) +Figure 3. The pipelines of Radar-to-Video Spatial Alignment and Temporal Matching. "MatMul" and "Hadamard" denote matrix product and Hadamard product, respectively. $\mathrm{FFT}_T$ represents FFT along the temporal dimension. + +scores. Considering that the final predicted 1D PPG signals are generated by the integration of the spatial and channel dimensions, we propose treating spatial tokens as distinct candidates for PPG signal prediction at different spatial locations and then projecting them from the 2D space into a unified 1D time series space by compressing their channel dimensions. Specifically, we leverage linear projections to transform $F_{v}$ and $F_{r}$ into the query matrix $S_{v} \in \mathbb{R}^{hw\times t_{1}}$ and the key matrix $S_{r} \in \mathbb{R}^{s\times t_{1}}$ , respectively. Afterwards, the alignment between video and radar spatial tokens can be performed in the time series space and the resulting aligned radar features are computed as follows: + +$$ +\mathrm {S A} _ {r 2 v} \left(F _ {v}, F _ {r}\right) = \operatorname {S o f t m a x} \left(\frac {S _ {v} S _ {r} ^ {T}}{\sqrt {t _ {1}}}\right) F _ {r} \tag {4} +$$ + +Note that the matrix multiplication between the attention score and $F_{r}$ is performed channel-wise. + +Now let us revisit the way to transform the spatial tokens. Two alternative approaches involve compressing the temporal dimension or flattening the channel and temporal dimensions of spatial tokens. However, the former variant risks irreversible information loss due to excessive compression of temporal signals, which is paramount for the prediction of PPG signals. The second approach, while straightforward, fails to account for the inherent differences between the channel and temporal dimensions. In contrast, our spatial alignment, which operates in the time series space, adheres to the paradigm of PPG signal regression and has the advantage of preserving temporal information. A quantitative comparison is provided in Sec. 4.4. + +# 3.2.2. Temporal Matching + +Although video and radar features are temporally synchronized and ideally encompass consistent frequency information, i.e. heart rate, the inherent disparity between wave + +forms derived from skin color changes and chest mechanical movements introduces a significant modality gap in the temporal dimension. To this end, we further utilize Temporal Matching to refine the cross-modal features. According to the Matched Filter theory, given an input signal $r(t)$ composed of the target signal $s(t)$ and the noise $n(t)$ , an output signal with the highest SNR at time step $T$ could be obtained by applying the filter $h(t) = s(T - t)$ . Inspired by this, we formulate the spatially aligned features as $r(t)$ and the waveform discrepancy as $n(t)$ , and the refined feature could be obtained through matched filtering. Specifically, given the video feature $F_{v} \in \mathbb{R}^{C \times t_{1} \times h \times w}$ and the spatially aligned radar feature $F_{r}^{\prime} \in \mathbb{R}^{C \times t_{1} \times h \times w}$ , the refined radar-to-video feature can be computed by: + +$$ +\mathrm {T M} _ {r 2 v} \left(F _ {v}, F _ {r} ^ {\prime}\right) = \mathrm {D W c o n v} \left(\mathrm {B N} \left(\mathcal {F} _ {T} ^ {- 1} \left(\overline {{\mathcal {F} _ {T} \left(F _ {v}\right)}} \mathcal {F} _ {T} \left(F _ {r} ^ {\prime}\right)\right)\right) \right. \tag {5} +$$ + +where $\mathcal{F}_T(\cdot)$ denotes Fast Fourier Transform (FFT) along the temporal dimension and $\overline{\mathcal{F}_T(\cdot)}$ represents FFT followed by the conjugate operation. DWConv and BN are depthwise convolution and batch normalization, respectively. + +# 3.2.3. Cross-Modal Fusion + +Overall, the forward process of the Radar-to-Video SATM adapter can be represented as: + +$$ +F _ {r 2 v} = \mathrm {T M} _ {r 2 v} \left(F _ {v}, \mathrm {S A} _ {r 2 v} \left(F _ {v}, F _ {r}\right)\right) +$$ + +$$ +\operatorname {S A T M} _ {r 2 v} \left(F _ {v}, F _ {r}\right) = W _ {u p} \left(\operatorname {R e L U} \left(W _ {\text {d o w n}} \left(F _ {v} + \alpha \cdot F _ {r 2 v}\right)\right)\right) \tag {6} +$$ + +where $W_{down}, W_{up}$ are the down-projection and up-projection layers of the vanilla adapter [12] and $\alpha$ is a learnable scaling factor. Afterward, the forward process of a video UniFormer block can be reformulated as: + +$$ +Z _ {v} ^ {(l)} = \operatorname {M H R A} \left(Y _ {v} ^ {(l)}\right) + Y _ {v} ^ {(l)} + \operatorname {S A T M} _ {r 2 v} \left(Y _ {v} ^ {(l)}, Y _ {r} ^ {(l)}\right) +$$ + +$$ +F _ {v} ^ {(l)} = \operatorname {F F N} \left(Z _ {v} ^ {(l)}\right) + Z _ {v} ^ {(l)} + \operatorname {S A T M} _ {r 2 v} \left(Z _ {v} ^ {(l)}, Z _ {r} ^ {(l)}\right) \tag {7} +$$ + +The forward processes of the Video-to-Radar SATM adapter and the fine-tuned radar UniFormer blocks can also be computed similarly to Eq. (6) and Eq. (7). + +In summary, Spatial Alignment aligns the spatial distributions of video and radar features in a unified time series space, while Temporal Matching utilizes a matched filter to handle the waveform difference between their signals. They complement each other by processing the cross-modal features along spatial and temporal dimensions separately, thereby effectively narrowing the substantial modality gap between video and radar features. According to [9], it is essential to balance the learning of unimodal and cross-modal features. Consequently, we only insert our SATM adapters into the last two stages of UniFormers while utilizing vanilla adapters in the first two stages. The output features of both backbones are integrated by a lightweight fusion stem and subsequently passed through a regression module to generate the predicted PPG signal. + +# 4. Experiments + +# 4.1. Dataset + +We utilize the EquiPleth [41] dataset to evaluate our method, which contains 533 30-second samples collected from 28 light, 49 medium, and 14 dark skin tone volunteers. All samples are recorded under motionless conditions and with sufficient illumination. + +In addition, considering that the EquiPleth dataset mainly focuses on constrained scenarios and shows limited practical diversity, we built a new dataset called MMRPM to promote a more comprehensive evaluation of our method. MMRPM comprises 180 60-second samples from 31 subjects, all of which are recorded under varying illumination conditions and subject motions. + +Specifically, MMRPM encompasses three illumination conditions: dark, medium, and light. The "dark" and "medium" settings correspond to illumination levels of approximately 2.5 and 6.3 lux respectively, while "light" represents samples are collected under sufficient sunlight and indoor illumination. Furthermore, two types of subject activities are considered: sitting still and free head movement. Overall, MMRPM covers six scenarios, where low-light conditions are challenging to the video modality and head motions have a greater impact on the radar modality. $^{1}$ + +# 4.2. Implementation Details + +Our proposed method is implemented using PyTorch. For video preprocessing, we first employ MTCNN to crop the face region in the initial frame and fix the region across subsequent frames. Then, we sample each video into clips with a time window of 160 frames and a step size of 80 frames, which are then resized to $96 \times 96$ pixels. In terms of radar preprocessing, the angular range selected for beamforming is $66^{\circ} \leq \Theta_{A} \leq 114^{\circ}, -24^{\circ} \leq \Theta_{E} \leq 24^{\circ}$ . + +We utilize UniFormer-S [18] and MAM in MAGIC [55] as our feature extractor and fusion stem respectively. As for the regression module, we follow the design in ADDP [19]. The video backbone is jointly pre-trained on VIPL [31, 32], MMPD [40], UBFC [1], PURE [39] and BUAA [47], and we use the dataset introduced in [53] for radar pre-trianing. + +During multimodal fine-tuning, both backbones are frozen, and only the adapters, fusion stem and regression module are trained. We employ the loss functions introduced in [51] and utilize the Adam optimizer for training, with the learning rate and weight decay set to 5e-4 and 5e-5 respectively. The model is trained for 40 epochs with a batch size of 2. Following [38], we apply random horizontal flipping, spatially resized crop and random intensity noise to augment the input videos. For radar augmentation, we randomly shift the range bins of the range-time matrices. + +Following [41], we perform our evaluation of the PPG signals over 300 samples windows with a stride of 128 samples. The heart rate corresponding to a specific signal clip is calculated using a combination of band-pass filtering, detrending and peak detection in the frequency domain. Finally, the standard deviation of the error (Std), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (R) are adopted to evaluate the predicted HR error. The unit of Std, MAE and RMSE is beats per minute (bpm). We split the MMRPM dataset into 5 subject-exclusive folds and use the official split for the EquiPleth dataset. + +# 4.3. Main Results + +We first compare our SATM with existing video-radar RPM methods [6, 41, 46] and several state-of-the-art multimodal adapter methods [2, 22, 28, 48] on the EquiPleth [41] dataset. To ensure a fair comparison, we provide extra results of improved versions of existing RPM methods by applying unimodal pre-training and replacing their original backbones with UniFormers. Since the radar used in the EquiPleth dataset activates only 1 transmitter and 1 receiver, we do not perform beamforming during radar preprocessing for this dataset. For existing video-radar RPM methods, we use the official implementation of EquiPleth, and re-implement Fusion-Vital and CardiacMamba, as the codes have not been released by the authors. As shown in Tab. 1, all of these methods exhibit RMSEs exceeding 2.9 bpm when trained from scratch. Their performance consistently improves when unimodal pre-training is applied, underscoring the significance of unimodal pre-training when video-radar data is limited. However, they still display MAEs greater than 0.64 bpm, regardless of whether using their original backbones or UniFormers, indicating that their full fine-tuning-based fusion strategies are suboptimal for the small-scale multimodal RPM dataset. In contrast, our method employs SATM adapters for multimodal fine-tuning, effectively leveraging pre-trained knowledge and enabling superior cross-modal fusion. + +Furthermore, our SATM also demonstrates superior performance compared to existing multimodal adapter methods. The subpar performance of UniAdapter and MMA suggests that their parameter-sharing strategies are ill-suited for multimodal RPM, as they impede the learning of modality-specific features. To adapt BAT for video-radar RPM, we employ an additional linear projection layer to align the token numbers of the video and radar branches. However, it fails to handle the significant modality gap between video and radar, thereby resulting in an RMSE exceeding 1.65 bpm. While the best baseline, LAVISH, achieves an RMSE of 1.498 bpm, our SATM yields a considerably lower RMSE of 1.242 bpm. + +We have also evaluated SATM on our self-collected + +
MethodsOriginal ModalityUnimodal Pre-trainedStd↓MAE↓RMSE↓R↑Trainable Params (M)↓
EquiPleth[41]Video-Radar×3.2741.0633.2830.9573.95
Fusion-Vital[6]Video-Radar×9.2883.7509.2970.709715.26
CardiacMamba[46]Video-Radar×2.9240.9012.9270.9759.93
EquiPlethVideo-Radar√2.9790.9682.9900.9653.95
Fusion-VitalVideo-Radar√9.1213.7509.1300.735715.26
CardiacMambaVideo-Radar√2.7970.8682.7710.9759.93
EquiPleth*Video-Radar√1.4350.6711.4360.99245.31
Fusion-Vital*Video-Radar√2.1430.6502.1570.983194.61
CardiacMamba*Video-Radar√2.3130.7442.3240.98033.62
UniAdapter[28]Vision-Language√1.6530.5901.6560.99010.60
MMA[48]Vision-Language√1.8350.6261.8420.98811.09
BAT[2]RGB-TIR√1.6790.6301.6840.99085.35
LAVISH[22]Audio-Visual√1.4930.5571.4980.99110.85
SATM (Ours)Video-Radar√1.2170.5401.2420.99410.78
+ +Table 1. Overall performance on the EquiPleth dataset. The best and second best results are in bold and underlined respectively. * denotes that we replace the methods' original backbones with UniFormers while preserving their fusion strategies. Blue and red indicate the existing video-radar RPM methods and existing adapter methods for other multimodal tasks, respectively. + +
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth7.6032.4217.6030.715
Fusion-Vital9.1395.1659.3150.501
CardiacMamba6.0992.1606.1070.790
EquiPleth*5.9641.7395.9740.806
Fusion-Vital*5.5501.6625.5890.827
CardiacMamba*5.5261.8435.5640.829
UniAdapter5.5091.5885.5100.831
MMA5.6941.7225.6940.819
BAT6.4322.1516.4330.760
LAVISH5.6601.4545.3120.844
SATM (Ours)5.0631.2925.0740.858
+ +MMRPM dataset, which covers more challenging and diverse scenarios than the EquiPleth dataset. As illustrated in Tab. 2, existing video-radar RPM methods exhibit poor results on this dataset, highlighting their limited capability to handle complex scenarios. In comparison, our SATM consistently outperforms the baseline methods in all metrics. + +We further conduct a cross-dataset evaluation by training the models on the EquiPleth dataset and evaluating them on the MMRPM dataset. As shown in Tab. 3, the improved versions of existing multimodal RPM methods exhibit subpar performance compared to adapter-based methods. This is primarily because they rely on full fine-tuning-based fusion strategies and are prone to overfitting small-scale training sets. In contrast, our method, built upon SATM adapters, achieves the best generalization performance and suffers + +Table 2. Overall performance on the MMRPM dataset. + +
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth*8.7874.2569.0820.629
Fusion-Vital*8.5384.4578.8970.642
CardiacMamba*9.0485.0389.3080.603
LAVISH8.0263.6878.2320.712
SATM (Ours)7.5423.4237.6560.738
+ +Table 3. Results of the cross-dataset evaluation from the EquiPleth dataset to the MMRPM dataset. + +minimal performance degradation compared to other baselines. Although it still lags behind the performance in the intra-dataset evaluation, we argue that this is reasonable, given the significant differences between these two datasets.[2] + +# 4.4. Ablation Study + +Key Components. Our SATM adapter mainly consists of two key components: Spatial Alignment and Temporal Matching. In Tab. 4, we conduct an ablation study on the MMRPM dataset to assess the effectiveness of these modules. First, compared to the vanilla adapter (the first row), the adapter with Spatial Alignment reduces the MAE and RMSE by 0.087 bpm and 0.149 bpm respectively. This can be attributed to its capability to facilitate modality complementarity by extracting spatially aligned cross-modal features to the corresponding backbones. The MAE and RMSE can be further reduced to 1.292 bpm and 5.074 bpm when Temporal Matching is inserted, highlighting the necessity + +
SATMStd↓MAE↓RMSE↓R↑
××5.4521.5415.4530.836
✓×5.2981.4545.3040.845
✓✓5.0631.2925.0740.858
+ +Table 4. Ablation study on the key components. "SA" and "TM" denote Spatial Alignment and Temporal Matching respectively. + +
VariantsStd↓MAE↓RMSE↓R↑
Interpolate6.0411.8766.0520.795
Linear5.5791.6285.5820.828
Channel6.0871.8076.0910.792
Flatten5.5881.6935.5950826
Ours5.2981.4545.3040.845
+ +to refine the aligned features along the temporal dimension. + +Alternative Approaches for Spatial Alignment. We also investigate four alternative approaches for Spatial Alignment. Specifically, "Interpolate" and "Linear" refer to methods that directly align the sizes of video and radar features through interpolation and linear projection instead of cross-attention. The "Channel" and "Flatten" indicate that the alignment is performed in the channel space and on the flattened spatial tokens respectively. As shown in Tab. 5, both "Interpolate" and "Linear" exhibit suboptimal performance, achieving MAEs of 1.876 bpm and 1.628 bpm. These methods only align spatial sizes while neglecting inherent differences in spatial distributions between video and radar features, thereby failing to address the modality gap. + +The "Channel" variant compresses the temporal dimensions of spatial tokens for the computation of the attention scores, focusing solely on the overall temporal period while neglecting fine-grained temporal variability. This results in degraded performance with an RMSE exceeding 6 bpm. The "Flatten" variant preserves temporal information, but conflates temporal and channel dimensions, overlooking their intrinsic differences. Therefore, it ultimately yields subpar performance with an MAE of 1.693 bpm and an RMSE of 5.595 bpm. In contrast, Spatial Alignment performed in the time series space effectively avoids temporal information loss and dimensional distortion, showing superior performance compared to these four variants. + +Temporal Matching vs. Temporal Attention. Since attention is a straightforward method for cross-modal fusion, we compare our Temporal Matching (TM) with Temporal Attention (TA). Specifically, we replace the matched filtering operation with a cross-attention along the temporal dimension. As shown in Tab. 6, our TM consistently outperforms TA across all metrics. Moreover, since the time complexity of FFT is $\mathcal{O}(N\log N)$ , TM is also more com + +Table 5. Ablation study on the design of Spatial Alignment. + +
TemporalStd↓MAE↓RMSE↓R↑
Attention5.1551.3675.1300.853
Matching5.0631.2925.0740.858
+ +Table 6. Comparison between Temporal Matching and Temporal Attention. + +
PositionsStd↓MAE↓RMSE↓R↑
SATM-45.1681.3865.1790.852
SATM-35.1071.3435.1090.857
SATM-25.0631.2925.0740.858
SATM-15.5261.6145.5260.830
+ +Table 7. Ablation study on the position to add SATM adapters. + +putationally efficient compared to the quadratic complexity of TA. + +Position to Insert Our SATM Adapters. We also explore the impact of different positions to insert our SATM adapters. We present 4 positions for insertion, formulated as "SATM-n", where "n" denotes that we use SATM adapters in the last "n" stages of the UniFormers and employ vanilla adapters in earlier stages. As reported in Tab. 7, SATM-1 displays an MAE higher than the naive baseline (1.614 bpm vs. 1.541 bpm), indicating that inserting the cross-modal module too late would hinder performance due to the lack of cross-modal interaction. Meanwhile, the performance of SATM-3 and SATM-4 is also limited due to insufficient modality-specific modeling. In contrast, the results demonstrate that SATM-2 strikes a good trade-off between the learning of unimodal and corss-modal features. + +# 5. Conclusion + +In this paper, we focus on video-radar RPM measurement. To overcome the data scarcity challenge, we make full use of unimodal pre-training and present the SATM adapter to effectively fine-tune unimodal backbones into a multimodal RPM model. The SATM adapter leverages Spatial Alignment and Temporal Matching to collaboratively narrow the modality gap between video and radar and effectively extract cross-modal features for both backbones. To thoroughly evaluate our method, we collect a challenging dataset that encompasses diverse illumination conditions and subject motions. Extensive experiments conducted on our self-collected and existing public datasets demonstrate that SATM outperforms existing methods. + +Limitations. While SATM effectively constructs a robust video-radar RPM model, it primarily focuses on cross-modal interaction and still relies on vanilla adapters for unimodal modeling during fine-tuning. In the future, we hope to develop a unified framework that harmonizes cross-modal and unimodal modeling. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China under Grant 62172381. + +# References + +[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6, 1 +[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 2, 3, 6, 7 +[3] Jinbo Chen, Dongheng Zhang, Zhi Wu, Fang Zhou, Qibin Sun, and Yan Chen. Contactless electrocardiogram monitoring with millimeter wave radar. IEEE Transactions on Mobile Computing, 23(1):270-285, 2022. 1, 2, 4 +[4] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 349-365, 2018. 1 +[5] Li-Wen Chiu, Yang-Ren Chou, Yi-Chiao Wu, and Bing-Fei Wu. Deep-learning based remote photoplethysmography measurement in driving scenarios with color and near-infrared images. IEEE Transactions on Instrumentation and Measurement, 2023. 2 +[6] Jae-Ho Choi, Ki-Bong Kang, and Kyung-Tae Kim. Fusion-vital: Video-rf fusion transformer for advanced remote physiological measurement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1344-1352, 2024. 1, 2, 3, 6, 7 +[7] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE transactions on biomedical engineering, 60(10):2878-2886, 2013. 2 +[8] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiological measurement, 35(9):1913, 2014. 2 +[9] Chenzhuang Du, Jiaye Teng, Tingle Li, Yichen Liu, Tianyuan Yuan, Yue Wang, Yang Yuan, and Hang Zhao. On uni-modal feature learning in supervised multi-modal learning. In International Conference on Machine Learning, pages 8632-8656. PMLR, 2023. 5 +[10] Jingda Du, Si-Qi Liu, Bochao Zhang, and Pong C Yuen. Dual-bridging with adversarial noise generation for domain adaptive rpgg estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10355–10364, 2023. 1 +[11] Unsoo Ha, Salah Assana, and Fadel Adib. Contactless seismocardiography via deep learning radars. In Proceedings of the 26th annual international conference on mobile computing and networking, pages 1-14, 2020. 2 +[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3, 5 + +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[14] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, XueQing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 3 +[15] Jitesh Joshi and Youngjun Cho. ibvp dataset: Rgb-thermal rppg dataset with high resolution signal quality labels. Electronics, 13(7):1334, 2024. 2 +[16] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 3 +[17] Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17615-17626, 2023. 2 +[18] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 3, 6, 2 +[19] Qian Liang, Yan Chen, and Yang Hu. Continual learning for remote physiological measurement: Minimize forgetting and simplify inference. In European conference on computer vision, pages 126-144. Springer, 2024. 3, 6 +[20] Zhu Lianghui, Liao Bencheng, Zhang Qian, Wang Xinlong, Liu Wenyu, and Wang Xinggang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv. Org, pages arXiv-org, 2024. 3 +[21] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 2 +[22] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2299–2309, 2023. 3, 6, 7 +[23] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2 +[24] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 1 +[25] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3343-3347. IEEE, 2021. 2 + +[26] Yunfei Long, Abhinav Kumar, Daniel Morris, Xiaoming Liu, Marcos Castro, and Punarjay Chakravarty. Radiant: Radar-image association network for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1808-1816, 2023. 2 +[27] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 2 +[28] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 2, 3, 6, 7 +[29] Hao Lu, Zitong Yu, Xuesong Niu, and Ying-Cong Chen. Neuron structure modeling for generalizable remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18589-18599, 2023. 1 +[30] Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2021. 2 +[31] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14, pages 562–576. Springer, 2019. 2, 6, 1 +[32] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 2, 6, 1 +[33] Xuesong Niu, Xingyuan Zhao, Hu Han, Abhijit Das, Antitza Dantcheva, Shiguang Shan, and Xin Chen. Robust remote heart rate estimation from face utilizing spatial-temporal attention. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-8. IEEE, 2019. 1 +[34] Soyeon Park, Bo-Kyeong Kim, and Suh-Yeon Dong. Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation. IEEE Transactions on Instrumentation and Measurement, 71:1-10, 2022. 2 +[35] Vladimir L Petrović, Milica M Janković, Anita V Lupšić, Veljko R Mihajlović, and Jelena S Popovic-Božović. High-accuracy real-time monitoring of heart rate variability using 24 ghz continuous-wave doppler radar. *IEEE Access*, 7: 74721–74733, 2019. 1 +[36] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10):10762-10774, 2010. 2 +[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 2 + +[38] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 6 +[39] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6, 1 +[40] Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, and Xin Liu. Mmpd: multidomain mobile video physiology dataset. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5. IEEE, 2023. 2, 6, 1 +[41] Alexander Vilesov, Pradyumna Chari, Adnan Armouti, Anirudh Bindiganavale Harish, Kimaya Kulkarni, Ananya Deoghare, Laleh Jalilian, and Achuta Kadambi. Blending camera and 77 ghz radar sensing for equitable, robust plethysmography. ACM Trans. Graph., 41(4):36-1, 2022. 1, 2, 6, 7 +[42] Fengyu Wang, Xiaolu Zeng, Chenshu Wu, Beibei Wang, and KJ Ray Liu. mmhrv: Contactless heart rate variability monitoring using millimeter-wave radio. IEEE Internet of Things Journal, 8(22):16623-16636, 2021. 1 +[43] Haoyu Wang, Jinbo Chen, Dongheng Zhang, Zhi Lu, Changwei Wu, Yang Hu, Qibin Sun, and Yan Chen. Contactless radar heart rate variability monitoring via deep spatiotemporal modeling. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 111-115. IEEE, 2024. 1, 2, 4 +[44] Kai Wang, Yapeng Tian, and Dimitrios Hatzinakos. Towards efficient audio-visual learners via empowering pre-trained vision transformers with cross-modal adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1837–1846, 2024. 3 +[45] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE transactions on Biomedical Engineering, 62(2): 415-425, 2014. 2 +[46] Zheng Wu, Yiping Xie, Bo Zhao, Jiguang He, Fei Luo, Ning Deng, and Zitong Yu. Cardiacmamba: A multimodal rgb-rf fusion framework with state space models for remote physiological measurement. arXiv preprint arXiv:2502.13624, 2025. 1, 2, 3, 6, 7 +[47] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6, 1 +[48] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2, 3, 6, 7 +[49] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from fa + +cial videos using spatio-temporal networks. arXiv preprint arXiv:1905.02419, 2019. 1, 2 +[50] Zitong Yu, Xiaobai Li, Xuesong Niu, Jingang Shi, and Guoying Zhao. Autohr: A strong end-to-end baseline for remote heart rate measurement with neural searching. IEEE Signal Processing Letters, 27:1245-1249, 2020. 1 +[51] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip HS Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4186-4196, 2022. 1, 2, 6 +[52] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021.3 +[53] Bin-Bin Zhang, Dongheng Zhang, Yadong Li, Zhi Lu, Jinbo Chen, Haoyu Wang, Fang Zhou, Yu Pu, Yang Hu, Li-Kun Ma, et al. Monitoring long-term cardiac activity with contactless radio frequency signals. Nature Communications, 15(1):1-11, 2024. 2, 6, 1 +[54] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 1 +[55] Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. Centering the value of every modality: Towards efficient and resilient modality-agnostic semantic segmentation. In European Conference on Computer Vision, pages 192–212. Springer, 2024. 6 \ No newline at end of file diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/images.zip b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f6c45c1662b343e77a979a1a90d2bbed20234c7 --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75ac4ea20a419ed3ae3053c28e6833f28b325fc35390b5b2cadaa2a63521d93f +size 528695 diff --git a/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/layout.json b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f59eec4f0b4c2efb460301b1ba1e99953cf068d9 --- /dev/null +++ b/2025/Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement/layout.json @@ -0,0 +1,7943 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "Spatial Alignment and Temporal Matching Adapter for Video-Radar Remote Physiological Measurement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "spans": [ + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": "Qian Liang" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Ruixu Geng" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Jinbo Chen" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Haoyu Wang" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Yan Chen" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": ", Yang Hu" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": "University of Science and Technology of China, " + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 160, + 512, + 219 + ], + "type": "text", + "content": "Nanyang Technological University \n{qianliang, gengruixu, why1999}@mail.ustc.edu.cn, jinbo.chen@ntu.edu.sg \n{eecyan, eeyhu}@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 571 + ], + "type": "text", + "content": "Remote physiological measurement (RPM) based on video and radar has made significant progress in recent years. However, unimodal methods based solely on video or radar sensor have notable limitations due to their measurement principles, and multimodal RPM that combines these modalities has emerged as a promising direction. Despite its potential, the lack of large-scale multimodal data and the significant modality gap between video and radar pose substantial challenges in building robust videoradar RPM models. To handle these problems, we suggest leveraging unimodal pre-training and present the Spatial alignment and Temporal Matching (SATM) Adapter to effectively fine-tune pre-trained unimodal backbones into a multimodal RPM model. Given the distinct measurement principles of video- and radar-based methods, we propose Spatial Alignment to align the spatial distribution of their features. Furthermore, Temporal Matching is applied to mitigate waveform discrepancies between video and radar signals. By integrating these two modules into adapters, the unimodal backbones could retain their modality-specific knowledge while effectively extracting complementary features from each other. Extensive experiments across various challenging scenarios, including low light conditions and head motions, demonstrate that our approach significantly surpasses the state-of-the-art methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 137, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "Remote physiological measurement has attracted a lot of attention recently owing to its ability to detect physiological signals without contact. A prominent line of research focuses on video-based remote photoplethysmography (rPPG), which uses ordinary cameras to extract vital signals by detecting periodic changes in skin color induced by heartbeats [4, 10, 24, 29, 49, 51]. Beyond video-based" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 555, + 350 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 350 + ], + "type": "image", + "image_path": "84f65275c7434545850e7505102f360569ce9f991d5cbf0f4a54965fb7b916f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 357, + 556, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 357, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 556, + 437 + ], + "type": "text", + "content": "Figure 1. Frameworks of video-radar RPM methods: (a) EquiPleth [41] trains unimodal backbones on the multimodal dataset and leverage a late fusion model to ensemble their predictions. (b) Fusion-Vital [6] and CardiacMamba [46] train their end-to-end models from scratch on the multimodal dataset. (c) Our method makes full use of unimodal pre-training and propose the SATM adapter to enable effective multimodal adaptation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 459, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 495 + ], + "type": "text", + "content": "methods, an alternative approach involves the use of radar sensors to capture cardiac activities through subtle mechanical motions of the human chest [3, 35, 42, 43]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 605 + ], + "type": "text", + "content": "However, both video- and radar-based methods have notable limitations. Video-based methods are highly sensitive to light conditions and may be influenced by subjects' skin tones [41]. On the other hand, radar-based methods are much more susceptible to subject movements, and most existing methods are designed for strict-constraint scenarios. Therefore, it is necessary and promising to combine video and radar for RPM measurement to overcome the limitations of unimodal approaches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "Nevertheless, constructing a robust video-radar RPM model remains a significant challenge. Due to the complexity of recording synchronized videos, radar data, and ground-truth PPG signals, video-radar RPM datasets are mostly limited in scale. Consequently, training a model from scratch on such datasets is prone to severe overfitting. Unfortunately, existing multimodal RPM works [6, 41, 46] overlook this problem, leading to limited robustness and generalization ability (See Sec. 4.3). In addition, video" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8623" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "based methods focus primarily on color changes of facial skin, whereas radar-based methods capture mechanical displacements of the chest. This implies that these two modalities capture distinct signals from different regions of interest (ROIs), introducing a substantial modality gap and making it difficult to effectively fuse their features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 146, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 295, + 419 + ], + "type": "text", + "content": "For the data scarcity challenge, we suggest making full use of unimodal pre-training, as unimodal data are easier to collect and large-scale unimodal datasets for remote physiological measurement are already available [31, 40, 53]. As illustrated in Fig. 1 (c), we first utilize large-scale unimodal datasets to pre-train our backbones. Subsequently, instead of fully fine-tuning the entire network on limited multimodal data, we propose leveraging cross-modal adapters to fine-tune unimodal backbones into a multimodal model while preserving the pre-trained knowledge. Several previous studies have introduced adapter techniques into multimodal learning [2, 28, 48]. For instance, UniAdapter [28] and MMA [48] design parameter-sharing adapters to facilitate modality alignment in vision-language models, while BAT [2] proposes a bi-directional adapter to directly extract cross-modal features for multimodal tracking. However, multimodal adapter for video-radar RPM has not been explored before. Furthermore, the aforementioned methods are suboptimal for multimodal RPM for two reasons: (1) crude modality alignment would hamper the learning of modality-specific features in this context, and (2) the significant modality gap between video and radar, as discussed earlier, impedes the direct fusion of their features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 421, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 660 + ], + "type": "text", + "content": "To this end, we propose a novel SATM adapter to enable effective multimodal fine-tuning on the video-radar RPM dataset. Specifically, the Spatial Alignment (SA) and Temporal Matching (TM) modules are devised to manage the large modality gap between video and radar. To mitigate the gap introduced by their distinct ROIs, Spatial Alignment projects latent features from different modalities into a unified time series space and then utilizes cross-attention to align their spatial distributions. Additionally, to handle the gap caused by different waveforms of detected signals from video and radar, Temporal Matching further applies a matched filter to refine the spatially aligned features along the temporal dimension. These two modules narrow the modality gap collaboratively and are integrated into adapters to effectively extract cross-modal features. To validate the effectiveness of our approach, we collect a challenging dataset, MMRPM, which includes diverse light conditions and subject head motions. Comprehensive experiments conducted on the MMRPM and EquiPleth [41] datasets demonstrate the superiority of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 295, + 685 + ], + "type": "text", + "content": "The key contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to highlight the dilemma of data hungry and data scarcity in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "multimodal RPM. We propose leveraging unimodal pretraining and multimodal adapter to tackle this problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 96, + 553, + 192 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "content": "- To handle the significant modality gap between video and radar, we design the Spatial Alignment and Temporal Matching modules. These modules are further integrated into a novel SATM adapter, enabling effective multimodal fine-tuning while mitigating overfitting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 156, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 192 + ], + "type": "text", + "content": "- Extensive experiments across various challenging scenarios show the superiority of the proposed method compared to existing approaches." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 552, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 552, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 552, + 237 + ], + "type": "text", + "content": "2.1. Unimodal Remote Physiological Measurement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 434 + ], + "type": "text", + "content": "Most existing remote physiological measurement methods focus on video and radar sensors. Given that the optical absorption of hemoglobin molecules fluctuates along with the cardiac cycle, physiological signals can be remotely measured by detecting changes in skin color in RGB videos. Early works [7, 8, 36, 45] mainly utilize blind source separation and color space transformations to enhance the signal-to-noise ratio (SNR) of extracted rPPG signals. However, these methods are based on restrictive assumptions and struggle to handle complex scenarios. With the development of deep learning, researchers have employed 3D-CNN or modified vision transformer (ViT) to extract rPPG signals from raw videos [49, 51] or specifically designed STMaps [27, 32]. In addition to RGB videos, several studies have also explored rPPG measurement using near-infrared and thermal-infrared videos [5, 15, 34]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 590 + ], + "type": "text", + "content": "On the other hand, radar sensors can extract vital signals by detecting subtle movements of the human chest induced by physiological activities. [11] combines 4D beamforming and CNN to capture seismocardiography from radar sensors. [3] leverages the connection between cardiac mechanical activity and electrical activity to introduce the first radar-based electrocardiography (ECG) monitoring framework. [43] proposes a spatio-temporal network for radar-based Heart Rate Variability (HRV) detection. However, due to the minuscule scale of cardiac motions, radar-based physiological measurement often requires subjects to remain in a supine position and keep still, limiting its flexibility for real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 599, + 492, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 599, + 492, + 611 + ], + "spans": [ + { + "bbox": [ + 314, + 599, + 492, + 611 + ], + "type": "text", + "content": "2.2. Video-Radar RPM Measurement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Given the complementary properties of vision and radar modalities, vision-radar fusion has been widely explored in tasks such as object detection [17, 23, 26, 30] and depth estimation [21, 25, 37]. In contrast, video-radar RPM measurement remains an understudied area, with only three related works reported recently [6, 41, 46]. EquiPleth [41] focuses on the skin tone bias problem of video-based methods. They design a late fusion model to ensemble predic" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8624" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "tions from video and radar backbones and propose an adversarial training strategy to enhance the fairness of the final predictions. Fusion-Vital [6] first projects video and radar inputs into a shared time-difference domain and then employs attention mechanism for cross-modal fusion, showcasing comparable performance even in modality corruption scenarios. Similarly, CardiacMamba [46] adopts a weight-sharing VisionMamba [20] to develop an end-to-end video-radar RPM model. However, these works train their models from scratch on small-scale video-radar datasets, making them prone to overfitting the training sets. To overcome this limitation, we propose leveraging the relatively large-scale unimodal datasets for pre-training and design the SATM adapter to facilitate effective multimodal finetuning while preserving pre-trained knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 178, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 178, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 178, + 272 + ], + "type": "text", + "content": "2.3. Adapter Fine-Tuning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 276, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 563 + ], + "type": "text", + "content": "The increasing scale of foundation models has brought about the challenge of prohibitive cost and substantial data requirements to fully fine-tune these models. Therefore, parameter-efficient fine-tuning (PEFT) has attracted growing attention [12, 13, 16, 52]. Adapter [12], comprising a pair of projection matrices and a nonlinear activation function, is one of the most representative PEFT methods. Although originally developed for natural language processing (NLP), recent works have extended its application to multimodal settings [14, 28, 44, 48]. [28, 48] leverage parameter-sharing adapters to fine-tune vision-language models and promote modality alignment, which is particularly crucial in vision-text tasks. [2] proposes a bi-directional adapter to foster modality complementarity in multimodal tracking. Considering cross-attention is an effective, yet computationally inefficient mechanism for cross-modal fusion, [22] utilizes latent tokens to efficiently integrate cross-attention into adapters. However, these existing approaches are not well-suited for video-radar RPM. This is primarily because naive modality alignment would hinder the learning of modality-specific features, and the substantial modality gap between video and radar further impedes direct fusion of their features. To this end, we present the SATM adapter, specifically tailored for video-radar RPM fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 111, + 586 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we focus on constructing a robust multimodal RPM model on the video-radar RPM dataset. To this end, we first leverage existing large-scale unimodal datasets to pre-train our backbones. Afterwards, we freeze the backbones to preserve pre-trained knowledge and employ both vanilla and SATM adapters to fine-tune them into a video-radar RPM model. The proposed SATM adapter consists of two core components, Spatial Alignment and Temporal Matching, which can effectively narrow the modality gap and extract cross-modal features. The general framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 107 + ], + "type": "text", + "content": "for video-radar fine-tuning is depicted in Fig. 2. We will present our method in more detail in the following parts of this section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 114, + 419, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 419, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 419, + 126 + ], + "type": "text", + "content": "3.1. Backbone Design" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 131, + 411, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 411, + 142 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 411, + 142 + ], + "type": "text", + "content": "3.1.1. Video Backbone" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "content": "Following [19], we utilize the combination of DiffNorm and UniFormer [18] as our video backbone. DiffNorm is a lightweight module that utilizes difference operation and normalization to efficiently integrate the appearance and dynamic features of input videos. UniFormer is an enhanced version of the Vision Transformer (ViT) designed for video analysis. Each UniFormer block consists of Dynamic Position Embedding (DPE), Multi-Head Relation Aggregator (MHRA), and Feed Forward Network (FFN). MHRA is the core component that unifies convolution and attention in a concise transformer format. It employs convolution to tackle feature redundancy in shallow layers while utilizing self-attention for global modeling in deep layers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "text", + "content": "Formally, given an input video " + }, + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "inline_equation", + "content": "X_{v} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}" + }, + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "text", + "content": ", the DiffNorm module first transforms it into " + }, + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "inline_equation", + "content": "F_{v}^{(0)} \\in \\mathbb{R}^{3 \\times T_{1} \\times H \\times W}" + }, + { + "bbox": [ + 313, + 301, + 555, + 352 + ], + "type": "text", + "content": " as the input of UniFormer. Then the forward process of a UniFormer block can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 359, + 487, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 359, + 487, + 374 + ], + "spans": [ + { + "bbox": [ + 358, + 359, + 487, + 374 + ], + "type": "interline_equation", + "content": "Y _ {v} ^ {(l)} = \\mathrm {D P E} \\left(F _ {v} ^ {(l - 1)}\\right) + F _ {v} ^ {(l - 1)}", + "image_path": "cfaf705ebcc28ef3224c8842ff2d17fe61c7ba78d2dc6666e3555f3ca717f521.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 376, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 376, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 359, + 376, + 553, + 392 + ], + "type": "interline_equation", + "content": "Z _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(\\operatorname {N o r m} \\left(Y _ {v} ^ {(l)}\\right)\\right) + Y _ {v} ^ {(l)} \\tag {1}", + "image_path": "fd2b5588cfcd744c34fada3a08018b640b27fb705853ecaec1784b6ec0f8a66c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 393, + 498, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 393, + 498, + 408 + ], + "spans": [ + { + "bbox": [ + 359, + 393, + 498, + 408 + ], + "type": "interline_equation", + "content": "F _ {v} ^ {(l)} = \\mathrm {F F N} (\\mathrm {N o r m} (Z _ {v} ^ {(l)})) + Z _ {v} ^ {(l)}", + "image_path": "e7ab59f2834019054b4a96f0920977be5ec4f7eaa5cabfeaa9ff45bcfda67341.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "inline_equation", + "content": "F_{v}^{(l - 1)}" + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "inline_equation", + "content": "F_{v}^{(l)}" + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "content": " are the output of the " + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "inline_equation", + "content": "l - 1" + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "content": "-th and the " + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 411, + 554, + 450 + ], + "type": "text", + "content": "-th UniFormer block, respectively. We will omit the normalization layers in the rest of the sections for brevity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 495, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 495, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 495, + 465 + ], + "type": "text", + "content": "3.1.2. Radar Preprocessing and Backbone" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 577 + ], + "type": "text", + "content": "The fundamental principle of radar-based physiological measurement involves capturing the subtle chest movement caused by cardiac activity. To extract displacement information from the radar intermediate frequency (IF) signal, we first perform range-FFT and coarse range gating to generate a range-time matrix and filter out irrelevant reflections. Afterwards, given that the 2D antenna array enables signal sampling in the spatial domain, we leverage beamforming to separate the reflections from different spatial positions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 580, + 500, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 580, + 500, + 594 + ], + "spans": [ + { + "bbox": [ + 372, + 580, + 500, + 594 + ], + "type": "interline_equation", + "content": "\\boldsymbol {B} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\boldsymbol {W} ^ {H} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {M}", + "image_path": "eaa7bde8e6ef40b41c8237d1ccc36b3a3050db78cd904568a0f57590d12976f3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 597, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 597, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 325, + 597, + 553, + 613 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\left[ \\dots , \\exp \\left(- j \\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) \\boldsymbol {d} _ {r} ^ {(i)}\\right), \\dots \\right] ^ {T} \\tag {2}", + "image_path": "592ffb48c9bced7473902c5fb35a801c7d8551f9bfcc6e5c06fb299a42a05568.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 614, + 536, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 536, + 637 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 536, + 637 + ], + "type": "interline_equation", + "content": "\\boldsymbol {k} \\left(\\theta_ {A}, \\theta_ {E}\\right) = \\frac {2 \\pi}{\\lambda} \\left[ \\cos \\theta_ {E} \\cos \\theta_ {A}, \\cos \\theta_ {E} \\sin \\theta_ {A}, \\sin \\theta_ {E} \\right]", + "image_path": "75338c96fcc12efa7998d4a23c8b712d4ed2696a491d2ce03abb11fe70bc8cf6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{N_C \\times R \\times T_2}" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " is the range-time matrix, with " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "N_C, R, T_2" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " indicating the number of virtual channels, range bins and frames, respectively. " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{W}(\\theta_A, \\theta_E) \\in \\mathbb{R}^{N_C}" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " is the steering vector in the direction of azimuth angle " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_A" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " and elevation angle " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta_E" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{d}_r^{(i)}" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " is the position of the i-th virtual channel. The shape of the beamformed output " + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 313, + 640, + 555, + 713 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8625" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 68, + 531, + 277 + ], + "blocks": [ + { + "bbox": [ + 81, + 68, + 531, + 277 + ], + "lines": [ + { + "bbox": [ + 81, + 68, + 531, + 277 + ], + "spans": [ + { + "bbox": [ + 81, + 68, + 531, + 277 + ], + "type": "image", + "image_path": "4622109d85be9d7d0bbc8cd42166a8a686f3a9afe126f93932a171eff27701d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. The overall framework of our model: Both backbones are based on UniFormer to facilitate cross-modal fusion. Vanilla adapters and SATM adapters are inserted into the shallow and deep layers of UniFormers, respectively, for multimodal fine-tuning. Each SATM adapter comprises two key components: Spatial Alignment and Temporal Matching. The Spatial Alignment module serves to align the spatial distributions of video and radar features, while the Temporal Matching module further refines the aligned features along the temporal dimension. Finally, the refined cross-modal features are passed through a vanilla adapter and added to the corresponding backbone as complementary features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\Theta_A \\times \\Theta_E) \\times R \\times T_2" + }, + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\Theta_A \\times \\Theta_E)" + }, + { + "bbox": [ + 55, + 369, + 295, + 418 + ], + "type": "text", + "content": " is the number of angular bins. Following [3, 43], we further utilize a second-order differentiator to amplify cardiac movements while suppressing noise signals as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 426, + 294, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 426, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 90, + 426, + 294, + 453 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {r, t} = \\left(\\psi_ {t - 3} + \\psi_ {t + 3}\\right) + 2 \\left(\\psi_ {t - 2} + \\psi_ {t + 2}\\right) \\tag {3} \\\\ - \\left(\\psi_ {t - 1} + \\psi_ {t + 1}\\right) - 4 \\psi_ {t} \\\\ \\end{array}", + "image_path": "ed7c27e6ca75af51adaf5321ecefbd5636c259e0190a0ee2ad782dfe1a5abed8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\psi_t" + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": " is the phase of " + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{B}" + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": " at the t-th frame. The output " + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "inline_equation", + "content": "F_r^{(0)} \\in \\mathbb{R}^{(\\Theta_A \\times \\Theta_E) \\times R \\times T_2}" + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": " is reshaped into " + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1 \\times S \\times T_2}" + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": " to serve as the input of the radar backbone, where " + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "inline_equation", + "content": "S = \\Theta_A \\times \\Theta_E \\times R" + }, + { + "bbox": [ + 55, + 460, + 295, + 510 + ], + "type": "text", + "content": " is the spatial dimension of the radar modality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 582 + ], + "type": "text", + "content": "Considering the necessity to model the mechanical motion of the human chest across both spatial and temporal dimensions [43], we adopt 2D UniFormer as our radar backbone. Overall, our video and radar backbones are symmetrically structured, which greatly facilitates cross-modal fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 151, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 151, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 151, + 601 + ], + "type": "text", + "content": "3.2. SATM Adapter" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "To enable effectively fine-tuning of the unimodal backbones and facilitate cross-modal modeling, we propose the SATM adapter. As illustrated in Fig. 2, Radar-to-Video SATM adapters and Video-to-Radar SATM adapters are inserted in parallel with the MHRA and FFN modules of the video and radar UniFormers respectively to extract complementary features from the other modality. As discussed in Sec. 1, the fundamental differences in measurement principles between video- and radar-based methods introduce" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 478 + ], + "type": "text", + "content": "a significant modality gap between their features and implicate cross-modal fusion. Considering the different ROIs of video and radar (face vs. chest), we present the Spatial Alignment (SA) module to align the spatial distributions of their features. As for the differences in captured signals (color changes vs. mechanical displacements), we design the Temporal Matching (TM) module to reduce waveform discrepancies and refine the spatially aligned features. Next, we will provide a detailed explanation of these modules." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 421, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 421, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 421, + 495 + ], + "type": "text", + "content": "3.2.1. Spatial Alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": "The physiological information contained in the face and chest is inherently distributed in distinct forms, resulting in different spatial distributions of video and radar features. We would like to emphasize that this dissimilarity is not merely reflected in the different spatial dimensions of " + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": "), and naively aligning their size through interpolation or linear projection is insufficient." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "To address this challenge, we design the Spatial Alignment module. Given the video feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and radar feature " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{r} \\in \\mathbb{R}^{C \\times s \\times t_{2}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", the core idea involves utilizing cross-attention to align the spatial distributions of their spatial tokens, i.e. " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{v} \\in \\mathbb{R}^{C \\times t_{1}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{r} \\in \\mathbb{R}^{C \\times t_{2}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Without loss of generality, we use the Radar-to-Video Spatial Alignment as an example for explanation, with its overall pipeline illustrated in Fig. 3 (a). First, bilinear interpolation is applied to align the temporal dimensions of " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". The primary challenge then lies in effectively transforming the 2D spatial tokens for the calculation of the attention" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8626" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 296, + 219 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "type": "image", + "image_path": "792d1c5ed9721d78fcf6a702d10073cd07d5ab30d9cc87a916cecadcf81003ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 227, + 296, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 271 + ], + "type": "text", + "content": "Figure 3. The pipelines of Radar-to-Video Spatial Alignment and Temporal Matching. \"MatMul\" and \"Hadamard\" denote matrix product and Hadamard product, respectively. " + }, + { + "bbox": [ + 55, + 227, + 296, + 271 + ], + "type": "inline_equation", + "content": "\\mathrm{FFT}_T" + }, + { + "bbox": [ + 55, + 227, + 296, + 271 + ], + "type": "text", + "content": " represents FFT along the temporal dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "content": "scores. Considering that the final predicted 1D PPG signals are generated by the integration of the spatial and channel dimensions, we propose treating spatial tokens as distinct candidates for PPG signal prediction at different spatial locations and then projecting them from the 2D space into a unified 1D time series space by compressing their channel dimensions. Specifically, we leverage linear projections to transform " + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "content": " into the query matrix " + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "inline_equation", + "content": "S_{v} \\in \\mathbb{R}^{hw\\times t_{1}}" + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "content": " and the key matrix " + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "inline_equation", + "content": "S_{r} \\in \\mathbb{R}^{s\\times t_{1}}" + }, + { + "bbox": [ + 54, + 292, + 296, + 435 + ], + "type": "text", + "content": ", respectively. Afterwards, the alignment between video and radar spatial tokens can be performed in the time series space and the resulting aligned radar features are computed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 443, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 443, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 96, + 443, + 295, + 471 + ], + "type": "interline_equation", + "content": "\\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {S _ {v} S _ {r} ^ {T}}{\\sqrt {t _ {1}}}\\right) F _ {r} \\tag {4}", + "image_path": "d2eb5b054b7fd7b68ea44a313ecf65629609c41373286411061143653087aae0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 501 + ], + "type": "text", + "content": "Note that the matrix multiplication between the attention score and " + }, + { + "bbox": [ + 55, + 477, + 295, + 501 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 55, + 477, + 295, + 501 + ], + "type": "text", + "content": " is performed channel-wise." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 501, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 501, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 54, + 501, + 295, + 657 + ], + "type": "text", + "content": "Now let us revisit the way to transform the spatial tokens. Two alternative approaches involve compressing the temporal dimension or flattening the channel and temporal dimensions of spatial tokens. However, the former variant risks irreversible information loss due to excessive compression of temporal signals, which is paramount for the prediction of PPG signals. The second approach, while straightforward, fails to account for the inherent differences between the channel and temporal dimensions. In contrast, our spatial alignment, which operates in the time series space, adheres to the paradigm of PPG signal regression and has the advantage of preserving temporal information. A quantitative comparison is provided in Sec. 4.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 662, + 169, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 169, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 169, + 674 + ], + "type": "text", + "content": "3.2.2. Temporal Matching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Although video and radar features are temporally synchronized and ideally encompass consistent frequency information, i.e. heart rate, the inherent disparity between wave" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "forms derived from skin color changes and chest mechanical movements introduces a significant modality gap in the temporal dimension. To this end, we further utilize Temporal Matching to refine the cross-modal features. According to the Matched Filter theory, given an input signal " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " composed of the target signal " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and the noise " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "n(t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", an output signal with the highest SNR at time step " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " could be obtained by applying the filter " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "h(t) = s(T - t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ". Inspired by this, we formulate the spatially aligned features as " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "r(t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and the waveform discrepancy as " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "n(t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", and the refined feature could be obtained through matched filtering. Specifically, given the video feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "F_{v} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and the spatially aligned radar feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "F_{r}^{\\prime} \\in \\mathbb{R}^{C \\times t_{1} \\times h \\times w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", the refined radar-to-video feature can be computed by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 244, + 553, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 269 + ], + "type": "interline_equation", + "content": "\\mathrm {T M} _ {r 2 v} \\left(F _ {v}, F _ {r} ^ {\\prime}\\right) = \\mathrm {D W c o n v} \\left(\\mathrm {B N} \\left(\\mathcal {F} _ {T} ^ {- 1} \\left(\\overline {{\\mathcal {F} _ {T} \\left(F _ {v}\\right)}} \\mathcal {F} _ {T} \\left(F _ {r} ^ {\\prime}\\right)\\right)\\right) \\right. \\tag {5}", + "image_path": "a59e5250d8060046339658854bb621af3a7a1c02682d357e67c6071dc2160b9b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_T(\\cdot)" + }, + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "text", + "content": " denotes Fast Fourier Transform (FFT) along the temporal dimension and " + }, + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{F}_T(\\cdot)}" + }, + { + "bbox": [ + 313, + 270, + 553, + 318 + ], + "type": "text", + "content": " represents FFT followed by the conjugate operation. DWConv and BN are depthwise convolution and batch normalization, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 322, + 429, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 429, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 429, + 333 + ], + "type": "text", + "content": "3.2.3. Cross-Modal Fusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "content": "Overall, the forward process of the Radar-to-Video SATM adapter can be represented as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 366, + 518, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 366, + 518, + 380 + ], + "spans": [ + { + "bbox": [ + 370, + 366, + 518, + 380 + ], + "type": "interline_equation", + "content": "F _ {r 2 v} = \\mathrm {T M} _ {r 2 v} \\left(F _ {v}, \\mathrm {S A} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right)\\right)", + "image_path": "cb83c0d03448fdb225a703a107dff6b9fa2002f9f5d1be5dd1a09147f0650e8b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 381, + 558, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 558, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 558, + 405 + ], + "type": "interline_equation", + "content": "\\operatorname {S A T M} _ {r 2 v} \\left(F _ {v}, F _ {r}\\right) = W _ {u p} \\left(\\operatorname {R e L U} \\left(W _ {\\text {d o w n}} \\left(F _ {v} + \\alpha \\cdot F _ {r 2 v}\\right)\\right)\\right) \\tag {6}", + "image_path": "e7213ced2e0ed781e36418d221d1b946c1e70c25dcf56dd768bdb715b7ccc57d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "W_{down}, W_{up}" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": " are the down-projection and up-projection layers of the vanilla adapter [12] and " + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 405, + 555, + 453 + ], + "type": "text", + "content": " is a learnable scaling factor. Afterward, the forward process of a video UniFormer block can be reformulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 455, + 543, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 455, + 543, + 471 + ], + "spans": [ + { + "bbox": [ + 324, + 455, + 543, + 471 + ], + "type": "interline_equation", + "content": "Z _ {v} ^ {(l)} = \\operatorname {M H R A} \\left(Y _ {v} ^ {(l)}\\right) + Y _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Y _ {v} ^ {(l)}, Y _ {r} ^ {(l)}\\right)", + "image_path": "0c8206e95dac9b62558708c7d4ac99ff5d8606772f7e67fc8078831041ece454.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 324, + 472, + 553, + 498 + ], + "type": "interline_equation", + "content": "F _ {v} ^ {(l)} = \\operatorname {F F N} \\left(Z _ {v} ^ {(l)}\\right) + Z _ {v} ^ {(l)} + \\operatorname {S A T M} _ {r 2 v} \\left(Z _ {v} ^ {(l)}, Z _ {r} ^ {(l)}\\right) \\tag {7}", + "image_path": "2dba21c2f1eb65968337753425414803d02920c48aedd03f37b05343eb0e4e93.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 533 + ], + "type": "text", + "content": "The forward processes of the Video-to-Radar SATM adapter and the fine-tuned radar UniFormer blocks can also be computed similarly to Eq. (6) and Eq. (7)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In summary, Spatial Alignment aligns the spatial distributions of video and radar features in a unified time series space, while Temporal Matching utilizes a matched filter to handle the waveform difference between their signals. They complement each other by processing the cross-modal features along spatial and temporal dimensions separately, thereby effectively narrowing the substantial modality gap between video and radar features. According to [9], it is essential to balance the learning of unimodal and cross-modal features. Consequently, we only insert our SATM adapters into the last two stages of UniFormers while utilizing vanilla adapters in the first two stages. The output features of both backbones are integrated by a lightweight fusion stem and subsequently passed through a regression module to generate the predicted PPG signal." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8627" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 136, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 115, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 115, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 115, + 102 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "type": "text", + "content": "We utilize the EquiPleth [41] dataset to evaluate our method, which contains 533 30-second samples collected from 28 light, 49 medium, and 14 dark skin tone volunteers. All samples are recorded under motionless conditions and with sufficient illumination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 168, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 168, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 168, + 295, + 252 + ], + "type": "text", + "content": "In addition, considering that the EquiPleth dataset mainly focuses on constrained scenarios and shows limited practical diversity, we built a new dataset called MMRPM to promote a more comprehensive evaluation of our method. MMRPM comprises 180 60-second samples from 31 subjects, all of which are recorded under varying illumination conditions and subject motions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 252, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 373 + ], + "type": "text", + "content": "Specifically, MMRPM encompasses three illumination conditions: dark, medium, and light. The \"dark\" and \"medium\" settings correspond to illumination levels of approximately 2.5 and 6.3 lux respectively, while \"light\" represents samples are collected under sufficient sunlight and indoor illumination. Furthermore, two types of subject activities are considered: sitting still and free head movement. Overall, MMRPM covers six scenarios, where low-light conditions are challenging to the video modality and head motions have a greater impact on the radar modality." + }, + { + "bbox": [ + 55, + 252, + 295, + 373 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 188, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 188, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 188, + 392 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "text", + "content": "Our proposed method is implemented using PyTorch. For video preprocessing, we first employ MTCNN to crop the face region in the initial frame and fix the region across subsequent frames. Then, we sample each video into clips with a time window of 160 frames and a step size of 80 frames, which are then resized to " + }, + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "inline_equation", + "content": "96 \\times 96" + }, + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "text", + "content": " pixels. In terms of radar preprocessing, the angular range selected for beamforming is " + }, + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "inline_equation", + "content": "66^{\\circ} \\leq \\Theta_{A} \\leq 114^{\\circ}, -24^{\\circ} \\leq \\Theta_{E} \\leq 24^{\\circ}" + }, + { + "bbox": [ + 55, + 397, + 295, + 493 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 493, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 566 + ], + "type": "text", + "content": "We utilize UniFormer-S [18] and MAM in MAGIC [55] as our feature extractor and fusion stem respectively. As for the regression module, we follow the design in ADDP [19]. The video backbone is jointly pre-trained on VIPL [31, 32], MMPD [40], UBFC [1], PURE [39] and BUAA [47], and we use the dataset introduced in [53] for radar pre-trianing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 566, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 685 + ], + "type": "text", + "content": "During multimodal fine-tuning, both backbones are frozen, and only the adapters, fusion stem and regression module are trained. We employ the loss functions introduced in [51] and utilize the Adam optimizer for training, with the learning rate and weight decay set to 5e-4 and 5e-5 respectively. The model is trained for 40 epochs with a batch size of 2. Following [38], we apply random horizontal flipping, spatially resized crop and random intensity noise to augment the input videos. For radar augmentation, we randomly shift the range bins of the range-time matrices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "Following [41], we perform our evaluation of the PPG signals over 300 samples windows with a stride of 128 samples. The heart rate corresponding to a specific signal clip is calculated using a combination of band-pass filtering, detrending and peak detection in the frequency domain. Finally, the standard deviation of the error (Std), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (R) are adopted to evaluate the predicted HR error. The unit of Std, MAE and RMSE is beats per minute (bpm). We split the MMRPM dataset into 5 subject-exclusive folds and use the official split for the EquiPleth dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 226, + 399, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 399, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 399, + 238 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 244, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 544 + ], + "type": "text", + "content": "We first compare our SATM with existing video-radar RPM methods [6, 41, 46] and several state-of-the-art multimodal adapter methods [2, 22, 28, 48] on the EquiPleth [41] dataset. To ensure a fair comparison, we provide extra results of improved versions of existing RPM methods by applying unimodal pre-training and replacing their original backbones with UniFormers. Since the radar used in the EquiPleth dataset activates only 1 transmitter and 1 receiver, we do not perform beamforming during radar preprocessing for this dataset. For existing video-radar RPM methods, we use the official implementation of EquiPleth, and re-implement Fusion-Vital and CardiacMamba, as the codes have not been released by the authors. As shown in Tab. 1, all of these methods exhibit RMSEs exceeding 2.9 bpm when trained from scratch. Their performance consistently improves when unimodal pre-training is applied, underscoring the significance of unimodal pre-training when video-radar data is limited. However, they still display MAEs greater than 0.64 bpm, regardless of whether using their original backbones or UniFormers, indicating that their full fine-tuning-based fusion strategies are suboptimal for the small-scale multimodal RPM dataset. In contrast, our method employs SATM adapters for multimodal fine-tuning, effectively leveraging pre-trained knowledge and enabling superior cross-modal fusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "content": "Furthermore, our SATM also demonstrates superior performance compared to existing multimodal adapter methods. The subpar performance of UniAdapter and MMA suggests that their parameter-sharing strategies are ill-suited for multimodal RPM, as they impede the learning of modality-specific features. To adapt BAT for video-radar RPM, we employ an additional linear projection layer to align the token numbers of the video and radar branches. However, it fails to handle the significant modality gap between video and radar, thereby resulting in an RMSE exceeding 1.65 bpm. While the best baseline, LAVISH, achieves an RMSE of 1.498 bpm, our SATM yields a considerably lower RMSE of 1.242 bpm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 712 + ], + "type": "text", + "content": "We have also evaluated SATM on our self-collected" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "For more details about our MMRPM dataset and pre-training, please refer to the supplemental materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8628" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 529, + 271 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 529, + 271 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 529, + 271 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 529, + 271 + ], + "type": "table", + "html": "
MethodsOriginal ModalityUnimodal Pre-trainedStd↓MAE↓RMSE↓R↑Trainable Params (M)↓
EquiPleth[41]Video-Radar×3.2741.0633.2830.9573.95
Fusion-Vital[6]Video-Radar×9.2883.7509.2970.709715.26
CardiacMamba[46]Video-Radar×2.9240.9012.9270.9759.93
EquiPlethVideo-Radar√2.9790.9682.9900.9653.95
Fusion-VitalVideo-Radar√9.1213.7509.1300.735715.26
CardiacMambaVideo-Radar√2.7970.8682.7710.9759.93
EquiPleth*Video-Radar√1.4350.6711.4360.99245.31
Fusion-Vital*Video-Radar√2.1430.6502.1570.983194.61
CardiacMamba*Video-Radar√2.3130.7442.3240.98033.62
UniAdapter[28]Vision-Language√1.6530.5901.6560.99010.60
MMA[48]Vision-Language√1.8350.6261.8420.98811.09
BAT[2]RGB-TIR√1.6790.6301.6840.99085.35
LAVISH[22]Audio-Visual√1.4930.5571.4980.99110.85
SATM (Ours)Video-Radar√1.2170.5401.2420.99410.78
", + "image_path": "9eaa949fdcc92ac672026c95905a0c3993b9849deaf0108bec73e6d05c53f49b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 331, + 293, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 313 + ], + "type": "text", + "content": "Table 1. Overall performance on the EquiPleth dataset. The best and second best results are in bold and underlined respectively. * denotes that we replace the methods' original backbones with UniFormers while preserving their fusion strategies. Blue and red indicate the existing video-radar RPM methods and existing adapter methods for other multimodal tasks, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 331, + 293, + 485 + ], + "lines": [ + { + "bbox": [ + 59, + 331, + 293, + 485 + ], + "spans": [ + { + "bbox": [ + 59, + 331, + 293, + 485 + ], + "type": "table", + "html": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth7.6032.4217.6030.715
Fusion-Vital9.1395.1659.3150.501
CardiacMamba6.0992.1606.1070.790
EquiPleth*5.9641.7395.9740.806
Fusion-Vital*5.5501.6625.5890.827
CardiacMamba*5.5261.8435.5640.829
UniAdapter5.5091.5885.5100.831
MMA5.6941.7225.6940.819
BAT6.4322.1516.4330.760
LAVISH5.6601.4545.3120.844
SATM (Ours)5.0631.2925.0740.858
", + "image_path": "236595545519b7861ca36304020c4239d0b1469fe6c9f66cc14bb78548035d1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 603 + ], + "type": "text", + "content": "MMRPM dataset, which covers more challenging and diverse scenarios than the EquiPleth dataset. As illustrated in Tab. 2, existing video-radar RPM methods exhibit poor results on this dataset, highlighting their limited capability to handle complex scenarios. In comparison, our SATM consistently outperforms the baseline methods in all metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "We further conduct a cross-dataset evaluation by training the models on the EquiPleth dataset and evaluating them on the MMRPM dataset. As shown in Tab. 3, the improved versions of existing multimodal RPM methods exhibit subpar performance compared to adapter-based methods. This is primarily because they rely on full fine-tuning-based fusion strategies and are prone to overfitting small-scale training sets. In contrast, our method, built upon SATM adapters, achieves the best generalization performance and suffers" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 331, + 552, + 414 + ], + "blocks": [ + { + "bbox": [ + 75, + 493, + 275, + 504 + ], + "lines": [ + { + "bbox": [ + 75, + 493, + 275, + 504 + ], + "spans": [ + { + "bbox": [ + 75, + 493, + 275, + 504 + ], + "type": "text", + "content": "Table 2. Overall performance on the MMRPM dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 331, + 552, + 414 + ], + "lines": [ + { + "bbox": [ + 318, + 331, + 552, + 414 + ], + "spans": [ + { + "bbox": [ + 318, + 331, + 552, + 414 + ], + "type": "table", + "html": "
MethodsStd↓MAE↓RMSE↓R↑
EquiPleth*8.7874.2569.0820.629
Fusion-Vital*8.5384.4578.8970.642
CardiacMamba*9.0485.0389.3080.603
LAVISH8.0263.6878.2320.712
SATM (Ours)7.5423.4237.6560.738
", + "image_path": "169b3ea8b28cd7e7ecf73632a2b9298223cfaced9ce02c2abd757a8175c31a73.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 421, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 444 + ], + "type": "text", + "content": "Table 3. Results of the cross-dataset evaluation from the EquiPleth dataset to the MMRPM dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 465, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 524 + ], + "type": "text", + "content": "minimal performance degradation compared to other baselines. Although it still lags behind the performance in the intra-dataset evaluation, we argue that this is reasonable, given the significant differences between these two datasets.[2]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 409, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 409, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 409, + 546 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 551, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 555, + 696 + ], + "type": "text", + "content": "Key Components. Our SATM adapter mainly consists of two key components: Spatial Alignment and Temporal Matching. In Tab. 4, we conduct an ablation study on the MMRPM dataset to assess the effectiveness of these modules. First, compared to the vanilla adapter (the first row), the adapter with Spatial Alignment reduces the MAE and RMSE by 0.087 bpm and 0.149 bpm respectively. This can be attributed to its capability to facilitate modality complementarity by extracting spatially aligned cross-modal features to the corresponding backbones. The MAE and RMSE can be further reduced to 1.292 bpm and 5.074 bpm when Temporal Matching is inserted, highlighting the necessity" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 539, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 539, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 539, + 714 + ], + "type": "text", + "content": "2Details about cross-dataset evaluation are provided the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8629" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 278, + 126 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 278, + 126 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 278, + 126 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 278, + 126 + ], + "type": "table", + "html": "
SATMStd↓MAE↓RMSE↓R↑
××5.4521.5415.4530.836
✓×5.2981.4545.3040.845
✓✓5.0631.2925.0740.858
", + "image_path": "fe2bf6e8af76ae772070089209b1bf1bbc0dc7708066e34d8d1250731256ca0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 169, + 281, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 157 + ], + "type": "text", + "content": "Table 4. Ablation study on the key components. \"SA\" and \"TM\" denote Spatial Alignment and Temporal Matching respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 169, + 281, + 251 + ], + "lines": [ + { + "bbox": [ + 70, + 169, + 281, + 251 + ], + "spans": [ + { + "bbox": [ + 70, + 169, + 281, + 251 + ], + "type": "table", + "html": "
VariantsStd↓MAE↓RMSE↓R↑
Interpolate6.0411.8766.0520.795
Linear5.5791.6285.5820.828
Channel6.0871.8076.0910.792
Flatten5.5881.6935.5950826
Ours5.2981.4545.3040.845
", + "image_path": "91f6c66af950c80a533335e6b8c8954e745fdac5dd366c806f91b72f6e97de16.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 304 + ], + "type": "text", + "content": "to refine the aligned features along the temporal dimension." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 305, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 461 + ], + "type": "text", + "content": "Alternative Approaches for Spatial Alignment. We also investigate four alternative approaches for Spatial Alignment. Specifically, \"Interpolate\" and \"Linear\" refer to methods that directly align the sizes of video and radar features through interpolation and linear projection instead of cross-attention. The \"Channel\" and \"Flatten\" indicate that the alignment is performed in the channel space and on the flattened spatial tokens respectively. As shown in Tab. 5, both \"Interpolate\" and \"Linear\" exhibit suboptimal performance, achieving MAEs of 1.876 bpm and 1.628 bpm. These methods only align spatial sizes while neglecting inherent differences in spatial distributions between video and radar features, thereby failing to address the modality gap." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 617 + ], + "type": "text", + "content": "The \"Channel\" variant compresses the temporal dimensions of spatial tokens for the computation of the attention scores, focusing solely on the overall temporal period while neglecting fine-grained temporal variability. This results in degraded performance with an RMSE exceeding 6 bpm. The \"Flatten\" variant preserves temporal information, but conflates temporal and channel dimensions, overlooking their intrinsic differences. Therefore, it ultimately yields subpar performance with an MAE of 1.693 bpm and an RMSE of 5.595 bpm. In contrast, Spatial Alignment performed in the time series space effectively avoids temporal information loss and dimensional distortion, showing superior performance compared to these four variants." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Temporal Matching vs. Temporal Attention. Since attention is a straightforward method for cross-modal fusion, we compare our Temporal Matching (TM) with Temporal Attention (TA). Specifically, we replace the matched filtering operation with a cross-attention along the temporal dimension. As shown in Tab. 6, our TM consistently outperforms TA across all metrics. Moreover, since the time complexity of FFT is " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N\\log N)" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": ", TM is also more com" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 538, + 115 + ], + "blocks": [ + { + "bbox": [ + 67, + 258, + 284, + 270 + ], + "lines": [ + { + "bbox": [ + 67, + 258, + 284, + 270 + ], + "spans": [ + { + "bbox": [ + 67, + 258, + 284, + 270 + ], + "type": "text", + "content": "Table 5. Ablation study on the design of Spatial Alignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 70, + 538, + 115 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 538, + 115 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 538, + 115 + ], + "type": "table", + "html": "
TemporalStd↓MAE↓RMSE↓R↑
Attention5.1551.3675.1300.853
Matching5.0631.2925.0740.858
", + "image_path": "e2953e1bd1bfddaffef3be02d3511cafdadb7b0bbbc702191ca5c199621c3933.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 333, + 156, + 537, + 224 + ], + "blocks": [ + { + "bbox": [ + 314, + 123, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 314, + 123, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 123, + 553, + 144 + ], + "type": "text", + "content": "Table 6. Comparison between Temporal Matching and Temporal Attention." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 156, + 537, + 224 + ], + "lines": [ + { + "bbox": [ + 333, + 156, + 537, + 224 + ], + "spans": [ + { + "bbox": [ + 333, + 156, + 537, + 224 + ], + "type": "table", + "html": "
PositionsStd↓MAE↓RMSE↓R↑
SATM-45.1681.3865.1790.852
SATM-35.1071.3435.1090.857
SATM-25.0631.2925.0740.858
SATM-15.5261.6145.5260.830
", + "image_path": "600e2dfc610c370a073f7aa51e1a993c04ae9a037cbccadc30188fa239c7fba3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 232, + 547, + 244 + ], + "lines": [ + { + "bbox": [ + 320, + 232, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 547, + 244 + ], + "type": "text", + "content": "Table 7. Ablation study on the position to add SATM adapters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "text", + "content": "putationally efficient compared to the quadratic complexity of TA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 456 + ], + "type": "text", + "content": "Position to Insert Our SATM Adapters. We also explore the impact of different positions to insert our SATM adapters. We present 4 positions for insertion, formulated as \"SATM-n\", where \"n\" denotes that we use SATM adapters in the last \"n\" stages of the UniFormers and employ vanilla adapters in earlier stages. As reported in Tab. 7, SATM-1 displays an MAE higher than the naive baseline (1.614 bpm vs. 1.541 bpm), indicating that inserting the cross-modal module too late would hinder performance due to the lack of cross-modal interaction. Meanwhile, the performance of SATM-3 and SATM-4 is also limited due to insufficient modality-specific modeling. In contrast, the results demonstrate that SATM-2 strikes a good trade-off between the learning of unimodal and corss-modal features." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 641 + ], + "type": "text", + "content": "In this paper, we focus on video-radar RPM measurement. To overcome the data scarcity challenge, we make full use of unimodal pre-training and present the SATM adapter to effectively fine-tune unimodal backbones into a multimodal RPM model. The SATM adapter leverages Spatial Alignment and Temporal Matching to collaboratively narrow the modality gap between video and radar and effectively extract cross-modal features for both backbones. To thoroughly evaluate our method, we collect a challenging dataset that encompasses diverse illumination conditions and subject motions. Extensive experiments conducted on our self-collected and existing public datasets demonstrate that SATM outperforms existing methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Limitations. While SATM effectively constructs a robust video-radar RPM model, it primarily focuses on cross-modal interaction and still relies on vanilla adapters for unimodal modeling during fine-tuning. In the future, we hope to develop a unified framework that harmonizes cross-modal and unimodal modeling." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8630" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 114 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62172381." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "text", + "content": "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 296, + 240 + ], + "type": "text", + "content": "[2] Bing Cao, Junliang Guo, Pengfei Zhu, and Qinghua Hu. Bidirectional adapter for multimodal tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 927-935, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 241, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 285 + ], + "type": "text", + "content": "[3] Jinbo Chen, Dongheng Zhang, Zhi Wu, Fang Zhou, Qibin Sun, and Yan Chen. Contactless electrocardiogram monitoring with millimeter wave radar. IEEE Transactions on Mobile Computing, 23(1):270-285, 2022. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 329 + ], + "type": "text", + "content": "[4] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 349-365, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 296, + 384 + ], + "type": "text", + "content": "[5] Li-Wen Chiu, Yang-Ren Chou, Yi-Chiao Wu, and Bing-Fei Wu. Deep-learning based remote photoplethysmography measurement in driving scenarios with color and near-infrared images. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "text", + "content": "[6] Jae-Ho Choi, Ki-Bong Kang, and Kyung-Tae Kim. Fusion-vital: Video-rf fusion transformer for advanced remote physiological measurement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1344-1352, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 296, + 472 + ], + "type": "text", + "content": "[7] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE transactions on biomedical engineering, 60(10):2878-2886, 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 472, + 296, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 505 + ], + "type": "text", + "content": "[8] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiological measurement, 35(9):1913, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 559 + ], + "type": "text", + "content": "[9] Chenzhuang Du, Jiaye Teng, Tingle Li, Yichen Liu, Tianyuan Yuan, Yue Wang, Yang Yuan, and Hang Zhao. On uni-modal feature learning in supervised multi-modal learning. In International Conference on Machine Learning, pages 8632-8656. PMLR, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 296, + 614 + ], + "type": "text", + "content": "[10] Jingda Du, Si-Qi Liu, Bochao Zhang, and Pong C Yuen. Dual-bridging with adversarial noise generation for domain adaptive rpgg estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10355–10364, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "text", + "content": "[11] Unsoo Ha, Salah Assana, and Fadel Adib. Contactless seismocardiography via deep learning radars. In Proceedings of the 26th annual international conference on mobile computing and networking, pages 1-14, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 714 + ], + "type": "text", + "content": "[12] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[14] Xiaojie Jin, Bowen Zhang, Weibo Gong, Kai Xu, XueQing Deng, Peng Wang, Zhao Zhang, Xiaohui Shen, and Jiashi Feng. Mv-adapter: Multimodal video transfer learning for video text retrieval. arXiv preprint arXiv:2301.07868, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 172, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 205 + ], + "type": "text", + "content": "[15] Jitesh Joshi and Youngjun Cho. ibvp dataset: Rgb-thermal rppg dataset with high resolution signal quality labels. Electronics, 13(7):1334, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "text", + "content": "[16] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[17] Youngseok Kim, Juyeb Shin, Sanmin Kim, In-Jae Lee, Jun Won Choi, and Dongsuk Kum. Crn: Camera radar net for accurate, robust, efficient 3d perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17615-17626, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "text", + "content": "[18] Kunchang Li, Yali Wang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unified transformer for efficient spatiotemporal representation learning. arXiv preprint arXiv:2201.04676, 2022. 3, 6, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 393 + ], + "type": "text", + "content": "[19] Qian Liang, Yan Chen, and Yang Hu. Continual learning for remote physiological measurement: Minimize forgetting and simplify inference. In European conference on computer vision, pages 126-144. Springer, 2024. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[20] Zhu Lianghui, Liao Bencheng, Zhang Qian, Wang Xinlong, Liu Wenyu, and Wang Xinggang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv. Org, pages arXiv-org, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[21] Juan-Ting Lin, Dengxin Dai, and Luc Van Gool. Depth estimation from monocular images and sparse radar data. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10233-10240. IEEE, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "text", + "content": "[22] Yan-Bo Lin, Yi-Lin Sung, Jie Lei, Mohit Bansal, and Gedas Bertasius. Vision transformers are parameter-efficient audiovisual learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2299–2309, 2023. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 614 + ], + "type": "text", + "content": "[23] Zhiwei Lin, Zhe Liu, Zhongyu Xia, Xinhao Wang, Yongtao Wang, Shengxiang Qi, Yang Dong, Nan Dong, Le Zhang, and Ce Zhu. Rcbevdet: Radar-camera fusion in bird's eye view for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14928-14937, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[24] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. Advances in Neural Information Processing Systems, 33:19400-19411, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[25] Chen-Chou Lo and Patrick Vandewalle. Depth estimation from monocular images and sparse radar using deep ordinal regression network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3343-3347. IEEE, 2021. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8631" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Yunfei Long, Abhinav Kumar, Daniel Morris, Xiaoming Liu, Marcos Castro, and Punarjay Chakravarty. Radiant: Radar-image association network for 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1808-1816, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[27] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12404-12413, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[28] Haoyu Lu, Yuqi Huo, Guoxing Yang, Zhiwu Lu, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling. arXiv preprint arXiv:2302.06605, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[29] Hao Lu, Zitong Yu, Xuesong Niu, and Ying-Cong Chen. Neuron structure modeling for generalizable remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18589-18599, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[30] Ramin Nabati and Hairong Qi. Centerfusion: Center-based radar and camera fusion for 3d object detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 387 + ], + "type": "text", + "content": "[31] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Vipl-hr: A multi-modal database for pulse estimation from less-constrained face video. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part V 14, pages 562–576. Springer, 2019. 2, 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[32] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2019. 2, 6, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 498 + ], + "type": "text", + "content": "[33] Xuesong Niu, Xingyuan Zhao, Hu Han, Abhijit Das, Antitza Dantcheva, Shiguang Shan, and Xin Chen. Robust remote heart rate estimation from face utilizing spatial-temporal attention. In 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019), pages 1-8. IEEE, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[34] Soyeon Park, Bo-Kyeong Kim, and Suh-Yeon Dong. Self-supervised rgb-nir fusion video vision transformer framework for rppg estimation. IEEE Transactions on Instrumentation and Measurement, 71:1-10, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "text", + "content": "[35] Vladimir L Petrović, Milica M Janković, Anita V Lupšić, Veljko R Mihajlović, and Jelena S Popovic-Božović. High-accuracy real-time monitoring of heart rate variability using 24 ghz continuous-wave doppler radar. *IEEE Access*, 7: 74721–74733, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 645 + ], + "type": "text", + "content": "[36] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics express, 18(10):10762-10774, 2010. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[37] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[38] Jeremy Speth, Nathan Vance, Patrick Flynn, and Adam Czajka. Non-contrastive unsupervised learning of physiological signals from video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14464–14474, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[39] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In The 23rd IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062. IEEE, 2014. 6, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "type": "text", + "content": "[40] Jiankai Tang, Kequan Chen, Yuntao Wang, Yuanchun Shi, Shwetak Patel, Daniel McDuff, and Xin Liu. Mmpd: multidomain mobile video physiology dataset. In 2023 45th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 1-5. IEEE, 2023. 2, 6, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "type": "text", + "content": "[41] Alexander Vilesov, Pradyumna Chari, Adnan Armouti, Anirudh Bindiganavale Harish, Kimaya Kulkarni, Ananya Deoghare, Laleh Jalilian, and Achuta Kadambi. Blending camera and 77 ghz radar sensing for equitable, robust plethysmography. ACM Trans. Graph., 41(4):36-1, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[42] Fengyu Wang, Xiaolu Zeng, Chenshu Wu, Beibei Wang, and KJ Ray Liu. mmhrv: Contactless heart rate variability monitoring using millimeter-wave radio. IEEE Internet of Things Journal, 8(22):16623-16636, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "type": "text", + "content": "[43] Haoyu Wang, Jinbo Chen, Dongheng Zhang, Zhi Lu, Changwei Wu, Yang Hu, Qibin Sun, and Yan Chen. Contactless radar heart rate variability monitoring via deep spatiotemporal modeling. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 111-115. IEEE, 2024. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[44] Kai Wang, Yapeng Tian, and Dimitrios Hatzinakos. Towards efficient audio-visual learners via empowering pre-trained vision transformers with cross-modal adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1837–1846, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 482, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 523 + ], + "type": "text", + "content": "[45] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE transactions on Biomedical Engineering, 62(2): 415-425, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "text", + "content": "[46] Zheng Wu, Yiping Xie, Bo Zhao, Jiguang He, Fei Luo, Ning Deng, and Zitong Yu. Cardiacmamba: A multimodal rgb-rf fusion framework with state space models for remote physiological measurement. arXiv preprint arXiv:2502.13624, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "text", + "content": "[47] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pages 1-7. IEEE, 2020. 6, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[48] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[49] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from fa" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8632" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 397 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "cial videos using spatio-temporal networks. arXiv preprint arXiv:1905.02419, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[50] Zitong Yu, Xiaobai Li, Xuesong Niu, Jingang Shi, and Guoying Zhao. Autohr: A strong end-to-end baseline for remote heart rate measurement with neural searching. IEEE Signal Processing Letters, 27:1245-1249, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 206 + ], + "type": "text", + "content": "[51] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip HS Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4186-4196, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[52] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[53] Bin-Bin Zhang, Dongheng Zhang, Yadong Li, Zhi Lu, Jinbo Chen, Haoyu Wang, Fang Zhou, Yu Pu, Yang Hu, Li-Kun Ma, et al. Monitoring long-term cardiac activity with contactless radio frequency signals. Nature Communications, 15(1):1-11, 2024. 2, 6, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 342 + ], + "type": "text", + "content": "[54] Zijun Zhang. Improved adam optimizer for deep neural networks. In 2018 IEEE/ACM 26th international symposium on quality of service (IWQoS), pages 1-2. IEEE, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[55] Xu Zheng, Yuanhuiyi Lyu, Jiazhou Zhou, and Lin Wang. Centering the value of every modality: Towards efficient and resilient modality-agnostic semantic segmentation. In European Conference on Computer Vision, pages 192–212. Springer, 2024. 6" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8633" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_content_list.json b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce4aac52f1fe7718bee1e6c411b9408f33c41427 --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Spatial Preference Rewarding for MLLMs Spatial Understanding", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Han Qiu $^{1}$ , Peng Gao $^{2}$ , Lewei Lu $^{3}$ , Xiaqin Zhang $^{4}$ , Ling Shao $^{5}$ , Shijian Lu $^{1*}$", + "bbox": [ + 197, + 179, + 805, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ Shanghai AI Laboratory $^{3}$ Sensetime Research, $^{4}$ Zhejiang University of Technology", + "bbox": [ + 222, + 214, + 776, + 248 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ UCAS-Terminus AI Lab, University of Chinese Academy of Sciences", + "bbox": [ + 220, + 250, + 774, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "han023@e.ntu.edu.sg, Shijian.Lu@ntu.edu.sg", + "bbox": [ + 307, + 272, + 683, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding (SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR.", + "bbox": [ + 88, + 353, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 821, + 220, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) [3, 15, 18, 31, 32, 51, 58, 70] have achieved remarkable success by inte-", + "bbox": [ + 89, + 845, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "grating pretrained large language model [2, 14, 49] with vision encoders [8, 37, 42], leading to significant advancements in a wide range of general vision-language tasks. By combining visual and language signals, MLLMs have demonstrated superior capabilities in multimodal understanding, reasoning, and interaction as compared with traditional vision models. Recently, several studies have further injected spatial knowledge into MLLMs, thereby improving MLLMs' fine-grained perception of visual inputs and enabling tasks such as referential dialogue [10, 16], grounding captioning [35, 58, 62], region description [31, 52], and object detection [61], etc. These advances have paved the way for MLLMs to serve as versatile visual assistants supporting a wider range of applications.", + "bbox": [ + 511, + 323, + 906, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advancements, MLLMs still face challenges in fine-grained spatial understanding, with responses not aligned with human preferences. As illustrated in 1, the generated grounded region descriptions are often vague with inaccurate object localizations, and models may fail to focus on the queried region, distracted from other regions in the image. The issue in spatial understanding could be attributed to the lack of positive-negative preference feedback in existing instruction-tuned MLLMs. Specifically, instruction fine-tuning (SFT) directly optimizes MLLMs to mimic ground truth positive samples, but it cannot impose any penalties if the model produces inaccurate negative samples for localization during actual inference. As a result, MLLMs may struggle to generate positive descriptions with accurate object localization and instead produce negative and inaccurate descriptions, leading to responses that do not align with user expectations. In addition, optimization using positive and negative samples has been proven crucial for spatial understanding in traditional object detection algorithms [7, 29, 45], highlighting a significant gap in the current MLLM training on spatial understanding.", + "bbox": [ + 511, + 536, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several studies [38, 48, 50, 59, 65, 69] attempt to introduce preference optimization for better MLLM alignment, where the preference data are constructed by collect", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "720", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ecce25348467d0b662edc637ca4686592b5c5f5e4c1881e8c418ea9b917eb06.jpg", + "image_caption": [ + "User Prompt about Regions in Images:", + "MLLM Response:" + ], + "image_footnote": [], + "bbox": [ + 135, + 104, + 308, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object.", + "bbox": [ + 316, + 127, + 483, + 171 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ab0ab80098fa3b1f68e0bff559b8f3ead7d3acf9823acc9f234ed7153636732b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 103, + 671, + 191 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object.", + "bbox": [ + 678, + 128, + 844, + 171 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The area [676, 229, 847, 1000] is occupied by a woman in a blue shirt who is holding a whistle [537, 530, 568, 606]. She is also wearing a black clipboard [801, 538, 825, 581] around her neck ....", + "bbox": [ + 143, + 229, + 321, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fail to Ground", + "bbox": [ + 248, + 285, + 310, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cbb76db1ef6c4cb5e22b0e979ca279eb75aa5bc2962748530c13bf7e0a1d3843.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 209, + 496, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The region [638, 157, 775, 709] is a trail located on a mountain [166, 365, 996, 996]. of people [619, 250, 799, 990] walking on it. The trail also has a marker [166, 339, 229, 439] to guide hikers. Inaccurate Localization", + "bbox": [ + 504, + 228, + 678, + 292 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/69c00219318dcdb064f0d69d1853b96aa7de07ffad9584a65699ac43e62540df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 209, + 859, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tuning with Spatial Preference Rewarding (SPR)", + "bbox": [ + 377, + 305, + 651, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The area [676, 229, 847, 998] in the image features a woman in a blue shirt with a name tag, .... She [676, 229, 817, 901] appears to .... The grassy area where she is standing [0, 385, 998, 998] seems to be .... activities. There's also a red cone [676, 909, 720, 998] ....", + "bbox": [ + 140, + 340, + 318, + 424 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5310376b17b7b49b037e7ebc730e6eaecf0df6b387ac830d8c7570c712df2ee1.jpg", + "image_caption": [ + "Responses from SPR-empowered MLLM:", + "Figure 1. The proposed Spatial Preference Rewarding (SPR) mitigates the distracted and inaccurate region descriptions generated by MLLMs. Given an image and a user-specified region of interest, MLLMs often fail to focus on the queried region. They may be distracted by objects outside the specified region, failing to ground the queried objects, or providing inaccurate localization. Tuning MLLMs with our proposed SPR leads to more accurate object localization and detailed object descriptions." + ], + "image_footnote": [], + "bbox": [ + 326, + 339, + 496, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The region [638, 157, 775, 709] is where a man [575, 150, 787, 993] is standing.... The water bottle [732, 321, 796, 481] suggests that he is well-equipped for his journey. Moreover, his grey jacket [575, 214, 720, 658] indicates that he is ready for the outdoor activity.", + "bbox": [ + 503, + 343, + 681, + 425 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad9d796c0ce9fd398c59abbb0cf3aadad020b7a8d46e09fefaf97857fe32e153.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 338, + 859, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing MLLM-generated image descriptions and scoring them by human or LLMs. However, these methods primarily leverage preferences to improve image-level coarse alignment, and most of them target mitigating hallucinations in MLLMs. The problem of fine-grained alignment for spatial understanding, such as detailed region descriptions and accurate object localization, has been largely neglected.", + "bbox": [ + 89, + 512, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this gap, we design SPR, a Spatial Preference Rewarding framework that enhances MLLM spatial understanding capabilities by rewarding detailed responses with accurate object localization over vague or inaccurate responses. Specifically, SPR selects random image regions containing multiple objects and prompts MLLMs in diverse ways to generate grounded region descriptions. In reward modeling, it introduces both semantic and localization scores to evaluate the alignment between the region description and the region semantics, as well as how detailed region objects are described. We also refine the grounded object in the generated description to enhance its localization accuracy. Finally, the best-scored refined description and the response of the lowest score are paired as preferred and rejected data for direct preference optimization (DPO) [43] training with LORA [17]. By aligning MLLMs with detailed and accurate responses, SPR mitigates MLLMs' incompetence in accurate localization and spatial understand", + "bbox": [ + 89, + 628, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing as required in many real-world tasks.", + "bbox": [ + 511, + 512, + 784, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate the effectiveness of SPR in enhancing MLLMs' spatial understanding capabilities with minimal overhead in training. Compared to the baseline, SPR enhances MLLMs on both referring and grounding benchmarks, especially under higher IoU thresholds which demand higher localization accuracy. In addition, SPR can improve MLLM trustworthiness and reduce MLLM hallucinations as well. Our experiments highlight the importance of incorporating preference-based feedback to enhance the fine-grained spatial understanding abilities in MLLMs.", + "bbox": [ + 511, + 529, + 906, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are summarized as follows:", + "bbox": [ + 513, + 683, + 898, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Spatial Preference Rewarding (SPR) framework to enhance the fine-grained spatial understanding of MLLMs via direct preference optimization (DPO), enhancing MLLMs' capabilities in precise region referring and accurate object localization in images.", + "- We develop an automated pipeline that creates preference data by constructing random region prompts and scoring model responses for spatial understanding. The pipeline requires no other MLLMs or human labours, making it scalable in future training.", + "- Extensive experiments show that the proposed SPR improves MLLMs' spatial understanding capabilities consistently across multiple public benchmarks." + ], + "bbox": [ + 511, + 702, + 906, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "721", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, the success of large language models (LLMs) [14, 49, 51] has been extended into the multimodal domain, resulting in models that demonstrate impressive performance in integrating vision and language [1, 15, 20, 32, 70]. These models treat visual signals as a special form of language, establishing multimodal understanding, reasoning, and interaction capabilities by combining visual encoders [37, 42] with pre-trained large language models, or by directly feeding encoded visual signals into LLMs [6, 53]. Most current MLLMs follow a two-step training process. The first step is pre-training, where large-scale vision-language datasets [9] are used to align visual features to the same space as language features. This enables the model to bridge visual and language embeddings effectively. The second step involves instruction-following finetuning, where high-quality vision-language datasets [25, 33, 70] are used to further enhance the MLLMs' capabilities to follow user instructions and comprehend multimodal information. These methods often convert existing datasets into an instruction-following format or adopt leading MLLMs like GPT to generate high-quality training instruction data for MLLMs [11, 33]. Despite their success, current MLLMs still face challenges that may generate undesired responses toward human preferences. For instance, these models are prone to generating hallucinated content [4, 30, 56, 59] or providing responses that do not fully meet user expectations. Improving the quality of MLLM responses and aligning them more closely with user preferences has thus become a surging focus of research in the community. Our work aims to improve the spatial understanding capabilities of MLLMs, aligning their behaviors better with human preferences.", + "bbox": [ + 93, + 114, + 480, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLMs for Spatial Understanding. Spatial understanding capabilities [7, 13, 22, 67], such as object detection, referring, and grounding description tasks, have long been a fundamental research topic in the field of computer vision. Recent efforts attempt to empower MLLMs with dense visual perception and spatial understanding abilities by integrating region-level data in MLLM training or modifying MLLM architectures. For example, Kosmos-2 [39] and Shikra [10] directly represent the object coordinates in text, constructing instruction datasets to inject spatial knowledge into MLLMs. LLava-Grounding [63] and GroundingGPT [27] construct large-scale grounding datasets to enhance multimodal grounding capabilities. To better facilitate localization within images, RegionGPT [16], GPT4ROI [66], Ferret [58], and Groma [35] encode region features as direct inputs to LLMs, facilitating explicit attention to specific image regions. The Griffon [60, 61] series focuses on dense detection, enabling MLLMs to achieve performance comparable to traditional object detectors. LocVLM [44] explores the. LocVLM [44]", + "bbox": [ + 93, + 599, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "explores key factors in instruction tuning for spatial understanding, such as coordinate representation, which improves MLLM's spatial awareness. However, these efforts primarily concentrate on the instruction-tuning phase and lack direct feedback on MLLMs' responses. To fill this gap, we propose a Spatial Preference Rewarding (SPR) framework, which constructs preference data based on MLLMs generated grounded region descriptions for MLLM tuning.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference Optimization for MLLMs. Preference alignment has recently emerged as a promising direction to align model responses with human preferences. One widely explored approach is to employ Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) to improve the trustworthiness of MLLMs and reduce hallucinations in their responses. For example, LLaVA-RLHF [48] and RLHF-V [59] leverage human annotators to evaluate model responses and construct preference data for fine-tuning. POVID [68] and Silkie [26] use external models, such as GPT, as evaluators to build preference datasets. CLIP-DPO [38] and CSR [69] use CLIP to rank model responses to avoid resource-intensive human or MLLM annotations. AMP [65] introduced a multi-level preference framework to enable MLLMs to better model differences between preference data. mDPO [50] introduced additional preference data pairs with corrupted images to avoid over-optimization on language-only preferences. Unlike these existing studies that primarily aim to reduce hallucinations in MLLMs, our proposed SPR framework focuses on optimizing MLLM responses related to spatial reasoning and understanding. Specifically, SPR focuses on fine-grained alignment with visual inputs and facilitates MLLMs in distinguishing between high-quality object localization (positive samples) and inaccurate localization (negative samples), thereby improving the spatial understanding capabilities of MLLMs.", + "bbox": [ + 511, + 213, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 632, + 611, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our proposed Spatial Preference Rewarding (SPR) framework. Following a typical DPO pipeline, SPR adopts a three-step process in MLLM finetuning, including collecting MLLMs' raw responses (Sec.3.1), evaluating the raw responses to construct preference data (Sec.3.2), and preference optimization (Sec.3.3). The details are elaborated in the following subsections.", + "bbox": [ + 511, + 657, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Grounded Region Description Generation", + "text_level": 1, + "bbox": [ + 511, + 772, + 870, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step of our pipeline is to collect diverse model responses that will later be ranked to construct preference data. Since our primary objective is to enhance MLLMs' localization capabilities and achieve fine-grained alignment to visual inputs, we choose the task of region description with grounding to evaluate MLLMs' object localization capabilities. However, existing datasets [23, 57] for region", + "bbox": [ + 511, + 794, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "722", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5accb4f5af8f7bbedc29d2987ae4baba93d938c8a84f49ae1eafd1114988169e.jpg", + "image_caption": [ + "Figure 2. We leverage the generated object references and cropped image region to build a variety of multimodal prompts to enhance the diversity of generated region descriptions." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 885, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "descriptions are often too simple, involving queried regions with only one or two objects and short phrases such as 'vehicles parked on the street' or 'bicycles are parked on the sidewalk.' Such simple prompts are inadequate for generating diverse responses to construct preferred and rejected preference data with sufficient divergence, which might hinder the effectiveness of DPO training [54]. To address this issue, we generate queried regions from scratch instead of using existing region description datasets.", + "bbox": [ + 88, + 354, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region Query Construction. We design a simple approach to generate randomly queried regions based on images and object annotations. Take the Objects365 dataset as an example. We first filter out images with few objects, ensuring that the data contains rich visual content. Then, given the annotated object bounding boxes in each image, we randomly select one of the objects as the starting region. From there, we iteratively expand the region by incorporating the nearest objects. The expansion stops randomly once more than four objects are involved in the region. The resulting region then prompts MLLMs to generate a detailed region description. Through this process, we simulate the human-like, dynamic attention across different parts of an image, encouraging the MLLM to adaptively focus on arbitrary image regions based on the given prompts.", + "bbox": [ + 89, + 491, + 483, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Grounded Region Description Generation. As shown in Fig. 2, we build a variety of prompts for MLLMs to generate several region descriptions for each image, serving as candidate responses for preference data. Since the original MLLM sometimes struggles to generate detailed responses following region prompts, we utilize cropped region images along with object references constructed from annotations to guide MLLMs to attend to the region's content and details. These prompts help the model focus more effectively on the specified region and produce detailed descriptions that might better align with human preference. In this way, we encourage the MLLM to generate responses that are dis", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tinct in content but consistent in language style, which is then used for constructing preference data.", + "bbox": [ + 511, + 354, + 906, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preference Data Ranking and Construction", + "text_level": 1, + "bbox": [ + 511, + 396, + 880, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next step is to rank the generated descriptions to obtain preferred and rejected data pairs. An ideal region description should meet at least two key criteria: (1) the text description should accurately match the semantics of the queried region and the surrounding image content, (2) it should provide detailed descriptions with accurate localization of objects within the region. To address these two criteria, we propose a semantic score and localization score to rank the responses. The descriptions with the highest and the lowest scores are paired to form preference data for DPO training.", + "bbox": [ + 511, + 417, + 906, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Score. We introduce the semantic score to evaluate the relevance between the generated descriptions and the semantics of queried image regions. We leverage a pretrained CLIP model [42] to compute the cosine similarities of text and visual embeddings as defined in Eq. (1):", + "bbox": [ + 511, + 585, + 905, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (I, T) = \\alpha * \\cos \\left(\\mathcal {F} _ {\\text {r e g i o n}} (I), \\mathcal {F} _ {\\text {t e x t}} (T)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 678, + 903, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\alpha$ is the scale of similarities, which is set as 5 in our work to balance the range of the semantic score, $I$ and $T$ are the input image and MLLM generated region description with grounding text removed; $\\mathcal{F}_{region}$ and $\\mathcal{F}_{text}$ denotes the visual embedding for image region and text embeddings, respectively.", + "bbox": [ + 511, + 703, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When extracting image region embeddings, a straightforward approach is to crop the image region $I_{crop}$ and directly extract visual embeddings. However, the similarity score with such embedding tends to overly focus on the region's details while neglecting the image's surrounding context. To address this limitation, we supplement it with similarities $S_{local}$ from visual embeddings of intact images", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "723", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that incorporate local attention. Specifically, we feed the original image into CLIP and replace the final layer of the vision encoder that aggregates the embeddings with a local-attention layer. This modification allows the model to better account for the context around the region of interest. As defined in Eq. (2), we then use the average of the cropped image's similarity score and the full image's similarity score with local attention as the final semantic score, which effectively evaluates the extent of fine-grained alignment between the region description and local visual semantics.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {s e m} = \\frac {1}{2} \\left(S \\left(I _ {c r o p}, T\\right) + S _ {l o c a l} (I, T)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 253, + 483, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Localization Score. We propose a localization score to evaluate how detailed the MLLM responds in describing objects within the queried region and its grounding accuracy. This score is calculated based on the number of objects mentioned in the description that match the ground truth objects in the region. In practice, we use Grounding DINO [34] and the cropped image region to extract bounding boxes for objects mentioned in the description. The extracted objects are then combined with the original object annotations to form a set of ground truth objects within the region. Next, we extract the grounding results from MLLM-generated descriptions and combine them with the results from Grounding DINO to form the predicted objects. Finally, we compute the average IoU between the predicted objects and the ground truth as the localization score. The detailed process is outlined in Algorithm 1.", + "bbox": [ + 89, + 287, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The localization score encourages the model to include more detailed descriptions for involved objects and accurately localize them in its responses. Finally, we combine the semantic and localization scores for each grounded region description:", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = \\lambda S _ {s e m} + (1 - \\lambda) S _ {l o c} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 619, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to 0.8 in our implementation. Then, the descriptions with the highest and lowest scores are paired as preferred and rejected data for preference optimization.", + "bbox": [ + 89, + 643, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grounded Region Description Refinement. After obtaining the preference data pairs, we further enhance the divergence of the grounding results of the preferred and rejected descriptions to encourage the model to distinguish between accurate and inaccurate object localization. To achieve this, we refine the grounding results in the preferred descriptions while keeping the rejected ones unchanged. In practice, we leverage the results obtained while computing the localization score, including the object box predictions $B_{pred}$ and ground-truth object boxes $B_{gt}$ . We retain only those predictions that match the ground truths (IoU > 0.5) and replace their bounding boxes with the matched ones. Then, we remove duplicates of predictions based on their textual position in the description and IoUs. Finally, we insert the", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Computing Localization Score", + "text_level": 1, + "bbox": [ + 514, + 90, + 808, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Cropped Image Region $I_{crop}$ , Grounded Region Description $T$ generated by MLLMs, Object Bounding Box Annotations $B_{anno}$ for the Queried Region.", + "bbox": [ + 514, + 109, + 903, + 155 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Localization Score: $S_{loc}$ .", + "bbox": [ + 514, + 155, + 746, + 170 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Extract bounding boxes $B_{text}$ from the description $T$ and get the plain text $T_{plain}$ .", + "2: Leverage Grounding DINO to get grounded object results $B_{\\text{ground}}$ from $T_{\\text{plain}}$ .", + "3: Get the set of ground truth object boxes $B_{gt}$ by aggregating $B_{ground}$ and $B_{anno}$ and removing duplicated boxes.", + "4: Get the set of object box predictions $B_{pred}$ for the description $T$ by aggregating $B_{ground}$ and $B_{text}$ and removing duplicated boxes.", + "5: Computing IoU matrix $\\mathbf{m}[i,j] = IoU(B_{gt}^i, B_{pred}^j)$ between $B_{gt}$ and $B_{pred}$ .", + "6: Filter the IoU by a threshold of 0.5." + ], + "bbox": [ + 521, + 170, + 903, + 367 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} [ i, j ] = \\left\\{ \\begin{array}{l l} 0 & \\mathbf {m} [ i, j ] < 0. 5 \\\\ \\mathbf {m} [ i, j ] & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 375, + 836, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "7: return $S_{loc} = \\frac{1}{n}\\sum_{i}^{n}\\max_{j}\\mathbf{p}[i,j]$", + "bbox": [ + 522, + 422, + 764, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "refined object box predictions into the region description, resulting in an improved grounded region description with more precise coordinates.", + "bbox": [ + 511, + 478, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Preference Optimization", + "text_level": 1, + "bbox": [ + 511, + 532, + 738, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After curating the preference dataset, we finetune MLLMs through DPO and adopt LORA to save the training cost. The loss for optimizing MLLMs is defined as:", + "bbox": [ + 511, + 555, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right)} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {*} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {*} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 622, + 906, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{w}$ and $y_{l}$ are the preferred and rejected description data; $\\pi_{r}ef(y|x)$ is the base reference policy model, i.e., the initial instruction-tuned MLLM which is frozen during the training; $\\pi_{*}(y|x)$ denotes the policy model which inherits from the instruction-tuned model with its LORA weights updated in the training process.", + "bbox": [ + 511, + 659, + 905, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 763, + 645, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setups.", + "text_level": 1, + "bbox": [ + 511, + 787, + 700, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In this work, we experiment with the proposed SPR with three MLLMs with spatial understanding capabilities, including Ferret [58], LLaVOneVision [24], and CogVLM-Grounding [52]. To construct preference data, we randomly select 10k images with object annotations from the training set of Objects365", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "724", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b8902624abe45f07c9c0d954ef56eb7d469c7b906736909ba1624f3caedd4ec6.jpg", + "table_caption": [ + "Table 1. Experiments on the Referring Expression Comprehension task (Acc@0.5) on datasets RefCOCO/+/g, and the Phrase Grounding task (Recall@1) on Flickr30k Entities dataset. “-” indicates results are unavailable or that MLLMs do not support multi-object grounding." + ], + "table_footnote": [], + "table_body": "
MethodRefCOCORefCOCO+RefCOCOgFlickr30k Entities
valtestAtestBvaltestAtestBvaltestvaltest
UNITER [12]81.4187.0474.1775.9081.4566.7074.0268.67--
UniTAB [55]86.3288.8480.6178.7083.2269.4879.9679.9778.7679.58
MDETR [21]86.7589.5881.4179.5284.0970.6281.6480.8982.383.8
MiniGPT-v2-7B [70]88.0691.2984.3079.5885.5273.3284.1984.31--
VistaLLM [41]88.191.583.082.989.874.883.684.4--
LLaVA-Grounding [64]89.16--81.68--84.82-83.0383.62
Shikra-7B [10]87.0190.6180.2481.6087.3672.1282.2782.1975.8476.54
Shikra-13B [10]87.8391.1181.8182.8987.7974.4182.6483.1677.4178.44
Griffon-13B [60]89.492.584.683.388.476.085.186.183.784.2
LLava-OV-7B [24]74.7782.5964.0470.1779.8558.4872.3471.39--
+ SPR76.6682.5265.9771.6279.8759.9972.9871.55--
Ferret-7B [58]87.4991.3582.4580.7887.3873.1483.9384.7680.3982.21
+ SPR88.3991.6783.9182.0787.8474.1985.5885.7581.5383.34
Ferret-13B [58]89.4892.4184.3682.8188.1475.1785.8386.3481.1384.76
+ SPR89.9493.0685.1283.2988.8975.7486.4686.9281.8283.75
CogVLM-Grounding-17B [52]92.7694.7588.9988.6892.9183.3989.7590.79--
+ SPR92.9594.8789.1588.8392.9583.8490.0190.96--
", + "bbox": [ + 148, + 128, + 844, + 391 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8a40c68cfb19c35c91c9f99da53dc5dc054d9a5ff8ccc0a7f61354427e00f939.jpg", + "table_caption": [ + "Table 2. Experiments on Referring Expression Comprehension task under different IoU thresholds. The results are the average on RefCOCO, RefCOCO+, and RefCOCOg datasets." + ], + "table_footnote": [], + "table_body": "
IoU Threshold0.50.60.70.80.9
Ferret-7B83.9181.2876.7267.0243.25
+SPR84.9382.3678.4270.0952.21
Ferret-13B85.5682.9478.5770.0449.55
+SPR86.1883.6379.9372.0353.61
", + "bbox": [ + 112, + 454, + 459, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset [46], then construct random regions to query models to generate grounded region descriptions. We adopt LORA [17] for tuning MLLMs. The training is conducted on one A100 GPU, which takes around 3 and 5 hours for Ferret 7B and 13B models, respectively. Please refer to Appendix for more details on preference data construction and hyperparameter selection.", + "bbox": [ + 88, + 564, + 482, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Benchmarks We evaluate our method on three types of benchmarks: (1) Grounding tasks that evaluate the localization accuracy, including referring expression comprehension (REC) and phrase grounding; (2) Region description task on Refcocog [22] and visual genome [23], and Ferret Bench [58] for comprehensive spatial understanding; (3) General benchmarks TextVQA [47], GQA [19], LLaVA-Bench [33], and hallucination benchmark POPE [56].", + "bbox": [ + 88, + 670, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiments on REC", + "text_level": 1, + "bbox": [ + 89, + 816, + 290, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our method on the referring expression comprehension (REC) task on RefCOCO [22], RefCOCO+ [22], and Refcocog [36]. The task requires the model to locate the object or region given a short de", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scription, which evaluates the model's fine-grained visual grounding abilities under the single-object referent scenarios. As shown in Tab. 1, our proposed SPR framework consistently improves the performance of three baseline MLLMs on different datasets for all model sizes. Considering that the REC results are based on an IoU threshold of 0.5, the improvement on its performance indicates that the model localized more objects successfully. Hence, this improvement can be largely attributed to the introduction of localization scores when constructing the preference data in SPR. Region descriptions that accurately mention more objects could achieve higher localization scores in SPR and be more likely to serve as preferred data, facilitating the model to attend to more objects and their locations in the image.", + "bbox": [ + 511, + 404, + 906, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better evaluate the impact of SPR on the localization capability of MLLMs, we also conduct REC experiments with higher IoU thresholds by gradually increasing the IoU threshold of valid REC results from the default value of 0.5 to 0.9. As shown in Tab. 2, the improvements brought by SPR significantly increase as the threshold rises, with accuracy gains of 8.96 and 4.06 for the 7B model and the 13B model, respectively, when the IoU threshold rises to 0.9. With SPR, the localization accuracy of the objects in the model's response is greatly improved. Equipped with the grounded region description refinement and the supervision of preferred-rejected localization data in SPR, the model can respond more accurately to grounding object locations, demonstrating the effectiveness of incorporating preference optimization for region description and object localization in the fine-grained spatial understanding of MLLMs.", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "725", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac696c35aea149b803e666d55e392faab323115fd65792ebe04de87cf42da289.jpg", + "table_caption": [ + "Table 3. Experiments on Phrase Grounding task under different IoU thresholds. The results are averaged over the validation and test set of the Flickr30k dataset." + ], + "table_footnote": [], + "table_body": "
IoU Threshold0.50.60.70.80.9
Ferret-7B81.376.1467.5553.8629.98
+SPR82.4477.1469.2556.1933.99
Ferret-13B82.9476.6268.3455.7432.60
+SPR82.7877.2369.7456.9634.18
", + "bbox": [ + 112, + 142, + 460, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb7ad5990e0e3cc11e5565b7648a41dee8bdc037dac7c12493a20760cdfe9735.jpg", + "table_caption": [ + "Table 4. Experiments on the region captioning task on Refcocog and Visual Genome datasets." + ], + "table_footnote": [], + "table_body": "
MethodRefcocogVisual Genome
METEORROUGE_LMETEORROUGE_L
Ferret-7B12.315.617.429.6
+SPR13.520.417.629.7
Ferret-13B12.926.417.931.0
+SPR13.327.218.231.3
", + "bbox": [ + 94, + 279, + 478, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experiments on Phrase Grounding", + "text_level": 1, + "bbox": [ + 89, + 401, + 397, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we experiment with the phrase grounding task on Flickr30k Entity [40]. In phrase grounding, the queried object phrases are combined in a single question, requiring MLLMs to detect the locations of multiple objects in a single response, which makes it more challenging than the single-object referring task like REC. Following [58], we adopt the question \"What are the locations of [phrases]?\" and evaluate the result using the MERGE-BOXES mode [21]. Since LLaVA-OneVision and CogVLM do not support multi-object detection, we report only the results for Ferret. As shown in Tab. 1, SPR effectively improves Ferret's performance in multi-object referent scenarios, especially for the 7B model, whose performance is even comparable to that of the 13B model.", + "bbox": [ + 88, + 425, + 482, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then experiment with the phrase grounding task under higher IoU thresholds. We found that the multi-object referencing setting in phrase grounding is more challenging than the single-object referencing in REC. As the IoU threshold increases, the performance drops more rapidly, indicating a significant demand for MLLM to improve the capabilities of more accurate localization. Our approach can significantly alleviate this issue. As shown in Tab. 3, SPR improves progressively with higher IoU thresholds, reaching a maximum gain of 4.01 and 1.58 Recall@1 for the 7B and 13B models, respectively. This experiment demonstrates the superiority of SPR in pursuing detailed descriptions with high-precision object localization.", + "bbox": [ + 89, + 637, + 482, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Experiments on Region Captioning", + "text_level": 1, + "bbox": [ + 89, + 847, + 398, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond the grounding task, we also verify our proposed SPR in improving the text qualities of MLLMs' outputs", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/30dcbc7d2afd9616c2e1b50bb29ea629056d15dd1d91f96d5af2e50c1c570e18.jpg", + "table_caption": [ + "Table 5. Experiments on the Ferret Bench. \"Description\", \"Reasoning\", and \"Grounding\" denote the Referring Description, Referring Reasoning, and Grounding in Conversation tasks." + ], + "table_footnote": [], + "table_body": "
ModelDescriptionReasoningGroundingAvg.
Ferret-7B68.767.357.564.5
+ SPR70.068.458.165.5
Ferret-13B70.668.759.766.3
+ SPR70.872.660.267.9
", + "bbox": [ + 529, + 142, + 887, + 226 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73ada29cfcffc83458d2dbbd1823dc76d5c91cf28e520a445739f531eccbc146.jpg", + "table_caption": [ + "Table 6. Experiments on the general and hallucination benchmarks. We report the accuracy for GQA and VQA and the F1 score on POPE." + ], + "table_footnote": [], + "table_body": "
ModelVQATGQALLaVAPOPE
Ferret-7B--64.785.36
+ SPR--66.385.69
LLaVA-OV-7B75.8962.2188.988.12
+ SPR76.0762.4291.488.49
", + "bbox": [ + 547, + 291, + 870, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on fine-grained spatial understanding. We conduct experiments on the RefCOCOg and Visual Genome benchmarks. We prompt MLLMs with the question \"Describe the region [region] in the image.\" to generate region captions and then evaluate the response quality using METEOR [5] and ROUGE_L [28] metrics. Tab. 4 shows that SPR effectively improves the quality of MLLM-generated region captions. After tuning with SPR, MLLMs are able to effectively attend to the user-specified regions and generate captions that better reflect the details of the region content.", + "bbox": [ + 511, + 398, + 906, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Experiments on Ferret Bench", + "text_level": 1, + "bbox": [ + 511, + 560, + 777, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ferret benchmark, proposed by [58], aims to evaluate MLLMs' fine-grained multimodal conversational capabilities such as referring description, referring reasoning, and grounded conversation. We follow the pipeline in [58] to prompt MLLMs with questions and employ GPT to evaluate the responses. As shown in Tab. 5, the proposed SPR can facilitate MLLMs in achieving better conversational qualities for fine-grained multimodal understanding, especially for the referring reasoning task, with an accuracy gain of about 3.9 for the 13B model. Equipped with SPR, MLLM can focus on more detailed visual information and generate responses that align better with human preferences.", + "bbox": [ + 511, + 582, + 906, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Experiments on General Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 773, + 841, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate SPR on three general benchmarks to validate the benefits of improving MLLMs' spatial capabilities. TextVQA and GQA require MLLMs to answer questions or perform reasoning based on specific text, objects, and image content. LLaVA bench evaluates MLLMs comprehensive capabilities in conversation, description, and reasoning. As shown in Tab. 6, improving MLLMs' spa", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "726", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f399fe37d719590d6ab6af66e0309d1f89849786f23f8f7558e5ff6152a85c8.jpg", + "table_caption": [ + "Table 7. Ablation Studies on the refinement of grounded region descriptions, and ratio $\\lambda$ between semantic and localization scores in ranking MLLM responses. We report the average results on Referring expression comprehension and phrase grounding tasks." + ], + "table_footnote": [], + "table_body": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ SPR84.9382.44
w/o Refinement84.4191.38
λ = 0.084.2581.83
λ = 0.484.3481.95
λ = 0.684.6682.13
λ = 0.884.9382.44
λ = 1.084.4581.87
", + "bbox": [ + 145, + 156, + 431, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07baba7eb626995d30778c63ea1b821a0d03f194033b65ac90903de3736d8099.jpg", + "table_caption": [ + "Table 8. Ablation studies on the training strategy." + ], + "table_footnote": [], + "table_body": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ Instruction Finetuning84.3581.72
+ DPO training84.9382.44
", + "bbox": [ + 127, + 325, + 441, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tial understanding capabilities consistently enhances their comprehension and reasoning abilities across diverse general scenarios, leading to performance gains on all three benchmarks. We also experiment on hallucination benchmark POPE, where SPR improves both Ferret and LLaVA-OneVision. This can be attributed to the preference data construction in SPR, where semantic and localization scores are applied to select region descriptions that better align with region content and reject those related to content outside the region or that contain hallucinations, thus effectively helping mitigate the hallucinations in MLLMs.", + "bbox": [ + 89, + 401, + 483, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 575, + 254, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies over the two designs in SPR and evaluate the performance of Ferret-7B on the REC (Refcoco $+ / \\mathrm{g}$ ) and the Flickr30k phrase grounding tasks.", + "bbox": [ + 89, + 598, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Score Ratio $\\lambda$ . In Sec. 3.2, we combine the semantic and localization scores to rank MLLMs generated descriptions with a score ratio $\\lambda$ . As shown in Tab. 7, we vary the $\\lambda$ from 0 to 1, and the trained models outperform the baseline model consistently. When $\\lambda$ equals zero, SPR achieves minimal gain, as the model might overly reward descriptions that simply list object names or fail to align with the region's semantics. On the other hand, when $\\lambda$ is set to 1, SPR completely disregards measuring how detail the MLLM describes the objects in the region. Under such situations, the model encourages coarse region descriptions with fewer objects involved and reduces the corresponding object bounding box texts in the preferred data, thereby hindering the training of MLLMs' localization capability. As the experiments show, a relatively high value of 0.8 achieves the best results and is set as the default value in SPR.", + "bbox": [ + 89, + 643, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Refinement of Grounded Region Description. After con", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "structuring the preferred and rejected data pairs, we further refine the localization results in the preferred descriptions by completing bounding boxes for objects in the description that were not grounded and refining the existing bounding boxes. Tab. 7 shows the results of this refinement. We found that the refinement leads to greater improvements in the multi-object referring task of phrase grounding. This could be attributed to the fact that the baseline model often fails to follow instructions for providing bounding boxes for each mentioned object when generating region descriptions. After tuning by the refined descriptions, MLLMs could faithfully ground the mentioned objects, thereby improving the multi-object phrase grounding clearly.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. Comparison with SFT", + "text_level": 1, + "bbox": [ + 511, + 296, + 720, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we adopt DPO with accept-reject preference data to optimize MLLMs for spatial understanding, whereas prior work [10, 44, 61], primarily focuses on the stage of supervised instruction fine-tuning (SFT). In Tab. 8, we compare these two training approaches, where SFT is trained using only the accepted data. The results show that while SFT could improve MLLMs' localization capabilities, its performance gains are significantly lower than DPO. DPO optimizes MLLM by contrasting accepted and rejected data pairs, similar to the positive-negative sample training mechanism in traditional object detection algorithms. This approach helps models distinguish between accurate and inaccurate localizations and facilitates MLLM in spatial understanding more effectively. However, it is important to note that DPO training also relies on a well-trained SFT model as a foundation, making these two approaches complementary. In future work, we will further explore how to integrate SFT and DPO to enhance MLLMs' spatial understanding.", + "bbox": [ + 511, + 318, + 906, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 603, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose SPR, a Spatial Preference Rewarding framework to enhance MLLM's fine-grained spatial understanding capabilities. We introduce a complete pipeline that includes (1) Constructing random region queries; (2) Prompting MLLMs to generate diverse grounded region descriptions; (3) Proposing semantic scores and localization scores to rank the descriptions comprehensively; (4) Refining the localization quality of preference data; (5) Finetuning MLLMs to optimize against detailed and accurate spatial understanding. The entire framework does not require additional human labor or external MLLMs, with minimal overhead on training costs. SPR addresses the lack of direct optimization for positive and negative localization samples in MLLM training, enhancing their localization capabilities and promoting better alignment with human preferences. Experiments demonstrate that SPR significantly improves MLLMs' performance on standard referring and grounding tasks for spatial understanding.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "727", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 114, + 483, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study is also supported by the MOE Tier-2 project, with project number MOE-T2EP20123-0003.", + "bbox": [ + 89, + 174, + 482, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 229, + 187, + 244 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3", + "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1", + "[4] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3", + "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7", + "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 3", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with Transformers. In ECCV, 2020. 1, 3", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1", + "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3", + "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 6, 8", + "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v:" + ], + "bbox": [ + 93, + 255, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3", + "[12] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 6", + "[13] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2024. 3", + "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 1, 3", + "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[16] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 1, 3", + "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6", + "[18] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 1", + "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6", + "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3", + "[21] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 6, 7", + "[22] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "728", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3, 6", + "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5, 6", + "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[26] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3", + "[27] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundingpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3", + "[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7", + "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 1", + "[30] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3", + "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3, 6", + "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[35] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 1, 3", + "[36] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6", + "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3", + "[38] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. arXiv preprint arXiv:2408.10433, 2024. 1, 3", + "[39] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3", + "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 7", + "[41] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 6", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4", + "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[44] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12987, 2024. 3, 8", + "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1", + "[46] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6", + "[47] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Con-" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "729", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 6", + "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 1, 3", + "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3", + "[50] Fei Wang, Wenxuan Zhou, James Y Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. arXiv preprint arXiv:2406.11839, 2024. 1, 3", + "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3", + "[52] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1, 5, 6", + "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3", + "[54] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. beta-dpo: Direct preference optimization with dynamic beta. arXiv preprint arXiv:2407.08639, 2024. 4", + "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 6", + "[56] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Zhao Wayne Xin, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 3, 6", + "[57] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3", + "[58] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 1, 3, 5, 6, 7", + "[59] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. arxiv, 2023. 1, 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 3, 6", + "[61] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 1, 3, 8", + "[62] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024.1", + "[63] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 3", + "[64] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 6", + "[65] Mengxi Zhang and Kang Rong. Automated multi-level preference for mllms. arXiv preprint arXiv:2405.11165, 2024. 1, 3", + "[66] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 3", + "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6578-6587, 2019. 3", + "[68] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024.3", + "[69] Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622, 2024. 1,3", + "[70] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 903, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "730", + "bbox": [ + 486, + 944, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_model.json b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2fcd9be952383ed06b4acb3b32efb8a0fb0802c3 --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Spatial Preference Rewarding for MLLMs Spatial Understanding" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.18, + 0.807, + 0.201 + ], + "angle": 0, + "content": "Han Qiu\\(^{1}\\), Peng Gao\\(^{2}\\), Lewei Lu\\(^{3}\\), Xiaqin Zhang\\(^{4}\\), Ling Shao\\(^{5}\\), Shijian Lu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.215, + 0.777, + 0.249 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University, \\(^{2}\\)Shanghai AI Laboratory \\(^{3}\\)Sensetime Research, \\(^{4}\\)Zhejiang University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.251, + 0.776, + 0.269 + ], + "angle": 0, + "content": "\\(^{5}\\)UCAS-Terminus AI Lab, University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.273, + 0.684, + 0.287 + ], + "angle": 0, + "content": "han023@e.ntu.edu.sg, Shijian.Lu@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.794 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding (SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.822, + 0.222, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.483, + 0.878 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) [3, 15, 18, 31, 32, 51, 58, 70] have achieved remarkable success by inte-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.908, + 0.534 + ], + "angle": 0, + "content": "grating pretrained large language model [2, 14, 49] with vision encoders [8, 37, 42], leading to significant advancements in a wide range of general vision-language tasks. By combining visual and language signals, MLLMs have demonstrated superior capabilities in multimodal understanding, reasoning, and interaction as compared with traditional vision models. Recently, several studies have further injected spatial knowledge into MLLMs, thereby improving MLLMs' fine-grained perception of visual inputs and enabling tasks such as referential dialogue [10, 16], grounding captioning [35, 58, 62], region description [31, 52], and object detection [61], etc. These advances have paved the way for MLLMs to serve as versatile visual assistants supporting a wider range of applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Despite recent advancements, MLLMs still face challenges in fine-grained spatial understanding, with responses not aligned with human preferences. As illustrated in 1, the generated grounded region descriptions are often vague with inaccurate object localizations, and models may fail to focus on the queried region, distracted from other regions in the image. The issue in spatial understanding could be attributed to the lack of positive-negative preference feedback in existing instruction-tuned MLLMs. Specifically, instruction fine-tuning (SFT) directly optimizes MLLMs to mimic ground truth positive samples, but it cannot impose any penalties if the model produces inaccurate negative samples for localization during actual inference. As a result, MLLMs may struggle to generate positive descriptions with accurate object localization and instead produce negative and inaccurate descriptions, leading to responses that do not align with user expectations. In addition, optimization using positive and negative samples has been proven crucial for spatial understanding in traditional object detection algorithms [7, 29, 45], highlighting a significant gap in the current MLLM training on spatial understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Several studies [38, 48, 50, 59, 65, 69] attempt to introduce preference optimization for better MLLM alignment, where the preference data are constructed by collect" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "720" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.09, + 0.345, + 0.104 + ], + "angle": 0, + "content": "User Prompt about Regions in Images:" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.105, + 0.31, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.128, + 0.484, + 0.172 + ], + "angle": 0, + "content": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.104, + 0.672, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.679, + 0.129, + 0.846, + 0.172 + ], + "angle": 0, + "content": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object." + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.197, + 0.243, + 0.21 + ], + "angle": 0, + "content": "MLLM Response:" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.23, + 0.323, + 0.285 + ], + "angle": 0, + "content": "The area [676, 229, 847, 1000] is occupied by a woman in a blue shirt who is holding a whistle [537, 530, 568, 606]. She is also wearing a black clipboard [801, 538, 825, 581] around her neck ...." + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.286, + 0.311, + 0.294 + ], + "angle": 0, + "content": "Fail to Ground" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.21, + 0.498, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.229, + 0.679, + 0.294 + ], + "angle": 0, + "content": "The region [638, 157, 775, 709] is a trail located on a mountain [166, 365, 996, 996]. of people [619, 250, 799, 990] walking on it. The trail also has a marker [166, 339, 229, 439] to guide hikers. Inaccurate Localization" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.21, + 0.861, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.306, + 0.653, + 0.32 + ], + "angle": 0, + "content": "Tuning with Spatial Preference Rewarding (SPR)" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.326, + 0.376, + 0.34 + ], + "angle": 0, + "content": "Responses from SPR-empowered MLLM:" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.342, + 0.319, + 0.425 + ], + "angle": 0, + "content": "The area [676, 229, 847, 998] in the image features a woman in a blue shirt with a name tag, .... She [676, 229, 817, 901] appears to .... The grassy area where she is standing [0, 385, 998, 998] seems to be .... activities. There's also a red cone [676, 909, 720, 998] ...." + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.34, + 0.498, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.344, + 0.683, + 0.426 + ], + "angle": 0, + "content": "The region [638, 157, 775, 709] is where a man [575, 150, 787, 993] is standing.... The water bottle [732, 321, 796, 481] suggests that he is well-equipped for his journey. Moreover, his grey jacket [575, 214, 720, 658] indicates that he is ready for the outdoor activity." + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.339, + 0.861, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.443, + 0.907, + 0.499 + ], + "angle": 0, + "content": "Figure 1. The proposed Spatial Preference Rewarding (SPR) mitigates the distracted and inaccurate region descriptions generated by MLLMs. Given an image and a user-specified region of interest, MLLMs often fail to focus on the queried region. They may be distracted by objects outside the specified region, failing to ground the queried objects, or providing inaccurate localization. Tuning MLLMs with our proposed SPR leads to more accurate object localization and detailed object descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.483, + 0.62 + ], + "angle": 0, + "content": "ing MLLM-generated image descriptions and scoring them by human or LLMs. However, these methods primarily leverage preferences to improve image-level coarse alignment, and most of them target mitigating hallucinations in MLLMs. The problem of fine-grained alignment for spatial understanding, such as detailed region descriptions and accurate object localization, has been largely neglected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address this gap, we design SPR, a Spatial Preference Rewarding framework that enhances MLLM spatial understanding capabilities by rewarding detailed responses with accurate object localization over vague or inaccurate responses. Specifically, SPR selects random image regions containing multiple objects and prompts MLLMs in diverse ways to generate grounded region descriptions. In reward modeling, it introduces both semantic and localization scores to evaluate the alignment between the region description and the region semantics, as well as how detailed region objects are described. We also refine the grounded object in the generated description to enhance its localization accuracy. Finally, the best-scored refined description and the response of the lowest score are paired as preferred and rejected data for direct preference optimization (DPO) [43] training with LORA [17]. By aligning MLLMs with detailed and accurate responses, SPR mitigates MLLMs' incompetence in accurate localization and spatial understand" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.785, + 0.528 + ], + "angle": 0, + "content": "ing as required in many real-world tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.682 + ], + "angle": 0, + "content": "We validate the effectiveness of SPR in enhancing MLLMs' spatial understanding capabilities with minimal overhead in training. Compared to the baseline, SPR enhances MLLMs on both referring and grounding benchmarks, especially under higher IoU thresholds which demand higher localization accuracy. In addition, SPR can improve MLLM trustworthiness and reduce MLLM hallucinations as well. Our experiments highlight the importance of incorporating preference-based feedback to enhance the fine-grained spatial understanding abilities in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.684, + 0.9, + 0.699 + ], + "angle": 0, + "content": "The contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.703, + 0.906, + 0.778 + ], + "angle": 0, + "content": "- We propose a Spatial Preference Rewarding (SPR) framework to enhance the fine-grained spatial understanding of MLLMs via direct preference optimization (DPO), enhancing MLLMs' capabilities in precise region referring and accurate object localization in images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.779, + 0.907, + 0.853 + ], + "angle": 0, + "content": "- We develop an automated pipeline that creates preference data by constructing random region prompts and scoring model responses for spatial understanding. The pipeline requires no other MLLMs or human labours, making it scalable in future training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.853, + 0.905, + 0.899 + ], + "angle": 0, + "content": "- Extensive experiments show that the proposed SPR improves MLLMs' spatial understanding capabilities consistently across multiple public benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.703, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "721" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.597 + ], + "angle": 0, + "content": "Recently, the success of large language models (LLMs) [14, 49, 51] has been extended into the multimodal domain, resulting in models that demonstrate impressive performance in integrating vision and language [1, 15, 20, 32, 70]. These models treat visual signals as a special form of language, establishing multimodal understanding, reasoning, and interaction capabilities by combining visual encoders [37, 42] with pre-trained large language models, or by directly feeding encoded visual signals into LLMs [6, 53]. Most current MLLMs follow a two-step training process. The first step is pre-training, where large-scale vision-language datasets [9] are used to align visual features to the same space as language features. This enables the model to bridge visual and language embeddings effectively. The second step involves instruction-following finetuning, where high-quality vision-language datasets [25, 33, 70] are used to further enhance the MLLMs' capabilities to follow user instructions and comprehend multimodal information. These methods often convert existing datasets into an instruction-following format or adopt leading MLLMs like GPT to generate high-quality training instruction data for MLLMs [11, 33]. Despite their success, current MLLMs still face challenges that may generate undesired responses toward human preferences. For instance, these models are prone to generating hallucinated content [4, 30, 56, 59] or providing responses that do not fully meet user expectations. Improving the quality of MLLM responses and aligning them more closely with user preferences has thus become a surging focus of research in the community. Our work aims to improve the spatial understanding capabilities of MLLMs, aligning their behaviors better with human preferences." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.6, + 0.482, + 0.901 + ], + "angle": 0, + "content": "MLLMs for Spatial Understanding. Spatial understanding capabilities [7, 13, 22, 67], such as object detection, referring, and grounding description tasks, have long been a fundamental research topic in the field of computer vision. Recent efforts attempt to empower MLLMs with dense visual perception and spatial understanding abilities by integrating region-level data in MLLM training or modifying MLLM architectures. For example, Kosmos-2 [39] and Shikra [10] directly represent the object coordinates in text, constructing instruction datasets to inject spatial knowledge into MLLMs. LLava-Grounding [63] and GroundingGPT [27] construct large-scale grounding datasets to enhance multimodal grounding capabilities. To better facilitate localization within images, RegionGPT [16], GPT4ROI [66], Ferret [58], and Groma [35] encode region features as direct inputs to LLMs, facilitating explicit attention to specific image regions. The Griffon [60, 61] series focuses on dense detection, enabling MLLMs to achieve performance comparable to traditional object detectors. LocVLM [44] explores the. LocVLM [44]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "explores key factors in instruction tuning for spatial understanding, such as coordinate representation, which improves MLLM's spatial awareness. However, these efforts primarily concentrate on the instruction-tuning phase and lack direct feedback on MLLMs' responses. To fill this gap, we propose a Spatial Preference Rewarding (SPR) framework, which constructs preference data based on MLLMs generated grounded region descriptions for MLLM tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.62 + ], + "angle": 0, + "content": "Preference Optimization for MLLMs. Preference alignment has recently emerged as a promising direction to align model responses with human preferences. One widely explored approach is to employ Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) to improve the trustworthiness of MLLMs and reduce hallucinations in their responses. For example, LLaVA-RLHF [48] and RLHF-V [59] leverage human annotators to evaluate model responses and construct preference data for fine-tuning. POVID [68] and Silkie [26] use external models, such as GPT, as evaluators to build preference datasets. CLIP-DPO [38] and CSR [69] use CLIP to rank model responses to avoid resource-intensive human or MLLM annotations. AMP [65] introduced a multi-level preference framework to enable MLLMs to better model differences between preference data. mDPO [50] introduced additional preference data pairs with corrupted images to avoid over-optimization on language-only preferences. Unlike these existing studies that primarily aim to reduce hallucinations in MLLMs, our proposed SPR framework focuses on optimizing MLLM responses related to spatial reasoning and understanding. Specifically, SPR focuses on fine-grained alignment with visual inputs and facilitates MLLMs in distinguishing between high-quality object localization (positive samples) and inaccurate localization (negative samples), thereby improving the spatial understanding capabilities of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.612, + 0.649 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.905, + 0.765 + ], + "angle": 0, + "content": "This section presents our proposed Spatial Preference Rewarding (SPR) framework. Following a typical DPO pipeline, SPR adopts a three-step process in MLLM finetuning, including collecting MLLMs' raw responses (Sec.3.1), evaluating the raw responses to construct preference data (Sec.3.2), and preference optimization (Sec.3.3). The details are elaborated in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.871, + 0.79 + ], + "angle": 0, + "content": "3.1. Grounded Region Description Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.903 + ], + "angle": 0, + "content": "The first step of our pipeline is to collect diverse model responses that will later be ranked to construct preference data. Since our primary objective is to enhance MLLMs' localization capabilities and achieve fine-grained alignment to visual inputs, we choose the task of region description with grounding to evaluate MLLMs' object localization capabilities. However, existing datasets [23, 57] for region" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "722" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.087, + 0.887, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.329 + ], + "angle": 0, + "content": "Figure 2. We leverage the generated object references and cropped image region to build a variety of multimodal prompts to enhance the diversity of generated region descriptions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.484, + 0.49 + ], + "angle": 0, + "content": "descriptions are often too simple, involving queried regions with only one or two objects and short phrases such as 'vehicles parked on the street' or 'bicycles are parked on the sidewalk.' Such simple prompts are inadequate for generating diverse responses to construct preferred and rejected preference data with sufficient divergence, which might hinder the effectiveness of DPO training [54]. To address this issue, we generate queried regions from scratch instead of using existing region description datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Region Query Construction. We design a simple approach to generate randomly queried regions based on images and object annotations. Take the Objects365 dataset as an example. We first filter out images with few objects, ensuring that the data contains rich visual content. Then, given the annotated object bounding boxes in each image, we randomly select one of the objects as the starting region. From there, we iteratively expand the region by incorporating the nearest objects. The expansion stops randomly once more than four objects are involved in the region. The resulting region then prompts MLLMs to generate a detailed region description. Through this process, we simulate the human-like, dynamic attention across different parts of an image, encouraging the MLLM to adaptively focus on arbitrary image regions based on the given prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Grounded Region Description Generation. As shown in Fig. 2, we build a variety of prompts for MLLMs to generate several region descriptions for each image, serving as candidate responses for preference data. Since the original MLLM sometimes struggles to generate detailed responses following region prompts, we utilize cropped region images along with object references constructed from annotations to guide MLLMs to attend to the region's content and details. These prompts help the model focus more effectively on the specified region and produce detailed descriptions that might better align with human preference. In this way, we encourage the MLLM to generate responses that are dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.385 + ], + "angle": 0, + "content": "tinct in content but consistent in language style, which is then used for constructing preference data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.397, + 0.882, + 0.413 + ], + "angle": 0, + "content": "3.2. Preference Data Ranking and Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.585 + ], + "angle": 0, + "content": "The next step is to rank the generated descriptions to obtain preferred and rejected data pairs. An ideal region description should meet at least two key criteria: (1) the text description should accurately match the semantics of the queried region and the surrounding image content, (2) it should provide detailed descriptions with accurate localization of objects within the region. To address these two criteria, we propose a semantic score and localization score to rank the responses. The descriptions with the highest and the lowest scores are paired to form preference data for DPO training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Semantic Score. We introduce the semantic score to evaluate the relevance between the generated descriptions and the semantics of queried image regions. We leverage a pretrained CLIP model [42] to compute the cosine similarities of text and visual embeddings as defined in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.679, + 0.905, + 0.695 + ], + "angle": 0, + "content": "\\[\nS (I, T) = \\alpha * \\cos \\left(\\mathcal {F} _ {\\text {r e g i o n}} (I), \\mathcal {F} _ {\\text {t e x t}} (T)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Where \\(\\alpha\\) is the scale of similarities, which is set as 5 in our work to balance the range of the semantic score, \\(I\\) and \\(T\\) are the input image and MLLM generated region description with grounding text removed; \\(\\mathcal{F}_{region}\\) and \\(\\mathcal{F}_{text}\\) denotes the visual embedding for image region and text embeddings, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "When extracting image region embeddings, a straightforward approach is to crop the image region \\( I_{crop} \\) and directly extract visual embeddings. However, the similarity score with such embedding tends to overly focus on the region's details while neglecting the image's surrounding context. To address this limitation, we supplement it with similarities \\( S_{local} \\) from visual embeddings of intact images" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "723" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "that incorporate local attention. Specifically, we feed the original image into CLIP and replace the final layer of the vision encoder that aggregates the embeddings with a local-attention layer. This modification allows the model to better account for the context around the region of interest. As defined in Eq. (2), we then use the average of the cropped image's similarity score and the full image's similarity score with local attention as the final semantic score, which effectively evaluates the extent of fine-grained alignment between the region description and local visual semantics." + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.254, + 0.484, + 0.284 + ], + "angle": 0, + "content": "\\[\nS _ {s e m} = \\frac {1}{2} \\left(S \\left(I _ {c r o p}, T\\right) + S _ {l o c a l} (I, T)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.485, + 0.529 + ], + "angle": 0, + "content": "Localization Score. We propose a localization score to evaluate how detailed the MLLM responds in describing objects within the queried region and its grounding accuracy. This score is calculated based on the number of objects mentioned in the description that match the ground truth objects in the region. In practice, we use Grounding DINO [34] and the cropped image region to extract bounding boxes for objects mentioned in the description. The extracted objects are then combined with the original object annotations to form a set of ground truth objects within the region. Next, we extract the grounding results from MLLM-generated descriptions and combine them with the results from Grounding DINO to form the predicted objects. Finally, we compute the average IoU between the predicted objects and the ground truth as the localization score. The detailed process is outlined in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.606 + ], + "angle": 0, + "content": "The localization score encourages the model to include more detailed descriptions for involved objects and accurately localize them in its responses. Finally, we combine the semantic and localization scores for each grounded region description:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.621, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\nS = \\lambda S _ {s e m} + (1 - \\lambda) S _ {l o c} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to 0.8 in our implementation. Then, the descriptions with the highest and lowest scores are paired as preferred and rejected data for preference optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Grounded Region Description Refinement. After obtaining the preference data pairs, we further enhance the divergence of the grounding results of the preferred and rejected descriptions to encourage the model to distinguish between accurate and inaccurate object localization. To achieve this, we refine the grounding results in the preferred descriptions while keeping the rejected ones unchanged. In practice, we leverage the results obtained while computing the localization score, including the object box predictions \\( B_{pred} \\) and ground-truth object boxes \\( B_{gt} \\). We retain only those predictions that match the ground truths (IoU > 0.5) and replace their bounding boxes with the matched ones. Then, we remove duplicates of predictions based on their textual position in the description and IoUs. Finally, we insert the" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.092, + 0.809, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Computing Localization Score" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.11, + 0.905, + 0.156 + ], + "angle": 0, + "content": "Input: Cropped Image Region \\( I_{crop} \\), Grounded Region Description \\( T \\) generated by MLLMs, Object Bounding Box Annotations \\( B_{anno} \\) for the Queried Region." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.156, + 0.747, + 0.171 + ], + "angle": 0, + "content": "Output: Localization Score: \\( S_{loc} \\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.171, + 0.905, + 0.202 + ], + "angle": 0, + "content": "1: Extract bounding boxes \\( B_{text} \\) from the description \\( T \\) and get the plain text \\( T_{plain} \\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.202, + 0.905, + 0.232 + ], + "angle": 0, + "content": "2: Leverage Grounding DINO to get grounded object results \\( B_{\\text{ground}} \\) from \\( T_{\\text{plain}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.232, + 0.905, + 0.275 + ], + "angle": 0, + "content": "3: Get the set of ground truth object boxes \\( B_{gt} \\) by aggregating \\( B_{ground} \\) and \\( B_{anno} \\) and removing duplicated boxes." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.277, + 0.905, + 0.322 + ], + "angle": 0, + "content": "4: Get the set of object box predictions \\( B_{pred} \\) for the description \\( T \\) by aggregating \\( B_{ground} \\) and \\( B_{text} \\) and removing duplicated boxes." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.321, + 0.905, + 0.353 + ], + "angle": 0, + "content": "5: Computing IoU matrix \\(\\mathbf{m}[i,j] = IoU(B_{gt}^i, B_{pred}^j)\\) between \\(B_{gt}\\) and \\(B_{pred}\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.353, + 0.777, + 0.368 + ], + "angle": 0, + "content": "6: Filter the IoU by a threshold of 0.5." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.171, + 0.905, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.376, + 0.838, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} [ i, j ] = \\left\\{ \\begin{array}{l l} 0 & \\mathbf {m} [ i, j ] < 0. 5 \\\\ \\mathbf {m} [ i, j ] & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.424, + 0.766, + 0.449 + ], + "angle": 0, + "content": "7: return \\( S_{loc} = \\frac{1}{n}\\sum_{i}^{n}\\max_{j}\\mathbf{p}[i,j] \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.525 + ], + "angle": 0, + "content": "refined object box predictions into the region description, resulting in an improved grounded region description with more precise coordinates." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.74, + 0.55 + ], + "angle": 0, + "content": "3.3. Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.601 + ], + "angle": 0, + "content": "After curating the preference dataset, we finetune MLLMs through DPO and adopt LORA to save the training cost. The loss for optimizing MLLMs is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.623, + 0.907, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right)} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {*} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {*} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.751 + ], + "angle": 0, + "content": "where \\( y_{w} \\) and \\( y_{l} \\) are the preferred and rejected description data; \\( \\pi_{r}ef(y|x) \\) is the base reference policy model, i.e., the initial instruction-tuned MLLM which is frozen during the training; \\( \\pi_{*}(y|x) \\) denotes the policy model which inherits from the instruction-tuned model with its LORA weights updated in the training process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.764, + 0.646, + 0.78 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.701, + 0.805 + ], + "angle": 0, + "content": "4.1. Experiment Setups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. In this work, we experiment with the proposed SPR with three MLLMs with spatial understanding capabilities, including Ferret [58], LLaVOneVision [24], and CogVLM-Grounding [52]. To construct preference data, we randomly select 10k images with object annotations from the training set of Objects365" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "724" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. Experiments on the Referring Expression Comprehension task (Acc@0.5) on datasets RefCOCO/+/g, and the Phrase Grounding task (Recall@1) on Flickr30k Entities dataset. “-” indicates results are unavailable or that MLLMs do not support multi-object grounding." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.13, + 0.845, + 0.392 + ], + "angle": 0, + "content": "
MethodRefCOCORefCOCO+RefCOCOgFlickr30k Entities
valtestAtestBvaltestAtestBvaltestvaltest
UNITER [12]81.4187.0474.1775.9081.4566.7074.0268.67--
UniTAB [55]86.3288.8480.6178.7083.2269.4879.9679.9778.7679.58
MDETR [21]86.7589.5881.4179.5284.0970.6281.6480.8982.383.8
MiniGPT-v2-7B [70]88.0691.2984.3079.5885.5273.3284.1984.31--
VistaLLM [41]88.191.583.082.989.874.883.684.4--
LLaVA-Grounding [64]89.16--81.68--84.82-83.0383.62
Shikra-7B [10]87.0190.6180.2481.6087.3672.1282.2782.1975.8476.54
Shikra-13B [10]87.8391.1181.8182.8987.7974.4182.6483.1677.4178.44
Griffon-13B [60]89.492.584.683.388.476.085.186.183.784.2
LLava-OV-7B [24]74.7782.5964.0470.1779.8558.4872.3471.39--
+ SPR76.6682.5265.9771.6279.8759.9972.9871.55--
Ferret-7B [58]87.4991.3582.4580.7887.3873.1483.9384.7680.3982.21
+ SPR88.3991.6783.9182.0787.8474.1985.5885.7581.5383.34
Ferret-13B [58]89.4892.4184.3682.8188.1475.1785.8386.3481.1384.76
+ SPR89.9493.0685.1283.2988.8975.7486.4686.9281.8283.75
CogVLM-Grounding-17B [52]92.7694.7588.9988.6892.9183.3989.7590.79--
+ SPR92.9594.8789.1588.8392.9583.8490.0190.96--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.403, + 0.486, + 0.446 + ], + "angle": 0, + "content": "Table 2. Experiments on Referring Expression Comprehension task under different IoU thresholds. The results are the average on RefCOCO, RefCOCO+, and RefCOCOg datasets." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.455, + 0.46, + 0.54 + ], + "angle": 0, + "content": "
IoU Threshold0.50.60.70.80.9
Ferret-7B83.9181.2876.7267.0243.25
+SPR84.9382.3678.4270.0952.21
Ferret-13B85.5682.9478.5770.0449.55
+SPR86.1883.6379.9372.0353.61
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.565, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Dataset [46], then construct random regions to query models to generate grounded region descriptions. We adopt LORA [17] for tuning MLLMs. The training is conducted on one A100 GPU, which takes around 3 and 5 hours for Ferret 7B and 13B models, respectively. Please refer to Appendix for more details on preference data construction and hyperparameter selection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Evaluation Benchmarks We evaluate our method on three types of benchmarks: (1) Grounding tasks that evaluate the localization accuracy, including referring expression comprehension (REC) and phrase grounding; (2) Region description task on Refcocog [22] and visual genome [23], and Ferret Bench [58] for comprehensive spatial understanding; (3) General benchmarks TextVQA [47], GQA [19], LLaVA-Bench [33], and hallucination benchmark POPE [56]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.292, + 0.835 + ], + "angle": 0, + "content": "4.2. Experiments on REC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We first evaluate our method on the referring expression comprehension (REC) task on RefCOCO [22], RefCOCO+ [22], and Refcocog [36]. The task requires the model to locate the object or region given a short de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.908, + 0.618 + ], + "angle": 0, + "content": "scription, which evaluates the model's fine-grained visual grounding abilities under the single-object referent scenarios. As shown in Tab. 1, our proposed SPR framework consistently improves the performance of three baseline MLLMs on different datasets for all model sizes. Considering that the REC results are based on an IoU threshold of 0.5, the improvement on its performance indicates that the model localized more objects successfully. Hence, this improvement can be largely attributed to the introduction of localization scores when constructing the preference data in SPR. Region descriptions that accurately mention more objects could achieve higher localization scores in SPR and be more likely to serve as preferred data, facilitating the model to attend to more objects and their locations in the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To better evaluate the impact of SPR on the localization capability of MLLMs, we also conduct REC experiments with higher IoU thresholds by gradually increasing the IoU threshold of valid REC results from the default value of 0.5 to 0.9. As shown in Tab. 2, the improvements brought by SPR significantly increase as the threshold rises, with accuracy gains of 8.96 and 4.06 for the 7B model and the 13B model, respectively, when the IoU threshold rises to 0.9. With SPR, the localization accuracy of the objects in the model's response is greatly improved. Equipped with the grounded region description refinement and the supervision of preferred-rejected localization data in SPR, the model can respond more accurately to grounding object locations, demonstrating the effectiveness of incorporating preference optimization for region description and object localization in the fine-grained spatial understanding of MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "725" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.132 + ], + "angle": 0, + "content": "Table 3. Experiments on Phrase Grounding task under different IoU thresholds. The results are averaged over the validation and test set of the Flickr30k dataset." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.143, + 0.462, + 0.227 + ], + "angle": 0, + "content": "
IoU Threshold0.50.60.70.80.9
Ferret-7B81.376.1467.5553.8629.98
+SPR82.4477.1469.2556.1933.99
Ferret-13B82.9476.6268.3455.7432.60
+SPR82.7877.2369.7456.9634.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.241, + 0.483, + 0.27 + ], + "angle": 0, + "content": "Table 4. Experiments on the region captioning task on Refcocog and Visual Genome datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.28, + 0.479, + 0.377 + ], + "angle": 0, + "content": "
MethodRefcocogVisual Genome
METEORROUGE_LMETEORROUGE_L
Ferret-7B12.315.617.429.6
+SPR13.520.417.629.7
Ferret-13B12.926.417.931.0
+SPR13.327.218.231.3
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.402, + 0.398, + 0.42 + ], + "angle": 0, + "content": "4.3. Experiments on Phrase Grounding" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.483, + 0.637 + ], + "angle": 0, + "content": "Furthermore, we experiment with the phrase grounding task on Flickr30k Entity [40]. In phrase grounding, the queried object phrases are combined in a single question, requiring MLLMs to detect the locations of multiple objects in a single response, which makes it more challenging than the single-object referring task like REC. Following [58], we adopt the question \"What are the locations of [phrases]?\" and evaluate the result using the MERGE-BOXES mode [21]. Since LLaVA-OneVision and CogVLM do not support multi-object detection, we report only the results for Ferret. As shown in Tab. 1, SPR effectively improves Ferret's performance in multi-object referent scenarios, especially for the 7B model, whose performance is even comparable to that of the 13B model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.835 + ], + "angle": 0, + "content": "We then experiment with the phrase grounding task under higher IoU thresholds. We found that the multi-object referencing setting in phrase grounding is more challenging than the single-object referencing in REC. As the IoU threshold increases, the performance drops more rapidly, indicating a significant demand for MLLM to improve the capabilities of more accurate localization. Our approach can significantly alleviate this issue. As shown in Tab. 3, SPR improves progressively with higher IoU thresholds, reaching a maximum gain of 4.01 and 1.58 Recall@1 for the 7B and 13B models, respectively. This experiment demonstrates the superiority of SPR in pursuing detailed descriptions with high-precision object localization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.4, + 0.866 + ], + "angle": 0, + "content": "4.4. Experiments on Region Captioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Beyond the grounding task, we also verify our proposed SPR in improving the text qualities of MLLMs' outputs" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 5. Experiments on the Ferret Bench. \"Description\", \"Reasoning\", and \"Grounding\" denote the Referring Description, Referring Reasoning, and Grounding in Conversation tasks." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.143, + 0.888, + 0.227 + ], + "angle": 0, + "content": "
ModelDescriptionReasoningGroundingAvg.
Ferret-7B68.767.357.564.5
+ SPR70.068.458.165.5
Ferret-13B70.668.759.766.3
+ SPR70.872.660.267.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.239, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Table 6. Experiments on the general and hallucination benchmarks. We report the accuracy for GQA and VQA and the F1 score on POPE." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.292, + 0.871, + 0.377 + ], + "angle": 0, + "content": "
ModelVQATGQALLaVAPOPE
Ferret-7B--64.785.36
+ SPR--66.385.69
LLaVA-OV-7B75.8962.2188.988.12
+ SPR76.0762.4291.488.49
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.553 + ], + "angle": 0, + "content": "on fine-grained spatial understanding. We conduct experiments on the RefCOCOg and Visual Genome benchmarks. We prompt MLLMs with the question \"Describe the region [region] in the image.\" to generate region captions and then evaluate the response quality using METEOR [5] and ROUGE_L [28] metrics. Tab. 4 shows that SPR effectively improves the quality of MLLM-generated region captions. After tuning with SPR, MLLMs are able to effectively attend to the user-specified regions and generate captions that better reflect the details of the region content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.779, + 0.577 + ], + "angle": 0, + "content": "4.5. Experiments on Ferret Bench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Ferret benchmark, proposed by [58], aims to evaluate MLLMs' fine-grained multimodal conversational capabilities such as referring description, referring reasoning, and grounded conversation. We follow the pipeline in [58] to prompt MLLMs with questions and employ GPT to evaluate the responses. As shown in Tab. 5, the proposed SPR can facilitate MLLMs in achieving better conversational qualities for fine-grained multimodal understanding, especially for the referring reasoning task, with an accuracy gain of about 3.9 for the 13B model. Equipped with SPR, MLLM can focus on more detailed visual information and generate responses that align better with human preferences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.842, + 0.79 + ], + "angle": 0, + "content": "4.6. Experiments on General Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We further evaluate SPR on three general benchmarks to validate the benefits of improving MLLMs' spatial capabilities. TextVQA and GQA require MLLMs to answer questions or perform reasoning based on specific text, objects, and image content. LLaVA bench evaluates MLLMs comprehensive capabilities in conversation, description, and reasoning. As shown in Tab. 6, improving MLLMs' spa" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "726" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 7. Ablation Studies on the refinement of grounded region descriptions, and ratio \\(\\lambda\\) between semantic and localization scores in ranking MLLM responses. We report the average results on Referring expression comprehension and phrase grounding tasks." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.157, + 0.432, + 0.289 + ], + "angle": 0, + "content": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ SPR84.9382.44
w/o Refinement84.4191.38
λ = 0.084.2581.83
λ = 0.484.3481.95
λ = 0.684.6682.13
λ = 0.884.9382.44
λ = 1.084.4581.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.3, + 0.435, + 0.315 + ], + "angle": 0, + "content": "Table 8. Ablation studies on the training strategy." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.326, + 0.442, + 0.391 + ], + "angle": 0, + "content": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ Instruction Finetuning84.3581.72
+ DPO training84.9382.44
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.485, + 0.569 + ], + "angle": 0, + "content": "tial understanding capabilities consistently enhances their comprehension and reasoning abilities across diverse general scenarios, leading to performance gains on all three benchmarks. We also experiment on hallucination benchmark POPE, where SPR improves both Ferret and LLaVA-OneVision. This can be attributed to the preference data construction in SPR, where semantic and localization scores are applied to select region descriptions that better align with region content and reject those related to content outside the region or that contain hallucinations, thus effectively helping mitigate the hallucinations in MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.255, + 0.592 + ], + "angle": 0, + "content": "4.7. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.645 + ], + "angle": 0, + "content": "We conduct ablation studies over the two designs in SPR and evaluate the performance of Ferret-7B on the REC (Refcoco \\(+ / \\mathrm{g}\\)) and the Flickr30k phrase grounding tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Score Ratio \\(\\lambda\\). In Sec. 3.2, we combine the semantic and localization scores to rank MLLMs generated descriptions with a score ratio \\(\\lambda\\). As shown in Tab. 7, we vary the \\(\\lambda\\) from 0 to 1, and the trained models outperform the baseline model consistently. When \\(\\lambda\\) equals zero, SPR achieves minimal gain, as the model might overly reward descriptions that simply list object names or fail to align with the region's semantics. On the other hand, when \\(\\lambda\\) is set to 1, SPR completely disregards measuring how detail the MLLM describes the objects in the region. Under such situations, the model encourages coarse region descriptions with fewer objects involved and reduces the corresponding object bounding box texts in the preferred data, thereby hindering the training of MLLMs' localization capability. As the experiments show, a relatively high value of 0.8 achieves the best results and is set as the default value in SPR." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Refinement of Grounded Region Description. After con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "structuring the preferred and rejected data pairs, we further refine the localization results in the preferred descriptions by completing bounding boxes for objects in the description that were not grounded and refining the existing bounding boxes. Tab. 7 shows the results of this refinement. We found that the refinement leads to greater improvements in the multi-object referring task of phrase grounding. This could be attributed to the fact that the baseline model often fails to follow instructions for providing bounding boxes for each mentioned object when generating region descriptions. After tuning by the refined descriptions, MLLMs could faithfully ground the mentioned objects, thereby improving the multi-object phrase grounding clearly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.722, + 0.314 + ], + "angle": 0, + "content": "4.8. Comparison with SFT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.908, + 0.593 + ], + "angle": 0, + "content": "In this paper, we adopt DPO with accept-reject preference data to optimize MLLMs for spatial understanding, whereas prior work [10, 44, 61], primarily focuses on the stage of supervised instruction fine-tuning (SFT). In Tab. 8, we compare these two training approaches, where SFT is trained using only the accepted data. The results show that while SFT could improve MLLMs' localization capabilities, its performance gains are significantly lower than DPO. DPO optimizes MLLM by contrasting accepted and rejected data pairs, similar to the positive-negative sample training mechanism in traditional object detection algorithms. This approach helps models distinguish between accurate and inaccurate localizations and facilitates MLLM in spatial understanding more effectively. However, it is important to note that DPO training also relies on a well-trained SFT model as a foundation, making these two approaches complementary. In future work, we will further explore how to integrate SFT and DPO to enhance MLLMs' spatial understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.634, + 0.62 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose SPR, a Spatial Preference Rewarding framework to enhance MLLM's fine-grained spatial understanding capabilities. We introduce a complete pipeline that includes (1) Constructing random region queries; (2) Prompting MLLMs to generate diverse grounded region descriptions; (3) Proposing semantic scores and localization scores to rank the descriptions comprehensively; (4) Refining the localization quality of preference data; (5) Finetuning MLLMs to optimize against detailed and accurate spatial understanding. The entire framework does not require additional human labor or external MLLMs, with minimal overhead on training costs. SPR addresses the lack of direct optimization for positive and negative localization samples in MLLM training, enhancing their localization capabilities and promoting better alignment with human preferences. Experiments demonstrate that SPR significantly improves MLLMs' performance on standard referring and grounding tasks for spatial understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "727" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.175 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.175, + 0.483, + 0.205 + ], + "angle": 0, + "content": "This study is also supported by the MOE Tier-2 project, with project number MOE-T2EP20123-0003." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.23, + 0.188, + 0.246 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.256, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[4] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with Transformers. In ECCV, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.256, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[12] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[13] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[16] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[18] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.832 + ], + "angle": 0, + "content": "[21] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "728" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[26] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[27] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundingpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[30] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[35] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[38] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. arXiv preprint arXiv:2408.10433, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[39] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[41] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[44] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12987, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[46] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[47] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Con-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "729" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[50] Fei Wang, Wenxuan Zhou, James Y Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. arXiv preprint arXiv:2406.11839, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[52] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[54] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. beta-dpo: Direct preference optimization with dynamic beta. arXiv preprint arXiv:2407.08639, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[56] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Zhao Wayne Xin, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[57] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[58] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[59] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. arxiv, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[60] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[61] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[62] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[63] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.374, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[64] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.445, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[65] Mengxi Zhang and Kang Rong. Automated multi-level preference for mllms. arXiv preprint arXiv:2405.11165, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[66] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6578-6587, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[68] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.725 + ], + "angle": 0, + "content": "[69] Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622, 2024. 1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[70] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "730" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_origin.pdf b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..66477c6450b7b9fa8bb1a8f2037eb5785ec2615b --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/7f1d9c24-f4e8-482b-91ef-4b03709c5ebb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ffe2b314ab2743f806d3754c671c5f2c6bf7ec3eb7fd02a866f8c1bf34ee4e7 +size 8359488 diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/full.md b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3260ea3ccc7b9a53361254d6df528ead39c82911 --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/full.md @@ -0,0 +1,338 @@ +# Spatial Preference Rewarding for MLLMs Spatial Understanding + +Han Qiu $^{1}$ , Peng Gao $^{2}$ , Lewei Lu $^{3}$ , Xiaqin Zhang $^{4}$ , Ling Shao $^{5}$ , Shijian Lu $^{1*}$ + +$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ Shanghai AI Laboratory $^{3}$ Sensetime Research, $^{4}$ Zhejiang University of Technology + +$^{5}$ UCAS-Terminus AI Lab, University of Chinese Academy of Sciences + +han023@e.ntu.edu.sg, Shijian.Lu@ntu.edu.sg + +# Abstract + +Multimodal large language models (MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding (SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR. + +# 1. Introduction + +Multimodal large language models (MLLMs) [3, 15, 18, 31, 32, 51, 58, 70] have achieved remarkable success by inte- + +grating pretrained large language model [2, 14, 49] with vision encoders [8, 37, 42], leading to significant advancements in a wide range of general vision-language tasks. By combining visual and language signals, MLLMs have demonstrated superior capabilities in multimodal understanding, reasoning, and interaction as compared with traditional vision models. Recently, several studies have further injected spatial knowledge into MLLMs, thereby improving MLLMs' fine-grained perception of visual inputs and enabling tasks such as referential dialogue [10, 16], grounding captioning [35, 58, 62], region description [31, 52], and object detection [61], etc. These advances have paved the way for MLLMs to serve as versatile visual assistants supporting a wider range of applications. + +Despite recent advancements, MLLMs still face challenges in fine-grained spatial understanding, with responses not aligned with human preferences. As illustrated in 1, the generated grounded region descriptions are often vague with inaccurate object localizations, and models may fail to focus on the queried region, distracted from other regions in the image. The issue in spatial understanding could be attributed to the lack of positive-negative preference feedback in existing instruction-tuned MLLMs. Specifically, instruction fine-tuning (SFT) directly optimizes MLLMs to mimic ground truth positive samples, but it cannot impose any penalties if the model produces inaccurate negative samples for localization during actual inference. As a result, MLLMs may struggle to generate positive descriptions with accurate object localization and instead produce negative and inaccurate descriptions, leading to responses that do not align with user expectations. In addition, optimization using positive and negative samples has been proven crucial for spatial understanding in traditional object detection algorithms [7, 29, 45], highlighting a significant gap in the current MLLM training on spatial understanding. + +Several studies [38, 48, 50, 59, 65, 69] attempt to introduce preference optimization for better MLLM alignment, where the preference data are constructed by collect + +![](images/6ecce25348467d0b662edc637ca4686592b5c5f5e4c1881e8c418ea9b917eb06.jpg) +User Prompt about Regions in Images: +MLLM Response: + +Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object. + +![](images/ab0ab80098fa3b1f68e0bff559b8f3ead7d3acf9823acc9f234ed7153636732b.jpg) + +Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object. + +The area [676, 229, 847, 1000] is occupied by a woman in a blue shirt who is holding a whistle [537, 530, 568, 606]. She is also wearing a black clipboard [801, 538, 825, 581] around her neck .... + +Fail to Ground + +![](images/cbb76db1ef6c4cb5e22b0e979ca279eb75aa5bc2962748530c13bf7e0a1d3843.jpg) + +The region [638, 157, 775, 709] is a trail located on a mountain [166, 365, 996, 996]. of people [619, 250, 799, 990] walking on it. The trail also has a marker [166, 339, 229, 439] to guide hikers. Inaccurate Localization + +![](images/69c00219318dcdb064f0d69d1853b96aa7de07ffad9584a65699ac43e62540df.jpg) + +Tuning with Spatial Preference Rewarding (SPR) + +The area [676, 229, 847, 998] in the image features a woman in a blue shirt with a name tag, .... She [676, 229, 817, 901] appears to .... The grassy area where she is standing [0, 385, 998, 998] seems to be .... activities. There's also a red cone [676, 909, 720, 998] .... + +![](images/5310376b17b7b49b037e7ebc730e6eaecf0df6b387ac830d8c7570c712df2ee1.jpg) +Responses from SPR-empowered MLLM: +Figure 1. The proposed Spatial Preference Rewarding (SPR) mitigates the distracted and inaccurate region descriptions generated by MLLMs. Given an image and a user-specified region of interest, MLLMs often fail to focus on the queried region. They may be distracted by objects outside the specified region, failing to ground the queried objects, or providing inaccurate localization. Tuning MLLMs with our proposed SPR leads to more accurate object localization and detailed object descriptions. + +The region [638, 157, 775, 709] is where a man [575, 150, 787, 993] is standing.... The water bottle [732, 321, 796, 481] suggests that he is well-equipped for his journey. Moreover, his grey jacket [575, 214, 720, 658] indicates that he is ready for the outdoor activity. + +![](images/ad9d796c0ce9fd398c59abbb0cf3aadad020b7a8d46e09fefaf97857fe32e153.jpg) + +ing MLLM-generated image descriptions and scoring them by human or LLMs. However, these methods primarily leverage preferences to improve image-level coarse alignment, and most of them target mitigating hallucinations in MLLMs. The problem of fine-grained alignment for spatial understanding, such as detailed region descriptions and accurate object localization, has been largely neglected. + +To address this gap, we design SPR, a Spatial Preference Rewarding framework that enhances MLLM spatial understanding capabilities by rewarding detailed responses with accurate object localization over vague or inaccurate responses. Specifically, SPR selects random image regions containing multiple objects and prompts MLLMs in diverse ways to generate grounded region descriptions. In reward modeling, it introduces both semantic and localization scores to evaluate the alignment between the region description and the region semantics, as well as how detailed region objects are described. We also refine the grounded object in the generated description to enhance its localization accuracy. Finally, the best-scored refined description and the response of the lowest score are paired as preferred and rejected data for direct preference optimization (DPO) [43] training with LORA [17]. By aligning MLLMs with detailed and accurate responses, SPR mitigates MLLMs' incompetence in accurate localization and spatial understand + +ing as required in many real-world tasks. + +We validate the effectiveness of SPR in enhancing MLLMs' spatial understanding capabilities with minimal overhead in training. Compared to the baseline, SPR enhances MLLMs on both referring and grounding benchmarks, especially under higher IoU thresholds which demand higher localization accuracy. In addition, SPR can improve MLLM trustworthiness and reduce MLLM hallucinations as well. Our experiments highlight the importance of incorporating preference-based feedback to enhance the fine-grained spatial understanding abilities in MLLMs. + +The contributions of this work are summarized as follows: + +- We propose a Spatial Preference Rewarding (SPR) framework to enhance the fine-grained spatial understanding of MLLMs via direct preference optimization (DPO), enhancing MLLMs' capabilities in precise region referring and accurate object localization in images. +- We develop an automated pipeline that creates preference data by constructing random region prompts and scoring model responses for spatial understanding. The pipeline requires no other MLLMs or human labours, making it scalable in future training. +- Extensive experiments show that the proposed SPR improves MLLMs' spatial understanding capabilities consistently across multiple public benchmarks. + +# 2. Related Work + +Recently, the success of large language models (LLMs) [14, 49, 51] has been extended into the multimodal domain, resulting in models that demonstrate impressive performance in integrating vision and language [1, 15, 20, 32, 70]. These models treat visual signals as a special form of language, establishing multimodal understanding, reasoning, and interaction capabilities by combining visual encoders [37, 42] with pre-trained large language models, or by directly feeding encoded visual signals into LLMs [6, 53]. Most current MLLMs follow a two-step training process. The first step is pre-training, where large-scale vision-language datasets [9] are used to align visual features to the same space as language features. This enables the model to bridge visual and language embeddings effectively. The second step involves instruction-following finetuning, where high-quality vision-language datasets [25, 33, 70] are used to further enhance the MLLMs' capabilities to follow user instructions and comprehend multimodal information. These methods often convert existing datasets into an instruction-following format or adopt leading MLLMs like GPT to generate high-quality training instruction data for MLLMs [11, 33]. Despite their success, current MLLMs still face challenges that may generate undesired responses toward human preferences. For instance, these models are prone to generating hallucinated content [4, 30, 56, 59] or providing responses that do not fully meet user expectations. Improving the quality of MLLM responses and aligning them more closely with user preferences has thus become a surging focus of research in the community. Our work aims to improve the spatial understanding capabilities of MLLMs, aligning their behaviors better with human preferences. + +MLLMs for Spatial Understanding. Spatial understanding capabilities [7, 13, 22, 67], such as object detection, referring, and grounding description tasks, have long been a fundamental research topic in the field of computer vision. Recent efforts attempt to empower MLLMs with dense visual perception and spatial understanding abilities by integrating region-level data in MLLM training or modifying MLLM architectures. For example, Kosmos-2 [39] and Shikra [10] directly represent the object coordinates in text, constructing instruction datasets to inject spatial knowledge into MLLMs. LLava-Grounding [63] and GroundingGPT [27] construct large-scale grounding datasets to enhance multimodal grounding capabilities. To better facilitate localization within images, RegionGPT [16], GPT4ROI [66], Ferret [58], and Groma [35] encode region features as direct inputs to LLMs, facilitating explicit attention to specific image regions. The Griffon [60, 61] series focuses on dense detection, enabling MLLMs to achieve performance comparable to traditional object detectors. LocVLM [44] explores the. LocVLM [44] + +explores key factors in instruction tuning for spatial understanding, such as coordinate representation, which improves MLLM's spatial awareness. However, these efforts primarily concentrate on the instruction-tuning phase and lack direct feedback on MLLMs' responses. To fill this gap, we propose a Spatial Preference Rewarding (SPR) framework, which constructs preference data based on MLLMs generated grounded region descriptions for MLLM tuning. + +Preference Optimization for MLLMs. Preference alignment has recently emerged as a promising direction to align model responses with human preferences. One widely explored approach is to employ Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) to improve the trustworthiness of MLLMs and reduce hallucinations in their responses. For example, LLaVA-RLHF [48] and RLHF-V [59] leverage human annotators to evaluate model responses and construct preference data for fine-tuning. POVID [68] and Silkie [26] use external models, such as GPT, as evaluators to build preference datasets. CLIP-DPO [38] and CSR [69] use CLIP to rank model responses to avoid resource-intensive human or MLLM annotations. AMP [65] introduced a multi-level preference framework to enable MLLMs to better model differences between preference data. mDPO [50] introduced additional preference data pairs with corrupted images to avoid over-optimization on language-only preferences. Unlike these existing studies that primarily aim to reduce hallucinations in MLLMs, our proposed SPR framework focuses on optimizing MLLM responses related to spatial reasoning and understanding. Specifically, SPR focuses on fine-grained alignment with visual inputs and facilitates MLLMs in distinguishing between high-quality object localization (positive samples) and inaccurate localization (negative samples), thereby improving the spatial understanding capabilities of MLLMs. + +# 3. Methods + +This section presents our proposed Spatial Preference Rewarding (SPR) framework. Following a typical DPO pipeline, SPR adopts a three-step process in MLLM finetuning, including collecting MLLMs' raw responses (Sec.3.1), evaluating the raw responses to construct preference data (Sec.3.2), and preference optimization (Sec.3.3). The details are elaborated in the following subsections. + +# 3.1. Grounded Region Description Generation + +The first step of our pipeline is to collect diverse model responses that will later be ranked to construct preference data. Since our primary objective is to enhance MLLMs' localization capabilities and achieve fine-grained alignment to visual inputs, we choose the task of region description with grounding to evaluate MLLMs' object localization capabilities. However, existing datasets [23, 57] for region + +![](images/5accb4f5af8f7bbedc29d2987ae4baba93d938c8a84f49ae1eafd1114988169e.jpg) +Figure 2. We leverage the generated object references and cropped image region to build a variety of multimodal prompts to enhance the diversity of generated region descriptions. + +descriptions are often too simple, involving queried regions with only one or two objects and short phrases such as 'vehicles parked on the street' or 'bicycles are parked on the sidewalk.' Such simple prompts are inadequate for generating diverse responses to construct preferred and rejected preference data with sufficient divergence, which might hinder the effectiveness of DPO training [54]. To address this issue, we generate queried regions from scratch instead of using existing region description datasets. + +Region Query Construction. We design a simple approach to generate randomly queried regions based on images and object annotations. Take the Objects365 dataset as an example. We first filter out images with few objects, ensuring that the data contains rich visual content. Then, given the annotated object bounding boxes in each image, we randomly select one of the objects as the starting region. From there, we iteratively expand the region by incorporating the nearest objects. The expansion stops randomly once more than four objects are involved in the region. The resulting region then prompts MLLMs to generate a detailed region description. Through this process, we simulate the human-like, dynamic attention across different parts of an image, encouraging the MLLM to adaptively focus on arbitrary image regions based on the given prompts. + +Grounded Region Description Generation. As shown in Fig. 2, we build a variety of prompts for MLLMs to generate several region descriptions for each image, serving as candidate responses for preference data. Since the original MLLM sometimes struggles to generate detailed responses following region prompts, we utilize cropped region images along with object references constructed from annotations to guide MLLMs to attend to the region's content and details. These prompts help the model focus more effectively on the specified region and produce detailed descriptions that might better align with human preference. In this way, we encourage the MLLM to generate responses that are dis + +tinct in content but consistent in language style, which is then used for constructing preference data. + +# 3.2. Preference Data Ranking and Construction + +The next step is to rank the generated descriptions to obtain preferred and rejected data pairs. An ideal region description should meet at least two key criteria: (1) the text description should accurately match the semantics of the queried region and the surrounding image content, (2) it should provide detailed descriptions with accurate localization of objects within the region. To address these two criteria, we propose a semantic score and localization score to rank the responses. The descriptions with the highest and the lowest scores are paired to form preference data for DPO training. + +Semantic Score. We introduce the semantic score to evaluate the relevance between the generated descriptions and the semantics of queried image regions. We leverage a pretrained CLIP model [42] to compute the cosine similarities of text and visual embeddings as defined in Eq. (1): + +$$ +S (I, T) = \alpha * \cos \left(\mathcal {F} _ {\text {r e g i o n}} (I), \mathcal {F} _ {\text {t e x t}} (T)\right) \tag {1} +$$ + +Where $\alpha$ is the scale of similarities, which is set as 5 in our work to balance the range of the semantic score, $I$ and $T$ are the input image and MLLM generated region description with grounding text removed; $\mathcal{F}_{region}$ and $\mathcal{F}_{text}$ denotes the visual embedding for image region and text embeddings, respectively. + +When extracting image region embeddings, a straightforward approach is to crop the image region $I_{crop}$ and directly extract visual embeddings. However, the similarity score with such embedding tends to overly focus on the region's details while neglecting the image's surrounding context. To address this limitation, we supplement it with similarities $S_{local}$ from visual embeddings of intact images + +that incorporate local attention. Specifically, we feed the original image into CLIP and replace the final layer of the vision encoder that aggregates the embeddings with a local-attention layer. This modification allows the model to better account for the context around the region of interest. As defined in Eq. (2), we then use the average of the cropped image's similarity score and the full image's similarity score with local attention as the final semantic score, which effectively evaluates the extent of fine-grained alignment between the region description and local visual semantics. + +$$ +S _ {s e m} = \frac {1}{2} \left(S \left(I _ {c r o p}, T\right) + S _ {l o c a l} (I, T)\right) \tag {2} +$$ + +Localization Score. We propose a localization score to evaluate how detailed the MLLM responds in describing objects within the queried region and its grounding accuracy. This score is calculated based on the number of objects mentioned in the description that match the ground truth objects in the region. In practice, we use Grounding DINO [34] and the cropped image region to extract bounding boxes for objects mentioned in the description. The extracted objects are then combined with the original object annotations to form a set of ground truth objects within the region. Next, we extract the grounding results from MLLM-generated descriptions and combine them with the results from Grounding DINO to form the predicted objects. Finally, we compute the average IoU between the predicted objects and the ground truth as the localization score. The detailed process is outlined in Algorithm 1. + +The localization score encourages the model to include more detailed descriptions for involved objects and accurately localize them in its responses. Finally, we combine the semantic and localization scores for each grounded region description: + +$$ +S = \lambda S _ {s e m} + (1 - \lambda) S _ {l o c} \tag {3} +$$ + +where $\lambda$ is set to 0.8 in our implementation. Then, the descriptions with the highest and lowest scores are paired as preferred and rejected data for preference optimization. + +Grounded Region Description Refinement. After obtaining the preference data pairs, we further enhance the divergence of the grounding results of the preferred and rejected descriptions to encourage the model to distinguish between accurate and inaccurate object localization. To achieve this, we refine the grounding results in the preferred descriptions while keeping the rejected ones unchanged. In practice, we leverage the results obtained while computing the localization score, including the object box predictions $B_{pred}$ and ground-truth object boxes $B_{gt}$ . We retain only those predictions that match the ground truths (IoU > 0.5) and replace their bounding boxes with the matched ones. Then, we remove duplicates of predictions based on their textual position in the description and IoUs. Finally, we insert the + +# Algorithm 1 Computing Localization Score + +Input: Cropped Image Region $I_{crop}$ , Grounded Region Description $T$ generated by MLLMs, Object Bounding Box Annotations $B_{anno}$ for the Queried Region. + +Output: Localization Score: $S_{loc}$ . + +1: Extract bounding boxes $B_{text}$ from the description $T$ and get the plain text $T_{plain}$ . +2: Leverage Grounding DINO to get grounded object results $B_{\text{ground}}$ from $T_{\text{plain}}$ . +3: Get the set of ground truth object boxes $B_{gt}$ by aggregating $B_{ground}$ and $B_{anno}$ and removing duplicated boxes. +4: Get the set of object box predictions $B_{pred}$ for the description $T$ by aggregating $B_{ground}$ and $B_{text}$ and removing duplicated boxes. +5: Computing IoU matrix $\mathbf{m}[i,j] = IoU(B_{gt}^i, B_{pred}^j)$ between $B_{gt}$ and $B_{pred}$ . +6: Filter the IoU by a threshold of 0.5. + +$$ +\mathbf {p} [ i, j ] = \left\{ \begin{array}{l l} 0 & \mathbf {m} [ i, j ] < 0. 5 \\ \mathbf {m} [ i, j ] & \text {o t h e r w i s e} \end{array} \right. +$$ + +7: return $S_{loc} = \frac{1}{n}\sum_{i}^{n}\max_{j}\mathbf{p}[i,j]$ + +refined object box predictions into the region description, resulting in an improved grounded region description with more precise coordinates. + +# 3.3. Preference Optimization + +After curating the preference dataset, we finetune MLLMs through DPO and adopt LORA to save the training cost. The loss for optimizing MLLMs is defined as: + +$$ +\mathcal {L} = - \mathbb {E} _ {\left(x, y _ {w}, y _ {l}\right)} \left[ \log \sigma \left(\beta \log \frac {\pi_ {*} \left(y _ {w} \mid x\right)}{\pi_ {\operatorname {r e f}} \left(y _ {w} \mid x\right)} - \beta \log \frac {\pi_ {*} \left(y _ {l} \mid x\right)}{\pi_ {\operatorname {r e f}} \left(y _ {l} \mid x\right)}\right) \right] \tag {4} +$$ + +where $y_{w}$ and $y_{l}$ are the preferred and rejected description data; $\pi_{r}ef(y|x)$ is the base reference policy model, i.e., the initial instruction-tuned MLLM which is frozen during the training; $\pi_{*}(y|x)$ denotes the policy model which inherits from the instruction-tuned model with its LORA weights updated in the training process. + +# 4. Experiments + +# 4.1. Experiment Setups. + +Implementation Details. In this work, we experiment with the proposed SPR with three MLLMs with spatial understanding capabilities, including Ferret [58], LLaVOneVision [24], and CogVLM-Grounding [52]. To construct preference data, we randomly select 10k images with object annotations from the training set of Objects365 + +Table 1. Experiments on the Referring Expression Comprehension task (Acc@0.5) on datasets RefCOCO/+/g, and the Phrase Grounding task (Recall@1) on Flickr30k Entities dataset. “-” indicates results are unavailable or that MLLMs do not support multi-object grounding. + +
MethodRefCOCORefCOCO+RefCOCOgFlickr30k Entities
valtestAtestBvaltestAtestBvaltestvaltest
UNITER [12]81.4187.0474.1775.9081.4566.7074.0268.67--
UniTAB [55]86.3288.8480.6178.7083.2269.4879.9679.9778.7679.58
MDETR [21]86.7589.5881.4179.5284.0970.6281.6480.8982.383.8
MiniGPT-v2-7B [70]88.0691.2984.3079.5885.5273.3284.1984.31--
VistaLLM [41]88.191.583.082.989.874.883.684.4--
LLaVA-Grounding [64]89.16--81.68--84.82-83.0383.62
Shikra-7B [10]87.0190.6180.2481.6087.3672.1282.2782.1975.8476.54
Shikra-13B [10]87.8391.1181.8182.8987.7974.4182.6483.1677.4178.44
Griffon-13B [60]89.492.584.683.388.476.085.186.183.784.2
LLava-OV-7B [24]74.7782.5964.0470.1779.8558.4872.3471.39--
+ SPR76.6682.5265.9771.6279.8759.9972.9871.55--
Ferret-7B [58]87.4991.3582.4580.7887.3873.1483.9384.7680.3982.21
+ SPR88.3991.6783.9182.0787.8474.1985.5885.7581.5383.34
Ferret-13B [58]89.4892.4184.3682.8188.1475.1785.8386.3481.1384.76
+ SPR89.9493.0685.1283.2988.8975.7486.4686.9281.8283.75
CogVLM-Grounding-17B [52]92.7694.7588.9988.6892.9183.3989.7590.79--
+ SPR92.9594.8789.1588.8392.9583.8490.0190.96--
+ +Table 2. Experiments on Referring Expression Comprehension task under different IoU thresholds. The results are the average on RefCOCO, RefCOCO+, and RefCOCOg datasets. + +
IoU Threshold0.50.60.70.80.9
Ferret-7B83.9181.2876.7267.0243.25
+SPR84.9382.3678.4270.0952.21
Ferret-13B85.5682.9478.5770.0449.55
+SPR86.1883.6379.9372.0353.61
+ +Dataset [46], then construct random regions to query models to generate grounded region descriptions. We adopt LORA [17] for tuning MLLMs. The training is conducted on one A100 GPU, which takes around 3 and 5 hours for Ferret 7B and 13B models, respectively. Please refer to Appendix for more details on preference data construction and hyperparameter selection. + +Evaluation Benchmarks We evaluate our method on three types of benchmarks: (1) Grounding tasks that evaluate the localization accuracy, including referring expression comprehension (REC) and phrase grounding; (2) Region description task on Refcocog [22] and visual genome [23], and Ferret Bench [58] for comprehensive spatial understanding; (3) General benchmarks TextVQA [47], GQA [19], LLaVA-Bench [33], and hallucination benchmark POPE [56]. + +# 4.2. Experiments on REC + +We first evaluate our method on the referring expression comprehension (REC) task on RefCOCO [22], RefCOCO+ [22], and Refcocog [36]. The task requires the model to locate the object or region given a short de + +scription, which evaluates the model's fine-grained visual grounding abilities under the single-object referent scenarios. As shown in Tab. 1, our proposed SPR framework consistently improves the performance of three baseline MLLMs on different datasets for all model sizes. Considering that the REC results are based on an IoU threshold of 0.5, the improvement on its performance indicates that the model localized more objects successfully. Hence, this improvement can be largely attributed to the introduction of localization scores when constructing the preference data in SPR. Region descriptions that accurately mention more objects could achieve higher localization scores in SPR and be more likely to serve as preferred data, facilitating the model to attend to more objects and their locations in the image. + +To better evaluate the impact of SPR on the localization capability of MLLMs, we also conduct REC experiments with higher IoU thresholds by gradually increasing the IoU threshold of valid REC results from the default value of 0.5 to 0.9. As shown in Tab. 2, the improvements brought by SPR significantly increase as the threshold rises, with accuracy gains of 8.96 and 4.06 for the 7B model and the 13B model, respectively, when the IoU threshold rises to 0.9. With SPR, the localization accuracy of the objects in the model's response is greatly improved. Equipped with the grounded region description refinement and the supervision of preferred-rejected localization data in SPR, the model can respond more accurately to grounding object locations, demonstrating the effectiveness of incorporating preference optimization for region description and object localization in the fine-grained spatial understanding of MLLMs. + +Table 3. Experiments on Phrase Grounding task under different IoU thresholds. The results are averaged over the validation and test set of the Flickr30k dataset. + +
IoU Threshold0.50.60.70.80.9
Ferret-7B81.376.1467.5553.8629.98
+SPR82.4477.1469.2556.1933.99
Ferret-13B82.9476.6268.3455.7432.60
+SPR82.7877.2369.7456.9634.18
+ +Table 4. Experiments on the region captioning task on Refcocog and Visual Genome datasets. + +
MethodRefcocogVisual Genome
METEORROUGE_LMETEORROUGE_L
Ferret-7B12.315.617.429.6
+SPR13.520.417.629.7
Ferret-13B12.926.417.931.0
+SPR13.327.218.231.3
+ +# 4.3. Experiments on Phrase Grounding + +Furthermore, we experiment with the phrase grounding task on Flickr30k Entity [40]. In phrase grounding, the queried object phrases are combined in a single question, requiring MLLMs to detect the locations of multiple objects in a single response, which makes it more challenging than the single-object referring task like REC. Following [58], we adopt the question "What are the locations of [phrases]?" and evaluate the result using the MERGE-BOXES mode [21]. Since LLaVA-OneVision and CogVLM do not support multi-object detection, we report only the results for Ferret. As shown in Tab. 1, SPR effectively improves Ferret's performance in multi-object referent scenarios, especially for the 7B model, whose performance is even comparable to that of the 13B model. + +We then experiment with the phrase grounding task under higher IoU thresholds. We found that the multi-object referencing setting in phrase grounding is more challenging than the single-object referencing in REC. As the IoU threshold increases, the performance drops more rapidly, indicating a significant demand for MLLM to improve the capabilities of more accurate localization. Our approach can significantly alleviate this issue. As shown in Tab. 3, SPR improves progressively with higher IoU thresholds, reaching a maximum gain of 4.01 and 1.58 Recall@1 for the 7B and 13B models, respectively. This experiment demonstrates the superiority of SPR in pursuing detailed descriptions with high-precision object localization. + +# 4.4. Experiments on Region Captioning + +Beyond the grounding task, we also verify our proposed SPR in improving the text qualities of MLLMs' outputs + +Table 5. Experiments on the Ferret Bench. "Description", "Reasoning", and "Grounding" denote the Referring Description, Referring Reasoning, and Grounding in Conversation tasks. + +
ModelDescriptionReasoningGroundingAvg.
Ferret-7B68.767.357.564.5
+ SPR70.068.458.165.5
Ferret-13B70.668.759.766.3
+ SPR70.872.660.267.9
+ +Table 6. Experiments on the general and hallucination benchmarks. We report the accuracy for GQA and VQA and the F1 score on POPE. + +
ModelVQATGQALLaVAPOPE
Ferret-7B--64.785.36
+ SPR--66.385.69
LLaVA-OV-7B75.8962.2188.988.12
+ SPR76.0762.4291.488.49
+ +on fine-grained spatial understanding. We conduct experiments on the RefCOCOg and Visual Genome benchmarks. We prompt MLLMs with the question "Describe the region [region] in the image." to generate region captions and then evaluate the response quality using METEOR [5] and ROUGE_L [28] metrics. Tab. 4 shows that SPR effectively improves the quality of MLLM-generated region captions. After tuning with SPR, MLLMs are able to effectively attend to the user-specified regions and generate captions that better reflect the details of the region content. + +# 4.5. Experiments on Ferret Bench + +Ferret benchmark, proposed by [58], aims to evaluate MLLMs' fine-grained multimodal conversational capabilities such as referring description, referring reasoning, and grounded conversation. We follow the pipeline in [58] to prompt MLLMs with questions and employ GPT to evaluate the responses. As shown in Tab. 5, the proposed SPR can facilitate MLLMs in achieving better conversational qualities for fine-grained multimodal understanding, especially for the referring reasoning task, with an accuracy gain of about 3.9 for the 13B model. Equipped with SPR, MLLM can focus on more detailed visual information and generate responses that align better with human preferences. + +# 4.6. Experiments on General Benchmarks + +We further evaluate SPR on three general benchmarks to validate the benefits of improving MLLMs' spatial capabilities. TextVQA and GQA require MLLMs to answer questions or perform reasoning based on specific text, objects, and image content. LLaVA bench evaluates MLLMs comprehensive capabilities in conversation, description, and reasoning. As shown in Tab. 6, improving MLLMs' spa + +Table 7. Ablation Studies on the refinement of grounded region descriptions, and ratio $\lambda$ between semantic and localization scores in ranking MLLM responses. We report the average results on Referring expression comprehension and phrase grounding tasks. + +
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ SPR84.9382.44
w/o Refinement84.4191.38
λ = 0.084.2581.83
λ = 0.484.3481.95
λ = 0.684.6682.13
λ = 0.884.9382.44
λ = 1.084.4581.87
+ +Table 8. Ablation studies on the training strategy. + +
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ Instruction Finetuning84.3581.72
+ DPO training84.9382.44
+ +tial understanding capabilities consistently enhances their comprehension and reasoning abilities across diverse general scenarios, leading to performance gains on all three benchmarks. We also experiment on hallucination benchmark POPE, where SPR improves both Ferret and LLaVA-OneVision. This can be attributed to the preference data construction in SPR, where semantic and localization scores are applied to select region descriptions that better align with region content and reject those related to content outside the region or that contain hallucinations, thus effectively helping mitigate the hallucinations in MLLMs. + +# 4.7. Ablation Studies + +We conduct ablation studies over the two designs in SPR and evaluate the performance of Ferret-7B on the REC (Refcoco $+ / \mathrm{g}$ ) and the Flickr30k phrase grounding tasks. + +Score Ratio $\lambda$ . In Sec. 3.2, we combine the semantic and localization scores to rank MLLMs generated descriptions with a score ratio $\lambda$ . As shown in Tab. 7, we vary the $\lambda$ from 0 to 1, and the trained models outperform the baseline model consistently. When $\lambda$ equals zero, SPR achieves minimal gain, as the model might overly reward descriptions that simply list object names or fail to align with the region's semantics. On the other hand, when $\lambda$ is set to 1, SPR completely disregards measuring how detail the MLLM describes the objects in the region. Under such situations, the model encourages coarse region descriptions with fewer objects involved and reduces the corresponding object bounding box texts in the preferred data, thereby hindering the training of MLLMs' localization capability. As the experiments show, a relatively high value of 0.8 achieves the best results and is set as the default value in SPR. + +Refinement of Grounded Region Description. After con + +structuring the preferred and rejected data pairs, we further refine the localization results in the preferred descriptions by completing bounding boxes for objects in the description that were not grounded and refining the existing bounding boxes. Tab. 7 shows the results of this refinement. We found that the refinement leads to greater improvements in the multi-object referring task of phrase grounding. This could be attributed to the fact that the baseline model often fails to follow instructions for providing bounding boxes for each mentioned object when generating region descriptions. After tuning by the refined descriptions, MLLMs could faithfully ground the mentioned objects, thereby improving the multi-object phrase grounding clearly. + +# 4.8. Comparison with SFT + +In this paper, we adopt DPO with accept-reject preference data to optimize MLLMs for spatial understanding, whereas prior work [10, 44, 61], primarily focuses on the stage of supervised instruction fine-tuning (SFT). In Tab. 8, we compare these two training approaches, where SFT is trained using only the accepted data. The results show that while SFT could improve MLLMs' localization capabilities, its performance gains are significantly lower than DPO. DPO optimizes MLLM by contrasting accepted and rejected data pairs, similar to the positive-negative sample training mechanism in traditional object detection algorithms. This approach helps models distinguish between accurate and inaccurate localizations and facilitates MLLM in spatial understanding more effectively. However, it is important to note that DPO training also relies on a well-trained SFT model as a foundation, making these two approaches complementary. In future work, we will further explore how to integrate SFT and DPO to enhance MLLMs' spatial understanding. + +# 5. Conclusion + +In this work, we propose SPR, a Spatial Preference Rewarding framework to enhance MLLM's fine-grained spatial understanding capabilities. We introduce a complete pipeline that includes (1) Constructing random region queries; (2) Prompting MLLMs to generate diverse grounded region descriptions; (3) Proposing semantic scores and localization scores to rank the descriptions comprehensively; (4) Refining the localization quality of preference data; (5) Finetuning MLLMs to optimize against detailed and accurate spatial understanding. The entire framework does not require additional human labor or external MLLMs, with minimal overhead on training costs. SPR addresses the lack of direct optimization for positive and negative localization samples in MLLM training, enhancing their localization capabilities and promoting better alignment with human preferences. Experiments demonstrate that SPR significantly improves MLLMs' performance on standard referring and grounding tasks for spatial understanding. + +# Acknowledgement + +This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +This study is also supported by the MOE Tier-2 project, with project number MOE-T2EP20123-0003. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1 +[4] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3 +[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7 +[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 3 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with Transformers. In ECCV, 2020. 1, 3 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1 +[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3 +[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 6, 8 +[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: + +Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3 +[12] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 6 +[13] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2024. 3 +[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 1, 3 +[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[16] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 1, 3 +[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6 +[18] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 1 +[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6 +[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3 +[21] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 6, 7 +[22] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3, 6 + +[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3, 6 +[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5, 6 +[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[26] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3 +[27] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundingpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3 +[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7 +[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 1 +[30] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3 +[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3, 6 +[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[35] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 1, 3 +[36] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In + +Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6 +[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3 +[38] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. arXiv preprint arXiv:2408.10433, 2024. 1, 3 +[39] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3 +[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 7 +[41] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 6 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4 +[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2 +[44] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12987, 2024. 3, 8 +[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1 +[46] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6 +[47] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Con- + +ference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 6 +[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 1, 3 +[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3 +[50] Fei Wang, Wenxuan Zhou, James Y Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. arXiv preprint arXiv:2406.11839, 2024. 1, 3 +[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3 +[52] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1, 5, 6 +[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 +[54] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. beta-dpo: Direct preference optimization with dynamic beta. arXiv preprint arXiv:2407.08639, 2024. 4 +[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 6 +[56] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Zhao Wayne Xin, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 3, 6 +[57] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3 +[58] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 1, 3, 5, 6, 7 +[59] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. arxiv, 2023. 1, 3 + +[60] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 3, 6 +[61] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 1, 3, 8 +[62] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024.1 +[63] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 3 +[64] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 6 +[65] Mengxi Zhang and Kang Rong. Automated multi-level preference for mllms. arXiv preprint arXiv:2405.11165, 2024. 1, 3 +[66] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 3 +[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6578-6587, 2019. 3 +[68] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024.3 +[69] Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622, 2024. 1,3 +[70] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6 \ No newline at end of file diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/images.zip b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..85673468d1e1664001118b9c34606604a8d2dacc --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fd18347a3ab77d9447192ea60b071c76027b12e0b5764595994df2932a6e3ee +size 469579 diff --git a/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/layout.json b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dca7cf6ef96aeeadb773af676d6082a5a009659a --- /dev/null +++ b/2025/Spatial Preference Rewarding for MLLMs Spatial Understanding/layout.json @@ -0,0 +1,8646 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "text", + "content": "Spatial Preference Rewarding for MLLMs Spatial Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "spans": [ + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": "Han Qiu" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": ", Peng Gao" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": ", Lewei Lu" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": ", Xiaqin Zhang" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": ", Ling Shao" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "text", + "content": ", Shijian Lu" + }, + { + "bbox": [ + 121, + 142, + 493, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "spans": [ + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University, " + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "text", + "content": "Sensetime Research, " + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 136, + 170, + 475, + 197 + ], + "type": "text", + "content": "Zhejiang University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 198, + 474, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 198, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 198, + 474, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 135, + 198, + 474, + 213 + ], + "type": "text", + "content": "UCAS-Terminus AI Lab, University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 188, + 216, + 418, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 216, + 418, + 227 + ], + "spans": [ + { + "bbox": [ + 188, + 216, + 418, + 227 + ], + "type": "text", + "content": "han023@e.ntu.edu.sg, Shijian.Lu@ntu.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 280, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 628 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) have demonstrated promising spatial understanding capabilities, such as referencing and grounding object descriptions. Despite their successes, MLLMs still fall short in fine-grained spatial perception abilities, such as generating detailed region descriptions or accurately localizing objects. Additionally, they often fail to respond to the user's requirements for desired fine-grained spatial understanding. This issue might arise because existing approaches primarily focus on tuning MLLMs to model pre-annotated instruction data to inject spatial knowledge, without direct supervision of MLLMs' actual responses. We address this issue by SPR, a Spatial Preference Rewarding (SPR) approach that enhances MLLMs' spatial capabilities by rewarding MLLMs' detailed responses with precise object localization over vague or inaccurate responses. With randomly selected image regions and region descriptions from MLLMs, SPR introduces semantic and localization scores to comprehensively evaluate the text quality and localization quality in MLLM-generated descriptions. We also refine the MLLM descriptions with better localization accuracy and pair the best-scored refinement with the initial descriptions of the lowest score for direct preference optimization, thereby enhancing fine-grained alignment with visual input. Extensive experiments over standard referring and grounding benchmarks show that SPR improves MLLM spatial understanding capabilities effectively with minimal overhead in training. Data and code will be released at https://github.com/hanqiu-hq/SPR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 651, + 135, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 135, + 662 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 135, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 695 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) [3, 15, 18, 31, 32, 51, 58, 70] have achieved remarkable success by inte-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 256, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 422 + ], + "type": "text", + "content": "grating pretrained large language model [2, 14, 49] with vision encoders [8, 37, 42], leading to significant advancements in a wide range of general vision-language tasks. By combining visual and language signals, MLLMs have demonstrated superior capabilities in multimodal understanding, reasoning, and interaction as compared with traditional vision models. Recently, several studies have further injected spatial knowledge into MLLMs, thereby improving MLLMs' fine-grained perception of visual inputs and enabling tasks such as referential dialogue [10, 16], grounding captioning [35, 58, 62], region description [31, 52], and object detection [61], etc. These advances have paved the way for MLLMs to serve as versatile visual assistants supporting a wider range of applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 425, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 675 + ], + "type": "text", + "content": "Despite recent advancements, MLLMs still face challenges in fine-grained spatial understanding, with responses not aligned with human preferences. As illustrated in 1, the generated grounded region descriptions are often vague with inaccurate object localizations, and models may fail to focus on the queried region, distracted from other regions in the image. The issue in spatial understanding could be attributed to the lack of positive-negative preference feedback in existing instruction-tuned MLLMs. Specifically, instruction fine-tuning (SFT) directly optimizes MLLMs to mimic ground truth positive samples, but it cannot impose any penalties if the model produces inaccurate negative samples for localization during actual inference. As a result, MLLMs may struggle to generate positive descriptions with accurate object localization and instead produce negative and inaccurate descriptions, leading to responses that do not align with user expectations. In addition, optimization using positive and negative samples has been proven crucial for spatial understanding in traditional object detection algorithms [7, 29, 45], highlighting a significant gap in the current MLLM training on spatial understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Several studies [38, 48, 50, 59, 65, 69] attempt to introduce preference optimization for better MLLM alignment, where the preference data are constructed by collect" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "720" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 83, + 189, + 152 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 211, + 82 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 211, + 82 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 211, + 82 + ], + "type": "text", + "content": "User Prompt about Regions in Images:" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 83, + 83, + 189, + 152 + ], + "lines": [ + { + "bbox": [ + 83, + 83, + 189, + 152 + ], + "spans": [ + { + "bbox": [ + 83, + 83, + 189, + 152 + ], + "type": "image", + "image_path": "6ecce25348467d0b662edc637ca4686592b5c5f5e4c1881e8c418ea9b917eb06.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 156, + 148, + 166 + ], + "lines": [ + { + "bbox": [ + 83, + 156, + 148, + 166 + ], + "spans": [ + { + "bbox": [ + 83, + 156, + 148, + 166 + ], + "type": "text", + "content": "MLLM Response:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 194, + 101, + 296, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 101, + 296, + 136 + ], + "spans": [ + { + "bbox": [ + 194, + 101, + 296, + 136 + ], + "type": "text", + "content": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 82, + 411, + 152 + ], + "blocks": [ + { + "bbox": [ + 306, + 82, + 411, + 152 + ], + "lines": [ + { + "bbox": [ + 306, + 82, + 411, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 82, + 411, + 152 + ], + "type": "image", + "image_path": "ab0ab80098fa3b1f68e0bff559b8f3ead7d3acf9823acc9f234ed7153636732b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 415, + 102, + 517, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 415, + 102, + 517, + 136 + ], + "spans": [ + { + "bbox": [ + 415, + 102, + 517, + 136 + ], + "type": "text", + "content": "Prompt: Can you give a brief description of the specified area [676, 229, 847, 1000] in the image? Specify the location of each mentioned object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 182, + 197, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 182, + 197, + 225 + ], + "spans": [ + { + "bbox": [ + 88, + 182, + 197, + 225 + ], + "type": "text", + "content": "The area [676, 229, 847, 1000] is occupied by a woman in a blue shirt who is holding a whistle [537, 530, 568, 606]. She is also wearing a black clipboard [801, 538, 825, 581] around her neck ...." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 226, + 190, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 226, + 190, + 232 + ], + "spans": [ + { + "bbox": [ + 152, + 226, + 190, + 232 + ], + "type": "text", + "content": "Fail to Ground" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 200, + 166, + 304, + 236 + ], + "blocks": [ + { + "bbox": [ + 200, + 166, + 304, + 236 + ], + "lines": [ + { + "bbox": [ + 200, + 166, + 304, + 236 + ], + "spans": [ + { + "bbox": [ + 200, + 166, + 304, + 236 + ], + "type": "image", + "image_path": "cbb76db1ef6c4cb5e22b0e979ca279eb75aa5bc2962748530c13bf7e0a1d3843.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 309, + 181, + 415, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 181, + 415, + 232 + ], + "spans": [ + { + "bbox": [ + 309, + 181, + 415, + 232 + ], + "type": "text", + "content": "The region [638, 157, 775, 709] is a trail located on a mountain [166, 365, 996, 996]. of people [619, 250, 799, 990] walking on it. The trail also has a marker [166, 339, 229, 439] to guide hikers. Inaccurate Localization" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 420, + 166, + 526, + 236 + ], + "blocks": [ + { + "bbox": [ + 420, + 166, + 526, + 236 + ], + "lines": [ + { + "bbox": [ + 420, + 166, + 526, + 236 + ], + "spans": [ + { + "bbox": [ + 420, + 166, + 526, + 236 + ], + "type": "image", + "image_path": "69c00219318dcdb064f0d69d1853b96aa7de07ffad9584a65699ac43e62540df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 242, + 399, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 242, + 399, + 253 + ], + "spans": [ + { + "bbox": [ + 231, + 242, + 399, + 253 + ], + "type": "text", + "content": "Tuning with Spatial Preference Rewarding (SPR)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 270, + 195, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 270, + 195, + 336 + ], + "spans": [ + { + "bbox": [ + 86, + 270, + 195, + 336 + ], + "type": "text", + "content": "The area [676, 229, 847, 998] in the image features a woman in a blue shirt with a name tag, .... She [676, 229, 817, 901] appears to .... The grassy area where she is standing [0, 385, 998, 998] seems to be .... activities. There's also a red cone [676, 909, 720, 998] ...." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 200, + 269, + 304, + 338 + ], + "blocks": [ + { + "bbox": [ + 83, + 258, + 230, + 269 + ], + "lines": [ + { + "bbox": [ + 83, + 258, + 230, + 269 + ], + "spans": [ + { + "bbox": [ + 83, + 258, + 230, + 269 + ], + "type": "text", + "content": "Responses from SPR-empowered MLLM:" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 269, + 304, + 338 + ], + "lines": [ + { + "bbox": [ + 200, + 269, + 304, + 338 + ], + "spans": [ + { + "bbox": [ + 200, + 269, + 304, + 338 + ], + "type": "image", + "image_path": "5310376b17b7b49b037e7ebc730e6eaecf0df6b387ac830d8c7570c712df2ee1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 395 + ], + "type": "text", + "content": "Figure 1. The proposed Spatial Preference Rewarding (SPR) mitigates the distracted and inaccurate region descriptions generated by MLLMs. Given an image and a user-specified region of interest, MLLMs often fail to focus on the queried region. They may be distracted by objects outside the specified region, failing to ground the queried objects, or providing inaccurate localization. Tuning MLLMs with our proposed SPR leads to more accurate object localization and detailed object descriptions." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 272, + 417, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 417, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 417, + 337 + ], + "type": "text", + "content": "The region [638, 157, 775, 709] is where a man [575, 150, 787, 993] is standing.... The water bottle [732, 321, 796, 481] suggests that he is well-equipped for his journey. Moreover, his grey jacket [575, 214, 720, 658] indicates that he is ready for the outdoor activity." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 268, + 526, + 338 + ], + "blocks": [ + { + "bbox": [ + 421, + 268, + 526, + 338 + ], + "lines": [ + { + "bbox": [ + 421, + 268, + 526, + 338 + ], + "spans": [ + { + "bbox": [ + 421, + 268, + 526, + 338 + ], + "type": "image", + "image_path": "ad9d796c0ce9fd398c59abbb0cf3aadad020b7a8d46e09fefaf97857fe32e153.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 406, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 295, + 491 + ], + "type": "text", + "content": "ing MLLM-generated image descriptions and scoring them by human or LLMs. However, these methods primarily leverage preferences to improve image-level coarse alignment, and most of them target mitigating hallucinations in MLLMs. The problem of fine-grained alignment for spatial understanding, such as detailed region descriptions and accurate object localization, has been largely neglected." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": "To address this gap, we design SPR, a Spatial Preference Rewarding framework that enhances MLLM spatial understanding capabilities by rewarding detailed responses with accurate object localization over vague or inaccurate responses. Specifically, SPR selects random image regions containing multiple objects and prompts MLLMs in diverse ways to generate grounded region descriptions. In reward modeling, it introduces both semantic and localization scores to evaluate the alignment between the region description and the region semantics, as well as how detailed region objects are described. We also refine the grounded object in the generated description to enhance its localization accuracy. Finally, the best-scored refined description and the response of the lowest score are paired as preferred and rejected data for direct preference optimization (DPO) [43] training with LORA [17]. By aligning MLLMs with detailed and accurate responses, SPR mitigates MLLMs' incompetence in accurate localization and spatial understand" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 406, + 480, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 480, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 480, + 418 + ], + "type": "text", + "content": "ing as required in many real-world tasks." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 540 + ], + "type": "text", + "content": "We validate the effectiveness of SPR in enhancing MLLMs' spatial understanding capabilities with minimal overhead in training. Compared to the baseline, SPR enhances MLLMs on both referring and grounding benchmarks, especially under higher IoU thresholds which demand higher localization accuracy. In addition, SPR can improve MLLM trustworthiness and reduce MLLM hallucinations as well. Our experiments highlight the importance of incorporating preference-based feedback to enhance the fine-grained spatial understanding abilities in MLLMs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 541, + 550, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 541, + 550, + 553 + ], + "spans": [ + { + "bbox": [ + 314, + 541, + 550, + 553 + ], + "type": "text", + "content": "The contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 556, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 314, + 556, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 556, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 314, + 556, + 554, + 616 + ], + "type": "text", + "content": "- We propose a Spatial Preference Rewarding (SPR) framework to enhance the fine-grained spatial understanding of MLLMs via direct preference optimization (DPO), enhancing MLLMs' capabilities in precise region referring and accurate object localization in images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 616, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 675 + ], + "type": "text", + "content": "- We develop an automated pipeline that creates preference data by constructing random region prompts and scoring model responses for spatial understanding. The pipeline requires no other MLLMs or human labours, making it scalable in future training." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 675, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 553, + 712 + ], + "type": "text", + "content": "- Extensive experiments show that the proposed SPR improves MLLMs' spatial understanding capabilities consistently across multiple public benchmarks." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "721" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 294, + 472 + ], + "type": "text", + "content": "Recently, the success of large language models (LLMs) [14, 49, 51] has been extended into the multimodal domain, resulting in models that demonstrate impressive performance in integrating vision and language [1, 15, 20, 32, 70]. These models treat visual signals as a special form of language, establishing multimodal understanding, reasoning, and interaction capabilities by combining visual encoders [37, 42] with pre-trained large language models, or by directly feeding encoded visual signals into LLMs [6, 53]. Most current MLLMs follow a two-step training process. The first step is pre-training, where large-scale vision-language datasets [9] are used to align visual features to the same space as language features. This enables the model to bridge visual and language embeddings effectively. The second step involves instruction-following finetuning, where high-quality vision-language datasets [25, 33, 70] are used to further enhance the MLLMs' capabilities to follow user instructions and comprehend multimodal information. These methods often convert existing datasets into an instruction-following format or adopt leading MLLMs like GPT to generate high-quality training instruction data for MLLMs [11, 33]. Despite their success, current MLLMs still face challenges that may generate undesired responses toward human preferences. For instance, these models are prone to generating hallucinated content [4, 30, 56, 59] or providing responses that do not fully meet user expectations. Improving the quality of MLLM responses and aligning them more closely with user preferences has thus become a surging focus of research in the community. Our work aims to improve the spatial understanding capabilities of MLLMs, aligning their behaviors better with human preferences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 475, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 713 + ], + "type": "text", + "content": "MLLMs for Spatial Understanding. Spatial understanding capabilities [7, 13, 22, 67], such as object detection, referring, and grounding description tasks, have long been a fundamental research topic in the field of computer vision. Recent efforts attempt to empower MLLMs with dense visual perception and spatial understanding abilities by integrating region-level data in MLLM training or modifying MLLM architectures. For example, Kosmos-2 [39] and Shikra [10] directly represent the object coordinates in text, constructing instruction datasets to inject spatial knowledge into MLLMs. LLava-Grounding [63] and GroundingGPT [27] construct large-scale grounding datasets to enhance multimodal grounding capabilities. To better facilitate localization within images, RegionGPT [16], GPT4ROI [66], Ferret [58], and Groma [35] encode region features as direct inputs to LLMs, facilitating explicit attention to specific image regions. The Griffon [60, 61] series focuses on dense detection, enabling MLLMs to achieve performance comparable to traditional object detectors. LocVLM [44] explores the. LocVLM [44]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "explores key factors in instruction tuning for spatial understanding, such as coordinate representation, which improves MLLM's spatial awareness. However, these efforts primarily concentrate on the instruction-tuning phase and lack direct feedback on MLLMs' responses. To fill this gap, we propose a Spatial Preference Rewarding (SPR) framework, which constructs preference data based on MLLMs generated grounded region descriptions for MLLM tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 169, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 491 + ], + "type": "text", + "content": "Preference Optimization for MLLMs. Preference alignment has recently emerged as a promising direction to align model responses with human preferences. One widely explored approach is to employ Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO) to improve the trustworthiness of MLLMs and reduce hallucinations in their responses. For example, LLaVA-RLHF [48] and RLHF-V [59] leverage human annotators to evaluate model responses and construct preference data for fine-tuning. POVID [68] and Silkie [26] use external models, such as GPT, as evaluators to build preference datasets. CLIP-DPO [38] and CSR [69] use CLIP to rank model responses to avoid resource-intensive human or MLLM annotations. AMP [65] introduced a multi-level preference framework to enable MLLMs to better model differences between preference data. mDPO [50] introduced additional preference data pairs with corrupted images to avoid over-optimization on language-only preferences. Unlike these existing studies that primarily aim to reduce hallucinations in MLLMs, our proposed SPR framework focuses on optimizing MLLM responses related to spatial reasoning and understanding. Specifically, SPR focuses on fine-grained alignment with visual inputs and facilitates MLLMs in distinguishing between high-quality object localization (positive samples) and inaccurate localization (negative samples), thereby improving the spatial understanding capabilities of MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 501, + 374, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 374, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 374, + 514 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 605 + ], + "type": "text", + "content": "This section presents our proposed Spatial Preference Rewarding (SPR) framework. Following a typical DPO pipeline, SPR adopts a three-step process in MLLM finetuning, including collecting MLLMs' raw responses (Sec.3.1), evaluating the raw responses to construct preference data (Sec.3.2), and preference optimization (Sec.3.3). The details are elaborated in the following subsections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 612, + 533, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 533, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 533, + 625 + ], + "type": "text", + "content": "3.1. Grounded Region Description Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "The first step of our pipeline is to collect diverse model responses that will later be ranked to construct preference data. Since our primary objective is to enhance MLLMs' localization capabilities and achieve fine-grained alignment to visual inputs, we choose the task of region description with grounding to evaluate MLLMs' object localization capabilities. However, existing datasets [23, 57] for region" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "722" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 542, + 224 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 542, + 224 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 542, + 224 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 542, + 224 + ], + "type": "image", + "image_path": "5accb4f5af8f7bbedc29d2987ae4baba93d938c8a84f49ae1eafd1114988169e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 260 + ], + "type": "text", + "content": "Figure 2. We leverage the generated object references and cropped image region to build a variety of multimodal prompts to enhance the diversity of generated region descriptions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 281, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 296, + 388 + ], + "type": "text", + "content": "descriptions are often too simple, involving queried regions with only one or two objects and short phrases such as 'vehicles parked on the street' or 'bicycles are parked on the sidewalk.' Such simple prompts are inadequate for generating diverse responses to construct preferred and rejected preference data with sufficient divergence, which might hinder the effectiveness of DPO training [54]. To address this issue, we generate queried regions from scratch instead of using existing region description datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 569 + ], + "type": "text", + "content": "Region Query Construction. We design a simple approach to generate randomly queried regions based on images and object annotations. Take the Objects365 dataset as an example. We first filter out images with few objects, ensuring that the data contains rich visual content. Then, given the annotated object bounding boxes in each image, we randomly select one of the objects as the starting region. From there, we iteratively expand the region by incorporating the nearest objects. The expansion stops randomly once more than four objects are involved in the region. The resulting region then prompts MLLMs to generate a detailed region description. Through this process, we simulate the human-like, dynamic attention across different parts of an image, encouraging the MLLM to adaptively focus on arbitrary image regions based on the given prompts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Grounded Region Description Generation. As shown in Fig. 2, we build a variety of prompts for MLLMs to generate several region descriptions for each image, serving as candidate responses for preference data. Since the original MLLM sometimes struggles to generate detailed responses following region prompts, we utilize cropped region images along with object references constructed from annotations to guide MLLMs to attend to the region's content and details. These prompts help the model focus more effectively on the specified region and produce detailed descriptions that might better align with human preference. In this way, we encourage the MLLM to generate responses that are dis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 281, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 304 + ], + "type": "text", + "content": "tinct in content but consistent in language style, which is then used for constructing preference data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 314, + 539, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 539, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 539, + 327 + ], + "type": "text", + "content": "3.2. Preference Data Ranking and Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 463 + ], + "type": "text", + "content": "The next step is to rank the generated descriptions to obtain preferred and rejected data pairs. An ideal region description should meet at least two key criteria: (1) the text description should accurately match the semantics of the queried region and the surrounding image content, (2) it should provide detailed descriptions with accurate localization of objects within the region. To address these two criteria, we propose a semantic score and localization score to rank the responses. The descriptions with the highest and the lowest scores are paired to form preference data for DPO training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 464, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 524 + ], + "type": "text", + "content": "Semantic Score. We introduce the semantic score to evaluate the relevance between the generated descriptions and the semantics of queried image regions. We leverage a pretrained CLIP model [42] to compute the cosine similarities of text and visual embeddings as defined in Eq. (1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 349, + 537, + 553, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 537, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 349, + 537, + 553, + 550 + ], + "type": "interline_equation", + "content": "S (I, T) = \\alpha * \\cos \\left(\\mathcal {F} _ {\\text {r e g i o n}} (I), \\mathcal {F} _ {\\text {t e x t}} (T)\\right) \\tag {1}", + "image_path": "0c2ca81b6678d489d083298e7089382459a97d30362c9dfad5eac39235955a19.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " is the scale of similarities, which is set as 5 in our work to balance the range of the semantic score, " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " are the input image and MLLM generated region description with grounding text removed; " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{region}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{text}" + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": " denotes the visual embedding for image region and text embeddings, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "When extracting image region embeddings, a straightforward approach is to crop the image region " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "I_{crop}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " and directly extract visual embeddings. However, the similarity score with such embedding tends to overly focus on the region's details while neglecting the image's surrounding context. To address this limitation, we supplement it with similarities " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "S_{local}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " from visual embeddings of intact images" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "723" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "that incorporate local attention. Specifically, we feed the original image into CLIP and replace the final layer of the vision encoder that aggregates the embeddings with a local-attention layer. This modification allows the model to better account for the context around the region of interest. As defined in Eq. (2), we then use the average of the cropped image's similarity score and the full image's similarity score with local attention as the final semantic score, which effectively evaluates the extent of fine-grained alignment between the region description and local visual semantics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 201, + 296, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 201, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 95, + 201, + 296, + 224 + ], + "type": "interline_equation", + "content": "S _ {s e m} = \\frac {1}{2} \\left(S \\left(I _ {c r o p}, T\\right) + S _ {l o c a l} (I, T)\\right) \\tag {2}", + "image_path": "5d7368a38d3a8edde1317d2a52a772aa46d42c0799fca7091adcae4d74b619c1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 418 + ], + "type": "text", + "content": "Localization Score. We propose a localization score to evaluate how detailed the MLLM responds in describing objects within the queried region and its grounding accuracy. This score is calculated based on the number of objects mentioned in the description that match the ground truth objects in the region. In practice, we use Grounding DINO [34] and the cropped image region to extract bounding boxes for objects mentioned in the description. The extracted objects are then combined with the original object annotations to form a set of ground truth objects within the region. Next, we extract the grounding results from MLLM-generated descriptions and combine them with the results from Grounding DINO to form the predicted objects. Finally, we compute the average IoU between the predicted objects and the ground truth as the localization score. The detailed process is outlined in Algorithm 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "The localization score encourages the model to include more detailed descriptions for involved objects and accurately localize them in its responses. Finally, we combine the semantic and localization scores for each grounded region description:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 491, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 491, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 120, + 491, + 294, + 506 + ], + "type": "interline_equation", + "content": "S = \\lambda S _ {s e m} + (1 - \\lambda) S _ {l o c} \\tag {3}", + "image_path": "114b42ce334c1d5d8fab07c52dc1fae772fe8d48272c69ebdb6bde245ea84be8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": " is set to 0.8 in our implementation. Then, the descriptions with the highest and lowest scores are paired as preferred and rejected data for preference optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Grounded Region Description Refinement. After obtaining the preference data pairs, we further enhance the divergence of the grounding results of the preferred and rejected descriptions to encourage the model to distinguish between accurate and inaccurate object localization. To achieve this, we refine the grounding results in the preferred descriptions while keeping the rejected ones unchanged. In practice, we leverage the results obtained while computing the localization score, including the object box predictions " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "B_{pred}" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " and ground-truth object boxes " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "B_{gt}" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": ". We retain only those predictions that match the ground truths (IoU > 0.5) and replace their bounding boxes with the matched ones. Then, we remove duplicates of predictions based on their textual position in the description and IoUs. Finally, we insert the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 72, + 495, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 495, + 84 + ], + "type": "text", + "content": "Algorithm 1 Computing Localization Score" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "text", + "content": "Input: Cropped Image Region " + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "I_{crop}" + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "text", + "content": ", Grounded Region Description " + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "text", + "content": " generated by MLLMs, Object Bounding Box Annotations " + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "inline_equation", + "content": "B_{anno}" + }, + { + "bbox": [ + 315, + 87, + 553, + 123 + ], + "type": "text", + "content": " for the Queried Region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 123, + 457, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 123, + 457, + 135 + ], + "spans": [ + { + "bbox": [ + 315, + 123, + 457, + 135 + ], + "type": "text", + "content": "Output: Localization Score: " + }, + { + "bbox": [ + 315, + 123, + 457, + 135 + ], + "type": "inline_equation", + "content": "S_{loc}" + }, + { + "bbox": [ + 315, + 123, + 457, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 135, + 553, + 291 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "text", + "content": "1: Extract bounding boxes " + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "inline_equation", + "content": "B_{text}" + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "text", + "content": " from the description " + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "text", + "content": " and get the plain text " + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "inline_equation", + "content": "T_{plain}" + }, + { + "bbox": [ + 319, + 135, + 553, + 159 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "text", + "content": "2: Leverage Grounding DINO to get grounded object results " + }, + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "inline_equation", + "content": "B_{\\text{ground}}" + }, + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "inline_equation", + "content": "T_{\\text{plain}}" + }, + { + "bbox": [ + 319, + 159, + 553, + 183 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "text", + "content": "3: Get the set of ground truth object boxes " + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "inline_equation", + "content": "B_{gt}" + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "text", + "content": " by aggregating " + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "inline_equation", + "content": "B_{ground}" + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "inline_equation", + "content": "B_{anno}" + }, + { + "bbox": [ + 319, + 183, + 553, + 217 + ], + "type": "text", + "content": " and removing duplicated boxes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "content": "4: Get the set of object box predictions " + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "inline_equation", + "content": "B_{pred}" + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "content": " for the description " + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "content": " by aggregating " + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "inline_equation", + "content": "B_{ground}" + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "inline_equation", + "content": "B_{text}" + }, + { + "bbox": [ + 319, + 219, + 553, + 255 + ], + "type": "text", + "content": " and removing duplicated boxes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "text", + "content": "5: Computing IoU matrix " + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "inline_equation", + "content": "\\mathbf{m}[i,j] = IoU(B_{gt}^i, B_{pred}^j)" + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "inline_equation", + "content": "B_{gt}" + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "inline_equation", + "content": "B_{pred}" + }, + { + "bbox": [ + 319, + 254, + 553, + 279 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 279, + 475, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 279, + 475, + 291 + ], + "spans": [ + { + "bbox": [ + 319, + 279, + 475, + 291 + ], + "type": "text", + "content": "6: Filter the IoU by a threshold of 0.5." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 371, + 297, + 512, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 297, + 512, + 329 + ], + "spans": [ + { + "bbox": [ + 371, + 297, + 512, + 329 + ], + "type": "interline_equation", + "content": "\\mathbf {p} [ i, j ] = \\left\\{ \\begin{array}{l l} 0 & \\mathbf {m} [ i, j ] < 0. 5 \\\\ \\mathbf {m} [ i, j ] & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "7bf319fbf2b179cc9373618eecd2d97497451acc45bb584ad449d032051d680f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 335, + 468, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 335, + 468, + 355 + ], + "spans": [ + { + "bbox": [ + 320, + 335, + 468, + 355 + ], + "type": "text", + "content": "7: return " + }, + { + "bbox": [ + 320, + 335, + 468, + 355 + ], + "type": "inline_equation", + "content": "S_{loc} = \\frac{1}{n}\\sum_{i}^{n}\\max_{j}\\mathbf{p}[i,j]" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 415 + ], + "type": "text", + "content": "refined object box predictions into the region description, resulting in an improved grounded region description with more precise coordinates." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 422, + 452, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 452, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 452, + 435 + ], + "type": "text", + "content": "3.3. Preference Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 475 + ], + "type": "text", + "content": "After curating the preference dataset, we finetune MLLMs through DPO and adopt LORA to save the training cost. The loss for optimizing MLLMs is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 493, + 555, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 493, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 320, + 493, + 555, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\mathbb {E} _ {\\left(x, y _ {w}, y _ {l}\\right)} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {*} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {*} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\operatorname {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right] \\tag {4}", + "image_path": "f21925520afdeac3bfb95d24b1175bab0de0c0d462d053f80943f2e5f9b5b893.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "y_{w}" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " are the preferred and rejected description data; " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\pi_{r}ef(y|x)" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " is the base reference policy model, i.e., the initial instruction-tuned MLLM which is frozen during the training; " + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\pi_{*}(y|x)" + }, + { + "bbox": [ + 313, + 522, + 554, + 594 + ], + "type": "text", + "content": " denotes the policy model which inherits from the instruction-tuned model with its LORA weights updated in the training process." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 605, + 395, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 395, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 395, + 617 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "text", + "content": "4.1. Experiment Setups." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "Implementation Details. In this work, we experiment with the proposed SPR with three MLLMs with spatial understanding capabilities, including Ferret [58], LLaVOneVision [24], and CogVLM-Grounding [52]. To construct preference data, we randomly select 10k images with object annotations from the training set of Objects365" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "724" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 102, + 517, + 310 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Experiments on the Referring Expression Comprehension task (Acc@0.5) on datasets RefCOCO/+/g, and the Phrase Grounding task (Recall@1) on Flickr30k Entities dataset. “-” indicates results are unavailable or that MLLMs do not support multi-object grounding." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 102, + 517, + 310 + ], + "lines": [ + { + "bbox": [ + 91, + 102, + 517, + 310 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 517, + 310 + ], + "type": "table", + "html": "
MethodRefCOCORefCOCO+RefCOCOgFlickr30k Entities
valtestAtestBvaltestAtestBvaltestvaltest
UNITER [12]81.4187.0474.1775.9081.4566.7074.0268.67--
UniTAB [55]86.3288.8480.6178.7083.2269.4879.9679.9778.7679.58
MDETR [21]86.7589.5881.4179.5284.0970.6281.6480.8982.383.8
MiniGPT-v2-7B [70]88.0691.2984.3079.5885.5273.3284.1984.31--
VistaLLM [41]88.191.583.082.989.874.883.684.4--
LLaVA-Grounding [64]89.16--81.68--84.82-83.0383.62
Shikra-7B [10]87.0190.6180.2481.6087.3672.1282.2782.1975.8476.54
Shikra-13B [10]87.8391.1181.8182.8987.7974.4182.6483.1677.4178.44
Griffon-13B [60]89.492.584.683.388.476.085.186.183.784.2
LLava-OV-7B [24]74.7782.5964.0470.1779.8558.4872.3471.39--
+ SPR76.6682.5265.9771.6279.8759.9972.9871.55--
Ferret-7B [58]87.4991.3582.4580.7887.3873.1483.9384.7680.3982.21
+ SPR88.3991.6783.9182.0787.8474.1985.5885.7581.5383.34
Ferret-13B [58]89.4892.4184.3682.8188.1475.1785.8386.3481.1384.76
+ SPR89.9493.0685.1283.2988.8975.7486.4686.9281.8283.75
CogVLM-Grounding-17B [52]92.7694.7588.9988.6892.9183.3989.7590.79--
+ SPR92.9594.8789.1588.8392.9583.8490.0190.96--
", + "image_path": "b8902624abe45f07c9c0d954ef56eb7d469c7b906736909ba1624f3caedd4ec6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 360, + 281, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 297, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 297, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 297, + 353 + ], + "type": "text", + "content": "Table 2. Experiments on Referring Expression Comprehension task under different IoU thresholds. The results are the average on RefCOCO, RefCOCO+, and RefCOCOg datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 360, + 281, + 427 + ], + "lines": [ + { + "bbox": [ + 69, + 360, + 281, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 360, + 281, + 427 + ], + "type": "table", + "html": "
IoU Threshold0.50.60.70.80.9
Ferret-7B83.9181.2876.7267.0243.25
+SPR84.9382.3678.4270.0952.21
Ferret-13B85.5682.9478.5770.0449.55
+SPR86.1883.6379.9372.0353.61
", + "image_path": "8a40c68cfb19c35c91c9f99da53dc5dc054d9a5ff8ccc0a7f61354427e00f939.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 447, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 295, + 530 + ], + "type": "text", + "content": "Dataset [46], then construct random regions to query models to generate grounded region descriptions. We adopt LORA [17] for tuning MLLMs. The training is conducted on one A100 GPU, which takes around 3 and 5 hours for Ferret 7B and 13B models, respectively. Please refer to Appendix for more details on preference data construction and hyperparameter selection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 639 + ], + "type": "text", + "content": "Evaluation Benchmarks We evaluate our method on three types of benchmarks: (1) Grounding tasks that evaluate the localization accuracy, including referring expression comprehension (REC) and phrase grounding; (2) Region description task on Refcocog [22] and visual genome [23], and Ferret Bench [58] for comprehensive spatial understanding; (3) General benchmarks TextVQA [47], GQA [19], LLaVA-Bench [33], and hallucination benchmark POPE [56]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 178, + 661 + ], + "type": "text", + "content": "4.2. Experiments on REC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We first evaluate our method on the referring expression comprehension (REC) task on RefCOCO [22], RefCOCO+ [22], and Refcocog [36]. The task requires the model to locate the object or region given a short de" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 489 + ], + "type": "text", + "content": "scription, which evaluates the model's fine-grained visual grounding abilities under the single-object referent scenarios. As shown in Tab. 1, our proposed SPR framework consistently improves the performance of three baseline MLLMs on different datasets for all model sizes. Considering that the REC results are based on an IoU threshold of 0.5, the improvement on its performance indicates that the model localized more objects successfully. Hence, this improvement can be largely attributed to the introduction of localization scores when constructing the preference data in SPR. Region descriptions that accurately mention more objects could achieve higher localization scores in SPR and be more likely to serve as preferred data, facilitating the model to attend to more objects and their locations in the image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "To better evaluate the impact of SPR on the localization capability of MLLMs, we also conduct REC experiments with higher IoU thresholds by gradually increasing the IoU threshold of valid REC results from the default value of 0.5 to 0.9. As shown in Tab. 2, the improvements brought by SPR significantly increase as the threshold rises, with accuracy gains of 8.96 and 4.06 for the 7B model and the 13B model, respectively, when the IoU threshold rises to 0.9. With SPR, the localization accuracy of the objects in the model's response is greatly improved. Equipped with the grounded region description refinement and the supervision of preferred-rejected localization data in SPR, the model can respond more accurately to grounding object locations, demonstrating the effectiveness of incorporating preference optimization for region description and object localization in the fine-grained spatial understanding of MLLMs." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "725" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 113, + 282, + 179 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 104 + ], + "type": "text", + "content": "Table 3. Experiments on Phrase Grounding task under different IoU thresholds. The results are averaged over the validation and test set of the Flickr30k dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 113, + 282, + 179 + ], + "lines": [ + { + "bbox": [ + 69, + 113, + 282, + 179 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 282, + 179 + ], + "type": "table", + "html": "
IoU Threshold0.50.60.70.80.9
Ferret-7B81.376.1467.5553.8629.98
+SPR82.4477.1469.2556.1933.99
Ferret-13B82.9476.6268.3455.7432.60
+SPR82.7877.2369.7456.9634.18
", + "image_path": "ac696c35aea149b803e666d55e392faab323115fd65792ebe04de87cf42da289.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 221, + 293, + 298 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 213 + ], + "type": "text", + "content": "Table 4. Experiments on the region captioning task on Refcocog and Visual Genome datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 221, + 293, + 298 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 293, + 298 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 293, + 298 + ], + "type": "table", + "html": "
MethodRefcocogVisual Genome
METEORROUGE_LMETEORROUGE_L
Ferret-7B12.315.617.429.6
+SPR13.520.417.629.7
Ferret-13B12.926.417.931.0
+SPR13.327.218.231.3
", + "image_path": "cb7ad5990e0e3cc11e5565b7648a41dee8bdc037dac7c12493a20760cdfe9735.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 243, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 243, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 243, + 332 + ], + "type": "text", + "content": "4.3. Experiments on Phrase Grounding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 337, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 295, + 504 + ], + "type": "text", + "content": "Furthermore, we experiment with the phrase grounding task on Flickr30k Entity [40]. In phrase grounding, the queried object phrases are combined in a single question, requiring MLLMs to detect the locations of multiple objects in a single response, which makes it more challenging than the single-object referring task like REC. Following [58], we adopt the question \"What are the locations of [phrases]?\" and evaluate the result using the MERGE-BOXES mode [21]. Since LLaVA-OneVision and CogVLM do not support multi-object detection, we report only the results for Ferret. As shown in Tab. 1, SPR effectively improves Ferret's performance in multi-object referent scenarios, especially for the 7B model, whose performance is even comparable to that of the 13B model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 661 + ], + "type": "text", + "content": "We then experiment with the phrase grounding task under higher IoU thresholds. We found that the multi-object referencing setting in phrase grounding is more challenging than the single-object referencing in REC. As the IoU threshold increases, the performance drops more rapidly, indicating a significant demand for MLLM to improve the capabilities of more accurate localization. Our approach can significantly alleviate this issue. As shown in Tab. 3, SPR improves progressively with higher IoU thresholds, reaching a maximum gain of 4.01 and 1.58 Recall@1 for the 7B and 13B models, respectively. This experiment demonstrates the superiority of SPR in pursuing detailed descriptions with high-precision object localization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "type": "text", + "content": "4.4. Experiments on Region Captioning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Beyond the grounding task, we also verify our proposed SPR in improving the text qualities of MLLMs' outputs" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 324, + 113, + 543, + 179 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 5. Experiments on the Ferret Bench. \"Description\", \"Reasoning\", and \"Grounding\" denote the Referring Description, Referring Reasoning, and Grounding in Conversation tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 113, + 543, + 179 + ], + "lines": [ + { + "bbox": [ + 324, + 113, + 543, + 179 + ], + "spans": [ + { + "bbox": [ + 324, + 113, + 543, + 179 + ], + "type": "table", + "html": "
ModelDescriptionReasoningGroundingAvg.
Ferret-7B68.767.357.564.5
+ SPR70.068.458.165.5
Ferret-13B70.668.759.766.3
+ SPR70.872.660.267.9
", + "image_path": "30dcbc7d2afd9616c2e1b50bb29ea629056d15dd1d91f96d5af2e50c1c570e18.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 335, + 231, + 533, + 298 + ], + "blocks": [ + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 222 + ], + "type": "text", + "content": "Table 6. Experiments on the general and hallucination benchmarks. We report the accuracy for GQA and VQA and the F1 score on POPE." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 231, + 533, + 298 + ], + "lines": [ + { + "bbox": [ + 335, + 231, + 533, + 298 + ], + "spans": [ + { + "bbox": [ + 335, + 231, + 533, + 298 + ], + "type": "table", + "html": "
ModelVQATGQALLaVAPOPE
Ferret-7B--64.785.36
+ SPR--66.385.69
LLaVA-OV-7B75.8962.2188.988.12
+ SPR76.0762.4291.488.49
", + "image_path": "73ada29cfcffc83458d2dbbd1823dc76d5c91cf28e520a445739f531eccbc146.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 437 + ], + "type": "text", + "content": "on fine-grained spatial understanding. We conduct experiments on the RefCOCOg and Visual Genome benchmarks. We prompt MLLMs with the question \"Describe the region [region] in the image.\" to generate region captions and then evaluate the response quality using METEOR [5] and ROUGE_L [28] metrics. Tab. 4 shows that SPR effectively improves the quality of MLLM-generated region captions. After tuning with SPR, MLLMs are able to effectively attend to the user-specified regions and generate captions that better reflect the details of the region content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 444, + 476, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 476, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 476, + 456 + ], + "type": "text", + "content": "4.5. Experiments on Ferret Bench" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 605 + ], + "type": "text", + "content": "Ferret benchmark, proposed by [58], aims to evaluate MLLMs' fine-grained multimodal conversational capabilities such as referring description, referring reasoning, and grounded conversation. We follow the pipeline in [58] to prompt MLLMs with questions and employ GPT to evaluate the responses. As shown in Tab. 5, the proposed SPR can facilitate MLLMs in achieving better conversational qualities for fine-grained multimodal understanding, especially for the referring reasoning task, with an accuracy gain of about 3.9 for the 13B model. Equipped with SPR, MLLM can focus on more detailed visual information and generate responses that align better with human preferences." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 613, + 515, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 515, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 515, + 625 + ], + "type": "text", + "content": "4.6. Experiments on General Benchmarks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "We further evaluate SPR on three general benchmarks to validate the benefits of improving MLLMs' spatial capabilities. TextVQA and GQA require MLLMs to answer questions or perform reasoning based on specific text, objects, and image content. LLaVA bench evaluates MLLMs comprehensive capabilities in conversation, description, and reasoning. As shown in Tab. 6, improving MLLMs' spa" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "726" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 124, + 264, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 7. Ablation Studies on the refinement of grounded region descriptions, and ratio " + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": " between semantic and localization scores in ranking MLLM responses. We report the average results on Referring expression comprehension and phrase grounding tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 124, + 264, + 228 + ], + "lines": [ + { + "bbox": [ + 89, + 124, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 89, + 124, + 264, + 228 + ], + "type": "table", + "html": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ SPR84.9382.44
w/o Refinement84.4191.38
λ = 0.084.2581.83
λ = 0.484.3481.95
λ = 0.684.6682.13
λ = 0.884.9382.44
λ = 1.084.4581.87
", + "image_path": "8f399fe37d719590d6ab6af66e0309d1f89849786f23f8f7558e5ff6152a85c8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 78, + 258, + 270, + 309 + ], + "blocks": [ + { + "bbox": [ + 85, + 237, + 266, + 249 + ], + "lines": [ + { + "bbox": [ + 85, + 237, + 266, + 249 + ], + "spans": [ + { + "bbox": [ + 85, + 237, + 266, + 249 + ], + "type": "text", + "content": "Table 8. Ablation studies on the training strategy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 258, + 270, + 309 + ], + "lines": [ + { + "bbox": [ + 78, + 258, + 270, + 309 + ], + "spans": [ + { + "bbox": [ + 78, + 258, + 270, + 309 + ], + "type": "table", + "html": "
MethodRECPhrase Grounding
Ferret-7B83.9181.30
+ Instruction Finetuning84.3581.72
+ DPO training84.9382.44
", + "image_path": "07baba7eb626995d30778c63ea1b821a0d03f194033b65ac90903de3736d8099.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 450 + ], + "type": "text", + "content": "tial understanding capabilities consistently enhances their comprehension and reasoning abilities across diverse general scenarios, leading to performance gains on all three benchmarks. We also experiment on hallucination benchmark POPE, where SPR improves both Ferret and LLaVA-OneVision. This can be attributed to the preference data construction in SPR, where semantic and localization scores are applied to select region descriptions that better align with region content and reject those related to content outside the region or that contain hallucinations, thus effectively helping mitigate the hallucinations in MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 156, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 156, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 156, + 468 + ], + "type": "text", + "content": "4.7. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "content": "We conduct ablation studies over the two designs in SPR and evaluate the performance of Ferret-7B on the REC (Refcoco " + }, + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "inline_equation", + "content": "+ / \\mathrm{g}" + }, + { + "bbox": [ + 55, + 474, + 295, + 510 + ], + "type": "text", + "content": ") and the Flickr30k phrase grounding tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": "Score Ratio " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": ". In Sec. 3.2, we combine the semantic and localization scores to rank MLLMs generated descriptions with a score ratio " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": ". As shown in Tab. 7, we vary the " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": " from 0 to 1, and the trained models outperform the baseline model consistently. When " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": " equals zero, SPR achieves minimal gain, as the model might overly reward descriptions that simply list object names or fail to align with the region's semantics. On the other hand, when " + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": " is set to 1, SPR completely disregards measuring how detail the MLLM describes the objects in the region. Under such situations, the model encourages coarse region descriptions with fewer objects involved and reduces the corresponding object bounding box texts in the preferred data, thereby hindering the training of MLLMs' localization capability. As the experiments show, a relatively high value of 0.8 achieves the best results and is set as the default value in SPR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Refinement of Grounded Region Description. After con" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "structuring the preferred and rejected data pairs, we further refine the localization results in the preferred descriptions by completing bounding boxes for objects in the description that were not grounded and refining the existing bounding boxes. Tab. 7 shows the results of this refinement. We found that the refinement leads to greater improvements in the multi-object referring task of phrase grounding. This could be attributed to the fact that the baseline model often fails to follow instructions for providing bounding boxes for each mentioned object when generating region descriptions. After tuning by the refined descriptions, MLLMs could faithfully ground the mentioned objects, thereby improving the multi-object phrase grounding clearly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 235, + 441, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 441, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 441, + 248 + ], + "type": "text", + "content": "4.8. Comparison with SFT" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 252, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 469 + ], + "type": "text", + "content": "In this paper, we adopt DPO with accept-reject preference data to optimize MLLMs for spatial understanding, whereas prior work [10, 44, 61], primarily focuses on the stage of supervised instruction fine-tuning (SFT). In Tab. 8, we compare these two training approaches, where SFT is trained using only the accepted data. The results show that while SFT could improve MLLMs' localization capabilities, its performance gains are significantly lower than DPO. DPO optimizes MLLM by contrasting accepted and rejected data pairs, similar to the positive-negative sample training mechanism in traditional object detection algorithms. This approach helps models distinguish between accurate and inaccurate localizations and facilitates MLLM in spatial understanding more effectively. However, it is important to note that DPO training also relies on a well-trained SFT model as a foundation, making these two approaches complementary. In future work, we will further explore how to integrate SFT and DPO to enhance MLLMs' spatial understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose SPR, a Spatial Preference Rewarding framework to enhance MLLM's fine-grained spatial understanding capabilities. We introduce a complete pipeline that includes (1) Constructing random region queries; (2) Prompting MLLMs to generate diverse grounded region descriptions; (3) Proposing semantic scores and localization scores to rank the descriptions comprehensively; (4) Refining the localization quality of preference data; (5) Finetuning MLLMs to optimize against detailed and accurate spatial understanding. The entire framework does not require additional human labor or external MLLMs, with minimal overhead on training costs. SPR addresses the lack of direct optimization for positive and negative localization samples in MLLM training, enhancing their localization capabilities and promoting better alignment with human preferences. Experiments demonstrate that SPR significantly improves MLLMs' performance on standard referring and grounding tasks for spatial understanding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "727" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 162 + ], + "type": "text", + "content": "This study is also supported by the MOE Tier-2 project, with project number MOE-T2EP20123-0003." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 115, + 194 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 202, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 202, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 202, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 296, + 257 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 357 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 401 + ], + "type": "text", + "content": "[4] Zechen Bai, Pichao Wang, Tianjun Xiao, Tong He, Zongbo Han, Zheng Zhang, and Mike Zheng Shou. Hallucination of multimodal large language models: A survey. arXiv preprint arXiv:2404.18930, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 457, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 491 + ], + "type": "text", + "content": "[6] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşirlar. Introducing our multimodal models, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with Transformers. In ECCV, 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 296, + 646 + ], + "type": "text", + "content": "[9] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[10] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[12] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 205 + ], + "type": "text", + "content": "[13] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 261 + ], + "type": "text", + "content": "[14] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 206, + 553, + 261 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 206, + 553, + 261 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "text", + "content": "[15] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale N Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "text", + "content": "[16] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[17] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 492 + ], + "type": "text", + "content": "[18] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[19] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[20] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[21] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[22] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 3, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "728" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "text", + "content": "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 285 + ], + "type": "text", + "content": "[26] Lei Li, Zhihui Xie, Mukai Li, Shunian Chen, Peiyi Wang, Liang Chen, Yazheng Yang, Benyou Wang, and Lingpeng Kong. Silkie: Preference distillation for large visual language models. arXiv preprint arXiv:2312.10665, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 352 + ], + "type": "text", + "content": "[27] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundingpt: Language enhanced multi-modal grounding model. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6657-6678, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[28] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 419 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In ICCV, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 475 + ], + "type": "text", + "content": "[30] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. In The Twelfth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 510 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 544 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[35] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[36] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[38] Yassine Ouali, Adrian Bulat, Brais Martinez, and Georgios Tzimiropoulos. Clip-dpo: Vision-language models as a source of preference for fixing hallucinations in lvlms. arXiv preprint arXiv:2408.10433, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[39] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[40] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641–2649, 2015. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 386 + ], + "type": "text", + "content": "[41] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "text", + "content": "[43] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 576 + ], + "type": "text", + "content": "[44] Kanchana Ranasinghe, Satya Narayan Shukla, Omid Poursaeed, Michael S Ryoo, and Tsung-Yu Lin. Learning to localize objects improves spatial reasoning in visual-llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12987, 2024. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "text", + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[46] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[47] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Con-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "729" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 294, + 94 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[48] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[50] Fei Wang, Wenxuan Zhou, James Y Huang, Nan Xu, Sheng Zhang, Hoifung Poon, and Muhao Chen. mdpo: Conditional preference optimization for multimodal large language models. arXiv preprint arXiv:2406.11839, 2024. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "text", + "content": "[52] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[54] Junkang Wu, Yuexiang Xie, Zhengyi Yang, Jiancan Wu, Jinyang Gao, Bolin Ding, Xiang Wang, and Xiangnan He. beta-dpo: Direct preference optimization with dynamic beta. arXiv preprint arXiv:2407.08639, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[55] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 546 + ], + "type": "text", + "content": "[56] Li Yifan, Du Yifan, Zhou Kun, Wang Jinpeng, Zhao Wayne Xin, and Wen Ji-Rong. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[57] Guojun Yin, Lu Sheng, Bin Liu, Nenghai Yu, Xiaogang Wang, and Jing Shao. Context and attribute grounded dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6241-6250, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[58] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[59] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, and Tat-Seng Chua. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. arxiv, 2023. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 620 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[60] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[61] Yufei Zhan, Yousong Zhu, Zhiyang Chen, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon: Spelling out all object locations at any granularity with large language models. In European Conference on Computer Vision, pages 405-422. Springer, 2025. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 238 + ], + "type": "text", + "content": "[62] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024.1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[63] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 296, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 296, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 553, + 351 + ], + "type": "text", + "content": "[64] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Chunyuan Li, Jainwei Yang, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2025. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 352, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 352, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 553, + 384 + ], + "type": "text", + "content": "[65] Mengxi Zhang and Kang Rong. Automated multi-level preference for mllms. arXiv preprint arXiv:2405.11165, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "text", + "content": "[66] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 474 + ], + "type": "text", + "content": "[67] Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded video description. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6578-6587, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "text", + "content": "[68] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "text", + "content": "[69] Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. Calibrated self-rewarding vision language models. arXiv preprint arXiv:2405.14622, 2024. 1,3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 620 + ], + "type": "text", + "content": "[70] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 3, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 313, + 757 + ], + "type": "text", + "content": "730" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_content_list.json b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a12f5d54b17304db0f174b424a5faeadfbd42d0 --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_content_list.json @@ -0,0 +1,1792 @@ +[ + { + "type": "text", + "text": "Spatial-Temporal Aware Visuomotor Diffusion Policy Learning", + "text_level": 1, + "bbox": [ + 205, + 160, + 790, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyang Liu $^{1,2}$ , Yikai Wang $^{3}$ , Kuanning Wang $^{1}$ , Longfei Liang $^{4}$", + "bbox": [ + 259, + 205, + 733, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyang Xue $^{1,\\dagger}$ , Yanwei Fu $^{1,2\\dagger}$", + "bbox": [ + 383, + 223, + 617, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fudan University $^{2}$ Shanghai Innovation Institute $^{3}$ Nanyang Technological University $^{4}$ NeuHelium Co., Ltd", + "bbox": [ + 171, + 239, + 825, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liuzy24@m.fudan.edu.cn, yikai.wang@ntu.edu.sg, knwang24@m.fudan.edu.cn", + "bbox": [ + 212, + 258, + 782, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "longfei.liang@neuhelium.com, {xyxue,yanweifu}@fudan.edu.cn", + "bbox": [ + 263, + 275, + 733, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: ZhenyangLiu.github.io/DP4", + "bbox": [ + 344, + 289, + 651, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 335, + 341, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual imitation learning is effective for robots to learn versatile tasks. However, many existing methods rely on behavior cloning with supervised historical trajectories, limiting their 3D spatial and 4D spatiotemporal awareness. Consequently, these methods struggle to capture the 3D structures and 4D spatiotemporal relationships necessary for real-world deployment. In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that incorporates spatiotemporal awareness into diffusion-based policies. Unlike traditional approaches that rely on trajectory cloning, DP4 leverages a dynamic Gaussian world model to guide the learning of 3D spatial and 4D spatiotemporal perceptions from interactive environments. Our method constructs the current 3D scene from a single-view RGB-D observation and predicts the future 3D scene, optimizing trajectory generation by explicitly modeling both spatial and temporal dependencies. Extensive experiments across 17 simulation tasks with 173 variants and 3 real-world robotic tasks demonstrate that the 4D Diffusion Policy (DP4) outperforms baseline methods, improving the average simulation task success rate by $16.4\\%$ (Adroit), $14\\%$ (DexArt), and $6.45\\%$ (RLBench), and the average real-world robotic task success rate by $8.6\\%$ .", + "bbox": [ + 122, + 366, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 125, + 713, + 243, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Imitation learning has emerged as an effective method for training robots to perform complex tasks, such as object grasping [18, 35, 38], legged locomotion [28, 45], dexterous manipulation [12, 29], and mobile manipulation [6, 34]. Recent advances in computer vision have further enhanced visual imitation learning, enabling impress-", + "bbox": [ + 122, + 737, + 483, + 821 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f78a27e752a7b0442405a313f575cb9ed5a83b5392c3d5445cadaad9465b90a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 335, + 869, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6fc01632182ef501da5f9573c1b55e20f80f26291f4e90bbbc9d441f0255193.jpg", + "image_caption": [ + "Figure 1. Spatial-temporal awareness in the 4D Diffusion Policy (DP4). Previous methods train perception and decision-making with trajectory supervision, but trajectory cloning fails to capture the 3D spatial and 4D spatiotemporal relationships. In contrast, DP4 constructs the current 3D scene with 3D spatial supervision from a single RGB-D view and predicts future 3D scene candidates using 4D spatiotemporal supervision, optimizing trajectory generation by effectively capturing both 3D structures and 4D dependencies." + ], + "image_footnote": [], + "bbox": [ + 516, + 448, + 870, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sive performance across various applications [3, 13, 33, 35]. Progress in computer vision has enabled sophisticated visual imitation learning methods that demonstrate excellent performance across various applications [3, 13, 33, 35]. However, real-world robotic deployment demands robust generalization to unseen scenarios, necessitating a deeper understanding of the spatiotemporal physical rules that govern dynamic environments.", + "bbox": [ + 509, + 742, + 870, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their success, existing visual imitation learning", + "bbox": [ + 529, + 853, + 870, + 867 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 117, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 245, + 13, + 799, + 29 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 328, + 30, + 715, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 297, + 44, + 746, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors.", + "bbox": [ + 140, + 832, + 263, + 843 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\ddagger}$ Prof. Yanwei Fu is also with Institute of Trustworthy Embodied AI, and the School of Data Science, Fudan University.", + "bbox": [ + 125, + 844, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7122", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods struggle to capture the 3D structure of environments and 4D spatiotemporal dependencies of interactive tasks. These limitations arise mainly because current approaches rely on behavior cloning from static, supervised trajectories. Such methods often fail to infer the essential physical properties needed for spatial reasoning, object interaction, and environment-aware decision-making.", + "bbox": [ + 122, + 125, + 482, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that integrates spatial-temporal awareness into diffusion-based policies. Unlike prior works that passively mimic demonstrations exclusively from supervised trajectories, DP4 actively learns from a dynamic Gaussian world model [19, 41], which explicitly encodes 3D spatial structures and 4D spatiotemporal variations as part of the imitation learning process. Thus, as illustrated in Fig. 1, DP4 extends conventional imitation learning pipelines in the following ways:", + "bbox": [ + 122, + 222, + 482, + 359 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 3D spatial perception is learned via Gaussian Splitting (3DGS) [19], ensuring accurate reconstruction of the current scene's geometry from point cloud inputs.", + "- 4D spatiotemporal awareness is enforced by future 3DGS reconstruction, where the system learns to predict and adaptively refine future environmental states based on the agent's planned trajectories.", + "- Trajectory generation is conditioned on these structured spatial-temporal representations, allowing the diffusion policy model to reason over future states rather than just replicating past behaviors." + ], + "bbox": [ + 124, + 361, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "So our method first converts RGB-D observations into 3D point clouds, from which we extract global and local spatiotemporal representations. These serve as conditions for the diffusion policy model to guide next-step trajectory generation. To strengthen structured perception:", + "bbox": [ + 124, + 513, + 482, + 583 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel Gaussian regressor that learns to reconstruct the 3D scene from point cloud inputs, ensuring spatially consistent representations.", + "- We enforce temporal consistency by deforming the current scene's 3DGS into the predicted future state, supervising this process using ground-truth RGB-D reconstructions from the future scene." + ], + "bbox": [ + 124, + 585, + 482, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate DP4 extensively across 17 simulated tasks (173 variants) from four benchmarks, along with three real-world tasks, including the challenging task of dexterous manipulation of deformable objects. Our experiments show that DP4 significantly outperforms state-of-the-art multi-task visual imitation methods, achieving a $+16.4\\%$ success rate on Adroit, $+14\\%$ on DexArt, $+6.45\\%$ on RL-Bench, and $+8.6\\%$ improvement in real-world robotic tasks.", + "bbox": [ + 122, + 686, + 482, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work bridges the gap between visual imitation learning and spatial-temporal physical reasoning, demonstrating the importance of structured world models in improving generalization and robustness in real-world robotic applications, with these contributions:", + "bbox": [ + 122, + 797, + 482, + 866 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Spatiotemporal-centric visual imitation learning: Unlike existing methods that overlook 3D perception and dynamic interactions, DP4 explicitly models 3D spatial structures and 4D temporal dynamics, improving generalization in dynamic environments.", + "- 4D Diffusion Policy with structured awareness: We introduce a diffusion-based visual imitation learning framework that generates trajectories based on learned spatial-temporal world representations.", + "- Dynamic Gaussian World Model for structured supervision: Our model actively learns from real-world interactions, ensuring the embedding of 3D spatial and 4D temporal reasoning into policy learning.", + "- State-of-the-art performance across diverse tasks: DP4 sets unprecedented success rates across 17 simulated and real-world robotic tasks, setting a new benchmark for imitation learning in dynamic environments." + ], + "bbox": [ + 511, + 125, + 870, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 373, + 640, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual imitation learning [20-22] is an effective method for equipping robots with skills, typically relying on numerous expert demonstrations as observation-action pairs. Recent advanced visual imitation learning methods commonly employ 2D images[1, 5, 7, 26, 33, 38] as state representations for their rich global information and ease of acquisition from raw sensory inputs, whereas 3D representations (e.g., point clouds and object geometries) [47, 48] are used for their explicit spatial structure, enhancing the spatial reasoning. However, these methods overlook the spatiotemporal dynamics of scenes arising from behavior-environment interactions. DP4 overcomes this limitation by employing a dynamic Gaussian world model to derive scene-level 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction.", + "bbox": [ + 509, + 396, + 870, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models as policies Building on their success in image and video generation [4, 14, 44], diffusion models have emerged as effective policy frameworks in robotics [5, 27, 29, 43, 47]. They are primarily applied in reinforcement learning and visual imitation learning, with Diffusion Policy [5] and variants [40, 47] achieving state-of-the-art results by fusing visual inputs (e.g., 2D images, 3D point clouds) with complex, multi-modal action spaces. Nevertheless, these approaches often underutilize visual perception, relying heavily on behavior cloning from supervised trajectories. In contrast, DP4 introduces a Gaussian world model that reconstructs current scenes from RGB-D data via volumetric rendering for 3D spatial understanding, and employs a deformation Gaussian field to predict future scene distributions through environment interaction, enabling comprehensive 4D spatiotemporal perception.", + "bbox": [ + 509, + 604, + 870, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World model Recent advances demonstrate that world models effectively encode scene dynamics by predicting future states from current states and actions, with appli", + "bbox": [ + 509, + 825, + 870, + 867 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7123", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32729a1c19491f8baf91483ddd502ecc561f35e53eb99941670633d99d84a506.jpg", + "image_caption": [ + "Figure 2. The framework of our 4D Diffusion Policy (DP4). From a single-view RGB-D observation, we construct 3D point clouds and extract global and local features to enrich both holistic and focused perceptions. These multi-level representations condition the diffusion policy model to generate trajectories based on current robot states. We introduce a Gaussian world model in DP4 to capture 3D structures and 4D spatiotemporal relationships. The current observation's 3D Gaussian Splitting (3DGS) is derived via a generalizable Gaussian regressor from point clouds and multi-level features. By enforcing consistency between ground-truth and rendered RGB-D images from this 3DGS, we enhance 3D spatial awareness. Additionally, future 3DGS are predicted from current states using policy-generated trajectories, with rendered RGB-D consistency fostering 4D spatiotemporal awareness. This improved spatiotemporal representation significantly benefits complex tasks such as object grasping and dexterous manipulation." + ], + "image_footnote": [], + "bbox": [ + 169, + 125, + 830, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cations in autonomous driving [15, 16, 39] and robotic manipulation [24, 32, 42]. Early works [8-11, 15, 23] employed latent space autoencoding for future prediction, yielding strong results in both simulated and real environments. Nonetheless, learning accurate latent representations is data-intensive and generally confined to simple tasks due to limited implicit feature capacity. To address these challenges, this study integrates Gaussian Splating [19], known for its efficiency and rendering quality, into world model construction. The resulting 4D Diffusion Policy leverages a dynamic Gaussian world model to capture scene-level 3D structure and 4D spatiotemporal dynamics.", + "bbox": [ + 122, + 569, + 483, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 124, + 747, + 246, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce 4D Diffusion Policy (DP4), a novel visual imitation learning method that emphasizes spatial-temporal awareness in diffusion policies. DP4 employs 3D Gaussian Splatting for informative supervision, enabling the learning of 3D spatial and 4D spatiotemporal perceptions in interactive environments. The DP4 framework is illustrated in Figure 2.", + "bbox": [ + 122, + 770, + 483, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 511, + 568, + 651, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual imitation learning allows robots to efficiently develop human-like skills using observation-action pairs from expert demonstrations. The intelligent agent must interactively predict subsequent continuous trajectories based on the current observation. The visual observation at time step $t$ , denoted $\\mathbf{O}^{(t)} = (\\mathbf{C}^{(t)},\\mathbf{D}^{(t)})$ , is derived from expert demonstrations, where $\\mathbf{C}^{(t)}$ and $\\mathbf{D}^{(t)}$ represent single-view RGB and depth images, respectively. Using the visual observation $\\mathbf{O}^{(t)}$ , the visuomotor policy learning model predicts the optimal action $\\mathbf{a}^{(t)}$ at current time step and generates subsequent optimal continuous action sequence $\\mathbf{a}^{(t + 1)}$ .", + "bbox": [ + 509, + 589, + 870, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To effectively train a visuomotor policy learning model, expert demonstrations are supplied as offline datasets for visual imitation learning, consisting of sample triplets with visual observations and expert actions [26, 40]. Current methods utilize robust visual representations, such as Diffusion Policy and 3D Diffusion Policy, to derive informative latent features for optimal action prediction. However, these methods heavily depend on behavior cloning with supervised historical trajectories. This approach struggles to fully", + "bbox": [ + 509, + 742, + 872, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7124", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "capture 3D structures and often overlooks 4D spatiotemporal dynamics. Consequently, the predicted actions frequently fail to achieve complex human goals, particularly when accurate object interactions are absent. In this study, we introduce the 4D Diffusion Policy, which employs a dynamic Gaussian world model to extract 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction.", + "bbox": [ + 122, + 125, + 483, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-level 3D Spatial Awareness", + "text_level": 1, + "bbox": [ + 124, + 246, + 390, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To capture complex 3D structures in the interactive physical world, DP4 achieves multi-level spatial perception by generating 3D representations from local point clouds and global voxel grids. These 3D representations are derived from an RGB-D image captured by a single-view camera with known extrinsic and intrinsic parameters, enhancing practical applicability in real-world settings. DP4 avoids impractical setups, such as using multiple cameras during training and inference or requiring extensive prior environmental visual data. In both simulated and real-world settings, we acquire depth images from a single camera and convert them into complete point clouds using the camera's extrinsic and intrinsic parameters.", + "bbox": [ + 122, + 267, + 483, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning 3D representations To generate the 3D local representation, DP4 crops the point clouds, retaining only points within a specified bounding box. This allows DP4 to focus on critical scene elements (e.g., the robot arm) and exclude redundant points (e.g., the ground). DP4 encodes the cropped point clouds into compact 3D local representations using a lightweight MLP network, called the 3D Local Encoder. To obtain the 3D global representation, DP4 converts the point clouds at step $t$ into a $100^3$ voxel. The 3D Global Encoder encodes this voxel, outputting the 3D global representation $v^{(t)} \\in \\mathbb{R}^{100^3 \\times 64}$ .", + "bbox": [ + 122, + 448, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enhance 3D spatial awareness To enhance the representation $v^{(t)}$ with global structure and texture information, we employ a Generalizable Gaussian Regressor, which takes the deep volume $v$ as the scene representation to derive the Gaussian world model based on Gaussian Splatting [19]. The Generalizable Gaussian Regressor is learned by rendering RGB and depth images from the constructed Gaussian world model. The entire neural rendering process is described below. Define $v^{(t)}\\mathbf{x}\\in \\mathbb{R}^{128}$ as the sampled 3D feature at point $\\mathbf{x}$ from the 3D global representation $v^{(t)}$ . $v^{(t)}\\mathbf{x}$ is obtained through trilinear interpolation, given the discretized nature of the 3D global representation $v^{(t)}$ . The Generalizable Gaussian Regressor uses the 3D coordinates $\\mathbf{x}$ and the feature $v_{\\mathbf{x}}^{(t)}$ to parameterize the Gaussian primitive as $\\theta^{(t)} = (\\mu^{(t)},c^{(t)},r^{(t)},s^{(t)},\\sigma^{(t)})$ , representing the position, color, rotation, scale, and opacity of the Gaussian primitive. To render a single view, we project Gaussian primitives onto the 2D plane using differential tile-based rasterization. The pixel value $\\mathbf{p}$ is rendered using alpha", + "bbox": [ + 122, + 602, + 483, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "blend rendering:", + "bbox": [ + 511, + 125, + 614, + 138 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} c _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 145, + 870, + 191 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} d _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}),\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 189, + 807, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{i}^{(t)} = \\sigma_{i}^{(t)}\\exp -\\frac{1}{2} (\\mathbf{p} - \\mu_{i}^{(t)})^{\\top}\\Sigma_{i}^{-1}(\\mathbf{p} - \\mu_{i}^{(t)})$ defines the 2D density of Gaussian points during the splatting process, $C^{(t)}$ and $D^{(t)}$ represent the rendered RGB image and depth, respectively, $N$ is the number of Gaussians per tile, $d_{i}^{(t)}$ is the $z$ -value obtained by transforming coordinate $u$ from world to camera coordinates, and $\\Sigma_{i}$ denotes the covariance matrix derived from the Gaussian parameters' rotation and scale. Our Generalizable Gaussian Regressor and 3D representation are optimized to reconstruct RGB and depth images from a single view by minimizing a 3D loss function:", + "bbox": [ + 509, + 234, + 870, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {3 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t)} \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 396, + 870, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{C}^{*(t)}$ and $\\mathbf{D}^{*(t)}$ represent the ground truth color and depth at time step $t$ . This 3D loss allows the 3D representation to emphasize local visual features while incorporating global structural knowledge and texture information.", + "bbox": [ + 509, + 434, + 870, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 4D Spatiotemporal Awareness", + "text_level": 1, + "bbox": [ + 511, + 498, + 756, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To emphasize 4D spatiotemporal awareness in the diffusion policy, DP4 integrates dynamics into the Gaussian world model described in Section 3.2. This model parameterizes the Gaussian mixture distribution for dynamic Gaussian splatting, enabling future scene reconstruction via parameter propagation. Thus, the model gains valuable supervision in interactive environments by ensuring consistency between reconstructed and real feature scenes. Specifically, the model learns environmental dynamics by predicting the future state $s^{(t + 1)}$ from the current state $s^{(t)}$ and action $a^{(t)}$ at time step $t$ .", + "bbox": [ + 509, + 518, + 870, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the dynamic Gaussian world model, we define the current state as the visual observation at the current time step, with actions representing the robot components. These observations and actions predict visual scenes for the next time step, representing the future state. Building on the Gaussian world model, DP4 uses a deformable MLP $p_{\\phi}$ to estimate changes in Gaussian parameters during time propagation and a Gaussian renderer $\\mathcal{R}$ to generate RGB and depth images for the future state:", + "bbox": [ + 509, + 670, + 870, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta^ {(t)} = p _ {\\phi} \\left(\\theta^ {(t)}, a ^ {(t)}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 801, + 718, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {(t + 1)} = \\theta^ {(t)} + \\Delta \\theta^ {(t)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 828, + 870, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC ^ {(t + 1)} (\\mathbf {p}), D ^ {(t + 1)} (\\mathbf {p}) = \\mathcal {R} \\left(\\theta^ {(t + 1)}, \\mathbf {p}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 847, + 816, + 869 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7125", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{C}^{(t + 1)}(\\mathbf{p})$ and $\\mathbf{D}^{(t + 1)}(\\mathbf{p})$ represent the predicted RGB and depth images at time step $t + 1$ , $\\theta^{(t)}$ represents the 3D Gaussian primitive parameters at time step $t$ , $w$ denotes the camera pose used to project the Gaussian primitives, and $\\mathcal{R}$ is the rendering process from Equation 1. DP4 employs multi-head neural networks as the Gaussian regressor, with each head predicting distinct features of the Gaussian parameters. By estimating changes in position and rotation between time steps, we derive the Gaussian parameters for the future step. The Gaussian renderer then projects the propagated Gaussian distribution onto the single view to reconstruct the future scene. To model 4D spatiotemporal relationships in the physical world, we enforce consistency supervision between reconstructed and real scenes at time step $t + 1$ , embedding scene-level spatiotemporal dynamics into the Gaussian parameters. Specifically, the training objective aligns predicted future scenes, derived from current observations and actions, with real scenes, as formulated below:", + "bbox": [ + 122, + 123, + 486, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {4 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t + 1)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t + 1)} \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 393, + 483, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{C}^{(t + 1)}$ and $\\mathbf{D}^{(t + 1)}$ represent the real future counterparts. This additional task of predicting future scenes from current observations via the dynamic Gaussian world model serves as a spatiotemporal enhancement penalty. Training with this task, the multi-level 3D representation learns to encode physical properties and develop awareness of 4D spatiotemporal dynamics.", + "bbox": [ + 122, + 431, + 483, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Diffusion-based Decision", + "text_level": 1, + "bbox": [ + 124, + 537, + 333, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The multi-level 3D representation, including local and global 3D representations, is optimized to predict the robot's optimal actions. We model this process as a conditional denoising diffusion framework, conditioned on the multi-level 3D representation $r$ and the robot state $q$ , which converts randomly sampled Gaussian noise into actions $a$ . Starting with sampled Gaussian noise $a^K$ , the denoising network $\\epsilon_{\\theta}$ performs $K$ iterations to progressively convert it into the noise-free action $a^0$ .", + "bbox": [ + 122, + 558, + 485, + 683 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, \\mathrm {I}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 691, + 483, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(0,\\mathbf{I})$ represents the sampled Gaussian noise, and $\\alpha_{k},\\gamma_{k}$ , and $\\sigma_{k}$ are functions of $k$ defined by the noise scheduler. To train the denoising network $\\epsilon_{\\theta}$ , we randomly sample a data point $a^0$ from the dataset and use a diffusion process to generate the noise $\\epsilon^k$ at iteration $k$ . The training objective is to predict the noise added to the data.", + "bbox": [ + 122, + 719, + 483, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a c t i o n}} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha} _ {k} a ^ {0} + \\bar {\\beta} _ {k} \\epsilon^ {k}, k, r, q\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 811, + 483, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_k$ and $\\bar{\\beta}_k$ are noise schedule that performs one step noise adding.", + "bbox": [ + 122, + 838, + 483, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall learning objective for DP4 is as follows:", + "bbox": [ + 529, + 125, + 846, + 138 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D P 4}} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {3 D}} \\mathcal {L} _ {\\mathrm {3 D}} + \\lambda_ {\\mathrm {4 D}} \\mathcal {L} _ {\\mathrm {4 D}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 148, + 870, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{3\\mathrm{D}}$ and $\\lambda_{4\\mathrm{D}}$ balance the 3D loss for current scenes and the 4D loss for future scenes across different objectives. We train DP4 using a joint training strategy. Empirical observations show that this strategy enhances information fusion while learning shared features. During training, we introduce a warm-up phase, freezing the deformable mapper for the first 500 iterations to establish a stable multi-level 3D compact representation and Gaussian regressor. After the warm-up phase, we simultaneously train the entire Gaussian world model and the diffusion-based decision module. After optimization, DP4 generates optimal actions with spatiotemporal awareness, including both 3D spatial and 4D spatiotemporal perceptions.", + "bbox": [ + 509, + 174, + 872, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 367, + 633, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 390, + 673, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Simulations For simulation experiments, we gathered 22 tasks across various domains, covering a wide range of robotic skills. These tasks span challenging scenarios, such as bimanual, deformable object, and articulated object manipulation, as well as simpler tasks like parallel gripper manipulation. The tasks include Adroit tasks [30] (Door, Pen, Hammer), DexArt tasks [2] (Laptop, Faucet, Bucket, Toilet), and RLBench tasks [17] (10 challenging tasks, each with at least two variations, totaling 166).", + "bbox": [ + 509, + 410, + 872, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real robot We used a KINOVA GEN2 robot with a Realsense D455 depth camera mounted in an eye-to-hand configuration. In an indoor environment, we arranged several bottles and cups to enable the agent to generalize manipulation skills across scenes, designing three tasks, including Grasping bottles, Pouring the water from the bottle into the cup and stacking cups. The RealSense D455 camera records the entire scene and robot state from a third-person perspective. This setup supplies RGB-D observations for policy training and RGB and depth supervision for constructing the dynamic Gaussian world model.", + "bbox": [ + 509, + 535, + 872, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expert demonstration Expert trajectories for ManiSkill and Adroit are collected using RL-trained agents, using VRL3 [37] for Adroit and PPO [31] for other domains. We generate successful trajectories with RL agents and ensure that all visual imitation learning algorithms use identical demonstrations. Additionally, we collect 20 demonstrations for each RLBench task using the motion planner.", + "bbox": [ + 509, + 686, + 870, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines The main objective of DP4 is to highlight the importance of 3D spatial and 4D spatiotemporal awareness in diffusion policies. To this end, our primary baseline is the point cloud-based 3D Diffusion Policy (DP3) [47]. Additionally, we incorporate comparisons with other visual imitation learning methods: IBC [7], BCRNN [26], Diffusion", + "bbox": [ + 509, + 784, + 872, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7126", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy (DP) [5], PerAct [35], GNFactor [46], and Mani-Gaussian [25]. Note that image and depth resolutions for all 2D and 3D methods remain consistent across experiments, ensuring fair comparisons.", + "bbox": [ + 122, + 125, + 482, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric Each experiment is conducted using three seeds. For Adroit and DexArt simulation tasks, we assess 20 episodes every 200 epochs and calculate the mean of the top five success rates. For RLBench tasks, we evaluate 25 episodes per task at the final checkpoint for 10 challenging tasks. We report the mean and standard deviation of success rates across the three seeds.", + "bbox": [ + 122, + 181, + 482, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We implement a convolutional network-based diffusion policy using DDIM[36] as the noise scheduler, preferring sample prediction to enhance high-dimensional action generation. Training employs 100 timesteps, while inference uses 10. Models are trained for 3,000 epochs with a batch size of 32 on both simulated and real-world tasks. Loss hyperparameters $\\lambda_{\\mathrm{3D}}$ and $\\lambda_{\\mathrm{4D}}$ are set to 0.1 and 0.01, respectively (see Section 4.4 for details). For fair comparison with DP3[47], we adopt the same 3D Local Encoder: a three-layer MLP, max-pooling for order-equivariant aggregation, and a projection head generating compact 3D local features, with LayerNorm inserted for training stability. All experiments run on an NVIDIA H100 80GB GPU and a 192-vCPU Intel Xeon Platinum 8468.", + "bbox": [ + 122, + 279, + 482, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 124, + 483, + 336, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our DP4 with previous state-of-the-art methods on the Adroit, DexArt, and RLBench simulation tasks.", + "bbox": [ + 122, + 505, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Adroit simulation Table 1 presents the success rates for the Adroit simulation tasks. DP4 achieves an overall success rate of $84.7\\%$ , significantly outperforming 2D-based methods, including IBC [7], BCRNN [26], and Diffusion Policy [5], while surpassing 3D-based methods such as 3D Diffusion Policy [47]. Experimental results demonstrate that DP4 completes complex tasks more accurately by leveraging spatiotemporal perception.", + "bbox": [ + 122, + 532, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DexArt simulation Table 2 presents the success rates for various DexArt simulation tasks. The DP4 achieves a success rate of $82.5\\%$ , significantly outperforming 2D-based methods (e.g., IBC, BCRNN, and Diffusion Policy) by $33.5\\%$ and 3D-based methods (e.g., 3D Diffusion Policy) by $14.0\\%$ . Experimental results demonstrate that DP4 integrates 3D spatial and 4D spatiotemporal perception of the physical world with diffusion policies, achieving higher success rates across multiple tasks.", + "bbox": [ + 122, + 645, + 482, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RLBench simulation We compare DP4 with previous state-of-the-art methods on the RLBench task suite. Table 3 presents a comparison of average success rates across tasks. DP4 achieves state-of-the-art performance with average success rates of $63.3\\%$ and $39.9\\%$ , significantly outperforming prior methods, including both perception-based and generative approaches. Leveraging supervision from", + "bbox": [ + 122, + 770, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26ec84e32d91961f541de5582ca89d0229f91c577c06d4b12b2cb72a27b37f64.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHammerDoorPenOverall
IBC [7]0±00±09±23.0
BCRNN [26]0±00±09±33.0
DP [5]48±1750±525±441.0
DP3 [47]100±062±443±668.3
DP4 (ours)100±080±275±384.7
", + "bbox": [ + 514, + 122, + 870, + 213 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/41d8fe96f2ff47889d4244ee0f6dafaeb9468f7ae44711f4bd48176b45cf5d05.jpg", + "table_caption": [ + "Table 1. Success rates $(\\%)$ on the Adroit simulation. We compare our DP4 with more baselines in the Adroit simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + ], + "table_footnote": [], + "table_body": "
MethodLaptopFaucetBucketToiletOverall
IBC [7]3±27±114±10±06.0
BCRNN [26]3±31±05±50±02.3
DP [5]69±423±858±246±149.0
DP3 [47]83±163±282±446±268.5
DP4 (ours)92±284±190±164±282.5
", + "bbox": [ + 514, + 277, + 867, + 366 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61712f5da278a9047efda638a268e88fdc1d571b439824301cb618e6feb76a68.jpg", + "table_caption": [ + "Table 2. Success rates $(\\%)$ on the DexArt simulation. We compare our DP4 with more baselines in the DexArt simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + ], + "table_footnote": [], + "table_body": "
Methodclose jaropen drawersweep to dustpanmeat off grillturn tapOverall
PerAct [35]18.7±8.254.7±18.60.0±0.040.0±17.038.7±6.830.4
GNFactor [46]25.3±6.876.0±5.728.0±15.057.3±18.950.7±8.247.5
ManiGaussian [25]28.4±5.476.3±6.264.5±13.460.2±18.256.2±6.857.1
DP4 (ours)35.4±2.282.2±3.472.0±5.264.4±6.862.4±2.863.3
Methodslide blockput in drawerdrag stickpush buttonsstack blocksOverall
PerAct [35]18.7±13.62.7±3.35.3±5.018.7±12.46.7±1.910.4
GNFactor [46]20.0±15.00.0±0.037.3±13.218.7±10.04.0±3.316.0
ManiGaussian [25]24.3±12.816.6±3.292.3±11.420.4±12.212.4±3.233.2
DP4 (ours)33.4±4.422.8±1.294.4±6.228.3±6.821.0±2.239.9
", + "bbox": [ + 511, + 426, + 875, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Success rates $(\\%)$ on the RLBench simulation. We compare our DP4 with more baselines in the RLBench simulation. Baseline approaches include PerAct, GNFactor, and Mani-Gaussian.", + "bbox": [ + 509, + 566, + 870, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D geometry, texture, and 4D spatiotemporal data, DP4 accurately predicts optimal continuous trajectories from a single RGB-D observation of the current scene.", + "bbox": [ + 509, + 628, + 870, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 509, + 679, + 714, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization of trajectories We present three qualitative examples of action sequences generated by DP3 and our DP4 in Figure 3. The results show that the previous agent (DP3) frequently fails to complete tasks, as it imitates the expert's motions (e.g., hammering nails, sliding doors, and spinning pens) without fully understanding the scene. In contrast, DP4 significantly improves the success rate by accurately interpreting 3D scene structures (e.g., nail positions) and 4D dynamics (e.g., pen rotation).", + "bbox": [ + 509, + 700, + 870, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization of Gaussian world model Figure 4 and Figure 5 show the visualization results. We visualize the Gaussian world model via single-view rendering, observ", + "bbox": [ + 509, + 825, + 870, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7127", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33406afdd17829ceaf93beed74cf9ca0eb168bce9c918ed14d8bcbe980ba0604.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 123, + 480, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55abc5184a14d7a6e1bf3652b917f8b8eac57d2eec1b503c782c03e053d070df.jpg", + "image_caption": [ + "Figure 4. Visualization of Gaussian world model. After training DP4, we visualize its Gaussian world model by rendering a single-view observation. DP4 uses a Gaussian world model to provide supervision, facilitating the learning of 3D spatial perception." + ], + "image_footnote": [], + "bbox": [ + 127, + 250, + 478, + 373 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b0358591e8cc5afbac5a33bbb1b82c221d0c714080cb4774ccf1f026f580718.jpg", + "image_caption": [ + "Figure 3. Qualitative case study. The red mark indicates a pose that significantly deviates from the expert demonstration, while the green mark denotes a pose that aligns with the expert trajectory. The 4D Diffusion Policy (DP4) integrates 3D spatial and 4D spatiotemporal awareness with diffusion policies, successfully completing the tasks." + ], + "image_footnote": [], + "bbox": [ + 129, + 378, + 478, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing it captures fundamental spatial structures and spatiotemporal dynamics. Although DP4 operates under a realistic single-view constraint, resulting renderings lack fine detail, they effectively represent scene-level 3D structures and 4D dynamics. Crucially, such details have minimal impact on task success; for example, the nail's thread is irrelevant to the hammering task.", + "bbox": [ + 122, + 598, + 483, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 124, + 704, + 264, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Component ablation. Ablations are conducted on the Adroit simulation tasks in Table 4. Without the proposed components, we use 3D representation without any 3D spatial supervision (RGB and depth (structure)) and 4D spatiotemporal supervision (dynamics). In the experiment on the Adroit simulation, the performance of the \"Hammer\", \"Door\", and \"Pen\" tasks demonstrated improvements in success rates with the proposed components, while the inference time remained stable or only slightly increased.", + "bbox": [ + 122, + 724, + 483, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- For the Hammer task: The base model started with a", + "bbox": [ + 124, + 853, + 483, + 866 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/883f8590c4d5270f6a24269dc21164be532fed6fe62a7ed559c2bbe3f62c3aed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 125, + 867, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "94.0% success rate and a completion time of 6.40 seconds. After adding the Gaussian world model with RGB supervision, success improved to $96.0\\%$ , with a minimal increase in time (6.52 seconds). Introducing 3D spatial loss (RGB and Depth supervision) raised success to $95.0\\%$ , with stable time (6.53 seconds). With 4D spatiotemporal supervision (DP4), success reached $98.0\\%$ , and the time increased slightly to 6.56 seconds, indicating a clear performance gain with a small time increase.", + "bbox": [ + 522, + 375, + 870, + 498 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the Door task: Initially, the model achieved a $94.0\\%$ success rate and a completion time of 6.40 seconds. Adding the Gaussian world model with RGB supervision improved success to $96.0\\%$ , with no significant time change. Incorporating 3D spatial loss (RGB and Depth) increased success to $98.0\\%$ , with a slight increase in time (6.56 seconds). With 4D spatiotemporal supervision (DP4), the success rate reached $100.0\\%$ , with a minimal time increase (6.57 seconds), reflecting significant improvement from additional supervision.", + "- For the Pen task: The base model had a $45.0\\%$ success rate and a completion time of 6.46 seconds. After adding the Gaussian world model with RGB supervision, success rose to $48.0\\%$ , with time remaining at 6.48 seconds. Introducing 3D spatial loss (RGB and depth) led to a slight decrease in success $(47.0\\%)$ . With 4D spatiotemporal supervision (DP4), success jumped to $75.0\\%$ , with a small increase in time to 6.50 seconds, highlighting the effectiveness of 4D supervision." + ], + "bbox": [ + 511, + 500, + 870, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, DP4 does not substantially increase model inference time. This is because DP4 does not construct a Gaussian world model during inference. The 3D spatial and 4D spatiotemporal awareness from the Gaussian world model are efficiently integrated into the multi-level 3D representation during training.", + "bbox": [ + 509, + 768, + 870, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter analysis We conduct hyperparameter", + "bbox": [ + 511, + 853, + 870, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7128", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6635bc0adfca8423a5256aee603baba3cf156b81b45d6c002b5d7ddd0c35a977.jpg", + "image_caption": [ + "Figure 5. More visualizations of the learned world model. DP4 uses a Gaussian world model to provide supervision, facilitating the 4D spatiotemporal perceptions in interactive environments." + ], + "image_footnote": [], + "bbox": [ + 129, + 136, + 197, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f687a09c42ab7486e7610126eba30686b9c1537cebb78bcc95dccbdb2714b3b.jpg", + "image_caption": [ + "Ground-truth Prediction at current time step T at current time step T" + ], + "image_footnote": [], + "bbox": [ + 202, + 122, + 269, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/453b535e30b31baa9c589d11f4efc24c74d5502bd15fbd7dfb24f4a742049018.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 136, + 338, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cec220ca16b1cf6c315ae503fa2472c3ceed9e504c40570ab5461ecd4acbcd47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 123, + 408, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/957bce2bec6f7cab2553fa72623cdc8f29ef9b55a7a6696381516a5f70741d24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 136, + 477, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/03505aa83777b4f0a93c49bad4022764651375bdf477909cfe08ccf3e59fef4a.jpg", + "table_caption": [ + "Results on Adroit Simulation" + ], + "table_footnote": [], + "table_body": "
3D Component4D ComponentPerformance
RGBDepthDynamicsHammerDoorPen
94.0/6.4064.0/6.4445.0/6.46
✓96.0/6.5268.0/6.4548.0/6.48
✓95.0/6.5368.0/6.4647.0/6.48
✓✓98.0/6.5675.0/6.5172.0/6.47
✓✓✓100.0/6.5780.0/6.4575.0/6.50
", + "bbox": [ + 127, + 252, + 478, + 366 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/605edf3e11de05e6e61990f0b495689c2878ef6deb6e91c7e3de02c5dda5d763.jpg", + "image_caption": [ + "Table 4. Component ablation studies. We report results for the Adroit simulation tasks. In the performance metrics, the number to the left of / represents the success rate $(\\%)$ , and the number to the right denotes the inference speed (seconds per task).", + "Figure 6. Hyperparameter analysis of DP4. The effects of $\\lambda_{3\\mathrm{D}}$ and $\\lambda_{4\\mathrm{D}}$ on success rates, respectively." + ], + "image_footnote": [], + "bbox": [ + 127, + 421, + 305, + 534 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c23d20c7befb132e4a8f5b7d43aff860fc7336241d097142d17882a6d71e23f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 421, + 480, + 534 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/89cf71df164e91b6976c3076bf1790ee84f4910a3d2405a6d105968b4206dd1a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGrasping BottlesStacking CupsPouring WaterOverall
DP [5]36.044.028.036.0
DP3 [47]42.062.034.046.0
DP4 (ours)48.072.044.054.6
", + "bbox": [ + 127, + 563, + 482, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Success rates $(\\%)$ on the real-robot tasks. We compare our DP4 with additional baselines on the real-robot tasks. We include the Diffusion Policy (DP) and 3D Diffusion Policy (DP3).", + "bbox": [ + 122, + 643, + 483, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "analysis on the Pen task within the Adroit simulation. DP4 uses two key hyperparameters, $\\lambda_{3\\mathrm{D}}$ and $\\lambda_{4\\mathrm{D}}$ , to balance the 3D spatial loss for current scenes and the 4D spatiotemporal loss for future scenes across multiple objectives. As shown in Figure 6, we analyze the hyperparameters of DP4. We vary $\\lambda_{3\\mathrm{D}}$ from 0 to 1 and $\\lambda_{4\\mathrm{D}}$ from 0 to 0.1, measuring the corresponding success rates. We find that the success rate peaks when $\\lambda_{3\\mathrm{D}} = 0.1$ and $\\lambda_{4\\mathrm{D}} = 0.01$ , so we adopt these values as the hyperparameters.", + "bbox": [ + 122, + 685, + 483, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Real-Robot Analysis", + "text_level": 1, + "bbox": [ + 124, + 819, + 303, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess DP4's performance on real-world robotic tasks, we designed three tasks: Grasping Bottles, Stacking Cups,", + "bbox": [ + 122, + 839, + 483, + 868 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0006c5f3ef97680ed3f00549108ce39aea3b25706ca016c58abd799e9c5a29e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 125, + 633, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98f0d77bd8b174e1eba639289184742a9e61457dca5d4b8dcf066eb07ef8ae1d.jpg", + "image_caption": [ + "\"Grasping Bottles\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 125, + 746, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84f69a1a1bf6d24fed2441dfb88af1390ac474edccf3bf87883ccf0c8af9e3d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 126, + 859, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb06977a3c8925dfd7df526a5a5e029eecbe37dd617e96e06055036b4a6d5949.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 210, + 633, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd1e93b1926e181abac301acf7640fa77227c481de793693ab633cc4dd3e2c81.jpg", + "image_caption": [ + "\"Stacking Cups\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 210, + 745, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1e9fa26f7d2e318e2915001e96357d37d0444f893c3d4441254bc6ec8586319.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 210, + 857, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36e04d4ea11b207f564179f00f38c151034526724033846e24e2dcb27ddc58bf.jpg", + "image_caption": [ + "Figure 7. Visualization of DP4 performance on three real-world robotic tasks. DP4 demonstrates strong performance in real-world settings and effectively handles a variety of common tasks with a single view." + ], + "image_footnote": [], + "bbox": [ + 524, + 296, + 633, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab9c019b907ac951141fc7fe64dd502d14a44dda55550f9defa5f81f13d05780.jpg", + "image_caption": [ + "\"Pouring Water\"" + ], + "image_footnote": [], + "bbox": [ + 638, + 296, + 745, + 362 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f3e10bc985da41cf79c0927b4c6aa5a00f61487a1dc178e95246ce839fabd03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 296, + 857, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and Pouring Water. We focus on the model's performance at key frames, where success is defined as when the error is below a certain threshold. The results of the real-world robotic experiments are summarized in Table 5. Experimental results show that DP4 outperforms the Diffusion Policy (DP) and 3D Diffusion Policy (DP3) baselines across all tasks. These results confirm that DP4 effectively integrates spatiotemporal awareness with diffusion policies, improving the accuracy of generated trajectories. Additionally, visualizations of DP4's performance on real-world tasks are shown in Figure 7.", + "bbox": [ + 509, + 446, + 870, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 620, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents 4D Diffusion Policy (DP4), which integrates spatiotemporal awareness of the physical world with diffusion-based visual imitation learning. Utilizing a single-view RGB-D input, DP4 constructs a dynamic Gaussian world model that provides rich supervision for acquiring 3D spatial and 4D spatiotemporal perceptions from interactive environments. Extensive experiments on simulated and real-robot tasks validate the effectiveness and superiority.", + "bbox": [ + 509, + 643, + 870, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 772, + 674, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the Science and Technology Commission of Shanghai Municipality (No.24511103100), NSFC Project (62176061), Doubao Fund, and Shanghai Technology Development and Entrepreneurship Platform for Neuromorphic and AI SoC.", + "bbox": [ + 509, + 797, + 870, + 868 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7129", + "bbox": [ + 482, + 914, + 514, + 926 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 125, + 123, + 212, + 137 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sridhar Pandian Arunachalam, Sneha Silwal, Ben Evans, and Lerrel Pinto. Dexterous imitation made easy: A learning-based framework for efficient dexterous manipulation. In 2023 IEEE international conference on robotics and automation (icra), pages 5954-5961. IEEE, 2023. 2", + "[2] Chen Bao, Helin Xu, Yuzhe Qin, and Xiaolong Wang. Dexart: Benchmarking generalizable dexterous manipulation with articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21190-21200, 2023. 5", + "[3] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1", + "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 2", + "[5] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8", + "[6] Yuqing Du, Daniel Ho, Alex Alemi, Eric Jang, and Mohi Khansari. Bayesian imitation learning for end-to-end mobile manipulation. In International Conference on Machine Learning, pages 5531-5546. PMLR, 2022. 1", + "[7] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on robot learning, pages 158-168. PMLR, 2022. 2, 5, 6", + "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 3", + "[9] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019.", + "[10] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020.", + "[11] Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems, 35:26091-26104, 2022. 3", + "[12] Siddhant Haldar, Jyothish Pari, Anant Rai, and Lerrel Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations. arXiv preprint arXiv:2303.01497, 2023. 1", + "[13] Nicklas Hansen, Zhecheng Yuan, Yanjie Ze, Tongzhou Mu, Aravind Rajeswaran, Hao Su, Huazhe Xu, and Xiaolong Wang. On pre-training for visuo-motor control: Revisiting a learning-from-scratch baseline. arXiv preprint arXiv:2212.05749, 2022. 1", + "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif" + ], + "bbox": [ + 125, + 146, + 485, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[15] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems, 35:20703-20716, 2022. 3", + "[16] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3", + "[17] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5, 4", + "[18] Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE international conference on robotics and automation (ICRA), pages 4613-4619. IEEE, 2021. 1", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4", + "[20] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2", + "[21] Peiyan Li, Hongtao Wu, Yan Huang, Chilam Cheang, Liang Wang, and Tao Kong. Gr-mg: Leveraging partially-annotated data via multi-modal goal-conditioned policy. IEEE Robotics and Automation Letters, 2025.", + "[22] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 2", + "[23] Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Reasongrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3718-3727, 2025. 3", + "[24] Zhenyang Liu, Sixiao Zheng, Siyu Chen, Cairong Zhao, Longfei Liang, Xiangyang Xue, and Yanwei Fu. A neural representation framework with llm-driven spatial reasoning for open-vocabulary 3d visual grounding. arXiv preprint arXiv:2507.06719, 2025. 3", + "[25] Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In European Conference on Computer Vision, pages 349-366. Springer, 2024. 6", + "[26] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio" + ], + "bbox": [ + 513, + 126, + 870, + 867 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7130", + "bbox": [ + 483, + 915, + 514, + 926 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 2, 3, 5, 6", + "[27] Liyuan Mao, Haoran Xu, Xianyuan Zhan, Weinan Zhang, and Amy Zhang. Diffusion-dice: In-sample diffusion guidance for offline reinforcement learning. arXiv preprint arXiv:2407.20109, 2024. 2", + "[28] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020.1", + "[29] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 1, 2", + "[30] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087, 2017. 5", + "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5", + "[32] Younggyo Seo, Danijar Hafner, Hao Liu, Fangchen Liu, Stephen James, Kimin Lee, and Pieter Abbeel. Masked world models for visual control. In Conference on Robot Learning, pages 1332-1344. PMLR, 2023. 3", + "[33] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. 1, 2", + "[34] Nur Muhammad Mahi Shafullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, Soumith Chintala, and Lerrel Pinto. On bringing robots home. arXiv preprint arXiv:2311.16098, 2023. 1", + "[35] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 6", + "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 6", + "[37] Che Wang, Xufang Luo, Keith Ross, and Dongsheng Li. Vrl3: A data-driven framework for visual deep reinforcement learning. Advances in Neural Information Processing Systems, 35:32974-32988, 2022. 5", + "[38] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimi-cplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1, 2", + "[39] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 3" + ], + "bbox": [ + 125, + 125, + 483, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning. arXiv preprint arXiv:2208.06193, 2022. 2, 3", + "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2", + "[42] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226-2240. PMLR, 2023. 3", + "[43] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 2", + "[44] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2", + "[45] Ruihan Yang, Zhuoqun Chen, Jianhan Ma, Chongyi Zheng, Yiyu Chen, Quan Nguyen, and Xiaolong Wang. Generalized animal imitator: Agile locomotion with versatile motion prior. arXiv preprint arXiv:2310.01408, 2023. 1", + "[46] Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. In Conference on Robot Learning, pages 284-301. PMLR, 2023. 6", + "[47] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024. 2, 5, 6, 8, 3", + "[48] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37:77799-77830, 2025. 2" + ], + "bbox": [ + 514, + 126, + 870, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7131", + "bbox": [ + 483, + 915, + 513, + 926 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_model.json b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee0e483a27191f491d70527c91616e4d926bb53c --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_model.json @@ -0,0 +1,2464 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.118, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.03 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.33, + 0.031, + 0.717, + 0.044 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.299, + 0.045, + 0.748, + 0.058 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.161, + 0.792, + 0.182 + ], + "angle": 0, + "content": "Spatial-Temporal Aware Visuomotor Diffusion Policy Learning" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.206, + 0.735, + 0.224 + ], + "angle": 0, + "content": "Zhenyang Liu\\(^{1,2}\\), Yikai Wang\\(^{3}\\), Kuanning Wang\\(^{1}\\), Longfei Liang\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.385, + 0.224, + 0.619, + 0.24 + ], + "angle": 0, + "content": "Xiangyang Xue\\(^{1,\\dagger}\\), Yanwei Fu\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.24, + 0.826, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Fudan University \\(^{2}\\)Shanghai Innovation Institute \\(^{3}\\)Nanyang Technological University \\(^{4}\\)NeuHelium Co., Ltd" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.259, + 0.783, + 0.272 + ], + "angle": 0, + "content": "liuzy24@m.fudan.edu.cn, yikai.wang@ntu.edu.sg, knwang24@m.fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.276, + 0.734, + 0.289 + ], + "angle": 0, + "content": "longfei.liang@neuhelium.com, {xyxue,yanweifu}@fudan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.29, + 0.652, + 0.305 + ], + "angle": 0, + "content": "Project Page: ZhenyangLiu.github.io/DP4" + }, + { + "type": "title", + "bbox": [ + 0.269, + 0.337, + 0.342, + 0.351 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.367, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Visual imitation learning is effective for robots to learn versatile tasks. However, many existing methods rely on behavior cloning with supervised historical trajectories, limiting their 3D spatial and 4D spatiotemporal awareness. Consequently, these methods struggle to capture the 3D structures and 4D spatiotemporal relationships necessary for real-world deployment. In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that incorporates spatiotemporal awareness into diffusion-based policies. Unlike traditional approaches that rely on trajectory cloning, DP4 leverages a dynamic Gaussian world model to guide the learning of 3D spatial and 4D spatiotemporal perceptions from interactive environments. Our method constructs the current 3D scene from a single-view RGB-D observation and predicts the future 3D scene, optimizing trajectory generation by explicitly modeling both spatial and temporal dependencies. Extensive experiments across 17 simulation tasks with 173 variants and 3 real-world robotic tasks demonstrate that the 4D Diffusion Policy (DP4) outperforms baseline methods, improving the average simulation task success rate by \\(16.4\\%\\) (Adroit), \\(14\\%\\) (DexArt), and \\(6.45\\%\\) (RLBench), and the average real-world robotic task success rate by \\(8.6\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.714, + 0.245, + 0.728 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.738, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Visual Imitation learning has emerged as an effective method for training robots to perform complex tasks, such as object grasping [18, 35, 38], legged locomotion [28, 45], dexterous manipulation [12, 29], and mobile manipulation [6, 34]. Recent advances in computer vision have further enhanced visual imitation learning, enabling impress-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.337, + 0.87, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.449, + 0.871, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.624, + 0.874, + 0.739 + ], + "angle": 0, + "content": "Figure 1. Spatial-temporal awareness in the 4D Diffusion Policy (DP4). Previous methods train perception and decision-making with trajectory supervision, but trajectory cloning fails to capture the 3D spatial and 4D spatiotemporal relationships. In contrast, DP4 constructs the current 3D scene with 3D spatial supervision from a single RGB-D view and predicts future 3D scene candidates using 4D spatiotemporal supervision, optimizing trajectory generation by effectively capturing both 3D structures and 4D dependencies." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.872, + 0.853 + ], + "angle": 0, + "content": "sive performance across various applications [3, 13, 33, 35]. Progress in computer vision has enabled sophisticated visual imitation learning methods that demonstrate excellent performance across various applications [3, 13, 33, 35]. However, real-world robotic deployment demands robust generalization to unseen scenarios, necessitating a deeper understanding of the spatiotemporal physical rules that govern dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.854, + 0.872, + 0.868 + ], + "angle": 0, + "content": "Despite their success, existing visual imitation learning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.142, + 0.833, + 0.264, + 0.844 + ], + "angle": 0, + "content": "†Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.126, + 0.845, + 0.483, + 0.868 + ], + "angle": 0, + "content": "\\(^{\\ddagger}\\)Prof. Yanwei Fu is also with Institute of Trustworthy Embodied AI, and the School of Data Science, Fudan University." + }, + { + "type": "list", + "bbox": [ + 0.126, + 0.833, + 0.483, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7122" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.125, + 0.483, + 0.222 + ], + "angle": 0, + "content": "methods struggle to capture the 3D structure of environments and 4D spatiotemporal dependencies of interactive tasks. These limitations arise mainly because current approaches rely on behavior cloning from static, supervised trajectories. Such methods often fail to infer the essential physical properties needed for spatial reasoning, object interaction, and environment-aware decision-making." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.223, + 0.483, + 0.361 + ], + "angle": 0, + "content": "In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that integrates spatial-temporal awareness into diffusion-based policies. Unlike prior works that passively mimic demonstrations exclusively from supervised trajectories, DP4 actively learns from a dynamic Gaussian world model [19, 41], which explicitly encodes 3D spatial structures and 4D spatiotemporal variations as part of the imitation learning process. Thus, as illustrated in Fig. 1, DP4 extends conventional imitation learning pipelines in the following ways:" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "- 3D spatial perception is learned via Gaussian Splitting (3DGS) [19], ensuring accurate reconstruction of the current scene's geometry from point cloud inputs." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.403, + 0.483, + 0.459 + ], + "angle": 0, + "content": "- 4D spatiotemporal awareness is enforced by future 3DGS reconstruction, where the system learns to predict and adaptively refine future environmental states based on the agent's planned trajectories." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.459, + 0.483, + 0.514 + ], + "angle": 0, + "content": "- Trajectory generation is conditioned on these structured spatial-temporal representations, allowing the diffusion policy model to reason over future states rather than just replicating past behaviors." + }, + { + "type": "list", + "bbox": [ + 0.125, + 0.362, + 0.483, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.515, + 0.483, + 0.584 + ], + "angle": 0, + "content": "So our method first converts RGB-D observations into 3D point clouds, from which we extract global and local spatiotemporal representations. These serve as conditions for the diffusion policy model to guide next-step trajectory generation. To strengthen structured perception:" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.587, + 0.483, + 0.628 + ], + "angle": 0, + "content": "- We present a novel Gaussian regressor that learns to reconstruct the 3D scene from point cloud inputs, ensuring spatially consistent representations." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.629, + 0.483, + 0.683 + ], + "angle": 0, + "content": "- We enforce temporal consistency by deforming the current scene's 3DGS into the predicted future state, supervising this process using ground-truth RGB-D reconstructions from the future scene." + }, + { + "type": "list", + "bbox": [ + 0.125, + 0.587, + 0.483, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.687, + 0.483, + 0.798 + ], + "angle": 0, + "content": "We evaluate DP4 extensively across 17 simulated tasks (173 variants) from four benchmarks, along with three real-world tasks, including the challenging task of dexterous manipulation of deformable objects. Our experiments show that DP4 significantly outperforms state-of-the-art multi-task visual imitation methods, achieving a \\(+16.4\\%\\) success rate on Adroit, \\(+14\\%\\) on DexArt, \\(+6.45\\%\\) on RL-Bench, and \\(+8.6\\%\\) improvement in real-world robotic tasks." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.799, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Our work bridges the gap between visual imitation learning and spatial-temporal physical reasoning, demonstrating the importance of structured world models in improving generalization and robustness in real-world robotic applications, with these contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.126, + 0.871, + 0.194 + ], + "angle": 0, + "content": "- Spatiotemporal-centric visual imitation learning: Unlike existing methods that overlook 3D perception and dynamic interactions, DP4 explicitly models 3D spatial structures and 4D temporal dynamics, improving generalization in dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.195, + 0.871, + 0.25 + ], + "angle": 0, + "content": "- 4D Diffusion Policy with structured awareness: We introduce a diffusion-based visual imitation learning framework that generates trajectories based on learned spatial-temporal world representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.251, + 0.871, + 0.305 + ], + "angle": 0, + "content": "- Dynamic Gaussian World Model for structured supervision: Our model actively learns from real-world interactions, ensuring the embedding of 3D spatial and 4D temporal reasoning into policy learning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.306, + 0.871, + 0.36 + ], + "angle": 0, + "content": "- State-of-the-art performance across diverse tasks: DP4 sets unprecedented success rates across 17 simulated and real-world robotic tasks, setting a new benchmark for imitation learning in dynamic environments." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.126, + 0.871, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.374, + 0.642, + 0.388 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.397, + 0.872, + 0.604 + ], + "angle": 0, + "content": "Visual imitation learning [20-22] is an effective method for equipping robots with skills, typically relying on numerous expert demonstrations as observation-action pairs. Recent advanced visual imitation learning methods commonly employ 2D images[1, 5, 7, 26, 33, 38] as state representations for their rich global information and ease of acquisition from raw sensory inputs, whereas 3D representations (e.g., point clouds and object geometries) [47, 48] are used for their explicit spatial structure, enhancing the spatial reasoning. However, these methods overlook the spatiotemporal dynamics of scenes arising from behavior-environment interactions. DP4 overcomes this limitation by employing a dynamic Gaussian world model to derive scene-level 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.605, + 0.872, + 0.827 + ], + "angle": 0, + "content": "Diffusion models as policies Building on their success in image and video generation [4, 14, 44], diffusion models have emerged as effective policy frameworks in robotics [5, 27, 29, 43, 47]. They are primarily applied in reinforcement learning and visual imitation learning, with Diffusion Policy [5] and variants [40, 47] achieving state-of-the-art results by fusing visual inputs (e.g., 2D images, 3D point clouds) with complex, multi-modal action spaces. Nevertheless, these approaches often underutilize visual perception, relying heavily on behavior cloning from supervised trajectories. In contrast, DP4 introduces a Gaussian world model that reconstructs current scenes from RGB-D data via volumetric rendering for 3D spatial understanding, and employs a deformation Gaussian field to predict future scene distributions through environment interaction, enabling comprehensive 4D spatiotemporal perception." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.827, + 0.871, + 0.868 + ], + "angle": 0, + "content": "World model Recent advances demonstrate that world models effectively encode scene dynamics by predicting future states from current states and actions, with appli" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.515, + 0.927 + ], + "angle": 0, + "content": "7123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.125, + 0.831, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.447, + 0.875, + 0.55 + ], + "angle": 0, + "content": "Figure 2. The framework of our 4D Diffusion Policy (DP4). From a single-view RGB-D observation, we construct 3D point clouds and extract global and local features to enrich both holistic and focused perceptions. These multi-level representations condition the diffusion policy model to generate trajectories based on current robot states. We introduce a Gaussian world model in DP4 to capture 3D structures and 4D spatiotemporal relationships. The current observation's 3D Gaussian Splitting (3DGS) is derived via a generalizable Gaussian regressor from point clouds and multi-level features. By enforcing consistency between ground-truth and rendered RGB-D images from this 3DGS, we enhance 3D spatial awareness. Additionally, future 3DGS are predicted from current states using policy-generated trajectories, with rendered RGB-D consistency fostering 4D spatiotemporal awareness. This improved spatiotemporal representation significantly benefits complex tasks such as object grasping and dexterous manipulation." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.57, + 0.484, + 0.736 + ], + "angle": 0, + "content": "cations in autonomous driving [15, 16, 39] and robotic manipulation [24, 32, 42]. Early works [8-11, 15, 23] employed latent space autoencoding for future prediction, yielding strong results in both simulated and real environments. Nonetheless, learning accurate latent representations is data-intensive and generally confined to simple tasks due to limited implicit feature capacity. To address these challenges, this study integrates Gaussian Splating [19], known for its efficiency and rendering quality, into world model construction. The resulting 4D Diffusion Policy leverages a dynamic Gaussian world model to capture scene-level 3D structure and 4D spatiotemporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.748, + 0.248, + 0.763 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.771, + 0.484, + 0.868 + ], + "angle": 0, + "content": "In this section, we introduce 4D Diffusion Policy (DP4), a novel visual imitation learning method that emphasizes spatial-temporal awareness in diffusion policies. DP4 employs 3D Gaussian Splatting for informative supervision, enabling the learning of 3D spatial and 4D spatiotemporal perceptions in interactive environments. The DP4 framework is illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.569, + 0.652, + 0.584 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.59, + 0.872, + 0.743 + ], + "angle": 0, + "content": "Visual imitation learning allows robots to efficiently develop human-like skills using observation-action pairs from expert demonstrations. The intelligent agent must interactively predict subsequent continuous trajectories based on the current observation. The visual observation at time step \\( t \\), denoted \\( \\mathbf{O}^{(t)} = (\\mathbf{C}^{(t)},\\mathbf{D}^{(t)}) \\), is derived from expert demonstrations, where \\( \\mathbf{C}^{(t)} \\) and \\( \\mathbf{D}^{(t)} \\) represent single-view RGB and depth images, respectively. Using the visual observation \\( \\mathbf{O}^{(t)} \\), the visuomotor policy learning model predicts the optimal action \\( \\mathbf{a}^{(t)} \\) at current time step and generates subsequent optimal continuous action sequence \\( \\mathbf{a}^{(t + 1)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.873, + 0.869 + ], + "angle": 0, + "content": "To effectively train a visuomotor policy learning model, expert demonstrations are supplied as offline datasets for visual imitation learning, consisting of sample triplets with visual observations and expert actions [26, 40]. Current methods utilize robust visual representations, such as Diffusion Policy and 3D Diffusion Policy, to derive informative latent features for optimal action prediction. However, these methods heavily depend on behavior cloning with supervised historical trajectories. This approach struggles to fully" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.125, + 0.484, + 0.236 + ], + "angle": 0, + "content": "capture 3D structures and often overlooks 4D spatiotemporal dynamics. Consequently, the predicted actions frequently fail to achieve complex human goals, particularly when accurate object interactions are absent. In this study, we introduce the 4D Diffusion Policy, which employs a dynamic Gaussian world model to extract 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.247, + 0.391, + 0.262 + ], + "angle": 0, + "content": "3.2. Multi-level 3D Spatial Awareness" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.268, + 0.484, + 0.447 + ], + "angle": 0, + "content": "To capture complex 3D structures in the interactive physical world, DP4 achieves multi-level spatial perception by generating 3D representations from local point clouds and global voxel grids. These 3D representations are derived from an RGB-D image captured by a single-view camera with known extrinsic and intrinsic parameters, enhancing practical applicability in real-world settings. DP4 avoids impractical setups, such as using multiple cameras during training and inference or requiring extensive prior environmental visual data. In both simulated and real-world settings, we acquire depth images from a single camera and convert them into complete point clouds using the camera's extrinsic and intrinsic parameters." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.449, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Learning 3D representations To generate the 3D local representation, DP4 crops the point clouds, retaining only points within a specified bounding box. This allows DP4 to focus on critical scene elements (e.g., the robot arm) and exclude redundant points (e.g., the ground). DP4 encodes the cropped point clouds into compact 3D local representations using a lightweight MLP network, called the 3D Local Encoder. To obtain the 3D global representation, DP4 converts the point clouds at step \\( t \\) into a \\( 100^3 \\) voxel. The 3D Global Encoder encodes this voxel, outputting the 3D global representation \\( v^{(t)} \\in \\mathbb{R}^{100^3 \\times 64} \\)." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.603, + 0.484, + 0.868 + ], + "angle": 0, + "content": "Enhance 3D spatial awareness To enhance the representation \\(v^{(t)}\\) with global structure and texture information, we employ a Generalizable Gaussian Regressor, which takes the deep volume \\(v\\) as the scene representation to derive the Gaussian world model based on Gaussian Splatting [19]. The Generalizable Gaussian Regressor is learned by rendering RGB and depth images from the constructed Gaussian world model. The entire neural rendering process is described below. Define \\(v^{(t)}\\mathbf{x}\\in \\mathbb{R}^{128}\\) as the sampled 3D feature at point \\(\\mathbf{x}\\) from the 3D global representation \\(v^{(t)}\\). \\(v^{(t)}\\mathbf{x}\\) is obtained through trilinear interpolation, given the discretized nature of the 3D global representation \\(v^{(t)}\\). The Generalizable Gaussian Regressor uses the 3D coordinates \\(\\mathbf{x}\\) and the feature \\(v_{\\mathbf{x}}^{(t)}\\) to parameterize the Gaussian primitive as \\(\\theta^{(t)} = (\\mu^{(t)},c^{(t)},r^{(t)},s^{(t)},\\sigma^{(t)})\\), representing the position, color, rotation, scale, and opacity of the Gaussian primitive. To render a single view, we project Gaussian primitives onto the 2D plane using differential tile-based rasterization. The pixel value \\(\\mathbf{p}\\) is rendered using alpha" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.126, + 0.616, + 0.139 + ], + "angle": 0, + "content": "blend rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.146, + 0.872, + 0.193 + ], + "angle": 0, + "content": "\\[\nC ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} c _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.19, + 0.808, + 0.228 + ], + "angle": 0, + "content": "\\[\nD ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} d _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.235, + 0.872, + 0.392 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}^{(t)} = \\sigma_{i}^{(t)}\\exp -\\frac{1}{2} (\\mathbf{p} - \\mu_{i}^{(t)})^{\\top}\\Sigma_{i}^{-1}(\\mathbf{p} - \\mu_{i}^{(t)})\\) defines the 2D density of Gaussian points during the splatting process, \\(C^{(t)}\\) and \\(D^{(t)}\\) represent the rendered RGB image and depth, respectively, \\(N\\) is the number of Gaussians per tile, \\(d_{i}^{(t)}\\) is the \\(z\\)-value obtained by transforming coordinate \\(u\\) from world to camera coordinates, and \\(\\Sigma_{i}\\) denotes the covariance matrix derived from the Gaussian parameters' rotation and scale. Our Generalizable Gaussian Regressor and 3D representation are optimized to reconstruct RGB and depth images from a single view by minimizing a 3D loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.397, + 0.872, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {3 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t)} \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.435, + 0.872, + 0.492 + ], + "angle": 0, + "content": "where \\(\\mathbf{C}^{*(t)}\\) and \\(\\mathbf{D}^{*(t)}\\) represent the ground truth color and depth at time step \\(t\\). This 3D loss allows the 3D representation to emphasize local visual features while incorporating global structural knowledge and texture information." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.499, + 0.757, + 0.513 + ], + "angle": 0, + "content": "3.3. 4D Spatiotemporal Awareness" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.519, + 0.872, + 0.671 + ], + "angle": 0, + "content": "To emphasize 4D spatiotemporal awareness in the diffusion policy, DP4 integrates dynamics into the Gaussian world model described in Section 3.2. This model parameterizes the Gaussian mixture distribution for dynamic Gaussian splatting, enabling future scene reconstruction via parameter propagation. Thus, the model gains valuable supervision in interactive environments by ensuring consistency between reconstructed and real feature scenes. Specifically, the model learns environmental dynamics by predicting the future state \\( s^{(t + 1)} \\) from the current state \\( s^{(t)} \\) and action \\( a^{(t)} \\) at time step \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.671, + 0.872, + 0.796 + ], + "angle": 0, + "content": "In the dynamic Gaussian world model, we define the current state as the visual observation at the current time step, with actions representing the robot components. These observations and actions predict visual scenes for the next time step, representing the future state. Building on the Gaussian world model, DP4 uses a deformable MLP \\( p_{\\phi} \\) to estimate changes in Gaussian parameters during time propagation and a Gaussian renderer \\( \\mathcal{R} \\) to generate RGB and depth images for the future state:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.803, + 0.719, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta^ {(t)} = p _ {\\phi} \\left(\\theta^ {(t)}, a ^ {(t)}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.829, + 0.872, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\theta^ {(t + 1)} = \\theta^ {(t)} + \\Delta \\theta^ {(t)}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.848, + 0.817, + 0.871 + ], + "angle": 0, + "content": "\\[\nC ^ {(t + 1)} (\\mathbf {p}), D ^ {(t + 1)} (\\mathbf {p}) = \\mathcal {R} \\left(\\theta^ {(t + 1)}, \\mathbf {p}\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.515, + 0.927 + ], + "angle": 0, + "content": "7125" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.124, + 0.487, + 0.388 + ], + "angle": 0, + "content": "where \\(\\mathbf{C}^{(t + 1)}(\\mathbf{p})\\) and \\(\\mathbf{D}^{(t + 1)}(\\mathbf{p})\\) represent the predicted RGB and depth images at time step \\(t + 1\\), \\(\\theta^{(t)}\\) represents the 3D Gaussian primitive parameters at time step \\(t\\), \\(w\\) denotes the camera pose used to project the Gaussian primitives, and \\(\\mathcal{R}\\) is the rendering process from Equation 1. DP4 employs multi-head neural networks as the Gaussian regressor, with each head predicting distinct features of the Gaussian parameters. By estimating changes in position and rotation between time steps, we derive the Gaussian parameters for the future step. The Gaussian renderer then projects the propagated Gaussian distribution onto the single view to reconstruct the future scene. To model 4D spatiotemporal relationships in the physical world, we enforce consistency supervision between reconstructed and real scenes at time step \\(t + 1\\), embedding scene-level spatiotemporal dynamics into the Gaussian parameters. Specifically, the training objective aligns predicted future scenes, derived from current observations and actions, with real scenes, as formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.394, + 0.485, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {4 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t + 1)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t + 1)} \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.433, + 0.485, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathbf{C}^{(t + 1)}\\) and \\(\\mathbf{D}^{(t + 1)}\\) represent the real future counterparts. This additional task of predicting future scenes from current observations via the dynamic Gaussian world model serves as a spatiotemporal enhancement penalty. Training with this task, the multi-level 3D representation learns to encode physical properties and develop awareness of 4D spatiotemporal dynamics." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.539, + 0.334, + 0.552 + ], + "angle": 0, + "content": "3.4. Diffusion-based Decision" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.559, + 0.486, + 0.684 + ], + "angle": 0, + "content": "The multi-level 3D representation, including local and global 3D representations, is optimized to predict the robot's optimal actions. We model this process as a conditional denoising diffusion framework, conditioned on the multi-level 3D representation \\( r \\) and the robot state \\( q \\), which converts randomly sampled Gaussian noise into actions \\( a \\). Starting with sampled Gaussian noise \\( a^K \\), the denoising network \\( \\epsilon_{\\theta} \\) performs \\( K \\) iterations to progressively convert it into the noise-free action \\( a^0 \\)." + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.693, + 0.484, + 0.712 + ], + "angle": 0, + "content": "\\[\na ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, \\mathrm {I}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.72, + 0.485, + 0.805 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(0,\\mathbf{I})\\) represents the sampled Gaussian noise, and \\(\\alpha_{k},\\gamma_{k}\\), and \\(\\sigma_{k}\\) are functions of \\(k\\) defined by the noise scheduler. To train the denoising network \\(\\epsilon_{\\theta}\\), we randomly sample a data point \\(a^0\\) from the dataset and use a diffusion process to generate the noise \\(\\epsilon^k\\) at iteration \\(k\\). The training objective is to predict the noise added to the data." + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.813, + 0.484, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a c t i o n}} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha} _ {k} a ^ {0} + \\bar {\\beta} _ {k} \\epsilon^ {k}, k, r, q\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.839, + 0.485, + 0.869 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_k\\) and \\(\\bar{\\beta}_k\\) are noise schedule that performs one step noise adding." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.126, + 0.848, + 0.139 + ], + "angle": 0, + "content": "The overall learning objective for DP4 is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.15, + 0.872, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D P 4}} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {3 D}} \\mathcal {L} _ {\\mathrm {3 D}} + \\lambda_ {\\mathrm {4 D}} \\mathcal {L} _ {\\mathrm {4 D}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.175, + 0.873, + 0.356 + ], + "angle": 0, + "content": "where \\(\\lambda_{3\\mathrm{D}}\\) and \\(\\lambda_{4\\mathrm{D}}\\) balance the 3D loss for current scenes and the 4D loss for future scenes across different objectives. We train DP4 using a joint training strategy. Empirical observations show that this strategy enhances information fusion while learning shared features. During training, we introduce a warm-up phase, freezing the deformable mapper for the first 500 iterations to establish a stable multi-level 3D compact representation and Gaussian regressor. After the warm-up phase, we simultaneously train the entire Gaussian world model and the diffusion-based decision module. After optimization, DP4 generates optimal actions with spatiotemporal awareness, including both 3D spatial and 4D spatiotemporal perceptions." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.368, + 0.634, + 0.383 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.391, + 0.674, + 0.406 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.411, + 0.873, + 0.536 + ], + "angle": 0, + "content": "Simulations For simulation experiments, we gathered 22 tasks across various domains, covering a wide range of robotic skills. These tasks span challenging scenarios, such as bimanual, deformable object, and articulated object manipulation, as well as simpler tasks like parallel gripper manipulation. The tasks include Adroit tasks [30] (Door, Pen, Hammer), DexArt tasks [2] (Laptop, Faucet, Bucket, Toilet), and RLBench tasks [17] (10 challenging tasks, each with at least two variations, totaling 166)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.536, + 0.873, + 0.687 + ], + "angle": 0, + "content": "Real robot We used a KINOVA GEN2 robot with a Realsense D455 depth camera mounted in an eye-to-hand configuration. In an indoor environment, we arranged several bottles and cups to enable the agent to generalize manipulation skills across scenes, designing three tasks, including Grasping bottles, Pouring the water from the bottle into the cup and stacking cups. The RealSense D455 camera records the entire scene and robot state from a third-person perspective. This setup supplies RGB-D observations for policy training and RGB and depth supervision for constructing the dynamic Gaussian world model." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.688, + 0.872, + 0.785 + ], + "angle": 0, + "content": "Expert demonstration Expert trajectories for ManiSkill and Adroit are collected using RL-trained agents, using VRL3 [37] for Adroit and PPO [31] for other domains. We generate successful trajectories with RL agents and ensure that all visual imitation learning algorithms use identical demonstrations. Additionally, we collect 20 demonstrations for each RLBench task using the motion planner." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.785, + 0.873, + 0.868 + ], + "angle": 0, + "content": "Baselines The main objective of DP4 is to highlight the importance of 3D spatial and 4D spatiotemporal awareness in diffusion policies. To this end, our primary baseline is the point cloud-based 3D Diffusion Policy (DP3) [47]. Additionally, we incorporate comparisons with other visual imitation learning methods: IBC [7], BCRNN [26], Diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7126" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.125, + 0.483, + 0.18 + ], + "angle": 0, + "content": "Policy (DP) [5], PerAct [35], GNFactor [46], and Mani-Gaussian [25]. Note that image and depth resolutions for all 2D and 3D methods remain consistent across experiments, ensuring fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.182, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Evaluation metric Each experiment is conducted using three seeds. For Adroit and DexArt simulation tasks, we assess 20 episodes every 200 epochs and calculate the mean of the top five success rates. For RLBench tasks, we evaluate 25 episodes per task at the final checkpoint for 10 challenging tasks. We report the mean and standard deviation of success rates across the three seeds." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.28, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Implementation details We implement a convolutional network-based diffusion policy using DDIM[36] as the noise scheduler, preferring sample prediction to enhance high-dimensional action generation. Training employs 100 timesteps, while inference uses 10. Models are trained for 3,000 epochs with a batch size of 32 on both simulated and real-world tasks. Loss hyperparameters \\(\\lambda_{\\mathrm{3D}}\\) and \\(\\lambda_{\\mathrm{4D}}\\) are set to 0.1 and 0.01, respectively (see Section 4.4 for details). For fair comparison with DP3[47], we adopt the same 3D Local Encoder: a three-layer MLP, max-pooling for order-equivariant aggregation, and a projection head generating compact 3D local features, with LayerNorm inserted for training stability. All experiments run on an NVIDIA H100 80GB GPU and a 192-vCPU Intel Xeon Platinum 8468." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.484, + 0.338, + 0.499 + ], + "angle": 0, + "content": "4.2. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.506, + 0.483, + 0.533 + ], + "angle": 0, + "content": "We compare our DP4 with previous state-of-the-art methods on the Adroit, DexArt, and RLBench simulation tasks." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.534, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Adroit simulation Table 1 presents the success rates for the Adroit simulation tasks. DP4 achieves an overall success rate of \\(84.7\\%\\), significantly outperforming 2D-based methods, including IBC [7], BCRNN [26], and Diffusion Policy [5], while surpassing 3D-based methods such as 3D Diffusion Policy [47]. Experimental results demonstrate that DP4 completes complex tasks more accurately by leveraging spatiotemporal perception." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.646, + 0.483, + 0.77 + ], + "angle": 0, + "content": "DexArt simulation Table 2 presents the success rates for various DexArt simulation tasks. The DP4 achieves a success rate of \\(82.5\\%\\), significantly outperforming 2D-based methods (e.g., IBC, BCRNN, and Diffusion Policy) by \\(33.5\\%\\) and 3D-based methods (e.g., 3D Diffusion Policy) by \\(14.0\\%\\). Experimental results demonstrate that DP4 integrates 3D spatial and 4D spatiotemporal perception of the physical world with diffusion policies, achieving higher success rates across multiple tasks." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.771, + 0.483, + 0.868 + ], + "angle": 0, + "content": "RLBench simulation We compare DP4 with previous state-of-the-art methods on the RLBench task suite. Table 3 presents a comparison of average success rates across tasks. DP4 achieves state-of-the-art performance with average success rates of \\(63.3\\%\\) and \\(39.9\\%\\), significantly outperforming prior methods, including both perception-based and generative approaches. Leveraging supervision from" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.123, + 0.871, + 0.214 + ], + "angle": 0, + "content": "
MethodHammerDoorPenOverall
IBC [7]0±00±09±23.0
BCRNN [26]0±00±09±33.0
DP [5]48±1750±525±441.0
DP3 [47]100±062±443±668.3
DP4 (ours)100±080±275±384.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.217, + 0.871, + 0.268 + ], + "angle": 0, + "content": "Table 1. Success rates \\((\\%)\\) on the Adroit simulation. We compare our DP4 with more baselines in the Adroit simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.278, + 0.869, + 0.367 + ], + "angle": 0, + "content": "
MethodLaptopFaucetBucketToiletOverall
IBC [7]3±27±114±10±06.0
BCRNN [26]3±31±05±50±02.3
DP [5]69±423±858±246±149.0
DP3 [47]83±163±282±446±268.5
DP4 (ours)92±284±190±164±282.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.371, + 0.871, + 0.422 + ], + "angle": 0, + "content": "Table 2. Success rates \\((\\%)\\) on the DexArt simulation. We compare our DP4 with more baselines in the DexArt simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.427, + 0.876, + 0.566 + ], + "angle": 0, + "content": "
Methodclose jaropen drawersweep to dustpanmeat off grillturn tapOverall
PerAct [35]18.7±8.254.7±18.60.0±0.040.0±17.038.7±6.830.4
GNFactor [46]25.3±6.876.0±5.728.0±15.057.3±18.950.7±8.247.5
ManiGaussian [25]28.4±5.476.3±6.264.5±13.460.2±18.256.2±6.857.1
DP4 (ours)35.4±2.282.2±3.472.0±5.264.4±6.862.4±2.863.3
Methodslide blockput in drawerdrag stickpush buttonsstack blocksOverall
PerAct [35]18.7±13.62.7±3.35.3±5.018.7±12.46.7±1.910.4
GNFactor [46]20.0±15.00.0±0.037.3±13.218.7±10.04.0±3.316.0
ManiGaussian [25]24.3±12.816.6±3.292.3±11.420.4±12.212.4±3.233.2
DP4 (ours)33.4±4.422.8±1.294.4±6.228.3±6.821.0±2.239.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.568, + 0.871, + 0.618 + ], + "angle": 0, + "content": "Table 3. Success rates \\((\\%)\\) on the RLBench simulation. We compare our DP4 with more baselines in the RLBench simulation. Baseline approaches include PerAct, GNFactor, and Mani-Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.871, + 0.67 + ], + "angle": 0, + "content": "3D geometry, texture, and 4D spatiotemporal data, DP4 accurately predicts optimal continuous trajectories from a single RGB-D observation of the current scene." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.68, + 0.715, + 0.695 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.701, + 0.871, + 0.826 + ], + "angle": 0, + "content": "Visualization of trajectories We present three qualitative examples of action sequences generated by DP3 and our DP4 in Figure 3. The results show that the previous agent (DP3) frequently fails to complete tasks, as it imitates the expert's motions (e.g., hammering nails, sliding doors, and spinning pens) without fully understanding the scene. In contrast, DP4 significantly improves the success rate by accurately interpreting 3D scene structures (e.g., nail positions) and 4D dynamics (e.g., pen rotation)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.827, + 0.871, + 0.868 + ], + "angle": 0, + "content": "Visualization of Gaussian world model Figure 4 and Figure 5 show the visualization results. We visualize the Gaussian world model via single-view rendering, observ" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.515, + 0.927 + ], + "angle": 0, + "content": "7127" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.124, + 0.481, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.25, + 0.48, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.379, + 0.48, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.509, + 0.486, + 0.587 + ], + "angle": 0, + "content": "Figure 3. Qualitative case study. The red mark indicates a pose that significantly deviates from the expert demonstration, while the green mark denotes a pose that aligns with the expert trajectory. The 4D Diffusion Policy (DP4) integrates 3D spatial and 4D spatiotemporal awareness with diffusion policies, successfully completing the tasks." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.599, + 0.484, + 0.697 + ], + "angle": 0, + "content": "ing it captures fundamental spatial structures and spatiotemporal dynamics. Although DP4 operates under a realistic single-view constraint, resulting renderings lack fine detail, they effectively represent scene-level 3D structures and 4D dynamics. Crucially, such details have minimal impact on task success; for example, the nail's thread is irrelevant to the hammering task." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.705, + 0.265, + 0.72 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.726, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Component ablation. Ablations are conducted on the Adroit simulation tasks in Table 4. Without the proposed components, we use 3D representation without any 3D spatial supervision (RGB and depth (structure)) and 4D spatiotemporal supervision (dynamics). In the experiment on the Adroit simulation, the performance of the \"Hammer\", \"Door\", and \"Pen\" tasks demonstrated improvements in success rates with the proposed components, while the inference time remained stable or only slightly increased." + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.854, + 0.484, + 0.867 + ], + "angle": 0, + "content": "- For the Hammer task: The base model started with a" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.126, + 0.869, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.307, + 0.872, + 0.359 + ], + "angle": 0, + "content": "Figure 4. Visualization of Gaussian world model. After training DP4, we visualize its Gaussian world model by rendering a single-view observation. DP4 uses a Gaussian world model to provide supervision, facilitating the learning of 3D spatial perception." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.375, + 0.872, + 0.499 + ], + "angle": 0, + "content": "94.0% success rate and a completion time of 6.40 seconds. After adding the Gaussian world model with RGB supervision, success improved to \\(96.0\\%\\), with a minimal increase in time (6.52 seconds). Introducing 3D spatial loss (RGB and Depth supervision) raised success to \\(95.0\\%\\), with stable time (6.53 seconds). With 4D spatiotemporal supervision (DP4), success reached \\(98.0\\%\\), and the time increased slightly to 6.56 seconds, indicating a clear performance gain with a small time increase." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.872, + 0.638 + ], + "angle": 0, + "content": "- For the Door task: Initially, the model achieved a \\(94.0\\%\\) success rate and a completion time of 6.40 seconds. Adding the Gaussian world model with RGB supervision improved success to \\(96.0\\%\\), with no significant time change. Incorporating 3D spatial loss (RGB and Depth) increased success to \\(98.0\\%\\), with a slight increase in time (6.56 seconds). With 4D spatiotemporal supervision (DP4), the success rate reached \\(100.0\\%\\), with a minimal time increase (6.57 seconds), reflecting significant improvement from additional supervision." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.638, + 0.872, + 0.763 + ], + "angle": 0, + "content": "- For the Pen task: The base model had a \\(45.0\\%\\) success rate and a completion time of 6.46 seconds. After adding the Gaussian world model with RGB supervision, success rose to \\(48.0\\%\\), with time remaining at 6.48 seconds. Introducing 3D spatial loss (RGB and depth) led to a slight decrease in success \\((47.0\\%)\\). With 4D spatiotemporal supervision (DP4), success jumped to \\(75.0\\%\\), with a small increase in time to 6.50 seconds, highlighting the effectiveness of 4D supervision." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.5, + 0.872, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.769, + 0.872, + 0.852 + ], + "angle": 0, + "content": "Notably, DP4 does not substantially increase model inference time. This is because DP4 does not construct a Gaussian world model during inference. The 3D spatial and 4D spatiotemporal awareness from the Gaussian world model are efficiently integrated into the multi-level 3D representation during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.872, + 0.868 + ], + "angle": 0, + "content": "Hyperparameter analysis We conduct hyperparameter" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7128" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.121, + 0.276, + 0.136 + ], + "angle": 0, + "content": "Ground-truth Prediction at current time step T at current time step T" + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.137, + 0.198, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.123, + 0.271, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.137, + 0.339, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.124, + 0.409, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.137, + 0.478, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.194, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Figure 5. More visualizations of the learned world model. DP4 uses a Gaussian world model to provide supervision, facilitating the 4D spatiotemporal perceptions in interactive environments." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.239, + 0.388, + 0.251 + ], + "angle": 0, + "content": "Results on Adroit Simulation" + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.253, + 0.48, + 0.367 + ], + "angle": 0, + "content": "
3D Component4D ComponentPerformance
RGBDepthDynamicsHammerDoorPen
94.0/6.4064.0/6.4445.0/6.46
✓96.0/6.5268.0/6.4548.0/6.48
✓95.0/6.5368.0/6.4647.0/6.48
✓✓98.0/6.5675.0/6.5172.0/6.47
✓✓✓100.0/6.5780.0/6.4575.0/6.50
" + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.37, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Table 4. Component ablation studies. We report results for the Adroit simulation tasks. In the performance metrics, the number to the left of / represents the success rate \\((\\%)\\), and the number to the right denotes the inference speed (seconds per task)." + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.422, + 0.306, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.422, + 0.482, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.535, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Figure 6. Hyperparameter analysis of DP4. The effects of \\(\\lambda_{3\\mathrm{D}}\\) and \\(\\lambda_{4\\mathrm{D}}\\) on success rates, respectively." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.564, + 0.483, + 0.642 + ], + "angle": 0, + "content": "
MethodGrasping BottlesStacking CupsPouring WaterOverall
DP [5]36.044.028.036.0
DP3 [47]42.062.034.046.0
DP4 (ours)48.072.044.054.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.645, + 0.484, + 0.684 + ], + "angle": 0, + "content": "Table 5. Success rates \\((\\%)\\) on the real-robot tasks. We compare our DP4 with additional baselines on the real-robot tasks. We include the Diffusion Policy (DP) and 3D Diffusion Policy (DP3)." + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.686, + 0.485, + 0.812 + ], + "angle": 0, + "content": "analysis on the Pen task within the Adroit simulation. DP4 uses two key hyperparameters, \\(\\lambda_{3\\mathrm{D}}\\) and \\(\\lambda_{4\\mathrm{D}}\\), to balance the 3D spatial loss for current scenes and the 4D spatiotemporal loss for future scenes across multiple objectives. As shown in Figure 6, we analyze the hyperparameters of DP4. We vary \\(\\lambda_{3\\mathrm{D}}\\) from 0 to 1 and \\(\\lambda_{4\\mathrm{D}}\\) from 0 to 0.1, measuring the corresponding success rates. We find that the success rate peaks when \\(\\lambda_{3\\mathrm{D}} = 0.1\\) and \\(\\lambda_{4\\mathrm{D}} = 0.01\\), so we adopt these values as the hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.82, + 0.304, + 0.835 + ], + "angle": 0, + "content": "4.5. Real-Robot Analysis" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.84, + 0.484, + 0.869 + ], + "angle": 0, + "content": "To assess DP4's performance on real-world robotic tasks, we designed three tasks: Grasping Bottles, Stacking Cups," + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.125, + 0.635, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.126, + 0.747, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.127, + 0.861, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.194, + 0.734, + 0.203 + ], + "angle": 0, + "content": "\"Grasping Bottles\"" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.212, + 0.634, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.212, + 0.746, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.212, + 0.859, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.28, + 0.729, + 0.29 + ], + "angle": 0, + "content": "\"Stacking Cups\"" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.297, + 0.634, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.297, + 0.746, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.297, + 0.859, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.365, + 0.729, + 0.375 + ], + "angle": 0, + "content": "\"Pouring Water\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.385, + 0.872, + 0.436 + ], + "angle": 0, + "content": "Figure 7. Visualization of DP4 performance on three real-world robotic tasks. DP4 demonstrates strong performance in real-world settings and effectively handles a variety of common tasks with a single view." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.448, + 0.872, + 0.601 + ], + "angle": 0, + "content": "and Pouring Water. We focus on the model's performance at key frames, where success is defined as when the error is below a certain threshold. The results of the real-world robotic experiments are summarized in Table 5. Experimental results show that DP4 outperforms the Diffusion Policy (DP) and 3D Diffusion Policy (DP3) baselines across all tasks. These results confirm that DP4 effectively integrates spatiotemporal awareness with diffusion policies, improving the accuracy of generated trajectories. Additionally, visualizations of DP4's performance on real-world tasks are shown in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.618, + 0.622, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.644, + 0.872, + 0.756 + ], + "angle": 0, + "content": "This paper presents 4D Diffusion Policy (DP4), which integrates spatiotemporal awareness of the physical world with diffusion-based visual imitation learning. Utilizing a single-view RGB-D input, DP4 constructs a dynamic Gaussian world model that provides rich supervision for acquiring 3D spatial and 4D spatiotemporal perceptions from interactive environments. Extensive experiments on simulated and real-robot tasks validate the effectiveness and superiority." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.773, + 0.676, + 0.79 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.798, + 0.872, + 0.869 + ], + "angle": 0, + "content": "This work was supported in part by the Science and Technology Commission of Shanghai Municipality (No.24511103100), NSFC Project (62176061), Doubao Fund, and Shanghai Technology Development and Entrepreneurship Platform for Neuromorphic and AI SoC." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.915, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7129" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.127, + 0.124, + 0.214, + 0.138 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.147, + 0.486, + 0.21 + ], + "angle": 0, + "content": "[1] Sridhar Pandian Arunachalam, Sneha Silwal, Ben Evans, and Lerrel Pinto. Dexterous imitation made easy: A learning-based framework for efficient dexterous manipulation. In 2023 IEEE international conference on robotics and automation (icra), pages 5954-5961. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.212, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[2] Chen Bao, Helin Xu, Yuzhe Qin, and Xiaolong Wang. Dexart: Benchmarking generalizable dexterous manipulation with articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21190-21200, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.277, + 0.484, + 0.339 + ], + "angle": 0, + "content": "[3] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.341, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.392, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[5] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.457, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[6] Yuqing Du, Daniel Ho, Alex Alemi, Eric Jang, and Mohi Khansari. Bayesian imitation learning for end-to-end mobile manipulation. In International Conference on Machine Learning, pages 5531-5546. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.508, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[7] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on robot learning, pages 158-168. PMLR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.572, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.134, + 0.611, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[9] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.65, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[10] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.689, + 0.484, + 0.737 + ], + "angle": 0, + "content": "[11] Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems, 35:26091-26104, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.74, + 0.484, + 0.777 + ], + "angle": 0, + "content": "[12] Siddhant Haldar, Jyothish Pari, Anant Rai, and Lerrel Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations. arXiv preprint arXiv:2303.01497, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.779, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[13] Nicklas Hansen, Zhecheng Yuan, Yanjie Ze, Tongzhou Mu, Aravind Rajeswaran, Hao Su, Huazhe Xu, and Xiaolong Wang. On pre-training for visuo-motor control: Revisiting a learning-from-scratch baseline. arXiv preprint arXiv:2212.05749, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.843, + 0.484, + 0.867 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif" + }, + { + "type": "list", + "bbox": [ + 0.127, + 0.147, + 0.486, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.127, + 0.872, + 0.152 + ], + "angle": 0, + "content": "fusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.154, + 0.872, + 0.217 + ], + "angle": 0, + "content": "[15] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems, 35:20703-20716, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.219, + 0.872, + 0.28 + ], + "angle": 0, + "content": "[16] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.283, + 0.872, + 0.333 + ], + "angle": 0, + "content": "[17] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.336, + 0.872, + 0.386 + ], + "angle": 0, + "content": "[18] Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE international conference on robotics and automation (ICRA), pages 4613-4619. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.388, + 0.872, + 0.437 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.44, + 0.872, + 0.528 + ], + "angle": 0, + "content": "[20] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.531, + 0.872, + 0.58 + ], + "angle": 0, + "content": "[21] Peiyan Li, Hongtao Wu, Yan Huang, Chilam Cheang, Liang Wang, and Tao Kong. Gr-mg: Leveraging partially-annotated data via multi-modal goal-conditioned policy. IEEE Robotics and Automation Letters, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.582, + 0.872, + 0.633 + ], + "angle": 0, + "content": "[22] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.635, + 0.872, + 0.711 + ], + "angle": 0, + "content": "[23] Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Reasongrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3718-3727, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.713, + 0.872, + 0.775 + ], + "angle": 0, + "content": "[24] Zhenyang Liu, Sixiao Zheng, Siyu Chen, Cairong Zhao, Longfei Liang, Xiangyang Xue, and Yanwei Fu. A neural representation framework with llm-driven spatial reasoning for open-vocabulary 3d visual grounding. arXiv preprint arXiv:2507.06719, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.778, + 0.872, + 0.841 + ], + "angle": 0, + "content": "[25] Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In European Conference on Computer Vision, pages 349-366. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.843, + 0.872, + 0.868 + ], + "angle": 0, + "content": "[26] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.127, + 0.872, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.916, + 0.516, + 0.927 + ], + "angle": 0, + "content": "7130" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.156, + 0.126, + 0.484, + 0.177 + ], + "angle": 0, + "content": "Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.18, + 0.484, + 0.23 + ], + "angle": 0, + "content": "[27] Liyuan Mao, Haoran Xu, Xianyuan Zhan, Weinan Zhang, and Amy Zhang. Diffusion-dice: In-sample diffusion guidance for offline reinforcement learning. arXiv preprint arXiv:2407.20109, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.233, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[28] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.285, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[29] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.352, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[30] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.417, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.458, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[32] Younggyo Seo, Danijar Hafner, Hao Liu, Fangchen Liu, Stephen James, Kimin Lee, and Pieter Abbeel. Masked world models for visual control. In Conference on Robot Learning, pages 1332-1344. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.511, + 0.484, + 0.562 + ], + "angle": 0, + "content": "[33] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning \\(k\\) modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.565, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[34] Nur Muhammad Mahi Shafullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, Soumith Chintala, and Lerrel Pinto. On bringing robots home. arXiv preprint arXiv:2311.16098, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.617, + 0.484, + 0.667 + ], + "angle": 0, + "content": "[35] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.67, + 0.484, + 0.695 + ], + "angle": 0, + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.698, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[37] Che Wang, Xufang Luo, Keith Ross, and Dongsheng Li. Vrl3: A data-driven framework for visual deep reinforcement learning. Advances in Neural Information Processing Systems, 35:32974-32988, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.751, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[38] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimi-cplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.805, + 0.484, + 0.867 + ], + "angle": 0, + "content": "[39] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.127, + 0.126, + 0.484, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.127, + 0.871, + 0.176 + ], + "angle": 0, + "content": "[40] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning. arXiv preprint arXiv:2208.06193, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.179, + 0.871, + 0.242 + ], + "angle": 0, + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.244, + 0.871, + 0.293 + ], + "angle": 0, + "content": "[42] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226-2240. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.296, + 0.871, + 0.357 + ], + "angle": 0, + "content": "[43] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.36, + 0.871, + 0.422 + ], + "angle": 0, + "content": "[44] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.424, + 0.871, + 0.475 + ], + "angle": 0, + "content": "[45] Ruihan Yang, Zhuoqun Chen, Jianhan Ma, Chongyi Zheng, Yiyu Chen, Quan Nguyen, and Xiaolong Wang. Generalized animal imitator: Agile locomotion with versatile motion prior. arXiv preprint arXiv:2310.01408, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.477, + 0.871, + 0.539 + ], + "angle": 0, + "content": "[46] Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. In Conference on Robot Learning, pages 284-301. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.541, + 0.871, + 0.604 + ], + "angle": 0, + "content": "[47] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024. 2, 5, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.606, + 0.871, + 0.668 + ], + "angle": 0, + "content": "[48] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37:77799-77830, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.127, + 0.871, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.916, + 0.514, + 0.927 + ], + "angle": 0, + "content": "7131" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_origin.pdf b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d92671bf6fb52d40d1df834e37140fb4e5c22d49 --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/635ea516-8718-4cd2-bbf5-79fa32431478_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bec8dffab1e9f8b66eeb93cd627d610744d7014d5288d1a103b89d9d7b67bb16 +size 1157078 diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/full.md b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1412cd6aec165019c4f38c57ac629621f9d99d11 --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/full.md @@ -0,0 +1,348 @@ +# Spatial-Temporal Aware Visuomotor Diffusion Policy Learning + +Zhenyang Liu $^{1,2}$ , Yikai Wang $^{3}$ , Kuanning Wang $^{1}$ , Longfei Liang $^{4}$ + +Xiangyang Xue $^{1,\dagger}$ , Yanwei Fu $^{1,2\dagger}$ + +$^{1}$ Fudan University $^{2}$ Shanghai Innovation Institute $^{3}$ Nanyang Technological University $^{4}$ NeuHelium Co., Ltd + +liuzy24@m.fudan.edu.cn, yikai.wang@ntu.edu.sg, knwang24@m.fudan.edu.cn + +longfei.liang@neuhelium.com, {xyxue,yanweifu}@fudan.edu.cn + +Project Page: ZhenyangLiu.github.io/DP4 + +# Abstract + +Visual imitation learning is effective for robots to learn versatile tasks. However, many existing methods rely on behavior cloning with supervised historical trajectories, limiting their 3D spatial and 4D spatiotemporal awareness. Consequently, these methods struggle to capture the 3D structures and 4D spatiotemporal relationships necessary for real-world deployment. In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that incorporates spatiotemporal awareness into diffusion-based policies. Unlike traditional approaches that rely on trajectory cloning, DP4 leverages a dynamic Gaussian world model to guide the learning of 3D spatial and 4D spatiotemporal perceptions from interactive environments. Our method constructs the current 3D scene from a single-view RGB-D observation and predicts the future 3D scene, optimizing trajectory generation by explicitly modeling both spatial and temporal dependencies. Extensive experiments across 17 simulation tasks with 173 variants and 3 real-world robotic tasks demonstrate that the 4D Diffusion Policy (DP4) outperforms baseline methods, improving the average simulation task success rate by $16.4\%$ (Adroit), $14\%$ (DexArt), and $6.45\%$ (RLBench), and the average real-world robotic task success rate by $8.6\%$ . + +# 1. Introduction + +Visual Imitation learning has emerged as an effective method for training robots to perform complex tasks, such as object grasping [18, 35, 38], legged locomotion [28, 45], dexterous manipulation [12, 29], and mobile manipulation [6, 34]. Recent advances in computer vision have further enhanced visual imitation learning, enabling impress- + +![](images/0f78a27e752a7b0442405a313f575cb9ed5a83b5392c3d5445cadaad9465b90a.jpg) + +![](images/f6fc01632182ef501da5f9573c1b55e20f80f26291f4e90bbbc9d441f0255193.jpg) +Figure 1. Spatial-temporal awareness in the 4D Diffusion Policy (DP4). Previous methods train perception and decision-making with trajectory supervision, but trajectory cloning fails to capture the 3D spatial and 4D spatiotemporal relationships. In contrast, DP4 constructs the current 3D scene with 3D spatial supervision from a single RGB-D view and predicts future 3D scene candidates using 4D spatiotemporal supervision, optimizing trajectory generation by effectively capturing both 3D structures and 4D dependencies. + +sive performance across various applications [3, 13, 33, 35]. Progress in computer vision has enabled sophisticated visual imitation learning methods that demonstrate excellent performance across various applications [3, 13, 33, 35]. However, real-world robotic deployment demands robust generalization to unseen scenarios, necessitating a deeper understanding of the spatiotemporal physical rules that govern dynamic environments. + +Despite their success, existing visual imitation learning + +methods struggle to capture the 3D structure of environments and 4D spatiotemporal dependencies of interactive tasks. These limitations arise mainly because current approaches rely on behavior cloning from static, supervised trajectories. Such methods often fail to infer the essential physical properties needed for spatial reasoning, object interaction, and environment-aware decision-making. + +In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that integrates spatial-temporal awareness into diffusion-based policies. Unlike prior works that passively mimic demonstrations exclusively from supervised trajectories, DP4 actively learns from a dynamic Gaussian world model [19, 41], which explicitly encodes 3D spatial structures and 4D spatiotemporal variations as part of the imitation learning process. Thus, as illustrated in Fig. 1, DP4 extends conventional imitation learning pipelines in the following ways: + +- 3D spatial perception is learned via Gaussian Splitting (3DGS) [19], ensuring accurate reconstruction of the current scene's geometry from point cloud inputs. +- 4D spatiotemporal awareness is enforced by future 3DGS reconstruction, where the system learns to predict and adaptively refine future environmental states based on the agent's planned trajectories. +- Trajectory generation is conditioned on these structured spatial-temporal representations, allowing the diffusion policy model to reason over future states rather than just replicating past behaviors. + +So our method first converts RGB-D observations into 3D point clouds, from which we extract global and local spatiotemporal representations. These serve as conditions for the diffusion policy model to guide next-step trajectory generation. To strengthen structured perception: + +- We present a novel Gaussian regressor that learns to reconstruct the 3D scene from point cloud inputs, ensuring spatially consistent representations. +- We enforce temporal consistency by deforming the current scene's 3DGS into the predicted future state, supervising this process using ground-truth RGB-D reconstructions from the future scene. + +We evaluate DP4 extensively across 17 simulated tasks (173 variants) from four benchmarks, along with three real-world tasks, including the challenging task of dexterous manipulation of deformable objects. Our experiments show that DP4 significantly outperforms state-of-the-art multi-task visual imitation methods, achieving a $+16.4\%$ success rate on Adroit, $+14\%$ on DexArt, $+6.45\%$ on RL-Bench, and $+8.6\%$ improvement in real-world robotic tasks. + +Our work bridges the gap between visual imitation learning and spatial-temporal physical reasoning, demonstrating the importance of structured world models in improving generalization and robustness in real-world robotic applications, with these contributions: + +- Spatiotemporal-centric visual imitation learning: Unlike existing methods that overlook 3D perception and dynamic interactions, DP4 explicitly models 3D spatial structures and 4D temporal dynamics, improving generalization in dynamic environments. +- 4D Diffusion Policy with structured awareness: We introduce a diffusion-based visual imitation learning framework that generates trajectories based on learned spatial-temporal world representations. +- Dynamic Gaussian World Model for structured supervision: Our model actively learns from real-world interactions, ensuring the embedding of 3D spatial and 4D temporal reasoning into policy learning. +- State-of-the-art performance across diverse tasks: DP4 sets unprecedented success rates across 17 simulated and real-world robotic tasks, setting a new benchmark for imitation learning in dynamic environments. + +# 2. Related Work + +Visual imitation learning [20-22] is an effective method for equipping robots with skills, typically relying on numerous expert demonstrations as observation-action pairs. Recent advanced visual imitation learning methods commonly employ 2D images[1, 5, 7, 26, 33, 38] as state representations for their rich global information and ease of acquisition from raw sensory inputs, whereas 3D representations (e.g., point clouds and object geometries) [47, 48] are used for their explicit spatial structure, enhancing the spatial reasoning. However, these methods overlook the spatiotemporal dynamics of scenes arising from behavior-environment interactions. DP4 overcomes this limitation by employing a dynamic Gaussian world model to derive scene-level 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction. + +Diffusion models as policies Building on their success in image and video generation [4, 14, 44], diffusion models have emerged as effective policy frameworks in robotics [5, 27, 29, 43, 47]. They are primarily applied in reinforcement learning and visual imitation learning, with Diffusion Policy [5] and variants [40, 47] achieving state-of-the-art results by fusing visual inputs (e.g., 2D images, 3D point clouds) with complex, multi-modal action spaces. Nevertheless, these approaches often underutilize visual perception, relying heavily on behavior cloning from supervised trajectories. In contrast, DP4 introduces a Gaussian world model that reconstructs current scenes from RGB-D data via volumetric rendering for 3D spatial understanding, and employs a deformation Gaussian field to predict future scene distributions through environment interaction, enabling comprehensive 4D spatiotemporal perception. + +World model Recent advances demonstrate that world models effectively encode scene dynamics by predicting future states from current states and actions, with appli + +![](images/32729a1c19491f8baf91483ddd502ecc561f35e53eb99941670633d99d84a506.jpg) +Figure 2. The framework of our 4D Diffusion Policy (DP4). From a single-view RGB-D observation, we construct 3D point clouds and extract global and local features to enrich both holistic and focused perceptions. These multi-level representations condition the diffusion policy model to generate trajectories based on current robot states. We introduce a Gaussian world model in DP4 to capture 3D structures and 4D spatiotemporal relationships. The current observation's 3D Gaussian Splitting (3DGS) is derived via a generalizable Gaussian regressor from point clouds and multi-level features. By enforcing consistency between ground-truth and rendered RGB-D images from this 3DGS, we enhance 3D spatial awareness. Additionally, future 3DGS are predicted from current states using policy-generated trajectories, with rendered RGB-D consistency fostering 4D spatiotemporal awareness. This improved spatiotemporal representation significantly benefits complex tasks such as object grasping and dexterous manipulation. + +cations in autonomous driving [15, 16, 39] and robotic manipulation [24, 32, 42]. Early works [8-11, 15, 23] employed latent space autoencoding for future prediction, yielding strong results in both simulated and real environments. Nonetheless, learning accurate latent representations is data-intensive and generally confined to simple tasks due to limited implicit feature capacity. To address these challenges, this study integrates Gaussian Splating [19], known for its efficiency and rendering quality, into world model construction. The resulting 4D Diffusion Policy leverages a dynamic Gaussian world model to capture scene-level 3D structure and 4D spatiotemporal dynamics. + +# 3. Methodology + +In this section, we introduce 4D Diffusion Policy (DP4), a novel visual imitation learning method that emphasizes spatial-temporal awareness in diffusion policies. DP4 employs 3D Gaussian Splatting for informative supervision, enabling the learning of 3D spatial and 4D spatiotemporal perceptions in interactive environments. The DP4 framework is illustrated in Figure 2. + +# 3.1. Problem Setup + +Visual imitation learning allows robots to efficiently develop human-like skills using observation-action pairs from expert demonstrations. The intelligent agent must interactively predict subsequent continuous trajectories based on the current observation. The visual observation at time step $t$ , denoted $\mathbf{O}^{(t)} = (\mathbf{C}^{(t)},\mathbf{D}^{(t)})$ , is derived from expert demonstrations, where $\mathbf{C}^{(t)}$ and $\mathbf{D}^{(t)}$ represent single-view RGB and depth images, respectively. Using the visual observation $\mathbf{O}^{(t)}$ , the visuomotor policy learning model predicts the optimal action $\mathbf{a}^{(t)}$ at current time step and generates subsequent optimal continuous action sequence $\mathbf{a}^{(t + 1)}$ . + +To effectively train a visuomotor policy learning model, expert demonstrations are supplied as offline datasets for visual imitation learning, consisting of sample triplets with visual observations and expert actions [26, 40]. Current methods utilize robust visual representations, such as Diffusion Policy and 3D Diffusion Policy, to derive informative latent features for optimal action prediction. However, these methods heavily depend on behavior cloning with supervised historical trajectories. This approach struggles to fully + +capture 3D structures and often overlooks 4D spatiotemporal dynamics. Consequently, the predicted actions frequently fail to achieve complex human goals, particularly when accurate object interactions are absent. In this study, we introduce the 4D Diffusion Policy, which employs a dynamic Gaussian world model to extract 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction. + +# 3.2. Multi-level 3D Spatial Awareness + +To capture complex 3D structures in the interactive physical world, DP4 achieves multi-level spatial perception by generating 3D representations from local point clouds and global voxel grids. These 3D representations are derived from an RGB-D image captured by a single-view camera with known extrinsic and intrinsic parameters, enhancing practical applicability in real-world settings. DP4 avoids impractical setups, such as using multiple cameras during training and inference or requiring extensive prior environmental visual data. In both simulated and real-world settings, we acquire depth images from a single camera and convert them into complete point clouds using the camera's extrinsic and intrinsic parameters. + +Learning 3D representations To generate the 3D local representation, DP4 crops the point clouds, retaining only points within a specified bounding box. This allows DP4 to focus on critical scene elements (e.g., the robot arm) and exclude redundant points (e.g., the ground). DP4 encodes the cropped point clouds into compact 3D local representations using a lightweight MLP network, called the 3D Local Encoder. To obtain the 3D global representation, DP4 converts the point clouds at step $t$ into a $100^3$ voxel. The 3D Global Encoder encodes this voxel, outputting the 3D global representation $v^{(t)} \in \mathbb{R}^{100^3 \times 64}$ . + +Enhance 3D spatial awareness To enhance the representation $v^{(t)}$ with global structure and texture information, we employ a Generalizable Gaussian Regressor, which takes the deep volume $v$ as the scene representation to derive the Gaussian world model based on Gaussian Splatting [19]. The Generalizable Gaussian Regressor is learned by rendering RGB and depth images from the constructed Gaussian world model. The entire neural rendering process is described below. Define $v^{(t)}\mathbf{x}\in \mathbb{R}^{128}$ as the sampled 3D feature at point $\mathbf{x}$ from the 3D global representation $v^{(t)}$ . $v^{(t)}\mathbf{x}$ is obtained through trilinear interpolation, given the discretized nature of the 3D global representation $v^{(t)}$ . The Generalizable Gaussian Regressor uses the 3D coordinates $\mathbf{x}$ and the feature $v_{\mathbf{x}}^{(t)}$ to parameterize the Gaussian primitive as $\theta^{(t)} = (\mu^{(t)},c^{(t)},r^{(t)},s^{(t)},\sigma^{(t)})$ , representing the position, color, rotation, scale, and opacity of the Gaussian primitive. To render a single view, we project Gaussian primitives onto the 2D plane using differential tile-based rasterization. The pixel value $\mathbf{p}$ is rendered using alpha + +blend rendering: + +$$ +C ^ {(t)} (\mathbf {p}) = \sum_ {i = 1} ^ {N} \alpha_ {i} ^ {(t)} c _ {i} ^ {(t)} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} ^ {(t)}), \tag {1} +$$ + +$$ +D ^ {(t)} (\mathbf {p}) = \sum_ {i = 1} ^ {N} \alpha_ {i} ^ {(t)} d _ {i} ^ {(t)} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} ^ {(t)}), +$$ + +where $\alpha_{i}^{(t)} = \sigma_{i}^{(t)}\exp -\frac{1}{2} (\mathbf{p} - \mu_{i}^{(t)})^{\top}\Sigma_{i}^{-1}(\mathbf{p} - \mu_{i}^{(t)})$ defines the 2D density of Gaussian points during the splatting process, $C^{(t)}$ and $D^{(t)}$ represent the rendered RGB image and depth, respectively, $N$ is the number of Gaussians per tile, $d_{i}^{(t)}$ is the $z$ -value obtained by transforming coordinate $u$ from world to camera coordinates, and $\Sigma_{i}$ denotes the covariance matrix derived from the Gaussian parameters' rotation and scale. Our Generalizable Gaussian Regressor and 3D representation are optimized to reconstruct RGB and depth images from a single view by minimizing a 3D loss function: + +$$ +\mathcal {L} _ {\mathrm {3 D}} = \sum_ {\mathbf {p}} \| \mathbf {C} ^ {(t)} (\mathbf {p}) - \mathbf {C} ^ {* (t)} \| _ {2} ^ {2} + \| \mathbf {D} ^ {(t)} (\mathbf {p}) - \mathbf {D} ^ {* (t)} \| _ {2} ^ {2}, \tag {2} +$$ + +where $\mathbf{C}^{*(t)}$ and $\mathbf{D}^{*(t)}$ represent the ground truth color and depth at time step $t$ . This 3D loss allows the 3D representation to emphasize local visual features while incorporating global structural knowledge and texture information. + +# 3.3. 4D Spatiotemporal Awareness + +To emphasize 4D spatiotemporal awareness in the diffusion policy, DP4 integrates dynamics into the Gaussian world model described in Section 3.2. This model parameterizes the Gaussian mixture distribution for dynamic Gaussian splatting, enabling future scene reconstruction via parameter propagation. Thus, the model gains valuable supervision in interactive environments by ensuring consistency between reconstructed and real feature scenes. Specifically, the model learns environmental dynamics by predicting the future state $s^{(t + 1)}$ from the current state $s^{(t)}$ and action $a^{(t)}$ at time step $t$ . + +In the dynamic Gaussian world model, we define the current state as the visual observation at the current time step, with actions representing the robot components. These observations and actions predict visual scenes for the next time step, representing the future state. Building on the Gaussian world model, DP4 uses a deformable MLP $p_{\phi}$ to estimate changes in Gaussian parameters during time propagation and a Gaussian renderer $\mathcal{R}$ to generate RGB and depth images for the future state: + +$$ +\Delta \theta^ {(t)} = p _ {\phi} \left(\theta^ {(t)}, a ^ {(t)}\right), +$$ + +$$ +\theta^ {(t + 1)} = \theta^ {(t)} + \Delta \theta^ {(t)}, \tag {3} +$$ + +$$ +C ^ {(t + 1)} (\mathbf {p}), D ^ {(t + 1)} (\mathbf {p}) = \mathcal {R} \left(\theta^ {(t + 1)}, \mathbf {p}\right), +$$ + +where $\mathbf{C}^{(t + 1)}(\mathbf{p})$ and $\mathbf{D}^{(t + 1)}(\mathbf{p})$ represent the predicted RGB and depth images at time step $t + 1$ , $\theta^{(t)}$ represents the 3D Gaussian primitive parameters at time step $t$ , $w$ denotes the camera pose used to project the Gaussian primitives, and $\mathcal{R}$ is the rendering process from Equation 1. DP4 employs multi-head neural networks as the Gaussian regressor, with each head predicting distinct features of the Gaussian parameters. By estimating changes in position and rotation between time steps, we derive the Gaussian parameters for the future step. The Gaussian renderer then projects the propagated Gaussian distribution onto the single view to reconstruct the future scene. To model 4D spatiotemporal relationships in the physical world, we enforce consistency supervision between reconstructed and real scenes at time step $t + 1$ , embedding scene-level spatiotemporal dynamics into the Gaussian parameters. Specifically, the training objective aligns predicted future scenes, derived from current observations and actions, with real scenes, as formulated below: + +$$ +\mathcal {L} _ {\mathrm {4 D}} = \sum_ {\mathbf {p}} \| \mathbf {C} ^ {(t + 1)} (\mathbf {p}) - \mathbf {C} ^ {* (t + 1)} \| _ {2} ^ {2} + \| \mathbf {D} ^ {(t + 1)} (\mathbf {p}) - \mathbf {D} ^ {* (t + 1)} \| _ {2} ^ {2}, \tag {4} +$$ + +where $\mathbf{C}^{(t + 1)}$ and $\mathbf{D}^{(t + 1)}$ represent the real future counterparts. This additional task of predicting future scenes from current observations via the dynamic Gaussian world model serves as a spatiotemporal enhancement penalty. Training with this task, the multi-level 3D representation learns to encode physical properties and develop awareness of 4D spatiotemporal dynamics. + +# 3.4. Diffusion-based Decision + +The multi-level 3D representation, including local and global 3D representations, is optimized to predict the robot's optimal actions. We model this process as a conditional denoising diffusion framework, conditioned on the multi-level 3D representation $r$ and the robot state $q$ , which converts randomly sampled Gaussian noise into actions $a$ . Starting with sampled Gaussian noise $a^K$ , the denoising network $\epsilon_{\theta}$ performs $K$ iterations to progressively convert it into the noise-free action $a^0$ . + +$$ +a ^ {k - 1} = \alpha_ {k} \left(a ^ {k} - \gamma_ {k} \epsilon_ {\theta} \left(a ^ {k}, k, r, q\right)\right) + \sigma_ {k} \mathcal {N} (0, \mathrm {I}), \tag {5} +$$ + +where $\mathcal{N}(0,\mathbf{I})$ represents the sampled Gaussian noise, and $\alpha_{k},\gamma_{k}$ , and $\sigma_{k}$ are functions of $k$ defined by the noise scheduler. To train the denoising network $\epsilon_{\theta}$ , we randomly sample a data point $a^0$ from the dataset and use a diffusion process to generate the noise $\epsilon^k$ at iteration $k$ . The training objective is to predict the noise added to the data. + +$$ +\mathcal {L} _ {\text {a c t i o n}} = \operatorname {M S E} \left(\epsilon^ {k}, \epsilon_ {\theta} \left(\bar {\alpha} _ {k} a ^ {0} + \bar {\beta} _ {k} \epsilon^ {k}, k, r, q\right)\right), \tag {6} +$$ + +where $\bar{\alpha}_k$ and $\bar{\beta}_k$ are noise schedule that performs one step noise adding. + +The overall learning objective for DP4 is as follows: + +$$ +\mathcal {L} _ {\mathrm {D P 4}} = \mathcal {L} _ {\text {a c t i o n}} + \lambda_ {\mathrm {3 D}} \mathcal {L} _ {\mathrm {3 D}} + \lambda_ {\mathrm {4 D}} \mathcal {L} _ {\mathrm {4 D}}, \tag {7} +$$ + +where $\lambda_{3\mathrm{D}}$ and $\lambda_{4\mathrm{D}}$ balance the 3D loss for current scenes and the 4D loss for future scenes across different objectives. We train DP4 using a joint training strategy. Empirical observations show that this strategy enhances information fusion while learning shared features. During training, we introduce a warm-up phase, freezing the deformable mapper for the first 500 iterations to establish a stable multi-level 3D compact representation and Gaussian regressor. After the warm-up phase, we simultaneously train the entire Gaussian world model and the diffusion-based decision module. After optimization, DP4 generates optimal actions with spatiotemporal awareness, including both 3D spatial and 4D spatiotemporal perceptions. + +# 4. Experiments + +# 4.1. Experiment Setup + +Simulations For simulation experiments, we gathered 22 tasks across various domains, covering a wide range of robotic skills. These tasks span challenging scenarios, such as bimanual, deformable object, and articulated object manipulation, as well as simpler tasks like parallel gripper manipulation. The tasks include Adroit tasks [30] (Door, Pen, Hammer), DexArt tasks [2] (Laptop, Faucet, Bucket, Toilet), and RLBench tasks [17] (10 challenging tasks, each with at least two variations, totaling 166). + +Real robot We used a KINOVA GEN2 robot with a Realsense D455 depth camera mounted in an eye-to-hand configuration. In an indoor environment, we arranged several bottles and cups to enable the agent to generalize manipulation skills across scenes, designing three tasks, including Grasping bottles, Pouring the water from the bottle into the cup and stacking cups. The RealSense D455 camera records the entire scene and robot state from a third-person perspective. This setup supplies RGB-D observations for policy training and RGB and depth supervision for constructing the dynamic Gaussian world model. + +Expert demonstration Expert trajectories for ManiSkill and Adroit are collected using RL-trained agents, using VRL3 [37] for Adroit and PPO [31] for other domains. We generate successful trajectories with RL agents and ensure that all visual imitation learning algorithms use identical demonstrations. Additionally, we collect 20 demonstrations for each RLBench task using the motion planner. + +Baselines The main objective of DP4 is to highlight the importance of 3D spatial and 4D spatiotemporal awareness in diffusion policies. To this end, our primary baseline is the point cloud-based 3D Diffusion Policy (DP3) [47]. Additionally, we incorporate comparisons with other visual imitation learning methods: IBC [7], BCRNN [26], Diffusion + +Policy (DP) [5], PerAct [35], GNFactor [46], and Mani-Gaussian [25]. Note that image and depth resolutions for all 2D and 3D methods remain consistent across experiments, ensuring fair comparisons. + +Evaluation metric Each experiment is conducted using three seeds. For Adroit and DexArt simulation tasks, we assess 20 episodes every 200 epochs and calculate the mean of the top five success rates. For RLBench tasks, we evaluate 25 episodes per task at the final checkpoint for 10 challenging tasks. We report the mean and standard deviation of success rates across the three seeds. + +Implementation details We implement a convolutional network-based diffusion policy using DDIM[36] as the noise scheduler, preferring sample prediction to enhance high-dimensional action generation. Training employs 100 timesteps, while inference uses 10. Models are trained for 3,000 epochs with a batch size of 32 on both simulated and real-world tasks. Loss hyperparameters $\lambda_{\mathrm{3D}}$ and $\lambda_{\mathrm{4D}}$ are set to 0.1 and 0.01, respectively (see Section 4.4 for details). For fair comparison with DP3[47], we adopt the same 3D Local Encoder: a three-layer MLP, max-pooling for order-equivariant aggregation, and a projection head generating compact 3D local features, with LayerNorm inserted for training stability. All experiments run on an NVIDIA H100 80GB GPU and a 192-vCPU Intel Xeon Platinum 8468. + +# 4.2. Quantitative Comparison + +We compare our DP4 with previous state-of-the-art methods on the Adroit, DexArt, and RLBench simulation tasks. + +Adroit simulation Table 1 presents the success rates for the Adroit simulation tasks. DP4 achieves an overall success rate of $84.7\%$ , significantly outperforming 2D-based methods, including IBC [7], BCRNN [26], and Diffusion Policy [5], while surpassing 3D-based methods such as 3D Diffusion Policy [47]. Experimental results demonstrate that DP4 completes complex tasks more accurately by leveraging spatiotemporal perception. + +DexArt simulation Table 2 presents the success rates for various DexArt simulation tasks. The DP4 achieves a success rate of $82.5\%$ , significantly outperforming 2D-based methods (e.g., IBC, BCRNN, and Diffusion Policy) by $33.5\%$ and 3D-based methods (e.g., 3D Diffusion Policy) by $14.0\%$ . Experimental results demonstrate that DP4 integrates 3D spatial and 4D spatiotemporal perception of the physical world with diffusion policies, achieving higher success rates across multiple tasks. + +RLBench simulation We compare DP4 with previous state-of-the-art methods on the RLBench task suite. Table 3 presents a comparison of average success rates across tasks. DP4 achieves state-of-the-art performance with average success rates of $63.3\%$ and $39.9\%$ , significantly outperforming prior methods, including both perception-based and generative approaches. Leveraging supervision from + +
MethodHammerDoorPenOverall
IBC [7]0±00±09±23.0
BCRNN [26]0±00±09±33.0
DP [5]48±1750±525±441.0
DP3 [47]100±062±443±668.3
DP4 (ours)100±080±275±384.7
+ +Table 1. Success rates $(\%)$ on the Adroit simulation. We compare our DP4 with more baselines in the Adroit simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3). + +
MethodLaptopFaucetBucketToiletOverall
IBC [7]3±27±114±10±06.0
BCRNN [26]3±31±05±50±02.3
DP [5]69±423±858±246±149.0
DP3 [47]83±163±282±446±268.5
DP4 (ours)92±284±190±164±282.5
+ +Table 2. Success rates $(\%)$ on the DexArt simulation. We compare our DP4 with more baselines in the DexArt simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3). + +
Methodclose jaropen drawersweep to dustpanmeat off grillturn tapOverall
PerAct [35]18.7±8.254.7±18.60.0±0.040.0±17.038.7±6.830.4
GNFactor [46]25.3±6.876.0±5.728.0±15.057.3±18.950.7±8.247.5
ManiGaussian [25]28.4±5.476.3±6.264.5±13.460.2±18.256.2±6.857.1
DP4 (ours)35.4±2.282.2±3.472.0±5.264.4±6.862.4±2.863.3
Methodslide blockput in drawerdrag stickpush buttonsstack blocksOverall
PerAct [35]18.7±13.62.7±3.35.3±5.018.7±12.46.7±1.910.4
GNFactor [46]20.0±15.00.0±0.037.3±13.218.7±10.04.0±3.316.0
ManiGaussian [25]24.3±12.816.6±3.292.3±11.420.4±12.212.4±3.233.2
DP4 (ours)33.4±4.422.8±1.294.4±6.228.3±6.821.0±2.239.9
+ +Table 3. Success rates $(\%)$ on the RLBench simulation. We compare our DP4 with more baselines in the RLBench simulation. Baseline approaches include PerAct, GNFactor, and Mani-Gaussian. + +3D geometry, texture, and 4D spatiotemporal data, DP4 accurately predicts optimal continuous trajectories from a single RGB-D observation of the current scene. + +# 4.3. Qualitative Comparison + +Visualization of trajectories We present three qualitative examples of action sequences generated by DP3 and our DP4 in Figure 3. The results show that the previous agent (DP3) frequently fails to complete tasks, as it imitates the expert's motions (e.g., hammering nails, sliding doors, and spinning pens) without fully understanding the scene. In contrast, DP4 significantly improves the success rate by accurately interpreting 3D scene structures (e.g., nail positions) and 4D dynamics (e.g., pen rotation). + +Visualization of Gaussian world model Figure 4 and Figure 5 show the visualization results. We visualize the Gaussian world model via single-view rendering, observ + +![](images/33406afdd17829ceaf93beed74cf9ca0eb168bce9c918ed14d8bcbe980ba0604.jpg) + +![](images/55abc5184a14d7a6e1bf3652b917f8b8eac57d2eec1b503c782c03e053d070df.jpg) +Figure 4. Visualization of Gaussian world model. After training DP4, we visualize its Gaussian world model by rendering a single-view observation. DP4 uses a Gaussian world model to provide supervision, facilitating the learning of 3D spatial perception. + +![](images/6b0358591e8cc5afbac5a33bbb1b82c221d0c714080cb4774ccf1f026f580718.jpg) +Figure 3. Qualitative case study. The red mark indicates a pose that significantly deviates from the expert demonstration, while the green mark denotes a pose that aligns with the expert trajectory. The 4D Diffusion Policy (DP4) integrates 3D spatial and 4D spatiotemporal awareness with diffusion policies, successfully completing the tasks. + +ing it captures fundamental spatial structures and spatiotemporal dynamics. Although DP4 operates under a realistic single-view constraint, resulting renderings lack fine detail, they effectively represent scene-level 3D structures and 4D dynamics. Crucially, such details have minimal impact on task success; for example, the nail's thread is irrelevant to the hammering task. + +# 4.4. Ablation Study + +Component ablation. Ablations are conducted on the Adroit simulation tasks in Table 4. Without the proposed components, we use 3D representation without any 3D spatial supervision (RGB and depth (structure)) and 4D spatiotemporal supervision (dynamics). In the experiment on the Adroit simulation, the performance of the "Hammer", "Door", and "Pen" tasks demonstrated improvements in success rates with the proposed components, while the inference time remained stable or only slightly increased. + +- For the Hammer task: The base model started with a + +![](images/883f8590c4d5270f6a24269dc21164be532fed6fe62a7ed559c2bbe3f62c3aed.jpg) + +94.0% success rate and a completion time of 6.40 seconds. After adding the Gaussian world model with RGB supervision, success improved to $96.0\%$ , with a minimal increase in time (6.52 seconds). Introducing 3D spatial loss (RGB and Depth supervision) raised success to $95.0\%$ , with stable time (6.53 seconds). With 4D spatiotemporal supervision (DP4), success reached $98.0\%$ , and the time increased slightly to 6.56 seconds, indicating a clear performance gain with a small time increase. + +- For the Door task: Initially, the model achieved a $94.0\%$ success rate and a completion time of 6.40 seconds. Adding the Gaussian world model with RGB supervision improved success to $96.0\%$ , with no significant time change. Incorporating 3D spatial loss (RGB and Depth) increased success to $98.0\%$ , with a slight increase in time (6.56 seconds). With 4D spatiotemporal supervision (DP4), the success rate reached $100.0\%$ , with a minimal time increase (6.57 seconds), reflecting significant improvement from additional supervision. +- For the Pen task: The base model had a $45.0\%$ success rate and a completion time of 6.46 seconds. After adding the Gaussian world model with RGB supervision, success rose to $48.0\%$ , with time remaining at 6.48 seconds. Introducing 3D spatial loss (RGB and depth) led to a slight decrease in success $(47.0\%)$ . With 4D spatiotemporal supervision (DP4), success jumped to $75.0\%$ , with a small increase in time to 6.50 seconds, highlighting the effectiveness of 4D supervision. + +Notably, DP4 does not substantially increase model inference time. This is because DP4 does not construct a Gaussian world model during inference. The 3D spatial and 4D spatiotemporal awareness from the Gaussian world model are efficiently integrated into the multi-level 3D representation during training. + +Hyperparameter analysis We conduct hyperparameter + +![](images/6635bc0adfca8423a5256aee603baba3cf156b81b45d6c002b5d7ddd0c35a977.jpg) +Figure 5. More visualizations of the learned world model. DP4 uses a Gaussian world model to provide supervision, facilitating the 4D spatiotemporal perceptions in interactive environments. + +![](images/3f687a09c42ab7486e7610126eba30686b9c1537cebb78bcc95dccbdb2714b3b.jpg) +Ground-truth Prediction at current time step T at current time step T + +![](images/453b535e30b31baa9c589d11f4efc24c74d5502bd15fbd7dfb24f4a742049018.jpg) + +![](images/cec220ca16b1cf6c315ae503fa2472c3ceed9e504c40570ab5461ecd4acbcd47.jpg) + +![](images/957bce2bec6f7cab2553fa72623cdc8f29ef9b55a7a6696381516a5f70741d24.jpg) + +Results on Adroit Simulation + +
3D Component4D ComponentPerformance
RGBDepthDynamicsHammerDoorPen
94.0/6.4064.0/6.4445.0/6.46
✓96.0/6.5268.0/6.4548.0/6.48
✓95.0/6.5368.0/6.4647.0/6.48
✓✓98.0/6.5675.0/6.5172.0/6.47
✓✓✓100.0/6.5780.0/6.4575.0/6.50
+ +![](images/605edf3e11de05e6e61990f0b495689c2878ef6deb6e91c7e3de02c5dda5d763.jpg) +Table 4. Component ablation studies. We report results for the Adroit simulation tasks. In the performance metrics, the number to the left of / represents the success rate $(\%)$ , and the number to the right denotes the inference speed (seconds per task). +Figure 6. Hyperparameter analysis of DP4. The effects of $\lambda_{3\mathrm{D}}$ and $\lambda_{4\mathrm{D}}$ on success rates, respectively. + +![](images/c23d20c7befb132e4a8f5b7d43aff860fc7336241d097142d17882a6d71e23f8.jpg) + +
MethodGrasping BottlesStacking CupsPouring WaterOverall
DP [5]36.044.028.036.0
DP3 [47]42.062.034.046.0
DP4 (ours)48.072.044.054.6
+ +Table 5. Success rates $(\%)$ on the real-robot tasks. We compare our DP4 with additional baselines on the real-robot tasks. We include the Diffusion Policy (DP) and 3D Diffusion Policy (DP3). + +analysis on the Pen task within the Adroit simulation. DP4 uses two key hyperparameters, $\lambda_{3\mathrm{D}}$ and $\lambda_{4\mathrm{D}}$ , to balance the 3D spatial loss for current scenes and the 4D spatiotemporal loss for future scenes across multiple objectives. As shown in Figure 6, we analyze the hyperparameters of DP4. We vary $\lambda_{3\mathrm{D}}$ from 0 to 1 and $\lambda_{4\mathrm{D}}$ from 0 to 0.1, measuring the corresponding success rates. We find that the success rate peaks when $\lambda_{3\mathrm{D}} = 0.1$ and $\lambda_{4\mathrm{D}} = 0.01$ , so we adopt these values as the hyperparameters. + +# 4.5. Real-Robot Analysis + +To assess DP4's performance on real-world robotic tasks, we designed three tasks: Grasping Bottles, Stacking Cups, + +![](images/d0006c5f3ef97680ed3f00549108ce39aea3b25706ca016c58abd799e9c5a29e.jpg) + +![](images/98f0d77bd8b174e1eba639289184742a9e61457dca5d4b8dcf066eb07ef8ae1d.jpg) +"Grasping Bottles" + +![](images/84f69a1a1bf6d24fed2441dfb88af1390ac474edccf3bf87883ccf0c8af9e3d6.jpg) + +![](images/cb06977a3c8925dfd7df526a5a5e029eecbe37dd617e96e06055036b4a6d5949.jpg) + +![](images/dd1e93b1926e181abac301acf7640fa77227c481de793693ab633cc4dd3e2c81.jpg) +"Stacking Cups" + +![](images/b1e9fa26f7d2e318e2915001e96357d37d0444f893c3d4441254bc6ec8586319.jpg) + +![](images/36e04d4ea11b207f564179f00f38c151034526724033846e24e2dcb27ddc58bf.jpg) +Figure 7. Visualization of DP4 performance on three real-world robotic tasks. DP4 demonstrates strong performance in real-world settings and effectively handles a variety of common tasks with a single view. + +![](images/ab9c019b907ac951141fc7fe64dd502d14a44dda55550f9defa5f81f13d05780.jpg) +"Pouring Water" + +![](images/7f3e10bc985da41cf79c0927b4c6aa5a00f61487a1dc178e95246ce839fabd03.jpg) + +and Pouring Water. We focus on the model's performance at key frames, where success is defined as when the error is below a certain threshold. The results of the real-world robotic experiments are summarized in Table 5. Experimental results show that DP4 outperforms the Diffusion Policy (DP) and 3D Diffusion Policy (DP3) baselines across all tasks. These results confirm that DP4 effectively integrates spatiotemporal awareness with diffusion policies, improving the accuracy of generated trajectories. Additionally, visualizations of DP4's performance on real-world tasks are shown in Figure 7. + +# 5. Conclusion + +This paper presents 4D Diffusion Policy (DP4), which integrates spatiotemporal awareness of the physical world with diffusion-based visual imitation learning. Utilizing a single-view RGB-D input, DP4 constructs a dynamic Gaussian world model that provides rich supervision for acquiring 3D spatial and 4D spatiotemporal perceptions from interactive environments. Extensive experiments on simulated and real-robot tasks validate the effectiveness and superiority. + +# 6. Acknowledgments + +This work was supported in part by the Science and Technology Commission of Shanghai Municipality (No.24511103100), NSFC Project (62176061), Doubao Fund, and Shanghai Technology Development and Entrepreneurship Platform for Neuromorphic and AI SoC. + +# References + +[1] Sridhar Pandian Arunachalam, Sneha Silwal, Ben Evans, and Lerrel Pinto. Dexterous imitation made easy: A learning-based framework for efficient dexterous manipulation. In 2023 IEEE international conference on robotics and automation (icra), pages 5954-5961. IEEE, 2023. 2 +[2] Chen Bao, Helin Xu, Yuzhe Qin, and Xiaolong Wang. Dexart: Benchmarking generalizable dexterous manipulation with articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21190-21200, 2023. 5 +[3] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1 +[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 2 +[5] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8 +[6] Yuqing Du, Daniel Ho, Alex Alemi, Eric Jang, and Mohi Khansari. Bayesian imitation learning for end-to-end mobile manipulation. In International Conference on Machine Learning, pages 5531-5546. PMLR, 2022. 1 +[7] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on robot learning, pages 158-168. PMLR, 2022. 2, 5, 6 +[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 3 +[9] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. +[10] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. +[11] Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems, 35:26091-26104, 2022. 3 +[12] Siddhant Haldar, Jyothish Pari, Anant Rai, and Lerrel Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations. arXiv preprint arXiv:2303.01497, 2023. 1 +[13] Nicklas Hansen, Zhecheng Yuan, Yanjie Ze, Tongzhou Mu, Aravind Rajeswaran, Hao Su, Huazhe Xu, and Xiaolong Wang. On pre-training for visuo-motor control: Revisiting a learning-from-scratch baseline. arXiv preprint arXiv:2212.05749, 2022. 1 +[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif + +fusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[15] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems, 35:20703-20716, 2022. 3 +[16] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3 +[17] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5, 4 +[18] Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE international conference on robotics and automation (ICRA), pages 4613-4619. IEEE, 2021. 1 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4 +[20] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2 +[21] Peiyan Li, Hongtao Wu, Yan Huang, Chilam Cheang, Liang Wang, and Tao Kong. Gr-mg: Leveraging partially-annotated data via multi-modal goal-conditioned policy. IEEE Robotics and Automation Letters, 2025. +[22] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 2 +[23] Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Reasongrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3718-3727, 2025. 3 +[24] Zhenyang Liu, Sixiao Zheng, Siyu Chen, Cairong Zhao, Longfei Liang, Xiangyang Xue, and Yanwei Fu. A neural representation framework with llm-driven spatial reasoning for open-vocabulary 3d visual grounding. arXiv preprint arXiv:2507.06719, 2025. 3 +[25] Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In European Conference on Computer Vision, pages 349-366. Springer, 2024. 6 +[26] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio + +Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 2, 3, 5, 6 +[27] Liyuan Mao, Haoran Xu, Xianyuan Zhan, Weinan Zhang, and Amy Zhang. Diffusion-dice: In-sample diffusion guidance for offline reinforcement learning. arXiv preprint arXiv:2407.20109, 2024. 2 +[28] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020.1 +[29] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 1, 2 +[30] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087, 2017. 5 +[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5 +[32] Younggyo Seo, Danijar Hafner, Hao Liu, Fangchen Liu, Stephen James, Kimin Lee, and Pieter Abbeel. Masked world models for visual control. In Conference on Robot Learning, pages 1332-1344. PMLR, 2023. 3 +[33] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. 1, 2 +[34] Nur Muhammad Mahi Shafullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, Soumith Chintala, and Lerrel Pinto. On bringing robots home. arXiv preprint arXiv:2311.16098, 2023. 1 +[35] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 6 +[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 6 +[37] Che Wang, Xufang Luo, Keith Ross, and Dongsheng Li. Vrl3: A data-driven framework for visual deep reinforcement learning. Advances in Neural Information Processing Systems, 35:32974-32988, 2022. 5 +[38] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimi-cplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1, 2 +[39] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 3 + +[40] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning. arXiv preprint arXiv:2208.06193, 2022. 2, 3 +[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2 +[42] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226-2240. PMLR, 2023. 3 +[43] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 2 +[44] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2 +[45] Ruihan Yang, Zhuoqun Chen, Jianhan Ma, Chongyi Zheng, Yiyu Chen, Quan Nguyen, and Xiaolong Wang. Generalized animal imitator: Agile locomotion with versatile motion prior. arXiv preprint arXiv:2310.01408, 2023. 1 +[46] Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. In Conference on Robot Learning, pages 284-301. PMLR, 2023. 6 +[47] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024. 2, 5, 6, 8, 3 +[48] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37:77799-77830, 2025. 2 \ No newline at end of file diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/images.zip b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab65b92d10bcc5a663f78cb97eb9c38cd57761ba --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8c87686e8f2699764c4ca5b5c468e4160a5850c07c3f4971d374e29ced580fd +size 480595 diff --git a/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/layout.json b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c77818ec6219baaf68683e1f534f5bf40bc9b472 --- /dev/null +++ b/2025/Spatial-Temporal Aware Visuomotor Diffusion Policy Learning/layout.json @@ -0,0 +1,9948 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 127, + 484, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 127, + 484, + 144 + ], + "spans": [ + { + "bbox": [ + 126, + 127, + 484, + 144 + ], + "type": "text", + "content": "Spatial-Temporal Aware Visuomotor Diffusion Policy Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "spans": [ + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "text", + "content": "Zhenyang Liu" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "text", + "content": ", Yikai Wang" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "text", + "content": ", Kuanning Wang" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "text", + "content": ", Longfei Liang" + }, + { + "bbox": [ + 159, + 163, + 449, + 177 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "spans": [ + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "type": "text", + "content": "Xiangyang Xue" + }, + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "type": "text", + "content": ", Yanwei Fu" + }, + { + "bbox": [ + 235, + 177, + 378, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "text", + "content": "Shanghai Innovation Institute " + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "type": "text", + "content": "NeuHelium Co., Ltd" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 205, + 479, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 205, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 479, + 215 + ], + "type": "text", + "content": "liuzy24@m.fudan.edu.cn, yikai.wang@ntu.edu.sg, knwang24@m.fudan.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 161, + 218, + 449, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 218, + 449, + 228 + ], + "spans": [ + { + "bbox": [ + 161, + 218, + 449, + 228 + ], + "type": "text", + "content": "longfei.liang@neuhelium.com, {xyxue,yanweifu}@fudan.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 211, + 229, + 399, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 229, + 399, + 241 + ], + "spans": [ + { + "bbox": [ + 211, + 229, + 399, + 241 + ], + "type": "text", + "content": "Project Page: ZhenyangLiu.github.io/DP4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 164, + 266, + 209, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 266, + 209, + 277 + ], + "spans": [ + { + "bbox": [ + 164, + 266, + 209, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "content": "Visual imitation learning is effective for robots to learn versatile tasks. However, many existing methods rely on behavior cloning with supervised historical trajectories, limiting their 3D spatial and 4D spatiotemporal awareness. Consequently, these methods struggle to capture the 3D structures and 4D spatiotemporal relationships necessary for real-world deployment. In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that incorporates spatiotemporal awareness into diffusion-based policies. Unlike traditional approaches that rely on trajectory cloning, DP4 leverages a dynamic Gaussian world model to guide the learning of 3D spatial and 4D spatiotemporal perceptions from interactive environments. Our method constructs the current 3D scene from a single-view RGB-D observation and predicts the future 3D scene, optimizing trajectory generation by explicitly modeling both spatial and temporal dependencies. Extensive experiments across 17 simulation tasks with 173 variants and 3 real-world robotic tasks demonstrate that the 4D Diffusion Policy (DP4) outperforms baseline methods, improving the average simulation task success rate by " + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "inline_equation", + "content": "16.4\\%" + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "content": " (Adroit), " + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "content": " (DexArt), and " + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "inline_equation", + "content": "6.45\\%" + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "content": " (RLBench), and the average real-world robotic task success rate by " + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 75, + 290, + 297, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 565, + 149, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 565, + 149, + 576 + ], + "spans": [ + { + "bbox": [ + 77, + 565, + 149, + 576 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 75, + 584, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 584, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 75, + 584, + 296, + 651 + ], + "type": "text", + "content": "Visual Imitation learning has emerged as an effective method for training robots to perform complex tasks, such as object grasping [18, 35, 38], legged locomotion [28, 45], dexterous manipulation [12, 29], and mobile manipulation [6, 34]. Recent advances in computer vision have further enhanced visual imitation learning, enabling impress-" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 266, + 532, + 353 + ], + "blocks": [ + { + "bbox": [ + 317, + 266, + 532, + 353 + ], + "lines": [ + { + "bbox": [ + 317, + 266, + 532, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 532, + 353 + ], + "type": "image", + "image_path": "0f78a27e752a7b0442405a313f575cb9ed5a83b5392c3d5445cadaad9465b90a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 355, + 533, + 487 + ], + "blocks": [ + { + "bbox": [ + 316, + 355, + 533, + 487 + ], + "lines": [ + { + "bbox": [ + 316, + 355, + 533, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 533, + 487 + ], + "type": "image", + "image_path": "f6fc01632182ef501da5f9573c1b55e20f80f26291f4e90bbbc9d441f0255193.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 494, + 534, + 585 + ], + "lines": [ + { + "bbox": [ + 312, + 494, + 534, + 585 + ], + "spans": [ + { + "bbox": [ + 312, + 494, + 534, + 585 + ], + "type": "text", + "content": "Figure 1. Spatial-temporal awareness in the 4D Diffusion Policy (DP4). Previous methods train perception and decision-making with trajectory supervision, but trajectory cloning fails to capture the 3D spatial and 4D spatiotemporal relationships. In contrast, DP4 constructs the current 3D scene with 3D spatial supervision from a single RGB-D view and predicts future 3D scene candidates using 4D spatiotemporal supervision, optimizing trajectory generation by effectively capturing both 3D structures and 4D dependencies." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 588, + 533, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 533, + 675 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 533, + 675 + ], + "type": "text", + "content": "sive performance across various applications [3, 13, 33, 35]. Progress in computer vision has enabled sophisticated visual imitation learning methods that demonstrate excellent performance across various applications [3, 13, 33, 35]. However, real-world robotic deployment demands robust generalization to unseen scenarios, necessitating a deeper understanding of the spatiotemporal physical rules that govern dynamic environments." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 676, + 533, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 676, + 533, + 687 + ], + "spans": [ + { + "bbox": [ + 324, + 676, + 533, + 687 + ], + "type": "text", + "content": "Despite their success, existing visual imitation learning" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 72, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 72, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "spans": [ + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "spans": [ + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 659, + 161, + 668 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 659, + 161, + 668 + ], + "spans": [ + { + "bbox": [ + 86, + 659, + 161, + 668 + ], + "type": "text", + "content": "†Corresponding authors." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 77, + 669, + 295, + 687 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 669, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 77, + 669, + 295, + 687 + ], + "type": "inline_equation", + "content": "^{\\ddagger}" + }, + { + "bbox": [ + 77, + 669, + 295, + 687 + ], + "type": "text", + "content": "Prof. Yanwei Fu is also with Institute of Trustworthy Embodied AI, and the School of Data Science, Fudan University." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7122" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 99, + 295, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 99, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 75, + 99, + 295, + 175 + ], + "type": "text", + "content": "methods struggle to capture the 3D structure of environments and 4D spatiotemporal dependencies of interactive tasks. These limitations arise mainly because current approaches rely on behavior cloning from static, supervised trajectories. Such methods often fail to infer the essential physical properties needed for spatial reasoning, object interaction, and environment-aware decision-making." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 176, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 176, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 75, + 176, + 295, + 285 + ], + "type": "text", + "content": "In this work, we propose 4D Diffusion Policy (DP4), a novel visual imitation learning method that integrates spatial-temporal awareness into diffusion-based policies. Unlike prior works that passively mimic demonstrations exclusively from supervised trajectories, DP4 actively learns from a dynamic Gaussian world model [19, 41], which explicitly encodes 3D spatial structures and 4D spatiotemporal variations as part of the imitation learning process. Thus, as illustrated in Fig. 1, DP4 extends conventional imitation learning pipelines in the following ways:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 286, + 295, + 407 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 286, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 76, + 286, + 295, + 319 + ], + "type": "text", + "content": "- 3D spatial perception is learned via Gaussian Splitting (3DGS) [19], ensuring accurate reconstruction of the current scene's geometry from point cloud inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 319, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 319, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 76, + 319, + 295, + 363 + ], + "type": "text", + "content": "- 4D spatiotemporal awareness is enforced by future 3DGS reconstruction, where the system learns to predict and adaptively refine future environmental states based on the agent's planned trajectories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 363, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 363, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 76, + 363, + 295, + 407 + ], + "type": "text", + "content": "- Trajectory generation is conditioned on these structured spatial-temporal representations, allowing the diffusion policy model to reason over future states rather than just replicating past behaviors." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 76, + 407, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 407, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 76, + 407, + 295, + 462 + ], + "type": "text", + "content": "So our method first converts RGB-D observations into 3D point clouds, from which we extract global and local spatiotemporal representations. These serve as conditions for the diffusion policy model to guide next-step trajectory generation. To strengthen structured perception:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 464, + 295, + 540 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 464, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 464, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 76, + 464, + 295, + 497 + ], + "type": "text", + "content": "- We present a novel Gaussian regressor that learns to reconstruct the 3D scene from point cloud inputs, ensuring spatially consistent representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 76, + 498, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 498, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 76, + 498, + 295, + 540 + ], + "type": "text", + "content": "- We enforce temporal consistency by deforming the current scene's 3DGS into the predicted future state, supervising this process using ground-truth RGB-D reconstructions from the future scene." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "content": "We evaluate DP4 extensively across 17 simulated tasks (173 variants) from four benchmarks, along with three real-world tasks, including the challenging task of dexterous manipulation of deformable objects. Our experiments show that DP4 significantly outperforms state-of-the-art multi-task visual imitation methods, achieving a " + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "inline_equation", + "content": "+16.4\\%" + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "content": " success rate on Adroit, " + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "inline_equation", + "content": "+14\\%" + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "content": " on DexArt, " + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "inline_equation", + "content": "+6.45\\%" + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "content": " on RL-Bench, and " + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "inline_equation", + "content": "+8.6\\%" + }, + { + "bbox": [ + 75, + 544, + 295, + 632 + ], + "type": "text", + "content": " improvement in real-world robotic tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 632, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 632, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 75, + 632, + 295, + 686 + ], + "type": "text", + "content": "Our work bridges the gap between visual imitation learning and spatial-temporal physical reasoning, demonstrating the importance of structured world models in improving generalization and robustness in real-world robotic applications, with these contributions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 99, + 533, + 285 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 99, + 533, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 533, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 533, + 153 + ], + "type": "text", + "content": "- Spatiotemporal-centric visual imitation learning: Unlike existing methods that overlook 3D perception and dynamic interactions, DP4 explicitly models 3D spatial structures and 4D temporal dynamics, improving generalization in dynamic environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 154, + 533, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 533, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 533, + 198 + ], + "type": "text", + "content": "- 4D Diffusion Policy with structured awareness: We introduce a diffusion-based visual imitation learning framework that generates trajectories based on learned spatial-temporal world representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 198, + 533, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 533, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 533, + 241 + ], + "type": "text", + "content": "- Dynamic Gaussian World Model for structured supervision: Our model actively learns from real-world interactions, ensuring the embedding of 3D spatial and 4D temporal reasoning into policy learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 242, + 533, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 533, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 533, + 285 + ], + "type": "text", + "content": "- State-of-the-art performance across diverse tasks: DP4 sets unprecedented success rates across 17 simulated and real-world robotic tasks, setting a new benchmark for imitation learning in dynamic environments." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 296, + 392, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 392, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 392, + 307 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 314, + 533, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 314, + 533, + 478 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 533, + 478 + ], + "type": "text", + "content": "Visual imitation learning [20-22] is an effective method for equipping robots with skills, typically relying on numerous expert demonstrations as observation-action pairs. Recent advanced visual imitation learning methods commonly employ 2D images[1, 5, 7, 26, 33, 38] as state representations for their rich global information and ease of acquisition from raw sensory inputs, whereas 3D representations (e.g., point clouds and object geometries) [47, 48] are used for their explicit spatial structure, enhancing the spatial reasoning. However, these methods overlook the spatiotemporal dynamics of scenes arising from behavior-environment interactions. DP4 overcomes this limitation by employing a dynamic Gaussian world model to derive scene-level 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 479, + 533, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 533, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 533, + 654 + ], + "type": "text", + "content": "Diffusion models as policies Building on their success in image and video generation [4, 14, 44], diffusion models have emerged as effective policy frameworks in robotics [5, 27, 29, 43, 47]. They are primarily applied in reinforcement learning and visual imitation learning, with Diffusion Policy [5] and variants [40, 47] achieving state-of-the-art results by fusing visual inputs (e.g., 2D images, 3D point clouds) with complex, multi-modal action spaces. Nevertheless, these approaches often underutilize visual perception, relying heavily on behavior cloning from supervised trajectories. In contrast, DP4 introduces a Gaussian world model that reconstructs current scenes from RGB-D data via volumetric rendering for 3D spatial understanding, and employs a deformation Gaussian field to predict future scene distributions through environment interaction, enabling comprehensive 4D spatiotemporal perception." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "type": "text", + "content": "World model Recent advances demonstrate that world models effectively encode scene dynamics by predicting future states from current states and actions, with appli" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7123" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 99, + 508, + 350 + ], + "blocks": [ + { + "bbox": [ + 104, + 99, + 508, + 350 + ], + "lines": [ + { + "bbox": [ + 104, + 99, + 508, + 350 + ], + "spans": [ + { + "bbox": [ + 104, + 99, + 508, + 350 + ], + "type": "image", + "image_path": "32729a1c19491f8baf91483ddd502ecc561f35e53eb99941670633d99d84a506.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 354, + 535, + 435 + ], + "lines": [ + { + "bbox": [ + 75, + 354, + 535, + 435 + ], + "spans": [ + { + "bbox": [ + 75, + 354, + 535, + 435 + ], + "type": "text", + "content": "Figure 2. The framework of our 4D Diffusion Policy (DP4). From a single-view RGB-D observation, we construct 3D point clouds and extract global and local features to enrich both holistic and focused perceptions. These multi-level representations condition the diffusion policy model to generate trajectories based on current robot states. We introduce a Gaussian world model in DP4 to capture 3D structures and 4D spatiotemporal relationships. The current observation's 3D Gaussian Splitting (3DGS) is derived via a generalizable Gaussian regressor from point clouds and multi-level features. By enforcing consistency between ground-truth and rendered RGB-D images from this 3DGS, we enhance 3D spatial awareness. Additionally, future 3DGS are predicted from current states using policy-generated trajectories, with rendered RGB-D consistency fostering 4D spatiotemporal awareness. This improved spatiotemporal representation significantly benefits complex tasks such as object grasping and dexterous manipulation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 451, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 451, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 75, + 451, + 296, + 582 + ], + "type": "text", + "content": "cations in autonomous driving [15, 16, 39] and robotic manipulation [24, 32, 42]. Early works [8-11, 15, 23] employed latent space autoencoding for future prediction, yielding strong results in both simulated and real environments. Nonetheless, learning accurate latent representations is data-intensive and generally confined to simple tasks due to limited implicit feature capacity. To address these challenges, this study integrates Gaussian Splating [19], known for its efficiency and rendering quality, into world model construction. The resulting 4D Diffusion Policy leverages a dynamic Gaussian world model to capture scene-level 3D structure and 4D spatiotemporal dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 592, + 151, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 592, + 151, + 604 + ], + "spans": [ + { + "bbox": [ + 76, + 592, + 151, + 604 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 610, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 610, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 75, + 610, + 296, + 687 + ], + "type": "text", + "content": "In this section, we introduce 4D Diffusion Policy (DP4), a novel visual imitation learning method that emphasizes spatial-temporal awareness in diffusion policies. DP4 employs 3D Gaussian Splatting for informative supervision, enabling the learning of 3D spatial and 4D spatiotemporal perceptions in interactive environments. The DP4 framework is illustrated in Figure 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 450, + 399, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 399, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 399, + 462 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": "Visual imitation learning allows robots to efficiently develop human-like skills using observation-action pairs from expert demonstrations. The intelligent agent must interactively predict subsequent continuous trajectories based on the current observation. The visual observation at time step " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": ", denoted " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{(t)} = (\\mathbf{C}^{(t)},\\mathbf{D}^{(t)})" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": ", is derived from expert demonstrations, where " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(t)}" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{(t)}" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": " represent single-view RGB and depth images, respectively. Using the visual observation " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^{(t)}" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": ", the visuomotor policy learning model predicts the optimal action " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{(t)}" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": " at current time step and generates subsequent optimal continuous action sequence " + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{(t + 1)}" + }, + { + "bbox": [ + 312, + 467, + 533, + 588 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 588, + 534, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 534, + 688 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 534, + 688 + ], + "type": "text", + "content": "To effectively train a visuomotor policy learning model, expert demonstrations are supplied as offline datasets for visual imitation learning, consisting of sample triplets with visual observations and expert actions [26, 40]. Current methods utilize robust visual representations, such as Diffusion Policy and 3D Diffusion Policy, to derive informative latent features for optimal action prediction. However, these methods heavily depend on behavior cloning with supervised historical trajectories. This approach struggles to fully" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7124" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 99, + 296, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 99, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 99, + 296, + 186 + ], + "type": "text", + "content": "capture 3D structures and often overlooks 4D spatiotemporal dynamics. Consequently, the predicted actions frequently fail to achieve complex human goals, particularly when accurate object interactions are absent. In this study, we introduce the 4D Diffusion Policy, which employs a dynamic Gaussian world model to extract 3D structures from current scene reconstruction and 4D spatiotemporal dynamics from future scene reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 195, + 239, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 195, + 239, + 207 + ], + "spans": [ + { + "bbox": [ + 76, + 195, + 239, + 207 + ], + "type": "text", + "content": "3.2. Multi-level 3D Spatial Awareness" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 212, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 212, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 75, + 212, + 296, + 354 + ], + "type": "text", + "content": "To capture complex 3D structures in the interactive physical world, DP4 achieves multi-level spatial perception by generating 3D representations from local point clouds and global voxel grids. These 3D representations are derived from an RGB-D image captured by a single-view camera with known extrinsic and intrinsic parameters, enhancing practical applicability in real-world settings. DP4 avoids impractical setups, such as using multiple cameras during training and inference or requiring extensive prior environmental visual data. In both simulated and real-world settings, we acquire depth images from a single camera and convert them into complete point clouds using the camera's extrinsic and intrinsic parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "text", + "content": "Learning 3D representations To generate the 3D local representation, DP4 crops the point clouds, retaining only points within a specified bounding box. This allows DP4 to focus on critical scene elements (e.g., the robot arm) and exclude redundant points (e.g., the ground). DP4 encodes the cropped point clouds into compact 3D local representations using a lightweight MLP network, called the 3D Local Encoder. To obtain the 3D global representation, DP4 converts the point clouds at step " + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "100^3" + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "text", + "content": " voxel. The 3D Global Encoder encodes this voxel, outputting the 3D global representation " + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "v^{(t)} \\in \\mathbb{R}^{100^3 \\times 64}" + }, + { + "bbox": [ + 75, + 355, + 296, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": "Enhance 3D spatial awareness To enhance the representation " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v^{(t)}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " with global structure and texture information, we employ a Generalizable Gaussian Regressor, which takes the deep volume " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " as the scene representation to derive the Gaussian world model based on Gaussian Splatting [19]. The Generalizable Gaussian Regressor is learned by rendering RGB and depth images from the constructed Gaussian world model. The entire neural rendering process is described below. Define " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v^{(t)}\\mathbf{x}\\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " as the sampled 3D feature at point " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " from the 3D global representation " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v^{(t)}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v^{(t)}\\mathbf{x}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " is obtained through trilinear interpolation, given the discretized nature of the 3D global representation " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v^{(t)}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": ". The Generalizable Gaussian Regressor uses the 3D coordinates " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " and the feature " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "v_{\\mathbf{x}}^{(t)}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " to parameterize the Gaussian primitive as " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\theta^{(t)} = (\\mu^{(t)},c^{(t)},r^{(t)},s^{(t)},\\sigma^{(t)})" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": ", representing the position, color, rotation, scale, and opacity of the Gaussian primitive. To render a single view, we project Gaussian primitives onto the 2D plane using differential tile-based rasterization. The pixel value " + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 75, + 477, + 296, + 687 + ], + "type": "text", + "content": " is rendered using alpha" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 99, + 376, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 376, + 110 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 376, + 110 + ], + "type": "text", + "content": "blend rendering:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 352, + 115, + 533, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 115, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 352, + 115, + 533, + 152 + ], + "type": "interline_equation", + "content": "C ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} c _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}), \\tag {1}", + "image_path": "d5c3d8114baad7605d230ba68c68c11d979403e4f677ed3f6b421fc1dccea03d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 150, + 494, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 150, + 494, + 180 + ], + "spans": [ + { + "bbox": [ + 353, + 150, + 494, + 180 + ], + "type": "interline_equation", + "content": "D ^ {(t)} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} \\alpha_ {i} ^ {(t)} d _ {i} ^ {(t)} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} ^ {(t)}),", + "image_path": "0d4a7d0b56e94722ce92489e5e380a11ea9bf9f45ce9c2f057bdec9127e6a385.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "spans": [ + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "\\alpha_{i}^{(t)} = \\sigma_{i}^{(t)}\\exp -\\frac{1}{2} (\\mathbf{p} - \\mu_{i}^{(t)})^{\\top}\\Sigma_{i}^{-1}(\\mathbf{p} - \\mu_{i}^{(t)})" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " defines the 2D density of Gaussian points during the splatting process, " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "C^{(t)}" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "D^{(t)}" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " represent the rendered RGB image and depth, respectively, " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " is the number of Gaussians per tile, " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "d_{i}^{(t)}" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": "-value obtained by transforming coordinate " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " from world to camera coordinates, and " + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 312, + 186, + 533, + 310 + ], + "type": "text", + "content": " denotes the covariance matrix derived from the Gaussian parameters' rotation and scale. Our Generalizable Gaussian Regressor and 3D representation are optimized to reconstruct RGB and depth images from a single view by minimizing a 3D loss function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 314, + 533, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 314, + 533, + 338 + ], + "spans": [ + { + "bbox": [ + 318, + 314, + 533, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {3 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t)} \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "9056a961bf162a31bac2ee9bea08c7937e3dbb23b765c3c7fd0e50d7c6af6d51.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "spans": [ + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{*(t)}" + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{*(t)}" + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "text", + "content": " represent the ground truth color and depth at time step " + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 312, + 344, + 533, + 389 + ], + "type": "text", + "content": ". This 3D loss allows the 3D representation to emphasize local visual features while incorporating global structural knowledge and texture information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 395, + 463, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 463, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 463, + 406 + ], + "type": "text", + "content": "3.3. 4D Spatiotemporal Awareness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "content": "To emphasize 4D spatiotemporal awareness in the diffusion policy, DP4 integrates dynamics into the Gaussian world model described in Section 3.2. This model parameterizes the Gaussian mixture distribution for dynamic Gaussian splatting, enabling future scene reconstruction via parameter propagation. Thus, the model gains valuable supervision in interactive environments by ensuring consistency between reconstructed and real feature scenes. Specifically, the model learns environmental dynamics by predicting the future state " + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "inline_equation", + "content": "s^{(t + 1)}" + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "content": " from the current state " + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "inline_equation", + "content": "s^{(t)}" + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "content": " and action " + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "inline_equation", + "content": "a^{(t)}" + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 312, + 411, + 533, + 531 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "text", + "content": "In the dynamic Gaussian world model, we define the current state as the visual observation at the current time step, with actions representing the robot components. These observations and actions predict visual scenes for the next time step, representing the future state. Building on the Gaussian world model, DP4 uses a deformable MLP " + }, + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "text", + "content": " to estimate changes in Gaussian parameters during time propagation and a Gaussian renderer " + }, + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 312, + 531, + 533, + 630 + ], + "type": "text", + "content": " to generate RGB and depth images for the future state:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 635, + 440, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 635, + 440, + 654 + ], + "spans": [ + { + "bbox": [ + 347, + 635, + 440, + 654 + ], + "type": "interline_equation", + "content": "\\Delta \\theta^ {(t)} = p _ {\\phi} \\left(\\theta^ {(t)}, a ^ {(t)}\\right),", + "image_path": "36d1cefb10bd5359eda77c5c36553dd1825591b160da637d305a80ed713f1b1e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 656, + 533, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 656, + 533, + 668 + ], + "spans": [ + { + "bbox": [ + 347, + 656, + 533, + 668 + ], + "type": "interline_equation", + "content": "\\theta^ {(t + 1)} = \\theta^ {(t)} + \\Delta \\theta^ {(t)}, \\tag {3}", + "image_path": "e379b9e676918318b92be445209c721735d74d8939a488f80763f38f55f42ddb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 671, + 500, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 671, + 500, + 689 + ], + "spans": [ + { + "bbox": [ + 347, + 671, + 500, + 689 + ], + "type": "interline_equation", + "content": "C ^ {(t + 1)} (\\mathbf {p}), D ^ {(t + 1)} (\\mathbf {p}) = \\mathcal {R} \\left(\\theta^ {(t + 1)}, \\mathbf {p}\\right),", + "image_path": "8093441a9cb1b8bfca8b27ecddd024e560d58789fffa023abd62ee6218d4a495.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7125" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "spans": [ + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(t + 1)}(\\mathbf{p})" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{(t + 1)}(\\mathbf{p})" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": " represent the predicted RGB and depth images at time step " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "\\theta^{(t)}" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": " represents the 3D Gaussian primitive parameters at time step " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": " denotes the camera pose used to project the Gaussian primitives, and " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": " is the rendering process from Equation 1. DP4 employs multi-head neural networks as the Gaussian regressor, with each head predicting distinct features of the Gaussian parameters. By estimating changes in position and rotation between time steps, we derive the Gaussian parameters for the future step. The Gaussian renderer then projects the propagated Gaussian distribution onto the single view to reconstruct the future scene. To model 4D spatiotemporal relationships in the physical world, we enforce consistency supervision between reconstructed and real scenes at time step " + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 75, + 98, + 298, + 307 + ], + "type": "text", + "content": ", embedding scene-level spatiotemporal dynamics into the Gaussian parameters. Specifically, the training objective aligns predicted future scenes, derived from current observations and actions, with real scenes, as formulated below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 312, + 296, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 312, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 77, + 312, + 296, + 343 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {4 D}} = \\sum_ {\\mathbf {p}} \\| \\mathbf {C} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {C} ^ {* (t + 1)} \\| _ {2} ^ {2} + \\| \\mathbf {D} ^ {(t + 1)} (\\mathbf {p}) - \\mathbf {D} ^ {* (t + 1)} \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "02cd6373a58f083f2d44ac3cc51da08630335e8f08c9c096e718d04837c7727d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^{(t + 1)}" + }, + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^{(t + 1)}" + }, + { + "bbox": [ + 75, + 342, + 296, + 421 + ], + "type": "text", + "content": " represent the real future counterparts. This additional task of predicting future scenes from current observations via the dynamic Gaussian world model serves as a spatiotemporal enhancement penalty. Training with this task, the multi-level 3D representation learns to encode physical properties and develop awareness of 4D spatiotemporal dynamics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 426, + 204, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 426, + 204, + 437 + ], + "spans": [ + { + "bbox": [ + 76, + 426, + 204, + 437 + ], + "type": "text", + "content": "3.4. Diffusion-based Decision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": "The multi-level 3D representation, including local and global 3D representations, is optimized to predict the robot's optimal actions. We model this process as a conditional denoising diffusion framework, conditioned on the multi-level 3D representation " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": " and the robot state " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": ", which converts randomly sampled Gaussian noise into actions " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": ". Starting with sampled Gaussian noise " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "a^K" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": ", the denoising network " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": " performs " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": " iterations to progressively convert it into the noise-free action " + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "inline_equation", + "content": "a^0" + }, + { + "bbox": [ + 75, + 442, + 297, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 548, + 296, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 548, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 86, + 548, + 296, + 563 + ], + "type": "interline_equation", + "content": "a ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, \\mathrm {I}), \\tag {5}", + "image_path": "b7b8a3852a656f3029aa3c24feaf8446ac8806177fe7196190a0765cb8148a1b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": " represents the sampled Gaussian noise, and " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\alpha_{k},\\gamma_{k}" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": " are functions of " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": " defined by the noise scheduler. To train the denoising network " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": ", we randomly sample a data point " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "a^0" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": " from the dataset and use a diffusion process to generate the noise " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\epsilon^k" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 75, + 570, + 296, + 637 + ], + "type": "text", + "content": ". The training objective is to predict the noise added to the data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 643, + 296, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 643, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 99, + 643, + 296, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a c t i o n}} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha} _ {k} a ^ {0} + \\bar {\\beta} _ {k} \\epsilon^ {k}, k, r, q\\right)\\right), \\tag {6}", + "image_path": "02ebebcc2588d216cfb52e6da06d1bd91ee14e22e60dedd38e74ffa4cd31f470.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_k" + }, + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\bar{\\beta}_k" + }, + { + "bbox": [ + 75, + 664, + 296, + 688 + ], + "type": "text", + "content": " are noise schedule that performs one step noise adding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 99, + 518, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 99, + 518, + 110 + ], + "spans": [ + { + "bbox": [ + 324, + 99, + 518, + 110 + ], + "type": "text", + "content": "The overall learning objective for DP4 is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 118, + 533, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 118, + 533, + 130 + ], + "spans": [ + { + "bbox": [ + 354, + 118, + 533, + 130 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D P 4}} = \\mathcal {L} _ {\\text {a c t i o n}} + \\lambda_ {\\mathrm {3 D}} \\mathcal {L} _ {\\mathrm {3 D}} + \\lambda_ {\\mathrm {4 D}} \\mathcal {L} _ {\\mathrm {4 D}}, \\tag {7}", + "image_path": "e25626062dca39e6eda103d7f477469d69c9eff1f8445ac28d0b721acf7c1f89.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "inline_equation", + "content": "\\lambda_{3\\mathrm{D}}" + }, + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "inline_equation", + "content": "\\lambda_{4\\mathrm{D}}" + }, + { + "bbox": [ + 312, + 138, + 534, + 281 + ], + "type": "text", + "content": " balance the 3D loss for current scenes and the 4D loss for future scenes across different objectives. We train DP4 using a joint training strategy. Empirical observations show that this strategy enhances information fusion while learning shared features. During training, we introduce a warm-up phase, freezing the deformable mapper for the first 500 iterations to establish a stable multi-level 3D compact representation and Gaussian regressor. After the warm-up phase, we simultaneously train the entire Gaussian world model and the diffusion-based decision module. After optimization, DP4 generates optimal actions with spatiotemporal awareness, including both 3D spatial and 4D spatiotemporal perceptions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 291, + 388, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 388, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 388, + 303 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 309, + 412, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 412, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 412, + 321 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 325, + 534, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 325, + 534, + 424 + ], + "spans": [ + { + "bbox": [ + 312, + 325, + 534, + 424 + ], + "type": "text", + "content": "Simulations For simulation experiments, we gathered 22 tasks across various domains, covering a wide range of robotic skills. These tasks span challenging scenarios, such as bimanual, deformable object, and articulated object manipulation, as well as simpler tasks like parallel gripper manipulation. The tasks include Adroit tasks [30] (Door, Pen, Hammer), DexArt tasks [2] (Laptop, Faucet, Bucket, Toilet), and RLBench tasks [17] (10 challenging tasks, each with at least two variations, totaling 166)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 424, + 534, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 424, + 534, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 424, + 534, + 544 + ], + "type": "text", + "content": "Real robot We used a KINOVA GEN2 robot with a Realsense D455 depth camera mounted in an eye-to-hand configuration. In an indoor environment, we arranged several bottles and cups to enable the agent to generalize manipulation skills across scenes, designing three tasks, including Grasping bottles, Pouring the water from the bottle into the cup and stacking cups. The RealSense D455 camera records the entire scene and robot state from a third-person perspective. This setup supplies RGB-D observations for policy training and RGB and depth supervision for constructing the dynamic Gaussian world model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 544, + 533, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 533, + 621 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 533, + 621 + ], + "type": "text", + "content": "Expert demonstration Expert trajectories for ManiSkill and Adroit are collected using RL-trained agents, using VRL3 [37] for Adroit and PPO [31] for other domains. We generate successful trajectories with RL agents and ensure that all visual imitation learning algorithms use identical demonstrations. Additionally, we collect 20 demonstrations for each RLBench task using the motion planner." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 621, + 534, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 621, + 534, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 621, + 534, + 687 + ], + "type": "text", + "content": "Baselines The main objective of DP4 is to highlight the importance of 3D spatial and 4D spatiotemporal awareness in diffusion policies. To this end, our primary baseline is the point cloud-based 3D Diffusion Policy (DP3) [47]. Additionally, we incorporate comparisons with other visual imitation learning methods: IBC [7], BCRNN [26], Diffusion" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7126" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 99, + 295, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 99, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 75, + 99, + 295, + 142 + ], + "type": "text", + "content": "Policy (DP) [5], PerAct [35], GNFactor [46], and Mani-Gaussian [25]. Note that image and depth resolutions for all 2D and 3D methods remain consistent across experiments, ensuring fair comparisons." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 144, + 295, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 144, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 75, + 144, + 295, + 219 + ], + "type": "text", + "content": "Evaluation metric Each experiment is conducted using three seeds. For Adroit and DexArt simulation tasks, we assess 20 episodes every 200 epochs and calculate the mean of the top five success rates. For RLBench tasks, we evaluate 25 episodes per task at the final checkpoint for 10 challenging tasks. We report the mean and standard deviation of success rates across the three seeds." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "text", + "content": "Implementation details We implement a convolutional network-based diffusion policy using DDIM[36] as the noise scheduler, preferring sample prediction to enhance high-dimensional action generation. Training employs 100 timesteps, while inference uses 10. Models are trained for 3,000 epochs with a batch size of 32 on both simulated and real-world tasks. Loss hyperparameters " + }, + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{3D}}" + }, + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{4D}}" + }, + { + "bbox": [ + 75, + 221, + 295, + 374 + ], + "type": "text", + "content": " are set to 0.1 and 0.01, respectively (see Section 4.4 for details). For fair comparison with DP3[47], we adopt the same 3D Local Encoder: a three-layer MLP, max-pooling for order-equivariant aggregation, and a projection head generating compact 3D local features, with LayerNorm inserted for training stability. All experiments run on an NVIDIA H100 80GB GPU and a 192-vCPU Intel Xeon Platinum 8468." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 383, + 206, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 383, + 206, + 395 + ], + "spans": [ + { + "bbox": [ + 76, + 383, + 206, + 395 + ], + "type": "text", + "content": "4.2. Quantitative Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 400, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 400, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 75, + 400, + 295, + 422 + ], + "type": "text", + "content": "We compare our DP4 with previous state-of-the-art methods on the Adroit, DexArt, and RLBench simulation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 422, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 422, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 75, + 422, + 295, + 510 + ], + "type": "text", + "content": "Adroit simulation Table 1 presents the success rates for the Adroit simulation tasks. DP4 achieves an overall success rate of " + }, + { + "bbox": [ + 75, + 422, + 295, + 510 + ], + "type": "inline_equation", + "content": "84.7\\%" + }, + { + "bbox": [ + 75, + 422, + 295, + 510 + ], + "type": "text", + "content": ", significantly outperforming 2D-based methods, including IBC [7], BCRNN [26], and Diffusion Policy [5], while surpassing 3D-based methods such as 3D Diffusion Policy [47]. Experimental results demonstrate that DP4 completes complex tasks more accurately by leveraging spatiotemporal perception." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "text", + "content": "DexArt simulation Table 2 presents the success rates for various DexArt simulation tasks. The DP4 achieves a success rate of " + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "inline_equation", + "content": "82.5\\%" + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "text", + "content": ", significantly outperforming 2D-based methods (e.g., IBC, BCRNN, and Diffusion Policy) by " + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "inline_equation", + "content": "33.5\\%" + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "text", + "content": " and 3D-based methods (e.g., 3D Diffusion Policy) by " + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "inline_equation", + "content": "14.0\\%" + }, + { + "bbox": [ + 75, + 511, + 295, + 609 + ], + "type": "text", + "content": ". Experimental results demonstrate that DP4 integrates 3D spatial and 4D spatiotemporal perception of the physical world with diffusion policies, achieving higher success rates across multiple tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "text", + "content": "RLBench simulation We compare DP4 with previous state-of-the-art methods on the RLBench task suite. Table 3 presents a comparison of average success rates across tasks. DP4 achieves state-of-the-art performance with average success rates of " + }, + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "inline_equation", + "content": "63.3\\%" + }, + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "inline_equation", + "content": "39.9\\%" + }, + { + "bbox": [ + 75, + 610, + 295, + 687 + ], + "type": "text", + "content": ", significantly outperforming prior methods, including both perception-based and generative approaches. Leveraging supervision from" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 97, + 533, + 169 + ], + "blocks": [ + { + "bbox": [ + 315, + 97, + 533, + 169 + ], + "lines": [ + { + "bbox": [ + 315, + 97, + 533, + 169 + ], + "spans": [ + { + "bbox": [ + 315, + 97, + 533, + 169 + ], + "type": "table", + "html": "
MethodHammerDoorPenOverall
IBC [7]0±00±09±23.0
BCRNN [26]0±00±09±33.0
DP [5]48±1750±525±441.0
DP3 [47]100±062±443±668.3
DP4 (ours)100±080±275±384.7
", + "image_path": "26ec84e32d91961f541de5582ca89d0229f91c577c06d4b12b2cb72a27b37f64.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 220, + 531, + 290 + ], + "blocks": [ + { + "bbox": [ + 312, + 171, + 533, + 212 + ], + "lines": [ + { + "bbox": [ + 312, + 171, + 533, + 212 + ], + "spans": [ + { + "bbox": [ + 312, + 171, + 533, + 212 + ], + "type": "text", + "content": "Table 1. Success rates " + }, + { + "bbox": [ + 312, + 171, + 533, + 212 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 312, + 171, + 533, + 212 + ], + "type": "text", + "content": " on the Adroit simulation. We compare our DP4 with more baselines in the Adroit simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 220, + 531, + 290 + ], + "lines": [ + { + "bbox": [ + 315, + 220, + 531, + 290 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 531, + 290 + ], + "type": "table", + "html": "
MethodLaptopFaucetBucketToiletOverall
IBC [7]3±27±114±10±06.0
BCRNN [26]3±31±05±50±02.3
DP [5]69±423±858±246±149.0
DP3 [47]83±163±282±446±268.5
DP4 (ours)92±284±190±164±282.5
", + "image_path": "41d8fe96f2ff47889d4244ee0f6dafaeb9468f7ae44711f4bd48176b45cf5d05.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 313, + 338, + 536, + 448 + ], + "blocks": [ + { + "bbox": [ + 312, + 293, + 533, + 334 + ], + "lines": [ + { + "bbox": [ + 312, + 293, + 533, + 334 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 533, + 334 + ], + "type": "text", + "content": "Table 2. Success rates " + }, + { + "bbox": [ + 312, + 293, + 533, + 334 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 312, + 293, + 533, + 334 + ], + "type": "text", + "content": " on the DexArt simulation. We compare our DP4 with more baselines in the DexArt simulation. The baselines include Implicit Behavioral Cloning (IBC), BCRNN, Diffusion Policy (DP), and 3D Diffusion Policy (DP3)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 338, + 536, + 448 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 536, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 536, + 448 + ], + "type": "table", + "html": "
Methodclose jaropen drawersweep to dustpanmeat off grillturn tapOverall
PerAct [35]18.7±8.254.7±18.60.0±0.040.0±17.038.7±6.830.4
GNFactor [46]25.3±6.876.0±5.728.0±15.057.3±18.950.7±8.247.5
ManiGaussian [25]28.4±5.476.3±6.264.5±13.460.2±18.256.2±6.857.1
DP4 (ours)35.4±2.282.2±3.472.0±5.264.4±6.862.4±2.863.3
Methodslide blockput in drawerdrag stickpush buttonsstack blocksOverall
PerAct [35]18.7±13.62.7±3.35.3±5.018.7±12.46.7±1.910.4
GNFactor [46]20.0±15.00.0±0.037.3±13.218.7±10.04.0±3.316.0
ManiGaussian [25]24.3±12.816.6±3.292.3±11.420.4±12.212.4±3.233.2
DP4 (ours)33.4±4.422.8±1.294.4±6.228.3±6.821.0±2.239.9
", + "image_path": "61712f5da278a9047efda638a268e88fdc1d571b439824301cb618e6feb76a68.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 449, + 533, + 489 + ], + "lines": [ + { + "bbox": [ + 312, + 449, + 533, + 489 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 533, + 489 + ], + "type": "text", + "content": "Table 3. Success rates " + }, + { + "bbox": [ + 312, + 449, + 533, + 489 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 312, + 449, + 533, + 489 + ], + "type": "text", + "content": " on the RLBench simulation. We compare our DP4 with more baselines in the RLBench simulation. Baseline approaches include PerAct, GNFactor, and Mani-Gaussian." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 498, + 533, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 533, + 530 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 533, + 530 + ], + "type": "text", + "content": "3D geometry, texture, and 4D spatiotemporal data, DP4 accurately predicts optimal continuous trajectories from a single RGB-D observation of the current scene." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 538, + 437, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 538, + 437, + 550 + ], + "spans": [ + { + "bbox": [ + 312, + 538, + 437, + 550 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 555, + 533, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 533, + 654 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 533, + 654 + ], + "type": "text", + "content": "Visualization of trajectories We present three qualitative examples of action sequences generated by DP3 and our DP4 in Figure 3. The results show that the previous agent (DP3) frequently fails to complete tasks, as it imitates the expert's motions (e.g., hammering nails, sliding doors, and spinning pens) without fully understanding the scene. In contrast, DP4 significantly improves the success rate by accurately interpreting 3D scene structures (e.g., nail positions) and 4D dynamics (e.g., pen rotation)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 654, + 533, + 687 + ], + "type": "text", + "content": "Visualization of Gaussian world model Figure 4 and Figure 5 show the visualization results. We visualize the Gaussian world model via single-view rendering, observ" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7127" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 98, + 294, + 195 + ], + "blocks": [ + { + "bbox": [ + 78, + 98, + 294, + 195 + ], + "lines": [ + { + "bbox": [ + 78, + 98, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 78, + 98, + 294, + 195 + ], + "type": "image", + "image_path": "33406afdd17829ceaf93beed74cf9ca0eb168bce9c918ed14d8bcbe980ba0604.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 198, + 293, + 296 + ], + "blocks": [ + { + "bbox": [ + 78, + 198, + 293, + 296 + ], + "lines": [ + { + "bbox": [ + 78, + 198, + 293, + 296 + ], + "spans": [ + { + "bbox": [ + 78, + 198, + 293, + 296 + ], + "type": "image", + "image_path": "55abc5184a14d7a6e1bf3652b917f8b8eac57d2eec1b503c782c03e053d070df.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 243, + 533, + 284 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 533, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 533, + 284 + ], + "type": "text", + "content": "Figure 4. Visualization of Gaussian world model. After training DP4, we visualize its Gaussian world model by rendering a single-view observation. DP4 uses a Gaussian world model to provide supervision, facilitating the learning of 3D spatial perception." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 79, + 300, + 293, + 397 + ], + "blocks": [ + { + "bbox": [ + 79, + 300, + 293, + 397 + ], + "lines": [ + { + "bbox": [ + 79, + 300, + 293, + 397 + ], + "spans": [ + { + "bbox": [ + 79, + 300, + 293, + 397 + ], + "type": "image", + "image_path": "6b0358591e8cc5afbac5a33bbb1b82c221d0c714080cb4774ccf1f026f580718.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 403, + 297, + 464 + ], + "lines": [ + { + "bbox": [ + 75, + 403, + 297, + 464 + ], + "spans": [ + { + "bbox": [ + 75, + 403, + 297, + 464 + ], + "type": "text", + "content": "Figure 3. Qualitative case study. The red mark indicates a pose that significantly deviates from the expert demonstration, while the green mark denotes a pose that aligns with the expert trajectory. The 4D Diffusion Policy (DP4) integrates 3D spatial and 4D spatiotemporal awareness with diffusion policies, successfully completing the tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 474, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 474, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 75, + 474, + 296, + 552 + ], + "type": "text", + "content": "ing it captures fundamental spatial structures and spatiotemporal dynamics. Although DP4 operates under a realistic single-view constraint, resulting renderings lack fine detail, they effectively represent scene-level 3D structures and 4D dynamics. Crucially, such details have minimal impact on task success; for example, the nail's thread is irrelevant to the hammering task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 558, + 162, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 558, + 162, + 570 + ], + "spans": [ + { + "bbox": [ + 76, + 558, + 162, + 570 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 574, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 574, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 75, + 574, + 296, + 673 + ], + "type": "text", + "content": "Component ablation. Ablations are conducted on the Adroit simulation tasks in Table 4. Without the proposed components, we use 3D representation without any 3D spatial supervision (RGB and depth (structure)) and 4D spatiotemporal supervision (dynamics). In the experiment on the Adroit simulation, the performance of the \"Hammer\", \"Door\", and \"Pen\" tasks demonstrated improvements in success rates with the proposed components, while the inference time remained stable or only slightly increased." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 676, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 676, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 76, + 676, + 296, + 686 + ], + "type": "text", + "content": "- For the Hammer task: The base model started with a" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 99, + 531, + 236 + ], + "blocks": [ + { + "bbox": [ + 317, + 99, + 531, + 236 + ], + "lines": [ + { + "bbox": [ + 317, + 99, + 531, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 99, + 531, + 236 + ], + "type": "image", + "image_path": "883f8590c4d5270f6a24269dc21164be532fed6fe62a7ed559c2bbe3f62c3aed.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "text", + "content": "94.0% success rate and a completion time of 6.40 seconds. After adding the Gaussian world model with RGB supervision, success improved to " + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "inline_equation", + "content": "96.0\\%" + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "text", + "content": ", with a minimal increase in time (6.52 seconds). Introducing 3D spatial loss (RGB and Depth supervision) raised success to " + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "inline_equation", + "content": "95.0\\%" + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "text", + "content": ", with stable time (6.53 seconds). With 4D spatiotemporal supervision (DP4), success reached " + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "inline_equation", + "content": "98.0\\%" + }, + { + "bbox": [ + 320, + 297, + 533, + 395 + ], + "type": "text", + "content": ", and the time increased slightly to 6.56 seconds, indicating a clear performance gain with a small time increase." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 533, + 604 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "content": "- For the Door task: Initially, the model achieved a " + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "inline_equation", + "content": "94.0\\%" + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "content": " success rate and a completion time of 6.40 seconds. Adding the Gaussian world model with RGB supervision improved success to " + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "inline_equation", + "content": "96.0\\%" + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "content": ", with no significant time change. Incorporating 3D spatial loss (RGB and Depth) increased success to " + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "inline_equation", + "content": "98.0\\%" + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "content": ", with a slight increase in time (6.56 seconds). With 4D spatiotemporal supervision (DP4), the success rate reached " + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "inline_equation", + "content": "100.0\\%" + }, + { + "bbox": [ + 313, + 396, + 533, + 505 + ], + "type": "text", + "content": ", with a minimal time increase (6.57 seconds), reflecting significant improvement from additional supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "content": "- For the Pen task: The base model had a " + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "inline_equation", + "content": "45.0\\%" + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "content": " success rate and a completion time of 6.46 seconds. After adding the Gaussian world model with RGB supervision, success rose to " + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "inline_equation", + "content": "48.0\\%" + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "content": ", with time remaining at 6.48 seconds. Introducing 3D spatial loss (RGB and depth) led to a slight decrease in success " + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "inline_equation", + "content": "(47.0\\%)" + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "content": ". With 4D spatiotemporal supervision (DP4), success jumped to " + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "inline_equation", + "content": "75.0\\%" + }, + { + "bbox": [ + 313, + 505, + 533, + 604 + ], + "type": "text", + "content": ", with a small increase in time to 6.50 seconds, highlighting the effectiveness of 4D supervision." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 609, + 533, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 609, + 533, + 674 + ], + "spans": [ + { + "bbox": [ + 312, + 609, + 533, + 674 + ], + "type": "text", + "content": "Notably, DP4 does not substantially increase model inference time. This is because DP4 does not construct a Gaussian world model during inference. The 3D spatial and 4D spatiotemporal awareness from the Gaussian world model are efficiently integrated into the multi-level 3D representation during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 676, + 533, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 533, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 533, + 687 + ], + "type": "text", + "content": "Hyperparameter analysis We conduct hyperparameter" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7128" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 108, + 121, + 148 + ], + "blocks": [ + { + "bbox": [ + 79, + 108, + 121, + 148 + ], + "lines": [ + { + "bbox": [ + 79, + 108, + 121, + 148 + ], + "spans": [ + { + "bbox": [ + 79, + 108, + 121, + 148 + ], + "type": "image", + "image_path": "6635bc0adfca8423a5256aee603baba3cf156b81b45d6c002b5d7ddd0c35a977.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 153, + 296, + 183 + ], + "lines": [ + { + "bbox": [ + 76, + 153, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 76, + 153, + 296, + 183 + ], + "type": "text", + "content": "Figure 5. More visualizations of the learned world model. DP4 uses a Gaussian world model to provide supervision, facilitating the 4D spatiotemporal perceptions in interactive environments." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 124, + 97, + 165, + 148 + ], + "blocks": [ + { + "bbox": [ + 78, + 95, + 168, + 107 + ], + "lines": [ + { + "bbox": [ + 78, + 95, + 168, + 107 + ], + "spans": [ + { + "bbox": [ + 78, + 95, + 168, + 107 + ], + "type": "text", + "content": "Ground-truth Prediction at current time step T at current time step T" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 124, + 97, + 165, + 148 + ], + "lines": [ + { + "bbox": [ + 124, + 97, + 165, + 148 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 165, + 148 + ], + "type": "image", + "image_path": "3f687a09c42ab7486e7610126eba30686b9c1537cebb78bcc95dccbdb2714b3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 108, + 207, + 148 + ], + "blocks": [ + { + "bbox": [ + 168, + 108, + 207, + 148 + ], + "lines": [ + { + "bbox": [ + 168, + 108, + 207, + 148 + ], + "spans": [ + { + "bbox": [ + 168, + 108, + 207, + 148 + ], + "type": "image", + "image_path": "453b535e30b31baa9c589d11f4efc24c74d5502bd15fbd7dfb24f4a742049018.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 210, + 98, + 250, + 148 + ], + "blocks": [ + { + "bbox": [ + 210, + 98, + 250, + 148 + ], + "lines": [ + { + "bbox": [ + 210, + 98, + 250, + 148 + ], + "spans": [ + { + "bbox": [ + 210, + 98, + 250, + 148 + ], + "type": "image", + "image_path": "cec220ca16b1cf6c315ae503fa2472c3ceed9e504c40570ab5461ecd4acbcd47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 252, + 108, + 292, + 148 + ], + "blocks": [ + { + "bbox": [ + 252, + 108, + 292, + 148 + ], + "lines": [ + { + "bbox": [ + 252, + 108, + 292, + 148 + ], + "spans": [ + { + "bbox": [ + 252, + 108, + 292, + 148 + ], + "type": "image", + "image_path": "957bce2bec6f7cab2553fa72623cdc8f29ef9b55a7a6696381516a5f70741d24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 78, + 200, + 293, + 290 + ], + "blocks": [ + { + "bbox": [ + 133, + 189, + 237, + 198 + ], + "lines": [ + { + "bbox": [ + 133, + 189, + 237, + 198 + ], + "spans": [ + { + "bbox": [ + 133, + 189, + 237, + 198 + ], + "type": "text", + "content": "Results on Adroit Simulation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 200, + 293, + 290 + ], + "lines": [ + { + "bbox": [ + 78, + 200, + 293, + 290 + ], + "spans": [ + { + "bbox": [ + 78, + 200, + 293, + 290 + ], + "type": "table", + "html": "
3D Component4D ComponentPerformance
RGBDepthDynamicsHammerDoorPen
94.0/6.4064.0/6.4445.0/6.46
✓96.0/6.5268.0/6.4548.0/6.48
✓95.0/6.5368.0/6.4647.0/6.48
✓✓98.0/6.5675.0/6.5172.0/6.47
✓✓✓100.0/6.5780.0/6.4575.0/6.50
", + "image_path": "03505aa83777b4f0a93c49bad4022764651375bdf477909cfe08ccf3e59fef4a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 78, + 334, + 187, + 423 + ], + "blocks": [ + { + "bbox": [ + 76, + 293, + 296, + 334 + ], + "lines": [ + { + "bbox": [ + 76, + 293, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 76, + 293, + 296, + 334 + ], + "type": "text", + "content": "Table 4. Component ablation studies. We report results for the Adroit simulation tasks. In the performance metrics, the number to the left of / represents the success rate " + }, + { + "bbox": [ + 76, + 293, + 296, + 334 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 76, + 293, + 296, + 334 + ], + "type": "text", + "content": ", and the number to the right denotes the inference speed (seconds per task)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 334, + 187, + 423 + ], + "lines": [ + { + "bbox": [ + 78, + 334, + 187, + 423 + ], + "spans": [ + { + "bbox": [ + 78, + 334, + 187, + 423 + ], + "type": "image", + "image_path": "605edf3e11de05e6e61990f0b495689c2878ef6deb6e91c7e3de02c5dda5d763.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "lines": [ + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "type": "text", + "content": "Figure 6. Hyperparameter analysis of DP4. The effects of " + }, + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{3\\mathrm{D}}" + }, + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\lambda_{4\\mathrm{D}}" + }, + { + "bbox": [ + 76, + 423, + 295, + 443 + ], + "type": "text", + "content": " on success rates, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 187, + 334, + 294, + 423 + ], + "blocks": [ + { + "bbox": [ + 187, + 334, + 294, + 423 + ], + "lines": [ + { + "bbox": [ + 187, + 334, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 187, + 334, + 294, + 423 + ], + "type": "image", + "image_path": "c23d20c7befb132e4a8f5b7d43aff860fc7336241d097142d17882a6d71e23f8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 78, + 446, + 295, + 508 + ], + "blocks": [ + { + "bbox": [ + 78, + 446, + 295, + 508 + ], + "lines": [ + { + "bbox": [ + 78, + 446, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 78, + 446, + 295, + 508 + ], + "type": "table", + "html": "
MethodGrasping BottlesStacking CupsPouring WaterOverall
DP [5]36.044.028.036.0
DP3 [47]42.062.034.046.0
DP4 (ours)48.072.044.054.6
", + "image_path": "89cf71df164e91b6976c3076bf1790ee84f4910a3d2405a6d105968b4206dd1a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 75, + 510, + 296, + 541 + ], + "lines": [ + { + "bbox": [ + 75, + 510, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 75, + 510, + 296, + 541 + ], + "type": "text", + "content": "Table 5. Success rates " + }, + { + "bbox": [ + 75, + 510, + 296, + 541 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 75, + 510, + 296, + 541 + ], + "type": "text", + "content": " on the real-robot tasks. We compare our DP4 with additional baselines on the real-robot tasks. We include the Diffusion Policy (DP) and 3D Diffusion Policy (DP3)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": "analysis on the Pen task within the Adroit simulation. DP4 uses two key hyperparameters, " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{3\\mathrm{D}}" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{4\\mathrm{D}}" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": ", to balance the 3D spatial loss for current scenes and the 4D spatiotemporal loss for future scenes across multiple objectives. As shown in Figure 6, we analyze the hyperparameters of DP4. We vary " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{3\\mathrm{D}}" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": " from 0 to 1 and " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{4\\mathrm{D}}" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": " from 0 to 0.1, measuring the corresponding success rates. We find that the success rate peaks when " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{3\\mathrm{D}} = 0.1" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{4\\mathrm{D}} = 0.01" + }, + { + "bbox": [ + 75, + 543, + 296, + 643 + ], + "type": "text", + "content": ", so we adopt these values as the hyperparameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 76, + 649, + 186, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 649, + 186, + 661 + ], + "spans": [ + { + "bbox": [ + 76, + 649, + 186, + 661 + ], + "type": "text", + "content": "4.5. Real-Robot Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 75, + 665, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 665, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 75, + 665, + 296, + 688 + ], + "type": "text", + "content": "To assess DP4's performance on real-world robotic tasks, we designed three tasks: Grasping Bottles, Stacking Cups," + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 320, + 99, + 388, + 152 + ], + "blocks": [ + { + "bbox": [ + 320, + 99, + 388, + 152 + ], + "lines": [ + { + "bbox": [ + 320, + 99, + 388, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 99, + 388, + 152 + ], + "type": "image", + "image_path": "d0006c5f3ef97680ed3f00549108ce39aea3b25706ca016c58abd799e9c5a29e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 391, + 99, + 457, + 152 + ], + "blocks": [ + { + "bbox": [ + 391, + 99, + 457, + 152 + ], + "lines": [ + { + "bbox": [ + 391, + 99, + 457, + 152 + ], + "spans": [ + { + "bbox": [ + 391, + 99, + 457, + 152 + ], + "type": "image", + "image_path": "98f0d77bd8b174e1eba639289184742a9e61457dca5d4b8dcf066eb07ef8ae1d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 153, + 449, + 160 + ], + "lines": [ + { + "bbox": [ + 399, + 153, + 449, + 160 + ], + "spans": [ + { + "bbox": [ + 399, + 153, + 449, + 160 + ], + "type": "text", + "content": "\"Grasping Bottles\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 460, + 100, + 526, + 152 + ], + "blocks": [ + { + "bbox": [ + 460, + 100, + 526, + 152 + ], + "lines": [ + { + "bbox": [ + 460, + 100, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 460, + 100, + 526, + 152 + ], + "type": "image", + "image_path": "84f69a1a1bf6d24fed2441dfb88af1390ac474edccf3bf87883ccf0c8af9e3d6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 321, + 167, + 388, + 220 + ], + "blocks": [ + { + "bbox": [ + 321, + 167, + 388, + 220 + ], + "lines": [ + { + "bbox": [ + 321, + 167, + 388, + 220 + ], + "spans": [ + { + "bbox": [ + 321, + 167, + 388, + 220 + ], + "type": "image", + "image_path": "cb06977a3c8925dfd7df526a5a5e029eecbe37dd617e96e06055036b4a6d5949.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 391, + 167, + 456, + 220 + ], + "blocks": [ + { + "bbox": [ + 391, + 167, + 456, + 220 + ], + "lines": [ + { + "bbox": [ + 391, + 167, + 456, + 220 + ], + "spans": [ + { + "bbox": [ + 391, + 167, + 456, + 220 + ], + "type": "image", + "image_path": "dd1e93b1926e181abac301acf7640fa77227c481de793693ab633cc4dd3e2c81.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 221, + 446, + 229 + ], + "lines": [ + { + "bbox": [ + 403, + 221, + 446, + 229 + ], + "spans": [ + { + "bbox": [ + 403, + 221, + 446, + 229 + ], + "type": "text", + "content": "\"Stacking Cups\"" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 460, + 167, + 525, + 220 + ], + "blocks": [ + { + "bbox": [ + 460, + 167, + 525, + 220 + ], + "lines": [ + { + "bbox": [ + 460, + 167, + 525, + 220 + ], + "spans": [ + { + "bbox": [ + 460, + 167, + 525, + 220 + ], + "type": "image", + "image_path": "b1e9fa26f7d2e318e2915001e96357d37d0444f893c3d4441254bc6ec8586319.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 321, + 235, + 388, + 287 + ], + "blocks": [ + { + "bbox": [ + 321, + 235, + 388, + 287 + ], + "lines": [ + { + "bbox": [ + 321, + 235, + 388, + 287 + ], + "spans": [ + { + "bbox": [ + 321, + 235, + 388, + 287 + ], + "type": "image", + "image_path": "36e04d4ea11b207f564179f00f38c151034526724033846e24e2dcb27ddc58bf.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 304, + 533, + 345 + ], + "lines": [ + { + "bbox": [ + 312, + 304, + 533, + 345 + ], + "spans": [ + { + "bbox": [ + 312, + 304, + 533, + 345 + ], + "type": "text", + "content": "Figure 7. Visualization of DP4 performance on three real-world robotic tasks. DP4 demonstrates strong performance in real-world settings and effectively handles a variety of common tasks with a single view." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 391, + 235, + 456, + 287 + ], + "blocks": [ + { + "bbox": [ + 391, + 235, + 456, + 287 + ], + "lines": [ + { + "bbox": [ + 391, + 235, + 456, + 287 + ], + "spans": [ + { + "bbox": [ + 391, + 235, + 456, + 287 + ], + "type": "image", + "image_path": "ab9c019b907ac951141fc7fe64dd502d14a44dda55550f9defa5f81f13d05780.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 289, + 446, + 297 + ], + "lines": [ + { + "bbox": [ + 403, + 289, + 446, + 297 + ], + "spans": [ + { + "bbox": [ + 403, + 289, + 446, + 297 + ], + "type": "text", + "content": "\"Pouring Water\"" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 460, + 235, + 525, + 287 + ], + "blocks": [ + { + "bbox": [ + 460, + 235, + 525, + 287 + ], + "lines": [ + { + "bbox": [ + 460, + 235, + 525, + 287 + ], + "spans": [ + { + "bbox": [ + 460, + 235, + 525, + 287 + ], + "type": "image", + "image_path": "7f3e10bc985da41cf79c0927b4c6aa5a00f61487a1dc178e95246ce839fabd03.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 354, + 533, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 533, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 533, + 475 + ], + "type": "text", + "content": "and Pouring Water. We focus on the model's performance at key frames, where success is defined as when the error is below a certain threshold. The results of the real-world robotic experiments are summarized in Table 5. Experimental results show that DP4 outperforms the Diffusion Policy (DP) and 3D Diffusion Policy (DP3) baselines across all tasks. These results confirm that DP4 effectively integrates spatiotemporal awareness with diffusion policies, improving the accuracy of generated trajectories. Additionally, visualizations of DP4's performance on real-world tasks are shown in Figure 7." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 489, + 380, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 380, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 380, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 510, + 533, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 533, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 533, + 598 + ], + "type": "text", + "content": "This paper presents 4D Diffusion Policy (DP4), which integrates spatiotemporal awareness of the physical world with diffusion-based visual imitation learning. Utilizing a single-view RGB-D input, DP4 constructs a dynamic Gaussian world model that provides rich supervision for acquiring 3D spatial and 4D spatiotemporal perceptions from interactive environments. Extensive experiments on simulated and real-robot tasks validate the effectiveness and superiority." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 612, + 413, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 413, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 413, + 625 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 632, + 533, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 632, + 533, + 688 + ], + "spans": [ + { + "bbox": [ + 312, + 632, + 533, + 688 + ], + "type": "text", + "content": "This work was supported in part by the Science and Technology Commission of Shanghai Municipality (No.24511103100), NSFC Project (62176061), Doubao Fund, and Shanghai Technology Development and Entrepreneurship Platform for Neuromorphic and AI SoC." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 295, + 724, + 315, + 734 + ], + "type": "text", + "content": "7129" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 98, + 130, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 98, + 130, + 109 + ], + "spans": [ + { + "bbox": [ + 77, + 98, + 130, + 109 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 116, + 297, + 686 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 116, + 297, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 116, + 297, + 166 + ], + "spans": [ + { + "bbox": [ + 81, + 116, + 297, + 166 + ], + "type": "text", + "content": "[1] Sridhar Pandian Arunachalam, Sneha Silwal, Ben Evans, and Lerrel Pinto. Dexterous imitation made easy: A learning-based framework for efficient dexterous manipulation. In 2023 IEEE international conference on robotics and automation (icra), pages 5954-5961. IEEE, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 167, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 167, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 81, + 167, + 296, + 217 + ], + "type": "text", + "content": "[2] Chen Bao, Helin Xu, Yuzhe Qin, and Xiaolong Wang. Dexart: Benchmarking generalizable dexterous manipulation with articulated objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21190-21200, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 219, + 296, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 219, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 82, + 219, + 296, + 268 + ], + "type": "text", + "content": "[3] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 270, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 270, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 82, + 270, + 296, + 308 + ], + "type": "text", + "content": "[4] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 310, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 310, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 82, + 310, + 296, + 360 + ], + "type": "text", + "content": "[5] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 361, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 361, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 82, + 361, + 296, + 401 + ], + "type": "text", + "content": "[6] Yuqing Du, Daniel Ho, Alex Alemi, Eric Jang, and Mohi Khansari. Bayesian imitation learning for end-to-end mobile manipulation. In International Conference on Machine Learning, pages 5531-5546. PMLR, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 402, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 402, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 82, + 402, + 296, + 451 + ], + "type": "text", + "content": "[7] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Conference on robot learning, pages 158-168. PMLR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 453, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 453, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 82, + 453, + 296, + 482 + ], + "type": "text", + "content": "[8] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 483, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 483, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 82, + 483, + 296, + 513 + ], + "type": "text", + "content": "[9] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 514, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 514, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 77, + 514, + 296, + 544 + ], + "type": "text", + "content": "[10] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 545, + 296, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 545, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 77, + 545, + 296, + 583 + ], + "type": "text", + "content": "[11] Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems, 35:26091-26104, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 586, + 296, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 586, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 77, + 586, + 296, + 615 + ], + "type": "text", + "content": "[12] Siddhant Haldar, Jyothish Pari, Anant Rai, and Lerrel Pinto. Teach a robot to fish: Versatile imitation from one minute of demonstrations. arXiv preprint arXiv:2303.01497, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 616, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 616, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 77, + 616, + 296, + 666 + ], + "type": "text", + "content": "[13] Nicklas Hansen, Zhecheng Yuan, Yanjie Ze, Tongzhou Mu, Aravind Rajeswaran, Hao Su, Huazhe Xu, and Xiaolong Wang. On pre-training for visuo-motor control: Revisiting a learning-from-scratch baseline. arXiv preprint arXiv:2212.05749, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 667, + 296, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 667, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 77, + 667, + 296, + 686 + ], + "type": "text", + "content": "[14] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 314, + 100, + 533, + 687 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 331, + 100, + 533, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 100, + 533, + 120 + ], + "spans": [ + { + "bbox": [ + 331, + 100, + 533, + 120 + ], + "type": "text", + "content": "fusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 121, + 533, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 121, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 121, + 533, + 171 + ], + "type": "text", + "content": "[15] Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. Advances in Neural Information Processing Systems, 35:20703-20716, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 173, + 533, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 173, + 533, + 221 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 533, + 221 + ], + "type": "text", + "content": "[16] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 224, + 533, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 224, + 533, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 224, + 533, + 263 + ], + "type": "text", + "content": "[17] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J Davison. Rlbench: The robot learning benchmark & learning environment. IEEE Robotics and Automation Letters, 5(2):3019-3026, 2020. 5, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 266, + 533, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 266, + 533, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 533, + 305 + ], + "type": "text", + "content": "[18] Edward Johns. Coarse-to-fine imitation learning: Robot manipulation from a single demonstration. In 2021 IEEE international conference on robotics and automation (ICRA), pages 4613-4619. IEEE, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 307, + 533, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 307, + 533, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 307, + 533, + 346 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 348, + 533, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 348, + 533, + 418 + ], + "spans": [ + { + "bbox": [ + 315, + 348, + 533, + 418 + ], + "type": "text", + "content": "[20] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 420, + 533, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 420, + 533, + 459 + ], + "spans": [ + { + "bbox": [ + 315, + 420, + 533, + 459 + ], + "type": "text", + "content": "[21] Peiyan Li, Hongtao Wu, Yan Huang, Chilam Cheang, Liang Wang, and Tao Kong. Gr-mg: Leveraging partially-annotated data via multi-modal goal-conditioned policy. IEEE Robotics and Automation Letters, 2025." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 460, + 533, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 460, + 533, + 501 + ], + "spans": [ + { + "bbox": [ + 315, + 460, + 533, + 501 + ], + "type": "text", + "content": "[22] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 502, + 533, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 502, + 533, + 563 + ], + "spans": [ + { + "bbox": [ + 315, + 502, + 533, + 563 + ], + "type": "text", + "content": "[23] Zhenyang Liu, Yikai Wang, Sixiao Zheng, Tongying Pan, Longfei Liang, Yanwei Fu, and Xiangyang Xue. Reasongrounder: Lvlm-guided hierarchical feature splatting for open-vocabulary 3d visual grounding and reasoning. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3718-3727, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 564, + 533, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 564, + 533, + 613 + ], + "spans": [ + { + "bbox": [ + 315, + 564, + 533, + 613 + ], + "type": "text", + "content": "[24] Zhenyang Liu, Sixiao Zheng, Siyu Chen, Cairong Zhao, Longfei Liang, Xiangyang Xue, and Yanwei Fu. A neural representation framework with llm-driven spatial reasoning for open-vocabulary 3d visual grounding. arXiv preprint arXiv:2507.06719, 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 616, + 533, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 616, + 533, + 666 + ], + "spans": [ + { + "bbox": [ + 315, + 616, + 533, + 666 + ], + "type": "text", + "content": "[25] Guanxing Lu, Shiyi Zhang, Ziwei Wang, Changliu Liu, Jiwen Lu, and Yansong Tang. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In European Conference on Computer Vision, pages 349-366. Springer, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 667, + 533, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 667, + 533, + 687 + ], + "spans": [ + { + "bbox": [ + 315, + 667, + 533, + 687 + ], + "type": "text", + "content": "[26] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 725, + 315, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 725, + 315, + 734 + ], + "spans": [ + { + "bbox": [ + 296, + 725, + 315, + 734 + ], + "type": "text", + "content": "7130" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 99, + 296, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 95, + 99, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 99, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 95, + 99, + 296, + 140 + ], + "type": "text", + "content": "Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. arXiv preprint arXiv:2108.03298, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 142, + 296, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 296, + 182 + ], + "type": "text", + "content": "[27] Liyuan Mao, Haoran Xu, Xianyuan Zhan, Weinan Zhang, and Amy Zhang. Diffusion-dice: In-sample diffusion guidance for offline reinforcement learning. arXiv preprint arXiv:2407.20109, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 184, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 184, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 77, + 184, + 296, + 223 + ], + "type": "text", + "content": "[28] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020.1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 225, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 225, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 77, + 225, + 296, + 276 + ], + "type": "text", + "content": "[29] Yuzhe Qin, Yueh-Hua Wu, Shaowei Liu, Hanwen Jiang, Ruihan Yang, Yang Fu, and Xiaolong Wang. Dexamv: Imitation learning for dexterous manipulation from human videos. In European Conference on Computer Vision, pages 570-587. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 278, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 278, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 77, + 278, + 296, + 328 + ], + "type": "text", + "content": "[30] Aravind Rajeswaran, Vikash Kumar, Abhishek Gupta, Giulia Vezzani, John Schulman, Emanuel Todorov, and Sergey Levine. Learning complex dexterous manipulation with deep reinforcement learning and demonstrations. arXiv preprint arXiv:1709.10087, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 330, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 330, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 77, + 330, + 296, + 360 + ], + "type": "text", + "content": "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 362, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 362, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 77, + 362, + 296, + 403 + ], + "type": "text", + "content": "[32] Younggyo Seo, Danijar Hafner, Hao Liu, Fangchen Liu, Stephen James, Kimin Lee, and Pieter Abbeel. Masked world models for visual control. In Conference on Robot Learning, pages 1332-1344. PMLR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 404, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 404, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 77, + 404, + 296, + 445 + ], + "type": "text", + "content": "[33] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning " + }, + { + "bbox": [ + 77, + 404, + 296, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 77, + 404, + 296, + 445 + ], + "type": "text", + "content": " modes with one stone. Advances in neural information processing systems, 35:22955-22968, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 447, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 447, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 77, + 447, + 296, + 485 + ], + "type": "text", + "content": "[34] Nur Muhammad Mahi Shafullah, Anant Rai, Haritheja Etukuru, Yiqian Liu, Ishan Misra, Soumith Chintala, and Lerrel Pinto. On bringing robots home. arXiv preprint arXiv:2311.16098, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 488, + 296, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 488, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 77, + 488, + 296, + 528 + ], + "type": "text", + "content": "[35] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 530, + 296, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 530, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 77, + 530, + 296, + 550 + ], + "type": "text", + "content": "[36] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv:2010.02502, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 77, + 552, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 552, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 77, + 552, + 296, + 592 + ], + "type": "text", + "content": "[37] Che Wang, Xufang Luo, Keith Ross, and Dongsheng Li. Vrl3: A data-driven framework for visual deep reinforcement learning. Advances in Neural Information Processing Systems, 35:32974-32988, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 77, + 594, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 594, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 77, + 594, + 296, + 635 + ], + "type": "text", + "content": "[38] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimi-cplay: Long-horizon imitation learning by watching human play. arXiv preprint arXiv:2302.12422, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 77, + 637, + 296, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 637, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 77, + 637, + 296, + 686 + ], + "type": "text", + "content": "[39] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-drive world models for autonomous driving. In European Conference on Computer Vision, pages 55-72. Springer, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 100, + 533, + 529 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 100, + 533, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 100, + 533, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 100, + 533, + 139 + ], + "type": "text", + "content": "[40] Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning. arXiv preprint arXiv:2208.06193, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 141, + 533, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 141, + 533, + 191 + ], + "spans": [ + { + "bbox": [ + 315, + 141, + 533, + 191 + ], + "type": "text", + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20310-20320, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 193, + 533, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 193, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 193, + 533, + 232 + ], + "type": "text", + "content": "[42] Philipp Wu, Alejandro Escontrela, Danijar Hafner, Pieter Abbeel, and Ken Goldberg. Daydreamer: World models for physical robot learning. In Conference on robot learning, pages 2226-2240. PMLR, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 234, + 533, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 234, + 533, + 282 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 533, + 282 + ], + "type": "text", + "content": "[43] Long Yang, Zhixiong Huang, Fenghao Lei, Yucun Zhong, Yiming Yang, Cong Fang, Shiting Wen, Binbin Zhou, and Zhouchen Lin. Policy representation via diffusion probability model for reinforcement learning. arXiv preprint arXiv:2305.13122, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 285, + 533, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 533, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 533, + 334 + ], + "type": "text", + "content": "[44] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4): 1-39, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 335, + 533, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 533, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 533, + 376 + ], + "type": "text", + "content": "[45] Ruihan Yang, Zhuoqun Chen, Jianhan Ma, Chongyi Zheng, Yiyu Chen, Quan Nguyen, and Xiaolong Wang. Generalized animal imitator: Agile locomotion with versatile motion prior. arXiv preprint arXiv:2310.01408, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 377, + 533, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 377, + 533, + 426 + ], + "spans": [ + { + "bbox": [ + 315, + 377, + 533, + 426 + ], + "type": "text", + "content": "[46] Yanjie Ze, Ge Yan, Yueh-Hua Wu, Annabella Macaluso, Yuying Ge, Jianglong Ye, Nicklas Hansen, Li Erran Li, and Xiaolong Wang. Gnfactor: Multi-task real robot learning with generalizable neural feature fields. In Conference on Robot Learning, pages 284-301. PMLR, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 428, + 533, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 428, + 533, + 478 + ], + "spans": [ + { + "bbox": [ + 315, + 428, + 533, + 478 + ], + "type": "text", + "content": "[47] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In Proceedings of Robotics: Science and Systems (RSS), 2024. 2, 5, 6, 8, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 479, + 533, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 479, + 533, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 479, + 533, + 529 + ], + "type": "text", + "content": "[48] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37:77799-77830, 2025. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 725, + 314, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 725, + 314, + 734 + ], + "spans": [ + { + "bbox": [ + 296, + 725, + 314, + 734 + ], + "type": "text", + "content": "7131" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_content_list.json b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f93dde7aeb2504f6a25140f32cd2e119a9c0b0cb --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Spatial-Temporal Forgery Trace based Forgery Image Identification", + "text_level": 1, + "bbox": [ + 153, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yilin Wang $^{1,2\\dagger}$ , Zunlei Feng $^{1,2,3\\dagger}$ , Jiachi Wang $^{1,2}$ , Hengrui Lou $^{1,2}$ , Binjia Zhou $^{1,2}$ , Jie Lei $^{4}$ , Mingli Song $^{1,2}$ , Yijun Bei $^{1,2*}$", + "bbox": [ + 181, + 178, + 815, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1State Key Laboratory of Blockchain and Data Security, Zhejiang University", + "bbox": [ + 197, + 218, + 800, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Software Technology, Zhejiang University", + "bbox": [ + 285, + 237, + 712, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Zhejiang Provincial Key Laboratory of Service Robot", + "bbox": [ + 282, + 253, + 715, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4College of Computer Science, Zhejiang University of Technology", + "bbox": [ + 233, + 272, + 759, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 323, + 326, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid development of AIGC technology has enabled highly realistic forged images to deceive human perception, posing serious risks across many areas. Current deepfake image detection methods primarily identify forgeries by extracting handcrafted features, deep features, and frequency-domain features. While these features contain forgery traces, they also include a substantial amount of the image's semantic information, which interferes with the precision and generalization of forgery detection models. To tackle these challenges, this paper introduces a novel forgery image identification method based on the Spatial-Temporal Forgery Trace (STFT). Motivated by the fact that forgery images are more easily fitted to a specific distribution than real images, the STFT method approaches the issue from a forged image distribution modeling perspective, employing generative diffusion models to meticulously capture the temporal distribution of images. It further models the relationship between temporal feature variations and spatially corresponding temporal features, treating them as temporal and spatial forgery traces. Moreover, STFT incorporates frequency-domain features as weighting factors to accelerate the localization of spatio-temporal forgery traces. Experiments demonstrate that by integrating spatial, temporal, and frequency perspectives within the latent space, STFT effectively captures subtle spatio-temporal forgery traces, exhibiting strong robustness and generalizability. It outperforms state-of-the-art methods on major benchmark datasets in the field. The source code for STFT is available at https://github.com/GCLion/STFT.", + "bbox": [ + 86, + 354, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 821, + 220, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of Artificial Intelligence Generated Content (AIGC) technology [5], particularly the ma", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "turity of generative models like diffusion models, AIGC has been widely applied in fields such as image and video generation, bringing significant innovation and convenience. However, the widespread application of AIGC has also brought about a series of serious risks. AIGC technology has made it extremely easy to generate realistic images. This has led to the proliferation of false information, malicious tampering, and other harmful activities. Additionally, by generating forged images, malicious attackers may impersonate others to carry out fraud, causing economic losses and a crisis of social trust [51].", + "bbox": [ + 511, + 325, + 906, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current deepfake image detection methods primarily rely on extracting handcrafted features, deep features, and frequency domain features for forgery identification. Handcrafted feature-based approaches [10, 18] typically extract statistical features such as color histograms, local binary patterns (LBP), and histograms of oriented gradients (HOG) to detect fake images. While these methods may be effective for early deepfake images, they exhibit poor adaptability to emerging deepfake techniques, such as images generated by diffusion models, making them less effective against continuously evolving attack strategies. Methods based on deep features [17, 41] typically utilize deep neural networks such as CNN, ResNet, and EfficientNet to extract high-level features from images for classification and discrimination. While these approaches can improve detection performance to some extent, they often learn the semantic features of the image itself rather than focusing specifically on forgery traces, which affects the generalization ability of the forgery detection model. Frequency domain-based methods [14, 47] extract features from high-frequency or specific frequency bands using frequency domain analysis techniques such as Fourier transform and wavelet transform to detect forgery traces. However, since different AIGC generation methods may use different generation strategies, the stability of frequency domain features is insufficient, which limits the generalization ability of these methods.", + "bbox": [ + 511, + 492, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The features extracted by the most of aforementioned", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. † Equal contribution to this work.", + "bbox": [ + 107, + 886, + 415, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods not only contain forgery trace characteristics but also include a large amount of semantic information from the image itself, which interferes with the accuracy and generalization ability of forgery detection models.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the generative mechanisms of contemporary deep image synthesis models, these models create images by assimilating particular data distributions. However, their gradient-based optimization processes inherently display a propensity for \"shortcut learning\"—favoring the acquisition of commonly encountered feature distributions while neglecting those that are more infrequent. This phenomenon is driven by the observation that forged image features are easily assimilated, whereas the subtle, rare features inherent in genuine images often evade capture.", + "bbox": [ + 88, + 152, + 482, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper adopts a perspective of forgery image distribution modeling, using generative diffusion models to obtain image temporal distribution features. Then, we mine forgery trace features from the temporal distribution feature space that do not contain semantic information interference, aiming to achieve high accuracy and high generalization in image forgery detection. By leveraging the mapped latent temporal distribution of images, this paper elucidates the evolving patterns of temporal distribution features, identifying them as temporal forgery indicators. In the spatial dimension, based on the hypothesis that forgery images tend to model the approximate semantic parts with either exactly the same or completely different distributions, we extract the correlation between the temporal features of corresponding image regions as spatial forgery traces. Furthermore, generated images typically expose forgery traces more easily in high-frequency regions. We obtain high-frequency information as weights to help the model more quickly identify spatio-temporal forgery traces. This approach provides an innovative perspective for research in the detection of forged images using deep learning models. Our main contributions are as follows:", + "bbox": [ + 91, + 306, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an innovative spatial-temporal forgery trace extraction framework designed for the identification of forged images through latent diffusion distribution. Rather than relying on mixed features, the temporal evolution patterns of individual temporal features and the interrelationships among multiple temporal features are modeled as temporal and spatial forgery traces, enabling robust and precise forgery identification.", + "- Instead of utilizing frequency features directly, high-frequency components are employed as weighting factors to enhance the forgery detection model's ability to swiftly localize spatio-temporal forgery traces.", + "- Comprehensive experiments demonstrate that the proposed method surpasses the limitations of prior approaches in feature extraction, and exhibits exceptional robustness and generalization capabilities against various forgery techniques." + ], + "bbox": [ + 89, + 641, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Forgery Technology", + "text_level": 1, + "bbox": [ + 513, + 114, + 754, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of image generation was initially led by technologies based on Generative Adversarial Networks (GANs) and autoregressive models. Through the continuous evolution of these technologies, milestone products such as BigGAN [3] and DALL-E [13] emerged. However, these methods face certain limitations in content generation control, with stability and complexity issues being particularly prominent. This has led to the emergence of a new type of image generation method—the diffusion model. The Denoising Diffusion Probabilistic Model (DDPM) [19], inspired by nonequilibrium thermodynamics, was born and demonstrated excellent generative performance, further driving in-depth research on diffusion models. Subsequently, researchers have focused on optimizing its structure and improving sampling efficiency. The Latent Diffusion Model (LDM) [38] applies the diffusion process in the latent space, not only improving model efficiency but also enhancing the ability to generate text-conditioned images through a cross-attention mechanism. LDM has become a significant advancement in the field of image generation, leading to the development of well-known generative models such as Stable Diffusion [38] and LCM-RoLA [27].", + "bbox": [ + 511, + 138, + 906, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image Forgery Detection", + "text_level": 1, + "bbox": [ + 513, + 482, + 741, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For traditional Photoshop-based forgery methods, early research primarily relied on extracting handcrafted features, deep features, and frequency-domain features for forgery identification. For deepfake images, different approaches have been explored based on distinct feature perspectives. Below, we introduce related works from three aspects: temporal, spatial, and frequency domains.", + "bbox": [ + 511, + 506, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In forgery detection methods based on spatial features, Sha et al. [39] analyze unique characteristics of generated images and perform model attribution, aiding in source identification and combating the misuse of forged images. The method in [45] detects forged images by comparing the input image with its reconstruction from a diffusion model, demonstrating strong performance in detecting forgeries under unknown diffusion models and various perturbations. Overall, spatial-based forgery detection methods mainly extract image features using deep models. However, these extracted features often mix semantic information with forgery traces, where the presence of semantic features increases the learning difficulty and reduces detection accuracy, ultimately leading to poor generalization performance of the detection models.", + "bbox": [ + 511, + 613, + 908, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In forgery detection methods based on temporal features, [29] effectively identifies images generated by diffusion models by leveraging the deterministic inversion and denoising computation errors of diffusion models, combined", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with statistical and neural network techniques. Similarly, [9] analyzes forensic features of GAN-generated images, which can also be used to evaluate the performance of existing GAN detectors on diffusion model images, further improving detection accuracy. Although forgery detection methods based on temporal features utilize diffusion models to extract temporal characteristics, they still rely on supervised signals to extract mixed features from the aggregated temporal features. Like spatial feature-based forgery detection, these methods suffer from interference with model accuracy and poor generalization ability.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In forgery detection based on image frequency domain features, [46] extracts frequency domain features using the Fourier transform and finds that forged images often exhibit different frequency distributions compared to real images. A convolutional neural network is then used to classify high-frequency features. [42] enhances detection performance by learning forgery-specific features in the frequency domain space and analyzing the abnormal patterns of forged images across different frequency bands. These methods determine authenticity based on the quantity and distribution of high-frequency features. However, due to the diversity of image types, high-frequency features also vary, leading to limited cross-domain generalization.", + "bbox": [ + 89, + 257, + 482, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The methods mentioned above are mostly designed for GAN-generated images. Although Corvi et al. [8, 9, 36] also found that spectral artifacts may exist in diffusion-generated images, the effectiveness of spectral artifact-based forgery detection methods in practice remains limited. To address this issue, researchers have begun to analyzing the characteristics of the diffusion generation process itself. For example, DIRE [45] and De-fake [39] use the reconstruction error of the diffusion model to detect forged images, while LaRE2 [28] and AEROBLADE [37] focus on the reconstruction error in the latent space.", + "bbox": [ + 89, + 454, + 482, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these methods are highly dependent on pretrained reconstruction models and are susceptible to the limitations of the training data characteristics. For example, if the reconstruction model is trained on a specific category (such as cats) but is used to detect forged images from different categories (such as dogs or plants), it may result in misclassification. The core issue here lies in using reconstruction errors as the basis for detection, without fully exploring the characteristics of the diffusion generation mechanism itself, thus affecting the understanding of the forgery image generation patterns.", + "bbox": [ + 89, + 619, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 799, + 228, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define the key notations of the latent diffusion model and outline its basic principles. In LDM [27], given an image $x \\in \\mathbb{R}^{H \\times W \\times 3}$ in the RGB space ( $H$ and $W$ denote the height and width of the image $x$ ), the encoder $\\mathcal{E}$ encodes $x$ into a latent representation $z = \\mathcal{E}(x)$ ,", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while the decoder $\\mathcal{D}$ reconstructs the image from the latent representation, yielding $\\tilde{x} = \\mathcal{D}(z) = \\mathcal{D}(\\mathcal{E}(x))$ , where ideally $z \\in \\mathbb{R}^{h \\times w \\times c}$ ( $h, w$ , and $c$ denote the height, width, and channel of the latent representation $z$ ). This diffusion model consists of a $T$ -step diffusion process $q$ for $1 \\leq t \\leq T$ and a denoising process, which is defined as follows:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 191, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 218, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z_{t}$ is the diffusion result at time step $t$ , $\\beta_{t}$ is the noise factor at time step $t$ , $I$ is the identity matrix, and $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ represent the mean and covariance matrix of the denoising distribution, respectively. After training the model, step-by-step denoising can be performed starting from the standard normal distribution $z_{T} \\sim \\mathcal{N}(0, I)$ :", + "bbox": [ + 511, + 239, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {1}{\\sqrt {1 - \\beta_ {t}}} \\left(z _ {t} - \\beta_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) + \\sigma_ {t} \\xi , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 339, + 903, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\xi \\sim \\mathcal{N}(0, I)$ is the random noise used to control sampling quality, and $\\sigma_t$ regulates the randomness of the denoising process. Through repeated iterations, the final representation $z_0$ is obtained, which is then decoded by $\\mathcal{D}$ to reconstruct the synthesized image. During the forward diffusion process, high-frequency components are gradually suppressed by noise. By progressively adding noise in the forward process and denoising in the reverse process, high-frequency components become increasingly compressed. Notably, at high noise levels, the loss of high-frequency details becomes inevitable.", + "bbox": [ + 511, + 378, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 558, + 604, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As depicted in Figure 1, this method employs a diffusion process to map the image into the latent distribution space, subsequently extracting forgery traces from this latent distribution through a comprehensive spatial-temporal analysis. The overall framework consists of the following three main modules: Temporal Prior Correlation Modeling analyzes the latent space representations at different time steps of the diffusion process to capture their temporal variation characteristics. Spatial Correlation Modeling extracts features from different latent dimensions and computes spatial correlations using a self-attention mechanism to model spatial dependencies. Frequency-Enhanced Attention Mechanism utilizes frequency domain information to guide temporal prior correlation computation and spatial correlation analysis, thereby expediting the localization of forgery traces and enhancing generalization capabilities by disregarding the interference of inconsequential features.", + "bbox": [ + 511, + 583, + 906, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Image Latent Distribution Modeling", + "text_level": 1, + "bbox": [ + 511, + 848, + 834, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the limitations of existing methods and gain a deeper understanding of the underlying differences between", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6bc072cf4fd90f648d1333f7c61848b229eaf54039847ad2f0313cf0c92bf0b5.jpg", + "image_caption": [ + "Figure 1. The framework of STFT. It begins by mapping the input image into a latent diffusion distribution space. Subsequently, temporal and spatial forgery traces are extracted by capturing patterns of temporal variation and spatial correlations within the latent diffusion distribution. High-frequency components are utilized as weight to facilitate the rapid localization of forgery traces." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "forged and normal images, we propose a forgery detection approach from a distributional perspective. Considering that LDM maps data into a low-dimensional latent space $z_{t}$ through an encoder $\\mathcal{E}$ , making the distribution $p(z)$ more compact and easier to model, while the high-dimensional pixel space $p(x)$ remains complex and difficult to model directly, we extract the latent representations $z \\in \\mathbb{R}^{h \\times w \\times c}$ during the diffusion process of the latent diffusion model for further analysis.", + "bbox": [ + 88, + 412, + 480, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To facilitate the modeling of distributional changes from a temporal perspective, we reshape the latent representation $z_{t}$ at each diffusion step into a one-dimensional vector $\\vec{z}_{t} \\in \\mathbb{R}^{1 \\times d}$ as sequential data at time step $t$ , where $d = h \\times w \\times c$ , $t \\in \\{1,2,\\dots,T\\}$ . The sequential vectors from step 1 to $T$ are concatenated vertically into a group of tokens $X = \\{\\vec{z}_{t}\\}_{t=1}^{T} \\in \\mathbb{R}^{T \\times d}$ , which serves as input for both temporal and frequency-domain modeling.", + "bbox": [ + 89, + 549, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Temporal Prior Association Modeling", + "text_level": 1, + "bbox": [ + 89, + 681, + 416, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the assumption that the latent space distribution of forged images tends to be smoother with weaker temporal variations, we draw inspiration from the unsupervised temporal anomaly representation learning [48] and employ a trainable Gaussian kernel to compute prior associations between time steps. We introduce an adjustable scale parameter $\\sigma$ to adapt to different temporal sequence patterns. The unimodal property of the Gaussian kernel enhances the association between local time steps, making the non-stationary characteristics of real images more prominent, while the smooth evolution pattern of forged images becomes more distinguishable. This improves the temporal modeling capability for forgery trace detection.", + "bbox": [ + 88, + 703, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the latent space representations $\\{\\vec{z}_t\\}_{t=1}^T$ at different time steps, we treat the diffusion process as a sequential series, concatenate them into $X \\in \\mathbb{R}^{T \\times d}$ , and compute prior associations between time steps. Specifically, we define the prior association matrix $P$ as follows:", + "bbox": [ + 511, + 412, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP = \\operatorname {R e s c a l e} \\left(\\frac {1}{\\sqrt {2 \\pi} \\sigma_ {t}} \\exp \\left(- \\frac {\\left| t ^ {\\prime} - t \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 497, + 906, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{t}$ represent the learnable scale parameters for the $t$ -th time step. The prior association is computed using a Gaussian kernel, with $G(|t' - t|; \\sigma_t) = \\frac{1}{\\sqrt{2\\pi} \\sigma_t} \\exp \\left( -\\frac{|t' - t|^2}{2 \\sigma_t^2} \\right)$ representing the association weight calculated based on the distance $|t' - t|$ between time points $t$ and $t'$ . The function Rescale(\\cdot) normalizes the association weights into a discrete distribution $P$ .", + "bbox": [ + 511, + 537, + 908, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Spatial Correlation Modeling", + "text_level": 1, + "bbox": [ + 511, + 657, + 774, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Forged images often exhibit abnormal spatial similarities (e.g., excessive symmetry) or inconsistencies (e.g., color or texture mismatches). To address this, we extract features from each dimension of the latent space and apply a self-attention mechanism to compute spatial correlations within the latent space. This helps amplify regions with abnormal similarities or inconsistencies in forged images, thereby enhancing forgery trace detection.", + "bbox": [ + 511, + 679, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the feature matrix $X$ , the vertical vector $u_{i} = X[(:,i]$ where $u_{i}\\in \\mathbb{R}^{T\\times 1}$ and $i\\in \\{1,2,3,\\dots ,d\\}$ , represents the pixel- or region-wise temporal features. Consequently, the spatial token group $X^{\\top} = \\{u_{i}^{\\top}\\}_{i = 1}^{d}\\in \\mathbb{R}^{d\\times T}$ can be utilized to compute the spatial correlation matrix $S$ as follows:", + "bbox": [ + 511, + 799, + 908, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ, K, V = X ^ {\\top} W _ {Q}, X ^ {\\top} W _ {K}, X ^ {\\top} W _ {V}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 882, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\left(S o f t m a x \\left(\\frac {Q K ^ {\\top}}{\\sqrt {T}}\\right) V\\right) ^ {\\top} A, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 88, + 483, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\top$ denotes the transpose operation, and $Q,K,V$ along with $W_{Q},W_{K},W_{V}$ represent the query, key, and value matrices and their corresponding parameters, respectively. The term $A$ signifies the frequency-enhanced attention weight, which directs the computation of the spatial correlation matrix to emphasize pixel- or region-wise temporal vectors exhibiting significant dynamic evolution patterns. Further elaboration and the corresponding equations for the frequency-enhanced attention matrix $A$ are discussed in the subsequent subsection.", + "bbox": [ + 89, + 128, + 483, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Frequency-Enhanced Attention Mechanism", + "text_level": 1, + "bbox": [ + 89, + 287, + 464, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In frequency domain analysis, the high-frequency details of forged images tend to be eliminated during gradient optimization, resulting in significant differences in frequency characteristics compared to real images. To address this, we extract the frequency-domain information of latent space features using the Fast Fourier Transform (FFT) and compute frequency-domain correlations to capture cross-scale feature interactions. Since high-frequency information is lost in forged images, the frequency distribution of real and forged images exhibits distinct patterns after conjugate product computation. We then apply the Inverse Fourier Transform (IFFT) to map the correlation back to the spatiotemporal domain, allowing the model to focus on anomalies caused by missing frequency-domain information.", + "bbox": [ + 89, + 309, + 483, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the group of tokens $X = \\{\\vec{z}_t\\}_{t=1}^T$ , the self-attention with query matrix $\\overline{Q}$ and key matrix $\\overline{K}$ for $X$ is adopted to find the feature temporal similarity of different tokens $\\{\\vec{z}_t\\}_{t=1}^T$ . Furthermore, we perform the Fourier Transform on the query matrix $\\overline{Q}$ and key matrix $\\overline{K}$ to extract frequency-domain features:", + "bbox": [ + 89, + 520, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {Q} = \\mathcal {F} (\\overline {{Q}}), \\widehat {K} = \\mathcal {F} (\\overline {{K}}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 618, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}()$ denotes the Fast Fourier Transform. Next, we compute the conjugate product of the query and key in the frequency domain to obtain cross-scale interaction information $C$ :", + "bbox": [ + 89, + 645, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC = \\widehat {Q} \\cdot \\widehat {K} ^ {*}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 703, + 480, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where “.” denotes element-wise multiplication of matrices, $\\widehat{K}^*$ represents the complex conjugate, ensuring correct interaction in the frequency domain and highlighting the different frequency distributions of real and forged images. Then, we apply the Inverse Fourier Transform $\\mathcal{F}^{-1}$ to restore the result to the spatio-temporal domain and perform Softmax normalization to ensure that the weights maintain appropriate scales in the spatio-temporal domain, yielding the frequency-enhanced spatio-temporal attention weights:", + "bbox": [ + 89, + 726, + 483, + 863 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F} ^ {- 1} (C)}{\\sqrt {T}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 869, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, $A$ serves as the frequency-enhanced temporal attention weight to improve generalization ability.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Learning Strategy", + "text_level": 1, + "bbox": [ + 511, + 130, + 689, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we assess the distribution characteristics of the input time series $X$ across multiple levels by computing the association difference $(AssD)$ between the prior association $P$ and the frequency-domain association $A$ . Specifically, we employ the Kullback-Leibler divergence (KL divergence) to quantify the discrepancy between two discrete distributions, defined as follows:", + "bbox": [ + 511, + 152, + 905, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA s s D (P, A) = K L (P [ t,: ] \\| A [ t,: ]) + K L (A [ t,: ] \\| P [ t,: ]),\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 268, + 901, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $KL(A[t,:]\\| P[t,:])$ denotes the KL divergence, measuring the informational disparity between distributions $A[t,:]$ and $P[t,:]$ . Here, $P[t,:]$ and $A[t,:]$ represent the prior association and frequency-domain association distributions at time step $t$ , respectively. The final association difference $AssD(P,A)$ captures the overall variation in association patterns.", + "bbox": [ + 511, + 295, + 905, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the output of the spatial correlation modeling module, we introduce a global optimization objective based on cross-entropy loss to steer the spatial correlation and frequency-domain enhancement modules in distinguishing between forged and authentic images.", + "bbox": [ + 511, + 402, + 905, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We formulate distinct optimization objectives tailored to different sample types. For forged images, we enforce a deviation between the prior association $P$ and the frequency-domain association $A$ , thereby accentuating the intricate and challenging-to-model characteristics of the real image distribution:", + "bbox": [ + 511, + 477, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f a k e}} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 578, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term represents the cross-entropy loss between the classification output $\\hat{y}$ of the spatial correlation modeling module and the ground truth labels $y$ , enhancing the module's classification performance. The second term regulates the relationship between the prior association $P$ and the frequency-domain association $A$ using the association difference $AssD()$ for forged images, thereby improving the model's temporal modeling capability. Here, $A_{fix}$ represents the fixed frequency-domain association. The hyperparameter $\\lambda$ modulates the contribution of the $AssD()$ loss.", + "bbox": [ + 511, + 604, + 905, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For real images, we allow the prior association $P$ to closely align with the frequency-domain association $A$ , emphasizing the smoothness of forged images in the frequency domain as follows:", + "bbox": [ + 511, + 771, + 905, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e a l}} = C E (y, \\hat {y}) + \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 843, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the second term ensures that the prior association $P$ remains distant from the frequency-domain association $A$", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/86007e8979d48ebffdd952337aaf332324256919b0506cb1ae84c7b0dfb39632.jpg", + "image_caption": [ + "(a) Performance comparisons on Genimage (ACC, %)." + ], + "image_footnote": [], + "bbox": [ + 116, + 90, + 450, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fabdbd33fd9f418b3779111ff79717d3103e790e37a24df61b065c6ab6bc74eb.jpg", + "image_caption": [ + "(b) Performance comparisons on Deepfacegen (AUC, %).", + "Figure 2. The radar chart shows the forgery identification performance of STFT in comparison to 12 alternative methods on Genimage (a) and 10 other methods on Deepfacegen (b). Solid lines depict the performance of the comparative methods, while a dashed line signifies the performance of STFT." + ], + "image_footnote": [], + "bbox": [ + 547, + 88, + 883, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for real images, thereby enhancing the model's discriminative capability.", + "bbox": [ + 88, + 376, + 482, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a batch comprising both forged and real images, in addition to the aforementioned losses $\\mathcal{L}_{\\text{fake}}$ and $\\mathcal{L}_{\\text{real}}$ , we ensure that $A$ remains distinct from $P$ , thereby directing greater attention to non-adjacent regions and enhancing the flexibility of global associations:", + "bbox": [ + 88, + 407, + 483, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m i x} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P _ {f i x}, A) \\|, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 494, + 482, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{fix}$ denotes the fixed prior association matrix. This alternating optimization strategy stabilizes the training process and utilizes the frequency-domain association $A$ to guide the model in focusing on more nuanced frequency-domain details, effectively distinguishing between genuine and forged images.", + "bbox": [ + 88, + 523, + 483, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 630, + 225, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first part of this section, we introduce the detailed experimental setup. Next, in the second part, we evaluate and compare STFT with other forgery detection methods, including SOTA approaches. In the fourth part, we assess the generalization performance of STFT. Finally, we conduct an ablation study on the proposed model.", + "bbox": [ + 88, + 655, + 483, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 757, + 282, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics and Datasets. In alignment with the established comparison standards, we evaluate accuracy (ACC) and area under the curve (AUC) as performance metrics for the GenImage [53] and DeepFaceGen [2] datasets, respectively. The GenImage [53] dataset consists of a total of 2,681,167 images, including 1,331,167 real images and 1,350,000 generated images. These generated images are sourced from eight different generative models: BigGAN", + "bbox": [ + 88, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[3], GLIDE [30], VQDM [16], Stable Diffusion V1 and V1.5 [38], ADM [12], Midjourney [21], and Wukong [15]. All fake images are generated using the template prompt \"photo of CLS,\" where \"CLS\" is replaced by one of the 1,000 labels from ImageNet [11]. The dataset is divided into eight subsets, each containing a portion of real images along with all fake images generated by the aforementioned models. In this study, we follow the official dataset split, using 2,581,167 images for training and the remaining 100,000 images for validation.", + "bbox": [ + 511, + 376, + 906, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DeepFaceGen [2] is a facial forgery dataset that includes both video and image modalities. We select face forgery data generated using novel prompt-guided forgery methods. These methods cover 12 diffusion- and autoregressive-based models across text-to-image and image-to-image modalities. The dataset includes face forgeries across various ages, genders, and environments, with generated images sourced from popular models such as Midjourney [21], DALL-E [13], and SDXL [38].", + "bbox": [ + 511, + 547, + 908, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We downloaded the pretrained LDM model, fed images into it, and extracted the latent space data. We set the default diffusion steps to 20, obtaining 20 latent representations of size $3 \\times 64 \\times 64$ . Each latent representation was then processed using average pooling to reduce its size to $3 \\times 8 \\times 8$ and flattened into a one-dimensional vector, ultimately forming sequential data of size $20 \\times 192$ . We set the batch size to 256, the number of channels in the hidden state to 512, and the number of attention heads to 8. The hyperparameter $\\lambda$ was set to 2 for all datasets. We used the ADAM optimizer [23] with an initial learning rate of $1 \\times 10^{-4}$ . All experiments were conducted on an Nvidia A40 GPU.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17072", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2e02365c6c670331305792add9b6d48f6bc3e215b69e7949b0169b40b2336fad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYearMidjourneySDv1.4SDv1.5ADMGLIDEWukongVQDMBigGAN
DIRE [45]202399.5 / 72.699.7 / 68.899.7 / 68.899.2 / 67.999.8 / 52.299.7 / 68.199.9 / 51.9100 / 50.1
Defake [39]202398.8 / 78.799.8 / 81.5100 / 82.398.7 / 82.599.8 / 73.199.3 / 75.699.1 / 82.098.7 / 76.6
PatchCraft [52]202390.1 / 81.189.5 / 81.390.1 / 81.390.4 / 79.390.6 / 80.492.6 / 79.390.3 / 74.980.9 / 73.7
AIDE [49]202498.9 / 86.099.7 / 84.999.6 / 84.095.3 / 79.197.6 / 81.898.8 / 80.492.3 / 83.096.3 / 81.7
DRCT [6]202499.9 / 89.395.5 / 86.696.8 / 89.195.1 / 81.896.6 / 83.899.6 / 84.494.3 / 82.498.8 / 84.0
STFT (Ours)-99.4 / 92.399.6 / 89.199.3 / 90.499.5 / 82.599.8 / 87.0100 / 84.899.7 / 85.7100 / 85.6
", + "bbox": [ + 94, + 88, + 903, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 89, + 281, + 390, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GenImage Benchmark. To verify the model's generalization ability across different generative models, this section follows the GenImage evaluation standard and compares our approach with 12 mainstream forgery detection models, including CNNSpot [34], F3Net [33], CLIP/RN50 [34], GramNet [24], De-fake [39], Conv-B [26], SwinT [25], UnivFD [31], DIRE [45], PathCraft [52], AIDE [49], and DRCT [6]. According to the GenImage evaluation protocol, all models were trained on the SDv1.4 subset of GenImage and evaluated on multiple test sets. As shown in Figure 2a, nearly all methods achieve high accuracy on the Wukong, SDv1.4, and SDv1.5 datasets but perform poorly on GLIDE, VQDM, Midjourney, ADM, and BigGAN datasets. Since BigGAN is a non-diffusion model dataset, all methods perform the worst on it. In contrast, our proposed STFT achieves over $89\\%$ accuracy on the BigGAN dataset and obtains the highest ACC on 7 out of 8 sub-datasets, demonstrating a significant advantage over existing forgery detection methods (see Appendix Table 4 for complete results).", + "bbox": [ + 91, + 321, + 483, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DeepFaceGen Benchmark. Following the DeepFaceGen evaluation standard, we compare STFT with mainstream generative image detection methods, including Xception [7], EfficientNet-B0 [44], F3-Net [33], RECCE [4], DNADet [50], DIRE [45], DRCT [6], UnivFD [31], NPR [43], and FreqNet [42]. Training and testing were conducted on the same dataset, which includes real images and images generated by 12 different forgery methods such as DALL-E and SD. The dataset was split into training, validation, and test sets with a 7:1:2 ratio. As shown in Figure 2b, STFT outperforms all methods across all 12 subsets, with a particularly notable advantage in distinguishing difficult-to-detect forged images generated by DALL-E 1, DALL-E 3, Midjourney, and Wenxin. Overall, STFT achieves an AUC above $90\\%$ for all forgery categories, with an average AUC of $95.21\\%$ (detailed results are provided in Appendix Table 5).", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d4dbd4e1040f9dd048ce4ede268885377a28c0142030e5846a8ece784ce813d.jpg", + "table_caption": [ + "Table 1. Performance comparison (ACC, %) of various methods. Each value is presented as (Score1 / Score2), where Score1 corresponds to the accuracy achieved when trained and tested on the same subset, and Score2 represents the average accuracy when trained on the current subset and evaluated across other subsets. Bold values highlight the best performance." + ], + "table_footnote": [], + "table_body": "
Train OnCatHorseBedroom
Test OnHorse/BedroomCat/BedroomCat/Horse
DIRE [45]73.0165.4555.04
DRCT [6]74.4868.7867.62
STFT (Ours)90.3289.7091.51
", + "bbox": [ + 517, + 279, + 908, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Generalization performance (AUC,%) of different methods on non-face images.", + "bbox": [ + 511, + 385, + 906, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Generalization Performance", + "text_level": 1, + "bbox": [ + 511, + 440, + 767, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalization capability of STFT, we designed two experimental scenarios: first, testing the model when the pre-trained model does not align with the target detection task, and second, assessing its adaptability to unknown forgery methods.", + "bbox": [ + 511, + 462, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptation to Unknown Forgery Methods. We trained STFT on a single subset of the GenImage dataset and tested it across all eight subsets. As shown in Table 1, we compared STFT with the top four models from the baseline evaluations. STFT outperformed the existing SOTA method in GenImage, DRCT, and achieved the best generalization performance across all eight subsets. When confronted with unseen forgery methods, all models experienced performance degradation in various detection metrics. However, STFT still demonstrated strong generalization ability. The superior generalization of STFT is primarily attributed to its frequency-domain-enhanced spatio-temporal correlation modeling, which captures multi-perspective features of forged images. This enhances the model's adaptability to different types of forgeries, allowing it to effectively recognize anomalous patterns even when facing previously unseen forgery methods.", + "bbox": [ + 511, + 537, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalization Performance on Non-face Images. Reconstruction-error-based forgery detection methods often fail when the test data differs from the reconstruction model's pre-training domain. For instance, applying a model trained on animals to plant images may misclassify real images as fake due to high reconstruction errors. To assess STFT's robustness in such cases, we used pre-trained", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17073", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/51844763180338786bd7969f6ab4d167bbd269bfb512c9cb64b0fb3c49c3647f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model VariantTemporalSpatialFrequencyAUC(%)
Full Model✓✓✓92.02
w/o Temporal✓✓86.28
w/o Spatial✓✓79.71
w/o Frequency✓✓89.75
", + "bbox": [ + 96, + 88, + 480, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on different modules.", + "bbox": [ + 148, + 176, + 419, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models from [12] on LSUN Bedroom, Cat, and Horse categories, and compared STFT with DIRE and DRCT. Following the Benchmark 1 protocol on 12 prompt-based forgery datasets, results in Table 2 show that while reconstruction-based methods suffered major performance drops, STFT consistently achieved AUC scores over $90\\%$ , highlighting its strong generalization ability.", + "bbox": [ + 89, + 217, + 483, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 334, + 243, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation experiments to quantify the contribution of each design component in our model and analyze the impact of key hyperparameters. Specifically, we examine the effect of the latent space sampling steps and feature map resolution and the inclusion of the frequency-enhanced attention module.", + "bbox": [ + 89, + 354, + 483, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Latent Space Sampling Steps and Feature Map Resolution. To investigate the impact of sampling steps and feature resolution in latent space, we evaluate detection performance with sampling steps set to 10, 15, 20, 25, and 30, and test four resolutions by downsampling latent features via average pooling. As shown in Figure 3, the model performs best with 20 sampling steps and a feature resolution of $3 \\times 8 \\times 8$ .", + "bbox": [ + 89, + 446, + 483, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This indicates that 20 steps sufficiently preserve temporal cues while preventing over-smoothing. A $3 \\times 8 \\times 8$ resolution balances detail retention and generalization. Lower resolutions (e.g., $3 \\times 4 \\times 4$ ) lose crucial spatial information, reducing accuracy, while higher resolutions introduce redundancy, affecting robustness. Thus, 20 sampling steps and $3 \\times 8 \\times 8$ resolution are adopted as default settings for optimal accuracy and efficiency.", + "bbox": [ + 89, + 566, + 483, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact Analysis of Model Components. To analyze the contributions of different STFT components, we conduct ablation experiments by sequentially removing the temporal prior correlation module, spatial correlation module, and frequency-enhanced attention module on DALL-E 3 Dataset. The results, presented in Table 3, confirm their complementary roles in forgery detection.", + "bbox": [ + 89, + 688, + 483, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Removing the temporal module significantly weakens the model's ability to capture sequential forgery patterns, leading to a notable performance drop. This aligns with our observation that latent diffusion smooths temporal transitions in forged images. Without spatial correlation modeling, the model struggles to detect local modifications, increasing false positives and reducing robustness. Frequency", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b833215d04f584db9a9227dad50305edda3acc39b74bcedc312382a20fe5f2cd.jpg", + "image_caption": [ + "Figure 3. The ablation study on diffusion steps (a) and latent feature map size (b). Solid lines in both cases represent the performance results corresponding to the final adopted values." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 697, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c04413263ed1255f384ca91c423de398ac5b97699cc35f3db8c515ee3dd0438.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 88, + 905, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features are crucial for detecting high-frequency artifacts in forgeries. Removing this module degrades performance, especially against unknown forgery methods, highlighting its role in generalization.", + "bbox": [ + 511, + 306, + 906, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 380, + 633, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes the Spatial-Temporal Forgery Trace (STFT) method for forgery image identification. Leveraging the tendency of forged images to fit specific distributions more easily than real ones, STFT employs generative diffusion models to capture temporal distributions and models relationships between temporal and spatial features as forgery traces. It also integrates frequency-domain features to accelerate trace localization. Experiments show that STFT effectively identifies subtle spatio-temporal traces by combining spatial, temporal, and frequency perspectives, demonstrating strong robustness and generalizability.", + "bbox": [ + 511, + 406, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works. Although the proposed STFT achieves high accuracy and demonstrates commendable generalization capabilities, it is not without its limitations. Firstly, the process of mapping the original image into the latent diffusion distribution is computationally intensive and time-consuming. Additionally, the method appears to exhibit suboptimal performance on GAN-based forgery images, likely due to fundamental differences in the generative mechanisms of GANs and generative diffusion models. In future work, we aim to enhance the efficiency of the proposed method and further extend its generalization ability.", + "bbox": [ + 511, + 571, + 908, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 752, + 692, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is funded by National Key Research and Development Project (Grant No: 2022YFB2703100), Ant Group through CCF-Ant Research Fund, and Information Technology Center, ZheJiang University.", + "bbox": [ + 511, + 776, + 906, + 838 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17074", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Baidu AI. Wenxin large model for text-to-image generation. Baidu AI Blog, 2023. 2", + "[2] Yijun Bei, Hengrui Lou, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Song, Mingli Song, and Zunlei Feng. A largescale universal evaluation benchmark for face forgery detection. arXiv preprint arXiv:2406.09181, 2024. 6", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2, 6", + "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4113-4122, 2022. 7", + "[5] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1", + "[6] Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In Forty-first International Conference on Machine Learning, 2024. 7, 2", + "[7] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 7", + "[8] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 973-982, 2023. 3", + "[9] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3", + "[10] Edmar RS De Rezende, Guilherme CS Ruppert, Antonio Theophilo, Eric K Tokuda, and Tiago Carvalho. Exposing computer generated images by using deep convolutional neural networks. Signal Processing: Image Communication, 66: 113-126, 2018. 1", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 8", + "[13] Tira Nur Fitria. Artificial intelligence (ai) technology in openai chatgpt application: A review of chatgpt in writing english essay. In ELT Forum: Journal of English Language Teaching, pages 44-58, 2023. 2, 6" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 1", + "[15] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Niu Minzhe, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, et al. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems, 35:26418-26431, 2022. 6, 2", + "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 6", + "[17] Weinan Guan, Wei Wang, Jing Dong, Bo Peng, and Tieniu Tan. Collaborative feature learning for fine-grained facial forgery detection and segmentation. arXiv preprint arXiv:2304.08078, 2023. 1", + "[18] Tahir Hassan, Aras Asaad, Dashti Ali, and Sabah Jassim. Artificial image tampering distorts spatial distribution of texture landmarks and quality characteristics. arXiv preprint arXiv:2208.02710, 2022. 1", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[21] Anna Jaruga-Rozdolska. Artificial intelligence as part of future practices in the architect's work: Midjourney generative tool as part of a process of creating an architectural form. Architectus, (3 (71):95-104, 2022. 6, 2", + "[22] Wei et al. Jiang. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16423-16432, 2021. 2", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[24] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 7, 2", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7, 2", + "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17075", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2", + "[28] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 3", + "[29] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. arXiv preprint arXiv:2307.06272, 2023. 2", + "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 2", + "[31] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 7, 2", + "[32] OpenAI. Dall-e 3: Openai's latest text-to-image model. OpenAI Blog, 2023. 2", + "[33] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7, 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 7, 2", + "[35] Aditya et al. Ramesh. Zero-shot text-to-image generation. arXiv preprint arXiv:2102.12092, 2021. 2", + "[36] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. arXiv preprint arXiv:2210.14571, 2022. 3", + "[37] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 3", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6", + "[39] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text-to-image generation models. In Proceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3418-3432, 2023. 2, 3, 7", + "[40] Stability.ai. Stable Diffusion, 2023. 2", + "[41] Eren Tahir and Mert Bal. Deep image composition meets image forgery. arXiv preprint arXiv:2404.02897, 2024. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5052-5060, 2024. 3, 7", + "[43] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7", + "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7", + "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2, 3, 7", + "[46] Jing Weng. Local frequency analysis for diffusion-generated image detection. In International Conference on Image Processing and Artificial Intelligence (ICIPAL 2024), pages 66-73. SPIE, 2024. 3", + "[47] Moritz Wolter, Felix Blanke, Raoul Heese, and Jochen Garcke. Wavelet-packets for deepfake image analysis and detection. Machine Learning, 111(11):4295-4327, 2022. 1", + "[48] Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. arXiv preprint arXiv:2110.02642, 2021. 4", + "[49] Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection. arXiv preprint arXiv:2406.19435, 2024. 7, 2", + "[50] Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4662-4670, 2022. 7", + "[51] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023. 1", + "[52] Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection. arXiv preprint arXiv:2311.12397, 2023. 7, 2", + "[53] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36:77771-77782, 2023. 6" + ], + "bbox": [ + 516, + 90, + 903, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17076", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_model.json b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59fcbefd36dea2194849768b904cbca7ff598a88 --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "Spatial-Temporal Forgery Trace based Forgery Image Identification" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.179, + 0.816, + 0.217 + ], + "angle": 0, + "content": "Yilin Wang\\(^{1,2\\dagger}\\), Zunlei Feng\\(^{1,2,3\\dagger}\\), Jiachi Wang\\(^{1,2}\\), Hengrui Lou\\(^{1,2}\\), Binjia Zhou\\(^{1,2}\\), Jie Lei\\(^{4}\\), Mingli Song\\(^{1,2}\\), Yijun Bei\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.219, + 0.802, + 0.237 + ], + "angle": 0, + "content": "1State Key Laboratory of Blockchain and Data Security, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.238, + 0.714, + 0.255 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Software Technology, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.255, + 0.716, + 0.273 + ], + "angle": 0, + "content": "3Zhejiang Provincial Key Laboratory of Service Robot" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.273, + 0.76, + 0.291 + ], + "angle": 0, + "content": "4College of Computer Science, Zhejiang University of Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.324, + 0.327, + 0.34 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.356, + 0.486, + 0.794 + ], + "angle": 0, + "content": "The rapid development of AIGC technology has enabled highly realistic forged images to deceive human perception, posing serious risks across many areas. Current deepfake image detection methods primarily identify forgeries by extracting handcrafted features, deep features, and frequency-domain features. While these features contain forgery traces, they also include a substantial amount of the image's semantic information, which interferes with the precision and generalization of forgery detection models. To tackle these challenges, this paper introduces a novel forgery image identification method based on the Spatial-Temporal Forgery Trace (STFT). Motivated by the fact that forgery images are more easily fitted to a specific distribution than real images, the STFT method approaches the issue from a forged image distribution modeling perspective, employing generative diffusion models to meticulously capture the temporal distribution of images. It further models the relationship between temporal feature variations and spatially corresponding temporal features, treating them as temporal and spatial forgery traces. Moreover, STFT incorporates frequency-domain features as weighting factors to accelerate the localization of spatio-temporal forgery traces. Experiments demonstrate that by integrating spatial, temporal, and frequency perspectives within the latent space, STFT effectively captures subtle spatio-temporal forgery traces, exhibiting strong robustness and generalizability. It outperforms state-of-the-art methods on major benchmark datasets in the field. The source code for STFT is available at https://github.com/GCLion/STFT." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.823, + 0.222, + 0.838 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.879 + ], + "angle": 0, + "content": "With the rapid development of Artificial Intelligence Generated Content (AIGC) technology [5], particularly the ma" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.908, + 0.492 + ], + "angle": 0, + "content": "turity of generative models like diffusion models, AIGC has been widely applied in fields such as image and video generation, bringing significant innovation and convenience. However, the widespread application of AIGC has also brought about a series of serious risks. AIGC technology has made it extremely easy to generate realistic images. This has led to the proliferation of false information, malicious tampering, and other harmful activities. Additionally, by generating forged images, malicious attackers may impersonate others to carry out fraud, causing economic losses and a crisis of social trust [51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Current deepfake image detection methods primarily rely on extracting handcrafted features, deep features, and frequency domain features for forgery identification. Handcrafted feature-based approaches [10, 18] typically extract statistical features such as color histograms, local binary patterns (LBP), and histograms of oriented gradients (HOG) to detect fake images. While these methods may be effective for early deepfake images, they exhibit poor adaptability to emerging deepfake techniques, such as images generated by diffusion models, making them less effective against continuously evolving attack strategies. Methods based on deep features [17, 41] typically utilize deep neural networks such as CNN, ResNet, and EfficientNet to extract high-level features from images for classification and discrimination. While these approaches can improve detection performance to some extent, they often learn the semantic features of the image itself rather than focusing specifically on forgery traces, which affects the generalization ability of the forgery detection model. Frequency domain-based methods [14, 47] extract features from high-frequency or specific frequency bands using frequency domain analysis techniques such as Fourier transform and wavelet transform to detect forgery traces. However, since different AIGC generation methods may use different generation strategies, the stability of frequency domain features is insufficient, which limits the generalization ability of these methods." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The features extracted by the most of aforementioned" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.416, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. † Equal contribution to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17067" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "methods not only contain forgery trace characteristics but also include a large amount of semantic information from the image itself, which interferes with the accuracy and generalization ability of forgery detection models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.154, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Regarding the generative mechanisms of contemporary deep image synthesis models, these models create images by assimilating particular data distributions. However, their gradient-based optimization processes inherently display a propensity for \"shortcut learning\"—favoring the acquisition of commonly encountered feature distributions while neglecting those that are more infrequent. This phenomenon is driven by the observation that forged image features are easily assimilated, whereas the subtle, rare features inherent in genuine images often evade capture." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.485, + 0.639 + ], + "angle": 0, + "content": "This paper adopts a perspective of forgery image distribution modeling, using generative diffusion models to obtain image temporal distribution features. Then, we mine forgery trace features from the temporal distribution feature space that do not contain semantic information interference, aiming to achieve high accuracy and high generalization in image forgery detection. By leveraging the mapped latent temporal distribution of images, this paper elucidates the evolving patterns of temporal distribution features, identifying them as temporal forgery indicators. In the spatial dimension, based on the hypothesis that forgery images tend to model the approximate semantic parts with either exactly the same or completely different distributions, we extract the correlation between the temporal features of corresponding image regions as spatial forgery traces. Furthermore, generated images typically expose forgery traces more easily in high-frequency regions. We obtain high-frequency information as weights to help the model more quickly identify spatio-temporal forgery traces. This approach provides an innovative perspective for research in the detection of forged images using deep learning models. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.642, + 0.483, + 0.761 + ], + "angle": 0, + "content": "- We introduce an innovative spatial-temporal forgery trace extraction framework designed for the identification of forged images through latent diffusion distribution. Rather than relying on mixed features, the temporal evolution patterns of individual temporal features and the interrelationships among multiple temporal features are modeled as temporal and spatial forgery traces, enabling robust and precise forgery identification." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.483, + 0.822 + ], + "angle": 0, + "content": "- Instead of utilizing frequency features directly, high-frequency components are employed as weighting factors to enhance the forgery detection model's ability to swiftly localize spatio-temporal forgery traces." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.823, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate that the proposed method surpasses the limitations of prior approaches in feature extraction, and exhibits exceptional robustness and generalization capabilities against various forgery techniques." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.642, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.755, + 0.132 + ], + "angle": 0, + "content": "2.1. Image Forgery Technology" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.907, + 0.472 + ], + "angle": 0, + "content": "The field of image generation was initially led by technologies based on Generative Adversarial Networks (GANs) and autoregressive models. Through the continuous evolution of these technologies, milestone products such as BigGAN [3] and DALL-E [13] emerged. However, these methods face certain limitations in content generation control, with stability and complexity issues being particularly prominent. This has led to the emergence of a new type of image generation method—the diffusion model. The Denoising Diffusion Probabilistic Model (DDPM) [19], inspired by nonequilibrium thermodynamics, was born and demonstrated excellent generative performance, further driving in-depth research on diffusion models. Subsequently, researchers have focused on optimizing its structure and improving sampling efficiency. The Latent Diffusion Model (LDM) [38] applies the diffusion process in the latent space, not only improving model efficiency but also enhancing the ability to generate text-conditioned images through a cross-attention mechanism. LDM has become a significant advancement in the field of image generation, leading to the development of well-known generative models such as Stable Diffusion [38] and LCM-RoLA [27]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.742, + 0.5 + ], + "angle": 0, + "content": "2.2. Image Forgery Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.612 + ], + "angle": 0, + "content": "For traditional Photoshop-based forgery methods, early research primarily relied on extracting handcrafted features, deep features, and frequency-domain features for forgery identification. For deepfake images, different approaches have been explored based on distinct feature perspectives. Below, we introduce related works from three aspects: temporal, spatial, and frequency domains." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.909, + 0.839 + ], + "angle": 0, + "content": "In forgery detection methods based on spatial features, Sha et al. [39] analyze unique characteristics of generated images and perform model attribution, aiding in source identification and combating the misuse of forged images. The method in [45] detects forged images by comparing the input image with its reconstruction from a diffusion model, demonstrating strong performance in detecting forgeries under unknown diffusion models and various perturbations. Overall, spatial-based forgery detection methods mainly extract image features using deep models. However, these extracted features often mix semantic information with forgery traces, where the presence of semantic features increases the learning difficulty and reduces detection accuracy, ultimately leading to poor generalization performance of the detection models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In forgery detection methods based on temporal features, [29] effectively identifies images generated by diffusion models by leveraging the deterministic inversion and denoising computation errors of diffusion models, combined" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "with statistical and neural network techniques. Similarly, [9] analyzes forensic features of GAN-generated images, which can also be used to evaluate the performance of existing GAN detectors on diffusion model images, further improving detection accuracy. Although forgery detection methods based on temporal features utilize diffusion models to extract temporal characteristics, they still rely on supervised signals to extract mixed features from the aggregated temporal features. Like spatial feature-based forgery detection, these methods suffer from interference with model accuracy and poor generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.454 + ], + "angle": 0, + "content": "In forgery detection based on image frequency domain features, [46] extracts frequency domain features using the Fourier transform and finds that forged images often exhibit different frequency distributions compared to real images. A convolutional neural network is then used to classify high-frequency features. [42] enhances detection performance by learning forgery-specific features in the frequency domain space and analyzing the abnormal patterns of forged images across different frequency bands. These methods determine authenticity based on the quantity and distribution of high-frequency features. However, due to the diversity of image types, high-frequency features also vary, leading to limited cross-domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.62 + ], + "angle": 0, + "content": "The methods mentioned above are mostly designed for GAN-generated images. Although Corvi et al. [8, 9, 36] also found that spectral artifacts may exist in diffusion-generated images, the effectiveness of spectral artifact-based forgery detection methods in practice remains limited. To address this issue, researchers have begun to analyzing the characteristics of the diffusion generation process itself. For example, DIRE [45] and De-fake [39] use the reconstruction error of the diffusion model to detect forged images, while LaRE2 [28] and AEROBLADE [37] focus on the reconstruction error in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.786 + ], + "angle": 0, + "content": "However, these methods are highly dependent on pretrained reconstruction models and are susceptible to the limitations of the training data characteristics. For example, if the reconstruction model is trained on a specific category (such as cats) but is used to detect forged images from different categories (such as dogs or plants), it may result in misclassification. The core issue here lies in using reconstruction errors as the basis for detection, without fully exploring the characteristics of the diffusion generation mechanism itself, thus affecting the understanding of the forgery image generation patterns." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.229, + 0.815 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we define the key notations of the latent diffusion model and outline its basic principles. In LDM [27], given an image \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) in the RGB space (\\(H\\) and \\(W\\) denote the height and width of the image \\(x\\)), the encoder \\(\\mathcal{E}\\) encodes \\(x\\) into a latent representation \\(z = \\mathcal{E}(x)\\)," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "while the decoder \\(\\mathcal{D}\\) reconstructs the image from the latent representation, yielding \\(\\tilde{x} = \\mathcal{D}(z) = \\mathcal{D}(\\mathcal{E}(x))\\), where ideally \\(z \\in \\mathbb{R}^{h \\times w \\times c}\\) (\\(h, w\\), and \\(c\\) denote the height, width, and channel of the latent representation \\(z\\)). This diffusion model consists of a \\(T\\)-step diffusion process \\(q\\) for \\(1 \\leq t \\leq T\\) and a denoising process, which is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.192, + 0.905, + 0.211 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.219, + 0.905, + 0.236 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.905, + 0.332 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is the diffusion result at time step \\( t \\), \\( \\beta_{t} \\) is the noise factor at time step \\( t \\), \\( I \\) is the identity matrix, and \\( \\mu_{\\theta} \\) and \\( \\Sigma_{\\theta} \\) represent the mean and covariance matrix of the denoising distribution, respectively. After training the model, step-by-step denoising can be performed starting from the standard normal distribution \\( z_{T} \\sim \\mathcal{N}(0, I) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.34, + 0.905, + 0.372 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {1}{\\sqrt {1 - \\beta_ {t}}} \\left(z _ {t} - \\beta_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) + \\sigma_ {t} \\xi , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.545 + ], + "angle": 0, + "content": "where \\(\\xi \\sim \\mathcal{N}(0, I)\\) is the random noise used to control sampling quality, and \\(\\sigma_t\\) regulates the randomness of the denoising process. Through repeated iterations, the final representation \\(z_0\\) is obtained, which is then decoded by \\(\\mathcal{D}\\) to reconstruct the synthesized image. During the forward diffusion process, high-frequency components are gradually suppressed by noise. By progressively adding noise in the forward process and denoising in the reverse process, high-frequency components become increasingly compressed. Notably, at high noise levels, the loss of high-frequency details becomes inevitable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.605, + 0.573 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.841 + ], + "angle": 0, + "content": "As depicted in Figure 1, this method employs a diffusion process to map the image into the latent distribution space, subsequently extracting forgery traces from this latent distribution through a comprehensive spatial-temporal analysis. The overall framework consists of the following three main modules: Temporal Prior Correlation Modeling analyzes the latent space representations at different time steps of the diffusion process to capture their temporal variation characteristics. Spatial Correlation Modeling extracts features from different latent dimensions and computes spatial correlations using a self-attention mechanism to model spatial dependencies. Frequency-Enhanced Attention Mechanism utilizes frequency domain information to guide temporal prior correlation computation and spatial correlation analysis, thereby expediting the localization of forgery traces and enhancing generalization capabilities by disregarding the interference of inconsequential features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.836, + 0.866 + ], + "angle": 0, + "content": "4.1. Image Latent Distribution Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To overcome the limitations of existing methods and gain a deeper understanding of the underlying differences between" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17069" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.912, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.908, + 0.389 + ], + "angle": 0, + "content": "Figure 1. The framework of STFT. It begins by mapping the input image into a latent diffusion distribution space. Subsequently, temporal and spatial forgery traces are extracted by capturing patterns of temporal variation and spatial correlations within the latent diffusion distribution. High-frequency components are utilized as weight to facilitate the rapid localization of forgery traces." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.482, + 0.549 + ], + "angle": 0, + "content": "forged and normal images, we propose a forgery detection approach from a distributional perspective. Considering that LDM maps data into a low-dimensional latent space \\( z_{t} \\) through an encoder \\( \\mathcal{E} \\), making the distribution \\( p(z) \\) more compact and easier to model, while the high-dimensional pixel space \\( p(x) \\) remains complex and difficult to model directly, we extract the latent representations \\( z \\in \\mathbb{R}^{h \\times w \\times c} \\) during the diffusion process of the latent diffusion model for further analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.673 + ], + "angle": 0, + "content": "To facilitate the modeling of distributional changes from a temporal perspective, we reshape the latent representation \\( z_{t} \\) at each diffusion step into a one-dimensional vector \\( \\vec{z}_{t} \\in \\mathbb{R}^{1 \\times d} \\) as sequential data at time step \\( t \\), where \\( d = h \\times w \\times c \\), \\( t \\in \\{1,2,\\dots,T\\} \\). The sequential vectors from step 1 to \\( T \\) are concatenated vertically into a group of tokens \\( X = \\{\\vec{z}_{t}\\}_{t=1}^{T} \\in \\mathbb{R}^{T \\times d} \\), which serves as input for both temporal and frequency-domain modeling." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.682, + 0.418, + 0.699 + ], + "angle": 0, + "content": "4.2. Temporal Prior Association Modeling" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Based on the assumption that the latent space distribution of forged images tends to be smoother with weaker temporal variations, we draw inspiration from the unsupervised temporal anomaly representation learning [48] and employ a trainable Gaussian kernel to compute prior associations between time steps. We introduce an adjustable scale parameter \\(\\sigma\\) to adapt to different temporal sequence patterns. The unimodal property of the Gaussian kernel enhances the association between local time steps, making the non-stationary characteristics of real images more prominent, while the smooth evolution pattern of forged images becomes more distinguishable. This improves the temporal modeling capability for forgery trace detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.492 + ], + "angle": 0, + "content": "For the latent space representations \\(\\{\\vec{z}_t\\}_{t=1}^T\\) at different time steps, we treat the diffusion process as a sequential series, concatenate them into \\(X \\in \\mathbb{R}^{T \\times d}\\), and compute prior associations between time steps. Specifically, we define the prior association matrix \\(P\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.498, + 0.907, + 0.533 + ], + "angle": 0, + "content": "\\[\nP = \\operatorname {R e s c a l e} \\left(\\frac {1}{\\sqrt {2 \\pi} \\sigma_ {t}} \\exp \\left(- \\frac {\\left| t ^ {\\prime} - t \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.65 + ], + "angle": 0, + "content": "where \\(\\sigma_{t}\\) represent the learnable scale parameters for the \\(t\\)-th time step. The prior association is computed using a Gaussian kernel, with \\(G(|t' - t|; \\sigma_t) = \\frac{1}{\\sqrt{2\\pi} \\sigma_t} \\exp \\left( -\\frac{|t' - t|^2}{2 \\sigma_t^2} \\right)\\) representing the association weight calculated based on the distance \\(|t' - t|\\) between time points \\(t\\) and \\(t'\\). The function Rescale(\\cdot) normalizes the association weights into a discrete distribution \\(P\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.776, + 0.675 + ], + "angle": 0, + "content": "4.3. Spatial Correlation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Forged images often exhibit abnormal spatial similarities (e.g., excessive symmetry) or inconsistencies (e.g., color or texture mismatches). To address this, we extract features from each dimension of the latent space and apply a self-attention mechanism to compute spatial correlations within the latent space. This helps amplify regions with abnormal similarities or inconsistencies in forged images, thereby enhancing forgery trace detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.8, + 0.909, + 0.877 + ], + "angle": 0, + "content": "For the feature matrix \\(X\\), the vertical vector \\(u_{i} = X[(:,i]\\) where \\(u_{i}\\in \\mathbb{R}^{T\\times 1}\\) and \\(i\\in \\{1,2,3,\\dots ,d\\}\\), represents the pixel- or region-wise temporal features. Consequently, the spatial token group \\(X^{\\top} = \\{u_{i}^{\\top}\\}_{i = 1}^{d}\\in \\mathbb{R}^{d\\times T}\\) can be utilized to compute the spatial correlation matrix \\(S\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.883, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nQ, K, V = X ^ {\\top} W _ {Q}, X ^ {\\top} W _ {K}, X ^ {\\top} W _ {V}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17070" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.165, + 0.089, + 0.484, + 0.126 + ], + "angle": 0, + "content": "\\[\nS = \\left(S o f t m a x \\left(\\frac {Q K ^ {\\top}}{\\sqrt {T}}\\right) V\\right) ^ {\\top} A, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.281 + ], + "angle": 0, + "content": "where \\(\\top\\) denotes the transpose operation, and \\(Q,K,V\\) along with \\(W_{Q},W_{K},W_{V}\\) represent the query, key, and value matrices and their corresponding parameters, respectively. The term \\(A\\) signifies the frequency-enhanced attention weight, which directs the computation of the spatial correlation matrix to emphasize pixel- or region-wise temporal vectors exhibiting significant dynamic evolution patterns. Further elaboration and the corresponding equations for the frequency-enhanced attention matrix \\(A\\) are discussed in the subsequent subsection." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.288, + 0.465, + 0.304 + ], + "angle": 0, + "content": "4.4. Frequency-Enhanced Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.521 + ], + "angle": 0, + "content": "In frequency domain analysis, the high-frequency details of forged images tend to be eliminated during gradient optimization, resulting in significant differences in frequency characteristics compared to real images. To address this, we extract the frequency-domain information of latent space features using the Fast Fourier Transform (FFT) and compute frequency-domain correlations to capture cross-scale feature interactions. Since high-frequency information is lost in forged images, the frequency distribution of real and forged images exhibits distinct patterns after conjugate product computation. We then apply the Inverse Fourier Transform (IFFT) to map the correlation back to the spatiotemporal domain, allowing the model to focus on anomalies caused by missing frequency-domain information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.612 + ], + "angle": 0, + "content": "For the group of tokens \\(X = \\{\\vec{z}_t\\}_{t=1}^T\\), the self-attention with query matrix \\(\\overline{Q}\\) and key matrix \\(\\overline{K}\\) for \\(X\\) is adopted to find the feature temporal similarity of different tokens \\(\\{\\vec{z}_t\\}_{t=1}^T\\). Furthermore, we perform the Fourier Transform on the query matrix \\(\\overline{Q}\\) and key matrix \\(\\overline{K}\\) to extract frequency-domain features:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.619, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\widehat {Q} = \\mathcal {F} (\\overline {{Q}}), \\widehat {K} = \\mathcal {F} (\\overline {{K}}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.705 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}()\\) denotes the Fast Fourier Transform. Next, we compute the conjugate product of the query and key in the frequency domain to obtain cross-scale interaction information \\(C\\):" + }, + { + "type": "equation", + "bbox": [ + 0.242, + 0.704, + 0.482, + 0.722 + ], + "angle": 0, + "content": "\\[\nC = \\widehat {Q} \\cdot \\widehat {K} ^ {*}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.864 + ], + "angle": 0, + "content": "where “.” denotes element-wise multiplication of matrices, \\(\\widehat{K}^*\\) represents the complex conjugate, ensuring correct interaction in the frequency domain and highlighting the different frequency distributions of real and forged images. Then, we apply the Inverse Fourier Transform \\(\\mathcal{F}^{-1}\\) to restore the result to the spatio-temporal domain and perform Softmax normalization to ensure that the weights maintain appropriate scales in the spatio-temporal domain, yielding the frequency-enhanced spatio-temporal attention weights:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.871, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F} ^ {- 1} (C)}{\\sqrt {T}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Finally, \\(A\\) serves as the frequency-enhanced temporal attention weight to improve generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.131, + 0.691, + 0.148 + ], + "angle": 0, + "content": "4.5. Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.258 + ], + "angle": 0, + "content": "First, we assess the distribution characteristics of the input time series \\(X\\) across multiple levels by computing the association difference \\((AssD)\\) between the prior association \\(P\\) and the frequency-domain association \\(A\\). Specifically, we employ the Kullback-Leibler divergence (KL divergence) to quantify the discrepancy between two discrete distributions, defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.269, + 0.903, + 0.288 + ], + "angle": 0, + "content": "\\[\nA s s D (P, A) = K L (P [ t,: ] \\| A [ t,: ]) + K L (A [ t,: ] \\| P [ t,: ]),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.906, + 0.403 + ], + "angle": 0, + "content": "where \\(KL(A[t,:]\\| P[t,:])\\) denotes the KL divergence, measuring the informational disparity between distributions \\(A[t,:]\\) and \\(P[t,:]\\). Here, \\(P[t,:]\\) and \\(A[t,:]\\) represent the prior association and frequency-domain association distributions at time step \\(t\\), respectively. The final association difference \\(AssD(P,A)\\) captures the overall variation in association patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.478 + ], + "angle": 0, + "content": "For the output of the spatial correlation modeling module, we introduce a global optimization objective based on cross-entropy loss to steer the spatial correlation and frequency-domain enhancement modules in distinguishing between forged and authentic images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.568 + ], + "angle": 0, + "content": "We formulate distinct optimization objectives tailored to different sample types. For forged images, we enforce a deviation between the prior association \\( P \\) and the frequency-domain association \\( A \\), thereby accentuating the intricate and challenging-to-model characteristics of the real image distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.579, + 0.905, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f a k e}} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.772 + ], + "angle": 0, + "content": "where the first term represents the cross-entropy loss between the classification output \\(\\hat{y}\\) of the spatial correlation modeling module and the ground truth labels \\(y\\), enhancing the module's classification performance. The second term regulates the relationship between the prior association \\(P\\) and the frequency-domain association \\(A\\) using the association difference \\(AssD()\\) for forged images, thereby improving the model's temporal modeling capability. Here, \\(A_{fix}\\) represents the fixed frequency-domain association. The hyperparameter \\(\\lambda\\) modulates the contribution of the \\(AssD()\\) loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.832 + ], + "angle": 0, + "content": "For real images, we allow the prior association \\( P \\) to closely align with the frequency-domain association \\( A \\), emphasizing the smoothness of forged images in the frequency domain as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.844, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e a l}} = C E (y, \\hat {y}) + \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where the second term ensures that the prior association \\(P\\) remains distant from the frequency-domain association \\(A\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17071" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.092, + 0.452, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.284, + 0.428, + 0.297 + ], + "angle": 0, + "content": "(a) Performance comparisons on Genimage (ACC, %)." + }, + { + "type": "image", + "bbox": [ + 0.549, + 0.089, + 0.885, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.284, + 0.868, + 0.297 + ], + "angle": 0, + "content": "(b) Performance comparisons on Deepfacegen (AUC, %)." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.308, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. The radar chart shows the forgery identification performance of STFT in comparison to 12 alternative methods on Genimage (a) and 10 other methods on Deepfacegen (b). Solid lines depict the performance of the comparative methods, while a dashed line signifies the performance of STFT." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.483, + 0.408 + ], + "angle": 0, + "content": "for real images, thereby enhancing the model's discriminative capability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.408, + 0.484, + 0.483 + ], + "angle": 0, + "content": "For a batch comprising both forged and real images, in addition to the aforementioned losses \\(\\mathcal{L}_{\\text{fake}}\\) and \\(\\mathcal{L}_{\\text{real}}\\), we ensure that \\(A\\) remains distinct from \\(P\\), thereby directing greater attention to non-adjacent regions and enhancing the flexibility of global associations:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.496, + 0.483, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m i x} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P _ {f i x}, A) \\|, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.525, + 0.484, + 0.616 + ], + "angle": 0, + "content": "where \\( P_{fix} \\) denotes the fixed prior association matrix. This alternating optimization strategy stabilizes the training process and utilizes the frequency-domain association \\( A \\) to guide the model in focusing on more nuanced frequency-domain details, effectively distinguishing between genuine and forged images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.631, + 0.226, + 0.648 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.484, + 0.747 + ], + "angle": 0, + "content": "In the first part of this section, we introduce the detailed experimental setup. Next, in the second part, we evaluate and compare STFT with other forgery detection methods, including SOTA approaches. In the fourth part, we assess the generalization performance of STFT. Finally, we conduct an ablation study on the proposed model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.283, + 0.775 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics and Datasets. In alignment with the established comparison standards, we evaluate accuracy (ACC) and area under the curve (AUC) as performance metrics for the GenImage [53] and DeepFaceGen [2] datasets, respectively. The GenImage [53] dataset consists of a total of 2,681,167 images, including 1,331,167 real images and 1,350,000 generated images. These generated images are sourced from eight different generative models: BigGAN" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.528 + ], + "angle": 0, + "content": "[3], GLIDE [30], VQDM [16], Stable Diffusion V1 and V1.5 [38], ADM [12], Midjourney [21], and Wukong [15]. All fake images are generated using the template prompt \"photo of CLS,\" where \"CLS\" is replaced by one of the 1,000 labels from ImageNet [11]. The dataset is divided into eight subsets, each containing a portion of real images along with all fake images generated by the aforementioned models. In this study, we follow the official dataset split, using 2,581,167 images for training and the remaining 100,000 images for validation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.685 + ], + "angle": 0, + "content": "DeepFaceGen [2] is a facial forgery dataset that includes both video and image modalities. We select face forgery data generated using novel prompt-guided forgery methods. These methods cover 12 diffusion- and autoregressive-based models across text-to-image and image-to-image modalities. The dataset includes face forgeries across various ages, genders, and environments, with generated images sourced from popular models such as Midjourney [21], DALL-E [13], and SDXL [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We downloaded the pretrained LDM model, fed images into it, and extracted the latent space data. We set the default diffusion steps to 20, obtaining 20 latent representations of size \\(3 \\times 64 \\times 64\\). Each latent representation was then processed using average pooling to reduce its size to \\(3 \\times 8 \\times 8\\) and flattened into a one-dimensional vector, ultimately forming sequential data of size \\(20 \\times 192\\). We set the batch size to 256, the number of channels in the hidden state to 512, and the number of attention heads to 8. The hyperparameter \\(\\lambda\\) was set to 2 for all datasets. We used the ADAM optimizer [23] with an initial learning rate of \\(1 \\times 10^{-4}\\). All experiments were conducted on an Nvidia A40 GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17072" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.205 + ], + "angle": 0, + "content": "
MethodYearMidjourneySDv1.4SDv1.5ADMGLIDEWukongVQDMBigGAN
DIRE [45]202399.5 / 72.699.7 / 68.899.7 / 68.899.2 / 67.999.8 / 52.299.7 / 68.199.9 / 51.9100 / 50.1
Defake [39]202398.8 / 78.799.8 / 81.5100 / 82.398.7 / 82.599.8 / 73.199.3 / 75.699.1 / 82.098.7 / 76.6
PatchCraft [52]202390.1 / 81.189.5 / 81.390.1 / 81.390.4 / 79.390.6 / 80.492.6 / 79.390.3 / 74.980.9 / 73.7
AIDE [49]202498.9 / 86.099.7 / 84.999.6 / 84.095.3 / 79.197.6 / 81.898.8 / 80.492.3 / 83.096.3 / 81.7
DRCT [6]202499.9 / 89.395.5 / 86.696.8 / 89.195.1 / 81.896.6 / 83.899.6 / 84.494.3 / 82.498.8 / 84.0
STFT (Ours)-99.4 / 92.399.6 / 89.199.3 / 90.499.5 / 82.599.8 / 87.0100 / 84.899.7 / 85.7100 / 85.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.214, + 0.907, + 0.258 + ], + "angle": 0, + "content": "Table 1. Performance comparison (ACC, %) of various methods. Each value is presented as (Score1 / Score2), where Score1 corresponds to the accuracy achieved when trained and tested on the same subset, and Score2 represents the average accuracy when trained on the current subset and evaluated across other subsets. Bold values highlight the best performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.391, + 0.299 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.626 + ], + "angle": 0, + "content": "GenImage Benchmark. To verify the model's generalization ability across different generative models, this section follows the GenImage evaluation standard and compares our approach with 12 mainstream forgery detection models, including CNNSpot [34], F3Net [33], CLIP/RN50 [34], GramNet [24], De-fake [39], Conv-B [26], SwinT [25], UnivFD [31], DIRE [45], PathCraft [52], AIDE [49], and DRCT [6]. According to the GenImage evaluation protocol, all models were trained on the SDv1.4 subset of GenImage and evaluated on multiple test sets. As shown in Figure 2a, nearly all methods achieve high accuracy on the Wukong, SDv1.4, and SDv1.5 datasets but perform poorly on GLIDE, VQDM, Midjourney, ADM, and BigGAN datasets. Since BigGAN is a non-diffusion model dataset, all methods perform the worst on it. In contrast, our proposed STFT achieves over \\(89\\%\\) accuracy on the BigGAN dataset and obtains the highest ACC on 7 out of 8 sub-datasets, demonstrating a significant advantage over existing forgery detection methods (see Appendix Table 4 for complete results)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "DeepFaceGen Benchmark. Following the DeepFaceGen evaluation standard, we compare STFT with mainstream generative image detection methods, including Xception [7], EfficientNet-B0 [44], F3-Net [33], RECCE [4], DNADet [50], DIRE [45], DRCT [6], UnivFD [31], NPR [43], and FreqNet [42]. Training and testing were conducted on the same dataset, which includes real images and images generated by 12 different forgery methods such as DALL-E and SD. The dataset was split into training, validation, and test sets with a 7:1:2 ratio. As shown in Figure 2b, STFT outperforms all methods across all 12 subsets, with a particularly notable advantage in distinguishing difficult-to-detect forged images generated by DALL-E 1, DALL-E 3, Midjourney, and Wenxin. Overall, STFT achieves an AUC above \\(90\\%\\) for all forgery categories, with an average AUC of \\(95.21\\%\\) (detailed results are provided in Appendix Table 5)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.28, + 0.909, + 0.378 + ], + "angle": 0, + "content": "
Train OnCatHorseBedroom
Test OnHorse/BedroomCat/BedroomCat/Horse
DIRE [45]73.0165.4555.04
DRCT [6]74.4868.7867.62
STFT (Ours)90.3289.7091.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.386, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Table 2. Generalization performance (AUC,%) of different methods on non-face images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.441, + 0.768, + 0.456 + ], + "angle": 0, + "content": "5.3. Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.538 + ], + "angle": 0, + "content": "To evaluate the generalization capability of STFT, we designed two experimental scenarios: first, testing the model when the pre-trained model does not align with the target detection task, and second, assessing its adaptability to unknown forgery methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Adaptation to Unknown Forgery Methods. We trained STFT on a single subset of the GenImage dataset and tested it across all eight subsets. As shown in Table 1, we compared STFT with the top four models from the baseline evaluations. STFT outperformed the existing SOTA method in GenImage, DRCT, and achieved the best generalization performance across all eight subsets. When confronted with unseen forgery methods, all models experienced performance degradation in various detection metrics. However, STFT still demonstrated strong generalization ability. The superior generalization of STFT is primarily attributed to its frequency-domain-enhanced spatio-temporal correlation modeling, which captures multi-perspective features of forged images. This enhances the model's adaptability to different types of forgeries, allowing it to effectively recognize anomalous patterns even when facing previously unseen forgery methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Generalization Performance on Non-face Images. Reconstruction-error-based forgery detection methods often fail when the test data differs from the reconstruction model's pre-training domain. For instance, applying a model trained on animals to plant images may misclassify real images as fake due to high reconstruction errors. To assess STFT's robustness in such cases, we used pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17073" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.482, + 0.166 + ], + "angle": 0, + "content": "
Model VariantTemporalSpatialFrequencyAUC(%)
Full Model✓✓✓92.02
w/o Temporal✓✓86.28
w/o Spatial✓✓79.71
w/o Frequency✓✓89.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.149, + 0.177, + 0.42, + 0.19 + ], + "angle": 0, + "content": "Table 3. Ablation study on different modules." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.324 + ], + "angle": 0, + "content": "models from [12] on LSUN Bedroom, Cat, and Horse categories, and compared STFT with DIRE and DRCT. Following the Benchmark 1 protocol on 12 prompt-based forgery datasets, results in Table 2 show that while reconstruction-based methods suffered major performance drops, STFT consistently achieved AUC scores over \\(90\\%\\), highlighting its strong generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.335, + 0.245, + 0.35 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.446 + ], + "angle": 0, + "content": "In this section, we conduct ablation experiments to quantify the contribution of each design component in our model and analyze the impact of key hyperparameters. Specifically, we examine the effect of the latent space sampling steps and feature map resolution and the inclusion of the frequency-enhanced attention module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.567 + ], + "angle": 0, + "content": "Effect of Latent Space Sampling Steps and Feature Map Resolution. To investigate the impact of sampling steps and feature resolution in latent space, we evaluate detection performance with sampling steps set to 10, 15, 20, 25, and 30, and test four resolutions by downsampling latent features via average pooling. As shown in Figure 3, the model performs best with 20 sampling steps and a feature resolution of \\(3 \\times 8 \\times 8\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.689 + ], + "angle": 0, + "content": "This indicates that 20 steps sufficiently preserve temporal cues while preventing over-smoothing. A \\(3 \\times 8 \\times 8\\) resolution balances detail retention and generalization. Lower resolutions (e.g., \\(3 \\times 4 \\times 4\\)) lose crucial spatial information, reducing accuracy, while higher resolutions introduce redundancy, affecting robustness. Thus, 20 sampling steps and \\(3 \\times 8 \\times 8\\) resolution are adopted as default settings for optimal accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Impact Analysis of Model Components. To analyze the contributions of different STFT components, we conduct ablation experiments by sequentially removing the temporal prior correlation module, spatial correlation module, and frequency-enhanced attention module on DALL-E 3 Dataset. The results, presented in Table 3, confirm their complementary roles in forgery detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Removing the temporal module significantly weakens the model's ability to capture sequential forgery patterns, leading to a notable performance drop. This aligns with our observation that latent diffusion smooths temporal transitions in forged images. Without spatial correlation modeling, the model struggles to detect local modifications, increasing false positives and reducing robustness. Frequency" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.698, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.906, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.238, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Figure 3. The ablation study on diffusion steps (a) and latent feature map size (b). Solid lines in both cases represent the performance results corresponding to the final adopted values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.368 + ], + "angle": 0, + "content": "features are crucial for detecting high-frequency artifacts in forgeries. Removing this module degrades performance, especially against unknown forgery methods, highlighting its role in generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.634, + 0.397 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.572 + ], + "angle": 0, + "content": "This paper proposes the Spatial-Temporal Forgery Trace (STFT) method for forgery image identification. Leveraging the tendency of forged images to fit specific distributions more easily than real ones, STFT employs generative diffusion models to capture temporal distributions and models relationships between temporal and spatial features as forgery traces. It also integrates frequency-domain features to accelerate trace localization. Experiments show that STFT effectively identifies subtle spatio-temporal traces by combining spatial, temporal, and frequency perspectives, demonstrating strong robustness and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.74 + ], + "angle": 0, + "content": "Limitations and Future Works. Although the proposed STFT achieves high accuracy and demonstrates commendable generalization capabilities, it is not without its limitations. Firstly, the process of mapping the original image into the latent diffusion distribution is computationally intensive and time-consuming. Additionally, the method appears to exhibit suboptimal performance on GAN-based forgery images, likely due to fundamental differences in the generative mechanisms of GANs and generative diffusion models. In future work, we aim to enhance the efficiency of the proposed method and further extend its generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.753, + 0.693, + 0.77 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.839 + ], + "angle": 0, + "content": "This work is funded by National Key Research and Development Project (Grant No: 2022YFB2703100), Ant Group through CCF-Ant Research Fund, and Information Technology Center, ZheJiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17074" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] Baidu AI. Wenxin large model for text-to-image generation. Baidu AI Blog, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.146, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Yijun Bei, Hengrui Lou, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Song, Mingli Song, and Zunlei Feng. A largescale universal evaluation benchmark for face forgery detection. arXiv preprint arXiv:2406.09181, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4113-4122, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.483, + 0.373 + ], + "angle": 0, + "content": "[5] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In Forty-first International Conference on Machine Learning, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[7] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[8] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 973-982, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[9] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[10] Edmar RS De Rezende, Guilherme CS Ruppert, Antonio Theophilo, Eric K Tokuda, and Tiago Carvalho. Exposing computer generated images by using deep convolutional neural networks. Signal Processing: Image Communication, 66: 113-126, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Tira Nur Fitria. Artificial intelligence (ai) technology in openai chatgpt application: A review of chatgpt in writing english essay. In ELT Forum: Journal of English Language Teaching, pages 44-58, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[14] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Niu Minzhe, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, et al. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems, 35:26418-26431, 2022. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[17] Weinan Guan, Wei Wang, Jing Dong, Bo Peng, and Tieniu Tan. Collaborative feature learning for fine-grained facial forgery detection and segmentation. arXiv preprint arXiv:2304.08078, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[18] Tahir Hassan, Aras Asaad, Dashti Ali, and Sabah Jassim. Artificial image tampering distorts spatial distribution of texture landmarks and quality characteristics. arXiv preprint arXiv:2208.02710, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[21] Anna Jaruga-Rozdolska. Artificial intelligence as part of future practices in the architect's work: Midjourney generative tool as part of a process of creating an architectural form. Architectus, (3 (71):95-104, 2022. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[22] Wei et al. Jiang. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16423-16432, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[24] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17075" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[28] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[29] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. arXiv preprint arXiv:2307.06272, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[31] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] OpenAI. Dall-e 3: Openai's latest text-to-image model. OpenAI Blog, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[33] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Aditya et al. Ramesh. Zero-shot text-to-image generation. arXiv preprint arXiv:2102.12092, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[36] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. arXiv preprint arXiv:2210.14571, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[37] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[39] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text-to-image generation models. In Proceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3418-3432, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.858, + 0.35, + 0.871 + ], + "angle": 0, + "content": "[40] Stability.ai. Stable Diffusion, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Eren Tahir and Mert Bal. Deep image composition meets image forgery. arXiv preprint arXiv:2404.02897, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[42] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5052-5060, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[43] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[46] Jing Weng. Local frequency analysis for diffusion-generated image detection. In International Conference on Image Processing and Artificial Intelligence (ICIPAL 2024), pages 66-73. SPIE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[47] Moritz Wolter, Felix Blanke, Raoul Heese, and Jochen Garcke. Wavelet-packets for deepfake image analysis and detection. Machine Learning, 111(11):4295-4327, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.528 + ], + "angle": 0, + "content": "[48] Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. arXiv preprint arXiv:2110.02642, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[49] Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection. arXiv preprint arXiv:2406.19435, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[50] Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4662-4670, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[51] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[52] Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection. arXiv preprint arXiv:2311.12397, 2023. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.757, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[53] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36:77771-77782, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17076" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_origin.pdf b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1dd5ce1a6abfa20c89aa81463b0c383677cfdd6f --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/b7964ec0-870b-444a-822c-785edaff983a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7249f9355b4992d1b2ce935fefccae076fbdeca48ae34a60ec433be4a2df990 +size 2251091 diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/full.md b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..805a0a1a7370700e931712624b4194117e1c50dd --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/full.md @@ -0,0 +1,321 @@ +# Spatial-Temporal Forgery Trace based Forgery Image Identification + +Yilin Wang $^{1,2\dagger}$ , Zunlei Feng $^{1,2,3\dagger}$ , Jiachi Wang $^{1,2}$ , Hengrui Lou $^{1,2}$ , Binjia Zhou $^{1,2}$ , Jie Lei $^{4}$ , Mingli Song $^{1,2}$ , Yijun Bei $^{1,2*}$ + +1State Key Laboratory of Blockchain and Data Security, Zhejiang University + +$^{2}$ School of Software Technology, Zhejiang University + +3Zhejiang Provincial Key Laboratory of Service Robot + +4College of Computer Science, Zhejiang University of Technology + +# Abstract + +The rapid development of AIGC technology has enabled highly realistic forged images to deceive human perception, posing serious risks across many areas. Current deepfake image detection methods primarily identify forgeries by extracting handcrafted features, deep features, and frequency-domain features. While these features contain forgery traces, they also include a substantial amount of the image's semantic information, which interferes with the precision and generalization of forgery detection models. To tackle these challenges, this paper introduces a novel forgery image identification method based on the Spatial-Temporal Forgery Trace (STFT). Motivated by the fact that forgery images are more easily fitted to a specific distribution than real images, the STFT method approaches the issue from a forged image distribution modeling perspective, employing generative diffusion models to meticulously capture the temporal distribution of images. It further models the relationship between temporal feature variations and spatially corresponding temporal features, treating them as temporal and spatial forgery traces. Moreover, STFT incorporates frequency-domain features as weighting factors to accelerate the localization of spatio-temporal forgery traces. Experiments demonstrate that by integrating spatial, temporal, and frequency perspectives within the latent space, STFT effectively captures subtle spatio-temporal forgery traces, exhibiting strong robustness and generalizability. It outperforms state-of-the-art methods on major benchmark datasets in the field. The source code for STFT is available at https://github.com/GCLion/STFT. + +# 1. Introduction + +With the rapid development of Artificial Intelligence Generated Content (AIGC) technology [5], particularly the ma + +turity of generative models like diffusion models, AIGC has been widely applied in fields such as image and video generation, bringing significant innovation and convenience. However, the widespread application of AIGC has also brought about a series of serious risks. AIGC technology has made it extremely easy to generate realistic images. This has led to the proliferation of false information, malicious tampering, and other harmful activities. Additionally, by generating forged images, malicious attackers may impersonate others to carry out fraud, causing economic losses and a crisis of social trust [51]. + +Current deepfake image detection methods primarily rely on extracting handcrafted features, deep features, and frequency domain features for forgery identification. Handcrafted feature-based approaches [10, 18] typically extract statistical features such as color histograms, local binary patterns (LBP), and histograms of oriented gradients (HOG) to detect fake images. While these methods may be effective for early deepfake images, they exhibit poor adaptability to emerging deepfake techniques, such as images generated by diffusion models, making them less effective against continuously evolving attack strategies. Methods based on deep features [17, 41] typically utilize deep neural networks such as CNN, ResNet, and EfficientNet to extract high-level features from images for classification and discrimination. While these approaches can improve detection performance to some extent, they often learn the semantic features of the image itself rather than focusing specifically on forgery traces, which affects the generalization ability of the forgery detection model. Frequency domain-based methods [14, 47] extract features from high-frequency or specific frequency bands using frequency domain analysis techniques such as Fourier transform and wavelet transform to detect forgery traces. However, since different AIGC generation methods may use different generation strategies, the stability of frequency domain features is insufficient, which limits the generalization ability of these methods. + +The features extracted by the most of aforementioned + +methods not only contain forgery trace characteristics but also include a large amount of semantic information from the image itself, which interferes with the accuracy and generalization ability of forgery detection models. + +Regarding the generative mechanisms of contemporary deep image synthesis models, these models create images by assimilating particular data distributions. However, their gradient-based optimization processes inherently display a propensity for "shortcut learning"—favoring the acquisition of commonly encountered feature distributions while neglecting those that are more infrequent. This phenomenon is driven by the observation that forged image features are easily assimilated, whereas the subtle, rare features inherent in genuine images often evade capture. + +This paper adopts a perspective of forgery image distribution modeling, using generative diffusion models to obtain image temporal distribution features. Then, we mine forgery trace features from the temporal distribution feature space that do not contain semantic information interference, aiming to achieve high accuracy and high generalization in image forgery detection. By leveraging the mapped latent temporal distribution of images, this paper elucidates the evolving patterns of temporal distribution features, identifying them as temporal forgery indicators. In the spatial dimension, based on the hypothesis that forgery images tend to model the approximate semantic parts with either exactly the same or completely different distributions, we extract the correlation between the temporal features of corresponding image regions as spatial forgery traces. Furthermore, generated images typically expose forgery traces more easily in high-frequency regions. We obtain high-frequency information as weights to help the model more quickly identify spatio-temporal forgery traces. This approach provides an innovative perspective for research in the detection of forged images using deep learning models. Our main contributions are as follows: + +- We introduce an innovative spatial-temporal forgery trace extraction framework designed for the identification of forged images through latent diffusion distribution. Rather than relying on mixed features, the temporal evolution patterns of individual temporal features and the interrelationships among multiple temporal features are modeled as temporal and spatial forgery traces, enabling robust and precise forgery identification. +- Instead of utilizing frequency features directly, high-frequency components are employed as weighting factors to enhance the forgery detection model's ability to swiftly localize spatio-temporal forgery traces. +- Comprehensive experiments demonstrate that the proposed method surpasses the limitations of prior approaches in feature extraction, and exhibits exceptional robustness and generalization capabilities against various forgery techniques. + +# 2. Related Work + +# 2.1. Image Forgery Technology + +The field of image generation was initially led by technologies based on Generative Adversarial Networks (GANs) and autoregressive models. Through the continuous evolution of these technologies, milestone products such as BigGAN [3] and DALL-E [13] emerged. However, these methods face certain limitations in content generation control, with stability and complexity issues being particularly prominent. This has led to the emergence of a new type of image generation method—the diffusion model. The Denoising Diffusion Probabilistic Model (DDPM) [19], inspired by nonequilibrium thermodynamics, was born and demonstrated excellent generative performance, further driving in-depth research on diffusion models. Subsequently, researchers have focused on optimizing its structure and improving sampling efficiency. The Latent Diffusion Model (LDM) [38] applies the diffusion process in the latent space, not only improving model efficiency but also enhancing the ability to generate text-conditioned images through a cross-attention mechanism. LDM has become a significant advancement in the field of image generation, leading to the development of well-known generative models such as Stable Diffusion [38] and LCM-RoLA [27]. + +# 2.2. Image Forgery Detection + +For traditional Photoshop-based forgery methods, early research primarily relied on extracting handcrafted features, deep features, and frequency-domain features for forgery identification. For deepfake images, different approaches have been explored based on distinct feature perspectives. Below, we introduce related works from three aspects: temporal, spatial, and frequency domains. + +In forgery detection methods based on spatial features, Sha et al. [39] analyze unique characteristics of generated images and perform model attribution, aiding in source identification and combating the misuse of forged images. The method in [45] detects forged images by comparing the input image with its reconstruction from a diffusion model, demonstrating strong performance in detecting forgeries under unknown diffusion models and various perturbations. Overall, spatial-based forgery detection methods mainly extract image features using deep models. However, these extracted features often mix semantic information with forgery traces, where the presence of semantic features increases the learning difficulty and reduces detection accuracy, ultimately leading to poor generalization performance of the detection models. + +In forgery detection methods based on temporal features, [29] effectively identifies images generated by diffusion models by leveraging the deterministic inversion and denoising computation errors of diffusion models, combined + +with statistical and neural network techniques. Similarly, [9] analyzes forensic features of GAN-generated images, which can also be used to evaluate the performance of existing GAN detectors on diffusion model images, further improving detection accuracy. Although forgery detection methods based on temporal features utilize diffusion models to extract temporal characteristics, they still rely on supervised signals to extract mixed features from the aggregated temporal features. Like spatial feature-based forgery detection, these methods suffer from interference with model accuracy and poor generalization ability. + +In forgery detection based on image frequency domain features, [46] extracts frequency domain features using the Fourier transform and finds that forged images often exhibit different frequency distributions compared to real images. A convolutional neural network is then used to classify high-frequency features. [42] enhances detection performance by learning forgery-specific features in the frequency domain space and analyzing the abnormal patterns of forged images across different frequency bands. These methods determine authenticity based on the quantity and distribution of high-frequency features. However, due to the diversity of image types, high-frequency features also vary, leading to limited cross-domain generalization. + +The methods mentioned above are mostly designed for GAN-generated images. Although Corvi et al. [8, 9, 36] also found that spectral artifacts may exist in diffusion-generated images, the effectiveness of spectral artifact-based forgery detection methods in practice remains limited. To address this issue, researchers have begun to analyzing the characteristics of the diffusion generation process itself. For example, DIRE [45] and De-fake [39] use the reconstruction error of the diffusion model to detect forged images, while LaRE2 [28] and AEROBLADE [37] focus on the reconstruction error in the latent space. + +However, these methods are highly dependent on pretrained reconstruction models and are susceptible to the limitations of the training data characteristics. For example, if the reconstruction model is trained on a specific category (such as cats) but is used to detect forged images from different categories (such as dogs or plants), it may result in misclassification. The core issue here lies in using reconstruction errors as the basis for detection, without fully exploring the characteristics of the diffusion generation mechanism itself, thus affecting the understanding of the forgery image generation patterns. + +# 3. Preliminaries + +In this section, we define the key notations of the latent diffusion model and outline its basic principles. In LDM [27], given an image $x \in \mathbb{R}^{H \times W \times 3}$ in the RGB space ( $H$ and $W$ denote the height and width of the image $x$ ), the encoder $\mathcal{E}$ encodes $x$ into a latent representation $z = \mathcal{E}(x)$ , + +while the decoder $\mathcal{D}$ reconstructs the image from the latent representation, yielding $\tilde{x} = \mathcal{D}(z) = \mathcal{D}(\mathcal{E}(x))$ , where ideally $z \in \mathbb{R}^{h \times w \times c}$ ( $h, w$ , and $c$ denote the height, width, and channel of the latent representation $z$ ). This diffusion model consists of a $T$ -step diffusion process $q$ for $1 \leq t \leq T$ and a denoising process, which is defined as follows: + +$$ +q \left(z _ {t} \mid z _ {t - 1}\right) = \mathcal {N} \left(z _ {t}; \sqrt {1 - \beta_ {t}} z _ {t - 1}, \beta_ {t} I\right), \tag {1} +$$ + +$$ +p _ {\theta} \left(z _ {t - 1} \mid z _ {t}\right) = \mathcal {N} \left(z _ {t - 1}; \mu_ {\theta} \left(z _ {t}, t\right), \Sigma_ {\theta} \left(z _ {t}, t\right)\right), \tag {2} +$$ + +where $z_{t}$ is the diffusion result at time step $t$ , $\beta_{t}$ is the noise factor at time step $t$ , $I$ is the identity matrix, and $\mu_{\theta}$ and $\Sigma_{\theta}$ represent the mean and covariance matrix of the denoising distribution, respectively. After training the model, step-by-step denoising can be performed starting from the standard normal distribution $z_{T} \sim \mathcal{N}(0, I)$ : + +$$ +z _ {t - 1} = \frac {1}{\sqrt {1 - \beta_ {t}}} \left(z _ {t} - \beta_ {t} \epsilon_ {\theta} \left(z _ {t}, t\right)\right) + \sigma_ {t} \xi , \tag {3} +$$ + +where $\xi \sim \mathcal{N}(0, I)$ is the random noise used to control sampling quality, and $\sigma_t$ regulates the randomness of the denoising process. Through repeated iterations, the final representation $z_0$ is obtained, which is then decoded by $\mathcal{D}$ to reconstruct the synthesized image. During the forward diffusion process, high-frequency components are gradually suppressed by noise. By progressively adding noise in the forward process and denoising in the reverse process, high-frequency components become increasingly compressed. Notably, at high noise levels, the loss of high-frequency details becomes inevitable. + +# 4. Method + +As depicted in Figure 1, this method employs a diffusion process to map the image into the latent distribution space, subsequently extracting forgery traces from this latent distribution through a comprehensive spatial-temporal analysis. The overall framework consists of the following three main modules: Temporal Prior Correlation Modeling analyzes the latent space representations at different time steps of the diffusion process to capture their temporal variation characteristics. Spatial Correlation Modeling extracts features from different latent dimensions and computes spatial correlations using a self-attention mechanism to model spatial dependencies. Frequency-Enhanced Attention Mechanism utilizes frequency domain information to guide temporal prior correlation computation and spatial correlation analysis, thereby expediting the localization of forgery traces and enhancing generalization capabilities by disregarding the interference of inconsequential features. + +# 4.1. Image Latent Distribution Modeling + +To overcome the limitations of existing methods and gain a deeper understanding of the underlying differences between + +![](images/6bc072cf4fd90f648d1333f7c61848b229eaf54039847ad2f0313cf0c92bf0b5.jpg) +Figure 1. The framework of STFT. It begins by mapping the input image into a latent diffusion distribution space. Subsequently, temporal and spatial forgery traces are extracted by capturing patterns of temporal variation and spatial correlations within the latent diffusion distribution. High-frequency components are utilized as weight to facilitate the rapid localization of forgery traces. + +forged and normal images, we propose a forgery detection approach from a distributional perspective. Considering that LDM maps data into a low-dimensional latent space $z_{t}$ through an encoder $\mathcal{E}$ , making the distribution $p(z)$ more compact and easier to model, while the high-dimensional pixel space $p(x)$ remains complex and difficult to model directly, we extract the latent representations $z \in \mathbb{R}^{h \times w \times c}$ during the diffusion process of the latent diffusion model for further analysis. + +To facilitate the modeling of distributional changes from a temporal perspective, we reshape the latent representation $z_{t}$ at each diffusion step into a one-dimensional vector $\vec{z}_{t} \in \mathbb{R}^{1 \times d}$ as sequential data at time step $t$ , where $d = h \times w \times c$ , $t \in \{1,2,\dots,T\}$ . The sequential vectors from step 1 to $T$ are concatenated vertically into a group of tokens $X = \{\vec{z}_{t}\}_{t=1}^{T} \in \mathbb{R}^{T \times d}$ , which serves as input for both temporal and frequency-domain modeling. + +# 4.2. Temporal Prior Association Modeling + +Based on the assumption that the latent space distribution of forged images tends to be smoother with weaker temporal variations, we draw inspiration from the unsupervised temporal anomaly representation learning [48] and employ a trainable Gaussian kernel to compute prior associations between time steps. We introduce an adjustable scale parameter $\sigma$ to adapt to different temporal sequence patterns. The unimodal property of the Gaussian kernel enhances the association between local time steps, making the non-stationary characteristics of real images more prominent, while the smooth evolution pattern of forged images becomes more distinguishable. This improves the temporal modeling capability for forgery trace detection. + +For the latent space representations $\{\vec{z}_t\}_{t=1}^T$ at different time steps, we treat the diffusion process as a sequential series, concatenate them into $X \in \mathbb{R}^{T \times d}$ , and compute prior associations between time steps. Specifically, we define the prior association matrix $P$ as follows: + +$$ +P = \operatorname {R e s c a l e} \left(\frac {1}{\sqrt {2 \pi} \sigma_ {t}} \exp \left(- \frac {\left| t ^ {\prime} - t \right| ^ {2}}{2 \sigma_ {t} ^ {2}}\right)\right), \tag {4} +$$ + +where $\sigma_{t}$ represent the learnable scale parameters for the $t$ -th time step. The prior association is computed using a Gaussian kernel, with $G(|t' - t|; \sigma_t) = \frac{1}{\sqrt{2\pi} \sigma_t} \exp \left( -\frac{|t' - t|^2}{2 \sigma_t^2} \right)$ representing the association weight calculated based on the distance $|t' - t|$ between time points $t$ and $t'$ . The function Rescale(\cdot) normalizes the association weights into a discrete distribution $P$ . + +# 4.3. Spatial Correlation Modeling + +Forged images often exhibit abnormal spatial similarities (e.g., excessive symmetry) or inconsistencies (e.g., color or texture mismatches). To address this, we extract features from each dimension of the latent space and apply a self-attention mechanism to compute spatial correlations within the latent space. This helps amplify regions with abnormal similarities or inconsistencies in forged images, thereby enhancing forgery trace detection. + +For the feature matrix $X$ , the vertical vector $u_{i} = X[(:,i]$ where $u_{i}\in \mathbb{R}^{T\times 1}$ and $i\in \{1,2,3,\dots ,d\}$ , represents the pixel- or region-wise temporal features. Consequently, the spatial token group $X^{\top} = \{u_{i}^{\top}\}_{i = 1}^{d}\in \mathbb{R}^{d\times T}$ can be utilized to compute the spatial correlation matrix $S$ as follows: + +$$ +Q, K, V = X ^ {\top} W _ {Q}, X ^ {\top} W _ {K}, X ^ {\top} W _ {V}, \tag {5} +$$ + +$$ +S = \left(S o f t m a x \left(\frac {Q K ^ {\top}}{\sqrt {T}}\right) V\right) ^ {\top} A, \tag {6} +$$ + +where $\top$ denotes the transpose operation, and $Q,K,V$ along with $W_{Q},W_{K},W_{V}$ represent the query, key, and value matrices and their corresponding parameters, respectively. The term $A$ signifies the frequency-enhanced attention weight, which directs the computation of the spatial correlation matrix to emphasize pixel- or region-wise temporal vectors exhibiting significant dynamic evolution patterns. Further elaboration and the corresponding equations for the frequency-enhanced attention matrix $A$ are discussed in the subsequent subsection. + +# 4.4. Frequency-Enhanced Attention Mechanism + +In frequency domain analysis, the high-frequency details of forged images tend to be eliminated during gradient optimization, resulting in significant differences in frequency characteristics compared to real images. To address this, we extract the frequency-domain information of latent space features using the Fast Fourier Transform (FFT) and compute frequency-domain correlations to capture cross-scale feature interactions. Since high-frequency information is lost in forged images, the frequency distribution of real and forged images exhibits distinct patterns after conjugate product computation. We then apply the Inverse Fourier Transform (IFFT) to map the correlation back to the spatiotemporal domain, allowing the model to focus on anomalies caused by missing frequency-domain information. + +For the group of tokens $X = \{\vec{z}_t\}_{t=1}^T$ , the self-attention with query matrix $\overline{Q}$ and key matrix $\overline{K}$ for $X$ is adopted to find the feature temporal similarity of different tokens $\{\vec{z}_t\}_{t=1}^T$ . Furthermore, we perform the Fourier Transform on the query matrix $\overline{Q}$ and key matrix $\overline{K}$ to extract frequency-domain features: + +$$ +\widehat {Q} = \mathcal {F} (\overline {{Q}}), \widehat {K} = \mathcal {F} (\overline {{K}}), \tag {7} +$$ + +where $\mathcal{F}()$ denotes the Fast Fourier Transform. Next, we compute the conjugate product of the query and key in the frequency domain to obtain cross-scale interaction information $C$ : + +$$ +C = \widehat {Q} \cdot \widehat {K} ^ {*}, \tag {8} +$$ + +where “.” denotes element-wise multiplication of matrices, $\widehat{K}^*$ represents the complex conjugate, ensuring correct interaction in the frequency domain and highlighting the different frequency distributions of real and forged images. Then, we apply the Inverse Fourier Transform $\mathcal{F}^{-1}$ to restore the result to the spatio-temporal domain and perform Softmax normalization to ensure that the weights maintain appropriate scales in the spatio-temporal domain, yielding the frequency-enhanced spatio-temporal attention weights: + +$$ +A = \operatorname {S o f t m a x} \left(\frac {\mathcal {F} ^ {- 1} (C)}{\sqrt {T}}\right). \tag {9} +$$ + +Finally, $A$ serves as the frequency-enhanced temporal attention weight to improve generalization ability. + +# 4.5. Learning Strategy + +First, we assess the distribution characteristics of the input time series $X$ across multiple levels by computing the association difference $(AssD)$ between the prior association $P$ and the frequency-domain association $A$ . Specifically, we employ the Kullback-Leibler divergence (KL divergence) to quantify the discrepancy between two discrete distributions, defined as follows: + +$$ +A s s D (P, A) = K L (P [ t,: ] \| A [ t,: ]) + K L (A [ t,: ] \| P [ t,: ]), +$$ + +where $KL(A[t,:]\| P[t,:])$ denotes the KL divergence, measuring the informational disparity between distributions $A[t,:]$ and $P[t,:]$ . Here, $P[t,:]$ and $A[t,:]$ represent the prior association and frequency-domain association distributions at time step $t$ , respectively. The final association difference $AssD(P,A)$ captures the overall variation in association patterns. + +For the output of the spatial correlation modeling module, we introduce a global optimization objective based on cross-entropy loss to steer the spatial correlation and frequency-domain enhancement modules in distinguishing between forged and authentic images. + +We formulate distinct optimization objectives tailored to different sample types. For forged images, we enforce a deviation between the prior association $P$ and the frequency-domain association $A$ , thereby accentuating the intricate and challenging-to-model characteristics of the real image distribution: + +$$ +\mathcal {L} _ {\text {f a k e}} = C E (y, \hat {y}) - \lambda \| A s s D (P, A _ {\text {f i x}}) \|, \tag {10} +$$ + +where the first term represents the cross-entropy loss between the classification output $\hat{y}$ of the spatial correlation modeling module and the ground truth labels $y$ , enhancing the module's classification performance. The second term regulates the relationship between the prior association $P$ and the frequency-domain association $A$ using the association difference $AssD()$ for forged images, thereby improving the model's temporal modeling capability. Here, $A_{fix}$ represents the fixed frequency-domain association. The hyperparameter $\lambda$ modulates the contribution of the $AssD()$ loss. + +For real images, we allow the prior association $P$ to closely align with the frequency-domain association $A$ , emphasizing the smoothness of forged images in the frequency domain as follows: + +$$ +\mathcal {L} _ {\text {r e a l}} = C E (y, \hat {y}) + \lambda \| A s s D (P, A _ {\text {f i x}}) \|, \tag {11} +$$ + +where the second term ensures that the prior association $P$ remains distant from the frequency-domain association $A$ + +![](images/86007e8979d48ebffdd952337aaf332324256919b0506cb1ae84c7b0dfb39632.jpg) +(a) Performance comparisons on Genimage (ACC, %). + +![](images/fabdbd33fd9f418b3779111ff79717d3103e790e37a24df61b065c6ab6bc74eb.jpg) +(b) Performance comparisons on Deepfacegen (AUC, %). +Figure 2. The radar chart shows the forgery identification performance of STFT in comparison to 12 alternative methods on Genimage (a) and 10 other methods on Deepfacegen (b). Solid lines depict the performance of the comparative methods, while a dashed line signifies the performance of STFT. + +for real images, thereby enhancing the model's discriminative capability. + +For a batch comprising both forged and real images, in addition to the aforementioned losses $\mathcal{L}_{\text{fake}}$ and $\mathcal{L}_{\text{real}}$ , we ensure that $A$ remains distinct from $P$ , thereby directing greater attention to non-adjacent regions and enhancing the flexibility of global associations: + +$$ +\mathcal {L} _ {m i x} = C E (y, \hat {y}) - \lambda \| A s s D (P _ {f i x}, A) \|, \tag {12} +$$ + +where $P_{fix}$ denotes the fixed prior association matrix. This alternating optimization strategy stabilizes the training process and utilizes the frequency-domain association $A$ to guide the model in focusing on more nuanced frequency-domain details, effectively distinguishing between genuine and forged images. + +# 5. Experiments + +In the first part of this section, we introduce the detailed experimental setup. Next, in the second part, we evaluate and compare STFT with other forgery detection methods, including SOTA approaches. In the fourth part, we assess the generalization performance of STFT. Finally, we conduct an ablation study on the proposed model. + +# 5.1. Experimental Setup + +Evaluation Metrics and Datasets. In alignment with the established comparison standards, we evaluate accuracy (ACC) and area under the curve (AUC) as performance metrics for the GenImage [53] and DeepFaceGen [2] datasets, respectively. The GenImage [53] dataset consists of a total of 2,681,167 images, including 1,331,167 real images and 1,350,000 generated images. These generated images are sourced from eight different generative models: BigGAN + +[3], GLIDE [30], VQDM [16], Stable Diffusion V1 and V1.5 [38], ADM [12], Midjourney [21], and Wukong [15]. All fake images are generated using the template prompt "photo of CLS," where "CLS" is replaced by one of the 1,000 labels from ImageNet [11]. The dataset is divided into eight subsets, each containing a portion of real images along with all fake images generated by the aforementioned models. In this study, we follow the official dataset split, using 2,581,167 images for training and the remaining 100,000 images for validation. + +DeepFaceGen [2] is a facial forgery dataset that includes both video and image modalities. We select face forgery data generated using novel prompt-guided forgery methods. These methods cover 12 diffusion- and autoregressive-based models across text-to-image and image-to-image modalities. The dataset includes face forgeries across various ages, genders, and environments, with generated images sourced from popular models such as Midjourney [21], DALL-E [13], and SDXL [38]. + +Implementation Details. We downloaded the pretrained LDM model, fed images into it, and extracted the latent space data. We set the default diffusion steps to 20, obtaining 20 latent representations of size $3 \times 64 \times 64$ . Each latent representation was then processed using average pooling to reduce its size to $3 \times 8 \times 8$ and flattened into a one-dimensional vector, ultimately forming sequential data of size $20 \times 192$ . We set the batch size to 256, the number of channels in the hidden state to 512, and the number of attention heads to 8. The hyperparameter $\lambda$ was set to 2 for all datasets. We used the ADAM optimizer [23] with an initial learning rate of $1 \times 10^{-4}$ . All experiments were conducted on an Nvidia A40 GPU. + +
MethodYearMidjourneySDv1.4SDv1.5ADMGLIDEWukongVQDMBigGAN
DIRE [45]202399.5 / 72.699.7 / 68.899.7 / 68.899.2 / 67.999.8 / 52.299.7 / 68.199.9 / 51.9100 / 50.1
Defake [39]202398.8 / 78.799.8 / 81.5100 / 82.398.7 / 82.599.8 / 73.199.3 / 75.699.1 / 82.098.7 / 76.6
PatchCraft [52]202390.1 / 81.189.5 / 81.390.1 / 81.390.4 / 79.390.6 / 80.492.6 / 79.390.3 / 74.980.9 / 73.7
AIDE [49]202498.9 / 86.099.7 / 84.999.6 / 84.095.3 / 79.197.6 / 81.898.8 / 80.492.3 / 83.096.3 / 81.7
DRCT [6]202499.9 / 89.395.5 / 86.696.8 / 89.195.1 / 81.896.6 / 83.899.6 / 84.494.3 / 82.498.8 / 84.0
STFT (Ours)-99.4 / 92.399.6 / 89.199.3 / 90.499.5 / 82.599.8 / 87.0100 / 84.899.7 / 85.7100 / 85.6
+ +# 5.2. Comparison with State-of-the-arts + +GenImage Benchmark. To verify the model's generalization ability across different generative models, this section follows the GenImage evaluation standard and compares our approach with 12 mainstream forgery detection models, including CNNSpot [34], F3Net [33], CLIP/RN50 [34], GramNet [24], De-fake [39], Conv-B [26], SwinT [25], UnivFD [31], DIRE [45], PathCraft [52], AIDE [49], and DRCT [6]. According to the GenImage evaluation protocol, all models were trained on the SDv1.4 subset of GenImage and evaluated on multiple test sets. As shown in Figure 2a, nearly all methods achieve high accuracy on the Wukong, SDv1.4, and SDv1.5 datasets but perform poorly on GLIDE, VQDM, Midjourney, ADM, and BigGAN datasets. Since BigGAN is a non-diffusion model dataset, all methods perform the worst on it. In contrast, our proposed STFT achieves over $89\%$ accuracy on the BigGAN dataset and obtains the highest ACC on 7 out of 8 sub-datasets, demonstrating a significant advantage over existing forgery detection methods (see Appendix Table 4 for complete results). + +DeepFaceGen Benchmark. Following the DeepFaceGen evaluation standard, we compare STFT with mainstream generative image detection methods, including Xception [7], EfficientNet-B0 [44], F3-Net [33], RECCE [4], DNADet [50], DIRE [45], DRCT [6], UnivFD [31], NPR [43], and FreqNet [42]. Training and testing were conducted on the same dataset, which includes real images and images generated by 12 different forgery methods such as DALL-E and SD. The dataset was split into training, validation, and test sets with a 7:1:2 ratio. As shown in Figure 2b, STFT outperforms all methods across all 12 subsets, with a particularly notable advantage in distinguishing difficult-to-detect forged images generated by DALL-E 1, DALL-E 3, Midjourney, and Wenxin. Overall, STFT achieves an AUC above $90\%$ for all forgery categories, with an average AUC of $95.21\%$ (detailed results are provided in Appendix Table 5). + +Table 1. Performance comparison (ACC, %) of various methods. Each value is presented as (Score1 / Score2), where Score1 corresponds to the accuracy achieved when trained and tested on the same subset, and Score2 represents the average accuracy when trained on the current subset and evaluated across other subsets. Bold values highlight the best performance. + +
Train OnCatHorseBedroom
Test OnHorse/BedroomCat/BedroomCat/Horse
DIRE [45]73.0165.4555.04
DRCT [6]74.4868.7867.62
STFT (Ours)90.3289.7091.51
+ +Table 2. Generalization performance (AUC,%) of different methods on non-face images. + +# 5.3. Generalization Performance + +To evaluate the generalization capability of STFT, we designed two experimental scenarios: first, testing the model when the pre-trained model does not align with the target detection task, and second, assessing its adaptability to unknown forgery methods. + +Adaptation to Unknown Forgery Methods. We trained STFT on a single subset of the GenImage dataset and tested it across all eight subsets. As shown in Table 1, we compared STFT with the top four models from the baseline evaluations. STFT outperformed the existing SOTA method in GenImage, DRCT, and achieved the best generalization performance across all eight subsets. When confronted with unseen forgery methods, all models experienced performance degradation in various detection metrics. However, STFT still demonstrated strong generalization ability. The superior generalization of STFT is primarily attributed to its frequency-domain-enhanced spatio-temporal correlation modeling, which captures multi-perspective features of forged images. This enhances the model's adaptability to different types of forgeries, allowing it to effectively recognize anomalous patterns even when facing previously unseen forgery methods. + +Generalization Performance on Non-face Images. Reconstruction-error-based forgery detection methods often fail when the test data differs from the reconstruction model's pre-training domain. For instance, applying a model trained on animals to plant images may misclassify real images as fake due to high reconstruction errors. To assess STFT's robustness in such cases, we used pre-trained + +
Model VariantTemporalSpatialFrequencyAUC(%)
Full Model✓✓✓92.02
w/o Temporal✓✓86.28
w/o Spatial✓✓79.71
w/o Frequency✓✓89.75
+ +Table 3. Ablation study on different modules. + +models from [12] on LSUN Bedroom, Cat, and Horse categories, and compared STFT with DIRE and DRCT. Following the Benchmark 1 protocol on 12 prompt-based forgery datasets, results in Table 2 show that while reconstruction-based methods suffered major performance drops, STFT consistently achieved AUC scores over $90\%$ , highlighting its strong generalization ability. + +# 5.4. Ablation Study + +In this section, we conduct ablation experiments to quantify the contribution of each design component in our model and analyze the impact of key hyperparameters. Specifically, we examine the effect of the latent space sampling steps and feature map resolution and the inclusion of the frequency-enhanced attention module. + +Effect of Latent Space Sampling Steps and Feature Map Resolution. To investigate the impact of sampling steps and feature resolution in latent space, we evaluate detection performance with sampling steps set to 10, 15, 20, 25, and 30, and test four resolutions by downsampling latent features via average pooling. As shown in Figure 3, the model performs best with 20 sampling steps and a feature resolution of $3 \times 8 \times 8$ . + +This indicates that 20 steps sufficiently preserve temporal cues while preventing over-smoothing. A $3 \times 8 \times 8$ resolution balances detail retention and generalization. Lower resolutions (e.g., $3 \times 4 \times 4$ ) lose crucial spatial information, reducing accuracy, while higher resolutions introduce redundancy, affecting robustness. Thus, 20 sampling steps and $3 \times 8 \times 8$ resolution are adopted as default settings for optimal accuracy and efficiency. + +Impact Analysis of Model Components. To analyze the contributions of different STFT components, we conduct ablation experiments by sequentially removing the temporal prior correlation module, spatial correlation module, and frequency-enhanced attention module on DALL-E 3 Dataset. The results, presented in Table 3, confirm their complementary roles in forgery detection. + +Removing the temporal module significantly weakens the model's ability to capture sequential forgery patterns, leading to a notable performance drop. This aligns with our observation that latent diffusion smooths temporal transitions in forged images. Without spatial correlation modeling, the model struggles to detect local modifications, increasing false positives and reducing robustness. Frequency + +![](images/b833215d04f584db9a9227dad50305edda3acc39b74bcedc312382a20fe5f2cd.jpg) +Figure 3. The ablation study on diffusion steps (a) and latent feature map size (b). Solid lines in both cases represent the performance results corresponding to the final adopted values. + +![](images/9c04413263ed1255f384ca91c423de398ac5b97699cc35f3db8c515ee3dd0438.jpg) + +features are crucial for detecting high-frequency artifacts in forgeries. Removing this module degrades performance, especially against unknown forgery methods, highlighting its role in generalization. + +# 6. Conclusion + +This paper proposes the Spatial-Temporal Forgery Trace (STFT) method for forgery image identification. Leveraging the tendency of forged images to fit specific distributions more easily than real ones, STFT employs generative diffusion models to capture temporal distributions and models relationships between temporal and spatial features as forgery traces. It also integrates frequency-domain features to accelerate trace localization. Experiments show that STFT effectively identifies subtle spatio-temporal traces by combining spatial, temporal, and frequency perspectives, demonstrating strong robustness and generalizability. + +Limitations and Future Works. Although the proposed STFT achieves high accuracy and demonstrates commendable generalization capabilities, it is not without its limitations. Firstly, the process of mapping the original image into the latent diffusion distribution is computationally intensive and time-consuming. Additionally, the method appears to exhibit suboptimal performance on GAN-based forgery images, likely due to fundamental differences in the generative mechanisms of GANs and generative diffusion models. In future work, we aim to enhance the efficiency of the proposed method and further extend its generalization ability. + +# 7. Acknowledgments + +This work is funded by National Key Research and Development Project (Grant No: 2022YFB2703100), Ant Group through CCF-Ant Research Fund, and Information Technology Center, ZheJiang University. + +# References + +[1] Baidu AI. Wenxin large model for text-to-image generation. Baidu AI Blog, 2023. 2 +[2] Yijun Bei, Hengrui Lou, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Song, Mingli Song, and Zunlei Feng. A largescale universal evaluation benchmark for face forgery detection. arXiv preprint arXiv:2406.09181, 2024. 6 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2, 6 +[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4113-4122, 2022. 7 +[5] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1 +[6] Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In Forty-first International Conference on Machine Learning, 2024. 7, 2 +[7] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 7 +[8] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 973-982, 2023. 3 +[9] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3 +[10] Edmar RS De Rezende, Guilherme CS Ruppert, Antonio Theophilo, Eric K Tokuda, and Tiago Carvalho. Exposing computer generated images by using deep convolutional neural networks. Signal Processing: Image Communication, 66: 113-126, 2018. 1 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 8 +[13] Tira Nur Fitria. Artificial intelligence (ai) technology in openai chatgpt application: A review of chatgpt in writing english essay. In ELT Forum: Journal of English Language Teaching, pages 44-58, 2023. 2, 6 + +[14] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 1 +[15] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Niu Minzhe, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, et al. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems, 35:26418-26431, 2022. 6, 2 +[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 6 +[17] Weinan Guan, Wei Wang, Jing Dong, Bo Peng, and Tieniu Tan. Collaborative feature learning for fine-grained facial forgery detection and segmentation. arXiv preprint arXiv:2304.08078, 2023. 1 +[18] Tahir Hassan, Aras Asaad, Dashti Ali, and Sabah Jassim. Artificial image tampering distorts spatial distribution of texture landmarks and quality characteristics. arXiv preprint arXiv:2208.02710, 2022. 1 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[21] Anna Jaruga-Rozdolska. Artificial intelligence as part of future practices in the architect's work: Midjourney generative tool as part of a process of creating an architectural form. Architectus, (3 (71):95-104, 2022. 6, 2 +[22] Wei et al. Jiang. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16423-16432, 2021. 2 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[24] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 7, 2 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7, 2 +[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7, 2 + +[27] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2 +[28] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 3 +[29] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. arXiv preprint arXiv:2307.06272, 2023. 2 +[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 2 +[31] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 7, 2 +[32] OpenAI. Dall-e 3: Openai's latest text-to-image model. OpenAI Blog, 2023. 2 +[33] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7, 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 7, 2 +[35] Aditya et al. Ramesh. Zero-shot text-to-image generation. arXiv preprint arXiv:2102.12092, 2021. 2 +[36] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. arXiv preprint arXiv:2210.14571, 2022. 3 +[37] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 3 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6 +[39] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text-to-image generation models. In Proceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3418-3432, 2023. 2, 3, 7 +[40] Stability.ai. Stable Diffusion, 2023. 2 +[41] Eren Tahir and Mert Bal. Deep image composition meets image forgery. arXiv preprint arXiv:2404.02897, 2024. 1 + +[42] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5052-5060, 2024. 3, 7 +[43] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7 +[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7 +[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2, 3, 7 +[46] Jing Weng. Local frequency analysis for diffusion-generated image detection. In International Conference on Image Processing and Artificial Intelligence (ICIPAL 2024), pages 66-73. SPIE, 2024. 3 +[47] Moritz Wolter, Felix Blanke, Raoul Heese, and Jochen Garcke. Wavelet-packets for deepfake image analysis and detection. Machine Learning, 111(11):4295-4327, 2022. 1 +[48] Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. arXiv preprint arXiv:2110.02642, 2021. 4 +[49] Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection. arXiv preprint arXiv:2406.19435, 2024. 7, 2 +[50] Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4662-4670, 2022. 7 +[51] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023. 1 +[52] Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection. arXiv preprint arXiv:2311.12397, 2023. 7, 2 +[53] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36:77771-77782, 2023. 6 \ No newline at end of file diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/images.zip b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..150576757a9d93c217f61a63351ee80d9a944bd1 --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44ab73c2de76d1357e44f9a77ab04d46df99e042f09fa00d75a29069c4acfaa4 +size 346886 diff --git a/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/layout.json b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3807374ac979d219d50ee6a84fcac5db0d785dde --- /dev/null +++ b/2025/Spatial-Temporal Forgery Trace based Forgery Image Identification/layout.json @@ -0,0 +1,9045 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "text", + "content": "Spatial-Temporal Forgery Trace based Forgery Image Identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": "Yilin Wang" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Zunlei Feng" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Jiachi Wang" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Hengrui Lou" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Binjia Zhou" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Jie Lei" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Mingli Song" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "text", + "content": ", Yijun Bei" + }, + { + "bbox": [ + 111, + 141, + 499, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 173, + 490, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 173, + 490, + 187 + ], + "spans": [ + { + "bbox": [ + 121, + 173, + 490, + 187 + ], + "type": "text", + "content": "1State Key Laboratory of Blockchain and Data Security, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 188, + 436, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 188, + 436, + 201 + ], + "spans": [ + { + "bbox": [ + 175, + 188, + 436, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 188, + 436, + 201 + ], + "type": "text", + "content": "School of Software Technology, Zhejiang University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 173, + 201, + 438, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 201, + 438, + 216 + ], + "spans": [ + { + "bbox": [ + 173, + 201, + 438, + 216 + ], + "type": "text", + "content": "3Zhejiang Provincial Key Laboratory of Service Robot" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 216, + 465, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 465, + 230 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 465, + 230 + ], + "type": "text", + "content": "4College of Computer Science, Zhejiang University of Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 256, + 200, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 256, + 200, + 269 + ], + "spans": [ + { + "bbox": [ + 151, + 256, + 200, + 269 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 281, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 297, + 628 + ], + "type": "text", + "content": "The rapid development of AIGC technology has enabled highly realistic forged images to deceive human perception, posing serious risks across many areas. Current deepfake image detection methods primarily identify forgeries by extracting handcrafted features, deep features, and frequency-domain features. While these features contain forgery traces, they also include a substantial amount of the image's semantic information, which interferes with the precision and generalization of forgery detection models. To tackle these challenges, this paper introduces a novel forgery image identification method based on the Spatial-Temporal Forgery Trace (STFT). Motivated by the fact that forgery images are more easily fitted to a specific distribution than real images, the STFT method approaches the issue from a forged image distribution modeling perspective, employing generative diffusion models to meticulously capture the temporal distribution of images. It further models the relationship between temporal feature variations and spatially corresponding temporal features, treating them as temporal and spatial forgery traces. Moreover, STFT incorporates frequency-domain features as weighting factors to accelerate the localization of spatio-temporal forgery traces. Experiments demonstrate that by integrating spatial, temporal, and frequency perspectives within the latent space, STFT effectively captures subtle spatio-temporal forgery traces, exhibiting strong robustness and generalizability. It outperforms state-of-the-art methods on major benchmark datasets in the field. The source code for STFT is available at https://github.com/GCLion/STFT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 651, + 135, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 135, + 663 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 135, + 663 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "With the rapid development of Artificial Intelligence Generated Content (AIGC) technology [5], particularly the ma" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 389 + ], + "type": "text", + "content": "turity of generative models like diffusion models, AIGC has been widely applied in fields such as image and video generation, bringing significant innovation and convenience. However, the widespread application of AIGC has also brought about a series of serious risks. AIGC technology has made it extremely easy to generate realistic images. This has led to the proliferation of false information, malicious tampering, and other harmful activities. Additionally, by generating forged images, malicious attackers may impersonate others to carry out fraud, causing economic losses and a crisis of social trust [51]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 700 + ], + "type": "text", + "content": "Current deepfake image detection methods primarily rely on extracting handcrafted features, deep features, and frequency domain features for forgery identification. Handcrafted feature-based approaches [10, 18] typically extract statistical features such as color histograms, local binary patterns (LBP), and histograms of oriented gradients (HOG) to detect fake images. While these methods may be effective for early deepfake images, they exhibit poor adaptability to emerging deepfake techniques, such as images generated by diffusion models, making them less effective against continuously evolving attack strategies. Methods based on deep features [17, 41] typically utilize deep neural networks such as CNN, ResNet, and EfficientNet to extract high-level features from images for classification and discrimination. While these approaches can improve detection performance to some extent, they often learn the semantic features of the image itself rather than focusing specifically on forgery traces, which affects the generalization ability of the forgery detection model. Frequency domain-based methods [14, 47] extract features from high-frequency or specific frequency bands using frequency domain analysis techniques such as Fourier transform and wavelet transform to detect forgery traces. However, since different AIGC generation methods may use different generation strategies, the stability of frequency domain features is insufficient, which limits the generalization ability of these methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "The features extracted by the most of aforementioned" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 254, + 713 + ], + "type": "text", + "content": "*Corresponding author. † Equal contribution to this work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17067" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "methods not only contain forgery trace characteristics but also include a large amount of semantic information from the image itself, which interferes with the accuracy and generalization ability of forgery detection models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 242 + ], + "type": "text", + "content": "Regarding the generative mechanisms of contemporary deep image synthesis models, these models create images by assimilating particular data distributions. However, their gradient-based optimization processes inherently display a propensity for \"shortcut learning\"—favoring the acquisition of commonly encountered feature distributions while neglecting those that are more infrequent. This phenomenon is driven by the observation that forged image features are easily assimilated, whereas the subtle, rare features inherent in genuine images often evade capture." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 506 + ], + "type": "text", + "content": "This paper adopts a perspective of forgery image distribution modeling, using generative diffusion models to obtain image temporal distribution features. Then, we mine forgery trace features from the temporal distribution feature space that do not contain semantic information interference, aiming to achieve high accuracy and high generalization in image forgery detection. By leveraging the mapped latent temporal distribution of images, this paper elucidates the evolving patterns of temporal distribution features, identifying them as temporal forgery indicators. In the spatial dimension, based on the hypothesis that forgery images tend to model the approximate semantic parts with either exactly the same or completely different distributions, we extract the correlation between the temporal features of corresponding image regions as spatial forgery traces. Furthermore, generated images typically expose forgery traces more easily in high-frequency regions. We obtain high-frequency information as weights to help the model more quickly identify spatio-temporal forgery traces. This approach provides an innovative perspective for research in the detection of forged images using deep learning models. Our main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 508, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 508, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 602 + ], + "type": "text", + "content": "- We introduce an innovative spatial-temporal forgery trace extraction framework designed for the identification of forged images through latent diffusion distribution. Rather than relying on mixed features, the temporal evolution patterns of individual temporal features and the interrelationships among multiple temporal features are modeled as temporal and spatial forgery traces, enabling robust and precise forgery identification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 604, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 651 + ], + "type": "text", + "content": "- Instead of utilizing frequency features directly, high-frequency components are employed as weighting factors to enhance the forgery detection model's ability to swiftly localize spatio-temporal forgery traces." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 711 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate that the proposed method surpasses the limitations of prior approaches in feature extraction, and exhibits exceptional robustness and generalization capabilities against various forgery techniques." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 91, + 462, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 462, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 462, + 104 + ], + "type": "text", + "content": "2.1. Image Forgery Technology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 110, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 373 + ], + "type": "text", + "content": "The field of image generation was initially led by technologies based on Generative Adversarial Networks (GANs) and autoregressive models. Through the continuous evolution of these technologies, milestone products such as BigGAN [3] and DALL-E [13] emerged. However, these methods face certain limitations in content generation control, with stability and complexity issues being particularly prominent. This has led to the emergence of a new type of image generation method—the diffusion model. The Denoising Diffusion Probabilistic Model (DDPM) [19], inspired by nonequilibrium thermodynamics, was born and demonstrated excellent generative performance, further driving in-depth research on diffusion models. Subsequently, researchers have focused on optimizing its structure and improving sampling efficiency. The Latent Diffusion Model (LDM) [38] applies the diffusion process in the latent space, not only improving model efficiency but also enhancing the ability to generate text-conditioned images through a cross-attention mechanism. LDM has become a significant advancement in the field of image generation, leading to the development of well-known generative models such as Stable Diffusion [38] and LCM-RoLA [27]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 382, + 454, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 454, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 454, + 396 + ], + "type": "text", + "content": "2.2. Image Forgery Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 484 + ], + "type": "text", + "content": "For traditional Photoshop-based forgery methods, early research primarily relied on extracting handcrafted features, deep features, and frequency-domain features for forgery identification. For deepfake images, different approaches have been explored based on distinct feature perspectives. Below, we introduce related works from three aspects: temporal, spatial, and frequency domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 664 + ], + "type": "text", + "content": "In forgery detection methods based on spatial features, Sha et al. [39] analyze unique characteristics of generated images and perform model attribution, aiding in source identification and combating the misuse of forged images. The method in [45] detects forged images by comparing the input image with its reconstruction from a diffusion model, demonstrating strong performance in detecting forgeries under unknown diffusion models and various perturbations. Overall, spatial-based forgery detection methods mainly extract image features using deep models. However, these extracted features often mix semantic information with forgery traces, where the presence of semantic features increases the learning difficulty and reduces detection accuracy, ultimately leading to poor generalization performance of the detection models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "In forgery detection methods based on temporal features, [29] effectively identifies images generated by diffusion models by leveraging the deterministic inversion and denoising computation errors of diffusion models, combined" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17068" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "with statistical and neural network techniques. Similarly, [9] analyzes forensic features of GAN-generated images, which can also be used to evaluate the performance of existing GAN detectors on diffusion model images, further improving detection accuracy. Although forgery detection methods based on temporal features utilize diffusion models to extract temporal characteristics, they still rely on supervised signals to extract mixed features from the aggregated temporal features. Like spatial feature-based forgery detection, these methods suffer from interference with model accuracy and poor generalization ability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 359 + ], + "type": "text", + "content": "In forgery detection based on image frequency domain features, [46] extracts frequency domain features using the Fourier transform and finds that forged images often exhibit different frequency distributions compared to real images. A convolutional neural network is then used to classify high-frequency features. [42] enhances detection performance by learning forgery-specific features in the frequency domain space and analyzing the abnormal patterns of forged images across different frequency bands. These methods determine authenticity based on the quantity and distribution of high-frequency features. However, due to the diversity of image types, high-frequency features also vary, leading to limited cross-domain generalization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 491 + ], + "type": "text", + "content": "The methods mentioned above are mostly designed for GAN-generated images. Although Corvi et al. [8, 9, 36] also found that spectral artifacts may exist in diffusion-generated images, the effectiveness of spectral artifact-based forgery detection methods in practice remains limited. To address this issue, researchers have begun to analyzing the characteristics of the diffusion generation process itself. For example, DIRE [45] and De-fake [39] use the reconstruction error of the diffusion model to detect forged images, while LaRE2 [28] and AEROBLADE [37] focus on the reconstruction error in the latent space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 622 + ], + "type": "text", + "content": "However, these methods are highly dependent on pretrained reconstruction models and are susceptible to the limitations of the training data characteristics. For example, if the reconstruction model is trained on a specific category (such as cats) but is used to detect forged images from different categories (such as dogs or plants), it may result in misclassification. The core issue here lies in using reconstruction errors as the basis for detection, without fully exploring the characteristics of the diffusion generation mechanism itself, thus affecting the understanding of the forgery image generation patterns." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 633, + 140, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 140, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 140, + 645 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "In this section, we define the key notations of the latent diffusion model and outline its basic principles. In LDM [27], given an image " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " in the RGB space (" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " denote the height and width of the image " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "), the encoder " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " into a latent representation " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(x)" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "while the decoder " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " reconstructs the image from the latent representation, yielding " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{x} = \\mathcal{D}(z) = \\mathcal{D}(\\mathcal{E}(x))" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", where ideally " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " denote the height, width, and channel of the latent representation " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "). This diffusion model consists of a " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "-step diffusion process " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "1 \\leq t \\leq T" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and a denoising process, which is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 351, + 152, + 553, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 152, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 351, + 152, + 553, + 167 + ], + "type": "interline_equation", + "content": "q \\left(z _ {t} \\mid z _ {t - 1}\\right) = \\mathcal {N} \\left(z _ {t}; \\sqrt {1 - \\beta_ {t}} z _ {t - 1}, \\beta_ {t} I\\right), \\tag {1}", + "image_path": "3f60f0c40daaa686b05be8f9ff7fc7741d3c1b5b2d3813d625c07d38a4f9951b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 173, + 553, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 173, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 343, + 173, + 553, + 186 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(z _ {t - 1} \\mid z _ {t}\\right) = \\mathcal {N} \\left(z _ {t - 1}; \\mu_ {\\theta} \\left(z _ {t}, t\\right), \\Sigma_ {\\theta} \\left(z _ {t}, t\\right)\\right), \\tag {2}", + "image_path": "ad30a1c984304946dbea73d8e0ed9edc5be22a71f574d73041641f0f19365332.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": " is the diffusion result at time step " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": " is the noise factor at time step " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": " is the identity matrix, and " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": " represent the mean and covariance matrix of the denoising distribution, respectively. After training the model, step-by-step denoising can be performed starting from the standard normal distribution " + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "inline_equation", + "content": "z_{T} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 190, + 553, + 262 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 269, + 553, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 269, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 348, + 269, + 553, + 294 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {1}{\\sqrt {1 - \\beta_ {t}}} \\left(z _ {t} - \\beta_ {t} \\epsilon_ {\\theta} \\left(z _ {t}, t\\right)\\right) + \\sigma_ {t} \\xi , \\tag {3}", + "image_path": "afa75089db13179f624a26eca049d747227088fdee3149f105642c7adfe7e248.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\xi \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " is the random noise used to control sampling quality, and " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " regulates the randomness of the denoising process. Through repeated iterations, the final representation " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " is obtained, which is then decoded by " + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 300, + 555, + 431 + ], + "type": "text", + "content": " to reconstruct the synthesized image. During the forward diffusion process, high-frequency components are gradually suppressed by noise. By progressively adding noise in the forward process and denoising in the reverse process, high-frequency components become increasingly compressed. Notably, at high noise levels, the loss of high-frequency details becomes inevitable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 442, + 370, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 370, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 370, + 453 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 666 + ], + "type": "text", + "content": "As depicted in Figure 1, this method employs a diffusion process to map the image into the latent distribution space, subsequently extracting forgery traces from this latent distribution through a comprehensive spatial-temporal analysis. The overall framework consists of the following three main modules: Temporal Prior Correlation Modeling analyzes the latent space representations at different time steps of the diffusion process to capture their temporal variation characteristics. Spatial Correlation Modeling extracts features from different latent dimensions and computes spatial correlations using a self-attention mechanism to model spatial dependencies. Frequency-Enhanced Attention Mechanism utilizes frequency domain information to guide temporal prior correlation computation and spatial correlation analysis, thereby expediting the localization of forgery traces and enhancing generalization capabilities by disregarding the interference of inconsequential features." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 672, + 511, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 511, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 511, + 685 + ], + "type": "text", + "content": "4.1. Image Latent Distribution Modeling" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To overcome the limitations of existing methods and gain a deeper understanding of the underlying differences between" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17069" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 266 + ], + "type": "image", + "image_path": "6bc072cf4fd90f648d1333f7c61848b229eaf54039847ad2f0313cf0c92bf0b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 308 + ], + "type": "text", + "content": "Figure 1. The framework of STFT. It begins by mapping the input image into a latent diffusion distribution space. Subsequently, temporal and spatial forgery traces are extracted by capturing patterns of temporal variation and spatial correlations within the latent diffusion distribution. High-frequency components are utilized as weight to facilitate the rapid localization of forgery traces." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": "forged and normal images, we propose a forgery detection approach from a distributional perspective. Considering that LDM maps data into a low-dimensional latent space " + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": " through an encoder " + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": ", making the distribution " + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": " more compact and easier to model, while the high-dimensional pixel space " + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": " remains complex and difficult to model directly, we extract the latent representations " + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 54, + 327, + 294, + 434 + ], + "type": "text", + "content": " during the diffusion process of the latent diffusion model for further analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": "To facilitate the modeling of distributional changes from a temporal perspective, we reshape the latent representation " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": " at each diffusion step into a one-dimensional vector " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\vec{z}_{t} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": " as sequential data at time step " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "d = h \\times w \\times c" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "t \\in \\{1,2,\\dots,T\\}" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": ". The sequential vectors from step 1 to " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": " are concatenated vertically into a group of tokens " + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "inline_equation", + "content": "X = \\{\\vec{z}_{t}\\}_{t=1}^{T} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 55, + 435, + 295, + 533 + ], + "type": "text", + "content": ", which serves as input for both temporal and frequency-domain modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 255, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 255, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 255, + 553 + ], + "type": "text", + "content": "4.2. Temporal Prior Association Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 714 + ], + "type": "text", + "content": "Based on the assumption that the latent space distribution of forged images tends to be smoother with weaker temporal variations, we draw inspiration from the unsupervised temporal anomaly representation learning [48] and employ a trainable Gaussian kernel to compute prior associations between time steps. We introduce an adjustable scale parameter " + }, + { + "bbox": [ + 54, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 557, + 295, + 714 + ], + "type": "text", + "content": " to adapt to different temporal sequence patterns. The unimodal property of the Gaussian kernel enhances the association between local time steps, making the non-stationary characteristics of real images more prominent, while the smooth evolution pattern of forged images becomes more distinguishable. This improves the temporal modeling capability for forgery trace detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "text", + "content": "For the latent space representations " + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\{\\vec{z}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "text", + "content": " at different time steps, we treat the diffusion process as a sequential series, concatenate them into " + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "text", + "content": ", and compute prior associations between time steps. Specifically, we define the prior association matrix " + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 327, + 555, + 389 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 394, + 555, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 394, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 353, + 394, + 555, + 422 + ], + "type": "interline_equation", + "content": "P = \\operatorname {R e s c a l e} \\left(\\frac {1}{\\sqrt {2 \\pi} \\sigma_ {t}} \\exp \\left(- \\frac {\\left| t ^ {\\prime} - t \\right| ^ {2}}{2 \\sigma_ {t} ^ {2}}\\right)\\right), \\tag {4}", + "image_path": "64e96347b3b84dc1bb3a57fa252cb2fd07251fc8eb2ecd21ab2f020dfd5ad50f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": " represent the learnable scale parameters for the " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": "-th time step. The prior association is computed using a Gaussian kernel, with " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "G(|t' - t|; \\sigma_t) = \\frac{1}{\\sqrt{2\\pi} \\sigma_t} \\exp \\left( -\\frac{|t' - t|^2}{2 \\sigma_t^2} \\right)" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": " representing the association weight calculated based on the distance " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "|t' - t|" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": " between time points " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": ". The function Rescale(\\cdot) normalizes the association weights into a discrete distribution " + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 426, + 556, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 474, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 474, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 474, + 534 + ], + "type": "text", + "content": "4.3. Spatial Correlation Modeling" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 633 + ], + "type": "text", + "content": "Forged images often exhibit abnormal spatial similarities (e.g., excessive symmetry) or inconsistencies (e.g., color or texture mismatches). To address this, we extract features from each dimension of the latent space and apply a self-attention mechanism to compute spatial correlations within the latent space. This helps amplify regions with abnormal similarities or inconsistencies in forged images, thereby enhancing forgery trace detection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": "For the feature matrix " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": ", the vertical vector " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "u_{i} = X[(:,i]" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "u_{i}\\in \\mathbb{R}^{T\\times 1}" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3,\\dots ,d\\}" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": ", represents the pixel- or region-wise temporal features. Consequently, the spatial token group " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "X^{\\top} = \\{u_{i}^{\\top}\\}_{i = 1}^{d}\\in \\mathbb{R}^{d\\times T}" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": " can be utilized to compute the spatial correlation matrix " + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 633, + 556, + 694 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 699, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 699, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 354, + 699, + 555, + 715 + ], + "type": "interline_equation", + "content": "Q, K, V = X ^ {\\top} W _ {Q}, X ^ {\\top} W _ {K}, X ^ {\\top} W _ {V}, \\tag {5}", + "image_path": "29b50ee3ed5b915c8a934c6778b7da93b4d2b2d929c7d1fdae09eec9d7fe6f0d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17070" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 100, + 70, + 296, + 99 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 70, + 296, + 99 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 296, + 99 + ], + "type": "interline_equation", + "content": "S = \\left(S o f t m a x \\left(\\frac {Q K ^ {\\top}}{\\sqrt {T}}\\right) V\\right) ^ {\\top} A, \\tag {6}", + "image_path": "bcf239dfc92052807a5cc8dabc3adae4b811d038c0018ab7775c7acd5de18bd9.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\top" + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": " denotes the transpose operation, and " + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "inline_equation", + "content": "Q,K,V" + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "inline_equation", + "content": "W_{Q},W_{K},W_{V}" + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": " represent the query, key, and value matrices and their corresponding parameters, respectively. The term " + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": " signifies the frequency-enhanced attention weight, which directs the computation of the spatial correlation matrix to emphasize pixel- or region-wise temporal vectors exhibiting significant dynamic evolution patterns. Further elaboration and the corresponding equations for the frequency-enhanced attention matrix " + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 102, + 296, + 222 + ], + "type": "text", + "content": " are discussed in the subsequent subsection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 284, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 284, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 284, + 240 + ], + "type": "text", + "content": "4.4. Frequency-Enhanced Attention Mechanism" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 412 + ], + "type": "text", + "content": "In frequency domain analysis, the high-frequency details of forged images tend to be eliminated during gradient optimization, resulting in significant differences in frequency characteristics compared to real images. To address this, we extract the frequency-domain information of latent space features using the Fast Fourier Transform (FFT) and compute frequency-domain correlations to capture cross-scale feature interactions. Since high-frequency information is lost in forged images, the frequency distribution of real and forged images exhibits distinct patterns after conjugate product computation. We then apply the Inverse Fourier Transform (IFFT) to map the correlation back to the spatiotemporal domain, allowing the model to focus on anomalies caused by missing frequency-domain information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": "For the group of tokens " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "X = \\{\\vec{z}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": ", the self-attention with query matrix " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\overline{Q}" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\overline{K}" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": " is adopted to find the feature temporal similarity of different tokens " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\{\\vec{z}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": ". Furthermore, we perform the Fourier Transform on the query matrix " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\overline{Q}" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "inline_equation", + "content": "\\overline{K}" + }, + { + "bbox": [ + 55, + 412, + 296, + 484 + ], + "type": "text", + "content": " to extract frequency-domain features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 490, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 490, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 123, + 490, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\widehat {Q} = \\mathcal {F} (\\overline {{Q}}), \\widehat {K} = \\mathcal {F} (\\overline {{K}}), \\tag {7}", + "image_path": "a2c3758cb0000395b95a81e453d66aade8ffb4823fa9731426a7ac2e53ad9ff1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{F}()" + }, + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "text", + "content": " denotes the Fast Fourier Transform. Next, we compute the conjugate product of the query and key in the frequency domain to obtain cross-scale interaction information " + }, + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 511, + 296, + 558 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 148, + 557, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 557, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 148, + 557, + 294, + 571 + ], + "type": "interline_equation", + "content": "C = \\widehat {Q} \\cdot \\widehat {K} ^ {*}, \\tag {8}", + "image_path": "e7c547c222980ef9a7eab2e6333c875ad6fad46d6228b5b3e17ee760d28bf1f8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "text", + "content": "where “.” denotes element-wise multiplication of matrices, " + }, + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\widehat{K}^*" + }, + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "text", + "content": " represents the complex conjugate, ensuring correct interaction in the frequency domain and highlighting the different frequency distributions of real and forged images. Then, we apply the Inverse Fourier Transform " + }, + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 55, + 575, + 296, + 684 + ], + "type": "text", + "content": " to restore the result to the spatio-temporal domain and perform Softmax normalization to ensure that the weights maintain appropriate scales in the spatio-temporal domain, yielding the frequency-enhanced spatio-temporal attention weights:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 689, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 689, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 117, + 689, + 296, + 717 + ], + "type": "interline_equation", + "content": "A = \\operatorname {S o f t m a x} \\left(\\frac {\\mathcal {F} ^ {- 1} (C)}{\\sqrt {T}}\\right). \\tag {9}", + "image_path": "0eea5c3a1c1f873b901d3348ad9ad3252abf8eb910ad8afb80e2a0d27c672ee8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " serves as the frequency-enhanced temporal attention weight to improve generalization ability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 103, + 422, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 422, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 422, + 117 + ], + "type": "text", + "content": "4.5. Learning Strategy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "content": "First, we assess the distribution characteristics of the input time series " + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "content": " across multiple levels by computing the association difference " + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "inline_equation", + "content": "(AssD)" + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "content": " between the prior association " + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "content": " and the frequency-domain association " + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 121, + 554, + 204 + ], + "type": "text", + "content": ". Specifically, we employ the Kullback-Leibler divergence (KL divergence) to quantify the discrepancy between two discrete distributions, defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 213, + 552, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 213, + 552, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 552, + 228 + ], + "type": "interline_equation", + "content": "A s s D (P, A) = K L (P [ t,: ] \\| A [ t,: ]) + K L (A [ t,: ] \\| P [ t,: ]),", + "image_path": "ebfc8287c44a5cd91abde5eec0565f951149f4936b1ff34de4069b11c9b72a7b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "KL(A[t,:]\\| P[t,:])" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": " denotes the KL divergence, measuring the informational disparity between distributions " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "A[t,:]" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "P[t,:]" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "P[t,:]" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "A[t,:]" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": " represent the prior association and frequency-domain association distributions at time step " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": ", respectively. The final association difference " + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "inline_equation", + "content": "AssD(P,A)" + }, + { + "bbox": [ + 313, + 234, + 554, + 319 + ], + "type": "text", + "content": " captures the overall variation in association patterns." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 378 + ], + "type": "text", + "content": "For the output of the spatial correlation modeling module, we introduce a global optimization objective based on cross-entropy loss to steer the spatial correlation and frequency-domain enhancement modules in distinguishing between forged and authentic images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "content": "We formulate distinct optimization objectives tailored to different sample types. For forged images, we enforce a deviation between the prior association " + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "content": " and the frequency-domain association " + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 378, + 554, + 449 + ], + "type": "text", + "content": ", thereby accentuating the intricate and challenging-to-model characteristics of the real image distribution:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 338, + 458, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 458, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 338, + 458, + 553, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f a k e}} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {10}", + "image_path": "1e956bbcb6c04a9ca3b8a640efcf08e9a834011b26c9c02486c8a22882184296.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": "where the first term represents the cross-entropy loss between the classification output " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " of the spatial correlation modeling module and the ground truth labels " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": ", enhancing the module's classification performance. The second term regulates the relationship between the prior association " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " and the frequency-domain association " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " using the association difference " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "AssD()" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " for forged images, thereby improving the model's temporal modeling capability. Here, " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "A_{fix}" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " represents the fixed frequency-domain association. The hyperparameter " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " modulates the contribution of the " + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "inline_equation", + "content": "AssD()" + }, + { + "bbox": [ + 313, + 479, + 554, + 611 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "text", + "content": "For real images, we allow the prior association " + }, + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "text", + "content": " to closely align with the frequency-domain association " + }, + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 611, + 554, + 658 + ], + "type": "text", + "content": ", emphasizing the smoothness of forged images in the frequency domain as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 348, + 668, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 668, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 348, + 668, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e a l}} = C E (y, \\hat {y}) + \\lambda \\| A s s D (P, A _ {\\text {f i x}}) \\|, \\tag {11}", + "image_path": "b55d813b7cce27e7e2e6cb6712bb407566c28fbe99d02d6fe8bac69ac7a6b512.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where the second term ensures that the prior association " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " remains distant from the frequency-domain association " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "A" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17071" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 72, + 276, + 220 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 276, + 220 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 276, + 220 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 276, + 220 + ], + "type": "image", + "image_path": "86007e8979d48ebffdd952337aaf332324256919b0506cb1ae84c7b0dfb39632.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 224, + 261, + 235 + ], + "lines": [ + { + "bbox": [ + 84, + 224, + 261, + 235 + ], + "spans": [ + { + "bbox": [ + 84, + 224, + 261, + 235 + ], + "type": "text", + "content": "(a) Performance comparisons on Genimage (ACC, %)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 335, + 70, + 541, + 221 + ], + "blocks": [ + { + "bbox": [ + 335, + 70, + 541, + 221 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 541, + 221 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 541, + 221 + ], + "type": "image", + "image_path": "fabdbd33fd9f418b3779111ff79717d3103e790e37a24df61b065c6ab6bc74eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 224, + 531, + 235 + ], + "lines": [ + { + "bbox": [ + 343, + 224, + 531, + 235 + ], + "spans": [ + { + "bbox": [ + 343, + 224, + 531, + 235 + ], + "type": "text", + "content": "(b) Performance comparisons on Deepfacegen (AUC, %)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. The radar chart shows the forgery identification performance of STFT in comparison to 12 alternative methods on Genimage (a) and 10 other methods on Deepfacegen (b). Solid lines depict the performance of the comparative methods, while a dashed line signifies the performance of STFT." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 298, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 295, + 323 + ], + "type": "text", + "content": "for real images, thereby enhancing the model's discriminative capability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "content": "For a batch comprising both forged and real images, in addition to the aforementioned losses " + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{fake}}" + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{real}}" + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "content": ", we ensure that " + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "content": " remains distinct from " + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 323, + 296, + 382 + ], + "type": "text", + "content": ", thereby directing greater attention to non-adjacent regions and enhancing the flexibility of global associations:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 392, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 392, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 89, + 392, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m i x} = C E (y, \\hat {y}) - \\lambda \\| A s s D (P _ {f i x}, A) \\|, \\tag {12}", + "image_path": "2273935a7f32f978ce1a3f2f5f125ee130b913a3ae881f2902cf5f25f4b18efe.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "inline_equation", + "content": "P_{fix}" + }, + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "text", + "content": " denotes the fixed prior association matrix. This alternating optimization strategy stabilizes the training process and utilizes the frequency-domain association " + }, + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 54, + 415, + 296, + 487 + ], + "type": "text", + "content": " to guide the model in focusing on more nuanced frequency-domain details, effectively distinguishing between genuine and forged images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 499, + 138, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 138, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 138, + 513 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 519, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 296, + 591 + ], + "type": "text", + "content": "In the first part of this section, we introduce the detailed experimental setup. Next, in the second part, we evaluate and compare STFT with other forgery detection methods, including SOTA approaches. In the fourth part, we assess the generalization performance of STFT. Finally, we conduct an ablation study on the proposed model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 173, + 613 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 296, + 715 + ], + "type": "text", + "content": "Evaluation Metrics and Datasets. In alignment with the established comparison standards, we evaluate accuracy (ACC) and area under the curve (AUC) as performance metrics for the GenImage [53] and DeepFaceGen [2] datasets, respectively. The GenImage [53] dataset consists of a total of 2,681,167 images, including 1,331,167 real images and 1,350,000 generated images. These generated images are sourced from eight different generative models: BigGAN" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 418 + ], + "type": "text", + "content": "[3], GLIDE [30], VQDM [16], Stable Diffusion V1 and V1.5 [38], ADM [12], Midjourney [21], and Wukong [15]. All fake images are generated using the template prompt \"photo of CLS,\" where \"CLS\" is replaced by one of the 1,000 labels from ImageNet [11]. The dataset is divided into eight subsets, each containing a portion of real images along with all fake images generated by the aforementioned models. In this study, we follow the official dataset split, using 2,581,167 images for training and the remaining 100,000 images for validation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "text", + "content": "DeepFaceGen [2] is a facial forgery dataset that includes both video and image modalities. We select face forgery data generated using novel prompt-guided forgery methods. These methods cover 12 diffusion- and autoregressive-based models across text-to-image and image-to-image modalities. The dataset includes face forgeries across various ages, genders, and environments, with generated images sourced from popular models such as Midjourney [21], DALL-E [13], and SDXL [38]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details. We downloaded the pretrained LDM model, fed images into it, and extracted the latent space data. We set the default diffusion steps to 20, obtaining 20 latent representations of size " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 64 \\times 64" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ". Each latent representation was then processed using average pooling to reduce its size to " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 8" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " and flattened into a one-dimensional vector, ultimately forming sequential data of size " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "20 \\times 192" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ". We set the batch size to 256, the number of channels in the hidden state to 512, and the number of attention heads to 8. The hyperparameter " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " was set to 2 for all datasets. We used the ADAM optimizer [23] with an initial learning rate of " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": ". All experiments were conducted on an Nvidia A40 GPU." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17072" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 162 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 162 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 162 + ], + "type": "table", + "html": "
MethodYearMidjourneySDv1.4SDv1.5ADMGLIDEWukongVQDMBigGAN
DIRE [45]202399.5 / 72.699.7 / 68.899.7 / 68.899.2 / 67.999.8 / 52.299.7 / 68.199.9 / 51.9100 / 50.1
Defake [39]202398.8 / 78.799.8 / 81.5100 / 82.398.7 / 82.599.8 / 73.199.3 / 75.699.1 / 82.098.7 / 76.6
PatchCraft [52]202390.1 / 81.189.5 / 81.390.1 / 81.390.4 / 79.390.6 / 80.492.6 / 79.390.3 / 74.980.9 / 73.7
AIDE [49]202498.9 / 86.099.7 / 84.999.6 / 84.095.3 / 79.197.6 / 81.898.8 / 80.492.3 / 83.096.3 / 81.7
DRCT [6]202499.9 / 89.395.5 / 86.696.8 / 89.195.1 / 81.896.6 / 83.899.6 / 84.494.3 / 82.498.8 / 84.0
STFT (Ours)-99.4 / 92.399.6 / 89.199.3 / 90.499.5 / 82.599.8 / 87.0100 / 84.899.7 / 85.7100 / 85.6
", + "image_path": "2e02365c6c670331305792add9b6d48f6bc3e215b69e7949b0169b40b2336fad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 239, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 239, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 239, + 236 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 255, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 495 + ], + "type": "text", + "content": "GenImage Benchmark. To verify the model's generalization ability across different generative models, this section follows the GenImage evaluation standard and compares our approach with 12 mainstream forgery detection models, including CNNSpot [34], F3Net [33], CLIP/RN50 [34], GramNet [24], De-fake [39], Conv-B [26], SwinT [25], UnivFD [31], DIRE [45], PathCraft [52], AIDE [49], and DRCT [6]. According to the GenImage evaluation protocol, all models were trained on the SDv1.4 subset of GenImage and evaluated on multiple test sets. As shown in Figure 2a, nearly all methods achieve high accuracy on the Wukong, SDv1.4, and SDv1.5 datasets but perform poorly on GLIDE, VQDM, Midjourney, ADM, and BigGAN datasets. Since BigGAN is a non-diffusion model dataset, all methods perform the worst on it. In contrast, our proposed STFT achieves over " + }, + { + "bbox": [ + 56, + 255, + 296, + 495 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 56, + 255, + 296, + 495 + ], + "type": "text", + "content": " accuracy on the BigGAN dataset and obtains the highest ACC on 7 out of 8 sub-datasets, demonstrating a significant advantage over existing forgery detection methods (see Appendix Table 4 for complete results)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "DeepFaceGen Benchmark. Following the DeepFaceGen evaluation standard, we compare STFT with mainstream generative image detection methods, including Xception [7], EfficientNet-B0 [44], F3-Net [33], RECCE [4], DNADet [50], DIRE [45], DRCT [6], UnivFD [31], NPR [43], and FreqNet [42]. Training and testing were conducted on the same dataset, which includes real images and images generated by 12 different forgery methods such as DALL-E and SD. The dataset was split into training, validation, and test sets with a 7:1:2 ratio. As shown in Figure 2b, STFT outperforms all methods across all 12 subsets, with a particularly notable advantage in distinguishing difficult-to-detect forged images generated by DALL-E 1, DALL-E 3, Midjourney, and Wenxin. Overall, STFT achieves an AUC above " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " for all forgery categories, with an average AUC of " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "95.21\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " (detailed results are provided in Appendix Table 5)." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 221, + 556, + 299 + ], + "blocks": [ + { + "bbox": [ + 55, + 169, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 555, + 204 + ], + "type": "text", + "content": "Table 1. Performance comparison (ACC, %) of various methods. Each value is presented as (Score1 / Score2), where Score1 corresponds to the accuracy achieved when trained and tested on the same subset, and Score2 represents the average accuracy when trained on the current subset and evaluated across other subsets. Bold values highlight the best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 221, + 556, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 221, + 556, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 556, + 299 + ], + "type": "table", + "html": "
Train OnCatHorseBedroom
Test OnHorse/BedroomCat/BedroomCat/Horse
DIRE [45]73.0165.4555.04
DRCT [6]74.4868.7867.62
STFT (Ours)90.3289.7091.51
", + "image_path": "4d4dbd4e1040f9dd048ce4ede268885377a28c0142030e5846a8ece784ce813d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 305, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 328 + ], + "type": "text", + "content": "Table 2. Generalization performance (AUC,%) of different methods on non-face images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 349, + 470, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 470, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 470, + 361 + ], + "type": "text", + "content": "5.3. Generalization Performance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 426 + ], + "type": "text", + "content": "To evaluate the generalization capability of STFT, we designed two experimental scenarios: first, testing the model when the pre-trained model does not align with the target detection task, and second, assessing its adaptability to unknown forgery methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 629 + ], + "type": "text", + "content": "Adaptation to Unknown Forgery Methods. We trained STFT on a single subset of the GenImage dataset and tested it across all eight subsets. As shown in Table 1, we compared STFT with the top four models from the baseline evaluations. STFT outperformed the existing SOTA method in GenImage, DRCT, and achieved the best generalization performance across all eight subsets. When confronted with unseen forgery methods, all models experienced performance degradation in various detection metrics. However, STFT still demonstrated strong generalization ability. The superior generalization of STFT is primarily attributed to its frequency-domain-enhanced spatio-temporal correlation modeling, which captures multi-perspective features of forged images. This enhances the model's adaptability to different types of forgeries, allowing it to effectively recognize anomalous patterns even when facing previously unseen forgery methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "Generalization Performance on Non-face Images. Reconstruction-error-based forgery detection methods often fail when the test data differs from the reconstruction model's pre-training domain. For instance, applying a model trained on animals to plant images may misclassify real images as fake due to high reconstruction errors. To assess STFT's robustness in such cases, we used pre-trained" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17073" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 131 + ], + "type": "table", + "html": "
Model VariantTemporalSpatialFrequencyAUC(%)
Full Model✓✓✓92.02
w/o Temporal✓✓86.28
w/o Spatial✓✓79.71
w/o Frequency✓✓89.75
", + "image_path": "51844763180338786bd7969f6ab4d167bbd269bfb512c9cb64b0fb3c49c3647f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 140, + 257, + 150 + ], + "lines": [ + { + "bbox": [ + 91, + 140, + 257, + 150 + ], + "spans": [ + { + "bbox": [ + 91, + 140, + 257, + 150 + ], + "type": "text", + "content": "Table 3. Ablation study on different modules." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 172, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 256 + ], + "type": "text", + "content": "models from [12] on LSUN Bedroom, Cat, and Horse categories, and compared STFT with DIRE and DRCT. Following the Benchmark 1 protocol on 12 prompt-based forgery datasets, results in Table 2 show that while reconstruction-based methods suffered major performance drops, STFT consistently achieved AUC scores over " + }, + { + "bbox": [ + 55, + 172, + 296, + 256 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 172, + 296, + 256 + ], + "type": "text", + "content": ", highlighting its strong generalization ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 149, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 149, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 149, + 277 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 353 + ], + "type": "text", + "content": "In this section, we conduct ablation experiments to quantify the contribution of each design component in our model and analyze the impact of key hyperparameters. Specifically, we examine the effect of the latent space sampling steps and feature map resolution and the inclusion of the frequency-enhanced attention module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "Effect of Latent Space Sampling Steps and Feature Map Resolution. To investigate the impact of sampling steps and feature resolution in latent space, we evaluate detection performance with sampling steps set to 10, 15, 20, 25, and 30, and test four resolutions by downsampling latent features via average pooling. As shown in Figure 3, the model performs best with 20 sampling steps and a feature resolution of " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 8" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "text", + "content": "This indicates that 20 steps sufficiently preserve temporal cues while preventing over-smoothing. A " + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 8" + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "text", + "content": " resolution balances detail retention and generalization. Lower resolutions (e.g., " + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "inline_equation", + "content": "3 \\times 4 \\times 4" + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "text", + "content": ") lose crucial spatial information, reducing accuracy, while higher resolutions introduce redundancy, affecting robustness. Thus, 20 sampling steps and " + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 8" + }, + { + "bbox": [ + 55, + 449, + 296, + 545 + ], + "type": "text", + "content": " resolution are adopted as default settings for optimal accuracy and efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "type": "text", + "content": "Impact Analysis of Model Components. To analyze the contributions of different STFT components, we conduct ablation experiments by sequentially removing the temporal prior correlation module, spatial correlation module, and frequency-enhanced attention module on DALL-E 3 Dataset. The results, presented in Table 3, confirm their complementary roles in forgery detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Removing the temporal module significantly weakens the model's ability to capture sequential forgery patterns, leading to a notable performance drop. This aligns with our observation that latent diffusion smooths temporal transitions in forged images. Without spatial correlation modeling, the model struggles to detect local modifications, increasing false positives and reducing robustness. Frequency" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 427, + 180 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 427, + 180 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 427, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 427, + 180 + ], + "type": "image", + "image_path": "b833215d04f584db9a9227dad50305edda3acc39b74bcedc312382a20fe5f2cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 188, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 555, + 222 + ], + "type": "text", + "content": "Figure 3. The ablation study on diffusion steps (a) and latent feature map size (b). Solid lines in both cases represent the performance results corresponding to the final adopted values." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 554, + 180 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 554, + 180 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 554, + 180 + ], + "type": "image", + "image_path": "9c04413263ed1255f384ca91c423de398ac5b97699cc35f3db8c515ee3dd0438.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 291 + ], + "type": "text", + "content": "features are crucial for detecting high-frequency artifacts in forgeries. Removing this module degrades performance, especially against unknown forgery methods, highlighting its role in generalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 301, + 388, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 388, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 388, + 314 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "content": "This paper proposes the Spatial-Temporal Forgery Trace (STFT) method for forgery image identification. Leveraging the tendency of forged images to fit specific distributions more easily than real ones, STFT employs generative diffusion models to capture temporal distributions and models relationships between temporal and spatial features as forgery traces. It also integrates frequency-domain features to accelerate trace localization. Experiments show that STFT effectively identifies subtle spatio-temporal traces by combining spatial, temporal, and frequency perspectives, demonstrating strong robustness and generalizability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 453, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 586 + ], + "type": "text", + "content": "Limitations and Future Works. Although the proposed STFT achieves high accuracy and demonstrates commendable generalization capabilities, it is not without its limitations. Firstly, the process of mapping the original image into the latent diffusion distribution is computationally intensive and time-consuming. Additionally, the method appears to exhibit suboptimal performance on GAN-based forgery images, likely due to fundamental differences in the generative mechanisms of GANs and generative diffusion models. In future work, we aim to enhance the efficiency of the proposed method and further extend its generalization ability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 596, + 424, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 424, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 424, + 609 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 664 + ], + "type": "text", + "content": "This work is funded by National Key Research and Development Project (Grant No: 2022YFB2703100), Ant Group through CCF-Ant Research Fund, and Information Technology Center, ZheJiang University." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17074" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] Baidu AI. Wenxin large model for text-to-image generation. Baidu AI Blog, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 115, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 115, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 115, + 295, + 158 + ], + "type": "text", + "content": "[2] Yijun Bei, Hengrui Lou, Jinsong Geng, Erteng Liu, Lechao Cheng, Jie Song, Mingli Song, and Zunlei Feng. A largescale universal evaluation benchmark for face forgery detection. arXiv preprint arXiv:2406.09181, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 294, + 193 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 249 + ], + "type": "text", + "content": "[4] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4113-4122, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 295 + ], + "type": "text", + "content": "[5] Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt. arXiv preprint arXiv:2303.04226, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 350 + ], + "type": "text", + "content": "[6] Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. Drct: Diffusion reconstruction contrastive training towards universal detection of diffusion generated images. In Forty-first International Conference on Machine Learning, 2024. 7, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "text", + "content": "[7] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 463 + ], + "type": "text", + "content": "[8] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 973-982, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 530 + ], + "type": "text", + "content": "[9] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 586 + ], + "type": "text", + "content": "[10] Edmar RS De Rezende, Guilherme CS Ruppert, Antonio Theophilo, Eric K Tokuda, and Tiago Carvalho. Exposing computer generated images by using deep convolutional neural networks. Signal Processing: Image Communication, 66: 113-126, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 667 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Tira Nur Fitria. Artificial intelligence (ai) technology in openai chatgpt application: A review of chatgpt in writing english essay. In ELT Forum: Journal of English Language Teaching, pages 44-58, 2023. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[14] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning, pages 3247-3258. PMLR, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "text", + "content": "[15] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Niu Minzhe, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, et al. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. Advances in Neural Information Processing Systems, 35:26418-26431, 2022. 6, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 553, + 251 + ], + "type": "text", + "content": "[16] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 253, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 295 + ], + "type": "text", + "content": "[17] Weinan Guan, Wei Wang, Jing Dong, Bo Peng, and Tieniu Tan. Collaborative feature learning for fine-grained facial forgery detection and segmentation. arXiv preprint arXiv:2304.08078, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 341 + ], + "type": "text", + "content": "[18] Tahir Hassan, Aras Asaad, Dashti Ali, and Sabah Jassim. Artificial image tampering distorts spatial distribution of texture landmarks and quality characteristics. arXiv preprint arXiv:2208.02710, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 375 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "type": "text", + "content": "[20] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "text", + "content": "[21] Anna Jaruga-Rozdolska. Artificial intelligence as part of future practices in the architect's work: Midjourney generative tool as part of a process of creating an architectural form. Architectus, (3 (71):95-104, 2022. 6, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[22] Wei et al. Jiang. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16423-16432, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 555 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "type": "text", + "content": "[24] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8060-8069, 2020. 7, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[26] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17075" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[27] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[28] Yunpeng Luo, Junlong Du, Ke Yan, and Shouhong Ding. Lare^2: Latent reconstruction error based method for diffusion-generated image detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17006-17015, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[29] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. arXiv preprint arXiv:2307.06272, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[30] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 6, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 318 + ], + "type": "text", + "content": "[31] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24480-24489, 2023. 7, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 342 + ], + "type": "text", + "content": "[32] OpenAI. Dall-e 3: Openai's latest text-to-image model. OpenAI Blog, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[33] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 7, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 7, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 476 + ], + "type": "text", + "content": "[35] Aditya et al. Ramesh. Zero-shot text-to-image generation. arXiv preprint arXiv:2102.12092, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "text", + "content": "[36] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. arXiv preprint arXiv:2210.14571, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 566 + ], + "type": "text", + "content": "[37] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[39] Zeyang Sha, Zheng Li, Ning Yu, and Yang Zhang. De-fake: Detection and attribution of fake images generated by text-to-image generation models. In Proceedings of the 2023 ACM SIGSAC conference on computer and communications security, pages 3418-3432, 2023. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 679, + 214, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 679, + 214, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 679, + 214, + 689 + ], + "type": "text", + "content": "[40] Stability.ai. Stable Diffusion, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[41] Eren Tahir and Mert Bal. Deep image composition meets image forgery. arXiv preprint arXiv:2404.02897, 2024. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[42] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Frequency-aware deepfake detection: Improving generalizability through frequency space domain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5052-5060, 2024. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[43] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 196, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 238 + ], + "type": "text", + "content": "[44] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[45] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "type": "text", + "content": "[46] Jing Weng. Local frequency analysis for diffusion-generated image detection. In International Conference on Image Processing and Artificial Intelligence (ICIPAL 2024), pages 66-73. SPIE, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[47] Moritz Wolter, Felix Blanke, Raoul Heese, and Jochen Garcke. Wavelet-packets for deepfake image analysis and detection. Machine Learning, 111(11):4295-4327, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 418 + ], + "type": "text", + "content": "[48] Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. arXiv preprint arXiv:2110.02642, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 463 + ], + "type": "text", + "content": "[49] Shilin Yan, Ouxiang Li, Jiayin Cai, Yanbin Hao, Xiaolong Jiang, Yao Hu, and Weidi Xie. A sanity check for ai-generated image detection. arXiv preprint arXiv:2406.19435, 2024. 7, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[50] Tianyun Yang, Ziyao Huang, Juan Cao, Lei Li, and Xirong Li. Deepfake network architecture attribution. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4662-4670, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[51] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[52] Nan Zhong, Yiran Xu, Sheng Li, Zhenxing Qian, and Xinpeng Zhang. Patchcraft: Exploring texture patch for efficient ai-generated image detection. arXiv preprint arXiv:2311.12397, 2023. 7, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 599, + 553, + 654 + ], + "type": "text", + "content": "[53] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. Advances in Neural Information Processing Systems, 36:77771-77782, 2023. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17076" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_content_list.json b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9f4b13f5c8fb1c01ffdc048eaf16dc27c4659e4 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_content_list.json @@ -0,0 +1,1837 @@ +[ + { + "type": "text", + "text": "SpatialCrafter: Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations", + "text_level": 1, + "bbox": [ + 148, + 128, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songchun Zhang $^{1,2}$ Huiyao Xu $^{1}$ Sitong Guo $^{1}$ Zhongwei Xie $^{2}$ \nHujun Bao $^{1}$ Weiwei Xu $^{1}$ Changqing Zou $^{1,3*}$ $^{1}\\mathrm{ZJU}$ ${}^{2}\\mathrm{HKUST}$ ${}^{3}\\mathrm{Zhejiang Lab}$", + "bbox": [ + 227, + 202, + 769, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70f4b821144851b348978c011d78cc1e1ec9d2397c7b94d94aefe37562ef2577.jpg", + "image_caption": [ + "Figure 1. Overview. Our method, SpatialCrafter, generates additional novel views of a scene from few inputs by leveraging camera trajectory-guided video diffusion models, which alleviates the ambiguity of sparse view reconstruction and robustly reconstructs the scene from the generated video frames. It demonstrates impressive performance in both indoor and outdoor scenes, while also exhibiting generalization capabilities in real scenes and synthetic images. More visualization results can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 94, + 270, + 330, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d656c76546a0eef9f4f920cd6ee3fa42d7eb94acd124dc8a5e8c87e507cd749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 270, + 905, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 681, + 326, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis (NVS) boosts immersive experiences in computer vision and graphics. Existing techniques, though progressed, rely on dense multi-view observations, restricting their application. We tackle the task of reconstructing photorealistic 3D scenes from only one or a few input views. We introduce SpatialCrafter, a framework that leverages the rich knowledge in video diffusion models to generate plausible additional observations, thereby alleviating reconstruction ambiguity. Through a trainable camera encoder and an epipolar attention mechanism for explicit geometric constraints, we achieve precise camera control and 3D consis", + "bbox": [ + 89, + 713, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tency, further reinforced by a unified scale estimation strategy to handle scale discrepancies across datasets. Furthermore, by integrating monocular depth priors with semantic features in the video latent space, our framework directly regresses 3D Gaussian primitives and efficiently processes long-sequence features using a hybrid network structure. Extensive experiments show our method enhances sparse view reconstruction and restores the realistic appearance of 3D scenes. Project page: https://franklinz233.github.io/projects/spatialcrafter/.", + "bbox": [ + 511, + 683, + 908, + 835 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 114, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27794", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis (NVS), as a key technology in computer vision and graphics, provides crucial support for immersive experiences in fields such as video games and mixed reality. Although neural reconstruction techniques have made significant progress in recent years [25, 39, 42, 62, 66], these methods typically rely on dense multiview observation data, which face numerous limitations in practical applications. Therefore, this paper focuses on a more practically valuable challenge: how to achieve high-quality 3D scene reconstruction and synthesize realistic novel views from sparse or even single-view observation.", + "bbox": [ + 89, + 114, + 480, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This problem is hard because (1) real-world scenes are diverse and complex, (2) occluded regions are unseen in sparse captures, and (3) geometric cues are too few to constrain reconstruction. To address these challenges, some methods [17, 26, 29, 40, 84, 88] constrain the optimization process by introducing regularization terms, but these methods need to be optimized for each scene, are not generalizable, and are difficult to cope with complex scenarios. Recently, some feed-forward methods [7, 8, 22, 50, 55, 60, 79, 80] have achieved generalizable sparse view scene reconstruction by directly predicting 3D Gaussian primitives for each pixel. However, such methods produce severe artifacts when reconstructing occluded regions and applying them to the extrapolation setting. To address this challenge, some methods [35, 45, 64, 65] have introduced the priors from the diffusion model to achieve sparse or single-view NVS, but they perform poorly on scene-level data and lack precise pose control and 3D consistency. This is due to the fact that low-dimensional camera information such as Euler angles and extrinsic struggle to provide comprehensive control signals to the generative models.", + "bbox": [ + 89, + 281, + 482, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a framework named SpatialCrafter for scene reconstruction and novel view synthesis based on sparse or single-view inputs. Our innovation lies in leveraging the rich physical world knowledge embedded in video diffusion models to provide plausible additional observations for scene reconstruction, thereby effectively reducing the problem complexity. Specifically, we first focus on enhancing precise camera control and 3D consistency in generated videos. To achieve this, we parameterize camera settings using ray embeddings [78] or metric depth-warped images [77] and incorporate them into the video diffusion model via a trainable camera encoder. Furthermore, we propose an epipolar attention mechanism that improves 3D consistency between video frames through explicit geometric constraints. Previous methods [19, 64] often experience performance degradation and difficulty generating large-motion videos when trained on multiple scene datasets. We identified that these issues primarily stem from scale ambiguity in scene datasets. To overcome this, we introduce a unified scale estimator to calibrate scene dataset,", + "bbox": [ + 89, + 599, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enabling effective joint training across multiple datasets.", + "bbox": [ + 511, + 90, + 883, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although we can generate visually coherent video sequences, relying solely on generated frames to reconstruct general scenes often leads to suboptimal solutions, particularly in large-scale outdoor environments and stylized scenes. To address this, we propose incorporating monocular depth priors with rich semantic features extracted from the video latent space. Leveraging these latent features, our method directly regresses 3D Gaussian primitives of the scene via a feed-forward network. Furthermore, to efficiently handle long-sequence feature interactions, we design a hybrid architecture that integrates Mamba blocks with Transformer blocks. Experiments show that our method not only improves the sparse-view reconstruction, but also accurately restores the appearance of 3D scenes, especially when extrapolating from a single view and when there are few sparse view overlaps. In conclusion, our key contributions can be summarized as follows:", + "bbox": [ + 511, + 108, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce a framework that effectively utilizes the physical-world knowledge embedded in video diffusion models to provide additional plausible observations for sparse-view scene reconstruction, thus reducing the ambiguity of sparse view scene reconstruction.", + "bbox": [ + 513, + 364, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- To address the scale ambiguity problem that occurs in joint training across datasets, we develop a unified scale estimation approach for trajectory calibration. This solves the performance degradation problem, thus enabling effective multi-dataset training.", + "bbox": [ + 513, + 441, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We combine monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives through a feed-forward manner. Meanwhile, we propose a hybrid architecture integrating Mamba blocks with Transformer blocks to efficiently handle long-sequence feature interactions.", + "bbox": [ + 513, + 516, + 903, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 625, + 653, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Sparse-View Scene Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 650, + 815, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF [62] and Gaussian Splatting [25] have achieved photorealistic representations of 3D scenes, but require optimization on densely collected image sets for each scene. To address this problem, some approaches [17, 29, 71, 88] focus on constraining the optimization process of 3D representations via manually designed regularization. Some other methods [7, 8, 23, 54, 58, 69, 72, 83] are trained on large-scale datasets and directly predict the 3D representation in a feed-forward manner. Recently, some approaches [13, 50] based on end-to-end 3D reconstruction models [7, 28, 60], have achieved efficient and high-quality 3D reconstruction. However, these methods cannot handle occluded regions, and reconstruction failures (e.g., under large viewpoint changes) can severely affect the quality of novel view synthesis. More relevant to our work, some", + "bbox": [ + 511, + 672, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27795", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d9e64276501703ed894cc268c7aa2232e8544bf6a501f4edad082b4305bd1ca4.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. Our generative reconstruction pipeline consists of two parts: camera-controlled video generation, and then video-based reconstruction. First, we set the exploration path based on the input views. In the video generation module, ray embeddings are used to parameterize the camera trajectories, and a cross-attention mechanism with epipolar constraints is introduced to improve the 3D consistency of the generated video. The video-based scene reconstruction pipeline integrates monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives via a feed-forward manner." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "works [31, 34, 57] have employed priors from video diffusion models to enhance the performance of sparse-view reconstruction.", + "bbox": [ + 88, + 464, + 482, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Diffusion-based Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 517, + 415, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since text-to-image diffusion models [44] contain rich natural image priors, some methods [41, 63] directly optimize 3D representations [25, 62] via score distillation sampling [41]. Several approaches [6, 35, 45, 56] enable zero-shot novel view synthesis by fine-tuning the image and pose-conditioned generative models on large-scale 3D object [10] or scene [56, 87] datasets. However, they still have difficulty synthesizing consistent novel views. To solve this, some approaches [5, 15, 36, 38, 47, 65, 85] generate multi-views simultaneously and model the correlation between multiple views with ground-truth poses. Some recent methods [30, 37] proposed to use large-scale unposed images to train NVS methods. In addition, some methods [9, 12, 14, 48, 76, 82] utilize depth-based warping to synthesize novel views and employ the T2I model to inpaint the warped images. However, the novel views generated by these methods tend to suffer from artifacts and cumulative errors, limiting their use in generalized scenes.", + "bbox": [ + 88, + 539, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Controllable Video Generation", + "text_level": 1, + "bbox": [ + 89, + 818, + 364, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While recent text-to-video diffusion models have achieved remarkable progress, they inherit the controllability limitations of their text-to-image counterparts, often requiring additional conditioning mechanisms to align generated", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "videos with user intent. Several works [18, 24, 61, 70] have been carried out to introduce a variety of conditions into video generation models. Recently, attention has focused on controlling the camera motion of the video. Some approaches [18, 53] employs motion-specific training for predefined camera movements, though this framework struggles with complex trajectory synthesis. Some methods [1, 19, 20, 64, 67] enable more complex camera control by parameterizing the camera trajectory and injecting it into a pre-trained video diffusion model via a trainable camera encoder. Another line of work [34, 43, 74, 77] advances this direction by enabling pixel-accurate view compositing through a point-based rendering approach.", + "bbox": [ + 511, + 464, + 906, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 675, + 604, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section begins by explaining our camera-controlled video generation model (Section 3.1), which employs ray embeddings for precise pose control. Then, we describe how to use epipolar geometry constraints to enhance 3D consistency in the video frames. Next, we introduce our scene reconstruction pipeline (Section 3.2), which aims to perform stable 3D reconstruction based on the generated video latents.", + "bbox": [ + 511, + 700, + 906, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Camera-Conditioned Video Generation", + "text_level": 1, + "bbox": [ + 511, + 832, + 852, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale Alignment. To address this challenge, our approach first employs VGGT [58] to estimate initial camera parameters, represented by the extrinsic matrix $\\mathbf{P} = [\\mathbf{R}|\\mathbf{T}]$ , and", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27796", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a corresponding depth map $d_v$ for each video frame. To establish a consistent scale across these datasets, we leverage the Metric3D [73] to infer a canonical metric depth map, denoted $d_m$ . We then align our initial depth prediction, $d_v$ , to the canonical scale by calculating a scale factor $s$ , defined as the ratio of the inter-percentile ranges (IPR) of the two depth maps: $s = \\mathrm{IPR}_{0.8,0.2}(d_v) / \\mathrm{IPR}_{0.8,0.2}(d_m)$ . This scale factor is then used to adjust the translation component of the camera extrinsics, resulting in $\\mathbf{P}_{\\text{scaled}} = [\\mathbf{R}|s \\cdot \\mathbf{T}]$ . This procedure yields camera extrinsics with a consistent absolute scale, effectively resolving alignment issues when fusing data from different sources.", + "bbox": [ + 89, + 90, + 483, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Camera Injection. Inspired by recent methods [19, 49, 78], we chose to use ray embeddings [19] or the depth warping frames [77] to represent the camera information. Specifically, for a ray defined by the origin $o \\in \\mathbb{R}^3$ and the normalized direction $d \\in \\mathbb{R}^3$ , we can represent it as $(o \\times d, d) \\in \\mathbb{R}^6$ . Given the camera parameters, the direction of the ray $d'$ corresponding to the pixel coordinates $(u, v)$ can be calculated as $d' = \\mathbf{RK}^{-1}(u, v, 1)^{\\top} + \\mathbf{T}$ . For depth warping frames, we leverage the pretrained depth estimation model [59] to map the reference image into a 3D point cloud. Subsequently, we render novel views at specified camera poses by projecting this point cloud via the target camera parameters. Afterwards, we use a trainable conditional encoder to inject the camera information into the video model instead of fine-tuning all model parameters. The training objective is:", + "bbox": [ + 89, + 272, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}; z _ {0}, t, \\phi (C)\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 508, + 480, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi (C)$ represents the camera conditional encodings, $z_{0}$ denotes the latent embeddings of the reference frame, and $t$ indicates the timestamps.", + "bbox": [ + 89, + 537, + 482, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Epipolar Feature Aggregation. Camera embedding enhances control but 3D video consistency remains challenging due to dense self-attention allowing unrestricted cross-frame pixel interactions. We address this by incorporating epipolar geometric constraints, and aggregating features along epipolar lines. For a pixel $\\pmb{p} = (u,v)$ in frame $i$ , its corresponding epipolar line $l_{ik}$ in frame $k$ is computed as $l_{ik}(\\pmb{p}) = \\pmb{F}_{ik} \\cdot \\tilde{\\pmb{p}}$ , where $\\tilde{\\pmb{p}} = (u,v,1)^{\\top}$ are the homogeneous coordinates, and $\\pmb{F}_{ik} \\in \\mathbb{R}^{3 \\times 3}$ is the fundamental matrix. The fundamental matrix is decomposed as $\\pmb{F}_{ik} = \\pmb{K}_k^{-\\top} \\pmb{E}_{ik} \\pmb{K}_i^{-1}$ , where $\\pmb{K}_i, \\pmb{K}_k \\in \\mathbb{R}^{3 \\times 3}$ are the camera intrinsics, and $\\pmb{E}_{ik}$ is the essential matrix. We then convert epipolar lines into attention masks by computing per-pixel distances and applying a threshold, restricting attention to geometrically valid regions.", + "bbox": [ + 89, + 584, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse-View Setting. To adapt to the sparse view input setting, we formulate the task as a video frame interpolation problem conditioned on the given start and end frames. To maximally preserve the priors from the pre-trained models [4], we inject boundary frame conditions in both latent and semantic space. Specifically, we combine latent fea", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tures of the first and last frames with their noisy latent representations, and then concatenate the extracted CLIP embeddings from both frames for cross-attentional feature injection. The training objective is:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, z _ {n} \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; z _ {0}, z _ {n}, t, \\phi (C)) \\| _ {2} ^ {2} \\right] \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 162, + 906, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{n}$ and $z_{0}$ are the latent embeddings of the final and initial frames, respectively. During the inference phase, we use VGGT [58] to estimate the extrinsic and intrinsic parameters for the input views.", + "bbox": [ + 511, + 200, + 905, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Video-based Scene Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 273, + 815, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. There are several challenges in reconstructing scenes directly from generated videos. First, the limited number of generated frames makes it difficult to capture complete scene information. In addition, the diverse styles of video frames can pose a challenge to traditional reconstruction techniques. Furthermore, the generated video frames may contain imperfect or low-quality areas, making the reconstruction process unstable. Moreover, In stylized scenes, previous methods often yield poor results, as recovering poses and sparse point clouds from video frames is difficult, leading to many artifacts. To address this, a powerful reconstruction module specifically designed to handle the generated videos is required.", + "bbox": [ + 511, + 295, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latent Feature Fusion. Given input video latents $\\mathbf{z} \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ and camera poses encoded as ray embeddings $\\mathbf{p} \\in \\mathbb{R}^{T \\times H \\times W \\times 6}$ , we transform them into token sequences through a patchification process. Specifically, we apply spatial patchification to latent features to obtain latent tokens $\\mathbf{z}_t \\in \\mathbb{R}^{N \\times d_z}$ , while ray embeddings undergo 3D-patchification to produce pose tokens $\\mathbf{p}_t \\in \\mathbb{R}^{N \\times d_p}$ . To incorporate explicit geometric guidance, we leverage monocular depth estimation within our pipeline. For each input RGB video frame, we estimate dense depth maps $\\mathbf{D} \\in \\mathbb{R}^{T \\times H \\times W \\times 1}$ , which are processed by a dedicated depth encoder to yield depth tokens $\\mathbf{d}_t \\in \\mathbb{R}^{N \\times d_d}$ . The three token sets are channel-wise concatenated to form a unified representation $\\mathbf{x} = [\\mathbf{z}_t; \\mathbf{p}_t; \\mathbf{d}_t] \\in \\mathbb{R}^{N \\times (d_z + d_p + d_d)}$ , which is then linearly projected to a lower-dimensional space $\\mathbf{x}' \\in \\mathbb{R}^{N \\times d}$ before being fed into a sequence of mamba and transformer blocks. Mamba [16] achieves the same token sequence processing functionality as Transformer, but reduces computational complexity from $\\mathcal{O}(L^2)$ to $\\mathcal{O}(L)$ , making it particularly suitable for dense reconstruction tasks. In our Mamba blocks, we implement bi-directional scanning across the token sequence. First, we compute state parameters $\\mathbf{A}, \\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times d}$ from the input tokens using a linear projection. Then, we execute the State Space Model in both forward ( $\\mathbf{y}_f$ ) and backward ( $\\mathbf{y}_b$ ) directions. The final output of each Mamba block is computed as $\\mathbf{y} = \\mathbf{y}_f + \\mathbf{y}_b$ followed by a final linear transformation.", + "bbox": [ + 511, + 493, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27797", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e753cdbf55dd4c8f8b1103458c77bd44f1c8332731ebfb8f624ac09d27ad0a35.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of generated videos. Compared to other benchmark methods, the videos generated by our method have better foreground object-background consistency, while also being able to generate videos with larger motion amplitude." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 357, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0bc77672b096c01e27dc68ae50d8f7993d0a47a9d6745badef542dc876c60e4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 617, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a81772d829b6c6620a3627e29da7927d39bdbef2a99355459a4ab12757be764e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 89, + 903, + 261 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/462d41b077ad1dc79923f54d03f5ffd8915910edd97044606efd0de7b105cdfa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFVD↓FID↓Rerr↓Terr↓LPIPS↓PSNR↑SSIM↑
RealEstate10K
MotionCtrl [64]22.65230.120.2340.7980.29914.720.404
CameraCtrl [19]21.48188.210.0540.1270.23017.330.516
ViewCrafter [77]20.96204.180.0550.1530.21518.950.503
Ours18.25183.250.0520.1030.20719.210.523
Tanks-and-Temples
MotionCtrl [64]30.25290.380.8381.5050.31514.640.388
CameraCtrl [19]24.41244.820.1180.2940.28515.380.469
ViewCrafter [77]22.50231.350.1260.3070.24716.240.508
Ours20.17192.520.0970.1750.22816.120.501
DL3DV
MotionCtrl [64]25.65249.510.4731.1180.31214.380.386
CameraCtrl [19]22.76233.540.0950.2380.26216.330.489
ViewCrafter [77]20.59211.240.0930.2440.24217.120.521
Ours18.21171.520.0630.1340.22517.020.537
", + "bbox": [ + 91, + 324, + 483, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison to the camera conditioned video generation method on RealEstate10K [86], DL3DV [32], and Tanks-and-Temple [27] dataset.", + "bbox": [ + 89, + 496, + 482, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Decoding. We design a lightweight decoder module that efficiently transforms output feature tokens into per-pixel Gaussian parameters. The decoding module consists of 3D-DeConv layers, which generates refined Gaussian feature map $G \\in \\mathbb{R}^{(T \\times H \\times W) \\times 12}$ . This 12-channel representation precisely encodes the complete set of Gaussian parameters: RGB color (3 channels), scale factors (3 channels), rotation quaternion (4 channels), opacity (1 channel), and ray distance (1 channel). The final output of the model is the merge of 3D Gaussians from all input video frames.", + "bbox": [ + 88, + 561, + 482, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective. During training, we render images from predicted Gaussians using randomly selected supervision views. Our approach employs a composite loss that integrates three components:", + "bbox": [ + 89, + 715, + 483, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {m s e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {d e p t h}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 792, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{mse}$ represents pixel-wise mean squared error, $\\mathcal{L}_{perc}$ denotes perceptual loss, and $\\mathcal{L}_{depth}$ enforces depth consistency. This formulation jointly optimizes for photometric accuracy, high-level perceptual fidelity, and geometric coherence.", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 325, + 638, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment settings", + "text_level": 1, + "bbox": [ + 511, + 354, + 702, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To comprehensively capture the underlying distribution of real-world scenarios, we trained our video diffusion model on three diverse datasets: RealEstate-10K [86], ACID [33], and DL3DV-10K [32]. The Re10K dataset from YouTube comprises 67,477 training and 7,289 testing indoor and outdoor camera trajectories. The ACID dataset focuses on natural landscapes, with 11,075 training and 1,972 testing scenes. The DL3DV-10K dataset is a large-scale collection featuring 10,510 scenes captured under controlled, standardized conditions.", + "bbox": [ + 511, + 380, + 906, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our video generation model is based on SVD [3], an image-video diffusion model based on UNet. In our experiments, we use a relative camera system in which all camera poses are converted to poses relative to the first frame. The camera in the first frame is located at the world origin. In the first stage, we train the model at a resolution of $320 \\times 512$ for 50,000 iterations with the frame length set to 25. Subsequently, we train on $576 \\times 1024$ for 10,000 iterations to adapt to high resolution. The learning rate is set to $1e - 5$ with a warmup of 1,000 steps, using the Adam optimizer. We chose Lightning as the training framework, using mixed-precision fp16 and DeepSpeed ZeRO-2. We trained the proposed method and its variants on 16 NVIDIA A800 GPUs with a batch size of 32. During inference, we adopt DDIM sampler [51] with classifier-free guidance [21].", + "bbox": [ + 511, + 534, + 908, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Progressive Training. Controlling a video generation model to generate arbitrary motion trajectories remains a challenging task. To enable robust arbitrary trajectory generation, we employ a three-stage curriculum learning strategy. The model first trains on smooth camera motions with small temporal intervals, then progressively adapts to more complex motion patterns through linear interval scheduling, and finally incorporates random sampling intervals.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27798", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d022385becdeb84df97778ce77a32a18b926b872fb16b84edc4be7c98dc9c977.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on DL3DV and Tank-and-Temples dataset. Our method reconstructs better than baselines, even with limited image overlap and in non-overlapping regions." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 491, + 327 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e21525f7f905fc1c3db96aa87bffc3f160ca5def693bbdbfefe13ba1e1f922f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3-view6-view9-view
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Mip-NeRF360
Zip-NeRF [2]12.770.2710.70513.610.2840.66314.300.3120.633
ReconFusion [65]15.500.3580.58516.930.4010.54418.190.4320.511
CAT3D [15]16.620.3770.51517.720.4250.48218.670.4600.460
ReconX [34]17.160.4350.40719.200.4730.37820.130.4820.356
ViewCrafter [77]14.510.3150.68215.870.3360.66516.450.3480.655
Ours17.320.4390.41819.420.5100.37120.190.5360.345
DTU
Zip-NeRF [2]9.180.6010.3838.840.5890.3709.230.5920.364
FSGS [88]17.340.8180.16921.550.8800.12724.330.9110.106
ReconFusion [65]20.740.8750.12423.620.9040.10524.620.9210.094
CAT3D [15]22.020.8440.12124.280.8990.09525.920.9280.073
ViewCrafter [77]15.630.5220.38315.030.5250.40814.920.4870.452
Ours27.920.8790.10328.820.8950.08229.030.9050.065
LLFF
Zip-NeRF [2]17.230.5740.37320.710.7640.22123.630.8300.166
FSGS [88]20.310.6520.28824.200.8110.17325.320.8560.136
ReconFusion [65]21.340.7240.20324.250.8150.15225.210.8480.134
CAT3D [15]21.580.7310.18124.710.8330.12125.630.8600.107
ViewCrafter [77]17.730.5210.33217.430.5120.34517.330.4880.351
Ours22.040.7410.16525.110.8140.12425.950.8380.103
", + "bbox": [ + 93, + 397, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This gradual transition from simple to complex trajectories proves crucial for high-quality video generation with arbitrary trajectories.", + "bbox": [ + 89, + 619, + 483, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 89, + 676, + 230, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Controllable Video Generation", + "text_level": 1, + "bbox": [ + 89, + 699, + 352, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark and Metrics. We evaluate our video generation model against three baselines [19, 64, 77] on three benchmark datasets: RealEstate10K [86] (500 randomly selected videos with first frame as image condition and subsequent frames at stride 3 for pose guidance), DL3DV-140 [32] (500 video clips at stride 2), and Tanks-and-Temples [27] (100 sequences at stride 4 from 14 scenes with COLMAP-annotated poses) for out-of-domain generalization testing.", + "bbox": [ + 89, + 719, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Our evaluation framework employs multiple metrics to assess performance: visual quality measured through Fréchet Video Distance (FVD) and Fréchet Inception Dis", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d557e1f8a2c4c4460246b66655ce13750d2a53e1f471757a24ae090f0b468848.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on Re10K [86]. Compared to baselines, we obtain superior reconstruction from limited overlap, and enhanced geometry reconstruction in non-overlapping regions." + ], + "image_footnote": [], + "bbox": [ + 506, + 92, + 903, + 313 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/107a1fca9a81bb69ea2abe54c37652461b137594e1acbe595b70c3df125dccce.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of sparse view 3D reconstruction on Out-of-Domain Datasets." + ], + "table_footnote": [], + "table_body": "
MethodRealEstate10KACID
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
pixelNeRF [75]20.430.5890.55020.970.5470.533
GPNR [52]24.110.7930.25525.280.7640.332
AttnRend [11]24.780.8200.21326.880.7990.218
MuRF [68]26.100.8580.14328.090.8410.155
pixelSplat [7]25.890.8580.14228.140.8390.150
MVSplat [8]26.390.8690.12828.250.8430.144
GS-LRM [79]28.100.8920.114---
DepthSplat [69]24.230.7900.217---
ReconX [34]28.310.9120.08828.840.8910.101
ViewCrafter [77]24.220.7880.21823.480.6600.299
Ours28.350.8620.12128.930.8990.116
", + "bbox": [ + 514, + 397, + 906, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative comparison of two-view sparse view reconstruction methods on RealEstate10K [86] and ACID [33] dataset.", + "bbox": [ + 511, + 547, + 906, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tance (FID); camera control precision quantified by rotation error $(R_{err})$ and translation error $(T_{err})$ computed from camera poses extracted via COLMAP and normalized relative to the first frame; and visual consistency evaluated using PSNR, SSIM, and LPIPS [81] between generated and ground-truth views. To ensure fair comparison across methods with varying output capabilities, we restrict visual similarity assessment to the first 14 frames, as generated content typically diverges progressively from the conditional single-view input as the scene extends.", + "bbox": [ + 511, + 612, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison. Table 1 shows that our method excels in qualitative results, outperforming other methods in terms of visual quality, pose control, and 3D consistency. Figure 3 further demonstrates our advantage in 3D consistency and visual quality. This performance improvement is attributed to the introduction of a camera parameterization method and an epipolar attention module, which respectively enhance pose control ability and the 3D consistency of generated videos. In addition, training on a diverse scene dataset sig", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27799", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c246ae6ab063c5553f37ed2fba94bf0165351eb4e500fdb5ae9a5efe582114af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSmallLarge
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
pixelNeRF [75]18.4170.6010.52620.8690.6390.458
AttnNeRF [11]19.1510.6630.36825.8970.8450.229
pixelSplat [7]20.2630.7170.26627.1510.8790.122
MVSplat [8]20.3530.7240.25427.4080.8840.116
DUSt3R [60]14.1010.4320.46816.4270.4530.402
CoPoNeRF [22]17.3930.5850.46220.4640.6520.358
Ours22.5140.7840.21327.4110.9130.109
", + "bbox": [ + 91, + 88, + 483, + 207 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ccda7f8f10ff2093f62910894dd50e1b694acae1b1a6c51ac1cca20c1865d2cc.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on the RealEstate10K [86] dataset. Small and Large refer to input images with low and high overlap ratios, respectively. Greater overlap means greater temporal coherence between adjacent images." + ], + "table_footnote": [], + "table_body": "
DatasetTanks-and-TemplesDL3DV
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LucidDreamer [9]14.5520.3640.41515.1260.4520.431
ZeroNVS [45]14.7340.3810.48015.1630.4640.466
MotionCtrl [64]15.3220.4260.40416.8890.5280.392
ViewCrafter [77]21.2860.6540.19321.3730.6860.244
Ours22.3310.7420.21322.7820.7910.202
", + "bbox": [ + 91, + 276, + 483, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of single-view novel view synthesis on DL3DV [32] and Tank-and-Templets [27] benchmarks.", + "bbox": [ + 89, + 373, + 482, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nificantly improves the generalization ability of the model.", + "bbox": [ + 89, + 415, + 478, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Sparse View Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 441, + 333, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark and Metrics. To comprehensively evaluate the performance of our method, we compare it with two categories of reconstruction methods: (1) two-view feedforward reconstruction methods [7, 8, 11, 52, 69, 79], and (2) optimization-based sparse-view reconstruction methods [13, 15, 34, 77]. The evaluation employs two types of test sets: in-domain test sets (from RealEstate10K [86] and Acid [33]) and out-of-domain test sets. To ensure a fair comparison, the out-of-domain test set is the same as that used in previous work [15, 65]. For evaluating scene reconstruction quality, we adopt three metrics: PSNR, SSIM, and LPIPS [81]. Additionally, we categorize the evaluation into two groups based on the overlap between input views, with detailed overlap estimation methods provided in the supplementary materials. The evaluation protocol follows a consistent procedure: first, we reconstruct the 3D scene based on the input images, then render from novel views, and compute the metrics between the rendered images and the reference images.", + "bbox": [ + 88, + 460, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison. As shown in Tables 2-4 and Figures 4-5, our method performs well on various evaluation metrics, particularly in challenging scenarios with occlusions and limited view overlap. This improvement can be attributed to our approach that utilizes scene priors and sparse input views to generate helpful additional observations, thereby enhancing visual correlations between different viewpoints and helping to reduce the complexity of sparse-view reconstruction. Compared to the method [77] conditioning on point-based rendering, our approach shows advantages in reconstructing", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d8a4bd7ccf0d825517c755150babf0b111e8e0f9c59490f351fdc46f275f662.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of novel views rendered from scenes reconstructed using other 3D generation methods." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 906, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scenes with thin structures, leading to improved 3D reconstruction results.", + "bbox": [ + 511, + 373, + 906, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3. Single View 3D Generation", + "text_level": 1, + "bbox": [ + 511, + 411, + 748, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark and Metrics. In single-view generation, we compare our method with several generative methods [45, 46, 64, 77] on the Tank-and-Temple [27] and DL3DV-140 [32] datasets. Noted that we begin by reconstructing the 3D scene from the given images, followed by calculating the metrics using renderings from novel views. Similarly, we used the PSNR, SSIM, and LPIPS [81] metrics to evaluate our results. Evaluation in this underconstrained setting is challenging, as multiple 3D scenes can produce consistent generations for a given view. Thus, we measure metrics using only temporally adjacent frames to the conditional image, 14 sampled frames and poses captured after the conditional image.", + "bbox": [ + 511, + 430, + 906, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison. As shown in Table 5, our method consistently achieves superior performance on image quality metrics compared to the baselines. For a fair comparison with ZeroNVS [45] and LucidDreamer [9], which are limited to square image inputs, we crop the generated novel views before computing the quantitative metrics. Figure 6 shows that scenes reconstructed using our method have less noise and distortion in occluded areas. For more visualizations, see Figure 7.", + "bbox": [ + 511, + 627, + 908, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 772, + 666, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Video Diffusion Model. We evaluate the effectiveness of the design choices. As shown in Table 6, ray embedding achieves significant improvements in camera control accuracy and rendering quality compared to other camera encoding methods by introducing denser position encoding and optimizing geometric correspondences. Specifically, cross-frame consistency shows notable en", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27800", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9ef69e940233b92e69350b942ec21da862ad7e49e00c29b50c8025277ed02c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 102, + 225, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7e6a23f3bf8a5ddb9fd75e05b3cc953f94de2bbcf3e2bf8ae5c92ae9cbe46dd.jpg", + "image_caption": [ + "Generated Videos" + ], + "image_footnote": [], + "bbox": [ + 227, + 103, + 359, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00e15124fa0801ee5d20cab739ba44509f5a609072445804cffe0eb4273e742a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 103, + 493, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ea890071fbf10b26b612d57323cd2adac9d1114ed6881f4653c805ffa6dd168.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 103, + 627, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03227ef2a56faa7695e3af222ff1e808fa9c3458e850f111f14cdca46b032eba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 103, + 761, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3555563403fa9f3d2e3e7dcf76a14d1499ba9989e25cc536f42f90017987af5.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 764, + 102, + 901, + 161 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88f1c0e74a6eda5f0043e8c0533954ebcb32ac197924fca9a493f0a72c6d0017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 162, + 225, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13794b7537442d126366394a030cfb6c75bb4b9f746aee10383254301f4c99ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 162, + 359, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7283d0c6dd546430b9df1cee4f84daeb2a02b3cd12e814b3e3e194a2d90ba94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 162, + 493, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f9881c5634ab7a4b58509264bfcceb87eb1472f2cfce3749b166237d9d28685.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 162, + 627, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29aed6526ba28a213fa59370562645fa62291dfb8e486f7c423f54d3d678613f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 162, + 761, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2349bfa7ab4e79770a429fd08eecd297f32bf1944cf9b2dede94cb860388c8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 162, + 901, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2c2da9932025e41bbe6986fd28e1659d93ca5352b39c1641485d2483ffa6f21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 223, + 225, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45bae437aa03a483ac6ea8d5bf4ce2149f2138448235f57d066e1edf0956da76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 223, + 359, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a79896b89416be8df2129509237e929c1db219653b30c3254a1ecaeac60d729.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 223, + 493, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1ef049494fdf5c36c7c3b8c247527f4d7663b9159911ad967fb08aa6b61b34a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 223, + 627, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39392d2c6448e620d7658946199f4a37a3319f8cd4928e0d198b0c076f164b7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 223, + 761, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fe593d8c4b5641733e60cbe8481e8c17b27387272825c7f55127678a5440f3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 223, + 901, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9aebd835454c7bf2de2b4d77db7ce5befa62e41b6027d3807a5257ddb82ad45.jpg", + "image_caption": [ + "Figure 7. Additional visualizations of generated videos and their corresponding 3DGS representations." + ], + "image_footnote": [], + "bbox": [ + 94, + 285, + 225, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/affc0dcc4ba8d5df3016827ca36cbb33a1b8470ec2b60abd6802e123c0e1c694.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 285, + 359, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/328c9665967acf4c342c619c6378edd9c0c7f703f5454961b753b34140b8eadc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 285, + 493, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d8bdd67b36f6b474f503a377d3a551dd55a3d661737a698f67bbb36fab14e7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 285, + 627, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d178091b024c2ce8fc6257fa3945b83fe6547d105301d9a7b7f16d76bff60326.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 285, + 761, + 344 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91919ad4b8b4cd7140536b8760164d39163f1f3cc2e66f5de4e7171009dc8dfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 284, + 901, + 345 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/de51b3e51a7e805523d2db398733b5a89b5e5a9d5937fa69f0fe7ebcd4a112c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTerr↓Rerr↓COLMAPerr↓FVD↓
Raw Value Emb.1.5922.37613.5%81.854
Quaternion Emb.1.4572.39413.1%80.368
W/o Ray Emb.2.6833.14515.7%112.843
W/o Epipolar1.1252.23512.2%79.547
W/o Scale Align.1.7122.49814.3%81.679
Full model1.0142.2184.2%78.132
", + "bbox": [ + 93, + 396, + 483, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the video diffusion model variants.", + "bbox": [ + 99, + 505, + 473, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7e200a74cd5ea78bd13b664e454b9b2bbde9a8d21b991594f4e48696cd20bfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 395, + 645, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6807d3a5a8cbdf96992d799baf51ec2bae3bf887eba7bc45386137f502c8ce08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 396, + 769, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d2179cf6fec3340966eb0d9edaafa15d78df20be333ba4593970f87325b7178.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 396, + 893, + 462 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd6e15a969446c95e3eecda69a1c14b0263753fd06bbc1d20cd5d98426bd2c0b.jpg", + "image_caption": [ + "Figure 9. Ablation on the feed-forward Reconstruction Module." + ], + "image_footnote": [], + "bbox": [ + 521, + 462, + 645, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82dd46a36944c72d03983280fe7d077e804e660ae707accf0e85e4508ee90615.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 462, + 769, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1cb64f90e35bb41b39e079266e97ccc03b03ea4b50947162a8c3cae6068c3ac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 462, + 893, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed22943e74657066c6571145654cabbe6fb25d234c396acea7ee37dd709f1e22.jpg", + "image_caption": [ + "Figure 8. Ablation on the Scale Alignment. The video reconstruction results obtained based on scale-alignment training exhibit reduced artifacts while improving the 3D consistency of the scene." + ], + "image_footnote": [], + "bbox": [ + 91, + 527, + 493, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hancement. Furthermore, when incorporating geometric constraints through Epipolar Attention, the system achieves optimal performance in terms of camera controllability and 3D consistency.", + "bbox": [ + 89, + 732, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scale Alignment. As illustrated in Table 6 and Figure 8, the incorporation of dataset metric scale alignment significantly improves the pose control accuracy and 3D consistency of the generated video.", + "bbox": [ + 89, + 794, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Latent-based Reconstruction. As demonstrated in Figure 9, incorporating generative priors and depth features in our method significantly enhances the fidelity of fine", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "grained geometric details in scene reconstruction. By introducing priors from video generative models, we reduced the difficulty of sparse-view reconstruction and achieved reasonable completion of unseen regions.", + "bbox": [ + 511, + 563, + 905, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 645, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SpatialCrafter, a framework for scene reconstruction and novel view synthesis from sparse or single-view inputs. By leveraging video diffusion models to generate plausible additional observations, we effectively reduce the complexity of sparse-view scene reconstruction. Our key contributions include: (1) precise camera control via ray embeddings or depth-warped images with a trainable condition encoder; (2) a unified scale estimator solving the scale ambiguity in multi-dataset training; and (3) a hybrid Mamba-Transformer architecture that combines monocular depth priors with semantic features from video latent space to directly regress 3D Gaussian primitives. Experiments show our method outperforms existing methods, especially in single-view extrapolation and scenarios with little overlaps. Future work will focus on extending to dynamic scene.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 6", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 13", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4", + "[5] Chenjie Cao, Chaohui Yu, Shang Liu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvgenmaster: Scaling multi-view generation from any image via 3d priors enhanced diffusion model. arXiv preprint arXiv:2411.16157, 2024. 3", + "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 3", + "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 2, 6, 7, 13", + "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 2, 6, 7, 13", + "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 7", + "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3", + "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 6, 7", + "[12] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 3", + "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv:2403.20309, 2024. 2, 7, 13", + "[14] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3, 6, 7, 14", + "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4", + "[17] Guangcong, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2", + "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3", + "[19] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3, 4, 5, 6", + "[20] Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models. arXiv preprint arXiv:2503.10592, 2025. 3", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5, 13", + "[22] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence pose and nerf for generalized pose-free novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20196-20206, 2024. 2, 7", + "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 2", + "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3", + "[26] Mijeong Kim, Seonguk Seo, and Bohyung Han. Informer: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2", + "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 6, 7, 18" + ], + "bbox": [ + 517, + 93, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27802", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2", + "[29] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 13", + "[30] Lingen Li, Zhaoyang Zhang, Yaowei Li, Jiale Xu, Wenbo Hu, Xiaoyu Li, Weihao Cheng, Jinwei Gu, Tianfan Xue, and Ying Shan. Nvcomposer: Boosting generative novel view synthesis with multiple sparse and unposed images. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 777-787, 2025. 3", + "[31] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 798-810, 2025. 3", + "[32] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In CVPR, 2024. 5, 6, 7, 18", + "[33] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 5, 6, 7", + "[34] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3, 6, 7, 14", + "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2, 3", + "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3", + "[37] Baorui Ma, Huachen Gao, Haoge Deng, Zhengxiong Luo, Tiejun Huang, Lulu Tang, and Xinlong Wang. You see it, you got it: Learning 3d creation on pose-free videos at scale. arXiv preprint arXiv:2412.06699, 2024. 3", + "[38] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In CVPR, 2024. 3", + "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2", + "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3", + "[42] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2", + "[43] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6121-6132, 2025. 3", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[45] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 2, 3, 7", + "[46] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seunglyong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 7", + "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3", + "[48] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3", + "[49] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 4", + "[50] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2", + "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5, 13", + "[52] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 6, 7", + "[53] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3, 14", + "[54] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward gener" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "alisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2", + "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[56] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 3", + "[57] Hanyang Wang, Fangfu Liu, Jiawei Chi, and Yueqi Duan. Videoscene: Distilling video diffusion model to generate 3d scenes in one step. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16475-16485, 2025. 3", + "[58] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 4", + "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 4", + "[60] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 7", + "[61] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[62] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 3", + "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[64] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH Conference, 2024. 2, 3, 5, 6, 7", + "[65] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 2, 3, 6, 7", + "[66] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 2", + "[67] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 6", + "[69] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 6, 7", + "[70] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3", + "[71] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2", + "[72] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2", + "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 4", + "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3", + "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 6, 7", + "[76] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 3", + "[77] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 5, 6, 7, 13", + "[78] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 4", + "[79] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 6, 7" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27804", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2", + "[81] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7", + "[82] Songchun Zhang, Yibo Zhang, Quan Zheng, Rui Ma, Wei Hua, Hujun Bao, Weiwei Xu, and Changqing Zou. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10170-10180, 2024. 3", + "[83] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21936-21947, 2025. 2", + "[84] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2", + "[85] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3", + "[86] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph, 37, 2018. 5, 6, 7", + "[87] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 3", + "[88] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 2, 6" + ], + "bbox": [ + 91, + 92, + 482, + 628 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "27805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_model.json b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf07fcca387a9ef591990cbf07e4870101100308 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_model.json @@ -0,0 +1,2754 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.13, + 0.85, + 0.175 + ], + "angle": 0, + "content": "SpatialCrafter: Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.203, + 0.77, + 0.258 + ], + "angle": 0, + "content": "Songchun Zhang\\(^{1,2}\\) Huiyao Xu\\(^{1}\\) Sitong Guo\\(^{1}\\) Zhongwei Xie\\(^{2}\\) \nHujun Bao\\(^{1}\\) Weiwei Xu\\(^{1}\\) Changqing Zou\\(^{1,3*}\\) \n\\(^{1}\\mathrm{ZJU}\\) \\({}^{2}\\mathrm{HKUST}\\) \\({}^{3}\\mathrm{Zhejiang Lab}\\)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.271, + 0.331, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.271, + 0.906, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.908, + 0.649 + ], + "angle": 0, + "content": "Figure 1. Overview. Our method, SpatialCrafter, generates additional novel views of a scene from few inputs by leveraging camera trajectory-guided video diffusion models, which alleviates the ambiguity of sparse view reconstruction and robustly reconstructs the scene from the generated video frames. It demonstrates impressive performance in both indoor and outdoor scenes, while also exhibiting generalization capabilities in real scenes and synthetic images. More visualization results can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.682, + 0.327, + 0.699 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Novel view synthesis (NVS) boosts immersive experiences in computer vision and graphics. Existing techniques, though progressed, rely on dense multi-view observations, restricting their application. We tackle the task of reconstructing photorealistic 3D scenes from only one or a few input views. We introduce SpatialCrafter, a framework that leverages the rich knowledge in video diffusion models to generate plausible additional observations, thereby alleviating reconstruction ambiguity. Through a trainable camera encoder and an epipolar attention mechanism for explicit geometric constraints, we achieve precise camera control and 3D consis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.909, + 0.837 + ], + "angle": 0, + "content": "tency, further reinforced by a unified scale estimation strategy to handle scale discrepancies across datasets. Furthermore, by integrating monocular depth priors with semantic features in the video latent space, our framework directly regresses 3D Gaussian primitives and efficiently processes long-sequence features using a hybrid network structure. Extensive experiments show our method enhances sparse view reconstruction and restores the realistic appearance of 3D scenes. Project page: https://franklinz233.github.io/projects/spatialcrafter/." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "27794" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.281 + ], + "angle": 0, + "content": "Novel View Synthesis (NVS), as a key technology in computer vision and graphics, provides crucial support for immersive experiences in fields such as video games and mixed reality. Although neural reconstruction techniques have made significant progress in recent years [25, 39, 42, 62, 66], these methods typically rely on dense multiview observation data, which face numerous limitations in practical applications. Therefore, this paper focuses on a more practically valuable challenge: how to achieve high-quality 3D scene reconstruction and synthesize realistic novel views from sparse or even single-view observation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.599 + ], + "angle": 0, + "content": "This problem is hard because (1) real-world scenes are diverse and complex, (2) occluded regions are unseen in sparse captures, and (3) geometric cues are too few to constrain reconstruction. To address these challenges, some methods [17, 26, 29, 40, 84, 88] constrain the optimization process by introducing regularization terms, but these methods need to be optimized for each scene, are not generalizable, and are difficult to cope with complex scenarios. Recently, some feed-forward methods [7, 8, 22, 50, 55, 60, 79, 80] have achieved generalizable sparse view scene reconstruction by directly predicting 3D Gaussian primitives for each pixel. However, such methods produce severe artifacts when reconstructing occluded regions and applying them to the extrapolation setting. To address this challenge, some methods [35, 45, 64, 65] have introduced the priors from the diffusion model to achieve sparse or single-view NVS, but they perform poorly on scene-level data and lack precise pose control and 3D consistency. This is due to the fact that low-dimensional camera information such as Euler angles and extrinsic struggle to provide comprehensive control signals to the generative models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a framework named SpatialCrafter for scene reconstruction and novel view synthesis based on sparse or single-view inputs. Our innovation lies in leveraging the rich physical world knowledge embedded in video diffusion models to provide plausible additional observations for scene reconstruction, thereby effectively reducing the problem complexity. Specifically, we first focus on enhancing precise camera control and 3D consistency in generated videos. To achieve this, we parameterize camera settings using ray embeddings [78] or metric depth-warped images [77] and incorporate them into the video diffusion model via a trainable camera encoder. Furthermore, we propose an epipolar attention mechanism that improves 3D consistency between video frames through explicit geometric constraints. Previous methods [19, 64] often experience performance degradation and difficulty generating large-motion videos when trained on multiple scene datasets. We identified that these issues primarily stem from scale ambiguity in scene datasets. To overcome this, we introduce a unified scale estimator to calibrate scene dataset," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.107 + ], + "angle": 0, + "content": "enabling effective joint training across multiple datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.109, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Although we can generate visually coherent video sequences, relying solely on generated frames to reconstruct general scenes often leads to suboptimal solutions, particularly in large-scale outdoor environments and stylized scenes. To address this, we propose incorporating monocular depth priors with rich semantic features extracted from the video latent space. Leveraging these latent features, our method directly regresses 3D Gaussian primitives of the scene via a feed-forward network. Furthermore, to efficiently handle long-sequence feature interactions, we design a hybrid architecture that integrates Mamba blocks with Transformer blocks. Experiments show that our method not only improves the sparse-view reconstruction, but also accurately restores the appearance of 3D scenes, especially when extrapolating from a single view and when there are few sparse view overlaps. In conclusion, our key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.366, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- We introduce a framework that effectively utilizes the physical-world knowledge embedded in video diffusion models to provide additional plausible observations for sparse-view scene reconstruction, thus reducing the ambiguity of sparse view scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.442, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- To address the scale ambiguity problem that occurs in joint training across datasets, we develop a unified scale estimation approach for trajectory calibration. This solves the performance degradation problem, thus enabling effective multi-dataset training." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.517, + 0.905, + 0.607 + ], + "angle": 0, + "content": "- We combine monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives through a feed-forward manner. Meanwhile, we propose a hybrid architecture integrating Mamba blocks with Transformer blocks to efficiently handle long-sequence feature interactions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.655, + 0.641 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.816, + 0.667 + ], + "angle": 0, + "content": "2.1. Sparse-View Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "NeRF [62] and Gaussian Splatting [25] have achieved photorealistic representations of 3D scenes, but require optimization on densely collected image sets for each scene. To address this problem, some approaches [17, 29, 71, 88] focus on constraining the optimization process of 3D representations via manually designed regularization. Some other methods [7, 8, 23, 54, 58, 69, 72, 83] are trained on large-scale datasets and directly predict the 3D representation in a feed-forward manner. Recently, some approaches [13, 50] based on end-to-end 3D reconstruction models [7, 28, 60], have achieved efficient and high-quality 3D reconstruction. However, these methods cannot handle occluded regions, and reconstruction failures (e.g., under large viewpoint changes) can severely affect the quality of novel view synthesis. More relevant to our work, some" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.907, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.376, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. Our generative reconstruction pipeline consists of two parts: camera-controlled video generation, and then video-based reconstruction. First, we set the exploration path based on the input views. In the video generation module, ray embeddings are used to parameterize the camera trajectories, and a cross-attention mechanism with epipolar constraints is introduced to improve the 3D consistency of the generated video. The video-based scene reconstruction pipeline integrates monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives via a feed-forward manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.465, + 0.483, + 0.51 + ], + "angle": 0, + "content": "works [31, 34, 57] have employed priors from video diffusion models to enhance the performance of sparse-view reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.518, + 0.416, + 0.533 + ], + "angle": 0, + "content": "2.2. Diffusion-based Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Since text-to-image diffusion models [44] contain rich natural image priors, some methods [41, 63] directly optimize 3D representations [25, 62] via score distillation sampling [41]. Several approaches [6, 35, 45, 56] enable zero-shot novel view synthesis by fine-tuning the image and pose-conditioned generative models on large-scale 3D object [10] or scene [56, 87] datasets. However, they still have difficulty synthesizing consistent novel views. To solve this, some approaches [5, 15, 36, 38, 47, 65, 85] generate multi-views simultaneously and model the correlation between multiple views with ground-truth poses. Some recent methods [30, 37] proposed to use large-scale unposed images to train NVS methods. In addition, some methods [9, 12, 14, 48, 76, 82] utilize depth-based warping to synthesize novel views and employ the T2I model to inpaint the warped images. However, the novel views generated by these methods tend to suffer from artifacts and cumulative errors, limiting their use in generalized scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.365, + 0.833 + ], + "angle": 0, + "content": "2.3. Controllable Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "While recent text-to-video diffusion models have achieved remarkable progress, they inherit the controllability limitations of their text-to-image counterparts, often requiring additional conditioning mechanisms to align generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.662 + ], + "angle": 0, + "content": "videos with user intent. Several works [18, 24, 61, 70] have been carried out to introduce a variety of conditions into video generation models. Recently, attention has focused on controlling the camera motion of the video. Some approaches [18, 53] employs motion-specific training for predefined camera movements, though this framework struggles with complex trajectory synthesis. Some methods [1, 19, 20, 64, 67] enable more complex camera control by parameterizing the camera trajectory and injecting it into a pre-trained video diffusion model via a trainable camera encoder. Another line of work [34, 43, 74, 77] advances this direction by enabling pixel-accurate view compositing through a point-based rendering approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.676, + 0.605, + 0.691 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.822 + ], + "angle": 0, + "content": "This section begins by explaining our camera-controlled video generation model (Section 3.1), which employs ray embeddings for precise pose control. Then, we describe how to use epipolar geometry constraints to enhance 3D consistency in the video frames. Next, we introduce our scene reconstruction pipeline (Section 3.2), which aims to perform stable 3D reconstruction based on the generated video latents." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.854, + 0.849 + ], + "angle": 0, + "content": "3.1. Camera-Conditioned Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Scale Alignment. To address this challenge, our approach first employs VGGT [58] to estimate initial camera parameters, represented by the extrinsic matrix \\(\\mathbf{P} = [\\mathbf{R}|\\mathbf{T}]\\), and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27796" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.272 + ], + "angle": 0, + "content": "a corresponding depth map \\( d_v \\) for each video frame. To establish a consistent scale across these datasets, we leverage the Metric3D [73] to infer a canonical metric depth map, denoted \\( d_m \\). We then align our initial depth prediction, \\( d_v \\), to the canonical scale by calculating a scale factor \\( s \\), defined as the ratio of the inter-percentile ranges (IPR) of the two depth maps: \\( s = \\mathrm{IPR}_{0.8,0.2}(d_v) / \\mathrm{IPR}_{0.8,0.2}(d_m) \\). This scale factor is then used to adjust the translation component of the camera extrinsics, resulting in \\( \\mathbf{P}_{\\text{scaled}} = [\\mathbf{R}|s \\cdot \\mathbf{T}] \\). This procedure yields camera extrinsics with a consistent absolute scale, effectively resolving alignment issues when fusing data from different sources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Camera Injection. Inspired by recent methods [19, 49, 78], we chose to use ray embeddings [19] or the depth warping frames [77] to represent the camera information. Specifically, for a ray defined by the origin \\(o \\in \\mathbb{R}^3\\) and the normalized direction \\(d \\in \\mathbb{R}^3\\), we can represent it as \\((o \\times d, d) \\in \\mathbb{R}^6\\). Given the camera parameters, the direction of the ray \\(d'\\) corresponding to the pixel coordinates \\((u, v)\\) can be calculated as \\(d' = \\mathbf{RK}^{-1}(u, v, 1)^{\\top} + \\mathbf{T}\\). For depth warping frames, we leverage the pretrained depth estimation model [59] to map the reference image into a 3D point cloud. Subsequently, we render novel views at specified camera poses by projecting this point cloud via the target camera parameters. Afterwards, we use a trainable conditional encoder to inject the camera information into the video model instead of fine-tuning all model parameters. The training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.51, + 0.482, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}; z _ {0}, t, \\phi (C)\\right) \\| _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.584 + ], + "angle": 0, + "content": "where \\(\\phi (C)\\) represents the camera conditional encodings, \\(z_{0}\\) denotes the latent embeddings of the reference frame, and \\(t\\) indicates the timestamps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Epipolar Feature Aggregation. Camera embedding enhances control but 3D video consistency remains challenging due to dense self-attention allowing unrestricted cross-frame pixel interactions. We address this by incorporating epipolar geometric constraints, and aggregating features along epipolar lines. For a pixel \\(\\pmb{p} = (u,v)\\) in frame \\(i\\), its corresponding epipolar line \\(l_{ik}\\) in frame \\(k\\) is computed as \\(l_{ik}(\\pmb{p}) = \\pmb{F}_{ik} \\cdot \\tilde{\\pmb{p}}\\), where \\(\\tilde{\\pmb{p}} = (u,v,1)^{\\top}\\) are the homogeneous coordinates, and \\(\\pmb{F}_{ik} \\in \\mathbb{R}^{3 \\times 3}\\) is the fundamental matrix. The fundamental matrix is decomposed as \\(\\pmb{F}_{ik} = \\pmb{K}_k^{-\\top} \\pmb{E}_{ik} \\pmb{K}_i^{-1}\\), where \\(\\pmb{K}_i, \\pmb{K}_k \\in \\mathbb{R}^{3 \\times 3}\\) are the camera intrinsics, and \\(\\pmb{E}_{ik}\\) is the essential matrix. We then convert epipolar lines into attention masks by computing per-pixel distances and applying a threshold, restricting attention to geometrically valid regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Sparse-View Setting. To adapt to the sparse view input setting, we formulate the task as a video frame interpolation problem conditioned on the given start and end frames. To maximally preserve the priors from the pre-trained models [4], we inject boundary frame conditions in both latent and semantic space. Specifically, we combine latent fea" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "tures of the first and last frames with their noisy latent representations, and then concatenate the extracted CLIP embeddings from both frames for cross-attentional feature injection. The training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.164, + 0.907, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, z _ {n} \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; z _ {0}, z _ {n}, t, \\phi (C)) \\| _ {2} ^ {2} \\right] \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.202, + 0.906, + 0.262 + ], + "angle": 0, + "content": "where \\( z_{n} \\) and \\( z_{0} \\) are the latent embeddings of the final and initial frames, respectively. During the inference phase, we use VGGT [58] to estimate the extrinsic and intrinsic parameters for the input views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.274, + 0.816, + 0.289 + ], + "angle": 0, + "content": "3.2. Video-based Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Motivation. There are several challenges in reconstructing scenes directly from generated videos. First, the limited number of generated frames makes it difficult to capture complete scene information. In addition, the diverse styles of video frames can pose a challenge to traditional reconstruction techniques. Furthermore, the generated video frames may contain imperfect or low-quality areas, making the reconstruction process unstable. Moreover, In stylized scenes, previous methods often yield poor results, as recovering poses and sparse point clouds from video frames is difficult, leading to many artifacts. To address this, a powerful reconstruction module specifically designed to handle the generated videos is required." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Latent Feature Fusion. Given input video latents \\(\\mathbf{z} \\in \\mathbb{R}^{T \\times H \\times W \\times C}\\) and camera poses encoded as ray embeddings \\(\\mathbf{p} \\in \\mathbb{R}^{T \\times H \\times W \\times 6}\\), we transform them into token sequences through a patchification process. Specifically, we apply spatial patchification to latent features to obtain latent tokens \\(\\mathbf{z}_t \\in \\mathbb{R}^{N \\times d_z}\\), while ray embeddings undergo 3D-patchification to produce pose tokens \\(\\mathbf{p}_t \\in \\mathbb{R}^{N \\times d_p}\\). To incorporate explicit geometric guidance, we leverage monocular depth estimation within our pipeline. For each input RGB video frame, we estimate dense depth maps \\(\\mathbf{D} \\in \\mathbb{R}^{T \\times H \\times W \\times 1}\\), which are processed by a dedicated depth encoder to yield depth tokens \\(\\mathbf{d}_t \\in \\mathbb{R}^{N \\times d_d}\\). The three token sets are channel-wise concatenated to form a unified representation \\(\\mathbf{x} = [\\mathbf{z}_t; \\mathbf{p}_t; \\mathbf{d}_t] \\in \\mathbb{R}^{N \\times (d_z + d_p + d_d)}\\), which is then linearly projected to a lower-dimensional space \\(\\mathbf{x}' \\in \\mathbb{R}^{N \\times d}\\) before being fed into a sequence of mamba and transformer blocks. Mamba [16] achieves the same token sequence processing functionality as Transformer, but reduces computational complexity from \\(\\mathcal{O}(L^2)\\) to \\(\\mathcal{O}(L)\\), making it particularly suitable for dense reconstruction tasks. In our Mamba blocks, we implement bi-directional scanning across the token sequence. First, we compute state parameters \\(\\mathbf{A}, \\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times d}\\) from the input tokens using a linear projection. Then, we execute the State Space Model in both forward (\\(\\mathbf{y}_f\\)) and backward (\\(\\mathbf{y}_b\\)) directions. The final output of each Mamba block is computed as \\(\\mathbf{y} = \\mathbf{y}_f + \\mathbf{y}_b\\) followed by a final linear transformation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27797" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.359, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.089, + 0.618, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.09, + 0.905, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.28, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of generated videos. Compared to other benchmark methods, the videos generated by our method have better foreground object-background consistency, while also being able to generate videos with larger motion amplitude." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.325, + 0.485, + 0.493 + ], + "angle": 0, + "content": "
MethodFVD↓FID↓Rerr↓Terr↓LPIPS↓PSNR↑SSIM↑
RealEstate10K
MotionCtrl [64]22.65230.120.2340.7980.29914.720.404
CameraCtrl [19]21.48188.210.0540.1270.23017.330.516
ViewCrafter [77]20.96204.180.0550.1530.21518.950.503
Ours18.25183.250.0520.1030.20719.210.523
Tanks-and-Temples
MotionCtrl [64]30.25290.380.8381.5050.31514.640.388
CameraCtrl [19]24.41244.820.1180.2940.28515.380.469
ViewCrafter [77]22.50231.350.1260.3070.24716.240.508
Ours20.17192.520.0970.1750.22816.120.501
DL3DV
MotionCtrl [64]25.65249.510.4731.1180.31214.380.386
CameraCtrl [19]22.76233.540.0950.2380.26216.330.489
ViewCrafter [77]20.59211.240.0930.2440.24217.120.521
Ours18.21171.520.0630.1340.22517.020.537
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.54 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison to the camera conditioned video generation method on RealEstate10K [86], DL3DV [32], and Tanks-and-Temple [27] dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.563, + 0.483, + 0.714 + ], + "angle": 0, + "content": "Gaussian Decoding. We design a lightweight decoder module that efficiently transforms output feature tokens into per-pixel Gaussian parameters. The decoding module consists of 3D-DeConv layers, which generates refined Gaussian feature map \\( G \\in \\mathbb{R}^{(T \\times H \\times W) \\times 12} \\). This 12-channel representation precisely encodes the complete set of Gaussian parameters: RGB color (3 channels), scale factors (3 channels), rotation quaternion (4 channels), opacity (1 channel), and ray distance (1 channel). The final output of the model is the merge of 3D Gaussians from all input video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Training Objective. During training, we render images from predicted Gaussians using randomly selected supervision views. Our approach employs a composite loss that integrates three components:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.794, + 0.483, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {m s e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {d e p t h}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{mse}\\) represents pixel-wise mean squared error, \\(\\mathcal{L}_{perc}\\) denotes perceptual loss, and \\(\\mathcal{L}_{depth}\\) enforces depth consistency. This formulation jointly optimizes for photometric accuracy, high-level perceptual fidelity, and geometric coherence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.327, + 0.64, + 0.344 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.704, + 0.372 + ], + "angle": 0, + "content": "4.1. Experiment settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Datasets. To comprehensively capture the underlying distribution of real-world scenarios, we trained our video diffusion model on three diverse datasets: RealEstate-10K [86], ACID [33], and DL3DV-10K [32]. The Re10K dataset from YouTube comprises 67,477 training and 7,289 testing indoor and outdoor camera trajectories. The ACID dataset focuses on natural landscapes, with 11,075 training and 1,972 testing scenes. The DL3DV-10K dataset is a large-scale collection featuring 10,510 scenes captured under controlled, standardized conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Implementation Details. Our video generation model is based on SVD [3], an image-video diffusion model based on UNet. In our experiments, we use a relative camera system in which all camera poses are converted to poses relative to the first frame. The camera in the first frame is located at the world origin. In the first stage, we train the model at a resolution of \\(320 \\times 512\\) for 50,000 iterations with the frame length set to 25. Subsequently, we train on \\(576 \\times 1024\\) for 10,000 iterations to adapt to high resolution. The learning rate is set to \\(1e - 5\\) with a warmup of 1,000 steps, using the Adam optimizer. We chose Lightning as the training framework, using mixed-precision fp16 and DeepSpeed ZeRO-2. We trained the proposed method and its variants on 16 NVIDIA A800 GPUs with a batch size of 32. During inference, we adopt DDIM sampler [51] with classifier-free guidance [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Progressive Training. Controlling a video generation model to generate arbitrary motion trajectories remains a challenging task. To enable robust arbitrary trajectory generation, we employ a three-stage curriculum learning strategy. The model first trains on smooth camera motions with small temporal intervals, then progressively adapts to more complex motion patterns through linear interval scheduling, and finally incorporates random sampling intervals." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.493, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.493, + 0.388 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on Re10K [86]. Compared to baselines, we obtain superior reconstruction from limited overlap, and enhanced geometry reconstruction in non-overlapping regions." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.398, + 0.484, + 0.582 + ], + "angle": 0, + "content": "
Method3-view6-view9-view
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Mip-NeRF360
Zip-NeRF [2]12.770.2710.70513.610.2840.66314.300.3120.633
ReconFusion [65]15.500.3580.58516.930.4010.54418.190.4320.511
CAT3D [15]16.620.3770.51517.720.4250.48218.670.4600.460
ReconX [34]17.160.4350.40719.200.4730.37820.130.4820.356
ViewCrafter [77]14.510.3150.68215.870.3360.66516.450.3480.655
Ours17.320.4390.41819.420.5100.37120.190.5360.345
DTU
Zip-NeRF [2]9.180.6010.3838.840.5890.3709.230.5920.364
FSGS [88]17.340.8180.16921.550.8800.12724.330.9110.106
ReconFusion [65]20.740.8750.12423.620.9040.10524.620.9210.094
CAT3D [15]22.020.8440.12124.280.8990.09525.920.9280.073
ViewCrafter [77]15.630.5220.38315.030.5250.40814.920.4870.452
Ours27.920.8790.10328.820.8950.08229.030.9050.065
LLFF
Zip-NeRF [2]17.230.5740.37320.710.7640.22123.630.8300.166
FSGS [88]20.310.6520.28824.200.8110.17325.320.8560.136
ReconFusion [65]21.340.7240.20324.250.8150.15225.210.8480.134
CAT3D [15]21.580.7310.18124.710.8330.12125.630.8600.107
ViewCrafter [77]17.730.5210.33217.430.5120.34517.330.4880.351
Ours22.040.7410.16525.110.8140.12425.950.8380.103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of sparse view 3D reconstruction on Out-of-Domain Datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.667 + ], + "angle": 0, + "content": "This gradual transition from simple to complex trajectories proves crucial for high-quality video generation with arbitrary trajectories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.231, + 0.694 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.353, + 0.715 + ], + "angle": 0, + "content": "4.2.1. Controllable Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Benchmark and Metrics. We evaluate our video generation model against three baselines [19, 64, 77] on three benchmark datasets: RealEstate10K [86] (500 randomly selected videos with first frame as image condition and subsequent frames at stride 3 for pose guidance), DL3DV-140 [32] (500 video clips at stride 2), and Tanks-and-Temples [27] (100 sequences at stride 4 from 14 scenes with COLMAP-annotated poses) for out-of-domain generalization testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Metrics. Our evaluation framework employs multiple metrics to assess performance: visual quality measured through Fréchet Video Distance (FVD) and Fréchet Inception Dis" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.093, + 0.905, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.334, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on DL3DV and Tank-and-Temples dataset. Our method reconstructs better than baselines, even with limited image overlap and in non-overlapping regions." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.398, + 0.907, + 0.545 + ], + "angle": 0, + "content": "
MethodRealEstate10KACID
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
pixelNeRF [75]20.430.5890.55020.970.5470.533
GPNR [52]24.110.7930.25525.280.7640.332
AttnRend [11]24.780.8200.21326.880.7990.218
MuRF [68]26.100.8580.14328.090.8410.155
pixelSplat [7]25.890.8580.14228.140.8390.150
MVSplat [8]26.390.8690.12828.250.8430.144
GS-LRM [79]28.100.8920.114---
DepthSplat [69]24.230.7900.217---
ReconX [34]28.310.9120.08828.840.8910.101
ViewCrafter [77]24.220.7880.21823.480.6600.299
Ours28.350.8620.12128.930.8990.116
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.549, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of two-view sparse view reconstruction methods on RealEstate10K [86] and ACID [33] dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.764 + ], + "angle": 0, + "content": "tance (FID); camera control precision quantified by rotation error \\((R_{err})\\) and translation error \\((T_{err})\\) computed from camera poses extracted via COLMAP and normalized relative to the first frame; and visual consistency evaluated using PSNR, SSIM, and LPIPS [81] between generated and ground-truth views. To ensure fair comparison across methods with varying output capabilities, we restrict visual similarity assessment to the first 14 frames, as generated content typically diverges progressively from the conditional single-view input as the scene extends." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparison. Table 1 shows that our method excels in qualitative results, outperforming other methods in terms of visual quality, pose control, and 3D consistency. Figure 3 further demonstrates our advantage in 3D consistency and visual quality. This performance improvement is attributed to the introduction of a camera parameterization method and an epipolar attention module, which respectively enhance pose control ability and the 3D consistency of generated videos. In addition, training on a diverse scene dataset sig" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27799" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.208 + ], + "angle": 0, + "content": "
MethodSmallLarge
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
pixelNeRF [75]18.4170.6010.52620.8690.6390.458
AttnNeRF [11]19.1510.6630.36825.8970.8450.229
pixelSplat [7]20.2630.7170.26627.1510.8790.122
MVSplat [8]20.3530.7240.25427.4080.8840.116
DUSt3R [60]14.1010.4320.46816.4270.4530.402
CoPoNeRF [22]17.3930.5850.46220.4640.6520.358
Ours22.5140.7840.21327.4110.9130.109
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.27 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on the RealEstate10K [86] dataset. Small and Large refer to input images with low and high overlap ratios, respectively. Greater overlap means greater temporal coherence between adjacent images." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.277, + 0.485, + 0.37 + ], + "angle": 0, + "content": "
DatasetTanks-and-TemplesDL3DV
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LucidDreamer [9]14.5520.3640.41515.1260.4520.431
ZeroNVS [45]14.7340.3810.48015.1630.4640.466
MotionCtrl [64]15.3220.4260.40416.8890.5280.392
ViewCrafter [77]21.2860.6540.19321.3730.6860.244
Ours22.3310.7420.21322.7820.7910.202
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of single-view novel view synthesis on DL3DV [32] and Tank-and-Templets [27] benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.479, + 0.432 + ], + "angle": 0, + "content": "nificantly improves the generalization ability of the model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.442, + 0.334, + 0.457 + ], + "angle": 0, + "content": "4.2.2. Sparse View Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Benchmark and Metrics. To comprehensively evaluate the performance of our method, we compare it with two categories of reconstruction methods: (1) two-view feedforward reconstruction methods [7, 8, 11, 52, 69, 79], and (2) optimization-based sparse-view reconstruction methods [13, 15, 34, 77]. The evaluation employs two types of test sets: in-domain test sets (from RealEstate10K [86] and Acid [33]) and out-of-domain test sets. To ensure a fair comparison, the out-of-domain test set is the same as that used in previous work [15, 65]. For evaluating scene reconstruction quality, we adopt three metrics: PSNR, SSIM, and LPIPS [81]. Additionally, we categorize the evaluation into two groups based on the overlap between input views, with detailed overlap estimation methods provided in the supplementary materials. The evaluation protocol follows a consistent procedure: first, we reconstruct the 3D scene based on the input images, then render from novel views, and compute the metrics between the rendered images and the reference images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison. As shown in Tables 2-4 and Figures 4-5, our method performs well on various evaluation metrics, particularly in challenging scenarios with occlusions and limited view overlap. This improvement can be attributed to our approach that utilizes scene priors and sparse input views to generate helpful additional observations, thereby enhancing visual correlations between different viewpoints and helping to reduce the complexity of sparse-view reconstruction. Compared to the method [77] conditioning on point-based rendering, our approach shows advantages in reconstructing" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.907, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.333, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of novel views rendered from scenes reconstructed using other 3D generation methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.907, + 0.403 + ], + "angle": 0, + "content": "scenes with thin structures, leading to improved 3D reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.412, + 0.749, + 0.427 + ], + "angle": 0, + "content": "4.2.3. Single View 3D Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Benchmark and Metrics. In single-view generation, we compare our method with several generative methods [45, 46, 64, 77] on the Tank-and-Temple [27] and DL3DV-140 [32] datasets. Noted that we begin by reconstructing the 3D scene from the given images, followed by calculating the metrics using renderings from novel views. Similarly, we used the PSNR, SSIM, and LPIPS [81] metrics to evaluate our results. Evaluation in this underconstrained setting is challenging, as multiple 3D scenes can produce consistent generations for a given view. Thus, we measure metrics using only temporally adjacent frames to the conditional image, 14 sampled frames and poses captured after the conditional image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.764 + ], + "angle": 0, + "content": "Comparison. As shown in Table 5, our method consistently achieves superior performance on image quality metrics compared to the baselines. For a fair comparison with ZeroNVS [45] and LucidDreamer [9], which are limited to square image inputs, we crop the generated novel views before computing the quantitative metrics. Figure 6 shows that scenes reconstructed using our method have less noise and distortion in occluded areas. For more visualizations, see Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.667, + 0.789 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation on Video Diffusion Model. We evaluate the effectiveness of the design choices. As shown in Table 6, ray embedding achieves significant improvements in camera control accuracy and rendering quality compared to other camera encoding methods by introducing denser position encoding and optimizing geometric correspondences. Specifically, cross-frame consistency shows notable en" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27800" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.089, + 0.491, + 0.101 + ], + "angle": 0, + "content": "Generated Videos" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.089, + 0.854, + 0.101 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.103, + 0.227, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.104, + 0.36, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.104, + 0.495, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.104, + 0.628, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.104, + 0.762, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.103, + 0.903, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.164, + 0.227, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.164, + 0.36, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.164, + 0.494, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.164, + 0.628, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.164, + 0.762, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.164, + 0.903, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.224, + 0.227, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.224, + 0.36, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.224, + 0.494, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.224, + 0.628, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.224, + 0.762, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.224, + 0.903, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.286, + 0.227, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.286, + 0.36, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.286, + 0.494, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.286, + 0.628, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.286, + 0.762, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.285, + 0.903, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.365, + 0.802, + 0.38 + ], + "angle": 0, + "content": "Figure 7. Additional visualizations of generated videos and their corresponding 3DGS representations." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.397, + 0.484, + 0.502 + ], + "angle": 0, + "content": "
MethodTerr↓Rerr↓COLMAPerr↓FVD↓
Raw Value Emb.1.5922.37613.5%81.854
Quaternion Emb.1.4572.39413.1%80.368
W/o Ray Emb.2.6833.14515.7%112.843
W/o Epipolar1.1252.23512.2%79.547
W/o Scale Align.1.7122.49814.3%81.679
Full model1.0142.2184.2%78.132
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.506, + 0.474, + 0.52 + ], + "angle": 0, + "content": "Table 6. Ablation study on the video diffusion model variants." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.396, + 0.646, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.397, + 0.77, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.397, + 0.895, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.463, + 0.646, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.463, + 0.77, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.463, + 0.895, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.523, + 0.899, + 0.537 + ], + "angle": 0, + "content": "Figure 9. Ablation on the feed-forward Reconstruction Module." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.528, + 0.494, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Figure 8. Ablation on the Scale Alignment. The video reconstruction results obtained based on scale-alignment training exhibit reduced artifacts while improving the 3D consistency of the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.794 + ], + "angle": 0, + "content": "hancement. Furthermore, when incorporating geometric constraints through Epipolar Attention, the system achieves optimal performance in terms of camera controllability and 3D consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Scale Alignment. As illustrated in Table 6 and Figure 8, the incorporation of dataset metric scale alignment significantly improves the pose control accuracy and 3D consistency of the generated video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video Latent-based Reconstruction. As demonstrated in Figure 9, incorporating generative priors and depth features in our method significantly enhances the fidelity of fine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.906, + 0.624 + ], + "angle": 0, + "content": "grained geometric details in scene reconstruction. By introducing priors from video generative models, we reduced the difficulty of sparse-view reconstruction and achieved reasonable completion of unseen regions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.634, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present SpatialCrafter, a framework for scene reconstruction and novel view synthesis from sparse or single-view inputs. By leveraging video diffusion models to generate plausible additional observations, we effectively reduce the complexity of sparse-view scene reconstruction. Our key contributions include: (1) precise camera control via ray embeddings or depth-warped images with a trainable condition encoder; (2) a unified scale estimator solving the scale ambiguity in multi-dataset training; and (3) a hybrid Mamba-Transformer architecture that combines monocular depth priors with semantic features from video latent space to directly regress 3D Gaussian primitives. Experiments show our method outperforms existing methods, especially in single-view extrapolation and scenarios with little overlaps. Future work will focus on extending to dynamic scene." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27801" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[5] Chenjie Cao, Chaohui Yu, Shang Liu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvgenmaster: Scaling multi-view generation from any image via 3d priors enhanced diffusion model. arXiv preprint arXiv:2411.16157, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 2, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 2, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.827 + ], + "angle": 0, + "content": "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv:2403.20309, 2024. 2, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[14] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[17] Guangcong, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[19] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[20] Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models. arXiv preprint arXiv:2503.10592, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.904, + 0.561 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[22] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence pose and nerf for generalized pose-free novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20196-20206, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[26] Mijeong Kim, Seonguk Seo, and Bohyung Han. Informer: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 6, 7, 18" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[29] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[30] Lingen Li, Zhaoyang Zhang, Yaowei Li, Jiale Xu, Wenbo Hu, Xiaoyu Li, Weihao Cheng, Jinwei Gu, Tianfan Xue, and Ying Shan. Nvcomposer: Boosting generative novel view synthesis with multiple sparse and unposed images. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 777-787, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[31] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 798-810, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In CVPR, 2024. 5, 6, 7, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[33] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[34] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[37] Baorui Ma, Huachen Gao, Haoge Deng, Zhengxiong Luo, Tiejun Huang, Lulu Tang, and Xinlong Wang. You see it, you got it: Learning 3d creation on pose-free videos at scale. arXiv preprint arXiv:2412.06699, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[38] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[42] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[43] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6121-6132, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[45] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[46] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seunglyong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[48] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[49] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[50] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[53] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward gener" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27803" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "alisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[56] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[57] Hanyang Wang, Fangfu Liu, Jiawei Chi, and Yueqi Duan. Videoscene: Distilling video diffusion model to generate 3d scenes in one step. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16475-16485, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[58] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[60] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[61] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[62] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[64] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH Conference, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[65] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[66] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[67] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[68] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[69] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[70] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[71] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[72] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[76] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[77] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 5, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[78] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[79] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[80] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[81] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[82] Songchun Zhang, Yibo Zhang, Quan Zheng, Rui Ma, Wei Hua, Hujun Bao, Weiwei Xu, and Changqing Zou. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10170-10180, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[83] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21936-21947, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[84] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[85] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[86] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph, 37, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[87] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[88] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "27805" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_origin.pdf b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..967a97fa83a8e1453e7de63b86ffb7c5358272f5 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/359a0256-c24c-45c3-9f07-0205db934b25_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:695378ca0a158440c98363924132b233b5fdd5d1ea1afbf5ea4cf4d0f12a6a03 +size 15757510 diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/full.md b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..38d03c56ab67e7ed3cb6765c9db9e259bf8b25e7 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/full.md @@ -0,0 +1,363 @@ +# SpatialCrafter: Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations + +Songchun Zhang $^{1,2}$ Huiyao Xu $^{1}$ Sitong Guo $^{1}$ Zhongwei Xie $^{2}$ +Hujun Bao $^{1}$ Weiwei Xu $^{1}$ Changqing Zou $^{1,3*}$ $^{1}\mathrm{ZJU}$ ${}^{2}\mathrm{HKUST}$ ${}^{3}\mathrm{Zhejiang Lab}$ + +![](images/70f4b821144851b348978c011d78cc1e1ec9d2397c7b94d94aefe37562ef2577.jpg) +Figure 1. Overview. Our method, SpatialCrafter, generates additional novel views of a scene from few inputs by leveraging camera trajectory-guided video diffusion models, which alleviates the ambiguity of sparse view reconstruction and robustly reconstructs the scene from the generated video frames. It demonstrates impressive performance in both indoor and outdoor scenes, while also exhibiting generalization capabilities in real scenes and synthetic images. More visualization results can be found in the supplementary material. + +![](images/9d656c76546a0eef9f4f920cd6ee3fa42d7eb94acd124dc8a5e8c87e507cd749.jpg) + +# Abstract + +Novel view synthesis (NVS) boosts immersive experiences in computer vision and graphics. Existing techniques, though progressed, rely on dense multi-view observations, restricting their application. We tackle the task of reconstructing photorealistic 3D scenes from only one or a few input views. We introduce SpatialCrafter, a framework that leverages the rich knowledge in video diffusion models to generate plausible additional observations, thereby alleviating reconstruction ambiguity. Through a trainable camera encoder and an epipolar attention mechanism for explicit geometric constraints, we achieve precise camera control and 3D consis + +tency, further reinforced by a unified scale estimation strategy to handle scale discrepancies across datasets. Furthermore, by integrating monocular depth priors with semantic features in the video latent space, our framework directly regresses 3D Gaussian primitives and efficiently processes long-sequence features using a hybrid network structure. Extensive experiments show our method enhances sparse view reconstruction and restores the realistic appearance of 3D scenes. Project page: https://franklinz233.github.io/projects/spatialcrafter/. + +# 1. Introduction + +Novel View Synthesis (NVS), as a key technology in computer vision and graphics, provides crucial support for immersive experiences in fields such as video games and mixed reality. Although neural reconstruction techniques have made significant progress in recent years [25, 39, 42, 62, 66], these methods typically rely on dense multiview observation data, which face numerous limitations in practical applications. Therefore, this paper focuses on a more practically valuable challenge: how to achieve high-quality 3D scene reconstruction and synthesize realistic novel views from sparse or even single-view observation. + +This problem is hard because (1) real-world scenes are diverse and complex, (2) occluded regions are unseen in sparse captures, and (3) geometric cues are too few to constrain reconstruction. To address these challenges, some methods [17, 26, 29, 40, 84, 88] constrain the optimization process by introducing regularization terms, but these methods need to be optimized for each scene, are not generalizable, and are difficult to cope with complex scenarios. Recently, some feed-forward methods [7, 8, 22, 50, 55, 60, 79, 80] have achieved generalizable sparse view scene reconstruction by directly predicting 3D Gaussian primitives for each pixel. However, such methods produce severe artifacts when reconstructing occluded regions and applying them to the extrapolation setting. To address this challenge, some methods [35, 45, 64, 65] have introduced the priors from the diffusion model to achieve sparse or single-view NVS, but they perform poorly on scene-level data and lack precise pose control and 3D consistency. This is due to the fact that low-dimensional camera information such as Euler angles and extrinsic struggle to provide comprehensive control signals to the generative models. + +In this paper, we propose a framework named SpatialCrafter for scene reconstruction and novel view synthesis based on sparse or single-view inputs. Our innovation lies in leveraging the rich physical world knowledge embedded in video diffusion models to provide plausible additional observations for scene reconstruction, thereby effectively reducing the problem complexity. Specifically, we first focus on enhancing precise camera control and 3D consistency in generated videos. To achieve this, we parameterize camera settings using ray embeddings [78] or metric depth-warped images [77] and incorporate them into the video diffusion model via a trainable camera encoder. Furthermore, we propose an epipolar attention mechanism that improves 3D consistency between video frames through explicit geometric constraints. Previous methods [19, 64] often experience performance degradation and difficulty generating large-motion videos when trained on multiple scene datasets. We identified that these issues primarily stem from scale ambiguity in scene datasets. To overcome this, we introduce a unified scale estimator to calibrate scene dataset, + +enabling effective joint training across multiple datasets. + +Although we can generate visually coherent video sequences, relying solely on generated frames to reconstruct general scenes often leads to suboptimal solutions, particularly in large-scale outdoor environments and stylized scenes. To address this, we propose incorporating monocular depth priors with rich semantic features extracted from the video latent space. Leveraging these latent features, our method directly regresses 3D Gaussian primitives of the scene via a feed-forward network. Furthermore, to efficiently handle long-sequence feature interactions, we design a hybrid architecture that integrates Mamba blocks with Transformer blocks. Experiments show that our method not only improves the sparse-view reconstruction, but also accurately restores the appearance of 3D scenes, especially when extrapolating from a single view and when there are few sparse view overlaps. In conclusion, our key contributions can be summarized as follows: + +- We introduce a framework that effectively utilizes the physical-world knowledge embedded in video diffusion models to provide additional plausible observations for sparse-view scene reconstruction, thus reducing the ambiguity of sparse view scene reconstruction. + +- To address the scale ambiguity problem that occurs in joint training across datasets, we develop a unified scale estimation approach for trajectory calibration. This solves the performance degradation problem, thus enabling effective multi-dataset training. + +- We combine monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives through a feed-forward manner. Meanwhile, we propose a hybrid architecture integrating Mamba blocks with Transformer blocks to efficiently handle long-sequence feature interactions. + +# 2. Related Work + +# 2.1. Sparse-View Scene Reconstruction + +NeRF [62] and Gaussian Splatting [25] have achieved photorealistic representations of 3D scenes, but require optimization on densely collected image sets for each scene. To address this problem, some approaches [17, 29, 71, 88] focus on constraining the optimization process of 3D representations via manually designed regularization. Some other methods [7, 8, 23, 54, 58, 69, 72, 83] are trained on large-scale datasets and directly predict the 3D representation in a feed-forward manner. Recently, some approaches [13, 50] based on end-to-end 3D reconstruction models [7, 28, 60], have achieved efficient and high-quality 3D reconstruction. However, these methods cannot handle occluded regions, and reconstruction failures (e.g., under large viewpoint changes) can severely affect the quality of novel view synthesis. More relevant to our work, some + +![](images/d9e64276501703ed894cc268c7aa2232e8544bf6a501f4edad082b4305bd1ca4.jpg) +Figure 2. Overview of our pipeline. Our generative reconstruction pipeline consists of two parts: camera-controlled video generation, and then video-based reconstruction. First, we set the exploration path based on the input views. In the video generation module, ray embeddings are used to parameterize the camera trajectories, and a cross-attention mechanism with epipolar constraints is introduced to improve the 3D consistency of the generated video. The video-based scene reconstruction pipeline integrates monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives via a feed-forward manner. + +works [31, 34, 57] have employed priors from video diffusion models to enhance the performance of sparse-view reconstruction. + +# 2.2. Diffusion-based Novel View Synthesis + +Since text-to-image diffusion models [44] contain rich natural image priors, some methods [41, 63] directly optimize 3D representations [25, 62] via score distillation sampling [41]. Several approaches [6, 35, 45, 56] enable zero-shot novel view synthesis by fine-tuning the image and pose-conditioned generative models on large-scale 3D object [10] or scene [56, 87] datasets. However, they still have difficulty synthesizing consistent novel views. To solve this, some approaches [5, 15, 36, 38, 47, 65, 85] generate multi-views simultaneously and model the correlation between multiple views with ground-truth poses. Some recent methods [30, 37] proposed to use large-scale unposed images to train NVS methods. In addition, some methods [9, 12, 14, 48, 76, 82] utilize depth-based warping to synthesize novel views and employ the T2I model to inpaint the warped images. However, the novel views generated by these methods tend to suffer from artifacts and cumulative errors, limiting their use in generalized scenes. + +# 2.3. Controllable Video Generation + +While recent text-to-video diffusion models have achieved remarkable progress, they inherit the controllability limitations of their text-to-image counterparts, often requiring additional conditioning mechanisms to align generated + +videos with user intent. Several works [18, 24, 61, 70] have been carried out to introduce a variety of conditions into video generation models. Recently, attention has focused on controlling the camera motion of the video. Some approaches [18, 53] employs motion-specific training for predefined camera movements, though this framework struggles with complex trajectory synthesis. Some methods [1, 19, 20, 64, 67] enable more complex camera control by parameterizing the camera trajectory and injecting it into a pre-trained video diffusion model via a trainable camera encoder. Another line of work [34, 43, 74, 77] advances this direction by enabling pixel-accurate view compositing through a point-based rendering approach. + +# 3. Method + +This section begins by explaining our camera-controlled video generation model (Section 3.1), which employs ray embeddings for precise pose control. Then, we describe how to use epipolar geometry constraints to enhance 3D consistency in the video frames. Next, we introduce our scene reconstruction pipeline (Section 3.2), which aims to perform stable 3D reconstruction based on the generated video latents. + +# 3.1. Camera-Conditioned Video Generation + +Scale Alignment. To address this challenge, our approach first employs VGGT [58] to estimate initial camera parameters, represented by the extrinsic matrix $\mathbf{P} = [\mathbf{R}|\mathbf{T}]$ , and + +a corresponding depth map $d_v$ for each video frame. To establish a consistent scale across these datasets, we leverage the Metric3D [73] to infer a canonical metric depth map, denoted $d_m$ . We then align our initial depth prediction, $d_v$ , to the canonical scale by calculating a scale factor $s$ , defined as the ratio of the inter-percentile ranges (IPR) of the two depth maps: $s = \mathrm{IPR}_{0.8,0.2}(d_v) / \mathrm{IPR}_{0.8,0.2}(d_m)$ . This scale factor is then used to adjust the translation component of the camera extrinsics, resulting in $\mathbf{P}_{\text{scaled}} = [\mathbf{R}|s \cdot \mathbf{T}]$ . This procedure yields camera extrinsics with a consistent absolute scale, effectively resolving alignment issues when fusing data from different sources. + +Camera Injection. Inspired by recent methods [19, 49, 78], we chose to use ray embeddings [19] or the depth warping frames [77] to represent the camera information. Specifically, for a ray defined by the origin $o \in \mathbb{R}^3$ and the normalized direction $d \in \mathbb{R}^3$ , we can represent it as $(o \times d, d) \in \mathbb{R}^6$ . Given the camera parameters, the direction of the ray $d'$ corresponding to the pixel coordinates $(u, v)$ can be calculated as $d' = \mathbf{RK}^{-1}(u, v, 1)^{\top} + \mathbf{T}$ . For depth warping frames, we leverage the pretrained depth estimation model [59] to map the reference image into a 3D point cloud. Subsequently, we render novel views at specified camera poses by projecting this point cloud via the target camera parameters. Afterwards, we use a trainable conditional encoder to inject the camera information into the video model instead of fine-tuning all model parameters. The training objective is: + +$$ +\mathcal {L} = \mathbb {E} _ {z, z _ {0}, \epsilon , C, t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}; z _ {0}, t, \phi (C)\right) \| _ {2} ^ {2} \right] \tag {1} +$$ + +where $\phi (C)$ represents the camera conditional encodings, $z_{0}$ denotes the latent embeddings of the reference frame, and $t$ indicates the timestamps. + +Epipolar Feature Aggregation. Camera embedding enhances control but 3D video consistency remains challenging due to dense self-attention allowing unrestricted cross-frame pixel interactions. We address this by incorporating epipolar geometric constraints, and aggregating features along epipolar lines. For a pixel $\pmb{p} = (u,v)$ in frame $i$ , its corresponding epipolar line $l_{ik}$ in frame $k$ is computed as $l_{ik}(\pmb{p}) = \pmb{F}_{ik} \cdot \tilde{\pmb{p}}$ , where $\tilde{\pmb{p}} = (u,v,1)^{\top}$ are the homogeneous coordinates, and $\pmb{F}_{ik} \in \mathbb{R}^{3 \times 3}$ is the fundamental matrix. The fundamental matrix is decomposed as $\pmb{F}_{ik} = \pmb{K}_k^{-\top} \pmb{E}_{ik} \pmb{K}_i^{-1}$ , where $\pmb{K}_i, \pmb{K}_k \in \mathbb{R}^{3 \times 3}$ are the camera intrinsics, and $\pmb{E}_{ik}$ is the essential matrix. We then convert epipolar lines into attention masks by computing per-pixel distances and applying a threshold, restricting attention to geometrically valid regions. + +Sparse-View Setting. To adapt to the sparse view input setting, we formulate the task as a video frame interpolation problem conditioned on the given start and end frames. To maximally preserve the priors from the pre-trained models [4], we inject boundary frame conditions in both latent and semantic space. Specifically, we combine latent fea + +tures of the first and last frames with their noisy latent representations, and then concatenate the extracted CLIP embeddings from both frames for cross-attentional feature injection. The training objective is: + +$$ +\mathcal {L} = \mathbb {E} _ {z, z _ {0}, z _ {n} \epsilon , C, t} \left[ \| \epsilon - \epsilon_ {\theta} (z _ {t}; z _ {0}, z _ {n}, t, \phi (C)) \| _ {2} ^ {2} \right] \quad (2) +$$ + +where $z_{n}$ and $z_{0}$ are the latent embeddings of the final and initial frames, respectively. During the inference phase, we use VGGT [58] to estimate the extrinsic and intrinsic parameters for the input views. + +# 3.2. Video-based Scene Reconstruction + +Motivation. There are several challenges in reconstructing scenes directly from generated videos. First, the limited number of generated frames makes it difficult to capture complete scene information. In addition, the diverse styles of video frames can pose a challenge to traditional reconstruction techniques. Furthermore, the generated video frames may contain imperfect or low-quality areas, making the reconstruction process unstable. Moreover, In stylized scenes, previous methods often yield poor results, as recovering poses and sparse point clouds from video frames is difficult, leading to many artifacts. To address this, a powerful reconstruction module specifically designed to handle the generated videos is required. + +Latent Feature Fusion. Given input video latents $\mathbf{z} \in \mathbb{R}^{T \times H \times W \times C}$ and camera poses encoded as ray embeddings $\mathbf{p} \in \mathbb{R}^{T \times H \times W \times 6}$ , we transform them into token sequences through a patchification process. Specifically, we apply spatial patchification to latent features to obtain latent tokens $\mathbf{z}_t \in \mathbb{R}^{N \times d_z}$ , while ray embeddings undergo 3D-patchification to produce pose tokens $\mathbf{p}_t \in \mathbb{R}^{N \times d_p}$ . To incorporate explicit geometric guidance, we leverage monocular depth estimation within our pipeline. For each input RGB video frame, we estimate dense depth maps $\mathbf{D} \in \mathbb{R}^{T \times H \times W \times 1}$ , which are processed by a dedicated depth encoder to yield depth tokens $\mathbf{d}_t \in \mathbb{R}^{N \times d_d}$ . The three token sets are channel-wise concatenated to form a unified representation $\mathbf{x} = [\mathbf{z}_t; \mathbf{p}_t; \mathbf{d}_t] \in \mathbb{R}^{N \times (d_z + d_p + d_d)}$ , which is then linearly projected to a lower-dimensional space $\mathbf{x}' \in \mathbb{R}^{N \times d}$ before being fed into a sequence of mamba and transformer blocks. Mamba [16] achieves the same token sequence processing functionality as Transformer, but reduces computational complexity from $\mathcal{O}(L^2)$ to $\mathcal{O}(L)$ , making it particularly suitable for dense reconstruction tasks. In our Mamba blocks, we implement bi-directional scanning across the token sequence. First, we compute state parameters $\mathbf{A}, \mathbf{B}, \mathbf{C} \in \mathbb{R}^{N \times d}$ from the input tokens using a linear projection. Then, we execute the State Space Model in both forward ( $\mathbf{y}_f$ ) and backward ( $\mathbf{y}_b$ ) directions. The final output of each Mamba block is computed as $\mathbf{y} = \mathbf{y}_f + \mathbf{y}_b$ followed by a final linear transformation. + +![](images/e753cdbf55dd4c8f8b1103458c77bd44f1c8332731ebfb8f624ac09d27ad0a35.jpg) +Figure 3. Qualitative comparison of generated videos. Compared to other benchmark methods, the videos generated by our method have better foreground object-background consistency, while also being able to generate videos with larger motion amplitude. + +![](images/0bc77672b096c01e27dc68ae50d8f7993d0a47a9d6745badef542dc876c60e4d.jpg) + +![](images/a81772d829b6c6620a3627e29da7927d39bdbef2a99355459a4ab12757be764e.jpg) + +
MethodFVD↓FID↓Rerr↓Terr↓LPIPS↓PSNR↑SSIM↑
RealEstate10K
MotionCtrl [64]22.65230.120.2340.7980.29914.720.404
CameraCtrl [19]21.48188.210.0540.1270.23017.330.516
ViewCrafter [77]20.96204.180.0550.1530.21518.950.503
Ours18.25183.250.0520.1030.20719.210.523
Tanks-and-Temples
MotionCtrl [64]30.25290.380.8381.5050.31514.640.388
CameraCtrl [19]24.41244.820.1180.2940.28515.380.469
ViewCrafter [77]22.50231.350.1260.3070.24716.240.508
Ours20.17192.520.0970.1750.22816.120.501
DL3DV
MotionCtrl [64]25.65249.510.4731.1180.31214.380.386
CameraCtrl [19]22.76233.540.0950.2380.26216.330.489
ViewCrafter [77]20.59211.240.0930.2440.24217.120.521
Ours18.21171.520.0630.1340.22517.020.537
+ +Table 1. Quantitative comparison to the camera conditioned video generation method on RealEstate10K [86], DL3DV [32], and Tanks-and-Temple [27] dataset. + +Gaussian Decoding. We design a lightweight decoder module that efficiently transforms output feature tokens into per-pixel Gaussian parameters. The decoding module consists of 3D-DeConv layers, which generates refined Gaussian feature map $G \in \mathbb{R}^{(T \times H \times W) \times 12}$ . This 12-channel representation precisely encodes the complete set of Gaussian parameters: RGB color (3 channels), scale factors (3 channels), rotation quaternion (4 channels), opacity (1 channel), and ray distance (1 channel). The final output of the model is the merge of 3D Gaussians from all input video frames. + +Training Objective. During training, we render images from predicted Gaussians using randomly selected supervision views. Our approach employs a composite loss that integrates three components: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \lambda_ {1} \mathcal {L} _ {\text {m s e}} + \lambda_ {2} \mathcal {L} _ {\text {p e r c}} + \lambda_ {3} \mathcal {L} _ {\text {d e p t h}}, \tag {3} +$$ + +where $\mathcal{L}_{mse}$ represents pixel-wise mean squared error, $\mathcal{L}_{perc}$ denotes perceptual loss, and $\mathcal{L}_{depth}$ enforces depth consistency. This formulation jointly optimizes for photometric accuracy, high-level perceptual fidelity, and geometric coherence. + +# 4. Experiment + +# 4.1. Experiment settings + +Datasets. To comprehensively capture the underlying distribution of real-world scenarios, we trained our video diffusion model on three diverse datasets: RealEstate-10K [86], ACID [33], and DL3DV-10K [32]. The Re10K dataset from YouTube comprises 67,477 training and 7,289 testing indoor and outdoor camera trajectories. The ACID dataset focuses on natural landscapes, with 11,075 training and 1,972 testing scenes. The DL3DV-10K dataset is a large-scale collection featuring 10,510 scenes captured under controlled, standardized conditions. + +Implementation Details. Our video generation model is based on SVD [3], an image-video diffusion model based on UNet. In our experiments, we use a relative camera system in which all camera poses are converted to poses relative to the first frame. The camera in the first frame is located at the world origin. In the first stage, we train the model at a resolution of $320 \times 512$ for 50,000 iterations with the frame length set to 25. Subsequently, we train on $576 \times 1024$ for 10,000 iterations to adapt to high resolution. The learning rate is set to $1e - 5$ with a warmup of 1,000 steps, using the Adam optimizer. We chose Lightning as the training framework, using mixed-precision fp16 and DeepSpeed ZeRO-2. We trained the proposed method and its variants on 16 NVIDIA A800 GPUs with a batch size of 32. During inference, we adopt DDIM sampler [51] with classifier-free guidance [21]. + +Progressive Training. Controlling a video generation model to generate arbitrary motion trajectories remains a challenging task. To enable robust arbitrary trajectory generation, we employ a three-stage curriculum learning strategy. The model first trains on smooth camera motions with small temporal intervals, then progressively adapts to more complex motion patterns through linear interval scheduling, and finally incorporates random sampling intervals. + +![](images/d022385becdeb84df97778ce77a32a18b926b872fb16b84edc4be7c98dc9c977.jpg) +Figure 5. Qualitative comparison on DL3DV and Tank-and-Temples dataset. Our method reconstructs better than baselines, even with limited image overlap and in non-overlapping regions. + +
Method3-view6-view9-view
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Mip-NeRF360
Zip-NeRF [2]12.770.2710.70513.610.2840.66314.300.3120.633
ReconFusion [65]15.500.3580.58516.930.4010.54418.190.4320.511
CAT3D [15]16.620.3770.51517.720.4250.48218.670.4600.460
ReconX [34]17.160.4350.40719.200.4730.37820.130.4820.356
ViewCrafter [77]14.510.3150.68215.870.3360.66516.450.3480.655
Ours17.320.4390.41819.420.5100.37120.190.5360.345
DTU
Zip-NeRF [2]9.180.6010.3838.840.5890.3709.230.5920.364
FSGS [88]17.340.8180.16921.550.8800.12724.330.9110.106
ReconFusion [65]20.740.8750.12423.620.9040.10524.620.9210.094
CAT3D [15]22.020.8440.12124.280.8990.09525.920.9280.073
ViewCrafter [77]15.630.5220.38315.030.5250.40814.920.4870.452
Ours27.920.8790.10328.820.8950.08229.030.9050.065
LLFF
Zip-NeRF [2]17.230.5740.37320.710.7640.22123.630.8300.166
FSGS [88]20.310.6520.28824.200.8110.17325.320.8560.136
ReconFusion [65]21.340.7240.20324.250.8150.15225.210.8480.134
CAT3D [15]21.580.7310.18124.710.8330.12125.630.8600.107
ViewCrafter [77]17.730.5210.33217.430.5120.34517.330.4880.351
Ours22.040.7410.16525.110.8140.12425.950.8380.103
+ +This gradual transition from simple to complex trajectories proves crucial for high-quality video generation with arbitrary trajectories. + +# 4.2. Comparisons + +# 4.2.1. Controllable Video Generation + +Benchmark and Metrics. We evaluate our video generation model against three baselines [19, 64, 77] on three benchmark datasets: RealEstate10K [86] (500 randomly selected videos with first frame as image condition and subsequent frames at stride 3 for pose guidance), DL3DV-140 [32] (500 video clips at stride 2), and Tanks-and-Temples [27] (100 sequences at stride 4 from 14 scenes with COLMAP-annotated poses) for out-of-domain generalization testing. + +Metrics. Our evaluation framework employs multiple metrics to assess performance: visual quality measured through Fréchet Video Distance (FVD) and Fréchet Inception Dis + +![](images/d557e1f8a2c4c4460246b66655ce13750d2a53e1f471757a24ae090f0b468848.jpg) +Figure 4. Qualitative comparison on Re10K [86]. Compared to baselines, we obtain superior reconstruction from limited overlap, and enhanced geometry reconstruction in non-overlapping regions. + +Table 2. Quantitative comparison of sparse view 3D reconstruction on Out-of-Domain Datasets. + +
MethodRealEstate10KACID
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
pixelNeRF [75]20.430.5890.55020.970.5470.533
GPNR [52]24.110.7930.25525.280.7640.332
AttnRend [11]24.780.8200.21326.880.7990.218
MuRF [68]26.100.8580.14328.090.8410.155
pixelSplat [7]25.890.8580.14228.140.8390.150
MVSplat [8]26.390.8690.12828.250.8430.144
GS-LRM [79]28.100.8920.114---
DepthSplat [69]24.230.7900.217---
ReconX [34]28.310.9120.08828.840.8910.101
ViewCrafter [77]24.220.7880.21823.480.6600.299
Ours28.350.8620.12128.930.8990.116
+ +Table 3. Quantitative comparison of two-view sparse view reconstruction methods on RealEstate10K [86] and ACID [33] dataset. + +tance (FID); camera control precision quantified by rotation error $(R_{err})$ and translation error $(T_{err})$ computed from camera poses extracted via COLMAP and normalized relative to the first frame; and visual consistency evaluated using PSNR, SSIM, and LPIPS [81] between generated and ground-truth views. To ensure fair comparison across methods with varying output capabilities, we restrict visual similarity assessment to the first 14 frames, as generated content typically diverges progressively from the conditional single-view input as the scene extends. + +Comparison. Table 1 shows that our method excels in qualitative results, outperforming other methods in terms of visual quality, pose control, and 3D consistency. Figure 3 further demonstrates our advantage in 3D consistency and visual quality. This performance improvement is attributed to the introduction of a camera parameterization method and an epipolar attention module, which respectively enhance pose control ability and the 3D consistency of generated videos. In addition, training on a diverse scene dataset sig + +
MethodSmallLarge
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
pixelNeRF [75]18.4170.6010.52620.8690.6390.458
AttnNeRF [11]19.1510.6630.36825.8970.8450.229
pixelSplat [7]20.2630.7170.26627.1510.8790.122
MVSplat [8]20.3530.7240.25427.4080.8840.116
DUSt3R [60]14.1010.4320.46816.4270.4530.402
CoPoNeRF [22]17.3930.5850.46220.4640.6520.358
Ours22.5140.7840.21327.4110.9130.109
+ +Table 4. Quantitative comparison on the RealEstate10K [86] dataset. Small and Large refer to input images with low and high overlap ratios, respectively. Greater overlap means greater temporal coherence between adjacent images. + +
DatasetTanks-and-TemplesDL3DV
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LucidDreamer [9]14.5520.3640.41515.1260.4520.431
ZeroNVS [45]14.7340.3810.48015.1630.4640.466
MotionCtrl [64]15.3220.4260.40416.8890.5280.392
ViewCrafter [77]21.2860.6540.19321.3730.6860.244
Ours22.3310.7420.21322.7820.7910.202
+ +Table 5. Quantitative comparison of single-view novel view synthesis on DL3DV [32] and Tank-and-Templets [27] benchmarks. + +nificantly improves the generalization ability of the model. + +# 4.2.2. Sparse View Reconstruction + +Benchmark and Metrics. To comprehensively evaluate the performance of our method, we compare it with two categories of reconstruction methods: (1) two-view feedforward reconstruction methods [7, 8, 11, 52, 69, 79], and (2) optimization-based sparse-view reconstruction methods [13, 15, 34, 77]. The evaluation employs two types of test sets: in-domain test sets (from RealEstate10K [86] and Acid [33]) and out-of-domain test sets. To ensure a fair comparison, the out-of-domain test set is the same as that used in previous work [15, 65]. For evaluating scene reconstruction quality, we adopt three metrics: PSNR, SSIM, and LPIPS [81]. Additionally, we categorize the evaluation into two groups based on the overlap between input views, with detailed overlap estimation methods provided in the supplementary materials. The evaluation protocol follows a consistent procedure: first, we reconstruct the 3D scene based on the input images, then render from novel views, and compute the metrics between the rendered images and the reference images. + +Comparison. As shown in Tables 2-4 and Figures 4-5, our method performs well on various evaluation metrics, particularly in challenging scenarios with occlusions and limited view overlap. This improvement can be attributed to our approach that utilizes scene priors and sparse input views to generate helpful additional observations, thereby enhancing visual correlations between different viewpoints and helping to reduce the complexity of sparse-view reconstruction. Compared to the method [77] conditioning on point-based rendering, our approach shows advantages in reconstructing + +![](images/0d8a4bd7ccf0d825517c755150babf0b111e8e0f9c59490f351fdc46f275f662.jpg) +Figure 6. Qualitative comparisons of novel views rendered from scenes reconstructed using other 3D generation methods. + +scenes with thin structures, leading to improved 3D reconstruction results. + +# 4.2.3. Single View 3D Generation + +Benchmark and Metrics. In single-view generation, we compare our method with several generative methods [45, 46, 64, 77] on the Tank-and-Temple [27] and DL3DV-140 [32] datasets. Noted that we begin by reconstructing the 3D scene from the given images, followed by calculating the metrics using renderings from novel views. Similarly, we used the PSNR, SSIM, and LPIPS [81] metrics to evaluate our results. Evaluation in this underconstrained setting is challenging, as multiple 3D scenes can produce consistent generations for a given view. Thus, we measure metrics using only temporally adjacent frames to the conditional image, 14 sampled frames and poses captured after the conditional image. + +Comparison. As shown in Table 5, our method consistently achieves superior performance on image quality metrics compared to the baselines. For a fair comparison with ZeroNVS [45] and LucidDreamer [9], which are limited to square image inputs, we crop the generated novel views before computing the quantitative metrics. Figure 6 shows that scenes reconstructed using our method have less noise and distortion in occluded areas. For more visualizations, see Figure 7. + +# 4.3. Ablation Study + +Ablation on Video Diffusion Model. We evaluate the effectiveness of the design choices. As shown in Table 6, ray embedding achieves significant improvements in camera control accuracy and rendering quality compared to other camera encoding methods by introducing denser position encoding and optimizing geometric correspondences. Specifically, cross-frame consistency shows notable en + +![](images/b9ef69e940233b92e69350b942ec21da862ad7e49e00c29b50c8025277ed02c5.jpg) + +![](images/d7e6a23f3bf8a5ddb9fd75e05b3cc953f94de2bbcf3e2bf8ae5c92ae9cbe46dd.jpg) +Generated Videos + +![](images/00e15124fa0801ee5d20cab739ba44509f5a609072445804cffe0eb4273e742a.jpg) + +![](images/9ea890071fbf10b26b612d57323cd2adac9d1114ed6881f4653c805ffa6dd168.jpg) + +![](images/03227ef2a56faa7695e3af222ff1e808fa9c3458e850f111f14cdca46b032eba.jpg) + +![](images/e3555563403fa9f3d2e3e7dcf76a14d1499ba9989e25cc536f42f90017987af5.jpg) +3DGS + +![](images/88f1c0e74a6eda5f0043e8c0533954ebcb32ac197924fca9a493f0a72c6d0017.jpg) + +![](images/13794b7537442d126366394a030cfb6c75bb4b9f746aee10383254301f4c99ca.jpg) + +![](images/b7283d0c6dd546430b9df1cee4f84daeb2a02b3cd12e814b3e3e194a2d90ba94.jpg) + +![](images/6f9881c5634ab7a4b58509264bfcceb87eb1472f2cfce3749b166237d9d28685.jpg) + +![](images/29aed6526ba28a213fa59370562645fa62291dfb8e486f7c423f54d3d678613f.jpg) + +![](images/f2349bfa7ab4e79770a429fd08eecd297f32bf1944cf9b2dede94cb860388c8f.jpg) + +![](images/e2c2da9932025e41bbe6986fd28e1659d93ca5352b39c1641485d2483ffa6f21.jpg) + +![](images/45bae437aa03a483ac6ea8d5bf4ce2149f2138448235f57d066e1edf0956da76.jpg) + +![](images/8a79896b89416be8df2129509237e929c1db219653b30c3254a1ecaeac60d729.jpg) + +![](images/a1ef049494fdf5c36c7c3b8c247527f4d7663b9159911ad967fb08aa6b61b34a.jpg) + +![](images/39392d2c6448e620d7658946199f4a37a3319f8cd4928e0d198b0c076f164b7b.jpg) + +![](images/6fe593d8c4b5641733e60cbe8481e8c17b27387272825c7f55127678a5440f3a.jpg) + +![](images/f9aebd835454c7bf2de2b4d77db7ce5befa62e41b6027d3807a5257ddb82ad45.jpg) +Figure 7. Additional visualizations of generated videos and their corresponding 3DGS representations. + +![](images/affc0dcc4ba8d5df3016827ca36cbb33a1b8470ec2b60abd6802e123c0e1c694.jpg) + +![](images/328c9665967acf4c342c619c6378edd9c0c7f703f5454961b753b34140b8eadc.jpg) + +![](images/5d8bdd67b36f6b474f503a377d3a551dd55a3d661737a698f67bbb36fab14e7d.jpg) + +![](images/d178091b024c2ce8fc6257fa3945b83fe6547d105301d9a7b7f16d76bff60326.jpg) + +![](images/91919ad4b8b4cd7140536b8760164d39163f1f3cc2e66f5de4e7171009dc8dfc.jpg) + +
MethodTerr↓Rerr↓COLMAPerr↓FVD↓
Raw Value Emb.1.5922.37613.5%81.854
Quaternion Emb.1.4572.39413.1%80.368
W/o Ray Emb.2.6833.14515.7%112.843
W/o Epipolar1.1252.23512.2%79.547
W/o Scale Align.1.7122.49814.3%81.679
Full model1.0142.2184.2%78.132
+ +Table 6. Ablation study on the video diffusion model variants. + +![](images/f7e200a74cd5ea78bd13b664e454b9b2bbde9a8d21b991594f4e48696cd20bfd.jpg) + +![](images/6807d3a5a8cbdf96992d799baf51ec2bae3bf887eba7bc45386137f502c8ce08.jpg) + +![](images/9d2179cf6fec3340966eb0d9edaafa15d78df20be333ba4593970f87325b7178.jpg) + +![](images/bd6e15a969446c95e3eecda69a1c14b0263753fd06bbc1d20cd5d98426bd2c0b.jpg) +Figure 9. Ablation on the feed-forward Reconstruction Module. + +![](images/82dd46a36944c72d03983280fe7d077e804e660ae707accf0e85e4508ee90615.jpg) + +![](images/1cb64f90e35bb41b39e079266e97ccc03b03ea4b50947162a8c3cae6068c3ac4.jpg) + +![](images/ed22943e74657066c6571145654cabbe6fb25d234c396acea7ee37dd709f1e22.jpg) +Figure 8. Ablation on the Scale Alignment. The video reconstruction results obtained based on scale-alignment training exhibit reduced artifacts while improving the 3D consistency of the scene. + +hancement. Furthermore, when incorporating geometric constraints through Epipolar Attention, the system achieves optimal performance in terms of camera controllability and 3D consistency. + +Scale Alignment. As illustrated in Table 6 and Figure 8, the incorporation of dataset metric scale alignment significantly improves the pose control accuracy and 3D consistency of the generated video. + +Video Latent-based Reconstruction. As demonstrated in Figure 9, incorporating generative priors and depth features in our method significantly enhances the fidelity of fine + +grained geometric details in scene reconstruction. By introducing priors from video generative models, we reduced the difficulty of sparse-view reconstruction and achieved reasonable completion of unseen regions. + +# 5. Conclusion + +We present SpatialCrafter, a framework for scene reconstruction and novel view synthesis from sparse or single-view inputs. By leveraging video diffusion models to generate plausible additional observations, we effectively reduce the complexity of sparse-view scene reconstruction. Our key contributions include: (1) precise camera control via ray embeddings or depth-warped images with a trainable condition encoder; (2) a unified scale estimator solving the scale ambiguity in multi-dataset training; and (3) a hybrid Mamba-Transformer architecture that combines monocular depth priors with semantic features from video latent space to directly regress 3D Gaussian primitives. Experiments show our method outperforms existing methods, especially in single-view extrapolation and scenarios with little overlaps. Future work will focus on extending to dynamic scene. + +# References + +[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 6 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 13 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4 +[5] Chenjie Cao, Chaohui Yu, Shang Liu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvgenmaster: Scaling multi-view generation from any image via 3d priors enhanced diffusion model. arXiv preprint arXiv:2411.16157, 2024. 3 +[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 3 +[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 2, 6, 7, 13 +[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 2, 6, 7, 13 +[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 7 +[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3 +[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 6, 7 +[12] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 3 +[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, + +Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv:2403.20309, 2024. 2, 7, 13 +[14] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3, 6, 7, 14 +[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4 +[17] Guangcong, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2 +[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3 +[19] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3, 4, 5, 6 +[20] Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models. arXiv preprint arXiv:2503.10592, 2025. 3 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5, 13 +[22] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence pose and nerf for generalized pose-free novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20196-20206, 2024. 2, 7 +[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 2 +[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3 +[26] Mijeong Kim, Seonguk Seo, and Bohyung Han. Informer: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2 +[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 6, 7, 18 + +[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2 +[29] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 13 +[30] Lingen Li, Zhaoyang Zhang, Yaowei Li, Jiale Xu, Wenbo Hu, Xiaoyu Li, Weihao Cheng, Jinwei Gu, Tianfan Xue, and Ying Shan. Nvcomposer: Boosting generative novel view synthesis with multiple sparse and unposed images. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 777-787, 2025. 3 +[31] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 798-810, 2025. 3 +[32] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In CVPR, 2024. 5, 6, 7, 18 +[33] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 5, 6, 7 +[34] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3, 6, 7, 14 +[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2, 3 +[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[37] Baorui Ma, Huachen Gao, Haoge Deng, Zhengxiong Luo, Tiejun Huang, Lulu Tang, and Xinlong Wang. You see it, you got it: Learning 3d creation on pose-free videos at scale. arXiv preprint arXiv:2412.06699, 2024. 3 +[38] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In CVPR, 2024. 3 +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg + +nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2 +[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3 +[42] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2 +[43] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6121-6132, 2025. 3 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[45] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 2, 3, 7 +[46] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seunglyong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 7 +[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3 +[48] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3 +[49] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 4 +[50] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2 +[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5, 13 +[52] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 6, 7 +[53] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3, 14 +[54] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward gener + +alisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2 +[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[56] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 3 +[57] Hanyang Wang, Fangfu Liu, Jiawei Chi, and Yueqi Duan. Videoscene: Distilling video diffusion model to generate 3d scenes in one step. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16475-16485, 2025. 3 +[58] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 4 +[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 4 +[60] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 7 +[61] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3 +[62] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 3 +[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[64] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH Conference, 2024. 2, 3, 5, 6, 7 +[65] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 2, 3, 6, 7 +[66] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 2 +[67] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 3 + +[68] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 6 +[69] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 6, 7 +[70] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3 +[71] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2 +[72] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2 +[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 4 +[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3 +[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 6, 7 +[76] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 3 +[77] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 5, 6, 7, 13 +[78] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 4 +[79] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 6, 7 + +[80] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2 +[81] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7 +[82] Songchun Zhang, Yibo Zhang, Quan Zheng, Rui Ma, Wei Hua, Hujun Bao, Weiwei Xu, and Changqing Zou. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10170-10180, 2024. 3 +[83] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21936-21947, 2025. 2 +[84] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2 +[85] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3 +[86] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph, 37, 2018. 5, 6, 7 +[87] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 3 +[88] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 2, 6 \ No newline at end of file diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/images.zip b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..57d11beddbc69bcde58d557cfd6f0ee862bb2aa9 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5f119385e2d591a46f3966eeb361272921862a39f8198c5a5525efaa54d91d4 +size 1167560 diff --git a/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/layout.json b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e84aabc49dd8134c51a5da29dfffafec4d61503 --- /dev/null +++ b/2025/SpatialCrafter_ Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations/layout.json @@ -0,0 +1,9930 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 91, + 102, + 520, + 138 + ], + "type": "text", + "content": "SpatialCrafter: Unleashing the Imagination of Video Diffusion Models for Scene Reconstruction from Limited Observations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": "Songchun Zhang" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " Huiyao Xu" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " Sitong Guo" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " Zhongwei Xie" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " \nHujun Bao" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " Weiwei Xu" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "text", + "content": " Changqing Zou" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "^{1}\\mathrm{ZJU}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "{}^{2}\\mathrm{HKUST}" + }, + { + "bbox": [ + 139, + 160, + 471, + 204 + ], + "type": "inline_equation", + "content": "{}^{3}\\mathrm{Zhejiang Lab}" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 214, + 202, + 458 + ], + "blocks": [ + { + "bbox": [ + 58, + 214, + 202, + 458 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 202, + 458 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 202, + 458 + ], + "type": "image", + "image_path": "70f4b821144851b348978c011d78cc1e1ec9d2397c7b94d94aefe37562ef2577.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 555, + 514 + ], + "type": "text", + "content": "Figure 1. Overview. Our method, SpatialCrafter, generates additional novel views of a scene from few inputs by leveraging camera trajectory-guided video diffusion models, which alleviates the ambiguity of sparse view reconstruction and robustly reconstructs the scene from the generated video frames. It demonstrates impressive performance in both indoor and outdoor scenes, while also exhibiting generalization capabilities in real scenes and synthetic images. More visualization results can be found in the supplementary material." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 214, + 554, + 457 + ], + "blocks": [ + { + "bbox": [ + 209, + 214, + 554, + 457 + ], + "lines": [ + { + "bbox": [ + 209, + 214, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 209, + 214, + 554, + 457 + ], + "type": "image", + "image_path": "9d656c76546a0eef9f4f920cd6ee3fa42d7eb94acd124dc8a5e8c87e507cd749.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 540, + 200, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 540, + 200, + 553 + ], + "spans": [ + { + "bbox": [ + 152, + 540, + 200, + 553 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 565, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 696 + ], + "type": "text", + "content": "Novel view synthesis (NVS) boosts immersive experiences in computer vision and graphics. Existing techniques, though progressed, rely on dense multi-view observations, restricting their application. We tackle the task of reconstructing photorealistic 3D scenes from only one or a few input views. We introduce SpatialCrafter, a framework that leverages the rich knowledge in video diffusion models to generate plausible additional observations, thereby alleviating reconstruction ambiguity. Through a trainable camera encoder and an epipolar attention mechanism for explicit geometric constraints, we achieve precise camera control and 3D consis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 541, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 556, + 662 + ], + "type": "text", + "content": "tency, further reinforced by a unified scale estimation strategy to handle scale discrepancies across datasets. Furthermore, by integrating monocular depth priors with semantic features in the video latent space, our framework directly regresses 3D Gaussian primitives and efficiently processes long-sequence features using a hybrid network structure. Extensive experiments show our method enhances sparse view reconstruction and restores the realistic appearance of 3D scenes. Project page: https://franklinz233.github.io/projects/spatialcrafter/." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27794" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 222 + ], + "type": "text", + "content": "Novel View Synthesis (NVS), as a key technology in computer vision and graphics, provides crucial support for immersive experiences in fields such as video games and mixed reality. Although neural reconstruction techniques have made significant progress in recent years [25, 39, 42, 62, 66], these methods typically rely on dense multiview observation data, which face numerous limitations in practical applications. Therefore, this paper focuses on a more practically valuable challenge: how to achieve high-quality 3D scene reconstruction and synthesize realistic novel views from sparse or even single-view observation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 474 + ], + "type": "text", + "content": "This problem is hard because (1) real-world scenes are diverse and complex, (2) occluded regions are unseen in sparse captures, and (3) geometric cues are too few to constrain reconstruction. To address these challenges, some methods [17, 26, 29, 40, 84, 88] constrain the optimization process by introducing regularization terms, but these methods need to be optimized for each scene, are not generalizable, and are difficult to cope with complex scenarios. Recently, some feed-forward methods [7, 8, 22, 50, 55, 60, 79, 80] have achieved generalizable sparse view scene reconstruction by directly predicting 3D Gaussian primitives for each pixel. However, such methods produce severe artifacts when reconstructing occluded regions and applying them to the extrapolation setting. To address this challenge, some methods [35, 45, 64, 65] have introduced the priors from the diffusion model to achieve sparse or single-view NVS, but they perform poorly on scene-level data and lack precise pose control and 3D consistency. This is due to the fact that low-dimensional camera information such as Euler angles and extrinsic struggle to provide comprehensive control signals to the generative models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a framework named SpatialCrafter for scene reconstruction and novel view synthesis based on sparse or single-view inputs. Our innovation lies in leveraging the rich physical world knowledge embedded in video diffusion models to provide plausible additional observations for scene reconstruction, thereby effectively reducing the problem complexity. Specifically, we first focus on enhancing precise camera control and 3D consistency in generated videos. To achieve this, we parameterize camera settings using ray embeddings [78] or metric depth-warped images [77] and incorporate them into the video diffusion model via a trainable camera encoder. Furthermore, we propose an epipolar attention mechanism that improves 3D consistency between video frames through explicit geometric constraints. Previous methods [19, 64] often experience performance degradation and difficulty generating large-motion videos when trained on multiple scene datasets. We identified that these issues primarily stem from scale ambiguity in scene datasets. To overcome this, we introduce a unified scale estimator to calibrate scene dataset," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "content": "enabling effective joint training across multiple datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 86, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 86, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 86, + 553, + 288 + ], + "type": "text", + "content": "Although we can generate visually coherent video sequences, relying solely on generated frames to reconstruct general scenes often leads to suboptimal solutions, particularly in large-scale outdoor environments and stylized scenes. To address this, we propose incorporating monocular depth priors with rich semantic features extracted from the video latent space. Leveraging these latent features, our method directly regresses 3D Gaussian primitives of the scene via a feed-forward network. Furthermore, to efficiently handle long-sequence feature interactions, we design a hybrid architecture that integrates Mamba blocks with Transformer blocks. Experiments show that our method not only improves the sparse-view reconstruction, but also accurately restores the appearance of 3D scenes, especially when extrapolating from a single view and when there are few sparse view overlaps. In conclusion, our key contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 289, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 553, + 348 + ], + "type": "text", + "content": "- We introduce a framework that effectively utilizes the physical-world knowledge embedded in video diffusion models to provide additional plausible observations for sparse-view scene reconstruction, thus reducing the ambiguity of sparse view scene reconstruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 350, + 553, + 408 + ], + "type": "text", + "content": "- To address the scale ambiguity problem that occurs in joint training across datasets, we develop a unified scale estimation approach for trajectory calibration. This solves the performance degradation problem, thus enabling effective multi-dataset training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 409, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 553, + 480 + ], + "type": "text", + "content": "- We combine monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives through a feed-forward manner. Meanwhile, we propose a hybrid architecture integrating Mamba blocks with Transformer blocks to efficiently handle long-sequence feature interactions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 515, + 499, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 499, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 499, + 528 + ], + "type": "text", + "content": "2.1. Sparse-View Scene Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 713 + ], + "type": "text", + "content": "NeRF [62] and Gaussian Splatting [25] have achieved photorealistic representations of 3D scenes, but require optimization on densely collected image sets for each scene. To address this problem, some approaches [17, 29, 71, 88] focus on constraining the optimization process of 3D representations via manually designed regularization. Some other methods [7, 8, 23, 54, 58, 69, 72, 83] are trained on large-scale datasets and directly predict the 3D representation in a feed-forward manner. Recently, some approaches [13, 50] based on end-to-end 3D reconstruction models [7, 28, 60], have achieved efficient and high-quality 3D reconstruction. However, these methods cannot handle occluded regions, and reconstruction failures (e.g., under large viewpoint changes) can severely affect the quality of novel view synthesis. More relevant to our work, some" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27795" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 555, + 285 + ], + "type": "image", + "image_path": "d9e64276501703ed894cc268c7aa2232e8544bf6a501f4edad082b4305bd1ca4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 297, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 297, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. Our generative reconstruction pipeline consists of two parts: camera-controlled video generation, and then video-based reconstruction. First, we set the exploration path based on the input views. In the video generation module, ray embeddings are used to parameterize the camera trajectories, and a cross-attention mechanism with epipolar constraints is introduced to improve the 3D consistency of the generated video. The video-based scene reconstruction pipeline integrates monocular depth priors with semantic features extracted from the video latent space, and directly regress 3D Gaussian primitives via a feed-forward manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 368, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 368, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 368, + 295, + 403 + ], + "type": "text", + "content": "works [31, 34, 57] have employed priors from video diffusion models to enhance the performance of sparse-view reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 254, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 254, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 254, + 422 + ], + "type": "text", + "content": "2.2. Diffusion-based Novel View Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 427, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 295, + 643 + ], + "type": "text", + "content": "Since text-to-image diffusion models [44] contain rich natural image priors, some methods [41, 63] directly optimize 3D representations [25, 62] via score distillation sampling [41]. Several approaches [6, 35, 45, 56] enable zero-shot novel view synthesis by fine-tuning the image and pose-conditioned generative models on large-scale 3D object [10] or scene [56, 87] datasets. However, they still have difficulty synthesizing consistent novel views. To solve this, some approaches [5, 15, 36, 38, 47, 65, 85] generate multi-views simultaneously and model the correlation between multiple views with ground-truth poses. Some recent methods [30, 37] proposed to use large-scale unposed images to train NVS methods. In addition, some methods [9, 12, 14, 48, 76, 82] utilize depth-based warping to synthesize novel views and employ the T2I model to inpaint the warped images. However, the novel views generated by these methods tend to suffer from artifacts and cumulative errors, limiting their use in generalized scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 223, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 223, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 223, + 659 + ], + "type": "text", + "content": "2.3. Controllable Video Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "While recent text-to-video diffusion models have achieved remarkable progress, they inherit the controllability limitations of their text-to-image counterparts, often requiring additional conditioning mechanisms to align generated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 368, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 524 + ], + "type": "text", + "content": "videos with user intent. Several works [18, 24, 61, 70] have been carried out to introduce a variety of conditions into video generation models. Recently, attention has focused on controlling the camera motion of the video. Some approaches [18, 53] employs motion-specific training for predefined camera movements, though this framework struggles with complex trajectory synthesis. Some methods [1, 19, 20, 64, 67] enable more complex camera control by parameterizing the camera trajectory and injecting it into a pre-trained video diffusion model via a trainable camera encoder. Another line of work [34, 43, 74, 77] advances this direction by enabling pixel-accurate view compositing through a point-based rendering approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 535, + 370, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 370, + 547 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 370, + 547 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "content": "This section begins by explaining our camera-controlled video generation model (Section 3.1), which employs ray embeddings for precise pose control. Then, we describe how to use epipolar geometry constraints to enhance 3D consistency in the video frames. Next, we introduce our scene reconstruction pipeline (Section 3.2), which aims to perform stable 3D reconstruction based on the generated video latents." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 522, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 522, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 522, + 672 + ], + "type": "text", + "content": "3.1. Camera-Conditioned Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Scale Alignment. To address this challenge, our approach first employs VGGT [58] to estimate initial camera parameters, represented by the extrinsic matrix " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = [\\mathbf{R}|\\mathbf{T}]" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27796" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": "a corresponding depth map " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": " for each video frame. To establish a consistent scale across these datasets, we leverage the Metric3D [73] to infer a canonical metric depth map, denoted " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "d_m" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". We then align our initial depth prediction, " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", to the canonical scale by calculating a scale factor " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ", defined as the ratio of the inter-percentile ranges (IPR) of the two depth maps: " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "s = \\mathrm{IPR}_{0.8,0.2}(d_v) / \\mathrm{IPR}_{0.8,0.2}(d_m)" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". This scale factor is then used to adjust the translation component of the camera extrinsics, resulting in " + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\text{scaled}} = [\\mathbf{R}|s \\cdot \\mathbf{T}]" + }, + { + "bbox": [ + 55, + 72, + 296, + 215 + ], + "type": "text", + "content": ". This procedure yields camera extrinsics with a consistent absolute scale, effectively resolving alignment issues when fusing data from different sources." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": "Camera Injection. Inspired by recent methods [19, 49, 78], we chose to use ray embeddings [19] or the depth warping frames [77] to represent the camera information. Specifically, for a ray defined by the origin " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": " and the normalized direction " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": ", we can represent it as " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "(o \\times d, d) \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": ". Given the camera parameters, the direction of the ray " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "d'" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": " corresponding to the pixel coordinates " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": " can be calculated as " + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "inline_equation", + "content": "d' = \\mathbf{RK}^{-1}(u, v, 1)^{\\top} + \\mathbf{T}" + }, + { + "bbox": [ + 55, + 216, + 296, + 403 + ], + "type": "text", + "content": ". For depth warping frames, we leverage the pretrained depth estimation model [59] to map the reference image into a 3D point cloud. Subsequently, we render novel views at specified camera poses by projecting this point cloud via the target camera parameters. Afterwards, we use a trainable conditional encoder to inject the camera information into the video model instead of fine-tuning all model parameters. The training objective is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 403, + 294, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 403, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 86, + 403, + 294, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}; z _ {0}, t, \\phi (C)\\right) \\| _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "6f32815a1d624ee34fd1e76c7c8c564be68ff6b4cd1acc631d9d5c57316730f5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "inline_equation", + "content": "\\phi (C)" + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": " represents the camera conditional encodings, " + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": " denotes the latent embeddings of the reference frame, and " + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 426, + 295, + 462 + ], + "type": "text", + "content": " indicates the timestamps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": "Epipolar Feature Aggregation. Camera embedding enhances control but 3D video consistency remains challenging due to dense self-attention allowing unrestricted cross-frame pixel interactions. We address this by incorporating epipolar geometric constraints, and aggregating features along epipolar lines. For a pixel " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{p} = (u,v)" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": ", its corresponding epipolar line " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "l_{ik}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "l_{ik}(\\pmb{p}) = \\pmb{F}_{ik} \\cdot \\tilde{\\pmb{p}}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}} = (u,v,1)^{\\top}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " are the homogeneous coordinates, and " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{ik} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " is the fundamental matrix. The fundamental matrix is decomposed as " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{ik} = \\pmb{K}_k^{-\\top} \\pmb{E}_{ik} \\pmb{K}_i^{-1}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{K}_i, \\pmb{K}_k \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " are the camera intrinsics, and " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{E}_{ik}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " is the essential matrix. We then convert epipolar lines into attention masks by computing per-pixel distances and applying a threshold, restricting attention to geometrically valid regions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Sparse-View Setting. To adapt to the sparse view input setting, we formulate the task as a video frame interpolation problem conditioned on the given start and end frames. To maximally preserve the priors from the pre-trained models [4], we inject boundary frame conditions in both latent and semantic space. Specifically, we combine latent fea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "tures of the first and last frames with their noisy latent representations, and then concatenate the extracted CLIP embeddings from both frames for cross-attentional feature injection. The training objective is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 129, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 129, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 326, + 129, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {z, z _ {0}, z _ {n} \\epsilon , C, t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (z _ {t}; z _ {0}, z _ {n}, t, \\phi (C)) \\| _ {2} ^ {2} \\right] \\quad (2)", + "image_path": "1c66131bba67c1196f8746a003a18c09837ac2b929d58fcfdf2dd8e5dbc02e3f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "inline_equation", + "content": "z_{n}" + }, + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 159, + 554, + 207 + ], + "type": "text", + "content": " are the latent embeddings of the final and initial frames, respectively. During the inference phase, we use VGGT [58] to estimate the extrinsic and intrinsic parameters for the input views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 217, + 499, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 499, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 499, + 228 + ], + "type": "text", + "content": "3.2. Video-based Scene Reconstruction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 389 + ], + "type": "text", + "content": "Motivation. There are several challenges in reconstructing scenes directly from generated videos. First, the limited number of generated frames makes it difficult to capture complete scene information. In addition, the diverse styles of video frames can pose a challenge to traditional reconstruction techniques. Furthermore, the generated video frames may contain imperfect or low-quality areas, making the reconstruction process unstable. Moreover, In stylized scenes, previous methods often yield poor results, as recovering poses and sparse point clouds from video frames is difficult, leading to many artifacts. To address this, a powerful reconstruction module specifically designed to handle the generated videos is required." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": "Latent Feature Fusion. Given input video latents " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": " and camera poses encoded as ray embeddings " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{T \\times H \\times W \\times 6}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ", we transform them into token sequences through a patchification process. Specifically, we apply spatial patchification to latent features to obtain latent tokens " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t \\in \\mathbb{R}^{N \\times d_z}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ", while ray embeddings undergo 3D-patchification to produce pose tokens " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t \\in \\mathbb{R}^{N \\times d_p}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ". To incorporate explicit geometric guidance, we leverage monocular depth estimation within our pipeline. For each input RGB video frame, we estimate dense depth maps " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D} \\in \\mathbb{R}^{T \\times H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ", which are processed by a dedicated depth encoder to yield depth tokens " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_t \\in \\mathbb{R}^{N \\times d_d}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ". The three token sets are channel-wise concatenated to form a unified representation " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [\\mathbf{z}_t; \\mathbf{p}_t; \\mathbf{d}_t] \\in \\mathbb{R}^{N \\times (d_z + d_p + d_d)}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ", which is then linearly projected to a lower-dimensional space " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": " before being fed into a sequence of mamba and transformer blocks. Mamba [16] achieves the same token sequence processing functionality as Transformer, but reduces computational complexity from " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(L^2)" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(L)" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ", making it particularly suitable for dense reconstruction tasks. In our Mamba blocks, we implement bi-directional scanning across the token sequence. First, we compute state parameters " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}, \\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": " from the input tokens using a linear projection. Then, we execute the State Space Model in both forward (" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_f" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ") and backward (" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_b" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": ") directions. The final output of each Mamba block is computed as " + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\mathbf{y}_f + \\mathbf{y}_b" + }, + { + "bbox": [ + 313, + 391, + 556, + 713 + ], + "type": "text", + "content": " followed by a final linear transformation." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27797" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 219, + 207 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 219, + 207 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 219, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 219, + 207 + ], + "type": "image", + "image_path": "e753cdbf55dd4c8f8b1103458c77bd44f1c8332731ebfb8f624ac09d27ad0a35.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 221, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 221, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 555, + 246 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of generated videos. Compared to other benchmark methods, the videos generated by our method have better foreground object-background consistency, while also being able to generate videos with larger motion amplitude." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 378, + 207 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 378, + 207 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 378, + 207 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 378, + 207 + ], + "type": "image", + "image_path": "0bc77672b096c01e27dc68ae50d8f7993d0a47a9d6745badef542dc876c60e4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 71, + 553, + 207 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 553, + 207 + ], + "type": "image", + "image_path": "a81772d829b6c6620a3627e29da7927d39bdbef2a99355459a4ab12757be764e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 257, + 296, + 390 + ], + "blocks": [ + { + "bbox": [ + 56, + 257, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 56, + 257, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 296, + 390 + ], + "type": "table", + "html": "
MethodFVD↓FID↓Rerr↓Terr↓LPIPS↓PSNR↑SSIM↑
RealEstate10K
MotionCtrl [64]22.65230.120.2340.7980.29914.720.404
CameraCtrl [19]21.48188.210.0540.1270.23017.330.516
ViewCrafter [77]20.96204.180.0550.1530.21518.950.503
Ours18.25183.250.0520.1030.20719.210.523
Tanks-and-Temples
MotionCtrl [64]30.25290.380.8381.5050.31514.640.388
CameraCtrl [19]24.41244.820.1180.2940.28515.380.469
ViewCrafter [77]22.50231.350.1260.3070.24716.240.508
Ours20.17192.520.0970.1750.22816.120.501
DL3DV
MotionCtrl [64]25.65249.510.4731.1180.31214.380.386
CameraCtrl [19]22.76233.540.0950.2380.26216.330.489
ViewCrafter [77]20.59211.240.0930.2440.24217.120.521
Ours18.21171.520.0630.1340.22517.020.537
", + "image_path": "462d41b077ad1dc79923f54d03f5ffd8915910edd97044606efd0de7b105cdfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 295, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 427 + ], + "type": "text", + "content": "Table 1. Quantitative comparison to the camera conditioned video generation method on RealEstate10K [86], DL3DV [32], and Tanks-and-Temple [27] dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 445, + 295, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 565 + ], + "type": "text", + "content": "Gaussian Decoding. We design a lightweight decoder module that efficiently transforms output feature tokens into per-pixel Gaussian parameters. The decoding module consists of 3D-DeConv layers, which generates refined Gaussian feature map " + }, + { + "bbox": [ + 54, + 445, + 295, + 565 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{(T \\times H \\times W) \\times 12}" + }, + { + "bbox": [ + 54, + 445, + 295, + 565 + ], + "type": "text", + "content": ". This 12-channel representation precisely encodes the complete set of Gaussian parameters: RGB color (3 channels), scale factors (3 channels), rotation quaternion (4 channels), opacity (1 channel), and ray distance (1 channel). The final output of the model is the merge of 3D Gaussians from all input video frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 615 + ], + "type": "text", + "content": "Training Objective. During training, we render images from predicted Gaussians using randomly selected supervision views. Our approach employs a composite loss that integrates three components:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 628, + 295, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 628, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 88, + 628, + 295, + 642 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\lambda_ {1} \\mathcal {L} _ {\\text {m s e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p e r c}} + \\lambda_ {3} \\mathcal {L} _ {\\text {d e p t h}}, \\tag {3}", + "image_path": "44db7de6f5b7ba89e52b1c84ab934b56ef563cfbd34771b178f56732e482ee87.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mse}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " represents pixel-wise mean squared error, " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{perc}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " denotes perceptual loss, and " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{depth}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": " enforces depth consistency. This formulation jointly optimizes for photometric accuracy, high-level perceptual fidelity, and geometric coherence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 258, + 391, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 258, + 391, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 258, + 391, + 272 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 430, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 430, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 430, + 294 + ], + "type": "text", + "content": "4.1. Experiment settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 421 + ], + "type": "text", + "content": "Datasets. To comprehensively capture the underlying distribution of real-world scenarios, we trained our video diffusion model on three diverse datasets: RealEstate-10K [86], ACID [33], and DL3DV-10K [32]. The Re10K dataset from YouTube comprises 67,477 training and 7,289 testing indoor and outdoor camera trajectories. The ACID dataset focuses on natural landscapes, with 11,075 training and 1,972 testing scenes. The DL3DV-10K dataset is a large-scale collection featuring 10,510 scenes captured under controlled, standardized conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "text", + "content": "Implementation Details. Our video generation model is based on SVD [3], an image-video diffusion model based on UNet. In our experiments, we use a relative camera system in which all camera poses are converted to poses relative to the first frame. The camera in the first frame is located at the world origin. In the first stage, we train the model at a resolution of " + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "inline_equation", + "content": "320 \\times 512" + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "text", + "content": " for 50,000 iterations with the frame length set to 25. Subsequently, we train on " + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "inline_equation", + "content": "576 \\times 1024" + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "text", + "content": " for 10,000 iterations to adapt to high resolution. The learning rate is set to " + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 313, + 423, + 556, + 616 + ], + "type": "text", + "content": " with a warmup of 1,000 steps, using the Adam optimizer. We chose Lightning as the training framework, using mixed-precision fp16 and DeepSpeed ZeRO-2. We trained the proposed method and its variants on 16 NVIDIA A800 GPUs with a batch size of 32. During inference, we adopt DDIM sampler [51] with classifier-free guidance [21]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Progressive Training. Controlling a video generation model to generate arbitrary motion trajectories remains a challenging task. To enable robust arbitrary trajectory generation, we employ a three-stage curriculum learning strategy. The model first trains on smooth camera motions with small temporal intervals, then progressively adapts to more complex motion patterns through linear interval scheduling, and finally incorporates random sampling intervals." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27798" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 301, + 259 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 301, + 259 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 301, + 259 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 301, + 259 + ], + "type": "image", + "image_path": "d022385becdeb84df97778ce77a32a18b926b872fb16b84edc4be7c98dc9c977.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 264, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 307, + 264, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 553, + 299 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on DL3DV and Tank-and-Temples dataset. Our method reconstructs better than baselines, even with limited image overlap and in non-overlapping regions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 315, + 296, + 460 + ], + "blocks": [ + { + "bbox": [ + 57, + 315, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 57, + 315, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 57, + 315, + 296, + 460 + ], + "type": "table", + "html": "
Method3-view6-view9-view
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Mip-NeRF360
Zip-NeRF [2]12.770.2710.70513.610.2840.66314.300.3120.633
ReconFusion [65]15.500.3580.58516.930.4010.54418.190.4320.511
CAT3D [15]16.620.3770.51517.720.4250.48218.670.4600.460
ReconX [34]17.160.4350.40719.200.4730.37820.130.4820.356
ViewCrafter [77]14.510.3150.68215.870.3360.66516.450.3480.655
Ours17.320.4390.41819.420.5100.37120.190.5360.345
DTU
Zip-NeRF [2]9.180.6010.3838.840.5890.3709.230.5920.364
FSGS [88]17.340.8180.16921.550.8800.12724.330.9110.106
ReconFusion [65]20.740.8750.12423.620.9040.10524.620.9210.094
CAT3D [15]22.020.8440.12124.280.8990.09525.920.9280.073
ViewCrafter [77]15.630.5220.38315.030.5250.40814.920.4870.452
Ours27.920.8790.10328.820.8950.08229.030.9050.065
LLFF
Zip-NeRF [2]17.230.5740.37320.710.7640.22123.630.8300.166
FSGS [88]20.310.6520.28824.200.8110.17325.320.8560.136
ReconFusion [65]21.340.7240.20324.250.8150.15225.210.8480.134
CAT3D [15]21.580.7310.18124.710.8330.12125.630.8600.107
ViewCrafter [77]17.730.5210.33217.430.5120.34517.330.4880.351
Ours22.040.7410.16525.110.8140.12425.950.8380.103
", + "image_path": "3e21525f7f905fc1c3db96aa87bffc3f160ca5def693bbdbfefe13ba1e1f922f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "type": "text", + "content": "This gradual transition from simple to complex trajectories proves crucial for high-quality video generation with arbitrary trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 141, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 141, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 141, + 549 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 216, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 216, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 216, + 566 + ], + "type": "text", + "content": "4.2.1. Controllable Video Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "type": "text", + "content": "Benchmark and Metrics. We evaluate our video generation model against three baselines [19, 64, 77] on three benchmark datasets: RealEstate10K [86] (500 randomly selected videos with first frame as image condition and subsequent frames at stride 3 for pose guidance), DL3DV-140 [32] (500 video clips at stride 2), and Tanks-and-Temples [27] (100 sequences at stride 4 from 14 scenes with COLMAP-annotated poses) for out-of-domain generalization testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Metrics. Our evaluation framework employs multiple metrics to assess performance: visual quality measured through Fréchet Video Distance (FVD) and Fréchet Inception Dis" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 73, + 553, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 274, + 301, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 301, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 301, + 307 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on Re10K [86]. Compared to baselines, we obtain superior reconstruction from limited overlap, and enhanced geometry reconstruction in non-overlapping regions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 73, + 553, + 248 + ], + "lines": [ + { + "bbox": [ + 310, + 73, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 310, + 73, + 553, + 248 + ], + "type": "image", + "image_path": "d557e1f8a2c4c4460246b66655ce13750d2a53e1f471757a24ae090f0b468848.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 315, + 555, + 431 + ], + "blocks": [ + { + "bbox": [ + 55, + 464, + 295, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 485 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of sparse view 3D reconstruction on Out-of-Domain Datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 315, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 315, + 315, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 315, + 555, + 431 + ], + "type": "table", + "html": "
MethodRealEstate10KACID
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
pixelNeRF [75]20.430.5890.55020.970.5470.533
GPNR [52]24.110.7930.25525.280.7640.332
AttnRend [11]24.780.8200.21326.880.7990.218
MuRF [68]26.100.8580.14328.090.8410.155
pixelSplat [7]25.890.8580.14228.140.8390.150
MVSplat [8]26.390.8690.12828.250.8430.144
GS-LRM [79]28.100.8920.114---
DepthSplat [69]24.230.7900.217---
ReconX [34]28.310.9120.08828.840.8910.101
ViewCrafter [77]24.220.7880.21823.480.6600.299
Ours28.350.8620.12128.930.8990.116
", + "image_path": "107a1fca9a81bb69ea2abe54c37652461b137594e1acbe595b70c3df125dccce.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 467 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of two-view sparse view reconstruction methods on RealEstate10K [86] and ACID [33] dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": "tance (FID); camera control precision quantified by rotation error " + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "inline_equation", + "content": "(R_{err})" + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": " and translation error " + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "inline_equation", + "content": "(T_{err})" + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": " computed from camera poses extracted via COLMAP and normalized relative to the first frame; and visual consistency evaluated using PSNR, SSIM, and LPIPS [81] between generated and ground-truth views. To ensure fair comparison across methods with varying output capabilities, we restrict visual similarity assessment to the first 14 frames, as generated content typically diverges progressively from the conditional single-view input as the scene extends." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Comparison. Table 1 shows that our method excels in qualitative results, outperforming other methods in terms of visual quality, pose control, and 3D consistency. Figure 3 further demonstrates our advantage in 3D consistency and visual quality. This performance improvement is attributed to the introduction of a camera parameterization method and an epipolar attention module, which respectively enhance pose control ability and the 3D consistency of generated videos. In addition, training on a diverse scene dataset sig" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27799" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 164 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 164 + ], + "type": "table", + "html": "
MethodSmallLarge
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
pixelNeRF [75]18.4170.6010.52620.8690.6390.458
AttnNeRF [11]19.1510.6630.36825.8970.8450.229
pixelSplat [7]20.2630.7170.26627.1510.8790.122
MVSplat [8]20.3530.7240.25427.4080.8840.116
DUSt3R [60]14.1010.4320.46816.4270.4530.402
CoPoNeRF [22]17.3930.5850.46220.4640.6520.358
Ours22.5140.7840.21327.4110.9130.109
", + "image_path": "c246ae6ab063c5553f37ed2fba94bf0165351eb4e500fdb5ae9a5efe582114af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 219, + 296, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 168, + 296, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 213 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on the RealEstate10K [86] dataset. Small and Large refer to input images with low and high overlap ratios, respectively. Greater overlap means greater temporal coherence between adjacent images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 293 + ], + "type": "table", + "html": "
DatasetTanks-and-TemplesDL3DV
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LucidDreamer [9]14.5520.3640.41515.1260.4520.431
ZeroNVS [45]14.7340.3810.48015.1630.4640.466
MotionCtrl [64]15.3220.4260.40416.8890.5280.392
ViewCrafter [77]21.2860.6540.19321.3730.6860.244
Ours22.3310.7420.21322.7820.7910.202
", + "image_path": "ccda7f8f10ff2093f62910894dd50e1b694acae1b1a6c51ac1cca20c1865d2cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 319 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of single-view novel view synthesis on DL3DV [32] and Tank-and-Templets [27] benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 329, + 293, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 293, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 293, + 342 + ], + "type": "text", + "content": "nificantly improves the generalization ability of the model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 204, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 204, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 204, + 361 + ], + "type": "text", + "content": "4.2.2. Sparse View Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 365, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 593 + ], + "type": "text", + "content": "Benchmark and Metrics. To comprehensively evaluate the performance of our method, we compare it with two categories of reconstruction methods: (1) two-view feedforward reconstruction methods [7, 8, 11, 52, 69, 79], and (2) optimization-based sparse-view reconstruction methods [13, 15, 34, 77]. The evaluation employs two types of test sets: in-domain test sets (from RealEstate10K [86] and Acid [33]) and out-of-domain test sets. To ensure a fair comparison, the out-of-domain test set is the same as that used in previous work [15, 65]. For evaluating scene reconstruction quality, we adopt three metrics: PSNR, SSIM, and LPIPS [81]. Additionally, we categorize the evaluation into two groups based on the overlap between input views, with detailed overlap estimation methods provided in the supplementary materials. The evaluation protocol follows a consistent procedure: first, we reconstruct the 3D scene based on the input images, then render from novel views, and compute the metrics between the rendered images and the reference images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Comparison. As shown in Tables 2-4 and Figures 4-5, our method performs well on various evaluation metrics, particularly in challenging scenarios with occlusions and limited view overlap. This improvement can be attributed to our approach that utilizes scene priors and sparse input views to generate helpful additional observations, thereby enhancing visual correlations between different viewpoints and helping to reduce the complexity of sparse-view reconstruction. Compared to the method [77] conditioning on point-based rendering, our approach shows advantages in reconstructing" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 555, + 248 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 555, + 248 + ], + "type": "image", + "image_path": "0d8a4bd7ccf0d825517c755150babf0b111e8e0f9c59490f351fdc46f275f662.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 263, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 286 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of novel views rendered from scenes reconstructed using other 3D generation methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 319 + ], + "type": "text", + "content": "scenes with thin structures, leading to improved 3D reconstruction results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 326, + 458, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 458, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 458, + 338 + ], + "type": "text", + "content": "4.2.3. Single View 3D Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 496 + ], + "type": "text", + "content": "Benchmark and Metrics. In single-view generation, we compare our method with several generative methods [45, 46, 64, 77] on the Tank-and-Temple [27] and DL3DV-140 [32] datasets. Noted that we begin by reconstructing the 3D scene from the given images, followed by calculating the metrics using renderings from novel views. Similarly, we used the PSNR, SSIM, and LPIPS [81] metrics to evaluate our results. Evaluation in this underconstrained setting is challenging, as multiple 3D scenes can produce consistent generations for a given view. Thus, we measure metrics using only temporally adjacent frames to the conditional image, 14 sampled frames and poses captured after the conditional image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 605 + ], + "type": "text", + "content": "Comparison. As shown in Table 5, our method consistently achieves superior performance on image quality metrics compared to the baselines. For a fair comparison with ZeroNVS [45] and LucidDreamer [9], which are limited to square image inputs, we crop the generated novel views before computing the quantitative metrics. Figure 6 shows that scenes reconstructed using our method have less noise and distortion in occluded areas. For more visualizations, see Figure 7." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 408, + 624 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Ablation on Video Diffusion Model. We evaluate the effectiveness of the design choices. As shown in Table 6, ray embedding achieves significant improvements in camera control accuracy and rendering quality compared to other camera encoding methods by introducing denser position encoding and optimizing geometric correspondences. Specifically, cross-frame consistency shows notable en" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27800" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 81, + 138, + 128 + ], + "blocks": [ + { + "bbox": [ + 57, + 81, + 138, + 128 + ], + "lines": [ + { + "bbox": [ + 57, + 81, + 138, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 81, + 138, + 128 + ], + "type": "image", + "image_path": "b9ef69e940233b92e69350b942ec21da862ad7e49e00c29b50c8025277ed02c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 139, + 82, + 220, + 128 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 300, + 79 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 300, + 79 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 300, + 79 + ], + "type": "text", + "content": "Generated Videos" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 82, + 220, + 128 + ], + "lines": [ + { + "bbox": [ + 139, + 82, + 220, + 128 + ], + "spans": [ + { + "bbox": [ + 139, + 82, + 220, + 128 + ], + "type": "image", + "image_path": "d7e6a23f3bf8a5ddb9fd75e05b3cc953f94de2bbcf3e2bf8ae5c92ae9cbe46dd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 82, + 302, + 128 + ], + "blocks": [ + { + "bbox": [ + 221, + 82, + 302, + 128 + ], + "lines": [ + { + "bbox": [ + 221, + 82, + 302, + 128 + ], + "spans": [ + { + "bbox": [ + 221, + 82, + 302, + 128 + ], + "type": "image", + "image_path": "00e15124fa0801ee5d20cab739ba44509f5a609072445804cffe0eb4273e742a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 303, + 82, + 384, + 128 + ], + "blocks": [ + { + "bbox": [ + 303, + 82, + 384, + 128 + ], + "lines": [ + { + "bbox": [ + 303, + 82, + 384, + 128 + ], + "spans": [ + { + "bbox": [ + 303, + 82, + 384, + 128 + ], + "type": "image", + "image_path": "9ea890071fbf10b26b612d57323cd2adac9d1114ed6881f4653c805ffa6dd168.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 385, + 82, + 466, + 128 + ], + "blocks": [ + { + "bbox": [ + 385, + 82, + 466, + 128 + ], + "lines": [ + { + "bbox": [ + 385, + 82, + 466, + 128 + ], + "spans": [ + { + "bbox": [ + 385, + 82, + 466, + 128 + ], + "type": "image", + "image_path": "03227ef2a56faa7695e3af222ff1e808fa9c3458e850f111f14cdca46b032eba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 468, + 81, + 552, + 128 + ], + "blocks": [ + { + "bbox": [ + 497, + 70, + 522, + 79 + ], + "lines": [ + { + "bbox": [ + 497, + 70, + 522, + 79 + ], + "spans": [ + { + "bbox": [ + 497, + 70, + 522, + 79 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 81, + 552, + 128 + ], + "lines": [ + { + "bbox": [ + 468, + 81, + 552, + 128 + ], + "spans": [ + { + "bbox": [ + 468, + 81, + 552, + 128 + ], + "type": "image", + "image_path": "e3555563403fa9f3d2e3e7dcf76a14d1499ba9989e25cc536f42f90017987af5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 129, + 138, + 176 + ], + "blocks": [ + { + "bbox": [ + 58, + 129, + 138, + 176 + ], + "lines": [ + { + "bbox": [ + 58, + 129, + 138, + 176 + ], + "spans": [ + { + "bbox": [ + 58, + 129, + 138, + 176 + ], + "type": "image", + "image_path": "88f1c0e74a6eda5f0043e8c0533954ebcb32ac197924fca9a493f0a72c6d0017.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 139, + 129, + 220, + 176 + ], + "blocks": [ + { + "bbox": [ + 139, + 129, + 220, + 176 + ], + "lines": [ + { + "bbox": [ + 139, + 129, + 220, + 176 + ], + "spans": [ + { + "bbox": [ + 139, + 129, + 220, + 176 + ], + "type": "image", + "image_path": "13794b7537442d126366394a030cfb6c75bb4b9f746aee10383254301f4c99ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 221, + 129, + 302, + 176 + ], + "blocks": [ + { + "bbox": [ + 221, + 129, + 302, + 176 + ], + "lines": [ + { + "bbox": [ + 221, + 129, + 302, + 176 + ], + "spans": [ + { + "bbox": [ + 221, + 129, + 302, + 176 + ], + "type": "image", + "image_path": "b7283d0c6dd546430b9df1cee4f84daeb2a02b3cd12e814b3e3e194a2d90ba94.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 303, + 129, + 384, + 176 + ], + "blocks": [ + { + "bbox": [ + 303, + 129, + 384, + 176 + ], + "lines": [ + { + "bbox": [ + 303, + 129, + 384, + 176 + ], + "spans": [ + { + "bbox": [ + 303, + 129, + 384, + 176 + ], + "type": "image", + "image_path": "6f9881c5634ab7a4b58509264bfcceb87eb1472f2cfce3749b166237d9d28685.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 384, + 129, + 466, + 176 + ], + "blocks": [ + { + "bbox": [ + 384, + 129, + 466, + 176 + ], + "lines": [ + { + "bbox": [ + 384, + 129, + 466, + 176 + ], + "spans": [ + { + "bbox": [ + 384, + 129, + 466, + 176 + ], + "type": "image", + "image_path": "29aed6526ba28a213fa59370562645fa62291dfb8e486f7c423f54d3d678613f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 468, + 129, + 552, + 176 + ], + "blocks": [ + { + "bbox": [ + 468, + 129, + 552, + 176 + ], + "lines": [ + { + "bbox": [ + 468, + 129, + 552, + 176 + ], + "spans": [ + { + "bbox": [ + 468, + 129, + 552, + 176 + ], + "type": "image", + "image_path": "f2349bfa7ab4e79770a429fd08eecd297f32bf1944cf9b2dede94cb860388c8f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 58, + 177, + 138, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 177, + 138, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 138, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 138, + 224 + ], + "type": "image", + "image_path": "e2c2da9932025e41bbe6986fd28e1659d93ca5352b39c1641485d2483ffa6f21.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 139, + 177, + 220, + 224 + ], + "blocks": [ + { + "bbox": [ + 139, + 177, + 220, + 224 + ], + "lines": [ + { + "bbox": [ + 139, + 177, + 220, + 224 + ], + "spans": [ + { + "bbox": [ + 139, + 177, + 220, + 224 + ], + "type": "image", + "image_path": "45bae437aa03a483ac6ea8d5bf4ce2149f2138448235f57d066e1edf0956da76.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 221, + 177, + 302, + 224 + ], + "blocks": [ + { + "bbox": [ + 221, + 177, + 302, + 224 + ], + "lines": [ + { + "bbox": [ + 221, + 177, + 302, + 224 + ], + "spans": [ + { + "bbox": [ + 221, + 177, + 302, + 224 + ], + "type": "image", + "image_path": "8a79896b89416be8df2129509237e929c1db219653b30c3254a1ecaeac60d729.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 303, + 177, + 384, + 224 + ], + "blocks": [ + { + "bbox": [ + 303, + 177, + 384, + 224 + ], + "lines": [ + { + "bbox": [ + 303, + 177, + 384, + 224 + ], + "spans": [ + { + "bbox": [ + 303, + 177, + 384, + 224 + ], + "type": "image", + "image_path": "a1ef049494fdf5c36c7c3b8c247527f4d7663b9159911ad967fb08aa6b61b34a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 384, + 177, + 466, + 224 + ], + "blocks": [ + { + "bbox": [ + 384, + 177, + 466, + 224 + ], + "lines": [ + { + "bbox": [ + 384, + 177, + 466, + 224 + ], + "spans": [ + { + "bbox": [ + 384, + 177, + 466, + 224 + ], + "type": "image", + "image_path": "39392d2c6448e620d7658946199f4a37a3319f8cd4928e0d198b0c076f164b7b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 177, + 552, + 224 + ], + "blocks": [ + { + "bbox": [ + 468, + 177, + 552, + 224 + ], + "lines": [ + { + "bbox": [ + 468, + 177, + 552, + 224 + ], + "spans": [ + { + "bbox": [ + 468, + 177, + 552, + 224 + ], + "type": "image", + "image_path": "6fe593d8c4b5641733e60cbe8481e8c17b27387272825c7f55127678a5440f3a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 226, + 138, + 273 + ], + "blocks": [ + { + "bbox": [ + 58, + 226, + 138, + 273 + ], + "lines": [ + { + "bbox": [ + 58, + 226, + 138, + 273 + ], + "spans": [ + { + "bbox": [ + 58, + 226, + 138, + 273 + ], + "type": "image", + "image_path": "f9aebd835454c7bf2de2b4d77db7ce5befa62e41b6027d3807a5257ddb82ad45.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 289, + 490, + 300 + ], + "lines": [ + { + "bbox": [ + 118, + 289, + 490, + 300 + ], + "spans": [ + { + "bbox": [ + 118, + 289, + 490, + 300 + ], + "type": "text", + "content": "Figure 7. Additional visualizations of generated videos and their corresponding 3DGS representations." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 139, + 226, + 220, + 273 + ], + "blocks": [ + { + "bbox": [ + 139, + 226, + 220, + 273 + ], + "lines": [ + { + "bbox": [ + 139, + 226, + 220, + 273 + ], + "spans": [ + { + "bbox": [ + 139, + 226, + 220, + 273 + ], + "type": "image", + "image_path": "affc0dcc4ba8d5df3016827ca36cbb33a1b8470ec2b60abd6802e123c0e1c694.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 221, + 226, + 302, + 273 + ], + "blocks": [ + { + "bbox": [ + 221, + 226, + 302, + 273 + ], + "lines": [ + { + "bbox": [ + 221, + 226, + 302, + 273 + ], + "spans": [ + { + "bbox": [ + 221, + 226, + 302, + 273 + ], + "type": "image", + "image_path": "328c9665967acf4c342c619c6378edd9c0c7f703f5454961b753b34140b8eadc.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 303, + 226, + 384, + 273 + ], + "blocks": [ + { + "bbox": [ + 303, + 226, + 384, + 273 + ], + "lines": [ + { + "bbox": [ + 303, + 226, + 384, + 273 + ], + "spans": [ + { + "bbox": [ + 303, + 226, + 384, + 273 + ], + "type": "image", + "image_path": "5d8bdd67b36f6b474f503a377d3a551dd55a3d661737a698f67bbb36fab14e7d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 384, + 226, + 466, + 273 + ], + "blocks": [ + { + "bbox": [ + 384, + 226, + 466, + 273 + ], + "lines": [ + { + "bbox": [ + 384, + 226, + 466, + 273 + ], + "spans": [ + { + "bbox": [ + 384, + 226, + 466, + 273 + ], + "type": "image", + "image_path": "d178091b024c2ce8fc6257fa3945b83fe6547d105301d9a7b7f16d76bff60326.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 468, + 225, + 552, + 274 + ], + "blocks": [ + { + "bbox": [ + 468, + 225, + 552, + 274 + ], + "lines": [ + { + "bbox": [ + 468, + 225, + 552, + 274 + ], + "spans": [ + { + "bbox": [ + 468, + 225, + 552, + 274 + ], + "type": "image", + "image_path": "91919ad4b8b4cd7140536b8760164d39163f1f3cc2e66f5de4e7171009dc8dfc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 57, + 314, + 296, + 397 + ], + "blocks": [ + { + "bbox": [ + 57, + 314, + 296, + 397 + ], + "lines": [ + { + "bbox": [ + 57, + 314, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 57, + 314, + 296, + 397 + ], + "type": "table", + "html": "
MethodTerr↓Rerr↓COLMAPerr↓FVD↓
Raw Value Emb.1.5922.37613.5%81.854
Quaternion Emb.1.4572.39413.1%80.368
W/o Ray Emb.2.6833.14515.7%112.843
W/o Epipolar1.1252.23512.2%79.547
W/o Scale Align.1.7122.49814.3%81.679
Full model1.0142.2184.2%78.132
", + "image_path": "de51b3e51a7e805523d2db398733b5a89b5e5a9d5937fa69f0fe7ebcd4a112c5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 61, + 400, + 290, + 411 + ], + "lines": [ + { + "bbox": [ + 61, + 400, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 400, + 290, + 411 + ], + "type": "text", + "content": "Table 6. Ablation study on the video diffusion model variants." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 313, + 395, + 366 + ], + "blocks": [ + { + "bbox": [ + 318, + 313, + 395, + 366 + ], + "lines": [ + { + "bbox": [ + 318, + 313, + 395, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 313, + 395, + 366 + ], + "type": "image", + "image_path": "f7e200a74cd5ea78bd13b664e454b9b2bbde9a8d21b991594f4e48696cd20bfd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 395, + 314, + 471, + 366 + ], + "blocks": [ + { + "bbox": [ + 395, + 314, + 471, + 366 + ], + "lines": [ + { + "bbox": [ + 395, + 314, + 471, + 366 + ], + "spans": [ + { + "bbox": [ + 395, + 314, + 471, + 366 + ], + "type": "image", + "image_path": "6807d3a5a8cbdf96992d799baf51ec2bae3bf887eba7bc45386137f502c8ce08.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 471, + 314, + 547, + 366 + ], + "blocks": [ + { + "bbox": [ + 471, + 314, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 471, + 314, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 471, + 314, + 547, + 366 + ], + "type": "image", + "image_path": "9d2179cf6fec3340966eb0d9edaafa15d78df20be333ba4593970f87325b7178.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 319, + 366, + 395, + 411 + ], + "blocks": [ + { + "bbox": [ + 319, + 366, + 395, + 411 + ], + "lines": [ + { + "bbox": [ + 319, + 366, + 395, + 411 + ], + "spans": [ + { + "bbox": [ + 319, + 366, + 395, + 411 + ], + "type": "image", + "image_path": "bd6e15a969446c95e3eecda69a1c14b0263753fd06bbc1d20cd5d98426bd2c0b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 414, + 550, + 425 + ], + "lines": [ + { + "bbox": [ + 318, + 414, + 550, + 425 + ], + "spans": [ + { + "bbox": [ + 318, + 414, + 550, + 425 + ], + "type": "text", + "content": "Figure 9. Ablation on the feed-forward Reconstruction Module." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 395, + 366, + 471, + 411 + ], + "blocks": [ + { + "bbox": [ + 395, + 366, + 471, + 411 + ], + "lines": [ + { + "bbox": [ + 395, + 366, + 471, + 411 + ], + "spans": [ + { + "bbox": [ + 395, + 366, + 471, + 411 + ], + "type": "image", + "image_path": "82dd46a36944c72d03983280fe7d077e804e660ae707accf0e85e4508ee90615.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 471, + 366, + 547, + 411 + ], + "blocks": [ + { + "bbox": [ + 471, + 366, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 471, + 366, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 471, + 366, + 547, + 411 + ], + "type": "image", + "image_path": "1cb64f90e35bb41b39e079266e97ccc03b03ea4b50947162a8c3cae6068c3ac4.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 56, + 418, + 302, + 506 + ], + "blocks": [ + { + "bbox": [ + 56, + 418, + 302, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 418, + 302, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 302, + 506 + ], + "type": "image", + "image_path": "ed22943e74657066c6571145654cabbe6fb25d234c396acea7ee37dd709f1e22.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 295, + 563 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 563 + ], + "type": "text", + "content": "Figure 8. Ablation on the Scale Alignment. The video reconstruction results obtained based on scale-alignment training exhibit reduced artifacts while improving the 3D consistency of the scene." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 628 + ], + "type": "text", + "content": "hancement. Furthermore, when incorporating geometric constraints through Epipolar Attention, the system achieves optimal performance in terms of camera controllability and 3D consistency." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "content": "Scale Alignment. As illustrated in Table 6 and Figure 8, the incorporation of dataset metric scale alignment significantly improves the pose control accuracy and 3D consistency of the generated video." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Video Latent-based Reconstruction. As demonstrated in Figure 9, incorporating generative priors and depth features in our method significantly enhances the fidelity of fine" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 446, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 494 + ], + "type": "text", + "content": "grained geometric details in scene reconstruction. By introducing priors from video generative models, we reduced the difficulty of sparse-view reconstruction and achieved reasonable completion of unseen regions." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 388, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We present SpatialCrafter, a framework for scene reconstruction and novel view synthesis from sparse or single-view inputs. By leveraging video diffusion models to generate plausible additional observations, we effectively reduce the complexity of sparse-view scene reconstruction. Our key contributions include: (1) precise camera control via ray embeddings or depth-warped images with a trainable condition encoder; (2) a unified scale estimator solving the scale ambiguity in multi-dataset training; and (3) a hybrid Mamba-Transformer architecture that combines monocular depth priors with semantic features from video latent space to directly regress 3D Gaussian primitives. Experiments show our method outperforms existing methods, especially in single-view extrapolation and scenarios with little overlaps. Future work will focus on extending to dynamic scene." + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27801" + } + ] + } + ], + "index": 44 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Sherwin Bahmani, Ivan Skorokhodov, Guocheng Qian, Aliaksandr Siarohin, Willi Menapace, Andrea Tagliasacchi, David B Lindell, and Sergey Tulyakov. Ac3d: Analyzing and improving 3d camera control in video diffusion transformers. arXiv preprint arXiv:2411.18673, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 5, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 360 + ], + "type": "text", + "content": "[5] Chenjie Cao, Chaohui Yu, Shang Liu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvgenmaster: Scaling multi-view generation from any image via 3d priors enhanced diffusion model. arXiv preprint arXiv:2411.16157, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 416 + ], + "type": "text", + "content": "[6] Eric R Chan, Koki Nagano, Matthew A Chan, Alexander W Bergman, Jeong Joon Park, Axel Levy, Miika Aittala, Shalini De Mello, Tero Karras, and Gordon Wetzstein. Generative novel view synthesis with 3d-aware diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 461 + ], + "type": "text", + "content": "[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, 2024. 2, 6, 7, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "text", + "content": "[8] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2024. 2, 6, 7, 13" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 294, + 552 + ], + "type": "text", + "content": "[9] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "text", + "content": "[10] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 654 + ], + "type": "text", + "content": "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 294, + 689 + ], + "type": "text", + "content": "[12] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 106 + ], + "type": "text", + "content": "Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv:2403.20309, 2024. 2, 7, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[14] Rafail Fridman, Amit Abecasis, Yoni Kasten, and Tali Dekel. Scenescape: Text-driven consistent scene generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 206 + ], + "type": "text", + "content": "[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 3, 6, 7, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 239 + ], + "type": "text", + "content": "[16] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "text", + "content": "[17] Guangcong, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In ICCV, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "text", + "content": "[18] Yuwei Guo, Ceyuan Yang, Anyi Rao, Yaohui Wang, Yu Qiao, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "text", + "content": "[19] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation. arXiv preprint arXiv:2404.02101, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "text", + "content": "[20] Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractrl ii: Dynamic scene exploration via camera-controlled video diffusion models. arXiv preprint arXiv:2503.10592, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5, 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 510 + ], + "type": "text", + "content": "[22] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence pose and nerf for generalized pose-free novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20196-20206, 2024. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 555 + ], + "type": "text", + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In ICLR, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 557, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 600 + ], + "type": "text", + "content": "[24] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "text", + "content": "[26] Mijeong Kim, Seonguk Seo, and Bohyung Han. Informer: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4), 2017. 5, 6, 7, 18" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27802" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[28] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[29] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In CVPR, 2024. 2, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 218 + ], + "type": "text", + "content": "[30] Lingen Li, Zhaoyang Zhang, Yaowei Li, Jiale Xu, Wenbo Hu, Xiaoyu Li, Weihao Cheng, Jinwei Gu, Tianfan Xue, and Ying Shan. Nvcomposer: Boosting generative novel view synthesis with multiple sparse and unposed images. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 777-787, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "text", + "content": "[31] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos N Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 798-810, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[32] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In CVPR, 2024. 5, 6, 7, 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "text", + "content": "[33] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[34] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3, 6, 7, 14" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 509 + ], + "type": "text", + "content": "[35] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "text", + "content": "[36] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 600 + ], + "type": "text", + "content": "[37] Baorui Ma, Huachen Gao, Haoge Deng, Zhengxiong Luo, Tiejun Huang, Lulu Tang, and Xinlong Wang. You see it, you got it: Learning 3d creation on pose-free videos at scale. arXiv preprint arXiv:2412.06699, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[38] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In CVPR, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[40] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Reg" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "nerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 128 + ], + "type": "text", + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[42] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlp's. In ICCV, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 165, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 553, + 230 + ], + "type": "text", + "content": "[43] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6121-6132, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "type": "text", + "content": "[45] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. ZeroNVS: Zero-shot 360-degree view synthesis from a single real image. In CVPR, 2024. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 375 + ], + "type": "text", + "content": "[46] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seunglyong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[48] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[49] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34: 19313-19325, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[50] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 589 + ], + "type": "text", + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 5, 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[52] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and Ameesh Makadia. Generalizable patch-based neural rendering. In European Conference on Computer Vision, pages 156-174. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[53] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3, 14" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[54] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward gener" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27803" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "alisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[55] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[56] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. arXiv preprint arXiv:2406.11819, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[57] Hanyang Wang, Fangfu Liu, Jiawei Chi, and Yueqi Duan. Videoscene: Distilling video diffusion model to generate 3d scenes in one step. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16475-16485, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[58] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[59] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[60] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[61] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[62] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf-: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "text", + "content": "[64] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH Conference, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[65] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[66] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[67] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[68] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "text", + "content": "[69] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 240 + ], + "type": "text", + "content": "[70] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 296 + ], + "type": "text", + "content": "[71] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "text", + "content": "[72] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 398 + ], + "type": "text", + "content": "[73] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043-9053, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 432 + ], + "type": "text", + "content": "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 487 + ], + "type": "text", + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 555 + ], + "type": "text", + "content": "[76] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 611 + ], + "type": "text", + "content": "[77] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 4, 5, 6, 7, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "text", + "content": "[78] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 713 + ], + "type": "text", + "content": "[79] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27804" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[80] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[81] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 238 + ], + "type": "text", + "content": "[82] Songchun Zhang, Yibo Zhang, Quan Zheng, Rui Ma, Wei Hua, Hujun Bao, Weiwei Xu, and Changqing Zou. 3d-scenedreamer: Text-driven 3d-consistent scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10170-10180, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "type": "text", + "content": "[83] Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 21936-21947, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 339 + ], + "type": "text", + "content": "[84] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 384 + ], + "type": "text", + "content": "[85] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "text", + "content": "[86] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. ACM Trans. Graph, 37, 2018. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 464 + ], + "type": "text", + "content": "[87] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snively. Stereo magnification: Learning view synthesis using multiplane images. ACM TOG, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "type": "text", + "content": "[88] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, 2024. 2, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27805" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_content_list.json b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eba1b84722884b174b5f18eac772e59feff0e335 --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_content_list.json @@ -0,0 +1,2210 @@ +[ + { + "type": "text", + "text": "SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Sheng $^{1}$ , Jiajun Deng $^{2\\dagger}$ , Xinran Zhang $^{1}$ , Yu Zhang $^{1}$ , Bei Hua $^{1}$ , Yanyong Zhang $^{1}$ , Jianmin Ji $^{1\\dagger}$ $^{1}$ University of Science and Technology of China, $^{2}$ The University of Adelaide", + "bbox": [ + 122, + 178, + 875, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shengyu724, zxrr}@mail.ustc.edu.cn, jiajun.deng@adelaide.edu.au", + "bbox": [ + 218, + 218, + 777, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangyu,bhua,yanyongz,jianmin}@ustc.edu.cn †: Corresponding Author", + "bbox": [ + 218, + 234, + 772, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3c34c3daae2b917c496310f5f0b8e4dc6853fe327dbc647ebc24d84048344a9.jpg", + "image_caption": [ + "Figure 1. Comparison between previous methods and our SpatialSplat. (a): Previous methods predict pixel-wise Gaussians, associating each primitive with compressed semantic feature. (b): Our SpatialSplat avoids generating redundant Gaussian primitives for overlapping pixels, and it represents semantics with a dual-field architecture to better preserve the information. (c): SpatialSplat outperforms the state-of-the-art method LSM [15] for both novel-view rendering quality and semantic segmentation while being more efficient." + ], + "image_footnote": [], + "bbox": [ + 109, + 290, + 405, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa58bb773a8f08e4e6a8f901883311c541a0e7032df8b973f9f276c359595d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 290, + 723, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3daf1c37c562445242a2565d5df8e75f09140300e699cd05c9308e8989f6a5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 290, + 887, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 539, + 326, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A major breakthrough in 3D reconstruction is the feedforward paradigm to generate pixel-wise 3D points or Gaussian primitives from sparse, unposed images. To further incorporate semantics while avoiding the significant memory and storage costs of high-dimensional semantic features, existing methods extend this paradigm by associating each primitive with a compressed semantic feature vector. However, these methods have two major limitations: (a) the naively compressed feature compromises expressiveness, affecting the model's ability to capture fine-grained semantics, and (b) the pixel-wise primitive prediction introduces redundancy in overlapping areas, causing unnecessary memory overhead. To this end, we introduce SpatialSplat, a feedforward framework that produces redundancy-aware Gaussians and capitalizes on a dual-field semantic representation. Particularly, with the insight that primitives within the same instance exhibit high semantic consistency, we decompose the semantic representation into a coarse feature field that encodes uncompressed semantics with minimal primitives, and a fine-grained yet low-dimensional feature field that captures detailed inter", + "bbox": [ + 86, + 583, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instance relationships. Moreover, we propose a selective Gaussian mechanism, which retains only essential Gaussians in the scene, effectively eliminating redundant primitives. Our proposed Spatialplat learns accurate semantic information and detailed instances prior with more compact 3D Gaussians, making semantic 3D reconstruction more applicable. We conduct extensive experiments to evaluate our method, demonstrating a remarkable $60\\%$ reduction in scene representation parameters while achieving superior performance over state-of-the-art methods. The code will be made available for future investigation.", + "bbox": [ + 511, + 541, + 908, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 738, + 645, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing and understanding 3D scenes from 2D images [5, 9, 32, 36, 53] is a fundamental topic in computer vision, aiming to obtain semantic-aware 3D structure from low-cost devices, i.e., RGB cameras. This technique is significant for various applications, such as robotics, autonomous driving, and VR/AR. Recently, with the emergence of the new 3D representation, i.e., Neural Radiance Fields (NeRF) [29] and 3D Gaussian Splatting (3DGS) [19], a popular paradigm of performing semantic-aware 3D re", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26404", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "construction is to distill the feature field of NeRF or 3DGS with powerful 2D vision-language foundation models [20, 31, 33, 52]. However, these methods typically rely on perscense optimization and complex multi-step preprocessing, limiting their ability to generalize across multiple scenes within a single model.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent breakthroughs [3, 7, 49] in 3D reconstruction have incorporated feed-forward networks to improve the generalization of 3DGS models and accelerate reconstruction. These methods follow a paradigm that generates pixel-wise Gaussian primitives from sparse posed images. Building on this, further efforts [38, 46] have extended the paradigm to reconstruct scenes from unposed images. Moreover, methods [15, 43] have advanced this approach by integrating semantic understanding, enabling semantic 3D reconstruction from sparse unposed images.", + "bbox": [ + 89, + 184, + 480, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant progress, these methods have two major limitations. First, as shown in Fig.1 (a), pixelwise Gaussian prediction introduces substantial redundancy in overlapping regions while providing minimal accuracy gains. Method [44] attempts to mitigate this issue by projecting primitives onto a plane and identifying overlaps, it depends on precise camera extrinsics, which are often difficult to obtain in real-world scenarios—especially when only a few sparse, textureless images are available. Second, methods [15, 31, 33, 43, 52] enforce per-primitive semantic learning to form a dense semantic feature field, as shown in Fig.1 (a). This primitive-level approach incurs significant memory and storage costs due to high-dimensional (512 or more) language features. To mitigate this, the methods often compress features into lower-dimensional representations (e.g., 64-128), but this compression leads to irreversible information loss [52], leading to suboptimal scene understanding performance.", + "bbox": [ + 91, + 337, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we take a slightly different viewpoint. First, inspired by MASt3R [25], which successfully extracts 3D geometry and pixel correspondences from images, we observe that redundant primitives often share similar geometry and appearance, allowing them to be identified directly from images without relying on complex geometric priors. Second, we find that per-primitive semantics are not essential for high-performance scene understanding. Instead, a coarse semantic representation, when combined with fine-grained instance information, is sufficient for strong performance, as Gaussian primitives within the same instance exhibit high semantic consistency.", + "bbox": [ + 89, + 611, + 480, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By consolidating this idea, we introduce SpatialSplat, a feed-forward network generating compact yet expressive semantic 3D representations from unposed images. As depicted in Fig. 1 (b), SpatialSplat introduces a dual-field architecture that decomposes the dense semantic feature field into two components: a coarse feature field encoding uncompressed instance-level semantics through mini", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mal Gaussians, and a low-dimensional fine-grained feature field capturing inter-instance relationships. Guided by multiple 2D foundation models, our method learns both accurate semantic and instance priors, achieving superior performance while reducing representation parameters. Additionally, we introduce a Selective Gaussian Mechanism (SGM) to eliminate redundancy in overlapping areas caused by pixelwise representations, along with a novel loss function that jointly optimizes redundancy-aware Gaussians and scene fidelity. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple downstream 3D tasks while using only $40\\%$ of the representation parameters required by the baseline, all without any 3D supervision. Our contributions are threefold:", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel feed-forward 3DGS framework that, to the best of our knowledge, is the first to simultaneously learn semantic and instance priors with guidance from foundation models while maintaining a high-fidelity radiance field.", + "- A dual-field semantic representation that significantly reduces storage consumption while enhancing open-vocabulary 3D segmentation performance.", + "- A mechanism that identifies redundant primitives from images without relying on any geometric priors." + ], + "bbox": [ + 511, + 303, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 455, + 653, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Feed-forward 3D Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 479, + 802, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of NeRF [29] and 3DGS [19] has significantly transformed the paradigm of 3D reconstruction from images. Despite achieving remarkable results in 3D reconstruction and novel view synthesis, these methods rely on time-consuming per-scene optimization, with even improved techniques [4, 16, 30] still requiring several minutes to hours. To overcome these limitations, recent approaches [3, 7, 11, 18, 41, 45, 48, 49] have explored using a single feed-forward network to learn reconstruction priors from large datasets, enabling generalizable 3D reconstruction. DUSt3R serials [25, 42] further simplifies the reconstruction pipeline, making 3D reconstruction possible from images without relying on camera extrinsics. This advancement has inspired a new wave of methods [13, 38, 46] capable of achieving dense reconstruction and view synthesis without camera extrinsics.", + "bbox": [ + 511, + 503, + 903, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Feature Field Distillation", + "text_level": 1, + "bbox": [ + 511, + 757, + 741, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches of feature field distillation were largely based on NeRF [29]. For example, methods [37, 51] successfully embedded 2D label data into 3D space, achieving sharp and precise 3D segmentation. Later approaches [12, 20, 22, 36, 40] incorporated more complex pixel-aligned feature vectors from technologies like LSeg [26] or DINO [2] into NeRF frameworks, enabling open-vocabulary 3D segmentation. With the rise of", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26405", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ca3c1e5b776457dda56ff2b09058f990e66f96beb2ee228f79f5ad9e4ab5796.jpg", + "image_caption": [ + "Figure 2. Pipeline of SpatialSplat. The SpatialSplat processes unposed images along with their intrinsicities through a 3D geometry transformer. The extracted features from the geometry transformer and the pretrained 2D model are then fed into separate Dense Prediction Transformer (DPT) heads to predict different Gaussian attributes, resulting in a fine-grained instance-aware radiance field $\\mathcal{F}_I$ and a coarse semantic feature field $\\mathcal{F}_S$ . This enables the synthesis of RGB and feature maps from novel viewpoints." + ], + "image_footnote": [], + "bbox": [ + 96, + 118, + 903, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS [19], methods such as [17, 31, 33, 52] have extended this concept by embedding language features into Gaussian attributes to enhance semantic understanding. Large Spatial Model (LSM) [15] was the first to achieve a generalizable semantic radiance field based on feed-forward 3DGS. However, due to the explicit nature of 3DGS representations, attaching feature vectors to each primitive incurs high memory costs. To mitigate this, existing methods often compress semantic features into lower-dimensional representations, inevitably resulting in information loss.", + "bbox": [ + 88, + 362, + 480, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our approach decomposes semantics into coarse instance-level semantics and fine-grained instance information, achieving superior performance with significantly lower storage requirements.", + "bbox": [ + 89, + 513, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Compact 3D Gaussian", + "text_level": 1, + "bbox": [ + 89, + 582, + 299, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work is also related to compact 3D Gaussian representation. Recent works have attempted to compress 3DGS by exploring the spatial relationships among Gaussians. Scaffold-GS [27] introduces anchor-centered features to reduce parameter counts, while HAC [6] builds on this by incorporating a hash grid to better organize scene primitives. Methods [14, 24] utilize techniques like distillation and Gaussian pruning to minimize storage requirements.", + "bbox": [ + 89, + 603, + 482, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, our focus is on addressing the additional representational redundancy introduced by the feed-forward pipeline (i.e., overleaping Gaussians) rather than the inherent redundancy of 3DGS itself. This issue is prevalent across existing generalizable 3DGS methods, and the aforementioned compression techniques are not suitable for these frameworks.", + "bbox": [ + 89, + 724, + 482, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 842, + 181, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. As illustrated in Fig.2, SpatialSplat leverages a standard Vision Transformer (ViT) [10] backbone with a", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "few Dense Prediction Transformer (DPT) [35] heads. Given $V$ unposed images at resolution $H\\times W$ with their intrinsic $\\{I^v,K^v\\}_{v = 1}^V$ , SpatialSplat learn a mapping $f_{\\theta}$ parametered by $\\theta$ that transforms the input images to an fine-grained instance-aware radiance field $\\mathcal{F}_I$ and a coarse semantic feature field $\\mathcal{F}_S$ , which can be formulated as:", + "bbox": [ + 511, + 363, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta}: \\left\\{\\boldsymbol {I} ^ {v}, \\boldsymbol {K} ^ {v} \\right\\} _ {v = 1} ^ {V} \\rightarrow \\mathcal {F} _ {I} + \\mathcal {F} _ {S},\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 463, + 745, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {\\boldsymbol {I}} = \\left\\{\\cup \\left(\\boldsymbol {f} _ {\\boldsymbol {I}} ^ {i}, \\boldsymbol {\\beta} ^ {i}, \\boldsymbol {\\mu} ^ {i}, \\boldsymbol {\\alpha} ^ {i}, \\boldsymbol {r} ^ {i}, \\boldsymbol {s} ^ {i}, \\boldsymbol {c} ^ {i}\\right) \\right\\} _ {i = 1, \\dots , V \\times H \\times W}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 484, + 903, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} _ {S} = \\{\\cup (\\boldsymbol {f} _ {S} ^ {j}, \\boldsymbol {\\mu} ^ {j}, \\boldsymbol {\\alpha} ^ {j}, \\boldsymbol {r} ^ {j}, \\boldsymbol {s} ^ {j}, \\boldsymbol {c} ^ {j}) \\} _ {j = 1, \\dots , V \\times \\frac {H}{S} \\times \\frac {W}{S}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 506, + 864, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\beta \\in \\mathbb{R}$ is the importance score, which will be explained in Section 3.2. $\\pmb{f}_I \\in \\mathbb{R}^N$ is the learnable instance feature and $\\pmb{f}_S \\in \\mathbb{R}^M$ is the learnable semantic feature, as detailed in Section 3.3. The remains represent the vanilla Gaussian parameters [19], including the center position $\\pmb{\\mu} \\in \\mathbb{R}^3$ , opacity $\\alpha \\in \\mathbb{R}$ , rotation factor in quaternion $\\pmb{r} \\in \\mathbb{R}^4$ , scale $\\pmb{s} \\in \\mathbb{R}^3$ , and spherical harmonics (SH) $\\pmb{c} \\in \\mathbb{R}^k$ with $k$ degrees of freedom. In the following sections, we provide a detailed explanation of each component of our method.", + "bbox": [ + 511, + 536, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Geometry Prediction", + "text_level": 1, + "bbox": [ + 511, + 681, + 736, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both the encoder and decoder in our geometric prediction module are built on pure ViT structures, requiring no geometric priors as in previous methods [3, 7]. The input image is patchified and flattened into image sequences, which along with the camera intrinsics processed by a linear layer, are fed into the encoder. The encoder weights are shared across different input views. The features from encoder are then passed to a ViT-based decoder, where cross-attention is applied to better capture spatial relationships and aggregate information across views. Finally, the features from different views are separately fed into a series of DPT heads to predict point maps and other Gaussian parameters. The point maps serve as Gaussians centers. Since the input im", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26406", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ages are unposed, the Gaussians centers suffer from scale ambiguity. Some methods [15, 38] rely on depth supervision to obtain per-scene scale, but this introduces additional errors due to imperfect depth. To overcome this limitation, we follow NoPoSplat's [46] approach by injecting camera intrinsics to resolve scale ambiguity. Experiments show that SpatialSplat effectively learns 3D priors from sparse unposed images without depth supervision, even while jointly learning multiple parameters and features.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Selective Gaussian Mechanism", + "text_level": 1, + "bbox": [ + 89, + 239, + 362, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods [3, 7, 15, 38, 46] directly use the pixelwise pointmap as Gaussian centers which introduce significant redundancy. To address this, we propose a selective Gaussian mechanism that assigns each primitive an importance score to quantify its necessity for the scene representation. Primitives with importance scores below a certain threshold are considered redundant and discarded. To effectively supervise the learning of importance scores, we analyze Gaussian properties and the rendering pipeline, designing a simple yet effective strategy. Specifically, a primitive's influence can be reduced by decreasing its opacity or size [19]. However, we observe that even primitives with near-zero size still contribute to the rendering by occupying a pixel. Therefore, we multiply the importance score to the opacity of each primitive and modify the alpha blending formulation as follows:", + "bbox": [ + 89, + 261, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\boldsymbol {\\alpha} _ {i} \\boldsymbol {\\beta} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\boldsymbol {\\alpha} _ {j} \\boldsymbol {\\beta} _ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 513, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the final intensities calculated by blending $n$ ordered Gaussians overlapping the pixels, and $\\beta_{i}$ is the importance score processed through a sigmoid activation function. This ensures that primitives with low importance have minimal impact on the final intensities. Removing important primitives leads to a significant increase in photometric loss, whereas discarding redundant ones has little effect. Therefore, we optimize $\\beta_{i}$ through photometric loss minimization. Following methods [15, 46], we define the photometric loss as a weighted sum of MSE and LPIPS [50]:", + "bbox": [ + 89, + 568, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) = \\mathcal {L} _ {M S E} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda \\mathcal {L} _ {L P I P S} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 746, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C$ and $\\hat{C}$ are rasterized and GT pixel intensities, and $\\lambda = 0.05$ . Since redundant primitives are directly discarded during inference, we encourage their importance scores to approach zero to minimize the discrepancy between training and inference. Simultaneously, we aim to reduce the number of primitives used, striving to eliminate redundant ones as much as possible. To achieve these objectives, we first design the importance score with inspiration from Leaky", + "bbox": [ + 89, + 777, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ReLU [28], formulated as follows:", + "bbox": [ + 513, + 90, + 743, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\beta} _ {i} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\beta} _ {i} & \\text {i f} \\boldsymbol {\\beta} _ {i} > \\tau \\\\ \\boldsymbol {\\beta} _ {i} \\times 1 0 ^ {- 3} & \\text {i f} \\boldsymbol {\\beta} _ {i} < \\tau \\end{array} \\right.. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 114, + 906, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This ensures that the importance scores of primitives below the threshold $\\tau$ are close to zero while preventing gradient vanishing. Additionally, we introduce a Binary Cross Entropy (BCE) loss with a regularization term that pushes $\\beta_{i}$ to either 0 or 1 while penalizing excessive use of primitives:", + "bbox": [ + 511, + 166, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} (\\boldsymbol {S}) = \\mathcal {L} _ {B C E} (\\boldsymbol {S}, \\hat {\\boldsymbol {S}}) + \\frac {1}{\\| \\boldsymbol {S} \\|} \\sum_ {\\boldsymbol {\\beta} _ {i} \\in \\boldsymbol {S}} \\beta_ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 251, + 906, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 295, + 558, + 308 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {S} = \\{\\beta_ {i} \\} _ {i = 1} ^ {H \\times W}, \\quad \\hat {\\boldsymbol {S}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\beta_ {i} > \\tau \\\\ 0 & \\text {i f} \\beta_ {i} < \\tau \\end{array} \\right..\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 311, + 856, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a result, SpatialSplat achieves a compact yet high-fidelity scene representation by selectively retaining only the most essential primitives.", + "bbox": [ + 511, + 362, + 905, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dual-field Architecture", + "text_level": 1, + "bbox": [ + 513, + 416, + 728, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While distilling 2D semantic features into dense 3D representations has proven effective for scene understanding, conventional per-primitive compressed feature assignments [31, 36, 52] result in information loss and suboptimal performance. To mitigate this loss without increasing storage costs, we propose a dual-field architecture that decouples semantic representation into: 1) a fine-grained instance-aware radiance field, capturing scene geometry, textures, and instance correspondences, and 2) a coarse, uncompressed semantic feature field that reduces storage requirements while preserving semantic accuracy.", + "bbox": [ + 511, + 436, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-grained Instance-Aware Radiance Field. Our instance-aware radiance field builds upon Vanilla 3DGS, initializing the redundancy-aware Gaussians augmented with learnable instance features $\\pmb{f}_I$ . We train the network using contrastive learning [1] to pull rendered instance features of the same instance closer while push those of different instances further apart. To achieve this, we lift multiview 2D segmentation priors into 3D to help SpatialSplat learn inter-instance relationships. Given a target image $I$ , we first use Segment Anything Model (SAM) [21] to extract a set of masks $\\mathcal{M} = \\{M^k \\mid k = 1, 2, \\dots, m\\}$ . The instance feature map rendered at target image view with Eq. 2 is denoted as $F_I$ . The contrastive loss is then defined as:", + "bbox": [ + 511, + 606, + 908, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C} \\left(\\boldsymbol {F} _ {I}, \\mathcal {M}\\right) = - \\log \\frac {1}{\\| \\boldsymbol {I} \\| ^ {2}} \\sum_ {u \\in \\boldsymbol {I}} \\sum_ {u ^ {\\prime} \\in \\boldsymbol {I} ^ {+}} \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right) \\\\ + \\sum_ {u ^ {\\prime} \\notin I ^ {+}} \\left(1 - \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right)\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 811, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/edc322fb8c5cf97cc7b9e8d09f8efe65e247459ed11f28c3282d8af028f9dec2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSemanticForwardSource ViewTarget View
mIoU ↑Acc. ↑mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
L-Seg [26]✓X0.55410.78240.55580.7856N/AN/AN/A
NeRF-DFF [23]✓X0.53810.77240.51370.758222.490.76500.2829
Feature-3DGS [52]✓X0.49920.73620.32230.567417.960.58120.4894
NoPoSplat [46]X✓N/AN/AN/AN/A25.700.81590.1875
LSM [15]✓✓0.51410.77190.51040.766224.120.79610.2526
SpatialSplat-Lite✓✓0.52720.78010.52650.769325.450.80320.2039
SpatialSplat✓✓0.55930.78140.55870.792425.460.80450.2046
", + "bbox": [ + 94, + 99, + 903, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparison in 3D Tasks on Scannet dataset. Our method outperforms both the latest SOTA semantic-aware feed-forward approach and per-scene optimization methods. \"Lite\": replacing the pretrained 2D model LSeg with CLIP ViT-B/16.", + "bbox": [ + 89, + 277, + 906, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I^{+}$ is the set of pixels that belong to the same instance mask, $F_{I}(u)$ is the feature vector of the $F_{I}$ at coordinate $u$ and $sim(. )$ is the cosine similarity. With this weak supervision, SpatialSplat effectively clusters radiance primitives by instance affiliation. Notably, the loss complexity in Eq. 6 scales quadratically with the number of pixels, making training infeasible. To address this, we employ a trick (detailed in the appendix) that enables efficient loss estimation with linear complexity.", + "bbox": [ + 88, + 320, + 480, + 457 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse Semantic Feature Field. Since primitives within the same instance share semantic consistency, SpatialSplat learns instance-level semantics using only a subset of primitives, avoiding the redundancy of per-primitive assignments. Manually selecting primitives for each instance, however, disrupts the network's cohesion and simplicity while increasing additional computational costs. To address this, we uniformly downsample the Gaussian centers from $\\mathcal{F}_I$ to form a coarse feature field, where each selected primitive is assigned a semantic feature $f_{S}$ . To enhance semantic learning, we inject image features from a pretrained 2D model into feature heads, as shown in Fig. 2. A 2D semantic feature map $F_{S}$ is rendered using alpha blending. We minimize the loss between the rendered feature map at a novel view and the feature map $\\hat{F}_{S}$ of the ground truth image extracted from the pretrained 2D model:", + "bbox": [ + 91, + 459, + 483, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) = M S E \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 708, + 482, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During open-vocabulary querying, we first cluster primitives in $\\mathcal{F}_I$ by instance feature similarity, assigning each instance the average feature of its clustered primitives as its label. For primitives in $\\mathcal{F}_S$ , their labels correspond to the instance features before sampling, and their semantic features are assigned to the instance with the highest cosine similarity. If multiple primitives contribute semantic features to the same instance, we take their average to ensure consistency. This approach enables the network to acquire sharper and clearer semantics compared to the previous methods. An additional advantage of dual-field architecture is that, un", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "like dense semantic supervision from LSeg [26], it allows the use of a much lighter pretrained model (e.g., CLIP ViT-B/16 [34]), significantly improving inference speed.", + "bbox": [ + 511, + 321, + 906, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objective", + "text_level": 1, + "bbox": [ + 511, + 375, + 694, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous methods [15], we perform end-to-end training to optimize our model with the following objective:", + "bbox": [ + 511, + 397, + 906, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda_ {1} \\mathcal {L} _ {I} (\\boldsymbol {S}) \\tag {8} \\\\ + \\lambda_ {2} \\mathcal {L} _ {C} (\\boldsymbol {F} _ {I}, \\mathcal {M}) + \\lambda_ {3} \\mathcal {L} _ {S} (\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 450, + 903, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The parameters $\\lambda_1, \\lambda_2, \\lambda_3$ are set to 0.01, 0.2, and 1.0, respectively. To save training time, the instance masks $\\mathcal{M}$ are generated by SAM [21] prior to training, while the semantic feature map $\\hat{\\pmb{F}}_S$ is predicted during training.", + "bbox": [ + 511, + 500, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 574, + 638, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 599, + 702, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following LSM[15], we primarily train our model on a combination of ScanNet++[47] and ScanNet [8]. We filter out bad scenes and those with incomplete extrinsic parameters, resulting in a training dataset of approximately 1,500 scenes. For evaluation, we follow LSM and select 40 unseen scenes from ScanNet to assess our model's performance. We test our method's cross-dataset generalization on the synthesized Replica Dataset [39].", + "bbox": [ + 511, + 623, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate our method on two downstream tasks: novel view synthesis (NVS) and 3D open-vocabulary segmentation (OVS). For NVS, we adopt the standard metrics: PSNR, SSIM, and LPIPS [50]. For OVS, we evaluate performance using class-wise intersection over union (mIoU) and average pixel accuracy (mAcc).", + "bbox": [ + 511, + 747, + 905, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We primarily compare our method with LSM [15], the latest state-of-the-art (SOTA) approach for generalizable semantic 3D reconstruction. Additionally, we include the following baselines: 1): Feature-3DGS [52]", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26408", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d60ac0a66109e44e4515ecaa35ad6ec3713a7e144d6a2720f2d920be15de4595.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison in NVs. SpatialSplat can synthesize realistic novel views. In challenging cases where LSM fails, such as the table legs in the first two rows and the corners in the last two rows, our method achieves significantly better results." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 890, + 404 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1eb7ff0960d9899eb4dd6f8b09c2d0b8814762289a2e596c15bd2162d407e0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SceneLSMSpatialSplat (Ours)
mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
room00.50240.742317.610.52470.32280.51310.793422.090.67600.1794
room10.46610.734914.340.59130.31740.46910.762820.620.67740.2265
office30.54840.762721.280.75670.20840.55820.799722.520.79130.1436
office40.50650.772319.510.69620.27790.50160.769521.130.72710.2072
", + "bbox": [ + 94, + 448, + 901, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Out-of-distribution (OOD) comparison on Replica dataset. SpatialSplat generalizes well on OOD data.", + "bbox": [ + 155, + 569, + 839, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and NeRF-DFF [23], pre-scene optimization methods for semantic 3D reconstruction based on 3DGS [19] and NeRF [29], respectively. 2): NoPoSplat [46], a latest SOTA method for generalizable novel view synthesis. 3): LSeg [26], a 2D open-vocabulary segmenter used for feature distillation in all compared semantic-aware methods. Unless otherwise specified, our method also uses image features from LSeg for supervision.", + "bbox": [ + 88, + 606, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For the 3D geometry prediction module, we use ViT-Large with a patch size of 16 as the encoder and ViT-Base as the decoder, both initialized with pre-trained weights from MAST3R [25], while the remaining modules are randomly initialized. We modify the 3DGS CUDA renderer to support instance and semantic feature map rendering. The importance score threshold $\\tau$ is set to 0.5, the downsampling ratio to 8, the instance feature dimension $N$ to 8, and the semantic feature dimension $M$ to 512. For a fair comparison, we train the model at a resolution of $256 \\times 256$ , consistent with our baseline.", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 606, + 709, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results of Novel View Synthesis. Following previous methods [3, 46], we select two images from each test scene as input view for the generalizable 3DGS methods, and choose four additional images located between these two as target views to evaluate the performance of different methods in Novel View Synthesis. Notably, Feature-3DGS [52] and NeRF-DFF [23] relies on dense view inputs and performs poorly with only two input images. To account for this, we use five out of the six available images (the two input images and four target views) for training and evaluate on the remaining image. As shown in Tab. 1, Spatial-Splat significantly outperforms latest SOTA method LSM. We observe that LSM struggles in certain scenes due to its reliance on accurate depth for aligning views during training as shown in Fig. 3. Although NeRF-DFF uses more images for training, our method still outperforms it and even demonstrates competitive performance with very re", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26409", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7adbd88e415b1782e08ad32a8f14d479bbf0e68e4557668c63e7be70bf2cdb3c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 97, + 220, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05e630337a72659c63d80ba405039ffdadc056ee42a3674e1f6c227a7a9f9c87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 183, + 218, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d9d4de1b8eee7075564d3559b1785e0fd6e4d528152a49ca518944d74f94ea9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 111, + 268, + 218, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae5f6f914002c52e64e75b90b61acc2934abc0d60f8a7690b403fab0f51f859b.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 111, + 354, + 218, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf0987a70ae53850481e6c4f6079568bcd111d97a8681093799a04a705c6b051.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 97, + 331, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eaada965a6447fc2a4be6c2ab36021da83df6ca5cd3dbc9d73d9d77c366d4948.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 183, + 331, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71c786ad40dd8fb57717e468be0b87160b5de383244522178ee88a1a87edcec0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 268, + 330, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/948d722f0c36bac7ae8f756dbe36c22ccfc3cc9b319eee1236c4b343bc1d83f0.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 223, + 354, + 330, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/742965fb04db8c4e71e45147e48e6cfcd01df17995266a4f2fe17efae9d9d0f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 97, + 441, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d80f53b12b2cc57c0b9203b63a158f794477d99d1049859b7fbeb81a95ef3819.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 183, + 441, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/acd795a1bd846dbabab26ab7f40a7c5a036c03a7cef9d7c00013e034f0e0b4f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 268, + 441, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dce9bd8eb625ce66fb3b054a307ddd5ba2fff89f6edb21c3976df8c37fef1db2.jpg", + "image_caption": [ + "LSeg" + ], + "image_footnote": [], + "bbox": [ + 334, + 354, + 441, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3000a16cb5513a700e503069b235891466071606b3593991d0bcb6f58bc898d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 97, + 553, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f485edf0cceea1ee0e27e8b57302d15cd81e1a2bcf7efc6742eeb8565a95c60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 183, + 552, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/274a31ba1ceb5e58489c73bffb190e186df071e66432d531f9ca9b953f3823f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 268, + 550, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63f52a2b804dd52190abe7cabb11316e5efcc89e069ffe9b9d0f018234706f19.jpg", + "image_caption": [ + "NeRF-DFF" + ], + "image_footnote": [], + "bbox": [ + 444, + 354, + 550, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce5c16875e7b76a252830491b175905df469c89b4fdd964848c36c53e0960686.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 97, + 663, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb1643c48ce1459c8892a458436df37095e1d51d9c732f5c82cef36ba45651d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 183, + 663, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/baa599d73b8f42a465999ae40dfd6460d1c080b39451c50b054c0975cf77e3be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 268, + 663, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51912050a7bbb4bf759cfa00207049a87b4220d627932c74d7c813ae5993ce47.jpg", + "image_caption": [ + "Feature-3DGS" + ], + "image_footnote": [], + "bbox": [ + 555, + 354, + 663, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/868f9008bd2025ac29fac86dc02c73e67974bf29ed731141da7fb3afda7e2e2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 97, + 774, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b2e79df87e59878db9648cf460cc0cb3e14764839fc38daca1142d92c0338ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 183, + 774, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f9e250c5ef6b67abd6d4da20aae36aafdb1423788bf0df45d12cbb395b7044f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 268, + 774, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c95b2d84afba7b2d5c8e40b028fb0b906e8daf9d397aaf43163714ad46b34c22.jpg", + "image_caption": [ + "LSM" + ], + "image_footnote": [], + "bbox": [ + 666, + 354, + 774, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ea02be0c05163d27521690c3739ce0aa84102404d904d5e5598e988d0159a0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 97, + 885, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e6e71891dfe7d144520c276881144b2b5e88fd7eb482ceab9b96dbcb6f6b10b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 183, + 885, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7faabe2266a7e0afd7c2a90f5fe6294431916025fe68be408d11b7b726c6de34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 268, + 885, + 352 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b29f3d03b5c4681e1d810b9481cc8b132d72bcac1ed8cfb348ee3dd7b4f3425b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 777, + 354, + 885, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Wall", + "bbox": [ + 124, + 460, + 158, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Floor", + "bbox": [ + 220, + 460, + 256, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ceiling", + "bbox": [ + 316, + 460, + 364, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Chair", + "bbox": [ + 424, + 460, + 462, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table", + "bbox": [ + 527, + 460, + 565, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sofa", + "bbox": [ + 633, + 460, + 665, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bed", + "bbox": [ + 738, + 460, + 766, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Others", + "bbox": [ + 831, + 460, + 879, + 472 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca7c174c9fb3d07f184ee41347586bbfa6f39dd3739b5cb98f78dd63e782f838.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison in OVS. SpatialSplat achieves sharper and more precise segmentation results compared to previous methods. Notably, our method excels in challenging details, such as distinguishing table legs from cabinet legs (the second the fourth rows), benefiting from our dual-field architecture that captures detailed instance information and uncompressed semantics." + ], + "image_footnote": [], + "bbox": [ + 106, + 551, + 194, + 619 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c113e95933606ccd2d4127d5f76d5e69bd0790d38a029607e350be288b2a603.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 553, + 287, + 619 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c16f72faef771f03962282832f99cd57bca07de508a8d58e73f5c79a9c14ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 553, + 375, + 619 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/36b23c56dbe38b8e8418290f0857dfa2893980787e123ee5de8ea9c1ca7da502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 553, + 467, + 619 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d9c75b686eb848c6d1f20f9f86827ce070e1b65a8e58fd5c2fa70854aa21a5a.jpg", + "image_caption": [ + "Input views" + ], + "image_footnote": [], + "bbox": [ + 107, + 625, + 192, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5b380a8713197993c942e565f6164fb9dbfaf9aa29a9de7efbe5f0ff34b47ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 625, + 287, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af49e80d9c25a03f76b1186f7fc03f6a30b46ffe7a3d842092e4957ef3899a6b.jpg", + "image_caption": [ + "Rendered Feature Field" + ], + "image_footnote": [], + "bbox": [ + 292, + 625, + 375, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/890a7fff8cd638e4343a3b08e0e7f567dfe2c929ce548d82bdd6ec3235964f98.jpg", + "image_caption": [ + "Figure 5. The rendered instance features. SpatialSplat predicts clear and consistent instance features across different views." + ], + "image_footnote": [], + "bbox": [ + 379, + 625, + 467, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cent SOTA methods designed specifically for novel view synthesis. This highlights that our method efficiently learns spatial priors from large-scale 2D data.", + "bbox": [ + 89, + 760, + 482, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Open-vocabulary 3D Segmentation. Following LSM's approach, we map category labels from the Scannet dataset into common categories: Wall, Floor, Ceiling, Chair, Table, Bed, Sofa, Others. As shown in Tab. 1, SpatialSplat outperforms all compared methods, even surpassing L-Seg, which provides semantic feature supervision", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecfc7f0f9a6cfff07e5803e25f02598567cec7a8cb7f954f217a8ffb4bd89b9f.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 529, + 551, + 599, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1d69e9a6c208387cf34ac932f4e7fa58a70fcf7d1d91d5a8403b122ed7fbbe6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 553, + 669, + 606 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90a840d68ed14cab321e553e3778c2fd63aaaf9c077ecc1f233cdac6d598b033.jpg", + "image_caption": [ + "LSM" + ], + "image_footnote": [], + "bbox": [ + 602, + 608, + 669, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ba58f3ba2c4146e33b6eed91c252afbea4f2252342e35c0455dc3f894ac5970.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 553, + 743, + 607 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4d8cc431ce151221ac3b1a628fe788e828f2068225a9186128bf10edb55c251.jpg", + "image_caption": [ + "Ours", + "Figure 6. Qualitative results of cross-dataset generalization. Zoom out for clearer visualization." + ], + "image_footnote": [], + "bbox": [ + 673, + 609, + 743, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1e3b597510dec67cf74c1e6234c02f36267ef681a17af71fd4bfa63563ddddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 553, + 816, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4db8a01d7bd9e668291b666fe07a180868b7690153095e7398b2b40dbd053a4a.jpg", + "image_caption": [ + "LSM" + ], + "image_footnote": [], + "bbox": [ + 746, + 609, + 816, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0235ce2885724e6deaa21ddcc12c2db85541c222f452b0ac90a3a66afd6d6ffa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 553, + 888, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57d4e09f8473e4a30f7c46cccbeb00cf64d159431fde016c8f913692c7a664c2.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 820, + 608, + 888, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for other compared methods. As illustrated in Fig. 4, SpatialSplat produces sharp and clear 3D semantic segmentation results. This is achieved by lifting both 2D instance and uncompressed semantic into 3D space, enabling more efficient semantic learning while mitigating the blurred contours and information loss associated with per-primitive compressed semantics. The visualizations in Fig. 5 show that SpatialSplat effectively learns a highly consistent instance prior with only weak supervision from 2D masks.", + "bbox": [ + 511, + 731, + 906, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results of Cross-Dataset Generalization. As shown in Tab. 2, our method significantly outperforms LSM on the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26410", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/110b60505e6c0e8e2a5d73b2d2d68a13dc0c645fb04d1e476700ccd03eb96b8d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 107, + 102, + 194, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c17300b3b9cedf4b8fe5829d22b52477b3c2a1de4159f1a88d42633e140fce23.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 197, + 102, + 285, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c902820742d14d40e0584a5dc351de0673b65c7c299421fcf585c3df1868331f.jpg", + "image_caption": [ + "(c)", + "Figure 7. Qualitative results of ablations. (a) and (b) Qualitative results of importance score prediction, with red color indicating an importance score of 1 and blue indicating 0. (c) The rendered semantic feature without the dual-field. (d) The rendered semantic feature with the dual-field, which appears clearer and sharper." + ], + "image_footnote": [], + "bbox": [ + 290, + 102, + 375, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40a6872b75f51c93a8078623b49001509f0c67302839ef6653f4809d65606e28.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 380, + 102, + 467, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/45d1347a0067c604f6fae59c36eecb2e43f00ca92c5db8ad09743c833a6c6864.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLatency↓Gaussian Size ↓Num. ↓
Feature-3DGS [52]1069 s229.65 MB476K
NoPoSplat [46]0.044 s30.93 MB131K
LSM [15]0.104 s63.00 MB131K
SpatialSplat-Lite0.058 s25.58 MB87.7K
SpatialSplat0.071 s25.58 MB87.7K
", + "bbox": [ + 94, + 267, + 478, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "unseen Replica dataset in both novel view synthesis and open-vocabulary segmentation. Notably, since Replica is a synthetic dataset with a different data modality from our training set, this underscores the strong generalization ability of our approach. The visualization in Fig. 6 further demonstrates that LSM struggles with coarse modeling at scene edges, leading to artifacts in synthesized views, whereas our method generates photorealistic renderings.", + "bbox": [ + 88, + 409, + 482, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Model Efficiency. We compare the model efficiency of different 3DGS-based methods. As shown in Tab. 3, our method is faster than LSM thanks to its streamlined architecture, requiring only $40\\%$ of the storage size and $65\\%$ of primitive number. Remarkably, our approach requires even fewer Gaussians parameters than NoPoSplat, which lacks semantic awareness. This demonstrates the effectiveness of our proposed dual-field architecture and selective Gaussian mechanism, which capture accurate semantics while eliminating redundant Gaussians. Furthermore, as our method does not rely on dense semantic supervision, we leverage a lightweight pretrained 2D model, significantly accelerating inference speed.", + "bbox": [ + 89, + 532, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations and Analysis.", + "text_level": 1, + "bbox": [ + 89, + 739, + 307, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform ablations to answer the following questions: (1) Are the primitives removed by our selective Gaussian mechanism truly redundant? (2) Does the improved scene understanding stem from the proposed dual-field architecture? (3) As the number of views increases, how does redundancy grow? Can our method effectively alleviate it?", + "bbox": [ + 89, + 761, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of selective Gaussian mechanism. To answer Question 1, we remove the SGM and retrain our model using the same configuration. As shown in Tab. 4, the SGM", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/64655833e635cc077102c02b0be24c5b524147e6f5a8eaa97f1f4f1198ecf635.jpg", + "table_caption": [ + "Table 3. Model efficiency comparison. \"Lite\": the model replaces LSeg with CLIP ViT-B/16." + ], + "table_footnote": [], + "table_body": "
ModelmIoU↑Acc.↑PSNR↑Num. ↓
No SGM0.55690.797225.55131k
No dual0.50270.778625.4687.6k
3-view0.51250.779324.9297.5k
SpatialSplat0.55770.792025.4787.7k
", + "bbox": [ + 519, + 89, + 901, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on design choices. The proposed Adaptive Gaussian mechanism and dual-field architecture jointly enable the construction of high-quality representations with significantly reduced dimensionality.", + "bbox": [ + 511, + 184, + 903, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eliminates approximately $35\\%$ of Gaussian primitives while causing only a 0.08 drop in PSNR, indicating that the removed primitives have minimal impact on rendering quality. The visualization in Fig. 7 (a) and (b) further demonstrates that the SGM accurately localizes overlapping areas in the input images.", + "bbox": [ + 511, + 257, + 903, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of dual-field architecture. To address Question 2, we replace our dual-field architecture with single-level per-primitive semantic learning. Following the approach in LSM, we set the semantic feature dimension to 64 and use a convolutional layer to expand it to 512. As shown in Tab. 4, this results in a significant drop of $5\\%$ in mIoU for open-vocabulary segmentation performance. The primary issue is that per-primitive semantic learning struggles to maintain accurate semantics and fails to preserve clear instance boundaries, as illustrated in Fig. 7 (c) and (d).", + "bbox": [ + 511, + 352, + 905, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extend to more views. Our experiments primarily focus on two-view settings to ensure a fair comparison with our baseline. However, our method can be extended to more views for reconstructing larger scenes. As shown in Tab. 4, our approach remains effective even with three input views. Moreover, compared to pixel-wise methods, which require approximately 197K primitives to represent the scene, our method achieves a more compact representation by reducing about $50\\%$ primitives. More qualitative details can be found in the appendix.", + "bbox": [ + 511, + 507, + 906, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present an efficient feed-forward network that generates compact yet high-performance representations for 3D reconstruction and scene understanding from sparse, unposed images. Our proposed dual-field architecture achieves state-of-the-art performance while significantly reducing the number of representation parameters. Additionally, we introduce a Selective Gaussian Mechanism that directly identifies redundant Gaussians caused by pixel-wise predictions without requiring geometric priors. SpatialSplat achieves impressive results in 3D reconstruction and scene understanding without relying on any 3D data during training or inference, making semantic 3D reconstruction from sparse, unposed images more practical.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26411", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by Hunan Province Major Scientific and Technological Project No. 2024QK2001, National Key R&D Program of China No. 2023YFB4704500.", + "bbox": [ + 89, + 113, + 485, + 170 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 199, + 187, + 214 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yash Bhalgat, Iro Laina, João F. Henriques, Andrea Vedaldi, and Andrew Zisserman. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9630-9640. IEEE, 2021. 2", + "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19457-19467. IEEE, 2024. 2, 3, 4, 6", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2", + "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. CoRR, abs/2401.03890, 2024. 1", + "[6] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: hash-grid assisted context for 3d gaussian splatting compression. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part VII, pages 422-438. Springer, 2024. 3", + "[7] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXI, pages 370-386. Springer, 2024. 2, 3, 4", + "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5", + "[9] Jianmei Dai, Zhilong Zhang, Shiwen Mao, and Danpu Liu. A view synthesis-based $360^{\\circ}$ VR caching system over mec-enabled C-RAN. IEEE Trans. Circuits Syst. Video Technol., 30(10):3843-3855, 2020. 1" + ], + "bbox": [ + 99, + 223, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3", + "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4970-4980. IEEE, 2023. 2", + "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2", + "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. CoRR, abs/2403.20309, 2024. 2", + "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ FPS. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 3", + "[15] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, Boris Ivanovic, and Marco Pavone. Large spatial model: End-to-end unposed images to semantic 3d. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 1, 2, 3, 4, 5, 8", + "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 12479-12488. IEEE, 2023. 2", + "[17] Qiao Gu, Zhaoyang Lv, Duncan P. Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIII, pages 382-400. Springer, 2024. 3", + "[18] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18344-18347. IEEE, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26412", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 4, 6", + "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19672-19682. IEEE, 2023. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3992-4003. IEEE, 2023. 4, 5", + "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 2", + "[23] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 5, 6", + "[24] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21719-21728. IEEE, 2024. 3", + "[25] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 6", + "[26] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 5, 6", + "[27] Luca De Luigi, Damiano Bolognini, Federico Domeniconi, Daniele De Gregorio, Matteo Poggi, and Luigi Di Stefano. Scannerf: a scalable benchmark for neural radiance fields. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 816-825. IEEE, 2023. 3", + "[28] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 4", + "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2022. 1, 2, 6", + "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[31] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20051-20060. IEEE, 2024. 2, 3, 4", + "[32] Ri-Zhao Qiu, Yafei Hu, Ge Yang, Yuchen Song, Yang Fu, Jianglong Ye, Jiteng Mu, Ruihan Yang, Nikolay Atanasov, Sebastian A. Scherer, and Xiaolong Wang. Learning generalizable feature fields for mobile manipulation. CoRR, abs/2403.07563, 2024. 1", + "[33] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLI, pages 368-383. Springer, 2024. 2, 3", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 5", + "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 12159-12168. IEEE, 2021. 3", + "[36] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In Conference on Robot Learning, CoRL 2023, 6-9 November 2023, Atlanta, GA, USA, pages 405-424. PMLR, 2023. 1, 2, 4", + "[37] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 9043-9052. IEEE, 2023. 2", + "[38] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2, 4, 3", + "[39] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26413", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In International Conference on 3D Vision, 3DV 2022, Prague, Czech Republic, September 12-16, 2022, pages 443-453. IEEE, 2022. 2", + "[41] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4690-4699. Computer Vision Foundation / IEEE, 2021. 2", + "[42] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2", + "[43] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs. CoRR, abs/2412.16932, 2024. 2", + "[44] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2, 1", + "[45] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20041-20050. IEEE, 2024. 2", + "[46] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4, 5, 6, 8, 3", + "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 5", + "[48] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4578-4587. Computer Vision Foundation / IEEE, 2021. 2", + "[49] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. CoRR, abs/2408.13770, 2024. 2", + "[50] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 5", + "[51] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15818-15827. IEEE, 2021. 2", + "[52] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21676-21685. IEEE, 2024. 2, 3, 4, 5, 6, 8", + "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 1" + ], + "bbox": [ + 516, + 92, + 903, + 357 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26414", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_model.json b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..76e21ac5649f5998ae8130ab8d6c94df24b9d6c3 --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_model.json @@ -0,0 +1,2884 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.179, + 0.877, + 0.217 + ], + "angle": 0, + "content": "Yu Sheng\\(^{1}\\), Jiajun Deng\\(^{2\\dagger}\\), Xinran Zhang\\(^{1}\\), Yu Zhang\\(^{1}\\), Bei Hua\\(^{1}\\), Yanyong Zhang\\(^{1}\\), Jianmin Ji\\(^{1\\dagger}\\) \n\\(^{1}\\)University of Science and Technology of China, \\(^{2}\\)The University of Adelaide" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.219, + 0.778, + 0.234 + ], + "angle": 0, + "content": "{shengyu724, zxrr}@mail.ustc.edu.cn, jiajun.deng@adelaide.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.236, + 0.774, + 0.252 + ], + "angle": 0, + "content": "{zhangyu,bhua,yanyongz,jianmin}@ustc.edu.cn †: Corresponding Author" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.291, + 0.406, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.291, + 0.725, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.291, + 0.888, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.908, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Comparison between previous methods and our SpatialSplat. (a): Previous methods predict pixel-wise Gaussians, associating each primitive with compressed semantic feature. (b): Our SpatialSplat avoids generating redundant Gaussian primitives for overlapping pixels, and it represents semantics with a dual-field architecture to better preserve the information. (c): SpatialSplat outperforms the state-of-the-art method LSM [15] for both novel-view rendering quality and semantic segmentation while being more efficient." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.54, + 0.327, + 0.556 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "A major breakthrough in 3D reconstruction is the feedforward paradigm to generate pixel-wise 3D points or Gaussian primitives from sparse, unposed images. To further incorporate semantics while avoiding the significant memory and storage costs of high-dimensional semantic features, existing methods extend this paradigm by associating each primitive with a compressed semantic feature vector. However, these methods have two major limitations: (a) the naively compressed feature compromises expressiveness, affecting the model's ability to capture fine-grained semantics, and (b) the pixel-wise primitive prediction introduces redundancy in overlapping areas, causing unnecessary memory overhead. To this end, we introduce SpatialSplat, a feedforward framework that produces redundancy-aware Gaussians and capitalizes on a dual-field semantic representation. Particularly, with the insight that primitives within the same instance exhibit high semantic consistency, we decompose the semantic representation into a coarse feature field that encodes uncompressed semantics with minimal primitives, and a fine-grained yet low-dimensional feature field that captures detailed inter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.709 + ], + "angle": 0, + "content": "instance relationships. Moreover, we propose a selective Gaussian mechanism, which retains only essential Gaussians in the scene, effectively eliminating redundant primitives. Our proposed Spatialplat learns accurate semantic information and detailed instances prior with more compact 3D Gaussians, making semantic 3D reconstruction more applicable. We conduct extensive experiments to evaluate our method, demonstrating a remarkable \\(60\\%\\) reduction in scene representation parameters while achieving superior performance over state-of-the-art methods. The code will be made available for future investigation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.739, + 0.646, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reconstructing and understanding 3D scenes from 2D images [5, 9, 32, 36, 53] is a fundamental topic in computer vision, aiming to obtain semantic-aware 3D structure from low-cost devices, i.e., RGB cameras. This technique is significant for various applications, such as robotics, autonomous driving, and VR/AR. Recently, with the emergence of the new 3D representation, i.e., Neural Radiance Fields (NeRF) [29] and 3D Gaussian Splatting (3DGS) [19], a popular paradigm of performing semantic-aware 3D re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "construction is to distill the feature field of NeRF or 3DGS with powerful 2D vision-language foundation models [20, 31, 33, 52]. However, these methods typically rely on perscense optimization and complex multi-step preprocessing, limiting their ability to generalize across multiple scenes within a single model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.482, + 0.336 + ], + "angle": 0, + "content": "Recent breakthroughs [3, 7, 49] in 3D reconstruction have incorporated feed-forward networks to improve the generalization of 3DGS models and accelerate reconstruction. These methods follow a paradigm that generates pixel-wise Gaussian primitives from sparse posed images. Building on this, further efforts [38, 46] have extended the paradigm to reconstruct scenes from unposed images. Moreover, methods [15, 43] have advanced this approach by integrating semantic understanding, enabling semantic 3D reconstruction from sparse unposed images." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Despite significant progress, these methods have two major limitations. First, as shown in Fig.1 (a), pixelwise Gaussian prediction introduces substantial redundancy in overlapping regions while providing minimal accuracy gains. Method [44] attempts to mitigate this issue by projecting primitives onto a plane and identifying overlaps, it depends on precise camera extrinsics, which are often difficult to obtain in real-world scenarios—especially when only a few sparse, textureless images are available. Second, methods [15, 31, 33, 43, 52] enforce per-primitive semantic learning to form a dense semantic feature field, as shown in Fig.1 (a). This primitive-level approach incurs significant memory and storage costs due to high-dimensional (512 or more) language features. To mitigate this, the methods often compress features into lower-dimensional representations (e.g., 64-128), but this compression leads to irreversible information loss [52], leading to suboptimal scene understanding performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.482, + 0.793 + ], + "angle": 0, + "content": "In this paper, we take a slightly different viewpoint. First, inspired by MASt3R [25], which successfully extracts 3D geometry and pixel correspondences from images, we observe that redundant primitives often share similar geometry and appearance, allowing them to be identified directly from images without relying on complex geometric priors. Second, we find that per-primitive semantics are not essential for high-performance scene understanding. Instead, a coarse semantic representation, when combined with fine-grained instance information, is sufficient for strong performance, as Gaussian primitives within the same instance exhibit high semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "By consolidating this idea, we introduce SpatialSplat, a feed-forward network generating compact yet expressive semantic 3D representations from unposed images. As depicted in Fig. 1 (b), SpatialSplat introduces a dual-field architecture that decomposes the dense semantic feature field into two components: a coarse feature field encoding uncompressed instance-level semantics through mini" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "mal Gaussians, and a low-dimensional fine-grained feature field capturing inter-instance relationships. Guided by multiple 2D foundation models, our method learns both accurate semantic and instance priors, achieving superior performance while reducing representation parameters. Additionally, we introduce a Selective Gaussian Mechanism (SGM) to eliminate redundancy in overlapping areas caused by pixelwise representations, along with a novel loss function that jointly optimizes redundancy-aware Gaussians and scene fidelity. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple downstream 3D tasks while using only \\(40\\%\\) of the representation parameters required by the baseline, all without any 3D supervision. Our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.304, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- A novel feed-forward 3DGS framework that, to the best of our knowledge, is the first to simultaneously learn semantic and instance priors with guidance from foundation models while maintaining a high-fidelity radiance field." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.904, + 0.409 + ], + "angle": 0, + "content": "- A dual-field semantic representation that significantly reduces storage consumption while enhancing open-vocabulary 3D segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.904, + 0.44 + ], + "angle": 0, + "content": "- A mechanism that identifies redundant primitives from images without relying on any geometric priors." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.304, + 0.905, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.456, + 0.655, + 0.471 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.803, + 0.496 + ], + "angle": 0, + "content": "2.1. Feed-forward 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.905, + 0.744 + ], + "angle": 0, + "content": "The emergence of NeRF [29] and 3DGS [19] has significantly transformed the paradigm of 3D reconstruction from images. Despite achieving remarkable results in 3D reconstruction and novel view synthesis, these methods rely on time-consuming per-scene optimization, with even improved techniques [4, 16, 30] still requiring several minutes to hours. To overcome these limitations, recent approaches [3, 7, 11, 18, 41, 45, 48, 49] have explored using a single feed-forward network to learn reconstruction priors from large datasets, enabling generalizable 3D reconstruction. DUSt3R serials [25, 42] further simplifies the reconstruction pipeline, making 3D reconstruction possible from images without relying on camera extrinsics. This advancement has inspired a new wave of methods [13, 38, 46] capable of achieving dense reconstruction and view synthesis without camera extrinsics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.742, + 0.772 + ], + "angle": 0, + "content": "2.2. Feature Field Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Early approaches of feature field distillation were largely based on NeRF [29]. For example, methods [37, 51] successfully embedded 2D label data into 3D space, achieving sharp and precise 3D segmentation. Later approaches [12, 20, 22, 36, 40] incorporated more complex pixel-aligned feature vectors from technologies like LSeg [26] or DINO [2] into NeRF frameworks, enabling open-vocabulary 3D segmentation. With the rise of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26405" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.119, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.3, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Pipeline of SpatialSplat. The SpatialSplat processes unposed images along with their intrinsicities through a 3D geometry transformer. The extracted features from the geometry transformer and the pretrained 2D model are then fed into separate Dense Prediction Transformer (DPT) heads to predict different Gaussian attributes, resulting in a fine-grained instance-aware radiance field \\(\\mathcal{F}_I\\) and a coarse semantic feature field \\(\\mathcal{F}_S\\). This enables the synthesis of RGB and feature maps from novel viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.482, + 0.514 + ], + "angle": 0, + "content": "3DGS [19], methods such as [17, 31, 33, 52] have extended this concept by embedding language features into Gaussian attributes to enhance semantic understanding. Large Spatial Model (LSM) [15] was the first to achieve a generalizable semantic radiance field based on feed-forward 3DGS. However, due to the explicit nature of 3DGS representations, attaching feature vectors to each primitive incurs high memory costs. To mitigate this, existing methods often compress semantic features into lower-dimensional representations, inevitably resulting in information loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.576 + ], + "angle": 0, + "content": "In contrast, our approach decomposes semantics into coarse instance-level semantics and fine-grained instance information, achieving superior performance with significantly lower storage requirements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.3, + 0.599 + ], + "angle": 0, + "content": "2.3. Compact 3D Gaussian" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Our work is also related to compact 3D Gaussian representation. Recent works have attempted to compress 3DGS by exploring the spatial relationships among Gaussians. Scaffold-GS [27] introduces anchor-centered features to reduce parameter counts, while HAC [6] builds on this by incorporating a hash grid to better organize scene primitives. Methods [14, 24] utilize techniques like distillation and Gaussian pruning to minimize storage requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.83 + ], + "angle": 0, + "content": "Notably, our focus is on addressing the additional representational redundancy introduced by the feed-forward pipeline (i.e., overleaping Gaussians) rather than the inherent redundancy of 3DGS itself. This issue is prevalent across existing generalizable 3DGS methods, and the aforementioned compression techniques are not suitable for these frameworks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.182, + 0.858 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Overview. As illustrated in Fig.2, SpatialSplat leverages a standard Vision Transformer (ViT) [10] backbone with a" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.907, + 0.455 + ], + "angle": 0, + "content": "few Dense Prediction Transformer (DPT) [35] heads. Given \\(V\\) unposed images at resolution \\(H\\times W\\) with their intrinsic \\(\\{I^v,K^v\\}_{v = 1}^V\\), SpatialSplat learn a mapping \\(f_{\\theta}\\) parametered by \\(\\theta\\) that transforms the input images to an fine-grained instance-aware radiance field \\(\\mathcal{F}_I\\) and a coarse semantic feature field \\(\\mathcal{F}_S\\), which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.464, + 0.746, + 0.483 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta}: \\left\\{\\boldsymbol {I} ^ {v}, \\boldsymbol {K} ^ {v} \\right\\} _ {v = 1} ^ {V} \\rightarrow \\mathcal {F} _ {I} + \\mathcal {F} _ {S},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.486, + 0.905, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {\\boldsymbol {I}} = \\left\\{\\cup \\left(\\boldsymbol {f} _ {\\boldsymbol {I}} ^ {i}, \\boldsymbol {\\beta} ^ {i}, \\boldsymbol {\\mu} ^ {i}, \\boldsymbol {\\alpha} ^ {i}, \\boldsymbol {r} ^ {i}, \\boldsymbol {s} ^ {i}, \\boldsymbol {c} ^ {i}\\right) \\right\\} _ {i = 1, \\dots , V \\times H \\times W}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.507, + 0.866, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} _ {S} = \\{\\cup (\\boldsymbol {f} _ {S} ^ {j}, \\boldsymbol {\\mu} ^ {j}, \\boldsymbol {\\alpha} ^ {j}, \\boldsymbol {r} ^ {j}, \\boldsymbol {s} ^ {j}, \\boldsymbol {c} ^ {j}) \\} _ {j = 1, \\dots , V \\times \\frac {H}{S} \\times \\frac {W}{S}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.674 + ], + "angle": 0, + "content": "\\(\\beta \\in \\mathbb{R}\\) is the importance score, which will be explained in Section 3.2. \\(\\pmb{f}_I \\in \\mathbb{R}^N\\) is the learnable instance feature and \\(\\pmb{f}_S \\in \\mathbb{R}^M\\) is the learnable semantic feature, as detailed in Section 3.3. The remains represent the vanilla Gaussian parameters [19], including the center position \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), opacity \\(\\alpha \\in \\mathbb{R}\\), rotation factor in quaternion \\(\\pmb{r} \\in \\mathbb{R}^4\\), scale \\(\\pmb{s} \\in \\mathbb{R}^3\\), and spherical harmonics (SH) \\(\\pmb{c} \\in \\mathbb{R}^k\\) with \\(k\\) degrees of freedom. In the following sections, we provide a detailed explanation of each component of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.738, + 0.699 + ], + "angle": 0, + "content": "3.1. 3D Geometry Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Both the encoder and decoder in our geometric prediction module are built on pure ViT structures, requiring no geometric priors as in previous methods [3, 7]. The input image is patchified and flattened into image sequences, which along with the camera intrinsics processed by a linear layer, are fed into the encoder. The encoder weights are shared across different input views. The features from encoder are then passed to a ViT-based decoder, where cross-attention is applied to better capture spatial relationships and aggregate information across views. Finally, the features from different views are separately fed into a series of DPT heads to predict point maps and other Gaussian parameters. The point maps serve as Gaussians centers. Since the input im" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "ages are unposed, the Gaussians centers suffer from scale ambiguity. Some methods [15, 38] rely on depth supervision to obtain per-scene scale, but this introduces additional errors due to imperfect depth. To overcome this limitation, we follow NoPoSplat's [46] approach by injecting camera intrinsics to resolve scale ambiguity. Experiments show that SpatialSplat effectively learns 3D priors from sparse unposed images without depth supervision, even while jointly learning multiple parameters and features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.364, + 0.255 + ], + "angle": 0, + "content": "3.2. Selective Gaussian Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.504 + ], + "angle": 0, + "content": "Previous methods [3, 7, 15, 38, 46] directly use the pixelwise pointmap as Gaussian centers which introduce significant redundancy. To address this, we propose a selective Gaussian mechanism that assigns each primitive an importance score to quantify its necessity for the scene representation. Primitives with importance scores below a certain threshold are considered redundant and discarded. To effectively supervise the learning of importance scores, we analyze Gaussian properties and the rendering pipeline, designing a simple yet effective strategy. Specifically, a primitive's influence can be reduced by decreasing its opacity or size [19]. However, we observe that even primitives with near-zero size still contribute to the rendering by occupying a pixel. Therefore, we multiply the importance score to the opacity of each primitive and modify the alpha blending formulation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.515, + 0.483, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\boldsymbol {\\alpha} _ {i} \\boldsymbol {\\beta} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\boldsymbol {\\alpha} _ {j} \\boldsymbol {\\beta} _ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.721 + ], + "angle": 0, + "content": "where \\( c \\) is the final intensities calculated by blending \\( n \\) ordered Gaussians overlapping the pixels, and \\( \\beta_{i} \\) is the importance score processed through a sigmoid activation function. This ensures that primitives with low importance have minimal impact on the final intensities. Removing important primitives leads to a significant increase in photometric loss, whereas discarding redundant ones has little effect. Therefore, we optimize \\( \\beta_{i} \\) through photometric loss minimization. Following methods [15, 46], we define the photometric loss as a weighted sum of MSE and LPIPS [50]:" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.747, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) = \\mathcal {L} _ {M S E} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda \\mathcal {L} _ {L P I P S} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( C \\) and \\( \\hat{C} \\) are rasterized and GT pixel intensities, and \\( \\lambda = 0.05 \\). Since redundant primitives are directly discarded during inference, we encourage their importance scores to approach zero to minimize the discrepancy between training and inference. Simultaneously, we aim to reduce the number of primitives used, striving to eliminate redundant ones as much as possible. To achieve these objectives, we first design the importance score with inspiration from Leaky" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.745, + 0.106 + ], + "angle": 0, + "content": "ReLU [28], formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.116, + 0.907, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\beta} _ {i} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\beta} _ {i} & \\text {i f} \\boldsymbol {\\beta} _ {i} > \\tau \\\\ \\boldsymbol {\\beta} _ {i} \\times 1 0 ^ {- 3} & \\text {i f} \\boldsymbol {\\beta} _ {i} < \\tau \\end{array} \\right.. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.167, + 0.907, + 0.243 + ], + "angle": 0, + "content": "This ensures that the importance scores of primitives below the threshold \\(\\tau\\) are close to zero while preventing gradient vanishing. Additionally, we introduce a Binary Cross Entropy (BCE) loss with a regularization term that pushes \\(\\beta_{i}\\) to either 0 or 1 while penalizing excessive use of primitives:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.252, + 0.907, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} (\\boldsymbol {S}) = \\mathcal {L} _ {B C E} (\\boldsymbol {S}, \\hat {\\boldsymbol {S}}) + \\frac {1}{\\| \\boldsymbol {S} \\|} \\sum_ {\\boldsymbol {\\beta} _ {i} \\in \\boldsymbol {S}} \\beta_ {i}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.296, + 0.56, + 0.309 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.312, + 0.857, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {S} = \\{\\beta_ {i} \\} _ {i = 1} ^ {H \\times W}, \\quad \\hat {\\boldsymbol {S}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\beta_ {i} > \\tau \\\\ 0 & \\text {i f} \\beta_ {i} < \\tau \\end{array} \\right..\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.363, + 0.906, + 0.408 + ], + "angle": 0, + "content": "As a result, SpatialSplat achieves a compact yet high-fidelity scene representation by selectively retaining only the most essential primitives." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.417, + 0.729, + 0.431 + ], + "angle": 0, + "content": "3.3. Dual-field Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.605 + ], + "angle": 0, + "content": "While distilling 2D semantic features into dense 3D representations has proven effective for scene understanding, conventional per-primitive compressed feature assignments [31, 36, 52] result in information loss and suboptimal performance. To mitigate this loss without increasing storage costs, we propose a dual-field architecture that decouples semantic representation into: 1) a fine-grained instance-aware radiance field, capturing scene geometry, textures, and instance correspondences, and 2) a coarse, uncompressed semantic feature field that reduces storage requirements while preserving semantic accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.607, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Fine-grained Instance-Aware Radiance Field. Our instance-aware radiance field builds upon Vanilla 3DGS, initializing the redundancy-aware Gaussians augmented with learnable instance features \\( \\pmb{f}_I \\). We train the network using contrastive learning [1] to pull rendered instance features of the same instance closer while push those of different instances further apart. To achieve this, we lift multiview 2D segmentation priors into 3D to help SpatialSplat learn inter-instance relationships. Given a target image \\( I \\), we first use Segment Anything Model (SAM) [21] to extract a set of masks \\( \\mathcal{M} = \\{M^k \\mid k = 1, 2, \\dots, m\\} \\). The instance feature map rendered at target image view with Eq. 2 is denoted as \\( F_I \\). The contrastive loss is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.812, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C} \\left(\\boldsymbol {F} _ {I}, \\mathcal {M}\\right) = - \\log \\frac {1}{\\| \\boldsymbol {I} \\| ^ {2}} \\sum_ {u \\in \\boldsymbol {I}} \\sum_ {u ^ {\\prime} \\in \\boldsymbol {I} ^ {+}} \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right) \\\\ + \\sum_ {u ^ {\\prime} \\notin I ^ {+}} \\left(1 - \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right)\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26407" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.1, + 0.904, + 0.276 + ], + "angle": 0, + "content": "
MethodsSemanticForwardSource ViewTarget View
mIoU ↑Acc. ↑mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
L-Seg [26]✓X0.55410.78240.55580.7856N/AN/AN/A
NeRF-DFF [23]✓X0.53810.77240.51370.758222.490.76500.2829
Feature-3DGS [52]✓X0.49920.73620.32230.567417.960.58120.4894
NoPoSplat [46]X✓N/AN/AN/AN/A25.700.81590.1875
LSM [15]✓✓0.51410.77190.51040.766224.120.79610.2526
SpatialSplat-Lite✓✓0.52720.78010.52650.769325.450.80320.2039
SpatialSplat✓✓0.55930.78140.55870.792425.460.80450.2046
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.278, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison in 3D Tasks on Scannet dataset. Our method outperforms both the latest SOTA semantic-aware feed-forward approach and per-scene optimization methods. \"Lite\": replacing the pretrained 2D model LSeg with CLIP ViT-B/16." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.482, + 0.458 + ], + "angle": 0, + "content": "where \\( I^{+} \\) is the set of pixels that belong to the same instance mask, \\( F_{I}(u) \\) is the feature vector of the \\( F_{I} \\) at coordinate \\( u \\) and \\( sim(. ) \\) is the cosine similarity. With this weak supervision, SpatialSplat effectively clusters radiance primitives by instance affiliation. Notably, the loss complexity in Eq. 6 scales quadratically with the number of pixels, making training infeasible. To address this, we employ a trick (detailed in the appendix) that enables efficient loss estimation with linear complexity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.485, + 0.702 + ], + "angle": 0, + "content": "Coarse Semantic Feature Field. Since primitives within the same instance share semantic consistency, SpatialSplat learns instance-level semantics using only a subset of primitives, avoiding the redundancy of per-primitive assignments. Manually selecting primitives for each instance, however, disrupts the network's cohesion and simplicity while increasing additional computational costs. To address this, we uniformly downsample the Gaussian centers from \\(\\mathcal{F}_I\\) to form a coarse feature field, where each selected primitive is assigned a semantic feature \\(f_{S}\\). To enhance semantic learning, we inject image features from a pretrained 2D model into feature heads, as shown in Fig. 2. A 2D semantic feature map \\(F_{S}\\) is rendered using alpha blending. We minimize the loss between the rendered feature map at a novel view and the feature map \\(\\hat{F}_{S}\\) of the ground truth image extracted from the pretrained 2D model:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.709, + 0.483, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) = M S E \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "During open-vocabulary querying, we first cluster primitives in \\(\\mathcal{F}_I\\) by instance feature similarity, assigning each instance the average feature of its clustered primitives as its label. For primitives in \\(\\mathcal{F}_S\\), their labels correspond to the instance features before sampling, and their semantic features are assigned to the instance with the highest cosine similarity. If multiple primitives contribute semantic features to the same instance, we take their average to ensure consistency. This approach enables the network to acquire sharper and clearer semantics compared to the previous methods. An additional advantage of dual-field architecture is that, un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.368 + ], + "angle": 0, + "content": "like dense semantic supervision from LSeg [26], it allows the use of a much lighter pretrained model (e.g., CLIP ViT-B/16 [34]), significantly improving inference speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.696, + 0.393 + ], + "angle": 0, + "content": "3.4. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Following previous methods [15], we perform end-to-end training to optimize our model with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.451, + 0.905, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda_ {1} \\mathcal {L} _ {I} (\\boldsymbol {S}) \\tag {8} \\\\ + \\lambda_ {2} \\mathcal {L} _ {C} (\\boldsymbol {F} _ {I}, \\mathcal {M}) + \\lambda_ {3} \\mathcal {L} _ {S} (\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.562 + ], + "angle": 0, + "content": "The parameters \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are set to 0.01, 0.2, and 1.0, respectively. To save training time, the instance masks \\(\\mathcal{M}\\) are generated by SAM [21] prior to training, while the semantic feature map \\(\\hat{\\pmb{F}}_S\\) is predicted during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.64, + 0.592 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.704, + 0.617 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.746 + ], + "angle": 0, + "content": "Datasets. Following LSM[15], we primarily train our model on a combination of ScanNet++[47] and ScanNet [8]. We filter out bad scenes and those with incomplete extrinsic parameters, resulting in a training dataset of approximately 1,500 scenes. For evaluation, we follow LSM and select 40 unseen scenes from ScanNet to assess our model's performance. We test our method's cross-dataset generalization on the synthesized Replica Dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate our method on two downstream tasks: novel view synthesis (NVS) and 3D open-vocabulary segmentation (OVS). For NVS, we adopt the standard metrics: PSNR, SSIM, and LPIPS [50]. For OVS, we evaluate performance using class-wise intersection over union (mIoU) and average pixel accuracy (mAcc)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Baselines. We primarily compare our method with LSM [15], the latest state-of-the-art (SOTA) approach for generalizable semantic 3D reconstruction. Additionally, we include the following baselines: 1): Feature-3DGS [52]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.891, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.407, + 0.908, + 0.438 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison in NVs. SpatialSplat can synthesize realistic novel views. In challenging cases where LSM fails, such as the table legs in the first two rows and the corners in the last two rows, our method achieves significantly better results." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.449, + 0.903, + 0.569 + ], + "angle": 0, + "content": "
SceneLSMSpatialSplat (Ours)
mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
room00.50240.742317.610.52470.32280.51310.793422.090.67600.1794
room10.46610.734914.340.59130.31740.46910.762820.620.67740.2265
office30.54840.762721.280.75670.20840.55820.799722.520.79130.1436
office40.50650.772319.510.69620.27790.50160.769521.130.72710.2072
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.57, + 0.84, + 0.585 + ], + "angle": 0, + "content": "Table 2. Out-of-distribution (OOD) comparison on Replica dataset. SpatialSplat generalizes well on OOD data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.607, + 0.483, + 0.729 + ], + "angle": 0, + "content": "and NeRF-DFF [23], pre-scene optimization methods for semantic 3D reconstruction based on 3DGS [19] and NeRF [29], respectively. 2): NoPoSplat [46], a latest SOTA method for generalizable novel view synthesis. 3): LSeg [26], a 2D open-vocabulary segmenter used for feature distillation in all compared semantic-aware methods. Unless otherwise specified, our method also uses image features from LSeg for supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation details. For the 3D geometry prediction module, we use ViT-Large with a patch size of 16 as the encoder and ViT-Base as the decoder, both initialized with pre-trained weights from MAST3R [25], while the remaining modules are randomly initialized. We modify the 3DGS CUDA renderer to support instance and semantic feature map rendering. The importance score threshold \\(\\tau\\) is set to 0.5, the downsampling ratio to 8, the instance feature dimension \\(N\\) to 8, and the semantic feature dimension \\(M\\) to 512. For a fair comparison, we train the model at a resolution of \\(256 \\times 256\\), consistent with our baseline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.71, + 0.623 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Results of Novel View Synthesis. Following previous methods [3, 46], we select two images from each test scene as input view for the generalizable 3DGS methods, and choose four additional images located between these two as target views to evaluate the performance of different methods in Novel View Synthesis. Notably, Feature-3DGS [52] and NeRF-DFF [23] relies on dense view inputs and performs poorly with only two input images. To account for this, we use five out of the six available images (the two input images and four target views) for training and evaluate on the remaining image. As shown in Tab. 1, Spatial-Splat significantly outperforms latest SOTA method LSM. We observe that LSM struggles in certain scenes due to its reliance on accurate depth for aligning views during training as shown in Fig. 3. Although NeRF-DFF uses more images for training, our method still outperforms it and even demonstrates competitive performance with very re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.098, + 0.222, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.184, + 0.22, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.269, + 0.219, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.356, + 0.219, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.439, + 0.186, + 0.448 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.098, + 0.332, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.184, + 0.332, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.269, + 0.331, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.355, + 0.331, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.439, + 0.291, + 0.448 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.098, + 0.442, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.184, + 0.442, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.269, + 0.442, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.355, + 0.442, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.439, + 0.412, + 0.45 + ], + "angle": 0, + "content": "LSeg" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.098, + 0.554, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.184, + 0.553, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.269, + 0.552, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.355, + 0.552, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.439, + 0.532, + 0.45 + ], + "angle": 0, + "content": "NeRF-DFF" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.098, + 0.664, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.184, + 0.664, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.269, + 0.664, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.355, + 0.664, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.439, + 0.658, + 0.449 + ], + "angle": 0, + "content": "Feature-3DGS" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.098, + 0.776, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.184, + 0.776, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.269, + 0.776, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.355, + 0.776, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.439, + 0.742, + 0.449 + ], + "angle": 0, + "content": "LSM" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.098, + 0.887, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.184, + 0.887, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.269, + 0.887, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.355, + 0.887, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.439, + 0.853, + 0.449 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.461, + 0.159, + 0.472 + ], + "angle": 0, + "content": "Wall" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.461, + 0.258, + 0.472 + ], + "angle": 0, + "content": "Floor" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.462, + 0.366, + 0.475 + ], + "angle": 0, + "content": "Ceiling" + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.462, + 0.464, + 0.473 + ], + "angle": 0, + "content": "Chair" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.462, + 0.566, + 0.473 + ], + "angle": 0, + "content": "Table" + }, + { + "type": "text", + "bbox": [ + 0.634, + 0.462, + 0.666, + 0.473 + ], + "angle": 0, + "content": "Sofa" + }, + { + "type": "text", + "bbox": [ + 0.74, + 0.462, + 0.767, + 0.472 + ], + "angle": 0, + "content": "Bed" + }, + { + "type": "text", + "bbox": [ + 0.833, + 0.462, + 0.88, + 0.473 + ], + "angle": 0, + "content": "Others" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.488, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison in OVS. SpatialSplat achieves sharper and more precise segmentation results compared to previous methods. Notably, our method excels in challenging details, such as distinguishing table legs from cabinet legs (the second the fourth rows), benefiting from our dual-field architecture that captures detailed instance information and uncompressed semantics." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.553, + 0.195, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.554, + 0.288, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.554, + 0.377, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.554, + 0.468, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.626, + 0.194, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.626, + 0.288, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.625, + 0.377, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.626, + 0.468, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.695, + 0.234, + 0.706 + ], + "angle": 0, + "content": "Input views" + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.695, + 0.443, + 0.706 + ], + "angle": 0, + "content": "Rendered Feature Field" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Figure 5. The rendered instance features. SpatialSplat predicts clear and consistent instance features across different views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.806 + ], + "angle": 0, + "content": "cent SOTA methods designed specifically for novel view synthesis. This highlights that our method efficiently learns spatial priors from large-scale 2D data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results of Open-vocabulary 3D Segmentation. Following LSM's approach, we map category labels from the Scannet dataset into common categories: Wall, Floor, Ceiling, Chair, Table, Bed, Sofa, Others. As shown in Tab. 1, SpatialSplat outperforms all compared methods, even surpassing L-Seg, which provides semantic feature supervision" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.553, + 0.6, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.664, + 0.581, + 0.673 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.554, + 0.671, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.609, + 0.671, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.663, + 0.657, + 0.673 + ], + "angle": 0, + "content": "LSM" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.554, + 0.744, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.61, + 0.744, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.663, + 0.727, + 0.673 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.554, + 0.817, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.61, + 0.817, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.663, + 0.8, + 0.673 + ], + "angle": 0, + "content": "LSM" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.554, + 0.89, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.609, + 0.89, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.846, + 0.663, + 0.875, + 0.673 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.676, + 0.905, + 0.703 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of cross-dataset generalization. Zoom out for clearer visualization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.868 + ], + "angle": 0, + "content": "for other compared methods. As illustrated in Fig. 4, SpatialSplat produces sharp and clear 3D semantic segmentation results. This is achieved by lifting both 2D instance and uncompressed semantic into 3D space, enabling more efficient semantic learning while mitigating the blurred contours and information loss associated with per-primitive compressed semantics. The visualizations in Fig. 5 show that SpatialSplat effectively learns a highly consistent instance prior with only weak supervision from 2D masks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Results of Cross-Dataset Generalization. As shown in Tab. 2, our method significantly outperforms LSM on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26410" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.103, + 0.195, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.171, + 0.16, + 0.182 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.103, + 0.286, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.171, + 0.25, + 0.182 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.103, + 0.377, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.171, + 0.341, + 0.182 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.103, + 0.468, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.171, + 0.433, + 0.182 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.256 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of ablations. (a) and (b) Qualitative results of importance score prediction, with red color indicating an importance score of 1 and blue indicating 0. (c) The rendered semantic feature without the dual-field. (d) The rendered semantic feature with the dual-field, which appears clearer and sharper." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.268, + 0.48, + 0.366 + ], + "angle": 0, + "content": "
MethodLatency↓Gaussian Size ↓Num. ↓
Feature-3DGS [52]1069 s229.65 MB476K
NoPoSplat [46]0.044 s30.93 MB131K
LSM [15]0.104 s63.00 MB131K
SpatialSplat-Lite0.058 s25.58 MB87.7K
SpatialSplat0.071 s25.58 MB87.7K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.369, + 0.482, + 0.397 + ], + "angle": 0, + "content": "Table 3. Model efficiency comparison. \"Lite\": the model replaces LSeg with CLIP ViT-B/16." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.41, + 0.483, + 0.531 + ], + "angle": 0, + "content": "unseen Replica dataset in both novel view synthesis and open-vocabulary segmentation. Notably, since Replica is a synthetic dataset with a different data modality from our training set, this underscores the strong generalization ability of our approach. The visualization in Fig. 6 further demonstrates that LSM struggles with coarse modeling at scene edges, leading to artifacts in synthesized views, whereas our method generates photorealistic renderings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Analysis of Model Efficiency. We compare the model efficiency of different 3DGS-based methods. As shown in Tab. 3, our method is faster than LSM thanks to its streamlined architecture, requiring only \\(40\\%\\) of the storage size and \\(65\\%\\) of primitive number. Remarkably, our approach requires even fewer Gaussians parameters than NoPoSplat, which lacks semantic awareness. This demonstrates the effectiveness of our proposed dual-field architecture and selective Gaussian mechanism, which capture accurate semantics while eliminating redundant Gaussians. Furthermore, as our method does not rely on dense semantic supervision, we leverage a lightweight pretrained 2D model, significantly accelerating inference speed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.308, + 0.757 + ], + "angle": 0, + "content": "4.3. Ablations and Analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We perform ablations to answer the following questions: (1) Are the primitives removed by our selective Gaussian mechanism truly redundant? (2) Does the improved scene understanding stem from the proposed dual-field architecture? (3) As the number of views increases, how does redundancy grow? Can our method effectively alleviate it?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of selective Gaussian mechanism. To answer Question 1, we remove the SGM and retrain our model using the same configuration. As shown in Tab. 4, the SGM" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.09, + 0.903, + 0.184 + ], + "angle": 0, + "content": "
ModelmIoU↑Acc.↑PSNR↑Num. ↓
No SGM0.55690.797225.55131k
No dual0.50270.778625.4687.6k
3-view0.51250.779324.9297.5k
SpatialSplat0.55770.792025.4787.7k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.242 + ], + "angle": 0, + "content": "Table 4. Ablation study on design choices. The proposed Adaptive Gaussian mechanism and dual-field architecture jointly enable the construction of high-quality representations with significantly reduced dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.349 + ], + "angle": 0, + "content": "eliminates approximately \\(35\\%\\) of Gaussian primitives while causing only a 0.08 drop in PSNR, indicating that the removed primitives have minimal impact on rendering quality. The visualization in Fig. 7 (a) and (b) further demonstrates that the SGM accurately localizes overlapping areas in the input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.504 + ], + "angle": 0, + "content": "Effect of dual-field architecture. To address Question 2, we replace our dual-field architecture with single-level per-primitive semantic learning. Following the approach in LSM, we set the semantic feature dimension to 64 and use a convolutional layer to expand it to 512. As shown in Tab. 4, this results in a significant drop of \\(5\\%\\) in mIoU for open-vocabulary segmentation performance. The primary issue is that per-primitive semantic learning struggles to maintain accurate semantics and fails to preserve clear instance boundaries, as illustrated in Fig. 7 (c) and (d)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Extend to more views. Our experiments primarily focus on two-view settings to ensure a fair comparison with our baseline. However, our method can be extended to more views for reconstructing larger scenes. As shown in Tab. 4, our approach remains effective even with three input views. Moreover, compared to pixel-wise methods, which require approximately 197K primitives to represent the scene, our method achieves a more compact representation by reducing about \\(50\\%\\) primitives. More qualitative details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.693 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present an efficient feed-forward network that generates compact yet high-performance representations for 3D reconstruction and scene understanding from sparse, unposed images. Our proposed dual-field architecture achieves state-of-the-art performance while significantly reducing the number of representation parameters. Additionally, we introduce a Selective Gaussian Mechanism that directly identifies redundant Gaussians caused by pixel-wise predictions without requiring geometric priors. SpatialSplat achieves impressive results in 3D reconstruction and scene understanding without relying on any 3D data during training or inference, making semantic 3D reconstruction from sparse, unposed images more practical." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26411" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.171 + ], + "angle": 0, + "content": "This work was partially supported by Hunan Province Major Scientific and Technological Project No. 2024QK2001, National Key R&D Program of China No. 2023YFB4704500." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.2, + 0.188, + 0.215 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[1] Yash Bhalgat, Iro Laina, João F. Henriques, Andrea Vedaldi, and Andrew Zisserman. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.323, + 0.482, + 0.406 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9630-9640. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.482, + 0.49 + ], + "angle": 0, + "content": "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19457-19467. IEEE, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. CoRR, abs/2401.03890, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[6] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: hash-grid assisted context for 3d gaussian splatting compression. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part VII, pages 422-438. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[7] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXI, pages 370-386. Springer, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[9] Jianmei Dai, Zhilong Zhang, Shiwen Mao, and Danpu Liu. A view synthesis-based \\(360^{\\circ}\\) VR caching system over mec-enabled C-RAN. IEEE Trans. Circuits Syst. Video Technol., 30(10):3843-3855, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.224, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.285 + ], + "angle": 0, + "content": "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4970-4980. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. CoRR, abs/2403.20309, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and \\(200+\\) FPS. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[15] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, Boris Ivanovic, and Marco Pavone. Large spatial model: End-to-end unposed images to semantic 3d. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 1, 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 12479-12488. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[17] Qiao Gu, Zhaoyang Lv, Duncan P. Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIII, pages 382-400. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[18] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18344-18347. IEEE, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26412" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19672-19682. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3992-4003. IEEE, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[23] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[24] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21719-21728. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[25] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[26] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[27] Luca De Luigi, Damiano Bolognini, Federico Domeniconi, Daniele De Gregorio, Matteo Poggi, and Luigi Di Stefano. Scannerf: a scalable benchmark for neural radiance fields. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 816-825. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[28] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[31] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20051-20060. IEEE, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[32] Ri-Zhao Qiu, Yafei Hu, Ge Yang, Yuchen Song, Yang Fu, Jianglong Ye, Jiteng Mu, Ruihan Yang, Nikolay Atanasov, Sebastian A. Scherer, and Xiaolong Wang. Learning generalizable feature fields for mobile manipulation. CoRR, abs/2403.07563, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[33] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLI, pages 368-383. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 12159-12168. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[36] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In Conference on Robot Learning, CoRL 2023, 6-9 November 2023, Atlanta, GA, USA, pages 405-424. PMLR, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.607, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[37] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 9043-9052. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[38] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2, 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[39] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26413" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[40] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In International Conference on 3D Vision, 3DV 2022, Prague, Czech Republic, September 12-16, 2022, pages 443-453. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[41] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4690-4699. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[42] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[43] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs. CoRR, abs/2412.16932, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[44] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[45] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20041-20050. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[46] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4, 5, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[48] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4578-4587. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[49] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. CoRR, abs/2408.13770, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[50] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[51] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15818-15827. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[52] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21676-21685. IEEE, 2024. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26414" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_origin.pdf b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5c3ec0f8ed07a731fb65e135d912d4f4d0a6c32f --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/b5246924-8fb7-4786-baa2-29a57614062d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee8ae3ce934f10acda08fe6ff3a2b45eda73d6780caf14b6cdf1e9a7c70946be +size 5909309 diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/full.md b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e78146d22b66685a4d7a628a4128b5630dd1000d --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/full.md @@ -0,0 +1,425 @@ +# SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images + +Yu Sheng $^{1}$ , Jiajun Deng $^{2\dagger}$ , Xinran Zhang $^{1}$ , Yu Zhang $^{1}$ , Bei Hua $^{1}$ , Yanyong Zhang $^{1}$ , Jianmin Ji $^{1\dagger}$ $^{1}$ University of Science and Technology of China, $^{2}$ The University of Adelaide + +{shengyu724, zxrr}@mail.ustc.edu.cn, jiajun.deng@adelaide.edu.au + +{zhangyu,bhua,yanyongz,jianmin}@ustc.edu.cn †: Corresponding Author + +![](images/f3c34c3daae2b917c496310f5f0b8e4dc6853fe327dbc647ebc24d84048344a9.jpg) +Figure 1. Comparison between previous methods and our SpatialSplat. (a): Previous methods predict pixel-wise Gaussians, associating each primitive with compressed semantic feature. (b): Our SpatialSplat avoids generating redundant Gaussian primitives for overlapping pixels, and it represents semantics with a dual-field architecture to better preserve the information. (c): SpatialSplat outperforms the state-of-the-art method LSM [15] for both novel-view rendering quality and semantic segmentation while being more efficient. + +![](images/fa58bb773a8f08e4e6a8f901883311c541a0e7032df8b973f9f276c359595d6b.jpg) + +![](images/b3daf1c37c562445242a2565d5df8e75f09140300e699cd05c9308e8989f6a5b.jpg) + +# Abstract + +A major breakthrough in 3D reconstruction is the feedforward paradigm to generate pixel-wise 3D points or Gaussian primitives from sparse, unposed images. To further incorporate semantics while avoiding the significant memory and storage costs of high-dimensional semantic features, existing methods extend this paradigm by associating each primitive with a compressed semantic feature vector. However, these methods have two major limitations: (a) the naively compressed feature compromises expressiveness, affecting the model's ability to capture fine-grained semantics, and (b) the pixel-wise primitive prediction introduces redundancy in overlapping areas, causing unnecessary memory overhead. To this end, we introduce SpatialSplat, a feedforward framework that produces redundancy-aware Gaussians and capitalizes on a dual-field semantic representation. Particularly, with the insight that primitives within the same instance exhibit high semantic consistency, we decompose the semantic representation into a coarse feature field that encodes uncompressed semantics with minimal primitives, and a fine-grained yet low-dimensional feature field that captures detailed inter + +instance relationships. Moreover, we propose a selective Gaussian mechanism, which retains only essential Gaussians in the scene, effectively eliminating redundant primitives. Our proposed Spatialplat learns accurate semantic information and detailed instances prior with more compact 3D Gaussians, making semantic 3D reconstruction more applicable. We conduct extensive experiments to evaluate our method, demonstrating a remarkable $60\%$ reduction in scene representation parameters while achieving superior performance over state-of-the-art methods. The code will be made available for future investigation. + +# 1. Introduction + +Reconstructing and understanding 3D scenes from 2D images [5, 9, 32, 36, 53] is a fundamental topic in computer vision, aiming to obtain semantic-aware 3D structure from low-cost devices, i.e., RGB cameras. This technique is significant for various applications, such as robotics, autonomous driving, and VR/AR. Recently, with the emergence of the new 3D representation, i.e., Neural Radiance Fields (NeRF) [29] and 3D Gaussian Splatting (3DGS) [19], a popular paradigm of performing semantic-aware 3D re + +construction is to distill the feature field of NeRF or 3DGS with powerful 2D vision-language foundation models [20, 31, 33, 52]. However, these methods typically rely on perscense optimization and complex multi-step preprocessing, limiting their ability to generalize across multiple scenes within a single model. + +Recent breakthroughs [3, 7, 49] in 3D reconstruction have incorporated feed-forward networks to improve the generalization of 3DGS models and accelerate reconstruction. These methods follow a paradigm that generates pixel-wise Gaussian primitives from sparse posed images. Building on this, further efforts [38, 46] have extended the paradigm to reconstruct scenes from unposed images. Moreover, methods [15, 43] have advanced this approach by integrating semantic understanding, enabling semantic 3D reconstruction from sparse unposed images. + +Despite significant progress, these methods have two major limitations. First, as shown in Fig.1 (a), pixelwise Gaussian prediction introduces substantial redundancy in overlapping regions while providing minimal accuracy gains. Method [44] attempts to mitigate this issue by projecting primitives onto a plane and identifying overlaps, it depends on precise camera extrinsics, which are often difficult to obtain in real-world scenarios—especially when only a few sparse, textureless images are available. Second, methods [15, 31, 33, 43, 52] enforce per-primitive semantic learning to form a dense semantic feature field, as shown in Fig.1 (a). This primitive-level approach incurs significant memory and storage costs due to high-dimensional (512 or more) language features. To mitigate this, the methods often compress features into lower-dimensional representations (e.g., 64-128), but this compression leads to irreversible information loss [52], leading to suboptimal scene understanding performance. + +In this paper, we take a slightly different viewpoint. First, inspired by MASt3R [25], which successfully extracts 3D geometry and pixel correspondences from images, we observe that redundant primitives often share similar geometry and appearance, allowing them to be identified directly from images without relying on complex geometric priors. Second, we find that per-primitive semantics are not essential for high-performance scene understanding. Instead, a coarse semantic representation, when combined with fine-grained instance information, is sufficient for strong performance, as Gaussian primitives within the same instance exhibit high semantic consistency. + +By consolidating this idea, we introduce SpatialSplat, a feed-forward network generating compact yet expressive semantic 3D representations from unposed images. As depicted in Fig. 1 (b), SpatialSplat introduces a dual-field architecture that decomposes the dense semantic feature field into two components: a coarse feature field encoding uncompressed instance-level semantics through mini + +mal Gaussians, and a low-dimensional fine-grained feature field capturing inter-instance relationships. Guided by multiple 2D foundation models, our method learns both accurate semantic and instance priors, achieving superior performance while reducing representation parameters. Additionally, we introduce a Selective Gaussian Mechanism (SGM) to eliminate redundancy in overlapping areas caused by pixelwise representations, along with a novel loss function that jointly optimizes redundancy-aware Gaussians and scene fidelity. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple downstream 3D tasks while using only $40\%$ of the representation parameters required by the baseline, all without any 3D supervision. Our contributions are threefold: + +- A novel feed-forward 3DGS framework that, to the best of our knowledge, is the first to simultaneously learn semantic and instance priors with guidance from foundation models while maintaining a high-fidelity radiance field. +- A dual-field semantic representation that significantly reduces storage consumption while enhancing open-vocabulary 3D segmentation performance. +- A mechanism that identifies redundant primitives from images without relying on any geometric priors. + +# 2. Related Work + +# 2.1. Feed-forward 3D Reconstruction + +The emergence of NeRF [29] and 3DGS [19] has significantly transformed the paradigm of 3D reconstruction from images. Despite achieving remarkable results in 3D reconstruction and novel view synthesis, these methods rely on time-consuming per-scene optimization, with even improved techniques [4, 16, 30] still requiring several minutes to hours. To overcome these limitations, recent approaches [3, 7, 11, 18, 41, 45, 48, 49] have explored using a single feed-forward network to learn reconstruction priors from large datasets, enabling generalizable 3D reconstruction. DUSt3R serials [25, 42] further simplifies the reconstruction pipeline, making 3D reconstruction possible from images without relying on camera extrinsics. This advancement has inspired a new wave of methods [13, 38, 46] capable of achieving dense reconstruction and view synthesis without camera extrinsics. + +# 2.2. Feature Field Distillation + +Early approaches of feature field distillation were largely based on NeRF [29]. For example, methods [37, 51] successfully embedded 2D label data into 3D space, achieving sharp and precise 3D segmentation. Later approaches [12, 20, 22, 36, 40] incorporated more complex pixel-aligned feature vectors from technologies like LSeg [26] or DINO [2] into NeRF frameworks, enabling open-vocabulary 3D segmentation. With the rise of + +![](images/3ca3c1e5b776457dda56ff2b09058f990e66f96beb2ee228f79f5ad9e4ab5796.jpg) +Figure 2. Pipeline of SpatialSplat. The SpatialSplat processes unposed images along with their intrinsicities through a 3D geometry transformer. The extracted features from the geometry transformer and the pretrained 2D model are then fed into separate Dense Prediction Transformer (DPT) heads to predict different Gaussian attributes, resulting in a fine-grained instance-aware radiance field $\mathcal{F}_I$ and a coarse semantic feature field $\mathcal{F}_S$ . This enables the synthesis of RGB and feature maps from novel viewpoints. + +3DGS [19], methods such as [17, 31, 33, 52] have extended this concept by embedding language features into Gaussian attributes to enhance semantic understanding. Large Spatial Model (LSM) [15] was the first to achieve a generalizable semantic radiance field based on feed-forward 3DGS. However, due to the explicit nature of 3DGS representations, attaching feature vectors to each primitive incurs high memory costs. To mitigate this, existing methods often compress semantic features into lower-dimensional representations, inevitably resulting in information loss. + +In contrast, our approach decomposes semantics into coarse instance-level semantics and fine-grained instance information, achieving superior performance with significantly lower storage requirements. + +# 2.3. Compact 3D Gaussian + +Our work is also related to compact 3D Gaussian representation. Recent works have attempted to compress 3DGS by exploring the spatial relationships among Gaussians. Scaffold-GS [27] introduces anchor-centered features to reduce parameter counts, while HAC [6] builds on this by incorporating a hash grid to better organize scene primitives. Methods [14, 24] utilize techniques like distillation and Gaussian pruning to minimize storage requirements. + +Notably, our focus is on addressing the additional representational redundancy introduced by the feed-forward pipeline (i.e., overleaping Gaussians) rather than the inherent redundancy of 3DGS itself. This issue is prevalent across existing generalizable 3DGS methods, and the aforementioned compression techniques are not suitable for these frameworks. + +# 3. Method + +Overview. As illustrated in Fig.2, SpatialSplat leverages a standard Vision Transformer (ViT) [10] backbone with a + +few Dense Prediction Transformer (DPT) [35] heads. Given $V$ unposed images at resolution $H\times W$ with their intrinsic $\{I^v,K^v\}_{v = 1}^V$ , SpatialSplat learn a mapping $f_{\theta}$ parametered by $\theta$ that transforms the input images to an fine-grained instance-aware radiance field $\mathcal{F}_I$ and a coarse semantic feature field $\mathcal{F}_S$ , which can be formulated as: + +$$ +f _ {\theta}: \left\{\boldsymbol {I} ^ {v}, \boldsymbol {K} ^ {v} \right\} _ {v = 1} ^ {V} \rightarrow \mathcal {F} _ {I} + \mathcal {F} _ {S}, +$$ + +$$ +\mathcal {F} _ {\boldsymbol {I}} = \left\{\cup \left(\boldsymbol {f} _ {\boldsymbol {I}} ^ {i}, \boldsymbol {\beta} ^ {i}, \boldsymbol {\mu} ^ {i}, \boldsymbol {\alpha} ^ {i}, \boldsymbol {r} ^ {i}, \boldsymbol {s} ^ {i}, \boldsymbol {c} ^ {i}\right) \right\} _ {i = 1, \dots , V \times H \times W}, \tag {1} +$$ + +$$ +\mathcal {F} _ {S} = \{\cup (\boldsymbol {f} _ {S} ^ {j}, \boldsymbol {\mu} ^ {j}, \boldsymbol {\alpha} ^ {j}, \boldsymbol {r} ^ {j}, \boldsymbol {s} ^ {j}, \boldsymbol {c} ^ {j}) \} _ {j = 1, \dots , V \times \frac {H}{S} \times \frac {W}{S}}. +$$ + +$\beta \in \mathbb{R}$ is the importance score, which will be explained in Section 3.2. $\pmb{f}_I \in \mathbb{R}^N$ is the learnable instance feature and $\pmb{f}_S \in \mathbb{R}^M$ is the learnable semantic feature, as detailed in Section 3.3. The remains represent the vanilla Gaussian parameters [19], including the center position $\pmb{\mu} \in \mathbb{R}^3$ , opacity $\alpha \in \mathbb{R}$ , rotation factor in quaternion $\pmb{r} \in \mathbb{R}^4$ , scale $\pmb{s} \in \mathbb{R}^3$ , and spherical harmonics (SH) $\pmb{c} \in \mathbb{R}^k$ with $k$ degrees of freedom. In the following sections, we provide a detailed explanation of each component of our method. + +# 3.1. 3D Geometry Prediction + +Both the encoder and decoder in our geometric prediction module are built on pure ViT structures, requiring no geometric priors as in previous methods [3, 7]. The input image is patchified and flattened into image sequences, which along with the camera intrinsics processed by a linear layer, are fed into the encoder. The encoder weights are shared across different input views. The features from encoder are then passed to a ViT-based decoder, where cross-attention is applied to better capture spatial relationships and aggregate information across views. Finally, the features from different views are separately fed into a series of DPT heads to predict point maps and other Gaussian parameters. The point maps serve as Gaussians centers. Since the input im + +ages are unposed, the Gaussians centers suffer from scale ambiguity. Some methods [15, 38] rely on depth supervision to obtain per-scene scale, but this introduces additional errors due to imperfect depth. To overcome this limitation, we follow NoPoSplat's [46] approach by injecting camera intrinsics to resolve scale ambiguity. Experiments show that SpatialSplat effectively learns 3D priors from sparse unposed images without depth supervision, even while jointly learning multiple parameters and features. + +# 3.2. Selective Gaussian Mechanism + +Previous methods [3, 7, 15, 38, 46] directly use the pixelwise pointmap as Gaussian centers which introduce significant redundancy. To address this, we propose a selective Gaussian mechanism that assigns each primitive an importance score to quantify its necessity for the scene representation. Primitives with importance scores below a certain threshold are considered redundant and discarded. To effectively supervise the learning of importance scores, we analyze Gaussian properties and the rendering pipeline, designing a simple yet effective strategy. Specifically, a primitive's influence can be reduced by decreasing its opacity or size [19]. However, we observe that even primitives with near-zero size still contribute to the rendering by occupying a pixel. Therefore, we multiply the importance score to the opacity of each primitive and modify the alpha blending formulation as follows: + +$$ +\boldsymbol {c} = \sum_ {i = 1} ^ {n} \boldsymbol {c} _ {i} \boldsymbol {\alpha} _ {i} \boldsymbol {\beta} _ {i} \prod_ {j = 1} ^ {i - 1} (1 - \boldsymbol {\alpha} _ {j} \boldsymbol {\beta} _ {j}), \tag {2} +$$ + +where $c$ is the final intensities calculated by blending $n$ ordered Gaussians overlapping the pixels, and $\beta_{i}$ is the importance score processed through a sigmoid activation function. This ensures that primitives with low importance have minimal impact on the final intensities. Removing important primitives leads to a significant increase in photometric loss, whereas discarding redundant ones has little effect. Therefore, we optimize $\beta_{i}$ through photometric loss minimization. Following methods [15, 46], we define the photometric loss as a weighted sum of MSE and LPIPS [50]: + +$$ +\mathcal {L} _ {P} (\boldsymbol {C}, \hat {\boldsymbol {C}}) = \mathcal {L} _ {M S E} (\boldsymbol {C}, \hat {\boldsymbol {C}}) + \lambda \mathcal {L} _ {L P I P S} (\boldsymbol {C}, \hat {\boldsymbol {C}}), \tag {3} +$$ + +where $C$ and $\hat{C}$ are rasterized and GT pixel intensities, and $\lambda = 0.05$ . Since redundant primitives are directly discarded during inference, we encourage their importance scores to approach zero to minimize the discrepancy between training and inference. Simultaneously, we aim to reduce the number of primitives used, striving to eliminate redundant ones as much as possible. To achieve these objectives, we first design the importance score with inspiration from Leaky + +ReLU [28], formulated as follows: + +$$ +\boldsymbol {\beta} _ {i} = \left\{ \begin{array}{l l} \boldsymbol {\beta} _ {i} & \text {i f} \boldsymbol {\beta} _ {i} > \tau \\ \boldsymbol {\beta} _ {i} \times 1 0 ^ {- 3} & \text {i f} \boldsymbol {\beta} _ {i} < \tau \end{array} \right.. \tag {4} +$$ + +This ensures that the importance scores of primitives below the threshold $\tau$ are close to zero while preventing gradient vanishing. Additionally, we introduce a Binary Cross Entropy (BCE) loss with a regularization term that pushes $\beta_{i}$ to either 0 or 1 while penalizing excessive use of primitives: + +$$ +\mathcal {L} _ {I} (\boldsymbol {S}) = \mathcal {L} _ {B C E} (\boldsymbol {S}, \hat {\boldsymbol {S}}) + \frac {1}{\| \boldsymbol {S} \|} \sum_ {\boldsymbol {\beta} _ {i} \in \boldsymbol {S}} \beta_ {i}, \tag {5} +$$ + +where + +$$ +\boldsymbol {S} = \{\beta_ {i} \} _ {i = 1} ^ {H \times W}, \quad \hat {\boldsymbol {S}} = \left\{ \begin{array}{l l} 1 & \text {i f} \beta_ {i} > \tau \\ 0 & \text {i f} \beta_ {i} < \tau \end{array} \right.. +$$ + +As a result, SpatialSplat achieves a compact yet high-fidelity scene representation by selectively retaining only the most essential primitives. + +# 3.3. Dual-field Architecture + +While distilling 2D semantic features into dense 3D representations has proven effective for scene understanding, conventional per-primitive compressed feature assignments [31, 36, 52] result in information loss and suboptimal performance. To mitigate this loss without increasing storage costs, we propose a dual-field architecture that decouples semantic representation into: 1) a fine-grained instance-aware radiance field, capturing scene geometry, textures, and instance correspondences, and 2) a coarse, uncompressed semantic feature field that reduces storage requirements while preserving semantic accuracy. + +Fine-grained Instance-Aware Radiance Field. Our instance-aware radiance field builds upon Vanilla 3DGS, initializing the redundancy-aware Gaussians augmented with learnable instance features $\pmb{f}_I$ . We train the network using contrastive learning [1] to pull rendered instance features of the same instance closer while push those of different instances further apart. To achieve this, we lift multiview 2D segmentation priors into 3D to help SpatialSplat learn inter-instance relationships. Given a target image $I$ , we first use Segment Anything Model (SAM) [21] to extract a set of masks $\mathcal{M} = \{M^k \mid k = 1, 2, \dots, m\}$ . The instance feature map rendered at target image view with Eq. 2 is denoted as $F_I$ . The contrastive loss is then defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {C} \left(\boldsymbol {F} _ {I}, \mathcal {M}\right) = - \log \frac {1}{\| \boldsymbol {I} \| ^ {2}} \sum_ {u \in \boldsymbol {I}} \sum_ {u ^ {\prime} \in \boldsymbol {I} ^ {+}} \operatorname {s i m} \left(\boldsymbol {F} _ {I} (u), \boldsymbol {F} _ {I} \left(u ^ {\prime}\right)\right) \\ + \sum_ {u ^ {\prime} \notin I ^ {+}} \left(1 - \operatorname {s i m} \left(\boldsymbol {F} _ {I} (u), \boldsymbol {F} _ {I} \left(u ^ {\prime}\right)\right)\right), \tag {6} \\ \end{array} +$$ + +
MethodsSemanticForwardSource ViewTarget View
mIoU ↑Acc. ↑mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
L-Seg [26]✓X0.55410.78240.55580.7856N/AN/AN/A
NeRF-DFF [23]✓X0.53810.77240.51370.758222.490.76500.2829
Feature-3DGS [52]✓X0.49920.73620.32230.567417.960.58120.4894
NoPoSplat [46]X✓N/AN/AN/AN/A25.700.81590.1875
LSM [15]✓✓0.51410.77190.51040.766224.120.79610.2526
SpatialSplat-Lite✓✓0.52720.78010.52650.769325.450.80320.2039
SpatialSplat✓✓0.55930.78140.55870.792425.460.80450.2046
+ +Table 1. Quantitative Comparison in 3D Tasks on Scannet dataset. Our method outperforms both the latest SOTA semantic-aware feed-forward approach and per-scene optimization methods. "Lite": replacing the pretrained 2D model LSeg with CLIP ViT-B/16. + +where $I^{+}$ is the set of pixels that belong to the same instance mask, $F_{I}(u)$ is the feature vector of the $F_{I}$ at coordinate $u$ and $sim(. )$ is the cosine similarity. With this weak supervision, SpatialSplat effectively clusters radiance primitives by instance affiliation. Notably, the loss complexity in Eq. 6 scales quadratically with the number of pixels, making training infeasible. To address this, we employ a trick (detailed in the appendix) that enables efficient loss estimation with linear complexity. + +Coarse Semantic Feature Field. Since primitives within the same instance share semantic consistency, SpatialSplat learns instance-level semantics using only a subset of primitives, avoiding the redundancy of per-primitive assignments. Manually selecting primitives for each instance, however, disrupts the network's cohesion and simplicity while increasing additional computational costs. To address this, we uniformly downsample the Gaussian centers from $\mathcal{F}_I$ to form a coarse feature field, where each selected primitive is assigned a semantic feature $f_{S}$ . To enhance semantic learning, we inject image features from a pretrained 2D model into feature heads, as shown in Fig. 2. A 2D semantic feature map $F_{S}$ is rendered using alpha blending. We minimize the loss between the rendered feature map at a novel view and the feature map $\hat{F}_{S}$ of the ground truth image extracted from the pretrained 2D model: + +$$ +\mathcal {L} _ {S} \left(\boldsymbol {F} _ {S}, \hat {\boldsymbol {F}} _ {S}\right) = M S E \left(\boldsymbol {F} _ {S}, \hat {\boldsymbol {F}} _ {S}\right) \tag {7} +$$ + +During open-vocabulary querying, we first cluster primitives in $\mathcal{F}_I$ by instance feature similarity, assigning each instance the average feature of its clustered primitives as its label. For primitives in $\mathcal{F}_S$ , their labels correspond to the instance features before sampling, and their semantic features are assigned to the instance with the highest cosine similarity. If multiple primitives contribute semantic features to the same instance, we take their average to ensure consistency. This approach enables the network to acquire sharper and clearer semantics compared to the previous methods. An additional advantage of dual-field architecture is that, un + +like dense semantic supervision from LSeg [26], it allows the use of a much lighter pretrained model (e.g., CLIP ViT-B/16 [34]), significantly improving inference speed. + +# 3.4. Training Objective + +Following previous methods [15], we perform end-to-end training to optimize our model with the following objective: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {P} (\boldsymbol {C}, \hat {\boldsymbol {C}}) + \lambda_ {1} \mathcal {L} _ {I} (\boldsymbol {S}) \tag {8} \\ + \lambda_ {2} \mathcal {L} _ {C} (\boldsymbol {F} _ {I}, \mathcal {M}) + \lambda_ {3} \mathcal {L} _ {S} (\boldsymbol {F} _ {S}, \hat {\boldsymbol {F}} _ {S}). \\ \end{array} +$$ + +The parameters $\lambda_1, \lambda_2, \lambda_3$ are set to 0.01, 0.2, and 1.0, respectively. To save training time, the instance masks $\mathcal{M}$ are generated by SAM [21] prior to training, while the semantic feature map $\hat{\pmb{F}}_S$ is predicted during training. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. Following LSM[15], we primarily train our model on a combination of ScanNet++[47] and ScanNet [8]. We filter out bad scenes and those with incomplete extrinsic parameters, resulting in a training dataset of approximately 1,500 scenes. For evaluation, we follow LSM and select 40 unseen scenes from ScanNet to assess our model's performance. We test our method's cross-dataset generalization on the synthesized Replica Dataset [39]. + +Evaluation Metrics. We evaluate our method on two downstream tasks: novel view synthesis (NVS) and 3D open-vocabulary segmentation (OVS). For NVS, we adopt the standard metrics: PSNR, SSIM, and LPIPS [50]. For OVS, we evaluate performance using class-wise intersection over union (mIoU) and average pixel accuracy (mAcc). + +Baselines. We primarily compare our method with LSM [15], the latest state-of-the-art (SOTA) approach for generalizable semantic 3D reconstruction. Additionally, we include the following baselines: 1): Feature-3DGS [52] + +![](images/d60ac0a66109e44e4515ecaa35ad6ec3713a7e144d6a2720f2d920be15de4595.jpg) +Figure 3. Qualitative comparison in NVs. SpatialSplat can synthesize realistic novel views. In challenging cases where LSM fails, such as the table legs in the first two rows and the corners in the last two rows, our method achieves significantly better results. + +
SceneLSMSpatialSplat (Ours)
mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
room00.50240.742317.610.52470.32280.51310.793422.090.67600.1794
room10.46610.734914.340.59130.31740.46910.762820.620.67740.2265
office30.54840.762721.280.75670.20840.55820.799722.520.79130.1436
office40.50650.772319.510.69620.27790.50160.769521.130.72710.2072
+ +Table 2. Out-of-distribution (OOD) comparison on Replica dataset. SpatialSplat generalizes well on OOD data. + +and NeRF-DFF [23], pre-scene optimization methods for semantic 3D reconstruction based on 3DGS [19] and NeRF [29], respectively. 2): NoPoSplat [46], a latest SOTA method for generalizable novel view synthesis. 3): LSeg [26], a 2D open-vocabulary segmenter used for feature distillation in all compared semantic-aware methods. Unless otherwise specified, our method also uses image features from LSeg for supervision. + +Implementation details. For the 3D geometry prediction module, we use ViT-Large with a patch size of 16 as the encoder and ViT-Base as the decoder, both initialized with pre-trained weights from MAST3R [25], while the remaining modules are randomly initialized. We modify the 3DGS CUDA renderer to support instance and semantic feature map rendering. The importance score threshold $\tau$ is set to 0.5, the downsampling ratio to 8, the instance feature dimension $N$ to 8, and the semantic feature dimension $M$ to 512. For a fair comparison, we train the model at a resolution of $256 \times 256$ , consistent with our baseline. + +# 4.2. Results and Analysis + +Results of Novel View Synthesis. Following previous methods [3, 46], we select two images from each test scene as input view for the generalizable 3DGS methods, and choose four additional images located between these two as target views to evaluate the performance of different methods in Novel View Synthesis. Notably, Feature-3DGS [52] and NeRF-DFF [23] relies on dense view inputs and performs poorly with only two input images. To account for this, we use five out of the six available images (the two input images and four target views) for training and evaluate on the remaining image. As shown in Tab. 1, Spatial-Splat significantly outperforms latest SOTA method LSM. We observe that LSM struggles in certain scenes due to its reliance on accurate depth for aligning views during training as shown in Fig. 3. Although NeRF-DFF uses more images for training, our method still outperforms it and even demonstrates competitive performance with very re + +![](images/7adbd88e415b1782e08ad32a8f14d479bbf0e68e4557668c63e7be70bf2cdb3c.jpg) + +![](images/05e630337a72659c63d80ba405039ffdadc056ee42a3674e1f6c227a7a9f9c87.jpg) + +![](images/8d9d4de1b8eee7075564d3559b1785e0fd6e4d528152a49ca518944d74f94ea9.jpg) + +![](images/ae5f6f914002c52e64e75b90b61acc2934abc0d60f8a7690b403fab0f51f859b.jpg) +RGB + +![](images/cf0987a70ae53850481e6c4f6079568bcd111d97a8681093799a04a705c6b051.jpg) + +![](images/eaada965a6447fc2a4be6c2ab36021da83df6ca5cd3dbc9d73d9d77c366d4948.jpg) + +![](images/71c786ad40dd8fb57717e468be0b87160b5de383244522178ee88a1a87edcec0.jpg) + +![](images/948d722f0c36bac7ae8f756dbe36c22ccfc3cc9b319eee1236c4b343bc1d83f0.jpg) +GT + +![](images/742965fb04db8c4e71e45147e48e6cfcd01df17995266a4f2fe17efae9d9d0f0.jpg) + +![](images/d80f53b12b2cc57c0b9203b63a158f794477d99d1049859b7fbeb81a95ef3819.jpg) + +![](images/acd795a1bd846dbabab26ab7f40a7c5a036c03a7cef9d7c00013e034f0e0b4f5.jpg) + +![](images/dce9bd8eb625ce66fb3b054a307ddd5ba2fff89f6edb21c3976df8c37fef1db2.jpg) +LSeg + +![](images/f3000a16cb5513a700e503069b235891466071606b3593991d0bcb6f58bc898d.jpg) + +![](images/1f485edf0cceea1ee0e27e8b57302d15cd81e1a2bcf7efc6742eeb8565a95c60.jpg) + +![](images/274a31ba1ceb5e58489c73bffb190e186df071e66432d531f9ca9b953f3823f6.jpg) + +![](images/63f52a2b804dd52190abe7cabb11316e5efcc89e069ffe9b9d0f018234706f19.jpg) +NeRF-DFF + +![](images/ce5c16875e7b76a252830491b175905df469c89b4fdd964848c36c53e0960686.jpg) + +![](images/fb1643c48ce1459c8892a458436df37095e1d51d9c732f5c82cef36ba45651d0.jpg) + +![](images/baa599d73b8f42a465999ae40dfd6460d1c080b39451c50b054c0975cf77e3be.jpg) + +![](images/51912050a7bbb4bf759cfa00207049a87b4220d627932c74d7c813ae5993ce47.jpg) +Feature-3DGS + +![](images/868f9008bd2025ac29fac86dc02c73e67974bf29ed731141da7fb3afda7e2e2f.jpg) + +![](images/7b2e79df87e59878db9648cf460cc0cb3e14764839fc38daca1142d92c0338ba.jpg) + +![](images/7f9e250c5ef6b67abd6d4da20aae36aafdb1423788bf0df45d12cbb395b7044f.jpg) + +![](images/c95b2d84afba7b2d5c8e40b028fb0b906e8daf9d397aaf43163714ad46b34c22.jpg) +LSM + +![](images/1ea02be0c05163d27521690c3739ce0aa84102404d904d5e5598e988d0159a0d.jpg) + +![](images/2e6e71891dfe7d144520c276881144b2b5e88fd7eb482ceab9b96dbcb6f6b10b.jpg) + +![](images/7faabe2266a7e0afd7c2a90f5fe6294431916025fe68be408d11b7b726c6de34.jpg) + +![](images/b29f3d03b5c4681e1d810b9481cc8b132d72bcac1ed8cfb348ee3dd7b4f3425b.jpg) +Ours + +Wall + +Floor + +Ceiling + +Chair + +Table + +Sofa + +Bed + +Others + +![](images/ca7c174c9fb3d07f184ee41347586bbfa6f39dd3739b5cb98f78dd63e782f838.jpg) +Figure 4. Qualitative comparison in OVS. SpatialSplat achieves sharper and more precise segmentation results compared to previous methods. Notably, our method excels in challenging details, such as distinguishing table legs from cabinet legs (the second the fourth rows), benefiting from our dual-field architecture that captures detailed instance information and uncompressed semantics. + +![](images/4c113e95933606ccd2d4127d5f76d5e69bd0790d38a029607e350be288b2a603.jpg) + +![](images/3c16f72faef771f03962282832f99cd57bca07de508a8d58e73f5c79a9c14ea5.jpg) + +![](images/36b23c56dbe38b8e8418290f0857dfa2893980787e123ee5de8ea9c1ca7da502.jpg) + +![](images/7d9c75b686eb848c6d1f20f9f86827ce070e1b65a8e58fd5c2fa70854aa21a5a.jpg) +Input views + +![](images/b5b380a8713197993c942e565f6164fb9dbfaf9aa29a9de7efbe5f0ff34b47ae.jpg) + +![](images/af49e80d9c25a03f76b1186f7fc03f6a30b46ffe7a3d842092e4957ef3899a6b.jpg) +Rendered Feature Field + +![](images/890a7fff8cd638e4343a3b08e0e7f567dfe2c929ce548d82bdd6ec3235964f98.jpg) +Figure 5. The rendered instance features. SpatialSplat predicts clear and consistent instance features across different views. + +cent SOTA methods designed specifically for novel view synthesis. This highlights that our method efficiently learns spatial priors from large-scale 2D data. + +Results of Open-vocabulary 3D Segmentation. Following LSM's approach, we map category labels from the Scannet dataset into common categories: Wall, Floor, Ceiling, Chair, Table, Bed, Sofa, Others. As shown in Tab. 1, SpatialSplat outperforms all compared methods, even surpassing L-Seg, which provides semantic feature supervision + +![](images/ecfc7f0f9a6cfff07e5803e25f02598567cec7a8cb7f954f217a8ffb4bd89b9f.jpg) +GT + +![](images/f1d69e9a6c208387cf34ac932f4e7fa58a70fcf7d1d91d5a8403b122ed7fbbe6.jpg) + +![](images/90a840d68ed14cab321e553e3778c2fd63aaaf9c077ecc1f233cdac6d598b033.jpg) +LSM + +![](images/6ba58f3ba2c4146e33b6eed91c252afbea4f2252342e35c0455dc3f894ac5970.jpg) + +![](images/b4d8cc431ce151221ac3b1a628fe788e828f2068225a9186128bf10edb55c251.jpg) +Ours +Figure 6. Qualitative results of cross-dataset generalization. Zoom out for clearer visualization. + +![](images/b1e3b597510dec67cf74c1e6234c02f36267ef681a17af71fd4bfa63563ddddc.jpg) + +![](images/4db8a01d7bd9e668291b666fe07a180868b7690153095e7398b2b40dbd053a4a.jpg) +LSM + +![](images/0235ce2885724e6deaa21ddcc12c2db85541c222f452b0ac90a3a66afd6d6ffa.jpg) + +![](images/57d4e09f8473e4a30f7c46cccbeb00cf64d159431fde016c8f913692c7a664c2.jpg) +Ours + +for other compared methods. As illustrated in Fig. 4, SpatialSplat produces sharp and clear 3D semantic segmentation results. This is achieved by lifting both 2D instance and uncompressed semantic into 3D space, enabling more efficient semantic learning while mitigating the blurred contours and information loss associated with per-primitive compressed semantics. The visualizations in Fig. 5 show that SpatialSplat effectively learns a highly consistent instance prior with only weak supervision from 2D masks. + +Results of Cross-Dataset Generalization. As shown in Tab. 2, our method significantly outperforms LSM on the + +![](images/110b60505e6c0e8e2a5d73b2d2d68a13dc0c645fb04d1e476700ccd03eb96b8d.jpg) +(a) + +![](images/c17300b3b9cedf4b8fe5829d22b52477b3c2a1de4159f1a88d42633e140fce23.jpg) +(b) + +![](images/c902820742d14d40e0584a5dc351de0673b65c7c299421fcf585c3df1868331f.jpg) +(c) +Figure 7. Qualitative results of ablations. (a) and (b) Qualitative results of importance score prediction, with red color indicating an importance score of 1 and blue indicating 0. (c) The rendered semantic feature without the dual-field. (d) The rendered semantic feature with the dual-field, which appears clearer and sharper. + +![](images/40a6872b75f51c93a8078623b49001509f0c67302839ef6653f4809d65606e28.jpg) +(d) + +
MethodLatency↓Gaussian Size ↓Num. ↓
Feature-3DGS [52]1069 s229.65 MB476K
NoPoSplat [46]0.044 s30.93 MB131K
LSM [15]0.104 s63.00 MB131K
SpatialSplat-Lite0.058 s25.58 MB87.7K
SpatialSplat0.071 s25.58 MB87.7K
+ +unseen Replica dataset in both novel view synthesis and open-vocabulary segmentation. Notably, since Replica is a synthetic dataset with a different data modality from our training set, this underscores the strong generalization ability of our approach. The visualization in Fig. 6 further demonstrates that LSM struggles with coarse modeling at scene edges, leading to artifacts in synthesized views, whereas our method generates photorealistic renderings. + +Analysis of Model Efficiency. We compare the model efficiency of different 3DGS-based methods. As shown in Tab. 3, our method is faster than LSM thanks to its streamlined architecture, requiring only $40\%$ of the storage size and $65\%$ of primitive number. Remarkably, our approach requires even fewer Gaussians parameters than NoPoSplat, which lacks semantic awareness. This demonstrates the effectiveness of our proposed dual-field architecture and selective Gaussian mechanism, which capture accurate semantics while eliminating redundant Gaussians. Furthermore, as our method does not rely on dense semantic supervision, we leverage a lightweight pretrained 2D model, significantly accelerating inference speed. + +# 4.3. Ablations and Analysis. + +We perform ablations to answer the following questions: (1) Are the primitives removed by our selective Gaussian mechanism truly redundant? (2) Does the improved scene understanding stem from the proposed dual-field architecture? (3) As the number of views increases, how does redundancy grow? Can our method effectively alleviate it? + +Effect of selective Gaussian mechanism. To answer Question 1, we remove the SGM and retrain our model using the same configuration. As shown in Tab. 4, the SGM + +Table 3. Model efficiency comparison. "Lite": the model replaces LSeg with CLIP ViT-B/16. + +
ModelmIoU↑Acc.↑PSNR↑Num. ↓
No SGM0.55690.797225.55131k
No dual0.50270.778625.4687.6k
3-view0.51250.779324.9297.5k
SpatialSplat0.55770.792025.4787.7k
+ +Table 4. Ablation study on design choices. The proposed Adaptive Gaussian mechanism and dual-field architecture jointly enable the construction of high-quality representations with significantly reduced dimensionality. + +eliminates approximately $35\%$ of Gaussian primitives while causing only a 0.08 drop in PSNR, indicating that the removed primitives have minimal impact on rendering quality. The visualization in Fig. 7 (a) and (b) further demonstrates that the SGM accurately localizes overlapping areas in the input images. + +Effect of dual-field architecture. To address Question 2, we replace our dual-field architecture with single-level per-primitive semantic learning. Following the approach in LSM, we set the semantic feature dimension to 64 and use a convolutional layer to expand it to 512. As shown in Tab. 4, this results in a significant drop of $5\%$ in mIoU for open-vocabulary segmentation performance. The primary issue is that per-primitive semantic learning struggles to maintain accurate semantics and fails to preserve clear instance boundaries, as illustrated in Fig. 7 (c) and (d). + +Extend to more views. Our experiments primarily focus on two-view settings to ensure a fair comparison with our baseline. However, our method can be extended to more views for reconstructing larger scenes. As shown in Tab. 4, our approach remains effective even with three input views. Moreover, compared to pixel-wise methods, which require approximately 197K primitives to represent the scene, our method achieves a more compact representation by reducing about $50\%$ primitives. More qualitative details can be found in the appendix. + +# 5. Conclusion + +We present an efficient feed-forward network that generates compact yet high-performance representations for 3D reconstruction and scene understanding from sparse, unposed images. Our proposed dual-field architecture achieves state-of-the-art performance while significantly reducing the number of representation parameters. Additionally, we introduce a Selective Gaussian Mechanism that directly identifies redundant Gaussians caused by pixel-wise predictions without requiring geometric priors. SpatialSplat achieves impressive results in 3D reconstruction and scene understanding without relying on any 3D data during training or inference, making semantic 3D reconstruction from sparse, unposed images more practical. + +# Acknowledgments + +This work was partially supported by Hunan Province Major Scientific and Technological Project No. 2024QK2001, National Key R&D Program of China No. 2023YFB4704500. + +# References + +[1] Yash Bhalgat, Iro Laina, João F. Henriques, Andrea Vedaldi, and Andrew Zisserman. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9630-9640. IEEE, 2021. 2 +[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19457-19467. IEEE, 2024. 2, 3, 4, 6 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2 +[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. CoRR, abs/2401.03890, 2024. 1 +[6] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: hash-grid assisted context for 3d gaussian splatting compression. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part VII, pages 422-438. Springer, 2024. 3 +[7] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXI, pages 370-386. Springer, 2024. 2, 3, 4 +[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5 +[9] Jianmei Dai, Zhilong Zhang, Shiwen Mao, and Danpu Liu. A view synthesis-based $360^{\circ}$ VR caching system over mec-enabled C-RAN. IEEE Trans. Circuits Syst. Video Technol., 30(10):3843-3855, 2020. 1 + +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4970-4980. IEEE, 2023. 2 +[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2 +[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. CoRR, abs/2403.20309, 2024. 2 +[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ FPS. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 3 +[15] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, Boris Ivanovic, and Marco Pavone. Large spatial model: End-to-end unposed images to semantic 3d. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 1, 2, 3, 4, 5, 8 +[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 12479-12488. IEEE, 2023. 2 +[17] Qiao Gu, Zhaoyang Lv, Duncan P. Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIII, pages 382-400. Springer, 2024. 3 +[18] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18344-18347. IEEE, 2022. 2 + +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 4, 6 +[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19672-19682. IEEE, 2023. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3992-4003. IEEE, 2023. 4, 5 +[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 2 +[23] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 5, 6 +[24] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21719-21728. IEEE, 2024. 3 +[25] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 6 +[26] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 5, 6 +[27] Luca De Luigi, Damiano Bolognini, Federico Domeniconi, Daniele De Gregorio, Matteo Poggi, and Luigi Di Stefano. Scannerf: a scalable benchmark for neural radiance fields. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 816-825. IEEE, 2023. 3 +[28] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 4 +[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2022. 1, 2, 6 +[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires + +olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[31] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20051-20060. IEEE, 2024. 2, 3, 4 +[32] Ri-Zhao Qiu, Yafei Hu, Ge Yang, Yuchen Song, Yang Fu, Jianglong Ye, Jiteng Mu, Ruihan Yang, Nikolay Atanasov, Sebastian A. Scherer, and Xiaolong Wang. Learning generalizable feature fields for mobile manipulation. CoRR, abs/2403.07563, 2024. 1 +[33] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLI, pages 368-383. Springer, 2024. 2, 3 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 5 +[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 12159-12168. IEEE, 2021. 3 +[36] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In Conference on Robot Learning, CoRL 2023, 6-9 November 2023, Atlanta, GA, USA, pages 405-424. PMLR, 2023. 1, 2, 4 +[37] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 9043-9052. IEEE, 2023. 2 +[38] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2, 4, 3 +[39] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5 + +[40] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In International Conference on 3D Vision, 3DV 2022, Prague, Czech Republic, September 12-16, 2022, pages 443-453. IEEE, 2022. 2 +[41] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4690-4699. Computer Vision Foundation / IEEE, 2021. 2 +[42] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2 +[43] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs. CoRR, abs/2412.16932, 2024. 2 +[44] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2, 1 +[45] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20041-20050. IEEE, 2024. 2 +[46] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4, 5, 6, 8, 3 +[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 5 +[48] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4578-4587. Computer Vision Foundation / IEEE, 2021. 2 +[49] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. CoRR, abs/2408.13770, 2024. 2 +[50] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586- + +595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 5 +[51] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15818-15827. IEEE, 2021. 2 +[52] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21676-21685. IEEE, 2024. 2, 3, 4, 5, 6, 8 +[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 1 \ No newline at end of file diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/images.zip b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68aada784007c8f95980b6cfdeb176749170e7a1 --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba738b92401784d13a1174434a57d4592c02cbe1c6535f34c1ade493e8a8785b +size 705656 diff --git a/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/layout.json b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b18048906d19fd93e2a7bce5d9617a11c868b681 --- /dev/null +++ b/2025/SpatialSplat_ Efficient Semantic 3D from Sparse Unposed Images/layout.json @@ -0,0 +1,10605 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "text", + "content": "SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": "Yu Sheng" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Jiajun Deng" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Xinran Zhang" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Yu Zhang" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Bei Hua" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Yanyong Zhang" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": ", Jianmin Ji" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": "University of Science and Technology of China, " + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 141, + 536, + 171 + ], + "type": "text", + "content": "The University of Adelaide" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 173, + 476, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 173, + 476, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 173, + 476, + 185 + ], + "type": "text", + "content": "{shengyu724, zxrr}@mail.ustc.edu.cn, jiajun.deng@adelaide.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 134, + 186, + 473, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 186, + 473, + 199 + ], + "spans": [ + { + "bbox": [ + 134, + 186, + 473, + 199 + ], + "type": "text", + "content": "{zhangyu,bhua,yanyongz,jianmin}@ustc.edu.cn †: Corresponding Author" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 230, + 248, + 361 + ], + "blocks": [ + { + "bbox": [ + 67, + 230, + 248, + 361 + ], + "lines": [ + { + "bbox": [ + 67, + 230, + 248, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 230, + 248, + 361 + ], + "type": "image", + "image_path": "f3c34c3daae2b917c496310f5f0b8e4dc6853fe327dbc647ebc24d84048344a9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Comparison between previous methods and our SpatialSplat. (a): Previous methods predict pixel-wise Gaussians, associating each primitive with compressed semantic feature. (b): Our SpatialSplat avoids generating redundant Gaussian primitives for overlapping pixels, and it represents semantics with a dual-field architecture to better preserve the information. (c): SpatialSplat outperforms the state-of-the-art method LSM [15] for both novel-view rendering quality and semantic segmentation while being more efficient." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 250, + 230, + 443, + 361 + ], + "blocks": [ + { + "bbox": [ + 250, + 230, + 443, + 361 + ], + "lines": [ + { + "bbox": [ + 250, + 230, + 443, + 361 + ], + "spans": [ + { + "bbox": [ + 250, + 230, + 443, + 361 + ], + "type": "image", + "image_path": "fa58bb773a8f08e4e6a8f901883311c541a0e7032df8b973f9f276c359595d6b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 446, + 230, + 543, + 361 + ], + "blocks": [ + { + "bbox": [ + 446, + 230, + 543, + 361 + ], + "lines": [ + { + "bbox": [ + 446, + 230, + 543, + 361 + ], + "spans": [ + { + "bbox": [ + 446, + 230, + 543, + 361 + ], + "type": "image", + "image_path": "b3daf1c37c562445242a2565d5df8e75f09140300e699cd05c9308e8989f6a5b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "spans": [ + { + "bbox": [ + 152, + 427, + 200, + 440 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "content": "A major breakthrough in 3D reconstruction is the feedforward paradigm to generate pixel-wise 3D points or Gaussian primitives from sparse, unposed images. To further incorporate semantics while avoiding the significant memory and storage costs of high-dimensional semantic features, existing methods extend this paradigm by associating each primitive with a compressed semantic feature vector. However, these methods have two major limitations: (a) the naively compressed feature compromises expressiveness, affecting the model's ability to capture fine-grained semantics, and (b) the pixel-wise primitive prediction introduces redundancy in overlapping areas, causing unnecessary memory overhead. To this end, we introduce SpatialSplat, a feedforward framework that produces redundancy-aware Gaussians and capitalizes on a dual-field semantic representation. Particularly, with the insight that primitives within the same instance exhibit high semantic consistency, we decompose the semantic representation into a coarse feature field that encodes uncompressed semantics with minimal primitives, and a fine-grained yet low-dimensional feature field that captures detailed inter" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "content": "instance relationships. Moreover, we propose a selective Gaussian mechanism, which retains only essential Gaussians in the scene, effectively eliminating redundant primitives. Our proposed Spatialplat learns accurate semantic information and detailed instances prior with more compact 3D Gaussians, making semantic 3D reconstruction more applicable. We conduct extensive experiments to evaluate our method, demonstrating a remarkable " + }, + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "content": " reduction in scene representation parameters while achieving superior performance over state-of-the-art methods. The code will be made available for future investigation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 585, + 395, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 585, + 395, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 585, + 395, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Reconstructing and understanding 3D scenes from 2D images [5, 9, 32, 36, 53] is a fundamental topic in computer vision, aiming to obtain semantic-aware 3D structure from low-cost devices, i.e., RGB cameras. This technique is significant for various applications, such as robotics, autonomous driving, and VR/AR. Recently, with the emergence of the new 3D representation, i.e., Neural Radiance Fields (NeRF) [29] and 3D Gaussian Splatting (3DGS) [19], a popular paradigm of performing semantic-aware 3D re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26404" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "construction is to distill the feature field of NeRF or 3DGS with powerful 2D vision-language foundation models [20, 31, 33, 52]. However, these methods typically rely on perscense optimization and complex multi-step preprocessing, limiting their ability to generalize across multiple scenes within a single model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 294, + 266 + ], + "type": "text", + "content": "Recent breakthroughs [3, 7, 49] in 3D reconstruction have incorporated feed-forward networks to improve the generalization of 3DGS models and accelerate reconstruction. These methods follow a paradigm that generates pixel-wise Gaussian primitives from sparse posed images. Building on this, further efforts [38, 46] have extended the paradigm to reconstruct scenes from unposed images. Moreover, methods [15, 43] have advanced this approach by integrating semantic understanding, enabling semantic 3D reconstruction from sparse unposed images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 267, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 295, + 483 + ], + "type": "text", + "content": "Despite significant progress, these methods have two major limitations. First, as shown in Fig.1 (a), pixelwise Gaussian prediction introduces substantial redundancy in overlapping regions while providing minimal accuracy gains. Method [44] attempts to mitigate this issue by projecting primitives onto a plane and identifying overlaps, it depends on precise camera extrinsics, which are often difficult to obtain in real-world scenarios—especially when only a few sparse, textureless images are available. Second, methods [15, 31, 33, 43, 52] enforce per-primitive semantic learning to form a dense semantic feature field, as shown in Fig.1 (a). This primitive-level approach incurs significant memory and storage costs due to high-dimensional (512 or more) language features. To mitigate this, the methods often compress features into lower-dimensional representations (e.g., 64-128), but this compression leads to irreversible information loss [52], leading to suboptimal scene understanding performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 294, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 294, + 628 + ], + "type": "text", + "content": "In this paper, we take a slightly different viewpoint. First, inspired by MASt3R [25], which successfully extracts 3D geometry and pixel correspondences from images, we observe that redundant primitives often share similar geometry and appearance, allowing them to be identified directly from images without relying on complex geometric priors. Second, we find that per-primitive semantics are not essential for high-performance scene understanding. Instead, a coarse semantic representation, when combined with fine-grained instance information, is sufficient for strong performance, as Gaussian primitives within the same instance exhibit high semantic consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "By consolidating this idea, we introduce SpatialSplat, a feed-forward network generating compact yet expressive semantic 3D representations from unposed images. As depicted in Fig. 1 (b), SpatialSplat introduces a dual-field architecture that decomposes the dense semantic feature field into two components: a coarse feature field encoding uncompressed instance-level semantics through mini" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "mal Gaussians, and a low-dimensional fine-grained feature field capturing inter-instance relationships. Guided by multiple 2D foundation models, our method learns both accurate semantic and instance priors, achieving superior performance while reducing representation parameters. Additionally, we introduce a Selective Gaussian Mechanism (SGM) to eliminate redundancy in overlapping areas caused by pixelwise representations, along with a novel loss function that jointly optimizes redundancy-aware Gaussians and scene fidelity. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple downstream 3D tasks while using only " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " of the representation parameters required by the baseline, all without any 3D supervision. Our contributions are threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 240, + 553, + 348 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 240, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 553, + 287 + ], + "type": "text", + "content": "- A novel feed-forward 3DGS framework that, to the best of our knowledge, is the first to simultaneously learn semantic and instance priors with guidance from foundation models while maintaining a high-fidelity radiance field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 323 + ], + "type": "text", + "content": "- A dual-field semantic representation that significantly reduces storage consumption while enhancing open-vocabulary 3D segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 348 + ], + "type": "text", + "content": "- A mechanism that identifies redundant primitives from images without relying on any geometric priors." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 361, + 400, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 400, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 400, + 373 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 491, + 392 + ], + "type": "text", + "content": "2.1. Feed-forward 3D Reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 553, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 589 + ], + "type": "text", + "content": "The emergence of NeRF [29] and 3DGS [19] has significantly transformed the paradigm of 3D reconstruction from images. Despite achieving remarkable results in 3D reconstruction and novel view synthesis, these methods rely on time-consuming per-scene optimization, with even improved techniques [4, 16, 30] still requiring several minutes to hours. To overcome these limitations, recent approaches [3, 7, 11, 18, 41, 45, 48, 49] have explored using a single feed-forward network to learn reconstruction priors from large datasets, enabling generalizable 3D reconstruction. DUSt3R serials [25, 42] further simplifies the reconstruction pipeline, making 3D reconstruction possible from images without relying on camera extrinsics. This advancement has inspired a new wave of methods [13, 38, 46] capable of achieving dense reconstruction and view synthesis without camera extrinsics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 600, + 454, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 454, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 454, + 611 + ], + "type": "text", + "content": "2.2. Feature Field Distillation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Early approaches of feature field distillation were largely based on NeRF [29]. For example, methods [37, 51] successfully embedded 2D label data into 3D space, achieving sharp and precise 3D segmentation. Later approaches [12, 20, 22, 36, 40] incorporated more complex pixel-aligned feature vectors from technologies like LSeg [26] or DINO [2] into NeRF frameworks, enabling open-vocabulary 3D segmentation. With the rise of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26405" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 94, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 59, + 94, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 59, + 94, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 59, + 94, + 553, + 235 + ], + "type": "image", + "image_path": "3ca3c1e5b776457dda56ff2b09058f990e66f96beb2ee228f79f5ad9e4ab5796.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "type": "text", + "content": "Figure 2. Pipeline of SpatialSplat. The SpatialSplat processes unposed images along with their intrinsicities through a 3D geometry transformer. The extracted features from the geometry transformer and the pretrained 2D model are then fed into separate Dense Prediction Transformer (DPT) heads to predict different Gaussian attributes, resulting in a fine-grained instance-aware radiance field " + }, + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_I" + }, + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "type": "text", + "content": " and a coarse semantic feature field " + }, + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_S" + }, + { + "bbox": [ + 54, + 237, + 555, + 283 + ], + "type": "text", + "content": ". This enables the synthesis of RGB and feature maps from novel viewpoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 287, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 294, + 407 + ], + "type": "text", + "content": "3DGS [19], methods such as [17, 31, 33, 52] have extended this concept by embedding language features into Gaussian attributes to enhance semantic understanding. Large Spatial Model (LSM) [15] was the first to achieve a generalizable semantic radiance field based on feed-forward 3DGS. However, due to the explicit nature of 3DGS representations, attaching feature vectors to each primitive incurs high memory costs. To mitigate this, existing methods often compress semantic features into lower-dimensional representations, inevitably resulting in information loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 456 + ], + "type": "text", + "content": "In contrast, our approach decomposes semantics into coarse instance-level semantics and fine-grained instance information, achieving superior performance with significantly lower storage requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 183, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 183, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 183, + 474 + ], + "type": "text", + "content": "2.3. Compact 3D Gaussian" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 574 + ], + "type": "text", + "content": "Our work is also related to compact 3D Gaussian representation. Recent works have attempted to compress 3DGS by exploring the spatial relationships among Gaussians. Scaffold-GS [27] introduces anchor-centered features to reduce parameter counts, while HAC [6] builds on this by incorporating a hash grid to better organize scene primitives. Methods [14, 24] utilize techniques like distillation and Gaussian pruning to minimize storage requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 295, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 657 + ], + "type": "text", + "content": "Notably, our focus is on addressing the additional representational redundancy introduced by the feed-forward pipeline (i.e., overleaping Gaussians) rather than the inherent redundancy of 3DGS itself. This issue is prevalent across existing generalizable 3DGS methods, and the aforementioned compression techniques are not suitable for these frameworks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 667, + 111, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 111, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 111, + 679 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Overview. As illustrated in Fig.2, SpatialSplat leverages a standard Vision Transformer (ViT) [10] backbone with a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": "few Dense Prediction Transformer (DPT) [35] heads. Given " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": " unposed images at resolution " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": " with their intrinsic " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\{I^v,K^v\\}_{v = 1}^V" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": ", SpatialSplat learn a mapping " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": " parametered by " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": " that transforms the input images to an fine-grained instance-aware radiance field " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_I" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": " and a coarse semantic feature field " + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_S" + }, + { + "bbox": [ + 313, + 288, + 555, + 360 + ], + "type": "text", + "content": ", which can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 367, + 456, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 367, + 456, + 382 + ], + "spans": [ + { + "bbox": [ + 321, + 367, + 456, + 382 + ], + "type": "interline_equation", + "content": "f _ {\\theta}: \\left\\{\\boldsymbol {I} ^ {v}, \\boldsymbol {K} ^ {v} \\right\\} _ {v = 1} ^ {V} \\rightarrow \\mathcal {F} _ {I} + \\mathcal {F} _ {S},", + "image_path": "5b3e0e4d9e258ff8e1e80dafe7af7aac9c759eb4e6217639854a6f2c4b2a4fdb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 384, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 384, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 322, + 384, + 553, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {\\boldsymbol {I}} = \\left\\{\\cup \\left(\\boldsymbol {f} _ {\\boldsymbol {I}} ^ {i}, \\boldsymbol {\\beta} ^ {i}, \\boldsymbol {\\mu} ^ {i}, \\boldsymbol {\\alpha} ^ {i}, \\boldsymbol {r} ^ {i}, \\boldsymbol {s} ^ {i}, \\boldsymbol {c} ^ {i}\\right) \\right\\} _ {i = 1, \\dots , V \\times H \\times W}, \\tag {1}", + "image_path": "dbb6a9627e5287a742d4ec094ebda7d43d2f39e7c3eaeef7d4190208ee5aa052.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 401, + 529, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 401, + 529, + 418 + ], + "spans": [ + { + "bbox": [ + 322, + 401, + 529, + 418 + ], + "type": "interline_equation", + "content": "\\mathcal {F} _ {S} = \\{\\cup (\\boldsymbol {f} _ {S} ^ {j}, \\boldsymbol {\\mu} ^ {j}, \\boldsymbol {\\alpha} ^ {j}, \\boldsymbol {r} ^ {j}, \\boldsymbol {s} ^ {j}, \\boldsymbol {c} ^ {j}) \\} _ {j = 1, \\dots , V \\times \\frac {H}{S} \\times \\frac {W}{S}}.", + "image_path": "53d574480a9ce46aac297c237829d5c766bf311cc5b48f4deed15a76bc24b247.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": " is the importance score, which will be explained in Section 3.2. " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{f}_I \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": " is the learnable instance feature and " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{f}_S \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": " is the learnable semantic feature, as detailed in Section 3.3. The remains represent the vanilla Gaussian parameters [19], including the center position " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": ", rotation factor in quaternion " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{r} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": ", and spherical harmonics (SH) " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\pmb{c} \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 425, + 555, + 533 + ], + "type": "text", + "content": " degrees of freedom. In the following sections, we provide a detailed explanation of each component of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 540, + 451, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 451, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 451, + 553 + ], + "type": "text", + "content": "3.1. 3D Geometry Prediction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Both the encoder and decoder in our geometric prediction module are built on pure ViT structures, requiring no geometric priors as in previous methods [3, 7]. The input image is patchified and flattened into image sequences, which along with the camera intrinsics processed by a linear layer, are fed into the encoder. The encoder weights are shared across different input views. The features from encoder are then passed to a ViT-based decoder, where cross-attention is applied to better capture spatial relationships and aggregate information across views. Finally, the features from different views are separately fed into a series of DPT heads to predict point maps and other Gaussian parameters. The point maps serve as Gaussians centers. Since the input im" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26406" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "ages are unposed, the Gaussians centers suffer from scale ambiguity. Some methods [15, 38] rely on depth supervision to obtain per-scene scale, but this introduces additional errors due to imperfect depth. To overcome this limitation, we follow NoPoSplat's [46] approach by injecting camera intrinsics to resolve scale ambiguity. Experiments show that SpatialSplat effectively learns 3D priors from sparse unposed images without depth supervision, even while jointly learning multiple parameters and features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 222, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 222, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 222, + 201 + ], + "type": "text", + "content": "3.2. Selective Gaussian Mechanism" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 399 + ], + "type": "text", + "content": "Previous methods [3, 7, 15, 38, 46] directly use the pixelwise pointmap as Gaussian centers which introduce significant redundancy. To address this, we propose a selective Gaussian mechanism that assigns each primitive an importance score to quantify its necessity for the scene representation. Primitives with importance scores below a certain threshold are considered redundant and discarded. To effectively supervise the learning of importance scores, we analyze Gaussian properties and the rendering pipeline, designing a simple yet effective strategy. Specifically, a primitive's influence can be reduced by decreasing its opacity or size [19]. However, we observe that even primitives with near-zero size still contribute to the rendering by occupying a pixel. Therefore, we multiply the importance score to the opacity of each primitive and modify the alpha blending formulation as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 407, + 295, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 407, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 108, + 407, + 295, + 441 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} = \\sum_ {i = 1} ^ {n} \\boldsymbol {c} _ {i} \\boldsymbol {\\alpha} _ {i} \\boldsymbol {\\beta} _ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\boldsymbol {\\alpha} _ {j} \\boldsymbol {\\beta} _ {j}), \\tag {2}", + "image_path": "c94b36f8e7e13b15f64c2d2f83efef89d8bad75ac7c0a61887407c354a956c34.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "content": " is the final intensities calculated by blending " + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "content": " ordered Gaussians overlapping the pixels, and " + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "content": " is the importance score processed through a sigmoid activation function. This ensures that primitives with low importance have minimal impact on the final intensities. Removing important primitives leads to a significant increase in photometric loss, whereas discarding redundant ones has little effect. Therefore, we optimize " + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 55, + 450, + 296, + 571 + ], + "type": "text", + "content": " through photometric loss minimization. Following methods [15, 46], we define the photometric loss as a weighted sum of MSE and LPIPS [50]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 591, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 591, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 68, + 591, + 295, + 607 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) = \\mathcal {L} _ {M S E} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda \\mathcal {L} _ {L P I P S} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}), \\tag {3}", + "image_path": "c5f98944dc8033418061a3d52a7d7dfad470c2df19c2cb081c49c11be493772f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "text", + "content": " are rasterized and GT pixel intensities, and " + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 55, + 616, + 296, + 714 + ], + "type": "text", + "content": ". Since redundant primitives are directly discarded during inference, we encourage their importance scores to approach zero to minimize the discrepancy between training and inference. Simultaneously, we aim to reduce the number of primitives used, striving to eliminate redundant ones as much as possible. To achieve these objectives, we first design the importance score with inspiration from Leaky" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 455, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 455, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 455, + 83 + ], + "type": "text", + "content": "ReLU [28], formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 356, + 91, + 555, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 91, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 356, + 91, + 555, + 125 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\beta} _ {i} = \\left\\{ \\begin{array}{l l} \\boldsymbol {\\beta} _ {i} & \\text {i f} \\boldsymbol {\\beta} _ {i} > \\tau \\\\ \\boldsymbol {\\beta} _ {i} \\times 1 0 ^ {- 3} & \\text {i f} \\boldsymbol {\\beta} _ {i} < \\tau \\end{array} \\right.. \\tag {4}", + "image_path": "ea7b87e3e0d8bd421475b6b8ec52c6e3eb072dac3eb59dfdad2014a991ac4ec3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "content": "This ensures that the importance scores of primitives below the threshold " + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "content": " are close to zero while preventing gradient vanishing. Additionally, we introduce a Binary Cross Entropy (BCE) loss with a regularization term that pushes " + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 313, + 132, + 555, + 192 + ], + "type": "text", + "content": " to either 0 or 1 while penalizing excessive use of primitives:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 199, + 555, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 199, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 351, + 199, + 555, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} (\\boldsymbol {S}) = \\mathcal {L} _ {B C E} (\\boldsymbol {S}, \\hat {\\boldsymbol {S}}) + \\frac {1}{\\| \\boldsymbol {S} \\|} \\sum_ {\\boldsymbol {\\beta} _ {i} \\in \\boldsymbol {S}} \\beta_ {i}, \\tag {5}", + "image_path": "db56a3878ed4f61ba71c19dacacd71f3e1b1c4c785aa540d33087c05d40f1628.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 234, + 342, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 234, + 342, + 244 + ], + "spans": [ + { + "bbox": [ + 314, + 234, + 342, + 244 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 247, + 524, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 247, + 524, + 279 + ], + "spans": [ + { + "bbox": [ + 343, + 247, + 524, + 279 + ], + "type": "interline_equation", + "content": "\\boldsymbol {S} = \\{\\beta_ {i} \\} _ {i = 1} ^ {H \\times W}, \\quad \\hat {\\boldsymbol {S}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\beta_ {i} > \\tau \\\\ 0 & \\text {i f} \\beta_ {i} < \\tau \\end{array} \\right..", + "image_path": "8c660a0260d5033c9359e3e3b8ccbb12d2293d55953677512ac1cc127501050e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": "As a result, SpatialSplat achieves a compact yet high-fidelity scene representation by selectively retaining only the most essential primitives." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 330, + 446, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 330, + 446, + 341 + ], + "spans": [ + { + "bbox": [ + 314, + 330, + 446, + 341 + ], + "type": "text", + "content": "3.3. Dual-field Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 346, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 479 + ], + "type": "text", + "content": "While distilling 2D semantic features into dense 3D representations has proven effective for scene understanding, conventional per-primitive compressed feature assignments [31, 36, 52] result in information loss and suboptimal performance. To mitigate this loss without increasing storage costs, we propose a dual-field architecture that decouples semantic representation into: 1) a fine-grained instance-aware radiance field, capturing scene geometry, textures, and instance correspondences, and 2) a coarse, uncompressed semantic feature field that reduces storage requirements while preserving semantic accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": "Fine-grained Instance-Aware Radiance Field. Our instance-aware radiance field builds upon Vanilla 3DGS, initializing the redundancy-aware Gaussians augmented with learnable instance features " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{f}_I" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": ". We train the network using contrastive learning [1] to pull rendered instance features of the same instance closer while push those of different instances further apart. To achieve this, we lift multiview 2D segmentation priors into 3D to help SpatialSplat learn inter-instance relationships. Given a target image " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": ", we first use Segment Anything Model (SAM) [21] to extract a set of masks " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M^k \\mid k = 1, 2, \\dots, m\\}" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": ". The instance feature map rendered at target image view with Eq. 2 is denoted as " + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "inline_equation", + "content": "F_I" + }, + { + "bbox": [ + 313, + 480, + 556, + 635 + ], + "type": "text", + "content": ". The contrastive loss is then defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 643, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 643, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 643, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {C} \\left(\\boldsymbol {F} _ {I}, \\mathcal {M}\\right) = - \\log \\frac {1}{\\| \\boldsymbol {I} \\| ^ {2}} \\sum_ {u \\in \\boldsymbol {I}} \\sum_ {u ^ {\\prime} \\in \\boldsymbol {I} ^ {+}} \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right) \\\\ + \\sum_ {u ^ {\\prime} \\notin I ^ {+}} \\left(1 - \\operatorname {s i m} \\left(\\boldsymbol {F} _ {I} (u), \\boldsymbol {F} _ {I} \\left(u ^ {\\prime}\\right)\\right)\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "524300a46f67d0d69b55bc2247bbb5fd7032775f94abf88318983b2c893f0778.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26407" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 79, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 58, + 79, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 553, + 218 + ], + "type": "table", + "html": "
MethodsSemanticForwardSource ViewTarget View
mIoU ↑Acc. ↑mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
L-Seg [26]✓X0.55410.78240.55580.7856N/AN/AN/A
NeRF-DFF [23]✓X0.53810.77240.51370.758222.490.76500.2829
Feature-3DGS [52]✓X0.49920.73620.32230.567417.960.58120.4894
NoPoSplat [46]X✓N/AN/AN/AN/A25.700.81590.1875
LSM [15]✓✓0.51410.77190.51040.766224.120.79610.2526
SpatialSplat-Lite✓✓0.52720.78010.52650.769325.450.80320.2039
SpatialSplat✓✓0.55930.78140.55870.792425.460.80450.2046
", + "image_path": "edc322fb8c5cf97cc7b9e8d09f8efe65e247459ed11f28c3282d8af028f9dec2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison in 3D Tasks on Scannet dataset. Our method outperforms both the latest SOTA semantic-aware feed-forward approach and per-scene optimization methods. \"Lite\": replacing the pretrained 2D model LSeg with CLIP ViT-B/16." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "inline_equation", + "content": "I^{+}" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": " is the set of pixels that belong to the same instance mask, " + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "inline_equation", + "content": "F_{I}(u)" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": " is the feature vector of the " + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "inline_equation", + "content": "F_{I}" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": " at coordinate " + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "inline_equation", + "content": "sim(. )" + }, + { + "bbox": [ + 54, + 254, + 294, + 362 + ], + "type": "text", + "content": " is the cosine similarity. With this weak supervision, SpatialSplat effectively clusters radiance primitives by instance affiliation. Notably, the loss complexity in Eq. 6 scales quadratically with the number of pixels, making training infeasible. To address this, we employ a trick (detailed in the appendix) that enables efficient loss estimation with linear complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "content": "Coarse Semantic Feature Field. Since primitives within the same instance share semantic consistency, SpatialSplat learns instance-level semantics using only a subset of primitives, avoiding the redundancy of per-primitive assignments. Manually selecting primitives for each instance, however, disrupts the network's cohesion and simplicity while increasing additional computational costs. To address this, we uniformly downsample the Gaussian centers from " + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_I" + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "content": " to form a coarse feature field, where each selected primitive is assigned a semantic feature " + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "content": ". To enhance semantic learning, we inject image features from a pretrained 2D model into feature heads, as shown in Fig. 2. A 2D semantic feature map " + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "content": " is rendered using alpha blending. We minimize the loss between the rendered feature map at a novel view and the feature map " + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\hat{F}_{S}" + }, + { + "bbox": [ + 56, + 364, + 296, + 555 + ], + "type": "text", + "content": " of the ground truth image extracted from the pretrained 2D model:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 561, + 295, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 561, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 107, + 561, + 295, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) = M S E \\left(\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}\\right) \\tag {7}", + "image_path": "a2459a042675a8fb2fb82071205aa506b2b3f4c326a1e93ee33472c726e76395.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "During open-vocabulary querying, we first cluster primitives in " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_I" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " by instance feature similarity, assigning each instance the average feature of its clustered primitives as its label. For primitives in " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_S" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", their labels correspond to the instance features before sampling, and their semantic features are assigned to the instance with the highest cosine similarity. If multiple primitives contribute semantic features to the same instance, we take their average to ensure consistency. This approach enables the network to acquire sharper and clearer semantics compared to the previous methods. An additional advantage of dual-field architecture is that, un" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 255, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 291 + ], + "type": "text", + "content": "like dense semantic supervision from LSeg [26], it allows the use of a much lighter pretrained model (e.g., CLIP ViT-B/16 [34]), significantly improving inference speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 297, + 425, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 425, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 425, + 311 + ], + "type": "text", + "content": "3.4. Training Objective" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 339 + ], + "type": "text", + "content": "Following previous methods [15], we perform end-to-end training to optimize our model with the following objective:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 352, + 357, + 553, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 357, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 352, + 357, + 553, + 388 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {P} (\\boldsymbol {C}, \\hat {\\boldsymbol {C}}) + \\lambda_ {1} \\mathcal {L} _ {I} (\\boldsymbol {S}) \\tag {8} \\\\ + \\lambda_ {2} \\mathcal {L} _ {C} (\\boldsymbol {F} _ {I}, \\mathcal {M}) + \\lambda_ {3} \\mathcal {L} _ {S} (\\boldsymbol {F} _ {S}, \\hat {\\boldsymbol {F}} _ {S}). \\\\ \\end{array}", + "image_path": "e1f0c496270b7718f8c268c0266660a8023dfa1a51b115c7f706991904d4d369.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "text", + "content": "The parameters " + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "text", + "content": " are set to 0.01, 0.2, and 1.0, respectively. To save training time, the instance masks " + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "text", + "content": " are generated by SAM [21] prior to training, while the semantic feature map " + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{F}}_S" + }, + { + "bbox": [ + 313, + 396, + 554, + 445 + ], + "type": "text", + "content": " is predicted during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 455, + 391, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 391, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 391, + 468 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 430, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 430, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 430, + 488 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 494, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 590 + ], + "type": "text", + "content": "Datasets. Following LSM[15], we primarily train our model on a combination of ScanNet++[47] and ScanNet [8]. We filter out bad scenes and those with incomplete extrinsic parameters, resulting in a training dataset of approximately 1,500 scenes. For evaluation, we follow LSM and select 40 unseen scenes from ScanNet to assess our model's performance. We test our method's cross-dataset generalization on the synthesized Replica Dataset [39]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 664 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate our method on two downstream tasks: novel view synthesis (NVS) and 3D open-vocabulary segmentation (OVS). For NVS, we adopt the standard metrics: PSNR, SSIM, and LPIPS [50]. For OVS, we evaluate performance using class-wise intersection over union (mIoU) and average pixel accuracy (mAcc)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Baselines. We primarily compare our method with LSM [15], the latest state-of-the-art (SOTA) approach for generalizable semantic 3D reconstruction. Additionally, we include the following baselines: 1): Feature-3DGS [52]" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26408" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 545, + 320 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 320 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 320 + ], + "type": "image", + "image_path": "d60ac0a66109e44e4515ecaa35ad6ec3713a7e144d6a2720f2d920be15de4595.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison in NVs. SpatialSplat can synthesize realistic novel views. In challenging cases where LSM fails, such as the table legs in the first two rows and the corners in the last two rows, our method achieves significantly better results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 355, + 552, + 450 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 552, + 450 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 552, + 450 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 552, + 450 + ], + "type": "table", + "html": "
SceneLSMSpatialSplat (Ours)
mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓mIoU ↑Acc. ↑PSNR ↑SSIM ↑LPIPS ↓
room00.50240.742317.610.52470.32280.51310.793422.090.67600.1794
room10.46610.734914.340.59130.31740.46910.762820.620.67740.2265
office30.54840.762721.280.75670.20840.55820.799722.520.79130.1436
office40.50650.772319.510.69620.27790.50160.769521.130.72710.2072
", + "image_path": "e1eb7ff0960d9899eb4dd6f8b09c2d0b8814762289a2e596c15bd2162d407e0e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 451, + 514, + 463 + ], + "lines": [ + { + "bbox": [ + 95, + 451, + 514, + 463 + ], + "spans": [ + { + "bbox": [ + 95, + 451, + 514, + 463 + ], + "type": "text", + "content": "Table 2. Out-of-distribution (OOD) comparison on Replica dataset. SpatialSplat generalizes well on OOD data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 480, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 480, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 480, + 295, + 577 + ], + "type": "text", + "content": "and NeRF-DFF [23], pre-scene optimization methods for semantic 3D reconstruction based on 3DGS [19] and NeRF [29], respectively. 2): NoPoSplat [46], a latest SOTA method for generalizable novel view synthesis. 3): LSeg [26], a 2D open-vocabulary segmenter used for feature distillation in all compared semantic-aware methods. Unless otherwise specified, our method also uses image features from LSeg for supervision." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Implementation details. For the 3D geometry prediction module, we use ViT-Large with a patch size of 16 as the encoder and ViT-Base as the decoder, both initialized with pre-trained weights from MAST3R [25], while the remaining modules are randomly initialized. We modify the 3DGS CUDA renderer to support instance and semantic feature map rendering. The importance score threshold " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " is set to 0.5, the downsampling ratio to 8, the instance feature dimension " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " to 8, and the semantic feature dimension " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": " to 512. For a fair comparison, we train the model at a resolution of " + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": ", consistent with our baseline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 480, + 434, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 434, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 434, + 493 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "Results of Novel View Synthesis. Following previous methods [3, 46], we select two images from each test scene as input view for the generalizable 3DGS methods, and choose four additional images located between these two as target views to evaluate the performance of different methods in Novel View Synthesis. Notably, Feature-3DGS [52] and NeRF-DFF [23] relies on dense view inputs and performs poorly with only two input images. To account for this, we use five out of the six available images (the two input images and four target views) for training and evaluate on the remaining image. As shown in Tab. 1, Spatial-Splat significantly outperforms latest SOTA method LSM. We observe that LSM struggles in certain scenes due to its reliance on accurate depth for aligning views during training as shown in Fig. 3. Although NeRF-DFF uses more images for training, our method still outperforms it and even demonstrates competitive performance with very re" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26409" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 77, + 135, + 144 + ], + "blocks": [ + { + "bbox": [ + 67, + 77, + 135, + 144 + ], + "lines": [ + { + "bbox": [ + 67, + 77, + 135, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 77, + 135, + 144 + ], + "type": "image", + "image_path": "7adbd88e415b1782e08ad32a8f14d479bbf0e68e4557668c63e7be70bf2cdb3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 145, + 134, + 211 + ], + "blocks": [ + { + "bbox": [ + 69, + 145, + 134, + 211 + ], + "lines": [ + { + "bbox": [ + 69, + 145, + 134, + 211 + ], + "spans": [ + { + "bbox": [ + 69, + 145, + 134, + 211 + ], + "type": "image", + "image_path": "05e630337a72659c63d80ba405039ffdadc056ee42a3674e1f6c227a7a9f9c87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 68, + 213, + 134, + 279 + ], + "blocks": [ + { + "bbox": [ + 68, + 213, + 134, + 279 + ], + "lines": [ + { + "bbox": [ + 68, + 213, + 134, + 279 + ], + "spans": [ + { + "bbox": [ + 68, + 213, + 134, + 279 + ], + "type": "image", + "image_path": "8d9d4de1b8eee7075564d3559b1785e0fd6e4d528152a49ca518944d74f94ea9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 68, + 281, + 134, + 346 + ], + "blocks": [ + { + "bbox": [ + 68, + 281, + 134, + 346 + ], + "lines": [ + { + "bbox": [ + 68, + 281, + 134, + 346 + ], + "spans": [ + { + "bbox": [ + 68, + 281, + 134, + 346 + ], + "type": "image", + "image_path": "ae5f6f914002c52e64e75b90b61acc2934abc0d60f8a7690b403fab0f51f859b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 347, + 113, + 354 + ], + "lines": [ + { + "bbox": [ + 95, + 347, + 113, + 354 + ], + "spans": [ + { + "bbox": [ + 95, + 347, + 113, + 354 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 77, + 203, + 144 + ], + "blocks": [ + { + "bbox": [ + 137, + 77, + 203, + 144 + ], + "lines": [ + { + "bbox": [ + 137, + 77, + 203, + 144 + ], + "spans": [ + { + "bbox": [ + 137, + 77, + 203, + 144 + ], + "type": "image", + "image_path": "cf0987a70ae53850481e6c4f6079568bcd111d97a8681093799a04a705c6b051.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 145, + 203, + 211 + ], + "blocks": [ + { + "bbox": [ + 137, + 145, + 203, + 211 + ], + "lines": [ + { + "bbox": [ + 137, + 145, + 203, + 211 + ], + "spans": [ + { + "bbox": [ + 137, + 145, + 203, + 211 + ], + "type": "image", + "image_path": "eaada965a6447fc2a4be6c2ab36021da83df6ca5cd3dbc9d73d9d77c366d4948.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 137, + 213, + 202, + 279 + ], + "blocks": [ + { + "bbox": [ + 137, + 213, + 202, + 279 + ], + "lines": [ + { + "bbox": [ + 137, + 213, + 202, + 279 + ], + "spans": [ + { + "bbox": [ + 137, + 213, + 202, + 279 + ], + "type": "image", + "image_path": "71c786ad40dd8fb57717e468be0b87160b5de383244522178ee88a1a87edcec0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 137, + 281, + 202, + 346 + ], + "blocks": [ + { + "bbox": [ + 137, + 281, + 202, + 346 + ], + "lines": [ + { + "bbox": [ + 137, + 281, + 202, + 346 + ], + "spans": [ + { + "bbox": [ + 137, + 281, + 202, + 346 + ], + "type": "image", + "image_path": "948d722f0c36bac7ae8f756dbe36c22ccfc3cc9b319eee1236c4b343bc1d83f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 347, + 178, + 354 + ], + "lines": [ + { + "bbox": [ + 165, + 347, + 178, + 354 + ], + "spans": [ + { + "bbox": [ + 165, + 347, + 178, + 354 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 77, + 270, + 144 + ], + "blocks": [ + { + "bbox": [ + 205, + 77, + 270, + 144 + ], + "lines": [ + { + "bbox": [ + 205, + 77, + 270, + 144 + ], + "spans": [ + { + "bbox": [ + 205, + 77, + 270, + 144 + ], + "type": "image", + "image_path": "742965fb04db8c4e71e45147e48e6cfcd01df17995266a4f2fe17efae9d9d0f0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 205, + 145, + 270, + 211 + ], + "blocks": [ + { + "bbox": [ + 205, + 145, + 270, + 211 + ], + "lines": [ + { + "bbox": [ + 205, + 145, + 270, + 211 + ], + "spans": [ + { + "bbox": [ + 205, + 145, + 270, + 211 + ], + "type": "image", + "image_path": "d80f53b12b2cc57c0b9203b63a158f794477d99d1049859b7fbeb81a95ef3819.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 205, + 213, + 270, + 279 + ], + "blocks": [ + { + "bbox": [ + 205, + 213, + 270, + 279 + ], + "lines": [ + { + "bbox": [ + 205, + 213, + 270, + 279 + ], + "spans": [ + { + "bbox": [ + 205, + 213, + 270, + 279 + ], + "type": "image", + "image_path": "acd795a1bd846dbabab26ab7f40a7c5a036c03a7cef9d7c00013e034f0e0b4f5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 205, + 281, + 270, + 346 + ], + "blocks": [ + { + "bbox": [ + 205, + 281, + 270, + 346 + ], + "lines": [ + { + "bbox": [ + 205, + 281, + 270, + 346 + ], + "spans": [ + { + "bbox": [ + 205, + 281, + 270, + 346 + ], + "type": "image", + "image_path": "dce9bd8eb625ce66fb3b054a307ddd5ba2fff89f6edb21c3976df8c37fef1db2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 347, + 252, + 356 + ], + "lines": [ + { + "bbox": [ + 231, + 347, + 252, + 356 + ], + "spans": [ + { + "bbox": [ + 231, + 347, + 252, + 356 + ], + "type": "text", + "content": "LSeg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 272, + 77, + 339, + 144 + ], + "blocks": [ + { + "bbox": [ + 272, + 77, + 339, + 144 + ], + "lines": [ + { + "bbox": [ + 272, + 77, + 339, + 144 + ], + "spans": [ + { + "bbox": [ + 272, + 77, + 339, + 144 + ], + "type": "image", + "image_path": "f3000a16cb5513a700e503069b235891466071606b3593991d0bcb6f58bc898d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 272, + 145, + 338, + 211 + ], + "blocks": [ + { + "bbox": [ + 272, + 145, + 338, + 211 + ], + "lines": [ + { + "bbox": [ + 272, + 145, + 338, + 211 + ], + "spans": [ + { + "bbox": [ + 272, + 145, + 338, + 211 + ], + "type": "image", + "image_path": "1f485edf0cceea1ee0e27e8b57302d15cd81e1a2bcf7efc6742eeb8565a95c60.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 272, + 213, + 337, + 279 + ], + "blocks": [ + { + "bbox": [ + 272, + 213, + 337, + 279 + ], + "lines": [ + { + "bbox": [ + 272, + 213, + 337, + 279 + ], + "spans": [ + { + "bbox": [ + 272, + 213, + 337, + 279 + ], + "type": "image", + "image_path": "274a31ba1ceb5e58489c73bffb190e186df071e66432d531f9ca9b953f3823f6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 272, + 281, + 337, + 346 + ], + "blocks": [ + { + "bbox": [ + 272, + 281, + 337, + 346 + ], + "lines": [ + { + "bbox": [ + 272, + 281, + 337, + 346 + ], + "spans": [ + { + "bbox": [ + 272, + 281, + 337, + 346 + ], + "type": "image", + "image_path": "63f52a2b804dd52190abe7cabb11316e5efcc89e069ffe9b9d0f018234706f19.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 347, + 325, + 356 + ], + "lines": [ + { + "bbox": [ + 285, + 347, + 325, + 356 + ], + "spans": [ + { + "bbox": [ + 285, + 347, + 325, + 356 + ], + "type": "text", + "content": "NeRF-DFF" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 340, + 77, + 406, + 144 + ], + "blocks": [ + { + "bbox": [ + 340, + 77, + 406, + 144 + ], + "lines": [ + { + "bbox": [ + 340, + 77, + 406, + 144 + ], + "spans": [ + { + "bbox": [ + 340, + 77, + 406, + 144 + ], + "type": "image", + "image_path": "ce5c16875e7b76a252830491b175905df469c89b4fdd964848c36c53e0960686.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 340, + 145, + 406, + 211 + ], + "blocks": [ + { + "bbox": [ + 340, + 145, + 406, + 211 + ], + "lines": [ + { + "bbox": [ + 340, + 145, + 406, + 211 + ], + "spans": [ + { + "bbox": [ + 340, + 145, + 406, + 211 + ], + "type": "image", + "image_path": "fb1643c48ce1459c8892a458436df37095e1d51d9c732f5c82cef36ba45651d0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 340, + 213, + 406, + 279 + ], + "blocks": [ + { + "bbox": [ + 340, + 213, + 406, + 279 + ], + "lines": [ + { + "bbox": [ + 340, + 213, + 406, + 279 + ], + "spans": [ + { + "bbox": [ + 340, + 213, + 406, + 279 + ], + "type": "image", + "image_path": "baa599d73b8f42a465999ae40dfd6460d1c080b39451c50b054c0975cf77e3be.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 340, + 281, + 406, + 346 + ], + "blocks": [ + { + "bbox": [ + 340, + 281, + 406, + 346 + ], + "lines": [ + { + "bbox": [ + 340, + 281, + 406, + 346 + ], + "spans": [ + { + "bbox": [ + 340, + 281, + 406, + 346 + ], + "type": "image", + "image_path": "51912050a7bbb4bf759cfa00207049a87b4220d627932c74d7c813ae5993ce47.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 347, + 402, + 355 + ], + "lines": [ + { + "bbox": [ + 345, + 347, + 402, + 355 + ], + "spans": [ + { + "bbox": [ + 345, + 347, + 402, + 355 + ], + "type": "text", + "content": "Feature-3DGS" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 408, + 77, + 474, + 144 + ], + "blocks": [ + { + "bbox": [ + 408, + 77, + 474, + 144 + ], + "lines": [ + { + "bbox": [ + 408, + 77, + 474, + 144 + ], + "spans": [ + { + "bbox": [ + 408, + 77, + 474, + 144 + ], + "type": "image", + "image_path": "868f9008bd2025ac29fac86dc02c73e67974bf29ed731141da7fb3afda7e2e2f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 408, + 145, + 474, + 211 + ], + "blocks": [ + { + "bbox": [ + 408, + 145, + 474, + 211 + ], + "lines": [ + { + "bbox": [ + 408, + 145, + 474, + 211 + ], + "spans": [ + { + "bbox": [ + 408, + 145, + 474, + 211 + ], + "type": "image", + "image_path": "7b2e79df87e59878db9648cf460cc0cb3e14764839fc38daca1142d92c0338ba.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 408, + 213, + 474, + 279 + ], + "blocks": [ + { + "bbox": [ + 408, + 213, + 474, + 279 + ], + "lines": [ + { + "bbox": [ + 408, + 213, + 474, + 279 + ], + "spans": [ + { + "bbox": [ + 408, + 213, + 474, + 279 + ], + "type": "image", + "image_path": "7f9e250c5ef6b67abd6d4da20aae36aafdb1423788bf0df45d12cbb395b7044f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 408, + 281, + 474, + 346 + ], + "blocks": [ + { + "bbox": [ + 408, + 281, + 474, + 346 + ], + "lines": [ + { + "bbox": [ + 408, + 281, + 474, + 346 + ], + "spans": [ + { + "bbox": [ + 408, + 281, + 474, + 346 + ], + "type": "image", + "image_path": "c95b2d84afba7b2d5c8e40b028fb0b906e8daf9d397aaf43163714ad46b34c22.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 347, + 454, + 355 + ], + "lines": [ + { + "bbox": [ + 435, + 347, + 454, + 355 + ], + "spans": [ + { + "bbox": [ + 435, + 347, + 454, + 355 + ], + "type": "text", + "content": "LSM" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 476, + 77, + 542, + 144 + ], + "blocks": [ + { + "bbox": [ + 476, + 77, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 476, + 77, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 476, + 77, + 542, + 144 + ], + "type": "image", + "image_path": "1ea02be0c05163d27521690c3739ce0aa84102404d904d5e5598e988d0159a0d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 476, + 145, + 542, + 211 + ], + "blocks": [ + { + "bbox": [ + 476, + 145, + 542, + 211 + ], + "lines": [ + { + "bbox": [ + 476, + 145, + 542, + 211 + ], + "spans": [ + { + "bbox": [ + 476, + 145, + 542, + 211 + ], + "type": "image", + "image_path": "2e6e71891dfe7d144520c276881144b2b5e88fd7eb482ceab9b96dbcb6f6b10b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 476, + 213, + 542, + 279 + ], + "blocks": [ + { + "bbox": [ + 476, + 213, + 542, + 279 + ], + "lines": [ + { + "bbox": [ + 476, + 213, + 542, + 279 + ], + "spans": [ + { + "bbox": [ + 476, + 213, + 542, + 279 + ], + "type": "image", + "image_path": "7faabe2266a7e0afd7c2a90f5fe6294431916025fe68be408d11b7b726c6de34.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 476, + 281, + 542, + 346 + ], + "blocks": [ + { + "bbox": [ + 476, + 281, + 542, + 346 + ], + "lines": [ + { + "bbox": [ + 476, + 281, + 542, + 346 + ], + "spans": [ + { + "bbox": [ + 476, + 281, + 542, + 346 + ], + "type": "image", + "image_path": "b29f3d03b5c4681e1d810b9481cc8b132d72bcac1ed8cfb348ee3dd7b4f3425b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 502, + 347, + 522, + 355 + ], + "lines": [ + { + "bbox": [ + 502, + 347, + 522, + 355 + ], + "spans": [ + { + "bbox": [ + 502, + 347, + 522, + 355 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 76, + 365, + 97, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 365, + 97, + 373 + ], + "spans": [ + { + "bbox": [ + 76, + 365, + 97, + 373 + ], + "type": "text", + "content": "Wall" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 135, + 365, + 157, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 365, + 157, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 365, + 157, + 373 + ], + "type": "text", + "content": "Floor" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 194, + 365, + 223, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 365, + 223, + 376 + ], + "spans": [ + { + "bbox": [ + 194, + 365, + 223, + 376 + ], + "type": "text", + "content": "Ceiling" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 260, + 365, + 283, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 365, + 283, + 374 + ], + "spans": [ + { + "bbox": [ + 260, + 365, + 283, + 374 + ], + "type": "text", + "content": "Chair" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 323, + 365, + 346, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 365, + 346, + 374 + ], + "spans": [ + { + "bbox": [ + 323, + 365, + 346, + 374 + ], + "type": "text", + "content": "Table" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 388, + 365, + 407, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 365, + 407, + 374 + ], + "spans": [ + { + "bbox": [ + 388, + 365, + 407, + 374 + ], + "type": "text", + "content": "Sofa" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 452, + 365, + 469, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 452, + 365, + 469, + 373 + ], + "spans": [ + { + "bbox": [ + 452, + 365, + 469, + 373 + ], + "type": "text", + "content": "Bed" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 509, + 365, + 538, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 509, + 365, + 538, + 374 + ], + "spans": [ + { + "bbox": [ + 509, + 365, + 538, + 374 + ], + "type": "text", + "content": "Others" + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 65, + 437, + 119, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 386, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 386, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 555, + 419 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison in OVS. SpatialSplat achieves sharper and more precise segmentation results compared to previous methods. Notably, our method excels in challenging details, such as distinguishing table legs from cabinet legs (the second the fourth rows), benefiting from our dual-field architecture that captures detailed instance information and uncompressed semantics." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 437, + 119, + 491 + ], + "lines": [ + { + "bbox": [ + 65, + 437, + 119, + 491 + ], + "spans": [ + { + "bbox": [ + 65, + 437, + 119, + 491 + ], + "type": "image", + "image_path": "ca7c174c9fb3d07f184ee41347586bbfa6f39dd3739b5cb98f78dd63e782f838.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 121, + 438, + 176, + 491 + ], + "blocks": [ + { + "bbox": [ + 121, + 438, + 176, + 491 + ], + "lines": [ + { + "bbox": [ + 121, + 438, + 176, + 491 + ], + "spans": [ + { + "bbox": [ + 121, + 438, + 176, + 491 + ], + "type": "image", + "image_path": "4c113e95933606ccd2d4127d5f76d5e69bd0790d38a029607e350be288b2a603.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 178, + 438, + 230, + 491 + ], + "blocks": [ + { + "bbox": [ + 178, + 438, + 230, + 491 + ], + "lines": [ + { + "bbox": [ + 178, + 438, + 230, + 491 + ], + "spans": [ + { + "bbox": [ + 178, + 438, + 230, + 491 + ], + "type": "image", + "image_path": "3c16f72faef771f03962282832f99cd57bca07de508a8d58e73f5c79a9c14ea5.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 233, + 438, + 286, + 491 + ], + "blocks": [ + { + "bbox": [ + 233, + 438, + 286, + 491 + ], + "lines": [ + { + "bbox": [ + 233, + 438, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 233, + 438, + 286, + 491 + ], + "type": "image", + "image_path": "36b23c56dbe38b8e8418290f0857dfa2893980787e123ee5de8ea9c1ca7da502.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 66, + 495, + 118, + 548 + ], + "blocks": [ + { + "bbox": [ + 66, + 495, + 118, + 548 + ], + "lines": [ + { + "bbox": [ + 66, + 495, + 118, + 548 + ], + "spans": [ + { + "bbox": [ + 66, + 495, + 118, + 548 + ], + "type": "image", + "image_path": "7d9c75b686eb848c6d1f20f9f86827ce070e1b65a8e58fd5c2fa70854aa21a5a.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 550, + 143, + 559 + ], + "lines": [ + { + "bbox": [ + 102, + 550, + 143, + 559 + ], + "spans": [ + { + "bbox": [ + 102, + 550, + 143, + 559 + ], + "type": "text", + "content": "Input views" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 121, + 495, + 176, + 548 + ], + "blocks": [ + { + "bbox": [ + 121, + 495, + 176, + 548 + ], + "lines": [ + { + "bbox": [ + 121, + 495, + 176, + 548 + ], + "spans": [ + { + "bbox": [ + 121, + 495, + 176, + 548 + ], + "type": "image", + "image_path": "b5b380a8713197993c942e565f6164fb9dbfaf9aa29a9de7efbe5f0ff34b47ae.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 179, + 495, + 230, + 548 + ], + "blocks": [ + { + "bbox": [ + 179, + 495, + 230, + 548 + ], + "lines": [ + { + "bbox": [ + 179, + 495, + 230, + 548 + ], + "spans": [ + { + "bbox": [ + 179, + 495, + 230, + 548 + ], + "type": "image", + "image_path": "af49e80d9c25a03f76b1186f7fc03f6a30b46ffe7a3d842092e4957ef3899a6b.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 550, + 271, + 559 + ], + "lines": [ + { + "bbox": [ + 190, + 550, + 271, + 559 + ], + "spans": [ + { + "bbox": [ + 190, + 550, + 271, + 559 + ], + "type": "text", + "content": "Rendered Feature Field" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 232, + 495, + 286, + 548 + ], + "blocks": [ + { + "bbox": [ + 232, + 495, + 286, + 548 + ], + "lines": [ + { + "bbox": [ + 232, + 495, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 232, + 495, + 286, + 548 + ], + "type": "image", + "image_path": "890a7fff8cd638e4343a3b08e0e7f567dfe2c929ce548d82bdd6ec3235964f98.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 565, + 295, + 586 + ], + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 586 + ], + "type": "text", + "content": "Figure 5. The rendered instance features. SpatialSplat predicts clear and consistent instance features across different views." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "bbox": [ + 55, + 602, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 638 + ], + "type": "text", + "content": "cent SOTA methods designed specifically for novel view synthesis. This highlights that our method efficiently learns spatial priors from large-scale 2D data." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Results of Open-vocabulary 3D Segmentation. Following LSM's approach, we map category labels from the Scannet dataset into common categories: Wall, Floor, Ceiling, Chair, Table, Bed, Sofa, Others. As shown in Tab. 1, SpatialSplat outperforms all compared methods, even surpassing L-Seg, which provides semantic feature supervision" + } + ] + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 324, + 437, + 367, + 524 + ], + "blocks": [ + { + "bbox": [ + 324, + 437, + 367, + 524 + ], + "lines": [ + { + "bbox": [ + 324, + 437, + 367, + 524 + ], + "spans": [ + { + "bbox": [ + 324, + 437, + 367, + 524 + ], + "type": "image", + "image_path": "ecfc7f0f9a6cfff07e5803e25f02598567cec7a8cb7f954f217a8ffb4bd89b9f.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 525, + 355, + 533 + ], + "lines": [ + { + "bbox": [ + 344, + 525, + 355, + 533 + ], + "spans": [ + { + "bbox": [ + 344, + 525, + 355, + 533 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 369, + 438, + 410, + 480 + ], + "blocks": [ + { + "bbox": [ + 369, + 438, + 410, + 480 + ], + "lines": [ + { + "bbox": [ + 369, + 438, + 410, + 480 + ], + "spans": [ + { + "bbox": [ + 369, + 438, + 410, + 480 + ], + "type": "image", + "image_path": "f1d69e9a6c208387cf34ac932f4e7fa58a70fcf7d1d91d5a8403b122ed7fbbe6.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 369, + 482, + 410, + 524 + ], + "blocks": [ + { + "bbox": [ + 369, + 482, + 410, + 524 + ], + "lines": [ + { + "bbox": [ + 369, + 482, + 410, + 524 + ], + "spans": [ + { + "bbox": [ + 369, + 482, + 410, + 524 + ], + "type": "image", + "image_path": "90a840d68ed14cab321e553e3778c2fd63aaaf9c077ecc1f233cdac6d598b033.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 525, + 402, + 533 + ], + "lines": [ + { + "bbox": [ + 386, + 525, + 402, + 533 + ], + "spans": [ + { + "bbox": [ + 386, + 525, + 402, + 533 + ], + "type": "text", + "content": "LSM" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 412, + 438, + 455, + 481 + ], + "blocks": [ + { + "bbox": [ + 412, + 438, + 455, + 481 + ], + "lines": [ + { + "bbox": [ + 412, + 438, + 455, + 481 + ], + "spans": [ + { + "bbox": [ + 412, + 438, + 455, + 481 + ], + "type": "image", + "image_path": "6ba58f3ba2c4146e33b6eed91c252afbea4f2252342e35c0455dc3f894ac5970.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 412, + 483, + 455, + 524 + ], + "blocks": [ + { + "bbox": [ + 412, + 483, + 455, + 524 + ], + "lines": [ + { + "bbox": [ + 412, + 483, + 455, + 524 + ], + "spans": [ + { + "bbox": [ + 412, + 483, + 455, + 524 + ], + "type": "image", + "image_path": "b4d8cc431ce151221ac3b1a628fe788e828f2068225a9186128bf10edb55c251.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 525, + 444, + 533 + ], + "lines": [ + { + "bbox": [ + 427, + 525, + 444, + 533 + ], + "spans": [ + { + "bbox": [ + 427, + 525, + 444, + 533 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 535, + 553, + 556 + ], + "lines": [ + { + "bbox": [ + 313, + 535, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 553, + 556 + ], + "type": "text", + "content": "Figure 6. Qualitative results of cross-dataset generalization. Zoom out for clearer visualization." + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 457, + 438, + 500, + 479 + ], + "blocks": [ + { + "bbox": [ + 457, + 438, + 500, + 479 + ], + "lines": [ + { + "bbox": [ + 457, + 438, + 500, + 479 + ], + "spans": [ + { + "bbox": [ + 457, + 438, + 500, + 479 + ], + "type": "image", + "image_path": "b1e3b597510dec67cf74c1e6234c02f36267ef681a17af71fd4bfa63563ddddc.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 457, + 483, + 500, + 524 + ], + "blocks": [ + { + "bbox": [ + 457, + 483, + 500, + 524 + ], + "lines": [ + { + "bbox": [ + 457, + 483, + 500, + 524 + ], + "spans": [ + { + "bbox": [ + 457, + 483, + 500, + 524 + ], + "type": "image", + "image_path": "4db8a01d7bd9e668291b666fe07a180868b7690153095e7398b2b40dbd053a4a.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 525, + 489, + 533 + ], + "lines": [ + { + "bbox": [ + 474, + 525, + 489, + 533 + ], + "spans": [ + { + "bbox": [ + 474, + 525, + 489, + 533 + ], + "type": "text", + "content": "LSM" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 501, + 438, + 544, + 479 + ], + "blocks": [ + { + "bbox": [ + 501, + 438, + 544, + 479 + ], + "lines": [ + { + "bbox": [ + 501, + 438, + 544, + 479 + ], + "spans": [ + { + "bbox": [ + 501, + 438, + 544, + 479 + ], + "type": "image", + "image_path": "0235ce2885724e6deaa21ddcc12c2db85541c222f452b0ac90a3a66afd6d6ffa.jpg" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_body" + } + ], + "index": 68 + }, + { + "type": "image", + "bbox": [ + 502, + 482, + 544, + 523 + ], + "blocks": [ + { + "bbox": [ + 502, + 482, + 544, + 523 + ], + "lines": [ + { + "bbox": [ + 502, + 482, + 544, + 523 + ], + "spans": [ + { + "bbox": [ + 502, + 482, + 544, + 523 + ], + "type": "image", + "image_path": "57d4e09f8473e4a30f7c46cccbeb00cf64d159431fde016c8f913692c7a664c2.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 517, + 525, + 535, + 533 + ], + "lines": [ + { + "bbox": [ + 517, + 525, + 535, + 533 + ], + "spans": [ + { + "bbox": [ + 517, + 525, + 535, + 533 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "bbox": [ + 313, + 579, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 687 + ], + "type": "text", + "content": "for other compared methods. As illustrated in Fig. 4, SpatialSplat produces sharp and clear 3D semantic segmentation results. This is achieved by lifting both 2D instance and uncompressed semantic into 3D space, enabling more efficient semantic learning while mitigating the blurred contours and information loss associated with per-primitive compressed semantics. The visualizations in Fig. 5 show that SpatialSplat effectively learns a highly consistent instance prior with only weak supervision from 2D masks." + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Results of Cross-Dataset Generalization. As shown in Tab. 2, our method significantly outperforms LSM on the" + } + ] + } + ], + "index": 73 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26410" + } + ] + } + ], + "index": 74 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 81, + 119, + 134 + ], + "blocks": [ + { + "bbox": [ + 66, + 81, + 119, + 134 + ], + "lines": [ + { + "bbox": [ + 66, + 81, + 119, + 134 + ], + "spans": [ + { + "bbox": [ + 66, + 81, + 119, + 134 + ], + "type": "image", + "image_path": "110b60505e6c0e8e2a5d73b2d2d68a13dc0c645fb04d1e476700ccd03eb96b8d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 135, + 97, + 144 + ], + "lines": [ + { + "bbox": [ + 86, + 135, + 97, + 144 + ], + "spans": [ + { + "bbox": [ + 86, + 135, + 97, + 144 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 81, + 175, + 134 + ], + "blocks": [ + { + "bbox": [ + 121, + 81, + 175, + 134 + ], + "lines": [ + { + "bbox": [ + 121, + 81, + 175, + 134 + ], + "spans": [ + { + "bbox": [ + 121, + 81, + 175, + 134 + ], + "type": "image", + "image_path": "c17300b3b9cedf4b8fe5829d22b52477b3c2a1de4159f1a88d42633e140fce23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 135, + 153, + 144 + ], + "lines": [ + { + "bbox": [ + 141, + 135, + 153, + 144 + ], + "spans": [ + { + "bbox": [ + 141, + 135, + 153, + 144 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 178, + 81, + 230, + 134 + ], + "blocks": [ + { + "bbox": [ + 178, + 81, + 230, + 134 + ], + "lines": [ + { + "bbox": [ + 178, + 81, + 230, + 134 + ], + "spans": [ + { + "bbox": [ + 178, + 81, + 230, + 134 + ], + "type": "image", + "image_path": "c902820742d14d40e0584a5dc351de0673b65c7c299421fcf585c3df1868331f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 135, + 208, + 144 + ], + "lines": [ + { + "bbox": [ + 198, + 135, + 208, + 144 + ], + "spans": [ + { + "bbox": [ + 198, + 135, + 208, + 144 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 202 + ], + "type": "text", + "content": "Figure 7. Qualitative results of ablations. (a) and (b) Qualitative results of importance score prediction, with red color indicating an importance score of 1 and blue indicating 0. (c) The rendered semantic feature without the dual-field. (d) The rendered semantic feature with the dual-field, which appears clearer and sharper." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 233, + 81, + 286, + 134 + ], + "blocks": [ + { + "bbox": [ + 233, + 81, + 286, + 134 + ], + "lines": [ + { + "bbox": [ + 233, + 81, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 233, + 81, + 286, + 134 + ], + "type": "image", + "image_path": "40a6872b75f51c93a8078623b49001509f0c67302839ef6653f4809d65606e28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 135, + 264, + 144 + ], + "lines": [ + { + "bbox": [ + 254, + 135, + 264, + 144 + ], + "spans": [ + { + "bbox": [ + 254, + 135, + 264, + 144 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 212, + 293, + 289 + ], + "blocks": [ + { + "bbox": [ + 58, + 212, + 293, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 212, + 293, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 212, + 293, + 289 + ], + "type": "table", + "html": "
MethodLatency↓Gaussian Size ↓Num. ↓
Feature-3DGS [52]1069 s229.65 MB476K
NoPoSplat [46]0.044 s30.93 MB131K
LSM [15]0.104 s63.00 MB131K
SpatialSplat-Lite0.058 s25.58 MB87.7K
SpatialSplat0.071 s25.58 MB87.7K
", + "image_path": "45d1347a0067c604f6fae59c36eecb2e43f00ca92c5db8ad09743c833a6c6864.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 324, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 324, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 295, + 420 + ], + "type": "text", + "content": "unseen Replica dataset in both novel view synthesis and open-vocabulary segmentation. Notably, since Replica is a synthetic dataset with a different data modality from our training set, this underscores the strong generalization ability of our approach. The visualization in Fig. 6 further demonstrates that LSM struggles with coarse modeling at scene edges, leading to artifacts in synthesized views, whereas our method generates photorealistic renderings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "text", + "content": "Analysis of Model Efficiency. We compare the model efficiency of different 3DGS-based methods. As shown in Tab. 3, our method is faster than LSM thanks to its streamlined architecture, requiring only " + }, + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "text", + "content": " of the storage size and " + }, + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 55, + 422, + 295, + 579 + ], + "type": "text", + "content": " of primitive number. Remarkably, our approach requires even fewer Gaussians parameters than NoPoSplat, which lacks semantic awareness. This demonstrates the effectiveness of our proposed dual-field architecture and selective Gaussian mechanism, which capture accurate semantics while eliminating redundant Gaussians. Furthermore, as our method does not rely on dense semantic supervision, we leverage a lightweight pretrained 2D model, significantly accelerating inference speed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 586, + 188, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 188, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 188, + 599 + ], + "type": "text", + "content": "4.3. Ablations and Analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 603, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 676 + ], + "type": "text", + "content": "We perform ablations to answer the following questions: (1) Are the primitives removed by our selective Gaussian mechanism truly redundant? (2) Does the improved scene understanding stem from the proposed dual-field architecture? (3) As the number of views increases, how does redundancy grow? Can our method effectively alleviate it?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Effect of selective Gaussian mechanism. To answer Question 1, we remove the SGM and retrain our model using the same configuration. As shown in Tab. 4, the SGM" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 552, + 145 + ], + "blocks": [ + { + "bbox": [ + 55, + 292, + 294, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 294, + 314 + ], + "type": "text", + "content": "Table 3. Model efficiency comparison. \"Lite\": the model replaces LSeg with CLIP ViT-B/16." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 71, + 552, + 145 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 145 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 145 + ], + "type": "table", + "html": "
ModelmIoU↑Acc.↑PSNR↑Num. ↓
No SGM0.55690.797225.55131k
No dual0.50270.778625.4687.6k
3-view0.51250.779324.9297.5k
SpatialSplat0.55770.792025.4787.7k
", + "image_path": "64655833e635cc077102c02b0be24c5b524147e6f5a8eaa97f1f4f1198ecf635.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 146, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 146, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 553, + 191 + ], + "type": "text", + "content": "Table 4. Ablation study on design choices. The proposed Adaptive Gaussian mechanism and dual-field architecture jointly enable the construction of high-quality representations with significantly reduced dimensionality." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": "eliminates approximately " + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": " of Gaussian primitives while causing only a 0.08 drop in PSNR, indicating that the removed primitives have minimal impact on rendering quality. The visualization in Fig. 7 (a) and (b) further demonstrates that the SGM accurately localizes overlapping areas in the input images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 279, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 399 + ], + "type": "text", + "content": "Effect of dual-field architecture. To address Question 2, we replace our dual-field architecture with single-level per-primitive semantic learning. Following the approach in LSM, we set the semantic feature dimension to 64 and use a convolutional layer to expand it to 512. As shown in Tab. 4, this results in a significant drop of " + }, + { + "bbox": [ + 313, + 279, + 554, + 399 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 279, + 554, + 399 + ], + "type": "text", + "content": " in mIoU for open-vocabulary segmentation performance. The primary issue is that per-primitive semantic learning struggles to maintain accurate semantics and fails to preserve clear instance boundaries, as illustrated in Fig. 7 (c) and (d)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 402, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 521 + ], + "type": "text", + "content": "Extend to more views. Our experiments primarily focus on two-view settings to ensure a fair comparison with our baseline. However, our method can be extended to more views for reconstructing larger scenes. As shown in Tab. 4, our approach remains effective even with three input views. Moreover, compared to pixel-wise methods, which require approximately 197K primitives to represent the scene, our method achieves a more compact representation by reducing about " + }, + { + "bbox": [ + 313, + 402, + 555, + 521 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 402, + 555, + 521 + ], + "type": "text", + "content": " primitives. More qualitative details can be found in the appendix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "We present an efficient feed-forward network that generates compact yet high-performance representations for 3D reconstruction and scene understanding from sparse, unposed images. Our proposed dual-field architecture achieves state-of-the-art performance while significantly reducing the number of representation parameters. Additionally, we introduce a Selective Gaussian Mechanism that directly identifies redundant Gaussians caused by pixel-wise predictions without requiring geometric priors. SpatialSplat achieves impressive results in 3D reconstruction and scene understanding without relying on any 3D data during training or inference, making semantic 3D reconstruction from sparse, unposed images more practical." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26411" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "content": "This work was partially supported by Hunan Province Major Scientific and Technological Project No. 2024QK2001, National Key R&D Program of China No. 2023YFB4704500." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 158, + 115, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 115, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 115, + 170 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 177, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 177, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 294, + 253 + ], + "type": "text", + "content": "[1] Yash Bhalgat, Iro Laina, João F. Henriques, Andrea Vedaldi, and Andrew Zisserman. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 294, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 294, + 321 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9630-9640. IEEE, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 388 + ], + "type": "text", + "content": "[3] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 19457-19467. IEEE, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 389, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 294, + 444 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXII, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 467 + ], + "type": "text", + "content": "[5] Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. CoRR, abs/2401.03890, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 534 + ], + "type": "text", + "content": "[6] Yihang Chen, Qianyi Wu, Weiyao Lin, Mehrtash Harandi, and Jianfei Cai. HAC: hash-grid assisted context for 3d gaussian splatting compression. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part VII, pages 422-438. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 611 + ], + "type": "text", + "content": "[7] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XXI, pages 370-386. Springer, 2024. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 667 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[9] Jianmei Dai, Zhilong Zhang, Shiwen Mao, and Danpu Liu. A view synthesis-based " + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": " VR caching system over mec-enabled C-RAN. IEEE Trans. Circuits Syst. Video Technol., 30(10):3843-3855, 2020. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 160 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 162, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 225 + ], + "type": "text", + "content": "[11] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4970-4980. IEEE, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "text", + "content": "[12] Zhiwen Fan, Peihao Wang, Yifan Jiang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Nerf-sos: Any-view self-supervised object segmentation on complex scenes. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "text", + "content": "[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. CoRR, abs/2403.20309, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "text", + "content": "[14] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and " + }, + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 317, + 361, + 553, + 436 + ], + "type": "text", + "content": " FPS. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 437, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 525 + ], + "type": "text", + "content": "[15] Zhiwen Fan, Jian Zhang, Wenyan Cong, Peihao Wang, Renjie Li, Kairun Wen, Shijie Zhou, Achuta Kadambi, Zhangyang Wang, Danfei Xu, Boris Ivanovic, and Marco Pavone. Large spatial model: End-to-end unposed images to semantic 3d. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 1, 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 526, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 591 + ], + "type": "text", + "content": "[16] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 12479-12488. IEEE, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 658 + ], + "type": "text", + "content": "[17] Qiao Gu, Zhaoyang Lv, Duncan P. Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d segmentation for egocentric perception. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLIII, pages 382-400. Springer, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[18] Mohammad Mahdi Johari, Yann Lepoittevin, and François Fleuret. Geonerf: Generalizing nerf with geometry priors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 18344-18347. IEEE, 2022. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26412" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[20] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. LERF: language embedded radiance fields. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 19672-19682. IEEE, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 250 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 3992-4003. IEEE, 2023. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 317 + ], + "type": "text", + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 385 + ], + "type": "text", + "content": "[23] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 442 + ], + "type": "text", + "content": "[24] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21719-21728. IEEE, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 444, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 485 + ], + "type": "text", + "content": "[25] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 543 + ], + "type": "text", + "content": "[26] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "type": "text", + "content": "[27] Luca De Luigi, Damiano Bolognini, Federico Domeniconi, Daniele De Gregorio, Matteo Poggi, and Luigi Di Stefano. Scannerf: a scalable benchmark for neural radiance fields. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 816-825. IEEE, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "text", + "content": "[28] Andrew L Maas, Awni Y Hannun, Andrew Y Ng, et al. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, page 3. Atlanta, GA, 2013. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: representing scenes as neural radiance fields for view synthesis. Commun. ACM, 65(1):99-106, 2022. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[30] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multires" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "olution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "type": "text", + "content": "[31] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20051-20060. IEEE, 2024. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 207 + ], + "type": "text", + "content": "[32] Ri-Zhao Qiu, Yafei Hu, Ge Yang, Yuchen Song, Yang Fu, Jianglong Ye, Jiteng Mu, Ruihan Yang, Nikolay Atanasov, Sebastian A. Scherer, and Xiaolong Wang. Learning generalizable feature fields for mobile manipulation. CoRR, abs/2403.07563, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 274 + ], + "type": "text", + "content": "[33] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part XLI, pages 368-383. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 277, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 364 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "text", + "content": "[35] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 12159-12168. IEEE, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "type": "text", + "content": "[36] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. In Conference on Robot Learning, CoRL 2023, 6-9 November 2023, Atlanta, GA, USA, pages 405-424. PMLR, 2023. 1, 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 480, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 480, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 480, + 553, + 555 + ], + "type": "text", + "content": "[37] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 9043-9052. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 601 + ], + "type": "text", + "content": "[38] Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs. arXiv preprint arXiv:2408.13912, 2024. 2, 4, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 604, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 713 + ], + "type": "text", + "content": "[39] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 5" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26413" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[40] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In International Conference on 3D Vision, 3DV 2022, Prague, Czech Republic, September 12-16, 2022, pages 443-453. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "text", + "content": "[41] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4690-4699. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[42] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 306 + ], + "type": "text", + "content": "[43] Xingrui Wang, Cuiling Lan, Hanxin Zhu, Zhibo Chen, and Yan Lu. Gsemsplat: Generalizable semantic 3d gaussian splatting from uncalibrated image pairs. CoRR, abs/2412.16932, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 385 + ], + "type": "text", + "content": "[44] Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024. 2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[45] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 20041-20050. IEEE, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 455, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 508 + ], + "type": "text", + "content": "[46] Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. arXiv preprint arXiv:2410.24207, 2024. 2, 4, 5, 6, 8, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "text", + "content": "[47] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[48] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 4578-4587. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[49] Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. CoRR, abs/2408.13770, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[50] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 586-" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 283 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "595. Computer Vision Foundation / IEEE Computer Society, 2018. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[51] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 15818-15827. IEEE, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 239 + ], + "type": "text", + "content": "[52] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21676-21685. IEEE, 2024. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "text", + "content": "[53] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26414" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_content_list.json b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ec23b081013b9102478b7e841c0db5f8976ae893 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion", + "text_level": 1, + "bbox": [ + 99, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxi Xiao $^{1*}$ Jianyuan Wang $^{2}$ Nan Xue $^{3}$ Nikita Karaev $^{2,4}$ Yuri Makarov $^{4}$ Bingyi Kang $^{5}$ Xing Zhu $^{3}$ Hujun Bao $^{1}$ Yujun Shen $^{3}$ Xiaowei Zhou $^{1\\dagger}$", + "bbox": [ + 109, + 178, + 885, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University $^{2}$ Oxford $^{3}$ Ant Group $^{4}$ Pixelwise AI $^{5}$ Bytedance Seed", + "bbox": [ + 127, + 223, + 872, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a96c785c085d36ec1b1d6856714fabcf7afe3e82dad86dabac1fdf80f452a789.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 253, + 496, + 478 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd13107edc76019a3015327cc619496bc75ec554b56371a447106172748ce6f4.jpg", + "image_caption": [ + "Figure 1. SpatialTrackerV2 produces consistent 3D scene geometry, camera poses, and 3D point trajectories all at once from monocular videos of arbitrary scenarios, e.g., robotic manipulation, first-person egocentric views, and dynamic sports (drifting and skating) shown in this figure. Try our online demo at https://huggingface.co/spaces/Yuxihenry/SpatialTrackerV2." + ], + "image_footnote": [], + "bbox": [ + 101, + 477, + 493, + 686 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1613cdebd69c7240e2eb167eac480000ed1ad30259aaf5a6003d7a18d3d99f08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 251, + 887, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b08c78df9be708b49dceb74a24a922cc3794f52032ad900d5839ca838149d598.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 481, + 887, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 741, + 328, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SpatialTrackerV2, a feed-forward 3D point tracking method for monocular videos. Going beyond modular pipelines built on off-the-shelf components for 3D tracking, our approach unifies the intrinsic connections between point tracking, monocular depth, and camera pose estimation into a high-performing and feedforward 3D point tracker. It decomposes world-space 3D motion", + "bbox": [ + 89, + 772, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into scene geometry, camera ego-motion, and pixel-wise object motion, with a fully differentiable and end-to-end architecture, allowing scalable training across a wide range of datasets, including synthetic sequences, posed RGB-D videos, and unlabeled in-the-wild footage. By learning geometry and motion jointly from such heterogeneous data, SpatialTrackerV2 outperforms existing 3D tracking methods by $30\\%$ , and matches the accuracy of leading dynamic 3D reconstruction approaches while running $50\\times$ faster.", + "bbox": [ + 511, + 746, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Partially completed during Ant internship. † Corresponding author.", + "bbox": [ + 88, + 886, + 455, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6726", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D point tracking aims to recover long-term 3D trajectories of arbitrary points from monocular videos. As a universal dynamic scene representation, it has recently shown strong potential in diverse applications, including robotics [48, 56, 67], video generation [22, 29, 57, 81], and 3D/4D reconstruction [10, 34, 38, 76]. Compared to parametric motion models (e.g., SMPL [47], MANO [63], skeletons, or 3D bounding boxes), it offers greater flexibility and generalization over various real-world scenes, as shown in Fig. 1.", + "bbox": [ + 89, + 114, + 480, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing solutions [1, 4, 54, 75, 86] of 3D point tracking extensively explored the well-developed low/mid-level vision models, such as optical flow [45, 68] and monocular depth estimation [27, 87], and took benefits from 2D point tracking models [15, 23, 33]. Among them, optimization-based methods [42, 75, 95] distill the optical flow, monocular depth models and camera motions for each given monocular video with promising results obtained, while being computationally expensive due to their per-scene optimization designs. SpatialTracker [86] moved forward to efficient 3D point tracking with a feed-forward model, and the more recent works [1, 4, 54] explored different architecture designs and rendering constraints to achieve higher-quality 3D tracking. Nevertheless, the feed-forward solutions are limited to training data scalability issues due to the need for ground-truth 3D tracks as supervision, which downgrades the tracking quality in real-world casual captures. Moreover, overlooking the inherent interplay between camera motion, object motion, and scene geometry results in error entanglement and accumulation across modules.", + "bbox": [ + 91, + 266, + 482, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These limitations motivate our core insights: (1) The reliance on ground-truth 3D trajectories constrains the scalability of existing feed-forward models, highlighting the need for designs that can generalize across diverse and weakly-supervised data sources. (2) The absence of joint reasoning over scene geometry, camera motion, and object motion leads to compounded errors and degraded performance, underscoring the importance of disentangling and explicitly modeling these motion components. To address these challenges, we decompose 3D point tracking into three distinct components: video depth, ego (camera) motion, and object motion, and integrate them within a fully differentiable pipeline that supports scalable joint training across heterogeneous data.", + "bbox": [ + 89, + 568, + 482, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our SpatialTrackerV2, a front-end and back-end architecture is proposed. The front end is a video depth estimator and camera pose initializer, adapted from typical monocular depth prediction frameworks [88] with attention-based temporal information encoding [74]. The predicted video depths and camera poses are then fused through a scale-shift estimation module, which ensures consistency between the depth and motion predictions. The back end", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "consists of a proposed Joint Motion Optimization Module, which takes the video depth and coarse camera trajectories as input and iteratively estimates 2D and 3D trajectories, along with trajectory-wise dynamics and visibility scores. This enables an efficient bundle adjustment process for optimizing camera poses in the loop. At its core lies a novel SyncFormer, which separately models the 2D and 3D correlations in two branches, connected by multiple cross-attention layers. This design mitigates mutual interference between 2D and 3D embeddings and allows the model to update representations in two distinct spaces, namely the image (UV) space and the camera coordinate space. Furthermore, benefiting from this dual-branch design, bundle adjustment can be effectively applied to jointly optimize camera poses as well as the 2D and 3D trajectories.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This unified and differentiable pipeline makes large-scale training on diverse datasets possible. For RGB-D datasets with camera poses, we jointly train 3D tracking using consistency constraints from ground-truth depth and camera poses for static points, while dynamic points seamlessly contribute to the optimization. For video datasets that provide only camera pose annotations and lack depth information, we leverage consistency among camera poses, and 2D and 3D point tracking to drive the model's optimization. Relying on this framework, we successfully scale up training of the entire pipeline across 17 datasets.", + "bbox": [ + 511, + 316, + 903, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluations on the TAPVid-3D benchmark [37] show that our method sets a new state-of-the-art in 3D point tracking, achieving 21.2 AJ and $31.0\\mathrm{APD}_{3D}$ , surpassing DELTA [54] with relative improvements of $61.8\\%$ and $50.5\\%$ , respectively. Additionally, extensive experiments on dynamic reconstruction show our superior results on consistent video depth and camera poses estimation. Specifically, SpatialTrackerV2 beats the best dynamic reconstruction method, MegaSAM [42], on most video depth datasets and achieves comparable results on various camera pose benchmarks, while its inference speed is $50\\times$ faster.", + "bbox": [ + 511, + 483, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 662, + 651, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section covers relevant literature on 3D tracking, depth, and camera pose estimation.", + "bbox": [ + 511, + 688, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Point tracking", + "text_level": 1, + "bbox": [ + 511, + 727, + 661, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PIPs [23] revisited the 2D point tracking task first introduced in [64] and proposed a deep learning approach to solve it. TAP-Vid [14] redefined the problem and introduced both a benchmark and a simple architecture, TAP-Net. Subsequently, the performance was improved in TAPIR [15] by combining the global matching capabilities of TAP-Net with the local refinement offered by PIPs. CoTracker [33] pioneered tracking through occlusions using a transformer architecture combined with joint attention, followed by TAPTR [40] and LocoTrack [11], which im", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6727", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proved efficiency and introduced 4D correlation volumes. Recently, BootsTAPIR [16] and CoTracker3 [32] explored the use of unlabeled data to achieve better performance.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While 2D point tracking has been extensively studied, 3D point tracking remains a relatively new field. The first method to demonstrate 3D point tracking capabilities was the test-time optimization-based OmniMotion [75]. Later, SpatialTracker [86] introduced the first feed-forward 3D point tracker by combining a 2D point tracker [33] with depth priors from a monocular depth estimator [3]. SceneTracker [71] proposed a new architecture for 3D tracking with depth priors, while DELTA [54] improved efficiency and achieved dense 3D tracking. Recently, TAPIP3D [90] improved 3D tracking robustness by lifting image features into the world coordinate space and performing tracking there. All of these 3D tracking models were trained on small synthetic datasets.", + "bbox": [ + 88, + 137, + 482, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike these methods, we present a scalable 3D tracking framework trained on a collection of both real and synthetic datasets, while also explicitly modeling camera motion to improve performance on egocentric videos.", + "bbox": [ + 89, + 349, + 483, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Depth estimation", + "text_level": 1, + "bbox": [ + 89, + 420, + 261, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early methods such as Eigen et al. [17] introduced single-view depth estimation using CNNs, but were limited by dataset scale and poor generalization [18, 89]. MiDaS [5] improved this by mixing datasets for broader coverage, and ZoeDepth [3] adapted it for metric depth, though ambiguity from missing camera intrinsics remained. Later, Metric3D [26] and UniDepth [60] addressed this by jointly estimating intrinsics and normalized depth.", + "bbox": [ + 89, + 441, + 482, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With large-scale pretraining (e.g., diffusion [13, 41], DINO [8]), recent models like Marigold [70] and DepthAnything [87, 88] have significantly advanced zero-shot depth. Extensions to video [9, 27] have also emerged.", + "bbox": [ + 89, + 564, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we build on DepthAnythingV2 and extend it to video, aiming for not just consistent depth, but a unified framework that also predicts camera poses and tracks, with proper scale alignment—posing new challenges beyond static depth estimation.", + "bbox": [ + 89, + 625, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Camera estimation", + "text_level": 1, + "bbox": [ + 89, + 710, + 272, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional camera pose estimation methods [24, 58] rely on image-to-image point correspondences using keypoint detectors (e.g., SIFT [49, 50], SURF [2]) and matching techniques such as nearest neighbors, followed by geometric algorithms like the five-point and eight-point methods [24, 25, 39, 55, 85]. Bundle Adjustment [69] is also commonly employed to further enhance accuracy. Recently, direct regression approaches using neural networks [35, 43, 51, 84, 92] have emerged, aiming to overcome limitations in sparse-view scenarios or when correspondences are unreliable. Diffusion models, such as PoseDiffusion [72] and RayDif", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fusion [93], have also been explored, offering strong accuracy but suffering from high inference costs. In contrast, the camera head of VGGsfm [73] or VGGT [74] adopts an iterative refinement paradigm similar to RAFT [68], striking a balance between accuracy and inference cost, and enabling estimation of both extrinsic and intrinsic parameters.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 194, + 604, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video with $T$ frames $(\\mathcal{I}^t)_{t=1}^T$ and $N$ query points $\\mathcal{Q}_i = (x_i, y_i) \\in \\mathbb{R}^2$ , $i = 1, \\dots, N$ , the goal of 3D tracking is to recover pixel-wise 3D trajectories $\\mathcal{T} = (\\mathcal{T}_i^t)_{i=1,\\dots,N}$ for each query. In order to account for static and dynamic parts of the scene, we decompose $\\mathcal{T}$ into two components the ego camera motion $\\mathcal{T}_{\\mathrm{ego}}$ and pixelwise object motion $\\mathcal{T}_{\\mathrm{object}}$ . Fig. 2 illustrates the pipeline of our method.", + "bbox": [ + 511, + 219, + 906, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Ego Motion Component", + "text_level": 1, + "bbox": [ + 511, + 335, + 735, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ego motion, represented by camera trajectories $\\mathcal{T}_{\\mathrm{ego}}$ , is a major contributor to the 3D flow in camera coordinates. To compute the 3D tracks induced by ego motion, we need to estimate scale-aligned camera trajectories and video depth. Video Depth. Monocular depth estimation models, such as DepthAnything [87], typically follow an encoder-decoder design with a vision encoder like DINO [8] and a DPT-style decoder. Following VGGT [74], we extend the monocular encoder into a temporal encoder using an alternating-attention mechanism. This mechanism alternates between intra-frame self-attention and inter-frame attention over flattened video tokens, effectively balancing performance and efficiency. Furthermore, two learnable tokens, $\\mathbf{P}_{\\mathrm{tok}}$ and $\\mathbf{S}_{\\mathrm{tok}}$ , are incorporated into the alternating-attention layers to capture high-level semantics for pose and scale regression. Camera Tracker. Similar to [73, 74], we adopt a differentiable pose head, $\\mathcal{H}$ , to decode pose, scale and shift directly:", + "bbox": [ + 511, + 357, + 906, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} ^ {t}, a, b = \\mathcal {H} \\left(\\mathbf {P} _ {\\mathrm {t o k}}, \\mathbf {S} _ {\\mathrm {t o k}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 628, + 906, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^t\\in \\mathbb{R}^{T\\times 8}$ is the camera encoding parameterized by a quaternion, a translation vector, and the normalized focal length concatenated together; The parameters $a$ and $b$ represent the scale and shift used to align the depth with the estimated camera poses. After that, we can easily obtain the 3D trajectories induced by ego-motion $\\mathcal{T}_{\\mathrm{ego}}$ :", + "bbox": [ + 511, + 650, + 905, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\text {e g o}} = \\mathcal {W} (\\mathcal {P}, a * \\mathcal {D} _ {\\text {n o r m}} + b), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 753, + 906, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\mathrm{norm}}$ is the raw results after DPT head with activation functions and $\\mathcal{W}$ is the camera transformation.", + "bbox": [ + 511, + 779, + 905, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Motion Optimization.", + "text_level": 1, + "bbox": [ + 511, + 816, + 761, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After ego-motion initialization, we jointly estimate 2D trajectories $\\mathcal{T}^{2\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 2}$ in UV space and their corresponding 3D trajectories $\\mathcal{T}^{3\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 3}$ in the camera coordinate system using an iterative transformer module, referred", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6728", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1bf9ec33a00fa908e2c02013dff54ac6ca106be1f5690bfc43cafb2fda30050.jpg", + "image_caption": [ + "Figure 2. Pipeline Overview. Our method adopts a front-end and back-end architecture. The front-end estimates scale-aligned depth and camera poses from the input video, which are used to construct initial static 3D tracks. The back-end then iteratively refines both tracks and poses via joint motion optimization." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/476246a2ee49336015912e61cf61e16111fb83964215b0c5bea46f381660547a.jpg", + "image_caption": [ + "Figure 3. SyncFormer. The model takes previous estimates and their corresponding embeddings as input, and updates them iteratively. The 2D and 3D embeddings are processed in separate branches that interact via cross-attention." + ], + "image_footnote": [], + "bbox": [ + 94, + 446, + 480, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to as SyncFormer. In parallel, the SyncFormer also dynamically estimates the visibility probability $p^{\\mathrm{vis}}$ and dynamic probability $p^{\\mathrm{dyn}}$ for each trajectory, enabling an efficient bundle adjustment process to refine the camera poses $\\mathcal{P}$ .", + "bbox": [ + 89, + 665, + 482, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SyncFormer. As illustrated in Fig. 3, in every iteration, SyncFormer takes 2D embeddings, 3D embeddings, and camera poses as input, and updates the 2D trajectories $\\mathcal{T}^{2\\mathrm{d}}$ , 3D trajectories $\\mathcal{T}^{3\\mathrm{d}}$ , dynamic probabilities $p^{\\mathrm{dyn}}$ , and visibility scores $p^{\\mathrm{vis}}$ :", + "bbox": [ + 89, + 726, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {k + 1} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k + 1} ^ {3 \\mathrm {d}}, p _ {k + 1} ^ {\\mathrm {d y n}}, p _ {k + 1} ^ {\\mathrm {v i s}} = f _ {\\text {s y n c}} \\left(\\mathcal {T} _ {k} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k} ^ {3 \\mathrm {d}}, p _ {k} ^ {\\mathrm {d y n}}, p _ {k} ^ {\\mathrm {v i s}}, \\mathcal {P} _ {k}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 811, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{sync}}$ denotes the transformer-based update function. To better capture the distinct characteristics of 2D and 3D motion, the 2D and 3D trajectory updaters are modeled using separate attention layers. To reduce computational cost,", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "correlation embeddings are first encoded into a compact set of 2D and 3D proxy tokens using cross-attention. Then, information is exchanged between the 2D and 3D branches via a cross-attention layer between the respective proxy tokens. This design decouples the mutual influence between 2D and 3D tracking, which are updated in two different domains: the UV space for 2D trajectories and the camera coordinate space for 3D trajectories.", + "bbox": [ + 511, + 446, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2D and 3D Embeddings. As the input for SyncFormer, the 2D and 3D embeddings encode the status of current estimations while recording the neighbourhood information for updating. For 2D embeddings, we keep the same to Cotracker3 [33]. The 3D embeddings $\\mathbf{E}^{3\\mathrm{d}} = (Corr_{3\\mathrm{D}}, \\mathbf{e}^{\\mathrm{time}}, \\mathbf{e}^{\\mathrm{Gpos}}, p^{\\mathrm{dyn}}, p^{\\mathrm{vis}})$ contains the 3D correlation features $Corr_{3\\mathrm{D}}$ , the time embeddings $\\mathbf{e}^{\\mathrm{time}}$ , global position embeddings $\\mathbf{e}^{\\mathrm{Gpos}}$ and dynamic-visibility scores. The 3D correlation $Corr_{3\\mathrm{D}}$ is the main features for the 3D embedding. Different to 2D correlations, we expect 3D tracking branch updates the 3D position in the camera coordinate space. Therefore, we calculate the 3D correlations on the normalized point maps from front-end instead of depth map. Similar to 2D correlations, we construct multi-resolution point maps and compute the relative translations between each point and its neighbors within radius of 3. We then apply harmonic positional encoding to project these relative translations into high-dimensional feature representations, which are combined with semantic features to compute the final 3D correlations:", + "bbox": [ + 511, + 566, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC o r r _ {3 \\mathrm {D}} = \\left[ K \\left(\\frac {\\mathbf {x}}{k s} + \\delta , \\frac {\\mathbf {y}}{k s} + \\delta\\right): \\delta \\in \\mathbb {Z}, \\| \\delta \\| _ {\\infty} \\leq \\Delta \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 877, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6729", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d5dcde056f3c29d26c85be8c5f2d86ee291ffdd649590b0e25f164e1f6ca84aa.jpg", + "image_caption": [ + "Figure 4. Iterative Process of SyncFormer. The left subfigure shows the convergence curve of reprojection error, illustrating rapid reprojection error reduction. The right subfigures visualize the progressive alignment between 2D tracking results (in green) and projections of 3D (world) tracking transformed by camera poses (in red)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\Delta$ set to 3 and at multiple scales $s = 1,2,3,4$ , $K$ denotes the operator for encoding relative translation features. Besides, in order to encode the camera pose to assist the tracking estimations, we propose $\\mathbf{e}^{\\mathrm{Gpos}}$ which transforms the query points $\\mathcal{Q}_i$ into each frames with the current camera poses, denoted as anchor points $\\mathcal{Q}_i^{\\mathrm{anc}}$ . $\\mathbf{e}^{\\mathrm{Gpos}}$ is calculated from the relative translation between the current position and $\\mathbf{e}^{\\mathrm{Gpos}}$ , and they are then projected into high dimension with same tricks as above.", + "bbox": [ + 88, + 335, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Camera Motion Optimization. After each iteration, the updates of $\\mathcal{T}_{k + 1}^{2\\mathrm{d}},\\mathcal{T}_{k + 1}^{3\\mathrm{d}},p_{k + 1}^{\\mathrm{dyn}},p_{k + 1}^{\\mathrm{vis}}$ naturally form self-consistency constraints with the camera poses. These intrinsic constraints can be seamlessly incorporated into a bundle adjustment formulation. Specifically, we apply a weighted Procrustes analysis to register $\\mathcal{T}_{k + 1}^{3\\mathrm{d}}$ into the world coordinate frame, where the alignment weights are given by the dynamic scores. This coarse alignment process is differentiable and provides effective supervision for learning the dynamic probability. Then we fuse the aligned $\\mathcal{T}_{k + 1}^{3\\mathrm{d}}$ to world points $\\mathbf{P}_{k + 1}^{\\mathrm{world}}\\in \\mathbb{R}^{N\\times 3}$ , where the dynamic points are filtered with the estimated dynamic scores. After this, we apply a direct bundle adjustment to optimize the camera poses with paired $\\mathbf{P}_{k + 1}^{\\mathrm{world}},\\mathcal{T}_{k + 1}^{2\\mathrm{d}}$ and $p_{k + 1}^{\\mathrm{vis}}$ . In the next iteration, the updated camera poses $\\mathcal{P}_{k + 1}$ influence the subsequent loop through global position embedding. It is worth noting that the 2D and 3D trajectories are not updated via bundle adjustment; instead, the entire system is primarily driven by SyncFormer. As shown in Fig. 4, the reprojection errors decrease rapidly over iterations, while the 2D trajectories gradually become consistent with the 3D projections transformed by camera motion.", + "bbox": [ + 91, + 473, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 816, + 194, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. Our model is trained on a large collection of 17 datasets, each providing different forms of supervision. The training data can be broadly categorized into three types: (1) Posed RGB-D with tracking annotations,", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Posed RGB-D, (3) Pose-only or unlabeled data. For (1) containing Kubric [21], PointOdyssey [96] and Dynamic Replica [31], we add the full supervisions of camera poses, video depth, dynamic segmentation and 2D, 3D tracking. Within category (2), including VKITTI [7], TartanAir [80], Spring [53], DL3DV [44], BEDLAM [6], MVS-Synth [28], CO3Dv2 [61], COP3D [65], WildRGBD [83], ScanNet++ [12] and OmniObject3D [82], we leverage depth supervision to improve the video depth estimator. In addition, joint training losses are applied to ensure that the 3D tracking remains consistent with the estimated depth. As for category (3), it includes HOI4D [46], Ego4D [20], and Stereo4D [30]. For this type of data, we apply camera loss and joint training losses here, and follow [9] by using a monocular depth model [78] as a teacher to preserve relative depth accuracy. It is worth noting that Stereo4D [30] provides only very sparse depth on valid 3D tracks. Therefore, we choose not to use its annotations and instead treat it as pose-only data, given that it is sourced from Internet videos with rich scene diversity.", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our training recipe consists of several stages. Stage 1. We first train the front end model to jointly estimate video depth and camera poses on category (1) and (2) datasets which sums up to 14 datasets. Training is conducted with mixed BF16 precision, while the DPT and camera tracker modules use full precision to ensure stable optimization. We use the AdamW optimizer with a learning rate of $5 \\times 10^{-5}$ and apply gradient clipping with a threshold of 0.1. Stage 1 training is performed for 200k iterations on 64 H20 GPUs. We shuffle the video length from 1-24 during the training. Stage 2. At this stage, we initialize SyncFormer using the category (1) datasets, where ground-truth depth is provided and camera poses are initialized as identity matrices. We load the Cotracker3 [32] checkpoint to initialize the 2D tracking branch. This stage training takes 3 days of 100k iterations on 8 H20 GPUs. The video length is shuffle from 12-48 during the training. Stage 3. Finally, we fixed the alternation-attention layers in front end and train the whole pipeline in all datasets for 20 hours to converge.", + "bbox": [ + 511, + 393, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 708, + 645, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our model across all sub-tasks, including 3D tracking (Sec. 4.1), and dynamic 3D reconstruction (Sec. 4.2). In addition, we conduct comprehensive ablation studies (Sec. 4.3) to analyze the impact of key design choices and to demonstrate the effectiveness of unified modeling and scaling up training.", + "bbox": [ + 511, + 733, + 906, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. 3D Point Tracking", + "text_level": 1, + "bbox": [ + 511, + 833, + 692, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We evaluate our model and compare it with existing baselines on TAPVid-3D [37], a comprehensive benchmark spanning diverse scenarios including Driving, Ego", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6730", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6582915d64767e3dc0648d192ca010228b6df6d5957b5fc696bf7bf02ed09c20.jpg", + "table_caption": [ + "Table 1. 3D tracking results on the TAPVid-3D benchmark. We report the 3D Average Jaccard (AJ), Average 3D Position Accuracy $(\\mathrm{APD}_{3D})$ , and Occlusion Accuracy (OA) across the Aria, DriveTrack, and PStudio subsets. $offl^{+}$ and $offl^{-}$ denote our offline model with/without considering the camera motion, respectively. COL, Univ2, and Mega are abbreviations for COLMAP, UnidepthV2, and MegaSAM. The best and the second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodsTypeDepth / Cam PoseAriaDriveTrackPStudioAverage
AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑
BootsTAPIRType ICOL9.114.578.611.818.683.86.911.681.89.314.981.4
TAPTRType IUniv215.724.287.812.419.184.87.313.584.311.818.985.6
LocoTrackType IUniv215.124.083.513.019.882.87.213.180.111.819.082.3
CoTracker3Type ICOL8.012.378.611.719.181.78.113.577.29.315.079.1
Univ215.824.488.913.519.987.19.213.884.212.819.486.7
Mega20.430.189.814.120.388.517.427.285.017.325.987.8
SpatialTrackerType IIUniv213.620.990.58.314.582.88.015.075.810.016.883.0
Mega15.923.890.17.713.585.215.325.278.113.020.884.5
SceneTrackerType IIUniv2-23.1-6.8--12.7---14.2-
DELTAType IIUniv216.624.486.814.622.585.88.215.076.413.120.683.0
Ours-offl-Type IIUniv218.626.390.816.424.390.218.127.686.717.726.089.2
Mega22.332.293.715.823.090.018.228.687.318.727.990.5
TAPIP3DType IIIMega23.532.891.214.921.882.618.127.785.518.827.486.4
Ours-offl+Type IIIMega24.735.293.916.023.490.118.628.786.119.829.190.0
Full-ours24.634.793.617.626.190.821.932.187.421.231.090.6
", + "bbox": [ + 94, + 145, + 901, + 387 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/999151d994b3970693ff4379e799ce3cddaa0fbbf6a835953cfda96d7b2a7a59.jpg", + "table_caption": [ + "Table 2. Video depth evaluation. Type I represents the methods specialized in video depth estimation, while Type II are neural reconstruction model, jointly recovering the geometry and camera motion from the video. Type III denotes the SoTA optimization-based method. The best and the second best results are highlighted." + ], + "table_footnote": [], + "table_body": "
Method / MetricsAverageKITTI [19]TUM Dyn [66]Bonn [59]Sintel [52]
AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)
DepthCrafter [27]0.1430.8570.1110.8850.1230.8730.0660.9790.2720.693
VDA [9]0.1540.8820.0800.9510.1180.9200.0490.9820.3700.674
DUSt3R [79]0.2400.7660.1240.8490.1870.7920.1740.8350.4750.591
MonST3R [91]0.1710.8020.0830.9340.1970.7260.0610.9540.3430.594
CUT3R [77]0.1860.8140.1040.8990.1080.8470.0680.9500.4660.560
VGGT [74]0.1040.8810.0510.9660.0680.9390.0560.9630.2420.659
MegaSAM [74]0.0930.8940.0690.9160.0810.9350.0370.9770.1850.746
Ours0.0810.9100.0520.9730.0450.9760.0280.9880.1990.703
", + "bbox": [ + 99, + 441, + 895, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "centric, and Studio. This benchmark consists of 4,569 evaluation videos, where the video length varies from 25 to 300 frames, and three 3D point tracking metrics are reported. Specifically, Occlusion Accuracy (OA) measures the precision of occlusion predictions; $\\mathrm{APD}_{3D}$ denotes the average percentage of estimated errors within multiple threshold scales $\\delta$ ; and Average Jaccard (AJ) quantifies the accuracy of both position and occlusion estimation.", + "bbox": [ + 88, + 619, + 480, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Settings. The existing baselines can be broadly categorized into three types. Type I: 2D trackers followed by depth lifting. We report the current state-of-the-art 2D tracking models, CoTracker3 [32] and BootsTAPIR [16], as representatives of this category. Type II: 3D trackers in camera space. We include SpatialTrackerV1 [86] and DELTA [54], evaluated with two state-of-the-art depth estimators, i.e., UnidepthV2 [60] and MegaSAM [42]. Notably, MegaSAM is an optimization-based SLAM system for estimating consistent video depth,", + "bbox": [ + 88, + 750, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which generally performs better than feedforward models such as UnidepthV2. To ensure fair comparisons under the same camera-space 3D tracking setup, our 3D tracker module is also evaluated in combination with these depth estimators. Type III: 3D trackers in world space. We compare our fully end-to-end model with the recently released TAPIP3D [90], which requires consistent video depth and camera poses as input, provided by MegaSAM. For a thorough and fair comparison, we also report results where our depth and pose estimations are replaced by those from MegaSAM.", + "bbox": [ + 511, + 619, + 906, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results and Analysis. In the TAPVid-3D benchmark, our method achieves the best results across various settings, consistently outperforming all baselines under comparable conditions as in Tab. 1. From those detailed comparisons, we can draw several conclusions from it.", + "bbox": [ + 511, + 789, + 906, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Better depth estimation is crucial for 3D tracking: For Type I methods, the 3D tracking performance largely re", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6731", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "flects the accuracy of the underlying 2D tracks and the associated depth predictions. Intriguingly, CoTracker3 + MegaSAM achieves substantial improvements over CoTracker3 + UniDepthV2, with 17.3 vs. 12.8 in AJ, 25.9 vs. 19.4 in $APD_{3D}$ , and 87.8 vs. 86.7 in OA. This represents the best performance among Type I methods. However, our method Ours-offl $^{-}$ , which uses the weaker depth estimator UniDepthV2, still outperforms CoTracker3 with MegaSAM, achieving 18.7 vs. 17.3 in AJ and 27.4 vs. 25.9 in $APD_{3D}$ . This is because Type I methods rely purely on back-projection, making them more sensitive to depth inconsistencies. In contrast, our method directly predicts 3D trajectories using a spatial-temporal transformer, which provides inherent temporal smoothness and robustness to noisy inputs.", + "bbox": [ + 102, + 90, + 482, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Camera motion decomposition improves 3D tracking: As shown in the comparison between Ours-offl $^+$ MegaSAM and Ours-offl $^+$ MegaSAM, incorporating camera pose estimation clearly improves 3D tracking accuracy. For example, on the Aria subset, our method achieves 24.7 vs. 22.3 in AJ and 35.2 vs. 32.2 in $APD_{3D}$ . This improvement is largely due to the design of the TAPVid-3D benchmark, which includes a substantial number of background points for evaluation in Aria. These background points primarily reflect camera motion and are particularly challenging for methods that only track in camera space, especially due to frequent out-of-view scenarios. In DriveTrack and Pstudio, the improvements are less, as DriveTrack includes only dynamic points on moving vehicles, while Pstudio contains static scenes with no camera motion.", + "bbox": [ + 91, + 318, + 482, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Dynamic 3D Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 570, + 344, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Video Depth Evaluation", + "text_level": 1, + "bbox": [ + 89, + 593, + 303, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. To illustrate the effectiveness and generalization ability of our method, we evaluate our video depth estimation method across four mainstream datasets, i.e., KITTI [19], Sintel [52], Bonn [59], TUM Dynamics [66]. These datasets encompass both indoor and outdoor scenes, featuring video sequences ranging from 50 to 110 frames, providing a comprehensive benchmark for assessing depth estimation consistency across varied environments.", + "bbox": [ + 89, + 612, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare our approach against three categories of methods. Type I includes SoTA video depth methods, DepthCrafter [27] and Video Depth Anything [87]. Type II consists of SoTA deep reconstruction approaches, i.e. DUST3R [79], MonST3R [91], CUT3R [77] and VGGT [74] Type III denotes the SoTA dynamic Structurefrom-Motion (SfM) system, MegaSAM [42]. It leverages an optimization-based paradigm to jointly estimate consistent video depths and camera poses, with constraints enforced by optical flow and monocular depth priors.", + "bbox": [ + 89, + 734, + 482, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We evaluate the performance of video depth esti", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c368c64a7247880b9c14f32b1f9e9c1bcaec64faa186aa10109a03e68f2fc872.jpg", + "table_caption": [ + "Table 3. Evaluation on Camera Pose Estimation on TUMdynamic [66], Lightspeed [12], and Sintel [52]. All values are absolute trajectory error (ATE), relative pose error (RPE) for translation and rotation. The best and the second best are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodTUM-dynamic (ATE / RPEt / RER)Lightspeed (ATE / RPEt / RER)Sintel (ATE / RPEt / RER)
Particle-SfM [95]-0.185 / 0.075 / 2.9900.129 / 0.031 / 0.535
Robust-CVD [36]0.153 / 0.026 / 3.528-/-/-0.360 / 0.154 / 3.443
CasualSAM [94]0.071 / 0.010 / 1.712-/-/-0.141 / 0.035 / 0.615
DUST3R [79]0.140 / 0.106 / 3.2860.412 / 0.177 / 20.1000.290 / 0.132 / 7.869
CUT3R [77]0.046 / 0.015 / 0.4730.274 / 0.067 / 1.5610.213 / 0.066 / 0.621
MonST3R [91]0.098 / 0.019 / 0.9350.149 / 0.046 / 1.2100.078 / 0.038 / 0.490
VGGT [74]0.021 / 0.013 / 0.3270.226 / 0.086 / 1.7290.082 / 0.043 / 1.253
MegaSAM [42]0.013 / 0.011 / 0.3400.105 / 0.040 / 0.9960.023 / 0.008 / 0.060
Ours-Front0.038 / 0.022 / 0.4800.203 / 0.079 / 1.6890.075 / 0.045 / 0.805
Ours0.012 / 0.010 / 0.3050.134 / 0.039 / 1.3400.054 / 0.027 / 0.288
", + "bbox": [ + 514, + 156, + 903, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mation using geometric accuracy metrics. To ensure fair comparisons with previous works [9, 77, 91], we follow their approach by aligning the predicted video depth to the ground truth using a shared scale and shift, and then compute the Absolute Relative Error (AbsRel) and $\\delta_{1}$ metrics.", + "bbox": [ + 511, + 328, + 903, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Tab. 2, our method clearly outperforms all existing approaches. Specifically, we surpass our baseline in Type II, VGGT [74], by a significant margin. On average, our method achieves an AbsRel of 0.081 compared to 0.104 $(-22.1\\%)$ , and a $\\delta_{1.25}$ of 0.910 compared to 0.881 $(+3.3\\%)$ . Besides, compared to MegaSAM (Type III), our method also maintain clear advantages with 0.081 v.s. 0.093 in AbsRel, and 0.910 v.s. 0.894 in $\\delta_{1.25}$ . It is important to note that MegaSAM [42] usually needs 5-10 min for a 100 frames video, while our method only takes 5-10 seconds for each which nearly achieves $50 \\times$ faster.", + "bbox": [ + 511, + 404, + 906, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Camera Poses", + "text_level": 1, + "bbox": [ + 511, + 592, + 658, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the accuracy of camera motion estimation, we conduct the comparisons on Sintel [52], TUM Dynamics [66], and Lightspeed [62]. Sintel and Lightspeed contains numerous challenging dynamic scenes with large egomotion and object motions, and they are both the synthetic data. TUM Dynamics [66] is a real data, captured by well-calibrated RGBD sensors. We report Absolute Translation Error (ATE), Relative Rotation Error (RPE rot) and Relative Translation Error (RTE) after Sim(3) alignment with the ground truth, as in [77, 91]. Shown in Tab. 3, our method outperforms all regression-based methods, much better than our baseline, VGGT [74] and on par with MegaSAM [42]. Besides, the table illustrates the significance of back end optimization. After joint motion optimization, the pose estimations become nearly twice accurate than before.", + "bbox": [ + 511, + 611, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 511, + 848, + 687, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our ablation study investigates the impact of training data, joint training, and the SyncFormer design on 2D and 3D", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6732", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78cdb4cc883ead795b66e2d91c83c06eed037579c7ec3850282206577a85a974.jpg", + "image_caption": [ + "(a) Query" + ], + "image_footnote": [], + "bbox": [ + 102, + 97, + 165, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/700498ee5487a950ac2670fbaf2347360f162b6bee00609b7673d92b1a1fb551.jpg", + "image_caption": [ + "(b) w/o Joint" + ], + "image_footnote": [], + "bbox": [ + 178, + 88, + 241, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20afaee421b5967959cc40756a53ce0e38b63b7556a6e3deb097816019d29e5e.jpg", + "image_caption": [ + "(c) w/ Joint", + "Figure 5. The influence of Joint Training." + ], + "image_footnote": [], + "bbox": [ + 256, + 88, + 318, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9790ce95f5932ba6507ea61a019295a5602bbb843c9f1894780619e9b32404f.jpg", + "image_caption": [ + "(d) w/o Joint" + ], + "image_footnote": [], + "bbox": [ + 331, + 88, + 395, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5db04851e56198cd2ecba82f7013500950e96773d52fa83975c606d90788caff.jpg", + "image_caption": [ + "(e) w/ Joint" + ], + "image_footnote": [], + "bbox": [ + 410, + 88, + 470, + 133 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3543dbd07e45ff76b438e97c0eedf2d1f90faef6b200eaa1c8691dc1cf703d8e.jpg", + "table_caption": [ + "Table 4. TAP-Vid DAVIS results. Cotracker3-3D is a fine-tuned model by adding simple 3D project layer. Mean represents the average of AJ and ${\\delta }_{\\text{avg }}^{\\text{vis }}$ ." + ], + "table_footnote": [], + "table_body": "
MethodLift 3DAJ ↑δvisavg ↑OA ↑Mean ↑
TAPTRX63.076.191.169.5
LocoTrackX62.975.387.269.1
CoTracker3X64.476.991.270.7
DELTA3D-Dec62.776.788.269.7
SpatialTrackerTri-plane61.176.389.568.7
CoTracker3-3D3D-Proj51.665.285.358.4
OursSyncFormer64.977.591.071.2
", + "bbox": [ + 91, + 244, + 480, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "point tracking. For depth estimation, we analyze different loss functions and training strategies to improve the generalizability of video depth prediction.", + "bbox": [ + 89, + 398, + 482, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Point Tracking. To illustrate the gains brought by scaling up the 3D point tracking to a wider range of data, we naively train a base model (Base@K in Table 5), and an advanced one, Base@V,K,P, which is trained on VKITTI, Kubric, and PointOdyssey. As the table shows, Base@V,K,P outperforms Base@K by a clear margin. Meanwhile, our final version is clearly better than these two. Besides, different from Base@K, Base@V,K,P is jointly trained on VKITTI, which brings very significant improvements on a similar type of real data: DriveTrack shows 14.7 vs. 7.4 in AJ and 21.9 vs. 13.3 in $\\mathrm{APD}_{3D}$ . Fig. 5 qualitatively demonstrates the meaning of joint training, i.e., joint training contributes to minimizing the 3D tracking drifts when the model is trained on new patterns of data.", + "bbox": [ + 89, + 445, + 483, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2D Point Tracking. We report a naive alternative to SyncFormer by modifying the final output layer of CoTracker3 [33] to output 2D and 3D tracking. The 2D and 3D embeddings are directly concatenated and projected back to the original dimension using an inserted, learnable linear layer. We name this naive baseline as Cotacker3-3D. As shown in Tab. 4, a simple adaptation of 3D lifting leads to a significant drop in 2D tracking accuracy—AJ drops from 64.4 to 51.6, $\\delta$ from 76.9 to 65.2, and OA from 91.2 to 85.3. This degradation occurs because the 2D and 3D correlation signals become entangled, disrupting the model's ability to focus on reliable features. Moreover, the update dynamics in 2D and 3D differ substantially, as they occur in separate spaces—UVD space for 2D and camera space for 3D. We also report comparisons with other 3D lifting techniques. These results validate our SyncFormer design, which effec", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tively lifts tracking into 3D while preserving—or even improving—accuracy in 2D.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Depth Estimation. We ablate different training datasets and their coverage to assess their impact on model performance. Additionally, we study the influence of different loss functions applied to real data, i.e., the pearson loss and full depth losses. As shown in Tab. 2, if we applied the full losses for depth will cause the obvious performance drop in the synthetic data, i.e. Sintel. On the contrary, if we only used the synthetic data to train the model, we found the performance in the real dataset, especially KITTI will heavily drop. Therefore, we leverage the strategy of using different losses for real and synthetic data to offset the influence of domain gap and different errors distribution. As shown in Tab. 2, Ours-Synthetic is our method trained only on the synthetic datasets. Ours-Real-Full is our method using the full depth loss in the real data, where the model was influenced by the noised distributions in the real dataset. The noised of real data has the negative influence on the model's influence on the synthetic data. Meanwhile, the different errors pattern also worsen the model's zero-shot capabilities in the real data.", + "bbox": [ + 511, + 122, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/50945381b81a3748a4e251fb7fa520d5e1d5baa1f255422abf9f0707823cbeb3.jpg", + "table_caption": [ + "Table 5. Ablation of Training Data and Joint Training. K, V, P represents Kubric [21], VKITTI [7] and PointOdyssey [96] in training. Inference depths are provided by Unidepthv2." + ], + "table_footnote": [], + "table_body": "
MethodJoint TrainingAriaDriveTrackPStudio
AJAPD3DAJAPD3DAJAPD3D
Base@KNo16.024.47.413.312.620.3
Base@V, K, PYes15.724.114.721.917.227.4
", + "bbox": [ + 514, + 470, + 903, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces SpatialTrackerV2, a feedforward and scalable approach for 3D point tracking in monocular videos. Built upon a deep exploration of widely-used low- and mid-level representations of motion and scene geometry, our method unifies consistent scene geometry, camera motion, and pixel-wise 3D motion into a fully differentiable end-to-end pipeline. SpatialTrackerV2 accurately reconstructs 3D trajectories from monocular videos, achieving strong quantitative results on public benchmarks and demonstrating robust performance on casually captured Internet videos. We believe SpatialTrackerV2 establishes a solid foundation for real-world motion understanding and brings us a step closer to spatial intelligence by exploring large-scale vision data.", + "bbox": [ + 511, + 568, + 906, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 797, + 665, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by Ant Group Research Intern Program, Zhejiang Provincial Natural Science Foundation of China (No. LR25F020003) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6733", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abhishek Badki, Hang Su, Bowen Wen, and Orazio Gallo. L4p: Low-level 4d vision perception unified. arXiv preprint arXiv:2502.13078, 2025. 2", + "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 3", + "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3", + "[4] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. CoRR, abs/2501.02690, 2025. 2", + "[5] Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3.1 - A model zoo for robust monocular relative depth estimation. CoRR, abs/2307.14460, 2023. 3", + "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737. IEEE, 2023. 5", + "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 8", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[9] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. CoRR, abs/2501.12375, 2025. 3, 5, 6, 7", + "[10] Zequn Chen, Jiezhi Yang, and Heng Yang. Pref3r: Pose-free feed-forward 3d gaussian splatting from variable-length image sequence. CoRR, abs/2411.16877, 2024. 2", + "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In European Conference on Computer Vision, pages 306–325. Springer, 2024. 2", + "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 7", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3", + "[14] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 2", + "[15] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2", + "[16] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, pages 3257-3274, 2024. 3, 6", + "[17] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 3", + "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011. Computer Vision Foundation / IEEE Computer Society, 2018. 3", + "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6, 7", + "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 5", + "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5, 8", + "[22] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, Wenping Wang, and Yuan Liu. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. CoRR, abs/2501.03847, 2025. 2", + "[23] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2", + "[24] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3", + "[25] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 3", + "[26] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE TPAMI, 46(12):10579-10596, 2024. 3", + "[27] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. CoRR, abs/2409.02095, 2024. 2, 3, 6, 7", + "[28] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6734", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "stereopsis. In CVPR, pages 2821-2830. Computer Vision Foundation / IEEE Computer Society, 2018. 5", + "[29] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy J. Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. CoRR, abs/2412.06016, 2024. 2", + "[30] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint, 2024. 5", + "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 5", + "[32] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3, 5, 6", + "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 4, 8", + "[34] Yoni Kasten, Wuyue Lu, and Haggai Maron. Fast encoder-based 3d from casual videos via point track processing. arXiv preprint arXiv:2404.07097, 2024. 2", + "[35] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In Proceedings of the IEEE international conference on computer vision, pages 1521–1529, 2017. 3", + "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1611-1621. Computer Vision Foundation / IEEE, 2021. 7", + "[37] Skanda Koppula, Ignacio Rocco, Yi Yang, Joseph Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. In NeurIPS, 2024. 2, 5", + "[38] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2", + "[39] Hongdong Li and Richard Hartley. Five-point motion estimation made easy. In 18th International Conference on Pattern Recognition (ICPR'06), pages 630-633. IEEE, 2006. 3", + "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2024. 2", + "[41] Xiaodi Li, Zongxin Yang, Ruijie Quan, and Yi Yang. Drip: Unleashing diffusion priors for joint foreground and alpha prediction in image matting. Advances in Neural Information Processing Systems, 37:79868-79888, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In CVPR, 2025. 2, 6, 7", + "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3", + "[44] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 5", + "[45] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE transactions on pattern analysis and machine intelligence, 33(5):978-994, 2010. 2", + "[46] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In CVPR, pages 20981-20990. IEEE, 2022. 5", + "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2", + "[48] Haozhe Lou, Yurong Liu, Yike Pan, Yiran Geng, Jianteng Chen, Wenlong Ma, Chenglong Li, Lin Wang, Hengzhen Feng, Lu Shi, Liyi Luo, and Yongliang Shi. Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation. CoRR, abs/2408.14873, 2024. 2", + "[49] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 3", + "[50] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 3", + "[51] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3", + "[52] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 6, 7", + "[53] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 5", + "[54] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. arXiv preprint arXiv:2410.24211, 2024. 2, 3, 6" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6735", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 3", + "[56] Dantong Niu, Yuvan Sharma, Haoru Xue, Giscard Biamby, Junyi Zhang, Ziteng Ji, Trevor Darrell, and Roei Herzig. Pretraining auto-regressive robotic models with 4d representations. arXiv preprint arXiv:2502.13142, 2025. 2", + "[57] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8089-8099, 2024. 2", + "[58] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 3", + "[59] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 6, 7", + "[60] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segù, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116. IEEE, 2024. 3, 6", + "[61] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV. IEEE, 2021. 5", + "[62] Chris Rockwell, Joseph Tung, Tsung-Yi Lin, Ming-Yu Liu, David F. Fouhey, and Chen-Hsuan Lin. Dynamic camera poses and where to find them. In CVPR, 2025. 7", + "[63] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2", + "[64] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2", + "[65] Samarth Sinha, Roman Shapovalov, Jeremy Reizenstein, Ignacio Rocco, Natalia Neverova, Andrea Vedaldi, and David Novotny. Common pets in 3d: Dynamic new-view synthesis of real-life deformable categories. In CVPR, pages 4881-4891. IEEE, 2023. 5", + "[66] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS, pages 573-580. IEEE, 2012. 6, 7", + "[67] Tutian Tang, Minghao Liu, Wenqiang Xu, and Cewu Lu. Kalib: Markerless hand-eye calibration with keypoint tracking. CoRR, abs/2408.10562, 2024. 2", + "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3", + "[69] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a mod-" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21-22, 1999 Proceedings, pages 298-372. Springer, 2000. 3", + "[70] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. CoRR, abs/2412.13389, 2024. 3", + "[71] Bo Wang, Jian Li, Yang Yu, Li Liu, Zhenping Sun, and Dewen Hu. Scenetracker: Long-term scene flow estimation network. arXiv preprint arXiv:2403.19924, 2024. 3", + "[72] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3", + "[73] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 3", + "[74] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 6, 7", + "[75] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 3", + "[76] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2", + "[77] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025.6, 7", + "[78] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 5", + "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 6, 7", + "[80] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5", + "[81] Zhouxia Wang, Yushi Lan, Shangchen Zhou, and Chen Change Loy. Objctrl-2.5 d: Training-free object control with camera poses. arXiv preprint arXiv:2412.07721, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6736", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR, pages 803-814. IEEE, 2023. 5", + "[83] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBD objects in the wild: Scaling real-world 3d object learning from RGB-D videos. In CVPR, pages 22378-22389. IEEE, 2024. 5", + "[84] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 3", + "[85] Yuxi Xiao, Nan Xue, Tianfu Wu, and Gui-Song Xia. Level-S $^2$ fM: Structure From Motion on Neural Level Set of Implicit Surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17205-17214, 2023. 3", + "[86] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2, 3, 6", + "[87] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3, 7", + "[88] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2025. 2, 3", + "[89] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3906-3915. IEEE, 2022. 3", + "[90] Bowei Zhang, Lei Ke, Adam W Harley, and Katerina Fragkiadaki. Tapip3d: Tracking any point in persistent 3d geometry. arXiv preprint arXiv:2504.14717, 2025. 3, 6", + "[91] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 6, 7", + "[92] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3", + "[93] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 3", + "[94] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 7" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In *ECCV*. Springer, 2022. 2, 7", + "[96] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 8" + ], + "bbox": [ + 516, + 90, + 903, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6737", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_model.json b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80a639cd75de310b39d3754649c1b805a1959764 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_model.json @@ -0,0 +1,2633 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.179, + 0.887, + 0.221 + ], + "angle": 0, + "content": "Yuxi Xiao\\(^{1*}\\) Jianyuan Wang\\(^{2}\\) Nan Xue\\(^{3}\\) Nikita Karaev\\(^{2,4}\\) Yuri Makarov\\(^{4}\\) Bingyi Kang\\(^{5}\\) Xing Zhu\\(^{3}\\) Hujun Bao\\(^{1}\\) Yujun Shen\\(^{3}\\) Xiaowei Zhou\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.224, + 0.873, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University \\(^{2}\\)Oxford \\(^{3}\\)Ant Group \\(^{4}\\)Pixelwise AI \\(^{5}\\)Bytedance Seed" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.254, + 0.497, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.478, + 0.495, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.252, + 0.888, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.482, + 0.888, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.691, + 0.908, + 0.734 + ], + "angle": 0, + "content": "Figure 1. SpatialTrackerV2 produces consistent 3D scene geometry, camera poses, and 3D point trajectories all at once from monocular videos of arbitrary scenarios, e.g., robotic manipulation, first-person egocentric views, and dynamic sports (drifting and skating) shown in this figure. Try our online demo at https://huggingface.co/spaces/Yuxihenry/SpatialTrackerV2." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.742, + 0.329, + 0.758 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.486, + 0.88 + ], + "angle": 0, + "content": "We present SpatialTrackerV2, a feed-forward 3D point tracking method for monocular videos. Going beyond modular pipelines built on off-the-shelf components for 3D tracking, our approach unifies the intrinsic connections between point tracking, monocular depth, and camera pose estimation into a high-performing and feedforward 3D point tracker. It decomposes world-space 3D motion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.909, + 0.884 + ], + "angle": 0, + "content": "into scene geometry, camera ego-motion, and pixel-wise object motion, with a fully differentiable and end-to-end architecture, allowing scalable training across a wide range of datasets, including synthetic sequences, posed RGB-D videos, and unlabeled in-the-wild footage. By learning geometry and motion jointly from such heterogeneous data, SpatialTrackerV2 outperforms existing 3D tracking methods by \\(30\\%\\), and matches the accuracy of leading dynamic 3D reconstruction approaches while running \\(50\\times\\) faster." + }, + { + "type": "page_footnote", + "bbox": [ + 0.089, + 0.887, + 0.456, + 0.902 + ], + "angle": 0, + "content": "* Partially completed during Ant internship. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6726" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.266 + ], + "angle": 0, + "content": "3D point tracking aims to recover long-term 3D trajectories of arbitrary points from monocular videos. As a universal dynamic scene representation, it has recently shown strong potential in diverse applications, including robotics [48, 56, 67], video generation [22, 29, 57, 81], and 3D/4D reconstruction [10, 34, 38, 76]. Compared to parametric motion models (e.g., SMPL [47], MANO [63], skeletons, or 3D bounding boxes), it offers greater flexibility and generalization over various real-world scenes, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.267, + 0.483, + 0.568 + ], + "angle": 0, + "content": "Existing solutions [1, 4, 54, 75, 86] of 3D point tracking extensively explored the well-developed low/mid-level vision models, such as optical flow [45, 68] and monocular depth estimation [27, 87], and took benefits from 2D point tracking models [15, 23, 33]. Among them, optimization-based methods [42, 75, 95] distill the optical flow, monocular depth models and camera motions for each given monocular video with promising results obtained, while being computationally expensive due to their per-scene optimization designs. SpatialTracker [86] moved forward to efficient 3D point tracking with a feed-forward model, and the more recent works [1, 4, 54] explored different architecture designs and rendering constraints to achieve higher-quality 3D tracking. Nevertheless, the feed-forward solutions are limited to training data scalability issues due to the need for ground-truth 3D tracks as supervision, which downgrades the tracking quality in real-world casual captures. Moreover, overlooking the inherent interplay between camera motion, object motion, and scene geometry results in error entanglement and accumulation across modules." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.78 + ], + "angle": 0, + "content": "These limitations motivate our core insights: (1) The reliance on ground-truth 3D trajectories constrains the scalability of existing feed-forward models, highlighting the need for designs that can generalize across diverse and weakly-supervised data sources. (2) The absence of joint reasoning over scene geometry, camera motion, and object motion leads to compounded errors and degraded performance, underscoring the importance of disentangling and explicitly modeling these motion components. To address these challenges, we decompose 3D point tracking into three distinct components: video depth, ego (camera) motion, and object motion, and integrate them within a fully differentiable pipeline that supports scalable joint training across heterogeneous data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our SpatialTrackerV2, a front-end and back-end architecture is proposed. The front end is a video depth estimator and camera pose initializer, adapted from typical monocular depth prediction frameworks [88] with attention-based temporal information encoding [74]. The predicted video depths and camera poses are then fused through a scale-shift estimation module, which ensures consistency between the depth and motion predictions. The back end" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "consists of a proposed Joint Motion Optimization Module, which takes the video depth and coarse camera trajectories as input and iteratively estimates 2D and 3D trajectories, along with trajectory-wise dynamics and visibility scores. This enables an efficient bundle adjustment process for optimizing camera poses in the loop. At its core lies a novel SyncFormer, which separately models the 2D and 3D correlations in two branches, connected by multiple cross-attention layers. This design mitigates mutual interference between 2D and 3D embeddings and allows the model to update representations in two distinct spaces, namely the image (UV) space and the camera coordinate space. Furthermore, benefiting from this dual-branch design, bundle adjustment can be effectively applied to jointly optimize camera poses as well as the 2D and 3D trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.483 + ], + "angle": 0, + "content": "This unified and differentiable pipeline makes large-scale training on diverse datasets possible. For RGB-D datasets with camera poses, we jointly train 3D tracking using consistency constraints from ground-truth depth and camera poses for static points, while dynamic points seamlessly contribute to the optimization. For video datasets that provide only camera pose annotations and lack depth information, we leverage consistency among camera poses, and 2D and 3D point tracking to drive the model's optimization. Relying on this framework, we successfully scale up training of the entire pipeline across 17 datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.905, + 0.651 + ], + "angle": 0, + "content": "Evaluations on the TAPVid-3D benchmark [37] show that our method sets a new state-of-the-art in 3D point tracking, achieving 21.2 AJ and \\(31.0\\mathrm{APD}_{3D}\\), surpassing DELTA [54] with relative improvements of \\(61.8\\%\\) and \\(50.5\\%\\), respectively. Additionally, extensive experiments on dynamic reconstruction show our superior results on consistent video depth and camera poses estimation. Specifically, SpatialTrackerV2 beats the best dynamic reconstruction method, MegaSAM [42], on most video depth datasets and achieves comparable results on various camera pose benchmarks, while its inference speed is \\(50\\times\\) faster." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.652, + 0.68 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.719 + ], + "angle": 0, + "content": "This section covers relevant literature on 3D tracking, depth, and camera pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.663, + 0.744 + ], + "angle": 0, + "content": "2.1. Point tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "PIPs [23] revisited the 2D point tracking task first introduced in [64] and proposed a deep learning approach to solve it. TAP-Vid [14] redefined the problem and introduced both a benchmark and a simple architecture, TAP-Net. Subsequently, the performance was improved in TAPIR [15] by combining the global matching capabilities of TAP-Net with the local refinement offered by PIPs. CoTracker [33] pioneered tracking through occlusions using a transformer architecture combined with joint attention, followed by TAPTR [40] and LocoTrack [11], which im" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6727" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "proved efficiency and introduced 4D correlation volumes. Recently, BootsTAPIR [16] and CoTracker3 [32] explored the use of unlabeled data to achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.483, + 0.348 + ], + "angle": 0, + "content": "While 2D point tracking has been extensively studied, 3D point tracking remains a relatively new field. The first method to demonstrate 3D point tracking capabilities was the test-time optimization-based OmniMotion [75]. Later, SpatialTracker [86] introduced the first feed-forward 3D point tracker by combining a 2D point tracker [33] with depth priors from a monocular depth estimator [3]. SceneTracker [71] proposed a new architecture for 3D tracking with depth priors, while DELTA [54] improved efficiency and achieved dense 3D tracking. Recently, TAPIP3D [90] improved 3D tracking robustness by lifting image features into the world coordinate space and performing tracking there. All of these 3D tracking models were trained on small synthetic datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Unlike these methods, we present a scalable 3D tracking framework trained on a collection of both real and synthetic datasets, while also explicitly modeling camera motion to improve performance on egocentric videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.421, + 0.262, + 0.437 + ], + "angle": 0, + "content": "2.2. Depth estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.564 + ], + "angle": 0, + "content": "Early methods such as Eigen et al. [17] introduced single-view depth estimation using CNNs, but were limited by dataset scale and poor generalization [18, 89]. MiDaS [5] improved this by mixing datasets for broader coverage, and ZoeDepth [3] adapted it for metric depth, though ambiguity from missing camera intrinsics remained. Later, Metric3D [26] and UniDepth [60] addressed this by jointly estimating intrinsics and normalized depth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.626 + ], + "angle": 0, + "content": "With large-scale pretraining (e.g., diffusion [13, 41], DINO [8]), recent models like Marigold [70] and DepthAnything [87, 88] have significantly advanced zero-shot depth. Extensions to video [9, 27] have also emerged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.702 + ], + "angle": 0, + "content": "In this work, we build on DepthAnythingV2 and extend it to video, aiming for not just consistent depth, but a unified framework that also predicts camera poses and tracks, with proper scale alignment—posing new challenges beyond static depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.274, + 0.727 + ], + "angle": 0, + "content": "2.3. Camera estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Traditional camera pose estimation methods [24, 58] rely on image-to-image point correspondences using keypoint detectors (e.g., SIFT [49, 50], SURF [2]) and matching techniques such as nearest neighbors, followed by geometric algorithms like the five-point and eight-point methods [24, 25, 39, 55, 85]. Bundle Adjustment [69] is also commonly employed to further enhance accuracy. Recently, direct regression approaches using neural networks [35, 43, 51, 84, 92] have emerged, aiming to overcome limitations in sparse-view scenarios or when correspondences are unreliable. Diffusion models, such as PoseDiffusion [72] and RayDif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "fusion [93], have also been explored, offering strong accuracy but suffering from high inference costs. In contrast, the camera head of VGGsfm [73] or VGGT [74] adopts an iterative refinement paradigm similar to RAFT [68], striking a balance between accuracy and inference cost, and enabling estimation of both extrinsic and intrinsic parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.605, + 0.21 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Given a video with \\(T\\) frames \\((\\mathcal{I}^t)_{t=1}^T\\) and \\(N\\) query points \\(\\mathcal{Q}_i = (x_i, y_i) \\in \\mathbb{R}^2\\), \\(i = 1, \\dots, N\\), the goal of 3D tracking is to recover pixel-wise 3D trajectories \\(\\mathcal{T} = (\\mathcal{T}_i^t)_{i=1,\\dots,N}\\) for each query. In order to account for static and dynamic parts of the scene, we decompose \\(\\mathcal{T}\\) into two components the ego camera motion \\(\\mathcal{T}_{\\mathrm{ego}}\\) and pixelwise object motion \\(\\mathcal{T}_{\\mathrm{object}}\\). Fig. 2 illustrates the pipeline of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.337, + 0.736, + 0.353 + ], + "angle": 0, + "content": "3.1. Ego Motion Component" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Ego motion, represented by camera trajectories \\(\\mathcal{T}_{\\mathrm{ego}}\\), is a major contributor to the 3D flow in camera coordinates. To compute the 3D tracks induced by ego motion, we need to estimate scale-aligned camera trajectories and video depth. Video Depth. Monocular depth estimation models, such as DepthAnything [87], typically follow an encoder-decoder design with a vision encoder like DINO [8] and a DPT-style decoder. Following VGGT [74], we extend the monocular encoder into a temporal encoder using an alternating-attention mechanism. This mechanism alternates between intra-frame self-attention and inter-frame attention over flattened video tokens, effectively balancing performance and efficiency. Furthermore, two learnable tokens, \\(\\mathbf{P}_{\\mathrm{tok}}\\) and \\(\\mathbf{S}_{\\mathrm{tok}}\\), are incorporated into the alternating-attention layers to capture high-level semantics for pose and scale regression. Camera Tracker. Similar to [73, 74], we adopt a differentiable pose head, \\(\\mathcal{H}\\), to decode pose, scale and shift directly:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.63, + 0.907, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} ^ {t}, a, b = \\mathcal {H} \\left(\\mathbf {P} _ {\\mathrm {t o k}}, \\mathbf {S} _ {\\mathrm {t o k}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.906, + 0.744 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^t\\in \\mathbb{R}^{T\\times 8}\\) is the camera encoding parameterized by a quaternion, a translation vector, and the normalized focal length concatenated together; The parameters \\(a\\) and \\(b\\) represent the scale and shift used to align the depth with the estimated camera poses. After that, we can easily obtain the 3D trajectories induced by ego-motion \\(\\mathcal{T}_{\\mathrm{ego}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.754, + 0.907, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\text {e g o}} = \\mathcal {W} (\\mathcal {P}, a * \\mathcal {D} _ {\\text {n o r m}} + b), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\mathrm{norm}}\\) is the raw results after DPT head with activation functions and \\(\\mathcal{W}\\) is the camera transformation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.762, + 0.835 + ], + "angle": 0, + "content": "3.2. Joint Motion Optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "After ego-motion initialization, we jointly estimate 2D trajectories \\(\\mathcal{T}^{2\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 2}\\) in UV space and their corresponding 3D trajectories \\(\\mathcal{T}^{3\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 3}\\) in the camera coordinate system using an iterative transformer module, referred" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.908, + 0.421 + ], + "angle": 0, + "content": "Figure 2. Pipeline Overview. Our method adopts a front-end and back-end architecture. The front-end estimates scale-aligned depth and camera poses from the input video, which are used to construct initial static 3D tracks. The back-end then iteratively refines both tracks and poses via joint motion optimization." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.447, + 0.481, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Figure 3. SyncFormer. The model takes previous estimates and their corresponding embeddings as input, and updates them iteratively. The 2D and 3D embeddings are processed in separate branches that interact via cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.726 + ], + "angle": 0, + "content": "to as SyncFormer. In parallel, the SyncFormer also dynamically estimates the visibility probability \\( p^{\\mathrm{vis}} \\) and dynamic probability \\( p^{\\mathrm{dyn}} \\) for each trajectory, enabling an efficient bundle adjustment process to refine the camera poses \\( \\mathcal{P} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.803 + ], + "angle": 0, + "content": "SyncFormer. As illustrated in Fig. 3, in every iteration, SyncFormer takes 2D embeddings, 3D embeddings, and camera poses as input, and updates the 2D trajectories \\(\\mathcal{T}^{2\\mathrm{d}}\\), 3D trajectories \\(\\mathcal{T}^{3\\mathrm{d}}\\), dynamic probabilities \\(p^{\\mathrm{dyn}}\\), and visibility scores \\(p^{\\mathrm{vis}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.812, + 0.483, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {k + 1} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k + 1} ^ {3 \\mathrm {d}}, p _ {k + 1} ^ {\\mathrm {d y n}}, p _ {k + 1} ^ {\\mathrm {v i s}} = f _ {\\text {s y n c}} \\left(\\mathcal {T} _ {k} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k} ^ {3 \\mathrm {d}}, p _ {k} ^ {\\mathrm {d y n}}, p _ {k} ^ {\\mathrm {v i s}}, \\mathcal {P} _ {k}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{sync}} \\) denotes the transformer-based update function. To better capture the distinct characteristics of 2D and 3D motion, the 2D and 3D trajectory updaters are modeled using separate attention layers. To reduce computational cost," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.568 + ], + "angle": 0, + "content": "correlation embeddings are first encoded into a compact set of 2D and 3D proxy tokens using cross-attention. Then, information is exchanged between the 2D and 3D branches via a cross-attention layer between the respective proxy tokens. This design decouples the mutual influence between 2D and 3D tracking, which are updated in two different domains: the UV space for 2D trajectories and the camera coordinate space for 3D trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.87 + ], + "angle": 0, + "content": "2D and 3D Embeddings. As the input for SyncFormer, the 2D and 3D embeddings encode the status of current estimations while recording the neighbourhood information for updating. For 2D embeddings, we keep the same to Cotracker3 [33]. The 3D embeddings \\(\\mathbf{E}^{3\\mathrm{d}} = (Corr_{3\\mathrm{D}}, \\mathbf{e}^{\\mathrm{time}}, \\mathbf{e}^{\\mathrm{Gpos}}, p^{\\mathrm{dyn}}, p^{\\mathrm{vis}})\\) contains the 3D correlation features \\(Corr_{3\\mathrm{D}}\\), the time embeddings \\(\\mathbf{e}^{\\mathrm{time}}\\), global position embeddings \\(\\mathbf{e}^{\\mathrm{Gpos}}\\) and dynamic-visibility scores. The 3D correlation \\(Corr_{3\\mathrm{D}}\\) is the main features for the 3D embedding. Different to 2D correlations, we expect 3D tracking branch updates the 3D position in the camera coordinate space. Therefore, we calculate the 3D correlations on the normalized point maps from front-end instead of depth map. Similar to 2D correlations, we construct multi-resolution point maps and compute the relative translations between each point and its neighbors within radius of 3. We then apply harmonic positional encoding to project these relative translations into high-dimensional feature representations, which are combined with semantic features to compute the final 3D correlations:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.878, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\nC o r r _ {3 \\mathrm {D}} = \\left[ K \\left(\\frac {\\mathbf {x}}{k s} + \\delta , \\frac {\\mathbf {y}}{k s} + \\delta\\right): \\delta \\in \\mathbb {Z}, \\| \\delta \\| _ {\\infty} \\leq \\Delta \\right], \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.485, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Iterative Process of SyncFormer. The left subfigure shows the convergence curve of reprojection error, illustrating rapid reprojection error reduction. The right subfigures visualize the progressive alignment between 2D tracking results (in green) and projections of 3D (world) tracking transformed by camera poses (in red)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.483, + 0.472 + ], + "angle": 0, + "content": "with \\(\\Delta\\) set to 3 and at multiple scales \\(s = 1,2,3,4\\), \\(K\\) denotes the operator for encoding relative translation features. Besides, in order to encode the camera pose to assist the tracking estimations, we propose \\(\\mathbf{e}^{\\mathrm{Gpos}}\\) which transforms the query points \\(\\mathcal{Q}_i\\) into each frames with the current camera poses, denoted as anchor points \\(\\mathcal{Q}_i^{\\mathrm{anc}}\\). \\(\\mathbf{e}^{\\mathrm{Gpos}}\\) is calculated from the relative translation between the current position and \\(\\mathbf{e}^{\\mathrm{Gpos}}\\), and they are then projected into high dimension with same tricks as above." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.474, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Camera Motion Optimization. After each iteration, the updates of \\(\\mathcal{T}_{k + 1}^{2\\mathrm{d}},\\mathcal{T}_{k + 1}^{3\\mathrm{d}},p_{k + 1}^{\\mathrm{dyn}},p_{k + 1}^{\\mathrm{vis}}\\) naturally form self-consistency constraints with the camera poses. These intrinsic constraints can be seamlessly incorporated into a bundle adjustment formulation. Specifically, we apply a weighted Procrustes analysis to register \\(\\mathcal{T}_{k + 1}^{3\\mathrm{d}}\\) into the world coordinate frame, where the alignment weights are given by the dynamic scores. This coarse alignment process is differentiable and provides effective supervision for learning the dynamic probability. Then we fuse the aligned \\(\\mathcal{T}_{k + 1}^{3\\mathrm{d}}\\) to world points \\(\\mathbf{P}_{k + 1}^{\\mathrm{world}}\\in \\mathbb{R}^{N\\times 3}\\), where the dynamic points are filtered with the estimated dynamic scores. After this, we apply a direct bundle adjustment to optimize the camera poses with paired \\(\\mathbf{P}_{k + 1}^{\\mathrm{world}},\\mathcal{T}_{k + 1}^{2\\mathrm{d}}\\) and \\(p_{k + 1}^{\\mathrm{vis}}\\). In the next iteration, the updated camera poses \\(\\mathcal{P}_{k + 1}\\) influence the subsequent loop through global position embedding. It is worth noting that the 2D and 3D trajectories are not updated via bundle adjustment; instead, the entire system is primarily driven by SyncFormer. As shown in Fig. 4, the reprojection errors decrease rapidly over iterations, while the 2D trajectories gradually become consistent with the 3D projections transformed by camera motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.196, + 0.835 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Training Datasets. Our model is trained on a large collection of 17 datasets, each providing different forms of supervision. The training data can be broadly categorized into three types: (1) Posed RGB-D with tracking annotations," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.393 + ], + "angle": 0, + "content": "(2) Posed RGB-D, (3) Pose-only or unlabeled data. For (1) containing Kubric [21], PointOdyssey [96] and Dynamic Replica [31], we add the full supervisions of camera poses, video depth, dynamic segmentation and 2D, 3D tracking. Within category (2), including VKITTI [7], TartanAir [80], Spring [53], DL3DV [44], BEDLAM [6], MVS-Synth [28], CO3Dv2 [61], COP3D [65], WildRGBD [83], ScanNet++ [12] and OmniObject3D [82], we leverage depth supervision to improve the video depth estimator. In addition, joint training losses are applied to ensure that the 3D tracking remains consistent with the estimated depth. As for category (3), it includes HOI4D [46], Ego4D [20], and Stereo4D [30]. For this type of data, we apply camera loss and joint training losses here, and follow [9] by using a monocular depth model [78] as a teacher to preserve relative depth accuracy. It is worth noting that Stereo4D [30] provides only very sparse depth on valid 3D tracks. Therefore, we choose not to use its annotations and instead treat it as pose-only data, given that it is sourced from Internet videos with rich scene diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Implementation details. Our training recipe consists of several stages. Stage 1. We first train the front end model to jointly estimate video depth and camera poses on category (1) and (2) datasets which sums up to 14 datasets. Training is conducted with mixed BF16 precision, while the DPT and camera tracker modules use full precision to ensure stable optimization. We use the AdamW optimizer with a learning rate of \\(5 \\times 10^{-5}\\) and apply gradient clipping with a threshold of 0.1. Stage 1 training is performed for 200k iterations on 64 H20 GPUs. We shuffle the video length from 1-24 during the training. Stage 2. At this stage, we initialize SyncFormer using the category (1) datasets, where ground-truth depth is provided and camera poses are initialized as identity matrices. We load the Cotracker3 [32] checkpoint to initialize the 2D tracking branch. This stage training takes 3 days of 100k iterations on 8 H20 GPUs. The video length is shuffle from 12-48 during the training. Stage 3. Finally, we fixed the alternation-attention layers in front end and train the whole pipeline in all datasets for 20 hours to converge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.646, + 0.725 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.826 + ], + "angle": 0, + "content": "We evaluate our model across all sub-tasks, including 3D tracking (Sec. 4.1), and dynamic 3D reconstruction (Sec. 4.2). In addition, we conduct comprehensive ablation studies (Sec. 4.3) to analyze the impact of key design choices and to demonstrate the effectiveness of unified modeling and scaling up training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.693, + 0.85 + ], + "angle": 0, + "content": "4.1. 3D Point Tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Dataset. We evaluate our model and compare it with existing baselines on TAPVid-3D [37], a comprehensive benchmark spanning diverse scenarios including Driving, Ego" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6730" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. 3D tracking results on the TAPVid-3D benchmark. We report the 3D Average Jaccard (AJ), Average 3D Position Accuracy \\((\\mathrm{APD}_{3D})\\), and Occlusion Accuracy (OA) across the Aria, DriveTrack, and PStudio subsets. \\(offl^{+}\\) and \\(offl^{-}\\) denote our offline model with/without considering the camera motion, respectively. COL, Univ2, and Mega are abbreviations for COLMAP, UnidepthV2, and MegaSAM. The best and the second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.146, + 0.903, + 0.388 + ], + "angle": 0, + "content": "
MethodsTypeDepth / Cam PoseAriaDriveTrackPStudioAverage
AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑
BootsTAPIRType ICOL9.114.578.611.818.683.86.911.681.89.314.981.4
TAPTRType IUniv215.724.287.812.419.184.87.313.584.311.818.985.6
LocoTrackType IUniv215.124.083.513.019.882.87.213.180.111.819.082.3
CoTracker3Type ICOL8.012.378.611.719.181.78.113.577.29.315.079.1
Univ215.824.488.913.519.987.19.213.884.212.819.486.7
Mega20.430.189.814.120.388.517.427.285.017.325.987.8
SpatialTrackerType IIUniv213.620.990.58.314.582.88.015.075.810.016.883.0
Mega15.923.890.17.713.585.215.325.278.113.020.884.5
SceneTrackerType IIUniv2-23.1-6.8--12.7---14.2-
DELTAType IIUniv216.624.486.814.622.585.88.215.076.413.120.683.0
Ours-offl-Type IIUniv218.626.390.816.424.390.218.127.686.717.726.089.2
Mega22.332.293.715.823.090.018.228.687.318.727.990.5
TAPIP3DType IIIMega23.532.891.214.921.882.618.127.785.518.827.486.4
Ours-offl+Type IIIMega24.735.293.916.023.490.118.628.786.119.829.190.0
Full-ours24.634.793.617.626.190.821.932.187.421.231.090.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Table 2. Video depth evaluation. Type I represents the methods specialized in video depth estimation, while Type II are neural reconstruction model, jointly recovering the geometry and camera motion from the video. Type III denotes the SoTA optimization-based method. The best and the second best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.442, + 0.897, + 0.595 + ], + "angle": 0, + "content": "
Method / MetricsAverageKITTI [19]TUM Dyn [66]Bonn [59]Sintel [52]
AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)
DepthCrafter [27]0.1430.8570.1110.8850.1230.8730.0660.9790.2720.693
VDA [9]0.1540.8820.0800.9510.1180.9200.0490.9820.3700.674
DUSt3R [79]0.2400.7660.1240.8490.1870.7920.1740.8350.4750.591
MonST3R [91]0.1710.8020.0830.9340.1970.7260.0610.9540.3430.594
CUT3R [77]0.1860.8140.1040.8990.1080.8470.0680.9500.4660.560
VGGT [74]0.1040.8810.0510.9660.0680.9390.0560.9630.2420.659
MegaSAM [74]0.0930.8940.0690.9160.0810.9350.0370.9770.1850.746
Ours0.0810.9100.0520.9730.0450.9760.0280.9880.1990.703
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.62, + 0.482, + 0.742 + ], + "angle": 0, + "content": "centric, and Studio. This benchmark consists of 4,569 evaluation videos, where the video length varies from 25 to 300 frames, and three 3D point tracking metrics are reported. Specifically, Occlusion Accuracy (OA) measures the precision of occlusion predictions; \\(\\mathrm{APD}_{3D}\\) denotes the average percentage of estimated errors within multiple threshold scales \\(\\delta\\); and Average Jaccard (AJ) quantifies the accuracy of both position and occlusion estimation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Baselines and Settings. The existing baselines can be broadly categorized into three types. Type I: 2D trackers followed by depth lifting. We report the current state-of-the-art 2D tracking models, CoTracker3 [32] and BootsTAPIR [16], as representatives of this category. Type II: 3D trackers in camera space. We include SpatialTrackerV1 [86] and DELTA [54], evaluated with two state-of-the-art depth estimators, i.e., UnidepthV2 [60] and MegaSAM [42]. Notably, MegaSAM is an optimization-based SLAM system for estimating consistent video depth," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.786 + ], + "angle": 0, + "content": "which generally performs better than feedforward models such as UnidepthV2. To ensure fair comparisons under the same camera-space 3D tracking setup, our 3D tracker module is also evaluated in combination with these depth estimators. Type III: 3D trackers in world space. We compare our fully end-to-end model with the recently released TAPIP3D [90], which requires consistent video depth and camera poses as input, provided by MegaSAM. For a thorough and fair comparison, we also report results where our depth and pose estimations are replaced by those from MegaSAM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Quantitative Results and Analysis. In the TAPVid-3D benchmark, our method achieves the best results across various settings, consistently outperforming all baselines under comparable conditions as in Tab. 1. From those detailed comparisons, we can draw several conclusions from it." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "- Better depth estimation is crucial for 3D tracking: For Type I methods, the 3D tracking performance largely re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "6731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.091, + 0.483, + 0.318 + ], + "angle": 0, + "content": "flects the accuracy of the underlying 2D tracks and the associated depth predictions. Intriguingly, CoTracker3 + MegaSAM achieves substantial improvements over CoTracker3 + UniDepthV2, with 17.3 vs. 12.8 in AJ, 25.9 vs. 19.4 in \\(APD_{3D}\\), and 87.8 vs. 86.7 in OA. This represents the best performance among Type I methods. However, our method Ours-offl\\(^{-}\\), which uses the weaker depth estimator UniDepthV2, still outperforms CoTracker3 with MegaSAM, achieving 18.7 vs. 17.3 in AJ and 27.4 vs. 25.9 in \\(APD_{3D}\\). This is because Type I methods rely purely on back-projection, making them more sensitive to depth inconsistencies. In contrast, our method directly predicts 3D trajectories using a spatial-temporal transformer, which provides inherent temporal smoothness and robustness to noisy inputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.319, + 0.483, + 0.56 + ], + "angle": 0, + "content": "- Camera motion decomposition improves 3D tracking: As shown in the comparison between Ours-offl \\(^+\\) MegaSAM and Ours-offl \\(^+\\) MegaSAM, incorporating camera pose estimation clearly improves 3D tracking accuracy. For example, on the Aria subset, our method achieves 24.7 vs. 22.3 in AJ and 35.2 vs. 32.2 in \\(APD_{3D}\\). This improvement is largely due to the design of the TAPVid-3D benchmark, which includes a substantial number of background points for evaluation in Aria. These background points primarily reflect camera motion and are particularly challenging for methods that only track in camera space, especially due to frequent out-of-view scenarios. In DriveTrack and Pstudio, the improvements are less, as DriveTrack includes only dynamic points on moving vehicles, while Pstudio contains static scenes with no camera motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.571, + 0.345, + 0.587 + ], + "angle": 0, + "content": "4.2. Dynamic 3D Reconstruction" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.594, + 0.305, + 0.609 + ], + "angle": 0, + "content": "4.2.1. Video Depth Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Datasets. To illustrate the effectiveness and generalization ability of our method, we evaluate our video depth estimation method across four mainstream datasets, i.e., KITTI [19], Sintel [52], Bonn [59], TUM Dynamics [66]. These datasets encompass both indoor and outdoor scenes, featuring video sequences ranging from 50 to 110 frames, providing a comprehensive benchmark for assessing depth estimation consistency across varied environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Baselines. We compare our approach against three categories of methods. Type I includes SoTA video depth methods, DepthCrafter [27] and Video Depth Anything [87]. Type II consists of SoTA deep reconstruction approaches, i.e. DUST3R [79], MonST3R [91], CUT3R [77] and VGGT [74] Type III denotes the SoTA dynamic Structurefrom-Motion (SfM) system, MegaSAM [42]. It leverages an optimization-based paradigm to jointly estimate consistent video depths and camera poses, with constraints enforced by optical flow and monocular depth priors." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics. We evaluate the performance of video depth esti" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Table 3. Evaluation on Camera Pose Estimation on TUMdynamic [66], Lightspeed [12], and Sintel [52]. All values are absolute trajectory error (ATE), relative pose error (RPE) for translation and rotation. The best and the second best are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.157, + 0.905, + 0.314 + ], + "angle": 0, + "content": "
MethodTUM-dynamic (ATE / RPEt / RER)Lightspeed (ATE / RPEt / RER)Sintel (ATE / RPEt / RER)
Particle-SfM [95]-0.185 / 0.075 / 2.9900.129 / 0.031 / 0.535
Robust-CVD [36]0.153 / 0.026 / 3.528-/-/-0.360 / 0.154 / 3.443
CasualSAM [94]0.071 / 0.010 / 1.712-/-/-0.141 / 0.035 / 0.615
DUST3R [79]0.140 / 0.106 / 3.2860.412 / 0.177 / 20.1000.290 / 0.132 / 7.869
CUT3R [77]0.046 / 0.015 / 0.4730.274 / 0.067 / 1.5610.213 / 0.066 / 0.621
MonST3R [91]0.098 / 0.019 / 0.9350.149 / 0.046 / 1.2100.078 / 0.038 / 0.490
VGGT [74]0.021 / 0.013 / 0.3270.226 / 0.086 / 1.7290.082 / 0.043 / 1.253
MegaSAM [42]0.013 / 0.011 / 0.3400.105 / 0.040 / 0.9960.023 / 0.008 / 0.060
Ours-Front0.038 / 0.022 / 0.4800.203 / 0.079 / 1.6890.075 / 0.045 / 0.805
Ours0.012 / 0.010 / 0.3050.134 / 0.039 / 1.3400.054 / 0.027 / 0.288
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.404 + ], + "angle": 0, + "content": "mation using geometric accuracy metrics. To ensure fair comparisons with previous works [9, 77, 91], we follow their approach by aligning the predicted video depth to the ground truth using a shared scale and shift, and then compute the Absolute Relative Error (AbsRel) and \\(\\delta_{1}\\) metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Tab. 2, our method clearly outperforms all existing approaches. Specifically, we surpass our baseline in Type II, VGGT [74], by a significant margin. On average, our method achieves an AbsRel of 0.081 compared to 0.104 \\((-22.1\\%)\\), and a \\(\\delta_{1.25}\\) of 0.910 compared to 0.881 \\((+3.3\\%)\\). Besides, compared to MegaSAM (Type III), our method also maintain clear advantages with 0.081 v.s. 0.093 in AbsRel, and 0.910 v.s. 0.894 in \\(\\delta_{1.25}\\). It is important to note that MegaSAM [42] usually needs 5-10 min for a 100 frames video, while our method only takes 5-10 seconds for each which nearly achieves \\(50 \\times\\) faster." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.593, + 0.659, + 0.607 + ], + "angle": 0, + "content": "4.2.2. Camera Poses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.84 + ], + "angle": 0, + "content": "To evaluate the accuracy of camera motion estimation, we conduct the comparisons on Sintel [52], TUM Dynamics [66], and Lightspeed [62]. Sintel and Lightspeed contains numerous challenging dynamic scenes with large egomotion and object motions, and they are both the synthetic data. TUM Dynamics [66] is a real data, captured by well-calibrated RGBD sensors. We report Absolute Translation Error (ATE), Relative Rotation Error (RPE rot) and Relative Translation Error (RTE) after Sim(3) alignment with the ground truth, as in [77, 91]. Shown in Tab. 3, our method outperforms all regression-based methods, much better than our baseline, VGGT [74] and on par with MegaSAM [42]. Besides, the table illustrates the significance of back end optimization. After joint motion optimization, the pose estimations become nearly twice accurate than before." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.688, + 0.865 + ], + "angle": 0, + "content": "4.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Our ablation study investigates the impact of training data, joint training, and the SyncFormer design on 2D and 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6732" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.098, + 0.166, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.137, + 0.16, + 0.148 + ], + "angle": 0, + "content": "(a) Query" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.089, + 0.243, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.137, + 0.245, + 0.147 + ], + "angle": 0, + "content": "(b) w/o Joint" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.089, + 0.319, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.137, + 0.318, + 0.147 + ], + "angle": 0, + "content": "(c) w/ Joint" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.089, + 0.396, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.137, + 0.398, + 0.147 + ], + "angle": 0, + "content": "(d) w/o Joint" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.089, + 0.472, + 0.134 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.137, + 0.471, + 0.147 + ], + "angle": 0, + "content": "(e) w/ Joint" + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.16, + 0.411, + 0.175 + ], + "angle": 0, + "content": "Figure 5. The influence of Joint Training." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.193, + 0.483, + 0.236 + ], + "angle": 0, + "content": "Table 4. TAP-Vid DAVIS results. Cotracker3-3D is a fine-tuned model by adding simple 3D project layer. Mean represents the average of AJ and \\( {\\delta }_{\\text{avg }}^{\\text{vis }} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.246, + 0.481, + 0.37 + ], + "angle": 0, + "content": "
MethodLift 3DAJ ↑δvisavg ↑OA ↑Mean ↑
TAPTRX63.076.191.169.5
LocoTrackX62.975.387.269.1
CoTracker3X64.476.991.270.7
DELTA3D-Dec62.776.788.269.7
SpatialTrackerTri-plane61.176.389.568.7
CoTracker3-3D3D-Proj51.665.285.358.4
OursSyncFormer64.977.591.071.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.445 + ], + "angle": 0, + "content": "point tracking. For depth estimation, we analyze different loss functions and training strategies to improve the generalizability of video depth prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.657 + ], + "angle": 0, + "content": "3D Point Tracking. To illustrate the gains brought by scaling up the 3D point tracking to a wider range of data, we naively train a base model (Base@K in Table 5), and an advanced one, Base@V,K,P, which is trained on VKITTI, Kubric, and PointOdyssey. As the table shows, Base@V,K,P outperforms Base@K by a clear margin. Meanwhile, our final version is clearly better than these two. Besides, different from Base@K, Base@V,K,P is jointly trained on VKITTI, which brings very significant improvements on a similar type of real data: DriveTrack shows 14.7 vs. 7.4 in AJ and 21.9 vs. 13.3 in \\(\\mathrm{APD}_{3D}\\). Fig. 5 qualitatively demonstrates the meaning of joint training, i.e., joint training contributes to minimizing the 3D tracking drifts when the model is trained on new patterns of data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2D Point Tracking. We report a naive alternative to SyncFormer by modifying the final output layer of CoTracker3 [33] to output 2D and 3D tracking. The 2D and 3D embeddings are directly concatenated and projected back to the original dimension using an inserted, learnable linear layer. We name this naive baseline as Cotacker3-3D. As shown in Tab. 4, a simple adaptation of 3D lifting leads to a significant drop in 2D tracking accuracy—AJ drops from 64.4 to 51.6, \\(\\delta\\) from 76.9 to 65.2, and OA from 91.2 to 85.3. This degradation occurs because the 2D and 3D correlation signals become entangled, disrupting the model's ability to focus on reliable features. Moreover, the update dynamics in 2D and 3D differ substantially, as they occur in separate spaces—UVD space for 2D and camera space for 3D. We also report comparisons with other 3D lifting techniques. These results validate our SyncFormer design, which effec" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "tively lifts tracking into 3D while preserving—or even improving—accuracy in 2D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Depth Estimation. We ablate different training datasets and their coverage to assess their impact on model performance. Additionally, we study the influence of different loss functions applied to real data, i.e., the pearson loss and full depth losses. As shown in Tab. 2, if we applied the full losses for depth will cause the obvious performance drop in the synthetic data, i.e. Sintel. On the contrary, if we only used the synthetic data to train the model, we found the performance in the real dataset, especially KITTI will heavily drop. Therefore, we leverage the strategy of using different losses for real and synthetic data to offset the influence of domain gap and different errors distribution. As shown in Tab. 2, Ours-Synthetic is our method trained only on the synthetic datasets. Ours-Real-Full is our method using the full depth loss in the real data, where the model was influenced by the noised distributions in the real dataset. The noised of real data has the negative influence on the model's influence on the synthetic data. Meanwhile, the different errors pattern also worsen the model's zero-shot capabilities in the real data." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.428, + 0.905, + 0.47 + ], + "angle": 0, + "content": "Table 5. Ablation of Training Data and Joint Training. K, V, P represents Kubric [21], VKITTI [7] and PointOdyssey [96] in training. Inference depths are provided by Unidepthv2." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.472, + 0.904, + 0.521 + ], + "angle": 0, + "content": "
MethodJoint TrainingAriaDriveTrackPStudio
AJAPD3DAJAPD3DAJAPD3D
Base@KNo16.024.47.413.312.620.3
Base@V, K, PYes15.724.114.721.917.227.4
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.781 + ], + "angle": 0, + "content": "This work introduces SpatialTrackerV2, a feedforward and scalable approach for 3D point tracking in monocular videos. Built upon a deep exploration of widely-used low- and mid-level representations of motion and scene geometry, our method unifies consistent scene geometry, camera motion, and pixel-wise 3D motion into a fully differentiable end-to-end pipeline. SpatialTrackerV2 accurately reconstructs 3D trajectories from monocular videos, achieving strong quantitative results on public benchmarks and demonstrating robust performance on casually captured Internet videos. We believe SpatialTrackerV2 establishes a solid foundation for real-world motion understanding and brings us a step closer to spatial intelligence by exploring large-scale vision data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.799, + 0.666, + 0.816 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work was partially supported by Ant Group Research Intern Program, Zhejiang Provincial Natural Science Foundation of China (No. LR25F020003) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Abhishek Badki, Hang Su, Bowen Wen, and Orazio Gallo. L4p: Low-level 4d vision perception unified. arXiv preprint arXiv:2502.13078, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[4] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. CoRR, abs/2501.02690, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[5] Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3.1 - A model zoo for robust monocular relative depth estimation. CoRR, abs/2307.14460, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.444 + ], + "angle": 0, + "content": "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[9] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. CoRR, abs/2501.12375, 2025. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[10] Zequn Chen, Jiezhi Yang, and Heng Yang. Pref3r: Pose-free feed-forward 3d gaussian splatting from variable-length image sequence. CoRR, abs/2411.16877, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In European Conference on Computer Vision, pages 306–325. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[14] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[16] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, pages 3257-3274, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[22] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, Wenping Wang, and Yuan Liu. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. CoRR, abs/2501.03847, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[23] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[25] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[26] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE TPAMI, 46(12):10579-10596, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. CoRR, abs/2409.02095, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "stereopsis. In CVPR, pages 2821-2830. Computer Vision Foundation / IEEE Computer Society, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[29] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy J. Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. CoRR, abs/2412.06016, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[30] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[32] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[34] Yoni Kasten, Wuyue Lu, and Haggai Maron. Fast encoder-based 3d from casual videos via point track processing. arXiv preprint arXiv:2404.07097, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[35] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In Proceedings of the IEEE international conference on computer vision, pages 1521–1529, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1611-1621. Computer Vision Foundation / IEEE, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[37] Skanda Koppula, Ignacio Rocco, Yi Yang, Joseph Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. In NeurIPS, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[38] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[39] Hongdong Li and Richard Hartley. Five-point motion estimation made easy. In 18th International Conference on Pattern Recognition (ICPR'06), pages 630-633. IEEE, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Xiaodi Li, Zongxin Yang, Ruijie Quan, and Yi Yang. Drip: Unleashing diffusion priors for joint foreground and alpha prediction in image matting. Advances in Neural Information Processing Systems, 37:79868-79888, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[42] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In CVPR, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[44] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[45] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE transactions on pattern analysis and machine intelligence, 33(5):978-994, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[46] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In CVPR, pages 20981-20990. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[48] Haozhe Lou, Yurong Liu, Yike Pan, Yiran Geng, Jianteng Chen, Wenlong Ma, Chenglong Li, Lin Wang, Hengzhen Feng, Lu Shi, Liyi Luo, and Yongliang Shi. Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation. CoRR, abs/2408.14873, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[49] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[50] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[51] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[52] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[53] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[54] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. arXiv preprint arXiv:2410.24211, 2024. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[55] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[56] Dantong Niu, Yuvan Sharma, Haoru Xue, Giscard Biamby, Junyi Zhang, Ziteng Ji, Trevor Darrell, and Roei Herzig. Pretraining auto-regressive robotic models with 4d representations. arXiv preprint arXiv:2502.13142, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[57] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8089-8099, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[58] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[59] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[60] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segù, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116. IEEE, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[61] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[62] Chris Rockwell, Joseph Tung, Tsung-Yi Lin, Ming-Yu Liu, David F. Fouhey, and Chen-Hsuan Lin. Dynamic camera poses and where to find them. In CVPR, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[63] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[64] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[65] Samarth Sinha, Roman Shapovalov, Jeremy Reizenstein, Ignacio Rocco, Natalia Neverova, Andrea Vedaldi, and David Novotny. Common pets in 3d: Dynamic new-view synthesis of real-life deformable categories. In CVPR, pages 4881-4891. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[66] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS, pages 573-580. IEEE, 2012. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[67] Tutian Tang, Minghao Liu, Wenqiang Xu, and Cewu Lu. Kalib: Markerless hand-eye calibration with keypoint tracking. CoRR, abs/2408.10562, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[69] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a mod-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "ern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21-22, 1999 Proceedings, pages 298-372. Springer, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[70] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. CoRR, abs/2412.13389, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[71] Bo Wang, Jian Li, Yang Yu, Li Liu, Zhenping Sun, and Dewen Hu. Scenetracker: Long-term scene flow estimation network. arXiv preprint arXiv:2403.19924, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.314 + ], + "angle": 0, + "content": "[72] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.384 + ], + "angle": 0, + "content": "[73] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.454 + ], + "angle": 0, + "content": "[74] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[75] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[76] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[77] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[78] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[80] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[81] Zhouxia Wang, Yushi Lan, Shangchen Zhou, and Chen Change Loy. Objctrl-2.5 d: Training-free object control with camera poses. arXiv preprint arXiv:2412.07721, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6736" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[82] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR, pages 803-814. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[83] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBD objects in the wild: Scaling real-world 3d object learning from RGB-D videos. In CVPR, pages 22378-22389. IEEE, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[84] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.349 + ], + "angle": 0, + "content": "[85] Yuxi Xiao, Nan Xue, Tianfu Wu, and Gui-Song Xia. Level-S\\(^2\\)fM: Structure From Motion on Neural Level Set of Implicit Surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17205-17214, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[86] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.424, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[87] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[88] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.609 + ], + "angle": 0, + "content": "[89] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3906-3915. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.613, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[90] Bowei Zhang, Lei Ke, Adam W Harley, and Katerina Fragkiadaki. Tapip3d: Tracking any point in persistent 3d geometry. arXiv preprint arXiv:2504.14717, 2025. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[91] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[92] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[93] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[94] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[95] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In *ECCV*. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[96] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6737" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_origin.pdf b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05590699e326bbaa0ed045d49c180f420af41a31 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/925f4fd6-2e60-44c7-a108-57075a837941_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb44f0a74f1ce40589514f2288d4207d94ff4db99bbca585c2424ca014ae818f +size 13577183 diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/full.md b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f658c0b35fff67df7d63f64709ea6b76d48d0f2 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/full.md @@ -0,0 +1,334 @@ +# SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion + +Yuxi Xiao $^{1*}$ Jianyuan Wang $^{2}$ Nan Xue $^{3}$ Nikita Karaev $^{2,4}$ Yuri Makarov $^{4}$ Bingyi Kang $^{5}$ Xing Zhu $^{3}$ Hujun Bao $^{1}$ Yujun Shen $^{3}$ Xiaowei Zhou $^{1\dagger}$ + +$^{1}$ Zhejiang University $^{2}$ Oxford $^{3}$ Ant Group $^{4}$ Pixelwise AI $^{5}$ Bytedance Seed + +![](images/a96c785c085d36ec1b1d6856714fabcf7afe3e82dad86dabac1fdf80f452a789.jpg) + +![](images/dd13107edc76019a3015327cc619496bc75ec554b56371a447106172748ce6f4.jpg) +Figure 1. SpatialTrackerV2 produces consistent 3D scene geometry, camera poses, and 3D point trajectories all at once from monocular videos of arbitrary scenarios, e.g., robotic manipulation, first-person egocentric views, and dynamic sports (drifting and skating) shown in this figure. Try our online demo at https://huggingface.co/spaces/Yuxihenry/SpatialTrackerV2. + +![](images/1613cdebd69c7240e2eb167eac480000ed1ad30259aaf5a6003d7a18d3d99f08.jpg) + +![](images/b08c78df9be708b49dceb74a24a922cc3794f52032ad900d5839ca838149d598.jpg) + +# Abstract + +We present SpatialTrackerV2, a feed-forward 3D point tracking method for monocular videos. Going beyond modular pipelines built on off-the-shelf components for 3D tracking, our approach unifies the intrinsic connections between point tracking, monocular depth, and camera pose estimation into a high-performing and feedforward 3D point tracker. It decomposes world-space 3D motion + +into scene geometry, camera ego-motion, and pixel-wise object motion, with a fully differentiable and end-to-end architecture, allowing scalable training across a wide range of datasets, including synthetic sequences, posed RGB-D videos, and unlabeled in-the-wild footage. By learning geometry and motion jointly from such heterogeneous data, SpatialTrackerV2 outperforms existing 3D tracking methods by $30\%$ , and matches the accuracy of leading dynamic 3D reconstruction approaches while running $50\times$ faster. + +# 1. Introduction + +3D point tracking aims to recover long-term 3D trajectories of arbitrary points from monocular videos. As a universal dynamic scene representation, it has recently shown strong potential in diverse applications, including robotics [48, 56, 67], video generation [22, 29, 57, 81], and 3D/4D reconstruction [10, 34, 38, 76]. Compared to parametric motion models (e.g., SMPL [47], MANO [63], skeletons, or 3D bounding boxes), it offers greater flexibility and generalization over various real-world scenes, as shown in Fig. 1. + +Existing solutions [1, 4, 54, 75, 86] of 3D point tracking extensively explored the well-developed low/mid-level vision models, such as optical flow [45, 68] and monocular depth estimation [27, 87], and took benefits from 2D point tracking models [15, 23, 33]. Among them, optimization-based methods [42, 75, 95] distill the optical flow, monocular depth models and camera motions for each given monocular video with promising results obtained, while being computationally expensive due to their per-scene optimization designs. SpatialTracker [86] moved forward to efficient 3D point tracking with a feed-forward model, and the more recent works [1, 4, 54] explored different architecture designs and rendering constraints to achieve higher-quality 3D tracking. Nevertheless, the feed-forward solutions are limited to training data scalability issues due to the need for ground-truth 3D tracks as supervision, which downgrades the tracking quality in real-world casual captures. Moreover, overlooking the inherent interplay between camera motion, object motion, and scene geometry results in error entanglement and accumulation across modules. + +These limitations motivate our core insights: (1) The reliance on ground-truth 3D trajectories constrains the scalability of existing feed-forward models, highlighting the need for designs that can generalize across diverse and weakly-supervised data sources. (2) The absence of joint reasoning over scene geometry, camera motion, and object motion leads to compounded errors and degraded performance, underscoring the importance of disentangling and explicitly modeling these motion components. To address these challenges, we decompose 3D point tracking into three distinct components: video depth, ego (camera) motion, and object motion, and integrate them within a fully differentiable pipeline that supports scalable joint training across heterogeneous data. + +In our SpatialTrackerV2, a front-end and back-end architecture is proposed. The front end is a video depth estimator and camera pose initializer, adapted from typical monocular depth prediction frameworks [88] with attention-based temporal information encoding [74]. The predicted video depths and camera poses are then fused through a scale-shift estimation module, which ensures consistency between the depth and motion predictions. The back end + +consists of a proposed Joint Motion Optimization Module, which takes the video depth and coarse camera trajectories as input and iteratively estimates 2D and 3D trajectories, along with trajectory-wise dynamics and visibility scores. This enables an efficient bundle adjustment process for optimizing camera poses in the loop. At its core lies a novel SyncFormer, which separately models the 2D and 3D correlations in two branches, connected by multiple cross-attention layers. This design mitigates mutual interference between 2D and 3D embeddings and allows the model to update representations in two distinct spaces, namely the image (UV) space and the camera coordinate space. Furthermore, benefiting from this dual-branch design, bundle adjustment can be effectively applied to jointly optimize camera poses as well as the 2D and 3D trajectories. + +This unified and differentiable pipeline makes large-scale training on diverse datasets possible. For RGB-D datasets with camera poses, we jointly train 3D tracking using consistency constraints from ground-truth depth and camera poses for static points, while dynamic points seamlessly contribute to the optimization. For video datasets that provide only camera pose annotations and lack depth information, we leverage consistency among camera poses, and 2D and 3D point tracking to drive the model's optimization. Relying on this framework, we successfully scale up training of the entire pipeline across 17 datasets. + +Evaluations on the TAPVid-3D benchmark [37] show that our method sets a new state-of-the-art in 3D point tracking, achieving 21.2 AJ and $31.0\mathrm{APD}_{3D}$ , surpassing DELTA [54] with relative improvements of $61.8\%$ and $50.5\%$ , respectively. Additionally, extensive experiments on dynamic reconstruction show our superior results on consistent video depth and camera poses estimation. Specifically, SpatialTrackerV2 beats the best dynamic reconstruction method, MegaSAM [42], on most video depth datasets and achieves comparable results on various camera pose benchmarks, while its inference speed is $50\times$ faster. + +# 2. Related work + +This section covers relevant literature on 3D tracking, depth, and camera pose estimation. + +# 2.1. Point tracking + +PIPs [23] revisited the 2D point tracking task first introduced in [64] and proposed a deep learning approach to solve it. TAP-Vid [14] redefined the problem and introduced both a benchmark and a simple architecture, TAP-Net. Subsequently, the performance was improved in TAPIR [15] by combining the global matching capabilities of TAP-Net with the local refinement offered by PIPs. CoTracker [33] pioneered tracking through occlusions using a transformer architecture combined with joint attention, followed by TAPTR [40] and LocoTrack [11], which im + +proved efficiency and introduced 4D correlation volumes. Recently, BootsTAPIR [16] and CoTracker3 [32] explored the use of unlabeled data to achieve better performance. + +While 2D point tracking has been extensively studied, 3D point tracking remains a relatively new field. The first method to demonstrate 3D point tracking capabilities was the test-time optimization-based OmniMotion [75]. Later, SpatialTracker [86] introduced the first feed-forward 3D point tracker by combining a 2D point tracker [33] with depth priors from a monocular depth estimator [3]. SceneTracker [71] proposed a new architecture for 3D tracking with depth priors, while DELTA [54] improved efficiency and achieved dense 3D tracking. Recently, TAPIP3D [90] improved 3D tracking robustness by lifting image features into the world coordinate space and performing tracking there. All of these 3D tracking models were trained on small synthetic datasets. + +Unlike these methods, we present a scalable 3D tracking framework trained on a collection of both real and synthetic datasets, while also explicitly modeling camera motion to improve performance on egocentric videos. + +# 2.2. Depth estimation + +Early methods such as Eigen et al. [17] introduced single-view depth estimation using CNNs, but were limited by dataset scale and poor generalization [18, 89]. MiDaS [5] improved this by mixing datasets for broader coverage, and ZoeDepth [3] adapted it for metric depth, though ambiguity from missing camera intrinsics remained. Later, Metric3D [26] and UniDepth [60] addressed this by jointly estimating intrinsics and normalized depth. + +With large-scale pretraining (e.g., diffusion [13, 41], DINO [8]), recent models like Marigold [70] and DepthAnything [87, 88] have significantly advanced zero-shot depth. Extensions to video [9, 27] have also emerged. + +In this work, we build on DepthAnythingV2 and extend it to video, aiming for not just consistent depth, but a unified framework that also predicts camera poses and tracks, with proper scale alignment—posing new challenges beyond static depth estimation. + +# 2.3. Camera estimation + +Traditional camera pose estimation methods [24, 58] rely on image-to-image point correspondences using keypoint detectors (e.g., SIFT [49, 50], SURF [2]) and matching techniques such as nearest neighbors, followed by geometric algorithms like the five-point and eight-point methods [24, 25, 39, 55, 85]. Bundle Adjustment [69] is also commonly employed to further enhance accuracy. Recently, direct regression approaches using neural networks [35, 43, 51, 84, 92] have emerged, aiming to overcome limitations in sparse-view scenarios or when correspondences are unreliable. Diffusion models, such as PoseDiffusion [72] and RayDif + +fusion [93], have also been explored, offering strong accuracy but suffering from high inference costs. In contrast, the camera head of VGGsfm [73] or VGGT [74] adopts an iterative refinement paradigm similar to RAFT [68], striking a balance between accuracy and inference cost, and enabling estimation of both extrinsic and intrinsic parameters. + +# 3. Method + +Given a video with $T$ frames $(\mathcal{I}^t)_{t=1}^T$ and $N$ query points $\mathcal{Q}_i = (x_i, y_i) \in \mathbb{R}^2$ , $i = 1, \dots, N$ , the goal of 3D tracking is to recover pixel-wise 3D trajectories $\mathcal{T} = (\mathcal{T}_i^t)_{i=1,\dots,N}$ for each query. In order to account for static and dynamic parts of the scene, we decompose $\mathcal{T}$ into two components the ego camera motion $\mathcal{T}_{\mathrm{ego}}$ and pixelwise object motion $\mathcal{T}_{\mathrm{object}}$ . Fig. 2 illustrates the pipeline of our method. + +# 3.1. Ego Motion Component + +Ego motion, represented by camera trajectories $\mathcal{T}_{\mathrm{ego}}$ , is a major contributor to the 3D flow in camera coordinates. To compute the 3D tracks induced by ego motion, we need to estimate scale-aligned camera trajectories and video depth. Video Depth. Monocular depth estimation models, such as DepthAnything [87], typically follow an encoder-decoder design with a vision encoder like DINO [8] and a DPT-style decoder. Following VGGT [74], we extend the monocular encoder into a temporal encoder using an alternating-attention mechanism. This mechanism alternates between intra-frame self-attention and inter-frame attention over flattened video tokens, effectively balancing performance and efficiency. Furthermore, two learnable tokens, $\mathbf{P}_{\mathrm{tok}}$ and $\mathbf{S}_{\mathrm{tok}}$ , are incorporated into the alternating-attention layers to capture high-level semantics for pose and scale regression. Camera Tracker. Similar to [73, 74], we adopt a differentiable pose head, $\mathcal{H}$ , to decode pose, scale and shift directly: + +$$ +\mathcal {P} ^ {t}, a, b = \mathcal {H} \left(\mathbf {P} _ {\mathrm {t o k}}, \mathbf {S} _ {\mathrm {t o k}}\right), \tag {1} +$$ + +where $\mathcal{P}^t\in \mathbb{R}^{T\times 8}$ is the camera encoding parameterized by a quaternion, a translation vector, and the normalized focal length concatenated together; The parameters $a$ and $b$ represent the scale and shift used to align the depth with the estimated camera poses. After that, we can easily obtain the 3D trajectories induced by ego-motion $\mathcal{T}_{\mathrm{ego}}$ : + +$$ +\mathcal {T} _ {\text {e g o}} = \mathcal {W} (\mathcal {P}, a * \mathcal {D} _ {\text {n o r m}} + b), \tag {2} +$$ + +where $\mathcal{D}_{\mathrm{norm}}$ is the raw results after DPT head with activation functions and $\mathcal{W}$ is the camera transformation. + +# 3.2. Joint Motion Optimization. + +After ego-motion initialization, we jointly estimate 2D trajectories $\mathcal{T}^{2\mathrm{d}} \in \mathbb{R}^{T \times N \times 2}$ in UV space and their corresponding 3D trajectories $\mathcal{T}^{3\mathrm{d}} \in \mathbb{R}^{T \times N \times 3}$ in the camera coordinate system using an iterative transformer module, referred + +![](images/f1bf9ec33a00fa908e2c02013dff54ac6ca106be1f5690bfc43cafb2fda30050.jpg) +Figure 2. Pipeline Overview. Our method adopts a front-end and back-end architecture. The front-end estimates scale-aligned depth and camera poses from the input video, which are used to construct initial static 3D tracks. The back-end then iteratively refines both tracks and poses via joint motion optimization. + +![](images/476246a2ee49336015912e61cf61e16111fb83964215b0c5bea46f381660547a.jpg) +Figure 3. SyncFormer. The model takes previous estimates and their corresponding embeddings as input, and updates them iteratively. The 2D and 3D embeddings are processed in separate branches that interact via cross-attention. + +to as SyncFormer. In parallel, the SyncFormer also dynamically estimates the visibility probability $p^{\mathrm{vis}}$ and dynamic probability $p^{\mathrm{dyn}}$ for each trajectory, enabling an efficient bundle adjustment process to refine the camera poses $\mathcal{P}$ . + +SyncFormer. As illustrated in Fig. 3, in every iteration, SyncFormer takes 2D embeddings, 3D embeddings, and camera poses as input, and updates the 2D trajectories $\mathcal{T}^{2\mathrm{d}}$ , 3D trajectories $\mathcal{T}^{3\mathrm{d}}$ , dynamic probabilities $p^{\mathrm{dyn}}$ , and visibility scores $p^{\mathrm{vis}}$ : + +$$ +\mathcal {T} _ {k + 1} ^ {2 \mathrm {d}}, \mathcal {T} _ {k + 1} ^ {3 \mathrm {d}}, p _ {k + 1} ^ {\mathrm {d y n}}, p _ {k + 1} ^ {\mathrm {v i s}} = f _ {\text {s y n c}} \left(\mathcal {T} _ {k} ^ {2 \mathrm {d}}, \mathcal {T} _ {k} ^ {3 \mathrm {d}}, p _ {k} ^ {\mathrm {d y n}}, p _ {k} ^ {\mathrm {v i s}}, \mathcal {P} _ {k}\right), \tag {3} +$$ + +where $f_{\mathrm{sync}}$ denotes the transformer-based update function. To better capture the distinct characteristics of 2D and 3D motion, the 2D and 3D trajectory updaters are modeled using separate attention layers. To reduce computational cost, + +correlation embeddings are first encoded into a compact set of 2D and 3D proxy tokens using cross-attention. Then, information is exchanged between the 2D and 3D branches via a cross-attention layer between the respective proxy tokens. This design decouples the mutual influence between 2D and 3D tracking, which are updated in two different domains: the UV space for 2D trajectories and the camera coordinate space for 3D trajectories. + +2D and 3D Embeddings. As the input for SyncFormer, the 2D and 3D embeddings encode the status of current estimations while recording the neighbourhood information for updating. For 2D embeddings, we keep the same to Cotracker3 [33]. The 3D embeddings $\mathbf{E}^{3\mathrm{d}} = (Corr_{3\mathrm{D}}, \mathbf{e}^{\mathrm{time}}, \mathbf{e}^{\mathrm{Gpos}}, p^{\mathrm{dyn}}, p^{\mathrm{vis}})$ contains the 3D correlation features $Corr_{3\mathrm{D}}$ , the time embeddings $\mathbf{e}^{\mathrm{time}}$ , global position embeddings $\mathbf{e}^{\mathrm{Gpos}}$ and dynamic-visibility scores. The 3D correlation $Corr_{3\mathrm{D}}$ is the main features for the 3D embedding. Different to 2D correlations, we expect 3D tracking branch updates the 3D position in the camera coordinate space. Therefore, we calculate the 3D correlations on the normalized point maps from front-end instead of depth map. Similar to 2D correlations, we construct multi-resolution point maps and compute the relative translations between each point and its neighbors within radius of 3. We then apply harmonic positional encoding to project these relative translations into high-dimensional feature representations, which are combined with semantic features to compute the final 3D correlations: + +$$ +C o r r _ {3 \mathrm {D}} = \left[ K \left(\frac {\mathbf {x}}{k s} + \delta , \frac {\mathbf {y}}{k s} + \delta\right): \delta \in \mathbb {Z}, \| \delta \| _ {\infty} \leq \Delta \right], \tag {4} +$$ + +![](images/d5dcde056f3c29d26c85be8c5f2d86ee291ffdd649590b0e25f164e1f6ca84aa.jpg) +Figure 4. Iterative Process of SyncFormer. The left subfigure shows the convergence curve of reprojection error, illustrating rapid reprojection error reduction. The right subfigures visualize the progressive alignment between 2D tracking results (in green) and projections of 3D (world) tracking transformed by camera poses (in red). + +with $\Delta$ set to 3 and at multiple scales $s = 1,2,3,4$ , $K$ denotes the operator for encoding relative translation features. Besides, in order to encode the camera pose to assist the tracking estimations, we propose $\mathbf{e}^{\mathrm{Gpos}}$ which transforms the query points $\mathcal{Q}_i$ into each frames with the current camera poses, denoted as anchor points $\mathcal{Q}_i^{\mathrm{anc}}$ . $\mathbf{e}^{\mathrm{Gpos}}$ is calculated from the relative translation between the current position and $\mathbf{e}^{\mathrm{Gpos}}$ , and they are then projected into high dimension with same tricks as above. + +Camera Motion Optimization. After each iteration, the updates of $\mathcal{T}_{k + 1}^{2\mathrm{d}},\mathcal{T}_{k + 1}^{3\mathrm{d}},p_{k + 1}^{\mathrm{dyn}},p_{k + 1}^{\mathrm{vis}}$ naturally form self-consistency constraints with the camera poses. These intrinsic constraints can be seamlessly incorporated into a bundle adjustment formulation. Specifically, we apply a weighted Procrustes analysis to register $\mathcal{T}_{k + 1}^{3\mathrm{d}}$ into the world coordinate frame, where the alignment weights are given by the dynamic scores. This coarse alignment process is differentiable and provides effective supervision for learning the dynamic probability. Then we fuse the aligned $\mathcal{T}_{k + 1}^{3\mathrm{d}}$ to world points $\mathbf{P}_{k + 1}^{\mathrm{world}}\in \mathbb{R}^{N\times 3}$ , where the dynamic points are filtered with the estimated dynamic scores. After this, we apply a direct bundle adjustment to optimize the camera poses with paired $\mathbf{P}_{k + 1}^{\mathrm{world}},\mathcal{T}_{k + 1}^{2\mathrm{d}}$ and $p_{k + 1}^{\mathrm{vis}}$ . In the next iteration, the updated camera poses $\mathcal{P}_{k + 1}$ influence the subsequent loop through global position embedding. It is worth noting that the 2D and 3D trajectories are not updated via bundle adjustment; instead, the entire system is primarily driven by SyncFormer. As shown in Fig. 4, the reprojection errors decrease rapidly over iterations, while the 2D trajectories gradually become consistent with the 3D projections transformed by camera motion. + +# 3.3. Training + +Training Datasets. Our model is trained on a large collection of 17 datasets, each providing different forms of supervision. The training data can be broadly categorized into three types: (1) Posed RGB-D with tracking annotations, + +(2) Posed RGB-D, (3) Pose-only or unlabeled data. For (1) containing Kubric [21], PointOdyssey [96] and Dynamic Replica [31], we add the full supervisions of camera poses, video depth, dynamic segmentation and 2D, 3D tracking. Within category (2), including VKITTI [7], TartanAir [80], Spring [53], DL3DV [44], BEDLAM [6], MVS-Synth [28], CO3Dv2 [61], COP3D [65], WildRGBD [83], ScanNet++ [12] and OmniObject3D [82], we leverage depth supervision to improve the video depth estimator. In addition, joint training losses are applied to ensure that the 3D tracking remains consistent with the estimated depth. As for category (3), it includes HOI4D [46], Ego4D [20], and Stereo4D [30]. For this type of data, we apply camera loss and joint training losses here, and follow [9] by using a monocular depth model [78] as a teacher to preserve relative depth accuracy. It is worth noting that Stereo4D [30] provides only very sparse depth on valid 3D tracks. Therefore, we choose not to use its annotations and instead treat it as pose-only data, given that it is sourced from Internet videos with rich scene diversity. + +Implementation details. Our training recipe consists of several stages. Stage 1. We first train the front end model to jointly estimate video depth and camera poses on category (1) and (2) datasets which sums up to 14 datasets. Training is conducted with mixed BF16 precision, while the DPT and camera tracker modules use full precision to ensure stable optimization. We use the AdamW optimizer with a learning rate of $5 \times 10^{-5}$ and apply gradient clipping with a threshold of 0.1. Stage 1 training is performed for 200k iterations on 64 H20 GPUs. We shuffle the video length from 1-24 during the training. Stage 2. At this stage, we initialize SyncFormer using the category (1) datasets, where ground-truth depth is provided and camera poses are initialized as identity matrices. We load the Cotracker3 [32] checkpoint to initialize the 2D tracking branch. This stage training takes 3 days of 100k iterations on 8 H20 GPUs. The video length is shuffle from 12-48 during the training. Stage 3. Finally, we fixed the alternation-attention layers in front end and train the whole pipeline in all datasets for 20 hours to converge. + +# 4. Experiments + +We evaluate our model across all sub-tasks, including 3D tracking (Sec. 4.1), and dynamic 3D reconstruction (Sec. 4.2). In addition, we conduct comprehensive ablation studies (Sec. 4.3) to analyze the impact of key design choices and to demonstrate the effectiveness of unified modeling and scaling up training. + +# 4.1. 3D Point Tracking + +Dataset. We evaluate our model and compare it with existing baselines on TAPVid-3D [37], a comprehensive benchmark spanning diverse scenarios including Driving, Ego + +Table 1. 3D tracking results on the TAPVid-3D benchmark. We report the 3D Average Jaccard (AJ), Average 3D Position Accuracy $(\mathrm{APD}_{3D})$ , and Occlusion Accuracy (OA) across the Aria, DriveTrack, and PStudio subsets. $offl^{+}$ and $offl^{-}$ denote our offline model with/without considering the camera motion, respectively. COL, Univ2, and Mega are abbreviations for COLMAP, UnidepthV2, and MegaSAM. The best and the second best are highlighted. + +
MethodsTypeDepth / Cam PoseAriaDriveTrackPStudioAverage
AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑
BootsTAPIRType ICOL9.114.578.611.818.683.86.911.681.89.314.981.4
TAPTRType IUniv215.724.287.812.419.184.87.313.584.311.818.985.6
LocoTrackType IUniv215.124.083.513.019.882.87.213.180.111.819.082.3
CoTracker3Type ICOL8.012.378.611.719.181.78.113.577.29.315.079.1
Univ215.824.488.913.519.987.19.213.884.212.819.486.7
Mega20.430.189.814.120.388.517.427.285.017.325.987.8
SpatialTrackerType IIUniv213.620.990.58.314.582.88.015.075.810.016.883.0
Mega15.923.890.17.713.585.215.325.278.113.020.884.5
SceneTrackerType IIUniv2-23.1-6.8--12.7---14.2-
DELTAType IIUniv216.624.486.814.622.585.88.215.076.413.120.683.0
Ours-offl-Type IIUniv218.626.390.816.424.390.218.127.686.717.726.089.2
Mega22.332.293.715.823.090.018.228.687.318.727.990.5
TAPIP3DType IIIMega23.532.891.214.921.882.618.127.785.518.827.486.4
Ours-offl+Type IIIMega24.735.293.916.023.490.118.628.786.119.829.190.0
Full-ours24.634.793.617.626.190.821.932.187.421.231.090.6
+ +Table 2. Video depth evaluation. Type I represents the methods specialized in video depth estimation, while Type II are neural reconstruction model, jointly recovering the geometry and camera motion from the video. Type III denotes the SoTA optimization-based method. The best and the second best results are highlighted. + +
Method / MetricsAverageKITTI [19]TUM Dyn [66]Bonn [59]Sintel [52]
AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)
DepthCrafter [27]0.1430.8570.1110.8850.1230.8730.0660.9790.2720.693
VDA [9]0.1540.8820.0800.9510.1180.9200.0490.9820.3700.674
DUSt3R [79]0.2400.7660.1240.8490.1870.7920.1740.8350.4750.591
MonST3R [91]0.1710.8020.0830.9340.1970.7260.0610.9540.3430.594
CUT3R [77]0.1860.8140.1040.8990.1080.8470.0680.9500.4660.560
VGGT [74]0.1040.8810.0510.9660.0680.9390.0560.9630.2420.659
MegaSAM [74]0.0930.8940.0690.9160.0810.9350.0370.9770.1850.746
Ours0.0810.9100.0520.9730.0450.9760.0280.9880.1990.703
+ +centric, and Studio. This benchmark consists of 4,569 evaluation videos, where the video length varies from 25 to 300 frames, and three 3D point tracking metrics are reported. Specifically, Occlusion Accuracy (OA) measures the precision of occlusion predictions; $\mathrm{APD}_{3D}$ denotes the average percentage of estimated errors within multiple threshold scales $\delta$ ; and Average Jaccard (AJ) quantifies the accuracy of both position and occlusion estimation. + +Baselines and Settings. The existing baselines can be broadly categorized into three types. Type I: 2D trackers followed by depth lifting. We report the current state-of-the-art 2D tracking models, CoTracker3 [32] and BootsTAPIR [16], as representatives of this category. Type II: 3D trackers in camera space. We include SpatialTrackerV1 [86] and DELTA [54], evaluated with two state-of-the-art depth estimators, i.e., UnidepthV2 [60] and MegaSAM [42]. Notably, MegaSAM is an optimization-based SLAM system for estimating consistent video depth, + +which generally performs better than feedforward models such as UnidepthV2. To ensure fair comparisons under the same camera-space 3D tracking setup, our 3D tracker module is also evaluated in combination with these depth estimators. Type III: 3D trackers in world space. We compare our fully end-to-end model with the recently released TAPIP3D [90], which requires consistent video depth and camera poses as input, provided by MegaSAM. For a thorough and fair comparison, we also report results where our depth and pose estimations are replaced by those from MegaSAM. + +Quantitative Results and Analysis. In the TAPVid-3D benchmark, our method achieves the best results across various settings, consistently outperforming all baselines under comparable conditions as in Tab. 1. From those detailed comparisons, we can draw several conclusions from it. + +- Better depth estimation is crucial for 3D tracking: For Type I methods, the 3D tracking performance largely re + +flects the accuracy of the underlying 2D tracks and the associated depth predictions. Intriguingly, CoTracker3 + MegaSAM achieves substantial improvements over CoTracker3 + UniDepthV2, with 17.3 vs. 12.8 in AJ, 25.9 vs. 19.4 in $APD_{3D}$ , and 87.8 vs. 86.7 in OA. This represents the best performance among Type I methods. However, our method Ours-offl $^{-}$ , which uses the weaker depth estimator UniDepthV2, still outperforms CoTracker3 with MegaSAM, achieving 18.7 vs. 17.3 in AJ and 27.4 vs. 25.9 in $APD_{3D}$ . This is because Type I methods rely purely on back-projection, making them more sensitive to depth inconsistencies. In contrast, our method directly predicts 3D trajectories using a spatial-temporal transformer, which provides inherent temporal smoothness and robustness to noisy inputs. + +- Camera motion decomposition improves 3D tracking: As shown in the comparison between Ours-offl $^+$ MegaSAM and Ours-offl $^+$ MegaSAM, incorporating camera pose estimation clearly improves 3D tracking accuracy. For example, on the Aria subset, our method achieves 24.7 vs. 22.3 in AJ and 35.2 vs. 32.2 in $APD_{3D}$ . This improvement is largely due to the design of the TAPVid-3D benchmark, which includes a substantial number of background points for evaluation in Aria. These background points primarily reflect camera motion and are particularly challenging for methods that only track in camera space, especially due to frequent out-of-view scenarios. In DriveTrack and Pstudio, the improvements are less, as DriveTrack includes only dynamic points on moving vehicles, while Pstudio contains static scenes with no camera motion. + +# 4.2. Dynamic 3D Reconstruction + +# 4.2.1. Video Depth Evaluation + +Datasets. To illustrate the effectiveness and generalization ability of our method, we evaluate our video depth estimation method across four mainstream datasets, i.e., KITTI [19], Sintel [52], Bonn [59], TUM Dynamics [66]. These datasets encompass both indoor and outdoor scenes, featuring video sequences ranging from 50 to 110 frames, providing a comprehensive benchmark for assessing depth estimation consistency across varied environments. + +Baselines. We compare our approach against three categories of methods. Type I includes SoTA video depth methods, DepthCrafter [27] and Video Depth Anything [87]. Type II consists of SoTA deep reconstruction approaches, i.e. DUST3R [79], MonST3R [91], CUT3R [77] and VGGT [74] Type III denotes the SoTA dynamic Structurefrom-Motion (SfM) system, MegaSAM [42]. It leverages an optimization-based paradigm to jointly estimate consistent video depths and camera poses, with constraints enforced by optical flow and monocular depth priors. + +Metrics. We evaluate the performance of video depth esti + +Table 3. Evaluation on Camera Pose Estimation on TUMdynamic [66], Lightspeed [12], and Sintel [52]. All values are absolute trajectory error (ATE), relative pose error (RPE) for translation and rotation. The best and the second best are highlighted. + +
MethodTUM-dynamic (ATE / RPEt / RER)Lightspeed (ATE / RPEt / RER)Sintel (ATE / RPEt / RER)
Particle-SfM [95]-0.185 / 0.075 / 2.9900.129 / 0.031 / 0.535
Robust-CVD [36]0.153 / 0.026 / 3.528-/-/-0.360 / 0.154 / 3.443
CasualSAM [94]0.071 / 0.010 / 1.712-/-/-0.141 / 0.035 / 0.615
DUST3R [79]0.140 / 0.106 / 3.2860.412 / 0.177 / 20.1000.290 / 0.132 / 7.869
CUT3R [77]0.046 / 0.015 / 0.4730.274 / 0.067 / 1.5610.213 / 0.066 / 0.621
MonST3R [91]0.098 / 0.019 / 0.9350.149 / 0.046 / 1.2100.078 / 0.038 / 0.490
VGGT [74]0.021 / 0.013 / 0.3270.226 / 0.086 / 1.7290.082 / 0.043 / 1.253
MegaSAM [42]0.013 / 0.011 / 0.3400.105 / 0.040 / 0.9960.023 / 0.008 / 0.060
Ours-Front0.038 / 0.022 / 0.4800.203 / 0.079 / 1.6890.075 / 0.045 / 0.805
Ours0.012 / 0.010 / 0.3050.134 / 0.039 / 1.3400.054 / 0.027 / 0.288
+ +mation using geometric accuracy metrics. To ensure fair comparisons with previous works [9, 77, 91], we follow their approach by aligning the predicted video depth to the ground truth using a shared scale and shift, and then compute the Absolute Relative Error (AbsRel) and $\delta_{1}$ metrics. + +Quantitative Results. As shown in Tab. 2, our method clearly outperforms all existing approaches. Specifically, we surpass our baseline in Type II, VGGT [74], by a significant margin. On average, our method achieves an AbsRel of 0.081 compared to 0.104 $(-22.1\%)$ , and a $\delta_{1.25}$ of 0.910 compared to 0.881 $(+3.3\%)$ . Besides, compared to MegaSAM (Type III), our method also maintain clear advantages with 0.081 v.s. 0.093 in AbsRel, and 0.910 v.s. 0.894 in $\delta_{1.25}$ . It is important to note that MegaSAM [42] usually needs 5-10 min for a 100 frames video, while our method only takes 5-10 seconds for each which nearly achieves $50 \times$ faster. + +# 4.2.2. Camera Poses + +To evaluate the accuracy of camera motion estimation, we conduct the comparisons on Sintel [52], TUM Dynamics [66], and Lightspeed [62]. Sintel and Lightspeed contains numerous challenging dynamic scenes with large egomotion and object motions, and they are both the synthetic data. TUM Dynamics [66] is a real data, captured by well-calibrated RGBD sensors. We report Absolute Translation Error (ATE), Relative Rotation Error (RPE rot) and Relative Translation Error (RTE) after Sim(3) alignment with the ground truth, as in [77, 91]. Shown in Tab. 3, our method outperforms all regression-based methods, much better than our baseline, VGGT [74] and on par with MegaSAM [42]. Besides, the table illustrates the significance of back end optimization. After joint motion optimization, the pose estimations become nearly twice accurate than before. + +# 4.3. Ablation Analysis + +Our ablation study investigates the impact of training data, joint training, and the SyncFormer design on 2D and 3D + +![](images/78cdb4cc883ead795b66e2d91c83c06eed037579c7ec3850282206577a85a974.jpg) +(a) Query + +![](images/700498ee5487a950ac2670fbaf2347360f162b6bee00609b7673d92b1a1fb551.jpg) +(b) w/o Joint + +![](images/20afaee421b5967959cc40756a53ce0e38b63b7556a6e3deb097816019d29e5e.jpg) +(c) w/ Joint +Figure 5. The influence of Joint Training. + +![](images/d9790ce95f5932ba6507ea61a019295a5602bbb843c9f1894780619e9b32404f.jpg) +(d) w/o Joint + +![](images/5db04851e56198cd2ecba82f7013500950e96773d52fa83975c606d90788caff.jpg) +(e) w/ Joint + +Table 4. TAP-Vid DAVIS results. Cotracker3-3D is a fine-tuned model by adding simple 3D project layer. Mean represents the average of AJ and ${\delta }_{\text{avg }}^{\text{vis }}$ . + +
MethodLift 3DAJ ↑δvisavg ↑OA ↑Mean ↑
TAPTRX63.076.191.169.5
LocoTrackX62.975.387.269.1
CoTracker3X64.476.991.270.7
DELTA3D-Dec62.776.788.269.7
SpatialTrackerTri-plane61.176.389.568.7
CoTracker3-3D3D-Proj51.665.285.358.4
OursSyncFormer64.977.591.071.2
+ +point tracking. For depth estimation, we analyze different loss functions and training strategies to improve the generalizability of video depth prediction. + +3D Point Tracking. To illustrate the gains brought by scaling up the 3D point tracking to a wider range of data, we naively train a base model (Base@K in Table 5), and an advanced one, Base@V,K,P, which is trained on VKITTI, Kubric, and PointOdyssey. As the table shows, Base@V,K,P outperforms Base@K by a clear margin. Meanwhile, our final version is clearly better than these two. Besides, different from Base@K, Base@V,K,P is jointly trained on VKITTI, which brings very significant improvements on a similar type of real data: DriveTrack shows 14.7 vs. 7.4 in AJ and 21.9 vs. 13.3 in $\mathrm{APD}_{3D}$ . Fig. 5 qualitatively demonstrates the meaning of joint training, i.e., joint training contributes to minimizing the 3D tracking drifts when the model is trained on new patterns of data. + +2D Point Tracking. We report a naive alternative to SyncFormer by modifying the final output layer of CoTracker3 [33] to output 2D and 3D tracking. The 2D and 3D embeddings are directly concatenated and projected back to the original dimension using an inserted, learnable linear layer. We name this naive baseline as Cotacker3-3D. As shown in Tab. 4, a simple adaptation of 3D lifting leads to a significant drop in 2D tracking accuracy—AJ drops from 64.4 to 51.6, $\delta$ from 76.9 to 65.2, and OA from 91.2 to 85.3. This degradation occurs because the 2D and 3D correlation signals become entangled, disrupting the model's ability to focus on reliable features. Moreover, the update dynamics in 2D and 3D differ substantially, as they occur in separate spaces—UVD space for 2D and camera space for 3D. We also report comparisons with other 3D lifting techniques. These results validate our SyncFormer design, which effec + +tively lifts tracking into 3D while preserving—or even improving—accuracy in 2D. + +Depth Estimation. We ablate different training datasets and their coverage to assess their impact on model performance. Additionally, we study the influence of different loss functions applied to real data, i.e., the pearson loss and full depth losses. As shown in Tab. 2, if we applied the full losses for depth will cause the obvious performance drop in the synthetic data, i.e. Sintel. On the contrary, if we only used the synthetic data to train the model, we found the performance in the real dataset, especially KITTI will heavily drop. Therefore, we leverage the strategy of using different losses for real and synthetic data to offset the influence of domain gap and different errors distribution. As shown in Tab. 2, Ours-Synthetic is our method trained only on the synthetic datasets. Ours-Real-Full is our method using the full depth loss in the real data, where the model was influenced by the noised distributions in the real dataset. The noised of real data has the negative influence on the model's influence on the synthetic data. Meanwhile, the different errors pattern also worsen the model's zero-shot capabilities in the real data. + +Table 5. Ablation of Training Data and Joint Training. K, V, P represents Kubric [21], VKITTI [7] and PointOdyssey [96] in training. Inference depths are provided by Unidepthv2. + +
MethodJoint TrainingAriaDriveTrackPStudio
AJAPD3DAJAPD3DAJAPD3D
Base@KNo16.024.47.413.312.620.3
Base@V, K, PYes15.724.114.721.917.227.4
+ +# 5. Conclusion + +This work introduces SpatialTrackerV2, a feedforward and scalable approach for 3D point tracking in monocular videos. Built upon a deep exploration of widely-used low- and mid-level representations of motion and scene geometry, our method unifies consistent scene geometry, camera motion, and pixel-wise 3D motion into a fully differentiable end-to-end pipeline. SpatialTrackerV2 accurately reconstructs 3D trajectories from monocular videos, achieving strong quantitative results on public benchmarks and demonstrating robust performance on casually captured Internet videos. We believe SpatialTrackerV2 establishes a solid foundation for real-world motion understanding and brings us a step closer to spatial intelligence by exploring large-scale vision data. + +# Acknowledgment + +This work was partially supported by Ant Group Research Intern Program, Zhejiang Provincial Natural Science Foundation of China (No. LR25F020003) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University. + +# References + +[1] Abhishek Badki, Hang Su, Bowen Wen, and Orazio Gallo. L4p: Low-level 4d vision perception unified. arXiv preprint arXiv:2502.13078, 2025. 2 +[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 3 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3 +[4] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. CoRR, abs/2501.02690, 2025. 2 +[5] Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3.1 - A model zoo for robust monocular relative depth estimation. CoRR, abs/2307.14460, 2023. 3 +[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737. IEEE, 2023. 5 +[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 8 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[9] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. CoRR, abs/2501.12375, 2025. 3, 5, 6, 7 +[10] Zequn Chen, Jiezhi Yang, and Heng Yang. Pref3r: Pose-free feed-forward 3d gaussian splatting from variable-length image sequence. CoRR, abs/2411.16877, 2024. 2 +[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In European Conference on Computer Vision, pages 306–325. Springer, 2024. 2 +[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 7 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3 +[14] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 2 +[15] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and + +temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2 +[16] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, pages 3257-3274, 2024. 3, 6 +[17] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 3 +[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011. Computer Vision Foundation / IEEE Computer Society, 2018. 3 +[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6, 7 +[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 5 +[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5, 8 +[22] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, Wenping Wang, and Yuan Liu. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. CoRR, abs/2501.03847, 2025. 2 +[23] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2 +[24] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3 +[25] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 3 +[26] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE TPAMI, 46(12):10579-10596, 2024. 3 +[27] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. CoRR, abs/2409.02095, 2024. 2, 3, 6, 7 +[28] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view + +stereopsis. In CVPR, pages 2821-2830. Computer Vision Foundation / IEEE Computer Society, 2018. 5 +[29] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy J. Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. CoRR, abs/2412.06016, 2024. 2 +[30] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint, 2024. 5 +[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 5 +[32] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3, 5, 6 +[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 4, 8 +[34] Yoni Kasten, Wuyue Lu, and Haggai Maron. Fast encoder-based 3d from casual videos via point track processing. arXiv preprint arXiv:2404.07097, 2024. 2 +[35] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In Proceedings of the IEEE international conference on computer vision, pages 1521–1529, 2017. 3 +[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1611-1621. Computer Vision Foundation / IEEE, 2021. 7 +[37] Skanda Koppula, Ignacio Rocco, Yi Yang, Joseph Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. In NeurIPS, 2024. 2, 5 +[38] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2 +[39] Hongdong Li and Richard Hartley. Five-point motion estimation made easy. In 18th International Conference on Pattern Recognition (ICPR'06), pages 630-633. IEEE, 2006. 3 +[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2024. 2 +[41] Xiaodi Li, Zongxin Yang, Ruijie Quan, and Yi Yang. Drip: Unleashing diffusion priors for joint foreground and alpha prediction in image matting. Advances in Neural Information Processing Systems, 37:79868-79888, 2024. 3 + +[42] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In CVPR, 2025. 2, 6, 7 +[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3 +[44] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 5 +[45] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE transactions on pattern analysis and machine intelligence, 33(5):978-994, 2010. 2 +[46] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In CVPR, pages 20981-20990. IEEE, 2022. 5 +[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2 +[48] Haozhe Lou, Yurong Liu, Yike Pan, Yiran Geng, Jianteng Chen, Wenlong Ma, Chenglong Li, Lin Wang, Hengzhen Feng, Lu Shi, Liyi Luo, and Yongliang Shi. Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation. CoRR, abs/2408.14873, 2024. 2 +[49] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 3 +[50] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 3 +[51] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3 +[52] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 6, 7 +[53] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 5 +[54] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. arXiv preprint arXiv:2410.24211, 2024. 2, 3, 6 + +[55] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 3 +[56] Dantong Niu, Yuvan Sharma, Haoru Xue, Giscard Biamby, Junyi Zhang, Ziteng Ji, Trevor Darrell, and Roei Herzig. Pretraining auto-regressive robotic models with 4d representations. arXiv preprint arXiv:2502.13142, 2025. 2 +[57] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8089-8099, 2024. 2 +[58] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 3 +[59] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 6, 7 +[60] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segù, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116. IEEE, 2024. 3, 6 +[61] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV. IEEE, 2021. 5 +[62] Chris Rockwell, Joseph Tung, Tsung-Yi Lin, Ming-Yu Liu, David F. Fouhey, and Chen-Hsuan Lin. Dynamic camera poses and where to find them. In CVPR, 2025. 7 +[63] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2 +[64] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2 +[65] Samarth Sinha, Roman Shapovalov, Jeremy Reizenstein, Ignacio Rocco, Natalia Neverova, Andrea Vedaldi, and David Novotny. Common pets in 3d: Dynamic new-view synthesis of real-life deformable categories. In CVPR, pages 4881-4891. IEEE, 2023. 5 +[66] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS, pages 573-580. IEEE, 2012. 6, 7 +[67] Tutian Tang, Minghao Liu, Wenqiang Xu, and Cewu Lu. Kalib: Markerless hand-eye calibration with keypoint tracking. CoRR, abs/2408.10562, 2024. 2 +[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3 +[69] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a mod- + +ern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21-22, 1999 Proceedings, pages 298-372. Springer, 2000. 3 +[70] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. CoRR, abs/2412.13389, 2024. 3 +[71] Bo Wang, Jian Li, Yang Yu, Li Liu, Zhenping Sun, and Dewen Hu. Scenetracker: Long-term scene flow estimation network. arXiv preprint arXiv:2403.19924, 2024. 3 +[72] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3 +[73] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 3 +[74] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 6, 7 +[75] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 3 +[76] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2 +[77] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025.6, 7 +[78] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 5 +[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 6, 7 +[80] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5 +[81] Zhouxia Wang, Yushi Lan, Shangchen Zhou, and Chen Change Loy. Objctrl-2.5 d: Training-free object control with camera poses. arXiv preprint arXiv:2412.07721, 2024. 2 + +[82] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR, pages 803-814. IEEE, 2023. 5 +[83] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBD objects in the wild: Scaling real-world 3d object learning from RGB-D videos. In CVPR, pages 22378-22389. IEEE, 2024. 5 +[84] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 3 +[85] Yuxi Xiao, Nan Xue, Tianfu Wu, and Gui-Song Xia. Level-S $^2$ fM: Structure From Motion on Neural Level Set of Implicit Surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17205-17214, 2023. 3 +[86] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2, 3, 6 +[87] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3, 7 +[88] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2025. 2, 3 +[89] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3906-3915. IEEE, 2022. 3 +[90] Bowei Zhang, Lei Ke, Adam W Harley, and Katerina Fragkiadaki. Tapip3d: Tracking any point in persistent 3d geometry. arXiv preprint arXiv:2504.14717, 2025. 3, 6 +[91] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 6, 7 +[92] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3 +[93] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 3 +[94] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 7 + +[95] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In *ECCV*. Springer, 2022. 2, 7 +[96] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 8 \ No newline at end of file diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/images.zip b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b1155a95f7d4e048d8f053769d583d9d87f04cc1 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b99a5b87ac01f929ef11a986b8c91efe51447059c36c955bf85476a847a28995 +size 731271 diff --git a/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/layout.json b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f6759c1456207876642eb7d249053363bfbb74a6 --- /dev/null +++ b/2025/SpatialTrackerV2_ Advancing 3D Point Tracking with Explicit Camera Motion/layout.json @@ -0,0 +1,9840 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "type": "text", + "content": "SpatialTrackerV2: Advancing 3D Point Tracking with Explicit Camera Motion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": "Yuxi Xiao" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Jianyuan Wang" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Nan Xue" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Nikita Karaev" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Yuri Makarov" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Bingyi Kang" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Xing Zhu" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Hujun Bao" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Yujun Shen" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "text", + "content": " Xiaowei Zhou" + }, + { + "bbox": [ + 67, + 141, + 542, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "spans": [ + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "content": "Oxford " + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "content": "Ant Group " + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "content": "Pixelwise AI " + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 78, + 177, + 534, + 194 + ], + "type": "text", + "content": "Bytedance Seed" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 201, + 304, + 379 + ], + "blocks": [ + { + "bbox": [ + 60, + 201, + 304, + 379 + ], + "lines": [ + { + "bbox": [ + 60, + 201, + 304, + 379 + ], + "spans": [ + { + "bbox": [ + 60, + 201, + 304, + 379 + ], + "type": "image", + "image_path": "a96c785c085d36ec1b1d6856714fabcf7afe3e82dad86dabac1fdf80f452a789.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 378, + 302, + 544 + ], + "blocks": [ + { + "bbox": [ + 62, + 378, + 302, + 544 + ], + "lines": [ + { + "bbox": [ + 62, + 378, + 302, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 302, + 544 + ], + "type": "image", + "image_path": "dd13107edc76019a3015327cc619496bc75ec554b56371a447106172748ce6f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 547, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 54, + 547, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 547, + 555, + 581 + ], + "type": "text", + "content": "Figure 1. SpatialTrackerV2 produces consistent 3D scene geometry, camera poses, and 3D point trajectories all at once from monocular videos of arbitrary scenarios, e.g., robotic manipulation, first-person egocentric views, and dynamic sports (drifting and skating) shown in this figure. Try our online demo at https://huggingface.co/spaces/Yuxihenry/SpatialTrackerV2." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 199, + 543, + 380 + ], + "blocks": [ + { + "bbox": [ + 305, + 199, + 543, + 380 + ], + "lines": [ + { + "bbox": [ + 305, + 199, + 543, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 543, + 380 + ], + "type": "image", + "image_path": "1613cdebd69c7240e2eb167eac480000ed1ad30259aaf5a6003d7a18d3d99f08.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 381, + 543, + 544 + ], + "blocks": [ + { + "bbox": [ + 306, + 381, + 543, + 544 + ], + "lines": [ + { + "bbox": [ + 306, + 381, + 543, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 543, + 544 + ], + "type": "image", + "image_path": "b08c78df9be708b49dceb74a24a922cc3794f52032ad900d5839ca838149d598.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 587, + 201, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 587, + 201, + 600 + ], + "spans": [ + { + "bbox": [ + 151, + 587, + 201, + 600 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 612, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 297, + 696 + ], + "type": "text", + "content": "We present SpatialTrackerV2, a feed-forward 3D point tracking method for monocular videos. Going beyond modular pipelines built on off-the-shelf components for 3D tracking, our approach unifies the intrinsic connections between point tracking, monocular depth, and camera pose estimation into a high-performing and feedforward 3D point tracker. It decomposes world-space 3D motion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "text", + "content": "into scene geometry, camera ego-motion, and pixel-wise object motion, with a fully differentiable and end-to-end architecture, allowing scalable training across a wide range of datasets, including synthetic sequences, posed RGB-D videos, and unlabeled in-the-wild footage. By learning geometry and motion jointly from such heterogeneous data, SpatialTrackerV2 outperforms existing 3D tracking methods by " + }, + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "text", + "content": ", and matches the accuracy of leading dynamic 3D reconstruction approaches while running " + }, + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "inline_equation", + "content": "50\\times" + }, + { + "bbox": [ + 313, + 591, + 556, + 700 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 702, + 279, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 702, + 279, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 702, + 279, + 714 + ], + "type": "text", + "content": "* Partially completed during Ant internship. † Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6726" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 210 + ], + "type": "text", + "content": "3D point tracking aims to recover long-term 3D trajectories of arbitrary points from monocular videos. As a universal dynamic scene representation, it has recently shown strong potential in diverse applications, including robotics [48, 56, 67], video generation [22, 29, 57, 81], and 3D/4D reconstruction [10, 34, 38, 76]. Compared to parametric motion models (e.g., SMPL [47], MANO [63], skeletons, or 3D bounding boxes), it offers greater flexibility and generalization over various real-world scenes, as shown in Fig. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 211, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 295, + 449 + ], + "type": "text", + "content": "Existing solutions [1, 4, 54, 75, 86] of 3D point tracking extensively explored the well-developed low/mid-level vision models, such as optical flow [45, 68] and monocular depth estimation [27, 87], and took benefits from 2D point tracking models [15, 23, 33]. Among them, optimization-based methods [42, 75, 95] distill the optical flow, monocular depth models and camera motions for each given monocular video with promising results obtained, while being computationally expensive due to their per-scene optimization designs. SpatialTracker [86] moved forward to efficient 3D point tracking with a feed-forward model, and the more recent works [1, 4, 54] explored different architecture designs and rendering constraints to achieve higher-quality 3D tracking. Nevertheless, the feed-forward solutions are limited to training data scalability issues due to the need for ground-truth 3D tracks as supervision, which downgrades the tracking quality in real-world casual captures. Moreover, overlooking the inherent interplay between camera motion, object motion, and scene geometry results in error entanglement and accumulation across modules." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "These limitations motivate our core insights: (1) The reliance on ground-truth 3D trajectories constrains the scalability of existing feed-forward models, highlighting the need for designs that can generalize across diverse and weakly-supervised data sources. (2) The absence of joint reasoning over scene geometry, camera motion, and object motion leads to compounded errors and degraded performance, underscoring the importance of disentangling and explicitly modeling these motion components. To address these challenges, we decompose 3D point tracking into three distinct components: video depth, ego (camera) motion, and object motion, and integrate them within a fully differentiable pipeline that supports scalable joint training across heterogeneous data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "In our SpatialTrackerV2, a front-end and back-end architecture is proposed. The front end is a video depth estimator and camera pose initializer, adapted from typical monocular depth prediction frameworks [88] with attention-based temporal information encoding [74]. The predicted video depths and camera poses are then fused through a scale-shift estimation module, which ensures consistency between the depth and motion predictions. The back end" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "consists of a proposed Joint Motion Optimization Module, which takes the video depth and coarse camera trajectories as input and iteratively estimates 2D and 3D trajectories, along with trajectory-wise dynamics and visibility scores. This enables an efficient bundle adjustment process for optimizing camera poses in the loop. At its core lies a novel SyncFormer, which separately models the 2D and 3D correlations in two branches, connected by multiple cross-attention layers. This design mitigates mutual interference between 2D and 3D embeddings and allows the model to update representations in two distinct spaces, namely the image (UV) space and the camera coordinate space. Furthermore, benefiting from this dual-branch design, bundle adjustment can be effectively applied to jointly optimize camera poses as well as the 2D and 3D trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 251, + 553, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 382 + ], + "type": "text", + "content": "This unified and differentiable pipeline makes large-scale training on diverse datasets possible. For RGB-D datasets with camera poses, we jointly train 3D tracking using consistency constraints from ground-truth depth and camera poses for static points, while dynamic points seamlessly contribute to the optimization. For video datasets that provide only camera pose annotations and lack depth information, we leverage consistency among camera poses, and 2D and 3D point tracking to drive the model's optimization. Relying on this framework, we successfully scale up training of the entire pipeline across 17 datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "content": "Evaluations on the TAPVid-3D benchmark [37] show that our method sets a new state-of-the-art in 3D point tracking, achieving 21.2 AJ and " + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "inline_equation", + "content": "31.0\\mathrm{APD}_{3D}" + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "content": ", surpassing DELTA [54] with relative improvements of " + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "inline_equation", + "content": "61.8\\%" + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "inline_equation", + "content": "50.5\\%" + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "content": ", respectively. Additionally, extensive experiments on dynamic reconstruction show our superior results on consistent video depth and camera poses estimation. Specifically, SpatialTrackerV2 beats the best dynamic reconstruction method, MegaSAM [42], on most video depth datasets and achieves comparable results on various camera pose benchmarks, while its inference speed is " + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "inline_equation", + "content": "50\\times" + }, + { + "bbox": [ + 313, + 383, + 553, + 515 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 525, + 399, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 399, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 399, + 538 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 569 + ], + "type": "text", + "content": "This section covers relevant literature on 3D tracking, depth, and camera pose estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 576, + 405, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 405, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 405, + 589 + ], + "type": "text", + "content": "2.1. Point tracking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "PIPs [23] revisited the 2D point tracking task first introduced in [64] and proposed a deep learning approach to solve it. TAP-Vid [14] redefined the problem and introduced both a benchmark and a simple architecture, TAP-Net. Subsequently, the performance was improved in TAPIR [15] by combining the global matching capabilities of TAP-Net with the local refinement offered by PIPs. CoTracker [33] pioneered tracking through occlusions using a transformer architecture combined with joint attention, followed by TAPTR [40] and LocoTrack [11], which im" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "6727" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "proved efficiency and introduced 4D correlation volumes. Recently, BootsTAPIR [16] and CoTracker3 [32] explored the use of unlabeled data to achieve better performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 295, + 275 + ], + "type": "text", + "content": "While 2D point tracking has been extensively studied, 3D point tracking remains a relatively new field. The first method to demonstrate 3D point tracking capabilities was the test-time optimization-based OmniMotion [75]. Later, SpatialTracker [86] introduced the first feed-forward 3D point tracker by combining a 2D point tracker [33] with depth priors from a monocular depth estimator [3]. SceneTracker [71] proposed a new architecture for 3D tracking with depth priors, while DELTA [54] improved efficiency and achieved dense 3D tracking. Recently, TAPIP3D [90] improved 3D tracking robustness by lifting image features into the world coordinate space and performing tracking there. All of these 3D tracking models were trained on small synthetic datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 324 + ], + "type": "text", + "content": "Unlike these methods, we present a scalable 3D tracking framework trained on a collection of both real and synthetic datasets, while also explicitly modeling camera motion to improve performance on egocentric videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 333, + 160, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 160, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 160, + 346 + ], + "type": "text", + "content": "2.2. Depth estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 446 + ], + "type": "text", + "content": "Early methods such as Eigen et al. [17] introduced single-view depth estimation using CNNs, but were limited by dataset scale and poor generalization [18, 89]. MiDaS [5] improved this by mixing datasets for broader coverage, and ZoeDepth [3] adapted it for metric depth, though ambiguity from missing camera intrinsics remained. Later, Metric3D [26] and UniDepth [60] addressed this by jointly estimating intrinsics and normalized depth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 495 + ], + "type": "text", + "content": "With large-scale pretraining (e.g., diffusion [13, 41], DINO [8]), recent models like Marigold [70] and DepthAnything [87, 88] have significantly advanced zero-shot depth. Extensions to video [9, 27] have also emerged." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 555 + ], + "type": "text", + "content": "In this work, we build on DepthAnythingV2 and extend it to video, aiming for not just consistent depth, but a unified framework that also predicts camera poses and tracks, with proper scale alignment—posing new challenges beyond static depth estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "text", + "content": "2.3. Camera estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Traditional camera pose estimation methods [24, 58] rely on image-to-image point correspondences using keypoint detectors (e.g., SIFT [49, 50], SURF [2]) and matching techniques such as nearest neighbors, followed by geometric algorithms like the five-point and eight-point methods [24, 25, 39, 55, 85]. Bundle Adjustment [69] is also commonly employed to further enhance accuracy. Recently, direct regression approaches using neural networks [35, 43, 51, 84, 92] have emerged, aiming to overcome limitations in sparse-view scenarios or when correspondences are unreliable. Diffusion models, such as PoseDiffusion [72] and RayDif" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "fusion [93], have also been explored, offering strong accuracy but suffering from high inference costs. In contrast, the camera head of VGGsfm [73] or VGGT [74] adopts an iterative refinement paradigm similar to RAFT [68], striking a balance between accuracy and inference cost, and enabling estimation of both extrinsic and intrinsic parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 370, + 166 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": "Given a video with " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}^t)_{t=1}^T" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " query points " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i = (x_i, y_i) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "i = 1, \\dots, N" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": ", the goal of 3D tracking is to recover pixel-wise 3D trajectories " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = (\\mathcal{T}_i^t)_{i=1,\\dots,N}" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " for each query. In order to account for static and dynamic parts of the scene, we decompose " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " into two components the ego camera motion " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ego}}" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": " and pixelwise object motion " + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{object}}" + }, + { + "bbox": [ + 313, + 174, + 555, + 260 + ], + "type": "text", + "content": ". Fig. 2 illustrates the pipeline of our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 450, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 450, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 450, + 279 + ], + "type": "text", + "content": "3.1. Ego Motion Component" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "content": "Ego motion, represented by camera trajectories " + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ego}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "content": ", is a major contributor to the 3D flow in camera coordinates. To compute the 3D tracks induced by ego motion, we need to estimate scale-aligned camera trajectories and video depth. Video Depth. Monocular depth estimation models, such as DepthAnything [87], typically follow an encoder-decoder design with a vision encoder like DINO [8] and a DPT-style decoder. Following VGGT [74], we extend the monocular encoder into a temporal encoder using an alternating-attention mechanism. This mechanism alternates between intra-frame self-attention and inter-frame attention over flattened video tokens, effectively balancing performance and efficiency. Furthermore, two learnable tokens, " + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{tok}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{tok}}" + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "content": ", are incorporated into the alternating-attention layers to capture high-level semantics for pose and scale regression. Camera Tracker. Similar to [73, 74], we adopt a differentiable pose head, " + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 283, + 555, + 498 + ], + "type": "text", + "content": ", to decode pose, scale and shift directly:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 498, + 555, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 498, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 382, + 498, + 555, + 512 + ], + "type": "interline_equation", + "content": "\\mathcal {P} ^ {t}, a, b = \\mathcal {H} \\left(\\mathbf {P} _ {\\mathrm {t o k}}, \\mathbf {S} _ {\\mathrm {t o k}}\\right), \\tag {1}", + "image_path": "12128fed7beb336dae00b557f8ad072c23ebcbd9824a950b52a0670e950f4db8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t\\in \\mathbb{R}^{T\\times 8}" + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "content": " is the camera encoding parameterized by a quaternion, a translation vector, and the normalized focal length concatenated together; The parameters " + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "content": " represent the scale and shift used to align the depth with the estimated camera poses. After that, we can easily obtain the 3D trajectories induced by ego-motion " + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{ego}}" + }, + { + "bbox": [ + 313, + 515, + 554, + 589 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 374, + 597, + 555, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 597, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 374, + 597, + 555, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\text {e g o}} = \\mathcal {W} (\\mathcal {P}, a * \\mathcal {D} _ {\\text {n o r m}} + b), \\tag {2}", + "image_path": "d008ce2991e60ab7bc90aaf13097d3ee61a6eca418bbfb3aa5bf6d674f2191ba.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{norm}}" + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "content": " is the raw results after DPT head with activation functions and " + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 617, + 554, + 640 + ], + "type": "text", + "content": " is the camera transformation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 647, + 466, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 466, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 466, + 661 + ], + "type": "text", + "content": "3.2. Joint Motion Optimization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "After ego-motion initialization, we jointly estimate 2D trajectories " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{2\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 2}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " in UV space and their corresponding 3D trajectories " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{3\\mathrm{d}} \\in \\mathbb{R}^{T \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " in the camera coordinate system using an iterative transformer module, referred" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6728" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 289 + ], + "type": "image", + "image_path": "f1bf9ec33a00fa908e2c02013dff54ac6ca106be1f5690bfc43cafb2fda30050.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 555, + 333 + ], + "type": "text", + "content": "Figure 2. Pipeline Overview. Our method adopts a front-end and back-end architecture. The front-end estimates scale-aligned depth and camera poses from the input video, which are used to construct initial static 3D tracks. The back-end then iteratively refines both tracks and poses via joint motion optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 354, + 294, + 453 + ], + "blocks": [ + { + "bbox": [ + 58, + 354, + 294, + 453 + ], + "lines": [ + { + "bbox": [ + 58, + 354, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 294, + 453 + ], + "type": "image", + "image_path": "476246a2ee49336015912e61cf61e16111fb83964215b0c5bea46f381660547a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 462, + 295, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 506 + ], + "type": "text", + "content": "Figure 3. SyncFormer. The model takes previous estimates and their corresponding embeddings as input, and updates them iteratively. The 2D and 3D embeddings are processed in separate branches that interact via cross-attention." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": "to as SyncFormer. In parallel, the SyncFormer also dynamically estimates the visibility probability " + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": " and dynamic probability " + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": " for each trajectory, enabling an efficient bundle adjustment process to refine the camera poses " + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 527, + 295, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": "SyncFormer. As illustrated in Fig. 3, in every iteration, SyncFormer takes 2D embeddings, 3D embeddings, and camera poses as input, and updates the 2D trajectories " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{2\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": ", 3D trajectories " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{T}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": ", dynamic probabilities " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": ", and visibility scores " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 643, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 643, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 61, + 643, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {k + 1} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k + 1} ^ {3 \\mathrm {d}}, p _ {k + 1} ^ {\\mathrm {d y n}}, p _ {k + 1} ^ {\\mathrm {v i s}} = f _ {\\text {s y n c}} \\left(\\mathcal {T} _ {k} ^ {2 \\mathrm {d}}, \\mathcal {T} _ {k} ^ {3 \\mathrm {d}}, p _ {k} ^ {\\mathrm {d y n}}, p _ {k} ^ {\\mathrm {v i s}}, \\mathcal {P} _ {k}\\right), \\tag {3}", + "image_path": "392a2e44fa0121384246ea01475f9f5cf24c9b640c6a8d1ca87996dbc37b4700.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{sync}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes the transformer-based update function. To better capture the distinct characteristics of 2D and 3D motion, the 2D and 3D trajectory updaters are modeled using separate attention layers. To reduce computational cost," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 449 + ], + "type": "text", + "content": "correlation embeddings are first encoded into a compact set of 2D and 3D proxy tokens using cross-attention. Then, information is exchanged between the 2D and 3D branches via a cross-attention layer between the respective proxy tokens. This design decouples the mutual influence between 2D and 3D tracking, which are updated in two different domains: the UV space for 2D trajectories and the camera coordinate space for 3D trajectories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": "2D and 3D Embeddings. As the input for SyncFormer, the 2D and 3D embeddings encode the status of current estimations while recording the neighbourhood information for updating. For 2D embeddings, we keep the same to Cotracker3 [33]. The 3D embeddings " + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{3\\mathrm{d}} = (Corr_{3\\mathrm{D}}, \\mathbf{e}^{\\mathrm{time}}, \\mathbf{e}^{\\mathrm{Gpos}}, p^{\\mathrm{dyn}}, p^{\\mathrm{vis}})" + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": " contains the 3D correlation features " + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "inline_equation", + "content": "Corr_{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": ", the time embeddings " + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{time}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": ", global position embeddings " + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{Gpos}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": " and dynamic-visibility scores. The 3D correlation " + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "inline_equation", + "content": "Corr_{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 449, + 556, + 689 + ], + "type": "text", + "content": " is the main features for the 3D embedding. Different to 2D correlations, we expect 3D tracking branch updates the 3D position in the camera coordinate space. Therefore, we calculate the 3D correlations on the normalized point maps from front-end instead of depth map. Similar to 2D correlations, we construct multi-resolution point maps and compute the relative translations between each point and its neighbors within radius of 3. We then apply harmonic positional encoding to project these relative translations into high-dimensional feature representations, which are combined with semantic features to compute the final 3D correlations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 695, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 695, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 319, + 695, + 555, + 717 + ], + "type": "interline_equation", + "content": "C o r r _ {3 \\mathrm {D}} = \\left[ K \\left(\\frac {\\mathbf {x}}{k s} + \\delta , \\frac {\\mathbf {y}}{k s} + \\delta\\right): \\delta \\in \\mathbb {Z}, \\| \\delta \\| _ {\\infty} \\leq \\Delta \\right], \\tag {4}", + "image_path": "ef2414a33f6aca91ff179d6fa4107f3e630d1b8fc8b4a6b44d89da49faaf0088.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6729" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 170 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 170 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 170 + ], + "type": "image", + "image_path": "d5dcde056f3c29d26c85be8c5f2d86ee291ffdd649590b0e25f164e1f6ca84aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 296, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 245 + ], + "type": "text", + "content": "Figure 4. Iterative Process of SyncFormer. The left subfigure shows the convergence curve of reprojection error, illustrating rapid reprojection error reduction. The right subfigures visualize the progressive alignment between 2D tracking results (in green) and projections of 3D (world) tracking transformed by camera poses (in red)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": " set to 3 and at multiple scales " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "s = 1,2,3,4" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": " denotes the operator for encoding relative translation features. Besides, in order to encode the camera pose to assist the tracking estimations, we propose " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{Gpos}}" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": " which transforms the query points " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": " into each frames with the current camera poses, denoted as anchor points " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_i^{\\mathrm{anc}}" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{Gpos}}" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": " is calculated from the relative translation between the current position and " + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{\\mathrm{Gpos}}" + }, + { + "bbox": [ + 54, + 266, + 295, + 373 + ], + "type": "text", + "content": ", and they are then projected into high dimension with same tricks as above." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": "Camera Motion Optimization. After each iteration, the updates of " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{k + 1}^{2\\mathrm{d}},\\mathcal{T}_{k + 1}^{3\\mathrm{d}},p_{k + 1}^{\\mathrm{dyn}},p_{k + 1}^{\\mathrm{vis}}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": " naturally form self-consistency constraints with the camera poses. These intrinsic constraints can be seamlessly incorporated into a bundle adjustment formulation. Specifically, we apply a weighted Procrustes analysis to register " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{k + 1}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": " into the world coordinate frame, where the alignment weights are given by the dynamic scores. This coarse alignment process is differentiable and provides effective supervision for learning the dynamic probability. Then we fuse the aligned " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{k + 1}^{3\\mathrm{d}}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": " to world points " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{k + 1}^{\\mathrm{world}}\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": ", where the dynamic points are filtered with the estimated dynamic scores. After this, we apply a direct bundle adjustment to optimize the camera poses with paired " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{k + 1}^{\\mathrm{world}},\\mathcal{T}_{k + 1}^{2\\mathrm{d}}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "p_{k + 1}^{\\mathrm{vis}}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": ". In the next iteration, the updated camera poses " + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{k + 1}" + }, + { + "bbox": [ + 56, + 375, + 296, + 639 + ], + "type": "text", + "content": " influence the subsequent loop through global position embedding. It is worth noting that the 2D and 3D trajectories are not updated via bundle adjustment; instead, the entire system is primarily driven by SyncFormer. As shown in Fig. 4, the reprojection errors decrease rapidly over iterations, while the 2D trajectories gradually become consistent with the 3D projections transformed by camera motion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 647, + 119, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 119, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 119, + 661 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Training Datasets. Our model is trained on a large collection of 17 datasets, each providing different forms of supervision. The training data can be broadly categorized into three types: (1) Posed RGB-D with tracking annotations," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "(2) Posed RGB-D, (3) Pose-only or unlabeled data. For (1) containing Kubric [21], PointOdyssey [96] and Dynamic Replica [31], we add the full supervisions of camera poses, video depth, dynamic segmentation and 2D, 3D tracking. Within category (2), including VKITTI [7], TartanAir [80], Spring [53], DL3DV [44], BEDLAM [6], MVS-Synth [28], CO3Dv2 [61], COP3D [65], WildRGBD [83], ScanNet++ [12] and OmniObject3D [82], we leverage depth supervision to improve the video depth estimator. In addition, joint training losses are applied to ensure that the 3D tracking remains consistent with the estimated depth. As for category (3), it includes HOI4D [46], Ego4D [20], and Stereo4D [30]. For this type of data, we apply camera loss and joint training losses here, and follow [9] by using a monocular depth model [78] as a teacher to preserve relative depth accuracy. It is worth noting that Stereo4D [30] provides only very sparse depth on valid 3D tracks. Therefore, we choose not to use its annotations and instead treat it as pose-only data, given that it is sourced from Internet videos with rich scene diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 551 + ], + "type": "text", + "content": "Implementation details. Our training recipe consists of several stages. Stage 1. We first train the front end model to jointly estimate video depth and camera poses on category (1) and (2) datasets which sums up to 14 datasets. Training is conducted with mixed BF16 precision, while the DPT and camera tracker modules use full precision to ensure stable optimization. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 312, + 555, + 551 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 312, + 555, + 551 + ], + "type": "text", + "content": " and apply gradient clipping with a threshold of 0.1. Stage 1 training is performed for 200k iterations on 64 H20 GPUs. We shuffle the video length from 1-24 during the training. Stage 2. At this stage, we initialize SyncFormer using the category (1) datasets, where ground-truth depth is provided and camera poses are initialized as identity matrices. We load the Cotracker3 [32] checkpoint to initialize the 2D tracking branch. This stage training takes 3 days of 100k iterations on 8 H20 GPUs. The video length is shuffle from 12-48 during the training. Stage 3. Finally, we fixed the alternation-attention layers in front end and train the whole pipeline in all datasets for 20 hours to converge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 561, + 395, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 395, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 395, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 581, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 654 + ], + "type": "text", + "content": "We evaluate our model across all sub-tasks, including 3D tracking (Sec. 4.1), and dynamic 3D reconstruction (Sec. 4.2). In addition, we conduct comprehensive ablation studies (Sec. 4.3) to analyze the impact of key design choices and to demonstrate the effectiveness of unified modeling and scaling up training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 660, + 424, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 424, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 424, + 673 + ], + "type": "text", + "content": "4.1. 3D Point Tracking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Dataset. We evaluate our model and compare it with existing baselines on TAPVid-3D [37], a comprehensive benchmark spanning diverse scenarios including Driving, Ego" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6730" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 115, + 552, + 307 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. 3D tracking results on the TAPVid-3D benchmark. We report the 3D Average Jaccard (AJ), Average 3D Position Accuracy " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "(\\mathrm{APD}_{3D})" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": ", and Occlusion Accuracy (OA) across the Aria, DriveTrack, and PStudio subsets. " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "offl^{+}" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "offl^{-}" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " denote our offline model with/without considering the camera motion, respectively. COL, Univ2, and Mega are abbreviations for COLMAP, UnidepthV2, and MegaSAM. The best and the second best are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 115, + 552, + 307 + ], + "lines": [ + { + "bbox": [ + 58, + 115, + 552, + 307 + ], + "spans": [ + { + "bbox": [ + 58, + 115, + 552, + 307 + ], + "type": "table", + "html": "
MethodsTypeDepth / Cam PoseAriaDriveTrackPStudioAverage
AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑AJ ↑APD3D ↑OA ↑
BootsTAPIRType ICOL9.114.578.611.818.683.86.911.681.89.314.981.4
TAPTRType IUniv215.724.287.812.419.184.87.313.584.311.818.985.6
LocoTrackType IUniv215.124.083.513.019.882.87.213.180.111.819.082.3
CoTracker3Type ICOL8.012.378.611.719.181.78.113.577.29.315.079.1
Univ215.824.488.913.519.987.19.213.884.212.819.486.7
Mega20.430.189.814.120.388.517.427.285.017.325.987.8
SpatialTrackerType IIUniv213.620.990.58.314.582.88.015.075.810.016.883.0
Mega15.923.890.17.713.585.215.325.278.113.020.884.5
SceneTrackerType IIUniv2-23.1-6.8--12.7---14.2-
DELTAType IIUniv216.624.486.814.622.585.88.215.076.413.120.683.0
Ours-offl-Type IIUniv218.626.390.816.424.390.218.127.686.717.726.089.2
Mega22.332.293.715.823.090.018.228.687.318.727.990.5
TAPIP3DType IIIMega23.532.891.214.921.882.618.127.785.518.827.486.4
Ours-offl+Type IIIMega24.735.293.916.023.490.118.628.786.119.829.190.0
Full-ours24.634.793.617.626.190.821.932.187.421.231.090.6
", + "image_path": "6582915d64767e3dc0648d192ca010228b6df6d5957b5fc696bf7bf02ed09c20.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 350, + 548, + 471 + ], + "blocks": [ + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 350 + ], + "type": "text", + "content": "Table 2. Video depth evaluation. Type I represents the methods specialized in video depth estimation, while Type II are neural reconstruction model, jointly recovering the geometry and camera motion from the video. Type III denotes the SoTA optimization-based method. The best and the second best results are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 350, + 548, + 471 + ], + "lines": [ + { + "bbox": [ + 61, + 350, + 548, + 471 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 548, + 471 + ], + "type": "table", + "html": "
Method / MetricsAverageKITTI [19]TUM Dyn [66]Bonn [59]Sintel [52]
AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)AbsRel (↓)δ1.25 (↑)
DepthCrafter [27]0.1430.8570.1110.8850.1230.8730.0660.9790.2720.693
VDA [9]0.1540.8820.0800.9510.1180.9200.0490.9820.3700.674
DUSt3R [79]0.2400.7660.1240.8490.1870.7920.1740.8350.4750.591
MonST3R [91]0.1710.8020.0830.9340.1970.7260.0610.9540.3430.594
CUT3R [77]0.1860.8140.1040.8990.1080.8470.0680.9500.4660.560
VGGT [74]0.1040.8810.0510.9660.0680.9390.0560.9630.2420.659
MegaSAM [74]0.0930.8940.0690.9160.0810.9350.0370.9770.1850.746
Ours0.0810.9100.0520.9730.0450.9760.0280.9880.1990.703
", + "image_path": "999151d994b3970693ff4379e799ce3cddaa0fbbf6a835953cfda96d7b2a7a59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "text", + "content": "centric, and Studio. This benchmark consists of 4,569 evaluation videos, where the video length varies from 25 to 300 frames, and three 3D point tracking metrics are reported. Specifically, Occlusion Accuracy (OA) measures the precision of occlusion predictions; " + }, + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_{3D}" + }, + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "text", + "content": " denotes the average percentage of estimated errors within multiple threshold scales " + }, + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 491, + 294, + 587 + ], + "type": "text", + "content": "; and Average Jaccard (AJ) quantifies the accuracy of both position and occlusion estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 715 + ], + "type": "text", + "content": "Baselines and Settings. The existing baselines can be broadly categorized into three types. Type I: 2D trackers followed by depth lifting. We report the current state-of-the-art 2D tracking models, CoTracker3 [32] and BootsTAPIR [16], as representatives of this category. Type II: 3D trackers in camera space. We include SpatialTrackerV1 [86] and DELTA [54], evaluated with two state-of-the-art depth estimators, i.e., UnidepthV2 [60] and MegaSAM [42]. Notably, MegaSAM is an optimization-based SLAM system for estimating consistent video depth," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": "which generally performs better than feedforward models such as UnidepthV2. To ensure fair comparisons under the same camera-space 3D tracking setup, our 3D tracker module is also evaluated in combination with these depth estimators. Type III: 3D trackers in world space. We compare our fully end-to-end model with the recently released TAPIP3D [90], which requires consistent video depth and camera poses as input, provided by MegaSAM. For a thorough and fair comparison, we also report results where our depth and pose estimations are replaced by those from MegaSAM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 625, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 555, + 686 + ], + "type": "text", + "content": "Quantitative Results and Analysis. In the TAPVid-3D benchmark, our method achieves the best results across various settings, consistently outperforming all baselines under comparable conditions as in Tab. 1. From those detailed comparisons, we can draw several conclusions from it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "- Better depth estimation is crucial for 3D tracking: For Type I methods, the 3D tracking performance largely re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "6731" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "text", + "content": "flects the accuracy of the underlying 2D tracks and the associated depth predictions. Intriguingly, CoTracker3 + MegaSAM achieves substantial improvements over CoTracker3 + UniDepthV2, with 17.3 vs. 12.8 in AJ, 25.9 vs. 19.4 in " + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "inline_equation", + "content": "APD_{3D}" + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "text", + "content": ", and 87.8 vs. 86.7 in OA. This represents the best performance among Type I methods. However, our method Ours-offl" + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "text", + "content": ", which uses the weaker depth estimator UniDepthV2, still outperforms CoTracker3 with MegaSAM, achieving 18.7 vs. 17.3 in AJ and 27.4 vs. 25.9 in " + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "inline_equation", + "content": "APD_{3D}" + }, + { + "bbox": [ + 63, + 72, + 295, + 251 + ], + "type": "text", + "content": ". This is because Type I methods rely purely on back-projection, making them more sensitive to depth inconsistencies. In contrast, our method directly predicts 3D trajectories using a spatial-temporal transformer, which provides inherent temporal smoothness and robustness to noisy inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "text", + "content": "- Camera motion decomposition improves 3D tracking: As shown in the comparison between Ours-offl " + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "text", + "content": " MegaSAM and Ours-offl " + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "text", + "content": " MegaSAM, incorporating camera pose estimation clearly improves 3D tracking accuracy. For example, on the Aria subset, our method achieves 24.7 vs. 22.3 in AJ and 35.2 vs. 32.2 in " + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "inline_equation", + "content": "APD_{3D}" + }, + { + "bbox": [ + 56, + 252, + 295, + 443 + ], + "type": "text", + "content": ". This improvement is largely due to the design of the TAPVid-3D benchmark, which includes a substantial number of background points for evaluation in Aria. These background points primarily reflect camera motion and are particularly challenging for methods that only track in camera space, especially due to frequent out-of-view scenarios. In DriveTrack and Pstudio, the improvements are less, as DriveTrack includes only dynamic points on moving vehicles, while Pstudio contains static scenes with no camera motion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 452, + 211, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 211, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 211, + 464 + ], + "type": "text", + "content": "4.2. Dynamic 3D Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 470, + 186, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 186, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 186, + 482 + ], + "type": "text", + "content": "4.2.1. Video Depth Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 581 + ], + "type": "text", + "content": "Datasets. To illustrate the effectiveness and generalization ability of our method, we evaluate our video depth estimation method across four mainstream datasets, i.e., KITTI [19], Sintel [52], Bonn [59], TUM Dynamics [66]. These datasets encompass both indoor and outdoor scenes, featuring video sequences ranging from 50 to 110 frames, providing a comprehensive benchmark for assessing depth estimation consistency across varied environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 701 + ], + "type": "text", + "content": "Baselines. We compare our approach against three categories of methods. Type I includes SoTA video depth methods, DepthCrafter [27] and Video Depth Anything [87]. Type II consists of SoTA deep reconstruction approaches, i.e. DUST3R [79], MonST3R [91], CUT3R [77] and VGGT [74] Type III denotes the SoTA dynamic Structurefrom-Motion (SfM) system, MegaSAM [42]. It leverages an optimization-based paradigm to jointly estimate consistent video depths and camera poses, with constraints enforced by optical flow and monocular depth priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Metrics. We evaluate the performance of video depth esti" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 124, + 553, + 248 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "type": "text", + "content": "Table 3. Evaluation on Camera Pose Estimation on TUMdynamic [66], Lightspeed [12], and Sintel [52]. All values are absolute trajectory error (ATE), relative pose error (RPE) for translation and rotation. The best and the second best are highlighted." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 124, + 553, + 248 + ], + "lines": [ + { + "bbox": [ + 315, + 124, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 315, + 124, + 553, + 248 + ], + "type": "table", + "html": "
MethodTUM-dynamic (ATE / RPEt / RER)Lightspeed (ATE / RPEt / RER)Sintel (ATE / RPEt / RER)
Particle-SfM [95]-0.185 / 0.075 / 2.9900.129 / 0.031 / 0.535
Robust-CVD [36]0.153 / 0.026 / 3.528-/-/-0.360 / 0.154 / 3.443
CasualSAM [94]0.071 / 0.010 / 1.712-/-/-0.141 / 0.035 / 0.615
DUST3R [79]0.140 / 0.106 / 3.2860.412 / 0.177 / 20.1000.290 / 0.132 / 7.869
CUT3R [77]0.046 / 0.015 / 0.4730.274 / 0.067 / 1.5610.213 / 0.066 / 0.621
MonST3R [91]0.098 / 0.019 / 0.9350.149 / 0.046 / 1.2100.078 / 0.038 / 0.490
VGGT [74]0.021 / 0.013 / 0.3270.226 / 0.086 / 1.7290.082 / 0.043 / 1.253
MegaSAM [42]0.013 / 0.011 / 0.3400.105 / 0.040 / 0.9960.023 / 0.008 / 0.060
Ours-Front0.038 / 0.022 / 0.4800.203 / 0.079 / 1.6890.075 / 0.045 / 0.805
Ours0.012 / 0.010 / 0.3050.134 / 0.039 / 1.3400.054 / 0.027 / 0.288
", + "image_path": "c368c64a7247880b9c14f32b1f9e9c1bcaec64faa186aa10109a03e68f2fc872.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 260, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 319 + ], + "type": "text", + "content": "mation using geometric accuracy metrics. To ensure fair comparisons with previous works [9, 77, 91], we follow their approach by aligning the predicted video depth to the ground truth using a shared scale and shift, and then compute the Absolute Relative Error (AbsRel) and " + }, + { + "bbox": [ + 313, + 260, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 260, + 553, + 319 + ], + "type": "text", + "content": " metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": "Quantitative Results. As shown in Tab. 2, our method clearly outperforms all existing approaches. Specifically, we surpass our baseline in Type II, VGGT [74], by a significant margin. On average, our method achieves an AbsRel of 0.081 compared to 0.104 " + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "inline_equation", + "content": "(-22.1\\%)" + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": ", and a " + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\delta_{1.25}" + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": " of 0.910 compared to 0.881 " + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "inline_equation", + "content": "(+3.3\\%)" + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": ". Besides, compared to MegaSAM (Type III), our method also maintain clear advantages with 0.081 v.s. 0.093 in AbsRel, and 0.910 v.s. 0.894 in " + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "inline_equation", + "content": "\\delta_{1.25}" + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": ". It is important to note that MegaSAM [42] usually needs 5-10 min for a 100 frames video, while our method only takes 5-10 seconds for each which nearly achieves " + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 313, + 320, + 555, + 462 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 469, + 403, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 403, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 403, + 480 + ], + "type": "text", + "content": "4.2.2. Camera Poses" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 665 + ], + "type": "text", + "content": "To evaluate the accuracy of camera motion estimation, we conduct the comparisons on Sintel [52], TUM Dynamics [66], and Lightspeed [62]. Sintel and Lightspeed contains numerous challenging dynamic scenes with large egomotion and object motions, and they are both the synthetic data. TUM Dynamics [66] is a real data, captured by well-calibrated RGBD sensors. We report Absolute Translation Error (ATE), Relative Rotation Error (RPE rot) and Relative Translation Error (RTE) after Sim(3) alignment with the ground truth, as in [77, 91]. Shown in Tab. 3, our method outperforms all regression-based methods, much better than our baseline, VGGT [74] and on par with MegaSAM [42]. Besides, the table illustrates the significance of back end optimization. After joint motion optimization, the pose estimations become nearly twice accurate than before." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 421, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 421, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 421, + 685 + ], + "type": "text", + "content": "4.3. Ablation Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Our ablation study investigates the impact of training data, joint training, and the SyncFormer design on 2D and 3D" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6732" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 77, + 101, + 106 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 101, + 106 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 101, + 106 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 101, + 106 + ], + "type": "image", + "image_path": "78cdb4cc883ead795b66e2d91c83c06eed037579c7ec3850282206577a85a974.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 108, + 97, + 117 + ], + "lines": [ + { + "bbox": [ + 66, + 108, + 97, + 117 + ], + "spans": [ + { + "bbox": [ + 66, + 108, + 97, + 117 + ], + "type": "text", + "content": "(a) Query" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 70, + 148, + 106 + ], + "blocks": [ + { + "bbox": [ + 109, + 70, + 148, + 106 + ], + "lines": [ + { + "bbox": [ + 109, + 70, + 148, + 106 + ], + "spans": [ + { + "bbox": [ + 109, + 70, + 148, + 106 + ], + "type": "image", + "image_path": "700498ee5487a950ac2670fbaf2347360f162b6bee00609b7673d92b1a1fb551.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 108, + 149, + 116 + ], + "lines": [ + { + "bbox": [ + 108, + 108, + 149, + 116 + ], + "spans": [ + { + "bbox": [ + 108, + 108, + 149, + 116 + ], + "type": "text", + "content": "(b) w/o Joint" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 70, + 195, + 106 + ], + "blocks": [ + { + "bbox": [ + 157, + 70, + 195, + 106 + ], + "lines": [ + { + "bbox": [ + 157, + 70, + 195, + 106 + ], + "spans": [ + { + "bbox": [ + 157, + 70, + 195, + 106 + ], + "type": "image", + "image_path": "20afaee421b5967959cc40756a53ce0e38b63b7556a6e3deb097816019d29e5e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 108, + 194, + 116 + ], + "lines": [ + { + "bbox": [ + 157, + 108, + 194, + 116 + ], + "spans": [ + { + "bbox": [ + 157, + 108, + 194, + 116 + ], + "type": "text", + "content": "(c) w/ Joint" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 126, + 251, + 138 + ], + "lines": [ + { + "bbox": [ + 99, + 126, + 251, + 138 + ], + "spans": [ + { + "bbox": [ + 99, + 126, + 251, + 138 + ], + "type": "text", + "content": "Figure 5. The influence of Joint Training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 70, + 242, + 106 + ], + "blocks": [ + { + "bbox": [ + 203, + 70, + 242, + 106 + ], + "lines": [ + { + "bbox": [ + 203, + 70, + 242, + 106 + ], + "spans": [ + { + "bbox": [ + 203, + 70, + 242, + 106 + ], + "type": "image", + "image_path": "d9790ce95f5932ba6507ea61a019295a5602bbb843c9f1894780619e9b32404f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 108, + 243, + 116 + ], + "lines": [ + { + "bbox": [ + 202, + 108, + 243, + 116 + ], + "spans": [ + { + "bbox": [ + 202, + 108, + 243, + 116 + ], + "type": "text", + "content": "(d) w/o Joint" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 70, + 288, + 106 + ], + "blocks": [ + { + "bbox": [ + 251, + 70, + 288, + 106 + ], + "lines": [ + { + "bbox": [ + 251, + 70, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 251, + 70, + 288, + 106 + ], + "type": "image", + "image_path": "5db04851e56198cd2ecba82f7013500950e96773d52fa83975c606d90788caff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 108, + 288, + 116 + ], + "lines": [ + { + "bbox": [ + 252, + 108, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 252, + 108, + 288, + 116 + ], + "type": "text", + "content": "(e) w/ Joint" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 194, + 294, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 295, + 186 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 186 + ], + "type": "text", + "content": "Table 4. TAP-Vid DAVIS results. Cotracker3-3D is a fine-tuned model by adding simple 3D project layer. Mean represents the average of AJ and " + }, + { + "bbox": [ + 55, + 152, + 295, + 186 + ], + "type": "inline_equation", + "content": "{\\delta }_{\\text{avg }}^{\\text{vis }}" + }, + { + "bbox": [ + 55, + 152, + 295, + 186 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 194, + 294, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 293 + ], + "type": "table", + "html": "
MethodLift 3DAJ ↑δvisavg ↑OA ↑Mean ↑
TAPTRX63.076.191.169.5
LocoTrackX62.975.387.269.1
CoTracker3X64.476.991.270.7
DELTA3D-Dec62.776.788.269.7
SpatialTrackerTri-plane61.176.389.568.7
CoTracker3-3D3D-Proj51.665.285.358.4
OursSyncFormer64.977.591.071.2
", + "image_path": "3543dbd07e45ff76b438e97c0eedf2d1f90faef6b200eaa1c8691dc1cf703d8e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 316, + 295, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 352 + ], + "type": "text", + "content": "point tracking. For depth estimation, we analyze different loss functions and training strategies to improve the generalizability of video depth prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 353, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 520 + ], + "type": "text", + "content": "3D Point Tracking. To illustrate the gains brought by scaling up the 3D point tracking to a wider range of data, we naively train a base model (Base@K in Table 5), and an advanced one, Base@V,K,P, which is trained on VKITTI, Kubric, and PointOdyssey. As the table shows, Base@V,K,P outperforms Base@K by a clear margin. Meanwhile, our final version is clearly better than these two. Besides, different from Base@K, Base@V,K,P is jointly trained on VKITTI, which brings very significant improvements on a similar type of real data: DriveTrack shows 14.7 vs. 7.4 in AJ and 21.9 vs. 13.3 in " + }, + { + "bbox": [ + 55, + 353, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathrm{APD}_{3D}" + }, + { + "bbox": [ + 55, + 353, + 296, + 520 + ], + "type": "text", + "content": ". Fig. 5 qualitatively demonstrates the meaning of joint training, i.e., joint training contributes to minimizing the 3D tracking drifts when the model is trained on new patterns of data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "2D Point Tracking. We report a naive alternative to SyncFormer by modifying the final output layer of CoTracker3 [33] to output 2D and 3D tracking. The 2D and 3D embeddings are directly concatenated and projected back to the original dimension using an inserted, learnable linear layer. We name this naive baseline as Cotacker3-3D. As shown in Tab. 4, a simple adaptation of 3D lifting leads to a significant drop in 2D tracking accuracy—AJ drops from 64.4 to 51.6, " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " from 76.9 to 65.2, and OA from 91.2 to 85.3. This degradation occurs because the 2D and 3D correlation signals become entangled, disrupting the model's ability to focus on reliable features. Moreover, the update dynamics in 2D and 3D differ substantially, as they occur in separate spaces—UVD space for 2D and camera space for 3D. We also report comparisons with other 3D lifting techniques. These results validate our SyncFormer design, which effec" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "tively lifts tracking into 3D while preserving—or even improving—accuracy in 2D." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "type": "text", + "content": "Depth Estimation. We ablate different training datasets and their coverage to assess their impact on model performance. Additionally, we study the influence of different loss functions applied to real data, i.e., the pearson loss and full depth losses. As shown in Tab. 2, if we applied the full losses for depth will cause the obvious performance drop in the synthetic data, i.e. Sintel. On the contrary, if we only used the synthetic data to train the model, we found the performance in the real dataset, especially KITTI will heavily drop. Therefore, we leverage the strategy of using different losses for real and synthetic data to offset the influence of domain gap and different errors distribution. As shown in Tab. 2, Ours-Synthetic is our method trained only on the synthetic datasets. Ours-Real-Full is our method using the full depth loss in the real data, where the model was influenced by the noised distributions in the real dataset. The noised of real data has the negative influence on the model's influence on the synthetic data. Meanwhile, the different errors pattern also worsen the model's zero-shot capabilities in the real data." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 315, + 373, + 553, + 412 + ], + "blocks": [ + { + "bbox": [ + 313, + 338, + 553, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 372 + ], + "type": "text", + "content": "Table 5. Ablation of Training Data and Joint Training. K, V, P represents Kubric [21], VKITTI [7] and PointOdyssey [96] in training. Inference depths are provided by Unidepthv2." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 373, + 553, + 412 + ], + "lines": [ + { + "bbox": [ + 315, + 373, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 315, + 373, + 553, + 412 + ], + "type": "table", + "html": "
MethodJoint TrainingAriaDriveTrackPStudio
AJAPD3DAJAPD3DAJAPD3D
Base@KNo16.024.47.413.312.620.3
Base@V, K, PYes15.724.114.721.917.227.4
", + "image_path": "50945381b81a3748a4e251fb7fa520d5e1d5baa1f255422abf9f0707823cbeb3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 618 + ], + "type": "text", + "content": "This work introduces SpatialTrackerV2, a feedforward and scalable approach for 3D point tracking in monocular videos. Built upon a deep exploration of widely-used low- and mid-level representations of motion and scene geometry, our method unifies consistent scene geometry, camera motion, and pixel-wise 3D motion into a fully differentiable end-to-end pipeline. SpatialTrackerV2 accurately reconstructs 3D trajectories from monocular videos, achieving strong quantitative results on public benchmarks and demonstrating robust performance on casually captured Internet videos. We believe SpatialTrackerV2 establishes a solid foundation for real-world motion understanding and brings us a step closer to spatial intelligence by exploring large-scale vision data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 632, + 407, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 632, + 407, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 632, + 407, + 646 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "This work was partially supported by Ant Group Research Intern Program, Zhejiang Provincial Natural Science Foundation of China (No. LR25F020003) and Information Technology Center and State Key Lab of CAD&CG, Zhejiang University." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6733" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Abhishek Badki, Hang Su, Bowen Wen, and Orazio Gallo. L4p: Low-level 4d vision perception unified. arXiv preprint arXiv:2502.13078, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "text", + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-Up Robust Features (SURF). CVIU, 110(3), 2008. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 248 + ], + "type": "text", + "content": "[4] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. CoRR, abs/2501.02690, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 282 + ], + "type": "text", + "content": "[5] Reiner Birkl, Diana Wofk, and Matthias Müller. Midas v3.1 - A model zoo for robust monocular relative depth estimation. CoRR, abs/2307.14460, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 327 + ], + "type": "text", + "content": "[6] Michael J. Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. BEDLAM: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR, pages 8726-8737. IEEE, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 329, + 295, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 351 + ], + "type": "text", + "content": "[7] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual kitti 2, 2020. 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "text", + "content": "[9] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zi long Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. CoRR, abs/2501.12375, 2025. 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 454, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 295, + 487 + ], + "type": "text", + "content": "[10] Zequn Chen, Jiezhi Yang, and Heng Yang. Pref3r: Pose-free feed-forward 3d gaussian splatting from variable-length image sequence. CoRR, abs/2411.16877, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "text", + "content": "[11] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seun-gryong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In European Conference on Computer Vision, pages 306–325. Springer, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 589 + ], + "type": "text", + "content": "[12] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[14] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610–13626, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[15] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 105 + ], + "type": "text", + "content": "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 106, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 554, + 171 + ], + "type": "text", + "content": "[16] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In Proceedings of the Asian Conference on Computer Vision, pages 3257-3274, 2024. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[17] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. In NeurIPS, pages 2366-2374, 2014. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "text", + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011. Computer Vision Foundation / IEEE Computer Society, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "text", + "content": "[19] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 371 + ], + "type": "text", + "content": "[20] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18995-19012, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 436 + ], + "type": "text", + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3749-3761, 2022. 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[22] Zekai Gu, Rui Yan, Jiahao Lu, Peng Li, Zhiyang Dou, Chenyang Si, Zhen Dong, Qifeng Liu, Cheng Lin, Ziwei Liu, Wenping Wang, and Yuan Liu. Diffusion as shader: 3d-aware video diffusion for versatile video generation control. CoRR, abs/2501.03847, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "text", + "content": "[23] Adam W. Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "text", + "content": "[24] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, ISBN: 0521540518, second edition, 2004. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 590 + ], + "type": "text", + "content": "[25] Richard I Hartley. In defense of the eight-point algorithm. IEEE Transactions on pattern analysis and machine intelligence, 19(6):580-593, 1997. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "text", + "content": "[26] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE TPAMI, 46(12):10579-10596, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[27] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. CoRR, abs/2409.02095, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[28] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6734" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "stereopsis. In CVPR, pages 2821-2830. Computer Vision Foundation / IEEE Computer Society, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[29] Hyeonho Jeong, Chun-Hao Paul Huang, Jong Chul Ye, Niloy J. Mitra, and Duygu Ceylan. Track4gen: Teaching video diffusion models to track points improves video generation. CoRR, abs/2412.06016, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[30] Linyi Jin, Richard Tucker, Zhengqi Li, David Fouhey, Noah Snavely, and Aleksander Holynski. Stereo4d: Learning how things move in 3d from internet stereo videos. arXiv preprint, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[31] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 307 + ], + "type": "text", + "content": "[32] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "text", + "content": "[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "text", + "content": "[34] Yoni Kasten, Wuyue Lu, and Haggai Maron. Fast encoder-based 3d from casual videos via point track processing. arXiv preprint arXiv:2404.07097, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[35] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In Proceedings of the IEEE international conference on computer vision, pages 1521–1529, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "text", + "content": "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1611-1621. Computer Vision Foundation / IEEE, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[37] Skanda Koppula, Ignacio Rocco, Yi Yang, Joseph Heyward, João Carreira, Andrew Zisserman, Gabriel Brostow, and Carl Doersch. Tapvid-3d: A benchmark for tracking any point in 3d. In NeurIPS, 2024. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 588 + ], + "type": "text", + "content": "[38] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 624 + ], + "type": "text", + "content": "[39] Hongdong Li and Richard Hartley. Five-point motion estimation made easy. In 18th International Conference on Pattern Recognition (ICPR'06), pages 630-633. IEEE, 2006. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[40] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. Taptr: Tracking any point with transformers as detection. In European Conference on Computer Vision, pages 57-75. Springer, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[41] Xiaodi Li, Zongxin Yang, Ruijie Quan, and Yi Yang. Drip: Unleashing diffusion priors for joint foreground and alpha prediction in image matting. Advances in Neural Information Processing Systems, 37:79868-79888, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[42] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snively. Megasam: Accurate, fast and robust structure and motion from casual dynamic videos. In CVPR, 2025. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "type": "text", + "content": "[44] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[45] Ce Liu, Jenny Yuen, and Antonio Torralba. Sift flow: Dense correspondence across scenes and its applications. IEEE transactions on pattern analysis and machine intelligence, 33(5):978-994, 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "text", + "content": "[46] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. HOI4D: A 4d egocentric dataset for category-level human-object interaction. In CVPR, pages 20981-20990. IEEE, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[47] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi-person linear model. ACM Trans. Graphics (Proc. SIGGRAPH Asia), 34(6):248:1-248:16, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 441 + ], + "type": "text", + "content": "[48] Haozhe Lou, Yurong Liu, Yike Pan, Yiran Geng, Jianteng Chen, Wenlong Ma, Chenglong Li, Lin Wang, Hengzhen Feng, Lu Shi, Liyi Luo, and Yongliang Shi. Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation. CoRR, abs/2408.14873, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 465 + ], + "type": "text", + "content": "[49] David G. Lowe. Object Recognition from Local Scale-Invariant Features. In Proc. ICCV, 1999. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 488 + ], + "type": "text", + "content": "[50] David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints. IJCV, 60(2), 2004. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 555, + 544 + ], + "type": "text", + "content": "[51] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 555, + 612 + ], + "type": "text", + "content": "[52] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4040-4048, 2016. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[53] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[54] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. Delta: Dense efficient long-range 3d tracking for any video. arXiv preprint arXiv:2410.24211, 2024. 2, 3, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6735" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[55] David Nister. An efficient solution to the five-point relative pose problem. IEEE transactions on pattern analysis and machine intelligence, 26(6):756-770, 2004. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[56] Dantong Niu, Yuvan Sharma, Haoru Xue, Giscard Biamby, Junyi Zhang, Ziteng Ji, Trevor Darrell, and Roei Herzig. Pretraining auto-regressive robotic models with 4d representations. arXiv preprint arXiv:2502.13142, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 217 + ], + "type": "text", + "content": "[57] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8089-8099, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 251 + ], + "type": "text", + "content": "[58] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "text", + "content": "[59] E. Palazzolo, J. Behley, P. Lottes, P. Giguère, and C. Stachniss. ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals. In IROS, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 342 + ], + "type": "text", + "content": "[60] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segù, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116. IEEE, 2024. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "type": "text", + "content": "[61] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In ICCV. IEEE, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 420 + ], + "type": "text", + "content": "[62] Chris Rockwell, Joseph Tung, Tsung-Yi Lin, Ming-Yu Liu, David F. Fouhey, and Chen-Hsuan Lin. Dynamic camera poses and where to find them. In CVPR, 2025. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[63] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "text", + "content": "[64] Peter Sand and Seth Teller. Particle video: Long-range motion estimation using point trajectories. International journal of computer vision, 80:72-91, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 555 + ], + "type": "text", + "content": "[65] Samarth Sinha, Roman Shapovalov, Jeremy Reizenstein, Ignacio Rocco, Natalia Neverova, Andrea Vedaldi, and David Novotny. Common pets in 3d: Dynamic new-view synthesis of real-life deformable categories. In CVPR, pages 4881-4891. IEEE, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 599 + ], + "type": "text", + "content": "[66] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In IROS, pages 573-580. IEEE, 2012. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 634 + ], + "type": "text", + "content": "[67] Tutian Tang, Minghao Liu, Wenqiang Xu, and Cewu Lu. Kalib: Markerless hand-eye calibration with keypoint tracking. CoRR, abs/2408.10562, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 689 + ], + "type": "text", + "content": "[68] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[69] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a mod-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 117 + ], + "type": "text", + "content": "ern synthesis. In Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms Corfu, Greece, September 21-22, 1999 Proceedings, pages 298-372. Springer, 2000. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "text", + "content": "[70] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. CoRR, abs/2412.13389, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 554, + 194 + ], + "type": "text", + "content": "[71] Bo Wang, Jian Li, Yang Yu, Li Liu, Zhenping Sun, and Dewen Hu. Scenetracker: Long-term scene flow estimation network. arXiv preprint arXiv:2403.19924, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 554, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 248 + ], + "type": "text", + "content": "[72] Jianyuan Wang, Christian Rupprecht, and David Novotny. Posediffusion: Solving pose estimation via diffusion-aided bundle adjustment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9773–9783, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 304 + ], + "type": "text", + "content": "[73] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 554, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 359 + ], + "type": "text", + "content": "[74] Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 5294-5306, 2025. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 554, + 415 + ], + "type": "text", + "content": "[75] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19795-19806, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "text", + "content": "[76] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 502 + ], + "type": "text", + "content": "[77] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025.6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "text", + "content": "[78] Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 601 + ], + "type": "text", + "content": "[79] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 668 + ], + "type": "text", + "content": "[80] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[81] Zhouxia Wang, Yushi Lan, Shangchen Zhou, and Chen Change Loy. Objctrl-2.5 d: Training-free object control with camera poses. arXiv preprint arXiv:2412.07721, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6736" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[82] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Jiawei Ren, Liang Pan, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In CVPR, pages 803-814. IEEE, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[83] Hongchi Xia, Yang Fu, Sifei Liu, and Xiaolong Wang. RGBD objects in the wild: Scaling real-world 3d object learning from RGB-D videos. In CVPR, pages 22378-22389. IEEE, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "type": "text", + "content": "[84] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 276 + ], + "type": "text", + "content": "[85] Yuxi Xiao, Nan Xue, Tianfu Wu, and Gui-Song Xia. Level-S" + }, + { + "bbox": [ + 56, + 222, + 294, + 276 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 56, + 222, + 294, + 276 + ], + "type": "text", + "content": "fM: Structure From Motion on Neural Level Set of Implicit Surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17205-17214, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 334 + ], + "type": "text", + "content": "[86] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "type": "text", + "content": "[87] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[88] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2025. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 482 + ], + "type": "text", + "content": "[89] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3906-3915. IEEE, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 485, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 485, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 294, + 518 + ], + "type": "text", + "content": "[90] Bowei Zhang, Lei Ke, Adam W Harley, and Katerina Fragkiadaki. Tapip3d: Tracking any point in persistent 3d geometry. arXiv preprint arXiv:2504.14717, 2025. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "text", + "content": "[91] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "type": "text", + "content": "[92] Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. In ECCV, pages 592-611. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "text", + "content": "[93] Jason Y Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. arXiv preprint arXiv:2402.14817, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[94] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T. Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 173 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[95] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In *ECCV*. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[96] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19855-19865, 2023. 5, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6737" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_content_list.json b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5233333c2790e2acbbd8507d3661c9112f270b0a --- /dev/null +++ b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_content_list.json @@ -0,0 +1,1541 @@ +[ + { + "type": "text", + "text": "Spatially-Varying Autofocus", + "text_level": 1, + "bbox": [ + 354, + 130, + 643, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingsi Qin, Aswin C. Sankaranarayanan, Matthew O'Toole \nCarnegie Mellon University", + "text_level": 1, + "bbox": [ + 264, + 180, + 736, + 218 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6910b5c6e496bb7f5864b70689d67320d430e5ebc335c730cf86c413d67ae569.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 234, + 385, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1feb8953730af0ecc694309fc73e9813636df0ec6a4413ea1c8f4736ce3736f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 236, + 495, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ff39be67e1b2def1ef0bc33818478bf8ea37577f533dae22e3ffaa806d4570c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 236, + 792, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/219357258e66128b3b89ac4dbe69a0e5c962884804880117b634620c1c938f22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 236, + 903, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/521b4623084adb37145d0a36210ec7e425c385fa032a9e261725fac5b6722290.jpg", + "image_caption": [ + "(a) Conventional photo and its confined focal plane" + ], + "image_footnote": [], + "bbox": [ + 93, + 364, + 493, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd69f3b1fd271c48cd468c4476822c2c5086eabbc9890d4f5a2d2501598467fa.jpg", + "image_caption": [ + "(b) All-In-Focus photo and its spatially-varying autofocused focal surface (ours)", + "Figure 1. Spatially-varying autofocus to produce an optical all-in-focus image. Left: A conventional photo with a regular lens, where objects at a single focal plane appear sharp. Right: An all-in-focus photo captured through spatially-varying autofocus. To achieve this, we combine (i) a programmable lens with spatially-varying control over focus, and (ii) a spatially-varying autofocus algorithm to drive the focus of this lens. Note that this is an optically-captured image of a real scene with no post-capture processing used." + ], + "image_footnote": [], + "bbox": [ + 501, + 364, + 903, + 430 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 521, + 326, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A lens brings a single plane into focus on a planar sensor; hence, parts of the scene that are outside this planar focus plane are resolved under defocus. Can we break this precept by enabling a \"lens\" that can change its depth of field arbitrarily? This work investigates the design and implementation of such a computational lens with spatially-selective focusing. Our design uses an optical arrangement of a Lohmann lens and a phase-only spatial light modulator to allow each pixel to focus at a different depth. We extend classical autofocusing techniques to the spatially-varying scenario where the depth map is iteratively estimated using contrast and disparity cues, enabling the camera to progressively shape its depth-of-field to the scene's depth. By obtaining an all-in-focus image optically, our technique advances upon prior work in two key aspects: the ability to bring an entire scene in focus simultaneously, and the ability to maintain the highest possible spatial resolution.", + "bbox": [ + 88, + 554, + 485, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 843, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Core to any imaging system is the lens: an optical component designed to gather rays of light from a scene and", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "form a focused image on a sensor. The focusing ability of a lens, however, only applies to a single plane in the scene. To form a focused image, the subject should be positioned at the focal plane, i.e., at a fixed depth from the camera. Points that do not lie on this focal plane appear blurry, and the amount of defocus increases progressively as the points move further away from the plane.", + "bbox": [ + 511, + 523, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reducing the size of a lens' aperture decreases the amount of defocus and increases the depth of field (i.e., the region near the focal plane where points appear to be in sharp focus). However, this comes at the cost of reduced light throughput. Furthermore, smaller apertures increase diffractive blur [8], making content within the depth of field less sharp. These restrictions on focus are attributed to the traditional design of a lens, which offers the ability to move the focal plane (e.g., by adjusting a focus ring) but maintains its shape. Hence, we raise the question: in place of a single focal plane, is it possible to optically program a focal surface that can adapt to any scene geometry?", + "bbox": [ + 511, + 628, + 908, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper advances the design and implementation of a computational lens capable of spatially-varying focus—one that allows a scene in its entirety to be simultaneously in focus on an image sensor even when the scene is highly non-planar. Our approach relies on adapting the so-called Lohmann lens [21], which is a focus-tunable lens produced", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24645", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/84f65a6cfc7d0ee4e679af6ada1641d36befcb5dc7f4c030eff1713f564e4da3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
optical sharpness# of images requiredall-in-focus generationoutputs depth
small aperturelowoneopticalno
cubic phase plate [9]lowonedeconvolutionno
focal sweep [16, 22]lowonedeconvolutionno
focal stack [23, 35]highmanycontrast metricyes
coded aperture [7, 11, 17, 18, 26, 27, 37, 41]lowonedepth-dependent deconv.yes
light field cameras [4, 10, 19, 24, 40]lowonecontrast metricyes
dual-pixel image deblurring [1, 2, 38, 42]lowonehard inverse problemyes
spatially-varying autofocus (ours)hightwoopticalyes
", + "bbox": [ + 101, + 88, + 893, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of all-in-focus imaging techniques. Optical sharpness: Most methods either use a small effective aperture (increasing the amount of diffraction blur), or intentionally blur the photos (e.g., to create depth-invariant blur). Our approach forms all-in-focus images by bringing each scene point into focus optically, while maintaining a large aperture. # of images required: Our method requires at least one image to approximate the scene geometry, and a second image to form the all-in-focus image. Moreover, our method is well suited for dynamic settings, where each frame determines the focus for the next frame. All-in-focus generation: Unlike most techniques, our approach forms images using an all-optical process; no additional computational post-processing is required. Outputs depth: A useful byproduct of several methods is the ability to recover a scene's depth map.", + "bbox": [ + 89, + 233, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by relative movement between two cubic lenses. Prior work by Qin et al. [28], in the context of near-eye virtual reality (VR) displays, has shown that a rearrangement of the Lohmann lens, along with the use of a phase-only spatial light modulator (SLM), can control the perceived depth of pixels on a display. Our work extends this concept to the imaging scenario to provide unprecedented control over a camera's focusing capabilities, and introduces a novel category to the solution space for all-in-focus imaging.", + "bbox": [ + 88, + 347, + 480, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. This paper proposes a programmable, spatially-varying lens for optical all-in-focus imaging and flexible depth-of-field manipulation.", + "bbox": [ + 89, + 494, + 482, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical all-in-focus (AIF) image. The centerpiece of our contribution is the ability to acquire an optical AIF image given knowledge of the scene depth, i.e., our technique uses knowledge of the depth to resolve all scene points in sharp focus on the sensor. A hallmark of this result is that, unlike prior work in all-in-focus imaging, our imaging process does not require computational post-processing.", + "bbox": [ + 89, + 547, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spatially-varying autofocusing. To recover the depth map of the scene, we extend traditional ideas in contrast- and phase-based autofocus to their spatially-varying counterparts. We show that our system can progressively bring the entire scene into focus using as few as two images. This has the benefit of matching the quality and depth resolution of the focus stacking without requiring a large image set.", + "bbox": [ + 89, + 662, + 482, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Code, datasets, and real-time video demonstrations of dynamic scenes are available on our project website: https://imaging.cs.cmu.edu/svaf [29].", + "bbox": [ + 89, + 768, + 482, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Limitation. Our current optical prototype is light inefficient due to the use of polarization-based phase modulation and a beamsplitter; note that at most one-eighth of the incident light reaches the sensor, due to light passing through a polarizer once and through a 50/50 beamsplitter twice.", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly discuss techniques used for extended depth-of-field imaging, with a focus on key differences from our proposed approach, as summarized in Tab. 1.", + "bbox": [ + 511, + 371, + 903, + 416 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Coded aperture systems. One of the classic problems in computer vision is that of depth from defocus. For a traditional lens, this problem is ill-conditioned since the circular shape of the defocus kernel is not sufficiently discriminative. To resolve this, coded-aperture systems reshape the aperture of the lens using amplitude or phase masks. A seminal work in this space is that of Dowski and Cathey [9], who show that a cubic phase mask has a depth-invariant blur kernel that can be used to deblur and obtain the AIF image. More recent work have concentrated on enabling better discriminability of depth, either by using amplitude apertures [17, 37] or phase-based ones [7, 11, 18, 26, 27, 41].", + "bbox": [ + 511, + 424, + 905, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Also loosely falling under the broad umbrella of coded aperture system are dual-pixel (DP) sensors, where each pixel has two sub-pixels under a single microlenslet. The resulting system acquires two images simultaneously, each from different halves of the lens aperture, thereby emulating a small baseline stereo setup. DP sensors have found extensive use in autofocusing systems [13]; however, their ability to provide stereo pairs has also enabled their use in AIF image and depth estimation [1, 2, 38, 42].", + "bbox": [ + 511, + 606, + 905, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed work differs from this class of techniques in two distinct ways. First, we do not need any computational post-processing since we optically form an all-in-focus image on the sensor. Second, since we optically focus on scene points, our images are sharper, modulo non-idealities of the optics.", + "bbox": [ + 511, + 742, + 903, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Focus stacking. One of the ways to regularize the depth-from-defocus problem is to capture a dense focal stack—that is, a collection of images obtained by sweeping the focus plane through the scene. Using local contrast as a cue,", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24646", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we can estimate depth and construct an all-in-focus image of the scene [23, 35]. However, this technique is slow, due to the need to capture a collection of images, and does not handle scene dynamics well.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Focal sweep and flexible depth-of-field imaging. Nagahara et al. [22] explore the idea of capturing a single image, where the focal plane changes throughout its exposure. One of their results is the construction of a near-invariant defocus kernel by sweeping focus, linearly in diopter space, within the exposure; similar to Dowski and Cathey's work, this creates a depth-invariant blur that can be computationally removed. A different result enables a flexible and nonplanar depth of field by synchronizing the focus plane to a rolling shutter [16]. Our work goes beyond these results by avoiding post-process computation, as well as enabling a freeform shape for the depth of field. Such control of the depth of field was previously only possible via post-processing, for example, a focal stack [12, 31, 44].", + "bbox": [ + 89, + 167, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Light fields. Light field cameras [4, 10, 19] sample the incident 4D space of light rays, and enables refocusing and AIF image synthesis to be done as a post-processing operation. Light fields can be measured by placing a microlens array in front of a sensor [24], or using a camera array [40], both of which sacrifice the spatial resolution of the sensors to increase angular resolution. A high angular resolution is critical for better depth selectivity, but this comes at a commensurate loss in spatial resolution for the reconstructed images.", + "bbox": [ + 89, + 395, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autofocusing systems. Our spatially-selective focusing technique revisits classical ideas in passive autofocusing to estimate the depth map of the scene, in order to obtain an all-in-focus image. Specifically, we develop spatially-varying counterparts to the traditional contrast and phase detection autofocusing techniques.", + "bbox": [ + 89, + 563, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrast detection autofocus (CDAF) is one of the primary methods used by digital cameras to focus a lens. The approach involves adjusting the focus settings of the lens until the camera detects the highest contrast (usually at a few select locations). CDAF techniques can be sped up using various hill-climbing techniques [15, 33, 36] to estimate the peak contrast using a sparser set of photographs.", + "bbox": [ + 89, + 655, + 483, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Phase detection autofocus (PDAF) is an alternate technique that is commonly available in cameras with a DP sensor. When a scene point is in focus, the two images captured by the corresponding sub-pixels match. Otherwise, disparity is introduced between the two views. The signed disparity determines the lens focus for the scene point.", + "bbox": [ + 89, + 763, + 483, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we extend both techniques to our spatially-selective framework, where we apply them to bring all regions of the scene into focus simultaneously.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8857848adaaab66b828af5a80ed2c60f77d19f3f2840e1ae8c1dbf4a61b9994c.jpg", + "image_caption": [ + "Figure 2. Split-Lohmann computational lens expands upon the Lohmann lens in two key steps. (a) First, it achieves a programmable focus-tunable lens by collocating the two cubic phase plates of the Lohmann lens using a 4f relay and placing a phase ramp at the Fourier plane whose slope controls the effective focal length. (b) We can optically collocate an image sensor with the SLM. Since each pixel on the sensor is resolved on the SLM, dispalying a spatially-varying phase ramp allows local focus control." + ], + "image_footnote": [], + "bbox": [ + 535, + 87, + 885, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Split-Lohmann Computational Lens", + "text_level": 1, + "bbox": [ + 511, + 417, + 879, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ideas of this paper are inspired by a recent result on the design of a VR display [28], which proposes the Split-Lohmann lens, a computational lens that can spatially vary the focal length. For completeness, we briefly discuss this prior work before delineating our specific contributions.", + "bbox": [ + 511, + 446, + 905, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Split-Lohmann design relies on a specific kind of focus-tunable lens called a Lohmann or Alvarez lens [5, 21] consisting of two translating cubic phase plates. Suppose that the optical profiles of the two cubic plates are given by $h_1(x) = \\kappa x^3$ and $h_2(x) = -\\kappa x^3$ , where $\\kappa$ is a curvature-related parameter. When stacked together with a lateral offset $\\Delta$ , the resulting phase modulation is given as", + "bbox": [ + 511, + 526, + 905, + 632 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {1} (x + \\Delta) + h _ {2} (x - \\Delta) = \\kappa \\left(6 \\Delta x ^ {2} + 2 \\Delta^ {3}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 650, + 903, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ignoring the constant term that is independent of $x$ , we get a phase modulation that is quadratic in $x$ , i.e., a lens whose focal length is inversely proportional to $\\Delta$ . Hence, lenses of different focal lengths can be obtained by changing the amount of translation between the cubic phase plates (see Fig. 2). The Split-Lohmann display advances the Lohmann lens by first removing the mechanical translation required, and second, achieving independent local control of focal length for different regions on an OLED display.", + "bbox": [ + 511, + 686, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose inverting the function of this optical system, by replacing the OLED display with a camera sensor and adding a camera lens. The result is a Split-Lohmann computational lens that now offers a camera the ability to spatially-vary its focus.", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24647", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Spatially-Varying Autofocus", + "text_level": 1, + "bbox": [ + 89, + 89, + 352, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Programming a Split-Lohmann computational lens to form all-in-focus images requires solving a spatially-varying autofocusing problem. While we can use a second device devoted to depth estimation—either in the form of a passive stereo camera, a structured light 3D scanner, or a time-of-flight device—we consider a self-contained autofocus loop where our device progressively estimates depth and revises its focus setting. We take inspiration from existing AF solutions used in conventional cameras to drive lens focus: Contrast Detection Autofocus (CDAF) and Phase Detection Autofocus (PDAF). Our goal here is to design the spatially-varying counterparts to these techniques, where we recover a depth image that is used to compute the necessary SLM pattern to perform simultaneous local autofocus and bring the entire field of view in focus simultaneously.", + "bbox": [ + 89, + 114, + 483, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Contrast Detection Autofocus (CDAF)", + "text_level": 1, + "bbox": [ + 89, + 349, + 419, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrast is a popular focus metric for driving focus, as it can be readily implemented on most imaging systems. The basic premise of this approach relies on the observation that contrast of a local region (say, a patch) is maximized when it is in focus. CDAF techniques, hence, search for the focus setting that maximizes the contrast of a pre-determined patch. We extend CDAF to its spatially-varying counterpart by identifying an independent focus parameter for every region that maximizes its image contrast.", + "bbox": [ + 88, + 372, + 482, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach relies on the insight that contrast of a patch as a function of depth is often smooth and, more importantly, unimodal. This property allows us to design an efficient search strategy by progressively narrowing down the focus/depth range where the mode can lie. Suppose that the total working range of focus, as measured in diopters (the reciprocal of depth), is from 0 to $W$ diopters. We obtain three images that correspond to the focus at $\\frac{W}{4}$ , $\\frac{W}{2}$ , and $\\frac{3W}{4}$ diopters. We can estimate the contrast at each local patch across these three images, and the focus which has the maximum contrast (across the three images) allows us to reduce the range centering the contrast maxima. For example, if the maximum value occurs at $\\frac{W}{4}$ , then the true maxima must be between the 0 and $\\frac{W}{2}$ diopters. In each iteration, the search range reduces by half, and we can repeat the technique over the refined focus range. Note that, this procedure happens in parallel and independently at each patch or region in the image, enabled by the spatially-selective focusing capability of our device. Since each iteration comprises of taking three images spanning a reduced search range, we obtain the same performance of linear search, but with a number of images that is logarithmic in comparison.", + "bbox": [ + 89, + 508, + 483, + 843 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/faeca5e9a5d791837bd2cc063a60db175c8c882578a834b97bd2d9af49f951dc.jpg", + "image_caption": [ + "(a) The search algorithm for local peak contrast" + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 732, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b6e6839ea79e222d539abf757b03252252257c6d04c57db4c0caea8fadc72744.jpg", + "image_caption": [ + "(b) Example focus search images" + ], + "image_footnote": [], + "bbox": [ + 738, + 88, + 905, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db89c2ac042817ed36a76b0c658a7dba1359574d73299299ebe61f8489440de1.jpg", + "image_caption": [ + "(c) Search results at increasing steps", + "Figure 3. Contrast-based autofocusing pipeline. (a) The objective is to identify a per-superpixel focus (or depth) setting that maximizes contrast. At step $K$ , we capture three measurements and identify the focus setting that maximizes contrast. We then refine the search around the next best candidate focus. (b) Example focus searches from the first two iterations. (c) Focus (or depth) maps and corresponding images captured after $K$ iterations." + ], + "image_footnote": [], + "bbox": [ + 733, + 185, + 903, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Patching strategy.** A key challenge in the technique described above is the definition of the local region or patch where the search technique maximizes contrast. In particular, we would like each patch to have minimal depth variations and especially avoid depth discontinuities, so that a single focus setting suffices for all pixels in the patch. We achieve this by performing superpixel segmentation; since depth edges invariably align with texture edges, this strategy of patching based on superpixels allows us to avoid having large depth variations or discontinuities inside a superpixel. We compute the superpixels using the k-means clustering-based SLIC method [3]. We update the superpixels after each autofocus iteration to benefit from the improved sharpness in the captured photographs, which in turn allows for finer segmentation of depth boundaries. When the updated superpixel contains pixels previously assigned to different depths, we assign the new depth to the most common (contrast-maximizing) depth in the superpixel.", + "bbox": [ + 511, + 449, + 906, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate our spatially-varying CDAF algorithm in Fig. 3. We also provide pseudocode in the supplemental pdf. Contrast-based autofocusing often requires multiple captures to identify the focus plane with the highest contrast. After establishing focus, one can then refine the existing focal surface over time to accommodate dynamic scenes, which is far less expensive than recapturing a full focal stack. In contrast, the technique we present next, PDAF, offers a single-shot approach to determining focus, at the cost of requiring a specialized sensor that uses dual pixel autofocus technology.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1In reality, two images suffice as the center focus in an iteration is one of the three focus settings used in a previous iteration; however, the specific patching strategy that we adopted complicates the reuse of previously captured images, thereby requiring a recapture of the center focus.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24648", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Phase Detection Autofocus (PDAF)", + "text_level": 1, + "bbox": [ + 89, + 90, + 393, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent advancements in DP sensors enabled Phase Detection Autofocus (PDAF) in consumer cameras and smartphones. A DP sensor consists of an array of microlenses, with two (or more) photodiodes under each lens. These photodiodes capture the light traveling through different parts of the aperture, producing a stereo image pair. The disparity observed across the two images determines the (signed) distance of the scene point from the focal plane (see Fig. 14 in supplemental pdf). When a scene point is sharply focused on the sensor, there is no disparity. When a scene point is out of focus, i.e., focused onto a plane in front of or behind the sensor, the disparity appears directionally proportional to the direction and depth of the focused plane. To solve for the signed disparity, we adapt the conjugate gradient-based optical flow solver by Liu [20]. From the magnitude and direction of optical flow, we can calculate the focus correction needed to bring each region into focus, thereby enabling a single-shot approach to autofocus the entire field of view.", + "bbox": [ + 89, + 114, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Challenges at depth boundaries. Since disparity in a stereo configuration is always horizontal, this suggests that much of our disparity estimation is driven by pixels with strong vertical gradients. However, in the context of focusing, we would also need to associate this disparity to the two regions on either side of the vertical gradient. As an example, suppose that we observe an edge that exhibits a disparity of a few pixels. When this edge corresponds to a depth boundary (which often have strong texture gradients), all we can observe is the relative shifts between the patches on either side of the boundary, and we cannot—just with local context—identify the exact shifts for each. To resolve this, we resort to segmenting the scene into semantically meaningful regions, and computing disparity for each region in isolation. In particular, we segment the total DP image using SegmentAnything [14] into masked labeled regions (layers), independently compute the masked optical flow for each layer, and sum all layers into a result flow map. The result flow map is used to drive the next autofocus. We illustrate this layered optical flow algorithm in Fig. 4 and detail the pseudocode in the supplemental pdf.", + "bbox": [ + 89, + 412, + 483, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PDAF has many desirable properties. Unlike CDAF, which requires a search procedure to identify the sharpest focus setting, PDAF requires a single image to identify the spatially-varying focus map; this allows it to adapt to scene dynamics and is less likely to get stuck in local minima. When the working range is large, we can expect significant defocus in regions with larger disparity, which degrades the quality of the disparity estimate; even in this case, as long as the sign of the disparity is accurate, the technique converges within a few steps. All of these benefits come with the requirement of a specialized sensor that has dual pixels.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9ab1b9a0d7ed0f75c4d783dbce9e5c19adc703a71415562cae43cdc5e7979552.jpg", + "image_caption": [ + "(a) Example phase-based AIF captures" + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 906, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77411819968f21612ca975cfe507146fb0c271b2573800e7b980f0c9f3588599.jpg", + "image_caption": [ + "(b) Phase-based autofocus algorithm", + "Figure 4. Phase-based autofocus pipeline. (a) Example 1st and 4th captures from phase-based autofocus. (b) The algorithm starts by capturing an initial DP image, and segments the result into semantically meaningful layers. For each layer, we use optical flow to compute a smooth flow map, that indicates how to adjust the focus settings. After updating focus, we capture another DP image and repeat the process." + ], + "image_footnote": [], + "bbox": [ + 516, + 171, + 903, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 465, + 599, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prototype. We show our prototype in Fig. 5. The Point Spread Function (PSF) of the prototype, shown in Fig. 6, is captured for a dot placed at varying distances spanning the entire tilting range we use of the SLM. To obtain DP images, we use the Canon R10 camera sensor with a $6000 \\times 4000$ resolution and a $3.72\\mu \\mathrm{m}$ pixel pitch. The sensor is a dual-pixel sensor which we use for our spatially-varying PDAF algorithm. We use gphoto2 commands to programmatically acquire raw images and the LibRaw library to extract the DP views from the raw image. To perform spatially-varying autofocusing, we use the Holoeye GAEA-2 Phase-Only SLM for phase modulation, which has a pixel pitch of $\\delta_{\\mathrm{SLM}} = 3.74\\mu \\mathrm{m}$ and a resolution of $4160 \\times 2464$ . The cubic phase plate is custom fabricated using subtractive laser etching. The relay lenses consist of three Samyang $85\\mathrm{mm}$ f/1.4 AS lenses. Our objective lens is a AF-S DX Micro NIKKOR $40\\mathrm{mm}$ f/2.8G.", + "bbox": [ + 511, + 491, + 908, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Freeform depth-of-field photography. The ability to spatially-vary focus opens up a lot of composition capabilities beyond all-in-focus photography. After obtaining a depth map, our system can also intentionally add freeform defocus to the scene. For example, we can perform tilt-shift photography without requiring a Scheimpflug adapter to tilt the plane of focus, as seen in Fig. 7. We also show in Fig. 7 the ability to selectively focus on isolated regions at different depths, while the rest of the scene remains defocused.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24649", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a5ccc80dc3623a754b62ca9427c7865b82f7dc92db6f35e1fd631c7a8e675cd.jpg", + "image_caption": [ + "Figure 5. Our prototype camera (see supplemental pdf for details)." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 473, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5154c4617d83d8c71742e066e617b44bef523c602976d9cdebf4293260ab6f76.jpg", + "image_caption": [ + "Figure 6. Point-spread functions of our prototype camera. Left: Images of a point placed at different distances from the camera, created by puncturing black paper with a needle and back-illuminating it with full-spectrum white light. The dot is 10-by-10 pixels, or $37.2\\mu \\mathrm{m}$ by $-37.2\\mu \\mathrm{m}$ on the sensor. We label the distance of each dot in diopters. Right: Focused images of the dot, captured with an SLM spatial frequency $v$ . Bottom: Focused images of multiple dots at $v = 0.09$ spanning the camera's FOV (see supplement for zoom-ins)." + ], + "image_footnote": [], + "bbox": [ + 94, + 262, + 480, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another application of our freeform depth-of-field selection can be seen in Fig. 8, where we suppress the presence of thin structures from a photo by optically defocusing only those particular pixels. We achieve this by selecting a focus that is far from its depth.", + "bbox": [ + 89, + 554, + 483, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All-in-focus (AIF) imaging. We capture AIF results for six scenes: Adventure, Planes, Flowers, Bunny, Ship, and Rainbow. The full results gallery is in the supplemental pdf.", + "bbox": [ + 89, + 643, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Adventure and Rainbow, we provide a quantitative analysis of AIF methods in Fig. 11 using PSNR and SSIM [39] as metrics. To obtain the ground truth target, we capture a dense focal stack with 69 depth planes, with each focus setting having 2 repeated captures averaged to reduce noise. We then combine these images to form a computational AIF image from this focal stack to serve as our ground truth. Given the recovered depth map, we also program our lens to capture an AIF image optically, shown in our qualitative comparisons. When evaluating all-in-focus results with respect to number of photos captured, we observe phase-based autofocusing having the best performance, followed by both contrast-based autofocus and capturing a focal stack. For focal sweep, we search for the", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abf165e3e8754c10d699c61c4dfdca38ebe77ee0af2caced0f67daa14b08dd0c.jpg", + "image_caption": [ + "Figure 7. Example freeform depth-of-field captures of Arc de Triomphé displayed on a vertically tilted OLED. We show the capability of capturing an optical AIF image (top right), Scheimpflug-focusing to scale the defocus horizontally (bottom left), and selective focusing where a user specifies select regions to be in focus while all other regions are defocused (bottom right)." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 342 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36d974a3d6e5a67f2bf94a6eedd87922d8f432f9f76ad8d3085d314e129618c5.jpg", + "image_caption": [ + "Figure 8. A thin structure removal example using freeform depth-of-field. The character stands in front of a thin wire mesh, far away from the lions print background. Left: A conventional capture shows the character in focus but with a visible wire mesh over the background. Right: Our proposed prototype optically removes the wire mesh by focusing on the background. The large defocus blur of the wire mesh, therefore, makes it hardly visible." + ], + "image_footnote": [], + "bbox": [ + 514, + 441, + 903, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "regularization parameter that produces the best result.", + "bbox": [ + 511, + 684, + 867, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all scenes, we show qualitative results for contrast-based after 3 steps (10 photos) and phase-based after 3 steps (4 photos). We also perform focal sweep (1 photo) and reconstruct AIF images via focus stacking both computationally and optically (20 photos). For the Planes scene, we also capture a small aperture photo with a separate camera and a long exposure ( $f/36$ with a $55\\mathrm{mm}$ lens on Nikon Z5).", + "bbox": [ + 511, + 699, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial resolution. We assess the spatial resolution of our imaging prototype, and the difference in performance across autofocus methods. Fig. 12 compares the modulation transfer function (MTF) of our spatially-varying autofocus methods to focal stack techniques and focal sweep. Both of our phase- and contrast-based autofocus methods show", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24650", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2ecba02bc49dbfa6c47f0c4890e926ae85b28fdfc6c0dcc5145deb50ba3ef62.jpg", + "image_caption": [ + "Initial photo" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 367, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2153fcf5f52749e18b460d5c5a0340a569d18cc993dc1faf655555a8cc34493.jpg", + "image_caption": [ + "Contrast-based result (ours)" + ], + "image_footnote": [], + "bbox": [ + 372, + 88, + 630, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf605895f5924117f90a338828b5d0a34f440648e6cf0e58b17a0adf3fb69cfb.jpg", + "image_caption": [ + "Phase-based result (ours)" + ], + "image_footnote": [], + "bbox": [ + 632, + 88, + 890, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ada5c7bd7047640ab6b2475d5096b209396d0184f4e92872b4b743b8ca90a7a8.jpg", + "image_caption": [ + "Small aperture $(f / 36)$" + ], + "image_footnote": [], + "bbox": [ + 109, + 268, + 367, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/856ad0cbfee069556b1cdf80888f64fc5bbea04958c7dc8202a72074ac4c9792.jpg", + "image_caption": [ + "Focal stack (computational)" + ], + "image_footnote": [], + "bbox": [ + 372, + 268, + 629, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42a33fd30ffea5fbd188263f718bfb11b07e377baa6592863f8b45cc0718a75c.jpg", + "image_caption": [ + "Focal stack (optical)" + ], + "image_footnote": [], + "bbox": [ + 632, + 268, + 890, + 430 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11ef3ec0bb66d8d0fce2f5a79424f890aa521a9301377fb0ddf4876efa730e0f.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison for the Planes scene. Each recovered depth map is shown in the top-left corner. All insets are blurrier for small aperture $(f / 36)$ when compared to our results due to diffraction (orange inset) and extreme depth range (green and blue insets)." + ], + "image_footnote": [], + "bbox": [ + 107, + 500, + 367, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/570e582d88cd00cc17568d4654321988ba21ab8ae72d9d35377dc473b9bc0465.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 614, + 367, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d5ad358656cb56c027235ae3e9908cfab4e79cccadeaf0d8202aaab59d17c22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 500, + 629, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4220e1ec342102e9767713aab14920d57ca4c7d85f3a35fe8956e69dd1cc0b44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 614, + 629, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e444246d41a81abfb85f81cf0aec1c5268a336d376f4167abb5d6e08776a6964.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 500, + 890, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16cd2906cad6fe314b30ca96de535aabed17eb639b8ae2fba69206b48a9dcda2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 614, + 890, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/996340d54d21d219fca076d954091b3adb3b1ffe953204d138d2ad1857572ac5.jpg", + "image_caption": [ + "Initial photo" + ], + "image_footnote": [], + "bbox": [ + 109, + 679, + 367, + 792 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b35d6cea0d20f956381655284a3c3cc53e49facc3685428a9ece0c255ecac533.jpg", + "image_caption": [ + "Focal sweep" + ], + "image_footnote": [], + "bbox": [ + 109, + 794, + 367, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f057cba3c9019c2f0aaa46abe0822852c98a64035d8b45fea58fcad072ca0db.jpg", + "image_caption": [ + "Contrast-based result (ours)" + ], + "image_footnote": [], + "bbox": [ + 372, + 679, + 629, + 792 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e54d5d946ec94a5a97c07710ec572b78206a56ab47fe9f02a7454b1e09b155f.jpg", + "image_caption": [ + "Focal stack (computational)", + "Figure 10. Qualitative comparison for the Flowers scene. We include a focal sweep photo (does not output depth) for additional comparison." + ], + "image_footnote": [], + "bbox": [ + 372, + 794, + 629, + 842 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f015d1cdd7b9ef3064fa8e4422646f4d719dd47cf5e8c692bb353697960704e6.jpg", + "image_caption": [ + "Phase-based result (ours)" + ], + "image_footnote": [], + "bbox": [ + 632, + 679, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37631c4454f8e2bfacf941f6b07dab0dff0b3eb30c0311a78b5567a80e7dd7a7.jpg", + "image_caption": [ + "Focal stack (optical)" + ], + "image_footnote": [], + "bbox": [ + 632, + 794, + 890, + 842 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24651", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc8b4b71dbbf2cf0de4c3ec11b1b9b62b70a48fc999a56148c2d13a4c0277180.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 94, + 478, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02f9015b1f4fd9d303642784b5fdf1143500bc6c6416daaabaf66b589d0d2b8f.jpg", + "image_caption": [ + "Figure 11. Quantitative analysis of all-in-focus imaging methods, evaluated as a function of number of photos. These plots represent the average performance over the Adventure and Rainbow scenes." + ], + "image_footnote": [], + "bbox": [ + 94, + 205, + 478, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparable performance with focal stack techniques, while focal sweep has consistently lower performance.", + "bbox": [ + 89, + 378, + 483, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 424, + 205, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aperture. We estimate that our prototype imaging system has a rest state f number of $f / 6.8$ (i.e., the slope of the phase ramp is 0). When a phase ramp is applied to the SLM, the cubic phase plates are optically translated relative to one another, resulting in a smaller effective aperture. In other words, as our system moves the focus away from the nominal plane, this changes the shape of the effective aperture (impacting the bokeh) and reduces its area (limiting the total amount of light transmitted through the system). At the most extreme focus settings, we calculate that light throughput is $76\\%$ when compared to the system in the rest state. See the supplemental materials for additional details.", + "bbox": [ + 88, + 450, + 482, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is important to note that the f-number depends on both the diameter and curvature of the cubic phase plate, among other optics. In principle, we can increase the light throughput by using a cubic phase plate with a large diameter.", + "bbox": [ + 89, + 632, + 483, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aberration correction. As with the continual evolution of optical systems from singlets to achromatic doublets, apochromatic designs, and modern computer-optimized multi-element systems, the optical design of our proposed system can be improved to correct for aberrations. This includes the following: (i) redesign the lenses used in our system in an end-to-end fashion, and (ii) optimize the cubic phase plate profile. For the former, Sitzmann et al. [32] and Sun et al. [34] have shown that end-to-end optical system design allows aberrations to be fixed with a higher degree of freedom. Such methods suggest that lenses can be jointly optimized with parametric SLM patterns to improve all axes of performance of the imaging system: chro", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/915a90bf4c4f9ae159251b16987d8a34d377d710a717cfb38e01d1791b2777ad.jpg", + "image_caption": [ + "(a) Images captured using different AIF methods" + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 890, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0e1164fed1817b71b8d1b110784b17809e47d96a2d2689ea14fff852f0d3e71.jpg", + "image_caption": [ + "(b) MTF curves comparing different AIF methods", + "Figure 12. MTF of our prototype. (a) We arranged printed USAF targets at three different depths and illuminated them with full-spectrum white light, and computed AIF images for different methods. (b) Each MTF is computed and averaged from both horizontal and vertical line pairs across the three depths. We observe that our methods, including phase-based, contrast-based, and focus stack optical, perform consistently comparable to the focus stack computational AIF method, while focal sweep performs consistently worse than all of them." + ], + "image_footnote": [], + "bbox": [ + 531, + 262, + 890, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "matic focal shift, artifacts at depth discontinuities, and f-number. For the latter, one can potentially optimize for a hybrid refractive-diffractive phase plate using the fully differentiable hybrid ray-tracing and wave-propagation (ray-wave) model recently proposed by Yang et al. [43], which significantly improved aberration correction.", + "bbox": [ + 511, + 617, + 906, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a first-of-its-kind imaging technique, one that provides spatially-varying autofocusing capabilities. The proposed imaging system and the autofocusing algorithms can be interpreted as an optical optimizer of image contrast; we show that standard autofocusing techniques based on contrast and phase can be readily adopted to the spatially-varying setting. In general, we believe that this novel approach to imaging has widespread applications where focus is of paramount importance, such as long-range surveillance, machine vision, and microscopy.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24652", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the Air Force Office of Scientific Research (FA 95502410244), a NSF CAREER award (IIS 2238485), and a James Sprague Presidential Fellowship.", + "bbox": [ + 89, + 90, + 485, + 154 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 162, + 187, + 179 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In ECCV, 2020. 2", + "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[3] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11):2274-2282, 2012. 4", + "[4] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 2, 3", + "[5] Luis W Alvarez and William E Humphrey. Variable-power lens and system, 1970. US Patent 3,507,565. 3", + "[6] Terry A Bartlett, William C McDonald, and James N Hall. Adapting texas instruments dlp technology to demonstrate a phase spatial light modulator. In Emerging Digital Micromirror Device Based Systems and Applications XI, pages 161-173. SPIE, 2019. 14", + "[7] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10193-10202, 2019. 2", + "[8] Julie Chang, Isaac Kauvar, Xuemei Hu, and Gordon Wetzstein. Variable aperture light field photography: overcoming the diffraction-limited spatio-angular resolution tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3737-3745, 2016. 1", + "[9] Edward R Dowski and W Thomas Cathey. Extended depth of field through wave-front coding. Applied optics, 34(11): 1859-1866, 1995. 2", + "[10] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2, 3", + "[11] Adam Greengard, Yoav Y Schechner, and Rafael Piestun. Depth from diffracted rotation. Optics letters, 31(2):181-183, 2006. 2", + "[12] David E Jacobs, Jongmin Baek, and Marc Levoy. Focal stack compositing for depth of field control. Stanford Computer Graphics Laboratory Technical Report, 1(1):2012, 2012. 3", + "[13] Jinbeum Jang, Yoonjong Yoo, Jongheon Kim, and Joonki Paik. Sensor-based auto-focusing system using multi-scale feature extraction and phase correlation matching. Sensors, 15(3):5747-5762, 2015. 2", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and" + ], + "bbox": [ + 93, + 188, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 5, 14", + "[15] Eric P Krotkov. Active computer vision by cooperative focus and stereo. Springer Science & Business Media, 2012. 3", + "[16] Sujit Kuthirummal, Hajime Nagahara, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):58-71, 2010. 2, 3", + "[17] Anat Levin, Rob Fergus, Frédo Durand, and William T Freeman. Image and depth from a conventional camera with a coded aperture. ACM transactions on graphics (TOG), 26 (3):70-es, 2007. 2", + "[18] Anat Levin, Samuel W Hasinoff, Paul Green, Frédo Durand, and William T Freeman. 4d frequency analysis of computational cameras for depth of field extension. ACM Transactions on Graphics (TOG), 28(3):1-14, 2009. 2", + "[19] Marc Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2, 3", + "[20] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, USA, 2009. AAI0822221. 5, 14, 15", + "[21] Adolf W Lohmann. A new class of varifocal lenses. Applied Optics, 9(7):1669-1671, 1970. 1, 3", + "[22] Hajime Nagahara, Sujit Kuthirummal, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. In European Conference on Computer Vision, pages 60-73. Springer, 2008. 2, 3", + "[23] Shree K Nayar and Yasuo Nakagawa. Shape from focus. IEEE Transactions on Pattern analysis and machine intelligence, 16(8):824-831, 1994. 2, 3", + "[24] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 3", + "[25] Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning features by watching objects move. In CVPR, 2017. 14", + "[26] Sri Rama Prasanna Pavani and Rafael Piestun. Three dimensional tracking of fluorescent microparticles using a photon-limited double-helix response system. Optics express, 16 (26):22048-22057, 2008. 2", + "[27] Sri Rama Prasanna Pavani, Michael A Thompson, Julie S Biteen, Samuel J Lord, Na Liu, Robert J Twieg, Rafael Piestun, and William E Moerner. Three-dimensional, single-molecule fluorescence imaging beyond the diffraction limit by using a double-helix point spread function. Proceedings of the National Academy of Sciences, 106(9):2995-2999, 2009. 2", + "[28] Yingsi Qin, Wei-Yu Chen, Matthew O'Toole, and Aswin C. Sankaranarayanan. Split-lohmann multifocal displays. ACM Trans. Graph., 42(4), 2023. 2, 3, 11, 14", + "[29] Yingsi Qin, Aswin C. Sankaranarayanan, and Matthew O'Toole. Project page: Spatially-varying autofocus. https://imaging.cs.cmu.edu/svaf, 2025.2", + "[30] Daniel J. Reiley. Nikon AF Nikkor 85mm f/1.4D IF Lens prescription design. https://www.lens-designs.com/photographic-primes, 2017. Licensed under the MIT License. 14" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24653", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Parikshit Sakurikar and P. J. Narayanan. Focal stack representation and focus manipulation. In 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR), pages 250–255, 2017. 3", + "[32] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 8, 14", + "[33] Muralidhara Subbarao and J-K Tyan. Selecting the optimal focus measure for autofocusing and depth-from-focus. IEEE transactions on pattern analysis and machine intelligence, 20(8):864-870, 1998. 3", + "[34] Qilin Sun, Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. End-to-end complex lens design with differentiate ray tracing. ACM Trans. Graph., 40(4), 2021. 8, 14", + "[35] Supasorn Suwajanakorn, Carlos Hernandez, and Steven M Seitz. Depth from focus with your mobile phone. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2015. 2, 3", + "[36] Dong-Chen Tsai and Homer H. Chen. Reciprocal focus profile. IEEE Transactions on Image Processing, 21(2):459-468, 2012. 3", + "[37] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69, 2007. 2", + "[38] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2", + "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6", + "[40] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 2, 3", + "[41] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin C. Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — Learning phase masks for passive single view depth estimation. In IEEE Intl. Conf. Computational Photography (ICCP), 2019. 2", + "[42] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 2", + "[43] Xinge Yang, Matheus Souza, Kunyi Wang, Praneeth Chakravarthula, Qiang Fu, and Wolfgang Heidrich. End-to-end hybrid refractive-diffractive lens design with differentiable ray-wave model. In SIGGRAPH Asia 2024 Con" + ], + "bbox": [ + 89, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 8, 14", + "[44] Changyin Zhou, Daniel Miau, and Shree K Nayar. Focal sweep camera for space-time refocusing. 2012. 3" + ], + "bbox": [ + 516, + 92, + 903, + 148 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24654", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_model.json b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a894b8a148f64b5ef8e5bf88add034514fcaf30c --- /dev/null +++ b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.355, + 0.131, + 0.644, + 0.154 + ], + "angle": 0, + "content": "Spatially-Varying Autofocus" + }, + { + "type": "title", + "bbox": [ + 0.265, + 0.181, + 0.737, + 0.219 + ], + "angle": 0, + "content": "Yingsi Qin, Aswin C. Sankaranarayanan, Matthew O'Toole \nCarnegie Mellon University" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.236, + 0.387, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.237, + 0.496, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.237, + 0.794, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.237, + 0.904, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.365, + 0.495, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.434, + 0.424, + 0.446 + ], + "angle": 0, + "content": "(a) Conventional photo and its confined focal plane" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.365, + 0.904, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.433, + 0.902, + 0.446 + ], + "angle": 0, + "content": "(b) All-In-Focus photo and its spatially-varying autofocused focal surface (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.453, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 1. Spatially-varying autofocus to produce an optical all-in-focus image. Left: A conventional photo with a regular lens, where objects at a single focal plane appear sharp. Right: An all-in-focus photo captured through spatially-varying autofocus. To achieve this, we combine (i) a programmable lens with spatially-varying control over focus, and (ii) a spatially-varying autofocus algorithm to drive the focus of this lens. Note that this is an optically-captured image of a real scene with no post-capture processing used." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.522, + 0.327, + 0.538 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.555, + 0.486, + 0.812 + ], + "angle": 0, + "content": "A lens brings a single plane into focus on a planar sensor; hence, parts of the scene that are outside this planar focus plane are resolved under defocus. Can we break this precept by enabling a \"lens\" that can change its depth of field arbitrarily? This work investigates the design and implementation of such a computational lens with spatially-selective focusing. Our design uses an optical arrangement of a Lohmann lens and a phase-only spatial light modulator to allow each pixel to focus at a different depth. We extend classical autofocusing techniques to the spatially-varying scenario where the depth map is iteratively estimated using contrast and disparity cues, enabling the camera to progressively shape its depth-of-field to the scene's depth. By obtaining an all-in-focus image optically, our technique advances upon prior work in two key aspects: the ability to bring an entire scene in focus simultaneously, and the ability to maintain the highest possible spatial resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.844, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Core to any imaging system is the lens: an optical component designed to gather rays of light from a scene and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.629 + ], + "angle": 0, + "content": "form a focused image on a sensor. The focusing ability of a lens, however, only applies to a single plane in the scene. To form a focused image, the subject should be positioned at the focal plane, i.e., at a fixed depth from the camera. Points that do not lie on this focal plane appear blurry, and the amount of defocus increases progressively as the points move further away from the plane." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.811 + ], + "angle": 0, + "content": "Reducing the size of a lens' aperture decreases the amount of defocus and increases the depth of field (i.e., the region near the focal plane where points appear to be in sharp focus). However, this comes at the cost of reduced light throughput. Furthermore, smaller apertures increase diffractive blur [8], making content within the depth of field less sharp. These restrictions on focus are attributed to the traditional design of a lens, which offers the ability to move the focal plane (e.g., by adjusting a focus ring) but maintains its shape. Hence, we raise the question: in place of a single focal plane, is it possible to optically program a focal surface that can adapt to any scene geometry?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "This paper advances the design and implementation of a computational lens capable of spatially-varying focus—one that allows a scene in its entirety to be simultaneously in focus on an image sensor even when the scene is highly non-planar. Our approach relies on adapting the so-called Lohmann lens [21], which is a focus-tunable lens produced" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24645" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.224 + ], + "angle": 0, + "content": "
optical sharpness# of images requiredall-in-focus generationoutputs depth
small aperturelowoneopticalno
cubic phase plate [9]lowonedeconvolutionno
focal sweep [16, 22]lowonedeconvolutionno
focal stack [23, 35]highmanycontrast metricyes
coded aperture [7, 11, 17, 18, 26, 27, 37, 41]lowonedepth-dependent deconv.yes
light field cameras [4, 10, 19, 24, 40]lowonecontrast metricyes
dual-pixel image deblurring [1, 2, 38, 42]lowonehard inverse problemyes
spatially-varying autofocus (ours)hightwoopticalyes
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Table 1. Comparison of all-in-focus imaging techniques. Optical sharpness: Most methods either use a small effective aperture (increasing the amount of diffraction blur), or intentionally blur the photos (e.g., to create depth-invariant blur). Our approach forms all-in-focus images by bringing each scene point into focus optically, while maintaining a large aperture. # of images required: Our method requires at least one image to approximate the scene geometry, and a second image to form the all-in-focus image. Moreover, our method is well suited for dynamic settings, where each frame determines the focus for the next frame. All-in-focus generation: Unlike most techniques, our approach forms images using an all-optical process; no additional computational post-processing is required. Outputs depth: A useful byproduct of several methods is the ability to recover a scene's depth map." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.482, + 0.485 + ], + "angle": 0, + "content": "by relative movement between two cubic lenses. Prior work by Qin et al. [28], in the context of near-eye virtual reality (VR) displays, has shown that a rearrangement of the Lohmann lens, along with the use of a phase-only spatial light modulator (SLM), can control the perceived depth of pixels on a display. Our work extends this concept to the imaging scenario to provide unprecedented control over a camera's focusing capabilities, and introduces a novel category to the solution space for all-in-focus imaging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.542 + ], + "angle": 0, + "content": "Contributions. This paper proposes a programmable, spatially-varying lens for optical all-in-focus imaging and flexible depth-of-field manipulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Optical all-in-focus (AIF) image. The centerpiece of our contribution is the ability to acquire an optical AIF image given knowledge of the scene depth, i.e., our technique uses knowledge of the depth to resolve all scene points in sharp focus on the sensor. A hallmark of this result is that, unlike prior work in all-in-focus imaging, our imaging process does not require computational post-processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Spatially-varying autofocusing. To recover the depth map of the scene, we extend traditional ideas in contrast- and phase-based autofocus to their spatially-varying counterparts. We show that our system can progressively bring the entire scene into focus using as few as two images. This has the benefit of matching the quality and depth resolution of the focus stacking without requiring a large image set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Code, datasets, and real-time video demonstrations of dynamic scenes are available on our project website: https://imaging.cs.cmu.edu/svaf [29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Limitation. Our current optical prototype is light inefficient due to the use of polarization-based phase modulation and a beamsplitter; note that at most one-eighth of the incident light reaches the sensor, due to light passing through a polarizer once and through a 50/50 beamsplitter twice." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.346, + 0.655, + 0.361 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.417 + ], + "angle": 0, + "content": "We briefly discuss techniques used for extended depth-of-field imaging, with a focus on key differences from our proposed approach, as summarized in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.606 + ], + "angle": 0, + "content": "Coded aperture systems. One of the classic problems in computer vision is that of depth from defocus. For a traditional lens, this problem is ill-conditioned since the circular shape of the defocus kernel is not sufficiently discriminative. To resolve this, coded-aperture systems reshape the aperture of the lens using amplitude or phase masks. A seminal work in this space is that of Dowski and Cathey [9], who show that a cubic phase mask has a depth-invariant blur kernel that can be used to deblur and obtain the AIF image. More recent work have concentrated on enabling better discriminability of depth, either by using amplitude apertures [17, 37] or phase-based ones [7, 11, 18, 26, 27, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.741 + ], + "angle": 0, + "content": "Also loosely falling under the broad umbrella of coded aperture system are dual-pixel (DP) sensors, where each pixel has two sub-pixels under a single microlenslet. The resulting system acquires two images simultaneously, each from different halves of the lens aperture, thereby emulating a small baseline stereo setup. DP sensors have found extensive use in autofocusing systems [13]; however, their ability to provide stereo pairs has also enabled their use in AIF image and depth estimation [1, 2, 38, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.905, + 0.833 + ], + "angle": 0, + "content": "The proposed work differs from this class of techniques in two distinct ways. First, we do not need any computational post-processing since we optically form an all-in-focus image on the sensor. Second, since we optically focus on scene points, our images are sharper, modulo non-idealities of the optics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Focus stacking. One of the ways to regularize the depth-from-defocus problem is to capture a dense focal stack—that is, a collection of images obtained by sweeping the focus plane through the scene. Using local contrast as a cue," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24646" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "we can estimate depth and construct an all-in-focus image of the scene [23, 35]. However, this technique is slow, due to the need to capture a collection of images, and does not handle scene dynamics well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.485, + 0.381 + ], + "angle": 0, + "content": "Focal sweep and flexible depth-of-field imaging. Nagahara et al. [22] explore the idea of capturing a single image, where the focal plane changes throughout its exposure. One of their results is the construction of a near-invariant defocus kernel by sweeping focus, linearly in diopter space, within the exposure; similar to Dowski and Cathey's work, this creates a depth-invariant blur that can be computationally removed. A different result enables a flexible and nonplanar depth of field by synchronizing the focus plane to a rolling shutter [16]. Our work goes beyond these results by avoiding post-process computation, as well as enabling a freeform shape for the depth of field. Such control of the depth of field was previously only possible via post-processing, for example, a focal stack [12, 31, 44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.548 + ], + "angle": 0, + "content": "Light fields. Light field cameras [4, 10, 19] sample the incident 4D space of light rays, and enables refocusing and AIF image synthesis to be done as a post-processing operation. Light fields can be measured by placing a microlens array in front of a sensor [24], or using a camera array [40], both of which sacrifice the spatial resolution of the sensors to increase angular resolution. A high angular resolution is critical for better depth selectivity, but this comes at a commensurate loss in spatial resolution for the reconstructed images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Autofocusing systems. Our spatially-selective focusing technique revisits classical ideas in passive autofocusing to estimate the depth map of the scene, in order to obtain an all-in-focus image. Specifically, we develop spatially-varying counterparts to the traditional contrast and phase detection autofocusing techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Contrast detection autofocus (CDAF) is one of the primary methods used by digital cameras to focus a lens. The approach involves adjusting the focus settings of the lens until the camera detects the highest contrast (usually at a few select locations). CDAF techniques can be sped up using various hill-climbing techniques [15, 33, 36] to estimate the peak contrast using a sparser set of photographs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Phase detection autofocus (PDAF) is an alternate technique that is commonly available in cameras with a DP sensor. When a scene point is in focus, the two images captured by the corresponding sub-pixels match. Otherwise, disparity is introduced between the two views. The signed disparity determines the lens focus for the scene point." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we extend both techniques to our spatially-selective framework, where we apply them to bring all regions of the scene into focus simultaneously." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.088, + 0.887, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.287, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Split-Lohmann computational lens expands upon the Lohmann lens in two key steps. (a) First, it achieves a programmable focus-tunable lens by collocating the two cubic phase plates of the Lohmann lens using a 4f relay and placing a phase ramp at the Fourier plane whose slope controls the effective focal length. (b) We can optically collocate an image sensor with the SLM. Since each pixel on the sensor is resolved on the SLM, dispalying a spatially-varying phase ramp allows local focus control." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.419, + 0.88, + 0.438 + ], + "angle": 0, + "content": "3. The Split-Lohmann Computational Lens" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.524 + ], + "angle": 0, + "content": "The ideas of this paper are inspired by a recent result on the design of a VR display [28], which proposes the Split-Lohmann lens, a computational lens that can spatially vary the focal length. For completeness, we briefly discuss this prior work before delineating our specific contributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.633 + ], + "angle": 0, + "content": "The Split-Lohmann design relies on a specific kind of focus-tunable lens called a Lohmann or Alvarez lens [5, 21] consisting of two translating cubic phase plates. Suppose that the optical profiles of the two cubic plates are given by \\( h_1(x) = \\kappa x^3 \\) and \\( h_2(x) = -\\kappa x^3 \\), where \\( \\kappa \\) is a curvature-related parameter. When stacked together with a lateral offset \\( \\Delta \\), the resulting phase modulation is given as" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.651, + 0.905, + 0.67 + ], + "angle": 0, + "content": "\\[\nh _ {1} (x + \\Delta) + h _ {2} (x - \\Delta) = \\kappa \\left(6 \\Delta x ^ {2} + 2 \\Delta^ {3}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.822 + ], + "angle": 0, + "content": "Ignoring the constant term that is independent of \\( x \\), we get a phase modulation that is quadratic in \\( x \\), i.e., a lens whose focal length is inversely proportional to \\( \\Delta \\). Hence, lenses of different focal lengths can be obtained by changing the amount of translation between the cubic phase plates (see Fig. 2). The Split-Lohmann display advances the Lohmann lens by first removing the mechanical translation required, and second, achieving independent local control of focal length for different regions on an OLED display." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We propose inverting the function of this optical system, by replacing the OLED display with a camera sensor and adding a camera lens. The result is a Split-Lohmann computational lens that now offers a camera the ability to spatially-vary its focus." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24647" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.353, + 0.108 + ], + "angle": 0, + "content": "4. Spatially-Varying Autofocus" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.342 + ], + "angle": 0, + "content": "Programming a Split-Lohmann computational lens to form all-in-focus images requires solving a spatially-varying autofocusing problem. While we can use a second device devoted to depth estimation—either in the form of a passive stereo camera, a structured light 3D scanner, or a time-of-flight device—we consider a self-contained autofocus loop where our device progressively estimates depth and revises its focus setting. We take inspiration from existing AF solutions used in conventional cameras to drive lens focus: Contrast Detection Autofocus (CDAF) and Phase Detection Autofocus (PDAF). Our goal here is to design the spatially-varying counterparts to these techniques, where we recover a depth image that is used to compute the necessary SLM pattern to perform simultaneous local autofocus and bring the entire field of view in focus simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.351, + 0.42, + 0.367 + ], + "angle": 0, + "content": "4.1. Contrast Detection Autofocus (CDAF)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.483, + 0.509 + ], + "angle": 0, + "content": "Contrast is a popular focus metric for driving focus, as it can be readily implemented on most imaging systems. The basic premise of this approach relies on the observation that contrast of a local region (say, a patch) is maximized when it is in focus. CDAF techniques, hence, search for the focus setting that maximizes the contrast of a pre-determined patch. We extend CDAF to its spatially-varying counterpart by identifying an independent focus parameter for every region that maximizes its image contrast." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Our approach relies on the insight that contrast of a patch as a function of depth is often smooth and, more importantly, unimodal. This property allows us to design an efficient search strategy by progressively narrowing down the focus/depth range where the mode can lie. Suppose that the total working range of focus, as measured in diopters (the reciprocal of depth), is from 0 to \\( W \\) diopters. We obtain three images that correspond to the focus at \\( \\frac{W}{4} \\), \\( \\frac{W}{2} \\), and \\( \\frac{3W}{4} \\) diopters. We can estimate the contrast at each local patch across these three images, and the focus which has the maximum contrast (across the three images) allows us to reduce the range centering the contrast maxima. For example, if the maximum value occurs at \\( \\frac{W}{4} \\), then the true maxima must be between the 0 and \\( \\frac{W}{2} \\) diopters. In each iteration, the search range reduces by half, and we can repeat the technique over the refined focus range. Note that, this procedure happens in parallel and independently at each patch or region in the image, enabled by the spatially-selective focusing capability of our device. Since each iteration comprises of taking three images spanning a reduced search range, we obtain the same performance of linear search, but with a number of images that is logarithmic in comparison." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.733, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.3, + 0.728, + 0.31 + ], + "angle": 0, + "content": "(a) The search algorithm for local peak contrast" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.089, + 0.906, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.174, + 0.895, + 0.184 + ], + "angle": 0, + "content": "(b) Example focus search images" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.186, + 0.905, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.299, + 0.901, + 0.31 + ], + "angle": 0, + "content": "(c) Search results at increasing steps" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.419 + ], + "angle": 0, + "content": "Figure 3. Contrast-based autofocusing pipeline. (a) The objective is to identify a per-superpixel focus (or depth) setting that maximizes contrast. At step \\( K \\), we capture three measurements and identify the focus setting that maximizes contrast. We then refine the search around the next best candidate focus. (b) Example focus searches from the first two iterations. (c) Focus (or depth) maps and corresponding images captured after \\( K \\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.723 + ], + "angle": 0, + "content": "**Patching strategy.** A key challenge in the technique described above is the definition of the local region or patch where the search technique maximizes contrast. In particular, we would like each patch to have minimal depth variations and especially avoid depth discontinuities, so that a single focus setting suffices for all pixels in the patch. We achieve this by performing superpixel segmentation; since depth edges invariably align with texture edges, this strategy of patching based on superpixels allows us to avoid having large depth variations or discontinuities inside a superpixel. We compute the superpixels using the k-means clustering-based SLIC method [3]. We update the superpixels after each autofocus iteration to benefit from the improved sharpness in the captured photographs, which in turn allows for finer segmentation of depth boundaries. When the updated superpixel contains pixels previously assigned to different depths, we assign the new depth to the most common (contrast-maximizing) depth in the superpixel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We illustrate our spatially-varying CDAF algorithm in Fig. 3. We also provide pseudocode in the supplemental pdf. Contrast-based autofocusing often requires multiple captures to identify the focus plane with the highest contrast. After establishing focus, one can then refine the existing focal surface over time to accommodate dynamic scenes, which is far less expensive than recapturing a full focal stack. In contrast, the technique we present next, PDAF, offers a single-shot approach to determining focus, at the cost of requiring a specialized sensor that uses dual pixel autofocus technology." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1In reality, two images suffice as the center focus in an iteration is one of the three focus settings used in a previous iteration; however, the specific patching strategy that we adopted complicates the reuse of previously captured images, thereby requiring a recapture of the center focus." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24648" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.395, + 0.106 + ], + "angle": 0, + "content": "4.2. Phase Detection Autofocus (PDAF)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.39 + ], + "angle": 0, + "content": "Recent advancements in DP sensors enabled Phase Detection Autofocus (PDAF) in consumer cameras and smartphones. A DP sensor consists of an array of microlenses, with two (or more) photodiodes under each lens. These photodiodes capture the light traveling through different parts of the aperture, producing a stereo image pair. The disparity observed across the two images determines the (signed) distance of the scene point from the focal plane (see Fig. 14 in supplemental pdf). When a scene point is sharply focused on the sensor, there is no disparity. When a scene point is out of focus, i.e., focused onto a plane in front of or behind the sensor, the disparity appears directionally proportional to the direction and depth of the focused plane. To solve for the signed disparity, we adapt the conjugate gradient-based optical flow solver by Liu [20]. From the magnitude and direction of optical flow, we can calculate the focus correction needed to bring each region into focus, thereby enabling a single-shot approach to autofocus the entire field of view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Challenges at depth boundaries. Since disparity in a stereo configuration is always horizontal, this suggests that much of our disparity estimation is driven by pixels with strong vertical gradients. However, in the context of focusing, we would also need to associate this disparity to the two regions on either side of the vertical gradient. As an example, suppose that we observe an edge that exhibits a disparity of a few pixels. When this edge corresponds to a depth boundary (which often have strong texture gradients), all we can observe is the relative shifts between the patches on either side of the boundary, and we cannot—just with local context—identify the exact shifts for each. To resolve this, we resort to segmenting the scene into semantically meaningful regions, and computing disparity for each region in isolation. In particular, we segment the total DP image using SegmentAnything [14] into masked labeled regions (layers), independently compute the masked optical flow for each layer, and sum all layers into a result flow map. The result flow map is used to drive the next autofocus. We illustrate this layered optical flow algorithm in Fig. 4 and detail the pseudocode in the supplemental pdf." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "PDAF has many desirable properties. Unlike CDAF, which requires a search procedure to identify the sharpest focus setting, PDAF requires a single image to identify the spatially-varying focus map; this allows it to adapt to scene dynamics and is less likely to get stuck in local minima. When the working range is large, we can expect significant defocus in regions with larger disparity, which degrades the quality of the disparity estimate; even in this case, as long as the sign of the disparity is accurate, the technique converges within a few steps. All of these benefits come with the requirement of a specialized sensor that has dual pixels." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.907, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.159, + 0.799, + 0.17 + ], + "angle": 0, + "content": "(a) Example phase-based AIF captures" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.172, + 0.905, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.337, + 0.786, + 0.348 + ], + "angle": 0, + "content": "(b) Phase-based autofocus algorithm" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.359, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 4. Phase-based autofocus pipeline. (a) Example 1st and 4th captures from phase-based autofocus. (b) The algorithm starts by capturing an initial DP image, and segments the result into semantically meaningful layers. For each layer, we use optical flow to compute a smooth flow map, that indicates how to adjust the focus settings. After updating focus, we capture another DP image and repeat the process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.466, + 0.6, + 0.481 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Prototype. We show our prototype in Fig. 5. The Point Spread Function (PSF) of the prototype, shown in Fig. 6, is captured for a dot placed at varying distances spanning the entire tilting range we use of the SLM. To obtain DP images, we use the Canon R10 camera sensor with a \\(6000 \\times 4000\\) resolution and a \\(3.72\\mu \\mathrm{m}\\) pixel pitch. The sensor is a dual-pixel sensor which we use for our spatially-varying PDAF algorithm. We use gphoto2 commands to programmatically acquire raw images and the LibRaw library to extract the DP views from the raw image. To perform spatially-varying autofocusing, we use the Holoeye GAEA-2 Phase-Only SLM for phase modulation, which has a pixel pitch of \\(\\delta_{\\mathrm{SLM}} = 3.74\\mu \\mathrm{m}\\) and a resolution of \\(4160 \\times 2464\\). The cubic phase plate is custom fabricated using subtractive laser etching. The relay lenses consist of three Samyang \\(85\\mathrm{mm}\\) f/1.4 AS lenses. Our objective lens is a AF-S DX Micro NIKKOR \\(40\\mathrm{mm}\\) f/2.8G." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Freeform depth-of-field photography. The ability to spatially-vary focus opens up a lot of composition capabilities beyond all-in-focus photography. After obtaining a depth map, our system can also intentionally add freeform defocus to the scene. For example, we can perform tilt-shift photography without requiring a Scheimpflug adapter to tilt the plane of focus, as seen in Fig. 7. We also show in Fig. 7 the ability to selectively focus on isolated regions at different depths, while the rest of the scene remains defocused." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24649" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.475, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.242, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Figure 5. Our prototype camera (see supplemental pdf for details)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.263, + 0.482, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Figure 6. Point-spread functions of our prototype camera. Left: Images of a point placed at different distances from the camera, created by puncturing black paper with a needle and back-illuminating it with full-spectrum white light. The dot is 10-by-10 pixels, or \\(37.2\\mu \\mathrm{m}\\) by \\(-37.2\\mu \\mathrm{m}\\) on the sensor. We label the distance of each dot in diopters. Right: Focused images of the dot, captured with an SLM spatial frequency \\(v\\). Bottom: Focused images of multiple dots at \\(v = 0.09\\) spanning the camera's FOV (see supplement for zoom-ins)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Another application of our freeform depth-of-field selection can be seen in Fig. 8, where we suppress the presence of thin structures from a photo by optically defocusing only those particular pixels. We achieve this by selecting a focus that is far from its depth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.689 + ], + "angle": 0, + "content": "All-in-focus (AIF) imaging. We capture AIF results for six scenes: Adventure, Planes, Flowers, Bunny, Ship, and Rainbow. The full results gallery is in the supplemental pdf." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For Adventure and Rainbow, we provide a quantitative analysis of AIF methods in Fig. 11 using PSNR and SSIM [39] as metrics. To obtain the ground truth target, we capture a dense focal stack with 69 depth planes, with each focus setting having 2 repeated captures averaged to reduce noise. We then combine these images to form a computational AIF image from this focal stack to serve as our ground truth. Given the recovered depth map, we also program our lens to capture an AIF image optically, shown in our qualitative comparisons. When evaluating all-in-focus results with respect to number of photos captured, we observe phase-based autofocusing having the best performance, followed by both contrast-based autofocus and capturing a focal stack. For focal sweep, we search for the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.355, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Figure 7. Example freeform depth-of-field captures of Arc de Triomphé displayed on a vertically tilted OLED. We show the capability of capturing an optical AIF image (top right), Scheimpflug-focusing to scale the defocus horizontally (bottom left), and selective focusing where a user specifies select regions to be in focus while all other regions are defocused (bottom right)." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.442, + 0.905, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.676 + ], + "angle": 0, + "content": "Figure 8. A thin structure removal example using freeform depth-of-field. The character stands in front of a thin wire mesh, far away from the lions print background. Left: A conventional capture shows the character in focus but with a visible wire mesh over the background. Right: Our proposed prototype optically removes the wire mesh by focusing on the background. The large defocus blur of the wire mesh, therefore, makes it hardly visible." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.685, + 0.868, + 0.699 + ], + "angle": 0, + "content": "regularization parameter that produces the best result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.806 + ], + "angle": 0, + "content": "For all scenes, we show qualitative results for contrast-based after 3 steps (10 photos) and phase-based after 3 steps (4 photos). We also perform focal sweep (1 photo) and reconstruct AIF images via focus stacking both computationally and optically (20 photos). For the Planes scene, we also capture a small aperture photo with a separate camera and a long exposure (\\(f/36\\) with a \\(55\\mathrm{mm}\\) lens on Nikon Z5)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Spatial resolution. We assess the spatial resolution of our imaging prototype, and the difference in performance across autofocus methods. Fig. 12 compares the modulation transfer function (MTF) of our spatially-varying autofocus methods to focal stack techniques and focal sweep. Both of our phase- and contrast-based autofocus methods show" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.368, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.255, + 0.281, + 0.268 + ], + "angle": 0, + "content": "Initial photo" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.631, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.255, + 0.595, + 0.268 + ], + "angle": 0, + "content": "Contrast-based result (ours)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.089, + 0.892, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.255, + 0.846, + 0.268 + ], + "angle": 0, + "content": "Phase-based result (ours)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.269, + 0.368, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.433, + 0.313, + 0.449 + ], + "angle": 0, + "content": "Small aperture \\((f / 36)\\)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.269, + 0.63, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.434, + 0.595, + 0.449 + ], + "angle": 0, + "content": "Focal stack (computational)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.269, + 0.891, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.434, + 0.83, + 0.449 + ], + "angle": 0, + "content": "Focal stack (optical)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.461, + 0.908, + 0.49 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison for the Planes scene. Each recovered depth map is shown in the top-left corner. All insets are blurrier for small aperture \\((f / 36)\\) when compared to our results due to diffraction (orange inset) and extreme depth range (green and blue insets)." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.5, + 0.368, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.616, + 0.368, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.666, + 0.281, + 0.68 + ], + "angle": 0, + "content": "Initial photo" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.5, + 0.63, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.616, + 0.63, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.666, + 0.595, + 0.68 + ], + "angle": 0, + "content": "Contrast-based result (ours)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.5, + 0.891, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.616, + 0.891, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.666, + 0.846, + 0.68 + ], + "angle": 0, + "content": "Phase-based result (ours)" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.68, + 0.368, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.795, + 0.368, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.846, + 0.281, + 0.86 + ], + "angle": 0, + "content": "Focal sweep" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.68, + 0.63, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.795, + 0.63, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.845, + 0.595, + 0.86 + ], + "angle": 0, + "content": "Focal stack (computational)" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.68, + 0.891, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.795, + 0.891, + 0.843 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.845, + 0.83, + 0.86 + ], + "angle": 0, + "content": "Focal stack (optical)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.873, + 0.905, + 0.888 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison for the Flowers scene. We include a focal sweep photo (does not output depth) for additional comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.095, + 0.48, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.207, + 0.48, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Figure 11. Quantitative analysis of all-in-focus imaging methods, evaluated as a function of number of photos. These plots represent the average performance over the Adventure and Rainbow scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.412 + ], + "angle": 0, + "content": "comparable performance with focal stack techniques, while focal sweep has consistently lower performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.206, + 0.44 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Aperture. We estimate that our prototype imaging system has a rest state f number of \\(f / 6.8\\) (i.e., the slope of the phase ramp is 0). When a phase ramp is applied to the SLM, the cubic phase plates are optically translated relative to one another, resulting in a smaller effective aperture. In other words, as our system moves the focus away from the nominal plane, this changes the shape of the effective aperture (impacting the bokeh) and reduces its area (limiting the total amount of light transmitted through the system). At the most extreme focus settings, we calculate that light throughput is \\(76\\%\\) when compared to the system in the rest state. See the supplemental materials for additional details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.695 + ], + "angle": 0, + "content": "It is important to note that the f-number depends on both the diameter and curvature of the cubic phase plate, among other optics. In principle, we can increase the light throughput by using a cubic phase plate with a large diameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Aberration correction. As with the continual evolution of optical systems from singlets to achromatic doublets, apochromatic designs, and modern computer-optimized multi-element systems, the optical design of our proposed system can be improved to correct for aberrations. This includes the following: (i) redesign the lenses used in our system in an end-to-end fashion, and (ii) optimize the cubic phase plate profile. For the former, Sitzmann et al. [32] and Sun et al. [34] have shown that end-to-end optical system design allows aberrations to be fixed with a higher degree of freedom. Such methods suggest that lenses can be jointly optimized with parametric SLM patterns to improve all axes of performance of the imaging system: chro" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.892, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.251, + 0.822, + 0.262 + ], + "angle": 0, + "content": "(a) Images captured using different AIF methods" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.263, + 0.891, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.453, + 0.825, + 0.465 + ], + "angle": 0, + "content": "(b) MTF curves comparing different AIF methods" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 12. MTF of our prototype. (a) We arranged printed USAF targets at three different depths and illuminated them with full-spectrum white light, and computed AIF images for different methods. (b) Each MTF is computed and averaged from both horizontal and vertical line pairs across the three depths. We observe that our methods, including phase-based, contrast-based, and focus stack optical, perform consistently comparable to the focus stack computational AIF method, while focal sweep performs consistently worse than all of them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.907, + 0.71 + ], + "angle": 0, + "content": "matic focal shift, artifacts at depth discontinuities, and f-number. For the latter, one can potentially optimize for a hybrid refractive-diffractive phase plate using the fully differentiable hybrid ray-tracing and wave-propagation (ray-wave) model recently proposed by Yang et al. [43], which significantly improved aberration correction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.74 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper introduces a first-of-its-kind imaging technique, one that provides spatially-varying autofocusing capabilities. The proposed imaging system and the autofocusing algorithms can be interpreted as an optical optimizer of image contrast; we show that standard autofocusing techniques based on contrast and phase can be readily adopted to the spatially-varying setting. In general, we believe that this novel approach to imaging has widespread applications where focus is of paramount importance, such as long-range surveillance, machine vision, and microscopy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.155 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the Air Force Office of Scientific Research (FA 95502410244), a NSF CAREER award (IIS 2238485), and a James Sprague Presidential Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.164, + 0.188, + 0.18 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.189, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.218, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.287, + 0.483, + 0.355 + ], + "angle": 0, + "content": "[3] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11):2274-2282, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[5] Luis W Alvarez and William E Humphrey. Variable-power lens and system, 1970. US Patent 3,507,565. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[6] Terry A Bartlett, William C McDonald, and James N Hall. Adapting texas instruments dlp technology to demonstrate a phase spatial light modulator. In Emerging Digital Micromirror Device Based Systems and Applications XI, pages 161-173. SPIE, 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[7] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10193-10202, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[8] Julie Chang, Isaac Kauvar, Xuemei Hu, and Gordon Wetzstein. Variable aperture light field photography: overcoming the diffraction-limited spatio-angular resolution tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3737-3745, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[9] Edward R Dowski and W Thomas Cathey. Extended depth of field through wave-front coding. Applied optics, 34(11): 1859-1866, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[11] Adam Greengard, Yoav Y Schechner, and Rafael Piestun. Depth from diffracted rotation. Optics letters, 31(2):181-183, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[12] David E Jacobs, Jongmin Baek, and Marc Levoy. Focal stack compositing for depth of field control. Stanford Computer Graphics Laboratory Technical Report, 1(1):2012, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[13] Jinbeum Jang, Yoonjong Yoo, Jongheon Kim, and Joonki Paik. Sensor-based auto-focusing system using multi-scale feature extraction and phase correlation matching. Sensors, 15(3):5747-5762, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.189, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[15] Eric P Krotkov. Active computer vision by cooperative focus and stereo. Springer Science & Business Media, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[16] Sujit Kuthirummal, Hajime Nagahara, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):58-71, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] Anat Levin, Rob Fergus, Frédo Durand, and William T Freeman. Image and depth from a conventional camera with a coded aperture. ACM transactions on graphics (TOG), 26 (3):70-es, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[18] Anat Levin, Samuel W Hasinoff, Paul Green, Frédo Durand, and William T Freeman. 4d frequency analysis of computational cameras for depth of field extension. ACM Transactions on Graphics (TOG), 28(3):1-14, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[19] Marc Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[20] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, USA, 2009. AAI0822221. 5, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[21] Adolf W Lohmann. A new class of varifocal lenses. Applied Optics, 9(7):1669-1671, 1970. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[22] Hajime Nagahara, Sujit Kuthirummal, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. In European Conference on Computer Vision, pages 60-73. Springer, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.51 + ], + "angle": 0, + "content": "[23] Shree K Nayar and Yasuo Nakagawa. Shape from focus. IEEE Transactions on Pattern analysis and machine intelligence, 16(8):824-831, 1994. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[24] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[25] Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning features by watching objects move. In CVPR, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[26] Sri Rama Prasanna Pavani and Rafael Piestun. Three dimensional tracking of fluorescent microparticles using a photon-limited double-helix response system. Optics express, 16 (26):22048-22057, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[27] Sri Rama Prasanna Pavani, Michael A Thompson, Julie S Biteen, Samuel J Lord, Na Liu, Robert J Twieg, Rafael Piestun, and William E Moerner. Three-dimensional, single-molecule fluorescence imaging beyond the diffraction limit by using a double-helix point spread function. Proceedings of the National Academy of Sciences, 106(9):2995-2999, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[28] Yingsi Qin, Wei-Yu Chen, Matthew O'Toole, and Aswin C. Sankaranarayanan. Split-lohmann multifocal displays. ACM Trans. Graph., 42(4), 2023. 2, 3, 11, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[29] Yingsi Qin, Aswin C. Sankaranarayanan, and Matthew O'Toole. Project page: Spatially-varying autofocus. https://imaging.cs.cmu.edu/svaf, 2025.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Daniel J. Reiley. Nikon AF Nikkor 85mm f/1.4D IF Lens prescription design. https://www.lens-designs.com/photographic-primes, 2017. Licensed under the MIT License. 14" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24653" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[31] Parikshit Sakurikar and P. J. Narayanan. Focal stack representation and focus manipulation. In 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR), pages 250–255, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[32] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[33] Muralidhara Subbarao and J-K Tyan. Selecting the optimal focus measure for autofocusing and depth-from-focus. IEEE transactions on pattern analysis and machine intelligence, 20(8):864-870, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[34] Qilin Sun, Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. End-to-end complex lens design with differentiate ray tracing. ACM Trans. Graph., 40(4), 2021. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[35] Supasorn Suwajanakorn, Carlos Hernandez, and Steven M Seitz. Depth from focus with your mobile phone. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.378, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[36] Dong-Chen Tsai and Homer H. Chen. Reciprocal focus profile. IEEE Transactions on Image Processing, 21(2):459-468, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.421, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[37] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.492, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[38] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.563, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.62, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[40] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[41] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin C. Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — Learning phase masks for passive single view depth estimation. In IEEE Intl. Conf. Computational Photography (ICCP), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[42] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[43] Xinge Yang, Matheus Souza, Kunyi Wang, Praneeth Chakravarthula, Qiang Fu, and Wolfgang Heidrich. End-to-end hybrid refractive-diffractive lens design with differentiable ray-wave model. In SIGGRAPH Asia 2024 Con" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.15 + ], + "angle": 0, + "content": "[44] Changyin Zhou, Daniel Miau, and Shree K Nayar. Focal sweep camera for space-time refocusing. 2012. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24654" + } + ] +] \ No newline at end of file diff --git a/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_origin.pdf b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd2f481758175d8724f7b75ed6f780d3123024c7 --- /dev/null +++ b/2025/Spatially-Varying Autofocus/812fa83e-886a-4a5e-b27e-3f560550ef19_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:821fe6bdee34721eac6ca605b757405572ced0f0444c53250455dcde091424e1 +size 49379944 diff --git a/2025/Spatially-Varying Autofocus/full.md b/2025/Spatially-Varying Autofocus/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3f1f748a51029a67b3904f66294c4edcb2de95c1 --- /dev/null +++ b/2025/Spatially-Varying Autofocus/full.md @@ -0,0 +1,293 @@ +# Spatially-Varying Autofocus + +# Yingsi Qin, Aswin C. Sankaranarayanan, Matthew O'Toole +Carnegie Mellon University + +![](images/6910b5c6e496bb7f5864b70689d67320d430e5ebc335c730cf86c413d67ae569.jpg) + +![](images/1feb8953730af0ecc694309fc73e9813636df0ec6a4413ea1c8f4736ce3736f7.jpg) + +![](images/8ff39be67e1b2def1ef0bc33818478bf8ea37577f533dae22e3ffaa806d4570c.jpg) + +![](images/219357258e66128b3b89ac4dbe69a0e5c962884804880117b634620c1c938f22.jpg) + +![](images/521b4623084adb37145d0a36210ec7e425c385fa032a9e261725fac5b6722290.jpg) +(a) Conventional photo and its confined focal plane + +![](images/cd69f3b1fd271c48cd468c4476822c2c5086eabbc9890d4f5a2d2501598467fa.jpg) +(b) All-In-Focus photo and its spatially-varying autofocused focal surface (ours) +Figure 1. Spatially-varying autofocus to produce an optical all-in-focus image. Left: A conventional photo with a regular lens, where objects at a single focal plane appear sharp. Right: An all-in-focus photo captured through spatially-varying autofocus. To achieve this, we combine (i) a programmable lens with spatially-varying control over focus, and (ii) a spatially-varying autofocus algorithm to drive the focus of this lens. Note that this is an optically-captured image of a real scene with no post-capture processing used. + +# Abstract + +A lens brings a single plane into focus on a planar sensor; hence, parts of the scene that are outside this planar focus plane are resolved under defocus. Can we break this precept by enabling a "lens" that can change its depth of field arbitrarily? This work investigates the design and implementation of such a computational lens with spatially-selective focusing. Our design uses an optical arrangement of a Lohmann lens and a phase-only spatial light modulator to allow each pixel to focus at a different depth. We extend classical autofocusing techniques to the spatially-varying scenario where the depth map is iteratively estimated using contrast and disparity cues, enabling the camera to progressively shape its depth-of-field to the scene's depth. By obtaining an all-in-focus image optically, our technique advances upon prior work in two key aspects: the ability to bring an entire scene in focus simultaneously, and the ability to maintain the highest possible spatial resolution. + +# 1. Introduction + +Core to any imaging system is the lens: an optical component designed to gather rays of light from a scene and + +form a focused image on a sensor. The focusing ability of a lens, however, only applies to a single plane in the scene. To form a focused image, the subject should be positioned at the focal plane, i.e., at a fixed depth from the camera. Points that do not lie on this focal plane appear blurry, and the amount of defocus increases progressively as the points move further away from the plane. + +Reducing the size of a lens' aperture decreases the amount of defocus and increases the depth of field (i.e., the region near the focal plane where points appear to be in sharp focus). However, this comes at the cost of reduced light throughput. Furthermore, smaller apertures increase diffractive blur [8], making content within the depth of field less sharp. These restrictions on focus are attributed to the traditional design of a lens, which offers the ability to move the focal plane (e.g., by adjusting a focus ring) but maintains its shape. Hence, we raise the question: in place of a single focal plane, is it possible to optically program a focal surface that can adapt to any scene geometry? + +This paper advances the design and implementation of a computational lens capable of spatially-varying focus—one that allows a scene in its entirety to be simultaneously in focus on an image sensor even when the scene is highly non-planar. Our approach relies on adapting the so-called Lohmann lens [21], which is a focus-tunable lens produced + +
optical sharpness# of images requiredall-in-focus generationoutputs depth
small aperturelowoneopticalno
cubic phase plate [9]lowonedeconvolutionno
focal sweep [16, 22]lowonedeconvolutionno
focal stack [23, 35]highmanycontrast metricyes
coded aperture [7, 11, 17, 18, 26, 27, 37, 41]lowonedepth-dependent deconv.yes
light field cameras [4, 10, 19, 24, 40]lowonecontrast metricyes
dual-pixel image deblurring [1, 2, 38, 42]lowonehard inverse problemyes
spatially-varying autofocus (ours)hightwoopticalyes
+ +Table 1. Comparison of all-in-focus imaging techniques. Optical sharpness: Most methods either use a small effective aperture (increasing the amount of diffraction blur), or intentionally blur the photos (e.g., to create depth-invariant blur). Our approach forms all-in-focus images by bringing each scene point into focus optically, while maintaining a large aperture. # of images required: Our method requires at least one image to approximate the scene geometry, and a second image to form the all-in-focus image. Moreover, our method is well suited for dynamic settings, where each frame determines the focus for the next frame. All-in-focus generation: Unlike most techniques, our approach forms images using an all-optical process; no additional computational post-processing is required. Outputs depth: A useful byproduct of several methods is the ability to recover a scene's depth map. + +by relative movement between two cubic lenses. Prior work by Qin et al. [28], in the context of near-eye virtual reality (VR) displays, has shown that a rearrangement of the Lohmann lens, along with the use of a phase-only spatial light modulator (SLM), can control the perceived depth of pixels on a display. Our work extends this concept to the imaging scenario to provide unprecedented control over a camera's focusing capabilities, and introduces a novel category to the solution space for all-in-focus imaging. + +Contributions. This paper proposes a programmable, spatially-varying lens for optical all-in-focus imaging and flexible depth-of-field manipulation. + +Optical all-in-focus (AIF) image. The centerpiece of our contribution is the ability to acquire an optical AIF image given knowledge of the scene depth, i.e., our technique uses knowledge of the depth to resolve all scene points in sharp focus on the sensor. A hallmark of this result is that, unlike prior work in all-in-focus imaging, our imaging process does not require computational post-processing. + +Spatially-varying autofocusing. To recover the depth map of the scene, we extend traditional ideas in contrast- and phase-based autofocus to their spatially-varying counterparts. We show that our system can progressively bring the entire scene into focus using as few as two images. This has the benefit of matching the quality and depth resolution of the focus stacking without requiring a large image set. + +Code, datasets, and real-time video demonstrations of dynamic scenes are available on our project website: https://imaging.cs.cmu.edu/svaf [29]. + +Limitation. Our current optical prototype is light inefficient due to the use of polarization-based phase modulation and a beamsplitter; note that at most one-eighth of the incident light reaches the sensor, due to light passing through a polarizer once and through a 50/50 beamsplitter twice. + +# 2. Related Work + +We briefly discuss techniques used for extended depth-of-field imaging, with a focus on key differences from our proposed approach, as summarized in Tab. 1. + +Coded aperture systems. One of the classic problems in computer vision is that of depth from defocus. For a traditional lens, this problem is ill-conditioned since the circular shape of the defocus kernel is not sufficiently discriminative. To resolve this, coded-aperture systems reshape the aperture of the lens using amplitude or phase masks. A seminal work in this space is that of Dowski and Cathey [9], who show that a cubic phase mask has a depth-invariant blur kernel that can be used to deblur and obtain the AIF image. More recent work have concentrated on enabling better discriminability of depth, either by using amplitude apertures [17, 37] or phase-based ones [7, 11, 18, 26, 27, 41]. + +Also loosely falling under the broad umbrella of coded aperture system are dual-pixel (DP) sensors, where each pixel has two sub-pixels under a single microlenslet. The resulting system acquires two images simultaneously, each from different halves of the lens aperture, thereby emulating a small baseline stereo setup. DP sensors have found extensive use in autofocusing systems [13]; however, their ability to provide stereo pairs has also enabled their use in AIF image and depth estimation [1, 2, 38, 42]. + +The proposed work differs from this class of techniques in two distinct ways. First, we do not need any computational post-processing since we optically form an all-in-focus image on the sensor. Second, since we optically focus on scene points, our images are sharper, modulo non-idealities of the optics. + +Focus stacking. One of the ways to regularize the depth-from-defocus problem is to capture a dense focal stack—that is, a collection of images obtained by sweeping the focus plane through the scene. Using local contrast as a cue, + +we can estimate depth and construct an all-in-focus image of the scene [23, 35]. However, this technique is slow, due to the need to capture a collection of images, and does not handle scene dynamics well. + +Focal sweep and flexible depth-of-field imaging. Nagahara et al. [22] explore the idea of capturing a single image, where the focal plane changes throughout its exposure. One of their results is the construction of a near-invariant defocus kernel by sweeping focus, linearly in diopter space, within the exposure; similar to Dowski and Cathey's work, this creates a depth-invariant blur that can be computationally removed. A different result enables a flexible and nonplanar depth of field by synchronizing the focus plane to a rolling shutter [16]. Our work goes beyond these results by avoiding post-process computation, as well as enabling a freeform shape for the depth of field. Such control of the depth of field was previously only possible via post-processing, for example, a focal stack [12, 31, 44]. + +Light fields. Light field cameras [4, 10, 19] sample the incident 4D space of light rays, and enables refocusing and AIF image synthesis to be done as a post-processing operation. Light fields can be measured by placing a microlens array in front of a sensor [24], or using a camera array [40], both of which sacrifice the spatial resolution of the sensors to increase angular resolution. A high angular resolution is critical for better depth selectivity, but this comes at a commensurate loss in spatial resolution for the reconstructed images. + +Autofocusing systems. Our spatially-selective focusing technique revisits classical ideas in passive autofocusing to estimate the depth map of the scene, in order to obtain an all-in-focus image. Specifically, we develop spatially-varying counterparts to the traditional contrast and phase detection autofocusing techniques. + +Contrast detection autofocus (CDAF) is one of the primary methods used by digital cameras to focus a lens. The approach involves adjusting the focus settings of the lens until the camera detects the highest contrast (usually at a few select locations). CDAF techniques can be sped up using various hill-climbing techniques [15, 33, 36] to estimate the peak contrast using a sparser set of photographs. + +Phase detection autofocus (PDAF) is an alternate technique that is commonly available in cameras with a DP sensor. When a scene point is in focus, the two images captured by the corresponding sub-pixels match. Otherwise, disparity is introduced between the two views. The signed disparity determines the lens focus for the scene point. + +In this work, we extend both techniques to our spatially-selective framework, where we apply them to bring all regions of the scene into focus simultaneously. + +![](images/8857848adaaab66b828af5a80ed2c60f77d19f3f2840e1ae8c1dbf4a61b9994c.jpg) +Figure 2. Split-Lohmann computational lens expands upon the Lohmann lens in two key steps. (a) First, it achieves a programmable focus-tunable lens by collocating the two cubic phase plates of the Lohmann lens using a 4f relay and placing a phase ramp at the Fourier plane whose slope controls the effective focal length. (b) We can optically collocate an image sensor with the SLM. Since each pixel on the sensor is resolved on the SLM, dispalying a spatially-varying phase ramp allows local focus control. + +# 3. The Split-Lohmann Computational Lens + +The ideas of this paper are inspired by a recent result on the design of a VR display [28], which proposes the Split-Lohmann lens, a computational lens that can spatially vary the focal length. For completeness, we briefly discuss this prior work before delineating our specific contributions. + +The Split-Lohmann design relies on a specific kind of focus-tunable lens called a Lohmann or Alvarez lens [5, 21] consisting of two translating cubic phase plates. Suppose that the optical profiles of the two cubic plates are given by $h_1(x) = \kappa x^3$ and $h_2(x) = -\kappa x^3$ , where $\kappa$ is a curvature-related parameter. When stacked together with a lateral offset $\Delta$ , the resulting phase modulation is given as + +$$ +h _ {1} (x + \Delta) + h _ {2} (x - \Delta) = \kappa \left(6 \Delta x ^ {2} + 2 \Delta^ {3}\right) \tag {1} +$$ + +Ignoring the constant term that is independent of $x$ , we get a phase modulation that is quadratic in $x$ , i.e., a lens whose focal length is inversely proportional to $\Delta$ . Hence, lenses of different focal lengths can be obtained by changing the amount of translation between the cubic phase plates (see Fig. 2). The Split-Lohmann display advances the Lohmann lens by first removing the mechanical translation required, and second, achieving independent local control of focal length for different regions on an OLED display. + +We propose inverting the function of this optical system, by replacing the OLED display with a camera sensor and adding a camera lens. The result is a Split-Lohmann computational lens that now offers a camera the ability to spatially-vary its focus. + +# 4. Spatially-Varying Autofocus + +Programming a Split-Lohmann computational lens to form all-in-focus images requires solving a spatially-varying autofocusing problem. While we can use a second device devoted to depth estimation—either in the form of a passive stereo camera, a structured light 3D scanner, or a time-of-flight device—we consider a self-contained autofocus loop where our device progressively estimates depth and revises its focus setting. We take inspiration from existing AF solutions used in conventional cameras to drive lens focus: Contrast Detection Autofocus (CDAF) and Phase Detection Autofocus (PDAF). Our goal here is to design the spatially-varying counterparts to these techniques, where we recover a depth image that is used to compute the necessary SLM pattern to perform simultaneous local autofocus and bring the entire field of view in focus simultaneously. + +# 4.1. Contrast Detection Autofocus (CDAF) + +Contrast is a popular focus metric for driving focus, as it can be readily implemented on most imaging systems. The basic premise of this approach relies on the observation that contrast of a local region (say, a patch) is maximized when it is in focus. CDAF techniques, hence, search for the focus setting that maximizes the contrast of a pre-determined patch. We extend CDAF to its spatially-varying counterpart by identifying an independent focus parameter for every region that maximizes its image contrast. + +Our approach relies on the insight that contrast of a patch as a function of depth is often smooth and, more importantly, unimodal. This property allows us to design an efficient search strategy by progressively narrowing down the focus/depth range where the mode can lie. Suppose that the total working range of focus, as measured in diopters (the reciprocal of depth), is from 0 to $W$ diopters. We obtain three images that correspond to the focus at $\frac{W}{4}$ , $\frac{W}{2}$ , and $\frac{3W}{4}$ diopters. We can estimate the contrast at each local patch across these three images, and the focus which has the maximum contrast (across the three images) allows us to reduce the range centering the contrast maxima. For example, if the maximum value occurs at $\frac{W}{4}$ , then the true maxima must be between the 0 and $\frac{W}{2}$ diopters. In each iteration, the search range reduces by half, and we can repeat the technique over the refined focus range. Note that, this procedure happens in parallel and independently at each patch or region in the image, enabled by the spatially-selective focusing capability of our device. Since each iteration comprises of taking three images spanning a reduced search range, we obtain the same performance of linear search, but with a number of images that is logarithmic in comparison. + +![](images/faeca5e9a5d791837bd2cc063a60db175c8c882578a834b97bd2d9af49f951dc.jpg) +(a) The search algorithm for local peak contrast + +![](images/b6e6839ea79e222d539abf757b03252252257c6d04c57db4c0caea8fadc72744.jpg) +(b) Example focus search images + +![](images/db89c2ac042817ed36a76b0c658a7dba1359574d73299299ebe61f8489440de1.jpg) +(c) Search results at increasing steps +Figure 3. Contrast-based autofocusing pipeline. (a) The objective is to identify a per-superpixel focus (or depth) setting that maximizes contrast. At step $K$ , we capture three measurements and identify the focus setting that maximizes contrast. We then refine the search around the next best candidate focus. (b) Example focus searches from the first two iterations. (c) Focus (or depth) maps and corresponding images captured after $K$ iterations. + +**Patching strategy.** A key challenge in the technique described above is the definition of the local region or patch where the search technique maximizes contrast. In particular, we would like each patch to have minimal depth variations and especially avoid depth discontinuities, so that a single focus setting suffices for all pixels in the patch. We achieve this by performing superpixel segmentation; since depth edges invariably align with texture edges, this strategy of patching based on superpixels allows us to avoid having large depth variations or discontinuities inside a superpixel. We compute the superpixels using the k-means clustering-based SLIC method [3]. We update the superpixels after each autofocus iteration to benefit from the improved sharpness in the captured photographs, which in turn allows for finer segmentation of depth boundaries. When the updated superpixel contains pixels previously assigned to different depths, we assign the new depth to the most common (contrast-maximizing) depth in the superpixel. + +We illustrate our spatially-varying CDAF algorithm in Fig. 3. We also provide pseudocode in the supplemental pdf. Contrast-based autofocusing often requires multiple captures to identify the focus plane with the highest contrast. After establishing focus, one can then refine the existing focal surface over time to accommodate dynamic scenes, which is far less expensive than recapturing a full focal stack. In contrast, the technique we present next, PDAF, offers a single-shot approach to determining focus, at the cost of requiring a specialized sensor that uses dual pixel autofocus technology. + +# 4.2. Phase Detection Autofocus (PDAF) + +Recent advancements in DP sensors enabled Phase Detection Autofocus (PDAF) in consumer cameras and smartphones. A DP sensor consists of an array of microlenses, with two (or more) photodiodes under each lens. These photodiodes capture the light traveling through different parts of the aperture, producing a stereo image pair. The disparity observed across the two images determines the (signed) distance of the scene point from the focal plane (see Fig. 14 in supplemental pdf). When a scene point is sharply focused on the sensor, there is no disparity. When a scene point is out of focus, i.e., focused onto a plane in front of or behind the sensor, the disparity appears directionally proportional to the direction and depth of the focused plane. To solve for the signed disparity, we adapt the conjugate gradient-based optical flow solver by Liu [20]. From the magnitude and direction of optical flow, we can calculate the focus correction needed to bring each region into focus, thereby enabling a single-shot approach to autofocus the entire field of view. + +Challenges at depth boundaries. Since disparity in a stereo configuration is always horizontal, this suggests that much of our disparity estimation is driven by pixels with strong vertical gradients. However, in the context of focusing, we would also need to associate this disparity to the two regions on either side of the vertical gradient. As an example, suppose that we observe an edge that exhibits a disparity of a few pixels. When this edge corresponds to a depth boundary (which often have strong texture gradients), all we can observe is the relative shifts between the patches on either side of the boundary, and we cannot—just with local context—identify the exact shifts for each. To resolve this, we resort to segmenting the scene into semantically meaningful regions, and computing disparity for each region in isolation. In particular, we segment the total DP image using SegmentAnything [14] into masked labeled regions (layers), independently compute the masked optical flow for each layer, and sum all layers into a result flow map. The result flow map is used to drive the next autofocus. We illustrate this layered optical flow algorithm in Fig. 4 and detail the pseudocode in the supplemental pdf. + +PDAF has many desirable properties. Unlike CDAF, which requires a search procedure to identify the sharpest focus setting, PDAF requires a single image to identify the spatially-varying focus map; this allows it to adapt to scene dynamics and is less likely to get stuck in local minima. When the working range is large, we can expect significant defocus in regions with larger disparity, which degrades the quality of the disparity estimate; even in this case, as long as the sign of the disparity is accurate, the technique converges within a few steps. All of these benefits come with the requirement of a specialized sensor that has dual pixels. + +![](images/9ab1b9a0d7ed0f75c4d783dbce9e5c19adc703a71415562cae43cdc5e7979552.jpg) +(a) Example phase-based AIF captures + +![](images/77411819968f21612ca975cfe507146fb0c271b2573800e7b980f0c9f3588599.jpg) +(b) Phase-based autofocus algorithm +Figure 4. Phase-based autofocus pipeline. (a) Example 1st and 4th captures from phase-based autofocus. (b) The algorithm starts by capturing an initial DP image, and segments the result into semantically meaningful layers. For each layer, we use optical flow to compute a smooth flow map, that indicates how to adjust the focus settings. After updating focus, we capture another DP image and repeat the process. + +# 5. Results + +Prototype. We show our prototype in Fig. 5. The Point Spread Function (PSF) of the prototype, shown in Fig. 6, is captured for a dot placed at varying distances spanning the entire tilting range we use of the SLM. To obtain DP images, we use the Canon R10 camera sensor with a $6000 \times 4000$ resolution and a $3.72\mu \mathrm{m}$ pixel pitch. The sensor is a dual-pixel sensor which we use for our spatially-varying PDAF algorithm. We use gphoto2 commands to programmatically acquire raw images and the LibRaw library to extract the DP views from the raw image. To perform spatially-varying autofocusing, we use the Holoeye GAEA-2 Phase-Only SLM for phase modulation, which has a pixel pitch of $\delta_{\mathrm{SLM}} = 3.74\mu \mathrm{m}$ and a resolution of $4160 \times 2464$ . The cubic phase plate is custom fabricated using subtractive laser etching. The relay lenses consist of three Samyang $85\mathrm{mm}$ f/1.4 AS lenses. Our objective lens is a AF-S DX Micro NIKKOR $40\mathrm{mm}$ f/2.8G. + +Freeform depth-of-field photography. The ability to spatially-vary focus opens up a lot of composition capabilities beyond all-in-focus photography. After obtaining a depth map, our system can also intentionally add freeform defocus to the scene. For example, we can perform tilt-shift photography without requiring a Scheimpflug adapter to tilt the plane of focus, as seen in Fig. 7. We also show in Fig. 7 the ability to selectively focus on isolated regions at different depths, while the rest of the scene remains defocused. + +![](images/4a5ccc80dc3623a754b62ca9427c7865b82f7dc92db6f35e1fd631c7a8e675cd.jpg) +Figure 5. Our prototype camera (see supplemental pdf for details). + +![](images/5154c4617d83d8c71742e066e617b44bef523c602976d9cdebf4293260ab6f76.jpg) +Figure 6. Point-spread functions of our prototype camera. Left: Images of a point placed at different distances from the camera, created by puncturing black paper with a needle and back-illuminating it with full-spectrum white light. The dot is 10-by-10 pixels, or $37.2\mu \mathrm{m}$ by $-37.2\mu \mathrm{m}$ on the sensor. We label the distance of each dot in diopters. Right: Focused images of the dot, captured with an SLM spatial frequency $v$ . Bottom: Focused images of multiple dots at $v = 0.09$ spanning the camera's FOV (see supplement for zoom-ins). + +Another application of our freeform depth-of-field selection can be seen in Fig. 8, where we suppress the presence of thin structures from a photo by optically defocusing only those particular pixels. We achieve this by selecting a focus that is far from its depth. + +All-in-focus (AIF) imaging. We capture AIF results for six scenes: Adventure, Planes, Flowers, Bunny, Ship, and Rainbow. The full results gallery is in the supplemental pdf. + +For Adventure and Rainbow, we provide a quantitative analysis of AIF methods in Fig. 11 using PSNR and SSIM [39] as metrics. To obtain the ground truth target, we capture a dense focal stack with 69 depth planes, with each focus setting having 2 repeated captures averaged to reduce noise. We then combine these images to form a computational AIF image from this focal stack to serve as our ground truth. Given the recovered depth map, we also program our lens to capture an AIF image optically, shown in our qualitative comparisons. When evaluating all-in-focus results with respect to number of photos captured, we observe phase-based autofocusing having the best performance, followed by both contrast-based autofocus and capturing a focal stack. For focal sweep, we search for the + +![](images/abf165e3e8754c10d699c61c4dfdca38ebe77ee0af2caced0f67daa14b08dd0c.jpg) +Figure 7. Example freeform depth-of-field captures of Arc de Triomphé displayed on a vertically tilted OLED. We show the capability of capturing an optical AIF image (top right), Scheimpflug-focusing to scale the defocus horizontally (bottom left), and selective focusing where a user specifies select regions to be in focus while all other regions are defocused (bottom right). + +![](images/36d974a3d6e5a67f2bf94a6eedd87922d8f432f9f76ad8d3085d314e129618c5.jpg) +Figure 8. A thin structure removal example using freeform depth-of-field. The character stands in front of a thin wire mesh, far away from the lions print background. Left: A conventional capture shows the character in focus but with a visible wire mesh over the background. Right: Our proposed prototype optically removes the wire mesh by focusing on the background. The large defocus blur of the wire mesh, therefore, makes it hardly visible. + +regularization parameter that produces the best result. + +For all scenes, we show qualitative results for contrast-based after 3 steps (10 photos) and phase-based after 3 steps (4 photos). We also perform focal sweep (1 photo) and reconstruct AIF images via focus stacking both computationally and optically (20 photos). For the Planes scene, we also capture a small aperture photo with a separate camera and a long exposure ( $f/36$ with a $55\mathrm{mm}$ lens on Nikon Z5). + +Spatial resolution. We assess the spatial resolution of our imaging prototype, and the difference in performance across autofocus methods. Fig. 12 compares the modulation transfer function (MTF) of our spatially-varying autofocus methods to focal stack techniques and focal sweep. Both of our phase- and contrast-based autofocus methods show + +![](images/c2ecba02bc49dbfa6c47f0c4890e926ae85b28fdfc6c0dcc5145deb50ba3ef62.jpg) +Initial photo + +![](images/a2153fcf5f52749e18b460d5c5a0340a569d18cc993dc1faf655555a8cc34493.jpg) +Contrast-based result (ours) + +![](images/bf605895f5924117f90a338828b5d0a34f440648e6cf0e58b17a0adf3fb69cfb.jpg) +Phase-based result (ours) + +![](images/ada5c7bd7047640ab6b2475d5096b209396d0184f4e92872b4b743b8ca90a7a8.jpg) +Small aperture $(f / 36)$ + +![](images/856ad0cbfee069556b1cdf80888f64fc5bbea04958c7dc8202a72074ac4c9792.jpg) +Focal stack (computational) + +![](images/42a33fd30ffea5fbd188263f718bfb11b07e377baa6592863f8b45cc0718a75c.jpg) +Focal stack (optical) + +![](images/11ef3ec0bb66d8d0fce2f5a79424f890aa521a9301377fb0ddf4876efa730e0f.jpg) +Figure 9. Qualitative comparison for the Planes scene. Each recovered depth map is shown in the top-left corner. All insets are blurrier for small aperture $(f / 36)$ when compared to our results due to diffraction (orange inset) and extreme depth range (green and blue insets). + +![](images/570e582d88cd00cc17568d4654321988ba21ab8ae72d9d35377dc473b9bc0465.jpg) + +![](images/2d5ad358656cb56c027235ae3e9908cfab4e79cccadeaf0d8202aaab59d17c22.jpg) + +![](images/4220e1ec342102e9767713aab14920d57ca4c7d85f3a35fe8956e69dd1cc0b44.jpg) + +![](images/e444246d41a81abfb85f81cf0aec1c5268a336d376f4167abb5d6e08776a6964.jpg) + +![](images/16cd2906cad6fe314b30ca96de535aabed17eb639b8ae2fba69206b48a9dcda2.jpg) + +![](images/996340d54d21d219fca076d954091b3adb3b1ffe953204d138d2ad1857572ac5.jpg) +Initial photo + +![](images/b35d6cea0d20f956381655284a3c3cc53e49facc3685428a9ece0c255ecac533.jpg) +Focal sweep + +![](images/5f057cba3c9019c2f0aaa46abe0822852c98a64035d8b45fea58fcad072ca0db.jpg) +Contrast-based result (ours) + +![](images/3e54d5d946ec94a5a97c07710ec572b78206a56ab47fe9f02a7454b1e09b155f.jpg) +Focal stack (computational) +Figure 10. Qualitative comparison for the Flowers scene. We include a focal sweep photo (does not output depth) for additional comparison. + +![](images/f015d1cdd7b9ef3064fa8e4422646f4d719dd47cf5e8c692bb353697960704e6.jpg) +Phase-based result (ours) + +![](images/37631c4454f8e2bfacf941f6b07dab0dff0b3eb30c0311a78b5567a80e7dd7a7.jpg) +Focal stack (optical) + +![](images/dc8b4b71dbbf2cf0de4c3ec11b1b9b62b70a48fc999a56148c2d13a4c0277180.jpg) + +![](images/02f9015b1f4fd9d303642784b5fdf1143500bc6c6416daaabaf66b589d0d2b8f.jpg) +Figure 11. Quantitative analysis of all-in-focus imaging methods, evaluated as a function of number of photos. These plots represent the average performance over the Adventure and Rainbow scenes. + +comparable performance with focal stack techniques, while focal sweep has consistently lower performance. + +# 6. Discussion + +Aperture. We estimate that our prototype imaging system has a rest state f number of $f / 6.8$ (i.e., the slope of the phase ramp is 0). When a phase ramp is applied to the SLM, the cubic phase plates are optically translated relative to one another, resulting in a smaller effective aperture. In other words, as our system moves the focus away from the nominal plane, this changes the shape of the effective aperture (impacting the bokeh) and reduces its area (limiting the total amount of light transmitted through the system). At the most extreme focus settings, we calculate that light throughput is $76\%$ when compared to the system in the rest state. See the supplemental materials for additional details. + +It is important to note that the f-number depends on both the diameter and curvature of the cubic phase plate, among other optics. In principle, we can increase the light throughput by using a cubic phase plate with a large diameter. + +Aberration correction. As with the continual evolution of optical systems from singlets to achromatic doublets, apochromatic designs, and modern computer-optimized multi-element systems, the optical design of our proposed system can be improved to correct for aberrations. This includes the following: (i) redesign the lenses used in our system in an end-to-end fashion, and (ii) optimize the cubic phase plate profile. For the former, Sitzmann et al. [32] and Sun et al. [34] have shown that end-to-end optical system design allows aberrations to be fixed with a higher degree of freedom. Such methods suggest that lenses can be jointly optimized with parametric SLM patterns to improve all axes of performance of the imaging system: chro + +![](images/915a90bf4c4f9ae159251b16987d8a34d377d710a717cfb38e01d1791b2777ad.jpg) +(a) Images captured using different AIF methods + +![](images/a0e1164fed1817b71b8d1b110784b17809e47d96a2d2689ea14fff852f0d3e71.jpg) +(b) MTF curves comparing different AIF methods +Figure 12. MTF of our prototype. (a) We arranged printed USAF targets at three different depths and illuminated them with full-spectrum white light, and computed AIF images for different methods. (b) Each MTF is computed and averaged from both horizontal and vertical line pairs across the three depths. We observe that our methods, including phase-based, contrast-based, and focus stack optical, perform consistently comparable to the focus stack computational AIF method, while focal sweep performs consistently worse than all of them. + +matic focal shift, artifacts at depth discontinuities, and f-number. For the latter, one can potentially optimize for a hybrid refractive-diffractive phase plate using the fully differentiable hybrid ray-tracing and wave-propagation (ray-wave) model recently proposed by Yang et al. [43], which significantly improved aberration correction. + +# 7. Conclusion + +This paper introduces a first-of-its-kind imaging technique, one that provides spatially-varying autofocusing capabilities. The proposed imaging system and the autofocusing algorithms can be interpreted as an optical optimizer of image contrast; we show that standard autofocusing techniques based on contrast and phase can be readily adopted to the spatially-varying setting. In general, we believe that this novel approach to imaging has widespread applications where focus is of paramount importance, such as long-range surveillance, machine vision, and microscopy. + +Acknowledgements. This work was supported by the Air Force Office of Scientific Research (FA 95502410244), a NSF CAREER award (IIS 2238485), and a James Sprague Presidential Fellowship. + +# References + +[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In ECCV, 2020. 2 +[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In IEEE/CVF International Conference on Computer Vision, 2021. 2 +[3] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11):2274-2282, 2012. 4 +[4] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 2, 3 +[5] Luis W Alvarez and William E Humphrey. Variable-power lens and system, 1970. US Patent 3,507,565. 3 +[6] Terry A Bartlett, William C McDonald, and James N Hall. Adapting texas instruments dlp technology to demonstrate a phase spatial light modulator. In Emerging Digital Micromirror Device Based Systems and Applications XI, pages 161-173. SPIE, 2019. 14 +[7] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10193-10202, 2019. 2 +[8] Julie Chang, Isaac Kauvar, Xuemei Hu, and Gordon Wetzstein. Variable aperture light field photography: overcoming the diffraction-limited spatio-angular resolution tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3737-3745, 2016. 1 +[9] Edward R Dowski and W Thomas Cathey. Extended depth of field through wave-front coding. Applied optics, 34(11): 1859-1866, 1995. 2 +[10] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2, 3 +[11] Adam Greengard, Yoav Y Schechner, and Rafael Piestun. Depth from diffracted rotation. Optics letters, 31(2):181-183, 2006. 2 +[12] David E Jacobs, Jongmin Baek, and Marc Levoy. Focal stack compositing for depth of field control. Stanford Computer Graphics Laboratory Technical Report, 1(1):2012, 2012. 3 +[13] Jinbeum Jang, Yoonjong Yoo, Jongheon Kim, and Joonki Paik. Sensor-based auto-focusing system using multi-scale feature extraction and phase correlation matching. Sensors, 15(3):5747-5762, 2015. 2 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and + +Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 5, 14 +[15] Eric P Krotkov. Active computer vision by cooperative focus and stereo. Springer Science & Business Media, 2012. 3 +[16] Sujit Kuthirummal, Hajime Nagahara, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):58-71, 2010. 2, 3 +[17] Anat Levin, Rob Fergus, Frédo Durand, and William T Freeman. Image and depth from a conventional camera with a coded aperture. ACM transactions on graphics (TOG), 26 (3):70-es, 2007. 2 +[18] Anat Levin, Samuel W Hasinoff, Paul Green, Frédo Durand, and William T Freeman. 4d frequency analysis of computational cameras for depth of field extension. ACM Transactions on Graphics (TOG), 28(3):1-14, 2009. 2 +[19] Marc Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2, 3 +[20] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, USA, 2009. AAI0822221. 5, 14, 15 +[21] Adolf W Lohmann. A new class of varifocal lenses. Applied Optics, 9(7):1669-1671, 1970. 1, 3 +[22] Hajime Nagahara, Sujit Kuthirummal, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. In European Conference on Computer Vision, pages 60-73. Springer, 2008. 2, 3 +[23] Shree K Nayar and Yasuo Nakagawa. Shape from focus. IEEE Transactions on Pattern analysis and machine intelligence, 16(8):824-831, 1994. 2, 3 +[24] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 3 +[25] Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning features by watching objects move. In CVPR, 2017. 14 +[26] Sri Rama Prasanna Pavani and Rafael Piestun. Three dimensional tracking of fluorescent microparticles using a photon-limited double-helix response system. Optics express, 16 (26):22048-22057, 2008. 2 +[27] Sri Rama Prasanna Pavani, Michael A Thompson, Julie S Biteen, Samuel J Lord, Na Liu, Robert J Twieg, Rafael Piestun, and William E Moerner. Three-dimensional, single-molecule fluorescence imaging beyond the diffraction limit by using a double-helix point spread function. Proceedings of the National Academy of Sciences, 106(9):2995-2999, 2009. 2 +[28] Yingsi Qin, Wei-Yu Chen, Matthew O'Toole, and Aswin C. Sankaranarayanan. Split-lohmann multifocal displays. ACM Trans. Graph., 42(4), 2023. 2, 3, 11, 14 +[29] Yingsi Qin, Aswin C. Sankaranarayanan, and Matthew O'Toole. Project page: Spatially-varying autofocus. https://imaging.cs.cmu.edu/svaf, 2025.2 +[30] Daniel J. Reiley. Nikon AF Nikkor 85mm f/1.4D IF Lens prescription design. https://www.lens-designs.com/photographic-primes, 2017. Licensed under the MIT License. 14 + +[31] Parikshit Sakurikar and P. J. Narayanan. Focal stack representation and focus manipulation. In 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR), pages 250–255, 2017. 3 +[32] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 8, 14 +[33] Muralidhara Subbarao and J-K Tyan. Selecting the optimal focus measure for autofocusing and depth-from-focus. IEEE transactions on pattern analysis and machine intelligence, 20(8):864-870, 1998. 3 +[34] Qilin Sun, Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. End-to-end complex lens design with differentiate ray tracing. ACM Trans. Graph., 40(4), 2021. 8, 14 +[35] Supasorn Suwajanakorn, Carlos Hernandez, and Steven M Seitz. Depth from focus with your mobile phone. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2015. 2, 3 +[36] Dong-Chen Tsai and Homer H. Chen. Reciprocal focus profile. IEEE Transactions on Image Processing, 21(2):459-468, 2012. 3 +[37] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69, 2007. 2 +[38] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2 +[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6 +[40] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 2, 3 +[41] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin C. Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — Learning phase masks for passive single view depth estimation. In IEEE Intl. Conf. Computational Photography (ICCP), 2019. 2 +[42] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 2 +[43] Xinge Yang, Matheus Souza, Kunyi Wang, Praneeth Chakravarthula, Qiang Fu, and Wolfgang Heidrich. End-to-end hybrid refractive-diffractive lens design with differentiable ray-wave model. In SIGGRAPH Asia 2024 Con + +ference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 8, 14 +[44] Changyin Zhou, Daniel Miau, and Shree K Nayar. Focal sweep camera for space-time refocusing. 2012. 3 \ No newline at end of file diff --git a/2025/Spatially-Varying Autofocus/images.zip b/2025/Spatially-Varying Autofocus/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2331f4fa64e2200ceae1ebc97348ca7b659a7851 --- /dev/null +++ b/2025/Spatially-Varying Autofocus/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e9a0014794d11d9d9b9ded2cee738d2f2b962e484095f7bbf92041b5f8b509f +size 1070949 diff --git a/2025/Spatially-Varying Autofocus/layout.json b/2025/Spatially-Varying Autofocus/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6b57bb6add0ed2774eb650859eaac1fc7f1165d --- /dev/null +++ b/2025/Spatially-Varying Autofocus/layout.json @@ -0,0 +1,7720 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 217, + 103, + 394, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 103, + 394, + 121 + ], + "spans": [ + { + "bbox": [ + 217, + 103, + 394, + 121 + ], + "type": "text", + "content": "Spatially-Varying Autofocus" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 143, + 451, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 143, + 451, + 173 + ], + "spans": [ + { + "bbox": [ + 162, + 143, + 451, + 173 + ], + "type": "text", + "content": "Yingsi Qin, Aswin C. Sankaranarayanan, Matthew O'Toole \nCarnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 186, + 236, + 288 + ], + "blocks": [ + { + "bbox": [ + 57, + 186, + 236, + 288 + ], + "lines": [ + { + "bbox": [ + 57, + 186, + 236, + 288 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 236, + 288 + ], + "type": "image", + "image_path": "6910b5c6e496bb7f5864b70689d67320d430e5ebc335c730cf86c413d67ae569.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 187, + 303, + 286 + ], + "blocks": [ + { + "bbox": [ + 238, + 187, + 303, + 286 + ], + "lines": [ + { + "bbox": [ + 238, + 187, + 303, + 286 + ], + "spans": [ + { + "bbox": [ + 238, + 187, + 303, + 286 + ], + "type": "image", + "image_path": "1feb8953730af0ecc694309fc73e9813636df0ec6a4413ea1c8f4736ce3736f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 187, + 485, + 288 + ], + "blocks": [ + { + "bbox": [ + 307, + 187, + 485, + 288 + ], + "lines": [ + { + "bbox": [ + 307, + 187, + 485, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 187, + 485, + 288 + ], + "type": "image", + "image_path": "8ff39be67e1b2def1ef0bc33818478bf8ea37577f533dae22e3ffaa806d4570c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 487, + 187, + 553, + 287 + ], + "blocks": [ + { + "bbox": [ + 487, + 187, + 553, + 287 + ], + "lines": [ + { + "bbox": [ + 487, + 187, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 487, + 187, + 553, + 287 + ], + "type": "image", + "image_path": "219357258e66128b3b89ac4dbe69a0e5c962884804880117b634620c1c938f22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 289, + 302, + 341 + ], + "blocks": [ + { + "bbox": [ + 57, + 289, + 302, + 341 + ], + "lines": [ + { + "bbox": [ + 57, + 289, + 302, + 341 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 302, + 341 + ], + "type": "image", + "image_path": "521b4623084adb37145d0a36210ec7e425c385fa032a9e261725fac5b6722290.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 343, + 259, + 353 + ], + "lines": [ + { + "bbox": [ + 101, + 343, + 259, + 353 + ], + "spans": [ + { + "bbox": [ + 101, + 343, + 259, + 353 + ], + "type": "text", + "content": "(a) Conventional photo and its confined focal plane" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 289, + 553, + 341 + ], + "blocks": [ + { + "bbox": [ + 307, + 289, + 553, + 341 + ], + "lines": [ + { + "bbox": [ + 307, + 289, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 553, + 341 + ], + "type": "image", + "image_path": "cd69f3b1fd271c48cd468c4476822c2c5086eabbc9890d4f5a2d2501598467fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 342, + 552, + 353 + ], + "lines": [ + { + "bbox": [ + 308, + 342, + 552, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 552, + 353 + ], + "type": "text", + "content": "(b) All-In-Focus photo and its spatially-varying autofocused focal surface (ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 555, + 404 + ], + "type": "text", + "content": "Figure 1. Spatially-varying autofocus to produce an optical all-in-focus image. Left: A conventional photo with a regular lens, where objects at a single focal plane appear sharp. Right: An all-in-focus photo captured through spatially-varying autofocus. To achieve this, we combine (i) a programmable lens with spatially-varying control over focus, and (ii) a spatially-varying autofocus algorithm to drive the focus of this lens. Note that this is an optically-captured image of a real scene with no post-capture processing used." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "spans": [ + { + "bbox": [ + 151, + 413, + 200, + 426 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 439, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 439, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 439, + 297, + 643 + ], + "type": "text", + "content": "A lens brings a single plane into focus on a planar sensor; hence, parts of the scene that are outside this planar focus plane are resolved under defocus. Can we break this precept by enabling a \"lens\" that can change its depth of field arbitrarily? This work investigates the design and implementation of such a computational lens with spatially-selective focusing. Our design uses an optical arrangement of a Lohmann lens and a phase-only spatial light modulator to allow each pixel to focus at a different depth. We extend classical autofocusing techniques to the spatially-varying scenario where the depth map is iteratively estimated using contrast and disparity cues, enabling the camera to progressively shape its depth-of-field to the scene's depth. By obtaining an all-in-focus image optically, our technique advances upon prior work in two key aspects: the ability to bring an entire scene in focus simultaneously, and the ability to maintain the highest possible spatial resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 668, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Core to any imaging system is the lens: an optical component designed to gather rays of light from a scene and" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 498 + ], + "type": "text", + "content": "form a focused image on a sensor. The focusing ability of a lens, however, only applies to a single plane in the scene. To form a focused image, the subject should be positioned at the focal plane, i.e., at a fixed depth from the camera. Points that do not lie on this focal plane appear blurry, and the amount of defocus increases progressively as the points move further away from the plane." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 642 + ], + "type": "text", + "content": "Reducing the size of a lens' aperture decreases the amount of defocus and increases the depth of field (i.e., the region near the focal plane where points appear to be in sharp focus). However, this comes at the cost of reduced light throughput. Furthermore, smaller apertures increase diffractive blur [8], making content within the depth of field less sharp. These restrictions on focus are attributed to the traditional design of a lens, which offers the ability to move the focal plane (e.g., by adjusting a focus ring) but maintains its shape. Hence, we raise the question: in place of a single focal plane, is it possible to optically program a focal surface that can adapt to any scene geometry?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "This paper advances the design and implementation of a computational lens capable of spatially-varying focus—one that allows a scene in its entirety to be simultaneously in focus on an image sensor even when the scene is highly non-planar. Our approach relies on adapting the so-called Lohmann lens [21], which is a focus-tunable lens produced" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24645" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 177 + ], + "type": "table", + "html": "
optical sharpness# of images requiredall-in-focus generationoutputs depth
small aperturelowoneopticalno
cubic phase plate [9]lowonedeconvolutionno
focal sweep [16, 22]lowonedeconvolutionno
focal stack [23, 35]highmanycontrast metricyes
coded aperture [7, 11, 17, 18, 26, 27, 37, 41]lowonedepth-dependent deconv.yes
light field cameras [4, 10, 19, 24, 40]lowonecontrast metricyes
dual-pixel image deblurring [1, 2, 38, 42]lowonehard inverse problemyes
spatially-varying autofocus (ours)hightwoopticalyes
", + "image_path": "84f65a6cfc7d0ee4e679af6ada1641d36befcb5dc7f4c030eff1713f564e4da3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 185, + 555, + 263 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 263 + ], + "type": "text", + "content": "Table 1. Comparison of all-in-focus imaging techniques. Optical sharpness: Most methods either use a small effective aperture (increasing the amount of diffraction blur), or intentionally blur the photos (e.g., to create depth-invariant blur). Our approach forms all-in-focus images by bringing each scene point into focus optically, while maintaining a large aperture. # of images required: Our method requires at least one image to approximate the scene geometry, and a second image to form the all-in-focus image. Moreover, our method is well suited for dynamic settings, where each frame determines the focus for the next frame. All-in-focus generation: Unlike most techniques, our approach forms images using an all-optical process; no additional computational post-processing is required. Outputs depth: A useful byproduct of several methods is the ability to recover a scene's depth map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 275, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 294, + 384 + ], + "type": "text", + "content": "by relative movement between two cubic lenses. Prior work by Qin et al. [28], in the context of near-eye virtual reality (VR) displays, has shown that a rearrangement of the Lohmann lens, along with the use of a phase-only spatial light modulator (SLM), can control the perceived depth of pixels on a display. Our work extends this concept to the imaging scenario to provide unprecedented control over a camera's focusing capabilities, and introduces a novel category to the solution space for all-in-focus imaging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 429 + ], + "type": "text", + "content": "Contributions. This paper proposes a programmable, spatially-varying lens for optical all-in-focus imaging and flexible depth-of-field manipulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": "Optical all-in-focus (AIF) image. The centerpiece of our contribution is the ability to acquire an optical AIF image given knowledge of the scene depth, i.e., our technique uses knowledge of the depth to resolve all scene points in sharp focus on the sensor. A hallmark of this result is that, unlike prior work in all-in-focus imaging, our imaging process does not require computational post-processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 608 + ], + "type": "text", + "content": "Spatially-varying autofocusing. To recover the depth map of the scene, we extend traditional ideas in contrast- and phase-based autofocus to their spatially-varying counterparts. We show that our system can progressively bring the entire scene into focus using as few as two images. This has the benefit of matching the quality and depth resolution of the focus stacking without requiring a large image set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": "Code, datasets, and real-time video demonstrations of dynamic scenes are available on our project website: https://imaging.cs.cmu.edu/svaf [29]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Limitation. Our current optical prototype is light inefficient due to the use of polarization-based phase modulation and a beamsplitter; note that at most one-eighth of the incident light reaches the sensor, due to light passing through a polarizer once and through a 50/50 beamsplitter twice." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 274, + 400, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 285 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 294, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 330 + ], + "type": "text", + "content": "We briefly discuss techniques used for extended depth-of-field imaging, with a focus on key differences from our proposed approach, as summarized in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 479 + ], + "type": "text", + "content": "Coded aperture systems. One of the classic problems in computer vision is that of depth from defocus. For a traditional lens, this problem is ill-conditioned since the circular shape of the defocus kernel is not sufficiently discriminative. To resolve this, coded-aperture systems reshape the aperture of the lens using amplitude or phase masks. A seminal work in this space is that of Dowski and Cathey [9], who show that a cubic phase mask has a depth-invariant blur kernel that can be used to deblur and obtain the AIF image. More recent work have concentrated on enabling better discriminability of depth, either by using amplitude apertures [17, 37] or phase-based ones [7, 11, 18, 26, 27, 41]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 586 + ], + "type": "text", + "content": "Also loosely falling under the broad umbrella of coded aperture system are dual-pixel (DP) sensors, where each pixel has two sub-pixels under a single microlenslet. The resulting system acquires two images simultaneously, each from different halves of the lens aperture, thereby emulating a small baseline stereo setup. DP sensors have found extensive use in autofocusing systems [13]; however, their ability to provide stereo pairs has also enabled their use in AIF image and depth estimation [1, 2, 38, 42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 588, + 553, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 553, + 659 + ], + "type": "text", + "content": "The proposed work differs from this class of techniques in two distinct ways. First, we do not need any computational post-processing since we optically form an all-in-focus image on the sensor. Second, since we optically focus on scene points, our images are sharper, modulo non-idealities of the optics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Focus stacking. One of the ways to regularize the depth-from-defocus problem is to capture a dense focal stack—that is, a collection of images obtained by sweeping the focus plane through the scene. Using local contrast as a cue," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24646" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "we can estimate depth and construct an all-in-focus image of the scene [23, 35]. However, this technique is slow, due to the need to capture a collection of images, and does not handle scene dynamics well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 301 + ], + "type": "text", + "content": "Focal sweep and flexible depth-of-field imaging. Nagahara et al. [22] explore the idea of capturing a single image, where the focal plane changes throughout its exposure. One of their results is the construction of a near-invariant defocus kernel by sweeping focus, linearly in diopter space, within the exposure; similar to Dowski and Cathey's work, this creates a depth-invariant blur that can be computationally removed. A different result enables a flexible and nonplanar depth of field by synchronizing the focus plane to a rolling shutter [16]. Our work goes beyond these results by avoiding post-process computation, as well as enabling a freeform shape for the depth of field. Such control of the depth of field was previously only possible via post-processing, for example, a focal stack [12, 31, 44]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 434 + ], + "type": "text", + "content": "Light fields. Light field cameras [4, 10, 19] sample the incident 4D space of light rays, and enables refocusing and AIF image synthesis to be done as a post-processing operation. Light fields can be measured by placing a microlens array in front of a sensor [24], or using a camera array [40], both of which sacrifice the spatial resolution of the sensors to increase angular resolution. A high angular resolution is critical for better depth selectivity, but this comes at a commensurate loss in spatial resolution for the reconstructed images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 518 + ], + "type": "text", + "content": "Autofocusing systems. Our spatially-selective focusing technique revisits classical ideas in passive autofocusing to estimate the depth map of the scene, in order to obtain an all-in-focus image. Specifically, we develop spatially-varying counterparts to the traditional contrast and phase detection autofocusing techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 519, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 604 + ], + "type": "text", + "content": "Contrast detection autofocus (CDAF) is one of the primary methods used by digital cameras to focus a lens. The approach involves adjusting the focus settings of the lens until the camera detects the highest contrast (usually at a few select locations). CDAF techniques can be sped up using various hill-climbing techniques [15, 33, 36] to estimate the peak contrast using a sparser set of photographs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": "Phase detection autofocus (PDAF) is an alternate technique that is commonly available in cameras with a DP sensor. When a scene point is in focus, the two images captured by the corresponding sub-pixels match. Otherwise, disparity is introduced between the two views. The signed disparity determines the lens focus for the scene point." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In this work, we extend both techniques to our spatially-selective framework, where we apply them to bring all regions of the scene into focus simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 69, + 542, + 220 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 542, + 220 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 542, + 220 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 542, + 220 + ], + "type": "image", + "image_path": "8857848adaaab66b828af5a80ed2c60f77d19f3f2840e1ae8c1dbf4a61b9994c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 316 + ], + "type": "text", + "content": "Figure 2. Split-Lohmann computational lens expands upon the Lohmann lens in two key steps. (a) First, it achieves a programmable focus-tunable lens by collocating the two cubic phase plates of the Lohmann lens using a 4f relay and placing a phase ramp at the Fourier plane whose slope controls the effective focal length. (b) We can optically collocate an image sensor with the SLM. Since each pixel on the sensor is resolved on the SLM, dispalying a spatially-varying phase ramp allows local focus control." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 538, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 538, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 538, + 346 + ], + "type": "text", + "content": "3. The Split-Lohmann Computational Lens" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": "The ideas of this paper are inspired by a recent result on the design of a VR display [28], which proposes the Split-Lohmann lens, a computational lens that can spatially vary the focal length. For completeness, we briefly discuss this prior work before delineating our specific contributions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "content": "The Split-Lohmann design relies on a specific kind of focus-tunable lens called a Lohmann or Alvarez lens [5, 21] consisting of two translating cubic phase plates. Suppose that the optical profiles of the two cubic plates are given by " + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "inline_equation", + "content": "h_1(x) = \\kappa x^3" + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "inline_equation", + "content": "h_2(x) = -\\kappa x^3" + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "content": " is a curvature-related parameter. When stacked together with a lateral offset " + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 417, + 554, + 501 + ], + "type": "text", + "content": ", the resulting phase modulation is given as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 340, + 515, + 553, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 515, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 340, + 515, + 553, + 530 + ], + "type": "interline_equation", + "content": "h _ {1} (x + \\Delta) + h _ {2} (x - \\Delta) = \\kappa \\left(6 \\Delta x ^ {2} + 2 \\Delta^ {3}\\right) \\tag {1}", + "image_path": "ad0f94fe5bb7ff6c4bd82da7bd962929ff0f77299e14aca25ef0abfde9ee3cb3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": "Ignoring the constant term that is independent of " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": ", we get a phase modulation that is quadratic in " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": ", i.e., a lens whose focal length is inversely proportional to " + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 544, + 554, + 651 + ], + "type": "text", + "content": ". Hence, lenses of different focal lengths can be obtained by changing the amount of translation between the cubic phase plates (see Fig. 2). The Split-Lohmann display advances the Lohmann lens by first removing the mechanical translation required, and second, achieving independent local control of focal length for different regions on an OLED display." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "We propose inverting the function of this optical system, by replacing the OLED display with a camera sensor and adding a camera lens. The result is a Split-Lohmann computational lens that now offers a camera the ability to spatially-vary its focus." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24647" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 216, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 216, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 216, + 85 + ], + "type": "text", + "content": "4. Spatially-Varying Autofocus" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 270 + ], + "type": "text", + "content": "Programming a Split-Lohmann computational lens to form all-in-focus images requires solving a spatially-varying autofocusing problem. While we can use a second device devoted to depth estimation—either in the form of a passive stereo camera, a structured light 3D scanner, or a time-of-flight device—we consider a self-contained autofocus loop where our device progressively estimates depth and revises its focus setting. We take inspiration from existing AF solutions used in conventional cameras to drive lens focus: Contrast Detection Autofocus (CDAF) and Phase Detection Autofocus (PDAF). Our goal here is to design the spatially-varying counterparts to these techniques, where we recover a depth image that is used to compute the necessary SLM pattern to perform simultaneous local autofocus and bring the entire field of view in focus simultaneously." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 257, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 257, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 257, + 290 + ], + "type": "text", + "content": "4.1. Contrast Detection Autofocus (CDAF)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 295, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 295, + 403 + ], + "type": "text", + "content": "Contrast is a popular focus metric for driving focus, as it can be readily implemented on most imaging systems. The basic premise of this approach relies on the observation that contrast of a local region (say, a patch) is maximized when it is in focus. CDAF techniques, hence, search for the focus setting that maximizes the contrast of a pre-determined patch. We extend CDAF to its spatially-varying counterpart by identifying an independent focus parameter for every region that maximizes its image contrast." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": "Our approach relies on the insight that contrast of a patch as a function of depth is often smooth and, more importantly, unimodal. This property allows us to design an efficient search strategy by progressively narrowing down the focus/depth range where the mode can lie. Suppose that the total working range of focus, as measured in diopters (the reciprocal of depth), is from 0 to " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": " diopters. We obtain three images that correspond to the focus at " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\frac{W}{4}" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\frac{W}{2}" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\frac{3W}{4}" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": " diopters. We can estimate the contrast at each local patch across these three images, and the focus which has the maximum contrast (across the three images) allows us to reduce the range centering the contrast maxima. For example, if the maximum value occurs at " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\frac{W}{4}" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": ", then the true maxima must be between the 0 and " + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\frac{W}{2}" + }, + { + "bbox": [ + 55, + 403, + 296, + 668 + ], + "type": "text", + "content": " diopters. In each iteration, the search range reduces by half, and we can repeat the technique over the refined focus range. Note that, this procedure happens in parallel and independently at each patch or region in the image, enabled by the spatially-selective focusing capability of our device. Since each iteration comprises of taking three images spanning a reduced search range, we obtain the same performance of linear search, but with a number of images that is logarithmic in comparison." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 448, + 233 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 448, + 233 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 448, + 233 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 448, + 233 + ], + "type": "image", + "image_path": "faeca5e9a5d791837bd2cc063a60db175c8c882578a834b97bd2d9af49f951dc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 237, + 445, + 245 + ], + "lines": [ + { + "bbox": [ + 321, + 237, + 445, + 245 + ], + "spans": [ + { + "bbox": [ + 321, + 237, + 445, + 245 + ], + "type": "text", + "content": "(a) The search algorithm for local peak contrast" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 452, + 70, + 554, + 137 + ], + "blocks": [ + { + "bbox": [ + 452, + 70, + 554, + 137 + ], + "lines": [ + { + "bbox": [ + 452, + 70, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 452, + 70, + 554, + 137 + ], + "type": "image", + "image_path": "b6e6839ea79e222d539abf757b03252252257c6d04c57db4c0caea8fadc72744.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 137, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 460, + 137, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 460, + 137, + 547, + 145 + ], + "type": "text", + "content": "(b) Example focus search images" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 449, + 147, + 553, + 233 + ], + "blocks": [ + { + "bbox": [ + 449, + 147, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 449, + 147, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 449, + 147, + 553, + 233 + ], + "type": "image", + "image_path": "db89c2ac042817ed36a76b0c658a7dba1359574d73299299ebe61f8489440de1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 236, + 551, + 245 + ], + "lines": [ + { + "bbox": [ + 456, + 236, + 551, + 245 + ], + "spans": [ + { + "bbox": [ + 456, + 236, + 551, + 245 + ], + "type": "text", + "content": "(c) Search results at increasing steps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "type": "text", + "content": "Figure 3. Contrast-based autofocusing pipeline. (a) The objective is to identify a per-superpixel focus (or depth) setting that maximizes contrast. At step " + }, + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "type": "text", + "content": ", we capture three measurements and identify the focus setting that maximizes contrast. We then refine the search around the next best candidate focus. (b) Example focus searches from the first two iterations. (c) Focus (or depth) maps and corresponding images captured after " + }, + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 254, + 555, + 331 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 356, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 572 + ], + "type": "text", + "content": "**Patching strategy.** A key challenge in the technique described above is the definition of the local region or patch where the search technique maximizes contrast. In particular, we would like each patch to have minimal depth variations and especially avoid depth discontinuities, so that a single focus setting suffices for all pixels in the patch. We achieve this by performing superpixel segmentation; since depth edges invariably align with texture edges, this strategy of patching based on superpixels allows us to avoid having large depth variations or discontinuities inside a superpixel. We compute the superpixels using the k-means clustering-based SLIC method [3]. We update the superpixels after each autofocus iteration to benefit from the improved sharpness in the captured photographs, which in turn allows for finer segmentation of depth boundaries. When the updated superpixel contains pixels previously assigned to different depths, we assign the new depth to the most common (contrast-maximizing) depth in the superpixel." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We illustrate our spatially-varying CDAF algorithm in Fig. 3. We also provide pseudocode in the supplemental pdf. Contrast-based autofocusing often requires multiple captures to identify the focus plane with the highest contrast. After establishing focus, one can then refine the existing focal surface over time to accommodate dynamic scenes, which is far less expensive than recapturing a full focal stack. In contrast, the technique we present next, PDAF, offers a single-shot approach to determining focus, at the cost of requiring a specialized sensor that uses dual pixel autofocus technology." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "1In reality, two images suffice as the center focus in an iteration is one of the three focus settings used in a previous iteration; however, the specific patching strategy that we adopted complicates the reuse of previously captured images, thereby requiring a recapture of the center focus." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24648" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 241, + 83 + ], + "type": "text", + "content": "4.2. Phase Detection Autofocus (PDAF)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 308 + ], + "type": "text", + "content": "Recent advancements in DP sensors enabled Phase Detection Autofocus (PDAF) in consumer cameras and smartphones. A DP sensor consists of an array of microlenses, with two (or more) photodiodes under each lens. These photodiodes capture the light traveling through different parts of the aperture, producing a stereo image pair. The disparity observed across the two images determines the (signed) distance of the scene point from the focal plane (see Fig. 14 in supplemental pdf). When a scene point is sharply focused on the sensor, there is no disparity. When a scene point is out of focus, i.e., focused onto a plane in front of or behind the sensor, the disparity appears directionally proportional to the direction and depth of the focused plane. To solve for the signed disparity, we adapt the conjugate gradient-based optical flow solver by Liu [20]. From the magnitude and direction of optical flow, we can calculate the focus correction needed to bring each region into focus, thereby enabling a single-shot approach to autofocus the entire field of view." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "type": "text", + "content": "Challenges at depth boundaries. Since disparity in a stereo configuration is always horizontal, this suggests that much of our disparity estimation is driven by pixels with strong vertical gradients. However, in the context of focusing, we would also need to associate this disparity to the two regions on either side of the vertical gradient. As an example, suppose that we observe an edge that exhibits a disparity of a few pixels. When this edge corresponds to a depth boundary (which often have strong texture gradients), all we can observe is the relative shifts between the patches on either side of the boundary, and we cannot—just with local context—identify the exact shifts for each. To resolve this, we resort to segmenting the scene into semantically meaningful regions, and computing disparity for each region in isolation. In particular, we segment the total DP image using SegmentAnything [14] into masked labeled regions (layers), independently compute the masked optical flow for each layer, and sum all layers into a result flow map. The result flow map is used to drive the next autofocus. We illustrate this layered optical flow algorithm in Fig. 4 and detail the pseudocode in the supplemental pdf." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "PDAF has many desirable properties. Unlike CDAF, which requires a search procedure to identify the sharpest focus setting, PDAF requires a single image to identify the spatially-varying focus map; this allows it to adapt to scene dynamics and is less likely to get stuck in local minima. When the working range is large, we can expect significant defocus in regions with larger disparity, which degrades the quality of the disparity estimate; even in this case, as long as the sign of the disparity is accurate, the technique converges within a few steps. All of these benefits come with the requirement of a specialized sensor that has dual pixels." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 555, + 125 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 555, + 125 + ], + "type": "image", + "image_path": "9ab1b9a0d7ed0f75c4d783dbce9e5c19adc703a71415562cae43cdc5e7979552.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 125, + 488, + 134 + ], + "lines": [ + { + "bbox": [ + 379, + 125, + 488, + 134 + ], + "spans": [ + { + "bbox": [ + 379, + 125, + 488, + 134 + ], + "type": "text", + "content": "(a) Example phase-based AIF captures" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 136, + 553, + 265 + ], + "blocks": [ + { + "bbox": [ + 316, + 136, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 316, + 136, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 136, + 553, + 265 + ], + "type": "image", + "image_path": "77411819968f21612ca975cfe507146fb0c271b2573800e7b980f0c9f3588599.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 266, + 481, + 275 + ], + "lines": [ + { + "bbox": [ + 377, + 266, + 481, + 275 + ], + "spans": [ + { + "bbox": [ + 377, + 266, + 481, + 275 + ], + "type": "text", + "content": "(b) Phase-based autofocus algorithm" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 284, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 361 + ], + "type": "text", + "content": "Figure 4. Phase-based autofocus pipeline. (a) Example 1st and 4th captures from phase-based autofocus. (b) The algorithm starts by capturing an initial DP image, and segments the result into semantically meaningful layers. For each layer, we use optical flow to compute a smooth flow map, that indicates how to adjust the focus settings. After updating focus, we capture another DP image and repeat the process." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 369, + 367, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 367, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 367, + 380 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": "Prototype. We show our prototype in Fig. 5. The Point Spread Function (PSF) of the prototype, shown in Fig. 6, is captured for a dot placed at varying distances spanning the entire tilting range we use of the SLM. To obtain DP images, we use the Canon R10 camera sensor with a " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "6000 \\times 4000" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": " resolution and a " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "3.72\\mu \\mathrm{m}" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": " pixel pitch. The sensor is a dual-pixel sensor which we use for our spatially-varying PDAF algorithm. We use gphoto2 commands to programmatically acquire raw images and the LibRaw library to extract the DP views from the raw image. To perform spatially-varying autofocusing, we use the Holoeye GAEA-2 Phase-Only SLM for phase modulation, which has a pixel pitch of " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{SLM}} = 3.74\\mu \\mathrm{m}" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": " and a resolution of " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "4160 \\times 2464" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": ". The cubic phase plate is custom fabricated using subtractive laser etching. The relay lenses consist of three Samyang " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "85\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": " f/1.4 AS lenses. Our objective lens is a AF-S DX Micro NIKKOR " + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "inline_equation", + "content": "40\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 389, + 556, + 594 + ], + "type": "text", + "content": " f/2.8G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Freeform depth-of-field photography. The ability to spatially-vary focus opens up a lot of composition capabilities beyond all-in-focus photography. After obtaining a depth map, our system can also intentionally add freeform defocus to the scene. For example, we can perform tilt-shift photography without requiring a Scheimpflug adapter to tilt the plane of focus, as seen in Fig. 7. We also show in Fig. 7 the ability to selectively focus on isolated regions at different depths, while the rest of the scene remains defocused." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24649" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 290, + 182 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 290, + 182 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 290, + 182 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 290, + 182 + ], + "type": "image", + "image_path": "4a5ccc80dc3623a754b62ca9427c7865b82f7dc92db6f35e1fd631c7a8e675cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 191, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 56, + 191, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 294, + 203 + ], + "type": "text", + "content": "Figure 5. Our prototype camera (see supplemental pdf for details)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 208, + 294, + 316 + ], + "blocks": [ + { + "bbox": [ + 58, + 208, + 294, + 316 + ], + "lines": [ + { + "bbox": [ + 58, + 208, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 58, + 208, + 294, + 316 + ], + "type": "image", + "image_path": "5154c4617d83d8c71742e066e617b44bef523c602976d9cdebf4293260ab6f76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "text", + "content": "Figure 6. Point-spread functions of our prototype camera. Left: Images of a point placed at different distances from the camera, created by puncturing black paper with a needle and back-illuminating it with full-spectrum white light. The dot is 10-by-10 pixels, or " + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "inline_equation", + "content": "37.2\\mu \\mathrm{m}" + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "inline_equation", + "content": "-37.2\\mu \\mathrm{m}" + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "text", + "content": " on the sensor. We label the distance of each dot in diopters. Right: Focused images of the dot, captured with an SLM spatial frequency " + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "text", + "content": ". Bottom: Focused images of multiple dots at " + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "inline_equation", + "content": "v = 0.09" + }, + { + "bbox": [ + 55, + 324, + 296, + 424 + ], + "type": "text", + "content": " spanning the camera's FOV (see supplement for zoom-ins)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 499 + ], + "type": "text", + "content": "Another application of our freeform depth-of-field selection can be seen in Fig. 8, where we suppress the presence of thin structures from a photo by optically defocusing only those particular pixels. We achieve this by selecting a focus that is far from its depth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 545 + ], + "type": "text", + "content": "All-in-focus (AIF) imaging. We capture AIF results for six scenes: Adventure, Planes, Flowers, Bunny, Ship, and Rainbow. The full results gallery is in the supplemental pdf." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "For Adventure and Rainbow, we provide a quantitative analysis of AIF methods in Fig. 11 using PSNR and SSIM [39] as metrics. To obtain the ground truth target, we capture a dense focal stack with 69 depth planes, with each focus setting having 2 repeated captures averaged to reduce noise. We then combine these images to form a computational AIF image from this focal stack to serve as our ground truth. Given the recovered depth map, we also program our lens to capture an AIF image optically, shown in our qualitative comparisons. When evaluating all-in-focus results with respect to number of photos captured, we observe phase-based autofocusing having the best performance, followed by both contrast-based autofocus and capturing a focal stack. For focal sweep, we search for the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 271 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 271 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 271 + ], + "type": "image", + "image_path": "abf165e3e8754c10d699c61c4dfdca38ebe77ee0af2caced0f67daa14b08dd0c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 281, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 346 + ], + "type": "text", + "content": "Figure 7. Example freeform depth-of-field captures of Arc de Triomphé displayed on a vertically tilted OLED. We show the capability of capturing an optical AIF image (top right), Scheimpflug-focusing to scale the defocus horizontally (bottom left), and selective focusing where a user specifies select regions to be in focus while all other regions are defocused (bottom right)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 350, + 553, + 449 + ], + "blocks": [ + { + "bbox": [ + 315, + 350, + 553, + 449 + ], + "lines": [ + { + "bbox": [ + 315, + 350, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 315, + 350, + 553, + 449 + ], + "type": "image", + "image_path": "36d974a3d6e5a67f2bf94a6eedd87922d8f432f9f76ad8d3085d314e129618c5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 457, + 555, + 535 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 535 + ], + "type": "text", + "content": "Figure 8. A thin structure removal example using freeform depth-of-field. The character stands in front of a thin wire mesh, far away from the lions print background. Left: A conventional capture shows the character in focus but with a visible wire mesh over the background. Right: Our proposed prototype optically removes the wire mesh by focusing on the background. The large defocus blur of the wire mesh, therefore, makes it hardly visible." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 542, + 531, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 531, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 531, + 553 + ], + "type": "text", + "content": "regularization parameter that produces the best result." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "text", + "content": "For all scenes, we show qualitative results for contrast-based after 3 steps (10 photos) and phase-based after 3 steps (4 photos). We also perform focal sweep (1 photo) and reconstruct AIF images via focus stacking both computationally and optically (20 photos). For the Planes scene, we also capture a small aperture photo with a separate camera and a long exposure (" + }, + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "inline_equation", + "content": "f/36" + }, + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "text", + "content": " with a " + }, + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "inline_equation", + "content": "55\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 554, + 555, + 638 + ], + "type": "text", + "content": " lens on Nikon Z5)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Spatial resolution. We assess the spatial resolution of our imaging prototype, and the difference in performance across autofocus methods. Fig. 12 compares the modulation transfer function (MTF) of our spatially-varying autofocus methods to focal stack techniques and focal sweep. Both of our phase- and contrast-based autofocus methods show" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24650" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 225, + 199 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 225, + 199 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 225, + 199 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 225, + 199 + ], + "type": "image", + "image_path": "c2ecba02bc49dbfa6c47f0c4890e926ae85b28fdfc6c0dcc5145deb50ba3ef62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 201, + 171, + 212 + ], + "lines": [ + { + "bbox": [ + 120, + 201, + 171, + 212 + ], + "spans": [ + { + "bbox": [ + 120, + 201, + 171, + 212 + ], + "type": "text", + "content": "Initial photo" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 386, + 199 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 386, + 199 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 386, + 199 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 386, + 199 + ], + "type": "image", + "image_path": "a2153fcf5f52749e18b460d5c5a0340a569d18cc993dc1faf655555a8cc34493.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 201, + 364, + 212 + ], + "lines": [ + { + "bbox": [ + 249, + 201, + 364, + 212 + ], + "spans": [ + { + "bbox": [ + 249, + 201, + 364, + 212 + ], + "type": "text", + "content": "Contrast-based result (ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 70, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 387, + 70, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 545, + 200 + ], + "type": "image", + "image_path": "bf605895f5924117f90a338828b5d0a34f440648e6cf0e58b17a0adf3fb69cfb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 201, + 517, + 212 + ], + "lines": [ + { + "bbox": [ + 413, + 201, + 517, + 212 + ], + "spans": [ + { + "bbox": [ + 413, + 201, + 517, + 212 + ], + "type": "text", + "content": "Phase-based result (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 213, + 225, + 341 + ], + "blocks": [ + { + "bbox": [ + 67, + 213, + 225, + 341 + ], + "lines": [ + { + "bbox": [ + 67, + 213, + 225, + 341 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 225, + 341 + ], + "type": "image", + "image_path": "ada5c7bd7047640ab6b2475d5096b209396d0184f4e92872b4b743b8ca90a7a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 342, + 191, + 355 + ], + "lines": [ + { + "bbox": [ + 100, + 342, + 191, + 355 + ], + "spans": [ + { + "bbox": [ + 100, + 342, + 191, + 355 + ], + "type": "text", + "content": "Small aperture " + }, + { + "bbox": [ + 100, + 342, + 191, + 355 + ], + "type": "inline_equation", + "content": "(f / 36)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 228, + 213, + 385, + 341 + ], + "blocks": [ + { + "bbox": [ + 228, + 213, + 385, + 341 + ], + "lines": [ + { + "bbox": [ + 228, + 213, + 385, + 341 + ], + "spans": [ + { + "bbox": [ + 228, + 213, + 385, + 341 + ], + "type": "image", + "image_path": "856ad0cbfee069556b1cdf80888f64fc5bbea04958c7dc8202a72074ac4c9792.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 343, + 364, + 355 + ], + "lines": [ + { + "bbox": [ + 250, + 343, + 364, + 355 + ], + "spans": [ + { + "bbox": [ + 250, + 343, + 364, + 355 + ], + "type": "text", + "content": "Focal stack (computational)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 387, + 213, + 545, + 341 + ], + "blocks": [ + { + "bbox": [ + 387, + 213, + 545, + 341 + ], + "lines": [ + { + "bbox": [ + 387, + 213, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 387, + 213, + 545, + 341 + ], + "type": "image", + "image_path": "42a33fd30ffea5fbd188263f718bfb11b07e377baa6592863f8b45cc0718a75c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 343, + 507, + 355 + ], + "lines": [ + { + "bbox": [ + 424, + 343, + 507, + 355 + ], + "spans": [ + { + "bbox": [ + 424, + 343, + 507, + 355 + ], + "type": "text", + "content": "Focal stack (optical)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 66, + 396, + 225, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison for the Planes scene. Each recovered depth map is shown in the top-left corner. All insets are blurrier for small aperture " + }, + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "inline_equation", + "content": "(f / 36)" + }, + { + "bbox": [ + 55, + 365, + 555, + 388 + ], + "type": "text", + "content": " when compared to our results due to diffraction (orange inset) and extreme depth range (green and blue insets)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 396, + 225, + 487 + ], + "lines": [ + { + "bbox": [ + 66, + 396, + 225, + 487 + ], + "spans": [ + { + "bbox": [ + 66, + 396, + 225, + 487 + ], + "type": "image", + "image_path": "11ef3ec0bb66d8d0fce2f5a79424f890aa521a9301377fb0ddf4876efa730e0f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 66, + 487, + 225, + 525 + ], + "blocks": [ + { + "bbox": [ + 66, + 487, + 225, + 525 + ], + "lines": [ + { + "bbox": [ + 66, + 487, + 225, + 525 + ], + "spans": [ + { + "bbox": [ + 66, + 487, + 225, + 525 + ], + "type": "image", + "image_path": "570e582d88cd00cc17568d4654321988ba21ab8ae72d9d35377dc473b9bc0465.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 228, + 396, + 385, + 487 + ], + "blocks": [ + { + "bbox": [ + 228, + 396, + 385, + 487 + ], + "lines": [ + { + "bbox": [ + 228, + 396, + 385, + 487 + ], + "spans": [ + { + "bbox": [ + 228, + 396, + 385, + 487 + ], + "type": "image", + "image_path": "2d5ad358656cb56c027235ae3e9908cfab4e79cccadeaf0d8202aaab59d17c22.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 228, + 487, + 385, + 525 + ], + "blocks": [ + { + "bbox": [ + 228, + 487, + 385, + 525 + ], + "lines": [ + { + "bbox": [ + 228, + 487, + 385, + 525 + ], + "spans": [ + { + "bbox": [ + 228, + 487, + 385, + 525 + ], + "type": "image", + "image_path": "4220e1ec342102e9767713aab14920d57ca4c7d85f3a35fe8956e69dd1cc0b44.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 387, + 396, + 545, + 487 + ], + "blocks": [ + { + "bbox": [ + 387, + 396, + 545, + 487 + ], + "lines": [ + { + "bbox": [ + 387, + 396, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 387, + 396, + 545, + 487 + ], + "type": "image", + "image_path": "e444246d41a81abfb85f81cf0aec1c5268a336d376f4167abb5d6e08776a6964.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 387, + 487, + 545, + 525 + ], + "blocks": [ + { + "bbox": [ + 387, + 487, + 545, + 525 + ], + "lines": [ + { + "bbox": [ + 387, + 487, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 387, + 487, + 545, + 525 + ], + "type": "image", + "image_path": "16cd2906cad6fe314b30ca96de535aabed17eb639b8ae2fba69206b48a9dcda2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 67, + 538, + 225, + 628 + ], + "blocks": [ + { + "bbox": [ + 121, + 527, + 171, + 538 + ], + "lines": [ + { + "bbox": [ + 121, + 527, + 171, + 538 + ], + "spans": [ + { + "bbox": [ + 121, + 527, + 171, + 538 + ], + "type": "text", + "content": "Initial photo" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 538, + 225, + 628 + ], + "lines": [ + { + "bbox": [ + 67, + 538, + 225, + 628 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 225, + 628 + ], + "type": "image", + "image_path": "996340d54d21d219fca076d954091b3adb3b1ffe953204d138d2ad1857572ac5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 67, + 629, + 225, + 666 + ], + "blocks": [ + { + "bbox": [ + 67, + 629, + 225, + 666 + ], + "lines": [ + { + "bbox": [ + 67, + 629, + 225, + 666 + ], + "spans": [ + { + "bbox": [ + 67, + 629, + 225, + 666 + ], + "type": "image", + "image_path": "b35d6cea0d20f956381655284a3c3cc53e49facc3685428a9ece0c255ecac533.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 670, + 171, + 681 + ], + "lines": [ + { + "bbox": [ + 120, + 670, + 171, + 681 + ], + "spans": [ + { + "bbox": [ + 120, + 670, + 171, + 681 + ], + "type": "text", + "content": "Focal sweep" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 228, + 538, + 385, + 628 + ], + "blocks": [ + { + "bbox": [ + 249, + 527, + 364, + 538 + ], + "lines": [ + { + "bbox": [ + 249, + 527, + 364, + 538 + ], + "spans": [ + { + "bbox": [ + 249, + 527, + 364, + 538 + ], + "type": "text", + "content": "Contrast-based result (ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 538, + 385, + 628 + ], + "lines": [ + { + "bbox": [ + 228, + 538, + 385, + 628 + ], + "spans": [ + { + "bbox": [ + 228, + 538, + 385, + 628 + ], + "type": "image", + "image_path": "5f057cba3c9019c2f0aaa46abe0822852c98a64035d8b45fea58fcad072ca0db.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 228, + 629, + 385, + 667 + ], + "blocks": [ + { + "bbox": [ + 228, + 629, + 385, + 667 + ], + "lines": [ + { + "bbox": [ + 228, + 629, + 385, + 667 + ], + "spans": [ + { + "bbox": [ + 228, + 629, + 385, + 667 + ], + "type": "image", + "image_path": "3e54d5d946ec94a5a97c07710ec572b78206a56ab47fe9f02a7454b1e09b155f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 669, + 364, + 681 + ], + "lines": [ + { + "bbox": [ + 249, + 669, + 364, + 681 + ], + "spans": [ + { + "bbox": [ + 249, + 669, + 364, + 681 + ], + "type": "text", + "content": "Focal stack (computational)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 691, + 553, + 703 + ], + "lines": [ + { + "bbox": [ + 55, + 691, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 691, + 553, + 703 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison for the Flowers scene. We include a focal sweep photo (does not output depth) for additional comparison." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 387, + 538, + 545, + 628 + ], + "blocks": [ + { + "bbox": [ + 413, + 527, + 517, + 538 + ], + "lines": [ + { + "bbox": [ + 413, + 527, + 517, + 538 + ], + "spans": [ + { + "bbox": [ + 413, + 527, + 517, + 538 + ], + "type": "text", + "content": "Phase-based result (ours)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 538, + 545, + 628 + ], + "lines": [ + { + "bbox": [ + 387, + 538, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 387, + 538, + 545, + 628 + ], + "type": "image", + "image_path": "f015d1cdd7b9ef3064fa8e4422646f4d719dd47cf5e8c692bb353697960704e6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 387, + 629, + 545, + 667 + ], + "blocks": [ + { + "bbox": [ + 387, + 629, + 545, + 667 + ], + "lines": [ + { + "bbox": [ + 387, + 629, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 387, + 629, + 545, + 667 + ], + "type": "image", + "image_path": "37631c4454f8e2bfacf941f6b07dab0dff0b3eb30c0311a78b5567a80e7dd7a7.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 669, + 507, + 681 + ], + "lines": [ + { + "bbox": [ + 424, + 669, + 507, + 681 + ], + "spans": [ + { + "bbox": [ + 424, + 669, + 507, + 681 + ], + "type": "text", + "content": "Focal stack (optical)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24651" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 75, + 293, + 163 + ], + "blocks": [ + { + "bbox": [ + 59, + 75, + 293, + 163 + ], + "lines": [ + { + "bbox": [ + 59, + 75, + 293, + 163 + ], + "spans": [ + { + "bbox": [ + 59, + 75, + 293, + 163 + ], + "type": "image", + "image_path": "dc8b4b71dbbf2cf0de4c3ec11b1b9b62b70a48fc999a56148c2d13a4c0277180.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 163, + 293, + 251 + ], + "blocks": [ + { + "bbox": [ + 58, + 163, + 293, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 163, + 293, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 163, + 293, + 251 + ], + "type": "image", + "image_path": "02f9015b1f4fd9d303642784b5fdf1143500bc6c6416daaabaf66b589d0d2b8f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 293 + ], + "type": "text", + "content": "Figure 11. Quantitative analysis of all-in-focus imaging methods, evaluated as a function of number of photos. These plots represent the average performance over the Adventure and Rainbow scenes." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 326 + ], + "type": "text", + "content": "comparable performance with focal stack techniques, while focal sweep has consistently lower performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 126, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 126, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 126, + 348 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "text", + "content": "Aperture. We estimate that our prototype imaging system has a rest state f number of " + }, + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "inline_equation", + "content": "f / 6.8" + }, + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "text", + "content": " (i.e., the slope of the phase ramp is 0). When a phase ramp is applied to the SLM, the cubic phase plates are optically translated relative to one another, resulting in a smaller effective aperture. In other words, as our system moves the focus away from the nominal plane, this changes the shape of the effective aperture (impacting the bokeh) and reduces its area (limiting the total amount of light transmitted through the system). At the most extreme focus settings, we calculate that light throughput is " + }, + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 54, + 357, + 295, + 501 + ], + "type": "text", + "content": " when compared to the system in the rest state. See the supplemental materials for additional details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 501, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 550 + ], + "type": "text", + "content": "It is important to note that the f-number depends on both the diameter and curvature of the cubic phase plate, among other optics. In principle, we can increase the light throughput by using a cubic phase plate with a large diameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Aberration correction. As with the continual evolution of optical systems from singlets to achromatic doublets, apochromatic designs, and modern computer-optimized multi-element systems, the optical design of our proposed system can be improved to correct for aberrations. This includes the following: (i) redesign the lenses used in our system in an end-to-end fashion, and (ii) optimize the cubic phase plate profile. For the former, Sitzmann et al. [32] and Sun et al. [34] have shown that end-to-end optical system design allows aberrations to be fixed with a higher degree of freedom. Such methods suggest that lenses can be jointly optimized with parametric SLM patterns to improve all axes of performance of the imaging system: chro" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 545, + 196 + ], + "type": "image", + "image_path": "915a90bf4c4f9ae159251b16987d8a34d377d710a717cfb38e01d1791b2777ad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 198, + 503, + 207 + ], + "lines": [ + { + "bbox": [ + 367, + 198, + 503, + 207 + ], + "spans": [ + { + "bbox": [ + 367, + 198, + 503, + 207 + ], + "type": "text", + "content": "(a) Images captured using different AIF methods" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 208, + 545, + 356 + ], + "blocks": [ + { + "bbox": [ + 325, + 208, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 325, + 208, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 325, + 208, + 545, + 356 + ], + "type": "image", + "image_path": "a0e1164fed1817b71b8d1b110784b17809e47d96a2d2689ea14fff852f0d3e71.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 358, + 504, + 368 + ], + "lines": [ + { + "bbox": [ + 365, + 358, + 504, + 368 + ], + "spans": [ + { + "bbox": [ + 365, + 358, + 504, + 368 + ], + "type": "text", + "content": "(b) MTF curves comparing different AIF methods" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 376, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 477 + ], + "type": "text", + "content": "Figure 12. MTF of our prototype. (a) We arranged printed USAF targets at three different depths and illuminated them with full-spectrum white light, and computed AIF images for different methods. (b) Each MTF is computed and averaged from both horizontal and vertical line pairs across the three depths. We observe that our methods, including phase-based, contrast-based, and focus stack optical, perform consistently comparable to the focus stack computational AIF method, while focal sweep performs consistently worse than all of them." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 489, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 562 + ], + "type": "text", + "content": "matic focal shift, artifacts at depth discontinuities, and f-number. For the latter, one can potentially optimize for a hybrid refractive-diffractive phase plate using the fully differentiable hybrid ray-tracing and wave-propagation (ray-wave) model recently proposed by Yang et al. [43], which significantly improved aberration correction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "This paper introduces a first-of-its-kind imaging technique, one that provides spatially-varying autofocusing capabilities. The proposed imaging system and the autofocusing algorithms can be interpreted as an optical optimizer of image contrast; we show that standard autofocusing techniques based on contrast and phase can be readily adopted to the spatially-varying setting. In general, we believe that this novel approach to imaging has widespread applications where focus is of paramount importance, such as long-range surveillance, machine vision, and microscopy." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24652" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 122 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the Air Force Office of Scientific Research (FA 95502410244), a NSF CAREER award (IIS 2238485), and a James Sprague Presidential Fellowship." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 149, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 149, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 149, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 149, + 294, + 171 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In ECCV, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 172, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 172, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 295, + 225 + ], + "type": "text", + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 227, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 227, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 295, + 281 + ], + "type": "text", + "content": "[3] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(11):2274-2282, 2012. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 295, + 315 + ], + "type": "text", + "content": "[4] Edward H Adelson, James R Bergen, et al. The plenoptic function and the elements of early vision. Computational models of visual processing, 1(2):3-20, 1991. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 337 + ], + "type": "text", + "content": "[5] Luis W Alvarez and William E Humphrey. Variable-power lens and system, 1970. US Patent 3,507,565. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 295, + 392 + ], + "type": "text", + "content": "[6] Terry A Bartlett, William C McDonald, and James N Hall. Adapting texas instruments dlp technology to demonstrate a phase spatial light modulator. In Emerging Digital Micromirror Device Based Systems and Applications XI, pages 161-173. SPIE, 2019. 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 295, + 437 + ], + "type": "text", + "content": "[7] Julie Chang and Gordon Wetzstein. Deep optics for monocular depth estimation and 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10193-10202, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "type": "text", + "content": "[8] Julie Chang, Isaac Kauvar, Xuemei Hu, and Gordon Wetzstein. Variable aperture light field photography: overcoming the diffraction-limited spatio-angular resolution tradeoff. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3737-3745, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 525 + ], + "type": "text", + "content": "[9] Edward R Dowski and W Thomas Cathey. Extended depth of field through wave-front coding. Applied optics, 34(11): 1859-1866, 1995. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 526, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 295, + 568 + ], + "type": "text", + "content": "[10] Steven J Gortler, Radek Grzeszcuk, Richard Szeliski, and Michael F Cohen. The lumigraph. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 453-464. 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 601 + ], + "type": "text", + "content": "[11] Adam Greengard, Yoav Y Schechner, and Rafael Piestun. Depth from diffracted rotation. Optics letters, 31(2):181-183, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 635 + ], + "type": "text", + "content": "[12] David E Jacobs, Jongmin Baek, and Marc Levoy. Focal stack compositing for depth of field control. Stanford Computer Graphics Laboratory Technical Report, 1(1):2012, 2012. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "text", + "content": "[13] Jinbeum Jang, Yoonjong Yoo, Jongheon Kim, and Joonki Paik. Sensor-based auto-focusing system using multi-scale feature extraction and phase correlation matching. Sensors, 15(3):5747-5762, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 5, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "text", + "content": "[15] Eric P Krotkov. Active computer vision by cooperative focus and stereo. Springer Science & Business Media, 2012. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[16] Sujit Kuthirummal, Hajime Nagahara, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1):58-71, 2010. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[17] Anat Levin, Rob Fergus, Frédo Durand, and William T Freeman. Image and depth from a conventional camera with a coded aperture. ACM transactions on graphics (TOG), 26 (3):70-es, 2007. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[18] Anat Levin, Samuel W Hasinoff, Paul Green, Frédo Durand, and William T Freeman. 4d frequency analysis of computational cameras for depth of field extension. ACM Transactions on Graphics (TOG), 28(3):1-14, 2009. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "type": "text", + "content": "[19] Marc Levoy. Light fields and computational imaging. Computer, 39(8):46-55, 2006. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "text", + "content": "[20] Ce Liu. Beyond pixels: exploring new representations and applications for motion analysis. PhD thesis, USA, 2009. AAI0822221. 5, 14, 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "type": "text", + "content": "[21] Adolf W Lohmann. A new class of varifocal lenses. Applied Optics, 9(7):1669-1671, 1970. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[22] Hajime Nagahara, Sujit Kuthirummal, Changyin Zhou, and Shree K Nayar. Flexible depth of field photography. In European Conference on Computer Vision, pages 60-73. Springer, 2008. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[23] Shree K Nayar and Yasuo Nakagawa. Shape from focus. IEEE Transactions on Pattern analysis and machine intelligence, 16(8):824-831, 1994. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 405, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 447 + ], + "type": "text", + "content": "[24] Ren Ng, Marc Levoy, Mathieu Brédif, Gene Duval, Mark Horowitz, and Pat Hanrahan. Light field photography with a hand-held plenoptic camera. PhD thesis, Stanford university, 2005. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[25] Deepak Pathak, Ross Girshick, Piotr Dollar, Trevor Darrell, and Bharath Hariharan. Learning features by watching objects move. In CVPR, 2017. 14" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 525 + ], + "type": "text", + "content": "[26] Sri Rama Prasanna Pavani and Rafael Piestun. Three dimensional tracking of fluorescent microparticles using a photon-limited double-helix response system. Optics express, 16 (26):22048-22057, 2008. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "type": "text", + "content": "[27] Sri Rama Prasanna Pavani, Michael A Thompson, Julie S Biteen, Samuel J Lord, Na Liu, Robert J Twieg, Rafael Piestun, and William E Moerner. Three-dimensional, single-molecule fluorescence imaging beyond the diffraction limit by using a double-helix point spread function. Proceedings of the National Academy of Sciences, 106(9):2995-2999, 2009. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[28] Yingsi Qin, Wei-Yu Chen, Matthew O'Toole, and Aswin C. Sankaranarayanan. Split-lohmann multifocal displays. ACM Trans. Graph., 42(4), 2023. 2, 3, 11, 14" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[29] Yingsi Qin, Aswin C. Sankaranarayanan, and Matthew O'Toole. Project page: Spatially-varying autofocus. https://imaging.cs.cmu.edu/svaf, 2025.2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Daniel J. Reiley. Nikon AF Nikkor 85mm f/1.4D IF Lens prescription design. https://www.lens-designs.com/photographic-primes, 2017. Licensed under the MIT License. 14" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24653" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[31] Parikshit Sakurikar and P. J. Narayanan. Focal stack representation and focus manipulation. In 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR), pages 250–255, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 294, + 173 + ], + "type": "text", + "content": "[32] Vincent Sitzmann, Steven Diamond, Yifan Peng, Xiong Dun, Stephen Boyd, Wolfgang Heidrich, Felix Heide, and Gordon Wetzstein. End-to-end optimization of optics and image processing for achromatic extended depth of field and superresolution imaging. ACM Trans. Graph., 37(4), 2018. 8, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 217 + ], + "type": "text", + "content": "[33] Muralidhara Subbarao and J-K Tyan. Selecting the optimal focus measure for autofocusing and depth-from-focus. IEEE transactions on pattern analysis and machine intelligence, 20(8):864-870, 1998. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 252 + ], + "type": "text", + "content": "[34] Qilin Sun, Congli Wang, Qiang Fu, Xiong Dun, and Wolfgang Heidrich. End-to-end complex lens design with differentiate ray tracing. ACM Trans. Graph., 40(4), 2021. 8, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 295, + 297 + ], + "type": "text", + "content": "[35] Supasorn Suwajanakorn, Carlos Hernandez, and Steven M Seitz. Depth from focus with your mobile phone. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3497-3506, 2015. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 299, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 299, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 299, + 294, + 330 + ], + "type": "text", + "content": "[36] Dong-Chen Tsai and Homer H. Chen. Reciprocal focus profile. IEEE Transactions on Image Processing, 21(2):459-468, 2012. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 333, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 333, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 294, + 386 + ], + "type": "text", + "content": "[37] Ashok Veeraraghavan, Ramesh Raskar, Amit Agrawal, Ankit Mohan, and Jack Tumblin. Dappled photography: Mask enhanced cameras for heterodyned light fields and coded aperture refocusing. ACM Trans. Graph., 26(3):69, 2007. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 389, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 389, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 57, + 389, + 294, + 443 + ], + "type": "text", + "content": "[38] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 445, + 294, + 488 + ], + "type": "text", + "content": "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 544 + ], + "type": "text", + "content": "[40] Bennett Wilburn, Neel Joshi, Vaibhav Vaish, Eino-Ville Talvala, Emilio Antunez, Adam Barth, Andrew Adams, Mark Horowitz, and Marc Levoy. High performance imaging using large camera arrays. In ACM SIGGRAPH 2005 Papers, pages 765-776. 2005. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 600 + ], + "type": "text", + "content": "[41] Yicheng Wu, Vivek Boominathan, Huaijin Chen, Aswin C. Sankaranarayanan, and Ashok Veeraraghavan. Phasecam3d — Learning phase masks for passive single view depth estimation. In IEEE Intl. Conf. Computational Photography (ICCP), 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "type": "text", + "content": "[42] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[43] Xinge Yang, Matheus Souza, Kunyi Wang, Praneeth Chakravarthula, Qiang Fu, and Wolfgang Heidrich. End-to-end hybrid refractive-diffractive lens design with differentiable ray-wave model. In SIGGRAPH Asia 2024 Con" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 118 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 8, 14" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "text", + "content": "[44] Changyin Zhou, Daniel Miau, and Shree K Nayar. Focal sweep camera for space-time refocusing. 2012. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24654" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_content_list.json b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..36dfc5113b38d37dc2ba24a56ef616f80b8c8328 --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_content_list.json @@ -0,0 +1,1546 @@ +[ + { + "type": "text", + "text": "Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image", + "text_level": 1, + "bbox": [ + 138, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shin Ishihara Imari Sato", + "bbox": [ + 385, + 203, + 611, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "National Institute of Informatics", + "bbox": [ + 370, + 222, + 625, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan", + "bbox": [ + 316, + 239, + 681, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sishihara,imarik}@nii.ac.jp", + "bbox": [ + 372, + 258, + 620, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral imaging has proven effective for appearance inspection because it can identify material compositions and reveal hidden features. Similarly, direct/indirect separation provides essential information about surface appearance and internal conditions, including layer structures and scattering behaviors. This paper presents a novel illumination system incorporating dispersive optics to unify both advantages for scene analysis. In general, achieving distinct direct/indirect separation requires multiple images with varying patterns. In a hyperspectral scenario, using a hyperspectral camera or tunable filters extends exposure and measurement times, hindering practical application. Our proposed system enables the illumination of a wavelength-dependent, spatially shifted pattern. With proper consideration of reflectance differences, we demonstrate that robust separation of direct and indirect components for each wavelength can be achieved using a single hyperspectral image acquired under our single spatiospectral pattern illumination.", + "bbox": [ + 89, + 340, + 483, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 309, + 643, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyperspectral imaging enables pixel-level analysis of objects and is widely used in fields such as medical imaging, food quality control, environmental monitoring, and art conservation [7, 17, 31, 35]. Spectral rendering is gaining traction, with modern renderers adopting wavelength-dependent models [10], and reflection/scattering analysis aiding in material editing [8]. Analyzing layer structures is also essential for inspecting an object's surface and internal conditions. Techniques for separating a scene's direct and indirect reflection components are often employed to accurately capture surface reflection colors and shapes, as well as to visualize an object's internal structure. Nayar et al. developed a groundbreaking fast separation technique [24] by measuring the scene under spatially high-frequency illumination, assuming that the indirect reflection component of a scene changes gradually in space, which means it is predominantly low-frequency.", + "bbox": [ + 511, + 338, + 906, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unifying both advantages for scene analysis, spectral layer structure analysis effectively separates surface and internal details, offering notable benefits such as accurate", + "bbox": [ + 511, + 599, + 906, + 646 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1715ca1ce09a95b21f48a800591fcc298bdf462559a6481de2eebe85884c68cd.jpg", + "image_caption": [ + "Figure 1. The proposed system captures a single hyperspectral image using a wavelength-dependent, spatially shifted illumination pattern generated by a projector with a diffraction grating. The input shows the observed image under the proposed illumination system (displayed as an RGB image) along with the separated direct and indirect reflectance components for each wavelength (\"Direct/Indirect\"). Rough textures and specular reflections are observed in the direct component, while soft textures, such as internal scattering, are observed in the indirect component. The leftmost plots show the comparison of our estimation with the reference at points $p_1$ and $p_2$ ." + ], + "image_footnote": [], + "bbox": [ + 96, + 652, + 901, + 801 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26827", + "bbox": [ + 478, + 935, + 519, + 946 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modeling, realistic rendering, and advanced editing. It also provides a non-destructive, non-invasive way to measure surface and internal spectral properties. This enables a wide range of applications, including medical analysis, such as diagnosing skin abnormalities [40], food inspection [20], and painting analysis for revealing the layered structure of pigments [18]. These analyses can be further enhanced by separating the direct and indirect components, where faster examination is preferred. It also reduces burdens on human subjects and minimizes light exposure for sensitive targets like art and biomedical samples. However, achieving this remains challenging, as a hyperspectral camera requires longer acquisition times than a conventional camera, making it unsuitable for computational photography approaches that rely on multiple pattern irradiations and acquisitions. In recent years, there have been attempts to separate reflection components from a single captured image by methods based on basis representation [30] and deep learning [9], but their method requires a dataset of a large number of images and which is hard to apply for hyperspectral scenarios due to the measurement time of hyperspectral images.", + "bbox": [ + 89, + 90, + 480, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problem of time-consuming hyperspectral imaging, CASSI (Coded Aperture Snapshot Spectral Imaging System) [1] which enables hyperspectral image snapshots [3, 36], techniques for estimating information on other modalities from spectral images [16, 22, 25], and methods for achieving both [2, 29] have been proposed. Snapshot hyperspectral imaging methods are effective for hyperspectral image reconstruction; however, their implementation often requires complex imaging systems, and achieving measurements with sufficient accuracy for further scene analysis necessitates considerable effort. There is an approach that, given a single hyperspectral image, separates the first and second bounce components based on the shape of the observed spectral distribution at each point [25]. However, this method is limited to scenes composed of only a few colors, imposing constraints on its applicability. A more general approach capable of handling complex scenes has yet to be realized. In summary, hyperspectral reconstruction and scene analysis using hyperspectral images have been studied in the past, but a robust solution for achieving spectral direct/indirect separation of complex scenes composed of multiple materials using a single hyperspectral image has yet to be realized.", + "bbox": [ + 91, + 412, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this challenging problem, we develop a novel illumination system that integrates dispersive optics to generate a wavelength-dependent, spatially shifted illumination pattern. Similar to color photometric stereo, where varying lighting directions are achieved through different color channels, our system leverages observed wavelength differences to achieve a shifted illumination pattern. We demonstrate that the challenges associated with hyperspectral-based direct/indirect separation are resolved by our system.", + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By accounting for spectral reflectance differences appropriately, we demonstrate that a single hyperspectral image captured under one illumination pattern can achieve robust separation of direct and indirect components for each wavelength. Furthermore, we show that analyzing variations across wavelengths aids in estimating scene depth.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution is threefold:", + "bbox": [ + 513, + 181, + 709, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel illumination system incorporating dispersive optics is proposed, enabling the illumination of a wavelength-dependent, spatially shifted pattern.", + "- By carefully addressing the differences in spectral reflectance, we formulate the observed pixel values based on these spectral-dependent illumination patterns and derive a solution for direct/indirect separation at each wavelength using a single hyperspectral image with one illumination pattern.", + "- We develop a real system and demonstrate robust spectral direct/indirect separation across various objects. Additionally, we show that analyzing differences across wavelengths aids in estimating scene depth." + ], + "bbox": [ + 513, + 196, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 407, + 661, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Separation of Direct and Indirect Reflection Components via Pattern Illumination", + "text_level": 1, + "bbox": [ + 511, + 431, + 903, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nayar et al. [24] proposed separating the direct and indirect reflection components of a scene using high-frequency illumination patterns, and various extended methods have also been proposed. As the applications for acquiring rich information, O'Toole et al. [26, 27] proposed an image measurement method that can observe light transport paths in a scene by modulating both the illumination pattern of the projector and the camera response. As the high-speed application, separation of direct and indirect reflections by time dithering of a DLP projector with a high-speed camera was proposed [23], and extended to multiple colors [34]. More recently, a fast-capturing and processing method based on machine learning using event data was proposed [38] by a hybrid camera system combining an event camera and an RGB camera. Thus, many methods have been proposed to speed up the separation method of direct and indirect reflection components using pattern illumination, and extensions to multispectral with a small number of channels have been made, but no extension of patterns to hyperspectral image acquisition or spectral dimension has been achieved.", + "bbox": [ + 511, + 462, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, several methods have been proposed to separate the direct and indirect reflection components of a scene using a single image. Subpa-Asa et al. [30] propose to evaluate the data-independent Fourier basis and learning-based principal component analysis (PCA) basis to find a better linear basis representation of the direct and indirect components. Duan et al. [9] proposed a deep learning solution for the separation of reflective components from a single image by designing a non-binary pattern with 4 or", + "bbox": [ + 511, + 765, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26828", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "9 intensities instead of a binary grid pattern. Although the method of separating reflection components from a single image has theoretical applicability to hyperspectral images, in practice it is difficult to apply because it requires a data set with many images.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decomposition-based Spectral Reflectance Separation", + "text_level": 1, + "bbox": [ + 89, + 170, + 470, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Separation of the direct and indirect spectral reflectance can also be treated as a spectral decomposition problem. Several studies based on spectral decomposition problems using hyperspectral images have been proposed. In the field of remote sensing, linear or nonlinear spectral decomposition methods have been proposed to address the problems of spectral variation due to the atmosphere and/or spectral mixing due to the imaging system [4]. In the field of computer vision, Zheng et al. [37] has enabled spectral separation between illumination and reflectance in hyperspectral images taken under common spectral illumination and extended to multiple light sources by Chen et al. [5]. Regarding the separation of specular reflection components on the target surface using spectral information, an analysis of chromaticity-intensity space in diffuse and specular reflections [33], a non-iterative method based on a dichroic reflection model and PCA [11], and a method based on blind source separation using non-negative matrix factorization [39] have been proposed.", + "bbox": [ + 91, + 185, + 483, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the application of separating indirect reflection, Okawa et al. [25] proposed a method to separate interreflection components from a single hyperspectral image when the object color is one or two colors, based on the fact that not only surfaces of different colors but also surfaces of the same color change the observed spectrum due to interreflection. Moreover, Fu et al. [12] proposed a method to separate the reflection and fluorescence emission spectra in a scene by using high-frequency illumination in the spectral domain. Their method achieves high-frequency illumination in the spectral domain using an equalized light source, whose wavelength characteristics can be freely controlled, but it is uniform in the spatial domain and cannot be used to separate direct and indirect reflection components. Thus, while several possible ways to separate reflection components using hyperspectral images exist, only interreflection or fluorescence has been considered for the separation of indirect components.", + "bbox": [ + 91, + 474, + 483, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dispersive Optics", + "text_level": 1, + "bbox": [ + 89, + 750, + 217, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dispersive optics, such as prisms and diffraction gratings, have often been used in camera and display design, especially for hyperspectral imaging. Mohan et al. [21] proposed a spectral mask that can control the spectral power distribution of a camera or projector using a diffraction grating and an attenuation mask. Hostettler et al. [15] used multiple prisms and calculated masks to realize a projector system that projects the desired color image without a conventional projector. CASSI uses dispersion optics and coding", + "bbox": [ + 89, + 763, + 480, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "masks to acquire masked spectral images with wavelength-dependent transformations and reconstructs the hyperspectral data cube using an iterative algorithm [1, 36]. Cao et al. [3] achieved fast hyperspectral image acquisition with a hybrid camera system constructed from an RGB camera, a grayscale camera, and multiple optical elements including prisms and coding masks.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several methods have been proposed to acquire spectral and 3D depth information using dispersive optics. Baek et al. [2] proposed the diffractive optical element (DOE) generates a unique point spread function (PSF) that changes with the scene's spectrum and depth, enabling single-shot hyperspectral 3D imaging through blur analysis. Shin et al. [29] proposed a system to obtain spectral and 3D depth information by analyzing which structured light patterns disperse and illuminate a scene, using a compact setting that adds diffraction grating film to a projector-camera system.", + "bbox": [ + 511, + 196, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dispersive optics have been utilized for projector/camera enhancement, hyperspectral image acquisition, and depth estimation. This field holds great potential for various applications. In this study, we demonstrate the effectiveness of direct/indirect separation.", + "bbox": [ + 511, + 348, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 434, + 604, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We separate the direct and indirect reflection components of a scene from a single-shot hyperspectral image taken under a wavelength-dependent, spatially shifted illumination pattern generated by a DMD (Digital Micromirror Device) projector with a diffraction grating (Figure 1). We first revisit the traditional approach to direct and indirect reflections, utilizing varying high-frequency projected patterns, and then extend the discussion to our case, which involves a wavelength-dependent, spatially shifted illumination pattern.", + "bbox": [ + 511, + 459, + 905, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Direct/Indirect Separation Using Sinusoidal Pattern Illumination", + "text_level": 1, + "bbox": [ + 511, + 617, + 903, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a single light source illuminates a scene, the direct component $I_{d}(x)$ is defined as the light that travels directly from the source, reflects once at a surface point, and reaches the camera, while the indirect component $I_{i}(x)$ encompasses all light that undergoes multiple interactions, such as subsurface scattering, interreflections, and volumetric scattering. Nayar et al. [24]. The observed intensity $I(x)$ by a camera is represented as", + "bbox": [ + 511, + 654, + 905, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x) = I _ {d} (x) + I _ {i} (x). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 790, + 903, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming that the indirect reflection component of a scene changes gradually in space and is predominantly low-frequency, their proposed approach [24] demonstrated that the scene's direct and indirect reflection components can be separated by irradiating three spatially high-frequency sinusoidal patterns.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26829", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14a5663ce8d325f61e68b049760b56f8df24e403ff3e83ef01870b50a816ad07.jpg", + "image_caption": [ + "Figure 2. The outline of the proposed spatio-spectral illumination system. (a) The (DMD) projector generates a sinusoidal pattern that has a broad spectral band of light. (b) The diffraction grating disperses the generated light pattern, then the sinusoidal pattern is spatially shifted according to wavelength. (c) The spatio-spectral shifted patterns are observed at each point of the scene, and the sinusoidal pattern is resampled in the spectral domain." + ], + "image_footnote": [], + "bbox": [ + 93, + 93, + 908, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under the high-frequency sinusoidal pattern illuminations, the observed images $I_{n}(x,k)$ are represented as:", + "bbox": [ + 89, + 333, + 482, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {n} (x, k) = \\frac {1}{2} I _ {d} (x) (\\sin (\\frac {2}{3} k \\pi + \\phi) + 1) + \\frac {1}{2} I _ {i} (x), \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 375, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi \\in [0,2\\pi]$ denotes an arbitrary phase of the sinusoidal function and $k = \\{1,2,3\\}$ represents the index of the pattern. This problem has three unknowns $\\{I_d,I_i,\\phi \\}$ and can be solved theoretically for each pixel using three images taken under three different sinusoidal pattern illumination. In practice, this approach requires pixel-by-pixel camera-projector alignment or additional patterns to avoid artifacts and ensure accuracy.", + "bbox": [ + 89, + 414, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spectral Direct/Indirect Separation", + "text_level": 1, + "bbox": [ + 89, + 545, + 398, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let us further extend the discussion to separation at the wavelength level by considering the spectral reflection of a scene. Here, we represent the direct reflection component $I_{d}(x,\\lambda)$ as the product of normalized spectral reflectance of direct component $R_{d}(x,\\lambda)$ , its magnitude $\\alpha_{d}(x)$ , and the intensity of the spectral light source $L_{d}(x,\\lambda)$ at the point $x$ .", + "bbox": [ + 89, + 568, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [24], the indirect reflection component is represented as:", + "bbox": [ + 89, + 659, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I _ {i} (x, \\lambda) = \\sum_ {x ^ {\\prime} \\in S (x)} A ([ x, x ^ {\\prime} ], \\lambda) L ([ x, x ^ {\\prime} ], \\lambda) \\tag {3} \\\\ = \\alpha_ {i} (x) R _ {i} (x, \\lambda), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 696, + 482, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S(x)$ denotes the set of surrounding pixels for $x$ , $L([x, x'], \\lambda)$ represents the spectral radiance at $x'$ in the direction of $x$ , and $A([x, x'], \\lambda)$ incorporates the spectral BRDF of $x$ as well as the relative geometric configuration between $x$ and $x'$ .1 $I_i(x, \\lambda)$ is further expressed in terms of the normalized spectrum $R_i(x, \\lambda)$ which is computed from the total sum of light contributions coming to $x$ from surrounding pixels $S(x)$ and its magnitude $\\alpha_i(x)$ . From this,", + "bbox": [ + 89, + 743, + 483, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the spectral reflection model of a scene $I(x,\\lambda)$ is given as:", + "bbox": [ + 511, + 333, + 901, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I (x, \\lambda) = I _ {d} (x, \\lambda) + I _ {i} (x, \\lambda) \\\\ = \\alpha_ {d} (x) R _ {d} (x, \\lambda) L _ {d} (x, \\lambda) + \\alpha_ {i} (x) R _ {i} (x, \\lambda). \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 359, + 916, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following, we introduce a spatial phase shift by designing the lighting pattern to shift horizontally with wavelength.", + "bbox": [ + 511, + 407, + 905, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Spatially Shifting Illumination According to Wavelength", + "text_level": 1, + "bbox": [ + 511, + 463, + 905, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a spatio-spectral pattern illumination system that combines a conventional projector and a diffraction grating to create spatially shifting illumination patterns according to wavelength. The configuration of the illumination system is shown in Figure 2. Let the spatially high-frequency sinusoidal pattern $P_0(v)$ generated by the conventional projector (Figure 2 (a)) be written as a sinusoidal function of the horizontal pixel position $v$ of the projector:", + "bbox": [ + 511, + 500, + 905, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {0} (v) = \\frac {l _ {0}}{2} \\left(\\sin \\left(f _ {0} v + \\phi_ {0}\\right) + 1\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 633, + 905, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l_{0}$ is the intensity of the projector, $f_{0}$ describes the frequency of the sinusoidal function, and $\\phi_0$ denotes the phase of the sinusoidal function. Then the grating diffracts the incoming pattern light and reflects it at slightly different angles for each wavelength of light (Figure 2 (b)). In the case of first-order diffracted light, the diffraction angle of the outgoing light, i.e. the emission angle, varies monotonically in the visible light range depending on the wavelength of the incident light. Thus, the illumination pattern varies in the spatio-spectral domain, similar to a rainbow. Therefore, as shown in Figure 3, when the diffracted pattern is observed at a scene point, it is also a sinusoidal pattern in the spectral domain (Figure 2 (c)). We proceed with the discussion assuming that projector's origin is the virtual origin of the pattern reflected by the grating. We also set the cam", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "In our system, the scene brightness has been calibrated by considering the camera's spectral sensitivity.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26830", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "era's origin at the same location to enable discussion within a unified coordinate system.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d064957930dd665ff8b729a3c4eb8977dc9ba020ab92c81794fb1ecdfc16716b.jpg", + "image_caption": [ + "Figure 3. After passing through the diffraction grating, the sinusoidal pattern generated by the conventional projector (following yellow arrow) is resampled in the spectral domain at each point of the scene (following red arrow)." + ], + "image_footnote": [], + "bbox": [ + 94, + 133, + 475, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spectral pattern $P(x, \\lambda)$ is expressed as the function of wavelength $\\lambda$ at the scene point corresponding to the projector's horizontal pixel position $x$ as in Equation (6):", + "bbox": [ + 89, + 316, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (x, \\lambda) = \\frac {l (\\lambda)}{2} \\left(\\sin \\left(f _ {p} (x) \\lambda + \\phi (x)\\right) + 1\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 368, + 482, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $l(\\lambda)$ be the spectral intensity of the projector, $f_{p}(x)$ denote the frequency of the sinusoidal function, and $\\phi (x)$ represent the phase of the sinusoidal function. We calibrate the spectral intensity of the projector $l(\\lambda)$ in advance and describe it as the constant value $l$ .", + "bbox": [ + 89, + 402, + 483, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Simple Case", + "text_level": 1, + "bbox": [ + 89, + 484, + 223, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin our discussion with the simplest case: illuminating our pattern onto a gray object with flat spectral reflectance and spatially varying albedo. In this case, the normalized spectra $R_{d}$ and $R_{i}$ of the direct and indirect components can be integrated and represented as the respective magnitudes $\\alpha_{d}$ and $\\alpha_{i}$ . When illuminating our pattern onto this object, the direct light component $L_{d}(x,\\lambda)$ in Equation (4) is replaced by $P(x,\\lambda)$ , and the indirect reflection component $I_{i}(x,\\lambda)$ is reduced to half intensity and pattern independent following [24]. Hence, the observed reflection $I_{p}(x,\\lambda)$ under the spatio-spectral pattern illumination $P(x,\\lambda)$ is modeled as:", + "bbox": [ + 89, + 506, + 483, + 689 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p} (x, \\lambda) = \\alpha_ {d} (x) P (x, \\lambda) + \\frac {1}{2} \\alpha_ {i} (x). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 693, + 482, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expanding Equation (7) based on Equation (6) yields:", + "bbox": [ + 89, + 726, + 447, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p} (x, \\lambda) = \\frac {l \\alpha_ {d} (x)}{2} (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) + \\frac {\\alpha_ {i} (x)}{2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 746, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this case, we can obtain a sinusoidal pattern in the spectral domain along wavelength axis at any point $x$ as shown in Figure 4 (a).", + "bbox": [ + 89, + 780, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the number of unknowns is four $\\{\\alpha_{d}(x),f_{p}(x),\\phi (x),\\alpha_{i}(x)\\}$ , the direct and indirect reflection components of a scene can be separated by acquiring spectral images at least 4 spectral channels in the simple case.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5e46589b634ec966a4df529ed3d186e4536c3caf47cfc6002ff72a02551e00a1.jpg", + "image_caption": [ + "Figure 4. Our illumination pattern and the observation as it is projected onto the scene. (a) For gray objects, a sinusoidal pattern is observed in the spectral domain. (b) For colored objects, a pattern that varies with the spectral reflectance of the direct and indirect components of the object is observed in the spectral domain." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 624, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6ece85625d22b329786e6cf62d587ac491bc95261557d072622a400e03247db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 90, + 763, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c7a0cdda9049d4f793a423b4695d3ffe63b042834e79b05011d413c122730b99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 90, + 898, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Complex Case", + "text_level": 1, + "bbox": [ + 511, + 301, + 661, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, let us examine the case of the object that has non-flat, spatially varying spectral reflectance. As shown in Figure 4 (b), the spectrum observed for a scene with non-flat spectral reflectance is affected by the spectral reflectance of the scene, unlike the simple flat reflectance case in Figure 4 (a).", + "bbox": [ + 511, + 324, + 905, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Equation (4), a scene can be composed of direct and indirect reflection components, each of which can have its spectral distribution. The observed intensity $I_{p}(x,\\lambda)$ at the scene point $x$ under the spatio-spectral pattern illumination $P(x,\\lambda)$ can be expressed:", + "bbox": [ + 511, + 415, + 905, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {p} (x, \\lambda) = \\alpha_ {d} (x) R _ {d} (x, \\lambda) P (x, \\lambda) + \\frac {\\alpha_ {i} (x)}{2} R _ {i} (x, \\lambda) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 503, + 906, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expanding Equation (9) based on Equation (6) yields:", + "bbox": [ + 511, + 542, + 870, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I _ {p} (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) / 2 \\\\ + \\quad \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 570, + 916, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, Equation (10) can be expressed in the form of a general sinusoidal function for $\\lambda$ as follows:", + "bbox": [ + 511, + 616, + 906, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI_{p}(x,\\lambda) = A(x,\\lambda)\\{\\sin (B(x)\\lambda +C(x)) + 1\\} +D(x,\\lambda),\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 657, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) / 2, B (x) = f _ {p} (x),\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 689, + 864, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (x) = \\phi (x), D (x, \\lambda) = \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2.\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 705, + 857, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(11)", + "bbox": [ + 874, + 720, + 903, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Equation (11) shows that the amplitude of the sinusoidal function is related to the direct reflection component and the constant term is related to the indirect reflection component.", + "bbox": [ + 511, + 734, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then discuss how to separate the direct and indirect reflection components from the observed values based on Equation (11). The number of observations at each pixel obtained in a single hyperspectral image measurement equals the number of spectral wavelength observations. Therefore, if either spectral reflectance $R_{d}$ and $R_{i}$ are unknown, solving Equation (11) from a single multi- or hyperspectral input image is an ill-posed problem. We address this problem", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26831", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by representing $R_{d}$ and $R_{i}$ as the combination of the spectral reflectance basis, following the work of Han et al. [14]:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {R}} _ {\\boldsymbol {d}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {d}} (\\boldsymbol {x}), \\bar {\\boldsymbol {R}} _ {\\boldsymbol {i}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {i}} (\\boldsymbol {x}), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 128, + 482, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\bar{R}_d(x)$ and $\\bar{R}_i(x) \\in \\mathbb{R}^f$ are the vectorized form of the normalized spectral reflectances at the pixel $x$ approximated by the basis representation, $B \\in \\mathbb{R}^{f \\times k}$ is a basis matrix consisting of $k$ basis vectors, and $c_d(x)$ and $c_i(x) \\in \\mathbb{R}^k$ are the basis coefficients at the pixel $x$ . $\\bar{R}_i(x)$ is not pure spectral reflectance, but it is the sum of the spectral BRDFs in the surrounding pixels and can therefore be expressed on the same basis. Based on the public dataset [28], in which the 1257 spectral reflectances of Munsell Book of Color: Matte Finish Collection [6] are recorded, $k$ spectral reflectance basis vectors are extracted using singular value decomposition. Approximation by the basis reduces the number of unknowns related to the spectral reflectance to $k$ . Thus, unknowns of our problem are $\\{c_d(x), c_i(x), \\alpha_d(x), f_p(x), \\phi(x), \\alpha_i(x)\\}$ and its total degree of freedom is $2k + 4$ .", + "bbox": [ + 89, + 155, + 483, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Optimization", + "text_level": 1, + "bbox": [ + 89, + 407, + 230, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We solve the problem described in Section 3.5 by optimization, using a single hyperspectral image captured by our spatio-spectral pattern illumination system as input. For simplicity, let us focus on a single pixel, with all terms written while omitting $x$ . The error function between the input spectral image and the model is computed as:", + "bbox": [ + 89, + 429, + 483, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nE _ {s} (\\lambda) = \\left| I _ {p} (\\lambda) - \\frac {l \\alpha_ {d}}{2} \\bar {R} _ {d} (\\lambda) \\left\\{\\sin \\left(f _ {p} \\lambda + \\phi\\right) + 1 \\right\\} - \\frac {\\alpha_ {i}}{2} \\bar {R} _ {i} (\\lambda) \\right|. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 527, + 482, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\bar{R}_d(\\lambda)$ and $\\bar{R}_i(\\lambda)$ denote the scalar form of the normalized spectral reflectances approximated by the basis representation. The objective function of the problem is:", + "bbox": [ + 89, + 569, + 483, + 614 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, \\alpha_ {d}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 625, + 482, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As mentioned earlier, our problem setup has $2k + 4$ unknowns, so the simplest solution is to give a hyperspectral image with $2k + 4$ spectral channels as the input image. From previous study [14], the number of bases for approximating the spectral reflectance of a general object with a basis representation is about 8. Therefore, when we follow them and let $k = 8$ , we require a hyperspectral image with 20 spectral channels to solve. When applying hyperspectral imaging to digital archiving and other applications, it is common to work in visible light, with wavelengths ranging from $400\\mathrm{nm}$ to $700\\mathrm{nm}$ . Therefore, acquiring multi- or hyperspectral images with a wavelength interval of at least 15 nm in the visible light range is sufficient.", + "bbox": [ + 89, + 674, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For robust optimization, we use initial values and ranges of estimation values for some of the parameters. As shown", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in the figure, the period of the sinusoidal function of the spectral domain at each pixel can be predicted from the local maxima and minima of the observed signal. The period obtained from the median of the intervals at the local maxima and minima is the initial value of the period $f_{p}$ at that pixel, and the search range for optimization is $\\pm 5\\%$ of the initial value to account for discretization error. Also, the search range of the phase $\\phi$ is $[-\\pi, \\pi]$ and the magnitude of the direct and indirect components $\\alpha_{d}$ and $\\alpha_{i}$ are $[0, \\infty]$ .", + "bbox": [ + 511, + 90, + 906, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a1bb872fc1433ff66bfc24296c2926eb3f5d29862117543c801aac4b8aaacafc.jpg", + "image_caption": [ + "Figure 5. Examples of the observed signal in the spectral domain. Orange and green circles represent local maxima and minima." + ], + "image_footnote": [], + "bbox": [ + 519, + 244, + 900, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, since spectral reflectance is known to be smooth for most common objects [13], we introduce a loss function that reflects the assumption of smoothness of spectral reflectance, following prior works [19, 32]:", + "bbox": [ + 511, + 390, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda) + \\gamma_ {d} | \\boldsymbol {B c} _ {d} \\boldsymbol {\\Gamma} | + \\gamma_ {i} | \\boldsymbol {B c} _ {i} \\boldsymbol {\\Gamma} |. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 464, + 905, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The weights $\\gamma_{d}$ and $\\gamma_{i}$ are chosen empirically and $\\Gamma$ denotes a matrix that gives smoothness by the subtraction between the neighbor elements.", + "bbox": [ + 511, + 517, + 905, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 579, + 718, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup and Conditions", + "text_level": 1, + "bbox": [ + 511, + 604, + 823, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our system consists of a xenon light source (Asahi Spectra LAX-100), a DMD (Ajile AJD-4500), a hyperspectral camera (EBA JAPAN NH-8), lenses for the DMD and hyperspectral camera (Fujinon HF25SA-1), and a diffraction grating (Edmund Optics 1200 Grooves, $50\\mathrm{mm}$ Square, $500\\mathrm{nm}$ Ruled Diffraction Grating), as shown in Figure 1. The hyperspectral camera captures spectral images at $5\\mathrm{nm}$ intervals within the $380 - 1000\\mathrm{nm}$ wavelength range. Considering the effective wavelength range of the diffraction grating, we performed separation in the $450 - 650\\mathrm{nm}$ range.", + "bbox": [ + 511, + 627, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The minimum number of observations required to solve our problem is 20, as discussed in Section 3.6. However, actual spectral images contain significant noise, and factors such as projector and camera calibration errors further impact accuracy, preventing us from achieving stable results with the minimum number of observations. In the following experiments, we use 41 spectral bands obtained at $5\\mathrm{nm}$ intervals to ensure robust separation against noise.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26832", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on Real Scenes", + "text_level": 1, + "bbox": [ + 89, + 90, + 302, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 8 shows the results of estimating the spectral reflectance of the direct and indirect reflectance components in various real-world objects. For reference, we captured a set of images of the scene using a hyperspectral camera with different sinusoidal pattern illuminations and provided the separation results using Nayar's method [24], as described in Equation (2). Hyperspectral images contain significant noise, making it challenging to obtain stable separation results free from artifacts by Nayar's method [24] unless numerous hyperspectral images are captured under various sinusoidal pattern illuminations. To mitigate such artifacts, the spatial resolution of the input images for the reference results, but not ours, is downsampled for visual comparison. All input hyperspectral images and estimated results are available in the supplementary material.", + "bbox": [ + 89, + 112, + 483, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For all four scenes in Figure 8, the estimated direct $(I_d)$ and indirect $(I_i)$ components closely resemble the corresponding reference across different wavelengths. From these examples, we observe that wavelength-dependent changes in the direct and indirect components are extracted. In particular, in the results for the Potato and Ball (top 2 rows) and the Candle (bottom 2 rows), the scattering inside the objects becomes broader and stronger for longer wavelengths. Fine textures can be seen in the direct reflection component and rarely in the indirect reflection component. The leftmost graph compares our estimation and reference for points $p_1$ and $p_2$ . The estimated spectra well represent that of the reference; nevertheless, they are derived solely from a single hyperspectral image. Moreover, Figure 6 shows spatial maps and values of error metrics. We used MSE (mean squared error) and SAM (spectral angle mapper) as error metrics.", + "bbox": [ + 89, + 339, + 483, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Depth Dependence of the Pattern's Period", + "text_level": 1, + "bbox": [ + 89, + 608, + 449, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 shows the spatial distribution of the estimated period $\\phi$ in Equation (15). The period $\\phi$ increases as the distance from the projector becomes longer, making it highly correlated with object distance. In Figure 7, 'DAY' is placed closer to the camera, and 'HAP' is farther away. This example demonstrates that the estimated $\\phi$ contains valuable information about the scene's depth.", + "bbox": [ + 89, + 631, + 483, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 753, + 209, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we proposed a novel method for separating the direct and indirect reflection components of a scene from a single hyperspectral image. By introducing a projection system incorporating dispersive optics, we enabled the projection of a wavelength-dependent, spatially shifted pattern. We formulated the observed pixel values under these spectral-dependent illumination patterns and derived a solution for direct and indirect reflection separation at each", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a92b5762de635944f01719392b9a1a340c309b9151352a3a1bfd5bffc1a21b24.jpg", + "image_caption": [ + "Figure 6. Error maps and metrics between the reference and the proposed method. The values represent the mean and variance of the MSE and SAM. Note that some stripe patterns in the SAM maps are affected by the artifacts of the reference." + ], + "image_footnote": [], + "bbox": [ + 516, + 85, + 903, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32579dbc8e51b3945fb1fc27c38a51786605a9040eb48fb54323874eaceb8086.jpg", + "image_caption": [ + "Figure 7. Spatial map of the estimated period $\\phi$ for a scene. 'DAY' is placed closer to the camera, and 'HAP' is farther away." + ], + "image_footnote": [], + "bbox": [ + 524, + 500, + 903, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "wavelength using only a single hyperspectral image with one projection pattern. Through experiments, we demonstrated the effectiveness of our system in achieving robust spectral direct/indirect separation even for objects with complex spectral reflectance. Furthermore, our analysis of differences across wavelengths was leveraged for scene depth estimation.", + "bbox": [ + 511, + 705, + 906, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 828, + 671, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This research is supported by JSPS Kakenhi Grant number 25K21236 and JST-Mirai Program Grant Number JP-MJM123G1.", + "bbox": [ + 511, + 854, + 905, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26833", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15acc3230c29c7aa979a14b3f2ef66c2d15b148061df512cf918d9d524a3433e.jpg", + "image_caption": [ + "Figure 8. Separation results of direct and indirect reflection components in various real-world objects. Top row, from left to right: scene captured by the hyperspectral camera (RGB converted), results of separation of direct and indirect reflection components $(I_d(480), I_d(620), I_i(480), I_i(620))$ by Nayar's method [24] as a reference, direct reflectance component (red) at point $p_1$ and indirect, comparison of spectral reflectance of direct (red) and indirect (green) components at point $p_1$ for the reference (dotted line) and the proposed method (solid line). Bottom row, from left to right: part of the input spectral images in the proposed method are superimposed on a pseudo-RGB, estimation results of direct and indirect reflection components $(I_d(480), I_d(620), I_i(480), I_i(620))$ by the proposed method, comparison of spectral reflectance at $p_2$ ." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 851, + 785 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26834", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gonzalo R Arce, David J Brady, Lawrence Carin, Henry Arguello, and David S Kittle. Compressive coded aperture spectral imaging: An introduction. IEEE Signal Processing Magazine, 31(1):105-115, 2013. 2, 3", + "[2] Seung-Hwan Baek, Hayato Ikoma, Daniel S Jeon, Yuqi Li, Wolfgang Heidrich, Gordon Wetzstein, and Min H Kim. Single-shot hyperspectral-depth imaging with learned diffractive optics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2651-2660, 2021. 2, 3", + "[3] Xun Cao, Xin Tong, Qionghai Dai, and Stephen Lin. High resolution multispectral video capture with a hybrid camera system. In CVPR 2011, pages 297-304. IEEE, 2011. 2, 3", + "[4] Xavier Ceamanos and Silvia Valero. 4 - processing hyperspectral images. In Optical Remote Sensing of Land Surface, pages 163-200. Elsevier, 2016. 3", + "[5] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 3", + "[6] Munsell Color. Munsell book of color: Matte finish collection: Kollmorgen corporation. Baltimore, Maryland, 1976. 6", + "[7] Costanza Cucci, John K Delaney, and Marcello Picollo. Reflectance hyperspectral imaging for investigation of works of art: old master paintings and illuminated manuscripts. Accounts of chemical research, 49(10):2070-2079, 2016. 1", + "[8] Craig Donner and Henrik Wann Jensen. A spectral bssrfd for shading human skin. Rendering techniques, 2006:409-418, 2006. 1", + "[9] Zhaoliang Duan, James Bieron, and Pieter Peers. Deep Separation of Direct and Global Components from a Single Photograph under Structured Lighting. Computer Graphics Forum, 2020. 2", + "[10] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Transactions on graphics (TOG), 37(6):1-14, 2018. 1", + "[11] Gamal ElMasry, Pere Gou, and Salim Al-Rejaie. Effectiveness of specularity removal from hyperspectral images on the quality of spectral signatures of food products. Journal of Food Engineering, 289:110148, 2021. 3", + "[12] Ying Fu, Antony Lam, Imari Sato, Takahiro Okabe, and Yoichi Sato. Separating reflective and fluorescent components using high frequency illumination in the spectral domain. In Proceedings of the IEEE International Conference on Computer Vision, pages 457-464, 2013. 3", + "[13] Lewis D Griffin. Reconciling the statistics of spectral reflectance and colour. PLoS One, 14(11):e0223069, 2019. 6", + "[14] Shuai Han, Imari Sato, Takahiro Okabe, and Yoichi Sato. Fast spectral reflectance recovery using dlp projector. International journal of computer vision, 110:172-184, 2014. 6", + "[15] Rafael Hostettler, Ralf Habel, Markus Gross, and Wojciech Jarosz. Dispersion-based color projection using masked prisms. In Computer Graphics Forum, pages 329-338. Wiley Online Library, 2015. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Shin Ishihara, Yuta Asano, Yinqiang Zheng, and Imari Sato. Underwater scene recovery using wavelength-dependent refraction of light. In 2020 International Conference on 3D Vision (3DV), pages 32-40. IEEE, 2020. 2", + "[17] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31 (4):1-11, 2012. 1", + "[18] Paul Kubelka and Franz Munk. An article on optics of paint layers. Z. Tech. Phys, 12(593-601):259-274, 1931. 2", + "[19] Wei Liu, Chengxun He, and Le Sun. Spectral-smoothness and non-local self-similarity regularized subspace low-rank learning method for hyperspectral mixed denoising. Remote Sensing, 13(16):3196, 2021. 6", + "[20] Lembe Samukelo Magwaza and Umezuruike Linus Opara. Analytical methods for determination of sugars and sweetness of horticultural products—a review. Scientia Horticulturae, 184:179-192, 2015. 2", + "[21] Ankit Mohan, Ramesh Raskar, and Jack Tumblin. Agile spectrum imaging: Programmable wavelength modulation for cameras and projectors. In Computer graphics forum, pages 709-717. Wiley Online Library, 2008. 3", + "[22] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE computer graphics and applications, 34(6):57-68, 2014. 2", + "[23] Srinivasa G Narasimhan, Sanjeev J Koppal, and Shuntaro Yamazaki. Temporal dithering of illumination for fast active vision. In Computer Vision-ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10, pages 830-844. Springer, 2008. 2", + "[24] Shree K Nayar, Gurunandan Krishnan, Michael D Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. In ACM SIGGRAPH 2006 Papers, pages 935-944. 2006. 1, 2, 3, 4, 5, 7, 8", + "[25] Hiroki Okawa, Yinqiang Zheng, Antony Lam, and Imari Sato. Spectral reflectance recovery with interreflection using a hyperspectral image. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13, pages 52-67. Springer, 2017. 2, 3", + "[26] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2", + "[27] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 2", + "[28] Jussi PS Parkkinen, Jarmo Hallikainen, and Timo Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical society of America A, 6(2):318-322, 1989. 6", + "[29] Suhyun Shin, Seokjun Choi, Felix Heide, and Seung-Hwan Baek. Dispersed structured light for hyperspectral 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24997-25006, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26835", + "bbox": [ + 478, + 935, + 517, + 946 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Art Subpa-asa, Ying Fu, Yinqiang Zheng, Toshiyuki Amano, and Imari Sato. Separating the direct and global components of a single image. Journal of Information Processing, 26: 755-767, 2018. 2", + "[31] Da-Wen Sun. Hyperspectral imaging for food quality analysis and control. Elsevier, 2010. 1", + "[32] Le Sun, Qihao Cheng, and Zhiguo Chen. Hyperspectral image super-resolution method based on spectral smoothing prior and tensor tubal row-sparse representation. Remote Sensing, 14(9):2142, 2022. 6", + "[33] Robby T Tan, Ko Nishino, and Katsushi Ikeuchi. Separating reflection components based on chromaticity and noise analysis. IEEE transactions on pattern analysis and machine intelligence, 26(10):1373-1379, 2004. 3", + "[34] Moriaki Torii, Takahiro Okabe, and Toshiyuki Amano. Multispectral direct-global separation of dynamic scenes. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1923-1931. IEEE, 2019. 2", + "[35] Mahesh Kumar Tripathi and H Govil. Evaluation of aviris-ng hyperspectral images for mineral identification and mapping. Heliyon, 5(11), 2019. 1", + "[36] Ashwin Wagadarikar, Renu John, Rebecca Willett, and David Brady. Single disperser design for coded aperture snapshot spectral imaging. Applied optics, 47(10):B44-B51, 2008. 2, 3", + "[37] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 3", + "[38] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. Evdig: Event-guided direct and global components separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9612-9621, 2024. 2", + "[39] Mustapha Zokay and Hicham Saylani. Removing specular reflection in multispectral dermatological images using blind source separation. In Annual Conference on Medical Image Understanding and Analysis, pages 734-750. Springer, 2022. 3", + "[40] George Zonios, Julie Bykowski, and Nikiforos Kollias. Skin melanin, hemoglobin, and light scattering properties can be quantitatively assessed in vivo using diffuse reflectance spectroscopy. Journal of Investigative Dermatology, 117(6): 1452-1457, 2001. 2" + ], + "bbox": [ + 91, + 92, + 480, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26836", + "bbox": [ + 478, + 936, + 517, + 946 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_model.json b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2830eafb0aca3bd7772576d143db91f9aac0af54 --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.13, + 0.86, + 0.177 + ], + "angle": 0, + "content": "Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.204, + 0.612, + 0.221 + ], + "angle": 0, + "content": "Shin Ishihara Imari Sato" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.223, + 0.627, + 0.238 + ], + "angle": 0, + "content": "National Institute of Informatics" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.24, + 0.682, + 0.258 + ], + "angle": 0, + "content": "2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.26, + 0.621, + 0.275 + ], + "angle": 0, + "content": "{sishihara,imarik}@nii.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.629 + ], + "angle": 0, + "content": "Hyperspectral imaging has proven effective for appearance inspection because it can identify material compositions and reveal hidden features. Similarly, direct/indirect separation provides essential information about surface appearance and internal conditions, including layer structures and scattering behaviors. This paper presents a novel illumination system incorporating dispersive optics to unify both advantages for scene analysis. In general, achieving distinct direct/indirect separation requires multiple images with varying patterns. In a hyperspectral scenario, using a hyperspectral camera or tunable filters extends exposure and measurement times, hindering practical application. Our proposed system enables the illumination of a wavelength-dependent, spatially shifted pattern. With proper consideration of reflectance differences, we demonstrate that robust separation of direct and indirect components for each wavelength can be achieved using a single hyperspectral image acquired under our single spatiospectral pattern illumination." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.31, + 0.645, + 0.325 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Hyperspectral imaging enables pixel-level analysis of objects and is widely used in fields such as medical imaging, food quality control, environmental monitoring, and art conservation [7, 17, 31, 35]. Spectral rendering is gaining traction, with modern renderers adopting wavelength-dependent models [10], and reflection/scattering analysis aiding in material editing [8]. Analyzing layer structures is also essential for inspecting an object's surface and internal conditions. Techniques for separating a scene's direct and indirect reflection components are often employed to accurately capture surface reflection colors and shapes, as well as to visualize an object's internal structure. Nayar et al. developed a groundbreaking fast separation technique [24] by measuring the scene under spatially high-frequency illumination, assuming that the indirect reflection component of a scene changes gradually in space, which means it is predominantly low-frequency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.601, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Unifying both advantages for scene analysis, spectral layer structure analysis effectively separates surface and internal details, offering notable benefits such as accurate" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.654, + 0.903, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.814, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Figure 1. The proposed system captures a single hyperspectral image using a wavelength-dependent, spatially shifted illumination pattern generated by a projector with a diffraction grating. The input shows the observed image under the proposed illumination system (displayed as an RGB image) along with the separated direct and indirect reflectance components for each wavelength (\"Direct/Indirect\"). Rough textures and specular reflections are observed in the direct component, while soft textures, such as internal scattering, are observed in the indirect component. The leftmost plots show the comparison of our estimation with the reference at points \\( p_1 \\) and \\( p_2 \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.936, + 0.521, + 0.948 + ], + "angle": 0, + "content": "26827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.409 + ], + "angle": 0, + "content": "modeling, realistic rendering, and advanced editing. It also provides a non-destructive, non-invasive way to measure surface and internal spectral properties. This enables a wide range of applications, including medical analysis, such as diagnosing skin abnormalities [40], food inspection [20], and painting analysis for revealing the layered structure of pigments [18]. These analyses can be further enhanced by separating the direct and indirect components, where faster examination is preferred. It also reduces burdens on human subjects and minimizes light exposure for sensitive targets like art and biomedical samples. However, achieving this remains challenging, as a hyperspectral camera requires longer acquisition times than a conventional camera, making it unsuitable for computational photography approaches that rely on multiple pattern irradiations and acquisitions. In recent years, there have been attempts to separate reflection components from a single captured image by methods based on basis representation [30] and deep learning [9], but their method requires a dataset of a large number of images and which is hard to apply for hyperspectral scenarios due to the measurement time of hyperspectral images." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.76 + ], + "angle": 0, + "content": "To address the problem of time-consuming hyperspectral imaging, CASSI (Coded Aperture Snapshot Spectral Imaging System) [1] which enables hyperspectral image snapshots [3, 36], techniques for estimating information on other modalities from spectral images [16, 22, 25], and methods for achieving both [2, 29] have been proposed. Snapshot hyperspectral imaging methods are effective for hyperspectral image reconstruction; however, their implementation often requires complex imaging systems, and achieving measurements with sufficient accuracy for further scene analysis necessitates considerable effort. There is an approach that, given a single hyperspectral image, separates the first and second bounce components based on the shape of the observed spectral distribution at each point [25]. However, this method is limited to scenes composed of only a few colors, imposing constraints on its applicability. A more general approach capable of handling complex scenes has yet to be realized. In summary, hyperspectral reconstruction and scene analysis using hyperspectral images have been studied in the past, but a robust solution for achieving spectral direct/indirect separation of complex scenes composed of multiple materials using a single hyperspectral image has yet to be realized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To tackle this challenging problem, we develop a novel illumination system that integrates dispersive optics to generate a wavelength-dependent, spatially shifted illumination pattern. Similar to color photometric stereo, where varying lighting directions are achieved through different color channels, our system leverages observed wavelength differences to achieve a shifted illumination pattern. We demonstrate that the challenges associated with hyperspectral-based direct/indirect separation are resolved by our system." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "By accounting for spectral reflectance differences appropriately, we demonstrate that a single hyperspectral image captured under one illumination pattern can achieve robust separation of direct and indirect components for each wavelength. Furthermore, we show that analyzing variations across wavelengths aids in estimating scene depth." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.71, + 0.196 + ], + "angle": 0, + "content": "Our contribution is threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- A novel illumination system incorporating dispersive optics is proposed, enabling the illumination of a wavelength-dependent, spatially shifted pattern." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.244, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- By carefully addressing the differences in spectral reflectance, we formulate the observed pixel values based on these spectral-dependent illumination patterns and derive a solution for direct/indirect separation at each wavelength using a single hyperspectral image with one illumination pattern." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We develop a real system and demonstrate robust spectral direct/indirect separation across various objects. Additionally, we show that analyzing differences across wavelengths aids in estimating scene depth." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.663, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Separation of Direct and Indirect Reflection Components via Pattern Illumination" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Nayar et al. [24] proposed separating the direct and indirect reflection components of a scene using high-frequency illumination patterns, and various extended methods have also been proposed. As the applications for acquiring rich information, O'Toole et al. [26, 27] proposed an image measurement method that can observe light transport paths in a scene by modulating both the illumination pattern of the projector and the camera response. As the high-speed application, separation of direct and indirect reflections by time dithering of a DLP projector with a high-speed camera was proposed [23], and extended to multiple colors [34]. More recently, a fast-capturing and processing method based on machine learning using event data was proposed [38] by a hybrid camera system combining an event camera and an RGB camera. Thus, many methods have been proposed to speed up the separation method of direct and indirect reflection components using pattern illumination, and extensions to multispectral with a small number of channels have been made, but no extension of patterns to hyperspectral image acquisition or spectral dimension has been achieved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.902 + ], + "angle": 0, + "content": "On the other hand, several methods have been proposed to separate the direct and indirect reflection components of a scene using a single image. Subpa-Asa et al. [30] propose to evaluate the data-independent Fourier basis and learning-based principal component analysis (PCA) basis to find a better linear basis representation of the direct and indirect components. Duan et al. [9] proposed a deep learning solution for the separation of reflective components from a single image by designing a non-binary pattern with 4 or" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.518, + 0.948 + ], + "angle": 0, + "content": "26828" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "9 intensities instead of a binary grid pattern. Although the method of separating reflection components from a single image has theoretical applicability to hyperspectral images, in practice it is difficult to apply because it requires a data set with many images." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.171, + 0.472, + 0.185 + ], + "angle": 0, + "content": "Decomposition-based Spectral Reflectance Separation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.473 + ], + "angle": 0, + "content": "Separation of the direct and indirect spectral reflectance can also be treated as a spectral decomposition problem. Several studies based on spectral decomposition problems using hyperspectral images have been proposed. In the field of remote sensing, linear or nonlinear spectral decomposition methods have been proposed to address the problems of spectral variation due to the atmosphere and/or spectral mixing due to the imaging system [4]. In the field of computer vision, Zheng et al. [37] has enabled spectral separation between illumination and reflectance in hyperspectral images taken under common spectral illumination and extended to multiple light sources by Chen et al. [5]. Regarding the separation of specular reflection components on the target surface using spectral information, an analysis of chromaticity-intensity space in diffuse and specular reflections [33], a non-iterative method based on a dichroic reflection model and PCA [11], and a method based on blind source separation using non-negative matrix factorization [39] have been proposed." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.747 + ], + "angle": 0, + "content": "As the application of separating indirect reflection, Okawa et al. [25] proposed a method to separate interreflection components from a single hyperspectral image when the object color is one or two colors, based on the fact that not only surfaces of different colors but also surfaces of the same color change the observed spectrum due to interreflection. Moreover, Fu et al. [12] proposed a method to separate the reflection and fluorescence emission spectra in a scene by using high-frequency illumination in the spectral domain. Their method achieves high-frequency illumination in the spectral domain using an equalized light source, whose wavelength characteristics can be freely controlled, but it is uniform in the spatial domain and cannot be used to separate direct and indirect reflection components. Thus, while several possible ways to separate reflection components using hyperspectral images exist, only interreflection or fluorescence has been considered for the separation of indirect components." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.218, + 0.765 + ], + "angle": 0, + "content": "Dispersive Optics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Dispersive optics, such as prisms and diffraction gratings, have often been used in camera and display design, especially for hyperspectral imaging. Mohan et al. [21] proposed a spectral mask that can control the spectral power distribution of a camera or projector using a diffraction grating and an attenuation mask. Hostettler et al. [15] used multiple prisms and calculated masks to realize a projector system that projects the desired color image without a conventional projector. CASSI uses dispersion optics and coding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "masks to acquire masked spectral images with wavelength-dependent transformations and reconstructs the hyperspectral data cube using an iterative algorithm [1, 36]. Cao et al. [3] achieved fast hyperspectral image acquisition with a hybrid camera system constructed from an RGB camera, a grayscale camera, and multiple optical elements including prisms and coding masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Several methods have been proposed to acquire spectral and 3D depth information using dispersive optics. Baek et al. [2] proposed the diffractive optical element (DOE) generates a unique point spread function (PSF) that changes with the scene's spectrum and depth, enabling single-shot hyperspectral 3D imaging through blur analysis. Shin et al. [29] proposed a system to obtain spectral and 3D depth information by analyzing which structured light patterns disperse and illuminate a scene, using a compact setting that adds diffraction grating film to a projector-camera system." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Dispersive optics have been utilized for projector/camera enhancement, hyperspectral image acquisition, and depth estimation. This field holds great potential for various applications. In this study, we demonstrate the effectiveness of direct/indirect separation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.605, + 0.451 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.61 + ], + "angle": 0, + "content": "We separate the direct and indirect reflection components of a scene from a single-shot hyperspectral image taken under a wavelength-dependent, spatially shifted illumination pattern generated by a DMD (Digital Micromirror Device) projector with a diffraction grating (Figure 1). We first revisit the traditional approach to direct and indirect reflections, utilizing varying high-frequency projected patterns, and then extend the discussion to our case, which involves a wavelength-dependent, spatially shifted illumination pattern." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.905, + 0.648 + ], + "angle": 0, + "content": "3.1. Direct/Indirect Separation Using Sinusoidal Pattern Illumination" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.777 + ], + "angle": 0, + "content": "When a single light source illuminates a scene, the direct component \\( I_{d}(x) \\) is defined as the light that travels directly from the source, reflects once at a surface point, and reaches the camera, while the indirect component \\( I_{i}(x) \\) encompasses all light that undergoes multiple interactions, such as subsurface scattering, interreflections, and volumetric scattering. Nayar et al. [24]. The observed intensity \\( I(x) \\) by a camera is represented as" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.791, + 0.905, + 0.807 + ], + "angle": 0, + "content": "\\[\nI (x) = I _ {d} (x) + I _ {i} (x). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Assuming that the indirect reflection component of a scene changes gradually in space and is predominantly low-frequency, their proposed approach [24] demonstrated that the scene's direct and indirect reflection components can be separated by irradiating three spatially high-frequency sinusoidal patterns." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.519, + 0.948 + ], + "angle": 0, + "content": "26829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.094, + 0.91, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.238, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Figure 2. The outline of the proposed spatio-spectral illumination system. (a) The (DMD) projector generates a sinusoidal pattern that has a broad spectral band of light. (b) The diffraction grating disperses the generated light pattern, then the sinusoidal pattern is spatially shifted according to wavelength. (c) The spatio-spectral shifted patterns are observed at each point of the scene, and the sinusoidal pattern is resampled in the spectral domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Under the high-frequency sinusoidal pattern illuminations, the observed images \\( I_{n}(x,k) \\) are represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.375, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\nI _ {n} (x, k) = \\frac {1}{2} I _ {d} (x) (\\sin (\\frac {2}{3} k \\pi + \\phi) + 1) + \\frac {1}{2} I _ {i} (x), \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.537 + ], + "angle": 0, + "content": "where \\(\\phi \\in [0,2\\pi]\\) denotes an arbitrary phase of the sinusoidal function and \\(k = \\{1,2,3\\}\\) represents the index of the pattern. This problem has three unknowns \\(\\{I_d,I_i,\\phi \\}\\) and can be solved theoretically for each pixel using three images taken under three different sinusoidal pattern illumination. In practice, this approach requires pixel-by-pixel camera-projector alignment or additional patterns to avoid artifacts and ensure accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.399, + 0.563 + ], + "angle": 0, + "content": "3.2. Spectral Direct/Indirect Separation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Let us further extend the discussion to separation at the wavelength level by considering the spectral reflection of a scene. Here, we represent the direct reflection component \\( I_{d}(x,\\lambda) \\) as the product of normalized spectral reflectance of direct component \\( R_{d}(x,\\lambda) \\), its magnitude \\( \\alpha_{d}(x) \\), and the intensity of the spectral light source \\( L_{d}(x,\\lambda) \\) at the point \\( x \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Following [24], the indirect reflection component is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.698, + 0.483, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I _ {i} (x, \\lambda) = \\sum_ {x ^ {\\prime} \\in S (x)} A ([ x, x ^ {\\prime} ], \\lambda) L ([ x, x ^ {\\prime} ], \\lambda) \\tag {3} \\\\ = \\alpha_ {i} (x) R _ {i} (x, \\lambda), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.866 + ], + "angle": 0, + "content": "where \\( S(x) \\) denotes the set of surrounding pixels for \\( x \\), \\( L([x, x'], \\lambda) \\) represents the spectral radiance at \\( x' \\) in the direction of \\( x \\), and \\( A([x, x'], \\lambda) \\) incorporates the spectral BRDF of \\( x \\) as well as the relative geometric configuration between \\( x \\) and \\( x' \\).1 \\( I_i(x, \\lambda) \\) is further expressed in terms of the normalized spectrum \\( R_i(x, \\lambda) \\) which is computed from the total sum of light contributions coming to \\( x \\) from surrounding pixels \\( S(x) \\) and its magnitude \\( \\alpha_i(x) \\). From this," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our system, the scene brightness has been calibrated by considering the camera's spectral sensitivity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.902, + 0.35 + ], + "angle": 0, + "content": "the spectral reflection model of a scene \\(I(x,\\lambda)\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.361, + 0.918, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I (x, \\lambda) = I _ {d} (x, \\lambda) + I _ {i} (x, \\lambda) \\\\ = \\alpha_ {d} (x) R _ {d} (x, \\lambda) L _ {d} (x, \\lambda) + \\alpha_ {i} (x) R _ {i} (x, \\lambda). \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.906, + 0.454 + ], + "angle": 0, + "content": "In the following, we introduce a spatial phase shift by designing the lighting pattern to shift horizontally with wavelength." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.906, + 0.495 + ], + "angle": 0, + "content": "3.3. Spatially Shifting Illumination According to Wavelength" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.623 + ], + "angle": 0, + "content": "We propose a spatio-spectral pattern illumination system that combines a conventional projector and a diffraction grating to create spatially shifting illumination patterns according to wavelength. The configuration of the illumination system is shown in Figure 2. Let the spatially high-frequency sinusoidal pattern \\( P_0(v) \\) generated by the conventional projector (Figure 2 (a)) be written as a sinusoidal function of the horizontal pixel position \\( v \\) of the projector:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.634, + 0.906, + 0.665 + ], + "angle": 0, + "content": "\\[\nP _ {0} (v) = \\frac {l _ {0}}{2} \\left(\\sin \\left(f _ {0} v + \\phi_ {0}\\right) + 1\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(l_{0}\\) is the intensity of the projector, \\(f_{0}\\) describes the frequency of the sinusoidal function, and \\(\\phi_0\\) denotes the phase of the sinusoidal function. Then the grating diffracts the incoming pattern light and reflects it at slightly different angles for each wavelength of light (Figure 2 (b)). In the case of first-order diffracted light, the diffraction angle of the outgoing light, i.e. the emission angle, varies monotonically in the visible light range depending on the wavelength of the incident light. Thus, the illumination pattern varies in the spatio-spectral domain, similar to a rainbow. Therefore, as shown in Figure 3, when the diffracted pattern is observed at a scene point, it is also a sinusoidal pattern in the spectral domain (Figure 2 (c)). We proceed with the discussion assuming that projector's origin is the virtual origin of the pattern reflected by the grating. We also set the cam" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.936, + 0.519, + 0.948 + ], + "angle": 0, + "content": "26830" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "era's origin at the same location to enable discussion within a unified coordinate system." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.135, + 0.477, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.236, + 0.485, + 0.293 + ], + "angle": 0, + "content": "Figure 3. After passing through the diffraction grating, the sinusoidal pattern generated by the conventional projector (following yellow arrow) is resampled in the spectral domain at each point of the scene (following red arrow)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.366 + ], + "angle": 0, + "content": "The spectral pattern \\( P(x, \\lambda) \\) is expressed as the function of wavelength \\( \\lambda \\) at the scene point corresponding to the projector's horizontal pixel position \\( x \\) as in Equation (6):" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.369, + 0.483, + 0.4 + ], + "angle": 0, + "content": "\\[\nP (x, \\lambda) = \\frac {l (\\lambda)}{2} \\left(\\sin \\left(f _ {p} (x) \\lambda + \\phi (x)\\right) + 1\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Let \\( l(\\lambda) \\) be the spectral intensity of the projector, \\( f_{p}(x) \\) denote the frequency of the sinusoidal function, and \\( \\phi (x) \\) represent the phase of the sinusoidal function. We calibrate the spectral intensity of the projector \\( l(\\lambda) \\) in advance and describe it as the constant value \\( l \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.224, + 0.502 + ], + "angle": 0, + "content": "3.4. Simple Case" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.69 + ], + "angle": 0, + "content": "We begin our discussion with the simplest case: illuminating our pattern onto a gray object with flat spectral reflectance and spatially varying albedo. In this case, the normalized spectra \\( R_{d} \\) and \\( R_{i} \\) of the direct and indirect components can be integrated and represented as the respective magnitudes \\( \\alpha_{d} \\) and \\( \\alpha_{i} \\). When illuminating our pattern onto this object, the direct light component \\( L_{d}(x,\\lambda) \\) in Equation (4) is replaced by \\( P(x,\\lambda) \\), and the indirect reflection component \\( I_{i}(x,\\lambda) \\) is reduced to half intensity and pattern independent following [24]. Hence, the observed reflection \\( I_{p}(x,\\lambda) \\) under the spatio-spectral pattern illumination \\( P(x,\\lambda) \\) is modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.694, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\nI _ {p} (x, \\lambda) = \\alpha_ {d} (x) P (x, \\lambda) + \\frac {1}{2} \\alpha_ {i} (x). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.727, + 0.449, + 0.743 + ], + "angle": 0, + "content": "Expanding Equation (7) based on Equation (6) yields:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.747, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\[\nI _ {p} (x, \\lambda) = \\frac {l \\alpha_ {d} (x)}{2} (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) + \\frac {\\alpha_ {i} (x)}{2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In this case, we can obtain a sinusoidal pattern in the spectral domain along wavelength axis at any point \\( x \\) as shown in Figure 4 (a)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since the number of unknowns is four \\(\\{\\alpha_{d}(x),f_{p}(x),\\phi (x),\\alpha_{i}(x)\\}\\) , the direct and indirect reflection components of a scene can be separated by acquiring spectral images at least 4 spectral channels in the simple case." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.625, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.091, + 0.764, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.091, + 0.899, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Figure 4. Our illumination pattern and the observation as it is projected onto the scene. (a) For gray objects, a sinusoidal pattern is observed in the spectral domain. (b) For colored objects, a pattern that varies with the spectral reflectance of the direct and indirect components of the object is observed in the spectral domain." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.662, + 0.319 + ], + "angle": 0, + "content": "3.5. Complex Case" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.906, + 0.415 + ], + "angle": 0, + "content": "Next, let us examine the case of the object that has non-flat, spatially varying spectral reflectance. As shown in Figure 4 (b), the spectrum observed for a scene with non-flat spectral reflectance is affected by the spectral reflectance of the scene, unlike the simple flat reflectance case in Figure 4 (a)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Based on Equation (4), a scene can be composed of direct and indirect reflection components, each of which can have its spectral distribution. The observed intensity \\( I_{p}(x,\\lambda) \\) at the scene point \\( x \\) under the spatio-spectral pattern illumination \\( P(x,\\lambda) \\) can be expressed:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.504, + 0.907, + 0.535 + ], + "angle": 0, + "content": "\\[\nI _ {p} (x, \\lambda) = \\alpha_ {d} (x) R _ {d} (x, \\lambda) P (x, \\lambda) + \\frac {\\alpha_ {i} (x)}{2} R _ {i} (x, \\lambda) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.871, + 0.56 + ], + "angle": 0, + "content": "Expanding Equation (9) based on Equation (6) yields:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.571, + 0.918, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I _ {p} (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) / 2 \\\\ + \\quad \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Furthermore, Equation (10) can be expressed in the form of a general sinusoidal function for \\(\\lambda\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.658, + 0.907, + 0.688 + ], + "angle": 0, + "content": "\\[\nI_{p}(x,\\lambda) = A(x,\\lambda)\\{\\sin (B(x)\\lambda +C(x)) + 1\\} +D(x,\\lambda),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.69, + 0.866, + 0.706 + ], + "angle": 0, + "content": "\\[\nA (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) / 2, B (x) = f _ {p} (x),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.706, + 0.859, + 0.721 + ], + "angle": 0, + "content": "\\[\nC (x) = \\phi (x), D (x, \\lambda) = \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.875, + 0.721, + 0.905, + 0.733 + ], + "angle": 0, + "content": "(11)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Equation (11) shows that the amplitude of the sinusoidal function is related to the direct reflection component and the constant term is related to the indirect reflection component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We then discuss how to separate the direct and indirect reflection components from the observed values based on Equation (11). The number of observations at each pixel obtained in a single hyperspectral image measurement equals the number of spectral wavelength observations. Therefore, if either spectral reflectance \\( R_{d} \\) and \\( R_{i} \\) are unknown, solving Equation (11) from a single multi- or hyperspectral input image is an ill-posed problem. We address this problem" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.936, + 0.518, + 0.948 + ], + "angle": 0, + "content": "26831" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "by representing \\(R_{d}\\) and \\(R_{i}\\) as the combination of the spectral reflectance basis, following the work of Han et al. [14]:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.13, + 0.483, + 0.149 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {R}} _ {\\boldsymbol {d}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {d}} (\\boldsymbol {x}), \\bar {\\boldsymbol {R}} _ {\\boldsymbol {i}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {i}} (\\boldsymbol {x}), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.484, + 0.399 + ], + "angle": 0, + "content": "where \\(\\bar{R}_d(x)\\) and \\(\\bar{R}_i(x) \\in \\mathbb{R}^f\\) are the vectorized form of the normalized spectral reflectances at the pixel \\(x\\) approximated by the basis representation, \\(B \\in \\mathbb{R}^{f \\times k}\\) is a basis matrix consisting of \\(k\\) basis vectors, and \\(c_d(x)\\) and \\(c_i(x) \\in \\mathbb{R}^k\\) are the basis coefficients at the pixel \\(x\\). \\(\\bar{R}_i(x)\\) is not pure spectral reflectance, but it is the sum of the spectral BRDFs in the surrounding pixels and can therefore be expressed on the same basis. Based on the public dataset [28], in which the 1257 spectral reflectances of Munsell Book of Color: Matte Finish Collection [6] are recorded, \\(k\\) spectral reflectance basis vectors are extracted using singular value decomposition. Approximation by the basis reduces the number of unknowns related to the spectral reflectance to \\(k\\). Thus, unknowns of our problem are \\(\\{c_d(x), c_i(x), \\alpha_d(x), f_p(x), \\phi(x), \\alpha_i(x)\\}\\) and its total degree of freedom is \\(2k + 4\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.231, + 0.423 + ], + "angle": 0, + "content": "3.6. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.52 + ], + "angle": 0, + "content": "We solve the problem described in Section 3.5 by optimization, using a single hyperspectral image captured by our spatio-spectral pattern illumination system as input. For simplicity, let us focus on a single pixel, with all terms written while omitting \\( x \\). The error function between the input spectral image and the model is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.529, + 0.483, + 0.569 + ], + "angle": 0, + "content": "\\[\nE _ {s} (\\lambda) = \\left| I _ {p} (\\lambda) - \\frac {l \\alpha_ {d}}{2} \\bar {R} _ {d} (\\lambda) \\left\\{\\sin \\left(f _ {p} \\lambda + \\phi\\right) + 1 \\right\\} - \\frac {\\alpha_ {i}}{2} \\bar {R} _ {i} (\\lambda) \\right|. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Where \\(\\bar{R}_d(\\lambda)\\) and \\(\\bar{R}_i(\\lambda)\\) denote the scalar form of the normalized spectral reflectances approximated by the basis representation. The objective function of the problem is:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.625, + 0.483, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, \\alpha_ {d}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.87 + ], + "angle": 0, + "content": "As mentioned earlier, our problem setup has \\(2k + 4\\) unknowns, so the simplest solution is to give a hyperspectral image with \\(2k + 4\\) spectral channels as the input image. From previous study [14], the number of bases for approximating the spectral reflectance of a general object with a basis representation is about 8. Therefore, when we follow them and let \\(k = 8\\), we require a hyperspectral image with 20 spectral channels to solve. When applying hyperspectral imaging to digital archiving and other applications, it is common to work in visible light, with wavelengths ranging from \\(400\\mathrm{nm}\\) to \\(700\\mathrm{nm}\\). Therefore, acquiring multi- or hyperspectral images with a wavelength interval of at least 15 nm in the visible light range is sufficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For robust optimization, we use initial values and ranges of estimation values for some of the parameters. As shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.23 + ], + "angle": 0, + "content": "in the figure, the period of the sinusoidal function of the spectral domain at each pixel can be predicted from the local maxima and minima of the observed signal. The period obtained from the median of the intervals at the local maxima and minima is the initial value of the period \\( f_{p} \\) at that pixel, and the search range for optimization is \\( \\pm 5\\% \\) of the initial value to account for discretization error. Also, the search range of the phase \\( \\phi \\) is \\( [-\\pi, \\pi] \\) and the magnitude of the direct and indirect components \\( \\alpha_{d} \\) and \\( \\alpha_{i} \\) are \\( [0, \\infty] \\)." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.245, + 0.901, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.332, + 0.905, + 0.36 + ], + "angle": 0, + "content": "Figure 5. Examples of the observed signal in the spectral domain. Orange and green circles represent local maxima and minima." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Furthermore, since spectral reflectance is known to be smooth for most common objects [13], we introduce a loss function that reflects the assumption of smoothness of spectral reflectance, following prior works [19, 32]:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.465, + 0.906, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda) + \\gamma_ {d} | \\boldsymbol {B c} _ {d} \\boldsymbol {\\Gamma} | + \\gamma_ {i} | \\boldsymbol {B c} _ {i} \\boldsymbol {\\Gamma} |. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.518, + 0.906, + 0.564 + ], + "angle": 0, + "content": "The weights \\(\\gamma_{d}\\) and \\(\\gamma_{i}\\) are chosen empirically and \\(\\Gamma\\) denotes a matrix that gives smoothness by the subtraction between the neighbor elements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.58, + 0.72, + 0.597 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.825, + 0.621 + ], + "angle": 0, + "content": "4.1. Experimental Setup and Conditions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Our system consists of a xenon light source (Asahi Spectra LAX-100), a DMD (Ajile AJD-4500), a hyperspectral camera (EBA JAPAN NH-8), lenses for the DMD and hyperspectral camera (Fujinon HF25SA-1), and a diffraction grating (Edmund Optics 1200 Grooves, \\(50\\mathrm{mm}\\) Square, \\(500\\mathrm{nm}\\) Ruled Diffraction Grating), as shown in Figure 1. The hyperspectral camera captures spectral images at \\(5\\mathrm{nm}\\) intervals within the \\(380 - 1000\\mathrm{nm}\\) wavelength range. Considering the effective wavelength range of the diffraction grating, we performed separation in the \\(450 - 650\\mathrm{nm}\\) range." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The minimum number of observations required to solve our problem is 20, as discussed in Section 3.6. However, actual spectral images contain significant noise, and factors such as projector and camera calibration errors further impact accuracy, preventing us from achieving stable results with the minimum number of observations. In the following experiments, we use 41 spectral bands obtained at \\(5\\mathrm{nm}\\) intervals to ensure robust separation against noise." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.519, + 0.948 + ], + "angle": 0, + "content": "26832" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.303, + 0.106 + ], + "angle": 0, + "content": "4.2. Results on Real Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.34 + ], + "angle": 0, + "content": "Figure 8 shows the results of estimating the spectral reflectance of the direct and indirect reflectance components in various real-world objects. For reference, we captured a set of images of the scene using a hyperspectral camera with different sinusoidal pattern illuminations and provided the separation results using Nayar's method [24], as described in Equation (2). Hyperspectral images contain significant noise, making it challenging to obtain stable separation results free from artifacts by Nayar's method [24] unless numerous hyperspectral images are captured under various sinusoidal pattern illuminations. To mitigate such artifacts, the spatial resolution of the input images for the reference results, but not ours, is downsampled for visual comparison. All input hyperspectral images and estimated results are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.599 + ], + "angle": 0, + "content": "For all four scenes in Figure 8, the estimated direct \\((I_d)\\) and indirect \\((I_i)\\) components closely resemble the corresponding reference across different wavelengths. From these examples, we observe that wavelength-dependent changes in the direct and indirect components are extracted. In particular, in the results for the Potato and Ball (top 2 rows) and the Candle (bottom 2 rows), the scattering inside the objects becomes broader and stronger for longer wavelengths. Fine textures can be seen in the direct reflection component and rarely in the indirect reflection component. The leftmost graph compares our estimation and reference for points \\(p_1\\) and \\(p_2\\). The estimated spectra well represent that of the reference; nevertheless, they are derived solely from a single hyperspectral image. Moreover, Figure 6 shows spatial maps and values of error metrics. We used MSE (mean squared error) and SAM (spectral angle mapper) as error metrics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.609, + 0.45, + 0.626 + ], + "angle": 0, + "content": "4.3. Depth Dependence of the Pattern's Period" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Figure 7 shows the spatial distribution of the estimated period \\(\\phi\\) in Equation (15). The period \\(\\phi\\) increases as the distance from the projector becomes longer, making it highly correlated with object distance. In Figure 7, 'DAY' is placed closer to the camera, and 'HAP' is farther away. This example demonstrates that the estimated \\(\\phi\\) contains valuable information about the scene's depth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.21, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we proposed a novel method for separating the direct and indirect reflection components of a scene from a single hyperspectral image. By introducing a projection system incorporating dispersive optics, we enabled the projection of a wavelength-dependent, spatially shifted pattern. We formulated the observed pixel values under these spectral-dependent illumination patterns and derived a solution for direct and indirect reflection separation at each" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.905, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.413, + 0.908, + 0.469 + ], + "angle": 0, + "content": "Figure 6. Error maps and metrics between the reference and the proposed method. The values represent the mean and variance of the MSE and SAM. Note that some stripe patterns in the SAM maps are affected by the artifacts of the reference." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.501, + 0.905, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.636, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Figure 7. Spatial map of the estimated period \\(\\phi\\) for a scene. 'DAY' is placed closer to the camera, and 'HAP' is farther away." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.812 + ], + "angle": 0, + "content": "wavelength using only a single hyperspectral image with one projection pattern. Through experiments, we demonstrated the effectiveness of our system in achieving robust spectral direct/indirect separation even for objects with complex spectral reflectance. Furthermore, our analysis of differences across wavelengths was leveraged for scene depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.829, + 0.673, + 0.846 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This research is supported by JSPS Kakenhi Grant number 25K21236 and JST-Mirai Program Grant Number JP-MJM123G1." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.518, + 0.948 + ], + "angle": 0, + "content": "26833" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.088, + 0.852, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.812, + 0.907, + 0.911 + ], + "angle": 0, + "content": "Figure 8. Separation results of direct and indirect reflection components in various real-world objects. Top row, from left to right: scene captured by the hyperspectral camera (RGB converted), results of separation of direct and indirect reflection components \\((I_d(480), I_d(620), I_i(480), I_i(620))\\) by Nayar's method [24] as a reference, direct reflectance component (red) at point \\(p_1\\) and indirect, comparison of spectral reflectance of direct (red) and indirect (green) components at point \\(p_1\\) for the reference (dotted line) and the proposed method (solid line). Bottom row, from left to right: part of the input spectral images in the proposed method are superimposed on a pseudo-RGB, estimation results of direct and indirect reflection components \\((I_d(480), I_d(620), I_i(480), I_i(620))\\) by the proposed method, comparison of spectral reflectance at \\(p_2\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.518, + 0.948 + ], + "angle": 0, + "content": "26834" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Gonzalo R Arce, David J Brady, Lawrence Carin, Henry Arguello, and David S Kittle. Compressive coded aperture spectral imaging: An introduction. IEEE Signal Processing Magazine, 31(1):105-115, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Seung-Hwan Baek, Hayato Ikoma, Daniel S Jeon, Yuqi Li, Wolfgang Heidrich, Gordon Wetzstein, and Min H Kim. Single-shot hyperspectral-depth imaging with learned diffractive optics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2651-2660, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Xun Cao, Xin Tong, Qionghai Dai, and Stephen Lin. High resolution multispectral video capture with a hybrid camera system. In CVPR 2011, pages 297-304. IEEE, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Xavier Ceamanos and Silvia Valero. 4 - processing hyperspectral images. In Optical Remote Sensing of Land Surface, pages 163-200. Elsevier, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[5] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[6] Munsell Color. Munsell book of color: Matte finish collection: Kollmorgen corporation. Baltimore, Maryland, 1976. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[7] Costanza Cucci, John K Delaney, and Marcello Picollo. Reflectance hyperspectral imaging for investigation of works of art: old master paintings and illuminated manuscripts. Accounts of chemical research, 49(10):2070-2079, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.541 + ], + "angle": 0, + "content": "[8] Craig Donner and Henrik Wann Jensen. A spectral bssrfd for shading human skin. Rendering techniques, 2006:409-418, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[9] Zhaoliang Duan, James Bieron, and Pieter Peers. Deep Separation of Direct and Global Components from a Single Photograph under Structured Lighting. Computer Graphics Forum, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[10] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Transactions on graphics (TOG), 37(6):1-14, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[11] Gamal ElMasry, Pere Gou, and Salim Al-Rejaie. Effectiveness of specularity removal from hyperspectral images on the quality of spectral signatures of food products. Journal of Food Engineering, 289:110148, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[12] Ying Fu, Antony Lam, Imari Sato, Takahiro Okabe, and Yoichi Sato. Separating reflective and fluorescent components using high frequency illumination in the spectral domain. In Proceedings of the IEEE International Conference on Computer Vision, pages 457-464, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[13] Lewis D Griffin. Reconciling the statistics of spectral reflectance and colour. PLoS One, 14(11):e0223069, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[14] Shuai Han, Imari Sato, Takahiro Okabe, and Yoichi Sato. Fast spectral reflectance recovery using dlp projector. International journal of computer vision, 110:172-184, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Rafael Hostettler, Ralf Habel, Markus Gross, and Wojciech Jarosz. Dispersion-based color projection using masked prisms. In Computer Graphics Forum, pages 329-338. Wiley Online Library, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[16] Shin Ishihara, Yuta Asano, Yinqiang Zheng, and Imari Sato. Underwater scene recovery using wavelength-dependent refraction of light. In 2020 International Conference on 3D Vision (3DV), pages 32-40. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[17] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31 (4):1-11, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[18] Paul Kubelka and Franz Munk. An article on optics of paint layers. Z. Tech. Phys, 12(593-601):259-274, 1931. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.906, + 0.301 + ], + "angle": 0, + "content": "[19] Wei Liu, Chengxun He, and Le Sun. Spectral-smoothness and non-local self-similarity regularized subspace low-rank learning method for hyperspectral mixed denoising. Remote Sensing, 13(16):3196, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[20] Lembe Samukelo Magwaza and Umezuruike Linus Opara. Analytical methods for determination of sugars and sweetness of horticultural products—a review. Scientia Horticulturae, 184:179-192, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[21] Ankit Mohan, Ramesh Raskar, and Jack Tumblin. Agile spectrum imaging: Programmable wavelength modulation for cameras and projectors. In Computer graphics forum, pages 709-717. Wiley Online Library, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[22] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE computer graphics and applications, 34(6):57-68, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[23] Srinivasa G Narasimhan, Sanjeev J Koppal, and Shuntaro Yamazaki. Temporal dithering of illumination for fast active vision. In Computer Vision-ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10, pages 830-844. Springer, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[24] Shree K Nayar, Gurunandan Krishnan, Michael D Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. In ACM SIGGRAPH 2006 Papers, pages 935-944. 2006. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[25] Hiroki Okawa, Yinqiang Zheng, Antony Lam, and Imari Sato. Spectral reflectance recovery with interreflection using a hyperspectral image. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13, pages 52-67. Springer, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[26] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[27] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.908, + 0.831 + ], + "angle": 0, + "content": "[28] Jussi PS Parkkinen, Jarmo Hallikainen, and Timo Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical society of America A, 6(2):318-322, 1989. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[29] Suhyun Shin, Seokjun Choi, Felix Heide, and Seung-Hwan Baek. Dispersed structured light for hyperspectral 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24997-25006, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.936, + 0.519, + 0.948 + ], + "angle": 0, + "content": "26835" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Art Subpa-asa, Ying Fu, Yinqiang Zheng, Toshiyuki Amano, and Imari Sato. Separating the direct and global components of a single image. Journal of Information Processing, 26: 755-767, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[31] Da-Wen Sun. Hyperspectral imaging for food quality analysis and control. Elsevier, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[32] Le Sun, Qihao Cheng, and Zhiguo Chen. Hyperspectral image super-resolution method based on spectral smoothing prior and tensor tubal row-sparse representation. Remote Sensing, 14(9):2142, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[33] Robby T Tan, Ko Nishino, and Katsushi Ikeuchi. Separating reflection components based on chromaticity and noise analysis. IEEE transactions on pattern analysis and machine intelligence, 26(10):1373-1379, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[34] Moriaki Torii, Takahiro Okabe, and Toshiyuki Amano. Multispectral direct-global separation of dynamic scenes. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1923-1931. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[35] Mahesh Kumar Tripathi and H Govil. Evaluation of aviris-ng hyperspectral images for mineral identification and mapping. Heliyon, 5(11), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[36] Ashwin Wagadarikar, Renu John, Rebecca Willett, and David Brady. Single disperser design for coded aperture snapshot spectral imaging. Applied optics, 47(10):B44-B51, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[37] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[38] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. Evdig: Event-guided direct and global components separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9612-9621, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.656 + ], + "angle": 0, + "content": "[39] Mustapha Zokay and Hicham Saylani. Removing specular reflection in multispectral dermatological images using blind source separation. In Annual Conference on Medical Image Understanding and Analysis, pages 734-750. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[40] George Zonios, Julie Bykowski, and Nikiforos Kollias. Skin melanin, hemoglobin, and light scattering properties can be quantitatively assessed in vivo using diffuse reflectance spectroscopy. Journal of Investigative Dermatology, 117(6): 1452-1457, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.937, + 0.519, + 0.948 + ], + "angle": 0, + "content": "26836" + } + ] +] \ No newline at end of file diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_origin.pdf b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc4bd0a414f9ec00dee52ee692e4f57911b26eab --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/c416257e-42cc-4603-837a-a5cb6ea0c724_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edb4ac090eaf2b4988aab088d02f29a1ac54772a33c7e222fc20f890aea3cc17 +size 15473322 diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/full.md b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..170e4a0a5ce94d6f317fd62980654dc11f1290ba --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/full.md @@ -0,0 +1,312 @@ +# Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image + +Shin Ishihara Imari Sato + +National Institute of Informatics + +2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan + +{sishihara,imarik}@nii.ac.jp + +# Abstract + +Hyperspectral imaging has proven effective for appearance inspection because it can identify material compositions and reveal hidden features. Similarly, direct/indirect separation provides essential information about surface appearance and internal conditions, including layer structures and scattering behaviors. This paper presents a novel illumination system incorporating dispersive optics to unify both advantages for scene analysis. In general, achieving distinct direct/indirect separation requires multiple images with varying patterns. In a hyperspectral scenario, using a hyperspectral camera or tunable filters extends exposure and measurement times, hindering practical application. Our proposed system enables the illumination of a wavelength-dependent, spatially shifted pattern. With proper consideration of reflectance differences, we demonstrate that robust separation of direct and indirect components for each wavelength can be achieved using a single hyperspectral image acquired under our single spatiospectral pattern illumination. + +# 1. Introduction + +Hyperspectral imaging enables pixel-level analysis of objects and is widely used in fields such as medical imaging, food quality control, environmental monitoring, and art conservation [7, 17, 31, 35]. Spectral rendering is gaining traction, with modern renderers adopting wavelength-dependent models [10], and reflection/scattering analysis aiding in material editing [8]. Analyzing layer structures is also essential for inspecting an object's surface and internal conditions. Techniques for separating a scene's direct and indirect reflection components are often employed to accurately capture surface reflection colors and shapes, as well as to visualize an object's internal structure. Nayar et al. developed a groundbreaking fast separation technique [24] by measuring the scene under spatially high-frequency illumination, assuming that the indirect reflection component of a scene changes gradually in space, which means it is predominantly low-frequency. + +Unifying both advantages for scene analysis, spectral layer structure analysis effectively separates surface and internal details, offering notable benefits such as accurate + +![](images/1715ca1ce09a95b21f48a800591fcc298bdf462559a6481de2eebe85884c68cd.jpg) +Figure 1. The proposed system captures a single hyperspectral image using a wavelength-dependent, spatially shifted illumination pattern generated by a projector with a diffraction grating. The input shows the observed image under the proposed illumination system (displayed as an RGB image) along with the separated direct and indirect reflectance components for each wavelength ("Direct/Indirect"). Rough textures and specular reflections are observed in the direct component, while soft textures, such as internal scattering, are observed in the indirect component. The leftmost plots show the comparison of our estimation with the reference at points $p_1$ and $p_2$ . + +modeling, realistic rendering, and advanced editing. It also provides a non-destructive, non-invasive way to measure surface and internal spectral properties. This enables a wide range of applications, including medical analysis, such as diagnosing skin abnormalities [40], food inspection [20], and painting analysis for revealing the layered structure of pigments [18]. These analyses can be further enhanced by separating the direct and indirect components, where faster examination is preferred. It also reduces burdens on human subjects and minimizes light exposure for sensitive targets like art and biomedical samples. However, achieving this remains challenging, as a hyperspectral camera requires longer acquisition times than a conventional camera, making it unsuitable for computational photography approaches that rely on multiple pattern irradiations and acquisitions. In recent years, there have been attempts to separate reflection components from a single captured image by methods based on basis representation [30] and deep learning [9], but their method requires a dataset of a large number of images and which is hard to apply for hyperspectral scenarios due to the measurement time of hyperspectral images. + +To address the problem of time-consuming hyperspectral imaging, CASSI (Coded Aperture Snapshot Spectral Imaging System) [1] which enables hyperspectral image snapshots [3, 36], techniques for estimating information on other modalities from spectral images [16, 22, 25], and methods for achieving both [2, 29] have been proposed. Snapshot hyperspectral imaging methods are effective for hyperspectral image reconstruction; however, their implementation often requires complex imaging systems, and achieving measurements with sufficient accuracy for further scene analysis necessitates considerable effort. There is an approach that, given a single hyperspectral image, separates the first and second bounce components based on the shape of the observed spectral distribution at each point [25]. However, this method is limited to scenes composed of only a few colors, imposing constraints on its applicability. A more general approach capable of handling complex scenes has yet to be realized. In summary, hyperspectral reconstruction and scene analysis using hyperspectral images have been studied in the past, but a robust solution for achieving spectral direct/indirect separation of complex scenes composed of multiple materials using a single hyperspectral image has yet to be realized. + +To tackle this challenging problem, we develop a novel illumination system that integrates dispersive optics to generate a wavelength-dependent, spatially shifted illumination pattern. Similar to color photometric stereo, where varying lighting directions are achieved through different color channels, our system leverages observed wavelength differences to achieve a shifted illumination pattern. We demonstrate that the challenges associated with hyperspectral-based direct/indirect separation are resolved by our system. + +By accounting for spectral reflectance differences appropriately, we demonstrate that a single hyperspectral image captured under one illumination pattern can achieve robust separation of direct and indirect components for each wavelength. Furthermore, we show that analyzing variations across wavelengths aids in estimating scene depth. + +Our contribution is threefold: + +- A novel illumination system incorporating dispersive optics is proposed, enabling the illumination of a wavelength-dependent, spatially shifted pattern. +- By carefully addressing the differences in spectral reflectance, we formulate the observed pixel values based on these spectral-dependent illumination patterns and derive a solution for direct/indirect separation at each wavelength using a single hyperspectral image with one illumination pattern. +- We develop a real system and demonstrate robust spectral direct/indirect separation across various objects. Additionally, we show that analyzing differences across wavelengths aids in estimating scene depth. + +# 2. Related Works + +# Separation of Direct and Indirect Reflection Components via Pattern Illumination + +Nayar et al. [24] proposed separating the direct and indirect reflection components of a scene using high-frequency illumination patterns, and various extended methods have also been proposed. As the applications for acquiring rich information, O'Toole et al. [26, 27] proposed an image measurement method that can observe light transport paths in a scene by modulating both the illumination pattern of the projector and the camera response. As the high-speed application, separation of direct and indirect reflections by time dithering of a DLP projector with a high-speed camera was proposed [23], and extended to multiple colors [34]. More recently, a fast-capturing and processing method based on machine learning using event data was proposed [38] by a hybrid camera system combining an event camera and an RGB camera. Thus, many methods have been proposed to speed up the separation method of direct and indirect reflection components using pattern illumination, and extensions to multispectral with a small number of channels have been made, but no extension of patterns to hyperspectral image acquisition or spectral dimension has been achieved. + +On the other hand, several methods have been proposed to separate the direct and indirect reflection components of a scene using a single image. Subpa-Asa et al. [30] propose to evaluate the data-independent Fourier basis and learning-based principal component analysis (PCA) basis to find a better linear basis representation of the direct and indirect components. Duan et al. [9] proposed a deep learning solution for the separation of reflective components from a single image by designing a non-binary pattern with 4 or + +9 intensities instead of a binary grid pattern. Although the method of separating reflection components from a single image has theoretical applicability to hyperspectral images, in practice it is difficult to apply because it requires a data set with many images. + +# Decomposition-based Spectral Reflectance Separation + +Separation of the direct and indirect spectral reflectance can also be treated as a spectral decomposition problem. Several studies based on spectral decomposition problems using hyperspectral images have been proposed. In the field of remote sensing, linear or nonlinear spectral decomposition methods have been proposed to address the problems of spectral variation due to the atmosphere and/or spectral mixing due to the imaging system [4]. In the field of computer vision, Zheng et al. [37] has enabled spectral separation between illumination and reflectance in hyperspectral images taken under common spectral illumination and extended to multiple light sources by Chen et al. [5]. Regarding the separation of specular reflection components on the target surface using spectral information, an analysis of chromaticity-intensity space in diffuse and specular reflections [33], a non-iterative method based on a dichroic reflection model and PCA [11], and a method based on blind source separation using non-negative matrix factorization [39] have been proposed. + +As the application of separating indirect reflection, Okawa et al. [25] proposed a method to separate interreflection components from a single hyperspectral image when the object color is one or two colors, based on the fact that not only surfaces of different colors but also surfaces of the same color change the observed spectrum due to interreflection. Moreover, Fu et al. [12] proposed a method to separate the reflection and fluorescence emission spectra in a scene by using high-frequency illumination in the spectral domain. Their method achieves high-frequency illumination in the spectral domain using an equalized light source, whose wavelength characteristics can be freely controlled, but it is uniform in the spatial domain and cannot be used to separate direct and indirect reflection components. Thus, while several possible ways to separate reflection components using hyperspectral images exist, only interreflection or fluorescence has been considered for the separation of indirect components. + +# Dispersive Optics + +Dispersive optics, such as prisms and diffraction gratings, have often been used in camera and display design, especially for hyperspectral imaging. Mohan et al. [21] proposed a spectral mask that can control the spectral power distribution of a camera or projector using a diffraction grating and an attenuation mask. Hostettler et al. [15] used multiple prisms and calculated masks to realize a projector system that projects the desired color image without a conventional projector. CASSI uses dispersion optics and coding + +masks to acquire masked spectral images with wavelength-dependent transformations and reconstructs the hyperspectral data cube using an iterative algorithm [1, 36]. Cao et al. [3] achieved fast hyperspectral image acquisition with a hybrid camera system constructed from an RGB camera, a grayscale camera, and multiple optical elements including prisms and coding masks. + +Several methods have been proposed to acquire spectral and 3D depth information using dispersive optics. Baek et al. [2] proposed the diffractive optical element (DOE) generates a unique point spread function (PSF) that changes with the scene's spectrum and depth, enabling single-shot hyperspectral 3D imaging through blur analysis. Shin et al. [29] proposed a system to obtain spectral and 3D depth information by analyzing which structured light patterns disperse and illuminate a scene, using a compact setting that adds diffraction grating film to a projector-camera system. + +Dispersive optics have been utilized for projector/camera enhancement, hyperspectral image acquisition, and depth estimation. This field holds great potential for various applications. In this study, we demonstrate the effectiveness of direct/indirect separation. + +# 3. Method + +We separate the direct and indirect reflection components of a scene from a single-shot hyperspectral image taken under a wavelength-dependent, spatially shifted illumination pattern generated by a DMD (Digital Micromirror Device) projector with a diffraction grating (Figure 1). We first revisit the traditional approach to direct and indirect reflections, utilizing varying high-frequency projected patterns, and then extend the discussion to our case, which involves a wavelength-dependent, spatially shifted illumination pattern. + +# 3.1. Direct/Indirect Separation Using Sinusoidal Pattern Illumination + +When a single light source illuminates a scene, the direct component $I_{d}(x)$ is defined as the light that travels directly from the source, reflects once at a surface point, and reaches the camera, while the indirect component $I_{i}(x)$ encompasses all light that undergoes multiple interactions, such as subsurface scattering, interreflections, and volumetric scattering. Nayar et al. [24]. The observed intensity $I(x)$ by a camera is represented as + +$$ +I (x) = I _ {d} (x) + I _ {i} (x). \tag {1} +$$ + +Assuming that the indirect reflection component of a scene changes gradually in space and is predominantly low-frequency, their proposed approach [24] demonstrated that the scene's direct and indirect reflection components can be separated by irradiating three spatially high-frequency sinusoidal patterns. + +![](images/14a5663ce8d325f61e68b049760b56f8df24e403ff3e83ef01870b50a816ad07.jpg) +Figure 2. The outline of the proposed spatio-spectral illumination system. (a) The (DMD) projector generates a sinusoidal pattern that has a broad spectral band of light. (b) The diffraction grating disperses the generated light pattern, then the sinusoidal pattern is spatially shifted according to wavelength. (c) The spatio-spectral shifted patterns are observed at each point of the scene, and the sinusoidal pattern is resampled in the spectral domain. + +Under the high-frequency sinusoidal pattern illuminations, the observed images $I_{n}(x,k)$ are represented as: + +$$ +I _ {n} (x, k) = \frac {1}{2} I _ {d} (x) (\sin (\frac {2}{3} k \pi + \phi) + 1) + \frac {1}{2} I _ {i} (x), \quad (2) +$$ + +where $\phi \in [0,2\pi]$ denotes an arbitrary phase of the sinusoidal function and $k = \{1,2,3\}$ represents the index of the pattern. This problem has three unknowns $\{I_d,I_i,\phi \}$ and can be solved theoretically for each pixel using three images taken under three different sinusoidal pattern illumination. In practice, this approach requires pixel-by-pixel camera-projector alignment or additional patterns to avoid artifacts and ensure accuracy. + +# 3.2. Spectral Direct/Indirect Separation + +Let us further extend the discussion to separation at the wavelength level by considering the spectral reflection of a scene. Here, we represent the direct reflection component $I_{d}(x,\lambda)$ as the product of normalized spectral reflectance of direct component $R_{d}(x,\lambda)$ , its magnitude $\alpha_{d}(x)$ , and the intensity of the spectral light source $L_{d}(x,\lambda)$ at the point $x$ . + +Following [24], the indirect reflection component is represented as: + +$$ +\begin{array}{l} I _ {i} (x, \lambda) = \sum_ {x ^ {\prime} \in S (x)} A ([ x, x ^ {\prime} ], \lambda) L ([ x, x ^ {\prime} ], \lambda) \tag {3} \\ = \alpha_ {i} (x) R _ {i} (x, \lambda), \\ \end{array} +$$ + +where $S(x)$ denotes the set of surrounding pixels for $x$ , $L([x, x'], \lambda)$ represents the spectral radiance at $x'$ in the direction of $x$ , and $A([x, x'], \lambda)$ incorporates the spectral BRDF of $x$ as well as the relative geometric configuration between $x$ and $x'$ .1 $I_i(x, \lambda)$ is further expressed in terms of the normalized spectrum $R_i(x, \lambda)$ which is computed from the total sum of light contributions coming to $x$ from surrounding pixels $S(x)$ and its magnitude $\alpha_i(x)$ . From this, + +the spectral reflection model of a scene $I(x,\lambda)$ is given as: + +$$ +\begin{array}{l} I (x, \lambda) = I _ {d} (x, \lambda) + I _ {i} (x, \lambda) \\ = \alpha_ {d} (x) R _ {d} (x, \lambda) L _ {d} (x, \lambda) + \alpha_ {i} (x) R _ {i} (x, \lambda). \tag {4} \\ \end{array} +$$ + +In the following, we introduce a spatial phase shift by designing the lighting pattern to shift horizontally with wavelength. + +# 3.3. Spatially Shifting Illumination According to Wavelength + +We propose a spatio-spectral pattern illumination system that combines a conventional projector and a diffraction grating to create spatially shifting illumination patterns according to wavelength. The configuration of the illumination system is shown in Figure 2. Let the spatially high-frequency sinusoidal pattern $P_0(v)$ generated by the conventional projector (Figure 2 (a)) be written as a sinusoidal function of the horizontal pixel position $v$ of the projector: + +$$ +P _ {0} (v) = \frac {l _ {0}}{2} \left(\sin \left(f _ {0} v + \phi_ {0}\right) + 1\right), \tag {5} +$$ + +where $l_{0}$ is the intensity of the projector, $f_{0}$ describes the frequency of the sinusoidal function, and $\phi_0$ denotes the phase of the sinusoidal function. Then the grating diffracts the incoming pattern light and reflects it at slightly different angles for each wavelength of light (Figure 2 (b)). In the case of first-order diffracted light, the diffraction angle of the outgoing light, i.e. the emission angle, varies monotonically in the visible light range depending on the wavelength of the incident light. Thus, the illumination pattern varies in the spatio-spectral domain, similar to a rainbow. Therefore, as shown in Figure 3, when the diffracted pattern is observed at a scene point, it is also a sinusoidal pattern in the spectral domain (Figure 2 (c)). We proceed with the discussion assuming that projector's origin is the virtual origin of the pattern reflected by the grating. We also set the cam + +era's origin at the same location to enable discussion within a unified coordinate system. + +![](images/d064957930dd665ff8b729a3c4eb8977dc9ba020ab92c81794fb1ecdfc16716b.jpg) +Figure 3. After passing through the diffraction grating, the sinusoidal pattern generated by the conventional projector (following yellow arrow) is resampled in the spectral domain at each point of the scene (following red arrow). + +The spectral pattern $P(x, \lambda)$ is expressed as the function of wavelength $\lambda$ at the scene point corresponding to the projector's horizontal pixel position $x$ as in Equation (6): + +$$ +P (x, \lambda) = \frac {l (\lambda)}{2} \left(\sin \left(f _ {p} (x) \lambda + \phi (x)\right) + 1\right). \tag {6} +$$ + +Let $l(\lambda)$ be the spectral intensity of the projector, $f_{p}(x)$ denote the frequency of the sinusoidal function, and $\phi (x)$ represent the phase of the sinusoidal function. We calibrate the spectral intensity of the projector $l(\lambda)$ in advance and describe it as the constant value $l$ . + +# 3.4. Simple Case + +We begin our discussion with the simplest case: illuminating our pattern onto a gray object with flat spectral reflectance and spatially varying albedo. In this case, the normalized spectra $R_{d}$ and $R_{i}$ of the direct and indirect components can be integrated and represented as the respective magnitudes $\alpha_{d}$ and $\alpha_{i}$ . When illuminating our pattern onto this object, the direct light component $L_{d}(x,\lambda)$ in Equation (4) is replaced by $P(x,\lambda)$ , and the indirect reflection component $I_{i}(x,\lambda)$ is reduced to half intensity and pattern independent following [24]. Hence, the observed reflection $I_{p}(x,\lambda)$ under the spatio-spectral pattern illumination $P(x,\lambda)$ is modeled as: + +$$ +I _ {p} (x, \lambda) = \alpha_ {d} (x) P (x, \lambda) + \frac {1}{2} \alpha_ {i} (x). \tag {7} +$$ + +Expanding Equation (7) based on Equation (6) yields: + +$$ +I _ {p} (x, \lambda) = \frac {l \alpha_ {d} (x)}{2} (\sin (f _ {p} (x) \lambda + \phi (x)) + 1) + \frac {\alpha_ {i} (x)}{2}. \tag {8} +$$ + +In this case, we can obtain a sinusoidal pattern in the spectral domain along wavelength axis at any point $x$ as shown in Figure 4 (a). + +Since the number of unknowns is four $\{\alpha_{d}(x),f_{p}(x),\phi (x),\alpha_{i}(x)\}$ , the direct and indirect reflection components of a scene can be separated by acquiring spectral images at least 4 spectral channels in the simple case. + +![](images/5e46589b634ec966a4df529ed3d186e4536c3caf47cfc6002ff72a02551e00a1.jpg) +Figure 4. Our illumination pattern and the observation as it is projected onto the scene. (a) For gray objects, a sinusoidal pattern is observed in the spectral domain. (b) For colored objects, a pattern that varies with the spectral reflectance of the direct and indirect components of the object is observed in the spectral domain. + +![](images/f6ece85625d22b329786e6cf62d587ac491bc95261557d072622a400e03247db.jpg) + +![](images/c7a0cdda9049d4f793a423b4695d3ffe63b042834e79b05011d413c122730b99.jpg) + +# 3.5. Complex Case + +Next, let us examine the case of the object that has non-flat, spatially varying spectral reflectance. As shown in Figure 4 (b), the spectrum observed for a scene with non-flat spectral reflectance is affected by the spectral reflectance of the scene, unlike the simple flat reflectance case in Figure 4 (a). + +Based on Equation (4), a scene can be composed of direct and indirect reflection components, each of which can have its spectral distribution. The observed intensity $I_{p}(x,\lambda)$ at the scene point $x$ under the spatio-spectral pattern illumination $P(x,\lambda)$ can be expressed: + +$$ +I _ {p} (x, \lambda) = \alpha_ {d} (x) R _ {d} (x, \lambda) P (x, \lambda) + \frac {\alpha_ {i} (x)}{2} R _ {i} (x, \lambda) \tag {9} +$$ + +Expanding Equation (9) based on Equation (6) yields: + +$$ +\begin{array}{l} I _ {p} (x, \lambda) = l \alpha_ {d} (x) R _ {d} (x, \lambda) (\sin (f _ {p} (x) \lambda + \phi (x)) + 1) / 2 \\ + \quad \alpha_ {i} (x) R _ {i} (x, \lambda) / 2. \tag {10} \\ \end{array} +$$ + +Furthermore, Equation (10) can be expressed in the form of a general sinusoidal function for $\lambda$ as follows: + +$$ +I_{p}(x,\lambda) = A(x,\lambda)\{\sin (B(x)\lambda +C(x)) + 1\} +D(x,\lambda), +$$ + +$$ +A (x, \lambda) = l \alpha_ {d} (x) R _ {d} (x, \lambda) / 2, B (x) = f _ {p} (x), +$$ + +$$ +C (x) = \phi (x), D (x, \lambda) = \alpha_ {i} (x) R _ {i} (x, \lambda) / 2. +$$ + +(11) + +Equation (11) shows that the amplitude of the sinusoidal function is related to the direct reflection component and the constant term is related to the indirect reflection component. + +We then discuss how to separate the direct and indirect reflection components from the observed values based on Equation (11). The number of observations at each pixel obtained in a single hyperspectral image measurement equals the number of spectral wavelength observations. Therefore, if either spectral reflectance $R_{d}$ and $R_{i}$ are unknown, solving Equation (11) from a single multi- or hyperspectral input image is an ill-posed problem. We address this problem + +by representing $R_{d}$ and $R_{i}$ as the combination of the spectral reflectance basis, following the work of Han et al. [14]: + +$$ +\bar {\boldsymbol {R}} _ {\boldsymbol {d}} (\boldsymbol {x}) = \boldsymbol {B} \boldsymbol {c} _ {\boldsymbol {d}} (\boldsymbol {x}), \bar {\boldsymbol {R}} _ {\boldsymbol {i}} (\boldsymbol {x}) = \boldsymbol {B} \boldsymbol {c} _ {\boldsymbol {i}} (\boldsymbol {x}), \tag {12} +$$ + +where $\bar{R}_d(x)$ and $\bar{R}_i(x) \in \mathbb{R}^f$ are the vectorized form of the normalized spectral reflectances at the pixel $x$ approximated by the basis representation, $B \in \mathbb{R}^{f \times k}$ is a basis matrix consisting of $k$ basis vectors, and $c_d(x)$ and $c_i(x) \in \mathbb{R}^k$ are the basis coefficients at the pixel $x$ . $\bar{R}_i(x)$ is not pure spectral reflectance, but it is the sum of the spectral BRDFs in the surrounding pixels and can therefore be expressed on the same basis. Based on the public dataset [28], in which the 1257 spectral reflectances of Munsell Book of Color: Matte Finish Collection [6] are recorded, $k$ spectral reflectance basis vectors are extracted using singular value decomposition. Approximation by the basis reduces the number of unknowns related to the spectral reflectance to $k$ . Thus, unknowns of our problem are $\{c_d(x), c_i(x), \alpha_d(x), f_p(x), \phi(x), \alpha_i(x)\}$ and its total degree of freedom is $2k + 4$ . + +# 3.6. Optimization + +We solve the problem described in Section 3.5 by optimization, using a single hyperspectral image captured by our spatio-spectral pattern illumination system as input. For simplicity, let us focus on a single pixel, with all terms written while omitting $x$ . The error function between the input spectral image and the model is computed as: + +$$ +E _ {s} (\lambda) = \left| I _ {p} (\lambda) - \frac {l \alpha_ {d}}{2} \bar {R} _ {d} (\lambda) \left\{\sin \left(f _ {p} \lambda + \phi\right) + 1 \right\} - \frac {\alpha_ {i}}{2} \bar {R} _ {i} (\lambda) \right|. \tag {13} +$$ + +Where $\bar{R}_d(\lambda)$ and $\bar{R}_i(\lambda)$ denote the scalar form of the normalized spectral reflectances approximated by the basis representation. The objective function of the problem is: + +$$ +\min _ {\boldsymbol {c} _ {d}, \boldsymbol {c} _ {i}, \alpha_ {d}, f _ {p}, \phi , \alpha_ {i}} \sum_ {\lambda} ^ {k} E _ {s} (\lambda). \tag {14} +$$ + +As mentioned earlier, our problem setup has $2k + 4$ unknowns, so the simplest solution is to give a hyperspectral image with $2k + 4$ spectral channels as the input image. From previous study [14], the number of bases for approximating the spectral reflectance of a general object with a basis representation is about 8. Therefore, when we follow them and let $k = 8$ , we require a hyperspectral image with 20 spectral channels to solve. When applying hyperspectral imaging to digital archiving and other applications, it is common to work in visible light, with wavelengths ranging from $400\mathrm{nm}$ to $700\mathrm{nm}$ . Therefore, acquiring multi- or hyperspectral images with a wavelength interval of at least 15 nm in the visible light range is sufficient. + +For robust optimization, we use initial values and ranges of estimation values for some of the parameters. As shown + +in the figure, the period of the sinusoidal function of the spectral domain at each pixel can be predicted from the local maxima and minima of the observed signal. The period obtained from the median of the intervals at the local maxima and minima is the initial value of the period $f_{p}$ at that pixel, and the search range for optimization is $\pm 5\%$ of the initial value to account for discretization error. Also, the search range of the phase $\phi$ is $[-\pi, \pi]$ and the magnitude of the direct and indirect components $\alpha_{d}$ and $\alpha_{i}$ are $[0, \infty]$ . + +![](images/a1bb872fc1433ff66bfc24296c2926eb3f5d29862117543c801aac4b8aaacafc.jpg) +Figure 5. Examples of the observed signal in the spectral domain. Orange and green circles represent local maxima and minima. + +Furthermore, since spectral reflectance is known to be smooth for most common objects [13], we introduce a loss function that reflects the assumption of smoothness of spectral reflectance, following prior works [19, 32]: + +$$ +\min _ {\boldsymbol {c} _ {d}, \boldsymbol {c} _ {i}, f _ {p}, \phi , \alpha_ {i}} \sum_ {\lambda} ^ {k} E _ {s} (\lambda) + \gamma_ {d} | \boldsymbol {B c} _ {d} \boldsymbol {\Gamma} | + \gamma_ {i} | \boldsymbol {B c} _ {i} \boldsymbol {\Gamma} |. \tag {15} +$$ + +The weights $\gamma_{d}$ and $\gamma_{i}$ are chosen empirically and $\Gamma$ denotes a matrix that gives smoothness by the subtraction between the neighbor elements. + +# 4. Experimental Results + +# 4.1. Experimental Setup and Conditions + +Our system consists of a xenon light source (Asahi Spectra LAX-100), a DMD (Ajile AJD-4500), a hyperspectral camera (EBA JAPAN NH-8), lenses for the DMD and hyperspectral camera (Fujinon HF25SA-1), and a diffraction grating (Edmund Optics 1200 Grooves, $50\mathrm{mm}$ Square, $500\mathrm{nm}$ Ruled Diffraction Grating), as shown in Figure 1. The hyperspectral camera captures spectral images at $5\mathrm{nm}$ intervals within the $380 - 1000\mathrm{nm}$ wavelength range. Considering the effective wavelength range of the diffraction grating, we performed separation in the $450 - 650\mathrm{nm}$ range. + +The minimum number of observations required to solve our problem is 20, as discussed in Section 3.6. However, actual spectral images contain significant noise, and factors such as projector and camera calibration errors further impact accuracy, preventing us from achieving stable results with the minimum number of observations. In the following experiments, we use 41 spectral bands obtained at $5\mathrm{nm}$ intervals to ensure robust separation against noise. + +# 4.2. Results on Real Scenes + +Figure 8 shows the results of estimating the spectral reflectance of the direct and indirect reflectance components in various real-world objects. For reference, we captured a set of images of the scene using a hyperspectral camera with different sinusoidal pattern illuminations and provided the separation results using Nayar's method [24], as described in Equation (2). Hyperspectral images contain significant noise, making it challenging to obtain stable separation results free from artifacts by Nayar's method [24] unless numerous hyperspectral images are captured under various sinusoidal pattern illuminations. To mitigate such artifacts, the spatial resolution of the input images for the reference results, but not ours, is downsampled for visual comparison. All input hyperspectral images and estimated results are available in the supplementary material. + +For all four scenes in Figure 8, the estimated direct $(I_d)$ and indirect $(I_i)$ components closely resemble the corresponding reference across different wavelengths. From these examples, we observe that wavelength-dependent changes in the direct and indirect components are extracted. In particular, in the results for the Potato and Ball (top 2 rows) and the Candle (bottom 2 rows), the scattering inside the objects becomes broader and stronger for longer wavelengths. Fine textures can be seen in the direct reflection component and rarely in the indirect reflection component. The leftmost graph compares our estimation and reference for points $p_1$ and $p_2$ . The estimated spectra well represent that of the reference; nevertheless, they are derived solely from a single hyperspectral image. Moreover, Figure 6 shows spatial maps and values of error metrics. We used MSE (mean squared error) and SAM (spectral angle mapper) as error metrics. + +# 4.3. Depth Dependence of the Pattern's Period + +Figure 7 shows the spatial distribution of the estimated period $\phi$ in Equation (15). The period $\phi$ increases as the distance from the projector becomes longer, making it highly correlated with object distance. In Figure 7, 'DAY' is placed closer to the camera, and 'HAP' is farther away. This example demonstrates that the estimated $\phi$ contains valuable information about the scene's depth. + +# 5. Conclusion + +In this work, we proposed a novel method for separating the direct and indirect reflection components of a scene from a single hyperspectral image. By introducing a projection system incorporating dispersive optics, we enabled the projection of a wavelength-dependent, spatially shifted pattern. We formulated the observed pixel values under these spectral-dependent illumination patterns and derived a solution for direct and indirect reflection separation at each + +![](images/a92b5762de635944f01719392b9a1a340c309b9151352a3a1bfd5bffc1a21b24.jpg) +Figure 6. Error maps and metrics between the reference and the proposed method. The values represent the mean and variance of the MSE and SAM. Note that some stripe patterns in the SAM maps are affected by the artifacts of the reference. + +![](images/32579dbc8e51b3945fb1fc27c38a51786605a9040eb48fb54323874eaceb8086.jpg) +Figure 7. Spatial map of the estimated period $\phi$ for a scene. 'DAY' is placed closer to the camera, and 'HAP' is farther away. + +wavelength using only a single hyperspectral image with one projection pattern. Through experiments, we demonstrated the effectiveness of our system in achieving robust spectral direct/indirect separation even for objects with complex spectral reflectance. Furthermore, our analysis of differences across wavelengths was leveraged for scene depth estimation. + +# Acknowledgments + +This research is supported by JSPS Kakenhi Grant number 25K21236 and JST-Mirai Program Grant Number JP-MJM123G1. + +![](images/15acc3230c29c7aa979a14b3f2ef66c2d15b148061df512cf918d9d524a3433e.jpg) +Figure 8. Separation results of direct and indirect reflection components in various real-world objects. Top row, from left to right: scene captured by the hyperspectral camera (RGB converted), results of separation of direct and indirect reflection components $(I_d(480), I_d(620), I_i(480), I_i(620))$ by Nayar's method [24] as a reference, direct reflectance component (red) at point $p_1$ and indirect, comparison of spectral reflectance of direct (red) and indirect (green) components at point $p_1$ for the reference (dotted line) and the proposed method (solid line). Bottom row, from left to right: part of the input spectral images in the proposed method are superimposed on a pseudo-RGB, estimation results of direct and indirect reflection components $(I_d(480), I_d(620), I_i(480), I_i(620))$ by the proposed method, comparison of spectral reflectance at $p_2$ . + +# References + +[1] Gonzalo R Arce, David J Brady, Lawrence Carin, Henry Arguello, and David S Kittle. Compressive coded aperture spectral imaging: An introduction. IEEE Signal Processing Magazine, 31(1):105-115, 2013. 2, 3 +[2] Seung-Hwan Baek, Hayato Ikoma, Daniel S Jeon, Yuqi Li, Wolfgang Heidrich, Gordon Wetzstein, and Min H Kim. Single-shot hyperspectral-depth imaging with learned diffractive optics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2651-2660, 2021. 2, 3 +[3] Xun Cao, Xin Tong, Qionghai Dai, and Stephen Lin. High resolution multispectral video capture with a hybrid camera system. In CVPR 2011, pages 297-304. IEEE, 2011. 2, 3 +[4] Xavier Ceamanos and Silvia Valero. 4 - processing hyperspectral images. In Optical Remote Sensing of Land Surface, pages 163-200. Elsevier, 2016. 3 +[5] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 3 +[6] Munsell Color. Munsell book of color: Matte finish collection: Kollmorgen corporation. Baltimore, Maryland, 1976. 6 +[7] Costanza Cucci, John K Delaney, and Marcello Picollo. Reflectance hyperspectral imaging for investigation of works of art: old master paintings and illuminated manuscripts. Accounts of chemical research, 49(10):2070-2079, 2016. 1 +[8] Craig Donner and Henrik Wann Jensen. A spectral bssrfd for shading human skin. Rendering techniques, 2006:409-418, 2006. 1 +[9] Zhaoliang Duan, James Bieron, and Pieter Peers. Deep Separation of Direct and Global Components from a Single Photograph under Structured Lighting. Computer Graphics Forum, 2020. 2 +[10] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Transactions on graphics (TOG), 37(6):1-14, 2018. 1 +[11] Gamal ElMasry, Pere Gou, and Salim Al-Rejaie. Effectiveness of specularity removal from hyperspectral images on the quality of spectral signatures of food products. Journal of Food Engineering, 289:110148, 2021. 3 +[12] Ying Fu, Antony Lam, Imari Sato, Takahiro Okabe, and Yoichi Sato. Separating reflective and fluorescent components using high frequency illumination in the spectral domain. In Proceedings of the IEEE International Conference on Computer Vision, pages 457-464, 2013. 3 +[13] Lewis D Griffin. Reconciling the statistics of spectral reflectance and colour. PLoS One, 14(11):e0223069, 2019. 6 +[14] Shuai Han, Imari Sato, Takahiro Okabe, and Yoichi Sato. Fast spectral reflectance recovery using dlp projector. International journal of computer vision, 110:172-184, 2014. 6 +[15] Rafael Hostettler, Ralf Habel, Markus Gross, and Wojciech Jarosz. Dispersion-based color projection using masked prisms. In Computer Graphics Forum, pages 329-338. Wiley Online Library, 2015. 3 + +[16] Shin Ishihara, Yuta Asano, Yinqiang Zheng, and Imari Sato. Underwater scene recovery using wavelength-dependent refraction of light. In 2020 International Conference on 3D Vision (3DV), pages 32-40. IEEE, 2020. 2 +[17] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31 (4):1-11, 2012. 1 +[18] Paul Kubelka and Franz Munk. An article on optics of paint layers. Z. Tech. Phys, 12(593-601):259-274, 1931. 2 +[19] Wei Liu, Chengxun He, and Le Sun. Spectral-smoothness and non-local self-similarity regularized subspace low-rank learning method for hyperspectral mixed denoising. Remote Sensing, 13(16):3196, 2021. 6 +[20] Lembe Samukelo Magwaza and Umezuruike Linus Opara. Analytical methods for determination of sugars and sweetness of horticultural products—a review. Scientia Horticulturae, 184:179-192, 2015. 2 +[21] Ankit Mohan, Ramesh Raskar, and Jack Tumblin. Agile spectrum imaging: Programmable wavelength modulation for cameras and projectors. In Computer graphics forum, pages 709-717. Wiley Online Library, 2008. 3 +[22] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE computer graphics and applications, 34(6):57-68, 2014. 2 +[23] Srinivasa G Narasimhan, Sanjeev J Koppal, and Shuntaro Yamazaki. Temporal dithering of illumination for fast active vision. In Computer Vision-ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10, pages 830-844. Springer, 2008. 2 +[24] Shree K Nayar, Gurunandan Krishnan, Michael D Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. In ACM SIGGRAPH 2006 Papers, pages 935-944. 2006. 1, 2, 3, 4, 5, 7, 8 +[25] Hiroki Okawa, Yinqiang Zheng, Antony Lam, and Imari Sato. Spectral reflectance recovery with interreflection using a hyperspectral image. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13, pages 52-67. Springer, 2017. 2, 3 +[26] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2 +[27] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 2 +[28] Jussi PS Parkkinen, Jarmo Hallikainen, and Timo Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical society of America A, 6(2):318-322, 1989. 6 +[29] Suhyun Shin, Seokjun Choi, Felix Heide, and Seung-Hwan Baek. Dispersed structured light for hyperspectral 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24997-25006, 2024. 2, 3 + +[30] Art Subpa-asa, Ying Fu, Yinqiang Zheng, Toshiyuki Amano, and Imari Sato. Separating the direct and global components of a single image. Journal of Information Processing, 26: 755-767, 2018. 2 +[31] Da-Wen Sun. Hyperspectral imaging for food quality analysis and control. Elsevier, 2010. 1 +[32] Le Sun, Qihao Cheng, and Zhiguo Chen. Hyperspectral image super-resolution method based on spectral smoothing prior and tensor tubal row-sparse representation. Remote Sensing, 14(9):2142, 2022. 6 +[33] Robby T Tan, Ko Nishino, and Katsushi Ikeuchi. Separating reflection components based on chromaticity and noise analysis. IEEE transactions on pattern analysis and machine intelligence, 26(10):1373-1379, 2004. 3 +[34] Moriaki Torii, Takahiro Okabe, and Toshiyuki Amano. Multispectral direct-global separation of dynamic scenes. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1923-1931. IEEE, 2019. 2 +[35] Mahesh Kumar Tripathi and H Govil. Evaluation of aviris-ng hyperspectral images for mineral identification and mapping. Heliyon, 5(11), 2019. 1 +[36] Ashwin Wagadarikar, Renu John, Rebecca Willett, and David Brady. Single disperser design for coded aperture snapshot spectral imaging. Applied optics, 47(10):B44-B51, 2008. 2, 3 +[37] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 3 +[38] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. Evdig: Event-guided direct and global components separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9612-9621, 2024. 2 +[39] Mustapha Zokay and Hicham Saylani. Removing specular reflection in multispectral dermatological images using blind source separation. In Annual Conference on Medical Image Understanding and Analysis, pages 734-750. Springer, 2022. 3 +[40] George Zonios, Julie Bykowski, and Nikiforos Kollias. Skin melanin, hemoglobin, and light scattering properties can be quantitatively assessed in vivo using diffuse reflectance spectroscopy. Journal of Investigative Dermatology, 117(6): 1452-1457, 2001. 2 \ No newline at end of file diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/images.zip b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..50d197f691c2ef87136237e60969c6d814f6943a --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ace5622d4b109cfd79e2b5786bfc7a48566b1666651cbb4890bc691fb9df152f +size 611428 diff --git a/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/layout.json b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc12e3b36034a14787e3ccbbde5618228619a4cd --- /dev/null +++ b/2025/Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image/layout.json @@ -0,0 +1,8337 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "text", + "content": "Spatio-Spectral Pattern Illumination for Direct and Indirect Separation from a Single Hyperspectral Image" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "spans": [ + { + "bbox": [ + 236, + 161, + 374, + 175 + ], + "type": "text", + "content": "Shin Ishihara Imari Sato" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 176, + 383, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 176, + 383, + 188 + ], + "spans": [ + { + "bbox": [ + 227, + 176, + 383, + 188 + ], + "type": "text", + "content": "National Institute of Informatics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 190, + 417, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 190, + 417, + 204 + ], + "spans": [ + { + "bbox": [ + 194, + 190, + 417, + 204 + ], + "type": "text", + "content": "2-1-2 Hitotsubashi, Chiyoda-ku, Tokyo, Japan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 205, + 380, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 205, + 380, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 205, + 380, + 217 + ], + "type": "text", + "content": "{sishihara,imarik}@nii.ac.jp" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": "Hyperspectral imaging has proven effective for appearance inspection because it can identify material compositions and reveal hidden features. Similarly, direct/indirect separation provides essential information about surface appearance and internal conditions, including layer structures and scattering behaviors. This paper presents a novel illumination system incorporating dispersive optics to unify both advantages for scene analysis. In general, achieving distinct direct/indirect separation requires multiple images with varying patterns. In a hyperspectral scenario, using a hyperspectral camera or tunable filters extends exposure and measurement times, hindering practical application. Our proposed system enables the illumination of a wavelength-dependent, spatially shifted pattern. With proper consideration of reflectance differences, we demonstrate that robust separation of direct and indirect components for each wavelength can be achieved using a single hyperspectral image acquired under our single spatiospectral pattern illumination." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 245, + 394, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 245, + 394, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 394, + 257 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 268, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 473 + ], + "type": "text", + "content": "Hyperspectral imaging enables pixel-level analysis of objects and is widely used in fields such as medical imaging, food quality control, environmental monitoring, and art conservation [7, 17, 31, 35]. Spectral rendering is gaining traction, with modern renderers adopting wavelength-dependent models [10], and reflection/scattering analysis aiding in material editing [8]. Analyzing layer structures is also essential for inspecting an object's surface and internal conditions. Techniques for separating a scene's direct and indirect reflection components are often employed to accurately capture surface reflection colors and shapes, as well as to visualize an object's internal structure. Nayar et al. developed a groundbreaking fast separation technique [24] by measuring the scene under spatially high-frequency illumination, assuming that the indirect reflection component of a scene changes gradually in space, which means it is predominantly low-frequency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 512 + ], + "type": "text", + "content": "Unifying both advantages for scene analysis, spectral layer structure analysis effectively separates surface and internal details, offering notable benefits such as accurate" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 517, + 552, + 635 + ], + "blocks": [ + { + "bbox": [ + 59, + 517, + 552, + 635 + ], + "lines": [ + { + "bbox": [ + 59, + 517, + 552, + 635 + ], + "spans": [ + { + "bbox": [ + 59, + 517, + 552, + 635 + ], + "type": "image", + "image_path": "1715ca1ce09a95b21f48a800591fcc298bdf462559a6481de2eebe85884c68cd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "lines": [ + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "type": "text", + "content": "Figure 1. The proposed system captures a single hyperspectral image using a wavelength-dependent, spatially shifted illumination pattern generated by a projector with a diffraction grating. The input shows the observed image under the proposed illumination system (displayed as an RGB image) along with the separated direct and indirect reflectance components for each wavelength (\"Direct/Indirect\"). Rough textures and specular reflections are observed in the direct component, while soft textures, such as internal scattering, are observed in the indirect component. The leftmost plots show the comparison of our estimation with the reference at points " + }, + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 644, + 555, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 741, + 318, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 318, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 318, + 750 + ], + "type": "text", + "content": "26827" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": "modeling, realistic rendering, and advanced editing. It also provides a non-destructive, non-invasive way to measure surface and internal spectral properties. This enables a wide range of applications, including medical analysis, such as diagnosing skin abnormalities [40], food inspection [20], and painting analysis for revealing the layered structure of pigments [18]. These analyses can be further enhanced by separating the direct and indirect components, where faster examination is preferred. It also reduces burdens on human subjects and minimizes light exposure for sensitive targets like art and biomedical samples. However, achieving this remains challenging, as a hyperspectral camera requires longer acquisition times than a conventional camera, making it unsuitable for computational photography approaches that rely on multiple pattern irradiations and acquisitions. In recent years, there have been attempts to separate reflection components from a single captured image by methods based on basis representation [30] and deep learning [9], but their method requires a dataset of a large number of images and which is hard to apply for hyperspectral scenarios due to the measurement time of hyperspectral images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 327, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 601 + ], + "type": "text", + "content": "To address the problem of time-consuming hyperspectral imaging, CASSI (Coded Aperture Snapshot Spectral Imaging System) [1] which enables hyperspectral image snapshots [3, 36], techniques for estimating information on other modalities from spectral images [16, 22, 25], and methods for achieving both [2, 29] have been proposed. Snapshot hyperspectral imaging methods are effective for hyperspectral image reconstruction; however, their implementation often requires complex imaging systems, and achieving measurements with sufficient accuracy for further scene analysis necessitates considerable effort. There is an approach that, given a single hyperspectral image, separates the first and second bounce components based on the shape of the observed spectral distribution at each point [25]. However, this method is limited to scenes composed of only a few colors, imposing constraints on its applicability. A more general approach capable of handling complex scenes has yet to be realized. In summary, hyperspectral reconstruction and scene analysis using hyperspectral images have been studied in the past, but a robust solution for achieving spectral direct/indirect separation of complex scenes composed of multiple materials using a single hyperspectral image has yet to be realized." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "content": "To tackle this challenging problem, we develop a novel illumination system that integrates dispersive optics to generate a wavelength-dependent, spatially shifted illumination pattern. Similar to color photometric stereo, where varying lighting directions are achieved through different color channels, our system leverages observed wavelength differences to achieve a shifted illumination pattern. We demonstrate that the challenges associated with hyperspectral-based direct/indirect separation are resolved by our system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "By accounting for spectral reflectance differences appropriately, we demonstrate that a single hyperspectral image captured under one illumination pattern can achieve robust separation of direct and indirect components for each wavelength. Furthermore, we show that analyzing variations across wavelengths aids in estimating scene depth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 144, + 434, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 434, + 155 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 434, + 155 + ], + "type": "text", + "content": "Our contribution is threefold:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 156, + 553, + 311 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 191 + ], + "type": "text", + "content": "- A novel illumination system incorporating dispersive optics is proposed, enabling the illumination of a wavelength-dependent, spatially shifted pattern." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 193, + 553, + 262 + ], + "type": "text", + "content": "- By carefully addressing the differences in spectral reflectance, we formulate the observed pixel values based on these spectral-dependent illumination patterns and derive a solution for direct/indirect separation at each wavelength using a single hyperspectral image with one illumination pattern." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 264, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 553, + 311 + ], + "type": "text", + "content": "- We develop a real system and demonstrate robust spectral direct/indirect separation across various objects. Additionally, we show that analyzing differences across wavelengths aids in estimating scene depth." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 405, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 365 + ], + "type": "text", + "content": "Separation of Direct and Indirect Reflection Components via Pattern Illumination" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 605 + ], + "type": "text", + "content": "Nayar et al. [24] proposed separating the direct and indirect reflection components of a scene using high-frequency illumination patterns, and various extended methods have also been proposed. As the applications for acquiring rich information, O'Toole et al. [26, 27] proposed an image measurement method that can observe light transport paths in a scene by modulating both the illumination pattern of the projector and the camera response. As the high-speed application, separation of direct and indirect reflections by time dithering of a DLP projector with a high-speed camera was proposed [23], and extended to multiple colors [34]. More recently, a fast-capturing and processing method based on machine learning using event data was proposed [38] by a hybrid camera system combining an event camera and an RGB camera. Thus, many methods have been proposed to speed up the separation method of direct and indirect reflection components using pattern illumination, and extensions to multispectral with a small number of channels have been made, but no extension of patterns to hyperspectral image acquisition or spectral dimension has been achieved." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 714 + ], + "type": "text", + "content": "On the other hand, several methods have been proposed to separate the direct and indirect reflection components of a scene using a single image. Subpa-Asa et al. [30] propose to evaluate the data-independent Fourier basis and learning-based principal component analysis (PCA) basis to find a better linear basis representation of the direct and indirect components. Duan et al. [9] proposed a deep learning solution for the separation of reflective components from a single image by designing a non-binary pattern with 4 or" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26828" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "9 intensities instead of a binary grid pattern. Although the method of separating reflection components from a single image has theoretical applicability to hyperspectral images, in practice it is difficult to apply because it requires a data set with many images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 288, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 288, + 146 + ], + "type": "text", + "content": "Decomposition-based Spectral Reflectance Separation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 147, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 374 + ], + "type": "text", + "content": "Separation of the direct and indirect spectral reflectance can also be treated as a spectral decomposition problem. Several studies based on spectral decomposition problems using hyperspectral images have been proposed. In the field of remote sensing, linear or nonlinear spectral decomposition methods have been proposed to address the problems of spectral variation due to the atmosphere and/or spectral mixing due to the imaging system [4]. In the field of computer vision, Zheng et al. [37] has enabled spectral separation between illumination and reflectance in hyperspectral images taken under common spectral illumination and extended to multiple light sources by Chen et al. [5]. Regarding the separation of specular reflection components on the target surface using spectral information, an analysis of chromaticity-intensity space in diffuse and specular reflections [33], a non-iterative method based on a dichroic reflection model and PCA [11], and a method based on blind source separation using non-negative matrix factorization [39] have been proposed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 376, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 591 + ], + "type": "text", + "content": "As the application of separating indirect reflection, Okawa et al. [25] proposed a method to separate interreflection components from a single hyperspectral image when the object color is one or two colors, based on the fact that not only surfaces of different colors but also surfaces of the same color change the observed spectrum due to interreflection. Moreover, Fu et al. [12] proposed a method to separate the reflection and fluorescence emission spectra in a scene by using high-frequency illumination in the spectral domain. Their method achieves high-frequency illumination in the spectral domain using an equalized light source, whose wavelength characteristics can be freely controlled, but it is uniform in the spatial domain and cannot be used to separate direct and indirect reflection components. Thus, while several possible ways to separate reflection components using hyperspectral images exist, only interreflection or fluorescence has been considered for the separation of indirect components." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 133, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 133, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 133, + 605 + ], + "type": "text", + "content": "Dispersive Optics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "content": "Dispersive optics, such as prisms and diffraction gratings, have often been used in camera and display design, especially for hyperspectral imaging. Mohan et al. [21] proposed a spectral mask that can control the spectral power distribution of a camera or projector using a diffraction grating and an attenuation mask. Hostettler et al. [15] used multiple prisms and calculated masks to realize a projector system that projects the desired color image without a conventional projector. CASSI uses dispersion optics and coding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "masks to acquire masked spectral images with wavelength-dependent transformations and reconstructs the hyperspectral data cube using an iterative algorithm [1, 36]. Cao et al. [3] achieved fast hyperspectral image acquisition with a hybrid camera system constructed from an RGB camera, a grayscale camera, and multiple optical elements including prisms and coding masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 276 + ], + "type": "text", + "content": "Several methods have been proposed to acquire spectral and 3D depth information using dispersive optics. Baek et al. [2] proposed the diffractive optical element (DOE) generates a unique point spread function (PSF) that changes with the scene's spectrum and depth, enabling single-shot hyperspectral 3D imaging through blur analysis. Shin et al. [29] proposed a system to obtain spectral and 3D depth information by analyzing which structured light patterns disperse and illuminate a scene, using a compact setting that adds diffraction grating film to a projector-camera system." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 335 + ], + "type": "text", + "content": "Dispersive optics have been utilized for projector/camera enhancement, hyperspectral image acquisition, and depth estimation. This field holds great potential for various applications. In this study, we demonstrate the effectiveness of direct/indirect separation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 344, + 370, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 370, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 370, + 357 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 364, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 483 + ], + "type": "text", + "content": "We separate the direct and indirect reflection components of a scene from a single-shot hyperspectral image taken under a wavelength-dependent, spatially shifted illumination pattern generated by a DMD (Digital Micromirror Device) projector with a diffraction grating (Figure 1). We first revisit the traditional approach to direct and indirect reflections, utilizing varying high-frequency projected patterns, and then extend the discussion to our case, which involves a wavelength-dependent, spatially shifted illumination pattern." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "type": "text", + "content": "3.1. Direct/Indirect Separation Using Sinusoidal Pattern Illumination" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "text", + "content": "When a single light source illuminates a scene, the direct component " + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "inline_equation", + "content": "I_{d}(x)" + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "text", + "content": " is defined as the light that travels directly from the source, reflects once at a surface point, and reaches the camera, while the indirect component " + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "inline_equation", + "content": "I_{i}(x)" + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "text", + "content": " encompasses all light that undergoes multiple interactions, such as subsurface scattering, interreflections, and volumetric scattering. Nayar et al. [24]. The observed intensity " + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "inline_equation", + "content": "I(x)" + }, + { + "bbox": [ + 313, + 518, + 554, + 615 + ], + "type": "text", + "content": " by a camera is represented as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 387, + 626, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 626, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 387, + 626, + 553, + 639 + ], + "type": "interline_equation", + "content": "I (x) = I _ {d} (x) + I _ {i} (x). \\tag {1}", + "image_path": "8405ac1c9a08ace111e64ca8f2ed447243e6d86a34117fdc2b54b30549f95c6a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Assuming that the indirect reflection component of a scene changes gradually in space and is predominantly low-frequency, their proposed approach [24] demonstrated that the scene's direct and indirect reflection components can be separated by irradiating three spatially high-frequency sinusoidal patterns." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26829" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 74, + 556, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 74, + 556, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 74, + 556, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 74, + 556, + 178 + ], + "type": "image", + "image_path": "14a5663ce8d325f61e68b049760b56f8df24e403ff3e83ef01870b50a816ad07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 188, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 54, + 188, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 54, + 188, + 555, + 234 + ], + "type": "text", + "content": "Figure 2. The outline of the proposed spatio-spectral illumination system. (a) The (DMD) projector generates a sinusoidal pattern that has a broad spectral band of light. (b) The diffraction grating disperses the generated light pattern, then the sinusoidal pattern is spatially shifted according to wavelength. (c) The spatio-spectral shifted patterns are observed at each point of the scene, and the sinusoidal pattern is resampled in the spectral domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "content": "Under the high-frequency sinusoidal pattern illuminations, the observed images " + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "inline_equation", + "content": "I_{n}(x,k)" + }, + { + "bbox": [ + 55, + 264, + 295, + 289 + ], + "type": "text", + "content": " are represented as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 297, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 295, + 321 + ], + "type": "interline_equation", + "content": "I _ {n} (x, k) = \\frac {1}{2} I _ {d} (x) (\\sin (\\frac {2}{3} k \\pi + \\phi) + 1) + \\frac {1}{2} I _ {i} (x), \\quad (2)", + "image_path": "eb22af5760dec937a08d5537359f7da2a325408807edf2bf034d21710dc951d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\phi \\in [0,2\\pi]" + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "text", + "content": " denotes an arbitrary phase of the sinusoidal function and " + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "inline_equation", + "content": "k = \\{1,2,3\\}" + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "text", + "content": " represents the index of the pattern. This problem has three unknowns " + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "inline_equation", + "content": "\\{I_d,I_i,\\phi \\}" + }, + { + "bbox": [ + 55, + 328, + 296, + 425 + ], + "type": "text", + "content": " and can be solved theoretically for each pixel using three images taken under three different sinusoidal pattern illumination. In practice, this approach requires pixel-by-pixel camera-projector alignment or additional patterns to avoid artifacts and ensure accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 244, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 244, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 244, + 445 + ], + "type": "text", + "content": "3.2. Spectral Direct/Indirect Separation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": "Let us further extend the discussion to separation at the wavelength level by considering the spectral reflection of a scene. Here, we represent the direct reflection component " + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "inline_equation", + "content": "I_{d}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": " as the product of normalized spectral reflectance of direct component " + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "inline_equation", + "content": "R_{d}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": ", its magnitude " + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\alpha_{d}(x)" + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": ", and the intensity of the spectral light source " + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "inline_equation", + "content": "L_{d}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": " at the point " + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 450, + 296, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "text", + "content": "Following [24], the indirect reflection component is represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 552, + 295, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 552, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 295, + 579 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I _ {i} (x, \\lambda) = \\sum_ {x ^ {\\prime} \\in S (x)} A ([ x, x ^ {\\prime} ], \\lambda) L ([ x, x ^ {\\prime} ], \\lambda) \\tag {3} \\\\ = \\alpha_ {i} (x) R _ {i} (x, \\lambda), \\\\ \\end{array}", + "image_path": "6bf69aca76055af178ab2c464479298112c2d4689c59f596fcef9d2599559ebd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " denotes the set of surrounding pixels for " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "L([x, x'], \\lambda)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " represents the spectral radiance at " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " in the direction of " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "A([x, x'], \\lambda)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " incorporates the spectral BRDF of " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " as well as the relative geometric configuration between " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": ".1 " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "I_i(x, \\lambda)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " is further expressed in terms of the normalized spectrum " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "R_i(x, \\lambda)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " which is computed from the total sum of light contributions coming to " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " from surrounding pixels " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "S(x)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": " and its magnitude " + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\alpha_i(x)" + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": ". From this," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 552, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 552, + 277 + ], + "type": "text", + "content": "the spectral reflection model of a scene " + }, + { + "bbox": [ + 313, + 264, + 552, + 277 + ], + "type": "inline_equation", + "content": "I(x,\\lambda)" + }, + { + "bbox": [ + 313, + 264, + 552, + 277 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 285, + 561, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 285, + 561, + 321 + ], + "spans": [ + { + "bbox": [ + 318, + 285, + 561, + 321 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I (x, \\lambda) = I _ {d} (x, \\lambda) + I _ {i} (x, \\lambda) \\\\ = \\alpha_ {d} (x) R _ {d} (x, \\lambda) L _ {d} (x, \\lambda) + \\alpha_ {i} (x) R _ {i} (x, \\lambda). \\tag {4} \\\\ \\end{array}", + "image_path": "c950508d48c63469183bb19a51989a2bdf8b7644f60acadf3aece21ccf4ae630.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 554, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 554, + 359 + ], + "type": "text", + "content": "In the following, we introduce a spatial phase shift by designing the lighting pattern to shift horizontally with wavelength." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 554, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 392 + ], + "type": "text", + "content": "3.3. Spatially Shifting Illumination According to Wavelength" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "text", + "content": "We propose a spatio-spectral pattern illumination system that combines a conventional projector and a diffraction grating to create spatially shifting illumination patterns according to wavelength. The configuration of the illumination system is shown in Figure 2. Let the spatially high-frequency sinusoidal pattern " + }, + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "inline_equation", + "content": "P_0(v)" + }, + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "text", + "content": " generated by the conventional projector (Figure 2 (a)) be written as a sinusoidal function of the horizontal pixel position " + }, + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 396, + 554, + 493 + ], + "type": "text", + "content": " of the projector:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 366, + 502, + 554, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 502, + 554, + 526 + ], + "spans": [ + { + "bbox": [ + 366, + 502, + 554, + 526 + ], + "type": "interline_equation", + "content": "P _ {0} (v) = \\frac {l _ {0}}{2} \\left(\\sin \\left(f _ {0} v + \\phi_ {0}\\right) + 1\\right), \\tag {5}", + "image_path": "3edd2d6d14dcf8c1137067de4cd25d54b5156a9919b1f8bea9aea53d3e7669de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " is the intensity of the projector, " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{0}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " describes the frequency of the sinusoidal function, and " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " denotes the phase of the sinusoidal function. Then the grating diffracts the incoming pattern light and reflects it at slightly different angles for each wavelength of light (Figure 2 (b)). In the case of first-order diffracted light, the diffraction angle of the outgoing light, i.e. the emission angle, varies monotonically in the visible light range depending on the wavelength of the incident light. Thus, the illumination pattern varies in the spatio-spectral domain, similar to a rainbow. Therefore, as shown in Figure 3, when the diffracted pattern is observed at a scene point, it is also a sinusoidal pattern in the spectral domain (Figure 2 (c)). We proceed with the discussion assuming that projector's origin is the virtual origin of the pattern reflected by the grating. We also set the cam" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "In our system, the scene brightness has been calibrated by considering the camera's spectral sensitivity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26830" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "era's origin at the same location to enable discussion within a unified coordinate system." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 106, + 291, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 106, + 291, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 291, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 291, + 177 + ], + "type": "image", + "image_path": "d064957930dd665ff8b729a3c4eb8977dc9ba020ab92c81794fb1ecdfc16716b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 296, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 232 + ], + "type": "text", + "content": "Figure 3. After passing through the diffraction grating, the sinusoidal pattern generated by the conventional projector (following yellow arrow) is resampled in the spectral domain at each point of the scene (following red arrow)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "text", + "content": "The spectral pattern " + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "inline_equation", + "content": "P(x, \\lambda)" + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "text", + "content": " is expressed as the function of wavelength " + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "text", + "content": " at the scene point corresponding to the projector's horizontal pixel position " + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 251, + 296, + 289 + ], + "type": "text", + "content": " as in Equation (6):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 292, + 295, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 292, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 295, + 316 + ], + "type": "interline_equation", + "content": "P (x, \\lambda) = \\frac {l (\\lambda)}{2} \\left(\\sin \\left(f _ {p} (x) \\lambda + \\phi (x)\\right) + 1\\right). \\tag {6}", + "image_path": "ac4bc47bdcf818c3bcfb48062e8e509bf30e612b4e495463af9306dae6b4d712.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "l(\\lambda)" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " be the spectral intensity of the projector, " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "f_{p}(x)" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " denote the frequency of the sinusoidal function, and " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "\\phi (x)" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " represent the phase of the sinusoidal function. We calibrate the spectral intensity of the projector " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "l(\\lambda)" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": " in advance and describe it as the constant value " + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 319, + 296, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 137, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 137, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 137, + 397 + ], + "type": "text", + "content": "3.4. Simple Case" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": "We begin our discussion with the simplest case: illuminating our pattern onto a gray object with flat spectral reflectance and spatially varying albedo. In this case, the normalized spectra " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "R_{d}" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " of the direct and indirect components can be integrated and represented as the respective magnitudes " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": ". When illuminating our pattern onto this object, the direct light component " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "L_{d}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " in Equation (4) is replaced by " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "P(x,\\lambda)" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": ", and the indirect reflection component " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "I_{i}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " is reduced to half intensity and pattern independent following [24]. Hence, the observed reflection " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "I_{p}(x,\\lambda)" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " under the spatio-spectral pattern illumination " + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "inline_equation", + "content": "P(x,\\lambda)" + }, + { + "bbox": [ + 55, + 401, + 296, + 546 + ], + "type": "text", + "content": " is modeled as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 549, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 549, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 100, + 549, + 295, + 572 + ], + "type": "interline_equation", + "content": "I _ {p} (x, \\lambda) = \\alpha_ {d} (x) P (x, \\lambda) + \\frac {1}{2} \\alpha_ {i} (x). \\tag {7}", + "image_path": "497adda5b19c994f9f4076b569698bb205bc962193e17256ec777326c6d8009f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 274, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 274, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 274, + 588 + ], + "type": "text", + "content": "Expanding Equation (7) based on Equation (6) yields:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 591, + 295, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 591, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 61, + 591, + 295, + 615 + ], + "type": "interline_equation", + "content": "I _ {p} (x, \\lambda) = \\frac {l \\alpha_ {d} (x)}{2} (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) + \\frac {\\alpha_ {i} (x)}{2}. \\tag {8}", + "image_path": "c6826f8f3aae7a4f33752b332d10b1e905caa2c04a9b0d4d04dfb03c18b88f15.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": "In this case, we can obtain a sinusoidal pattern in the spectral domain along wavelength axis at any point " + }, + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": " as shown in Figure 4 (a)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Since the number of unknowns is four " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{d}(x),f_{p}(x),\\phi (x),\\alpha_{i}(x)\\}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " , the direct and indirect reflection components of a scene can be separated by acquiring spectral images at least 4 spectral channels in the simple case." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 382, + 140 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 382, + 140 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 382, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 382, + 140 + ], + "type": "image", + "image_path": "5e46589b634ec966a4df529ed3d186e4536c3caf47cfc6002ff72a02551e00a1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 209 + ], + "type": "text", + "content": "Figure 4. Our illumination pattern and the observation as it is projected onto the scene. (a) For gray objects, a sinusoidal pattern is observed in the spectral domain. (b) For colored objects, a pattern that varies with the spectral reflectance of the direct and indirect components of the object is observed in the spectral domain." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 467, + 140 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 467, + 140 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 467, + 140 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 467, + 140 + ], + "type": "image", + "image_path": "f6ece85625d22b329786e6cf62d587ac491bc95261557d072622a400e03247db.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 467, + 72, + 550, + 140 + ], + "blocks": [ + { + "bbox": [ + 467, + 72, + 550, + 140 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 550, + 140 + ], + "type": "image", + "image_path": "c7a0cdda9049d4f793a423b4695d3ffe63b042834e79b05011d413c122730b99.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 239, + 405, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 405, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 405, + 252 + ], + "type": "text", + "content": "3.5. Complex Case" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 257, + 554, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 554, + 328 + ], + "type": "text", + "content": "Next, let us examine the case of the object that has non-flat, spatially varying spectral reflectance. As shown in Figure 4 (b), the spectrum observed for a scene with non-flat spectral reflectance is affected by the spectral reflectance of the scene, unlike the simple flat reflectance case in Figure 4 (a)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "content": "Based on Equation (4), a scene can be composed of direct and indirect reflection components, each of which can have its spectral distribution. The observed intensity " + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "inline_equation", + "content": "I_{p}(x,\\lambda)" + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "content": " at the scene point " + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "content": " under the spatio-spectral pattern illumination " + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "inline_equation", + "content": "P(x,\\lambda)" + }, + { + "bbox": [ + 313, + 329, + 554, + 389 + ], + "type": "text", + "content": " can be expressed:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 399, + 555, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 555, + 423 + ], + "type": "interline_equation", + "content": "I _ {p} (x, \\lambda) = \\alpha_ {d} (x) R _ {d} (x, \\lambda) P (x, \\lambda) + \\frac {\\alpha_ {i} (x)}{2} R _ {i} (x, \\lambda) \\tag {9}", + "image_path": "c30057ec8f8ae399544e6cbffdecec76ab3388ceb8556cce53be35e00199d288.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 430, + 533, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 533, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 533, + 443 + ], + "type": "text", + "content": "Expanding Equation (9) based on Equation (6) yields:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 452, + 561, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 452, + 561, + 487 + ], + "spans": [ + { + "bbox": [ + 318, + 452, + 561, + 487 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I _ {p} (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) (\\sin (f _ {p} (x) \\lambda + \\phi (x)) + 1) / 2 \\\\ + \\quad \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2. \\tag {10} \\\\ \\end{array}", + "image_path": "89ab8793927f8839300a3a0fb992b11b2071200c75c78000e976ee2ca9624df4.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 488, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 512 + ], + "type": "text", + "content": "Furthermore, Equation (10) can be expressed in the form of a general sinusoidal function for " + }, + { + "bbox": [ + 313, + 488, + 555, + 512 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 488, + 555, + 512 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 318, + 521, + 555, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 521, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 318, + 521, + 555, + 544 + ], + "type": "interline_equation", + "content": "I_{p}(x,\\lambda) = A(x,\\lambda)\\{\\sin (B(x)\\lambda +C(x)) + 1\\} +D(x,\\lambda),", + "image_path": "00a01e8ffd17225782ed505398d86de1b99cbbf5056e540385e31d5ef1255738.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 345, + 546, + 529, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 546, + 529, + 559 + ], + "spans": [ + { + "bbox": [ + 345, + 546, + 529, + 559 + ], + "type": "interline_equation", + "content": "A (x, \\lambda) = l \\alpha_ {d} (x) R _ {d} (x, \\lambda) / 2, B (x) = f _ {p} (x),", + "image_path": "3ddcc0b3db6dc387ccd69398d9860717b18d6d8acf03eaea6f149a3adb1851ab.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 350, + 559, + 525, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 559, + 525, + 571 + ], + "spans": [ + { + "bbox": [ + 350, + 559, + 525, + 571 + ], + "type": "interline_equation", + "content": "C (x) = \\phi (x), D (x, \\lambda) = \\alpha_ {i} (x) R _ {i} (x, \\lambda) / 2.", + "image_path": "8e26c9909cad50b80424fe2fa41fd73be03db1a88090ee953a5da4827a54174e.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 535, + 571, + 553, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 535, + 571, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 535, + 571, + 553, + 580 + ], + "type": "text", + "content": "(11)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 616 + ], + "type": "text", + "content": "Equation (11) shows that the amplitude of the sinusoidal function is related to the direct reflection component and the constant term is related to the indirect reflection component." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "We then discuss how to separate the direct and indirect reflection components from the observed values based on Equation (11). The number of observations at each pixel obtained in a single hyperspectral image measurement equals the number of spectral wavelength observations. Therefore, if either spectral reflectance " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "R_{d}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " are unknown, solving Equation (11) from a single multi- or hyperspectral input image is an ill-posed problem. We address this problem" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26831" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "by representing " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "R_{d}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " as the combination of the spectral reflectance basis, following the work of Han et al. [14]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 102, + 295, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 102, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 97, + 102, + 295, + 118 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {R}} _ {\\boldsymbol {d}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {d}} (\\boldsymbol {x}), \\bar {\\boldsymbol {R}} _ {\\boldsymbol {i}} (\\boldsymbol {x}) = \\boldsymbol {B} \\boldsymbol {c} _ {\\boldsymbol {i}} (\\boldsymbol {x}), \\tag {12}", + "image_path": "f541dadb5370e865108dbdc97a2308165ebff6b19b6a8a2e36c39f53fa8b015a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\bar{R}_d(x)" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\bar{R}_i(x) \\in \\mathbb{R}^f" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " are the vectorized form of the normalized spectral reflectances at the pixel " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " approximated by the basis representation, " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{f \\times k}" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " is a basis matrix consisting of " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " basis vectors, and " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "c_d(x)" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "c_i(x) \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " are the basis coefficients at the pixel " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\bar{R}_i(x)" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " is not pure spectral reflectance, but it is the sum of the spectral BRDFs in the surrounding pixels and can therefore be expressed on the same basis. Based on the public dataset [28], in which the 1257 spectral reflectances of Munsell Book of Color: Matte Finish Collection [6] are recorded, " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " spectral reflectance basis vectors are extracted using singular value decomposition. Approximation by the basis reduces the number of unknowns related to the spectral reflectance to " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": ". Thus, unknowns of our problem are " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\{c_d(x), c_i(x), \\alpha_d(x), f_p(x), \\phi(x), \\alpha_i(x)\\}" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": " and its total degree of freedom is " + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "inline_equation", + "content": "2k + 4" + }, + { + "bbox": [ + 55, + 123, + 296, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "type": "text", + "content": "3.6. Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "text", + "content": "We solve the problem described in Section 3.5 by optimization, using a single hyperspectral image captured by our spatio-spectral pattern illumination system as input. For simplicity, let us focus on a single pixel, with all terms written while omitting " + }, + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "text", + "content": ". The error function between the input spectral image and the model is computed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 295, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 450 + ], + "type": "interline_equation", + "content": "E _ {s} (\\lambda) = \\left| I _ {p} (\\lambda) - \\frac {l \\alpha_ {d}}{2} \\bar {R} _ {d} (\\lambda) \\left\\{\\sin \\left(f _ {p} \\lambda + \\phi\\right) + 1 \\right\\} - \\frac {\\alpha_ {i}}{2} \\bar {R} _ {i} (\\lambda) \\right|. \\tag {13}", + "image_path": "65cb7978fc7db77bfa35a9a6ea8509e77ec45d8bb92866d0461ff2fdee2353d1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\bar{R}_d(\\lambda)" + }, + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\bar{R}_i(\\lambda)" + }, + { + "bbox": [ + 55, + 451, + 296, + 487 + ], + "type": "text", + "content": " denote the scalar form of the normalized spectral reflectances approximated by the basis representation. The objective function of the problem is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 495, + 295, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 495, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 119, + 495, + 295, + 528 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, \\alpha_ {d}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda). \\tag {14}", + "image_path": "dc3b040096a8c3ed6bf06a85d32e9d2328f9d41f31b5412e2e54019783977f0a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": "As mentioned earlier, our problem setup has " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "2k + 4" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " unknowns, so the simplest solution is to give a hyperspectral image with " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "2k + 4" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " spectral channels as the input image. From previous study [14], the number of bases for approximating the spectral reflectance of a general object with a basis representation is about 8. Therefore, when we follow them and let " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": ", we require a hyperspectral image with 20 spectral channels to solve. When applying hyperspectral imaging to digital archiving and other applications, it is common to work in visible light, with wavelengths ranging from " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "400\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "inline_equation", + "content": "700\\mathrm{nm}" + }, + { + "bbox": [ + 55, + 534, + 295, + 689 + ], + "type": "text", + "content": ". Therefore, acquiring multi- or hyperspectral images with a wavelength interval of at least 15 nm in the visible light range is sufficient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "For robust optimization, we use initial values and ranges of estimation values for some of the parameters. As shown" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": "in the figure, the period of the sinusoidal function of the spectral domain at each pixel can be predicted from the local maxima and minima of the observed signal. The period obtained from the median of the intervals at the local maxima and minima is the initial value of the period " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " at that pixel, and the search range for optimization is " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\pm 5\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " of the initial value to account for discretization error. Also, the search range of the phase " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "[-\\pi, \\pi]" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " and the magnitude of the direct and indirect components " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\alpha_{d}" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "inline_equation", + "content": "[0, \\infty]" + }, + { + "bbox": [ + 313, + 72, + 555, + 182 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 194, + 551, + 251 + ], + "blocks": [ + { + "bbox": [ + 318, + 194, + 551, + 251 + ], + "lines": [ + { + "bbox": [ + 318, + 194, + 551, + 251 + ], + "spans": [ + { + "bbox": [ + 318, + 194, + 551, + 251 + ], + "type": "image", + "image_path": "a1bb872fc1433ff66bfc24296c2926eb3f5d29862117543c801aac4b8aaacafc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 262, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 285 + ], + "type": "text", + "content": "Figure 5. Examples of the observed signal in the spectral domain. Orange and green circles represent local maxima and minima." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "content": "Furthermore, since spectral reflectance is known to be smooth for most common objects [13], we introduce a loss function that reflects the assumption of smoothness of spectral reflectance, following prior works [19, 32]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 368, + 554, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 368, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 323, + 368, + 554, + 400 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {c} _ {d}, \\boldsymbol {c} _ {i}, f _ {p}, \\phi , \\alpha_ {i}} \\sum_ {\\lambda} ^ {k} E _ {s} (\\lambda) + \\gamma_ {d} | \\boldsymbol {B c} _ {d} \\boldsymbol {\\Gamma} | + \\gamma_ {i} | \\boldsymbol {B c} _ {i} \\boldsymbol {\\Gamma} |. \\tag {15}", + "image_path": "dc2f8d3f70503be8f0198ab21a0570381aa4707da824952a9309a4b5cc2eeb5b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": "The weights " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\gamma_{d}" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\gamma_{i}" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " are chosen empirically and " + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 410, + 554, + 446 + ], + "type": "text", + "content": " denotes a matrix that gives smoothness by the subtraction between the neighbor elements." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 459, + 440, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 440, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 440, + 472 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 479, + 504, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 504, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 504, + 491 + ], + "type": "text", + "content": "4.1. Experimental Setup and Conditions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": "Our system consists of a xenon light source (Asahi Spectra LAX-100), a DMD (Ajile AJD-4500), a hyperspectral camera (EBA JAPAN NH-8), lenses for the DMD and hyperspectral camera (Fujinon HF25SA-1), and a diffraction grating (Edmund Optics 1200 Grooves, " + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "inline_equation", + "content": "50\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": " Square, " + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "inline_equation", + "content": "500\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": " Ruled Diffraction Grating), as shown in Figure 1. The hyperspectral camera captures spectral images at " + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "inline_equation", + "content": "5\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": " intervals within the " + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "inline_equation", + "content": "380 - 1000\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": " wavelength range. Considering the effective wavelength range of the diffraction grating, we performed separation in the " + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "inline_equation", + "content": "450 - 650\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 497, + 555, + 617 + ], + "type": "text", + "content": " range." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "The minimum number of observations required to solve our problem is 20, as discussed in Section 3.6. However, actual spectral images contain significant noise, and factors such as projector and camera calibration errors further impact accuracy, preventing us from achieving stable results with the minimum number of observations. In the following experiments, we use 41 spectral bands obtained at " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "5\\mathrm{nm}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " intervals to ensure robust separation against noise." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26832" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "type": "text", + "content": "4.2. Results on Real Scenes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 269 + ], + "type": "text", + "content": "Figure 8 shows the results of estimating the spectral reflectance of the direct and indirect reflectance components in various real-world objects. For reference, we captured a set of images of the scene using a hyperspectral camera with different sinusoidal pattern illuminations and provided the separation results using Nayar's method [24], as described in Equation (2). Hyperspectral images contain significant noise, making it challenging to obtain stable separation results free from artifacts by Nayar's method [24] unless numerous hyperspectral images are captured under various sinusoidal pattern illuminations. To mitigate such artifacts, the spatial resolution of the input images for the reference results, but not ours, is downsampled for visual comparison. All input hyperspectral images and estimated results are available in the supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": "For all four scenes in Figure 8, the estimated direct " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "(I_d)" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " and indirect " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "(I_i)" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " components closely resemble the corresponding reference across different wavelengths. From these examples, we observe that wavelength-dependent changes in the direct and indirect components are extracted. In particular, in the results for the Potato and Ball (top 2 rows) and the Candle (bottom 2 rows), the scattering inside the objects becomes broader and stronger for longer wavelengths. Fine textures can be seen in the direct reflection component and rarely in the indirect reflection component. The leftmost graph compares our estimation and reference for points " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 269, + 296, + 474 + ], + "type": "text", + "content": ". The estimated spectra well represent that of the reference; nevertheless, they are derived solely from a single hyperspectral image. Moreover, Figure 6 shows spatial maps and values of error metrics. We used MSE (mean squared error) and SAM (spectral angle mapper) as error metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 482, + 275, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 275, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 275, + 495 + ], + "type": "text", + "content": "4.3. Depth Dependence of the Pattern's Period" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "text", + "content": "Figure 7 shows the spatial distribution of the estimated period " + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "text", + "content": " in Equation (15). The period " + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "text", + "content": " increases as the distance from the projector becomes longer, making it highly correlated with object distance. In Figure 7, 'DAY' is placed closer to the camera, and 'HAP' is farther away. This example demonstrates that the estimated " + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 500, + 296, + 584 + ], + "type": "text", + "content": " contains valuable information about the scene's depth." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "In this work, we proposed a novel method for separating the direct and indirect reflection components of a scene from a single hyperspectral image. By introducing a projection system incorporating dispersive optics, we enabled the projection of a wavelength-dependent, spatially shifted pattern. We formulated the observed pixel values under these spectral-dependent illumination patterns and derived a solution for direct and indirect reflection separation at each" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 68, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 553, + 319 + ], + "type": "image", + "image_path": "a92b5762de635944f01719392b9a1a340c309b9151352a3a1bfd5bffc1a21b24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 371 + ], + "type": "text", + "content": "Figure 6. Error maps and metrics between the reference and the proposed method. The values represent the mean and variance of the MSE and SAM. Note that some stripe patterns in the SAM maps are affected by the artifacts of the reference." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 396, + 553, + 496 + ], + "blocks": [ + { + "bbox": [ + 321, + 396, + 553, + 496 + ], + "lines": [ + { + "bbox": [ + 321, + 396, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 321, + 396, + 553, + 496 + ], + "type": "image", + "image_path": "32579dbc8e51b3945fb1fc27c38a51786605a9040eb48fb54323874eaceb8086.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 503, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 526 + ], + "type": "text", + "content": "Figure 7. Spatial map of the estimated period " + }, + { + "bbox": [ + 313, + 503, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 503, + 555, + 526 + ], + "type": "text", + "content": " for a scene. 'DAY' is placed closer to the camera, and 'HAP' is farther away." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 559, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 643 + ], + "type": "text", + "content": "wavelength using only a single hyperspectral image with one projection pattern. Through experiments, we demonstrated the effectiveness of our system in achieving robust spectral direct/indirect separation even for objects with complex spectral reflectance. Furthermore, our analysis of differences across wavelengths was leveraged for scene depth estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 656, + 411, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 656, + 411, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 656, + 411, + 670 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "content": "This research is supported by JSPS Kakenhi Grant number 25K21236 and JST-Mirai Program Grant Number JP-MJM123G1." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26833" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 521, + 622 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 521, + 622 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 521, + 622 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 521, + 622 + ], + "type": "image", + "image_path": "15acc3230c29c7aa979a14b3f2ef66c2d15b148061df512cf918d9d524a3433e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": "Figure 8. Separation results of direct and indirect reflection components in various real-world objects. Top row, from left to right: scene captured by the hyperspectral camera (RGB converted), results of separation of direct and indirect reflection components " + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "inline_equation", + "content": "(I_d(480), I_d(620), I_i(480), I_i(620))" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": " by Nayar's method [24] as a reference, direct reflectance component (red) at point " + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": " and indirect, comparison of spectral reflectance of direct (red) and indirect (green) components at point " + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": " for the reference (dotted line) and the proposed method (solid line). Bottom row, from left to right: part of the input spectral images in the proposed method are superimposed on a pseudo-RGB, estimation results of direct and indirect reflection components " + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "inline_equation", + "content": "(I_d(480), I_d(620), I_i(480), I_i(620))" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": " by the proposed method, comparison of spectral reflectance at " + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 55, + 643, + 555, + 721 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26834" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Gonzalo R Arce, David J Brady, Lawrence Carin, Henry Arguello, and David S Kittle. Compressive coded aperture spectral imaging: An introduction. IEEE Signal Processing Magazine, 31(1):105-115, 2013. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Seung-Hwan Baek, Hayato Ikoma, Daniel S Jeon, Yuqi Li, Wolfgang Heidrich, Gordon Wetzstein, and Min H Kim. Single-shot hyperspectral-depth imaging with learned diffractive optics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2651-2660, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "text", + "content": "[3] Xun Cao, Xin Tong, Qionghai Dai, and Stephen Lin. High resolution multispectral video capture with a hybrid camera system. In CVPR 2011, pages 297-304. IEEE, 2011. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "type": "text", + "content": "[4] Xavier Ceamanos and Silvia Valero. 4 - processing hyperspectral images. In Optical Remote Sensing of Land Surface, pages 163-200. Elsevier, 2016. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 316 + ], + "type": "text", + "content": "[5] Xiaochuan Chen, Mark S Drew, and Ze-Nian Li. Illumination and reflectance spectra separation of hyperspectral image data under multiple illumination conditions. *Electronic Imaging*, 29:194–199, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "type": "text", + "content": "[6] Munsell Color. Munsell book of color: Matte finish collection: Kollmorgen corporation. Baltimore, Maryland, 1976. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "text", + "content": "[7] Costanza Cucci, John K Delaney, and Marcello Picollo. Reflectance hyperspectral imaging for investigation of works of art: old master paintings and illuminated manuscripts. Accounts of chemical research, 49(10):2070-2079, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 295, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 428 + ], + "type": "text", + "content": "[8] Craig Donner and Henrik Wann Jensen. A spectral bssrfd for shading human skin. Rendering techniques, 2006:409-418, 2006. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 474 + ], + "type": "text", + "content": "[9] Zhaoliang Duan, James Bieron, and Pieter Peers. Deep Separation of Direct and Global Components from a Single Photograph under Structured Lighting. Computer Graphics Forum, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 295, + 509 + ], + "type": "text", + "content": "[10] Jonathan Dupuy and Wenzel Jakob. An adaptive parameterization for efficient material acquisition and rendering. ACM Transactions on graphics (TOG), 37(6):1-14, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 553 + ], + "type": "text", + "content": "[11] Gamal ElMasry, Pere Gou, and Salim Al-Rejaie. Effectiveness of specularity removal from hyperspectral images on the quality of spectral signatures of food products. Journal of Food Engineering, 289:110148, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 610 + ], + "type": "text", + "content": "[12] Ying Fu, Antony Lam, Imari Sato, Takahiro Okabe, and Yoichi Sato. Separating reflective and fluorescent components using high frequency illumination in the spectral domain. In Proceedings of the IEEE International Conference on Computer Vision, pages 457-464, 2013. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 633 + ], + "type": "text", + "content": "[13] Lewis D Griffin. Reconciling the statistics of spectral reflectance and colour. PLoS One, 14(11):e0223069, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[14] Shuai Han, Imari Sato, Takahiro Okabe, and Yoichi Sato. Fast spectral reflectance recovery using dlp projector. International journal of computer vision, 110:172-184, 2014. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[15] Rafael Hostettler, Ralf Habel, Markus Gross, and Wojciech Jarosz. Dispersion-based color projection using masked prisms. In Computer Graphics Forum, pages 329-338. Wiley Online Library, 2015. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[16] Shin Ishihara, Yuta Asano, Yinqiang Zheng, and Imari Sato. Underwater scene recovery using wavelength-dependent refraction of light. In 2020 International Conference on 3D Vision (3DV), pages 32-40. IEEE, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[17] Min H Kim, Todd Alan Harvey, David S Kittle, Holly Rushmeier, Julie Dorsey, Richard O Prum, and David J Brady. 3d imaging spectroscopy for measuring hyperspectral patterns on solid objects. ACM Transactions on Graphics (TOG), 31 (4):1-11, 2012. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "text", + "content": "[18] Paul Kubelka and Franz Munk. An article on optics of paint layers. Z. Tech. Phys, 12(593-601):259-274, 1931. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 195, + 554, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 554, + 238 + ], + "type": "text", + "content": "[19] Wei Liu, Chengxun He, and Le Sun. Spectral-smoothness and non-local self-similarity regularized subspace low-rank learning method for hyperspectral mixed denoising. Remote Sensing, 13(16):3196, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[20] Lembe Samukelo Magwaza and Umezuruike Linus Opara. Analytical methods for determination of sugars and sweetness of horticultural products—a review. Scientia Horticulturae, 184:179-192, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[21] Ankit Mohan, Ramesh Raskar, and Jack Tumblin. Agile spectrum imaging: Programmable wavelength modulation for cameras and projectors. In Computer graphics forum, pages 709-717. Wiley Online Library, 2008. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 360 + ], + "type": "text", + "content": "[22] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE computer graphics and applications, 34(6):57-68, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "text", + "content": "[23] Srinivasa G Narasimhan, Sanjeev J Koppal, and Shuntaro Yamazaki. Temporal dithering of illumination for fast active vision. In Computer Vision-ECCV 2008: 10th European Conference on Computer Vision, Marseille, France, October 12-18, 2008, Proceedings, Part IV 10, pages 830-844. Springer, 2008. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 427, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 554, + 480 + ], + "type": "text", + "content": "[24] Shree K Nayar, Gurunandan Krishnan, Michael D Grossberg, and Ramesh Raskar. Fast separation of direct and global components of a scene using high frequency illumination. In ACM SIGGRAPH 2006 Papers, pages 935-944. 2006. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 482, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 548 + ], + "type": "text", + "content": "[25] Hiroki Okawa, Yinqiang Zheng, Antony Lam, and Imari Sato. Spectral reflectance recovery with interreflection using a hyperspectral image. In Computer Vision-ACCV 2016: 13th Asian Conference on Computer Vision, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part IV 13, pages 52-67. Springer, 2017. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 581 + ], + "type": "text", + "content": "[26] Matthew O'Toole, Ramesh Raskar, and Kiriakos N Kutulakos. Primal-dual coding to probe light transport. ACM Trans. Graph., 31(4):39-1, 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 554, + 624 + ], + "type": "text", + "content": "[27] Matthew O'Toole, Supreeth Achar, Srinivasa G Narasimhan, and Kiriakos N Kutulakos. Homogeneous codes for energy-efficient illumination and imaging. ACM Transactions on Graphics (ToG), 34(4):1-13, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 625, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 555, + 658 + ], + "type": "text", + "content": "[28] Jussi PS Parkkinen, Jarmo Hallikainen, and Timo Jaaskelainen. Characteristic spectra of munsell colors. Journal of the Optical society of America A, 6(2):318-322, 1989. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 659, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 554, + 712 + ], + "type": "text", + "content": "[29] Suhyun Shin, Seokjun Choi, Felix Heide, and Seung-Hwan Baek. Dispersed structured light for hyperspectral 3d imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24997-25006, 2024. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 741, + 317, + 750 + ], + "type": "text", + "content": "26835" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 575 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[30] Art Subpa-asa, Ying Fu, Yinqiang Zheng, Toshiyuki Amano, and Imari Sato. Separating the direct and global components of a single image. Journal of Information Processing, 26: 755-767, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 138 + ], + "type": "text", + "content": "[31] Da-Wen Sun. Hyperspectral imaging for food quality analysis and control. Elsevier, 2010. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[32] Le Sun, Qihao Cheng, and Zhiguo Chen. Hyperspectral image super-resolution method based on spectral smoothing prior and tensor tubal row-sparse representation. Remote Sensing, 14(9):2142, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[33] Robby T Tan, Ko Nishino, and Katsushi Ikeuchi. Separating reflection components based on chromaticity and noise analysis. IEEE transactions on pattern analysis and machine intelligence, 26(10):1373-1379, 2004. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[34] Moriaki Torii, Takahiro Okabe, and Toshiyuki Amano. Multispectral direct-global separation of dynamic scenes. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1923-1931. IEEE, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[35] Mahesh Kumar Tripathi and H Govil. Evaluation of aviris-ng hyperspectral images for mineral identification and mapping. Heliyon, 5(11), 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[36] Ashwin Wagadarikar, Renu John, Rebecca Willett, and David Brady. Single disperser design for coded aperture snapshot spectral imaging. Applied optics, 47(10):B44-B51, 2008. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[37] Yinqiang Zheng, Imari Sato, and Yoichi Sato. Illumination and reflectance spectra separation of a hyperspectral image meets low-rank matrix factorization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1779-1787, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[38] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. Evdig: Event-guided direct and global components separation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9612-9621, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 519 + ], + "type": "text", + "content": "[39] Mustapha Zokay and Hicham Saylani. Removing specular reflection in multispectral dermatological images using blind source separation. In Annual Conference on Medical Image Understanding and Analysis, pages 734-750. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 575 + ], + "type": "text", + "content": "[40] George Zonios, Julie Bykowski, and Nikiforos Kollias. Skin melanin, hemoglobin, and light scattering properties can be quantitatively assessed in vivo using diffuse reflectance spectroscopy. Journal of Investigative Dermatology, 117(6): 1452-1457, 2001. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 742, + 317, + 750 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 742, + 317, + 750 + ], + "spans": [ + { + "bbox": [ + 293, + 742, + 317, + 750 + ], + "type": "text", + "content": "26836" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file